diff --git a/.gitattributes b/.gitattributes index a10611bacfffbb41d357bad88d204618c7788e3a..d43b8ce152afbe5249aaa59dfb15faefbba90001 100644 --- a/.gitattributes +++ b/.gitattributes @@ -212,3 +212,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Attention[[:space:]]Beats[[:space:]]Linear[[:space:]]for[[:space:]]Fast[[:space:]]Implicit[[:space:]]Neural[[:space:]]Representation[[:space:]]Generation/de0aa410-2d32-4dfd-bf03-15bed0a92794_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Attention[[:space:]]Decomposition[[:space:]]for[[:space:]]Cross-Domain[[:space:]]Semantic[[:space:]]Segmentation/67e3719a-47a6-4cf3-aed4-fa4f27d80e9b_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Attention[[:space:]]Prompting[[:space:]]on[[:space:]]Image[[:space:]]for[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/99bf19de-1eec-40d1-9eb4-8fada8842d06_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Attention-Challenging[[:space:]]Multiple[[:space:]]Instance[[:space:]]Learning[[:space:]]for[[:space:]]Whole[[:space:]]Slide[[:space:]]Image[[:space:]]Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/AttentionHand_[[:space:]]Text-driven[[:space:]]Controllable[[:space:]]Hand[[:space:]]Image[[:space:]]Generation[[:space:]]for[[:space:]]3D[[:space:]]Hand[[:space:]]Reconstruction[[:space:]]in[[:space:]]the[[:space:]]Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/AttnZero_[[:space:]]Efficient[[:space:]]Attention[[:space:]]Discovery[[:space:]]for[[:space:]]Vision[[:space:]]Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Audio-Synchronized[[:space:]]Visual[[:space:]]Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Audio-driven[[:space:]]Talking[[:space:]]Face[[:space:]]Generation[[:space:]]with[[:space:]]Stabilized[[:space:]]Synchronization[[:space:]]Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Audio-visual[[:space:]]Generalized[[:space:]]Zero-shot[[:space:]]Learning[[:space:]]the[[:space:]]Easy[[:space:]]Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/AugDETR_[[:space:]]Improving[[:space:]]Multi-scale[[:space:]]Learning[[:space:]]for[[:space:]]Detection[[:space:]]Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/AugUndo_[[:space:]]Scaling[[:space:]]Up[[:space:]]Augmentations[[:space:]]for[[:space:]]Monocular[[:space:]]Depth[[:space:]]Completion[[:space:]]and[[:space:]]Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Augmented[[:space:]]Neural[[:space:]]Fine-tuning[[:space:]]for[[:space:]]Efficient[[:space:]]Backdoor[[:space:]]Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Auto-DAS_[[:space:]]Automated[[:space:]]Proxy[[:space:]]Discovery[[:space:]]for[[:space:]]Training-free[[:space:]]Distillation-aware[[:space:]]Architecture[[:space:]]Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Auto-GAS_[[:space:]]Automated[[:space:]]Proxy[[:space:]]Discovery[[:space:]]for[[:space:]]Training-free[[:space:]]Generative[[:space:]]Architecture[[:space:]]Search/d262e2a7-fb7b-49f5-8763-ff8884371602_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/AutoDIR_[[:space:]]Automatic[[:space:]]All-in-One[[:space:]]Image[[:space:]]Restoration[[:space:]]with[[:space:]]Latent[[:space:]]Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/AutoEval-Video_[[:space:]]An[[:space:]]Automatic[[:space:]]Benchmark[[:space:]]for[[:space:]]Assessing[[:space:]]Large[[:space:]]Vision[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Open-Ended[[:space:]]Video[[:space:]]Question[[:space:]]Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Avatar[[:space:]]Fingerprinting[[:space:]]for[[:space:]]Authorized[[:space:]]Use[[:space:]]of[[:space:]]Synthetic[[:space:]]Talking-Head[[:space:]]Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/AvatarPose_[[:space:]]Avatar-guided[[:space:]]3D[[:space:]]Pose[[:space:]]Estimation[[:space:]]of[[:space:]]Close[[:space:]]Human[[:space:]]Interaction[[:space:]]from[[:space:]]Sparse[[:space:]]Multi-view[[:space:]]Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BAD-Gaussians_[[:space:]]Bundle[[:space:]]Adjusted[[:space:]]Deblur[[:space:]]Gaussian[[:space:]]Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BAFFLE_[[:space:]]A[[:space:]]Baseline[[:space:]]of[[:space:]]Backpropagation-Free[[:space:]]Federated[[:space:]]Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BAGS_[[:space:]]Blur[[:space:]]Agnostic[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]through[[:space:]]Multi-Scale[[:space:]]Kernel[[:space:]]Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BAM-DETR_[[:space:]]Boundary-Aligned[[:space:]]Moment[[:space:]]Detection[[:space:]]Transformer[[:space:]]for[[:space:]]Temporal[[:space:]]Sentence[[:space:]]Grounding[[:space:]]in[[:space:]]Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BAMM_[[:space:]]Bidirectional[[:space:]]Autoregressive[[:space:]]Motion[[:space:]]Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BEAF_[[:space:]]Observing[[:space:]]BEfore-AFter[[:space:]]Changes[[:space:]]to[[:space:]]Evaluate[[:space:]]Hallucination[[:space:]]in[[:space:]]Vision-language[[:space:]]Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BI-MDRG_[[:space:]]Bridging[[:space:]]Image[[:space:]]History[[:space:]]in[[:space:]]Multimodal[[:space:]]Dialogue[[:space:]]Response[[:space:]]Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BK-SDM_[[:space:]]A[[:space:]]Lightweight,[[:space:]]Fast,[[:space:]]and[[:space:]]Cheap[[:space:]]Version[[:space:]]of[[:space:]]Stable[[:space:]]Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BKDSNN_[[:space:]]Enhancing[[:space:]]the[[:space:]]Performance[[:space:]]of[[:space:]]Learning-based[[:space:]]Spiking[[:space:]]Neural[[:space:]]Networks[[:space:]]Training[[:space:]]with[[:space:]]Blurred[[:space:]]Knowledge[[:space:]]Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BLINK_[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Can[[:space:]]See[[:space:]]but[[:space:]]Not[[:space:]]Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BRAVE_[[:space:]]Broadening[[:space:]]the[[:space:]]visual[[:space:]]encoding[[:space:]]of[[:space:]]vision-language[[:space:]]models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BRIDGE_[[:space:]]Bridging[[:space:]]Gaps[[:space:]]in[[:space:]]Image[[:space:]]Captioning[[:space:]]Evaluation[[:space:]]with[[:space:]]Stronger[[:space:]]Visual[[:space:]]Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BaSIC_[[:space:]]BayesNet[[:space:]]Structure[[:space:]]Learning[[:space:]]for[[:space:]]Computational[[:space:]]Scalable[[:space:]]Neural[[:space:]]Image[[:space:]]Compression/360699b9-a034-4ed5-b537-07332b36832e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Background[[:space:]]Adaptation[[:space:]]with[[:space:]]Residual[[:space:]]Modeling[[:space:]]for[[:space:]]Exemplar-Free[[:space:]]Class-Incremental[[:space:]]Semantic[[:space:]]Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bad[[:space:]]Students[[:space:]]Make[[:space:]]Great[[:space:]]Teachers_[[:space:]]Active[[:space:]]Learning[[:space:]]Accelerates[[:space:]]Large-Scale[[:space:]]Visual[[:space:]]Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bayesian[[:space:]]Detector[[:space:]]Combination[[:space:]]for[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]Crowdsourced[[:space:]]Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bayesian[[:space:]]Evidential[[:space:]]Deep[[:space:]]Learning[[:space:]]for[[:space:]]Online[[:space:]]Action[[:space:]]Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bayesian[[:space:]]Self-Training[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]3D[[:space:]]Segmentation/546e4382-dd78-463e-8d89-435110642a17_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Be[[:space:]]Yourself_[[:space:]]Bounded[[:space:]]Attention[[:space:]]for[[:space:]]Multi-Subject[[:space:]]Text-to-Image[[:space:]]Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Be-Your-Outpainter_[[:space:]]Mastering[[:space:]]Video[[:space:]]Outpainting[[:space:]]through[[:space:]]Input-Specific[[:space:]]Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BeNeRF_Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Blurry[[:space:]]Image[[:space:]]and[[:space:]]Event[[:space:]]Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Beat-It_[[:space:]]Beat-Synchronized[[:space:]]Multi-Condition[[:space:]]3D[[:space:]]Dance[[:space:]]Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BenchLMM_[[:space:]]Benchmarking[[:space:]]Cross-style[[:space:]]Visual[[:space:]]Capability[[:space:]]of[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Benchmarking[[:space:]]Object[[:space:]]Detectors[[:space:]]with[[:space:]]COCO_[[:space:]]A[[:space:]]New[[:space:]]Path[[:space:]]Forward/64864c4f-d7d3-4bce-8017-e14998347c20_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Benchmarking[[:space:]]Spurious[[:space:]]Bias[[:space:]]in[[:space:]]Few-Shot[[:space:]]Image[[:space:]]Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Benchmarking[[:space:]]the[[:space:]]Robustness[[:space:]]of[[:space:]]Cross-view[[:space:]]Geo-localization[[:space:]]Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Benchmarks[[:space:]]and[[:space:]]Challenges[[:space:]]in[[:space:]]Pose[[:space:]]Estimation[[:space:]]for[[:space:]]Egocentric[[:space:]]Hand[[:space:]]Interactions[[:space:]]with[[:space:]]Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Beta-Tuned[[:space:]]Timestep[[:space:]]Diffusion[[:space:]]Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Betrayed[[:space:]]by[[:space:]]Attention_[[:space:]]A[[:space:]]Simple[[:space:]]yet[[:space:]]Effective[[:space:]]Approach[[:space:]]for[[:space:]]Self-supervised[[:space:]]Video[[:space:]]Object[[:space:]]Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Better[[:space:]]Call[[:space:]]SAL_[[:space:]]Towards[[:space:]]Learning[[:space:]]to[[:space:]]Segment[[:space:]]Anything[[:space:]]in[[:space:]]Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Better[[:space:]]Regression[[:space:]]Makes[[:space:]]Better[[:space:]]Test-time[[:space:]]Adaptive[[:space:]]3D[[:space:]]Object[[:space:]]Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Beyond[[:space:]]MOT_[[:space:]]Semantic[[:space:]]Multi-Object[[:space:]]Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Beyond[[:space:]]Pixels_[[:space:]]Semi-Supervised[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]with[[:space:]]a[[:space:]]Multi-scale[[:space:]]Patch-based[[:space:]]Multi-Label[[:space:]]Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Beyond[[:space:]]Prompt[[:space:]]Learning_[[:space:]]Continual[[:space:]]Adapter[[:space:]]for[[:space:]]Efficient[[:space:]]Rehearsal-Free[[:space:]]Continual[[:space:]]Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Beyond[[:space:]]Viewpoint_[[:space:]]Robust[[:space:]]3D[[:space:]]Object[[:space:]]Recognition[[:space:]]under[[:space:]]Arbitrary[[:space:]]Views[[:space:]]through[[:space:]]Joint[[:space:]]Multi-Part[[:space:]]Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Beyond[[:space:]]the[[:space:]]Contact_[[:space:]]Discovering[[:space:]]Comprehensive[[:space:]]Affordance[[:space:]]for[[:space:]]3D[[:space:]]Objects[[:space:]]from[[:space:]]Pre-trained[[:space:]]2D[[:space:]]Diffusion[[:space:]]Models/4cf86d6f-6634-4c59-b4d4-c22212431643_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Beyond[[:space:]]the[[:space:]]Data[[:space:]]Imbalance_[[:space:]]Employing[[:space:]]the[[:space:]]Heterogeneous[[:space:]]Datasets[[:space:]]for[[:space:]]Vehicle[[:space:]]Maneuver[[:space:]]Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BeyondScene_[[:space:]]Higher-Resolution[[:space:]]Human-Centric[[:space:]]Scene[[:space:]]Generation[[:space:]]With[[:space:]]Pretrained[[:space:]]Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bi-TTA_[[:space:]]Bidirectional[[:space:]]Test-Time[[:space:]]Adapter[[:space:]]for[[:space:]]Remote[[:space:]]Physiological[[:space:]]Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bi-directional[[:space:]]Contextual[[:space:]]Attention[[:space:]]for[[:space:]]3D[[:space:]]Dense[[:space:]]Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bidirectional[[:space:]]Progressive[[:space:]]Transformer[[:space:]]for[[:space:]]Interaction[[:space:]]Intention[[:space:]]Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bidirectional[[:space:]]Stereo[[:space:]]Image[[:space:]]Compression[[:space:]]with[[:space:]]Cross-Dimensional[[:space:]]Entropy[[:space:]]Model/e50db183-7c10-4237-a52b-c78459d9ff33_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bidirectional[[:space:]]Uncertainty-Based[[:space:]]Active[[:space:]]Learning[[:space:]]for[[:space:]]Open-Set[[:space:]]Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Binomial[[:space:]]Self-compensation[[:space:]]for[[:space:]]Motion[[:space:]]Error[[:space:]]in[[:space:]]Dynamic[[:space:]]3D[[:space:]]Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BlazeBVD_[[:space:]]Make[[:space:]]Scale-Time[[:space:]]Equalization[[:space:]]Great[[:space:]]Again[[:space:]]for[[:space:]]Blind[[:space:]]Video[[:space:]]Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BlenderAlchemy_[[:space:]]Editing[[:space:]]3D[[:space:]]Graphics[[:space:]]with[[:space:]]Vision-Language[[:space:]]Models/19f5daea-7c9c-46d5-b550-8a14478e364b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Blind[[:space:]]Image[[:space:]]Deconvolution[[:space:]]by[[:space:]]Generative-based[[:space:]]Kernel[[:space:]]Prior[[:space:]]and[[:space:]]Initializer[[:space:]]via[[:space:]]Latent[[:space:]]Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Blind[[:space:]]image[[:space:]]deblurring[[:space:]]with[[:space:]]noise-robust[[:space:]]kernel[[:space:]]estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BlinkVision_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Optical[[:space:]]Flow,[[:space:]]Scene[[:space:]]Flow[[:space:]]and[[:space:]]Point[[:space:]]Tracking[[:space:]]Estimation[[:space:]]using[[:space:]]RGB[[:space:]]Frames[[:space:]]and[[:space:]]Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_content_list.json b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5e7423e3a3593c69f9c091e640e70a64579afbc6 --- /dev/null +++ b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_content_list.json @@ -0,0 +1,1828 @@ +[ + { + "type": "text", + "text": "Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification", + "text_level": 1, + "bbox": [ + 218, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunlong Zhang $^{1,2,3}$ , Honglin Li $^{1,2,3}$ , Yunxuan Sun $^{1,2,3}$ , Sunyi Zheng $^{2,3}$ , Chenglu Zhu $^{2,3}$ , and Lin Yang $^{2,3\\star}$", + "bbox": [ + 245, + 210, + 754, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University \n $^{2}$ Research Center for Industries of the Future, Westlake University \n $^{3}$ School of Engineering, Westlake University", + "bbox": [ + 274, + 253, + 725, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In the application of Multiple Instance Learning (MIL) methods for Whole Slide Image (WSI) classification, attention mechanisms often focus on a subset of discriminative instances, which are closely linked to overfitting. To mitigate overfitting, we present Attention-Challenging MIL (ACMIL). ACMIL combines two techniques based on separate analyses for attention value concentration. Firstly, UMAP of instance features reveals various patterns among discriminative instances, with existing attention mechanisms capturing only some of them. To remedy this, we introduce Multiple Branch Attention (MBA) to capture more discriminative instances using multiple attention branches. Secondly, the examination of the cumulative value of Top-K attention scores indicates that a tiny number of instances dominate the majority of attention. In response, we present Stochastic Top-K Instance Masking (STKIM), which masks out a portion of instances with Top-K attention values and allocates their attention values to the remaining instances. The extensive experimental results on three WSI datasets with two pre-trained backbones reveal that our ACMIL outperforms state-of-the-art methods. Additionally, through heatmap visualization and UMAP visualization, this paper extensively illustrates ACMIL's effectiveness in suppressing attention value concentration and overcoming the overfitting challenge. The source code is available at https://github.com/dazhangyu123/ACMIL.", + "bbox": [ + 261, + 334, + 743, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Computational pathology $\\cdot$ Whole slide image $\\cdot$ Multiple instance learning $\\cdot$ Overfitting", + "bbox": [ + 261, + 638, + 743, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 693, + 375, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Whole slide image (WSI) classification is a critical undertaking in digital pathology, aiming to extract valuable information from high-resolution scanned images for precise diagnosis [23,33,49,52], prognosis [11,32,55,61], and treatment planning [14,35,37,40,41] of diseases. In recent years, multiple instance learning (MIL) [1,18,38] has emerged as a promising approach for WSI classification, treating each WSI as a \"bag\" and its extracted small patches as \"instances\"", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/413de03d7ce586999586037886ea55821489f8ebaa93793be039324604fa6390.jpg", + "image_caption": [ + "Fig. 1: The change of validation loss and entropy of attention values throughout the training of ABMIL. The results are reported on LBC with SSL pretrained features. There exists the strong negative correlation between loss and entropy." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 491, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0e4b5dfdf12709684e652abda20d7881fe07a5fabafeecf6a87f09ddb0931b52.jpg", + "image_caption": [ + "Fig.2: Comparison of AUC and entropy of attention values between ABMIL and ACMIL. One point denotes the result of a seed on LBC with SSL pretrained features. ACMIL achieves the higher AUC and entropy than ABMIL." + ], + "image_footnote": [], + "bbox": [ + 509, + 142, + 785, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "within the bag, thus enabling efficient classification of WSIs through assigning a single label to the entire slide.", + "bbox": [ + 212, + 398, + 784, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overfitting is a significant challenge in utilizing MIL methods for WSI classification [34, 46, 57]. Common WSI datasets exhibit intrinsic characteristics of limited data scale, ultra-high resolutions, and staining bias, which makes overfitting more likely [2]. Specifically, these datasets often consist of a relatively small number of slides, typically in the hundreds, with a typical resolution of $10,000 \\times 10,000$ [36]. Moreover, pathology images are susceptible to staining bias caused by variations in tissue preparations, staining protocols, and digital scanning methods [58], leading models to learn spurious correlations [34].", + "bbox": [ + 212, + 429, + 785, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the attention mechanism, attention values/heatmaps provide insights into the model's decision-making process. Multiple existing works [36,44,46,56] alongside our own experiments (as indicated in Sec. 4.3) have pointed out the excessive concentration of attention values in current MIL methods. Specifically, MIL's attention mechanisms often concentrate on a subset of discriminative instances (i.e., instances relevant to the bag label) while disregarding the remaining ones. We investigate the correlation between attention value concentration and overfitting, utilizing the entropy of attention values and validation loss. Fig. 1 depicts a negative correlation between loss and entropy throughout the training process, illustrating that over-concentration of attention values (indicated by lower entropy) significantly compromises the model's generalization ability (indicated by higher loss values). Moreover, in the field of natural image classification, recent studies [17,26,48] have demonstrated that models solely relying on a portion of discriminative features could be susceptible to overfitting. Transitioning to WSI classification, fixating on a subset of discriminative instances similarly impedes the model's ability to generalize. These findings highlight the tight connection between attention value concentration and overfitting.", + "bbox": [ + 212, + 551, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, numerous efforts have been made to address the overfitting challenge by enhancing representation quality [16, 24, 31, 36, 51], building spatial", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "instance correlations [21, 30, 44] and developing data augmentation methods [13, 42, 46, 54, 57]. Additionally, some of these studies [30, 36] suggest that reducing the concentration of attention values can enhance model interpretability. However, the investigation of attention values concentration for alleviating overfitting remains under-explored.", + "bbox": [ + 212, + 146, + 784, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate overfitting, we present two analyses for attention value concentration using UMAP and Top-K value statistics. Then, we introduce Attention-Challenging MIL (ACMIL), which combines two novel techniques based on these two analyses. First, by observing UMAP of instance features, we find that there are various patterns among discriminative instances, and attention mechanisms tend to capture some of them. To solve this, we introduce Multiple Branch Attention (MBA). MBA utilizes multiple attention branches, each focusing on capturing instances with a specific pattern, thereby ensuring that more discriminative instances contribute to the final prediction. Second, by analyzing the cumulative value of Top-K attention scores, we find that a tiny number of instances (e.g., $K = 10$ ) occupy majority attention, resulting in overlooking sophisticated discriminative instances. To suppress these instances, we propose Stochastic Top-K Instance Masking (STKIM). STKIM randomly masks out a portion of instances with Top-K attention values and assigns their attention values to the remaining instances. Combining MBA and STKIM, our ACMIL effectively alleviates the attention value concentration and suppresses overfitting (see Fig. 2).", + "bbox": [ + 212, + 222, + 787, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct experiments on three WSI datasets (i.e., CAMELYON16, BRACS, and our in-house LBC dataset) with two backbones (i.e., ImageNet pre-trained ResNet18 and SSL pre-trained ViT/S-16). Experimental results demonstrate the superiority of our ACMIL over existing state-of-the-art methods. We also present substantial experimental results, including heatmap visualization and UMAP visualization, to comprehensively demonstrate the effectiveness of ACMIL in suppressing attention value concentration and combatting overfitting.", + "bbox": [ + 212, + 463, + 792, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 590, + 387, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Combating Overfitting in WSI Analysis", + "text_level": 1, + "bbox": [ + 215, + 619, + 588, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the domain of WSI classification, combating the challenge of overfitting has received substantial attention. Next three paragraphs detail methods from three different aspects.", + "bbox": [ + 212, + 643, + 784, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some efforts have concentrated on enhancing the quality of instance representations. Early studies (e.g., [7, 27, 44]) rely on backbones pre-trained on the ImageNet dataset. However, the substantial domain gap between natural and pathological images hinders representation quality. Recent works (e.g., [16, 24, 36, 47, 51]) address this by emphasizing Self-Supervised Learning (SSL) to learn patch-level feature representations. In addition, efforts such as the work by Chen et al. [10] leverage hierarchical SSL for high-resolution image representations. Further, studies by Li et al. [31] and Wang et al. [50] demonstrate that fine-tuning the pre-trained encoder is essential for acquiring task-specific information.", + "bbox": [ + 212, + 689, + 787, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ACMIL", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another line of research has focused on establishing spatial instance correlations. DSMIL [30], $\\mathrm{H}^2\\mathrm{MIL}$ [25], and DAS-MIL [5] consider the hierarchical structure of patches and aggregate multi-scale representations in attention mechanisms. Furthermore, some studies introduce self-attention layers [43, 44, 53] and graph neural networks [9, 21, 59] to model correlations between different areas.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further strategies have concentrated on data augmentation. Examples include DTFD-MIL [57], which introduces pseudo-bags for expanding bag counts and employs a double-tier MIL framework. IPS [4], Zoom-In Network [29], and Top-K MIL [13] generate bag representations by aggregating the representations of salient patches. Remix [54] and RankMix [12] introduce instance representation mixup for MIL. MHIM-MIL [46] and WENO [42] augments bags by randomly masking salient instances.", + "bbox": [ + 212, + 222, + 784, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although our ACMIL shares a similar spirit with some of these works, the proposed ACMIL excels by further building on detailed analysis for attention value concentration. As a result, our ACMIL exhibits stronger interpretability against existing solutions.", + "bbox": [ + 212, + 327, + 785, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Over-Concentration of Attention Values", + "text_level": 1, + "bbox": [ + 214, + 409, + 591, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the realm of natural image classification, research has shown that an excessive focus on certain parts of an object can impede the overall effectiveness of model generalization [17,26,60]. To tackle this issue, various heuristic techniques have been proposed. For instance, Cutout [17, 60] is a valuable data augmentation method that randomly masks square regions of input during training. Another approach, RSC [26], involves regularization that eliminates salient features activated during training. This paper investigates the issue of attention value concentration in WSI classification tasks. We identify two specific phenomena related to attention value concentration existing in WSI classification and propose two techniques to address them respectively.", + "bbox": [ + 212, + 431, + 787, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 603, + 330, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the ABMIL (detailed in Sec. 3.1), we present ACMIL to alleviate the overfitting problem, which is built on two components: Multiple Branch Attention (MBA) and Stochastic Top-K Instance Masking (STKIM). We describe the details of two components in Sec. 3.2 and 3.3, respectively.", + "bbox": [ + 212, + 632, + 784, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 ABMIL for WSI Classification", + "text_level": 1, + "bbox": [ + 214, + 713, + 513, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the binary MIL classification problem [7], a bag of instance, $\\mathbf{X} = \\{\\mathbf{x}_n\\}_{n=1}^N$ , is associated with a single bag label, $\\mathbf{Y}$ . Each instance, $x_n$ , is associated with a single binary label, $y_n$ , which remains unknown during training. The assumption behind the MIL can be written as:", + "bbox": [ + 212, + 734, + 784, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Y} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f f} \\sum_ {n = 1} ^ {N} y _ {n} = 0 \\\\ 1, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 803, + 785, + 844 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e2ff3924e97d98c54d623d2df01d56a2c2c568af919a487ae618e6e2ce1d79a8.jpg", + "image_caption": [ + "Fig.3: Motivation of MBA. UMAP visualization [39] of tumor instance features from CAMELYON16 'test_113' case. There are various patterns/clusters among tumor instances, and relying on one single branch tends to capture a part of clusters. Three instances are selected to exhibit their texture differences." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 493, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f0fc29677695b752a1871f4da125dfed8bafae63b059ab1e3a22efe85972c58e.jpg", + "image_caption": [ + "Fig.4: Motivation of STKIM. Accumulation of Top-K attention values. Instances with Top-K attention values occupy majority attention. Results are derived from features extracted through supervised pretraining." + ], + "image_footnote": [], + "bbox": [ + 509, + 142, + 787, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the ABMIL [27], the multiple instance learning is modeled by a three-step process. i) Instance transformation into a low-dimensional embedding through neural networks: $\\pmb{h}_n = f(\\pmb{x}_n)$ . ii) Aggregation of all instance embeddings into the bag-level representation using an attention operator. Specifically, this operation is defined as:", + "bbox": [ + 212, + 426, + 782, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} = \\sum_ {n = 1} ^ {N} a _ {n} \\boldsymbol {h} _ {n} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 500, + 785, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $a_{n} = \\sigma(\\pmb{h}_{n})$ represents the attention values for $n$ -th instance, $\\pmb{h}_{n}$ . In the case of ABMIL, a gated attention (GA) mechanism [15] is adopted:", + "bbox": [ + 214, + 542, + 782, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma \\left(\\boldsymbol {h} _ {n}\\right) = \\frac {\\exp \\left\\{\\boldsymbol {w} ^ {\\mathrm {T}} \\left(\\tanh \\left(\\boldsymbol {V} _ {1} \\boldsymbol {h} _ {n}\\right) \\odot \\operatorname {s i g m} \\left(\\boldsymbol {V} _ {2} \\boldsymbol {h} _ {n}\\right)\\right) \\right\\}}{\\sum_ {j = 1} ^ {N} \\exp \\left\\{\\boldsymbol {w} ^ {\\mathrm {T}} \\left(\\tanh \\left(\\boldsymbol {V} _ {1} \\boldsymbol {h} _ {j}\\right) \\odot \\operatorname {s i g m} \\left(\\boldsymbol {V} _ {2} \\boldsymbol {h} _ {j}\\right)\\right) \\right\\}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 582, + 785, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{V}_1, \\mathbf{V}_2 \\in \\mathbb{R}^{L \\times M}$ , $\\mathbf{w} \\in \\mathbb{R}^{L \\times 1}$ are parameters, $\\odot$ is an element-wise multiplication and $\\operatorname{sigm}(\\cdot)$ is the sigmoid non-linearity. iii) The bag prediction is generated based on the aggregated bag embedding: $\\hat{\\mathbf{Y}} = g(\\mathbf{z})$ .", + "bbox": [ + 214, + 628, + 787, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Mutiple Branch Attention", + "text_level": 1, + "bbox": [ + 214, + 696, + 478, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivation. It is challenging to capture all discriminative instances using a single attention branch (see Fig. 3). This challenge arises due to variations in patterns among discriminative patches, stemming from differences in texture and morphology. Additionally, DNNs tend to exhibit a form of \"laziness\" where they prioritize capturing simpler patterns to minimize training loss, neglecting more intricate and challenging patterns [19, 20]. To tackle this issue, we design the MBA that captures more discriminative instances by multiple attention branches.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ACMIL", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As depicted in Fig. 5 top view, the MBA firstly captures $M$ patterns and then aggregates their embeddings to make predictions. Each pattern is captured by an attention branch. To maintain both the discriminative nature of patterns and semantic diversity between them, we introduce two regularization techniques: semantic regularization and diversity regularization. Firstly, to ensure capturing discriminative patterns, the semantic regularization is accomplished by hanging a MLP layer behind each pattern embedding, equipping with the cross entropy loss function:", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = - \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\boldsymbol {Y} \\log \\hat {\\boldsymbol {Y}} _ {i} + (1 - \\boldsymbol {Y}) \\log \\left(1 - \\hat {\\boldsymbol {Y}} _ {i}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 276, + 785, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{Y}}_i = g_i(\\mathbf{z}_i)$ is the prediction based on $i$ -th pattern embedding, $\\mathbf{z}_i$ . However, only equipping with cross-entropy loss may learn similar patterns and cannot dig out more discriminative information. To tackle this issue, we further introduce a diversity loss as follows:", + "bbox": [ + 214, + 328, + 784, + 390 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\frac {2}{M (M - 1)} \\sum_ {i = 1} ^ {M} \\sum_ {j = i + 1} ^ {M} \\cos \\left(\\boldsymbol {a} _ {i}, \\boldsymbol {a} _ {j}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 400, + 785, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{a}_i$ consists of all attention values of $i$ -th pattern, $\\pmb{a}_i = \\{a_{i1},\\dots ,a_{iN}\\}$ , also named heatmap as custom. The $\\cos (\\cdot)$ function is used to measure the similarity of the heatmaps between branches. By diversifying the heatmaps, the embedding of each branch can concentrate on different patterns.", + "bbox": [ + 212, + 453, + 784, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To aggregate the captured patterns to make predictions, the average of heatmaps is utilized as the heatmap of the whole bag:", + "bbox": [ + 212, + 513, + 785, + 545 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\boldsymbol {a} _ {i} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 555, + 785, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{a}$ is the heatmap of the whole bag, with a dimension of $N$ . Then, the bag embedding can be obtained by aggregating the instance features using averaged heatmap $\\pmb{a}$ . Moreover, since $\\sum_{n=1}^{N} \\left( \\frac{1}{M} \\sum_{i=1}^{M} a_{in} \\right) \\pmb{h}_n = \\frac{1}{M} \\sum_{i=1}^{M} \\left( \\sum_{n=1}^{N} a_{in} \\pmb{h}_n \\right)$ , the bag embedding also can be formulated by applying mean pooling operator to pattern embeddings. The top view of Fig. 5 adopts the latter formulation for brevity. The loss function for the bag classifier is defined as:", + "bbox": [ + 214, + 604, + 787, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b} = - \\boldsymbol {Y} \\log \\hat {\\boldsymbol {Y}} + (1 - \\boldsymbol {Y}) \\log \\left(1 - \\hat {\\boldsymbol {Y}}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 707, + 785, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the overall loss function for the ACMIL can be written as the combination of three loss terms defined in Eq. 4, 5 and 7,", + "bbox": [ + 214, + 739, + 784, + 771 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {b} + \\mathcal {L} _ {p} + \\mathcal {L} _ {d} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 782, + 785, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. It's important to highlight that when the parameter $M$ is set to 1 in MBA, it essentially mirrors the feature aggregation process of ABMIL, allowing", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f6a0513029876194d765ff31c29dd3a7fee35bc1f41bc92eaacb935f4b37ee3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 143, + 769, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e88986ad247e602ec3171684de6cf1c8ab668f665c3070ad37334e820913666b.jpg", + "image_caption": [ + "Fig. 5: Overview of the proposed MBA (top view) and STKIM (bottom view). In the MBA, $M$ discriminative patterns are extracted from patch features using the attention operator regularized by semantic and diversity regularization terms. Then, the mean operator is applied to these $M$ pattern features to produce the bag feature, which is utilized for bag-level prediction. In the STKIM, instances with Top-K attention values are randomly masked with a probability $p$ ." + ], + "image_footnote": [], + "bbox": [ + 233, + 253, + 769, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for the discernment of a single pattern. In this sense, MBA serves as an extension of ABMIL, specifically designed to capture a more diverse set of patterns. We further discuss the connection between MBA and Multiple-Head Attention (MHA). HIPT [10] has unveiled that distinct heads in MHA can effectively capture different visual concepts, akin to the role played by our MBA. However, these two techniques can be easily distinguished by: 1) MBA has diversity regularization, ensuring that different branches can learn different concepts. This is absent in MHA, resulting in different heads learning the same concept [10]. We demonstrate the critical role of $\\mathcal{L}_d$ for performance in Tab. 3b. 2) MHA is a type of attention formulation, while MBA operates independently of the attention formulation, accommodating MHA within its framework. Appendix Sec. 4.1 discusses the results of combining MHA and ACMIL.", + "bbox": [ + 212, + 484, + 787, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Stochastic Top-K Instance Masking", + "text_level": 1, + "bbox": [ + 214, + 690, + 555, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivation. A tiny number of instances will occupy the majority of attention in ABMIL while ignoring sophisticated discriminative instances. As depicted in Fig. 4, the sum of top-10 attention values is larger than 0.85 on all three datasets. However, the WSI typically involves more than 10 discriminative instances. For instance, in the CAMELYON16 dataset, 129 out of 155 tumor slides contain 10 to 20,000 cancerous instances. In essence, numerous discriminative instances are overlooked. To deal with this issue, the proposed STKIM aims to suppress the salient instances and assign more attention to the remaining instances.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ACMIL", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As depicted in Fig. 5 bottom view, STKIM introduces a masking operation into the attention mechanism, before feature aggregation and after attention values generation. The primary objective is to suppress Top-K salient instances. A straightforward solution to achieve this is to mask out all of the Top-K salient instances. However, this method poses certain challenges. It can result in the loss of information associated with key instances, which are crucial for discrimination. Furthermore, it might lead to a statistical mismatch between the feature representations before and after discarding these key instances. To address these issues, we draw inspiration from dropout [45] and cutout [17,60] commonly used in computer vision. Our proposed solution employs stochastic masking for instance features with Top-K attention values. Specifically, we begin by sorting all attention values from highest to lowest. Subsequently, we randomly set the attention values of the Top-K instances to 0, with a probability of $p$ . This process can be formulated as:", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\na _ {n} = \\left\\{ \\begin{array}{l l} 0, & \\text {w i t h p r o b a b i l i t y} p \\text {a n d w i t h i n T o p - K v a l u e s} \\\\ a _ {n}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 364, + 785, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p$ and $K$ are two hyperparameters that control the intensity of masking. Following Eq. 9, we will assign the attention values of masked instances to the remaining instances by $a_{n} \\to \\frac{1}{\\sum_{n=1}^{N} a_{n}} a_{n}$ . Notably, drawing inspiration from dropout and cutout, we remove STKIM at the inference time.", + "bbox": [ + 212, + 412, + 784, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. While STKIM, MHIM-MIL [46], and WENO [42] all adopt the technique that masks salient instances, there are notable technical distinctions between them. Firstly, our STKIM has the different masking strategy compared with WENO and MHIM-MIL. STKIM only masks a minority of instances (i.e., $K = 10$ ) with a probability of $p$ . As a comparison, the other two methods mask out a larger number of instances. WENO masks out 95 instances. MHIM-MIL masks $1\\%$ instances. In our framework, our scheme performs best in three strategies (see Appendix Tab. 2). Secondly, both MHIM-MIL and WENO necessitate a well-trained model for masking out salient instances, utilizing the remaining instances for model training. They both employ a teacher-student framework, wherein the teacher model needs to be pre-trained beforehand (the warm-up process in WENO and the pre-training stage in MHIM-MIL). In contrast, STKIM requires neither a teacher-student framework nor a pre-training process, thus highlighting simplicity and efficiency.", + "bbox": [ + 212, + 477, + 787, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 710, + 375, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Details", + "text_level": 1, + "bbox": [ + 215, + 739, + 437, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. The performance of ACMIL is evaluated on two public WSI datasets, i.e., CAMELYON16 [3] and BRACS [6], and one private benchmark, LBC. CAMELYON16 dataset consists of 400 WSIs in total, including 270 for training and 130 for testing. Following [30, 57], we further randomly split the training and validation sets from the official training set with a", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a392c591642781ff06f30f9fb6a931955b4b48671327d78568f4bb2f1561aa41.jpg", + "table_caption": [ + "Table 1: The performance of different MIL approaches across three datasets, two pre-trained methods, and two evaluation metrics. The most superior performance is highlighted in bold, while the second-best performance is indicated by underlining." + ], + "table_footnote": [], + "table_body": "
Performance MethodCAMELYON-16BRACSLBC
F1-scoreAUCF1-scoreAUCF1-scoreAUC
ResNet-18Max-pooling0.582±0.1700.620±0.1550.489±0.0470.738±0.0140.476±0.0330.775±0.010
Mean-pooling0.592±0.0260.597±0.0330.484±0.0290.685±0.0110.511±0.0220.797±0.011
Clam-SB0.742±0.0240.763±0.0490.521±0.0460.750±0.0390.514±0.0240.805±0.017
TransMIL0.643±0.0880.706±0.0760.444±0.0400.732±0.0430.385±0.0130.693±0.027
DSMIL0.736±0.0280.773±0.0340.511±0.0520.751±0.0280.458±0.0290.766±0.023
DTFD-MIL0.758±0.0510.815±0.0630.469±0.0160.717±0.0320.473±0.0210.776±0.021
IBMIL0.777±0.0090.799±0.0500.510±0.0430.726±0.0340.489±0.0170.791±0.021
MHIM-MIL0.752±0.0340.772±0.0260.511±0.0220.774±0.0210.543±0.0370.816±0.009
ABMIL0.757±0.0200.790±0.0270.523±0.0280.723±0.0350.465±0.0400.798±0.013
ACMIL(ours)0.798±0.0290.841±0.0300.552±0.0480.754±0.0080.546±0.0280.821±0.015
VIT-S/16Max-pooling0.903±0.0540.956±0.0290.596±0.0290.823±0.0330.590±0.0430.829±0.023
Mean-pooling0.577±0.0570.569±0.0810.522±0.0380.739±0.0070.559±0.0240.827±0.012
Clam-SB0.925±0.0350.969±0.0240.631±0.0340.863±0.0050.617±0.0220.865±0.018
TransMIL0.922±0.0190.943±0.0090.631±0.0300.841±0.0060.539±0.0280.805±0.010
DSMIL0.943±0.0070.966±0.0090.577±0.0280.816±0.0280.562±0.0280.820±0.033
DTFD-MIL0.948±0.0070.980±0.0110.612±0.0800.870±0.0220.612±0.0340.842±0.010
IBMIL0.912±0.0340.954±0.0220.645±0.0410.871±0.0140.604±0.0320.834±0.014
MHIM-MIL0.932±0.0240.970±0.0370.625±0.0600.865±0.0170.658±0.0410.872±0.022
ABMIL0.914±0.0310.945±0.0270.680±0.0510.866±0.0290.595±0.0360.831±0.022
ACMIL(ours)0.954±0.0120.974±0.0120.722±0.0300.888±0.0100.662±0.0430.901±0.011
", + "bbox": [ + 217, + 198, + 785, + 452 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ratio of 9:1. We do not resplit BRACS dataset as it has been officially split to 395 of training set, 65 of validation set, and 87 of test set. We follow the challenge for a 3-class WSI classification: benign tumor, atypical tumor, and malignant. The liquid-based cytology (LBC) dataset collected 1,989 WSIs and included 4 classes, i.e., Negative, ASC-US, LSIL, and ASC-H/HSIL. We randomly split the whole dataset into training, validation, and test sets with the ratio of 6:2:2. Following [31], macro-AUC and macro-F1 scores are reported since all three datasets are class imbalanced. Each of the main experiments is performed five times with random parameter initializations, and the average classification performance and standard deviation are reported. Besides, following [36,57], the test performance is reported in epochs with the best validation performance.", + "bbox": [ + 217, + 500, + 785, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We systematically assess the efficacy of our ACMIL approach by benchmarking it against conventional MIL pooling strategies, Max-pooling and Mean-pooling, as well as contemporary attention-based techniques such as AB-MIL [27], DSMIL [30], TransMIL [44], CLAM-SB [36], DTFD-MIL [57], MHIM-MIL [46], and IBMIL [34]. In pursuit of a comprehensive comparison across diverse aggregation operators, we utilize two distinct sets of features derived from ResNet-18 pre-trained on the ImageNet dataset [22] and ViT-S/16 pretrained using DINO [8, 28]. The results of all other methods are reproduced using the official code they provide under the same settings.", + "bbox": [ + 217, + 676, + 785, + 813 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. Implementation Details are in Appendix Sec. 3.", + "bbox": [ + 217, + 823, + 759, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ACMIL", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8c9d840be914cf556d82760eacc6c5040b0eeb601e2c0c4ebe1752ed51805fa7.jpg", + "image_caption": [ + "Fig. 6: Heatmap visualization of WSI examples produced by ABMIL [27] (baseline) and our ACMIL. The left part shows three tumor WSIs that come from CAMELYON16 dataset, and their tumor regions are delineated by red lines. ACMIL generates attention values that cover a more extensive portion of the tumor region compared to ABMIL. The right part shows three normal WSIs that come from CAMELYON16 dataset. ABMIL primarily focuses on a part of tissue such as adipose, while ACMIL extends its attention to the more normal tissues." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 772, + 361 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 WSI Classification Results", + "text_level": 1, + "bbox": [ + 215, + 502, + 480, + 517 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tab. 1 provides a thorough comparison of performance between ACMIL and existing MIL methods. This evaluation spans three diverse datasets, involves two different choices for pretraining methods, and employs two crucial evaluation metrics, resulting in a comprehensive assessment with a total of 12 terms.", + "bbox": [ + 212, + 536, + 787, + 595 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Considering the overall performance, ACMIL consistently outshines existing methods. It secures the top position in 10 out of the 12 metrics and holds the second position in the remaining 2 metrics. Specifically, for the CAMELYON16, ACMIL achieves outstanding results using ResNet-18 pre-trained on ImageNet embeddings, surpassing the runner-up by $2.1\\%$ and $2.6\\%$ in terms of F1-score and AUC, respectively. On the other hand, with ViT-S/16 SSL pretrained embeddings, existing attention-based MIL methods exhibit remarkable performance, boasting F1-scores and AUC values exceeding 0.9. Notably, ACMIL achieves comparable performance with the former best-performing method, DTFD-MIL, in this setup. For the BRACS, ACMIL demonstrates a substantial lead when utilizing ViT-S/16 SSL pre-trained embeddings, surpassing the second-best performance by margins of $4.2\\%$ and $1.7\\%$ in F1-score and AUC, respectively. Moreover, when employing ResNet-18 pre-trained on ImageNet embeddings, ACMIL achieves comparable performance with the previously top-performing method, MHIM-MIL. For the LBC, ACMIL stands out significantly among the other methods across all four metrics.", + "bbox": [ + 212, + 598, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8c6aa1b0e9856419966755d09cfa5cc50d870ce6fdf6d15f1b1d514d7bab3617.jpg", + "table_caption": [ + "Table 2: Comparison of FROC between ABMIL and ACMIL" + ], + "table_footnote": [], + "table_body": "
ABMILACMIL
FROC0.39870.4233
", + "bbox": [ + 372, + 171, + 624, + 204 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Localization Results", + "text_level": 1, + "bbox": [ + 215, + 233, + 428, + 247 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Heatmap visualization. Fig. 6 presents heatmap visualizations illustrating examples of our approach's performance in comparison to the baseline method, ABMIL [27]. Three tumor slides (left part) and three normal slides (right part) are selected to showcase the heatmap differences. Due to the space limitation, we present more visualization in Appendix Sec. 4.4 for further insights.", + "bbox": [ + 212, + 261, + 782, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For the tumor slides, ABMIL tends to concentrate its attention on only a fraction of the tumor regions, potentially overlooking other significant areas. In contrast, ACMIL allocates attention across a wider spectrum of tumor regions, resulting in better alignment with expert annotations. For the normal slides, ABMIL predominantly focuses on specific tissue types, such as adipose tissue. This will lead to misinterpretation that only the adipose tissue is the normal tissue and other normal regions are uncorrelated to the WSI label. On the other hand, ACMIL effectively distributes attention values to encompass all normal regions, ensuring all regions are correlated for the WSI label. This approach closely mimics human intuition and satisfies the definition of the MIL formulation.", + "bbox": [ + 212, + 338, + 784, + 488 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "FROC results. We employ the FROC metric suggested by CAMELYON16 challenge to evaluate the localization of tumor region quantitatively. As shown in Table 2, the proposed ACMIL achieves higher FROC than ABMIL.", + "bbox": [ + 212, + 513, + 784, + 559 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 584, + 387, + 598 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 7 illustrates the AUC scores of ACMIL across three datasets when utilizing a ViT/B-16 feature extractor and varying hyperparameter settings. Several key observations emerge from these experiments:", + "bbox": [ + 212, + 611, + 784, + 656 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of branches number $M$ in MBA. As shown in the first column, we find that the choice of $M$ affects performance significantly. Combining three datasets, setting $M = 5$ achieves the best performance.", + "bbox": [ + 212, + 657, + 784, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of masking probability $p$ in STKIM. As shown in the second column, we find that the choice of $M$ also affects performance significantly. Notably, setting $p = 1.0$ (masking all of Top-K instances) leads to performance deterioration across all three datasets. For LBC and CAMELYON, a $p = 1.0$ setting even results in performance lower than the blue dotted lines. Otherwise, $p = 0.6$ achieves the best performance on the BRACS dataset, whereas $p = 0.8$ achieves the best performance on the other two datasets.", + "bbox": [ + 212, + 703, + 784, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of number of masking instances $K$ in STKIM. The third column shows that hyperparameter $K$ exhibits minimal sensitivity, where different $K$", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ACMIL", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/54ac08292af7c9054c16ee6cc9477d59141448dece695701f16a2e880da4cbb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 142, + 419, + 246 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/339448ff6e229fbc134b0cf212d4daaa9f72d93452a73e4650027389658ef325.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 142, + 588, + 244 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f05a0557acd9d6e1d5764c934854d43c230cf030f389ff74de7a04bac83684f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 142, + 759, + 244 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9a0a32e90d00e30db54a3e8b19ab05655b69a83a355fafef165ef835a96385e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 246, + 421, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2749d22ecf04fb1d930f7689f98dd0811134f123ce8e5bcd4d6cf349c1c7d15c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 246, + 588, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/88fda8165a51d6c95aae69e0d0cc1022f1994e57fd2aadbe431454f333f0765a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 246, + 759, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b0ecfc5ff427dd98ffdf2f922cd0375f2be30632fe5f460f4c85dcfb531b78e6.jpg", + "image_caption": [ + "Fig. 7: Ablation study on features extracted through the SSL pre-training. The effect of three hyperparameters, $K$ , $p$ , $M$ , is investigated. Note that the orange dotted line denotes the performance of baseline, ABMIL, and the blue dotted line denotes the performance of ACMIL w/o MBA or STKIM. Five conclusions derived from the figure can be found in Sec. 4.4." + ], + "image_footnote": [], + "bbox": [ + 243, + 354, + 421, + 460 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/459fba53739ea91e79c4f8eb3a0ea28bd0a69de4b463bc6fff10256e66f19456.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 354, + 588, + 460 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f9cadb9552a5128e78f7bce3b6744575d58826ec239686f35f496355c9c09552.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 354, + 759, + 460 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "values result in a performance difference of less than $1.0\\%$ AUC. In practice, setting $K$ to 10 is generally sufficient for achieving near-optimal performance.", + "bbox": [ + 212, + 569, + 787, + 599 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Implementing either MBA or STKIM individually leads to significant performance improvements. The blue dotted lines represent ACMIL's AUC performance without MBA or STKIM, outperforming the orange dotted lines (ABMIL's performance) across all subfigures. Particularly noteworthy is the observation that MBA achieves better improvement than STKIM on all three datasets, with blue dotted lines in the last two columns surpassing those in the first column, especially on the CAMELYON and LBC datasets.", + "bbox": [ + 212, + 599, + 787, + 704 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Combining MBA with STKIM yields greater performance improvements compared to using either MBA or STKIM alone. The green dots represent ACMIL's performance under different hyperparameter combinations, with 39 out of 45 green dots exceeding blue dotted lines.", + "bbox": [ + 212, + 705, + 787, + 766 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Further Analysis", + "text_level": 1, + "bbox": [ + 214, + 786, + 401, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "MBA can capture diverse patterns. We employ UMAP [39] to visualize instance features within the tumor region of the CAMELYON16 'test_090' case.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/902a8dcf03da0490d6facae9731650205a18dcdc21083092fa767d8e65f7051c.jpg", + "image_caption": [ + "(a) ABMIL" + ], + "image_footnote": [], + "bbox": [ + 220, + 142, + 500, + 277 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/556a2bb6e937e5576b2e69bfe409993d68a87c9f3aacdcaec0360ef0ba8c1969.jpg", + "image_caption": [ + "(b) ABMIL with MBA", + "Fig. 8: UMAP visualization of tumor instance features from the CAMELYON16 'test_090' case. The tumor instances display distinct patterns, posing a challenge for a single branch to capture all of them. As a result, ABMIL overlooks the right pattern/cluster. In contrast, multiple branches in MBA capture different patterns separately, and their combination enables the activation of more tumor instances. An instance is considered active when its attention value surpasses $\\frac{1}{N}$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 142, + 784, + 279 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Fig. 8a, it's evident that the tumor instances exhibit two primary patterns. However, ABMIL primarily activates the left pattern (colored orange) and neglects the right one. On the other hand, as demonstrated in Fig. 8b, MBA's various branches (branch1, branch2, branch3, and branch5) collectively capture the substructures of the left pattern, while branch4 specifically captures the right pattern. Combining all branches can capture more comprehensive patterns.", + "bbox": [ + 212, + 425, + 784, + 516 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "STKIM can suppress the concentration of Top-K attention values. Fig. 9 illustrates a comparison of the cumulative sum of Top-K attention values with and without STKIM. The plot clearly shows that the use of STKIM helps mitigate the scenario where Top-K attention values excessively dominate in the attention mechanism. This effect is particularly pronounced for CAMELYON16 dataset, where the cumulative sum of the top-10 values decreases from 0.87 to 0.6.", + "bbox": [ + 212, + 518, + 514, + 699 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Do we need STKIM at the test phase? The answer is No. In Tab. 3a, we present the outcomes of ACMIL with and without STKIM during the test phase, along with the performance differences between these settings. Across 11", + "bbox": [ + 212, + 704, + 514, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "out of 12 evaluation metrics, the version of ACMIL without STKIM during testing outperforms the version with STKIM slightly. This suggests that STKIM is not necessary during the test phase.", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/53ee61c7fe7397ee4bbddb99184223db7989e93c2f2592b96e2caa11880ce52d.jpg", + "image_caption": [ + "Fig.9: Comparison of the accumulative sum of Top-K attention values with and without STKIM. The use of STKIM helps alleviate the issue of excessive concentration of attention values within the Top-K range." + ], + "image_footnote": [], + "bbox": [ + 535, + 549, + 777, + 667 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ACMIL", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: (a): Performance comparison between ACMIL with (w.) and without (w/o.) T-STKIM. T-STKIM means using STKIM at the test phase. The Gap column reports the performance difference between with and without T-STKIM. Using STKIM at the test phase slightly reduces performance. (b): Performance comparison between ACMIL with (w.) and without (w/o.) $\\mathcal{L}_d$ . The Gap column reports the performance difference between without and with $\\mathcal{L}_d$ . ACMIL without $\\mathcal{L}_d$ drastically reduces its performance.", + "bbox": [ + 212, + 143, + 787, + 228 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/e6fd9db65242f6f29916ae9aa9f2a43937247518354e66fa777a091134a93cd4.jpg", + "table_caption": [ + "(a) w. T-STKIM v.s. w/o. T-STKIM" + ], + "table_footnote": [], + "table_body": "
ViT-S/16 SSL pre-trained
DatasetMetricw. T-STKIMw/o. T-STKIMGap(%)
CamelyonF1-score0.927±0.0570.954±0.012+2.7
AUC0.967±0.0170.974±0.012+0.7
BRACSF1-score0.697±0.0330.722±0.030+2.5
AUC0.875±0.0120.888±0.010+1.3
LBCF1-score0.637±0.0340.662±0.043+2.5
AUC0.878±0.0120.901±0.011+2.3
ResNet-18Imagenet pre-trained
DatasetMetricw. T-STKIMw/o. T-STKIMGap(%)
CamelyonF1-score0.780±0.0260.798±0.029+1.8
AUC0.837±0.0280.841±0.030+0.4
BRACSF1-score0.566±0.0540.552±0.048-1.4
AUC0.750±0.0210.754±0.008+0.4
LBCF1-score0.535±0.0270.546±0.028+1.1
AUC0.808±0.0190.821±0.015+1.3
", + "bbox": [ + 215, + 273, + 506, + 439 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/daaf7724e7bc08ecb29816184e395ddc0b5e1e8753dd703bec25bab67d1b2a01.jpg", + "table_caption": [ + "(b) w. $\\mathcal{L}_d$ v.s. w/o. $\\mathcal{L}_d$" + ], + "table_footnote": [], + "table_body": "
ViT-S/16 SSL pre-trained
DatasetMetricw/o. Ldw. LdGap(%)
CamelyonF1-score0.901±0.0370.954±0.012+5.3
AUC0.943±0.0270.974±0.012+3.1
BRACSF1-score0.642±0.0460.722±0.030+8.0
AUC0.859±0.0200.888±0.010+2.9
LBCF1-score0.603±0.0230.662±0.043+5.9
AUC0.837±0.0090.901±0.011+6.4
ResNet-18Imagenet pre-trained
DatasetMetricw/o. Ldw. LdGap(%)
CamelyonF1-score0.747±0.0220.798±0.029+5.1
AUC0.796±0.0320.841±0.030+5.5
BRACSF1-score0.500±0.0310.552±0.048+5.2
AUC0.760±0.0260.754±0.008-0.6
LBCF1-score0.532±0.0190.546±0.028+1.4
AUC0.809±0.0180.821±0.015+1.2
", + "bbox": [ + 517, + 272, + 790, + 439 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Do we need diversity loss in MBA? The answer is Yes. In Tab. 3b, we present the outcomes of ACMIL with and without $\\mathcal{L}_d$ , along with the performance differences between these settings. Notably, the last column clearly indicates a significant performance drop for ACMIL without $\\mathcal{L}_d$ . This emphasizes the crucial role of $\\mathcal{L}_d$ in encouraging different branches to acquire distinctive discriminative knowledge within the MBA technique.", + "bbox": [ + 212, + 468, + 787, + 558 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Additional Analyses: Appendix presents further analyses that offer deeper insights into ACMIL's behavior. Specifically, Sec. 4.2 provides a detailed analysis of bag features, while Sec. 4.3 provides a comprehensive examination of evaluation metrics throughout the training process.", + "bbox": [ + 212, + 559, + 787, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 642, + 359, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Due to the intrinsic properties of WSI, MIL methods have often led to overfitting, limiting their applications. This paper reveals that the overly concentrated attention values in the heatmap are closely related to overfitting. To address this, we propose ACMIL, which is underpinned by two novel techniques: MBA and STKIM. Our experimental results on three datasets demonstrate that ACMIL significantly surpasses SOTA methods. Moreover, this paper provides comprehensive experiments confirming the effectiveness of ACMIL in suppressing the attention value concentration and alleviating overfitting. We hope that our work can inspire future exploration into leveraging attention values for a comprehensive analysis of attention mechanisms. We also hope that our ACMIL can be applied to a broader spectrum of WSI analysis tasks.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment. This study was partially supported by the National Natural Science Foundation of China (Grant No. 92270108), Zhejiang Provincial Natural Science Foundation of China (Grant No. XHD23F0201), and the Research Center for Industries of the Future (RCIF) at Westlake University.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Amores, J.: Multiple instance classification: Review, taxonomy and comparative study. Artificial intelligence 201, 81-105 (2013)", + "2. Bejani, M.M., Ghatee, M.: A systematic review on overfitting control in shallow and deep neural networks. Artificial Intelligence Review pp. 1-48 (2021)", + "3. Bejnordi, B.E., Veta, M., Van Diest, P.J., Van Ginneken, B., Karssemeijer, N., Litjens, G., Van Der Laak, J.A., Hermsen, M., Manson, Q.F., Balkenhol, M., et al.: Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama 318(22), 2199-2210 (2017)", + "4. Bergner, B., Lippert, C., Mahendran, A.: Iterative patch selection for high-resolution image recognition. arXiv preprint arXiv:2210.13007 (2022)", + "5. Bontempo, G., Bolelli, F., Porrello, A., Calderara, S., Ficarra, E.: A graph-based multi-scale approach with knowledge distillation for wsi classification. TMI (2023)", + "6. Brancati, N., Anniciello, A.M., Pati, P., Riccio, D., Scognamiglio, G., Jaume, G., De Pietro, G., Di Bonito, M., Foncubierta, A., Botti, G., et al.: Bracs: A dataset for breast carcinoma subtyping in h&e histology images. Database 2022, baac093 (2022)", + "7. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine 25(8), 1301-1309 (2019)", + "8. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "9. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15661-15670 (2023)", + "0. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144-16155 (2022)", + "1. Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4025 (2021)", + "2. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23936-23945 (2023)", + "3. Chikontwe, P., Kim, M., Nam, S.J., Go, H., Park, S.H.: Multiple instance learning with center embeddings for histopathology classification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International" + ], + "bbox": [ + 225, + 260, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ACMIL", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23. pp. 519-528. Springer (2020)", + "14. Cornish, T.C., Swapp, R.E., Kaplan, K.J.: Whole-slide imaging: routine pathologic diagnosis. Advances in anatomic pathology 19(3), 152-159 (2012)", + "15. Dauphin, Y.N., Fan, A., Auli, M., Grangier, D.: Language modeling with gated convolutional networks. In: International conference on machine learning. pp. 933-941. PMLR (2017)", + "16. Dehaene, O., Camara, A., Moindrot, O., de Lavergne, A., Courtiol, P.: Self-supervision closes the gap between weak and strong supervision in histology. arXiv preprint arXiv:2012.03583 (2020)", + "17. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)", + "18. Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T.: Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence 89(1-2), 31-71 (1997)", + "19. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)", + "20. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)", + "21. Guan, Y., Zhang, J., Tian, K., Yang, S., Dong, P., Xiang, J., Yang, W., Huang, J., Zhang, Y., Han, X.: Node-aligned graph convolutional network for whole-slide image representation and classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18813-18823 (2022)", + "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "23. He, L., Long, L.R., Antani, S., Thoma, G.R.: Histology image analysis for carcinoma detection and grading. Computer methods and programs in biomedicine 107(3), 538-556 (2012)", + "24. Holdenried-Krafft, S., Somers, P., Montes-Majarro, I.A., Silimon, D., Tarín, C., Fend, F., Lensch, H.: Dual-query multiple instance learning for dynamic metaembedding based tumor classification. arXiv preprint arXiv:2307.07482 (2023)", + "25. Hou, W., Yu, L., Lin, C., Huang, H., Yu, R., Qin, J., Wang, L.: $\\mathrm{H}^{\\wedge}$ 2-mil: Exploring hierarchical representation with heterogeneous multiple instance learning for whole slide image analysis. In: AAAI. vol. 36, pp. 933-941 (2022)", + "26. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 124-140. Springer (2020)", + "27. Ilse, M., Tomczak, J., Welling, M.: Attention-based deep multiple instance learning. In: International conference on machine learning. pp. 2127-2136. PMLR (2018)", + "28. Kang, M., Song, H., Park, S., Yoo, D., Pereira, S.: Benchmarking self-supervised learning on diverse pathology datasets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3344-3354 (2023)", + "29. Kong, F., Henao, R.: Efficient classification of very large images with tiny objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2384-2394 (2022)", + "30. Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Pro" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318-14328 (2021)", + "31. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7454-7463 (2023)", + "32. Li, R., Yao, J., Zhu, X., Li, Y., Huang, J.: Graph cnn for survival analysis on whole slide pathological images. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 174-182. Springer (2018)", + "33. Li, Y., Ping, W.: Cancer metastasis detection with neural conditional random field. arXiv preprint arXiv:1806.07064 (2018)", + "34. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19830-19839 (2023)", + "35. Litjens, G., Sánchez, C.I., Timofeeva, N., Hermsen, M., Nagtegaal, I., Kovacs, I., Hulsbergen-Van De Kaa, C., Bult, P., Van Ginneken, B., Van Der Laak, J.: Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis. Scientific reports 6(1), 26286 (2016)", + "36. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering 5(6), 555-570 (2021)", + "37. Madabhushi, A.: Digital pathology image analysis: opportunities and challenges. Imaging in medicine 1(1), 7 (2009)", + "38. Maron, O., Lozano-Pérez, T.: A framework for multiple-instance learning. Advances in neural information processing systems 10 (1997)", + "39. McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018)", + "40. Pantanowitz, L., Valenstein, P.N., Evans, A.J., Kaplan, K.J., Pfeifer, J.D., Wilbur, D.C., Collins, L.C., Colgan, T.J.: Review of the current state of whole slide imaging in pathology. Journal of pathology informatics 2(1), 36 (2011)", + "41. Pinckaers, H., Van Ginneken, B., Litjens, G.: Streaming convolutional neural networks for end-to-end learning with multi-megapixel images. IEEE transactions on pattern analysis and machine intelligence 44(3), 1581-1590 (2020)", + "42. Qu, L., Wang, M., Song, Z., et al.: Bi-directional weakly supervised knowledge distillation for whole slide image classification. Neurips 35, 15368-15381 (2022)", + "43. Qu, L., Yang, Z., Duan, M., Ma, Y., Wang, S., Wang, M., Song, Z.: Boosting whole slide image classification from the perspectives of distribution, correlation and magnification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21463-21473 (2023)", + "44. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems 34, 2136-2147 (2021)", + "45. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)", + "46. Tang, W., Huang, S., Zhang, X., Zhou, F., Zhang, Y., Liu, B.: Multiple instance learning framework with masked hard instance mining for whole slide image classification. arXiv preprint arXiv:2307.15254 (2023)", + "47. Tellez, D., Litjens, G., van der Laak, J., Ciompi, F.: Neural image compression for gigapixel histopathology image analysis. IEEE transactions on pattern analysis and machine intelligence 43(2), 567-578 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ACMIL", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Tiwari, R., Shenoy, P.: Overcoming simplicity bias in deep networks using a feature sieve (2023)", + "49. Wang, D., Khosla, A., Gargeya, R., Irshad, H., Beck, A.H.: Deep learning for identifying metastatic breast cancer. arXiv preprint arXiv:1606.05718 (2016)", + "50. Wang, H., Luo, L., Wang, F., Tong, R., Chen, Y.W., Hu, H., Lin, L., Chen, H.: Iteratively coupled multiple instance learning from instance to bag classifier for whole slide image classification. arXiv preprint arXiv:2303.15749 (2023)", + "51. Wang, X., Xiang, J., Zhang, J., Yang, S., Yang, Z., Wang, M.H., Zhang, J., Yang, W., Huang, J., Han, X.: Scl-wc: Cross-slide contrastive learning for weakly-supervised whole-slide image classification. Advances in neural information processing systems 35, 18009-18021 (2022)", + "52. Wang, Y., Kartasalo, K., Weitz, P., Acs, B., Valkonen, M., Larsson, C., Ruusuvuori, P., Hartman, J., Rantalainen, M.: Predicting molecular phenotypes from histopathology images: A transcriptome-wide expression-morphology analysis in breast cancer. Cancer research 81(19), 5115-5126 (2021)", + "53. Xiong, C., Chen, H., Sung, J., King, I.: Diagnose like a pathologist: Transformer-enabled hierarchical attention-guided multiple instance learning for whole slide image classification. arXiv preprint arXiv:2301.08125 (2023)", + "54. Yang, J., Chen, H., Zhao, Y., Yang, F., Zhang, Y., He, L., Yao, J.: Remix: A general and efficient framework for multiple instance learning based whole slide image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 35-45. Springer (2022)", + "55. Yao, J., Zhu, X., Jonnagaddala, J., Hawkins, N., Huang, J.: Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis 65, 101789 (2020)", + "56. Yufei, C., Liu, Z., Liu, X., Liu, X., Wang, C., Kuo, T.W., Xue, C.J., Chan, A.B.: Bayes-mil: A new probabilistic perspective on attention-based multiple instance learning for whole slide images. In: The Eleventh International Conference on Learning Representations (2022)", + "57. Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18802-18812 (2022)", + "58. Zhang, Y., Sun, Y., Li, H., Zheng, S., Zhu, C., Yang, L.: Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 242-252. Springer (2022)", + "59. Zhao, Y., Yang, F., Fang, Y., Liu, H., Zhou, N., Zhang, J., Sun, J., Yang, S., Menze, B., Fan, X., et al.: Predicting lymph node metastasis using histopathological images based on multiple instance learning with deep graph convolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4837-4846 (2020)", + "60. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13001-13008 (2020)", + "61. Zhu, X., Yao, J., Zhu, F., Huang, J.: Wsisa: Making survival prediction from whole slide histopathological images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7234-7242 (2017)" + ], + "bbox": [ + 217, + 147, + 784, + 811 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_model.json b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_model.json new file mode 100644 index 0000000000000000000000000000000000000000..528997ac9e83629233824bdb3d326bb16a34e4e3 --- /dev/null +++ b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_model.json @@ -0,0 +1,2546 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.212, + 0.756, + 0.245 + ], + "angle": 0, + "content": "Yunlong Zhang\\(^{1,2,3}\\), Honglin Li\\(^{1,2,3}\\), Yunxuan Sun\\(^{1,2,3}\\), Sunyi Zheng\\(^{2,3}\\), Chenglu Zhu\\(^{2,3}\\), and Lin Yang\\(^{2,3\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.255, + 0.726, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\) Zhejiang University \n\\(^{2}\\) Research Center for Industries of the Future, Westlake University \n\\(^{3}\\) School of Engineering, Westlake University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.335, + 0.744, + 0.628 + ], + "angle": 0, + "content": "Abstract. In the application of Multiple Instance Learning (MIL) methods for Whole Slide Image (WSI) classification, attention mechanisms often focus on a subset of discriminative instances, which are closely linked to overfitting. To mitigate overfitting, we present Attention-Challenging MIL (ACMIL). ACMIL combines two techniques based on separate analyses for attention value concentration. Firstly, UMAP of instance features reveals various patterns among discriminative instances, with existing attention mechanisms capturing only some of them. To remedy this, we introduce Multiple Branch Attention (MBA) to capture more discriminative instances using multiple attention branches. Secondly, the examination of the cumulative value of Top-K attention scores indicates that a tiny number of instances dominate the majority of attention. In response, we present Stochastic Top-K Instance Masking (STKIM), which masks out a portion of instances with Top-K attention values and allocates their attention values to the remaining instances. The extensive experimental results on three WSI datasets with two pre-trained backbones reveal that our ACMIL outperforms state-of-the-art methods. Additionally, through heatmap visualization and UMAP visualization, this paper extensively illustrates ACMIL's effectiveness in suppressing attention value concentration and overcoming the overfitting challenge. The source code is available at https://github.com/dazhangyu123/ACMIL." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.64, + 0.744, + 0.669 + ], + "angle": 0, + "content": "Keywords: Computational pathology \\(\\cdot\\) Whole slide image \\(\\cdot\\) Multiple instance learning \\(\\cdot\\) Overfitting" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.694, + 0.377, + 0.71 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Whole slide image (WSI) classification is a critical undertaking in digital pathology, aiming to extract valuable information from high-resolution scanned images for precise diagnosis [23,33,49,52], prognosis [11,32,55,61], and treatment planning [14,35,37,40,41] of diseases. In recent years, multiple instance learning (MIL) [1,18,38] has emerged as a promising approach for WSI classification, treating each WSI as a \"bag\" and its extracted small patches as \"instances\"" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.493, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.283, + 0.493, + 0.367 + ], + "angle": 0, + "content": "Fig. 1: The change of validation loss and entropy of attention values throughout the training of ABMIL. The results are reported on LBC with SSL pretrained features. There exists the strong negative correlation between loss and entropy." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.143, + 0.787, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.283, + 0.788, + 0.367 + ], + "angle": 0, + "content": "Fig.2: Comparison of AUC and entropy of attention values between ABMIL and ACMIL. One point denotes the result of a seed on LBC with SSL pretrained features. ACMIL achieves the higher AUC and entropy than ABMIL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.429 + ], + "angle": 0, + "content": "within the bag, thus enabling efficient classification of WSIs through assigning a single label to the entire slide." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.787, + 0.552 + ], + "angle": 0, + "content": "Overfitting is a significant challenge in utilizing MIL methods for WSI classification [34, 46, 57]. Common WSI datasets exhibit intrinsic characteristics of limited data scale, ultra-high resolutions, and staining bias, which makes overfitting more likely [2]. Specifically, these datasets often consist of a relatively small number of slides, typically in the hundreds, with a typical resolution of \\(10,000 \\times 10,000\\) [36]. Moreover, pathology images are susceptible to staining bias caused by variations in tissue preparations, staining protocols, and digital scanning methods [58], leading models to learn spurious correlations [34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In the attention mechanism, attention values/heatmaps provide insights into the model's decision-making process. Multiple existing works [36,44,46,56] alongside our own experiments (as indicated in Sec. 4.3) have pointed out the excessive concentration of attention values in current MIL methods. Specifically, MIL's attention mechanisms often concentrate on a subset of discriminative instances (i.e., instances relevant to the bag label) while disregarding the remaining ones. We investigate the correlation between attention value concentration and overfitting, utilizing the entropy of attention values and validation loss. Fig. 1 depicts a negative correlation between loss and entropy throughout the training process, illustrating that over-concentration of attention values (indicated by lower entropy) significantly compromises the model's generalization ability (indicated by higher loss values). Moreover, in the field of natural image classification, recent studies [17,26,48] have demonstrated that models solely relying on a portion of discriminative features could be susceptible to overfitting. Transitioning to WSI classification, fixating on a subset of discriminative instances similarly impedes the model's ability to generalize. These findings highlight the tight connection between attention value concentration and overfitting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Recently, numerous efforts have been made to address the overfitting challenge by enhancing representation quality [16, 24, 31, 36, 51], building spatial" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ACMIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.221 + ], + "angle": 0, + "content": "instance correlations [21, 30, 44] and developing data augmentation methods [13, 42, 46, 54, 57]. Additionally, some of these studies [30, 36] suggest that reducing the concentration of attention values can enhance model interpretability. However, the investigation of attention values concentration for alleviating overfitting remains under-explored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.463 + ], + "angle": 0, + "content": "To mitigate overfitting, we present two analyses for attention value concentration using UMAP and Top-K value statistics. Then, we introduce Attention-Challenging MIL (ACMIL), which combines two novel techniques based on these two analyses. First, by observing UMAP of instance features, we find that there are various patterns among discriminative instances, and attention mechanisms tend to capture some of them. To solve this, we introduce Multiple Branch Attention (MBA). MBA utilizes multiple attention branches, each focusing on capturing instances with a specific pattern, thereby ensuring that more discriminative instances contribute to the final prediction. Second, by analyzing the cumulative value of Top-K attention scores, we find that a tiny number of instances (e.g., \\( K = 10 \\)) occupy majority attention, resulting in overlooking sophisticated discriminative instances. To suppress these instances, we propose Stochastic Top-K Instance Masking (STKIM). STKIM randomly masks out a portion of instances with Top-K attention values and assigns their attention values to the remaining instances. Combining MBA and STKIM, our ACMIL effectively alleviates the attention value concentration and suppresses overfitting (see Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.794, + 0.571 + ], + "angle": 0, + "content": "We conduct experiments on three WSI datasets (i.e., CAMELYON16, BRACS, and our in-house LBC dataset) with two backbones (i.e., ImageNet pre-trained ResNet18 and SSL pre-trained ViT/S-16). Experimental results demonstrate the superiority of our ACMIL over existing state-of-the-art methods. We also present substantial experimental results, including heatmap visualization and UMAP visualization, to comprehensively demonstrate the effectiveness of ACMIL in suppressing attention value concentration and combatting overfitting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.388, + 0.607 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.589, + 0.636 + ], + "angle": 0, + "content": "2.1 Combating Overfitting in WSI Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.688 + ], + "angle": 0, + "content": "In the domain of WSI classification, combating the challenge of overfitting has received substantial attention. Next three paragraphs detail methods from three different aspects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.839 + ], + "angle": 0, + "content": "Some efforts have concentrated on enhancing the quality of instance representations. Early studies (e.g., [7, 27, 44]) rely on backbones pre-trained on the ImageNet dataset. However, the substantial domain gap between natural and pathological images hinders representation quality. Recent works (e.g., [16, 24, 36, 47, 51]) address this by emphasizing Self-Supervised Learning (SSL) to learn patch-level feature representations. In addition, efforts such as the work by Chen et al. [10] leverage hierarchical SSL for high-resolution image representations. Further, studies by Li et al. [31] and Wang et al. [50] demonstrate that fine-tuning the pre-trained encoder is essential for acquiring task-specific information." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "Another line of research has focused on establishing spatial instance correlations. DSMIL [30], \\(\\mathrm{H}^2\\mathrm{MIL}\\) [25], and DAS-MIL [5] consider the hierarchical structure of patches and aggregate multi-scale representations in attention mechanisms. Furthermore, some studies introduce self-attention layers [43, 44, 53] and graph neural networks [9, 21, 59] to model correlations between different areas." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.327 + ], + "angle": 0, + "content": "Further strategies have concentrated on data augmentation. Examples include DTFD-MIL [57], which introduces pseudo-bags for expanding bag counts and employs a double-tier MIL framework. IPS [4], Zoom-In Network [29], and Top-K MIL [13] generate bag representations by aggregating the representations of salient patches. Remix [54] and RankMix [12] introduce instance representation mixup for MIL. MHIM-MIL [46] and WENO [42] augments bags by randomly masking salient instances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.39 + ], + "angle": 0, + "content": "Although our ACMIL shares a similar spirit with some of these works, the proposed ACMIL excels by further building on detailed analysis for attention value concentration. As a result, our ACMIL exhibits stronger interpretability against existing solutions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.41, + 0.593, + 0.423 + ], + "angle": 0, + "content": "2.2 Over-Concentration of Attention Values" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.584 + ], + "angle": 0, + "content": "In the realm of natural image classification, research has shown that an excessive focus on certain parts of an object can impede the overall effectiveness of model generalization [17,26,60]. To tackle this issue, various heuristic techniques have been proposed. For instance, Cutout [17, 60] is a valuable data augmentation method that randomly masks square regions of input during training. Another approach, RSC [26], involves regularization that eliminates salient features activated during training. This paper investigates the issue of attention value concentration in WSI classification tasks. We identify two specific phenomena related to attention value concentration existing in WSI classification and propose two techniques to address them respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.331, + 0.62 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.785, + 0.695 + ], + "angle": 0, + "content": "Based on the ABMIL (detailed in Sec. 3.1), we present ACMIL to alleviate the overfitting problem, which is built on two components: Multiple Branch Attention (MBA) and Stochastic Top-K Instance Masking (STKIM). We describe the details of two components in Sec. 3.2 and 3.3, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.714, + 0.514, + 0.729 + ], + "angle": 0, + "content": "3.1 ABMIL for WSI Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.736, + 0.785, + 0.797 + ], + "angle": 0, + "content": "In the binary MIL classification problem [7], a bag of instance, \\( \\mathbf{X} = \\{\\mathbf{x}_n\\}_{n=1}^N \\), is associated with a single bag label, \\( \\mathbf{Y} \\). Each instance, \\( x_n \\), is associated with a single binary label, \\( y_n \\), which remains unknown during training. The assumption behind the MIL can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.804, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Y} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f f} \\sum_ {n = 1} ^ {N} y _ {n} = 0 \\\\ 1, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ACMIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.495, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.288, + 0.496, + 0.4 + ], + "angle": 0, + "content": "Fig.3: Motivation of MBA. UMAP visualization [39] of tumor instance features from CAMELYON16 'test_113' case. There are various patterns/clusters among tumor instances, and relying on one single branch tends to capture a part of clusters. Three instances are selected to exhibit their texture differences." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.143, + 0.789, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.288, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Fig.4: Motivation of STKIM. Accumulation of Top-K attention values. Instances with Top-K attention values occupy majority attention. Results are derived from features extracted through supervised pretraining." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.784, + 0.5 + ], + "angle": 0, + "content": "In the ABMIL [27], the multiple instance learning is modeled by a three-step process. i) Instance transformation into a low-dimensional embedding through neural networks: \\( \\pmb{h}_n = f(\\pmb{x}_n) \\). ii) Aggregation of all instance embeddings into the bag-level representation using an attention operator. Specifically, this operation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.5, + 0.786, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} = \\sum_ {n = 1} ^ {N} a _ {n} \\boldsymbol {h} _ {n} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.544, + 0.784, + 0.576 + ], + "angle": 0, + "content": "Here, \\(a_{n} = \\sigma(\\pmb{h}_{n})\\) represents the attention values for \\(n\\)-th instance, \\(\\pmb{h}_{n}\\). In the case of ABMIL, a gated attention (GA) mechanism [15] is adopted:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.583, + 0.787, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\sigma \\left(\\boldsymbol {h} _ {n}\\right) = \\frac {\\exp \\left\\{\\boldsymbol {w} ^ {\\mathrm {T}} \\left(\\tanh \\left(\\boldsymbol {V} _ {1} \\boldsymbol {h} _ {n}\\right) \\odot \\operatorname {s i g m} \\left(\\boldsymbol {V} _ {2} \\boldsymbol {h} _ {n}\\right)\\right) \\right\\}}{\\sum_ {j = 1} ^ {N} \\exp \\left\\{\\boldsymbol {w} ^ {\\mathrm {T}} \\left(\\tanh \\left(\\boldsymbol {V} _ {1} \\boldsymbol {h} _ {j}\\right) \\odot \\operatorname {s i g m} \\left(\\boldsymbol {V} _ {2} \\boldsymbol {h} _ {j}\\right)\\right) \\right\\}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.629, + 0.788, + 0.679 + ], + "angle": 0, + "content": "where \\(\\mathbf{V}_1, \\mathbf{V}_2 \\in \\mathbb{R}^{L \\times M}\\), \\(\\mathbf{w} \\in \\mathbb{R}^{L \\times 1}\\) are parameters, \\(\\odot\\) is an element-wise multiplication and \\(\\operatorname{sigm}(\\cdot)\\) is the sigmoid non-linearity. iii) The bag prediction is generated based on the aggregated bag embedding: \\(\\hat{\\mathbf{Y}} = g(\\mathbf{z})\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.48, + 0.712 + ], + "angle": 0, + "content": "3.2 Mutiple Branch Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Motivation. It is challenging to capture all discriminative instances using a single attention branch (see Fig. 3). This challenge arises due to variations in patterns among discriminative patches, stemming from differences in texture and morphology. Additionally, DNNs tend to exhibit a form of \"laziness\" where they prioritize capturing simpler patterns to minimize training loss, neglecting more intricate and challenging patterns [19, 20]. To tackle this issue, we design the MBA that captures more discriminative instances by multiple attention branches." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "As depicted in Fig. 5 top view, the MBA firstly captures \\( M \\) patterns and then aggregates their embeddings to make predictions. Each pattern is captured by an attention branch. To maintain both the discriminative nature of patterns and semantic diversity between them, we introduce two regularization techniques: semantic regularization and diversity regularization. Firstly, to ensure capturing discriminative patterns, the semantic regularization is accomplished by hanging a MLP layer behind each pattern embedding, equipping with the cross entropy loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.277, + 0.786, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = - \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\boldsymbol {Y} \\log \\hat {\\boldsymbol {Y}} _ {i} + (1 - \\boldsymbol {Y}) \\log \\left(1 - \\hat {\\boldsymbol {Y}} _ {i}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.329, + 0.785, + 0.391 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{Y}}_i = g_i(\\mathbf{z}_i)\\) is the prediction based on \\(i\\)-th pattern embedding, \\(\\mathbf{z}_i\\). However, only equipping with cross-entropy loss may learn similar patterns and cannot dig out more discriminative information. To tackle this issue, we further introduce a diversity loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.401, + 0.786, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\frac {2}{M (M - 1)} \\sum_ {i = 1} ^ {M} \\sum_ {j = i + 1} ^ {M} \\cos \\left(\\boldsymbol {a} _ {i}, \\boldsymbol {a} _ {j}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.785, + 0.514 + ], + "angle": 0, + "content": "where \\(\\pmb{a}_i\\) consists of all attention values of \\(i\\)-th pattern, \\(\\pmb{a}_i = \\{a_{i1},\\dots ,a_{iN}\\}\\), also named heatmap as custom. The \\(\\cos (\\cdot)\\) function is used to measure the similarity of the heatmaps between branches. By diversifying the heatmaps, the embedding of each branch can concentrate on different patterns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.787, + 0.546 + ], + "angle": 0, + "content": "To aggregate the captured patterns to make predictions, the average of heatmaps is utilized as the heatmap of the whole bag:" + }, + { + "type": "equation", + "bbox": [ + 0.452, + 0.556, + 0.786, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\boldsymbol {a} _ {i} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.606, + 0.788, + 0.698 + ], + "angle": 0, + "content": "where \\(\\pmb{a}\\) is the heatmap of the whole bag, with a dimension of \\(N\\). Then, the bag embedding can be obtained by aggregating the instance features using averaged heatmap \\(\\pmb{a}\\). Moreover, since \\(\\sum_{n=1}^{N} \\left( \\frac{1}{M} \\sum_{i=1}^{M} a_{in} \\right) \\pmb{h}_n = \\frac{1}{M} \\sum_{i=1}^{M} \\left( \\sum_{n=1}^{N} a_{in} \\pmb{h}_n \\right)\\), the bag embedding also can be formulated by applying mean pooling operator to pattern embeddings. The top view of Fig. 5 adopts the latter formulation for brevity. The loss function for the bag classifier is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.708, + 0.787, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b} = - \\boldsymbol {Y} \\log \\hat {\\boldsymbol {Y}} + (1 - \\boldsymbol {Y}) \\log \\left(1 - \\hat {\\boldsymbol {Y}}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.741, + 0.785, + 0.772 + ], + "angle": 0, + "content": "Finally, the overall loss function for the ACMIL can be written as the combination of three loss terms defined in Eq. 4, 5 and 7," + }, + { + "type": "equation", + "bbox": [ + 0.438, + 0.784, + 0.786, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {b} + \\mathcal {L} _ {p} + \\mathcal {L} _ {d} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Discussion. It's important to highlight that when the parameter \\(M\\) is set to 1 in MBA, it essentially mirrors the feature aggregation process of ABMIL, allowing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ACMIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.77, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.254, + 0.77, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.453 + ], + "angle": 0, + "content": "Fig. 5: Overview of the proposed MBA (top view) and STKIM (bottom view). In the MBA, \\( M \\) discriminative patterns are extracted from patch features using the attention operator regularized by semantic and diversity regularization terms. Then, the mean operator is applied to these \\( M \\) pattern features to produce the bag feature, which is utilized for bag-level prediction. In the STKIM, instances with Top-K attention values are randomly masked with a probability \\( p \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.667 + ], + "angle": 0, + "content": "for the discernment of a single pattern. In this sense, MBA serves as an extension of ABMIL, specifically designed to capture a more diverse set of patterns. We further discuss the connection between MBA and Multiple-Head Attention (MHA). HIPT [10] has unveiled that distinct heads in MHA can effectively capture different visual concepts, akin to the role played by our MBA. However, these two techniques can be easily distinguished by: 1) MBA has diversity regularization, ensuring that different branches can learn different concepts. This is absent in MHA, resulting in different heads learning the same concept [10]. We demonstrate the critical role of \\(\\mathcal{L}_d\\) for performance in Tab. 3b. 2) MHA is a type of attention formulation, while MBA operates independently of the attention formulation, accommodating MHA within its framework. Appendix Sec. 4.1 discusses the results of combining MHA and ACMIL." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.691, + 0.557, + 0.708 + ], + "angle": 0, + "content": "3.3 Stochastic Top-K Instance Masking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Motivation. A tiny number of instances will occupy the majority of attention in ABMIL while ignoring sophisticated discriminative instances. As depicted in Fig. 4, the sum of top-10 attention values is larger than 0.85 on all three datasets. However, the WSI typically involves more than 10 discriminative instances. For instance, in the CAMELYON16 dataset, 129 out of 155 tumor slides contain 10 to 20,000 cancerous instances. In essence, numerous discriminative instances are overlooked. To deal with this issue, the proposed STKIM aims to suppress the salient instances and assign more attention to the remaining instances." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "As depicted in Fig. 5 bottom view, STKIM introduces a masking operation into the attention mechanism, before feature aggregation and after attention values generation. The primary objective is to suppress Top-K salient instances. A straightforward solution to achieve this is to mask out all of the Top-K salient instances. However, this method poses certain challenges. It can result in the loss of information associated with key instances, which are crucial for discrimination. Furthermore, it might lead to a statistical mismatch between the feature representations before and after discarding these key instances. To address these issues, we draw inspiration from dropout [45] and cutout [17,60] commonly used in computer vision. Our proposed solution employs stochastic masking for instance features with Top-K attention values. Specifically, we begin by sorting all attention values from highest to lowest. Subsequently, we randomly set the attention values of the Top-K instances to 0, with a probability of \\( p \\). This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.366, + 0.786, + 0.406 + ], + "angle": 0, + "content": "\\[\na _ {n} = \\left\\{ \\begin{array}{l l} 0, & \\text {w i t h p r o b a b i l i t y} p \\text {a n d w i t h i n T o p - K v a l u e s} \\\\ a _ {n}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.785, + 0.476 + ], + "angle": 0, + "content": "where \\( p \\) and \\( K \\) are two hyperparameters that control the intensity of masking. Following Eq. 9, we will assign the attention values of masked instances to the remaining instances by \\( a_{n} \\to \\frac{1}{\\sum_{n=1}^{N} a_{n}} a_{n} \\). Notably, drawing inspiration from dropout and cutout, we remove STKIM at the inference time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Discussion. While STKIM, MHIM-MIL [46], and WENO [42] all adopt the technique that masks salient instances, there are notable technical distinctions between them. Firstly, our STKIM has the different masking strategy compared with WENO and MHIM-MIL. STKIM only masks a minority of instances (i.e., \\( K = 10 \\)) with a probability of \\( p \\). As a comparison, the other two methods mask out a larger number of instances. WENO masks out 95 instances. MHIM-MIL masks \\( 1\\% \\) instances. In our framework, our scheme performs best in three strategies (see Appendix Tab. 2). Secondly, both MHIM-MIL and WENO necessitate a well-trained model for masking out salient instances, utilizing the remaining instances for model training. They both employ a teacher-student framework, wherein the teacher model needs to be pre-trained beforehand (the warm-up process in WENO and the pre-training stage in MHIM-MIL). In contrast, STKIM requires neither a teacher-student framework nor a pre-training process, thus highlighting simplicity and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.376, + 0.728 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.438, + 0.756 + ], + "angle": 0, + "content": "4.1 Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. The performance of ACMIL is evaluated on two public WSI datasets, i.e., CAMELYON16 [3] and BRACS [6], and one private benchmark, LBC. CAMELYON16 dataset consists of 400 WSIs in total, including 270 for training and 130 for testing. Following [30, 57], we further randomly split the training and validation sets from the official training set with a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ACMIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.786, + 0.187 + ], + "angle": 0, + "content": "Table 1: The performance of different MIL approaches across three datasets, two pre-trained methods, and two evaluation metrics. The most superior performance is highlighted in bold, while the second-best performance is indicated by underlining." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.787, + 0.453 + ], + "angle": 0, + "content": "
Performance MethodCAMELYON-16BRACSLBC
F1-scoreAUCF1-scoreAUCF1-scoreAUC
ResNet-18Max-pooling0.582±0.1700.620±0.1550.489±0.0470.738±0.0140.476±0.0330.775±0.010
Mean-pooling0.592±0.0260.597±0.0330.484±0.0290.685±0.0110.511±0.0220.797±0.011
Clam-SB0.742±0.0240.763±0.0490.521±0.0460.750±0.0390.514±0.0240.805±0.017
TransMIL0.643±0.0880.706±0.0760.444±0.0400.732±0.0430.385±0.0130.693±0.027
DSMIL0.736±0.0280.773±0.0340.511±0.0520.751±0.0280.458±0.0290.766±0.023
DTFD-MIL0.758±0.0510.815±0.0630.469±0.0160.717±0.0320.473±0.0210.776±0.021
IBMIL0.777±0.0090.799±0.0500.510±0.0430.726±0.0340.489±0.0170.791±0.021
MHIM-MIL0.752±0.0340.772±0.0260.511±0.0220.774±0.0210.543±0.0370.816±0.009
ABMIL0.757±0.0200.790±0.0270.523±0.0280.723±0.0350.465±0.0400.798±0.013
ACMIL(ours)0.798±0.0290.841±0.0300.552±0.0480.754±0.0080.546±0.0280.821±0.015
VIT-S/16Max-pooling0.903±0.0540.956±0.0290.596±0.0290.823±0.0330.590±0.0430.829±0.023
Mean-pooling0.577±0.0570.569±0.0810.522±0.0380.739±0.0070.559±0.0240.827±0.012
Clam-SB0.925±0.0350.969±0.0240.631±0.0340.863±0.0050.617±0.0220.865±0.018
TransMIL0.922±0.0190.943±0.0090.631±0.0300.841±0.0060.539±0.0280.805±0.010
DSMIL0.943±0.0070.966±0.0090.577±0.0280.816±0.0280.562±0.0280.820±0.033
DTFD-MIL0.948±0.0070.980±0.0110.612±0.0800.870±0.0220.612±0.0340.842±0.010
IBMIL0.912±0.0340.954±0.0220.645±0.0410.871±0.0140.604±0.0320.834±0.014
MHIM-MIL0.932±0.0240.970±0.0370.625±0.0600.865±0.0170.658±0.0410.872±0.022
ABMIL0.914±0.0310.945±0.0270.680±0.0510.866±0.0290.595±0.0360.831±0.022
ACMIL(ours)0.954±0.0120.974±0.0120.722±0.0300.888±0.0100.662±0.0430.901±0.011
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.501, + 0.786, + 0.668 + ], + "angle": 0, + "content": "ratio of 9:1. We do not resplit BRACS dataset as it has been officially split to 395 of training set, 65 of validation set, and 87 of test set. We follow the challenge for a 3-class WSI classification: benign tumor, atypical tumor, and malignant. The liquid-based cytology (LBC) dataset collected 1,989 WSIs and included 4 classes, i.e., Negative, ASC-US, LSIL, and ASC-H/HSIL. We randomly split the whole dataset into training, validation, and test sets with the ratio of 6:2:2. Following [31], macro-AUC and macro-F1 scores are reported since all three datasets are class imbalanced. Each of the main experiments is performed five times with random parameter initializations, and the average classification performance and standard deviation are reported. Besides, following [36,57], the test performance is reported in epochs with the best validation performance." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.678, + 0.786, + 0.814 + ], + "angle": 0, + "content": "Baselines. We systematically assess the efficacy of our ACMIL approach by benchmarking it against conventional MIL pooling strategies, Max-pooling and Mean-pooling, as well as contemporary attention-based techniques such as AB-MIL [27], DSMIL [30], TransMIL [44], CLAM-SB [36], DTFD-MIL [57], MHIM-MIL [46], and IBMIL [34]. In pursuit of a comprehensive comparison across diverse aggregation operators, we utilize two distinct sets of features derived from ResNet-18 pre-trained on the ImageNet dataset [22] and ViT-S/16 pretrained using DINO [8, 28]. The results of all other methods are reproduced using the official code they provide under the same settings." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.824, + 0.761, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. Implementation Details are in Appendix Sec. 3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.774, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Fig. 6: Heatmap visualization of WSI examples produced by ABMIL [27] (baseline) and our ACMIL. The left part shows three tumor WSIs that come from CAMELYON16 dataset, and their tumor regions are delineated by red lines. ACMIL generates attention values that cover a more extensive portion of the tumor region compared to ABMIL. The right part shows three normal WSIs that come from CAMELYON16 dataset. ABMIL primarily focuses on a part of tissue such as adipose, while ACMIL extends its attention to the more normal tissues." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.482, + 0.518 + ], + "angle": 0, + "content": "4.2 WSI Classification Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.597 + ], + "angle": 0, + "content": "Tab. 1 provides a thorough comparison of performance between ACMIL and existing MIL methods. This evaluation spans three diverse datasets, involves two different choices for pretraining methods, and employs two crucial evaluation metrics, resulting in a comprehensive assessment with a total of 12 terms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Considering the overall performance, ACMIL consistently outshines existing methods. It secures the top position in 10 out of the 12 metrics and holds the second position in the remaining 2 metrics. Specifically, for the CAMELYON16, ACMIL achieves outstanding results using ResNet-18 pre-trained on ImageNet embeddings, surpassing the runner-up by \\(2.1\\%\\) and \\(2.6\\%\\) in terms of F1-score and AUC, respectively. On the other hand, with ViT-S/16 SSL pretrained embeddings, existing attention-based MIL methods exhibit remarkable performance, boasting F1-scores and AUC values exceeding 0.9. Notably, ACMIL achieves comparable performance with the former best-performing method, DTFD-MIL, in this setup. For the BRACS, ACMIL demonstrates a substantial lead when utilizing ViT-S/16 SSL pre-trained embeddings, surpassing the second-best performance by margins of \\(4.2\\%\\) and \\(1.7\\%\\) in F1-score and AUC, respectively. Moreover, when employing ResNet-18 pre-trained on ImageNet embeddings, ACMIL achieves comparable performance with the previously top-performing method, MHIM-MIL. For the LBC, ACMIL stands out significantly among the other methods across all four metrics." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ACMIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.294, + 0.145, + 0.709, + 0.159 + ], + "angle": 0, + "content": "Table 2: Comparison of FROC between ABMIL and ACMIL" + }, + { + "type": "table", + "bbox": [ + 0.374, + 0.172, + 0.625, + 0.205 + ], + "angle": 0, + "content": "
ABMILACMIL
FROC0.39870.4233
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.234, + 0.429, + 0.248 + ], + "angle": 0, + "content": "4.3 Localization Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.262, + 0.784, + 0.337 + ], + "angle": 0, + "content": "Heatmap visualization. Fig. 6 presents heatmap visualizations illustrating examples of our approach's performance in comparison to the baseline method, ABMIL [27]. Three tumor slides (left part) and three normal slides (right part) are selected to showcase the heatmap differences. Due to the space limitation, we present more visualization in Appendix Sec. 4.4 for further insights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.785, + 0.489 + ], + "angle": 0, + "content": "For the tumor slides, ABMIL tends to concentrate its attention on only a fraction of the tumor regions, potentially overlooking other significant areas. In contrast, ACMIL allocates attention across a wider spectrum of tumor regions, resulting in better alignment with expert annotations. For the normal slides, ABMIL predominantly focuses on specific tissue types, such as adipose tissue. This will lead to misinterpretation that only the adipose tissue is the normal tissue and other normal regions are uncorrelated to the WSI label. On the other hand, ACMIL effectively distributes attention values to encompass all normal regions, ensuring all regions are correlated for the WSI label. This approach closely mimics human intuition and satisfies the definition of the MIL formulation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.785, + 0.56 + ], + "angle": 0, + "content": "FROC results. We employ the FROC metric suggested by CAMELYON16 challenge to evaluate the localization of tumor region quantitatively. As shown in Table 2, the proposed ACMIL achieves higher FROC than ABMIL." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.585, + 0.388, + 0.599 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.657 + ], + "angle": 0, + "content": "Fig. 7 illustrates the AUC scores of ACMIL across three datasets when utilizing a ViT/B-16 feature extractor and varying hyperparameter settings. Several key observations emerge from these experiments:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.785, + 0.703 + ], + "angle": 0, + "content": "Effect of branches number \\( M \\) in MBA. As shown in the first column, we find that the choice of \\( M \\) affects performance significantly. Combining three datasets, setting \\( M = 5 \\) achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Effect of masking probability \\( p \\) in STKIM. As shown in the second column, we find that the choice of \\( M \\) also affects performance significantly. Notably, setting \\( p = 1.0 \\) (masking all of Top-K instances) leads to performance deterioration across all three datasets. For LBC and CAMELYON, a \\( p = 1.0 \\) setting even results in performance lower than the blue dotted lines. Otherwise, \\( p = 0.6 \\) achieves the best performance on the BRACS dataset, whereas \\( p = 0.8 \\) achieves the best performance on the other two datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Effect of number of masking instances \\( K \\) in STKIM. The third column shows that hyperparameter \\( K \\) exhibits minimal sensitivity, where different \\( K \\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.143, + 0.421, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.143, + 0.589, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.143, + 0.76, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.247, + 0.422, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.247, + 0.589, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.247, + 0.76, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.355, + 0.422, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.355, + 0.589, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.355, + 0.76, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Fig. 7: Ablation study on features extracted through the SSL pre-training. The effect of three hyperparameters, \\( K \\), \\( p \\), \\( M \\), is investigated. Note that the orange dotted line denotes the performance of baseline, ABMIL, and the blue dotted line denotes the performance of ACMIL w/o MBA or STKIM. Five conclusions derived from the figure can be found in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.6 + ], + "angle": 0, + "content": "values result in a performance difference of less than \\(1.0\\%\\) AUC. In practice, setting \\(K\\) to 10 is generally sufficient for achieving near-optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Implementing either MBA or STKIM individually leads to significant performance improvements. The blue dotted lines represent ACMIL's AUC performance without MBA or STKIM, outperforming the orange dotted lines (ABMIL's performance) across all subfigures. Particularly noteworthy is the observation that MBA achieves better improvement than STKIM on all three datasets, with blue dotted lines in the last two columns surpassing those in the first column, especially on the CAMELYON and LBC datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.788, + 0.767 + ], + "angle": 0, + "content": "Combining MBA with STKIM yields greater performance improvements compared to using either MBA or STKIM alone. The green dots represent ACMIL's performance under different hyperparameter combinations, with 39 out of 45 green dots exceeding blue dotted lines." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.403, + 0.803 + ], + "angle": 0, + "content": "4.5 Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "MBA can capture diverse patterns. We employ UMAP [39] to visualize instance features within the tumor region of the CAMELYON16 'test_090' case." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ACMIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.143, + 0.501, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.282, + 0.396, + 0.294 + ], + "angle": 0, + "content": "(a) ABMIL" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.143, + 0.785, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.282, + 0.71, + 0.293 + ], + "angle": 0, + "content": "(b) ABMIL with MBA" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.305, + 0.785, + 0.39 + ], + "angle": 0, + "content": "Fig. 8: UMAP visualization of tumor instance features from the CAMELYON16 'test_090' case. The tumor instances display distinct patterns, posing a challenge for a single branch to capture all of them. As a result, ABMIL overlooks the right pattern/cluster. In contrast, multiple branches in MBA capture different patterns separately, and their combination enables the activation of more tumor instances. An instance is considered active when its attention value surpasses \\(\\frac{1}{N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.785, + 0.517 + ], + "angle": 0, + "content": "In Fig. 8a, it's evident that the tumor instances exhibit two primary patterns. However, ABMIL primarily activates the left pattern (colored orange) and neglects the right one. On the other hand, as demonstrated in Fig. 8b, MBA's various branches (branch1, branch2, branch3, and branch5) collectively capture the substructures of the left pattern, while branch4 specifically captures the right pattern. Combining all branches can capture more comprehensive patterns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.516, + 0.7 + ], + "angle": 0, + "content": "STKIM can suppress the concentration of Top-K attention values. Fig. 9 illustrates a comparison of the cumulative sum of Top-K attention values with and without STKIM. The plot clearly shows that the use of STKIM helps mitigate the scenario where Top-K attention values excessively dominate in the attention mechanism. This effect is particularly pronounced for CAMELYON16 dataset, where the cumulative sum of the top-10 values decreases from 0.87 to 0.6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.516, + 0.794 + ], + "angle": 0, + "content": "Do we need STKIM at the test phase? The answer is No. In Tab. 3a, we present the outcomes of ACMIL with and without STKIM during the test phase, along with the performance differences between these settings. Across 11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "out of 12 evaluation metrics, the version of ACMIL without STKIM during testing outperforms the version with STKIM slightly. This suggests that STKIM is not necessary during the test phase." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.55, + 0.779, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.679, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Fig.9: Comparison of the accumulative sum of Top-K attention values with and without STKIM. The use of STKIM helps alleviate the issue of excessive concentration of attention values within the Top-K range." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.229 + ], + "angle": 0, + "content": "Table 3: (a): Performance comparison between ACMIL with (w.) and without (w/o.) T-STKIM. T-STKIM means using STKIM at the test phase. The Gap column reports the performance difference between with and without T-STKIM. Using STKIM at the test phase slightly reduces performance. (b): Performance comparison between ACMIL with (w.) and without (w/o.) \\(\\mathcal{L}_d\\). The Gap column reports the performance difference between without and with \\(\\mathcal{L}_d\\). ACMIL without \\(\\mathcal{L}_d\\) drastically reduces its performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.246, + 0.458, + 0.26 + ], + "angle": 0, + "content": "(a) w. T-STKIM v.s. w/o. T-STKIM" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.274, + 0.507, + 0.44 + ], + "angle": 0, + "content": "
ViT-S/16 SSL pre-trained
DatasetMetricw. T-STKIMw/o. T-STKIMGap(%)
CamelyonF1-score0.927±0.0570.954±0.012+2.7
AUC0.967±0.0170.974±0.012+0.7
BRACSF1-score0.697±0.0330.722±0.030+2.5
AUC0.875±0.0120.888±0.010+1.3
LBCF1-score0.637±0.0340.662±0.043+2.5
AUC0.878±0.0120.901±0.011+2.3
ResNet-18Imagenet pre-trained
DatasetMetricw. T-STKIMw/o. T-STKIMGap(%)
CamelyonF1-score0.780±0.0260.798±0.029+1.8
AUC0.837±0.0280.841±0.030+0.4
BRACSF1-score0.566±0.0540.552±0.048-1.4
AUC0.750±0.0210.754±0.008+0.4
LBCF1-score0.535±0.0270.546±0.028+1.1
AUC0.808±0.0190.821±0.015+1.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.246, + 0.719, + 0.26 + ], + "angle": 0, + "content": "(b) w. \\(\\mathcal{L}_d\\) v.s. w/o. \\(\\mathcal{L}_d\\)" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.273, + 0.792, + 0.44 + ], + "angle": 0, + "content": "
ViT-S/16 SSL pre-trained
DatasetMetricw/o. Ldw. LdGap(%)
CamelyonF1-score0.901±0.0370.954±0.012+5.3
AUC0.943±0.0270.974±0.012+3.1
BRACSF1-score0.642±0.0460.722±0.030+8.0
AUC0.859±0.0200.888±0.010+2.9
LBCF1-score0.603±0.0230.662±0.043+5.9
AUC0.837±0.0090.901±0.011+6.4
ResNet-18Imagenet pre-trained
DatasetMetricw/o. Ldw. LdGap(%)
CamelyonF1-score0.747±0.0220.798±0.029+5.1
AUC0.796±0.0320.841±0.030+5.5
BRACSF1-score0.500±0.0310.552±0.048+5.2
AUC0.760±0.0260.754±0.008-0.6
LBCF1-score0.532±0.0190.546±0.028+1.4
AUC0.809±0.0180.821±0.015+1.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.559 + ], + "angle": 0, + "content": "Do we need diversity loss in MBA? The answer is Yes. In Tab. 3b, we present the outcomes of ACMIL with and without \\(\\mathcal{L}_d\\), along with the performance differences between these settings. Notably, the last column clearly indicates a significant performance drop for ACMIL without \\(\\mathcal{L}_d\\). This emphasizes the crucial role of \\(\\mathcal{L}_d\\) in encouraging different branches to acquire distinctive discriminative knowledge within the MBA technique." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Additional Analyses: Appendix presents further analyses that offer deeper insights into ACMIL's behavior. Specifically, Sec. 4.2 provides a detailed analysis of bag features, while Sec. 4.3 provides a comprehensive examination of evaluation metrics throughout the training process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.36, + 0.658 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Due to the intrinsic properties of WSI, MIL methods have often led to overfitting, limiting their applications. This paper reveals that the overly concentrated attention values in the heatmap are closely related to overfitting. To address this, we propose ACMIL, which is underpinned by two novel techniques: MBA and STKIM. Our experimental results on three datasets demonstrate that ACMIL significantly surpasses SOTA methods. Moreover, this paper provides comprehensive experiments confirming the effectiveness of ACMIL in suppressing the attention value concentration and alleviating overfitting. We hope that our work can inspire future exploration into leveraging attention values for a comprehensive analysis of attention mechanisms. We also hope that our ACMIL can be applied to a broader spectrum of WSI analysis tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ACMIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Acknowledgment. This study was partially supported by the National Natural Science Foundation of China (Grant No. 92270108), Zhejiang Provincial Natural Science Foundation of China (Grant No. XHD23F0201), and the Research Center for Industries of the Future (RCIF) at Westlake University." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.288 + ], + "angle": 0, + "content": "1. Amores, J.: Multiple instance classification: Review, taxonomy and comparative study. Artificial intelligence 201, 81-105 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.316 + ], + "angle": 0, + "content": "2. Bejani, M.M., Ghatee, M.: A systematic review on overfitting control in shallow and deep neural networks. Artificial Intelligence Review pp. 1-48 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.317, + 0.785, + 0.371 + ], + "angle": 0, + "content": "3. Bejnordi, B.E., Veta, M., Van Diest, P.J., Van Ginneken, B., Karssemeijer, N., Litjens, G., Van Der Laak, J.A., Hermsen, M., Manson, Q.F., Balkenhol, M., et al.: Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama 318(22), 2199-2210 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.372, + 0.785, + 0.399 + ], + "angle": 0, + "content": "4. Bergner, B., Lippert, C., Mahendran, A.: Iterative patch selection for high-resolution image recognition. arXiv preprint arXiv:2210.13007 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.4, + 0.785, + 0.426 + ], + "angle": 0, + "content": "5. Bontempo, G., Bolelli, F., Porrello, A., Calderara, S., Ficarra, E.: A graph-based multi-scale approach with knowledge distillation for wsi classification. TMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.427, + 0.785, + 0.481 + ], + "angle": 0, + "content": "6. Brancati, N., Anniciello, A.M., Pati, P., Riccio, D., Scognamiglio, G., Jaume, G., De Pietro, G., Di Bonito, M., Foncubierta, A., Botti, G., et al.: Bracs: A dataset for breast carcinoma subtyping in h&e histology images. Database 2022, baac093 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.537 + ], + "angle": 0, + "content": "7. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine 25(8), 1301-1309 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.785, + 0.578 + ], + "angle": 0, + "content": "8. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.633 + ], + "angle": 0, + "content": "9. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15661-15670 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.689 + ], + "angle": 0, + "content": "0. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144-16155 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.744 + ], + "angle": 0, + "content": "1. Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4025 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "2. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23936-23945 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Chikontwe, P., Kim, M., Nam, S.J., Go, H., Park, S.H.: Multiple instance learning with center embeddings for histopathology classification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.128 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23. pp. 519-528. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "14. Cornish, T.C., Swapp, R.E., Kaplan, K.J.: Whole-slide imaging: routine pathologic diagnosis. Advances in anatomic pathology 19(3), 152-159 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "15. Dauphin, Y.N., Fan, A., Auli, M., Grangier, D.: Language modeling with gated convolutional networks. In: International conference on machine learning. pp. 933-941. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "16. Dehaene, O., Camara, A., Moindrot, O., de Lavergne, A., Courtiol, P.: Self-supervision closes the gap between weak and strong supervision in histology. arXiv preprint arXiv:2012.03583 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "17. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "18. Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T.: Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence 89(1-2), 31-71 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "19. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "20. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.48 + ], + "angle": 0, + "content": "21. Guan, Y., Zhang, J., Tian, K., Yang, S., Dong, P., Xiang, J., Yang, W., Huang, J., Zhang, Y., Han, X.: Node-aligned graph convolutional network for whole-slide image representation and classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18813-18823 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "23. He, L., Long, L.R., Antani, S., Thoma, G.R.: Histology image analysis for carcinoma detection and grading. Computer methods and programs in biomedicine 107(3), 538-556 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "24. Holdenried-Krafft, S., Somers, P., Montes-Majarro, I.A., Silimon, D., Tarín, C., Fend, F., Lensch, H.: Dual-query multiple instance learning for dynamic metaembedding based tumor classification. arXiv preprint arXiv:2307.07482 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "25. Hou, W., Yu, L., Lin, C., Huang, H., Yu, R., Qin, J., Wang, L.: \\(\\mathrm{H}^{\\wedge}\\) 2-mil: Exploring hierarchical representation with heterogeneous multiple instance learning for whole slide image analysis. In: AAAI. vol. 36, pp. 933-941 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.702 + ], + "angle": 0, + "content": "26. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 124-140. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "27. Ilse, M., Tomczak, J., Welling, M.: Attention-based deep multiple instance learning. In: International conference on machine learning. pp. 2127-2136. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "28. Kang, M., Song, H., Park, S., Yoo, D., Pereira, S.: Benchmarking self-supervised learning on diverse pathology datasets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3344-3354 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "29. Kong, F., Henao, R.: Efficient classification of very large images with tiny objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2384-2394 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Pro" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ACMIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318-14328 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.23 + ], + "angle": 0, + "content": "31. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7454-7463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.786, + 0.271 + ], + "angle": 0, + "content": "32. Li, R., Yao, J., Zhu, X., Li, Y., Huang, J.: Graph cnn for survival analysis on whole slide pathological images. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 174-182. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.786, + 0.298 + ], + "angle": 0, + "content": "33. Li, Y., Ping, W.: Cancer metastasis detection with neural conditional random field. arXiv preprint arXiv:1806.07064 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.786, + 0.338 + ], + "angle": 0, + "content": "34. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19830-19839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.786, + 0.393 + ], + "angle": 0, + "content": "35. Litjens, G., Sánchez, C.I., Timofeeva, N., Hermsen, M., Nagtegaal, I., Kovacs, I., Hulsbergen-Van De Kaa, C., Bult, P., Van Ginneken, B., Van Der Laak, J.: Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis. Scientific reports 6(1), 26286 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.786, + 0.434 + ], + "angle": 0, + "content": "36. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering 5(6), 555-570 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.786, + 0.461 + ], + "angle": 0, + "content": "37. Madabhushi, A.: Digital pathology image analysis: opportunities and challenges. Imaging in medicine 1(1), 7 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.786, + 0.487 + ], + "angle": 0, + "content": "38. Maron, O., Lozano-Pérez, T.: A framework for multiple-instance learning. Advances in neural information processing systems 10 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.786, + 0.514 + ], + "angle": 0, + "content": "39. McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.786, + 0.555 + ], + "angle": 0, + "content": "40. Pantanowitz, L., Valenstein, P.N., Evans, A.J., Kaplan, K.J., Pfeifer, J.D., Wilbur, D.C., Collins, L.C., Colgan, T.J.: Review of the current state of whole slide imaging in pathology. Journal of pathology informatics 2(1), 36 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.555, + 0.786, + 0.596 + ], + "angle": 0, + "content": "41. Pinckaers, H., Van Ginneken, B., Litjens, G.: Streaming convolutional neural networks for end-to-end learning with multi-megapixel images. IEEE transactions on pattern analysis and machine intelligence 44(3), 1581-1590 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.596, + 0.786, + 0.622 + ], + "angle": 0, + "content": "42. Qu, L., Wang, M., Song, Z., et al.: Bi-directional weakly supervised knowledge distillation for whole slide image classification. Neurips 35, 15368-15381 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.622, + 0.786, + 0.677 + ], + "angle": 0, + "content": "43. Qu, L., Yang, Z., Duan, M., Ma, Y., Wang, S., Wang, M., Song, Z.: Boosting whole slide image classification from the perspectives of distribution, correlation and magnification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21463-21473 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.677, + 0.786, + 0.718 + ], + "angle": 0, + "content": "44. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems 34, 2136-2147 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.786, + 0.758 + ], + "angle": 0, + "content": "45. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "46. Tang, W., Huang, S., Zhang, X., Zhou, F., Zhang, Y., Liu, B.: Multiple instance learning framework with masked hard instance mining for whole slide image classification. arXiv preprint arXiv:2307.15254 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "47. Tellez, D., Litjens, G., van der Laak, J., Ciompi, F.: Neural image compression for gigapixel histopathology image analysis. IEEE transactions on pattern analysis and machine intelligence 43(2), 567-578 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "48. Tiwari, R., Shenoy, P.: Overcoming simplicity bias in deep networks using a feature sieve (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "49. Wang, D., Khosla, A., Gargeya, R., Irshad, H., Beck, A.H.: Deep learning for identifying metastatic breast cancer. arXiv preprint arXiv:1606.05718 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.244 + ], + "angle": 0, + "content": "50. Wang, H., Luo, L., Wang, F., Tong, R., Chen, Y.W., Hu, H., Lin, L., Chen, H.: Iteratively coupled multiple instance learning from instance to bag classifier for whole slide image classification. arXiv preprint arXiv:2303.15749 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "51. Wang, X., Xiang, J., Zhang, J., Yang, S., Yang, Z., Wang, M.H., Zhang, J., Yang, W., Huang, J., Han, X.: Scl-wc: Cross-slide contrastive learning for weakly-supervised whole-slide image classification. Advances in neural information processing systems 35, 18009-18021 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.302, + 0.785, + 0.355 + ], + "angle": 0, + "content": "52. Wang, Y., Kartasalo, K., Weitz, P., Acs, B., Valkonen, M., Larsson, C., Ruusuvuori, P., Hartman, J., Rantalainen, M.: Predicting molecular phenotypes from histopathology images: A transcriptome-wide expression-morphology analysis in breast cancer. Cancer research 81(19), 5115-5126 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "53. Xiong, C., Chen, H., Sung, J., King, I.: Diagnose like a pathologist: Transformer-enabled hierarchical attention-guided multiple instance learning for whole slide image classification. arXiv preprint arXiv:2301.08125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.399, + 0.785, + 0.453 + ], + "angle": 0, + "content": "54. Yang, J., Chen, H., Zhao, Y., Yang, F., Zhang, Y., He, L., Yao, J.: Remix: A general and efficient framework for multiple instance learning based whole slide image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 35-45. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.455, + 0.785, + 0.493 + ], + "angle": 0, + "content": "55. Yao, J., Zhu, X., Jonnagaddala, J., Hawkins, N., Huang, J.: Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis 65, 101789 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.549 + ], + "angle": 0, + "content": "56. Yufei, C., Liu, Z., Liu, X., Liu, X., Wang, C., Kuo, T.W., Xue, C.J., Chan, A.B.: Bayes-mil: A new probabilistic perspective on attention-based multiple instance learning for whole slide images. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.551, + 0.785, + 0.605 + ], + "angle": 0, + "content": "57. Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18802-18812 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.607, + 0.785, + 0.659 + ], + "angle": 0, + "content": "58. Zhang, Y., Sun, Y., Li, H., Zheng, S., Zhu, C., Yang, L.: Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 242-252. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.785, + 0.729 + ], + "angle": 0, + "content": "59. Zhao, Y., Yang, F., Fang, Y., Liu, H., Zhou, N., Zhang, J., Sun, J., Yang, S., Menze, B., Fan, X., et al.: Predicting lymph node metastasis using histopathological images based on multiple instance learning with deep graph convolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4837-4846 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.731, + 0.785, + 0.77 + ], + "angle": 0, + "content": "60. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13001-13008 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "61. Zhu, X., Yao, J., Zhu, F., Huang, J.: Wsisa: Making survival prediction from whole slide histopathological images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7234-7242 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.812 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_origin.pdf b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..60a9f1dcbdbfc3d474d452e2f56716fe0cc5aaa2 --- /dev/null +++ b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/8ca43dc8-8fea-4515-a05f-f88a9e5bd536_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d997373a10e3e5d952de56ffb62244288320264bb8caad1acc0fecbcea17d549 +size 2018838 diff --git a/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/full.md b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2225da44b578d3ffe89cc00f4d89fdf4e0054050 --- /dev/null +++ b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/full.md @@ -0,0 +1,327 @@ +# Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification + +Yunlong Zhang $^{1,2,3}$ , Honglin Li $^{1,2,3}$ , Yunxuan Sun $^{1,2,3}$ , Sunyi Zheng $^{2,3}$ , Chenglu Zhu $^{2,3}$ , and Lin Yang $^{2,3\star}$ + +$^{1}$ Zhejiang University + $^{2}$ Research Center for Industries of the Future, Westlake University + $^{3}$ School of Engineering, Westlake University + +Abstract. In the application of Multiple Instance Learning (MIL) methods for Whole Slide Image (WSI) classification, attention mechanisms often focus on a subset of discriminative instances, which are closely linked to overfitting. To mitigate overfitting, we present Attention-Challenging MIL (ACMIL). ACMIL combines two techniques based on separate analyses for attention value concentration. Firstly, UMAP of instance features reveals various patterns among discriminative instances, with existing attention mechanisms capturing only some of them. To remedy this, we introduce Multiple Branch Attention (MBA) to capture more discriminative instances using multiple attention branches. Secondly, the examination of the cumulative value of Top-K attention scores indicates that a tiny number of instances dominate the majority of attention. In response, we present Stochastic Top-K Instance Masking (STKIM), which masks out a portion of instances with Top-K attention values and allocates their attention values to the remaining instances. The extensive experimental results on three WSI datasets with two pre-trained backbones reveal that our ACMIL outperforms state-of-the-art methods. Additionally, through heatmap visualization and UMAP visualization, this paper extensively illustrates ACMIL's effectiveness in suppressing attention value concentration and overcoming the overfitting challenge. The source code is available at https://github.com/dazhangyu123/ACMIL. + +Keywords: Computational pathology $\cdot$ Whole slide image $\cdot$ Multiple instance learning $\cdot$ Overfitting + +# 1 Introduction + +Whole slide image (WSI) classification is a critical undertaking in digital pathology, aiming to extract valuable information from high-resolution scanned images for precise diagnosis [23,33,49,52], prognosis [11,32,55,61], and treatment planning [14,35,37,40,41] of diseases. In recent years, multiple instance learning (MIL) [1,18,38] has emerged as a promising approach for WSI classification, treating each WSI as a "bag" and its extracted small patches as "instances" + +![](images/413de03d7ce586999586037886ea55821489f8ebaa93793be039324604fa6390.jpg) +Fig. 1: The change of validation loss and entropy of attention values throughout the training of ABMIL. The results are reported on LBC with SSL pretrained features. There exists the strong negative correlation between loss and entropy. + +![](images/0e4b5dfdf12709684e652abda20d7881fe07a5fabafeecf6a87f09ddb0931b52.jpg) +Fig.2: Comparison of AUC and entropy of attention values between ABMIL and ACMIL. One point denotes the result of a seed on LBC with SSL pretrained features. ACMIL achieves the higher AUC and entropy than ABMIL. + +within the bag, thus enabling efficient classification of WSIs through assigning a single label to the entire slide. + +Overfitting is a significant challenge in utilizing MIL methods for WSI classification [34, 46, 57]. Common WSI datasets exhibit intrinsic characteristics of limited data scale, ultra-high resolutions, and staining bias, which makes overfitting more likely [2]. Specifically, these datasets often consist of a relatively small number of slides, typically in the hundreds, with a typical resolution of $10,000 \times 10,000$ [36]. Moreover, pathology images are susceptible to staining bias caused by variations in tissue preparations, staining protocols, and digital scanning methods [58], leading models to learn spurious correlations [34]. + +In the attention mechanism, attention values/heatmaps provide insights into the model's decision-making process. Multiple existing works [36,44,46,56] alongside our own experiments (as indicated in Sec. 4.3) have pointed out the excessive concentration of attention values in current MIL methods. Specifically, MIL's attention mechanisms often concentrate on a subset of discriminative instances (i.e., instances relevant to the bag label) while disregarding the remaining ones. We investigate the correlation between attention value concentration and overfitting, utilizing the entropy of attention values and validation loss. Fig. 1 depicts a negative correlation between loss and entropy throughout the training process, illustrating that over-concentration of attention values (indicated by lower entropy) significantly compromises the model's generalization ability (indicated by higher loss values). Moreover, in the field of natural image classification, recent studies [17,26,48] have demonstrated that models solely relying on a portion of discriminative features could be susceptible to overfitting. Transitioning to WSI classification, fixating on a subset of discriminative instances similarly impedes the model's ability to generalize. These findings highlight the tight connection between attention value concentration and overfitting. + +Recently, numerous efforts have been made to address the overfitting challenge by enhancing representation quality [16, 24, 31, 36, 51], building spatial + +instance correlations [21, 30, 44] and developing data augmentation methods [13, 42, 46, 54, 57]. Additionally, some of these studies [30, 36] suggest that reducing the concentration of attention values can enhance model interpretability. However, the investigation of attention values concentration for alleviating overfitting remains under-explored. + +To mitigate overfitting, we present two analyses for attention value concentration using UMAP and Top-K value statistics. Then, we introduce Attention-Challenging MIL (ACMIL), which combines two novel techniques based on these two analyses. First, by observing UMAP of instance features, we find that there are various patterns among discriminative instances, and attention mechanisms tend to capture some of them. To solve this, we introduce Multiple Branch Attention (MBA). MBA utilizes multiple attention branches, each focusing on capturing instances with a specific pattern, thereby ensuring that more discriminative instances contribute to the final prediction. Second, by analyzing the cumulative value of Top-K attention scores, we find that a tiny number of instances (e.g., $K = 10$ ) occupy majority attention, resulting in overlooking sophisticated discriminative instances. To suppress these instances, we propose Stochastic Top-K Instance Masking (STKIM). STKIM randomly masks out a portion of instances with Top-K attention values and assigns their attention values to the remaining instances. Combining MBA and STKIM, our ACMIL effectively alleviates the attention value concentration and suppresses overfitting (see Fig. 2). + +We conduct experiments on three WSI datasets (i.e., CAMELYON16, BRACS, and our in-house LBC dataset) with two backbones (i.e., ImageNet pre-trained ResNet18 and SSL pre-trained ViT/S-16). Experimental results demonstrate the superiority of our ACMIL over existing state-of-the-art methods. We also present substantial experimental results, including heatmap visualization and UMAP visualization, to comprehensively demonstrate the effectiveness of ACMIL in suppressing attention value concentration and combatting overfitting. + +# 2 Related Work + +# 2.1 Combating Overfitting in WSI Analysis + +In the domain of WSI classification, combating the challenge of overfitting has received substantial attention. Next three paragraphs detail methods from three different aspects. + +Some efforts have concentrated on enhancing the quality of instance representations. Early studies (e.g., [7, 27, 44]) rely on backbones pre-trained on the ImageNet dataset. However, the substantial domain gap between natural and pathological images hinders representation quality. Recent works (e.g., [16, 24, 36, 47, 51]) address this by emphasizing Self-Supervised Learning (SSL) to learn patch-level feature representations. In addition, efforts such as the work by Chen et al. [10] leverage hierarchical SSL for high-resolution image representations. Further, studies by Li et al. [31] and Wang et al. [50] demonstrate that fine-tuning the pre-trained encoder is essential for acquiring task-specific information. + +Another line of research has focused on establishing spatial instance correlations. DSMIL [30], $\mathrm{H}^2\mathrm{MIL}$ [25], and DAS-MIL [5] consider the hierarchical structure of patches and aggregate multi-scale representations in attention mechanisms. Furthermore, some studies introduce self-attention layers [43, 44, 53] and graph neural networks [9, 21, 59] to model correlations between different areas. + +Further strategies have concentrated on data augmentation. Examples include DTFD-MIL [57], which introduces pseudo-bags for expanding bag counts and employs a double-tier MIL framework. IPS [4], Zoom-In Network [29], and Top-K MIL [13] generate bag representations by aggregating the representations of salient patches. Remix [54] and RankMix [12] introduce instance representation mixup for MIL. MHIM-MIL [46] and WENO [42] augments bags by randomly masking salient instances. + +Although our ACMIL shares a similar spirit with some of these works, the proposed ACMIL excels by further building on detailed analysis for attention value concentration. As a result, our ACMIL exhibits stronger interpretability against existing solutions. + +# 2.2 Over-Concentration of Attention Values + +In the realm of natural image classification, research has shown that an excessive focus on certain parts of an object can impede the overall effectiveness of model generalization [17,26,60]. To tackle this issue, various heuristic techniques have been proposed. For instance, Cutout [17, 60] is a valuable data augmentation method that randomly masks square regions of input during training. Another approach, RSC [26], involves regularization that eliminates salient features activated during training. This paper investigates the issue of attention value concentration in WSI classification tasks. We identify two specific phenomena related to attention value concentration existing in WSI classification and propose two techniques to address them respectively. + +# 3 Method + +Based on the ABMIL (detailed in Sec. 3.1), we present ACMIL to alleviate the overfitting problem, which is built on two components: Multiple Branch Attention (MBA) and Stochastic Top-K Instance Masking (STKIM). We describe the details of two components in Sec. 3.2 and 3.3, respectively. + +# 3.1 ABMIL for WSI Classification + +In the binary MIL classification problem [7], a bag of instance, $\mathbf{X} = \{\mathbf{x}_n\}_{n=1}^N$ , is associated with a single bag label, $\mathbf{Y}$ . Each instance, $x_n$ , is associated with a single binary label, $y_n$ , which remains unknown during training. The assumption behind the MIL can be written as: + +$$ +\boldsymbol {Y} = \left\{ \begin{array}{l l} 0, & \text {i f f} \sum_ {n = 1} ^ {N} y _ {n} = 0 \\ 1, & \text {o t h e r w i s e} \end{array} \right. \tag {1} +$$ + +![](images/e2ff3924e97d98c54d623d2df01d56a2c2c568af919a487ae618e6e2ce1d79a8.jpg) +Fig.3: Motivation of MBA. UMAP visualization [39] of tumor instance features from CAMELYON16 'test_113' case. There are various patterns/clusters among tumor instances, and relying on one single branch tends to capture a part of clusters. Three instances are selected to exhibit their texture differences. + +![](images/f0fc29677695b752a1871f4da125dfed8bafae63b059ab1e3a22efe85972c58e.jpg) +Fig.4: Motivation of STKIM. Accumulation of Top-K attention values. Instances with Top-K attention values occupy majority attention. Results are derived from features extracted through supervised pretraining. + +In the ABMIL [27], the multiple instance learning is modeled by a three-step process. i) Instance transformation into a low-dimensional embedding through neural networks: $\pmb{h}_n = f(\pmb{x}_n)$ . ii) Aggregation of all instance embeddings into the bag-level representation using an attention operator. Specifically, this operation is defined as: + +$$ +\boldsymbol {z} = \sum_ {n = 1} ^ {N} a _ {n} \boldsymbol {h} _ {n} \tag {2} +$$ + +Here, $a_{n} = \sigma(\pmb{h}_{n})$ represents the attention values for $n$ -th instance, $\pmb{h}_{n}$ . In the case of ABMIL, a gated attention (GA) mechanism [15] is adopted: + +$$ +\sigma \left(\boldsymbol {h} _ {n}\right) = \frac {\exp \left\{\boldsymbol {w} ^ {\mathrm {T}} \left(\tanh \left(\boldsymbol {V} _ {1} \boldsymbol {h} _ {n}\right) \odot \operatorname {s i g m} \left(\boldsymbol {V} _ {2} \boldsymbol {h} _ {n}\right)\right) \right\}}{\sum_ {j = 1} ^ {N} \exp \left\{\boldsymbol {w} ^ {\mathrm {T}} \left(\tanh \left(\boldsymbol {V} _ {1} \boldsymbol {h} _ {j}\right) \odot \operatorname {s i g m} \left(\boldsymbol {V} _ {2} \boldsymbol {h} _ {j}\right)\right) \right\}} \tag {3} +$$ + +where $\mathbf{V}_1, \mathbf{V}_2 \in \mathbb{R}^{L \times M}$ , $\mathbf{w} \in \mathbb{R}^{L \times 1}$ are parameters, $\odot$ is an element-wise multiplication and $\operatorname{sigm}(\cdot)$ is the sigmoid non-linearity. iii) The bag prediction is generated based on the aggregated bag embedding: $\hat{\mathbf{Y}} = g(\mathbf{z})$ . + +# 3.2 Mutiple Branch Attention + +Motivation. It is challenging to capture all discriminative instances using a single attention branch (see Fig. 3). This challenge arises due to variations in patterns among discriminative patches, stemming from differences in texture and morphology. Additionally, DNNs tend to exhibit a form of "laziness" where they prioritize capturing simpler patterns to minimize training loss, neglecting more intricate and challenging patterns [19, 20]. To tackle this issue, we design the MBA that captures more discriminative instances by multiple attention branches. + +As depicted in Fig. 5 top view, the MBA firstly captures $M$ patterns and then aggregates their embeddings to make predictions. Each pattern is captured by an attention branch. To maintain both the discriminative nature of patterns and semantic diversity between them, we introduce two regularization techniques: semantic regularization and diversity regularization. Firstly, to ensure capturing discriminative patterns, the semantic regularization is accomplished by hanging a MLP layer behind each pattern embedding, equipping with the cross entropy loss function: + +$$ +\mathcal {L} _ {p} = - \frac {1}{M} \sum_ {i = 1} ^ {M} \boldsymbol {Y} \log \hat {\boldsymbol {Y}} _ {i} + (1 - \boldsymbol {Y}) \log \left(1 - \hat {\boldsymbol {Y}} _ {i}\right) \tag {4} +$$ + +where $\hat{\mathbf{Y}}_i = g_i(\mathbf{z}_i)$ is the prediction based on $i$ -th pattern embedding, $\mathbf{z}_i$ . However, only equipping with cross-entropy loss may learn similar patterns and cannot dig out more discriminative information. To tackle this issue, we further introduce a diversity loss as follows: + +$$ +\mathcal {L} _ {d} = \frac {2}{M (M - 1)} \sum_ {i = 1} ^ {M} \sum_ {j = i + 1} ^ {M} \cos \left(\boldsymbol {a} _ {i}, \boldsymbol {a} _ {j}\right) \tag {5} +$$ + +where $\pmb{a}_i$ consists of all attention values of $i$ -th pattern, $\pmb{a}_i = \{a_{i1},\dots ,a_{iN}\}$ , also named heatmap as custom. The $\cos (\cdot)$ function is used to measure the similarity of the heatmaps between branches. By diversifying the heatmaps, the embedding of each branch can concentrate on different patterns. + +To aggregate the captured patterns to make predictions, the average of heatmaps is utilized as the heatmap of the whole bag: + +$$ +\boldsymbol {a} = \frac {1}{M} \sum_ {i = 1} ^ {M} \boldsymbol {a} _ {i} \tag {6} +$$ + +where $\pmb{a}$ is the heatmap of the whole bag, with a dimension of $N$ . Then, the bag embedding can be obtained by aggregating the instance features using averaged heatmap $\pmb{a}$ . Moreover, since $\sum_{n=1}^{N} \left( \frac{1}{M} \sum_{i=1}^{M} a_{in} \right) \pmb{h}_n = \frac{1}{M} \sum_{i=1}^{M} \left( \sum_{n=1}^{N} a_{in} \pmb{h}_n \right)$ , the bag embedding also can be formulated by applying mean pooling operator to pattern embeddings. The top view of Fig. 5 adopts the latter formulation for brevity. The loss function for the bag classifier is defined as: + +$$ +\mathcal {L} _ {b} = - \boldsymbol {Y} \log \hat {\boldsymbol {Y}} + (1 - \boldsymbol {Y}) \log \left(1 - \hat {\boldsymbol {Y}}\right) \tag {7} +$$ + +Finally, the overall loss function for the ACMIL can be written as the combination of three loss terms defined in Eq. 4, 5 and 7, + +$$ +\mathcal {L} = \mathcal {L} _ {b} + \mathcal {L} _ {p} + \mathcal {L} _ {d} \tag {8} +$$ + +Discussion. It's important to highlight that when the parameter $M$ is set to 1 in MBA, it essentially mirrors the feature aggregation process of ABMIL, allowing + +![](images/7f6a0513029876194d765ff31c29dd3a7fee35bc1f41bc92eaacb935f4b37ee3.jpg) + +![](images/e88986ad247e602ec3171684de6cf1c8ab668f665c3070ad37334e820913666b.jpg) +Fig. 5: Overview of the proposed MBA (top view) and STKIM (bottom view). In the MBA, $M$ discriminative patterns are extracted from patch features using the attention operator regularized by semantic and diversity regularization terms. Then, the mean operator is applied to these $M$ pattern features to produce the bag feature, which is utilized for bag-level prediction. In the STKIM, instances with Top-K attention values are randomly masked with a probability $p$ . + +for the discernment of a single pattern. In this sense, MBA serves as an extension of ABMIL, specifically designed to capture a more diverse set of patterns. We further discuss the connection between MBA and Multiple-Head Attention (MHA). HIPT [10] has unveiled that distinct heads in MHA can effectively capture different visual concepts, akin to the role played by our MBA. However, these two techniques can be easily distinguished by: 1) MBA has diversity regularization, ensuring that different branches can learn different concepts. This is absent in MHA, resulting in different heads learning the same concept [10]. We demonstrate the critical role of $\mathcal{L}_d$ for performance in Tab. 3b. 2) MHA is a type of attention formulation, while MBA operates independently of the attention formulation, accommodating MHA within its framework. Appendix Sec. 4.1 discusses the results of combining MHA and ACMIL. + +# 3.3 Stochastic Top-K Instance Masking + +Motivation. A tiny number of instances will occupy the majority of attention in ABMIL while ignoring sophisticated discriminative instances. As depicted in Fig. 4, the sum of top-10 attention values is larger than 0.85 on all three datasets. However, the WSI typically involves more than 10 discriminative instances. For instance, in the CAMELYON16 dataset, 129 out of 155 tumor slides contain 10 to 20,000 cancerous instances. In essence, numerous discriminative instances are overlooked. To deal with this issue, the proposed STKIM aims to suppress the salient instances and assign more attention to the remaining instances. + +As depicted in Fig. 5 bottom view, STKIM introduces a masking operation into the attention mechanism, before feature aggregation and after attention values generation. The primary objective is to suppress Top-K salient instances. A straightforward solution to achieve this is to mask out all of the Top-K salient instances. However, this method poses certain challenges. It can result in the loss of information associated with key instances, which are crucial for discrimination. Furthermore, it might lead to a statistical mismatch between the feature representations before and after discarding these key instances. To address these issues, we draw inspiration from dropout [45] and cutout [17,60] commonly used in computer vision. Our proposed solution employs stochastic masking for instance features with Top-K attention values. Specifically, we begin by sorting all attention values from highest to lowest. Subsequently, we randomly set the attention values of the Top-K instances to 0, with a probability of $p$ . This process can be formulated as: + +$$ +a _ {n} = \left\{ \begin{array}{l l} 0, & \text {w i t h p r o b a b i l i t y} p \text {a n d w i t h i n T o p - K v a l u e s} \\ a _ {n}, & \text {o t h e r w i s e} \end{array} \right. \tag {9} +$$ + +where $p$ and $K$ are two hyperparameters that control the intensity of masking. Following Eq. 9, we will assign the attention values of masked instances to the remaining instances by $a_{n} \to \frac{1}{\sum_{n=1}^{N} a_{n}} a_{n}$ . Notably, drawing inspiration from dropout and cutout, we remove STKIM at the inference time. + +Discussion. While STKIM, MHIM-MIL [46], and WENO [42] all adopt the technique that masks salient instances, there are notable technical distinctions between them. Firstly, our STKIM has the different masking strategy compared with WENO and MHIM-MIL. STKIM only masks a minority of instances (i.e., $K = 10$ ) with a probability of $p$ . As a comparison, the other two methods mask out a larger number of instances. WENO masks out 95 instances. MHIM-MIL masks $1\%$ instances. In our framework, our scheme performs best in three strategies (see Appendix Tab. 2). Secondly, both MHIM-MIL and WENO necessitate a well-trained model for masking out salient instances, utilizing the remaining instances for model training. They both employ a teacher-student framework, wherein the teacher model needs to be pre-trained beforehand (the warm-up process in WENO and the pre-training stage in MHIM-MIL). In contrast, STKIM requires neither a teacher-student framework nor a pre-training process, thus highlighting simplicity and efficiency. + +# 4 Experiments + +# 4.1 Experimental Details + +Datasets and Evaluation Metrics. The performance of ACMIL is evaluated on two public WSI datasets, i.e., CAMELYON16 [3] and BRACS [6], and one private benchmark, LBC. CAMELYON16 dataset consists of 400 WSIs in total, including 270 for training and 130 for testing. Following [30, 57], we further randomly split the training and validation sets from the official training set with a + +Table 1: The performance of different MIL approaches across three datasets, two pre-trained methods, and two evaluation metrics. The most superior performance is highlighted in bold, while the second-best performance is indicated by underlining. + +
Performance MethodCAMELYON-16BRACSLBC
F1-scoreAUCF1-scoreAUCF1-scoreAUC
ResNet-18Max-pooling0.582±0.1700.620±0.1550.489±0.0470.738±0.0140.476±0.0330.775±0.010
Mean-pooling0.592±0.0260.597±0.0330.484±0.0290.685±0.0110.511±0.0220.797±0.011
Clam-SB0.742±0.0240.763±0.0490.521±0.0460.750±0.0390.514±0.0240.805±0.017
TransMIL0.643±0.0880.706±0.0760.444±0.0400.732±0.0430.385±0.0130.693±0.027
DSMIL0.736±0.0280.773±0.0340.511±0.0520.751±0.0280.458±0.0290.766±0.023
DTFD-MIL0.758±0.0510.815±0.0630.469±0.0160.717±0.0320.473±0.0210.776±0.021
IBMIL0.777±0.0090.799±0.0500.510±0.0430.726±0.0340.489±0.0170.791±0.021
MHIM-MIL0.752±0.0340.772±0.0260.511±0.0220.774±0.0210.543±0.0370.816±0.009
ABMIL0.757±0.0200.790±0.0270.523±0.0280.723±0.0350.465±0.0400.798±0.013
ACMIL(ours)0.798±0.0290.841±0.0300.552±0.0480.754±0.0080.546±0.0280.821±0.015
VIT-S/16Max-pooling0.903±0.0540.956±0.0290.596±0.0290.823±0.0330.590±0.0430.829±0.023
Mean-pooling0.577±0.0570.569±0.0810.522±0.0380.739±0.0070.559±0.0240.827±0.012
Clam-SB0.925±0.0350.969±0.0240.631±0.0340.863±0.0050.617±0.0220.865±0.018
TransMIL0.922±0.0190.943±0.0090.631±0.0300.841±0.0060.539±0.0280.805±0.010
DSMIL0.943±0.0070.966±0.0090.577±0.0280.816±0.0280.562±0.0280.820±0.033
DTFD-MIL0.948±0.0070.980±0.0110.612±0.0800.870±0.0220.612±0.0340.842±0.010
IBMIL0.912±0.0340.954±0.0220.645±0.0410.871±0.0140.604±0.0320.834±0.014
MHIM-MIL0.932±0.0240.970±0.0370.625±0.0600.865±0.0170.658±0.0410.872±0.022
ABMIL0.914±0.0310.945±0.0270.680±0.0510.866±0.0290.595±0.0360.831±0.022
ACMIL(ours)0.954±0.0120.974±0.0120.722±0.0300.888±0.0100.662±0.0430.901±0.011
+ +ratio of 9:1. We do not resplit BRACS dataset as it has been officially split to 395 of training set, 65 of validation set, and 87 of test set. We follow the challenge for a 3-class WSI classification: benign tumor, atypical tumor, and malignant. The liquid-based cytology (LBC) dataset collected 1,989 WSIs and included 4 classes, i.e., Negative, ASC-US, LSIL, and ASC-H/HSIL. We randomly split the whole dataset into training, validation, and test sets with the ratio of 6:2:2. Following [31], macro-AUC and macro-F1 scores are reported since all three datasets are class imbalanced. Each of the main experiments is performed five times with random parameter initializations, and the average classification performance and standard deviation are reported. Besides, following [36,57], the test performance is reported in epochs with the best validation performance. + +Baselines. We systematically assess the efficacy of our ACMIL approach by benchmarking it against conventional MIL pooling strategies, Max-pooling and Mean-pooling, as well as contemporary attention-based techniques such as AB-MIL [27], DSMIL [30], TransMIL [44], CLAM-SB [36], DTFD-MIL [57], MHIM-MIL [46], and IBMIL [34]. In pursuit of a comprehensive comparison across diverse aggregation operators, we utilize two distinct sets of features derived from ResNet-18 pre-trained on the ImageNet dataset [22] and ViT-S/16 pretrained using DINO [8, 28]. The results of all other methods are reproduced using the official code they provide under the same settings. + +Implementation Details. Implementation Details are in Appendix Sec. 3. + +![](images/8c9d840be914cf556d82760eacc6c5040b0eeb601e2c0c4ebe1752ed51805fa7.jpg) +Fig. 6: Heatmap visualization of WSI examples produced by ABMIL [27] (baseline) and our ACMIL. The left part shows three tumor WSIs that come from CAMELYON16 dataset, and their tumor regions are delineated by red lines. ACMIL generates attention values that cover a more extensive portion of the tumor region compared to ABMIL. The right part shows three normal WSIs that come from CAMELYON16 dataset. ABMIL primarily focuses on a part of tissue such as adipose, while ACMIL extends its attention to the more normal tissues. + +# 4.2 WSI Classification Results + +Tab. 1 provides a thorough comparison of performance between ACMIL and existing MIL methods. This evaluation spans three diverse datasets, involves two different choices for pretraining methods, and employs two crucial evaluation metrics, resulting in a comprehensive assessment with a total of 12 terms. + +Considering the overall performance, ACMIL consistently outshines existing methods. It secures the top position in 10 out of the 12 metrics and holds the second position in the remaining 2 metrics. Specifically, for the CAMELYON16, ACMIL achieves outstanding results using ResNet-18 pre-trained on ImageNet embeddings, surpassing the runner-up by $2.1\%$ and $2.6\%$ in terms of F1-score and AUC, respectively. On the other hand, with ViT-S/16 SSL pretrained embeddings, existing attention-based MIL methods exhibit remarkable performance, boasting F1-scores and AUC values exceeding 0.9. Notably, ACMIL achieves comparable performance with the former best-performing method, DTFD-MIL, in this setup. For the BRACS, ACMIL demonstrates a substantial lead when utilizing ViT-S/16 SSL pre-trained embeddings, surpassing the second-best performance by margins of $4.2\%$ and $1.7\%$ in F1-score and AUC, respectively. Moreover, when employing ResNet-18 pre-trained on ImageNet embeddings, ACMIL achieves comparable performance with the previously top-performing method, MHIM-MIL. For the LBC, ACMIL stands out significantly among the other methods across all four metrics. + +Table 2: Comparison of FROC between ABMIL and ACMIL + +
ABMILACMIL
FROC0.39870.4233
+ +# 4.3 Localization Results + +Heatmap visualization. Fig. 6 presents heatmap visualizations illustrating examples of our approach's performance in comparison to the baseline method, ABMIL [27]. Three tumor slides (left part) and three normal slides (right part) are selected to showcase the heatmap differences. Due to the space limitation, we present more visualization in Appendix Sec. 4.4 for further insights. + +For the tumor slides, ABMIL tends to concentrate its attention on only a fraction of the tumor regions, potentially overlooking other significant areas. In contrast, ACMIL allocates attention across a wider spectrum of tumor regions, resulting in better alignment with expert annotations. For the normal slides, ABMIL predominantly focuses on specific tissue types, such as adipose tissue. This will lead to misinterpretation that only the adipose tissue is the normal tissue and other normal regions are uncorrelated to the WSI label. On the other hand, ACMIL effectively distributes attention values to encompass all normal regions, ensuring all regions are correlated for the WSI label. This approach closely mimics human intuition and satisfies the definition of the MIL formulation. + +FROC results. We employ the FROC metric suggested by CAMELYON16 challenge to evaluate the localization of tumor region quantitatively. As shown in Table 2, the proposed ACMIL achieves higher FROC than ABMIL. + +# 4.4 Ablation Study + +Fig. 7 illustrates the AUC scores of ACMIL across three datasets when utilizing a ViT/B-16 feature extractor and varying hyperparameter settings. Several key observations emerge from these experiments: + +Effect of branches number $M$ in MBA. As shown in the first column, we find that the choice of $M$ affects performance significantly. Combining three datasets, setting $M = 5$ achieves the best performance. + +Effect of masking probability $p$ in STKIM. As shown in the second column, we find that the choice of $M$ also affects performance significantly. Notably, setting $p = 1.0$ (masking all of Top-K instances) leads to performance deterioration across all three datasets. For LBC and CAMELYON, a $p = 1.0$ setting even results in performance lower than the blue dotted lines. Otherwise, $p = 0.6$ achieves the best performance on the BRACS dataset, whereas $p = 0.8$ achieves the best performance on the other two datasets. + +Effect of number of masking instances $K$ in STKIM. The third column shows that hyperparameter $K$ exhibits minimal sensitivity, where different $K$ + +![](images/54ac08292af7c9054c16ee6cc9477d59141448dece695701f16a2e880da4cbb0.jpg) + +![](images/339448ff6e229fbc134b0cf212d4daaa9f72d93452a73e4650027389658ef325.jpg) + +![](images/f05a0557acd9d6e1d5764c934854d43c230cf030f389ff74de7a04bac83684f2.jpg) + +![](images/9a0a32e90d00e30db54a3e8b19ab05655b69a83a355fafef165ef835a96385e4.jpg) + +![](images/2749d22ecf04fb1d930f7689f98dd0811134f123ce8e5bcd4d6cf349c1c7d15c.jpg) + +![](images/88fda8165a51d6c95aae69e0d0cc1022f1994e57fd2aadbe431454f333f0765a.jpg) + +![](images/b0ecfc5ff427dd98ffdf2f922cd0375f2be30632fe5f460f4c85dcfb531b78e6.jpg) +Fig. 7: Ablation study on features extracted through the SSL pre-training. The effect of three hyperparameters, $K$ , $p$ , $M$ , is investigated. Note that the orange dotted line denotes the performance of baseline, ABMIL, and the blue dotted line denotes the performance of ACMIL w/o MBA or STKIM. Five conclusions derived from the figure can be found in Sec. 4.4. + +![](images/459fba53739ea91e79c4f8eb3a0ea28bd0a69de4b463bc6fff10256e66f19456.jpg) + +![](images/f9cadb9552a5128e78f7bce3b6744575d58826ec239686f35f496355c9c09552.jpg) + +values result in a performance difference of less than $1.0\%$ AUC. In practice, setting $K$ to 10 is generally sufficient for achieving near-optimal performance. + +Implementing either MBA or STKIM individually leads to significant performance improvements. The blue dotted lines represent ACMIL's AUC performance without MBA or STKIM, outperforming the orange dotted lines (ABMIL's performance) across all subfigures. Particularly noteworthy is the observation that MBA achieves better improvement than STKIM on all three datasets, with blue dotted lines in the last two columns surpassing those in the first column, especially on the CAMELYON and LBC datasets. + +Combining MBA with STKIM yields greater performance improvements compared to using either MBA or STKIM alone. The green dots represent ACMIL's performance under different hyperparameter combinations, with 39 out of 45 green dots exceeding blue dotted lines. + +# 4.5 Further Analysis + +MBA can capture diverse patterns. We employ UMAP [39] to visualize instance features within the tumor region of the CAMELYON16 'test_090' case. + +![](images/902a8dcf03da0490d6facae9731650205a18dcdc21083092fa767d8e65f7051c.jpg) +(a) ABMIL + +![](images/556a2bb6e937e5576b2e69bfe409993d68a87c9f3aacdcaec0360ef0ba8c1969.jpg) +(b) ABMIL with MBA +Fig. 8: UMAP visualization of tumor instance features from the CAMELYON16 'test_090' case. The tumor instances display distinct patterns, posing a challenge for a single branch to capture all of them. As a result, ABMIL overlooks the right pattern/cluster. In contrast, multiple branches in MBA capture different patterns separately, and their combination enables the activation of more tumor instances. An instance is considered active when its attention value surpasses $\frac{1}{N}$ . + +In Fig. 8a, it's evident that the tumor instances exhibit two primary patterns. However, ABMIL primarily activates the left pattern (colored orange) and neglects the right one. On the other hand, as demonstrated in Fig. 8b, MBA's various branches (branch1, branch2, branch3, and branch5) collectively capture the substructures of the left pattern, while branch4 specifically captures the right pattern. Combining all branches can capture more comprehensive patterns. + +STKIM can suppress the concentration of Top-K attention values. Fig. 9 illustrates a comparison of the cumulative sum of Top-K attention values with and without STKIM. The plot clearly shows that the use of STKIM helps mitigate the scenario where Top-K attention values excessively dominate in the attention mechanism. This effect is particularly pronounced for CAMELYON16 dataset, where the cumulative sum of the top-10 values decreases from 0.87 to 0.6. + +Do we need STKIM at the test phase? The answer is No. In Tab. 3a, we present the outcomes of ACMIL with and without STKIM during the test phase, along with the performance differences between these settings. Across 11 + +out of 12 evaluation metrics, the version of ACMIL without STKIM during testing outperforms the version with STKIM slightly. This suggests that STKIM is not necessary during the test phase. + +![](images/53ee61c7fe7397ee4bbddb99184223db7989e93c2f2592b96e2caa11880ce52d.jpg) +Fig.9: Comparison of the accumulative sum of Top-K attention values with and without STKIM. The use of STKIM helps alleviate the issue of excessive concentration of attention values within the Top-K range. + +Table 3: (a): Performance comparison between ACMIL with (w.) and without (w/o.) T-STKIM. T-STKIM means using STKIM at the test phase. The Gap column reports the performance difference between with and without T-STKIM. Using STKIM at the test phase slightly reduces performance. (b): Performance comparison between ACMIL with (w.) and without (w/o.) $\mathcal{L}_d$ . The Gap column reports the performance difference between without and with $\mathcal{L}_d$ . ACMIL without $\mathcal{L}_d$ drastically reduces its performance. + +(a) w. T-STKIM v.s. w/o. T-STKIM + +
ViT-S/16 SSL pre-trained
DatasetMetricw. T-STKIMw/o. T-STKIMGap(%)
CamelyonF1-score0.927±0.0570.954±0.012+2.7
AUC0.967±0.0170.974±0.012+0.7
BRACSF1-score0.697±0.0330.722±0.030+2.5
AUC0.875±0.0120.888±0.010+1.3
LBCF1-score0.637±0.0340.662±0.043+2.5
AUC0.878±0.0120.901±0.011+2.3
ResNet-18Imagenet pre-trained
DatasetMetricw. T-STKIMw/o. T-STKIMGap(%)
CamelyonF1-score0.780±0.0260.798±0.029+1.8
AUC0.837±0.0280.841±0.030+0.4
BRACSF1-score0.566±0.0540.552±0.048-1.4
AUC0.750±0.0210.754±0.008+0.4
LBCF1-score0.535±0.0270.546±0.028+1.1
AUC0.808±0.0190.821±0.015+1.3
+ +(b) w. $\mathcal{L}_d$ v.s. w/o. $\mathcal{L}_d$ + +
ViT-S/16 SSL pre-trained
DatasetMetricw/o. Ldw. LdGap(%)
CamelyonF1-score0.901±0.0370.954±0.012+5.3
AUC0.943±0.0270.974±0.012+3.1
BRACSF1-score0.642±0.0460.722±0.030+8.0
AUC0.859±0.0200.888±0.010+2.9
LBCF1-score0.603±0.0230.662±0.043+5.9
AUC0.837±0.0090.901±0.011+6.4
ResNet-18Imagenet pre-trained
DatasetMetricw/o. Ldw. LdGap(%)
CamelyonF1-score0.747±0.0220.798±0.029+5.1
AUC0.796±0.0320.841±0.030+5.5
BRACSF1-score0.500±0.0310.552±0.048+5.2
AUC0.760±0.0260.754±0.008-0.6
LBCF1-score0.532±0.0190.546±0.028+1.4
AUC0.809±0.0180.821±0.015+1.2
+ +Do we need diversity loss in MBA? The answer is Yes. In Tab. 3b, we present the outcomes of ACMIL with and without $\mathcal{L}_d$ , along with the performance differences between these settings. Notably, the last column clearly indicates a significant performance drop for ACMIL without $\mathcal{L}_d$ . This emphasizes the crucial role of $\mathcal{L}_d$ in encouraging different branches to acquire distinctive discriminative knowledge within the MBA technique. + +Additional Analyses: Appendix presents further analyses that offer deeper insights into ACMIL's behavior. Specifically, Sec. 4.2 provides a detailed analysis of bag features, while Sec. 4.3 provides a comprehensive examination of evaluation metrics throughout the training process. + +# 5 Conclusion + +Due to the intrinsic properties of WSI, MIL methods have often led to overfitting, limiting their applications. This paper reveals that the overly concentrated attention values in the heatmap are closely related to overfitting. To address this, we propose ACMIL, which is underpinned by two novel techniques: MBA and STKIM. Our experimental results on three datasets demonstrate that ACMIL significantly surpasses SOTA methods. Moreover, this paper provides comprehensive experiments confirming the effectiveness of ACMIL in suppressing the attention value concentration and alleviating overfitting. We hope that our work can inspire future exploration into leveraging attention values for a comprehensive analysis of attention mechanisms. We also hope that our ACMIL can be applied to a broader spectrum of WSI analysis tasks. + +Acknowledgment. This study was partially supported by the National Natural Science Foundation of China (Grant No. 92270108), Zhejiang Provincial Natural Science Foundation of China (Grant No. XHD23F0201), and the Research Center for Industries of the Future (RCIF) at Westlake University. + +# References + +1. Amores, J.: Multiple instance classification: Review, taxonomy and comparative study. Artificial intelligence 201, 81-105 (2013) +2. Bejani, M.M., Ghatee, M.: A systematic review on overfitting control in shallow and deep neural networks. Artificial Intelligence Review pp. 1-48 (2021) +3. Bejnordi, B.E., Veta, M., Van Diest, P.J., Van Ginneken, B., Karssemeijer, N., Litjens, G., Van Der Laak, J.A., Hermsen, M., Manson, Q.F., Balkenhol, M., et al.: Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama 318(22), 2199-2210 (2017) +4. Bergner, B., Lippert, C., Mahendran, A.: Iterative patch selection for high-resolution image recognition. arXiv preprint arXiv:2210.13007 (2022) +5. Bontempo, G., Bolelli, F., Porrello, A., Calderara, S., Ficarra, E.: A graph-based multi-scale approach with knowledge distillation for wsi classification. TMI (2023) +6. Brancati, N., Anniciello, A.M., Pati, P., Riccio, D., Scognamiglio, G., Jaume, G., De Pietro, G., Di Bonito, M., Foncubierta, A., Botti, G., et al.: Bracs: A dataset for breast carcinoma subtyping in h&e histology images. Database 2022, baac093 (2022) +7. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine 25(8), 1301-1309 (2019) +8. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +9. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15661-15670 (2023) +0. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144-16155 (2022) +1. Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4025 (2021) +2. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23936-23945 (2023) +3. Chikontwe, P., Kim, M., Nam, S.J., Go, H., Park, S.H.: Multiple instance learning with center embeddings for histopathology classification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International + +Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23. pp. 519-528. Springer (2020) +14. Cornish, T.C., Swapp, R.E., Kaplan, K.J.: Whole-slide imaging: routine pathologic diagnosis. Advances in anatomic pathology 19(3), 152-159 (2012) +15. Dauphin, Y.N., Fan, A., Auli, M., Grangier, D.: Language modeling with gated convolutional networks. In: International conference on machine learning. pp. 933-941. PMLR (2017) +16. Dehaene, O., Camara, A., Moindrot, O., de Lavergne, A., Courtiol, P.: Self-supervision closes the gap between weak and strong supervision in histology. arXiv preprint arXiv:2012.03583 (2020) +17. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017) +18. Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T.: Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence 89(1-2), 31-71 (1997) +19. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020) +20. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018) +21. Guan, Y., Zhang, J., Tian, K., Yang, S., Dong, P., Xiang, J., Yang, W., Huang, J., Zhang, Y., Han, X.: Node-aligned graph convolutional network for whole-slide image representation and classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18813-18823 (2022) +22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +23. He, L., Long, L.R., Antani, S., Thoma, G.R.: Histology image analysis for carcinoma detection and grading. Computer methods and programs in biomedicine 107(3), 538-556 (2012) +24. Holdenried-Krafft, S., Somers, P., Montes-Majarro, I.A., Silimon, D., Tarín, C., Fend, F., Lensch, H.: Dual-query multiple instance learning for dynamic metaembedding based tumor classification. arXiv preprint arXiv:2307.07482 (2023) +25. Hou, W., Yu, L., Lin, C., Huang, H., Yu, R., Qin, J., Wang, L.: $\mathrm{H}^{\wedge}$ 2-mil: Exploring hierarchical representation with heterogeneous multiple instance learning for whole slide image analysis. In: AAAI. vol. 36, pp. 933-941 (2022) +26. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 124-140. Springer (2020) +27. Ilse, M., Tomczak, J., Welling, M.: Attention-based deep multiple instance learning. In: International conference on machine learning. pp. 2127-2136. PMLR (2018) +28. Kang, M., Song, H., Park, S., Yoo, D., Pereira, S.: Benchmarking self-supervised learning on diverse pathology datasets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3344-3354 (2023) +29. Kong, F., Henao, R.: Efficient classification of very large images with tiny objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2384-2394 (2022) +30. Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Pro + +ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318-14328 (2021) +31. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7454-7463 (2023) +32. Li, R., Yao, J., Zhu, X., Li, Y., Huang, J.: Graph cnn for survival analysis on whole slide pathological images. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 174-182. Springer (2018) +33. Li, Y., Ping, W.: Cancer metastasis detection with neural conditional random field. arXiv preprint arXiv:1806.07064 (2018) +34. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19830-19839 (2023) +35. Litjens, G., Sánchez, C.I., Timofeeva, N., Hermsen, M., Nagtegaal, I., Kovacs, I., Hulsbergen-Van De Kaa, C., Bult, P., Van Ginneken, B., Van Der Laak, J.: Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis. Scientific reports 6(1), 26286 (2016) +36. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering 5(6), 555-570 (2021) +37. Madabhushi, A.: Digital pathology image analysis: opportunities and challenges. Imaging in medicine 1(1), 7 (2009) +38. Maron, O., Lozano-Pérez, T.: A framework for multiple-instance learning. Advances in neural information processing systems 10 (1997) +39. McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018) +40. Pantanowitz, L., Valenstein, P.N., Evans, A.J., Kaplan, K.J., Pfeifer, J.D., Wilbur, D.C., Collins, L.C., Colgan, T.J.: Review of the current state of whole slide imaging in pathology. Journal of pathology informatics 2(1), 36 (2011) +41. Pinckaers, H., Van Ginneken, B., Litjens, G.: Streaming convolutional neural networks for end-to-end learning with multi-megapixel images. IEEE transactions on pattern analysis and machine intelligence 44(3), 1581-1590 (2020) +42. Qu, L., Wang, M., Song, Z., et al.: Bi-directional weakly supervised knowledge distillation for whole slide image classification. Neurips 35, 15368-15381 (2022) +43. Qu, L., Yang, Z., Duan, M., Ma, Y., Wang, S., Wang, M., Song, Z.: Boosting whole slide image classification from the perspectives of distribution, correlation and magnification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21463-21473 (2023) +44. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems 34, 2136-2147 (2021) +45. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) +46. Tang, W., Huang, S., Zhang, X., Zhou, F., Zhang, Y., Liu, B.: Multiple instance learning framework with masked hard instance mining for whole slide image classification. arXiv preprint arXiv:2307.15254 (2023) +47. Tellez, D., Litjens, G., van der Laak, J., Ciompi, F.: Neural image compression for gigapixel histopathology image analysis. IEEE transactions on pattern analysis and machine intelligence 43(2), 567-578 (2019) + +48. Tiwari, R., Shenoy, P.: Overcoming simplicity bias in deep networks using a feature sieve (2023) +49. Wang, D., Khosla, A., Gargeya, R., Irshad, H., Beck, A.H.: Deep learning for identifying metastatic breast cancer. arXiv preprint arXiv:1606.05718 (2016) +50. Wang, H., Luo, L., Wang, F., Tong, R., Chen, Y.W., Hu, H., Lin, L., Chen, H.: Iteratively coupled multiple instance learning from instance to bag classifier for whole slide image classification. arXiv preprint arXiv:2303.15749 (2023) +51. Wang, X., Xiang, J., Zhang, J., Yang, S., Yang, Z., Wang, M.H., Zhang, J., Yang, W., Huang, J., Han, X.: Scl-wc: Cross-slide contrastive learning for weakly-supervised whole-slide image classification. Advances in neural information processing systems 35, 18009-18021 (2022) +52. Wang, Y., Kartasalo, K., Weitz, P., Acs, B., Valkonen, M., Larsson, C., Ruusuvuori, P., Hartman, J., Rantalainen, M.: Predicting molecular phenotypes from histopathology images: A transcriptome-wide expression-morphology analysis in breast cancer. Cancer research 81(19), 5115-5126 (2021) +53. Xiong, C., Chen, H., Sung, J., King, I.: Diagnose like a pathologist: Transformer-enabled hierarchical attention-guided multiple instance learning for whole slide image classification. arXiv preprint arXiv:2301.08125 (2023) +54. Yang, J., Chen, H., Zhao, Y., Yang, F., Zhang, Y., He, L., Yao, J.: Remix: A general and efficient framework for multiple instance learning based whole slide image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 35-45. Springer (2022) +55. Yao, J., Zhu, X., Jonnagaddala, J., Hawkins, N., Huang, J.: Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis 65, 101789 (2020) +56. Yufei, C., Liu, Z., Liu, X., Liu, X., Wang, C., Kuo, T.W., Xue, C.J., Chan, A.B.: Bayes-mil: A new probabilistic perspective on attention-based multiple instance learning for whole slide images. In: The Eleventh International Conference on Learning Representations (2022) +57. Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18802-18812 (2022) +58. Zhang, Y., Sun, Y., Li, H., Zheng, S., Zhu, C., Yang, L.: Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 242-252. Springer (2022) +59. Zhao, Y., Yang, F., Fang, Y., Liu, H., Zhou, N., Zhang, J., Sun, J., Yang, S., Menze, B., Fan, X., et al.: Predicting lymph node metastasis using histopathological images based on multiple instance learning with deep graph convolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4837-4846 (2020) +60. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13001-13008 (2020) +61. Zhu, X., Yao, J., Zhu, F., Huang, J.: Wsisa: Making survival prediction from whole slide histopathological images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7234-7242 (2017) \ No newline at end of file diff --git a/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/images.zip b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6c52c87d4e11737837dbb3c78a77f9c25a3cd0b1 --- /dev/null +++ b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cd90134e645e752d588a4a59fadb84ba3e9179779e45d50ea77777290a22e15 +size 614651 diff --git a/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/layout.json b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..844bcb1214b484c9862c93bdc33081f1b38c0b7e --- /dev/null +++ b/2024/Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification/layout.json @@ -0,0 +1,9561 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 481, + 148 + ], + "type": "text", + "content": "Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "text", + "content": "Yunlong Zhang" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Honglin Li" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Yunxuan Sun" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Sunyi Zheng" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Chenglu Zhu" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "text", + "content": ", and Lin Yang" + }, + { + "bbox": [ + 150, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{2,3\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 168, + 201, + 444, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 201, + 444, + 236 + ], + "spans": [ + { + "bbox": [ + 168, + 201, + 444, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 201, + 444, + 236 + ], + "type": "text", + "content": " Zhejiang University \n" + }, + { + "bbox": [ + 168, + 201, + 444, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 168, + 201, + 444, + 236 + ], + "type": "text", + "content": " Research Center for Industries of the Future, Westlake University \n" + }, + { + "bbox": [ + 168, + 201, + 444, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 168, + 201, + 444, + 236 + ], + "type": "text", + "content": " School of Engineering, Westlake University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 265, + 455, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 265, + 455, + 497 + ], + "spans": [ + { + "bbox": [ + 160, + 265, + 455, + 497 + ], + "type": "text", + "content": "Abstract. In the application of Multiple Instance Learning (MIL) methods for Whole Slide Image (WSI) classification, attention mechanisms often focus on a subset of discriminative instances, which are closely linked to overfitting. To mitigate overfitting, we present Attention-Challenging MIL (ACMIL). ACMIL combines two techniques based on separate analyses for attention value concentration. Firstly, UMAP of instance features reveals various patterns among discriminative instances, with existing attention mechanisms capturing only some of them. To remedy this, we introduce Multiple Branch Attention (MBA) to capture more discriminative instances using multiple attention branches. Secondly, the examination of the cumulative value of Top-K attention scores indicates that a tiny number of instances dominate the majority of attention. In response, we present Stochastic Top-K Instance Masking (STKIM), which masks out a portion of instances with Top-K attention values and allocates their attention values to the remaining instances. The extensive experimental results on three WSI datasets with two pre-trained backbones reveal that our ACMIL outperforms state-of-the-art methods. Additionally, through heatmap visualization and UMAP visualization, this paper extensively illustrates ACMIL's effectiveness in suppressing attention value concentration and overcoming the overfitting challenge. The source code is available at https://github.com/dazhangyu123/ACMIL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 506, + 455, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 506, + 455, + 529 + ], + "spans": [ + { + "bbox": [ + 160, + 506, + 455, + 529 + ], + "type": "text", + "content": "Keywords: Computational pathology " + }, + { + "bbox": [ + 160, + 506, + 455, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 455, + 529 + ], + "type": "text", + "content": " Whole slide image " + }, + { + "bbox": [ + 160, + 506, + 455, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 455, + 529 + ], + "type": "text", + "content": " Multiple instance learning " + }, + { + "bbox": [ + 160, + 506, + 455, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 455, + 529 + ], + "type": "text", + "content": " Overfitting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 549, + 230, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 230, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 230, + 562 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Whole slide image (WSI) classification is a critical undertaking in digital pathology, aiming to extract valuable information from high-resolution scanned images for precise diagnosis [23,33,49,52], prognosis [11,32,55,61], and treatment planning [14,35,37,40,41] of diseases. In recent years, multiple instance learning (MIL) [1,18,38] has emerged as a promising approach for WSI classification, treating each WSI as a \"bag\" and its extracted small patches as \"instances\"" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 301, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 301, + 216 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 301, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 301, + 216 + ], + "type": "image", + "image_path": "413de03d7ce586999586037886ea55821489f8ebaa93793be039324604fa6390.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 301, + 290 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 301, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 301, + 290 + ], + "type": "text", + "content": "Fig. 1: The change of validation loss and entropy of attention values throughout the training of ABMIL. The results are reported on LBC with SSL pretrained features. There exists the strong negative correlation between loss and entropy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 113, + 481, + 216 + ], + "blocks": [ + { + "bbox": [ + 312, + 113, + 481, + 216 + ], + "lines": [ + { + "bbox": [ + 312, + 113, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 113, + 481, + 216 + ], + "type": "image", + "image_path": "0e4b5dfdf12709684e652abda20d7881fe07a5fabafeecf6a87f09ddb0931b52.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 224, + 482, + 290 + ], + "lines": [ + { + "bbox": [ + 311, + 224, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 311, + 224, + 482, + 290 + ], + "type": "text", + "content": "Fig.2: Comparison of AUC and entropy of attention values between ABMIL and ACMIL. One point denotes the result of a seed on LBC with SSL pretrained features. ACMIL achieves the higher AUC and entropy than ABMIL." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 339 + ], + "type": "text", + "content": "within the bag, thus enabling efficient classification of WSIs through assigning a single label to the entire slide." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 340, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 481, + 437 + ], + "type": "text", + "content": "Overfitting is a significant challenge in utilizing MIL methods for WSI classification [34, 46, 57]. Common WSI datasets exhibit intrinsic characteristics of limited data scale, ultra-high resolutions, and staining bias, which makes overfitting more likely [2]. Specifically, these datasets often consist of a relatively small number of slides, typically in the hundreds, with a typical resolution of " + }, + { + "bbox": [ + 130, + 340, + 481, + 437 + ], + "type": "inline_equation", + "content": "10,000 \\times 10,000" + }, + { + "bbox": [ + 130, + 340, + 481, + 437 + ], + "type": "text", + "content": " [36]. Moreover, pathology images are susceptible to staining bias caused by variations in tissue preparations, staining protocols, and digital scanning methods [58], leading models to learn spurious correlations [34]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "content": "In the attention mechanism, attention values/heatmaps provide insights into the model's decision-making process. Multiple existing works [36,44,46,56] alongside our own experiments (as indicated in Sec. 4.3) have pointed out the excessive concentration of attention values in current MIL methods. Specifically, MIL's attention mechanisms often concentrate on a subset of discriminative instances (i.e., instances relevant to the bag label) while disregarding the remaining ones. We investigate the correlation between attention value concentration and overfitting, utilizing the entropy of attention values and validation loss. Fig. 1 depicts a negative correlation between loss and entropy throughout the training process, illustrating that over-concentration of attention values (indicated by lower entropy) significantly compromises the model's generalization ability (indicated by higher loss values). Moreover, in the field of natural image classification, recent studies [17,26,48] have demonstrated that models solely relying on a portion of discriminative features could be susceptible to overfitting. Transitioning to WSI classification, fixating on a subset of discriminative instances similarly impedes the model's ability to generalize. These findings highlight the tight connection between attention value concentration and overfitting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Recently, numerous efforts have been made to address the overfitting challenge by enhancing representation quality [16, 24, 31, 36, 51], building spatial" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 175 + ], + "type": "text", + "content": "instance correlations [21, 30, 44] and developing data augmentation methods [13, 42, 46, 54, 57]. Additionally, some of these studies [30, 36] suggest that reducing the concentration of attention values can enhance model interpretability. However, the investigation of attention values concentration for alleviating overfitting remains under-explored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 366 + ], + "type": "text", + "content": "To mitigate overfitting, we present two analyses for attention value concentration using UMAP and Top-K value statistics. Then, we introduce Attention-Challenging MIL (ACMIL), which combines two novel techniques based on these two analyses. First, by observing UMAP of instance features, we find that there are various patterns among discriminative instances, and attention mechanisms tend to capture some of them. To solve this, we introduce Multiple Branch Attention (MBA). MBA utilizes multiple attention branches, each focusing on capturing instances with a specific pattern, thereby ensuring that more discriminative instances contribute to the final prediction. Second, by analyzing the cumulative value of Top-K attention scores, we find that a tiny number of instances (e.g., " + }, + { + "bbox": [ + 130, + 176, + 482, + 366 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 130, + 176, + 482, + 366 + ], + "type": "text", + "content": ") occupy majority attention, resulting in overlooking sophisticated discriminative instances. To suppress these instances, we propose Stochastic Top-K Instance Masking (STKIM). STKIM randomly masks out a portion of instances with Top-K attention values and assigns their attention values to the remaining instances. Combining MBA and STKIM, our ACMIL effectively alleviates the attention value concentration and suppresses overfitting (see Fig. 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 367, + 485, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 485, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 485, + 452 + ], + "type": "text", + "content": "We conduct experiments on three WSI datasets (i.e., CAMELYON16, BRACS, and our in-house LBC dataset) with two backbones (i.e., ImageNet pre-trained ResNet18 and SSL pre-trained ViT/S-16). Experimental results demonstrate the superiority of our ACMIL over existing state-of-the-art methods. We also present substantial experimental results, including heatmap visualization and UMAP visualization, to comprehensively demonstrate the effectiveness of ACMIL in suppressing attention value concentration and combatting overfitting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 468, + 237, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 237, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 237, + 480 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 491, + 360, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 360, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 360, + 503 + ], + "type": "text", + "content": "2.1 Combating Overfitting in WSI Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 544 + ], + "type": "text", + "content": "In the domain of WSI classification, combating the challenge of overfitting has received substantial attention. Next three paragraphs detail methods from three different aspects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 664 + ], + "type": "text", + "content": "Some efforts have concentrated on enhancing the quality of instance representations. Early studies (e.g., [7, 27, 44]) rely on backbones pre-trained on the ImageNet dataset. However, the substantial domain gap between natural and pathological images hinders representation quality. Recent works (e.g., [16, 24, 36, 47, 51]) address this by emphasizing Self-Supervised Learning (SSL) to learn patch-level feature representations. In addition, efforts such as the work by Chen et al. [10] leverage hierarchical SSL for high-resolution image representations. Further, studies by Li et al. [31] and Wang et al. [50] demonstrate that fine-tuning the pre-trained encoder is essential for acquiring task-specific information." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "ACMIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "Another line of research has focused on establishing spatial instance correlations. DSMIL [30], " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2\\mathrm{MIL}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " [25], and DAS-MIL [5] consider the hierarchical structure of patches and aggregate multi-scale representations in attention mechanisms. Furthermore, some studies introduce self-attention layers [43, 44, 53] and graph neural networks [9, 21, 59] to model correlations between different areas." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 258 + ], + "type": "text", + "content": "Further strategies have concentrated on data augmentation. Examples include DTFD-MIL [57], which introduces pseudo-bags for expanding bag counts and employs a double-tier MIL framework. IPS [4], Zoom-In Network [29], and Top-K MIL [13] generate bag representations by aggregating the representations of salient patches. Remix [54] and RankMix [12] introduce instance representation mixup for MIL. MHIM-MIL [46] and WENO [42] augments bags by randomly masking salient instances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "type": "text", + "content": "Although our ACMIL shares a similar spirit with some of these works, the proposed ACMIL excels by further building on detailed analysis for attention value concentration. As a result, our ACMIL exhibits stronger interpretability against existing solutions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 324, + 362, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 324, + 362, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 324, + 362, + 335 + ], + "type": "text", + "content": "2.2 Over-Concentration of Attention Values" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "type": "text", + "content": "In the realm of natural image classification, research has shown that an excessive focus on certain parts of an object can impede the overall effectiveness of model generalization [17,26,60]. To tackle this issue, various heuristic techniques have been proposed. For instance, Cutout [17, 60] is a valuable data augmentation method that randomly masks square regions of input during training. Another approach, RSC [26], involves regularization that eliminates salient features activated during training. This paper investigates the issue of attention value concentration in WSI classification tasks. We identify two specific phenomena related to attention value concentration existing in WSI classification and propose two techniques to address them respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 478, + 202, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 202, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 202, + 491 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 501, + 480, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 480, + 550 + ], + "type": "text", + "content": "Based on the ABMIL (detailed in Sec. 3.1), we present ACMIL to alleviate the overfitting problem, which is built on two components: Multiple Branch Attention (MBA) and Stochastic Top-K Instance Masking (STKIM). We describe the details of two components in Sec. 3.2 and 3.3, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 565, + 314, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 565, + 314, + 577 + ], + "spans": [ + { + "bbox": [ + 131, + 565, + 314, + 577 + ], + "type": "text", + "content": "3.1 ABMIL for WSI Classification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "text", + "content": "In the binary MIL classification problem [7], a bag of instance, " + }, + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{\\mathbf{x}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "text", + "content": ", is associated with a single bag label, " + }, + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "text", + "content": ". Each instance, " + }, + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "inline_equation", + "content": "x_n" + }, + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "text", + "content": ", is associated with a single binary label, " + }, + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "inline_equation", + "content": "y_n" + }, + { + "bbox": [ + 130, + 582, + 480, + 631 + ], + "type": "text", + "content": ", which remains unknown during training. The assumption behind the MIL can be written as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 245, + 636, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 636, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 245, + 636, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Y} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f f} \\sum_ {n = 1} ^ {N} y _ {n} = 0 \\\\ 1, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}", + "image_path": "43f4b95776a5750229aa2860d31704ce7f93003c602622fe65c938594cc68806.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 302, + 222 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 302, + 222 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 302, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 302, + 222 + ], + "type": "image", + "image_path": "e2ff3924e97d98c54d623d2df01d56a2c2c568af919a487ae618e6e2ce1d79a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 228, + 303, + 316 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 303, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 303, + 316 + ], + "type": "text", + "content": "Fig.3: Motivation of MBA. UMAP visualization [39] of tumor instance features from CAMELYON16 'test_113' case. There are various patterns/clusters among tumor instances, and relying on one single branch tends to capture a part of clusters. Three instances are selected to exhibit their texture differences." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 113, + 482, + 220 + ], + "blocks": [ + { + "bbox": [ + 312, + 113, + 482, + 220 + ], + "lines": [ + { + "bbox": [ + 312, + 113, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 312, + 113, + 482, + 220 + ], + "type": "image", + "image_path": "f0fc29677695b752a1871f4da125dfed8bafae63b059ab1e3a22efe85972c58e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 228, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 311, + 228, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 311, + 228, + 482, + 295 + ], + "type": "text", + "content": "Fig.4: Motivation of STKIM. Accumulation of Top-K attention values. Instances with Top-K attention values occupy majority attention. Results are derived from features extracted through supervised pretraining." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 479, + 396 + ], + "type": "text", + "content": "In the ABMIL [27], the multiple instance learning is modeled by a three-step process. i) Instance transformation into a low-dimensional embedding through neural networks: " + }, + { + "bbox": [ + 130, + 338, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\pmb{h}_n = f(\\pmb{x}_n)" + }, + { + "bbox": [ + 130, + 338, + 479, + 396 + ], + "type": "text", + "content": ". ii) Aggregation of all instance embeddings into the bag-level representation using an attention operator. Specifically, this operation is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 276, + 396, + 481, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 396, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 276, + 396, + 481, + 427 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} = \\sum_ {n = 1} ^ {N} a _ {n} \\boldsymbol {h} _ {n} \\tag {2}", + "image_path": "fefb291adbb9901d330d4a8c04850b3770e71ed065770d85d214d5e0c18cbcb2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 430, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 430, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 479, + 456 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 430, + 479, + 456 + ], + "type": "inline_equation", + "content": "a_{n} = \\sigma(\\pmb{h}_{n})" + }, + { + "bbox": [ + 131, + 430, + 479, + 456 + ], + "type": "text", + "content": " represents the attention values for " + }, + { + "bbox": [ + 131, + 430, + 479, + 456 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 430, + 479, + 456 + ], + "type": "text", + "content": "-th instance, " + }, + { + "bbox": [ + 131, + 430, + 479, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{h}_{n}" + }, + { + "bbox": [ + 131, + 430, + 479, + 456 + ], + "type": "text", + "content": ". In the case of ABMIL, a gated attention (GA) mechanism [15] is adopted:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 190, + 461, + 481, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 461, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 190, + 461, + 481, + 492 + ], + "type": "interline_equation", + "content": "\\sigma \\left(\\boldsymbol {h} _ {n}\\right) = \\frac {\\exp \\left\\{\\boldsymbol {w} ^ {\\mathrm {T}} \\left(\\tanh \\left(\\boldsymbol {V} _ {1} \\boldsymbol {h} _ {n}\\right) \\odot \\operatorname {s i g m} \\left(\\boldsymbol {V} _ {2} \\boldsymbol {h} _ {n}\\right)\\right) \\right\\}}{\\sum_ {j = 1} ^ {N} \\exp \\left\\{\\boldsymbol {w} ^ {\\mathrm {T}} \\left(\\tanh \\left(\\boldsymbol {V} _ {1} \\boldsymbol {h} _ {j}\\right) \\odot \\operatorname {s i g m} \\left(\\boldsymbol {V} _ {2} \\boldsymbol {h} _ {j}\\right)\\right) \\right\\}} \\tag {3}", + "image_path": "6a8d4e7da65ab9be8961982582bde0c6aecc97a52867dfaaff02c38dcf34305e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_1, \\mathbf{V}_2 \\in \\mathbb{R}^{L \\times M}" + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^{L \\times 1}" + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "text", + "content": " are parameters, " + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "text", + "content": " is an element-wise multiplication and " + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\operatorname{sigm}(\\cdot)" + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "text", + "content": " is the sigmoid non-linearity. iii) The bag prediction is generated based on the aggregated bag embedding: " + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}} = g(\\mathbf{z})" + }, + { + "bbox": [ + 131, + 498, + 482, + 537 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 552, + 293, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 293, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 293, + 563 + ], + "type": "text", + "content": "3.2 Mutiple Branch Attention" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Motivation. It is challenging to capture all discriminative instances using a single attention branch (see Fig. 3). This challenge arises due to variations in patterns among discriminative patches, stemming from differences in texture and morphology. Additionally, DNNs tend to exhibit a form of \"laziness\" where they prioritize capturing simpler patterns to minimize training loss, neglecting more intricate and challenging patterns [19, 20]. To tackle this issue, we design the MBA that captures more discriminative instances by multiple attention branches." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "ACMIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "As depicted in Fig. 5 top view, the MBA firstly captures " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " patterns and then aggregates their embeddings to make predictions. Each pattern is captured by an attention branch. To maintain both the discriminative nature of patterns and semantic diversity between them, we introduce two regularization techniques: semantic regularization and diversity regularization. Firstly, to ensure capturing discriminative patterns, the semantic regularization is accomplished by hanging a MLP layer behind each pattern embedding, equipping with the cross entropy loss function:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 219, + 481, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 219, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 206, + 219, + 481, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = - \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\boldsymbol {Y} \\log \\hat {\\boldsymbol {Y}} _ {i} + (1 - \\boldsymbol {Y}) \\log \\left(1 - \\hat {\\boldsymbol {Y}} _ {i}\\right) \\tag {4}", + "image_path": "f3e98e3470f470105ed458535c8ac07367014e7afd2ff64d3a0f8a12b70488ca.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 260, + 480, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 260, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 131, + 260, + 480, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 260, + 480, + 309 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_i = g_i(\\mathbf{z}_i)" + }, + { + "bbox": [ + 131, + 260, + 480, + 309 + ], + "type": "text", + "content": " is the prediction based on " + }, + { + "bbox": [ + 131, + 260, + 480, + 309 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 260, + 480, + 309 + ], + "type": "text", + "content": "-th pattern embedding, " + }, + { + "bbox": [ + 131, + 260, + 480, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 131, + 260, + 480, + 309 + ], + "type": "text", + "content": ". However, only equipping with cross-entropy loss may learn similar patterns and cannot dig out more discriminative information. To tackle this issue, we further introduce a diversity loss as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 226, + 317, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 317, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 226, + 317, + 481, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\frac {2}{M (M - 1)} \\sum_ {i = 1} ^ {M} \\sum_ {j = i + 1} ^ {M} \\cos \\left(\\boldsymbol {a} _ {i}, \\boldsymbol {a} _ {j}\\right) \\tag {5}", + "image_path": "40324271b549ce58c879c20aa0ef54aad3f0a98556ba49ee6782ce112d760c2b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "inline_equation", + "content": "\\pmb{a}_i" + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "content": " consists of all attention values of " + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "content": "-th pattern, " + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "inline_equation", + "content": "\\pmb{a}_i = \\{a_{i1},\\dots ,a_{iN}\\}" + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "content": ", also named heatmap as custom. The " + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "content": " function is used to measure the similarity of the heatmaps between branches. By diversifying the heatmaps, the embedding of each branch can concentrate on different patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "type": "text", + "content": "To aggregate the captured patterns to make predictions, the average of heatmaps is utilized as the heatmap of the whole bag:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 276, + 440, + 481, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 276, + 440, + 481, + 472 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\boldsymbol {a} _ {i} \\tag {6}", + "image_path": "1d267592b449ec6d9ee602660b2e3b70cfed5cd7fef81b4e22240a29c7f024a4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "text", + "content": " is the heatmap of the whole bag, with a dimension of " + }, + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "text", + "content": ". Then, the bag embedding can be obtained by aggregating the instance features using averaged heatmap " + }, + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "text", + "content": ". Moreover, since " + }, + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\sum_{n=1}^{N} \\left( \\frac{1}{M} \\sum_{i=1}^{M} a_{in} \\right) \\pmb{h}_n = \\frac{1}{M} \\sum_{i=1}^{M} \\left( \\sum_{n=1}^{N} a_{in} \\pmb{h}_n \\right)" + }, + { + "bbox": [ + 131, + 479, + 482, + 552 + ], + "type": "text", + "content": ", the bag embedding also can be formulated by applying mean pooling operator to pattern embeddings. The top view of Fig. 5 adopts the latter formulation for brevity. The loss function for the bag classifier is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 223, + 560, + 481, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 560, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 223, + 560, + 481, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b} = - \\boldsymbol {Y} \\log \\hat {\\boldsymbol {Y}} + (1 - \\boldsymbol {Y}) \\log \\left(1 - \\hat {\\boldsymbol {Y}}\\right) \\tag {7}", + "image_path": "63485ec7ffe96e26b8f2c4ae8364c6f5813ea742468032ff201bad5f3886cdc3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 586, + 480, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 480, + 611 + ], + "type": "text", + "content": "Finally, the overall loss function for the ACMIL can be written as the combination of three loss terms defined in Eq. 4, 5 and 7," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 268, + 620, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 620, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 268, + 620, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {b} + \\mathcal {L} _ {p} + \\mathcal {L} _ {d} \\tag {8}", + "image_path": "96112439e0e2023f55ccdb2fef892ffd57995ec65a49f3da52385c9977cfe229.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Discussion. It's important to highlight that when the parameter " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " is set to 1 in MBA, it essentially mirrors the feature aggregation process of ABMIL, allowing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 471, + 199 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 471, + 199 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 471, + 199 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 471, + 199 + ], + "type": "image", + "image_path": "7f6a0513029876194d765ff31c29dd3a7fee35bc1f41bc92eaacb935f4b37ee3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 201, + 471, + 281 + ], + "blocks": [ + { + "bbox": [ + 143, + 201, + 471, + 281 + ], + "lines": [ + { + "bbox": [ + 143, + 201, + 471, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 201, + 471, + 281 + ], + "type": "image", + "image_path": "e88986ad247e602ec3171684de6cf1c8ab668f665c3070ad37334e820913666b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "text", + "content": "Fig. 5: Overview of the proposed MBA (top view) and STKIM (bottom view). In the MBA, " + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "text", + "content": " discriminative patterns are extracted from patch features using the attention operator regularized by semantic and diversity regularization terms. Then, the mean operator is applied to these " + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "text", + "content": " pattern features to produce the bag feature, which is utilized for bag-level prediction. In the STKIM, instances with Top-K attention values are randomly masked with a probability " + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 384, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 528 + ], + "type": "text", + "content": "for the discernment of a single pattern. In this sense, MBA serves as an extension of ABMIL, specifically designed to capture a more diverse set of patterns. We further discuss the connection between MBA and Multiple-Head Attention (MHA). HIPT [10] has unveiled that distinct heads in MHA can effectively capture different visual concepts, akin to the role played by our MBA. However, these two techniques can be easily distinguished by: 1) MBA has diversity regularization, ensuring that different branches can learn different concepts. This is absent in MHA, resulting in different heads learning the same concept [10]. We demonstrate the critical role of " + }, + { + "bbox": [ + 130, + 384, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 130, + 384, + 482, + 528 + ], + "type": "text", + "content": " for performance in Tab. 3b. 2) MHA is a type of attention formulation, while MBA operates independently of the attention formulation, accommodating MHA within its framework. Appendix Sec. 4.1 discusses the results of combining MHA and ACMIL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 547, + 340, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 340, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 340, + 560 + ], + "type": "text", + "content": "3.3 Stochastic Top-K Instance Masking" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Motivation. A tiny number of instances will occupy the majority of attention in ABMIL while ignoring sophisticated discriminative instances. As depicted in Fig. 4, the sum of top-10 attention values is larger than 0.85 on all three datasets. However, the WSI typically involves more than 10 discriminative instances. For instance, in the CAMELYON16 dataset, 129 out of 155 tumor slides contain 10 to 20,000 cancerous instances. In essence, numerous discriminative instances are overlooked. To deal with this issue, the proposed STKIM aims to suppress the salient instances and assign more attention to the remaining instances." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "ACMIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "As depicted in Fig. 5 bottom view, STKIM introduces a masking operation into the attention mechanism, before feature aggregation and after attention values generation. The primary objective is to suppress Top-K salient instances. A straightforward solution to achieve this is to mask out all of the Top-K salient instances. However, this method poses certain challenges. It can result in the loss of information associated with key instances, which are crucial for discrimination. Furthermore, it might lead to a statistical mismatch between the feature representations before and after discarding these key instances. To address these issues, we draw inspiration from dropout [45] and cutout [17,60] commonly used in computer vision. Our proposed solution employs stochastic masking for instance features with Top-K attention values. Specifically, we begin by sorting all attention values from highest to lowest. Subsequently, we randomly set the attention values of the Top-K instances to 0, with a probability of " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": ". This process can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 289, + 481, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 289, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 182, + 289, + 481, + 321 + ], + "type": "interline_equation", + "content": "a _ {n} = \\left\\{ \\begin{array}{l l} 0, & \\text {w i t h p r o b a b i l i t y} p \\text {a n d w i t h i n T o p - K v a l u e s} \\\\ a _ {n}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}", + "image_path": "6e9ab7cab5a25deb51ff89c57aaf31693457f0a23465a544526788a977a5a0bd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "content": " are two hyperparameters that control the intensity of masking. Following Eq. 9, we will assign the attention values of masked instances to the remaining instances by " + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "inline_equation", + "content": "a_{n} \\to \\frac{1}{\\sum_{n=1}^{N} a_{n}} a_{n}" + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "content": ". Notably, drawing inspiration from dropout and cutout, we remove STKIM at the inference time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 547 + ], + "type": "text", + "content": "Discussion. While STKIM, MHIM-MIL [46], and WENO [42] all adopt the technique that masks salient instances, there are notable technical distinctions between them. Firstly, our STKIM has the different masking strategy compared with WENO and MHIM-MIL. STKIM only masks a minority of instances (i.e., " + }, + { + "bbox": [ + 130, + 378, + 482, + 547 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 130, + 378, + 482, + 547 + ], + "type": "text", + "content": ") with a probability of " + }, + { + "bbox": [ + 130, + 378, + 482, + 547 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 378, + 482, + 547 + ], + "type": "text", + "content": ". As a comparison, the other two methods mask out a larger number of instances. WENO masks out 95 instances. MHIM-MIL masks " + }, + { + "bbox": [ + 130, + 378, + 482, + 547 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 547 + ], + "type": "text", + "content": " instances. In our framework, our scheme performs best in three strategies (see Appendix Tab. 2). Secondly, both MHIM-MIL and WENO necessitate a well-trained model for masking out salient instances, utilizing the remaining instances for model training. They both employ a teacher-student framework, wherein the teacher model needs to be pre-trained beforehand (the warm-up process in WENO and the pre-training stage in MHIM-MIL). In contrast, STKIM requires neither a teacher-student framework nor a pre-training process, thus highlighting simplicity and efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 563, + 230, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 230, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 230, + 576 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 586, + 268, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 268, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 268, + 598 + ], + "type": "text", + "content": "4.1 Experimental Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. The performance of ACMIL is evaluated on two public WSI datasets, i.e., CAMELYON16 [3] and BRACS [6], and one private benchmark, LBC. CAMELYON16 dataset consists of 400 WSIs in total, including 270 for training and 130 for testing. Following [30, 57], we further randomly split the training and validation sets from the official training set with a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 481, + 358 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "type": "text", + "content": "Table 1: The performance of different MIL approaches across three datasets, two pre-trained methods, and two evaluation metrics. The most superior performance is highlighted in bold, while the second-best performance is indicated by underlining." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 481, + 358 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 481, + 358 + ], + "type": "table", + "html": "
Performance MethodCAMELYON-16BRACSLBC
F1-scoreAUCF1-scoreAUCF1-scoreAUC
ResNet-18Max-pooling0.582±0.1700.620±0.1550.489±0.0470.738±0.0140.476±0.0330.775±0.010
Mean-pooling0.592±0.0260.597±0.0330.484±0.0290.685±0.0110.511±0.0220.797±0.011
Clam-SB0.742±0.0240.763±0.0490.521±0.0460.750±0.0390.514±0.0240.805±0.017
TransMIL0.643±0.0880.706±0.0760.444±0.0400.732±0.0430.385±0.0130.693±0.027
DSMIL0.736±0.0280.773±0.0340.511±0.0520.751±0.0280.458±0.0290.766±0.023
DTFD-MIL0.758±0.0510.815±0.0630.469±0.0160.717±0.0320.473±0.0210.776±0.021
IBMIL0.777±0.0090.799±0.0500.510±0.0430.726±0.0340.489±0.0170.791±0.021
MHIM-MIL0.752±0.0340.772±0.0260.511±0.0220.774±0.0210.543±0.0370.816±0.009
ABMIL0.757±0.0200.790±0.0270.523±0.0280.723±0.0350.465±0.0400.798±0.013
ACMIL(ours)0.798±0.0290.841±0.0300.552±0.0480.754±0.0080.546±0.0280.821±0.015
VIT-S/16Max-pooling0.903±0.0540.956±0.0290.596±0.0290.823±0.0330.590±0.0430.829±0.023
Mean-pooling0.577±0.0570.569±0.0810.522±0.0380.739±0.0070.559±0.0240.827±0.012
Clam-SB0.925±0.0350.969±0.0240.631±0.0340.863±0.0050.617±0.0220.865±0.018
TransMIL0.922±0.0190.943±0.0090.631±0.0300.841±0.0060.539±0.0280.805±0.010
DSMIL0.943±0.0070.966±0.0090.577±0.0280.816±0.0280.562±0.0280.820±0.033
DTFD-MIL0.948±0.0070.980±0.0110.612±0.0800.870±0.0220.612±0.0340.842±0.010
IBMIL0.912±0.0340.954±0.0220.645±0.0410.871±0.0140.604±0.0320.834±0.014
MHIM-MIL0.932±0.0240.970±0.0370.625±0.0600.865±0.0170.658±0.0410.872±0.022
ABMIL0.914±0.0310.945±0.0270.680±0.0510.866±0.0290.595±0.0360.831±0.022
ACMIL(ours)0.954±0.0120.974±0.0120.722±0.0300.888±0.0100.662±0.0430.901±0.011
", + "image_path": "a392c591642781ff06f30f9fb6a931955b4b48671327d78568f4bb2f1561aa41.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 396, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 481, + 529 + ], + "type": "text", + "content": "ratio of 9:1. We do not resplit BRACS dataset as it has been officially split to 395 of training set, 65 of validation set, and 87 of test set. We follow the challenge for a 3-class WSI classification: benign tumor, atypical tumor, and malignant. The liquid-based cytology (LBC) dataset collected 1,989 WSIs and included 4 classes, i.e., Negative, ASC-US, LSIL, and ASC-H/HSIL. We randomly split the whole dataset into training, validation, and test sets with the ratio of 6:2:2. Following [31], macro-AUC and macro-F1 scores are reported since all three datasets are class imbalanced. Each of the main experiments is performed five times with random parameter initializations, and the average classification performance and standard deviation are reported. Besides, following [36,57], the test performance is reported in epochs with the best validation performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 536, + 481, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 481, + 644 + ], + "type": "text", + "content": "Baselines. We systematically assess the efficacy of our ACMIL approach by benchmarking it against conventional MIL pooling strategies, Max-pooling and Mean-pooling, as well as contemporary attention-based techniques such as AB-MIL [27], DSMIL [30], TransMIL [44], CLAM-SB [36], DTFD-MIL [57], MHIM-MIL [46], and IBMIL [34]. In pursuit of a comprehensive comparison across diverse aggregation operators, we utilize two distinct sets of features derived from ResNet-18 pre-trained on the ImageNet dataset [22] and ViT-S/16 pretrained using DINO [8, 28]. The results of all other methods are reproduced using the official code they provide under the same settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 652, + 465, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 465, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 465, + 665 + ], + "type": "text", + "content": "Implementation Details. Implementation Details are in Appendix Sec. 3." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "ACMIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 473, + 286 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 473, + 286 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 473, + 286 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 473, + 286 + ], + "type": "image", + "image_path": "8c9d840be914cf556d82760eacc6c5040b0eeb601e2c0c4ebe1752ed51805fa7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 294, + 482, + 371 + ], + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 371 + ], + "type": "text", + "content": "Fig. 6: Heatmap visualization of WSI examples produced by ABMIL [27] (baseline) and our ACMIL. The left part shows three tumor WSIs that come from CAMELYON16 dataset, and their tumor regions are delineated by red lines. ACMIL generates attention values that cover a more extensive portion of the tumor region compared to ABMIL. The right part shows three normal WSIs that come from CAMELYON16 dataset. ABMIL primarily focuses on a part of tissue such as adipose, while ACMIL extends its attention to the more normal tissues." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 398, + 294, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 294, + 410 + ], + "type": "text", + "content": "4.2 WSI Classification Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 425, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 472 + ], + "type": "text", + "content": "Tab. 1 provides a thorough comparison of performance between ACMIL and existing MIL methods. This evaluation spans three diverse datasets, involves two different choices for pretraining methods, and employs two crucial evaluation metrics, resulting in a comprehensive assessment with a total of 12 terms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "content": "Considering the overall performance, ACMIL consistently outshines existing methods. It secures the top position in 10 out of the 12 metrics and holds the second position in the remaining 2 metrics. Specifically, for the CAMELYON16, ACMIL achieves outstanding results using ResNet-18 pre-trained on ImageNet embeddings, surpassing the runner-up by " + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "content": " in terms of F1-score and AUC, respectively. On the other hand, with ViT-S/16 SSL pretrained embeddings, existing attention-based MIL methods exhibit remarkable performance, boasting F1-scores and AUC values exceeding 0.9. Notably, ACMIL achieves comparable performance with the former best-performing method, DTFD-MIL, in this setup. For the BRACS, ACMIL demonstrates a substantial lead when utilizing ViT-S/16 SSL pre-trained embeddings, surpassing the second-best performance by margins of " + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "content": " in F1-score and AUC, respectively. Moreover, when employing ResNet-18 pre-trained on ImageNet embeddings, ACMIL achieves comparable performance with the previously top-performing method, MHIM-MIL. For the LBC, ACMIL stands out significantly among the other methods across all four metrics." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 228, + 136, + 382, + 162 + ], + "blocks": [ + { + "bbox": [ + 179, + 114, + 433, + 125 + ], + "lines": [ + { + "bbox": [ + 179, + 114, + 433, + 125 + ], + "spans": [ + { + "bbox": [ + 179, + 114, + 433, + 125 + ], + "type": "text", + "content": "Table 2: Comparison of FROC between ABMIL and ACMIL" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 228, + 136, + 382, + 162 + ], + "lines": [ + { + "bbox": [ + 228, + 136, + 382, + 162 + ], + "spans": [ + { + "bbox": [ + 228, + 136, + 382, + 162 + ], + "type": "table", + "html": "
ABMILACMIL
FROC0.39870.4233
", + "image_path": "8c6aa1b0e9856419966755d09cfa5cc50d870ce6fdf6d15f1b1d514d7bab3617.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 262, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 262, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 262, + 196 + ], + "type": "text", + "content": "4.3 Localization Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 207, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 479, + 266 + ], + "type": "text", + "content": "Heatmap visualization. Fig. 6 presents heatmap visualizations illustrating examples of our approach's performance in comparison to the baseline method, ABMIL [27]. Three tumor slides (left part) and three normal slides (right part) are selected to showcase the heatmap differences. Due to the space limitation, we present more visualization in Appendix Sec. 4.4 for further insights." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 268, + 480, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 480, + 387 + ], + "type": "text", + "content": "For the tumor slides, ABMIL tends to concentrate its attention on only a fraction of the tumor regions, potentially overlooking other significant areas. In contrast, ACMIL allocates attention across a wider spectrum of tumor regions, resulting in better alignment with expert annotations. For the normal slides, ABMIL predominantly focuses on specific tissue types, such as adipose tissue. This will lead to misinterpretation that only the adipose tissue is the normal tissue and other normal regions are uncorrelated to the WSI label. On the other hand, ACMIL effectively distributes attention values to encompass all normal regions, ensuring all regions are correlated for the WSI label. This approach closely mimics human intuition and satisfies the definition of the MIL formulation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 443 + ], + "type": "text", + "content": "FROC results. We employ the FROC metric suggested by CAMELYON16 challenge to evaluate the localization of tumor region quantitatively. As shown in Table 2, the proposed ACMIL achieves higher FROC than ABMIL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 463, + 237, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 237, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 237, + 474 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 484, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 520 + ], + "type": "text", + "content": "Fig. 7 illustrates the AUC scores of ACMIL across three datasets when utilizing a ViT/B-16 feature extractor and varying hyperparameter settings. Several key observations emerge from these experiments:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 556 + ], + "type": "text", + "content": "Effect of branches number " + }, + { + "bbox": [ + 130, + 521, + 480, + 556 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 521, + 480, + 556 + ], + "type": "text", + "content": " in MBA. As shown in the first column, we find that the choice of " + }, + { + "bbox": [ + 130, + 521, + 480, + 556 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 521, + 480, + 556 + ], + "type": "text", + "content": " affects performance significantly. Combining three datasets, setting " + }, + { + "bbox": [ + 130, + 521, + 480, + 556 + ], + "type": "inline_equation", + "content": "M = 5" + }, + { + "bbox": [ + 130, + 521, + 480, + 556 + ], + "type": "text", + "content": " achieves the best performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": "Effect of masking probability " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": " in STKIM. As shown in the second column, we find that the choice of " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": " also affects performance significantly. Notably, setting " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "p = 1.0" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": " (masking all of Top-K instances) leads to performance deterioration across all three datasets. For LBC and CAMELYON, a " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "p = 1.0" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": " setting even results in performance lower than the blue dotted lines. Otherwise, " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "p = 0.6" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": " achieves the best performance on the BRACS dataset, whereas " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "p = 0.8" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": " achieves the best performance on the other two datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Effect of number of masking instances " + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": " in STKIM. The third column shows that hyperparameter " + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": " exhibits minimal sensitivity, where different " + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "ACMIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 113, + 257, + 195 + ], + "blocks": [ + { + "bbox": [ + 149, + 113, + 257, + 195 + ], + "lines": [ + { + "bbox": [ + 149, + 113, + 257, + 195 + ], + "spans": [ + { + "bbox": [ + 149, + 113, + 257, + 195 + ], + "type": "image", + "image_path": "54ac08292af7c9054c16ee6cc9477d59141448dece695701f16a2e880da4cbb0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 113, + 360, + 194 + ], + "blocks": [ + { + "bbox": [ + 258, + 113, + 360, + 194 + ], + "lines": [ + { + "bbox": [ + 258, + 113, + 360, + 194 + ], + "spans": [ + { + "bbox": [ + 258, + 113, + 360, + 194 + ], + "type": "image", + "image_path": "339448ff6e229fbc134b0cf212d4daaa9f72d93452a73e4650027389658ef325.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 361, + 113, + 465, + 194 + ], + "blocks": [ + { + "bbox": [ + 361, + 113, + 465, + 194 + ], + "lines": [ + { + "bbox": [ + 361, + 113, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 361, + 113, + 465, + 194 + ], + "type": "image", + "image_path": "f05a0557acd9d6e1d5764c934854d43c230cf030f389ff74de7a04bac83684f2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 195, + 258, + 281 + ], + "blocks": [ + { + "bbox": [ + 149, + 195, + 258, + 281 + ], + "lines": [ + { + "bbox": [ + 149, + 195, + 258, + 281 + ], + "spans": [ + { + "bbox": [ + 149, + 195, + 258, + 281 + ], + "type": "image", + "image_path": "9a0a32e90d00e30db54a3e8b19ab05655b69a83a355fafef165ef835a96385e4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 258, + 195, + 360, + 281 + ], + "blocks": [ + { + "bbox": [ + 258, + 195, + 360, + 281 + ], + "lines": [ + { + "bbox": [ + 258, + 195, + 360, + 281 + ], + "spans": [ + { + "bbox": [ + 258, + 195, + 360, + 281 + ], + "type": "image", + "image_path": "2749d22ecf04fb1d930f7689f98dd0811134f123ce8e5bcd4d6cf349c1c7d15c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 360, + 195, + 465, + 281 + ], + "blocks": [ + { + "bbox": [ + 360, + 195, + 465, + 281 + ], + "lines": [ + { + "bbox": [ + 360, + 195, + 465, + 281 + ], + "spans": [ + { + "bbox": [ + 360, + 195, + 465, + 281 + ], + "type": "image", + "image_path": "88fda8165a51d6c95aae69e0d0cc1022f1994e57fd2aadbe431454f333f0765a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 149, + 281, + 258, + 365 + ], + "blocks": [ + { + "bbox": [ + 149, + 281, + 258, + 365 + ], + "lines": [ + { + "bbox": [ + 149, + 281, + 258, + 365 + ], + "spans": [ + { + "bbox": [ + 149, + 281, + 258, + 365 + ], + "type": "image", + "image_path": "b0ecfc5ff427dd98ffdf2f922cd0375f2be30632fe5f460f4c85dcfb531b78e6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 373, + 482, + 429 + ], + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 429 + ], + "type": "text", + "content": "Fig. 7: Ablation study on features extracted through the SSL pre-training. The effect of three hyperparameters, " + }, + { + "bbox": [ + 130, + 373, + 482, + 429 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 373, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 482, + 429 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 373, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 482, + 429 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 373, + 482, + 429 + ], + "type": "text", + "content": ", is investigated. Note that the orange dotted line denotes the performance of baseline, ABMIL, and the blue dotted line denotes the performance of ACMIL w/o MBA or STKIM. Five conclusions derived from the figure can be found in Sec. 4.4." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 258, + 281, + 360, + 365 + ], + "blocks": [ + { + "bbox": [ + 258, + 281, + 360, + 365 + ], + "lines": [ + { + "bbox": [ + 258, + 281, + 360, + 365 + ], + "spans": [ + { + "bbox": [ + 258, + 281, + 360, + 365 + ], + "type": "image", + "image_path": "459fba53739ea91e79c4f8eb3a0ea28bd0a69de4b463bc6fff10256e66f19456.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 360, + 281, + 465, + 365 + ], + "blocks": [ + { + "bbox": [ + 360, + 281, + 465, + 365 + ], + "lines": [ + { + "bbox": [ + 360, + 281, + 465, + 365 + ], + "spans": [ + { + "bbox": [ + 360, + 281, + 465, + 365 + ], + "type": "image", + "image_path": "f9cadb9552a5128e78f7bce3b6744575d58826ec239686f35f496355c9c09552.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "content": "values result in a performance difference of less than " + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "content": " AUC. In practice, setting " + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "content": " to 10 is generally sufficient for achieving near-optimal performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 475, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 558 + ], + "type": "text", + "content": "Implementing either MBA or STKIM individually leads to significant performance improvements. The blue dotted lines represent ACMIL's AUC performance without MBA or STKIM, outperforming the orange dotted lines (ABMIL's performance) across all subfigures. Particularly noteworthy is the observation that MBA achieves better improvement than STKIM on all three datasets, with blue dotted lines in the last two columns surpassing those in the first column, especially on the CAMELYON and LBC datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 559, + 482, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 607 + ], + "type": "text", + "content": "Combining MBA with STKIM yields greater performance improvements compared to using either MBA or STKIM alone. The green dots represent ACMIL's performance under different hyperparameter combinations, with 39 out of 45 green dots exceeding blue dotted lines." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 623, + 246, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 246, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 246, + 635 + ], + "type": "text", + "content": "4.5 Further Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "MBA can capture diverse patterns. We employ UMAP [39] to visualize instance features within the tumor region of the CAMELYON16 'test_090' case." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 306, + 220 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 306, + 220 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 306, + 220 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 306, + 220 + ], + "type": "image", + "image_path": "902a8dcf03da0490d6facae9731650205a18dcdc21083092fa767d8e65f7051c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 223, + 242, + 232 + ], + "lines": [ + { + "bbox": [ + 200, + 223, + 242, + 232 + ], + "spans": [ + { + "bbox": [ + 200, + 223, + 242, + 232 + ], + "type": "text", + "content": "(a) ABMIL" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 113, + 480, + 221 + ], + "blocks": [ + { + "bbox": [ + 307, + 113, + 480, + 221 + ], + "lines": [ + { + "bbox": [ + 307, + 113, + 480, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 480, + 221 + ], + "type": "image", + "image_path": "556a2bb6e937e5576b2e69bfe409993d68a87c9f3aacdcaec0360ef0ba8c1969.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 223, + 434, + 232 + ], + "lines": [ + { + "bbox": [ + 352, + 223, + 434, + 232 + ], + "spans": [ + { + "bbox": [ + 352, + 223, + 434, + 232 + ], + "type": "text", + "content": "(b) ABMIL with MBA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 241, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 480, + 308 + ], + "type": "text", + "content": "Fig. 8: UMAP visualization of tumor instance features from the CAMELYON16 'test_090' case. The tumor instances display distinct patterns, posing a challenge for a single branch to capture all of them. As a result, ABMIL overlooks the right pattern/cluster. In contrast, multiple branches in MBA capture different patterns separately, and their combination enables the activation of more tumor instances. An instance is considered active when its attention value surpasses " + }, + { + "bbox": [ + 130, + 241, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}" + }, + { + "bbox": [ + 130, + 241, + 480, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 480, + 409 + ], + "type": "text", + "content": "In Fig. 8a, it's evident that the tumor instances exhibit two primary patterns. However, ABMIL primarily activates the left pattern (colored orange) and neglects the right one. On the other hand, as demonstrated in Fig. 8b, MBA's various branches (branch1, branch2, branch3, and branch5) collectively capture the substructures of the left pattern, while branch4 specifically captures the right pattern. Combining all branches can capture more comprehensive patterns." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 315, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 315, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 315, + 554 + ], + "type": "text", + "content": "STKIM can suppress the concentration of Top-K attention values. Fig. 9 illustrates a comparison of the cumulative sum of Top-K attention values with and without STKIM. The plot clearly shows that the use of STKIM helps mitigate the scenario where Top-K attention values excessively dominate in the attention mechanism. This effect is particularly pronounced for CAMELYON16 dataset, where the cumulative sum of the top-10 values decreases from 0.87 to 0.6." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 315, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 315, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 315, + 628 + ], + "type": "text", + "content": "Do we need STKIM at the test phase? The answer is No. In Tab. 3a, we present the outcomes of ACMIL with and without STKIM during the test phase, along with the performance differences between these settings. Across 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "out of 12 evaluation metrics, the version of ACMIL without STKIM during testing outperforms the version with STKIM slightly. This suggests that STKIM is not necessary during the test phase." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 435, + 476, + 529 + ], + "blocks": [ + { + "bbox": [ + 328, + 435, + 476, + 529 + ], + "lines": [ + { + "bbox": [ + 328, + 435, + 476, + 529 + ], + "spans": [ + { + "bbox": [ + 328, + 435, + 476, + 529 + ], + "type": "image", + "image_path": "53ee61c7fe7397ee4bbddb99184223db7989e93c2f2592b96e2caa11880ce52d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 537, + 482, + 604 + ], + "lines": [ + { + "bbox": [ + 321, + 537, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 321, + 537, + 482, + 604 + ], + "type": "text", + "content": "Fig.9: Comparison of the accumulative sum of Top-K attention values with and without STKIM. The use of STKIM helps alleviate the issue of excessive concentration of attention values within the Top-K range." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "ACMIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": "Table 3: (a): Performance comparison between ACMIL with (w.) and without (w/o.) T-STKIM. T-STKIM means using STKIM at the test phase. The Gap column reports the performance difference between with and without T-STKIM. Using STKIM at the test phase slightly reduces performance. (b): Performance comparison between ACMIL with (w.) and without (w/o.) " + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": ". The Gap column reports the performance difference between without and with " + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": ". ACMIL without " + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": " drastically reduces its performance." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 217, + 310, + 348 + ], + "blocks": [ + { + "bbox": [ + 147, + 194, + 280, + 205 + ], + "lines": [ + { + "bbox": [ + 147, + 194, + 280, + 205 + ], + "spans": [ + { + "bbox": [ + 147, + 194, + 280, + 205 + ], + "type": "text", + "content": "(a) w. T-STKIM v.s. w/o. T-STKIM" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 217, + 310, + 348 + ], + "lines": [ + { + "bbox": [ + 132, + 217, + 310, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 310, + 348 + ], + "type": "table", + "html": "
ViT-S/16 SSL pre-trained
DatasetMetricw. T-STKIMw/o. T-STKIMGap(%)
CamelyonF1-score0.927±0.0570.954±0.012+2.7
AUC0.967±0.0170.974±0.012+0.7
BRACSF1-score0.697±0.0330.722±0.030+2.5
AUC0.875±0.0120.888±0.010+1.3
LBCF1-score0.637±0.0340.662±0.043+2.5
AUC0.878±0.0120.901±0.011+2.3
ResNet-18Imagenet pre-trained
DatasetMetricw. T-STKIMw/o. T-STKIMGap(%)
CamelyonF1-score0.780±0.0260.798±0.029+1.8
AUC0.837±0.0280.841±0.030+0.4
BRACSF1-score0.566±0.0540.552±0.048-1.4
AUC0.750±0.0210.754±0.008+0.4
LBCF1-score0.535±0.0270.546±0.028+1.1
AUC0.808±0.0190.821±0.015+1.3
", + "image_path": "e6fd9db65242f6f29916ae9aa9f2a43937247518354e66fa777a091134a93cd4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 216, + 484, + 348 + ], + "blocks": [ + { + "bbox": [ + 356, + 194, + 440, + 205 + ], + "lines": [ + { + "bbox": [ + 356, + 194, + 440, + 205 + ], + "spans": [ + { + "bbox": [ + 356, + 194, + 440, + 205 + ], + "type": "text", + "content": "(b) w. " + }, + { + "bbox": [ + 356, + 194, + 440, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 356, + 194, + 440, + 205 + ], + "type": "text", + "content": " v.s. w/o. " + }, + { + "bbox": [ + 356, + 194, + 440, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 216, + 484, + 348 + ], + "lines": [ + { + "bbox": [ + 317, + 216, + 484, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 216, + 484, + 348 + ], + "type": "table", + "html": "
ViT-S/16 SSL pre-trained
DatasetMetricw/o. Ldw. LdGap(%)
CamelyonF1-score0.901±0.0370.954±0.012+5.3
AUC0.943±0.0270.974±0.012+3.1
BRACSF1-score0.642±0.0460.722±0.030+8.0
AUC0.859±0.0200.888±0.010+2.9
LBCF1-score0.603±0.0230.662±0.043+5.9
AUC0.837±0.0090.901±0.011+6.4
ResNet-18Imagenet pre-trained
DatasetMetricw/o. Ldw. LdGap(%)
CamelyonF1-score0.747±0.0220.798±0.029+5.1
AUC0.796±0.0320.841±0.030+5.5
BRACSF1-score0.500±0.0310.552±0.048+5.2
AUC0.760±0.0260.754±0.008-0.6
LBCF1-score0.532±0.0190.546±0.028+1.4
AUC0.809±0.0180.821±0.015+1.2
", + "image_path": "daaf7724e7bc08ecb29816184e395ddc0b5e1e8753dd703bec25bab67d1b2a01.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 371, + 482, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 442 + ], + "type": "text", + "content": "Do we need diversity loss in MBA? The answer is Yes. In Tab. 3b, we present the outcomes of ACMIL with and without " + }, + { + "bbox": [ + 130, + 371, + 482, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 130, + 371, + 482, + 442 + ], + "type": "text", + "content": ", along with the performance differences between these settings. Notably, the last column clearly indicates a significant performance drop for ACMIL without " + }, + { + "bbox": [ + 130, + 371, + 482, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 130, + 371, + 482, + 442 + ], + "type": "text", + "content": ". This emphasizes the crucial role of " + }, + { + "bbox": [ + 130, + 371, + 482, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 130, + 371, + 482, + 442 + ], + "type": "text", + "content": " in encouraging different branches to acquire distinctive discriminative knowledge within the MBA technique." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 443, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 491 + ], + "type": "text", + "content": "Additional Analyses: Appendix presents further analyses that offer deeper insights into ACMIL's behavior. Specifically, Sec. 4.2 provides a detailed analysis of bag features, while Sec. 4.3 provides a comprehensive examination of evaluation metrics throughout the training process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Due to the intrinsic properties of WSI, MIL methods have often led to overfitting, limiting their applications. This paper reveals that the overly concentrated attention values in the heatmap are closely related to overfitting. To address this, we propose ACMIL, which is underpinned by two novel techniques: MBA and STKIM. Our experimental results on three datasets demonstrate that ACMIL significantly surpasses SOTA methods. Moreover, this paper provides comprehensive experiments confirming the effectiveness of ACMIL in suppressing the attention value concentration and alleviating overfitting. We hope that our work can inspire future exploration into leveraging attention values for a comprehensive analysis of attention mechanisms. We also hope that our ACMIL can be applied to a broader spectrum of WSI analysis tasks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "Acknowledgment. This study was partially supported by the National Natural Science Foundation of China (Grant No. 92270108), Zhejiang Provincial Natural Science Foundation of China (Grant No. XHD23F0201), and the Research Center for Industries of the Future (RCIF) at Westlake University." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 206, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "type": "text", + "content": "1. Amores, J.: Multiple instance classification: Review, taxonomy and comparative study. Artificial intelligence 201, 81-105 (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 250 + ], + "type": "text", + "content": "2. Bejani, M.M., Ghatee, M.: A systematic review on overfitting control in shallow and deep neural networks. Artificial Intelligence Review pp. 1-48 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 293 + ], + "type": "text", + "content": "3. Bejnordi, B.E., Veta, M., Van Diest, P.J., Van Ginneken, B., Karssemeijer, N., Litjens, G., Van Der Laak, J.A., Hermsen, M., Manson, Q.F., Balkenhol, M., et al.: Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama 318(22), 2199-2210 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 294, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 294, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 294, + 480, + 316 + ], + "type": "text", + "content": "4. Bergner, B., Lippert, C., Mahendran, A.: Iterative patch selection for high-resolution image recognition. arXiv preprint arXiv:2210.13007 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 480, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 337 + ], + "type": "text", + "content": "5. Bontempo, G., Bolelli, F., Porrello, A., Calderara, S., Ficarra, E.: A graph-based multi-scale approach with knowledge distillation for wsi classification. TMI (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 338, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 480, + 380 + ], + "type": "text", + "content": "6. Brancati, N., Anniciello, A.M., Pati, P., Riccio, D., Scognamiglio, G., Jaume, G., De Pietro, G., Di Bonito, M., Foncubierta, A., Botti, G., et al.: Bracs: A dataset for breast carcinoma subtyping in h&e histology images. Database 2022, baac093 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 381, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 425 + ], + "type": "text", + "content": "7. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine 25(8), 1301-1309 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "type": "text", + "content": "8. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 458, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 501 + ], + "type": "text", + "content": "9. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15661-15670 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 545 + ], + "type": "text", + "content": "0. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144-16155 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 545, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 589 + ], + "type": "text", + "content": "1. Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4025 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "2. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23936-23945 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "3. Chikontwe, P., Kim, M., Nam, S.J., Go, H., Park, S.H.: Multiple instance learning with center embeddings for histopathology classification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "ACMIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 149, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 480, + 138 + ], + "type": "text", + "content": "Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23. pp. 519-528. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "text", + "content": "14. Cornish, T.C., Swapp, R.E., Kaplan, K.J.: Whole-slide imaging: routine pathologic diagnosis. Advances in anatomic pathology 19(3), 152-159 (2012)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "text", + "content": "15. Dauphin, Y.N., Fan, A., Auli, M., Grangier, D.: Language modeling with gated convolutional networks. In: International conference on machine learning. pp. 933-941. PMLR (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "16. Dehaene, O., Camara, A., Moindrot, O., de Lavergne, A., Courtiol, P.: Self-supervision closes the gap between weak and strong supervision in histology. arXiv preprint arXiv:2012.03583 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 480, + 248 + ], + "type": "text", + "content": "17. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 270 + ], + "type": "text", + "content": "18. Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T.: Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence 89(1-2), 31-71 (1997)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 480, + 303 + ], + "type": "text", + "content": "19. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "20. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "text", + "content": "21. Guan, Y., Zhang, J., Tian, K., Yang, S., Dong, P., Xiang, J., Yang, W., Huang, J., Zhang, Y., Han, X.: Node-aligned graph convolutional network for whole-slide image representation and classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18813-18823 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "23. He, L., Long, L.R., Antani, S., Thoma, G.R.: Histology image analysis for carcinoma detection and grading. Computer methods and programs in biomedicine 107(3), 538-556 (2012)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "24. Holdenried-Krafft, S., Somers, P., Montes-Majarro, I.A., Silimon, D., Tarín, C., Fend, F., Lensch, H.: Dual-query multiple instance learning for dynamic metaembedding based tumor classification. arXiv preprint arXiv:2307.07482 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "25. Hou, W., Yu, L., Lin, C., Huang, H., Yu, R., Qin, J., Wang, L.: " + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^{\\wedge}" + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": " 2-mil: Exploring hierarchical representation with heterogeneous multiple instance learning for whole slide image analysis. In: AAAI. vol. 36, pp. 933-941 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "text", + "content": "26. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 124-140. Springer (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "text", + "content": "27. Ilse, M., Tomczak, J., Welling, M.: Attention-based deep multiple instance learning. In: International conference on machine learning. pp. 2127-2136. PMLR (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "28. Kang, M., Song, H., Park, S., Yoo, D., Pereira, S.: Benchmarking self-supervised learning on diverse pathology datasets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3344-3354 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "29. Kong, F., Henao, R.: Efficient classification of very large images with tiny objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2384-2394 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "30. Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Pro" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 139 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318-14328 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "text", + "content": "31. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7454-7463 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "text", + "content": "32. Li, R., Yao, J., Zhu, X., Li, Y., Huang, J.: Graph cnn for survival analysis on whole slide pathological images. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 174-182. Springer (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 236 + ], + "type": "text", + "content": "33. Li, Y., Ping, W.: Cancer metastasis detection with neural conditional random field. arXiv preprint arXiv:1806.07064 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "text", + "content": "34. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19830-19839 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 268, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 311 + ], + "type": "text", + "content": "35. Litjens, G., Sánchez, C.I., Timofeeva, N., Hermsen, M., Nagtegaal, I., Kovacs, I., Hulsbergen-Van De Kaa, C., Bult, P., Van Ginneken, B., Van Der Laak, J.: Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis. Scientific reports 6(1), 26286 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "text", + "content": "36. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering 5(6), 555-570 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "type": "text", + "content": "37. Madabhushi, A.: Digital pathology image analysis: opportunities and challenges. Imaging in medicine 1(1), 7 (2009)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "text", + "content": "38. Maron, O., Lozano-Pérez, T.: A framework for multiple-instance learning. Advances in neural information processing systems 10 (1997)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "39. McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 481, + 439 + ], + "type": "text", + "content": "40. Pantanowitz, L., Valenstein, P.N., Evans, A.J., Kaplan, K.J., Pfeifer, J.D., Wilbur, D.C., Collins, L.C., Colgan, T.J.: Review of the current state of whole slide imaging in pathology. Journal of pathology informatics 2(1), 36 (2011)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 439, + 481, + 472 + ], + "type": "text", + "content": "41. Pinckaers, H., Van Ginneken, B., Litjens, G.: Streaming convolutional neural networks for end-to-end learning with multi-megapixel images. IEEE transactions on pattern analysis and machine intelligence 44(3), 1581-1590 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 472, + 481, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 481, + 492 + ], + "type": "text", + "content": "42. Qu, L., Wang, M., Song, Z., et al.: Bi-directional weakly supervised knowledge distillation for whole slide image classification. Neurips 35, 15368-15381 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 492, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 492, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 492, + 481, + 536 + ], + "type": "text", + "content": "43. Qu, L., Yang, Z., Duan, M., Ma, Y., Wang, S., Wang, M., Song, Z.: Boosting whole slide image classification from the perspectives of distribution, correlation and magnification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21463-21473 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 481, + 568 + ], + "type": "text", + "content": "44. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems 34, 2136-2147 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 568, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 481, + 600 + ], + "type": "text", + "content": "45. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "type": "text", + "content": "46. Tang, W., Huang, S., Zhang, X., Zhou, F., Zhang, Y., Liu, B.: Multiple instance learning framework with masked hard instance mining for whole slide image classification. arXiv preprint arXiv:2307.15254 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "text", + "content": "47. Tellez, D., Litjens, G., van der Laak, J., Ciompi, F.: Neural image compression for gigapixel histopathology image analysis. IEEE transactions on pattern analysis and machine intelligence 43(2), 567-578 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "ACMIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 643 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "48. Tiwari, R., Shenoy, P.: Overcoming simplicity bias in deep networks using a feature sieve (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "49. Wang, D., Khosla, A., Gargeya, R., Irshad, H., Beck, A.H.: Deep learning for identifying metastatic breast cancer. arXiv preprint arXiv:1606.05718 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 162, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 193 + ], + "type": "text", + "content": "50. Wang, H., Luo, L., Wang, F., Tong, R., Chen, Y.W., Hu, H., Lin, L., Chen, H.: Iteratively coupled multiple instance learning from instance to bag classifier for whole slide image classification. arXiv preprint arXiv:2303.15749 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "text", + "content": "51. Wang, X., Xiang, J., Zhang, J., Yang, S., Yang, Z., Wang, M.H., Zhang, J., Yang, W., Huang, J., Han, X.: Scl-wc: Cross-slide contrastive learning for weakly-supervised whole-slide image classification. Advances in neural information processing systems 35, 18009-18021 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 239, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 480, + 281 + ], + "type": "text", + "content": "52. Wang, Y., Kartasalo, K., Weitz, P., Acs, B., Valkonen, M., Larsson, C., Ruusuvuori, P., Hartman, J., Rantalainen, M.: Predicting molecular phenotypes from histopathology images: A transcriptome-wide expression-morphology analysis in breast cancer. Cancer research 81(19), 5115-5126 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 480, + 314 + ], + "type": "text", + "content": "53. Xiong, C., Chen, H., Sung, J., King, I.: Diagnose like a pathologist: Transformer-enabled hierarchical attention-guided multiple instance learning for whole slide image classification. arXiv preprint arXiv:2301.08125 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 316, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 316, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 133, + 316, + 480, + 358 + ], + "type": "text", + "content": "54. Yang, J., Chen, H., Zhao, Y., Yang, F., Zhang, Y., He, L., Yao, J.: Remix: A general and efficient framework for multiple instance learning based whole slide image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 35-45. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 360, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 360, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 360, + 480, + 390 + ], + "type": "text", + "content": "55. Yao, J., Zhu, X., Jonnagaddala, J., Hawkins, N., Huang, J.: Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis 65, 101789 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 392, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 434 + ], + "type": "text", + "content": "56. Yufei, C., Liu, Z., Liu, X., Liu, X., Wang, C., Kuo, T.W., Xue, C.J., Chan, A.B.: Bayes-mil: A new probabilistic perspective on attention-based multiple instance learning for whole slide images. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 436, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 436, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 436, + 480, + 479 + ], + "type": "text", + "content": "57. Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18802-18812 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 480, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 480, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 133, + 480, + 480, + 521 + ], + "type": "text", + "content": "58. Zhang, Y., Sun, Y., Li, H., Zheng, S., Zhu, C., Yang, L.: Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 242-252. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 523, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 480, + 577 + ], + "type": "text", + "content": "59. Zhao, Y., Yang, F., Fang, Y., Liu, H., Zhou, N., Zhang, J., Sun, J., Yang, S., Menze, B., Fan, X., et al.: Predicting lymph node metastasis using histopathological images based on multiple instance learning with deep graph convolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4837-4846 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 480, + 609 + ], + "type": "text", + "content": "60. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13001-13008 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "text", + "content": "61. Zhu, X., Yao, J., Zhu, F., Huang, J.: Wsisa: Making survival prediction from whole slide histopathological images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7234-7242 (2017)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_content_list.json b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fcc50c6df74b936048b5d6cf26073515afe7a3e1 --- /dev/null +++ b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_content_list.json @@ -0,0 +1,1681 @@ +[ + { + "type": "text", + "text": "AttentionHand: Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild", + "text_level": 1, + "bbox": [ + 218, + 141, + 782, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junho Park $^{1,2*}$ , Kyeongbo Kong $^{3*}$ , and Suk-Ju Kang $^{1*}$", + "bbox": [ + 279, + 238, + 723, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Electronic Engineering, Sogang University, South Korea $^{2}$ AI Lab, CTO Division, LG Electronics, South Korea", + "bbox": [ + 256, + 263, + 746, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Electrical & Electronics Engineering, Pusan National University, South Korea", + "bbox": [ + 228, + 292, + 772, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "junho18.park@gmail.com kbkong@pusan.ac.kr sj Kang@sogang.ac.kr https://github.com/redorangeyellow/AttentionHand", + "bbox": [ + 259, + 321, + 741, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recently, there has been a significant amount of research conducted on 3D hand reconstruction to use various forms of human-computer interaction. However, 3D hand reconstruction in the wild is challenging due to extreme lack of in-the-wild 3D hand datasets. Especially, when hands are in complex pose such as interacting hands, the problems like appearance similarity, self-handed occlusion and depth ambiguity make it more difficult. To overcome these issues, we propose AttentionHand, a novel method for text-driven controllable hand image generation. Since AttentionHand can generate various and numerous in-the-wild hand images well-aligned with 3D hand label, we can acquire a new 3D hand dataset, and can relieve the domain gap between indoor and outdoor scenes. Our method needs easy-to-use four modalities (i.e., an RGB image, a hand mesh image from 3D label, a bounding box, and a text prompt). These modalities are embedded into the latent space by the encoding phase. Then, through the text attention stage, hand-related tokens from the given text prompt are attended to highlight hand-related regions of the latent embedding. After the highlighted embedding is fed to the visual attention stage, hand-related regions in the embedding are attended by conditioning global and local hand mesh images with the diffusion-based pipeline. In the decoding phase, the final feature is decoded to new hand images, which are well-aligned with the given hand mesh image and text prompt. As a result, AttentionHand achieved state-of-the-art among text-to-hand image generation models, and the performance of 3D hand mesh reconstruction was improved by additionally training with hand images generated by AttentionHand.", + "bbox": [ + 261, + 412, + 740, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Hand Mesh Reconstruction $\\cdot$ Text-to-Image Generation", + "bbox": [ + 261, + 776, + 740, + 791 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 217, + 810, + 364, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\text{忍}$ Corresponding author.", + "bbox": [ + 217, + 824, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5c47883ff948277ba3d6cbf11820aa8d952ad9d23eb7b1625f6f6c0da0dd07c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 467, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/12d027893901489413cf8e86ba89e989cd97aa9c5d7ce4f1e9749609d3cf8f22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 146, + 782, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/704b89104c4cd11d2237f0c636f67bf6e102e61653250703e2a05833d4d0ea2d.jpg", + "image_caption": [ + "(b)", + "(c)", + "Fig. 1: Various acquisition types of 3D hand datasets. (a) In-the-wild dataset (i.e., MSCOCO [6]) is naively acquired with inaccurate pseudo annotation, (b) relighted dataset (i.e., Re:InterHand [5]) consists of unnatural hands with inharmonious background, and (c) our in-the-wild dataset from AttentionHand, which is annotated with accurate 3D labels, contains natural hands with harmonious background, easy to generate, and can be made infinitely." + ], + "image_footnote": [], + "bbox": [ + 220, + 252, + 782, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 537, + 375, + 553 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of 3D hand mesh reconstruction is to recover the 3D hand mesh from a single RGB image. It becomes difficult when hands are in the wild, due to insufficiency of in-the-wild 3D hand datasets. Compared to in-the-lab datasets [1-3], acquisition in-the-wild datasets is challenging due to unpredictable conditions such as weather, lighting, cost of sensors, and safety issues on crowded roads and public places. Even if an in-the-wild dataset is collected, data diversity would be poor due to the aforementioned severe constraints. Although arbitrary labels can be obtained through pseudo annotation, the precision and accuracy is still poor compared to in-the-lab datasets as shown in Fig. 1(a). To tackle this problem, several synthetic datasets [4, 5] have introduced. However, since the hand and background images are synthesized out of harmony, they consist of unnatural and unrealistic hand images as shown in Fig. 1(b). Hence, it is difficult to overcome the domain gap between indoor and outdoor scenes with synthetic datasets.", + "bbox": [ + 212, + 568, + 787, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, when hands are in a complex pose like interacting hands, it becomes even more challenging to reconstruct 3D hand meshes due to the appearance similarity, self-handed occlusion and depth ambiguity. Starting with InterHand2.6M [7], several works [8-15] have emerged to solve the complex hand", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/427ef0e5947b2e7350ebc2c61e31c19ce434208d0e894781630774db6a60b58c.jpg", + "image_caption": [ + "Fig. 2: Visualization of attention maps with corresponding tokens from given text prompts. Red and green boxes represent attention maps without and with Attention-Hand, respectively." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 782, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pose. However, they have been employed and evaluated primarily on in-the-lab scenes except for InterWild [16]; it tried to relieve the domain gap by leveraging the geometric features of the hand, which is not affected by the domain. Nevertheless, since InterWild was trained with MSCOCO [6], which is extremely lack of in-the-wild hand images with inaccurate 3D labels, there is earnestly need of more diverse, numerous, and accurately annotated in-the-wild datasets.", + "bbox": [ + 212, + 333, + 784, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address aforementioned issues, we propose AttentionHand, a new method for the text-driven controllable hand image generation. AttentionHand is designed based on Stable Diffusion (SD) [17] to create accurate, natural, realistic and harmonious in-the-wild hand images easily and infinitely as shown in Fig. 1(c). AttentionHand has a huge advantage: we can simply generate images with only four modalities - an RGB image, the corresponding hand mesh image, bounding box, and text prompt. Therefore, we can generate (1) various in-the-wild hand images with flexible text prompts, and (2) well-aligned hand images with 3D hand label. By generating new samples with AttentionHand, we can alleviate aforementioned issues of the 3D hand mesh reconstruction in the wild.", + "bbox": [ + 212, + 426, + 784, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train AttentionHand, we need to additionally prepare a local RGB image and local mesh image for attention of hand-focused region of the image. The preparation of local information is essential because hands commonly occupy relatively small region in the image. Hence, we obtain local RGB and mesh images by cropping and resizing original RGB and mesh images (i.e., we define them as global information) with the bounding box of hand region. After encoding prepared information in the encoding phase, encoded latent embeddings are fed to the conditioning phase, which is composed of the text attention stage (TAS) and visual attention stage (VAS).", + "bbox": [ + 212, + 580, + 784, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TAS attends on hand-related tokens from the given text prompt by leveraging attention maps as shown in Fig. 2. Specifically, TAS extracts hand-related attention maps (i.e., holding and hand), and these attention maps are updated to highlight hand-related regions by the refinement based on the softmax operation and Gaussian filter. With TAS, we can obtain more hand-focused images than before. On the other hand, VAS attends on hand-related regions by conditioning global and local hand mesh images with the SD-based pipeline. With global and local information, AttentionHand can be jointly optimized to reflect the global", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "AttentionHand", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "context (i.e., in-the-wild background) and local context (i.e., hand-focused foreground.) In the end of the conditioning phase, we finally get the diffusion feature, which is decoded to new hand images in the decoding phase. Hence, Attention-Hand can generate well-aligned hand images with the given mesh image and text prompt for the 3D hand mesh reconstruction in the wild.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To prove the excellence of AttentionHand, we conducted extensive experiments for the text-to-hand image generation and 3D hand mesh reconstruction. As a result, AttentionHand achieved state-of-the-art in the text-to-hand image generation, and the performance of 3D hand mesh reconstruction was considerably improved by additionally training with new hand samples generated by AttentionHand. Especially, the performance was enhanced significantly on in-the-wild datasets, which implies AttentionHand can generate various and well-annotated in-the-wild hand images. The summary of our contributions is as follows:", + "bbox": [ + 212, + 222, + 784, + 356 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel method, AttentionHand, which generates well-aligned in-the-wild hand images in a simple manner without laborious data acquisition.", + "- AttentionHand is designed based on a generative model that attends on hand-related tokens from the text prompt and hand-related regions from the hand mesh image, for generating hand-focused images.", + "- AttentionHand achieved state-of-the-art in the text-to-hand image generation, and we verified that utilizing the dataset generated from AttentionHand improves the performance on 3D hand mesh reconstruction in the wild." + ], + "bbox": [ + 225, + 367, + 784, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 510, + 387, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 215, + 542, + 477, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-to-image generation aims to synthesize high-resolution image from natural language descriptions. With the advent of diffusion models, various studies on text-to-image generation have been conducted in recent years [17-21]. Specifically, ControlNet [18] and T2I-Adapter [19] proposed novel approaches to incorporate arbitrary condition into the generation process. Recently, UniControlNet [20] presented a novel approach that allows for the simultaneous utilization of various conditions in a flexible and composable manner. Nevertheless, aforementioned models exhibited common limitations in generating hand images, due to the relatively small size of hands within the overall image resolution.", + "bbox": [ + 212, + 566, + 784, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Generative Models for Hand", + "text_level": 1, + "bbox": [ + 215, + 739, + 500, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GANs for Hand. There are several works [22-25] to tackle the hand image generation problem with the generative adversarial network (GAN) [26]. Specifically, a novel network for image-to-image translation [22] was proposed to make generated images follow the same statistical distribution as real-world hand images. GestureGAN [23] was designed to translate hand gesture-to-gesture with", + "bbox": [ + 212, + 763, + 784, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the explicit hand skeleton information through the color loss and the cycle-consistency loss. Moreover, the first model-aware gesture-to-gesture translation framework [24] was introduced with hand prior as the intermediate representation. Recently, a new method [25], which employs the expressive model-aware hand-object representation and leverages its inherent topology to build the unified surface space, was proposed. However, these works have a common limitation; they are confined to target gestures in generating new hand images. In other words, they are inappropriate to generate in-the-wild images focused on various hands.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion Models for Hand. Recently, some works [27-29] have been addressed hand-related problems with diffusion models. DiffHand [27] introduced the first diffusion-based framework that approaches hand mesh reconstruction as a denoising diffusion process. HandDiffuse [28] proposed a strong baseline for the controllable motion generation of interacting hands using various controllers by designing a diffusion-based model. HandRefiner [29] presented an inpainting pipeline to rectify malformed human hands in generated images with diffusion-based models. However, since these models are not text-driven methods, they cannot generate various in-the-wild hand images conditioned on language instructions.", + "bbox": [ + 212, + 301, + 787, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 476, + 330, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce AttentionHand, a novel framework for creating various and plausible hand images. AttentionHand is a SD-based framework that can generate new RGB images infinitely conditioned on hand mesh images and text prompts. The overall pipeline is shown in Fig. 3.", + "bbox": [ + 212, + 507, + 785, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Data Preparation Phase", + "text_level": 1, + "bbox": [ + 214, + 589, + 464, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the first box of Fig. 3, it just requires four inputs to train AttentionHand: (1) a global RGB hand image $I_{RGB}^{G} \\in \\mathbb{R}^{3 \\times 512 \\times 512}$ , (2) a global hand mesh image $I_{mesh}^{G} \\in \\mathbb{R}^{3 \\times 512 \\times 512}$ , (3) a bounding box of the hand region $B \\in \\mathbb{R}^{1 \\times 4}$ , and (4) a hand-related text prompt $U$ . However, since hands typically occupy small areas on in-the-wild scenes, we also obtain a local RGB hand image $I_{RGB}^{L} \\in \\mathbb{R}^{3 \\times 512 \\times 512}$ and a local hand mesh image $I_{mesh}^{L} \\in \\mathbb{R}^{3 \\times 512 \\times 512}$ by cropping and resizing $I_{RGB}^{G}$ and $I_{mesh}^{G}$ with $B$ . This combination of local and global information enhances hand image conditioning. Details will be explained in the supplementary materials.", + "bbox": [ + 212, + 612, + 787, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Encoding Phase", + "text_level": 1, + "bbox": [ + 214, + 770, + 397, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the diffusion process in latent space, encoding phase for $I_{RGB}^{G}$ , $I_{RGB}^{L}$ and $U$ is implemented by the encoder $\\mathcal{E}$ . It makes global and local latent image embeddings $X_0^G, X_0^L \\in \\mathbb{R}^{4\\times 64\\times 64}$ for $I_{RGB}^{G}$ and $I_{RGB}^{L}$ , and a latent text embedding", + "bbox": [ + 212, + 792, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "AttentionHand", + "bbox": [ + 629, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1a29bab4c2648fe0b78e25c48f45afa72b9c5c110b4b752b48b5c585ded6690c.jpg", + "image_caption": [ + "Fig. 3: Overall pipeline of AttentionHand. In the data preparation phase, we prepare global and local RGB images, global and local hand mesh images, bounding box, and text prompt. In the encoding phase, we get global and local latent image embeddings through VQ-GAN [30], and text embedding through CLIP [31]. In the conditioning phase, we refine image embeddings through the text attention stage, and obtain the diffusion feature through the visual attention stage. In the decoding phase, we generate a new hand image $\\hat{I}_{RGB}$ from $Y_{d}$ through VQ-GAN." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$K \\in \\mathbb{R}^{77 \\times 768}$ for $U$ . Specifically, $X_0^G$ and $X_0^L$ are obtained by VQ-GAN [30], and $K$ is obtained by CLIP [31] as shown in the second box of Fig. 3. These latent embeddings are fed as inputs to the conditioning phase, which will be introduced by the next subsection. The encoding phase is expressed as follows:", + "bbox": [ + 212, + 458, + 787, + 521 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {0} ^ {G}, X _ {0} ^ {L}, K = \\mathcal {E} \\left(I _ {R G B} ^ {G}, I _ {R G B} ^ {L}, U\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 530, + 785, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Conditioning Phase", + "text_level": 1, + "bbox": [ + 215, + 580, + 426, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For generating new hand images conditioned by given text prompt and mesh images, we design the text attention stage (TAS) and the visual attention stage (VAS) in the conditioning phase, as shown in the third box of Fig. 3. TAS is a stage of paying attention to tokens for the hand and its corresponding gesture in a given text. VAS is a stage of training the SD-based model specialized for hand image generation by conditioning global and local mesh images.", + "bbox": [ + 212, + 606, + 787, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text Attention Stage (TAS). TAS is a stage of attending tokens which represent hand or gestures in a given text prompt as shown in Fig. 4(a). First, by adding Gaussian noise to $X_0^G$ and $X_0^L$ with $t$ diffusion steps, the global noisy embedding $X_t^G$ and local noisy embedding $X_t^L$ are obtained. For simplicity, we define as $X_0 = (X_0^G, X_0^L)$ and $X_t = (X_t^G, X_t^L)$ . Then, $X_t$ and $K$ are fed to TAS as inputs. For the text attention of TAS, we utilize the cross attention [32]. Specifically, an attention map $A \\in \\mathbb{R}^{H \\times W \\times N}$ is obtained by calculating the key (i.e., $K$ ) and query (i.e., $Q$ , which is the linear projection of intermediate feature", + "bbox": [ + 212, + 718, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/28a6c76d0d51a82ec13193fa5912a81df0e28166b724a7c10d6fc70384cca3c4.jpg", + "image_caption": [ + "Fig. 4: (a) Overall process of the text attention stage (TAS). By leveraging the hand-related tagging and refinement, we can highlight hand-related attention maps, which leads to update noisy embeddings with $\\mathcal{L}^{TAS}$ . (b) Overall process of the visual attention stage (VAS). By utilizing the global and local information, we can obtain the harmonious diffusion feature, which leads to generate high-fidelity hand images." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 531, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03fe6f18fc5147b5951beb4fe431389fe84841795d89690833447bc8258e6e54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 150, + 779, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "map from $X_{t}$ of U-Net [33] in SD). $H$ and $W$ denote the height and width of $A$ , and $N$ denotes the number of all tokens of $K$ .", + "bbox": [ + 212, + 404, + 784, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, to extract hand-related attention maps $A_{k\\in K}\\in \\mathbb{R}^{H\\times W\\times N_k}$ in $A$ , we design the hand-related tagging $\\mathcal{H}_{tag}$ , which is based on part-of-speech tagging [34], where $N_{k}$ denotes the number of hand-related tokens in $K$ . Specifically, $\\mathcal{H}_{tag}$ determines if the input token indicates the hand-related word (i.e., holding, taking, or hand). With $\\mathcal{H}_{tag}$ , we can attend hand-related tokens $k$ to generate more hand-focused images. More details are in the supplementary materials.", + "bbox": [ + 212, + 436, + 787, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, we employ the softmax operation and Gaussian smoothing to maximize the effect of $A_{k}$ . Since the Gaussian filter effectively removes noise from images and preserves detailed information by using the average value of surrounding pixels, we fully exploit these advantages. Hence, $A_{k}$ is updated to $\\hat{A}_k\\in \\mathbb{R}^{H\\times W\\times N_k}$ by refining hand-related attention maps as follows:", + "bbox": [ + 212, + 530, + 787, + 606 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {A} _ {k} = \\text {G a u s s i a n} (\\operatorname {S o f t m a x} (A _ {k})). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 621, + 784, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For simplicity, we define $\\hat{A} \\in \\mathbb{R}^{H \\times W \\times N}$ as the concatenation of $\\hat{A}_{k \\in K}$ and $A_{l \\notin K} \\in \\mathbb{R}^{H \\times W \\times N_l}$ , where $N_l$ denotes the number of not hand-related tokens in $K$ , and $N = N_k + N_l$ .", + "bbox": [ + 212, + 652, + 787, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, optimization for evenly reflecting the image features of all attention maps is necessary. In other words, it is required to design an objective to prevent poor generation of the image feature for a specific token. Specifically, for arbitrary token $n \\in K$ , the highest value $s_n$ among all patches in the $n$ -th refined attention map $\\hat{A}_n$ is extracted, and it is subtracted from 1. This operation is implemented for all tokens in $K$ , and a novel loss, which named $\\mathcal{L}^{TAS}$ , is computed the largest value among them as follows:", + "bbox": [ + 212, + 702, + 787, + 808 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {T A S} = \\max _ {n \\in K} \\left(1 - s _ {n}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 821, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "AttentionHand", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on $\\mathcal{L}^{TAS}$ , $X_{t}$ is updated to $\\hat{X}_{t} = (\\hat{X}_{t}^{G},\\hat{X}_{t}^{L})$ as follows:", + "bbox": [ + 215, + 143, + 653, + 162 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {t} = X _ {t} - \\alpha_ {t} \\nabla_ {X _ {t}} \\mathcal {L} ^ {T A S}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 169, + 785, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ indicates the learning rate, and $\\hat{X}_t^G$ and $\\hat{X}_t^L$ indicate updated global and local noisy embedding, respectively.", + "bbox": [ + 214, + 195, + 785, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Attention Stage (VAS). VAS is a stage of training SD-based model by conditioning the aforementioned global and local mesh image. VAS is composed of two modules: one is the guidance module $\\phi_{g}$ , and the other is the diffusion module $\\phi_{d}$ as shown in Fig. 4(b). First, the diffusion module $\\phi_{d}$ is designed based on an U-Net network, consisting of 25 blocks: 8 blocks are downsampling and upsampling convolution layers, and the remaining 17 blocks consist of four ResNet [35] layers and two Vision Transformers [36]. We define the parameter set of $\\phi_{d}$ as $\\theta_{d}$ , which is fixed frozen to maintain the image generation performance of SD.", + "bbox": [ + 212, + 244, + 782, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the other hand, the guidance module $\\phi_{g}$ is also based on an U-Net network with 25 blocks of $\\phi_{d}$ . We define the parameter set of $\\phi_{g}$ as $\\theta_{g}$ , which is a copied version of $\\theta_{d}$ . Different from $\\theta_{d}$ , $\\theta_{g}$ is set to be learnable for generating images conditioned to $I_{mesh}^{G}$ and $I_{mesh}^{L}$ . Specifically, $\\phi_{g}$ has zero convolution $\\mathcal{Z}$ [18] at the front of the network, and last 12 blocks of the network consist of $\\mathcal{Z}$ . Since $\\mathcal{Z}$ is defined as a $1\\times 1$ convolution layer whose weights and bias are initialized to zero, the gradients of the weight and bias progressively grow from zeros to optimized parameters in a learnable manner. Hence, $\\mathcal{Z}$ helps generated images to be conditioned on $I_{mesh}^{G}$ and $I_{mesh}^{L}$ , while maintaining the quality of image generation. More specifically, $\\phi_{d}$ and $\\phi_{g}$ share weights at the beginning of training, because parameter sets of both modules, i.e., $\\theta_{d}$ and $\\theta_{g}$ , are initialized with the pre-trained SD. However, while continuing with training process, $\\theta_{g}$ is updated to learn $I_{mesh}^{G}$ and $I_{mesh}^{L}$ , whereas $\\theta_{d}$ is fixed frozen to preserve the performance of image generation. At the end of training, $\\theta_{d}$ and $\\theta_{g}$ are completely different from the beginning. Hence, $\\phi_{g}$ is formulated as follows:", + "bbox": [ + 212, + 382, + 787, + 609 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nY _ {g} = \\phi_ {g} (\\hat {X} _ {t}, I _ {\\text {m e s h}}, K, t; \\theta_ {g}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 618, + 784, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $I_{mesh} = (I_{mesh}^{G}, I_{mesh}^{L})$ denotes the concatenation of the global and local mesh image, $t$ denotes the diffusion step obtained by positional encoding, $Y_{g} = (Y_{g}^{G}, Y_{g}^{L}) \\in \\mathbb{R}^{2 \\times 4 \\times 64 \\times 64}$ denotes the concatenation of the global guidance feature $Y_{g}^{G}$ and local guidance feature $Y_{g}^{L}$ . Next, $\\phi_{d}$ is formulated as follows:", + "bbox": [ + 215, + 643, + 787, + 710 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nY _ {d} = \\phi_ {d} \\left(\\hat {X} _ {t}, K, t; \\theta_ {d}\\right) + Y _ {g}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 719, + 784, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $Y_{d} = (Y_{d}^{G},Y_{d}^{L})\\in \\mathbb{R}^{2\\times 4\\times 64\\times 64}$ indicates the concatenation of the global diffusion feature $Y_{d}^{G}$ and local diffusion feature $Y_{d}^{L}$ .", + "bbox": [ + 214, + 742, + 785, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimization. Since the diffusion model typically involves both forward and reverse processes, our AttentionHand also employs two processes. For the forward process, the noisy embedding $X_{t} = (X_{t}^{G},X_{t}^{L})$ is obtained by progressively", + "bbox": [ + 214, + 794, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "perturbing Gaussian noise $\\epsilon = (\\epsilon^{G},\\epsilon^{L})$ to the initial embedding $X_0 = (X_0^G,X_0^L)$ by $t$ diffusion steps. $\\epsilon^G$ and $\\epsilon^L$ denote the global and local noise added to $X_0^G$ and $X_0^L$ , respectively. Then, since $X_{t}$ is updated to $\\hat{X}_t$ by TAS, $\\epsilon$ is also updated to $\\hat{\\epsilon} = (\\hat{\\epsilon}^{G},\\hat{\\epsilon}^{L})$ . In other words, $\\hat{\\epsilon}$ is considered the residual noise between $X_0$ and $\\hat{X}_t$ . Thus, $\\epsilon^G$ and $\\epsilon^L$ denote the global and local residual noise. For the reverse process, AttentionHand learns to gradually remove residual noises with global and local denoising processes. Therefore, given text embedding $K$ , diffusion steps $t$ , and mesh images $I_{mesh}^{G}$ and $I_{mesh}^{L}$ , the diffusion training network $\\epsilon_{\\theta}$ is optimized to predict $\\hat{\\epsilon}^{G}$ and $\\hat{\\epsilon}^{L}$ jointly through the following objectives:", + "bbox": [ + 212, + 143, + 787, + 291 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {G} = \\mathbb {E} _ {X _ {0} ^ {G}, I _ {\\text {m e s h}} ^ {G}, K, t, \\hat {\\epsilon^ {G}} \\sim \\mathcal {N} (0, 1)} [ \\| \\hat {\\epsilon^ {G}} - \\epsilon_ {\\theta} (\\hat {X} _ {t} ^ {G}, I _ {\\text {m e s h}} ^ {G}, K, t) \\| _ {2} ^ {2} ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 305, + 785, + 329 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {L} = \\mathbb {E} _ {X _ {0} ^ {L}, I _ {\\text {m e s h}} ^ {L}, K, t, \\hat {\\epsilon} ^ {L} \\sim \\mathcal {N} (0, 1)} [ \\| \\hat {\\epsilon} ^ {L} - \\epsilon_ {\\theta} (\\hat {X} _ {t} ^ {L}, I _ {\\text {m e s h}} ^ {L}, K, t) \\| _ {2} ^ {2} ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 357, + 785, + 380 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{L}^G$ and $\\mathcal{L}^L$ indicate the cost function of global and local features, respectively. Thus, the final objective is defined as follows:", + "bbox": [ + 212, + 390, + 782, + 422 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda^ {G} \\mathcal {L} ^ {G} + \\lambda^ {L} \\mathcal {L} ^ {L}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 436, + 784, + 454 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda^G$ and $\\lambda^L$ are weighted coefficients of $\\mathcal{L}^G$ and $\\mathcal{L}^L$ .", + "bbox": [ + 214, + 467, + 630, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Decoding Phase", + "text_level": 1, + "bbox": [ + 214, + 515, + 395, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the decoding phase, we can generate a new RGB hand image $\\hat{I}_{RGB} \\in \\mathbb{R}^{3 \\times 512 \\times 512}$ by passing $Y_{d}$ through the decoder $\\mathcal{D}$ , as shown in the fourth box of Fig. 3. While $\\mathcal{E}$ encodes $X_{0}$ by downsampling $I_{RGB}$ in the latent space, $\\mathcal{D}$ decodes $\\hat{I}_{RGB}$ by upsampling $Y_{d}$ in the pixel space, conditioned to given text prompt and mesh images. The structure of $\\mathcal{D}$ is similar to the decoder of VQ-GAN. The decoding phase is expressed as follows:", + "bbox": [ + 212, + 547, + 795, + 642 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {R G B} = \\mathcal {D} \\left(Y _ {d}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 656, + 784, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 704, + 375, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 744, + 334, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the text-to-image generation, we adopted MSCOCO [6]. For the 3D hand mesh reconstruction, we adopted Hands-In-Action (HIC) [37], Re:InterHand (ReIH) [5], InterHand2.6M (IH2.6M) [7], and MSCOCO. Due to the page limit, details will be explained in the supplementary materials.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "AttentionHand", + "bbox": [ + 629, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/35b1c7021d3606cf52a73266f984b5bd118f647a27fc592a4602f75122249438.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons with state-of-the-art text-to-image generation models." + ], + "table_footnote": [], + "table_body": "
MethodsFID↓KID↓FID-H↓KID-H↓Hand Conf.↑MSE-2D↓MSE-3D↓User Pref.(%)↑
Stable Diffusion [17]40.520.0068450.780.025540.6512.9324.5915.864
Uni-ControlNet [20]30.340.0074437.770.020040.8552.1053.0398.796
T2I-Adapter [19]22.000.0076132.080.015680.9141.5462.45119.676
ControlNet [18]21.670.0065840.320.020980.8101.2522.1827.948
AttentionHand (w/o TAS)21.270.0033128.560.013900.9551.2112.04220.734
AttentionHand (w/TAS)20.710.0030127.090.012870.9651.0261.98636.905
", + "bbox": [ + 274, + 185, + 727, + 252 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e0ff79691e559c26ac675e737abd3f2b2167700d91520f12f79568ba09841603.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparisons with state-of-the-art text-to-image generation models. Red and green boxes in each sample indicate the wrong and current hand bounding box, respectively." + ], + "image_footnote": [], + "bbox": [ + 218, + 263, + 782, + 535 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluation Protocol", + "text_level": 1, + "bbox": [ + 215, + 633, + 426, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For the text-to-image generation, we adopted FID [38], FID-Hand (FID-H), KID [39], KID-Hand (KID-H), the hand confidence score (Hand Conf.) [40], the mean square error of 2D and 3D keypoints (MSE-2D, 3D), and the user preference (User Pref.). For the 3D hand mesh reconstruction, we adopted the mean pervertex position error (MPVPE), the right hand-relative vertex error (RRVE), and the mean relative-root position error (MRRPE). Due to the page limit, details will be explained in the supplementary materials.", + "bbox": [ + 212, + 657, + 787, + 765 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparisons with State-of-the-arts", + "text_level": 1, + "bbox": [ + 215, + 785, + 552, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Text-to-Image Generation. As shown in Table 1, our AttentionHand exhibited the highest performance in all metrics among state-of-the-arts [17-20]. This", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/30fcc7f79f45d4f17dbab3d6b375ab7e968a0ef80a9e618c31926d2591566373.jpg", + "table_caption": [ + "Table 2: Quantitative comparisons with state-of-the-art 3D hand mesh reconstruction methods with and without AttentionHand. The red subscripts indicate the difference in performance with and without AttentionHand." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsIn-the-wildIn-the-lab
HIC [37]ReIH [5]IH2.6M [7]
MPVPE↓RRVE↓MRRPE↓MPVPE↓RRVE↓MRRPE↓MPVPE↓RRVE↓MRRPE↓
IHMR [8]38.5745.51119.6430.9045.5598.4516.9421.9833.39
IHMR+AttentionHand36.73-1.8444.10-1.4194.63-25.0129.11-1.7943.12-2.4387.07-11.3815.09-1.8520.55-1.4332.21-1.18
InterShape [9]27.6634.69110.2527.8738.5680.0412.9717.3531.56
InterShape+AttentionHand25.04-2.6233.33-1.3680.17-30.0826.44-1.4336.54-2.0261.41-18.6311.90-1.0716.22-1.1330.04-1.52
IntagHand [10]23.0728.7452.4625.9030.0542.2212.3417.3229.31
IntagHand+AttentionHand21.87-1.2027.09-1.6547.11-5.3523.39-2.5128.77-1.2833.98-8.2411.42-0.9215.81-1.5129.18-0.13
DIR [13]21.8926.1143.1121.8229.6637.0110.2617.1128.98
DIR+AttentionHand20.66-1.2325.87-0.2440.54-2.5719.91-1.9126.67-2.9935.05-1.9610.09-0.1716.99-0.1228.02-0.96
InterWild [16]15.3021.3531.2613.9920.0722.3811.5219.7726.87
InterWild+AttentionHand14.74-0.5621.10-0.2529.26-2.0013.95-0.0419.94-0.1322.05-0.3310.62-0.9019.09-0.6825.74-1.13
", + "bbox": [ + 246, + 199, + 754, + 308 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/623fef012dd936e01fdf1be0e614dd203f50ff3cc40f47dee6578b3813ce51bc.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparisons on MSCOCO [6]. Red and green boxes indicate wrong and correct region of the reconstructed hand, respectively." + ], + "image_footnote": [], + "bbox": [ + 223, + 329, + 782, + 570 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "is particularly evident in the comparison of FID(-H) and KID(-H), which signify the quality of the generated images being on par with real RGB images. Furthermore, the lowest MSE-2D and MSE-3D indicates the remarkable alignment between the generated images and the corresponding hand mesh images. With respect to the user preference, AttentionHand scored the highest compared to other methods. It implies that most users acknowledged the outstanding quality of hand images generated by AttentionHand. In addition, as shown in Fig. 5, our AttentionHand generated the high-quality hand image which is well-corresponding with the given mesh image and fully reflected the given text prompt. Specifically, even when two-hands mesh image is given, which is more challenging than in the case of single-hand mesh image, AttentionHand generated the hand image robustly. It implies our AttentionHand is proper to generate", + "bbox": [ + 212, + 657, + 787, + 843 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "AttentionHand", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "well-aligned hand images with given mesh images and text prompt. Additional qualitative results are in the supplementary materials.", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3D Hand Mesh Reconstruction. To verify our AttentionHand extensively, we trained state-of-the-art hand pose networks [8-10, 13, 16] by additionally adding new data generated by AttentionHand. As shown in Table 2, the performance of all methods increased for all metrics. Specifically, with respect to the MPVPE, AttentionHand showed the dramatic performance improvement with InterWild [16] about $3.66\\%$ and $7.81\\%$ on HIC and ReIH, respectively. With respect to the RRVE, it increased by about $1.17\\%$ and $0.65\\%$ on HIC and ReIH, respectively. With respect to the MRRPE, it increased by about $6.40\\%$ and $1.47\\%$ on HIC and ReIH, respectively. These imply generated hand images help increasing the accuracy of the 3D hand mesh reconstruction. In addition, the qualitative performance for in-the-wild scenes is also verified as shown in Fig. 6. Although MSCOCO mainly contains in-the-wild situations, 3D hand mesh is reconstructed robustly. It implies that even for difficult situations, the performance of reconstruction can be improved by utilizing AttentionHand. Additional qualitative results are in the supplementary materials.", + "bbox": [ + 212, + 195, + 789, + 422 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 440, + 401, + 455 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Text Attention Stage (TAS). We deeply dived into TAS to verify its superiority. Firstly, as in the last two rows in Table 1, TAS showed its effectiveness in all metrics. In addition, as shown in Fig. 7, attention maps are well described their corresponding tokens in the case of with TAS. It implies that with TAS, AttentionHand can reflect hand-related tokens enough. Additional qualitative results are in the supplementary materials.", + "bbox": [ + 212, + 460, + 787, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Secondly, we conducted more experiments about Gaussian filters as follows: (1) no Gaussian filter, (2) random Gaussian filter, and (3) fixed Gaussian filter. As shown in the second, third, and fourth columns of Fig. 8, we found interesting results: in the case of (1), the hand was disappeared or its shape became strange. In the case of (2), generated images are not well-aligned with given hand mesh images. However, in the case of (3), generated images are well-aligned with given hand mesh images and look natural. Hence, we determined fixed Gaussian filter makes the generated image plausibly regardless of diffusion timestep $t$ .", + "bbox": [ + 212, + 551, + 787, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Thirdly, we compared our loss, $\\mathcal{L}^{T\\hat{A}S}$ , with the load balancing loss $(\\mathcal{L}^{LB})$ [41, 42]. Since $\\mathcal{L}^{LB}$ is an auxiliary loss for balancing loads among experts, it plays a similar role with $\\mathcal{L}^{TAS}$ , which evenly reflects the image features of all the attention maps. Therefore, we replaced $\\mathcal{L}^{TAS}$ to $\\mathcal{L}^{LB}$ and considered its feasibility as shown in the fifth column of Fig. 8. Unfortunately, in the case of $\\mathcal{L}^{LB}$ , generated images are not fit at all with given hand mesh images. We guess while $\\mathcal{L}^{TAS}$ updates the image embedding based on the spatial information of the attention map, $\\mathcal{L}^{LB}$ flattens the 2D attention map as 1D representation, leading to distort spatial knowledge.", + "bbox": [ + 212, + 672, + 787, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Last but not least, we explored the range of updated noise $(\\hat{\\epsilon})$ . According to [43], we set $\\alpha_{t}$ of Eq. 4 as gradually decreasing according to timestep $t$ (i.e.,", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/434ca597a084b46368eb5ad7a59d16ef094b44d688bee6d69ddf5f123d6b91e7.jpg", + "image_caption": [ + "Fig. 7: Ablation studies on the text attention stage (TAS). Attention maps with red and green box are results without and with TAS, respectively. Red and green bounding boxes indicate wrong and correct hand poses, respectively." + ], + "image_footnote": [], + "bbox": [ + 246, + 145, + 754, + 213 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b1339acda2dc8c2a7d3a3dee0ebf8a4eb52badf52d9345db80cf2389cdcf37c2.jpg", + "image_caption": [ + "Fig. 8: Ablation studies on Gaussian filter, losses (i.e. $\\mathcal{L}^{TAS}$ and $\\mathcal{L}^{LB}$ ), and the regularization of $\\hat{\\epsilon}$ for the text attention stage (TAS). Red and green bounding boxes indicate wrong and correct hand poses, respectively." + ], + "image_footnote": [], + "bbox": [ + 305, + 305, + 699, + 448 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "from 20 to 10) for regularization of $\\hat{\\epsilon}$ . However, if $\\alpha_{t}$ is randomly set, $\\hat{\\epsilon}$ tends to be out of distribution (i.e., Gaussian distribution) as shown in the sixth column of Fig. 8: in the case of w/o regularization, generated images are not aligned with given mesh images, or missed some hands. Therefore, it is necessary to regularize $\\hat{\\epsilon}$ for faithful hand image generation.", + "bbox": [ + 212, + 554, + 787, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Model Design Justification. To justify our model's superiority, we compared the characteristics of prior works including our model. As shown in Table 3, our model's distinctive and potential features compared to prior works are (1) harmonious preservation of locality (i.e., hand) with globality (i.e., in-the-wild scene), and (2) selective attention on hand-related tokens by cross attention. Specifically, to harmonize globality and locality, we developed global and local designs for the visual attention stage (VAS). Moreover, since the global and local branches are designed structurally same, we set them to share their weights for reducing the number of training parameters (about $20.2\\% \\downarrow$ ) and improving the generalizability (see two shaded rows in Table 4). We experimentally verified the effectiveness of our design as shown in Table 4.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "AttentionHand", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9d671812ffe92031b0cf663e6f6633823965852bb7875614b97849711596afaa.jpg", + "table_caption": [ + "Table 3: Network comparisons with prior works." + ], + "table_footnote": [], + "table_body": "
MethodsText PromptVisual PromptLocalityHand-related Token Attention
Stable Diffusion✓
Uni-ControlNet✓✓
T2I-Adapter✓✓
ControlNet✓✓
AttentionHand (Ours)✓✓✓✓
", + "bbox": [ + 305, + 171, + 697, + 232 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/22b3ecdc5eecd14733595abdb39c696f932c07f24578809a8613223d54ddd766.jpg", + "table_caption": [ + "Table 4: Ablation studies on the visual attention stage (VAS)." + ], + "table_footnote": [], + "table_body": "
GlobalityLocalityWeightsFID↓KID↓FID-H↓KID-H↓Hand Conf.↑MSE-2D↓MSE-3D↓
Shared40.520.0068450.780.025540.6512.9324.591
✓Shared21.670.0065840.320.020980.8101.2522.182
✓Shared52.980.0071332.110.016040.9111.5392.397
✓✓Shared20.710.0030127.090.012870.9651.0261.986
✓✓Separated21.900.0029326.890.013400.9601.1082.017
", + "bbox": [ + 305, + 273, + 699, + 332 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/609314d8b34e9448b4f346398543efe94a88cb6b367ec593c8abbae59c08e3a4.jpg", + "image_caption": [ + "Fig. 9: (a) Multiple generated hand images from same modalities. Green boxes indicate correct hand poses. (b) t-SNE distribution of AttentionHand and MSCOCO [6]." + ], + "image_footnote": [], + "bbox": [ + 246, + 345, + 754, + 443 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Robustness of Generated Dataset. To verify robustness of our generated dataset, we generated multiple hand images from same modalities as shown in Fig. 9(a). As a result, all generated images are perfectly well-aligned with given hand mesh images. Moreover, we found the t-SNE distribution [44] of AttentionHand is broader than MSCOCO as shown in Fig. 9(b). As a result, we believe that AttentionHand can contribute to the downstream task with our extensive in-the-wild hand images, leading to alleviate the domain gap between indoor and outdoor scenes.", + "bbox": [ + 212, + 527, + 787, + 647 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 671, + 359, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduced a novel text-to-hand image generation model, AttentionHand, which pays attention to the hand-related tokens from the text prompt and global and local mesh images. AttentionHand achieved state-of-the-art performance in text-to-hand image generation, and we demonstrated that training with the dataset generated by our AttentionHand improved the performance of 3D hand mesh reconstruction. However, the diversity may decrease as the generative model is trained to optimize hand mesh images. We expect for the emergence of outstanding diffusion model to improve the diversity and quality of the hand image.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00260091) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation) and Korea Institute for Advancement of Technology (KIAT) grant funded by the Korea Government (MOTIE) (P0020535, The Competency Development Program for Industry Specialist) and National Supercomputing Center with supercomputing resources including technical support (KSC-2023-CRE-0444).", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 287, + 323, + 305 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Hampali, Shreyas and Rad, Mahdi and Oberweger, Markus and Lepetit, Vincent. Honnotate: A method for 3D annotation of hand and object poses. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3196-3206, 2020.", + "2. Chao, Yu-Wei and Yang, Wei and Xiang, Yu and Molchanov, Pavlo and Handa, Ankur and Tremblay, Jonathan and Narang, Yashraj S and Van Wyk, Karl and Iqbal, Umar and Birchfield, Stan and Kautz, Jan and Fox, Dieter. DexYCB: A benchmark for capturing hand grasping of objects. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9044–9053, 2021.", + "3. Ohkawa, Takehiko and He, Kun and Sener, Fadime and Hodan, Tomas and Tran, Luan and Keskin, Cem. Assemblyhands: Towards egocentric activity understanding via 3D hand pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12999-13008, 2023.", + "4. Lin, Fanqing and Wilhelm, Connor and Martinez, Tony. Two-hand global 3d pose estimation using monocular rgb. pages 2373-2381, 2021.", + "5. Moon, Gyeongsik and Saito, Shunsuke and Xu, Weipeng and Joshi, Rohan and Buffalini, Julia and Bellan, Harley and Rosen, Nicholas and Richardson, Jesse and Mize, Mallorie and De Bree, Philippe and Simon, Tomas and Peng, Bo and Garg, Shubham and McPhail, Kevyn and Shiratori, Takaaki. A dataset of relighted 3d interacting hands. Adv. Neural Inform. Process. Syst., 36, 2023.", + "6. Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Dollar, Piotr and Zitnick, C Lawrence. Microsoft coco: Common objects in context. In *Eur. Conf. Comput. Vis.*, pages 740-755, 2014.", + "7. Moon, Gyeongsik and Yu, Shouu-I and Wen, He and Shiratori, Takaaki and Lee, Kyoung Mu. Interhand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single RGB image. In *Eur. Conf. Comput. Vis.*, pages 548-564, 2020.", + "8. Rong, Yu and Wang, Jingbo and Liu, Ziwei and Loy, Chen Change. Monocular 3D reconstruction of interacting hands via collision-aware factorized refinements. In 3DV, pages 432-441, 2021.", + "9. Zhang, Baowen and Wang, Yangang and Deng, Xiaoming and Zhang, Yinda and Tan, Ping and Ma, Cuixia and Wang, Hongan. Interacting two-hand 3D pose and shape reconstruction from single color image. In Int. Conf. Comput. Vis., pages 11354-11363, 2021.", + "10. Li, Mengcheng and An, Liang and Zhang, Hongwen and Wu, Lianpeng and Chen, Feng and Yu, Tao and Liu, Yebin. Interacting attention graph for single image two-hand reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2761-2770, 2022." + ], + "bbox": [ + 225, + 316, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "AttentionHand", + "bbox": [ + 629, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11. Hampali, Shreyas and Sarkar, Sayan Deb and Rad, Mahdi and Lepetit, Vincent. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3D pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11090-11100, 2022.", + "12. Meng, Hao and Jin, Sheng and Liu, Wentao and Qian, Chen and Lin, Mengxiang and Ouyang, Wanli and Luo, Ping. 3D interacting hand pose estimation by hand de-occlusion and removal. In Eur. Conf. Comput. Vis., pages 380-397, 2022.", + "13. Ren, Pengfei and Wen, Chao and Zheng, Xiaozheng and Xue, Zhou and Sun, Haifeng and Qi, Qi and Wang, Jingyu and Liao, Jianxin. Decoupled Iterative Refinement Framework for Interacting Hands Reconstruction from a Single RGB Image. In Int. Conf. Comput. Vis., pages 8014-8025, 2023.", + "14. Zuo, Binghui and Zhao, Zimeng and Sun, Wenqian and Xie, Wei and Xue, Zhou and Wang, Yangang. Reconstructing interacting hands with interaction prior from monocular images. In Int. Conf. Comput. Vis., pages 9054-9064, 2023.", + "15. Li, Lijun and Tian, Linrui and Zhang, Xindi and Wang, Qi and Zhang, Bang and Bo, Liefeng and Liu, Mengyuan and Chen, Chen. Renderih: A large-scale synthetic dataset for 3d interacting hand pose estimation. In Int. Conf. Comput. Vis., pages 20395-20405, 2023.", + "16. Moon, Gyeongsik. Bringing inputs to shared domains for 3D interacting hands recovery in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 17028-17037, 2023.", + "17. Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Björn. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022.", + "18. Zhang, Lvmin and Rao, Anyi and Agrawala, Maneesh. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., pages 3836-3847, 2023.", + "19. Mou, Chong and Wang, Xintao and Xie, Liangbin and Zhang, Jian and Qi, Zhonggang and Shan, Ying and Qie, Xiaohu. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023.", + "20. Zhao, Shihao and Chen, Dongdong and Chen, Yen-Chun and Bao, Jianmin and Hao, Shaozhe and Yuan, Lu and Wong, Kwan-Yee K. Uni-ControlNet: All-in-one control to text-to-image diffusion models. Adv. Neural Inform. Process. Syst., 2023.", + "21. Podell, Dustin and English, Zion and Lacey, Kyle and Blattmann, Andreas and Dockhorn, Tim and Müller, Jonas and Penna, Joe and Rombach, Robin. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023.", + "22. Mueller, Franziska and Bernard, Florian and Sotnychenko, Oleksandr and Mehta, Dushyant and Sridhar, Srinath and Casas, Dan and Theobalt, Christian. GANerated hands for real-time 3D hand tracking from monocular RGB. In IEEE Conf. Comput. Vis. Pattern Recog., pages 49-59, 2018.", + "23. Tang, Hao and Wang, Wei and Xu, Dan and Yan, Yan and Sebe, Nicu. GestureGAN for hand gesture-to-gesture translation in the wild. In ACM Int. Conf. Multimedia, pages 774-782, 2018.", + "24. Hu, Hezhen and Wang, Weilun and Zhou, Wengang and Zhao, Weichao and Li, Houqiang. Model-aware gesture-to-gesture translation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16428-16437, 2021.", + "25. Hu, Hezhen and Wang, Weilun and Zhou, Wengang and Li, Houqiang. Hand-object interaction image generation. Adv. Neural Inform. Process. Syst., 35:23805-23817, 2022." + ], + "bbox": [ + 217, + 146, + 785, + 838 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "26. Goodfellow, Ian and Pouget-Abadie, Jean and Mirza, Mehdi and Xu, Bing and Warde-Farley, David and Ozair, Sherjil and Courville, Aaron and Bengio, Yoshua. Generative adversarial nets. Adv. Neural Inform. Process. Syst., 27:2672-2680, 2014.", + "27. Li, Lijun and Zhuo, Li'an and Zhang, Bang and Bo, Liefeng and Chen, Chen. Diff-Hand: End-to-end hand mesh reconstruction via diffusion models. arXiv preprint arXiv:2305.13705, 2023.", + "28. Lin, Pei and Xu, Sihang and Yang, Hongdi and Liu, Yiran and Chen, Xin and Wang, Jingya and Yu, Jingyi and Xu, Lan. HandDiffuse: Generative controllers for two-hand interactions via diffusion models. arXiv preprint arXiv:2312.04867, 2023.", + "29. Lu, Wenquan and Xu, Yufei and Zhang, Jing and Wang, Chaoyue and Tao, Dacheng. HandRefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. arXiv preprint arXiv:2311.17957, 2023.", + "30. Esser, Patrick and Rombach, Robin and Ommer, Bjorn. Taming transformers for high-resolution image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12873-12883, 2021.", + "31. Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., pages 8748-8763, 2021.", + "32. Chen, Chun-Fu Richard and Fan, Quanfu and Panda, Rameswar. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Int. Conf. Comput. Vis., pages 357-366, 2021.", + "33. Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015.", + "34. Chiche, Alebachew and Yitagesu, Betselot. Part of speech tagging: a systematic review of deep learning and machine learning approaches. Journal of Big Data, 9(1):1-25, 2022.", + "35. He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016.", + "36. Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent., 2020.", + "37. Tzionas, Dimitrios and Ballan, Luca and Srikantha, Abhilash and Aponte, Pablo and Pollefeys, Marc and Gall, Juergen. Capturing hands in action using discriminative salient points and physics simulation. Int. J. Comput. Vis., 118:172-193, 2016.", + "38. Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Adv. Neural Inform. Process. Syst., 30:6626-6637, 2017.", + "39. Binkowski, Mikolaj and Sutherland, Danica J and Arbel, Michael and Gretton, Arthur. Demystifying MMD GANs. In Int. Conf. Learn. Represent., 2018.", + "40. Narasimhaswamy, Supreeth and Bhattacharya, Uttaran and Chen, Xiang and Dasgupta, Ishita and Mitra, Saayan and Hoai, Minh. Handiffuser: Text-to-image generation with realistic hand appearances. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2468-2479, 2024." + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "AttentionHand", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Zhou, Yanqi and Lei, Tao and Liu, Hanxiao and Du, Nan and Huang, Yanping and Zhao, Vincent and Dai, Andrew M and Le, Quoc V and Laudon, James. Mixture-of-experts with expert choice routing. Adv. Neural Inform. Process. Syst., 35:7103-7114, 2022.", + "42. Fedus, William and Zoph, Barret and Shazeer, Noam. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 23(120):1-39, 2022.", + "43. Chefer, Hila and Alaluf, Yuval and Vinker, Yael and Wolf, Lior and Cohen-Or, Daniel. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans. Graph., 42(4):1-10, 2023.", + "44. Van der Maaten, Laurens and Hinton, Geoffrey. Visualizing data using t-SNE. J. Mach. Learn. Res., 9(86):2579-2605, 2008." + ], + "bbox": [ + 215, + 146, + 785, + 314 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_model.json b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d8e6c02e793695b1b97d5d615c9121b7d61e9a0b --- /dev/null +++ b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_model.json @@ -0,0 +1,2271 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.142, + 0.784, + 0.207 + ], + "angle": 0, + "content": "AttentionHand: Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.239, + 0.724, + 0.255 + ], + "angle": 0, + "content": "Junho Park\\(^{1,2*}\\), Kyeongbo Kong\\(^{3*}\\), and Suk-Ju Kang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.265, + 0.747, + 0.293 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Electronic Engineering, Sogang University, South Korea \\(^{2}\\) AI Lab, CTO Division, LG Electronics, South Korea" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.293, + 0.774, + 0.321 + ], + "angle": 0, + "content": "\\(^{3}\\) Department of Electrical & Electronics Engineering, Pusan National University, South Korea" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.323, + 0.742, + 0.35 + ], + "angle": 0, + "content": "junho18.park@gmail.com kbkong@pusan.ac.kr sj Kang@sogang.ac.kr https://github.com/redorangeyellow/AttentionHand" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.414, + 0.741, + 0.761 + ], + "angle": 0, + "content": "Abstract. Recently, there has been a significant amount of research conducted on 3D hand reconstruction to use various forms of human-computer interaction. However, 3D hand reconstruction in the wild is challenging due to extreme lack of in-the-wild 3D hand datasets. Especially, when hands are in complex pose such as interacting hands, the problems like appearance similarity, self-handed occlusion and depth ambiguity make it more difficult. To overcome these issues, we propose AttentionHand, a novel method for text-driven controllable hand image generation. Since AttentionHand can generate various and numerous in-the-wild hand images well-aligned with 3D hand label, we can acquire a new 3D hand dataset, and can relieve the domain gap between indoor and outdoor scenes. Our method needs easy-to-use four modalities (i.e., an RGB image, a hand mesh image from 3D label, a bounding box, and a text prompt). These modalities are embedded into the latent space by the encoding phase. Then, through the text attention stage, hand-related tokens from the given text prompt are attended to highlight hand-related regions of the latent embedding. After the highlighted embedding is fed to the visual attention stage, hand-related regions in the embedding are attended by conditioning global and local hand mesh images with the diffusion-based pipeline. In the decoding phase, the final feature is decoded to new hand images, which are well-aligned with the given hand mesh image and text prompt. As a result, AttentionHand achieved state-of-the-art among text-to-hand image generation models, and the performance of 3D hand mesh reconstruction was improved by additionally training with hand images generated by AttentionHand." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.777, + 0.741, + 0.792 + ], + "angle": 0, + "content": "Keywords: 3D Hand Mesh Reconstruction \\(\\cdot\\) Text-to-Image Generation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.365, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.384, + 0.84 + ], + "angle": 0, + "content": "\\(\\text{忍}\\) Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.384, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.147, + 0.468, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.238, + 0.354, + 0.251 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.147, + 0.783, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.238, + 0.635, + 0.251 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.253, + 0.784, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.388, + 0.512, + 0.401 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.513 + ], + "angle": 0, + "content": "Fig. 1: Various acquisition types of 3D hand datasets. (a) In-the-wild dataset (i.e., MSCOCO [6]) is naively acquired with inaccurate pseudo annotation, (b) relighted dataset (i.e., Re:InterHand [5]) consists of unnatural hands with inharmonious background, and (c) our in-the-wild dataset from AttentionHand, which is annotated with accurate 3D labels, contains natural hands with harmonious background, easy to generate, and can be made infinitely." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.376, + 0.554 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.779 + ], + "angle": 0, + "content": "The goal of 3D hand mesh reconstruction is to recover the 3D hand mesh from a single RGB image. It becomes difficult when hands are in the wild, due to insufficiency of in-the-wild 3D hand datasets. Compared to in-the-lab datasets [1-3], acquisition in-the-wild datasets is challenging due to unpredictable conditions such as weather, lighting, cost of sensors, and safety issues on crowded roads and public places. Even if an in-the-wild dataset is collected, data diversity would be poor due to the aforementioned severe constraints. Although arbitrary labels can be obtained through pseudo annotation, the precision and accuracy is still poor compared to in-the-lab datasets as shown in Fig. 1(a). To tackle this problem, several synthetic datasets [4, 5] have introduced. However, since the hand and background images are synthesized out of harmony, they consist of unnatural and unrealistic hand images as shown in Fig. 1(b). Hence, it is difficult to overcome the domain gap between indoor and outdoor scenes with synthetic datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Moreover, when hands are in a complex pose like interacting hands, it becomes even more challenging to reconstruct 3D hand meshes due to the appearance similarity, self-handed occlusion and depth ambiguity. Starting with InterHand2.6M [7], several works [8-15] have emerged to solve the complex hand" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AttentionHand" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.784, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.256, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Fig. 2: Visualization of attention maps with corresponding tokens from given text prompts. Red and green boxes represent attention maps without and with Attention-Hand, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.785, + 0.424 + ], + "angle": 0, + "content": "pose. However, they have been employed and evaluated primarily on in-the-lab scenes except for InterWild [16]; it tried to relieve the domain gap by leveraging the geometric features of the hand, which is not affected by the domain. Nevertheless, since InterWild was trained with MSCOCO [6], which is extremely lack of in-the-wild hand images with inaccurate 3D labels, there is earnestly need of more diverse, numerous, and accurately annotated in-the-wild datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.785, + 0.577 + ], + "angle": 0, + "content": "To address aforementioned issues, we propose AttentionHand, a new method for the text-driven controllable hand image generation. AttentionHand is designed based on Stable Diffusion (SD) [17] to create accurate, natural, realistic and harmonious in-the-wild hand images easily and infinitely as shown in Fig. 1(c). AttentionHand has a huge advantage: we can simply generate images with only four modalities - an RGB image, the corresponding hand mesh image, bounding box, and text prompt. Therefore, we can generate (1) various in-the-wild hand images with flexible text prompts, and (2) well-aligned hand images with 3D hand label. By generating new samples with AttentionHand, we can alleviate aforementioned issues of the 3D hand mesh reconstruction in the wild." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.785, + 0.717 + ], + "angle": 0, + "content": "To train AttentionHand, we need to additionally prepare a local RGB image and local mesh image for attention of hand-focused region of the image. The preparation of local information is essential because hands commonly occupy relatively small region in the image. Hence, we obtain local RGB and mesh images by cropping and resizing original RGB and mesh images (i.e., we define them as global information) with the bounding box of hand region. After encoding prepared information in the encoding phase, encoded latent embeddings are fed to the conditioning phase, which is composed of the text attention stage (TAS) and visual attention stage (VAS)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "TAS attends on hand-related tokens from the given text prompt by leveraging attention maps as shown in Fig. 2. Specifically, TAS extracts hand-related attention maps (i.e., holding and hand), and these attention maps are updated to highlight hand-related regions by the refinement based on the softmax operation and Gaussian filter. With TAS, we can obtain more hand-focused images than before. On the other hand, VAS attends on hand-related regions by conditioning global and local hand mesh images with the SD-based pipeline. With global and local information, AttentionHand can be jointly optimized to reflect the global" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "context (i.e., in-the-wild background) and local context (i.e., hand-focused foreground.) In the end of the conditioning phase, we finally get the diffusion feature, which is decoded to new hand images in the decoding phase. Hence, Attention-Hand can generate well-aligned hand images with the given mesh image and text prompt for the 3D hand mesh reconstruction in the wild." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.357 + ], + "angle": 0, + "content": "To prove the excellence of AttentionHand, we conducted extensive experiments for the text-to-hand image generation and 3D hand mesh reconstruction. As a result, AttentionHand achieved state-of-the-art in the text-to-hand image generation, and the performance of 3D hand mesh reconstruction was considerably improved by additionally training with new hand samples generated by AttentionHand. Especially, the performance was enhanced significantly on in-the-wild datasets, which implies AttentionHand can generate various and well-annotated in-the-wild hand images. The summary of our contributions is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.368, + 0.784, + 0.397 + ], + "angle": 0, + "content": "- We propose a novel method, AttentionHand, which generates well-aligned in-the-wild hand images in a simple manner without laborious data acquisition." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.442 + ], + "angle": 0, + "content": "- AttentionHand is designed based on a generative model that attends on hand-related tokens from the text prompt and hand-related regions from the hand mesh image, for generating hand-focused images." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.444, + 0.784, + 0.487 + ], + "angle": 0, + "content": "- AttentionHand achieved state-of-the-art in the text-to-hand image generation, and we verified that utilizing the dataset generated from AttentionHand improves the performance on 3D hand mesh reconstruction in the wild." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.511, + 0.388, + 0.527 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.478, + 0.559 + ], + "angle": 0, + "content": "2.1 Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.716 + ], + "angle": 0, + "content": "Text-to-image generation aims to synthesize high-resolution image from natural language descriptions. With the advent of diffusion models, various studies on text-to-image generation have been conducted in recent years [17-21]. Specifically, ControlNet [18] and T2I-Adapter [19] proposed novel approaches to incorporate arbitrary condition into the generation process. Recently, UniControlNet [20] presented a novel approach that allows for the simultaneous utilization of various conditions in a flexible and composable manner. Nevertheless, aforementioned models exhibited common limitations in generating hand images, due to the relatively small size of hands within the overall image resolution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.5, + 0.755 + ], + "angle": 0, + "content": "2.2 Generative Models for Hand" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "GANs for Hand. There are several works [22-25] to tackle the hand image generation problem with the generative adversarial network (GAN) [26]. Specifically, a novel network for image-to-image translation [22] was proposed to make generated images follow the same statistical distribution as real-world hand images. GestureGAN [23] was designed to translate hand gesture-to-gesture with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "AttentionHand" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "the explicit hand skeleton information through the color loss and the cycle-consistency loss. Moreover, the first model-aware gesture-to-gesture translation framework [24] was introduced with hand prior as the intermediate representation. Recently, a new method [25], which employs the expressive model-aware hand-object representation and leverages its inherent topology to build the unified surface space, was proposed. However, these works have a common limitation; they are confined to target gestures in generating new hand images. In other words, they are inappropriate to generate in-the-wild images focused on various hands." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Diffusion Models for Hand. Recently, some works [27-29] have been addressed hand-related problems with diffusion models. DiffHand [27] introduced the first diffusion-based framework that approaches hand mesh reconstruction as a denoising diffusion process. HandDiffuse [28] proposed a strong baseline for the controllable motion generation of interacting hands using various controllers by designing a diffusion-based model. HandRefiner [29] presented an inpainting pipeline to rectify malformed human hands in generated images with diffusion-based models. However, since these models are not text-driven methods, they cannot generate various in-the-wild hand images conditioned on language instructions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.477, + 0.331, + 0.494 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.57 + ], + "angle": 0, + "content": "We introduce AttentionHand, a novel framework for creating various and plausible hand images. AttentionHand is a SD-based framework that can generate new RGB images infinitely conditioned on hand mesh images and text prompts. The overall pipeline is shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.465, + 0.607 + ], + "angle": 0, + "content": "3.1 Data Preparation Phase" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.751 + ], + "angle": 0, + "content": "As shown in the first box of Fig. 3, it just requires four inputs to train AttentionHand: (1) a global RGB hand image \\( I_{RGB}^{G} \\in \\mathbb{R}^{3 \\times 512 \\times 512} \\), (2) a global hand mesh image \\( I_{mesh}^{G} \\in \\mathbb{R}^{3 \\times 512 \\times 512} \\), (3) a bounding box of the hand region \\( B \\in \\mathbb{R}^{1 \\times 4} \\), and (4) a hand-related text prompt \\( U \\). However, since hands typically occupy small areas on in-the-wild scenes, we also obtain a local RGB hand image \\( I_{RGB}^{L} \\in \\mathbb{R}^{3 \\times 512 \\times 512} \\) and a local hand mesh image \\( I_{mesh}^{L} \\in \\mathbb{R}^{3 \\times 512 \\times 512} \\) by cropping and resizing \\( I_{RGB}^{G} \\) and \\( I_{mesh}^{G} \\) with \\( B \\). This combination of local and global information enhances hand image conditioning. Details will be explained in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.398, + 0.788 + ], + "angle": 0, + "content": "3.2 Encoding Phase" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.787, + 0.843 + ], + "angle": 0, + "content": "For the diffusion process in latent space, encoding phase for \\( I_{RGB}^{G} \\), \\( I_{RGB}^{L} \\) and \\( U \\) is implemented by the encoder \\( \\mathcal{E} \\). It makes global and local latent image embeddings \\( X_0^G, X_0^L \\in \\mathbb{R}^{4\\times 64\\times 64} \\) for \\( I_{RGB}^{G} \\) and \\( I_{RGB}^{L} \\), and a latent text embedding" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.784, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Fig. 3: Overall pipeline of AttentionHand. In the data preparation phase, we prepare global and local RGB images, global and local hand mesh images, bounding box, and text prompt. In the encoding phase, we get global and local latent image embeddings through VQ-GAN [30], and text embedding through CLIP [31]. In the conditioning phase, we refine image embeddings through the text attention stage, and obtain the diffusion feature through the visual attention stage. In the decoding phase, we generate a new hand image \\(\\hat{I}_{RGB}\\) from \\(Y_{d}\\) through VQ-GAN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.522 + ], + "angle": 0, + "content": "\\(K \\in \\mathbb{R}^{77 \\times 768}\\) for \\(U\\). Specifically, \\(X_0^G\\) and \\(X_0^L\\) are obtained by VQ-GAN [30], and \\(K\\) is obtained by CLIP [31] as shown in the second box of Fig. 3. These latent embeddings are fed as inputs to the conditioning phase, which will be introduced by the next subsection. The encoding phase is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.531, + 0.787, + 0.55 + ], + "angle": 0, + "content": "\\[\nX _ {0} ^ {G}, X _ {0} ^ {L}, K = \\mathcal {E} \\left(I _ {R G B} ^ {G}, I _ {R G B} ^ {L}, U\\right). \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.581, + 0.427, + 0.597 + ], + "angle": 0, + "content": "3.3 Conditioning Phase" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.699 + ], + "angle": 0, + "content": "For generating new hand images conditioned by given text prompt and mesh images, we design the text attention stage (TAS) and the visual attention stage (VAS) in the conditioning phase, as shown in the third box of Fig. 3. TAS is a stage of paying attention to tokens for the hand and its corresponding gesture in a given text. VAS is a stage of training the SD-based model specialized for hand image generation by conditioning global and local mesh images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Text Attention Stage (TAS). TAS is a stage of attending tokens which represent hand or gestures in a given text prompt as shown in Fig. 4(a). First, by adding Gaussian noise to \\( X_0^G \\) and \\( X_0^L \\) with \\( t \\) diffusion steps, the global noisy embedding \\( X_t^G \\) and local noisy embedding \\( X_t^L \\) are obtained. For simplicity, we define as \\( X_0 = (X_0^G, X_0^L) \\) and \\( X_t = (X_t^G, X_t^L) \\). Then, \\( X_t \\) and \\( K \\) are fed to TAS as inputs. For the text attention of TAS, we utilize the cross attention [32]. Specifically, an attention map \\( A \\in \\mathbb{R}^{H \\times W \\times N} \\) is obtained by calculating the key (i.e., \\( K \\)) and query (i.e., \\( Q \\), which is the linear projection of intermediate feature" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AttentionHand" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.532, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.151, + 0.78, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.302, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Fig. 4: (a) Overall process of the text attention stage (TAS). By leveraging the hand-related tagging and refinement, we can highlight hand-related attention maps, which leads to update noisy embeddings with \\(\\mathcal{L}^{TAS}\\). (b) Overall process of the visual attention stage (VAS). By utilizing the global and local information, we can obtain the harmonious diffusion feature, which leads to generate high-fidelity hand images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.785, + 0.435 + ], + "angle": 0, + "content": "map from \\( X_{t} \\) of U-Net [33] in SD). \\( H \\) and \\( W \\) denote the height and width of \\( A \\), and \\( N \\) denotes the number of all tokens of \\( K \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.529 + ], + "angle": 0, + "content": "Next, to extract hand-related attention maps \\(A_{k\\in K}\\in \\mathbb{R}^{H\\times W\\times N_k}\\) in \\(A\\), we design the hand-related tagging \\(\\mathcal{H}_{tag}\\), which is based on part-of-speech tagging [34], where \\(N_{k}\\) denotes the number of hand-related tokens in \\(K\\). Specifically, \\(\\mathcal{H}_{tag}\\) determines if the input token indicates the hand-related word (i.e., holding, taking, or hand). With \\(\\mathcal{H}_{tag}\\), we can attend hand-related tokens \\(k\\) to generate more hand-focused images. More details are in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.607 + ], + "angle": 0, + "content": "Then, we employ the softmax operation and Gaussian smoothing to maximize the effect of \\( A_{k} \\). Since the Gaussian filter effectively removes noise from images and preserves detailed information by using the average value of surrounding pixels, we fully exploit these advantages. Hence, \\( A_{k} \\) is updated to \\( \\hat{A}_k\\in \\mathbb{R}^{H\\times W\\times N_k} \\) by refining hand-related attention maps as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.622, + 0.785, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\hat {A} _ {k} = \\text {G a u s s i a n} (\\operatorname {S o f t m a x} (A _ {k})). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.701 + ], + "angle": 0, + "content": "For simplicity, we define \\(\\hat{A} \\in \\mathbb{R}^{H \\times W \\times N}\\) as the concatenation of \\(\\hat{A}_{k \\in K}\\) and \\(A_{l \\notin K} \\in \\mathbb{R}^{H \\times W \\times N_l}\\), where \\(N_l\\) denotes the number of not hand-related tokens in \\(K\\), and \\(N = N_k + N_l\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Moreover, optimization for evenly reflecting the image features of all attention maps is necessary. In other words, it is required to design an objective to prevent poor generation of the image feature for a specific token. Specifically, for arbitrary token \\( n \\in K \\), the highest value \\( s_n \\) among all patches in the \\( n \\)-th refined attention map \\( \\hat{A}_n \\) is extracted, and it is subtracted from 1. This operation is implemented for all tokens in \\( K \\), and a novel loss, which named \\( \\mathcal{L}^{TAS} \\), is computed the largest value among them as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.823, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {T A S} = \\max _ {n \\in K} \\left(1 - s _ {n}\\right). \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.145, + 0.655, + 0.163 + ], + "angle": 0, + "content": "Based on \\(\\mathcal{L}^{TAS}\\), \\(X_{t}\\) is updated to \\(\\hat{X}_{t} = (\\hat{X}_{t}^{G},\\hat{X}_{t}^{L})\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.17, + 0.786, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {t} = X _ {t} - \\alpha_ {t} \\nabla_ {X _ {t}} \\mathcal {L} ^ {T A S}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.196, + 0.786, + 0.227 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) indicates the learning rate, and \\(\\hat{X}_t^G\\) and \\(\\hat{X}_t^L\\) indicate updated global and local noisy embedding, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.784, + 0.382 + ], + "angle": 0, + "content": "Visual Attention Stage (VAS). VAS is a stage of training SD-based model by conditioning the aforementioned global and local mesh image. VAS is composed of two modules: one is the guidance module \\(\\phi_{g}\\), and the other is the diffusion module \\(\\phi_{d}\\) as shown in Fig. 4(b). First, the diffusion module \\(\\phi_{d}\\) is designed based on an U-Net network, consisting of 25 blocks: 8 blocks are downsampling and upsampling convolution layers, and the remaining 17 blocks consist of four ResNet [35] layers and two Vision Transformers [36]. We define the parameter set of \\(\\phi_{d}\\) as \\(\\theta_{d}\\), which is fixed frozen to maintain the image generation performance of SD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.61 + ], + "angle": 0, + "content": "On the other hand, the guidance module \\(\\phi_{g}\\) is also based on an U-Net network with 25 blocks of \\(\\phi_{d}\\). We define the parameter set of \\(\\phi_{g}\\) as \\(\\theta_{g}\\), which is a copied version of \\(\\theta_{d}\\). Different from \\(\\theta_{d}\\), \\(\\theta_{g}\\) is set to be learnable for generating images conditioned to \\(I_{mesh}^{G}\\) and \\(I_{mesh}^{L}\\). Specifically, \\(\\phi_{g}\\) has zero convolution \\(\\mathcal{Z}\\) [18] at the front of the network, and last 12 blocks of the network consist of \\(\\mathcal{Z}\\). Since \\(\\mathcal{Z}\\) is defined as a \\(1\\times 1\\) convolution layer whose weights and bias are initialized to zero, the gradients of the weight and bias progressively grow from zeros to optimized parameters in a learnable manner. Hence, \\(\\mathcal{Z}\\) helps generated images to be conditioned on \\(I_{mesh}^{G}\\) and \\(I_{mesh}^{L}\\), while maintaining the quality of image generation. More specifically, \\(\\phi_{d}\\) and \\(\\phi_{g}\\) share weights at the beginning of training, because parameter sets of both modules, i.e., \\(\\theta_{d}\\) and \\(\\theta_{g}\\), are initialized with the pre-trained SD. However, while continuing with training process, \\(\\theta_{g}\\) is updated to learn \\(I_{mesh}^{G}\\) and \\(I_{mesh}^{L}\\), whereas \\(\\theta_{d}\\) is fixed frozen to preserve the performance of image generation. At the end of training, \\(\\theta_{d}\\) and \\(\\theta_{g}\\) are completely different from the beginning. Hence, \\(\\phi_{g}\\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.619, + 0.785, + 0.637 + ], + "angle": 0, + "content": "\\[\nY _ {g} = \\phi_ {g} (\\hat {X} _ {t}, I _ {\\text {m e s h}}, K, t; \\theta_ {g}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.645, + 0.788, + 0.711 + ], + "angle": 0, + "content": "where \\( I_{mesh} = (I_{mesh}^{G}, I_{mesh}^{L}) \\) denotes the concatenation of the global and local mesh image, \\( t \\) denotes the diffusion step obtained by positional encoding, \\( Y_{g} = (Y_{g}^{G}, Y_{g}^{L}) \\in \\mathbb{R}^{2 \\times 4 \\times 64 \\times 64} \\) denotes the concatenation of the global guidance feature \\( Y_{g}^{G} \\) and local guidance feature \\( Y_{g}^{L} \\). Next, \\( \\phi_{d} \\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.72, + 0.785, + 0.737 + ], + "angle": 0, + "content": "\\[\nY _ {d} = \\phi_ {d} \\left(\\hat {X} _ {t}, K, t; \\theta_ {d}\\right) + Y _ {g}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.743, + 0.786, + 0.777 + ], + "angle": 0, + "content": "where \\( Y_{d} = (Y_{d}^{G},Y_{d}^{L})\\in \\mathbb{R}^{2\\times 4\\times 64\\times 64} \\) indicates the concatenation of the global diffusion feature \\( Y_{d}^{G} \\) and local diffusion feature \\( Y_{d}^{L} \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Optimization. Since the diffusion model typically involves both forward and reverse processes, our AttentionHand also employs two processes. For the forward process, the noisy embedding \\( X_{t} = (X_{t}^{G},X_{t}^{L}) \\) is obtained by progressively" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "AttentionHand" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.292 + ], + "angle": 0, + "content": "perturbing Gaussian noise \\(\\epsilon = (\\epsilon^{G},\\epsilon^{L})\\) to the initial embedding \\(X_0 = (X_0^G,X_0^L)\\) by \\(t\\) diffusion steps. \\(\\epsilon^G\\) and \\(\\epsilon^L\\) denote the global and local noise added to \\(X_0^G\\) and \\(X_0^L\\), respectively. Then, since \\(X_{t}\\) is updated to \\(\\hat{X}_t\\) by TAS, \\(\\epsilon\\) is also updated to \\(\\hat{\\epsilon} = (\\hat{\\epsilon}^{G},\\hat{\\epsilon}^{L})\\). In other words, \\(\\hat{\\epsilon}\\) is considered the residual noise between \\(X_0\\) and \\(\\hat{X}_t\\). Thus, \\(\\epsilon^G\\) and \\(\\epsilon^L\\) denote the global and local residual noise. For the reverse process, AttentionHand learns to gradually remove residual noises with global and local denoising processes. Therefore, given text embedding \\(K\\), diffusion steps \\(t\\), and mesh images \\(I_{mesh}^{G}\\) and \\(I_{mesh}^{L}\\), the diffusion training network \\(\\epsilon_{\\theta}\\) is optimized to predict \\(\\hat{\\epsilon}^{G}\\) and \\(\\hat{\\epsilon}^{L}\\) jointly through the following objectives:" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.306, + 0.786, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {G} = \\mathbb {E} _ {X _ {0} ^ {G}, I _ {\\text {m e s h}} ^ {G}, K, t, \\hat {\\epsilon^ {G}} \\sim \\mathcal {N} (0, 1)} [ \\| \\hat {\\epsilon^ {G}} - \\epsilon_ {\\theta} (\\hat {X} _ {t} ^ {G}, I _ {\\text {m e s h}} ^ {G}, K, t) \\| _ {2} ^ {2} ], \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.358, + 0.786, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {L} = \\mathbb {E} _ {X _ {0} ^ {L}, I _ {\\text {m e s h}} ^ {L}, K, t, \\hat {\\epsilon} ^ {L} \\sim \\mathcal {N} (0, 1)} [ \\| \\hat {\\epsilon} ^ {L} - \\epsilon_ {\\theta} (\\hat {X} _ {t} ^ {L}, I _ {\\text {m e s h}} ^ {L}, K, t) \\| _ {2} ^ {2} ], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.784, + 0.423 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}^G\\) and \\(\\mathcal{L}^L\\) indicate the cost function of global and local features, respectively. Thus, the final objective is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.437, + 0.785, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda^ {G} \\mathcal {L} ^ {G} + \\lambda^ {L} \\mathcal {L} ^ {L}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.468, + 0.631, + 0.485 + ], + "angle": 0, + "content": "where \\(\\lambda^G\\) and \\(\\lambda^L\\) are weighted coefficients of \\(\\mathcal{L}^G\\) and \\(\\mathcal{L}^L\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.516, + 0.396, + 0.532 + ], + "angle": 0, + "content": "3.4 Decoding Phase" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.796, + 0.643 + ], + "angle": 0, + "content": "In the decoding phase, we can generate a new RGB hand image \\(\\hat{I}_{RGB} \\in \\mathbb{R}^{3 \\times 512 \\times 512}\\) by passing \\(Y_{d}\\) through the decoder \\(\\mathcal{D}\\), as shown in the fourth box of Fig. 3. While \\(\\mathcal{E}\\) encodes \\(X_{0}\\) by downsampling \\(I_{RGB}\\) in the latent space, \\(\\mathcal{D}\\) decodes \\(\\hat{I}_{RGB}\\) by upsampling \\(Y_{d}\\) in the pixel space, conditioned to given text prompt and mesh images. The structure of \\(\\mathcal{D}\\) is similar to the decoder of VQ-GAN. The decoding phase is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.657, + 0.785, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {R G B} = \\mathcal {D} \\left(Y _ {d}\\right). \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.705, + 0.376, + 0.723 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.746, + 0.336, + 0.76 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "For the text-to-image generation, we adopted MSCOCO [6]. For the 3D hand mesh reconstruction, we adopted Hands-In-Action (HIC) [37], Re:InterHand (ReIH) [5], InterHand2.6M (IH2.6M) [7], and MSCOCO. Due to the page limit, details will be explained in the supplementary materials." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.172 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons with state-of-the-art text-to-image generation models." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.186, + 0.728, + 0.253 + ], + "angle": 0, + "content": "
MethodsFID↓KID↓FID-H↓KID-H↓Hand Conf.↑MSE-2D↓MSE-3D↓User Pref.(%)↑
Stable Diffusion [17]40.520.0068450.780.025540.6512.9324.5915.864
Uni-ControlNet [20]30.340.0074437.770.020040.8552.1053.0398.796
T2I-Adapter [19]22.000.0076132.080.015680.9141.5462.45119.676
ControlNet [18]21.670.0065840.320.020980.8101.2522.1827.948
AttentionHand (w/o TAS)21.270.0033128.560.013900.9551.2112.04220.734
AttentionHand (w/TAS)20.710.0030127.090.012870.9651.0261.98636.905
" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.264, + 0.784, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.607 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparisons with state-of-the-art text-to-image generation models. Red and green boxes in each sample indicate the wrong and current hand bounding box, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.427, + 0.648 + ], + "angle": 0, + "content": "4.2 Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.766 + ], + "angle": 0, + "content": "For the text-to-image generation, we adopted FID [38], FID-Hand (FID-H), KID [39], KID-Hand (KID-H), the hand confidence score (Hand Conf.) [40], the mean square error of 2D and 3D keypoints (MSE-2D, 3D), and the user preference (User Pref.). For the 3D hand mesh reconstruction, we adopted the mean pervertex position error (MPVPE), the right hand-relative vertex error (RRVE), and the mean relative-root position error (MRRPE). Due to the page limit, details will be explained in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.553, + 0.801 + ], + "angle": 0, + "content": "4.3 Comparisons with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Text-to-Image Generation. As shown in Table 1, our AttentionHand exhibited the highest performance in all metrics among state-of-the-arts [17-20]. This" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AttentionHand" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons with state-of-the-art 3D hand mesh reconstruction methods with and without AttentionHand. The red subscripts indicate the difference in performance with and without AttentionHand." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.2, + 0.756, + 0.309 + ], + "angle": 0, + "content": "
DatasetsMethodsIn-the-wildIn-the-lab
HIC [37]ReIH [5]IH2.6M [7]
MPVPE↓RRVE↓MRRPE↓MPVPE↓RRVE↓MRRPE↓MPVPE↓RRVE↓MRRPE↓
IHMR [8]38.5745.51119.6430.9045.5598.4516.9421.9833.39
IHMR+AttentionHand36.73-1.8444.10-1.4194.63-25.0129.11-1.7943.12-2.4387.07-11.3815.09-1.8520.55-1.4332.21-1.18
InterShape [9]27.6634.69110.2527.8738.5680.0412.9717.3531.56
InterShape+AttentionHand25.04-2.6233.33-1.3680.17-30.0826.44-1.4336.54-2.0261.41-18.6311.90-1.0716.22-1.1330.04-1.52
IntagHand [10]23.0728.7452.4625.9030.0542.2212.3417.3229.31
IntagHand+AttentionHand21.87-1.2027.09-1.6547.11-5.3523.39-2.5128.77-1.2833.98-8.2411.42-0.9215.81-1.5129.18-0.13
DIR [13]21.8926.1143.1121.8229.6637.0110.2617.1128.98
DIR+AttentionHand20.66-1.2325.87-0.2440.54-2.5719.91-1.9126.67-2.9935.05-1.9610.09-0.1716.99-0.1228.02-0.96
InterWild [16]15.3021.3531.2613.9920.0722.3811.5219.7726.87
InterWild+AttentionHand14.74-0.5621.10-0.2529.26-2.0013.95-0.0419.94-0.1322.05-0.3310.62-0.9019.09-0.6825.74-1.13
" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.33, + 0.784, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.594, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparisons on MSCOCO [6]. Red and green boxes indicate wrong and correct region of the reconstructed hand, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.844 + ], + "angle": 0, + "content": "is particularly evident in the comparison of FID(-H) and KID(-H), which signify the quality of the generated images being on par with real RGB images. Furthermore, the lowest MSE-2D and MSE-3D indicates the remarkable alignment between the generated images and the corresponding hand mesh images. With respect to the user preference, AttentionHand scored the highest compared to other methods. It implies that most users acknowledged the outstanding quality of hand images generated by AttentionHand. In addition, as shown in Fig. 5, our AttentionHand generated the high-quality hand image which is well-corresponding with the given mesh image and fully reflected the given text prompt. Specifically, even when two-hands mesh image is given, which is more challenging than in the case of single-hand mesh image, AttentionHand generated the hand image robustly. It implies our AttentionHand is proper to generate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "well-aligned hand images with given mesh images and text prompt. Additional qualitative results are in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.196, + 0.79, + 0.424 + ], + "angle": 0, + "content": "3D Hand Mesh Reconstruction. To verify our AttentionHand extensively, we trained state-of-the-art hand pose networks [8-10, 13, 16] by additionally adding new data generated by AttentionHand. As shown in Table 2, the performance of all methods increased for all metrics. Specifically, with respect to the MPVPE, AttentionHand showed the dramatic performance improvement with InterWild [16] about \\(3.66\\%\\) and \\(7.81\\%\\) on HIC and ReIH, respectively. With respect to the RRVE, it increased by about \\(1.17\\%\\) and \\(0.65\\%\\) on HIC and ReIH, respectively. With respect to the MRRPE, it increased by about \\(6.40\\%\\) and \\(1.47\\%\\) on HIC and ReIH, respectively. These imply generated hand images help increasing the accuracy of the 3D hand mesh reconstruction. In addition, the qualitative performance for in-the-wild scenes is also verified as shown in Fig. 6. Although MSCOCO mainly contains in-the-wild situations, 3D hand mesh is reconstructed robustly. It implies that even for difficult situations, the performance of reconstruction can be improved by utilizing AttentionHand. Additional qualitative results are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.441, + 0.403, + 0.456 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Text Attention Stage (TAS). We deeply dived into TAS to verify its superiority. Firstly, as in the last two rows in Table 1, TAS showed its effectiveness in all metrics. In addition, as shown in Fig. 7, attention maps are well described their corresponding tokens in the case of with TAS. It implies that with TAS, AttentionHand can reflect hand-related tokens enough. Additional qualitative results are in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Secondly, we conducted more experiments about Gaussian filters as follows: (1) no Gaussian filter, (2) random Gaussian filter, and (3) fixed Gaussian filter. As shown in the second, third, and fourth columns of Fig. 8, we found interesting results: in the case of (1), the hand was disappeared or its shape became strange. In the case of (2), generated images are not well-aligned with given hand mesh images. However, in the case of (3), generated images are well-aligned with given hand mesh images and look natural. Hence, we determined fixed Gaussian filter makes the generated image plausibly regardless of diffusion timestep \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Thirdly, we compared our loss, \\(\\mathcal{L}^{T\\hat{A}S}\\), with the load balancing loss \\((\\mathcal{L}^{LB})\\) [41, 42]. Since \\(\\mathcal{L}^{LB}\\) is an auxiliary loss for balancing loads among experts, it plays a similar role with \\(\\mathcal{L}^{TAS}\\), which evenly reflects the image features of all the attention maps. Therefore, we replaced \\(\\mathcal{L}^{TAS}\\) to \\(\\mathcal{L}^{LB}\\) and considered its feasibility as shown in the fifth column of Fig. 8. Unfortunately, in the case of \\(\\mathcal{L}^{LB}\\), generated images are not fit at all with given hand mesh images. We guess while \\(\\mathcal{L}^{TAS}\\) updates the image embedding based on the spatial information of the attention map, \\(\\mathcal{L}^{LB}\\) flattens the 2D attention map as 1D representation, leading to distort spatial knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Last but not least, we explored the range of updated noise \\((\\hat{\\epsilon})\\). According to [43], we set \\(\\alpha_{t}\\) of Eq. 4 as gradually decreasing according to timestep \\(t\\) (i.e.," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AttentionHand" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.756, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.24, + 0.788, + 0.281 + ], + "angle": 0, + "content": "Fig. 7: Ablation studies on the text attention stage (TAS). Attention maps with red and green box are results without and with TAS, respectively. Red and green bounding boxes indicate wrong and correct hand poses, respectively." + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.306, + 0.7, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.473, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Fig. 8: Ablation studies on Gaussian filter, losses (i.e. \\(\\mathcal{L}^{TAS}\\) and \\(\\mathcal{L}^{LB}\\)), and the regularization of \\(\\hat{\\epsilon}\\) for the text attention stage (TAS). Red and green bounding boxes indicate wrong and correct hand poses, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.631 + ], + "angle": 0, + "content": "from 20 to 10) for regularization of \\(\\hat{\\epsilon}\\). However, if \\(\\alpha_{t}\\) is randomly set, \\(\\hat{\\epsilon}\\) tends to be out of distribution (i.e., Gaussian distribution) as shown in the sixth column of Fig. 8: in the case of w/o regularization, generated images are not aligned with given mesh images, or missed some hands. Therefore, it is necessary to regularize \\(\\hat{\\epsilon}\\) for faithful hand image generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Model Design Justification. To justify our model's superiority, we compared the characteristics of prior works including our model. As shown in Table 3, our model's distinctive and potential features compared to prior works are (1) harmonious preservation of locality (i.e., hand) with globality (i.e., in-the-wild scene), and (2) selective attention on hand-related tokens by cross attention. Specifically, to harmonize globality and locality, we developed global and local designs for the visual attention stage (VAS). Moreover, since the global and local branches are designed structurally same, we set them to share their weights for reducing the number of training parameters (about \\(20.2\\% \\downarrow\\)) and improving the generalizability (see two shaded rows in Table 4). We experimentally verified the effectiveness of our design as shown in Table 4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.336, + 0.145, + 0.666, + 0.16 + ], + "angle": 0, + "content": "Table 3: Network comparisons with prior works." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.172, + 0.699, + 0.233 + ], + "angle": 0, + "content": "
MethodsText PromptVisual PromptLocalityHand-related Token Attention
Stable Diffusion✓
Uni-ControlNet✓✓
T2I-Adapter✓✓
ControlNet✓✓
AttentionHand (Ours)✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.246, + 0.712, + 0.261 + ], + "angle": 0, + "content": "Table 4: Ablation studies on the visual attention stage (VAS)." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.274, + 0.7, + 0.333 + ], + "angle": 0, + "content": "
GlobalityLocalityWeightsFID↓KID↓FID-H↓KID-H↓Hand Conf.↑MSE-2D↓MSE-3D↓
Shared40.520.0068450.780.025540.6512.9324.591
✓Shared21.670.0065840.320.020980.8101.2522.182
✓Shared52.980.0071332.110.016040.9111.5392.397
✓✓Shared20.710.0030127.090.012870.9651.0261.986
✓✓Separated21.900.0029326.890.013400.9601.1082.017
" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.347, + 0.755, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Fig. 9: (a) Multiple generated hand images from same modalities. Green boxes indicate correct hand poses. (b) t-SNE distribution of AttentionHand and MSCOCO [6]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.648 + ], + "angle": 0, + "content": "Robustness of Generated Dataset. To verify robustness of our generated dataset, we generated multiple hand images from same modalities as shown in Fig. 9(a). As a result, all generated images are perfectly well-aligned with given hand mesh images. Moreover, we found the t-SNE distribution [44] of AttentionHand is broader than MSCOCO as shown in Fig. 9(b). As a result, we believe that AttentionHand can contribute to the downstream task with our extensive in-the-wild hand images, leading to alleviate the domain gap between indoor and outdoor scenes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.36, + 0.687 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduced a novel text-to-hand image generation model, AttentionHand, which pays attention to the hand-related tokens from the text prompt and global and local mesh images. AttentionHand achieved state-of-the-art performance in text-to-hand image generation, and we demonstrated that training with the dataset generated by our AttentionHand improved the performance of 3D hand mesh reconstruction. However, the diversity may decrease as the generative model is trained to optimize hand mesh images. We expect for the emergence of outstanding diffusion model to improve the diversity and quality of the hand image." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "AttentionHand" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Acknowledgements. This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00260091) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation) and Korea Institute for Advancement of Technology (KIAT) grant funded by the Korea Government (MOTIE) (P0020535, The Competency Development Program for Industry Specialist) and National Supercomputing Center with supercomputing resources including technical support (KSC-2023-CRE-0444)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.289, + 0.325, + 0.306 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Hampali, Shreyas and Rad, Mahdi and Oberweger, Markus and Lepetit, Vincent. Honnotate: A method for 3D annotation of hand and object poses. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3196-3206, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.429 + ], + "angle": 0, + "content": "2. Chao, Yu-Wei and Yang, Wei and Xiang, Yu and Molchanov, Pavlo and Handa, Ankur and Tremblay, Jonathan and Narang, Yashraj S and Van Wyk, Karl and Iqbal, Umar and Birchfield, Stan and Kautz, Jan and Fox, Dieter. DexYCB: A benchmark for capturing hand grasping of objects. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9044–9053, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.483 + ], + "angle": 0, + "content": "3. Ohkawa, Takehiko and He, Kun and Sener, Fadime and Hodan, Tomas and Tran, Luan and Keskin, Cem. Assemblyhands: Towards egocentric activity understanding via 3D hand pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12999-13008, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.51 + ], + "angle": 0, + "content": "4. Lin, Fanqing and Wilhelm, Connor and Martinez, Tony. Two-hand global 3d pose estimation using monocular rgb. pages 2373-2381, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.579 + ], + "angle": 0, + "content": "5. Moon, Gyeongsik and Saito, Shunsuke and Xu, Weipeng and Joshi, Rohan and Buffalini, Julia and Bellan, Harley and Rosen, Nicholas and Richardson, Jesse and Mize, Mallorie and De Bree, Philippe and Simon, Tomas and Peng, Bo and Garg, Shubham and McPhail, Kevyn and Shiratori, Takaaki. A dataset of relighted 3d interacting hands. Adv. Neural Inform. Process. Syst., 36, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.633 + ], + "angle": 0, + "content": "6. Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Dollar, Piotr and Zitnick, C Lawrence. Microsoft coco: Common objects in context. In *Eur. Conf. Comput. Vis.*, pages 740-755, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.688 + ], + "angle": 0, + "content": "7. Moon, Gyeongsik and Yu, Shouu-I and Wen, He and Shiratori, Takaaki and Lee, Kyoung Mu. Interhand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single RGB image. In *Eur. Conf. Comput. Vis.*, pages 548-564, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.731 + ], + "angle": 0, + "content": "8. Rong, Yu and Wang, Jingbo and Liu, Ziwei and Loy, Chen Change. Monocular 3D reconstruction of interacting hands via collision-aware factorized refinements. In 3DV, pages 432-441, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.784 + ], + "angle": 0, + "content": "9. Zhang, Baowen and Wang, Yangang and Deng, Xiaoming and Zhang, Yinda and Tan, Ping and Ma, Cuixia and Wang, Hongan. Interacting two-hand 3D pose and shape reconstruction from single color image. In Int. Conf. Comput. Vis., pages 11354-11363, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.839 + ], + "angle": 0, + "content": "10. Li, Mengcheng and An, Liang and Zhang, Hongwen and Wu, Lianpeng and Chen, Feng and Yu, Tao and Liu, Yebin. Interacting attention graph for single image two-hand reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2761-2770, 2022." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "11. Hampali, Shreyas and Sarkar, Sayan Deb and Rad, Mahdi and Lepetit, Vincent. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3D pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11090-11100, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "12. Meng, Hao and Jin, Sheng and Liu, Wentao and Qian, Chen and Lin, Mengxiang and Ouyang, Wanli and Luo, Ping. 3D interacting hand pose estimation by hand de-occlusion and removal. In Eur. Conf. Comput. Vis., pages 380-397, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.787, + 0.298 + ], + "angle": 0, + "content": "13. Ren, Pengfei and Wen, Chao and Zheng, Xiaozheng and Xue, Zhou and Sun, Haifeng and Qi, Qi and Wang, Jingyu and Liao, Jianxin. Decoupled Iterative Refinement Framework for Interacting Hands Reconstruction from a Single RGB Image. In Int. Conf. Comput. Vis., pages 8014-8025, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "14. Zuo, Binghui and Zhao, Zimeng and Sun, Wenqian and Xie, Wei and Xue, Zhou and Wang, Yangang. Reconstructing interacting hands with interaction prior from monocular images. In Int. Conf. Comput. Vis., pages 9054-9064, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.787, + 0.392 + ], + "angle": 0, + "content": "15. Li, Lijun and Tian, Linrui and Zhang, Xindi and Wang, Qi and Zhang, Bang and Bo, Liefeng and Liu, Mengyuan and Chen, Chen. Renderih: A large-scale synthetic dataset for 3d interacting hand pose estimation. In Int. Conf. Comput. Vis., pages 20395-20405, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.392, + 0.787, + 0.433 + ], + "angle": 0, + "content": "16. Moon, Gyeongsik. Bringing inputs to shared domains for 3D interacting hands recovery in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 17028-17037, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "17. Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Björn. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.473, + 0.787, + 0.513 + ], + "angle": 0, + "content": "18. Zhang, Lvmin and Rao, Anyi and Agrawala, Maneesh. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., pages 3836-3847, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.513, + 0.787, + 0.568 + ], + "angle": 0, + "content": "19. Mou, Chong and Wang, Xintao and Xie, Liangbin and Zhang, Jian and Qi, Zhonggang and Shan, Ying and Qie, Xiaohu. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.568, + 0.787, + 0.609 + ], + "angle": 0, + "content": "20. Zhao, Shihao and Chen, Dongdong and Chen, Yen-Chun and Bao, Jianmin and Hao, Shaozhe and Yuan, Lu and Wong, Kwan-Yee K. Uni-ControlNet: All-in-one control to text-to-image diffusion models. Adv. Neural Inform. Process. Syst., 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.609, + 0.787, + 0.663 + ], + "angle": 0, + "content": "21. Podell, Dustin and English, Zion and Lacey, Kyle and Blattmann, Andreas and Dockhorn, Tim and Müller, Jonas and Penna, Joe and Rombach, Robin. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.663, + 0.787, + 0.718 + ], + "angle": 0, + "content": "22. Mueller, Franziska and Bernard, Florian and Sotnychenko, Oleksandr and Mehta, Dushyant and Sridhar, Srinath and Casas, Dan and Theobalt, Christian. GANerated hands for real-time 3D hand tracking from monocular RGB. In IEEE Conf. Comput. Vis. Pattern Recog., pages 49-59, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "23. Tang, Hao and Wang, Wei and Xu, Dan and Yan, Yan and Sebe, Nicu. GestureGAN for hand gesture-to-gesture translation in the wild. In ACM Int. Conf. Multimedia, pages 774-782, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "24. Hu, Hezhen and Wang, Weilun and Zhou, Wengang and Zhao, Weichao and Li, Houqiang. Model-aware gesture-to-gesture translation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16428-16437, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.839 + ], + "angle": 0, + "content": "25. Hu, Hezhen and Wang, Weilun and Zhou, Wengang and Li, Houqiang. Hand-object interaction image generation. Adv. Neural Inform. Process. Syst., 35:23805-23817, 2022." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AttentionHand" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.202 + ], + "angle": 0, + "content": "26. Goodfellow, Ian and Pouget-Abadie, Jean and Mirza, Mehdi and Xu, Bing and Warde-Farley, David and Ozair, Sherjil and Courville, Aaron and Bengio, Yoshua. Generative adversarial nets. Adv. Neural Inform. Process. Syst., 27:2672-2680, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.242 + ], + "angle": 0, + "content": "27. Li, Lijun and Zhuo, Li'an and Zhang, Bang and Bo, Liefeng and Chen, Chen. Diff-Hand: End-to-end hand mesh reconstruction via diffusion models. arXiv preprint arXiv:2305.13705, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.243, + 0.787, + 0.296 + ], + "angle": 0, + "content": "28. Lin, Pei and Xu, Sihang and Yang, Hongdi and Liu, Yiran and Chen, Xin and Wang, Jingya and Yu, Jingyi and Xu, Lan. HandDiffuse: Generative controllers for two-hand interactions via diffusion models. arXiv preprint arXiv:2312.04867, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "29. Lu, Wenquan and Xu, Yufei and Zhang, Jing and Wang, Chaoyue and Tao, Dacheng. HandRefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. arXiv preprint arXiv:2311.17957, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.339, + 0.787, + 0.379 + ], + "angle": 0, + "content": "30. Esser, Patrick and Rombach, Robin and Ommer, Bjorn. Taming transformers for high-resolution image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12873-12883, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.379, + 0.787, + 0.447 + ], + "angle": 0, + "content": "31. Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., pages 8748-8763, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.448, + 0.787, + 0.488 + ], + "angle": 0, + "content": "32. Chen, Chun-Fu Richard and Fan, Quanfu and Panda, Rameswar. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Int. Conf. Comput. Vis., pages 357-366, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.488, + 0.787, + 0.514 + ], + "angle": 0, + "content": "33. Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.515, + 0.787, + 0.554 + ], + "angle": 0, + "content": "34. Chiche, Alebachew and Yitagesu, Betselot. Part of speech tagging: a systematic review of deep learning and machine learning approaches. Journal of Big Data, 9(1):1-25, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.555, + 0.787, + 0.595 + ], + "angle": 0, + "content": "35. He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.596, + 0.787, + 0.663 + ], + "angle": 0, + "content": "36. Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.664, + 0.787, + 0.717 + ], + "angle": 0, + "content": "37. Tzionas, Dimitrios and Ballan, Luca and Srikantha, Abhilash and Aponte, Pablo and Pollefeys, Marc and Gall, Juergen. Capturing hands in action using discriminative salient points and physics simulation. Int. J. Comput. Vis., 118:172-193, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "38. Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Adv. Neural Inform. Process. Syst., 30:6626-6637, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.759, + 0.787, + 0.785 + ], + "angle": 0, + "content": "39. Binkowski, Mikolaj and Sutherland, Danica J and Arbel, Michael and Gretton, Arthur. Demystifying MMD GANs. In Int. Conf. Learn. Represent., 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "40. Narasimhaswamy, Supreeth and Bhattacharya, Uttaran and Chen, Xiang and Dasgupta, Ishita and Mitra, Saayan and Hoai, Minh. Handiffuser: Text-to-image generation with realistic hand appearances. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2468-2479, 2024." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "41. Zhou, Yanqi and Lei, Tao and Liu, Hanxiao and Du, Nan and Huang, Yanping and Zhao, Vincent and Dai, Andrew M and Le, Quoc V and Laudon, James. Mixture-of-experts with expert choice routing. Adv. Neural Inform. Process. Syst., 35:7103-7114, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "42. Fedus, William and Zoph, Barret and Shazeer, Noam. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 23(120):1-39, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "43. Chefer, Hila and Alaluf, Yuval and Vinker, Yael and Wolf, Lior and Cohen-Or, Daniel. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans. Graph., 42(4):1-10, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.315 + ], + "angle": 0, + "content": "44. Van der Maaten, Laurens and Hinton, Geoffrey. Visualizing data using t-SNE. J. Mach. Learn. Res., 9(86):2579-2605, 2008." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.315 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_origin.pdf b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f0cf88e2167cf83d3c249afa491f4807957836f --- /dev/null +++ b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/8e6d7b99-1133-4831-a79f-a95e2ffe1180_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83b6132bc2688d91bd9f106eda242a5d29573c013a08d64811e97b9018af2553 +size 9584043 diff --git a/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/full.md b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f217c94fda2d058662ff8c8f1c133eb8ef4d368c --- /dev/null +++ b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/full.md @@ -0,0 +1,287 @@ +# AttentionHand: Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild + +Junho Park $^{1,2*}$ , Kyeongbo Kong $^{3*}$ , and Suk-Ju Kang $^{1*}$ + +$^{1}$ Department of Electronic Engineering, Sogang University, South Korea $^{2}$ AI Lab, CTO Division, LG Electronics, South Korea + +$^{3}$ Department of Electrical & Electronics Engineering, Pusan National University, South Korea + +junho18.park@gmail.com kbkong@pusan.ac.kr sj Kang@sogang.ac.kr https://github.com/redorangeyellow/AttentionHand + +Abstract. Recently, there has been a significant amount of research conducted on 3D hand reconstruction to use various forms of human-computer interaction. However, 3D hand reconstruction in the wild is challenging due to extreme lack of in-the-wild 3D hand datasets. Especially, when hands are in complex pose such as interacting hands, the problems like appearance similarity, self-handed occlusion and depth ambiguity make it more difficult. To overcome these issues, we propose AttentionHand, a novel method for text-driven controllable hand image generation. Since AttentionHand can generate various and numerous in-the-wild hand images well-aligned with 3D hand label, we can acquire a new 3D hand dataset, and can relieve the domain gap between indoor and outdoor scenes. Our method needs easy-to-use four modalities (i.e., an RGB image, a hand mesh image from 3D label, a bounding box, and a text prompt). These modalities are embedded into the latent space by the encoding phase. Then, through the text attention stage, hand-related tokens from the given text prompt are attended to highlight hand-related regions of the latent embedding. After the highlighted embedding is fed to the visual attention stage, hand-related regions in the embedding are attended by conditioning global and local hand mesh images with the diffusion-based pipeline. In the decoding phase, the final feature is decoded to new hand images, which are well-aligned with the given hand mesh image and text prompt. As a result, AttentionHand achieved state-of-the-art among text-to-hand image generation models, and the performance of 3D hand mesh reconstruction was improved by additionally training with hand images generated by AttentionHand. + +Keywords: 3D Hand Mesh Reconstruction $\cdot$ Text-to-Image Generation + +![](images/b5c47883ff948277ba3d6cbf11820aa8d952ad9d23eb7b1625f6f6c0da0dd07c.jpg) +(a) + +![](images/12d027893901489413cf8e86ba89e989cd97aa9c5d7ce4f1e9749609d3cf8f22.jpg) + +![](images/704b89104c4cd11d2237f0c636f67bf6e102e61653250703e2a05833d4d0ea2d.jpg) +(b) +(c) +Fig. 1: Various acquisition types of 3D hand datasets. (a) In-the-wild dataset (i.e., MSCOCO [6]) is naively acquired with inaccurate pseudo annotation, (b) relighted dataset (i.e., Re:InterHand [5]) consists of unnatural hands with inharmonious background, and (c) our in-the-wild dataset from AttentionHand, which is annotated with accurate 3D labels, contains natural hands with harmonious background, easy to generate, and can be made infinitely. + +# 1 Introduction + +The goal of 3D hand mesh reconstruction is to recover the 3D hand mesh from a single RGB image. It becomes difficult when hands are in the wild, due to insufficiency of in-the-wild 3D hand datasets. Compared to in-the-lab datasets [1-3], acquisition in-the-wild datasets is challenging due to unpredictable conditions such as weather, lighting, cost of sensors, and safety issues on crowded roads and public places. Even if an in-the-wild dataset is collected, data diversity would be poor due to the aforementioned severe constraints. Although arbitrary labels can be obtained through pseudo annotation, the precision and accuracy is still poor compared to in-the-lab datasets as shown in Fig. 1(a). To tackle this problem, several synthetic datasets [4, 5] have introduced. However, since the hand and background images are synthesized out of harmony, they consist of unnatural and unrealistic hand images as shown in Fig. 1(b). Hence, it is difficult to overcome the domain gap between indoor and outdoor scenes with synthetic datasets. + +Moreover, when hands are in a complex pose like interacting hands, it becomes even more challenging to reconstruct 3D hand meshes due to the appearance similarity, self-handed occlusion and depth ambiguity. Starting with InterHand2.6M [7], several works [8-15] have emerged to solve the complex hand + +![](images/427ef0e5947b2e7350ebc2c61e31c19ce434208d0e894781630774db6a60b58c.jpg) +Fig. 2: Visualization of attention maps with corresponding tokens from given text prompts. Red and green boxes represent attention maps without and with Attention-Hand, respectively. + +pose. However, they have been employed and evaluated primarily on in-the-lab scenes except for InterWild [16]; it tried to relieve the domain gap by leveraging the geometric features of the hand, which is not affected by the domain. Nevertheless, since InterWild was trained with MSCOCO [6], which is extremely lack of in-the-wild hand images with inaccurate 3D labels, there is earnestly need of more diverse, numerous, and accurately annotated in-the-wild datasets. + +To address aforementioned issues, we propose AttentionHand, a new method for the text-driven controllable hand image generation. AttentionHand is designed based on Stable Diffusion (SD) [17] to create accurate, natural, realistic and harmonious in-the-wild hand images easily and infinitely as shown in Fig. 1(c). AttentionHand has a huge advantage: we can simply generate images with only four modalities - an RGB image, the corresponding hand mesh image, bounding box, and text prompt. Therefore, we can generate (1) various in-the-wild hand images with flexible text prompts, and (2) well-aligned hand images with 3D hand label. By generating new samples with AttentionHand, we can alleviate aforementioned issues of the 3D hand mesh reconstruction in the wild. + +To train AttentionHand, we need to additionally prepare a local RGB image and local mesh image for attention of hand-focused region of the image. The preparation of local information is essential because hands commonly occupy relatively small region in the image. Hence, we obtain local RGB and mesh images by cropping and resizing original RGB and mesh images (i.e., we define them as global information) with the bounding box of hand region. After encoding prepared information in the encoding phase, encoded latent embeddings are fed to the conditioning phase, which is composed of the text attention stage (TAS) and visual attention stage (VAS). + +TAS attends on hand-related tokens from the given text prompt by leveraging attention maps as shown in Fig. 2. Specifically, TAS extracts hand-related attention maps (i.e., holding and hand), and these attention maps are updated to highlight hand-related regions by the refinement based on the softmax operation and Gaussian filter. With TAS, we can obtain more hand-focused images than before. On the other hand, VAS attends on hand-related regions by conditioning global and local hand mesh images with the SD-based pipeline. With global and local information, AttentionHand can be jointly optimized to reflect the global + +context (i.e., in-the-wild background) and local context (i.e., hand-focused foreground.) In the end of the conditioning phase, we finally get the diffusion feature, which is decoded to new hand images in the decoding phase. Hence, Attention-Hand can generate well-aligned hand images with the given mesh image and text prompt for the 3D hand mesh reconstruction in the wild. + +To prove the excellence of AttentionHand, we conducted extensive experiments for the text-to-hand image generation and 3D hand mesh reconstruction. As a result, AttentionHand achieved state-of-the-art in the text-to-hand image generation, and the performance of 3D hand mesh reconstruction was considerably improved by additionally training with new hand samples generated by AttentionHand. Especially, the performance was enhanced significantly on in-the-wild datasets, which implies AttentionHand can generate various and well-annotated in-the-wild hand images. The summary of our contributions is as follows: + +- We propose a novel method, AttentionHand, which generates well-aligned in-the-wild hand images in a simple manner without laborious data acquisition. +- AttentionHand is designed based on a generative model that attends on hand-related tokens from the text prompt and hand-related regions from the hand mesh image, for generating hand-focused images. +- AttentionHand achieved state-of-the-art in the text-to-hand image generation, and we verified that utilizing the dataset generated from AttentionHand improves the performance on 3D hand mesh reconstruction in the wild. + +# 2 Related Work + +# 2.1 Text-to-Image Generation + +Text-to-image generation aims to synthesize high-resolution image from natural language descriptions. With the advent of diffusion models, various studies on text-to-image generation have been conducted in recent years [17-21]. Specifically, ControlNet [18] and T2I-Adapter [19] proposed novel approaches to incorporate arbitrary condition into the generation process. Recently, UniControlNet [20] presented a novel approach that allows for the simultaneous utilization of various conditions in a flexible and composable manner. Nevertheless, aforementioned models exhibited common limitations in generating hand images, due to the relatively small size of hands within the overall image resolution. + +# 2.2 Generative Models for Hand + +GANs for Hand. There are several works [22-25] to tackle the hand image generation problem with the generative adversarial network (GAN) [26]. Specifically, a novel network for image-to-image translation [22] was proposed to make generated images follow the same statistical distribution as real-world hand images. GestureGAN [23] was designed to translate hand gesture-to-gesture with + +the explicit hand skeleton information through the color loss and the cycle-consistency loss. Moreover, the first model-aware gesture-to-gesture translation framework [24] was introduced with hand prior as the intermediate representation. Recently, a new method [25], which employs the expressive model-aware hand-object representation and leverages its inherent topology to build the unified surface space, was proposed. However, these works have a common limitation; they are confined to target gestures in generating new hand images. In other words, they are inappropriate to generate in-the-wild images focused on various hands. + +Diffusion Models for Hand. Recently, some works [27-29] have been addressed hand-related problems with diffusion models. DiffHand [27] introduced the first diffusion-based framework that approaches hand mesh reconstruction as a denoising diffusion process. HandDiffuse [28] proposed a strong baseline for the controllable motion generation of interacting hands using various controllers by designing a diffusion-based model. HandRefiner [29] presented an inpainting pipeline to rectify malformed human hands in generated images with diffusion-based models. However, since these models are not text-driven methods, they cannot generate various in-the-wild hand images conditioned on language instructions. + +# 3 Method + +We introduce AttentionHand, a novel framework for creating various and plausible hand images. AttentionHand is a SD-based framework that can generate new RGB images infinitely conditioned on hand mesh images and text prompts. The overall pipeline is shown in Fig. 3. + +# 3.1 Data Preparation Phase + +As shown in the first box of Fig. 3, it just requires four inputs to train AttentionHand: (1) a global RGB hand image $I_{RGB}^{G} \in \mathbb{R}^{3 \times 512 \times 512}$ , (2) a global hand mesh image $I_{mesh}^{G} \in \mathbb{R}^{3 \times 512 \times 512}$ , (3) a bounding box of the hand region $B \in \mathbb{R}^{1 \times 4}$ , and (4) a hand-related text prompt $U$ . However, since hands typically occupy small areas on in-the-wild scenes, we also obtain a local RGB hand image $I_{RGB}^{L} \in \mathbb{R}^{3 \times 512 \times 512}$ and a local hand mesh image $I_{mesh}^{L} \in \mathbb{R}^{3 \times 512 \times 512}$ by cropping and resizing $I_{RGB}^{G}$ and $I_{mesh}^{G}$ with $B$ . This combination of local and global information enhances hand image conditioning. Details will be explained in the supplementary materials. + +# 3.2 Encoding Phase + +For the diffusion process in latent space, encoding phase for $I_{RGB}^{G}$ , $I_{RGB}^{L}$ and $U$ is implemented by the encoder $\mathcal{E}$ . It makes global and local latent image embeddings $X_0^G, X_0^L \in \mathbb{R}^{4\times 64\times 64}$ for $I_{RGB}^{G}$ and $I_{RGB}^{L}$ , and a latent text embedding + +![](images/1a29bab4c2648fe0b78e25c48f45afa72b9c5c110b4b752b48b5c585ded6690c.jpg) +Fig. 3: Overall pipeline of AttentionHand. In the data preparation phase, we prepare global and local RGB images, global and local hand mesh images, bounding box, and text prompt. In the encoding phase, we get global and local latent image embeddings through VQ-GAN [30], and text embedding through CLIP [31]. In the conditioning phase, we refine image embeddings through the text attention stage, and obtain the diffusion feature through the visual attention stage. In the decoding phase, we generate a new hand image $\hat{I}_{RGB}$ from $Y_{d}$ through VQ-GAN. + +$K \in \mathbb{R}^{77 \times 768}$ for $U$ . Specifically, $X_0^G$ and $X_0^L$ are obtained by VQ-GAN [30], and $K$ is obtained by CLIP [31] as shown in the second box of Fig. 3. These latent embeddings are fed as inputs to the conditioning phase, which will be introduced by the next subsection. The encoding phase is expressed as follows: + +$$ +X _ {0} ^ {G}, X _ {0} ^ {L}, K = \mathcal {E} \left(I _ {R G B} ^ {G}, I _ {R G B} ^ {L}, U\right). \tag {1} +$$ + +# 3.3 Conditioning Phase + +For generating new hand images conditioned by given text prompt and mesh images, we design the text attention stage (TAS) and the visual attention stage (VAS) in the conditioning phase, as shown in the third box of Fig. 3. TAS is a stage of paying attention to tokens for the hand and its corresponding gesture in a given text. VAS is a stage of training the SD-based model specialized for hand image generation by conditioning global and local mesh images. + +Text Attention Stage (TAS). TAS is a stage of attending tokens which represent hand or gestures in a given text prompt as shown in Fig. 4(a). First, by adding Gaussian noise to $X_0^G$ and $X_0^L$ with $t$ diffusion steps, the global noisy embedding $X_t^G$ and local noisy embedding $X_t^L$ are obtained. For simplicity, we define as $X_0 = (X_0^G, X_0^L)$ and $X_t = (X_t^G, X_t^L)$ . Then, $X_t$ and $K$ are fed to TAS as inputs. For the text attention of TAS, we utilize the cross attention [32]. Specifically, an attention map $A \in \mathbb{R}^{H \times W \times N}$ is obtained by calculating the key (i.e., $K$ ) and query (i.e., $Q$ , which is the linear projection of intermediate feature + +![](images/28a6c76d0d51a82ec13193fa5912a81df0e28166b724a7c10d6fc70384cca3c4.jpg) +Fig. 4: (a) Overall process of the text attention stage (TAS). By leveraging the hand-related tagging and refinement, we can highlight hand-related attention maps, which leads to update noisy embeddings with $\mathcal{L}^{TAS}$ . (b) Overall process of the visual attention stage (VAS). By utilizing the global and local information, we can obtain the harmonious diffusion feature, which leads to generate high-fidelity hand images. + +![](images/03fe6f18fc5147b5951beb4fe431389fe84841795d89690833447bc8258e6e54.jpg) + +map from $X_{t}$ of U-Net [33] in SD). $H$ and $W$ denote the height and width of $A$ , and $N$ denotes the number of all tokens of $K$ . + +Next, to extract hand-related attention maps $A_{k\in K}\in \mathbb{R}^{H\times W\times N_k}$ in $A$ , we design the hand-related tagging $\mathcal{H}_{tag}$ , which is based on part-of-speech tagging [34], where $N_{k}$ denotes the number of hand-related tokens in $K$ . Specifically, $\mathcal{H}_{tag}$ determines if the input token indicates the hand-related word (i.e., holding, taking, or hand). With $\mathcal{H}_{tag}$ , we can attend hand-related tokens $k$ to generate more hand-focused images. More details are in the supplementary materials. + +Then, we employ the softmax operation and Gaussian smoothing to maximize the effect of $A_{k}$ . Since the Gaussian filter effectively removes noise from images and preserves detailed information by using the average value of surrounding pixels, we fully exploit these advantages. Hence, $A_{k}$ is updated to $\hat{A}_k\in \mathbb{R}^{H\times W\times N_k}$ by refining hand-related attention maps as follows: + +$$ +\hat {A} _ {k} = \text {G a u s s i a n} (\operatorname {S o f t m a x} (A _ {k})). \tag {2} +$$ + +For simplicity, we define $\hat{A} \in \mathbb{R}^{H \times W \times N}$ as the concatenation of $\hat{A}_{k \in K}$ and $A_{l \notin K} \in \mathbb{R}^{H \times W \times N_l}$ , where $N_l$ denotes the number of not hand-related tokens in $K$ , and $N = N_k + N_l$ . + +Moreover, optimization for evenly reflecting the image features of all attention maps is necessary. In other words, it is required to design an objective to prevent poor generation of the image feature for a specific token. Specifically, for arbitrary token $n \in K$ , the highest value $s_n$ among all patches in the $n$ -th refined attention map $\hat{A}_n$ is extracted, and it is subtracted from 1. This operation is implemented for all tokens in $K$ , and a novel loss, which named $\mathcal{L}^{TAS}$ , is computed the largest value among them as follows: + +$$ +\mathcal {L} ^ {T A S} = \max _ {n \in K} \left(1 - s _ {n}\right). \tag {3} +$$ + +Based on $\mathcal{L}^{TAS}$ , $X_{t}$ is updated to $\hat{X}_{t} = (\hat{X}_{t}^{G},\hat{X}_{t}^{L})$ as follows: + +$$ +\hat {X} _ {t} = X _ {t} - \alpha_ {t} \nabla_ {X _ {t}} \mathcal {L} ^ {T A S}, \tag {4} +$$ + +where $\alpha_{t}$ indicates the learning rate, and $\hat{X}_t^G$ and $\hat{X}_t^L$ indicate updated global and local noisy embedding, respectively. + +Visual Attention Stage (VAS). VAS is a stage of training SD-based model by conditioning the aforementioned global and local mesh image. VAS is composed of two modules: one is the guidance module $\phi_{g}$ , and the other is the diffusion module $\phi_{d}$ as shown in Fig. 4(b). First, the diffusion module $\phi_{d}$ is designed based on an U-Net network, consisting of 25 blocks: 8 blocks are downsampling and upsampling convolution layers, and the remaining 17 blocks consist of four ResNet [35] layers and two Vision Transformers [36]. We define the parameter set of $\phi_{d}$ as $\theta_{d}$ , which is fixed frozen to maintain the image generation performance of SD. + +On the other hand, the guidance module $\phi_{g}$ is also based on an U-Net network with 25 blocks of $\phi_{d}$ . We define the parameter set of $\phi_{g}$ as $\theta_{g}$ , which is a copied version of $\theta_{d}$ . Different from $\theta_{d}$ , $\theta_{g}$ is set to be learnable for generating images conditioned to $I_{mesh}^{G}$ and $I_{mesh}^{L}$ . Specifically, $\phi_{g}$ has zero convolution $\mathcal{Z}$ [18] at the front of the network, and last 12 blocks of the network consist of $\mathcal{Z}$ . Since $\mathcal{Z}$ is defined as a $1\times 1$ convolution layer whose weights and bias are initialized to zero, the gradients of the weight and bias progressively grow from zeros to optimized parameters in a learnable manner. Hence, $\mathcal{Z}$ helps generated images to be conditioned on $I_{mesh}^{G}$ and $I_{mesh}^{L}$ , while maintaining the quality of image generation. More specifically, $\phi_{d}$ and $\phi_{g}$ share weights at the beginning of training, because parameter sets of both modules, i.e., $\theta_{d}$ and $\theta_{g}$ , are initialized with the pre-trained SD. However, while continuing with training process, $\theta_{g}$ is updated to learn $I_{mesh}^{G}$ and $I_{mesh}^{L}$ , whereas $\theta_{d}$ is fixed frozen to preserve the performance of image generation. At the end of training, $\theta_{d}$ and $\theta_{g}$ are completely different from the beginning. Hence, $\phi_{g}$ is formulated as follows: + +$$ +Y _ {g} = \phi_ {g} (\hat {X} _ {t}, I _ {\text {m e s h}}, K, t; \theta_ {g}), \tag {5} +$$ + +where $I_{mesh} = (I_{mesh}^{G}, I_{mesh}^{L})$ denotes the concatenation of the global and local mesh image, $t$ denotes the diffusion step obtained by positional encoding, $Y_{g} = (Y_{g}^{G}, Y_{g}^{L}) \in \mathbb{R}^{2 \times 4 \times 64 \times 64}$ denotes the concatenation of the global guidance feature $Y_{g}^{G}$ and local guidance feature $Y_{g}^{L}$ . Next, $\phi_{d}$ is formulated as follows: + +$$ +Y _ {d} = \phi_ {d} \left(\hat {X} _ {t}, K, t; \theta_ {d}\right) + Y _ {g}, \tag {6} +$$ + +where $Y_{d} = (Y_{d}^{G},Y_{d}^{L})\in \mathbb{R}^{2\times 4\times 64\times 64}$ indicates the concatenation of the global diffusion feature $Y_{d}^{G}$ and local diffusion feature $Y_{d}^{L}$ . + +Optimization. Since the diffusion model typically involves both forward and reverse processes, our AttentionHand also employs two processes. For the forward process, the noisy embedding $X_{t} = (X_{t}^{G},X_{t}^{L})$ is obtained by progressively + +perturbing Gaussian noise $\epsilon = (\epsilon^{G},\epsilon^{L})$ to the initial embedding $X_0 = (X_0^G,X_0^L)$ by $t$ diffusion steps. $\epsilon^G$ and $\epsilon^L$ denote the global and local noise added to $X_0^G$ and $X_0^L$ , respectively. Then, since $X_{t}$ is updated to $\hat{X}_t$ by TAS, $\epsilon$ is also updated to $\hat{\epsilon} = (\hat{\epsilon}^{G},\hat{\epsilon}^{L})$ . In other words, $\hat{\epsilon}$ is considered the residual noise between $X_0$ and $\hat{X}_t$ . Thus, $\epsilon^G$ and $\epsilon^L$ denote the global and local residual noise. For the reverse process, AttentionHand learns to gradually remove residual noises with global and local denoising processes. Therefore, given text embedding $K$ , diffusion steps $t$ , and mesh images $I_{mesh}^{G}$ and $I_{mesh}^{L}$ , the diffusion training network $\epsilon_{\theta}$ is optimized to predict $\hat{\epsilon}^{G}$ and $\hat{\epsilon}^{L}$ jointly through the following objectives: + +$$ +\mathcal {L} ^ {G} = \mathbb {E} _ {X _ {0} ^ {G}, I _ {\text {m e s h}} ^ {G}, K, t, \hat {\epsilon^ {G}} \sim \mathcal {N} (0, 1)} [ \| \hat {\epsilon^ {G}} - \epsilon_ {\theta} (\hat {X} _ {t} ^ {G}, I _ {\text {m e s h}} ^ {G}, K, t) \| _ {2} ^ {2} ], \tag {7} +$$ + +$$ +\mathcal {L} ^ {L} = \mathbb {E} _ {X _ {0} ^ {L}, I _ {\text {m e s h}} ^ {L}, K, t, \hat {\epsilon} ^ {L} \sim \mathcal {N} (0, 1)} [ \| \hat {\epsilon} ^ {L} - \epsilon_ {\theta} (\hat {X} _ {t} ^ {L}, I _ {\text {m e s h}} ^ {L}, K, t) \| _ {2} ^ {2} ], \tag {8} +$$ + +where $\mathcal{L}^G$ and $\mathcal{L}^L$ indicate the cost function of global and local features, respectively. Thus, the final objective is defined as follows: + +$$ +\mathcal {L} = \lambda^ {G} \mathcal {L} ^ {G} + \lambda^ {L} \mathcal {L} ^ {L}, \tag {9} +$$ + +where $\lambda^G$ and $\lambda^L$ are weighted coefficients of $\mathcal{L}^G$ and $\mathcal{L}^L$ . + +# 3.4 Decoding Phase + +In the decoding phase, we can generate a new RGB hand image $\hat{I}_{RGB} \in \mathbb{R}^{3 \times 512 \times 512}$ by passing $Y_{d}$ through the decoder $\mathcal{D}$ , as shown in the fourth box of Fig. 3. While $\mathcal{E}$ encodes $X_{0}$ by downsampling $I_{RGB}$ in the latent space, $\mathcal{D}$ decodes $\hat{I}_{RGB}$ by upsampling $Y_{d}$ in the pixel space, conditioned to given text prompt and mesh images. The structure of $\mathcal{D}$ is similar to the decoder of VQ-GAN. The decoding phase is expressed as follows: + +$$ +\hat {I} _ {R G B} = \mathcal {D} \left(Y _ {d}\right). \tag {10} +$$ + +# 4 Experiments + +# 4.1 Datasets + +For the text-to-image generation, we adopted MSCOCO [6]. For the 3D hand mesh reconstruction, we adopted Hands-In-Action (HIC) [37], Re:InterHand (ReIH) [5], InterHand2.6M (IH2.6M) [7], and MSCOCO. Due to the page limit, details will be explained in the supplementary materials. + +Table 1: Quantitative comparisons with state-of-the-art text-to-image generation models. + +
MethodsFID↓KID↓FID-H↓KID-H↓Hand Conf.↑MSE-2D↓MSE-3D↓User Pref.(%)↑
Stable Diffusion [17]40.520.0068450.780.025540.6512.9324.5915.864
Uni-ControlNet [20]30.340.0074437.770.020040.8552.1053.0398.796
T2I-Adapter [19]22.000.0076132.080.015680.9141.5462.45119.676
ControlNet [18]21.670.0065840.320.020980.8101.2522.1827.948
AttentionHand (w/o TAS)21.270.0033128.560.013900.9551.2112.04220.734
AttentionHand (w/TAS)20.710.0030127.090.012870.9651.0261.98636.905
+ +![](images/e0ff79691e559c26ac675e737abd3f2b2167700d91520f12f79568ba09841603.jpg) +Fig. 5: Qualitative comparisons with state-of-the-art text-to-image generation models. Red and green boxes in each sample indicate the wrong and current hand bounding box, respectively. + +# 4.2 Evaluation Protocol + +For the text-to-image generation, we adopted FID [38], FID-Hand (FID-H), KID [39], KID-Hand (KID-H), the hand confidence score (Hand Conf.) [40], the mean square error of 2D and 3D keypoints (MSE-2D, 3D), and the user preference (User Pref.). For the 3D hand mesh reconstruction, we adopted the mean pervertex position error (MPVPE), the right hand-relative vertex error (RRVE), and the mean relative-root position error (MRRPE). Due to the page limit, details will be explained in the supplementary materials. + +# 4.3 Comparisons with State-of-the-arts + +Text-to-Image Generation. As shown in Table 1, our AttentionHand exhibited the highest performance in all metrics among state-of-the-arts [17-20]. This + +Table 2: Quantitative comparisons with state-of-the-art 3D hand mesh reconstruction methods with and without AttentionHand. The red subscripts indicate the difference in performance with and without AttentionHand. + +
DatasetsMethodsIn-the-wildIn-the-lab
HIC [37]ReIH [5]IH2.6M [7]
MPVPE↓RRVE↓MRRPE↓MPVPE↓RRVE↓MRRPE↓MPVPE↓RRVE↓MRRPE↓
IHMR [8]38.5745.51119.6430.9045.5598.4516.9421.9833.39
IHMR+AttentionHand36.73-1.8444.10-1.4194.63-25.0129.11-1.7943.12-2.4387.07-11.3815.09-1.8520.55-1.4332.21-1.18
InterShape [9]27.6634.69110.2527.8738.5680.0412.9717.3531.56
InterShape+AttentionHand25.04-2.6233.33-1.3680.17-30.0826.44-1.4336.54-2.0261.41-18.6311.90-1.0716.22-1.1330.04-1.52
IntagHand [10]23.0728.7452.4625.9030.0542.2212.3417.3229.31
IntagHand+AttentionHand21.87-1.2027.09-1.6547.11-5.3523.39-2.5128.77-1.2833.98-8.2411.42-0.9215.81-1.5129.18-0.13
DIR [13]21.8926.1143.1121.8229.6637.0110.2617.1128.98
DIR+AttentionHand20.66-1.2325.87-0.2440.54-2.5719.91-1.9126.67-2.9935.05-1.9610.09-0.1716.99-0.1228.02-0.96
InterWild [16]15.3021.3531.2613.9920.0722.3811.5219.7726.87
InterWild+AttentionHand14.74-0.5621.10-0.2529.26-2.0013.95-0.0419.94-0.1322.05-0.3310.62-0.9019.09-0.6825.74-1.13
+ +![](images/623fef012dd936e01fdf1be0e614dd203f50ff3cc40f47dee6578b3813ce51bc.jpg) +Fig. 6: Qualitative comparisons on MSCOCO [6]. Red and green boxes indicate wrong and correct region of the reconstructed hand, respectively. + +is particularly evident in the comparison of FID(-H) and KID(-H), which signify the quality of the generated images being on par with real RGB images. Furthermore, the lowest MSE-2D and MSE-3D indicates the remarkable alignment between the generated images and the corresponding hand mesh images. With respect to the user preference, AttentionHand scored the highest compared to other methods. It implies that most users acknowledged the outstanding quality of hand images generated by AttentionHand. In addition, as shown in Fig. 5, our AttentionHand generated the high-quality hand image which is well-corresponding with the given mesh image and fully reflected the given text prompt. Specifically, even when two-hands mesh image is given, which is more challenging than in the case of single-hand mesh image, AttentionHand generated the hand image robustly. It implies our AttentionHand is proper to generate + +well-aligned hand images with given mesh images and text prompt. Additional qualitative results are in the supplementary materials. + +3D Hand Mesh Reconstruction. To verify our AttentionHand extensively, we trained state-of-the-art hand pose networks [8-10, 13, 16] by additionally adding new data generated by AttentionHand. As shown in Table 2, the performance of all methods increased for all metrics. Specifically, with respect to the MPVPE, AttentionHand showed the dramatic performance improvement with InterWild [16] about $3.66\%$ and $7.81\%$ on HIC and ReIH, respectively. With respect to the RRVE, it increased by about $1.17\%$ and $0.65\%$ on HIC and ReIH, respectively. With respect to the MRRPE, it increased by about $6.40\%$ and $1.47\%$ on HIC and ReIH, respectively. These imply generated hand images help increasing the accuracy of the 3D hand mesh reconstruction. In addition, the qualitative performance for in-the-wild scenes is also verified as shown in Fig. 6. Although MSCOCO mainly contains in-the-wild situations, 3D hand mesh is reconstructed robustly. It implies that even for difficult situations, the performance of reconstruction can be improved by utilizing AttentionHand. Additional qualitative results are in the supplementary materials. + +# 4.4 Ablation Studies + +Text Attention Stage (TAS). We deeply dived into TAS to verify its superiority. Firstly, as in the last two rows in Table 1, TAS showed its effectiveness in all metrics. In addition, as shown in Fig. 7, attention maps are well described their corresponding tokens in the case of with TAS. It implies that with TAS, AttentionHand can reflect hand-related tokens enough. Additional qualitative results are in the supplementary materials. + +Secondly, we conducted more experiments about Gaussian filters as follows: (1) no Gaussian filter, (2) random Gaussian filter, and (3) fixed Gaussian filter. As shown in the second, third, and fourth columns of Fig. 8, we found interesting results: in the case of (1), the hand was disappeared or its shape became strange. In the case of (2), generated images are not well-aligned with given hand mesh images. However, in the case of (3), generated images are well-aligned with given hand mesh images and look natural. Hence, we determined fixed Gaussian filter makes the generated image plausibly regardless of diffusion timestep $t$ . + +Thirdly, we compared our loss, $\mathcal{L}^{T\hat{A}S}$ , with the load balancing loss $(\mathcal{L}^{LB})$ [41, 42]. Since $\mathcal{L}^{LB}$ is an auxiliary loss for balancing loads among experts, it plays a similar role with $\mathcal{L}^{TAS}$ , which evenly reflects the image features of all the attention maps. Therefore, we replaced $\mathcal{L}^{TAS}$ to $\mathcal{L}^{LB}$ and considered its feasibility as shown in the fifth column of Fig. 8. Unfortunately, in the case of $\mathcal{L}^{LB}$ , generated images are not fit at all with given hand mesh images. We guess while $\mathcal{L}^{TAS}$ updates the image embedding based on the spatial information of the attention map, $\mathcal{L}^{LB}$ flattens the 2D attention map as 1D representation, leading to distort spatial knowledge. + +Last but not least, we explored the range of updated noise $(\hat{\epsilon})$ . According to [43], we set $\alpha_{t}$ of Eq. 4 as gradually decreasing according to timestep $t$ (i.e., + +![](images/434ca597a084b46368eb5ad7a59d16ef094b44d688bee6d69ddf5f123d6b91e7.jpg) +Fig. 7: Ablation studies on the text attention stage (TAS). Attention maps with red and green box are results without and with TAS, respectively. Red and green bounding boxes indicate wrong and correct hand poses, respectively. + +![](images/b1339acda2dc8c2a7d3a3dee0ebf8a4eb52badf52d9345db80cf2389cdcf37c2.jpg) +Fig. 8: Ablation studies on Gaussian filter, losses (i.e. $\mathcal{L}^{TAS}$ and $\mathcal{L}^{LB}$ ), and the regularization of $\hat{\epsilon}$ for the text attention stage (TAS). Red and green bounding boxes indicate wrong and correct hand poses, respectively. + +from 20 to 10) for regularization of $\hat{\epsilon}$ . However, if $\alpha_{t}$ is randomly set, $\hat{\epsilon}$ tends to be out of distribution (i.e., Gaussian distribution) as shown in the sixth column of Fig. 8: in the case of w/o regularization, generated images are not aligned with given mesh images, or missed some hands. Therefore, it is necessary to regularize $\hat{\epsilon}$ for faithful hand image generation. + +Model Design Justification. To justify our model's superiority, we compared the characteristics of prior works including our model. As shown in Table 3, our model's distinctive and potential features compared to prior works are (1) harmonious preservation of locality (i.e., hand) with globality (i.e., in-the-wild scene), and (2) selective attention on hand-related tokens by cross attention. Specifically, to harmonize globality and locality, we developed global and local designs for the visual attention stage (VAS). Moreover, since the global and local branches are designed structurally same, we set them to share their weights for reducing the number of training parameters (about $20.2\% \downarrow$ ) and improving the generalizability (see two shaded rows in Table 4). We experimentally verified the effectiveness of our design as shown in Table 4. + +Table 3: Network comparisons with prior works. + +
MethodsText PromptVisual PromptLocalityHand-related Token Attention
Stable Diffusion✓
Uni-ControlNet✓✓
T2I-Adapter✓✓
ControlNet✓✓
AttentionHand (Ours)✓✓✓✓
+ +Table 4: Ablation studies on the visual attention stage (VAS). + +
GlobalityLocalityWeightsFID↓KID↓FID-H↓KID-H↓Hand Conf.↑MSE-2D↓MSE-3D↓
Shared40.520.0068450.780.025540.6512.9324.591
✓Shared21.670.0065840.320.020980.8101.2522.182
✓Shared52.980.0071332.110.016040.9111.5392.397
✓✓Shared20.710.0030127.090.012870.9651.0261.986
✓✓Separated21.900.0029326.890.013400.9601.1082.017
+ +![](images/609314d8b34e9448b4f346398543efe94a88cb6b367ec593c8abbae59c08e3a4.jpg) +Fig. 9: (a) Multiple generated hand images from same modalities. Green boxes indicate correct hand poses. (b) t-SNE distribution of AttentionHand and MSCOCO [6]. + +Robustness of Generated Dataset. To verify robustness of our generated dataset, we generated multiple hand images from same modalities as shown in Fig. 9(a). As a result, all generated images are perfectly well-aligned with given hand mesh images. Moreover, we found the t-SNE distribution [44] of AttentionHand is broader than MSCOCO as shown in Fig. 9(b). As a result, we believe that AttentionHand can contribute to the downstream task with our extensive in-the-wild hand images, leading to alleviate the domain gap between indoor and outdoor scenes. + +# 5 Conclusion + +In this paper, we introduced a novel text-to-hand image generation model, AttentionHand, which pays attention to the hand-related tokens from the text prompt and global and local mesh images. AttentionHand achieved state-of-the-art performance in text-to-hand image generation, and we demonstrated that training with the dataset generated by our AttentionHand improved the performance of 3D hand mesh reconstruction. However, the diversity may decrease as the generative model is trained to optimize hand mesh images. We expect for the emergence of outstanding diffusion model to improve the diversity and quality of the hand image. + +Acknowledgements. This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00260091) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation) and Korea Institute for Advancement of Technology (KIAT) grant funded by the Korea Government (MOTIE) (P0020535, The Competency Development Program for Industry Specialist) and National Supercomputing Center with supercomputing resources including technical support (KSC-2023-CRE-0444). + +# References + +1. Hampali, Shreyas and Rad, Mahdi and Oberweger, Markus and Lepetit, Vincent. Honnotate: A method for 3D annotation of hand and object poses. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3196-3206, 2020. +2. Chao, Yu-Wei and Yang, Wei and Xiang, Yu and Molchanov, Pavlo and Handa, Ankur and Tremblay, Jonathan and Narang, Yashraj S and Van Wyk, Karl and Iqbal, Umar and Birchfield, Stan and Kautz, Jan and Fox, Dieter. DexYCB: A benchmark for capturing hand grasping of objects. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9044–9053, 2021. +3. Ohkawa, Takehiko and He, Kun and Sener, Fadime and Hodan, Tomas and Tran, Luan and Keskin, Cem. Assemblyhands: Towards egocentric activity understanding via 3D hand pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12999-13008, 2023. +4. Lin, Fanqing and Wilhelm, Connor and Martinez, Tony. Two-hand global 3d pose estimation using monocular rgb. pages 2373-2381, 2021. +5. Moon, Gyeongsik and Saito, Shunsuke and Xu, Weipeng and Joshi, Rohan and Buffalini, Julia and Bellan, Harley and Rosen, Nicholas and Richardson, Jesse and Mize, Mallorie and De Bree, Philippe and Simon, Tomas and Peng, Bo and Garg, Shubham and McPhail, Kevyn and Shiratori, Takaaki. A dataset of relighted 3d interacting hands. Adv. Neural Inform. Process. Syst., 36, 2023. +6. Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Dollar, Piotr and Zitnick, C Lawrence. Microsoft coco: Common objects in context. In *Eur. Conf. Comput. Vis.*, pages 740-755, 2014. +7. Moon, Gyeongsik and Yu, Shouu-I and Wen, He and Shiratori, Takaaki and Lee, Kyoung Mu. Interhand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single RGB image. In *Eur. Conf. Comput. Vis.*, pages 548-564, 2020. +8. Rong, Yu and Wang, Jingbo and Liu, Ziwei and Loy, Chen Change. Monocular 3D reconstruction of interacting hands via collision-aware factorized refinements. In 3DV, pages 432-441, 2021. +9. Zhang, Baowen and Wang, Yangang and Deng, Xiaoming and Zhang, Yinda and Tan, Ping and Ma, Cuixia and Wang, Hongan. Interacting two-hand 3D pose and shape reconstruction from single color image. In Int. Conf. Comput. Vis., pages 11354-11363, 2021. +10. Li, Mengcheng and An, Liang and Zhang, Hongwen and Wu, Lianpeng and Chen, Feng and Yu, Tao and Liu, Yebin. Interacting attention graph for single image two-hand reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2761-2770, 2022. + +11. Hampali, Shreyas and Sarkar, Sayan Deb and Rad, Mahdi and Lepetit, Vincent. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3D pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11090-11100, 2022. +12. Meng, Hao and Jin, Sheng and Liu, Wentao and Qian, Chen and Lin, Mengxiang and Ouyang, Wanli and Luo, Ping. 3D interacting hand pose estimation by hand de-occlusion and removal. In Eur. Conf. Comput. Vis., pages 380-397, 2022. +13. Ren, Pengfei and Wen, Chao and Zheng, Xiaozheng and Xue, Zhou and Sun, Haifeng and Qi, Qi and Wang, Jingyu and Liao, Jianxin. Decoupled Iterative Refinement Framework for Interacting Hands Reconstruction from a Single RGB Image. In Int. Conf. Comput. Vis., pages 8014-8025, 2023. +14. Zuo, Binghui and Zhao, Zimeng and Sun, Wenqian and Xie, Wei and Xue, Zhou and Wang, Yangang. Reconstructing interacting hands with interaction prior from monocular images. In Int. Conf. Comput. Vis., pages 9054-9064, 2023. +15. Li, Lijun and Tian, Linrui and Zhang, Xindi and Wang, Qi and Zhang, Bang and Bo, Liefeng and Liu, Mengyuan and Chen, Chen. Renderih: A large-scale synthetic dataset for 3d interacting hand pose estimation. In Int. Conf. Comput. Vis., pages 20395-20405, 2023. +16. Moon, Gyeongsik. Bringing inputs to shared domains for 3D interacting hands recovery in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 17028-17037, 2023. +17. Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Björn. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. +18. Zhang, Lvmin and Rao, Anyi and Agrawala, Maneesh. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., pages 3836-3847, 2023. +19. Mou, Chong and Wang, Xintao and Xie, Liangbin and Zhang, Jian and Qi, Zhonggang and Shan, Ying and Qie, Xiaohu. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. +20. Zhao, Shihao and Chen, Dongdong and Chen, Yen-Chun and Bao, Jianmin and Hao, Shaozhe and Yuan, Lu and Wong, Kwan-Yee K. Uni-ControlNet: All-in-one control to text-to-image diffusion models. Adv. Neural Inform. Process. Syst., 2023. +21. Podell, Dustin and English, Zion and Lacey, Kyle and Blattmann, Andreas and Dockhorn, Tim and Müller, Jonas and Penna, Joe and Rombach, Robin. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. +22. Mueller, Franziska and Bernard, Florian and Sotnychenko, Oleksandr and Mehta, Dushyant and Sridhar, Srinath and Casas, Dan and Theobalt, Christian. GANerated hands for real-time 3D hand tracking from monocular RGB. In IEEE Conf. Comput. Vis. Pattern Recog., pages 49-59, 2018. +23. Tang, Hao and Wang, Wei and Xu, Dan and Yan, Yan and Sebe, Nicu. GestureGAN for hand gesture-to-gesture translation in the wild. In ACM Int. Conf. Multimedia, pages 774-782, 2018. +24. Hu, Hezhen and Wang, Weilun and Zhou, Wengang and Zhao, Weichao and Li, Houqiang. Model-aware gesture-to-gesture translation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16428-16437, 2021. +25. Hu, Hezhen and Wang, Weilun and Zhou, Wengang and Li, Houqiang. Hand-object interaction image generation. Adv. Neural Inform. Process. Syst., 35:23805-23817, 2022. + +26. Goodfellow, Ian and Pouget-Abadie, Jean and Mirza, Mehdi and Xu, Bing and Warde-Farley, David and Ozair, Sherjil and Courville, Aaron and Bengio, Yoshua. Generative adversarial nets. Adv. Neural Inform. Process. Syst., 27:2672-2680, 2014. +27. Li, Lijun and Zhuo, Li'an and Zhang, Bang and Bo, Liefeng and Chen, Chen. Diff-Hand: End-to-end hand mesh reconstruction via diffusion models. arXiv preprint arXiv:2305.13705, 2023. +28. Lin, Pei and Xu, Sihang and Yang, Hongdi and Liu, Yiran and Chen, Xin and Wang, Jingya and Yu, Jingyi and Xu, Lan. HandDiffuse: Generative controllers for two-hand interactions via diffusion models. arXiv preprint arXiv:2312.04867, 2023. +29. Lu, Wenquan and Xu, Yufei and Zhang, Jing and Wang, Chaoyue and Tao, Dacheng. HandRefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. arXiv preprint arXiv:2311.17957, 2023. +30. Esser, Patrick and Rombach, Robin and Ommer, Bjorn. Taming transformers for high-resolution image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12873-12883, 2021. +31. Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., pages 8748-8763, 2021. +32. Chen, Chun-Fu Richard and Fan, Quanfu and Panda, Rameswar. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Int. Conf. Comput. Vis., pages 357-366, 2021. +33. Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. +34. Chiche, Alebachew and Yitagesu, Betselot. Part of speech tagging: a systematic review of deep learning and machine learning approaches. Journal of Big Data, 9(1):1-25, 2022. +35. He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. +36. Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent., 2020. +37. Tzionas, Dimitrios and Ballan, Luca and Srikantha, Abhilash and Aponte, Pablo and Pollefeys, Marc and Gall, Juergen. Capturing hands in action using discriminative salient points and physics simulation. Int. J. Comput. Vis., 118:172-193, 2016. +38. Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Adv. Neural Inform. Process. Syst., 30:6626-6637, 2017. +39. Binkowski, Mikolaj and Sutherland, Danica J and Arbel, Michael and Gretton, Arthur. Demystifying MMD GANs. In Int. Conf. Learn. Represent., 2018. +40. Narasimhaswamy, Supreeth and Bhattacharya, Uttaran and Chen, Xiang and Dasgupta, Ishita and Mitra, Saayan and Hoai, Minh. Handiffuser: Text-to-image generation with realistic hand appearances. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2468-2479, 2024. + +41. Zhou, Yanqi and Lei, Tao and Liu, Hanxiao and Du, Nan and Huang, Yanping and Zhao, Vincent and Dai, Andrew M and Le, Quoc V and Laudon, James. Mixture-of-experts with expert choice routing. Adv. Neural Inform. Process. Syst., 35:7103-7114, 2022. +42. Fedus, William and Zoph, Barret and Shazeer, Noam. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 23(120):1-39, 2022. +43. Chefer, Hila and Alaluf, Yuval and Vinker, Yael and Wolf, Lior and Cohen-Or, Daniel. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans. Graph., 42(4):1-10, 2023. +44. Van der Maaten, Laurens and Hinton, Geoffrey. Visualizing data using t-SNE. J. Mach. Learn. Res., 9(86):2579-2605, 2008. \ No newline at end of file diff --git a/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/images.zip b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9ef88a30917b9971c1f7b96afc4a28a6965c3ea2 --- /dev/null +++ b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c9cb1d895df364f80e8c461755981350bfccac3d5bec6a3e099aef3afffd380 +size 788311 diff --git a/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/layout.json b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..aed0df6cf6d96221189afc6dedfbbafe231beeef --- /dev/null +++ b/2024/AttentionHand_ Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild/layout.json @@ -0,0 +1,10691 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 112, + 479, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 112, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 112, + 479, + 163 + ], + "type": "text", + "content": "AttentionHand: Text-driven Controllable Hand Image Generation for 3D Hand Reconstruction in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 171, + 189, + 443, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 189, + 443, + 201 + ], + "spans": [ + { + "bbox": [ + 171, + 189, + 443, + 201 + ], + "type": "text", + "content": "Junho Park" + }, + { + "bbox": [ + 171, + 189, + 443, + 201 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 171, + 189, + 443, + 201 + ], + "type": "text", + "content": ", Kyeongbo Kong" + }, + { + "bbox": [ + 171, + 189, + 443, + 201 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 171, + 189, + 443, + 201 + ], + "type": "text", + "content": ", and Suk-Ju Kang" + }, + { + "bbox": [ + 171, + 189, + 443, + 201 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 209, + 457, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 209, + 457, + 232 + ], + "spans": [ + { + "bbox": [ + 157, + 209, + 457, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 209, + 457, + 232 + ], + "type": "text", + "content": " Department of Electronic Engineering, Sogang University, South Korea " + }, + { + "bbox": [ + 157, + 209, + 457, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 209, + 457, + 232 + ], + "type": "text", + "content": " AI Lab, CTO Division, LG Electronics, South Korea" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 232, + 473, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 232, + 473, + 254 + ], + "spans": [ + { + "bbox": [ + 140, + 232, + 473, + 254 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 232, + 473, + 254 + ], + "type": "text", + "content": " Department of Electrical & Electronics Engineering, Pusan National University, South Korea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 255, + 454, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 255, + 454, + 277 + ], + "spans": [ + { + "bbox": [ + 159, + 255, + 454, + 277 + ], + "type": "text", + "content": "junho18.park@gmail.com kbkong@pusan.ac.kr sj Kang@sogang.ac.kr https://github.com/redorangeyellow/AttentionHand" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 327, + 453, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 327, + 453, + 602 + ], + "spans": [ + { + "bbox": [ + 160, + 327, + 453, + 602 + ], + "type": "text", + "content": "Abstract. Recently, there has been a significant amount of research conducted on 3D hand reconstruction to use various forms of human-computer interaction. However, 3D hand reconstruction in the wild is challenging due to extreme lack of in-the-wild 3D hand datasets. Especially, when hands are in complex pose such as interacting hands, the problems like appearance similarity, self-handed occlusion and depth ambiguity make it more difficult. To overcome these issues, we propose AttentionHand, a novel method for text-driven controllable hand image generation. Since AttentionHand can generate various and numerous in-the-wild hand images well-aligned with 3D hand label, we can acquire a new 3D hand dataset, and can relieve the domain gap between indoor and outdoor scenes. Our method needs easy-to-use four modalities (i.e., an RGB image, a hand mesh image from 3D label, a bounding box, and a text prompt). These modalities are embedded into the latent space by the encoding phase. Then, through the text attention stage, hand-related tokens from the given text prompt are attended to highlight hand-related regions of the latent embedding. After the highlighted embedding is fed to the visual attention stage, hand-related regions in the embedding are attended by conditioning global and local hand mesh images with the diffusion-based pipeline. In the decoding phase, the final feature is decoded to new hand images, which are well-aligned with the given hand mesh image and text prompt. As a result, AttentionHand achieved state-of-the-art among text-to-hand image generation models, and the performance of 3D hand mesh reconstruction was improved by additionally training with hand images generated by AttentionHand." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 615, + 453, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 615, + 453, + 627 + ], + "spans": [ + { + "bbox": [ + 160, + 615, + 453, + 627 + ], + "type": "text", + "content": "Keywords: 3D Hand Mesh Reconstruction " + }, + { + "bbox": [ + 160, + 615, + 453, + 627 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 615, + 453, + 627 + ], + "type": "text", + "content": " Text-to-Image Generation" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 223, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 223, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 223, + 653 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 653, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 235, + 665 + ], + "type": "inline_equation", + "content": "\\text{忍}" + }, + { + "bbox": [ + 133, + 653, + 235, + 665 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 286, + 186 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 286, + 186 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 286, + 186 + ], + "type": "image", + "image_path": "b5c47883ff948277ba3d6cbf11820aa8d952ad9d23eb7b1625f6f6c0da0dd07c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 188, + 216, + 198 + ], + "lines": [ + { + "bbox": [ + 205, + 188, + 216, + 198 + ], + "spans": [ + { + "bbox": [ + 205, + 188, + 216, + 198 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 286, + 116, + 479, + 186 + ], + "blocks": [ + { + "bbox": [ + 286, + 116, + 479, + 186 + ], + "lines": [ + { + "bbox": [ + 286, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 286, + 116, + 479, + 186 + ], + "type": "image", + "image_path": "12d027893901489413cf8e86ba89e989cd97aa9c5d7ce4f1e9749609d3cf8f22.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 200, + 479, + 304 + ], + "blocks": [ + { + "bbox": [ + 376, + 188, + 388, + 198 + ], + "lines": [ + { + "bbox": [ + 376, + 188, + 388, + 198 + ], + "spans": [ + { + "bbox": [ + 376, + 188, + 388, + 198 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 200, + 479, + 304 + ], + "lines": [ + { + "bbox": [ + 135, + 200, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 135, + 200, + 479, + 304 + ], + "type": "image", + "image_path": "704b89104c4cd11d2237f0c636f67bf6e102e61653250703e2a05833d4d0ea2d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 307, + 313, + 317 + ], + "lines": [ + { + "bbox": [ + 301, + 307, + 313, + 317 + ], + "spans": [ + { + "bbox": [ + 301, + 307, + 313, + 317 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 338, + 482, + 406 + ], + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 406 + ], + "type": "text", + "content": "Fig. 1: Various acquisition types of 3D hand datasets. (a) In-the-wild dataset (i.e., MSCOCO [6]) is naively acquired with inaccurate pseudo annotation, (b) relighted dataset (i.e., Re:InterHand [5]) consists of unnatural hands with inharmonious background, and (c) our in-the-wild dataset from AttentionHand, which is annotated with accurate 3D labels, contains natural hands with harmonious background, easy to generate, and can be made infinitely." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 426, + 230, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 230, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 230, + 438 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 450, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 616 + ], + "type": "text", + "content": "The goal of 3D hand mesh reconstruction is to recover the 3D hand mesh from a single RGB image. It becomes difficult when hands are in the wild, due to insufficiency of in-the-wild 3D hand datasets. Compared to in-the-lab datasets [1-3], acquisition in-the-wild datasets is challenging due to unpredictable conditions such as weather, lighting, cost of sensors, and safety issues on crowded roads and public places. Even if an in-the-wild dataset is collected, data diversity would be poor due to the aforementioned severe constraints. Although arbitrary labels can be obtained through pseudo annotation, the precision and accuracy is still poor compared to in-the-lab datasets as shown in Fig. 1(a). To tackle this problem, several synthetic datasets [4, 5] have introduced. However, since the hand and background images are synthesized out of harmony, they consist of unnatural and unrealistic hand images as shown in Fig. 1(b). Hence, it is difficult to overcome the domain gap between indoor and outdoor scenes with synthetic datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Moreover, when hands are in a complex pose like interacting hands, it becomes even more challenging to reconstruct 3D hand meshes due to the appearance similarity, self-handed occlusion and depth ambiguity. Starting with InterHand2.6M [7], several works [8-15] have emerged to solve the complex hand" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 479, + 180 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 479, + 180 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 479, + 180 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 479, + 180 + ], + "type": "image", + "image_path": "427ef0e5947b2e7350ebc2c61e31c19ce434208d0e894781630774db6a60b58c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 202, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 480, + 236 + ], + "type": "text", + "content": "Fig. 2: Visualization of attention maps with corresponding tokens from given text prompts. Red and green boxes represent attention maps without and with Attention-Hand, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 264, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 480, + 335 + ], + "type": "text", + "content": "pose. However, they have been employed and evaluated primarily on in-the-lab scenes except for InterWild [16]; it tried to relieve the domain gap by leveraging the geometric features of the hand, which is not affected by the domain. Nevertheless, since InterWild was trained with MSCOCO [6], which is extremely lack of in-the-wild hand images with inaccurate 3D labels, there is earnestly need of more diverse, numerous, and accurately annotated in-the-wild datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 480, + 456 + ], + "type": "text", + "content": "To address aforementioned issues, we propose AttentionHand, a new method for the text-driven controllable hand image generation. AttentionHand is designed based on Stable Diffusion (SD) [17] to create accurate, natural, realistic and harmonious in-the-wild hand images easily and infinitely as shown in Fig. 1(c). AttentionHand has a huge advantage: we can simply generate images with only four modalities - an RGB image, the corresponding hand mesh image, bounding box, and text prompt. Therefore, we can generate (1) various in-the-wild hand images with flexible text prompts, and (2) well-aligned hand images with 3D hand label. By generating new samples with AttentionHand, we can alleviate aforementioned issues of the 3D hand mesh reconstruction in the wild." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 480, + 567 + ], + "type": "text", + "content": "To train AttentionHand, we need to additionally prepare a local RGB image and local mesh image for attention of hand-focused region of the image. The preparation of local information is essential because hands commonly occupy relatively small region in the image. Hence, we obtain local RGB and mesh images by cropping and resizing original RGB and mesh images (i.e., we define them as global information) with the bounding box of hand region. After encoding prepared information in the encoding phase, encoded latent embeddings are fed to the conditioning phase, which is composed of the text attention stage (TAS) and visual attention stage (VAS)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "TAS attends on hand-related tokens from the given text prompt by leveraging attention maps as shown in Fig. 2. Specifically, TAS extracts hand-related attention maps (i.e., holding and hand), and these attention maps are updated to highlight hand-related regions by the refinement based on the softmax operation and Gaussian filter. With TAS, we can obtain more hand-focused images than before. On the other hand, VAS attends on hand-related regions by conditioning global and local hand mesh images with the SD-based pipeline. With global and local information, AttentionHand can be jointly optimized to reflect the global" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "AttentionHand" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "context (i.e., in-the-wild background) and local context (i.e., hand-focused foreground.) In the end of the conditioning phase, we finally get the diffusion feature, which is decoded to new hand images in the decoding phase. Hence, Attention-Hand can generate well-aligned hand images with the given mesh image and text prompt for the 3D hand mesh reconstruction in the wild." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": "To prove the excellence of AttentionHand, we conducted extensive experiments for the text-to-hand image generation and 3D hand mesh reconstruction. As a result, AttentionHand achieved state-of-the-art in the text-to-hand image generation, and the performance of 3D hand mesh reconstruction was considerably improved by additionally training with new hand samples generated by AttentionHand. Especially, the performance was enhanced significantly on in-the-wild datasets, which implies AttentionHand can generate various and well-annotated in-the-wild hand images. The summary of our contributions is as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 291, + 480, + 385 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 291, + 479, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 479, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 479, + 314 + ], + "type": "text", + "content": "- We propose a novel method, AttentionHand, which generates well-aligned in-the-wild hand images in a simple manner without laborious data acquisition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 315, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 350 + ], + "type": "text", + "content": "- AttentionHand is designed based on a generative model that attends on hand-related tokens from the text prompt and hand-related regions from the hand mesh image, for generating hand-focused images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 351, + 479, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 479, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 479, + 385 + ], + "type": "text", + "content": "- AttentionHand achieved state-of-the-art in the text-to-hand image generation, and we verified that utilizing the dataset generated from AttentionHand improves the performance on 3D hand mesh reconstruction in the wild." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 404, + 237, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 237, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 237, + 417 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 430, + 292, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 292, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 292, + 442 + ], + "type": "text", + "content": "2.1 Text-to-Image Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 567 + ], + "type": "text", + "content": "Text-to-image generation aims to synthesize high-resolution image from natural language descriptions. With the advent of diffusion models, various studies on text-to-image generation have been conducted in recent years [17-21]. Specifically, ControlNet [18] and T2I-Adapter [19] proposed novel approaches to incorporate arbitrary condition into the generation process. Recently, UniControlNet [20] presented a novel approach that allows for the simultaneous utilization of various conditions in a flexible and composable manner. Nevertheless, aforementioned models exhibited common limitations in generating hand images, due to the relatively small size of hands within the overall image resolution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 586, + 306, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 306, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 306, + 597 + ], + "type": "text", + "content": "2.2 Generative Models for Hand" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "GANs for Hand. There are several works [22-25] to tackle the hand image generation problem with the generative adversarial network (GAN) [26]. Specifically, a novel network for image-to-image translation [22] was proposed to make generated images follow the same statistical distribution as real-world hand images. GestureGAN [23] was designed to translate hand gesture-to-gesture with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "the explicit hand skeleton information through the color loss and the cycle-consistency loss. Moreover, the first model-aware gesture-to-gesture translation framework [24] was introduced with hand prior as the intermediate representation. Recently, a new method [25], which employs the expressive model-aware hand-object representation and leverages its inherent topology to build the unified surface space, was proposed. However, these works have a common limitation; they are confined to target gestures in generating new hand images. In other words, they are inappropriate to generate in-the-wild images focused on various hands." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "text", + "content": "Diffusion Models for Hand. Recently, some works [27-29] have been addressed hand-related problems with diffusion models. DiffHand [27] introduced the first diffusion-based framework that approaches hand mesh reconstruction as a denoising diffusion process. HandDiffuse [28] proposed a strong baseline for the controllable motion generation of interacting hands using various controllers by designing a diffusion-based model. HandRefiner [29] presented an inpainting pipeline to rectify malformed human hands in generated images with diffusion-based models. However, since these models are not text-driven methods, they cannot generate various in-the-wild hand images conditioned on language instructions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 377, + 202, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 377, + 202, + 391 + ], + "spans": [ + { + "bbox": [ + 131, + 377, + 202, + 391 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 481, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 451 + ], + "type": "text", + "content": "We introduce AttentionHand, a novel framework for creating various and plausible hand images. AttentionHand is a SD-based framework that can generate new RGB images infinitely conditioned on hand mesh images and text prompts. The overall pipeline is shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 467, + 284, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 284, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 284, + 480 + ], + "type": "text", + "content": "3.1 Data Preparation Phase" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": "As shown in the first box of Fig. 3, it just requires four inputs to train AttentionHand: (1) a global RGB hand image " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_{RGB}^{G} \\in \\mathbb{R}^{3 \\times 512 \\times 512}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ", (2) a global hand mesh image " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_{mesh}^{G} \\in \\mathbb{R}^{3 \\times 512 \\times 512}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ", (3) a bounding box of the hand region " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{1 \\times 4}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ", and (4) a hand-related text prompt " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ". However, since hands typically occupy small areas on in-the-wild scenes, we also obtain a local RGB hand image " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_{RGB}^{L} \\in \\mathbb{R}^{3 \\times 512 \\times 512}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": " and a local hand mesh image " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_{mesh}^{L} \\in \\mathbb{R}^{3 \\times 512 \\times 512}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": " by cropping and resizing " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_{RGB}^{G}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_{mesh}^{G}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ". This combination of local and global information enhances hand image conditioning. Details will be explained in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 610, + 243, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 243, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 243, + 624 + ], + "type": "text", + "content": "3.2 Encoding Phase" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "text", + "content": "For the diffusion process in latent space, encoding phase for " + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "inline_equation", + "content": "I_{RGB}^{G}" + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "inline_equation", + "content": "I_{RGB}^{L}" + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "text", + "content": " is implemented by the encoder " + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "text", + "content": ". It makes global and local latent image embeddings " + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "inline_equation", + "content": "X_0^G, X_0^L \\in \\mathbb{R}^{4\\times 64\\times 64}" + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "inline_equation", + "content": "I_{RGB}^{G}" + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "inline_equation", + "content": "I_{RGB}^{L}" + }, + { + "bbox": [ + 130, + 628, + 481, + 667 + ], + "type": "text", + "content": ", and a latent text embedding" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "type": "text", + "content": "AttentionHand" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 244 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 244 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 244 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 244 + ], + "type": "image", + "image_path": "1a29bab4c2648fe0b78e25c48f45afa72b9c5c110b4b752b48b5c585ded6690c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 263, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 342 + ], + "type": "text", + "content": "Fig. 3: Overall pipeline of AttentionHand. In the data preparation phase, we prepare global and local RGB images, global and local hand mesh images, bounding box, and text prompt. In the encoding phase, we get global and local latent image embeddings through VQ-GAN [30], and text embedding through CLIP [31]. In the conditioning phase, we refine image embeddings through the text attention stage, and obtain the diffusion feature through the visual attention stage. In the decoding phase, we generate a new hand image " + }, + { + "bbox": [ + 130, + 263, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\hat{I}_{RGB}" + }, + { + "bbox": [ + 130, + 263, + 482, + 342 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 263, + 482, + 342 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 130, + 263, + 482, + 342 + ], + "type": "text", + "content": " through VQ-GAN." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{77 \\times 768}" + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "inline_equation", + "content": "X_0^G" + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "inline_equation", + "content": "X_0^L" + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "text", + "content": " are obtained by VQ-GAN [30], and " + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 363, + 482, + 413 + ], + "type": "text", + "content": " is obtained by CLIP [31] as shown in the second box of Fig. 3. These latent embeddings are fed as inputs to the conditioning phase, which will be introduced by the next subsection. The encoding phase is expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 233, + 420, + 481, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 420, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 233, + 420, + 481, + 435 + ], + "type": "interline_equation", + "content": "X _ {0} ^ {G}, X _ {0} ^ {L}, K = \\mathcal {E} \\left(I _ {R G B} ^ {G}, I _ {R G B} ^ {L}, U\\right). \\tag {1}", + "image_path": "3885c9dfdaa75208b40caa8fdc678ab7caadf41972c6d19b6fa60a9c8e615994.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 460, + 261, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 261, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 261, + 472 + ], + "type": "text", + "content": "3.3 Conditioning Phase" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 480, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 553 + ], + "type": "text", + "content": "For generating new hand images conditioned by given text prompt and mesh images, we design the text attention stage (TAS) and the visual attention stage (VAS) in the conditioning phase, as shown in the third box of Fig. 3. TAS is a stage of paying attention to tokens for the hand and its corresponding gesture in a given text. VAS is a stage of training the SD-based model specialized for hand image generation by conditioning global and local mesh images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": "Text Attention Stage (TAS). TAS is a stage of attending tokens which represent hand or gestures in a given text prompt as shown in Fig. 4(a). First, by adding Gaussian noise to " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "X_0^G" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "X_0^L" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " diffusion steps, the global noisy embedding " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "X_t^G" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " and local noisy embedding " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "X_t^L" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " are obtained. For simplicity, we define as " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "X_0 = (X_0^G, X_0^L)" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "X_t = (X_t^G, X_t^L)" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "X_t" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " are fed to TAS as inputs. For the text attention of TAS, we utilize the cross attention [32]. Specifically, an attention map " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{H \\times W \\times N}" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " is obtained by calculating the key (i.e., " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": ") and query (i.e., " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": ", which is the linear projection of intermediate feature" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 325, + 219 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 325, + 219 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 325, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 325, + 219 + ], + "type": "image", + "image_path": "28a6c76d0d51a82ec13193fa5912a81df0e28166b724a7c10d6fc70384cca3c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "text", + "content": "Fig. 4: (a) Overall process of the text attention stage (TAS). By leveraging the hand-related tagging and refinement, we can highlight hand-related attention maps, which leads to update noisy embeddings with " + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{TAS}" + }, + { + "bbox": [ + 130, + 239, + 482, + 294 + ], + "type": "text", + "content": ". (b) Overall process of the visual attention stage (VAS). By utilizing the global and local information, we can obtain the harmonious diffusion feature, which leads to generate high-fidelity hand images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 332, + 119, + 477, + 217 + ], + "blocks": [ + { + "bbox": [ + 332, + 119, + 477, + 217 + ], + "lines": [ + { + "bbox": [ + 332, + 119, + 477, + 217 + ], + "spans": [ + { + "bbox": [ + 332, + 119, + 477, + 217 + ], + "type": "image", + "image_path": "03fe6f18fc5147b5951beb4fe431389fe84841795d89690833447bc8258e6e54.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "text", + "content": "map from " + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "text", + "content": " of U-Net [33] in SD). " + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "text", + "content": " denote the height and width of " + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "text", + "content": " denotes the number of all tokens of " + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 320, + 480, + 344 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": "Next, to extract hand-related attention maps " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "A_{k\\in K}\\in \\mathbb{R}^{H\\times W\\times N_k}" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": ", we design the hand-related tagging " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{tag}" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": ", which is based on part-of-speech tagging [34], where " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " denotes the number of hand-related tokens in " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{tag}" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " determines if the input token indicates the hand-related word (i.e., holding, taking, or hand). With " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{tag}" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": ", we can attend hand-related tokens " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " to generate more hand-focused images. More details are in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 420, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 480 + ], + "type": "text", + "content": "Then, we employ the softmax operation and Gaussian smoothing to maximize the effect of " + }, + { + "bbox": [ + 130, + 420, + 482, + 480 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 130, + 420, + 482, + 480 + ], + "type": "text", + "content": ". Since the Gaussian filter effectively removes noise from images and preserves detailed information by using the average value of surrounding pixels, we fully exploit these advantages. Hence, " + }, + { + "bbox": [ + 130, + 420, + 482, + 480 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 130, + 420, + 482, + 480 + ], + "type": "text", + "content": " is updated to " + }, + { + "bbox": [ + 130, + 420, + 482, + 480 + ], + "type": "inline_equation", + "content": "\\hat{A}_k\\in \\mathbb{R}^{H\\times W\\times N_k}" + }, + { + "bbox": [ + 130, + 420, + 482, + 480 + ], + "type": "text", + "content": " by refining hand-related attention maps as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 235, + 492, + 480, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 492, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 235, + 492, + 480, + 507 + ], + "type": "interline_equation", + "content": "\\hat {A} _ {k} = \\text {G a u s s i a n} (\\operatorname {S o f t m a x} (A _ {k})). \\tag {2}", + "image_path": "e93c3a885712e0f361e742bec9290c67a0c192bcd6b0b255b195ba4788a35eb3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": "For simplicity, we define " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\hat{A} \\in \\mathbb{R}^{H \\times W \\times N}" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": " as the concatenation of " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\hat{A}_{k \\in K}" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "A_{l \\notin K} \\in \\mathbb{R}^{H \\times W \\times N_l}" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "N_l" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": " denotes the number of not hand-related tokens in " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "N = N_k + N_l" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": "Moreover, optimization for evenly reflecting the image features of all attention maps is necessary. In other words, it is required to design an objective to prevent poor generation of the image feature for a specific token. Specifically, for arbitrary token " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "n \\in K" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": ", the highest value " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": " among all patches in the " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": "-th refined attention map " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\hat{A}_n" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": " is extracted, and it is subtracted from 1. This operation is implemented for all tokens in " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": ", and a novel loss, which named " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{TAS}" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": ", is computed the largest value among them as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 247, + 651, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 651, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 247, + 651, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {T A S} = \\max _ {n \\in K} \\left(1 - s _ {n}\\right). \\tag {3}", + "image_path": "1bdffb0a50f6bd61be2beac7eba2da3fe979fb7100b9b9c00d2a2fd1dcaf5f81.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "AttentionHand" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "text", + "content": "Based on " + }, + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{TAS}" + }, + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "text", + "content": " is updated to " + }, + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t} = (\\hat{X}_{t}^{G},\\hat{X}_{t}^{L})" + }, + { + "bbox": [ + 132, + 114, + 400, + 129 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 134, + 481, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 134, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 255, + 134, + 481, + 148 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {t} = X _ {t} - \\alpha_ {t} \\nabla_ {X _ {t}} \\mathcal {L} ^ {T A S}, \\tag {4}", + "image_path": "9f3815a836bbab3a4373bb106b0766f7ed0dcf6cecc15d0bb384d0986f7b844e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 155, + 481, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 155, + 481, + 179 + ], + "spans": [ + { + "bbox": [ + 131, + 155, + 481, + 179 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 155, + 481, + 179 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 131, + 155, + 481, + 179 + ], + "type": "text", + "content": " indicates the learning rate, and " + }, + { + "bbox": [ + 131, + 155, + 481, + 179 + ], + "type": "inline_equation", + "content": "\\hat{X}_t^G" + }, + { + "bbox": [ + 131, + 155, + 481, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 155, + 481, + 179 + ], + "type": "inline_equation", + "content": "\\hat{X}_t^L" + }, + { + "bbox": [ + 131, + 155, + 481, + 179 + ], + "type": "text", + "content": " indicate updated global and local noisy embedding, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "text", + "content": "Visual Attention Stage (VAS). VAS is a stage of training SD-based model by conditioning the aforementioned global and local mesh image. VAS is composed of two modules: one is the guidance module " + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "inline_equation", + "content": "\\phi_{g}" + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "text", + "content": ", and the other is the diffusion module " + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "inline_equation", + "content": "\\phi_{d}" + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "text", + "content": " as shown in Fig. 4(b). First, the diffusion module " + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "inline_equation", + "content": "\\phi_{d}" + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "text", + "content": " is designed based on an U-Net network, consisting of 25 blocks: 8 blocks are downsampling and upsampling convolution layers, and the remaining 17 blocks consist of four ResNet [35] layers and two Vision Transformers [36]. We define the parameter set of " + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "inline_equation", + "content": "\\phi_{d}" + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "inline_equation", + "content": "\\theta_{d}" + }, + { + "bbox": [ + 130, + 194, + 479, + 302 + ], + "type": "text", + "content": ", which is fixed frozen to maintain the image generation performance of SD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": "On the other hand, the guidance module " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\phi_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " is also based on an U-Net network with 25 blocks of " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\phi_{d}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": ". We define the parameter set of " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\phi_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": ", which is a copied version of " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{d}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": ". Different from " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{d}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " is set to be learnable for generating images conditioned to " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "I_{mesh}^{G}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "I_{mesh}^{L}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\phi_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " has zero convolution " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " [18] at the front of the network, and last 12 blocks of the network consist of " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " is defined as a " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " convolution layer whose weights and bias are initialized to zero, the gradients of the weight and bias progressively grow from zeros to optimized parameters in a learnable manner. Hence, " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " helps generated images to be conditioned on " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "I_{mesh}^{G}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "I_{mesh}^{L}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": ", while maintaining the quality of image generation. More specifically, " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\phi_{d}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\phi_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " share weights at the beginning of training, because parameter sets of both modules, i.e., " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{d}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": ", are initialized with the pre-trained SD. However, while continuing with training process, " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " is updated to learn " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "I_{mesh}^{G}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "I_{mesh}^{L}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": ", whereas " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{d}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " is fixed frozen to preserve the performance of image generation. At the end of training, " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{d}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " are completely different from the beginning. Hence, " + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\phi_{g}" + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 490, + 480, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 490, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 247, + 490, + 480, + 504 + ], + "type": "interline_equation", + "content": "Y _ {g} = \\phi_ {g} (\\hat {X} _ {t}, I _ {\\text {m e s h}}, K, t; \\theta_ {g}), \\tag {5}", + "image_path": "73c6ddc319f58caf84a6c4bf678a32af9554f5405926acd2e17b17f4a582e79b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "inline_equation", + "content": "I_{mesh} = (I_{mesh}^{G}, I_{mesh}^{L})" + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "text", + "content": " denotes the concatenation of the global and local mesh image, " + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "text", + "content": " denotes the diffusion step obtained by positional encoding, " + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "inline_equation", + "content": "Y_{g} = (Y_{g}^{G}, Y_{g}^{L}) \\in \\mathbb{R}^{2 \\times 4 \\times 64 \\times 64}" + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "text", + "content": " denotes the concatenation of the global guidance feature " + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "inline_equation", + "content": "Y_{g}^{G}" + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "text", + "content": " and local guidance feature " + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "inline_equation", + "content": "Y_{g}^{L}" + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "text", + "content": ". Next, " + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\phi_{d}" + }, + { + "bbox": [ + 132, + 510, + 482, + 563 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 570, + 480, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 570, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 249, + 570, + 480, + 583 + ], + "type": "interline_equation", + "content": "Y _ {d} = \\phi_ {d} \\left(\\hat {X} _ {t}, K, t; \\theta_ {d}\\right) + Y _ {g}, \\tag {6}", + "image_path": "0a344b54228eb7b7f37cb3ef011e952de06a9401a1cf94a2344de75e8e5e24ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 588, + 481, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 588, + 481, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 588, + 481, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 588, + 481, + 615 + ], + "type": "inline_equation", + "content": "Y_{d} = (Y_{d}^{G},Y_{d}^{L})\\in \\mathbb{R}^{2\\times 4\\times 64\\times 64}" + }, + { + "bbox": [ + 131, + 588, + 481, + 615 + ], + "type": "text", + "content": " indicates the concatenation of the global diffusion feature " + }, + { + "bbox": [ + 131, + 588, + 481, + 615 + ], + "type": "inline_equation", + "content": "Y_{d}^{G}" + }, + { + "bbox": [ + 131, + 588, + 481, + 615 + ], + "type": "text", + "content": " and local diffusion feature " + }, + { + "bbox": [ + 131, + 588, + 481, + 615 + ], + "type": "inline_equation", + "content": "Y_{d}^{L}" + }, + { + "bbox": [ + 131, + 588, + 481, + 615 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "content": "Optimization. Since the diffusion model typically involves both forward and reverse processes, our AttentionHand also employs two processes. For the forward process, the noisy embedding " + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "X_{t} = (X_{t}^{G},X_{t}^{L})" + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "content": " is obtained by progressively" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": "perturbing Gaussian noise " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon = (\\epsilon^{G},\\epsilon^{L})" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " to the initial embedding " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "X_0 = (X_0^G,X_0^L)" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " diffusion steps. " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon^G" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon^L" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " denote the global and local noise added to " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "X_0^G" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "X_0^L" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": ", respectively. Then, since " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " is updated to " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\hat{X}_t" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " by TAS, " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " is also updated to " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon} = (\\hat{\\epsilon}^{G},\\hat{\\epsilon}^{L})" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": ". In other words, " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " is considered the residual noise between " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\hat{X}_t" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon^G" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon^L" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " denote the global and local residual noise. For the reverse process, AttentionHand learns to gradually remove residual noises with global and local denoising processes. Therefore, given text embedding " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": ", diffusion steps " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": ", and mesh images " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "I_{mesh}^{G}" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "I_{mesh}^{L}" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": ", the diffusion training network " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " is optimized to predict " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}^{G}" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}^{L}" + }, + { + "bbox": [ + 130, + 114, + 482, + 231 + ], + "type": "text", + "content": " jointly through the following objectives:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 180, + 242, + 481, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 242, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 180, + 242, + 481, + 261 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {G} = \\mathbb {E} _ {X _ {0} ^ {G}, I _ {\\text {m e s h}} ^ {G}, K, t, \\hat {\\epsilon^ {G}} \\sim \\mathcal {N} (0, 1)} [ \\| \\hat {\\epsilon^ {G}} - \\epsilon_ {\\theta} (\\hat {X} _ {t} ^ {G}, I _ {\\text {m e s h}} ^ {G}, K, t) \\| _ {2} ^ {2} ], \\tag {7}", + "image_path": "608ba70fc4e593d24e56a31e3660857861aa4f518021205b7cfaf8c77aa094f2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 283, + 481, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 283, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 181, + 283, + 481, + 301 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {L} = \\mathbb {E} _ {X _ {0} ^ {L}, I _ {\\text {m e s h}} ^ {L}, K, t, \\hat {\\epsilon} ^ {L} \\sim \\mathcal {N} (0, 1)} [ \\| \\hat {\\epsilon} ^ {L} - \\epsilon_ {\\theta} (\\hat {X} _ {t} ^ {L}, I _ {\\text {m e s h}} ^ {L}, K, t) \\| _ {2} ^ {2} ], \\tag {8}", + "image_path": "beab9464ba98711c3982eb33f73be3024ca7e2215474d9319a3b808d86795465.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 479, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 309, + 479, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^G" + }, + { + "bbox": [ + 130, + 309, + 479, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 309, + 479, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^L" + }, + { + "bbox": [ + 130, + 309, + 479, + 335 + ], + "type": "text", + "content": " indicate the cost function of global and local features, respectively. Thus, the final objective is defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 264, + 346, + 480, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 346, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 264, + 346, + 480, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda^ {G} \\mathcal {L} ^ {G} + \\lambda^ {L} \\mathcal {L} ^ {L}, \\tag {9}", + "image_path": "6082a784e1ab3b4bcaa64654ed070c2201ece94902b09ca98b9da7db60a3f615.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "inline_equation", + "content": "\\lambda^G" + }, + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "inline_equation", + "content": "\\lambda^L" + }, + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "text", + "content": " are weighted coefficients of " + }, + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^G" + }, + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^L" + }, + { + "bbox": [ + 131, + 370, + 386, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 408, + 242, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 408, + 242, + 421 + ], + "spans": [ + { + "bbox": [ + 131, + 408, + 242, + 421 + ], + "type": "text", + "content": "3.4 Decoding Phase" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": "In the decoding phase, we can generate a new RGB hand image " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "\\hat{I}_{RGB} \\in \\mathbb{R}^{3 \\times 512 \\times 512}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": " by passing " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": " through the decoder " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": ", as shown in the fourth box of Fig. 3. While " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": " encodes " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "X_{0}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": " by downsampling " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "I_{RGB}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": " in the latent space, " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": " decodes " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "\\hat{I}_{RGB}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": " by upsampling " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "Y_{d}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": " in the pixel space, conditioned to given text prompt and mesh images. The structure of " + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 434, + 487, + 509 + ], + "type": "text", + "content": " is similar to the decoder of VQ-GAN. The decoding phase is expressed as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 274, + 520, + 480, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 520, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 274, + 520, + 480, + 533 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {R G B} = \\mathcal {D} \\left(Y _ {d}\\right). \\tag {10}", + "image_path": "94ca038cf41a1015e7dc6531a3b89352aa6f7aa1726d6525ca2b24546dfd620e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 558, + 230, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 230, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 230, + 572 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 590, + 205, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 590, + 205, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 590, + 205, + 601 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "For the text-to-image generation, we adopted MSCOCO [6]. For the 3D hand mesh reconstruction, we adopted Hands-In-Action (HIC) [37], Re:InterHand (ReIH) [5], InterHand2.6M (IH2.6M) [7], and MSCOCO. Due to the page limit, details will be explained in the supplementary materials." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "type": "text", + "content": "AttentionHand" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 147, + 445, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons with state-of-the-art text-to-image generation models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 147, + 445, + 200 + ], + "lines": [ + { + "bbox": [ + 168, + 147, + 445, + 200 + ], + "spans": [ + { + "bbox": [ + 168, + 147, + 445, + 200 + ], + "type": "table", + "html": "
MethodsFID↓KID↓FID-H↓KID-H↓Hand Conf.↑MSE-2D↓MSE-3D↓User Pref.(%)↑
Stable Diffusion [17]40.520.0068450.780.025540.6512.9324.5915.864
Uni-ControlNet [20]30.340.0074437.770.020040.8552.1053.0398.796
T2I-Adapter [19]22.000.0076132.080.015680.9141.5462.45119.676
ControlNet [18]21.670.0065840.320.020980.8101.2522.1827.948
AttentionHand (w/o TAS)21.270.0033128.560.013900.9551.2112.04220.734
AttentionHand (w/TAS)20.710.0030127.090.012870.9651.0261.98636.905
", + "image_path": "35b1c7021d3606cf52a73266f984b5bd118f647a27fc592a4602f75122249438.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 209, + 479, + 424 + ], + "blocks": [ + { + "bbox": [ + 134, + 209, + 479, + 424 + ], + "lines": [ + { + "bbox": [ + 134, + 209, + 479, + 424 + ], + "spans": [ + { + "bbox": [ + 134, + 209, + 479, + 424 + ], + "type": "image", + "image_path": "e0ff79691e559c26ac675e737abd3f2b2167700d91520f12f79568ba09841603.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 446, + 482, + 480 + ], + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 480 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparisons with state-of-the-art text-to-image generation models. Red and green boxes in each sample indicate the wrong and current hand bounding box, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 502, + 261, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 261, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 261, + 513 + ], + "type": "text", + "content": "4.2 Evaluation Protocol" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 606 + ], + "type": "text", + "content": "For the text-to-image generation, we adopted FID [38], FID-Hand (FID-H), KID [39], KID-Hand (KID-H), the hand confidence score (Hand Conf.) [40], the mean square error of 2D and 3D keypoints (MSE-2D, 3D), and the user preference (User Pref.). For the 3D hand mesh reconstruction, we adopted the mean pervertex position error (MPVPE), the right hand-relative vertex error (RRVE), and the mean relative-root position error (MRRPE). Due to the page limit, details will be explained in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 622, + 338, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 338, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 338, + 634 + ], + "type": "text", + "content": "4.3 Comparisons with State-of-the-arts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Text-to-Image Generation. As shown in Table 1, our AttentionHand exhibited the highest performance in all metrics among state-of-the-arts [17-20]. This" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 158, + 462, + 244 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons with state-of-the-art 3D hand mesh reconstruction methods with and without AttentionHand. The red subscripts indicate the difference in performance with and without AttentionHand." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 158, + 462, + 244 + ], + "lines": [ + { + "bbox": [ + 151, + 158, + 462, + 244 + ], + "spans": [ + { + "bbox": [ + 151, + 158, + 462, + 244 + ], + "type": "table", + "html": "
DatasetsMethodsIn-the-wildIn-the-lab
HIC [37]ReIH [5]IH2.6M [7]
MPVPE↓RRVE↓MRRPE↓MPVPE↓RRVE↓MRRPE↓MPVPE↓RRVE↓MRRPE↓
IHMR [8]38.5745.51119.6430.9045.5598.4516.9421.9833.39
IHMR+AttentionHand36.73-1.8444.10-1.4194.63-25.0129.11-1.7943.12-2.4387.07-11.3815.09-1.8520.55-1.4332.21-1.18
InterShape [9]27.6634.69110.2527.8738.5680.0412.9717.3531.56
InterShape+AttentionHand25.04-2.6233.33-1.3680.17-30.0826.44-1.4336.54-2.0261.41-18.6311.90-1.0716.22-1.1330.04-1.52
IntagHand [10]23.0728.7452.4625.9030.0542.2212.3417.3229.31
IntagHand+AttentionHand21.87-1.2027.09-1.6547.11-5.3523.39-2.5128.77-1.2833.98-8.2411.42-0.9215.81-1.5129.18-0.13
DIR [13]21.8926.1143.1121.8229.6637.0110.2617.1128.98
DIR+AttentionHand20.66-1.2325.87-0.2440.54-2.5719.91-1.9126.67-2.9935.05-1.9610.09-0.1716.99-0.1228.02-0.96
InterWild [16]15.3021.3531.2613.9920.0722.3811.5219.7726.87
InterWild+AttentionHand14.74-0.5621.10-0.2529.26-2.0013.95-0.0419.94-0.1322.05-0.3310.62-0.9019.09-0.6825.74-1.13
", + "image_path": "30fcc7f79f45d4f17dbab3d6b375ab7e968a0ef80a9e618c31926d2591566373.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 261, + 479, + 452 + ], + "blocks": [ + { + "bbox": [ + 137, + 261, + 479, + 452 + ], + "lines": [ + { + "bbox": [ + 137, + 261, + 479, + 452 + ], + "spans": [ + { + "bbox": [ + 137, + 261, + 479, + 452 + ], + "type": "image", + "image_path": "623fef012dd936e01fdf1be0e614dd203f50ff3cc40f47dee6578b3813ce51bc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 470, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 132, + 470, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 482, + 495 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparisons on MSCOCO [6]. Red and green boxes indicate wrong and correct region of the reconstructed hand, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": "is particularly evident in the comparison of FID(-H) and KID(-H), which signify the quality of the generated images being on par with real RGB images. Furthermore, the lowest MSE-2D and MSE-3D indicates the remarkable alignment between the generated images and the corresponding hand mesh images. With respect to the user preference, AttentionHand scored the highest compared to other methods. It implies that most users acknowledged the outstanding quality of hand images generated by AttentionHand. In addition, as shown in Fig. 5, our AttentionHand generated the high-quality hand image which is well-corresponding with the given mesh image and fully reflected the given text prompt. Specifically, even when two-hands mesh image is given, which is more challenging than in the case of single-hand mesh image, AttentionHand generated the hand image robustly. It implies our AttentionHand is proper to generate" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "AttentionHand" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "well-aligned hand images with given mesh images and text prompt. Additional qualitative results are in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "text", + "content": "3D Hand Mesh Reconstruction. To verify our AttentionHand extensively, we trained state-of-the-art hand pose networks [8-10, 13, 16] by additionally adding new data generated by AttentionHand. As shown in Table 2, the performance of all methods increased for all metrics. Specifically, with respect to the MPVPE, AttentionHand showed the dramatic performance improvement with InterWild [16] about " + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "inline_equation", + "content": "3.66\\%" + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "inline_equation", + "content": "7.81\\%" + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "text", + "content": " on HIC and ReIH, respectively. With respect to the RRVE, it increased by about " + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "inline_equation", + "content": "1.17\\%" + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "inline_equation", + "content": "0.65\\%" + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "text", + "content": " on HIC and ReIH, respectively. With respect to the MRRPE, it increased by about " + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "inline_equation", + "content": "6.40\\%" + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "inline_equation", + "content": "1.47\\%" + }, + { + "bbox": [ + 130, + 155, + 483, + 335 + ], + "type": "text", + "content": " on HIC and ReIH, respectively. These imply generated hand images help increasing the accuracy of the 3D hand mesh reconstruction. In addition, the qualitative performance for in-the-wild scenes is also verified as shown in Fig. 6. Although MSCOCO mainly contains in-the-wild situations, 3D hand mesh is reconstructed robustly. It implies that even for difficult situations, the performance of reconstruction can be improved by utilizing AttentionHand. Additional qualitative results are in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 349, + 246, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 349, + 246, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 246, + 361 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": "Text Attention Stage (TAS). We deeply dived into TAS to verify its superiority. Firstly, as in the last two rows in Table 1, TAS showed its effectiveness in all metrics. In addition, as shown in Fig. 7, attention maps are well described their corresponding tokens in the case of with TAS. It implies that with TAS, AttentionHand can reflect hand-related tokens enough. Additional qualitative results are in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "content": "Secondly, we conducted more experiments about Gaussian filters as follows: (1) no Gaussian filter, (2) random Gaussian filter, and (3) fixed Gaussian filter. As shown in the second, third, and fourth columns of Fig. 8, we found interesting results: in the case of (1), the hand was disappeared or its shape became strange. In the case of (2), generated images are not well-aligned with given hand mesh images. However, in the case of (3), generated images are well-aligned with given hand mesh images and look natural. Hence, we determined fixed Gaussian filter makes the generated image plausibly regardless of diffusion timestep " + }, + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": "Thirdly, we compared our loss, " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{T\\hat{A}S}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": ", with the load balancing loss " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}^{LB})" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " [41, 42]. Since " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{LB}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " is an auxiliary loss for balancing loads among experts, it plays a similar role with " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{TAS}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": ", which evenly reflects the image features of all the attention maps. Therefore, we replaced " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{TAS}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{LB}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " and considered its feasibility as shown in the fifth column of Fig. 8. Unfortunately, in the case of " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{LB}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": ", generated images are not fit at all with given hand mesh images. We guess while " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{TAS}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " updates the image embedding based on the spatial information of the attention map, " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{LB}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " flattens the 2D attention map as 1D representation, leading to distort spatial knowledge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Last but not least, we explored the range of updated noise " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\hat{\\epsilon})" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ". According to [43], we set " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " of Eq. 4 as gradually decreasing according to timestep " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " (i.e.," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 462, + 169 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 462, + 169 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 462, + 169 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 462, + 169 + ], + "type": "image", + "image_path": "434ca597a084b46368eb5ad7a59d16ef094b44d688bee6d69ddf5f123d6b91e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 190, + 482, + 222 + ], + "lines": [ + { + "bbox": [ + 131, + 190, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 482, + 222 + ], + "type": "text", + "content": "Fig. 7: Ablation studies on the text attention stage (TAS). Attention maps with red and green box are results without and with TAS, respectively. Red and green bounding boxes indicate wrong and correct hand poses, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 187, + 242, + 428, + 355 + ], + "blocks": [ + { + "bbox": [ + 187, + 242, + 428, + 355 + ], + "lines": [ + { + "bbox": [ + 187, + 242, + 428, + 355 + ], + "spans": [ + { + "bbox": [ + 187, + 242, + 428, + 355 + ], + "type": "image", + "image_path": "b1339acda2dc8c2a7d3a3dee0ebf8a4eb52badf52d9345db80cf2389cdcf37c2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 374, + 482, + 409 + ], + "lines": [ + { + "bbox": [ + 131, + 374, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 482, + 409 + ], + "type": "text", + "content": "Fig. 8: Ablation studies on Gaussian filter, losses (i.e. " + }, + { + "bbox": [ + 131, + 374, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{TAS}" + }, + { + "bbox": [ + 131, + 374, + 482, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 374, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{LB}" + }, + { + "bbox": [ + 131, + 374, + 482, + 409 + ], + "type": "text", + "content": "), and the regularization of " + }, + { + "bbox": [ + 131, + 374, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 131, + 374, + 482, + 409 + ], + "type": "text", + "content": " for the text attention stage (TAS). Red and green bounding boxes indicate wrong and correct hand poses, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": "from 20 to 10) for regularization of " + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": ". However, if " + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": " is randomly set, " + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": " tends to be out of distribution (i.e., Gaussian distribution) as shown in the sixth column of Fig. 8: in the case of w/o regularization, generated images are not aligned with given mesh images, or missed some hands. Therefore, it is necessary to regularize " + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": " for faithful hand image generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "Model Design Justification. To justify our model's superiority, we compared the characteristics of prior works including our model. As shown in Table 3, our model's distinctive and potential features compared to prior works are (1) harmonious preservation of locality (i.e., hand) with globality (i.e., in-the-wild scene), and (2) selective attention on hand-related tokens by cross attention. Specifically, to harmonize globality and locality, we developed global and local designs for the visual attention stage (VAS). Moreover, since the global and local branches are designed structurally same, we set them to share their weights for reducing the number of training parameters (about " + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "inline_equation", + "content": "20.2\\% \\downarrow" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": ") and improving the generalizability (see two shaded rows in Table 4). We experimentally verified the effectiveness of our design as shown in Table 4." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "AttentionHand" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 136, + 427, + 184 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 407, + 126 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 407, + 126 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 407, + 126 + ], + "type": "text", + "content": "Table 3: Network comparisons with prior works." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 136, + 427, + 184 + ], + "lines": [ + { + "bbox": [ + 187, + 136, + 427, + 184 + ], + "spans": [ + { + "bbox": [ + 187, + 136, + 427, + 184 + ], + "type": "table", + "html": "
MethodsText PromptVisual PromptLocalityHand-related Token Attention
Stable Diffusion✓
Uni-ControlNet✓✓
T2I-Adapter✓✓
ControlNet✓✓
AttentionHand (Ours)✓✓✓✓
", + "image_path": "9d671812ffe92031b0cf663e6f6633823965852bb7875614b97849711596afaa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 187, + 217, + 428, + 263 + ], + "blocks": [ + { + "bbox": [ + 177, + 194, + 435, + 206 + ], + "lines": [ + { + "bbox": [ + 177, + 194, + 435, + 206 + ], + "spans": [ + { + "bbox": [ + 177, + 194, + 435, + 206 + ], + "type": "text", + "content": "Table 4: Ablation studies on the visual attention stage (VAS)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 217, + 428, + 263 + ], + "lines": [ + { + "bbox": [ + 187, + 217, + 428, + 263 + ], + "spans": [ + { + "bbox": [ + 187, + 217, + 428, + 263 + ], + "type": "table", + "html": "
GlobalityLocalityWeightsFID↓KID↓FID-H↓KID-H↓Hand Conf.↑MSE-2D↓MSE-3D↓
Shared40.520.0068450.780.025540.6512.9324.591
✓Shared21.670.0065840.320.020980.8101.2522.182
✓Shared52.980.0071332.110.016040.9111.5392.397
✓✓Shared20.710.0030127.090.012870.9651.0261.986
✓✓Separated21.900.0029326.890.013400.9601.1082.017
", + "image_path": "22b3ecdc5eecd14733595abdb39c696f932c07f24578809a8613223d54ddd766.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 151, + 274, + 462, + 351 + ], + "blocks": [ + { + "bbox": [ + 151, + 274, + 462, + 351 + ], + "lines": [ + { + "bbox": [ + 151, + 274, + 462, + 351 + ], + "spans": [ + { + "bbox": [ + 151, + 274, + 462, + 351 + ], + "type": "image", + "image_path": "609314d8b34e9448b4f346398543efe94a88cb6b367ec593c8abbae59c08e3a4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 371, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 395 + ], + "type": "text", + "content": "Fig. 9: (a) Multiple generated hand images from same modalities. Green boxes indicate correct hand poses. (b) t-SNE distribution of AttentionHand and MSCOCO [6]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": "Robustness of Generated Dataset. To verify robustness of our generated dataset, we generated multiple hand images from same modalities as shown in Fig. 9(a). As a result, all generated images are perfectly well-aligned with given hand mesh images. Moreover, we found the t-SNE distribution [44] of AttentionHand is broader than MSCOCO as shown in Fig. 9(b). As a result, we believe that AttentionHand can contribute to the downstream task with our extensive in-the-wild hand images, leading to alleviate the domain gap between indoor and outdoor scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduced a novel text-to-hand image generation model, AttentionHand, which pays attention to the hand-related tokens from the text prompt and global and local mesh images. AttentionHand achieved state-of-the-art performance in text-to-hand image generation, and we demonstrated that training with the dataset generated by our AttentionHand improved the performance of 3D hand mesh reconstruction. However, the diversity may decrease as the generative model is trained to optimize hand mesh images. We expect for the emergence of outstanding diffusion model to improve the diversity and quality of the hand image." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Acknowledgements. This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00260091) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation) and Korea Institute for Advancement of Technology (KIAT) grant funded by the Korea Government (MOTIE) (P0020535, The Competency Development Program for Industry Specialist) and National Supercomputing Center with supercomputing resources including technical support (KSC-2023-CRE-0444)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 228, + 198, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 198, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 198, + 242 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 251, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "text", + "content": "1. Hampali, Shreyas and Rad, Mahdi and Oberweger, Markus and Lepetit, Vincent. Honnotate: A method for 3D annotation of hand and object poses. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3196-3206, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 285, + 480, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 339 + ], + "type": "text", + "content": "2. Chao, Yu-Wei and Yang, Wei and Xiang, Yu and Molchanov, Pavlo and Handa, Ankur and Tremblay, Jonathan and Narang, Yashraj S and Van Wyk, Karl and Iqbal, Umar and Birchfield, Stan and Kautz, Jan and Fox, Dieter. DexYCB: A benchmark for capturing hand grasping of objects. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9044–9053, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 339, + 480, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 480, + 382 + ], + "type": "text", + "content": "3. Ohkawa, Takehiko and He, Kun and Sener, Fadime and Hodan, Tomas and Tran, Luan and Keskin, Cem. Assemblyhands: Towards egocentric activity understanding via 3D hand pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12999-13008, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "text", + "content": "4. Lin, Fanqing and Wilhelm, Connor and Martinez, Tony. Two-hand global 3d pose estimation using monocular rgb. pages 2373-2381, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 404, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 458 + ], + "type": "text", + "content": "5. Moon, Gyeongsik and Saito, Shunsuke and Xu, Weipeng and Joshi, Rohan and Buffalini, Julia and Bellan, Harley and Rosen, Nicholas and Richardson, Jesse and Mize, Mallorie and De Bree, Philippe and Simon, Tomas and Peng, Bo and Garg, Shubham and McPhail, Kevyn and Shiratori, Takaaki. A dataset of relighted 3d interacting hands. Adv. Neural Inform. Process. Syst., 36, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 459, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 501 + ], + "type": "text", + "content": "6. Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Dollar, Piotr and Zitnick, C Lawrence. Microsoft coco: Common objects in context. In *Eur. Conf. Comput. Vis.*, pages 740-755, 2014." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 502, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 544 + ], + "type": "text", + "content": "7. Moon, Gyeongsik and Yu, Shouu-I and Wen, He and Shiratori, Takaaki and Lee, Kyoung Mu. Interhand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single RGB image. In *Eur. Conf. Comput. Vis.*, pages 548-564, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "text", + "content": "8. Rong, Yu and Wang, Jingbo and Liu, Ziwei and Loy, Chen Change. Monocular 3D reconstruction of interacting hands via collision-aware factorized refinements. In 3DV, pages 432-441, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 578, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 620 + ], + "type": "text", + "content": "9. Zhang, Baowen and Wang, Yangang and Deng, Xiaoming and Zhang, Yinda and Tan, Ping and Ma, Cuixia and Wang, Hongan. Interacting two-hand 3D pose and shape reconstruction from single color image. In Int. Conf. Comput. Vis., pages 11354-11363, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 621, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 664 + ], + "type": "text", + "content": "10. Li, Mengcheng and An, Liang and Zhang, Hongwen and Wu, Lianpeng and Chen, Feng and Yu, Tao and Liu, Yebin. Interacting attention graph for single image two-hand reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2761-2770, 2022." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "type": "text", + "content": "AttentionHand" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "11. Hampali, Shreyas and Sarkar, Sayan Deb and Rad, Mahdi and Lepetit, Vincent. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3D pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11090-11100, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "12. Meng, Hao and Jin, Sheng and Liu, Wentao and Qian, Chen and Lin, Mengxiang and Ouyang, Wanli and Luo, Ping. 3D interacting hand pose estimation by hand de-occlusion and removal. In Eur. Conf. Comput. Vis., pages 380-397, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 481, + 236 + ], + "type": "text", + "content": "13. Ren, Pengfei and Wen, Chao and Zheng, Xiaozheng and Xue, Zhou and Sun, Haifeng and Qi, Qi and Wang, Jingyu and Liao, Jianxin. Decoupled Iterative Refinement Framework for Interacting Hands Reconstruction from a Single RGB Image. In Int. Conf. Comput. Vis., pages 8014-8025, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "text", + "content": "14. Zuo, Binghui and Zhao, Zimeng and Sun, Wenqian and Xie, Wei and Xue, Zhou and Wang, Yangang. Reconstructing interacting hands with interaction prior from monocular images. In Int. Conf. Comput. Vis., pages 9054-9064, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 267, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 481, + 310 + ], + "type": "text", + "content": "15. Li, Lijun and Tian, Linrui and Zhang, Xindi and Wang, Qi and Zhang, Bang and Bo, Liefeng and Liu, Mengyuan and Chen, Chen. Renderih: A large-scale synthetic dataset for 3d interacting hand pose estimation. In Int. Conf. Comput. Vis., pages 20395-20405, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 310, + 481, + 342 + ], + "type": "text", + "content": "16. Moon, Gyeongsik. Bringing inputs to shared domains for 3D interacting hands recovery in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 17028-17037, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "text", + "content": "17. Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Björn. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 374, + 481, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 374, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 481, + 406 + ], + "type": "text", + "content": "18. Zhang, Lvmin and Rao, Anyi and Agrawala, Maneesh. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., pages 3836-3847, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 406, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 406, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 406, + 481, + 449 + ], + "type": "text", + "content": "19. Mou, Chong and Wang, Xintao and Xie, Liangbin and Zhang, Jian and Qi, Zhonggang and Shan, Ying and Qie, Xiaohu. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 449, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 449, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 449, + 481, + 482 + ], + "type": "text", + "content": "20. Zhao, Shihao and Chen, Dongdong and Chen, Yen-Chun and Bao, Jianmin and Hao, Shaozhe and Yuan, Lu and Wong, Kwan-Yee K. Uni-ControlNet: All-in-one control to text-to-image diffusion models. Adv. Neural Inform. Process. Syst., 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 482, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 482, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 482, + 481, + 525 + ], + "type": "text", + "content": "21. Podell, Dustin and English, Zion and Lacey, Kyle and Blattmann, Andreas and Dockhorn, Tim and Müller, Jonas and Penna, Joe and Rombach, Robin. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 481, + 568 + ], + "type": "text", + "content": "22. Mueller, Franziska and Bernard, Florian and Sotnychenko, Oleksandr and Mehta, Dushyant and Sridhar, Srinath and Casas, Dan and Theobalt, Christian. GANerated hands for real-time 3D hand tracking from monocular RGB. In IEEE Conf. Comput. Vis. Pattern Recog., pages 49-59, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 481, + 601 + ], + "type": "text", + "content": "23. Tang, Hao and Wang, Wei and Xu, Dan and Yan, Yan and Sebe, Nicu. GestureGAN for hand gesture-to-gesture translation in the wild. In ACM Int. Conf. Multimedia, pages 774-782, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 481, + 632 + ], + "type": "text", + "content": "24. Hu, Hezhen and Wang, Weilun and Zhou, Wengang and Zhao, Weichao and Li, Houqiang. Model-aware gesture-to-gesture translation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16428-16437, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 632, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 664 + ], + "type": "text", + "content": "25. Hu, Hezhen and Wang, Weilun and Zhou, Wengang and Li, Houqiang. Hand-object interaction image generation. Adv. Neural Inform. Process. Syst., 35:23805-23817, 2022." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "text", + "content": "26. Goodfellow, Ian and Pouget-Abadie, Jean and Mirza, Mehdi and Xu, Bing and Warde-Farley, David and Ozair, Sherjil and Courville, Aaron and Bengio, Yoshua. Generative adversarial nets. Adv. Neural Inform. Process. Syst., 27:2672-2680, 2014." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 191 + ], + "type": "text", + "content": "27. Li, Lijun and Zhuo, Li'an and Zhang, Bang and Bo, Liefeng and Chen, Chen. Diff-Hand: End-to-end hand mesh reconstruction via diffusion models. arXiv preprint arXiv:2305.13705, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 481, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 234 + ], + "type": "text", + "content": "28. Lin, Pei and Xu, Sihang and Yang, Hongdi and Liu, Yiran and Chen, Xin and Wang, Jingya and Yu, Jingyi and Xu, Lan. HandDiffuse: Generative controllers for two-hand interactions via diffusion models. arXiv preprint arXiv:2312.04867, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "text", + "content": "29. Lu, Wenquan and Xu, Yufei and Zhang, Jing and Wang, Chaoyue and Tao, Dacheng. HandRefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. arXiv preprint arXiv:2311.17957, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "text", + "content": "30. Esser, Patrick and Rombach, Robin and Ommer, Bjorn. Taming transformers for high-resolution image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12873-12883, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 354 + ], + "type": "text", + "content": "31. Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., pages 8748-8763, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "32. Chen, Chun-Fu Richard and Fan, Quanfu and Panda, Rameswar. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Int. Conf. Comput. Vis., pages 357-366, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "33. Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "type": "text", + "content": "34. Chiche, Alebachew and Yitagesu, Betselot. Part of speech tagging: a systematic review of deep learning and machine learning approaches. Journal of Big Data, 9(1):1-25, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "text", + "content": "35. He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 525 + ], + "type": "text", + "content": "36. Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent., 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 525, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 567 + ], + "type": "text", + "content": "37. Tzionas, Dimitrios and Ballan, Luca and Srikantha, Abhilash and Aponte, Pablo and Pollefeys, Marc and Gall, Juergen. Capturing hands in action using discriminative salient points and physics simulation. Int. J. Comput. Vis., 118:172-193, 2016." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "38. Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Adv. Neural Inform. Process. Syst., 30:6626-6637, 2017." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 601, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 621 + ], + "type": "text", + "content": "39. Binkowski, Mikolaj and Sutherland, Danica J and Arbel, Michael and Gretton, Arthur. Demystifying MMD GANs. In Int. Conf. Learn. Represent., 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "40. Narasimhaswamy, Supreeth and Bhattacharya, Uttaran and Chen, Xiang and Dasgupta, Ishita and Mitra, Saayan and Hoai, Minh. Handiffuser: Text-to-image generation with realistic hand appearances. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2468-2479, 2024." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "AttentionHand" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 249 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "41. Zhou, Yanqi and Lei, Tao and Liu, Hanxiao and Du, Nan and Huang, Yanping and Zhao, Vincent and Dai, Andrew M and Le, Quoc V and Laudon, James. Mixture-of-experts with expert choice routing. Adv. Neural Inform. Process. Syst., 35:7103-7114, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "42. Fedus, William and Zoph, Barret and Shazeer, Noam. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 23(120):1-39, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "43. Chefer, Hila and Alaluf, Yuval and Vinker, Yael and Wolf, Lior and Cohen-Or, Daniel. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans. Graph., 42(4):1-10, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 249 + ], + "type": "text", + "content": "44. Van der Maaten, Laurens and Hinton, Geoffrey. Visualizing data using t-SNE. J. Mach. Learn. Res., 9(86):2579-2605, 2008." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_content_list.json b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..97fe387af1f736146e7f486226fe12b55e343054 --- /dev/null +++ b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_content_list.json @@ -0,0 +1,1747 @@ +[ + { + "type": "text", + "text": "AttnZero: Efficient Attention Discovery for Vision Transformers", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lujun Li $^{1}$ , Zimian Wei $^{2\\boxtimes}$ , Peijie Dong $^{3}$ , Wenhan Luo $^{1}$ , Wei Xue $^{1}$ , Qifeng Liu $^{1\\boxtimes}$ , and Yike Guo $^{1\\boxtimes}$", + "bbox": [ + 236, + 210, + 767, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong University of Science and Technology", + "bbox": [ + 310, + 253, + 687, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National University of Defense Technology", + "bbox": [ + 349, + 268, + 651, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 The Hong Kong University of Science and Technology (Guangzhou)", + "bbox": [ + 267, + 282, + 732, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lilujunai@gmail.com;weizimian16@nudt.edu.cn;pdong212@connect.hkust-gz.edu.cn;whluo.china@gmail.com;weixue@ust.hk;liuqifeng@ust.hk;yikeguo@ust.hk", + "bbox": [ + 215, + 297, + 810, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract In this paper, we present AttnZero, the first framework for automatically discovering efficient attention modules tailored for Vision Transformers (ViTs). While traditional self-attention in ViTs suffers from quadratic computation complexity, linear attention offers a more efficient alternative with linear complexity approximation. However, existing hand-crafted linear attention suffers from performance degradation. To address these issues, our AttnZero constructs search spaces and employs evolutionary algorithms to discover potential linear attention formulations. Specifically, our search space consists of six kinds of computation graphs and advanced activation, normalize, and binary operators. To enhance generality, we derive results of candidate attention applied to multiple advanced ViTs as the multi-objective for the evolutionary search. To expedite the search process, we utilize program checking and rejection protocols to filter out unpromising candidates swiftly. Additionally, we develop Attn-Bench-101, which provides precomputed performance of 2,000 attentions in the search spaces, enabling us to summarize attention design insights. Experimental results demonstrate that the discovered AttnZero module generalizes well to different tasks and consistently achieves improved performance across various ViTs. For instance, the tiny model of DeiT|PVT|Swin|CSwin trained with AttnZero on ImageNet reaches $74.9\\%$ |78.1%|82.1%|82.9% top-1 accuracy. Codes at: https://github.com/lliai/AttnZero.", + "bbox": [ + 259, + 357, + 743, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision Transformer $\\cdot$ Efficient Attention $\\cdot$ AutoML", + "bbox": [ + 261, + 674, + 689, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 710, + 375, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, Vision Transformers (ViTs) demonstrate remarkable success in various vision tasks, including image classification [16, 19, 58, 62], object detection [60, 75], semantic segmentation [47]. ViTs draw inspiration from the transformative impact of Transformers in natural language processing, utilizing self-attention mechanisms to capture global dependencies. However, self-attention", + "bbox": [ + 212, + 741, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors.", + "bbox": [ + 228, + 825, + 411, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "has a quadratic complexity that scales with the square of the input sequence length. This quadratic complexity limits the scalability of ViTs for handling large image inputs, thereby limiting their practical applications.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, researchers have been exploring the use of linear attention [57] as a more computationally efficient alternative. Unlike self-attention, linear attention has a linear complexity approximation, making it more suitable for handling large input sequences. However, existing linear attention methods [42, 54] suffer from several limitations: (1) Linear attention often leads to reduced performance than softmax attention due to its simplifying assumptions. The simple linear operators used in linear attention have limited flexibility and produce noisier activations than nonlinear functions [24]. These performance degradations challenge leveraging linear attention as a viable substitute for self-attention in ViTs. (2) Hand-crafted linear attention designs are often specific to certain tasks or architectures, lacking generality across different scenarios. Additionally, the manual design is labor-intensive, time-consuming, and heavily reliant on expert knowledge. Consequently, these drawbacks hinder the scalability and accessibility of linear attention in ViTs.", + "bbox": [ + 212, + 196, + 787, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we have made several observations. We notice that some large language models [4, 7] employ new operators in Transformers to improve performance. For example, LLaMA [59] currently uses RMSNorm and SiLU to prevent training crashes. These successes inspire us to assemble a series of advanced operators to enhance linear attention and explore their optimal combination formulations. However, the question remains: How can we efficiently solve this combination problem? Recent developments in AutoML, such as AutoML-Zero [50] and NAS methods [40, 46], offer a promising solution. These methods involve constructing an extensive search space encompassing various options and searching for optimal architectures. Nevertheless, existing NAS methods are not suitable for solving the formulation discovery problem in our case for several reasons. (1) Different search space. The search spaces in conventional NAS methods typically focus on network width/depth, block/cell types, and parametric operators [5, 56]. In contrast, our attention search space includes many non-parametric operators and different combinatorial types. This leads to an explosion in the search space and many unsuccessful candidates. (2) Different search algorithms. Most recent NAS methods employ weight-sharing strategies [14, 27] to reduce search costs. However, our search space is explosive and extremely sparse, meaning that using previous ways to find potential candidates is less efficient and effective. (3) Different search goals. Most NAS methods prioritize performance for a single task and may struggle to generalize well to different applications [65]. In contrast, our objective is to discover generalized attention mechanisms that can be applied to various ViT models and tasks. These challenges highlight that attention discovery is a new task that requires a unique search space and strategy compared to traditional NAS methods.", + "bbox": [ + 212, + 412, + 787, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above observations, we propose AttnZero, a new framework that aims to discover superior linear attention formulations that achieve a better", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63c60c60f0fc55f05083d1da86342b05f18793ce17e402d4d2b05d7c6a217c71.jpg", + "image_caption": [ + "Figure 1: Overall pipeline of AttnZero. Our framework obtains candidate attention populations by sampling from attention search space. With program checking and rejection protocols, we conduct validation results of these attentions on multiple ViTs. Then, we use NSGA-II to obtain the optimal solution and cross-mutate for the next population." + ], + "image_footnote": [], + "bbox": [ + 218, + 141, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "accuracy-efficiency tradeoff. As shown in Figure 1, our framework can be divided into two parts: a comprehensive attention search space and a multi-objective evolutionary search. Based on the basic form of linear attention, we develop six computation graphs to represent attention candidates in the search space. Then, we combine unary and binary operators with advanced functions, including various activation functions, normalization techniques, and matrix arithmetic functions. This approach allows us to obtain flexible and effective attention candidates. Our goal for the search algorithm is to find attention formulations applicable across different ViT models. To achieve this, we model the search task as a multi-objective optimization problem and employ the NSGA-II evolution method to obtain optimal solutions. During the search process, we validate candidate attention in multiple ViT models through multi-architecture evaluations. Based on these evaluations, our search engine exploits crossover and mutation for better populations. To accelerate the search process, we perform a fast program check and a rejection protocol to eliminate unpromising candidates during the search process. In this way, our AttnZero improves at least $5 \\times$ in search efficiency without the weight-sharing setting [22], making our method free of ranking issues in traditional NAS methods [9, 18].", + "bbox": [ + 212, + 414, + 787, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments to evaluate the transferability of our discovered attention mechanisms across different ViT models, datasets, and tasks without requiring additional searches. When applied to various ViT models, our AttnZero consistently outperforms the baseline method. For instance, our method achieves gains of $3.0\\% \\sim 1.6\\%$ on PVT-T to PVT-L on the ImageNet dataset. Furthermore, ViT models with AttnZero demonstrate significant improvements when considering different data domains. For example, DeiT+AttnZero achieves gains of $6.2\\%$ , $7.8\\%$ , $1.1\\%$ , and $2.7\\%$ on the Tiny-ImageNet, Flowers, Chaoyang, and ImageNet datasets, respectively. Additionally, we compile a new benchmark named Attn-Bench-101, which consists of 2,000 attentions from our", + "bbox": [ + 212, + 689, + 802, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "AttnZero: Efficient Attention Discovery for Vision Transformers", + "bbox": [ + 305, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "search space. The benchmark highlights the preference for advanced normalization techniques, activation functions, and well-performing attentions. In conclusion, our contributions can be summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To improve linear attention, we present a new task, efficient attention discovery, that generalizes to various ViT models and tasks. For this task, our AttnZero, the first framework, is developed to achieve effective attention search.", + "- We propose a comprehensive attention search space, including different types of computation graphs and advanced operators as options. We develop multi-objective searches with multiple ViTs evaluations to ensure generalisability.", + "- Our found attention can be well generalized to DeiT|PVT|Swin|CSwin and other ViTs. Extensive experiments on multiple datasets in classification, detection, and segmentation prove the superior performance of our AttnZero.", + "- We build Attn-Bench-101, providing many attentions and their results. Some attention design guidances are summarised based on these results." + ], + "bbox": [ + 225, + 205, + 784, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 412, + 397, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient Attention. The application of self-attention in computer vision has made significant progress [2,8,26,68,69] with the introduction of ViT [19]. Then, PVT [60] introduces a hierarchical structure while Swin Transformer [41] proposes window-based attention to model dependencies efficiently. However, softmax attention's quadratic complexity poses challenges. Linear attention reduces this but often degrades performance [6,30,64]. Efficient Attention [54] applies softmax to Q and K while SOFT [42] uses matrix decomposition. Castling-ViT [66] employs softmax attention during training but linear attention at inference. FLatten [24] preserves diversity through focused functions. While effective, these methods still have performance limitations and require complex manual design. In contrast, our AttnZero automatically discovers novel formulations to enhance linear attention. Rather than replacing prior work, we aim to improve the performance of the existing method while reducing manual effort. We build our search space based on previous work and provide feedback on new patterns. Thus, AttnZero complements rather than competes with manual methods.", + "bbox": [ + 212, + 446, + 787, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Architecture Search (NAS). NAS methods automate model design through search spaces and algorithms [49,77]. Early NAS works adopt reinforcement learning with expensive costs [78]. Recent one-shot NAS [15] and zero-shot NAS [11-13,36,37,73,76] use techniques like weight sharing [23] and training-free proxy evaluation to speed up search. Current NAS methods on ViTs [5,45,72] (e.g., AutoFormer [5]) search the number of depth/head of transformers. In contrast, AttnZero focuses on discovering attention mechanisms, not architectural dimensions. We also apply AttnZero to AutoFormer and observe consistent improvements, indicating our method is orthogonal to existing ViT-NAS methods. There are attempts to search entire attention blocks [67], SE-like attention [43,61], projection layers [21], and Transformers searches in other fields like", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CTR [70] and NLP [20,55]. We clarify that our AttnZero is the first to search for efficient attention in visual tasks, which clearly differs in spaces/strategies/goals from previous methods. Inspired by general idea in AutoML-Zero [51], we achieve many new designs in search space/strategies for new attention discovery task, and opens up new possibilities for enhancing the capabilities of ViTs.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology: Efficient Attention Discovery", + "text_level": 1, + "bbox": [ + 214, + 246, + 671, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first review the linear attention form. Then, we present our search space and algorithms. The overall process is shown in Figure 1.", + "bbox": [ + 212, + 279, + 782, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Recap of Self-Attention and Linear Attention", + "text_level": 1, + "bbox": [ + 214, + 333, + 635, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In self-attention, input tokens are projected into query $(Q)$ , key $(K)$ , and value $(V)$ matrices, respectively. The general form of self-attention is given by:", + "bbox": [ + 214, + 359, + 782, + 390 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {i} = \\sum_ {j = 1} ^ {N} \\frac {\\operatorname {S i m} \\left(Q _ {i} , K _ {j}\\right)}{\\sum_ {j = 1} ^ {N} \\operatorname {S i m} \\left(Q _ {i} , K _ {j}\\right)} V _ {j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 401, + 784, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $O_{i}$ represents the output of the $i$ -th token and Sim denotes the similarity function. The commonly used similarity function is the softmax function, $\\exp\\left(QK^{T}/\\sqrt{d}\\right)$ , where $d$ is the dimension of the input tokens. However, the quadratic time and space complexity of softmax attention, $\\mathcal{O}(N^{2}d)$ , becomes impractical for large numbers of tokens ( $N$ ). To address this, linear attention is explored as an efficient alternative, which can be expressed as", + "bbox": [ + 212, + 448, + 784, + 546 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {i} = \\frac {\\phi \\left(Q _ {i}\\right) \\left(\\sum_ {j = 1} ^ {N} \\phi \\left(K _ {j}\\right) ^ {T} V _ {j}\\right)}{\\phi \\left(Q _ {i}\\right) \\left(\\sum_ {j = 1} ^ {N} \\phi \\left(K _ {j}\\right) ^ {T}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 559, + 784, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi$ represents the kernel function. Linear attention reduces the computational complexity to $\\mathcal{O}(N)$ , making it scalable for large input sequences. This reduction is achieved by using carefully designed kernels to approximate the similarity function, enabling a change in the computation order from $(QK^T)V$ to $Q(K^TV)$ , exploiting the associative property of matrix multiplication.", + "bbox": [ + 212, + 608, + 782, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Efficient Attention Search Space", + "text_level": 1, + "bbox": [ + 214, + 707, + 529, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention Representation. In Figure 2, we construct six types of computational graphs using (Q, K, V) as inputs to represent our candidate attentions. Among them, our i3n2 attentions follow the same process as standard linear attention, and we improve it by searching for different unary operations. Partially inspired by Castling-ViT [66], we introduce novel attention mechanisms with three binary operator nodes based on basic linear attention. With additional operators with either $\\mathcal{O}(N\\times d)$ or $\\mathcal{O}(d\\times d)$ computational complexity, the", + "bbox": [ + 212, + 734, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "AttnZero: Efficient Attention Discovery for Vision Transformers", + "bbox": [ + 305, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/51fd63600db8eba61318eb39388f26ad2caf22a7c44ed344c4bb43d7e4a3bb06.jpg", + "image_caption": [ + "Figure 2: Schematic of the six computational graphs in our search space. We denote the graph with 3 inputs and 2 binary operations as i3n2. Similarly, i3n3 refers to a graph with 3 inputs and 3 binary operations. Since each binary operation takes two inputs, there are 5 possible graph structures for i3n3. We label these i3n301~i3n305, respectively. For simplicity, we do not display unary operators here." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 756, + 292 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3dc12ee9974e742361b089770d54da96cb41fc9c04e3908225292d45897921eb.jpg", + "table_caption": [ + "Table 1: Some operations in our attention search space. \"UOP\" and \"BOP\" denote the unary and binary operation, respectively. Full operations are in the Appendix." + ], + "table_footnote": [], + "table_body": "
OP IDOP NameExpression
UOP00no_opx
UOP01relumax(0,x)
UOP02scalex/√dx
UOP03sqrt√x
UOP04invert1/x+e-6
UOP05l2_normx-mean(x)/std(x)
UOP06sigmoid1/1+e-x
UOP07logsoftmaxln ex/∑i=1n e^xi
UOP08softmaxex/∑i=1n e^xi
UOP09softsignx/1+|x|
", + "bbox": [ + 222, + 426, + 431, + 613 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1ea2e9c8fae53c5335b4c85e190dfcc3069753959027e1e638bf222b5f897af5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OP IDOP NameExpression
UOP10eluelu(x) + 1
UOP11sigmoid_revert1 - sigmoid(x)
UOP12expex
UOP13abslog|ln x|
UOP14min_max_normx-min(x)/max(x)-min(x)
UOP15transposexT
UOP16swishx × sigmoid (x)
BOP01sumx + y
BOP02cosine_simcos(x,y)
BOP03productx ⊙ y
BOP04matrix-multix · y
BOP05euclidean_disd(x,y)
", + "bbox": [ + 475, + 431, + 723, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performance of these attention mechanisms can be enhanced without a significant increase in computational overhead. Based on the different binary nodes, these augmented attentions can be categorized into five types (i3n301 to i3n305), effectively enhancing the modeling capability. We divide the sub-search space according to these different types of attention. During the search process, we first search the solutions within the sub-search space in parallel and then combine them to obtain the final search result. Figure 4 (b) shows performance for different attention candidates. These designs allow our method to balance computational complexity and accuracy.", + "bbox": [ + 212, + 643, + 787, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Primitive Operations. Table 1 presents some unary and binary operators in our search space. These operators encompass both commonly used operations found in existing linear attentions [3, 24, 54, 66] and our proposed ones, including: (1) The activation operations are responsible for enhancing the non-", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 NSGA-II Evolution Search for AttnZero.", + "text_level": 1, + "bbox": [ + 215, + 146, + 614, + 160 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input: Search space $\\mathcal{A}$ , population size $N$ , crossover probability $p_c$ , mutation probability $p_m$ , max generations $G$ .", + "bbox": [ + 215, + 162, + 785, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Output: Set of Pareto-optimal attention formulations.", + "bbox": [ + 215, + 190, + 576, + 203 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize population $\\mathcal{P}_0$ with $N$ individuals generated randomly;", + "2: Evaluate the fitness of each individual in $\\mathcal{P}_0$ for all objectives;", + "3: for $g = 1$ to $G$ do", + "4: Perform non-dominated sorting on $\\mathcal{P}_{g - 1}$ to classify into fronts $F_{1},F_{2},\\ldots$", + "5: Calculate crowding distance for individuals in each front;", + "6: $\\mathcal{Q} \\coloneqq \\operatorname{Select} N$ individuals for mating via front rank and crowding distance;", + "7: Generate offspring $\\mathcal{Q}'$ via crossover & mutation on $\\mathcal{Q}$ with prob. $p_c$ and $p_m$ ;", + "8: Evaluate the fitness of each offspring in $\\mathcal{Q}'$ for all objectives;", + "9: Apply rejection protocols to $\\mathcal{Q}'$ to remove unpromising candidate and get $\\mathcal{Q}''$ .", + "10: $\\mathcal{R} := \\operatorname{Combine} \\mathcal{P}_{g-1}$ and $\\mathcal{Q}'$ ;", + "11: Perform non-dominated sorting on $\\mathcal{R}$", + "12: Update $\\mathcal{P}_g$ by selecting top $N$ individuals via front rank and crowding distance;", + "13: end for", + "14: $\\mathcal{P}^* \\coloneqq$ Extract the first front from $\\mathcal{P}_G$ as the set of Pareto-optimal solutions;" + ], + "bbox": [ + 215, + 205, + 784, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "linearity of attention. We include relu in FLatten [24] and introduce new activation functions such as swish and elu. (2) Normalization operations ensure a smoother distribution of attention scores. We provide options such as $l2\\_norm$ and $min\\_max\\_norm$ . (3) The binary operators facilitate information interaction. Alongside cosine_similarity in Hydra [3], we also consider euclidean_distance as a simplified calculation option. As shown in Figure 4 (c), these operators enhance the versatility and adaptability of attention expressions, resulting in improved performance. Some complex operations such as decomposition [64] and Gaussian kernel [6] would result in a highly sparse search space or need hardware-specific supports. Hence they have not been included in our investigation. In future work, we will continue to involve more promising operators in our search space.", + "bbox": [ + 212, + 435, + 799, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Multi-objective Evolutionary Search", + "text_level": 1, + "bbox": [ + 214, + 643, + 560, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After building the search space $\\mathcal{A}$ , we sample the attention candidates $a \\in \\mathbb{N}$ and apply it to the ViT model with weights $W$ . For single-object search, the aim is to find the optimal attention $a_{single}^{*}$ by evaluating candidates using the trained ViT model $(W^{*})$ :", + "bbox": [ + 212, + 672, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\na _ {s i n g l e} ^ {*} = \\underset {a \\in \\mathcal {A}} {\\operatorname {a r g m a x}} \\mathrm {A C C} _ {\\text {v a l}} \\left(\\mathcal {N} (a, W _ {\\mathcal {A}} ^ {*} (a))\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 744, + 785, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, $a_{\\text{single}}^*$ may perform well on the search model but fail to other ViT models. For example, attention searched in Swin, while superior on Swin, performs poorly on DeiT, PVT in Figure 5 (b). Such attentions suffer from limited usefulness because they are difficult to transfer directly to different tasks.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "AttnZero: Efficient Attention Discovery for Vision Transformers", + "bbox": [ + 305, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-objective Evolution. To improve generalization of attention candidates in different ViT models, our AttnZero framework employs a multi-objective evolutionary algorithm. This algorithm aims to maximize the results of multiple ViT models $1, 2, \\ldots, n$ under predefined computational complexity constraints. We define a fitness function as follows:", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\na _ {m u l t i} ^ {*} = \\underset {a \\in \\mathcal {A}} {\\operatorname {a r g m a x}} \\mathrm {A C C} _ {\\text {v a l}} \\left(1 (a, W _ {1 \\mathcal {A}} ^ {*} (a)), 2 (a, W _ {2 \\mathcal {A}} ^ {*} (a)), \\dots , n (a, W _ {n \\mathcal {A}} ^ {*} (a)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 227, + 785, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our framework incorporates validation results of attention in DeiT, AutoFormer, PVT, and Swin as objectives. This ensures that our attention module performs effectively in various ViT models, including pure ViT, ViT models with NAS search, ViT models with hierarchical design, and ViT models with local window configurations, respectively. Solving this multi-objective problem is challenging due to obstacles like conflicting optimals. To address this, we utilize NSGA-II method [10] in Alg. 1 as our search engine. Our search begins with an initial population of candidate attention modules and each candidate's fitness is evaluated based on performance across ViT models. Non-dominated sorting ranks the candidates and assigns them to fronts based on how they compare to others. Then, we use crowding distance [48] calculations to maintain diversity within the population. The selection process chooses candidates for the next generation, considering sorting and distance. Genetic operations like crossover and mutation are applied to the selected candidates. This introduces diversity and explores new areas of the search space. The least fit candidates are replaced with offspring, ensuring continuous gains. The algorithm iterates through these steps until it meets a termination criterion. During the search process, our algorithm takes FLOPs as the complexity constraint. Finally, we consider the real speed results in selecting final found attention from the Pareto-optimal solutions.", + "bbox": [ + 212, + 255, + 787, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Search Acceleration Strategies. To improve the search efficiency, we employ several strategies for search acceleration: (1) Program checking. We perform program checking by verifying if the single candidate attention module can successfully generate outputs using random tensor inputs. This helps us identify and filter out candidates with arithmetic errors or dimension mismatches. (2) Rejection protocols. During training each population, we reject candidates with collapsed optimisation (e.g., loss = nan) and historical average validation results falling far below baseline linear attention. These strategies allow more than $5 \\times$ search acceleration.", + "bbox": [ + 212, + 542, + 602, + 739 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17e2105aea9e874689a4ff32514b6b417c03aa985fd5f751f138a3dd9f23b47b.jpg", + "image_caption": [ + "Figure 3: Discovered attention $a_{\\text{AttnZero}}^*$ ." + ], + "image_footnote": [], + "bbox": [ + 625, + 560, + 772, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Discovered AttnZero Analysis", + "text_level": 1, + "bbox": [ + 215, + 758, + 509, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Formulas of discovered AttnZero. Figure 3 illustrates the optimal attention $a_{\\text{AttnZero}}^*$ in our search, whose detailed expressions are as follows:", + "bbox": [ + 212, + 780, + 787, + 811 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\na _ {A t t n Z e r o} ^ {*} = Q \\times \\eta_ {2} (\\eta_ {2} (Q) \\eta_ {2} (\\phi_ {1} (K) ^ {T} \\eta_ {1} (V))) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 816, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/520b46eac29d9fe4c997c4df847ec4863c00b63b6396bb20be371afb485809be.jpg", + "image_caption": [ + "Figure 4: (a) Performance distribution of subsets of Attn-Bench-101 (AutoFormer). (b) Results of various types of Attention. (c) Results of our default search space, default search space without advanced normalization operations, activation operations, and binary operations that we have newly introduced. More details are in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 411, + 262 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1d0eb401a8c88a87220f4d7bc85f5415ec4969ce51805676e1093ce9a119b45a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 146, + 598, + 262 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/062ebd17bcd295d0915d827f00f040f3e9c5abae1160fac9b879f827065a4bd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 146, + 784, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\phi_{1}$ utilizes the Exponential Linear Unit (ELU) with a shift of 1, introducing nonlinearities into the attention computation. The $\\eta_{1}$ performs min-max normalization, scaling the matrix values to a range between 0 and 1. This normalization step ensures that the attention weights are properly distributed and avoids potential issues with extreme values. Furthermore, the l2 normalization function $\\eta_{2}$ guarantees that the attention weights have a unit norm, facilitating better interpretation and utilization of the attention mechanism. Regarding computational complexity, AttnZero requires triple matrix multiplications (i.e., $\\mathcal{O}(d\\times d + N\\times d + N\\times d)$ . Despite this, the overall time and memory complexity of AttnZero remains $\\mathcal{O}(N)$ , which ensures AttnZero to handle computations for sequences of varying lengths efficiently.", + "bbox": [ + 212, + 359, + 787, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Analysis of Attn-Bench-101. Our Attn-Bench-101 collects 2,000 attention configurations and their training results in DeiT/AutoFormer/PVT/Swin on CIFAR-100 (see Appendix for details). We analyze some results on AutoFormer to understand better our search space: (1) In Figure 4 (a), we observe that our search space contains many candidates with stronger performance than baseline model. (2) Among various computed graphs, we identify i3n302 and i3n303, superior to the other types (see in Figure 4 (b)). This is attributed to their augmented topology allowing for a good tradeoff between performance and efficiency. (3) For ablations in Figure 4 (c), advanced activation and normalization operations offer more contributions compared to new binary operators.", + "bbox": [ + 212, + 526, + 787, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Guidance for Efficient Attention Design. Based on the above analysis, we can summarize that we prefer attention types of i3n302 & i3n303 and combinations of advanced normalization and activation operations in the detailed design.", + "bbox": [ + 212, + 676, + 787, + 723 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Search Experiments", + "text_level": 1, + "bbox": [ + 214, + 748, + 444, + 766 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Search Implementation and Results", + "text_level": 1, + "bbox": [ + 214, + 782, + 555, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation. Our attention search is performed on CIFAR-100, which are scaled to $224 \\times 224$ resolution. We search using the validation results to ensure", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "AttnZero: Efficient Attention Discovery for Vision Transformers", + "bbox": [ + 305, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/621aeb061e4e4d97300bc63b6c2f55a886325dc65605271b829bcd801dab81af.jpg", + "table_caption": [ + "Table 2: Top-1 accuracy (\\%) of hand-crafted attentions (e.g., Linear [28] and FLatten [24]) and attention with Random NAS method in our search space on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
ModelsBaselineLinear [28]FLatten [24]Random (NAS)AttnZero (NAS)
DeiT-T65.0857.8072.4576.0477.68 (+12.60)
AutoFormer66.5864.5875.4276.9478.61 (+12.03)
PVT-T67.4266.5874.8373.1276.68 (+9.26)
Swin-T68.2569.4574.3174.5875.90 (+7.65)
", + "bbox": [ + 225, + 184, + 776, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "fair comparisons, and our validation set does not overlap with the test set. For search settings, we configure $(N, p_c, p_m, G)$ in Alg. 1 as (20,0.9,0.1,100) for parallel search on DeiT|AutoFormer|PVT|Swin. To obtain validation results, we adopt standard training settings [31] (e.g., 300 epochs and AdamW optimizer). With our search acceleration strategies, the overall search only costs 1.05 days on the $8 \\times \\mathrm{V}100$ . More details are available in the Appendix.", + "bbox": [ + 212, + 301, + 782, + 391 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison Results with Hand-Crafted Attentions. The results in Table 2 demonstrate the effectiveness of our AttnZero in consistently achieving stable boosts across different models. Firstly, when compared to the baseline, our AttnZero achieves significant improvements. For example, the DeiT-T model with our AttnZero improves from a baseline accuracy of $65.08\\%$ to $77.68\\%$ . Similar improvements are observed for the AutoFormer, PVT-T, and Swin-T models, indicating the consistent effectiveness of our approach. Furthermore, our AttnZero shows improvements over the recent state-of-the-art method. For example, the gain over FLatten for the DeiT-T model is $5.2\\%$ . These findings emphasize the potential and significance of our AttnZero approach in improving the performance of various models.", + "bbox": [ + 212, + 392, + 784, + 556 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparing Random NAS Method. It is not easy to directly compare our method to other NAS methods because we use different search spaces and do not have weight sharing. To analyze our search strategy, we compare it to random search in our same search space. The results in Table 2 show that our AttnZero consistently improves model performance compared to random search. We achieve increases in accuracy ranging from $1.32\\%$ to $3.56\\%$ depending on the model. This demonstrates that our search strategy is more effective than a random search.", + "bbox": [ + 212, + 559, + 787, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 705, + 388, + 720 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Multi-Objective Search (MOS) vs. Single-Objective Search (SOS). We employ MOS to enable the discovery of attentions that exhibit strong generalization capabilities. Figure 5 (a) illustrates the optimal candidate resulting from our MOS process. The results demonstrate that the Top-3 candidates outperform the others consistently across various ViT models. In Figures 5 (b) and (c), we compare the generalization performance of MOS and SOS across different ViT models and datasets, respectively. The results indicate that while the SOS", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/866aef81ba3ff9d806d18681fc1ec2c3e9ef54b315cfae3221f34997ff6080bc.jpg", + "image_caption": [ + "Figure 5: (a) Candidates in multi-objective search. (b) Comparisons of multi-objective vs. single-objective on multiple architectures; (c) Comparisons of multi-objective vs. single-objective on multiple datasets. (d) Search curves of various settings." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 352, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/91972eb13952e83dca49bf8fbebe25537795c895ee2b32c7a69e8e7998adbf45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 143, + 475, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9f414fea34a7b711a1d59acd34d3941d88b3711145e4fbf847e83a01e7201f3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 143, + 602, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/84e7f537f864be2cfba5c99218f9e6f960184e189f830c407442e5ab42f073a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 143, + 784, + 244 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f62ecfe39aca8295b3d8de395364707c624e2e4a8be18f7139d68e163a6ceacd.jpg", + "table_caption": [ + "Table 3: Results of AttnZero for different ViT models and datasets." + ], + "table_footnote": [], + "table_body": "
ModelTop-1 Acc. (%)
Tiny ImageNet [1]Flowers [44]Chaoyang [74]
DeiT-T53.6250.0682.00
DeiT-T+AttnZero59.91 (+6.29)57.89 (+7.83)83.12 (+1.12)
AutoFormer-T56.3854.9882.84
AutoFormer-T+AttnZero62.56 (+6.18)61.58 (+6.60)83.68 (+0.84)
PVT-T58.4758.5782.46
PVT-T+AttnZero63.10 (+4.63)64.30 (+5.73)84.57 (+2.11)
Swin-T60.2258.8582.98
Swin-T+AttnZero64.55 (+4.33)65.13 (+6.28)85.11 (+2.13)
", + "bbox": [ + 215, + 339, + 787, + 497 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "method may yield better results on the searched models, MOS consistently outperforms SOS over a wider range of ViT architectures. Similarly, our AttnZero approach with MOS consistently demonstrates performance benefits across different dataset domains. For instance, our AttnZero with MOS achieves a $0.9\\%$ improvement over SOS on the ImageNet dataset.", + "bbox": [ + 212, + 527, + 787, + 602 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation of the Search Algorithm. In Figure 5 (d), we examine the evolutionary search and search acceleration strategies. The search results reveal the effectiveness of our acceleration strategy (i.e., program checking and rejection protocols) in filtering out unpromising candidates and greatly expediting the algorithm's convergence. These search curves also indicate that the evolutionary algorithm outperforms random search regarding efficiency in finding high-quality solutions within our framework.", + "bbox": [ + 212, + 603, + 787, + 708 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Transfer Experiments", + "text_level": 1, + "bbox": [ + 215, + 734, + 464, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Experimental Results on Tiny Datasets", + "text_level": 1, + "bbox": [ + 214, + 767, + 586, + 782 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation. To check if our discovered attention generalizes beyond the CIFAR-100 dataset that we searched on, we it to multiple datasets (i.e., Tiny ImageNet [1], Flowers [44], and Chaoyang [74]). We train DeiT|AutoFormer|PVT|Swin", + "bbox": [ + 212, + 794, + 803, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "AttnZero: Efficient Attention Discovery for Vision Transformers", + "bbox": [ + 305, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/feaccb8270bc685f4c85bd07d616372e7d4867cfffa599388e100038db3faac7.jpg", + "table_caption": [ + "Table 4: Results of AttnZero for DeiT & PVT on ImageNet $(224\\times 224)$" + ], + "table_footnote": [], + "table_body": "
MethodParamsFLOPsTop-1 (%)
DeiT-T [58]5.7M1.2G72.2
DeiT-T+AttnZero5.7M1.0G74.9 (+2.7)
PVT-T [60]13.2M1.9G75.1
PVT-T+AttnZero10.9M1.8G78.1 (+3.0)
PVT-S24.5M3.8G79.8
PVT-S+AttnZero19.8M3.6G81.4 (+1.6)
PVT-M44.2M6.7G81.2
PVT-M+AttnZero34.6M6.6G83.0 (+1.8)
PVT-L61.4M9.8G81.7
PVT-L+AttnZero47.1M9.7G83.3 (+1.6)
", + "bbox": [ + 218, + 184, + 491, + 320 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bc91a5f40fc87c648167841e14aaae841381cd858b2af305b2c25034d986b604.jpg", + "table_caption": [ + "Table 5: Results of AttnZero on ImageNet with $224\\times 224$ resolution inputs." + ], + "table_footnote": [], + "table_body": "
MethodParamsFLOPsTop-1 (%)
AutoFormer-T [58]5.7M1.3G74.7
AutoF.-T+AttnZero5.7M1.1G75.2 (+0.5)
Swin-T [41]29.0M4.5G81.3
Swin-T+AttnZero28.2M4.3G82.1 (+0.8)
Swin-S50.0M8.7G83.0
Swin-S+AttnZero49.5M8.5G83.2 (+0.2)
Swin-B88.0M15.4G83.5
Swin-B+AttnZero89.0M15.3G83.7 (+0.2)
CSwin-T [17]23.0M4.3G82.7
CSwin-T+AttnZero20.3M4.0G82.9 (+0.2)
", + "bbox": [ + 509, + 184, + 782, + 320 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/738b272af250039f96d0a86717eeb5a17863e8a3c790b3d4e5115deb4eebec9b.jpg", + "table_caption": [ + "Table 6: Comparison of different efficient attention methods on ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodResoParamsFLOPsTop-1 (%)
DeiT-T [58]22425.7M1.2G72.2
DeiT-T+Hydra Attn [3]22425.7M1.1G68.3
DeiT-T+Efficient Attn [54]22425.7M1.1G70.2
DeiT-T+Linear Angular [66]22425.7M1.1G70.8
DeiT-T+FLatten22426.1M1.1G74.1
DeiT-T+AttnZero22425.7M1.0G74.9 (+2.7)
", + "bbox": [ + 277, + 359, + 720, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "using discovered attention from scratch. The training process follows standard settings [31], including 300 training epochs, a cosine learning rate scheduler, and the AdamW optimizer. Detailed settings are shown in the Appendix.", + "bbox": [ + 212, + 503, + 785, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on Various Datasets. Table 3 highlights the effectiveness of AttnZero across various ViT models and datasets. It consistently surpasses the baselines, resulting in a $6.29\\%$ increase in accuracy for the DeiT-T model on Tiny-ImageNet. Similarly, AutoFormer and PVT-T models with AttnZero exhibit similar improvements. On the Flowers dataset, AttnZero achieves an accuracy of $57.89\\%$ for the DeiT-T model, surpassing the baseline by $7.83\\%$ . AutoFormer and PVT-T models also benefit from AttnZero. The results on the Chaoyang dataset further exemplify AttnZero's capability to enhance accuracies. On the DeiT-T model, AttnZero achieves an accuracy of $83.12\\%$ , compared to the $82\\%$ baseline. Moreover, the attention modules from AttnZero contribute to the improved performance of AutoFormer, PVT-T, and Swin-T models compared to their original versions. These findings demonstrate the promising ability of AttnZero across different datasets and models.", + "bbox": [ + 212, + 547, + 787, + 744 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Experimental Results on Large-scale Datasets", + "text_level": 1, + "bbox": [ + 214, + 768, + 640, + 784 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Implementation. We conduct the experiment on the ImageNet [52] dataset, a widely used large-scale dataset consisting of 1000 classes and 1.2 million training images. Note that we do not perform additional searches or modify the exist-", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2ae9fa38a4258ddaab24c505348c7b6d44ef7e010590f1e49c3c41793645129d.jpg", + "table_caption": [ + "Table 7: Results for Mask R-CNN on COCO dataset. We compute the FLOPs across the backbone, FPN, and detection head using an input resolution of $1280 \\times 800$ ." + ], + "table_footnote": [], + "table_body": "
MethodFLOPsSch.\\( AP^b \\)\\( AP_{50}^b \\)\\( AP_{75}^b \\)\\( AP^m \\)\\( AP_{50}^m \\)\\( AP_{75}^m \\)
PVT-T240G1x36.759.239.335.156.737.3
PVT-T+FLatten244G1x38.261.641.937.057.639.0
PVT-T+AttnZero209G1x40.062.643.137.459.440.1
PVT-S305G1x40.462.943.837.860.140.3
PVT-S+AttnZero248G1x41.063.445.038.260.441.1
", + "bbox": [ + 267, + 185, + 730, + 277 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ing classical models. Instead, we directly replace the self-attention mechanism with our discovered attention module. Furthermore, we ensure that our training settings remain consistent with the standard settings [17, 41, 58]. This includes training the models for 300 epochs, cosine learning rate scheduler, and AdamW optimizer. Specific training settings are shown in the Appendix.", + "bbox": [ + 212, + 311, + 782, + 387 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results on Various Backbones. Table 4 and Table 5 provides the performance results of our AttnZero on ImageNet. The results show that our method can transfer effectively to different ViT models. When applied to DeiT-T, AttnZero achieves a significant $2.7\\%$ gain over baseline. Similarly, AttnZero improves PVT-T accuracy to $78.1\\%$ , surpassing the baseline by $3.0\\%$ . This demonstrates that AttnZero can seamlessly migrate and consistently enhance performance across ViT architectures. AttnZero also provides stable gains across models of varying scales. Paired with PVT-S, it boosts accuracy by $1.6\\%$ . Applied to larger models like PVT-M, PVT-L, Swin-T, Swin-S, and CSwin-T, improvements of $1.8\\%$ , $1.6\\%$ , $0.8\\%$ , $0.2\\%$ and $0.2\\%$ respectively are observed. These findings highlight AttnZero's stability in boosting the performance of ViT models from small to large scale. These results highlight the stability of our AttnZero in enhancing the performance of ViT models across different scales.", + "bbox": [ + 212, + 388, + 784, + 583 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparison with Other Linear Attention. Table 6 compares various linear attention methods. The results demonstrate that AttnZero surpasses other linear attention methods and baseline when incorporated into the DeiT-T model. Notably, compared to the Softmax baseline, AttnZero achieves a $2.7\\%$ improvement. In contrast, other linear attention methods, such as Hydra Attn, Efficient Attn, Linear Angular, and FLatten, exhibit lower accuracies. These findings highlight the superiority of AttnZero in preserving or surpassing the performance achieved by Softmax attention. On the other hand, other linear attention methods tend to result in notable performance degradation.", + "bbox": [ + 212, + 585, + 785, + 722 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Transfer Experiments on Detection and Segmentation", + "text_level": 1, + "bbox": [ + 215, + 750, + 705, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Object Detection. We evaluate our methods for object detection and instance segmentation tasks on the COCO dataset [39]. Table 7 provides results on Mask R-CNN detector [25] with $1 \\times$ schedules. Our proposed PVT-T+AttnZero model achieves 2.3 box AP gains on PVT-T and significantly outperforms FLatten.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "AttnZero: Efficient Attention Discovery for Vision Transformers", + "bbox": [ + 305, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2ecdc32caf798eeb16f08d5b9b724b31dd92e20e41d66e8aad068079d24e2908.jpg", + "table_caption": [ + "Table 8: Results of semantic segmentation on ADE20K. We compute the FLOPs for both the encoders and decoders using input images at a resolution of $512 \\times 2048$ ." + ], + "table_footnote": [], + "table_body": "
BackboneMethodFLOPs#ParamsmIoUmAcc
PVT-TSemanticFPN158G17M36.5746.72
PVT-T+FLattenSemanticFPN169G16M37.2148.95
PVT-T+AttnZeroSemanticFPN126G15M39.1650.30
PVT-SSemanticFPN225G28M41.9553.02
PVT-S+AttnZeroSemanticFPN165G23M42.6254.02
", + "bbox": [ + 272, + 185, + 725, + 277 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "These enhancements demonstrate the benefits of incorporating our found attention module within the popular ViT model for object detection tasks.", + "bbox": [ + 215, + 306, + 782, + 335 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Semantic Segmentation. We assess the performance of our model using the ADE20K dataset [71]. Table 8 provides semantic segmentation results on SemanticFPN [29]. Among them, the Flatten module yields slight gains on the baseline PVT-T model. However, our AttnZero model achieves even higher mIoU (39.16) and mAcc (50.30) scores while reducing the FLOPs and parameters. Similarly, AttnZero achieves clear gains on PVT-S. These results highlight the compatibility of our AttnZero with different backbone architectures and its consistent improvements across various scenarios.", + "bbox": [ + 215, + 335, + 784, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 476, + 356, + 493 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce AttnZero, the first method that automates the discovery of efficient attention modules that can be applied to various ViT models. AttnZero constructs a comprehensive search space consisting of different unary and binary operators as potential attention mechanisms. It utilizes a multi-objective genetic algorithm to search within this design space efficiently. Additionally, we introduce Attn-Bench-101, which provides attention profile results to summarize the insights gained from the search. Extensive experiments demonstrate that the attention modules discovered by AttnZero can be successfully applied to advanced ViTs for image classification, segmentation, and detection tasks. For future work, we will consider extending the AttnZero with knowledge distillation methods [32-35, 38, 53, 63]. Limitation. We present search strategies but still involve some search costs following most AutoML methods. But these costs are worthwhile because our discovered attention can be effectively applied to different ViTs and tasks without additional searching (proven by our extensive transfer experiments). Social impact. Our AttnZero focuses on technology improvement without social and ethical implications.", + "bbox": [ + 215, + 505, + 785, + 746 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 766, + 401, + 784 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The research was supported by Theme-based Research Scheme (T45-205/21-N) from Hong Kong RGC, and Generative AI Research and Development Centre from InnoHK.", + "bbox": [ + 215, + 794, + 782, + 838 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. http://tiny-imagenet.herokuapp.com/", + "2. Ali, A., Touvron, H., Caron, M., Bojanowski, P., Douze, M., Joulin, A., Laptev, I., Neverova, N., Synnaeve, G., Verbeek, J., et al.: Xcit: Cross-covariance image transformers. NIPS (2021)", + "3. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Hoffman, J.: Hydra attention: Efficient attention with many heads. In: ECCVW (2022)", + "4. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. arXiv preprint arXiv:2005.14165 (2020)", + "5. Chen, M., Peng, H., Fu, J., Ling, H.: Autoformer: Searching transformers for visual recognition. In: ICCV. pp. 12270-12280 (2021)", + "6. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. In: ICLR (2021)", + "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. JMLR (2023)", + "8. Chu, X., Tian, Z., Zhang, B., Wang, X., Wei, X., Xia, H., Shen, C.: Conditional positional encodings for vision transformers. Arxiv preprint 2102.10882 (2021), https://arxiv.org/pdf/2102.10882.pdf", + "9. Chu, X., Zhang, B., Xu, R.: Fairnas: Rethinking evaluation fairness of weight sharing neural architecture search. In: ICCV (2021)", + "0. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary computation (2002)", + "1. Dong, P., Li, L., Tang, Z., Liu, X., Pan, X., Wang, Q., Chu, X.: Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. In: ICML (2024)", + "2. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023)", + "3. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV. pp. 17076-17086 (2023)", + "4. Dong, P., Niu, X., Li, L., Tian, Z., Wang, X., Wei, Z., Pan, H., Li, D.: Rd-nas: Enhancing one-shot supernet ranking ability via ranking distillation from zero-cost proxies. ICASSP (2023)", + "5. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022)", + "6. Dong, P., Niu, X., Tian, Z., Li, L., Wang, X., Wei, Z., Pan, H., Li, D.: Progressive meta-pooling learning for lightweight image classification model. In: ICASSP (2023)", + "7. Dong, X., Bao, J., Chen, D., Zhang, W., Yu, N., Yuan, L., Chen, D., Guo, B.: Cswin transformer: A general vision transformer backbone with cross-shaped windows. ArXiv abs/2107.00652 (2021)", + "8. Dong, X., Yang, Y.: Nas-bench-201: Extending the scope of reproducible neural architecture search (2020)", + "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "AttnZero: Efficient Attention Discovery for Vision Transformers", + "bbox": [ + 305, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Gao, J., Xu, H., Shi, H., Ren, X., Philip, L., Liang, X., Jiang, X., Li, Z.: Autobert-zero: Evolving bert backbone from scratch. In: AAAI (2022)", + "21. Guan, C., Wang, X., Zhu, W.: Autoattend: Automated attention representation search. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 3864-3874. PMLR (18-24 Jul 2021), http://proceedings.mlr.press/v139/guan21a.html", + "22. Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. arXiv preprint arXiv:1904.00420 (2019)", + "23. Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. In: ECCV. vol. abs/1904.00420, pp. 544-560. Springer (2019)", + "24. Han, D., Pan, X., Han, Y., Song, S., Huang, G.: Flatten transformer: Vision transformer using focused linear attention. In: ICCV (2023)", + "25. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: ICCV (2017)", + "26. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021)", + "27. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021)", + "28. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: ICML (2020)", + "29. Kirillov, A., Girshick, R., He, K., Dollár, P.: Panoptic feature pyramid networks. In: CVPR (2019)", + "30. Kitaev, N., Kaiser, L., Levskaya, A.: Reformer: The efficient transformer. In: ICLR (2020)", + "31. Li, K., Yu, R., Wang, Z., Yuan, L., Song, G., Chen, J.: Locality guidance for improving vision transformers on tiny datasets. In: ECCV. pp. 110-127. Springer (2022)", + "32. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022)", + "33. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024)", + "34. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023)", + "35. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022)", + "36. Li, L., Sun, H., Dong, P., Wei, Z., Shao, S.: Auto-das: Automated proxy discovery for training-free distillation-aware architecture search. In: ECCV (2024)", + "37. Li, L., Sun, H., Li, S., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Auto-gas: Automated proxy discovery for training-free generative architecture search. In: ECCV (2024)", + "38. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation (2020)", + "39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)", + "40. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019)", + "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. In: NeurIPS (2021)", + "43. Nakai, K., Matsubara, T., Uehara, K.: Att-darts: Differentiable neural architecture search for attention. In: IJCNN. IEEE", + "44. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008)", + "45. Peng, H., Du, H., Yu, H., Li, Q., Liao, J., Fu, J.: Cream of the crop: Distilling prioritized paths for one-shot neural architecture search. NIPS (2020)", + "46. Pham, H., Guan, M.Y., Zoph, B., Le, Q.V., Dean, J.: Efficient neural architecture search via parameter sharing. In: ICML (2018)", + "47. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI (2022)", + "48. Raquel, C.R., Naval Jr, P.C.: An effective use of crowding distance in multiobjective particle swarm optimization. In: CGEC (2005)", + "49. Real, E., Aggarwal, A., Huang, Y., Le, Q.V.: Regularized evolution for image classifier architecture search. In: AAAI (2019)", + "50. Real, E., Liang, C., So, D., Le, Q.: Automl-zero: Evolving machine learning algorithms from scratch. In: ICML (2020)", + "51. Real, E., Liang, C., So, D., Le, Q.: Automl-zero: Evolving machine learning algorithms from scratch. In: ICML (2020)", + "52. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Li, F.F.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision (2015)", + "53. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023)", + "54. Shen, Z., Zhang, M., Zhao, H., Yi, S., Li, H.: Efficient attention: Attention with linear complexities. In: WACV (2021)", + "55. So, D.R., Manke, W., Liu, H., Dai, Z., Shazeer, N., Le, Q.V.: Primer: Searching for efficient transformers for language modeling. arXiv preprint arXiv:2109.08668 (2021)", + "56. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)", + "57. Tay, Y., Dehghani, M., Bahri, D., Metzler, D.: Efficient transformers: A survey. arXiv preprint arXiv:2009.06732 (2020)", + "58. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021)", + "59. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. corr, abs/2302.13971, 2023. doi: 10.48550. arXiv preprint arXiv.2302.13971", + "60. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021)", + "61. Wang, X., Xiong, X., Neumann, M., Piergiovanni, A., Ryoo, M.S., Angelova, A., Kitani, K.M., Hua, W.: Attentionnas: Spatiotemporal attention cell search for video classification. In: ECCV (2020)", + "62. Wei, Z., Pan, H., Li, L.L., Lu, M., Niu, X., Dong, P., Li, D.: Convformer: Closing the gap between cnn and vision transformers. arXiv preprint arXiv:2209.07738 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "AttnZero: Efficient Attention Discovery for Vision Transformers", + "bbox": [ + 305, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching (2022)", + "64. Xiong, Y., Zeng, Z., Chakraborty, R., Tan, M., Fung, G., Li, Y., Singh, V.: Nystromformer: A nystrom-based algorithm for approximating self-attention. In: AAAI (2021)", + "65. Yang, A., Esperança, P.M., Carlucci, F.M.: Nas evaluation is frustratingly hard. arXiv preprint arXiv:1912.12522 (2019)", + "66. You, H., Xiong, Y., Dai, X., Wu, B., Zhang, P., Fan, H., Vajda, P., Lin, Y.C.: Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference. In: CVPR (2023)", + "67. You, S., Huang, T., Yang, M., Wang, F., Qian, C., Zhang, C.: Greedynas: Towards fast one-shot nas with greedy supernet. In: CVPR (2020)", + "68. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: CVPR (2022)", + "69. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token vit: Training vision transformers from scratch onImagenet. In: ICCV (2021)", + "70. Zheng, Z., Gao, X., Pan, J., Luo, Q., Chen, G., Liu, D., Jiang, J.: Autoattention: Automatic field pair selection for attention in user behavior modeling. In: 2022 IEEE International Conference on Data Mining (ICDM). pp. 803-812. IEEE (2022)", + "71. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019)", + "72. Zhou, Q., Sheng, K., Zheng, X., Li, K., Sun, X., Tian, Y., Chen, J., Ji, R.: Training-free transformer architecture search. In: CVPR. pp. 10894-10903 (2022)", + "73. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024)", + "74. Zhu, C., Chen, W., Peng, T., Wang, Y., Jin, M.: Hard sample aware noise robust learning for histopathology image classification. TMI", + "75. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)", + "76. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024)", + "77. Zoph, B., Le, Q.V.: Neural architecture search with reinforcement learning. In: ICLR (2017)", + "78. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 672 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_model.json b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fa65e479d6a81b8358a031f471c9f42dedccaaf7 --- /dev/null +++ b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_model.json @@ -0,0 +1,2777 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.185 + ], + "angle": 0, + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.212, + 0.768, + 0.244 + ], + "angle": 0, + "content": "Lujun Li\\(^{1}\\), Zimian Wei\\(^{2\\boxtimes}\\), Peijie Dong\\(^{3}\\), Wenhan Luo\\(^{1}\\), Wei Xue\\(^{1}\\), Qifeng Liu\\(^{1\\boxtimes}\\), and Yike Guo\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.255, + 0.689, + 0.269 + ], + "angle": 0, + "content": "1 The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.269, + 0.652, + 0.284 + ], + "angle": 0, + "content": "\\(^{2}\\) National University of Defense Technology" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.283, + 0.733, + 0.298 + ], + "angle": 0, + "content": "3 The Hong Kong University of Science and Technology (Guangzhou)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.298, + 0.812, + 0.326 + ], + "angle": 0, + "content": "lilujunai@gmail.com;weizimian16@nudt.edu.cn;pdong212@connect.hkust-gz.edu.cn;whluo.china@gmail.com;weixue@ust.hk;liuqifeng@ust.hk;yikeguo@ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.358, + 0.744, + 0.662 + ], + "angle": 0, + "content": "Abstract In this paper, we present AttnZero, the first framework for automatically discovering efficient attention modules tailored for Vision Transformers (ViTs). While traditional self-attention in ViTs suffers from quadratic computation complexity, linear attention offers a more efficient alternative with linear complexity approximation. However, existing hand-crafted linear attention suffers from performance degradation. To address these issues, our AttnZero constructs search spaces and employs evolutionary algorithms to discover potential linear attention formulations. Specifically, our search space consists of six kinds of computation graphs and advanced activation, normalize, and binary operators. To enhance generality, we derive results of candidate attention applied to multiple advanced ViTs as the multi-objective for the evolutionary search. To expedite the search process, we utilize program checking and rejection protocols to filter out unpromising candidates swiftly. Additionally, we develop Attn-Bench-101, which provides precomputed performance of 2,000 attentions in the search spaces, enabling us to summarize attention design insights. Experimental results demonstrate that the discovered AttnZero module generalizes well to different tasks and consistently achieves improved performance across various ViTs. For instance, the tiny model of DeiT|PVT|Swin|CSwin trained with AttnZero on ImageNet reaches \\(74.9\\%\\) |78.1%|82.1%|82.9% top-1 accuracy. Codes at: https://github.com/lliai/AttnZero." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.675, + 0.69, + 0.689 + ], + "angle": 0, + "content": "Keywords: Vision Transformer \\(\\cdot\\) Efficient Attention \\(\\cdot\\) AutoML" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.712, + 0.377, + 0.729 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.819 + ], + "angle": 0, + "content": "In recent years, Vision Transformers (ViTs) demonstrate remarkable success in various vision tasks, including image classification [16, 19, 58, 62], object detection [60, 75], semantic segmentation [47]. ViTs draw inspiration from the transformative impact of Transformers in natural language processing, utilizing self-attention mechanisms to capture global dependencies. However, self-attention" + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.826, + 0.412, + 0.841 + ], + "angle": 0, + "content": "Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "has a quadratic complexity that scales with the square of the input sequence length. This quadratic complexity limits the scalability of ViTs for handling large image inputs, thereby limiting their practical applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.409 + ], + "angle": 0, + "content": "To address this issue, researchers have been exploring the use of linear attention [57] as a more computationally efficient alternative. Unlike self-attention, linear attention has a linear complexity approximation, making it more suitable for handling large input sequences. However, existing linear attention methods [42, 54] suffer from several limitations: (1) Linear attention often leads to reduced performance than softmax attention due to its simplifying assumptions. The simple linear operators used in linear attention have limited flexibility and produce noisier activations than nonlinear functions [24]. These performance degradations challenge leveraging linear attention as a viable substitute for self-attention in ViTs. (2) Hand-crafted linear attention designs are often specific to certain tasks or architectures, lacking generality across different scenarios. Additionally, the manual design is labor-intensive, time-consuming, and heavily reliant on expert knowledge. Consequently, these drawbacks hinder the scalability and accessibility of linear attention in ViTs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.806 + ], + "angle": 0, + "content": "To address these challenges, we have made several observations. We notice that some large language models [4, 7] employ new operators in Transformers to improve performance. For example, LLaMA [59] currently uses RMSNorm and SiLU to prevent training crashes. These successes inspire us to assemble a series of advanced operators to enhance linear attention and explore their optimal combination formulations. However, the question remains: How can we efficiently solve this combination problem? Recent developments in AutoML, such as AutoML-Zero [50] and NAS methods [40, 46], offer a promising solution. These methods involve constructing an extensive search space encompassing various options and searching for optimal architectures. Nevertheless, existing NAS methods are not suitable for solving the formulation discovery problem in our case for several reasons. (1) Different search space. The search spaces in conventional NAS methods typically focus on network width/depth, block/cell types, and parametric operators [5, 56]. In contrast, our attention search space includes many non-parametric operators and different combinatorial types. This leads to an explosion in the search space and many unsuccessful candidates. (2) Different search algorithms. Most recent NAS methods employ weight-sharing strategies [14, 27] to reduce search costs. However, our search space is explosive and extremely sparse, meaning that using previous ways to find potential candidates is less efficient and effective. (3) Different search goals. Most NAS methods prioritize performance for a single task and may struggle to generalize well to different applications [65]. In contrast, our objective is to discover generalized attention mechanisms that can be applied to various ViT models and tasks. These challenges highlight that attention discovery is a new task that requires a unique search space and strategy compared to traditional NAS methods." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Based on the above observations, we propose AttnZero, a new framework that aims to discover superior linear attention formulations that achieve a better" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.307, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.142, + 0.788, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.785, + 0.379 + ], + "angle": 0, + "content": "Figure 1: Overall pipeline of AttnZero. Our framework obtains candidate attention populations by sampling from attention search space. With program checking and rejection protocols, we conduct validation results of these attentions on multiple ViTs. Then, we use NSGA-II to obtain the optimal solution and cross-mutate for the next population." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.687 + ], + "angle": 0, + "content": "accuracy-efficiency tradeoff. As shown in Figure 1, our framework can be divided into two parts: a comprehensive attention search space and a multi-objective evolutionary search. Based on the basic form of linear attention, we develop six computation graphs to represent attention candidates in the search space. Then, we combine unary and binary operators with advanced functions, including various activation functions, normalization techniques, and matrix arithmetic functions. This approach allows us to obtain flexible and effective attention candidates. Our goal for the search algorithm is to find attention formulations applicable across different ViT models. To achieve this, we model the search task as a multi-objective optimization problem and employ the NSGA-II evolution method to obtain optimal solutions. During the search process, we validate candidate attention in multiple ViT models through multi-architecture evaluations. Based on these evaluations, our search engine exploits crossover and mutation for better populations. To accelerate the search process, we perform a fast program check and a rejection protocol to eliminate unpromising candidates during the search process. In this way, our AttnZero improves at least \\(5 \\times\\) in search efficiency without the weight-sharing setting [22], making our method free of ranking issues in traditional NAS methods [9, 18]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.803, + 0.84 + ], + "angle": 0, + "content": "We conduct comprehensive experiments to evaluate the transferability of our discovered attention mechanisms across different ViT models, datasets, and tasks without requiring additional searches. When applied to various ViT models, our AttnZero consistently outperforms the baseline method. For instance, our method achieves gains of \\(3.0\\% \\sim 1.6\\%\\) on PVT-T to PVT-L on the ImageNet dataset. Furthermore, ViT models with AttnZero demonstrate significant improvements when considering different data domains. For example, DeiT+AttnZero achieves gains of \\(6.2\\%\\), \\(7.8\\%\\), \\(1.1\\%\\), and \\(2.7\\%\\) on the Tiny-ImageNet, Flowers, Chaoyang, and ImageNet datasets, respectively. Additionally, we compile a new benchmark named Attn-Bench-101, which consists of 2,000 attentions from our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "search space. The benchmark highlights the preference for advanced normalization techniques, activation functions, and well-performing attentions. In conclusion, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.206, + 0.784, + 0.265 + ], + "angle": 0, + "content": "- To improve linear attention, we present a new task, efficient attention discovery, that generalizes to various ViT models and tasks. For this task, our AttnZero, the first framework, is developed to achieve effective attention search." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.268, + 0.785, + 0.311 + ], + "angle": 0, + "content": "- We propose a comprehensive attention search space, including different types of computation graphs and advanced operators as options. We develop multi-objective searches with multiple ViTs evaluations to ensure generalisability." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.313, + 0.784, + 0.357 + ], + "angle": 0, + "content": "- Our found attention can be well generalized to DeiT|PVT|Swin|CSwin and other ViTs. Extensive experiments on multiple datasets in classification, detection, and segmentation prove the superior performance of our AttnZero." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.359, + 0.784, + 0.387 + ], + "angle": 0, + "content": "- We build Attn-Bench-101, providing many attentions and their results. Some attention design guidances are summarised based on these results." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.206, + 0.785, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.398, + 0.43 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Efficient Attention. The application of self-attention in computer vision has made significant progress [2,8,26,68,69] with the introduction of ViT [19]. Then, PVT [60] introduces a hierarchical structure while Swin Transformer [41] proposes window-based attention to model dependencies efficiently. However, softmax attention's quadratic complexity poses challenges. Linear attention reduces this but often degrades performance [6,30,64]. Efficient Attention [54] applies softmax to Q and K while SOFT [42] uses matrix decomposition. Castling-ViT [66] employs softmax attention during training but linear attention at inference. FLatten [24] preserves diversity through focused functions. While effective, these methods still have performance limitations and require complex manual design. In contrast, our AttnZero automatically discovers novel formulations to enhance linear attention. Rather than replacing prior work, we aim to improve the performance of the existing method while reducing manual effort. We build our search space based on previous work and provide feedback on new patterns. Thus, AttnZero complements rather than competes with manual methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Neural Architecture Search (NAS). NAS methods automate model design through search spaces and algorithms [49,77]. Early NAS works adopt reinforcement learning with expensive costs [78]. Recent one-shot NAS [15] and zero-shot NAS [11-13,36,37,73,76] use techniques like weight sharing [23] and training-free proxy evaluation to speed up search. Current NAS methods on ViTs [5,45,72] (e.g., AutoFormer [5]) search the number of depth/head of transformers. In contrast, AttnZero focuses on discovering attention mechanisms, not architectural dimensions. We also apply AttnZero to AutoFormer and observe consistent improvements, indicating our method is orthogonal to existing ViT-NAS methods. There are attempts to search entire attention blocks [67], SE-like attention [43,61], projection layers [21], and Transformers searches in other fields like" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.307, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "CTR [70] and NLP [20,55]. We clarify that our AttnZero is the first to search for efficient attention in visual tasks, which clearly differs in spaces/strategies/goals from previous methods. Inspired by general idea in AutoML-Zero [51], we achieve many new designs in search space/strategies for new attention discovery task, and opens up new possibilities for enhancing the capabilities of ViTs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.247, + 0.673, + 0.265 + ], + "angle": 0, + "content": "3 Methodology: Efficient Attention Discovery" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.784, + 0.311 + ], + "angle": 0, + "content": "In this section, we first review the linear attention form. Then, we present our search space and algorithms. The overall process is shown in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.334, + 0.636, + 0.349 + ], + "angle": 0, + "content": "3.1 Recap of Self-Attention and Linear Attention" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.36, + 0.784, + 0.391 + ], + "angle": 0, + "content": "In self-attention, input tokens are projected into query \\((Q)\\), key \\((K)\\), and value \\((V)\\) matrices, respectively. The general form of self-attention is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.402, + 0.785, + 0.44 + ], + "angle": 0, + "content": "\\[\nO _ {i} = \\sum_ {j = 1} ^ {N} \\frac {\\operatorname {S i m} \\left(Q _ {i} , K _ {j}\\right)}{\\sum_ {j = 1} ^ {N} \\operatorname {S i m} \\left(Q _ {i} , K _ {j}\\right)} V _ {j}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.785, + 0.547 + ], + "angle": 0, + "content": "where \\( O_{i} \\) represents the output of the \\( i \\)-th token and Sim denotes the similarity function. The commonly used similarity function is the softmax function, \\( \\exp\\left(QK^{T}/\\sqrt{d}\\right) \\), where \\( d \\) is the dimension of the input tokens. However, the quadratic time and space complexity of softmax attention, \\( \\mathcal{O}(N^{2}d) \\), becomes impractical for large numbers of tokens (\\( N \\)). To address this, linear attention is explored as an efficient alternative, which can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.56, + 0.785, + 0.606 + ], + "angle": 0, + "content": "\\[\nO _ {i} = \\frac {\\phi \\left(Q _ {i}\\right) \\left(\\sum_ {j = 1} ^ {N} \\phi \\left(K _ {j}\\right) ^ {T} V _ {j}\\right)}{\\phi \\left(Q _ {i}\\right) \\left(\\sum_ {j = 1} ^ {N} \\phi \\left(K _ {j}\\right) ^ {T}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.784, + 0.685 + ], + "angle": 0, + "content": "where \\(\\phi\\) represents the kernel function. Linear attention reduces the computational complexity to \\(\\mathcal{O}(N)\\), making it scalable for large input sequences. This reduction is achieved by using carefully designed kernels to approximate the similarity function, enabling a change in the computation order from \\((QK^T)V\\) to \\(Q(K^TV)\\), exploiting the associative property of matrix multiplication." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.53, + 0.724 + ], + "angle": 0, + "content": "3.2 Efficient Attention Search Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Attention Representation. In Figure 2, we construct six types of computational graphs using (Q, K, V) as inputs to represent our candidate attentions. Among them, our i3n2 attentions follow the same process as standard linear attention, and we improve it by searching for different unary operations. Partially inspired by Castling-ViT [66], we introduce novel attention mechanisms with three binary operator nodes based on basic linear attention. With additional operators with either \\(\\mathcal{O}(N\\times d)\\) or \\(\\mathcal{O}(d\\times d)\\) computational complexity, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.145, + 0.758, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.789, + 0.373 + ], + "angle": 0, + "content": "Figure 2: Schematic of the six computational graphs in our search space. We denote the graph with 3 inputs and 2 binary operations as i3n2. Similarly, i3n3 refers to a graph with 3 inputs and 3 binary operations. Since each binary operation takes two inputs, there are 5 possible graph structures for i3n3. We label these i3n301~i3n305, respectively. For simplicity, we do not display unary operators here." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.386, + 0.789, + 0.416 + ], + "angle": 0, + "content": "Table 1: Some operations in our attention search space. \"UOP\" and \"BOP\" denote the unary and binary operation, respectively. Full operations are in the Appendix." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.428, + 0.432, + 0.614 + ], + "angle": 0, + "content": "
OP IDOP NameExpression
UOP00no_opx
UOP01relumax(0,x)
UOP02scalex/√dx
UOP03sqrt√x
UOP04invert1/x+e-6
UOP05l2_normx-mean(x)/std(x)
UOP06sigmoid1/1+e-x
UOP07logsoftmaxln ex/∑i=1n e^xi
UOP08softmaxex/∑i=1n e^xi
UOP09softsignx/1+|x|
" + }, + { + "type": "table", + "bbox": [ + 0.477, + 0.432, + 0.724, + 0.611 + ], + "angle": 0, + "content": "
OP IDOP NameExpression
UOP10eluelu(x) + 1
UOP11sigmoid_revert1 - sigmoid(x)
UOP12expex
UOP13abslog|ln x|
UOP14min_max_normx-min(x)/max(x)-min(x)
UOP15transposexT
UOP16swishx × sigmoid (x)
BOP01sumx + y
BOP02cosine_simcos(x,y)
BOP03productx ⊙ y
BOP04matrix-multix · y
BOP05euclidean_disd(x,y)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.779 + ], + "angle": 0, + "content": "performance of these attention mechanisms can be enhanced without a significant increase in computational overhead. Based on the different binary nodes, these augmented attentions can be categorized into five types (i3n301 to i3n305), effectively enhancing the modeling capability. We divide the sub-search space according to these different types of attention. During the search process, we first search the solutions within the sub-search space in parallel and then combine them to obtain the final search result. Figure 4 (b) shows performance for different attention candidates. These designs allow our method to balance computational complexity and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Primitive Operations. Table 1 presents some unary and binary operators in our search space. These operators encompass both commonly used operations found in existing linear attentions [3, 24, 54, 66] and our proposed ones, including: (1) The activation operations are responsible for enhancing the non-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.307, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.616, + 0.161 + ], + "angle": 0, + "content": "Algorithm 1 NSGA-II Evolution Search for AttnZero." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.164, + 0.786, + 0.19 + ], + "angle": 0, + "content": "Input: Search space \\(\\mathcal{A}\\), population size \\(N\\), crossover probability \\(p_c\\), mutation probability \\(p_m\\), max generations \\(G\\)." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.191, + 0.578, + 0.204 + ], + "angle": 0, + "content": "Output: Set of Pareto-optimal attention formulations." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.207, + 0.672, + 0.221 + ], + "angle": 0, + "content": "1: Initialize population \\(\\mathcal{P}_0\\) with \\(N\\) individuals generated randomly;" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.222, + 0.656, + 0.235 + ], + "angle": 0, + "content": "2: Evaluate the fitness of each individual in \\(\\mathcal{P}_0\\) for all objectives;" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.236, + 0.369, + 0.248 + ], + "angle": 0, + "content": "3: for \\( g = 1 \\) to \\( G \\) do" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.249, + 0.746, + 0.262 + ], + "angle": 0, + "content": "4: Perform non-dominated sorting on \\(\\mathcal{P}_{g - 1}\\) to classify into fronts \\(F_{1},F_{2},\\ldots\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.263, + 0.64, + 0.276 + ], + "angle": 0, + "content": "5: Calculate crowding distance for individuals in each front;" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.277, + 0.761, + 0.29 + ], + "angle": 0, + "content": "6: \\(\\mathcal{Q} \\coloneqq \\operatorname{Select} N\\) individuals for mating via front rank and crowding distance;" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.291, + 0.764, + 0.304 + ], + "angle": 0, + "content": "7: Generate offspring \\(\\mathcal{Q}'\\) via crossover & mutation on \\(\\mathcal{Q}\\) with prob. \\(p_c\\) and \\(p_m\\);" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.305, + 0.663, + 0.318 + ], + "angle": 0, + "content": "8: Evaluate the fitness of each offspring in \\(\\mathcal{Q}'\\) for all objectives;" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.318, + 0.773, + 0.332 + ], + "angle": 0, + "content": "9: Apply rejection protocols to \\(\\mathcal{Q}'\\) to remove unpromising candidate and get \\(\\mathcal{Q}''\\)." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.333, + 0.459, + 0.346 + ], + "angle": 0, + "content": "10: \\(\\mathcal{R} := \\operatorname{Combine} \\mathcal{P}_{g-1}\\) and \\(\\mathcal{Q}'\\);" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.347, + 0.514, + 0.359 + ], + "angle": 0, + "content": "11: Perform non-dominated sorting on \\(\\mathcal{R}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.36, + 0.785, + 0.374 + ], + "angle": 0, + "content": "12: Update \\(\\mathcal{P}_g\\) by selecting top \\(N\\) individuals via front rank and crowding distance;" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.375, + 0.302, + 0.386 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.387, + 0.753, + 0.401 + ], + "angle": 0, + "content": "14: \\(\\mathcal{P}^* \\coloneqq\\) Extract the first front from \\(\\mathcal{P}_G\\) as the set of Pareto-optimal solutions;" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.207, + 0.785, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.8, + 0.618 + ], + "angle": 0, + "content": "linearity of attention. We include relu in FLatten [24] and introduce new activation functions such as swish and elu. (2) Normalization operations ensure a smoother distribution of attention scores. We provide options such as \\( l2\\_norm \\) and \\( min\\_max\\_norm \\). (3) The binary operators facilitate information interaction. Alongside cosine_similarity in Hydra [3], we also consider euclidean_distance as a simplified calculation option. As shown in Figure 4 (c), these operators enhance the versatility and adaptability of attention expressions, resulting in improved performance. Some complex operations such as decomposition [64] and Gaussian kernel [6] would result in a highly sparse search space or need hardware-specific supports. Hence they have not been included in our investigation. In future work, we will continue to involve more promising operators in our search space." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.645, + 0.562, + 0.66 + ], + "angle": 0, + "content": "3.3 Multi-objective Evolutionary Search" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.735 + ], + "angle": 0, + "content": "After building the search space \\(\\mathcal{A}\\), we sample the attention candidates \\(a \\in \\mathbb{N}\\) and apply it to the ViT model with weights \\(W\\). For single-object search, the aim is to find the optimal attention \\(a_{single}^{*}\\) by evaluating candidates using the trained ViT model \\((W^{*})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.746, + 0.786, + 0.769 + ], + "angle": 0, + "content": "\\[\na _ {s i n g l e} ^ {*} = \\underset {a \\in \\mathcal {A}} {\\operatorname {a r g m a x}} \\mathrm {A C C} _ {\\text {v a l}} \\left(\\mathcal {N} (a, W _ {\\mathcal {A}} ^ {*} (a))\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "However, \\( a_{\\text{single}}^* \\) may perform well on the search model but fail to other ViT models. For example, attention searched in Swin, while superior on Swin, performs poorly on DeiT, PVT in Figure 5 (b). Such attentions suffer from limited usefulness because they are difficult to transfer directly to different tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "Multi-objective Evolution. To improve generalization of attention candidates in different ViT models, our AttnZero framework employs a multi-objective evolutionary algorithm. This algorithm aims to maximize the results of multiple ViT models \\(1, 2, \\ldots, n\\) under predefined computational complexity constraints. We define a fitness function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.228, + 0.786, + 0.251 + ], + "angle": 0, + "content": "\\[\na _ {m u l t i} ^ {*} = \\underset {a \\in \\mathcal {A}} {\\operatorname {a r g m a x}} \\mathrm {A C C} _ {\\text {v a l}} \\left(1 (a, W _ {1 \\mathcal {A}} ^ {*} (a)), 2 (a, W _ {2 \\mathcal {A}} ^ {*} (a)), \\dots , n (a, W _ {n \\mathcal {A}} ^ {*} (a)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.789, + 0.544 + ], + "angle": 0, + "content": "Our framework incorporates validation results of attention in DeiT, AutoFormer, PVT, and Swin as objectives. This ensures that our attention module performs effectively in various ViT models, including pure ViT, ViT models with NAS search, ViT models with hierarchical design, and ViT models with local window configurations, respectively. Solving this multi-objective problem is challenging due to obstacles like conflicting optimals. To address this, we utilize NSGA-II method [10] in Alg. 1 as our search engine. Our search begins with an initial population of candidate attention modules and each candidate's fitness is evaluated based on performance across ViT models. Non-dominated sorting ranks the candidates and assigns them to fronts based on how they compare to others. Then, we use crowding distance [48] calculations to maintain diversity within the population. The selection process chooses candidates for the next generation, considering sorting and distance. Genetic operations like crossover and mutation are applied to the selected candidates. This introduces diversity and explores new areas of the search space. The least fit candidates are replaced with offspring, ensuring continuous gains. The algorithm iterates through these steps until it meets a termination criterion. During the search process, our algorithm takes FLOPs as the complexity constraint. Finally, we consider the real speed results in selecting final found attention from the Pareto-optimal solutions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.604, + 0.74 + ], + "angle": 0, + "content": "Search Acceleration Strategies. To improve the search efficiency, we employ several strategies for search acceleration: (1) Program checking. We perform program checking by verifying if the single candidate attention module can successfully generate outputs using random tensor inputs. This helps us identify and filter out candidates with arithmetic errors or dimension mismatches. (2) Rejection protocols. During training each population, we reject candidates with collapsed optimisation (e.g., loss = nan) and historical average validation results falling far below baseline linear attention. These strategies allow more than \\(5 \\times\\) search acceleration." + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.561, + 0.773, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.7, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Figure 3: Discovered attention \\( a_{\\text{AttnZero}}^* \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.759, + 0.51, + 0.775 + ], + "angle": 0, + "content": "3.4 Discovered AttnZero Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Formulas of discovered AttnZero. Figure 3 illustrates the optimal attention \\( a_{\\text{AttnZero}}^* \\) in our search, whose detailed expressions are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.818, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\na _ {A t t n Z e r o} ^ {*} = Q \\times \\eta_ {2} (\\eta_ {2} (Q) \\eta_ {2} (\\phi_ {1} (K) ^ {T} \\eta_ {1} (V))) \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.307, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.413, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.147, + 0.599, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.147, + 0.785, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.273, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Figure 4: (a) Performance distribution of subsets of Attn-Bench-101 (AutoFormer). (b) Results of various types of Attention. (c) Results of our default search space, default search space without advanced normalization operations, activation operations, and binary operations that we have newly introduced. More details are in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.526 + ], + "angle": 0, + "content": "where \\(\\phi_{1}\\) utilizes the Exponential Linear Unit (ELU) with a shift of 1, introducing nonlinearities into the attention computation. The \\(\\eta_{1}\\) performs min-max normalization, scaling the matrix values to a range between 0 and 1. This normalization step ensures that the attention weights are properly distributed and avoids potential issues with extreme values. Furthermore, the l2 normalization function \\(\\eta_{2}\\) guarantees that the attention weights have a unit norm, facilitating better interpretation and utilization of the attention mechanism. Regarding computational complexity, AttnZero requires triple matrix multiplications (i.e., \\(\\mathcal{O}(d\\times d + N\\times d + N\\times d)\\). Despite this, the overall time and memory complexity of AttnZero remains \\(\\mathcal{O}(N)\\), which ensures AttnZero to handle computations for sequences of varying lengths efficiently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Analysis of Attn-Bench-101. Our Attn-Bench-101 collects 2,000 attention configurations and their training results in DeiT/AutoFormer/PVT/Swin on CIFAR-100 (see Appendix for details). We analyze some results on AutoFormer to understand better our search space: (1) In Figure 4 (a), we observe that our search space contains many candidates with stronger performance than baseline model. (2) Among various computed graphs, we identify i3n302 and i3n303, superior to the other types (see in Figure 4 (b)). This is attributed to their augmented topology allowing for a good tradeoff between performance and efficiency. (3) For ablations in Figure 4 (c), advanced activation and normalization operations offer more contributions compared to new binary operators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Guidance for Efficient Attention Design. Based on the above analysis, we can summarize that we prefer attention types of i3n302 & i3n303 and combinations of advanced normalization and activation operations in the detailed design." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.749, + 0.446, + 0.767 + ], + "angle": 0, + "content": "4 Search Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.557, + 0.798 + ], + "angle": 0, + "content": "4.1 Search Implementation and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Implementation. Our attention search is performed on CIFAR-100, which are scaled to \\(224 \\times 224\\) resolution. We search using the validation results to ensure" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Top-1 accuracy (\\%) of hand-crafted attentions (e.g., Linear [28] and FLatten [24]) and attention with Random NAS method in our search space on CIFAR-100." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.185, + 0.777, + 0.273 + ], + "angle": 0, + "content": "
ModelsBaselineLinear [28]FLatten [24]Random (NAS)AttnZero (NAS)
DeiT-T65.0857.8072.4576.0477.68 (+12.60)
AutoFormer66.5864.5875.4276.9478.61 (+12.03)
PVT-T67.4266.5874.8373.1276.68 (+9.26)
Swin-T68.2569.4574.3174.5875.90 (+7.65)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.784, + 0.392 + ], + "angle": 0, + "content": "fair comparisons, and our validation set does not overlap with the test set. For search settings, we configure \\((N, p_c, p_m, G)\\) in Alg. 1 as (20,0.9,0.1,100) for parallel search on DeiT|AutoFormer|PVT|Swin. To obtain validation results, we adopt standard training settings [31] (e.g., 300 epochs and AdamW optimizer). With our search acceleration strategies, the overall search only costs 1.05 days on the \\(8 \\times \\mathrm{V}100\\). More details are available in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.785, + 0.558 + ], + "angle": 0, + "content": "Comparison Results with Hand-Crafted Attentions. The results in Table 2 demonstrate the effectiveness of our AttnZero in consistently achieving stable boosts across different models. Firstly, when compared to the baseline, our AttnZero achieves significant improvements. For example, the DeiT-T model with our AttnZero improves from a baseline accuracy of \\(65.08\\%\\) to \\(77.68\\%\\). Similar improvements are observed for the AutoFormer, PVT-T, and Swin-T models, indicating the consistent effectiveness of our approach. Furthermore, our AttnZero shows improvements over the recent state-of-the-art method. For example, the gain over FLatten for the DeiT-T model is \\(5.2\\%\\). These findings emphasize the potential and significance of our AttnZero approach in improving the performance of various models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.68 + ], + "angle": 0, + "content": "Comparing Random NAS Method. It is not easy to directly compare our method to other NAS methods because we use different search spaces and do not have weight sharing. To analyze our search strategy, we compare it to random search in our same search space. The results in Table 2 show that our AttnZero consistently improves model performance compared to random search. We achieve increases in accuracy ranging from \\(1.32\\%\\) to \\(3.56\\%\\) depending on the model. This demonstrates that our search strategy is more effective than a random search." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.706, + 0.39, + 0.722 + ], + "angle": 0, + "content": "4.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Multi-Objective Search (MOS) vs. Single-Objective Search (SOS). We employ MOS to enable the discovery of attentions that exhibit strong generalization capabilities. Figure 5 (a) illustrates the optimal candidate resulting from our MOS process. The results demonstrate that the Top-3 candidates outperform the others consistently across various ViT models. In Figures 5 (b) and (c), we compare the generalization performance of MOS and SOS across different ViT models and datasets, respectively. The results indicate that while the SOS" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.307, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.353, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.144, + 0.477, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.144, + 0.603, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.145, + 0.785, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.257, + 0.788, + 0.3 + ], + "angle": 0, + "content": "Figure 5: (a) Candidates in multi-objective search. (b) Comparisons of multi-objective vs. single-objective on multiple architectures; (c) Comparisons of multi-objective vs. single-objective on multiple datasets. (d) Search curves of various settings." + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.314, + 0.731, + 0.328 + ], + "angle": 0, + "content": "Table 3: Results of AttnZero for different ViT models and datasets." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.34, + 0.789, + 0.498 + ], + "angle": 0, + "content": "
ModelTop-1 Acc. (%)
Tiny ImageNet [1]Flowers [44]Chaoyang [74]
DeiT-T53.6250.0682.00
DeiT-T+AttnZero59.91 (+6.29)57.89 (+7.83)83.12 (+1.12)
AutoFormer-T56.3854.9882.84
AutoFormer-T+AttnZero62.56 (+6.18)61.58 (+6.60)83.68 (+0.84)
PVT-T58.4758.5782.46
PVT-T+AttnZero63.10 (+4.63)64.30 (+5.73)84.57 (+2.11)
Swin-T60.2258.8582.98
Swin-T+AttnZero64.55 (+4.33)65.13 (+6.28)85.11 (+2.13)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.603 + ], + "angle": 0, + "content": "method may yield better results on the searched models, MOS consistently outperforms SOS over a wider range of ViT architectures. Similarly, our AttnZero approach with MOS consistently demonstrates performance benefits across different dataset domains. For instance, our AttnZero with MOS achieves a \\(0.9\\%\\) improvement over SOS on the ImageNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Ablation of the Search Algorithm. In Figure 5 (d), we examine the evolutionary search and search acceleration strategies. The search results reveal the effectiveness of our acceleration strategy (i.e., program checking and rejection protocols) in filtering out unpromising candidates and greatly expediting the algorithm's convergence. These search curves also indicate that the evolutionary algorithm outperforms random search regarding efficiency in finding high-quality solutions within our framework." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.735, + 0.465, + 0.753 + ], + "angle": 0, + "content": "5 Transfer Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.768, + 0.588, + 0.784 + ], + "angle": 0, + "content": "5.1 Experimental Results on Tiny Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.804, + 0.843 + ], + "angle": 0, + "content": "Implementation. To check if our discovered attention generalizes beyond the CIFAR-100 dataset that we searched on, we it to multiple datasets (i.e., Tiny ImageNet [1], Flowers [44], and Chaoyang [74]). We train DeiT|AutoFormer|PVT|Swin" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.48, + 0.174 + ], + "angle": 0, + "content": "Table 4: Results of AttnZero for DeiT & PVT on ImageNet \\((224\\times 224)\\)" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.492, + 0.321 + ], + "angle": 0, + "content": "
MethodParamsFLOPsTop-1 (%)
DeiT-T [58]5.7M1.2G72.2
DeiT-T+AttnZero5.7M1.0G74.9 (+2.7)
PVT-T [60]13.2M1.9G75.1
PVT-T+AttnZero10.9M1.8G78.1 (+3.0)
PVT-S24.5M3.8G79.8
PVT-S+AttnZero19.8M3.6G81.4 (+1.6)
PVT-M44.2M6.7G81.2
PVT-M+AttnZero34.6M6.6G83.0 (+1.8)
PVT-L61.4M9.8G81.7
PVT-L+AttnZero47.1M9.7G83.3 (+1.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.145, + 0.795, + 0.174 + ], + "angle": 0, + "content": "Table 5: Results of AttnZero on ImageNet with \\(224\\times 224\\) resolution inputs." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.185, + 0.784, + 0.321 + ], + "angle": 0, + "content": "
MethodParamsFLOPsTop-1 (%)
AutoFormer-T [58]5.7M1.3G74.7
AutoF.-T+AttnZero5.7M1.1G75.2 (+0.5)
Swin-T [41]29.0M4.5G81.3
Swin-T+AttnZero28.2M4.3G82.1 (+0.8)
Swin-S50.0M8.7G83.0
Swin-S+AttnZero49.5M8.5G83.2 (+0.2)
Swin-B88.0M15.4G83.5
Swin-B+AttnZero89.0M15.3G83.7 (+0.2)
CSwin-T [17]23.0M4.3G82.7
CSwin-T+AttnZero20.3M4.0G82.9 (+0.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.25, + 0.334, + 0.751, + 0.349 + ], + "angle": 0, + "content": "Table 6: Comparison of different efficient attention methods on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.279, + 0.361, + 0.721, + 0.475 + ], + "angle": 0, + "content": "
MethodResoParamsFLOPsTop-1 (%)
DeiT-T [58]22425.7M1.2G72.2
DeiT-T+Hydra Attn [3]22425.7M1.1G68.3
DeiT-T+Efficient Attn [54]22425.7M1.1G70.2
DeiT-T+Linear Angular [66]22425.7M1.1G70.8
DeiT-T+FLatten22426.1M1.1G74.1
DeiT-T+AttnZero22425.7M1.0G74.9 (+2.7)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.786, + 0.548 + ], + "angle": 0, + "content": "using discovered attention from scratch. The training process follows standard settings [31], including 300 training epochs, a cosine learning rate scheduler, and the AdamW optimizer. Detailed settings are shown in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.745 + ], + "angle": 0, + "content": "Results on Various Datasets. Table 3 highlights the effectiveness of AttnZero across various ViT models and datasets. It consistently surpasses the baselines, resulting in a \\(6.29\\%\\) increase in accuracy for the DeiT-T model on Tiny-ImageNet. Similarly, AutoFormer and PVT-T models with AttnZero exhibit similar improvements. On the Flowers dataset, AttnZero achieves an accuracy of \\(57.89\\%\\) for the DeiT-T model, surpassing the baseline by \\(7.83\\%\\). AutoFormer and PVT-T models also benefit from AttnZero. The results on the Chaoyang dataset further exemplify AttnZero's capability to enhance accuracies. On the DeiT-T model, AttnZero achieves an accuracy of \\(83.12\\%\\), compared to the \\(82\\%\\) baseline. Moreover, the attention modules from AttnZero contribute to the improved performance of AutoFormer, PVT-T, and Swin-T models compared to their original versions. These findings demonstrate the promising ability of AttnZero across different datasets and models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.641, + 0.785 + ], + "angle": 0, + "content": "5.2 Experimental Results on Large-scale Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Implementation. We conduct the experiment on the ImageNet [52] dataset, a widely used large-scale dataset consisting of 1000 classes and 1.2 million training images. Note that we do not perform additional searches or modify the exist-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.307, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 7: Results for Mask R-CNN on COCO dataset. We compute the FLOPs across the backbone, FPN, and detection head using an input resolution of \\(1280 \\times 800\\)." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.186, + 0.731, + 0.279 + ], + "angle": 0, + "content": "
MethodFLOPsSch.\\( AP^b \\)\\( AP_{50}^b \\)\\( AP_{75}^b \\)\\( AP^m \\)\\( AP_{50}^m \\)\\( AP_{75}^m \\)
PVT-T240G1x36.759.239.335.156.737.3
PVT-T+FLatten244G1x38.261.641.937.057.639.0
PVT-T+AttnZero209G1x40.062.643.137.459.440.1
PVT-S305G1x40.462.943.837.860.140.3
PVT-S+AttnZero248G1x41.063.445.038.260.441.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.784, + 0.388 + ], + "angle": 0, + "content": "ing classical models. Instead, we directly replace the self-attention mechanism with our discovered attention module. Furthermore, we ensure that our training settings remain consistent with the standard settings [17, 41, 58]. This includes training the models for 300 epochs, cosine learning rate scheduler, and AdamW optimizer. Specific training settings are shown in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.785, + 0.584 + ], + "angle": 0, + "content": "Results on Various Backbones. Table 4 and Table 5 provides the performance results of our AttnZero on ImageNet. The results show that our method can transfer effectively to different ViT models. When applied to DeiT-T, AttnZero achieves a significant \\(2.7\\%\\) gain over baseline. Similarly, AttnZero improves PVT-T accuracy to \\(78.1\\%\\), surpassing the baseline by \\(3.0\\%\\). This demonstrates that AttnZero can seamlessly migrate and consistently enhance performance across ViT architectures. AttnZero also provides stable gains across models of varying scales. Paired with PVT-S, it boosts accuracy by \\(1.6\\%\\). Applied to larger models like PVT-M, PVT-L, Swin-T, Swin-S, and CSwin-T, improvements of \\(1.8\\%\\), \\(1.6\\%\\), \\(0.8\\%\\), \\(0.2\\%\\) and \\(0.2\\%\\) respectively are observed. These findings highlight AttnZero's stability in boosting the performance of ViT models from small to large scale. These results highlight the stability of our AttnZero in enhancing the performance of ViT models across different scales." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.787, + 0.723 + ], + "angle": 0, + "content": "Comparison with Other Linear Attention. Table 6 compares various linear attention methods. The results demonstrate that AttnZero surpasses other linear attention methods and baseline when incorporated into the DeiT-T model. Notably, compared to the Softmax baseline, AttnZero achieves a \\(2.7\\%\\) improvement. In contrast, other linear attention methods, such as Hydra Attn, Efficient Attn, Linear Angular, and FLatten, exhibit lower accuracies. These findings highlight the superiority of AttnZero in preserving or surpassing the performance achieved by Softmax attention. On the other hand, other linear attention methods tend to result in notable performance degradation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.75, + 0.707, + 0.765 + ], + "angle": 0, + "content": "5.3 Transfer Experiments on Detection and Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Object Detection. We evaluate our methods for object detection and instance segmentation tasks on the COCO dataset [39]. Table 7 provides results on Mask R-CNN detector [25] with \\(1 \\times\\) schedules. Our proposed PVT-T+AttnZero model achieves 2.3 box AP gains on PVT-T and significantly outperforms FLatten." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 8: Results of semantic segmentation on ADE20K. We compute the FLOPs for both the encoders and decoders using input images at a resolution of \\(512 \\times 2048\\)." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.186, + 0.726, + 0.279 + ], + "angle": 0, + "content": "
BackboneMethodFLOPs#ParamsmIoUmAcc
PVT-TSemanticFPN158G17M36.5746.72
PVT-T+FLattenSemanticFPN169G16M37.2148.95
PVT-T+AttnZeroSemanticFPN126G15M39.1650.30
PVT-SSemanticFPN225G28M41.9553.02
PVT-S+AttnZeroSemanticFPN165G23M42.6254.02
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.307, + 0.784, + 0.336 + ], + "angle": 0, + "content": "These enhancements demonstrate the benefits of incorporating our found attention module within the popular ViT model for object detection tasks." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.337, + 0.785, + 0.457 + ], + "angle": 0, + "content": "Semantic Segmentation. We assess the performance of our model using the ADE20K dataset [71]. Table 8 provides semantic segmentation results on SemanticFPN [29]. Among them, the Flatten module yields slight gains on the baseline PVT-T model. However, our AttnZero model achieves even higher mIoU (39.16) and mAcc (50.30) scores while reducing the FLOPs and parameters. Similarly, AttnZero achieves clear gains on PVT-S. These results highlight the compatibility of our AttnZero with different backbone architectures and its consistent improvements across various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.477, + 0.357, + 0.494 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.506, + 0.786, + 0.747 + ], + "angle": 0, + "content": "In this paper, we introduce AttnZero, the first method that automates the discovery of efficient attention modules that can be applied to various ViT models. AttnZero constructs a comprehensive search space consisting of different unary and binary operators as potential attention mechanisms. It utilizes a multi-objective genetic algorithm to search within this design space efficiently. Additionally, we introduce Attn-Bench-101, which provides attention profile results to summarize the insights gained from the search. Extensive experiments demonstrate that the attention modules discovered by AttnZero can be successfully applied to advanced ViTs for image classification, segmentation, and detection tasks. For future work, we will consider extending the AttnZero with knowledge distillation methods [32-35, 38, 53, 63]. Limitation. We present search strategies but still involve some search costs following most AutoML methods. But these costs are worthwhile because our discovered attention can be effectively applied to different ViTs and tasks without additional searching (proven by our extensive transfer experiments). Social impact. Our AttnZero focuses on technology improvement without social and ethical implications." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.767, + 0.402, + 0.785 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.784, + 0.839 + ], + "angle": 0, + "content": "The research was supported by Theme-based Research Scheme (T45-205/21-N) from Hong Kong RGC, and Generative AI Research and Development Centre from InnoHK." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.307, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.518, + 0.187 + ], + "angle": 0, + "content": "1. http://tiny-imagenet.herokuapp.com/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.187, + 0.787, + 0.228 + ], + "angle": 0, + "content": "2. Ali, A., Touvron, H., Caron, M., Bojanowski, P., Douze, M., Joulin, A., Laptev, I., Neverova, N., Synnaeve, G., Verbeek, J., et al.: Xcit: Cross-covariance image transformers. NIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.228, + 0.786, + 0.255 + ], + "angle": 0, + "content": "3. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Hoffman, J.: Hydra attention: Efficient attention with many heads. In: ECCVW (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.255, + 0.786, + 0.296 + ], + "angle": 0, + "content": "4. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. arXiv preprint arXiv:2005.14165 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.786, + 0.323 + ], + "angle": 0, + "content": "5. Chen, M., Peng, H., Fu, J., Ling, H.: Autoformer: Searching transformers for visual recognition. In: ICCV. pp. 12270-12280 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.786, + 0.364 + ], + "angle": 0, + "content": "6. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.786, + 0.405 + ], + "angle": 0, + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. JMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.446 + ], + "angle": 0, + "content": "8. Chu, X., Tian, Z., Zhang, B., Wang, X., Wei, X., Xia, H., Shen, C.: Conditional positional encodings for vision transformers. Arxiv preprint 2102.10882 (2021), https://arxiv.org/pdf/2102.10882.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.786, + 0.472 + ], + "angle": 0, + "content": "9. Chu, X., Zhang, B., Xu, R.: Fairnas: Rethinking evaluation fairness of weight sharing neural architecture search. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.472, + 0.786, + 0.5 + ], + "angle": 0, + "content": "0. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary computation (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.786, + 0.54 + ], + "angle": 0, + "content": "1. Dong, P., Li, L., Tang, Z., Liu, X., Pan, X., Wang, Q., Chu, X.: Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. In: ICML (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.54, + 0.786, + 0.568 + ], + "angle": 0, + "content": "2. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.786, + 0.609 + ], + "angle": 0, + "content": "3. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV. pp. 17076-17086 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.786, + 0.65 + ], + "angle": 0, + "content": "4. Dong, P., Niu, X., Li, L., Tian, Z., Wang, X., Wei, Z., Pan, H., Li, D.: Rd-nas: Enhancing one-shot supernet ranking ability via ranking distillation from zero-cost proxies. ICASSP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.786, + 0.676 + ], + "angle": 0, + "content": "5. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.786, + 0.717 + ], + "angle": 0, + "content": "6. Dong, P., Niu, X., Tian, Z., Li, L., Wang, X., Wei, Z., Pan, H., Li, D.: Progressive meta-pooling learning for lightweight image classification model. In: ICASSP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.786, + 0.758 + ], + "angle": 0, + "content": "7. Dong, X., Bao, J., Chen, D., Zhang, W., Yu, N., Yuan, L., Chen, D., Guo, B.: Cswin transformer: A general vision transformer backbone with cross-shaped windows. ArXiv abs/2107.00652 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.785 + ], + "angle": 0, + "content": "8. Dong, X., Yang, Y.: Nas-bench-201: Extending the scope of reproducible neural architecture search (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.786, + 0.84 + ], + "angle": 0, + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.174, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Gao, J., Xu, H., Shi, H., Ren, X., Philip, L., Liang, X., Jiang, X., Li, Z.: Autobert-zero: Evolving bert backbone from scratch. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.243 + ], + "angle": 0, + "content": "21. Guan, C., Wang, X., Zhu, W.: Autoattend: Automated attention representation search. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 3864-3874. PMLR (18-24 Jul 2021), http://proceedings.mlr.press/v139/guan21a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.785, + 0.285 + ], + "angle": 0, + "content": "22. Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. arXiv preprint arXiv:1904.00420 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.326 + ], + "angle": 0, + "content": "23. Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. In: ECCV. vol. abs/1904.00420, pp. 544-560. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.326, + 0.785, + 0.353 + ], + "angle": 0, + "content": "24. Han, D., Pan, X., Han, Y., Song, S., Huang, G.: Flatten transformer: Vision transformer using focused linear attention. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.743, + 0.366 + ], + "angle": 0, + "content": "25. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.393 + ], + "angle": 0, + "content": "26. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.42 + ], + "angle": 0, + "content": "27. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.447 + ], + "angle": 0, + "content": "28. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "29. Kirillov, A., Girshick, R., He, K., Dollár, P.: Panoptic feature pyramid networks. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.501 + ], + "angle": 0, + "content": "30. Kitaev, N., Kaiser, L., Levskaya, A.: Reformer: The efficient transformer. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "31. Li, K., Yu, R., Wang, Z., Yuan, L., Song, G., Chen, J.: Locality guidance for improving vision transformers on tiny datasets. In: ECCV. pp. 110-127. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.569 + ], + "angle": 0, + "content": "32. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.596 + ], + "angle": 0, + "content": "33. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.623 + ], + "angle": 0, + "content": "34. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.65 + ], + "angle": 0, + "content": "35. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.677 + ], + "angle": 0, + "content": "36. Li, L., Sun, H., Dong, P., Wei, Z., Shao, S.: Auto-das: Automated proxy discovery for training-free distillation-aware architecture search. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "37. Li, L., Sun, H., Li, S., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Auto-gas: Automated proxy discovery for training-free generative architecture search. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.746 + ], + "angle": 0, + "content": "38. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "40. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.307, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "42. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.201 + ], + "angle": 0, + "content": "43. Nakai, K., Matsubara, T., Uehara, K.: Att-darts: Differentiable neural architecture search for attention. In: IJCNN. IEEE" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "44. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.785, + 0.27 + ], + "angle": 0, + "content": "45. Peng, H., Du, H., Yu, H., Li, Q., Liao, J., Fu, J.: Cream of the crop: Distilling prioritized paths for one-shot neural architecture search. NIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "46. Pham, H., Guan, M.Y., Zoph, B., Le, Q.V., Dean, J.: Efficient neural architecture search via parameter sharing. In: ICML (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.785, + 0.324 + ], + "angle": 0, + "content": "47. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.326, + 0.785, + 0.351 + ], + "angle": 0, + "content": "48. Raquel, C.R., Naval Jr, P.C.: An effective use of crowding distance in multiobjective particle swarm optimization. In: CGEC (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.785, + 0.378 + ], + "angle": 0, + "content": "49. Real, E., Aggarwal, A., Huang, Y., Le, Q.V.: Regularized evolution for image classifier architecture search. In: AAAI (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.785, + 0.404 + ], + "angle": 0, + "content": "50. Real, E., Liang, C., So, D., Le, Q.: Automl-zero: Evolving machine learning algorithms from scratch. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.406, + 0.785, + 0.432 + ], + "angle": 0, + "content": "51. Real, E., Liang, C., So, D., Le, Q.: Automl-zero: Evolving machine learning algorithms from scratch. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "52. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Li, F.F.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.475, + 0.785, + 0.513 + ], + "angle": 0, + "content": "53. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.515, + 0.785, + 0.54 + ], + "angle": 0, + "content": "54. Shen, Z., Zhang, M., Zhao, H., Yi, S., Li, H.: Efficient attention: Attention with linear complexities. In: WACV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.542, + 0.785, + 0.581 + ], + "angle": 0, + "content": "55. So, D.R., Manke, W., Liu, H., Dai, Z., Shazeer, N., Le, Q.V.: Primer: Searching for efficient transformers for language modeling. arXiv preprint arXiv:2109.08668 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "56. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.61, + 0.785, + 0.636 + ], + "angle": 0, + "content": "57. Tay, Y., Dehghani, M., Bahri, D., Metzler, D.: Efficient transformers: A survey. arXiv preprint arXiv:2009.06732 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.637, + 0.785, + 0.663 + ], + "angle": 0, + "content": "58. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.664, + 0.785, + 0.716 + ], + "angle": 0, + "content": "59. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. corr, abs/2302.13971, 2023. doi: 10.48550. arXiv preprint arXiv.2302.13971" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "60. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "61. Wang, X., Xiong, X., Neumann, M., Piergiovanni, A., Ryoo, M.S., Angelova, A., Kitani, K.M., Hua, W.: Attentionnas: Spatiotemporal attention cell search for video classification. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "62. Wei, Z., Pan, H., Li, L.L., Lu, M., Niu, X., Dong, P., Li, D.: Convformer: Closing the gap between cnn and vision transformers. arXiv preprint arXiv:2209.07738 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "63. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "64. Xiong, Y., Zeng, Z., Chakraborty, R., Tan, M., Fung, G., Li, Y., Singh, V.: Nystromformer: A nystrom-based algorithm for approximating self-attention. In: AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "65. Yang, A., Esperança, P.M., Carlucci, F.M.: Nas evaluation is frustratingly hard. arXiv preprint arXiv:1912.12522 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "66. You, H., Xiong, Y., Dai, X., Wu, B., Zhang, P., Fan, H., Vajda, P., Lin, Y.C.: Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "67. You, S., Huang, T., Yang, M., Wang, F., Qian, C., Zhang, C.: Greedynas: Towards fast one-shot nas with greedy supernet. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "68. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "69. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token vit: Training vision transformers from scratch onImagenet. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "70. Zheng, Z., Gao, X., Pan, J., Luo, Q., Chen, G., Liu, D., Jiang, J.: Autoattention: Automatic field pair selection for attention in user behavior modeling. In: 2022 IEEE International Conference on Data Mining (ICDM). pp. 803-812. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "71. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "72. Zhou, Q., Sheng, K., Zheng, X., Li, K., Sun, X., Tian, Y., Chen, J., Ji, R.: Training-free transformer architecture search. In: CVPR. pp. 10894-10903 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.507 + ], + "angle": 0, + "content": "73. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.535 + ], + "angle": 0, + "content": "74. Zhu, C., Chen, W., Peng, T., Wang, Y., Jin, M.: Hard sample aware noise robust learning for histopathology image classification. TMI" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "75. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "76. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "77. Zoph, B., Le, Q.V.: Neural architecture search with reinforcement learning. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "78. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.674 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_origin.pdf b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c88e6c3071b451beed41a113e48f30ff420a365a --- /dev/null +++ b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/147bb716-8a72-46c3-9d8e-9429b38a70d7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d8a0416a33ebe2efd67753965a7f83bf938d96cb89bcd600410f159055d3f86 +size 4305970 diff --git a/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/full.md b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae7b596df440b8e166d8add612e05717a8ae7d61 --- /dev/null +++ b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/full.md @@ -0,0 +1,330 @@ +# AttnZero: Efficient Attention Discovery for Vision Transformers + +Lujun Li $^{1}$ , Zimian Wei $^{2\boxtimes}$ , Peijie Dong $^{3}$ , Wenhan Luo $^{1}$ , Wei Xue $^{1}$ , Qifeng Liu $^{1\boxtimes}$ , and Yike Guo $^{1\boxtimes}$ + +1 The Hong Kong University of Science and Technology + +$^{2}$ National University of Defense Technology + +3 The Hong Kong University of Science and Technology (Guangzhou) + +lilujunai@gmail.com;weizimian16@nudt.edu.cn;pdong212@connect.hkust-gz.edu.cn;whluo.china@gmail.com;weixue@ust.hk;liuqifeng@ust.hk;yikeguo@ust.hk + +Abstract In this paper, we present AttnZero, the first framework for automatically discovering efficient attention modules tailored for Vision Transformers (ViTs). While traditional self-attention in ViTs suffers from quadratic computation complexity, linear attention offers a more efficient alternative with linear complexity approximation. However, existing hand-crafted linear attention suffers from performance degradation. To address these issues, our AttnZero constructs search spaces and employs evolutionary algorithms to discover potential linear attention formulations. Specifically, our search space consists of six kinds of computation graphs and advanced activation, normalize, and binary operators. To enhance generality, we derive results of candidate attention applied to multiple advanced ViTs as the multi-objective for the evolutionary search. To expedite the search process, we utilize program checking and rejection protocols to filter out unpromising candidates swiftly. Additionally, we develop Attn-Bench-101, which provides precomputed performance of 2,000 attentions in the search spaces, enabling us to summarize attention design insights. Experimental results demonstrate that the discovered AttnZero module generalizes well to different tasks and consistently achieves improved performance across various ViTs. For instance, the tiny model of DeiT|PVT|Swin|CSwin trained with AttnZero on ImageNet reaches $74.9\%$ |78.1%|82.1%|82.9% top-1 accuracy. Codes at: https://github.com/lliai/AttnZero. + +Keywords: Vision Transformer $\cdot$ Efficient Attention $\cdot$ AutoML + +# 1 Introduction + +In recent years, Vision Transformers (ViTs) demonstrate remarkable success in various vision tasks, including image classification [16, 19, 58, 62], object detection [60, 75], semantic segmentation [47]. ViTs draw inspiration from the transformative impact of Transformers in natural language processing, utilizing self-attention mechanisms to capture global dependencies. However, self-attention + +has a quadratic complexity that scales with the square of the input sequence length. This quadratic complexity limits the scalability of ViTs for handling large image inputs, thereby limiting their practical applications. + +To address this issue, researchers have been exploring the use of linear attention [57] as a more computationally efficient alternative. Unlike self-attention, linear attention has a linear complexity approximation, making it more suitable for handling large input sequences. However, existing linear attention methods [42, 54] suffer from several limitations: (1) Linear attention often leads to reduced performance than softmax attention due to its simplifying assumptions. The simple linear operators used in linear attention have limited flexibility and produce noisier activations than nonlinear functions [24]. These performance degradations challenge leveraging linear attention as a viable substitute for self-attention in ViTs. (2) Hand-crafted linear attention designs are often specific to certain tasks or architectures, lacking generality across different scenarios. Additionally, the manual design is labor-intensive, time-consuming, and heavily reliant on expert knowledge. Consequently, these drawbacks hinder the scalability and accessibility of linear attention in ViTs. + +To address these challenges, we have made several observations. We notice that some large language models [4, 7] employ new operators in Transformers to improve performance. For example, LLaMA [59] currently uses RMSNorm and SiLU to prevent training crashes. These successes inspire us to assemble a series of advanced operators to enhance linear attention and explore their optimal combination formulations. However, the question remains: How can we efficiently solve this combination problem? Recent developments in AutoML, such as AutoML-Zero [50] and NAS methods [40, 46], offer a promising solution. These methods involve constructing an extensive search space encompassing various options and searching for optimal architectures. Nevertheless, existing NAS methods are not suitable for solving the formulation discovery problem in our case for several reasons. (1) Different search space. The search spaces in conventional NAS methods typically focus on network width/depth, block/cell types, and parametric operators [5, 56]. In contrast, our attention search space includes many non-parametric operators and different combinatorial types. This leads to an explosion in the search space and many unsuccessful candidates. (2) Different search algorithms. Most recent NAS methods employ weight-sharing strategies [14, 27] to reduce search costs. However, our search space is explosive and extremely sparse, meaning that using previous ways to find potential candidates is less efficient and effective. (3) Different search goals. Most NAS methods prioritize performance for a single task and may struggle to generalize well to different applications [65]. In contrast, our objective is to discover generalized attention mechanisms that can be applied to various ViT models and tasks. These challenges highlight that attention discovery is a new task that requires a unique search space and strategy compared to traditional NAS methods. + +Based on the above observations, we propose AttnZero, a new framework that aims to discover superior linear attention formulations that achieve a better + +![](images/63c60c60f0fc55f05083d1da86342b05f18793ce17e402d4d2b05d7c6a217c71.jpg) +Figure 1: Overall pipeline of AttnZero. Our framework obtains candidate attention populations by sampling from attention search space. With program checking and rejection protocols, we conduct validation results of these attentions on multiple ViTs. Then, we use NSGA-II to obtain the optimal solution and cross-mutate for the next population. + +accuracy-efficiency tradeoff. As shown in Figure 1, our framework can be divided into two parts: a comprehensive attention search space and a multi-objective evolutionary search. Based on the basic form of linear attention, we develop six computation graphs to represent attention candidates in the search space. Then, we combine unary and binary operators with advanced functions, including various activation functions, normalization techniques, and matrix arithmetic functions. This approach allows us to obtain flexible and effective attention candidates. Our goal for the search algorithm is to find attention formulations applicable across different ViT models. To achieve this, we model the search task as a multi-objective optimization problem and employ the NSGA-II evolution method to obtain optimal solutions. During the search process, we validate candidate attention in multiple ViT models through multi-architecture evaluations. Based on these evaluations, our search engine exploits crossover and mutation for better populations. To accelerate the search process, we perform a fast program check and a rejection protocol to eliminate unpromising candidates during the search process. In this way, our AttnZero improves at least $5 \times$ in search efficiency without the weight-sharing setting [22], making our method free of ranking issues in traditional NAS methods [9, 18]. + +We conduct comprehensive experiments to evaluate the transferability of our discovered attention mechanisms across different ViT models, datasets, and tasks without requiring additional searches. When applied to various ViT models, our AttnZero consistently outperforms the baseline method. For instance, our method achieves gains of $3.0\% \sim 1.6\%$ on PVT-T to PVT-L on the ImageNet dataset. Furthermore, ViT models with AttnZero demonstrate significant improvements when considering different data domains. For example, DeiT+AttnZero achieves gains of $6.2\%$ , $7.8\%$ , $1.1\%$ , and $2.7\%$ on the Tiny-ImageNet, Flowers, Chaoyang, and ImageNet datasets, respectively. Additionally, we compile a new benchmark named Attn-Bench-101, which consists of 2,000 attentions from our + +search space. The benchmark highlights the preference for advanced normalization techniques, activation functions, and well-performing attentions. In conclusion, our contributions can be summarized as follows: + +- To improve linear attention, we present a new task, efficient attention discovery, that generalizes to various ViT models and tasks. For this task, our AttnZero, the first framework, is developed to achieve effective attention search. +- We propose a comprehensive attention search space, including different types of computation graphs and advanced operators as options. We develop multi-objective searches with multiple ViTs evaluations to ensure generalisability. +- Our found attention can be well generalized to DeiT|PVT|Swin|CSwin and other ViTs. Extensive experiments on multiple datasets in classification, detection, and segmentation prove the superior performance of our AttnZero. +- We build Attn-Bench-101, providing many attentions and their results. Some attention design guidances are summarised based on these results. + +# 2 Related Works + +Efficient Attention. The application of self-attention in computer vision has made significant progress [2,8,26,68,69] with the introduction of ViT [19]. Then, PVT [60] introduces a hierarchical structure while Swin Transformer [41] proposes window-based attention to model dependencies efficiently. However, softmax attention's quadratic complexity poses challenges. Linear attention reduces this but often degrades performance [6,30,64]. Efficient Attention [54] applies softmax to Q and K while SOFT [42] uses matrix decomposition. Castling-ViT [66] employs softmax attention during training but linear attention at inference. FLatten [24] preserves diversity through focused functions. While effective, these methods still have performance limitations and require complex manual design. In contrast, our AttnZero automatically discovers novel formulations to enhance linear attention. Rather than replacing prior work, we aim to improve the performance of the existing method while reducing manual effort. We build our search space based on previous work and provide feedback on new patterns. Thus, AttnZero complements rather than competes with manual methods. + +Neural Architecture Search (NAS). NAS methods automate model design through search spaces and algorithms [49,77]. Early NAS works adopt reinforcement learning with expensive costs [78]. Recent one-shot NAS [15] and zero-shot NAS [11-13,36,37,73,76] use techniques like weight sharing [23] and training-free proxy evaluation to speed up search. Current NAS methods on ViTs [5,45,72] (e.g., AutoFormer [5]) search the number of depth/head of transformers. In contrast, AttnZero focuses on discovering attention mechanisms, not architectural dimensions. We also apply AttnZero to AutoFormer and observe consistent improvements, indicating our method is orthogonal to existing ViT-NAS methods. There are attempts to search entire attention blocks [67], SE-like attention [43,61], projection layers [21], and Transformers searches in other fields like + +CTR [70] and NLP [20,55]. We clarify that our AttnZero is the first to search for efficient attention in visual tasks, which clearly differs in spaces/strategies/goals from previous methods. Inspired by general idea in AutoML-Zero [51], we achieve many new designs in search space/strategies for new attention discovery task, and opens up new possibilities for enhancing the capabilities of ViTs. + +# 3 Methodology: Efficient Attention Discovery + +In this section, we first review the linear attention form. Then, we present our search space and algorithms. The overall process is shown in Figure 1. + +# 3.1 Recap of Self-Attention and Linear Attention + +In self-attention, input tokens are projected into query $(Q)$ , key $(K)$ , and value $(V)$ matrices, respectively. The general form of self-attention is given by: + +$$ +O _ {i} = \sum_ {j = 1} ^ {N} \frac {\operatorname {S i m} \left(Q _ {i} , K _ {j}\right)}{\sum_ {j = 1} ^ {N} \operatorname {S i m} \left(Q _ {i} , K _ {j}\right)} V _ {j}, \tag {1} +$$ + +where $O_{i}$ represents the output of the $i$ -th token and Sim denotes the similarity function. The commonly used similarity function is the softmax function, $\exp\left(QK^{T}/\sqrt{d}\right)$ , where $d$ is the dimension of the input tokens. However, the quadratic time and space complexity of softmax attention, $\mathcal{O}(N^{2}d)$ , becomes impractical for large numbers of tokens ( $N$ ). To address this, linear attention is explored as an efficient alternative, which can be expressed as + +$$ +O _ {i} = \frac {\phi \left(Q _ {i}\right) \left(\sum_ {j = 1} ^ {N} \phi \left(K _ {j}\right) ^ {T} V _ {j}\right)}{\phi \left(Q _ {i}\right) \left(\sum_ {j = 1} ^ {N} \phi \left(K _ {j}\right) ^ {T}\right)}, \tag {2} +$$ + +where $\phi$ represents the kernel function. Linear attention reduces the computational complexity to $\mathcal{O}(N)$ , making it scalable for large input sequences. This reduction is achieved by using carefully designed kernels to approximate the similarity function, enabling a change in the computation order from $(QK^T)V$ to $Q(K^TV)$ , exploiting the associative property of matrix multiplication. + +# 3.2 Efficient Attention Search Space + +Attention Representation. In Figure 2, we construct six types of computational graphs using (Q, K, V) as inputs to represent our candidate attentions. Among them, our i3n2 attentions follow the same process as standard linear attention, and we improve it by searching for different unary operations. Partially inspired by Castling-ViT [66], we introduce novel attention mechanisms with three binary operator nodes based on basic linear attention. With additional operators with either $\mathcal{O}(N\times d)$ or $\mathcal{O}(d\times d)$ computational complexity, the + +![](images/51fd63600db8eba61318eb39388f26ad2caf22a7c44ed344c4bb43d7e4a3bb06.jpg) +Figure 2: Schematic of the six computational graphs in our search space. We denote the graph with 3 inputs and 2 binary operations as i3n2. Similarly, i3n3 refers to a graph with 3 inputs and 3 binary operations. Since each binary operation takes two inputs, there are 5 possible graph structures for i3n3. We label these i3n301~i3n305, respectively. For simplicity, we do not display unary operators here. + +Table 1: Some operations in our attention search space. "UOP" and "BOP" denote the unary and binary operation, respectively. Full operations are in the Appendix. + +
OP IDOP NameExpression
UOP00no_opx
UOP01relumax(0,x)
UOP02scalex/√dx
UOP03sqrt√x
UOP04invert1/x+e-6
UOP05l2_normx-mean(x)/std(x)
UOP06sigmoid1/1+e-x
UOP07logsoftmaxln ex/∑i=1n e^xi
UOP08softmaxex/∑i=1n e^xi
UOP09softsignx/1+|x|
+ +
OP IDOP NameExpression
UOP10eluelu(x) + 1
UOP11sigmoid_revert1 - sigmoid(x)
UOP12expex
UOP13abslog|ln x|
UOP14min_max_normx-min(x)/max(x)-min(x)
UOP15transposexT
UOP16swishx × sigmoid (x)
BOP01sumx + y
BOP02cosine_simcos(x,y)
BOP03productx ⊙ y
BOP04matrix-multix · y
BOP05euclidean_disd(x,y)
+ +performance of these attention mechanisms can be enhanced without a significant increase in computational overhead. Based on the different binary nodes, these augmented attentions can be categorized into five types (i3n301 to i3n305), effectively enhancing the modeling capability. We divide the sub-search space according to these different types of attention. During the search process, we first search the solutions within the sub-search space in parallel and then combine them to obtain the final search result. Figure 4 (b) shows performance for different attention candidates. These designs allow our method to balance computational complexity and accuracy. + +Primitive Operations. Table 1 presents some unary and binary operators in our search space. These operators encompass both commonly used operations found in existing linear attentions [3, 24, 54, 66] and our proposed ones, including: (1) The activation operations are responsible for enhancing the non- + +# Algorithm 1 NSGA-II Evolution Search for AttnZero. + +Input: Search space $\mathcal{A}$ , population size $N$ , crossover probability $p_c$ , mutation probability $p_m$ , max generations $G$ . + +Output: Set of Pareto-optimal attention formulations. + +1: Initialize population $\mathcal{P}_0$ with $N$ individuals generated randomly; +2: Evaluate the fitness of each individual in $\mathcal{P}_0$ for all objectives; +3: for $g = 1$ to $G$ do +4: Perform non-dominated sorting on $\mathcal{P}_{g - 1}$ to classify into fronts $F_{1},F_{2},\ldots$ +5: Calculate crowding distance for individuals in each front; +6: $\mathcal{Q} \coloneqq \operatorname{Select} N$ individuals for mating via front rank and crowding distance; +7: Generate offspring $\mathcal{Q}'$ via crossover & mutation on $\mathcal{Q}$ with prob. $p_c$ and $p_m$ ; +8: Evaluate the fitness of each offspring in $\mathcal{Q}'$ for all objectives; +9: Apply rejection protocols to $\mathcal{Q}'$ to remove unpromising candidate and get $\mathcal{Q}''$ . +10: $\mathcal{R} := \operatorname{Combine} \mathcal{P}_{g-1}$ and $\mathcal{Q}'$ ; +11: Perform non-dominated sorting on $\mathcal{R}$ +12: Update $\mathcal{P}_g$ by selecting top $N$ individuals via front rank and crowding distance; +13: end for +14: $\mathcal{P}^* \coloneqq$ Extract the first front from $\mathcal{P}_G$ as the set of Pareto-optimal solutions; + +linearity of attention. We include relu in FLatten [24] and introduce new activation functions such as swish and elu. (2) Normalization operations ensure a smoother distribution of attention scores. We provide options such as $l2\_norm$ and $min\_max\_norm$ . (3) The binary operators facilitate information interaction. Alongside cosine_similarity in Hydra [3], we also consider euclidean_distance as a simplified calculation option. As shown in Figure 4 (c), these operators enhance the versatility and adaptability of attention expressions, resulting in improved performance. Some complex operations such as decomposition [64] and Gaussian kernel [6] would result in a highly sparse search space or need hardware-specific supports. Hence they have not been included in our investigation. In future work, we will continue to involve more promising operators in our search space. + +# 3.3 Multi-objective Evolutionary Search + +After building the search space $\mathcal{A}$ , we sample the attention candidates $a \in \mathbb{N}$ and apply it to the ViT model with weights $W$ . For single-object search, the aim is to find the optimal attention $a_{single}^{*}$ by evaluating candidates using the trained ViT model $(W^{*})$ : + +$$ +a _ {s i n g l e} ^ {*} = \underset {a \in \mathcal {A}} {\operatorname {a r g m a x}} \mathrm {A C C} _ {\text {v a l}} \left(\mathcal {N} (a, W _ {\mathcal {A}} ^ {*} (a))\right) \tag {3} +$$ + +However, $a_{\text{single}}^*$ may perform well on the search model but fail to other ViT models. For example, attention searched in Swin, while superior on Swin, performs poorly on DeiT, PVT in Figure 5 (b). Such attentions suffer from limited usefulness because they are difficult to transfer directly to different tasks. + +Multi-objective Evolution. To improve generalization of attention candidates in different ViT models, our AttnZero framework employs a multi-objective evolutionary algorithm. This algorithm aims to maximize the results of multiple ViT models $1, 2, \ldots, n$ under predefined computational complexity constraints. We define a fitness function as follows: + +$$ +a _ {m u l t i} ^ {*} = \underset {a \in \mathcal {A}} {\operatorname {a r g m a x}} \mathrm {A C C} _ {\text {v a l}} \left(1 (a, W _ {1 \mathcal {A}} ^ {*} (a)), 2 (a, W _ {2 \mathcal {A}} ^ {*} (a)), \dots , n (a, W _ {n \mathcal {A}} ^ {*} (a)\right). \tag {4} +$$ + +Our framework incorporates validation results of attention in DeiT, AutoFormer, PVT, and Swin as objectives. This ensures that our attention module performs effectively in various ViT models, including pure ViT, ViT models with NAS search, ViT models with hierarchical design, and ViT models with local window configurations, respectively. Solving this multi-objective problem is challenging due to obstacles like conflicting optimals. To address this, we utilize NSGA-II method [10] in Alg. 1 as our search engine. Our search begins with an initial population of candidate attention modules and each candidate's fitness is evaluated based on performance across ViT models. Non-dominated sorting ranks the candidates and assigns them to fronts based on how they compare to others. Then, we use crowding distance [48] calculations to maintain diversity within the population. The selection process chooses candidates for the next generation, considering sorting and distance. Genetic operations like crossover and mutation are applied to the selected candidates. This introduces diversity and explores new areas of the search space. The least fit candidates are replaced with offspring, ensuring continuous gains. The algorithm iterates through these steps until it meets a termination criterion. During the search process, our algorithm takes FLOPs as the complexity constraint. Finally, we consider the real speed results in selecting final found attention from the Pareto-optimal solutions. + +Search Acceleration Strategies. To improve the search efficiency, we employ several strategies for search acceleration: (1) Program checking. We perform program checking by verifying if the single candidate attention module can successfully generate outputs using random tensor inputs. This helps us identify and filter out candidates with arithmetic errors or dimension mismatches. (2) Rejection protocols. During training each population, we reject candidates with collapsed optimisation (e.g., loss = nan) and historical average validation results falling far below baseline linear attention. These strategies allow more than $5 \times$ search acceleration. + +![](images/17e2105aea9e874689a4ff32514b6b417c03aa985fd5f751f138a3dd9f23b47b.jpg) +Figure 3: Discovered attention $a_{\text{AttnZero}}^*$ . + +# 3.4 Discovered AttnZero Analysis + +Formulas of discovered AttnZero. Figure 3 illustrates the optimal attention $a_{\text{AttnZero}}^*$ in our search, whose detailed expressions are as follows: + +$$ +a _ {A t t n Z e r o} ^ {*} = Q \times \eta_ {2} (\eta_ {2} (Q) \eta_ {2} (\phi_ {1} (K) ^ {T} \eta_ {1} (V))) \tag {5} +$$ + +![](images/520b46eac29d9fe4c997c4df847ec4863c00b63b6396bb20be371afb485809be.jpg) +Figure 4: (a) Performance distribution of subsets of Attn-Bench-101 (AutoFormer). (b) Results of various types of Attention. (c) Results of our default search space, default search space without advanced normalization operations, activation operations, and binary operations that we have newly introduced. More details are in the Appendix. + +![](images/1d0eb401a8c88a87220f4d7bc85f5415ec4969ce51805676e1093ce9a119b45a.jpg) + +![](images/062ebd17bcd295d0915d827f00f040f3e9c5abae1160fac9b879f827065a4bd2.jpg) + +where $\phi_{1}$ utilizes the Exponential Linear Unit (ELU) with a shift of 1, introducing nonlinearities into the attention computation. The $\eta_{1}$ performs min-max normalization, scaling the matrix values to a range between 0 and 1. This normalization step ensures that the attention weights are properly distributed and avoids potential issues with extreme values. Furthermore, the l2 normalization function $\eta_{2}$ guarantees that the attention weights have a unit norm, facilitating better interpretation and utilization of the attention mechanism. Regarding computational complexity, AttnZero requires triple matrix multiplications (i.e., $\mathcal{O}(d\times d + N\times d + N\times d)$ . Despite this, the overall time and memory complexity of AttnZero remains $\mathcal{O}(N)$ , which ensures AttnZero to handle computations for sequences of varying lengths efficiently. + +Analysis of Attn-Bench-101. Our Attn-Bench-101 collects 2,000 attention configurations and their training results in DeiT/AutoFormer/PVT/Swin on CIFAR-100 (see Appendix for details). We analyze some results on AutoFormer to understand better our search space: (1) In Figure 4 (a), we observe that our search space contains many candidates with stronger performance than baseline model. (2) Among various computed graphs, we identify i3n302 and i3n303, superior to the other types (see in Figure 4 (b)). This is attributed to their augmented topology allowing for a good tradeoff between performance and efficiency. (3) For ablations in Figure 4 (c), advanced activation and normalization operations offer more contributions compared to new binary operators. + +Guidance for Efficient Attention Design. Based on the above analysis, we can summarize that we prefer attention types of i3n302 & i3n303 and combinations of advanced normalization and activation operations in the detailed design. + +# 4 Search Experiments + +# 4.1 Search Implementation and Results + +Implementation. Our attention search is performed on CIFAR-100, which are scaled to $224 \times 224$ resolution. We search using the validation results to ensure + +Table 2: Top-1 accuracy (\%) of hand-crafted attentions (e.g., Linear [28] and FLatten [24]) and attention with Random NAS method in our search space on CIFAR-100. + +
ModelsBaselineLinear [28]FLatten [24]Random (NAS)AttnZero (NAS)
DeiT-T65.0857.8072.4576.0477.68 (+12.60)
AutoFormer66.5864.5875.4276.9478.61 (+12.03)
PVT-T67.4266.5874.8373.1276.68 (+9.26)
Swin-T68.2569.4574.3174.5875.90 (+7.65)
+ +fair comparisons, and our validation set does not overlap with the test set. For search settings, we configure $(N, p_c, p_m, G)$ in Alg. 1 as (20,0.9,0.1,100) for parallel search on DeiT|AutoFormer|PVT|Swin. To obtain validation results, we adopt standard training settings [31] (e.g., 300 epochs and AdamW optimizer). With our search acceleration strategies, the overall search only costs 1.05 days on the $8 \times \mathrm{V}100$ . More details are available in the Appendix. + +Comparison Results with Hand-Crafted Attentions. The results in Table 2 demonstrate the effectiveness of our AttnZero in consistently achieving stable boosts across different models. Firstly, when compared to the baseline, our AttnZero achieves significant improvements. For example, the DeiT-T model with our AttnZero improves from a baseline accuracy of $65.08\%$ to $77.68\%$ . Similar improvements are observed for the AutoFormer, PVT-T, and Swin-T models, indicating the consistent effectiveness of our approach. Furthermore, our AttnZero shows improvements over the recent state-of-the-art method. For example, the gain over FLatten for the DeiT-T model is $5.2\%$ . These findings emphasize the potential and significance of our AttnZero approach in improving the performance of various models. + +Comparing Random NAS Method. It is not easy to directly compare our method to other NAS methods because we use different search spaces and do not have weight sharing. To analyze our search strategy, we compare it to random search in our same search space. The results in Table 2 show that our AttnZero consistently improves model performance compared to random search. We achieve increases in accuracy ranging from $1.32\%$ to $3.56\%$ depending on the model. This demonstrates that our search strategy is more effective than a random search. + +# 4.2 Ablation Study + +Multi-Objective Search (MOS) vs. Single-Objective Search (SOS). We employ MOS to enable the discovery of attentions that exhibit strong generalization capabilities. Figure 5 (a) illustrates the optimal candidate resulting from our MOS process. The results demonstrate that the Top-3 candidates outperform the others consistently across various ViT models. In Figures 5 (b) and (c), we compare the generalization performance of MOS and SOS across different ViT models and datasets, respectively. The results indicate that while the SOS + +![](images/866aef81ba3ff9d806d18681fc1ec2c3e9ef54b315cfae3221f34997ff6080bc.jpg) +Figure 5: (a) Candidates in multi-objective search. (b) Comparisons of multi-objective vs. single-objective on multiple architectures; (c) Comparisons of multi-objective vs. single-objective on multiple datasets. (d) Search curves of various settings. + +![](images/91972eb13952e83dca49bf8fbebe25537795c895ee2b32c7a69e8e7998adbf45.jpg) + +![](images/9f414fea34a7b711a1d59acd34d3941d88b3711145e4fbf847e83a01e7201f3f.jpg) + +![](images/84e7f537f864be2cfba5c99218f9e6f960184e189f830c407442e5ab42f073a0.jpg) + +Table 3: Results of AttnZero for different ViT models and datasets. + +
ModelTop-1 Acc. (%)
Tiny ImageNet [1]Flowers [44]Chaoyang [74]
DeiT-T53.6250.0682.00
DeiT-T+AttnZero59.91 (+6.29)57.89 (+7.83)83.12 (+1.12)
AutoFormer-T56.3854.9882.84
AutoFormer-T+AttnZero62.56 (+6.18)61.58 (+6.60)83.68 (+0.84)
PVT-T58.4758.5782.46
PVT-T+AttnZero63.10 (+4.63)64.30 (+5.73)84.57 (+2.11)
Swin-T60.2258.8582.98
Swin-T+AttnZero64.55 (+4.33)65.13 (+6.28)85.11 (+2.13)
+ +method may yield better results on the searched models, MOS consistently outperforms SOS over a wider range of ViT architectures. Similarly, our AttnZero approach with MOS consistently demonstrates performance benefits across different dataset domains. For instance, our AttnZero with MOS achieves a $0.9\%$ improvement over SOS on the ImageNet dataset. + +Ablation of the Search Algorithm. In Figure 5 (d), we examine the evolutionary search and search acceleration strategies. The search results reveal the effectiveness of our acceleration strategy (i.e., program checking and rejection protocols) in filtering out unpromising candidates and greatly expediting the algorithm's convergence. These search curves also indicate that the evolutionary algorithm outperforms random search regarding efficiency in finding high-quality solutions within our framework. + +# 5 Transfer Experiments + +# 5.1 Experimental Results on Tiny Datasets + +Implementation. To check if our discovered attention generalizes beyond the CIFAR-100 dataset that we searched on, we it to multiple datasets (i.e., Tiny ImageNet [1], Flowers [44], and Chaoyang [74]). We train DeiT|AutoFormer|PVT|Swin + +Table 4: Results of AttnZero for DeiT & PVT on ImageNet $(224\times 224)$ + +
MethodParamsFLOPsTop-1 (%)
DeiT-T [58]5.7M1.2G72.2
DeiT-T+AttnZero5.7M1.0G74.9 (+2.7)
PVT-T [60]13.2M1.9G75.1
PVT-T+AttnZero10.9M1.8G78.1 (+3.0)
PVT-S24.5M3.8G79.8
PVT-S+AttnZero19.8M3.6G81.4 (+1.6)
PVT-M44.2M6.7G81.2
PVT-M+AttnZero34.6M6.6G83.0 (+1.8)
PVT-L61.4M9.8G81.7
PVT-L+AttnZero47.1M9.7G83.3 (+1.6)
+ +Table 5: Results of AttnZero on ImageNet with $224\times 224$ resolution inputs. + +
MethodParamsFLOPsTop-1 (%)
AutoFormer-T [58]5.7M1.3G74.7
AutoF.-T+AttnZero5.7M1.1G75.2 (+0.5)
Swin-T [41]29.0M4.5G81.3
Swin-T+AttnZero28.2M4.3G82.1 (+0.8)
Swin-S50.0M8.7G83.0
Swin-S+AttnZero49.5M8.5G83.2 (+0.2)
Swin-B88.0M15.4G83.5
Swin-B+AttnZero89.0M15.3G83.7 (+0.2)
CSwin-T [17]23.0M4.3G82.7
CSwin-T+AttnZero20.3M4.0G82.9 (+0.2)
+ +Table 6: Comparison of different efficient attention methods on ImageNet. + +
MethodResoParamsFLOPsTop-1 (%)
DeiT-T [58]22425.7M1.2G72.2
DeiT-T+Hydra Attn [3]22425.7M1.1G68.3
DeiT-T+Efficient Attn [54]22425.7M1.1G70.2
DeiT-T+Linear Angular [66]22425.7M1.1G70.8
DeiT-T+FLatten22426.1M1.1G74.1
DeiT-T+AttnZero22425.7M1.0G74.9 (+2.7)
+ +using discovered attention from scratch. The training process follows standard settings [31], including 300 training epochs, a cosine learning rate scheduler, and the AdamW optimizer. Detailed settings are shown in the Appendix. + +Results on Various Datasets. Table 3 highlights the effectiveness of AttnZero across various ViT models and datasets. It consistently surpasses the baselines, resulting in a $6.29\%$ increase in accuracy for the DeiT-T model on Tiny-ImageNet. Similarly, AutoFormer and PVT-T models with AttnZero exhibit similar improvements. On the Flowers dataset, AttnZero achieves an accuracy of $57.89\%$ for the DeiT-T model, surpassing the baseline by $7.83\%$ . AutoFormer and PVT-T models also benefit from AttnZero. The results on the Chaoyang dataset further exemplify AttnZero's capability to enhance accuracies. On the DeiT-T model, AttnZero achieves an accuracy of $83.12\%$ , compared to the $82\%$ baseline. Moreover, the attention modules from AttnZero contribute to the improved performance of AutoFormer, PVT-T, and Swin-T models compared to their original versions. These findings demonstrate the promising ability of AttnZero across different datasets and models. + +# 5.2 Experimental Results on Large-scale Datasets + +Implementation. We conduct the experiment on the ImageNet [52] dataset, a widely used large-scale dataset consisting of 1000 classes and 1.2 million training images. Note that we do not perform additional searches or modify the exist- + +Table 7: Results for Mask R-CNN on COCO dataset. We compute the FLOPs across the backbone, FPN, and detection head using an input resolution of $1280 \times 800$ . + +
MethodFLOPsSch.\( AP^b \)\( AP_{50}^b \)\( AP_{75}^b \)\( AP^m \)\( AP_{50}^m \)\( AP_{75}^m \)
PVT-T240G1x36.759.239.335.156.737.3
PVT-T+FLatten244G1x38.261.641.937.057.639.0
PVT-T+AttnZero209G1x40.062.643.137.459.440.1
PVT-S305G1x40.462.943.837.860.140.3
PVT-S+AttnZero248G1x41.063.445.038.260.441.1
+ +ing classical models. Instead, we directly replace the self-attention mechanism with our discovered attention module. Furthermore, we ensure that our training settings remain consistent with the standard settings [17, 41, 58]. This includes training the models for 300 epochs, cosine learning rate scheduler, and AdamW optimizer. Specific training settings are shown in the Appendix. + +Results on Various Backbones. Table 4 and Table 5 provides the performance results of our AttnZero on ImageNet. The results show that our method can transfer effectively to different ViT models. When applied to DeiT-T, AttnZero achieves a significant $2.7\%$ gain over baseline. Similarly, AttnZero improves PVT-T accuracy to $78.1\%$ , surpassing the baseline by $3.0\%$ . This demonstrates that AttnZero can seamlessly migrate and consistently enhance performance across ViT architectures. AttnZero also provides stable gains across models of varying scales. Paired with PVT-S, it boosts accuracy by $1.6\%$ . Applied to larger models like PVT-M, PVT-L, Swin-T, Swin-S, and CSwin-T, improvements of $1.8\%$ , $1.6\%$ , $0.8\%$ , $0.2\%$ and $0.2\%$ respectively are observed. These findings highlight AttnZero's stability in boosting the performance of ViT models from small to large scale. These results highlight the stability of our AttnZero in enhancing the performance of ViT models across different scales. + +Comparison with Other Linear Attention. Table 6 compares various linear attention methods. The results demonstrate that AttnZero surpasses other linear attention methods and baseline when incorporated into the DeiT-T model. Notably, compared to the Softmax baseline, AttnZero achieves a $2.7\%$ improvement. In contrast, other linear attention methods, such as Hydra Attn, Efficient Attn, Linear Angular, and FLatten, exhibit lower accuracies. These findings highlight the superiority of AttnZero in preserving or surpassing the performance achieved by Softmax attention. On the other hand, other linear attention methods tend to result in notable performance degradation. + +# 5.3 Transfer Experiments on Detection and Segmentation + +Object Detection. We evaluate our methods for object detection and instance segmentation tasks on the COCO dataset [39]. Table 7 provides results on Mask R-CNN detector [25] with $1 \times$ schedules. Our proposed PVT-T+AttnZero model achieves 2.3 box AP gains on PVT-T and significantly outperforms FLatten. + +Table 8: Results of semantic segmentation on ADE20K. We compute the FLOPs for both the encoders and decoders using input images at a resolution of $512 \times 2048$ . + +
BackboneMethodFLOPs#ParamsmIoUmAcc
PVT-TSemanticFPN158G17M36.5746.72
PVT-T+FLattenSemanticFPN169G16M37.2148.95
PVT-T+AttnZeroSemanticFPN126G15M39.1650.30
PVT-SSemanticFPN225G28M41.9553.02
PVT-S+AttnZeroSemanticFPN165G23M42.6254.02
+ +These enhancements demonstrate the benefits of incorporating our found attention module within the popular ViT model for object detection tasks. + +Semantic Segmentation. We assess the performance of our model using the ADE20K dataset [71]. Table 8 provides semantic segmentation results on SemanticFPN [29]. Among them, the Flatten module yields slight gains on the baseline PVT-T model. However, our AttnZero model achieves even higher mIoU (39.16) and mAcc (50.30) scores while reducing the FLOPs and parameters. Similarly, AttnZero achieves clear gains on PVT-S. These results highlight the compatibility of our AttnZero with different backbone architectures and its consistent improvements across various scenarios. + +# 6 Conclusion + +In this paper, we introduce AttnZero, the first method that automates the discovery of efficient attention modules that can be applied to various ViT models. AttnZero constructs a comprehensive search space consisting of different unary and binary operators as potential attention mechanisms. It utilizes a multi-objective genetic algorithm to search within this design space efficiently. Additionally, we introduce Attn-Bench-101, which provides attention profile results to summarize the insights gained from the search. Extensive experiments demonstrate that the attention modules discovered by AttnZero can be successfully applied to advanced ViTs for image classification, segmentation, and detection tasks. For future work, we will consider extending the AttnZero with knowledge distillation methods [32-35, 38, 53, 63]. Limitation. We present search strategies but still involve some search costs following most AutoML methods. But these costs are worthwhile because our discovered attention can be effectively applied to different ViTs and tasks without additional searching (proven by our extensive transfer experiments). Social impact. Our AttnZero focuses on technology improvement without social and ethical implications. + +# Acknowledgements + +The research was supported by Theme-based Research Scheme (T45-205/21-N) from Hong Kong RGC, and Generative AI Research and Development Centre from InnoHK. + +# References + +1. http://tiny-imagenet.herokuapp.com/ +2. Ali, A., Touvron, H., Caron, M., Bojanowski, P., Douze, M., Joulin, A., Laptev, I., Neverova, N., Synnaeve, G., Verbeek, J., et al.: Xcit: Cross-covariance image transformers. NIPS (2021) +3. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Hoffman, J.: Hydra attention: Efficient attention with many heads. In: ECCVW (2022) +4. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. arXiv preprint arXiv:2005.14165 (2020) +5. Chen, M., Peng, H., Fu, J., Ling, H.: Autoformer: Searching transformers for visual recognition. In: ICCV. pp. 12270-12280 (2021) +6. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. In: ICLR (2021) +7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. JMLR (2023) +8. Chu, X., Tian, Z., Zhang, B., Wang, X., Wei, X., Xia, H., Shen, C.: Conditional positional encodings for vision transformers. Arxiv preprint 2102.10882 (2021), https://arxiv.org/pdf/2102.10882.pdf +9. Chu, X., Zhang, B., Xu, R.: Fairnas: Rethinking evaluation fairness of weight sharing neural architecture search. In: ICCV (2021) +0. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary computation (2002) +1. Dong, P., Li, L., Tang, Z., Liu, X., Pan, X., Wang, Q., Chu, X.: Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. In: ICML (2024) +2. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023) +3. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV. pp. 17076-17086 (2023) +4. Dong, P., Niu, X., Li, L., Tian, Z., Wang, X., Wei, Z., Pan, H., Li, D.: Rd-nas: Enhancing one-shot supernet ranking ability via ranking distillation from zero-cost proxies. ICASSP (2023) +5. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022) +6. Dong, P., Niu, X., Tian, Z., Li, L., Wang, X., Wei, Z., Pan, H., Li, D.: Progressive meta-pooling learning for lightweight image classification model. In: ICASSP (2023) +7. Dong, X., Bao, J., Chen, D., Zhang, W., Yu, N., Yuan, L., Chen, D., Guo, B.: Cswin transformer: A general vision transformer backbone with cross-shaped windows. ArXiv abs/2107.00652 (2021) +8. Dong, X., Yang, Y.: Nas-bench-201: Extending the scope of reproducible neural architecture search (2020) +9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) + +20. Gao, J., Xu, H., Shi, H., Ren, X., Philip, L., Liang, X., Jiang, X., Li, Z.: Autobert-zero: Evolving bert backbone from scratch. In: AAAI (2022) +21. Guan, C., Wang, X., Zhu, W.: Autoattend: Automated attention representation search. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 3864-3874. PMLR (18-24 Jul 2021), http://proceedings.mlr.press/v139/guan21a.html +22. Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. arXiv preprint arXiv:1904.00420 (2019) +23. Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. In: ECCV. vol. abs/1904.00420, pp. 544-560. Springer (2019) +24. Han, D., Pan, X., Han, Y., Song, S., Huang, G.: Flatten transformer: Vision transformer using focused linear attention. In: ICCV (2023) +25. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: ICCV (2017) +26. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021) +27. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021) +28. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: ICML (2020) +29. Kirillov, A., Girshick, R., He, K., Dollár, P.: Panoptic feature pyramid networks. In: CVPR (2019) +30. Kitaev, N., Kaiser, L., Levskaya, A.: Reformer: The efficient transformer. In: ICLR (2020) +31. Li, K., Yu, R., Wang, Z., Yuan, L., Song, G., Chen, J.: Locality guidance for improving vision transformers on tiny datasets. In: ECCV. pp. 110-127. Springer (2022) +32. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022) +33. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024) +34. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023) +35. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022) +36. Li, L., Sun, H., Dong, P., Wei, Z., Shao, S.: Auto-das: Automated proxy discovery for training-free distillation-aware architecture search. In: ECCV (2024) +37. Li, L., Sun, H., Li, S., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Auto-gas: Automated proxy discovery for training-free generative architecture search. In: ECCV (2024) +38. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation (2020) +39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) +40. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019) +41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) + +42. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. In: NeurIPS (2021) +43. Nakai, K., Matsubara, T., Uehara, K.: Att-darts: Differentiable neural architecture search for attention. In: IJCNN. IEEE +44. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008) +45. Peng, H., Du, H., Yu, H., Li, Q., Liao, J., Fu, J.: Cream of the crop: Distilling prioritized paths for one-shot neural architecture search. NIPS (2020) +46. Pham, H., Guan, M.Y., Zoph, B., Le, Q.V., Dean, J.: Efficient neural architecture search via parameter sharing. In: ICML (2018) +47. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI (2022) +48. Raquel, C.R., Naval Jr, P.C.: An effective use of crowding distance in multiobjective particle swarm optimization. In: CGEC (2005) +49. Real, E., Aggarwal, A., Huang, Y., Le, Q.V.: Regularized evolution for image classifier architecture search. In: AAAI (2019) +50. Real, E., Liang, C., So, D., Le, Q.: Automl-zero: Evolving machine learning algorithms from scratch. In: ICML (2020) +51. Real, E., Liang, C., So, D., Le, Q.: Automl-zero: Evolving machine learning algorithms from scratch. In: ICML (2020) +52. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Li, F.F.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision (2015) +53. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023) +54. Shen, Z., Zhang, M., Zhao, H., Yi, S., Li, H.: Efficient attention: Attention with linear complexities. In: WACV (2021) +55. So, D.R., Manke, W., Liu, H., Dai, Z., Shazeer, N., Le, Q.V.: Primer: Searching for efficient transformers for language modeling. arXiv preprint arXiv:2109.08668 (2021) +56. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019) +57. Tay, Y., Dehghani, M., Bahri, D., Metzler, D.: Efficient transformers: A survey. arXiv preprint arXiv:2009.06732 (2020) +58. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021) +59. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. corr, abs/2302.13971, 2023. doi: 10.48550. arXiv preprint arXiv.2302.13971 +60. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021) +61. Wang, X., Xiong, X., Neumann, M., Piergiovanni, A., Ryoo, M.S., Angelova, A., Kitani, K.M., Hua, W.: Attentionnas: Spatiotemporal attention cell search for video classification. In: ECCV (2020) +62. Wei, Z., Pan, H., Li, L.L., Lu, M., Niu, X., Dong, P., Li, D.: Convformer: Closing the gap between cnn and vision transformers. arXiv preprint arXiv:2209.07738 (2022) + +63. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching (2022) +64. Xiong, Y., Zeng, Z., Chakraborty, R., Tan, M., Fung, G., Li, Y., Singh, V.: Nystromformer: A nystrom-based algorithm for approximating self-attention. In: AAAI (2021) +65. Yang, A., Esperança, P.M., Carlucci, F.M.: Nas evaluation is frustratingly hard. arXiv preprint arXiv:1912.12522 (2019) +66. You, H., Xiong, Y., Dai, X., Wu, B., Zhang, P., Fan, H., Vajda, P., Lin, Y.C.: Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference. In: CVPR (2023) +67. You, S., Huang, T., Yang, M., Wang, F., Qian, C., Zhang, C.: Greedynas: Towards fast one-shot nas with greedy supernet. In: CVPR (2020) +68. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: CVPR (2022) +69. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token vit: Training vision transformers from scratch onImagenet. In: ICCV (2021) +70. Zheng, Z., Gao, X., Pan, J., Luo, Q., Chen, G., Liu, D., Jiang, J.: Autoattention: Automatic field pair selection for attention in user behavior modeling. In: 2022 IEEE International Conference on Data Mining (ICDM). pp. 803-812. IEEE (2022) +71. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019) +72. Zhou, Q., Sheng, K., Zheng, X., Li, K., Sun, X., Tian, Y., Chen, J., Ji, R.: Training-free transformer architecture search. In: CVPR. pp. 10894-10903 (2022) +73. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024) +74. Zhu, C., Chen, W., Peng, T., Wang, Y., Jin, M.: Hard sample aware noise robust learning for histopathology image classification. TMI +75. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) +76. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024) +77. Zoph, B., Le, Q.V.: Neural architecture search with reinforcement learning. In: ICLR (2017) +78. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018) \ No newline at end of file diff --git a/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/images.zip b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..477fbb1619eb839776074d221875b6c15c041080 --- /dev/null +++ b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de5d63ad09bed2079008623eac50c1a6675420f7b0245542a2c9ebae5dd9c4f8 +size 592221 diff --git a/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/layout.json b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fb1572bfaf249cdb6b832466f935898c6a9cea2b --- /dev/null +++ b/2024/AttnZero_ Efficient Attention Discovery for Vision Transformers/layout.json @@ -0,0 +1,10599 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 146 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 146 + ], + "type": "text", + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "spans": [ + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "text", + "content": "Lujun Li" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Zimian Wei" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Peijie Dong" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Wenhan Luo" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Wei Xue" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Qifeng Liu" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "text", + "content": ", and Yike Guo" + }, + { + "bbox": [ + 145, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 201, + 421, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 201, + 421, + 213 + ], + "spans": [ + { + "bbox": [ + 190, + 201, + 421, + 213 + ], + "type": "text", + "content": "1 The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 213, + 399, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 213, + 399, + 224 + ], + "spans": [ + { + "bbox": [ + 214, + 213, + 399, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 214, + 213, + 399, + 224 + ], + "type": "text", + "content": " National University of Defense Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 224, + 448, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 224, + 448, + 236 + ], + "spans": [ + { + "bbox": [ + 164, + 224, + 448, + 236 + ], + "type": "text", + "content": "3 The Hong Kong University of Science and Technology (Guangzhou)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 496, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 496, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 496, + 258 + ], + "type": "text", + "content": "lilujunai@gmail.com;weizimian16@nudt.edu.cn;pdong212@connect.hkust-gz.edu.cn;whluo.china@gmail.com;weixue@ust.hk;liuqifeng@ust.hk;yikeguo@ust.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 283, + 455, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 283, + 455, + 524 + ], + "spans": [ + { + "bbox": [ + 159, + 283, + 455, + 524 + ], + "type": "text", + "content": "Abstract In this paper, we present AttnZero, the first framework for automatically discovering efficient attention modules tailored for Vision Transformers (ViTs). While traditional self-attention in ViTs suffers from quadratic computation complexity, linear attention offers a more efficient alternative with linear complexity approximation. However, existing hand-crafted linear attention suffers from performance degradation. To address these issues, our AttnZero constructs search spaces and employs evolutionary algorithms to discover potential linear attention formulations. Specifically, our search space consists of six kinds of computation graphs and advanced activation, normalize, and binary operators. To enhance generality, we derive results of candidate attention applied to multiple advanced ViTs as the multi-objective for the evolutionary search. To expedite the search process, we utilize program checking and rejection protocols to filter out unpromising candidates swiftly. Additionally, we develop Attn-Bench-101, which provides precomputed performance of 2,000 attentions in the search spaces, enabling us to summarize attention design insights. Experimental results demonstrate that the discovered AttnZero module generalizes well to different tasks and consistently achieves improved performance across various ViTs. For instance, the tiny model of DeiT|PVT|Swin|CSwin trained with AttnZero on ImageNet reaches " + }, + { + "bbox": [ + 159, + 283, + 455, + 524 + ], + "type": "inline_equation", + "content": "74.9\\%" + }, + { + "bbox": [ + 159, + 283, + 455, + 524 + ], + "type": "text", + "content": " |78.1%|82.1%|82.9% top-1 accuracy. Codes at: https://github.com/lliai/AttnZero." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 534, + 422, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 534, + 422, + 545 + ], + "spans": [ + { + "bbox": [ + 160, + 534, + 422, + 545 + ], + "type": "text", + "content": "Keywords: Vision Transformer " + }, + { + "bbox": [ + 160, + 534, + 422, + 545 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 534, + 422, + 545 + ], + "type": "text", + "content": " Efficient Attention " + }, + { + "bbox": [ + 160, + 534, + 422, + 545 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 534, + 422, + 545 + ], + "type": "text", + "content": " AutoML" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 563, + 230, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 230, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 230, + 577 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 587, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 648 + ], + "type": "text", + "content": "In recent years, Vision Transformers (ViTs) demonstrate remarkable success in various vision tasks, including image classification [16, 19, 58, 62], object detection [60, 75], semantic segmentation [47]. ViTs draw inspiration from the transformative impact of Transformers in natural language processing, utilizing self-attention mechanisms to capture global dependencies. However, self-attention" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 654, + 252, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 654, + 252, + 666 + ], + "spans": [ + { + "bbox": [ + 140, + 654, + 252, + 666 + ], + "type": "text", + "content": "Corresponding authors." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "has a quadratic complexity that scales with the square of the input sequence length. This quadratic complexity limits the scalability of ViTs for handling large image inputs, thereby limiting their practical applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 323 + ], + "type": "text", + "content": "To address this issue, researchers have been exploring the use of linear attention [57] as a more computationally efficient alternative. Unlike self-attention, linear attention has a linear complexity approximation, making it more suitable for handling large input sequences. However, existing linear attention methods [42, 54] suffer from several limitations: (1) Linear attention often leads to reduced performance than softmax attention due to its simplifying assumptions. The simple linear operators used in linear attention have limited flexibility and produce noisier activations than nonlinear functions [24]. These performance degradations challenge leveraging linear attention as a viable substitute for self-attention in ViTs. (2) Hand-crafted linear attention designs are often specific to certain tasks or architectures, lacking generality across different scenarios. Additionally, the manual design is labor-intensive, time-consuming, and heavily reliant on expert knowledge. Consequently, these drawbacks hinder the scalability and accessibility of linear attention in ViTs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 638 + ], + "type": "text", + "content": "To address these challenges, we have made several observations. We notice that some large language models [4, 7] employ new operators in Transformers to improve performance. For example, LLaMA [59] currently uses RMSNorm and SiLU to prevent training crashes. These successes inspire us to assemble a series of advanced operators to enhance linear attention and explore their optimal combination formulations. However, the question remains: How can we efficiently solve this combination problem? Recent developments in AutoML, such as AutoML-Zero [50] and NAS methods [40, 46], offer a promising solution. These methods involve constructing an extensive search space encompassing various options and searching for optimal architectures. Nevertheless, existing NAS methods are not suitable for solving the formulation discovery problem in our case for several reasons. (1) Different search space. The search spaces in conventional NAS methods typically focus on network width/depth, block/cell types, and parametric operators [5, 56]. In contrast, our attention search space includes many non-parametric operators and different combinatorial types. This leads to an explosion in the search space and many unsuccessful candidates. (2) Different search algorithms. Most recent NAS methods employ weight-sharing strategies [14, 27] to reduce search costs. However, our search space is explosive and extremely sparse, meaning that using previous ways to find potential candidates is less efficient and effective. (3) Different search goals. Most NAS methods prioritize performance for a single task and may struggle to generalize well to different applications [65]. In contrast, our objective is to discover generalized attention mechanisms that can be applied to various ViT models and tasks. These challenges highlight that attention discovery is a new task that requires a unique search space and strategy compared to traditional NAS methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Based on the above observations, we propose AttnZero, a new framework that aims to discover superior linear attention formulations that achieve a better" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 112, + 482, + 236 + ], + "blocks": [ + { + "bbox": [ + 134, + 112, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 134, + 112, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 112, + 482, + 236 + ], + "type": "image", + "image_path": "63c60c60f0fc55f05083d1da86342b05f18793ce17e402d4d2b05d7c6a217c71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 480, + 300 + ], + "type": "text", + "content": "Figure 1: Overall pipeline of AttnZero. Our framework obtains candidate attention populations by sampling from attention search space. With program checking and rejection protocols, we conduct validation results of these attentions on multiple ViTs. Then, we use NSGA-II to obtain the optimal solution and cross-mutate for the next population." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 328, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 544 + ], + "type": "text", + "content": "accuracy-efficiency tradeoff. As shown in Figure 1, our framework can be divided into two parts: a comprehensive attention search space and a multi-objective evolutionary search. Based on the basic form of linear attention, we develop six computation graphs to represent attention candidates in the search space. Then, we combine unary and binary operators with advanced functions, including various activation functions, normalization techniques, and matrix arithmetic functions. This approach allows us to obtain flexible and effective attention candidates. Our goal for the search algorithm is to find attention formulations applicable across different ViT models. To achieve this, we model the search task as a multi-objective optimization problem and employ the NSGA-II evolution method to obtain optimal solutions. During the search process, we validate candidate attention in multiple ViT models through multi-architecture evaluations. Based on these evaluations, our search engine exploits crossover and mutation for better populations. To accelerate the search process, we perform a fast program check and a rejection protocol to eliminate unpromising candidates during the search process. In this way, our AttnZero improves at least " + }, + { + "bbox": [ + 130, + 328, + 482, + 544 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 130, + 328, + 482, + 544 + ], + "type": "text", + "content": " in search efficiency without the weight-sharing setting [22], making our method free of ranking issues in traditional NAS methods [9, 18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "text", + "content": "We conduct comprehensive experiments to evaluate the transferability of our discovered attention mechanisms across different ViT models, datasets, and tasks without requiring additional searches. When applied to various ViT models, our AttnZero consistently outperforms the baseline method. For instance, our method achieves gains of " + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "inline_equation", + "content": "3.0\\% \\sim 1.6\\%" + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "text", + "content": " on PVT-T to PVT-L on the ImageNet dataset. Furthermore, ViT models with AttnZero demonstrate significant improvements when considering different data domains. For example, DeiT+AttnZero achieves gains of " + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "inline_equation", + "content": "7.8\\%" + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 130, + 546, + 491, + 665 + ], + "type": "text", + "content": " on the Tiny-ImageNet, Flowers, Chaoyang, and ImageNet datasets, respectively. Additionally, we compile a new benchmark named Attn-Bench-101, which consists of 2,000 attentions from our" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "text", + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "search space. The benchmark highlights the preference for advanced normalization techniques, activation functions, and well-performing attentions. In conclusion, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 163, + 480, + 306 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 163, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 479, + 209 + ], + "type": "text", + "content": "- To improve linear attention, we present a new task, efficient attention discovery, that generalizes to various ViT models and tasks. For this task, our AttnZero, the first framework, is developed to achieve effective attention search." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 212, + 480, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 212, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 139, + 212, + 480, + 246 + ], + "type": "text", + "content": "- We propose a comprehensive attention search space, including different types of computation graphs and advanced operators as options. We develop multi-objective searches with multiple ViTs evaluations to ensure generalisability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 247, + 479, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 247, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 139, + 247, + 479, + 282 + ], + "type": "text", + "content": "- Our found attention can be well generalized to DeiT|PVT|Swin|CSwin and other ViTs. Extensive experiments on multiple datasets in classification, detection, and segmentation prove the superior performance of our AttnZero." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 284, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 284, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 139, + 284, + 479, + 306 + ], + "type": "text", + "content": "- We build Attn-Bench-101, providing many attentions and their results. Some attention design guidances are summarised based on these results." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 327, + 243, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 243, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 243, + 340 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "type": "text", + "content": "Efficient Attention. The application of self-attention in computer vision has made significant progress [2,8,26,68,69] with the introduction of ViT [19]. Then, PVT [60] introduces a hierarchical structure while Swin Transformer [41] proposes window-based attention to model dependencies efficiently. However, softmax attention's quadratic complexity poses challenges. Linear attention reduces this but often degrades performance [6,30,64]. Efficient Attention [54] applies softmax to Q and K while SOFT [42] uses matrix decomposition. Castling-ViT [66] employs softmax attention during training but linear attention at inference. FLatten [24] preserves diversity through focused functions. While effective, these methods still have performance limitations and require complex manual design. In contrast, our AttnZero automatically discovers novel formulations to enhance linear attention. Rather than replacing prior work, we aim to improve the performance of the existing method while reducing manual effort. We build our search space based on previous work and provide feedback on new patterns. Thus, AttnZero complements rather than competes with manual methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Neural Architecture Search (NAS). NAS methods automate model design through search spaces and algorithms [49,77]. Early NAS works adopt reinforcement learning with expensive costs [78]. Recent one-shot NAS [15] and zero-shot NAS [11-13,36,37,73,76] use techniques like weight sharing [23] and training-free proxy evaluation to speed up search. Current NAS methods on ViTs [5,45,72] (e.g., AutoFormer [5]) search the number of depth/head of transformers. In contrast, AttnZero focuses on discovering attention mechanisms, not architectural dimensions. We also apply AttnZero to AutoFormer and observe consistent improvements, indicating our method is orthogonal to existing ViT-NAS methods. There are attempts to search entire attention blocks [67], SE-like attention [43,61], projection layers [21], and Transformers searches in other fields like" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "CTR [70] and NLP [20,55]. We clarify that our AttnZero is the first to search for efficient attention in visual tasks, which clearly differs in spaces/strategies/goals from previous methods. Inspired by general idea in AutoML-Zero [51], we achieve many new designs in search space/strategies for new attention discovery task, and opens up new possibilities for enhancing the capabilities of ViTs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 195, + 411, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 195, + 411, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 195, + 411, + 209 + ], + "type": "text", + "content": "3 Methodology: Efficient Attention Discovery" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 479, + 246 + ], + "type": "text", + "content": "In this section, we first review the linear attention form. Then, we present our search space and algorithms. The overall process is shown in Figure 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 264, + 389, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 264, + 389, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 264, + 389, + 276 + ], + "type": "text", + "content": "3.1 Recap of Self-Attention and Linear Attention" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "type": "text", + "content": "In self-attention, input tokens are projected into query " + }, + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 131, + 285, + 479, + 309 + ], + "type": "text", + "content": " matrices, respectively. The general form of self-attention is given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 241, + 318, + 480, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 318, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 241, + 318, + 480, + 348 + ], + "type": "interline_equation", + "content": "O _ {i} = \\sum_ {j = 1} ^ {N} \\frac {\\operatorname {S i m} \\left(Q _ {i} , K _ {j}\\right)}{\\sum_ {j = 1} ^ {N} \\operatorname {S i m} \\left(Q _ {i} , K _ {j}\\right)} V _ {j}, \\tag {1}", + "image_path": "4a0e46b96f7c1bb01cdc527c28868c39cbfc1c1b8bc4dcada499b9926a1ee483.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "text", + "content": " represents the output of the " + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "text", + "content": "-th token and Sim denotes the similarity function. The commonly used similarity function is the softmax function, " + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "inline_equation", + "content": "\\exp\\left(QK^{T}/\\sqrt{d}\\right)" + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "text", + "content": " is the dimension of the input tokens. However, the quadratic time and space complexity of softmax attention, " + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N^{2}d)" + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "text", + "content": ", becomes impractical for large numbers of tokens (" + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 355, + 480, + 433 + ], + "type": "text", + "content": "). To address this, linear attention is explored as an efficient alternative, which can be expressed as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 241, + 443, + 480, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 443, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 241, + 443, + 480, + 479 + ], + "type": "interline_equation", + "content": "O _ {i} = \\frac {\\phi \\left(Q _ {i}\\right) \\left(\\sum_ {j = 1} ^ {N} \\phi \\left(K _ {j}\\right) ^ {T} V _ {j}\\right)}{\\phi \\left(Q _ {i}\\right) \\left(\\sum_ {j = 1} ^ {N} \\phi \\left(K _ {j}\\right) ^ {T}\\right)}, \\tag {2}", + "image_path": "c65f9cfdcfb864668e2aa00777beb50b1b477a3c114e04187b8c506e33ec1863.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "text", + "content": " represents the kernel function. Linear attention reduces the computational complexity to " + }, + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N)" + }, + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "text", + "content": ", making it scalable for large input sequences. This reduction is achieved by using carefully designed kernels to approximate the similarity function, enabling a change in the computation order from " + }, + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "inline_equation", + "content": "(QK^T)V" + }, + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "inline_equation", + "content": "Q(K^TV)" + }, + { + "bbox": [ + 130, + 482, + 479, + 542 + ], + "type": "text", + "content": ", exploiting the associative property of matrix multiplication." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 560, + 324, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 324, + 573 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 324, + 573 + ], + "type": "text", + "content": "3.2 Efficient Attention Search Space" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "content": "Attention Representation. In Figure 2, we construct six types of computational graphs using (Q, K, V) as inputs to represent our candidate attentions. Among them, our i3n2 attentions follow the same process as standard linear attention, and we improve it by searching for different unary operations. Partially inspired by Castling-ViT [66], we introduce novel attention mechanisms with three binary operator nodes based on basic linear attention. With additional operators with either " + }, + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N\\times d)" + }, + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(d\\times d)" + }, + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "content": " computational complexity, the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "text", + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 463, + 232 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 463, + 232 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 463, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 463, + 232 + ], + "type": "image", + "image_path": "51fd63600db8eba61318eb39388f26ad2caf22a7c44ed344c4bb43d7e4a3bb06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 295 + ], + "type": "text", + "content": "Figure 2: Schematic of the six computational graphs in our search space. We denote the graph with 3 inputs and 2 binary operations as i3n2. Similarly, i3n3 refers to a graph with 3 inputs and 3 binary operations. Since each binary operation takes two inputs, there are 5 possible graph structures for i3n3. We label these i3n301~i3n305, respectively. For simplicity, we do not display unary operators here." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 136, + 338, + 264, + 486 + ], + "blocks": [ + { + "bbox": [ + 131, + 305, + 482, + 329 + ], + "lines": [ + { + "bbox": [ + 131, + 305, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 305, + 482, + 329 + ], + "type": "text", + "content": "Table 1: Some operations in our attention search space. \"UOP\" and \"BOP\" denote the unary and binary operation, respectively. Full operations are in the Appendix." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 338, + 264, + 486 + ], + "lines": [ + { + "bbox": [ + 136, + 338, + 264, + 486 + ], + "spans": [ + { + "bbox": [ + 136, + 338, + 264, + 486 + ], + "type": "table", + "html": "
OP IDOP NameExpression
UOP00no_opx
UOP01relumax(0,x)
UOP02scalex/√dx
UOP03sqrt√x
UOP04invert1/x+e-6
UOP05l2_normx-mean(x)/std(x)
UOP06sigmoid1/1+e-x
UOP07logsoftmaxln ex/∑i=1n e^xi
UOP08softmaxex/∑i=1n e^xi
UOP09softsignx/1+|x|
", + "image_path": "3dc12ee9974e742361b089770d54da96cb41fc9c04e3908225292d45897921eb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 291, + 342, + 443, + 483 + ], + "blocks": [ + { + "bbox": [ + 291, + 342, + 443, + 483 + ], + "lines": [ + { + "bbox": [ + 291, + 342, + 443, + 483 + ], + "spans": [ + { + "bbox": [ + 291, + 342, + 443, + 483 + ], + "type": "table", + "html": "
OP IDOP NameExpression
UOP10eluelu(x) + 1
UOP11sigmoid_revert1 - sigmoid(x)
UOP12expex
UOP13abslog|ln x|
UOP14min_max_normx-min(x)/max(x)-min(x)
UOP15transposexT
UOP16swishx × sigmoid (x)
BOP01sumx + y
BOP02cosine_simcos(x,y)
BOP03productx ⊙ y
BOP04matrix-multix · y
BOP05euclidean_disd(x,y)
", + "image_path": "1ea2e9c8fae53c5335b4c85e190dfcc3069753959027e1e638bf222b5f897af5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "text", + "content": "performance of these attention mechanisms can be enhanced without a significant increase in computational overhead. Based on the different binary nodes, these augmented attentions can be categorized into five types (i3n301 to i3n305), effectively enhancing the modeling capability. We divide the sub-search space according to these different types of attention. During the search process, we first search the solutions within the sub-search space in parallel and then combine them to obtain the final search result. Figure 4 (b) shows performance for different attention candidates. These designs allow our method to balance computational complexity and accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Primitive Operations. Table 1 presents some unary and binary operators in our search space. These operators encompass both commonly used operations found in existing linear attentions [3, 24, 54, 66] and our proposed ones, including: (1) The activation operations are responsible for enhancing the non-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 376, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 376, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 376, + 127 + ], + "type": "text", + "content": "Algorithm 1 NSGA-II Evolution Search for AttnZero." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "text", + "content": "Input: Search space " + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "text", + "content": ", population size " + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "text", + "content": ", crossover probability " + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "inline_equation", + "content": "p_c" + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "text", + "content": ", mutation probability " + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "text", + "content": ", max generations " + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 132, + 129, + 481, + 150 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 151, + 353, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 353, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 353, + 161 + ], + "type": "text", + "content": "Output: Set of Pareto-optimal attention formulations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 163, + 480, + 317 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 136, + 163, + 411, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 163, + 411, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 163, + 411, + 175 + ], + "type": "text", + "content": "1: Initialize population " + }, + { + "bbox": [ + 136, + 163, + 411, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_0" + }, + { + "bbox": [ + 136, + 163, + 411, + 175 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 136, + 163, + 411, + 175 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 136, + 163, + 411, + 175 + ], + "type": "text", + "content": " individuals generated randomly;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 175, + 401, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 175, + 401, + 186 + ], + "spans": [ + { + "bbox": [ + 136, + 175, + 401, + 186 + ], + "type": "text", + "content": "2: Evaluate the fitness of each individual in " + }, + { + "bbox": [ + 136, + 175, + 401, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_0" + }, + { + "bbox": [ + 136, + 175, + 401, + 186 + ], + "type": "text", + "content": " for all objectives;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 186, + 225, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 186, + 225, + 196 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 225, + 196 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 136, + 186, + 225, + 196 + ], + "type": "inline_equation", + "content": "g = 1" + }, + { + "bbox": [ + 136, + 186, + 225, + 196 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 136, + 186, + 225, + 196 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 136, + 186, + 225, + 196 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 197, + 456, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 197, + 456, + 207 + ], + "spans": [ + { + "bbox": [ + 136, + 197, + 456, + 207 + ], + "type": "text", + "content": "4: Perform non-dominated sorting on " + }, + { + "bbox": [ + 136, + 197, + 456, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{g - 1}" + }, + { + "bbox": [ + 136, + 197, + 456, + 207 + ], + "type": "text", + "content": " to classify into fronts " + }, + { + "bbox": [ + 136, + 197, + 456, + 207 + ], + "type": "inline_equation", + "content": "F_{1},F_{2},\\ldots" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 208, + 391, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 208, + 391, + 218 + ], + "spans": [ + { + "bbox": [ + 136, + 208, + 391, + 218 + ], + "type": "text", + "content": "5: Calculate crowding distance for individuals in each front;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 219, + 465, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 219, + 465, + 229 + ], + "spans": [ + { + "bbox": [ + 136, + 219, + 465, + 229 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 136, + 219, + 465, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} \\coloneqq \\operatorname{Select} N" + }, + { + "bbox": [ + 136, + 219, + 465, + 229 + ], + "type": "text", + "content": " individuals for mating via front rank and crowding distance;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "spans": [ + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "text", + "content": "7: Generate offspring " + }, + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}'" + }, + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "text", + "content": " via crossover & mutation on " + }, + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "text", + "content": " with prob. " + }, + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "inline_equation", + "content": "p_c" + }, + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 136, + 230, + 467, + 240 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 241, + 405, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 241, + 405, + 251 + ], + "spans": [ + { + "bbox": [ + 136, + 241, + 405, + 251 + ], + "type": "text", + "content": "8: Evaluate the fitness of each offspring in " + }, + { + "bbox": [ + 136, + 241, + 405, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}'" + }, + { + "bbox": [ + 136, + 241, + 405, + 251 + ], + "type": "text", + "content": " for all objectives;" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 251, + 473, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 251, + 473, + 262 + ], + "spans": [ + { + "bbox": [ + 135, + 251, + 473, + 262 + ], + "type": "text", + "content": "9: Apply rejection protocols to " + }, + { + "bbox": [ + 135, + 251, + 473, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}'" + }, + { + "bbox": [ + 135, + 251, + 473, + 262 + ], + "type": "text", + "content": " to remove unpromising candidate and get " + }, + { + "bbox": [ + 135, + 251, + 473, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}''" + }, + { + "bbox": [ + 135, + 251, + 473, + 262 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 263, + 280, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 263, + 280, + 274 + ], + "spans": [ + { + "bbox": [ + 133, + 263, + 280, + 274 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 133, + 263, + 280, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{R} := \\operatorname{Combine} \\mathcal{P}_{g-1}" + }, + { + "bbox": [ + 133, + 263, + 280, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 263, + 280, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}'" + }, + { + "bbox": [ + 133, + 263, + 280, + 274 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 274, + 314, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 274, + 314, + 284 + ], + "spans": [ + { + "bbox": [ + 133, + 274, + 314, + 284 + ], + "type": "text", + "content": "11: Perform non-dominated sorting on " + }, + { + "bbox": [ + 133, + 274, + 314, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 285, + 480, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 285, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 480, + 296 + ], + "type": "text", + "content": "12: Update " + }, + { + "bbox": [ + 133, + 285, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_g" + }, + { + "bbox": [ + 133, + 285, + 480, + 296 + ], + "type": "text", + "content": " by selecting top " + }, + { + "bbox": [ + 133, + 285, + 480, + 296 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 133, + 285, + 480, + 296 + ], + "type": "text", + "content": " individuals via front rank and crowding distance;" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 297, + 184, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 297, + 184, + 305 + ], + "spans": [ + { + "bbox": [ + 133, + 297, + 184, + 305 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 306, + 460, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 460, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 460, + 317 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 132, + 306, + 460, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^* \\coloneqq" + }, + { + "bbox": [ + 132, + 306, + 460, + 317 + ], + "type": "text", + "content": " Extract the first front from " + }, + { + "bbox": [ + 132, + 306, + 460, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_G" + }, + { + "bbox": [ + 132, + 306, + 460, + 317 + ], + "type": "text", + "content": " as the set of Pareto-optimal solutions;" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 345, + 489, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 489, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 489, + 489 + ], + "type": "text", + "content": "linearity of attention. We include relu in FLatten [24] and introduce new activation functions such as swish and elu. (2) Normalization operations ensure a smoother distribution of attention scores. We provide options such as " + }, + { + "bbox": [ + 130, + 345, + 489, + 489 + ], + "type": "inline_equation", + "content": "l2\\_norm" + }, + { + "bbox": [ + 130, + 345, + 489, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 345, + 489, + 489 + ], + "type": "inline_equation", + "content": "min\\_max\\_norm" + }, + { + "bbox": [ + 130, + 345, + 489, + 489 + ], + "type": "text", + "content": ". (3) The binary operators facilitate information interaction. Alongside cosine_similarity in Hydra [3], we also consider euclidean_distance as a simplified calculation option. As shown in Figure 4 (c), these operators enhance the versatility and adaptability of attention expressions, resulting in improved performance. Some complex operations such as decomposition [64] and Gaussian kernel [6] would result in a highly sparse search space or need hardware-specific supports. Hence they have not been included in our investigation. In future work, we will continue to involve more promising operators in our search space." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 510, + 343, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 343, + 522 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 343, + 522 + ], + "type": "text", + "content": "3.3 Multi-objective Evolutionary Search" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": "After building the search space " + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": ", we sample the attention candidates " + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{N}" + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": " and apply it to the ViT model with weights " + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": ". For single-object search, the aim is to find the optimal attention " + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "inline_equation", + "content": "a_{single}^{*}" + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": " by evaluating candidates using the trained ViT model " + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "inline_equation", + "content": "(W^{*})" + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 224, + 590, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 590, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 224, + 590, + 481, + 609 + ], + "type": "interline_equation", + "content": "a _ {s i n g l e} ^ {*} = \\underset {a \\in \\mathcal {A}} {\\operatorname {a r g m a x}} \\mathrm {A C C} _ {\\text {v a l}} \\left(\\mathcal {N} (a, W _ {\\mathcal {A}} ^ {*} (a))\\right) \\tag {3}", + "image_path": "6344f412a0c201677d55d5300501a69e2c7a5c9e636d6102b335fa18f4853496.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "However, " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "a_{\\text{single}}^*" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " may perform well on the search model but fail to other ViT models. For example, attention searched in Swin, while superior on Swin, performs poorly on DeiT, PVT in Figure 5 (b). Such attentions suffer from limited usefulness because they are difficult to transfer directly to different tasks." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "text", + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "Multi-objective Evolution. To improve generalization of attention candidates in different ViT models, our AttnZero framework employs a multi-objective evolutionary algorithm. This algorithm aims to maximize the results of multiple ViT models " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "1, 2, \\ldots, n" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " under predefined computational complexity constraints. We define a fitness function as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 180, + 481, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 180, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 159, + 180, + 481, + 198 + ], + "type": "interline_equation", + "content": "a _ {m u l t i} ^ {*} = \\underset {a \\in \\mathcal {A}} {\\operatorname {a r g m a x}} \\mathrm {A C C} _ {\\text {v a l}} \\left(1 (a, W _ {1 \\mathcal {A}} ^ {*} (a)), 2 (a, W _ {2 \\mathcal {A}} ^ {*} (a)), \\dots , n (a, W _ {n \\mathcal {A}} ^ {*} (a)\\right). \\tag {4}", + "image_path": "f03ddc756dcd502b91707476b06b3c84cbbd8749d576f31d0041a67bab6454e8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 202, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 430 + ], + "type": "text", + "content": "Our framework incorporates validation results of attention in DeiT, AutoFormer, PVT, and Swin as objectives. This ensures that our attention module performs effectively in various ViT models, including pure ViT, ViT models with NAS search, ViT models with hierarchical design, and ViT models with local window configurations, respectively. Solving this multi-objective problem is challenging due to obstacles like conflicting optimals. To address this, we utilize NSGA-II method [10] in Alg. 1 as our search engine. Our search begins with an initial population of candidate attention modules and each candidate's fitness is evaluated based on performance across ViT models. Non-dominated sorting ranks the candidates and assigns them to fronts based on how they compare to others. Then, we use crowding distance [48] calculations to maintain diversity within the population. The selection process chooses candidates for the next generation, considering sorting and distance. Genetic operations like crossover and mutation are applied to the selected candidates. This introduces diversity and explores new areas of the search space. The least fit candidates are replaced with offspring, ensuring continuous gains. The algorithm iterates through these steps until it meets a termination criterion. During the search process, our algorithm takes FLOPs as the complexity constraint. Finally, we consider the real speed results in selecting final found attention from the Pareto-optimal solutions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 369, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 369, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 369, + 586 + ], + "type": "text", + "content": "Search Acceleration Strategies. To improve the search efficiency, we employ several strategies for search acceleration: (1) Program checking. We perform program checking by verifying if the single candidate attention module can successfully generate outputs using random tensor inputs. This helps us identify and filter out candidates with arithmetic errors or dimension mismatches. (2) Rejection protocols. During training each population, we reject candidates with collapsed optimisation (e.g., loss = nan) and historical average validation results falling far below baseline linear attention. These strategies allow more than " + }, + { + "bbox": [ + 130, + 430, + 369, + 586 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 130, + 430, + 369, + 586 + ], + "type": "text", + "content": " search acceleration." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 383, + 444, + 473, + 552 + ], + "blocks": [ + { + "bbox": [ + 383, + 444, + 473, + 552 + ], + "lines": [ + { + "bbox": [ + 383, + 444, + 473, + 552 + ], + "spans": [ + { + "bbox": [ + 383, + 444, + 473, + 552 + ], + "type": "image", + "image_path": "17e2105aea9e874689a4ff32514b6b417c03aa985fd5f751f138a3dd9f23b47b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 554, + 482, + 576 + ], + "lines": [ + { + "bbox": [ + 374, + 554, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 374, + 554, + 482, + 576 + ], + "type": "text", + "content": "Figure 3: Discovered attention " + }, + { + "bbox": [ + 374, + 554, + 482, + 576 + ], + "type": "inline_equation", + "content": "a_{\\text{AttnZero}}^*" + }, + { + "bbox": [ + 374, + 554, + 482, + 576 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 601, + 312, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 312, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 312, + 613 + ], + "type": "text", + "content": "3.4 Discovered AttnZero Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 618, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 482, + 643 + ], + "type": "text", + "content": "Formulas of discovered AttnZero. Figure 3 illustrates the optimal attention " + }, + { + "bbox": [ + 130, + 618, + 482, + 643 + ], + "type": "inline_equation", + "content": "a_{\\text{AttnZero}}^*" + }, + { + "bbox": [ + 130, + 618, + 482, + 643 + ], + "type": "text", + "content": " in our search, whose detailed expressions are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 196, + 647, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 647, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 196, + 647, + 481, + 669 + ], + "type": "interline_equation", + "content": "a _ {A t t n Z e r o} ^ {*} = Q \\times \\eta_ {2} (\\eta_ {2} (Q) \\eta_ {2} (\\phi_ {1} (K) ^ {T} \\eta_ {1} (V))) \\tag {5}", + "image_path": "e54e297586f32ab7e6917d2c093c679d7f31e31cc2b4dce150f4edf3b5f8d0a6.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 252, + 208 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 252, + 208 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 252, + 208 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 252, + 208 + ], + "type": "image", + "image_path": "520b46eac29d9fe4c997c4df847ec4863c00b63b6396bb20be371afb485809be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 216, + 480, + 260 + ], + "lines": [ + { + "bbox": [ + 131, + 216, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 480, + 260 + ], + "type": "text", + "content": "Figure 4: (a) Performance distribution of subsets of Attn-Bench-101 (AutoFormer). (b) Results of various types of Attention. (c) Results of our default search space, default search space without advanced normalization operations, activation operations, and binary operations that we have newly introduced. More details are in the Appendix." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 116, + 366, + 208 + ], + "blocks": [ + { + "bbox": [ + 253, + 116, + 366, + 208 + ], + "lines": [ + { + "bbox": [ + 253, + 116, + 366, + 208 + ], + "spans": [ + { + "bbox": [ + 253, + 116, + 366, + 208 + ], + "type": "image", + "image_path": "1d0eb401a8c88a87220f4d7bc85f5415ec4969ce51805676e1093ce9a119b45a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 366, + 116, + 480, + 208 + ], + "blocks": [ + { + "bbox": [ + 366, + 116, + 480, + 208 + ], + "lines": [ + { + "bbox": [ + 366, + 116, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 366, + 116, + 480, + 208 + ], + "type": "image", + "image_path": "062ebd17bcd295d0915d827f00f040f3e9c5abae1160fac9b879f827065a4bd2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "text", + "content": " utilizes the Exponential Linear Unit (ELU) with a shift of 1, introducing nonlinearities into the attention computation. The " + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\eta_{1}" + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "text", + "content": " performs min-max normalization, scaling the matrix values to a range between 0 and 1. This normalization step ensures that the attention weights are properly distributed and avoids potential issues with extreme values. Furthermore, the l2 normalization function " + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\eta_{2}" + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "text", + "content": " guarantees that the attention weights have a unit norm, facilitating better interpretation and utilization of the attention mechanism. Regarding computational complexity, AttnZero requires triple matrix multiplications (i.e., " + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(d\\times d + N\\times d + N\\times d)" + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "text", + "content": ". Despite this, the overall time and memory complexity of AttnZero remains " + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N)" + }, + { + "bbox": [ + 130, + 285, + 482, + 416 + ], + "type": "text", + "content": ", which ensures AttnZero to handle computations for sequences of varying lengths efficiently." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 417, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 536 + ], + "type": "text", + "content": "Analysis of Attn-Bench-101. Our Attn-Bench-101 collects 2,000 attention configurations and their training results in DeiT/AutoFormer/PVT/Swin on CIFAR-100 (see Appendix for details). We analyze some results on AutoFormer to understand better our search space: (1) In Figure 4 (a), we observe that our search space contains many candidates with stronger performance than baseline model. (2) Among various computed graphs, we identify i3n302 and i3n303, superior to the other types (see in Figure 4 (b)). This is attributed to their augmented topology allowing for a good tradeoff between performance and efficiency. (3) For ablations in Figure 4 (c), advanced activation and normalization operations offer more contributions compared to new binary operators." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 536, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 573 + ], + "type": "text", + "content": "Guidance for Efficient Attention Design. Based on the above analysis, we can summarize that we prefer attention types of i3n302 & i3n303 and combinations of advanced normalization and activation operations in the detailed design." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 593, + 272, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 593, + 272, + 607 + ], + "spans": [ + { + "bbox": [ + 131, + 593, + 272, + 607 + ], + "type": "text", + "content": "4 Search Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 620, + 340, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 340, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 340, + 632 + ], + "type": "text", + "content": "4.1 Search Implementation and Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Implementation. Our attention search is performed on CIFAR-100, which are scaled to " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " resolution. We search using the validation results to ensure" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "text", + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 146, + 475, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Top-1 accuracy (\\%) of hand-crafted attentions (e.g., Linear [28] and FLatten [24]) and attention with Random NAS method in our search space on CIFAR-100." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 146, + 475, + 216 + ], + "lines": [ + { + "bbox": [ + 138, + 146, + 475, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 146, + 475, + 216 + ], + "type": "table", + "html": "
ModelsBaselineLinear [28]FLatten [24]Random (NAS)AttnZero (NAS)
DeiT-T65.0857.8072.4576.0477.68 (+12.60)
AutoFormer66.5864.5875.4276.9478.61 (+12.03)
PVT-T67.4266.5874.8373.1276.68 (+9.26)
Swin-T68.2569.4574.3174.5875.90 (+7.65)
", + "image_path": "621aeb061e4e4d97300bc63b6c2f55a886325dc65605271b829bcd801dab81af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 310 + ], + "type": "text", + "content": "fair comparisons, and our validation set does not overlap with the test set. For search settings, we configure " + }, + { + "bbox": [ + 130, + 239, + 479, + 310 + ], + "type": "inline_equation", + "content": "(N, p_c, p_m, G)" + }, + { + "bbox": [ + 130, + 239, + 479, + 310 + ], + "type": "text", + "content": " in Alg. 1 as (20,0.9,0.1,100) for parallel search on DeiT|AutoFormer|PVT|Swin. To obtain validation results, we adopt standard training settings [31] (e.g., 300 epochs and AdamW optimizer). With our search acceleration strategies, the overall search only costs 1.05 days on the " + }, + { + "bbox": [ + 130, + 239, + 479, + 310 + ], + "type": "inline_equation", + "content": "8 \\times \\mathrm{V}100" + }, + { + "bbox": [ + 130, + 239, + 479, + 310 + ], + "type": "text", + "content": ". More details are available in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 311, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 480, + 441 + ], + "type": "text", + "content": "Comparison Results with Hand-Crafted Attentions. The results in Table 2 demonstrate the effectiveness of our AttnZero in consistently achieving stable boosts across different models. Firstly, when compared to the baseline, our AttnZero achieves significant improvements. For example, the DeiT-T model with our AttnZero improves from a baseline accuracy of " + }, + { + "bbox": [ + 130, + 311, + 480, + 441 + ], + "type": "inline_equation", + "content": "65.08\\%" + }, + { + "bbox": [ + 130, + 311, + 480, + 441 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 311, + 480, + 441 + ], + "type": "inline_equation", + "content": "77.68\\%" + }, + { + "bbox": [ + 130, + 311, + 480, + 441 + ], + "type": "text", + "content": ". Similar improvements are observed for the AutoFormer, PVT-T, and Swin-T models, indicating the consistent effectiveness of our approach. Furthermore, our AttnZero shows improvements over the recent state-of-the-art method. For example, the gain over FLatten for the DeiT-T model is " + }, + { + "bbox": [ + 130, + 311, + 480, + 441 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 130, + 311, + 480, + 441 + ], + "type": "text", + "content": ". These findings emphasize the potential and significance of our AttnZero approach in improving the performance of various models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 443, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 538 + ], + "type": "text", + "content": "Comparing Random NAS Method. It is not easy to directly compare our method to other NAS methods because we use different search spaces and do not have weight sharing. To analyze our search strategy, we compare it to random search in our same search space. The results in Table 2 show that our AttnZero consistently improves model performance compared to random search. We achieve increases in accuracy ranging from " + }, + { + "bbox": [ + 130, + 443, + 482, + 538 + ], + "type": "inline_equation", + "content": "1.32\\%" + }, + { + "bbox": [ + 130, + 443, + 482, + 538 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 443, + 482, + 538 + ], + "type": "inline_equation", + "content": "3.56\\%" + }, + { + "bbox": [ + 130, + 443, + 482, + 538 + ], + "type": "text", + "content": " depending on the model. This demonstrates that our search strategy is more effective than a random search." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 559, + 238, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 238, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 238, + 571 + ], + "type": "text", + "content": "4.2 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Multi-Objective Search (MOS) vs. Single-Objective Search (SOS). We employ MOS to enable the discovery of attentions that exhibit strong generalization capabilities. Figure 5 (a) illustrates the optimal candidate resulting from our MOS process. The results demonstrate that the Top-3 candidates outperform the others consistently across various ViT models. In Figures 5 (b) and (c), we compare the generalization performance of MOS and SOS across different ViT models and datasets, respectively. The results indicate that while the SOS" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 216, + 194 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 216, + 194 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 216, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 216, + 194 + ], + "type": "image", + "image_path": "866aef81ba3ff9d806d18681fc1ec2c3e9ef54b315cfae3221f34997ff6080bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 203, + 482, + 237 + ], + "lines": [ + { + "bbox": [ + 131, + 203, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 131, + 203, + 482, + 237 + ], + "type": "text", + "content": "Figure 5: (a) Candidates in multi-objective search. (b) Comparisons of multi-objective vs. single-objective on multiple architectures; (c) Comparisons of multi-objective vs. single-objective on multiple datasets. (d) Search curves of various settings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 114, + 291, + 194 + ], + "blocks": [ + { + "bbox": [ + 215, + 114, + 291, + 194 + ], + "lines": [ + { + "bbox": [ + 215, + 114, + 291, + 194 + ], + "spans": [ + { + "bbox": [ + 215, + 114, + 291, + 194 + ], + "type": "image", + "image_path": "91972eb13952e83dca49bf8fbebe25537795c895ee2b32c7a69e8e7998adbf45.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 292, + 114, + 369, + 194 + ], + "blocks": [ + { + "bbox": [ + 292, + 114, + 369, + 194 + ], + "lines": [ + { + "bbox": [ + 292, + 114, + 369, + 194 + ], + "spans": [ + { + "bbox": [ + 292, + 114, + 369, + 194 + ], + "type": "image", + "image_path": "9f414fea34a7b711a1d59acd34d3941d88b3711145e4fbf847e83a01e7201f3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 480, + 194 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 480, + 194 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 480, + 194 + ], + "type": "image", + "image_path": "84e7f537f864be2cfba5c99218f9e6f960184e189f830c407442e5ab42f073a0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 132, + 269, + 482, + 394 + ], + "blocks": [ + { + "bbox": [ + 165, + 248, + 447, + 259 + ], + "lines": [ + { + "bbox": [ + 165, + 248, + 447, + 259 + ], + "spans": [ + { + "bbox": [ + 165, + 248, + 447, + 259 + ], + "type": "text", + "content": "Table 3: Results of AttnZero for different ViT models and datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 269, + 482, + 394 + ], + "lines": [ + { + "bbox": [ + 132, + 269, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 482, + 394 + ], + "type": "table", + "html": "
ModelTop-1 Acc. (%)
Tiny ImageNet [1]Flowers [44]Chaoyang [74]
DeiT-T53.6250.0682.00
DeiT-T+AttnZero59.91 (+6.29)57.89 (+7.83)83.12 (+1.12)
AutoFormer-T56.3854.9882.84
AutoFormer-T+AttnZero62.56 (+6.18)61.58 (+6.60)83.68 (+0.84)
PVT-T58.4758.5782.46
PVT-T+AttnZero63.10 (+4.63)64.30 (+5.73)84.57 (+2.11)
Swin-T60.2258.8582.98
Swin-T+AttnZero64.55 (+4.33)65.13 (+6.28)85.11 (+2.13)
", + "image_path": "f62ecfe39aca8295b3d8de395364707c624e2e4a8be18f7139d68e163a6ceacd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 418, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 477 + ], + "type": "text", + "content": "method may yield better results on the searched models, MOS consistently outperforms SOS over a wider range of ViT architectures. Similarly, our AttnZero approach with MOS consistently demonstrates performance benefits across different dataset domains. For instance, our AttnZero with MOS achieves a " + }, + { + "bbox": [ + 130, + 418, + 482, + 477 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 418, + 482, + 477 + ], + "type": "text", + "content": " improvement over SOS on the ImageNet dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 478, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 561 + ], + "type": "text", + "content": "Ablation of the Search Algorithm. In Figure 5 (d), we examine the evolutionary search and search acceleration strategies. The search results reveal the effectiveness of our acceleration strategy (i.e., program checking and rejection protocols) in filtering out unpromising candidates and greatly expediting the algorithm's convergence. These search curves also indicate that the evolutionary algorithm outperforms random search regarding efficiency in finding high-quality solutions within our framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 582, + 284, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 284, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 284, + 596 + ], + "type": "text", + "content": "5 Transfer Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 608, + 359, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 608, + 359, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 608, + 359, + 620 + ], + "type": "text", + "content": "5.1 Experimental Results on Tiny Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 492, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 492, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 492, + 667 + ], + "type": "text", + "content": "Implementation. To check if our discovered attention generalizes beyond the CIFAR-100 dataset that we searched on, we it to multiple datasets (i.e., Tiny ImageNet [1], Flowers [44], and Chaoyang [74]). We train DeiT|AutoFormer|PVT|Swin" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "text", + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 301, + 254 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 293, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 293, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 293, + 137 + ], + "type": "text", + "content": "Table 4: Results of AttnZero for DeiT & PVT on ImageNet " + }, + { + "bbox": [ + 132, + 114, + 293, + 137 + ], + "type": "inline_equation", + "content": "(224\\times 224)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 301, + 254 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 301, + 254 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 301, + 254 + ], + "type": "table", + "html": "
MethodParamsFLOPsTop-1 (%)
DeiT-T [58]5.7M1.2G72.2
DeiT-T+AttnZero5.7M1.0G74.9 (+2.7)
PVT-T [60]13.2M1.9G75.1
PVT-T+AttnZero10.9M1.8G78.1 (+3.0)
PVT-S24.5M3.8G79.8
PVT-S+AttnZero19.8M3.6G81.4 (+1.6)
PVT-M44.2M6.7G81.2
PVT-M+AttnZero34.6M6.6G83.0 (+1.8)
PVT-L61.4M9.8G81.7
PVT-L+AttnZero47.1M9.7G83.3 (+1.6)
", + "image_path": "feaccb8270bc685f4c85bd07d616372e7d4867cfffa599388e100038db3faac7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 312, + 146, + 479, + 254 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 486, + 137 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 486, + 137 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 486, + 137 + ], + "type": "text", + "content": "Table 5: Results of AttnZero on ImageNet with " + }, + { + "bbox": [ + 309, + 114, + 486, + 137 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 309, + 114, + 486, + 137 + ], + "type": "text", + "content": " resolution inputs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 146, + 479, + 254 + ], + "lines": [ + { + "bbox": [ + 312, + 146, + 479, + 254 + ], + "spans": [ + { + "bbox": [ + 312, + 146, + 479, + 254 + ], + "type": "table", + "html": "
MethodParamsFLOPsTop-1 (%)
AutoFormer-T [58]5.7M1.3G74.7
AutoF.-T+AttnZero5.7M1.1G75.2 (+0.5)
Swin-T [41]29.0M4.5G81.3
Swin-T+AttnZero28.2M4.3G82.1 (+0.8)
Swin-S50.0M8.7G83.0
Swin-S+AttnZero49.5M8.5G83.2 (+0.2)
Swin-B88.0M15.4G83.5
Swin-B+AttnZero89.0M15.3G83.7 (+0.2)
CSwin-T [17]23.0M4.3G82.7
CSwin-T+AttnZero20.3M4.0G82.9 (+0.2)
", + "image_path": "bc91a5f40fc87c648167841e14aaae841381cd858b2af305b2c25034d986b604.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 170, + 285, + 441, + 376 + ], + "blocks": [ + { + "bbox": [ + 153, + 264, + 459, + 276 + ], + "lines": [ + { + "bbox": [ + 153, + 264, + 459, + 276 + ], + "spans": [ + { + "bbox": [ + 153, + 264, + 459, + 276 + ], + "type": "text", + "content": "Table 6: Comparison of different efficient attention methods on ImageNet." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 285, + 441, + 376 + ], + "lines": [ + { + "bbox": [ + 170, + 285, + 441, + 376 + ], + "spans": [ + { + "bbox": [ + 170, + 285, + 441, + 376 + ], + "type": "table", + "html": "
MethodResoParamsFLOPsTop-1 (%)
DeiT-T [58]22425.7M1.2G72.2
DeiT-T+Hydra Attn [3]22425.7M1.1G68.3
DeiT-T+Efficient Attn [54]22425.7M1.1G70.2
DeiT-T+Linear Angular [66]22425.7M1.1G70.8
DeiT-T+FLatten22426.1M1.1G74.1
DeiT-T+AttnZero22425.7M1.0G74.9 (+2.7)
", + "image_path": "738b272af250039f96d0a86717eeb5a17863e8a3c790b3d4e5115deb4eebec9b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 399, + 481, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 434 + ], + "type": "text", + "content": "using discovered attention from scratch. The training process follows standard settings [31], including 300 training epochs, a cosine learning rate scheduler, and the AdamW optimizer. Detailed settings are shown in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "text", + "content": "Results on Various Datasets. Table 3 highlights the effectiveness of AttnZero across various ViT models and datasets. It consistently surpasses the baselines, resulting in a " + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "inline_equation", + "content": "6.29\\%" + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "text", + "content": " increase in accuracy for the DeiT-T model on Tiny-ImageNet. Similarly, AutoFormer and PVT-T models with AttnZero exhibit similar improvements. On the Flowers dataset, AttnZero achieves an accuracy of " + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "inline_equation", + "content": "57.89\\%" + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "text", + "content": " for the DeiT-T model, surpassing the baseline by " + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "inline_equation", + "content": "7.83\\%" + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "text", + "content": ". AutoFormer and PVT-T models also benefit from AttnZero. The results on the Chaoyang dataset further exemplify AttnZero's capability to enhance accuracies. On the DeiT-T model, AttnZero achieves an accuracy of " + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "inline_equation", + "content": "83.12\\%" + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "text", + "content": ", compared to the " + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "inline_equation", + "content": "82\\%" + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "text", + "content": " baseline. Moreover, the attention modules from AttnZero contribute to the improved performance of AutoFormer, PVT-T, and Swin-T models compared to their original versions. These findings demonstrate the promising ability of AttnZero across different datasets and models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 609, + 392, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 392, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 392, + 621 + ], + "type": "text", + "content": "5.2 Experimental Results on Large-scale Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Implementation. We conduct the experiment on the ImageNet [52] dataset, a widely used large-scale dataset consisting of 1000 classes and 1.2 million training images. Note that we do not perform additional searches or modify the exist-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 147, + 447, + 220 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 7: Results for Mask R-CNN on COCO dataset. We compute the FLOPs across the backbone, FPN, and detection head using an input resolution of " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "1280 \\times 800" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 147, + 447, + 220 + ], + "lines": [ + { + "bbox": [ + 164, + 147, + 447, + 220 + ], + "spans": [ + { + "bbox": [ + 164, + 147, + 447, + 220 + ], + "type": "table", + "html": "
MethodFLOPsSch.\\( AP^b \\)\\( AP_{50}^b \\)\\( AP_{75}^b \\)\\( AP^m \\)\\( AP_{50}^m \\)\\( AP_{75}^m \\)
PVT-T240G1x36.759.239.335.156.737.3
PVT-T+FLatten244G1x38.261.641.937.057.639.0
PVT-T+AttnZero209G1x40.062.643.137.459.440.1
PVT-S305G1x40.462.943.837.860.140.3
PVT-S+AttnZero248G1x41.063.445.038.260.441.1
", + "image_path": "2ae9fa38a4258ddaab24c505348c7b6d44ef7e010590f1e49c3c41793645129d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 307 + ], + "type": "text", + "content": "ing classical models. Instead, we directly replace the self-attention mechanism with our discovered attention module. Furthermore, we ensure that our training settings remain consistent with the standard settings [17, 41, 58]. This includes training the models for 300 epochs, cosine learning rate scheduler, and AdamW optimizer. Specific training settings are shown in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": "Results on Various Backbones. Table 4 and Table 5 provides the performance results of our AttnZero on ImageNet. The results show that our method can transfer effectively to different ViT models. When applied to DeiT-T, AttnZero achieves a significant " + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": " gain over baseline. Similarly, AttnZero improves PVT-T accuracy to " + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "inline_equation", + "content": "78.1\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": ", surpassing the baseline by " + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": ". This demonstrates that AttnZero can seamlessly migrate and consistently enhance performance across ViT architectures. AttnZero also provides stable gains across models of varying scales. Paired with PVT-S, it boosts accuracy by " + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": ". Applied to larger models like PVT-M, PVT-L, Swin-T, Swin-S, and CSwin-T, improvements of " + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 462 + ], + "type": "text", + "content": " respectively are observed. These findings highlight AttnZero's stability in boosting the performance of ViT models from small to large scale. These results highlight the stability of our AttnZero in enhancing the performance of ViT models across different scales." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 481, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 572 + ], + "type": "text", + "content": "Comparison with Other Linear Attention. Table 6 compares various linear attention methods. The results demonstrate that AttnZero surpasses other linear attention methods and baseline when incorporated into the DeiT-T model. Notably, compared to the Softmax baseline, AttnZero achieves a " + }, + { + "bbox": [ + 130, + 464, + 481, + 572 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 130, + 464, + 481, + 572 + ], + "type": "text", + "content": " improvement. In contrast, other linear attention methods, such as Hydra Attn, Efficient Attn, Linear Angular, and FLatten, exhibit lower accuracies. These findings highlight the superiority of AttnZero in preserving or surpassing the performance achieved by Softmax attention. On the other hand, other linear attention methods tend to result in notable performance degradation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 594, + 432, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 432, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 432, + 605 + ], + "type": "text", + "content": "5.3 Transfer Experiments on Detection and Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Object Detection. We evaluate our methods for object detection and instance segmentation tasks on the COCO dataset [39]. Table 7 provides results on Mask R-CNN detector [25] with " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " schedules. Our proposed PVT-T+AttnZero model achieves 2.3 box AP gains on PVT-T and significantly outperforms FLatten." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 448, + 102 + ], + "type": "text", + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 147, + 444, + 220 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 8: Results of semantic segmentation on ADE20K. We compute the FLOPs for both the encoders and decoders using input images at a resolution of " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "512 \\times 2048" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 147, + 444, + 220 + ], + "lines": [ + { + "bbox": [ + 167, + 147, + 444, + 220 + ], + "spans": [ + { + "bbox": [ + 167, + 147, + 444, + 220 + ], + "type": "table", + "html": "
BackboneMethodFLOPs#ParamsmIoUmAcc
PVT-TSemanticFPN158G17M36.5746.72
PVT-T+FLattenSemanticFPN169G16M37.2148.95
PVT-T+AttnZeroSemanticFPN126G15M39.1650.30
PVT-SSemanticFPN225G28M41.9553.02
PVT-S+AttnZeroSemanticFPN165G23M42.6254.02
", + "image_path": "2ecdc32caf798eeb16f08d5b9b724b31dd92e20e41d66e8aad068079d24e2908.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 243, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 479, + 266 + ], + "type": "text", + "content": "These enhancements demonstrate the benefits of incorporating our found attention module within the popular ViT model for object detection tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 266, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 480, + 361 + ], + "type": "text", + "content": "Semantic Segmentation. We assess the performance of our model using the ADE20K dataset [71]. Table 8 provides semantic segmentation results on SemanticFPN [29]. Among them, the Flatten module yields slight gains on the baseline PVT-T model. However, our AttnZero model achieves even higher mIoU (39.16) and mAcc (50.30) scores while reducing the FLOPs and parameters. Similarly, AttnZero achieves clear gains on PVT-S. These results highlight the compatibility of our AttnZero with different backbone architectures and its consistent improvements across various scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 377, + 218, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 218, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 218, + 391 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 400, + 481, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 400, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 132, + 400, + 481, + 591 + ], + "type": "text", + "content": "In this paper, we introduce AttnZero, the first method that automates the discovery of efficient attention modules that can be applied to various ViT models. AttnZero constructs a comprehensive search space consisting of different unary and binary operators as potential attention mechanisms. It utilizes a multi-objective genetic algorithm to search within this design space efficiently. Additionally, we introduce Attn-Bench-101, which provides attention profile results to summarize the insights gained from the search. Extensive experiments demonstrate that the attention modules discovered by AttnZero can be successfully applied to advanced ViTs for image classification, segmentation, and detection tasks. For future work, we will consider extending the AttnZero with knowledge distillation methods [32-35, 38, 53, 63]. Limitation. We present search strategies but still involve some search costs following most AutoML methods. But these costs are worthwhile because our discovered attention can be effectively applied to different ViTs and tasks without additional searching (proven by our extensive transfer experiments). Social impact. Our AttnZero focuses on technology improvement without social and ethical implications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 607, + 246, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 607, + 246, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 607, + 246, + 621 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 629, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 479, + 664 + ], + "type": "text", + "content": "The research was supported by Theme-based Research Scheme (T45-205/21-N) from Hong Kong RGC, and Generative AI Research and Development Centre from InnoHK." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 137, + 317, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 317, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 317, + 148 + ], + "type": "text", + "content": "1. http://tiny-imagenet.herokuapp.com/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 148, + 481, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 148, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 148, + 481, + 180 + ], + "type": "text", + "content": "2. Ali, A., Touvron, H., Caron, M., Bojanowski, P., Douze, M., Joulin, A., Laptev, I., Neverova, N., Synnaeve, G., Verbeek, J., et al.: Xcit: Cross-covariance image transformers. NIPS (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 481, + 201 + ], + "type": "text", + "content": "3. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Hoffman, J.: Hydra attention: Efficient attention with many heads. In: ECCVW (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 201, + 481, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 481, + 234 + ], + "type": "text", + "content": "4. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. arXiv preprint arXiv:2005.14165 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 234, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 481, + 255 + ], + "type": "text", + "content": "5. Chen, M., Peng, H., Fu, J., Ling, H.: Autoformer: Searching transformers for visual recognition. In: ICCV. pp. 12270-12280 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 255, + 481, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 481, + 288 + ], + "type": "text", + "content": "6. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. In: ICLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 288, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 320 + ], + "type": "text", + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. JMLR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 320, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 353 + ], + "type": "text", + "content": "8. Chu, X., Tian, Z., Zhang, B., Wang, X., Wei, X., Xia, H., Shen, C.: Conditional positional encodings for vision transformers. Arxiv preprint 2102.10882 (2021), https://arxiv.org/pdf/2102.10882.pdf" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "type": "text", + "content": "9. Chu, X., Zhang, B., Xu, R.: Fairnas: Rethinking evaluation fairness of weight sharing neural architecture search. In: ICCV (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 373, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 396 + ], + "type": "text", + "content": "0. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary computation (2002)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 396, + 481, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 427 + ], + "type": "text", + "content": "1. Dong, P., Li, L., Tang, Z., Liu, X., Pan, X., Wang, Q., Chu, X.: Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. In: ICML (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 427, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 481, + 449 + ], + "type": "text", + "content": "2. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "text", + "content": "3. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV. pp. 17076-17086 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "type": "text", + "content": "4. Dong, P., Niu, X., Li, L., Tian, Z., Wang, X., Wei, Z., Pan, H., Li, D.: Rd-nas: Enhancing one-shot supernet ranking ability via ranking distillation from zero-cost proxies. ICASSP (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 514, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 535 + ], + "type": "text", + "content": "5. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 535, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 481, + 567 + ], + "type": "text", + "content": "6. Dong, P., Niu, X., Tian, Z., Li, L., Wang, X., Wei, Z., Pan, H., Li, D.: Progressive meta-pooling learning for lightweight image classification model. In: ICASSP (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "text", + "content": "7. Dong, X., Bao, J., Chen, D., Zhang, W., Yu, N., Yuan, L., Chen, D., Guo, B.: Cswin transformer: A general vision transformer backbone with cross-shaped windows. ArXiv abs/2107.00652 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 621 + ], + "type": "text", + "content": "8. Dong, X., Yang, Y.: Nas-bench-201: Extending the scope of reproducible neural architecture search (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 101 + ], + "type": "text", + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "20. Gao, J., Xu, H., Shi, H., Ren, X., Philip, L., Liang, X., Jiang, X., Li, Z.: Autobert-zero: Evolving bert backbone from scratch. In: AAAI (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 192 + ], + "type": "text", + "content": "21. Guan, C., Wang, X., Zhu, W.: Autoattend: Automated attention representation search. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 3864-3874. PMLR (18-24 Jul 2021), http://proceedings.mlr.press/v139/guan21a.html" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 480, + 225 + ], + "type": "text", + "content": "22. Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. arXiv preprint arXiv:1904.00420 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 258 + ], + "type": "text", + "content": "23. Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. In: ECCV. vol. abs/1904.00420, pp. 544-560. Springer (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 258, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 480, + 279 + ], + "type": "text", + "content": "24. Han, D., Pan, X., Han, Y., Song, S., Huang, G.: Flatten transformer: Vision transformer using focused linear attention. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 279, + 454, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 454, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 454, + 289 + ], + "type": "text", + "content": "25. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: ICCV (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "text", + "content": "26. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "text", + "content": "27. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "text", + "content": "28. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: ICML (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "29. Kirillov, A., Girshick, R., He, K., Dollár, P.: Panoptic feature pyramid networks. In: CVPR (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "30. Kitaev, N., Kaiser, L., Levskaya, A.: Reformer: The efficient transformer. In: ICLR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "text", + "content": "31. Li, K., Yu, R., Wang, Z., Yuan, L., Song, G., Chen, J.: Locality guidance for improving vision transformers on tiny datasets. In: ECCV. pp. 110-127. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "type": "text", + "content": "32. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "text", + "content": "33. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "34. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "text", + "content": "35. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "type": "text", + "content": "36. Li, L., Sun, H., Dong, P., Wei, Z., Shao, S.: Auto-das: Automated proxy discovery for training-free distillation-aware architecture search. In: ECCV (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "text", + "content": "37. Li, L., Sun, H., Li, S., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Auto-gas: Automated proxy discovery for training-free generative architecture search. In: ECCV (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "text", + "content": "38. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "40. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "42. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. In: NeurIPS (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "43. Nakai, K., Matsubara, T., Uehara, K.: Att-darts: Differentiable neural architecture search for attention. In: IJCNN. IEEE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "text", + "content": "44. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 193, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 480, + 213 + ], + "type": "text", + "content": "45. Peng, H., Du, H., Yu, H., Li, Q., Liao, J., Fu, J.: Cream of the crop: Distilling prioritized paths for one-shot neural architecture search. NIPS (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 235 + ], + "type": "text", + "content": "46. Pham, H., Guan, M.Y., Zoph, B., Le, Q.V., Dean, J.: Efficient neural architecture search via parameter sharing. In: ICML (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 236, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 480, + 256 + ], + "type": "text", + "content": "47. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 258, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 258, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 480, + 277 + ], + "type": "text", + "content": "48. Raquel, C.R., Naval Jr, P.C.: An effective use of crowding distance in multiobjective particle swarm optimization. In: CGEC (2005)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 278, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 480, + 299 + ], + "type": "text", + "content": "49. Real, E., Aggarwal, A., Huang, Y., Le, Q.V.: Regularized evolution for image classifier architecture search. In: AAAI (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 300, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 480, + 319 + ], + "type": "text", + "content": "50. Real, E., Liang, C., So, D., Le, Q.: Automl-zero: Evolving machine learning algorithms from scratch. In: ICML (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 321, + 480, + 342 + ], + "type": "text", + "content": "51. Real, E., Liang, C., So, D., Le, Q.: Automl-zero: Evolving machine learning algorithms from scratch. In: ICML (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 343, + 480, + 374 + ], + "type": "text", + "content": "52. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Li, F.F.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 376, + 480, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 376, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 133, + 376, + 480, + 406 + ], + "type": "text", + "content": "53. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 407, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 480, + 427 + ], + "type": "text", + "content": "54. Shen, Z., Zhang, M., Zhao, H., Yi, S., Li, H.: Efficient attention: Attention with linear complexities. In: WACV (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "type": "text", + "content": "55. So, D.R., Manke, W., Liu, H., Dai, Z., Shazeer, N., Le, Q.V.: Primer: Searching for efficient transformers for language modeling. arXiv preprint arXiv:2109.08668 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 461, + 480, + 482 + ], + "type": "text", + "content": "56. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 483, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 480, + 503 + ], + "type": "text", + "content": "57. Tay, Y., Dehghani, M., Bahri, D., Metzler, D.: Efficient transformers: A survey. arXiv preprint arXiv:2009.06732 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 504, + 480, + 525 + ], + "type": "text", + "content": "58. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 525, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 480, + 567 + ], + "type": "text", + "content": "59. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. corr, abs/2302.13971, 2023. doi: 10.48550. arXiv preprint arXiv.2302.13971" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 480, + 600 + ], + "type": "text", + "content": "60. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 632 + ], + "type": "text", + "content": "61. Wang, X., Xiong, X., Neumann, M., Piergiovanni, A., Ryoo, M.S., Angelova, A., Kitani, K.M., Hua, W.: Attentionnas: Spatiotemporal attention cell search for video classification. In: ECCV (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "62. Wei, Z., Pan, H., Li, L.L., Lu, M., Niu, X., Dong, P., Li, D.: Convformer: Closing the gap between cnn and vision transformers. arXiv preprint arXiv:2209.07738 (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 101 + ], + "type": "text", + "content": "AttnZero: Efficient Attention Discovery for Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 533 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "63. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "64. Xiong, Y., Zeng, Z., Chakraborty, R., Tan, M., Fung, G., Li, Y., Singh, V.: Nystromformer: A nystrom-based algorithm for approximating self-attention. In: AAAI (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "65. Yang, A., Esperança, P.M., Carlucci, F.M.: Nas evaluation is frustratingly hard. arXiv preprint arXiv:1912.12522 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "66. You, H., Xiong, Y., Dai, X., Wu, B., Zhang, P., Fan, H., Vajda, P., Lin, Y.C.: Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "67. You, S., Huang, T., Yang, M., Wang, F., Qian, C., Zhang, C.: Greedynas: Towards fast one-shot nas with greedy supernet. In: CVPR (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "68. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "69. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token vit: Training vision transformers from scratch onImagenet. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "70. Zheng, Z., Gao, X., Pan, J., Luo, Q., Chen, G., Liu, D., Jiang, J.: Autoattention: Automatic field pair selection for attention in user behavior modeling. In: 2022 IEEE International Conference on Data Mining (ICDM). pp. 803-812. IEEE (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "71. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "72. Zhou, Q., Sheng, K., Zheng, X., Li, K., Sun, X., Tian, Y., Chen, J., Ji, R.: Training-free transformer architecture search. In: CVPR. pp. 10894-10903 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 401 + ], + "type": "text", + "content": "73. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 402, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 423 + ], + "type": "text", + "content": "74. Zhu, C., Chen, W., Peng, T., Wang, Y., Jin, M.: Hard sample aware noise robust learning for histopathology image classification. TMI" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "75. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "76. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "text", + "content": "77. Zoph, B., Le, Q.V.: Neural architecture search with reinforcement learning. In: ICLR (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "78. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Audio-Synchronized Visual Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_content_list.json b/2024/Audio-Synchronized Visual Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c2c0a6d47b2457ad51b6914e640940f16a1da149 --- /dev/null +++ b/2024/Audio-Synchronized Visual Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_content_list.json @@ -0,0 +1,1832 @@ +[ + { + "type": "text", + "text": "Audio-Synchronized Visual Animation", + "text_level": 1, + "bbox": [ + 282, + 141, + 722, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lin Zhang $^{1}$ , Shentong Mo $^{2}$ , Yijing Zhang $^{1}$ , and Pedro Morgado $^{1}$", + "bbox": [ + 243, + 188, + 756, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Wisconsin-Madison", + "bbox": [ + 383, + 215, + 619, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Carnegie Mellon University", + "bbox": [ + 401, + 229, + 599, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/43af284ff5c1aba269647c60332aaf7fff45c9bfef3c0359af50f511ff04f78e.jpg", + "image_caption": [ + "Fig. 1: Given an audio and an image (green box), we produce animations beyond image stylization with complex but natural dynamics, synchronized with audio at each frame. Results were produced by our AVSyncD model trained on the proposed AVSync15. Project webpage: https://lzhangbj.github.io/projects/asma/asva/asva.html." + ], + "image_footnote": [], + "bbox": [ + 217, + 262, + 785, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current visual generation methods can produce high-quality videos guided by text prompts. However, effectively controlling object dynamics remains a challenge. This work explores audio as a cue to generate temporally synchronized image animations. We introduce Audio-Synchronized Visual Animation (ASVA), a task that aims to animate a static image of an object with motions temporally guided by audio clips. To this end, we present AVSync15, a dataset curated from VGGSound with videos featuring synchronized audio-visual events across 15 categories. We also present a diffusion model, AVSyncD, capable of generating audio-guided animations. Extensive evaluations validate AVSync15 as a reliable benchmark for synchronized generation and demonstrate our model's superior performance. We further explore AVSyncD's potential in a variety of audio-synchronized generation tasks, from generating full videos without a base image to controlling object motions with various sounds. We hope our established benchmark can open new avenues for controllable visual generation.", + "bbox": [ + 259, + 511, + 740, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Audio $\\cdot$ Video $\\cdot$ Synchronization $\\cdot$ Generation", + "bbox": [ + 261, + 744, + 653, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 781, + 375, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative modeling has witnessed remarkable progress recently, largely due to the emergence of diffusion models [17,34,38]. Conditional generation and, in par", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ticular, text-to-image generation [33,34], has been the focal point given its application potential and availability of high-quality datasets [1,36]. This success has also led to revived interest in video generation, such as text-to-video [5,21,37,45]. While text guidance has been thoroughly investigated, the unique advantages of audio-visual synchrony for video generation remain underexplored. Unlike text, which provides control over global semantics, audio offers both semantic control on videos and precise control at each moment in time.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing works on audio-to-visual generation are however either limited to semantic control [20, 22, 39, 40], or constrained on singular scenarios such as human talking face [9-11, 28, 30, 49, 51, 52]. The former focuses on ambient audio datasets lacking synchronization cues with environmental sounds (ranining, fire crakcling, wind) [22] where shifting the audio temporally does not lead to major changes in visual content. The input audio in such cases thus substitutes text to provide only global semantics. The latter, talking faces, while synchronized, are extremely limited in generation diversity and control.", + "bbox": [ + 212, + 257, + 787, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, we introduce Audio-Synchronized Visual Animation, ASVA, a task which aims to animate objects depicted in natural static images into a video, with clear motion dynamics that are semantically aligned and temporally synchronized with the input audio. ASVA requires a sophisticated understanding of the audio's temporal structure, as well as of how objects move in synchrony with sound. Prior attempts on visual generation guided by diverse sounds fall short in generated visual quality [23] and accurate synchronization control [20,48] due to two challenges: (1) the lack of high-quality training data and benchmarks for learning audio-synchronized visual dynamics; (2) the development of effective methods capable of generating highly synchronized video motions. Successfully addressing these challenges will expand the scope of current video generation methods to enable more fine-grained semantic and temporal control over the generation process via synchronized audio conditioning.", + "bbox": [ + 212, + 382, + 787, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We address the first challenge by constructing a high-quality diverse dataset with strong correlations between audio and object motions at each moment in time. In an ideal dataset, sound sources should be easily identifiable in the scene. Every visual motion in the video should highly correlate with the audio semantically and temporally, and vice versa. Moreover, the visual content should be of high quality for generation. However, existing audio-visual datasets either are too noisy [7, 13], containing a large number of unassociated audio-visual pairs [31], or predominantly featuring ambient sound categories that lack meaningful synchronized object dynamics [22, 41]. We thus curate a high-quality dataset from VGGSound [7] by deploying an efficient two-step curation pipeline. In the first step, we use a variety of signal processing techniques and foundation models to automatically filter out videos with poor semantic alignment or temporal synchronization, as well as those depicting static scenes or with too fast camera motions. Then, to ensure the highest possible quality as a benchmark, we narrow down the dataset to sound categories with strong audio-visual synchronization cues, and manually verify the quality of each video. We end up obtaining AVSync15 with 15 dynamic sound classes, each with 100 examples rich", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Lin Zhang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and accurate in semantics, object dynamics, and audio-visual synchronization. An overview of AVSync15 classes is in Figs. 2a and 2b and its comparison with prior audio-video generation datasets is in Suppl. Sec. 5.1.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second challenge pertains to the generation of audio-synchronized motions, which requires a detailed understanding of audio-visual correlations and object dynamics. Take, for instance, a video featuring a dog. To generate realistic video, the model is expected to not only synchronize the dog's mouth with the barking sound, but also accurately depict subtleties in the dog's head pose before and after barking. Furthermore, in a more challenging scenario, the model should discern which object to animate to preserve semantic consistency depending on the input sound. However, existing audio-conditioned visual generation frameworks [14, 39] primarily focus on semantic control, often encoding the audio into a single global semantic feature and thereby neglecting the audio's temporal domain. Even recent attempts at audio-synchronized video generation [20, 23, 48] have not fully realized the potential of audio for fine-grained temporal control, as they either rely on crude audio representations such as audio amplitude [23], learn from weakly synchronized [22], noisy datasets [7], or ignore object dynamics in the generation process [20, 22, 23, 48]. To this end, we introduce Audio-Video Synchronized Diffusion (AVSyncD), a framework improving a pre-trained image latent diffusion model [34] for enhanced audio guidance and motion generation. We employ the pre-trained ImageBind [14] encoder to encode audio into time-aware semantic tokens, then fuse them into each frame's latent features. This allows for precise audio guidance on video semantics and synchronization. To capture complex video motions, we add temporal attention layers to the diffusion model. Finally, to ensure faithful animation of the input image, we incorporate temporal convolutions and attention layers that always reference the input image, i.e., first-frame lookups.", + "bbox": [ + 212, + 191, + 787, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the carefully designed dataset and architecture, we are able to train a model specialized for ASVA and produce animations with more realistic and audio-synchronized contents than prior works (Figs. 1 and 4). We provide thorough experiments to validate the effectiveness of AVSync15 and AVSyncD and demonstrate how to deploy AVSyncD for controllable generation, including amplifying audio guidance and semantic-aware object animation (Secs. 5.3 and 5.4).", + "bbox": [ + 212, + 554, + 787, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 665, + 387, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Controllable Visual Generation", + "text_level": 1, + "bbox": [ + 214, + 695, + 519, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many conditional visual generation models based on diffusion process [17, 38] have emerged recently. Benefiting from more efficient architecture, large-scale training data [36], and aligned semantic space [32], Latent Diffusion Model (LDM) [34] has achieved great success to generate realistic images conditioned on text. This inspired researchers to explore various diffusion-based visual generation tasks, such as text-to-video [2, 5, 21, 44-46], audio-to-image [14, 39], and audio-to-video [20, 23, 40, 48]. The architectures can be training-free [21, 23, 45], fully-trained [5, 44], or trained partially, which augments a pre-trained LDM by", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Audio-Synchronized Visual Animation", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "carefully adding some trainable layers [2, 20, 46]. Extensive works have also attempted to control the semantics of the generated content [15, 25, 27, 50], while how to apply control in the temporal dimension remains under-explored.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we developed an image animation model AVSyncD to control generation semantically and temporally guided by audio. AVSyncD augments pre-trained StableDiffusion [34] with trainable temporal layers and audio conditioning mechanism, preserving training efficiency and generalizing well.", + "bbox": [ + 212, + 191, + 784, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Audio-to-Video Generation", + "text_level": 1, + "bbox": [ + 214, + 275, + 486, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditionally, audio has been used as a temporal cue for talking face generation [28, 30, 49, 51, 52], where face and lip actions should be synchronized with audio at each frame. Many works also rely on complex inputs such as 3D meshes and human poses [28, 49]. Although accurately synchronized, this line of research is extremely limited in scenarios and cannot generate videos for diverse audios.", + "bbox": [ + 212, + 300, + 782, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A series of works attempted to expand the class of audio by encoding sound into a global semantic condition for video generation [14, 39, 40], however often overlooked the temporal aspect inherent in audio. Some recent works, although divided audio features into time-aware segments as inputs [22, 35], failed to achieve promising visual quality or synchronize video motions with audios. AADiff [23] is a training-free method re-weighting the text-image cross-attention map in LDM using audio amplitude at each frame, however can only control styles of each frame. TPoS [20] learns segmented audio features aligned with CLIP [32] using sophisticated modules and training losses, and feeds them into a pre-trained text-to-image model [34] for video generation. TempoToken [48] also learns segmented audio features with a pre-trained audio encoder BEATs [8], and fuses them into a pre-trained text-to-video model [44]. However, primarily focused on monotonous sound classes in Landscapes [22] or noisy audio-visual data in VGGSound [7], these methods are limited to generating video semantics without capturing the natural and synchronized dynamics of video content. Frozen generation architectures also prevent them from generating natural motions.", + "bbox": [ + 212, + 376, + 784, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these limitations, we introduce AVSync15, a high-quality dataset specifically designed for ASVA. AVSync15 stands out from previous efforts by focusing on synchronization cues between audio and visual dynamics, allowing for generating motions beyond mere visual effects. Once trained on AVSync15, our AVSyncD can generalize to many applications to control video motions guided by audios, on which previous methods performed poorly.", + "bbox": [ + 212, + 618, + 784, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Audio-Synchronized Visual Animation", + "text_level": 1, + "bbox": [ + 214, + 732, + 624, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, the Audio-Synchronized Visual Animation (ASVA) task can be posed as follows. Given an audio clip $\\pmb{a}$ of length $T$ seconds and an image $\\pmb{x}_1$ , the goal is to generate the future video sequence $\\pmb{x}_2, \\dots, \\pmb{x}_{rT}$ (or $\\pmb{x}_{2:rT}$ for short), where $r$ is the desired frame rate. Despite the simple formulation, this is a challenging task as the generated video sequence should (1) be of high visual quality, (2)", + "bbox": [ + 212, + 763, + 784, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Lin Zhang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/835fa2fbaee3831d2b2532d55dd24815728e1f2e8b722dddb39ac274fe145976.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 442, + 248 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bb5dc3b06a9c08f25dae32a92972f5bd44df3ba2811b6f334d36e4c3a11217b6.jpg", + "image_caption": [ + "(b)", + "Fig. 2: (a): Overview of 15 categories in AVSync15. Categories are listed below x-axis on the right plot. (b): Category-wise averages of av-sync score $\\phi$ , IA, and IT on AVSync15 and equivalently sized subsets of VGGSS and AVSync-AC. Error bars for VGGSS and AVSync-AC are obtained from 3 random splits." + ], + "image_footnote": [], + "bbox": [ + 457, + 146, + 782, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "be semantically aligned with the image $\\pmb{x}_1$ and audio $\\pmb{a}$ , (3) exhibit temporal coherence and (4) natural object motions temporally synchronized with the audio $\\pmb{a}$ . To facilitate research in ASVA, we introduce a new benchmark that includes a curated high-quality dataset and a suite of evaluation metrics designed to capture the various components of audio-synchronized generation.", + "bbox": [ + 212, + 356, + 787, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 AVSync15: A High-Quality Dataset for Audio Synchronized Video Generation", + "text_level": 1, + "bbox": [ + 212, + 452, + 759, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing large-scale audio-visual datasets like VGGSound [6,7,19] and AudioSet [1] often contain amateur videos from platforms like YouTube. These videos, while diverse, can pose challenges for audio-synchronized video generation tasks due to rapid scene changes, camera motion, noisy audio, or out-of-frame sound sources. Prior work [22, 24] has addressed this by focusing on simpler videos, such as those depicting fire crackling or weather patterns. However, such videos often lack strong synchronization cues between audio and visual motion, making them unsuitable for ASVA. To facilitate research in this area, we introduce a high-quality dataset specifically designed for audio-synchronized video generation, ensuring a close synchronization between audio and visuals. More specifically, our selection criteria to create the dataset were: (1) High Correlation: Significant visual changes should be closely associated with audio at each timestamp, and vice versa. (2) Dynamic Content: We included content rich in temporal changes, excluding ambient or monotonous classes. (3) Quality and Relevance: Both video and audio needed to be clean, stable, and semantically aligned.", + "bbox": [ + 212, + 489, + 787, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preliminary Curation We start from VGGSound [7], a large-scale dataset with 309 diverse audio classes. Similar to VGGSoundSync [6], we first narrow down to 149 classes with potentially clear audio-visual synchronized events, removing ambient classes, which is referred as VGGSS. We then deploy a sequence of automatic cleaning steps and a final manual selection step to identify appropriate videos. The procedures are summarized below and detailed in Suppl. Sec. 5.2.", + "bbox": [ + 212, + 717, + 787, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Automatic Curation We first use PySceneDetect [4] to cut videos with sharp scene changes into different scenes, which are still likely to contain low-quality", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Audio-Synchronized Visual Animation", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "short clips. To maximize usage, we split each scene into 3-second clips with 0.5-second strides, and discard unsuitable clips based on the following metrics:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Raw Pixel Difference: We calculate average pixel distances between consecutive frames and remove clips with both small and large values, likely depicting either static or videos with excessive motion.", + "bbox": [ + 212, + 189, + 784, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image Semantics Difference: Complementing the raw pixel analysis above, we also compute distances on CLIP [32] image features, removing videos with small semantics changes such as zoom in/out or large semantic content transitions.", + "bbox": [ + 212, + 234, + 782, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audio Waveform Amplitude: We exclude clips whose maximum audio waveform amplitude is low, indicating weak audio cues.", + "bbox": [ + 214, + 281, + 782, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic Alignment: We compute the average image-audio (IA) and image-text (IT) alignment scores [32] in a video using ImageBind [14], removing clips with low scores to ensure cross-modal semantic alignment.", + "bbox": [ + 212, + 311, + 782, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audio-Video Synchronization: To measure audio-visual synchronization, we follow VGGSoundSync [6] to contrastively train an audio-visual synchronization classifier on VGGSS, ending up with a comparable $40.85\\%$ test accuracy. The model outputs an unbounded av-sync score $\\phi_{\\pmb{a}_i,\\pmb{v}_j}$ for an audio-video pair $(\\pmb{a}_i,\\pmb{v}_j)$ . During training, we compute $\\phi$ for a synchronized pair $(\\pmb{a}_i,\\pmb{v}_i)$ and its temporally-shifted pairs from the same instance. Contrastive loss is then applied to these shifted pairs to maximize the synchronization probability:", + "bbox": [ + 212, + 358, + 784, + 463 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {S y n c}} \\left(\\boldsymbol {a} _ {i}, \\boldsymbol {v} _ {i}\\right) = \\frac {\\exp \\left(\\phi_ {\\boldsymbol {a} _ {i} , \\boldsymbol {v} _ {i}} / \\tau\\right)}{\\sum_ {j} \\exp \\left(\\phi_ {\\boldsymbol {a} _ {i} , \\boldsymbol {v} _ {j}} / \\tau\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 477, + 784, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to distinguish the synchronized pair from shifted ones. We use $P_{\\mathrm{sync}}$ as a synchronization indicator to remove low-scoring clips. When computing $P_{\\mathrm{sync}}$ , we discard the temperature parameter $\\tau$ used to improve training efficacy. We detail the synchronization classifier and $P_{\\mathrm{sync}}$ in Suppl. Sec. 1 and Sec. 2.1, respectively.", + "bbox": [ + 212, + 522, + 782, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We empirically determine metrics' thresholds by prioritizing quality, acknowledging that some acceptable clips might be discarded. After automatic curation, we merge all continuous 3-second clips from each video and remove categories with less than 100 examples to avoid class imbalance, resulting in AVSync-AC (AVSync w/ Automatic Curation) with 76 categories and 39,902 examples.", + "bbox": [ + 212, + 594, + 782, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Manual Curation We further select 15 diverse categories with clear audiovisual motion cues from AVSync-AC for manual refinement. The categories range from animals and human actions to triggered tools and musical instruments. Manual curation once again seeks to identify appropriate videos for ASVA with the criteria above: high correlation, dynamic content, quality and relevance.", + "bbox": [ + 212, + 671, + 782, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Summary The final dataset, AVSync15, contains 90 training and 10 testing videos per category, each $2\\sim 10$ seconds long. We provide an overview of AVSync15 in Fig. 2a. To validate our curation pipeline, we randomly sample three 1500-video splits on the selected 15 categories from VGGSS and AVSync-AC, and quantitatively compare them with AVSync15 in Fig. 2b and Tab. 2b. We also compare AVSync15 with other audio-visual datasets in Suppl. Sec. 5.1.", + "bbox": [ + 212, + 750, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Lin Zhang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 146, + 419, + 159 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ASVA is a multi-faceted generation task, necessitating high quality at both image and video level. At the image level, we follow previous conventions [3,12] to use (1) Fréchet Inception Distance (FID) [16] to measure the quality of individual frames; (2) IA [14]/IT [47] to measure image-audio/image-text semantics alignment on CLIP [32] space. At the video level, we use Fréchet Video Distance (FVD) [42] to assess video quality. To measure audio-video synchronization, we compute the following metrics with the trained synchronization classifier:", + "bbox": [ + 212, + 171, + 782, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RelSync During testing, we use the ground truth audio-visual pair $(\\pmb{a},\\pmb{v})$ as a reference to measure synchronization of the generated video $\\hat{\\pmb{v}}$ as follows:", + "bbox": [ + 212, + 277, + 782, + 308 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e l S y n c} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v}) = \\frac {\\exp \\left(\\phi_ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}}}\\right)}{\\exp \\left(\\phi_ {\\boldsymbol {a} , \\boldsymbol {v}}\\right) + \\exp \\left(\\phi_ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}}}\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 319, + 784, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that while this reference-based metric normalizes the score by the synchronization of the reference pair $(\\pmb{a},\\pmb{v})$ , the metric can still be sensitive to the quality of the reference pair. In fact, evaluating synchronization on a dataset where even ground-truth audios and videos are ambiguously synchronized is less informative of the model capabilities, e.g., Landscapes [22].", + "bbox": [ + 212, + 361, + 782, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AlignSync The synchronization classifier is only trained on semantically-aligned and temporally-shifted audio-video pairs sampled from the same instance (See Suppl. Sec. 1.2). RelSync is thus implicitly conditioned on semantics alignment, i.e., P(Sync|Align). To jointly measure semantics alignment and synchronization, we first approximate $P_{\\mathrm{Align}}$ similarly as RelSync:", + "bbox": [ + 212, + 436, + 782, + 513 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {A l i g n}} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v} _ {1}) = \\frac {1}{b - 1} \\sum_ {i = 2.. b} \\frac {\\exp \\left(\\mathrm {I A} _ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}} _ {i}}\\right)}{\\exp \\left(\\mathrm {I A} _ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}} _ {i}}\\right) + \\exp \\left(\\mathrm {I A} _ {\\boldsymbol {a} , \\boldsymbol {v} _ {1}}\\right)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 525, + 784, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{v}_1$ is the first frame input for animation, and $b$ is the number of generated frames. The generated first frame $\\hat{\\pmb{v}}_1$ is eliminated because it is often a replicate of input $\\pmb{v}_1$ . By multiplying $P_{\\mathrm{Align}}$ with RelSync, we obtain the joint score:", + "bbox": [ + 212, + 571, + 782, + 616 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A l i g n S y n c} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v}) = P _ {\\text {A l i g n}} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v} _ {1}) \\cdot \\operatorname {R e l S y n c} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 628, + 784, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which jointly measures semantic alignment and temporal synchronization between $\\pmb{a}$ and $\\hat{\\pmb{v}}$ .", + "bbox": [ + 212, + 655, + 782, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These automated metrics are not always aligned with human preference. We therefore conduct a user study detailed in Sec. 5.2, asking human raters to compare videos generated by multiple models and select the best one.", + "bbox": [ + 212, + 686, + 782, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Audio-Video Synchronized Diffusion", + "text_level": 1, + "bbox": [ + 215, + 753, + 602, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Preliminary: Text-to-Image Latent Diffusion", + "text_level": 1, + "bbox": [ + 215, + 785, + 627, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-to-image latent diffusion models (LDMs [34]) encode images $\\pmb{x}$ into a lower-dimensional latent space $\\pmb{z} = \\mathcal{E}(\\pmb{x})$ using a pre-trained perceptual auto-encoder", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Audio-Synchronized Visual Animation", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathcal{E}(\\cdot)$ , and learn the conditional distribution $p(\\boldsymbol{z}|\\boldsymbol{\\tau})$ given a CLIP-encoded text prompt $\\boldsymbol{\\tau}$ . It gradually denoises latents $\\boldsymbol{z}^k$ , obtained by corrupting the image latent $\\boldsymbol{z}$ with Gaussian noise $\\epsilon$ , over $k$ time steps. A denoising UNet $\\epsilon_{\\theta}(\\cdot)$ parameterized by $\\theta$ is deployed to estimate the added noise $\\epsilon$ by minimizing", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L D M} = \\mathbb {E} _ {\\boldsymbol {z}, \\epsilon \\sim \\mathcal {N} (0, 1), k} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {z} ^ {k}, k, \\boldsymbol {\\tau}) \\| _ {2} ^ {2} \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 218, + 785, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During inference, LDMs start from a random Gaussian noise map $\\pmb{z}^{K}$ , and iterate over $K$ reverse diffusion steps [17, 38] to denoise the latents $\\pmb{z}^{k-1} = \\pmb{z}^{k} - \\epsilon_{\\theta}(\\pmb{z}^{k}, k, \\pmb{\\tau})$ , until $\\pmb{z}^{0}$ is obtained. LDMs then decode the latent into an image $\\pmb{x}^{0} = \\mathcal{D}(\\pmb{z}^{0})$ using the pre-trained decoder $\\mathcal{D}(\\cdot)$ . For simplicity, we refer to the images by their latents $\\pmb{z}$ rather than $\\pmb{x}$ throughout the rest of the paper.", + "bbox": [ + 214, + 244, + 787, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Proposed Architecture", + "text_level": 1, + "bbox": [ + 214, + 343, + 452, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we extend the capabilities of LDMs for ASVA, by focusing on learning video dynamics and temporal synchronization, unlike existing approaches [14, 39, 40] which primarily use audio to control global semantics. We propose the Audio-Video Synchronized Diffusion model (AVSyncD), which builds upon a pre-trained image LDM and integrates synchronized audio control and temporal layers for improved video consistency. The major component is a UNet $\\epsilon_{\\theta}$ aiming to denoise a video instead of an image. The UNet is trained on synchronized audio-video pairs, with the first frame $z_{1}$ , the corresponding audio $\\pmb{a}$ , and the CLIP encoded audio category name $\\pmb{\\tau}$ as input conditions. The LDM denoising objective in Eq. (5) is applied to the remaining frames $z_{2:rT}$ to be predicted. The architecture of AVSyncD is shown in Fig. 3 and discussed below. Further details, for example, highlighting the different attention modules used in the architecture are described in Suppl. Sec. 3.2.", + "bbox": [ + 212, + 367, + 787, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initial-frame Conditioning To condition LDM on an input image, we feed its latent $\\mathbf{z}_1$ without noise into the UNet at every diffusion timestep $k$ . For all subsequent frames, we adhere to the original LDM, using independently sampled noisy latents $\\mathbf{z}_{2:rT}^k$ as inputs and predicting the added noise $\\epsilon^k$ .", + "bbox": [ + 212, + 565, + 787, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Conditioning We retain the text cross-attention layers in the original LDMs [34] (without finetuning) to condition the model on the audio category. However, due to the limited text diversity in training, class conditioning does not bring significant gains (see Suppl. Sec. 6.5).", + "bbox": [ + 212, + 627, + 787, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Audio Conditioning To facilitate audio-synchronized generation, we condition the generation on ImageBind audio embeddings [14]. ImageBind computes an audio classification token, $\\pmb{a}^g$ , representing global semantics, and local patch tokens, $\\pmb{a}_{f,t}$ , across $T_{a}$ timestamps. The original ImageBind only uses $\\pmb{a}^g$ for contrastive learning and discards $\\pmb{a}_{f,t}$ . However, we found these frozen local tokens as efficient synchronization cues. We split the patch tokens temporally into $rT$ segments, corresponding to the same timestamps as the frames $z_{1:rT}$ and append the global token to each. Each frame $z_{t}$ then learns both semantics and synchronization guidance from its audio segment $\\pmb{a}_{t}$ via cross-attention [43]. In Suppl. Sec. 6.7, we compared ImageBind with CLAP and BEATs as encoders.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Lin Zhang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1bd66f98f8d4eae3fefeaba02b910f21dd921c8882d448d582879daf4a8ae176.jpg", + "image_caption": [ + "Fig. 3: AVSyncD overview. Left: We use ImageBind to encode audio into semantically aware time-dependent tokens $(\\pmb{a}_t)_{t=1}^{rT}$ and CLIP to encode the audio category into text embedding $\\pmb{\\tau}$ . In addition, the model receives the latent of the first frame $\\pmb{z}_1$ , and iteratively denoises noisy latents of the subsequent frames $\\pmb{z}_{2:rT}^k$ via reverse diffusion. The denoising UNet, based on LDMs [34], consists of a sequence of downsampling, bottleneck and up-sampling blocks, with structure detailed on the right. Right: Anatomy of a UNet block for frame $\\pmb{z}_t$ . LDM's original spatial conv, spatial attention and text cross attention layers are frozen, while its spatial self-attention layers are adjusted to first-frame spatial attentions, cross-attending to $\\pmb{z}_1$ instead. To learn video dynamics, we introduce temporal attention layers, and first-frame lookup temporal convolutions applied to input, output, and ResNet layers. We also train audio cross attentions for audio conditioning and synchronization. Trainable layers are marked with $\\oplus$ ." + ], + "image_footnote": [], + "bbox": [ + 261, + 146, + 743, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Spatial Convolution The original LDM's pre-trained spatial convolutional layers were frozen and used without modification.", + "bbox": [ + 212, + 501, + 784, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First-frame Temporal Convolution Each spatial convolution block was augmented with a 1D temporal convolution layer (kernel size 3) to capture temporal dependencies. To better adhere to the starting image, $\\mathbf{z}_1$ , we adjusted the receptive field at frame $t$ to include frames $(1, t - 1, t)$ as opposed to $(t - 1, t, t + 1)$ . These temporal convolutions with first-frame lookup were applied to three components in the UNet, namely the input/output conv layers and all ResNet convs.", + "bbox": [ + 212, + 532, + 787, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First-frame Spatial Attention We also leveraged the pre-trained LDM's spatial self-attention. Following [21], we modified the frozen spatial attention layers to cross-attend to the first frame rather than self-attend to the current frame, by computing the key-value pairs from the first frame and the queries from the current one.", + "bbox": [ + 212, + 625, + 787, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Temporal Attention To effectively model long-range visual dependencies, we incorporated temporal attention layers [2]. Each frame index $t$ was converted into a sinusoidal positional embedding, added to the corresponding frame's latents after a learnable projection [43]. Each frame's local representation at spatial position $(h,w)$ , $z_{hw}$ , was then updated by attending to all frames at the same position $(z_{hw1},z_{hw2},\\ldots ,z_{hw(rT)})$ through a standard attention mechanism.", + "bbox": [ + 212, + 702, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Classifier-free Audio Guidance Classifier-free guidance [18] is a technique used to control the impact of input prompts on the generated outputs. We extended it to amplify audio guidance for improved synchronization. To accomplish", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Audio-Synchronized Visual Animation", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "this, we trained the model for both audio-conditioned and unconditioned generation, by randomly replacing $\\mathbf{a}$ with a null audio embedding, $\\mathbf{a}_{\\emptyset}$ , with a 20% likelihood. $\\mathbf{a}_{\\emptyset}$ was computed by encoding an all-zero audio mel spectrogram via ImageBind. During inference, a factor $\\eta \\geq 1$ controls the audio guidance:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {2: r T} ^ {k - 1} = (1 - \\eta) \\cdot \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {2: r T} ^ {k}, k; \\boldsymbol {z} _ {1}, \\boldsymbol {a} _ {\\emptyset}, \\boldsymbol {\\tau}\\right) + \\eta \\cdot \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {2: r T} ^ {k}, k; \\boldsymbol {z} _ {1}, \\boldsymbol {a}, \\boldsymbol {\\tau}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 217, + 785, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\eta$ guides the denoising process towards latents congruent with audio-conditioned generation and away from those of audio-unconditional generation.", + "bbox": [ + 212, + 244, + 785, + 275 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 297, + 375, + 313 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Implementation", + "text_level": 1, + "bbox": [ + 215, + 327, + 392, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset We conducted experiments on three datasets. AVSync15: Our high-quality dataset curated from VGGSound [7], with 15 balanced categories, 1350 training videos, and 150 test videos. We also assessed our data curation pipeline by comparing it to models trained on Landscapes, TheGreatestHits, VGGSS and our AVSync-AC. Landscapes [22] is composed of 9 environmental sound classes. We followed the split in [35,48] with 900 clips for training and 100 for testing. Since Landscapes is full of ambient sounds without synchronized video motion, we mainly use it to evaluate visual quality. TheGreatestHits [29] is an audio-video dataset recording humans probing environments with a drumstick, with 733 videos for training and 244 for testing. The videos are synchronized with audio at the moments of impact but contain lots of static moments. It also offers limited diversity, featuring a singular motion of probing. VGGSS and our AVSync-AC were described in Sec. 3.", + "bbox": [ + 212, + 352, + 787, + 547 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data Preprocess We sampled 2-second synchronized audio-video pairs for experiments. Videos were sampled at 6 fps with 12 frames, and resized to $256 \\times 256$ on AVSync15/Landscapes or $128 \\times 256$ on TheGreatestHits. Following Image-Bind [14], audios were sampled at $16\\mathrm{kHz}$ and converted into 128-d spectrograms.", + "bbox": [ + 212, + 550, + 787, + 612 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines We first adopted a simple Static baseline by repeating the input frame into a video, then compared it to several state-of-the-art works. (1) Semantic audio-to-video generation (CoDi [40]): Image, text, and audio are encoded into a shared CLIP [32] space and summed, then fused into a video diffusion model trained on large-scale datasets. (2) Image-to-video generation (VideoCrafter [5]): A superior video diffusion model however without audio inputs. It animates images by fusing CLIP-encoded image and text features into the model via modality-dependent cross-attention layers. (3) Synchronized audio-to-video generation (TPoS [20], AADiff [23], TempoToken [48]): Audio is encoded into time-dependent segments and fused into frozen text-to-image [34] or text-to-video [44] models. We re-implemented AADiff, and used TPoS and TempoToken's pretrained checkpoints on VGGSound and Landscapes. More details are provided in Suppl. Sec. 4.", + "bbox": [ + 212, + 612, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training & Evaluation We adopted the pretrained Stable Diffusion-V1.5 [34] as the diffusion model and ImageBind [14] as the audio encoder. All models", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Lin Zhang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e8892f0690f70eb192e769bf36aa8c4a75a15945eab56609dfabd42aca9fd297.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 143, + 787, + 330 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/058c123d73accdbcd10bf911170d2915a113bd94be723b6cb8d3c6723a3fbc9c.jpg", + "image_caption": [ + "(a) Comparison on AVSync15. AVSyncD produces synchronized dynamic motions, such as muzzle flash, chicken and dog's head poses, and swirling in toilet.", + "(b) Comparison on Landscapes. Without synchronization (c) Generation by AVSyncD on The-cues in audio, AVSyncD still produces evolving visual ef- GreatestHits. AVSyncD generates hitting fecs and semantics, such as flames, waves, and lightning. actions at appropriate moments.", + "Fig. 4: Qualitative results on three datasets." + ], + "image_footnote": [], + "bbox": [ + 218, + 358, + 787, + 482 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "were trained using Adam optimizer with a batch size of 64 and a learning rate of 0.0001. Besides metrics in Sec. 3.2, we also provide results using metric in [26] in Suppl. Sec. 3.4. We evaluated on 3 clips uniformly sampled from each video.", + "bbox": [ + 212, + 569, + 787, + 614 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 633, + 372, + 647 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset Comparison In Tab. 1, the Static baseline is a good indicator of dataset attributes. It has similar IA/IT scores compared to Groundtruth, since Static is composed of a subset of groundtruth frames, which are obviously semantically aligned. FVD of Static on TheGreatestHits is relatively low, since the TheGreatestHits contains frequent static ground truth clips without any moving objects. More importantly, the RelSync and AlignSync of Static gradually increase from AVSync15, to TheGreatestHits, to Landscapes, with those on Landscapes even surpassing the ground truth. The fact that static videos perform well in terms of audio synchronization on TheGreatestHits and Landscapes datasets testify to the superiority of AVSync15 for audio-video synchronized generation. Model Comparison We compared to prior works on AVSync15 and Landscapes in Tabs. 1a and 1b and Figs. 4a and 4b. In Tab. 1a, CoDi achieves inferior", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Audio-Synchronized Visual Animation", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2549954ef17cff95afaeb680e73b54831e2fdcc177810270a119beaf4c856a75.jpg", + "table_caption": [ + "Table 1: Quantitative results. User study shows votes on 3 metrics: image quality, frame consistency, and synchronization. Inputs are combinations of image, text, audio." + ], + "table_footnote": [], + "table_body": "
InputModelFID↓IA↑IT↑FVD↓AlignSync↑RelSync↑User Study
IQ↑FC↑Sync↑
T+ATPoS [20]13.523.3824.832671.019.5242.50---
TempoToken [48]12.218.8417.454466.419.7444.05---
I+TVideoCrafter [5]11.8-29.87840.721.2843.16382012
I2VD12.1-30.35398.221.8043.92629091
I+T+ACoDi [40]14.528.1523.421522.619.5441.51---
TPoS [20]11.938.3630.731227.819.6739.62---
AADiff [23]18.834.2328.97978.022.1145.483745
AVSyncD η = 112.138.3630.34382.722.2544.81---
AVSyncD η = 411.738.5330.45349.122.6245.52163186192
AVSyncD η = 811.737.9930.27420.722.7445.88---
Static-39.7630.391220.421.8343.66
Groundtruth-40.0630.31-25.0450.00
", + "bbox": [ + 220, + 184, + 785, + 368 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f28e1f39d3679bb3be5ca79af3806a72aa6a2a5e56c6927b453ab4ad7e974bd5.jpg", + "table_caption": [ + "(a) Performance on AVSync15." + ], + "table_footnote": [], + "table_body": "
InputModelFID↓IA↑IT↑FVD↓AignSync↑RelSync↑
T+ATPoS [20]16.515.6126.702081.323.1248.15
TempoToken [48]16.422.5822.872480.024.2148.65
I+TI2VD16.7-22.56539.524.7449.89
I+T+ACoDi [40]20.522.6324.23982.922.6345.48
TPoS [20]16.223.5223.20789.623.5147.05
AADiff [23]70.722.0722.921186.326.7753.93
AVSyncD η = 116.522.2922.81463.124.8149.96
AVSyncD η = 416.222.4922.79415.224.8249.93
Static-23.6022.211177.525.7951.59
Groundtruth-23.6522.08-25.0150.00
", + "bbox": [ + 220, + 395, + 785, + 539 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a41fb580d1678c1a68134eb885e2b6773368fadbd3d12be2e7e87af093cc1b25.jpg", + "table_caption": [ + "(b) Performance on Landscapes." + ], + "table_footnote": [], + "table_body": "
InputModelFID↓IA↑IT↑FVD↓AignSync↑RelSync↑
I+TI2VD9.1-13.42425.022.0544.58
I+T+AAVSyncD η = 19.011.8513.18313.522.5945.52
AVSyncD η = 48.712.0713.31249.322.8345.95
Static-13.3316.56348.924.3648.73
Groundtruth-13.5216.49-25.0250.00
", + "bbox": [ + 220, + 565, + 785, + 652 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(c) Performance on TheGreatestHits.", + "bbox": [ + 392, + 654, + 609, + 666 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "results on almost all metrics. TPoS(I+T+A) shows strong image quality (FID), but is worse in video quality (FVD) and synchronization (RelSync). TempoToken, on the other hand, is better at synchronization rather than visual quality, likely due to the lack of image input. AADiff is competitive on synchronization but extremely bad on image quality (FID). This is expected as AADiff adjusts each frame using audio amplitude, producing visual changes that highly correlate to audio changes temporally but may be overwhelmed by noises, as shown in Fig. 4a. On Landscapes, its similar FVD to Static but abnormally higher Align-", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Lin Zhang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f86012a6a6f27ff20057e81e52d610c43f5acaf7d67b915b379a21a969852442.jpg", + "table_caption": [ + "Table 2: Effect of (a) first-frame lookups (b) data curation, evaluated on AVSync15." + ], + "table_footnote": [], + "table_body": "
FF-ConvFF-AttnFID↓FVD↓AignSync↑
✗✗11.8383.322.19
✓✗11.6347.422.24
✓✓11.8325.622.33
", + "bbox": [ + 220, + 170, + 493, + 229 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e3092b275de156e555f075753a3fbea48a5d43265a28cc46bcfeefbac52db02f.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
DatasetACMCFID↓FVD↓AignSync↑
VGGSS✗✗12.91307.921.50
AVSync-AC✓✗12.0428.822.09
AVSync15✓✓11.8325.622.33
", + "bbox": [ + 513, + 171, + 787, + 229 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 638, + 229, + 658, + 241 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sync and RelSync also suggest that it only applied minor modifications to the input image due to lack of sound changes, as shown in Fig. 4b. Without audio input, VideoCrafter performs poorly on synchronization. It also has difficulty faithfully adhering to the input frame, as in Fig. 4a. AVSyncD achieves the best animation results on almost all metrics. On Landscapes, AVSyncD also performs the best on FID and FVD, with other scores being similar to ground truth.", + "bbox": [ + 212, + 281, + 784, + 372 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "User Study We invited 15 participants to compare 4 animation models with top overall performance (VideoCrafter, AADiff, I2VD, AVSyncD) on AVSync15, based on 3 metrics in Tab. 1. The 4 models generated videos conditioned on the same test examples (audio+image). Each test example was independently evaluated by 2 participants to select their most preferred generation (vote) on each metric. In total, we evaluated all 150 test examples on AVSync15, collecting $150 \\times 2 = 300$ votes on each metric. Tab. 1a shows votes each model received.", + "bbox": [ + 212, + 373, + 787, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 498, + 401, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Audio Conditioning In Tab. 1, AVSyncD outperforms I2VD, especially on AlignSync and RelSync. AVSyncD did not improve RelSync on Landscapes probably due to the lack of synchronization cues on the dataset itself. These results show that audio condition enhances generation quality and synchronization.", + "bbox": [ + 212, + 521, + 784, + 582 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Audio Guidance Tab. 1 shows increasing the audio guidance factor $\\eta$ from 1 to 4 improves FID, IA, and FVD on all three datasets. As expected, audio guidance also improved AlignSync and RelSync significantly on AVSync15 and TheGreatestHits, but not on the less synchronized Landscapes dataset. Prior works [20, 23, 39] claimed that increasing audio amplitude can also lead to stronger visual effects. We compare this approach with our audio guidance in Fig. 5a. Audio guidance offers a better control mechanism than audio amplitude.", + "bbox": [ + 212, + 583, + 787, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "First-frame (FF) Lookups We validated FF Lookups by replacing them with standard temporal convolutions or spatial self-attention in Tab. 2a.", + "bbox": [ + 212, + 689, + 785, + 720 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Data Curation We compared to AVSyncD trained on random subsets from VGSS and AVSync-AC with equal data scale and balanced categories in Tab. 2b.", + "bbox": [ + 212, + 720, + 785, + 752 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Applications and Extensions", + "text_level": 1, + "bbox": [ + 214, + 771, + 496, + 786 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Animate Generated Images When lacking an image as input, we can use existing image generators to generate the image, which AVSyncD can also animate. Fig. 5b shows animations on images generated by StableDiffusion-V1.5 [34].", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Audio-Synchronized Visual Animation", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/741ecc577b7ccee586195bae66c66765ce4e24d40ce91146ec224f7e52f4b2e5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 372, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e79388744e8b04af778dbfc981cc23a4e57ed62dea088be901836af36782d789.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 195, + 405, + 226 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7d333ad9eafefe2bee8a92966f501b4cd6878d9632bca21c62c8fbc102aa6af0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 148, + 522, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c75d518219ade30adf5f7469acb6af97fb6f94098dfc792964082184db382563.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 534, + 155, + 656, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fe35110bccbdfb0f7473420051d4dd5a5eeded50301c2e6f19106de9be753876.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 663, + 156, + 784, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/54d348134abf2bec79da71e3be43ee2388ec590b4f4893083c0ca03158166b72.jpg", + "image_caption": [ + "Fig. 5: (a): Effects of audio amplitude vs. classifier-free audio guidance. top: original audio with $\\eta = 1$ ; mid: $100\\times$ amplified audio with $\\eta = 1$ ; bottom: original audio with $\\eta = 8$ . (b): animate generated images. (c): Animation with internet images and audios.", + "(a) Controllable animation with (un)related audiios." + ], + "image_footnote": [], + "bbox": [ + 217, + 339, + 784, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/09101c7e7b476f76d84dcf78fe06a7244022083f73c686fb67c9f7e246831bf2.jpg", + "image_caption": [ + "(b)Animate target objects on internet images with audios.", + "Fig. 6: Controllable image animation with audiios. Key frames are visualized." + ], + "image_footnote": [], + "bbox": [ + 217, + 420, + 784, + 473 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Animate Contents from Internet AVSyncD can also generalize well to unseen images and audio, as shown in Fig. 5c.", + "bbox": [ + 212, + 542, + 784, + 571 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Control Animations with (Un)Related Audios We can control the motion of an image to follow desired audio, e.g., animate a baby to not only cry but also bark like a dog or roar like a lion, as seen in Fig. 6a. When there is no object related to the audio, the animations do not demonstrate corresponding motion.", + "bbox": [ + 212, + 574, + 785, + 633 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Animate Target Objects with Audios When multiple objects exist in the image, a scenario not existing in training data, we can still use audios to only animate the related target object, as shown in Fig. 6b.", + "bbox": [ + 212, + 635, + 785, + 681 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 703, + 359, + 719 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We tackled the under-explored Audio-Synchronized Visual Animation task, with an emphasis on generating videos with audio-synchronized dynamics. We contributed the high-quality AVSync15 benchmark via careful data curation and proposed the AVSyncD model to animate images with realistic motions. Due to the scale of AVSync15, our work cannot generalize to all audio classes in the world, which requires several orders of magnitude larger datasets. However, we hope our research inspires further work in this direction.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Lin Zhang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE International Conference on Computer Vision (2021)", + "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "3. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)", + "4. Castellano, B.: Pyscenedetect. https://www.scenedetect.com/", + "5. Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., Weng, C., Shan, Y.: Videocrafter1: Open diffusion models for high-quality video generation (2023)", + "6. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronization in the wild. In: Proceedings of the British Machine Vision Conference (BMVC) (2021)", + "7. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset. In: ICASSP (2020)", + "8. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Wei, F.: Beats: Audio pre-training with acoustic tokenizers. In: ICML (2023)", + "9. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: ACCV (2016)", + "10. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV Workshop (2016)", + "1. Chung, S.W., Chung, J.S., Kang, H.G.: Perfect match: Improved cross-modal embeddings for audio-visual synchronisation. In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2019)", + "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)", + "3. Gemmeke, J.F., Ellis, D.P.W., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: Proc. IEEE ICASSP 2017 (2017)", + "4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023)", + "5. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2023)", + "6. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems (2017)", + "7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "8. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop on Deep Generative Models and Downstream Applications (2022)", + "9. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. In: British Machine Vision Conference (BMVC) (2022)", + "20. Jeong, Y., Ryoo, W., Lee, S., Seo, D., Byeon, W., Kim, S., Kim, J.: The power of sound (tpos): Audio reactive video generation with stable diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7822-7832 (2023)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Audio-Synchronized Visual Animation", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In: ICCV (2023)", + "22. Lee, S.H., Oh, G., Byeon, W., Bae, J., Kim, C., Ryoo, W.J., Yoon, S.H., Kim, J., Kim, S.: Sound-guided semantic video generation. In: ECCV (2022)", + "23. Lee, S., Kong, C., Jeon, D., Kwak, N.: Aadiff: Audio-aligned video synthesis with text-to-image diffusion. In: CVPR Workshop on Content Generation (2023)", + "24. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation. In: ICCV (2021)", + "25. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)", + "26. Luo, S., Yan, C., Hu, C., Zhao, H.: Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In: NeurIPS (2023)", + "27. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)", + "28. Ng, E., Romero, J., Bagautdinov, T., Bai, S., Darrell, T., Kanazawa, A., Richard, A.: From audio to photoreal embodiment: Synthesizing humans in conversations. In: ArXiv (2024)", + "29. Owens, A., Isola, P., McDermott, J., Torralba, A., Adelson, E.H., Freeman, W.T.: Visually indicated sounds. In: CVPR (2016)", + "30. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: AAAI Conference on Artificial Intelligence (AAAI) (2022)", + "31. Pedro Morgado, Ishan Misra, N.V.: Robust audio-visual instance discrimination. In: Computer Vision and Pattern Recognition (CVPR), IEEE/CVF Conf. on (2021)", + "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "33. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: arXiv (2022)", + "34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "35. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N.J., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: CVPR (2023)", + "36. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)", + "37. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data (2022)", + "38. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)", + "39. Sung-Bin, K., Senocak, A., Ha, H., Owens, A., Oh, T.H.: Sound to visual scene generation by audio-to-visual latent alignment. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "40. Tang, Z., Yang, Z., Zhu, C., Zeng, M., Bansal, M.: Any-to-any generation via composable diffusion. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=2EDqbSCnmF" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Lin Zhang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Tsuchida, S., Fukayama, S., Hamasaki, M., Goto, M.: Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In: Proceedings of the 20th International Society for Music Information Retrieval Conference, ISMIR 2019. Delft, Netherlands (Nov 2019)", + "42. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. In: arXiv (2019)", + "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "44. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report (2023)", + "45. Wang, W., Xie, k., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023)", + "46. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern by few-shot tuning a text-to-image diffusion model. arXiv preprint arXiv:2310.10769 (2023)", + "47. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: EMNLP (2021)", + "48. Yariv, G., Gat, I., Benaim, S., Wolf, L., Schwartz, I., Adi, Y.: Diverse and aligned audio-to-video generation via text-to-video model adaptation (2023)", + "49. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. In: ICLR (2023)", + "50. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE International Conference on Computer Vision (ICCV) (2023)", + "51. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: AAAI Conference on Artificial Intelligence (AAAI) (2019)", + "52. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 618 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Audio-Synchronized Visual Animation", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Audio-Synchronized Visual Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_model.json b/2024/Audio-Synchronized Visual Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d4eb8db668ed9cef10f210c4d7e31125158ebbfa --- /dev/null +++ b/2024/Audio-Synchronized Visual Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_model.json @@ -0,0 +1,2500 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.283, + 0.142, + 0.723, + 0.163 + ], + "angle": 0, + "content": "Audio-Synchronized Visual Animation" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.189, + 0.757, + 0.207 + ], + "angle": 0, + "content": "Lin Zhang\\(^{1}\\), Shentong Mo\\(^{2}\\), Yijing Zhang\\(^{1}\\), and Pedro Morgado\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.217, + 0.62, + 0.231 + ], + "angle": 0, + "content": "1 University of Wisconsin-Madison" + }, + { + "type": "text", + "bbox": [ + 0.403, + 0.231, + 0.6, + 0.246 + ], + "angle": 0, + "content": "2 Carnegie Mellon University" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.263, + 0.787, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 1: Given an audio and an image (green box), we produce animations beyond image stylization with complex but natural dynamics, synchronized with audio at each frame. Results were produced by our AVSyncD model trained on the proposed AVSync15. Project webpage: https://lzhangbj.github.io/projects/asma/asva/asva.html." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.512, + 0.741, + 0.734 + ], + "angle": 0, + "content": "Abstract. Current visual generation methods can produce high-quality videos guided by text prompts. However, effectively controlling object dynamics remains a challenge. This work explores audio as a cue to generate temporally synchronized image animations. We introduce Audio-Synchronized Visual Animation (ASVA), a task that aims to animate a static image of an object with motions temporally guided by audio clips. To this end, we present AVSync15, a dataset curated from VGGSound with videos featuring synchronized audio-visual events across 15 categories. We also present a diffusion model, AVSyncD, capable of generating audio-guided animations. Extensive evaluations validate AVSync15 as a reliable benchmark for synchronized generation and demonstrate our model's superior performance. We further explore AVSyncD's potential in a variety of audio-synchronized generation tasks, from generating full videos without a base image to controlling object motions with various sounds. We hope our established benchmark can open new avenues for controllable visual generation." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.746, + 0.655, + 0.761 + ], + "angle": 0, + "content": "Keywords: Audio \\(\\cdot\\) Video \\(\\cdot\\) Synchronization \\(\\cdot\\) Generation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.782, + 0.377, + 0.798 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Generative modeling has witnessed remarkable progress recently, largely due to the emergence of diffusion models [17,34,38]. Conditional generation and, in par" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Lin Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "ticular, text-to-image generation [33,34], has been the focal point given its application potential and availability of high-quality datasets [1,36]. This success has also led to revived interest in video generation, such as text-to-video [5,21,37,45]. While text guidance has been thoroughly investigated, the unique advantages of audio-visual synchrony for video generation remain underexplored. Unlike text, which provides control over global semantics, audio offers both semantic control on videos and precise control at each moment in time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Most existing works on audio-to-visual generation are however either limited to semantic control [20, 22, 39, 40], or constrained on singular scenarios such as human talking face [9-11, 28, 30, 49, 51, 52]. The former focuses on ambient audio datasets lacking synchronization cues with environmental sounds (ranining, fire crakcling, wind) [22] where shifting the audio temporally does not lead to major changes in visual content. The input audio in such cases thus substitutes text to provide only global semantics. The latter, talking faces, while synchronized, are extremely limited in generation diversity and control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.581 + ], + "angle": 0, + "content": "To bridge this gap, we introduce Audio-Synchronized Visual Animation, ASVA, a task which aims to animate objects depicted in natural static images into a video, with clear motion dynamics that are semantically aligned and temporally synchronized with the input audio. ASVA requires a sophisticated understanding of the audio's temporal structure, as well as of how objects move in synchrony with sound. Prior attempts on visual generation guided by diverse sounds fall short in generated visual quality [23] and accurate synchronization control [20,48] due to two challenges: (1) the lack of high-quality training data and benchmarks for learning audio-synchronized visual dynamics; (2) the development of effective methods capable of generating highly synchronized video motions. Successfully addressing these challenges will expand the scope of current video generation methods to enable more fine-grained semantic and temporal control over the generation process via synchronized audio conditioning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We address the first challenge by constructing a high-quality diverse dataset with strong correlations between audio and object motions at each moment in time. In an ideal dataset, sound sources should be easily identifiable in the scene. Every visual motion in the video should highly correlate with the audio semantically and temporally, and vice versa. Moreover, the visual content should be of high quality for generation. However, existing audio-visual datasets either are too noisy [7, 13], containing a large number of unassociated audio-visual pairs [31], or predominantly featuring ambient sound categories that lack meaningful synchronized object dynamics [22, 41]. We thus curate a high-quality dataset from VGGSound [7] by deploying an efficient two-step curation pipeline. In the first step, we use a variety of signal processing techniques and foundation models to automatically filter out videos with poor semantic alignment or temporal synchronization, as well as those depicting static scenes or with too fast camera motions. Then, to ensure the highest possible quality as a benchmark, we narrow down the dataset to sound categories with strong audio-visual synchronization cues, and manually verify the quality of each video. We end up obtaining AVSync15 with 15 dynamic sound classes, each with 100 examples rich" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-Synchronized Visual Animation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "and accurate in semantics, object dynamics, and audio-visual synchronization. An overview of AVSync15 classes is in Figs. 2a and 2b and its comparison with prior audio-video generation datasets is in Suppl. Sec. 5.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.555 + ], + "angle": 0, + "content": "The second challenge pertains to the generation of audio-synchronized motions, which requires a detailed understanding of audio-visual correlations and object dynamics. Take, for instance, a video featuring a dog. To generate realistic video, the model is expected to not only synchronize the dog's mouth with the barking sound, but also accurately depict subtleties in the dog's head pose before and after barking. Furthermore, in a more challenging scenario, the model should discern which object to animate to preserve semantic consistency depending on the input sound. However, existing audio-conditioned visual generation frameworks [14, 39] primarily focus on semantic control, often encoding the audio into a single global semantic feature and thereby neglecting the audio's temporal domain. Even recent attempts at audio-synchronized video generation [20, 23, 48] have not fully realized the potential of audio for fine-grained temporal control, as they either rely on crude audio representations such as audio amplitude [23], learn from weakly synchronized [22], noisy datasets [7], or ignore object dynamics in the generation process [20, 22, 23, 48]. To this end, we introduce Audio-Video Synchronized Diffusion (AVSyncD), a framework improving a pre-trained image latent diffusion model [34] for enhanced audio guidance and motion generation. We employ the pre-trained ImageBind [14] encoder to encode audio into time-aware semantic tokens, then fuse them into each frame's latent features. This allows for precise audio guidance on video semantics and synchronization. To capture complex video motions, we add temporal attention layers to the diffusion model. Finally, to ensure faithful animation of the input image, we incorporate temporal convolutions and attention layers that always reference the input image, i.e., first-frame lookups." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.646 + ], + "angle": 0, + "content": "With the carefully designed dataset and architecture, we are able to train a model specialized for ASVA and produce animations with more realistic and audio-synchronized contents than prior works (Figs. 1 and 4). We provide thorough experiments to validate the effectiveness of AVSync15 and AVSyncD and demonstrate how to deploy AVSyncD for controllable generation, including amplifying audio guidance and semantic-aware object animation (Secs. 5.3 and 5.4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.666, + 0.388, + 0.683 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.696, + 0.52, + 0.71 + ], + "angle": 0, + "content": "2.1 Controllable Visual Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Many conditional visual generation models based on diffusion process [17, 38] have emerged recently. Benefiting from more efficient architecture, large-scale training data [36], and aligned semantic space [32], Latent Diffusion Model (LDM) [34] has achieved great success to generate realistic images conditioned on text. This inspired researchers to explore various diffusion-based visual generation tasks, such as text-to-video [2, 5, 21, 44-46], audio-to-image [14, 39], and audio-to-video [20, 23, 40, 48]. The architectures can be training-free [21, 23, 45], fully-trained [5, 44], or trained partially, which augments a pre-trained LDM by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Lin Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "carefully adding some trainable layers [2, 20, 46]. Extensive works have also attempted to control the semantics of the generated content [15, 25, 27, 50], while how to apply control in the temporal dimension remains under-explored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.253 + ], + "angle": 0, + "content": "In this work, we developed an image animation model AVSyncD to control generation semantically and temporally guided by audio. AVSyncD augments pre-trained StableDiffusion [34] with trainable temporal layers and audio conditioning mechanism, preserving training efficiency and generalizing well." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.276, + 0.488, + 0.29 + ], + "angle": 0, + "content": "2.2 Audio-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.784, + 0.376 + ], + "angle": 0, + "content": "Traditionally, audio has been used as a temporal cue for talking face generation [28, 30, 49, 51, 52], where face and lip actions should be synchronized with audio at each frame. Many works also rely on complex inputs such as 3D meshes and human poses [28, 49]. Although accurately synchronized, this line of research is extremely limited in scenarios and cannot generate videos for diverse audios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.785, + 0.619 + ], + "angle": 0, + "content": "A series of works attempted to expand the class of audio by encoding sound into a global semantic condition for video generation [14, 39, 40], however often overlooked the temporal aspect inherent in audio. Some recent works, although divided audio features into time-aware segments as inputs [22, 35], failed to achieve promising visual quality or synchronize video motions with audios. AADiff [23] is a training-free method re-weighting the text-image cross-attention map in LDM using audio amplitude at each frame, however can only control styles of each frame. TPoS [20] learns segmented audio features aligned with CLIP [32] using sophisticated modules and training losses, and feeds them into a pre-trained text-to-image model [34] for video generation. TempoToken [48] also learns segmented audio features with a pre-trained audio encoder BEATs [8], and fuses them into a pre-trained text-to-video model [44]. However, primarily focused on monotonous sound classes in Landscapes [22] or noisy audio-visual data in VGGSound [7], these methods are limited to generating video semantics without capturing the natural and synchronized dynamics of video content. Frozen generation architectures also prevent them from generating natural motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.785, + 0.709 + ], + "angle": 0, + "content": "To address these limitations, we introduce AVSync15, a high-quality dataset specifically designed for ASVA. AVSync15 stands out from previous efforts by focusing on synchronization cues between audio and visual dynamics, allowing for generating motions beyond mere visual effects. Once trained on AVSync15, our AVSyncD can generalize to many applications to control video motions guided by audios, on which previous methods performed poorly." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.733, + 0.625, + 0.75 + ], + "angle": 0, + "content": "3 Audio-Synchronized Visual Animation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.843 + ], + "angle": 0, + "content": "Formally, the Audio-Synchronized Visual Animation (ASVA) task can be posed as follows. Given an audio clip \\( \\pmb{a} \\) of length \\( T \\) seconds and an image \\( \\pmb{x}_1 \\), the goal is to generate the future video sequence \\( \\pmb{x}_2, \\dots, \\pmb{x}_{rT} \\) (or \\( \\pmb{x}_{2:rT} \\) for short), where \\( r \\) is the desired frame rate. Despite the simple formulation, this is a challenging task as the generated video sequence should (1) be of high visual quality, (2)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-Synchronized Visual Animation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.444, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.251, + 0.34, + 0.263 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.459, + 0.147, + 0.784, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.251, + 0.631, + 0.263 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.331 + ], + "angle": 0, + "content": "Fig. 2: (a): Overview of 15 categories in AVSync15. Categories are listed below x-axis on the right plot. (b): Category-wise averages of av-sync score \\(\\phi\\), IA, and IT on AVSync15 and equivalently sized subsets of VGGSS and AVSync-AC. Error bars for VGGSS and AVSync-AC are obtained from 3 random splits." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.435 + ], + "angle": 0, + "content": "be semantically aligned with the image \\( \\pmb{x}_1 \\) and audio \\( \\pmb{a} \\), (3) exhibit temporal coherence and (4) natural object motions temporally synchronized with the audio \\( \\pmb{a} \\). To facilitate research in ASVA, we introduce a new benchmark that includes a curated high-quality dataset and a suite of evaluation metrics designed to capture the various components of audio-synchronized generation." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.453, + 0.76, + 0.483 + ], + "angle": 0, + "content": "3.1 AVSync15: A High-Quality Dataset for Audio Synchronized Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Existing large-scale audio-visual datasets like VGGSound [6,7,19] and AudioSet [1] often contain amateur videos from platforms like YouTube. These videos, while diverse, can pose challenges for audio-synchronized video generation tasks due to rapid scene changes, camera motion, noisy audio, or out-of-frame sound sources. Prior work [22, 24] has addressed this by focusing on simpler videos, such as those depicting fire crackling or weather patterns. However, such videos often lack strong synchronization cues between audio and visual motion, making them unsuitable for ASVA. To facilitate research in this area, we introduce a high-quality dataset specifically designed for audio-synchronized video generation, ensuring a close synchronization between audio and visuals. More specifically, our selection criteria to create the dataset were: (1) High Correlation: Significant visual changes should be closely associated with audio at each timestamp, and vice versa. (2) Dynamic Content: We included content rich in temporal changes, excluding ambient or monotonous classes. (3) Quality and Relevance: Both video and audio needed to be clean, stable, and semantically aligned." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Preliminary Curation We start from VGGSound [7], a large-scale dataset with 309 diverse audio classes. Similar to VGGSoundSync [6], we first narrow down to 149 classes with potentially clear audio-visual synchronized events, removing ambient classes, which is referred as VGGSS. We then deploy a sequence of automatic cleaning steps and a final manual selection step to identify appropriate videos. The procedures are summarized below and detailed in Suppl. Sec. 5.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Automatic Curation We first use PySceneDetect [4] to cut videos with sharp scene changes into different scenes, which are still likely to contain low-quality" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Lin Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "short clips. To maximize usage, we split each scene into 3-second clips with 0.5-second strides, and discard unsuitable clips based on the following metrics:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.19, + 0.785, + 0.234 + ], + "angle": 0, + "content": "Raw Pixel Difference: We calculate average pixel distances between consecutive frames and remove clips with both small and large values, likely depicting either static or videos with excessive motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.784, + 0.281 + ], + "angle": 0, + "content": "Image Semantics Difference: Complementing the raw pixel analysis above, we also compute distances on CLIP [32] image features, removing videos with small semantics changes such as zoom in/out or large semantic content transitions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.282, + 0.784, + 0.311 + ], + "angle": 0, + "content": "Audio Waveform Amplitude: We exclude clips whose maximum audio waveform amplitude is low, indicating weak audio cues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.784, + 0.357 + ], + "angle": 0, + "content": "Semantic Alignment: We compute the average image-audio (IA) and image-text (IT) alignment scores [32] in a video using ImageBind [14], removing clips with low scores to ensure cross-modal semantic alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.785, + 0.464 + ], + "angle": 0, + "content": "Audio-Video Synchronization: To measure audio-visual synchronization, we follow VGGSoundSync [6] to contrastively train an audio-visual synchronization classifier on VGGSS, ending up with a comparable \\(40.85\\%\\) test accuracy. The model outputs an unbounded av-sync score \\(\\phi_{\\pmb{a}_i,\\pmb{v}_j}\\) for an audio-video pair \\((\\pmb{a}_i,\\pmb{v}_j)\\). During training, we compute \\(\\phi\\) for a synchronized pair \\((\\pmb{a}_i,\\pmb{v}_i)\\) and its temporally-shifted pairs from the same instance. Contrastive loss is then applied to these shifted pairs to maximize the synchronization probability:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.478, + 0.785, + 0.511 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {S y n c}} \\left(\\boldsymbol {a} _ {i}, \\boldsymbol {v} _ {i}\\right) = \\frac {\\exp \\left(\\phi_ {\\boldsymbol {a} _ {i} , \\boldsymbol {v} _ {i}} / \\tau\\right)}{\\sum_ {j} \\exp \\left(\\phi_ {\\boldsymbol {a} _ {i} , \\boldsymbol {v} _ {j}} / \\tau\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.584 + ], + "angle": 0, + "content": "to distinguish the synchronized pair from shifted ones. We use \\( P_{\\mathrm{sync}} \\) as a synchronization indicator to remove low-scoring clips. When computing \\( P_{\\mathrm{sync}} \\), we discard the temperature parameter \\( \\tau \\) used to improve training efficacy. We detail the synchronization classifier and \\( P_{\\mathrm{sync}} \\) in Suppl. Sec. 1 and Sec. 2.1, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.784, + 0.67 + ], + "angle": 0, + "content": "We empirically determine metrics' thresholds by prioritizing quality, acknowledging that some acceptable clips might be discarded. After automatic curation, we merge all continuous 3-second clips from each video and remove categories with less than 100 examples to avoid class imbalance, resulting in AVSync-AC (AVSync w/ Automatic Curation) with 76 categories and 39,902 examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.784, + 0.748 + ], + "angle": 0, + "content": "Manual Curation We further select 15 diverse categories with clear audiovisual motion cues from AVSync-AC for manual refinement. The categories range from animals and human actions to triggered tools and musical instruments. Manual curation once again seeks to identify appropriate videos for ASVA with the criteria above: high correlation, dynamic content, quality and relevance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Dataset Summary The final dataset, AVSync15, contains 90 training and 10 testing videos per category, each \\(2\\sim 10\\) seconds long. We provide an overview of AVSync15 in Fig. 2a. To validate our curation pipeline, we randomly sample three 1500-video splits on the selected 15 categories from VGGSS and AVSync-AC, and quantitatively compare them with AVSync15 in Fig. 2b and Tab. 2b. We also compare AVSync15 with other audio-visual datasets in Suppl. Sec. 5.1." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-Synchronized Visual Animation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.42, + 0.16 + ], + "angle": 0, + "content": "3.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.784, + 0.276 + ], + "angle": 0, + "content": "ASVA is a multi-faceted generation task, necessitating high quality at both image and video level. At the image level, we follow previous conventions [3,12] to use (1) Fréchet Inception Distance (FID) [16] to measure the quality of individual frames; (2) IA [14]/IT [47] to measure image-audio/image-text semantics alignment on CLIP [32] space. At the video level, we use Fréchet Video Distance (FVD) [42] to assess video quality. To measure audio-video synchronization, we compute the following metrics with the trained synchronization classifier:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.784, + 0.309 + ], + "angle": 0, + "content": "RelSync During testing, we use the ground truth audio-visual pair \\((\\pmb{a},\\pmb{v})\\) as a reference to measure synchronization of the generated video \\(\\hat{\\pmb{v}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.32, + 0.785, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e l S y n c} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v}) = \\frac {\\exp \\left(\\phi_ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}}}\\right)}{\\exp \\left(\\phi_ {\\boldsymbol {a} , \\boldsymbol {v}}\\right) + \\exp \\left(\\phi_ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}}}\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.784, + 0.436 + ], + "angle": 0, + "content": "Note that while this reference-based metric normalizes the score by the synchronization of the reference pair \\((\\pmb{a},\\pmb{v})\\), the metric can still be sensitive to the quality of the reference pair. In fact, evaluating synchronization on a dataset where even ground-truth audios and videos are ambiguously synchronized is less informative of the model capabilities, e.g., Landscapes [22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.784, + 0.514 + ], + "angle": 0, + "content": "AlignSync The synchronization classifier is only trained on semantically-aligned and temporally-shifted audio-video pairs sampled from the same instance (See Suppl. Sec. 1.2). RelSync is thus implicitly conditioned on semantics alignment, i.e., P(Sync|Align). To jointly measure semantics alignment and synchronization, we first approximate \\(P_{\\mathrm{Align}}\\) similarly as RelSync:" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.526, + 0.785, + 0.561 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {A l i g n}} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v} _ {1}) = \\frac {1}{b - 1} \\sum_ {i = 2.. b} \\frac {\\exp \\left(\\mathrm {I A} _ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}} _ {i}}\\right)}{\\exp \\left(\\mathrm {I A} _ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}} _ {i}}\\right) + \\exp \\left(\\mathrm {I A} _ {\\boldsymbol {a} , \\boldsymbol {v} _ {1}}\\right)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.784, + 0.617 + ], + "angle": 0, + "content": "where \\(\\pmb{v}_1\\) is the first frame input for animation, and \\(b\\) is the number of generated frames. The generated first frame \\(\\hat{\\pmb{v}}_1\\) is eliminated because it is often a replicate of input \\(\\pmb{v}_1\\). By multiplying \\(P_{\\mathrm{Align}}\\) with RelSync, we obtain the joint score:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.629, + 0.785, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A l i g n S y n c} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v}) = P _ {\\text {A l i g n}} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v} _ {1}) \\cdot \\operatorname {R e l S y n c} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.783, + 0.684 + ], + "angle": 0, + "content": "which jointly measures semantic alignment and temporal synchronization between \\(\\pmb{a}\\) and \\(\\hat{\\pmb{v}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.784, + 0.731 + ], + "angle": 0, + "content": "These automated metrics are not always aligned with human preference. We therefore conduct a user study detailed in Sec. 5.2, asking human raters to compare videos generated by multiple models and select the best one." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.603, + 0.772 + ], + "angle": 0, + "content": "4 Audio-Video Synchronized Diffusion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.628, + 0.801 + ], + "angle": 0, + "content": "4.1 Preliminary: Text-to-Image Latent Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Text-to-image latent diffusion models (LDMs [34]) encode images \\( \\pmb{x} \\) into a lower-dimensional latent space \\( \\pmb{z} = \\mathcal{E}(\\pmb{x}) \\) using a pre-trained perceptual auto-encoder" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Lin Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "\\(\\mathcal{E}(\\cdot)\\), and learn the conditional distribution \\(p(\\boldsymbol{z}|\\boldsymbol{\\tau})\\) given a CLIP-encoded text prompt \\(\\boldsymbol{\\tau}\\). It gradually denoises latents \\(\\boldsymbol{z}^k\\), obtained by corrupting the image latent \\(\\boldsymbol{z}\\) with Gaussian noise \\(\\epsilon\\), over \\(k\\) time steps. A denoising UNet \\(\\epsilon_{\\theta}(\\cdot)\\) parameterized by \\(\\theta\\) is deployed to estimate the added noise \\(\\epsilon\\) by minimizing" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.219, + 0.786, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L D M} = \\mathbb {E} _ {\\boldsymbol {z}, \\epsilon \\sim \\mathcal {N} (0, 1), k} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {z} ^ {k}, k, \\boldsymbol {\\tau}) \\| _ {2} ^ {2} \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.246, + 0.789, + 0.324 + ], + "angle": 0, + "content": "During inference, LDMs start from a random Gaussian noise map \\( \\pmb{z}^{K} \\), and iterate over \\( K \\) reverse diffusion steps [17, 38] to denoise the latents \\( \\pmb{z}^{k-1} = \\pmb{z}^{k} - \\epsilon_{\\theta}(\\pmb{z}^{k}, k, \\pmb{\\tau}) \\), until \\( \\pmb{z}^{0} \\) is obtained. LDMs then decode the latent into an image \\( \\pmb{x}^{0} = \\mathcal{D}(\\pmb{z}^{0}) \\) using the pre-trained decoder \\( \\mathcal{D}(\\cdot) \\). For simplicity, we refer to the images by their latents \\( \\pmb{z} \\) rather than \\( \\pmb{x} \\) throughout the rest of the paper." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.344, + 0.453, + 0.359 + ], + "angle": 0, + "content": "4.2 Proposed Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.789, + 0.564 + ], + "angle": 0, + "content": "In this work, we extend the capabilities of LDMs for ASVA, by focusing on learning video dynamics and temporal synchronization, unlike existing approaches [14, 39, 40] which primarily use audio to control global semantics. We propose the Audio-Video Synchronized Diffusion model (AVSyncD), which builds upon a pre-trained image LDM and integrates synchronized audio control and temporal layers for improved video consistency. The major component is a UNet \\(\\epsilon_{\\theta}\\) aiming to denoise a video instead of an image. The UNet is trained on synchronized audio-video pairs, with the first frame \\(z_{1}\\), the corresponding audio \\(\\pmb{a}\\), and the CLIP encoded audio category name \\(\\pmb{\\tau}\\) as input conditions. The LDM denoising objective in Eq. (5) is applied to the remaining frames \\(z_{2:rT}\\) to be predicted. The architecture of AVSyncD is shown in Fig. 3 and discussed below. Further details, for example, highlighting the different attention modules used in the architecture are described in Suppl. Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.789, + 0.627 + ], + "angle": 0, + "content": "Initial-frame Conditioning To condition LDM on an input image, we feed its latent \\( \\mathbf{z}_1 \\) without noise into the UNet at every diffusion timestep \\( k \\). For all subsequent frames, we adhere to the original LDM, using independently sampled noisy latents \\( \\mathbf{z}_{2:rT}^k \\) as inputs and predicting the added noise \\( \\epsilon^k \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Text Conditioning We retain the text cross-attention layers in the original LDMs [34] (without finetuning) to condition the model on the audio category. However, due to the limited text diversity in training, class conditioning does not bring significant gains (see Suppl. Sec. 6.5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Audio Conditioning To facilitate audio-synchronized generation, we condition the generation on ImageBind audio embeddings [14]. ImageBind computes an audio classification token, \\( \\pmb{a}^g \\), representing global semantics, and local patch tokens, \\( \\pmb{a}_{f,t} \\), across \\( T_{a} \\) timestamps. The original ImageBind only uses \\( \\pmb{a}^g \\) for contrastive learning and discards \\( \\pmb{a}_{f,t} \\). However, we found these frozen local tokens as efficient synchronization cues. We split the patch tokens temporally into \\( rT \\) segments, corresponding to the same timestamps as the frames \\( z_{1:rT} \\) and append the global token to each. Each frame \\( z_{t} \\) then learns both semantics and synchronization guidance from its audio segment \\( \\pmb{a}_{t} \\) via cross-attention [43]. In Suppl. Sec. 6.7, we compared ImageBind with CLAP and BEATs as encoders." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-Synchronized Visual Animation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.147, + 0.745, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.305, + 0.788, + 0.472 + ], + "angle": 0, + "content": "Fig. 3: AVSyncD overview. Left: We use ImageBind to encode audio into semantically aware time-dependent tokens \\((\\pmb{a}_t)_{t=1}^{rT}\\) and CLIP to encode the audio category into text embedding \\(\\pmb{\\tau}\\). In addition, the model receives the latent of the first frame \\(\\pmb{z}_1\\), and iteratively denoises noisy latents of the subsequent frames \\(\\pmb{z}_{2:rT}^k\\) via reverse diffusion. The denoising UNet, based on LDMs [34], consists of a sequence of downsampling, bottleneck and up-sampling blocks, with structure detailed on the right. Right: Anatomy of a UNet block for frame \\(\\pmb{z}_t\\). LDM's original spatial conv, spatial attention and text cross attention layers are frozen, while its spatial self-attention layers are adjusted to first-frame spatial attentions, cross-attending to \\(\\pmb{z}_1\\) instead. To learn video dynamics, we introduce temporal attention layers, and first-frame lookup temporal convolutions applied to input, output, and ResNet layers. We also train audio cross attentions for audio conditioning and synchronization. Trainable layers are marked with \\(\\oplus\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.785, + 0.533 + ], + "angle": 0, + "content": "Spatial Convolution The original LDM's pre-trained spatial convolutional layers were frozen and used without modification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.625 + ], + "angle": 0, + "content": "First-frame Temporal Convolution Each spatial convolution block was augmented with a 1D temporal convolution layer (kernel size 3) to capture temporal dependencies. To better adhere to the starting image, \\( \\mathbf{z}_1 \\), we adjusted the receptive field at frame \\( t \\) to include frames \\( (1, t - 1, t) \\) as opposed to \\( (t - 1, t, t + 1) \\). These temporal convolutions with first-frame lookup were applied to three components in the UNet, namely the input/output conv layers and all ResNet convs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.701 + ], + "angle": 0, + "content": "First-frame Spatial Attention We also leveraged the pre-trained LDM's spatial self-attention. Following [21], we modified the frozen spatial attention layers to cross-attend to the first frame rather than self-attend to the current frame, by computing the key-value pairs from the first frame and the queries from the current one." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Temporal Attention To effectively model long-range visual dependencies, we incorporated temporal attention layers [2]. Each frame index \\( t \\) was converted into a sinusoidal positional embedding, added to the corresponding frame's latents after a learnable projection [43]. Each frame's local representation at spatial position \\( (h,w) \\), \\( z_{hw} \\), was then updated by attending to all frames at the same position \\( (z_{hw1},z_{hw2},\\ldots ,z_{hw(rT)}) \\) through a standard attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Classifier-free Audio Guidance Classifier-free guidance [18] is a technique used to control the impact of input prompts on the generated outputs. We extended it to amplify audio guidance for improved synchronization. To accomplish" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Lin Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "this, we trained the model for both audio-conditioned and unconditioned generation, by randomly replacing \\( \\mathbf{a} \\) with a null audio embedding, \\( \\mathbf{a}_{\\emptyset} \\), with a 20% likelihood. \\( \\mathbf{a}_{\\emptyset} \\) was computed by encoding an all-zero audio mel spectrogram via ImageBind. During inference, a factor \\( \\eta \\geq 1 \\) controls the audio guidance:" + }, + { + "type": "equation", + "bbox": [ + 0.281, + 0.218, + 0.786, + 0.236 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {2: r T} ^ {k - 1} = (1 - \\eta) \\cdot \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {2: r T} ^ {k}, k; \\boldsymbol {z} _ {1}, \\boldsymbol {a} _ {\\emptyset}, \\boldsymbol {\\tau}\\right) + \\eta \\cdot \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {2: r T} ^ {k}, k; \\boldsymbol {z} _ {1}, \\boldsymbol {a}, \\boldsymbol {\\tau}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.245, + 0.786, + 0.276 + ], + "angle": 0, + "content": "where \\(\\eta\\) guides the denoising process towards latents congruent with audio-conditioned generation and away from those of audio-unconditional generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.298, + 0.376, + 0.314 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.328, + 0.393, + 0.344 + ], + "angle": 0, + "content": "5.1 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.789, + 0.549 + ], + "angle": 0, + "content": "Dataset We conducted experiments on three datasets. AVSync15: Our high-quality dataset curated from VGGSound [7], with 15 balanced categories, 1350 training videos, and 150 test videos. We also assessed our data curation pipeline by comparing it to models trained on Landscapes, TheGreatestHits, VGGSS and our AVSync-AC. Landscapes [22] is composed of 9 environmental sound classes. We followed the split in [35,48] with 900 clips for training and 100 for testing. Since Landscapes is full of ambient sounds without synchronized video motion, we mainly use it to evaluate visual quality. TheGreatestHits [29] is an audio-video dataset recording humans probing environments with a drumstick, with 733 videos for training and 244 for testing. The videos are synchronized with audio at the moments of impact but contain lots of static moments. It also offers limited diversity, featuring a singular motion of probing. VGGSS and our AVSync-AC were described in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.789, + 0.613 + ], + "angle": 0, + "content": "Data Preprocess We sampled 2-second synchronized audio-video pairs for experiments. Videos were sampled at 6 fps with 12 frames, and resized to \\(256 \\times 256\\) on AVSync15/Landscapes or \\(128 \\times 256\\) on TheGreatestHits. Following Image-Bind [14], audios were sampled at \\(16\\mathrm{kHz}\\) and converted into 128-d spectrograms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.789, + 0.809 + ], + "angle": 0, + "content": "Baselines We first adopted a simple Static baseline by repeating the input frame into a video, then compared it to several state-of-the-art works. (1) Semantic audio-to-video generation (CoDi [40]): Image, text, and audio are encoded into a shared CLIP [32] space and summed, then fused into a video diffusion model trained on large-scale datasets. (2) Image-to-video generation (VideoCrafter [5]): A superior video diffusion model however without audio inputs. It animates images by fusing CLIP-encoded image and text features into the model via modality-dependent cross-attention layers. (3) Synchronized audio-to-video generation (TPoS [20], AADiff [23], TempoToken [48]): Audio is encoded into time-dependent segments and fused into frozen text-to-image [34] or text-to-video [44] models. We re-implemented AADiff, and used TPoS and TempoToken's pretrained checkpoints on VGGSound and Landscapes. More details are provided in Suppl. Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Training & Evaluation We adopted the pretrained Stable Diffusion-V1.5 [34] as the diffusion model and ImageBind [14] as the audio encoder. All models" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-Synchronized Visual Animation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.788, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.336, + 0.785, + 0.358 + ], + "angle": 0, + "content": "(a) Comparison on AVSync15. AVSyncD produces synchronized dynamic motions, such as muzzle flash, chicken and dog's head poses, and swirling in toilet." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.359, + 0.789, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.484, + 0.788, + 0.517 + ], + "angle": 0, + "content": "(b) Comparison on Landscapes. Without synchronization (c) Generation by AVSyncD on The-cues in audio, AVSyncD still produces evolving visual ef- GreatestHits. AVSyncD generates hitting fecs and semantics, such as flames, waves, and lightning. actions at appropriate moments." + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.528, + 0.651, + 0.542 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results on three datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.615 + ], + "angle": 0, + "content": "were trained using Adam optimizer with a batch size of 64 and a learning rate of 0.0001. Besides metrics in Sec. 3.2, we also provide results using metric in [26] in Suppl. Sec. 3.4. We evaluated on 3 clips uniformly sampled from each video." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.373, + 0.648 + ], + "angle": 0, + "content": "5.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Dataset Comparison In Tab. 1, the Static baseline is a good indicator of dataset attributes. It has similar IA/IT scores compared to Groundtruth, since Static is composed of a subset of groundtruth frames, which are obviously semantically aligned. FVD of Static on TheGreatestHits is relatively low, since the TheGreatestHits contains frequent static ground truth clips without any moving objects. More importantly, the RelSync and AlignSync of Static gradually increase from AVSync15, to TheGreatestHits, to Landscapes, with those on Landscapes even surpassing the ground truth. The fact that static videos perform well in terms of audio synchronization on TheGreatestHits and Landscapes datasets testify to the superiority of AVSync15 for audio-video synchronized generation. Model Comparison We compared to prior works on AVSync15 and Landscapes in Tabs. 1a and 1b and Figs. 4a and 4b. In Tab. 1a, CoDi achieves inferior" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Lin Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 1: Quantitative results. User study shows votes on 3 metrics: image quality, frame consistency, and synchronization. Inputs are combinations of image, text, audio." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.185, + 0.787, + 0.369 + ], + "angle": 0, + "content": "
InputModelFID↓IA↑IT↑FVD↓AlignSync↑RelSync↑User Study
IQ↑FC↑Sync↑
T+ATPoS [20]13.523.3824.832671.019.5242.50---
TempoToken [48]12.218.8417.454466.419.7444.05---
I+TVideoCrafter [5]11.8-29.87840.721.2843.16382012
I2VD12.1-30.35398.221.8043.92629091
I+T+ACoDi [40]14.528.1523.421522.619.5441.51---
TPoS [20]11.938.3630.731227.819.6739.62---
AADiff [23]18.834.2328.97978.022.1145.483745
AVSyncD η = 112.138.3630.34382.722.2544.81---
AVSyncD η = 411.738.5330.45349.122.6245.52163186192
AVSyncD η = 811.737.9930.27420.722.7445.88---
Static-39.7630.391220.421.8343.66
Groundtruth-40.0630.31-25.0450.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.411, + 0.372, + 0.591, + 0.383 + ], + "angle": 0, + "content": "(a) Performance on AVSync15." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.396, + 0.787, + 0.54 + ], + "angle": 0, + "content": "
InputModelFID↓IA↑IT↑FVD↓AignSync↑RelSync↑
T+ATPoS [20]16.515.6126.702081.323.1248.15
TempoToken [48]16.422.5822.872480.024.2148.65
I+TI2VD16.7-22.56539.524.7449.89
I+T+ACoDi [40]20.522.6324.23982.922.6345.48
TPoS [20]16.223.5223.20789.623.5147.05
AADiff [23]70.722.0722.921186.326.7753.93
AVSyncD η = 116.522.2922.81463.124.8149.96
AVSyncD η = 416.222.4922.79415.224.8249.93
Static-23.6022.211177.525.7951.59
Groundtruth-23.6522.08-25.0150.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.408, + 0.543, + 0.596, + 0.554 + ], + "angle": 0, + "content": "(b) Performance on Landscapes." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.566, + 0.787, + 0.654 + ], + "angle": 0, + "content": "
InputModelFID↓IA↑IT↑FVD↓AignSync↑RelSync↑
I+TI2VD9.1-13.42425.022.0544.58
I+T+AAVSyncD η = 19.011.8513.18313.522.5945.52
AVSyncD η = 48.712.0713.31249.322.8345.95
Static-13.3316.56348.924.3648.73
Groundtruth-13.5216.49-25.0250.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.393, + 0.655, + 0.61, + 0.667 + ], + "angle": 0, + "content": "(c) Performance on TheGreatestHits." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "results on almost all metrics. TPoS(I+T+A) shows strong image quality (FID), but is worse in video quality (FVD) and synchronization (RelSync). TempoToken, on the other hand, is better at synchronization rather than visual quality, likely due to the lack of image input. AADiff is competitive on synchronization but extremely bad on image quality (FID). This is expected as AADiff adjusts each frame using audio amplitude, producing visual changes that highly correlate to audio changes temporally but may be overwhelmed by noises, as shown in Fig. 4a. On Landscapes, its similar FVD to Static but abnormally higher Align-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-Synchronized Visual Animation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 2: Effect of (a) first-frame lookups (b) data curation, evaluated on AVSync15." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.171, + 0.495, + 0.23 + ], + "angle": 0, + "content": "
FF-ConvFF-AttnFID↓FVD↓AignSync↑
✗✗11.8383.322.19
✓✗11.6347.422.24
✓✓11.8325.622.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.346, + 0.232, + 0.365, + 0.244 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.172, + 0.788, + 0.23 + ], + "angle": 0, + "content": "
DatasetACMCFID↓FVD↓AignSync↑
VGGSS✗✗12.91307.921.50
AVSync-AC✓✗12.0428.822.09
AVSync15✓✓11.8325.622.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.639, + 0.231, + 0.66, + 0.242 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.785, + 0.373 + ], + "angle": 0, + "content": "Sync and RelSync also suggest that it only applied minor modifications to the input image due to lack of sound changes, as shown in Fig. 4b. Without audio input, VideoCrafter performs poorly on synchronization. It also has difficulty faithfully adhering to the input frame, as in Fig. 4a. AVSyncD achieves the best animation results on almost all metrics. On Landscapes, AVSyncD also performs the best on FID and FVD, with other scores being similar to ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.48 + ], + "angle": 0, + "content": "User Study We invited 15 participants to compare 4 animation models with top overall performance (VideoCrafter, AADiff, I2VD, AVSyncD) on AVSync15, based on 3 metrics in Tab. 1. The 4 models generated videos conditioned on the same test examples (audio+image). Each test example was independently evaluated by 2 participants to select their most preferred generation (vote) on each metric. In total, we evaluated all 150 test examples on AVSync15, collecting \\(150 \\times 2 = 300\\) votes on each metric. Tab. 1a shows votes each model received." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.499, + 0.402, + 0.513 + ], + "angle": 0, + "content": "5.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.583 + ], + "angle": 0, + "content": "Audio Conditioning In Tab. 1, AVSyncD outperforms I2VD, especially on AlignSync and RelSync. AVSyncD did not improve RelSync on Landscapes probably due to the lack of synchronization cues on the dataset itself. These results show that audio condition enhances generation quality and synchronization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Audio Guidance Tab. 1 shows increasing the audio guidance factor \\(\\eta\\) from 1 to 4 improves FID, IA, and FVD on all three datasets. As expected, audio guidance also improved AlignSync and RelSync significantly on AVSync15 and TheGreatestHits, but not on the less synchronized Landscapes dataset. Prior works [20, 23, 39] claimed that increasing audio amplitude can also lead to stronger visual effects. We compare this approach with our audio guidance in Fig. 5a. Audio guidance offers a better control mechanism than audio amplitude." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.721 + ], + "angle": 0, + "content": "First-frame (FF) Lookups We validated FF Lookups by replacing them with standard temporal convolutions or spatial self-attention in Tab. 2a." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.787, + 0.753 + ], + "angle": 0, + "content": "Data Curation We compared to AVSyncD trained on random subsets from VGSS and AVSync-AC with equal data scale and balanced categories in Tab. 2b." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.772, + 0.498, + 0.787 + ], + "angle": 0, + "content": "5.4 Applications and Extensions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Animate Generated Images When lacking an image as input, we can use existing image generators to generate the image, which AVSyncD can also animate. Fig. 5b shows animations on images generated by StableDiffusion-V1.5 [34]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Lin Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.373, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.196, + 0.406, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.257, + 0.382, + 0.269 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.15, + 0.523, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.156, + 0.657, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.258, + 0.605, + 0.269 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.157, + 0.785, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.257, + 0.734, + 0.269 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.28, + 0.788, + 0.324 + ], + "angle": 0, + "content": "Fig. 5: (a): Effects of audio amplitude vs. classifier-free audio guidance. top: original audio with \\(\\eta = 1\\); mid: \\(100\\times\\) amplified audio with \\(\\eta = 1\\); bottom: original audio with \\(\\eta = 8\\). (b): animate generated images. (c): Animation with internet images and audios." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.34, + 0.785, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.406, + 0.651, + 0.418 + ], + "angle": 0, + "content": "(a) Controllable animation with (un)related audiios." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.421, + 0.785, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.476, + 0.672, + 0.488 + ], + "angle": 0, + "content": "(b)Animate target objects on internet images with audios." + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.5, + 0.757, + 0.513 + ], + "angle": 0, + "content": "Fig. 6: Controllable image animation with audiios. Key frames are visualized." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.785, + 0.573 + ], + "angle": 0, + "content": "Animate Contents from Internet AVSyncD can also generalize well to unseen images and audio, as shown in Fig. 5c." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.787, + 0.634 + ], + "angle": 0, + "content": "Control Animations with (Un)Related Audios We can control the motion of an image to follow desired audio, e.g., animate a baby to not only cry but also bark like a dog or roar like a lion, as seen in Fig. 6a. When there is no object related to the audio, the animations do not demonstrate corresponding motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.787, + 0.682 + ], + "angle": 0, + "content": "Animate Target Objects with Audios When multiple objects exist in the image, a scenario not existing in training data, we can still use audios to only animate the related target object, as shown in Fig. 6b." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.36, + 0.72 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We tackled the under-explored Audio-Synchronized Visual Animation task, with an emphasis on generating videos with audio-synchronized dynamics. We contributed the high-quality AVSync15 benchmark via careful data curation and proposed the AVSyncD model to animate images with realistic motions. Due to the scale of AVSync15, our work cannot generalize to all audio classes in the world, which requires several orders of magnitude larger datasets. However, we hope our research inspires further work in this direction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-Synchronized Visual Animation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.287 + ], + "angle": 0, + "content": "3. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.668, + 0.302 + ], + "angle": 0, + "content": "4. Castellano, B.: Pyscenedetect. https://www.scenedetect.com/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.343 + ], + "angle": 0, + "content": "5. Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., Weng, C., Shan, Y.: Videocrafter1: Open diffusion models for high-quality video generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "6. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronization in the wild. In: Proceedings of the British Machine Vision Conference (BMVC) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.412 + ], + "angle": 0, + "content": "7. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset. In: ICASSP (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "8. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Wei, F.: Beats: Audio pre-training with acoustic tokenizers. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.708, + 0.453 + ], + "angle": 0, + "content": "9. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: ACCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.481 + ], + "angle": 0, + "content": "10. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV Workshop (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.523 + ], + "angle": 0, + "content": "1. Chung, S.W., Chung, J.S., Kang, H.G.: Perfect match: Improved cross-modal embeddings for audio-visual synchronisation. In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "3. Gemmeke, J.F., Ellis, D.P.W., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: Proc. IEEE ICASSP 2017 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.62 + ], + "angle": 0, + "content": "4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.647 + ], + "angle": 0, + "content": "5. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "6. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.744 + ], + "angle": 0, + "content": "8. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop on Deep Generative Models and Downstream Applications (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.786 + ], + "angle": 0, + "content": "9. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. In: British Machine Vision Conference (BMVC) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Jeong, Y., Ryoo, W., Lee, S., Seo, D., Byeon, W., Kim, S., Kim, J.: The power of sound (tpos): Audio reactive video generation with stable diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7822-7832 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Lin Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "21. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "22. Lee, S.H., Oh, G., Byeon, W., Bae, J., Kim, C., Ryoo, W.J., Yoon, S.H., Kim, J., Kim, S.: Sound-guided semantic video generation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "23. Lee, S., Kong, C., Jeon, D., Kwak, N.: Aadiff: Audio-aligned video synthesis with text-to-image diffusion. In: CVPR Workshop on Content Generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "24. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.785, + 0.297 + ], + "angle": 0, + "content": "25. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.785, + 0.324 + ], + "angle": 0, + "content": "26. Luo, S., Yan, C., Hu, C., Zhao, H.: Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.324, + 0.785, + 0.351 + ], + "angle": 0, + "content": "27. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.351, + 0.785, + 0.392 + ], + "angle": 0, + "content": "28. Ng, E., Romero, J., Bagautdinov, T., Bai, S., Darrell, T., Kanazawa, A., Richard, A.: From audio to photoreal embodiment: Synthesizing humans in conversations. In: ArXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.419 + ], + "angle": 0, + "content": "29. Owens, A., Isola, P., McDermott, J., Torralba, A., Adelson, E.H., Freeman, W.T.: Visually indicated sounds. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.785, + 0.46 + ], + "angle": 0, + "content": "30. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: AAAI Conference on Artificial Intelligence (AAAI) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "31. Pedro Morgado, Ishan Misra, N.V.: Robust audio-visual instance discrimination. In: Computer Vision and Pattern Recognition (CVPR), IEEE/CVF Conf. on (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.541 + ], + "angle": 0, + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.568 + ], + "angle": 0, + "content": "33. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.595 + ], + "angle": 0, + "content": "34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.636 + ], + "angle": 0, + "content": "35. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N.J., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.785, + 0.704 + ], + "angle": 0, + "content": "36. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.745 + ], + "angle": 0, + "content": "37. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.759 + ], + "angle": 0, + "content": "38. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "39. Sung-Bin, K., Senocak, A., Ha, H., Owens, A., Oh, T.H.: Sound to visual scene generation by audio-to-visual latent alignment. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "40. Tang, Z., Yang, Z., Zhu, C., Zeng, M., Bansal, M.: Any-to-any generation via composable diffusion. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=2EDqbSCnmF" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-Synchronized Visual Animation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "41. Tsuchida, S., Fukayama, S., Hamasaki, M., Goto, M.: Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In: Proceedings of the 20th International Society for Music Information Retrieval Conference, ISMIR 2019. Delft, Netherlands (Nov 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "42. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. In: arXiv (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "44. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "45. Wang, W., Xie, k., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "46. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern by few-shot tuning a text-to-image diffusion model. arXiv preprint arXiv:2310.10769 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.383, + 0.787, + 0.425 + ], + "angle": 0, + "content": "47. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: EMNLP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.452 + ], + "angle": 0, + "content": "48. Yariv, G., Gat, I., Benaim, S., Wolf, L., Schwartz, I., Adi, Y.: Diverse and aligned audio-to-video generation via text-to-video model adaptation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.481 + ], + "angle": 0, + "content": "49. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "50. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "51. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: AAAI Conference on Artificial Intelligence (AAAI) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.564, + 0.787, + 0.619 + ], + "angle": 0, + "content": "52. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.619 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Audio-Synchronized Visual Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_origin.pdf b/2024/Audio-Synchronized Visual Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..db4d7e8b65af5c3af7f9cb94214abee3260e9bae --- /dev/null +++ b/2024/Audio-Synchronized Visual Animation/c68a2a32-6ab9-4dd4-9cc7-ad90d05d4f52_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0f9a06200682996d770fe9b9131f6d126a1454a582654b5079415fcc8bfbec9 +size 11310831 diff --git a/2024/Audio-Synchronized Visual Animation/full.md b/2024/Audio-Synchronized Visual Animation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ccb7758ae1c4b55c28b98c236a8155659b158a45 --- /dev/null +++ b/2024/Audio-Synchronized Visual Animation/full.md @@ -0,0 +1,325 @@ +# Audio-Synchronized Visual Animation + +Lin Zhang $^{1}$ , Shentong Mo $^{2}$ , Yijing Zhang $^{1}$ , and Pedro Morgado $^{1}$ + +1 University of Wisconsin-Madison + +2 Carnegie Mellon University + +![](images/43af284ff5c1aba269647c60332aaf7fff45c9bfef3c0359af50f511ff04f78e.jpg) +Fig. 1: Given an audio and an image (green box), we produce animations beyond image stylization with complex but natural dynamics, synchronized with audio at each frame. Results were produced by our AVSyncD model trained on the proposed AVSync15. Project webpage: https://lzhangbj.github.io/projects/asma/asva/asva.html. + +Abstract. Current visual generation methods can produce high-quality videos guided by text prompts. However, effectively controlling object dynamics remains a challenge. This work explores audio as a cue to generate temporally synchronized image animations. We introduce Audio-Synchronized Visual Animation (ASVA), a task that aims to animate a static image of an object with motions temporally guided by audio clips. To this end, we present AVSync15, a dataset curated from VGGSound with videos featuring synchronized audio-visual events across 15 categories. We also present a diffusion model, AVSyncD, capable of generating audio-guided animations. Extensive evaluations validate AVSync15 as a reliable benchmark for synchronized generation and demonstrate our model's superior performance. We further explore AVSyncD's potential in a variety of audio-synchronized generation tasks, from generating full videos without a base image to controlling object motions with various sounds. We hope our established benchmark can open new avenues for controllable visual generation. + +Keywords: Audio $\cdot$ Video $\cdot$ Synchronization $\cdot$ Generation + +# 1 Introduction + +Generative modeling has witnessed remarkable progress recently, largely due to the emergence of diffusion models [17,34,38]. Conditional generation and, in par + +ticular, text-to-image generation [33,34], has been the focal point given its application potential and availability of high-quality datasets [1,36]. This success has also led to revived interest in video generation, such as text-to-video [5,21,37,45]. While text guidance has been thoroughly investigated, the unique advantages of audio-visual synchrony for video generation remain underexplored. Unlike text, which provides control over global semantics, audio offers both semantic control on videos and precise control at each moment in time. + +Most existing works on audio-to-visual generation are however either limited to semantic control [20, 22, 39, 40], or constrained on singular scenarios such as human talking face [9-11, 28, 30, 49, 51, 52]. The former focuses on ambient audio datasets lacking synchronization cues with environmental sounds (ranining, fire crakcling, wind) [22] where shifting the audio temporally does not lead to major changes in visual content. The input audio in such cases thus substitutes text to provide only global semantics. The latter, talking faces, while synchronized, are extremely limited in generation diversity and control. + +To bridge this gap, we introduce Audio-Synchronized Visual Animation, ASVA, a task which aims to animate objects depicted in natural static images into a video, with clear motion dynamics that are semantically aligned and temporally synchronized with the input audio. ASVA requires a sophisticated understanding of the audio's temporal structure, as well as of how objects move in synchrony with sound. Prior attempts on visual generation guided by diverse sounds fall short in generated visual quality [23] and accurate synchronization control [20,48] due to two challenges: (1) the lack of high-quality training data and benchmarks for learning audio-synchronized visual dynamics; (2) the development of effective methods capable of generating highly synchronized video motions. Successfully addressing these challenges will expand the scope of current video generation methods to enable more fine-grained semantic and temporal control over the generation process via synchronized audio conditioning. + +We address the first challenge by constructing a high-quality diverse dataset with strong correlations between audio and object motions at each moment in time. In an ideal dataset, sound sources should be easily identifiable in the scene. Every visual motion in the video should highly correlate with the audio semantically and temporally, and vice versa. Moreover, the visual content should be of high quality for generation. However, existing audio-visual datasets either are too noisy [7, 13], containing a large number of unassociated audio-visual pairs [31], or predominantly featuring ambient sound categories that lack meaningful synchronized object dynamics [22, 41]. We thus curate a high-quality dataset from VGGSound [7] by deploying an efficient two-step curation pipeline. In the first step, we use a variety of signal processing techniques and foundation models to automatically filter out videos with poor semantic alignment or temporal synchronization, as well as those depicting static scenes or with too fast camera motions. Then, to ensure the highest possible quality as a benchmark, we narrow down the dataset to sound categories with strong audio-visual synchronization cues, and manually verify the quality of each video. We end up obtaining AVSync15 with 15 dynamic sound classes, each with 100 examples rich + +and accurate in semantics, object dynamics, and audio-visual synchronization. An overview of AVSync15 classes is in Figs. 2a and 2b and its comparison with prior audio-video generation datasets is in Suppl. Sec. 5.1. + +The second challenge pertains to the generation of audio-synchronized motions, which requires a detailed understanding of audio-visual correlations and object dynamics. Take, for instance, a video featuring a dog. To generate realistic video, the model is expected to not only synchronize the dog's mouth with the barking sound, but also accurately depict subtleties in the dog's head pose before and after barking. Furthermore, in a more challenging scenario, the model should discern which object to animate to preserve semantic consistency depending on the input sound. However, existing audio-conditioned visual generation frameworks [14, 39] primarily focus on semantic control, often encoding the audio into a single global semantic feature and thereby neglecting the audio's temporal domain. Even recent attempts at audio-synchronized video generation [20, 23, 48] have not fully realized the potential of audio for fine-grained temporal control, as they either rely on crude audio representations such as audio amplitude [23], learn from weakly synchronized [22], noisy datasets [7], or ignore object dynamics in the generation process [20, 22, 23, 48]. To this end, we introduce Audio-Video Synchronized Diffusion (AVSyncD), a framework improving a pre-trained image latent diffusion model [34] for enhanced audio guidance and motion generation. We employ the pre-trained ImageBind [14] encoder to encode audio into time-aware semantic tokens, then fuse them into each frame's latent features. This allows for precise audio guidance on video semantics and synchronization. To capture complex video motions, we add temporal attention layers to the diffusion model. Finally, to ensure faithful animation of the input image, we incorporate temporal convolutions and attention layers that always reference the input image, i.e., first-frame lookups. + +With the carefully designed dataset and architecture, we are able to train a model specialized for ASVA and produce animations with more realistic and audio-synchronized contents than prior works (Figs. 1 and 4). We provide thorough experiments to validate the effectiveness of AVSync15 and AVSyncD and demonstrate how to deploy AVSyncD for controllable generation, including amplifying audio guidance and semantic-aware object animation (Secs. 5.3 and 5.4). + +# 2 Related Work + +# 2.1 Controllable Visual Generation + +Many conditional visual generation models based on diffusion process [17, 38] have emerged recently. Benefiting from more efficient architecture, large-scale training data [36], and aligned semantic space [32], Latent Diffusion Model (LDM) [34] has achieved great success to generate realistic images conditioned on text. This inspired researchers to explore various diffusion-based visual generation tasks, such as text-to-video [2, 5, 21, 44-46], audio-to-image [14, 39], and audio-to-video [20, 23, 40, 48]. The architectures can be training-free [21, 23, 45], fully-trained [5, 44], or trained partially, which augments a pre-trained LDM by + +carefully adding some trainable layers [2, 20, 46]. Extensive works have also attempted to control the semantics of the generated content [15, 25, 27, 50], while how to apply control in the temporal dimension remains under-explored. + +In this work, we developed an image animation model AVSyncD to control generation semantically and temporally guided by audio. AVSyncD augments pre-trained StableDiffusion [34] with trainable temporal layers and audio conditioning mechanism, preserving training efficiency and generalizing well. + +# 2.2 Audio-to-Video Generation + +Traditionally, audio has been used as a temporal cue for talking face generation [28, 30, 49, 51, 52], where face and lip actions should be synchronized with audio at each frame. Many works also rely on complex inputs such as 3D meshes and human poses [28, 49]. Although accurately synchronized, this line of research is extremely limited in scenarios and cannot generate videos for diverse audios. + +A series of works attempted to expand the class of audio by encoding sound into a global semantic condition for video generation [14, 39, 40], however often overlooked the temporal aspect inherent in audio. Some recent works, although divided audio features into time-aware segments as inputs [22, 35], failed to achieve promising visual quality or synchronize video motions with audios. AADiff [23] is a training-free method re-weighting the text-image cross-attention map in LDM using audio amplitude at each frame, however can only control styles of each frame. TPoS [20] learns segmented audio features aligned with CLIP [32] using sophisticated modules and training losses, and feeds them into a pre-trained text-to-image model [34] for video generation. TempoToken [48] also learns segmented audio features with a pre-trained audio encoder BEATs [8], and fuses them into a pre-trained text-to-video model [44]. However, primarily focused on monotonous sound classes in Landscapes [22] or noisy audio-visual data in VGGSound [7], these methods are limited to generating video semantics without capturing the natural and synchronized dynamics of video content. Frozen generation architectures also prevent them from generating natural motions. + +To address these limitations, we introduce AVSync15, a high-quality dataset specifically designed for ASVA. AVSync15 stands out from previous efforts by focusing on synchronization cues between audio and visual dynamics, allowing for generating motions beyond mere visual effects. Once trained on AVSync15, our AVSyncD can generalize to many applications to control video motions guided by audios, on which previous methods performed poorly. + +# 3 Audio-Synchronized Visual Animation + +Formally, the Audio-Synchronized Visual Animation (ASVA) task can be posed as follows. Given an audio clip $\pmb{a}$ of length $T$ seconds and an image $\pmb{x}_1$ , the goal is to generate the future video sequence $\pmb{x}_2, \dots, \pmb{x}_{rT}$ (or $\pmb{x}_{2:rT}$ for short), where $r$ is the desired frame rate. Despite the simple formulation, this is a challenging task as the generated video sequence should (1) be of high visual quality, (2) + +![](images/835fa2fbaee3831d2b2532d55dd24815728e1f2e8b722dddb39ac274fe145976.jpg) +(a) + +![](images/bb5dc3b06a9c08f25dae32a92972f5bd44df3ba2811b6f334d36e4c3a11217b6.jpg) +(b) +Fig. 2: (a): Overview of 15 categories in AVSync15. Categories are listed below x-axis on the right plot. (b): Category-wise averages of av-sync score $\phi$ , IA, and IT on AVSync15 and equivalently sized subsets of VGGSS and AVSync-AC. Error bars for VGGSS and AVSync-AC are obtained from 3 random splits. + +be semantically aligned with the image $\pmb{x}_1$ and audio $\pmb{a}$ , (3) exhibit temporal coherence and (4) natural object motions temporally synchronized with the audio $\pmb{a}$ . To facilitate research in ASVA, we introduce a new benchmark that includes a curated high-quality dataset and a suite of evaluation metrics designed to capture the various components of audio-synchronized generation. + +# 3.1 AVSync15: A High-Quality Dataset for Audio Synchronized Video Generation + +Existing large-scale audio-visual datasets like VGGSound [6,7,19] and AudioSet [1] often contain amateur videos from platforms like YouTube. These videos, while diverse, can pose challenges for audio-synchronized video generation tasks due to rapid scene changes, camera motion, noisy audio, or out-of-frame sound sources. Prior work [22, 24] has addressed this by focusing on simpler videos, such as those depicting fire crackling or weather patterns. However, such videos often lack strong synchronization cues between audio and visual motion, making them unsuitable for ASVA. To facilitate research in this area, we introduce a high-quality dataset specifically designed for audio-synchronized video generation, ensuring a close synchronization between audio and visuals. More specifically, our selection criteria to create the dataset were: (1) High Correlation: Significant visual changes should be closely associated with audio at each timestamp, and vice versa. (2) Dynamic Content: We included content rich in temporal changes, excluding ambient or monotonous classes. (3) Quality and Relevance: Both video and audio needed to be clean, stable, and semantically aligned. + +Preliminary Curation We start from VGGSound [7], a large-scale dataset with 309 diverse audio classes. Similar to VGGSoundSync [6], we first narrow down to 149 classes with potentially clear audio-visual synchronized events, removing ambient classes, which is referred as VGGSS. We then deploy a sequence of automatic cleaning steps and a final manual selection step to identify appropriate videos. The procedures are summarized below and detailed in Suppl. Sec. 5.2. + +Automatic Curation We first use PySceneDetect [4] to cut videos with sharp scene changes into different scenes, which are still likely to contain low-quality + +short clips. To maximize usage, we split each scene into 3-second clips with 0.5-second strides, and discard unsuitable clips based on the following metrics: + +Raw Pixel Difference: We calculate average pixel distances between consecutive frames and remove clips with both small and large values, likely depicting either static or videos with excessive motion. + +Image Semantics Difference: Complementing the raw pixel analysis above, we also compute distances on CLIP [32] image features, removing videos with small semantics changes such as zoom in/out or large semantic content transitions. + +Audio Waveform Amplitude: We exclude clips whose maximum audio waveform amplitude is low, indicating weak audio cues. + +Semantic Alignment: We compute the average image-audio (IA) and image-text (IT) alignment scores [32] in a video using ImageBind [14], removing clips with low scores to ensure cross-modal semantic alignment. + +Audio-Video Synchronization: To measure audio-visual synchronization, we follow VGGSoundSync [6] to contrastively train an audio-visual synchronization classifier on VGGSS, ending up with a comparable $40.85\%$ test accuracy. The model outputs an unbounded av-sync score $\phi_{\pmb{a}_i,\pmb{v}_j}$ for an audio-video pair $(\pmb{a}_i,\pmb{v}_j)$ . During training, we compute $\phi$ for a synchronized pair $(\pmb{a}_i,\pmb{v}_i)$ and its temporally-shifted pairs from the same instance. Contrastive loss is then applied to these shifted pairs to maximize the synchronization probability: + +$$ +P _ {\text {S y n c}} \left(\boldsymbol {a} _ {i}, \boldsymbol {v} _ {i}\right) = \frac {\exp \left(\phi_ {\boldsymbol {a} _ {i} , \boldsymbol {v} _ {i}} / \tau\right)}{\sum_ {j} \exp \left(\phi_ {\boldsymbol {a} _ {i} , \boldsymbol {v} _ {j}} / \tau\right)} \tag {1} +$$ + +to distinguish the synchronized pair from shifted ones. We use $P_{\mathrm{sync}}$ as a synchronization indicator to remove low-scoring clips. When computing $P_{\mathrm{sync}}$ , we discard the temperature parameter $\tau$ used to improve training efficacy. We detail the synchronization classifier and $P_{\mathrm{sync}}$ in Suppl. Sec. 1 and Sec. 2.1, respectively. + +We empirically determine metrics' thresholds by prioritizing quality, acknowledging that some acceptable clips might be discarded. After automatic curation, we merge all continuous 3-second clips from each video and remove categories with less than 100 examples to avoid class imbalance, resulting in AVSync-AC (AVSync w/ Automatic Curation) with 76 categories and 39,902 examples. + +Manual Curation We further select 15 diverse categories with clear audiovisual motion cues from AVSync-AC for manual refinement. The categories range from animals and human actions to triggered tools and musical instruments. Manual curation once again seeks to identify appropriate videos for ASVA with the criteria above: high correlation, dynamic content, quality and relevance. + +Dataset Summary The final dataset, AVSync15, contains 90 training and 10 testing videos per category, each $2\sim 10$ seconds long. We provide an overview of AVSync15 in Fig. 2a. To validate our curation pipeline, we randomly sample three 1500-video splits on the selected 15 categories from VGGSS and AVSync-AC, and quantitatively compare them with AVSync15 in Fig. 2b and Tab. 2b. We also compare AVSync15 with other audio-visual datasets in Suppl. Sec. 5.1. + +# 3.2 Evaluation Metrics + +ASVA is a multi-faceted generation task, necessitating high quality at both image and video level. At the image level, we follow previous conventions [3,12] to use (1) Fréchet Inception Distance (FID) [16] to measure the quality of individual frames; (2) IA [14]/IT [47] to measure image-audio/image-text semantics alignment on CLIP [32] space. At the video level, we use Fréchet Video Distance (FVD) [42] to assess video quality. To measure audio-video synchronization, we compute the following metrics with the trained synchronization classifier: + +RelSync During testing, we use the ground truth audio-visual pair $(\pmb{a},\pmb{v})$ as a reference to measure synchronization of the generated video $\hat{\pmb{v}}$ as follows: + +$$ +\operatorname {R e l S y n c} (\boldsymbol {a}, \hat {\boldsymbol {v}}, \boldsymbol {v}) = \frac {\exp \left(\phi_ {\boldsymbol {a} , \hat {\boldsymbol {v}}}\right)}{\exp \left(\phi_ {\boldsymbol {a} , \boldsymbol {v}}\right) + \exp \left(\phi_ {\boldsymbol {a} , \hat {\boldsymbol {v}}}\right)} \tag {2} +$$ + +Note that while this reference-based metric normalizes the score by the synchronization of the reference pair $(\pmb{a},\pmb{v})$ , the metric can still be sensitive to the quality of the reference pair. In fact, evaluating synchronization on a dataset where even ground-truth audios and videos are ambiguously synchronized is less informative of the model capabilities, e.g., Landscapes [22]. + +AlignSync The synchronization classifier is only trained on semantically-aligned and temporally-shifted audio-video pairs sampled from the same instance (See Suppl. Sec. 1.2). RelSync is thus implicitly conditioned on semantics alignment, i.e., P(Sync|Align). To jointly measure semantics alignment and synchronization, we first approximate $P_{\mathrm{Align}}$ similarly as RelSync: + +$$ +P _ {\text {A l i g n}} (\boldsymbol {a}, \hat {\boldsymbol {v}}, \boldsymbol {v} _ {1}) = \frac {1}{b - 1} \sum_ {i = 2.. b} \frac {\exp \left(\mathrm {I A} _ {\boldsymbol {a} , \hat {\boldsymbol {v}} _ {i}}\right)}{\exp \left(\mathrm {I A} _ {\boldsymbol {a} , \hat {\boldsymbol {v}} _ {i}}\right) + \exp \left(\mathrm {I A} _ {\boldsymbol {a} , \boldsymbol {v} _ {1}}\right)} \tag {3} +$$ + +where $\pmb{v}_1$ is the first frame input for animation, and $b$ is the number of generated frames. The generated first frame $\hat{\pmb{v}}_1$ is eliminated because it is often a replicate of input $\pmb{v}_1$ . By multiplying $P_{\mathrm{Align}}$ with RelSync, we obtain the joint score: + +$$ +\operatorname {A l i g n S y n c} (\boldsymbol {a}, \hat {\boldsymbol {v}}, \boldsymbol {v}) = P _ {\text {A l i g n}} (\boldsymbol {a}, \hat {\boldsymbol {v}}, \boldsymbol {v} _ {1}) \cdot \operatorname {R e l S y n c} (\boldsymbol {a}, \hat {\boldsymbol {v}}, \boldsymbol {v}) \tag {4} +$$ + +which jointly measures semantic alignment and temporal synchronization between $\pmb{a}$ and $\hat{\pmb{v}}$ . + +These automated metrics are not always aligned with human preference. We therefore conduct a user study detailed in Sec. 5.2, asking human raters to compare videos generated by multiple models and select the best one. + +# 4 Audio-Video Synchronized Diffusion + +# 4.1 Preliminary: Text-to-Image Latent Diffusion + +Text-to-image latent diffusion models (LDMs [34]) encode images $\pmb{x}$ into a lower-dimensional latent space $\pmb{z} = \mathcal{E}(\pmb{x})$ using a pre-trained perceptual auto-encoder + +$\mathcal{E}(\cdot)$ , and learn the conditional distribution $p(\boldsymbol{z}|\boldsymbol{\tau})$ given a CLIP-encoded text prompt $\boldsymbol{\tau}$ . It gradually denoises latents $\boldsymbol{z}^k$ , obtained by corrupting the image latent $\boldsymbol{z}$ with Gaussian noise $\epsilon$ , over $k$ time steps. A denoising UNet $\epsilon_{\theta}(\cdot)$ parameterized by $\theta$ is deployed to estimate the added noise $\epsilon$ by minimizing + +$$ +\mathcal {L} _ {L D M} = \mathbb {E} _ {\boldsymbol {z}, \epsilon \sim \mathcal {N} (0, 1), k} \left[ \| \epsilon - \epsilon_ {\theta} (\boldsymbol {z} ^ {k}, k, \boldsymbol {\tau}) \| _ {2} ^ {2} \right] \tag {5} +$$ + +During inference, LDMs start from a random Gaussian noise map $\pmb{z}^{K}$ , and iterate over $K$ reverse diffusion steps [17, 38] to denoise the latents $\pmb{z}^{k-1} = \pmb{z}^{k} - \epsilon_{\theta}(\pmb{z}^{k}, k, \pmb{\tau})$ , until $\pmb{z}^{0}$ is obtained. LDMs then decode the latent into an image $\pmb{x}^{0} = \mathcal{D}(\pmb{z}^{0})$ using the pre-trained decoder $\mathcal{D}(\cdot)$ . For simplicity, we refer to the images by their latents $\pmb{z}$ rather than $\pmb{x}$ throughout the rest of the paper. + +# 4.2 Proposed Architecture + +In this work, we extend the capabilities of LDMs for ASVA, by focusing on learning video dynamics and temporal synchronization, unlike existing approaches [14, 39, 40] which primarily use audio to control global semantics. We propose the Audio-Video Synchronized Diffusion model (AVSyncD), which builds upon a pre-trained image LDM and integrates synchronized audio control and temporal layers for improved video consistency. The major component is a UNet $\epsilon_{\theta}$ aiming to denoise a video instead of an image. The UNet is trained on synchronized audio-video pairs, with the first frame $z_{1}$ , the corresponding audio $\pmb{a}$ , and the CLIP encoded audio category name $\pmb{\tau}$ as input conditions. The LDM denoising objective in Eq. (5) is applied to the remaining frames $z_{2:rT}$ to be predicted. The architecture of AVSyncD is shown in Fig. 3 and discussed below. Further details, for example, highlighting the different attention modules used in the architecture are described in Suppl. Sec. 3.2. + +Initial-frame Conditioning To condition LDM on an input image, we feed its latent $\mathbf{z}_1$ without noise into the UNet at every diffusion timestep $k$ . For all subsequent frames, we adhere to the original LDM, using independently sampled noisy latents $\mathbf{z}_{2:rT}^k$ as inputs and predicting the added noise $\epsilon^k$ . + +Text Conditioning We retain the text cross-attention layers in the original LDMs [34] (without finetuning) to condition the model on the audio category. However, due to the limited text diversity in training, class conditioning does not bring significant gains (see Suppl. Sec. 6.5). + +Audio Conditioning To facilitate audio-synchronized generation, we condition the generation on ImageBind audio embeddings [14]. ImageBind computes an audio classification token, $\pmb{a}^g$ , representing global semantics, and local patch tokens, $\pmb{a}_{f,t}$ , across $T_{a}$ timestamps. The original ImageBind only uses $\pmb{a}^g$ for contrastive learning and discards $\pmb{a}_{f,t}$ . However, we found these frozen local tokens as efficient synchronization cues. We split the patch tokens temporally into $rT$ segments, corresponding to the same timestamps as the frames $z_{1:rT}$ and append the global token to each. Each frame $z_{t}$ then learns both semantics and synchronization guidance from its audio segment $\pmb{a}_{t}$ via cross-attention [43]. In Suppl. Sec. 6.7, we compared ImageBind with CLAP and BEATs as encoders. + +![](images/1bd66f98f8d4eae3fefeaba02b910f21dd921c8882d448d582879daf4a8ae176.jpg) +Fig. 3: AVSyncD overview. Left: We use ImageBind to encode audio into semantically aware time-dependent tokens $(\pmb{a}_t)_{t=1}^{rT}$ and CLIP to encode the audio category into text embedding $\pmb{\tau}$ . In addition, the model receives the latent of the first frame $\pmb{z}_1$ , and iteratively denoises noisy latents of the subsequent frames $\pmb{z}_{2:rT}^k$ via reverse diffusion. The denoising UNet, based on LDMs [34], consists of a sequence of downsampling, bottleneck and up-sampling blocks, with structure detailed on the right. Right: Anatomy of a UNet block for frame $\pmb{z}_t$ . LDM's original spatial conv, spatial attention and text cross attention layers are frozen, while its spatial self-attention layers are adjusted to first-frame spatial attentions, cross-attending to $\pmb{z}_1$ instead. To learn video dynamics, we introduce temporal attention layers, and first-frame lookup temporal convolutions applied to input, output, and ResNet layers. We also train audio cross attentions for audio conditioning and synchronization. Trainable layers are marked with $\oplus$ . + +Spatial Convolution The original LDM's pre-trained spatial convolutional layers were frozen and used without modification. + +First-frame Temporal Convolution Each spatial convolution block was augmented with a 1D temporal convolution layer (kernel size 3) to capture temporal dependencies. To better adhere to the starting image, $\mathbf{z}_1$ , we adjusted the receptive field at frame $t$ to include frames $(1, t - 1, t)$ as opposed to $(t - 1, t, t + 1)$ . These temporal convolutions with first-frame lookup were applied to three components in the UNet, namely the input/output conv layers and all ResNet convs. + +First-frame Spatial Attention We also leveraged the pre-trained LDM's spatial self-attention. Following [21], we modified the frozen spatial attention layers to cross-attend to the first frame rather than self-attend to the current frame, by computing the key-value pairs from the first frame and the queries from the current one. + +Temporal Attention To effectively model long-range visual dependencies, we incorporated temporal attention layers [2]. Each frame index $t$ was converted into a sinusoidal positional embedding, added to the corresponding frame's latents after a learnable projection [43]. Each frame's local representation at spatial position $(h,w)$ , $z_{hw}$ , was then updated by attending to all frames at the same position $(z_{hw1},z_{hw2},\ldots ,z_{hw(rT)})$ through a standard attention mechanism. + +Classifier-free Audio Guidance Classifier-free guidance [18] is a technique used to control the impact of input prompts on the generated outputs. We extended it to amplify audio guidance for improved synchronization. To accomplish + +this, we trained the model for both audio-conditioned and unconditioned generation, by randomly replacing $\mathbf{a}$ with a null audio embedding, $\mathbf{a}_{\emptyset}$ , with a 20% likelihood. $\mathbf{a}_{\emptyset}$ was computed by encoding an all-zero audio mel spectrogram via ImageBind. During inference, a factor $\eta \geq 1$ controls the audio guidance: + +$$ +\boldsymbol {z} _ {2: r T} ^ {k - 1} = (1 - \eta) \cdot \epsilon_ {\theta} \left(\boldsymbol {z} _ {2: r T} ^ {k}, k; \boldsymbol {z} _ {1}, \boldsymbol {a} _ {\emptyset}, \boldsymbol {\tau}\right) + \eta \cdot \epsilon_ {\theta} \left(\boldsymbol {z} _ {2: r T} ^ {k}, k; \boldsymbol {z} _ {1}, \boldsymbol {a}, \boldsymbol {\tau}\right) \tag {6} +$$ + +where $\eta$ guides the denoising process towards latents congruent with audio-conditioned generation and away from those of audio-unconditional generation. + +# 5 Experiments + +# 5.1 Implementation + +Dataset We conducted experiments on three datasets. AVSync15: Our high-quality dataset curated from VGGSound [7], with 15 balanced categories, 1350 training videos, and 150 test videos. We also assessed our data curation pipeline by comparing it to models trained on Landscapes, TheGreatestHits, VGGSS and our AVSync-AC. Landscapes [22] is composed of 9 environmental sound classes. We followed the split in [35,48] with 900 clips for training and 100 for testing. Since Landscapes is full of ambient sounds without synchronized video motion, we mainly use it to evaluate visual quality. TheGreatestHits [29] is an audio-video dataset recording humans probing environments with a drumstick, with 733 videos for training and 244 for testing. The videos are synchronized with audio at the moments of impact but contain lots of static moments. It also offers limited diversity, featuring a singular motion of probing. VGGSS and our AVSync-AC were described in Sec. 3. + +Data Preprocess We sampled 2-second synchronized audio-video pairs for experiments. Videos were sampled at 6 fps with 12 frames, and resized to $256 \times 256$ on AVSync15/Landscapes or $128 \times 256$ on TheGreatestHits. Following Image-Bind [14], audios were sampled at $16\mathrm{kHz}$ and converted into 128-d spectrograms. + +Baselines We first adopted a simple Static baseline by repeating the input frame into a video, then compared it to several state-of-the-art works. (1) Semantic audio-to-video generation (CoDi [40]): Image, text, and audio are encoded into a shared CLIP [32] space and summed, then fused into a video diffusion model trained on large-scale datasets. (2) Image-to-video generation (VideoCrafter [5]): A superior video diffusion model however without audio inputs. It animates images by fusing CLIP-encoded image and text features into the model via modality-dependent cross-attention layers. (3) Synchronized audio-to-video generation (TPoS [20], AADiff [23], TempoToken [48]): Audio is encoded into time-dependent segments and fused into frozen text-to-image [34] or text-to-video [44] models. We re-implemented AADiff, and used TPoS and TempoToken's pretrained checkpoints on VGGSound and Landscapes. More details are provided in Suppl. Sec. 4. + +Training & Evaluation We adopted the pretrained Stable Diffusion-V1.5 [34] as the diffusion model and ImageBind [14] as the audio encoder. All models + +![](images/e8892f0690f70eb192e769bf36aa8c4a75a15945eab56609dfabd42aca9fd297.jpg) + +![](images/058c123d73accdbcd10bf911170d2915a113bd94be723b6cb8d3c6723a3fbc9c.jpg) +(a) Comparison on AVSync15. AVSyncD produces synchronized dynamic motions, such as muzzle flash, chicken and dog's head poses, and swirling in toilet. +(b) Comparison on Landscapes. Without synchronization (c) Generation by AVSyncD on The-cues in audio, AVSyncD still produces evolving visual ef- GreatestHits. AVSyncD generates hitting fecs and semantics, such as flames, waves, and lightning. actions at appropriate moments. +Fig. 4: Qualitative results on three datasets. + +were trained using Adam optimizer with a batch size of 64 and a learning rate of 0.0001. Besides metrics in Sec. 3.2, we also provide results using metric in [26] in Suppl. Sec. 3.4. We evaluated on 3 clips uniformly sampled from each video. + +# 5.2 Main Results + +Dataset Comparison In Tab. 1, the Static baseline is a good indicator of dataset attributes. It has similar IA/IT scores compared to Groundtruth, since Static is composed of a subset of groundtruth frames, which are obviously semantically aligned. FVD of Static on TheGreatestHits is relatively low, since the TheGreatestHits contains frequent static ground truth clips without any moving objects. More importantly, the RelSync and AlignSync of Static gradually increase from AVSync15, to TheGreatestHits, to Landscapes, with those on Landscapes even surpassing the ground truth. The fact that static videos perform well in terms of audio synchronization on TheGreatestHits and Landscapes datasets testify to the superiority of AVSync15 for audio-video synchronized generation. Model Comparison We compared to prior works on AVSync15 and Landscapes in Tabs. 1a and 1b and Figs. 4a and 4b. In Tab. 1a, CoDi achieves inferior + +Table 1: Quantitative results. User study shows votes on 3 metrics: image quality, frame consistency, and synchronization. Inputs are combinations of image, text, audio. + +
InputModelFID↓IA↑IT↑FVD↓AlignSync↑RelSync↑User Study
IQ↑FC↑Sync↑
T+ATPoS [20]13.523.3824.832671.019.5242.50---
TempoToken [48]12.218.8417.454466.419.7444.05---
I+TVideoCrafter [5]11.8-29.87840.721.2843.16382012
I2VD12.1-30.35398.221.8043.92629091
I+T+ACoDi [40]14.528.1523.421522.619.5441.51---
TPoS [20]11.938.3630.731227.819.6739.62---
AADiff [23]18.834.2328.97978.022.1145.483745
AVSyncD η = 112.138.3630.34382.722.2544.81---
AVSyncD η = 411.738.5330.45349.122.6245.52163186192
AVSyncD η = 811.737.9930.27420.722.7445.88---
Static-39.7630.391220.421.8343.66
Groundtruth-40.0630.31-25.0450.00
+ +(a) Performance on AVSync15. + +
InputModelFID↓IA↑IT↑FVD↓AignSync↑RelSync↑
T+ATPoS [20]16.515.6126.702081.323.1248.15
TempoToken [48]16.422.5822.872480.024.2148.65
I+TI2VD16.7-22.56539.524.7449.89
I+T+ACoDi [40]20.522.6324.23982.922.6345.48
TPoS [20]16.223.5223.20789.623.5147.05
AADiff [23]70.722.0722.921186.326.7753.93
AVSyncD η = 116.522.2922.81463.124.8149.96
AVSyncD η = 416.222.4922.79415.224.8249.93
Static-23.6022.211177.525.7951.59
Groundtruth-23.6522.08-25.0150.00
+ +(b) Performance on Landscapes. + +
InputModelFID↓IA↑IT↑FVD↓AignSync↑RelSync↑
I+TI2VD9.1-13.42425.022.0544.58
I+T+AAVSyncD η = 19.011.8513.18313.522.5945.52
AVSyncD η = 48.712.0713.31249.322.8345.95
Static-13.3316.56348.924.3648.73
Groundtruth-13.5216.49-25.0250.00
+ +(c) Performance on TheGreatestHits. + +results on almost all metrics. TPoS(I+T+A) shows strong image quality (FID), but is worse in video quality (FVD) and synchronization (RelSync). TempoToken, on the other hand, is better at synchronization rather than visual quality, likely due to the lack of image input. AADiff is competitive on synchronization but extremely bad on image quality (FID). This is expected as AADiff adjusts each frame using audio amplitude, producing visual changes that highly correlate to audio changes temporally but may be overwhelmed by noises, as shown in Fig. 4a. On Landscapes, its similar FVD to Static but abnormally higher Align- + +Table 2: Effect of (a) first-frame lookups (b) data curation, evaluated on AVSync15. + +
FF-ConvFF-AttnFID↓FVD↓AignSync↑
✗✗11.8383.322.19
✓✗11.6347.422.24
✓✓11.8325.622.33
+ +(a) + +
DatasetACMCFID↓FVD↓AignSync↑
VGGSS✗✗12.91307.921.50
AVSync-AC✓✗12.0428.822.09
AVSync15✓✓11.8325.622.33
+ +(b) + +Sync and RelSync also suggest that it only applied minor modifications to the input image due to lack of sound changes, as shown in Fig. 4b. Without audio input, VideoCrafter performs poorly on synchronization. It also has difficulty faithfully adhering to the input frame, as in Fig. 4a. AVSyncD achieves the best animation results on almost all metrics. On Landscapes, AVSyncD also performs the best on FID and FVD, with other scores being similar to ground truth. + +User Study We invited 15 participants to compare 4 animation models with top overall performance (VideoCrafter, AADiff, I2VD, AVSyncD) on AVSync15, based on 3 metrics in Tab. 1. The 4 models generated videos conditioned on the same test examples (audio+image). Each test example was independently evaluated by 2 participants to select their most preferred generation (vote) on each metric. In total, we evaluated all 150 test examples on AVSync15, collecting $150 \times 2 = 300$ votes on each metric. Tab. 1a shows votes each model received. + +# 5.3 Ablation Studies + +Audio Conditioning In Tab. 1, AVSyncD outperforms I2VD, especially on AlignSync and RelSync. AVSyncD did not improve RelSync on Landscapes probably due to the lack of synchronization cues on the dataset itself. These results show that audio condition enhances generation quality and synchronization. + +Audio Guidance Tab. 1 shows increasing the audio guidance factor $\eta$ from 1 to 4 improves FID, IA, and FVD on all three datasets. As expected, audio guidance also improved AlignSync and RelSync significantly on AVSync15 and TheGreatestHits, but not on the less synchronized Landscapes dataset. Prior works [20, 23, 39] claimed that increasing audio amplitude can also lead to stronger visual effects. We compare this approach with our audio guidance in Fig. 5a. Audio guidance offers a better control mechanism than audio amplitude. + +First-frame (FF) Lookups We validated FF Lookups by replacing them with standard temporal convolutions or spatial self-attention in Tab. 2a. + +Data Curation We compared to AVSyncD trained on random subsets from VGSS and AVSync-AC with equal data scale and balanced categories in Tab. 2b. + +# 5.4 Applications and Extensions + +Animate Generated Images When lacking an image as input, we can use existing image generators to generate the image, which AVSyncD can also animate. Fig. 5b shows animations on images generated by StableDiffusion-V1.5 [34]. + +![](images/741ecc577b7ccee586195bae66c66765ce4e24d40ce91146ec224f7e52f4b2e5.jpg) +(a) + +![](images/e79388744e8b04af778dbfc981cc23a4e57ed62dea088be901836af36782d789.jpg) + +![](images/7d333ad9eafefe2bee8a92966f501b4cd6878d9632bca21c62c8fbc102aa6af0.jpg) + +![](images/c75d518219ade30adf5f7469acb6af97fb6f94098dfc792964082184db382563.jpg) +(b) + +![](images/fe35110bccbdfb0f7473420051d4dd5a5eeded50301c2e6f19106de9be753876.jpg) +(c) + +![](images/54d348134abf2bec79da71e3be43ee2388ec590b4f4893083c0ca03158166b72.jpg) +Fig. 5: (a): Effects of audio amplitude vs. classifier-free audio guidance. top: original audio with $\eta = 1$ ; mid: $100\times$ amplified audio with $\eta = 1$ ; bottom: original audio with $\eta = 8$ . (b): animate generated images. (c): Animation with internet images and audios. +(a) Controllable animation with (un)related audiios. + +![](images/09101c7e7b476f76d84dcf78fe06a7244022083f73c686fb67c9f7e246831bf2.jpg) +(b)Animate target objects on internet images with audios. +Fig. 6: Controllable image animation with audiios. Key frames are visualized. + +Animate Contents from Internet AVSyncD can also generalize well to unseen images and audio, as shown in Fig. 5c. + +Control Animations with (Un)Related Audios We can control the motion of an image to follow desired audio, e.g., animate a baby to not only cry but also bark like a dog or roar like a lion, as seen in Fig. 6a. When there is no object related to the audio, the animations do not demonstrate corresponding motion. + +Animate Target Objects with Audios When multiple objects exist in the image, a scenario not existing in training data, we can still use audios to only animate the related target object, as shown in Fig. 6b. + +# 6 Conclusion + +We tackled the under-explored Audio-Synchronized Visual Animation task, with an emphasis on generating videos with audio-synchronized dynamics. We contributed the high-quality AVSync15 benchmark via careful data curation and proposed the AVSyncD model to animate images with realistic motions. Due to the scale of AVSync15, our work cannot generalize to all audio classes in the world, which requires several orders of magnitude larger datasets. However, we hope our research inspires further work in this direction. + +# References + +1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE International Conference on Computer Vision (2021) +2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +3. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019) +4. Castellano, B.: Pyscenedetect. https://www.scenedetect.com/ +5. Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., Weng, C., Shan, Y.: Videocrafter1: Open diffusion models for high-quality video generation (2023) +6. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronization in the wild. In: Proceedings of the British Machine Vision Conference (BMVC) (2021) +7. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset. In: ICASSP (2020) +8. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Wei, F.: Beats: Audio pre-training with acoustic tokenizers. In: ICML (2023) +9. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: ACCV (2016) +10. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV Workshop (2016) +1. Chung, S.W., Chung, J.S., Kang, H.G.: Perfect match: Improved cross-modal embeddings for audio-visual synchronisation. In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2019) +2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021) +3. Gemmeke, J.F., Ellis, D.P.W., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: Proc. IEEE ICASSP 2017 (2017) +4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023) +5. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2023) +6. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems (2017) +7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +8. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop on Deep Generative Models and Downstream Applications (2022) +9. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. In: British Machine Vision Conference (BMVC) (2022) +20. Jeong, Y., Ryoo, W., Lee, S., Seo, D., Byeon, W., Kim, S., Kim, J.: The power of sound (tpos): Audio reactive video generation with stable diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7822-7832 (2023) + +21. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In: ICCV (2023) +22. Lee, S.H., Oh, G., Byeon, W., Bae, J., Kim, C., Ryoo, W.J., Yoon, S.H., Kim, J., Kim, S.: Sound-guided semantic video generation. In: ECCV (2022) +23. Lee, S., Kong, C., Jeon, D., Kwak, N.: Aadiff: Audio-aligned video synthesis with text-to-image diffusion. In: CVPR Workshop on Content Generation (2023) +24. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation. In: ICCV (2021) +25. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023) +26. Luo, S., Yan, C., Hu, C., Zhao, H.: Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In: NeurIPS (2023) +27. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023) +28. Ng, E., Romero, J., Bagautdinov, T., Bai, S., Darrell, T., Kanazawa, A., Richard, A.: From audio to photoreal embodiment: Synthesizing humans in conversations. In: ArXiv (2024) +29. Owens, A., Isola, P., McDermott, J., Torralba, A., Adelson, E.H., Freeman, W.T.: Visually indicated sounds. In: CVPR (2016) +30. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: AAAI Conference on Artificial Intelligence (AAAI) (2022) +31. Pedro Morgado, Ishan Misra, N.V.: Robust audio-visual instance discrimination. In: Computer Vision and Pattern Recognition (CVPR), IEEE/CVF Conf. on (2021) +32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) +33. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: arXiv (2022) +34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +35. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N.J., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: CVPR (2023) +36. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022) +37. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data (2022) +38. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) +39. Sung-Bin, K., Senocak, A., Ha, H., Owens, A., Oh, T.H.: Sound to visual scene generation by audio-to-visual latent alignment. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +40. Tang, Z., Yang, Z., Zhu, C., Zeng, M., Bansal, M.: Any-to-any generation via composable diffusion. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=2EDqbSCnmF + +41. Tsuchida, S., Fukayama, S., Hamasaki, M., Goto, M.: Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In: Proceedings of the 20th International Society for Music Information Retrieval Conference, ISMIR 2019. Delft, Netherlands (Nov 2019) +42. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. In: arXiv (2019) +43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +44. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report (2023) +45. Wang, W., Xie, k., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023) +46. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern by few-shot tuning a text-to-image diffusion model. arXiv preprint arXiv:2310.10769 (2023) +47. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: EMNLP (2021) +48. Yariv, G., Gat, I., Benaim, S., Wolf, L., Schwartz, I., Adi, Y.: Diverse and aligned audio-to-video generation via text-to-video model adaptation (2023) +49. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. In: ICLR (2023) +50. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE International Conference on Computer Vision (ICCV) (2023) +51. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: AAAI Conference on Artificial Intelligence (AAAI) (2019) +52. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) \ No newline at end of file diff --git a/2024/Audio-Synchronized Visual Animation/images.zip b/2024/Audio-Synchronized Visual Animation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ccfc6a5e51b66cf5d58e95946e9ef52e3e902b1 --- /dev/null +++ b/2024/Audio-Synchronized Visual Animation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:697383fc5f7072badc544051f09aa8120f8ac739d854108a63c5870a3f0daa08 +size 628409 diff --git a/2024/Audio-Synchronized Visual Animation/layout.json b/2024/Audio-Synchronized Visual Animation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e37d6ad9fd41c773cc260345ea6853348814c0a5 --- /dev/null +++ b/2024/Audio-Synchronized Visual Animation/layout.json @@ -0,0 +1,9834 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 112, + 442, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 112, + 442, + 129 + ], + "spans": [ + { + "bbox": [ + 173, + 112, + 442, + 129 + ], + "type": "text", + "content": "Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "spans": [ + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "type": "text", + "content": "Lin Zhang" + }, + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "type": "text", + "content": ", Shentong Mo" + }, + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "type": "text", + "content": ", Yijing Zhang" + }, + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "type": "text", + "content": ", and Pedro Morgado" + }, + { + "bbox": [ + 149, + 149, + 463, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 235, + 171, + 379, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 171, + 379, + 182 + ], + "spans": [ + { + "bbox": [ + 235, + 171, + 379, + 182 + ], + "type": "text", + "content": "1 University of Wisconsin-Madison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 246, + 182, + 367, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 182, + 367, + 194 + ], + "spans": [ + { + "bbox": [ + 246, + 182, + 367, + 194 + ], + "type": "text", + "content": "2 Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 208, + 481, + 338 + ], + "blocks": [ + { + "bbox": [ + 133, + 208, + 481, + 338 + ], + "lines": [ + { + "bbox": [ + 133, + 208, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 133, + 208, + 481, + 338 + ], + "type": "image", + "image_path": "43af284ff5c1aba269647c60332aaf7fff45c9bfef3c0359af50f511ff04f78e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 346, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 390 + ], + "type": "text", + "content": "Fig. 1: Given an audio and an image (green box), we produce animations beyond image stylization with complex but natural dynamics, synchronized with audio at each frame. Results were produced by our AVSyncD model trained on the proposed AVSync15. Project webpage: https://lzhangbj.github.io/projects/asma/asva/asva.html." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 405, + 453, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 405, + 453, + 581 + ], + "spans": [ + { + "bbox": [ + 159, + 405, + 453, + 581 + ], + "type": "text", + "content": "Abstract. Current visual generation methods can produce high-quality videos guided by text prompts. However, effectively controlling object dynamics remains a challenge. This work explores audio as a cue to generate temporally synchronized image animations. We introduce Audio-Synchronized Visual Animation (ASVA), a task that aims to animate a static image of an object with motions temporally guided by audio clips. To this end, we present AVSync15, a dataset curated from VGGSound with videos featuring synchronized audio-visual events across 15 categories. We also present a diffusion model, AVSyncD, capable of generating audio-guided animations. Extensive evaluations validate AVSync15 as a reliable benchmark for synchronized generation and demonstrate our model's superior performance. We further explore AVSyncD's potential in a variety of audio-synchronized generation tasks, from generating full videos without a base image to controlling object motions with various sounds. We hope our established benchmark can open new avenues for controllable visual generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 590, + 400, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 590, + 400, + 602 + ], + "spans": [ + { + "bbox": [ + 160, + 590, + 400, + 602 + ], + "type": "text", + "content": "Keywords: Audio " + }, + { + "bbox": [ + 160, + 590, + 400, + 602 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 590, + 400, + 602 + ], + "type": "text", + "content": " Video " + }, + { + "bbox": [ + 160, + 590, + 400, + 602 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 590, + 400, + 602 + ], + "type": "text", + "content": " Synchronization " + }, + { + "bbox": [ + 160, + 590, + 400, + 602 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 590, + 400, + 602 + ], + "type": "text", + "content": " Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 619, + 230, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 230, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 230, + 632 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Generative modeling has witnessed remarkable progress recently, largely due to the emergence of diffusion models [17,34,38]. Conditional generation and, in par" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "ticular, text-to-image generation [33,34], has been the focal point given its application potential and availability of high-quality datasets [1,36]. This success has also led to revived interest in video generation, such as text-to-video [5,21,37,45]. While text guidance has been thoroughly investigated, the unique advantages of audio-visual synchrony for video generation remain underexplored. Unlike text, which provides control over global semantics, audio offers both semantic control on videos and precise control at each moment in time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 204, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 300 + ], + "type": "text", + "content": "Most existing works on audio-to-visual generation are however either limited to semantic control [20, 22, 39, 40], or constrained on singular scenarios such as human talking face [9-11, 28, 30, 49, 51, 52]. The former focuses on ambient audio datasets lacking synchronization cues with environmental sounds (ranining, fire crakcling, wind) [22] where shifting the audio temporally does not lead to major changes in visual content. The input audio in such cases thus substitutes text to provide only global semantics. The latter, talking faces, while synchronized, are extremely limited in generation diversity and control." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 303, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 460 + ], + "type": "text", + "content": "To bridge this gap, we introduce Audio-Synchronized Visual Animation, ASVA, a task which aims to animate objects depicted in natural static images into a video, with clear motion dynamics that are semantically aligned and temporally synchronized with the input audio. ASVA requires a sophisticated understanding of the audio's temporal structure, as well as of how objects move in synchrony with sound. Prior attempts on visual generation guided by diverse sounds fall short in generated visual quality [23] and accurate synchronization control [20,48] due to two challenges: (1) the lack of high-quality training data and benchmarks for learning audio-synchronized visual dynamics; (2) the development of effective methods capable of generating highly synchronized video motions. Successfully addressing these challenges will expand the scope of current video generation methods to enable more fine-grained semantic and temporal control over the generation process via synchronized audio conditioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "We address the first challenge by constructing a high-quality diverse dataset with strong correlations between audio and object motions at each moment in time. In an ideal dataset, sound sources should be easily identifiable in the scene. Every visual motion in the video should highly correlate with the audio semantically and temporally, and vice versa. Moreover, the visual content should be of high quality for generation. However, existing audio-visual datasets either are too noisy [7, 13], containing a large number of unassociated audio-visual pairs [31], or predominantly featuring ambient sound categories that lack meaningful synchronized object dynamics [22, 41]. We thus curate a high-quality dataset from VGGSound [7] by deploying an efficient two-step curation pipeline. In the first step, we use a variety of signal processing techniques and foundation models to automatically filter out videos with poor semantic alignment or temporal synchronization, as well as those depicting static scenes or with too fast camera motions. Then, to ensure the highest possible quality as a benchmark, we narrow down the dataset to sound categories with strong audio-visual synchronization cues, and manually verify the quality of each video. We end up obtaining AVSync15 with 15 dynamic sound classes, each with 100 examples rich" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Lin Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "and accurate in semantics, object dynamics, and audio-visual synchronization. An overview of AVSync15 classes is in Figs. 2a and 2b and its comparison with prior audio-video generation datasets is in Suppl. Sec. 5.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 439 + ], + "type": "text", + "content": "The second challenge pertains to the generation of audio-synchronized motions, which requires a detailed understanding of audio-visual correlations and object dynamics. Take, for instance, a video featuring a dog. To generate realistic video, the model is expected to not only synchronize the dog's mouth with the barking sound, but also accurately depict subtleties in the dog's head pose before and after barking. Furthermore, in a more challenging scenario, the model should discern which object to animate to preserve semantic consistency depending on the input sound. However, existing audio-conditioned visual generation frameworks [14, 39] primarily focus on semantic control, often encoding the audio into a single global semantic feature and thereby neglecting the audio's temporal domain. Even recent attempts at audio-synchronized video generation [20, 23, 48] have not fully realized the potential of audio for fine-grained temporal control, as they either rely on crude audio representations such as audio amplitude [23], learn from weakly synchronized [22], noisy datasets [7], or ignore object dynamics in the generation process [20, 22, 23, 48]. To this end, we introduce Audio-Video Synchronized Diffusion (AVSyncD), a framework improving a pre-trained image latent diffusion model [34] for enhanced audio guidance and motion generation. We employ the pre-trained ImageBind [14] encoder to encode audio into time-aware semantic tokens, then fuse them into each frame's latent features. This allows for precise audio guidance on video semantics and synchronization. To capture complex video motions, we add temporal attention layers to the diffusion model. Finally, to ensure faithful animation of the input image, we incorporate temporal convolutions and attention layers that always reference the input image, i.e., first-frame lookups." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "type": "text", + "content": "With the carefully designed dataset and architecture, we are able to train a model specialized for ASVA and produce animations with more realistic and audio-synchronized contents than prior works (Figs. 1 and 4). We provide thorough experiments to validate the effectiveness of AVSync15 and AVSyncD and demonstrate how to deploy AVSyncD for controllable generation, including amplifying audio guidance and semantic-aware object animation (Secs. 5.3 and 5.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 527, + 237, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 527, + 237, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 237, + 540 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 551, + 318, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 318, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 318, + 562 + ], + "type": "text", + "content": "2.1 Controllable Visual Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Many conditional visual generation models based on diffusion process [17, 38] have emerged recently. Benefiting from more efficient architecture, large-scale training data [36], and aligned semantic space [32], Latent Diffusion Model (LDM) [34] has achieved great success to generate realistic images conditioned on text. This inspired researchers to explore various diffusion-based visual generation tasks, such as text-to-video [2, 5, 21, 44-46], audio-to-image [14, 39], and audio-to-video [20, 23, 40, 48]. The architectures can be training-free [21, 23, 45], fully-trained [5, 44], or trained partially, which augments a pre-trained LDM by" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "carefully adding some trainable layers [2, 20, 46]. Extensive works have also attempted to control the semantics of the generated content [15, 25, 27, 50], while how to apply control in the temporal dimension remains under-explored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 200 + ], + "type": "text", + "content": "In this work, we developed an image animation model AVSyncD to control generation semantically and temporally guided by audio. AVSyncD augments pre-trained StableDiffusion [34] with trainable temporal layers and audio conditioning mechanism, preserving training efficiency and generalizing well." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 218, + 298, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 218, + 298, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 218, + 298, + 229 + ], + "type": "text", + "content": "2.2 Audio-to-Video Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 238, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 479, + 297 + ], + "type": "text", + "content": "Traditionally, audio has been used as a temporal cue for talking face generation [28, 30, 49, 51, 52], where face and lip actions should be synchronized with audio at each frame. Many works also rely on complex inputs such as 3D meshes and human poses [28, 49]. Although accurately synchronized, this line of research is extremely limited in scenarios and cannot generate videos for diverse audios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 298, + 480, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 480, + 490 + ], + "type": "text", + "content": "A series of works attempted to expand the class of audio by encoding sound into a global semantic condition for video generation [14, 39, 40], however often overlooked the temporal aspect inherent in audio. Some recent works, although divided audio features into time-aware segments as inputs [22, 35], failed to achieve promising visual quality or synchronize video motions with audios. AADiff [23] is a training-free method re-weighting the text-image cross-attention map in LDM using audio amplitude at each frame, however can only control styles of each frame. TPoS [20] learns segmented audio features aligned with CLIP [32] using sophisticated modules and training losses, and feeds them into a pre-trained text-to-image model [34] for video generation. TempoToken [48] also learns segmented audio features with a pre-trained audio encoder BEATs [8], and fuses them into a pre-trained text-to-video model [44]. However, primarily focused on monotonous sound classes in Landscapes [22] or noisy audio-visual data in VGGSound [7], these methods are limited to generating video semantics without capturing the natural and synchronized dynamics of video content. Frozen generation architectures also prevent them from generating natural motions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 490, + 480, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 480, + 561 + ], + "type": "text", + "content": "To address these limitations, we introduce AVSync15, a high-quality dataset specifically designed for ASVA. AVSync15 stands out from previous efforts by focusing on synchronization cues between audio and visual dynamics, allowing for generating motions beyond mere visual effects. Once trained on AVSync15, our AVSyncD can generalize to many applications to control video motions guided by audios, on which previous methods performed poorly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 580, + 382, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 580, + 382, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 580, + 382, + 594 + ], + "type": "text", + "content": "3 Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "text", + "content": "Formally, the Audio-Synchronized Visual Animation (ASVA) task can be posed as follows. Given an audio clip " + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "text", + "content": " seconds and an image " + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_1" + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "text", + "content": ", the goal is to generate the future video sequence " + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_2, \\dots, \\pmb{x}_{rT}" + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{2:rT}" + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "text", + "content": " for short), where " + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 605, + 480, + 667 + ], + "type": "text", + "content": " is the desired frame rate. Despite the simple formulation, this is a challenging task as the generated video sequence should (1) be of high visual quality, (2)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Lin Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 271, + 197 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 271, + 197 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 271, + 197 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 271, + 197 + ], + "type": "image", + "image_path": "835fa2fbaee3831d2b2532d55dd24815728e1f2e8b722dddb39ac274fe145976.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 198, + 208, + 208 + ], + "lines": [ + { + "bbox": [ + 196, + 198, + 208, + 208 + ], + "spans": [ + { + "bbox": [ + 196, + 198, + 208, + 208 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 280, + 116, + 479, + 194 + ], + "blocks": [ + { + "bbox": [ + 280, + 116, + 479, + 194 + ], + "lines": [ + { + "bbox": [ + 280, + 116, + 479, + 194 + ], + "spans": [ + { + "bbox": [ + 280, + 116, + 479, + 194 + ], + "type": "image", + "image_path": "bb5dc3b06a9c08f25dae32a92972f5bd44df3ba2811b6f334d36e4c3a11217b6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 198, + 386, + 208 + ], + "lines": [ + { + "bbox": [ + 373, + 198, + 386, + 208 + ], + "spans": [ + { + "bbox": [ + 373, + 198, + 386, + 208 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "type": "text", + "content": "Fig. 2: (a): Overview of 15 categories in AVSync15. Categories are listed below x-axis on the right plot. (b): Category-wise averages of av-sync score " + }, + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "type": "text", + "content": ", IA, and IT on AVSync15 and equivalently sized subsets of VGGSS and AVSync-AC. Error bars for VGGSS and AVSync-AC are obtained from 3 random splits." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 344 + ], + "type": "text", + "content": "be semantically aligned with the image " + }, + { + "bbox": [ + 130, + 282, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\pmb{x}_1" + }, + { + "bbox": [ + 130, + 282, + 482, + 344 + ], + "type": "text", + "content": " and audio " + }, + { + "bbox": [ + 130, + 282, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 130, + 282, + 482, + 344 + ], + "type": "text", + "content": ", (3) exhibit temporal coherence and (4) natural object motions temporally synchronized with the audio " + }, + { + "bbox": [ + 130, + 282, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 130, + 282, + 482, + 344 + ], + "type": "text", + "content": ". To facilitate research in ASVA, we introduce a new benchmark that includes a curated high-quality dataset and a suite of evaluation metrics designed to capture the various components of audio-synchronized generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 358, + 465, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 465, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 465, + 382 + ], + "type": "text", + "content": "3.1 AVSync15: A High-Quality Dataset for Audio Synchronized Video Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 388, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 567 + ], + "type": "text", + "content": "Existing large-scale audio-visual datasets like VGGSound [6,7,19] and AudioSet [1] often contain amateur videos from platforms like YouTube. These videos, while diverse, can pose challenges for audio-synchronized video generation tasks due to rapid scene changes, camera motion, noisy audio, or out-of-frame sound sources. Prior work [22, 24] has addressed this by focusing on simpler videos, such as those depicting fire crackling or weather patterns. However, such videos often lack strong synchronization cues between audio and visual motion, making them unsuitable for ASVA. To facilitate research in this area, we introduce a high-quality dataset specifically designed for audio-synchronized video generation, ensuring a close synchronization between audio and visuals. More specifically, our selection criteria to create the dataset were: (1) High Correlation: Significant visual changes should be closely associated with audio at each timestamp, and vice versa. (2) Dynamic Content: We included content rich in temporal changes, excluding ambient or monotonous classes. (3) Quality and Relevance: Both video and audio needed to be clean, stable, and semantically aligned." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": "Preliminary Curation We start from VGGSound [7], a large-scale dataset with 309 diverse audio classes. Similar to VGGSoundSync [6], we first narrow down to 149 classes with potentially clear audio-visual synchronized events, removing ambient classes, which is referred as VGGSS. We then deploy a sequence of automatic cleaning steps and a final manual selection step to identify appropriate videos. The procedures are summarized below and detailed in Suppl. Sec. 5.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Automatic Curation We first use PySceneDetect [4] to cut videos with sharp scene changes into different scenes, which are still likely to contain low-quality" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "short clips. To maximize usage, we split each scene into 3-second clips with 0.5-second strides, and discard unsuitable clips based on the following metrics:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 480, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 480, + 185 + ], + "type": "text", + "content": "Raw Pixel Difference: We calculate average pixel distances between consecutive frames and remove clips with both small and large values, likely depicting either static or videos with excessive motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 479, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 479, + 222 + ], + "type": "text", + "content": "Image Semantics Difference: Complementing the raw pixel analysis above, we also compute distances on CLIP [32] image features, removing videos with small semantics changes such as zoom in/out or large semantic content transitions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 223, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 223, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 131, + 223, + 479, + 246 + ], + "type": "text", + "content": "Audio Waveform Amplitude: We exclude clips whose maximum audio waveform amplitude is low, indicating weak audio cues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 479, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 282 + ], + "type": "text", + "content": "Semantic Alignment: We compute the average image-audio (IA) and image-text (IT) alignment scores [32] in a video using ImageBind [14], removing clips with low scores to ensure cross-modal semantic alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "text", + "content": "Audio-Video Synchronization: To measure audio-visual synchronization, we follow VGGSoundSync [6] to contrastively train an audio-visual synchronization classifier on VGGSS, ending up with a comparable " + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "inline_equation", + "content": "40.85\\%" + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "text", + "content": " test accuracy. The model outputs an unbounded av-sync score " + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "inline_equation", + "content": "\\phi_{\\pmb{a}_i,\\pmb{v}_j}" + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "text", + "content": " for an audio-video pair " + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "inline_equation", + "content": "(\\pmb{a}_i,\\pmb{v}_j)" + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "text", + "content": ". During training, we compute " + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "text", + "content": " for a synchronized pair " + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "inline_equation", + "content": "(\\pmb{a}_i,\\pmb{v}_i)" + }, + { + "bbox": [ + 130, + 284, + 480, + 367 + ], + "type": "text", + "content": " and its temporally-shifted pairs from the same instance. Contrastive loss is then applied to these shifted pairs to maximize the synchronization probability:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 235, + 378, + 480, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 378, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 235, + 378, + 480, + 404 + ], + "type": "interline_equation", + "content": "P _ {\\text {S y n c}} \\left(\\boldsymbol {a} _ {i}, \\boldsymbol {v} _ {i}\\right) = \\frac {\\exp \\left(\\phi_ {\\boldsymbol {a} _ {i} , \\boldsymbol {v} _ {i}} / \\tau\\right)}{\\sum_ {j} \\exp \\left(\\phi_ {\\boldsymbol {a} _ {i} , \\boldsymbol {v} _ {j}} / \\tau\\right)} \\tag {1}", + "image_path": "f6be922934b8544223293f2c71c034467ac111ce33041d7822f2c529b4e8e332.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "text", + "content": "to distinguish the synchronized pair from shifted ones. We use " + }, + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{sync}}" + }, + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "text", + "content": " as a synchronization indicator to remove low-scoring clips. When computing " + }, + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{sync}}" + }, + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "text", + "content": ", we discard the temperature parameter " + }, + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "text", + "content": " used to improve training efficacy. We detail the synchronization classifier and " + }, + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{sync}}" + }, + { + "bbox": [ + 130, + 414, + 479, + 462 + ], + "type": "text", + "content": " in Suppl. Sec. 1 and Sec. 2.1, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 471, + 479, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 479, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 479, + 530 + ], + "type": "text", + "content": "We empirically determine metrics' thresholds by prioritizing quality, acknowledging that some acceptable clips might be discarded. After automatic curation, we merge all continuous 3-second clips from each video and remove categories with less than 100 examples to avoid class imbalance, resulting in AVSync-AC (AVSync w/ Automatic Curation) with 76 categories and 39,902 examples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 532, + 479, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 479, + 592 + ], + "type": "text", + "content": "Manual Curation We further select 15 diverse categories with clear audiovisual motion cues from AVSync-AC for manual refinement. The categories range from animals and human actions to triggered tools and musical instruments. Manual curation once again seeks to identify appropriate videos for ASVA with the criteria above: high correlation, dynamic content, quality and relevance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": "Dataset Summary The final dataset, AVSync15, contains 90 training and 10 testing videos per category, each " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "2\\sim 10" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": " seconds long. We provide an overview of AVSync15 in Fig. 2a. To validate our curation pipeline, we randomly sample three 1500-video splits on the selected 15 categories from VGGSS and AVSync-AC, and quantitatively compare them with AVSync15 in Fig. 2b and Tab. 2b. We also compare AVSync15 with other audio-visual datasets in Suppl. Sec. 5.1." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Lin Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 257, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 257, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 257, + 126 + ], + "type": "text", + "content": "3.2 Evaluation Metrics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 479, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 479, + 218 + ], + "type": "text", + "content": "ASVA is a multi-faceted generation task, necessitating high quality at both image and video level. At the image level, we follow previous conventions [3,12] to use (1) Fréchet Inception Distance (FID) [16] to measure the quality of individual frames; (2) IA [14]/IT [47] to measure image-audio/image-text semantics alignment on CLIP [32] space. At the video level, we use Fréchet Video Distance (FVD) [42] to assess video quality. To measure audio-video synchronization, we compute the following metrics with the trained synchronization classifier:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 479, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 479, + 244 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 479, + 244 + ], + "type": "text", + "content": "RelSync During testing, we use the ground truth audio-visual pair " + }, + { + "bbox": [ + 130, + 220, + 479, + 244 + ], + "type": "inline_equation", + "content": "(\\pmb{a},\\pmb{v})" + }, + { + "bbox": [ + 130, + 220, + 479, + 244 + ], + "type": "text", + "content": " as a reference to measure synchronization of the generated video " + }, + { + "bbox": [ + 130, + 220, + 479, + 244 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}" + }, + { + "bbox": [ + 130, + 220, + 479, + 244 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 253, + 480, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 253, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 216, + 253, + 480, + 278 + ], + "type": "interline_equation", + "content": "\\operatorname {R e l S y n c} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v}) = \\frac {\\exp \\left(\\phi_ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}}}\\right)}{\\exp \\left(\\phi_ {\\boldsymbol {a} , \\boldsymbol {v}}\\right) + \\exp \\left(\\phi_ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}}}\\right)} \\tag {2}", + "image_path": "331f2c540b8a0d3e293b5c69b1198fd981ef24e6450e7ffd031a488fc4f46a9f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 286, + 479, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 479, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 479, + 345 + ], + "type": "text", + "content": "Note that while this reference-based metric normalizes the score by the synchronization of the reference pair " + }, + { + "bbox": [ + 130, + 286, + 479, + 345 + ], + "type": "inline_equation", + "content": "(\\pmb{a},\\pmb{v})" + }, + { + "bbox": [ + 130, + 286, + 479, + 345 + ], + "type": "text", + "content": ", the metric can still be sensitive to the quality of the reference pair. In fact, evaluating synchronization on a dataset where even ground-truth audios and videos are ambiguously synchronized is less informative of the model capabilities, e.g., Landscapes [22]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 346, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 479, + 407 + ], + "type": "text", + "content": "AlignSync The synchronization classifier is only trained on semantically-aligned and temporally-shifted audio-video pairs sampled from the same instance (See Suppl. Sec. 1.2). RelSync is thus implicitly conditioned on semantics alignment, i.e., P(Sync|Align). To jointly measure semantics alignment and synchronization, we first approximate " + }, + { + "bbox": [ + 130, + 346, + 479, + 407 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{Align}}" + }, + { + "bbox": [ + 130, + 346, + 479, + 407 + ], + "type": "text", + "content": " similarly as RelSync:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 186, + 416, + 480, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 416, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 186, + 416, + 480, + 444 + ], + "type": "interline_equation", + "content": "P _ {\\text {A l i g n}} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v} _ {1}) = \\frac {1}{b - 1} \\sum_ {i = 2.. b} \\frac {\\exp \\left(\\mathrm {I A} _ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}} _ {i}}\\right)}{\\exp \\left(\\mathrm {I A} _ {\\boldsymbol {a} , \\hat {\\boldsymbol {v}} _ {i}}\\right) + \\exp \\left(\\mathrm {I A} _ {\\boldsymbol {a} , \\boldsymbol {v} _ {1}}\\right)} \\tag {3}", + "image_path": "0a2402d0bedc4cb76ebe801a3a5f484cd805182bf5f7132ca8ea14409426293d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "inline_equation", + "content": "\\pmb{v}_1" + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "text", + "content": " is the first frame input for animation, and " + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "text", + "content": " is the number of generated frames. The generated first frame " + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}_1" + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "text", + "content": " is eliminated because it is often a replicate of input " + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "inline_equation", + "content": "\\pmb{v}_1" + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "text", + "content": ". By multiplying " + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{Align}}" + }, + { + "bbox": [ + 130, + 453, + 479, + 488 + ], + "type": "text", + "content": " with RelSync, we obtain the joint score:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 190, + 498, + 480, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 498, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 190, + 498, + 480, + 510 + ], + "type": "interline_equation", + "content": "\\operatorname {A l i g n S y n c} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v}) = P _ {\\text {A l i g n}} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v} _ {1}) \\cdot \\operatorname {R e l S y n c} (\\boldsymbol {a}, \\hat {\\boldsymbol {v}}, \\boldsymbol {v}) \\tag {4}", + "image_path": "0ea24485a4ebb9cb39676eddd73aa24ad7e6645a88a01ba62497f60cf9e42f43.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "type": "text", + "content": "which jointly measures semantic alignment and temporal synchronization between " + }, + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}" + }, + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 544, + 479, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 479, + 578 + ], + "type": "text", + "content": "These automated metrics are not always aligned with human preference. We therefore conduct a user study detailed in Sec. 5.2, asking human raters to compare videos generated by multiple models and select the best one." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 597, + 369, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 369, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 369, + 611 + ], + "type": "text", + "content": "4 Audio-Video Synchronized Diffusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 622, + 384, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 384, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 384, + 634 + ], + "type": "text", + "content": "4.1 Preliminary: Text-to-Image Latent Diffusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Text-to-image latent diffusion models (LDMs [34]) encode images " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " into a lower-dimensional latent space " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{z} = \\mathcal{E}(\\pmb{x})" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " using a pre-trained perceptual auto-encoder" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", and learn the conditional distribution " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{z}|\\boldsymbol{\\tau})" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " given a CLIP-encoded text prompt " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\tau}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ". It gradually denoises latents " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}^k" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", obtained by corrupting the image latent " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " with Gaussian noise " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", over " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " time steps. A denoising UNet " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " is deployed to estimate the added noise " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " by minimizing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 173, + 481, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 173, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 216, + 173, + 481, + 188 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L D M} = \\mathbb {E} _ {\\boldsymbol {z}, \\epsilon \\sim \\mathcal {N} (0, 1), k} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {z} ^ {k}, k, \\boldsymbol {\\tau}) \\| _ {2} ^ {2} \\right] \\tag {5}", + "image_path": "f0573843b72abc9e65dc978c5bfaeb38880dc1758d0f424fd0a8558d2803c1ae.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "content": "During inference, LDMs start from a random Gaussian noise map " + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\pmb{z}^{K}" + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "content": ", and iterate over " + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "content": " reverse diffusion steps [17, 38] to denoise the latents " + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\pmb{z}^{k-1} = \\pmb{z}^{k} - \\epsilon_{\\theta}(\\pmb{z}^{k}, k, \\pmb{\\tau})" + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "content": ", until " + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\pmb{z}^{0}" + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "content": " is obtained. LDMs then decode the latent into an image " + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{0} = \\mathcal{D}(\\pmb{z}^{0})" + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "content": " using the pre-trained decoder " + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "content": ". For simplicity, we refer to the images by their latents " + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "content": " rather than " + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 131, + 194, + 482, + 256 + ], + "type": "text", + "content": " throughout the rest of the paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 272, + 277, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 272, + 277, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 272, + 277, + 284 + ], + "type": "text", + "content": "4.2 Proposed Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "text", + "content": "In this work, we extend the capabilities of LDMs for ASVA, by focusing on learning video dynamics and temporal synchronization, unlike existing approaches [14, 39, 40] which primarily use audio to control global semantics. We propose the Audio-Video Synchronized Diffusion model (AVSyncD), which builds upon a pre-trained image LDM and integrates synchronized audio control and temporal layers for improved video consistency. The major component is a UNet " + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "text", + "content": " aiming to denoise a video instead of an image. The UNet is trained on synchronized audio-video pairs, with the first frame " + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "inline_equation", + "content": "z_{1}" + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "text", + "content": ", the corresponding audio " + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "text", + "content": ", and the CLIP encoded audio category name " + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "text", + "content": " as input conditions. The LDM denoising objective in Eq. (5) is applied to the remaining frames " + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "inline_equation", + "content": "z_{2:rT}" + }, + { + "bbox": [ + 130, + 291, + 482, + 446 + ], + "type": "text", + "content": " to be predicted. The architecture of AVSyncD is shown in Fig. 3 and discussed below. Further details, for example, highlighting the different attention modules used in the architecture are described in Suppl. Sec. 3.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "text", + "content": "Initial-frame Conditioning To condition LDM on an input image, we feed its latent " + }, + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1" + }, + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "text", + "content": " without noise into the UNet at every diffusion timestep " + }, + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "text", + "content": ". For all subsequent frames, we adhere to the original LDM, using independently sampled noisy latents " + }, + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{2:rT}^k" + }, + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "text", + "content": " as inputs and predicting the added noise " + }, + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "inline_equation", + "content": "\\epsilon^k" + }, + { + "bbox": [ + 130, + 448, + 482, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 545 + ], + "type": "text", + "content": "Text Conditioning We retain the text cross-attention layers in the original LDMs [34] (without finetuning) to condition the model on the audio category. However, due to the limited text diversity in training, class conditioning does not bring significant gains (see Suppl. Sec. 6.5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Audio Conditioning To facilitate audio-synchronized generation, we condition the generation on ImageBind audio embeddings [14]. ImageBind computes an audio classification token, " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{a}^g" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", representing global semantics, and local patch tokens, " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{f,t}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", across " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{a}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " timestamps. The original ImageBind only uses " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{a}^g" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " for contrastive learning and discards " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{f,t}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". However, we found these frozen local tokens as efficient synchronization cues. We split the patch tokens temporally into " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "rT" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " segments, corresponding to the same timestamps as the frames " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{1:rT}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " and append the global token to each. Each frame " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " then learns both semantics and synchronization guidance from its audio segment " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{t}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " via cross-attention [43]. In Suppl. Sec. 6.7, we compared ImageBind with CLAP and BEATs as encoders." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Lin Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 116, + 455, + 232 + ], + "blocks": [ + { + "bbox": [ + 160, + 116, + 455, + 232 + ], + "lines": [ + { + "bbox": [ + 160, + 116, + 455, + 232 + ], + "spans": [ + { + "bbox": [ + 160, + 116, + 455, + 232 + ], + "type": "image", + "image_path": "1bd66f98f8d4eae3fefeaba02b910f21dd921c8882d448d582879daf4a8ae176.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": "Fig. 3: AVSyncD overview. Left: We use ImageBind to encode audio into semantically aware time-dependent tokens " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "(\\pmb{a}_t)_{t=1}^{rT}" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": " and CLIP to encode the audio category into text embedding " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": ". In addition, the model receives the latent of the first frame " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\pmb{z}_1" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": ", and iteratively denoises noisy latents of the subsequent frames " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{2:rT}^k" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": " via reverse diffusion. The denoising UNet, based on LDMs [34], consists of a sequence of downsampling, bottleneck and up-sampling blocks, with structure detailed on the right. Right: Anatomy of a UNet block for frame " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\pmb{z}_t" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": ". LDM's original spatial conv, spatial attention and text cross attention layers are frozen, while its spatial self-attention layers are adjusted to first-frame spatial attentions, cross-attending to " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\pmb{z}_1" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": " instead. To learn video dynamics, we introduce temporal attention layers, and first-frame lookup temporal convolutions applied to input, output, and ResNet layers. We also train audio cross attentions for audio conditioning and synchronization. Trainable layers are marked with " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 397, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 480, + 422 + ], + "type": "text", + "content": "Spatial Convolution The original LDM's pre-trained spatial convolutional layers were frozen and used without modification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": "First-frame Temporal Convolution Each spatial convolution block was augmented with a 1D temporal convolution layer (kernel size 3) to capture temporal dependencies. To better adhere to the starting image, " + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1" + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": ", we adjusted the receptive field at frame " + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": " to include frames " + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "inline_equation", + "content": "(1, t - 1, t)" + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": " as opposed to " + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "inline_equation", + "content": "(t - 1, t, t + 1)" + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": ". These temporal convolutions with first-frame lookup were applied to three components in the UNet, namely the input/output conv layers and all ResNet convs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "content": "First-frame Spatial Attention We also leveraged the pre-trained LDM's spatial self-attention. Following [21], we modified the frozen spatial attention layers to cross-attend to the first frame rather than self-attend to the current frame, by computing the key-value pairs from the first frame and the queries from the current one." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": "Temporal Attention To effectively model long-range visual dependencies, we incorporated temporal attention layers [2]. Each frame index " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": " was converted into a sinusoidal positional embedding, added to the corresponding frame's latents after a learnable projection [43]. Each frame's local representation at spatial position " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "z_{hw}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": ", was then updated by attending to all frames at the same position " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "(z_{hw1},z_{hw2},\\ldots ,z_{hw(rT)})" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": " through a standard attention mechanism." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Classifier-free Audio Guidance Classifier-free guidance [18] is a technique used to control the impact of input prompts on the generated outputs. We extended it to amplify audio guidance for improved synchronization. To accomplish" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "this, we trained the model for both audio-conditioned and unconditioned generation, by randomly replacing " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " with a null audio embedding, " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\emptyset}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ", with a 20% likelihood. " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\emptyset}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " was computed by encoding an all-zero audio mel spectrogram via ImageBind. During inference, a factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\eta \\geq 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " controls the audio guidance:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 171, + 172, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 172, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 171, + 172, + 481, + 186 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {2: r T} ^ {k - 1} = (1 - \\eta) \\cdot \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {2: r T} ^ {k}, k; \\boldsymbol {z} _ {1}, \\boldsymbol {a} _ {\\emptyset}, \\boldsymbol {\\tau}\\right) + \\eta \\cdot \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {2: r T} ^ {k}, k; \\boldsymbol {z} _ {1}, \\boldsymbol {a}, \\boldsymbol {\\tau}\\right) \\tag {6}", + "image_path": "d863e4b9c9f07d07fc8bcfa5e2aa3f876ce4d7b7be4db56035bccfccaf17a295.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 481, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 194, + 481, + 218 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 194, + 481, + 218 + ], + "type": "text", + "content": " guides the denoising process towards latents congruent with audio-conditioned generation and away from those of audio-unconditional generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 230, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 230, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 230, + 248 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 240, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 240, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 240, + 272 + ], + "type": "text", + "content": "5.1 Implementation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 279, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 434 + ], + "type": "text", + "content": "Dataset We conducted experiments on three datasets. AVSync15: Our high-quality dataset curated from VGGSound [7], with 15 balanced categories, 1350 training videos, and 150 test videos. We also assessed our data curation pipeline by comparing it to models trained on Landscapes, TheGreatestHits, VGGSS and our AVSync-AC. Landscapes [22] is composed of 9 environmental sound classes. We followed the split in [35,48] with 900 clips for training and 100 for testing. Since Landscapes is full of ambient sounds without synchronized video motion, we mainly use it to evaluate visual quality. TheGreatestHits [29] is an audio-video dataset recording humans probing environments with a drumstick, with 733 videos for training and 244 for testing. The videos are synchronized with audio at the moments of impact but contain lots of static moments. It also offers limited diversity, featuring a singular motion of probing. VGGSS and our AVSync-AC were described in Sec. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "content": "Data Preprocess We sampled 2-second synchronized audio-video pairs for experiments. Videos were sampled at 6 fps with 12 frames, and resized to " + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "content": " on AVSync15/Landscapes or " + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "content": " on TheGreatestHits. Following Image-Bind [14], audios were sampled at " + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "inline_equation", + "content": "16\\mathrm{kHz}" + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "content": " and converted into 128-d spectrograms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 640 + ], + "type": "text", + "content": "Baselines We first adopted a simple Static baseline by repeating the input frame into a video, then compared it to several state-of-the-art works. (1) Semantic audio-to-video generation (CoDi [40]): Image, text, and audio are encoded into a shared CLIP [32] space and summed, then fused into a video diffusion model trained on large-scale datasets. (2) Image-to-video generation (VideoCrafter [5]): A superior video diffusion model however without audio inputs. It animates images by fusing CLIP-encoded image and text features into the model via modality-dependent cross-attention layers. (3) Synchronized audio-to-video generation (TPoS [20], AADiff [23], TempoToken [48]): Audio is encoded into time-dependent segments and fused into frozen text-to-image [34] or text-to-video [44] models. We re-implemented AADiff, and used TPoS and TempoToken's pretrained checkpoints on VGGSound and Landscapes. More details are provided in Suppl. Sec. 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Training & Evaluation We adopted the pretrained Stable Diffusion-V1.5 [34] as the diffusion model and ImageBind [14] as the audio encoder. All models" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Lin Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 482, + 262 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 482, + 262 + ], + "type": "image", + "image_path": "e8892f0690f70eb192e769bf36aa8c4a75a15945eab56609dfabd42aca9fd297.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 284, + 482, + 382 + ], + "blocks": [ + { + "bbox": [ + 132, + 266, + 480, + 283 + ], + "lines": [ + { + "bbox": [ + 132, + 266, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 480, + 283 + ], + "type": "text", + "content": "(a) Comparison on AVSync15. AVSyncD produces synchronized dynamic motions, such as muzzle flash, chicken and dog's head poses, and swirling in toilet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 284, + 482, + 382 + ], + "lines": [ + { + "bbox": [ + 134, + 284, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 134, + 284, + 482, + 382 + ], + "type": "image", + "image_path": "058c123d73accdbcd10bf911170d2915a113bd94be723b6cb8d3c6723a3fbc9c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 383, + 482, + 409 + ], + "lines": [ + { + "bbox": [ + 132, + 383, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 482, + 409 + ], + "type": "text", + "content": "(b) Comparison on Landscapes. Without synchronization (c) Generation by AVSyncD on The-cues in audio, AVSyncD still produces evolving visual ef- GreatestHits. AVSyncD generates hitting fecs and semantics, such as flames, waves, and lightning. actions at appropriate moments." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 418, + 398, + 429 + ], + "lines": [ + { + "bbox": [ + 214, + 418, + 398, + 429 + ], + "spans": [ + { + "bbox": [ + 214, + 418, + 398, + 429 + ], + "type": "text", + "content": "Fig. 4: Qualitative results on three datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 451, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 487 + ], + "type": "text", + "content": "were trained using Adam optimizer with a batch size of 64 and a learning rate of 0.0001. Besides metrics in Sec. 3.2, we also provide results using metric in [26] in Suppl. Sec. 3.4. We evaluated on 3 clips uniformly sampled from each video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 228, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 228, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 228, + 513 + ], + "type": "text", + "content": "5.2 Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Dataset Comparison In Tab. 1, the Static baseline is a good indicator of dataset attributes. It has similar IA/IT scores compared to Groundtruth, since Static is composed of a subset of groundtruth frames, which are obviously semantically aligned. FVD of Static on TheGreatestHits is relatively low, since the TheGreatestHits contains frequent static ground truth clips without any moving objects. More importantly, the RelSync and AlignSync of Static gradually increase from AVSync15, to TheGreatestHits, to Landscapes, with those on Landscapes even surpassing the ground truth. The fact that static videos perform well in terms of audio synchronization on TheGreatestHits and Landscapes datasets testify to the superiority of AVSync15 for audio-video synchronized generation. Model Comparison We compared to prior works on AVSync15 and Landscapes in Tabs. 1a and 1b and Figs. 4a and 4b. In Tab. 1a, CoDi achieves inferior" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 481, + 292 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative results. User study shows votes on 3 metrics: image quality, frame consistency, and synchronization. Inputs are combinations of image, text, audio." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 481, + 292 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 481, + 292 + ], + "type": "table", + "html": "
InputModelFID↓IA↑IT↑FVD↓AlignSync↑RelSync↑User Study
IQ↑FC↑Sync↑
T+ATPoS [20]13.523.3824.832671.019.5242.50---
TempoToken [48]12.218.8417.454466.419.7444.05---
I+TVideoCrafter [5]11.8-29.87840.721.2843.16382012
I2VD12.1-30.35398.221.8043.92629091
I+T+ACoDi [40]14.528.1523.421522.619.5441.51---
TPoS [20]11.938.3630.731227.819.6739.62---
AADiff [23]18.834.2328.97978.022.1145.483745
AVSyncD η = 112.138.3630.34382.722.2544.81---
AVSyncD η = 411.738.5330.45349.122.6245.52163186192
AVSyncD η = 811.737.9930.27420.722.7445.88---
Static-39.7630.391220.421.8343.66
Groundtruth-40.0630.31-25.0450.00
", + "image_path": "2549954ef17cff95afaeb680e73b54831e2fdcc177810270a119beaf4c856a75.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 313, + 481, + 427 + ], + "blocks": [ + { + "bbox": [ + 251, + 294, + 361, + 303 + ], + "lines": [ + { + "bbox": [ + 251, + 294, + 361, + 303 + ], + "spans": [ + { + "bbox": [ + 251, + 294, + 361, + 303 + ], + "type": "text", + "content": "(a) Performance on AVSync15." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 313, + 481, + 427 + ], + "lines": [ + { + "bbox": [ + 135, + 313, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 135, + 313, + 481, + 427 + ], + "type": "table", + "html": "
InputModelFID↓IA↑IT↑FVD↓AignSync↑RelSync↑
T+ATPoS [20]16.515.6126.702081.323.1248.15
TempoToken [48]16.422.5822.872480.024.2148.65
I+TI2VD16.7-22.56539.524.7449.89
I+T+ACoDi [40]20.522.6324.23982.922.6345.48
TPoS [20]16.223.5223.20789.623.5147.05
AADiff [23]70.722.0722.921186.326.7753.93
AVSyncD η = 116.522.2922.81463.124.8149.96
AVSyncD η = 416.222.4922.79415.224.8249.93
Static-23.6022.211177.525.7951.59
Groundtruth-23.6522.08-25.0150.00
", + "image_path": "f28e1f39d3679bb3be5ca79af3806a72aa6a2a5e56c6927b453ab4ad7e974bd5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 135, + 448, + 481, + 517 + ], + "blocks": [ + { + "bbox": [ + 249, + 430, + 364, + 438 + ], + "lines": [ + { + "bbox": [ + 249, + 430, + 364, + 438 + ], + "spans": [ + { + "bbox": [ + 249, + 430, + 364, + 438 + ], + "type": "text", + "content": "(b) Performance on Landscapes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 448, + 481, + 517 + ], + "lines": [ + { + "bbox": [ + 135, + 448, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 135, + 448, + 481, + 517 + ], + "type": "table", + "html": "
InputModelFID↓IA↑IT↑FVD↓AignSync↑RelSync↑
I+TI2VD9.1-13.42425.022.0544.58
I+T+AAVSyncD η = 19.011.8513.18313.522.5945.52
AVSyncD η = 48.712.0713.31249.322.8345.95
Static-13.3316.56348.924.3648.73
Groundtruth-13.5216.49-25.0250.00
", + "image_path": "a41fb580d1678c1a68134eb885e2b6773368fadbd3d12be2e7e87af093cc1b25.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 240, + 518, + 373, + 528 + ], + "lines": [ + { + "bbox": [ + 240, + 518, + 373, + 528 + ], + "spans": [ + { + "bbox": [ + 240, + 518, + 373, + 528 + ], + "type": "text", + "content": "(c) Performance on TheGreatestHits." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "results on almost all metrics. TPoS(I+T+A) shows strong image quality (FID), but is worse in video quality (FVD) and synchronization (RelSync). TempoToken, on the other hand, is better at synchronization rather than visual quality, likely due to the lack of image input. AADiff is competitive on synchronization but extremely bad on image quality (FID). This is expected as AADiff adjusts each frame using audio amplitude, producing visual changes that highly correlate to audio changes temporally but may be overwhelmed by noises, as shown in Fig. 4a. On Landscapes, its similar FVD to Static but abnormally higher Align-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Lin Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 135, + 302, + 182 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 2: Effect of (a) first-frame lookups (b) data curation, evaluated on AVSync15." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 135, + 302, + 182 + ], + "lines": [ + { + "bbox": [ + 135, + 135, + 302, + 182 + ], + "spans": [ + { + "bbox": [ + 135, + 135, + 302, + 182 + ], + "type": "table", + "html": "
FF-ConvFF-AttnFID↓FVD↓AignSync↑
✗✗11.8383.322.19
✓✗11.6347.422.24
✓✓11.8325.622.33
", + "image_path": "f86012a6a6f27ff20057e81e52d610c43f5acaf7d67b915b379a21a969852442.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 314, + 136, + 482, + 182 + ], + "blocks": [ + { + "bbox": [ + 211, + 183, + 223, + 193 + ], + "lines": [ + { + "bbox": [ + 211, + 183, + 223, + 193 + ], + "spans": [ + { + "bbox": [ + 211, + 183, + 223, + 193 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 136, + 482, + 182 + ], + "lines": [ + { + "bbox": [ + 314, + 136, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 314, + 136, + 482, + 182 + ], + "type": "table", + "html": "
DatasetACMCFID↓FVD↓AignSync↑
VGGSS✗✗12.91307.921.50
AVSync-AC✓✗12.0428.822.09
AVSync15✓✓11.8325.622.33
", + "image_path": "e3092b275de156e555f075753a3fbea48a5d43265a28cc46bcfeefbac52db02f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 391, + 182, + 403, + 191 + ], + "lines": [ + { + "bbox": [ + 391, + 182, + 403, + 191 + ], + "spans": [ + { + "bbox": [ + 391, + 182, + 403, + 191 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 223, + 480, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 480, + 295 + ], + "type": "text", + "content": "Sync and RelSync also suggest that it only applied minor modifications to the input image due to lack of sound changes, as shown in Fig. 4b. Without audio input, VideoCrafter performs poorly on synchronization. It also has difficulty faithfully adhering to the input frame, as in Fig. 4a. AVSyncD achieves the best animation results on almost all metrics. On Landscapes, AVSyncD also performs the best on FID and FVD, with other scores being similar to ground truth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "text", + "content": "User Study We invited 15 participants to compare 4 animation models with top overall performance (VideoCrafter, AADiff, I2VD, AVSyncD) on AVSync15, based on 3 metrics in Tab. 1. The 4 models generated videos conditioned on the same test examples (audio+image). Each test example was independently evaluated by 2 participants to select their most preferred generation (vote) on each metric. In total, we evaluated all 150 test examples on AVSync15, collecting " + }, + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "inline_equation", + "content": "150 \\times 2 = 300" + }, + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "text", + "content": " votes on each metric. Tab. 1a shows votes each model received." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 395, + 246, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 395, + 246, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 395, + 246, + 406 + ], + "type": "text", + "content": "5.3 Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 413, + 480, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 461 + ], + "type": "text", + "content": "Audio Conditioning In Tab. 1, AVSyncD outperforms I2VD, especially on AlignSync and RelSync. AVSyncD did not improve RelSync on Landscapes probably due to the lack of synchronization cues on the dataset itself. These results show that audio condition enhances generation quality and synchronization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 462, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 545 + ], + "type": "text", + "content": "Audio Guidance Tab. 1 shows increasing the audio guidance factor " + }, + { + "bbox": [ + 130, + 462, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 462, + 482, + 545 + ], + "type": "text", + "content": " from 1 to 4 improves FID, IA, and FVD on all three datasets. As expected, audio guidance also improved AlignSync and RelSync significantly on AVSync15 and TheGreatestHits, but not on the less synchronized Landscapes dataset. Prior works [20, 23, 39] claimed that increasing audio amplitude can also lead to stronger visual effects. We compare this approach with our audio guidance in Fig. 5a. Audio guidance offers a better control mechanism than audio amplitude." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 546, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 571 + ], + "type": "text", + "content": "First-frame (FF) Lookups We validated FF Lookups by replacing them with standard temporal convolutions or spatial self-attention in Tab. 2a." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 571, + 481, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 481, + 596 + ], + "type": "text", + "content": "Data Curation We compared to AVSyncD trained on random subsets from VGSS and AVSync-AC with equal data scale and balanced categories in Tab. 2b." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 611, + 304, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 304, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 304, + 623 + ], + "type": "text", + "content": "5.4 Applications and Extensions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "Animate Generated Images When lacking an image as input, we can use existing image generators to generate the image, which AVSyncD can also animate. Fig. 5b shows animations on images generated by StableDiffusion-V1.5 [34]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 228, + 202 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 228, + 202 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 228, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 228, + 202 + ], + "type": "image", + "image_path": "741ecc577b7ccee586195bae66c66765ce4e24d40ce91146ec224f7e52f4b2e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 203, + 233, + 213 + ], + "lines": [ + { + "bbox": [ + 221, + 203, + 233, + 213 + ], + "spans": [ + { + "bbox": [ + 221, + 203, + 233, + 213 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 155, + 248, + 179 + ], + "blocks": [ + { + "bbox": [ + 223, + 155, + 248, + 179 + ], + "lines": [ + { + "bbox": [ + 223, + 155, + 248, + 179 + ], + "spans": [ + { + "bbox": [ + 223, + 155, + 248, + 179 + ], + "type": "image", + "image_path": "e79388744e8b04af778dbfc981cc23a4e57ed62dea088be901836af36782d789.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 248, + 118, + 320, + 202 + ], + "blocks": [ + { + "bbox": [ + 248, + 118, + 320, + 202 + ], + "lines": [ + { + "bbox": [ + 248, + 118, + 320, + 202 + ], + "spans": [ + { + "bbox": [ + 248, + 118, + 320, + 202 + ], + "type": "image", + "image_path": "7d333ad9eafefe2bee8a92966f501b4cd6878d9632bca21c62c8fbc102aa6af0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 327, + 123, + 402, + 202 + ], + "blocks": [ + { + "bbox": [ + 327, + 123, + 402, + 202 + ], + "lines": [ + { + "bbox": [ + 327, + 123, + 402, + 202 + ], + "spans": [ + { + "bbox": [ + 327, + 123, + 402, + 202 + ], + "type": "image", + "image_path": "c75d518219ade30adf5f7469acb6af97fb6f94098dfc792964082184db382563.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 204, + 370, + 213 + ], + "lines": [ + { + "bbox": [ + 357, + 204, + 370, + 213 + ], + "spans": [ + { + "bbox": [ + 357, + 204, + 370, + 213 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 406, + 124, + 480, + 202 + ], + "blocks": [ + { + "bbox": [ + 406, + 124, + 480, + 202 + ], + "lines": [ + { + "bbox": [ + 406, + 124, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 406, + 124, + 480, + 202 + ], + "type": "image", + "image_path": "fe35110bccbdfb0f7473420051d4dd5a5eeded50301c2e6f19106de9be753876.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 203, + 449, + 213 + ], + "lines": [ + { + "bbox": [ + 437, + 203, + 449, + 213 + ], + "spans": [ + { + "bbox": [ + 437, + 203, + 449, + 213 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 269, + 480, + 319 + ], + "blocks": [ + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "lines": [ + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "type": "text", + "content": "Fig. 5: (a): Effects of audio amplitude vs. classifier-free audio guidance. top: original audio with " + }, + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\eta = 1" + }, + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "type": "text", + "content": "; mid: " + }, + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "type": "inline_equation", + "content": "100\\times" + }, + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "type": "text", + "content": " amplified audio with " + }, + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\eta = 1" + }, + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "type": "text", + "content": "; bottom: original audio with " + }, + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\eta = 8" + }, + { + "bbox": [ + 131, + 221, + 482, + 256 + ], + "type": "text", + "content": ". (b): animate generated images. (c): Animation with internet images and audios." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 269, + 480, + 319 + ], + "lines": [ + { + "bbox": [ + 133, + 269, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 269, + 480, + 319 + ], + "type": "image", + "image_path": "54d348134abf2bec79da71e3be43ee2388ec590b4f4893083c0ca03158166b72.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 321, + 398, + 331 + ], + "lines": [ + { + "bbox": [ + 216, + 321, + 398, + 331 + ], + "spans": [ + { + "bbox": [ + 216, + 321, + 398, + 331 + ], + "type": "text", + "content": "(a) Controllable animation with (un)related audiios." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 133, + 333, + 480, + 375 + ], + "blocks": [ + { + "bbox": [ + 133, + 333, + 480, + 375 + ], + "lines": [ + { + "bbox": [ + 133, + 333, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 133, + 333, + 480, + 375 + ], + "type": "image", + "image_path": "09101c7e7b476f76d84dcf78fe06a7244022083f73c686fb67c9f7e246831bf2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 376, + 411, + 386 + ], + "lines": [ + { + "bbox": [ + 202, + 376, + 411, + 386 + ], + "spans": [ + { + "bbox": [ + 202, + 376, + 411, + 386 + ], + "type": "text", + "content": "(b)Animate target objects on internet images with audios." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 396, + 463, + 406 + ], + "lines": [ + { + "bbox": [ + 148, + 396, + 463, + 406 + ], + "spans": [ + { + "bbox": [ + 148, + 396, + 463, + 406 + ], + "type": "text", + "content": "Fig. 6: Controllable image animation with audiios. Key frames are visualized." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 430, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 453 + ], + "type": "text", + "content": "Animate Contents from Internet AVSyncD can also generalize well to unseen images and audio, as shown in Fig. 5c." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 455, + 481, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 481, + 502 + ], + "type": "text", + "content": "Control Animations with (Un)Related Audios We can control the motion of an image to follow desired audio, e.g., animate a baby to not only cry but also bark like a dog or roar like a lion, as seen in Fig. 6a. When there is no object related to the audio, the animations do not demonstrate corresponding motion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 503, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 481, + 540 + ], + "type": "text", + "content": "Animate Target Objects with Audios When multiple objects exist in the image, a scenario not existing in training data, we can still use audios to only animate the related target object, as shown in Fig. 6b." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 557, + 220, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 220, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 220, + 570 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "We tackled the under-explored Audio-Synchronized Visual Animation task, with an emphasis on generating videos with audio-synchronized dynamics. We contributed the high-quality AVSync15 benchmark via careful data curation and proposed the AVSyncD model to animate images with realistic motions. Due to the scale of AVSync15, our work cannot generalize to all audio classes in the world, which requires several orders of magnitude larger datasets. However, we hope our research inspires further work in this direction." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Lin Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "text", + "content": "3. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 408, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 408, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 408, + 239 + ], + "type": "text", + "content": "4. Castellano, B.: Pyscenedetect. https://www.scenedetect.com/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 239, + 480, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 271 + ], + "type": "text", + "content": "5. Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., Weng, C., Shan, Y.: Videocrafter1: Open diffusion models for high-quality video generation (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "text", + "content": "6. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronization in the wild. In: Proceedings of the British Machine Vision Conference (BMVC) (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "text", + "content": "7. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset. In: ICASSP (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "text", + "content": "8. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Wei, F.: Beats: Audio pre-training with acoustic tokenizers. In: ICML (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 348, + 433, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 433, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 433, + 358 + ], + "type": "text", + "content": "9. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: ACCV (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "text", + "content": "10. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV Workshop (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "text", + "content": "1. Chung, S.W., Chung, J.S., Kang, H.G.: Perfect match: Improved cross-modal embeddings for audio-visual synchronisation. In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "text", + "content": "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "text", + "content": "3. Gemmeke, J.F., Ellis, D.P.W., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: Proc. IEEE ICASSP 2017 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 468, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 491 + ], + "type": "text", + "content": "4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "type": "text", + "content": "5. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "6. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "8. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop on Deep Generative Models and Downstream Applications (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 589, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 622 + ], + "type": "text", + "content": "9. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. In: British Machine Vision Conference (BMVC) (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "20. Jeong, Y., Ryoo, W., Lee, S., Seo, D., Byeon, W., Kim, S., Kim, J.: The power of sound (tpos): Audio reactive video generation with stable diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7822-7832 (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "21. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In: ICCV (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "22. Lee, S.H., Oh, G., Byeon, W., Bae, J., Kim, C., Ryoo, W.J., Yoon, S.H., Kim, J., Kim, S.: Sound-guided semantic video generation. In: ECCV (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "text", + "content": "23. Lee, S., Kong, C., Jeon, D., Kwak, N.: Aadiff: Audio-aligned video synthesis with text-to-image diffusion. In: CVPR Workshop on Content Generation (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "text", + "content": "24. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 235 + ], + "type": "text", + "content": "25. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "type": "text", + "content": "26. Luo, S., Yan, C., Hu, C., Zhao, H.: Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In: NeurIPS (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 256, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 480, + 277 + ], + "type": "text", + "content": "27. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "type": "text", + "content": "28. Ng, E., Romero, J., Bagautdinov, T., Bai, S., Darrell, T., Kanazawa, A., Richard, A.: From audio to photoreal embodiment: Synthesizing humans in conversations. In: ArXiv (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 310, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 331 + ], + "type": "text", + "content": "29. Owens, A., Isola, P., McDermott, J., Torralba, A., Adelson, E.H., Freeman, W.T.: Visually indicated sounds. In: CVPR (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 331, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 364 + ], + "type": "text", + "content": "30. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: AAAI Conference on Artificial Intelligence (AAAI) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "31. Pedro Morgado, Ishan Misra, N.V.: Robust audio-visual instance discrimination. In: Computer Vision and Pattern Recognition (CVPR), IEEE/CVF Conf. on (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 396, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 428 + ], + "type": "text", + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "type": "text", + "content": "33. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: arXiv (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "text", + "content": "34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 471, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 503 + ], + "type": "text", + "content": "35. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N.J., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 503, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 480, + 557 + ], + "type": "text", + "content": "36. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "text", + "content": "37. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 601 + ], + "type": "text", + "content": "38. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "39. Sung-Bin, K., Senocak, A., Ha, H., Owens, A., Oh, T.H.: Sound to visual scene generation by audio-to-visual latent alignment. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "40. Tang, Z., Yang, Z., Zhu, C., Zeng, M., Bansal, M.: Any-to-any generation via composable diffusion. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=2EDqbSCnmF" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Lin Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 490 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "41. Tsuchida, S., Fukayama, S., Hamasaki, M., Goto, M.: Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In: Proceedings of the 20th International Society for Music Information Retrieval Conference, ISMIR 2019. Delft, Netherlands (Nov 2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "42. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. In: arXiv (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "44. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "45. Wang, W., Xie, k., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "46. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern by few-shot tuning a text-to-image diffusion model. arXiv preprint arXiv:2310.10769 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "text", + "content": "47. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: EMNLP (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "48. Yariv, G., Gat, I., Benaim, S., Wolf, L., Schwartz, I., Adi, Y.: Diverse and aligned audio-to-video generation via text-to-video model adaptation (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": "49. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. In: ICLR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "50. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "51. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: AAAI Conference on Artificial Intelligence (AAAI) (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "text", + "content": "52. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-Synchronized Visual Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_content_list.json b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..75ea707b3c234baa881dcc4e5fd359489e00b6c2 --- /dev/null +++ b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_content_list.json @@ -0,0 +1,1737 @@ +[ + { + "type": "text", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "text_level": 1, + "bbox": [ + 256, + 140, + 750, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dogucan Yaman1, Fevziye Irem Eyiokur1, Leonard Bärmann1, Hazim Kemal Ekenel2, and Alexander Waibel1,3", + "bbox": [ + 259, + 210, + 743, + 243 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Karlsruhe Institute of Technology, Karlsruhe, Germany", + "2 Istanbul Technical University, Istanbul, Turkey", + "$^{3}$ Carnegie Mellon University, Pittsburg PA, USA dogucan.yaman@kit.edu" + ], + "bbox": [ + 308, + 253, + 692, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Talking face generation aims to create realistic videos with accurate lip synchronization and high visual quality, using given audio and reference video while preserving identity and visual characteristics. In this paper, we start by identifying several issues with existing synchronization learning methods. These involve unstable training, lip synchronization, and visual quality issues caused by lip-sync loss, SyncNet, and lip leaking from the identity reference. To address these issues, we first tackle the lip leaking problem by introducing a silent-lip generator, which changes the lips of the identity reference to alleviate leakage. We then introduce stabilized synchronization loss and AVSyncNet to overcome problems caused by lip-sync loss and SyncNet. Experiments show that our model outperforms state-of-the-art methods in both visual quality and lip synchronization. Comprehensive ablation studies further validate our individual contributions and their cohesive effects.", + "bbox": [ + 261, + 340, + 740, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Talking face generation $\\cdot$ Lip synchronization $\\cdot$ Lip leaking", + "bbox": [ + 261, + 547, + 740, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 584, + 375, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Audio-driven talking face generation aims at generating a video with respect to given face and audio sequences. The objective is to achieve synchronized lip movements corresponding to the provided audio while preserving the identity and visual details. This task has recently attracted significant attention due to its versatile applications, including dubbing in the film industry, online education, enhancing video conferencing, and dubbing for various types of videos [71, 75].", + "bbox": [ + 212, + 613, + 785, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The talking face generation task comprises two primary aspects: (1) lip synchronization and (2) visual quality of the face. Since lips that are out-of-sync with audio can be easily identified by humans, synchronized lips are key to achieving natural and realistic talking-face generation. For this, the primary solution is to evaluate audio-lip synchronization and use it as a training objective. Wav2Lip [46] introduced an improved version of SyncNet [14], a pretrained model designed to measure audio-visual synchronization. This model is also used during training to extract features and compute lip-sync loss. Subsequently, many approaches have employed improved SyncNet [46] to guide the", + "bbox": [ + 212, + 705, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6647a73177df405c2a63197b23f09cfd59134fa1c42614fc06ded5f83ba822a4.jpg", + "image_caption": [ + "(a) SyncNet [46]" + ], + "image_footnote": [], + "bbox": [ + 228, + 148, + 403, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e2d05c640144d1b3a3edfb6edb5388406a3a9afbb650b2fb588f3d46d64aee2.jpg", + "image_caption": [ + "(b) AVSyncNet (ours)", + "Fig. 1: (a, b) Cosine similarity between GT audio-lip pairs on random LRS2 samples, showcasing the instability of SyncNet and more robust performance of AVSyncNet. (c) illustrates full mouth region / lip leaking from the reference, pose effect from the reference, and similar identity reference-target image scenarios." + ], + "image_footnote": [], + "bbox": [ + 424, + 148, + 599, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8847030a5549f6e1f2bba0af662de421fec6c445d6b2393eea2c472792327752.jpg", + "image_caption": [ + "(c) Leaking problems" + ], + "image_footnote": [], + "bbox": [ + 617, + 145, + 779, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model throughout training. The main concept is to provide audio-video pairs to SyncNet, which extracts features through its image and audio encoders. Subsequently, cosine similarity with binary cross-entropy loss is measured between these two features [46]. Since the model was trained in this manner, a higher cosine similarity is expected when the lips are synchronized with the given audio. Although existing approaches with this strategy mostly surpass other methods in lip synchronization, there are still challenges that must be addressed to enhance performance.", + "bbox": [ + 212, + 369, + 787, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we identify two main challenges in existing approaches that restrict models from achieving satisfactory performance in both lip synchronization and visual quality: SyncNet instabilities and lip leaking. First, in alignment with previous research [41], we observe instabilities in the performance of SyncNet [46] on pairs of ground-truth (GT) lip and audio (see Fig. 1a and App. A). Thus, when SyncNet is utilized as part of talking face generation training, it might provide an inadequate training signal, assigning low similarity scores to generated images even when the lips are synchronized. This problem causes unstable training, degrading the lip generation capability of the network, thereby ending up in out-of-sync lips or suboptimal lip-sync. Furthermore, the lip-sync loss [46] and reconstruction losses are conflicting [39]. This leads models to have either poor lip synchronization or degraded visual quality (sometimes even both), escalating further when lip-sync loss and SyncNet are employed with high-resolution (HR) data [39,73]. To address these problems, we first improve SyncNet and introduce AVSyncNet, demonstrating a more robust performance (see Fig. 1b) and also overcoming poor shift-invariance characteristics of SyncNet (see Fig. 6a). However, despite improved performance, the instability problem is not fully solved (see Fig. 1b). To overcome this problem further, we also introduce a stabilized synchronization loss. Specifically, instead of directly using the similarity of the (generated lips, audio) pair, we calculate the difference of the similarities between (GT lips, audio) and (generated lips, audio). We hypothesize that this alleviates the described problems since it guides the model to generate a lip movement with a similar synchronization score as for the GT face. Together with AVSync-", + "bbox": [ + 212, + 493, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Net, this method empirically enhances the lip synchronization performance as well as avoids visual quality issues caused by misleading lip-sync loss.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Second, we address another main problem in the talking face generation literature: lip leaking. The current gold standard in 2D-based methods is to input a bottom-half masked face image (\"pose reference\"), wherein the model is expected to generate the input face with proper lip movements. However, because of masking, the model requires a reference image to retain the identity and texture in the masked part. This is achieved by randomly choosing a face image from a different part of the video sequence, referred to as \"identity reference\". However, this introduces a new challenge: As it is randomly selected, the lip movements of the identity reference can frequently be similar to the GT lips during training (see last row of Fig. 1c). Hence, for faster convergence, the model tends to replicate the lip movements from the identity reference, resulting in poor lip synchronization or complete out-of-sync output. By following the literature [41,46], we term this phenomenon as lip leaking. To tackle this problem, we propose a simple yet effective technique: silent-lip generator. The concept involves modifying the identity reference to make the lips closed (thus \"silent-lip\") before feeding it to the talking face generator network. This method ensures consistently closed lips in the identity reference, effectively mitigating the lip leaking problem. Our contributions are as follows: (i) We identify and analyze various fundamental problems that harm lip synchronization learning and also cause visual quality issues. (ii) We present a robust and shift-invariant AVSyncNet, and stabilized synchronization loss to overcome the problems caused by lip-sync loss and SyncNet. (iii) We present a silent-lip generator to generate an identity reference with closed lips before feeding the talking face generator to alleviate the lip leaking.", + "bbox": [ + 212, + 178, + 787, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 549, + 387, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-driven Talking Face Generation. Initial studies mapped audio features to time-aligned facial motions [69] or predicted facial motions by HMM [6]. In [57], videos were generated by finding the images most aligned with the audio. ATVGNet [9] transfers audio to facial landmarks and uses pixel-wise loss with an attention mechanism to avoid the jittering problem and leaking of irrelevant speech. [16] and [80] use facial landmark representation to synthesize faces with synchronized lips. Recent papers address the task as conditional inpainting by masking the bottom half of the input face and feeding the network an identity reference from another time step of the same video, along with the audio segment. Wav2Lip [46] proposes SyncNet and lip-sync loss to predict lip synchronization and achieves significant performance. PC-AVS [78] introduces a pose-controllable audio-visual system, while GC-AVT [36], EAMM [29] and EVP [30] control the emotion by utilizing an emotion embedding. On the other hand, SyncTalkFace [45] uses Audio-Lip Memory to store lip motion features and retrieves them as visual hints for better synchronization. VideoReTalking [10] proposes to manipulate the reference image to have a face with canonical expression to alleviate the sensitivity of the model against the identity reference image.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, we introduce a silent-lip generator to implicitly learn to manipulate the lips of the identity reference to mitigate the lip leaking problem. Compared to the method proposed in [10], our method works more efficiently to preserve the identity and visual quality. More recently, LipFormer [64] uses a pre-learned facial codebook to generate HR videos, while DINet [73] proposes to use a deformation module to obtain deformed features to enhance lip synchronization and head pose alignment. IPLAP [76] shows satisfactory visual quality and stability by using intermediate landmark representation and motion field. Recently, TalkLip [63] introduces a global audio encoder, trained with self-supervised learning, to encode features by considering the entire content of the audio. Besides, they propose to use lip reading during the training as well as in the evaluation to control whether the content is preserved. SIDGAN [39] performs important analyzes, and introduces shift-invariant APS-SyncNet and training objectives along with the coarse-to-fine pyramid model for HR dubbed video generation. Recent works use diffusion model because of its stability and accuracy [50, 55]. Finally, in [61], talking face generation is used as a part of a full system to perform end-to-end face dubbing, involving speech recognition, translation, speech, and video generation. In contrast to the above, 3D-based and Neural Radiance Fields-based (NeRFs) methods typically generate the entire head, rather than just manipulating 2D images of the face, often involving manipulation of pose, emotion, and 3D face model [4, 5, 21, 37, 44, 49, 53, 58, 59, 62, 66–68, 70, 72, 74, 77, 80]. Similarly, portrait animation aims at utilizing a single input image to generate a video by predicting the pose and expression of the subsequent frames, along with ensuring synchronized lips. Nevertheless, these tasks significantly differ from face dubbing in terms of methodology, task definition, goals, and real-world applications.", + "bbox": [ + 212, + 146, + 787, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lip Synchronization. Some earlier works [23, 52] used hand-crafted features and statistical models to evaluate lip synchronization. Recent studies proposed to use mutual information between audio-visual features to produce sync or out-of-sync output for sound [8,26,43] or speech [2,12,15,32,33]. While some methods learn lip synchronization implicitly [9,21,28,34,57,66,74], other methods employ distance between landmarks or facial parameters [30,44,53,80]. Contrarily, the majority of works [17,20,36,45,46,54,56,60,62-64,67,73,77,78] extract audio-visual features with an additional network (mostly SyncNet [46]) for audio-lip synchronization prediction. Afterwards, while some of these methods utilize lip-sync loss [46], others benefit from contrastive learning by using infoNCE [42]. We follow a similar strategy by utilizing an additional network to calculate a synchronization loss. We propose a robust and shift-invariant AVSyncNet along with a stabilized synchronization loss to overcome existing challenges.", + "bbox": [ + 212, + 542, + 789, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Approach", + "text_level": 1, + "bbox": [ + 215, + 758, + 444, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Talking Face Generation", + "text_level": 1, + "bbox": [ + 215, + 787, + 464, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose an audio-driven talking face generation model $G_{L}$ with enhanced lip synchronization. As shown in Fig. 2a, our model incorporates: 1) an audio", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7bddec5e1b8bf354bf2c1facb6932189a4d173659ff17d8b7ae18c7a6091a6b3.jpg", + "image_caption": [ + "(a) Talking face generation model", + "(b) FD block" + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 663, + 309 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc4063384e5965c490218073f5107120f76fbb60a62779176997d73f71f710fe.jpg", + "image_caption": [ + "Fig. 2: Talking face generation model $G_{L}$ (a) and face-decoding (FD) block (b). Our model receives a pose reference sequence, mel-spectrogram of an audio snippet, and a silent identity reference, that is generated by our silent-lip generator $G_{S}$ , aiming to alleviate lip leaking problem. The model then synthesizes the talking face sequence to ensure lip synchronization. Subsequently, the employed loss functions are computed." + ], + "image_footnote": [], + "bbox": [ + 673, + 176, + 785, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encoder for processing the audio snippet, 2) an identity encoder for addressing an identity reference image, and 3) a pose encoder for utilizing a pose reference.", + "bbox": [ + 212, + 439, + 784, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Audio Encoder. The audio encoder $E_{A}$ generates phoneme-level embeddings, serving as conditions for the face generator to generate lip movements accurately. Our audio encoder extracts embeddings $F_{A} = E_{A}(A) \\in \\mathbb{R}^{1 \\times 1 \\times 512}$ from the given mel-spectrogram $A$ , representing the driving audio. In contrast to existing methods [10, 46, 73, 76], we propose using a pretrained, frozen audio encoder, concurrently trained with a face encoder to learn lip synchronization similar to the objective in SyncNet [46]. Thus, we can obtain improved embeddings during the talking face generation training by leveraging the capacity of the pretrained robust audio encoder. We obtain the best score with such frozen audio encoder.", + "bbox": [ + 212, + 496, + 787, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face Encoder. In accordance with the gold standard in the literature, we utilize an identity reference, $I^R$ , and a pose reference, $I$ , as inputs to the model. The identity reference is a face image of the subject that provides identity information. It is different from the pose reference and randomly selected from the same video. The pose reference is identical to the target image, except for the bottom half, which is masked, as the model is designed to focus on generating lip movements. Unlike most conventional methods that employ a joint encoder for processing identity and pose references, we use individual encoders to allow each encoder to focus solely on their respective tasks [39]. Therefore, we utilize two parallel CNN-based face encoders to process identity and pose references individually. This approach yields better feature representation and ultimately leads to improved performance.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "bbox": [ + 225, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a4914333acf4b1932424f458c6b84e04a0bbe41ccf0067ae0283a8ac6e78a5a9.jpg", + "image_caption": [ + "(a) Silent-lip generation model $G_{S}$" + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 480, + 299 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eeb47c49fc2ae167e741c0a489896d3cdd2383635a7473f2246d9b316b85104b.jpg", + "image_caption": [ + "(b) AVSyncNet", + "Fig. 3: $G_{S}$ in inference (a) and AVSyncNet training pipeline (b)." + ], + "image_footnote": [], + "bbox": [ + 552, + 162, + 790, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Silent-Lip Generator", + "text_level": 1, + "bbox": [ + 215, + 372, + 433, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Talking face generation involves using an identity reference to preserve the identity in the generated image. This is particularly important because the bottom half of the pose reference, specifically the mouth region, is masked. This masking is necessary as we aim to synthesize appropriate lip movements corresponding to the provided audio. However, models are unintentionally affected by the lip movement of the identity reference, rather than solely gathering identity information (see Fig. 1c and App. C for details). This behavior, which we refer to as lip leaking, leads to poor lip synchronization or occasionally even non-converged training. We hypothesize that there are two main reasons for this: First, the lip movement of the identity reference may occasionally resemble that of the lips in the target image (see last row of Fig. 1c). Hence, the model can lower the synchronization loss more quickly by undesirably replicating the lip movements of the identity reference. Second, the diversity of lip movements in the identity reference may yield the model to seek a correlation with the target lips. This causes a challenging disentanglement task for the identity encoder —namely distinguishing identity information and lip movement.", + "bbox": [ + 212, + 400, + 787, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To mitigate the aforementioned issue, we propose to use an additional model, called silent-lip generator $G_{S}$ , prior to the talking face generation model $G_{L}$ , aimed at modifying the lip shape of the identity reference. Specifically, we reconstruct the input face with closed, flat lips. This strategy reduces the likelihood of having lips similar to the target face in the identity reference and resolves the issue of diverse lip movements. Consequently, the model no longer replicates the lips from the identity reference, resulting in stable training & improved lip sync.", + "bbox": [ + 212, + 643, + 787, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To implement $G_{S}$ , we structure the task and the model similar to the talking face generation. Specifically, we input a bottom-half masked pose reference, an identity reference, and an audio snippet to train a GAN [18] for generating a talking face. The model is trained to reconstruct the face using both the pose and identity references. Notably, we exclude any synchronization loss and SyncNet during training and train the model under the weak condition. Consequently, the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model focuses solely on generating lip movements without synchronization when speech is present. Therefore, it implicitly learns to generate closed lips when the input audio is silent as shown in Fig. 3a. Furthermore, by eliminating synchronization loss, we overcome unstable training and the issue of lip leaking from the identity reference for this network. We choose this approach for its efficient utilization of the same model and data. Given the scarcity of closed-lip faces in the dataset, we avoid using these frames directly to maintain effective training and generalization. Please note that we use the same architecture for $G_{S}$ and $G_{L}$ . We initially train $G_{S}$ separately on the same training data (LRS2) and then incorporate it into the training of $G_{L}$ without further finetuning. Specifically, we then only pass silent audio to $G_{S}$ , so that it modifies given identity references to have silent lips for subsequent use as identity reference for $G_{L}$ .", + "bbox": [ + 212, + 146, + 787, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Video Generation", + "text_level": 1, + "bbox": [ + 214, + 347, + 410, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the talking face video generation, we employ the aforementioned components. Initially, we employ our pretrained silent-lip generator $G_{S}$ to synthesize the identity reference with closed lips. Subsequently, we input this silent identity reference to the identity encoder and the pose reference to the pose encoder in the talking face generation model $G_{L}$ . Similarly, we provide the mel-spectrogram of the corresponding audio snippet to the audio encoder. Next, we concatenate the embeddings from the identity and pose encoders, along with the depth dimension, and pass it through a $1 \\times 1$ convolution layer to reduce the depth. Finally, we concatenate this feature representation with the audio embeddings before feeding them into the face generator. The face generator generates the entire face with accurate lip movements by preserving the identity and the pose.", + "bbox": [ + 212, + 368, + 787, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We illustrate the talking face generation model $G_{L}$ in Fig. 2a. We integrate the U-Net architecture [47] for our overall design, leveraging its adequate performance in reconstruction tasks while ensuring computational efficiency. Our identity encoder and pose encoder share the same architecture, consisting of consecutive face-encoding (FE) blocks. Each block has a strided-convolutional layer followed by two non-strided convolutional layers, each paired with a batch normalization layer [27] and a ReLU activation function [35,40]. We also use the residual connection strategy [22] by summing the input and output of each block before forwarding it to the next layer. On the other hand, our face generator has consecutive face-decoding (FD) blocks. As shown in Fig. 2b, within each block, we utilize a transposed-convolutional layer, followed by two convolutional layers incorporating batch normalization and ReLU activation function. Moreover, we apply a skip connection between the reciprocal layers of the face encoders and the decoder to retain high-level features and enhance the training stability.", + "bbox": [ + 212, + 534, + 787, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GAN Loss. To train our model, we utilize GAN loss [18] and employ a discriminator, which is a straightforward CNN-based binary classification network to distinguish real and fake samples, designed with a balanced architecture aligned with our face encoders. We benefit from consecutive strided-convolutional layers followed by the Leaky ReLu activation function and spectral normalization [38].", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "bbox": [ + 225, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction Loss. We employ L1 loss in pixel space $L_{pixel} = ||I' - I^{GT}||_1$ between the generated $I'$ and the target faces $I^{GT}$ to ensure consistency in areas outside the lips and maintain the illumination condition. We further utilize perceptual loss [31] based on the pretrained VGG-19 [51]:", + "bbox": [ + 212, + 145, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p e r} = \\sum_ {i = 1} ^ {5} c _ {i} | | \\mathrm {V G G} ^ {\\phi_ {i}} \\left(I ^ {\\prime}\\right) - \\mathrm {V G G} ^ {\\phi_ {i}} \\left(I ^ {G T}\\right) | | _ {1} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 219, + 785, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $c_{i}$ are weight coefficients from [31], and $\\phi$ indicates the set of VGG layers.", + "bbox": [ + 215, + 270, + 784, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive Triplet Loss. Although our goal is to capture visual details from the identity reference, we observe that the model occasionally focuses on the visual details (e.g., illumination, pose) in the identity reference excessively and this tendency could potentially degrade the quality and stability of the generated face sequence, resulting in suboptimal performance. To tackle this, we exploit a triplet loss strategy [48], aiming to minimize the distance between the generated face and GT, while maximizing the distance between the generated face and the identity reference. However, the random selection of identity references increases the probability of choosing an image that closely resembles the GT. This scenario poses a challenge for the vanilla triplet loss, potentially degrading training and resulting in poor visual and pose quality. To mitigate this, we introduce an adaptive triplet loss that considers the similarity between the identity reference and GT during loss computation to alter its effect. The formula is as follows:", + "bbox": [ + 212, + 309, + 787, + 507 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a t} = \\left[ D (\\mathrm {V G G} (I ^ {\\prime}), \\mathrm {V G G} (I ^ {G T})) - \\frac {D (\\mathrm {V G G} (I ^ {\\prime}) , \\mathrm {V G G} (I ^ {R}))}{D (\\mathrm {V G G} (I ^ {G T}) , \\mathrm {V G G} (I ^ {R}))} + \\alpha \\right] _ {+} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 518, + 785, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $[\\cdot]_{+} = \\max(\\cdot, \\epsilon)$ , $D$ represents the L2 distance, and we empirically choose $\\alpha = 1$ . In this loss, we leverage the ratio of the similarity between the generated image and identity reference to that of the GT and the identity reference to adjust the loss value. As the identity reference becomes more similar to GT, the impact of the distance between the generated image and identity reference on the loss diminishes, since expecting a high distance in this case is not reasonable. Since our objective is to incorporate visual details from the identity reference, we opt for a very low coefficient to avoid conflicting with the primary goal.", + "bbox": [ + 212, + 563, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Learning Synchronization", + "text_level": 1, + "bbox": [ + 215, + 708, + 473, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The lip-sync loss [46] serves as a method to calculate synchronization between audio and video. Leveraging the pretrained SyncNet [46] for feature extraction from both audio and video inputs demonstrates reasonable performance in learning lip synchronization. However, the SyncNet is significantly unstable when measuring this similarity. Our evaluation using SyncNet on GT training data reveals notable fluctuations in cosine similarity between video and audio, contrary to the expected high scores (see Fig. 1a and App. A for details). Therefore, this", + "bbox": [ + 212, + 733, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "provides conflicting information to the system, resulting in poor lip synchronization, unstable training, and degraded visual quality. To tackle this problem, we present a more accurate, shift-invariant, and robust version of SyncNet, named AVSyncNet, and introduce a novel, stabilized synchronization loss.", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "AVSyncNet. We employ a ResNet-50-based [22] image encoder, known for its superior performance in face recognition, alongside a ResNetSE-34-based audio encoder [11], which is a modified version of ResNet-34 [22] designed to handle spectrogram inputs. This design makes the AVSyncNet model robust and shift-invariant compared to SyncNet [46]. We train our model on the LRS2 training data [1], calculating the cosine similarity between audio and lip features, followed by a binary cross-entropy loss, as shown in Fig. 3b. During each training step, we provide a set of images (5 images) along with the corresponding audio. For negative samples, we randomly select an audio snippet from the non-overlapping part of the video. Please note that as we feed the bottom half of the face to the image encoder, we adapt the first layer of ResNet-50 for an input size of $112 \\times 224$ .", + "bbox": [ + 212, + 226, + 787, + 395 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Stabilized Synchronization Loss $(L_{ss})$ . Although AVSyncNet shows improved performance compared to SyncNet [46] and alleviates existing instability problem that harms lip-sync and visual quality, the unstable performance is not fully solved due to the inherent challenges of the task (see Fig. 1b and App. A, B). Therefore, we introduce a stabilized synchronization loss $(L_{ss})$ to improve the lip synchronization performance further in conjunction with AVSyncNet by providing more stable and precise supervision. The formula is shown below:", + "bbox": [ + 212, + 412, + 787, + 521 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s s} = - \\log \\left(1 - \\frac {\\left| x - y \\right| + \\epsilon}{\\left| x - y \\right| + \\left| y - d \\right| + \\epsilon}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 527, + 785, + 561 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nx = \\operatorname {A V S I M} \\left(I ^ {\\prime}, A\\right), \\quad y = \\operatorname {A V S I M} \\left(I ^ {G T}, A\\right), \\quad d = \\operatorname {A V S I M} \\left(I ^ {R}, A\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 584, + 785, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $I^{\\prime}, I^{GT}$ , and $I^R$ are generated, GT, and identity reference lips, respectively, while $A$ is their corresponding audio. AVSIM( $I, A$ ) indicates the audio-visual similarity between a face image (bottom half only, i.e. lips) and audio, given by the cosine similarity of extracted image and audio features $\\phi_{AVS}^V(I)$ , $\\phi_{AVS}^A(A)$ of the respective AVSyncNet encoder.", + "bbox": [ + 212, + 611, + 782, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this formulation, $x$ followed by cross-entropy loss denotes the lip-sync loss [46]. However, to address the unstable and fluctuating performance, we utilize the relative distance in similarity between GT lips-audio and generated-audio pairs. Sometimes, randomly selected reference images may have a lip", + "bbox": [ + 212, + 686, + 787, + 750 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "bbox": [ + 225, + 113, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "This is reminiscent of distillation loss [25], as the actual value of the scores is neglected, and only their difference provides loss value for the training. However, initial experiments trying to directly minimize the distance between SyncNet (or AVSyncNet) image encoder features of generated and GT faces showed poor performance. We hypothesize that this is caused by SyncNet image features being only meaningful for comparing with corresponding audio features due to its training strategy.", + "bbox": [ + 217, + 753, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "movement similar to that of the target image. We have already introduced the silent-lip generator $G_{S}$ to mitigate this situation. However, AVSyncNet, while less severe than SyncNet, might still be unstable (e.g., unexpectedly high scores with incorrect pairs or vice versa), resulting in a minor lip leaking and stability issue. This problem arises when AVSyncNet assigns an erroneously high score to silent lip-audio pairs as well as when the target lip shape looks similar to closed lips. To mitigate this, we inject the similarity score between the identity reference and audio into the formulation. Specifically, we penalize the model more when the identity reference-audio pair shows higher similarity.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.5 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 306, + 457, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Combining all the presented contributions, the total loss is:", + "bbox": [ + 214, + 332, + 645, + 348 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {G A N} (G, D) + \\lambda_ {1} L _ {p i x e l} (G) + \\lambda_ {2} L _ {p e r} (G) + \\lambda_ {3} L _ {s s} (G) + \\lambda_ {4} L _ {a t} (G) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 361, + 785, + 378 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $G$ and $D$ indicate generator and discriminator outputs, respectively. We empirically found the best coefficients as $(\\lambda_1, \\dots, \\lambda_4) = (10, 1, 2, 0.5)$ . Please note that this is the loss function for training $G_L$ , while for $G_S$ we set $\\lambda_3 = \\lambda_4 = 0$ .", + "bbox": [ + 212, + 388, + 787, + 434 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We process videos by using 5 consecutive frames in each step to consider the temporal information. We detect faces with FAN [7], followed by acquiring tight crops and resizing to $96 \\times 96$ , as faces in LRS2 [1] are of low resolution. Our audio encoder receives a mel-spectrogram of size $16 \\times 80$ derived from $16\\mathrm{kHz}$ audio with a window size of 800 and a hop size of 200. We employ the Adam optimizer with $(\\beta_{1},\\beta_{2}) = (0.5,0.999)$ and set the learning rate to $1 \\times 10^{-4}$ for all models. Training our AVSyncNet is done similarly to SyncNet [46] on the LRS2 dataset, and then we freeze the audio encoder of AVSyncNet and use it in the training of $G_{S}$ and $G_{L}$ . At the end of talking face generation, we apply a post-processing step by using VQFR [19] to enhance the visual quality and the resolution, aiming to achieve HR videos. We train and test our models with a single NVIDIA RTX A6000 GPU.", + "bbox": [ + 212, + 435, + 789, + 617 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 214, + 640, + 460, + 657 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. We trained our silent-lip generator and talking face generator using Lip Reading Sentence 2 (LRS2) [1] training set as it is a well-known benchmark with extensive subject diversity. The evaluation was carried out on the LRS2 test set and extended to the LRW [13] test set and HTDF dataset [74] to demonstrate performance on unseen data.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics and Baseline. For visual quality, we employ widely used metrics: FID [24], SSIM [65], and PSNR. We also use inter-frame consistency (IFC), presented as a training objective in [79]. This is achieved by calculating the difference between the distances of the consecutive frames in the generated and GT videos. To evaluate lip synchronization, we follow the literature and use Landmark Distance (LMD) [9] in the mouth region and LSE-C & LSE-D metrics [46]", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/0d2282647bd1263b228bca740f4fa7f886bce9086ebc94ef7cd4195311061fbf.jpg", + "table_caption": [ + "Table 1: Quantitative results on the test sets of LRS2 and LRW." + ], + "table_footnote": [], + "table_body": "
LRS2LRW
MethodSSIM ↑PSNR ↑FID ↓IFC ↓LMD ↓LSE-C ↑LSE-D ↓SSIM ↑PSNR ↑FID ↓IFC ↓LMD ↓LSE-C ↑LSE-D ↓
Wav2Lip [46]0.8626.537.050.212.387.596.750.8525.146.810.202.147.496.51
PC-AVS [78]0.7328.2418.400.461.936.417.520.8132.2514.270.381.426.537.15
EAMM [29]0.6921.0184.650.513.543.319.930.7126.2244.160.482.614.329.04
VideoReTalking w/ FR [10]0.8425.589.280.222.617.496.820.8727.115.300.232.396.597.12
DINet [73]0.7824.354.260.252.305.378.370.8827.508.170.221.965.249.09
TalkLip [63]0.8626.114.940.242.348.536.080.8626.3415.730.261.837.286.48
IPLAP [76]0.8729.674.100.202.116.497.160.9130.458.400.211.645.947.76
Ours w/o FR0.9532.643.830.161.138.416.030.9231.454.460.181.227.866.24
Ours w/ FR (VQFR)0.9031.805.230.271.368.525.830.9030.217.050.211.417.926.00
", + "bbox": [ + 215, + 171, + 785, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "to measure the confidence and distance scores through a pretrained model [14]. We choose SOTA methods with publicly available codes and models to compare them fairly under the same conditions, as the implementation of the metrics and face cropping strategy before computing the metrics affect the scores.", + "bbox": [ + 212, + 309, + 784, + 369 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 391, + 431, + 406 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Tab. 1, we present quantitative results on test sets of two benchmark datasets, namely LRS2 and LRW. We achieve state-of-the-art results in all visual quality metrics excluding PSNR on LRW, which is a less informative metric compared to SSIM and FID. On IFC, we similarly outperform all compared methods, indicating that our model generates the most consistent videos in terms of temporal information and stability of the faces.", + "bbox": [ + 212, + 416, + 784, + 505 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In lip synchronization evaluation, we obtain the best performance in LMD. Nevertheless, it is important to highlight that LMD is sensitive to the changes in the image, as it does not disentangle synchronization and visual stability. For instance, affine transformations impact the LMD score even when the lips are synchronized, and vice versa. On the LRW dataset, we achieve SOTA results with more reliable confidence and distance metrics for lip synchronization: LSE-C & D. On the LRS2 dataset, TalkLip yields a slightly better score than our model with the LSE-C metric. Nevertheless, we outperform TalkLip and achieve a SOTA result with the LSE-D metric. All these results indicate the accuracy of our method in terms of visual quality and lip synchronization. Similarly, we achieve SOTA results for most metrics on unseen HDTF [74], see App. E.1.", + "bbox": [ + 212, + 507, + 785, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 694, + 419, + 709 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig. 4, we demonstrate a qualitative comparison with SOTA models and GT data. We use their respective publicly available models and generate videos from the HDTF dataset [74] to compare the models on unseen data since the presented models were trained on the LRS2 dataset, except for DINet (trained with HDTF). Due to SyncNet and lip-sync loss, TalkLip and Wav2Lip encounter generalization issues, sometimes leading to visual artifacts in the mouth region or face boundaries, especially when the pose of the identity reference differs from the pose reference, despite generating accurate lip movements. This observation", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f13553b11fc2a0811d7ce4e16d8f88304a730a7dc93c5a653549df43b3c96dbf.jpg", + "image_caption": [ + "VideoReTalking (Cheng et al. 2022)", + "DINet (Zhang et al. 2023)", + "TalkLip (Wang et al. 2023)", + "IP-LAP (Zhong et al. 2023)", + "Ours w/o FR", + "Ground Truth", + "Fig. 4: Qualitative comparison with the SOTA methods. Reference videos (from HDTF [74]) are randomly selected and not seen during training by our model. For more images and videos, please check App. E, F and https://yamand16.github.io/TalkingFaceGeneration/." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 433, + 376 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7ad44190c5c745ee88774d909f76ede959f038ddd976d407a9ad2f6304327dad.jpg", + "image_caption": [ + "Wav2Lip (Prajwal et al. 2020)" + ], + "image_footnote": [], + "bbox": [ + 568, + 143, + 785, + 376 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/25c7b450e8d51346fc90fb54de77d8f08fc3ebab9c236bb83278df2105a35526.jpg", + "table_caption": [ + "Table 2: Ablation studies on the LRS2 test set. See the text for details." + ], + "table_footnote": [], + "table_body": "
AblationSetupMethodPSNR ↑ SSIM ↑ FID ↓ LMD ↓ LSE-C ↑ LSE-D ↓ IFC ↓
ComponentsAGL+LSa26.3490.85312.252.4087.116 ± 1.927.396 ± 1.03
BA + Ea,S26.6140.8689.822.3257.271 ± 1.767.106 ± 0.98
CA + Ea,W26.5900.86910.562.2787.220 ± 1.757.158 ± 0.99
DB + GS27.1800.8728.161.7417.752 ± 1.716.413 ± 0.95
EGL + Ea,S + GS + LSs31.1660.9255.271.1408.370 ± 1.166.032 ± 0.59
FE + Lt30.6580.9176.241.2508.260 ± 1.346.176 ± 0.64
GE + Lat32.7550.9494.021.1358.382 ± 1.166.057 ± 0.61
HG w / AVSyncNet32.6400.9523.831.1308.410 ± 0.976.037 ± 0.55
Post-processingFR1Setup H + GPEN28.9910.91958.771.1977.6256.457
FR2Setup H + GFPGAN31.1690.91613.071.2197.6246.496
FR3Setup H + VQFR: full model31.8060.9055.231.3658.5285.838
Silent face generationVRT-SVideoReTalking silent data22.1240.64633.60--0.463
Ours-SOur silent data (GS)33.3280.9514.41--0.141
", + "bbox": [ + 272, + 484, + 730, + 602 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "clearly validates the motivation of our contributions. In contrast, our model generates consistent faces with comparably fewer artifacts, featuring appropriate lip movements that align with both the GT faces and the corresponding audio. However, VideoReTalking demonstrates comparable lip synchronization and visual quality performance to our model. On the other hand, IP-LAP shows sufficient visual quality, while less accurate lip synchronization.", + "bbox": [ + 212, + 628, + 787, + 720 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 739, + 390, + 757 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 2 and Fig. 5a show a comprehensive ablation study on the LRS2 test set, analyzing the individual impact of our contributions. We first train our $G_{L}$ model using SyncNet [46] and lip-sync loss [46] as a baseline. As expected, we encountered several issues with unstable training. Once our model converged after several random seeds, the results (Setup A) show that lip-sync loss can achieve", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/186ab939a02d824c0535dc5009730c3f298013352f616b8d6517c19652aa79e2.jpg", + "image_caption": [ + "(a) Components." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 447, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c56851318c7dd8c76bee7484ffb313fc82bf129adfee63911ef80a4eb9e03faf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 205, + 447, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3ea110fe614b3cb5d7616ff8a89b4325ab1a3455cb1d39242400b6ba5f6374fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 243, + 446, + 279 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1055a1553975ccbadddd88742894d815f9c8cb82185b62118ce28e857b118079.jpg", + "image_caption": [ + "(c) Generated samples for challenging cases: Pose, teeth, beard." + ], + "image_footnote": [], + "bbox": [ + 217, + 281, + 446, + 318 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/54419d10d011a30a52a30eda9f7adb39c169b01e7c7808035f8dc405e9a74aae.jpg", + "image_caption": [ + "(b) Face restoration methods." + ], + "image_footnote": [], + "bbox": [ + 511, + 145, + 759, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8c8874f5699470fd9854a1d10024b1607b80b9ca6d9ac031266f60e8f73b5b8d.jpg", + "image_caption": [ + "(d) Silent face comparison with VRT [10]. © European Central Bank (CC BY).", + "Fig. 5: Ablation studies of components (a), face restoration methods (b), and silent face generation (d). (c) demonstrates generated images in challenging cases." + ], + "image_footnote": [], + "bbox": [ + 511, + 205, + 784, + 318 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "decent synchronization performance, despite lower visual quality and training stability issues. In this setup, we train the audio encoder as a part of $G_{L}$ . Replacing this with our pretrained audio encoder enhances the synchronization and visual quality (Setup B). For further comparison, we also utilize the audio encoder of the Wav2Vec2 [3], presented in Setup C. However, it slightly decreases the scores, which validates our hypothesis about training the audio encoder for synchronization purposes. Thus, we continue with Setup B and add our silent-lip generator $G_{S}$ to generate silent identity references (Setup D). $G_{S}$ alleviates the lip leaking problem, makes the training more stable, and improves the scores noticeably. We further replace lip-sync loss with our stabilized synchronization loss, yielding drastically improved synchronization and visual quality scores (Setup E). Moreover, we observed that Setup E shows almost no instability in the training. In Setup F, we train Setup E including vanilla triplet loss and it enhances neither visual quality nor synchronization; in fact, it even causes detrimental effects. This shows the necessity of modifying the triplet loss and introducing the adaptive triplet loss. In Setup G, replacing vanilla triplet loss with adaptive triplet loss demonstrates a slight improvement in visual quality, while not having a negative impact on lip synchronization. In Setup H, we switch SyncNet with AVSyncNet and achieve slightly better visual quality and lip synchronization.", + "bbox": [ + 212, + 421, + 787, + 709 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We compare $G_{S}$ with VRT [10] silent face generation approach in Tab. 2. Our model surpasses VRT quantitatively and qualitatively (see Fig. 5d), preserving the visual details and identity while modifying the lips.", + "bbox": [ + 212, + 713, + 787, + 760 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We compare different face restoration methods as post processing and present the results in Tab. 2 and Fig. 5b. VQFR surpasses other methods in preserving lip synchronization as well as FID and PSNR. However, GPEN shows better performance in the remaining metrics (see App. F.2 for details). In summary, we employ VQFR for post-processing in our full model.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "bbox": [ + 225, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d386668efa85fb07cc9df1121e5b27b15feca8a88f5a440f1b773ab58667a0e9.jpg", + "image_caption": [ + "(a) Shifting [39].", + "Fig. 6: Graphs show the performance of SyncNet [46] and our AVSyncNet on LRS2 GT audio-lip pairs under certain transformations. See App. A for further details. We apply lip-sync loss [46] for the analyses to fairly compare SyncNet and our AVSyncNet independent from our loss function." + ], + "image_footnote": [], + "bbox": [ + 241, + 152, + 380, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5bc8583d456ce149bb0e7de200d22a85fa0f94408843731876bb88053aa0f174.jpg", + "image_caption": [ + "(b) Rotation." + ], + "image_footnote": [], + "bbox": [ + 398, + 152, + 539, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9305dd95a910a0bb9e933c9f260f615520d191e8bd05942eedca21ada26ab331.jpg", + "image_caption": [ + "(c) Further analysis." + ], + "image_footnote": [], + "bbox": [ + 560, + 147, + 743, + 234 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "AVSyncNet. Comprehensive experiments on LRS2 GT data pairs highlight AVSyncNet's superior performance (see Fig. 6). Furthermore, it demonstrates strong shift-invariance and robustness against affine transformations in the data due to AVSyncNet's design, particularly emphasizing its effectiveness in focusing on lip synchronization while being less affected by other factors. Moreover, AVSyncNet's performance is not affected by face pose unlike SyncNet [46]. Tab. 2 also demonstrates that our AVSyncNet improves the performance of our talking face generation model and works harmoniously with the proposed $L_{ss}$ .", + "bbox": [ + 212, + 344, + 787, + 467 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 489, + 359, + 506 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we improve audio-driven talking face generation by identifying problems in current approaches and mitigating them accordingly. Specifically, we introduce a silent-lip generator to mitigate lip leaking, which is a common problem that harms lip-sync and training stability. Furthermore, we propose stabilized synchronization loss along with AVSyncNet, which significantly improves the training stability, lip synchronization performance, and visual quality by solving the problems caused by lip-sync loss and SyncNet. Experimental results on benchmark datasets and a comprehensive ablation study show the merit of our method and contributions. Moreover, our detailed analyses reveal the main issues, support our claims, and validate proposed contributions.", + "bbox": [ + 212, + 522, + 787, + 674 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Despite the notable improvements, SyncNet's and AVSyncNet's unstable nature should be investigated further. Moreover, face restoration sometimes causes inconsistencies in the video. Silent-lip generator makes teeth invisible in identity references, occasionally resulting in suboptimal teeth generation.", + "bbox": [ + 212, + 695, + 787, + 758 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ethics & Social Impact. We believe that generating lip-synchronized faces holds significant benefits across a broad spectrum of applications. However, we acknowledge its vulnerability to potential misuse, particularly deepfake generation. We will utilize Watermarking and prevent uncontrolled usage of our model.", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the European Commission Project Meetween (101135798) under the call HORIZON-CL4-2023-HUMAN-01-03.", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 220, + 321, + 236 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Afouras, T., Chung, J.S., Senior, A., Vinyals, O., Zisserman, A.: Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence 44(12), 8717-8727 (2018)", + "2. Afouras, T., Owens, A., Chung, J.S., Zisserman, A.: Self-supervised learning of audio-visual objects from video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16. pp. 208–224. Springer (2020)", + "3. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020)", + "4. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 187-194 (1999)", + "5. Booth, J., Roussos, A., Ponniah, A., Dunaway, D., Zafeiriou, S.: Large scale 3d morphable models. International Journal of Computer Vision 126(2), 233-254 (2018)", + "6. Brand, M.: Voice puppetry. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 21-28 (1999)", + "7. Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In: International Conference on Computer Vision (2017)", + "8. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronisation in the wild. arXiv preprint arXiv:2112.04432 (2021)", + "9. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7832-7841 (2019)", + "0. Cheng, K., Cun, X., Zhang, Y., Xia, M., Yin, F., Zhu, M., Wang, X., Wang, J., Wang, N.: Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "1. Chung, J.S., Huh, J., Mun, S., Lee, M., Heo, H.S., Choe, S., Ham, C., Jung, S., Lee, B.J., Han, I.: In defence of metric learning for speaker recognition. arXiv preprint arXiv:2003.11982 (2020)", + "2. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017)", + "3. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017)", + "4. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 251-263. Springer (2017)" + ], + "bbox": [ + 225, + 257, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Chung, S.W., Chung, J.S., Kang, H.G.: Perfect match: Improved cross-modal embeddings for audio-visual synchronisation. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3965-3969. IEEE (2019)", + "16. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 408-424. Springer (2020)", + "17. Eskimez, S.E., Maddox, R.K., Xu, C., Duan, Z.: End-to-end generation of talking faces from noisy speech. In: ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 1948-1952. IEEE (2020)", + "18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)", + "19. Gu, Y., Wang, X., Xie, L., Dong, C., Li, G., Shan, Y., Cheng, M.M.: Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII. pp. 126–143. Springer (2022)", + "20. Guan, J., Zhang, Z., Zhou, H., Hu, T., Wang, K., He, D., Feng, H., Liu, J., Ding, E., Liu, Z., et al.: Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1505-1515 (2023)", + "21. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5784-5794 (2021)", + "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "23. Hershey, J., Movellan, J.: Audio vision: Using audio-visual synchrony to locate sounds. Advances in neural information processing systems 12 (1999)", + "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "25. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "26. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. arXiv preprint arXiv:2210.07055 (2022)", + "27. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015)", + "28. Jamaludin, A., Chung, J.S., Zisserman, A.: You said that?: Synthesising talking faces from audio. International Journal of Computer Vision 127, 1767-1779 (2019)", + "29. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)", + "30. Ji, X., Zhou, H., Wang, K., Wu, W., Loy, C.C., Cao, X., Xu, F.: Audio-driven emotional video portraits. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14080-14089 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)", + "32. Kadandale, V.S., Montesinos, J.F., Haro, G.: Vocalist: An audio-visual synchronisation model for lips and voices. arXiv preprint arXiv:2204.02090 (2022)", + "33. Kim, Y.J., Heo, H.S., Chung, S.W., Lee, B.J.: End-to-end lip synchronisation based on pattern classification. In: 2021 IEEE Spoken Language Technology Workshop (SLT). pp. 598-605. IEEE (2021)", + "34. KR, P., Mukhopadhyay, R., Philip, J., Jha, A., Namboodiri, V., Jawahar, C.: Towards automatic face-to-face translation. In: Proceedings of the 27th ACM international conference on multimedia. pp. 1428-1436 (2019)", + "35. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)", + "36. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3387-3396 (2022)", + "37. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 106-125. Springer (2022)", + "38. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018)", + "39. Muaz, U., Jang, W., Tripathi, R., Mani, S., Ouyang, W., Gadde, R.T., Gecer, B., Elizondo, S., Madad, R., Nair, N.: Sidgan: High-resolution dubbed video generation via shift-invariant learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7833-7842 (2023)", + "40. Nair, V., Hinton, G.E.: Rectified linear units improve restricted boltzmann machines. In: Proceedings of the 27th international conference on machine learning (ICML-10). pp. 807-814 (2010)", + "41. Nayak, S., Schuler, C., Saha, D., Baumann, T.: A deep dive into neural synchrony evaluation for audio-visual translation. In: Proceedings of the 2022 International Conference on Multimodal Interaction. pp. 642-647 (2022)", + "42. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "43. Owens, A., Efros, A.A.: Audio-visual scene analysis with self-supervised multisensory features. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 631-648 (2018)", + "44. Papantoniou, F.P., Filntisis, P.P., Maragos, P., Roussos, A.: Neural emotion director: Speech-preserving semantic control of facial expressions in\" in-the-wild\" videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18781-18790 (2022)", + "45. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2062-2070 (2022)", + "46. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020)" + ], + "bbox": [ + 215, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "48. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)", + "49. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII. pp. 666-682. Springer (2022)", + "50. Shen, S., Zhao, W., Meng, Z., Li, W., Zhu, Z., Zhou, J., Lu, J.: Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1982-1991 (2023)", + "51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "52. Slaney, M., Covell, M.: Facesync: A linear operator for measuring synchronization of video facial images and audio tracks. Advances in neural information processing systems 13 (2000)", + "53. Song, L., Wu, W., Qian, C., He, R., Loy, C.C.: Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security 17, 585-598 (2022)", + "54. Song, Y., Zhu, J., Li, D., Wang, X., Qi, H.: Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786 (2018)", + "55. Stypulkowski, M., Vougioukas, K., He, S., Zieba, M., Petridis, S., Pantic, M.: Diffused heads: Diffusion models beat gans on talking-face generation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5091-5100 (2024)", + "56. Sun, Y., Zhou, H., Wang, K., Wu, Q., Hong, Z., Liu, J., Ding, E., Wang, J., Liu, Z., Hideki, K.: Masked lip-sync prediction by audio-visual contextual exploitation in transformers. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1–9 (2022)", + "57. Suwajanakorn, S., Seitz, S.M., Kemelmacher-Shlizerman, I.: Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)", + "58. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022)", + "59. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 716-731. Springer (2020)", + "60. Vougioukas, K., Petridis, S., Pantic, M.: Realistic speech-driven facial animation with gans. International Journal of Computer Vision 128, 1398-1413 (2020)", + "61. Waibel, A., Behr, M., Yaman, D., Eyiokur, F.I., Nguyen, T.N., Mullov, C., Demirtas, M.A., Kantarci, A., Constantin, S., Ekenel, H.K.: Face-dubbing++: Lipsynchronous, voice preserving translation of videos. In: 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). pp. 1-5. IEEE (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Wang, D., Deng, Y., Yin, Z., Shum, H.Y., Wang, B.: Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17979-17989 (2023)", + "63. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14653-14662 (2023)", + "64. Wang, J., Zhao, K., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Lipformer: High-fidelity and generalizable talking face generation with a pre-learned facial codebook. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13844-13853 (2023)", + "65. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "66. Wu, H., Jia, J., Wang, H., Dou, Y., Duan, C., Deng, Q.: Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1478-1486 (2021)", + "67. Yao, S., Zhong, R., Yan, Y., Zhai, G., Yang, X.: Dfa-nerf: personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791 (2022)", + "68. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023)", + "69. Yehia, H., Rubin, P., Vatikiotis-Bateson, E.: Quantitative association of vocal-tract and facial behavior. Speech Communication 26(1-2), 23-43 (1998)", + "70. Yin, F., Zhang, Y., Cun, X., Cao, M., Fan, Y., Wang, X., Bai, Q., Wu, B., Wang, J., Yang, Y.: Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 85-101. Springer (2022)", + "71. Zhan, F., Yu, Y., Wu, R., Zhang, J., Lu, S.: Multimodal image synthesis and editing: A survey. arXiv preprint arXiv:2112.13592 (2021)", + "72. Zhang, C., Zhao, Y., Huang, Y., Zeng, M., Ni, S., Budagavi, M., Guo, X.: Facial: Synthesizing dynamic talking face with implicit attribute learning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3867-3876 (2021)", + "73. Zhang, Z., Hu, Z., Deng, W., Fan, C., Lv, T., Ding, Y.: Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. arXiv preprint arXiv:2303.03988 (2023)", + "74. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021)", + "75. Zhen, R., Song, W., He, Q., Cao, J., Shi, L., Luo, J.: Human-computer interaction system: A survey of talking-head generation. Electronics 12(1), 218 (2023)", + "76. Zhong, W., Fang, C., Cai, Y., Wei, P., Zhao, G., Lin, L., Li, G.: Identity-preserving talking face generation with landmark and appearance priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2023)", + "77. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 9299–9306 (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "78. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4176-4186 (2021)", + "79. Zhou, M., Bai, Y., Zhang, W., Yao, T., Zhao, T., Mei, T.: Responsive listening head generation: A benchmark dataset and baseline (2022)", + "80. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG) 39(6), 1-15 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 272 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "D. Yaman et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_model.json b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..690696e3c6048ab4d20c715341114218b7d5ab56 --- /dev/null +++ b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_model.json @@ -0,0 +1,2803 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.257, + 0.141, + 0.75, + 0.186 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.212, + 0.744, + 0.244 + ], + "angle": 0, + "content": "Dogucan Yaman1, Fevziye Irem Eyiokur1, Leonard Bärmann1, Hazim Kemal Ekenel2, and Alexander Waibel1,3" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.255, + 0.694, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Karlsruhe Institute of Technology, Karlsruhe, Germany" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.269, + 0.665, + 0.283 + ], + "angle": 0, + "content": "2 Istanbul Technical University, Istanbul, Turkey" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.283, + 0.668, + 0.311 + ], + "angle": 0, + "content": "\\(^{3}\\) Carnegie Mellon University, Pittsburg PA, USA dogucan.yaman@kit.edu" + }, + { + "type": "list", + "bbox": [ + 0.31, + 0.255, + 0.694, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.341, + 0.741, + 0.535 + ], + "angle": 0, + "content": "Abstract. Talking face generation aims to create realistic videos with accurate lip synchronization and high visual quality, using given audio and reference video while preserving identity and visual characteristics. In this paper, we start by identifying several issues with existing synchronization learning methods. These involve unstable training, lip synchronization, and visual quality issues caused by lip-sync loss, SyncNet, and lip leaking from the identity reference. To address these issues, we first tackle the lip leaking problem by introducing a silent-lip generator, which changes the lips of the identity reference to alleviate leakage. We then introduce stabilized synchronization loss and AVSyncNet to overcome problems caused by lip-sync loss and SyncNet. Experiments show that our model outperforms state-of-the-art methods in both visual quality and lip synchronization. Comprehensive ablation studies further validate our individual contributions and their cohesive effects." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.548, + 0.741, + 0.563 + ], + "angle": 0, + "content": "Keywords: Talking face generation \\(\\cdot\\) Lip synchronization \\(\\cdot\\) Lip leaking" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.585, + 0.377, + 0.601 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.705 + ], + "angle": 0, + "content": "Audio-driven talking face generation aims at generating a video with respect to given face and audio sequences. The objective is to achieve synchronized lip movements corresponding to the provided audio while preserving the identity and visual details. This task has recently attracted significant attention due to its versatile applications, including dubbing in the film industry, online education, enhancing video conferencing, and dubbing for various types of videos [71, 75]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The talking face generation task comprises two primary aspects: (1) lip synchronization and (2) visual quality of the face. Since lips that are out-of-sync with audio can be easily identified by humans, synchronized lips are key to achieving natural and realistic talking-face generation. For this, the primary solution is to evaluate audio-lip synchronization and use it as a training objective. Wav2Lip [46] introduced an improved version of SyncNet [14], a pretrained model designed to measure audio-visual synchronization. This model is also used during training to extract features and compute lip-sync loss. Subsequently, many approaches have employed improved SyncNet [46] to guide the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.15, + 0.404, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.26, + 0.371, + 0.272 + ], + "angle": 0, + "content": "(a) SyncNet [46]" + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.15, + 0.601, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.26, + 0.584, + 0.272 + ], + "angle": 0, + "content": "(b) AVSyncNet (ours)" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.146, + 0.781, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.26, + 0.762, + 0.272 + ], + "angle": 0, + "content": "(c) Leaking problems" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.339 + ], + "angle": 0, + "content": "Fig. 1: (a, b) Cosine similarity between GT audio-lip pairs on random LRS2 samples, showcasing the instability of SyncNet and more robust performance of AVSyncNet. (c) illustrates full mouth region / lip leaking from the reference, pose effect from the reference, and similar identity reference-target image scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.789, + 0.492 + ], + "angle": 0, + "content": "model throughout training. The main concept is to provide audio-video pairs to SyncNet, which extracts features through its image and audio encoders. Subsequently, cosine similarity with binary cross-entropy loss is measured between these two features [46]. Since the model was trained in this manner, a higher cosine similarity is expected when the lips are synchronized with the given audio. Although existing approaches with this strategy mostly surpass other methods in lip synchronization, there are still challenges that must be addressed to enhance performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this work, we identify two main challenges in existing approaches that restrict models from achieving satisfactory performance in both lip synchronization and visual quality: SyncNet instabilities and lip leaking. First, in alignment with previous research [41], we observe instabilities in the performance of SyncNet [46] on pairs of ground-truth (GT) lip and audio (see Fig. 1a and App. A). Thus, when SyncNet is utilized as part of talking face generation training, it might provide an inadequate training signal, assigning low similarity scores to generated images even when the lips are synchronized. This problem causes unstable training, degrading the lip generation capability of the network, thereby ending up in out-of-sync lips or suboptimal lip-sync. Furthermore, the lip-sync loss [46] and reconstruction losses are conflicting [39]. This leads models to have either poor lip synchronization or degraded visual quality (sometimes even both), escalating further when lip-sync loss and SyncNet are employed with high-resolution (HR) data [39,73]. To address these problems, we first improve SyncNet and introduce AVSyncNet, demonstrating a more robust performance (see Fig. 1b) and also overcoming poor shift-invariance characteristics of SyncNet (see Fig. 6a). However, despite improved performance, the instability problem is not fully solved (see Fig. 1b). To overcome this problem further, we also introduce a stabilized synchronization loss. Specifically, instead of directly using the similarity of the (generated lips, audio) pair, we calculate the difference of the similarities between (GT lips, audio) and (generated lips, audio). We hypothesize that this alleviates the described problems since it guides the model to generate a lip movement with a similar synchronization score as for the GT face. Together with AVSync-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Net, this method empirically enhances the lip synchronization performance as well as avoids visual quality issues caused by misleading lip-sync loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Second, we address another main problem in the talking face generation literature: lip leaking. The current gold standard in 2D-based methods is to input a bottom-half masked face image (\"pose reference\"), wherein the model is expected to generate the input face with proper lip movements. However, because of masking, the model requires a reference image to retain the identity and texture in the masked part. This is achieved by randomly choosing a face image from a different part of the video sequence, referred to as \"identity reference\". However, this introduces a new challenge: As it is randomly selected, the lip movements of the identity reference can frequently be similar to the GT lips during training (see last row of Fig. 1c). Hence, for faster convergence, the model tends to replicate the lip movements from the identity reference, resulting in poor lip synchronization or complete out-of-sync output. By following the literature [41,46], we term this phenomenon as lip leaking. To tackle this problem, we propose a simple yet effective technique: silent-lip generator. The concept involves modifying the identity reference to make the lips closed (thus \"silent-lip\") before feeding it to the talking face generator network. This method ensures consistently closed lips in the identity reference, effectively mitigating the lip leaking problem. Our contributions are as follows: (i) We identify and analyze various fundamental problems that harm lip synchronization learning and also cause visual quality issues. (ii) We present a robust and shift-invariant AVSyncNet, and stabilized synchronization loss to overcome the problems caused by lip-sync loss and SyncNet. (iii) We present a silent-lip generator to generate an identity reference with closed lips before feeding the talking face generator to alleviate the lip leaking." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.388, + 0.567 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation. Initial studies mapped audio features to time-aligned facial motions [69] or predicted facial motions by HMM [6]. In [57], videos were generated by finding the images most aligned with the audio. ATVGNet [9] transfers audio to facial landmarks and uses pixel-wise loss with an attention mechanism to avoid the jittering problem and leaking of irrelevant speech. [16] and [80] use facial landmark representation to synthesize faces with synchronized lips. Recent papers address the task as conditional inpainting by masking the bottom half of the input face and feeding the network an identity reference from another time step of the same video, along with the audio segment. Wav2Lip [46] proposes SyncNet and lip-sync loss to predict lip synchronization and achieves significant performance. PC-AVS [78] introduces a pose-controllable audio-visual system, while GC-AVT [36], EAMM [29] and EVP [30] control the emotion by utilizing an emotion embedding. On the other hand, SyncTalkFace [45] uses Audio-Lip Memory to store lip motion features and retrieves them as visual hints for better synchronization. VideoReTalking [10] proposes to manipulate the reference image to have a face with canonical expression to alleviate the sensitivity of the model against the identity reference image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Similarly, we introduce a silent-lip generator to implicitly learn to manipulate the lips of the identity reference to mitigate the lip leaking problem. Compared to the method proposed in [10], our method works more efficiently to preserve the identity and visual quality. More recently, LipFormer [64] uses a pre-learned facial codebook to generate HR videos, while DINet [73] proposes to use a deformation module to obtain deformed features to enhance lip synchronization and head pose alignment. IPLAP [76] shows satisfactory visual quality and stability by using intermediate landmark representation and motion field. Recently, TalkLip [63] introduces a global audio encoder, trained with self-supervised learning, to encode features by considering the entire content of the audio. Besides, they propose to use lip reading during the training as well as in the evaluation to control whether the content is preserved. SIDGAN [39] performs important analyzes, and introduces shift-invariant APS-SyncNet and training objectives along with the coarse-to-fine pyramid model for HR dubbed video generation. Recent works use diffusion model because of its stability and accuracy [50, 55]. Finally, in [61], talking face generation is used as a part of a full system to perform end-to-end face dubbing, involving speech recognition, translation, speech, and video generation. In contrast to the above, 3D-based and Neural Radiance Fields-based (NeRFs) methods typically generate the entire head, rather than just manipulating 2D images of the face, often involving manipulation of pose, emotion, and 3D face model [4, 5, 21, 37, 44, 49, 53, 58, 59, 62, 66–68, 70, 72, 74, 77, 80]. Similarly, portrait animation aims at utilizing a single input image to generate a video by predicting the pose and expression of the subsequent frames, along with ensuring synchronized lips. Nevertheless, these tasks significantly differ from face dubbing in terms of methodology, task definition, goals, and real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.79, + 0.742 + ], + "angle": 0, + "content": "Lip Synchronization. Some earlier works [23, 52] used hand-crafted features and statistical models to evaluate lip synchronization. Recent studies proposed to use mutual information between audio-visual features to produce sync or out-of-sync output for sound [8,26,43] or speech [2,12,15,32,33]. While some methods learn lip synchronization implicitly [9,21,28,34,57,66,74], other methods employ distance between landmarks or facial parameters [30,44,53,80]. Contrarily, the majority of works [17,20,36,45,46,54,56,60,62-64,67,73,77,78] extract audio-visual features with an additional network (mostly SyncNet [46]) for audio-lip synchronization prediction. Afterwards, while some of these methods utilize lip-sync loss [46], others benefit from contrastive learning by using infoNCE [42]. We follow a similar strategy by utilizing an additional network to calculate a synchronization loss. We propose a robust and shift-invariant AVSyncNet along with a stabilized synchronization loss to overcome existing challenges." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.76, + 0.446, + 0.779 + ], + "angle": 0, + "content": "3 Proposed Approach" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.465, + 0.804 + ], + "angle": 0, + "content": "3.1 Talking Face Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We propose an audio-driven talking face generation model \\( G_{L} \\) with enhanced lip synchronization. As shown in Fig. 2a, our model incorporates: 1) an audio" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.664, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.316, + 0.539, + 0.327 + ], + "angle": 0, + "content": "(a) Talking face generation model" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.178, + 0.787, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.315, + 0.771, + 0.327 + ], + "angle": 0, + "content": "(b) FD block" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 2: Talking face generation model \\( G_{L} \\) (a) and face-decoding (FD) block (b). Our model receives a pose reference sequence, mel-spectrogram of an audio snippet, and a silent identity reference, that is generated by our silent-lip generator \\( G_{S} \\), aiming to alleviate lip leaking problem. The model then synthesizes the talking face sequence to ensure lip synchronization. Subsequently, the employed loss functions are computed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.471 + ], + "angle": 0, + "content": "encoder for processing the audio snippet, 2) an identity encoder for addressing an identity reference image, and 3) a pose encoder for utilizing a pose reference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Audio Encoder. The audio encoder \\( E_{A} \\) generates phoneme-level embeddings, serving as conditions for the face generator to generate lip movements accurately. Our audio encoder extracts embeddings \\( F_{A} = E_{A}(A) \\in \\mathbb{R}^{1 \\times 1 \\times 512} \\) from the given mel-spectrogram \\( A \\), representing the driving audio. In contrast to existing methods [10, 46, 73, 76], we propose using a pretrained, frozen audio encoder, concurrently trained with a face encoder to learn lip synchronization similar to the objective in SyncNet [46]. Thus, we can obtain improved embeddings during the talking face generation training by leveraging the capacity of the pretrained robust audio encoder. We obtain the best score with such frozen audio encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Face Encoder. In accordance with the gold standard in the literature, we utilize an identity reference, \\( I^R \\), and a pose reference, \\( I \\), as inputs to the model. The identity reference is a face image of the subject that provides identity information. It is different from the pose reference and randomly selected from the same video. The pose reference is identical to the target image, except for the bottom half, which is masked, as the model is designed to focus on generating lip movements. Unlike most conventional methods that employ a joint encoder for processing identity and pose references, we use individual encoders to allow each encoder to focus solely on their respective tasks [39]. Therefore, we utilize two parallel CNN-based face encoders to process identity and pose references individually. This approach yields better feature representation and ultimately leads to improved performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.481, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.303, + 0.449, + 0.317 + ], + "angle": 0, + "content": "(a) Silent-lip generation model \\(G_{S}\\)" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.163, + 0.791, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.303, + 0.72, + 0.317 + ], + "angle": 0, + "content": "(b) AVSyncNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.327, + 0.72, + 0.343 + ], + "angle": 0, + "content": "Fig. 3: \\( G_{S} \\) in inference (a) and AVSyncNet training pipeline (b)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.434, + 0.389 + ], + "angle": 0, + "content": "3.2 Silent-Lip Generator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Talking face generation involves using an identity reference to preserve the identity in the generated image. This is particularly important because the bottom half of the pose reference, specifically the mouth region, is masked. This masking is necessary as we aim to synthesize appropriate lip movements corresponding to the provided audio. However, models are unintentionally affected by the lip movement of the identity reference, rather than solely gathering identity information (see Fig. 1c and App. C for details). This behavior, which we refer to as lip leaking, leads to poor lip synchronization or occasionally even non-converged training. We hypothesize that there are two main reasons for this: First, the lip movement of the identity reference may occasionally resemble that of the lips in the target image (see last row of Fig. 1c). Hence, the model can lower the synchronization loss more quickly by undesirably replicating the lip movements of the identity reference. Second, the diversity of lip movements in the identity reference may yield the model to seek a correlation with the target lips. This causes a challenging disentanglement task for the identity encoder —namely distinguishing identity information and lip movement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.749 + ], + "angle": 0, + "content": "To mitigate the aforementioned issue, we propose to use an additional model, called silent-lip generator \\( G_{S} \\), prior to the talking face generation model \\( G_{L} \\), aimed at modifying the lip shape of the identity reference. Specifically, we reconstruct the input face with closed, flat lips. This strategy reduces the likelihood of having lips similar to the target face in the identity reference and resolves the issue of diverse lip movements. Consequently, the model no longer replicates the lips from the identity reference, resulting in stable training & improved lip sync." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To implement \\( G_{S} \\), we structure the task and the model similar to the talking face generation. Specifically, we input a bottom-half masked pose reference, an identity reference, and an audio snippet to train a GAN [18] for generating a talking face. The model is trained to reconstruct the face using both the pose and identity references. Notably, we exclude any synchronization loss and SyncNet during training and train the model under the weak condition. Consequently, the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.33 + ], + "angle": 0, + "content": "model focuses solely on generating lip movements without synchronization when speech is present. Therefore, it implicitly learns to generate closed lips when the input audio is silent as shown in Fig. 3a. Furthermore, by eliminating synchronization loss, we overcome unstable training and the issue of lip leaking from the identity reference for this network. We choose this approach for its efficient utilization of the same model and data. Given the scarcity of closed-lip faces in the dataset, we avoid using these frames directly to maintain effective training and generalization. Please note that we use the same architecture for \\( G_{S} \\) and \\( G_{L} \\). We initially train \\( G_{S} \\) separately on the same training data (LRS2) and then incorporate it into the training of \\( G_{L} \\) without further finetuning. Specifically, we then only pass silent audio to \\( G_{S} \\), so that it modifies given identity references to have silent lips for subsequent use as identity reference for \\( G_{L} \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.348, + 0.411, + 0.362 + ], + "angle": 0, + "content": "3.3 Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.534 + ], + "angle": 0, + "content": "For the talking face video generation, we employ the aforementioned components. Initially, we employ our pretrained silent-lip generator \\( G_{S} \\) to synthesize the identity reference with closed lips. Subsequently, we input this silent identity reference to the identity encoder and the pose reference to the pose encoder in the talking face generation model \\( G_{L} \\). Similarly, we provide the mel-spectrogram of the corresponding audio snippet to the audio encoder. Next, we concatenate the embeddings from the identity and pose encoders, along with the depth dimension, and pass it through a \\( 1 \\times 1 \\) convolution layer to reduce the depth. Finally, we concatenate this feature representation with the audio embeddings before feeding them into the face generator. The face generator generates the entire face with accurate lip movements by preserving the identity and the pose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.789, + 0.747 + ], + "angle": 0, + "content": "We illustrate the talking face generation model \\( G_{L} \\) in Fig. 2a. We integrate the U-Net architecture [47] for our overall design, leveraging its adequate performance in reconstruction tasks while ensuring computational efficiency. Our identity encoder and pose encoder share the same architecture, consisting of consecutive face-encoding (FE) blocks. Each block has a strided-convolutional layer followed by two non-strided convolutional layers, each paired with a batch normalization layer [27] and a ReLU activation function [35,40]. We also use the residual connection strategy [22] by summing the input and output of each block before forwarding it to the next layer. On the other hand, our face generator has consecutive face-decoding (FD) blocks. As shown in Fig. 2b, within each block, we utilize a transposed-convolutional layer, followed by two convolutional layers incorporating batch normalization and ReLU activation function. Moreover, we apply a skip connection between the reciprocal layers of the face encoders and the decoder to retain high-level features and enhance the training stability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "GAN Loss. To train our model, we utilize GAN loss [18] and employ a discriminator, which is a straightforward CNN-based binary classification network to distinguish real and fake samples, designed with a balanced architecture aligned with our face encoders. We benefit from consecutive strided-convolutional layers followed by the Leaky ReLu activation function and spectral normalization [38]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.209 + ], + "angle": 0, + "content": "Reconstruction Loss. We employ L1 loss in pixel space \\(L_{pixel} = ||I' - I^{GT}||_1\\) between the generated \\(I'\\) and the target faces \\(I^{GT}\\) to ensure consistency in areas outside the lips and maintain the illumination condition. We further utilize perceptual loss [31] based on the pretrained VGG-19 [51]:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.22, + 0.786, + 0.261 + ], + "angle": 0, + "content": "\\[\nL _ {p e r} = \\sum_ {i = 1} ^ {5} c _ {i} | | \\mathrm {V G G} ^ {\\phi_ {i}} \\left(I ^ {\\prime}\\right) - \\mathrm {V G G} ^ {\\phi_ {i}} \\left(I ^ {G T}\\right) | | _ {1} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.271, + 0.785, + 0.288 + ], + "angle": 0, + "content": "where \\( c_{i} \\) are weight coefficients from [31], and \\( \\phi \\) indicates the set of VGG layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.789, + 0.508 + ], + "angle": 0, + "content": "Adaptive Triplet Loss. Although our goal is to capture visual details from the identity reference, we observe that the model occasionally focuses on the visual details (e.g., illumination, pose) in the identity reference excessively and this tendency could potentially degrade the quality and stability of the generated face sequence, resulting in suboptimal performance. To tackle this, we exploit a triplet loss strategy [48], aiming to minimize the distance between the generated face and GT, while maximizing the distance between the generated face and the identity reference. However, the random selection of identity references increases the probability of choosing an image that closely resembles the GT. This scenario poses a challenge for the vanilla triplet loss, potentially degrading training and resulting in poor visual and pose quality. To mitigate this, we introduce an adaptive triplet loss that considers the similarity between the identity reference and GT during loss computation to alter its effect. The formula is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.519, + 0.786, + 0.555 + ], + "angle": 0, + "content": "\\[\nL _ {a t} = \\left[ D (\\mathrm {V G G} (I ^ {\\prime}), \\mathrm {V G G} (I ^ {G T})) - \\frac {D (\\mathrm {V G G} (I ^ {\\prime}) , \\mathrm {V G G} (I ^ {R}))}{D (\\mathrm {V G G} (I ^ {G T}) , \\mathrm {V G G} (I ^ {R}))} + \\alpha \\right] _ {+} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.687 + ], + "angle": 0, + "content": "where \\([\\cdot]_{+} = \\max(\\cdot, \\epsilon)\\), \\(D\\) represents the L2 distance, and we empirically choose \\(\\alpha = 1\\). In this loss, we leverage the ratio of the similarity between the generated image and identity reference to that of the GT and the identity reference to adjust the loss value. As the identity reference becomes more similar to GT, the impact of the distance between the generated image and identity reference on the loss diminishes, since expecting a high distance in this case is not reasonable. Since our objective is to incorporate visual details from the identity reference, we opt for a very low coefficient to avoid conflicting with the primary goal." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.475, + 0.725 + ], + "angle": 0, + "content": "3.4 Learning Synchronization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The lip-sync loss [46] serves as a method to calculate synchronization between audio and video. Leveraging the pretrained SyncNet [46] for feature extraction from both audio and video inputs demonstrates reasonable performance in learning lip synchronization. However, the SyncNet is significantly unstable when measuring this similarity. Our evaluation using SyncNet on GT training data reveals notable fluctuations in cosine similarity between video and audio, contrary to the expected high scores (see Fig. 1a and App. A for details). Therefore, this" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "provides conflicting information to the system, resulting in poor lip synchronization, unstable training, and degraded visual quality. To tackle this problem, we present a more accurate, shift-invariant, and robust version of SyncNet, named AVSyncNet, and introduce a novel, stabilized synchronization loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.789, + 0.396 + ], + "angle": 0, + "content": "AVSyncNet. We employ a ResNet-50-based [22] image encoder, known for its superior performance in face recognition, alongside a ResNetSE-34-based audio encoder [11], which is a modified version of ResNet-34 [22] designed to handle spectrogram inputs. This design makes the AVSyncNet model robust and shift-invariant compared to SyncNet [46]. We train our model on the LRS2 training data [1], calculating the cosine similarity between audio and lip features, followed by a binary cross-entropy loss, as shown in Fig. 3b. During each training step, we provide a set of images (5 images) along with the corresponding audio. For negative samples, we randomly select an audio snippet from the non-overlapping part of the video. Please note that as we feed the bottom half of the face to the image encoder, we adapt the first layer of ResNet-50 for an input size of \\(112 \\times 224\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.522 + ], + "angle": 0, + "content": "Stabilized Synchronization Loss \\((L_{ss})\\). Although AVSyncNet shows improved performance compared to SyncNet [46] and alleviates existing instability problem that harms lip-sync and visual quality, the unstable performance is not fully solved due to the inherent challenges of the task (see Fig. 1b and App. A, B). Therefore, we introduce a stabilized synchronization loss \\((L_{ss})\\) to improve the lip synchronization performance further in conjunction with AVSyncNet by providing more stable and precise supervision. The formula is shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.529, + 0.787, + 0.562 + ], + "angle": 0, + "content": "\\[\nL _ {s s} = - \\log \\left(1 - \\frac {\\left| x - y \\right| + \\epsilon}{\\left| x - y \\right| + \\left| y - d \\right| + \\epsilon}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.585, + 0.787, + 0.605 + ], + "angle": 0, + "content": "\\[\nx = \\operatorname {A V S I M} \\left(I ^ {\\prime}, A\\right), \\quad y = \\operatorname {A V S I M} \\left(I ^ {G T}, A\\right), \\quad d = \\operatorname {A V S I M} \\left(I ^ {R}, A\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.784, + 0.687 + ], + "angle": 0, + "content": "where \\( I^{\\prime}, I^{GT} \\), and \\( I^R \\) are generated, GT, and identity reference lips, respectively, while \\( A \\) is their corresponding audio. AVSIM(\\( I, A \\)) indicates the audio-visual similarity between a face image (bottom half only, i.e. lips) and audio, given by the cosine similarity of extracted image and audio features \\( \\phi_{AVS}^V(I) \\), \\( \\phi_{AVS}^A(A) \\) of the respective AVSyncNet encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.751 + ], + "angle": 0, + "content": "In this formulation, \\( x \\) followed by cross-entropy loss denotes the lip-sync loss [46]. However, to address the unstable and fluctuating performance, we utilize the relative distance in similarity between GT lips-audio and generated-audio pairs. Sometimes, randomly selected reference images may have a lip" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.755, + 0.788, + 0.842 + ], + "angle": 0, + "content": "This is reminiscent of distillation loss [25], as the actual value of the scores is neglected, and only their difference provides loss value for the training. However, initial experiments trying to directly minimize the distance between SyncNet (or AVSyncNet) image encoder features of generated and GT faces showed poor performance. We hypothesize that this is caused by SyncNet image features being only meaningful for comparing with corresponding audio features due to its training strategy." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "movement similar to that of the target image. We have already introduced the silent-lip generator \\( G_{S} \\) to mitigate this situation. However, AVSyncNet, while less severe than SyncNet, might still be unstable (e.g., unexpectedly high scores with incorrect pairs or vice versa), resulting in a minor lip leaking and stability issue. This problem arises when AVSyncNet assigns an erroneously high score to silent lip-audio pairs as well as when the target lip shape looks similar to closed lips. To mitigate this, we inject the similarity score between the identity reference and audio into the formulation. Specifically, we penalize the model more when the identity reference-audio pair shows higher similarity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.307, + 0.459, + 0.323 + ], + "angle": 0, + "content": "3.5 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.333, + 0.646, + 0.349 + ], + "angle": 0, + "content": "Combining all the presented contributions, the total loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.362, + 0.786, + 0.379 + ], + "angle": 0, + "content": "\\[\nL = L _ {G A N} (G, D) + \\lambda_ {1} L _ {p i x e l} (G) + \\lambda_ {2} L _ {p e r} (G) + \\lambda_ {3} L _ {s s} (G) + \\lambda_ {4} L _ {a t} (G) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.789, + 0.435 + ], + "angle": 0, + "content": "where \\( G \\) and \\( D \\) indicate generator and discriminator outputs, respectively. We empirically found the best coefficients as \\( (\\lambda_1, \\dots, \\lambda_4) = (10, 1, 2, 0.5) \\). Please note that this is the loss function for training \\( G_L \\), while for \\( G_S \\) we set \\( \\lambda_3 = \\lambda_4 = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.79, + 0.618 + ], + "angle": 0, + "content": "We process videos by using 5 consecutive frames in each step to consider the temporal information. We detect faces with FAN [7], followed by acquiring tight crops and resizing to \\(96 \\times 96\\), as faces in LRS2 [1] are of low resolution. Our audio encoder receives a mel-spectrogram of size \\(16 \\times 80\\) derived from \\(16\\mathrm{kHz}\\) audio with a window size of 800 and a hop size of 200. We employ the Adam optimizer with \\((\\beta_{1},\\beta_{2}) = (0.5,0.999)\\) and set the learning rate to \\(1 \\times 10^{-4}\\) for all models. Training our AVSyncNet is done similarly to SyncNet [46] on the LRS2 dataset, and then we freeze the audio encoder of AVSyncNet and use it in the training of \\(G_{S}\\) and \\(G_{L}\\). At the end of talking face generation, we apply a post-processing step by using VQFR [19] to enhance the visual quality and the resolution, aiming to achieve HR videos. We train and test our models with a single NVIDIA RTX A6000 GPU." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.641, + 0.462, + 0.659 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Dataset. We trained our silent-lip generator and talking face generator using Lip Reading Sentence 2 (LRS2) [1] training set as it is a well-known benchmark with extensive subject diversity. The evaluation was carried out on the LRS2 test set and extended to the LRW [13] test set and HTDF dataset [74] to demonstrate performance on unseen data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Metrics and Baseline. For visual quality, we employ widely used metrics: FID [24], SSIM [65], and PSNR. We also use inter-frame consistency (IFC), presented as a training objective in [79]. This is achieved by calculating the difference between the distances of the consecutive frames in the generated and GT videos. To evaluate lip synchronization, we follow the literature and use Landmark Distance (LMD) [9] in the mouth region and LSE-C & LSE-D metrics [46]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.281, + 0.145, + 0.721, + 0.159 + ], + "angle": 0, + "content": "Table 1: Quantitative results on the test sets of LRS2 and LRW." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.172, + 0.787, + 0.283 + ], + "angle": 0, + "content": "
LRS2LRW
MethodSSIM ↑PSNR ↑FID ↓IFC ↓LMD ↓LSE-C ↑LSE-D ↓SSIM ↑PSNR ↑FID ↓IFC ↓LMD ↓LSE-C ↑LSE-D ↓
Wav2Lip [46]0.8626.537.050.212.387.596.750.8525.146.810.202.147.496.51
PC-AVS [78]0.7328.2418.400.461.936.417.520.8132.2514.270.381.426.537.15
EAMM [29]0.6921.0184.650.513.543.319.930.7126.2244.160.482.614.329.04
VideoReTalking w/ FR [10]0.8425.589.280.222.617.496.820.8727.115.300.232.396.597.12
DINet [73]0.7824.354.260.252.305.378.370.8827.508.170.221.965.249.09
TalkLip [63]0.8626.114.940.242.348.536.080.8626.3415.730.261.837.286.48
IPLAP [76]0.8729.674.100.202.116.497.160.9130.458.400.211.645.947.76
Ours w/o FR0.9532.643.830.161.138.416.030.9231.454.460.181.227.866.24
Ours w/ FR (VQFR)0.9031.805.230.271.368.525.830.9030.217.050.211.417.926.00
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.371 + ], + "angle": 0, + "content": "to measure the confidence and distance scores through a pretrained model [14]. We choose SOTA methods with publicly available codes and models to compare them fairly under the same conditions, as the implementation of the metrics and face cropping strategy before computing the metrics affect the scores." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.392, + 0.433, + 0.407 + ], + "angle": 0, + "content": "4.1 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.506 + ], + "angle": 0, + "content": "In Tab. 1, we present quantitative results on test sets of two benchmark datasets, namely LRS2 and LRW. We achieve state-of-the-art results in all visual quality metrics excluding PSNR on LRW, which is a less informative metric compared to SSIM and FID. On IFC, we similarly outperform all compared methods, indicating that our model generates the most consistent videos in terms of temporal information and stability of the faces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.675 + ], + "angle": 0, + "content": "In lip synchronization evaluation, we obtain the best performance in LMD. Nevertheless, it is important to highlight that LMD is sensitive to the changes in the image, as it does not disentangle synchronization and visual stability. For instance, affine transformations impact the LMD score even when the lips are synchronized, and vice versa. On the LRW dataset, we achieve SOTA results with more reliable confidence and distance metrics for lip synchronization: LSE-C & D. On the LRS2 dataset, TalkLip yields a slightly better score than our model with the LSE-C metric. Nevertheless, we outperform TalkLip and achieve a SOTA result with the LSE-D metric. All these results indicate the accuracy of our method in terms of visual quality and lip synchronization. Similarly, we achieve SOTA results for most metrics on unseen HDTF [74], see App. E.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.421, + 0.71 + ], + "angle": 0, + "content": "4.2 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In Fig. 4, we demonstrate a qualitative comparison with SOTA models and GT data. We use their respective publicly available models and generate videos from the HDTF dataset [74] to compare the models on unseen data since the presented models were trained on the LRS2 dataset, except for DINet (trained with HDTF). Due to SyncNet and lip-sync loss, TalkLip and Wav2Lip encounter generalization issues, sometimes leading to visual artifacts in the mouth region or face boundaries, especially when the pose of the identity reference differs from the pose reference, despite generating accurate lip movements. This observation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.434, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.152, + 0.55, + 0.174 + ], + "angle": 0, + "content": "Wav2Lip (Prajwal et al. 2020)" + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.184, + 0.548, + 0.207 + ], + "angle": 0, + "content": "VideoReTalking (Cheng et al. 2022)" + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.218, + 0.547, + 0.24 + ], + "angle": 0, + "content": "DINet (Zhang et al. 2023)" + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.25, + 0.546, + 0.272 + ], + "angle": 0, + "content": "TalkLip (Wang et al. 2023)" + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.283, + 0.548, + 0.306 + ], + "angle": 0, + "content": "IP-LAP (Zhong et al. 2023)" + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.321, + 0.532, + 0.332 + ], + "angle": 0, + "content": "Ours w/o FR" + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.355, + 0.533, + 0.365 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.144, + 0.787, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.443 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison with the SOTA methods. Reference videos (from HDTF [74]) are randomly selected and not seen during training by our model. For more images and videos, please check App. E, F and https://yamand16.github.io/TalkingFaceGeneration/." + }, + { + "type": "table_caption", + "bbox": [ + 0.259, + 0.457, + 0.742, + 0.471 + ], + "angle": 0, + "content": "Table 2: Ablation studies on the LRS2 test set. See the text for details." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.485, + 0.731, + 0.603 + ], + "angle": 0, + "content": "
AblationSetupMethodPSNR ↑ SSIM ↑ FID ↓ LMD ↓ LSE-C ↑ LSE-D ↓ IFC ↓
ComponentsAGL+LSa26.3490.85312.252.4087.116 ± 1.927.396 ± 1.03
BA + Ea,S26.6140.8689.822.3257.271 ± 1.767.106 ± 0.98
CA + Ea,W26.5900.86910.562.2787.220 ± 1.757.158 ± 0.99
DB + GS27.1800.8728.161.7417.752 ± 1.716.413 ± 0.95
EGL + Ea,S + GS + LSs31.1660.9255.271.1408.370 ± 1.166.032 ± 0.59
FE + Lt30.6580.9176.241.2508.260 ± 1.346.176 ± 0.64
GE + Lat32.7550.9494.021.1358.382 ± 1.166.057 ± 0.61
HG w / AVSyncNet32.6400.9523.831.1308.410 ± 0.976.037 ± 0.55
Post-processingFR1Setup H + GPEN28.9910.91958.771.1977.6256.457
FR2Setup H + GFPGAN31.1690.91613.071.2197.6246.496
FR3Setup H + VQFR: full model31.8060.9055.231.3658.5285.838
Silent face generationVRT-SVideoReTalking silent data22.1240.64633.60--0.463
Ours-SOur silent data (GS)33.3280.9514.41--0.141
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.722 + ], + "angle": 0, + "content": "clearly validates the motivation of our contributions. In contrast, our model generates consistent faces with comparably fewer artifacts, featuring appropriate lip movements that align with both the GT faces and the corresponding audio. However, VideoReTalking demonstrates comparable lip synchronization and visual quality performance to our model. On the other hand, IP-LAP shows sufficient visual quality, while less accurate lip synchronization." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.391, + 0.758 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Tab. 2 and Fig. 5a show a comprehensive ablation study on the LRS2 test set, analyzing the individual impact of our contributions. We first train our \\( G_{L} \\) model using SyncNet [46] and lip-sync loss [46] as a baseline. As expected, we encountered several issues with unstable training. Once our model converged after several random seeds, the results (Setup A) show that lip-sync loss can achieve" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.449, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.195, + 0.382, + 0.205 + ], + "angle": 0, + "content": "(a) Components." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.206, + 0.448, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.244, + 0.447, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.282, + 0.447, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.32, + 0.446, + 0.341 + ], + "angle": 0, + "content": "(c) Generated samples for challenging cases: Pose, teeth, beard." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.146, + 0.761, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.195, + 0.735, + 0.205 + ], + "angle": 0, + "content": "(b) Face restoration methods." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.206, + 0.785, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.32, + 0.786, + 0.342 + ], + "angle": 0, + "content": "(d) Silent face comparison with VRT [10]. © European Central Bank (CC BY)." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.353, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Fig. 5: Ablation studies of components (a), face restoration methods (b), and silent face generation (d). (c) demonstrates generated images in challenging cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.71 + ], + "angle": 0, + "content": "decent synchronization performance, despite lower visual quality and training stability issues. In this setup, we train the audio encoder as a part of \\( G_{L} \\). Replacing this with our pretrained audio encoder enhances the synchronization and visual quality (Setup B). For further comparison, we also utilize the audio encoder of the Wav2Vec2 [3], presented in Setup C. However, it slightly decreases the scores, which validates our hypothesis about training the audio encoder for synchronization purposes. Thus, we continue with Setup B and add our silent-lip generator \\( G_{S} \\) to generate silent identity references (Setup D). \\( G_{S} \\) alleviates the lip leaking problem, makes the training more stable, and improves the scores noticeably. We further replace lip-sync loss with our stabilized synchronization loss, yielding drastically improved synchronization and visual quality scores (Setup E). Moreover, we observed that Setup E shows almost no instability in the training. In Setup F, we train Setup E including vanilla triplet loss and it enhances neither visual quality nor synchronization; in fact, it even causes detrimental effects. This shows the necessity of modifying the triplet loss and introducing the adaptive triplet loss. In Setup G, replacing vanilla triplet loss with adaptive triplet loss demonstrates a slight improvement in visual quality, while not having a negative impact on lip synchronization. In Setup H, we switch SyncNet with AVSyncNet and achieve slightly better visual quality and lip synchronization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.714, + 0.788, + 0.761 + ], + "angle": 0, + "content": "We compare \\( G_{S} \\) with VRT [10] silent face generation approach in Tab. 2. Our model surpasses VRT quantitatively and qualitatively (see Fig. 5d), preserving the visual details and identity while modifying the lips." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We compare different face restoration methods as post processing and present the results in Tab. 2 and Fig. 5b. VQFR surpasses other methods in preserving lip synchronization as well as FID and PSNR. However, GPEN shows better performance in the remaining metrics (see App. F.2 for details). In summary, we employ VQFR for post-processing in our full model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.154, + 0.382, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.236, + 0.367, + 0.249 + ], + "angle": 0, + "content": "(a) Shifting [39]." + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.154, + 0.54, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.237, + 0.516, + 0.249 + ], + "angle": 0, + "content": "(b) Rotation." + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.148, + 0.744, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.237, + 0.719, + 0.248 + ], + "angle": 0, + "content": "(c) Further analysis." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 6: Graphs show the performance of SyncNet [46] and our AVSyncNet on LRS2 GT audio-lip pairs under certain transformations. See App. A for further details. We apply lip-sync loss [46] for the analyses to fairly compare SyncNet and our AVSyncNet independent from our loss function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.789, + 0.468 + ], + "angle": 0, + "content": "AVSyncNet. Comprehensive experiments on LRS2 GT data pairs highlight AVSyncNet's superior performance (see Fig. 6). Furthermore, it demonstrates strong shift-invariance and robustness against affine transformations in the data due to AVSyncNet's design, particularly emphasizing its effectiveness in focusing on lip synchronization while being less affected by other factors. Moreover, AVSyncNet's performance is not affected by face pose unlike SyncNet [46]. Tab. 2 also demonstrates that our AVSyncNet improves the performance of our talking face generation model and works harmoniously with the proposed \\(L_{ss}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.36, + 0.507 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.675 + ], + "angle": 0, + "content": "In this paper, we improve audio-driven talking face generation by identifying problems in current approaches and mitigating them accordingly. Specifically, we introduce a silent-lip generator to mitigate lip leaking, which is a common problem that harms lip-sync and training stability. Furthermore, we propose stabilized synchronization loss along with AVSyncNet, which significantly improves the training stability, lip synchronization performance, and visual quality by solving the problems caused by lip-sync loss and SyncNet. Experimental results on benchmark datasets and a comprehensive ablation study show the merit of our method and contributions. Moreover, our detailed analyses reveal the main issues, support our claims, and validate proposed contributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.789, + 0.759 + ], + "angle": 0, + "content": "Limitations. Despite the notable improvements, SyncNet's and AVSyncNet's unstable nature should be investigated further. Moreover, face restoration sometimes causes inconsistencies in the video. Silent-lip generator makes teeth invisible in identity references, occasionally resulting in suboptimal teeth generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Ethics & Social Impact. We believe that generating lip-synchronized faces holds significant benefits across a broad spectrum of applications. However, we acknowledge its vulnerability to potential misuse, particularly deepfake generation. We will utilize Watermarking and prevent uncontrolled usage of our model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the European Commission Project Meetween (101135798) under the call HORIZON-CL4-2023-HUMAN-01-03." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.221, + 0.323, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.258, + 0.785, + 0.3 + ], + "angle": 0, + "content": "1. Afouras, T., Chung, J.S., Senior, A., Vinyals, O., Zisserman, A.: Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence 44(12), 8717-8727 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.357 + ], + "angle": 0, + "content": "2. Afouras, T., Owens, A., Chung, J.S., Zisserman, A.: Self-supervised learning of audio-visual objects from video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16. pp. 208–224. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.399 + ], + "angle": 0, + "content": "3. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.441 + ], + "angle": 0, + "content": "4. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 187-194 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.443, + 0.785, + 0.471 + ], + "angle": 0, + "content": "5. Booth, J., Roussos, A., Ponniah, A., Dunaway, D., Zafeiriou, S.: Large scale 3d morphable models. International Journal of Computer Vision 126(2), 233-254 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.472, + 0.785, + 0.499 + ], + "angle": 0, + "content": "6. Brand, M.: Voice puppetry. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 21-28 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.785, + 0.542 + ], + "angle": 0, + "content": "7. Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In: International Conference on Computer Vision (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.543, + 0.785, + 0.571 + ], + "angle": 0, + "content": "8. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronisation in the wild. arXiv preprint arXiv:2112.04432 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.571, + 0.785, + 0.613 + ], + "angle": 0, + "content": "9. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7832-7841 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.614, + 0.785, + 0.655 + ], + "angle": 0, + "content": "0. Cheng, K., Cun, X., Zhang, Y., Xia, M., Yin, F., Zhu, M., Wang, X., Wang, J., Wang, N.: Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.657, + 0.785, + 0.698 + ], + "angle": 0, + "content": "1. Chung, J.S., Huh, J., Mun, S., Lee, M., Heo, H.S., Choe, S., Ham, C., Jung, S., Lee, B.J., Han, I.: In defence of metric learning for speaker recognition. arXiv preprint arXiv:2003.11982 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "2. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.785, + 0.784 + ], + "angle": 0, + "content": "3. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 251-263. Springer (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.258, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "15. Chung, S.W., Chung, J.S., Kang, H.G.: Perfect match: Improved cross-modal embeddings for audio-visual synchronisation. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3965-3969. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.26 + ], + "angle": 0, + "content": "16. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 408-424. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.302 + ], + "angle": 0, + "content": "17. Eskimez, S.E., Maddox, R.K., Xu, C., Duan, Z.: End-to-end generation of talking faces from noisy speech. In: ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 1948-1952. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.787, + 0.344 + ], + "angle": 0, + "content": "18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.401 + ], + "angle": 0, + "content": "19. Gu, Y., Wang, X., Xie, L., Dong, C., Li, G., Shan, Y., Cheng, M.M.: Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII. pp. 126–143. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.787, + 0.457 + ], + "angle": 0, + "content": "20. Guan, J., Zhang, Z., Zhou, H., Hu, T., Wang, K., He, D., Feng, H., Liu, J., Ding, E., Liu, Z., et al.: Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1505-1515 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.458, + 0.787, + 0.5 + ], + "angle": 0, + "content": "21. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5784-5794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.543 + ], + "angle": 0, + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.571 + ], + "angle": 0, + "content": "23. Hershey, J., Movellan, J.: Audio vision: Using audio-visual synchrony to locate sounds. Advances in neural information processing systems 12 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.787, + 0.613 + ], + "angle": 0, + "content": "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.642 + ], + "angle": 0, + "content": "25. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.684 + ], + "angle": 0, + "content": "26. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. arXiv preprint arXiv:2210.07055 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "27. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "28. Jamaludin, A., Chung, J.S., Zisserman, A.: You said that?: Synthesising talking faces from audio. International Journal of Computer Vision 127, 1767-1779 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "29. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Ji, X., Zhou, H., Wang, K., Wu, W., Loy, C.C., Cao, X., Xu, F.: Audio-driven emotional video portraits. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14080-14089 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "31. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.231 + ], + "angle": 0, + "content": "32. Kadandale, V.S., Montesinos, J.F., Haro, G.: Vocalist: An audio-visual synchronisation model for lips and voices. arXiv preprint arXiv:2204.02090 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "33. Kim, Y.J., Heo, H.S., Chung, S.W., Lee, B.J.: End-to-end lip synchronisation based on pattern classification. In: 2021 IEEE Spoken Language Technology Workshop (SLT). pp. 598-605. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "34. KR, P., Mukhopadhyay, R., Philip, J., Jha, A., Namboodiri, V., Jawahar, C.: Towards automatic face-to-face translation. In: Proceedings of the 27th ACM international conference on multimedia. pp. 1428-1436 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "35. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.412 + ], + "angle": 0, + "content": "36. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3387-3396 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "37. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 106-125. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.494 + ], + "angle": 0, + "content": "38. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.55 + ], + "angle": 0, + "content": "39. Muaz, U., Jang, W., Tripathi, R., Mani, S., Ouyang, W., Gadde, R.T., Gecer, B., Elizondo, S., Madad, R., Nair, N.: Sidgan: High-resolution dubbed video generation via shift-invariant learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7833-7842 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "40. Nair, V., Hinton, G.E.: Rectified linear units improve restricted boltzmann machines. In: Proceedings of the 27th international conference on machine learning (ICML-10). pp. 807-814 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "41. Nayak, S., Schuler, C., Saha, D., Baumann, T.: A deep dive into neural synchrony evaluation for audio-visual translation. In: Proceedings of the 2022 International Conference on Multimodal Interaction. pp. 642-647 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.66 + ], + "angle": 0, + "content": "42. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.702 + ], + "angle": 0, + "content": "43. Owens, A., Efros, A.A.: Audio-visual scene analysis with self-supervised multisensory features. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 631-648 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.758 + ], + "angle": 0, + "content": "44. Papantoniou, F.P., Filntisis, P.P., Maragos, P., Roussos, A.: Neural emotion director: Speech-preserving semantic control of facial expressions in\" in-the-wild\" videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18781-18790 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "45. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2062-2070 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "46. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.204 + ], + "angle": 0, + "content": "47. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "48. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.303 + ], + "angle": 0, + "content": "49. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII. pp. 666-682. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.358 + ], + "angle": 0, + "content": "50. Shen, S., Zhao, W., Meng, Z., Li, W., Zhu, Z., Zhou, J., Lu, J.: Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1982-1991 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.388 + ], + "angle": 0, + "content": "51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.43 + ], + "angle": 0, + "content": "52. Slaney, M., Covell, M.: Facesync: A linear operator for measuring synchronization of video facial images and audio tracks. Advances in neural information processing systems 13 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "53. Song, L., Wu, W., Qian, C., He, R., Loy, C.C.: Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security 17, 585-598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.502 + ], + "angle": 0, + "content": "54. Song, Y., Zhu, J., Li, D., Wang, X., Qi, H.: Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.503, + 0.787, + 0.558 + ], + "angle": 0, + "content": "55. Stypulkowski, M., Vougioukas, K., He, S., Zieba, M., Petridis, S., Pantic, M.: Diffused heads: Diffusion models beat gans on talking-face generation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5091-5100 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.601 + ], + "angle": 0, + "content": "56. Sun, Y., Zhou, H., Wang, K., Wu, Q., Hong, Z., Liu, J., Ding, E., Wang, J., Liu, Z., Hideki, K.: Masked lip-sync prediction by audio-visual contextual exploitation in transformers. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1–9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.787, + 0.643 + ], + "angle": 0, + "content": "57. Suwajanakorn, S., Seitz, S.M., Kemelmacher-Shlizerman, I.: Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.787, + 0.685 + ], + "angle": 0, + "content": "58. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.787, + 0.742 + ], + "angle": 0, + "content": "59. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 716-731. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.771 + ], + "angle": 0, + "content": "60. Vougioukas, K., Petridis, S., Pantic, M.: Realistic speech-driven facial animation with gans. International Journal of Computer Vision 128, 1398-1413 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.841 + ], + "angle": 0, + "content": "61. Waibel, A., Behr, M., Yaman, D., Eyiokur, F.I., Nguyen, T.N., Mullov, C., Demirtas, M.A., Kantarci, A., Constantin, S., Ekenel, H.K.: Face-dubbing++: Lipsynchronous, voice preserving translation of videos. In: 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). pp. 1-5. IEEE (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.203 + ], + "angle": 0, + "content": "62. Wang, D., Deng, Y., Yin, Z., Shum, H.Y., Wang, B.: Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17979-17989 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.243 + ], + "angle": 0, + "content": "63. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14653-14662 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.298 + ], + "angle": 0, + "content": "64. Wang, J., Zhao, K., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Lipformer: High-fidelity and generalizable talking face generation with a pre-learned facial codebook. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13844-13853 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "65. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "66. Wu, H., Jia, J., Wang, H., Dou, Y., Duan, C., Deng, Q.: Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1478-1486 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "67. Yao, S., Zhong, R., Yan, Y., Zhai, G., Yang, X.: Dfa-nerf: personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "68. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "69. Yehia, H., Rubin, P., Vatikiotis-Bateson, E.: Quantitative association of vocal-tract and facial behavior. Speech Communication 26(1-2), 23-43 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.555 + ], + "angle": 0, + "content": "70. Yin, F., Zhang, Y., Cun, X., Cao, M., Fan, Y., Wang, X., Bai, Q., Wu, B., Wang, J., Yang, Y.: Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 85-101. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.583 + ], + "angle": 0, + "content": "71. Zhan, F., Yu, Y., Wu, R., Zhang, J., Lu, S.: Multimodal image synthesis and editing: A survey. arXiv preprint arXiv:2112.13592 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.637 + ], + "angle": 0, + "content": "72. Zhang, C., Zhao, Y., Huang, Y., Zeng, M., Ni, S., Budagavi, M., Guo, X.: Facial: Synthesizing dynamic talking face with implicit attribute learning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3867-3876 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "73. Zhang, Z., Hu, Z., Deng, W., Fan, C., Lv, T., Ding, Y.: Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. arXiv preprint arXiv:2303.03988 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "74. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.745 + ], + "angle": 0, + "content": "75. Zhen, R., Song, W., He, Q., Cao, J., Shi, L., Luo, J.: Human-computer interaction system: A survey of talking-head generation. Electronics 12(1), 218 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "76. Zhong, W., Fang, C., Cai, Y., Wei, P., Zhao, G., Lin, L., Li, G.: Identity-preserving talking face generation with landmark and appearance priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "77. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 9299–9306 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "D. Yaman et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "78. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4176-4186 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "79. Zhou, M., Bai, Y., Zhang, W., Yao, T., Zhao, T., Mei, T.: Responsive listening head generation: A benchmark dataset and baseline (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "80. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG) 39(6), 1-15 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_origin.pdf b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..42525528883ed4499702219f9981c71484b25166 --- /dev/null +++ b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/c95102fb-f623-4070-b6c5-8b77c44a2fa7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8ee0b49a5fa635162329e34745bf399cc94af618b2ce563ca01fc1c15a9442c +size 3238280 diff --git a/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/full.md b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae271f83cda3444095d0b44dbca76019926ddc29 --- /dev/null +++ b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/full.md @@ -0,0 +1,316 @@ +# Audio-driven Talking Face Generation with Stabilized Synchronization Loss + +Dogucan Yaman1, Fevziye Irem Eyiokur1, Leonard Bärmann1, Hazim Kemal Ekenel2, and Alexander Waibel1,3 + +$^{1}$ Karlsruhe Institute of Technology, Karlsruhe, Germany +2 Istanbul Technical University, Istanbul, Turkey +$^{3}$ Carnegie Mellon University, Pittsburg PA, USA dogucan.yaman@kit.edu + +Abstract. Talking face generation aims to create realistic videos with accurate lip synchronization and high visual quality, using given audio and reference video while preserving identity and visual characteristics. In this paper, we start by identifying several issues with existing synchronization learning methods. These involve unstable training, lip synchronization, and visual quality issues caused by lip-sync loss, SyncNet, and lip leaking from the identity reference. To address these issues, we first tackle the lip leaking problem by introducing a silent-lip generator, which changes the lips of the identity reference to alleviate leakage. We then introduce stabilized synchronization loss and AVSyncNet to overcome problems caused by lip-sync loss and SyncNet. Experiments show that our model outperforms state-of-the-art methods in both visual quality and lip synchronization. Comprehensive ablation studies further validate our individual contributions and their cohesive effects. + +Keywords: Talking face generation $\cdot$ Lip synchronization $\cdot$ Lip leaking + +# 1 Introduction + +Audio-driven talking face generation aims at generating a video with respect to given face and audio sequences. The objective is to achieve synchronized lip movements corresponding to the provided audio while preserving the identity and visual details. This task has recently attracted significant attention due to its versatile applications, including dubbing in the film industry, online education, enhancing video conferencing, and dubbing for various types of videos [71, 75]. + +The talking face generation task comprises two primary aspects: (1) lip synchronization and (2) visual quality of the face. Since lips that are out-of-sync with audio can be easily identified by humans, synchronized lips are key to achieving natural and realistic talking-face generation. For this, the primary solution is to evaluate audio-lip synchronization and use it as a training objective. Wav2Lip [46] introduced an improved version of SyncNet [14], a pretrained model designed to measure audio-visual synchronization. This model is also used during training to extract features and compute lip-sync loss. Subsequently, many approaches have employed improved SyncNet [46] to guide the + +![](images/6647a73177df405c2a63197b23f09cfd59134fa1c42614fc06ded5f83ba822a4.jpg) +(a) SyncNet [46] + +![](images/2e2d05c640144d1b3a3edfb6edb5388406a3a9afbb650b2fb588f3d46d64aee2.jpg) +(b) AVSyncNet (ours) +Fig. 1: (a, b) Cosine similarity between GT audio-lip pairs on random LRS2 samples, showcasing the instability of SyncNet and more robust performance of AVSyncNet. (c) illustrates full mouth region / lip leaking from the reference, pose effect from the reference, and similar identity reference-target image scenarios. + +![](images/8847030a5549f6e1f2bba0af662de421fec6c445d6b2393eea2c472792327752.jpg) +(c) Leaking problems + +model throughout training. The main concept is to provide audio-video pairs to SyncNet, which extracts features through its image and audio encoders. Subsequently, cosine similarity with binary cross-entropy loss is measured between these two features [46]. Since the model was trained in this manner, a higher cosine similarity is expected when the lips are synchronized with the given audio. Although existing approaches with this strategy mostly surpass other methods in lip synchronization, there are still challenges that must be addressed to enhance performance. + +In this work, we identify two main challenges in existing approaches that restrict models from achieving satisfactory performance in both lip synchronization and visual quality: SyncNet instabilities and lip leaking. First, in alignment with previous research [41], we observe instabilities in the performance of SyncNet [46] on pairs of ground-truth (GT) lip and audio (see Fig. 1a and App. A). Thus, when SyncNet is utilized as part of talking face generation training, it might provide an inadequate training signal, assigning low similarity scores to generated images even when the lips are synchronized. This problem causes unstable training, degrading the lip generation capability of the network, thereby ending up in out-of-sync lips or suboptimal lip-sync. Furthermore, the lip-sync loss [46] and reconstruction losses are conflicting [39]. This leads models to have either poor lip synchronization or degraded visual quality (sometimes even both), escalating further when lip-sync loss and SyncNet are employed with high-resolution (HR) data [39,73]. To address these problems, we first improve SyncNet and introduce AVSyncNet, demonstrating a more robust performance (see Fig. 1b) and also overcoming poor shift-invariance characteristics of SyncNet (see Fig. 6a). However, despite improved performance, the instability problem is not fully solved (see Fig. 1b). To overcome this problem further, we also introduce a stabilized synchronization loss. Specifically, instead of directly using the similarity of the (generated lips, audio) pair, we calculate the difference of the similarities between (GT lips, audio) and (generated lips, audio). We hypothesize that this alleviates the described problems since it guides the model to generate a lip movement with a similar synchronization score as for the GT face. Together with AVSync- + +Net, this method empirically enhances the lip synchronization performance as well as avoids visual quality issues caused by misleading lip-sync loss. + +Second, we address another main problem in the talking face generation literature: lip leaking. The current gold standard in 2D-based methods is to input a bottom-half masked face image ("pose reference"), wherein the model is expected to generate the input face with proper lip movements. However, because of masking, the model requires a reference image to retain the identity and texture in the masked part. This is achieved by randomly choosing a face image from a different part of the video sequence, referred to as "identity reference". However, this introduces a new challenge: As it is randomly selected, the lip movements of the identity reference can frequently be similar to the GT lips during training (see last row of Fig. 1c). Hence, for faster convergence, the model tends to replicate the lip movements from the identity reference, resulting in poor lip synchronization or complete out-of-sync output. By following the literature [41,46], we term this phenomenon as lip leaking. To tackle this problem, we propose a simple yet effective technique: silent-lip generator. The concept involves modifying the identity reference to make the lips closed (thus "silent-lip") before feeding it to the talking face generator network. This method ensures consistently closed lips in the identity reference, effectively mitigating the lip leaking problem. Our contributions are as follows: (i) We identify and analyze various fundamental problems that harm lip synchronization learning and also cause visual quality issues. (ii) We present a robust and shift-invariant AVSyncNet, and stabilized synchronization loss to overcome the problems caused by lip-sync loss and SyncNet. (iii) We present a silent-lip generator to generate an identity reference with closed lips before feeding the talking face generator to alleviate the lip leaking. + +# 2 Related Work + +Audio-driven Talking Face Generation. Initial studies mapped audio features to time-aligned facial motions [69] or predicted facial motions by HMM [6]. In [57], videos were generated by finding the images most aligned with the audio. ATVGNet [9] transfers audio to facial landmarks and uses pixel-wise loss with an attention mechanism to avoid the jittering problem and leaking of irrelevant speech. [16] and [80] use facial landmark representation to synthesize faces with synchronized lips. Recent papers address the task as conditional inpainting by masking the bottom half of the input face and feeding the network an identity reference from another time step of the same video, along with the audio segment. Wav2Lip [46] proposes SyncNet and lip-sync loss to predict lip synchronization and achieves significant performance. PC-AVS [78] introduces a pose-controllable audio-visual system, while GC-AVT [36], EAMM [29] and EVP [30] control the emotion by utilizing an emotion embedding. On the other hand, SyncTalkFace [45] uses Audio-Lip Memory to store lip motion features and retrieves them as visual hints for better synchronization. VideoReTalking [10] proposes to manipulate the reference image to have a face with canonical expression to alleviate the sensitivity of the model against the identity reference image. + +Similarly, we introduce a silent-lip generator to implicitly learn to manipulate the lips of the identity reference to mitigate the lip leaking problem. Compared to the method proposed in [10], our method works more efficiently to preserve the identity and visual quality. More recently, LipFormer [64] uses a pre-learned facial codebook to generate HR videos, while DINet [73] proposes to use a deformation module to obtain deformed features to enhance lip synchronization and head pose alignment. IPLAP [76] shows satisfactory visual quality and stability by using intermediate landmark representation and motion field. Recently, TalkLip [63] introduces a global audio encoder, trained with self-supervised learning, to encode features by considering the entire content of the audio. Besides, they propose to use lip reading during the training as well as in the evaluation to control whether the content is preserved. SIDGAN [39] performs important analyzes, and introduces shift-invariant APS-SyncNet and training objectives along with the coarse-to-fine pyramid model for HR dubbed video generation. Recent works use diffusion model because of its stability and accuracy [50, 55]. Finally, in [61], talking face generation is used as a part of a full system to perform end-to-end face dubbing, involving speech recognition, translation, speech, and video generation. In contrast to the above, 3D-based and Neural Radiance Fields-based (NeRFs) methods typically generate the entire head, rather than just manipulating 2D images of the face, often involving manipulation of pose, emotion, and 3D face model [4, 5, 21, 37, 44, 49, 53, 58, 59, 62, 66–68, 70, 72, 74, 77, 80]. Similarly, portrait animation aims at utilizing a single input image to generate a video by predicting the pose and expression of the subsequent frames, along with ensuring synchronized lips. Nevertheless, these tasks significantly differ from face dubbing in terms of methodology, task definition, goals, and real-world applications. + +Lip Synchronization. Some earlier works [23, 52] used hand-crafted features and statistical models to evaluate lip synchronization. Recent studies proposed to use mutual information between audio-visual features to produce sync or out-of-sync output for sound [8,26,43] or speech [2,12,15,32,33]. While some methods learn lip synchronization implicitly [9,21,28,34,57,66,74], other methods employ distance between landmarks or facial parameters [30,44,53,80]. Contrarily, the majority of works [17,20,36,45,46,54,56,60,62-64,67,73,77,78] extract audio-visual features with an additional network (mostly SyncNet [46]) for audio-lip synchronization prediction. Afterwards, while some of these methods utilize lip-sync loss [46], others benefit from contrastive learning by using infoNCE [42]. We follow a similar strategy by utilizing an additional network to calculate a synchronization loss. We propose a robust and shift-invariant AVSyncNet along with a stabilized synchronization loss to overcome existing challenges. + +# 3 Proposed Approach + +# 3.1 Talking Face Generation + +We propose an audio-driven talking face generation model $G_{L}$ with enhanced lip synchronization. As shown in Fig. 2a, our model incorporates: 1) an audio + +![](images/7bddec5e1b8bf354bf2c1facb6932189a4d173659ff17d8b7ae18c7a6091a6b3.jpg) +(a) Talking face generation model +(b) FD block + +![](images/dc4063384e5965c490218073f5107120f76fbb60a62779176997d73f71f710fe.jpg) +Fig. 2: Talking face generation model $G_{L}$ (a) and face-decoding (FD) block (b). Our model receives a pose reference sequence, mel-spectrogram of an audio snippet, and a silent identity reference, that is generated by our silent-lip generator $G_{S}$ , aiming to alleviate lip leaking problem. The model then synthesizes the talking face sequence to ensure lip synchronization. Subsequently, the employed loss functions are computed. + +encoder for processing the audio snippet, 2) an identity encoder for addressing an identity reference image, and 3) a pose encoder for utilizing a pose reference. + +Audio Encoder. The audio encoder $E_{A}$ generates phoneme-level embeddings, serving as conditions for the face generator to generate lip movements accurately. Our audio encoder extracts embeddings $F_{A} = E_{A}(A) \in \mathbb{R}^{1 \times 1 \times 512}$ from the given mel-spectrogram $A$ , representing the driving audio. In contrast to existing methods [10, 46, 73, 76], we propose using a pretrained, frozen audio encoder, concurrently trained with a face encoder to learn lip synchronization similar to the objective in SyncNet [46]. Thus, we can obtain improved embeddings during the talking face generation training by leveraging the capacity of the pretrained robust audio encoder. We obtain the best score with such frozen audio encoder. + +Face Encoder. In accordance with the gold standard in the literature, we utilize an identity reference, $I^R$ , and a pose reference, $I$ , as inputs to the model. The identity reference is a face image of the subject that provides identity information. It is different from the pose reference and randomly selected from the same video. The pose reference is identical to the target image, except for the bottom half, which is masked, as the model is designed to focus on generating lip movements. Unlike most conventional methods that employ a joint encoder for processing identity and pose references, we use individual encoders to allow each encoder to focus solely on their respective tasks [39]. Therefore, we utilize two parallel CNN-based face encoders to process identity and pose references individually. This approach yields better feature representation and ultimately leads to improved performance. + +![](images/a4914333acf4b1932424f458c6b84e04a0bbe41ccf0067ae0283a8ac6e78a5a9.jpg) +(a) Silent-lip generation model $G_{S}$ + +![](images/eeb47c49fc2ae167e741c0a489896d3cdd2383635a7473f2246d9b316b85104b.jpg) +(b) AVSyncNet +Fig. 3: $G_{S}$ in inference (a) and AVSyncNet training pipeline (b). + +# 3.2 Silent-Lip Generator + +Talking face generation involves using an identity reference to preserve the identity in the generated image. This is particularly important because the bottom half of the pose reference, specifically the mouth region, is masked. This masking is necessary as we aim to synthesize appropriate lip movements corresponding to the provided audio. However, models are unintentionally affected by the lip movement of the identity reference, rather than solely gathering identity information (see Fig. 1c and App. C for details). This behavior, which we refer to as lip leaking, leads to poor lip synchronization or occasionally even non-converged training. We hypothesize that there are two main reasons for this: First, the lip movement of the identity reference may occasionally resemble that of the lips in the target image (see last row of Fig. 1c). Hence, the model can lower the synchronization loss more quickly by undesirably replicating the lip movements of the identity reference. Second, the diversity of lip movements in the identity reference may yield the model to seek a correlation with the target lips. This causes a challenging disentanglement task for the identity encoder —namely distinguishing identity information and lip movement. + +To mitigate the aforementioned issue, we propose to use an additional model, called silent-lip generator $G_{S}$ , prior to the talking face generation model $G_{L}$ , aimed at modifying the lip shape of the identity reference. Specifically, we reconstruct the input face with closed, flat lips. This strategy reduces the likelihood of having lips similar to the target face in the identity reference and resolves the issue of diverse lip movements. Consequently, the model no longer replicates the lips from the identity reference, resulting in stable training & improved lip sync. + +To implement $G_{S}$ , we structure the task and the model similar to the talking face generation. Specifically, we input a bottom-half masked pose reference, an identity reference, and an audio snippet to train a GAN [18] for generating a talking face. The model is trained to reconstruct the face using both the pose and identity references. Notably, we exclude any synchronization loss and SyncNet during training and train the model under the weak condition. Consequently, the + +model focuses solely on generating lip movements without synchronization when speech is present. Therefore, it implicitly learns to generate closed lips when the input audio is silent as shown in Fig. 3a. Furthermore, by eliminating synchronization loss, we overcome unstable training and the issue of lip leaking from the identity reference for this network. We choose this approach for its efficient utilization of the same model and data. Given the scarcity of closed-lip faces in the dataset, we avoid using these frames directly to maintain effective training and generalization. Please note that we use the same architecture for $G_{S}$ and $G_{L}$ . We initially train $G_{S}$ separately on the same training data (LRS2) and then incorporate it into the training of $G_{L}$ without further finetuning. Specifically, we then only pass silent audio to $G_{S}$ , so that it modifies given identity references to have silent lips for subsequent use as identity reference for $G_{L}$ . + +# 3.3 Video Generation + +For the talking face video generation, we employ the aforementioned components. Initially, we employ our pretrained silent-lip generator $G_{S}$ to synthesize the identity reference with closed lips. Subsequently, we input this silent identity reference to the identity encoder and the pose reference to the pose encoder in the talking face generation model $G_{L}$ . Similarly, we provide the mel-spectrogram of the corresponding audio snippet to the audio encoder. Next, we concatenate the embeddings from the identity and pose encoders, along with the depth dimension, and pass it through a $1 \times 1$ convolution layer to reduce the depth. Finally, we concatenate this feature representation with the audio embeddings before feeding them into the face generator. The face generator generates the entire face with accurate lip movements by preserving the identity and the pose. + +We illustrate the talking face generation model $G_{L}$ in Fig. 2a. We integrate the U-Net architecture [47] for our overall design, leveraging its adequate performance in reconstruction tasks while ensuring computational efficiency. Our identity encoder and pose encoder share the same architecture, consisting of consecutive face-encoding (FE) blocks. Each block has a strided-convolutional layer followed by two non-strided convolutional layers, each paired with a batch normalization layer [27] and a ReLU activation function [35,40]. We also use the residual connection strategy [22] by summing the input and output of each block before forwarding it to the next layer. On the other hand, our face generator has consecutive face-decoding (FD) blocks. As shown in Fig. 2b, within each block, we utilize a transposed-convolutional layer, followed by two convolutional layers incorporating batch normalization and ReLU activation function. Moreover, we apply a skip connection between the reciprocal layers of the face encoders and the decoder to retain high-level features and enhance the training stability. + +GAN Loss. To train our model, we utilize GAN loss [18] and employ a discriminator, which is a straightforward CNN-based binary classification network to distinguish real and fake samples, designed with a balanced architecture aligned with our face encoders. We benefit from consecutive strided-convolutional layers followed by the Leaky ReLu activation function and spectral normalization [38]. + +Reconstruction Loss. We employ L1 loss in pixel space $L_{pixel} = ||I' - I^{GT}||_1$ between the generated $I'$ and the target faces $I^{GT}$ to ensure consistency in areas outside the lips and maintain the illumination condition. We further utilize perceptual loss [31] based on the pretrained VGG-19 [51]: + +$$ +L _ {p e r} = \sum_ {i = 1} ^ {5} c _ {i} | | \mathrm {V G G} ^ {\phi_ {i}} \left(I ^ {\prime}\right) - \mathrm {V G G} ^ {\phi_ {i}} \left(I ^ {G T}\right) | | _ {1} \tag {1} +$$ + +where $c_{i}$ are weight coefficients from [31], and $\phi$ indicates the set of VGG layers. + +Adaptive Triplet Loss. Although our goal is to capture visual details from the identity reference, we observe that the model occasionally focuses on the visual details (e.g., illumination, pose) in the identity reference excessively and this tendency could potentially degrade the quality and stability of the generated face sequence, resulting in suboptimal performance. To tackle this, we exploit a triplet loss strategy [48], aiming to minimize the distance between the generated face and GT, while maximizing the distance between the generated face and the identity reference. However, the random selection of identity references increases the probability of choosing an image that closely resembles the GT. This scenario poses a challenge for the vanilla triplet loss, potentially degrading training and resulting in poor visual and pose quality. To mitigate this, we introduce an adaptive triplet loss that considers the similarity between the identity reference and GT during loss computation to alter its effect. The formula is as follows: + +$$ +L _ {a t} = \left[ D (\mathrm {V G G} (I ^ {\prime}), \mathrm {V G G} (I ^ {G T})) - \frac {D (\mathrm {V G G} (I ^ {\prime}) , \mathrm {V G G} (I ^ {R}))}{D (\mathrm {V G G} (I ^ {G T}) , \mathrm {V G G} (I ^ {R}))} + \alpha \right] _ {+} \tag {2} +$$ + +where $[\cdot]_{+} = \max(\cdot, \epsilon)$ , $D$ represents the L2 distance, and we empirically choose $\alpha = 1$ . In this loss, we leverage the ratio of the similarity between the generated image and identity reference to that of the GT and the identity reference to adjust the loss value. As the identity reference becomes more similar to GT, the impact of the distance between the generated image and identity reference on the loss diminishes, since expecting a high distance in this case is not reasonable. Since our objective is to incorporate visual details from the identity reference, we opt for a very low coefficient to avoid conflicting with the primary goal. + +# 3.4 Learning Synchronization + +The lip-sync loss [46] serves as a method to calculate synchronization between audio and video. Leveraging the pretrained SyncNet [46] for feature extraction from both audio and video inputs demonstrates reasonable performance in learning lip synchronization. However, the SyncNet is significantly unstable when measuring this similarity. Our evaluation using SyncNet on GT training data reveals notable fluctuations in cosine similarity between video and audio, contrary to the expected high scores (see Fig. 1a and App. A for details). Therefore, this + +provides conflicting information to the system, resulting in poor lip synchronization, unstable training, and degraded visual quality. To tackle this problem, we present a more accurate, shift-invariant, and robust version of SyncNet, named AVSyncNet, and introduce a novel, stabilized synchronization loss. + +AVSyncNet. We employ a ResNet-50-based [22] image encoder, known for its superior performance in face recognition, alongside a ResNetSE-34-based audio encoder [11], which is a modified version of ResNet-34 [22] designed to handle spectrogram inputs. This design makes the AVSyncNet model robust and shift-invariant compared to SyncNet [46]. We train our model on the LRS2 training data [1], calculating the cosine similarity between audio and lip features, followed by a binary cross-entropy loss, as shown in Fig. 3b. During each training step, we provide a set of images (5 images) along with the corresponding audio. For negative samples, we randomly select an audio snippet from the non-overlapping part of the video. Please note that as we feed the bottom half of the face to the image encoder, we adapt the first layer of ResNet-50 for an input size of $112 \times 224$ . + +Stabilized Synchronization Loss $(L_{ss})$ . Although AVSyncNet shows improved performance compared to SyncNet [46] and alleviates existing instability problem that harms lip-sync and visual quality, the unstable performance is not fully solved due to the inherent challenges of the task (see Fig. 1b and App. A, B). Therefore, we introduce a stabilized synchronization loss $(L_{ss})$ to improve the lip synchronization performance further in conjunction with AVSyncNet by providing more stable and precise supervision. The formula is shown below: + +$$ +L _ {s s} = - \log \left(1 - \frac {\left| x - y \right| + \epsilon}{\left| x - y \right| + \left| y - d \right| + \epsilon}\right) \tag {3} +$$ + +$$ +x = \operatorname {A V S I M} \left(I ^ {\prime}, A\right), \quad y = \operatorname {A V S I M} \left(I ^ {G T}, A\right), \quad d = \operatorname {A V S I M} \left(I ^ {R}, A\right) \tag {4} +$$ + +where $I^{\prime}, I^{GT}$ , and $I^R$ are generated, GT, and identity reference lips, respectively, while $A$ is their corresponding audio. AVSIM( $I, A$ ) indicates the audio-visual similarity between a face image (bottom half only, i.e. lips) and audio, given by the cosine similarity of extracted image and audio features $\phi_{AVS}^V(I)$ , $\phi_{AVS}^A(A)$ of the respective AVSyncNet encoder. + +In this formulation, $x$ followed by cross-entropy loss denotes the lip-sync loss [46]. However, to address the unstable and fluctuating performance, we utilize the relative distance in similarity between GT lips-audio and generated-audio pairs. Sometimes, randomly selected reference images may have a lip + +movement similar to that of the target image. We have already introduced the silent-lip generator $G_{S}$ to mitigate this situation. However, AVSyncNet, while less severe than SyncNet, might still be unstable (e.g., unexpectedly high scores with incorrect pairs or vice versa), resulting in a minor lip leaking and stability issue. This problem arises when AVSyncNet assigns an erroneously high score to silent lip-audio pairs as well as when the target lip shape looks similar to closed lips. To mitigate this, we inject the similarity score between the identity reference and audio into the formulation. Specifically, we penalize the model more when the identity reference-audio pair shows higher similarity. + +# 3.5 Implementation Details + +Combining all the presented contributions, the total loss is: + +$$ +L = L _ {G A N} (G, D) + \lambda_ {1} L _ {p i x e l} (G) + \lambda_ {2} L _ {p e r} (G) + \lambda_ {3} L _ {s s} (G) + \lambda_ {4} L _ {a t} (G) \tag {5} +$$ + +where $G$ and $D$ indicate generator and discriminator outputs, respectively. We empirically found the best coefficients as $(\lambda_1, \dots, \lambda_4) = (10, 1, 2, 0.5)$ . Please note that this is the loss function for training $G_L$ , while for $G_S$ we set $\lambda_3 = \lambda_4 = 0$ . + +We process videos by using 5 consecutive frames in each step to consider the temporal information. We detect faces with FAN [7], followed by acquiring tight crops and resizing to $96 \times 96$ , as faces in LRS2 [1] are of low resolution. Our audio encoder receives a mel-spectrogram of size $16 \times 80$ derived from $16\mathrm{kHz}$ audio with a window size of 800 and a hop size of 200. We employ the Adam optimizer with $(\beta_{1},\beta_{2}) = (0.5,0.999)$ and set the learning rate to $1 \times 10^{-4}$ for all models. Training our AVSyncNet is done similarly to SyncNet [46] on the LRS2 dataset, and then we freeze the audio encoder of AVSyncNet and use it in the training of $G_{S}$ and $G_{L}$ . At the end of talking face generation, we apply a post-processing step by using VQFR [19] to enhance the visual quality and the resolution, aiming to achieve HR videos. We train and test our models with a single NVIDIA RTX A6000 GPU. + +# 4 Experimental Results + +Dataset. We trained our silent-lip generator and talking face generator using Lip Reading Sentence 2 (LRS2) [1] training set as it is a well-known benchmark with extensive subject diversity. The evaluation was carried out on the LRS2 test set and extended to the LRW [13] test set and HTDF dataset [74] to demonstrate performance on unseen data. + +Metrics and Baseline. For visual quality, we employ widely used metrics: FID [24], SSIM [65], and PSNR. We also use inter-frame consistency (IFC), presented as a training objective in [79]. This is achieved by calculating the difference between the distances of the consecutive frames in the generated and GT videos. To evaluate lip synchronization, we follow the literature and use Landmark Distance (LMD) [9] in the mouth region and LSE-C & LSE-D metrics [46] + +Table 1: Quantitative results on the test sets of LRS2 and LRW. + +
LRS2LRW
MethodSSIM ↑PSNR ↑FID ↓IFC ↓LMD ↓LSE-C ↑LSE-D ↓SSIM ↑PSNR ↑FID ↓IFC ↓LMD ↓LSE-C ↑LSE-D ↓
Wav2Lip [46]0.8626.537.050.212.387.596.750.8525.146.810.202.147.496.51
PC-AVS [78]0.7328.2418.400.461.936.417.520.8132.2514.270.381.426.537.15
EAMM [29]0.6921.0184.650.513.543.319.930.7126.2244.160.482.614.329.04
VideoReTalking w/ FR [10]0.8425.589.280.222.617.496.820.8727.115.300.232.396.597.12
DINet [73]0.7824.354.260.252.305.378.370.8827.508.170.221.965.249.09
TalkLip [63]0.8626.114.940.242.348.536.080.8626.3415.730.261.837.286.48
IPLAP [76]0.8729.674.100.202.116.497.160.9130.458.400.211.645.947.76
Ours w/o FR0.9532.643.830.161.138.416.030.9231.454.460.181.227.866.24
Ours w/ FR (VQFR)0.9031.805.230.271.368.525.830.9030.217.050.211.417.926.00
+ +to measure the confidence and distance scores through a pretrained model [14]. We choose SOTA methods with publicly available codes and models to compare them fairly under the same conditions, as the implementation of the metrics and face cropping strategy before computing the metrics affect the scores. + +# 4.1 Quantitative Results + +In Tab. 1, we present quantitative results on test sets of two benchmark datasets, namely LRS2 and LRW. We achieve state-of-the-art results in all visual quality metrics excluding PSNR on LRW, which is a less informative metric compared to SSIM and FID. On IFC, we similarly outperform all compared methods, indicating that our model generates the most consistent videos in terms of temporal information and stability of the faces. + +In lip synchronization evaluation, we obtain the best performance in LMD. Nevertheless, it is important to highlight that LMD is sensitive to the changes in the image, as it does not disentangle synchronization and visual stability. For instance, affine transformations impact the LMD score even when the lips are synchronized, and vice versa. On the LRW dataset, we achieve SOTA results with more reliable confidence and distance metrics for lip synchronization: LSE-C & D. On the LRS2 dataset, TalkLip yields a slightly better score than our model with the LSE-C metric. Nevertheless, we outperform TalkLip and achieve a SOTA result with the LSE-D metric. All these results indicate the accuracy of our method in terms of visual quality and lip synchronization. Similarly, we achieve SOTA results for most metrics on unseen HDTF [74], see App. E.1. + +# 4.2 Qualitative Results + +In Fig. 4, we demonstrate a qualitative comparison with SOTA models and GT data. We use their respective publicly available models and generate videos from the HDTF dataset [74] to compare the models on unseen data since the presented models were trained on the LRS2 dataset, except for DINet (trained with HDTF). Due to SyncNet and lip-sync loss, TalkLip and Wav2Lip encounter generalization issues, sometimes leading to visual artifacts in the mouth region or face boundaries, especially when the pose of the identity reference differs from the pose reference, despite generating accurate lip movements. This observation + +![](images/f13553b11fc2a0811d7ce4e16d8f88304a730a7dc93c5a653549df43b3c96dbf.jpg) +VideoReTalking (Cheng et al. 2022) +DINet (Zhang et al. 2023) +TalkLip (Wang et al. 2023) +IP-LAP (Zhong et al. 2023) +Ours w/o FR +Ground Truth +Fig. 4: Qualitative comparison with the SOTA methods. Reference videos (from HDTF [74]) are randomly selected and not seen during training by our model. For more images and videos, please check App. E, F and https://yamand16.github.io/TalkingFaceGeneration/. + +![](images/7ad44190c5c745ee88774d909f76ede959f038ddd976d407a9ad2f6304327dad.jpg) +Wav2Lip (Prajwal et al. 2020) + +Table 2: Ablation studies on the LRS2 test set. See the text for details. + +
AblationSetupMethodPSNR ↑ SSIM ↑ FID ↓ LMD ↓ LSE-C ↑ LSE-D ↓ IFC ↓
ComponentsAGL+LSa26.3490.85312.252.4087.116 ± 1.927.396 ± 1.03
BA + Ea,S26.6140.8689.822.3257.271 ± 1.767.106 ± 0.98
CA + Ea,W26.5900.86910.562.2787.220 ± 1.757.158 ± 0.99
DB + GS27.1800.8728.161.7417.752 ± 1.716.413 ± 0.95
EGL + Ea,S + GS + LSs31.1660.9255.271.1408.370 ± 1.166.032 ± 0.59
FE + Lt30.6580.9176.241.2508.260 ± 1.346.176 ± 0.64
GE + Lat32.7550.9494.021.1358.382 ± 1.166.057 ± 0.61
HG w / AVSyncNet32.6400.9523.831.1308.410 ± 0.976.037 ± 0.55
Post-processingFR1Setup H + GPEN28.9910.91958.771.1977.6256.457
FR2Setup H + GFPGAN31.1690.91613.071.2197.6246.496
FR3Setup H + VQFR: full model31.8060.9055.231.3658.5285.838
Silent face generationVRT-SVideoReTalking silent data22.1240.64633.60--0.463
Ours-SOur silent data (GS)33.3280.9514.41--0.141
+ +clearly validates the motivation of our contributions. In contrast, our model generates consistent faces with comparably fewer artifacts, featuring appropriate lip movements that align with both the GT faces and the corresponding audio. However, VideoReTalking demonstrates comparable lip synchronization and visual quality performance to our model. On the other hand, IP-LAP shows sufficient visual quality, while less accurate lip synchronization. + +# 4.3 Ablation Study + +Tab. 2 and Fig. 5a show a comprehensive ablation study on the LRS2 test set, analyzing the individual impact of our contributions. We first train our $G_{L}$ model using SyncNet [46] and lip-sync loss [46] as a baseline. As expected, we encountered several issues with unstable training. Once our model converged after several random seeds, the results (Setup A) show that lip-sync loss can achieve + +![](images/186ab939a02d824c0535dc5009730c3f298013352f616b8d6517c19652aa79e2.jpg) +(a) Components. + +![](images/c56851318c7dd8c76bee7484ffb313fc82bf129adfee63911ef80a4eb9e03faf.jpg) + +![](images/3ea110fe614b3cb5d7616ff8a89b4325ab1a3455cb1d39242400b6ba5f6374fb.jpg) + +![](images/1055a1553975ccbadddd88742894d815f9c8cb82185b62118ce28e857b118079.jpg) +(c) Generated samples for challenging cases: Pose, teeth, beard. + +![](images/54419d10d011a30a52a30eda9f7adb39c169b01e7c7808035f8dc405e9a74aae.jpg) +(b) Face restoration methods. + +![](images/8c8874f5699470fd9854a1d10024b1607b80b9ca6d9ac031266f60e8f73b5b8d.jpg) +(d) Silent face comparison with VRT [10]. © European Central Bank (CC BY). +Fig. 5: Ablation studies of components (a), face restoration methods (b), and silent face generation (d). (c) demonstrates generated images in challenging cases. + +decent synchronization performance, despite lower visual quality and training stability issues. In this setup, we train the audio encoder as a part of $G_{L}$ . Replacing this with our pretrained audio encoder enhances the synchronization and visual quality (Setup B). For further comparison, we also utilize the audio encoder of the Wav2Vec2 [3], presented in Setup C. However, it slightly decreases the scores, which validates our hypothesis about training the audio encoder for synchronization purposes. Thus, we continue with Setup B and add our silent-lip generator $G_{S}$ to generate silent identity references (Setup D). $G_{S}$ alleviates the lip leaking problem, makes the training more stable, and improves the scores noticeably. We further replace lip-sync loss with our stabilized synchronization loss, yielding drastically improved synchronization and visual quality scores (Setup E). Moreover, we observed that Setup E shows almost no instability in the training. In Setup F, we train Setup E including vanilla triplet loss and it enhances neither visual quality nor synchronization; in fact, it even causes detrimental effects. This shows the necessity of modifying the triplet loss and introducing the adaptive triplet loss. In Setup G, replacing vanilla triplet loss with adaptive triplet loss demonstrates a slight improvement in visual quality, while not having a negative impact on lip synchronization. In Setup H, we switch SyncNet with AVSyncNet and achieve slightly better visual quality and lip synchronization. + +We compare $G_{S}$ with VRT [10] silent face generation approach in Tab. 2. Our model surpasses VRT quantitatively and qualitatively (see Fig. 5d), preserving the visual details and identity while modifying the lips. + +We compare different face restoration methods as post processing and present the results in Tab. 2 and Fig. 5b. VQFR surpasses other methods in preserving lip synchronization as well as FID and PSNR. However, GPEN shows better performance in the remaining metrics (see App. F.2 for details). In summary, we employ VQFR for post-processing in our full model. + +![](images/d386668efa85fb07cc9df1121e5b27b15feca8a88f5a440f1b773ab58667a0e9.jpg) +(a) Shifting [39]. +Fig. 6: Graphs show the performance of SyncNet [46] and our AVSyncNet on LRS2 GT audio-lip pairs under certain transformations. See App. A for further details. We apply lip-sync loss [46] for the analyses to fairly compare SyncNet and our AVSyncNet independent from our loss function. + +![](images/5bc8583d456ce149bb0e7de200d22a85fa0f94408843731876bb88053aa0f174.jpg) +(b) Rotation. + +![](images/9305dd95a910a0bb9e933c9f260f615520d191e8bd05942eedca21ada26ab331.jpg) +(c) Further analysis. + +AVSyncNet. Comprehensive experiments on LRS2 GT data pairs highlight AVSyncNet's superior performance (see Fig. 6). Furthermore, it demonstrates strong shift-invariance and robustness against affine transformations in the data due to AVSyncNet's design, particularly emphasizing its effectiveness in focusing on lip synchronization while being less affected by other factors. Moreover, AVSyncNet's performance is not affected by face pose unlike SyncNet [46]. Tab. 2 also demonstrates that our AVSyncNet improves the performance of our talking face generation model and works harmoniously with the proposed $L_{ss}$ . + +# 5 Conclusion + +In this paper, we improve audio-driven talking face generation by identifying problems in current approaches and mitigating them accordingly. Specifically, we introduce a silent-lip generator to mitigate lip leaking, which is a common problem that harms lip-sync and training stability. Furthermore, we propose stabilized synchronization loss along with AVSyncNet, which significantly improves the training stability, lip synchronization performance, and visual quality by solving the problems caused by lip-sync loss and SyncNet. Experimental results on benchmark datasets and a comprehensive ablation study show the merit of our method and contributions. Moreover, our detailed analyses reveal the main issues, support our claims, and validate proposed contributions. + +Limitations. Despite the notable improvements, SyncNet's and AVSyncNet's unstable nature should be investigated further. Moreover, face restoration sometimes causes inconsistencies in the video. Silent-lip generator makes teeth invisible in identity references, occasionally resulting in suboptimal teeth generation. + +Ethics & Social Impact. We believe that generating lip-synchronized faces holds significant benefits across a broad spectrum of applications. However, we acknowledge its vulnerability to potential misuse, particularly deepfake generation. We will utilize Watermarking and prevent uncontrolled usage of our model. + +Acknowledgements. This work was supported in part by the European Commission Project Meetween (101135798) under the call HORIZON-CL4-2023-HUMAN-01-03. + +# References + +1. Afouras, T., Chung, J.S., Senior, A., Vinyals, O., Zisserman, A.: Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence 44(12), 8717-8727 (2018) +2. Afouras, T., Owens, A., Chung, J.S., Zisserman, A.: Self-supervised learning of audio-visual objects from video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16. pp. 208–224. Springer (2020) +3. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020) +4. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 187-194 (1999) +5. Booth, J., Roussos, A., Ponniah, A., Dunaway, D., Zafeiriou, S.: Large scale 3d morphable models. International Journal of Computer Vision 126(2), 233-254 (2018) +6. Brand, M.: Voice puppetry. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 21-28 (1999) +7. Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In: International Conference on Computer Vision (2017) +8. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronisation in the wild. arXiv preprint arXiv:2112.04432 (2021) +9. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7832-7841 (2019) +0. Cheng, K., Cun, X., Zhang, Y., Xia, M., Yin, F., Zhu, M., Wang, X., Wang, J., Wang, N.: Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +1. Chung, J.S., Huh, J., Mun, S., Lee, M., Heo, H.S., Choe, S., Ham, C., Jung, S., Lee, B.J., Han, I.: In defence of metric learning for speaker recognition. arXiv preprint arXiv:2003.11982 (2020) +2. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017) +3. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017) +4. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 251-263. Springer (2017) + +15. Chung, S.W., Chung, J.S., Kang, H.G.: Perfect match: Improved cross-modal embeddings for audio-visual synchronisation. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3965-3969. IEEE (2019) +16. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 408-424. Springer (2020) +17. Eskimez, S.E., Maddox, R.K., Xu, C., Duan, Z.: End-to-end generation of talking faces from noisy speech. In: ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 1948-1952. IEEE (2020) +18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) +19. Gu, Y., Wang, X., Xie, L., Dong, C., Li, G., Shan, Y., Cheng, M.M.: Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII. pp. 126–143. Springer (2022) +20. Guan, J., Zhang, Z., Zhou, H., Hu, T., Wang, K., He, D., Feng, H., Liu, J., Ding, E., Liu, Z., et al.: Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1505-1515 (2023) +21. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5784-5794 (2021) +22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +23. Hershey, J., Movellan, J.: Audio vision: Using audio-visual synchrony to locate sounds. Advances in neural information processing systems 12 (1999) +24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +25. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +26. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. arXiv preprint arXiv:2210.07055 (2022) +27. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015) +28. Jamaludin, A., Chung, J.S., Zisserman, A.: You said that?: Synthesising talking faces from audio. International Journal of Computer Vision 127, 1767-1779 (2019) +29. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022) +30. Ji, X., Zhou, H., Wang, K., Wu, W., Loy, C.C., Cao, X., Xu, F.: Audio-driven emotional video portraits. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14080-14089 (2021) + +31. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) +32. Kadandale, V.S., Montesinos, J.F., Haro, G.: Vocalist: An audio-visual synchronisation model for lips and voices. arXiv preprint arXiv:2204.02090 (2022) +33. Kim, Y.J., Heo, H.S., Chung, S.W., Lee, B.J.: End-to-end lip synchronisation based on pattern classification. In: 2021 IEEE Spoken Language Technology Workshop (SLT). pp. 598-605. IEEE (2021) +34. KR, P., Mukhopadhyay, R., Philip, J., Jha, A., Namboodiri, V., Jawahar, C.: Towards automatic face-to-face translation. In: Proceedings of the 27th ACM international conference on multimedia. pp. 1428-1436 (2019) +35. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012) +36. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3387-3396 (2022) +37. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 106-125. Springer (2022) +38. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018) +39. Muaz, U., Jang, W., Tripathi, R., Mani, S., Ouyang, W., Gadde, R.T., Gecer, B., Elizondo, S., Madad, R., Nair, N.: Sidgan: High-resolution dubbed video generation via shift-invariant learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7833-7842 (2023) +40. Nair, V., Hinton, G.E.: Rectified linear units improve restricted boltzmann machines. In: Proceedings of the 27th international conference on machine learning (ICML-10). pp. 807-814 (2010) +41. Nayak, S., Schuler, C., Saha, D., Baumann, T.: A deep dive into neural synchrony evaluation for audio-visual translation. In: Proceedings of the 2022 International Conference on Multimodal Interaction. pp. 642-647 (2022) +42. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +43. Owens, A., Efros, A.A.: Audio-visual scene analysis with self-supervised multisensory features. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 631-648 (2018) +44. Papantoniou, F.P., Filntisis, P.P., Maragos, P., Roussos, A.: Neural emotion director: Speech-preserving semantic control of facial expressions in" in-the-wild" videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18781-18790 (2022) +45. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2062-2070 (2022) +46. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020) + +47. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +48. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015) +49. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII. pp. 666-682. Springer (2022) +50. Shen, S., Zhao, W., Meng, Z., Li, W., Zhu, Z., Zhou, J., Lu, J.: Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1982-1991 (2023) +51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +52. Slaney, M., Covell, M.: Facesync: A linear operator for measuring synchronization of video facial images and audio tracks. Advances in neural information processing systems 13 (2000) +53. Song, L., Wu, W., Qian, C., He, R., Loy, C.C.: Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security 17, 585-598 (2022) +54. Song, Y., Zhu, J., Li, D., Wang, X., Qi, H.: Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786 (2018) +55. Stypulkowski, M., Vougioukas, K., He, S., Zieba, M., Petridis, S., Pantic, M.: Diffused heads: Diffusion models beat gans on talking-face generation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5091-5100 (2024) +56. Sun, Y., Zhou, H., Wang, K., Wu, Q., Hong, Z., Liu, J., Ding, E., Wang, J., Liu, Z., Hideki, K.: Masked lip-sync prediction by audio-visual contextual exploitation in transformers. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1–9 (2022) +57. Suwajanakorn, S., Seitz, S.M., Kemelmacher-Shlizerman, I.: Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017) +58. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022) +59. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 716-731. Springer (2020) +60. Vougioukas, K., Petridis, S., Pantic, M.: Realistic speech-driven facial animation with gans. International Journal of Computer Vision 128, 1398-1413 (2020) +61. Waibel, A., Behr, M., Yaman, D., Eyiokur, F.I., Nguyen, T.N., Mullov, C., Demirtas, M.A., Kantarci, A., Constantin, S., Ekenel, H.K.: Face-dubbing++: Lipsynchronous, voice preserving translation of videos. In: 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). pp. 1-5. IEEE (2023) + +62. Wang, D., Deng, Y., Yin, Z., Shum, H.Y., Wang, B.: Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17979-17989 (2023) +63. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14653-14662 (2023) +64. Wang, J., Zhao, K., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Lipformer: High-fidelity and generalizable talking face generation with a pre-learned facial codebook. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13844-13853 (2023) +65. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +66. Wu, H., Jia, J., Wang, H., Dou, Y., Duan, C., Deng, Q.: Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1478-1486 (2021) +67. Yao, S., Zhong, R., Yan, Y., Zhai, G., Yang, X.: Dfa-nerf: personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791 (2022) +68. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023) +69. Yehia, H., Rubin, P., Vatikiotis-Bateson, E.: Quantitative association of vocal-tract and facial behavior. Speech Communication 26(1-2), 23-43 (1998) +70. Yin, F., Zhang, Y., Cun, X., Cao, M., Fan, Y., Wang, X., Bai, Q., Wu, B., Wang, J., Yang, Y.: Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 85-101. Springer (2022) +71. Zhan, F., Yu, Y., Wu, R., Zhang, J., Lu, S.: Multimodal image synthesis and editing: A survey. arXiv preprint arXiv:2112.13592 (2021) +72. Zhang, C., Zhao, Y., Huang, Y., Zeng, M., Ni, S., Budagavi, M., Guo, X.: Facial: Synthesizing dynamic talking face with implicit attribute learning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3867-3876 (2021) +73. Zhang, Z., Hu, Z., Deng, W., Fan, C., Lv, T., Ding, Y.: Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. arXiv preprint arXiv:2303.03988 (2023) +74. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021) +75. Zhen, R., Song, W., He, Q., Cao, J., Shi, L., Luo, J.: Human-computer interaction system: A survey of talking-head generation. Electronics 12(1), 218 (2023) +76. Zhong, W., Fang, C., Cai, Y., Wei, P., Zhao, G., Lin, L., Li, G.: Identity-preserving talking face generation with landmark and appearance priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2023) +77. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 9299–9306 (2019) + +78. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4176-4186 (2021) +79. Zhou, M., Bai, Y., Zhang, W., Yao, T., Zhao, T., Mei, T.: Responsive listening head generation: A benchmark dataset and baseline (2022) +80. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG) 39(6), 1-15 (2020) \ No newline at end of file diff --git a/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/images.zip b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e86255dd54aa7643b902fc56d674af54bbc8503 --- /dev/null +++ b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e28eb6e2e78c4f1acedcd3dc8480236ee8902fb58c4bb0f8073d6e6ea99f9ee +size 456684 diff --git a/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/layout.json b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a9b635e30020b0ec9aa2321070c1407a056ed6ac --- /dev/null +++ b/2024/Audio-driven Talking Face Generation with Stabilized Synchronization Loss/layout.json @@ -0,0 +1,9892 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 459, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 459, + 147 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 459, + 147 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 167, + 455, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 167, + 455, + 193 + ], + "spans": [ + { + "bbox": [ + 159, + 167, + 455, + 193 + ], + "type": "text", + "content": "Dogucan Yaman1, Fevziye Irem Eyiokur1, Leonard Bärmann1, Hazim Kemal Ekenel2, and Alexander Waibel1,3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 189, + 201, + 424, + 246 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "spans": [ + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "type": "text", + "content": " Karlsruhe Institute of Technology, Karlsruhe, Germany" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 213, + 406, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 213, + 406, + 224 + ], + "spans": [ + { + "bbox": [ + 206, + 213, + 406, + 224 + ], + "type": "text", + "content": "2 Istanbul Technical University, Istanbul, Turkey" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 224, + 408, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 224, + 408, + 246 + ], + "spans": [ + { + "bbox": [ + 205, + 224, + 408, + 246 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 205, + 224, + 408, + 246 + ], + "type": "text", + "content": " Carnegie Mellon University, Pittsburg PA, USA dogucan.yaman@kit.edu" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 160, + 270, + 453, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 270, + 453, + 423 + ], + "spans": [ + { + "bbox": [ + 160, + 270, + 453, + 423 + ], + "type": "text", + "content": "Abstract. Talking face generation aims to create realistic videos with accurate lip synchronization and high visual quality, using given audio and reference video while preserving identity and visual characteristics. In this paper, we start by identifying several issues with existing synchronization learning methods. These involve unstable training, lip synchronization, and visual quality issues caused by lip-sync loss, SyncNet, and lip leaking from the identity reference. To address these issues, we first tackle the lip leaking problem by introducing a silent-lip generator, which changes the lips of the identity reference to alleviate leakage. We then introduce stabilized synchronization loss and AVSyncNet to overcome problems caused by lip-sync loss and SyncNet. Experiments show that our model outperforms state-of-the-art methods in both visual quality and lip synchronization. Comprehensive ablation studies further validate our individual contributions and their cohesive effects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "spans": [ + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "text", + "content": "Keywords: Talking face generation " + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "text", + "content": " Lip synchronization " + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "text", + "content": " Lip leaking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 463, + 230, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 463, + 230, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 463, + 230, + 475 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 481, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 558 + ], + "type": "text", + "content": "Audio-driven talking face generation aims at generating a video with respect to given face and audio sequences. The objective is to achieve synchronized lip movements corresponding to the provided audio while preserving the identity and visual details. This task has recently attracted significant attention due to its versatile applications, including dubbing in the film industry, online education, enhancing video conferencing, and dubbing for various types of videos [71, 75]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 559, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 481, + 666 + ], + "type": "text", + "content": "The talking face generation task comprises two primary aspects: (1) lip synchronization and (2) visual quality of the face. Since lips that are out-of-sync with audio can be easily identified by humans, synchronized lips are key to achieving natural and realistic talking-face generation. For this, the primary solution is to evaluate audio-lip synchronization and use it as a training objective. Wav2Lip [46] introduced an improved version of SyncNet [14], a pretrained model designed to measure audio-visual synchronization. This model is also used during training to extract features and compute lip-sync loss. Subsequently, many approaches have employed improved SyncNet [46] to guide the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 118, + 247, + 203 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 247, + 203 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 247, + 203 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 247, + 203 + ], + "type": "image", + "image_path": "6647a73177df405c2a63197b23f09cfd59134fa1c42614fc06ded5f83ba822a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 205, + 227, + 215 + ], + "lines": [ + { + "bbox": [ + 167, + 205, + 227, + 215 + ], + "spans": [ + { + "bbox": [ + 167, + 205, + 227, + 215 + ], + "type": "text", + "content": "(a) SyncNet [46]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 118, + 367, + 204 + ], + "blocks": [ + { + "bbox": [ + 260, + 118, + 367, + 204 + ], + "lines": [ + { + "bbox": [ + 260, + 118, + 367, + 204 + ], + "spans": [ + { + "bbox": [ + 260, + 118, + 367, + 204 + ], + "type": "image", + "image_path": "2e2d05c640144d1b3a3edfb6edb5388406a3a9afbb650b2fb588f3d46d64aee2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 205, + 357, + 215 + ], + "lines": [ + { + "bbox": [ + 278, + 205, + 357, + 215 + ], + "spans": [ + { + "bbox": [ + 278, + 205, + 357, + 215 + ], + "type": "text", + "content": "(b) AVSyncNet (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 223, + 482, + 268 + ], + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 268 + ], + "type": "text", + "content": "Fig. 1: (a, b) Cosine similarity between GT audio-lip pairs on random LRS2 samples, showcasing the instability of SyncNet and more robust performance of AVSyncNet. (c) illustrates full mouth region / lip leaking from the reference, pose effect from the reference, and similar identity reference-target image scenarios." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 378, + 115, + 477, + 204 + ], + "blocks": [ + { + "bbox": [ + 378, + 115, + 477, + 204 + ], + "lines": [ + { + "bbox": [ + 378, + 115, + 477, + 204 + ], + "spans": [ + { + "bbox": [ + 378, + 115, + 477, + 204 + ], + "type": "image", + "image_path": "8847030a5549f6e1f2bba0af662de421fec6c445d6b2393eea2c472792327752.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 205, + 466, + 215 + ], + "lines": [ + { + "bbox": [ + 389, + 205, + 466, + 215 + ], + "spans": [ + { + "bbox": [ + 389, + 205, + 466, + 215 + ], + "type": "text", + "content": "(c) Leaking problems" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "type": "text", + "content": "model throughout training. The main concept is to provide audio-video pairs to SyncNet, which extracts features through its image and audio encoders. Subsequently, cosine similarity with binary cross-entropy loss is measured between these two features [46]. Since the model was trained in this manner, a higher cosine similarity is expected when the lips are synchronized with the given audio. Although existing approaches with this strategy mostly surpass other methods in lip synchronization, there are still challenges that must be addressed to enhance performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 391, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 666 + ], + "type": "text", + "content": "In this work, we identify two main challenges in existing approaches that restrict models from achieving satisfactory performance in both lip synchronization and visual quality: SyncNet instabilities and lip leaking. First, in alignment with previous research [41], we observe instabilities in the performance of SyncNet [46] on pairs of ground-truth (GT) lip and audio (see Fig. 1a and App. A). Thus, when SyncNet is utilized as part of talking face generation training, it might provide an inadequate training signal, assigning low similarity scores to generated images even when the lips are synchronized. This problem causes unstable training, degrading the lip generation capability of the network, thereby ending up in out-of-sync lips or suboptimal lip-sync. Furthermore, the lip-sync loss [46] and reconstruction losses are conflicting [39]. This leads models to have either poor lip synchronization or degraded visual quality (sometimes even both), escalating further when lip-sync loss and SyncNet are employed with high-resolution (HR) data [39,73]. To address these problems, we first improve SyncNet and introduce AVSyncNet, demonstrating a more robust performance (see Fig. 1b) and also overcoming poor shift-invariance characteristics of SyncNet (see Fig. 6a). However, despite improved performance, the instability problem is not fully solved (see Fig. 1b). To overcome this problem further, we also introduce a stabilized synchronization loss. Specifically, instead of directly using the similarity of the (generated lips, audio) pair, we calculate the difference of the similarities between (GT lips, audio) and (generated lips, audio). We hypothesize that this alleviates the described problems since it guides the model to generate a lip movement with a similar synchronization score as for the GT face. Together with AVSync-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Net, this method empirically enhances the lip synchronization performance as well as avoids visual quality issues caused by misleading lip-sync loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 416 + ], + "type": "text", + "content": "Second, we address another main problem in the talking face generation literature: lip leaking. The current gold standard in 2D-based methods is to input a bottom-half masked face image (\"pose reference\"), wherein the model is expected to generate the input face with proper lip movements. However, because of masking, the model requires a reference image to retain the identity and texture in the masked part. This is achieved by randomly choosing a face image from a different part of the video sequence, referred to as \"identity reference\". However, this introduces a new challenge: As it is randomly selected, the lip movements of the identity reference can frequently be similar to the GT lips during training (see last row of Fig. 1c). Hence, for faster convergence, the model tends to replicate the lip movements from the identity reference, resulting in poor lip synchronization or complete out-of-sync output. By following the literature [41,46], we term this phenomenon as lip leaking. To tackle this problem, we propose a simple yet effective technique: silent-lip generator. The concept involves modifying the identity reference to make the lips closed (thus \"silent-lip\") before feeding it to the talking face generator network. This method ensures consistently closed lips in the identity reference, effectively mitigating the lip leaking problem. Our contributions are as follows: (i) We identify and analyze various fundamental problems that harm lip synchronization learning and also cause visual quality issues. (ii) We present a robust and shift-invariant AVSyncNet, and stabilized synchronization loss to overcome the problems caused by lip-sync loss and SyncNet. (iii) We present a silent-lip generator to generate an identity reference with closed lips before feeding the talking face generator to alleviate the lip leaking." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 435, + 237, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 237, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 237, + 449 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation. Initial studies mapped audio features to time-aligned facial motions [69] or predicted facial motions by HMM [6]. In [57], videos were generated by finding the images most aligned with the audio. ATVGNet [9] transfers audio to facial landmarks and uses pixel-wise loss with an attention mechanism to avoid the jittering problem and leaking of irrelevant speech. [16] and [80] use facial landmark representation to synthesize faces with synchronized lips. Recent papers address the task as conditional inpainting by masking the bottom half of the input face and feeding the network an identity reference from another time step of the same video, along with the audio segment. Wav2Lip [46] proposes SyncNet and lip-sync loss to predict lip synchronization and achieves significant performance. PC-AVS [78] introduces a pose-controllable audio-visual system, while GC-AVT [36], EAMM [29] and EVP [30] control the emotion by utilizing an emotion embedding. On the other hand, SyncTalkFace [45] uses Audio-Lip Memory to store lip motion features and retrieves them as visual hints for better synchronization. VideoReTalking [10] proposes to manipulate the reference image to have a face with canonical expression to alleviate the sensitivity of the model against the identity reference image." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "type": "text", + "content": "Similarly, we introduce a silent-lip generator to implicitly learn to manipulate the lips of the identity reference to mitigate the lip leaking problem. Compared to the method proposed in [10], our method works more efficiently to preserve the identity and visual quality. More recently, LipFormer [64] uses a pre-learned facial codebook to generate HR videos, while DINet [73] proposes to use a deformation module to obtain deformed features to enhance lip synchronization and head pose alignment. IPLAP [76] shows satisfactory visual quality and stability by using intermediate landmark representation and motion field. Recently, TalkLip [63] introduces a global audio encoder, trained with self-supervised learning, to encode features by considering the entire content of the audio. Besides, they propose to use lip reading during the training as well as in the evaluation to control whether the content is preserved. SIDGAN [39] performs important analyzes, and introduces shift-invariant APS-SyncNet and training objectives along with the coarse-to-fine pyramid model for HR dubbed video generation. Recent works use diffusion model because of its stability and accuracy [50, 55]. Finally, in [61], talking face generation is used as a part of a full system to perform end-to-end face dubbing, involving speech recognition, translation, speech, and video generation. In contrast to the above, 3D-based and Neural Radiance Fields-based (NeRFs) methods typically generate the entire head, rather than just manipulating 2D images of the face, often involving manipulation of pose, emotion, and 3D face model [4, 5, 21, 37, 44, 49, 53, 58, 59, 62, 66–68, 70, 72, 74, 77, 80]. Similarly, portrait animation aims at utilizing a single input image to generate a video by predicting the pose and expression of the subsequent frames, along with ensuring synchronized lips. Nevertheless, these tasks significantly differ from face dubbing in terms of methodology, task definition, goals, and real-world applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 430, + 483, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 483, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 483, + 587 + ], + "type": "text", + "content": "Lip Synchronization. Some earlier works [23, 52] used hand-crafted features and statistical models to evaluate lip synchronization. Recent studies proposed to use mutual information between audio-visual features to produce sync or out-of-sync output for sound [8,26,43] or speech [2,12,15,32,33]. While some methods learn lip synchronization implicitly [9,21,28,34,57,66,74], other methods employ distance between landmarks or facial parameters [30,44,53,80]. Contrarily, the majority of works [17,20,36,45,46,54,56,60,62-64,67,73,77,78] extract audio-visual features with an additional network (mostly SyncNet [46]) for audio-lip synchronization prediction. Afterwards, while some of these methods utilize lip-sync loss [46], others benefit from contrastive learning by using infoNCE [42]. We follow a similar strategy by utilizing an additional network to calculate a synchronization loss. We propose a robust and shift-invariant AVSyncNet along with a stabilized synchronization loss to overcome existing challenges." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 601, + 272, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 272, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 272, + 616 + ], + "type": "text", + "content": "3 Proposed Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 624, + 284, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 284, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 284, + 636 + ], + "type": "text", + "content": "3.1 Talking Face Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We propose an audio-driven talking face generation model " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " with enhanced lip synchronization. As shown in Fig. 2a, our model incorporates: 1) an audio" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 406, + 245 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 406, + 245 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 406, + 245 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 406, + 245 + ], + "type": "image", + "image_path": "7bddec5e1b8bf354bf2c1facb6932189a4d173659ff17d8b7ae18c7a6091a6b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 250, + 329, + 258 + ], + "lines": [ + { + "bbox": [ + 209, + 250, + 329, + 258 + ], + "spans": [ + { + "bbox": [ + 209, + 250, + 329, + 258 + ], + "type": "text", + "content": "(a) Talking face generation model" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 422, + 249, + 471, + 258 + ], + "lines": [ + { + "bbox": [ + 422, + 249, + 471, + 258 + ], + "spans": [ + { + "bbox": [ + 422, + 249, + 471, + 258 + ], + "type": "text", + "content": "(b) FD block" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 412, + 140, + 481, + 205 + ], + "blocks": [ + { + "bbox": [ + 412, + 140, + 481, + 205 + ], + "lines": [ + { + "bbox": [ + 412, + 140, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 412, + 140, + 481, + 205 + ], + "type": "image", + "image_path": "dc4063384e5965c490218073f5107120f76fbb60a62779176997d73f71f710fe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 268, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 323 + ], + "type": "text", + "content": "Fig. 2: Talking face generation model " + }, + { + "bbox": [ + 130, + 268, + 482, + 323 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 268, + 482, + 323 + ], + "type": "text", + "content": " (a) and face-decoding (FD) block (b). Our model receives a pose reference sequence, mel-spectrogram of an audio snippet, and a silent identity reference, that is generated by our silent-lip generator " + }, + { + "bbox": [ + 130, + 268, + 482, + 323 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 268, + 482, + 323 + ], + "type": "text", + "content": ", aiming to alleviate lip leaking problem. The model then synthesizes the talking face sequence to ensure lip synchronization. Subsequently, the employed loss functions are computed." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 348, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 373 + ], + "type": "text", + "content": "encoder for processing the audio snippet, 2) an identity encoder for addressing an identity reference image, and 3) a pose encoder for utilizing a pose reference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "text", + "content": "Audio Encoder. The audio encoder " + }, + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "inline_equation", + "content": "E_{A}" + }, + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "text", + "content": " generates phoneme-level embeddings, serving as conditions for the face generator to generate lip movements accurately. Our audio encoder extracts embeddings " + }, + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "inline_equation", + "content": "F_{A} = E_{A}(A) \\in \\mathbb{R}^{1 \\times 1 \\times 512}" + }, + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "text", + "content": " from the given mel-spectrogram " + }, + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "text", + "content": ", representing the driving audio. In contrast to existing methods [10, 46, 73, 76], we propose using a pretrained, frozen audio encoder, concurrently trained with a face encoder to learn lip synchronization similar to the objective in SyncNet [46]. Thus, we can obtain improved embeddings during the talking face generation training by leveraging the capacity of the pretrained robust audio encoder. We obtain the best score with such frozen audio encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Face Encoder. In accordance with the gold standard in the literature, we utilize an identity reference, " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "I^R" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", and a pose reference, " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", as inputs to the model. The identity reference is a face image of the subject that provides identity information. It is different from the pose reference and randomly selected from the same video. The pose reference is identical to the target image, except for the bottom half, which is masked, as the model is designed to focus on generating lip movements. Unlike most conventional methods that employ a joint encoder for processing identity and pose references, we use individual encoders to allow each encoder to focus solely on their respective tasks [39]. Therefore, we utilize two parallel CNN-based face encoders to process identity and pose references individually. This approach yields better feature representation and ultimately leads to improved performance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 448, + 102 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 294, + 237 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 294, + 237 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 294, + 237 + ], + "type": "image", + "image_path": "a4914333acf4b1932424f458c6b84e04a0bbe41ccf0067ae0283a8ac6e78a5a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 239, + 274, + 251 + ], + "lines": [ + { + "bbox": [ + 149, + 239, + 274, + 251 + ], + "spans": [ + { + "bbox": [ + 149, + 239, + 274, + 251 + ], + "type": "text", + "content": "(a) Silent-lip generation model " + }, + { + "bbox": [ + 149, + 239, + 274, + 251 + ], + "type": "inline_equation", + "content": "G_{S}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 338, + 129, + 484, + 211 + ], + "blocks": [ + { + "bbox": [ + 338, + 129, + 484, + 211 + ], + "lines": [ + { + "bbox": [ + 338, + 129, + 484, + 211 + ], + "spans": [ + { + "bbox": [ + 338, + 129, + 484, + 211 + ], + "type": "image", + "image_path": "eeb47c49fc2ae167e741c0a489896d3cdd2383635a7473f2246d9b316b85104b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 239, + 440, + 251 + ], + "lines": [ + { + "bbox": [ + 382, + 239, + 440, + 251 + ], + "spans": [ + { + "bbox": [ + 382, + 239, + 440, + 251 + ], + "type": "text", + "content": "(b) AVSyncNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 258, + 440, + 271 + ], + "lines": [ + { + "bbox": [ + 173, + 258, + 440, + 271 + ], + "spans": [ + { + "bbox": [ + 173, + 258, + 440, + 271 + ], + "type": "text", + "content": "Fig. 3: " + }, + { + "bbox": [ + 173, + 258, + 440, + 271 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 173, + 258, + 440, + 271 + ], + "type": "text", + "content": " in inference (a) and AVSyncNet training pipeline (b)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 295, + 265, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 265, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 265, + 308 + ], + "type": "text", + "content": "3.2 Silent-Lip Generator" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 317, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 509 + ], + "type": "text", + "content": "Talking face generation involves using an identity reference to preserve the identity in the generated image. This is particularly important because the bottom half of the pose reference, specifically the mouth region, is masked. This masking is necessary as we aim to synthesize appropriate lip movements corresponding to the provided audio. However, models are unintentionally affected by the lip movement of the identity reference, rather than solely gathering identity information (see Fig. 1c and App. C for details). This behavior, which we refer to as lip leaking, leads to poor lip synchronization or occasionally even non-converged training. We hypothesize that there are two main reasons for this: First, the lip movement of the identity reference may occasionally resemble that of the lips in the target image (see last row of Fig. 1c). Hence, the model can lower the synchronization loss more quickly by undesirably replicating the lip movements of the identity reference. Second, the diversity of lip movements in the identity reference may yield the model to seek a correlation with the target lips. This causes a challenging disentanglement task for the identity encoder —namely distinguishing identity information and lip movement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": "To mitigate the aforementioned issue, we propose to use an additional model, called silent-lip generator " + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": ", prior to the talking face generation model " + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": ", aimed at modifying the lip shape of the identity reference. Specifically, we reconstruct the input face with closed, flat lips. This strategy reduces the likelihood of having lips similar to the target face in the identity reference and resolves the issue of diverse lip movements. Consequently, the model no longer replicates the lips from the identity reference, resulting in stable training & improved lip sync." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To implement " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", we structure the task and the model similar to the talking face generation. Specifically, we input a bottom-half masked pose reference, an identity reference, and an audio snippet to train a GAN [18] for generating a talking face. The model is trained to reconstruct the face using both the pose and identity references. Notably, we exclude any synchronization loss and SyncNet during training and train the model under the weak condition. Consequently, the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": "model focuses solely on generating lip movements without synchronization when speech is present. Therefore, it implicitly learns to generate closed lips when the input audio is silent as shown in Fig. 3a. Furthermore, by eliminating synchronization loss, we overcome unstable training and the issue of lip leaking from the identity reference for this network. We choose this approach for its efficient utilization of the same model and data. Given the scarcity of closed-lip faces in the dataset, we avoid using these frames directly to maintain effective training and generalization. Please note that we use the same architecture for " + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": ". We initially train " + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": " separately on the same training data (LRS2) and then incorporate it into the training of " + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": " without further finetuning. Specifically, we then only pass silent audio to " + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": ", so that it modifies given identity references to have silent lips for subsequent use as identity reference for " + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 275, + 251, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 251, + 286 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 251, + 286 + ], + "type": "text", + "content": "3.3 Video Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "text", + "content": "For the talking face video generation, we employ the aforementioned components. Initially, we employ our pretrained silent-lip generator " + }, + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "text", + "content": " to synthesize the identity reference with closed lips. Subsequently, we input this silent identity reference to the identity encoder and the pose reference to the pose encoder in the talking face generation model " + }, + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "text", + "content": ". Similarly, we provide the mel-spectrogram of the corresponding audio snippet to the audio encoder. Next, we concatenate the embeddings from the identity and pose encoders, along with the depth dimension, and pass it through a " + }, + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "text", + "content": " convolution layer to reduce the depth. Finally, we concatenate this feature representation with the audio embeddings before feeding them into the face generator. The face generator generates the entire face with accurate lip movements by preserving the identity and the pose." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "type": "text", + "content": "We illustrate the talking face generation model " + }, + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "type": "text", + "content": " in Fig. 2a. We integrate the U-Net architecture [47] for our overall design, leveraging its adequate performance in reconstruction tasks while ensuring computational efficiency. Our identity encoder and pose encoder share the same architecture, consisting of consecutive face-encoding (FE) blocks. Each block has a strided-convolutional layer followed by two non-strided convolutional layers, each paired with a batch normalization layer [27] and a ReLU activation function [35,40]. We also use the residual connection strategy [22] by summing the input and output of each block before forwarding it to the next layer. On the other hand, our face generator has consecutive face-decoding (FD) blocks. As shown in Fig. 2b, within each block, we utilize a transposed-convolutional layer, followed by two convolutional layers incorporating batch normalization and ReLU activation function. Moreover, we apply a skip connection between the reciprocal layers of the face encoders and the decoder to retain high-level features and enhance the training stability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "GAN Loss. To train our model, we utilize GAN loss [18] and employ a discriminator, which is a straightforward CNN-based binary classification network to distinguish real and fake samples, designed with a balanced architecture aligned with our face encoders. We benefit from consecutive strided-convolutional layers followed by the Leaky ReLu activation function and spectral normalization [38]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 448, + 102 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 165 + ], + "type": "text", + "content": "Reconstruction Loss. We employ L1 loss in pixel space " + }, + { + "bbox": [ + 130, + 115, + 482, + 165 + ], + "type": "inline_equation", + "content": "L_{pixel} = ||I' - I^{GT}||_1" + }, + { + "bbox": [ + 130, + 115, + 482, + 165 + ], + "type": "text", + "content": " between the generated " + }, + { + "bbox": [ + 130, + 115, + 482, + 165 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 130, + 115, + 482, + 165 + ], + "type": "text", + "content": " and the target faces " + }, + { + "bbox": [ + 130, + 115, + 482, + 165 + ], + "type": "inline_equation", + "content": "I^{GT}" + }, + { + "bbox": [ + 130, + 115, + 482, + 165 + ], + "type": "text", + "content": " to ensure consistency in areas outside the lips and maintain the illumination condition. We further utilize perceptual loss [31] based on the pretrained VGG-19 [51]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 174, + 481, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 174, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 211, + 174, + 481, + 206 + ], + "type": "interline_equation", + "content": "L _ {p e r} = \\sum_ {i = 1} ^ {5} c _ {i} | | \\mathrm {V G G} ^ {\\phi_ {i}} \\left(I ^ {\\prime}\\right) - \\mathrm {V G G} ^ {\\phi_ {i}} \\left(I ^ {G T}\\right) | | _ {1} \\tag {1}", + "image_path": "aa6e4308c6c1acd351dc499e59218d3b457c3ecb6fe5feb01785aff034c2f2f5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 214, + 480, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 214, + 480, + 228 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 132, + 214, + 480, + 228 + ], + "type": "text", + "content": " are weight coefficients from [31], and " + }, + { + "bbox": [ + 132, + 214, + 480, + 228 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 132, + 214, + 480, + 228 + ], + "type": "text", + "content": " indicates the set of VGG layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 245, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 402 + ], + "type": "text", + "content": "Adaptive Triplet Loss. Although our goal is to capture visual details from the identity reference, we observe that the model occasionally focuses on the visual details (e.g., illumination, pose) in the identity reference excessively and this tendency could potentially degrade the quality and stability of the generated face sequence, resulting in suboptimal performance. To tackle this, we exploit a triplet loss strategy [48], aiming to minimize the distance between the generated face and GT, while maximizing the distance between the generated face and the identity reference. However, the random selection of identity references increases the probability of choosing an image that closely resembles the GT. This scenario poses a challenge for the vanilla triplet loss, potentially degrading training and resulting in poor visual and pose quality. To mitigate this, we introduce an adaptive triplet loss that considers the similarity between the identity reference and GT during loss computation to alter its effect. The formula is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 149, + 411, + 481, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 411, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 149, + 411, + 481, + 439 + ], + "type": "interline_equation", + "content": "L _ {a t} = \\left[ D (\\mathrm {V G G} (I ^ {\\prime}), \\mathrm {V G G} (I ^ {G T})) - \\frac {D (\\mathrm {V G G} (I ^ {\\prime}) , \\mathrm {V G G} (I ^ {R}))}{D (\\mathrm {V G G} (I ^ {G T}) , \\mathrm {V G G} (I ^ {R}))} + \\alpha \\right] _ {+} \\tag {2}", + "image_path": "a1e9184afccb1bfec36b6a77ea29beb6ef0bac7f17e1588f8711060dc3bbce07.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 446, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 446, + 482, + 544 + ], + "type": "inline_equation", + "content": "[\\cdot]_{+} = \\max(\\cdot, \\epsilon)" + }, + { + "bbox": [ + 130, + 446, + 482, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 446, + 482, + 544 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 446, + 482, + 544 + ], + "type": "text", + "content": " represents the L2 distance, and we empirically choose " + }, + { + "bbox": [ + 130, + 446, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 446, + 482, + 544 + ], + "type": "text", + "content": ". In this loss, we leverage the ratio of the similarity between the generated image and identity reference to that of the GT and the identity reference to adjust the loss value. As the identity reference becomes more similar to GT, the impact of the distance between the generated image and identity reference on the loss diminishes, since expecting a high distance in this case is not reasonable. Since our objective is to incorporate visual details from the identity reference, we opt for a very low coefficient to avoid conflicting with the primary goal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 561, + 290, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 290, + 574 + ], + "type": "text", + "content": "3.4 Learning Synchronization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": "The lip-sync loss [46] serves as a method to calculate synchronization between audio and video. Leveraging the pretrained SyncNet [46] for feature extraction from both audio and video inputs demonstrates reasonable performance in learning lip synchronization. However, the SyncNet is significantly unstable when measuring this similarity. Our evaluation using SyncNet on GT training data reveals notable fluctuations in cosine similarity between video and audio, contrary to the expected high scores (see Fig. 1a and App. A for details). Therefore, this" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "provides conflicting information to the system, resulting in poor lip synchronization, unstable training, and degraded visual quality. To tackle this problem, we present a more accurate, shift-invariant, and robust version of SyncNet, named AVSyncNet, and introduce a novel, stabilized synchronization loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 179, + 482, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 482, + 313 + ], + "type": "text", + "content": "AVSyncNet. We employ a ResNet-50-based [22] image encoder, known for its superior performance in face recognition, alongside a ResNetSE-34-based audio encoder [11], which is a modified version of ResNet-34 [22] designed to handle spectrogram inputs. This design makes the AVSyncNet model robust and shift-invariant compared to SyncNet [46]. We train our model on the LRS2 training data [1], calculating the cosine similarity between audio and lip features, followed by a binary cross-entropy loss, as shown in Fig. 3b. During each training step, we provide a set of images (5 images) along with the corresponding audio. For negative samples, we randomly select an audio snippet from the non-overlapping part of the video. Please note that as we feed the bottom half of the face to the image encoder, we adapt the first layer of ResNet-50 for an input size of " + }, + { + "bbox": [ + 130, + 179, + 482, + 313 + ], + "type": "inline_equation", + "content": "112 \\times 224" + }, + { + "bbox": [ + 130, + 179, + 482, + 313 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 413 + ], + "type": "text", + "content": "Stabilized Synchronization Loss " + }, + { + "bbox": [ + 130, + 327, + 482, + 413 + ], + "type": "inline_equation", + "content": "(L_{ss})" + }, + { + "bbox": [ + 130, + 327, + 482, + 413 + ], + "type": "text", + "content": ". Although AVSyncNet shows improved performance compared to SyncNet [46] and alleviates existing instability problem that harms lip-sync and visual quality, the unstable performance is not fully solved due to the inherent challenges of the task (see Fig. 1b and App. A, B). Therefore, we introduce a stabilized synchronization loss " + }, + { + "bbox": [ + 130, + 327, + 482, + 413 + ], + "type": "inline_equation", + "content": "(L_{ss})" + }, + { + "bbox": [ + 130, + 327, + 482, + 413 + ], + "type": "text", + "content": " to improve the lip synchronization performance further in conjunction with AVSyncNet by providing more stable and precise supervision. The formula is shown below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 220, + 418, + 481, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 418, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 220, + 418, + 481, + 445 + ], + "type": "interline_equation", + "content": "L _ {s s} = - \\log \\left(1 - \\frac {\\left| x - y \\right| + \\epsilon}{\\left| x - y \\right| + \\left| y - d \\right| + \\epsilon}\\right) \\tag {3}", + "image_path": "11938d83e9efdeab25bc65957a8a9461dc0c6bc0d797ee2fac0cf779a430f739.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 463, + 481, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 463, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 159, + 463, + 481, + 479 + ], + "type": "interline_equation", + "content": "x = \\operatorname {A V S I M} \\left(I ^ {\\prime}, A\\right), \\quad y = \\operatorname {A V S I M} \\left(I ^ {G T}, A\\right), \\quad d = \\operatorname {A V S I M} \\left(I ^ {R}, A\\right) \\tag {4}", + "image_path": "b59540236171472749f9d6b6f05c6e709141beb341b51fe81739590ee6a00ffb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "inline_equation", + "content": "I^{\\prime}, I^{GT}" + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "inline_equation", + "content": "I^R" + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "content": " are generated, GT, and identity reference lips, respectively, while " + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "content": " is their corresponding audio. AVSIM(" + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "inline_equation", + "content": "I, A" + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "content": ") indicates the audio-visual similarity between a face image (bottom half only, i.e. lips) and audio, given by the cosine similarity of extracted image and audio features " + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "inline_equation", + "content": "\\phi_{AVS}^V(I)" + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "inline_equation", + "content": "\\phi_{AVS}^A(A)" + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "content": " of the respective AVSyncNet encoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "text", + "content": "In this formulation, " + }, + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "text", + "content": " followed by cross-entropy loss denotes the lip-sync loss [46]. However, to address the unstable and fluctuating performance, we utilize the relative distance in similarity between GT lips-audio and generated-audio pairs. Sometimes, randomly selected reference images may have a lip" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 138, + 90, + 448, + 103 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 597, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 597, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 597, + 482, + 666 + ], + "type": "text", + "content": "This is reminiscent of distillation loss [25], as the actual value of the scores is neglected, and only their difference provides loss value for the training. However, initial experiments trying to directly minimize the distance between SyncNet (or AVSyncNet) image encoder features of generated and GT faces showed poor performance. We hypothesize that this is caused by SyncNet image features being only meaningful for comparing with corresponding audio features due to its training strategy." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "movement similar to that of the target image. We have already introduced the silent-lip generator " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " to mitigate this situation. However, AVSyncNet, while less severe than SyncNet, might still be unstable (e.g., unexpectedly high scores with incorrect pairs or vice versa), resulting in a minor lip leaking and stability issue. This problem arises when AVSyncNet assigns an erroneously high score to silent lip-audio pairs as well as when the target lip shape looks similar to closed lips. To mitigate this, we inject the similarity score between the identity reference and audio into the formulation. Specifically, we penalize the model more when the identity reference-audio pair shows higher similarity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 243, + 280, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 280, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 280, + 255 + ], + "type": "text", + "content": "3.5 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 263, + 395, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 263, + 395, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 263, + 395, + 276 + ], + "type": "text", + "content": "Combining all the presented contributions, the total loss is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 150, + 286, + 481, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 286, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 150, + 286, + 481, + 300 + ], + "type": "interline_equation", + "content": "L = L _ {G A N} (G, D) + \\lambda_ {1} L _ {p i x e l} (G) + \\lambda_ {2} L _ {p e r} (G) + \\lambda_ {3} L _ {s s} (G) + \\lambda_ {4} L _ {a t} (G) \\tag {5}", + "image_path": "b3bf7db13e577cfee238aa09a51216dbf7c414a279bae3b57e95396486187ccd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "text", + "content": " indicate generator and discriminator outputs, respectively. We empirically found the best coefficients as " + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "inline_equation", + "content": "(\\lambda_1, \\dots, \\lambda_4) = (10, 1, 2, 0.5)" + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "text", + "content": ". Please note that this is the loss function for training " + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "inline_equation", + "content": "G_L" + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "text", + "content": ", while for " + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "inline_equation", + "content": "G_S" + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "text", + "content": " we set " + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\lambda_3 = \\lambda_4 = 0" + }, + { + "bbox": [ + 130, + 308, + 482, + 344 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "text", + "content": "We process videos by using 5 consecutive frames in each step to consider the temporal information. We detect faces with FAN [7], followed by acquiring tight crops and resizing to " + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "text", + "content": ", as faces in LRS2 [1] are of low resolution. Our audio encoder receives a mel-spectrogram of size " + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "inline_equation", + "content": "16 \\times 80" + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "text", + "content": " derived from " + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "inline_equation", + "content": "16\\mathrm{kHz}" + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "text", + "content": " audio with a window size of 800 and a hop size of 200. We employ the Adam optimizer with " + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "inline_equation", + "content": "(\\beta_{1},\\beta_{2}) = (0.5,0.999)" + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "text", + "content": " and set the learning rate to " + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "text", + "content": " for all models. Training our AVSyncNet is done similarly to SyncNet [46] on the LRS2 dataset, and then we freeze the audio encoder of AVSyncNet and use it in the training of " + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 345, + 483, + 489 + ], + "type": "text", + "content": ". At the end of talking face generation, we apply a post-processing step by using VQFR [19] to enhance the visual quality and the resolution, aiming to achieve HR videos. We train and test our models with a single NVIDIA RTX A6000 GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 507, + 282, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 507, + 282, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 507, + 282, + 521 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "Dataset. We trained our silent-lip generator and talking face generator using Lip Reading Sentence 2 (LRS2) [1] training set as it is a well-known benchmark with extensive subject diversity. The evaluation was carried out on the LRS2 test set and extended to the LRW [13] test set and HTDF dataset [74] to demonstrate performance on unseen data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Metrics and Baseline. For visual quality, we employ widely used metrics: FID [24], SSIM [65], and PSNR. We also use inter-frame consistency (IFC), presented as a training objective in [79]. This is achieved by calculating the difference between the distances of the consecutive frames in the generated and GT videos. To evaluate lip synchronization, we follow the literature and use Landmark Distance (LMD) [9] in the mouth region and LSE-C & LSE-D metrics [46]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 136, + 481, + 224 + ], + "blocks": [ + { + "bbox": [ + 171, + 114, + 441, + 125 + ], + "lines": [ + { + "bbox": [ + 171, + 114, + 441, + 125 + ], + "spans": [ + { + "bbox": [ + 171, + 114, + 441, + 125 + ], + "type": "text", + "content": "Table 1: Quantitative results on the test sets of LRS2 and LRW." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 136, + 481, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 136, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 481, + 224 + ], + "type": "table", + "html": "
LRS2LRW
MethodSSIM ↑PSNR ↑FID ↓IFC ↓LMD ↓LSE-C ↑LSE-D ↓SSIM ↑PSNR ↑FID ↓IFC ↓LMD ↓LSE-C ↑LSE-D ↓
Wav2Lip [46]0.8626.537.050.212.387.596.750.8525.146.810.202.147.496.51
PC-AVS [78]0.7328.2418.400.461.936.417.520.8132.2514.270.381.426.537.15
EAMM [29]0.6921.0184.650.513.543.319.930.7126.2244.160.482.614.329.04
VideoReTalking w/ FR [10]0.8425.589.280.222.617.496.820.8727.115.300.232.396.597.12
DINet [73]0.7824.354.260.252.305.378.370.8827.508.170.221.965.249.09
TalkLip [63]0.8626.114.940.242.348.536.080.8626.3415.730.261.837.286.48
IPLAP [76]0.8729.674.100.202.116.497.160.9130.458.400.211.645.947.76
Ours w/o FR0.9532.643.830.161.138.416.030.9231.454.460.181.227.866.24
Ours w/ FR (VQFR)0.9031.805.230.271.368.525.830.9030.217.050.211.417.926.00
", + "image_path": "0d2282647bd1263b228bca740f4fa7f886bce9086ebc94ef7cd4195311061fbf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 245, + 480, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 293 + ], + "type": "text", + "content": "to measure the confidence and distance scores through a pretrained model [14]. We choose SOTA methods with publicly available codes and models to compare them fairly under the same conditions, as the implementation of the metrics and face cropping strategy before computing the metrics affect the scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 310, + 264, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 264, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 264, + 322 + ], + "type": "text", + "content": "4.1 Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 330, + 480, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 400 + ], + "type": "text", + "content": "In Tab. 1, we present quantitative results on test sets of two benchmark datasets, namely LRS2 and LRW. We achieve state-of-the-art results in all visual quality metrics excluding PSNR on LRW, which is a less informative metric compared to SSIM and FID. On IFC, we similarly outperform all compared methods, indicating that our model generates the most consistent videos in terms of temporal information and stability of the faces." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 481, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 534 + ], + "type": "text", + "content": "In lip synchronization evaluation, we obtain the best performance in LMD. Nevertheless, it is important to highlight that LMD is sensitive to the changes in the image, as it does not disentangle synchronization and visual stability. For instance, affine transformations impact the LMD score even when the lips are synchronized, and vice versa. On the LRW dataset, we achieve SOTA results with more reliable confidence and distance metrics for lip synchronization: LSE-C & D. On the LRS2 dataset, TalkLip yields a slightly better score than our model with the LSE-C metric. Nevertheless, we outperform TalkLip and achieve a SOTA result with the LSE-D metric. All these results indicate the accuracy of our method in terms of visual quality and lip synchronization. Similarly, we achieve SOTA results for most metrics on unseen HDTF [74], see App. E.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 550, + 257, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 257, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 257, + 562 + ], + "type": "text", + "content": "4.2 Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "In Fig. 4, we demonstrate a qualitative comparison with SOTA models and GT data. We use their respective publicly available models and generate videos from the HDTF dataset [74] to compare the models on unseen data since the presented models were trained on the LRS2 dataset, except for DINet (trained with HDTF). Due to SyncNet and lip-sync loss, TalkLip and Wav2Lip encounter generalization issues, sometimes leading to visual artifacts in the mouth region or face boundaries, especially when the pose of the identity reference differs from the pose reference, despite generating accurate lip movements. This observation" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 265, + 298 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 265, + 298 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 265, + 298 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 265, + 298 + ], + "type": "image", + "image_path": "f13553b11fc2a0811d7ce4e16d8f88304a730a7dc93c5a653549df43b3c96dbf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 145, + 335, + 163 + ], + "lines": [ + { + "bbox": [ + 276, + 145, + 335, + 163 + ], + "spans": [ + { + "bbox": [ + 276, + 145, + 335, + 163 + ], + "type": "text", + "content": "VideoReTalking (Cheng et al. 2022)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 277, + 172, + 334, + 190 + ], + "lines": [ + { + "bbox": [ + 277, + 172, + 334, + 190 + ], + "spans": [ + { + "bbox": [ + 277, + 172, + 334, + 190 + ], + "type": "text", + "content": "DINet (Zhang et al. 2023)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 277, + 198, + 334, + 215 + ], + "lines": [ + { + "bbox": [ + 277, + 198, + 334, + 215 + ], + "spans": [ + { + "bbox": [ + 277, + 198, + 334, + 215 + ], + "type": "text", + "content": "TalkLip (Wang et al. 2023)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 277, + 224, + 335, + 242 + ], + "lines": [ + { + "bbox": [ + 277, + 224, + 335, + 242 + ], + "spans": [ + { + "bbox": [ + 277, + 224, + 335, + 242 + ], + "type": "text", + "content": "IP-LAP (Zhong et al. 2023)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 254, + 325, + 262 + ], + "lines": [ + { + "bbox": [ + 286, + 254, + 325, + 262 + ], + "spans": [ + { + "bbox": [ + 286, + 254, + 325, + 262 + ], + "type": "text", + "content": "Ours w/o FR" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 285, + 281, + 326, + 289 + ], + "lines": [ + { + "bbox": [ + 285, + 281, + 326, + 289 + ], + "spans": [ + { + "bbox": [ + 285, + 281, + 326, + 289 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 307, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 350 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison with the SOTA methods. Reference videos (from HDTF [74]) are randomly selected and not seen during training by our model. For more images and videos, please check App. E, F and https://yamand16.github.io/TalkingFaceGeneration/." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 348, + 114, + 481, + 298 + ], + "blocks": [ + { + "bbox": [ + 276, + 120, + 336, + 137 + ], + "lines": [ + { + "bbox": [ + 276, + 120, + 336, + 137 + ], + "spans": [ + { + "bbox": [ + 276, + 120, + 336, + 137 + ], + "type": "text", + "content": "Wav2Lip (Prajwal et al. 2020)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 348, + 114, + 481, + 298 + ], + "lines": [ + { + "bbox": [ + 348, + 114, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 348, + 114, + 481, + 298 + ], + "type": "image", + "image_path": "7ad44190c5c745ee88774d909f76ede959f038ddd976d407a9ad2f6304327dad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 167, + 384, + 447, + 477 + ], + "blocks": [ + { + "bbox": [ + 158, + 361, + 454, + 373 + ], + "lines": [ + { + "bbox": [ + 158, + 361, + 454, + 373 + ], + "spans": [ + { + "bbox": [ + 158, + 361, + 454, + 373 + ], + "type": "text", + "content": "Table 2: Ablation studies on the LRS2 test set. See the text for details." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 384, + 447, + 477 + ], + "lines": [ + { + "bbox": [ + 167, + 384, + 447, + 477 + ], + "spans": [ + { + "bbox": [ + 167, + 384, + 447, + 477 + ], + "type": "table", + "html": "
AblationSetupMethodPSNR ↑ SSIM ↑ FID ↓ LMD ↓ LSE-C ↑ LSE-D ↓ IFC ↓
ComponentsAGL+LSa26.3490.85312.252.4087.116 ± 1.927.396 ± 1.03
BA + Ea,S26.6140.8689.822.3257.271 ± 1.767.106 ± 0.98
CA + Ea,W26.5900.86910.562.2787.220 ± 1.757.158 ± 0.99
DB + GS27.1800.8728.161.7417.752 ± 1.716.413 ± 0.95
EGL + Ea,S + GS + LSs31.1660.9255.271.1408.370 ± 1.166.032 ± 0.59
FE + Lt30.6580.9176.241.2508.260 ± 1.346.176 ± 0.64
GE + Lat32.7550.9494.021.1358.382 ± 1.166.057 ± 0.61
HG w / AVSyncNet32.6400.9523.831.1308.410 ± 0.976.037 ± 0.55
Post-processingFR1Setup H + GPEN28.9910.91958.771.1977.6256.457
FR2Setup H + GFPGAN31.1690.91613.071.2197.6246.496
FR3Setup H + VQFR: full model31.8060.9055.231.3658.5285.838
Silent face generationVRT-SVideoReTalking silent data22.1240.64633.60--0.463
Ours-SOur silent data (GS)33.3280.9514.41--0.141
", + "image_path": "25c7b450e8d51346fc90fb54de77d8f08fc3ebab9c236bb83278df2105a35526.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 498, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 571 + ], + "type": "text", + "content": "clearly validates the motivation of our contributions. In contrast, our model generates consistent faces with comparably fewer artifacts, featuring appropriate lip movements that align with both the GT faces and the corresponding audio. However, VideoReTalking demonstrates comparable lip synchronization and visual quality performance to our model. On the other hand, IP-LAP shows sufficient visual quality, while less accurate lip synchronization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 586, + 239, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 239, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 239, + 600 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Tab. 2 and Fig. 5a show a comprehensive ablation study on the LRS2 test set, analyzing the individual impact of our contributions. We first train our " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " model using SyncNet [46] and lip-sync loss [46] as a baseline. As expected, we encountered several issues with unstable training. Once our model converged after several random seeds, the results (Setup A) show that lip-sync loss can achieve" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 274, + 152 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 274, + 152 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 274, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 274, + 152 + ], + "type": "image", + "image_path": "186ab939a02d824c0535dc5009730c3f298013352f616b8d6517c19652aa79e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 154, + 233, + 162 + ], + "lines": [ + { + "bbox": [ + 173, + 154, + 233, + 162 + ], + "spans": [ + { + "bbox": [ + 173, + 154, + 233, + 162 + ], + "type": "text", + "content": "(a) Components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 163, + 274, + 192 + ], + "blocks": [ + { + "bbox": [ + 133, + 163, + 274, + 192 + ], + "lines": [ + { + "bbox": [ + 133, + 163, + 274, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 274, + 192 + ], + "type": "image", + "image_path": "c56851318c7dd8c76bee7484ffb313fc82bf129adfee63911ef80a4eb9e03faf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 193, + 273, + 221 + ], + "blocks": [ + { + "bbox": [ + 133, + 193, + 273, + 221 + ], + "lines": [ + { + "bbox": [ + 133, + 193, + 273, + 221 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 273, + 221 + ], + "type": "image", + "image_path": "3ea110fe614b3cb5d7616ff8a89b4325ab1a3455cb1d39242400b6ba5f6374fb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 223, + 273, + 252 + ], + "blocks": [ + { + "bbox": [ + 133, + 223, + 273, + 252 + ], + "lines": [ + { + "bbox": [ + 133, + 223, + 273, + 252 + ], + "spans": [ + { + "bbox": [ + 133, + 223, + 273, + 252 + ], + "type": "image", + "image_path": "1055a1553975ccbadddd88742894d815f9c8cb82185b62118ce28e857b118079.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 253, + 272, + 270 + ], + "lines": [ + { + "bbox": [ + 132, + 253, + 272, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 272, + 270 + ], + "type": "text", + "content": "(c) Generated samples for challenging cases: Pose, teeth, beard." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 115, + 465, + 152 + ], + "blocks": [ + { + "bbox": [ + 313, + 115, + 465, + 152 + ], + "lines": [ + { + "bbox": [ + 313, + 115, + 465, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 465, + 152 + ], + "type": "image", + "image_path": "54419d10d011a30a52a30eda9f7adb39c169b01e7c7808035f8dc405e9a74aae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 154, + 449, + 162 + ], + "lines": [ + { + "bbox": [ + 343, + 154, + 449, + 162 + ], + "spans": [ + { + "bbox": [ + 343, + 154, + 449, + 162 + ], + "type": "text", + "content": "(b) Face restoration methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 163, + 480, + 252 + ], + "blocks": [ + { + "bbox": [ + 313, + 163, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 480, + 252 + ], + "type": "image", + "image_path": "8c8874f5699470fd9854a1d10024b1607b80b9ca6d9ac031266f60e8f73b5b8d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 253, + 481, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 253, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 481, + 270 + ], + "type": "text", + "content": "(d) Silent face comparison with VRT [10]. © European Central Bank (CC BY)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 279, + 480, + 301 + ], + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 301 + ], + "type": "text", + "content": "Fig. 5: Ablation studies of components (a), face restoration methods (b), and silent face generation (d). (c) demonstrates generated images in challenging cases." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 334, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 562 + ], + "type": "text", + "content": "decent synchronization performance, despite lower visual quality and training stability issues. In this setup, we train the audio encoder as a part of " + }, + { + "bbox": [ + 130, + 334, + 482, + 562 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 130, + 334, + 482, + 562 + ], + "type": "text", + "content": ". Replacing this with our pretrained audio encoder enhances the synchronization and visual quality (Setup B). For further comparison, we also utilize the audio encoder of the Wav2Vec2 [3], presented in Setup C. However, it slightly decreases the scores, which validates our hypothesis about training the audio encoder for synchronization purposes. Thus, we continue with Setup B and add our silent-lip generator " + }, + { + "bbox": [ + 130, + 334, + 482, + 562 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 334, + 482, + 562 + ], + "type": "text", + "content": " to generate silent identity references (Setup D). " + }, + { + "bbox": [ + 130, + 334, + 482, + 562 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 334, + 482, + 562 + ], + "type": "text", + "content": " alleviates the lip leaking problem, makes the training more stable, and improves the scores noticeably. We further replace lip-sync loss with our stabilized synchronization loss, yielding drastically improved synchronization and visual quality scores (Setup E). Moreover, we observed that Setup E shows almost no instability in the training. In Setup F, we train Setup E including vanilla triplet loss and it enhances neither visual quality nor synchronization; in fact, it even causes detrimental effects. This shows the necessity of modifying the triplet loss and introducing the adaptive triplet loss. In Setup G, replacing vanilla triplet loss with adaptive triplet loss demonstrates a slight improvement in visual quality, while not having a negative impact on lip synchronization. In Setup H, we switch SyncNet with AVSyncNet and achieve slightly better visual quality and lip synchronization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 565, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 482, + 602 + ], + "type": "text", + "content": "We compare " + }, + { + "bbox": [ + 130, + 565, + 482, + 602 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 130, + 565, + 482, + 602 + ], + "type": "text", + "content": " with VRT [10] silent face generation approach in Tab. 2. Our model surpasses VRT quantitatively and qualitatively (see Fig. 5d), preserving the visual details and identity while modifying the lips." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "We compare different face restoration methods as post processing and present the results in Tab. 2 and Fig. 5b. VQFR surpasses other methods in preserving lip synchronization as well as FID and PSNR. However, GPEN shows better performance in the remaining metrics (see App. F.2 for details). In summary, we employ VQFR for post-processing in our full model." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 448, + 102 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 121, + 233, + 186 + ], + "blocks": [ + { + "bbox": [ + 148, + 121, + 233, + 186 + ], + "lines": [ + { + "bbox": [ + 148, + 121, + 233, + 186 + ], + "spans": [ + { + "bbox": [ + 148, + 121, + 233, + 186 + ], + "type": "image", + "image_path": "d386668efa85fb07cc9df1121e5b27b15feca8a88f5a440f1b773ab58667a0e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 186, + 224, + 197 + ], + "lines": [ + { + "bbox": [ + 165, + 186, + 224, + 197 + ], + "spans": [ + { + "bbox": [ + 165, + 186, + 224, + 197 + ], + "type": "text", + "content": "(a) Shifting [39]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "text", + "content": "Fig. 6: Graphs show the performance of SyncNet [46] and our AVSyncNet on LRS2 GT audio-lip pairs under certain transformations. See App. A for further details. We apply lip-sync loss [46] for the analyses to fairly compare SyncNet and our AVSyncNet independent from our loss function." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 121, + 330, + 186 + ], + "blocks": [ + { + "bbox": [ + 244, + 121, + 330, + 186 + ], + "lines": [ + { + "bbox": [ + 244, + 121, + 330, + 186 + ], + "spans": [ + { + "bbox": [ + 244, + 121, + 330, + 186 + ], + "type": "image", + "image_path": "5bc8583d456ce149bb0e7de200d22a85fa0f94408843731876bb88053aa0f174.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 267, + 187, + 315, + 197 + ], + "lines": [ + { + "bbox": [ + 267, + 187, + 315, + 197 + ], + "spans": [ + { + "bbox": [ + 267, + 187, + 315, + 197 + ], + "type": "text", + "content": "(b) Rotation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 343, + 117, + 455, + 186 + ], + "blocks": [ + { + "bbox": [ + 343, + 117, + 455, + 186 + ], + "lines": [ + { + "bbox": [ + 343, + 117, + 455, + 186 + ], + "spans": [ + { + "bbox": [ + 343, + 117, + 455, + 186 + ], + "type": "image", + "image_path": "9305dd95a910a0bb9e933c9f260f615520d191e8bd05942eedca21ada26ab331.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 187, + 440, + 196 + ], + "lines": [ + { + "bbox": [ + 367, + 187, + 440, + 196 + ], + "spans": [ + { + "bbox": [ + 367, + 187, + 440, + 196 + ], + "type": "text", + "content": "(c) Further analysis." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 273, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 370 + ], + "type": "text", + "content": "AVSyncNet. Comprehensive experiments on LRS2 GT data pairs highlight AVSyncNet's superior performance (see Fig. 6). Furthermore, it demonstrates strong shift-invariance and robustness against affine transformations in the data due to AVSyncNet's design, particularly emphasizing its effectiveness in focusing on lip synchronization while being less affected by other factors. Moreover, AVSyncNet's performance is not affected by face pose unlike SyncNet [46]. Tab. 2 also demonstrates that our AVSyncNet improves the performance of our talking face generation model and works harmoniously with the proposed " + }, + { + "bbox": [ + 130, + 273, + 482, + 370 + ], + "type": "inline_equation", + "content": "L_{ss}" + }, + { + "bbox": [ + 130, + 273, + 482, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 388, + 220, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 220, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 220, + 401 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "content": "In this paper, we improve audio-driven talking face generation by identifying problems in current approaches and mitigating them accordingly. Specifically, we introduce a silent-lip generator to mitigate lip leaking, which is a common problem that harms lip-sync and training stability. Furthermore, we propose stabilized synchronization loss along with AVSyncNet, which significantly improves the training stability, lip synchronization performance, and visual quality by solving the problems caused by lip-sync loss and SyncNet. Experimental results on benchmark datasets and a comprehensive ablation study show the merit of our method and contributions. Moreover, our detailed analyses reveal the main issues, support our claims, and validate proposed contributions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 551, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 601 + ], + "type": "text", + "content": "Limitations. Despite the notable improvements, SyncNet's and AVSyncNet's unstable nature should be investigated further. Moreover, face restoration sometimes causes inconsistencies in the video. Silent-lip generator makes teeth invisible in identity references, occasionally resulting in suboptimal teeth generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Ethics & Social Impact. We believe that generating lip-synchronized faces holds significant benefits across a broad spectrum of applications. However, we acknowledge its vulnerability to potential misuse, particularly deepfake generation. We will utilize Watermarking and prevent uncontrolled usage of our model." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the European Commission Project Meetween (101135798) under the call HORIZON-CL4-2023-HUMAN-01-03." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 175, + 197, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 175, + 197, + 187 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 197, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 204, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 204, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 204, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 204, + 480, + 237 + ], + "type": "text", + "content": "1. Afouras, T., Chung, J.S., Senior, A., Vinyals, O., Zisserman, A.: Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence 44(12), 8717-8727 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 238, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 282 + ], + "type": "text", + "content": "2. Afouras, T., Owens, A., Chung, J.S., Zisserman, A.: Self-supervised learning of audio-visual objects from video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16. pp. 208–224. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 283, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 316 + ], + "type": "text", + "content": "3. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "text", + "content": "4. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 187-194 (1999)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 350, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 373 + ], + "type": "text", + "content": "5. Booth, J., Roussos, A., Ponniah, A., Dunaway, D., Zafeiriou, S.: Large scale 3d morphable models. International Journal of Computer Vision 126(2), 233-254 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 373, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 480, + 395 + ], + "type": "text", + "content": "6. Brand, M.: Voice puppetry. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 21-28 (1999)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "text", + "content": "7. Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In: International Conference on Computer Vision (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 430, + 480, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 430, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 138, + 430, + 480, + 452 + ], + "type": "text", + "content": "8. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronisation in the wild. arXiv preprint arXiv:2112.04432 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 452, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 452, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 452, + 480, + 485 + ], + "type": "text", + "content": "9. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7832-7841 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 486, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 486, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 138, + 486, + 480, + 518 + ], + "type": "text", + "content": "0. Cheng, K., Cun, X., Zhang, Y., Xia, M., Yin, F., Zhu, M., Wang, X., Wang, J., Wang, N.: Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 520, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 520, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 520, + 480, + 552 + ], + "type": "text", + "content": "1. Chung, J.S., Huh, J., Mun, S., Lee, M., Heo, H.S., Choe, S., Ham, C., Jung, S., Lee, B.J., Han, I.: In defence of metric learning for speaker recognition. arXiv preprint arXiv:2003.11982 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "text", + "content": "2. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 587, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 480, + 620 + ], + "type": "text", + "content": "3. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "4. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 251-263. Springer (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "15. Chung, S.W., Chung, J.S., Kang, H.G.: Perfect match: Improved cross-modal embeddings for audio-visual synchronisation. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3965-3969. IEEE (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "text", + "content": "16. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 408-424. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "text", + "content": "17. Eskimez, S.E., Maddox, R.K., Xu, C., Duan, Z.: End-to-end generation of talking faces from noisy speech. In: ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 1948-1952. IEEE (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 481, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 272 + ], + "type": "text", + "content": "18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 317 + ], + "type": "text", + "content": "19. Gu, Y., Wang, X., Xie, L., Dong, C., Li, G., Shan, Y., Cheng, M.M.: Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII. pp. 126–143. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 318, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 361 + ], + "type": "text", + "content": "20. Guan, J., Zhang, Z., Zhou, H., Hu, T., Wang, K., He, D., Feng, H., Liu, J., Ding, E., Liu, Z., et al.: Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1505-1515 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 362, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 362, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 481, + 396 + ], + "type": "text", + "content": "21. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5784-5794 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "text", + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 430, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 452 + ], + "type": "text", + "content": "23. Hershey, J., Movellan, J.: Audio vision: Using audio-visual synchrony to locate sounds. Advances in neural information processing systems 12 (1999)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "text", + "content": "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 486, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 508 + ], + "type": "text", + "content": "25. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "type": "text", + "content": "26. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. arXiv preprint arXiv:2210.07055 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "text", + "content": "27. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "text", + "content": "28. Jamaludin, A., Chung, J.S., Zisserman, A.: You said that?: Synthesising talking faces from audio. International Journal of Computer Vision 127, 1767-1779 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "29. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "30. Ji, X., Zhou, H., Wang, K., Wu, W., Loy, C.C., Cao, X., Xu, F.: Audio-driven emotional video portraits. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14080-14089 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "type": "text", + "content": "31. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 182 + ], + "type": "text", + "content": "32. Kadandale, V.S., Montesinos, J.F., Haro, G.: Vocalist: An audio-visual synchronisation model for lips and voices. arXiv preprint arXiv:2204.02090 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "33. Kim, Y.J., Heo, H.S., Chung, S.W., Lee, B.J.: End-to-end lip synchronisation based on pattern classification. In: 2021 IEEE Spoken Language Technology Workshop (SLT). pp. 598-605. IEEE (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "34. KR, P., Mukhopadhyay, R., Philip, J., Jha, A., Namboodiri, V., Jawahar, C.: Towards automatic face-to-face translation. In: Proceedings of the 27th ACM international conference on multimedia. pp. 1428-1436 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "35. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 326 + ], + "type": "text", + "content": "36. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3387-3396 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "37. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 106-125. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "type": "text", + "content": "38. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "text", + "content": "39. Muaz, U., Jang, W., Tripathi, R., Mani, S., Ouyang, W., Gadde, R.T., Gecer, B., Elizondo, S., Madad, R., Nair, N.: Sidgan: High-resolution dubbed video generation via shift-invariant learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7833-7842 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "40. Nair, V., Hinton, G.E.: Rectified linear units improve restricted boltzmann machines. In: Proceedings of the 27th international conference on machine learning (ICML-10). pp. 807-814 (2010)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "41. Nayak, S., Schuler, C., Saha, D., Baumann, T.: A deep dive into neural synchrony evaluation for audio-visual translation. In: Proceedings of the 2022 International Conference on Multimodal Interaction. pp. 642-647 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 522 + ], + "type": "text", + "content": "42. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "text", + "content": "43. Owens, A., Efros, A.A.: Audio-visual scene analysis with self-supervised multisensory features. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 631-648 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 600 + ], + "type": "text", + "content": "44. Papantoniou, F.P., Filntisis, P.P., Maragos, P., Roussos, A.: Neural emotion director: Speech-preserving semantic control of facial expressions in\" in-the-wild\" videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18781-18790 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "45. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2062-2070 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "46. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "text", + "content": "47. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "text", + "content": "48. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "text", + "content": "49. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII. pp. 666-682. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 240, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 283 + ], + "type": "text", + "content": "50. Shen, S., Zhao, W., Meng, Z., Li, W., Zhu, Z., Zhou, J., Lu, J.: Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1982-1991 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "text", + "content": "51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "type": "text", + "content": "52. Slaney, M., Covell, M.: Facesync: A linear operator for measuring synchronization of video facial images and audio tracks. Advances in neural information processing systems 13 (2000)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "text", + "content": "53. Song, L., Wu, W., Qian, C., He, R., Loy, C.C.: Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security 17, 585-598 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 375, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 397 + ], + "type": "text", + "content": "54. Song, Y., Zhu, J., Li, D., Wang, X., Qi, H.: Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 398, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 481, + 441 + ], + "type": "text", + "content": "55. Stypulkowski, M., Vougioukas, K., He, S., Zieba, M., Petridis, S., Pantic, M.: Diffused heads: Diffusion models beat gans on talking-face generation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5091-5100 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "56. Sun, Y., Zhou, H., Wang, K., Wu, Q., Hong, Z., Liu, J., Ding, E., Wang, J., Liu, Z., Hideki, K.: Masked lip-sync prediction by audio-visual contextual exploitation in transformers. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1–9 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 476, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 481, + 509 + ], + "type": "text", + "content": "57. Suwajanakorn, S., Seitz, S.M., Kemelmacher-Shlizerman, I.: Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "type": "text", + "content": "58. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 544, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 587 + ], + "type": "text", + "content": "59. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 716-731. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 588, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 610 + ], + "type": "text", + "content": "60. Vougioukas, K., Petridis, S., Pantic, M.: Realistic speech-driven facial animation with gans. International Journal of Computer Vision 128, 1398-1413 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 666 + ], + "type": "text", + "content": "61. Waibel, A., Behr, M., Yaman, D., Eyiokur, F.I., Nguyen, T.N., Mullov, C., Demirtas, M.A., Kantarci, A., Constantin, S., Ekenel, H.K.: Face-dubbing++: Lipsynchronous, voice preserving translation of videos. In: 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 160 + ], + "type": "text", + "content": "62. Wang, D., Deng, Y., Yin, Z., Shum, H.Y., Wang, B.: Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17979-17989 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 192 + ], + "type": "text", + "content": "63. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14653-14662 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 236 + ], + "type": "text", + "content": "64. Wang, J., Zhao, K., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Lipformer: High-fidelity and generalizable talking face generation with a pre-learned facial codebook. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13844-13853 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "65. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "66. Wu, H., Jia, J., Wang, H., Dou, Y., Duan, C., Deng, Q.: Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1478-1486 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "67. Yao, S., Zhong, R., Yan, Y., Zhai, G., Yang, X.: Dfa-nerf: personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "text", + "content": "68. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "69. Yehia, H., Rubin, P., Vatikiotis-Bateson, E.: Quantitative association of vocal-tract and facial behavior. Speech Communication 26(1-2), 23-43 (1998)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 385, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 439 + ], + "type": "text", + "content": "70. Yin, F., Zhang, Y., Cun, X., Cao, M., Fan, Y., Wang, X., Bai, Q., Wu, B., Wang, J., Yang, Y.: Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 85-101. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 439, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 461 + ], + "type": "text", + "content": "71. Zhan, F., Yu, Y., Wu, R., Zhang, J., Lu, S.: Multimodal image synthesis and editing: A survey. arXiv preprint arXiv:2112.13592 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 461, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 504 + ], + "type": "text", + "content": "72. Zhang, C., Zhao, Y., Huang, Y., Zeng, M., Ni, S., Budagavi, M., Guo, X.: Facial: Synthesizing dynamic talking face with implicit attribute learning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3867-3876 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "73. Zhang, Z., Hu, Z., Deng, W., Fan, C., Lv, T., Ding, Y.: Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. arXiv preprint arXiv:2303.03988 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "74. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "text", + "content": "75. Zhen, R., Song, W., He, Q., Cao, J., Shi, L., Luo, J.: Human-computer interaction system: A survey of talking-head generation. Electronics 12(1), 218 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 632 + ], + "type": "text", + "content": "76. Zhong, W., Fang, C., Cai, Y., Wei, P., Zhao, G., Lin, L., Li, G.: Identity-preserving talking face generation with landmark and appearance priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "77. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 9299–9306 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-driven Talking Face Generation with Stabilized Synchronization Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "text", + "content": "78. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4176-4186 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "text", + "content": "79. Zhou, M., Bai, Y., Zhang, W., Yao, T., Zhao, T., Mei, T.: Responsive listening head generation: A benchmark dataset and baseline (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "80. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG) 39(6), 1-15 (2020)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "D. Yaman et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_content_list.json b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04b81ef3036361775ffef3aac0e149b0b964ad85 --- /dev/null +++ b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_content_list.json @@ -0,0 +1,1541 @@ +[ + { + "type": "text", + "text": "Audio-visual Generalized Zero-shot Learning the Easy Way", + "text_level": 1, + "bbox": [ + 246, + 140, + 754, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shentong Mo1 and Pedro Morgado2", + "bbox": [ + 370, + 212, + 629, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University $^{2}$ University of Wisconsin-Madison", + "bbox": [ + 382, + 239, + 617, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Audio-visual generalized zero-shot learning is a rapidly advancing domain that seeks to understand the intricate relations between audio and visual cues within videos. The overarching goal is to leverage insights from seen classes to identify instances from previously unseen ones. Prior approaches primarily utilized synchronized auto-encoders to reconstruct audio-visual attributes, which were informed by cross-attention transformers and projected text embeddings. However, these methods fell short of effectively capturing the intricate relationship between cross-modal features and class-label embeddings inherent in pretrained language-aligned embeddings. To circumvent these bottlenecks, we introduce a simple yet effective framework for Easy Audio-Visual Generalized Zero-shot Learning, named EZ-AVGZL, that aligns audio-visual embeddings with transformed text representations. It utilizes a single supervised text audio-visual contrastive loss to learn an alignment between audio-visual and textual modalities, moving away from the conventional approach of reconstructing cross-modal features and text embeddings. Our key insight is that while class name embeddings are well aligned with language-based audio-visual features, they don't provide sufficient class separation to be useful for zero-shot learning. To address this, our method leverages differential optimization to transform class embeddings into a more discriminative space while preserving the semantic structure of language representations. We conduct extensive experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks. Our results demonstrate that our EZ-AVGZL achieves state-of-the-art performance in audio-visual generalized zero-shot learning.", + "bbox": [ + 261, + 311, + 738, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Audio-Visual Learning $\\cdot$ Zero-shot Learning $\\cdot$ Audio-Visual Generalized Zero-shot Learning", + "bbox": [ + 261, + 672, + 738, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 729, + 374, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of machine learning, the ability of models to recognize patterns and make predictions is crucial. Zero-shot learning exemplifies this capability, which leverages known information to understand the unknowns. To further explore its potential and applications, the principle of zero-shot learning has been applied in various contexts. One such context is Audio-visual generalized", + "bbox": [ + 217, + 763, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dbd5321a80463c291c70e52f3bbadec3715ee5b46d42d863b2891df9983574ee.jpg", + "image_caption": [ + "Fig.1: Comparison of our EZ-AVGZL with state-of-the-art methods on VGGSound-GZSL, UCF-GZSL, ActivityNet-GZSL benchmarks in terms of Harmonic Mean (higher is better) for seen and unseen classes. Our method significantly outperforms previous baselines in terms of all datasets." + ], + "image_footnote": [], + "bbox": [ + 289, + 147, + 709, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "zero-shot learning (AVGZSL), an emerging topic of study that predicts unseen audio-visual categories by leveraging knowledge learned from seen classes.", + "bbox": [ + 212, + 470, + 785, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Within the scope of AVGZSL, previous approaches [27, 28] have focused on designing autodecoders for reconstructing audio-visual attributes. For example, TCaF [27] adopted a temporal contrastive feature alignment strategy. This strategy aligns audio-visual cues with word2vec embeddings of class prompts by defining both cross-entropy and regression losses to reduce the distance between the output embedding for a sample and the corresponding projected word2vec embedding. These cross-modal and textual features were derived from cross-attention transformers and the projected text embeddings. However, the complexity of the network for AVGZSL, which is derived from the synergy between cross-attention transformers and the corresponding projected word2vec embeddings, makes it challenging to understand the deeper interplay between cross-modal features and class label embeddings.", + "bbox": [ + 212, + 503, + 787, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the context of audio-visual zero-shot learning, the main challenge lies in the absence of unseen classes during training. Without the supervision, the prediction of cross-modal class labels performs poorly on these audio-visual samples. To address the challenge, we propose a novel framework for Easy Audio-Visual Generalized Zero-shot Learning, referred to as EZ-AVGZL. This framework aligns audio-visual embeddings with transformed text representations in a simple yet effective manner. EZ-AVGZL learn well-separated text semantics across seen and unseen class names to achieve maximal separation while retaining the semantic relations between each other. During training, we leverage a simple supervised contrastive loss to learn the cross-modal alignment between audio-visual features", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Mo and P. Morgado", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and transformed text embeddings. EZ-AVGZL enhances class separability while preserving semantics through the use of differential optimization, thereby generating more distinct text embeddings. Furthermore, it employs a single supervised text to audio-visual contrastive loss, establishing an alignment between the audio-visual and textual modalities. This approach effectively eliminates the need for reconstructing cross-modal features and text embeddings.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Empirical experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks demonstrate the state-of-the-art performance against previous audio-visual generalized zero-shot learning baselines, as illustrated in Fig. 1. Extensive ablation studies further confirm the importance of class embedding optimization and supervised audio-visual language contrastive objectives in learning language alignment with audio-visual representations. Additionally, we demonstrate the robustness of our approach across various zero-shot learning tasks and observe the benefit of our approach when applied to a variety of foundation models pre-trained from unimodal and cross-modal.", + "bbox": [ + 212, + 237, + 787, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 238, + 373, + 633, + 387 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the EZ-AVGZL framework, a paradigm shift in audio-visual generalized zero-shot learning. This framework simplifies the process with a supervised text-audio-visual contrastive loss, enhancing effectiveness.", + "- We propose a novel text transformation technique designed to construct distinct text embeddings, paving the way for better alignment and, by extension, superior recognition capabilities.", + "- We demonstrate that our method achieves state-of-the-art performance on extensive benchmarks in audio-visual generalized zero-shot learning." + ], + "bbox": [ + 223, + 395, + 784, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 537, + 387, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-Visual Learning. Audio-visual learning problems have been addressed in many previous works [1,2,8,10,14,17,21,32,35,36,38,41,42,44,46,47,50,60-62] to learn the audio-visual correlation between two distinct modalities from videos. Such cross-modal alignments are beneficial for a variety audio-visual tasks, including audio-event localization [22, 24, 30, 31, 33, 39, 40, 54, 58], audio-visual spatialization [4, 11, 42, 45], audio-visual continual learning [34, 49], audio-visual navigation [4-6] and audio-visual parsing [23, 37, 53, 57]. In this work, our primary focus is on learning audio-visual representations that align with text embeddings for generalized zero-shot learning, a task that presents more challenges than the aforementioned tasks.", + "bbox": [ + 212, + 566, + 787, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Zero-shot Learning. Zero-shot learning aims to leverage semantic relationships, typically utilizing attributes or word embeddings associated with class labels to make informed predictions about previously unseen categories. This is often achieved by drawing parallels to the semantic information of known categories. Early audio-visual zero-shot learning methods, such as those proposed by Parida et al. [48] and Mazumder et al. [26], attempted to map video, audio, and text into the same feature space and compute the similarity across each other for comparison. For example, Audio-Visual Generalized Zero-shot Learning Network [26]", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Audio-visual Generalized Zero-shot Learning the Easy Way", + "bbox": [ + 336, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "introduced a reconstruction module to predict text representations from audio and visual features. More recently, ImageBind [12] proposed learning a joint embedding across six different modalities, especially on audio, text, and video. This approach demonstrated the ability to recognize or retrieve unseen categories without explicit training. Unlike these methods, our approach does not require large-scale pre-training. Instead, we present a simple yet effective approach based on unimodal or crossmodal weights pre-trained on large-scale data to achieve audio-visual generalized zero-shot learning.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Audio-Visual Generalized Zero-shot Learning. Audio-visual generalized zero-shot learning extends the zero-shot learning paradigm to scenarios where both seen and unseen classes are present at test time. The seminal work, AVCA [28] proposed an attention mechanism to selectively focus on discriminative audio and visual cues. By harnessing the combined strength of these cues, AVCA not only achieved superior performance on seen classes but also exhibited proficiency in recognizing unseen classes.", + "bbox": [ + 212, + 268, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following up, TCaF [27] introduced a temporal contrastive feature alignment strategy. This strategy aligns audio-visual cues across time and applies two separate decoders to reconstruct the audio-visual embedding and the projected word2vec embedding to be close to the original word2vec embedding of class labels. More recently, Hong et al. [16] incorporated cross-modality alignment between video and audio features in the hyperbolic space. They used multiple adaptive curvatures for hyperbolic projections to achieve curvature-aware geometric learning. In contrast, we develop a novel framework to aggregate the alignment between audio-visual representations and text embeddings with explicit class embedding optimization. To the best of our knowledge, this work is the first to leverage an explicit class embedding optimization mechanism for audio-visual generalized zero-shot learning. Our experiments in Section 4.2 also demonstrate the effectiveness of our method in classifying both seen and unseen categories.", + "bbox": [ + 212, + 378, + 787, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 612, + 328, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to learn a language-aligned representation for a video clip, which includes both the visual frame sequence and corresponding audio, to enable open-dictionary recognition of audio-visual events. To achieve this, we propose a simple yet effective approach, named Easy Audio-Visual Generalized Zero-shot Learning (EZ-AVGZL). This approach seeks to 1) construct multi-modal representations of the data and 2) align them with language representations using a supervised contrastive learning objective while leveraging strong foundation models. The proposed approach is based on two components: class embedding optimization described in Section 3.2 and supervised alignment between language and audio-visual representations described in Section 3.3. Despite its simplicity, EZ-AVGZL outperforms recent state-of-the-art methods on Audio-Visual Generalized Zero-shot Learning (AVGZSL) on a variety of datasets.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Mo and P. Morgado", + "bbox": [ + 271, + 114, + 423, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a752799a4de23b84bc0cf07e68a6196edc59bcdff539a01485a5edf264911d41.jpg", + "image_caption": [ + "Fig. 2: Illustration of the proposed Easy Audio-Visual Generalized Zero-shot Learning (EZ-AVGZL) framework. The initial class embeddings $\\pmb{t}_i$ from a frozen text encoder are optimized with maximal separability and preserved semantics to generate the new embeddings $\\pmb{w}_i$ . Then the cross-attention transformer $f_{\\theta}^{av}(\\cdot, \\cdot)$ takes visual and audio features $(\\pmb{v}_i, \\pmb{a}_i)$ from the unimodal encoder to generate the multi-modal representations $\\pmb{x}_i^{av}$ . Finally, a non-linear similarity function is applied to align representations $\\pmb{x}_i^{av}$ with the corresponding class embeddings $\\pmb{w}_{y_i}$ by minimizing the cross-entropy loss between the predicted similarity score and the target score as 1 in the target class entry, while the target is 0 for multi-modal representations $\\pmb{x}_j^{av}$ given visual features $\\pmb{v}_j$ of videos from other classes." + ], + "image_footnote": [], + "bbox": [ + 233, + 146, + 767, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries: Audio-Visual Generalized Zero-shot Learning (AVGZSL)", + "text_level": 1, + "bbox": [ + 212, + 506, + 767, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Statement. The objective of AVGZSL is to utilize natural language descriptions of audio-visual events $t_i$ and the semantic relationships to enable a model to recognize events from a set of audio-visual event classes $\\mathcal{T} = \\mathcal{T}_s \\cup \\mathcal{T}_u$ . This should be possible regardless of whether the model has seen them $(t_i \\in \\mathcal{T}_s)$ or not seen them $(t_i \\in \\mathcal{T}_u)$ during training.", + "bbox": [ + 212, + 550, + 787, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To accomplish this, a model $f_{\\theta}:\\mathcal{V}\\times \\mathcal{A}\\to \\mathcal{T}$ is trained on the seen dataset $\\mathcal{D}_s = \\{(v_i,a_i,t_i)\\}$ for audio-visual event recognition, where $v_{i}\\in \\mathcal{V}$ denotes the $i^{th}$ visual clip, typically multiple frames spanning a short time window, $a_{i}\\in \\mathcal{A}$ denotes the corresponding audio, and $t_i\\in \\mathcal{T}_s$ denotes the (human-readable) class name of the audio-visual event depicted in video $(v_{i},a_{i})$ . For effective zero-shot learning, a model needs to learn a mapping that can generalize beyond $\\mathcal{T}_s$ by inferring from the semantic relationships between seen and unseen classes. In other words, the AVGZSL setting evaluates models based on their classification performance on videos $(v_{i},a_{i})$ belonging to both seen $\\mathcal{T}_s$ and unseen classes $\\mathcal{T}_u$ . Architecture Following TCaF [27], we tackle AVGZSL using a variety of pretrained models to encode all input signals. We experiment with a wider variety of pre-trained models than in [27], in all cases, we train our system on a series of unimodal pre-extracted feature representations for the visual frames $\\pmb {v}_i = f_v(\\pmb {v}_i)$ , audio $\\pmb {a}_i = f_a(a_i)$ and class names $\\pmb {t}_i = f_t(t_i)$ . As our goal is multi-modal GZSL,", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Audio-visual Generalized Zero-shot Learning the Easy Way", + "bbox": [ + 334, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we further encode the unimodal audio and visual representations into a joint multi-modal representation using the fusion model $\\pmb{x}_i^{av} = f_{av}(\\pmb{v}_i, \\pmb{a}_i)$ proposed in [27], which consists of a sequence of $L$ transformer layers with cross-attention between audio and visual modalities.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quality of representations obtained from foundation models is crucial for effective zero-shot learning. However, the focus of this work is the language alignment step. While prior work often relies on elaborate alignment strategies, we demonstrate that a simple supervised contrastive objective between text and audio-visual representations can achieve state-of-the-art performance. This is possible as long as the class name embeddings for seen and unseen events are maximally separated while retaining the semantic relations between each other. An overview of the proposed approach is shown in Fig. 2.", + "bbox": [ + 212, + 207, + 787, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Apart from the quality of representations derived from foundation models, language alignment, which is the focus of this work, is a critical component for effective zero-shot learning. While prior work often relies on complex alignment strategies, we demonstrate that a simple supervised contrastive objective between text and audio-visual representations can achieve state-of-the-art performance, as long as the class name embeddings for seen and unseen events are maximally separated while retaining the semantic relations between each other. An overview of the proposed approach is shown in Fig. 2.", + "bbox": [ + 212, + 328, + 787, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Class Embedding Optimization", + "text_level": 1, + "bbox": [ + 215, + 469, + 517, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Language-aligned classifiers can achieve zero-shot generalization as class embeddings $t_1, \\ldots, t_{|\\mathcal{T}|}$ obtained from large-scale models like LLaMa or CLIP, encode semantic relationships between classes. However, while these class embeddings can effectively encode semantics, they are sub-optimal for classification due to a lack of class separability. To obtain well-separated semantic embeddings, we optimize a new set of class embeddings $\\boldsymbol{w}_1^*, \\ldots, \\boldsymbol{w}_{|\\mathcal{T}|}^*$ to satisfy two competing constraints: class separability and semantic preservation.", + "bbox": [ + 212, + 492, + 787, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Class Separability. Class embeddings that are maximally separated from each other are better suited for recognition of seen classes [43]. Thus, to achieve class embedding separation, we minimize the negative distance of each class embedding to its nearest neighbor", + "bbox": [ + 212, + 598, + 787, + 659 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e p} = - \\sum_ {c = 1} ^ {| \\tau |} \\min _ {k \\neq c} \\| \\boldsymbol {w} _ {c} - \\boldsymbol {w} _ {k} \\| _ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 667, + 787, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, maximizing separability alone would destroy the semantics of the pre-trained embeddings. In fact, the minimum of Eq. 1 (if norm constrained) would be achieved for a set of embeddings $\\pmb{w}_1, \\dots, \\pmb{w}_{|\\mathcal{T}|}$ following a \"Tammes problem\" configuration [52]. Note that the embeddings are normalized. When the embedding dimension $d$ is larger than the number of classes in the dictionary $\\mathcal{T}$ , the Tammes configuration provides a solution where all embeddings are equally spaced from each other, i.e. $\\| \\pmb{w}_i - \\pmb{w}_j \\|_2 = \\mathrm{Cte} \\quad \\forall i \\neq j$ . For example, the distance between the class \"a person playing guitar\" and the class \"a person", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Mo and P. Morgado", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "playing the violin\" would be the same as the distance to the class \"a dog barking\". Hence, since effective zero-shot recognition heavily relies on the semantic relations between seen and unseen classes, class separability must be balanced with the preservation of the semantics available in pre-trained embeddings.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic preservation To preserve the semantics of the initial embeddings $\\pmb{t}_i$ , a simple approach would be to use $\\pmb{t}_i$ as proxies and constrain the new embeddings $\\pmb{w}_i$ to be in their proximity", + "bbox": [ + 212, + 207, + 782, + 252 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e m} ^ {p r o x} = \\sum_ {c = 1} ^ {| \\mathcal {T} |} \\| \\boldsymbol {w} _ {c} - \\boldsymbol {t} _ {c} \\| _ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 260, + 785, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, this constraint would significantly limit the separability of class embeddings $\\mathbf{w}$ , given their direct ties to $\\pmb{t}$ . Instead, we seek new class embeddings $\\mathbf{w}$ that retain a similar ranking of similarities to other classes. This is accomplished using a margin ranking loss of the form", + "bbox": [ + 214, + 306, + 782, + 367 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e m} ^ {r a n k} = \\sum_ {c = 1} ^ {| \\mathcal {T} |} \\sum_ {\\substack {i = 1 \\\\ c \\neq i \\neq j}} ^ {| \\mathcal {T} |} \\sum_ {j = 1} ^ {| \\mathcal {T} |} \\max \\left\\{0, m - \\operatorname {sign} \\left(d _ {c i} ^ {\\boldsymbol {t}} - d _ {c j} ^ {\\boldsymbol {t}}\\right) \\left(d _ {c i} ^ {\\boldsymbol {w}} - d _ {c j} ^ {\\boldsymbol {w}}\\right) \\right\\} \\tag{3}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 375, + 785, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $m$ is a margin coefficient, and $d_{ci}^{\\pmb{t}} = \\| \\pmb{t}_{c} - \\pmb{t}_{i}\\|_{2}$ and $d_{ci}^{\\pmb{w}} = \\| \\pmb{w}_{c} - \\pmb{w}_{i}\\|_{2}$ are the $\\ell_2$ distances between pairs of class embeddings, using either the pre-trained embeddings $\\pmb{t}$ or the new optimized embeddings $\\pmb{w}$ , respectively. Intuitively, this ranking loss will seek to organize the new embeddings $\\pmb{w}$ such that if $\\pmb{t}_{c}$ is closer to $\\pmb{t}_{i}$ than to $\\pmb{t}_{j}$ , then $\\pmb{w}_{c}$ will also be closer to $\\pmb{w}_{i}$ than to $\\pmb{w}_{j}$ . Note that Eq. 3's complexity grows cubically with the number of classes. While manageable for the tested datasets, it can be impractical for large vocabularies. A potential solution is to sample a large (but fixed) number of class triplets per iteration, and replace Eq. 3 with an expectation over these triplets.", + "bbox": [ + 212, + 435, + 784, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To balance separability and semantic preservation, we obtain the class embeddings by minimizing the joint loss", + "bbox": [ + 212, + 571, + 782, + 599 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {w} _ {1}, \\dots , \\boldsymbol {w} _ {| \\mathcal {T} |}} {\\arg \\min } (1 - \\alpha) \\mathcal {L} _ {s e m} + \\alpha \\mathcal {L} _ {s e p}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 609, + 785, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha$ is set to 0.5. Given that the objective is differentiable, we optimize the new embeddings using gradient descent, starting from the initialization $\\pmb{w}_i = \\pmb{t}_i\\forall i$", + "bbox": [ + 214, + 640, + 782, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Supervised Audio-Visual Language Alignment", + "text_level": 1, + "bbox": [ + 214, + 689, + 638, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the enhanced class embeddings $\\pmb{w}_1, \\dots, \\pmb{w}_{|\\mathcal{T}|}$ (with improved class separability), we empirically show that a contrastive objective is sufficient to learn state-of-the-art models for generalized zero-shot learning of audio-visual events. Specifically, let $\\pmb{x}_i^{av} = f_\\theta^{av}(\\pmb{v}_i, \\pmb{a}_i)$ be the multi-modal representation of an audio-visual sample $(v_i, a_i)$ belonging to class $y_i$ . Then, we align representations $\\pmb{x}_i^{av}$ with the corresponding class embeddings $\\pmb{w}_{y_i}$ by minimizing the contrastive loss", + "bbox": [ + 212, + 710, + 782, + 800 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A V L A} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {x} _ {i} ^ {a v} , \\boldsymbol {w} _ {y _ {i}}\\right)\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {x} _ {i} ^ {a v}, \\boldsymbol {w} _ {y _ {k}}\\right)\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 809, + 785, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Audio-visual Generalized Zero-shot Learning the Easy Way", + "bbox": [ + 334, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(\\cdot, \\cdot)$ is a similarity function used to contrast the class embedding $\\mathbf{w}_{y_i}$ against others in the current batch of data $\\mathbf{w}_{y_k}$ .", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarity. While contrastive objectives typically rely on cosine similarity metrics for $\\sin (\\cdot ,\\cdot):\\mathcal{X}\\times \\mathcal{T}\\rightarrow \\Re$ , we found that non-linear similarity functions can better capture the relationship between audio-visual representations and class embeddings. Specifically, we draw inspiration from transformer attention modules and use a multi-head cross-attention layer. In this layer, class embeddings $\\pmb{w}_c$ are used as queries and audio-visual representations $\\pmb{x}_i^{av}$ as key-value pairs. After cross-attention, class similarities are computed through a linear projection $\\pmb{p}$ . Optimizing Eq. 5 pushes the learned embeddings closer to the text embeddings of the same category. During inference, we use the cross-attention transformer layer to output $C$ similarity scores corresponding to each category, including unseen ones. We then classify the audio-visual samples according to the class with highest similarity.", + "bbox": [ + 212, + 176, + 787, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 378, + 375, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed EZ-AVGZL on three benchmarks for audio-visual generalized zero-shot learning, including VGGSound-GZSL [28], UCF-GZSL [28], and ActivityNet-GZSL [28].", + "bbox": [ + 212, + 406, + 787, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 470, + 428, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. VGGSound-GZSL [28] contains 93,752 videos across 276 classes, selected from VGGSound [7], a large audio-visual dataset of 309 classes over 200k videos. We follow the prior work [27, 28] and use the same split for train/val(U)/test(U) over 138/69/69 classes. UCF-GZSL [28] includes 6,816 videos from 51 classes, chosen from UCF101 [51], the original video action recognition dataset, which consists of over 13k videos in 101 classes. We use the same train/val(U)/test(U) split over 30/12/9 classes in previous work [27, 28]. ActivityNet-GZSL [28] contains 20k videos in 200 classes of varying duration for action recognition [13]. We use the same train/val(U)/test(U) split over 99/51/50 classes as in the prior work [27, 28]", + "bbox": [ + 212, + 492, + 787, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metrics. In line with prior work [27, 28], we use the mean class accuracy to evaluate all models. For ZSL performance, we evaluate only on the subset of test samples from unseen classes. For the GZSL performance, we evaluate on the full test set, which includes both seen and unseen classes. We report results for the subsets of seen (S) and unseen (U) classes, as well as their harmonic mean.", + "bbox": [ + 212, + 643, + 787, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation. We follow previous work [27] and use 8 heads with a dimension of 64 per head for all attention blocks in the cross-attention transformer. The model is trained for 50 epochs with a batch size of 64, using the Adam optimizer [20] with a learning rate of 1e-4, running average coefficients $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , and a weight decay of 1e-5. Following the prior work [27, 28], we use the word2vec [29] embeddings for text, and VGGish [15] embeddings for audio, and C3D [56] features for video frames.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Mo and P. Morgado", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f01bebe3f4f8f105b14790dac24b5e379241fdabd6906e08f73ee0136dfc5065.jpg", + "table_caption": [ + "Table 1: Quantitative results of VGGSound-GZSL benchmark." + ], + "table_footnote": [], + "table_body": "
MethodSeenUnseenHarmonicMeanZSL
AttentionFusion [9]14.133.004.953.37
Perceiver [19]13.253.034.933.44
CJME [48]10.862.223.683.72
AVGZSLNet [26]15.023.195.264.81
AVCA [28]12.636.198.316.91
TCaF [27]12.636.728.777.41
Hyper-alignment [16]12.506.448.507.25
Hyper-single [16]12.565.037.185.47
Hyper-multiple [16]15.626.008.677.31
EZ-AVGZL (ours)17.268.6811.558.92
", + "bbox": [ + 333, + 172, + 674, + 315 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a0eaa7b65001b6431d141dae715c0ec2d9eed8bf8f723b75fb4750a959de35bc.jpg", + "table_caption": [ + "Table 2: Quantitative results of UCF-GZSL benchmark." + ], + "table_footnote": [], + "table_body": "
MethodSeenUnseenHarmonic MeanZSL
AttentionFusion [9]39.3418.2924.9720.21
Perceiver [19]46.8526.8234.1128.12
CJME [48]33.8924.8228.6529.01
AVGZSLNet [26]74.7924.1536.5131.51
AVCA [28]63.1530.7241.3437.72
TCaF [27]67.1440.8350.7844.64
Hyper-alignment [16]57.1333.8642.5239.80
Hyper-single [16]63.4734.8544.9939.86
Hyper-multiple [16]74.2635.7948.3052.11
EZ-AVGZL (ours)78.3246.3558.2448.27
", + "bbox": [ + 333, + 359, + 674, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparison to Prior Work", + "text_level": 1, + "bbox": [ + 215, + 531, + 483, + 546 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we propose a novel and effective framework for audio-visual generalized zero-shot learning. In order to demonstrate the effectiveness of the proposed EZ-AVGZL, we comprehensively compare it to previous audio-visual baselines [9, 19] adapted to our problem and current state-of-the-art audio-visual generalized zero-shot learning approaches [16, 26-28, 48].", + "bbox": [ + 212, + 566, + 785, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the VGGSound-GZSL dataset, we report the quantitative comparison results in Table 1. As shown, we achieve the best results in all metrics for both seen and unseen classes compared to previous audio-visual generalized zero-shot learning approaches. In particular, the proposed EZ-AVGZL outperforms Hyper-multiple [16], the current state-of-the-art audio-visual generalized zero-shot learning baseline, by 1.64@Seen, 2.68@Unseen, 2.88@Harmonic Mean, and 1.61@ZSL. Moreover, we also perform better than TCaF [27], a strong audio-visual generalized zero-shot learning baseline, by 4.63@Seen, 1.96@Unseen, 2.78@Harmonic Mean, and 1.51@ZSL. Furthermore, we achieve significant performance gains compared to AVCA [28], the first baseline using cross-modal attention and textual label embeddings. This indicates the importance of optimized text transformation and non-linear similarity functions in learning better transferrable knowledge for generalized zero-shot learning.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Audio-visual Generalized Zero-shot Learning the Easy Way", + "bbox": [ + 336, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/bfdd0eb3b4ff791d384782ef7f5343748031a76b7a0152aae5d98d8f2d1ff55a.jpg", + "table_caption": [ + "Table 3: Quantitative results of ActivityNet-GZSL benchmark." + ], + "table_footnote": [], + "table_body": "
MethodSeenUnseenHarmonic MeanZSL
AttentionFusion [9]11.153.375.184.88
Perceiver [19]18.254.276.924.47
CJME [48]10.755.557.326.29
AVGZSLNet [26]13.705.968.306.39
AVCA [28]16.777.049.927.58
TCaF [27]30.127.6512.207.96
Hyper-alignment [16]29.778.7713.559.13
Hyper-single [16]24.6110.1014.3210.37
Hyper-multiple [16]36.989.6015.2510.39
EZ-AVGZL (ours)38.5212.8519.2713.01
", + "bbox": [ + 333, + 172, + 674, + 313 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 4: Ablation studies on Class Embedding Optimization and Audio-Visual-Language Alignment.", + "bbox": [ + 212, + 327, + 464, + 385 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/567d73740cd3bc2f1655377c221faf81b15be485d8c5234ea48cec6485e1743d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CEPAVLASeenUnseenHarmonic MeanZSL
XX12.636.728.777.41
✓X15.357.9210.458.36
X✓14.677.239.697.91
✓✓17.268.6811.558.92
", + "bbox": [ + 218, + 398, + 464, + 454 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5e4da7ca994a41c6d4fa927a970dffd5354f20fdfcb1f4a7de3cfb878e263d31.jpg", + "table_caption": [ + "Table 5: Ablation studies on optimization objectives in Class Embedding Optimization." + ], + "table_footnote": [], + "table_body": "
LsemLsepSeenUnseenHarmonic MeanZSL
XX14.677.239.327.91
LproxsemX15.757.8310.468.26
X✓16.138.1710.858.45
Lproxsem✓17.268.6811.558.92
Lranksem✓18.0510.2813.1010.36
", + "bbox": [ + 473, + 372, + 785, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Additionally, we observe significant gains in the UCF-GZSL and ActivityNet-GZSL benchmarks, as shown in Table 2 and Table 3. Compared to TCaF [27], a strong baseline using three manually-designed objectives, we achieve result gains of 11.18@Seen, 5.52@Unseen, 7.46@Harmonic Mean, and 3.63@ZSL on UCF-GZSL dataset. When evaluated on the challenging ActivityNet-GZSL benchmark, our method still outperforms TCaF [27] by 8.40@Seen, 5.20@Unseen, 7.07@Harmonic Mean, and 5.05@ZSL. We also achieve better results than AVCA [28], a cross-modal baseline based on cross-modal attention between audio and visual embeddings. These results demonstrate the effectiveness of our approach in learning optimized textual label embeddings and non-linear similarity functions for generalized audio-visual zero-shot recognition.", + "bbox": [ + 212, + 483, + 787, + 650 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Experimental Analysis", + "text_level": 1, + "bbox": [ + 214, + 674, + 449, + 690 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to demonstrate the benefit of introducing the Class Embedding Optimization (CEO) and Audio-Visual Language Alignment (AVLA) modules. We also carry out extensive experiments to investigate optimization objectives in class embedding optimization, non-linear similarity functions, and generalization across diverse cross-modal settings.", + "bbox": [ + 212, + 703, + 787, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Class Embedding Optimization (CEO) & Audio-Visual Language Alignment (AVLA). In order to demonstrate the effectiveness of CEO and AVLA, we conduct an ablation study and report the quantitative results on VGGSound-GZSL dataset in Table 4.", + "bbox": [ + 212, + 779, + 789, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Mo and P. Morgado", + "bbox": [ + 269, + 114, + 424, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As observed, adding CEO to the vanilla baseline improves the results by 2.72@Seen, 1.20@Unseen, 1.68@Harmonic Mean, and 0.95@ZSL, which validates the benefit of optimizing class embeddings so as to obtain separated embeddings that preserve the semantic relationships between classes. On the other hand, introducing AVLA in the baseline increases the zero-shot performance across all metrics. More importantly, incorporating both CEP and AVLA into the baseline significantly raises the performance by 4.63@Seen, 1.96@Unseen, 2.78@Harmonic Mean, and 1.51@ZSL. These improvements validate the importance of text transformation and supervised text audio-visual contrastive in extracting separated textual label embeddings for AVGZSL.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Objectives in Class Embedding Optimization. Optimization objectives in class embedding transformation, including global semantic preservation and class separability, are essential for learning optimized textual class embeddings for zero-shot learning. To explore the effects of each objective more comprehensively, we ablate each text optimization objective and report the quantitative results in Table 5. We observe that adding the semantic preservation loss to the vanilla baseline improves the performance by 1.08@Seen, 0.60@Unseen, 1.14@Harmonic Mean, and 0.35@ZSL. On the other hand, introducing the class separability loss in the baseline enhances the zero-shot performance across all metrics. Furthermore, combining the semantic preservation and class separability objectives for the text transformer significantly increases the performance by 2.59@Seen, 3.05@Unseen, 3.78@Harmonic Mean, and 2.45@ZSL. This result validates the importance of class separability loss in balancing the preservation of the semantics available in pre-trained representations for zero-shot learning. Replacing the Euclidean-based semantic preservation loss $\\mathcal{L}_{sem}^{prox}$ with a ranking based loss $\\mathcal{L}_{sem}^{rank}$ further achieves performance gains of 0.79@Seen, 1.60@Unseen, 1.55@Harmonic Mean, and 1.44@ZSL. These results demonstrate the effectiveness of both semantic preservation and class separability objectives in class embedding optimization to generate meaningful text representations.", + "bbox": [ + 212, + 303, + 787, + 592 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Non-linear Similarity Functions. The non-linear similarity operator used in supervised audio-visual language alignment impacts the extracted cross-modal representations. To explore these effects more comprehensively, we varied the functions from $\\{\\mathrm{Cosine - similarity}$ , Linear, MLP, Cross-Attention\\}. For Cosine-Similarity, we directly compute the dot product between audio-visual embeddings $\\mathbf{x}_i^{av}$ and text embeddings $\\mathbf{t}_i$ . For Linear, we apply a linear layer to the concatenation of both embeddings to predict the similarity score. For MLP, we replace the linear layer with a multi-layer perceptron to generate the score. For Cross-Attention, we use text embeddings $\\mathbf{t}_i$ as the queries and audio-visual representations $\\mathbf{x}_i^{av}$ as key-value pairs in a transformer attention module to output the score. The comparison results of zero-shot performance on VGGSound-GZSL benchmark are reported in Table 6. When using Cosine-Similarity as the function, we achieve the worst results across all metrics. Replacing the cosine-similarity operator with both Linear increases the seen results by $2.55@Seen$ , but decreases the performance by $0.07@Unseen$ and $0.12@ZSL$ . This might be because adding complexity to the function causes the model to overfit the seen classes in the", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Audio-visual Generalized Zero-shot Learning the Easy Way", + "bbox": [ + 336, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e8bd875d74028fa1ee00aee93875e722b967c662b2259f144414a7d9727c3b15.jpg", + "table_caption": [ + "Table 6: Exploration studies on non-linear similarity functions $\\phi \\left( \\cdot \\right)$ ." + ], + "table_footnote": [], + "table_body": "
φ(·)SeenUnseenHarmonic MeanZSL
Cosine-Similarity13.978.5910.648.67
Linear16.528.5211.248.55
MLP17.038.4611.308.49
Cross-Attention17.268.6811.558.92
", + "bbox": [ + 375, + 172, + 630, + 229 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/429b555cd1816445683e7982b0bd9253569a16f2a49a7c99a1390331f077378e.jpg", + "table_caption": [ + "Table 7: Exploration studies on diverse cross-modal settings. CEP denotes Class Embedding Optimization." + ], + "table_footnote": [], + "table_body": "
Text EncoderAudio EncoderVisual EncoderCEPSeenUnseenHarmonic MeanZSL
word2vecVGGishC3D×12.636.728.777.41
✓17.268.6811.558.92
LLaMaAudioMAEDINO×18.3510.6313.4612.26
✓23.9513.3217.1213.73
CLAPCLAPX-CLIP×27.7914.0118.6317.39
✓33.2820.1525.1021.06
ImageBindImageBindImageBind×28.3214.6819.3419.87
✓35.5621.6326.9022.35
", + "bbox": [ + 346, + 289, + 660, + 401 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "training data, , leading to poorer generalization to unseen categories. Additionally, using MLP as the function operator significantly raises the performance by 3.06@Seen, but slightly deteriorates the results on unseen classes. Our method achieves the best results across all metrics by using the cross-attention operator to predict the similarity score. These improved results demonstrate the effectiveness of the cross-attention operator in balancing the trade-off between linear and non-linear similarity functions.", + "bbox": [ + 212, + 431, + 787, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Generalization to Diverse Cross-modal Settings. To validate the generalizability of the proposed method across a flexible number of feature encoders, we experiment with different settings: unimodal, paired (video-text, audio-text), and unified embedding space (image-audio-text). In the unimodal setting, we use DINO [3] for the visual encoder, AudioMAE [18] for the audio encoder, and LLaMa [55] for the text encoder. In the paired setting, we use X-CLIP [25] pre-trained with video-text objectives for the visual encoder, and CLAP [59] pre-trained with audio-text objectives for the audio and text encoder, respectively. In the unified embedding setting, we apply ImageBind [12] pre-trained with image-audio-text alignment in the same embedding space as the encoder for all three modalities. The quantitative results on the VGGSound-GZSL benchmark are compared in Table 7. As observed, the unimodal setting, which uses transformers with strong unimodal features, significantly outperforms the vanilla baseline that uses convolution networks, where EZ-AVGZL achieves performance gains of 6.69@Seen, 4.64@Unseen, 5.77@Harmonic Mean, and 4.81@ZSL using class embedding optimization. In the paired setting, the proposed method further increases the results by 9.33@Seen, 6.93@Unseen, 7.98@Harmonic Mean, and 7.331@ZSL, benefiting from the text-audio correspondence and text-video alignment pre-trained on large-scale data. In the unified embedding setting using ImageBind [12] embeddings, we achieve the best results across all metrics", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Mo and P. Morgado", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/067d722f8caa4b8fce2122c1c1c97e86ab50a7dc6f9f23558a0cf5cd30799d87.jpg", + "table_caption": [ + "Table 8: Comparison of class embeddings w/o and w optimization." + ], + "table_footnote": [], + "table_body": "
base classNN (w/o opt)D (w/o opt)NN (w opt)D (w opt)
alligators, crocodiles hissingchurch bell ringing0.455playing marimba, xylophone0.498
barn swallow callingblack capped chickadee calling0.307black capped chickadee calling0.465
basketball bounceplaying volleyball0.324bouncing on trampoline0.394
bee, wasp, etc. buzzingmotorboat, speedboat acceleration0.322fly, housefly buzzing0.333
bird squawkingpleasant crowing0.400duck quacking0.496
black capped chickadee callingbarn swallow calling0.307magpie calling0.436
bouncing on trampolinebasketball bounce0.416basketball bounce0.494
bowling impactplaying table tennis0.410striking bowling0.467
bull bellowingeagle screaming0.449dog growling0.532
canary callingwhale calling0.403black capped chickadee calling0.513
", + "bbox": [ + 274, + 171, + 728, + 267 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a40c9db7660021e7ee79bdb06dc55e75846e784520815d25cc3adc1c518db28b.jpg", + "image_caption": [ + "Fig. 3: Confusion matrices of zero-shot predictions on VGGSound-GZSL using model with and without class embedding optimization." + ], + "image_footnote": [], + "bbox": [ + 246, + 279, + 754, + 449 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "compared to other settings. This might be because text embeddings with unified training on the other two modalities strengthen the alignment for audio-visual zero-shot learning. Furthermore, when class embedding optimization is added to all settings, we improve the baseline based on the initial text embeddings, further demonstrating the effectiveness of the proposed class embedding optimization in generating text representations for AVZSL.", + "bbox": [ + 212, + 522, + 787, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Comparisons with and without Optimization. To provide insights into why the proposed loss terms improve the separability of class embeddings while keeping semantics, we conduct qualitative studies. These include comparing the visualization of the class embedding space with and without optimization. The comparisons of class embeddings with and without optimization are shown in Table 8. We report the base class and identify 1) the nearest neighbor (NN) class and 2) the distance (D) to the nearest neighbor. We observe larger distances and similar NN classes. The confusion matrices for class-wise performances of the model are reported in Fig. 3.", + "bbox": [ + 212, + 613, + 787, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of Hyper-parameters $\\alpha$ and Margin $m$ . To investigate the impact of $\\alpha$ and $m$ on the performance thoroughly, we set $\\alpha$ from $\\{0.1, 0.5, 0.9\\}$ and $m$ from $\\{0, 0.5, 1\\}$ . The quantitative comparisons on VGGSound-GZSL are reported in Table 9. As can be seen, with the increase of $\\alpha$ from 0.1 to 0.5, the proposed approach achieves better results regarding all metrics. However, when $\\alpha$ is increased to 0.9, the performance decreases a lot, which might be caused by not", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Audio-visual Generalized Zero-shot Learning the Easy Way", + "bbox": [ + 334, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/275b08c55abb8f5e882262c6fef78b1b9f89ede539743eea39b095fa9b0ba5cb.jpg", + "table_caption": [ + "Table 9: Study of $\\alpha$ and margin $m$ on VGGSound-GZSL." + ], + "table_footnote": [], + "table_body": "
αmSeenUnseenHarmonic MeanZSL
0.518.0510.2813.1010.36
0.1116.358.3911.098.58
0.916.268.7311.368.92
0.5017.829.7112.579.67
0.517.679.8512.659.95
", + "bbox": [ + 217, + 186, + 447, + 252 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/2d5d70319f8b39e970a5f01d21b7048a1982a34c5011cb27052bb9ea5ff697f9.jpg", + "table_caption": [ + "Table 10: Study of distance functions on VGGSound-GZSL." + ], + "table_footnote": [], + "table_body": "
DistanceSeenUnseenHarmonic MeanZSL
cosine18.0510.2813.1010.36
l217.589.3212.189.27
l117.358.7811.668.85
", + "bbox": [ + 473, + 188, + 787, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "much semantics preservation during training. Meanwhile, with the decrease of $m$ , we observe a clearly dropping trend in our metrics.", + "bbox": [ + 212, + 276, + 785, + 306 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablating Distance Functions for Eq. 1-2. Distance functions in Eq. 1-2 is crucial for us to maximize separation while preserving semantics. We experimented with different distance functions, $\\ell_2$ and $\\ell_1$ distances for Eq. 1-2. Note that while, Eq. 1-2 show $\\ell_2$ distances in the paper, the embeddings are normalized. The quantitative results are shown in Table 10. We can observe that using the cosine distance function performed the best regarding all metrics. However, when optimizing $\\ell_1$ distance functions, we achieve the worst results in terms of all metrics. These ablation results highlight the importance of designing appropriate distance functions in maximizing separation while preserving the semantics of text class embeddings.", + "bbox": [ + 212, + 306, + 787, + 458 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 478, + 359, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present a novel and effective framework, EZ-AVGZL, that aligns audio-visual representations with optimized text embeddings to achieve audio-visual generalized zero-shot learning. We leverage differential optimization to learn better-separated textual representations with maximal separability and semantic preservation. Additionally, we introduce supervised audio-visual language alignment to learn the correspondence between audio-visual features and textual embeddings, which captures the cross-modal dynamics with class labels as prompts. Experimental results on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL datasets demonstrate our method's superiority against previous baselines. Extensive ablation studies also validate the importance of class embedding optimization and supervised audio-visual language alignment for audio-visual generalized zero-shot learning. We also conduct comprehensive experiments to demonstrate the effectiveness of non-linear similarity functions and generalization to diverse cross-modal settings.", + "bbox": [ + 212, + 507, + 787, + 718 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Broader Impact. This work makes substantial progress in audio-visual generalized zero-shot learning. Our method is novel in its simplicity, demonstrating that a simple supervised contrastive objective, without \"bells and whistles\", outperforms complex alignment strategies. Without our work, progress in this area would likely be driven by increasingly complex methods, which are difficult to implement, reproduce and compare. Beyond the strong results and simplicity of our method, the class embedding optimization procedure is also technically novel to maximize semantic separation while keeping semantics preserved.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Mo and P. Morgado", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arandjelovic, R., Zisserman, A.: Look, listen and learn. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 609-617 (2017) 3", + "2. Aytar, Y., Vondrick, C., Torralba, A.: Soundnet: Learning sound representations from unlabeled video. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2016) 3", + "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) 12", + "4. Chen, C., Jain, U., Schissler, C., Gari, S.V.A., Al-Halah, Z., Ithapu, V.K., Robinson, P., Grauman, K.: Soundspaces: Audio-visual navigation in 3d environments. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 17-36 (2020) 3", + "5. Chen, C., Majumder, S., Ziad, A.H., Gao, R., Kumar Ramakrishnan, S., Grauman, K.: Learning to set waypoints for audio-visual navigation. In: Proceedings of International Conference on Learning Representations (ICLR) (2021) 3", + "6. Chen, C., Schissler, C., Garg, S., Kobernik, P., Clegg, A., Calamia, P., Batra, D., Robinson, P.W., Grauman, K.: Soundspaces 2.0: A simulation platform for visual-acoustic learning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track (2022) 3", + "7. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 721-725. IEEE (2020) 8", + "8. Ephrat, A., Mosseri, I., Lang, O., Dekel, T., Wilson, K., Hassidim, A., Freeman, W.T., Rubinstein, M.: Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619 (2018) 3", + "9. Fayek, H.M., Kumar, A.: Large scale audiovisual learning of sounds with weakly labeled data. In: Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (2021) 9, 10", + "0. Gan, C., Huang, D., Zhao, H., Tenenbaum, J.B., Torralba, A.: Music gesture for visual sound separation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10478-10487 (2020) 3", + "1. Gao, R., Grauman, K.: 2.5d visual sound. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 324-333 (2019) 3", + "2. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 4, 12", + "3. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 961-970 (2015) 8", + "4. Hershey, J., Casey, M.: Audio-visual sound separation via hidden markov models Advances in Neural Information Processing Systems 14 (2001) 3", + "5. Hershey, S., Chaudhuri, S., Ellis, D.P.W., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., Slaney, M., Weiss, R.J., Wilson, K.: Cnn architectures for large-scale audio classification. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2017) 8" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Audio-visual Generalized Zero-shot Learning the Easy Way", + "bbox": [ + 334, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hong, J., Hayden, Z., Han, J., Fang, P., Harandi, M., Petersson, L.: Hyperbolic audio-visual zero-shot learning. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023) 4, 9, 10", + "17. Hu, D., Nie, F., Li, X.: Deep multimodal clustering for unsupervised audiovisual learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9248-9257 (2019) 3", + "18. Huang, P.Y., Xu, H., Li, J., Baevski, A., Auli, M., Galuba, W., Metze, F., Feichtenhofer, C.: Masked autoencoders that listen. In: Proceedings of Advances In Neural Information Processing Systems (NeurIPS) (2022) 12", + "19. Jaegle, A., Gimeno, F., Brock, A., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver: General perception with iterative attention. In: Proceedings of the International Conference on Machine Learning (ICML) (2021) 9, 10", + "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 8", + "21. Korbar, B., Tran, D., Torresani, L.: Cooperative learning of audio and video models from self-supervised synchronization. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2018) 3", + "22. Lin, Y.B., Li, Y.J., Wang, Y.C.F.: Dual-modality seq2seq network for audio-visual event localization. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2002–2006 (2019) 3", + "23. Lin, Y.B., Tseng, H.Y., Lee, H.Y., Lin, Y.Y., Yang, M.H.: Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2021) 3", + "24. Lin, Y.B., Wang, Y.C.F.: Audiovisual transformer with instance attention for audiovisual event localization. In: Proceedings of the Asian Conference on Computer Vision (ACCV) (2020) 3", + "25. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP:: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285 (2022) 12", + "26. Mazumder, P., Sing, P., Kumar Parida, K., Namboodiri, V.P.: Avgzslnet: Audiovisual generalized zero-shot learning by reconstructing label features from multimodal embeddings. In: Proceedings of 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 3089-3098 (2021) 3, 9, 10", + "27. Mercea, O.B., Hummel, T., Koepke, A.S., Akata, Z.: Temporal and cross-modal attention for audio-visual zero-shot learning. In: Proceedings of European Conference on Computer Vision (ECCV) (2022) 2, 4, 5, 6, 8, 9, 10", + "28. Mercea, O.B., Riesch, L., Koepke, A.S., Akata, Z.: Audio-visual generalised zero-shot learning with cross-modal attention and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10553–10563 (2022) 2, 4, 8, 9, 10", + "29. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. In: Proceedings of International Conference on Learning Representations (ICLR) (2013) 8", + "30. Mo, S., Morgado, P.: Benchmarking weakly-supervised audio-visual sound localization. In: European Conference on Computer Vision (ECCV) Workshop (2022) 3", + "31. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) 3" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Mo and P. Morgado", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Mo, S., Morgado, P.: A unified audio-visual learning framework for localization, separation, and recognition. In: Proceedings of the International Conference on Machine Learning (ICML) (2023) 3", + "33. Mo, S., Morgado, P.: Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling. arXiv preprint arXiv:2312.01017 (2023) 3", + "34. Mo, S., Pian, W., Tian, Y.: Class-incremental grouping network for continual audio-visual learning. arXiv preprint arXiv:2309.05281 (2023) 3", + "35. Mo, S., Shi, J., Tian, Y.: DiffAVA: Personalized text-to-audio generation with visual alignment. arXiv preprint arXiv:2305.12903 (2023) 3", + "36. Mo, S., Shi, J., Tian, Y.: Text-to-audio generation synchronized with videos. arXiv preprint arXiv:2403.07938 (2024) 3", + "37. Mo, S., Tian, Y.: Multi-modal grouping network for weakly-supervised audiovisual video parsing. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) 3", + "38. Mo, S., Tian, Y.: Semantic-aware multi-modal grouping for weakly-supervised audio-visual video parsing. In: European Conference on Computer Vision (ECCV) Workshop (2022) 3", + "39. Mo, S., Tian, Y.: Audio-visual grouping network for sound localization from mixtures. arXiv preprint arXiv:2303.17056 (2023) 3", + "40. Mo, S., Tian, Y.: AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023) 3", + "41. Mo, S., Tian, Y.: Semantic grouping network for audio source separation. arXiv preprint arXiv:2407.03736 (2024) 3", + "42. Morgado, P., Li, Y., Vasconcelos, N.: Learning representations from audio-visual spatial alignment. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS). pp. 4733-4744 (2020) 3", + "43. Morgado, P., Li, Y., Costa Pereira, J., Saberian, M., Vasconcelos, N.: Deep hashing with hash-consistent large margin proxy embeddings. International Journal of Computer Vision (2020) 6", + "44. Morgado, P., Misra, I., Vasconcelos, N.: Robust audio-visual instance discrimination. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12934–12945 (2021) 3", + "45. Morgado, P., Vasconcelos, N., Langlois, T., Wang, O.: Self-supervised generation of spatial audio for 360 video. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2018) 3", + "46. Morgado, P., Vasconcelos, N., Misra, I.: Audio-visual instance discrimination with cross-modal agreement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12475–12486 (June 2021) 3", + "47. Owens, A., Wu, J., McDermott, J.H., Freeman, W.T., Torralba, A.: Ambient sound provides supervision for visual learning. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 801-816 (2016) 3", + "48. Parida, K.K., Matiyali, N., Guha, T., Sharma, G.: Coordinated joint multimodal embeddings for generalized audio-visual zero-shot classification and retrieval of videos. In: Proceedings of 2020 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 3240–3249 (2020) 3, 9, 10", + "49. Pian, W., Mo, S., Guo, Y., Tian, Y.: Audio-visual class-incremental learning. arXiv preprint arXiv:2308.11073 (2023) 3", + "50. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4358-4366 (2018) 3" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Audio-visual Generalized Zero-shot Learning the Easy Way", + "bbox": [ + 334, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) 8", + "52. Tammes, P.M.L.: On the origin of number and arrangement of the places of exit on the surface of pollen-grains. Recueil des travaux botaniques neerlandais 27(1), 1-84 (1930) 6", + "53. Tian, Y., Li, D., Xu, C.: Unified multisensory perception: Weakly-supervised audiovisual video parsing. In: Proceedings of European Conference on Computer Vision (ECCV). p. 436-454 (2020) 3", + "54. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of European Conference on Computer Vision (ECCV) (2018) 3", + "55. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 12", + "56. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2015) 8", + "57. Wu, Y., Yang, Y.: Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1326–1335 (2021) 3", + "58. Wu, Y., Zhu, L., Yan, Y., Yang, Y.: Dual attention matching for audio-visual event localization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 6291–6299 (2019) 3", + "59. Wu, Y., Chen, K., Zhang, T., Hui, Y., Berg-Kirkpatrick, T., Dubnov, S.: Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In: IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP (2023) 12", + "60. Zhang, L., Mo, S., Zhang, Y., Morgado, P.: Audio-synchronized visual animation arXiv preprint arXiv:2403.05659 (2024) 3", + "61. Zhao, H., Gan, C., Ma, W.C., Torralba, A.: The sound of motions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1735-1744 (2019) 3", + "62. Zhao, H., Gan, C., Rouditchenko, A., Vondrick, C., McDermott, J., Torralba, A.: The sound of pixels. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 570-586 (2018) 3" + ], + "bbox": [ + 215, + 147, + 785, + 756 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Mo and P. Morgado", + "bbox": [ + 271, + 114, + 423, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_model.json b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a7d915377202096de70ee64a22615b784d8e92ba --- /dev/null +++ b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_model.json @@ -0,0 +1,2271 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.248, + 0.141, + 0.756, + 0.186 + ], + "angle": 0, + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.213, + 0.63, + 0.228 + ], + "angle": 0, + "content": "Shentong Mo1 and Pedro Morgado2" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.24, + 0.618, + 0.268 + ], + "angle": 0, + "content": "\\(^{1}\\) Carnegie Mellon University \\(^{2}\\) University of Wisconsin-Madison" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.312, + 0.74, + 0.658 + ], + "angle": 0, + "content": "Abstract. Audio-visual generalized zero-shot learning is a rapidly advancing domain that seeks to understand the intricate relations between audio and visual cues within videos. The overarching goal is to leverage insights from seen classes to identify instances from previously unseen ones. Prior approaches primarily utilized synchronized auto-encoders to reconstruct audio-visual attributes, which were informed by cross-attention transformers and projected text embeddings. However, these methods fell short of effectively capturing the intricate relationship between cross-modal features and class-label embeddings inherent in pretrained language-aligned embeddings. To circumvent these bottlenecks, we introduce a simple yet effective framework for Easy Audio-Visual Generalized Zero-shot Learning, named EZ-AVGZL, that aligns audio-visual embeddings with transformed text representations. It utilizes a single supervised text audio-visual contrastive loss to learn an alignment between audio-visual and textual modalities, moving away from the conventional approach of reconstructing cross-modal features and text embeddings. Our key insight is that while class name embeddings are well aligned with language-based audio-visual features, they don't provide sufficient class separation to be useful for zero-shot learning. To address this, our method leverages differential optimization to transform class embeddings into a more discriminative space while preserving the semantic structure of language representations. We conduct extensive experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks. Our results demonstrate that our EZ-AVGZL achieves state-of-the-art performance in audio-visual generalized zero-shot learning." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.673, + 0.739, + 0.7 + ], + "angle": 0, + "content": "Keywords: Audio-Visual Learning \\(\\cdot\\) Zero-shot Learning \\(\\cdot\\) Audio-Visual Generalized Zero-shot Learning" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.73, + 0.375, + 0.745 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.765, + 0.786, + 0.84 + ], + "angle": 0, + "content": "In the field of machine learning, the ability of models to recognize patterns and make predictions is crucial. Zero-shot learning exemplifies this capability, which leverages known information to understand the unknowns. To further explore its potential and applications, the principle of zero-shot learning has been applied in various contexts. One such context is Audio-visual generalized" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "S. Mo and P. Morgado" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.148, + 0.71, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig.1: Comparison of our EZ-AVGZL with state-of-the-art methods on VGGSound-GZSL, UCF-GZSL, ActivityNet-GZSL benchmarks in terms of Harmonic Mean (higher is better) for seen and unseen classes. Our method significantly outperforms previous baselines in terms of all datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.786, + 0.502 + ], + "angle": 0, + "content": "zero-shot learning (AVGZSL), an emerging topic of study that predicts unseen audio-visual categories by leveraging knowledge learned from seen classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.788, + 0.686 + ], + "angle": 0, + "content": "Within the scope of AVGZSL, previous approaches [27, 28] have focused on designing autodecoders for reconstructing audio-visual attributes. For example, TCaF [27] adopted a temporal contrastive feature alignment strategy. This strategy aligns audio-visual cues with word2vec embeddings of class prompts by defining both cross-entropy and regression losses to reduce the distance between the output embedding for a sample and the corresponding projected word2vec embedding. These cross-modal and textual features were derived from cross-attention transformers and the projected text embeddings. However, the complexity of the network for AVGZSL, which is derived from the synergy between cross-attention transformers and the corresponding projected word2vec embeddings, makes it challenging to understand the deeper interplay between cross-modal features and class label embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In the context of audio-visual zero-shot learning, the main challenge lies in the absence of unseen classes during training. Without the supervision, the prediction of cross-modal class labels performs poorly on these audio-visual samples. To address the challenge, we propose a novel framework for Easy Audio-Visual Generalized Zero-shot Learning, referred to as EZ-AVGZL. This framework aligns audio-visual embeddings with transformed text representations in a simple yet effective manner. EZ-AVGZL learn well-separated text semantics across seen and unseen class names to achieve maximal separation while retaining the semantic relations between each other. During training, we leverage a simple supervised contrastive loss to learn the cross-modal alignment between audio-visual features" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.337, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "and transformed text embeddings. EZ-AVGZL enhances class separability while preserving semantics through the use of differential optimization, thereby generating more distinct text embeddings. Furthermore, it employs a single supervised text to audio-visual contrastive loss, establishing an alignment between the audio-visual and textual modalities. This approach effectively eliminates the need for reconstructing cross-modal features and text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Empirical experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks demonstrate the state-of-the-art performance against previous audio-visual generalized zero-shot learning baselines, as illustrated in Fig. 1. Extensive ablation studies further confirm the importance of class embedding optimization and supervised audio-visual language contrastive objectives in learning language alignment with audio-visual representations. Additionally, we demonstrate the robustness of our approach across various zero-shot learning tasks and observe the benefit of our approach when applied to a variety of foundation models pre-trained from unimodal and cross-modal." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.374, + 0.634, + 0.388 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.396, + 0.784, + 0.44 + ], + "angle": 0, + "content": "- We introduce the EZ-AVGZL framework, a paradigm shift in audio-visual generalized zero-shot learning. This framework simplifies the process with a supervised text-audio-visual contrastive loss, enhancing effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.442, + 0.785, + 0.485 + ], + "angle": 0, + "content": "- We propose a novel text transformation technique designed to construct distinct text embeddings, paving the way for better alignment and, by extension, superior recognition capabilities." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.487, + 0.784, + 0.514 + ], + "angle": 0, + "content": "- We demonstrate that our method achieves state-of-the-art performance on extensive benchmarks in audio-visual generalized zero-shot learning." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.396, + 0.785, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.388, + 0.554 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Audio-Visual Learning. Audio-visual learning problems have been addressed in many previous works [1,2,8,10,14,17,21,32,35,36,38,41,42,44,46,47,50,60-62] to learn the audio-visual correlation between two distinct modalities from videos. Such cross-modal alignments are beneficial for a variety audio-visual tasks, including audio-event localization [22, 24, 30, 31, 33, 39, 40, 54, 58], audio-visual spatialization [4, 11, 42, 45], audio-visual continual learning [34, 49], audio-visual navigation [4-6] and audio-visual parsing [23, 37, 53, 57]. In this work, our primary focus is on learning audio-visual representations that align with text embeddings for generalized zero-shot learning, a task that presents more challenges than the aforementioned tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Zero-shot Learning. Zero-shot learning aims to leverage semantic relationships, typically utilizing attributes or word embeddings associated with class labels to make informed predictions about previously unseen categories. This is often achieved by drawing parallels to the semantic information of known categories. Early audio-visual zero-shot learning methods, such as those proposed by Parida et al. [48] and Mazumder et al. [26], attempted to map video, audio, and text into the same feature space and compute the similarity across each other for comparison. For example, Audio-Visual Generalized Zero-shot Learning Network [26]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.424, + 0.128 + ], + "angle": 0, + "content": "S. Mo and P. Morgado" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "introduced a reconstruction module to predict text representations from audio and visual features. More recently, ImageBind [12] proposed learning a joint embedding across six different modalities, especially on audio, text, and video. This approach demonstrated the ability to recognize or retrieve unseen categories without explicit training. Unlike these methods, our approach does not require large-scale pre-training. Instead, we present a simple yet effective approach based on unimodal or crossmodal weights pre-trained on large-scale data to achieve audio-visual generalized zero-shot learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.376 + ], + "angle": 0, + "content": "Audio-Visual Generalized Zero-shot Learning. Audio-visual generalized zero-shot learning extends the zero-shot learning paradigm to scenarios where both seen and unseen classes are present at test time. The seminal work, AVCA [28] proposed an attention mechanism to selectively focus on discriminative audio and visual cues. By harnessing the combined strength of these cues, AVCA not only achieved superior performance on seen classes but also exhibited proficiency in recognizing unseen classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Following up, TCaF [27] introduced a temporal contrastive feature alignment strategy. This strategy aligns audio-visual cues across time and applies two separate decoders to reconstruct the audio-visual embedding and the projected word2vec embedding to be close to the original word2vec embedding of class labels. More recently, Hong et al. [16] incorporated cross-modality alignment between video and audio features in the hyperbolic space. They used multiple adaptive curvatures for hyperbolic projections to achieve curvature-aware geometric learning. In contrast, we develop a novel framework to aggregate the alignment between audio-visual representations and text embeddings with explicit class embedding optimization. To the best of our knowledge, this work is the first to leverage an explicit class embedding optimization mechanism for audio-visual generalized zero-shot learning. Our experiments in Section 4.2 also demonstrate the effectiveness of our method in classifying both seen and unseen categories." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.613, + 0.33, + 0.629 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We aim to learn a language-aligned representation for a video clip, which includes both the visual frame sequence and corresponding audio, to enable open-dictionary recognition of audio-visual events. To achieve this, we propose a simple yet effective approach, named Easy Audio-Visual Generalized Zero-shot Learning (EZ-AVGZL). This approach seeks to 1) construct multi-modal representations of the data and 2) align them with language representations using a supervised contrastive learning objective while leveraging strong foundation models. The proposed approach is based on two components: class embedding optimization described in Section 3.2 and supervised alignment between language and audio-visual representations described in Section 3.3. Despite its simplicity, EZ-AVGZL outperforms recent state-of-the-art methods on Audio-Visual Generalized Zero-shot Learning (AVGZSL) on a variety of datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.147, + 0.768, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Fig. 2: Illustration of the proposed Easy Audio-Visual Generalized Zero-shot Learning (EZ-AVGZL) framework. The initial class embeddings \\( \\pmb{t}_i \\) from a frozen text encoder are optimized with maximal separability and preserved semantics to generate the new embeddings \\( \\pmb{w}_i \\). Then the cross-attention transformer \\( f_{\\theta}^{av}(\\cdot, \\cdot) \\) takes visual and audio features \\( (\\pmb{v}_i, \\pmb{a}_i) \\) from the unimodal encoder to generate the multi-modal representations \\( \\pmb{x}_i^{av} \\). Finally, a non-linear similarity function is applied to align representations \\( \\pmb{x}_i^{av} \\) with the corresponding class embeddings \\( \\pmb{w}_{y_i} \\) by minimizing the cross-entropy loss between the predicted similarity score and the target score as 1 in the target class entry, while the target is 0 for multi-modal representations \\( \\pmb{x}_j^{av} \\) given visual features \\( \\pmb{v}_j \\) of videos from other classes." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.507, + 0.768, + 0.539 + ], + "angle": 0, + "content": "3.1 Preliminaries: Audio-Visual Generalized Zero-shot Learning (AVGZSL)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Problem Statement. The objective of AVGZSL is to utilize natural language descriptions of audio-visual events \\( t_i \\) and the semantic relationships to enable a model to recognize events from a set of audio-visual event classes \\( \\mathcal{T} = \\mathcal{T}_s \\cup \\mathcal{T}_u \\). This should be possible regardless of whether the model has seen them \\( (t_i \\in \\mathcal{T}_s) \\) or not seen them \\( (t_i \\in \\mathcal{T}_u) \\) during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To accomplish this, a model \\( f_{\\theta}:\\mathcal{V}\\times \\mathcal{A}\\to \\mathcal{T} \\) is trained on the seen dataset \\( \\mathcal{D}_s = \\{(v_i,a_i,t_i)\\} \\) for audio-visual event recognition, where \\( v_{i}\\in \\mathcal{V} \\) denotes the \\( i^{th} \\) visual clip, typically multiple frames spanning a short time window, \\( a_{i}\\in \\mathcal{A} \\) denotes the corresponding audio, and \\( t_i\\in \\mathcal{T}_s \\) denotes the (human-readable) class name of the audio-visual event depicted in video \\( (v_{i},a_{i}) \\). For effective zero-shot learning, a model needs to learn a mapping that can generalize beyond \\( \\mathcal{T}_s \\) by inferring from the semantic relationships between seen and unseen classes. In other words, the AVGZSL setting evaluates models based on their classification performance on videos \\( (v_{i},a_{i}) \\) belonging to both seen \\( \\mathcal{T}_s \\) and unseen classes \\( \\mathcal{T}_u \\). Architecture Following TCaF [27], we tackle AVGZSL using a variety of pretrained models to encode all input signals. We experiment with a wider variety of pre-trained models than in [27], in all cases, we train our system on a series of unimodal pre-extracted feature representations for the visual frames \\( \\pmb {v}_i = f_v(\\pmb {v}_i) \\), audio \\( \\pmb {a}_i = f_a(a_i) \\) and class names \\( \\pmb {t}_i = f_t(t_i) \\). As our goal is multi-modal GZSL," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "S. Mo and P. Morgado" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": "we further encode the unimodal audio and visual representations into a joint multi-modal representation using the fusion model \\( \\pmb{x}_i^{av} = f_{av}(\\pmb{v}_i, \\pmb{a}_i) \\) proposed in [27], which consists of a sequence of \\( L \\) transformer layers with cross-attention between audio and visual modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.328 + ], + "angle": 0, + "content": "The quality of representations obtained from foundation models is crucial for effective zero-shot learning. However, the focus of this work is the language alignment step. While prior work often relies on elaborate alignment strategies, we demonstrate that a simple supervised contrastive objective between text and audio-visual representations can achieve state-of-the-art performance. This is possible as long as the class name embeddings for seen and unseen events are maximally separated while retaining the semantic relations between each other. An overview of the proposed approach is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.45 + ], + "angle": 0, + "content": "Apart from the quality of representations derived from foundation models, language alignment, which is the focus of this work, is a critical component for effective zero-shot learning. While prior work often relies on complex alignment strategies, we demonstrate that a simple supervised contrastive objective between text and audio-visual representations can achieve state-of-the-art performance, as long as the class name embeddings for seen and unseen events are maximally separated while retaining the semantic relations between each other. An overview of the proposed approach is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.519, + 0.485 + ], + "angle": 0, + "content": "3.2 Class Embedding Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Language-aligned classifiers can achieve zero-shot generalization as class embeddings \\( t_1, \\ldots, t_{|\\mathcal{T}|} \\) obtained from large-scale models like LLaMa or CLIP, encode semantic relationships between classes. However, while these class embeddings can effectively encode semantics, they are sub-optimal for classification due to a lack of class separability. To obtain well-separated semantic embeddings, we optimize a new set of class embeddings \\( \\boldsymbol{w}_1^*, \\ldots, \\boldsymbol{w}_{|\\mathcal{T}|}^* \\) to satisfy two competing constraints: class separability and semantic preservation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.66 + ], + "angle": 0, + "content": "Class Separability. Class embeddings that are maximally separated from each other are better suited for recognition of seen classes [43]. Thus, to achieve class embedding separation, we minimize the negative distance of each class embedding to its nearest neighbor" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.669, + 0.788, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e p} = - \\sum_ {c = 1} ^ {| \\tau |} \\min _ {k \\neq c} \\| \\boldsymbol {w} _ {c} - \\boldsymbol {w} _ {k} \\| _ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "However, maximizing separability alone would destroy the semantics of the pre-trained embeddings. In fact, the minimum of Eq. 1 (if norm constrained) would be achieved for a set of embeddings \\(\\pmb{w}_1, \\dots, \\pmb{w}_{|\\mathcal{T}|}\\) following a \"Tammes problem\" configuration [52]. Note that the embeddings are normalized. When the embedding dimension \\(d\\) is larger than the number of classes in the dictionary \\(\\mathcal{T}\\), the Tammes configuration provides a solution where all embeddings are equally spaced from each other, i.e. \\(\\| \\pmb{w}_i - \\pmb{w}_j \\|_2 = \\mathrm{Cte} \\quad \\forall i \\neq j\\). For example, the distance between the class \"a person playing guitar\" and the class \"a person" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "playing the violin\" would be the same as the distance to the class \"a dog barking\". Hence, since effective zero-shot recognition heavily relies on the semantic relations between seen and unseen classes, class separability must be balanced with the preservation of the semantics available in pre-trained embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.784, + 0.253 + ], + "angle": 0, + "content": "Semantic preservation To preserve the semantics of the initial embeddings \\( \\pmb{t}_i \\), a simple approach would be to use \\( \\pmb{t}_i \\) as proxies and constrain the new embeddings \\( \\pmb{w}_i \\) to be in their proximity" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.261, + 0.786, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e m} ^ {p r o x} = \\sum_ {c = 1} ^ {| \\mathcal {T} |} \\| \\boldsymbol {w} _ {c} - \\boldsymbol {t} _ {c} \\| _ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.307, + 0.784, + 0.368 + ], + "angle": 0, + "content": "However, this constraint would significantly limit the separability of class embeddings \\(\\mathbf{w}\\), given their direct ties to \\(\\pmb{t}\\). Instead, we seek new class embeddings \\(\\mathbf{w}\\) that retain a similar ranking of similarities to other classes. This is accomplished using a margin ranking loss of the form" + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.375, + 0.786, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e m} ^ {r a n k} = \\sum_ {c = 1} ^ {| \\mathcal {T} |} \\sum_ {\\substack {i = 1 \\\\ c \\neq i \\neq j}} ^ {| \\mathcal {T} |} \\sum_ {j = 1} ^ {| \\mathcal {T} |} \\max \\left\\{0, m - \\operatorname {sign} \\left(d _ {c i} ^ {\\boldsymbol {t}} - d _ {c j} ^ {\\boldsymbol {t}}\\right) \\left(d _ {c i} ^ {\\boldsymbol {w}} - d _ {c j} ^ {\\boldsymbol {w}}\\right) \\right\\} \\tag{3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.785, + 0.571 + ], + "angle": 0, + "content": "where \\( m \\) is a margin coefficient, and \\( d_{ci}^{\\pmb{t}} = \\| \\pmb{t}_{c} - \\pmb{t}_{i}\\|_{2} \\) and \\( d_{ci}^{\\pmb{w}} = \\| \\pmb{w}_{c} - \\pmb{w}_{i}\\|_{2} \\) are the \\( \\ell_2 \\) distances between pairs of class embeddings, using either the pre-trained embeddings \\( \\pmb{t} \\) or the new optimized embeddings \\( \\pmb{w} \\), respectively. Intuitively, this ranking loss will seek to organize the new embeddings \\( \\pmb{w} \\) such that if \\( \\pmb{t}_{c} \\) is closer to \\( \\pmb{t}_{i} \\) than to \\( \\pmb{t}_{j} \\), then \\( \\pmb{w}_{c} \\) will also be closer to \\( \\pmb{w}_{i} \\) than to \\( \\pmb{w}_{j} \\). Note that Eq. 3's complexity grows cubically with the number of classes. While manageable for the tested datasets, it can be impractical for large vocabularies. A potential solution is to sample a large (but fixed) number of class triplets per iteration, and replace Eq. 3 with an expectation over these triplets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.784, + 0.601 + ], + "angle": 0, + "content": "To balance separability and semantic preservation, we obtain the class embeddings by minimizing the joint loss" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.61, + 0.786, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {w} _ {1}, \\dots , \\boldsymbol {w} _ {| \\mathcal {T} |}} {\\arg \\min } (1 - \\alpha) \\mathcal {L} _ {s e m} + \\alpha \\mathcal {L} _ {s e p}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.641, + 0.784, + 0.672 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is set to 0.5. Given that the objective is differentiable, we optimize the new embeddings using gradient descent, starting from the initialization \\(\\pmb{w}_i = \\pmb{t}_i\\forall i\\)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.69, + 0.64, + 0.705 + ], + "angle": 0, + "content": "3.3 Supervised Audio-Visual Language Alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.784, + 0.801 + ], + "angle": 0, + "content": "Given the enhanced class embeddings \\(\\pmb{w}_1, \\dots, \\pmb{w}_{|\\mathcal{T}|}\\) (with improved class separability), we empirically show that a contrastive objective is sufficient to learn state-of-the-art models for generalized zero-shot learning of audio-visual events. Specifically, let \\(\\pmb{x}_i^{av} = f_\\theta^{av}(\\pmb{v}_i, \\pmb{a}_i)\\) be the multi-modal representation of an audio-visual sample \\((v_i, a_i)\\) belonging to class \\(y_i\\). Then, we align representations \\(\\pmb{x}_i^{av}\\) with the corresponding class embeddings \\(\\pmb{w}_{y_i}\\) by minimizing the contrastive loss" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.81, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A V L A} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {x} _ {i} ^ {a v} , \\boldsymbol {w} _ {y _ {i}}\\right)\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {x} _ {i} ^ {a v}, \\boldsymbol {w} _ {y _ {k}}\\right)\\right)} \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "S. Mo and P. Morgado" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(\\cdot, \\cdot)\\) is a similarity function used to contrast the class embedding \\(\\mathbf{w}_{y_i}\\) against others in the current batch of data \\(\\mathbf{w}_{y_k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Similarity. While contrastive objectives typically rely on cosine similarity metrics for \\(\\sin (\\cdot ,\\cdot):\\mathcal{X}\\times \\mathcal{T}\\rightarrow \\Re\\), we found that non-linear similarity functions can better capture the relationship between audio-visual representations and class embeddings. Specifically, we draw inspiration from transformer attention modules and use a multi-head cross-attention layer. In this layer, class embeddings \\(\\pmb{w}_c\\) are used as queries and audio-visual representations \\(\\pmb{x}_i^{av}\\) as key-value pairs. After cross-attention, class similarities are computed through a linear projection \\(\\pmb{p}\\). Optimizing Eq. 5 pushes the learned embeddings closer to the text embeddings of the same category. During inference, we use the cross-attention transformer layer to output \\(C\\) similarity scores corresponding to each category, including unseen ones. We then classify the audio-visual samples according to the class with highest similarity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.377, + 0.396 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.454 + ], + "angle": 0, + "content": "In this section, we evaluate the proposed EZ-AVGZL on three benchmarks for audio-visual generalized zero-shot learning, including VGGSound-GZSL [28], UCF-GZSL [28], and ActivityNet-GZSL [28]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.471, + 0.429, + 0.487 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Datasets. VGGSound-GZSL [28] contains 93,752 videos across 276 classes, selected from VGGSound [7], a large audio-visual dataset of 309 classes over 200k videos. We follow the prior work [27, 28] and use the same split for train/val(U)/test(U) over 138/69/69 classes. UCF-GZSL [28] includes 6,816 videos from 51 classes, chosen from UCF101 [51], the original video action recognition dataset, which consists of over 13k videos in 101 classes. We use the same train/val(U)/test(U) split over 30/12/9 classes in previous work [27, 28]. ActivityNet-GZSL [28] contains 20k videos in 200 classes of varying duration for action recognition [13]. We use the same train/val(U)/test(U) split over 99/51/50 classes as in the prior work [27, 28]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Evaluation Metrics. In line with prior work [27, 28], we use the mean class accuracy to evaluate all models. For ZSL performance, we evaluate only on the subset of test samples from unseen classes. For the GZSL performance, we evaluate on the full test set, which includes both seen and unseen classes. We report results for the subsets of seen (S) and unseen (U) classes, as well as their harmonic mean." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Implementation. We follow previous work [27] and use 8 heads with a dimension of 64 per head for all attention blocks in the cross-attention transformer. The model is trained for 50 epochs with a batch size of 64, using the Adam optimizer [20] with a learning rate of 1e-4, running average coefficients \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), and a weight decay of 1e-5. Following the prior work [27, 28], we use the word2vec [29] embeddings for text, and VGGish [15] embeddings for audio, and C3D [56] features for video frames." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.337, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.145, + 0.729, + 0.16 + ], + "angle": 0, + "content": "Table 1: Quantitative results of VGGSound-GZSL benchmark." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.173, + 0.675, + 0.316 + ], + "angle": 0, + "content": "
MethodSeenUnseenHarmonicMeanZSL
AttentionFusion [9]14.133.004.953.37
Perceiver [19]13.253.034.933.44
CJME [48]10.862.223.683.72
AVGZSLNet [26]15.023.195.264.81
AVCA [28]12.636.198.316.91
TCaF [27]12.636.728.777.41
Hyper-alignment [16]12.506.448.507.25
Hyper-single [16]12.565.037.185.47
Hyper-multiple [16]15.626.008.677.31
EZ-AVGZL (ours)17.268.6811.558.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.294, + 0.332, + 0.706, + 0.347 + ], + "angle": 0, + "content": "Table 2: Quantitative results of UCF-GZSL benchmark." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.36, + 0.675, + 0.501 + ], + "angle": 0, + "content": "
MethodSeenUnseenHarmonic MeanZSL
AttentionFusion [9]39.3418.2924.9720.21
Perceiver [19]46.8526.8234.1128.12
CJME [48]33.8924.8228.6529.01
AVGZSLNet [26]74.7924.1536.5131.51
AVCA [28]63.1530.7241.3437.72
TCaF [27]67.1440.8350.7844.64
Hyper-alignment [16]57.1333.8642.5239.80
Hyper-single [16]63.4734.8544.9939.86
Hyper-multiple [16]74.2635.7948.3052.11
EZ-AVGZL (ours)78.3246.3558.2448.27
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.532, + 0.484, + 0.547 + ], + "angle": 0, + "content": "4.2 Comparison to Prior Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.787, + 0.643 + ], + "angle": 0, + "content": "In this work, we propose a novel and effective framework for audio-visual generalized zero-shot learning. In order to demonstrate the effectiveness of the proposed EZ-AVGZL, we comprehensively compare it to previous audio-visual baselines [9, 19] adapted to our problem and current state-of-the-art audio-visual generalized zero-shot learning approaches [16, 26-28, 48]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For the VGGSound-GZSL dataset, we report the quantitative comparison results in Table 1. As shown, we achieve the best results in all metrics for both seen and unseen classes compared to previous audio-visual generalized zero-shot learning approaches. In particular, the proposed EZ-AVGZL outperforms Hyper-multiple [16], the current state-of-the-art audio-visual generalized zero-shot learning baseline, by 1.64@Seen, 2.68@Unseen, 2.88@Harmonic Mean, and 1.61@ZSL. Moreover, we also perform better than TCaF [27], a strong audio-visual generalized zero-shot learning baseline, by 4.63@Seen, 1.96@Unseen, 2.78@Harmonic Mean, and 1.51@ZSL. Furthermore, we achieve significant performance gains compared to AVCA [28], the first baseline using cross-modal attention and textual label embeddings. This indicates the importance of optimized text transformation and non-linear similarity functions in learning better transferrable knowledge for generalized zero-shot learning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "S. Mo and P. Morgado" + }, + { + "type": "table_caption", + "bbox": [ + 0.269, + 0.145, + 0.731, + 0.16 + ], + "angle": 0, + "content": "Table 3: Quantitative results of ActivityNet-GZSL benchmark." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.173, + 0.675, + 0.314 + ], + "angle": 0, + "content": "
MethodSeenUnseenHarmonic MeanZSL
AttentionFusion [9]11.153.375.184.88
Perceiver [19]18.254.276.924.47
CJME [48]10.755.557.326.29
AVGZSLNet [26]13.705.968.306.39
AVCA [28]16.777.049.927.58
TCaF [27]30.127.6512.207.96
Hyper-alignment [16]29.778.7713.559.13
Hyper-single [16]24.6110.1014.3210.37
Hyper-multiple [16]36.989.6015.2510.39
EZ-AVGZL (ours)38.5212.8519.2713.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.328, + 0.465, + 0.386 + ], + "angle": 0, + "content": "Table 4: Ablation studies on Class Embedding Optimization and Audio-Visual-Language Alignment." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.4, + 0.465, + 0.455 + ], + "angle": 0, + "content": "
CEPAVLASeenUnseenHarmonic MeanZSL
XX12.636.728.777.41
✓X15.357.9210.458.36
X✓14.677.239.697.91
✓✓17.268.6811.558.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.47, + 0.329, + 0.787, + 0.361 + ], + "angle": 0, + "content": "Table 5: Ablation studies on optimization objectives in Class Embedding Optimization." + }, + { + "type": "table", + "bbox": [ + 0.475, + 0.373, + 0.787, + 0.451 + ], + "angle": 0, + "content": "
LsemLsepSeenUnseenHarmonic MeanZSL
XX14.677.239.327.91
LproxsemX15.757.8310.468.26
X✓16.138.1710.858.45
Lproxsem✓17.268.6811.558.92
Lranksem✓18.0510.2813.1010.36
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.789, + 0.651 + ], + "angle": 0, + "content": "Additionally, we observe significant gains in the UCF-GZSL and ActivityNet-GZSL benchmarks, as shown in Table 2 and Table 3. Compared to TCaF [27], a strong baseline using three manually-designed objectives, we achieve result gains of 11.18@Seen, 5.52@Unseen, 7.46@Harmonic Mean, and 3.63@ZSL on UCF-GZSL dataset. When evaluated on the challenging ActivityNet-GZSL benchmark, our method still outperforms TCaF [27] by 8.40@Seen, 5.20@Unseen, 7.07@Harmonic Mean, and 5.05@ZSL. We also achieve better results than AVCA [28], a cross-modal baseline based on cross-modal attention between audio and visual embeddings. These results demonstrate the effectiveness of our approach in learning optimized textual label embeddings and non-linear similarity functions for generalized audio-visual zero-shot recognition." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.675, + 0.45, + 0.691 + ], + "angle": 0, + "content": "4.3 Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.779 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to demonstrate the benefit of introducing the Class Embedding Optimization (CEO) and Audio-Visual Language Alignment (AVLA) modules. We also carry out extensive experiments to investigate optimization objectives in class embedding optimization, non-linear similarity functions, and generalization across diverse cross-modal settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.84 + ], + "angle": 0, + "content": "Class Embedding Optimization (CEO) & Audio-Visual Language Alignment (AVLA). In order to demonstrate the effectiveness of CEO and AVLA, we conduct an ablation study and report the quantitative results on VGGSound-GZSL dataset in Table 4." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.337, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "As observed, adding CEO to the vanilla baseline improves the results by 2.72@Seen, 1.20@Unseen, 1.68@Harmonic Mean, and 0.95@ZSL, which validates the benefit of optimizing class embeddings so as to obtain separated embeddings that preserve the semantic relationships between classes. On the other hand, introducing AVLA in the baseline increases the zero-shot performance across all metrics. More importantly, incorporating both CEP and AVLA into the baseline significantly raises the performance by 4.63@Seen, 1.96@Unseen, 2.78@Harmonic Mean, and 1.51@ZSL. These improvements validate the importance of text transformation and supervised text audio-visual contrastive in extracting separated textual label embeddings for AVGZSL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.593 + ], + "angle": 0, + "content": "Objectives in Class Embedding Optimization. Optimization objectives in class embedding transformation, including global semantic preservation and class separability, are essential for learning optimized textual class embeddings for zero-shot learning. To explore the effects of each objective more comprehensively, we ablate each text optimization objective and report the quantitative results in Table 5. We observe that adding the semantic preservation loss to the vanilla baseline improves the performance by 1.08@Seen, 0.60@Unseen, 1.14@Harmonic Mean, and 0.35@ZSL. On the other hand, introducing the class separability loss in the baseline enhances the zero-shot performance across all metrics. Furthermore, combining the semantic preservation and class separability objectives for the text transformer significantly increases the performance by 2.59@Seen, 3.05@Unseen, 3.78@Harmonic Mean, and 2.45@ZSL. This result validates the importance of class separability loss in balancing the preservation of the semantics available in pre-trained representations for zero-shot learning. Replacing the Euclidean-based semantic preservation loss \\(\\mathcal{L}_{sem}^{prox}\\) with a ranking based loss \\(\\mathcal{L}_{sem}^{rank}\\) further achieves performance gains of 0.79@Seen, 1.60@Unseen, 1.55@Harmonic Mean, and 1.44@ZSL. These results demonstrate the effectiveness of both semantic preservation and class separability objectives in class embedding optimization to generate meaningful text representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Non-linear Similarity Functions. The non-linear similarity operator used in supervised audio-visual language alignment impacts the extracted cross-modal representations. To explore these effects more comprehensively, we varied the functions from \\(\\{\\mathrm{Cosine - similarity}\\), Linear, MLP, Cross-Attention\\}. For Cosine-Similarity, we directly compute the dot product between audio-visual embeddings \\(\\mathbf{x}_i^{av}\\) and text embeddings \\(\\mathbf{t}_i\\). For Linear, we apply a linear layer to the concatenation of both embeddings to predict the similarity score. For MLP, we replace the linear layer with a multi-layer perceptron to generate the score. For Cross-Attention, we use text embeddings \\(\\mathbf{t}_i\\) as the queries and audio-visual representations \\(\\mathbf{x}_i^{av}\\) as key-value pairs in a transformer attention module to output the score. The comparison results of zero-shot performance on VGGSound-GZSL benchmark are reported in Table 6. When using Cosine-Similarity as the function, we achieve the worst results across all metrics. Replacing the cosine-similarity operator with both Linear increases the seen results by \\(2.55@Seen\\), but decreases the performance by \\(0.07@Unseen\\) and \\(0.12@ZSL\\). This might be because adding complexity to the function causes the model to overfit the seen classes in the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "S. Mo and P. Morgado" + }, + { + "type": "table_caption", + "bbox": [ + 0.254, + 0.145, + 0.748, + 0.161 + ], + "angle": 0, + "content": "Table 6: Exploration studies on non-linear similarity functions \\( \\phi \\left( \\cdot \\right) \\) ." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.173, + 0.632, + 0.231 + ], + "angle": 0, + "content": "
φ(·)SeenUnseenHarmonic MeanZSL
Cosine-Similarity13.978.5910.648.67
Linear16.528.5211.248.55
MLP17.038.4611.308.49
Cross-Attention17.268.6811.558.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.276 + ], + "angle": 0, + "content": "Table 7: Exploration studies on diverse cross-modal settings. CEP denotes Class Embedding Optimization." + }, + { + "type": "table", + "bbox": [ + 0.348, + 0.29, + 0.661, + 0.402 + ], + "angle": 0, + "content": "
Text EncoderAudio EncoderVisual EncoderCEPSeenUnseenHarmonic MeanZSL
word2vecVGGishC3D×12.636.728.777.41
✓17.268.6811.558.92
LLaMaAudioMAEDINO×18.3510.6313.4612.26
✓23.9513.3217.1213.73
CLAPCLAPX-CLIP×27.7914.0118.6317.39
✓33.2820.1525.1021.06
ImageBindImageBindImageBind×28.3214.6819.3419.87
✓35.5621.6326.9022.35
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.537 + ], + "angle": 0, + "content": "training data, , leading to poorer generalization to unseen categories. Additionally, using MLP as the function operator significantly raises the performance by 3.06@Seen, but slightly deteriorates the results on unseen classes. Our method achieves the best results across all metrics by using the cross-attention operator to predict the similarity score. These improved results demonstrate the effectiveness of the cross-attention operator in balancing the trade-off between linear and non-linear similarity functions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Generalization to Diverse Cross-modal Settings. To validate the generalizability of the proposed method across a flexible number of feature encoders, we experiment with different settings: unimodal, paired (video-text, audio-text), and unified embedding space (image-audio-text). In the unimodal setting, we use DINO [3] for the visual encoder, AudioMAE [18] for the audio encoder, and LLaMa [55] for the text encoder. In the paired setting, we use X-CLIP [25] pre-trained with video-text objectives for the visual encoder, and CLAP [59] pre-trained with audio-text objectives for the audio and text encoder, respectively. In the unified embedding setting, we apply ImageBind [12] pre-trained with image-audio-text alignment in the same embedding space as the encoder for all three modalities. The quantitative results on the VGGSound-GZSL benchmark are compared in Table 7. As observed, the unimodal setting, which uses transformers with strong unimodal features, significantly outperforms the vanilla baseline that uses convolution networks, where EZ-AVGZL achieves performance gains of 6.69@Seen, 4.64@Unseen, 5.77@Harmonic Mean, and 4.81@ZSL using class embedding optimization. In the paired setting, the proposed method further increases the results by 9.33@Seen, 6.93@Unseen, 7.98@Harmonic Mean, and 7.331@ZSL, benefiting from the text-audio correspondence and text-video alignment pre-trained on large-scale data. In the unified embedding setting using ImageBind [12] embeddings, we achieve the best results across all metrics" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.257, + 0.145, + 0.744, + 0.161 + ], + "angle": 0, + "content": "Table 8: Comparison of class embeddings w/o and w optimization." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.172, + 0.73, + 0.268 + ], + "angle": 0, + "content": "
base classNN (w/o opt)D (w/o opt)NN (w opt)D (w opt)
alligators, crocodiles hissingchurch bell ringing0.455playing marimba, xylophone0.498
barn swallow callingblack capped chickadee calling0.307black capped chickadee calling0.465
basketball bounceplaying volleyball0.324bouncing on trampoline0.394
bee, wasp, etc. buzzingmotorboat, speedboat acceleration0.322fly, housefly buzzing0.333
bird squawkingpleasant crowing0.400duck quacking0.496
black capped chickadee callingbarn swallow calling0.307magpie calling0.436
bouncing on trampolinebasketball bounce0.416basketball bounce0.494
bowling impactplaying table tennis0.410striking bowling0.467
bull bellowingeagle screaming0.449dog growling0.532
canary callingwhale calling0.403black capped chickadee calling0.513
" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.28, + 0.756, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.461, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 3: Confusion matrices of zero-shot predictions on VGGSound-GZSL using model with and without class embedding optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.613 + ], + "angle": 0, + "content": "compared to other settings. This might be because text embeddings with unified training on the other two modalities strengthen the alignment for audio-visual zero-shot learning. Furthermore, when class embedding optimization is added to all settings, we improve the baseline based on the initial text embeddings, further demonstrating the effectiveness of the proposed class embedding optimization in generating text representations for AVZSL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Qualitative Comparisons with and without Optimization. To provide insights into why the proposed loss terms improve the separability of class embeddings while keeping semantics, we conduct qualitative studies. These include comparing the visualization of the class embedding space with and without optimization. The comparisons of class embeddings with and without optimization are shown in Table 8. We report the base class and identify 1) the nearest neighbor (NN) class and 2) the distance (D) to the nearest neighbor. We observe larger distances and similar NN classes. The confusion matrices for class-wise performances of the model are reported in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Impact of Hyper-parameters \\(\\alpha\\) and Margin \\(m\\). To investigate the impact of \\(\\alpha\\) and \\(m\\) on the performance thoroughly, we set \\(\\alpha\\) from \\(\\{0.1, 0.5, 0.9\\}\\) and \\(m\\) from \\(\\{0, 0.5, 1\\}\\). The quantitative comparisons on VGGSound-GZSL are reported in Table 9. As can be seen, with the increase of \\(\\alpha\\) from 0.1 to 0.5, the proposed approach achieves better results regarding all metrics. However, when \\(\\alpha\\) is increased to 0.9, the performance decreases a lot, which might be caused by not" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "S. Mo and P. Morgado" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.447, + 0.174 + ], + "angle": 0, + "content": "Table 9: Study of \\( \\alpha \\) and margin \\( m \\) on VGGSound-GZSL." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.187, + 0.449, + 0.253 + ], + "angle": 0, + "content": "
αmSeenUnseenHarmonic MeanZSL
0.518.0510.2813.1010.36
0.1116.358.3911.098.58
0.916.268.7311.368.92
0.5017.829.7112.579.67
0.517.679.8512.659.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.47, + 0.146, + 0.785, + 0.176 + ], + "angle": 0, + "content": "Table 10: Study of distance functions on VGGSound-GZSL." + }, + { + "type": "table", + "bbox": [ + 0.474, + 0.189, + 0.788, + 0.251 + ], + "angle": 0, + "content": "
DistanceSeenUnseenHarmonic MeanZSL
cosine18.0510.2813.1010.36
l217.589.3212.189.27
l117.358.7811.668.85
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.786, + 0.307 + ], + "angle": 0, + "content": "much semantics preservation during training. Meanwhile, with the decrease of \\( m \\), we observe a clearly dropping trend in our metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.459 + ], + "angle": 0, + "content": "Ablating Distance Functions for Eq. 1-2. Distance functions in Eq. 1-2 is crucial for us to maximize separation while preserving semantics. We experimented with different distance functions, \\(\\ell_2\\) and \\(\\ell_1\\) distances for Eq. 1-2. Note that while, Eq. 1-2 show \\(\\ell_2\\) distances in the paper, the embeddings are normalized. The quantitative results are shown in Table 10. We can observe that using the cosine distance function performed the best regarding all metrics. However, when optimizing \\(\\ell_1\\) distance functions, we achieve the worst results in terms of all metrics. These ablation results highlight the importance of designing appropriate distance functions in maximizing separation while preserving the semantics of text class embeddings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.479, + 0.36, + 0.495 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.719 + ], + "angle": 0, + "content": "In this work, we present a novel and effective framework, EZ-AVGZL, that aligns audio-visual representations with optimized text embeddings to achieve audio-visual generalized zero-shot learning. We leverage differential optimization to learn better-separated textual representations with maximal separability and semantic preservation. Additionally, we introduce supervised audio-visual language alignment to learn the correspondence between audio-visual features and textual embeddings, which captures the cross-modal dynamics with class labels as prompts. Experimental results on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL datasets demonstrate our method's superiority against previous baselines. Extensive ablation studies also validate the importance of class embedding optimization and supervised audio-visual language alignment for audio-visual generalized zero-shot learning. We also conduct comprehensive experiments to demonstrate the effectiveness of non-linear similarity functions and generalization to diverse cross-modal settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Broader Impact. This work makes substantial progress in audio-visual generalized zero-shot learning. Our method is novel in its simplicity, demonstrating that a simple supervised contrastive objective, without \"bells and whistles\", outperforms complex alignment strategies. Without our work, progress in this area would likely be driven by increasingly complex methods, which are difficult to implement, reproduce and compare. Beyond the strong results and simplicity of our method, the class embedding optimization procedure is also technically novel to maximize semantic separation while keeping semantics preserved." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Arandjelovic, R., Zisserman, A.: Look, listen and learn. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 609-617 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.25 + ], + "angle": 0, + "content": "2. Aytar, Y., Vondrick, C., Torralba, A.: Soundnet: Learning sound representations from unlabeled video. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.787, + 0.347 + ], + "angle": 0, + "content": "4. Chen, C., Jain, U., Schissler, C., Gari, S.V.A., Al-Halah, Z., Ithapu, V.K., Robinson, P., Grauman, K.: Soundspaces: Audio-visual navigation in 3d environments. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 17-36 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.391 + ], + "angle": 0, + "content": "5. Chen, C., Majumder, S., Ziad, A.H., Gao, R., Kumar Ramakrishnan, S., Grauman, K.: Learning to set waypoints for audio-visual navigation. In: Proceedings of International Conference on Learning Representations (ICLR) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.447 + ], + "angle": 0, + "content": "6. Chen, C., Schissler, C., Garg, S., Kobernik, P., Clegg, A., Calamia, P., Batra, D., Robinson, P.W., Grauman, K.: Soundspaces 2.0: A simulation platform for visual-acoustic learning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.489 + ], + "angle": 0, + "content": "7. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 721-725. IEEE (2020) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.531 + ], + "angle": 0, + "content": "8. Ephrat, A., Mosseri, I., Lang, O., Dekel, T., Wilson, K., Hassidim, A., Freeman, W.T., Rubinstein, M.: Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.573 + ], + "angle": 0, + "content": "9. Fayek, H.M., Kumar, A.: Large scale audiovisual learning of sounds with weakly labeled data. In: Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (2021) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.615 + ], + "angle": 0, + "content": "0. Gan, C., Huang, D., Zhao, H., Tenenbaum, J.B., Torralba, A.: Music gesture for visual sound separation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10478-10487 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "1. Gao, R., Grauman, K.: 2.5d visual sound. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 324-333 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.713 + ], + "angle": 0, + "content": "2. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.756 + ], + "angle": 0, + "content": "3. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 961-970 (2015) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "4. Hershey, J., Casey, M.: Audio-visual sound separation via hidden markov models Advances in Neural Information Processing Systems 14 (2001) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Hershey, S., Chaudhuri, S., Ellis, D.P.W., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., Slaney, M., Weiss, R.J., Wilson, K.: Cnn architectures for large-scale audio classification. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2017) 8" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "S. Mo and P. Morgado" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Hong, J., Hayden, Z., Han, J., Fang, P., Harandi, M., Petersson, L.: Hyperbolic audio-visual zero-shot learning. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023) 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "17. Hu, D., Nie, F., Li, X.: Deep multimodal clustering for unsupervised audiovisual learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9248-9257 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "18. Huang, P.Y., Xu, H., Li, J., Baevski, A., Auli, M., Galuba, W., Metze, F., Feichtenhofer, C.: Masked autoencoders that listen. In: Proceedings of Advances In Neural Information Processing Systems (NeurIPS) (2022) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "19. Jaegle, A., Gimeno, F., Brock, A., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver: General perception with iterative attention. In: Proceedings of the International Conference on Machine Learning (ICML) (2021) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "21. Korbar, B., Tran, D., Torresani, L.: Cooperative learning of audio and video models from self-supervised synchronization. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.787, + 0.425 + ], + "angle": 0, + "content": "22. Lin, Y.B., Li, Y.J., Wang, Y.C.F.: Dual-modality seq2seq network for audio-visual event localization. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2002–2006 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.48 + ], + "angle": 0, + "content": "23. Lin, Y.B., Tseng, H.Y., Lee, H.Y., Lin, Y.Y., Yang, M.H.: Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "24. Lin, Y.B., Wang, Y.C.F.: Audiovisual transformer with instance attention for audiovisual event localization. In: Proceedings of the Asian Conference on Computer Vision (ACCV) (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "25. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP:: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285 (2022) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.619 + ], + "angle": 0, + "content": "26. Mazumder, P., Sing, P., Kumar Parida, K., Namboodiri, V.P.: Avgzslnet: Audiovisual generalized zero-shot learning by reconstructing label features from multimodal embeddings. In: Proceedings of 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 3089-3098 (2021) 3, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "27. Mercea, O.B., Hummel, T., Koepke, A.S., Akata, Z.: Temporal and cross-modal attention for audio-visual zero-shot learning. In: Proceedings of European Conference on Computer Vision (ECCV) (2022) 2, 4, 5, 6, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.715 + ], + "angle": 0, + "content": "28. Mercea, O.B., Riesch, L., Koepke, A.S., Akata, Z.: Audio-visual generalised zero-shot learning with cross-modal attention and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10553–10563 (2022) 2, 4, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "29. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. In: Proceedings of International Conference on Learning Representations (ICLR) (2013) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.797 + ], + "angle": 0, + "content": "30. Mo, S., Morgado, P.: Benchmarking weakly-supervised audio-visual sound localization. In: European Conference on Computer Vision (ECCV) Workshop (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.798, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Mo, S., Morgado, P.: A unified audio-visual learning framework for localization, separation, and recognition. In: Proceedings of the International Conference on Machine Learning (ICML) (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "33. Mo, S., Morgado, P.: Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling. arXiv preprint arXiv:2312.01017 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.787, + 0.257 + ], + "angle": 0, + "content": "34. Mo, S., Pian, W., Tian, Y.: Class-incremental grouping network for continual audio-visual learning. arXiv preprint arXiv:2309.05281 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.257, + 0.787, + 0.284 + ], + "angle": 0, + "content": "35. Mo, S., Shi, J., Tian, Y.: DiffAVA: Personalized text-to-audio generation with visual alignment. arXiv preprint arXiv:2305.12903 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.787, + 0.311 + ], + "angle": 0, + "content": "36. Mo, S., Shi, J., Tian, Y.: Text-to-audio generation synchronized with videos. arXiv preprint arXiv:2403.07938 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "37. Mo, S., Tian, Y.: Multi-modal grouping network for weakly-supervised audiovisual video parsing. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.787, + 0.392 + ], + "angle": 0, + "content": "38. Mo, S., Tian, Y.: Semantic-aware multi-modal grouping for weakly-supervised audio-visual video parsing. In: European Conference on Computer Vision (ECCV) Workshop (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.392, + 0.787, + 0.419 + ], + "angle": 0, + "content": "39. Mo, S., Tian, Y.: Audio-visual grouping network for sound localization from mixtures. arXiv preprint arXiv:2303.17056 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.419, + 0.787, + 0.446 + ], + "angle": 0, + "content": "40. Mo, S., Tian, Y.: AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.446, + 0.787, + 0.473 + ], + "angle": 0, + "content": "41. Mo, S., Tian, Y.: Semantic grouping network for audio source separation. arXiv preprint arXiv:2407.03736 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "42. Morgado, P., Li, Y., Vasconcelos, N.: Learning representations from audio-visual spatial alignment. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS). pp. 4733-4744 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.787, + 0.554 + ], + "angle": 0, + "content": "43. Morgado, P., Li, Y., Costa Pereira, J., Saberian, M., Vasconcelos, N.: Deep hashing with hash-consistent large margin proxy embeddings. International Journal of Computer Vision (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.554, + 0.787, + 0.595 + ], + "angle": 0, + "content": "44. Morgado, P., Misra, I., Vasconcelos, N.: Robust audio-visual instance discrimination. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12934–12945 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.595, + 0.787, + 0.636 + ], + "angle": 0, + "content": "45. Morgado, P., Vasconcelos, N., Langlois, T., Wang, O.: Self-supervised generation of spatial audio for 360 video. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.636, + 0.787, + 0.677 + ], + "angle": 0, + "content": "46. Morgado, P., Vasconcelos, N., Misra, I.: Audio-visual instance discrimination with cross-modal agreement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12475–12486 (June 2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.677, + 0.787, + 0.717 + ], + "angle": 0, + "content": "47. Owens, A., Wu, J., McDermott, J.H., Freeman, W.T., Torralba, A.: Ambient sound provides supervision for visual learning. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 801-816 (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.717, + 0.787, + 0.772 + ], + "angle": 0, + "content": "48. Parida, K.K., Matiyali, N., Guha, T., Sharma, G.: Coordinated joint multimodal embeddings for generalized audio-visual zero-shot classification and retrieval of videos. In: Proceedings of 2020 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 3240–3249 (2020) 3, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.8 + ], + "angle": 0, + "content": "49. Pian, W., Mo, S., Guo, Y., Tian, Y.: Audio-visual class-incremental learning. arXiv preprint arXiv:2308.11073 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "50. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4358-4366 (2018) 3" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.424, + 0.128 + ], + "angle": 0, + "content": "S. Mo and P. Morgado" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "51. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "52. Tammes, P.M.L.: On the origin of number and arrangement of the places of exit on the surface of pollen-grains. Recueil des travaux botaniques neerlandais 27(1), 1-84 (1930) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "53. Tian, Y., Li, D., Xu, C.: Unified multisensory perception: Weakly-supervised audiovisual video parsing. In: Proceedings of European Conference on Computer Vision (ECCV). p. 436-454 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "54. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of European Conference on Computer Vision (ECCV) (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.466 + ], + "angle": 0, + "content": "55. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "56. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2015) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "57. Wu, Y., Yang, Y.: Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1326–1335 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "58. Wu, Y., Zhu, L., Yan, Y., Yang, Y.: Dual attention matching for audio-visual event localization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 6291–6299 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.646 + ], + "angle": 0, + "content": "59. Wu, Y., Chen, K., Zhang, T., Hui, Y., Berg-Kirkpatrick, T., Dubnov, S.: Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In: IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "60. Zhang, L., Mo, S., Zhang, Y., Morgado, P.: Audio-synchronized visual animation arXiv preprint arXiv:2403.05659 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "61. Zhao, H., Gan, C., Ma, W.C., Torralba, A.: The sound of motions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1735-1744 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "62. Zhao, H., Gan, C., Rouditchenko, A., Vondrick, C., McDermott, J., Torralba, A.: The sound of pixels. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 570-586 (2018) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.757 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_origin.pdf b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c07039704555ffa2ece7183033442889b13ef025 --- /dev/null +++ b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/66ab4f1e-8b77-415e-a773-fa90d6ba9ef7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e37c94506d3d12fc5efd6fd9c94631bf418a804a21ef58620b8df5b84f9cd8a2 +size 1974675 diff --git a/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/full.md b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/full.md new file mode 100644 index 0000000000000000000000000000000000000000..212b66c82deabcb2c658ff8f3914df516a008b27 --- /dev/null +++ b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/full.md @@ -0,0 +1,274 @@ +# Audio-visual Generalized Zero-shot Learning the Easy Way + +Shentong Mo1 and Pedro Morgado2 + +$^{1}$ Carnegie Mellon University $^{2}$ University of Wisconsin-Madison + +Abstract. Audio-visual generalized zero-shot learning is a rapidly advancing domain that seeks to understand the intricate relations between audio and visual cues within videos. The overarching goal is to leverage insights from seen classes to identify instances from previously unseen ones. Prior approaches primarily utilized synchronized auto-encoders to reconstruct audio-visual attributes, which were informed by cross-attention transformers and projected text embeddings. However, these methods fell short of effectively capturing the intricate relationship between cross-modal features and class-label embeddings inherent in pretrained language-aligned embeddings. To circumvent these bottlenecks, we introduce a simple yet effective framework for Easy Audio-Visual Generalized Zero-shot Learning, named EZ-AVGZL, that aligns audio-visual embeddings with transformed text representations. It utilizes a single supervised text audio-visual contrastive loss to learn an alignment between audio-visual and textual modalities, moving away from the conventional approach of reconstructing cross-modal features and text embeddings. Our key insight is that while class name embeddings are well aligned with language-based audio-visual features, they don't provide sufficient class separation to be useful for zero-shot learning. To address this, our method leverages differential optimization to transform class embeddings into a more discriminative space while preserving the semantic structure of language representations. We conduct extensive experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks. Our results demonstrate that our EZ-AVGZL achieves state-of-the-art performance in audio-visual generalized zero-shot learning. + +Keywords: Audio-Visual Learning $\cdot$ Zero-shot Learning $\cdot$ Audio-Visual Generalized Zero-shot Learning + +# 1 Introduction + +In the field of machine learning, the ability of models to recognize patterns and make predictions is crucial. Zero-shot learning exemplifies this capability, which leverages known information to understand the unknowns. To further explore its potential and applications, the principle of zero-shot learning has been applied in various contexts. One such context is Audio-visual generalized + +![](images/dbd5321a80463c291c70e52f3bbadec3715ee5b46d42d863b2891df9983574ee.jpg) +Fig.1: Comparison of our EZ-AVGZL with state-of-the-art methods on VGGSound-GZSL, UCF-GZSL, ActivityNet-GZSL benchmarks in terms of Harmonic Mean (higher is better) for seen and unseen classes. Our method significantly outperforms previous baselines in terms of all datasets. + +zero-shot learning (AVGZSL), an emerging topic of study that predicts unseen audio-visual categories by leveraging knowledge learned from seen classes. + +Within the scope of AVGZSL, previous approaches [27, 28] have focused on designing autodecoders for reconstructing audio-visual attributes. For example, TCaF [27] adopted a temporal contrastive feature alignment strategy. This strategy aligns audio-visual cues with word2vec embeddings of class prompts by defining both cross-entropy and regression losses to reduce the distance between the output embedding for a sample and the corresponding projected word2vec embedding. These cross-modal and textual features were derived from cross-attention transformers and the projected text embeddings. However, the complexity of the network for AVGZSL, which is derived from the synergy between cross-attention transformers and the corresponding projected word2vec embeddings, makes it challenging to understand the deeper interplay between cross-modal features and class label embeddings. + +In the context of audio-visual zero-shot learning, the main challenge lies in the absence of unseen classes during training. Without the supervision, the prediction of cross-modal class labels performs poorly on these audio-visual samples. To address the challenge, we propose a novel framework for Easy Audio-Visual Generalized Zero-shot Learning, referred to as EZ-AVGZL. This framework aligns audio-visual embeddings with transformed text representations in a simple yet effective manner. EZ-AVGZL learn well-separated text semantics across seen and unseen class names to achieve maximal separation while retaining the semantic relations between each other. During training, we leverage a simple supervised contrastive loss to learn the cross-modal alignment between audio-visual features + +and transformed text embeddings. EZ-AVGZL enhances class separability while preserving semantics through the use of differential optimization, thereby generating more distinct text embeddings. Furthermore, it employs a single supervised text to audio-visual contrastive loss, establishing an alignment between the audio-visual and textual modalities. This approach effectively eliminates the need for reconstructing cross-modal features and text embeddings. + +Empirical experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks demonstrate the state-of-the-art performance against previous audio-visual generalized zero-shot learning baselines, as illustrated in Fig. 1. Extensive ablation studies further confirm the importance of class embedding optimization and supervised audio-visual language contrastive objectives in learning language alignment with audio-visual representations. Additionally, we demonstrate the robustness of our approach across various zero-shot learning tasks and observe the benefit of our approach when applied to a variety of foundation models pre-trained from unimodal and cross-modal. + +Our main contributions can be summarized as follows: + +- We introduce the EZ-AVGZL framework, a paradigm shift in audio-visual generalized zero-shot learning. This framework simplifies the process with a supervised text-audio-visual contrastive loss, enhancing effectiveness. +- We propose a novel text transformation technique designed to construct distinct text embeddings, paving the way for better alignment and, by extension, superior recognition capabilities. +- We demonstrate that our method achieves state-of-the-art performance on extensive benchmarks in audio-visual generalized zero-shot learning. + +# 2 Related Work + +Audio-Visual Learning. Audio-visual learning problems have been addressed in many previous works [1,2,8,10,14,17,21,32,35,36,38,41,42,44,46,47,50,60-62] to learn the audio-visual correlation between two distinct modalities from videos. Such cross-modal alignments are beneficial for a variety audio-visual tasks, including audio-event localization [22, 24, 30, 31, 33, 39, 40, 54, 58], audio-visual spatialization [4, 11, 42, 45], audio-visual continual learning [34, 49], audio-visual navigation [4-6] and audio-visual parsing [23, 37, 53, 57]. In this work, our primary focus is on learning audio-visual representations that align with text embeddings for generalized zero-shot learning, a task that presents more challenges than the aforementioned tasks. + +Zero-shot Learning. Zero-shot learning aims to leverage semantic relationships, typically utilizing attributes or word embeddings associated with class labels to make informed predictions about previously unseen categories. This is often achieved by drawing parallels to the semantic information of known categories. Early audio-visual zero-shot learning methods, such as those proposed by Parida et al. [48] and Mazumder et al. [26], attempted to map video, audio, and text into the same feature space and compute the similarity across each other for comparison. For example, Audio-Visual Generalized Zero-shot Learning Network [26] + +introduced a reconstruction module to predict text representations from audio and visual features. More recently, ImageBind [12] proposed learning a joint embedding across six different modalities, especially on audio, text, and video. This approach demonstrated the ability to recognize or retrieve unseen categories without explicit training. Unlike these methods, our approach does not require large-scale pre-training. Instead, we present a simple yet effective approach based on unimodal or crossmodal weights pre-trained on large-scale data to achieve audio-visual generalized zero-shot learning. + +Audio-Visual Generalized Zero-shot Learning. Audio-visual generalized zero-shot learning extends the zero-shot learning paradigm to scenarios where both seen and unseen classes are present at test time. The seminal work, AVCA [28] proposed an attention mechanism to selectively focus on discriminative audio and visual cues. By harnessing the combined strength of these cues, AVCA not only achieved superior performance on seen classes but also exhibited proficiency in recognizing unseen classes. + +Following up, TCaF [27] introduced a temporal contrastive feature alignment strategy. This strategy aligns audio-visual cues across time and applies two separate decoders to reconstruct the audio-visual embedding and the projected word2vec embedding to be close to the original word2vec embedding of class labels. More recently, Hong et al. [16] incorporated cross-modality alignment between video and audio features in the hyperbolic space. They used multiple adaptive curvatures for hyperbolic projections to achieve curvature-aware geometric learning. In contrast, we develop a novel framework to aggregate the alignment between audio-visual representations and text embeddings with explicit class embedding optimization. To the best of our knowledge, this work is the first to leverage an explicit class embedding optimization mechanism for audio-visual generalized zero-shot learning. Our experiments in Section 4.2 also demonstrate the effectiveness of our method in classifying both seen and unseen categories. + +# 3 Method + +We aim to learn a language-aligned representation for a video clip, which includes both the visual frame sequence and corresponding audio, to enable open-dictionary recognition of audio-visual events. To achieve this, we propose a simple yet effective approach, named Easy Audio-Visual Generalized Zero-shot Learning (EZ-AVGZL). This approach seeks to 1) construct multi-modal representations of the data and 2) align them with language representations using a supervised contrastive learning objective while leveraging strong foundation models. The proposed approach is based on two components: class embedding optimization described in Section 3.2 and supervised alignment between language and audio-visual representations described in Section 3.3. Despite its simplicity, EZ-AVGZL outperforms recent state-of-the-art methods on Audio-Visual Generalized Zero-shot Learning (AVGZSL) on a variety of datasets. + +![](images/a752799a4de23b84bc0cf07e68a6196edc59bcdff539a01485a5edf264911d41.jpg) +Fig. 2: Illustration of the proposed Easy Audio-Visual Generalized Zero-shot Learning (EZ-AVGZL) framework. The initial class embeddings $\pmb{t}_i$ from a frozen text encoder are optimized with maximal separability and preserved semantics to generate the new embeddings $\pmb{w}_i$ . Then the cross-attention transformer $f_{\theta}^{av}(\cdot, \cdot)$ takes visual and audio features $(\pmb{v}_i, \pmb{a}_i)$ from the unimodal encoder to generate the multi-modal representations $\pmb{x}_i^{av}$ . Finally, a non-linear similarity function is applied to align representations $\pmb{x}_i^{av}$ with the corresponding class embeddings $\pmb{w}_{y_i}$ by minimizing the cross-entropy loss between the predicted similarity score and the target score as 1 in the target class entry, while the target is 0 for multi-modal representations $\pmb{x}_j^{av}$ given visual features $\pmb{v}_j$ of videos from other classes. + +# 3.1 Preliminaries: Audio-Visual Generalized Zero-shot Learning (AVGZSL) + +Problem Statement. The objective of AVGZSL is to utilize natural language descriptions of audio-visual events $t_i$ and the semantic relationships to enable a model to recognize events from a set of audio-visual event classes $\mathcal{T} = \mathcal{T}_s \cup \mathcal{T}_u$ . This should be possible regardless of whether the model has seen them $(t_i \in \mathcal{T}_s)$ or not seen them $(t_i \in \mathcal{T}_u)$ during training. + +To accomplish this, a model $f_{\theta}:\mathcal{V}\times \mathcal{A}\to \mathcal{T}$ is trained on the seen dataset $\mathcal{D}_s = \{(v_i,a_i,t_i)\}$ for audio-visual event recognition, where $v_{i}\in \mathcal{V}$ denotes the $i^{th}$ visual clip, typically multiple frames spanning a short time window, $a_{i}\in \mathcal{A}$ denotes the corresponding audio, and $t_i\in \mathcal{T}_s$ denotes the (human-readable) class name of the audio-visual event depicted in video $(v_{i},a_{i})$ . For effective zero-shot learning, a model needs to learn a mapping that can generalize beyond $\mathcal{T}_s$ by inferring from the semantic relationships between seen and unseen classes. In other words, the AVGZSL setting evaluates models based on their classification performance on videos $(v_{i},a_{i})$ belonging to both seen $\mathcal{T}_s$ and unseen classes $\mathcal{T}_u$ . Architecture Following TCaF [27], we tackle AVGZSL using a variety of pretrained models to encode all input signals. We experiment with a wider variety of pre-trained models than in [27], in all cases, we train our system on a series of unimodal pre-extracted feature representations for the visual frames $\pmb {v}_i = f_v(\pmb {v}_i)$ , audio $\pmb {a}_i = f_a(a_i)$ and class names $\pmb {t}_i = f_t(t_i)$ . As our goal is multi-modal GZSL, + +we further encode the unimodal audio and visual representations into a joint multi-modal representation using the fusion model $\pmb{x}_i^{av} = f_{av}(\pmb{v}_i, \pmb{a}_i)$ proposed in [27], which consists of a sequence of $L$ transformer layers with cross-attention between audio and visual modalities. + +The quality of representations obtained from foundation models is crucial for effective zero-shot learning. However, the focus of this work is the language alignment step. While prior work often relies on elaborate alignment strategies, we demonstrate that a simple supervised contrastive objective between text and audio-visual representations can achieve state-of-the-art performance. This is possible as long as the class name embeddings for seen and unseen events are maximally separated while retaining the semantic relations between each other. An overview of the proposed approach is shown in Fig. 2. + +Apart from the quality of representations derived from foundation models, language alignment, which is the focus of this work, is a critical component for effective zero-shot learning. While prior work often relies on complex alignment strategies, we demonstrate that a simple supervised contrastive objective between text and audio-visual representations can achieve state-of-the-art performance, as long as the class name embeddings for seen and unseen events are maximally separated while retaining the semantic relations between each other. An overview of the proposed approach is shown in Fig. 2. + +# 3.2 Class Embedding Optimization + +Language-aligned classifiers can achieve zero-shot generalization as class embeddings $t_1, \ldots, t_{|\mathcal{T}|}$ obtained from large-scale models like LLaMa or CLIP, encode semantic relationships between classes. However, while these class embeddings can effectively encode semantics, they are sub-optimal for classification due to a lack of class separability. To obtain well-separated semantic embeddings, we optimize a new set of class embeddings $\boldsymbol{w}_1^*, \ldots, \boldsymbol{w}_{|\mathcal{T}|}^*$ to satisfy two competing constraints: class separability and semantic preservation. + +Class Separability. Class embeddings that are maximally separated from each other are better suited for recognition of seen classes [43]. Thus, to achieve class embedding separation, we minimize the negative distance of each class embedding to its nearest neighbor + +$$ +\mathcal {L} _ {s e p} = - \sum_ {c = 1} ^ {| \tau |} \min _ {k \neq c} \| \boldsymbol {w} _ {c} - \boldsymbol {w} _ {k} \| _ {2}. \tag {1} +$$ + +However, maximizing separability alone would destroy the semantics of the pre-trained embeddings. In fact, the minimum of Eq. 1 (if norm constrained) would be achieved for a set of embeddings $\pmb{w}_1, \dots, \pmb{w}_{|\mathcal{T}|}$ following a "Tammes problem" configuration [52]. Note that the embeddings are normalized. When the embedding dimension $d$ is larger than the number of classes in the dictionary $\mathcal{T}$ , the Tammes configuration provides a solution where all embeddings are equally spaced from each other, i.e. $\| \pmb{w}_i - \pmb{w}_j \|_2 = \mathrm{Cte} \quad \forall i \neq j$ . For example, the distance between the class "a person playing guitar" and the class "a person + +playing the violin" would be the same as the distance to the class "a dog barking". Hence, since effective zero-shot recognition heavily relies on the semantic relations between seen and unseen classes, class separability must be balanced with the preservation of the semantics available in pre-trained embeddings. + +Semantic preservation To preserve the semantics of the initial embeddings $\pmb{t}_i$ , a simple approach would be to use $\pmb{t}_i$ as proxies and constrain the new embeddings $\pmb{w}_i$ to be in their proximity + +$$ +\mathcal {L} _ {s e m} ^ {p r o x} = \sum_ {c = 1} ^ {| \mathcal {T} |} \| \boldsymbol {w} _ {c} - \boldsymbol {t} _ {c} \| _ {2}. \tag {2} +$$ + +However, this constraint would significantly limit the separability of class embeddings $\mathbf{w}$ , given their direct ties to $\pmb{t}$ . Instead, we seek new class embeddings $\mathbf{w}$ that retain a similar ranking of similarities to other classes. This is accomplished using a margin ranking loss of the form + +$$ +\mathcal {L} _ {s e m} ^ {r a n k} = \sum_ {c = 1} ^ {| \mathcal {T} |} \sum_ {\substack {i = 1 \\ c \neq i \neq j}} ^ {| \mathcal {T} |} \sum_ {j = 1} ^ {| \mathcal {T} |} \max \left\{0, m - \operatorname {sign} \left(d _ {c i} ^ {\boldsymbol {t}} - d _ {c j} ^ {\boldsymbol {t}}\right) \left(d _ {c i} ^ {\boldsymbol {w}} - d _ {c j} ^ {\boldsymbol {w}}\right) \right\} \tag{3} +$$ + +where $m$ is a margin coefficient, and $d_{ci}^{\pmb{t}} = \| \pmb{t}_{c} - \pmb{t}_{i}\|_{2}$ and $d_{ci}^{\pmb{w}} = \| \pmb{w}_{c} - \pmb{w}_{i}\|_{2}$ are the $\ell_2$ distances between pairs of class embeddings, using either the pre-trained embeddings $\pmb{t}$ or the new optimized embeddings $\pmb{w}$ , respectively. Intuitively, this ranking loss will seek to organize the new embeddings $\pmb{w}$ such that if $\pmb{t}_{c}$ is closer to $\pmb{t}_{i}$ than to $\pmb{t}_{j}$ , then $\pmb{w}_{c}$ will also be closer to $\pmb{w}_{i}$ than to $\pmb{w}_{j}$ . Note that Eq. 3's complexity grows cubically with the number of classes. While manageable for the tested datasets, it can be impractical for large vocabularies. A potential solution is to sample a large (but fixed) number of class triplets per iteration, and replace Eq. 3 with an expectation over these triplets. + +To balance separability and semantic preservation, we obtain the class embeddings by minimizing the joint loss + +$$ +\underset {\boldsymbol {w} _ {1}, \dots , \boldsymbol {w} _ {| \mathcal {T} |}} {\arg \min } (1 - \alpha) \mathcal {L} _ {s e m} + \alpha \mathcal {L} _ {s e p}, \tag {4} +$$ + +where $\alpha$ is set to 0.5. Given that the objective is differentiable, we optimize the new embeddings using gradient descent, starting from the initialization $\pmb{w}_i = \pmb{t}_i\forall i$ + +# 3.3 Supervised Audio-Visual Language Alignment + +Given the enhanced class embeddings $\pmb{w}_1, \dots, \pmb{w}_{|\mathcal{T}|}$ (with improved class separability), we empirically show that a contrastive objective is sufficient to learn state-of-the-art models for generalized zero-shot learning of audio-visual events. Specifically, let $\pmb{x}_i^{av} = f_\theta^{av}(\pmb{v}_i, \pmb{a}_i)$ be the multi-modal representation of an audio-visual sample $(v_i, a_i)$ belonging to class $y_i$ . Then, we align representations $\pmb{x}_i^{av}$ with the corresponding class embeddings $\pmb{w}_{y_i}$ by minimizing the contrastive loss + +$$ +\mathcal {L} _ {A V L A} = - \log \frac {\exp \left(\operatorname {s i m} \left(\boldsymbol {x} _ {i} ^ {a v} , \boldsymbol {w} _ {y _ {i}}\right)\right)}{\sum_ {k = 1} ^ {B} \exp \left(\operatorname {s i m} \left(\boldsymbol {x} _ {i} ^ {a v}, \boldsymbol {w} _ {y _ {k}}\right)\right)} \tag {5} +$$ + +where $\mathrm{sim}(\cdot, \cdot)$ is a similarity function used to contrast the class embedding $\mathbf{w}_{y_i}$ against others in the current batch of data $\mathbf{w}_{y_k}$ . + +Similarity. While contrastive objectives typically rely on cosine similarity metrics for $\sin (\cdot ,\cdot):\mathcal{X}\times \mathcal{T}\rightarrow \Re$ , we found that non-linear similarity functions can better capture the relationship between audio-visual representations and class embeddings. Specifically, we draw inspiration from transformer attention modules and use a multi-head cross-attention layer. In this layer, class embeddings $\pmb{w}_c$ are used as queries and audio-visual representations $\pmb{x}_i^{av}$ as key-value pairs. After cross-attention, class similarities are computed through a linear projection $\pmb{p}$ . Optimizing Eq. 5 pushes the learned embeddings closer to the text embeddings of the same category. During inference, we use the cross-attention transformer layer to output $C$ similarity scores corresponding to each category, including unseen ones. We then classify the audio-visual samples according to the class with highest similarity. + +# 4 Experiments + +In this section, we evaluate the proposed EZ-AVGZL on three benchmarks for audio-visual generalized zero-shot learning, including VGGSound-GZSL [28], UCF-GZSL [28], and ActivityNet-GZSL [28]. + +# 4.1 Experimental Setup + +Datasets. VGGSound-GZSL [28] contains 93,752 videos across 276 classes, selected from VGGSound [7], a large audio-visual dataset of 309 classes over 200k videos. We follow the prior work [27, 28] and use the same split for train/val(U)/test(U) over 138/69/69 classes. UCF-GZSL [28] includes 6,816 videos from 51 classes, chosen from UCF101 [51], the original video action recognition dataset, which consists of over 13k videos in 101 classes. We use the same train/val(U)/test(U) split over 30/12/9 classes in previous work [27, 28]. ActivityNet-GZSL [28] contains 20k videos in 200 classes of varying duration for action recognition [13]. We use the same train/val(U)/test(U) split over 99/51/50 classes as in the prior work [27, 28] + +Evaluation Metrics. In line with prior work [27, 28], we use the mean class accuracy to evaluate all models. For ZSL performance, we evaluate only on the subset of test samples from unseen classes. For the GZSL performance, we evaluate on the full test set, which includes both seen and unseen classes. We report results for the subsets of seen (S) and unseen (U) classes, as well as their harmonic mean. + +Implementation. We follow previous work [27] and use 8 heads with a dimension of 64 per head for all attention blocks in the cross-attention transformer. The model is trained for 50 epochs with a batch size of 64, using the Adam optimizer [20] with a learning rate of 1e-4, running average coefficients $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , and a weight decay of 1e-5. Following the prior work [27, 28], we use the word2vec [29] embeddings for text, and VGGish [15] embeddings for audio, and C3D [56] features for video frames. + +Table 1: Quantitative results of VGGSound-GZSL benchmark. + +
MethodSeenUnseenHarmonicMeanZSL
AttentionFusion [9]14.133.004.953.37
Perceiver [19]13.253.034.933.44
CJME [48]10.862.223.683.72
AVGZSLNet [26]15.023.195.264.81
AVCA [28]12.636.198.316.91
TCaF [27]12.636.728.777.41
Hyper-alignment [16]12.506.448.507.25
Hyper-single [16]12.565.037.185.47
Hyper-multiple [16]15.626.008.677.31
EZ-AVGZL (ours)17.268.6811.558.92
+ +Table 2: Quantitative results of UCF-GZSL benchmark. + +
MethodSeenUnseenHarmonic MeanZSL
AttentionFusion [9]39.3418.2924.9720.21
Perceiver [19]46.8526.8234.1128.12
CJME [48]33.8924.8228.6529.01
AVGZSLNet [26]74.7924.1536.5131.51
AVCA [28]63.1530.7241.3437.72
TCaF [27]67.1440.8350.7844.64
Hyper-alignment [16]57.1333.8642.5239.80
Hyper-single [16]63.4734.8544.9939.86
Hyper-multiple [16]74.2635.7948.3052.11
EZ-AVGZL (ours)78.3246.3558.2448.27
+ +# 4.2 Comparison to Prior Work + +In this work, we propose a novel and effective framework for audio-visual generalized zero-shot learning. In order to demonstrate the effectiveness of the proposed EZ-AVGZL, we comprehensively compare it to previous audio-visual baselines [9, 19] adapted to our problem and current state-of-the-art audio-visual generalized zero-shot learning approaches [16, 26-28, 48]. + +For the VGGSound-GZSL dataset, we report the quantitative comparison results in Table 1. As shown, we achieve the best results in all metrics for both seen and unseen classes compared to previous audio-visual generalized zero-shot learning approaches. In particular, the proposed EZ-AVGZL outperforms Hyper-multiple [16], the current state-of-the-art audio-visual generalized zero-shot learning baseline, by 1.64@Seen, 2.68@Unseen, 2.88@Harmonic Mean, and 1.61@ZSL. Moreover, we also perform better than TCaF [27], a strong audio-visual generalized zero-shot learning baseline, by 4.63@Seen, 1.96@Unseen, 2.78@Harmonic Mean, and 1.51@ZSL. Furthermore, we achieve significant performance gains compared to AVCA [28], the first baseline using cross-modal attention and textual label embeddings. This indicates the importance of optimized text transformation and non-linear similarity functions in learning better transferrable knowledge for generalized zero-shot learning. + +Table 3: Quantitative results of ActivityNet-GZSL benchmark. + +
MethodSeenUnseenHarmonic MeanZSL
AttentionFusion [9]11.153.375.184.88
Perceiver [19]18.254.276.924.47
CJME [48]10.755.557.326.29
AVGZSLNet [26]13.705.968.306.39
AVCA [28]16.777.049.927.58
TCaF [27]30.127.6512.207.96
Hyper-alignment [16]29.778.7713.559.13
Hyper-single [16]24.6110.1014.3210.37
Hyper-multiple [16]36.989.6015.2510.39
EZ-AVGZL (ours)38.5212.8519.2713.01
+ +Table 4: Ablation studies on Class Embedding Optimization and Audio-Visual-Language Alignment. + +
CEPAVLASeenUnseenHarmonic MeanZSL
XX12.636.728.777.41
✓X15.357.9210.458.36
X✓14.677.239.697.91
✓✓17.268.6811.558.92
+ +Table 5: Ablation studies on optimization objectives in Class Embedding Optimization. + +
LsemLsepSeenUnseenHarmonic MeanZSL
XX14.677.239.327.91
LproxsemX15.757.8310.468.26
X✓16.138.1710.858.45
Lproxsem✓17.268.6811.558.92
Lranksem✓18.0510.2813.1010.36
+ +Additionally, we observe significant gains in the UCF-GZSL and ActivityNet-GZSL benchmarks, as shown in Table 2 and Table 3. Compared to TCaF [27], a strong baseline using three manually-designed objectives, we achieve result gains of 11.18@Seen, 5.52@Unseen, 7.46@Harmonic Mean, and 3.63@ZSL on UCF-GZSL dataset. When evaluated on the challenging ActivityNet-GZSL benchmark, our method still outperforms TCaF [27] by 8.40@Seen, 5.20@Unseen, 7.07@Harmonic Mean, and 5.05@ZSL. We also achieve better results than AVCA [28], a cross-modal baseline based on cross-modal attention between audio and visual embeddings. These results demonstrate the effectiveness of our approach in learning optimized textual label embeddings and non-linear similarity functions for generalized audio-visual zero-shot recognition. + +# 4.3 Experimental Analysis + +In this section, we conduct ablation studies to demonstrate the benefit of introducing the Class Embedding Optimization (CEO) and Audio-Visual Language Alignment (AVLA) modules. We also carry out extensive experiments to investigate optimization objectives in class embedding optimization, non-linear similarity functions, and generalization across diverse cross-modal settings. + +Class Embedding Optimization (CEO) & Audio-Visual Language Alignment (AVLA). In order to demonstrate the effectiveness of CEO and AVLA, we conduct an ablation study and report the quantitative results on VGGSound-GZSL dataset in Table 4. + +As observed, adding CEO to the vanilla baseline improves the results by 2.72@Seen, 1.20@Unseen, 1.68@Harmonic Mean, and 0.95@ZSL, which validates the benefit of optimizing class embeddings so as to obtain separated embeddings that preserve the semantic relationships between classes. On the other hand, introducing AVLA in the baseline increases the zero-shot performance across all metrics. More importantly, incorporating both CEP and AVLA into the baseline significantly raises the performance by 4.63@Seen, 1.96@Unseen, 2.78@Harmonic Mean, and 1.51@ZSL. These improvements validate the importance of text transformation and supervised text audio-visual contrastive in extracting separated textual label embeddings for AVGZSL. + +Objectives in Class Embedding Optimization. Optimization objectives in class embedding transformation, including global semantic preservation and class separability, are essential for learning optimized textual class embeddings for zero-shot learning. To explore the effects of each objective more comprehensively, we ablate each text optimization objective and report the quantitative results in Table 5. We observe that adding the semantic preservation loss to the vanilla baseline improves the performance by 1.08@Seen, 0.60@Unseen, 1.14@Harmonic Mean, and 0.35@ZSL. On the other hand, introducing the class separability loss in the baseline enhances the zero-shot performance across all metrics. Furthermore, combining the semantic preservation and class separability objectives for the text transformer significantly increases the performance by 2.59@Seen, 3.05@Unseen, 3.78@Harmonic Mean, and 2.45@ZSL. This result validates the importance of class separability loss in balancing the preservation of the semantics available in pre-trained representations for zero-shot learning. Replacing the Euclidean-based semantic preservation loss $\mathcal{L}_{sem}^{prox}$ with a ranking based loss $\mathcal{L}_{sem}^{rank}$ further achieves performance gains of 0.79@Seen, 1.60@Unseen, 1.55@Harmonic Mean, and 1.44@ZSL. These results demonstrate the effectiveness of both semantic preservation and class separability objectives in class embedding optimization to generate meaningful text representations. + +Non-linear Similarity Functions. The non-linear similarity operator used in supervised audio-visual language alignment impacts the extracted cross-modal representations. To explore these effects more comprehensively, we varied the functions from $\{\mathrm{Cosine - similarity}$ , Linear, MLP, Cross-Attention\}. For Cosine-Similarity, we directly compute the dot product between audio-visual embeddings $\mathbf{x}_i^{av}$ and text embeddings $\mathbf{t}_i$ . For Linear, we apply a linear layer to the concatenation of both embeddings to predict the similarity score. For MLP, we replace the linear layer with a multi-layer perceptron to generate the score. For Cross-Attention, we use text embeddings $\mathbf{t}_i$ as the queries and audio-visual representations $\mathbf{x}_i^{av}$ as key-value pairs in a transformer attention module to output the score. The comparison results of zero-shot performance on VGGSound-GZSL benchmark are reported in Table 6. When using Cosine-Similarity as the function, we achieve the worst results across all metrics. Replacing the cosine-similarity operator with both Linear increases the seen results by $2.55@Seen$ , but decreases the performance by $0.07@Unseen$ and $0.12@ZSL$ . This might be because adding complexity to the function causes the model to overfit the seen classes in the + +Table 6: Exploration studies on non-linear similarity functions $\phi \left( \cdot \right)$ . + +
φ(·)SeenUnseenHarmonic MeanZSL
Cosine-Similarity13.978.5910.648.67
Linear16.528.5211.248.55
MLP17.038.4611.308.49
Cross-Attention17.268.6811.558.92
+ +Table 7: Exploration studies on diverse cross-modal settings. CEP denotes Class Embedding Optimization. + +
Text EncoderAudio EncoderVisual EncoderCEPSeenUnseenHarmonic MeanZSL
word2vecVGGishC3D×12.636.728.777.41
✓17.268.6811.558.92
LLaMaAudioMAEDINO×18.3510.6313.4612.26
✓23.9513.3217.1213.73
CLAPCLAPX-CLIP×27.7914.0118.6317.39
✓33.2820.1525.1021.06
ImageBindImageBindImageBind×28.3214.6819.3419.87
✓35.5621.6326.9022.35
+ +training data, , leading to poorer generalization to unseen categories. Additionally, using MLP as the function operator significantly raises the performance by 3.06@Seen, but slightly deteriorates the results on unseen classes. Our method achieves the best results across all metrics by using the cross-attention operator to predict the similarity score. These improved results demonstrate the effectiveness of the cross-attention operator in balancing the trade-off between linear and non-linear similarity functions. + +Generalization to Diverse Cross-modal Settings. To validate the generalizability of the proposed method across a flexible number of feature encoders, we experiment with different settings: unimodal, paired (video-text, audio-text), and unified embedding space (image-audio-text). In the unimodal setting, we use DINO [3] for the visual encoder, AudioMAE [18] for the audio encoder, and LLaMa [55] for the text encoder. In the paired setting, we use X-CLIP [25] pre-trained with video-text objectives for the visual encoder, and CLAP [59] pre-trained with audio-text objectives for the audio and text encoder, respectively. In the unified embedding setting, we apply ImageBind [12] pre-trained with image-audio-text alignment in the same embedding space as the encoder for all three modalities. The quantitative results on the VGGSound-GZSL benchmark are compared in Table 7. As observed, the unimodal setting, which uses transformers with strong unimodal features, significantly outperforms the vanilla baseline that uses convolution networks, where EZ-AVGZL achieves performance gains of 6.69@Seen, 4.64@Unseen, 5.77@Harmonic Mean, and 4.81@ZSL using class embedding optimization. In the paired setting, the proposed method further increases the results by 9.33@Seen, 6.93@Unseen, 7.98@Harmonic Mean, and 7.331@ZSL, benefiting from the text-audio correspondence and text-video alignment pre-trained on large-scale data. In the unified embedding setting using ImageBind [12] embeddings, we achieve the best results across all metrics + +Table 8: Comparison of class embeddings w/o and w optimization. + +
base classNN (w/o opt)D (w/o opt)NN (w opt)D (w opt)
alligators, crocodiles hissingchurch bell ringing0.455playing marimba, xylophone0.498
barn swallow callingblack capped chickadee calling0.307black capped chickadee calling0.465
basketball bounceplaying volleyball0.324bouncing on trampoline0.394
bee, wasp, etc. buzzingmotorboat, speedboat acceleration0.322fly, housefly buzzing0.333
bird squawkingpleasant crowing0.400duck quacking0.496
black capped chickadee callingbarn swallow calling0.307magpie calling0.436
bouncing on trampolinebasketball bounce0.416basketball bounce0.494
bowling impactplaying table tennis0.410striking bowling0.467
bull bellowingeagle screaming0.449dog growling0.532
canary callingwhale calling0.403black capped chickadee calling0.513
+ +![](images/a40c9db7660021e7ee79bdb06dc55e75846e784520815d25cc3adc1c518db28b.jpg) +Fig. 3: Confusion matrices of zero-shot predictions on VGGSound-GZSL using model with and without class embedding optimization. + +compared to other settings. This might be because text embeddings with unified training on the other two modalities strengthen the alignment for audio-visual zero-shot learning. Furthermore, when class embedding optimization is added to all settings, we improve the baseline based on the initial text embeddings, further demonstrating the effectiveness of the proposed class embedding optimization in generating text representations for AVZSL. + +Qualitative Comparisons with and without Optimization. To provide insights into why the proposed loss terms improve the separability of class embeddings while keeping semantics, we conduct qualitative studies. These include comparing the visualization of the class embedding space with and without optimization. The comparisons of class embeddings with and without optimization are shown in Table 8. We report the base class and identify 1) the nearest neighbor (NN) class and 2) the distance (D) to the nearest neighbor. We observe larger distances and similar NN classes. The confusion matrices for class-wise performances of the model are reported in Fig. 3. + +Impact of Hyper-parameters $\alpha$ and Margin $m$ . To investigate the impact of $\alpha$ and $m$ on the performance thoroughly, we set $\alpha$ from $\{0.1, 0.5, 0.9\}$ and $m$ from $\{0, 0.5, 1\}$ . The quantitative comparisons on VGGSound-GZSL are reported in Table 9. As can be seen, with the increase of $\alpha$ from 0.1 to 0.5, the proposed approach achieves better results regarding all metrics. However, when $\alpha$ is increased to 0.9, the performance decreases a lot, which might be caused by not + +Table 9: Study of $\alpha$ and margin $m$ on VGGSound-GZSL. + +
αmSeenUnseenHarmonic MeanZSL
0.518.0510.2813.1010.36
0.1116.358.3911.098.58
0.916.268.7311.368.92
0.5017.829.7112.579.67
0.517.679.8512.659.95
+ +Table 10: Study of distance functions on VGGSound-GZSL. + +
DistanceSeenUnseenHarmonic MeanZSL
cosine18.0510.2813.1010.36
l217.589.3212.189.27
l117.358.7811.668.85
+ +much semantics preservation during training. Meanwhile, with the decrease of $m$ , we observe a clearly dropping trend in our metrics. + +Ablating Distance Functions for Eq. 1-2. Distance functions in Eq. 1-2 is crucial for us to maximize separation while preserving semantics. We experimented with different distance functions, $\ell_2$ and $\ell_1$ distances for Eq. 1-2. Note that while, Eq. 1-2 show $\ell_2$ distances in the paper, the embeddings are normalized. The quantitative results are shown in Table 10. We can observe that using the cosine distance function performed the best regarding all metrics. However, when optimizing $\ell_1$ distance functions, we achieve the worst results in terms of all metrics. These ablation results highlight the importance of designing appropriate distance functions in maximizing separation while preserving the semantics of text class embeddings. + +# 5 Conclusion + +In this work, we present a novel and effective framework, EZ-AVGZL, that aligns audio-visual representations with optimized text embeddings to achieve audio-visual generalized zero-shot learning. We leverage differential optimization to learn better-separated textual representations with maximal separability and semantic preservation. Additionally, we introduce supervised audio-visual language alignment to learn the correspondence between audio-visual features and textual embeddings, which captures the cross-modal dynamics with class labels as prompts. Experimental results on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL datasets demonstrate our method's superiority against previous baselines. Extensive ablation studies also validate the importance of class embedding optimization and supervised audio-visual language alignment for audio-visual generalized zero-shot learning. We also conduct comprehensive experiments to demonstrate the effectiveness of non-linear similarity functions and generalization to diverse cross-modal settings. + +Broader Impact. This work makes substantial progress in audio-visual generalized zero-shot learning. Our method is novel in its simplicity, demonstrating that a simple supervised contrastive objective, without "bells and whistles", outperforms complex alignment strategies. Without our work, progress in this area would likely be driven by increasingly complex methods, which are difficult to implement, reproduce and compare. Beyond the strong results and simplicity of our method, the class embedding optimization procedure is also technically novel to maximize semantic separation while keeping semantics preserved. + +# References + +1. Arandjelovic, R., Zisserman, A.: Look, listen and learn. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 609-617 (2017) 3 +2. Aytar, Y., Vondrick, C., Torralba, A.: Soundnet: Learning sound representations from unlabeled video. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2016) 3 +3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) 12 +4. Chen, C., Jain, U., Schissler, C., Gari, S.V.A., Al-Halah, Z., Ithapu, V.K., Robinson, P., Grauman, K.: Soundspaces: Audio-visual navigation in 3d environments. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 17-36 (2020) 3 +5. Chen, C., Majumder, S., Ziad, A.H., Gao, R., Kumar Ramakrishnan, S., Grauman, K.: Learning to set waypoints for audio-visual navigation. In: Proceedings of International Conference on Learning Representations (ICLR) (2021) 3 +6. Chen, C., Schissler, C., Garg, S., Kobernik, P., Clegg, A., Calamia, P., Batra, D., Robinson, P.W., Grauman, K.: Soundspaces 2.0: A simulation platform for visual-acoustic learning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track (2022) 3 +7. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 721-725. IEEE (2020) 8 +8. Ephrat, A., Mosseri, I., Lang, O., Dekel, T., Wilson, K., Hassidim, A., Freeman, W.T., Rubinstein, M.: Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619 (2018) 3 +9. Fayek, H.M., Kumar, A.: Large scale audiovisual learning of sounds with weakly labeled data. In: Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (2021) 9, 10 +0. Gan, C., Huang, D., Zhao, H., Tenenbaum, J.B., Torralba, A.: Music gesture for visual sound separation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10478-10487 (2020) 3 +1. Gao, R., Grauman, K.: 2.5d visual sound. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 324-333 (2019) 3 +2. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 4, 12 +3. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 961-970 (2015) 8 +4. Hershey, J., Casey, M.: Audio-visual sound separation via hidden markov models Advances in Neural Information Processing Systems 14 (2001) 3 +5. Hershey, S., Chaudhuri, S., Ellis, D.P.W., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., Slaney, M., Weiss, R.J., Wilson, K.: Cnn architectures for large-scale audio classification. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2017) 8 + +16. Hong, J., Hayden, Z., Han, J., Fang, P., Harandi, M., Petersson, L.: Hyperbolic audio-visual zero-shot learning. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023) 4, 9, 10 +17. Hu, D., Nie, F., Li, X.: Deep multimodal clustering for unsupervised audiovisual learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9248-9257 (2019) 3 +18. Huang, P.Y., Xu, H., Li, J., Baevski, A., Auli, M., Galuba, W., Metze, F., Feichtenhofer, C.: Masked autoencoders that listen. In: Proceedings of Advances In Neural Information Processing Systems (NeurIPS) (2022) 12 +19. Jaegle, A., Gimeno, F., Brock, A., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver: General perception with iterative attention. In: Proceedings of the International Conference on Machine Learning (ICML) (2021) 9, 10 +20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 8 +21. Korbar, B., Tran, D., Torresani, L.: Cooperative learning of audio and video models from self-supervised synchronization. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2018) 3 +22. Lin, Y.B., Li, Y.J., Wang, Y.C.F.: Dual-modality seq2seq network for audio-visual event localization. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2002–2006 (2019) 3 +23. Lin, Y.B., Tseng, H.Y., Lee, H.Y., Lin, Y.Y., Yang, M.H.: Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2021) 3 +24. Lin, Y.B., Wang, Y.C.F.: Audiovisual transformer with instance attention for audiovisual event localization. In: Proceedings of the Asian Conference on Computer Vision (ACCV) (2020) 3 +25. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP:: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285 (2022) 12 +26. Mazumder, P., Sing, P., Kumar Parida, K., Namboodiri, V.P.: Avgzslnet: Audiovisual generalized zero-shot learning by reconstructing label features from multimodal embeddings. In: Proceedings of 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 3089-3098 (2021) 3, 9, 10 +27. Mercea, O.B., Hummel, T., Koepke, A.S., Akata, Z.: Temporal and cross-modal attention for audio-visual zero-shot learning. In: Proceedings of European Conference on Computer Vision (ECCV) (2022) 2, 4, 5, 6, 8, 9, 10 +28. Mercea, O.B., Riesch, L., Koepke, A.S., Akata, Z.: Audio-visual generalised zero-shot learning with cross-modal attention and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10553–10563 (2022) 2, 4, 8, 9, 10 +29. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. In: Proceedings of International Conference on Learning Representations (ICLR) (2013) 8 +30. Mo, S., Morgado, P.: Benchmarking weakly-supervised audio-visual sound localization. In: European Conference on Computer Vision (ECCV) Workshop (2022) 3 +31. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) 3 + +32. Mo, S., Morgado, P.: A unified audio-visual learning framework for localization, separation, and recognition. In: Proceedings of the International Conference on Machine Learning (ICML) (2023) 3 +33. Mo, S., Morgado, P.: Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling. arXiv preprint arXiv:2312.01017 (2023) 3 +34. Mo, S., Pian, W., Tian, Y.: Class-incremental grouping network for continual audio-visual learning. arXiv preprint arXiv:2309.05281 (2023) 3 +35. Mo, S., Shi, J., Tian, Y.: DiffAVA: Personalized text-to-audio generation with visual alignment. arXiv preprint arXiv:2305.12903 (2023) 3 +36. Mo, S., Shi, J., Tian, Y.: Text-to-audio generation synchronized with videos. arXiv preprint arXiv:2403.07938 (2024) 3 +37. Mo, S., Tian, Y.: Multi-modal grouping network for weakly-supervised audiovisual video parsing. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) 3 +38. Mo, S., Tian, Y.: Semantic-aware multi-modal grouping for weakly-supervised audio-visual video parsing. In: European Conference on Computer Vision (ECCV) Workshop (2022) 3 +39. Mo, S., Tian, Y.: Audio-visual grouping network for sound localization from mixtures. arXiv preprint arXiv:2303.17056 (2023) 3 +40. Mo, S., Tian, Y.: AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023) 3 +41. Mo, S., Tian, Y.: Semantic grouping network for audio source separation. arXiv preprint arXiv:2407.03736 (2024) 3 +42. Morgado, P., Li, Y., Vasconcelos, N.: Learning representations from audio-visual spatial alignment. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS). pp. 4733-4744 (2020) 3 +43. Morgado, P., Li, Y., Costa Pereira, J., Saberian, M., Vasconcelos, N.: Deep hashing with hash-consistent large margin proxy embeddings. International Journal of Computer Vision (2020) 6 +44. Morgado, P., Misra, I., Vasconcelos, N.: Robust audio-visual instance discrimination. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12934–12945 (2021) 3 +45. Morgado, P., Vasconcelos, N., Langlois, T., Wang, O.: Self-supervised generation of spatial audio for 360 video. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2018) 3 +46. Morgado, P., Vasconcelos, N., Misra, I.: Audio-visual instance discrimination with cross-modal agreement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12475–12486 (June 2021) 3 +47. Owens, A., Wu, J., McDermott, J.H., Freeman, W.T., Torralba, A.: Ambient sound provides supervision for visual learning. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 801-816 (2016) 3 +48. Parida, K.K., Matiyali, N., Guha, T., Sharma, G.: Coordinated joint multimodal embeddings for generalized audio-visual zero-shot classification and retrieval of videos. In: Proceedings of 2020 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 3240–3249 (2020) 3, 9, 10 +49. Pian, W., Mo, S., Guo, Y., Tian, Y.: Audio-visual class-incremental learning. arXiv preprint arXiv:2308.11073 (2023) 3 +50. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4358-4366 (2018) 3 + +51. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) 8 +52. Tammes, P.M.L.: On the origin of number and arrangement of the places of exit on the surface of pollen-grains. Recueil des travaux botaniques neerlandais 27(1), 1-84 (1930) 6 +53. Tian, Y., Li, D., Xu, C.: Unified multisensory perception: Weakly-supervised audiovisual video parsing. In: Proceedings of European Conference on Computer Vision (ECCV). p. 436-454 (2020) 3 +54. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of European Conference on Computer Vision (ECCV) (2018) 3 +55. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 12 +56. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2015) 8 +57. Wu, Y., Yang, Y.: Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1326–1335 (2021) 3 +58. Wu, Y., Zhu, L., Yan, Y., Yang, Y.: Dual attention matching for audio-visual event localization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 6291–6299 (2019) 3 +59. Wu, Y., Chen, K., Zhang, T., Hui, Y., Berg-Kirkpatrick, T., Dubnov, S.: Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In: IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP (2023) 12 +60. Zhang, L., Mo, S., Zhang, Y., Morgado, P.: Audio-synchronized visual animation arXiv preprint arXiv:2403.05659 (2024) 3 +61. Zhao, H., Gan, C., Ma, W.C., Torralba, A.: The sound of motions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1735-1744 (2019) 3 +62. Zhao, H., Gan, C., Rouditchenko, A., Vondrick, C., McDermott, J., Torralba, A.: The sound of pixels. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 570-586 (2018) 3 \ No newline at end of file diff --git a/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/images.zip b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d7ab1bf78e0256690d5bbc85b0cee22f4abc8430 --- /dev/null +++ b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8a66b8542af0b37f9f30b1e5f33fb6f79b65de24de6a388b1b71c6968f6b9d6 +size 412963 diff --git a/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/layout.json b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..40c4bf54ee22dd4bc1d45f89866a1e31b9cf2dbf --- /dev/null +++ b/2024/Audio-visual Generalized Zero-shot Learning the Easy Way/layout.json @@ -0,0 +1,9038 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 111, + 462, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 111, + 462, + 147 + ], + "spans": [ + { + "bbox": [ + 151, + 111, + 462, + 147 + ], + "type": "text", + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 227, + 168, + 385, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 168, + 385, + 180 + ], + "spans": [ + { + "bbox": [ + 227, + 168, + 385, + 180 + ], + "type": "text", + "content": "Shentong Mo1 and Pedro Morgado2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 234, + 190, + 378, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 190, + 378, + 212 + ], + "spans": [ + { + "bbox": [ + 234, + 190, + 378, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 234, + 190, + 378, + 212 + ], + "type": "text", + "content": " Carnegie Mellon University " + }, + { + "bbox": [ + 234, + 190, + 378, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 234, + 190, + 378, + 212 + ], + "type": "text", + "content": " University of Wisconsin-Madison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 247, + 452, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 247, + 452, + 521 + ], + "spans": [ + { + "bbox": [ + 160, + 247, + 452, + 521 + ], + "type": "text", + "content": "Abstract. Audio-visual generalized zero-shot learning is a rapidly advancing domain that seeks to understand the intricate relations between audio and visual cues within videos. The overarching goal is to leverage insights from seen classes to identify instances from previously unseen ones. Prior approaches primarily utilized synchronized auto-encoders to reconstruct audio-visual attributes, which were informed by cross-attention transformers and projected text embeddings. However, these methods fell short of effectively capturing the intricate relationship between cross-modal features and class-label embeddings inherent in pretrained language-aligned embeddings. To circumvent these bottlenecks, we introduce a simple yet effective framework for Easy Audio-Visual Generalized Zero-shot Learning, named EZ-AVGZL, that aligns audio-visual embeddings with transformed text representations. It utilizes a single supervised text audio-visual contrastive loss to learn an alignment between audio-visual and textual modalities, moving away from the conventional approach of reconstructing cross-modal features and text embeddings. Our key insight is that while class name embeddings are well aligned with language-based audio-visual features, they don't provide sufficient class separation to be useful for zero-shot learning. To address this, our method leverages differential optimization to transform class embeddings into a more discriminative space while preserving the semantic structure of language representations. We conduct extensive experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks. Our results demonstrate that our EZ-AVGZL achieves state-of-the-art performance in audio-visual generalized zero-shot learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 533, + 452, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 533, + 452, + 554 + ], + "spans": [ + { + "bbox": [ + 160, + 533, + 452, + 554 + ], + "type": "text", + "content": "Keywords: Audio-Visual Learning " + }, + { + "bbox": [ + 160, + 533, + 452, + 554 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 533, + 452, + 554 + ], + "type": "text", + "content": " Zero-shot Learning " + }, + { + "bbox": [ + 160, + 533, + 452, + 554 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 533, + 452, + 554 + ], + "type": "text", + "content": " Audio-Visual Generalized Zero-shot Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 578, + 229, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 229, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 229, + 590 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 605, + 481, + 665 + ], + "type": "text", + "content": "In the field of machine learning, the ability of models to recognize patterns and make predictions is crucial. Zero-shot learning exemplifies this capability, which leverages known information to understand the unknowns. To further explore its potential and applications, the principle of zero-shot learning has been applied in various contexts. One such context is Audio-visual generalized" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 117, + 434, + 284 + ], + "blocks": [ + { + "bbox": [ + 177, + 117, + 434, + 284 + ], + "lines": [ + { + "bbox": [ + 177, + 117, + 434, + 284 + ], + "spans": [ + { + "bbox": [ + 177, + 117, + 434, + 284 + ], + "type": "image", + "image_path": "dbd5321a80463c291c70e52f3bbadec3715ee5b46d42d863b2891df9983574ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 296, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 344 + ], + "type": "text", + "content": "Fig.1: Comparison of our EZ-AVGZL with state-of-the-art methods on VGGSound-GZSL, UCF-GZSL, ActivityNet-GZSL benchmarks in terms of Harmonic Mean (higher is better) for seen and unseen classes. Our method significantly outperforms previous baselines in terms of all datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 373, + 481, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 481, + 397 + ], + "type": "text", + "content": "zero-shot learning (AVGZSL), an emerging topic of study that predicts unseen audio-visual categories by leveraging knowledge learned from seen classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 543 + ], + "type": "text", + "content": "Within the scope of AVGZSL, previous approaches [27, 28] have focused on designing autodecoders for reconstructing audio-visual attributes. For example, TCaF [27] adopted a temporal contrastive feature alignment strategy. This strategy aligns audio-visual cues with word2vec embeddings of class prompts by defining both cross-entropy and regression losses to reduce the distance between the output embedding for a sample and the corresponding projected word2vec embedding. These cross-modal and textual features were derived from cross-attention transformers and the projected text embeddings. However, the complexity of the network for AVGZSL, which is derived from the synergy between cross-attention transformers and the corresponding projected word2vec embeddings, makes it challenging to understand the deeper interplay between cross-modal features and class label embeddings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "In the context of audio-visual zero-shot learning, the main challenge lies in the absence of unseen classes during training. Without the supervision, the prediction of cross-modal class labels performs poorly on these audio-visual samples. To address the challenge, we propose a novel framework for Easy Audio-Visual Generalized Zero-shot Learning, referred to as EZ-AVGZL. This framework aligns audio-visual embeddings with transformed text representations in a simple yet effective manner. EZ-AVGZL learn well-separated text semantics across seen and unseen class names to achieve maximal separation while retaining the semantic relations between each other. During training, we leverage a simple supervised contrastive loss to learn the cross-modal alignment between audio-visual features" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "S. Mo and P. Morgado" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "and transformed text embeddings. EZ-AVGZL enhances class separability while preserving semantics through the use of differential optimization, thereby generating more distinct text embeddings. Furthermore, it employs a single supervised text to audio-visual contrastive loss, establishing an alignment between the audio-visual and textual modalities. This approach effectively eliminates the need for reconstructing cross-modal features and text embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "content": "Empirical experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks demonstrate the state-of-the-art performance against previous audio-visual generalized zero-shot learning baselines, as illustrated in Fig. 1. Extensive ablation studies further confirm the importance of class embedding optimization and supervised audio-visual language contrastive objectives in learning language alignment with audio-visual representations. Additionally, we demonstrate the robustness of our approach across various zero-shot learning tasks and observe the benefit of our approach when applied to a variety of foundation models pre-trained from unimodal and cross-modal." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 296, + 388, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 296, + 388, + 307 + ], + "spans": [ + { + "bbox": [ + 146, + 296, + 388, + 307 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 313, + 480, + 407 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 313, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 479, + 348 + ], + "type": "text", + "content": "- We introduce the EZ-AVGZL framework, a paradigm shift in audio-visual generalized zero-shot learning. This framework simplifies the process with a supervised text-audio-visual contrastive loss, enhancing effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 350, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 350, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 137, + 350, + 480, + 384 + ], + "type": "text", + "content": "- We propose a novel text transformation technique designed to construct distinct text embeddings, paving the way for better alignment and, by extension, superior recognition capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 385, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 385, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 137, + 385, + 479, + 407 + ], + "type": "text", + "content": "- We demonstrate that our method achieves state-of-the-art performance on extensive benchmarks in audio-visual generalized zero-shot learning." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 426, + 237, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 237, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 237, + 438 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 568 + ], + "type": "text", + "content": "Audio-Visual Learning. Audio-visual learning problems have been addressed in many previous works [1,2,8,10,14,17,21,32,35,36,38,41,42,44,46,47,50,60-62] to learn the audio-visual correlation between two distinct modalities from videos. Such cross-modal alignments are beneficial for a variety audio-visual tasks, including audio-event localization [22, 24, 30, 31, 33, 39, 40, 54, 58], audio-visual spatialization [4, 11, 42, 45], audio-visual continual learning [34, 49], audio-visual navigation [4-6] and audio-visual parsing [23, 37, 53, 57]. In this work, our primary focus is on learning audio-visual representations that align with text embeddings for generalized zero-shot learning, a task that presents more challenges than the aforementioned tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Zero-shot Learning. Zero-shot learning aims to leverage semantic relationships, typically utilizing attributes or word embeddings associated with class labels to make informed predictions about previously unseen categories. This is often achieved by drawing parallels to the semantic information of known categories. Early audio-visual zero-shot learning methods, such as those proposed by Parida et al. [48] and Mazumder et al. [26], attempted to map video, audio, and text into the same feature space and compute the similarity across each other for comparison. For example, Audio-Visual Generalized Zero-shot Learning Network [26]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "introduced a reconstruction module to predict text representations from audio and visual features. More recently, ImageBind [12] proposed learning a joint embedding across six different modalities, especially on audio, text, and video. This approach demonstrated the ability to recognize or retrieve unseen categories without explicit training. Unlike these methods, our approach does not require large-scale pre-training. Instead, we present a simple yet effective approach based on unimodal or crossmodal weights pre-trained on large-scale data to achieve audio-visual generalized zero-shot learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "text", + "content": "Audio-Visual Generalized Zero-shot Learning. Audio-visual generalized zero-shot learning extends the zero-shot learning paradigm to scenarios where both seen and unseen classes are present at test time. The seminal work, AVCA [28] proposed an attention mechanism to selectively focus on discriminative audio and visual cues. By harnessing the combined strength of these cues, AVCA not only achieved superior performance on seen classes but also exhibited proficiency in recognizing unseen classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 456 + ], + "type": "text", + "content": "Following up, TCaF [27] introduced a temporal contrastive feature alignment strategy. This strategy aligns audio-visual cues across time and applies two separate decoders to reconstruct the audio-visual embedding and the projected word2vec embedding to be close to the original word2vec embedding of class labels. More recently, Hong et al. [16] incorporated cross-modality alignment between video and audio features in the hyperbolic space. They used multiple adaptive curvatures for hyperbolic projections to achieve curvature-aware geometric learning. In contrast, we develop a novel framework to aggregate the alignment between audio-visual representations and text embeddings with explicit class embedding optimization. To the best of our knowledge, this work is the first to leverage an explicit class embedding optimization mechanism for audio-visual generalized zero-shot learning. Our experiments in Section 4.2 also demonstrate the effectiveness of our method in classifying both seen and unseen categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 485, + 201, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 201, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 201, + 498 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "We aim to learn a language-aligned representation for a video clip, which includes both the visual frame sequence and corresponding audio, to enable open-dictionary recognition of audio-visual events. To achieve this, we propose a simple yet effective approach, named Easy Audio-Visual Generalized Zero-shot Learning (EZ-AVGZL). This approach seeks to 1) construct multi-modal representations of the data and 2) align them with language representations using a supervised contrastive learning objective while leveraging strong foundation models. The proposed approach is based on two components: class embedding optimization described in Section 3.2 and supervised alignment between language and audio-visual representations described in Section 3.3. Despite its simplicity, EZ-AVGZL outperforms recent state-of-the-art methods on Audio-Visual Generalized Zero-shot Learning (AVGZSL) on a variety of datasets." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "text", + "content": "S. Mo and P. Morgado" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 470, + 244 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 470, + 244 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 470, + 244 + ], + "type": "image", + "image_path": "a752799a4de23b84bc0cf07e68a6196edc59bcdff539a01485a5edf264911d41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": "Fig. 2: Illustration of the proposed Easy Audio-Visual Generalized Zero-shot Learning (EZ-AVGZL) framework. The initial class embeddings " + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": " from a frozen text encoder are optimized with maximal separability and preserved semantics to generate the new embeddings " + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\pmb{w}_i" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": ". Then the cross-attention transformer " + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{av}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": " takes visual and audio features " + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "inline_equation", + "content": "(\\pmb{v}_i, \\pmb{a}_i)" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": " from the unimodal encoder to generate the multi-modal representations " + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{av}" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": ". Finally, a non-linear similarity function is applied to align representations " + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{av}" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": " with the corresponding class embeddings " + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{y_i}" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": " by minimizing the cross-entropy loss between the predicted similarity score and the target score as 1 in the target class entry, while the target is 0 for multi-modal representations " + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\pmb{x}_j^{av}" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": " given visual features " + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\pmb{v}_j" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": " of videos from other classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 401, + 470, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 470, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 470, + 426 + ], + "type": "text", + "content": "3.1 Preliminaries: Audio-Visual Generalized Zero-shot Learning (AVGZSL)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "text", + "content": "Problem Statement. The objective of AVGZSL is to utilize natural language descriptions of audio-visual events " + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "text", + "content": " and the semantic relationships to enable a model to recognize events from a set of audio-visual event classes " + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\mathcal{T}_s \\cup \\mathcal{T}_u" + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "text", + "content": ". This should be possible regardless of whether the model has seen them " + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "inline_equation", + "content": "(t_i \\in \\mathcal{T}_s)" + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "text", + "content": " or not seen them " + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "inline_equation", + "content": "(t_i \\in \\mathcal{T}_u)" + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "text", + "content": " during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "To accomplish this, a model " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{\\theta}:\\mathcal{V}\\times \\mathcal{A}\\to \\mathcal{T}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " is trained on the seen dataset " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s = \\{(v_i,a_i,t_i)\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " for audio-visual event recognition, where " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_{i}\\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " visual clip, typically multiple frames spanning a short time window, " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "a_{i}\\in \\mathcal{A}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " denotes the corresponding audio, and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_i\\in \\mathcal{T}_s" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " denotes the (human-readable) class name of the audio-visual event depicted in video " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "(v_{i},a_{i})" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ". For effective zero-shot learning, a model needs to learn a mapping that can generalize beyond " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " by inferring from the semantic relationships between seen and unseen classes. In other words, the AVGZSL setting evaluates models based on their classification performance on videos " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "(v_{i},a_{i})" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " belonging to both seen " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " and unseen classes " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_u" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ". Architecture Following TCaF [27], we tackle AVGZSL using a variety of pretrained models to encode all input signals. We experiment with a wider variety of pre-trained models than in [27], in all cases, we train our system on a series of unimodal pre-extracted feature representations for the visual frames " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb {v}_i = f_v(\\pmb {v}_i)" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", audio " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb {a}_i = f_a(a_i)" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " and class names " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb {t}_i = f_t(t_i)" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ". As our goal is multi-modal GZSL," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "type": "text", + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "we further encode the unimodal audio and visual representations into a joint multi-modal representation using the fusion model " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{av} = f_{av}(\\pmb{v}_i, \\pmb{a}_i)" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " proposed in [27], which consists of a sequence of " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " transformer layers with cross-attention between audio and visual modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 259 + ], + "type": "text", + "content": "The quality of representations obtained from foundation models is crucial for effective zero-shot learning. However, the focus of this work is the language alignment step. While prior work often relies on elaborate alignment strategies, we demonstrate that a simple supervised contrastive objective between text and audio-visual representations can achieve state-of-the-art performance. This is possible as long as the class name embeddings for seen and unseen events are maximally separated while retaining the semantic relations between each other. An overview of the proposed approach is shown in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 356 + ], + "type": "text", + "content": "Apart from the quality of representations derived from foundation models, language alignment, which is the focus of this work, is a critical component for effective zero-shot learning. While prior work often relies on complex alignment strategies, we demonstrate that a simple supervised contrastive objective between text and audio-visual representations can achieve state-of-the-art performance, as long as the class name embeddings for seen and unseen events are maximally separated while retaining the semantic relations between each other. An overview of the proposed approach is shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 372, + 317, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 317, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 317, + 384 + ], + "type": "text", + "content": "3.2 Class Embedding Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "content": "Language-aligned classifiers can achieve zero-shot generalization as class embeddings " + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "inline_equation", + "content": "t_1, \\ldots, t_{|\\mathcal{T}|}" + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "content": " obtained from large-scale models like LLaMa or CLIP, encode semantic relationships between classes. However, while these class embeddings can effectively encode semantics, they are sub-optimal for classification due to a lack of class separability. To obtain well-separated semantic embeddings, we optimize a new set of class embeddings " + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_1^*, \\ldots, \\boldsymbol{w}_{|\\mathcal{T}|}^*" + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "content": " to satisfy two competing constraints: class separability and semantic preservation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 522 + ], + "type": "text", + "content": "Class Separability. Class embeddings that are maximally separated from each other are better suited for recognition of seen classes [43]. Thus, to achieve class embedding separation, we minimize the negative distance of each class embedding to its nearest neighbor" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 241, + 529, + 482, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 529, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 241, + 529, + 482, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e p} = - \\sum_ {c = 1} ^ {| \\tau |} \\min _ {k \\neq c} \\| \\boldsymbol {w} _ {c} - \\boldsymbol {w} _ {k} \\| _ {2}. \\tag {1}", + "image_path": "c93d8906e17dab87c9a9f469428544404d4df1867266e546dc2c2ce1a5d87ce3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "However, maximizing separability alone would destroy the semantics of the pre-trained embeddings. In fact, the minimum of Eq. 1 (if norm constrained) would be achieved for a set of embeddings " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{w}_1, \\dots, \\pmb{w}_{|\\mathcal{T}|}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " following a \"Tammes problem\" configuration [52]. Note that the embeddings are normalized. When the embedding dimension " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is larger than the number of classes in the dictionary " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", the Tammes configuration provides a solution where all embeddings are equally spaced from each other, i.e. " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\| \\pmb{w}_i - \\pmb{w}_j \\|_2 = \\mathrm{Cte} \\quad \\forall i \\neq j" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". For example, the distance between the class \"a person playing guitar\" and the class \"a person" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "S. Mo and P. Morgado" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "playing the violin\" would be the same as the distance to the class \"a dog barking\". Hence, since effective zero-shot recognition heavily relies on the semantic relations between seen and unseen classes, class separability must be balanced with the preservation of the semantics available in pre-trained embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 479, + 200 + ], + "type": "text", + "content": "Semantic preservation To preserve the semantics of the initial embeddings " + }, + { + "bbox": [ + 130, + 164, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i" + }, + { + "bbox": [ + 130, + 164, + 479, + 200 + ], + "type": "text", + "content": ", a simple approach would be to use " + }, + { + "bbox": [ + 130, + 164, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i" + }, + { + "bbox": [ + 130, + 164, + 479, + 200 + ], + "type": "text", + "content": " as proxies and constrain the new embeddings " + }, + { + "bbox": [ + 130, + 164, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\pmb{w}_i" + }, + { + "bbox": [ + 130, + 164, + 479, + 200 + ], + "type": "text", + "content": " to be in their proximity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 256, + 206, + 481, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 206, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 256, + 206, + 481, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e m} ^ {p r o x} = \\sum_ {c = 1} ^ {| \\mathcal {T} |} \\| \\boldsymbol {w} _ {c} - \\boldsymbol {t} _ {c} \\| _ {2}. \\tag {2}", + "image_path": "e15f934dab7137fd1117f588b3cbec24b5ca56d2932df873e47a568ed687d597.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 243, + 479, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 479, + 291 + ], + "type": "text", + "content": "However, this constraint would significantly limit the separability of class embeddings " + }, + { + "bbox": [ + 131, + 243, + 479, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 131, + 243, + 479, + 291 + ], + "type": "text", + "content": ", given their direct ties to " + }, + { + "bbox": [ + 131, + 243, + 479, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 131, + 243, + 479, + 291 + ], + "type": "text", + "content": ". Instead, we seek new class embeddings " + }, + { + "bbox": [ + 131, + 243, + 479, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 131, + 243, + 479, + 291 + ], + "type": "text", + "content": " that retain a similar ranking of similarities to other classes. This is accomplished using a margin ranking loss of the form" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 297, + 481, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 297, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 176, + 297, + 481, + 339 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e m} ^ {r a n k} = \\sum_ {c = 1} ^ {| \\mathcal {T} |} \\sum_ {\\substack {i = 1 \\\\ c \\neq i \\neq j}} ^ {| \\mathcal {T} |} \\sum_ {j = 1} ^ {| \\mathcal {T} |} \\max \\left\\{0, m - \\operatorname {sign} \\left(d _ {c i} ^ {\\boldsymbol {t}} - d _ {c j} ^ {\\boldsymbol {t}}\\right) \\left(d _ {c i} ^ {\\boldsymbol {w}} - d _ {c j} ^ {\\boldsymbol {w}}\\right) \\right\\} \\tag{3}", + "image_path": "2380abdb6975f42b10fe63d74ac8f5308dc76f8f338f7869fcbb182d8a552577.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " is a margin coefficient, and " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "d_{ci}^{\\pmb{t}} = \\| \\pmb{t}_{c} - \\pmb{t}_{i}\\|_{2}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "d_{ci}^{\\pmb{w}} = \\| \\pmb{w}_{c} - \\pmb{w}_{i}\\|_{2}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " distances between pairs of class embeddings, using either the pre-trained embeddings " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " or the new optimized embeddings " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": ", respectively. Intuitively, this ranking loss will seek to organize the new embeddings " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " such that if " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{c}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " is closer to " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{i}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " than to " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{j}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{c}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " will also be closer to " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{i}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": " than to " + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{j}" + }, + { + "bbox": [ + 130, + 345, + 480, + 452 + ], + "type": "text", + "content": ". Note that Eq. 3's complexity grows cubically with the number of classes. While manageable for the tested datasets, it can be impractical for large vocabularies. A potential solution is to sample a large (but fixed) number of class triplets per iteration, and replace Eq. 3 with an expectation over these triplets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 479, + 475 + ], + "type": "text", + "content": "To balance separability and semantic preservation, we obtain the class embeddings by minimizing the joint loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 241, + 483, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 483, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 241, + 483, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {w} _ {1}, \\dots , \\boldsymbol {w} _ {| \\mathcal {T} |}} {\\arg \\min } (1 - \\alpha) \\mathcal {L} _ {s e m} + \\alpha \\mathcal {L} _ {s e p}, \\tag {4}", + "image_path": "f8aef293d34ac4d826a094e214b63610cef249d7cb702464279e51b23204c4c2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 507, + 479, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 507, + 479, + 532 + ], + "spans": [ + { + "bbox": [ + 131, + 507, + 479, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 507, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 507, + 479, + 532 + ], + "type": "text", + "content": " is set to 0.5. Given that the objective is differentiable, we optimize the new embeddings using gradient descent, starting from the initialization " + }, + { + "bbox": [ + 131, + 507, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{w}_i = \\pmb{t}_i\\forall i" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 546, + 391, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 391, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 391, + 558 + ], + "type": "text", + "content": "3.3 Supervised Audio-Visual Language Alignment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "text", + "content": "Given the enhanced class embeddings " + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{w}_1, \\dots, \\pmb{w}_{|\\mathcal{T}|}" + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "text", + "content": " (with improved class separability), we empirically show that a contrastive objective is sufficient to learn state-of-the-art models for generalized zero-shot learning of audio-visual events. Specifically, let " + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{av} = f_\\theta^{av}(\\pmb{v}_i, \\pmb{a}_i)" + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "text", + "content": " be the multi-modal representation of an audio-visual sample " + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "inline_equation", + "content": "(v_i, a_i)" + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "text", + "content": " belonging to class " + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "text", + "content": ". Then, we align representations " + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{av}" + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "text", + "content": " with the corresponding class embeddings " + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{y_i}" + }, + { + "bbox": [ + 130, + 563, + 479, + 634 + ], + "type": "text", + "content": " by minimizing the contrastive loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 217, + 641, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 641, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 217, + 641, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A V L A} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {x} _ {i} ^ {a v} , \\boldsymbol {w} _ {y _ {i}}\\right)\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {x} _ {i} ^ {a v}, \\boldsymbol {w} _ {y _ {k}}\\right)\\right)} \\tag {5}", + "image_path": "be05e85a7b2275aaa8ceb1392f10eed4cff0b0a55649bdc9ca110e2ebd4c056d.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is a similarity function used to contrast the class embedding " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{y_i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " against others in the current batch of data " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{y_k}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "text", + "content": "Similarity. While contrastive objectives typically rely on cosine similarity metrics for " + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\sin (\\cdot ,\\cdot):\\mathcal{X}\\times \\mathcal{T}\\rightarrow \\Re" + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "text", + "content": ", we found that non-linear similarity functions can better capture the relationship between audio-visual representations and class embeddings. Specifically, we draw inspiration from transformer attention modules and use a multi-head cross-attention layer. In this layer, class embeddings " + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\pmb{w}_c" + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "text", + "content": " are used as queries and audio-visual representations " + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{av}" + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "text", + "content": " as key-value pairs. After cross-attention, class similarities are computed through a linear projection " + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "text", + "content": ". Optimizing Eq. 5 pushes the learned embeddings closer to the text embeddings of the same category. During inference, we use the cross-attention transformer layer to output " + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "text", + "content": " similarity scores corresponding to each category, including unseen ones. We then classify the audio-visual samples according to the class with highest similarity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 230, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 230, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 230, + 313 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 322, + 482, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 359 + ], + "type": "text", + "content": "In this section, we evaluate the proposed EZ-AVGZL on three benchmarks for audio-visual generalized zero-shot learning, including VGGSound-GZSL [28], UCF-GZSL [28], and ActivityNet-GZSL [28]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 373, + 262, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 262, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 262, + 385 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 390, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 510 + ], + "type": "text", + "content": "Datasets. VGGSound-GZSL [28] contains 93,752 videos across 276 classes, selected from VGGSound [7], a large audio-visual dataset of 309 classes over 200k videos. We follow the prior work [27, 28] and use the same split for train/val(U)/test(U) over 138/69/69 classes. UCF-GZSL [28] includes 6,816 videos from 51 classes, chosen from UCF101 [51], the original video action recognition dataset, which consists of over 13k videos in 101 classes. We use the same train/val(U)/test(U) split over 30/12/9 classes in previous work [27, 28]. ActivityNet-GZSL [28] contains 20k videos in 200 classes of varying duration for action recognition [13]. We use the same train/val(U)/test(U) split over 99/51/50 classes as in the prior work [27, 28]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "content": "Evaluation Metrics. In line with prior work [27, 28], we use the mean class accuracy to evaluate all models. For ZSL performance, we evaluate only on the subset of test samples from unseen classes. For the GZSL performance, we evaluate on the full test set, which includes both seen and unseen classes. We report results for the subsets of seen (S) and unseen (U) classes, as well as their harmonic mean." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Implementation. We follow previous work [27] and use 8 heads with a dimension of 64 per head for all attention blocks in the cross-attention transformer. The model is trained for 50 epochs with a batch size of 64, using the Adam optimizer [20] with a learning rate of 1e-4, running average coefficients " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", and a weight decay of 1e-5. Following the prior work [27, 28], we use the word2vec [29] embeddings for text, and VGGish [15] embeddings for audio, and C3D [56] features for video frames." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "S. Mo and P. Morgado" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 204, + 137, + 413, + 250 + ], + "blocks": [ + { + "bbox": [ + 165, + 114, + 446, + 126 + ], + "lines": [ + { + "bbox": [ + 165, + 114, + 446, + 126 + ], + "spans": [ + { + "bbox": [ + 165, + 114, + 446, + 126 + ], + "type": "text", + "content": "Table 1: Quantitative results of VGGSound-GZSL benchmark." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 204, + 137, + 413, + 250 + ], + "lines": [ + { + "bbox": [ + 204, + 137, + 413, + 250 + ], + "spans": [ + { + "bbox": [ + 204, + 137, + 413, + 250 + ], + "type": "table", + "html": "
MethodSeenUnseenHarmonicMeanZSL
AttentionFusion [9]14.133.004.953.37
Perceiver [19]13.253.034.933.44
CJME [48]10.862.223.683.72
AVGZSLNet [26]15.023.195.264.81
AVCA [28]12.636.198.316.91
TCaF [27]12.636.728.777.41
Hyper-alignment [16]12.506.448.507.25
Hyper-single [16]12.565.037.185.47
Hyper-multiple [16]15.626.008.677.31
EZ-AVGZL (ours)17.268.6811.558.92
", + "image_path": "f01bebe3f4f8f105b14790dac24b5e379241fdabd6906e08f73ee0136dfc5065.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 204, + 285, + 413, + 396 + ], + "blocks": [ + { + "bbox": [ + 179, + 262, + 432, + 274 + ], + "lines": [ + { + "bbox": [ + 179, + 262, + 432, + 274 + ], + "spans": [ + { + "bbox": [ + 179, + 262, + 432, + 274 + ], + "type": "text", + "content": "Table 2: Quantitative results of UCF-GZSL benchmark." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 204, + 285, + 413, + 396 + ], + "lines": [ + { + "bbox": [ + 204, + 285, + 413, + 396 + ], + "spans": [ + { + "bbox": [ + 204, + 285, + 413, + 396 + ], + "type": "table", + "html": "
MethodSeenUnseenHarmonic MeanZSL
AttentionFusion [9]39.3418.2924.9720.21
Perceiver [19]46.8526.8234.1128.12
CJME [48]33.8924.8228.6529.01
AVGZSLNet [26]74.7924.1536.5131.51
AVCA [28]63.1530.7241.3437.72
TCaF [27]67.1440.8350.7844.64
Hyper-alignment [16]57.1333.8642.5239.80
Hyper-single [16]63.4734.8544.9939.86
Hyper-multiple [16]74.2635.7948.3052.11
EZ-AVGZL (ours)78.3246.3558.2448.27
", + "image_path": "a0eaa7b65001b6431d141dae715c0ec2d9eed8bf8f723b75fb4750a959de35bc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 421, + 296, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 421, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 296, + 433 + ], + "type": "text", + "content": "4.2 Comparison to Prior Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 509 + ], + "type": "text", + "content": "In this work, we propose a novel and effective framework for audio-visual generalized zero-shot learning. In order to demonstrate the effectiveness of the proposed EZ-AVGZL, we comprehensively compare it to previous audio-visual baselines [9, 19] adapted to our problem and current state-of-the-art audio-visual generalized zero-shot learning approaches [16, 26-28, 48]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "For the VGGSound-GZSL dataset, we report the quantitative comparison results in Table 1. As shown, we achieve the best results in all metrics for both seen and unseen classes compared to previous audio-visual generalized zero-shot learning approaches. In particular, the proposed EZ-AVGZL outperforms Hyper-multiple [16], the current state-of-the-art audio-visual generalized zero-shot learning baseline, by 1.64@Seen, 2.68@Unseen, 2.88@Harmonic Mean, and 1.61@ZSL. Moreover, we also perform better than TCaF [27], a strong audio-visual generalized zero-shot learning baseline, by 4.63@Seen, 1.96@Unseen, 2.78@Harmonic Mean, and 1.51@ZSL. Furthermore, we achieve significant performance gains compared to AVCA [28], the first baseline using cross-modal attention and textual label embeddings. This indicates the importance of optimized text transformation and non-linear similarity functions in learning better transferrable knowledge for generalized zero-shot learning." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 204, + 137, + 413, + 248 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 447, + 126 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 447, + 126 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 447, + 126 + ], + "type": "text", + "content": "Table 3: Quantitative results of ActivityNet-GZSL benchmark." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 204, + 137, + 413, + 248 + ], + "lines": [ + { + "bbox": [ + 204, + 137, + 413, + 248 + ], + "spans": [ + { + "bbox": [ + 204, + 137, + 413, + 248 + ], + "type": "table", + "html": "
MethodSeenUnseenHarmonic MeanZSL
AttentionFusion [9]11.153.375.184.88
Perceiver [19]18.254.276.924.47
CJME [48]10.755.557.326.29
AVGZSLNet [26]13.705.968.306.39
AVCA [28]16.777.049.927.58
TCaF [27]30.127.6512.207.96
Hyper-alignment [16]29.778.7713.559.13
Hyper-single [16]24.6110.1014.3210.37
Hyper-multiple [16]36.989.6015.2510.39
EZ-AVGZL (ours)38.5212.8519.2713.01
", + "image_path": "bfdd0eb3b4ff791d384782ef7f5343748031a76b7a0152aae5d98d8f2d1ff55a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 284, + 305 + ], + "lines": [ + { + "bbox": [ + 130, + 259, + 284, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 284, + 305 + ], + "type": "text", + "content": "Table 4: Ablation studies on Class Embedding Optimization and Audio-Visual-Language Alignment." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 134, + 316, + 284, + 360 + ], + "blocks": [ + { + "bbox": [ + 134, + 316, + 284, + 360 + ], + "lines": [ + { + "bbox": [ + 134, + 316, + 284, + 360 + ], + "spans": [ + { + "bbox": [ + 134, + 316, + 284, + 360 + ], + "type": "table", + "html": "
CEPAVLASeenUnseenHarmonic MeanZSL
XX12.636.728.777.41
✓X15.357.9210.458.36
X✓14.677.239.697.91
✓✓17.268.6811.558.92
", + "image_path": "567d73740cd3bc2f1655377c221faf81b15be485d8c5234ea48cec6485e1743d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 290, + 295, + 481, + 357 + ], + "blocks": [ + { + "bbox": [ + 287, + 260, + 481, + 285 + ], + "lines": [ + { + "bbox": [ + 287, + 260, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 287, + 260, + 481, + 285 + ], + "type": "text", + "content": "Table 5: Ablation studies on optimization objectives in Class Embedding Optimization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 290, + 295, + 481, + 357 + ], + "lines": [ + { + "bbox": [ + 290, + 295, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 290, + 295, + 481, + 357 + ], + "type": "table", + "html": "
LsemLsepSeenUnseenHarmonic MeanZSL
XX14.677.239.327.91
LproxsemX15.757.8310.468.26
X✓16.138.1710.858.45
Lproxsem✓17.268.6811.558.92
Lranksem✓18.0510.2813.1010.36
", + "image_path": "5e4da7ca994a41c6d4fa927a970dffd5354f20fdfcb1f4a7de3cfb878e263d31.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 383, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 515 + ], + "type": "text", + "content": "Additionally, we observe significant gains in the UCF-GZSL and ActivityNet-GZSL benchmarks, as shown in Table 2 and Table 3. Compared to TCaF [27], a strong baseline using three manually-designed objectives, we achieve result gains of 11.18@Seen, 5.52@Unseen, 7.46@Harmonic Mean, and 3.63@ZSL on UCF-GZSL dataset. When evaluated on the challenging ActivityNet-GZSL benchmark, our method still outperforms TCaF [27] by 8.40@Seen, 5.20@Unseen, 7.07@Harmonic Mean, and 5.05@ZSL. We also achieve better results than AVCA [28], a cross-modal baseline based on cross-modal attention between audio and visual embeddings. These results demonstrate the effectiveness of our approach in learning optimized textual label embeddings and non-linear similarity functions for generalized audio-visual zero-shot recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 534, + 275, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 534, + 275, + 547 + ], + "spans": [ + { + "bbox": [ + 131, + 534, + 275, + 547 + ], + "type": "text", + "content": "4.3 Experimental Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 616 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to demonstrate the benefit of introducing the Class Embedding Optimization (CEO) and Audio-Visual Language Alignment (AVLA) modules. We also carry out extensive experiments to investigate optimization objectives in class embedding optimization, non-linear similarity functions, and generalization across diverse cross-modal settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 665 + ], + "type": "text", + "content": "Class Embedding Optimization (CEO) & Audio-Visual Language Alignment (AVLA). In order to demonstrate the effectiveness of CEO and AVLA, we conduct an ablation study and report the quantitative results on VGGSound-GZSL dataset in Table 4." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 260, + 102 + ], + "type": "text", + "content": "S. Mo and P. Morgado" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "As observed, adding CEO to the vanilla baseline improves the results by 2.72@Seen, 1.20@Unseen, 1.68@Harmonic Mean, and 0.95@ZSL, which validates the benefit of optimizing class embeddings so as to obtain separated embeddings that preserve the semantic relationships between classes. On the other hand, introducing AVLA in the baseline increases the zero-shot performance across all metrics. More importantly, incorporating both CEP and AVLA into the baseline significantly raises the performance by 4.63@Seen, 1.96@Unseen, 2.78@Harmonic Mean, and 1.51@ZSL. These improvements validate the importance of text transformation and supervised text audio-visual contrastive in extracting separated textual label embeddings for AVGZSL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 240, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 469 + ], + "type": "text", + "content": "Objectives in Class Embedding Optimization. Optimization objectives in class embedding transformation, including global semantic preservation and class separability, are essential for learning optimized textual class embeddings for zero-shot learning. To explore the effects of each objective more comprehensively, we ablate each text optimization objective and report the quantitative results in Table 5. We observe that adding the semantic preservation loss to the vanilla baseline improves the performance by 1.08@Seen, 0.60@Unseen, 1.14@Harmonic Mean, and 0.35@ZSL. On the other hand, introducing the class separability loss in the baseline enhances the zero-shot performance across all metrics. Furthermore, combining the semantic preservation and class separability objectives for the text transformer significantly increases the performance by 2.59@Seen, 3.05@Unseen, 3.78@Harmonic Mean, and 2.45@ZSL. This result validates the importance of class separability loss in balancing the preservation of the semantics available in pre-trained representations for zero-shot learning. Replacing the Euclidean-based semantic preservation loss " + }, + { + "bbox": [ + 130, + 240, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sem}^{prox}" + }, + { + "bbox": [ + 130, + 240, + 482, + 469 + ], + "type": "text", + "content": " with a ranking based loss " + }, + { + "bbox": [ + 130, + 240, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sem}^{rank}" + }, + { + "bbox": [ + 130, + 240, + 482, + 469 + ], + "type": "text", + "content": " further achieves performance gains of 0.79@Seen, 1.60@Unseen, 1.55@Harmonic Mean, and 1.44@ZSL. These results demonstrate the effectiveness of both semantic preservation and class separability objectives in class embedding optimization to generate meaningful text representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "Non-linear Similarity Functions. The non-linear similarity operator used in supervised audio-visual language alignment impacts the extracted cross-modal representations. To explore these effects more comprehensively, we varied the functions from " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{Cosine - similarity}" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": ", Linear, MLP, Cross-Attention\\}. For Cosine-Similarity, we directly compute the dot product between audio-visual embeddings " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{av}" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " and text embeddings " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": ". For Linear, we apply a linear layer to the concatenation of both embeddings to predict the similarity score. For MLP, we replace the linear layer with a multi-layer perceptron to generate the score. For Cross-Attention, we use text embeddings " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " as the queries and audio-visual representations " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{av}" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " as key-value pairs in a transformer attention module to output the score. The comparison results of zero-shot performance on VGGSound-GZSL benchmark are reported in Table 6. When using Cosine-Similarity as the function, we achieve the worst results across all metrics. Replacing the cosine-similarity operator with both Linear increases the seen results by " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.55@Seen" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": ", but decreases the performance by " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.07@Unseen" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.12@ZSL" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": ". This might be because adding complexity to the function causes the model to overfit the seen classes in the" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 230, + 137, + 386, + 182 + ], + "blocks": [ + { + "bbox": [ + 155, + 114, + 457, + 127 + ], + "lines": [ + { + "bbox": [ + 155, + 114, + 457, + 127 + ], + "spans": [ + { + "bbox": [ + 155, + 114, + 457, + 127 + ], + "type": "text", + "content": "Table 6: Exploration studies on non-linear similarity functions " + }, + { + "bbox": [ + 155, + 114, + 457, + 127 + ], + "type": "inline_equation", + "content": "\\phi \\left( \\cdot \\right)" + }, + { + "bbox": [ + 155, + 114, + 457, + 127 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 230, + 137, + 386, + 182 + ], + "lines": [ + { + "bbox": [ + 230, + 137, + 386, + 182 + ], + "spans": [ + { + "bbox": [ + 230, + 137, + 386, + 182 + ], + "type": "table", + "html": "
φ(·)SeenUnseenHarmonic MeanZSL
Cosine-Similarity13.978.5910.648.67
Linear16.528.5211.248.55
MLP17.038.4611.308.49
Cross-Attention17.268.6811.558.92
", + "image_path": "e8bd875d74028fa1ee00aee93875e722b967c662b2259f144414a7d9727c3b15.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 212, + 229, + 404, + 318 + ], + "blocks": [ + { + "bbox": [ + 130, + 194, + 482, + 218 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 218 + ], + "type": "text", + "content": "Table 7: Exploration studies on diverse cross-modal settings. CEP denotes Class Embedding Optimization." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 229, + 404, + 318 + ], + "lines": [ + { + "bbox": [ + 212, + 229, + 404, + 318 + ], + "spans": [ + { + "bbox": [ + 212, + 229, + 404, + 318 + ], + "type": "table", + "html": "
Text EncoderAudio EncoderVisual EncoderCEPSeenUnseenHarmonic MeanZSL
word2vecVGGishC3D×12.636.728.777.41
✓17.268.6811.558.92
LLaMaAudioMAEDINO×18.3510.6313.4612.26
✓23.9513.3217.1213.73
CLAPCLAPX-CLIP×27.7914.0118.6317.39
✓33.2820.1525.1021.06
ImageBindImageBindImageBind×28.3214.6819.3419.87
✓35.5621.6326.9022.35
", + "image_path": "429b555cd1816445683e7982b0bd9253569a16f2a49a7c99a1390331f077378e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 425 + ], + "type": "text", + "content": "training data, , leading to poorer generalization to unseen categories. Additionally, using MLP as the function operator significantly raises the performance by 3.06@Seen, but slightly deteriorates the results on unseen classes. Our method achieves the best results across all metrics by using the cross-attention operator to predict the similarity score. These improved results demonstrate the effectiveness of the cross-attention operator in balancing the trade-off between linear and non-linear similarity functions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "Generalization to Diverse Cross-modal Settings. To validate the generalizability of the proposed method across a flexible number of feature encoders, we experiment with different settings: unimodal, paired (video-text, audio-text), and unified embedding space (image-audio-text). In the unimodal setting, we use DINO [3] for the visual encoder, AudioMAE [18] for the audio encoder, and LLaMa [55] for the text encoder. In the paired setting, we use X-CLIP [25] pre-trained with video-text objectives for the visual encoder, and CLAP [59] pre-trained with audio-text objectives for the audio and text encoder, respectively. In the unified embedding setting, we apply ImageBind [12] pre-trained with image-audio-text alignment in the same embedding space as the encoder for all three modalities. The quantitative results on the VGGSound-GZSL benchmark are compared in Table 7. As observed, the unimodal setting, which uses transformers with strong unimodal features, significantly outperforms the vanilla baseline that uses convolution networks, where EZ-AVGZL achieves performance gains of 6.69@Seen, 4.64@Unseen, 5.77@Harmonic Mean, and 4.81@ZSL using class embedding optimization. In the paired setting, the proposed method further increases the results by 9.33@Seen, 6.93@Unseen, 7.98@Harmonic Mean, and 7.331@ZSL, benefiting from the text-audio correspondence and text-video alignment pre-trained on large-scale data. In the unified embedding setting using ImageBind [12] embeddings, we achieve the best results across all metrics" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "S. Mo and P. Morgado" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 136, + 446, + 212 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 455, + 127 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 455, + 127 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 455, + 127 + ], + "type": "text", + "content": "Table 8: Comparison of class embeddings w/o and w optimization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 136, + 446, + 212 + ], + "lines": [ + { + "bbox": [ + 168, + 136, + 446, + 212 + ], + "spans": [ + { + "bbox": [ + 168, + 136, + 446, + 212 + ], + "type": "table", + "html": "
base classNN (w/o opt)D (w/o opt)NN (w opt)D (w opt)
alligators, crocodiles hissingchurch bell ringing0.455playing marimba, xylophone0.498
barn swallow callingblack capped chickadee calling0.307black capped chickadee calling0.465
basketball bounceplaying volleyball0.324bouncing on trampoline0.394
bee, wasp, etc. buzzingmotorboat, speedboat acceleration0.322fly, housefly buzzing0.333
bird squawkingpleasant crowing0.400duck quacking0.496
black capped chickadee callingbarn swallow calling0.307magpie calling0.436
bouncing on trampolinebasketball bounce0.416basketball bounce0.494
bowling impactplaying table tennis0.410striking bowling0.467
bull bellowingeagle screaming0.449dog growling0.532
canary callingwhale calling0.403black capped chickadee calling0.513
", + "image_path": "067d722f8caa4b8fce2122c1c1c97e86ab50a7dc6f9f23558a0cf5cd30799d87.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 221, + 462, + 356 + ], + "blocks": [ + { + "bbox": [ + 151, + 221, + 462, + 356 + ], + "lines": [ + { + "bbox": [ + 151, + 221, + 462, + 356 + ], + "spans": [ + { + "bbox": [ + 151, + 221, + 462, + 356 + ], + "type": "image", + "image_path": "a40c9db7660021e7ee79bdb06dc55e75846e784520815d25cc3adc1c518db28b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 365, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 131, + 365, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 482, + 390 + ], + "type": "text", + "content": "Fig. 3: Confusion matrices of zero-shot predictions on VGGSound-GZSL using model with and without class embedding optimization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "type": "text", + "content": "compared to other settings. This might be because text embeddings with unified training on the other two modalities strengthen the alignment for audio-visual zero-shot learning. Furthermore, when class embedding optimization is added to all settings, we improve the baseline based on the initial text embeddings, further demonstrating the effectiveness of the proposed class embedding optimization in generating text representations for AVZSL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "content": "Qualitative Comparisons with and without Optimization. To provide insights into why the proposed loss terms improve the separability of class embeddings while keeping semantics, we conduct qualitative studies. These include comparing the visualization of the class embedding space with and without optimization. The comparisons of class embeddings with and without optimization are shown in Table 8. We report the base class and identify 1) the nearest neighbor (NN) class and 2) the distance (D) to the nearest neighbor. We observe larger distances and similar NN classes. The confusion matrices for class-wise performances of the model are reported in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Impact of Hyper-parameters " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " and Margin " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". To investigate the impact of " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " on the performance thoroughly, we set " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.5, 0.9\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\{0, 0.5, 1\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". The quantitative comparisons on VGGSound-GZSL are reported in Table 9. As can be seen, with the increase of " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " from 0.1 to 0.5, the proposed approach achieves better results regarding all metrics. However, when " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " is increased to 0.9, the performance decreases a lot, which might be caused by not" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 148, + 274, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 273, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 273, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 273, + 137 + ], + "type": "text", + "content": "Table 9: Study of " + }, + { + "bbox": [ + 132, + 114, + 273, + 137 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 114, + 273, + 137 + ], + "type": "text", + "content": " and margin " + }, + { + "bbox": [ + 132, + 114, + 273, + 137 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 132, + 114, + 273, + 137 + ], + "type": "text", + "content": " on VGGSound-GZSL." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 148, + 274, + 200 + ], + "lines": [ + { + "bbox": [ + 133, + 148, + 274, + 200 + ], + "spans": [ + { + "bbox": [ + 133, + 148, + 274, + 200 + ], + "type": "table", + "html": "
αmSeenUnseenHarmonic MeanZSL
0.518.0510.2813.1010.36
0.1116.358.3911.098.58
0.916.268.7311.368.92
0.5017.829.7112.579.67
0.517.679.8512.659.95
", + "image_path": "275b08c55abb8f5e882262c6fef78b1b9f89ede539743eea39b095fa9b0ba5cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 290, + 149, + 482, + 198 + ], + "blocks": [ + { + "bbox": [ + 287, + 115, + 480, + 139 + ], + "lines": [ + { + "bbox": [ + 287, + 115, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 287, + 115, + 480, + 139 + ], + "type": "text", + "content": "Table 10: Study of distance functions on VGGSound-GZSL." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 290, + 149, + 482, + 198 + ], + "lines": [ + { + "bbox": [ + 290, + 149, + 482, + 198 + ], + "spans": [ + { + "bbox": [ + 290, + 149, + 482, + 198 + ], + "type": "table", + "html": "
DistanceSeenUnseenHarmonic MeanZSL
cosine18.0510.2813.1010.36
l217.589.3212.189.27
l117.358.7811.668.85
", + "image_path": "2d5d70319f8b39e970a5f01d21b7048a1982a34c5011cb27052bb9ea5ff697f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 219, + 481, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 481, + 243 + ], + "type": "text", + "content": "much semantics preservation during training. Meanwhile, with the decrease of " + }, + { + "bbox": [ + 130, + 219, + 481, + 243 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 219, + 481, + 243 + ], + "type": "text", + "content": ", we observe a clearly dropping trend in our metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "text", + "content": "Ablating Distance Functions for Eq. 1-2. Distance functions in Eq. 1-2 is crucial for us to maximize separation while preserving semantics. We experimented with different distance functions, " + }, + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "text", + "content": " distances for Eq. 1-2. Note that while, Eq. 1-2 show " + }, + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "text", + "content": " distances in the paper, the embeddings are normalized. The quantitative results are shown in Table 10. We can observe that using the cosine distance function performed the best regarding all metrics. However, when optimizing " + }, + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 243, + 482, + 363 + ], + "type": "text", + "content": " distance functions, we achieve the worst results in terms of all metrics. These ablation results highlight the importance of designing appropriate distance functions in maximizing separation while preserving the semantics of text class embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 379, + 220, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 379, + 220, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 379, + 220, + 392 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": "In this work, we present a novel and effective framework, EZ-AVGZL, that aligns audio-visual representations with optimized text embeddings to achieve audio-visual generalized zero-shot learning. We leverage differential optimization to learn better-separated textual representations with maximal separability and semantic preservation. Additionally, we introduce supervised audio-visual language alignment to learn the correspondence between audio-visual features and textual embeddings, which captures the cross-modal dynamics with class labels as prompts. Experimental results on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL datasets demonstrate our method's superiority against previous baselines. Extensive ablation studies also validate the importance of class embedding optimization and supervised audio-visual language alignment for audio-visual generalized zero-shot learning. We also conduct comprehensive experiments to demonstrate the effectiveness of non-linear similarity functions and generalization to diverse cross-modal settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Broader Impact. This work makes substantial progress in audio-visual generalized zero-shot learning. Our method is novel in its simplicity, demonstrating that a simple supervised contrastive objective, without \"bells and whistles\", outperforms complex alignment strategies. Without our work, progress in this area would likely be driven by increasingly complex methods, which are difficult to implement, reproduce and compare. Beyond the strong results and simplicity of our method, the class embedding optimization procedure is also technically novel to maximize semantic separation while keeping semantics preserved." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "S. Mo and P. Morgado" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "text", + "content": "1. Arandjelovic, R., Zisserman, A.: Look, listen and learn. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 609-617 (2017) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "text", + "content": "2. Aytar, Y., Vondrick, C., Torralba, A.: Soundnet: Learning sound representations from unlabeled video. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2016) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "text", + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 481, + 274 + ], + "type": "text", + "content": "4. Chen, C., Jain, U., Schissler, C., Gari, S.V.A., Al-Halah, Z., Ithapu, V.K., Robinson, P., Grauman, K.: Soundspaces: Audio-visual navigation in 3d environments. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 17-36 (2020) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 276, + 480, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 309 + ], + "type": "text", + "content": "5. Chen, C., Majumder, S., Ziad, A.H., Gao, R., Kumar Ramakrishnan, S., Grauman, K.: Learning to set waypoints for audio-visual navigation. In: Proceedings of International Conference on Learning Representations (ICLR) (2021) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 354 + ], + "type": "text", + "content": "6. Chen, C., Schissler, C., Garg, S., Kobernik, P., Clegg, A., Calamia, P., Batra, D., Robinson, P.W., Grauman, K.: Soundspaces 2.0: A simulation platform for visual-acoustic learning. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track (2022) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 387 + ], + "type": "text", + "content": "7. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 721-725. IEEE (2020) 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "text", + "content": "8. Ephrat, A., Mosseri, I., Lang, O., Dekel, T., Wilson, K., Hassidim, A., Freeman, W.T., Rubinstein, M.: Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619 (2018) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "text", + "content": "9. Fayek, H.M., Kumar, A.: Large scale audiovisual learning of sounds with weakly labeled data. In: Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (2021) 9, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "type": "text", + "content": "0. Gan, C., Huang, D., Zhao, H., Tenenbaum, J.B., Torralba, A.: Music gesture for visual sound separation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10478-10487 (2020) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "text", + "content": "1. Gao, R., Grauman, K.: 2.5d visual sound. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 324-333 (2019) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 564 + ], + "type": "text", + "content": "2. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 4, 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 598 + ], + "type": "text", + "content": "3. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 961-970 (2015) 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "4. Hershey, J., Casey, M.: Audio-visual sound separation via hidden markov models Advances in Neural Information Processing Systems 14 (2001) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "5. Hershey, S., Chaudhuri, S., Ellis, D.P.W., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., Slaney, M., Weiss, R.J., Wilson, K.: Cnn architectures for large-scale audio classification. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2017) 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Hong, J., Hayden, Z., Han, J., Fang, P., Harandi, M., Petersson, L.: Hyperbolic audio-visual zero-shot learning. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023) 4, 9, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "17. Hu, D., Nie, F., Li, X.: Deep multimodal clustering for unsupervised audiovisual learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9248-9257 (2019) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 216 + ], + "type": "text", + "content": "18. Huang, P.Y., Xu, H., Li, J., Baevski, A., Auli, M., Galuba, W., Metze, F., Feichtenhofer, C.: Masked autoencoders that listen. In: Proceedings of Advances In Neural Information Processing Systems (NeurIPS) (2022) 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "text", + "content": "19. Jaegle, A., Gimeno, F., Brock, A., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver: General perception with iterative attention. In: Proceedings of the International Conference on Machine Learning (ICML) (2021) 9, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "21. Korbar, B., Tran, D., Torresani, L.: Cooperative learning of audio and video models from self-supervised synchronization. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2018) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "text", + "content": "22. Lin, Y.B., Li, Y.J., Wang, Y.C.F.: Dual-modality seq2seq network for audio-visual event localization. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2002–2006 (2019) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "text", + "content": "23. Lin, Y.B., Tseng, H.Y., Lee, H.Y., Lin, Y.Y., Yang, M.H.: Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2021) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "24. Lin, Y.B., Wang, Y.C.F.: Audiovisual transformer with instance attention for audiovisual event localization. In: Proceedings of the Asian Conference on Computer Vision (ACCV) (2020) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "25. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP:: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285 (2022) 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "text", + "content": "26. Mazumder, P., Sing, P., Kumar Parida, K., Namboodiri, V.P.: Avgzslnet: Audiovisual generalized zero-shot learning by reconstructing label features from multimodal embeddings. In: Proceedings of 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 3089-3098 (2021) 3, 9, 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "27. Mercea, O.B., Hummel, T., Koepke, A.S., Akata, Z.: Temporal and cross-modal attention for audio-visual zero-shot learning. In: Proceedings of European Conference on Computer Vision (ECCV) (2022) 2, 4, 5, 6, 8, 9, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "text", + "content": "28. Mercea, O.B., Riesch, L., Koepke, A.S., Akata, Z.: Audio-visual generalised zero-shot learning with cross-modal attention and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10553–10563 (2022) 2, 4, 8, 9, 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "29. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. In: Proceedings of International Conference on Learning Representations (ICLR) (2013) 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 631 + ], + "type": "text", + "content": "30. Mo, S., Morgado, P.: Benchmarking weakly-supervised audio-visual sound localization. In: European Conference on Computer Vision (ECCV) Workshop (2022) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "31. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "S. Mo and P. Morgado" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Mo, S., Morgado, P.: A unified audio-visual learning framework for localization, separation, and recognition. In: Proceedings of the International Conference on Machine Learning (ICML) (2023) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "text", + "content": "33. Mo, S., Morgado, P.: Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling. arXiv preprint arXiv:2312.01017 (2023) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 481, + 203 + ], + "type": "text", + "content": "34. Mo, S., Pian, W., Tian, Y.: Class-incremental grouping network for continual audio-visual learning. arXiv preprint arXiv:2309.05281 (2023) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 481, + 224 + ], + "type": "text", + "content": "35. Mo, S., Shi, J., Tian, Y.: DiffAVA: Personalized text-to-audio generation with visual alignment. arXiv preprint arXiv:2305.12903 (2023) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 246 + ], + "type": "text", + "content": "36. Mo, S., Shi, J., Tian, Y.: Text-to-audio generation synchronized with videos. arXiv preprint arXiv:2403.07938 (2024) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 481, + 278 + ], + "type": "text", + "content": "37. Mo, S., Tian, Y.: Multi-modal grouping network for weakly-supervised audiovisual video parsing. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 278, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 481, + 310 + ], + "type": "text", + "content": "38. Mo, S., Tian, Y.: Semantic-aware multi-modal grouping for weakly-supervised audio-visual video parsing. In: European Conference on Computer Vision (ECCV) Workshop (2022) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 310, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 481, + 331 + ], + "type": "text", + "content": "39. Mo, S., Tian, Y.: Audio-visual grouping network for sound localization from mixtures. arXiv preprint arXiv:2303.17056 (2023) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 331, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 353 + ], + "type": "text", + "content": "40. Mo, S., Tian, Y.: AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 353, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 481, + 374 + ], + "type": "text", + "content": "41. Mo, S., Tian, Y.: Semantic grouping network for audio source separation. arXiv preprint arXiv:2407.03736 (2024) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "type": "text", + "content": "42. Morgado, P., Li, Y., Vasconcelos, N.: Learning representations from audio-visual spatial alignment. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS). pp. 4733-4744 (2020) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 407, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 438 + ], + "type": "text", + "content": "43. Morgado, P., Li, Y., Costa Pereira, J., Saberian, M., Vasconcelos, N.: Deep hashing with hash-consistent large margin proxy embeddings. International Journal of Computer Vision (2020) 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 438, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 481, + 471 + ], + "type": "text", + "content": "44. Morgado, P., Misra, I., Vasconcelos, N.: Robust audio-visual instance discrimination. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12934–12945 (2021) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 471, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 481, + 503 + ], + "type": "text", + "content": "45. Morgado, P., Vasconcelos, N., Langlois, T., Wang, O.: Self-supervised generation of spatial audio for 360 video. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2018) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 503, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 481, + 536 + ], + "type": "text", + "content": "46. Morgado, P., Vasconcelos, N., Misra, I.: Audio-visual instance discrimination with cross-modal agreement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12475–12486 (June 2021) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 567 + ], + "type": "text", + "content": "47. Owens, A., Wu, J., McDermott, J.H., Freeman, W.T., Torralba, A.: Ambient sound provides supervision for visual learning. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 801-816 (2016) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 567, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 611 + ], + "type": "text", + "content": "48. Parida, K.K., Matiyali, N., Guha, T., Sharma, G.: Coordinated joint multimodal embeddings for generalized audio-visual zero-shot classification and retrieval of videos. In: Proceedings of 2020 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 3240–3249 (2020) 3, 9, 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 633 + ], + "type": "text", + "content": "49. Pian, W., Mo, S., Guo, Y., Tian, Y.: Audio-visual class-incremental learning. arXiv preprint arXiv:2308.11073 (2023) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "type": "text", + "content": "50. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4358-4366 (2018) 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "text", + "content": "Audio-visual Generalized Zero-shot Learning the Easy Way" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 599 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "51. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "type": "text", + "content": "52. Tammes, P.M.L.: On the origin of number and arrangement of the places of exit on the surface of pollen-grains. Recueil des travaux botaniques neerlandais 27(1), 1-84 (1930) 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "53. Tian, Y., Li, D., Xu, C.: Unified multisensory perception: Weakly-supervised audiovisual video parsing. In: Proceedings of European Conference on Computer Vision (ECCV). p. 436-454 (2020) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "54. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of European Conference on Computer Vision (ECCV) (2018) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 369 + ], + "type": "text", + "content": "55. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "56. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2015) 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "57. Wu, Y., Yang, Y.: Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1326–1335 (2021) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "58. Wu, Y., Zhu, L., Yan, Y., Yang, Y.: Dual attention matching for audio-visual event localization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 6291–6299 (2019) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "text", + "content": "59. Wu, Y., Chen, K., Zhang, T., Hui, Y., Berg-Kirkpatrick, T., Dubnov, S.: Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In: IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP (2023) 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "60. Zhang, L., Mo, S., Zhang, Y., Morgado, P.: Audio-synchronized visual animation arXiv preprint arXiv:2403.05659 (2024) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "61. Zhao, H., Gan, C., Ma, W.C., Torralba, A.: The sound of motions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1735-1744 (2019) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "62. Zhao, H., Gan, C., Rouditchenko, A., Vondrick, C., McDermott, J., Torralba, A.: The sound of pixels. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 570-586 (2018) 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "text", + "content": "S. Mo and P. Morgado" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_content_list.json b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0bb9c93a08167036cba5bae1075c59be1b4af829 --- /dev/null +++ b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_content_list.json @@ -0,0 +1,1712 @@ +[ + { + "type": "text", + "text": "AugDETR: Improving Multi-scale Learning for Detection Transformer", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinpeng Dong, Yutong Lin, Chen Li, Sanping Zhou, Nanning Zheng*", + "bbox": [ + 251, + 212, + 746, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University {djp1235a, yutonglin, edward82}@stu.xjtu.edu.cn {spzhou, nnzheng} $@$ xjtu.edu.cn", + "bbox": [ + 220, + 239, + 781, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current end-to-end detectors typically exploit transformers to detect objects and show promising performance. Among them, Deformable DETR is a representative paradigm that effectively exploits multi-scale features. However, small local receptive fields and limited query-encoder interactions weaken multi-scale learning. In this paper, we analyze local feature enhancement and multi-level encoder exploitation for improved multi-scale learning and construct a novel detection transformer detector named Augmented DETR (AugDETR) to realize them. Specifically, AugDETR consists of two components: Hybrid Attention Encoder and Encoder-Mixing Cross-Attention. Hybrid Attention Encoder enlarges the receptive field of the deformable encoder and introduces global context features to enhance feature representation. Encoder-Mixing Cross-Attention adaptively leverages multi-level encoders based on query features for more discriminative object features and faster convergence. By combining AugDETR with DETR-based detectors such as DINO, AlignDETR, DDQ, our models achieve performance improvements of 1.2, 1.1, and 1.0 AP in the COCO under the ResNet-50-4scale and 12 epochs setting, respectively.", + "bbox": [ + 261, + 345, + 738, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Object detection $\\cdot$ Detection transformer $\\cdot$ Hybrid attention $\\cdot$ Multi-level encoder", + "bbox": [ + 261, + 609, + 740, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 662, + 375, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many CNN-based detectors [1,12,20,27,34,37,47] have been proposed to improve detection performance in the past decade. Despite their promising progress, these detectors suffer from many hand-crafted components, such as non-maximum suppression and anchor design. In contrast, CNN-based paradigm is broken by DETR [4], a novel transformer detector. It gets rid of hand-crafted components and implements an end-to-end paradigm. Despite its novel paradigm, DETR suffers from some limitations, including slow convergence and inferior performance. To address these limitations, many methods [2, 6, 14, 24, 25, 29, 42, 50]", + "bbox": [ + 212, + 694, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 218, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have been proposed from different perspectives, such as introducing spatial priors, hybrid matching training, designing efficient attention modules, assigning sample weights, etc. With these optimizations, DINO [42] achieved a new record on the COCO benchmark [22].", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although DETR-based detectors have achieved promising performance, multiscale object detection has not received sufficient attention so far, which may limit the performance of detectors. Scale variation of objects remains one of the crucial challenges in object detection. Objects at different scales require inconsistent semantic features, e.g., large objects require features with larger receptive fields. To the best of our knowledge, Deformable-DETR [50] has made some effort to remedy this issue. Deformable-DETR uses multi-scale backbone features to improve multi-scale learning by deformable attention. In addition to the use of multi-scale backbone features, we believe that the multi-scale learning of Deformable-DETR has not been fully exploited. As shown in Table 1, although Deformable DETR is 2.5 AP higher than the DETR in overall performance, the performance on large objects is still 1.5 AP lower than the DETR. We present two techniques that can further improve multi-scale learning as follows:", + "bbox": [ + 212, + 208, + 789, + 405 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/b113959c313a5d2b507d5191bd488dd2c01065f154ada899f3063d6a55bae984.jpg", + "table_caption": [ + "Table 1: Performance comparison between DETR and Deformable DETR." + ], + "table_footnote": [], + "table_body": "
MethodEpochsAPAPSAPMAPL
DETR [4]50042.020.545.861.1
Deformable DETR [50]5044.527.147.659.6
", + "bbox": [ + 227, + 488, + 522, + 550 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fb497dd95c6ea9c3b931e0bed1014de3ee7a801f9a8ee6cc6daa3c33284e6d5d.jpg", + "image_caption": [ + "Fig. 1: Comparison of receptive fields of different encoders." + ], + "image_footnote": [], + "bbox": [ + 566, + 473, + 661, + 564 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/525e51d10ad7af2bec6a9d5e98da01acc439e4779e73eabc0ce5f33971791407.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 473, + 785, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Local feature enhancement. Deformable-DETR [50] replaces the original self-attention (global dense attention) with deformable attention (sparse attention) to reduce the complexity and slow convergence of dense attention. Deformable attention is implemented by a fixed number of deformable sampling points. Limited by local sampling points, deformable attention is difficult to achieve a larger receptive field compared to self-attention. This is not conducive to the detection of large objects. Further stacking of deformable attention still improves performance, as analyzed in the ablations. In addition, deformable attention loses the ability to perceive global context, while multi-scale object detection in some complex scenes may rely on global context information. Increasing the sampling points degrades the performance of the deformable attention, as analyzed in the ablations. Therefore, we introduce global contextual information into the deformable encoder to enhance multi-scale learning.", + "bbox": [ + 212, + 595, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-level encoder exploitation. DETR-based detectors typically stack six encoder layers and six decoder layers. Encoder layers enhance the backbone features, and decoder layers extract features from last encoder layer to decode", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Dong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "queries. Previous methods ignore the use of multi-level encoder features. More discriminative object features can be obtained by fully exploiting multi-level encoder features with different receptive fields and feature preferences. We refer to RepLKNet [7] for analysis of receptive field. As shown in Figure 1, the comparison of encoder 1 and 6 in DINO [42] shows that different encoders have varying feature preferences and receptive fields. In addition, the fact that multilevel encoders are all involved in object feature generation can speed up convergence because the supervisory signals are available to multiple levels of encoders, rather than being available only to the last level encoder. Utilizing multi-level encoder features also can introduce redundancy to mitigate the risk of overfitting. Although multi-level encoder features is beneficial for object detection, using multi-level encoder features fixed for all objects is not conducive to multiscale learning. Since objects at different scales have different feature preferences, adaptive use of multi-level encoder features by learning weights based on query features will facilitate multi-scale learning for DETR-based detectors.", + "bbox": [ + 212, + 146, + 787, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose AugDETR, a simple yet effective DETR-based detector that uses two different components to improve multi-scale learning. First, Hybrid Attention Encoder is proposed to augment local features in the deformable encoder by partially combining dense attention with deformable attention. Hybrid Attention Encoder can enlarge the receptive field and introduce global context, which can improve detection performance on large objects. Second, Encoder-Mixing Cross-Attention is introduced to better exploit features from different encoder layers by weighted aggregation. Objects at different scales can learn different aggregation weights based on object features to facilitate multi-scale learning. With this design, Encoder-Mixing Cross-Attention can generate more discriminative object features for subsequent classification and localization tasks. We validate the effectiveness of AugDETR on the COCO benchmark. Our method achieves 50.2 AP (+1.2 AP), 51.3 AP (+1.1 AP), and 51.8 AP (+1.0 AP) based on DINO [42], AlignDETR [2], and DDQ [45] with ResNet-50-4scale under 12 epochs settings, respectively.", + "bbox": [ + 212, + 373, + 787, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 240, + 599, + 557, + 613 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We analyze local feature enhancement and multi-level encoder exploitation techniques for improved multi-scale learning.", + "- A new detection transformer detector named AugDETR is proposed to achieve the techniques with Hybrid Attention Encoder and Encoder-Mixing Cross-Attention.", + "- We validate AugDETR equipped with various DETR-based detectors and backbones on the COCO benchmark, and it consistently yields significant performance improvements over DETR-based detectors." + ], + "bbox": [ + 225, + 619, + 784, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 758, + 387, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Object Detection", + "text_level": 1, + "bbox": [ + 215, + 787, + 405, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early deep-learning detectors can be roughly classified into anchor-based and anchor-free methods. Anchor-based methods use predefined anchors to locate", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "AugDETR", + "bbox": [ + 656, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "objects, including Faster R-CNN [34], SSD [27], RetinaNet [21], and their variants [1, 13, 31, 44]. Anchor-free methods use points near object centers to locate objects, including CenterNet [47], FCOS [37], and others [9, 16, 17, 48, 49].", + "bbox": [ + 212, + 145, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, DETR [4] proposes a simple transformer pipeline for object detection, effectively alleviating the need for heuristic components. Compared to CNN-based work, DETR is rather novel but still suffers from some limitations, including slow convergence and inferior performance. Many follow-up works focus on alleviating the limitations of DETR from different perspectives. Conditional DETR [29] proposes to decouple the context embedding and position embedding, which helps DETR speed up convergence. Anchor DETR [38] uses the anchor points as the initial queries for fast convergence. DAB DETR [24] uses 4D anchor boxes that are dynamically updated with the decoder layer to represent queries. DN DETR [18] proposes a denoising training mechanism to speed up convergence. DINO [42] further extends DN DETR with contrastive denoising training, mixed query, and look-forward twice techniques. Group DETR [5], HDETR [14], and CO-DETR [51] propose various one-to-one matching and one-to-many matching combination methods to stabilize the matching process of DETR. Align DETR [2], Stable DINO [25], and Rank DETR [32] design different schemes to weight the samples by the IoU-aware classification loss. Plain DETR [23] proposes a box-to-pixel relative position bias and a masked image modeling [39] pre-training scheme to improve the DETR detector. Unlike these methods that design spatial priors, hybrid matching, sample weight assignments, etc., our approach focuses on improving the multi-scale learning of DETR detectors. In addition, we conducted experiments using some SOTA methods as the baseline models. Our techniques are complementary to theirs.", + "bbox": [ + 212, + 191, + 787, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Multi-Scale Learning for Object Detection", + "text_level": 1, + "bbox": [ + 214, + 544, + 611, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the early era, image pyramid is a common solution to improve multi-scale learning. Instead of image pyramid, some methods utilize multi-scale features to mitigate scale variation. A representative work among these methods is the feature pyramid network (FPN [20]), which proposes a top-down pathway and lateral connections to construct the feature pyramid. Inspired by FPN, multi-scale feature extraction has been widely studied, such as BiFPN [36], PAFPN [26], NAS-FPN [11], AugFPN [12], and other variants [8,15,33]. Recently, some works have attempted to design multi-scale feature extraction modules in DETR. Deformable DETR [50] proposes deformable attention to utilize multi-scale backbone features. Dynamic DETR [6] proposes a CNN-based dynamic encoder with scale-aware attention, spatial-aware attention, and task-aware attention. SMCA [10] generates the Gaussian-like weight map as the spatial prior and the scale weights as the scale prior to modulated multi-scale features. IMFA [40] designs a new scheme to adaptively sample sparse multi-scale features. Different from them, our approach enhances multi-scale learning by introducing global attention to expand the receptive field and integrate global context. To use global attention effectively, we propose hybrid scale and hybrid layer strategies to construct the hybrid attention encoder. In addition, we adaptively use multi-level", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Dong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "encoder features based on query features to improve multi-scale learning, which is also distinctly different from previous methods.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6895ecd5450bb6c6fd1c867f6df6a2f3535d04fac7a5c463ac50becb05f28ad1.jpg", + "image_caption": [ + "Fig. 2: Left: The pipeline of conventional DETR-based detectors. The backbone network and the stacked encoder layers are used to extract and enhance the image features. Then, the object queries utilize the decoder layers to interact with the enhanced image features. Finally, the interacted object queries are fed into the detection head to obtain the predictions. Only the features from the last encoder layer can be used to interact with the object queries. Right: The pipeline of Our AugDETR. Our AugDETR restructures the conventional encoder-decoder pipeline. The object queries can interact with features from all encoder layers. For clarity, only the 3-layer structure is shown." + ], + "image_footnote": [], + "bbox": [ + 240, + 204, + 491, + 377 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af94f65b0cfca1973e3f69f84960b73ca9b669627c87dbd622f987dedaf472b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 202, + 767, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 553, + 330, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In section 3.1, we first briefly revisit the detection transformer and then introduce the pipeline of AugDETR. In section 3.2, we describe the details of the proposed Hybrid Attention Encoder. In section 3.3, we introduce the Encoder-Mixing Cross-Attention used in the transformer decoder.", + "bbox": [ + 212, + 583, + 787, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 A Revisit of DETR-based Detector", + "text_level": 1, + "bbox": [ + 214, + 664, + 552, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our proposed method is constructed on top of the recent DETR-based object detectors, we first briefly review the pipeline of the DETR-based detectors, taking the Deformable DETR [50] as an example.", + "bbox": [ + 212, + 688, + 785, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pipeline of Deformable DETR is shown in Figure 2 (left). Deformable DETR consists of 3 subcomponents: a backbone, a encoder, and a decoder. Given an input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , a backbone network generates multi-scale features. We denote the multi-scale features as $\\{C_2, C_3, C_4, C_5\\}$ . The spatial resolutions of these features are usually $1/4^2$ , $1/8^2$ , $1/16^2$ , and $1/32^2$ of the given image. Then these features reduced to the same channel by a linear projection are denoted as $\\{P_2, P_3, P_4, P_5\\}$ . Next, a transformer encoder is used to further enhance these", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "AugDETR", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "multi-scale features. The transformer encoder is typically stacked with 6 encoder layers. The encoder layer mainly consists of a self-attention and a feed-forward network. For Deformable DETR, multi-scale deformable attention is employed as self-attention to enhance multi-scale features. The encoded features of the different layers in the encoder are denoted as $\\{E_1,E_2,E_3,E_4,E_5,E_6\\}$ . Finally, a transformer decoder is used to produce object predictions from a set of object queries. Typically, only the last level of encoder features are fed into the decoder to interact with the query. The transformer decoder is also typically stacked with 6 decoder layers for better detection performance. The decoder layer mainly consists of a self-attention, a cross-attention, and a feed-forward network. For deformable DETR, multi-scale deformable attention is used as cross-attention to achieve query-feature interaction. As shown in Figure 2 (right), our pipeline changes the fact that the decoder can utilize only the last level of encoder features, but rather all encoder features can be utilized by the decoder. This design lays the foundation for cross-attention to utilize multi-level encoder features.", + "bbox": [ + 212, + 146, + 787, + 373 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da531c7b14acbdc100eff6e9ddeff03fc1cce099a8797bc45fd719259127b523.jpg", + "image_caption": [ + "Fig. 3: (a) is the process of hybrid attention in the Hybrid Attention Layer. Dense attention is first applied to only the top features, and then sparse attention is applied to the multi-scale features. (b) is the details of Encoder-Mixing Cross-Attention. Deformable attention extracts multiple object features from multi-level encoder and then the extracted features are aggregated based on the adaptive fusion weights learned from the query features." + ], + "image_footnote": [], + "bbox": [ + 225, + 400, + 495, + 491 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9296197508af022d0cdbf0a2fea8b4c9ea6db2b72940e377c5b285df478185cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 398, + 777, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Hybrid Attention Encoder", + "text_level": 1, + "bbox": [ + 215, + 638, + 480, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The encoder layer in the original DETR [4] uses the self-attention as follows:", + "bbox": [ + 215, + 662, + 784, + 679 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = W _ {q} P _ {5}, \\mathbf {K} = W _ {k} P _ {5}, \\mathbf {V} = W _ {v} P _ {5},\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 704, + 599, + 720 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S e l f A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {Q K}) \\mathbf {V}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 722, + 784, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{5}$ are the last level features of the backbone network; $\\mathbf{Q}$ , $\\mathbf{K}$ , and $\\mathbf{V}$ are query, key, and value features, respectively.", + "bbox": [ + 212, + 750, + 784, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The standard self-attention performs dense computation that involves all features. Applying standard self-attention to a large feature map will cause high complexity due to its dense computation, so the multi-scale features of the backbone network are not utilized in the original DETR. In addition, dense compu", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Dong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tation will lead to slow convergence. The original DETR required 500 epochs of training to achieve good performance.", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To exploit the multi-scale features of the backbone, Deformable DETR [50] proposes deformable attention. The multi-scale deformable self-attention can be simply formulated as:", + "bbox": [ + 212, + 176, + 782, + 222 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\operatorname {C o n c a t} \\left(P _ {3}, P _ {4}, P _ {5}\\right), \\mathbf {A} = W _ {a} \\mathbf {Q},\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 236, + 591, + 250 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {V} = \\operatorname {S a m p} \\left(W _ {v} \\mathbf {Q}, r + \\Delta r\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 253, + 612, + 268 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {D e f o r m a b l e A t t e n t i o n} (\\mathbf {Q}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 273, + 784, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $P_{3}, P_{4}, P_{5}$ are multi-scale features of the backbone, $\\mathbf{Q}$ are a set of queries, $\\Delta r$ are the offsets of reference points, $r$ are the reference points, $\\mathbf{A}$ are the attention weights of sample features, Samp is a function that extracts the features corresponding to the location $(r + \\Delta r)$ by bilinear interpolation.", + "bbox": [ + 212, + 300, + 782, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The deformable attention performs sparse computation that involves only a predefined number of surrounding features. Because deformable attention does not produce high complexity when combined with multi-scale features, it can utilize the multi-scale features of the backbone to detect multi-scale objects. In addition, sparse computation will lead to fast convergence. Deformable DETR converges 10 times faster than DETR. However, limited by sparse computation, deformable attention typically does not have a large receptive field and does not use global context information. This is not conducive to detecting large objects and objects in complex scenes.", + "bbox": [ + 212, + 361, + 784, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address these issues, we propose the Hybrid Attention Encoder that combines standard self-attention and deformable attention in the encoder layer. The hybrid attention layer is shown in Figure 3 (a). Introducing dense attention into the deformable attention encoder can enlarge the receptive field and perceive the global context. This will be more beneficial for multi-scale object detection. However, how to introduce dense attention into deformable attention without increasing too much complexity and slowing convergence is a key challenge. In order not to introduce too much complexity, we propose the hybrid-scale strategy. This strategy means that only the last scale features receive dense attention, not all scale features. With this strategy, our method does not introduce too much computation. Our hybrid attention can be simply formulated as:", + "bbox": [ + 212, + 497, + 784, + 662 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {\\prime} = \\mathbf {K} ^ {\\prime} = \\mathbf {V} ^ {\\prime} = P _ {5},\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 676, + 483, + 689 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP _ {5} ^ {\\prime} = \\operatorname {S e l f A t t e n t i o n} \\left(\\mathbf {Q} ^ {\\prime}, \\mathbf {K} ^ {\\prime}, \\mathbf {V} ^ {\\prime}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 694, + 562, + 709 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\operatorname {C o n c a t} \\left(P _ {3}, P _ {4}, P _ {5} ^ {\\prime}\\right), \\mathbf {A} = W _ {a} \\mathbf {Q},\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 714, + 591, + 729 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {V} = \\operatorname {S a m p} \\left(W _ {v} \\mathbf {Q}, r + \\Delta r\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 733, + 612, + 748 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {D e f o r m a b l e A t t e n t i o n} (\\mathbf {Q}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 751, + 782, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order not to slow down the convergence, we propose the hybrid-layer strategy. This strategy means that instead of adopting hybrid attention at all encoder layers, it is only adopted at some layers. From the ablation experiments, it can be seen that adopting hybrid attention in all the encoder layers will have no", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "AugDETR", + "bbox": [ + 656, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance gain due to slow convergence. In our practical use, we introduce hybrid attention only in the first two encoder layers in order to balance complexity and performance. With the above design, we build the Hybrid Attention Encoder that can enlarge the receptive field and utilize the global context.", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Encoder-Mixing Cross-Attention", + "text_level": 1, + "bbox": [ + 215, + 227, + 532, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The cross-attention in the decoder layer of Deformable DETR [50] uses multiscale deformable attention to achieve query-feature interaction. The multi-scale deformable cross-attention can be simply formulated as:", + "bbox": [ + 212, + 250, + 787, + 297 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} = W _ {a} \\mathbf {Q}, \\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {V} = \\operatorname {S a m p} (W _ {v} E _ {6}, r + \\Delta r),\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 306, + 683, + 321 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {D e f o r m a b l e A t t e n t i o n} (\\mathbf {Q}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 325, + 784, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}$ are a set of decoder queries, $\\mathbf{A}$ is the attention weights, $\\Delta r$ is the offsets of reference points, $r$ is the reference points, $E_6$ is the multi-scale features of last encoder layer, Samp is the sample function.", + "bbox": [ + 212, + 349, + 784, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DETR-based detectors typically stack six encoder layers to enhance the backbone features. However, only features of the last encoder layer are utilized in cross-attention. Although features of the last encoder layer have larger receptive fields and stronger semantics, it is not optimal for multi-scale objects due to the fact that objects have different favors for features. Features from different encoder layers have different receptive fields, and if they are fully utilized they will be more beneficial for multi-scale object detection.", + "bbox": [ + 212, + 395, + 787, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address these issues, cross-attention interacting with multi-level encoder can obtain more discriminative features. However, using the same multi-level encoder for all objects is not conducive to multi-scale learning since multi-scale objects have different feature preferences. We propose Encoder-Mixing Cross-Attention (EMCA) to adaptively exploit features of different encoders based on query features. The EMCA is shown in Figure 3 (b). Specifically, we first extract features from each encoder layer by using multi-scale deformable attention. Then, we generate fusion weights for extracted features from each encoder based on the query features. Finally, we adaptively aggregate features from each encoder according to the fusion weights. The EMCA can be simply formulated as:", + "bbox": [ + 212, + 500, + 787, + 652 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} = W _ {a} \\mathbf {Q}, \\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {w} _ {e} = \\sigma (W _ {e} \\mathbf {Q})\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 662, + 629, + 676 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {l} = \\operatorname {S a m p} \\left(W _ {v l} E _ {l}, r + \\Delta r\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 681, + 558, + 696 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E M C A} (\\mathbf {Q}, \\mathbf {E}) = \\sum_ {l = 1} ^ {L} w _ {e} ^ {l} \\cdot \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V} _ {l}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 700, + 784, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\sigma$ means the Sigmoid, $E_{l}$ is the features of the $l$ level encoder, $\\mathbf{E}$ is a set of features from each level encoders, $w_{e}^{l}$ is the weights of the $l$ level encoder, $\\mathbf{w}_{e}$ is a set of weights for all encoders. $L$ is the number of encoder layers.", + "bbox": [ + 212, + 750, + 784, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use the EMCA to replace all the cross-attention in the decoder layer. With this design, each object can adaptively utilize multi-level encoder features to obtain more discriminative features for improved multi-scale learning.", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Dong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5451531f21b457b6f7b8c5841ba6d0ec69911b2597411b60c35c91343068d4f5.jpg", + "table_caption": [ + "Table 2: Comparison with the state-of-the-art DETR variants on COCO val2017 with the ResNet-50-4scale backbone. * indicates the result is from our implemented model." + ], + "table_footnote": [], + "table_body": "
ModelBackbone#epochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Conditional-DETR [29]R5010843.064.045.722.746.761.5
SAM-DETR [41]R505039.861.841.620.543.459.6
Anchor-DETR [38]R505042.163.144.922.346.260.0
Dynamic-DETR [6]R501242.961.046.324.644.954.4
SMCA-DETR [10]R5010845.665.549.125.949.362.6
CF-DETR [3]R503647.866.552.431.250.662.8
Sparse-DETR [35]R505046.366.050.129.049.560.8
BoxeR-2D [30]R505050.067.954.730.952.862.6
Deformable-DETR [50]R505046.965.651.029.650.161.6
DAB-Deformable-DETR [24]R505046.866.050.429.149.862.3
DN-Deformable-DETR [18]R505048.667.452.731.052.063.7
H-DETR [14]R501248.766.452.931.251.563.5
Co-DETR [51]R501249.567.654.332.452.763.7
DINO [42]R501249.066.653.532.052.363.0
DINO [42]R502450.468.354.833.353.764.8
DDQ* [45]R501250.867.956.134.654.365.5
Focus DETR [46]R503650.468.555.034.053.564.4
AlignDETR [2]R501250.267.854.432.953.365.0
Stable-DINO [25]R501250.467.455.032.954.065.5
Rank-DETR [32]R501250.267.755.034.153.664.0
DINO+OursR501250.2 (+1.2)67.855.032.353.264.7
DINO+OursR502451.3 (+0.9)69.056.233.554.766.1
AlignDETR+OursR501251.3 (+1.1)68.855.833.354.866.6
DDQ+OursR501251.8 (+1.0)69.256.934.855.066.7
", + "bbox": [ + 215, + 185, + 785, + 526 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 551, + 375, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset and Metrics. We conduct all experiments on the challenging COCO 2017 detection dataset. It is split into training set, validation set, and test-dev set with 115K, 5K, and 20K images respectively. We train models on the training set and evaluate performance on the validation set and test-dev set. All evaluated results follow the COCO-style Average Precision (AP) metrics.", + "bbox": [ + 215, + 583, + 785, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We use DINO [42], Align DETR [2], and DDQ [45] as our baseline methods. All ablation studies are conducted on DINO-4scale with ResNet-50 for 12 epochs. We employ AdamW with $1 \\times 10^{-4}$ weight decay as the optimizer to train our models. The batch size is set as 16 for all experiments. The initial learning rate is set as $1 \\times 10^{-4}$ and it decreases by multiplying 0.1 after the 11th epoch for the 12 epochs setting. We use a random seed 0 in all our experiments. All other hyper-parameters follow the default from the codebase.", + "bbox": [ + 215, + 659, + 785, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Main Results", + "text_level": 1, + "bbox": [ + 215, + 785, + 375, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we evaluate AugDETR on the COCO val2017 set and compare it with the state-of-the-art DETR variants. All results with ResNet50 are sum-", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "AugDETR", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c872ad3436994c6dcd5b8eaf014eacd763d4a35519c5bb43fe7851e799202d25.jpg", + "table_caption": [ + "Table 3: Comparison with the DETR variants on val2017 set with the Swin-T." + ], + "table_footnote": [], + "table_body": "
MethodBackbone#epochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
DINO [42]Swin-T1251.369.056.034.554.466.0
OursSwin-T1252.3 (+1.0)70.257.036.855.367.2
AlignDETR [2]Swin-T1252.570.157.235.955.968.4
OursSwin-T1253.5 (+1.0)71.358.136.956.669.1
", + "bbox": [ + 220, + 171, + 776, + 252 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "marized in Table 2. Combining our components with DINO [42], AugDETR achieves 50.2 AP, which is 1.2 AP higher than the DINO baseline with the same setting. Recent work on replacing the focal loss in DINO with the IoU-aware focal loss [19, 21, 43] has shown good performance. We chose a typical work AlignDETR [2] from these studies as baseline to validate the generality of our approach. Combining our components with AlignDETR, AugDETR achieves 51.3 AP, which is 1.1 AP higher than the AlignDETR baseline. Besides, we conduct experiments in the latest SOTA model DDQ [45]. Combining the our components with DDQ, AugDETR achieves 51.8 AP, which is 1.0 AP higher than the DDQ baseline. To demonstrate the effectiveness of our method under longer training, we conducted experiments under the $2 \\times$ schedule. Our method achieves 51.3 AP, which is 0.9 AP higher than the baseline DINO. Besides, AugDETR can consistently achieve performance improvements even with more advanced backbone networks. As shown in Table 3, when using Swin-T [28] as the backbone, our method achieves performance improvements of 1.0 AP and 1.0 AP over DINO and AlignDETR, respectively. These results demonstrate the effectiveness and generalizability of our method.", + "bbox": [ + 212, + 277, + 787, + 534 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6f6802950ce75c2c5160aae6cce808a55974bc5172e637f5e9c7e99b14206cf5.jpg", + "table_caption": [ + "Table 4: Ablations on each component of AugDETR. \"HAE\" and \"EMCA\" represent Hybrid Attention Encoder, and Encoder-Mixing Cross-Attention, respectively." + ], + "table_footnote": [], + "table_body": "
HAEEMCAAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
49.066.653.532.052.363.0
✓49.667.054.231.852.963.8
✓49.867.454.633.052.864.7
✓✓50.267.855.032.353.264.7
", + "bbox": [ + 305, + 599, + 692, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Ablations", + "text_level": 1, + "bbox": [ + 215, + 727, + 341, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ablation studies on the effectiveness of each component. To analyze the contributions of individual components in AugDETR, Hybrid Attention Encoder (HAE), and Encoder-Mixing Cross-Attention (EMCA) are gradually introduced to the baseline. As shown in Table 4, HAE brings 0.6 AP improvement to the baseline model. This benefits from the HAE enlarges the receptive field and introduces the global context features to improve semantic representation. The", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Dong et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "improvements of $AP_{M}$ $(+0.6\\mathrm{AP})$ and $AP_{L}$ $(+0.8\\mathrm{AP})$ contribute to the final performance improvement. These results are good demonstrations of our motivation and design. The EMCA improves the detection performance from 49.0 AP to 49.8 AP. It can be seen that the results on small, medium, and large objects are all improved, which means that object queries adaptively interacting with multi-level encoder features by cross-attention can extract more discriminative object features for improved multi-scale learning. When both components are included in the baseline model, it achieves 50.2 AP with 1.2 AP performance improvement. The results show that the two components are complementary and address different issues in DETR [4].", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3ff066831c3d2c3f05970021a16146a94feec72c0e811e4f28c37c5e9853fbaa.jpg", + "table_caption": [ + "Table 5: Comparisons of ours HAE with Deformable Encoder variants." + ], + "table_footnote": [], + "table_body": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)Flops
Deform. Enc. (4 Points)49.066.653.532.052.363.0245G
Deform. Enc. (6 Points)48.465.952.731.051.662.9250G
Deform. Enc. (8 Points)48.566.152.830.551.862.8255G
HAE (Ours)49.667.054.231.852.963.8248G
", + "bbox": [ + 250, + 352, + 750, + 433 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparisons of HAE with Deformable Encoder variants. To make a fair comparison with the deformable encoder [50], we conduct experiments on increasing sampling points in the deformable encoder. The results are shown in Table 5. The performance of baseline with 4 sampling points is 49.0 AP. Performance decreases as sample points in the deformable encoder increases to 6 and 8, which are 48.4 AP and 48.5 AP, respectively. Even increasing the number of sample points in the deformable encoder does not improve performance for large and medium objects. This may be because deformable attention determines sampling weights based on query alone, leading to inaccurate weights as the number of points increases. However, our HAE can achieve performance gains (+0.6 AP) with less computational consumption (248G vs 250G). We think the possible reason is that the weight generation in standard self-attention, which calculates the similarity by query and key, is more suitable for long-distance sampling. These results support the motivation and design of our hybrid attention encoder.", + "bbox": [ + 212, + 460, + 787, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The number of Hybrid Attention Layers. Introducing too much self-attention may lead to slow convergence. We analyze the effects of the number of hybrid attention layers. The results are shown in Table 6. When only a hybrid attention layer is introduced, the model achieves 49.3 AP with 0.3 AP improvement. The improvement is mainly contributed by large and medium objects. Increasing the number of hybrid attention layers by 2 or 3 results in performance gains of 0.6 and 0.7 AP, respectively. It is still the large and medium objects that contribute the most performance gains. As the number of hybrid attention layers increases, the performance gains begin to decrease. In particular, there is no performance gain when increasing to 6 hybrid attention layers. This may be because the negative effect of slow convergence becomes dominant.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "AugDETR", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a786e5eaadb9ccb949e2a675afa982b3b94ef6695a55bb85a2758ff121ca8d3c.jpg", + "table_caption": [ + "Table 6: Ablations on the number of hybrid attention layers in HAE." + ], + "table_footnote": [], + "table_body": "
NumberAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
-49.066.653.532.052.363.0
149.366.653.731.152.763.7
249.667.054.231.852.963.8
349.767.254.332.352.864.0
449.367.053.831.952.763.5
549.366.854.031.452.863.1
648.866.353.430.752.263.1
", + "bbox": [ + 326, + 171, + 671, + 292 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e71a30fc47c2f0d942e20fab1ad587cba660c169cb1e3a2d05e58a72b628382b.jpg", + "table_caption": [ + "Table 7: Comparisons of ours EMCA with other decoder variants. * indicates that these models scale to the similar complexity as EMCA." + ], + "table_footnote": [], + "table_body": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_{2th} \\)\\( AP_{6th} \\)\\( AP_{10th} \\)Flops
DINO [42]49.066.653.534.243.245.7245G
+Iterative Enc [40]47.765.252.133.442.044.6245G
+Iterative Enc*47.465.151.633.242.044.1284G
+Memory Fusion [25]49.367.053.735.743.746.2253G
+Memory Fusion*49.366.953.935.743.546.2290G
+EMCA (Ours)49.867.454.636.044.546.6283G
", + "bbox": [ + 241, + 349, + 754, + 459 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparisons of EMCA with decoder variants. We compare our EMCA with other decoder variants. As shown in Table 7, using the Iterative Encoder [40] degrades the performance of the baseline. There are 0.3 AP performance gains when the Memory Fusion [25] is introduced, but there is no performance gain from the scaling of the approach. The performance gain of Memory Fusion is 0.5 AP less than our EMCA. These results demonstrate the effectiveness of our method design, which adaptively fuses multi-level encoder features based on query features for multi-scale objects, rather than using the same features for all objects. In addition, to show that our method converges faster, we also show results for the 2nd, 6th, and 10th epochs. It can be seen that our method outperforms the other methods throughout the training process, indicating the faster convergence of our method.", + "bbox": [ + 212, + 489, + 785, + 671 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparisons of EMCA with DINO variants. For a fair comparison, we compared with more complex baseline models due to the increased computational complexity introduced by EMCA. More complex baseline models are implemented by introducing more encoders and features. The results are shown in Table 8. By increasing the original baseline from 6 encoders to 7 encoders, the model reached 49.3 AP with 268G. When using 8 encoders, the model achieves 49.5 AP with 292G. While performance can be improved by stacking encoders, our EMCA utilizes encoder redundancy information and is a good complement to encoders. When using 10 encoders, the model suffers from overfitting and achieves only 47.8 AP with 339G. However, our EMCA with 10 encoders still has performance gains to reach 50.6 AP. This results show that our EMCA can", + "bbox": [ + 212, + 672, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Dong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1f0593880f19c08a67f2968fa7224fce85c85be2f7d1c54fd98476d1814e5277.jpg", + "table_caption": [ + "Table 8: Comparisons of ours EMCA with DINO variants." + ], + "table_footnote": [], + "table_body": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)Flops
6 encoder (Baseline)49.066.653.532.052.363.0245G
7 encoder49.366.753.732.052.463.7268G
8 encoder49.566.954.031.952.863.4292G
10 encoder47.864.852.130.650.961.5339G
5 scale49.466.953.832.352.563.9721G
6encoder+EMCA (Ours)49.867.454.633.052.864.7283G
8encoder+EMCA (Ours)50.267.854.632.453.364.6346G
10encoder+EMCA (Ours)50.668.155.333.553.965.3409G
", + "bbox": [ + 241, + 170, + 756, + 308 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "mitigate the risk of overfitting. A very large amount of computation (721G) is required when 5 scales of backbone features are used, and then the performance reaches only 49.4 AP. These results show that our EMCA is a good complement to encoders and can mitigate the risk of overfitting.", + "bbox": [ + 212, + 339, + 782, + 398 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitative analysis of EMCA. To analyze the weights of EMCA, we count the fusion weights in the last decoder layer from the COCO val2017 set. First, we analyzed the weights of different encoder layers, and the results are shown in Figure 4. It can be seen that our method learns different weights for encoder layers. Specifically, the first encoder layer plays the most important role, the 2nd, 3rd, and 4th encoder layers have decreasing importance, and the 4th, 5th, and 6th encoder layers have similar roles. As shown in Table 9, using the 2nd, 4th, and 6th encoder layers as inputs, our approach drops 0.3 AP and saves 23G of computational overhead. When only the 1st and 6th encoder layers are used, the computation is reduced by 30G while the performance is dropped by 0.2 AP. We call the EMCA with only the 1st and 6th encoder layers as EMCA-Lite. We then analyze the weights of the objects at different scales to illustrate that our method improves multi-scale learning. As shown in Figure 5, objects of different sizes are given different weights, and the larger the scale, the higher the weights. See the Appendix for object scale intervals. These results illustrate the effectiveness of fusion weights based on queries.", + "bbox": [ + 212, + 401, + 787, + 643 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e519da46a9aa1b3a905a5dd92e71b061434d391bd679246b694a1cb54e6c989.jpg", + "image_caption": [ + "Fig. 4: Statistical fusion weights for different encoders. Layer #6 means the last-level encoder." + ], + "image_footnote": [], + "bbox": [ + 227, + 672, + 478, + 781 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e5df4497c2d98d95ddb2dcfb7ee420f32edcf0ed38029689c5ecf32e78402bd6.jpg", + "image_caption": [ + "Fig.5: Statistical fusion weights for different scale objects. Size #1 to size #6 are ordered by scale from small to large." + ], + "image_footnote": [], + "bbox": [ + 521, + 672, + 774, + 781 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "AugDETR", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/fc6cbe0f80a5abb32c773f8e7905a0949d076dc53c784bd227f02571524dddc1.jpg", + "table_caption": [ + "Table 9: Ablations on the layer of input encoders in EMCA." + ], + "table_footnote": [], + "table_body": "
EncoderAPAP50AP75Flops
-49.066.653.5245G
12345649.867.454.6283G
24649.567.253.9260G
2649.367.253.8253G
1649.667.454.3253G
", + "bbox": [ + 235, + 184, + 470, + 280 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/2f239f909b7fed2d463c90387c53ad1bdcdd7438ad5d1ce6dfbf51f6d9c88f42.jpg", + "table_caption": [ + "Table 10: Ablations on the fusion methods in EMCA." + ], + "table_footnote": [], + "table_body": "
Fusion MethodAPAPSAPMAPL
-49.032.052.363.0
sum49.331.952.563.9
max48.630.951.663.2
sigmoid weighted49.833.052.864.7
softmax weighted49.532.152.864.5
", + "bbox": [ + 511, + 184, + 779, + 280 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fusion Methods in EMCA. The fusion methods we explore include sum fusion, max fusion, and adaptive weighted fusion (sigmoid and softmax). All results are shown in Table 10. By using sum fusion, our method improves the baseline by 0.3 AP. This result shows that multi-level encoder features are richer in semantic information than single-level encoder features. When using max fusion, our method achieves 48.6 AP below the baseline. It can be seen that the performance decreases significantly for small and medium objects, but increases slightly for large objects. This may be because large objects tend to have higher activation so the features of small and medium objects are weakened during max fusion. When using adaptive weighted fusion, our method improves the baseline by 0.8 AP and 0.5 AP, respectively. It can be seen that the performance of small, medium, and large objects is all improved. In particular, the performance on large objects has improved significantly (+1.7 AP) by sigmoid weighted fusion. This may be because the larger the object, the more information is activated at all levels of the encoder, and therefore the more information can be used. In addition, the corresponding query features of large objects are richer so that the learned fusion weights are more accurate. These results suggest that sigmoid weighted fusion is a better fusion method.", + "bbox": [ + 212, + 316, + 787, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 633, + 359, + 650 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we analyze the limitations of DETR detectors for multi-scale learning and propose local feature enhancement and multi-level encoder exploitation techniques to address the limitations. Based on these, we construct a novel detection transformer detector named Augmented DETR (AugDETR), which consists of two components: Hybrid Attention Encoder and Encoder-Mixing Cross-Attention. HAE enlarges the receptive field and introduces global context to enhance the local feature of the deformable encoder. EMCA uses multi-level encoders to obtain more discriminative features and speed up convergence. Extensive experiments with various DETR-based detectors and backbones validate the effectiveness of our method.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Dong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China (NSFC) under Grant 62088102. This work is partly supported by China Telecom Corporation Ltd. Data&Al Technology Company, Beijing, China and China Telecom - Xi'an Jiaotong University Joint Innovation Institute for Science-Education Integration of Intelligent Cloud Network. We also appreciate Jingwen Fu, He Zhang and Tao Yang for their insightful discussions. We thank all the anonymous reviewers and Area Chairs for their constructive and helpful comments, which have significantly improved the quality of the paper.", + "bbox": [ + 212, + 179, + 787, + 301 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 327, + 321, + 343 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154-6162 (2018)", + "2. Cai, Z., Liu, S., Wang, G., Ge, Z., Zhang, X., Huang, D.: Align-detr: Improving detr with simple iou-aware bce loss. arXiv preprint arXiv:2304.07527 (2023)", + "3. Cao, X., Yuan, P., Feng, B., Niu, K.: Cf-detr: Coarse-to-fine transformers for end-to-end object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 185-193 (2022)", + "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "5. Chen, Q., Chen, X., Zeng, G., Wang, J.: Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085 (2022)", + "6. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2988-2997 (2021)", + "7. Ding, X., Zhang, X., Han, J., Ding, G.: Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11963-11975 (2022)", + "8. Dong, J., Huang, Y., Zhang, S., Chen, S., Zheng, N.: Construct effective geometry aware feature pyramid network for multi-scale object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 534-541 (2022)", + "9. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019)", + "10. Gao, P., Zheng, M., Wang, X., Dai, J., Li, H.: Fast convergence of detr with spatially modulated co-attention. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3621-3630 (2021)", + "11. Ghiasi, G., Lin, T.Y., Le, Q.V.: Nas-fpn: Learning scalable feature pyramid architecture for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7036-7045 (2019)", + "12. Guo, C., Fan, B., Zhang, Q., Xiang, S., Pan, C.: Augfpn: Improving multi-scale feature learning for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12595-12604 (2020)" + ], + "bbox": [ + 218, + 361, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "AugDETR", + "bbox": [ + 656, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. He, Y., Zhu, C., Wang, J., Savvides, M., Zhang, X.: Bounding box regression with uncertainty for accurate object detection. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 2888-2897 (2019)", + "14. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: Detrs with hybrid matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19702-19712 (2023)", + "15. Kim, S.W., Kook, H.K., Sun, J.Y., Kang, M.C., Ko, S.J.: Parallel feature pyramid network for object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 234-250 (2018)", + "16. Kong, T., Sun, F., Liu, H., Jiang, Y., Li, L., Shi, J.: Foveabox: Beyound anchor-based object detection. IEEE Transactions on Image Processing 29, 7389-7398 (2020)", + "17. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. In: Proceedings of the European conference on computer vision (ECCV). pp. 734-750 (2018)", + "18. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)", + "19. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems 33, 21002-21012 (2020)", + "20. Lin, T.Y., Dólar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)", + "21. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "23. Lin, Y., Yuan, Y., Zhang, Z., Li, C., Zheng, N., Hu, H.: Detr does not need multiscale or locality design. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6545-6554 (2023)", + "24. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. In: International Conference on Learning Representations (2021)", + "25. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., et al.: Detection transformer with stable matching. arXiv preprint arXiv:2304.04742 (2023)", + "26. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8759-8768 (2018)", + "27. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 21-37. Springer (2016)", + "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Dong et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)", + "30. Nguyen, D.K., Ju, J., Booij, O., Oswald, M.R., Snoek, C.G.: Boxer: Box-attention for 2d and 3d transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4773-4782 (2022)", + "31. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra r-cnn: Towards balanced learning for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 821-830 (2019)", + "32. Pu, Y., Liang, W., Hao, Y., Yuan, Y., Yang, Y., Zhang, C., Hu, H., Huang, G.: Rank-detr for high quality object detection. Advances in Neural Information Processing Systems 36 (2024)", + "33. Qiao, S., Chen, L.C., Yuille, A.: Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10213-10224 (2021)", + "34. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 39(6), 1137-1149 (2016)", + "35. Roh, B., Shin, J., Shin, W., Kim, S.: Sparse detr: Efficient end-to-end object detection with learnable sparsity. In: International Conference on Learning Representations (2021)", + "36. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020)", + "37. Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9627-9636 (2019)", + "38. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022)", + "39. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)", + "40. Zhang, G., Luo, Z., Tian, Z., Zhang, J., Zhang, X., Lu, S.: Towards efficient use of multi-scale features in transformer-based object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6206-6216 (2023)", + "41. Zhang, G., Luo, Z., Yu, Y., Cui, K., Lu, S.: Accelerating detr convergence via semantic-aligned matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 949-958 (2022)", + "42. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: The Eleventh International Conference on Learning Representations (2022)", + "43. Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8514-8523 (2021)", + "44. Zhang, H., Chang, H., Ma, B., Wang, N., Chen, X.: Dynamic r-cnn: Towards high quality object detection via dynamic training. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 260-275. Springer (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "AugDETR", + "bbox": [ + 656, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7329-7338 (2023)", + "46. Zheng, D., Dong, W., Hu, H., Chen, X., Wang, Y.: Less is more: Focus attention for efficient detr. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6674-6683 (2023)", + "47. Zhou, X., Wang, D., Krähenbuhl, P.: Objects as points. arXiv preprint arXiv:1904.07850 (2019)", + "48. Zhou, X., Zhuo, J., Krahenbuhl, P.: Bottom-up object detection by grouping extreme and center points. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 850-859 (2019)", + "49. Zhu, C., He, Y., Savvides, M.: Feature selective anchor-free module for single-shot object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 840-849 (2019)", + "50. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)", + "51. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6748-6758 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 439 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Dong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_model.json b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..60fb80bbbb033ad0f52768d2375b412dd23149d3 --- /dev/null +++ b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_model.json @@ -0,0 +1,2337 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "AugDETR: Improving Multi-scale Learning for Detection Transformer" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.213, + 0.747, + 0.229 + ], + "angle": 0, + "content": "Jinpeng Dong, Yutong Lin, Chen Li, Sanping Zhou, Nanning Zheng*" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.24, + 0.782, + 0.311 + ], + "angle": 0, + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University {djp1235a, yutonglin, edward82}@stu.xjtu.edu.cn {spzhou, nnzheng} \\(@\\) xjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.347, + 0.74, + 0.596 + ], + "angle": 0, + "content": "Abstract. Current end-to-end detectors typically exploit transformers to detect objects and show promising performance. Among them, Deformable DETR is a representative paradigm that effectively exploits multi-scale features. However, small local receptive fields and limited query-encoder interactions weaken multi-scale learning. In this paper, we analyze local feature enhancement and multi-level encoder exploitation for improved multi-scale learning and construct a novel detection transformer detector named Augmented DETR (AugDETR) to realize them. Specifically, AugDETR consists of two components: Hybrid Attention Encoder and Encoder-Mixing Cross-Attention. Hybrid Attention Encoder enlarges the receptive field of the deformable encoder and introduces global context features to enhance feature representation. Encoder-Mixing Cross-Attention adaptively leverages multi-level encoders based on query features for more discriminative object features and faster convergence. By combining AugDETR with DETR-based detectors such as DINO, AlignDETR, DDQ, our models achieve performance improvements of 1.2, 1.1, and 1.0 AP in the COCO under the ResNet-50-4scale and 12 epochs setting, respectively." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.61, + 0.741, + 0.636 + ], + "angle": 0, + "content": "Keywords: Object detection \\(\\cdot\\) Detection transformer \\(\\cdot\\) Hybrid attention \\(\\cdot\\) Multi-level encoder" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.664, + 0.377, + 0.68 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Many CNN-based detectors [1,12,20,27,34,37,47] have been proposed to improve detection performance in the past decade. Despite their promising progress, these detectors suffer from many hand-crafted components, such as non-maximum suppression and anchor design. In contrast, CNN-based paradigm is broken by DETR [4], a novel transformer detector. It gets rid of hand-crafted components and implements an end-to-end paradigm. Despite its novel paradigm, DETR suffers from some limitations, including slow convergence and inferior performance. To address these limitations, many methods [2, 6, 14, 24, 25, 29, 42, 50]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "J. Dong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "have been proposed from different perspectives, such as introducing spatial priors, hybrid matching training, designing efficient attention modules, assigning sample weights, etc. With these optimizations, DINO [42] achieved a new record on the COCO benchmark [22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.79, + 0.406 + ], + "angle": 0, + "content": "Although DETR-based detectors have achieved promising performance, multiscale object detection has not received sufficient attention so far, which may limit the performance of detectors. Scale variation of objects remains one of the crucial challenges in object detection. Objects at different scales require inconsistent semantic features, e.g., large objects require features with larger receptive fields. To the best of our knowledge, Deformable-DETR [50] has made some effort to remedy this issue. Deformable-DETR uses multi-scale backbone features to improve multi-scale learning by deformable attention. In addition to the use of multi-scale backbone features, we believe that the multi-scale learning of Deformable-DETR has not been fully exploited. As shown in Table 1, although Deformable DETR is 2.5 AP higher than the DETR in overall performance, the performance on large objects is still 1.5 AP lower than the DETR. We present two techniques that can further improve multi-scale learning as follows:" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.45, + 0.533, + 0.478 + ], + "angle": 0, + "content": "Table 1: Performance comparison between DETR and Deformable DETR." + }, + { + "type": "table", + "bbox": [ + 0.228, + 0.489, + 0.524, + 0.551 + ], + "angle": 0, + "content": "
MethodEpochsAPAPSAPMAPL
DETR [4]50042.020.545.861.1
Deformable DETR [50]5044.527.147.659.6
" + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.435, + 0.788, + 0.463 + ], + "angle": 0, + "content": "Fig. 1: Comparison of receptive fields of different encoders." + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.474, + 0.663, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.474, + 0.787, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Local feature enhancement. Deformable-DETR [50] replaces the original self-attention (global dense attention) with deformable attention (sparse attention) to reduce the complexity and slow convergence of dense attention. Deformable attention is implemented by a fixed number of deformable sampling points. Limited by local sampling points, deformable attention is difficult to achieve a larger receptive field compared to self-attention. This is not conducive to the detection of large objects. Further stacking of deformable attention still improves performance, as analyzed in the ablations. In addition, deformable attention loses the ability to perceive global context, while multi-scale object detection in some complex scenes may rely on global context information. Increasing the sampling points degrades the performance of the deformable attention, as analyzed in the ablations. Therefore, we introduce global contextual information into the deformable encoder to enhance multi-scale learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Multi-level encoder exploitation. DETR-based detectors typically stack six encoder layers and six decoder layers. Encoder layers enhance the backbone features, and decoder layers extract features from last encoder layer to decode" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "AugDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "queries. Previous methods ignore the use of multi-level encoder features. More discriminative object features can be obtained by fully exploiting multi-level encoder features with different receptive fields and feature preferences. We refer to RepLKNet [7] for analysis of receptive field. As shown in Figure 1, the comparison of encoder 1 and 6 in DINO [42] shows that different encoders have varying feature preferences and receptive fields. In addition, the fact that multilevel encoders are all involved in object feature generation can speed up convergence because the supervisory signals are available to multiple levels of encoders, rather than being available only to the last level encoder. Utilizing multi-level encoder features also can introduce redundancy to mitigate the risk of overfitting. Although multi-level encoder features is beneficial for object detection, using multi-level encoder features fixed for all objects is not conducive to multiscale learning. Since objects at different scales have different feature preferences, adaptive use of multi-level encoder features by learning weights based on query features will facilitate multi-scale learning for DETR-based detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.6 + ], + "angle": 0, + "content": "In this paper, we propose AugDETR, a simple yet effective DETR-based detector that uses two different components to improve multi-scale learning. First, Hybrid Attention Encoder is proposed to augment local features in the deformable encoder by partially combining dense attention with deformable attention. Hybrid Attention Encoder can enlarge the receptive field and introduce global context, which can improve detection performance on large objects. Second, Encoder-Mixing Cross-Attention is introduced to better exploit features from different encoder layers by weighted aggregation. Objects at different scales can learn different aggregation weights based on object features to facilitate multi-scale learning. With this design, Encoder-Mixing Cross-Attention can generate more discriminative object features for subsequent classification and localization tasks. We validate the effectiveness of AugDETR on the COCO benchmark. Our method achieves 50.2 AP (+1.2 AP), 51.3 AP (+1.1 AP), and 51.8 AP (+1.0 AP) based on DINO [42], AlignDETR [2], and DDQ [45] with ResNet-50-4scale under 12 epochs settings, respectively." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.601, + 0.558, + 0.614 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.65 + ], + "angle": 0, + "content": "- We analyze local feature enhancement and multi-level encoder exploitation techniques for improved multi-scale learning." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.651, + 0.785, + 0.693 + ], + "angle": 0, + "content": "- A new detection transformer detector named AugDETR is proposed to achieve the techniques with Hybrid Attention Encoder and Encoder-Mixing Cross-Attention." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.695, + 0.785, + 0.739 + ], + "angle": 0, + "content": "- We validate AugDETR equipped with various DETR-based detectors and backbones on the COCO benchmark, and it consistently yields significant performance improvements over DETR-based detectors." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.759, + 0.388, + 0.775 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.406, + 0.803 + ], + "angle": 0, + "content": "2.1 Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Early deep-learning detectors can be roughly classified into anchor-based and anchor-free methods. Anchor-based methods use predefined anchors to locate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "J. Dong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.192 + ], + "angle": 0, + "content": "objects, including Faster R-CNN [34], SSD [27], RetinaNet [21], and their variants [1, 13, 31, 44]. Anchor-free methods use points near object centers to locate objects, including CenterNet [47], FCOS [37], and others [9, 16, 17, 48, 49]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Recently, DETR [4] proposes a simple transformer pipeline for object detection, effectively alleviating the need for heuristic components. Compared to CNN-based work, DETR is rather novel but still suffers from some limitations, including slow convergence and inferior performance. Many follow-up works focus on alleviating the limitations of DETR from different perspectives. Conditional DETR [29] proposes to decouple the context embedding and position embedding, which helps DETR speed up convergence. Anchor DETR [38] uses the anchor points as the initial queries for fast convergence. DAB DETR [24] uses 4D anchor boxes that are dynamically updated with the decoder layer to represent queries. DN DETR [18] proposes a denoising training mechanism to speed up convergence. DINO [42] further extends DN DETR with contrastive denoising training, mixed query, and look-forward twice techniques. Group DETR [5], HDETR [14], and CO-DETR [51] propose various one-to-one matching and one-to-many matching combination methods to stabilize the matching process of DETR. Align DETR [2], Stable DINO [25], and Rank DETR [32] design different schemes to weight the samples by the IoU-aware classification loss. Plain DETR [23] proposes a box-to-pixel relative position bias and a masked image modeling [39] pre-training scheme to improve the DETR detector. Unlike these methods that design spatial priors, hybrid matching, sample weight assignments, etc., our approach focuses on improving the multi-scale learning of DETR detectors. In addition, we conducted experiments using some SOTA methods as the baseline models. Our techniques are complementary to theirs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.612, + 0.56 + ], + "angle": 0, + "content": "2.2 Multi-Scale Learning for Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In the early era, image pyramid is a common solution to improve multi-scale learning. Instead of image pyramid, some methods utilize multi-scale features to mitigate scale variation. A representative work among these methods is the feature pyramid network (FPN [20]), which proposes a top-down pathway and lateral connections to construct the feature pyramid. Inspired by FPN, multi-scale feature extraction has been widely studied, such as BiFPN [36], PAFPN [26], NAS-FPN [11], AugFPN [12], and other variants [8,15,33]. Recently, some works have attempted to design multi-scale feature extraction modules in DETR. Deformable DETR [50] proposes deformable attention to utilize multi-scale backbone features. Dynamic DETR [6] proposes a CNN-based dynamic encoder with scale-aware attention, spatial-aware attention, and task-aware attention. SMCA [10] generates the Gaussian-like weight map as the spatial prior and the scale weights as the scale prior to modulated multi-scale features. IMFA [40] designs a new scheme to adaptively sample sparse multi-scale features. Different from them, our approach enhances multi-scale learning by introducing global attention to expand the receptive field and integrate global context. To use global attention effectively, we propose hybrid scale and hybrid layer strategies to construct the hybrid attention encoder. In addition, we adaptively use multi-level" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "encoder features based on query features to improve multi-scale learning, which is also distinctly different from previous methods." + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.205, + 0.492, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.203, + 0.768, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig. 2: Left: The pipeline of conventional DETR-based detectors. The backbone network and the stacked encoder layers are used to extract and enhance the image features. Then, the object queries utilize the decoder layers to interact with the enhanced image features. Finally, the interacted object queries are fed into the detection head to obtain the predictions. Only the features from the last encoder layer can be used to interact with the object queries. Right: The pipeline of Our AugDETR. Our AugDETR restructures the conventional encoder-decoder pipeline. The object queries can interact with features from all encoder layers. For clarity, only the 3-layer structure is shown." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.331, + 0.57 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.645 + ], + "angle": 0, + "content": "In section 3.1, we first briefly revisit the detection transformer and then introduce the pipeline of AugDETR. In section 3.2, we describe the details of the proposed Hybrid Attention Encoder. In section 3.3, we introduce the Encoder-Mixing Cross-Attention used in the transformer decoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.553, + 0.68 + ], + "angle": 0, + "content": "3.1 A Revisit of DETR-based Detector" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.734 + ], + "angle": 0, + "content": "Since our proposed method is constructed on top of the recent DETR-based object detectors, we first briefly review the pipeline of the DETR-based detectors, taking the Deformable DETR [50] as an example." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The pipeline of Deformable DETR is shown in Figure 2 (left). Deformable DETR consists of 3 subcomponents: a backbone, a encoder, and a decoder. Given an input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), a backbone network generates multi-scale features. We denote the multi-scale features as \\(\\{C_2, C_3, C_4, C_5\\}\\). The spatial resolutions of these features are usually \\(1/4^2\\), \\(1/8^2\\), \\(1/16^2\\), and \\(1/32^2\\) of the given image. Then these features reduced to the same channel by a linear projection are denoted as \\(\\{P_2, P_3, P_4, P_5\\}\\). Next, a transformer encoder is used to further enhance these" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "J. Dong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.374 + ], + "angle": 0, + "content": "multi-scale features. The transformer encoder is typically stacked with 6 encoder layers. The encoder layer mainly consists of a self-attention and a feed-forward network. For Deformable DETR, multi-scale deformable attention is employed as self-attention to enhance multi-scale features. The encoded features of the different layers in the encoder are denoted as \\(\\{E_1,E_2,E_3,E_4,E_5,E_6\\}\\). Finally, a transformer decoder is used to produce object predictions from a set of object queries. Typically, only the last level of encoder features are fed into the decoder to interact with the query. The transformer decoder is also typically stacked with 6 decoder layers for better detection performance. The decoder layer mainly consists of a self-attention, a cross-attention, and a feed-forward network. For deformable DETR, multi-scale deformable attention is used as cross-attention to achieve query-feature interaction. As shown in Figure 2 (right), our pipeline changes the fact that the decoder can utilize only the last level of encoder features, but rather all encoder features can be utilized by the decoder. This design lays the foundation for cross-attention to utilize multi-level encoder features." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.401, + 0.496, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.4, + 0.778, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.505, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Fig. 3: (a) is the process of hybrid attention in the Hybrid Attention Layer. Dense attention is first applied to only the top features, and then sparse attention is applied to the multi-scale features. (b) is the details of Encoder-Mixing Cross-Attention. Deformable attention extracts multiple object features from multi-level encoder and then the extracted features are aggregated based on the adaptive fusion weights learned from the query features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.64, + 0.482, + 0.654 + ], + "angle": 0, + "content": "3.2 Hybrid Attention Encoder" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.664, + 0.785, + 0.68 + ], + "angle": 0, + "content": "The encoder layer in the original DETR [4] uses the self-attention as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.705, + 0.6, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = W _ {q} P _ {5}, \\mathbf {K} = W _ {k} P _ {5}, \\mathbf {V} = W _ {v} P _ {5},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.723, + 0.785, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S e l f A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {Q K}) \\mathbf {V}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.779 + ], + "angle": 0, + "content": "where \\(P_{5}\\) are the last level features of the backbone network; \\(\\mathbf{Q}\\), \\(\\mathbf{K}\\), and \\(\\mathbf{V}\\) are query, key, and value features, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The standard self-attention performs dense computation that involves all features. Applying standard self-attention to a large feature map will cause high complexity due to its dense computation, so the multi-scale features of the backbone network are not utilized in the original DETR. In addition, dense compu" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "AugDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "tation will lead to slow convergence. The original DETR required 500 epochs of training to achieve good performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.223 + ], + "angle": 0, + "content": "To exploit the multi-scale features of the backbone, Deformable DETR [50] proposes deformable attention. The multi-scale deformable self-attention can be simply formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.237, + 0.593, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\operatorname {C o n c a t} \\left(P _ {3}, P _ {4}, P _ {5}\\right), \\mathbf {A} = W _ {a} \\mathbf {Q},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.255, + 0.614, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {V} = \\operatorname {S a m p} \\left(W _ {v} \\mathbf {Q}, r + \\Delta r\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.274, + 0.785, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\operatorname {D e f o r m a b l e A t t e n t i o n} (\\mathbf {Q}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.784, + 0.361 + ], + "angle": 0, + "content": "where \\( P_{3}, P_{4}, P_{5} \\) are multi-scale features of the backbone, \\( \\mathbf{Q} \\) are a set of queries, \\( \\Delta r \\) are the offsets of reference points, \\( r \\) are the reference points, \\( \\mathbf{A} \\) are the attention weights of sample features, Samp is a function that extracts the features corresponding to the location \\( (r + \\Delta r) \\) by bilinear interpolation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.785, + 0.497 + ], + "angle": 0, + "content": "The deformable attention performs sparse computation that involves only a predefined number of surrounding features. Because deformable attention does not produce high complexity when combined with multi-scale features, it can utilize the multi-scale features of the backbone to detect multi-scale objects. In addition, sparse computation will lead to fast convergence. Deformable DETR converges 10 times faster than DETR. However, limited by sparse computation, deformable attention typically does not have a large receptive field and does not use global context information. This is not conducive to detecting large objects and objects in complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.785, + 0.664 + ], + "angle": 0, + "content": "To address these issues, we propose the Hybrid Attention Encoder that combines standard self-attention and deformable attention in the encoder layer. The hybrid attention layer is shown in Figure 3 (a). Introducing dense attention into the deformable attention encoder can enlarge the receptive field and perceive the global context. This will be more beneficial for multi-scale object detection. However, how to introduce dense attention into deformable attention without increasing too much complexity and slowing convergence is a key challenge. In order not to introduce too much complexity, we propose the hybrid-scale strategy. This strategy means that only the last scale features receive dense attention, not all scale features. With this strategy, our method does not introduce too much computation. Our hybrid attention can be simply formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.677, + 0.484, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {\\prime} = \\mathbf {K} ^ {\\prime} = \\mathbf {V} ^ {\\prime} = P _ {5},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.695, + 0.563, + 0.71 + ], + "angle": 0, + "content": "\\[\nP _ {5} ^ {\\prime} = \\operatorname {S e l f A t t e n t i o n} \\left(\\mathbf {Q} ^ {\\prime}, \\mathbf {K} ^ {\\prime}, \\mathbf {V} ^ {\\prime}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.715, + 0.593, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\operatorname {C o n c a t} \\left(P _ {3}, P _ {4}, P _ {5} ^ {\\prime}\\right), \\mathbf {A} = W _ {a} \\mathbf {Q},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.734, + 0.614, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {V} = \\operatorname {S a m p} \\left(W _ {v} \\mathbf {Q}, r + \\Delta r\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.752, + 0.784, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\text {D e f o r m a b l e A t t e n t i o n} (\\mathbf {Q}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In order not to slow down the convergence, we propose the hybrid-layer strategy. This strategy means that instead of adopting hybrid attention at all encoder layers, it is only adopted at some layers. From the ablation experiments, it can be seen that adopting hybrid attention in all the encoder layers will have no" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "J. Dong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "performance gain due to slow convergence. In our practical use, we introduce hybrid attention only in the first two encoder layers in order to balance complexity and performance. With the above design, we build the Hybrid Attention Encoder that can enlarge the receptive field and utilize the global context." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.228, + 0.533, + 0.243 + ], + "angle": 0, + "content": "3.3 Encoder-Mixing Cross-Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.788, + 0.298 + ], + "angle": 0, + "content": "The cross-attention in the decoder layer of Deformable DETR [50] uses multiscale deformable attention to achieve query-feature interaction. The multi-scale deformable cross-attention can be simply formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.308, + 0.684, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} = W _ {a} \\mathbf {Q}, \\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {V} = \\operatorname {S a m p} (W _ {v} E _ {6}, r + \\Delta r),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.326, + 0.785, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\operatorname {D e f o r m a b l e A t t e n t i o n} (\\mathbf {Q}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.785, + 0.395 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}\\) are a set of decoder queries, \\(\\mathbf{A}\\) is the attention weights, \\(\\Delta r\\) is the offsets of reference points, \\(r\\) is the reference points, \\(E_6\\) is the multi-scale features of last encoder layer, Samp is the sample function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.501 + ], + "angle": 0, + "content": "DETR-based detectors typically stack six encoder layers to enhance the backbone features. However, only features of the last encoder layer are utilized in cross-attention. Although features of the last encoder layer have larger receptive fields and stronger semantics, it is not optimal for multi-scale objects due to the fact that objects have different favors for features. Features from different encoder layers have different receptive fields, and if they are fully utilized they will be more beneficial for multi-scale object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.653 + ], + "angle": 0, + "content": "To address these issues, cross-attention interacting with multi-level encoder can obtain more discriminative features. However, using the same multi-level encoder for all objects is not conducive to multi-scale learning since multi-scale objects have different feature preferences. We propose Encoder-Mixing Cross-Attention (EMCA) to adaptively exploit features of different encoders based on query features. The EMCA is shown in Figure 3 (b). Specifically, we first extract features from each encoder layer by using multi-scale deformable attention. Then, we generate fusion weights for extracted features from each encoder based on the query features. Finally, we adaptively aggregate features from each encoder according to the fusion weights. The EMCA can be simply formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.663, + 0.63, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} = W _ {a} \\mathbf {Q}, \\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {w} _ {e} = \\sigma (W _ {e} \\mathbf {Q})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.682, + 0.559, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {l} = \\operatorname {S a m p} \\left(W _ {v l} E _ {l}, r + \\Delta r\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E M C A} (\\mathbf {Q}, \\mathbf {E}) = \\sum_ {l = 1} ^ {L} w _ {e} ^ {l} \\cdot \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V} _ {l}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.795 + ], + "angle": 0, + "content": "where \\(\\sigma\\) means the Sigmoid, \\(E_{l}\\) is the features of the \\(l\\) level encoder, \\(\\mathbf{E}\\) is a set of features from each level encoders, \\(w_{e}^{l}\\) is the weights of the \\(l\\) level encoder, \\(\\mathbf{w}_{e}\\) is a set of weights for all encoders. \\(L\\) is the number of encoder layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We use the EMCA to replace all the cross-attention in the decoder layer. With this design, each object can adaptively utilize multi-level encoder features to obtain more discriminative features for improved multi-scale learning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 2: Comparison with the state-of-the-art DETR variants on COCO val2017 with the ResNet-50-4scale backbone. * indicates the result is from our implemented model." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.186, + 0.787, + 0.527 + ], + "angle": 0, + "content": "
ModelBackbone#epochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Conditional-DETR [29]R5010843.064.045.722.746.761.5
SAM-DETR [41]R505039.861.841.620.543.459.6
Anchor-DETR [38]R505042.163.144.922.346.260.0
Dynamic-DETR [6]R501242.961.046.324.644.954.4
SMCA-DETR [10]R5010845.665.549.125.949.362.6
CF-DETR [3]R503647.866.552.431.250.662.8
Sparse-DETR [35]R505046.366.050.129.049.560.8
BoxeR-2D [30]R505050.067.954.730.952.862.6
Deformable-DETR [50]R505046.965.651.029.650.161.6
DAB-Deformable-DETR [24]R505046.866.050.429.149.862.3
DN-Deformable-DETR [18]R505048.667.452.731.052.063.7
H-DETR [14]R501248.766.452.931.251.563.5
Co-DETR [51]R501249.567.654.332.452.763.7
DINO [42]R501249.066.653.532.052.363.0
DINO [42]R502450.468.354.833.353.764.8
DDQ* [45]R501250.867.956.134.654.365.5
Focus DETR [46]R503650.468.555.034.053.564.4
AlignDETR [2]R501250.267.854.432.953.365.0
Stable-DINO [25]R501250.467.455.032.954.065.5
Rank-DETR [32]R501250.267.755.034.153.664.0
DINO+OursR501250.2 (+1.2)67.855.032.353.264.7
DINO+OursR502451.3 (+0.9)69.056.233.554.766.1
AlignDETR+OursR501251.3 (+1.1)68.855.833.354.866.6
DDQ+OursR501251.8 (+1.0)69.256.934.855.066.7
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.553, + 0.377, + 0.571 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.584, + 0.787, + 0.659 + ], + "angle": 0, + "content": "Dataset and Metrics. We conduct all experiments on the challenging COCO 2017 detection dataset. It is split into training set, validation set, and test-dev set with 115K, 5K, and 20K images respectively. We train models on the training set and evaluate performance on the validation set and test-dev set. All evaluated results follow the COCO-style Average Precision (AP) metrics." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.765 + ], + "angle": 0, + "content": "Implementation Details. We use DINO [42], Align DETR [2], and DDQ [45] as our baseline methods. All ablation studies are conducted on DINO-4scale with ResNet-50 for 12 epochs. We employ AdamW with \\(1 \\times 10^{-4}\\) weight decay as the optimizer to train our models. The batch size is set as 16 for all experiments. The initial learning rate is set as \\(1 \\times 10^{-4}\\) and it decreases by multiplying 0.1 after the 11th epoch for the 12 epochs setting. We use a random seed 0 in all our experiments. All other hyper-parameters follow the default from the codebase." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.786, + 0.377, + 0.801 + ], + "angle": 0, + "content": "4.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we evaluate AugDETR on the COCO val2017 set and compare it with the state-of-the-art DETR variants. All results with ResNet50 are sum-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Dong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.233, + 0.145, + 0.768, + 0.159 + ], + "angle": 0, + "content": "Table 3: Comparison with the DETR variants on val2017 set with the Swin-T." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.172, + 0.777, + 0.253 + ], + "angle": 0, + "content": "
MethodBackbone#epochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
DINO [42]Swin-T1251.369.056.034.554.466.0
OursSwin-T1252.3 (+1.0)70.257.036.855.367.2
AlignDETR [2]Swin-T1252.570.157.235.955.968.4
OursSwin-T1253.5 (+1.0)71.358.136.956.669.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.788, + 0.535 + ], + "angle": 0, + "content": "marized in Table 2. Combining our components with DINO [42], AugDETR achieves 50.2 AP, which is 1.2 AP higher than the DINO baseline with the same setting. Recent work on replacing the focal loss in DINO with the IoU-aware focal loss [19, 21, 43] has shown good performance. We chose a typical work AlignDETR [2] from these studies as baseline to validate the generality of our approach. Combining our components with AlignDETR, AugDETR achieves 51.3 AP, which is 1.1 AP higher than the AlignDETR baseline. Besides, we conduct experiments in the latest SOTA model DDQ [45]. Combining the our components with DDQ, AugDETR achieves 51.8 AP, which is 1.0 AP higher than the DDQ baseline. To demonstrate the effectiveness of our method under longer training, we conducted experiments under the \\(2 \\times\\) schedule. Our method achieves 51.3 AP, which is 0.9 AP higher than the baseline DINO. Besides, AugDETR can consistently achieve performance improvements even with more advanced backbone networks. As shown in Table 3, when using Swin-T [28] as the backbone, our method achieves performance improvements of 1.0 AP and 1.0 AP over DINO and AlignDETR, respectively. These results demonstrate the effectiveness and generalizability of our method." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.589 + ], + "angle": 0, + "content": "Table 4: Ablations on each component of AugDETR. \"HAE\" and \"EMCA\" represent Hybrid Attention Encoder, and Encoder-Mixing Cross-Attention, respectively." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.6, + 0.694, + 0.683 + ], + "angle": 0, + "content": "
HAEEMCAAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
49.066.653.532.052.363.0
✓49.667.054.231.852.963.8
✓49.867.454.633.052.864.7
✓✓50.267.855.032.353.264.7
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.728, + 0.342, + 0.741 + ], + "angle": 0, + "content": "4.2 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ablation studies on the effectiveness of each component. To analyze the contributions of individual components in AugDETR, Hybrid Attention Encoder (HAE), and Encoder-Mixing Cross-Attention (EMCA) are gradually introduced to the baseline. As shown in Table 4, HAE brings 0.6 AP improvement to the baseline model. This benefits from the HAE enlarges the receptive field and introduces the global context features to improve semantic representation. The" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "improvements of \\(AP_{M}\\) \\((+0.6\\mathrm{AP})\\) and \\(AP_{L}\\) \\((+0.8\\mathrm{AP})\\) contribute to the final performance improvement. These results are good demonstrations of our motivation and design. The EMCA improves the detection performance from 49.0 AP to 49.8 AP. It can be seen that the results on small, medium, and large objects are all improved, which means that object queries adaptively interacting with multi-level encoder features by cross-attention can extract more discriminative object features for improved multi-scale learning. When both components are included in the baseline model, it achieves 50.2 AP with 1.2 AP performance improvement. The results show that the two components are complementary and address different issues in DETR [4]." + }, + { + "type": "table_caption", + "bbox": [ + 0.261, + 0.326, + 0.741, + 0.34 + ], + "angle": 0, + "content": "Table 5: Comparisons of ours HAE with Deformable Encoder variants." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.353, + 0.75, + 0.434 + ], + "angle": 0, + "content": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)Flops
Deform. Enc. (4 Points)49.066.653.532.052.363.0245G
Deform. Enc. (6 Points)48.465.952.731.051.662.9250G
Deform. Enc. (8 Points)48.566.152.830.551.862.8255G
HAE (Ours)49.667.054.231.852.963.8248G
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Comparisons of HAE with Deformable Encoder variants. To make a fair comparison with the deformable encoder [50], we conduct experiments on increasing sampling points in the deformable encoder. The results are shown in Table 5. The performance of baseline with 4 sampling points is 49.0 AP. Performance decreases as sample points in the deformable encoder increases to 6 and 8, which are 48.4 AP and 48.5 AP, respectively. Even increasing the number of sample points in the deformable encoder does not improve performance for large and medium objects. This may be because deformable attention determines sampling weights based on query alone, leading to inaccurate weights as the number of points increases. However, our HAE can achieve performance gains (+0.6 AP) with less computational consumption (248G vs 250G). We think the possible reason is that the weight generation in standard self-attention, which calculates the similarity by query and key, is more suitable for long-distance sampling. These results support the motivation and design of our hybrid attention encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The number of Hybrid Attention Layers. Introducing too much self-attention may lead to slow convergence. We analyze the effects of the number of hybrid attention layers. The results are shown in Table 6. When only a hybrid attention layer is introduced, the model achieves 49.3 AP with 0.3 AP improvement. The improvement is mainly contributed by large and medium objects. Increasing the number of hybrid attention layers by 2 or 3 results in performance gains of 0.6 and 0.7 AP, respectively. It is still the large and medium objects that contribute the most performance gains. As the number of hybrid attention layers increases, the performance gains begin to decrease. In particular, there is no performance gain when increasing to 6 hybrid attention layers. This may be because the negative effect of slow convergence becomes dominant." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "J. Dong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.268, + 0.145, + 0.734, + 0.159 + ], + "angle": 0, + "content": "Table 6: Ablations on the number of hybrid attention layers in HAE." + }, + { + "type": "table", + "bbox": [ + 0.327, + 0.172, + 0.672, + 0.294 + ], + "angle": 0, + "content": "
NumberAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
-49.066.653.532.052.363.0
149.366.653.731.152.763.7
249.667.054.231.852.963.8
349.767.254.332.352.864.0
449.367.053.831.952.763.5
549.366.854.031.452.863.1
648.866.353.430.752.263.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.311, + 0.787, + 0.339 + ], + "angle": 0, + "content": "Table 7: Comparisons of ours EMCA with other decoder variants. * indicates that these models scale to the similar complexity as EMCA." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.351, + 0.756, + 0.46 + ], + "angle": 0, + "content": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_{2th} \\)\\( AP_{6th} \\)\\( AP_{10th} \\)Flops
DINO [42]49.066.653.534.243.245.7245G
+Iterative Enc [40]47.765.252.133.442.044.6245G
+Iterative Enc*47.465.151.633.242.044.1284G
+Memory Fusion [25]49.367.053.735.743.746.2253G
+Memory Fusion*49.366.953.935.743.546.2290G
+EMCA (Ours)49.867.454.636.044.546.6283G
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.787, + 0.672 + ], + "angle": 0, + "content": "Comparisons of EMCA with decoder variants. We compare our EMCA with other decoder variants. As shown in Table 7, using the Iterative Encoder [40] degrades the performance of the baseline. There are 0.3 AP performance gains when the Memory Fusion [25] is introduced, but there is no performance gain from the scaling of the approach. The performance gain of Memory Fusion is 0.5 AP less than our EMCA. These results demonstrate the effectiveness of our method design, which adaptively fuses multi-level encoder features based on query features for multi-scale objects, rather than using the same features for all objects. In addition, to show that our method converges faster, we also show results for the 2nd, 6th, and 10th epochs. It can be seen that our method outperforms the other methods throughout the training process, indicating the faster convergence of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Comparisons of EMCA with DINO variants. For a fair comparison, we compared with more complex baseline models due to the increased computational complexity introduced by EMCA. More complex baseline models are implemented by introducing more encoders and features. The results are shown in Table 8. By increasing the original baseline from 6 encoders to 7 encoders, the model reached 49.3 AP with 268G. When using 8 encoders, the model achieves 49.5 AP with 292G. While performance can be improved by stacking encoders, our EMCA utilizes encoder redundancy information and is a good complement to encoders. When using 10 encoders, the model suffers from overfitting and achieves only 47.8 AP with 339G. However, our EMCA with 10 encoders still has performance gains to reach 50.6 AP. This results show that our EMCA can" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.302, + 0.145, + 0.699, + 0.16 + ], + "angle": 0, + "content": "Table 8: Comparisons of ours EMCA with DINO variants." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.171, + 0.758, + 0.309 + ], + "angle": 0, + "content": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)Flops
6 encoder (Baseline)49.066.653.532.052.363.0245G
7 encoder49.366.753.732.052.463.7268G
8 encoder49.566.954.031.952.863.4292G
10 encoder47.864.852.130.650.961.5339G
5 scale49.466.953.832.352.563.9721G
6encoder+EMCA (Ours)49.867.454.633.052.864.7283G
8encoder+EMCA (Ours)50.267.854.632.453.364.6346G
10encoder+EMCA (Ours)50.668.155.333.553.965.3409G
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.784, + 0.4 + ], + "angle": 0, + "content": "mitigate the risk of overfitting. A very large amount of computation (721G) is required when 5 scales of backbone features are used, and then the performance reaches only 49.4 AP. These results show that our EMCA is a good complement to encoders and can mitigate the risk of overfitting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Quantitative analysis of EMCA. To analyze the weights of EMCA, we count the fusion weights in the last decoder layer from the COCO val2017 set. First, we analyzed the weights of different encoder layers, and the results are shown in Figure 4. It can be seen that our method learns different weights for encoder layers. Specifically, the first encoder layer plays the most important role, the 2nd, 3rd, and 4th encoder layers have decreasing importance, and the 4th, 5th, and 6th encoder layers have similar roles. As shown in Table 9, using the 2nd, 4th, and 6th encoder layers as inputs, our approach drops 0.3 AP and saves 23G of computational overhead. When only the 1st and 6th encoder layers are used, the computation is reduced by 30G while the performance is dropped by 0.2 AP. We call the EMCA with only the 1st and 6th encoder layers as EMCA-Lite. We then analyze the weights of the objects at different scales to illustrate that our method improves multi-scale learning. As shown in Figure 5, objects of different sizes are given different weights, and the larger the scale, the higher the weights. See the Appendix for object scale intervals. These results illustrate the effectiveness of fusion weights based on queries." + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.674, + 0.48, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.794, + 0.49, + 0.836 + ], + "angle": 0, + "content": "Fig. 4: Statistical fusion weights for different encoders. Layer #6 means the last-level encoder." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.674, + 0.776, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.794, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig.5: Statistical fusion weights for different scale objects. Size #1 to size #6 are ordered by scale from small to large." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "J. Dong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.49, + 0.172 + ], + "angle": 0, + "content": "Table 9: Ablations on the layer of input encoders in EMCA." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.185, + 0.472, + 0.281 + ], + "angle": 0, + "content": "
EncoderAPAP50AP75Flops
-49.066.653.5245G
12345649.867.454.6283G
24649.567.253.9260G
2649.367.253.8253G
1649.667.454.3253G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.145, + 0.785, + 0.172 + ], + "angle": 0, + "content": "Table 10: Ablations on the fusion methods in EMCA." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.185, + 0.78, + 0.281 + ], + "angle": 0, + "content": "
Fusion MethodAPAPSAPMAPL
-49.032.052.363.0
sum49.331.952.563.9
max48.630.951.663.2
sigmoid weighted49.833.052.864.7
softmax weighted49.532.152.864.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.789, + 0.591 + ], + "angle": 0, + "content": "Fusion Methods in EMCA. The fusion methods we explore include sum fusion, max fusion, and adaptive weighted fusion (sigmoid and softmax). All results are shown in Table 10. By using sum fusion, our method improves the baseline by 0.3 AP. This result shows that multi-level encoder features are richer in semantic information than single-level encoder features. When using max fusion, our method achieves 48.6 AP below the baseline. It can be seen that the performance decreases significantly for small and medium objects, but increases slightly for large objects. This may be because large objects tend to have higher activation so the features of small and medium objects are weakened during max fusion. When using adaptive weighted fusion, our method improves the baseline by 0.8 AP and 0.5 AP, respectively. It can be seen that the performance of small, medium, and large objects is all improved. In particular, the performance on large objects has improved significantly (+1.7 AP) by sigmoid weighted fusion. This may be because the larger the object, the more information is activated at all levels of the encoder, and therefore the more information can be used. In addition, the corresponding query features of large objects are richer so that the learned fusion weights are more accurate. These results suggest that sigmoid weighted fusion is a better fusion method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.36, + 0.651 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we analyze the limitations of DETR detectors for multi-scale learning and propose local feature enhancement and multi-level encoder exploitation techniques to address the limitations. Based on these, we construct a novel detection transformer detector named Augmented DETR (AugDETR), which consists of two components: Hybrid Attention Encoder and Encoder-Mixing Cross-Attention. HAE enlarges the receptive field and introduces global context to enhance the local feature of the deformable encoder. EMCA uses multi-level encoders to obtain more discriminative features and speed up convergence. Extensive experiments with various DETR-based detectors and backbones validate the effectiveness of our method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "AugDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.302 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (NSFC) under Grant 62088102. This work is partly supported by China Telecom Corporation Ltd. Data&Al Technology Company, Beijing, China and China Telecom - Xi'an Jiaotong University Joint Innovation Institute for Science-Education Integration of Intelligent Cloud Network. We also appreciate Jingwen Fu, He Zhang and Tao Yang for their insightful discussions. We thank all the anonymous reviewers and Area Chairs for their constructive and helpful comments, which have significantly improved the quality of the paper." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.328, + 0.323, + 0.344 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.785, + 0.404 + ], + "angle": 0, + "content": "1. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154-6162 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "2. Cai, Z., Liu, S., Wang, G., Ge, Z., Zhang, X., Huang, D.: Align-detr: Improving detr with simple iou-aware bce loss. arXiv preprint arXiv:2304.07527 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "3. Cao, X., Yuan, P., Feng, B., Niu, K.: Cf-detr: Coarse-to-fine transformers for end-to-end object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 185-193 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.517 + ], + "angle": 0, + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.544 + ], + "angle": 0, + "content": "5. Chen, Q., Chen, X., Zeng, G., Wang, J.: Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.587 + ], + "angle": 0, + "content": "6. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2988-2997 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.629 + ], + "angle": 0, + "content": "7. Ding, X., Zhang, X., Han, J., Ding, G.: Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11963-11975 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.785, + 0.672 + ], + "angle": 0, + "content": "8. Dong, J., Huang, Y., Zhang, S., Chen, S., Zheng, N.: Construct effective geometry aware feature pyramid network for multi-scale object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 534-541 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.713 + ], + "angle": 0, + "content": "9. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.715, + 0.785, + 0.755 + ], + "angle": 0, + "content": "10. Gao, P., Zheng, M., Wang, X., Dai, J., Li, H.: Fast convergence of detr with spatially modulated co-attention. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3621-3630 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "11. Ghiasi, G., Lin, T.Y., Le, Q.V.: Nas-fpn: Learning scalable feature pyramid architecture for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7036-7045 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Guo, C., Fan, B., Zhang, Q., Xiang, S., Pan, C.: Augfpn: Improving multi-scale feature learning for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12595-12604 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.362, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Dong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. He, Y., Zhu, C., Wang, J., Savvides, M., Zhang, X.: Bounding box regression with uncertainty for accurate object detection. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 2888-2897 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "14. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: Detrs with hybrid matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19702-19712 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "15. Kim, S.W., Kook, H.K., Sun, J.Y., Kang, M.C., Ko, S.J.: Parallel feature pyramid network for object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 234-250 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "16. Kong, T., Sun, F., Liu, H., Jiang, Y., Li, L., Shi, J.: Foveabox: Beyound anchor-based object detection. IEEE Transactions on Image Processing 29, 7389-7398 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.787, + 0.338 + ], + "angle": 0, + "content": "17. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. In: Proceedings of the European conference on computer vision (ECCV). pp. 734-750 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "18. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.787, + 0.433 + ], + "angle": 0, + "content": "19. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems 33, 21002-21012 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "20. Lin, T.Y., Dólar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "21. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.787, + 0.569 + ], + "angle": 0, + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "23. Lin, Y., Yuan, Y., Zhang, Z., Li, C., Zheng, N., Hu, H.: Detr does not need multiscale or locality design. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6545-6554 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "24. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.65, + 0.787, + 0.69 + ], + "angle": 0, + "content": "25. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., et al.: Detection transformer with stable matching. arXiv preprint arXiv:2304.04742 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.69, + 0.787, + 0.731 + ], + "angle": 0, + "content": "26. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8759-8768 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.731, + 0.787, + 0.786 + ], + "angle": 0, + "content": "27. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 21-37. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "AugDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "30. Nguyen, D.K., Ju, J., Booij, O., Oswald, M.R., Snoek, C.G.: Boxer: Box-attention for 2d and 3d transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4773-4782 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "31. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra r-cnn: Towards balanced learning for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 821-830 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "32. Pu, Y., Liang, W., Hao, Y., Yuan, Y., Yang, Y., Zhang, C., Hu, H., Huang, G.: Rank-detr for high quality object detection. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.357 + ], + "angle": 0, + "content": "33. Qiao, S., Chen, L.C., Yuille, A.: Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10213-10224 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.398 + ], + "angle": 0, + "content": "34. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 39(6), 1137-1149 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "35. Roh, B., Shin, J., Shin, W., Kim, S.: Sparse detr: Efficient end-to-end object detection with learnable sparsity. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.481 + ], + "angle": 0, + "content": "36. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "37. Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9627-9636 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.563 + ], + "angle": 0, + "content": "38. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "39. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.66 + ], + "angle": 0, + "content": "40. Zhang, G., Luo, Z., Tian, Z., Zhang, J., Zhang, X., Lu, S.: Towards efficient use of multi-scale features in transformer-based object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6206-6216 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "41. Zhang, G., Luo, Z., Yu, Y., Cui, K., Lu, S.: Accelerating detr convergence via semantic-aligned matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 949-958 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "42. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.744, + 0.788, + 0.785 + ], + "angle": 0, + "content": "43. Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8514-8523 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.84 + ], + "angle": 0, + "content": "44. Zhang, H., Chang, H., Ma, B., Wang, N., Chen, X.: Dynamic r-cnn: Towards high quality object detection via dynamic training. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 260-275. Springer (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "J. Dong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "45. Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7329-7338 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "46. Zheng, D., Dong, W., Hu, H., Chen, X., Wang, Y.: Less is more: Focus attention for efficient detr. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6674-6683 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.273 + ], + "angle": 0, + "content": "47. Zhou, X., Wang, D., Krähenbuhl, P.: Objects as points. arXiv preprint arXiv:1904.07850 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "48. Zhou, X., Zhuo, J., Krahenbuhl, P.: Bottom-up object detection by grouping extreme and center points. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 850-859 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "49. Zhu, C., He, Y., Savvides, M.: Feature selective anchor-free module for single-shot object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 840-849 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.396 + ], + "angle": 0, + "content": "50. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.787, + 0.44 + ], + "angle": 0, + "content": "51. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6748-6758 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.44 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_origin.pdf b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..698d542d16ecd6e6ae14517f947ba3a42c437f23 --- /dev/null +++ b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/1ec0e321-b82f-41e0-8f96-ae2631ea464b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0d117f5a07a19c9f05c734f307445b9b8f814265097e9febb186c860ac97390 +size 2103662 diff --git a/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/full.md b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3ffeb33df96234d54686b75c8f24464e9acd52f2 --- /dev/null +++ b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/full.md @@ -0,0 +1,313 @@ +# AugDETR: Improving Multi-scale Learning for Detection Transformer + +Jinpeng Dong, Yutong Lin, Chen Li, Sanping Zhou, Nanning Zheng* + +National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University {djp1235a, yutonglin, edward82}@stu.xjtu.edu.cn {spzhou, nnzheng} $@$ xjtu.edu.cn + +Abstract. Current end-to-end detectors typically exploit transformers to detect objects and show promising performance. Among them, Deformable DETR is a representative paradigm that effectively exploits multi-scale features. However, small local receptive fields and limited query-encoder interactions weaken multi-scale learning. In this paper, we analyze local feature enhancement and multi-level encoder exploitation for improved multi-scale learning and construct a novel detection transformer detector named Augmented DETR (AugDETR) to realize them. Specifically, AugDETR consists of two components: Hybrid Attention Encoder and Encoder-Mixing Cross-Attention. Hybrid Attention Encoder enlarges the receptive field of the deformable encoder and introduces global context features to enhance feature representation. Encoder-Mixing Cross-Attention adaptively leverages multi-level encoders based on query features for more discriminative object features and faster convergence. By combining AugDETR with DETR-based detectors such as DINO, AlignDETR, DDQ, our models achieve performance improvements of 1.2, 1.1, and 1.0 AP in the COCO under the ResNet-50-4scale and 12 epochs setting, respectively. + +Keywords: Object detection $\cdot$ Detection transformer $\cdot$ Hybrid attention $\cdot$ Multi-level encoder + +# 1 Introduction + +Many CNN-based detectors [1,12,20,27,34,37,47] have been proposed to improve detection performance in the past decade. Despite their promising progress, these detectors suffer from many hand-crafted components, such as non-maximum suppression and anchor design. In contrast, CNN-based paradigm is broken by DETR [4], a novel transformer detector. It gets rid of hand-crafted components and implements an end-to-end paradigm. Despite its novel paradigm, DETR suffers from some limitations, including slow convergence and inferior performance. To address these limitations, many methods [2, 6, 14, 24, 25, 29, 42, 50] + +have been proposed from different perspectives, such as introducing spatial priors, hybrid matching training, designing efficient attention modules, assigning sample weights, etc. With these optimizations, DINO [42] achieved a new record on the COCO benchmark [22]. + +Although DETR-based detectors have achieved promising performance, multiscale object detection has not received sufficient attention so far, which may limit the performance of detectors. Scale variation of objects remains one of the crucial challenges in object detection. Objects at different scales require inconsistent semantic features, e.g., large objects require features with larger receptive fields. To the best of our knowledge, Deformable-DETR [50] has made some effort to remedy this issue. Deformable-DETR uses multi-scale backbone features to improve multi-scale learning by deformable attention. In addition to the use of multi-scale backbone features, we believe that the multi-scale learning of Deformable-DETR has not been fully exploited. As shown in Table 1, although Deformable DETR is 2.5 AP higher than the DETR in overall performance, the performance on large objects is still 1.5 AP lower than the DETR. We present two techniques that can further improve multi-scale learning as follows: + +Table 1: Performance comparison between DETR and Deformable DETR. + +
MethodEpochsAPAPSAPMAPL
DETR [4]50042.020.545.861.1
Deformable DETR [50]5044.527.147.659.6
+ +![](images/fb497dd95c6ea9c3b931e0bed1014de3ee7a801f9a8ee6cc6daa3c33284e6d5d.jpg) +Fig. 1: Comparison of receptive fields of different encoders. + +![](images/525e51d10ad7af2bec6a9d5e98da01acc439e4779e73eabc0ce5f33971791407.jpg) + +Local feature enhancement. Deformable-DETR [50] replaces the original self-attention (global dense attention) with deformable attention (sparse attention) to reduce the complexity and slow convergence of dense attention. Deformable attention is implemented by a fixed number of deformable sampling points. Limited by local sampling points, deformable attention is difficult to achieve a larger receptive field compared to self-attention. This is not conducive to the detection of large objects. Further stacking of deformable attention still improves performance, as analyzed in the ablations. In addition, deformable attention loses the ability to perceive global context, while multi-scale object detection in some complex scenes may rely on global context information. Increasing the sampling points degrades the performance of the deformable attention, as analyzed in the ablations. Therefore, we introduce global contextual information into the deformable encoder to enhance multi-scale learning. + +Multi-level encoder exploitation. DETR-based detectors typically stack six encoder layers and six decoder layers. Encoder layers enhance the backbone features, and decoder layers extract features from last encoder layer to decode + +queries. Previous methods ignore the use of multi-level encoder features. More discriminative object features can be obtained by fully exploiting multi-level encoder features with different receptive fields and feature preferences. We refer to RepLKNet [7] for analysis of receptive field. As shown in Figure 1, the comparison of encoder 1 and 6 in DINO [42] shows that different encoders have varying feature preferences and receptive fields. In addition, the fact that multilevel encoders are all involved in object feature generation can speed up convergence because the supervisory signals are available to multiple levels of encoders, rather than being available only to the last level encoder. Utilizing multi-level encoder features also can introduce redundancy to mitigate the risk of overfitting. Although multi-level encoder features is beneficial for object detection, using multi-level encoder features fixed for all objects is not conducive to multiscale learning. Since objects at different scales have different feature preferences, adaptive use of multi-level encoder features by learning weights based on query features will facilitate multi-scale learning for DETR-based detectors. + +In this paper, we propose AugDETR, a simple yet effective DETR-based detector that uses two different components to improve multi-scale learning. First, Hybrid Attention Encoder is proposed to augment local features in the deformable encoder by partially combining dense attention with deformable attention. Hybrid Attention Encoder can enlarge the receptive field and introduce global context, which can improve detection performance on large objects. Second, Encoder-Mixing Cross-Attention is introduced to better exploit features from different encoder layers by weighted aggregation. Objects at different scales can learn different aggregation weights based on object features to facilitate multi-scale learning. With this design, Encoder-Mixing Cross-Attention can generate more discriminative object features for subsequent classification and localization tasks. We validate the effectiveness of AugDETR on the COCO benchmark. Our method achieves 50.2 AP (+1.2 AP), 51.3 AP (+1.1 AP), and 51.8 AP (+1.0 AP) based on DINO [42], AlignDETR [2], and DDQ [45] with ResNet-50-4scale under 12 epochs settings, respectively. + +We summarize our contributions as follows: + +- We analyze local feature enhancement and multi-level encoder exploitation techniques for improved multi-scale learning. +- A new detection transformer detector named AugDETR is proposed to achieve the techniques with Hybrid Attention Encoder and Encoder-Mixing Cross-Attention. +- We validate AugDETR equipped with various DETR-based detectors and backbones on the COCO benchmark, and it consistently yields significant performance improvements over DETR-based detectors. + +# 2 Related Work + +# 2.1 Object Detection + +Early deep-learning detectors can be roughly classified into anchor-based and anchor-free methods. Anchor-based methods use predefined anchors to locate + +objects, including Faster R-CNN [34], SSD [27], RetinaNet [21], and their variants [1, 13, 31, 44]. Anchor-free methods use points near object centers to locate objects, including CenterNet [47], FCOS [37], and others [9, 16, 17, 48, 49]. + +Recently, DETR [4] proposes a simple transformer pipeline for object detection, effectively alleviating the need for heuristic components. Compared to CNN-based work, DETR is rather novel but still suffers from some limitations, including slow convergence and inferior performance. Many follow-up works focus on alleviating the limitations of DETR from different perspectives. Conditional DETR [29] proposes to decouple the context embedding and position embedding, which helps DETR speed up convergence. Anchor DETR [38] uses the anchor points as the initial queries for fast convergence. DAB DETR [24] uses 4D anchor boxes that are dynamically updated with the decoder layer to represent queries. DN DETR [18] proposes a denoising training mechanism to speed up convergence. DINO [42] further extends DN DETR with contrastive denoising training, mixed query, and look-forward twice techniques. Group DETR [5], HDETR [14], and CO-DETR [51] propose various one-to-one matching and one-to-many matching combination methods to stabilize the matching process of DETR. Align DETR [2], Stable DINO [25], and Rank DETR [32] design different schemes to weight the samples by the IoU-aware classification loss. Plain DETR [23] proposes a box-to-pixel relative position bias and a masked image modeling [39] pre-training scheme to improve the DETR detector. Unlike these methods that design spatial priors, hybrid matching, sample weight assignments, etc., our approach focuses on improving the multi-scale learning of DETR detectors. In addition, we conducted experiments using some SOTA methods as the baseline models. Our techniques are complementary to theirs. + +# 2.2 Multi-Scale Learning for Object Detection + +In the early era, image pyramid is a common solution to improve multi-scale learning. Instead of image pyramid, some methods utilize multi-scale features to mitigate scale variation. A representative work among these methods is the feature pyramid network (FPN [20]), which proposes a top-down pathway and lateral connections to construct the feature pyramid. Inspired by FPN, multi-scale feature extraction has been widely studied, such as BiFPN [36], PAFPN [26], NAS-FPN [11], AugFPN [12], and other variants [8,15,33]. Recently, some works have attempted to design multi-scale feature extraction modules in DETR. Deformable DETR [50] proposes deformable attention to utilize multi-scale backbone features. Dynamic DETR [6] proposes a CNN-based dynamic encoder with scale-aware attention, spatial-aware attention, and task-aware attention. SMCA [10] generates the Gaussian-like weight map as the spatial prior and the scale weights as the scale prior to modulated multi-scale features. IMFA [40] designs a new scheme to adaptively sample sparse multi-scale features. Different from them, our approach enhances multi-scale learning by introducing global attention to expand the receptive field and integrate global context. To use global attention effectively, we propose hybrid scale and hybrid layer strategies to construct the hybrid attention encoder. In addition, we adaptively use multi-level + +encoder features based on query features to improve multi-scale learning, which is also distinctly different from previous methods. + +![](images/6895ecd5450bb6c6fd1c867f6df6a2f3535d04fac7a5c463ac50becb05f28ad1.jpg) +Fig. 2: Left: The pipeline of conventional DETR-based detectors. The backbone network and the stacked encoder layers are used to extract and enhance the image features. Then, the object queries utilize the decoder layers to interact with the enhanced image features. Finally, the interacted object queries are fed into the detection head to obtain the predictions. Only the features from the last encoder layer can be used to interact with the object queries. Right: The pipeline of Our AugDETR. Our AugDETR restructures the conventional encoder-decoder pipeline. The object queries can interact with features from all encoder layers. For clarity, only the 3-layer structure is shown. + +![](images/af94f65b0cfca1973e3f69f84960b73ca9b669627c87dbd622f987dedaf472b0.jpg) + +# 3 Method + +In section 3.1, we first briefly revisit the detection transformer and then introduce the pipeline of AugDETR. In section 3.2, we describe the details of the proposed Hybrid Attention Encoder. In section 3.3, we introduce the Encoder-Mixing Cross-Attention used in the transformer decoder. + +# 3.1 A Revisit of DETR-based Detector + +Since our proposed method is constructed on top of the recent DETR-based object detectors, we first briefly review the pipeline of the DETR-based detectors, taking the Deformable DETR [50] as an example. + +The pipeline of Deformable DETR is shown in Figure 2 (left). Deformable DETR consists of 3 subcomponents: a backbone, a encoder, and a decoder. Given an input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , a backbone network generates multi-scale features. We denote the multi-scale features as $\{C_2, C_3, C_4, C_5\}$ . The spatial resolutions of these features are usually $1/4^2$ , $1/8^2$ , $1/16^2$ , and $1/32^2$ of the given image. Then these features reduced to the same channel by a linear projection are denoted as $\{P_2, P_3, P_4, P_5\}$ . Next, a transformer encoder is used to further enhance these + +multi-scale features. The transformer encoder is typically stacked with 6 encoder layers. The encoder layer mainly consists of a self-attention and a feed-forward network. For Deformable DETR, multi-scale deformable attention is employed as self-attention to enhance multi-scale features. The encoded features of the different layers in the encoder are denoted as $\{E_1,E_2,E_3,E_4,E_5,E_6\}$ . Finally, a transformer decoder is used to produce object predictions from a set of object queries. Typically, only the last level of encoder features are fed into the decoder to interact with the query. The transformer decoder is also typically stacked with 6 decoder layers for better detection performance. The decoder layer mainly consists of a self-attention, a cross-attention, and a feed-forward network. For deformable DETR, multi-scale deformable attention is used as cross-attention to achieve query-feature interaction. As shown in Figure 2 (right), our pipeline changes the fact that the decoder can utilize only the last level of encoder features, but rather all encoder features can be utilized by the decoder. This design lays the foundation for cross-attention to utilize multi-level encoder features. + +![](images/da531c7b14acbdc100eff6e9ddeff03fc1cce099a8797bc45fd719259127b523.jpg) +Fig. 3: (a) is the process of hybrid attention in the Hybrid Attention Layer. Dense attention is first applied to only the top features, and then sparse attention is applied to the multi-scale features. (b) is the details of Encoder-Mixing Cross-Attention. Deformable attention extracts multiple object features from multi-level encoder and then the extracted features are aggregated based on the adaptive fusion weights learned from the query features. + +![](images/9296197508af022d0cdbf0a2fea8b4c9ea6db2b72940e377c5b285df478185cd.jpg) + +# 3.2 Hybrid Attention Encoder + +The encoder layer in the original DETR [4] uses the self-attention as follows: + +$$ +\mathbf {Q} = W _ {q} P _ {5}, \mathbf {K} = W _ {k} P _ {5}, \mathbf {V} = W _ {v} P _ {5}, +$$ + +$$ +\operatorname {S e l f A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t m a x} (\mathbf {Q K}) \mathbf {V}, \tag {1} +$$ + +where $P_{5}$ are the last level features of the backbone network; $\mathbf{Q}$ , $\mathbf{K}$ , and $\mathbf{V}$ are query, key, and value features, respectively. + +The standard self-attention performs dense computation that involves all features. Applying standard self-attention to a large feature map will cause high complexity due to its dense computation, so the multi-scale features of the backbone network are not utilized in the original DETR. In addition, dense compu + +tation will lead to slow convergence. The original DETR required 500 epochs of training to achieve good performance. + +To exploit the multi-scale features of the backbone, Deformable DETR [50] proposes deformable attention. The multi-scale deformable self-attention can be simply formulated as: + +$$ +\mathbf {Q} = \operatorname {C o n c a t} \left(P _ {3}, P _ {4}, P _ {5}\right), \mathbf {A} = W _ {a} \mathbf {Q}, +$$ + +$$ +\Delta r = W _ {p} \mathbf {Q}, \mathbf {V} = \operatorname {S a m p} \left(W _ {v} \mathbf {Q}, r + \Delta r\right), +$$ + +$$ +\operatorname {D e f o r m a b l e A t t e n t i o n} (\mathbf {Q}, \mathbf {V}) = \operatorname {S o f t m a x} (\mathbf {A}) \mathbf {V}, \tag {2} +$$ + +where $P_{3}, P_{4}, P_{5}$ are multi-scale features of the backbone, $\mathbf{Q}$ are a set of queries, $\Delta r$ are the offsets of reference points, $r$ are the reference points, $\mathbf{A}$ are the attention weights of sample features, Samp is a function that extracts the features corresponding to the location $(r + \Delta r)$ by bilinear interpolation. + +The deformable attention performs sparse computation that involves only a predefined number of surrounding features. Because deformable attention does not produce high complexity when combined with multi-scale features, it can utilize the multi-scale features of the backbone to detect multi-scale objects. In addition, sparse computation will lead to fast convergence. Deformable DETR converges 10 times faster than DETR. However, limited by sparse computation, deformable attention typically does not have a large receptive field and does not use global context information. This is not conducive to detecting large objects and objects in complex scenes. + +To address these issues, we propose the Hybrid Attention Encoder that combines standard self-attention and deformable attention in the encoder layer. The hybrid attention layer is shown in Figure 3 (a). Introducing dense attention into the deformable attention encoder can enlarge the receptive field and perceive the global context. This will be more beneficial for multi-scale object detection. However, how to introduce dense attention into deformable attention without increasing too much complexity and slowing convergence is a key challenge. In order not to introduce too much complexity, we propose the hybrid-scale strategy. This strategy means that only the last scale features receive dense attention, not all scale features. With this strategy, our method does not introduce too much computation. Our hybrid attention can be simply formulated as: + +$$ +\mathbf {Q} ^ {\prime} = \mathbf {K} ^ {\prime} = \mathbf {V} ^ {\prime} = P _ {5}, +$$ + +$$ +P _ {5} ^ {\prime} = \operatorname {S e l f A t t e n t i o n} \left(\mathbf {Q} ^ {\prime}, \mathbf {K} ^ {\prime}, \mathbf {V} ^ {\prime}\right), +$$ + +$$ +\mathbf {Q} = \operatorname {C o n c a t} \left(P _ {3}, P _ {4}, P _ {5} ^ {\prime}\right), \mathbf {A} = W _ {a} \mathbf {Q}, +$$ + +$$ +\Delta r = W _ {p} \mathbf {Q}, \mathbf {V} = \operatorname {S a m p} \left(W _ {v} \mathbf {Q}, r + \Delta r\right), +$$ + +$$ +\text {D e f o r m a b l e A t t e n t i o n} (\mathbf {Q}, \mathbf {V}) = \operatorname {S o f t m a x} (\mathbf {A}) \mathbf {V}, \tag {3} +$$ + +In order not to slow down the convergence, we propose the hybrid-layer strategy. This strategy means that instead of adopting hybrid attention at all encoder layers, it is only adopted at some layers. From the ablation experiments, it can be seen that adopting hybrid attention in all the encoder layers will have no + +performance gain due to slow convergence. In our practical use, we introduce hybrid attention only in the first two encoder layers in order to balance complexity and performance. With the above design, we build the Hybrid Attention Encoder that can enlarge the receptive field and utilize the global context. + +# 3.3 Encoder-Mixing Cross-Attention + +The cross-attention in the decoder layer of Deformable DETR [50] uses multiscale deformable attention to achieve query-feature interaction. The multi-scale deformable cross-attention can be simply formulated as: + +$$ +\mathbf {A} = W _ {a} \mathbf {Q}, \Delta r = W _ {p} \mathbf {Q}, \mathbf {V} = \operatorname {S a m p} (W _ {v} E _ {6}, r + \Delta r), +$$ + +$$ +\operatorname {D e f o r m a b l e A t t e n t i o n} (\mathbf {Q}, \mathbf {V}) = \operatorname {S o f t m a x} (\mathbf {A}) \mathbf {V}, \tag {4} +$$ + +where $\mathbf{Q}$ are a set of decoder queries, $\mathbf{A}$ is the attention weights, $\Delta r$ is the offsets of reference points, $r$ is the reference points, $E_6$ is the multi-scale features of last encoder layer, Samp is the sample function. + +DETR-based detectors typically stack six encoder layers to enhance the backbone features. However, only features of the last encoder layer are utilized in cross-attention. Although features of the last encoder layer have larger receptive fields and stronger semantics, it is not optimal for multi-scale objects due to the fact that objects have different favors for features. Features from different encoder layers have different receptive fields, and if they are fully utilized they will be more beneficial for multi-scale object detection. + +To address these issues, cross-attention interacting with multi-level encoder can obtain more discriminative features. However, using the same multi-level encoder for all objects is not conducive to multi-scale learning since multi-scale objects have different feature preferences. We propose Encoder-Mixing Cross-Attention (EMCA) to adaptively exploit features of different encoders based on query features. The EMCA is shown in Figure 3 (b). Specifically, we first extract features from each encoder layer by using multi-scale deformable attention. Then, we generate fusion weights for extracted features from each encoder based on the query features. Finally, we adaptively aggregate features from each encoder according to the fusion weights. The EMCA can be simply formulated as: + +$$ +\mathbf {A} = W _ {a} \mathbf {Q}, \Delta r = W _ {p} \mathbf {Q}, \mathbf {w} _ {e} = \sigma (W _ {e} \mathbf {Q}) +$$ + +$$ +\mathbf {V} _ {l} = \operatorname {S a m p} \left(W _ {v l} E _ {l}, r + \Delta r\right), +$$ + +$$ +\operatorname {E M C A} (\mathbf {Q}, \mathbf {E}) = \sum_ {l = 1} ^ {L} w _ {e} ^ {l} \cdot \operatorname {S o f t m a x} (\mathbf {A}) \mathbf {V} _ {l}, \tag {5} +$$ + +where $\sigma$ means the Sigmoid, $E_{l}$ is the features of the $l$ level encoder, $\mathbf{E}$ is a set of features from each level encoders, $w_{e}^{l}$ is the weights of the $l$ level encoder, $\mathbf{w}_{e}$ is a set of weights for all encoders. $L$ is the number of encoder layers. + +We use the EMCA to replace all the cross-attention in the decoder layer. With this design, each object can adaptively utilize multi-level encoder features to obtain more discriminative features for improved multi-scale learning. + +Table 2: Comparison with the state-of-the-art DETR variants on COCO val2017 with the ResNet-50-4scale backbone. * indicates the result is from our implemented model. + +
ModelBackbone#epochsAP\( AP_{50} \)\( AP_{75} \)\( AP_S \)\( AP_M \)\( AP_L \)
Conditional-DETR [29]R5010843.064.045.722.746.761.5
SAM-DETR [41]R505039.861.841.620.543.459.6
Anchor-DETR [38]R505042.163.144.922.346.260.0
Dynamic-DETR [6]R501242.961.046.324.644.954.4
SMCA-DETR [10]R5010845.665.549.125.949.362.6
CF-DETR [3]R503647.866.552.431.250.662.8
Sparse-DETR [35]R505046.366.050.129.049.560.8
BoxeR-2D [30]R505050.067.954.730.952.862.6
Deformable-DETR [50]R505046.965.651.029.650.161.6
DAB-Deformable-DETR [24]R505046.866.050.429.149.862.3
DN-Deformable-DETR [18]R505048.667.452.731.052.063.7
H-DETR [14]R501248.766.452.931.251.563.5
Co-DETR [51]R501249.567.654.332.452.763.7
DINO [42]R501249.066.653.532.052.363.0
DINO [42]R502450.468.354.833.353.764.8
DDQ* [45]R501250.867.956.134.654.365.5
Focus DETR [46]R503650.468.555.034.053.564.4
AlignDETR [2]R501250.267.854.432.953.365.0
Stable-DINO [25]R501250.467.455.032.954.065.5
Rank-DETR [32]R501250.267.755.034.153.664.0
DINO+OursR501250.2 (+1.2)67.855.032.353.264.7
DINO+OursR502451.3 (+0.9)69.056.233.554.766.1
AlignDETR+OursR501251.3 (+1.1)68.855.833.354.866.6
DDQ+OursR501251.8 (+1.0)69.256.934.855.066.7
+ +# 4 Experiments + +Dataset and Metrics. We conduct all experiments on the challenging COCO 2017 detection dataset. It is split into training set, validation set, and test-dev set with 115K, 5K, and 20K images respectively. We train models on the training set and evaluate performance on the validation set and test-dev set. All evaluated results follow the COCO-style Average Precision (AP) metrics. + +Implementation Details. We use DINO [42], Align DETR [2], and DDQ [45] as our baseline methods. All ablation studies are conducted on DINO-4scale with ResNet-50 for 12 epochs. We employ AdamW with $1 \times 10^{-4}$ weight decay as the optimizer to train our models. The batch size is set as 16 for all experiments. The initial learning rate is set as $1 \times 10^{-4}$ and it decreases by multiplying 0.1 after the 11th epoch for the 12 epochs setting. We use a random seed 0 in all our experiments. All other hyper-parameters follow the default from the codebase. + +# 4.1 Main Results + +In this section, we evaluate AugDETR on the COCO val2017 set and compare it with the state-of-the-art DETR variants. All results with ResNet50 are sum- + +Table 3: Comparison with the DETR variants on val2017 set with the Swin-T. + +
MethodBackbone#epochsAP\( AP_{50} \)\( AP_{75} \)\( AP_S \)\( AP_M \)\( AP_L \)
DINO [42]Swin-T1251.369.056.034.554.466.0
OursSwin-T1252.3 (+1.0)70.257.036.855.367.2
AlignDETR [2]Swin-T1252.570.157.235.955.968.4
OursSwin-T1253.5 (+1.0)71.358.136.956.669.1
+ +marized in Table 2. Combining our components with DINO [42], AugDETR achieves 50.2 AP, which is 1.2 AP higher than the DINO baseline with the same setting. Recent work on replacing the focal loss in DINO with the IoU-aware focal loss [19, 21, 43] has shown good performance. We chose a typical work AlignDETR [2] from these studies as baseline to validate the generality of our approach. Combining our components with AlignDETR, AugDETR achieves 51.3 AP, which is 1.1 AP higher than the AlignDETR baseline. Besides, we conduct experiments in the latest SOTA model DDQ [45]. Combining the our components with DDQ, AugDETR achieves 51.8 AP, which is 1.0 AP higher than the DDQ baseline. To demonstrate the effectiveness of our method under longer training, we conducted experiments under the $2 \times$ schedule. Our method achieves 51.3 AP, which is 0.9 AP higher than the baseline DINO. Besides, AugDETR can consistently achieve performance improvements even with more advanced backbone networks. As shown in Table 3, when using Swin-T [28] as the backbone, our method achieves performance improvements of 1.0 AP and 1.0 AP over DINO and AlignDETR, respectively. These results demonstrate the effectiveness and generalizability of our method. + +Table 4: Ablations on each component of AugDETR. "HAE" and "EMCA" represent Hybrid Attention Encoder, and Encoder-Mixing Cross-Attention, respectively. + +
HAEEMCAAP\( AP_{50} \)\( AP_{75} \)\( AP_S \)\( AP_M \)\( AP_L \)
49.066.653.532.052.363.0
✓49.667.054.231.852.963.8
✓49.867.454.633.052.864.7
✓✓50.267.855.032.353.264.7
+ +# 4.2 Ablations + +Ablation studies on the effectiveness of each component. To analyze the contributions of individual components in AugDETR, Hybrid Attention Encoder (HAE), and Encoder-Mixing Cross-Attention (EMCA) are gradually introduced to the baseline. As shown in Table 4, HAE brings 0.6 AP improvement to the baseline model. This benefits from the HAE enlarges the receptive field and introduces the global context features to improve semantic representation. The + +improvements of $AP_{M}$ $(+0.6\mathrm{AP})$ and $AP_{L}$ $(+0.8\mathrm{AP})$ contribute to the final performance improvement. These results are good demonstrations of our motivation and design. The EMCA improves the detection performance from 49.0 AP to 49.8 AP. It can be seen that the results on small, medium, and large objects are all improved, which means that object queries adaptively interacting with multi-level encoder features by cross-attention can extract more discriminative object features for improved multi-scale learning. When both components are included in the baseline model, it achieves 50.2 AP with 1.2 AP performance improvement. The results show that the two components are complementary and address different issues in DETR [4]. + +Table 5: Comparisons of ours HAE with Deformable Encoder variants. + +
MethodAP\( AP_{50} \)\( AP_{75} \)\( AP_S \)\( AP_M \)\( AP_L \)Flops
Deform. Enc. (4 Points)49.066.653.532.052.363.0245G
Deform. Enc. (6 Points)48.465.952.731.051.662.9250G
Deform. Enc. (8 Points)48.566.152.830.551.862.8255G
HAE (Ours)49.667.054.231.852.963.8248G
+ +Comparisons of HAE with Deformable Encoder variants. To make a fair comparison with the deformable encoder [50], we conduct experiments on increasing sampling points in the deformable encoder. The results are shown in Table 5. The performance of baseline with 4 sampling points is 49.0 AP. Performance decreases as sample points in the deformable encoder increases to 6 and 8, which are 48.4 AP and 48.5 AP, respectively. Even increasing the number of sample points in the deformable encoder does not improve performance for large and medium objects. This may be because deformable attention determines sampling weights based on query alone, leading to inaccurate weights as the number of points increases. However, our HAE can achieve performance gains (+0.6 AP) with less computational consumption (248G vs 250G). We think the possible reason is that the weight generation in standard self-attention, which calculates the similarity by query and key, is more suitable for long-distance sampling. These results support the motivation and design of our hybrid attention encoder. + +The number of Hybrid Attention Layers. Introducing too much self-attention may lead to slow convergence. We analyze the effects of the number of hybrid attention layers. The results are shown in Table 6. When only a hybrid attention layer is introduced, the model achieves 49.3 AP with 0.3 AP improvement. The improvement is mainly contributed by large and medium objects. Increasing the number of hybrid attention layers by 2 or 3 results in performance gains of 0.6 and 0.7 AP, respectively. It is still the large and medium objects that contribute the most performance gains. As the number of hybrid attention layers increases, the performance gains begin to decrease. In particular, there is no performance gain when increasing to 6 hybrid attention layers. This may be because the negative effect of slow convergence becomes dominant. + +Table 6: Ablations on the number of hybrid attention layers in HAE. + +
NumberAP\( AP_{50} \)\( AP_{75} \)\( AP_S \)\( AP_M \)\( AP_L \)
-49.066.653.532.052.363.0
149.366.653.731.152.763.7
249.667.054.231.852.963.8
349.767.254.332.352.864.0
449.367.053.831.952.763.5
549.366.854.031.452.863.1
648.866.353.430.752.263.1
+ +Table 7: Comparisons of ours EMCA with other decoder variants. * indicates that these models scale to the similar complexity as EMCA. + +
MethodAP\( AP_{50} \)\( AP_{75} \)\( AP_{2th} \)\( AP_{6th} \)\( AP_{10th} \)Flops
DINO [42]49.066.653.534.243.245.7245G
+Iterative Enc [40]47.765.252.133.442.044.6245G
+Iterative Enc*47.465.151.633.242.044.1284G
+Memory Fusion [25]49.367.053.735.743.746.2253G
+Memory Fusion*49.366.953.935.743.546.2290G
+EMCA (Ours)49.867.454.636.044.546.6283G
+ +Comparisons of EMCA with decoder variants. We compare our EMCA with other decoder variants. As shown in Table 7, using the Iterative Encoder [40] degrades the performance of the baseline. There are 0.3 AP performance gains when the Memory Fusion [25] is introduced, but there is no performance gain from the scaling of the approach. The performance gain of Memory Fusion is 0.5 AP less than our EMCA. These results demonstrate the effectiveness of our method design, which adaptively fuses multi-level encoder features based on query features for multi-scale objects, rather than using the same features for all objects. In addition, to show that our method converges faster, we also show results for the 2nd, 6th, and 10th epochs. It can be seen that our method outperforms the other methods throughout the training process, indicating the faster convergence of our method. + +Comparisons of EMCA with DINO variants. For a fair comparison, we compared with more complex baseline models due to the increased computational complexity introduced by EMCA. More complex baseline models are implemented by introducing more encoders and features. The results are shown in Table 8. By increasing the original baseline from 6 encoders to 7 encoders, the model reached 49.3 AP with 268G. When using 8 encoders, the model achieves 49.5 AP with 292G. While performance can be improved by stacking encoders, our EMCA utilizes encoder redundancy information and is a good complement to encoders. When using 10 encoders, the model suffers from overfitting and achieves only 47.8 AP with 339G. However, our EMCA with 10 encoders still has performance gains to reach 50.6 AP. This results show that our EMCA can + +Table 8: Comparisons of ours EMCA with DINO variants. + +
MethodAP\( AP_{50} \)\( AP_{75} \)\( AP_S \)\( AP_M \)\( AP_L \)Flops
6 encoder (Baseline)49.066.653.532.052.363.0245G
7 encoder49.366.753.732.052.463.7268G
8 encoder49.566.954.031.952.863.4292G
10 encoder47.864.852.130.650.961.5339G
5 scale49.466.953.832.352.563.9721G
6encoder+EMCA (Ours)49.867.454.633.052.864.7283G
8encoder+EMCA (Ours)50.267.854.632.453.364.6346G
10encoder+EMCA (Ours)50.668.155.333.553.965.3409G
+ +mitigate the risk of overfitting. A very large amount of computation (721G) is required when 5 scales of backbone features are used, and then the performance reaches only 49.4 AP. These results show that our EMCA is a good complement to encoders and can mitigate the risk of overfitting. + +Quantitative analysis of EMCA. To analyze the weights of EMCA, we count the fusion weights in the last decoder layer from the COCO val2017 set. First, we analyzed the weights of different encoder layers, and the results are shown in Figure 4. It can be seen that our method learns different weights for encoder layers. Specifically, the first encoder layer plays the most important role, the 2nd, 3rd, and 4th encoder layers have decreasing importance, and the 4th, 5th, and 6th encoder layers have similar roles. As shown in Table 9, using the 2nd, 4th, and 6th encoder layers as inputs, our approach drops 0.3 AP and saves 23G of computational overhead. When only the 1st and 6th encoder layers are used, the computation is reduced by 30G while the performance is dropped by 0.2 AP. We call the EMCA with only the 1st and 6th encoder layers as EMCA-Lite. We then analyze the weights of the objects at different scales to illustrate that our method improves multi-scale learning. As shown in Figure 5, objects of different sizes are given different weights, and the larger the scale, the higher the weights. See the Appendix for object scale intervals. These results illustrate the effectiveness of fusion weights based on queries. + +![](images/3e519da46a9aa1b3a905a5dd92e71b061434d391bd679246b694a1cb54e6c989.jpg) +Fig. 4: Statistical fusion weights for different encoders. Layer #6 means the last-level encoder. + +![](images/e5df4497c2d98d95ddb2dcfb7ee420f32edcf0ed38029689c5ecf32e78402bd6.jpg) +Fig.5: Statistical fusion weights for different scale objects. Size #1 to size #6 are ordered by scale from small to large. + +Table 9: Ablations on the layer of input encoders in EMCA. + +
EncoderAPAP50AP75Flops
-49.066.653.5245G
12345649.867.454.6283G
24649.567.253.9260G
2649.367.253.8253G
1649.667.454.3253G
+ +Table 10: Ablations on the fusion methods in EMCA. + +
Fusion MethodAPAPSAPMAPL
-49.032.052.363.0
sum49.331.952.563.9
max48.630.951.663.2
sigmoid weighted49.833.052.864.7
softmax weighted49.532.152.864.5
+ +Fusion Methods in EMCA. The fusion methods we explore include sum fusion, max fusion, and adaptive weighted fusion (sigmoid and softmax). All results are shown in Table 10. By using sum fusion, our method improves the baseline by 0.3 AP. This result shows that multi-level encoder features are richer in semantic information than single-level encoder features. When using max fusion, our method achieves 48.6 AP below the baseline. It can be seen that the performance decreases significantly for small and medium objects, but increases slightly for large objects. This may be because large objects tend to have higher activation so the features of small and medium objects are weakened during max fusion. When using adaptive weighted fusion, our method improves the baseline by 0.8 AP and 0.5 AP, respectively. It can be seen that the performance of small, medium, and large objects is all improved. In particular, the performance on large objects has improved significantly (+1.7 AP) by sigmoid weighted fusion. This may be because the larger the object, the more information is activated at all levels of the encoder, and therefore the more information can be used. In addition, the corresponding query features of large objects are richer so that the learned fusion weights are more accurate. These results suggest that sigmoid weighted fusion is a better fusion method. + +# 5 Conclusion + +In this paper, we analyze the limitations of DETR detectors for multi-scale learning and propose local feature enhancement and multi-level encoder exploitation techniques to address the limitations. Based on these, we construct a novel detection transformer detector named Augmented DETR (AugDETR), which consists of two components: Hybrid Attention Encoder and Encoder-Mixing Cross-Attention. HAE enlarges the receptive field and introduces global context to enhance the local feature of the deformable encoder. EMCA uses multi-level encoders to obtain more discriminative features and speed up convergence. Extensive experiments with various DETR-based detectors and backbones validate the effectiveness of our method. + +# Acknowledgements + +This work is supported by National Natural Science Foundation of China (NSFC) under Grant 62088102. This work is partly supported by China Telecom Corporation Ltd. Data&Al Technology Company, Beijing, China and China Telecom - Xi'an Jiaotong University Joint Innovation Institute for Science-Education Integration of Intelligent Cloud Network. We also appreciate Jingwen Fu, He Zhang and Tao Yang for their insightful discussions. We thank all the anonymous reviewers and Area Chairs for their constructive and helpful comments, which have significantly improved the quality of the paper. + +# References + +1. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154-6162 (2018) +2. Cai, Z., Liu, S., Wang, G., Ge, Z., Zhang, X., Huang, D.: Align-detr: Improving detr with simple iou-aware bce loss. arXiv preprint arXiv:2304.07527 (2023) +3. Cao, X., Yuan, P., Feng, B., Niu, K.: Cf-detr: Coarse-to-fine transformers for end-to-end object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 185-193 (2022) +4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +5. Chen, Q., Chen, X., Zeng, G., Wang, J.: Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085 (2022) +6. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2988-2997 (2021) +7. Ding, X., Zhang, X., Han, J., Ding, G.: Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11963-11975 (2022) +8. Dong, J., Huang, Y., Zhang, S., Chen, S., Zheng, N.: Construct effective geometry aware feature pyramid network for multi-scale object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 534-541 (2022) +9. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019) +10. Gao, P., Zheng, M., Wang, X., Dai, J., Li, H.: Fast convergence of detr with spatially modulated co-attention. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3621-3630 (2021) +11. Ghiasi, G., Lin, T.Y., Le, Q.V.: Nas-fpn: Learning scalable feature pyramid architecture for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7036-7045 (2019) +12. Guo, C., Fan, B., Zhang, Q., Xiang, S., Pan, C.: Augfpn: Improving multi-scale feature learning for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12595-12604 (2020) + +13. He, Y., Zhu, C., Wang, J., Savvides, M., Zhang, X.: Bounding box regression with uncertainty for accurate object detection. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 2888-2897 (2019) +14. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: Detrs with hybrid matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19702-19712 (2023) +15. Kim, S.W., Kook, H.K., Sun, J.Y., Kang, M.C., Ko, S.J.: Parallel feature pyramid network for object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 234-250 (2018) +16. Kong, T., Sun, F., Liu, H., Jiang, Y., Li, L., Shi, J.: Foveabox: Beyound anchor-based object detection. IEEE Transactions on Image Processing 29, 7389-7398 (2020) +17. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. In: Proceedings of the European conference on computer vision (ECCV). pp. 734-750 (2018) +18. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022) +19. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems 33, 21002-21012 (2020) +20. Lin, T.Y., Dólar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017) +21. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +23. Lin, Y., Yuan, Y., Zhang, Z., Li, C., Zheng, N., Hu, H.: Detr does not need multiscale or locality design. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6545-6554 (2023) +24. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. In: International Conference on Learning Representations (2021) +25. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., et al.: Detection transformer with stable matching. arXiv preprint arXiv:2304.04742 (2023) +26. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8759-8768 (2018) +27. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 21-37. Springer (2016) +28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) + +29. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021) +30. Nguyen, D.K., Ju, J., Booij, O., Oswald, M.R., Snoek, C.G.: Boxer: Box-attention for 2d and 3d transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4773-4782 (2022) +31. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra r-cnn: Towards balanced learning for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 821-830 (2019) +32. Pu, Y., Liang, W., Hao, Y., Yuan, Y., Yang, Y., Zhang, C., Hu, H., Huang, G.: Rank-detr for high quality object detection. Advances in Neural Information Processing Systems 36 (2024) +33. Qiao, S., Chen, L.C., Yuille, A.: Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10213-10224 (2021) +34. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 39(6), 1137-1149 (2016) +35. Roh, B., Shin, J., Shin, W., Kim, S.: Sparse detr: Efficient end-to-end object detection with learnable sparsity. In: International Conference on Learning Representations (2021) +36. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020) +37. Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9627-9636 (2019) +38. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022) +39. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022) +40. Zhang, G., Luo, Z., Tian, Z., Zhang, J., Zhang, X., Lu, S.: Towards efficient use of multi-scale features in transformer-based object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6206-6216 (2023) +41. Zhang, G., Luo, Z., Yu, Y., Cui, K., Lu, S.: Accelerating detr convergence via semantic-aligned matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 949-958 (2022) +42. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: The Eleventh International Conference on Learning Representations (2022) +43. Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8514-8523 (2021) +44. Zhang, H., Chang, H., Ma, B., Wang, N., Chen, X.: Dynamic r-cnn: Towards high quality object detection via dynamic training. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 260-275. Springer (2020) + +45. Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7329-7338 (2023) +46. Zheng, D., Dong, W., Hu, H., Chen, X., Wang, Y.: Less is more: Focus attention for efficient detr. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6674-6683 (2023) +47. Zhou, X., Wang, D., Krähenbuhl, P.: Objects as points. arXiv preprint arXiv:1904.07850 (2019) +48. Zhou, X., Zhuo, J., Krahenbuhl, P.: Bottom-up object detection by grouping extreme and center points. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 850-859 (2019) +49. Zhu, C., He, Y., Savvides, M.: Feature selective anchor-free module for single-shot object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 840-849 (2019) +50. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) +51. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6748-6758 (2023) \ No newline at end of file diff --git a/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/images.zip b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..26706133745b2fc7f7b2295f91a8310b661b3df2 --- /dev/null +++ b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8bc26eddb5d1e42453213eced44bee9db925a2141c9a153882dd7cd469d6dee +size 608343 diff --git a/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/layout.json b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4c2831d02b24847db76868b63f22a302c9f321e3 --- /dev/null +++ b/2024/AugDETR_ Improving Multi-scale Learning for Detection Transformer/layout.json @@ -0,0 +1,8067 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "text", + "content": "AugDETR: Improving Multi-scale Learning for Detection Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 168, + 457, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 168, + 457, + 181 + ], + "spans": [ + { + "bbox": [ + 154, + 168, + 457, + 181 + ], + "type": "text", + "content": "Jinpeng Dong, Yutong Lin, Chen Li, Sanping Zhou, Nanning Zheng*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 190, + 478, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 190, + 478, + 246 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 478, + 246 + ], + "type": "text", + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University {djp1235a, yutonglin, edward82}@stu.xjtu.edu.cn {spzhou, nnzheng} " + }, + { + "bbox": [ + 135, + 190, + 478, + 246 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 135, + 190, + 478, + 246 + ], + "type": "text", + "content": " xjtu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 274, + 452, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 274, + 452, + 472 + ], + "spans": [ + { + "bbox": [ + 160, + 274, + 452, + 472 + ], + "type": "text", + "content": "Abstract. Current end-to-end detectors typically exploit transformers to detect objects and show promising performance. Among them, Deformable DETR is a representative paradigm that effectively exploits multi-scale features. However, small local receptive fields and limited query-encoder interactions weaken multi-scale learning. In this paper, we analyze local feature enhancement and multi-level encoder exploitation for improved multi-scale learning and construct a novel detection transformer detector named Augmented DETR (AugDETR) to realize them. Specifically, AugDETR consists of two components: Hybrid Attention Encoder and Encoder-Mixing Cross-Attention. Hybrid Attention Encoder enlarges the receptive field of the deformable encoder and introduces global context features to enhance feature representation. Encoder-Mixing Cross-Attention adaptively leverages multi-level encoders based on query features for more discriminative object features and faster convergence. By combining AugDETR with DETR-based detectors such as DINO, AlignDETR, DDQ, our models achieve performance improvements of 1.2, 1.1, and 1.0 AP in the COCO under the ResNet-50-4scale and 12 epochs setting, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 483, + 453, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 483, + 453, + 503 + ], + "spans": [ + { + "bbox": [ + 160, + 483, + 453, + 503 + ], + "type": "text", + "content": "Keywords: Object detection " + }, + { + "bbox": [ + 160, + 483, + 453, + 503 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 483, + 453, + 503 + ], + "type": "text", + "content": " Detection transformer " + }, + { + "bbox": [ + 160, + 483, + 453, + 503 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 483, + 453, + 503 + ], + "type": "text", + "content": " Hybrid attention " + }, + { + "bbox": [ + 160, + 483, + 453, + 503 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 483, + 453, + 503 + ], + "type": "text", + "content": " Multi-level encoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 525, + 230, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 230, + 538 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 230, + 538 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": "Many CNN-based detectors [1,12,20,27,34,37,47] have been proposed to improve detection performance in the past decade. Despite their promising progress, these detectors suffer from many hand-crafted components, such as non-maximum suppression and anchor design. In contrast, CNN-based paradigm is broken by DETR [4], a novel transformer detector. It gets rid of hand-crafted components and implements an end-to-end paradigm. Despite its novel paradigm, DETR suffers from some limitations, including slow convergence and inferior performance. To address these limitations, many methods [2, 6, 14, 24, 25, 29, 42, 50]" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "have been proposed from different perspectives, such as introducing spatial priors, hybrid matching training, designing efficient attention modules, assigning sample weights, etc. With these optimizations, DINO [42] achieved a new record on the COCO benchmark [22]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 483, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 483, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 483, + 321 + ], + "type": "text", + "content": "Although DETR-based detectors have achieved promising performance, multiscale object detection has not received sufficient attention so far, which may limit the performance of detectors. Scale variation of objects remains one of the crucial challenges in object detection. Objects at different scales require inconsistent semantic features, e.g., large objects require features with larger receptive fields. To the best of our knowledge, Deformable-DETR [50] has made some effort to remedy this issue. Deformable-DETR uses multi-scale backbone features to improve multi-scale learning by deformable attention. In addition to the use of multi-scale backbone features, we believe that the multi-scale learning of Deformable-DETR has not been fully exploited. As shown in Table 1, although Deformable DETR is 2.5 AP higher than the DETR in overall performance, the performance on large objects is still 1.5 AP lower than the DETR. We present two techniques that can further improve multi-scale learning as follows:" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 139, + 387, + 320, + 436 + ], + "blocks": [ + { + "bbox": [ + 132, + 356, + 326, + 378 + ], + "lines": [ + { + "bbox": [ + 132, + 356, + 326, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 356, + 326, + 378 + ], + "type": "text", + "content": "Table 1: Performance comparison between DETR and Deformable DETR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 139, + 387, + 320, + 436 + ], + "lines": [ + { + "bbox": [ + 139, + 387, + 320, + 436 + ], + "spans": [ + { + "bbox": [ + 139, + 387, + 320, + 436 + ], + "type": "table", + "html": "
MethodEpochsAPAPSAPMAPL
DETR [4]50042.020.545.861.1
Deformable DETR [50]5044.527.147.659.6
", + "image_path": "b113959c313a5d2b507d5191bd488dd2c01065f154ada899f3063d6a55bae984.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 347, + 375, + 405, + 447 + ], + "blocks": [ + { + "bbox": [ + 346, + 344, + 482, + 366 + ], + "lines": [ + { + "bbox": [ + 346, + 344, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 346, + 344, + 482, + 366 + ], + "type": "text", + "content": "Fig. 1: Comparison of receptive fields of different encoders." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 375, + 405, + 447 + ], + "lines": [ + { + "bbox": [ + 347, + 375, + 405, + 447 + ], + "spans": [ + { + "bbox": [ + 347, + 375, + 405, + 447 + ], + "type": "image", + "image_path": "fb497dd95c6ea9c3b931e0bed1014de3ee7a801f9a8ee6cc6daa3c33284e6d5d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 422, + 375, + 481, + 447 + ], + "blocks": [ + { + "bbox": [ + 422, + 375, + 481, + 447 + ], + "lines": [ + { + "bbox": [ + 422, + 375, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 422, + 375, + 481, + 447 + ], + "type": "image", + "image_path": "525e51d10ad7af2bec6a9d5e98da01acc439e4779e73eabc0ce5f33971791407.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 472, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 628 + ], + "type": "text", + "content": "Local feature enhancement. Deformable-DETR [50] replaces the original self-attention (global dense attention) with deformable attention (sparse attention) to reduce the complexity and slow convergence of dense attention. Deformable attention is implemented by a fixed number of deformable sampling points. Limited by local sampling points, deformable attention is difficult to achieve a larger receptive field compared to self-attention. This is not conducive to the detection of large objects. Further stacking of deformable attention still improves performance, as analyzed in the ablations. In addition, deformable attention loses the ability to perceive global context, while multi-scale object detection in some complex scenes may rely on global context information. Increasing the sampling points degrades the performance of the deformable attention, as analyzed in the ablations. Therefore, we introduce global contextual information into the deformable encoder to enhance multi-scale learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Multi-level encoder exploitation. DETR-based detectors typically stack six encoder layers and six decoder layers. Encoder layers enhance the backbone features, and decoder layers extract features from last encoder layer to decode" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": "queries. Previous methods ignore the use of multi-level encoder features. More discriminative object features can be obtained by fully exploiting multi-level encoder features with different receptive fields and feature preferences. We refer to RepLKNet [7] for analysis of receptive field. As shown in Figure 1, the comparison of encoder 1 and 6 in DINO [42] shows that different encoders have varying feature preferences and receptive fields. In addition, the fact that multilevel encoders are all involved in object feature generation can speed up convergence because the supervisory signals are available to multiple levels of encoders, rather than being available only to the last level encoder. Utilizing multi-level encoder features also can introduce redundancy to mitigate the risk of overfitting. Although multi-level encoder features is beneficial for object detection, using multi-level encoder features fixed for all objects is not conducive to multiscale learning. Since objects at different scales have different feature preferences, adaptive use of multi-level encoder features by learning weights based on query features will facilitate multi-scale learning for DETR-based detectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "content": "In this paper, we propose AugDETR, a simple yet effective DETR-based detector that uses two different components to improve multi-scale learning. First, Hybrid Attention Encoder is proposed to augment local features in the deformable encoder by partially combining dense attention with deformable attention. Hybrid Attention Encoder can enlarge the receptive field and introduce global context, which can improve detection performance on large objects. Second, Encoder-Mixing Cross-Attention is introduced to better exploit features from different encoder layers by weighted aggregation. Objects at different scales can learn different aggregation weights based on object features to facilitate multi-scale learning. With this design, Encoder-Mixing Cross-Attention can generate more discriminative object features for subsequent classification and localization tasks. We validate the effectiveness of AugDETR on the COCO benchmark. Our method achieves 50.2 AP (+1.2 AP), 51.3 AP (+1.1 AP), and 51.8 AP (+1.0 AP) based on DINO [42], AlignDETR [2], and DDQ [45] with ResNet-50-4scale under 12 epochs settings, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 475, + 341, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 475, + 341, + 486 + ], + "spans": [ + { + "bbox": [ + 147, + 475, + 341, + 486 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 491, + 480, + 585 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 491, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 514 + ], + "type": "text", + "content": "- We analyze local feature enhancement and multi-level encoder exploitation techniques for improved multi-scale learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 515, + 480, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 515, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 138, + 515, + 480, + 548 + ], + "type": "text", + "content": "- A new detection transformer detector named AugDETR is proposed to achieve the techniques with Hybrid Attention Encoder and Encoder-Mixing Cross-Attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 550, + 480, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 550, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 138, + 550, + 480, + 585 + ], + "type": "text", + "content": "- We validate AugDETR equipped with various DETR-based detectors and backbones on the COCO benchmark, and it consistently yields significant performance improvements over DETR-based detectors." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 601, + 237, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 237, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 237, + 613 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 248, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 248, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 248, + 635 + ], + "type": "text", + "content": "2.1 Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "Early deep-learning detectors can be roughly classified into anchor-based and anchor-free methods. Anchor-based methods use predefined anchors to locate" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "text", + "content": "AugDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": "objects, including Faster R-CNN [34], SSD [27], RetinaNet [21], and their variants [1, 13, 31, 44]. Anchor-free methods use points near object centers to locate objects, including CenterNet [47], FCOS [37], and others [9, 16, 17, 48, 49]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 416 + ], + "type": "text", + "content": "Recently, DETR [4] proposes a simple transformer pipeline for object detection, effectively alleviating the need for heuristic components. Compared to CNN-based work, DETR is rather novel but still suffers from some limitations, including slow convergence and inferior performance. Many follow-up works focus on alleviating the limitations of DETR from different perspectives. Conditional DETR [29] proposes to decouple the context embedding and position embedding, which helps DETR speed up convergence. Anchor DETR [38] uses the anchor points as the initial queries for fast convergence. DAB DETR [24] uses 4D anchor boxes that are dynamically updated with the decoder layer to represent queries. DN DETR [18] proposes a denoising training mechanism to speed up convergence. DINO [42] further extends DN DETR with contrastive denoising training, mixed query, and look-forward twice techniques. Group DETR [5], HDETR [14], and CO-DETR [51] propose various one-to-one matching and one-to-many matching combination methods to stabilize the matching process of DETR. Align DETR [2], Stable DINO [25], and Rank DETR [32] design different schemes to weight the samples by the IoU-aware classification loss. Plain DETR [23] proposes a box-to-pixel relative position bias and a masked image modeling [39] pre-training scheme to improve the DETR detector. Unlike these methods that design spatial priors, hybrid matching, sample weight assignments, etc., our approach focuses on improving the multi-scale learning of DETR detectors. In addition, we conducted experiments using some SOTA methods as the baseline models. Our techniques are complementary to theirs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 431, + 374, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 374, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 374, + 443 + ], + "type": "text", + "content": "2.2 Multi-Scale Learning for Object Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "In the early era, image pyramid is a common solution to improve multi-scale learning. Instead of image pyramid, some methods utilize multi-scale features to mitigate scale variation. A representative work among these methods is the feature pyramid network (FPN [20]), which proposes a top-down pathway and lateral connections to construct the feature pyramid. Inspired by FPN, multi-scale feature extraction has been widely studied, such as BiFPN [36], PAFPN [26], NAS-FPN [11], AugFPN [12], and other variants [8,15,33]. Recently, some works have attempted to design multi-scale feature extraction modules in DETR. Deformable DETR [50] proposes deformable attention to utilize multi-scale backbone features. Dynamic DETR [6] proposes a CNN-based dynamic encoder with scale-aware attention, spatial-aware attention, and task-aware attention. SMCA [10] generates the Gaussian-like weight map as the spatial prior and the scale weights as the scale prior to modulated multi-scale features. IMFA [40] designs a new scheme to adaptively sample sparse multi-scale features. Different from them, our approach enhances multi-scale learning by introducing global attention to expand the receptive field and integrate global context. To use global attention effectively, we propose hybrid scale and hybrid layer strategies to construct the hybrid attention encoder. In addition, we adaptively use multi-level" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "encoder features based on query features to improve multi-scale learning, which is also distinctly different from previous methods." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 147, + 162, + 301, + 299 + ], + "blocks": [ + { + "bbox": [ + 147, + 162, + 301, + 299 + ], + "lines": [ + { + "bbox": [ + 147, + 162, + 301, + 299 + ], + "spans": [ + { + "bbox": [ + 147, + 162, + 301, + 299 + ], + "type": "image", + "image_path": "6895ecd5450bb6c6fd1c867f6df6a2f3535d04fac7a5c463ac50becb05f28ad1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 309, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 399 + ], + "type": "text", + "content": "Fig. 2: Left: The pipeline of conventional DETR-based detectors. The backbone network and the stacked encoder layers are used to extract and enhance the image features. Then, the object queries utilize the decoder layers to interact with the enhanced image features. Finally, the interacted object queries are fed into the detection head to obtain the predictions. Only the features from the last encoder layer can be used to interact with the object queries. Right: The pipeline of Our AugDETR. Our AugDETR restructures the conventional encoder-decoder pipeline. The object queries can interact with features from all encoder layers. For clarity, only the 3-layer structure is shown." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 314, + 160, + 470, + 299 + ], + "blocks": [ + { + "bbox": [ + 314, + 160, + 470, + 299 + ], + "lines": [ + { + "bbox": [ + 314, + 160, + 470, + 299 + ], + "spans": [ + { + "bbox": [ + 314, + 160, + 470, + 299 + ], + "type": "image", + "image_path": "af94f65b0cfca1973e3f69f84960b73ca9b669627c87dbd622f987dedaf472b0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 438, + 202, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 202, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 202, + 451 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 510 + ], + "type": "text", + "content": "In section 3.1, we first briefly revisit the detection transformer and then introduce the pipeline of AugDETR. In section 3.2, we describe the details of the proposed Hybrid Attention Encoder. In section 3.3, we introduce the Encoder-Mixing Cross-Attention used in the transformer decoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 526, + 338, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 338, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 338, + 538 + ], + "type": "text", + "content": "3.1 A Revisit of DETR-based Detector" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 581 + ], + "type": "text", + "content": "Since our proposed method is constructed on top of the recent DETR-based object detectors, we first briefly review the pipeline of the DETR-based detectors, taking the Deformable DETR [50] as an example." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "The pipeline of Deformable DETR is shown in Figure 2 (left). Deformable DETR consists of 3 subcomponents: a backbone, a encoder, and a decoder. Given an input image " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", a backbone network generates multi-scale features. We denote the multi-scale features as " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{C_2, C_3, C_4, C_5\\}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". The spatial resolutions of these features are usually " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/4^2" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/8^2" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/16^2" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/32^2" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " of the given image. Then these features reduced to the same channel by a linear projection are denoted as " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{P_2, P_3, P_4, P_5\\}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". Next, a transformer encoder is used to further enhance these" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": "multi-scale features. The transformer encoder is typically stacked with 6 encoder layers. The encoder layer mainly consists of a self-attention and a feed-forward network. For Deformable DETR, multi-scale deformable attention is employed as self-attention to enhance multi-scale features. The encoded features of the different layers in the encoder are denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\{E_1,E_2,E_3,E_4,E_5,E_6\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": ". Finally, a transformer decoder is used to produce object predictions from a set of object queries. Typically, only the last level of encoder features are fed into the decoder to interact with the query. The transformer decoder is also typically stacked with 6 decoder layers for better detection performance. The decoder layer mainly consists of a self-attention, a cross-attention, and a feed-forward network. For deformable DETR, multi-scale deformable attention is used as cross-attention to achieve query-feature interaction. As shown in Figure 2 (right), our pipeline changes the fact that the decoder can utilize only the last level of encoder features, but rather all encoder features can be utilized by the decoder. This design lays the foundation for cross-attention to utilize multi-level encoder features." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 317, + 303, + 389 + ], + "blocks": [ + { + "bbox": [ + 138, + 317, + 303, + 389 + ], + "lines": [ + { + "bbox": [ + 138, + 317, + 303, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 303, + 389 + ], + "type": "image", + "image_path": "da531c7b14acbdc100eff6e9ddeff03fc1cce099a8797bc45fd719259127b523.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 399, + 482, + 467 + ], + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 467 + ], + "type": "text", + "content": "Fig. 3: (a) is the process of hybrid attention in the Hybrid Attention Layer. Dense attention is first applied to only the top features, and then sparse attention is applied to the multi-scale features. (b) is the details of Encoder-Mixing Cross-Attention. Deformable attention extracts multiple object features from multi-level encoder and then the extracted features are aggregated based on the adaptive fusion weights learned from the query features." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 316, + 476, + 389 + ], + "blocks": [ + { + "bbox": [ + 309, + 316, + 476, + 389 + ], + "lines": [ + { + "bbox": [ + 309, + 316, + 476, + 389 + ], + "spans": [ + { + "bbox": [ + 309, + 316, + 476, + 389 + ], + "type": "image", + "image_path": "9296197508af022d0cdbf0a2fea8b4c9ea6db2b72940e377c5b285df478185cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 506, + 294, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 294, + 517 + ], + "type": "text", + "content": "3.2 Hybrid Attention Encoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 525, + 480, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 538 + ], + "type": "text", + "content": "The encoder layer in the original DETR [4] uses the self-attention as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 210, + 558, + 367, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 558, + 367, + 571 + ], + "spans": [ + { + "bbox": [ + 210, + 558, + 367, + 571 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = W _ {q} P _ {5}, \\mathbf {K} = W _ {k} P _ {5}, \\mathbf {V} = W _ {v} P _ {5},", + "image_path": "2899cec8f609aa08a447e9234a2ebb5bd899ede08e7cbba56b871114dcdd0a3b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 210, + 572, + 480, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 572, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 210, + 572, + 480, + 586 + ], + "type": "interline_equation", + "content": "\\operatorname {S e l f A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {Q K}) \\mathbf {V}, \\tag {1}", + "image_path": "bd0aad23dd7f3b428f71a5eac568deeeef2eed9cb5baf157bccbb2806858ae63.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "inline_equation", + "content": "P_{5}" + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": " are the last level features of the backbone network; " + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": " are query, key, and value features, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The standard self-attention performs dense computation that involves all features. Applying standard self-attention to a large feature map will cause high complexity due to its dense computation, so the multi-scale features of the backbone network are not utilized in the original DETR. In addition, dense compu" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "tation will lead to slow convergence. The original DETR required 500 epochs of training to achieve good performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "content": "To exploit the multi-scale features of the backbone, Deformable DETR [50] proposes deformable attention. The multi-scale deformable self-attention can be simply formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 208, + 187, + 362, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 187, + 362, + 198 + ], + "spans": [ + { + "bbox": [ + 208, + 187, + 362, + 198 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\operatorname {C o n c a t} \\left(P _ {3}, P _ {4}, P _ {5}\\right), \\mathbf {A} = W _ {a} \\mathbf {Q},", + "image_path": "6549ae88748a012809b0351655eed95022f910632045190eb84c13a6d742c821.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 201, + 375, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 201, + 375, + 213 + ], + "spans": [ + { + "bbox": [ + 208, + 201, + 375, + 213 + ], + "type": "interline_equation", + "content": "\\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {V} = \\operatorname {S a m p} \\left(W _ {v} \\mathbf {Q}, r + \\Delta r\\right),", + "image_path": "4977acad173d65eeeacafeb3c321e98006cf311fac3333e5f2a8d3e75987c09e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 208, + 217, + 480, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 217, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 208, + 217, + 480, + 228 + ], + "type": "interline_equation", + "content": "\\operatorname {D e f o r m a b l e A t t e n t i o n} (\\mathbf {Q}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V}, \\tag {2}", + "image_path": "987f996d5fad5cf08f18297823772ce0bc519974edd6086ff66a5eb1c867198b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "inline_equation", + "content": "P_{3}, P_{4}, P_{5}" + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "text", + "content": " are multi-scale features of the backbone, " + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "text", + "content": " are a set of queries, " + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\Delta r" + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "text", + "content": " are the offsets of reference points, " + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "text", + "content": " are the reference points, " + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "text", + "content": " are the attention weights of sample features, Samp is a function that extracts the features corresponding to the location " + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "inline_equation", + "content": "(r + \\Delta r)" + }, + { + "bbox": [ + 130, + 238, + 479, + 285 + ], + "type": "text", + "content": " by bilinear interpolation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 286, + 480, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 480, + 393 + ], + "type": "text", + "content": "The deformable attention performs sparse computation that involves only a predefined number of surrounding features. Because deformable attention does not produce high complexity when combined with multi-scale features, it can utilize the multi-scale features of the backbone to detect multi-scale objects. In addition, sparse computation will lead to fast convergence. Deformable DETR converges 10 times faster than DETR. However, limited by sparse computation, deformable attention typically does not have a large receptive field and does not use global context information. This is not conducive to detecting large objects and objects in complex scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 394, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 480, + 525 + ], + "type": "text", + "content": "To address these issues, we propose the Hybrid Attention Encoder that combines standard self-attention and deformable attention in the encoder layer. The hybrid attention layer is shown in Figure 3 (a). Introducing dense attention into the deformable attention encoder can enlarge the receptive field and perceive the global context. This will be more beneficial for multi-scale object detection. However, how to introduce dense attention into deformable attention without increasing too much complexity and slowing convergence is a key challenge. In order not to introduce too much complexity, we propose the hybrid-scale strategy. This strategy means that only the last scale features receive dense attention, not all scale features. With this strategy, our method does not introduce too much computation. Our hybrid attention can be simply formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 208, + 536, + 296, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 536, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 208, + 536, + 296, + 546 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {\\prime} = \\mathbf {K} ^ {\\prime} = \\mathbf {V} ^ {\\prime} = P _ {5},", + "image_path": "2661f7d95ec66eedafd9700690cc3e80e474c5136b8b7aafd7e289f1049333eb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 208, + 550, + 344, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 550, + 344, + 562 + ], + "spans": [ + { + "bbox": [ + 208, + 550, + 344, + 562 + ], + "type": "interline_equation", + "content": "P _ {5} ^ {\\prime} = \\operatorname {S e l f A t t e n t i o n} \\left(\\mathbf {Q} ^ {\\prime}, \\mathbf {K} ^ {\\prime}, \\mathbf {V} ^ {\\prime}\\right),", + "image_path": "b7e7b5163f7e0e2bac6766889fa86b9c86dd66aca980d9ff374b88b27744e088.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 208, + 566, + 362, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 566, + 362, + 578 + ], + "spans": [ + { + "bbox": [ + 208, + 566, + 362, + 578 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\operatorname {C o n c a t} \\left(P _ {3}, P _ {4}, P _ {5} ^ {\\prime}\\right), \\mathbf {A} = W _ {a} \\mathbf {Q},", + "image_path": "064af64f9cf09e4130fd23e585050bda44dbf46a8f873d962ce0f282940f29e8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 208, + 581, + 375, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 581, + 375, + 593 + ], + "spans": [ + { + "bbox": [ + 208, + 581, + 375, + 593 + ], + "type": "interline_equation", + "content": "\\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {V} = \\operatorname {S a m p} \\left(W _ {v} \\mathbf {Q}, r + \\Delta r\\right),", + "image_path": "66604758bf922e2bab167f5e9addb6cf9297f4c9743fa6f89f9dd5b382da1290.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 207, + 595, + 479, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 595, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 207, + 595, + 479, + 608 + ], + "type": "interline_equation", + "content": "\\text {D e f o r m a b l e A t t e n t i o n} (\\mathbf {Q}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V}, \\tag {3}", + "image_path": "348f6ac5127e5613fe6ca5ada69376d64608f5ab171a5a32c5cca34aae9d3b1c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "In order not to slow down the convergence, we propose the hybrid-layer strategy. This strategy means that instead of adopting hybrid attention at all encoder layers, it is only adopted at some layers. From the ablation experiments, it can be seen that adopting hybrid attention in all the encoder layers will have no" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "text", + "content": "AugDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "performance gain due to slow convergence. In our practical use, we introduce hybrid attention only in the first two encoder layers in order to balance complexity and performance. With the above design, we build the Hybrid Attention Encoder that can enlarge the receptive field and utilize the global context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 180, + 326, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 180, + 326, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 326, + 192 + ], + "type": "text", + "content": "3.3 Encoder-Mixing Cross-Attention" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 236 + ], + "type": "text", + "content": "The cross-attention in the decoder layer of Deformable DETR [50] uses multiscale deformable attention to achieve query-feature interaction. The multi-scale deformable cross-attention can be simply formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 243, + 418, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 243, + 418, + 255 + ], + "spans": [ + { + "bbox": [ + 197, + 243, + 418, + 255 + ], + "type": "interline_equation", + "content": "\\mathbf {A} = W _ {a} \\mathbf {Q}, \\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {V} = \\operatorname {S a m p} (W _ {v} E _ {6}, r + \\Delta r),", + "image_path": "f5244b4caeccb935a766ee180b2c52bf2a6b09f11b4241aac12d8785c1053c69.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 197, + 258, + 480, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 258, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 197, + 258, + 480, + 270 + ], + "type": "interline_equation", + "content": "\\operatorname {D e f o r m a b l e A t t e n t i o n} (\\mathbf {Q}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V}, \\tag {4}", + "image_path": "154e5b64c9452ffc0d0b056223ecd640f14862c3f7886cefbcefacaabbc5be4d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": " are a set of decoder queries, " + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": " is the attention weights, " + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\Delta r" + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": " is the offsets of reference points, " + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": " is the reference points, " + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "inline_equation", + "content": "E_6" + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": " is the multi-scale features of last encoder layer, Samp is the sample function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 313, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 396 + ], + "type": "text", + "content": "DETR-based detectors typically stack six encoder layers to enhance the backbone features. However, only features of the last encoder layer are utilized in cross-attention. Although features of the last encoder layer have larger receptive fields and stronger semantics, it is not optimal for multi-scale objects due to the fact that objects have different favors for features. Features from different encoder layers have different receptive fields, and if they are fully utilized they will be more beneficial for multi-scale object detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 396, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 517 + ], + "type": "text", + "content": "To address these issues, cross-attention interacting with multi-level encoder can obtain more discriminative features. However, using the same multi-level encoder for all objects is not conducive to multi-scale learning since multi-scale objects have different feature preferences. We propose Encoder-Mixing Cross-Attention (EMCA) to adaptively exploit features of different encoders based on query features. The EMCA is shown in Figure 3 (b). Specifically, we first extract features from each encoder layer by using multi-scale deformable attention. Then, we generate fusion weights for extracted features from each encoder based on the query features. Finally, we adaptively aggregate features from each encoder according to the fusion weights. The EMCA can be simply formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 220, + 525, + 385, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 525, + 385, + 536 + ], + "spans": [ + { + "bbox": [ + 220, + 525, + 385, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {A} = W _ {a} \\mathbf {Q}, \\Delta r = W _ {p} \\mathbf {Q}, \\mathbf {w} _ {e} = \\sigma (W _ {e} \\mathbf {Q})", + "image_path": "cd86ab82369a31b7fdfa79c0b829cec63cd242ee0d069290c2d248d5faeb68ce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 220, + 540, + 342, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 540, + 342, + 552 + ], + "spans": [ + { + "bbox": [ + 220, + 540, + 342, + 552 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {l} = \\operatorname {S a m p} \\left(W _ {v l} E _ {l}, r + \\Delta r\\right),", + "image_path": "acebea11d71ec4cd77e135d9536fef62981e4216e2eff600cc70bd8865276973.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 220, + 555, + 480, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 220, + 555, + 480, + 587 + ], + "type": "interline_equation", + "content": "\\operatorname {E M C A} (\\mathbf {Q}, \\mathbf {E}) = \\sum_ {l = 1} ^ {L} w _ {e} ^ {l} \\cdot \\operatorname {S o f t m a x} (\\mathbf {A}) \\mathbf {V} _ {l}, \\tag {5}", + "image_path": "19fa17bc6e13301665522dbb388d40f3c8492dbbe815ca267c4b4fa2af3d1501.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": " means the Sigmoid, " + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "inline_equation", + "content": "E_{l}" + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": " is the features of the " + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": " level encoder, " + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": " is a set of features from each level encoders, " + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "inline_equation", + "content": "w_{e}^{l}" + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": " is the weights of the " + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": " level encoder, " + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{e}" + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": " is a set of weights for all encoders. " + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": " is the number of encoder layers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "We use the EMCA to replace all the cross-attention in the decoder layer. With this design, each object can adaptively utilize multi-level encoder features to obtain more discriminative features for improved multi-scale learning." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 147, + 481, + 417 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 2: Comparison with the state-of-the-art DETR variants on COCO val2017 with the ResNet-50-4scale backbone. * indicates the result is from our implemented model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 147, + 481, + 417 + ], + "lines": [ + { + "bbox": [ + 132, + 147, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 147, + 481, + 417 + ], + "type": "table", + "html": "
ModelBackbone#epochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Conditional-DETR [29]R5010843.064.045.722.746.761.5
SAM-DETR [41]R505039.861.841.620.543.459.6
Anchor-DETR [38]R505042.163.144.922.346.260.0
Dynamic-DETR [6]R501242.961.046.324.644.954.4
SMCA-DETR [10]R5010845.665.549.125.949.362.6
CF-DETR [3]R503647.866.552.431.250.662.8
Sparse-DETR [35]R505046.366.050.129.049.560.8
BoxeR-2D [30]R505050.067.954.730.952.862.6
Deformable-DETR [50]R505046.965.651.029.650.161.6
DAB-Deformable-DETR [24]R505046.866.050.429.149.862.3
DN-Deformable-DETR [18]R505048.667.452.731.052.063.7
H-DETR [14]R501248.766.452.931.251.563.5
Co-DETR [51]R501249.567.654.332.452.763.7
DINO [42]R501249.066.653.532.052.363.0
DINO [42]R502450.468.354.833.353.764.8
DDQ* [45]R501250.867.956.134.654.365.5
Focus DETR [46]R503650.468.555.034.053.564.4
AlignDETR [2]R501250.267.854.432.953.365.0
Stable-DINO [25]R501250.467.455.032.954.065.5
Rank-DETR [32]R501250.267.755.034.153.664.0
DINO+OursR501250.2 (+1.2)67.855.032.353.264.7
DINO+OursR502451.3 (+0.9)69.056.233.554.766.1
AlignDETR+OursR501251.3 (+1.1)68.855.833.354.866.6
DDQ+OursR501251.8 (+1.0)69.256.934.855.066.7
", + "image_path": "5451531f21b457b6f7b8c5841ba6d0ec69911b2597411b60c35c91343068d4f5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 437, + 230, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 230, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 230, + 452 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 462, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 481, + 521 + ], + "type": "text", + "content": "Dataset and Metrics. We conduct all experiments on the challenging COCO 2017 detection dataset. It is split into training set, validation set, and test-dev set with 115K, 5K, and 20K images respectively. We train models on the training set and evaluate performance on the validation set and test-dev set. All evaluated results follow the COCO-style Average Precision (AP) metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 522, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 605 + ], + "type": "text", + "content": "Implementation Details. We use DINO [42], Align DETR [2], and DDQ [45] as our baseline methods. All ablation studies are conducted on DINO-4scale with ResNet-50 for 12 epochs. We employ AdamW with " + }, + { + "bbox": [ + 132, + 522, + 481, + 605 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 132, + 522, + 481, + 605 + ], + "type": "text", + "content": " weight decay as the optimizer to train our models. The batch size is set as 16 for all experiments. The initial learning rate is set as " + }, + { + "bbox": [ + 132, + 522, + 481, + 605 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 132, + 522, + 481, + 605 + ], + "type": "text", + "content": " and it decreases by multiplying 0.1 after the 11th epoch for the 12 epochs setting. We use a random seed 0 in all our experiments. All other hyper-parameters follow the default from the codebase." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 622, + 230, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 230, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 230, + 634 + ], + "type": "text", + "content": "4.1 Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we evaluate AugDETR on the COCO val2017 set and compare it with the state-of-the-art DETR variants. All results with ResNet50 are sum-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 136, + 475, + 200 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 470, + 125 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 470, + 125 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 470, + 125 + ], + "type": "text", + "content": "Table 3: Comparison with the DETR variants on val2017 set with the Swin-T." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 136, + 475, + 200 + ], + "lines": [ + { + "bbox": [ + 135, + 136, + 475, + 200 + ], + "spans": [ + { + "bbox": [ + 135, + 136, + 475, + 200 + ], + "type": "table", + "html": "
MethodBackbone#epochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
DINO [42]Swin-T1251.369.056.034.554.466.0
OursSwin-T1252.3 (+1.0)70.257.036.855.367.2
AlignDETR [2]Swin-T1252.570.157.235.955.968.4
OursSwin-T1253.5 (+1.0)71.358.136.956.669.1
", + "image_path": "c872ad3436994c6dcd5b8eaf014eacd763d4a35519c5bb43fe7851e799202d25.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 423 + ], + "type": "text", + "content": "marized in Table 2. Combining our components with DINO [42], AugDETR achieves 50.2 AP, which is 1.2 AP higher than the DINO baseline with the same setting. Recent work on replacing the focal loss in DINO with the IoU-aware focal loss [19, 21, 43] has shown good performance. We chose a typical work AlignDETR [2] from these studies as baseline to validate the generality of our approach. Combining our components with AlignDETR, AugDETR achieves 51.3 AP, which is 1.1 AP higher than the AlignDETR baseline. Besides, we conduct experiments in the latest SOTA model DDQ [45]. Combining the our components with DDQ, AugDETR achieves 51.8 AP, which is 1.0 AP higher than the DDQ baseline. To demonstrate the effectiveness of our method under longer training, we conducted experiments under the " + }, + { + "bbox": [ + 130, + 220, + 482, + 423 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 130, + 220, + 482, + 423 + ], + "type": "text", + "content": " schedule. Our method achieves 51.3 AP, which is 0.9 AP higher than the baseline DINO. Besides, AugDETR can consistently achieve performance improvements even with more advanced backbone networks. As shown in Table 3, when using Swin-T [28] as the backbone, our method achieves performance improvements of 1.0 AP and 1.0 AP over DINO and AlignDETR, respectively. These results demonstrate the effectiveness and generalizability of our method." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 187, + 475, + 424, + 540 + ], + "blocks": [ + { + "bbox": [ + 130, + 442, + 482, + 466 + ], + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 466 + ], + "type": "text", + "content": "Table 4: Ablations on each component of AugDETR. \"HAE\" and \"EMCA\" represent Hybrid Attention Encoder, and Encoder-Mixing Cross-Attention, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 475, + 424, + 540 + ], + "lines": [ + { + "bbox": [ + 187, + 475, + 424, + 540 + ], + "spans": [ + { + "bbox": [ + 187, + 475, + 424, + 540 + ], + "type": "table", + "html": "
HAEEMCAAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
49.066.653.532.052.363.0
✓49.667.054.231.852.963.8
✓49.867.454.633.052.864.7
✓✓50.267.855.032.353.264.7
", + "image_path": "6f6802950ce75c2c5160aae6cce808a55974bc5172e637f5e9c7e99b14206cf5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 576, + 209, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 209, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 209, + 586 + ], + "type": "text", + "content": "4.2 Ablations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Ablation studies on the effectiveness of each component. To analyze the contributions of individual components in AugDETR, Hybrid Attention Encoder (HAE), and Encoder-Mixing Cross-Attention (EMCA) are gradually introduced to the baseline. As shown in Table 4, HAE brings 0.6 AP improvement to the baseline model. This benefits from the HAE enlarges the receptive field and introduces the global context features to improve semantic representation. The" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "improvements of " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "AP_{M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "(+0.6\\mathrm{AP})" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "AP_{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "(+0.8\\mathrm{AP})" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " contribute to the final performance improvement. These results are good demonstrations of our motivation and design. The EMCA improves the detection performance from 49.0 AP to 49.8 AP. It can be seen that the results on small, medium, and large objects are all improved, which means that object queries adaptively interacting with multi-level encoder features by cross-attention can extract more discriminative object features for improved multi-scale learning. When both components are included in the baseline model, it achieves 50.2 AP with 1.2 AP performance improvement. The results show that the two components are complementary and address different issues in DETR [4]." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 153, + 279, + 459, + 343 + ], + "blocks": [ + { + "bbox": [ + 159, + 258, + 453, + 269 + ], + "lines": [ + { + "bbox": [ + 159, + 258, + 453, + 269 + ], + "spans": [ + { + "bbox": [ + 159, + 258, + 453, + 269 + ], + "type": "text", + "content": "Table 5: Comparisons of ours HAE with Deformable Encoder variants." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 279, + 459, + 343 + ], + "lines": [ + { + "bbox": [ + 153, + 279, + 459, + 343 + ], + "spans": [ + { + "bbox": [ + 153, + 279, + 459, + 343 + ], + "type": "table", + "html": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)Flops
Deform. Enc. (4 Points)49.066.653.532.052.363.0245G
Deform. Enc. (6 Points)48.465.952.731.051.662.9250G
Deform. Enc. (8 Points)48.566.152.830.551.862.8255G
HAE (Ours)49.667.054.231.852.963.8248G
", + "image_path": "3ff066831c3d2c3f05970021a16146a94feec72c0e811e4f28c37c5e9853fbaa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 533 + ], + "type": "text", + "content": "Comparisons of HAE with Deformable Encoder variants. To make a fair comparison with the deformable encoder [50], we conduct experiments on increasing sampling points in the deformable encoder. The results are shown in Table 5. The performance of baseline with 4 sampling points is 49.0 AP. Performance decreases as sample points in the deformable encoder increases to 6 and 8, which are 48.4 AP and 48.5 AP, respectively. Even increasing the number of sample points in the deformable encoder does not improve performance for large and medium objects. This may be because deformable attention determines sampling weights based on query alone, leading to inaccurate weights as the number of points increases. However, our HAE can achieve performance gains (+0.6 AP) with less computational consumption (248G vs 250G). We think the possible reason is that the weight generation in standard self-attention, which calculates the similarity by query and key, is more suitable for long-distance sampling. These results support the motivation and design of our hybrid attention encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "The number of Hybrid Attention Layers. Introducing too much self-attention may lead to slow convergence. We analyze the effects of the number of hybrid attention layers. The results are shown in Table 6. When only a hybrid attention layer is introduced, the model achieves 49.3 AP with 0.3 AP improvement. The improvement is mainly contributed by large and medium objects. Increasing the number of hybrid attention layers by 2 or 3 results in performance gains of 0.6 and 0.7 AP, respectively. It is still the large and medium objects that contribute the most performance gains. As the number of hybrid attention layers increases, the performance gains begin to decrease. In particular, there is no performance gain when increasing to 6 hybrid attention layers. This may be because the negative effect of slow convergence becomes dominant." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 200, + 136, + 411, + 232 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 449, + 125 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 449, + 125 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 449, + 125 + ], + "type": "text", + "content": "Table 6: Ablations on the number of hybrid attention layers in HAE." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 136, + 411, + 232 + ], + "lines": [ + { + "bbox": [ + 200, + 136, + 411, + 232 + ], + "spans": [ + { + "bbox": [ + 200, + 136, + 411, + 232 + ], + "type": "table", + "html": "
NumberAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
-49.066.653.532.052.363.0
149.366.653.731.152.763.7
249.667.054.231.852.963.8
349.767.254.332.352.864.0
449.367.053.831.952.763.5
549.366.854.031.452.863.1
648.866.353.430.752.263.1
", + "image_path": "a786e5eaadb9ccb949e2a675afa982b3b94ef6695a55bb85a2758ff121ca8d3c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 148, + 277, + 462, + 364 + ], + "blocks": [ + { + "bbox": [ + 132, + 246, + 481, + 268 + ], + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 268 + ], + "type": "text", + "content": "Table 7: Comparisons of ours EMCA with other decoder variants. * indicates that these models scale to the similar complexity as EMCA." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 277, + 462, + 364 + ], + "lines": [ + { + "bbox": [ + 148, + 277, + 462, + 364 + ], + "spans": [ + { + "bbox": [ + 148, + 277, + 462, + 364 + ], + "type": "table", + "html": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_{2th} \\)\\( AP_{6th} \\)\\( AP_{10th} \\)Flops
DINO [42]49.066.653.534.243.245.7245G
+Iterative Enc [40]47.765.252.133.442.044.6245G
+Iterative Enc*47.465.151.633.242.044.1284G
+Memory Fusion [25]49.367.053.735.743.746.2253G
+Memory Fusion*49.366.953.935.743.546.2290G
+EMCA (Ours)49.867.454.636.044.546.6283G
", + "image_path": "e71a30fc47c2f0d942e20fab1ad587cba660c169cb1e3a2d05e58a72b628382b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 481, + 532 + ], + "type": "text", + "content": "Comparisons of EMCA with decoder variants. We compare our EMCA with other decoder variants. As shown in Table 7, using the Iterative Encoder [40] degrades the performance of the baseline. There are 0.3 AP performance gains when the Memory Fusion [25] is introduced, but there is no performance gain from the scaling of the approach. The performance gain of Memory Fusion is 0.5 AP less than our EMCA. These results demonstrate the effectiveness of our method design, which adaptively fuses multi-level encoder features based on query features for multi-scale objects, rather than using the same features for all objects. In addition, to show that our method converges faster, we also show results for the 2nd, 6th, and 10th epochs. It can be seen that our method outperforms the other methods throughout the training process, indicating the faster convergence of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": "Comparisons of EMCA with DINO variants. For a fair comparison, we compared with more complex baseline models due to the increased computational complexity introduced by EMCA. More complex baseline models are implemented by introducing more encoders and features. The results are shown in Table 8. By increasing the original baseline from 6 encoders to 7 encoders, the model reached 49.3 AP with 268G. When using 8 encoders, the model achieves 49.5 AP with 292G. While performance can be improved by stacking encoders, our EMCA utilizes encoder redundancy information and is a good complement to encoders. When using 10 encoders, the model suffers from overfitting and achieves only 47.8 AP with 339G. However, our EMCA with 10 encoders still has performance gains to reach 50.6 AP. This results show that our EMCA can" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 135, + 463, + 244 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 427, + 126 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 427, + 126 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 427, + 126 + ], + "type": "text", + "content": "Table 8: Comparisons of ours EMCA with DINO variants." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 135, + 463, + 244 + ], + "lines": [ + { + "bbox": [ + 148, + 135, + 463, + 244 + ], + "spans": [ + { + "bbox": [ + 148, + 135, + 463, + 244 + ], + "type": "table", + "html": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)Flops
6 encoder (Baseline)49.066.653.532.052.363.0245G
7 encoder49.366.753.732.052.463.7268G
8 encoder49.566.954.031.952.863.4292G
10 encoder47.864.852.130.650.961.5339G
5 scale49.466.953.832.352.563.9721G
6encoder+EMCA (Ours)49.867.454.633.052.864.7283G
8encoder+EMCA (Ours)50.267.854.632.453.364.6346G
10encoder+EMCA (Ours)50.668.155.333.553.965.3409G
", + "image_path": "1f0593880f19c08a67f2968fa7224fce85c85be2f7d1c54fd98476d1814e5277.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 269, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 479, + 316 + ], + "type": "text", + "content": "mitigate the risk of overfitting. A very large amount of computation (721G) is required when 5 scales of backbone features are used, and then the performance reaches only 49.4 AP. These results show that our EMCA is a good complement to encoders and can mitigate the risk of overfitting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 318, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 510 + ], + "type": "text", + "content": "Quantitative analysis of EMCA. To analyze the weights of EMCA, we count the fusion weights in the last decoder layer from the COCO val2017 set. First, we analyzed the weights of different encoder layers, and the results are shown in Figure 4. It can be seen that our method learns different weights for encoder layers. Specifically, the first encoder layer plays the most important role, the 2nd, 3rd, and 4th encoder layers have decreasing importance, and the 4th, 5th, and 6th encoder layers have similar roles. As shown in Table 9, using the 2nd, 4th, and 6th encoder layers as inputs, our approach drops 0.3 AP and saves 23G of computational overhead. When only the 1st and 6th encoder layers are used, the computation is reduced by 30G while the performance is dropped by 0.2 AP. We call the EMCA with only the 1st and 6th encoder layers as EMCA-Lite. We then analyze the weights of the objects at different scales to illustrate that our method improves multi-scale learning. As shown in Figure 5, objects of different sizes are given different weights, and the larger the scale, the higher the weights. See the Appendix for object scale intervals. These results illustrate the effectiveness of fusion weights based on queries." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 139, + 533, + 293, + 619 + ], + "blocks": [ + { + "bbox": [ + 139, + 533, + 293, + 619 + ], + "lines": [ + { + "bbox": [ + 139, + 533, + 293, + 619 + ], + "spans": [ + { + "bbox": [ + 139, + 533, + 293, + 619 + ], + "type": "image", + "image_path": "3e519da46a9aa1b3a905a5dd92e71b061434d391bd679246b694a1cb54e6c989.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 628, + 299, + 662 + ], + "lines": [ + { + "bbox": [ + 131, + 628, + 299, + 662 + ], + "spans": [ + { + "bbox": [ + 131, + 628, + 299, + 662 + ], + "type": "text", + "content": "Fig. 4: Statistical fusion weights for different encoders. Layer #6 means the last-level encoder." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 533, + 474, + 619 + ], + "blocks": [ + { + "bbox": [ + 319, + 533, + 474, + 619 + ], + "lines": [ + { + "bbox": [ + 319, + 533, + 474, + 619 + ], + "spans": [ + { + "bbox": [ + 319, + 533, + 474, + 619 + ], + "type": "image", + "image_path": "e5df4497c2d98d95ddb2dcfb7ee420f32edcf0ed38029689c5ecf32e78402bd6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 628, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 313, + 628, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 482, + 662 + ], + "type": "text", + "content": "Fig.5: Statistical fusion weights for different scale objects. Size #1 to size #6 are ordered by scale from small to large." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 146, + 288, + 222 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 299, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 299, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 299, + 136 + ], + "type": "text", + "content": "Table 9: Ablations on the layer of input encoders in EMCA." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 146, + 288, + 222 + ], + "lines": [ + { + "bbox": [ + 144, + 146, + 288, + 222 + ], + "spans": [ + { + "bbox": [ + 144, + 146, + 288, + 222 + ], + "type": "table", + "html": "
EncoderAPAP50AP75Flops
-49.066.653.5245G
12345649.867.454.6283G
24649.567.253.9260G
2649.367.253.8253G
1649.667.454.3253G
", + "image_path": "fc6cbe0f80a5abb32c773f8e7905a0949d076dc53c784bd227f02571524dddc1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 313, + 146, + 477, + 222 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 480, + 136 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 480, + 136 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 480, + 136 + ], + "type": "text", + "content": "Table 10: Ablations on the fusion methods in EMCA." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 146, + 477, + 222 + ], + "lines": [ + { + "bbox": [ + 313, + 146, + 477, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 477, + 222 + ], + "type": "table", + "html": "
Fusion MethodAPAPSAPMAPL
-49.032.052.363.0
sum49.331.952.563.9
max48.630.951.663.2
sigmoid weighted49.833.052.864.7
softmax weighted49.532.152.864.5
", + "image_path": "2f239f909b7fed2d463c90387c53ad1bdcdd7438ad5d1ce6dfbf51f6d9c88f42.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 468 + ], + "type": "text", + "content": "Fusion Methods in EMCA. The fusion methods we explore include sum fusion, max fusion, and adaptive weighted fusion (sigmoid and softmax). All results are shown in Table 10. By using sum fusion, our method improves the baseline by 0.3 AP. This result shows that multi-level encoder features are richer in semantic information than single-level encoder features. When using max fusion, our method achieves 48.6 AP below the baseline. It can be seen that the performance decreases significantly for small and medium objects, but increases slightly for large objects. This may be because large objects tend to have higher activation so the features of small and medium objects are weakened during max fusion. When using adaptive weighted fusion, our method improves the baseline by 0.8 AP and 0.5 AP, respectively. It can be seen that the performance of small, medium, and large objects is all improved. In particular, the performance on large objects has improved significantly (+1.7 AP) by sigmoid weighted fusion. This may be because the larger the object, the more information is activated at all levels of the encoder, and therefore the more information can be used. In addition, the corresponding query features of large objects are richer so that the learned fusion weights are more accurate. These results suggest that sigmoid weighted fusion is a better fusion method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 502, + 220, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 220, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 220, + 515 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we analyze the limitations of DETR detectors for multi-scale learning and propose local feature enhancement and multi-level encoder exploitation techniques to address the limitations. Based on these, we construct a novel detection transformer detector named Augmented DETR (AugDETR), which consists of two components: Hybrid Attention Encoder and Encoder-Mixing Cross-Attention. HAE enlarges the receptive field and introduces global context to enhance the local feature of the deformable encoder. EMCA uses multi-level encoders to obtain more discriminative features and speed up convergence. Extensive experiments with various DETR-based detectors and backbones validate the effectiveness of our method." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 239 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (NSFC) under Grant 62088102. This work is partly supported by China Telecom Corporation Ltd. Data&Al Technology Company, Beijing, China and China Telecom - Xi'an Jiaotong University Joint Innovation Institute for Science-Education Integration of Intelligent Cloud Network. We also appreciate Jingwen Fu, He Zhang and Tao Yang for their insightful discussions. We thank all the anonymous reviewers and Area Chairs for their constructive and helpful comments, which have significantly improved the quality of the paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 259, + 197, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 197, + 272 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 197, + 272 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 286, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 286, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 480, + 319 + ], + "type": "text", + "content": "1. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154-6162 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "text", + "content": "2. Cai, Z., Liu, S., Wang, G., Ge, Z., Zhang, X., Huang, D.: Align-detr: Improving detr with simple iou-aware bce loss. arXiv preprint arXiv:2304.07527 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "text", + "content": "3. Cao, X., Yuan, P., Feng, B., Niu, K.: Cf-detr: Coarse-to-fine transformers for end-to-end object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 185-193 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 376, + 480, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 409 + ], + "type": "text", + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "text", + "content": "5. Chen, Q., Chen, X., Zeng, G., Wang, J.: Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "text", + "content": "6. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2988-2997 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "text", + "content": "7. Ding, X., Zhang, X., Han, J., Ding, G.: Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11963-11975 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 499, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 480, + 532 + ], + "type": "text", + "content": "8. Dong, J., Huang, Y., Zhang, S., Chen, S., Zheng, N.: Construct effective geometry aware feature pyramid network for multi-scale object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 534-541 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "text", + "content": "9. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 566, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 566, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 566, + 480, + 597 + ], + "type": "text", + "content": "10. Gao, P., Zheng, M., Wang, X., Dai, J., Li, H.: Fast convergence of detr with spatially modulated co-attention. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3621-3630 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "text", + "content": "11. Ghiasi, G., Lin, T.Y., Le, Q.V.: Nas-fpn: Learning scalable feature pyramid architecture for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7036-7045 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "12. Guo, C., Fan, B., Zhang, Q., Xiang, S., Pan, C.: Augfpn: Improving multi-scale feature learning for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12595-12604 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "text", + "content": "AugDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. He, Y., Zhu, C., Wang, J., Savvides, M., Zhang, X.: Bounding box regression with uncertainty for accurate object detection. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 2888-2897 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "14. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: Detrs with hybrid matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19702-19712 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "text", + "content": "15. Kim, S.W., Kook, H.K., Sun, J.Y., Kang, M.C., Ko, S.J.: Parallel feature pyramid network for object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 234-250 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "text", + "content": "16. Kong, T., Sun, F., Liu, H., Jiang, Y., Li, L., Shi, J.: Foveabox: Beyound anchor-based object detection. IEEE Transactions on Image Processing 29, 7389-7398 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 481, + 267 + ], + "type": "text", + "content": "17. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. In: Proceedings of the European conference on computer vision (ECCV). pp. 734-750 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "type": "text", + "content": "18. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 300, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 481, + 342 + ], + "type": "text", + "content": "19. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems 33, 21002-21012 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "text", + "content": "20. Lin, T.Y., Dólar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "type": "text", + "content": "21. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 407, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 481, + 450 + ], + "type": "text", + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 481, + 482 + ], + "type": "text", + "content": "23. Lin, Y., Yuan, Y., Zhang, Z., Li, C., Zheng, N., Hu, H.: Detr does not need multiscale or locality design. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6545-6554 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "type": "text", + "content": "24. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 481, + 546 + ], + "type": "text", + "content": "25. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., et al.: Detection transformer with stable matching. arXiv preprint arXiv:2304.04742 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 546, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 546, + 481, + 578 + ], + "type": "text", + "content": "26. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8759-8768 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 578, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 481, + 622 + ], + "type": "text", + "content": "27. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 21-37. Springer (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "text", + "content": "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "30. Nguyen, D.K., Ju, J., Booij, O., Oswald, M.R., Snoek, C.G.: Boxer: Box-attention for 2d and 3d transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4773-4782 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "text", + "content": "31. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra r-cnn: Towards balanced learning for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 821-830 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "text", + "content": "32. Pu, Y., Liang, W., Hao, Y., Yuan, Y., Yang, Y., Zhang, C., Hu, H., Huang, G.: Rank-detr for high quality object detection. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "type": "text", + "content": "33. Qiao, S., Chen, L.C., Yuille, A.: Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10213-10224 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "type": "text", + "content": "34. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 39(6), 1137-1149 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "text", + "content": "35. Roh, B., Shin, J., Shin, W., Kim, S.: Sparse detr: Efficient end-to-end object detection with learnable sparsity. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "text", + "content": "36. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "37. Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9627-9636 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "text", + "content": "38. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "39. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "text", + "content": "40. Zhang, G., Luo, Z., Tian, Z., Zhang, J., Zhang, X., Lu, S.: Towards efficient use of multi-scale features in transformer-based object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6206-6216 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "41. Zhang, G., Luo, Z., Yu, Y., Cui, K., Lu, S.: Accelerating detr convergence via semantic-aligned matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 949-958 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "text", + "content": "42. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 589, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 482, + 621 + ], + "type": "text", + "content": "43. Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8514-8523 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "type": "text", + "content": "44. Zhang, H., Chang, H., Ma, B., Wang, N., Chen, X.: Dynamic r-cnn: Towards high quality object detection via dynamic training. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 260-275. Springer (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "text", + "content": "AugDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 348 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "45. Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7329-7338 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "46. Zheng, D., Dong, W., Hu, H., Chen, X., Wang, Y.: Less is more: Focus attention for efficient detr. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6674-6683 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "47. Zhou, X., Wang, D., Krähenbuhl, P.: Objects as points. arXiv preprint arXiv:1904.07850 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "48. Zhou, X., Zhuo, J., Krahenbuhl, P.: Bottom-up object detection by grouping extreme and center points. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 850-859 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "49. Zhu, C., He, Y., Savvides, M.: Feature selective anchor-free module for single-shot object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 840-849 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 313 + ], + "type": "text", + "content": "50. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 314, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 481, + 348 + ], + "type": "text", + "content": "51. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6748-6758 (2023)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_content_list.json b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fb33ec4e323505099ebfa21daf50cff94f7ad992 --- /dev/null +++ b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_content_list.json @@ -0,0 +1,1530 @@ +[ + { + "type": "text", + "text": "AugUndo: Scaling Up Augmentations for Monocular Depth Completion and Estimation", + "text_level": 1, + "bbox": [ + 238, + 140, + 764, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yangchao Wu $^{1}$ , Tian Yu Liu $^{1}$ , Hyoungseob Park $^{2}$ , Stefano Soatto $^{1}$ , Dong Lao $^{1}$ , and Alex Wong $^{2}$", + "bbox": [ + 303, + 210, + 699, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ UCLA Vision Lab, Los Angeles, CA 90095, USA {wuyangchao1997, tianyu, soatto, lao}@cs.ucla.edu", + "bbox": [ + 308, + 253, + 691, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Yale Vision Lab, New Haven, CT 06511, USA {hyoungseob.park, alex.wong}@yale.edu", + "bbox": [ + 339, + 282, + 661, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Unsupervised depth completion and estimation methods are trained by minimizing reconstruction error. Block artifacts from resampling, intensity saturation, and occlusions are amongst the many undesirable by-products of common data augmentation schemes that affect image reconstruction quality, and thus the training signal. Hence, typical augmentations on images viewed as essential to training pipelines in other vision tasks have seen limited use beyond small image intensity changes and flipping. The sparse depth modality in depth completion have seen even less use as intensity transformations alter the scale of the 3D scene, and geometric transformations may decimate the sparse points during resampling. We propose a method that unlocks a wide range of previously-infeasible geometric augmentations for unsupervised depth completion and estimation. This is achieved by reversing, or \"undo\"-ing, geometric transformations to the coordinates of the output depth, warping the depth map back to the original reference frame. This enables computing the reconstruction losses using the original images and sparse depth maps, eliminating the pitfalls of naive loss computation on the augmented inputs and allowing us to scale up augmentations to boost performance. We demonstrate our method on indoor (VOID) and outdoor (KITTI) datasets, where we consistently improve upon recent methods across both datasets as well as generalization to four other datasets. Code available at: https://github.com/alexklwong/augundo", + "bbox": [ + 261, + 345, + 743, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: depth completion $\\cdot$ monocular depth $\\cdot$ augmentations", + "bbox": [ + 261, + 664, + 705, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 703, + 375, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data augmentation is essential to training machine learning models; it plays a role in performance and generalization [40, 51, 55]. One common axiom of choosing augmentations is that the task output should remain invariant to the augmentation. For example, image flipping is a viable augmentation for classifying animals, since it does not alter the label. Conversely, flipping road signs can alter their meanings; hence, such augmentation can be detrimental to tasks involving road sign recognition. For geometric tasks, the range of augmentations", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is more restricted due to constraints of problem formulation: Stereo assumes pairs of frontoparallel rectified images; hence, in-plane rotations are not viable. Image-guided sparse depth completion relies on sparse points to ground estimates to metric scale; therefore, intensity transformations on sparse depth maps that alter the scale of the 3-dimensional (3D) scene are infeasible. Unsupervised learning of depth completion and estimation further limit the use of augmentations as the supervision signal comes from reconstructing the inputs, where augmenting the input introduces artifacts that impact reconstruction quality and therefore the supervision (see Sec. D, Fig. 1, Tab. 8, 9 in the Supp. Mat. for examples of artifacts and extended discussion on their effect on learning). Moreover, video-based unsupervised training assumes rigid motion, so augmentations that introduce padding (e.g., translation, rotation) will yield constant or edge extended borders across images (i.e., no motion), preventing the model from properly learning depth and pose - leaving few augmentations viable. While simulating nuisances is desirable, naively applying augmentations may do more harm than good; thus, it is unsurprising that existing work in unsupervised depth completion [33, 36, 50, 63-65, 67, 73] and estimation [14, 15, 35, 81, 85] primarily rely on a small range of photometric augmentations and flipping.", + "bbox": [ + 212, + 146, + 787, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, photometric augmentations help model the diverse range of illumination conditions and colors of objects that may populate the scene; geometric augmentations can simulate the various camera parameters, i.e., image resizing (zooming) can model changes in focal length, and scene arrangements, i.e., image flipping. However, block artifacts, loss during resampling, and intensity saturation are just some of the many undesirable side-effects of traditional augmentations to the image and sparse depth map for unsupervised learning of geometric tasks. To avoid compromising the supervision signal, we compute the typical reconstruction loss on the original input image and sparse depth map instead of the augmented inputs, which bypasses negative effects of reconstruction artifacts due to photometric and geometric augmentations. However, there exists a mis-alignment between the original input (e.g., image, sparse depth), and the model depth estimate as geometric augmentations induce a change in coordinates. Hence, we undo the geometric augmentations by inverting them in the output space to align the model estimate with the training target.", + "bbox": [ + 212, + 425, + 787, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Amongst the many geometric tasks, we focus on unsupervised depth completion, the task of inferring a dense depth map from an image and sparse depth map, where augmentations have seen limited use. Here, a training sample includes the input sparse depth map, its associated image and additional images of neighboring views of the same 3D scene. Our method is also applicable to unsupervised monocular depth estimation, which omits the sparse depth modality. Augmentations have traditionally been restricted to a limited range of photometric transformations and flipping - due to the need to preserve photometric consistency across a sequence of video frames used during training, and the sparse set of 3D points projected onto the image frame as a 2.5D range map; degradation to either modalities directly impacts the supervision signal as the loss function is conventionally computed on the augmented inputs. By using our", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/83c8538fd14440956fcbe2378d683cabcd091e86b8f88e1b9e7d80338566e3b0.jpg", + "image_caption": [ + "Fig. 1: Overview. We apply photometric augmentations to the input image, and the same set of geometric augmentations to both the input image and sparse depth map. We warp the output depth back to the original reference frame with the inverse geometric transformations. This enables image and sparse depth reconstruction losses to be computed on the original inputs while unlocking previously-infeasible augmentations." + ], + "image_footnote": [], + "bbox": [ + 295, + 143, + 702, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "method, loss functions involving sparse depth and image reconstruction from other views can be computed on the original inputs while applying augmentations that were previously not viable for the task. Our hypothesis: By \"undoing\" the augmentations, one can expand the viable set and scale up their use in training, leading to improved model performance and generalizability.", + "bbox": [ + 212, + 416, + 782, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we introduce AugUndo, an augmentation framework that enables one to apply a wide range of photometric and geometric transformations on the inputs, and to \"undo\" them during loss computation. This allows computing the unsupervised loss on the original images and sparse depth maps, free of artifacts, through a warping of the output depth map - obtained from augmented input onto the input frame of reference based on the inverse geometric transformation. In addition to group transformations that allow for output alignment, we combine them with commonly employed photometric augmentations. To the best of our knowledge, we are the first to propose a unified augmentation scheme for photometric and geometric augmentations for unsupervised depth completion and estimation. We demonstrate AugUndo on recent methods on indoor and outdoor settings, where we consistently improve all methods across all datasets.", + "bbox": [ + 212, + 492, + 784, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are as follows: (1) We propose AugUndo, a simple-yet-effective framework to scale up photometric and geometric augmentations for unsupervised depth completion and estimation, without compromising the supervision signal; AugUndo can be applied in a plug-and-play manner to existing methods with negligible increase in computational costs during training. (2) We enable previously-infeasible augmentations to be used for training unsupervised methods and comprehensively ablate combinations of eleven types of augmentations to study the performance benefits of each. (3) We show that AugUndo can consistently improve robustness to shifts in sparse point densities for completion, model performance as well as zero-shot generalization for both depth completion and estimation for indoor and outdoor scenarios; thus, validating our hypothesis.", + "bbox": [ + 212, + 674, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation", + "bbox": [ + 235, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data augmentation for depth completion and estimation. While photometric transformations such as color jitter are often applied for unsupervised depth completion [36,63-65], and estimation [10,14-16,26,27,41,47,66,79,80], geometric augmentations other than simple flipping are less commonly adopted. Most supervised depth completion methods [8,18,19,21,23,29,31,43-45,69,73,82,83] similarly limit their augmentations to color jitter and horizontal flips due to sparse depth maps being decimated by rotation and scaling, causing points to be interpolated away. Nevertheless, for supervised training, it is straightforward to directly apply the same transformation to the ground truth for training. Indeed, some supervised depth completion methods [2,3,9,38,39,53,59,78] adopted random scaling, translation, in-plane rotation. Translation augmentations are also commonly applied in supervised depth estimation methods [6,7,25,30,32, 46,47,58,62,68,71,76]. We posit that the reason that such transformations can be applied to the ground truth in supervised settings is due to artifacts caused by the transformation of a piece-wise smooth depth map being less severe than those of an RGB image and its intensities. However, such assumptions do not hold in unsupervised training. These artifacts would affect the training signal for unsupervised methods, which rely on photometric correspondences and observed sparse points. Our approach aims to resolve this to allow diverse geometric augmentations to be applied in a plug-and-play fashion in unsupervised training.", + "bbox": [ + 212, + 175, + 787, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unsupervised depth completion methods [36, 50, 63-65, 70, 73] leverage image and sparse depth reconstructions as training signals by minimizing errors between the input image and its reconstruction from other views, and errors between the input sparse depth map and the predicted depth along with a local smoothness regularizer. [36] used Perspective-n-Point [28] and RANSAC [11] to align consecutive video frames. [73] learns a depth prior conditioned on the image. [63] uses synthetic data to learn a prior on the shapes populating a scene, while [34] translates synthetic data to real domain to leverage rendered depth. [64] proposed an adaptive scheme to reduce penalties incurred on occluded regions. [67] maps the image onto the 3D scene through calibrated backprojection. [70] decouples structure and scale. [20] uses line features from visual SLAM and [33] introduced monitored distillation for positive congruent training. Augmentations for these methods are limited to a small range of photometric perturbations and image flipping. Operations such as rotation, resizing, and translation require resampling, which creates artifacts and affects the reconstruction quality. This causes performance degradation since loss is typically computed on the augmented images. Loss in sparse depth maps is further impacted as resampling and interpolation may cause loss of sparse points. Contrary to these limited augmentation schemes, we enable a large range of photometric and geometric augmentations to be introduced during training.", + "bbox": [ + 212, + 478, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unsupervised monocular depth estimation, like depth completion, also minimizes photometric reconstruction error. [12] frames depth estimation as a novel view synthesis problem. [14] improves [12] by imposing a consistency loss on the depth predicted from left and right images. [85] uses a pose network to", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "enable unsupervised training on video sequences. [15] introduced auto-masking and min reprojection loss. To improve the supervision signal, [4, 56, 61] leverage noisy proxy labels and [42] uses a trinocular assumption. Additional loss terms based on visual odometry [60], iterative closest point [37], surface normals [75], and semantic segmentation [17, 24] were also introduced; [41, 72] further included predictive uncertainty. To handle rigid and non-rigid motions in the scene, previous works explored multi-task learning to include optical flow and moving object estimation [1, 48, 74, 77, 86], and used semantics to filter out outlier regions [22]. [35] redesigned the skip connection and decoders to extract high-resolution features, [84] combined global and local representations and [81] introduced a lightweight architecture with dilated convolution and attention. In addition to depth completion, our method also shows consistent improvements on monocular depth estimation for [15, 35, 81].", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method Formulation", + "text_level": 1, + "bbox": [ + 215, + 364, + 455, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $I:\\varOmega\\subset\\mathbb{R}^2\\to\\mathbb{R}_+^3$ be an RGB image captured by a calibrated camera, $z:\\varOmega_z\\subset\\varOmega\\rightarrow\\mathbb{R}_+$ the corresponding sparse point cloud projected onto the image plane as a sparse depth map, and $K\\in \\mathbb{R}^{3\\times 3}$ the intrinsic calibration matrix. Given an image and its sparse depth map, depth completion aims to learn a function $f_{\\theta}(I,z)$ that recovers the distance between the camera to points in the 3D scene as a dense depth map. In another mode, if sparse depth maps are not given, then the problem reduces to monocular depth estimation which learns a function $f_{\\theta}(I)$ to map a single image to a depth map $\\varOmega\\rightarrow\\mathbb{R}_{+}$ . For the ease of notation, we denote the output depth map for both depth completion and estimation as $\\hat{d}\\in \\mathbb{R}_{+}^{H\\times W}$ where $H$ and $W$ are its height and width.", + "bbox": [ + 212, + 396, + 787, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unsupervised depth completion relies on photometric and sparse depth reconstruction errors as its primary supervision signal. Without loss of generality, we assume an input of $(I_t,z_t)$ for an RGB image and associated sparse depth map captured at time $t$ and during training, an additional set of temporally adjacent images $I_{\\tau}$ for $\\tau \\in \\Upsilon \\doteq \\{t - 1,t + 1\\}$ . The reconstruction $\\hat{I}_{t\\tau}$ of $I_{t}$ from image $I_{\\tau}$ is given by the reprojection based on estimated depth $\\hat{d}_t\\coloneqq f_\\theta (\\cdot)$", + "bbox": [ + 212, + 547, + 787, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {t \\tau} (x) = I _ {\\tau} \\left(\\pi g _ {\\tau t} K ^ {- 1} \\bar {x} \\hat {d} _ {t} (x)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 652, + 784, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{x} = [x^{\\top},1]^{\\top}$ is the homogeneous coordinates of $x\\in \\Omega$ , $g_{\\tau t}\\in SE(3)$ the relative pose of the camera from time $t$ to $\\tau$ , $K$ the intrinsic calibration matrix, and $\\pi$ the canonical perspective projection.", + "bbox": [ + 212, + 679, + 787, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using Eq. (1), a depth completion or estimation network $f_{\\theta}$ minimizes", + "bbox": [ + 238, + 726, + 750, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\theta} \\sum_ {\\tau \\in \\Upsilon} \\sum_ {x \\in \\Omega} \\alpha \\rho (\\hat {I} _ {t \\tau} (x), I _ {t} (x)) + \\sum_ {x \\in \\Omega_ {z}} \\beta \\psi (\\hat {d} _ {t} (x), z _ {t} (x)) + \\lambda R (\\hat {d} _ {t}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 753, + 785, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho$ denotes the photometric reconstruction error, typically $L_{1}$ difference in pixel values and/or structural similarity (SSIM), $\\psi$ the sparse depth reconstruction error, typically $L_{1}$ or $L_{2}$ , $R$ the regularizer that biases the depth map to", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation", + "bbox": [ + 233, + 113, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "be piece-wise smooth with depth discontinuities aligned with edges in the image (commonly used by previous works [15,36,65,67]), and $\\alpha$ , $\\beta$ and $\\lambda$ their respective weighting. Note: monocular depth omits sparse depth term, i.e., $\\beta = 0$ .", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inverse transformation. Let $\\mathcal{A}_{pt}$ be the set of possible photometric transformations, and $\\mathcal{A}_{ge}$ be the set of all geometric transformations. Given $T_{ge} \\in \\mathcal{A}_{ge}$ , we wish to obtain an inverse transformation $T_{ge}^{-1}$ such that $T_{ge} \\circ T_{ge}^{-1} \\approx Id$ the identity function3. At each time step, we can sample a sequence of transformations $\\{T_{pt}^{1} \\ldots T_{pt}^{k}\\}$ and $\\{T_{ge}^{1} \\ldots T_{ge}^{m}\\}$ respectively from $\\mathcal{A}_{pt}$ and $\\mathcal{A}_{ge}$ to construct transformations $T_{pt} = T_{pt}^{1} \\circ \\dots \\circ T_{pt}^{k}$ and $T_{ge} = T_{ge}^{1} \\circ \\dots \\circ T_{ge}^{m}$ . We denote the composition of a collection of augmentation transformations by $T = T_{ge} \\circ T_{pt}$ where $T_{pt}$ denotes photometric transformation, and $T_{ge}$ denotes geometric transformation. Furthermore, we denote the inverse geometric transformations by $T_{ge}^{-1} = (T_{ge}^{m})^{-1} \\circ (T_{ge}^{m-1})^{-1} \\circ \\dots \\circ (T_{ge}^{1})^{-1}$ , which operates on the space of depth maps to reverse the geometric transformation so that we can warp the output depth map onto the reference frame of the original image. Through inverse warping, this process is differentiable. In practice, some transformations cause border regions of the image to be warped out of frame, i.e., translated off the image plane or cropped away. Hence, after warping our output depth map back to the original frame of reference using the inverse geometric transformations, border extensions (edge paddings) are introduced to handle out-of-frame regions.", + "bbox": [ + 212, + 191, + 784, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AugUndo. We apply each geometric transformation over image coordinates and resample (bilinear for image and nearest neighbor for sparse depth):", + "bbox": [ + 212, + 450, + 782, + 479 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ x ^ {\\prime} 1 \\right] ^ {\\top} = T _ {g e} \\left[ x 1 \\right] ^ {\\top} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 489, + 784, + 510 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {\\prime} \\left(x ^ {\\prime}\\right) = T _ {p t} (I) (x); z ^ {\\prime} \\left(x ^ {\\prime}\\right) = z (x) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 512, + 784, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T_{ge}$ is the geometric transformation, $x \\in \\Omega$ and $x' \\in \\Omega$ are coordinates in the image grid, and $I'$ is the image after the transformation; for ease of notation, we hereafter denote $I' = T(I) = T_{ge} \\circ T_{pt}(I)$ to include both augmentations through composition. Note that $x$ is in the original image reference frame and $x'$ is in the transformed image reference frame. Naturally, this process can be extended to multiple geometric transformations by composing them, i.e., $T_{ge} = T_{ge}^{1} \\circ T_{ge}^{2} \\circ \\dots \\circ T_{ge}^{m}$ . The reverse process is simply inverting the transformations where $T_{ge}^{-1} = (T_{ge}^{m})^{-1} \\circ (T_{ge}^{m-1})^{-1} \\circ \\dots \\circ (T_{ge}^{1})^{-1}$ :", + "bbox": [ + 214, + 537, + 782, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ x ^ {\\prime \\prime} 1 \\right] ^ {\\top} = T _ {g e} ^ {- 1} \\left[ x ^ {\\prime} 1 \\right] ^ {\\top} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 672, + 784, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {d} \\left(x ^ {\\prime \\prime}\\right) = \\hat {d} ^ {\\prime} \\left(x ^ {\\prime}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 696, + 784, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{d}'$ is the depth map inferred from augmented inputs $(I', z')$ . Once reverted back to the original reference frame, $\\hat{d}$ is used to reconstruct $I_t$ from $I_\\tau$ for $\\tau \\in \\Upsilon$ as in Eq. (1). More details can be found in Alg. 1 in the Supp. Mat.", + "bbox": [ + 212, + 724, + 782, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By modeling $T_{ge}^{-1}$ , Eq. (3)-Eq. (6) allow us to apply a wide range of augmentations, while still establishing the correspondence between $I_{t}$ and $I_{\\tau}$ . Specifically,", + "bbox": [ + 212, + 772, + 782, + 803 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3 Note that in practice, not all transformations are bijective. For instance, image rescaling with a fixed size image plane; hence strict equality is not always possible.", + "bbox": [ + 217, + 809, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for minimizing Eq. (2), one may simply augment the image and sparse depth by $T$ to yield $(I_t', z_t')$ as input, and reconstruct the original image and sparse depth $(I_t, z_t)$ from other views $I_\\tau$ and the aligned output depth $\\hat{d}$ (Eq. (6)). We note that the inverse transformation is critical for enabling the sparse depth reconstruction term to be computed properly in Eq. (2); if computed in the transformed reference frame, i.e., on $z_t'$ , many of the sparse points would be decimated by interpolation during rotation and resizing (downsampling) augmentations - leaving a lack of supervision from the sparse depth term of the loss function. We note that the original image sequence is fed to the pose estimator, rather than the augmented images. This ensures the estimated relative pose used during reprojection (Eq. (1)) also corresponds to the original images.", + "bbox": [ + 212, + 146, + 787, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modeling geometric augmentations. Depth completion differs from typical image-based (e.g., monocular) depth estimation, where operations like resizing are susceptible to scale ambiguity and can either correspond to changes in focal length, or distance from the camera (i.e., due to changes in camera pose). In depth completion, the sparse depth map grounds image pixels to specific depth values. Hence, there is no ambiguity in distance of each sparse point from the camera. Our goal is to synthesize new training data via geometric transformations that are consistent with the original 3D scene. Given that there is no scale ambiguity, this can be achieved by either (1) changes in camera pose (extrinsics), or (2) changes in camera parameters (intrinsics).", + "bbox": [ + 212, + 316, + 787, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Augmentation via (1) would require warping the inputs for view synthesis. Since we operate under the unsupervised setting, accurate warping (preserving 3D scene) without access to dense ground truth depth values is difficult to achieve. An exception to this is rotation, which we address below. On the other hand, (2) can be achieved via standard 2D image transformations while preserving the 3D structure of the scene. This motivates our modeling of geometric transformations, i.e., resizing, translation, as changes in camera parameters.", + "bbox": [ + 212, + 470, + 787, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For example, augmenting focal length (i.e. zooming in/out) is akin to \"resizing\". Similarly, translation can be used to model shifting of the optic center, i.e. capturing the same 3D scene using cameras with different principal point offsets. Note: image rotation can be seen as rotating a camera about its optical axis, which does not change the distance of a point from the camera. Thus, the camera position and the scene are kept constant across augmentations, eliminating the need for adjusting depth values via warping during the process. Instead, we only have to realign the output depth back to the original frame of reference (Eqs. (5) and (6)), which makes our method computationally efficient.", + "bbox": [ + 212, + 578, + 787, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Augmentations. Our choices are based on common nuisance variabilities, i.e., changes in illumination, occlusions, and object color, scene layout (flip), and camera parameters (resize, translation) and orientation (rotation).", + "bbox": [ + 212, + 717, + 787, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Photometric. We include brightness, contrast, saturation, and hue, where all follow standard augmentation pipeline in existing works [15,65,67,81]. Applying the inverse of photometric augmentation can be viewed as recovering the original image; hence, we directly use original image instead of applying transformations to the intensities (which are not recoverable if saturated at the max value).", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation", + "bbox": [ + 235, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Occlusion. We consider patch removal and sparse point removal. For the former, we randomly select a percentage of pixels $x \\in \\Omega$ in the image and remove (with zero-fill) an arbitrary-sized patch around it. For the latter, we randomly sample a percentage of points $x \\in \\Omega_z$ in the sparse depth map. The inverse transformation of this is simply the original image and sparse depth map, both used in loss computation before augmentation.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Flip. We consider horizontal and vertical flips. When applied, the same flip operation is used for both input image and sparse depth maps. We record the flip type during data augmentation. During loss computation, we reverse the flip direction to align output depth with the original image and sparse depth map.", + "bbox": [ + 212, + 237, + 784, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Resize. We define a new image plane of the input resolution and generate a random scaling factor to be applied along both height and width directions. The image is warped to the new image plane, where any point warped out of the plane is excluded; borders of the warped image are extended to the bounds of the image plane by edge replication. Sparse points occupying multiple pixels are eroded to a single point for resizing with a factor greater than 1. During loss computation, we warp the output depth map onto a new image plane of the same dimensions by the inverse scaling matrix of the recorded scaling factors; borders of the warped depth map are extended by edge replication if necessary. We view resizing factors greater and less than 1 (zooming in and out) as two separate forms of augmentations to distinguish their contributions.", + "bbox": [ + 212, + 297, + 785, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Rotation. Naive rotation leads to the loss of large areas of the image, i.e., cropped away to retain the same-size image, discarding large portions of possible co-visible regions and also supervisory signals. To preserve the entire image, we first warp the image by a randomly generated angle to a new (larger) image plane, so that the rotated image fits tightly within the new image. As image sizes within a batch can vary depending on the rotation angle, we center-pad (uniformly on each side) each image in the batch to the maximum width and height of the augmented batch. To reverse the rotation on the output depth, we warp the output depth map back by the inverse rotation matrix, then perform a center crop on the depth map to align with the original input.", + "bbox": [ + 212, + 465, + 785, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Translation. We define a new image plane of the same dimensions as the input and generate random height and width translation factors. The coordinates of the input are translated and its pixels are inverse warped onto the new image plane. Any pixel warped out of the image plane is excluded. Borders of the warped image are extended to the bounds of the image plane by edge replication. During loss computation, we warp the output depth map onto a new image plane of the same dimensions by the inverse translation matrix. Borders of the warped depth map are extended to the bounds of the image plane by edge replication.", + "bbox": [ + 212, + 616, + 785, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 761, + 375, + 777 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "We demonstrate AugUndo on recent unsupervised depth completion (VOICED [65], FusionNet [63], KBNet [67]) and estimation (Monodepth2 [15], HRDepth [35], LiteMono [81]) methods on two datasets (KITTI [13, 57], Void [65]) here.", + "bbox": [ + 214, + 794, + 784, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to space limitations, we defer the results of MonDi [33] and DesNet [70] to Sec. K in the Supp. Mat. To evaluate zero-shot generalization, we test on three additional datasets for each task: NYUv2 [52], ScanNet [5], Waymo [54], and Make3d [49] - where we transfer models on Void to NYUv2 (Tab. 3) and ScanNet, and from KITTI to Waymo (Tab. 4 in Supp. Mat.) for completion and Make3d [49] (Tab. 5 in Supp. Mat.) for estimation. We also test model sensitivity to different sparse depth input densities in Tab. 1 (right) - see Sec. H of the Supp. Mat for an extensive study. Additionally, we present an comprehensive ablation study in Sec. G in the Supp. Mat. to test the effect of each augmentation. We also provide results on modeling AugUndo as changes in camera motion or parameters in Tab. 7 of the Supp. Mat.; interpreting geometric augmentations as different camera parameters yields better performance. Tables. 8, 9 in Supp. Mat. show that unsupervised models cannot be trained with naive geometric augmentations. See Sec. E of the Supp. Mat. for descriptions of datasets.", + "bbox": [ + 212, + 145, + 787, + 358 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Experiment setup. We followed the original settings of the open-sourced code for each work and modified the data handling and loss function to incorporate AugUndo. We perform 4 independent trials for each experiment and report their means and standard deviations. To ensure a fair comparison, we train all models from scratch. Below, we report the best combination of augmentations found empirically through extensive experiments on each dataset. See Secs. A and B in Supp. Mat. for implementation details, and evaluation metrics, respectively.", + "bbox": [ + 212, + 362, + 787, + 469 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Augmentations. Through a search over augmentation types and values, we found a consistent set that tends to yield improvements across all methods with small changes to degree of augmentation catered to each method. See Sec. C in Supp. Mat. for details of augmentation parameters. Rows in Tabs. 1 to 6 marked with \"+\" AugUndo\" denote models trained with our method. We note that performance gain can be obtained by typical set and ranges of augmentations and does not require a meticulous selection of hyper-parameters (see Sec. J of Supp. Mat. for a sensitivity study). Results reported here aims to study how far one can push performance and generalization, purely from augmentations.", + "bbox": [ + 212, + 473, + 787, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results on VOID. We begin by presenting quantitative results for AugUndo on unsupervised depth completion. Tab. 1 (left, VOID1500) shows our main results on the VOID benchmark. By training the models with AugUndo, we observe an average overall improvement of $18.3\\%$ across all methods and metrics on VOID1500. Specifically, we improve VOICED by $26.4\\%$ , FusionNet by $16.3\\%$ , and KBNet by $12.2\\%$ . These experiments validate our hypothesis that by applying a wider range of augmentations, we are able to improve the baseline model's performance. They also illustrate the lack in use of augmentations in existing works: incorporating standard augmentations (albeit requiring modification to augmentation and loss computation pipelines) can yield a large performance boost. Fig. 2 shows a head-to-head comparison between KBNet trained using standard procedure in [67] and KBNet trained using AugUndo. We observe qualitative improvements from AugUndo where we improve KBNet in homogeneous regions and image borders, i.e., pillar (left), cabinet (middle), wall (right). Applying geometric augmentations yields models with fewer border", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation", + "bbox": [ + 235, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f62ad62a1427835dca122bd750d2c99ee3b70cd8a06dd04334d2da87f8abc09f.jpg", + "table_caption": [ + "Table 1: Depth completion on Void. AugUndo improves performance by an average of $18.3\\%$ across all methods and metrics on Void1500. When models trained on Void1500 are tested Void500, AugUndo improves by $23.1\\%$ , as translation, resizing, and occlusion augmentations vary sparsity by removing sparse points from the input." + ], + "table_footnote": [], + "table_body": "
MethodVOID1500VOID500
MAE ↓RMSE ↓iMAE ↓iRMSE ↓MAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED74.78±2.69139.75±4.5739.20±1.4671.98±2.54137.01±4.23235.80±7.8271.36±1.86130.63±5.66
+ AugUndo52.73±0.41111.09±0.9226.93±0.5454.46±0.3892.99±1.11176.94±1.3846.43±0.8591.10±1.64
FusionNet52.11±0.44113.30±1.1828.53±0.5258.79±2.0197.73±0.73194.32±1.3658.65±1.31122.95±3.04
+ AugUndo41.16±0.1899.21±0.3922.23±0.3553.07±1.3074.97±0.69162.71±2.3940.44±1.3992.11±5.79
KBNet38.11±0.7795.22±1.7219.51±0.1446.70±0.4878.44±1.39178.17±3.2737.56±0.6183.43±1.89
+ AugUndo33.32±0.1885.67±0.3916.61±0.2941.24±0.6066.97±0.81151.55±2.0331.63±0.5371.90±0.82
", + "bbox": [ + 220, + 213, + 785, + 321 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/80be05bc519fb2d33fb4934d9933db4e29ba94e2eb280e65aa1a27a78ce9b7e9.jpg", + "image_caption": [ + "Fig. 2: Depth completion on VOID. We compare KBNet trained with standard augmentations and with AugUndo. AugUndo consistently produces lower errors with reduced border artifacts and improves on homogeneous regions, i.e., pillar (left), cabinet (middle), wall of staircase (right). Error maps are highlighted for comparison." + ], + "image_footnote": [], + "bbox": [ + 232, + 339, + 772, + 498 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "artifacts as we apply random translation, which warps part of the image out of frame. This allows us to simulate border occlusion, where there lacks correspondence in an adjacent frame. While training with standard protocol results in failures to recover structures near the image border, training with AugUndo can render models robust to them by computing the loss on the original frame of reference (where we do have correspondence) through our inverse transformations of the depth map. Additionally, resizing allows the model to learn multiple resolutions of the input, akin to zooming in and out, which can also impose smoothness in homogeneous regions through the scale-space transition; whereas, rotation can simulate diverse camera orientations. Fig. 2 shows that translation, rotation, and resizing can model these effect in the input space to yield models robust to these nuisance variability.", + "bbox": [ + 212, + 597, + 787, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "AugUndo also is applicable for monocular depth estimation. Tab. 2 shows a comparison using the standard augmentation procedure of Monodepth2, HR-Depth, and Lite-Mono and using AugUndo. Tab. 2 shows that AugUndo consistently improves all models across all error and accuracy metrics. Thus, validating", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c9044bedbcfc894358f8c46ae9b8a3fb4344dc65c02606238c139d1f55971422.jpg", + "table_caption": [ + "Table 2: Monocular depth estimation on VOID. AugUndo is applicable for monocular depth estimation and consistently improves three monocular depth models." + ], + "table_footnote": [], + "table_body": "
MethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.252 ↑δ < 1.253 ↑
Monodepth2283.861±3.732395.947±5.7280.183±0.0020.100±0.0030.717±0.0050.922±0.0040.975±0.002
+ AugUndo277.696±4.861388.088±5.7680.178±0.0030.095±0.0040.724±0.0070.925±0.0040.978±0.002
HR-Depth286.282±7.059399.112±9.1840.185±0.0040.100±0.0040.714±0.0120.919±0.0060.975±0.002
+ AugUndo283.086±6.787394.261±9.1330.181±0.0050.097±0.0050.718±0.0130.922±0.0040.977±0.002
Lite-Mono319.910±15.00446.005±22.970.209±0.0130.129±0.0190.669±0.0160.892±0.0110.963±0.006
+ AugUndo308.010±0.859426.626±0.4840.200±0.0030.114±0.0010.674±0.0050.901±0.0020.969±0.002
", + "bbox": [ + 217, + 186, + 781, + 281 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the hypothesis that AugUndo can be applied generically to improve monocular depth estimation. Specifically, we observe a boost in the most difficult accuracy metric $(\\delta < 1.25)$ , where Monodepth2 improves from 0.717 to 0.724, HR-Depth from 0.714 to 0.718 and Lite-Mono from 0.669 to 0.674. At the same time, for AbsRel metric, Monodepth2 improves from 0.183 to 0.178, HR-depth improves from 0.185 to 0.181, and Lite-Mono improves from 0.209 to 0.200.", + "bbox": [ + 212, + 319, + 784, + 409 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sensitivity study. One limitation of existing depth completion training pipeline is that there are little to no augmentations applied to sparse depth modality. However, in real-world applications, sparse depth has, in fact, high variability, i.e., features tracked in SLAM/VIO systems will vary depending on the scene (points are dropped or added to the state), and point cloud densities returned by a sensor will differ based on specifications. To further examine the effect of AugUndo on sparse depth, we study the sensitivity to changes in sparse depth density by testing models on VOID1500 ( $\\approx$ 1500 points) on VOID500 ( $\\approx$ 500). For the $3 \\times$ reduction in sparse points, AugUndo improves robustness by an average of $23.1\\%$ across all methods. These improvements result from the geometric and occlusion augmentations made possible via AugUndo, which greatly increases sparse depth variations, i.e., decimating them through resizing, reorienting their configuration through rotation, translating them out of frame, and randomly occluding them, to avoid overfitting particular sparse depth configurations. We further push their limits in Sec. H of the Supp. Mat., where we test them on VOID150 with a $10 \\times$ reduction in sparse points from the training set (VOID1500) and observe the same trend of improvements. This shows that AugUndo significantly improves robustness of models to changes in sparse depth.", + "bbox": [ + 212, + 412, + 787, + 685 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zero-shot generalization. We test depth completion models trained on VOID on NYUv2 and ScanNet. Tab. 3 shows an average of $23.2\\%$ improvement on NYUv2 and $27.6\\%$ on ScanNet. Applying AugUndo to VOICED greatly improve its generalizability to both NYUv2 and ScanNet. This is likely due to the scaffolding densification employed by VOICED, where the network can overfit to foldings of particular sparse depth configurations and therefore, does not generalize well when presented with sparse depth maps with different configurations. Like our sensitivity study (Fig. 2, VOID500), occlusion and geometric augmentations introduce variation into the sparse depth configurations and densities, which alleviates overfitting to specific point clouds or 3D scenes; hence,", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation", + "bbox": [ + 235, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/918fd9242c5bf27726d4ec0939ba28293eb7aa4d14fe4697edf36847e9a1b6a8.jpg", + "table_caption": [ + "Table 3: Zero-shot transfer from VOID to NYUv2 and ScanNet for depth completion. AugUndo improves generalization of models trained on VOID to novel datasets by an average of $25.4\\%$ for all evaluation metrics (in millimeters) across both datasets." + ], + "table_footnote": [], + "table_body": "
MethodNYUv2ScanNet
MAE ↓RMSE ↓iMAE ↓iRMSE ↓MAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED2240±143.902427±143.49211±9.60238±10.891562±136.791764±146.39270±17.25311±17.36
+ AugUndo990±82.481181±65.55110±6.11132±6.01638±44.74791±79.48131±7.68170±6.32
FusionNet132.24±2.12236.16±4.5928.68±0.4261.87±1.20109.47±3.01206.33±6.1155.45±1.56122.52±2.04
+ AugUndo124.93±3.05227.23±4.9625.70±0.4154.09±0.87100.64±2.31195.85±5.8545.98±0.7899.90±5.00
KBNet138.31±5.74257.99±10.3625.48±0.6351.77±0.99103.05±4.99217.12±13.3536.23±1.1276.55±2.90
+ AugUndo118.60±3.44231.13±8.8522.06±0.3147.07±0.7082.53±4.33175.30±11.1329.87±1.0663.78±1.30
", + "bbox": [ + 222, + 200, + 781, + 292 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6566b079c467874bbcf71b39e9d696ef012be4ef4b67463d877a83a3107f1b01.jpg", + "table_caption": [ + "Table 4: Zero-shot transfer from VOID to NYUv2 and ScanNet for depth estimation. AugUndo consistently improves generalization (in meters) for monocular depth models." + ], + "table_footnote": [], + "table_body": "
DatasetMethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.25² ↑δ < 1.25³ ↑
NYUv2Monodepth20.432±0.0030.556±0.0050.205±0.0010.159±0.0010.683±0.0050.907±0.0010.975±0.001
+ AugUndo0.415±0.0050.537±0.0060.196±0.0020.148±0.0030.700±0.0080.915±0.0020.977±0.001
HR-Depth0.424±0.0030.549±0.0030.201±0.0020.154±0.0020.692±0.0030.910±0.0020.976±0.001
+ AugUndo0.421±0.0090.542±0.0110.199±0.0050.152±0.0060.696±0.0090.913±0.0040.976±0.001
Lite-Mono0.480±0.0120.616±0.0180.231±0.0080.199±0.0150.637±0.0090.879±0.0090.964±0.004
+ AugUndo0.468±0.0030.595±0.0030.225±0.0040.187±0.0050.646±0.0020.889±0.0030.968±0.001
ScanNetMonodepth20.284±0.0040.368±0.0050.177±0.0020.097±0.0020.741±0.0060.931±0.0030.980±0.001
+ AugUndo0.270±0.0030.351±0.0040.169±0.0010.088±0.0020.759±0.0040.937±0.0010.982±0.001
HR-Depth0.282±0.0040.366±0.0050.175±0.0020.095±0.0020.743±0.0050.929±0.0030.979±0.002
+ AugUndo0.274±0.0070.357±0.0090.172±0.0050.092±0.0050.754±0.0090.935±0.0030.981±0.001
Lite-Mono0.296±0.0020.388±0.0070.185±0.0030.109±0.0060.731±0.0020.921±0.0020.976±0.001
+ AugUndo0.296±0.0010.382±0.0010.185±0.0010.105±0.0020.728±0.0020.924±0.0010.977±0.001
", + "bbox": [ + 228, + 356, + 769, + 498 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "improving VOICED from $49.9\\%$ and $52.6\\%$ on NYUv2 and ScanNet, respectively. For FusionNet and KBNet, we still see nontrivial improvement: FusionNet improves by $8\\%$ and $12.1\\%$ on NYUv2 and ScanNet, and KBNet by $11.7\\%$ and $18.3\\%$ . This further validates our hypothesis that by applying a more diverse set of transformations, we are able to improve generalization to new datasets.", + "bbox": [ + 212, + 532, + 782, + 608 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For depth estimation, we tested Monodepth2, HR-Depth, Lite-Mono (trained on VOID) for zero-shot generalizability to NYUv2 and ScanNet. The results are shown in Tab. 4, where training with AugUndo consistently improves generalization errors across all evaluation metrics for both NYUv2 and ScanNet. Notably, for RMSE metric, Monodepth2 improves from 0.556 to 0.537, HR-Depth from 0.549 to 0.542, Lite-Mono from 0.616 to 0.595. For ScanNet, similar improvement on RMSE can also be observed, where Monodepth2 improves from 0.368 to 0.351, HR-Depth from 0.366 to 0.357, and Lite-Mono from 0.388 to 0.382. The improvement in RMSE metric, which is sensitive to outliers, highlights AugUndo's ability to model different input data distributions, collected by a different camera of different orientation with different object colors and layouts.", + "bbox": [ + 212, + 611, + 784, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on KITTI. We begin with quantitative results for depth completion. While AugUndo consistently improves all methods (Tab. 5), we note that the improvement is less, but respectable, in this case: $\\approx 5.2\\%$ overall with the largest gain in FusionNet of $6.41\\%$ . This is largely due to the small scene varia", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/34ee28e45e1e111a811318a35d542e375a3c2d75a1fa4c1d2dbea29ebbdea5b2.jpg", + "table_caption": [ + "Table 5: Depth completion on KITTI. AugUndo consistently improves all methods for all metrics (in millimeters) by approximately $5.2\\%$ on average." + ], + "table_footnote": [], + "table_body": "
MethodMAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED318.59±7.741,213.60±17.491.30±0.053.72±0.04
+ AugUndo295.41±0.301,159.27±5.441.20±0.023.49±0.03
FusionNet285.55±2.161,174.47±10.671.20±0.033.45±0.08
+ AugUndo267.69±1.851,157.07±4.611.08±0.023.19±0.03
KBNet263.90±3.631,130.66±6.221.05±0.013.24±0.04
+ AugUndo256.37±1.001,114.53±3.791.01±0.013.13±0.03
", + "bbox": [ + 346, + 186, + 651, + 276 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b7613559793937b55ff355bff53e46aa37439847fb190cf5e49589c5b6fab1a6.jpg", + "image_caption": [ + "Fig. 3: Depth completion on KITTI. We compare KBNet and FusionNet trained with standard augmentations and with AugUndo. AugUndo consistently produces lower errors in highlighted regions where structures may have arbitrary orientation (vegetation) and regions near image border that typically lacks correspondence during training." + ], + "image_footnote": [], + "bbox": [ + 245, + 295, + 759, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "tions in the outdoor driving scenarios, i.e., ground plane with vehicles, buildings on the sides, horizontal lidar scans, and largely planar motion. The dataset bias is strong enough to render vertical flip detrimental to performance. This is also evident in existing works as none utilizes vertical flip as augmentation. Nonetheless, AugUndo still improves performance, where point removal models different lidar returns patterns, and resizing simulates large variations in scales of objects observed in road scenes. Similar to indoors, translations help model occlusions by shifting the projection with different principal points. Fig. 3 compares FusionNet and KBNet trained using conventional augmentations and AugUndo. AugUndo improves over objects that may have diverse orientations (i.e., tree branches and vegetation), thanks to rotation augmentations. Improvements are also observed in \"small\" objects at far regions where resizing can simulate zooming in/out to emulate objects of different sizes. Through translation and occlusion, AugUndo also improves on occlusion boundaries during training (highlighted), making estimates near image borders and object boundaries more robust.", + "bbox": [ + 212, + 489, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We additionally show results for monocular depth estimation on KITTI in Tab. 6, where we compare the standard augmentation pipelines used by [15,35, 81] and AugUndo. We observe similar trends in performance gain as we did in depth completion trained on KITTI: Applying our set of augmentations improves most metrics for all methods. We observe notable improvements in $\\delta < 1.25$ , the most difficult accuracy metric, from 0.869 to 0.879 in Monodepth2, from 0.879 to 0.883 in HR-Depth, and 0.862 to 0.863 in Lite-Mono. For Monodepth2, we also observe a large reduction in AbsRel error, improving it from 0.117 to 0.113.", + "bbox": [ + 212, + 719, + 789, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation", + "bbox": [ + 233, + 113, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4cb5693e85de13c761bae8d356bc13f091333e5f9c3f029456ced9d1b2d0115f.jpg", + "table_caption": [ + "Table 6: Monocular depth estimation on KITTI. AugUndo consistently improve on evaluation metrics (in meters) across different models." + ], + "table_footnote": [], + "table_body": "
MethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.252↑δ < 1.253↑
Monodepth2 + AugUndo2.315±0.0054.794±0.0350.117±0.0010.845±0.0300.869±0.0040.959±0.0010.982±0.001
2.237±0.0144.739±0.0320.113±0.0000.862±0.0300.879±0.0020.960±0.0010.982±0.001
HR-Depth +AugUndo2.226±0.0044.626±0.0320.113±0.0010.797±0.0220.879±0.0020.961±0.0010.982±0.000
2.185±0.0134.610±0.0290.111±0.0010.794±0.0210.883±0.0010.962±0.0010.983±0.001
Lite-Mono + AugUndo2.338±0.0054.821±0.0270.121±0.0010.875±0.0070.862±0.0020.955±0.0010.980±0.000
2.314±0.0304.780±0.0550.120±0.0010.849±0.0190.863±0.0040.955±0.0010.981±0.001
", + "bbox": [ + 240, + 186, + 756, + 275 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Similarly, for Lite-Mono, we also boosted performance across all metrics, with a particularly high reduction in SqRel from 0.875 to 0.849.", + "bbox": [ + 212, + 305, + 782, + 334 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Remarks. As AugUndo only augments the inputs and modifies the loss computations, which add negligible time to training, we note that the performance improvements from it are obtained nearly for free. Yet, the percentage gain, however, is similar to improvements obtained by each successive state-of-the-art: For example, the improvement of Lite-Mono over HR-Depth, and that of HR-Depth over PackNet for monocular depth estimation; similarly, the improvement of FusionNet over VOICED and KBNet over FusionNet for depth completion.", + "bbox": [ + 212, + 335, + 784, + 441 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 462, + 354, + 478 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conventionally, data augmentation aims to seek visual invariance and create a collection of equivalent classes, i.e., identifying an image and its augmented variant as the same. For geometric tasks, the underlying equivalence is in the 3D scene structures under various illumination conditions, camera, viewpoints, occlusion, etc. Assuming a rigid scene, the shapes populating it should persist regardless of the nuisance variables. This motivates the use of geometric augmentations, as it simulates the nuisances within the image. However, adoption of geometric augmentations in unsupervised geometric tasks are obstructed by artifacts introduced during transformations (see Sec. D of Supp. Mat.). AugUndo lifts this obstacle by \"undo-ing\" the augmentations before computing the loss.", + "bbox": [ + 212, + 492, + 784, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While AugUndo enables scaling up augmentations for unsupervised training, i.e., depth completion, it may also be applicable for supervised methods; though, we posit that the gain to be less as the artifacts induced from photometric and geometric augmentations of an image tend to be larger than those of a piecewise smooth depth map. AugUndo is also limited to 2D augmentations; whereas, nuisances modeled by it are projections of the 3D scene (see limitations in Sec. M in Supp. Mat.). We leave extensions to 3D for future work. We also only consider a single image and sparse depth map as input. Likewise, extensions can be made towards multi-frame tasks such as stereo, optical flow, pose estimation, etc., but one must account for their specifics and problem setups, i.e., stereo assumes frontoparallel views. This is outside of our scope, so we leave them as future directions. This work paves way for the empirical success in unsupervised geometric tasks that we have observed in other visual recognition tasks.", + "bbox": [ + 212, + 643, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by was supported by NSF 2112562 Athena AI Institute.", + "bbox": [ + 215, + 146, + 787, + 176 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 203, + 325, + 219 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7063-7072 (2019)", + "2. Cheng, X., Wang, P., Guan, C., Yang, R.: Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10615-10622 (2020)", + "3. Cheng, X., Wang, P., Yang, R.: Depth estimation via affinity learned with convolutional spatial propagation network. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 103-119 (2018)", + "4. Choi, H., Lee, H., Kim, S., Kim, S., Kim, S., Sohn, K., Min, D.: Adaptive confidence thresholding for monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12808-12818 (2021)", + "5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)", + "6. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE international conference on computer vision. pp. 2650-2658 (2015)", + "7. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014)", + "8. Eldesokey, A., Felsberg, M., Holmquist, K., Persson, M.: Uncertainty-aware cnns for depth completion: Uncertainty from beginning to end. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12014-12023 (2020)", + "9. Ezhov, V., Park, H., Zhang, Z., Upadhyay, R., Zhang, H., Chandrappa, C.C., Kadambi, A., Ba, Y., Dorsey, J., Wong, A.: All-day depth completion. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE (2024)", + "10. Fei, X., Wong, A., Soatto, S.: Geo-supervised visual depth prediction. IEEE Robotics and Automation Letters 4(2), 1661-1668 (2019)", + "1. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)", + "12. Garg, R., Bg, V.K., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 740-756. Springer (2016)", + "13. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012)" + ], + "bbox": [ + 218, + 237, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation", + "bbox": [ + 235, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 270-279 (2017)", + "15. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3828-3838 (2019)", + "16. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2485-2494 (2020)", + "17. Guizilini, V., Hou, R., Li, J., Ambrus, R., Gaidon, A.: Semantically-guided representation learning for self-supervised monocular depth. In: International Conference on Learning Representations (2019)", + "18. Hu, M., Wang, S., Li, B., Ning, S., Fan, L., Gong, X.: Penet: Towards precise and efficient image guided depth completion. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13656-13662. IEEE (2021)", + "19. Jaritz, M., De Charette, R., Wirbel, E., Perrotton, X., Nashashibi, F.: Sparse and dense data with cnns: Depth completion and semantic segmentation. In: 2018 International Conference on 3D Vision (3DV). pp. 52-60. IEEE (2018)", + "20. Jeon, J., Lim, H., Seo, D.U., Myung, H.: Struct-mdc: Mesh-refined unsupervised depth completion leveraging structural regularities from visual slam. IEEE Robotics and Automation Letters 7(3), 6391-6398 (2022)", + "21. Kam, J., Kim, J., Kim, S., Park, J., Lee, S.: Costdcnet: Cost volume based depth completion for a single rgb-d image. In: European Conference on Computer Vision. pp. 257-274. Springer (2022)", + "22. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 582-600. Springer (2020)", + "23. Krishna, S., Vandrotti, B.S.: Deepsmooth: Efficient and smooth depth completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3357-3366 (2023)", + "24. Kumar, V.R., Klingner, M., Yogamani, S., Milz, S., Fingscheidt, T., Mader, P.: Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 61-71 (2021)", + "25. Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., Navab, N.: Deeper depth prediction with fully convolutional residual networks. In: 2016 Fourth international conference on 3D vision (3DV). pp. 239-248. IEEE (2016)", + "26. Lao, D., Wu, Y., Liu, T.Y., Wong, A., Soatto, S.: Sub-token vit embedding via stochastic resonance transformers. In: International Conference on Machine Learning. PMLR (2024)", + "27. Lao, D., Yang, F., Wang, D., Park, H., Lu, S., Wong, A., Soatto, S.: On the viability of monocular depth pre-training for semantic segmentation. In: European Conference on Computer Vision. Springer (2024)", + "28. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision 81(2), 155 (2009)", + "29. Li, A., Yuan, Z., Ling, Y., Chi, W., Zhang, C., et al.: A multi-scale guided cascade hourglass network for depth completion. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 32-40 (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Li, B., Shen, C., Dai, Y., Van Den Hengel, A., He, M.: Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1119-1127 (2015)", + "31. Lin, Y., Cheng, T., Zhong, Q., Zhou, W., Yang, H.: Dynamic spatial propagation network for depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1638-1646 (2022)", + "32. Liu, F., Shen, C., Lin, G., Reid, I.: Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence 38(10), 2024-2039 (2015)", + "33. Liu, T.Y., Agrawal, P., Chen, A., Hong, B.W., Wong, A.: Monitored distillation for positive congruent depth completion. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II. pp. 35-53. Springer (2022)", + "34. Lopez-Rodriguez, A., Busam, B., Mikolajczyk, K.: Project to adapt: Domain adaptation for depth completion from noisy and sparse sensor data. In: Proceedings of the Asian Conference on Computer Vision (2020)", + "35. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hr-depth: High resolution self-supervised monocular depth estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2294-2301 (2021)", + "36. Ma, F., Cavalheiro, G.V., Karaman, S.: Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In: International Conference on Robotics and Automation (ICRA). pp. 3288-3295. IEEE (2019)", + "37. Mahjourian, R., Wicke, M., Angelova, A.: Unsupervised learning of depth and egomotion from monocular video using 3d geometric constraints. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5667-5675 (2018)", + "38. Park, H., Gupta, A., Wong, A.: Test-time adaptation for depth completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20519-20529 (2024)", + "39. Park, J., Joo, K., Hu, Z., Liu, C.K., Kweon, I.S.: Non-local spatial propagation network for depth completion. In: European Conference on Computer Vision, ECCV 2020. European Conference on Computer Vision (2020)", + "40. Perez, L., Wang, J.: The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.04621 (2017)", + "41. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3227-3237 (2020)", + "42. Poggi, M., Tosi, F., Mattoccia, S.: Learning monocular depth estimation with unsupervised trinocular assumptions. In: 2018 International conference on 3d vision (3DV). pp. 324-333. IEEE (2018)", + "43. Qiu, J., Cui, Z., Zhang, Y., Zhang, X., Liu, S., Zeng, B., Pollefeys, M.: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3313-3322 (2019)", + "44. Qu, C., Liu, W., Taylor, C.J.: Bayesian deep basis fitting for depth completion with uncertainty. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16147-16157 (2021)", + "45. Qu, C., Nguyen, T., Taylor, C.: Depth completion via deep basis fitting. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 71-80 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation", + "bbox": [ + 235, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)", + "47. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020)", + "48. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12240-12249 (2019)", + "49. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. IEEE transactions on pattern analysis and machine intelligence 31(5), 824-840 (2008)", + "50. Shivakumar, S.S., Nguyen, T., Miller, I.D., Chen, S.W., Kumar, V., Taylor, C.J.: Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 13-20. IEEE (2019)", + "51. Shorten, C., Khoshgoftaar, T.M.: A survey on image data augmentation for deep learning. Journal of big data 6(1), 1-48 (2019)", + "52. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: European conference on computer vision. pp. 746-760. Springer (2012)", + "53. Singh, A.D., Ba, Y., Sarker, A., Zhang, H., Kadambi, A., Soatto, S., Srivastava, M., Wong, A.: Depth estimation from camera image and mmwave radar point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9275-9285 (2023)", + "54. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)", + "55. Taylor, L., Nitschke, G.: Improving deep learning with generic data augmentation. In: 2018 IEEE symposium series on computational intelligence (SSCI). pp. 1542-1547. IEEE (2018)", + "56. Tosi, F., Aleotti, F., Poggi, M., Mattoccia, S.: Learning monocular depth estimation infusing traditional stereo knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9799-9809 (2019)", + "57. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 2017 international conference on 3D Vision (3DV). pp. 11-20. IEEE (2017)", + "58. Upadhyay, R., Zhang, H., Ba, Y., Yang, E., Gella, B., Jiang, S., Wong, A., Kadambi, A.: Enhancing diffusion models with 3d perspective geometry constraints. ACM Transactions on Graphics (TOG) 42(6), 1-15 (2023)", + "59. Van Gansbeke, W., Neven, D., De Brabandere, B., Van Gool, L.: Sparse and noisy lidar completion with rgb guidance and uncertainty. In: 2019 16th International Conference on Machine Vision Applications (MVA). pp. 1-6. IEEE (2019)", + "60. Wang, C., Buenapasada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2022-2030 (2018)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2162-2171 (2019)", + "62. Wong, A., Cicek, S., Soatto, S.: Targeted adversarial perturbations for monocular depth prediction. Advances in neural information processing systems 33, 8486-8497 (2020)", + "63. Wong, A., Cicek, S., Soatto, S.: Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 1495-1502 (2021)", + "64. Wong, A., Fei, X., Hong, B.W., Soatto, S.: An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 3120-3127 (2021)", + "65. Wong, A., Fei, X., Tsuei, S., Soatto, S.: Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters 5(2), 1899-1906 (2020)", + "66. Wong, A., Soatto, S.: Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5644-5653 (2019)", + "67. Wong, A., Soatto, S.: Unsupervised depth completion with calibrated backprojection layers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12747-12756 (2021)", + "68. Xu, D., Ricci, E., Ouyang, W., Wang, X., Sebe, N.: Multi-scale continuous crfs as sequential deep networks for monocular depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5354-5362 (2017)", + "69. Xu, Y., Zhu, X., Shi, J., Zhang, G., Bao, H., Li, H.: Depth completion from sparse lidar data with depth-normal constraints. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2811-2820 (2019)", + "70. Yan, Z., Wang, K., Li, X., Zhang, Z., Li, J., Yang, J.: Desnet: Decomposed scale-consistent network for unsupervised depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3109-3117 (2023)", + "71. Yang, F., Feng, C., Chen, Z., Park, H., Wang, D., Dou, Y., Zeng, Z., Chen, X., Gangopadhyay, R., Owens, A., et al.: Binding touch to everything: Learning unified multimodal tactile representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26340-26353 (2024)", + "72. Yang, N., Stumberg, L.v., Wang, R., Cremers, D.: D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1281-1292 (2020)", + "73. Yang, Y., Wong, A., Soatto, S.: Dense depth posterior (ddp) from single image and sparse range. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3353-3362 (2019)", + "74. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: Every pixel counts: Unsupervised geometry learning with holistic 3d motion understanding. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)", + "75. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: Lego: Learning edge with geometry all at once by watching videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 225-234 (2018)", + "76. Yin, W., Liu, Y., Shen, C., Yan, Y.: Enforcing geometric constraints of virtual normal for depth prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5684-5693 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation", + "bbox": [ + 235, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "77. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1983-1992 (2018)", + "78. Yu, Z., Sheng, Z., Zhou, Z., Luo, L., Cao, S.Y., Gu, H., Zhang, H., Shen, H.L.: Aggregating feature point cloud for depth completion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8732-8743 (2023)", + "79. Zeng, Z., Wang, D., Yang, F., Park, H., Soatto, S., Lao, D., Wong, A.: Wordepth: Variational language prior for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9708-9719 (2024)", + "80. Zhan, H., Garg, R., Weerasekera, C.S., Li, K., Agarwal, H., Reid, I.: Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 340-349 (2018)", + "81. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18537-18546 (2023)", + "82. Zhang, Y., Funkhouser, T.: Deep depth completion of a single rgb-d image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 175-185 (2018)", + "83. Zhang, Y., Guo, X., Poggi, M., Zhu, Z., Huang, G., Mattoccia, S.: Completionformer: Depth completion with convolutions and vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18527-18536 (2023)", + "84. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 2022 International Conference on 3D Vision (3DV). pp. 668-678. IEEE (2022)", + "85. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1851-1858 (2017)", + "86. Zou, Y., Luo, Z., Huang, J.B.: Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In: Proceedings of the European conference on computer vision (ECCV). pp. 36-53 (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 632 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Y. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_model.json b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_model.json new file mode 100644 index 0000000000000000000000000000000000000000..804af75e211badd9eb97c233227ea49e81311a04 --- /dev/null +++ b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_model.json @@ -0,0 +1,2462 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.765, + 0.187 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Monocular Depth Completion and Estimation" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.212, + 0.7, + 0.244 + ], + "angle": 0, + "content": "Yangchao Wu\\(^{1}\\), Tian Yu Liu\\(^{1}\\), Hyoungseob Park\\(^{2}\\), Stefano Soatto\\(^{1}\\), Dong Lao\\(^{1}\\), and Alex Wong\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.254, + 0.692, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) UCLA Vision Lab, Los Angeles, CA 90095, USA {wuyangchao1997, tianyu, soatto, lao}@cs.ucla.edu" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.283, + 0.662, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) Yale Vision Lab, New Haven, CT 06511, USA {hyoungseob.park, alex.wong}@yale.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.346, + 0.744, + 0.653 + ], + "angle": 0, + "content": "Abstract. Unsupervised depth completion and estimation methods are trained by minimizing reconstruction error. Block artifacts from resampling, intensity saturation, and occlusions are amongst the many undesirable by-products of common data augmentation schemes that affect image reconstruction quality, and thus the training signal. Hence, typical augmentations on images viewed as essential to training pipelines in other vision tasks have seen limited use beyond small image intensity changes and flipping. The sparse depth modality in depth completion have seen even less use as intensity transformations alter the scale of the 3D scene, and geometric transformations may decimate the sparse points during resampling. We propose a method that unlocks a wide range of previously-infeasible geometric augmentations for unsupervised depth completion and estimation. This is achieved by reversing, or \"undo\"-ing, geometric transformations to the coordinates of the output depth, warping the depth map back to the original reference frame. This enables computing the reconstruction losses using the original images and sparse depth maps, eliminating the pitfalls of naive loss computation on the augmented inputs and allowing us to scale up augmentations to boost performance. We demonstrate our method on indoor (VOID) and outdoor (KITTI) datasets, where we consistently improve upon recent methods across both datasets as well as generalization to four other datasets. Code available at: https://github.com/alexklwong/augundo" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.665, + 0.707, + 0.679 + ], + "angle": 0, + "content": "Keywords: depth completion \\(\\cdot\\) monocular depth \\(\\cdot\\) augmentations" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.704, + 0.377, + 0.72 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Data augmentation is essential to training machine learning models; it plays a role in performance and generalization [40, 51, 55]. One common axiom of choosing augmentations is that the task output should remain invariant to the augmentation. For example, image flipping is a viable augmentation for classifying animals, since it does not alter the label. Conversely, flipping road signs can alter their meanings; hence, such augmentation can be detrimental to tasks involving road sign recognition. For geometric tasks, the range of augmentations" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.42 + ], + "angle": 0, + "content": "is more restricted due to constraints of problem formulation: Stereo assumes pairs of frontoparallel rectified images; hence, in-plane rotations are not viable. Image-guided sparse depth completion relies on sparse points to ground estimates to metric scale; therefore, intensity transformations on sparse depth maps that alter the scale of the 3-dimensional (3D) scene are infeasible. Unsupervised learning of depth completion and estimation further limit the use of augmentations as the supervision signal comes from reconstructing the inputs, where augmenting the input introduces artifacts that impact reconstruction quality and therefore the supervision (see Sec. D, Fig. 1, Tab. 8, 9 in the Supp. Mat. for examples of artifacts and extended discussion on their effect on learning). Moreover, video-based unsupervised training assumes rigid motion, so augmentations that introduce padding (e.g., translation, rotation) will yield constant or edge extended borders across images (i.e., no motion), preventing the model from properly learning depth and pose - leaving few augmentations viable. While simulating nuisances is desirable, naively applying augmentations may do more harm than good; thus, it is unsurprising that existing work in unsupervised depth completion [33, 36, 50, 63-65, 67, 73] and estimation [14, 15, 35, 81, 85] primarily rely on a small range of photometric augmentations and flipping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Nevertheless, photometric augmentations help model the diverse range of illumination conditions and colors of objects that may populate the scene; geometric augmentations can simulate the various camera parameters, i.e., image resizing (zooming) can model changes in focal length, and scene arrangements, i.e., image flipping. However, block artifacts, loss during resampling, and intensity saturation are just some of the many undesirable side-effects of traditional augmentations to the image and sparse depth map for unsupervised learning of geometric tasks. To avoid compromising the supervision signal, we compute the typical reconstruction loss on the original input image and sparse depth map instead of the augmented inputs, which bypasses negative effects of reconstruction artifacts due to photometric and geometric augmentations. However, there exists a mis-alignment between the original input (e.g., image, sparse depth), and the model depth estimate as geometric augmentations induce a change in coordinates. Hence, we undo the geometric augmentations by inverting them in the output space to align the model estimate with the training target." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Amongst the many geometric tasks, we focus on unsupervised depth completion, the task of inferring a dense depth map from an image and sparse depth map, where augmentations have seen limited use. Here, a training sample includes the input sparse depth map, its associated image and additional images of neighboring views of the same 3D scene. Our method is also applicable to unsupervised monocular depth estimation, which omits the sparse depth modality. Augmentations have traditionally been restricted to a limited range of photometric transformations and flipping - due to the need to preserve photometric consistency across a sequence of video frames used during training, and the sparse set of 3D points projected onto the image frame as a 2.5D range map; degradation to either modalities directly impacts the supervision signal as the loss function is conventionally computed on the augmented inputs. By using our" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.145, + 0.703, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.316, + 0.784, + 0.386 + ], + "angle": 0, + "content": "Fig. 1: Overview. We apply photometric augmentations to the input image, and the same set of geometric augmentations to both the input image and sparse depth map. We warp the output depth back to the original reference frame with the inverse geometric transformations. This enables image and sparse depth reconstruction losses to be computed on the original inputs while unlocking previously-infeasible augmentations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.784, + 0.492 + ], + "angle": 0, + "content": "method, loss functions involving sparse depth and image reconstruction from other views can be computed on the original inputs while applying augmentations that were previously not viable for the task. Our hypothesis: By \"undoing\" the augmentations, one can expand the viable set and scale up their use in training, leading to improved model performance and generalizability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.785, + 0.674 + ], + "angle": 0, + "content": "To this end, we introduce AugUndo, an augmentation framework that enables one to apply a wide range of photometric and geometric transformations on the inputs, and to \"undo\" them during loss computation. This allows computing the unsupervised loss on the original images and sparse depth maps, free of artifacts, through a warping of the output depth map - obtained from augmented input onto the input frame of reference based on the inverse geometric transformation. In addition to group transformations that allow for output alignment, we combine them with commonly employed photometric augmentations. To the best of our knowledge, we are the first to propose a unified augmentation scheme for photometric and geometric augmentations for unsupervised depth completion and estimation. We demonstrate AugUndo on recent methods on indoor and outdoor settings, where we consistently improve all methods across all datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Our contributions are as follows: (1) We propose AugUndo, a simple-yet-effective framework to scale up photometric and geometric augmentations for unsupervised depth completion and estimation, without compromising the supervision signal; AugUndo can be applied in a plug-and-play manner to existing methods with negligible increase in computational costs during training. (2) We enable previously-infeasible augmentations to be used for training unsupervised methods and comprehensively ablate combinations of eleven types of augmentations to study the performance benefits of each. (3) We show that AugUndo can consistently improve robustness to shifts in sparse point densities for completion, model performance as well as zero-shot generalization for both depth completion and estimation for indoor and outdoor scenarios; thus, validating our hypothesis." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.478 + ], + "angle": 0, + "content": "Data augmentation for depth completion and estimation. While photometric transformations such as color jitter are often applied for unsupervised depth completion [36,63-65], and estimation [10,14-16,26,27,41,47,66,79,80], geometric augmentations other than simple flipping are less commonly adopted. Most supervised depth completion methods [8,18,19,21,23,29,31,43-45,69,73,82,83] similarly limit their augmentations to color jitter and horizontal flips due to sparse depth maps being decimated by rotation and scaling, causing points to be interpolated away. Nevertheless, for supervised training, it is straightforward to directly apply the same transformation to the ground truth for training. Indeed, some supervised depth completion methods [2,3,9,38,39,53,59,78] adopted random scaling, translation, in-plane rotation. Translation augmentations are also commonly applied in supervised depth estimation methods [6,7,25,30,32, 46,47,58,62,68,71,76]. We posit that the reason that such transformations can be applied to the ground truth in supervised settings is due to artifacts caused by the transformation of a piece-wise smooth depth map being less severe than those of an RGB image and its intensities. However, such assumptions do not hold in unsupervised training. These artifacts would affect the training signal for unsupervised methods, which rely on photometric correspondences and observed sparse points. Our approach aims to resolve this to allow diverse geometric augmentations to be applied in a plug-and-play fashion in unsupervised training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Unsupervised depth completion methods [36, 50, 63-65, 70, 73] leverage image and sparse depth reconstructions as training signals by minimizing errors between the input image and its reconstruction from other views, and errors between the input sparse depth map and the predicted depth along with a local smoothness regularizer. [36] used Perspective-n-Point [28] and RANSAC [11] to align consecutive video frames. [73] learns a depth prior conditioned on the image. [63] uses synthetic data to learn a prior on the shapes populating a scene, while [34] translates synthetic data to real domain to leverage rendered depth. [64] proposed an adaptive scheme to reduce penalties incurred on occluded regions. [67] maps the image onto the 3D scene through calibrated backprojection. [70] decouples structure and scale. [20] uses line features from visual SLAM and [33] introduced monitored distillation for positive congruent training. Augmentations for these methods are limited to a small range of photometric perturbations and image flipping. Operations such as rotation, resizing, and translation require resampling, which creates artifacts and affects the reconstruction quality. This causes performance degradation since loss is typically computed on the augmented images. Loss in sparse depth maps is further impacted as resampling and interpolation may cause loss of sparse points. Contrary to these limited augmentation schemes, we enable a large range of photometric and geometric augmentations to be introduced during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Unsupervised monocular depth estimation, like depth completion, also minimizes photometric reconstruction error. [12] frames depth estimation as a novel view synthesis problem. [14] improves [12] by imposing a consistency loss on the depth predicted from left and right images. [85] uses a pose network to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "enable unsupervised training on video sequences. [15] introduced auto-masking and min reprojection loss. To improve the supervision signal, [4, 56, 61] leverage noisy proxy labels and [42] uses a trinocular assumption. Additional loss terms based on visual odometry [60], iterative closest point [37], surface normals [75], and semantic segmentation [17, 24] were also introduced; [41, 72] further included predictive uncertainty. To handle rigid and non-rigid motions in the scene, previous works explored multi-task learning to include optical flow and moving object estimation [1, 48, 74, 77, 86], and used semantics to filter out outlier regions [22]. [35] redesigned the skip connection and decoders to extract high-resolution features, [84] combined global and local representations and [81] introduced a lightweight architecture with dilated convolution and attention. In addition to depth completion, our method also shows consistent improvements on monocular depth estimation for [15, 35, 81]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.366, + 0.456, + 0.383 + ], + "angle": 0, + "content": "3 Method Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.789, + 0.548 + ], + "angle": 0, + "content": "Let \\(I:\\varOmega\\subset\\mathbb{R}^2\\to\\mathbb{R}_+^3\\) be an RGB image captured by a calibrated camera, \\(z:\\varOmega_z\\subset\\varOmega\\rightarrow\\mathbb{R}_+\\) the corresponding sparse point cloud projected onto the image plane as a sparse depth map, and \\(K\\in \\mathbb{R}^{3\\times 3}\\) the intrinsic calibration matrix. Given an image and its sparse depth map, depth completion aims to learn a function \\(f_{\\theta}(I,z)\\) that recovers the distance between the camera to points in the 3D scene as a dense depth map. In another mode, if sparse depth maps are not given, then the problem reduces to monocular depth estimation which learns a function \\(f_{\\theta}(I)\\) to map a single image to a depth map \\(\\varOmega\\rightarrow\\mathbb{R}_{+}\\). For the ease of notation, we denote the output depth map for both depth completion and estimation as \\(\\hat{d}\\in \\mathbb{R}_{+}^{H\\times W}\\) where \\(H\\) and \\(W\\) are its height and width." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Unsupervised depth completion relies on photometric and sparse depth reconstruction errors as its primary supervision signal. Without loss of generality, we assume an input of \\((I_t,z_t)\\) for an RGB image and associated sparse depth map captured at time \\(t\\) and during training, an additional set of temporally adjacent images \\(I_{\\tau}\\) for \\(\\tau \\in \\Upsilon \\doteq \\{t - 1,t + 1\\}\\). The reconstruction \\(\\hat{I}_{t\\tau}\\) of \\(I_{t}\\) from image \\(I_{\\tau}\\) is given by the reprojection based on estimated depth \\(\\hat{d}_t\\coloneqq f_\\theta (\\cdot)\\)" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.653, + 0.785, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {t \\tau} (x) = I _ {\\tau} \\left(\\pi g _ {\\tau t} K ^ {- 1} \\bar {x} \\hat {d} _ {t} (x)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.725 + ], + "angle": 0, + "content": "where \\(\\bar{x} = [x^{\\top},1]^{\\top}\\) is the homogeneous coordinates of \\(x\\in \\Omega\\), \\(g_{\\tau t}\\in SE(3)\\) the relative pose of the camera from time \\(t\\) to \\(\\tau\\), \\(K\\) the intrinsic calibration matrix, and \\(\\pi\\) the canonical perspective projection." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.727, + 0.75, + 0.743 + ], + "angle": 0, + "content": "Using Eq. (1), a depth completion or estimation network \\( f_{\\theta} \\) minimizes" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.754, + 0.786, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\theta} \\sum_ {\\tau \\in \\Upsilon} \\sum_ {x \\in \\Omega} \\alpha \\rho (\\hat {I} _ {t \\tau} (x), I _ {t} (x)) + \\sum_ {x \\in \\Omega_ {z}} \\beta \\psi (\\hat {d} _ {t} (x), z _ {t} (x)) + \\lambda R (\\hat {d} _ {t}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\rho\\) denotes the photometric reconstruction error, typically \\(L_{1}\\) difference in pixel values and/or structural similarity (SSIM), \\(\\psi\\) the sparse depth reconstruction error, typically \\(L_{1}\\) or \\(L_{2}\\), \\(R\\) the regularizer that biases the depth map to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "be piece-wise smooth with depth discontinuities aligned with edges in the image (commonly used by previous works [15,36,65,67]), and \\(\\alpha\\), \\(\\beta\\) and \\(\\lambda\\) their respective weighting. Note: monocular depth omits sparse depth term, i.e., \\(\\beta = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.45 + ], + "angle": 0, + "content": "Inverse transformation. Let \\(\\mathcal{A}_{pt}\\) be the set of possible photometric transformations, and \\(\\mathcal{A}_{ge}\\) be the set of all geometric transformations. Given \\(T_{ge} \\in \\mathcal{A}_{ge}\\), we wish to obtain an inverse transformation \\(T_{ge}^{-1}\\) such that \\(T_{ge} \\circ T_{ge}^{-1} \\approx Id\\) the identity function3. At each time step, we can sample a sequence of transformations \\(\\{T_{pt}^{1} \\ldots T_{pt}^{k}\\}\\) and \\(\\{T_{ge}^{1} \\ldots T_{ge}^{m}\\}\\) respectively from \\(\\mathcal{A}_{pt}\\) and \\(\\mathcal{A}_{ge}\\) to construct transformations \\(T_{pt} = T_{pt}^{1} \\circ \\dots \\circ T_{pt}^{k}\\) and \\(T_{ge} = T_{ge}^{1} \\circ \\dots \\circ T_{ge}^{m}\\). We denote the composition of a collection of augmentation transformations by \\(T = T_{ge} \\circ T_{pt}\\) where \\(T_{pt}\\) denotes photometric transformation, and \\(T_{ge}\\) denotes geometric transformation. Furthermore, we denote the inverse geometric transformations by \\(T_{ge}^{-1} = (T_{ge}^{m})^{-1} \\circ (T_{ge}^{m-1})^{-1} \\circ \\dots \\circ (T_{ge}^{1})^{-1}\\), which operates on the space of depth maps to reverse the geometric transformation so that we can warp the output depth map onto the reference frame of the original image. Through inverse warping, this process is differentiable. In practice, some transformations cause border regions of the image to be warped out of frame, i.e., translated off the image plane or cropped away. Hence, after warping our output depth map back to the original frame of reference using the inverse geometric transformations, border extensions (edge paddings) are introduced to handle out-of-frame regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.784, + 0.481 + ], + "angle": 0, + "content": "AugUndo. We apply each geometric transformation over image coordinates and resample (bilinear for image and nearest neighbor for sparse depth):" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.491, + 0.785, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\left[ x ^ {\\prime} 1 \\right] ^ {\\top} = T _ {g e} \\left[ x 1 \\right] ^ {\\top} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.513, + 0.785, + 0.53 + ], + "angle": 0, + "content": "\\[\nI ^ {\\prime} \\left(x ^ {\\prime}\\right) = T _ {p t} (I) (x); z ^ {\\prime} \\left(x ^ {\\prime}\\right) = z (x) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.539, + 0.784, + 0.663 + ], + "angle": 0, + "content": "where \\( T_{ge} \\) is the geometric transformation, \\( x \\in \\Omega \\) and \\( x' \\in \\Omega \\) are coordinates in the image grid, and \\( I' \\) is the image after the transformation; for ease of notation, we hereafter denote \\( I' = T(I) = T_{ge} \\circ T_{pt}(I) \\) to include both augmentations through composition. Note that \\( x \\) is in the original image reference frame and \\( x' \\) is in the transformed image reference frame. Naturally, this process can be extended to multiple geometric transformations by composing them, i.e., \\( T_{ge} = T_{ge}^{1} \\circ T_{ge}^{2} \\circ \\dots \\circ T_{ge}^{m} \\). The reverse process is simply inverting the transformations where \\( T_{ge}^{-1} = (T_{ge}^{m})^{-1} \\circ (T_{ge}^{m-1})^{-1} \\circ \\dots \\circ (T_{ge}^{1})^{-1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.673, + 0.785, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\left[ x ^ {\\prime \\prime} 1 \\right] ^ {\\top} = T _ {g e} ^ {- 1} \\left[ x ^ {\\prime} 1 \\right] ^ {\\top} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.697, + 0.785, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\hat {d} \\left(x ^ {\\prime \\prime}\\right) = \\hat {d} ^ {\\prime} \\left(x ^ {\\prime}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.784, + 0.772 + ], + "angle": 0, + "content": "where \\(\\hat{d}'\\) is the depth map inferred from augmented inputs \\((I', z')\\). Once reverted back to the original reference frame, \\(\\hat{d}\\) is used to reconstruct \\(I_t\\) from \\(I_\\tau\\) for \\(\\tau \\in \\Upsilon\\) as in Eq. (1). More details can be found in Alg. 1 in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.784, + 0.804 + ], + "angle": 0, + "content": "By modeling \\( T_{ge}^{-1} \\), Eq. (3)-Eq. (6) allow us to apply a wide range of augmentations, while still establishing the correspondence between \\( I_{t} \\) and \\( I_{\\tau} \\). Specifically," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "3 Note that in practice, not all transformations are bijective. For instance, image rescaling with a fixed size image plane; hence strict equality is not always possible." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.316 + ], + "angle": 0, + "content": "for minimizing Eq. (2), one may simply augment the image and sparse depth by \\(T\\) to yield \\((I_t', z_t')\\) as input, and reconstruct the original image and sparse depth \\((I_t, z_t)\\) from other views \\(I_\\tau\\) and the aligned output depth \\(\\hat{d}\\) (Eq. (6)). We note that the inverse transformation is critical for enabling the sparse depth reconstruction term to be computed properly in Eq. (2); if computed in the transformed reference frame, i.e., on \\(z_t'\\), many of the sparse points would be decimated by interpolation during rotation and resizing (downsampling) augmentations - leaving a lack of supervision from the sparse depth term of the loss function. We note that the original image sequence is fed to the pose estimator, rather than the augmented images. This ensures the estimated relative pose used during reprojection (Eq. (1)) also corresponds to the original images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Modeling geometric augmentations. Depth completion differs from typical image-based (e.g., monocular) depth estimation, where operations like resizing are susceptible to scale ambiguity and can either correspond to changes in focal length, or distance from the camera (i.e., due to changes in camera pose). In depth completion, the sparse depth map grounds image pixels to specific depth values. Hence, there is no ambiguity in distance of each sparse point from the camera. Our goal is to synthesize new training data via geometric transformations that are consistent with the original 3D scene. Given that there is no scale ambiguity, this can be achieved by either (1) changes in camera pose (extrinsics), or (2) changes in camera parameters (intrinsics)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Augmentation via (1) would require warping the inputs for view synthesis. Since we operate under the unsupervised setting, accurate warping (preserving 3D scene) without access to dense ground truth depth values is difficult to achieve. An exception to this is rotation, which we address below. On the other hand, (2) can be achieved via standard 2D image transformations while preserving the 3D structure of the scene. This motivates our modeling of geometric transformations, i.e., resizing, translation, as changes in camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.716 + ], + "angle": 0, + "content": "For example, augmenting focal length (i.e. zooming in/out) is akin to \"resizing\". Similarly, translation can be used to model shifting of the optic center, i.e. capturing the same 3D scene using cameras with different principal point offsets. Note: image rotation can be seen as rotating a camera about its optical axis, which does not change the distance of a point from the camera. Thus, the camera position and the scene are kept constant across augmentations, eliminating the need for adjusting depth values via warping during the process. Instead, we only have to realign the output depth back to the original frame of reference (Eqs. (5) and (6)), which makes our method computationally efficient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Augmentations. Our choices are based on common nuisance variabilities, i.e., changes in illumination, occlusions, and object color, scene layout (flip), and camera parameters (resize, translation) and orientation (rotation)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Photometric. We include brightness, contrast, saturation, and hue, where all follow standard augmentation pipeline in existing works [15,65,67,81]. Applying the inverse of photometric augmentation can be viewed as recovering the original image; hence, we directly use original image instead of applying transformations to the intensities (which are not recoverable if saturated at the max value)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "Occlusion. We consider patch removal and sparse point removal. For the former, we randomly select a percentage of pixels \\( x \\in \\Omega \\) in the image and remove (with zero-fill) an arbitrary-sized patch around it. For the latter, we randomly sample a percentage of points \\( x \\in \\Omega_z \\) in the sparse depth map. The inverse transformation of this is simply the original image and sparse depth map, both used in loss computation before augmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Flip. We consider horizontal and vertical flips. When applied, the same flip operation is used for both input image and sparse depth maps. We record the flip type during data augmentation. During loss computation, we reverse the flip direction to align output depth with the original image and sparse depth map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.463 + ], + "angle": 0, + "content": "Resize. We define a new image plane of the input resolution and generate a random scaling factor to be applied along both height and width directions. The image is warped to the new image plane, where any point warped out of the plane is excluded; borders of the warped image are extended to the bounds of the image plane by edge replication. Sparse points occupying multiple pixels are eroded to a single point for resizing with a factor greater than 1. During loss computation, we warp the output depth map onto a new image plane of the same dimensions by the inverse scaling matrix of the recorded scaling factors; borders of the warped depth map are extended by edge replication if necessary. We view resizing factors greater and less than 1 (zooming in and out) as two separate forms of augmentations to distinguish their contributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.787, + 0.616 + ], + "angle": 0, + "content": "Rotation. Naive rotation leads to the loss of large areas of the image, i.e., cropped away to retain the same-size image, discarding large portions of possible co-visible regions and also supervisory signals. To preserve the entire image, we first warp the image by a randomly generated angle to a new (larger) image plane, so that the rotated image fits tightly within the new image. As image sizes within a batch can vary depending on the rotation angle, we center-pad (uniformly on each side) each image in the batch to the maximum width and height of the augmented batch. To reverse the rotation on the output depth, we warp the output depth map back by the inverse rotation matrix, then perform a center crop on the depth map to align with the original input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.787, + 0.737 + ], + "angle": 0, + "content": "Translation. We define a new image plane of the same dimensions as the input and generate random height and width translation factors. The coordinates of the input are translated and its pixels are inverse warped onto the new image plane. Any pixel warped out of the image plane is excluded. Borders of the warped image are extended to the bounds of the image plane by edge replication. During loss computation, we warp the output depth map onto a new image plane of the same dimensions by the inverse translation matrix. Borders of the warped depth map are extended to the bounds of the image plane by edge replication." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.762, + 0.376, + 0.779 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.795, + 0.785, + 0.842 + ], + "angle": 0, + "content": "We demonstrate AugUndo on recent unsupervised depth completion (VOICED [65], FusionNet [63], KBNet [67]) and estimation (Monodepth2 [15], HRDepth [35], LiteMono [81]) methods on two datasets (KITTI [13, 57], Void [65]) here." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Due to space limitations, we defer the results of MonDi [33] and DesNet [70] to Sec. K in the Supp. Mat. To evaluate zero-shot generalization, we test on three additional datasets for each task: NYUv2 [52], ScanNet [5], Waymo [54], and Make3d [49] - where we transfer models on Void to NYUv2 (Tab. 3) and ScanNet, and from KITTI to Waymo (Tab. 4 in Supp. Mat.) for completion and Make3d [49] (Tab. 5 in Supp. Mat.) for estimation. We also test model sensitivity to different sparse depth input densities in Tab. 1 (right) - see Sec. H of the Supp. Mat for an extensive study. Additionally, we present an comprehensive ablation study in Sec. G in the Supp. Mat. to test the effect of each augmentation. We also provide results on modeling AugUndo as changes in camera motion or parameters in Tab. 7 of the Supp. Mat.; interpreting geometric augmentations as different camera parameters yields better performance. Tables. 8, 9 in Supp. Mat. show that unsupervised models cannot be trained with naive geometric augmentations. See Sec. E of the Supp. Mat. for descriptions of datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.47 + ], + "angle": 0, + "content": "Experiment setup. We followed the original settings of the open-sourced code for each work and modified the data handling and loss function to incorporate AugUndo. We perform 4 independent trials for each experiment and report their means and standard deviations. To ensure a fair comparison, we train all models from scratch. Below, we report the best combination of augmentations found empirically through extensive experiments on each dataset. See Secs. A and B in Supp. Mat. for implementation details, and evaluation metrics, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Augmentations. Through a search over augmentation types and values, we found a consistent set that tends to yield improvements across all methods with small changes to degree of augmentation catered to each method. See Sec. C in Supp. Mat. for details of augmentation parameters. Rows in Tabs. 1 to 6 marked with \"+\" AugUndo\" denote models trained with our method. We note that performance gain can be obtained by typical set and ranges of augmentations and does not require a meticulous selection of hyper-parameters (see Sec. J of Supp. Mat. for a sensitivity study). Results reported here aims to study how far one can push performance and generalization, purely from augmentations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Results on VOID. We begin by presenting quantitative results for AugUndo on unsupervised depth completion. Tab. 1 (left, VOID1500) shows our main results on the VOID benchmark. By training the models with AugUndo, we observe an average overall improvement of \\(18.3\\%\\) across all methods and metrics on VOID1500. Specifically, we improve VOICED by \\(26.4\\%\\), FusionNet by \\(16.3\\%\\), and KBNet by \\(12.2\\%\\). These experiments validate our hypothesis that by applying a wider range of augmentations, we are able to improve the baseline model's performance. They also illustrate the lack in use of augmentations in existing works: incorporating standard augmentations (albeit requiring modification to augmentation and loss computation pipelines) can yield a large performance boost. Fig. 2 shows a head-to-head comparison between KBNet trained using standard procedure in [67] and KBNet trained using AugUndo. We observe qualitative improvements from AugUndo where we improve KBNet in homogeneous regions and image borders, i.e., pillar (left), cabinet (middle), wall (right). Applying geometric augmentations yields models with fewer border" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: Depth completion on Void. AugUndo improves performance by an average of \\(18.3\\%\\) across all methods and metrics on Void1500. When models trained on Void1500 are tested Void500, AugUndo improves by \\(23.1\\%\\), as translation, resizing, and occlusion augmentations vary sparsity by removing sparse points from the input." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.214, + 0.787, + 0.322 + ], + "angle": 0, + "content": "
MethodVOID1500VOID500
MAE ↓RMSE ↓iMAE ↓iRMSE ↓MAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED74.78±2.69139.75±4.5739.20±1.4671.98±2.54137.01±4.23235.80±7.8271.36±1.86130.63±5.66
+ AugUndo52.73±0.41111.09±0.9226.93±0.5454.46±0.3892.99±1.11176.94±1.3846.43±0.8591.10±1.64
FusionNet52.11±0.44113.30±1.1828.53±0.5258.79±2.0197.73±0.73194.32±1.3658.65±1.31122.95±3.04
+ AugUndo41.16±0.1899.21±0.3922.23±0.3553.07±1.3074.97±0.69162.71±2.3940.44±1.3992.11±5.79
KBNet38.11±0.7795.22±1.7219.51±0.1446.70±0.4878.44±1.39178.17±3.2737.56±0.6183.43±1.89
+ AugUndo33.32±0.1885.67±0.3916.61±0.2941.24±0.6066.97±0.81151.55±2.0331.63±0.5371.90±0.82
" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.34, + 0.773, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.509, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Fig. 2: Depth completion on VOID. We compare KBNet trained with standard augmentations and with AugUndo. AugUndo consistently produces lower errors with reduced border artifacts and improves on homogeneous regions, i.e., pillar (left), cabinet (middle), wall of staircase (right). Error maps are highlighted for comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.779 + ], + "angle": 0, + "content": "artifacts as we apply random translation, which warps part of the image out of frame. This allows us to simulate border occlusion, where there lacks correspondence in an adjacent frame. While training with standard protocol results in failures to recover structures near the image border, training with AugUndo can render models robust to them by computing the loss on the original frame of reference (where we do have correspondence) through our inverse transformations of the depth map. Additionally, resizing allows the model to learn multiple resolutions of the input, akin to zooming in and out, which can also impose smoothness in homogeneous regions through the scale-space transition; whereas, rotation can simulate diverse camera orientations. Fig. 2 shows that translation, rotation, and resizing can model these effect in the input space to yield models robust to these nuisance variability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "AugUndo also is applicable for monocular depth estimation. Tab. 2 shows a comparison using the standard augmentation procedure of Monodepth2, HR-Depth, and Lite-Mono and using AugUndo. Tab. 2 shows that AugUndo consistently improves all models across all error and accuracy metrics. Thus, validating" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 2: Monocular depth estimation on VOID. AugUndo is applicable for monocular depth estimation and consistently improves three monocular depth models." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.187, + 0.782, + 0.282 + ], + "angle": 0, + "content": "
MethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.252 ↑δ < 1.253 ↑
Monodepth2283.861±3.732395.947±5.7280.183±0.0020.100±0.0030.717±0.0050.922±0.0040.975±0.002
+ AugUndo277.696±4.861388.088±5.7680.178±0.0030.095±0.0040.724±0.0070.925±0.0040.978±0.002
HR-Depth286.282±7.059399.112±9.1840.185±0.0040.100±0.0040.714±0.0120.919±0.0060.975±0.002
+ AugUndo283.086±6.787394.261±9.1330.181±0.0050.097±0.0050.718±0.0130.922±0.0040.977±0.002
Lite-Mono319.910±15.00446.005±22.970.209±0.0130.129±0.0190.669±0.0160.892±0.0110.963±0.006
+ AugUndo308.010±0.859426.626±0.4840.200±0.0030.114±0.0010.674±0.0050.901±0.0020.969±0.002
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.785, + 0.41 + ], + "angle": 0, + "content": "the hypothesis that AugUndo can be applied generically to improve monocular depth estimation. Specifically, we observe a boost in the most difficult accuracy metric \\((\\delta < 1.25)\\), where Monodepth2 improves from 0.717 to 0.724, HR-Depth from 0.714 to 0.718 and Lite-Mono from 0.669 to 0.674. At the same time, for AbsRel metric, Monodepth2 improves from 0.183 to 0.178, HR-depth improves from 0.185 to 0.181, and Lite-Mono improves from 0.209 to 0.200." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.686 + ], + "angle": 0, + "content": "Sensitivity study. One limitation of existing depth completion training pipeline is that there are little to no augmentations applied to sparse depth modality. However, in real-world applications, sparse depth has, in fact, high variability, i.e., features tracked in SLAM/VIO systems will vary depending on the scene (points are dropped or added to the state), and point cloud densities returned by a sensor will differ based on specifications. To further examine the effect of AugUndo on sparse depth, we study the sensitivity to changes in sparse depth density by testing models on VOID1500 (\\(\\approx\\)1500 points) on VOID500 (\\(\\approx\\)500). For the \\(3 \\times\\) reduction in sparse points, AugUndo improves robustness by an average of \\(23.1\\%\\) across all methods. These improvements result from the geometric and occlusion augmentations made possible via AugUndo, which greatly increases sparse depth variations, i.e., decimating them through resizing, reorienting their configuration through rotation, translating them out of frame, and randomly occluding them, to avoid overfitting particular sparse depth configurations. We further push their limits in Sec. H of the Supp. Mat., where we test them on VOID150 with a \\(10 \\times\\) reduction in sparse points from the training set (VOID1500) and observe the same trend of improvements. This shows that AugUndo significantly improves robustness of models to changes in sparse depth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Zero-shot generalization. We test depth completion models trained on VOID on NYUv2 and ScanNet. Tab. 3 shows an average of \\(23.2\\%\\) improvement on NYUv2 and \\(27.6\\%\\) on ScanNet. Applying AugUndo to VOICED greatly improve its generalizability to both NYUv2 and ScanNet. This is likely due to the scaffolding densification employed by VOICED, where the network can overfit to foldings of particular sparse depth configurations and therefore, does not generalize well when presented with sparse depth maps with different configurations. Like our sensitivity study (Fig. 2, VOID500), occlusion and geometric augmentations introduce variation into the sparse depth configurations and densities, which alleviates overfitting to specific point clouds or 3D scenes; hence," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Zero-shot transfer from VOID to NYUv2 and ScanNet for depth completion. AugUndo improves generalization of models trained on VOID to novel datasets by an average of \\(25.4\\%\\) for all evaluation metrics (in millimeters) across both datasets." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.201, + 0.782, + 0.294 + ], + "angle": 0, + "content": "
MethodNYUv2ScanNet
MAE ↓RMSE ↓iMAE ↓iRMSE ↓MAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED2240±143.902427±143.49211±9.60238±10.891562±136.791764±146.39270±17.25311±17.36
+ AugUndo990±82.481181±65.55110±6.11132±6.01638±44.74791±79.48131±7.68170±6.32
FusionNet132.24±2.12236.16±4.5928.68±0.4261.87±1.20109.47±3.01206.33±6.1155.45±1.56122.52±2.04
+ AugUndo124.93±3.05227.23±4.9625.70±0.4154.09±0.87100.64±2.31195.85±5.8545.98±0.7899.90±5.00
KBNet138.31±5.74257.99±10.3625.48±0.6351.77±0.99103.05±4.99217.12±13.3536.23±1.1276.55±2.90
+ AugUndo118.60±3.44231.13±8.8522.06±0.3147.07±0.7082.53±4.33175.30±11.1329.87±1.0663.78±1.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.314, + 0.784, + 0.342 + ], + "angle": 0, + "content": "Table 4: Zero-shot transfer from VOID to NYUv2 and ScanNet for depth estimation. AugUndo consistently improves generalization (in meters) for monocular depth models." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.357, + 0.77, + 0.499 + ], + "angle": 0, + "content": "
DatasetMethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.25² ↑δ < 1.25³ ↑
NYUv2Monodepth20.432±0.0030.556±0.0050.205±0.0010.159±0.0010.683±0.0050.907±0.0010.975±0.001
+ AugUndo0.415±0.0050.537±0.0060.196±0.0020.148±0.0030.700±0.0080.915±0.0020.977±0.001
HR-Depth0.424±0.0030.549±0.0030.201±0.0020.154±0.0020.692±0.0030.910±0.0020.976±0.001
+ AugUndo0.421±0.0090.542±0.0110.199±0.0050.152±0.0060.696±0.0090.913±0.0040.976±0.001
Lite-Mono0.480±0.0120.616±0.0180.231±0.0080.199±0.0150.637±0.0090.879±0.0090.964±0.004
+ AugUndo0.468±0.0030.595±0.0030.225±0.0040.187±0.0050.646±0.0020.889±0.0030.968±0.001
ScanNetMonodepth20.284±0.0040.368±0.0050.177±0.0020.097±0.0020.741±0.0060.931±0.0030.980±0.001
+ AugUndo0.270±0.0030.351±0.0040.169±0.0010.088±0.0020.759±0.0040.937±0.0010.982±0.001
HR-Depth0.282±0.0040.366±0.0050.175±0.0020.095±0.0020.743±0.0050.929±0.0030.979±0.002
+ AugUndo0.274±0.0070.357±0.0090.172±0.0050.092±0.0050.754±0.0090.935±0.0030.981±0.001
Lite-Mono0.296±0.0020.388±0.0070.185±0.0030.109±0.0060.731±0.0020.921±0.0020.976±0.001
+ AugUndo0.296±0.0010.382±0.0010.185±0.0010.105±0.0020.728±0.0020.924±0.0010.977±0.001
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.784, + 0.609 + ], + "angle": 0, + "content": "improving VOICED from \\(49.9\\%\\) and \\(52.6\\%\\) on NYUv2 and ScanNet, respectively. For FusionNet and KBNet, we still see nontrivial improvement: FusionNet improves by \\(8\\%\\) and \\(12.1\\%\\) on NYUv2 and ScanNet, and KBNet by \\(11.7\\%\\) and \\(18.3\\%\\). This further validates our hypothesis that by applying a more diverse set of transformations, we are able to improve generalization to new datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.778 + ], + "angle": 0, + "content": "For depth estimation, we tested Monodepth2, HR-Depth, Lite-Mono (trained on VOID) for zero-shot generalizability to NYUv2 and ScanNet. The results are shown in Tab. 4, where training with AugUndo consistently improves generalization errors across all evaluation metrics for both NYUv2 and ScanNet. Notably, for RMSE metric, Monodepth2 improves from 0.556 to 0.537, HR-Depth from 0.549 to 0.542, Lite-Mono from 0.616 to 0.595. For ScanNet, similar improvement on RMSE can also be observed, where Monodepth2 improves from 0.368 to 0.351, HR-Depth from 0.366 to 0.357, and Lite-Mono from 0.388 to 0.382. The improvement in RMSE metric, which is sensitive to outliers, highlights AugUndo's ability to model different input data distributions, collected by a different camera of different orientation with different object colors and layouts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Results on KITTI. We begin with quantitative results for depth completion. While AugUndo consistently improves all methods (Tab. 5), we note that the improvement is less, but respectable, in this case: \\(\\approx 5.2\\%\\) overall with the largest gain in FusionNet of \\(6.41\\%\\). This is largely due to the small scene varia" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 5: Depth completion on KITTI. AugUndo consistently improves all methods for all metrics (in millimeters) by approximately \\(5.2\\%\\) on average." + }, + { + "type": "table", + "bbox": [ + 0.348, + 0.187, + 0.653, + 0.277 + ], + "angle": 0, + "content": "
MethodMAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED318.59±7.741,213.60±17.491.30±0.053.72±0.04
+ AugUndo295.41±0.301,159.27±5.441.20±0.023.49±0.03
FusionNet285.55±2.161,174.47±10.671.20±0.033.45±0.08
+ AugUndo267.69±1.851,157.07±4.611.08±0.023.19±0.03
KBNet263.90±3.631,130.66±6.221.05±0.013.24±0.04
+ AugUndo256.37±1.001,114.53±3.791.01±0.013.13±0.03
" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.296, + 0.761, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.398, + 0.789, + 0.456 + ], + "angle": 0, + "content": "Fig. 3: Depth completion on KITTI. We compare KBNet and FusionNet trained with standard augmentations and with AugUndo. AugUndo consistently produces lower errors in highlighted regions where structures may have arbitrary orientation (vegetation) and regions near image border that typically lacks correspondence during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.718 + ], + "angle": 0, + "content": "tions in the outdoor driving scenarios, i.e., ground plane with vehicles, buildings on the sides, horizontal lidar scans, and largely planar motion. The dataset bias is strong enough to render vertical flip detrimental to performance. This is also evident in existing works as none utilizes vertical flip as augmentation. Nonetheless, AugUndo still improves performance, where point removal models different lidar returns patterns, and resizing simulates large variations in scales of objects observed in road scenes. Similar to indoors, translations help model occlusions by shifting the projection with different principal points. Fig. 3 compares FusionNet and KBNet trained using conventional augmentations and AugUndo. AugUndo improves over objects that may have diverse orientations (i.e., tree branches and vegetation), thanks to rotation augmentations. Improvements are also observed in \"small\" objects at far regions where resizing can simulate zooming in/out to emulate objects of different sizes. Through translation and occlusion, AugUndo also improves on occlusion boundaries during training (highlighted), making estimates near image borders and object boundaries more robust." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.79, + 0.841 + ], + "angle": 0, + "content": "We additionally show results for monocular depth estimation on KITTI in Tab. 6, where we compare the standard augmentation pipelines used by [15,35, 81] and AugUndo. We observe similar trends in performance gain as we did in depth completion trained on KITTI: Applying our set of augmentations improves most metrics for all methods. We observe notable improvements in \\(\\delta < 1.25\\), the most difficult accuracy metric, from 0.869 to 0.879 in Monodepth2, from 0.879 to 0.883 in HR-Depth, and 0.862 to 0.863 in Lite-Mono. For Monodepth2, we also observe a large reduction in AbsRel error, improving it from 0.117 to 0.113." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.173 + ], + "angle": 0, + "content": "Table 6: Monocular depth estimation on KITTI. AugUndo consistently improve on evaluation metrics (in meters) across different models." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.187, + 0.758, + 0.276 + ], + "angle": 0, + "content": "
MethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.252↑δ < 1.253↑
Monodepth2 + AugUndo2.315±0.0054.794±0.0350.117±0.0010.845±0.0300.869±0.0040.959±0.0010.982±0.001
2.237±0.0144.739±0.0320.113±0.0000.862±0.0300.879±0.0020.960±0.0010.982±0.001
HR-Depth +AugUndo2.226±0.0044.626±0.0320.113±0.0010.797±0.0220.879±0.0020.961±0.0010.982±0.000
2.185±0.0134.610±0.0290.111±0.0010.794±0.0210.883±0.0010.962±0.0010.983±0.001
Lite-Mono + AugUndo2.338±0.0054.821±0.0270.121±0.0010.875±0.0070.862±0.0020.955±0.0010.980±0.000
2.314±0.0304.780±0.0550.120±0.0010.849±0.0190.863±0.0040.955±0.0010.981±0.001
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.784, + 0.335 + ], + "angle": 0, + "content": "Similarly, for Lite-Mono, we also boosted performance across all metrics, with a particularly high reduction in SqRel from 0.875 to 0.849." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.785, + 0.442 + ], + "angle": 0, + "content": "Remarks. As AugUndo only augments the inputs and modifies the loss computations, which add negligible time to training, we note that the performance improvements from it are obtained nearly for free. Yet, the percentage gain, however, is similar to improvements obtained by each successive state-of-the-art: For example, the improvement of Lite-Mono over HR-Depth, and that of HR-Depth over PackNet for monocular depth estimation; similarly, the improvement of FusionNet over VOICED and KBNet over FusionNet for depth completion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.355, + 0.479 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.785, + 0.644 + ], + "angle": 0, + "content": "Conventionally, data augmentation aims to seek visual invariance and create a collection of equivalent classes, i.e., identifying an image and its augmented variant as the same. For geometric tasks, the underlying equivalence is in the 3D scene structures under various illumination conditions, camera, viewpoints, occlusion, etc. Assuming a rigid scene, the shapes populating it should persist regardless of the nuisance variables. This motivates the use of geometric augmentations, as it simulates the nuisances within the image. However, adoption of geometric augmentations in unsupervised geometric tasks are obstructed by artifacts introduced during transformations (see Sec. D of Supp. Mat.). AugUndo lifts this obstacle by \"undo-ing\" the augmentations before computing the loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.841 + ], + "angle": 0, + "content": "While AugUndo enables scaling up augmentations for unsupervised training, i.e., depth completion, it may also be applicable for supervised methods; though, we posit that the gain to be less as the artifacts induced from photometric and geometric augmentations of an image tend to be larger than those of a piecewise smooth depth map. AugUndo is also limited to 2D augmentations; whereas, nuisances modeled by it are projections of the 3D scene (see limitations in Sec. M in Supp. Mat.). We leave extensions to 3D for future work. We also only consider a single image and sparse depth map as input. Likewise, extensions can be made towards multi-frame tasks such as stereo, optical flow, pose estimation, etc., but one must account for their specifics and problem setups, i.e., stereo assumes frontoparallel views. This is outside of our scope, so we leave them as future directions. This work paves way for the empirical success in unsupervised geometric tasks that we have observed in other visual recognition tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.177 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by was supported by NSF 2112562 Athena AI Institute." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.204, + 0.326, + 0.22 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.238, + 0.787, + 0.294 + ], + "angle": 0, + "content": "1. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7063-7072 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.295, + 0.787, + 0.35 + ], + "angle": 0, + "content": "2. Cheng, X., Wang, P., Guan, C., Yang, R.: Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10615-10622 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.787, + 0.392 + ], + "angle": 0, + "content": "3. Cheng, X., Wang, P., Yang, R.: Depth estimation via affinity learned with convolutional spatial propagation network. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 103-119 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.787, + 0.434 + ], + "angle": 0, + "content": "4. Choi, H., Lee, H., Kim, S., Kim, S., Kim, S., Sohn, K., Min, D.: Adaptive confidence thresholding for monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12808-12818 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.435, + 0.787, + 0.476 + ], + "angle": 0, + "content": "5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.477, + 0.787, + 0.518 + ], + "angle": 0, + "content": "6. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE international conference on computer vision. pp. 2650-2658 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.787, + 0.56 + ], + "angle": 0, + "content": "7. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.787, + 0.615 + ], + "angle": 0, + "content": "8. Eldesokey, A., Felsberg, M., Holmquist, K., Persson, M.: Uncertainty-aware cnns for depth completion: Uncertainty from beginning to end. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12014-12023 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.787, + 0.672 + ], + "angle": 0, + "content": "9. Ezhov, V., Park, H., Zhang, Z., Upadhyay, R., Zhang, H., Chandrappa, C.C., Kadambi, A., Ba, Y., Dorsey, J., Wong, A.: All-day depth completion. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.7 + ], + "angle": 0, + "content": "10. Fei, X., Wong, A., Soatto, S.: Geo-supervised visual depth prediction. IEEE Robotics and Automation Letters 4(2), 1661-1668 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.787, + 0.742 + ], + "angle": 0, + "content": "1. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.787, + 0.798 + ], + "angle": 0, + "content": "12. Garg, R., Bg, V.K., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 740-756. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "13. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.238, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 270-279 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "15. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3828-3838 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.274 + ], + "angle": 0, + "content": "16. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2485-2494 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "17. Guizilini, V., Hou, R., Li, J., Ambrus, R., Gaidon, A.: Semantically-guided representation learning for self-supervised monocular depth. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.359 + ], + "angle": 0, + "content": "18. Hu, M., Wang, S., Li, B., Ning, S., Fan, L., Gong, X.: Penet: Towards precise and efficient image guided depth completion. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13656-13662. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.785, + 0.402 + ], + "angle": 0, + "content": "19. Jaritz, M., De Charette, R., Wirbel, E., Perrotton, X., Nashashibi, F.: Sparse and dense data with cnns: Depth completion and semantic segmentation. In: 2018 International Conference on 3D Vision (3DV). pp. 52-60. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.445 + ], + "angle": 0, + "content": "20. Jeon, J., Lim, H., Seo, D.U., Myung, H.: Struct-mdc: Mesh-refined unsupervised depth completion leveraging structural regularities from visual slam. IEEE Robotics and Automation Letters 7(3), 6391-6398 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.487 + ], + "angle": 0, + "content": "21. Kam, J., Kim, J., Kim, S., Park, J., Lee, S.: Costdcnet: Cost volume based depth completion for a single rgb-d image. In: European Conference on Computer Vision. pp. 257-274. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.543 + ], + "angle": 0, + "content": "22. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 582-600. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.585 + ], + "angle": 0, + "content": "23. Krishna, S., Vandrotti, B.S.: Deepsmooth: Efficient and smooth depth completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3357-3366 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.642 + ], + "angle": 0, + "content": "24. Kumar, V.R., Klingner, M., Yogamani, S., Milz, S., Fingscheidt, T., Mader, P.: Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 61-71 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.684 + ], + "angle": 0, + "content": "25. Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., Navab, N.: Deeper depth prediction with fully convolutional residual networks. In: 2016 Fourth international conference on 3D vision (3DV). pp. 239-248. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.685, + 0.785, + 0.727 + ], + "angle": 0, + "content": "26. Lao, D., Wu, Y., Liu, T.Y., Wong, A., Soatto, S.: Sub-token vit embedding via stochastic resonance transformers. In: International Conference on Machine Learning. PMLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.769 + ], + "angle": 0, + "content": "27. Lao, D., Yang, F., Wang, D., Park, H., Lu, S., Wong, A., Soatto, S.: On the viability of monocular depth pre-training for semantic segmentation. In: European Conference on Computer Vision. Springer (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.798 + ], + "angle": 0, + "content": "28. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision 81(2), 155 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Li, A., Yuan, Z., Ling, Y., Chi, W., Zhang, C., et al.: A multi-scale guided cascade hourglass network for depth completion. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 32-40 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.203 + ], + "angle": 0, + "content": "30. Li, B., Shen, C., Dai, Y., Van Den Hengel, A., He, M.: Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1119-1127 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "31. Lin, Y., Cheng, T., Zhong, Q., Zhou, W., Yang, H.: Dynamic spatial propagation network for depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1638-1646 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.787, + 0.284 + ], + "angle": 0, + "content": "32. Liu, F., Shen, C., Lin, G., Reid, I.: Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence 38(10), 2024-2039 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.338 + ], + "angle": 0, + "content": "33. Liu, T.Y., Agrawal, P., Chen, A., Hong, B.W., Wong, A.: Monitored distillation for positive congruent depth completion. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II. pp. 35-53. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.787, + 0.379 + ], + "angle": 0, + "content": "34. Lopez-Rodriguez, A., Busam, B., Mikolajczyk, K.: Project to adapt: Domain adaptation for depth completion from noisy and sparse sensor data. In: Proceedings of the Asian Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "35. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hr-depth: High resolution self-supervised monocular depth estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2294-2301 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "36. Ma, F., Cavalheiro, G.V., Karaman, S.: Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In: International Conference on Robotics and Automation (ICRA). pp. 3288-3295. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.514 + ], + "angle": 0, + "content": "37. Mahjourian, R., Wicke, M., Angelova, A.: Unsupervised learning of depth and egomotion from monocular video using 3d geometric constraints. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5667-5675 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "38. Park, H., Gupta, A., Wong, A.: Test-time adaptation for depth completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20519-20529 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.597 + ], + "angle": 0, + "content": "39. Park, J., Joo, K., Hu, Z., Liu, C.K., Kweon, I.S.: Non-local spatial propagation network for depth completion. In: European Conference on Computer Vision, ECCV 2020. European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.623 + ], + "angle": 0, + "content": "40. Perez, L., Wang, J.: The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.04621 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.664 + ], + "angle": 0, + "content": "41. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3227-3237 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "42. Poggi, M., Tosi, F., Mattoccia, S.: Learning monocular depth estimation with unsupervised trinocular assumptions. In: 2018 International conference on 3d vision (3DV). pp. 324-333. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.759 + ], + "angle": 0, + "content": "43. Qiu, J., Cui, Z., Zhang, Y., Zhang, X., Liu, S., Zeng, B., Pollefeys, M.: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3313-3322 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "44. Qu, C., Liu, W., Taylor, C.J.: Bayesian deep basis fitting for depth completion with uncertainty. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16147-16157 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "45. Qu, C., Nguyen, T., Taylor, C.: Depth completion via deep basis fitting. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 71-80 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "46. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.787, + 0.245 + ], + "angle": 0, + "content": "47. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.302 + ], + "angle": 0, + "content": "48. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12240-12249 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.344 + ], + "angle": 0, + "content": "49. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. IEEE transactions on pattern analysis and machine intelligence 31(5), 824-840 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.401 + ], + "angle": 0, + "content": "50. Shivakumar, S.S., Nguyen, T., Miller, I.D., Chen, S.W., Kumar, V., Taylor, C.J.: Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 13-20. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.43 + ], + "angle": 0, + "content": "51. Shorten, C., Khoshgoftaar, T.M.: A survey on image data augmentation for deep learning. Journal of big data 6(1), 1-48 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.785, + 0.471 + ], + "angle": 0, + "content": "52. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: European conference on computer vision. pp. 746-760. Springer (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.528 + ], + "angle": 0, + "content": "53. Singh, A.D., Ba, Y., Sarker, A., Zhang, H., Kadambi, A., Soatto, S., Srivastava, M., Wong, A.: Depth estimation from camera image and mmwave radar point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9275-9285 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.585 + ], + "angle": 0, + "content": "54. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.627 + ], + "angle": 0, + "content": "55. Taylor, L., Nitschke, G.: Improving deep learning with generic data augmentation. In: 2018 IEEE symposium series on computational intelligence (SSCI). pp. 1542-1547. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.67 + ], + "angle": 0, + "content": "56. Tosi, F., Aleotti, F., Poggi, M., Mattoccia, S.: Learning monocular depth estimation infusing traditional stereo knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9799-9809 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.712 + ], + "angle": 0, + "content": "57. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 2017 international conference on 3D Vision (3DV). pp. 11-20. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "58. Upadhyay, R., Zhang, H., Ba, Y., Yang, E., Gella, B., Jiang, S., Wong, A., Kadambi, A.: Enhancing diffusion models with 3d perspective geometry constraints. ACM Transactions on Graphics (TOG) 42(6), 1-15 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "59. Van Gansbeke, W., Neven, D., De Brabandere, B., Van Gool, L.: Sparse and noisy lidar completion with rgb guidance and uncertainty. In: 2019 16th International Conference on Machine Vision Applications (MVA). pp. 1-6. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "60. Wang, C., Buenapasada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2022-2030 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "61. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2162-2171 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "62. Wong, A., Cicek, S., Soatto, S.: Targeted adversarial perturbations for monocular depth prediction. Advances in neural information processing systems 33, 8486-8497 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.786, + 0.274 + ], + "angle": 0, + "content": "63. Wong, A., Cicek, S., Soatto, S.: Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 1495-1502 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.786, + 0.317 + ], + "angle": 0, + "content": "64. Wong, A., Fei, X., Hong, B.W., Soatto, S.: An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 3120-3127 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.786, + 0.346 + ], + "angle": 0, + "content": "65. Wong, A., Fei, X., Tsuei, S., Soatto, S.: Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters 5(2), 1899-1906 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.786, + 0.389 + ], + "angle": 0, + "content": "66. Wong, A., Soatto, S.: Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5644-5653 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.786, + 0.431 + ], + "angle": 0, + "content": "67. Wong, A., Soatto, S.: Unsupervised depth completion with calibrated backprojection layers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12747-12756 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.786, + 0.487 + ], + "angle": 0, + "content": "68. Xu, D., Ricci, E., Ouyang, W., Wang, X., Sebe, N.: Multi-scale continuous crfs as sequential deep networks for monocular depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5354-5362 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.786, + 0.53 + ], + "angle": 0, + "content": "69. Xu, Y., Zhu, X., Shi, J., Zhang, G., Bao, H., Li, H.: Depth completion from sparse lidar data with depth-normal constraints. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2811-2820 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.786, + 0.572 + ], + "angle": 0, + "content": "70. Yan, Z., Wang, K., Li, X., Zhang, Z., Li, J., Yang, J.: Desnet: Decomposed scale-consistent network for unsupervised depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3109-3117 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.786, + 0.628 + ], + "angle": 0, + "content": "71. Yang, F., Feng, C., Chen, Z., Park, H., Wang, D., Dou, Y., Zeng, Z., Chen, X., Gangopadhyay, R., Owens, A., et al.: Binding touch to everything: Learning unified multimodal tactile representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26340-26353 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.786, + 0.671 + ], + "angle": 0, + "content": "72. Yang, N., Stumberg, L.v., Wang, R., Cremers, D.: D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1281-1292 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.786, + 0.713 + ], + "angle": 0, + "content": "73. Yang, Y., Wong, A., Soatto, S.: Dense depth posterior (ddp) from single image and sparse range. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3353-3362 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.786, + 0.756 + ], + "angle": 0, + "content": "74. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: Every pixel counts: Unsupervised geometry learning with holistic 3d motion understanding. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.786, + 0.798 + ], + "angle": 0, + "content": "75. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: Lego: Learning edge with geometry all at once by watching videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 225-234 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "76. Yin, W., Liu, Y., Shen, C., Yan, Y.: Enforcing geometric constraints of virtual normal for depth prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5684-5693 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "77. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1983-1992 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "78. Yu, Z., Sheng, Z., Zhou, Z., Luo, L., Cao, S.Y., Gu, H., Zhang, H., Shen, H.L.: Aggregating feature point cloud for depth completion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8732-8743 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "79. Zeng, Z., Wang, D., Yang, F., Park, H., Soatto, S., Lao, D., Wong, A.: Wordepth: Variational language prior for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9708-9719 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "80. Zhan, H., Garg, R., Weerasekera, C.S., Li, K., Agarwal, H., Reid, I.: Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 340-349 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.397 + ], + "angle": 0, + "content": "81. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18537-18546 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "82. Zhang, Y., Funkhouser, T.: Deep depth completion of a single rgb-d image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 175-185 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.493 + ], + "angle": 0, + "content": "83. Zhang, Y., Guo, X., Poggi, M., Zhu, Z., Huang, G., Mattoccia, S.: Completionformer: Depth completion with convolutions and vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18527-18536 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.549 + ], + "angle": 0, + "content": "84. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 2022 International Conference on 3D Vision (3DV). pp. 668-678. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.785, + 0.591 + ], + "angle": 0, + "content": "85. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1851-1858 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.633 + ], + "angle": 0, + "content": "86. Zou, Y., Luo, Z., Huang, J.B.: Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In: Proceedings of the European conference on computer vision (ECCV). pp. 36-53 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.633 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_origin.pdf b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d1a8830aea7916afe622735cf129509d2737b553 --- /dev/null +++ b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/60d02a90-d56a-484e-a4c2-1efc2af7b5db_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:232817fdbb09cf8275ef79ca3f61d53160da03e373d32497dce676972e99f964 +size 1992882 diff --git a/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/full.md b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c4fb32992df4300ba0456bbfa0df246def00369d --- /dev/null +++ b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/full.md @@ -0,0 +1,278 @@ +# AugUndo: Scaling Up Augmentations for Monocular Depth Completion and Estimation + +Yangchao Wu $^{1}$ , Tian Yu Liu $^{1}$ , Hyoungseob Park $^{2}$ , Stefano Soatto $^{1}$ , Dong Lao $^{1}$ , and Alex Wong $^{2}$ + +$^{1}$ UCLA Vision Lab, Los Angeles, CA 90095, USA {wuyangchao1997, tianyu, soatto, lao}@cs.ucla.edu + +$^{2}$ Yale Vision Lab, New Haven, CT 06511, USA {hyoungseob.park, alex.wong}@yale.edu + +Abstract. Unsupervised depth completion and estimation methods are trained by minimizing reconstruction error. Block artifacts from resampling, intensity saturation, and occlusions are amongst the many undesirable by-products of common data augmentation schemes that affect image reconstruction quality, and thus the training signal. Hence, typical augmentations on images viewed as essential to training pipelines in other vision tasks have seen limited use beyond small image intensity changes and flipping. The sparse depth modality in depth completion have seen even less use as intensity transformations alter the scale of the 3D scene, and geometric transformations may decimate the sparse points during resampling. We propose a method that unlocks a wide range of previously-infeasible geometric augmentations for unsupervised depth completion and estimation. This is achieved by reversing, or "undo"-ing, geometric transformations to the coordinates of the output depth, warping the depth map back to the original reference frame. This enables computing the reconstruction losses using the original images and sparse depth maps, eliminating the pitfalls of naive loss computation on the augmented inputs and allowing us to scale up augmentations to boost performance. We demonstrate our method on indoor (VOID) and outdoor (KITTI) datasets, where we consistently improve upon recent methods across both datasets as well as generalization to four other datasets. Code available at: https://github.com/alexklwong/augundo + +Keywords: depth completion $\cdot$ monocular depth $\cdot$ augmentations + +# 1 Introduction + +Data augmentation is essential to training machine learning models; it plays a role in performance and generalization [40, 51, 55]. One common axiom of choosing augmentations is that the task output should remain invariant to the augmentation. For example, image flipping is a viable augmentation for classifying animals, since it does not alter the label. Conversely, flipping road signs can alter their meanings; hence, such augmentation can be detrimental to tasks involving road sign recognition. For geometric tasks, the range of augmentations + +is more restricted due to constraints of problem formulation: Stereo assumes pairs of frontoparallel rectified images; hence, in-plane rotations are not viable. Image-guided sparse depth completion relies on sparse points to ground estimates to metric scale; therefore, intensity transformations on sparse depth maps that alter the scale of the 3-dimensional (3D) scene are infeasible. Unsupervised learning of depth completion and estimation further limit the use of augmentations as the supervision signal comes from reconstructing the inputs, where augmenting the input introduces artifacts that impact reconstruction quality and therefore the supervision (see Sec. D, Fig. 1, Tab. 8, 9 in the Supp. Mat. for examples of artifacts and extended discussion on their effect on learning). Moreover, video-based unsupervised training assumes rigid motion, so augmentations that introduce padding (e.g., translation, rotation) will yield constant or edge extended borders across images (i.e., no motion), preventing the model from properly learning depth and pose - leaving few augmentations viable. While simulating nuisances is desirable, naively applying augmentations may do more harm than good; thus, it is unsurprising that existing work in unsupervised depth completion [33, 36, 50, 63-65, 67, 73] and estimation [14, 15, 35, 81, 85] primarily rely on a small range of photometric augmentations and flipping. + +Nevertheless, photometric augmentations help model the diverse range of illumination conditions and colors of objects that may populate the scene; geometric augmentations can simulate the various camera parameters, i.e., image resizing (zooming) can model changes in focal length, and scene arrangements, i.e., image flipping. However, block artifacts, loss during resampling, and intensity saturation are just some of the many undesirable side-effects of traditional augmentations to the image and sparse depth map for unsupervised learning of geometric tasks. To avoid compromising the supervision signal, we compute the typical reconstruction loss on the original input image and sparse depth map instead of the augmented inputs, which bypasses negative effects of reconstruction artifacts due to photometric and geometric augmentations. However, there exists a mis-alignment between the original input (e.g., image, sparse depth), and the model depth estimate as geometric augmentations induce a change in coordinates. Hence, we undo the geometric augmentations by inverting them in the output space to align the model estimate with the training target. + +Amongst the many geometric tasks, we focus on unsupervised depth completion, the task of inferring a dense depth map from an image and sparse depth map, where augmentations have seen limited use. Here, a training sample includes the input sparse depth map, its associated image and additional images of neighboring views of the same 3D scene. Our method is also applicable to unsupervised monocular depth estimation, which omits the sparse depth modality. Augmentations have traditionally been restricted to a limited range of photometric transformations and flipping - due to the need to preserve photometric consistency across a sequence of video frames used during training, and the sparse set of 3D points projected onto the image frame as a 2.5D range map; degradation to either modalities directly impacts the supervision signal as the loss function is conventionally computed on the augmented inputs. By using our + +![](images/83c8538fd14440956fcbe2378d683cabcd091e86b8f88e1b9e7d80338566e3b0.jpg) +Fig. 1: Overview. We apply photometric augmentations to the input image, and the same set of geometric augmentations to both the input image and sparse depth map. We warp the output depth back to the original reference frame with the inverse geometric transformations. This enables image and sparse depth reconstruction losses to be computed on the original inputs while unlocking previously-infeasible augmentations. + +method, loss functions involving sparse depth and image reconstruction from other views can be computed on the original inputs while applying augmentations that were previously not viable for the task. Our hypothesis: By "undoing" the augmentations, one can expand the viable set and scale up their use in training, leading to improved model performance and generalizability. + +To this end, we introduce AugUndo, an augmentation framework that enables one to apply a wide range of photometric and geometric transformations on the inputs, and to "undo" them during loss computation. This allows computing the unsupervised loss on the original images and sparse depth maps, free of artifacts, through a warping of the output depth map - obtained from augmented input onto the input frame of reference based on the inverse geometric transformation. In addition to group transformations that allow for output alignment, we combine them with commonly employed photometric augmentations. To the best of our knowledge, we are the first to propose a unified augmentation scheme for photometric and geometric augmentations for unsupervised depth completion and estimation. We demonstrate AugUndo on recent methods on indoor and outdoor settings, where we consistently improve all methods across all datasets. + +Our contributions are as follows: (1) We propose AugUndo, a simple-yet-effective framework to scale up photometric and geometric augmentations for unsupervised depth completion and estimation, without compromising the supervision signal; AugUndo can be applied in a plug-and-play manner to existing methods with negligible increase in computational costs during training. (2) We enable previously-infeasible augmentations to be used for training unsupervised methods and comprehensively ablate combinations of eleven types of augmentations to study the performance benefits of each. (3) We show that AugUndo can consistently improve robustness to shifts in sparse point densities for completion, model performance as well as zero-shot generalization for both depth completion and estimation for indoor and outdoor scenarios; thus, validating our hypothesis. + +# 2 Related Work + +Data augmentation for depth completion and estimation. While photometric transformations such as color jitter are often applied for unsupervised depth completion [36,63-65], and estimation [10,14-16,26,27,41,47,66,79,80], geometric augmentations other than simple flipping are less commonly adopted. Most supervised depth completion methods [8,18,19,21,23,29,31,43-45,69,73,82,83] similarly limit their augmentations to color jitter and horizontal flips due to sparse depth maps being decimated by rotation and scaling, causing points to be interpolated away. Nevertheless, for supervised training, it is straightforward to directly apply the same transformation to the ground truth for training. Indeed, some supervised depth completion methods [2,3,9,38,39,53,59,78] adopted random scaling, translation, in-plane rotation. Translation augmentations are also commonly applied in supervised depth estimation methods [6,7,25,30,32, 46,47,58,62,68,71,76]. We posit that the reason that such transformations can be applied to the ground truth in supervised settings is due to artifacts caused by the transformation of a piece-wise smooth depth map being less severe than those of an RGB image and its intensities. However, such assumptions do not hold in unsupervised training. These artifacts would affect the training signal for unsupervised methods, which rely on photometric correspondences and observed sparse points. Our approach aims to resolve this to allow diverse geometric augmentations to be applied in a plug-and-play fashion in unsupervised training. + +Unsupervised depth completion methods [36, 50, 63-65, 70, 73] leverage image and sparse depth reconstructions as training signals by minimizing errors between the input image and its reconstruction from other views, and errors between the input sparse depth map and the predicted depth along with a local smoothness regularizer. [36] used Perspective-n-Point [28] and RANSAC [11] to align consecutive video frames. [73] learns a depth prior conditioned on the image. [63] uses synthetic data to learn a prior on the shapes populating a scene, while [34] translates synthetic data to real domain to leverage rendered depth. [64] proposed an adaptive scheme to reduce penalties incurred on occluded regions. [67] maps the image onto the 3D scene through calibrated backprojection. [70] decouples structure and scale. [20] uses line features from visual SLAM and [33] introduced monitored distillation for positive congruent training. Augmentations for these methods are limited to a small range of photometric perturbations and image flipping. Operations such as rotation, resizing, and translation require resampling, which creates artifacts and affects the reconstruction quality. This causes performance degradation since loss is typically computed on the augmented images. Loss in sparse depth maps is further impacted as resampling and interpolation may cause loss of sparse points. Contrary to these limited augmentation schemes, we enable a large range of photometric and geometric augmentations to be introduced during training. + +Unsupervised monocular depth estimation, like depth completion, also minimizes photometric reconstruction error. [12] frames depth estimation as a novel view synthesis problem. [14] improves [12] by imposing a consistency loss on the depth predicted from left and right images. [85] uses a pose network to + +enable unsupervised training on video sequences. [15] introduced auto-masking and min reprojection loss. To improve the supervision signal, [4, 56, 61] leverage noisy proxy labels and [42] uses a trinocular assumption. Additional loss terms based on visual odometry [60], iterative closest point [37], surface normals [75], and semantic segmentation [17, 24] were also introduced; [41, 72] further included predictive uncertainty. To handle rigid and non-rigid motions in the scene, previous works explored multi-task learning to include optical flow and moving object estimation [1, 48, 74, 77, 86], and used semantics to filter out outlier regions [22]. [35] redesigned the skip connection and decoders to extract high-resolution features, [84] combined global and local representations and [81] introduced a lightweight architecture with dilated convolution and attention. In addition to depth completion, our method also shows consistent improvements on monocular depth estimation for [15, 35, 81]. + +# 3 Method Formulation + +Let $I:\varOmega\subset\mathbb{R}^2\to\mathbb{R}_+^3$ be an RGB image captured by a calibrated camera, $z:\varOmega_z\subset\varOmega\rightarrow\mathbb{R}_+$ the corresponding sparse point cloud projected onto the image plane as a sparse depth map, and $K\in \mathbb{R}^{3\times 3}$ the intrinsic calibration matrix. Given an image and its sparse depth map, depth completion aims to learn a function $f_{\theta}(I,z)$ that recovers the distance between the camera to points in the 3D scene as a dense depth map. In another mode, if sparse depth maps are not given, then the problem reduces to monocular depth estimation which learns a function $f_{\theta}(I)$ to map a single image to a depth map $\varOmega\rightarrow\mathbb{R}_{+}$ . For the ease of notation, we denote the output depth map for both depth completion and estimation as $\hat{d}\in \mathbb{R}_{+}^{H\times W}$ where $H$ and $W$ are its height and width. + +Unsupervised depth completion relies on photometric and sparse depth reconstruction errors as its primary supervision signal. Without loss of generality, we assume an input of $(I_t,z_t)$ for an RGB image and associated sparse depth map captured at time $t$ and during training, an additional set of temporally adjacent images $I_{\tau}$ for $\tau \in \Upsilon \doteq \{t - 1,t + 1\}$ . The reconstruction $\hat{I}_{t\tau}$ of $I_{t}$ from image $I_{\tau}$ is given by the reprojection based on estimated depth $\hat{d}_t\coloneqq f_\theta (\cdot)$ + +$$ +\hat {I} _ {t \tau} (x) = I _ {\tau} \left(\pi g _ {\tau t} K ^ {- 1} \bar {x} \hat {d} _ {t} (x)\right) \tag {1} +$$ + +where $\bar{x} = [x^{\top},1]^{\top}$ is the homogeneous coordinates of $x\in \Omega$ , $g_{\tau t}\in SE(3)$ the relative pose of the camera from time $t$ to $\tau$ , $K$ the intrinsic calibration matrix, and $\pi$ the canonical perspective projection. + +Using Eq. (1), a depth completion or estimation network $f_{\theta}$ minimizes + +$$ +\arg \min _ {\theta} \sum_ {\tau \in \Upsilon} \sum_ {x \in \Omega} \alpha \rho (\hat {I} _ {t \tau} (x), I _ {t} (x)) + \sum_ {x \in \Omega_ {z}} \beta \psi (\hat {d} _ {t} (x), z _ {t} (x)) + \lambda R (\hat {d} _ {t}) \tag {2} +$$ + +where $\rho$ denotes the photometric reconstruction error, typically $L_{1}$ difference in pixel values and/or structural similarity (SSIM), $\psi$ the sparse depth reconstruction error, typically $L_{1}$ or $L_{2}$ , $R$ the regularizer that biases the depth map to + +be piece-wise smooth with depth discontinuities aligned with edges in the image (commonly used by previous works [15,36,65,67]), and $\alpha$ , $\beta$ and $\lambda$ their respective weighting. Note: monocular depth omits sparse depth term, i.e., $\beta = 0$ . + +Inverse transformation. Let $\mathcal{A}_{pt}$ be the set of possible photometric transformations, and $\mathcal{A}_{ge}$ be the set of all geometric transformations. Given $T_{ge} \in \mathcal{A}_{ge}$ , we wish to obtain an inverse transformation $T_{ge}^{-1}$ such that $T_{ge} \circ T_{ge}^{-1} \approx Id$ the identity function3. At each time step, we can sample a sequence of transformations $\{T_{pt}^{1} \ldots T_{pt}^{k}\}$ and $\{T_{ge}^{1} \ldots T_{ge}^{m}\}$ respectively from $\mathcal{A}_{pt}$ and $\mathcal{A}_{ge}$ to construct transformations $T_{pt} = T_{pt}^{1} \circ \dots \circ T_{pt}^{k}$ and $T_{ge} = T_{ge}^{1} \circ \dots \circ T_{ge}^{m}$ . We denote the composition of a collection of augmentation transformations by $T = T_{ge} \circ T_{pt}$ where $T_{pt}$ denotes photometric transformation, and $T_{ge}$ denotes geometric transformation. Furthermore, we denote the inverse geometric transformations by $T_{ge}^{-1} = (T_{ge}^{m})^{-1} \circ (T_{ge}^{m-1})^{-1} \circ \dots \circ (T_{ge}^{1})^{-1}$ , which operates on the space of depth maps to reverse the geometric transformation so that we can warp the output depth map onto the reference frame of the original image. Through inverse warping, this process is differentiable. In practice, some transformations cause border regions of the image to be warped out of frame, i.e., translated off the image plane or cropped away. Hence, after warping our output depth map back to the original frame of reference using the inverse geometric transformations, border extensions (edge paddings) are introduced to handle out-of-frame regions. + +AugUndo. We apply each geometric transformation over image coordinates and resample (bilinear for image and nearest neighbor for sparse depth): + +$$ +\left[ x ^ {\prime} 1 \right] ^ {\top} = T _ {g e} \left[ x 1 \right] ^ {\top} \tag {3} +$$ + +$$ +I ^ {\prime} \left(x ^ {\prime}\right) = T _ {p t} (I) (x); z ^ {\prime} \left(x ^ {\prime}\right) = z (x) \tag {4} +$$ + +where $T_{ge}$ is the geometric transformation, $x \in \Omega$ and $x' \in \Omega$ are coordinates in the image grid, and $I'$ is the image after the transformation; for ease of notation, we hereafter denote $I' = T(I) = T_{ge} \circ T_{pt}(I)$ to include both augmentations through composition. Note that $x$ is in the original image reference frame and $x'$ is in the transformed image reference frame. Naturally, this process can be extended to multiple geometric transformations by composing them, i.e., $T_{ge} = T_{ge}^{1} \circ T_{ge}^{2} \circ \dots \circ T_{ge}^{m}$ . The reverse process is simply inverting the transformations where $T_{ge}^{-1} = (T_{ge}^{m})^{-1} \circ (T_{ge}^{m-1})^{-1} \circ \dots \circ (T_{ge}^{1})^{-1}$ : + +$$ +\left[ x ^ {\prime \prime} 1 \right] ^ {\top} = T _ {g e} ^ {- 1} \left[ x ^ {\prime} 1 \right] ^ {\top} \tag {5} +$$ + +$$ +\hat {d} \left(x ^ {\prime \prime}\right) = \hat {d} ^ {\prime} \left(x ^ {\prime}\right) \tag {6} +$$ + +where $\hat{d}'$ is the depth map inferred from augmented inputs $(I', z')$ . Once reverted back to the original reference frame, $\hat{d}$ is used to reconstruct $I_t$ from $I_\tau$ for $\tau \in \Upsilon$ as in Eq. (1). More details can be found in Alg. 1 in the Supp. Mat. + +By modeling $T_{ge}^{-1}$ , Eq. (3)-Eq. (6) allow us to apply a wide range of augmentations, while still establishing the correspondence between $I_{t}$ and $I_{\tau}$ . Specifically, + +for minimizing Eq. (2), one may simply augment the image and sparse depth by $T$ to yield $(I_t', z_t')$ as input, and reconstruct the original image and sparse depth $(I_t, z_t)$ from other views $I_\tau$ and the aligned output depth $\hat{d}$ (Eq. (6)). We note that the inverse transformation is critical for enabling the sparse depth reconstruction term to be computed properly in Eq. (2); if computed in the transformed reference frame, i.e., on $z_t'$ , many of the sparse points would be decimated by interpolation during rotation and resizing (downsampling) augmentations - leaving a lack of supervision from the sparse depth term of the loss function. We note that the original image sequence is fed to the pose estimator, rather than the augmented images. This ensures the estimated relative pose used during reprojection (Eq. (1)) also corresponds to the original images. + +Modeling geometric augmentations. Depth completion differs from typical image-based (e.g., monocular) depth estimation, where operations like resizing are susceptible to scale ambiguity and can either correspond to changes in focal length, or distance from the camera (i.e., due to changes in camera pose). In depth completion, the sparse depth map grounds image pixels to specific depth values. Hence, there is no ambiguity in distance of each sparse point from the camera. Our goal is to synthesize new training data via geometric transformations that are consistent with the original 3D scene. Given that there is no scale ambiguity, this can be achieved by either (1) changes in camera pose (extrinsics), or (2) changes in camera parameters (intrinsics). + +Augmentation via (1) would require warping the inputs for view synthesis. Since we operate under the unsupervised setting, accurate warping (preserving 3D scene) without access to dense ground truth depth values is difficult to achieve. An exception to this is rotation, which we address below. On the other hand, (2) can be achieved via standard 2D image transformations while preserving the 3D structure of the scene. This motivates our modeling of geometric transformations, i.e., resizing, translation, as changes in camera parameters. + +For example, augmenting focal length (i.e. zooming in/out) is akin to "resizing". Similarly, translation can be used to model shifting of the optic center, i.e. capturing the same 3D scene using cameras with different principal point offsets. Note: image rotation can be seen as rotating a camera about its optical axis, which does not change the distance of a point from the camera. Thus, the camera position and the scene are kept constant across augmentations, eliminating the need for adjusting depth values via warping during the process. Instead, we only have to realign the output depth back to the original frame of reference (Eqs. (5) and (6)), which makes our method computationally efficient. + +Augmentations. Our choices are based on common nuisance variabilities, i.e., changes in illumination, occlusions, and object color, scene layout (flip), and camera parameters (resize, translation) and orientation (rotation). + +Photometric. We include brightness, contrast, saturation, and hue, where all follow standard augmentation pipeline in existing works [15,65,67,81]. Applying the inverse of photometric augmentation can be viewed as recovering the original image; hence, we directly use original image instead of applying transformations to the intensities (which are not recoverable if saturated at the max value). + +Occlusion. We consider patch removal and sparse point removal. For the former, we randomly select a percentage of pixels $x \in \Omega$ in the image and remove (with zero-fill) an arbitrary-sized patch around it. For the latter, we randomly sample a percentage of points $x \in \Omega_z$ in the sparse depth map. The inverse transformation of this is simply the original image and sparse depth map, both used in loss computation before augmentation. + +Flip. We consider horizontal and vertical flips. When applied, the same flip operation is used for both input image and sparse depth maps. We record the flip type during data augmentation. During loss computation, we reverse the flip direction to align output depth with the original image and sparse depth map. + +Resize. We define a new image plane of the input resolution and generate a random scaling factor to be applied along both height and width directions. The image is warped to the new image plane, where any point warped out of the plane is excluded; borders of the warped image are extended to the bounds of the image plane by edge replication. Sparse points occupying multiple pixels are eroded to a single point for resizing with a factor greater than 1. During loss computation, we warp the output depth map onto a new image plane of the same dimensions by the inverse scaling matrix of the recorded scaling factors; borders of the warped depth map are extended by edge replication if necessary. We view resizing factors greater and less than 1 (zooming in and out) as two separate forms of augmentations to distinguish their contributions. + +Rotation. Naive rotation leads to the loss of large areas of the image, i.e., cropped away to retain the same-size image, discarding large portions of possible co-visible regions and also supervisory signals. To preserve the entire image, we first warp the image by a randomly generated angle to a new (larger) image plane, so that the rotated image fits tightly within the new image. As image sizes within a batch can vary depending on the rotation angle, we center-pad (uniformly on each side) each image in the batch to the maximum width and height of the augmented batch. To reverse the rotation on the output depth, we warp the output depth map back by the inverse rotation matrix, then perform a center crop on the depth map to align with the original input. + +Translation. We define a new image plane of the same dimensions as the input and generate random height and width translation factors. The coordinates of the input are translated and its pixels are inverse warped onto the new image plane. Any pixel warped out of the image plane is excluded. Borders of the warped image are extended to the bounds of the image plane by edge replication. During loss computation, we warp the output depth map onto a new image plane of the same dimensions by the inverse translation matrix. Borders of the warped depth map are extended to the bounds of the image plane by edge replication. + +# 4 Experiments + +We demonstrate AugUndo on recent unsupervised depth completion (VOICED [65], FusionNet [63], KBNet [67]) and estimation (Monodepth2 [15], HRDepth [35], LiteMono [81]) methods on two datasets (KITTI [13, 57], Void [65]) here. + +Due to space limitations, we defer the results of MonDi [33] and DesNet [70] to Sec. K in the Supp. Mat. To evaluate zero-shot generalization, we test on three additional datasets for each task: NYUv2 [52], ScanNet [5], Waymo [54], and Make3d [49] - where we transfer models on Void to NYUv2 (Tab. 3) and ScanNet, and from KITTI to Waymo (Tab. 4 in Supp. Mat.) for completion and Make3d [49] (Tab. 5 in Supp. Mat.) for estimation. We also test model sensitivity to different sparse depth input densities in Tab. 1 (right) - see Sec. H of the Supp. Mat for an extensive study. Additionally, we present an comprehensive ablation study in Sec. G in the Supp. Mat. to test the effect of each augmentation. We also provide results on modeling AugUndo as changes in camera motion or parameters in Tab. 7 of the Supp. Mat.; interpreting geometric augmentations as different camera parameters yields better performance. Tables. 8, 9 in Supp. Mat. show that unsupervised models cannot be trained with naive geometric augmentations. See Sec. E of the Supp. Mat. for descriptions of datasets. + +Experiment setup. We followed the original settings of the open-sourced code for each work and modified the data handling and loss function to incorporate AugUndo. We perform 4 independent trials for each experiment and report their means and standard deviations. To ensure a fair comparison, we train all models from scratch. Below, we report the best combination of augmentations found empirically through extensive experiments on each dataset. See Secs. A and B in Supp. Mat. for implementation details, and evaluation metrics, respectively. + +Augmentations. Through a search over augmentation types and values, we found a consistent set that tends to yield improvements across all methods with small changes to degree of augmentation catered to each method. See Sec. C in Supp. Mat. for details of augmentation parameters. Rows in Tabs. 1 to 6 marked with "+" AugUndo" denote models trained with our method. We note that performance gain can be obtained by typical set and ranges of augmentations and does not require a meticulous selection of hyper-parameters (see Sec. J of Supp. Mat. for a sensitivity study). Results reported here aims to study how far one can push performance and generalization, purely from augmentations. + +Results on VOID. We begin by presenting quantitative results for AugUndo on unsupervised depth completion. Tab. 1 (left, VOID1500) shows our main results on the VOID benchmark. By training the models with AugUndo, we observe an average overall improvement of $18.3\%$ across all methods and metrics on VOID1500. Specifically, we improve VOICED by $26.4\%$ , FusionNet by $16.3\%$ , and KBNet by $12.2\%$ . These experiments validate our hypothesis that by applying a wider range of augmentations, we are able to improve the baseline model's performance. They also illustrate the lack in use of augmentations in existing works: incorporating standard augmentations (albeit requiring modification to augmentation and loss computation pipelines) can yield a large performance boost. Fig. 2 shows a head-to-head comparison between KBNet trained using standard procedure in [67] and KBNet trained using AugUndo. We observe qualitative improvements from AugUndo where we improve KBNet in homogeneous regions and image borders, i.e., pillar (left), cabinet (middle), wall (right). Applying geometric augmentations yields models with fewer border + +Table 1: Depth completion on Void. AugUndo improves performance by an average of $18.3\%$ across all methods and metrics on Void1500. When models trained on Void1500 are tested Void500, AugUndo improves by $23.1\%$ , as translation, resizing, and occlusion augmentations vary sparsity by removing sparse points from the input. + +
MethodVOID1500VOID500
MAE ↓RMSE ↓iMAE ↓iRMSE ↓MAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED74.78±2.69139.75±4.5739.20±1.4671.98±2.54137.01±4.23235.80±7.8271.36±1.86130.63±5.66
+ AugUndo52.73±0.41111.09±0.9226.93±0.5454.46±0.3892.99±1.11176.94±1.3846.43±0.8591.10±1.64
FusionNet52.11±0.44113.30±1.1828.53±0.5258.79±2.0197.73±0.73194.32±1.3658.65±1.31122.95±3.04
+ AugUndo41.16±0.1899.21±0.3922.23±0.3553.07±1.3074.97±0.69162.71±2.3940.44±1.3992.11±5.79
KBNet38.11±0.7795.22±1.7219.51±0.1446.70±0.4878.44±1.39178.17±3.2737.56±0.6183.43±1.89
+ AugUndo33.32±0.1885.67±0.3916.61±0.2941.24±0.6066.97±0.81151.55±2.0331.63±0.5371.90±0.82
+ +![](images/80be05bc519fb2d33fb4934d9933db4e29ba94e2eb280e65aa1a27a78ce9b7e9.jpg) +Fig. 2: Depth completion on VOID. We compare KBNet trained with standard augmentations and with AugUndo. AugUndo consistently produces lower errors with reduced border artifacts and improves on homogeneous regions, i.e., pillar (left), cabinet (middle), wall of staircase (right). Error maps are highlighted for comparison. + +artifacts as we apply random translation, which warps part of the image out of frame. This allows us to simulate border occlusion, where there lacks correspondence in an adjacent frame. While training with standard protocol results in failures to recover structures near the image border, training with AugUndo can render models robust to them by computing the loss on the original frame of reference (where we do have correspondence) through our inverse transformations of the depth map. Additionally, resizing allows the model to learn multiple resolutions of the input, akin to zooming in and out, which can also impose smoothness in homogeneous regions through the scale-space transition; whereas, rotation can simulate diverse camera orientations. Fig. 2 shows that translation, rotation, and resizing can model these effect in the input space to yield models robust to these nuisance variability. + +AugUndo also is applicable for monocular depth estimation. Tab. 2 shows a comparison using the standard augmentation procedure of Monodepth2, HR-Depth, and Lite-Mono and using AugUndo. Tab. 2 shows that AugUndo consistently improves all models across all error and accuracy metrics. Thus, validating + +Table 2: Monocular depth estimation on VOID. AugUndo is applicable for monocular depth estimation and consistently improves three monocular depth models. + +
MethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.252 ↑δ < 1.253 ↑
Monodepth2283.861±3.732395.947±5.7280.183±0.0020.100±0.0030.717±0.0050.922±0.0040.975±0.002
+ AugUndo277.696±4.861388.088±5.7680.178±0.0030.095±0.0040.724±0.0070.925±0.0040.978±0.002
HR-Depth286.282±7.059399.112±9.1840.185±0.0040.100±0.0040.714±0.0120.919±0.0060.975±0.002
+ AugUndo283.086±6.787394.261±9.1330.181±0.0050.097±0.0050.718±0.0130.922±0.0040.977±0.002
Lite-Mono319.910±15.00446.005±22.970.209±0.0130.129±0.0190.669±0.0160.892±0.0110.963±0.006
+ AugUndo308.010±0.859426.626±0.4840.200±0.0030.114±0.0010.674±0.0050.901±0.0020.969±0.002
+ +the hypothesis that AugUndo can be applied generically to improve monocular depth estimation. Specifically, we observe a boost in the most difficult accuracy metric $(\delta < 1.25)$ , where Monodepth2 improves from 0.717 to 0.724, HR-Depth from 0.714 to 0.718 and Lite-Mono from 0.669 to 0.674. At the same time, for AbsRel metric, Monodepth2 improves from 0.183 to 0.178, HR-depth improves from 0.185 to 0.181, and Lite-Mono improves from 0.209 to 0.200. + +Sensitivity study. One limitation of existing depth completion training pipeline is that there are little to no augmentations applied to sparse depth modality. However, in real-world applications, sparse depth has, in fact, high variability, i.e., features tracked in SLAM/VIO systems will vary depending on the scene (points are dropped or added to the state), and point cloud densities returned by a sensor will differ based on specifications. To further examine the effect of AugUndo on sparse depth, we study the sensitivity to changes in sparse depth density by testing models on VOID1500 ( $\approx$ 1500 points) on VOID500 ( $\approx$ 500). For the $3 \times$ reduction in sparse points, AugUndo improves robustness by an average of $23.1\%$ across all methods. These improvements result from the geometric and occlusion augmentations made possible via AugUndo, which greatly increases sparse depth variations, i.e., decimating them through resizing, reorienting their configuration through rotation, translating them out of frame, and randomly occluding them, to avoid overfitting particular sparse depth configurations. We further push their limits in Sec. H of the Supp. Mat., where we test them on VOID150 with a $10 \times$ reduction in sparse points from the training set (VOID1500) and observe the same trend of improvements. This shows that AugUndo significantly improves robustness of models to changes in sparse depth. + +Zero-shot generalization. We test depth completion models trained on VOID on NYUv2 and ScanNet. Tab. 3 shows an average of $23.2\%$ improvement on NYUv2 and $27.6\%$ on ScanNet. Applying AugUndo to VOICED greatly improve its generalizability to both NYUv2 and ScanNet. This is likely due to the scaffolding densification employed by VOICED, where the network can overfit to foldings of particular sparse depth configurations and therefore, does not generalize well when presented with sparse depth maps with different configurations. Like our sensitivity study (Fig. 2, VOID500), occlusion and geometric augmentations introduce variation into the sparse depth configurations and densities, which alleviates overfitting to specific point clouds or 3D scenes; hence, + +Table 3: Zero-shot transfer from VOID to NYUv2 and ScanNet for depth completion. AugUndo improves generalization of models trained on VOID to novel datasets by an average of $25.4\%$ for all evaluation metrics (in millimeters) across both datasets. + +
MethodNYUv2ScanNet
MAE ↓RMSE ↓iMAE ↓iRMSE ↓MAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED2240±143.902427±143.49211±9.60238±10.891562±136.791764±146.39270±17.25311±17.36
+ AugUndo990±82.481181±65.55110±6.11132±6.01638±44.74791±79.48131±7.68170±6.32
FusionNet132.24±2.12236.16±4.5928.68±0.4261.87±1.20109.47±3.01206.33±6.1155.45±1.56122.52±2.04
+ AugUndo124.93±3.05227.23±4.9625.70±0.4154.09±0.87100.64±2.31195.85±5.8545.98±0.7899.90±5.00
KBNet138.31±5.74257.99±10.3625.48±0.6351.77±0.99103.05±4.99217.12±13.3536.23±1.1276.55±2.90
+ AugUndo118.60±3.44231.13±8.8522.06±0.3147.07±0.7082.53±4.33175.30±11.1329.87±1.0663.78±1.30
+ +Table 4: Zero-shot transfer from VOID to NYUv2 and ScanNet for depth estimation. AugUndo consistently improves generalization (in meters) for monocular depth models. + +
DatasetMethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.25² ↑δ < 1.25³ ↑
NYUv2Monodepth20.432±0.0030.556±0.0050.205±0.0010.159±0.0010.683±0.0050.907±0.0010.975±0.001
+ AugUndo0.415±0.0050.537±0.0060.196±0.0020.148±0.0030.700±0.0080.915±0.0020.977±0.001
HR-Depth0.424±0.0030.549±0.0030.201±0.0020.154±0.0020.692±0.0030.910±0.0020.976±0.001
+ AugUndo0.421±0.0090.542±0.0110.199±0.0050.152±0.0060.696±0.0090.913±0.0040.976±0.001
Lite-Mono0.480±0.0120.616±0.0180.231±0.0080.199±0.0150.637±0.0090.879±0.0090.964±0.004
+ AugUndo0.468±0.0030.595±0.0030.225±0.0040.187±0.0050.646±0.0020.889±0.0030.968±0.001
ScanNetMonodepth20.284±0.0040.368±0.0050.177±0.0020.097±0.0020.741±0.0060.931±0.0030.980±0.001
+ AugUndo0.270±0.0030.351±0.0040.169±0.0010.088±0.0020.759±0.0040.937±0.0010.982±0.001
HR-Depth0.282±0.0040.366±0.0050.175±0.0020.095±0.0020.743±0.0050.929±0.0030.979±0.002
+ AugUndo0.274±0.0070.357±0.0090.172±0.0050.092±0.0050.754±0.0090.935±0.0030.981±0.001
Lite-Mono0.296±0.0020.388±0.0070.185±0.0030.109±0.0060.731±0.0020.921±0.0020.976±0.001
+ AugUndo0.296±0.0010.382±0.0010.185±0.0010.105±0.0020.728±0.0020.924±0.0010.977±0.001
+ +improving VOICED from $49.9\%$ and $52.6\%$ on NYUv2 and ScanNet, respectively. For FusionNet and KBNet, we still see nontrivial improvement: FusionNet improves by $8\%$ and $12.1\%$ on NYUv2 and ScanNet, and KBNet by $11.7\%$ and $18.3\%$ . This further validates our hypothesis that by applying a more diverse set of transformations, we are able to improve generalization to new datasets. + +For depth estimation, we tested Monodepth2, HR-Depth, Lite-Mono (trained on VOID) for zero-shot generalizability to NYUv2 and ScanNet. The results are shown in Tab. 4, where training with AugUndo consistently improves generalization errors across all evaluation metrics for both NYUv2 and ScanNet. Notably, for RMSE metric, Monodepth2 improves from 0.556 to 0.537, HR-Depth from 0.549 to 0.542, Lite-Mono from 0.616 to 0.595. For ScanNet, similar improvement on RMSE can also be observed, where Monodepth2 improves from 0.368 to 0.351, HR-Depth from 0.366 to 0.357, and Lite-Mono from 0.388 to 0.382. The improvement in RMSE metric, which is sensitive to outliers, highlights AugUndo's ability to model different input data distributions, collected by a different camera of different orientation with different object colors and layouts. + +Results on KITTI. We begin with quantitative results for depth completion. While AugUndo consistently improves all methods (Tab. 5), we note that the improvement is less, but respectable, in this case: $\approx 5.2\%$ overall with the largest gain in FusionNet of $6.41\%$ . This is largely due to the small scene varia + +Table 5: Depth completion on KITTI. AugUndo consistently improves all methods for all metrics (in millimeters) by approximately $5.2\%$ on average. + +
MethodMAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED318.59±7.741,213.60±17.491.30±0.053.72±0.04
+ AugUndo295.41±0.301,159.27±5.441.20±0.023.49±0.03
FusionNet285.55±2.161,174.47±10.671.20±0.033.45±0.08
+ AugUndo267.69±1.851,157.07±4.611.08±0.023.19±0.03
KBNet263.90±3.631,130.66±6.221.05±0.013.24±0.04
+ AugUndo256.37±1.001,114.53±3.791.01±0.013.13±0.03
+ +![](images/b7613559793937b55ff355bff53e46aa37439847fb190cf5e49589c5b6fab1a6.jpg) +Fig. 3: Depth completion on KITTI. We compare KBNet and FusionNet trained with standard augmentations and with AugUndo. AugUndo consistently produces lower errors in highlighted regions where structures may have arbitrary orientation (vegetation) and regions near image border that typically lacks correspondence during training. + +tions in the outdoor driving scenarios, i.e., ground plane with vehicles, buildings on the sides, horizontal lidar scans, and largely planar motion. The dataset bias is strong enough to render vertical flip detrimental to performance. This is also evident in existing works as none utilizes vertical flip as augmentation. Nonetheless, AugUndo still improves performance, where point removal models different lidar returns patterns, and resizing simulates large variations in scales of objects observed in road scenes. Similar to indoors, translations help model occlusions by shifting the projection with different principal points. Fig. 3 compares FusionNet and KBNet trained using conventional augmentations and AugUndo. AugUndo improves over objects that may have diverse orientations (i.e., tree branches and vegetation), thanks to rotation augmentations. Improvements are also observed in "small" objects at far regions where resizing can simulate zooming in/out to emulate objects of different sizes. Through translation and occlusion, AugUndo also improves on occlusion boundaries during training (highlighted), making estimates near image borders and object boundaries more robust. + +We additionally show results for monocular depth estimation on KITTI in Tab. 6, where we compare the standard augmentation pipelines used by [15,35, 81] and AugUndo. We observe similar trends in performance gain as we did in depth completion trained on KITTI: Applying our set of augmentations improves most metrics for all methods. We observe notable improvements in $\delta < 1.25$ , the most difficult accuracy metric, from 0.869 to 0.879 in Monodepth2, from 0.879 to 0.883 in HR-Depth, and 0.862 to 0.863 in Lite-Mono. For Monodepth2, we also observe a large reduction in AbsRel error, improving it from 0.117 to 0.113. + +Table 6: Monocular depth estimation on KITTI. AugUndo consistently improve on evaluation metrics (in meters) across different models. + +
MethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.252↑δ < 1.253↑
Monodepth2 + AugUndo2.315±0.0054.794±0.0350.117±0.0010.845±0.0300.869±0.0040.959±0.0010.982±0.001
2.237±0.0144.739±0.0320.113±0.0000.862±0.0300.879±0.0020.960±0.0010.982±0.001
HR-Depth +AugUndo2.226±0.0044.626±0.0320.113±0.0010.797±0.0220.879±0.0020.961±0.0010.982±0.000
2.185±0.0134.610±0.0290.111±0.0010.794±0.0210.883±0.0010.962±0.0010.983±0.001
Lite-Mono + AugUndo2.338±0.0054.821±0.0270.121±0.0010.875±0.0070.862±0.0020.955±0.0010.980±0.000
2.314±0.0304.780±0.0550.120±0.0010.849±0.0190.863±0.0040.955±0.0010.981±0.001
+ +Similarly, for Lite-Mono, we also boosted performance across all metrics, with a particularly high reduction in SqRel from 0.875 to 0.849. + +Remarks. As AugUndo only augments the inputs and modifies the loss computations, which add negligible time to training, we note that the performance improvements from it are obtained nearly for free. Yet, the percentage gain, however, is similar to improvements obtained by each successive state-of-the-art: For example, the improvement of Lite-Mono over HR-Depth, and that of HR-Depth over PackNet for monocular depth estimation; similarly, the improvement of FusionNet over VOICED and KBNet over FusionNet for depth completion. + +# 5 Discussion + +Conventionally, data augmentation aims to seek visual invariance and create a collection of equivalent classes, i.e., identifying an image and its augmented variant as the same. For geometric tasks, the underlying equivalence is in the 3D scene structures under various illumination conditions, camera, viewpoints, occlusion, etc. Assuming a rigid scene, the shapes populating it should persist regardless of the nuisance variables. This motivates the use of geometric augmentations, as it simulates the nuisances within the image. However, adoption of geometric augmentations in unsupervised geometric tasks are obstructed by artifacts introduced during transformations (see Sec. D of Supp. Mat.). AugUndo lifts this obstacle by "undo-ing" the augmentations before computing the loss. + +While AugUndo enables scaling up augmentations for unsupervised training, i.e., depth completion, it may also be applicable for supervised methods; though, we posit that the gain to be less as the artifacts induced from photometric and geometric augmentations of an image tend to be larger than those of a piecewise smooth depth map. AugUndo is also limited to 2D augmentations; whereas, nuisances modeled by it are projections of the 3D scene (see limitations in Sec. M in Supp. Mat.). We leave extensions to 3D for future work. We also only consider a single image and sparse depth map as input. Likewise, extensions can be made towards multi-frame tasks such as stereo, optical flow, pose estimation, etc., but one must account for their specifics and problem setups, i.e., stereo assumes frontoparallel views. This is outside of our scope, so we leave them as future directions. This work paves way for the empirical success in unsupervised geometric tasks that we have observed in other visual recognition tasks. + +Acknowledgements. This work was supported by was supported by NSF 2112562 Athena AI Institute. + +# References + +1. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7063-7072 (2019) +2. Cheng, X., Wang, P., Guan, C., Yang, R.: Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10615-10622 (2020) +3. Cheng, X., Wang, P., Yang, R.: Depth estimation via affinity learned with convolutional spatial propagation network. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 103-119 (2018) +4. Choi, H., Lee, H., Kim, S., Kim, S., Kim, S., Sohn, K., Min, D.: Adaptive confidence thresholding for monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12808-12818 (2021) +5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) +6. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE international conference on computer vision. pp. 2650-2658 (2015) +7. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014) +8. Eldesokey, A., Felsberg, M., Holmquist, K., Persson, M.: Uncertainty-aware cnns for depth completion: Uncertainty from beginning to end. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12014-12023 (2020) +9. Ezhov, V., Park, H., Zhang, Z., Upadhyay, R., Zhang, H., Chandrappa, C.C., Kadambi, A., Ba, Y., Dorsey, J., Wong, A.: All-day depth completion. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE (2024) +10. Fei, X., Wong, A., Soatto, S.: Geo-supervised visual depth prediction. IEEE Robotics and Automation Letters 4(2), 1661-1668 (2019) +1. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981) +12. Garg, R., Bg, V.K., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 740-756. Springer (2016) +13. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012) + +14. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 270-279 (2017) +15. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3828-3838 (2019) +16. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2485-2494 (2020) +17. Guizilini, V., Hou, R., Li, J., Ambrus, R., Gaidon, A.: Semantically-guided representation learning for self-supervised monocular depth. In: International Conference on Learning Representations (2019) +18. Hu, M., Wang, S., Li, B., Ning, S., Fan, L., Gong, X.: Penet: Towards precise and efficient image guided depth completion. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13656-13662. IEEE (2021) +19. Jaritz, M., De Charette, R., Wirbel, E., Perrotton, X., Nashashibi, F.: Sparse and dense data with cnns: Depth completion and semantic segmentation. In: 2018 International Conference on 3D Vision (3DV). pp. 52-60. IEEE (2018) +20. Jeon, J., Lim, H., Seo, D.U., Myung, H.: Struct-mdc: Mesh-refined unsupervised depth completion leveraging structural regularities from visual slam. IEEE Robotics and Automation Letters 7(3), 6391-6398 (2022) +21. Kam, J., Kim, J., Kim, S., Park, J., Lee, S.: Costdcnet: Cost volume based depth completion for a single rgb-d image. In: European Conference on Computer Vision. pp. 257-274. Springer (2022) +22. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 582-600. Springer (2020) +23. Krishna, S., Vandrotti, B.S.: Deepsmooth: Efficient and smooth depth completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3357-3366 (2023) +24. Kumar, V.R., Klingner, M., Yogamani, S., Milz, S., Fingscheidt, T., Mader, P.: Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 61-71 (2021) +25. Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., Navab, N.: Deeper depth prediction with fully convolutional residual networks. In: 2016 Fourth international conference on 3D vision (3DV). pp. 239-248. IEEE (2016) +26. Lao, D., Wu, Y., Liu, T.Y., Wong, A., Soatto, S.: Sub-token vit embedding via stochastic resonance transformers. In: International Conference on Machine Learning. PMLR (2024) +27. Lao, D., Yang, F., Wang, D., Park, H., Lu, S., Wong, A., Soatto, S.: On the viability of monocular depth pre-training for semantic segmentation. In: European Conference on Computer Vision. Springer (2024) +28. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision 81(2), 155 (2009) +29. Li, A., Yuan, Z., Ling, Y., Chi, W., Zhang, C., et al.: A multi-scale guided cascade hourglass network for depth completion. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 32-40 (2020) + +30. Li, B., Shen, C., Dai, Y., Van Den Hengel, A., He, M.: Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1119-1127 (2015) +31. Lin, Y., Cheng, T., Zhong, Q., Zhou, W., Yang, H.: Dynamic spatial propagation network for depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1638-1646 (2022) +32. Liu, F., Shen, C., Lin, G., Reid, I.: Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence 38(10), 2024-2039 (2015) +33. Liu, T.Y., Agrawal, P., Chen, A., Hong, B.W., Wong, A.: Monitored distillation for positive congruent depth completion. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II. pp. 35-53. Springer (2022) +34. Lopez-Rodriguez, A., Busam, B., Mikolajczyk, K.: Project to adapt: Domain adaptation for depth completion from noisy and sparse sensor data. In: Proceedings of the Asian Conference on Computer Vision (2020) +35. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hr-depth: High resolution self-supervised monocular depth estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2294-2301 (2021) +36. Ma, F., Cavalheiro, G.V., Karaman, S.: Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In: International Conference on Robotics and Automation (ICRA). pp. 3288-3295. IEEE (2019) +37. Mahjourian, R., Wicke, M., Angelova, A.: Unsupervised learning of depth and egomotion from monocular video using 3d geometric constraints. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5667-5675 (2018) +38. Park, H., Gupta, A., Wong, A.: Test-time adaptation for depth completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20519-20529 (2024) +39. Park, J., Joo, K., Hu, Z., Liu, C.K., Kweon, I.S.: Non-local spatial propagation network for depth completion. In: European Conference on Computer Vision, ECCV 2020. European Conference on Computer Vision (2020) +40. Perez, L., Wang, J.: The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.04621 (2017) +41. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3227-3237 (2020) +42. Poggi, M., Tosi, F., Mattoccia, S.: Learning monocular depth estimation with unsupervised trinocular assumptions. In: 2018 International conference on 3d vision (3DV). pp. 324-333. IEEE (2018) +43. Qiu, J., Cui, Z., Zhang, Y., Zhang, X., Liu, S., Zeng, B., Pollefeys, M.: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3313-3322 (2019) +44. Qu, C., Liu, W., Taylor, C.J.: Bayesian deep basis fitting for depth completion with uncertainty. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16147-16157 (2021) +45. Qu, C., Nguyen, T., Taylor, C.: Depth completion via deep basis fitting. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 71-80 (2020) + +46. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021) +47. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020) +48. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12240-12249 (2019) +49. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. IEEE transactions on pattern analysis and machine intelligence 31(5), 824-840 (2008) +50. Shivakumar, S.S., Nguyen, T., Miller, I.D., Chen, S.W., Kumar, V., Taylor, C.J.: Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 13-20. IEEE (2019) +51. Shorten, C., Khoshgoftaar, T.M.: A survey on image data augmentation for deep learning. Journal of big data 6(1), 1-48 (2019) +52. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: European conference on computer vision. pp. 746-760. Springer (2012) +53. Singh, A.D., Ba, Y., Sarker, A., Zhang, H., Kadambi, A., Soatto, S., Srivastava, M., Wong, A.: Depth estimation from camera image and mmwave radar point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9275-9285 (2023) +54. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) +55. Taylor, L., Nitschke, G.: Improving deep learning with generic data augmentation. In: 2018 IEEE symposium series on computational intelligence (SSCI). pp. 1542-1547. IEEE (2018) +56. Tosi, F., Aleotti, F., Poggi, M., Mattoccia, S.: Learning monocular depth estimation infusing traditional stereo knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9799-9809 (2019) +57. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 2017 international conference on 3D Vision (3DV). pp. 11-20. IEEE (2017) +58. Upadhyay, R., Zhang, H., Ba, Y., Yang, E., Gella, B., Jiang, S., Wong, A., Kadambi, A.: Enhancing diffusion models with 3d perspective geometry constraints. ACM Transactions on Graphics (TOG) 42(6), 1-15 (2023) +59. Van Gansbeke, W., Neven, D., De Brabandere, B., Van Gool, L.: Sparse and noisy lidar completion with rgb guidance and uncertainty. In: 2019 16th International Conference on Machine Vision Applications (MVA). pp. 1-6. IEEE (2019) +60. Wang, C., Buenapasada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2022-2030 (2018) + +61. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2162-2171 (2019) +62. Wong, A., Cicek, S., Soatto, S.: Targeted adversarial perturbations for monocular depth prediction. Advances in neural information processing systems 33, 8486-8497 (2020) +63. Wong, A., Cicek, S., Soatto, S.: Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 1495-1502 (2021) +64. Wong, A., Fei, X., Hong, B.W., Soatto, S.: An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 3120-3127 (2021) +65. Wong, A., Fei, X., Tsuei, S., Soatto, S.: Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters 5(2), 1899-1906 (2020) +66. Wong, A., Soatto, S.: Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5644-5653 (2019) +67. Wong, A., Soatto, S.: Unsupervised depth completion with calibrated backprojection layers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12747-12756 (2021) +68. Xu, D., Ricci, E., Ouyang, W., Wang, X., Sebe, N.: Multi-scale continuous crfs as sequential deep networks for monocular depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5354-5362 (2017) +69. Xu, Y., Zhu, X., Shi, J., Zhang, G., Bao, H., Li, H.: Depth completion from sparse lidar data with depth-normal constraints. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2811-2820 (2019) +70. Yan, Z., Wang, K., Li, X., Zhang, Z., Li, J., Yang, J.: Desnet: Decomposed scale-consistent network for unsupervised depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3109-3117 (2023) +71. Yang, F., Feng, C., Chen, Z., Park, H., Wang, D., Dou, Y., Zeng, Z., Chen, X., Gangopadhyay, R., Owens, A., et al.: Binding touch to everything: Learning unified multimodal tactile representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26340-26353 (2024) +72. Yang, N., Stumberg, L.v., Wang, R., Cremers, D.: D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1281-1292 (2020) +73. Yang, Y., Wong, A., Soatto, S.: Dense depth posterior (ddp) from single image and sparse range. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3353-3362 (2019) +74. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: Every pixel counts: Unsupervised geometry learning with holistic 3d motion understanding. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018) +75. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: Lego: Learning edge with geometry all at once by watching videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 225-234 (2018) +76. Yin, W., Liu, Y., Shen, C., Yan, Y.: Enforcing geometric constraints of virtual normal for depth prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5684-5693 (2019) + +77. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1983-1992 (2018) +78. Yu, Z., Sheng, Z., Zhou, Z., Luo, L., Cao, S.Y., Gu, H., Zhang, H., Shen, H.L.: Aggregating feature point cloud for depth completion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8732-8743 (2023) +79. Zeng, Z., Wang, D., Yang, F., Park, H., Soatto, S., Lao, D., Wong, A.: Wordepth: Variational language prior for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9708-9719 (2024) +80. Zhan, H., Garg, R., Weerasekera, C.S., Li, K., Agarwal, H., Reid, I.: Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 340-349 (2018) +81. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18537-18546 (2023) +82. Zhang, Y., Funkhouser, T.: Deep depth completion of a single rgb-d image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 175-185 (2018) +83. Zhang, Y., Guo, X., Poggi, M., Zhu, Z., Huang, G., Mattoccia, S.: Completionformer: Depth completion with convolutions and vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18527-18536 (2023) +84. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 2022 International Conference on 3D Vision (3DV). pp. 668-678. IEEE (2022) +85. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1851-1858 (2017) +86. Zou, Y., Luo, Z., Huang, J.B.: Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In: Proceedings of the European conference on computer vision (ECCV). pp. 36-53 (2018) \ No newline at end of file diff --git a/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/images.zip b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b7b15a9bee60f70e69b3bc6490d2b138f3b3d7e --- /dev/null +++ b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be8faf9ff8d7434508f705052d401ee39047e89a2562017fd8359fad5fd64340 +size 476026 diff --git a/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/layout.json b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c970ee4cc190e1bd51466a1ab44e4c45a9c3aa5e --- /dev/null +++ b/2024/AugUndo_ Scaling Up Augmentations for Monocular Depth Completion and Estimation/layout.json @@ -0,0 +1,9671 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 468, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 468, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 468, + 148 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Monocular Depth Completion and Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "spans": [ + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "text", + "content": "Yangchao Wu" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "text", + "content": ", Tian Yu Liu" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "text", + "content": ", Hyoungseob Park" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "text", + "content": ", Stefano Soatto" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "text", + "content": ", Dong Lao" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "text", + "content": ", and Alex Wong" + }, + { + "bbox": [ + 186, + 167, + 428, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 189, + 201, + 423, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 201, + 423, + 224 + ], + "spans": [ + { + "bbox": [ + 189, + 201, + 423, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 189, + 201, + 423, + 224 + ], + "type": "text", + "content": " UCLA Vision Lab, Los Angeles, CA 90095, USA {wuyangchao1997, tianyu, soatto, lao}@cs.ucla.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 208, + 224, + 405, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 224, + 405, + 246 + ], + "spans": [ + { + "bbox": [ + 208, + 224, + 405, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 208, + 224, + 405, + 246 + ], + "type": "text", + "content": " Yale Vision Lab, New Haven, CT 06511, USA {hyoungseob.park, alex.wong}@yale.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 274, + 455, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 274, + 455, + 517 + ], + "spans": [ + { + "bbox": [ + 160, + 274, + 455, + 517 + ], + "type": "text", + "content": "Abstract. Unsupervised depth completion and estimation methods are trained by minimizing reconstruction error. Block artifacts from resampling, intensity saturation, and occlusions are amongst the many undesirable by-products of common data augmentation schemes that affect image reconstruction quality, and thus the training signal. Hence, typical augmentations on images viewed as essential to training pipelines in other vision tasks have seen limited use beyond small image intensity changes and flipping. The sparse depth modality in depth completion have seen even less use as intensity transformations alter the scale of the 3D scene, and geometric transformations may decimate the sparse points during resampling. We propose a method that unlocks a wide range of previously-infeasible geometric augmentations for unsupervised depth completion and estimation. This is achieved by reversing, or \"undo\"-ing, geometric transformations to the coordinates of the output depth, warping the depth map back to the original reference frame. This enables computing the reconstruction losses using the original images and sparse depth maps, eliminating the pitfalls of naive loss computation on the augmented inputs and allowing us to scale up augmentations to boost performance. We demonstrate our method on indoor (VOID) and outdoor (KITTI) datasets, where we consistently improve upon recent methods across both datasets as well as generalization to four other datasets. Code available at: https://github.com/alexklwong/augundo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 526, + 432, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 526, + 432, + 537 + ], + "spans": [ + { + "bbox": [ + 160, + 526, + 432, + 537 + ], + "type": "text", + "content": "Keywords: depth completion " + }, + { + "bbox": [ + 160, + 526, + 432, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 526, + 432, + 537 + ], + "type": "text", + "content": " monocular depth " + }, + { + "bbox": [ + 160, + 526, + 432, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 526, + 432, + 537 + ], + "type": "text", + "content": " augmentations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Data augmentation is essential to training machine learning models; it plays a role in performance and generalization [40, 51, 55]. One common axiom of choosing augmentations is that the task output should remain invariant to the augmentation. For example, image flipping is a viable augmentation for classifying animals, since it does not alter the label. Conversely, flipping road signs can alter their meanings; hence, such augmentation can be detrimental to tasks involving road sign recognition. For geometric tasks, the range of augmentations" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": "is more restricted due to constraints of problem formulation: Stereo assumes pairs of frontoparallel rectified images; hence, in-plane rotations are not viable. Image-guided sparse depth completion relies on sparse points to ground estimates to metric scale; therefore, intensity transformations on sparse depth maps that alter the scale of the 3-dimensional (3D) scene are infeasible. Unsupervised learning of depth completion and estimation further limit the use of augmentations as the supervision signal comes from reconstructing the inputs, where augmenting the input introduces artifacts that impact reconstruction quality and therefore the supervision (see Sec. D, Fig. 1, Tab. 8, 9 in the Supp. Mat. for examples of artifacts and extended discussion on their effect on learning). Moreover, video-based unsupervised training assumes rigid motion, so augmentations that introduce padding (e.g., translation, rotation) will yield constant or edge extended borders across images (i.e., no motion), preventing the model from properly learning depth and pose - leaving few augmentations viable. While simulating nuisances is desirable, naively applying augmentations may do more harm than good; thus, it is unsurprising that existing work in unsupervised depth completion [33, 36, 50, 63-65, 67, 73] and estimation [14, 15, 35, 81, 85] primarily rely on a small range of photometric augmentations and flipping." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 337, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 517 + ], + "type": "text", + "content": "Nevertheless, photometric augmentations help model the diverse range of illumination conditions and colors of objects that may populate the scene; geometric augmentations can simulate the various camera parameters, i.e., image resizing (zooming) can model changes in focal length, and scene arrangements, i.e., image flipping. However, block artifacts, loss during resampling, and intensity saturation are just some of the many undesirable side-effects of traditional augmentations to the image and sparse depth map for unsupervised learning of geometric tasks. To avoid compromising the supervision signal, we compute the typical reconstruction loss on the original input image and sparse depth map instead of the augmented inputs, which bypasses negative effects of reconstruction artifacts due to photometric and geometric augmentations. However, there exists a mis-alignment between the original input (e.g., image, sparse depth), and the model depth estimate as geometric augmentations induce a change in coordinates. Hence, we undo the geometric augmentations by inverting them in the output space to align the model estimate with the training target." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Amongst the many geometric tasks, we focus on unsupervised depth completion, the task of inferring a dense depth map from an image and sparse depth map, where augmentations have seen limited use. Here, a training sample includes the input sparse depth map, its associated image and additional images of neighboring views of the same 3D scene. Our method is also applicable to unsupervised monocular depth estimation, which omits the sparse depth modality. Augmentations have traditionally been restricted to a limited range of photometric transformations and flipping - due to the need to preserve photometric consistency across a sequence of video frames used during training, and the sparse set of 3D points projected onto the image frame as a 2.5D range map; degradation to either modalities directly impacts the supervision signal as the loss function is conventionally computed on the augmented inputs. By using our" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 181, + 114, + 430, + 240 + ], + "blocks": [ + { + "bbox": [ + 181, + 114, + 430, + 240 + ], + "lines": [ + { + "bbox": [ + 181, + 114, + 430, + 240 + ], + "spans": [ + { + "bbox": [ + 181, + 114, + 430, + 240 + ], + "type": "image", + "image_path": "83c8538fd14440956fcbe2378d683cabcd091e86b8f88e1b9e7d80338566e3b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 250, + 479, + 305 + ], + "lines": [ + { + "bbox": [ + 130, + 250, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 479, + 305 + ], + "type": "text", + "content": "Fig. 1: Overview. We apply photometric augmentations to the input image, and the same set of geometric augmentations to both the input image and sparse depth map. We warp the output depth back to the original reference frame with the inverse geometric transformations. This enables image and sparse depth reconstruction losses to be computed on the original inputs while unlocking previously-infeasible augmentations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 330, + 479, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 479, + 389 + ], + "type": "text", + "content": "method, loss functions involving sparse depth and image reconstruction from other views can be computed on the original inputs while applying augmentations that were previously not viable for the task. Our hypothesis: By \"undoing\" the augmentations, one can expand the viable set and scale up their use in training, leading to improved model performance and generalizability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 480, + 533 + ], + "type": "text", + "content": "To this end, we introduce AugUndo, an augmentation framework that enables one to apply a wide range of photometric and geometric transformations on the inputs, and to \"undo\" them during loss computation. This allows computing the unsupervised loss on the original images and sparse depth maps, free of artifacts, through a warping of the output depth map - obtained from augmented input onto the input frame of reference based on the inverse geometric transformation. In addition to group transformations that allow for output alignment, we combine them with commonly employed photometric augmentations. To the best of our knowledge, we are the first to propose a unified augmentation scheme for photometric and geometric augmentations for unsupervised depth completion and estimation. We demonstrate AugUndo on recent methods on indoor and outdoor settings, where we consistently improve all methods across all datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "content": "Our contributions are as follows: (1) We propose AugUndo, a simple-yet-effective framework to scale up photometric and geometric augmentations for unsupervised depth completion and estimation, without compromising the supervision signal; AugUndo can be applied in a plug-and-play manner to existing methods with negligible increase in computational costs during training. (2) We enable previously-infeasible augmentations to be used for training unsupervised methods and comprehensively ablate combinations of eleven types of augmentations to study the performance benefits of each. (3) We show that AugUndo can consistently improve robustness to shifts in sparse point densities for completion, model performance as well as zero-shot generalization for both depth completion and estimation for indoor and outdoor scenarios; thus, validating our hypothesis." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 378 + ], + "type": "text", + "content": "Data augmentation for depth completion and estimation. While photometric transformations such as color jitter are often applied for unsupervised depth completion [36,63-65], and estimation [10,14-16,26,27,41,47,66,79,80], geometric augmentations other than simple flipping are less commonly adopted. Most supervised depth completion methods [8,18,19,21,23,29,31,43-45,69,73,82,83] similarly limit their augmentations to color jitter and horizontal flips due to sparse depth maps being decimated by rotation and scaling, causing points to be interpolated away. Nevertheless, for supervised training, it is straightforward to directly apply the same transformation to the ground truth for training. Indeed, some supervised depth completion methods [2,3,9,38,39,53,59,78] adopted random scaling, translation, in-plane rotation. Translation augmentations are also commonly applied in supervised depth estimation methods [6,7,25,30,32, 46,47,58,62,68,71,76]. We posit that the reason that such transformations can be applied to the ground truth in supervised settings is due to artifacts caused by the transformation of a piece-wise smooth depth map being less severe than those of an RGB image and its intensities. However, such assumptions do not hold in unsupervised training. These artifacts would affect the training signal for unsupervised methods, which rely on photometric correspondences and observed sparse points. Our approach aims to resolve this to allow diverse geometric augmentations to be applied in a plug-and-play fashion in unsupervised training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 379, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 617 + ], + "type": "text", + "content": "Unsupervised depth completion methods [36, 50, 63-65, 70, 73] leverage image and sparse depth reconstructions as training signals by minimizing errors between the input image and its reconstruction from other views, and errors between the input sparse depth map and the predicted depth along with a local smoothness regularizer. [36] used Perspective-n-Point [28] and RANSAC [11] to align consecutive video frames. [73] learns a depth prior conditioned on the image. [63] uses synthetic data to learn a prior on the shapes populating a scene, while [34] translates synthetic data to real domain to leverage rendered depth. [64] proposed an adaptive scheme to reduce penalties incurred on occluded regions. [67] maps the image onto the 3D scene through calibrated backprojection. [70] decouples structure and scale. [20] uses line features from visual SLAM and [33] introduced monitored distillation for positive congruent training. Augmentations for these methods are limited to a small range of photometric perturbations and image flipping. Operations such as rotation, resizing, and translation require resampling, which creates artifacts and affects the reconstruction quality. This causes performance degradation since loss is typically computed on the augmented images. Loss in sparse depth maps is further impacted as resampling and interpolation may cause loss of sparse points. Contrary to these limited augmentation schemes, we enable a large range of photometric and geometric augmentations to be introduced during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Unsupervised monocular depth estimation, like depth completion, also minimizes photometric reconstruction error. [12] frames depth estimation as a novel view synthesis problem. [14] improves [12] by imposing a consistency loss on the depth predicted from left and right images. [85] uses a pose network to" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": "enable unsupervised training on video sequences. [15] introduced auto-masking and min reprojection loss. To improve the supervision signal, [4, 56, 61] leverage noisy proxy labels and [42] uses a trinocular assumption. Additional loss terms based on visual odometry [60], iterative closest point [37], surface normals [75], and semantic segmentation [17, 24] were also introduced; [41, 72] further included predictive uncertainty. To handle rigid and non-rigid motions in the scene, previous works explored multi-task learning to include optical flow and moving object estimation [1, 48, 74, 77, 86], and used semantics to filter out outlier regions [22]. [35] redesigned the skip connection and decoders to extract high-resolution features, [84] combined global and local representations and [81] introduced a lightweight architecture with dilated convolution and attention. In addition to depth completion, our method also shows consistent improvements on monocular depth estimation for [15, 35, 81]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 289, + 279, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 279, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 279, + 303 + ], + "type": "text", + "content": "3 Method Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "inline_equation", + "content": "I:\\varOmega\\subset\\mathbb{R}^2\\to\\mathbb{R}_+^3" + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": " be an RGB image captured by a calibrated camera, " + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "inline_equation", + "content": "z:\\varOmega_z\\subset\\varOmega\\rightarrow\\mathbb{R}_+" + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": " the corresponding sparse point cloud projected onto the image plane as a sparse depth map, and " + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "inline_equation", + "content": "K\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": " the intrinsic calibration matrix. Given an image and its sparse depth map, depth completion aims to learn a function " + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "inline_equation", + "content": "f_{\\theta}(I,z)" + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": " that recovers the distance between the camera to points in the 3D scene as a dense depth map. In another mode, if sparse depth maps are not given, then the problem reduces to monocular depth estimation which learns a function " + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "inline_equation", + "content": "f_{\\theta}(I)" + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": " to map a single image to a depth map " + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\varOmega\\rightarrow\\mathbb{R}_{+}" + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": ". For the ease of notation, we denote the output depth map for both depth completion and estimation as " + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\hat{d}\\in \\mathbb{R}_{+}^{H\\times W}" + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 314, + 482, + 434 + ], + "type": "text", + "content": " are its height and width." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": "Unsupervised depth completion relies on photometric and sparse depth reconstruction errors as its primary supervision signal. Without loss of generality, we assume an input of " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "(I_t,z_t)" + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": " for an RGB image and associated sparse depth map captured at time " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": " and during training, an additional set of temporally adjacent images " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "I_{\\tau}" + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\tau \\in \\Upsilon \\doteq \\{t - 1,t + 1\\}" + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": ". The reconstruction " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t\\tau}" + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": " from image " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "I_{\\tau}" + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": " is given by the reprojection based on estimated depth " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\hat{d}_t\\coloneqq f_\\theta (\\cdot)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 246, + 517, + 480, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 517, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 246, + 517, + 480, + 531 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {t \\tau} (x) = I _ {\\tau} \\left(\\pi g _ {\\tau t} K ^ {- 1} \\bar {x} \\hat {d} _ {t} (x)\\right) \\tag {1}", + "image_path": "e3ee054c97c6844685d09f73468527f027a99b0fa97f7eec4f00dad62de3a46e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\bar{x} = [x^{\\top},1]^{\\top}" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " is the homogeneous coordinates of " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "x\\in \\Omega" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "g_{\\tau t}\\in SE(3)" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " the relative pose of the camera from time " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " the intrinsic calibration matrix, and " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " the canonical perspective projection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 575, + 459, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 575, + 459, + 588 + ], + "spans": [ + { + "bbox": [ + 146, + 575, + 459, + 588 + ], + "type": "text", + "content": "Using Eq. (1), a depth completion or estimation network " + }, + { + "bbox": [ + 146, + 575, + 459, + 588 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 146, + 575, + 459, + 588 + ], + "type": "text", + "content": " minimizes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 162, + 597, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 597, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 162, + 597, + 481, + 622 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\theta} \\sum_ {\\tau \\in \\Upsilon} \\sum_ {x \\in \\Omega} \\alpha \\rho (\\hat {I} _ {t \\tau} (x), I _ {t} (x)) + \\sum_ {x \\in \\Omega_ {z}} \\beta \\psi (\\hat {d} _ {t} (x), z _ {t} (x)) + \\lambda R (\\hat {d} _ {t}) \\tag {2}", + "image_path": "4366841c98864d66a571f61eaa06650e4cabaed1cdcfe1e7e864f8c0815be5ea.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " denotes the photometric reconstruction error, typically " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " difference in pixel values and/or structural similarity (SSIM), " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " the sparse depth reconstruction error, typically " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " the regularizer that biases the depth map to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "be piece-wise smooth with depth discontinuities aligned with edges in the image (commonly used by previous works [15,36,65,67]), and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " their respective weighting. Note: monocular depth omits sparse depth term, i.e., " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": "Inverse transformation. Let " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{pt}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " be the set of possible photometric transformations, and " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{ge}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " be the set of all geometric transformations. Given " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "T_{ge} \\in \\mathcal{A}_{ge}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": ", we wish to obtain an inverse transformation " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "T_{ge}^{-1}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "T_{ge} \\circ T_{ge}^{-1} \\approx Id" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " the identity function3. At each time step, we can sample a sequence of transformations " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "\\{T_{pt}^{1} \\ldots T_{pt}^{k}\\}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "\\{T_{ge}^{1} \\ldots T_{ge}^{m}\\}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " respectively from " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{pt}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{ge}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " to construct transformations " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "T_{pt} = T_{pt}^{1} \\circ \\dots \\circ T_{pt}^{k}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "T_{ge} = T_{ge}^{1} \\circ \\dots \\circ T_{ge}^{m}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": ". We denote the composition of a collection of augmentation transformations by " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "T = T_{ge} \\circ T_{pt}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "T_{pt}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " denotes photometric transformation, and " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "T_{ge}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": " denotes geometric transformation. Furthermore, we denote the inverse geometric transformations by " + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "inline_equation", + "content": "T_{ge}^{-1} = (T_{ge}^{m})^{-1} \\circ (T_{ge}^{m-1})^{-1} \\circ \\dots \\circ (T_{ge}^{1})^{-1}" + }, + { + "bbox": [ + 130, + 152, + 480, + 356 + ], + "type": "text", + "content": ", which operates on the space of depth maps to reverse the geometric transformation so that we can warp the output depth map onto the reference frame of the original image. Through inverse warping, this process is differentiable. In practice, some transformations cause border regions of the image to be warped out of frame, i.e., translated off the image plane or cropped away. Hence, after warping our output depth map back to the original frame of reference using the inverse geometric transformations, border extensions (edge paddings) are introduced to handle out-of-frame regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 357, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 479, + 380 + ], + "type": "text", + "content": "AugUndo. We apply each geometric transformation over image coordinates and resample (bilinear for image and nearest neighbor for sparse depth):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 233, + 388, + 480, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 388, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 233, + 388, + 480, + 404 + ], + "type": "interline_equation", + "content": "\\left[ x ^ {\\prime} 1 \\right] ^ {\\top} = T _ {g e} \\left[ x 1 \\right] ^ {\\top} \\tag {3}", + "image_path": "eeb7661feaf752782e3185690ef9e2e7ef5b6bbc61084b5318e024f815030944.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 240, + 406, + 480, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 406, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 240, + 406, + 480, + 419 + ], + "type": "interline_equation", + "content": "I ^ {\\prime} \\left(x ^ {\\prime}\\right) = T _ {p t} (I) (x); z ^ {\\prime} \\left(x ^ {\\prime}\\right) = z (x) \\tag {4}", + "image_path": "a3f8238e4e719d264308e9b4b13bdfe4c9ece0d021c679d908d40c09f486a937.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "inline_equation", + "content": "T_{ge}" + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": " is the geometric transformation, " + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "inline_equation", + "content": "x \\in \\Omega" + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "inline_equation", + "content": "x' \\in \\Omega" + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": " are coordinates in the image grid, and " + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": " is the image after the transformation; for ease of notation, we hereafter denote " + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "inline_equation", + "content": "I' = T(I) = T_{ge} \\circ T_{pt}(I)" + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": " to include both augmentations through composition. Note that " + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": " is in the original image reference frame and " + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": " is in the transformed image reference frame. Naturally, this process can be extended to multiple geometric transformations by composing them, i.e., " + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "inline_equation", + "content": "T_{ge} = T_{ge}^{1} \\circ T_{ge}^{2} \\circ \\dots \\circ T_{ge}^{m}" + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": ". The reverse process is simply inverting the transformations where " + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "inline_equation", + "content": "T_{ge}^{-1} = (T_{ge}^{m})^{-1} \\circ (T_{ge}^{m-1})^{-1} \\circ \\dots \\circ (T_{ge}^{1})^{-1}" + }, + { + "bbox": [ + 131, + 426, + 479, + 525 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 259, + 533, + 480, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 533, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 259, + 533, + 480, + 548 + ], + "type": "interline_equation", + "content": "\\left[ x ^ {\\prime \\prime} 1 \\right] ^ {\\top} = T _ {g e} ^ {- 1} \\left[ x ^ {\\prime} 1 \\right] ^ {\\top} \\tag {5}", + "image_path": "d7d388cfb7953e67a1e92eed55337de1160542b0d316aa18a7caadbd5eec0e38.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 268, + 552, + 480, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 552, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 268, + 552, + 480, + 565 + ], + "type": "interline_equation", + "content": "\\hat {d} \\left(x ^ {\\prime \\prime}\\right) = \\hat {d} ^ {\\prime} \\left(x ^ {\\prime}\\right) \\tag {6}", + "image_path": "edb407feb736ada53d08d5233085aa74442a18fbfe7043db49d27a81e3759dff.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "inline_equation", + "content": "\\hat{d}'" + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "text", + "content": " is the depth map inferred from augmented inputs " + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "inline_equation", + "content": "(I', z')" + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "text", + "content": ". Once reverted back to the original reference frame, " + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "inline_equation", + "content": "\\hat{d}" + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "text", + "content": " is used to reconstruct " + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "inline_equation", + "content": "I_\\tau" + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "inline_equation", + "content": "\\tau \\in \\Upsilon" + }, + { + "bbox": [ + 130, + 574, + 479, + 611 + ], + "type": "text", + "content": " as in Eq. (1). More details can be found in Alg. 1 in the Supp. Mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 612, + 479, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 479, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 479, + 636 + ], + "type": "text", + "content": "By modeling " + }, + { + "bbox": [ + 130, + 612, + 479, + 636 + ], + "type": "inline_equation", + "content": "T_{ge}^{-1}" + }, + { + "bbox": [ + 130, + 612, + 479, + 636 + ], + "type": "text", + "content": ", Eq. (3)-Eq. (6) allow us to apply a wide range of augmentations, while still establishing the correspondence between " + }, + { + "bbox": [ + 130, + 612, + 479, + 636 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 612, + 479, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 612, + 479, + 636 + ], + "type": "inline_equation", + "content": "I_{\\tau}" + }, + { + "bbox": [ + 130, + 612, + 479, + 636 + ], + "type": "text", + "content": ". Specifically," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "text", + "content": "3 Note that in practice, not all transformations are bijective. For instance, image rescaling with a fixed size image plane; hence strict equality is not always possible." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "text", + "content": "for minimizing Eq. (2), one may simply augment the image and sparse depth by " + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "text", + "content": " to yield " + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "inline_equation", + "content": "(I_t', z_t')" + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "text", + "content": " as input, and reconstruct the original image and sparse depth " + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "inline_equation", + "content": "(I_t, z_t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "text", + "content": " from other views " + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "inline_equation", + "content": "I_\\tau" + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "text", + "content": " and the aligned output depth " + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "inline_equation", + "content": "\\hat{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "text", + "content": " (Eq. (6)). We note that the inverse transformation is critical for enabling the sparse depth reconstruction term to be computed properly in Eq. (2); if computed in the transformed reference frame, i.e., on " + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "inline_equation", + "content": "z_t'" + }, + { + "bbox": [ + 130, + 116, + 482, + 250 + ], + "type": "text", + "content": ", many of the sparse points would be decimated by interpolation during rotation and resizing (downsampling) augmentations - leaving a lack of supervision from the sparse depth term of the loss function. We note that the original image sequence is fed to the pose estimator, rather than the augmented images. This ensures the estimated relative pose used during reprojection (Eq. (1)) also corresponds to the original images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 370 + ], + "type": "text", + "content": "Modeling geometric augmentations. Depth completion differs from typical image-based (e.g., monocular) depth estimation, where operations like resizing are susceptible to scale ambiguity and can either correspond to changes in focal length, or distance from the camera (i.e., due to changes in camera pose). In depth completion, the sparse depth map grounds image pixels to specific depth values. Hence, there is no ambiguity in distance of each sparse point from the camera. Our goal is to synthesize new training data via geometric transformations that are consistent with the original 3D scene. Given that there is no scale ambiguity, this can be achieved by either (1) changes in camera pose (extrinsics), or (2) changes in camera parameters (intrinsics)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 373, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 456 + ], + "type": "text", + "content": "Augmentation via (1) would require warping the inputs for view synthesis. Since we operate under the unsupervised setting, accurate warping (preserving 3D scene) without access to dense ground truth depth values is difficult to achieve. An exception to this is rotation, which we address below. On the other hand, (2) can be achieved via standard 2D image transformations while preserving the 3D structure of the scene. This motivates our modeling of geometric transformations, i.e., resizing, translation, as changes in camera parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 458, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 567 + ], + "type": "text", + "content": "For example, augmenting focal length (i.e. zooming in/out) is akin to \"resizing\". Similarly, translation can be used to model shifting of the optic center, i.e. capturing the same 3D scene using cameras with different principal point offsets. Note: image rotation can be seen as rotating a camera about its optical axis, which does not change the distance of a point from the camera. Thus, the camera position and the scene are kept constant across augmentations, eliminating the need for adjusting depth values via warping during the process. Instead, we only have to realign the output depth back to the original frame of reference (Eqs. (5) and (6)), which makes our method computationally efficient." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 568, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 604 + ], + "type": "text", + "content": "Augmentations. Our choices are based on common nuisance variabilities, i.e., changes in illumination, occlusions, and object color, scene layout (flip), and camera parameters (resize, translation) and orientation (rotation)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Photometric. We include brightness, contrast, saturation, and hue, where all follow standard augmentation pipeline in existing works [15,65,67,81]. Applying the inverse of photometric augmentation can be viewed as recovering the original image; hence, we directly use original image instead of applying transformations to the intensities (which are not recoverable if saturated at the max value)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "Occlusion. We consider patch removal and sparse point removal. For the former, we randomly select a percentage of pixels " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "x \\in \\Omega" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " in the image and remove (with zero-fill) an arbitrary-sized patch around it. For the latter, we randomly sample a percentage of points " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "x \\in \\Omega_z" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " in the sparse depth map. The inverse transformation of this is simply the original image and sparse depth map, both used in loss computation before augmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 236 + ], + "type": "text", + "content": "Flip. We consider horizontal and vertical flips. When applied, the same flip operation is used for both input image and sparse depth maps. We record the flip type during data augmentation. During loss computation, we reverse the flip direction to align output depth with the original image and sparse depth map." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 366 + ], + "type": "text", + "content": "Resize. We define a new image plane of the input resolution and generate a random scaling factor to be applied along both height and width directions. The image is warped to the new image plane, where any point warped out of the plane is excluded; borders of the warped image are extended to the bounds of the image plane by edge replication. Sparse points occupying multiple pixels are eroded to a single point for resizing with a factor greater than 1. During loss computation, we warp the output depth map onto a new image plane of the same dimensions by the inverse scaling matrix of the recorded scaling factors; borders of the warped depth map are extended by edge replication if necessary. We view resizing factors greater and less than 1 (zooming in and out) as two separate forms of augmentations to distinguish their contributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 481, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 487 + ], + "type": "text", + "content": "Rotation. Naive rotation leads to the loss of large areas of the image, i.e., cropped away to retain the same-size image, discarding large portions of possible co-visible regions and also supervisory signals. To preserve the entire image, we first warp the image by a randomly generated angle to a new (larger) image plane, so that the rotated image fits tightly within the new image. As image sizes within a batch can vary depending on the rotation angle, we center-pad (uniformly on each side) each image in the batch to the maximum width and height of the augmented batch. To reverse the rotation on the output depth, we warp the output depth map back by the inverse rotation matrix, then perform a center crop on the depth map to align with the original input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 488, + 481, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 481, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 481, + 583 + ], + "type": "text", + "content": "Translation. We define a new image plane of the same dimensions as the input and generate random height and width translation factors. The coordinates of the input are translated and its pixels are inverse warped onto the new image plane. Any pixel warped out of the image plane is excluded. Borders of the warped image are extended to the bounds of the image plane by edge replication. During loss computation, we warp the output depth map onto a new image plane of the same dimensions by the inverse translation matrix. Borders of the warped depth map are extended to the bounds of the image plane by edge replication." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 603, + 230, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 230, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 230, + 616 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "text", + "content": "We demonstrate AugUndo on recent unsupervised depth completion (VOICED [65], FusionNet [63], KBNet [67]) and estimation (Monodepth2 [15], HRDepth [35], LiteMono [81]) methods on two datasets (KITTI [13, 57], Void [65]) here." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": "Due to space limitations, we defer the results of MonDi [33] and DesNet [70] to Sec. K in the Supp. Mat. To evaluate zero-shot generalization, we test on three additional datasets for each task: NYUv2 [52], ScanNet [5], Waymo [54], and Make3d [49] - where we transfer models on Void to NYUv2 (Tab. 3) and ScanNet, and from KITTI to Waymo (Tab. 4 in Supp. Mat.) for completion and Make3d [49] (Tab. 5 in Supp. Mat.) for estimation. We also test model sensitivity to different sparse depth input densities in Tab. 1 (right) - see Sec. H of the Supp. Mat for an extensive study. Additionally, we present an comprehensive ablation study in Sec. G in the Supp. Mat. to test the effect of each augmentation. We also provide results on modeling AugUndo as changes in camera motion or parameters in Tab. 7 of the Supp. Mat.; interpreting geometric augmentations as different camera parameters yields better performance. Tables. 8, 9 in Supp. Mat. show that unsupervised models cannot be trained with naive geometric augmentations. See Sec. E of the Supp. Mat. for descriptions of datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 287, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 372 + ], + "type": "text", + "content": "Experiment setup. We followed the original settings of the open-sourced code for each work and modified the data handling and loss function to incorporate AugUndo. We perform 4 independent trials for each experiment and report their means and standard deviations. To ensure a fair comparison, we train all models from scratch. Below, we report the best combination of augmentations found empirically through extensive experiments on each dataset. See Secs. A and B in Supp. Mat. for implementation details, and evaluation metrics, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "text", + "content": "Augmentations. Through a search over augmentation types and values, we found a consistent set that tends to yield improvements across all methods with small changes to degree of augmentation catered to each method. See Sec. C in Supp. Mat. for details of augmentation parameters. Rows in Tabs. 1 to 6 marked with \"+\" AugUndo\" denote models trained with our method. We note that performance gain can be obtained by typical set and ranges of augmentations and does not require a meticulous selection of hyper-parameters (see Sec. J of Supp. Mat. for a sensitivity study). Results reported here aims to study how far one can push performance and generalization, purely from augmentations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Results on VOID. We begin by presenting quantitative results for AugUndo on unsupervised depth completion. Tab. 1 (left, VOID1500) shows our main results on the VOID benchmark. By training the models with AugUndo, we observe an average overall improvement of " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " across all methods and metrics on VOID1500. Specifically, we improve VOICED by " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "26.4\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", FusionNet by " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", and KBNet by " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "12.2\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". These experiments validate our hypothesis that by applying a wider range of augmentations, we are able to improve the baseline model's performance. They also illustrate the lack in use of augmentations in existing works: incorporating standard augmentations (albeit requiring modification to augmentation and loss computation pipelines) can yield a large performance boost. Fig. 2 shows a head-to-head comparison between KBNet trained using standard procedure in [67] and KBNet trained using AugUndo. We observe qualitative improvements from AugUndo where we improve KBNet in homogeneous regions and image borders, i.e., pillar (left), cabinet (middle), wall (right). Applying geometric augmentations yields models with fewer border" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 169, + 481, + 255 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: Depth completion on Void. AugUndo improves performance by an average of " + }, + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "type": "text", + "content": " across all methods and metrics on Void1500. When models trained on Void1500 are tested Void500, AugUndo improves by " + }, + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "23.1\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "type": "text", + "content": ", as translation, resizing, and occlusion augmentations vary sparsity by removing sparse points from the input." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 169, + 481, + 255 + ], + "lines": [ + { + "bbox": [ + 135, + 169, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 135, + 169, + 481, + 255 + ], + "type": "table", + "html": "
MethodVOID1500VOID500
MAE ↓RMSE ↓iMAE ↓iRMSE ↓MAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED74.78±2.69139.75±4.5739.20±1.4671.98±2.54137.01±4.23235.80±7.8271.36±1.86130.63±5.66
+ AugUndo52.73±0.41111.09±0.9226.93±0.5454.46±0.3892.99±1.11176.94±1.3846.43±0.8591.10±1.64
FusionNet52.11±0.44113.30±1.1828.53±0.5258.79±2.0197.73±0.73194.32±1.3658.65±1.31122.95±3.04
+ AugUndo41.16±0.1899.21±0.3922.23±0.3553.07±1.3074.97±0.69162.71±2.3940.44±1.3992.11±5.79
KBNet38.11±0.7795.22±1.7219.51±0.1446.70±0.4878.44±1.39178.17±3.2737.56±0.6183.43±1.89
+ AugUndo33.32±0.1885.67±0.3916.61±0.2941.24±0.6066.97±0.81151.55±2.0331.63±0.5371.90±0.82
", + "image_path": "f62ad62a1427835dca122bd750d2c99ee3b70cd8a06dd04334d2da87f8abc09f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 142, + 269, + 473, + 395 + ], + "blocks": [ + { + "bbox": [ + 142, + 269, + 473, + 395 + ], + "lines": [ + { + "bbox": [ + 142, + 269, + 473, + 395 + ], + "spans": [ + { + "bbox": [ + 142, + 269, + 473, + 395 + ], + "type": "image", + "image_path": "80be05bc519fb2d33fb4934d9933db4e29ba94e2eb280e65aa1a27a78ce9b7e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 403, + 482, + 449 + ], + "lines": [ + { + "bbox": [ + 132, + 403, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 482, + 449 + ], + "type": "text", + "content": "Fig. 2: Depth completion on VOID. We compare KBNet trained with standard augmentations and with AugUndo. AugUndo consistently produces lower errors with reduced border artifacts and improves on homogeneous regions, i.e., pillar (left), cabinet (middle), wall of staircase (right). Error maps are highlighted for comparison." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 616 + ], + "type": "text", + "content": "artifacts as we apply random translation, which warps part of the image out of frame. This allows us to simulate border occlusion, where there lacks correspondence in an adjacent frame. While training with standard protocol results in failures to recover structures near the image border, training with AugUndo can render models robust to them by computing the loss on the original frame of reference (where we do have correspondence) through our inverse transformations of the depth map. Additionally, resizing allows the model to learn multiple resolutions of the input, akin to zooming in and out, which can also impose smoothness in homogeneous regions through the scale-space transition; whereas, rotation can simulate diverse camera orientations. Fig. 2 shows that translation, rotation, and resizing can model these effect in the input space to yield models robust to these nuisance variability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "AugUndo also is applicable for monocular depth estimation. Tab. 2 shows a comparison using the standard augmentation procedure of Monodepth2, HR-Depth, and Lite-Mono and using AugUndo. Tab. 2 shows that AugUndo consistently improves all models across all error and accuracy metrics. Thus, validating" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 148, + 478, + 223 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Monocular depth estimation on VOID. AugUndo is applicable for monocular depth estimation and consistently improves three monocular depth models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 148, + 478, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 148, + 478, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 148, + 478, + 223 + ], + "type": "table", + "html": "
MethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.252 ↑δ < 1.253 ↑
Monodepth2283.861±3.732395.947±5.7280.183±0.0020.100±0.0030.717±0.0050.922±0.0040.975±0.002
+ AugUndo277.696±4.861388.088±5.7680.178±0.0030.095±0.0040.724±0.0070.925±0.0040.978±0.002
HR-Depth286.282±7.059399.112±9.1840.185±0.0040.100±0.0040.714±0.0120.919±0.0060.975±0.002
+ AugUndo283.086±6.787394.261±9.1330.181±0.0050.097±0.0050.718±0.0130.922±0.0040.977±0.002
Lite-Mono319.910±15.00446.005±22.970.209±0.0130.129±0.0190.669±0.0160.892±0.0110.963±0.006
+ AugUndo308.010±0.859426.626±0.4840.200±0.0030.114±0.0010.674±0.0050.901±0.0020.969±0.002
", + "image_path": "c9044bedbcfc894358f8c46ae9b8a3fb4344dc65c02606238c139d1f55971422.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 480, + 324 + ], + "type": "text", + "content": "the hypothesis that AugUndo can be applied generically to improve monocular depth estimation. Specifically, we observe a boost in the most difficult accuracy metric " + }, + { + "bbox": [ + 130, + 253, + 480, + 324 + ], + "type": "inline_equation", + "content": "(\\delta < 1.25)" + }, + { + "bbox": [ + 130, + 253, + 480, + 324 + ], + "type": "text", + "content": ", where Monodepth2 improves from 0.717 to 0.724, HR-Depth from 0.714 to 0.718 and Lite-Mono from 0.669 to 0.674. At the same time, for AbsRel metric, Monodepth2 improves from 0.183 to 0.178, HR-depth improves from 0.185 to 0.181, and Lite-Mono improves from 0.209 to 0.200." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "text", + "content": "Sensitivity study. One limitation of existing depth completion training pipeline is that there are little to no augmentations applied to sparse depth modality. However, in real-world applications, sparse depth has, in fact, high variability, i.e., features tracked in SLAM/VIO systems will vary depending on the scene (points are dropped or added to the state), and point cloud densities returned by a sensor will differ based on specifications. To further examine the effect of AugUndo on sparse depth, we study the sensitivity to changes in sparse depth density by testing models on VOID1500 (" + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "text", + "content": "1500 points) on VOID500 (" + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "text", + "content": "500). For the " + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "text", + "content": " reduction in sparse points, AugUndo improves robustness by an average of " + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "inline_equation", + "content": "23.1\\%" + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "text", + "content": " across all methods. These improvements result from the geometric and occlusion augmentations made possible via AugUndo, which greatly increases sparse depth variations, i.e., decimating them through resizing, reorienting their configuration through rotation, translating them out of frame, and randomly occluding them, to avoid overfitting particular sparse depth configurations. We further push their limits in Sec. H of the Supp. Mat., where we test them on VOID150 with a " + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 130, + 327, + 482, + 543 + ], + "type": "text", + "content": " reduction in sparse points from the training set (VOID1500) and observe the same trend of improvements. This shows that AugUndo significantly improves robustness of models to changes in sparse depth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Zero-shot generalization. We test depth completion models trained on VOID on NYUv2 and ScanNet. Tab. 3 shows an average of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "23.2\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " improvement on NYUv2 and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "27.6\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " on ScanNet. Applying AugUndo to VOICED greatly improve its generalizability to both NYUv2 and ScanNet. This is likely due to the scaffolding densification employed by VOICED, where the network can overfit to foldings of particular sparse depth configurations and therefore, does not generalize well when presented with sparse depth maps with different configurations. Like our sensitivity study (Fig. 2, VOID500), occlusion and geometric augmentations introduce variation into the sparse depth configurations and densities, which alleviates overfitting to specific point clouds or 3D scenes; hence," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 159, + 478, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Zero-shot transfer from VOID to NYUv2 and ScanNet for depth completion. AugUndo improves generalization of models trained on VOID to novel datasets by an average of " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "25.4\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " for all evaluation metrics (in millimeters) across both datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 159, + 478, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 159, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 159, + 478, + 232 + ], + "type": "table", + "html": "
MethodNYUv2ScanNet
MAE ↓RMSE ↓iMAE ↓iRMSE ↓MAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED2240±143.902427±143.49211±9.60238±10.891562±136.791764±146.39270±17.25311±17.36
+ AugUndo990±82.481181±65.55110±6.11132±6.01638±44.74791±79.48131±7.68170±6.32
FusionNet132.24±2.12236.16±4.5928.68±0.4261.87±1.20109.47±3.01206.33±6.1155.45±1.56122.52±2.04
+ AugUndo124.93±3.05227.23±4.9625.70±0.4154.09±0.87100.64±2.31195.85±5.8545.98±0.7899.90±5.00
KBNet138.31±5.74257.99±10.3625.48±0.6351.77±0.99103.05±4.99217.12±13.3536.23±1.1276.55±2.90
+ AugUndo118.60±3.44231.13±8.8522.06±0.3147.07±0.7082.53±4.33175.30±11.1329.87±1.0663.78±1.30
", + "image_path": "918fd9242c5bf27726d4ec0939ba28293eb7aa4d14fe4697edf36847e9a1b6a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 140, + 282, + 471, + 395 + ], + "blocks": [ + { + "bbox": [ + 132, + 248, + 479, + 270 + ], + "lines": [ + { + "bbox": [ + 132, + 248, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 479, + 270 + ], + "type": "text", + "content": "Table 4: Zero-shot transfer from VOID to NYUv2 and ScanNet for depth estimation. AugUndo consistently improves generalization (in meters) for monocular depth models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 282, + 471, + 395 + ], + "lines": [ + { + "bbox": [ + 140, + 282, + 471, + 395 + ], + "spans": [ + { + "bbox": [ + 140, + 282, + 471, + 395 + ], + "type": "table", + "html": "
DatasetMethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.25² ↑δ < 1.25³ ↑
NYUv2Monodepth20.432±0.0030.556±0.0050.205±0.0010.159±0.0010.683±0.0050.907±0.0010.975±0.001
+ AugUndo0.415±0.0050.537±0.0060.196±0.0020.148±0.0030.700±0.0080.915±0.0020.977±0.001
HR-Depth0.424±0.0030.549±0.0030.201±0.0020.154±0.0020.692±0.0030.910±0.0020.976±0.001
+ AugUndo0.421±0.0090.542±0.0110.199±0.0050.152±0.0060.696±0.0090.913±0.0040.976±0.001
Lite-Mono0.480±0.0120.616±0.0180.231±0.0080.199±0.0150.637±0.0090.879±0.0090.964±0.004
+ AugUndo0.468±0.0030.595±0.0030.225±0.0040.187±0.0050.646±0.0020.889±0.0030.968±0.001
ScanNetMonodepth20.284±0.0040.368±0.0050.177±0.0020.097±0.0020.741±0.0060.931±0.0030.980±0.001
+ AugUndo0.270±0.0030.351±0.0040.169±0.0010.088±0.0020.759±0.0040.937±0.0010.982±0.001
HR-Depth0.282±0.0040.366±0.0050.175±0.0020.095±0.0020.743±0.0050.929±0.0030.979±0.002
+ AugUndo0.274±0.0070.357±0.0090.172±0.0050.092±0.0050.754±0.0090.935±0.0030.981±0.001
Lite-Mono0.296±0.0020.388±0.0070.185±0.0030.109±0.0060.731±0.0020.921±0.0020.976±0.001
+ AugUndo0.296±0.0010.382±0.0010.185±0.0010.105±0.0020.728±0.0020.924±0.0010.977±0.001
", + "image_path": "6566b079c467874bbcf71b39e9d696ef012be4ef4b67463d877a83a3107f1b01.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "text", + "content": "improving VOICED from " + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "inline_equation", + "content": "49.9\\%" + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "inline_equation", + "content": "52.6\\%" + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "text", + "content": " on NYUv2 and ScanNet, respectively. For FusionNet and KBNet, we still see nontrivial improvement: FusionNet improves by " + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "inline_equation", + "content": "12.1\\%" + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "text", + "content": " on NYUv2 and ScanNet, and KBNet by " + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 130, + 422, + 479, + 482 + ], + "type": "text", + "content": ". This further validates our hypothesis that by applying a more diverse set of transformations, we are able to improve generalization to new datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 616 + ], + "type": "text", + "content": "For depth estimation, we tested Monodepth2, HR-Depth, Lite-Mono (trained on VOID) for zero-shot generalizability to NYUv2 and ScanNet. The results are shown in Tab. 4, where training with AugUndo consistently improves generalization errors across all evaluation metrics for both NYUv2 and ScanNet. Notably, for RMSE metric, Monodepth2 improves from 0.556 to 0.537, HR-Depth from 0.549 to 0.542, Lite-Mono from 0.616 to 0.595. For ScanNet, similar improvement on RMSE can also be observed, where Monodepth2 improves from 0.368 to 0.351, HR-Depth from 0.366 to 0.357, and Lite-Mono from 0.388 to 0.382. The improvement in RMSE metric, which is sensitive to outliers, highlights AugUndo's ability to model different input data distributions, collected by a different camera of different orientation with different object colors and layouts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Results on KITTI. We begin with quantitative results for depth completion. While AugUndo consistently improves all methods (Tab. 5), we note that the improvement is less, but respectable, in this case: " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\approx 5.2\\%" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " overall with the largest gain in FusionNet of " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "6.41\\%" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ". This is largely due to the small scene varia" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 212, + 148, + 399, + 219 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 5: Depth completion on KITTI. AugUndo consistently improves all methods for all metrics (in millimeters) by approximately " + }, + { + "bbox": [ + 131, + 114, + 482, + 138 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 131, + 114, + 482, + 138 + ], + "type": "text", + "content": " on average." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 148, + 399, + 219 + ], + "lines": [ + { + "bbox": [ + 212, + 148, + 399, + 219 + ], + "spans": [ + { + "bbox": [ + 212, + 148, + 399, + 219 + ], + "type": "table", + "html": "
MethodMAE ↓RMSE ↓iMAE ↓iRMSE ↓
VOICED318.59±7.741,213.60±17.491.30±0.053.72±0.04
+ AugUndo295.41±0.301,159.27±5.441.20±0.023.49±0.03
FusionNet285.55±2.161,174.47±10.671.20±0.033.45±0.08
+ AugUndo267.69±1.851,157.07±4.611.08±0.023.19±0.03
KBNet263.90±3.631,130.66±6.221.05±0.013.24±0.04
+ AugUndo256.37±1.001,114.53±3.791.01±0.013.13±0.03
", + "image_path": "34ee28e45e1e111a811318a35d542e375a3c2d75a1fa4c1d2dbea29ebbdea5b2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 150, + 234, + 465, + 308 + ], + "blocks": [ + { + "bbox": [ + 150, + 234, + 465, + 308 + ], + "lines": [ + { + "bbox": [ + 150, + 234, + 465, + 308 + ], + "spans": [ + { + "bbox": [ + 150, + 234, + 465, + 308 + ], + "type": "image", + "image_path": "b7613559793937b55ff355bff53e46aa37439847fb190cf5e49589c5b6fab1a6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 315, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 361 + ], + "type": "text", + "content": "Fig. 3: Depth completion on KITTI. We compare KBNet and FusionNet trained with standard augmentations and with AugUndo. AugUndo consistently produces lower errors in highlighted regions where structures may have arbitrary orientation (vegetation) and regions near image border that typically lacks correspondence during training." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 568 + ], + "type": "text", + "content": "tions in the outdoor driving scenarios, i.e., ground plane with vehicles, buildings on the sides, horizontal lidar scans, and largely planar motion. The dataset bias is strong enough to render vertical flip detrimental to performance. This is also evident in existing works as none utilizes vertical flip as augmentation. Nonetheless, AugUndo still improves performance, where point removal models different lidar returns patterns, and resizing simulates large variations in scales of objects observed in road scenes. Similar to indoors, translations help model occlusions by shifting the projection with different principal points. Fig. 3 compares FusionNet and KBNet trained using conventional augmentations and AugUndo. AugUndo improves over objects that may have diverse orientations (i.e., tree branches and vegetation), thanks to rotation augmentations. Improvements are also observed in \"small\" objects at far regions where resizing can simulate zooming in/out to emulate objects of different sizes. Through translation and occlusion, AugUndo also improves on occlusion boundaries during training (highlighted), making estimates near image borders and object boundaries more robust." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": "We additionally show results for monocular depth estimation on KITTI in Tab. 6, where we compare the standard augmentation pipelines used by [15,35, 81] and AugUndo. We observe similar trends in performance gain as we did in depth completion trained on KITTI: Applying our set of augmentations improves most metrics for all methods. We observe notable improvements in " + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": ", the most difficult accuracy metric, from 0.869 to 0.879 in Monodepth2, from 0.879 to 0.883 in HR-Depth, and 0.862 to 0.863 in Lite-Mono. For Monodepth2, we also observe a large reduction in AbsRel error, improving it from 0.117 to 0.113." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 447, + 103 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 148, + 463, + 218 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 6: Monocular depth estimation on KITTI. AugUndo consistently improve on evaluation metrics (in meters) across different models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 148, + 463, + 218 + ], + "lines": [ + { + "bbox": [ + 147, + 148, + 463, + 218 + ], + "spans": [ + { + "bbox": [ + 147, + 148, + 463, + 218 + ], + "type": "table", + "html": "
MethodMAE ↓RMSE ↓AbsRel ↓SqRel ↓δ < 1.25 ↑δ < 1.252↑δ < 1.253↑
Monodepth2 + AugUndo2.315±0.0054.794±0.0350.117±0.0010.845±0.0300.869±0.0040.959±0.0010.982±0.001
2.237±0.0144.739±0.0320.113±0.0000.862±0.0300.879±0.0020.960±0.0010.982±0.001
HR-Depth +AugUndo2.226±0.0044.626±0.0320.113±0.0010.797±0.0220.879±0.0020.961±0.0010.982±0.000
2.185±0.0134.610±0.0290.111±0.0010.794±0.0210.883±0.0010.962±0.0010.983±0.001
Lite-Mono + AugUndo2.338±0.0054.821±0.0270.121±0.0010.875±0.0070.862±0.0020.955±0.0010.980±0.000
2.314±0.0304.780±0.0550.120±0.0010.849±0.0190.863±0.0040.955±0.0010.981±0.001
", + "image_path": "4cb5693e85de13c761bae8d356bc13f091333e5f9c3f029456ced9d1b2d0115f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 242, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 479, + 265 + ], + "type": "text", + "content": "Similarly, for Lite-Mono, we also boosted performance across all metrics, with a particularly high reduction in SqRel from 0.875 to 0.849." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 350 + ], + "type": "text", + "content": "Remarks. As AugUndo only augments the inputs and modifies the loss computations, which add negligible time to training, we note that the performance improvements from it are obtained nearly for free. Yet, the percentage gain, however, is similar to improvements obtained by each successive state-of-the-art: For example, the improvement of Lite-Mono over HR-Depth, and that of HR-Depth over PackNet for monocular depth estimation; similarly, the improvement of FusionNet over VOICED and KBNet over FusionNet for depth completion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 366, + 217, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 217, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 217, + 379 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 390, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 480, + 510 + ], + "type": "text", + "content": "Conventionally, data augmentation aims to seek visual invariance and create a collection of equivalent classes, i.e., identifying an image and its augmented variant as the same. For geometric tasks, the underlying equivalence is in the 3D scene structures under various illumination conditions, camera, viewpoints, occlusion, etc. Assuming a rigid scene, the shapes populating it should persist regardless of the nuisance variables. This motivates the use of geometric augmentations, as it simulates the nuisances within the image. However, adoption of geometric augmentations in unsupervised geometric tasks are obstructed by artifacts introduced during transformations (see Sec. D of Supp. Mat.). AugUndo lifts this obstacle by \"undo-ing\" the augmentations before computing the loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "content": "While AugUndo enables scaling up augmentations for unsupervised training, i.e., depth completion, it may also be applicable for supervised methods; though, we posit that the gain to be less as the artifacts induced from photometric and geometric augmentations of an image tend to be larger than those of a piecewise smooth depth map. AugUndo is also limited to 2D augmentations; whereas, nuisances modeled by it are projections of the 3D scene (see limitations in Sec. M in Supp. Mat.). We leave extensions to 3D for future work. We also only consider a single image and sparse depth map as input. Likewise, extensions can be made towards multi-frame tasks such as stereo, optical flow, pose estimation, etc., but one must account for their specifics and problem setups, i.e., stereo assumes frontoparallel views. This is outside of our scope, so we leave them as future directions. This work paves way for the empirical success in unsupervised geometric tasks that we have observed in other visual recognition tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by was supported by NSF 2112562 Athena AI Institute." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 199, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 199, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 199, + 174 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 188, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 188, + 481, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 481, + 232 + ], + "type": "text", + "content": "1. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7063-7072 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 233, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 481, + 277 + ], + "type": "text", + "content": "2. Cheng, X., Wang, P., Guan, C., Yang, R.: Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10615-10622 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "text", + "content": "3. Cheng, X., Wang, P., Yang, R.: Depth estimation via affinity learned with convolutional spatial propagation network. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 103-119 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 481, + 343 + ], + "type": "text", + "content": "4. Choi, H., Lee, H., Kim, S., Kim, S., Kim, S., Sohn, K., Min, D.: Adaptive confidence thresholding for monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12808-12818 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 344, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 344, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 481, + 376 + ], + "type": "text", + "content": "5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 377, + 481, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 377, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 377, + 481, + 410 + ], + "type": "text", + "content": "6. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE international conference on computer vision. pp. 2650-2658 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 411, + 481, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 443 + ], + "type": "text", + "content": "7. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 444, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 481, + 487 + ], + "type": "text", + "content": "8. Eldesokey, A., Felsberg, M., Holmquist, K., Persson, M.: Uncertainty-aware cnns for depth completion: Uncertainty from beginning to end. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12014-12023 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 487, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 532 + ], + "type": "text", + "content": "9. Ezhov, V., Park, H., Zhang, Z., Upadhyay, R., Zhang, H., Chandrappa, C.C., Kadambi, A., Ba, Y., Dorsey, J., Wong, A.: All-day depth completion. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 533, + 481, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 554 + ], + "type": "text", + "content": "10. Fei, X., Wong, A., Soatto, S.: Geo-supervised visual depth prediction. IEEE Robotics and Automation Letters 4(2), 1661-1668 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "1. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 588, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 481, + 632 + ], + "type": "text", + "content": "12. Garg, R., Bg, V.K., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 740-756. Springer (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "13. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 270-279 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "15. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3828-3838 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "type": "text", + "content": "16. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2485-2494 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "text", + "content": "17. Guizilini, V., Hou, R., Li, J., Ambrus, R., Gaidon, A.: Semantically-guided representation learning for self-supervised monocular depth. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "text", + "content": "18. Hu, M., Wang, S., Li, B., Ning, S., Fan, L., Gong, X.: Penet: Towards precise and efficient image guided depth completion. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13656-13662. IEEE (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "text", + "content": "19. Jaritz, M., De Charette, R., Wirbel, E., Perrotton, X., Nashashibi, F.: Sparse and dense data with cnns: Depth completion and semantic segmentation. In: 2018 International Conference on 3D Vision (3DV). pp. 52-60. IEEE (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "text", + "content": "20. Jeon, J., Lim, H., Seo, D.U., Myung, H.: Struct-mdc: Mesh-refined unsupervised depth completion leveraging structural regularities from visual slam. IEEE Robotics and Automation Letters 7(3), 6391-6398 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "type": "text", + "content": "21. Kam, J., Kim, J., Kim, S., Park, J., Lee, S.: Costdcnet: Cost volume based depth completion for a single rgb-d image. In: European Conference on Computer Vision. pp. 257-274. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 430 + ], + "type": "text", + "content": "22. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 582-600. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 463 + ], + "type": "text", + "content": "23. Krishna, S., Vandrotti, B.S.: Deepsmooth: Efficient and smooth depth completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3357-3366 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 508 + ], + "type": "text", + "content": "24. Kumar, V.R., Klingner, M., Yogamani, S., Milz, S., Fingscheidt, T., Mader, P.: Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 61-71 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 480, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 541 + ], + "type": "text", + "content": "25. Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., Navab, N.: Deeper depth prediction with fully convolutional residual networks. In: 2016 Fourth international conference on 3D vision (3DV). pp. 239-248. IEEE (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 542, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 480, + 575 + ], + "type": "text", + "content": "26. Lao, D., Wu, Y., Liu, T.Y., Wong, A., Soatto, S.: Sub-token vit embedding via stochastic resonance transformers. In: International Conference on Machine Learning. PMLR (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "text", + "content": "27. Lao, D., Yang, F., Wang, D., Park, H., Lu, S., Wong, A., Soatto, S.: On the viability of monocular depth pre-training for semantic segmentation. In: European Conference on Computer Vision. Springer (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "text", + "content": "28. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision 81(2), 155 (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "29. Li, A., Yuan, Z., Ling, Y., Chi, W., Zhang, C., et al.: A multi-scale guided cascade hourglass network for depth completion. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 32-40 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "text", + "content": "30. Li, B., Shen, C., Dai, Y., Van Den Hengel, A., He, M.: Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1119-1127 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "type": "text", + "content": "31. Lin, Y., Cheng, T., Zhong, Q., Zhou, W., Yang, H.: Dynamic spatial propagation network for depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1638-1646 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "text", + "content": "32. Liu, F., Shen, C., Lin, G., Reid, I.: Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence 38(10), 2024-2039 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 267 + ], + "type": "text", + "content": "33. Liu, T.Y., Agrawal, P., Chen, A., Hong, B.W., Wong, A.: Monitored distillation for positive congruent depth completion. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II. pp. 35-53. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "text", + "content": "34. Lopez-Rodriguez, A., Busam, B., Mikolajczyk, K.: Project to adapt: Domain adaptation for depth completion from noisy and sparse sensor data. In: Proceedings of the Asian Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "35. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hr-depth: High resolution self-supervised monocular depth estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2294-2301 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "36. Ma, F., Cavalheiro, G.V., Karaman, S.: Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In: International Conference on Robotics and Automation (ICRA). pp. 3288-3295. IEEE (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "text", + "content": "37. Mahjourian, R., Wicke, M., Angelova, A.: Unsupervised learning of depth and egomotion from monocular video using 3d geometric constraints. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5667-5675 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "38. Park, H., Gupta, A., Wong, A.: Test-time adaptation for depth completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20519-20529 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "39. Park, J., Joo, K., Hu, Z., Liu, C.K., Kweon, I.S.: Non-local spatial propagation network for depth completion. In: European Conference on Computer Vision, ECCV 2020. European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "text", + "content": "40. Perez, L., Wang, J.: The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.04621 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "41. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3227-3237 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "42. Poggi, M., Tosi, F., Mattoccia, S.: Learning monocular depth estimation with unsupervised trinocular assumptions. In: 2018 International conference on 3d vision (3DV). pp. 324-333. IEEE (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "text", + "content": "43. Qiu, J., Cui, Z., Zhang, Y., Zhang, X., Liu, S., Zeng, B., Pollefeys, M.: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3313-3322 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "44. Qu, C., Liu, W., Taylor, C.J.: Bayesian deep basis fitting for depth completion with uncertainty. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16147-16157 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "45. Qu, C., Nguyen, T., Taylor, C.: Depth completion via deep basis fitting. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 71-80 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "46. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 194 + ], + "type": "text", + "content": "47. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 239 + ], + "type": "text", + "content": "48. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12240-12249 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 240, + 480, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 272 + ], + "type": "text", + "content": "49. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. IEEE transactions on pattern analysis and machine intelligence 31(5), 824-840 (2008)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 317 + ], + "type": "text", + "content": "50. Shivakumar, S.S., Nguyen, T., Miller, I.D., Chen, S.W., Kumar, V., Taylor, C.J.: Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 13-20. IEEE (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 319, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 340 + ], + "type": "text", + "content": "51. Shorten, C., Khoshgoftaar, T.M.: A survey on image data augmentation for deep learning. Journal of big data 6(1), 1-48 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "text", + "content": "52. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: European conference on computer vision. pp. 746-760. Springer (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 418 + ], + "type": "text", + "content": "53. Singh, A.D., Ba, Y., Sarker, A., Zhang, H., Kadambi, A., Soatto, S., Srivastava, M., Wong, A.: Depth estimation from camera image and mmwave radar point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9275-9285 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 463 + ], + "type": "text", + "content": "54. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 480, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 496 + ], + "type": "text", + "content": "55. Taylor, L., Nitschke, G.: Improving deep learning with generic data augmentation. In: 2018 IEEE symposium series on computational intelligence (SSCI). pp. 1542-1547. IEEE (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "text", + "content": "56. Tosi, F., Aleotti, F., Poggi, M., Mattoccia, S.: Learning monocular depth estimation infusing traditional stereo knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9799-9809 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "text", + "content": "57. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 2017 international conference on 3D Vision (3DV). pp. 11-20. IEEE (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "text", + "content": "58. Upadhyay, R., Zhang, H., Ba, Y., Yang, E., Gella, B., Jiang, S., Wong, A., Kadambi, A.: Enhancing diffusion models with 3d perspective geometry constraints. ACM Transactions on Graphics (TOG) 42(6), 1-15 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "59. Van Gansbeke, W., Neven, D., De Brabandere, B., Van Gool, L.: Sparse and noisy lidar completion with rgb guidance and uncertainty. In: 2019 16th International Conference on Machine Vision Applications (MVA). pp. 1-6. IEEE (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "60. Wang, C., Buenapasada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2022-2030 (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "61. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2162-2171 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "text", + "content": "62. Wong, A., Cicek, S., Soatto, S.: Targeted adversarial perturbations for monocular depth prediction. Advances in neural information processing systems 33, 8486-8497 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "text", + "content": "63. Wong, A., Cicek, S., Soatto, S.: Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 1495-1502 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "64. Wong, A., Fei, X., Hong, B.W., Soatto, S.: An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 3120-3127 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "text", + "content": "65. Wong, A., Fei, X., Tsuei, S., Soatto, S.: Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters 5(2), 1899-1906 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "66. Wong, A., Soatto, S.: Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5644-5653 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "67. Wong, A., Soatto, S.: Unsupervised depth completion with calibrated backprojection layers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12747-12756 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 385 + ], + "type": "text", + "content": "68. Xu, D., Ricci, E., Ouyang, W., Wang, X., Sebe, N.: Multi-scale continuous crfs as sequential deep networks for monocular depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5354-5362 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "text", + "content": "69. Xu, Y., Zhu, X., Shi, J., Zhang, G., Bao, H., Li, H.: Depth completion from sparse lidar data with depth-normal constraints. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2811-2820 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "text", + "content": "70. Yan, Z., Wang, K., Li, X., Zhang, Z., Li, J., Yang, J.: Desnet: Decomposed scale-consistent network for unsupervised depth completion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3109-3117 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "text", + "content": "71. Yang, F., Feng, C., Chen, Z., Park, H., Wang, D., Dou, Y., Zeng, Z., Chen, X., Gangopadhyay, R., Owens, A., et al.: Binding touch to everything: Learning unified multimodal tactile representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26340-26353 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 531 + ], + "type": "text", + "content": "72. Yang, N., Stumberg, L.v., Wang, R., Cremers, D.: D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1281-1292 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "type": "text", + "content": "73. Yang, Y., Wong, A., Soatto, S.: Dense depth posterior (ddp) from single image and sparse range. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3353-3362 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "text", + "content": "74. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: Every pixel counts: Unsupervised geometry learning with holistic 3d motion understanding. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "75. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: Lego: Learning edge with geometry all at once by watching videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 225-234 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "76. Yin, W., Liu, Y., Shen, C., Yan, Y.: Enforcing geometric constraints of virtual normal for depth prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5684-5693 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "text", + "content": "AugUndo: Scaling Up Augmentations for Depth Completion & Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 501 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "77. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1983-1992 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "78. Yu, Z., Sheng, Z., Zhou, Z., Luo, L., Cao, S.Y., Gu, H., Zhang, H., Shen, H.L.: Aggregating feature point cloud for depth completion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8732-8743 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "text", + "content": "79. Zeng, Z., Wang, D., Yang, F., Park, H., Soatto, S., Lao, D., Wong, A.: Wordepth: Variational language prior for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9708-9719 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "80. Zhan, H., Garg, R., Weerasekera, C.S., Li, K., Agarwal, H., Reid, I.: Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 340-349 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "type": "text", + "content": "81. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18537-18546 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "82. Zhang, Y., Funkhouser, T.: Deep depth completion of a single rgb-d image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 175-185 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "text", + "content": "83. Zhang, Y., Guo, X., Poggi, M., Zhu, Z., Huang, G., Mattoccia, S.: Completionformer: Depth completion with convolutions and vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18527-18536 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 391, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 434 + ], + "type": "text", + "content": "84. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 2022 International Conference on 3D Vision (3DV). pp. 668-678. IEEE (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 434, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 480, + 468 + ], + "type": "text", + "content": "85. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1851-1858 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "86. Zou, Y., Luo, Z., Huang, J.B.: Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In: Proceedings of the European conference on computer vision (ECCV). pp. 36-53 (2018)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_content_list.json b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7af0e6bf559a33f2931606ec1b3606bc9d0c5322 --- /dev/null +++ b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_content_list.json @@ -0,0 +1,1698 @@ +[ + { + "type": "text", + "text": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", + "text_level": 1, + "bbox": [ + 125, + 48, + 871, + 104 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nazmul Karim\\*,1, Abdullah Al Arafat\\*,2, Umar Khalid\\* Zhishan Guo\\* and Nazanin Rahnavard\\*", + "bbox": [ + 79, + 139, + 915, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Central Florida 2North Carolina State University", + "bbox": [ + 171, + 193, + 823, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent studies have revealed the vulnerability of deep neural networks (DNNs) to various backdoor attacks, where the behavior of DNNs can be compromised by utilizing certain types of triggers or poisoning mechanisms. State-of-the-art (SOTA) defenses employ too-sophisticated mechanisms that require either a computationally expensive adversarial search module for reverse-engineering the trigger distribution or an over-sensitive hyper-parameter selection module. Moreover, they offer sub-par performance in challenging scenarios, e.g., limited validation data and strong attacks. In this paper, we propose—Neural mask Fine-Tuning (NFT)—with an aim to optimally re-organize the neuron activities in a way that the effect of the backdoor is removed. Utilizing a simple data augmentation like MixUp, NFT relaxes the trigger synthesis process and eliminates the requirement of the adversarial search module. Our study further reveals that direct weight fine-tuning under limited validation data results in poor post-purification clean test accuracy, primarily due to overfitting issue. To overcome this, we propose to fine-tune neural masks instead of model weights. In addition, a mask regularizer has been devised to further mitigate the model drift during the purification process. The distinct characteristics of NFT render it highly efficient in both runtime and sample usage, as it can remove the backdoor even when a single sample is available from each class. We validate the effectiveness of NFT through extensive experiments covering the tasks of image classification, object detection, video action recognition, 3D point cloud, and natural language processing. We evaluate our method against 14 different attacks (LIRA, WaNet, etc.) on 11 benchmark data sets (ImageNet, UCF101, Pascal VOC, ModelNet, OpenSubtitles2012, etc.). Our code is available online in this GitHub Repository.", + "bbox": [ + 147, + 261, + 849, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 79, + 778, + 309, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning and computer vision algorithms are increasingly common in safety-critical applications [3, 20], necessitating the design of secure and robust learning algorithms. Backdoor attack [9,17] on deep neural network (DNN) models is one of the heavily studied branches of AI safety and robustness. Backdoor defenses can generally be categorized into two major groups based on whether", + "bbox": [ + 79, + 819, + 917, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution", + "bbox": [ + 84, + 930, + 292, + 947 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the defense is done during training or test. Training time defense (e.g., [25,27,49]) focuses on training a benign model on the poisonous data, while test time defense (e.g., [6,29,34,54]) deals with purifying backdoor model after it has already been trained. In this work, our goal is to develop an efficient test-time defense. Some test-time defenses focus on synthesizing trigger patterns [7,51,52] followed by vanilla weight fine-tuning. Most of these defenses aim to synthesize class-specific triggers independently or use additional models to generate the triggers simultaneously. Recent state-of-the-art (SOTA) backdoor defense methods, e.g., ANP [54], I-BAU [59], AWM [6], employ very similar techniques and do not work well without an expensive adversarial search module. For example, ANP [54] performs an adversarial search to find vulnerable neurons responsible for backdoor behavior. Identifying and pruning vulnerable neurons require an exhaustive adversarial search, resulting in high computational costs. Similar to ANP, AWM and I-BAU also resort to trigger synthesizing with a modified adversarial search process. Another very recent technique, FT-SAM [66] uses sharpness-aware minimization (SAM) [15] to fine-tune model weights. SAM is a recently proposed optimizer that utilizes Stochastic Gradient Descent (SGD), which penalizes sudden changes in the loss surface by constraining the search area to a compact region. Since SAM performs a double forward pass to compute the loss gradient twice, it results in a notable runtime increase for FT-SAM. In our work, we aim to develop an effective backdoor defense system that neither requires an expensive adversarial search process to recover the trigger nor a special type of optimizer with a runtime bottleneck.", + "bbox": [ + 75, + 55, + 922, + 502 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this goal, we propose a simple yet effective approach Neural mask Fine-Tuning (NFT), to remove backdoor through augmented fine-tuning of cost-efficient neural masks. We start with replacing the expensive adversarial search-based trigger synthesis process with a simple data augmentation technique—MixUp [60]. In general, the backdoor is inserted by forcing the model to memorize the trigger distribution. Intuitively, synthesizing and unlearning that trigger distribution would effectively remove the backdoor. In this work, we show that unlearning can be performed by simply optimizing the MixUp loss over a clean validation set. Our theoretical analysis suggests that MixUp loss is an upper bound on the standard loss obtained from triggered (synthesized or already known) validation data, termed as ideal purification loss. As the minimization of ideal loss guarantees backdoor purification, minimizing the MixUp loss would effectively remove the backdoor (Sec. 4.1). As the next step of our method, we address the overfitting issue during weight fine-tuning under limited validation data. In general, the outcome of such overfitting is poor post-purification test accuracy, which is not desirable for any backdoor defense. To this end, we propose to fine-tune a set of neural masks instead of the model weights, as this type of soft-masking enables us to reprogram the neurons affected by the backdoor without significantly altering the original backdoor model. As an added step to this, a mask regularizer has been introduced to further mitigate model drift during the purification process. In addition, we deploy a mask scheduling function to have better control over the purification process. Our experimental", + "bbox": [ + 75, + 521, + 922, + 949 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 79, + 16, + 94, + 30 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "N. Karim et al.", + "bbox": [ + 159, + 14, + 309, + 30 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "results indicate that these straightforward yet intuitive steps significantly improve the post-purification test accuracy as compared to previous SOTA. Our contributions can be summarized as follows:", + "bbox": [ + 77, + 53, + 920, + 113 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose a novel backdoor removal framework utilizing simple MixUp-based model fine-tuning. Our thorough analysis shows how minimizing the MixUp loss eliminates the requirement of an expensive trigger synthesis process while effectively removing the backdoor (Sec. 4.1).", + "bbox": [ + 89, + 123, + 917, + 201 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- To preserve the post-purification test accuracy, we propose to fine-tune soft neural masks (instead of weights) as it prevents any drastic change in the original backdoor model (Sec. 4.2). Additionally, a novel mask regularizer has been introduced that further encourages the purified model to retain the class separability of the original model. In addition to being computationally efficient, our proposed method shows significant improvement in sample efficiency as it can purify backdoor even with one-shot fine-tuning, i.e., only a single sample is available from each class (Sec. 4.3).", + "bbox": [ + 89, + 201, + 920, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- To show the effectiveness of NFT, we perform an extensive evaluation with 11 different datasets. Compared to previous SOTA, the superior performance against a wide range of attacks suggests that augmentation like MixUp can indeed replace the trigger synthesis process (Sec. 5).", + "bbox": [ + 89, + 357, + 920, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 77, + 461, + 331, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backdoor Attack. Neural networks are intrinsically vulnerable to backdoor attacks [38, 57]. A substantial number of studies have investigated the possibility of backdoor attacks after the initial studies [9, 17, 36] found the existence of backdoors in DNNs. Generally, backdoor attacks are categorized into two types: clean-label attacks and poison-label attacks. A clean-label backdoor attack does not alter the label [41, 50, 64], while a poison-label attack aims at specific target classes such that the DNN misclassifies to those classes in the presence of a trigger [28]. As for trigger types, researchers have studied numerous types of triggering patterns in their respective attacks [9, 13, 17, 30]. Such triggers can exist in the form of dynamic patterns [30] or as simple as a single pixel [49]. Some of the more complex backdoor triggers that have been proposed in the literature are sinusoidal strips [2], adversarial patterns [64], and blending backgrounds [9]. Besides, backdoor attacks exist for many different tasks, e.g., multi-label clean image attack [8] has been proposed that alters the label distribution to insert triggers into the model, which works well in multi-label (e.g., detection) settings; domain adaptation [1] setting while adversary source can successfully insert backdoor to the target domains, etc.", + "bbox": [ + 77, + 500, + 920, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backdoor Defense. Generally, the backdoor defense methods are categorized into two types: Training Time Defense, and Test Time Defense. Regarding training time defense techniques (a few to mention [16, 19, 36, 49]), the researchers have proposed numerous defense methods through input pre-processing [36], poison-suppression [19], model diagnosis [30], network pruning [36, 55], and model reconstruction [63], etc. Notably, DeepSweep [43] explores different augmentations to", + "bbox": [ + 77, + 832, + 920, + 949 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "purify a backdoor model and rectify the triggered samples. Although DeepSweep revealed that different augmentation functions could be leveraged to invert the backdoor effect of a model or erase the trigger from a trigger-embedded image, our work re-purposes the usage of augmentation differently to cover the approximate (unknown) trigger distribution during the purification phase. Moreover, DeepSweep assumes that backdoor triggers are known to the defender, which is hardly a practical assumption. In the case of test time defenses, besides the works mentioned in the introduction related to reverse-engineering of backdoor triggers in the input samples [7, 51, 52], several recent works explored the model vulnerability/sensitivity towards adversarially perturbed neurons [54], weights [6], or network channels [65]. However, these approaches require expensive adversarial search processes to be effective. A concurrent work FIP [21] studied the loss-surface smoothness of the backdoor model and developed a purification method by regularizing the spectral norm of the model.", + "bbox": [ + 77, + 53, + 922, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Threat Model", + "text_level": 1, + "bbox": [ + 79, + 357, + 329, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attack Model. Our work considers the most commonly used data poisoning attacks. Consider $\\mathbb{D}_{\\mathrm{train}} = \\{x_i, y_i\\}_{i=1}^N$ as the training data where $x_i \\in \\mathbb{R}^d$ is an input sample labeled as $y_i \\in \\{0, \\dots, c-1\\}$ sampled from unknown distribution $\\mathcal{D}$ of the task to be learned. Here, $N$ is the total number of samples, and $d$ is the dimension of the input sample. In addition, we assume there are $c$ number of classes in the input data. Let $f_{\\theta^*}: \\mathbb{R}^d \\to \\mathbb{R}^c$ be a benign (ideal) DNN trained with $\\mathbb{D}_{\\mathrm{train}} \\sim \\mathcal{D}$ . Here, $\\theta^*$ is the DNN parameters that is to be optimized using a suitable loss function $\\ell(.,.)$ . The total empirical loss can be defined as,", + "bbox": [ + 77, + 400, + 922, + 557 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\theta^ {*}, \\mathbb {D} _ {\\text {t r a i n}}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ \\ell \\left(y _ {i}, f _ {\\theta^ {*}} \\left(x _ {i}\\right)\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 575, + 915, + 624 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, consider an adversary inserts backdoor to a model $f_{\\theta}(.)$ through modifying a small subset of $\\mathbb{D}_{\\mathrm{train}}$ as $\\{\\hat{x}_i,\\hat{y}_i\\}$ such that $\\hat{y}_i = \\arg \\max f_\\theta (\\hat{x}_i)$ preserving $y_{i} = \\arg \\max f_{\\theta}(x_{i}),\\forall (x_{i},y_{i})\\in \\mathbb{D}_{\\mathrm{train}}.$ Here, $\\hat{x}_i = x_i + \\delta$ is the triggered input with adversary set target label $\\hat{y}_i\\neq y_i$ , where $\\delta \\in \\mathbb{R}^d$ represents trigger pattern. Defense Objective. Consider a defense model where defender removes backdoor from $f_{\\theta}(.)$ using a small validation data $\\mathbb{D}_{\\mathrm{val}} = \\{x_i,y_i\\}_{i = 1}^{N_{val}}$ such that $y_{i} = \\arg \\max f_{\\theta_{c}}(\\hat{x}_{i})$ , where $y_{i}\\neq \\hat{y}_{i}$ .", + "bbox": [ + 77, + 639, + 922, + 778 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Neural Fine-Tuning (NFT)", + "text_level": 1, + "bbox": [ + 79, + 806, + 513, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us consider a fully-connected DNN, $f_{\\theta}:\\mathbb{R}^{d}\\to \\mathbb{R}^{c}$ , that receives a datapoint $x\\in \\mathbb{R}^d$ and predicts a probability distribution $p\\in \\mathbb{R}^c$ ; where $c$ is the number of classes. In general, $x$ goes through a multi-layer DNN architecture before the DNN model predicts an output class $i = \\arg \\max p$ . Let us consider a multi-layer DNN architecture with $L$ layers in which the $l$ -th layer contains $k_{l}$ neurons.", + "bbox": [ + 77, + 850, + 922, + 947 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 79, + 16, + 94, + 29 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "N. Karim et al.", + "bbox": [ + 159, + 14, + 309, + 30 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The neurons of each layer produce activations $\\psi_{l}\\in \\mathbb{R}^{k_{l}}$ based on the output activations of previous layer $\\psi_{l - 1}\\in \\mathbb{R}^{k_{l - 1}}$ . To be specific,", + "bbox": [ + 75, + 55, + 917, + 95 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {l} := \\sigma \\left(\\Theta_ {l} ^ {T} \\cdot \\psi_ {l - 1} + b _ {l}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 105, + 917, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma(.)$ is a non-linear activation function, matrix $\\Theta_l = [\\Theta_l^{(1)}\\dots \\Theta_l^{(k_l)}]\\in \\mathbb{R}^{k_{l - 1}\\times k_l}$ , for $l = 1,2,\\ldots ,L$ , includes the weights of the $l$ -th layer, and $b_{l}\\in \\mathbb{R}^{k_{l}}$ is the bias vector. Here, $\\Theta_l^{(j)}\\in \\mathbb{R}^{k_{l - 1}}$ denotes the weights vector corresponding to the activation of the $j$ -th neuron of the $l$ -th layer. Model parameters $\\theta$ can be expressed as $\\theta = \\{\\Theta_1,\\dots ,\\Theta_L\\}$ . This type of multi-layer DNN architecture is also valid for convolutional neural networks, where we use multiple 2-D arrays of neurons (i.e., filters) instead of a 1-D array.", + "bbox": [ + 72, + 136, + 922, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Backdoor Suppressor", + "text_level": 1, + "bbox": [ + 75, + 304, + 404, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our objective is to make the backdoor model forget about poison distribution while retaining the knowledge of clean distribution. To understand how we achieve this objective, let us first revisit the process of generating triggered data $(\\hat{x})$ . In general, $\\hat{x}$ is created by adding minor modifications (i.e., adding triggers $\\delta$ ) to clean data $(x)$ . Note that the backdoor is inserted by forcing the model to learn the mapping, $\\hat{x} \\rightarrow \\hat{y}$ . Here, $(\\hat{x}, \\hat{y})$ is the poison data. If we were to change the mapping from $(\\hat{x} \\rightarrow \\hat{y})$ to $(\\hat{x} \\rightarrow y)$ , we would have a robust clean model instead of a backdoor model. This is because, in this case, the model would treat $\\delta$ as one type of augmentation and $\\hat{x}$ as the augmented clean data. In summary, the backdoor insertion process functions as an augmentation process if we simply use $y$ instead of $\\hat{y}$ . Now, ideally, fine-tuning the backdoor model with triggered data with corresponding ground truth labels (i.e., $\\{\\hat{x}, y\\}$ ) would remove the backdoor effect from the model. Let us consider this ideal scenario where we have access to the trigger $\\delta$ during purification. We define the ideal purification loss as:", + "bbox": [ + 72, + 331, + 922, + 624 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {i d e a l}} (\\theta , \\mathbb {D} _ {\\text {v a l}}) = \\frac {1}{N _ {\\text {v a l}}} \\sum_ {i = 1} ^ {N _ {\\text {v a l}}} \\ell \\left(y _ {i}, f _ {\\theta} \\left(\\hat {x} _ {i}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 637, + 917, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{x} = x + \\delta$ and $y$ is the ground truth label of $x$ . In our work, as we do not have access to $\\delta$ or adversarially reverse engineer it [6,54,59], we relax this process by strongly augmenting the clean validation data, i.e., creating augmented $\\mathbb{D}_{\\mathrm{val}}$ with already known augmentation technique such as MixUp [60]. For MixUp, we can easily perform $\\tilde{x}_{i,j} = \\lambda x_i + (1 - \\lambda)x_j$ and $\\tilde{y}_{i,j} = \\lambda y_i + (1 - \\lambda)y_j$ for $\\lambda \\in [0,1]$ ; here $\\tilde{y}_{i,j}$ represents the linear combination of one-hot vectors corresponding to $y_i$ and $y_j$ . The loss after the MixUp becomes:", + "bbox": [ + 72, + 695, + 922, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {m i x}} (\\theta , \\mathbb {D} _ {\\text {v a l}}) = \\frac {1}{N _ {\\text {v a l}} ^ {2}} \\sum_ {i, j = 1} ^ {N _ {\\text {v a l}}} \\underset {\\lambda \\sim \\mathcal {D} _ {\\lambda}} {\\mathbb {E}} \\ell \\left(\\tilde {y} _ {i, j}, f _ {\\theta} \\left(\\tilde {x} _ {i, j}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 845, + 917, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{\\lambda}$ is a distribution supported on $[0,1]$ . In our work, we consider the widely used $\\mathcal{D}_{\\lambda}$ - Beta distribution $Beta(\\alpha, \\beta)$ for $\\alpha, \\beta > 0$ . We provide both", + "bbox": [ + 75, + 908, + 922, + 949 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", + "bbox": [ + 179, + 14, + 837, + 34 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "empirical (Sec. 5) and theoretical proof for a binary classification problem on why minimizing Eq. (4) would effectively remove the backdoor.", + "bbox": [ + 72, + 55, + 917, + 95 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theoretical Justifications. For a fully-connected neural network (NN) with logistic loss $\\ell(y, f_{\\theta}(x)) = \\log(1 + \\exp(f_{\\theta}(x))) - yf_{\\theta}(x)$ with $y \\in \\{0, 1\\}$ , it can be shown that $\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})$ is an upper-bound of the second order Taylor expansion of the ideal loss $\\mathcal{L}^{\\mathrm{ideal}}(\\theta, \\mathbb{D}_{\\mathrm{val}})$ . With the nonlinearity $\\sigma$ for ReLU and max-pooling in NN, the function $f_{\\theta}$ satisfies that $f_{\\theta}(x) = \\nabla f_{\\theta}(x)^T x$ and $\\nabla^2 f_{\\theta}(x) = 0$ almost everywhere, where the gradient is taken with respect to the input $x$ .", + "bbox": [ + 72, + 95, + 922, + 211 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first rewrite the $\\mathcal{L}^{\\mathrm{ideal}}(\\theta, \\mathbb{D}_{\\mathrm{val}})$ using Taylor series approximation. The second-order Taylor expansion of $\\ell(y, f_{\\theta}(x + \\delta))$ is given by,", + "bbox": [ + 72, + 211, + 922, + 250 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\ell (y, f _ {\\theta} (x + \\delta)) = \\ell (y, f _ {\\theta} (x)) + (g (f _ {\\theta} (x)) - y) (f _ {\\theta} (\\delta)) + \\frac {1}{2} g (f _ {\\theta} (x)) (1 - g (f _ {\\theta} (x))) (f _ {\\theta} (\\delta)) ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 75, + 261, + 937, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $g(x) = \\frac{e^x}{1 + e^x}$ is the logistic function. Based on the MixUp related analysis in prior works [5, 61], the following can be derived for $\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})$ using the second-order Taylor series expansion,", + "bbox": [ + 72, + 308, + 922, + 370 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 1. Assuming $f_{\\theta}(x) = \\nabla f_{\\theta}(x)^{T}x$ and $\\nabla^2 f_{\\theta}(x) = 0$ (which are satisfied by ReLU and max-pooling activation functions), $\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})$ can be expressed as,", + "bbox": [ + 75, + 378, + 922, + 435 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {m i x}} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) = \\mathcal {L} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) + \\mathcal {R} _ {1} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) + \\mathcal {R} _ {2} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 435, + 917, + 461 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where,", + "bbox": [ + 77, + 465, + 150, + 482 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {R} _ {1} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\geq \\frac {R c _ {x} \\mathbb {E} _ {\\lambda} [ (1 - \\lambda) ] \\sqrt {d}}{N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} | g (f _ {\\theta} (x _ {i})) - y _ {i} | \\cdot | | \\nabla f _ {\\theta} (x _ {i}) | | _ {2} \\\\ \\mathcal {R} _ {2} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\geq \\frac {R ^ {2} c _ {x} ^ {2} \\mathbb {E} _ {\\lambda} [ (1 - \\lambda) ] ^ {2} d}{2 N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} | g (f _ {\\theta} (x _ {i})) (1 - g (f _ {\\theta} (x _ {i}))) | \\cdot | | \\nabla f _ {\\theta} (x _ {i}) | | _ {2} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 92, + 494, + 903, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $R = \\min_{i\\in [N_{\\mathrm{val}}]}\\langle \\nabla f_{\\theta}(x_i),x_i\\rangle /||\\nabla f_{\\theta}(x_i)||\\cdot ||x_i||$ and $c_{x} > 0$ is a constant.", + "bbox": [ + 77, + 617, + 905, + 639 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By comparing $\\ell(y, f_{\\theta}(x + \\delta))$ and $\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})$ for a fully connected NN, we can prove the following.", + "bbox": [ + 72, + 645, + 917, + 682 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 1. Suppose that $f_{\\theta}(x) = \\nabla f_{\\theta}(x)^{T}x$ , $\\nabla^2 f_{\\theta}(x) = 0$ and there exists a constant $c_{x} > 0$ such that $\\| x_{i}\\|_{2}\\geq c_{x}\\sqrt{d}$ for all $i\\in \\{1,\\ldots ,N_{\\mathrm{val}}\\}$ . Then, for any $f_{\\theta}$ , we have", + "bbox": [ + 75, + 694, + 922, + 754 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {m i x}} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\geq \\frac {1}{N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} \\ell \\left(y _ {i}, f _ {\\theta} (x _ {i} + \\varepsilon_ {i})\\right) \\geq \\frac {1}{N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} \\ell \\left(y _ {i}, f _ {\\theta} (x _ {i} + \\varepsilon)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 762, + 866, + 817 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\varepsilon_{i} = R_{i}c_{x}\\mathbb{E}_{\\lambda \\sim \\mathcal{D}_{\\lambda}}[1 - \\lambda ]\\sqrt{d}$ with $R_{i} = \\langle \\nabla f_{\\theta}(x_{i}),x_{i}\\rangle /\\| \\nabla f_{\\theta}(x_{i})\\| \\cdot \\| x_{i}\\|$ and $\\varepsilon = \\min \\{\\varepsilon_i\\}$ .", + "bbox": [ + 75, + 827, + 922, + 871 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. is provided in Supplementary.", + "bbox": [ + 77, + 879, + 472, + 899 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 1 implies that as long as $||\\delta ||\\leq \\varepsilon$ holds, the MixUp loss $\\mathcal{L}^{\\mathrm{mix}}(\\theta ,\\mathbb{D}_{\\mathrm{val}})$ can be considered as an upper-bound of $\\mathcal{L}^{\\mathrm{ideal}}(\\theta ,\\mathbb{D}_{\\mathrm{val}})$ .", + "bbox": [ + 72, + 908, + 922, + 949 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 77, + 16, + 94, + 30 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "N. Karim et al.", + "bbox": [ + 157, + 14, + 312, + 30 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Clean Accuracy Retainer", + "text_level": 1, + "bbox": [ + 75, + 56, + 450, + 74 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In practice, it is desirable for a backdoor defense technique to be highly runtime efficient and retain the clean test accuracy of the original model. For better runtime efficiency and to retain clean accuracy, we choose to apply neural mask fine-tuning instead of fine-tuning the entire model, which can be formulated as,", + "bbox": [ + 72, + 87, + 922, + 167 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {M} = \\quad \\arg \\min \\quad \\mathcal {L} ^ {\\operatorname {m i x}} (\\theta \\odot M, \\mathbb {D} _ {\\text {v a l}}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 268, + 177, + 917, + 201 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nM \\mid m _ {l} ^ {(i)} \\in [ \\mu (l), 1 ], \\forall l, i\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 201, + 532, + 219 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We only optimize for neural masks $M$ using $\\mathbb{D}_{\\mathrm{val}}$ and define $\\theta \\odot M$ as,", + "bbox": [ + 75, + 234, + 825, + 253 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\theta \\odot M := \\left\\{\\Theta_ {1} \\odot M _ {1}, \\Theta_ {2} \\odot M _ {2}, \\dots , \\Theta_ {L} \\odot M _ {L} \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 269, + 917, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\theta \\coloneqq \\{\\Theta_1,\\dots ,\\Theta_L\\}$ , $M\\coloneqq \\{M_1,\\ldots ,M_L\\}$ , and $M_{l} = [m_{l}^{(1)}\\dots m_{l}^{k_{l}}]^{T}\\in \\mathbb{R}^{k_{l}}$ . We have", + "bbox": [ + 72, + 305, + 917, + 344 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta_ {l} \\odot M _ {l} := \\left[ m _ {l} ^ {(1)} \\Theta_ {l} ^ {(1)} \\dotsm m _ {l} ^ {\\left(k _ {l}\\right)} \\Theta_ {l} ^ {\\left(k _ {l}\\right)} \\right] \\in \\mathbb {R} ^ {k _ {l - 1} \\times k _ {l}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 344, + 917, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note, in Eq. (8), a scalar mask $m_{l}^{(i)}$ is applied to the weight vector $\\Theta_{l}^{(i)}$ corresponding to the $i^{th}$ neuron of $l^{th}$ layer. In our work, we formulate a constraint optimization problem where $M$ depends on a mask scheduling function $\\mu(l): [1, L] \\to [0, 1]$ . Notice that $\\mu(l)$ provides the lower limit of possible $M$ 's for the $l^{th}$ layer's neurons' mask. We find the suitable function for $\\mu(l)$ by analyzing commonly used mathematical functions (e.g., cosine, logarithmic, cubic, etc.). We analyze the impact of these functions in Section 5.3 and choose an exponential formulation (e.g., $\\alpha \\cdot e^{-\\beta \\cdot l}$ ) for $\\mu(l)$ since it produces the best possible outcome in terms of backdoor removal performance. Such formulation significantly reduces the mask search space, which leads to reduced runtime. Furthermore, the overall formulation for $M$ encourages relatively small changes to the original backdoor model's parameters. This helps us retain on-par clean test accuracy after backdoor removal, which is highly desirable for a defense technique. To this end, we aim to purify the backdoor model by optimizing for the best possible mask $\\widehat{M}$ that suppresses backdoor-affected neurons and bolsters the neurons responsible for clean test accuracy. Here, $\\widehat{M}$ should give us a purified model as, $f_{\\theta_c}(.)$ , where $\\theta_c = \\theta \\odot \\widehat{M}$ . Noteworthy, we do not change the bias as it may harm the classification accuracy.", + "bbox": [ + 72, + 378, + 920, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Sample Efficiency of NFT", + "text_level": 1, + "bbox": [ + 75, + 760, + 457, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we discuss how careful modifications to the optimization scenario can make NFT highly sample-efficient. For this analysis, we first train a backdoor model (PreActResNet-18 [18]) on a poisoned CIFAR10 dataset for 200 epochs. For poisoning the dataset, we use TrojanNet [35] backdoor attack with a poison rate of $10\\%$ . In Fig. 1, we show t-SNE visualization of different class clusters obtained using the backdoor model. Instead of 10 clusters, we have an additional cluster (red color, labeled \"11\") that sits closely with the original target class cluster (in this case, the target class is \"0\"). We name this cluster \"poison cluster\",", + "bbox": [ + 72, + 791, + 920, + 949 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", + "bbox": [ + 179, + 14, + 837, + 34 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/661d214f21bf33f101b31e1d7c849ca255d3f6bf4039d4385dea1cfe3859d108.jpg", + "image_caption": [ + "(a) Backdoor Model" + ], + "image_footnote": [], + "bbox": [ + 84, + 52, + 358, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ea7a8cde66c1e057de8cc687d9d6c2923d24fe44c2d8f7c205b9aeaf214366e.jpg", + "image_caption": [ + "(b) Purification w/o Reg.", + "Fig. 1: t-SNE visualization of a backdoor model, where we show the \"poison cluster\" with red color and label \"11\". Since the attack target class is \"0\", cluster \"0\" and the poison cluster sit closely with each other (Fig. 1a). After purification, the cluster should break, and all triggered samples should be classified according to their original label. In Fig. 1b, we perform one-shot NFT without employing the regularizer. Due to the overfitting issue, the clean clusters lose their separability that can be established with Mask regularizer, which tackles this issue (larger cluster gaps as compared to scenarios in Fig. 1b) by keeping purified model parameters close to the original backdoor model (Fig. 1c). This, in turn, produces better clean test accuracy. For evaluation, we train a PreActResNet18 [18] on CIFAR10 dataset with a poison rate of $10\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 363, + 52, + 634, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0143fd2bc8fd83fcc35e7784a9c45b38fcf4f56f9b330947f453d39d70d5fad1.jpg", + "image_caption": [ + "(c) Purification with Reg." + ], + "image_footnote": [], + "bbox": [ + 634, + 50, + 910, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "whereas other clusters are \"clean clusters\". This cluster contains the embeddings of attacked or triggered samples from all other classes. The goal of any defense system is to break the formation of the poison cluster so that poison samples return to their original clusters.", + "bbox": [ + 72, + 448, + 920, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One-Shot NFT. Let us consider that there is only 1 sample (per class) available for the validation set $\\mathbb{D}_{\\mathrm{val}}$ . Applying NFT with this $\\mathbb{D}_{\\mathrm{val}}$ forms the clusters shown in Fig. 1b. Notice that the poison cluster breaks even with one-shot fine-tuning, indicating the backdoor's effect is removed successfully. However, since only one sample is available per class, the model easily overfits $\\mathbb{D}_{\\mathrm{val}}$ , reducing margins between clean clusters. Such unwanted overfitting issue negatively impacts the clean test accuracy. To combat this issue in scenarios where very few samples are available, we add a simple regularizer term as follows,", + "bbox": [ + 72, + 526, + 920, + 684 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {M \\mid m _ {l} ^ {(i)} \\in [ \\mu (l), 1 ], \\forall l, i} {\\arg \\min } \\mathcal {L} ^ {\\operatorname {m i x}} \\left(\\theta \\odot M, \\mathbb {D} _ {\\text {v a l}}\\right) + \\eta_ {c} \\| M _ {0} - M \\| _ {1} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 699, + 917, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $M_0$ are the initial mask values (initialized as 1's) and $\\eta_c$ is the regularizer coefficient. By minimizing the $\\ell_1$ -norm of the mask differences, we try to keep the purified model parameters $(\\theta \\odot M)$ close to the original backdoor model parameters $(\\theta \\odot M_0)$ . We hope to preserve the original decision boundary between clean clusters by keeping these parameters as close as possible. Note that the overfitting issue is not as prominent whenever we have a reasonably sized $\\mathbb{D}_{\\mathrm{val}}$ (e.g., $1\\%$ of $\\mathbb{D}_{\\mathrm{train}}$ ). Therefore, we choose the value of $\\eta_c$ to be $5e^{-4} / n_c$ , which dynamically changes based on the number of samples available per class $(n_c)$ . As the number of samples increases, the impact of the regularizer reduces. Note that, Eq. (9) represents the final optimization function for our proposed method.", + "bbox": [ + 75, + 754, + 920, + 949 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 77, + 16, + 94, + 29 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "N. Karim et al.", + "bbox": [ + 159, + 14, + 309, + 30 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a34836711a5956bf4dc669f1b4c06d9a6d0865ab075d6cd032ac36345e50d895.jpg", + "table_caption": [ + "Table 1: Comparison of different defense methods for CIFAR10 and ImageNet. Average drop $(\\downarrow)$ indicates the $\\%$ changes in ASR/ACC compared to the baseline, i.e., ASR/ACC of No Defense. A good defense should have a large ASR drop with a small ACC drop. Attacks are implemented with a poison rate of $10\\%$ ." + ], + "table_footnote": [], + "table_body": "
DatasetMethodNo DefenseANPI-BAUAWMFT-SAMRNPNFT (Ours)
AttacksASRACCASRACCASRACCASRACCASRACCASRACCASRACC
CIFAR-10Benign095.21092.28093.98093.56093.80093.16094.10
Badnets10092.964.8785.922.8485.965.7287.854.3486.172.7588.461.7490.82
Blend10094.114.7787.613.8189.105.5390.842.1388.930.9191.530.3193.17
Troj-one10089.573.7882.185.4786.206.9187.245.4186.453.8487.391.6487.71
Troj-all10088.333.9181.955.5384.896.8285.944.4284.604.0285.801.7987.10
SIG10088.641.0481.920.3783.604.1283.570.9085.380.5186.460.1287.16
Dyn-one10092.524.7388.611.7887.267.4891.163.3588.418.6190.051.3790.81
Dyn-all10092.614.2888.322.1984.517.3089.742.4687.7210.5790.281.4291.53
CLB10092.780.8387.411.4185.075.7886.701.8984.186.1290.381.0490.37
CBA93.2090.1727.8083.7945.1185.6336.1285.0538.8185.5817.7286.4021.6087.97
FBA10090.787.9582.9066.7087.4210.6687.3522.3187.069.4887.636.2188.56
WaNet98.6492.295.8186.703.1889.247.7286.942.9688.458.1090.262.3889.65
ISSBA99.8092.786.7685.423.8289.2012.4890.034.5789.597.5888.624.2490.18
LIRA99.2592.157.3487.414.5189.616.1388.503.8689.2211.8387.591.5390.57
BPPA99.7093.829.9490.2310.4690.579.9490.6810.6090.889.7491.375.0491.78
Avg. Drop--92.63↓ 5.94↓ 88.10↓ 4.66↓ 91.21↓ 3.71↓ 92.61↓ 4.26↓ 92.06↓ 2.95↓ 95.56↓ 1.81
ImageNetBenign077.06073.52071.85074.21071.63075.20075.51
Badnets99.2474.535.9169.376.3166.282.8769.464.1869.447.5870.493.6170.96
Troj-one99.2174.027.6369.157.7367.145.7469.352.8670.622.9472.173.1672.37
Troj-all97.5874.459.1869.867.5468.206.0269.643.2769.854.8171.452.6872.13
Blend10074.426.4370.207.7968.517.4568.618.1568.915.6970.243.8371.52
SIG94.6674.691.2369.824.2866.085.3770.023.4769.744.3670.732.9472.36
CLB95.0874.146.7169.194.3766.417.6469.703.5069.329.4471.523.0572.25
Dyn-one98.2474.806.6869.658.3269.618.6270.174.4270.0512.5670.392.6271.91
Dyn-all98.5675.0813.4970.189.8268.9212.6870.244.8169.9014.1869.473.7771.62
LIRA96.0474.6112.8669.2212.0869.8013.2769.353.1612.3170.5071.382.6270.73
WaNet97.6074.486.3468.345.6767.236.3170.024.4266.827.7871.624.7171.63
ISSBA98.2374.387.6168.424.5067.928.2169.513.3568.029.7470.812.0670.67
Avg. Drop--90.08↓ 5.17↓ 88.90↓ 7.41↓ 90.01↓ 4.72↓ 92.24↓ 5.61↓ 89.37↓ 3.66↓ 94.03↓ 2.84
", + "bbox": [ + 84, + 141, + 922, + 562 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experimental Results", + "text_level": 1, + "bbox": [ + 75, + 591, + 440, + 613 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Evaluation Settings", + "text_level": 1, + "bbox": [ + 75, + 627, + 382, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We evaluate the proposed method through a range of experiments on two widely used datasets for backdoor attack study: CIFAR10 [22] with 10 classes, GTSRB [47] with 43 classes. For the scalability test of our method, we also consider Tiny-ImageNet [24] with 100,000 images distributed among 200 classes and ImageNet [11] with 1.28M images distributed among 1000 classes. For multi-label clean-image backdoor attacks, we use object detection datasets Pascal VOC [14] and MS-COCO [33]. UCF-101 [46] and HMDB51 [23] have been used for evaluating in action recognition task. The ModelNet [56] dataset was also utilized to assess the performance of a 3D point cloud classifier. In addition to vision, we also consider attacks on natural language generation and use WMT2014 En-De [4] machine translation and OpenSubtitles2012 [48] dialogue generation datasets (Results are in the Supplementary).", + "bbox": [ + 72, + 656, + 917, + 889 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Attacks Configurations. Here, we first briefly overview the attack configurations on single-label image recognition datasets. We consider 14 state-of-the-art backdoor attacks: 1) Badnets [17], 2) Blend attack [9], 3 & 4) TrojanNet", + "bbox": [ + 72, + 891, + 917, + 949 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", + "bbox": [ + 179, + 14, + 837, + 34 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1d3a12e3b1916942cb0a610e3de458dbe0f40a0fd376629a8bc65dde13629c0d.jpg", + "table_caption": [ + "Table 2: Performance analysis for the multi-label backdoor attack [8]. We choose 3 object detection datasets [14,33] and ML-decoder [44] network architecture for this evaluation. Mean average precision (mAP) and ASR of the model, with and without defenses, have been shown." + ], + "table_footnote": [], + "table_body": "
DatasetNo defenseANPAWMRNPFT-SAMNFT (Ours)
ASRmAPASRmAPASRmAPASRmAPASRmAPASRmAP
VOC0786.492.521.786.926.687.319.287.619.386.817.389.1
VOC1284.891.918.685.319.085.913.886.414.687.114.288.4
MS-COCO85.688.019.784.122.683.417.184.319.283.816.685.8
", + "bbox": [ + 162, + 141, + 832, + 239 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/899fcd9db2b7eda0d8bcb04ae8a9655fb304b306f0712328d8f8e8c4f0aee82a.jpg", + "table_caption": [ + "Table 3: Performance analysis for action recognition tasks where we choose 2 video datasets for evaluation. We consider a clean-label attack [64], where we need to generate adversarial perturbations for each input frame." + ], + "table_footnote": [], + "table_body": "
DatasetNo defenseI-BAUAWMRNPFT-SAMNFT (Ours)
ASRACCASRACCASRACCASRACCASRACCASRACC
UCF-10181.375.620.470.620.870.117.070.315.971.613.371.2
HMDB-5180.245.017.541.115.240.912.640.410.841.79.440.8
", + "bbox": [ + 181, + 333, + 815, + 414 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(Troj-one & Troj-all) [35], 5) Sinusoidal signal attack (SIG) [2], 6 & 7) Input-Aware Attack (Dyn-one and Dyn-all) [40], 8) Clean-label attack (CLB) [50], 9) Composite backdoor (CBA) [32], 10) Deep feature space attack (FBA) [10], 11) Warping-based backdoor attack (WaNet) [39], 12) Invisible triggers based backdoor attack (ISSBA) [31], 13) Imperceptible backdoor attack (LIRA) [12], and 14) Quantization and contrastive learning-based attack (BPPA) [53]. In order to facilitate a fair comparison, we adopt trigger patterns and settings similar to those used in the original papers. Specifically, for both Trojan-one and Dyn-one attacks, we set all triggered images to have the same target label (i.e., all2one), whereas, for Trojan-all and Dyn-all attacks, we have uniformly distributed the target labels across all classes (i.e., all2all). Details on the hyper-parameters and overall training settings can be found in the Supplementary. We measure the success of an attack using two metrics: clean test accuracy (ACC) defined as the percentage of clean samples that are classified to their original target label and attack success rate (ASR) defined as the percentage of poison test samples ( $\\hat{x}$ ) that are classified to the target label ( $\\hat{y}$ ).", + "bbox": [ + 72, + 458, + 920, + 770 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Defenses Configurations. We compare our approach with 10 existing backdoor mitigation methods: 1) $FT-SAM$ [66]; 2) Adversarial Neural Pruning (ANP) [54]; 3) Implicit Backdoor Adversarial Unlearning ( $I-BAU$ ) [59]; 4) Adversarial Weight Masking ( $AWM$ ) [6]; 5) Reconstructive Neuron Pruning (RNP) [29]; 6) Fine-Pruning ( $FP$ ) [36]; 7) Mode Connectivity Repair ( $MCR$ ) [63]; 8) Neural Attention Distillation ( $NAD$ ) [28], 9) Causality-inspired Backdoor Defense ( $CBD$ ) [62]), 10) Anti Backdoor Learning ( $ABL$ ) [27]. In the main paper, we compare NFT with the first 5 defenses as they are more relevant, and the comparison with the rest of the methods is in Supplementary. To apply NFT, we", + "bbox": [ + 72, + 773, + 920, + 949 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "N. Karim et al.", + "bbox": [ + 157, + 14, + 312, + 30 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/fb3e8caf6671e7a1f3bffe231407c4a16f93e71b24f163ddef478eeee0a5717d.jpg", + "table_caption": [ + "Table 4: Removal performance (\\%) of NFT against backdoor attacks on 3D point cloud classifiers. The attack methods [26], namely Poison-Label Backdoor Attack (PointPBA) with interaction trigger (PointPBA-I), PointPBA with orientation trigger (PointPBA-O), and Clean-Label Backdoor Attack (PointCBA) were considered, as well as the \"backdoor points\" based attack (3DPC-BA) outlined in prior work [58]." + ], + "table_footnote": [], + "table_body": "
AttackNo DefenseANPAWMRNPFT-SAMNFT (Ours)
ASRACCASRACCASRACCASRACCASRACCASRACC
PointBA-I98.689.113.682.615.483.98.184.08.884.59.685.7
PointBA-O94.789.814.882.013.182.49.483.88.285.07.585.3
PointCBA66.088.721.283.321.583.818.684.620.384.719.486.1
3DPC-BA93.891.216.884.715.685.913.985.713.186.312.687.7
", + "bbox": [ + 176, + 159, + 820, + 269 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "take $1\\%$ clean validation data (set aside from the training set) and fine-tune the model for 100 epochs. An SGD optimizer has been employed with a learning rate of 0.05 and a momentum of 0.95. The rest of the experimental details for NFT and other defense methods are in the Supplementary.", + "bbox": [ + 72, + 308, + 917, + 388 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Performance Comparison of NFT", + "text_level": 1, + "bbox": [ + 75, + 422, + 552, + 443 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we compare the performance of NFT with other defenses in various scenarios: single-label (i.e., image classification), multi-label (i.e., object detection), video action recognition, and 3D point cloud.", + "bbox": [ + 72, + 461, + 917, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Single-Label Backdoor attack. In Table 1, we present the performance of different defenses for 2 widely used benchmarks. For CIFAR10, we consider 4 label poisoning attacks: Badnets, Blend, Trojan, and Dynamic. For all of these attacks, NFT shows significant performance improvement over other baseline methods. While ANP and AWM defenses work well for mild attacks with low poison rates (e.g., $5\\%$ ), the performance deteriorates for attacks with high poison rates (e.g., $\\geq 10\\%$ ). It is observable that NFT performs well across all attack scenarios, e.g., obtaining a $99.69\\%$ drop in ASR for blend attack while also performing well for clean data (only $0.94\\%$ of ACC drop). For Trojan and Dynamic attacks, we consider two different versions of attacks based on label-mapping criteria (all2one and all2all). The drop in attack success rate shows the effectiveness of NFT against such attacks. Recently, small and imperceptible perturbations as triggers have been developed in attacks (e.g., WaNet, LIRA, etc.) to fool the defense systems. While AWM generates perturbations as a proxy for these imperceptible triggers, NFT does a better job in this regard. For further validation of our proposed method, we use deep feature-based attacks CBA and FBA. Both of these attacks manipulate features to insert backdoor behavior. Overall, we achieve an average drop of $95.56\\%$ in ASR while sacrificing an ACC of $1.81\\%$ . For the scalability test, we consider a large and widely used dataset in vision tasks, ImageNet. In consistency with other datasets, NFT also obtains SOTA performance in this particular dataset. Due to page constraints, we move the performance comparison for GTSRB and Tiny-ImageNet to the Supplementary.", + "bbox": [ + 72, + 521, + 920, + 949 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 891, + 16, + 915, + 30 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/739caefc6cc6f472a3e3fa20536eceebcdcc00859c38125ba0e0bdc9ce58bd04.jpg", + "image_caption": [ + "Fig.2: Ablation with different Mask Scheduling Function $(\\mu)$" + ], + "image_footnote": [], + "bbox": [ + 77, + 53, + 268, + 154 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0a630b1ad7cd024747d42e8992d218f11393921131a29a5c2041a784a595ed3a.jpg", + "image_caption": [ + "Fig.3: Mask Distribution of AWM (left) and NFT (right)." + ], + "image_footnote": [], + "bbox": [ + 288, + 50, + 481, + 152 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/33c18e501e4c85b3e148d7a67d4a97b2dcb5ee0ed683ec77f0ed7c64e534259a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 515, + 50, + 709, + 152 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/75b5989a31239d4f1ac6b0417f212a1d00d58ffa8377f9dcaa9aded1fd44feba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 726, + 52, + 920, + 152 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Multi-label Backdoor Attack. In Table 2, we also evaluate our proposed method on multi-label clean-image backdoor attack [8]. In general, we put a trigger on the image and change the corresponding ground truth of that image. However, a certain combination of objects has been used as a trigger pattern. For example, if a combination of car, person, and truck is present in the image, it will fool the model into misclassifying it. Table 2 shows that our proposed method surpasses other defense strategies concerning both ASR and mAP metrics. Notably, ANP and AWM, which rely on adversarial search limited applicability in multi-label scenarios. This limitation arises from the less accurate process of approximating triggers for object detection. Conversely, FT-SAM's optimization driven by sharpness proves effective in removing backdoors, yet it achieves a lower mAP post-purification. This outcome is not ideal since the goal is to eliminate backdoors without significantly compromising clean accuracy.", + "bbox": [ + 72, + 253, + 920, + 508 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Action Recognition Model. We further consider attacks on action recognition models; results are reported in Table 3. We use two widely used datasets, UCF-101 [46] and HMDB51 [23], with a CNN+LSTM network architecture. An ImageNet pre-trained ResNet50 network has been used for the CNN, and a sequential input-based Long Short Term Memory (LSTM) [45] network has been put on top of it. We subsample the input video by keeping one out of every 5 frames and use a fixed frame resolution of $224 \\times 224$ . We choose a trigger size of $20 \\times 20$ . Following [64], we create the required perturbation for clean-label attack by running projected gradient descent (PGD) [37] for 2000 steps with a perturbation norm of $\\epsilon = 16$ . Note that our proposed augmentation strategies for image classification are directly applicable to video recognition. During training, we keep $5\\%$ samples from each class to use them later as the clean validation set. Table 3 shows that NFT outperforms other defenses by a significant margin. In the case of a high number of classes and multiple image frames in the same input, it is a challenging task to optimize for the proper trigger pattern through the adversarial search described in I-BAU and AWM. Without a good approximation of the trigger, these methods seem to underperform in most cases.", + "bbox": [ + 72, + 513, + 917, + 845 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3D Point Cloud. In this phase of our work, we assess NFT's resilience against attacks on 3D point cloud classifiers [26,58]. To evaluate, we utilize the ModelNet dataset [56] and the PointNet++ architecture [42]. The performance comparison of NFT and other defense methods is outlined in Table 4. NFT outperforms other defenses due to its unique formulations of the objective function.", + "bbox": [ + 72, + 850, + 920, + 949 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "N. Karim et al.", + "bbox": [ + 157, + 14, + 312, + 30 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/32f052f28caedc9430265e0477f4fdb9dc23ba00839e742fb0a38011776b51fb.jpg", + "table_caption": [ + "Table 5: Average runtime of different defense methods against all 14 attacks. An NVIDIA RTX3090 GPU is used." + ], + "table_footnote": [], + "table_body": "
Method|ANP I-BAU AWM FT-SAM RNP NFT (Ours)
Runtime (sec.)118.192.5112.598.7102.628.4
", + "bbox": [ + 82, + 126, + 493, + 165 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2bec722c56030f5539df398af269e04255d514e0e2db51e05e2a258f48cb8f7f.jpg", + "table_caption": [ + "Table 6: Sensitivity analysis of $\\alpha$ and $\\beta$ for LIRA on CIFAR10." + ], + "table_footnote": [], + "table_body": "
α0.90.90.90.80.80.80.70.70.7
β0.250.500.750.250.500.750.250.500.75
ASR3.653.424.872.731.533.764.184.925.23
ACC88.5288.3489.1289.9790.5789.3487.6487.7888.10
", + "bbox": [ + 527, + 105, + 903, + 170 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 75, + 208, + 331, + 229 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For all ablation studies, we consider the CIFAR10 dataset.", + "bbox": [ + 72, + 245, + 699, + 263 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Runtime Analysis. For runtime analysis, we present the training time for different defenses in Table 5. ANP and AWM both employ computationally expensive adversarial search procedures to prune neurons, which makes them almost $4\\mathrm{x}$ slower than our method. However, NFT offers a computationally less expensive defense with SOTA performance in major benchmarks.", + "bbox": [ + 72, + 266, + 920, + 362 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Choice of Scheduling Function, $\\mu$ . For choosing the suitable function for $\\mu$ , we conduct a detailed study with commonly used mathematical functions. Note that, we only consider a family of functions that decreases over the depth of the network (shown in Fig. 2). This allows more variations for deeper layer weights, making sense as they are more responsible for DNN decision-making. In our work, we choose an exponential formulation for $\\mu$ as it offers superior performance. We also perform sensitivity analysis of scheduling parameters $\\alpha$ and $\\beta$ in Table 6. In Fig. 3, we show the generated mask distributions of AWM and NFT. Compared to AWM, NFT produces more uniformly distributed masks that seem helpful for backdoor purification. We show the impact of $\\mu$ in Table 7. Nature of Optimization. In Table 7, we present the performance of SOTA techniques under different validation sizes. Even with 10 samples (single-shot), NFT performs reasonably well and offers better performance as compared to AWM. This again shows that the trigger generation process is less accurate and effective for a very small validation set. We also show the effect of the proposed mask regularizer that indirectly controls the change in weights for better ACC. Although AWM employs a similar $\\ell_1$ regularizer for masks, our proposed regularizer is more intuitive and specifically designed for better ACC preservation. While AWM encourages sparse solutions for $M$ (shown in the left subfigure of Fig. 3), it helps with ASR but heavily compromises ACC. We also show the performance of NFT without augmentations.", + "bbox": [ + 72, + 364, + 920, + 772 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Label Correction Rate. In the standard backdoor removal metric, it is sufficient for backdoored images to be classified as a non-target class (any class other than $\\hat{y}$ ). However, we also consider another metric, label correction rate (LCR), for quantifying the success of a defense. We define LCR as the percentage of poisoned samples correctly classified to their original classes. Any method with the highest value of LCR is considered to be the best defense method. For this evaluation, we use CIFAR10 dataset and 6 backdoor attacks. Initially, the correction rate is $0\\%$ with no defense as the ASR is close to $100\\%$ . Table 8 shows that NFT obtains better performance in terms of LCR.", + "bbox": [ + 72, + 773, + 920, + 947 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", + "bbox": [ + 179, + 14, + 837, + 34 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b605384c58315c47c95786660a4f8b428bbfdfc5499453ba738e237a02b92779.jpg", + "table_caption": [ + "Table 7: Purification performance (\\%) for various validation data sizes. NFT performs reasonably well even with as few as 10 samples, i.e., one sample (shot) per class for CIFAR10. We also show the impact of the mask regularizer, mask scheduling function $\\mu$ , and augmentations on performance, which resonates with Fig. 1. Mask regularizer has the most impact on the clean test accuracy (around $7\\%$ worse without the regularizer). Without strong augmentations, we have a better ACC with a slightly worse ASR (around $6\\%$ drop)." + ], + "table_footnote": [], + "table_body": "
AttackDynamicWaNetLIRA
Samples101001010010100
MethodASRACCASRACCASRACCASRACCASRACCASRACC
No Defense10092.5210092.5298.6492.2998.6492.2999.2592.1599.2592.15
AWM86.7455.739.1685.3383.0162.217.2384.3891.4566.6410.8385.87
FT-SAM8.3573.495.7284.709.3575.985.5686.6311.8372.404.8588.82
NFT w/o Reg.5.6776.741.3682.214.1876.723.0283.314.8374.582.3283.61
NFT w/o Aug.11.9181.8610.5989.5310.3683.107.8189.6812.2381.059.1688.74
NFT w/o μ(l)5.1180.323.0488.585.8582.464.6488.026.4881.944.3388.75
NFT4.8380.511.7290.084.4183.582.9689.155.1882.722.0489.34
", + "bbox": [ + 113, + 193, + 883, + 386 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/c284aa376c7fa0ebec1a8f7d62c6585f5869ed55645823f49678f6447177e32b.jpg", + "table_caption": [ + "Table 8: Label Correction Rate (%) for different defense techniques, defined as the percentage of backdoor samples that are correctly classified to their original ground truth label." + ], + "table_footnote": [], + "table_body": "
DefenseBadnetsTrojanBlendSIGCLBWaNet
No Defense000000
ANP84.7480.5281.3853.3582.7280.23
I-BAU78.4177.1277.5639.4678.0780.65
AWM79.3778.2479.8144.5179.8679.18
FT-SAM85.5680.6984.4957.6482.0483.62
NFT (Ours)86.8281.1585.6155.1886.2385.70
", + "bbox": [ + 256, + 473, + 743, + 611 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 75, + 645, + 290, + 668 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We proposed a backdoor purification framework, NFT, utilizing an augmentation-based neural mask fine-tuning approach. NFT can change the backdoor model weights in a computationally efficient manner while ensuring SOTA purification performance. Our proposed method showed that the addition of MixUp during fine-tuning replaces the need for a computationally expensive trigger synthesis process. Furthermore, we proposed a novel mask regularizer that helps us preserve the cluster separability of the original backdoor model. By preserving this separability, the proposed method offers better clean test accuracy compared to SOTA methods. Furthermore, we suggested using a mask scheduling function that reduces the mask search space and improves the computational efficiency further. Our extensive experiments on 5 different tasks validate the efficiency and efficacy of the proposed backdoor purification method. We also conducted a detailed ablation study to explain the reasoning behind our design choices.", + "bbox": [ + 72, + 687, + 924, + 941 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "N. Karim et al.", + "bbox": [ + 157, + 14, + 312, + 32 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 77, + 53, + 351, + 74 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Science Foundation under Grant ECCS-1810256 and CMMI-2246672.", + "bbox": [ + 75, + 94, + 920, + 131 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 77, + 162, + 234, + 182 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahmed, S., Al Arafat, A., Rizve, M.N., Hossain, R., Guo, Z., Rakin, A.S.: Ssda: Secure source-free domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19180-19190 (2023)", + "2. Barni, M., Kallas, K., Tondi, B.: A new backdoor attack in cnns by training set corruption without label poisoning. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 101-105. IEEE (2019)", + "3. Bian, J., Al Arafat, A., Xiong, H., Li, J., Li, L., Chen, H., Wang, J., Dou, D., Guo, Z.: Machine learning in real-time internet of things (iot) systems: A survey. IEEE Internet of Things Journal 9(11), 8364-8386 (2022)", + "4. Bojar, O., Buck, C., Federmann, C., Haddow, B., Koehn, P., Leveling, J., Monz, C., Pecina, P., Post, M., Saint-Amand, H., Soricut, R., Specia, L., Tamchyna, A.: Findings of the 2014 workshop on statistical machine translation. In: Proceedings of the Ninth Workshop on Statistical Machine Translation. pp. 12-58. Association for Computational Linguistics, Baltimore, Maryland, USA (Jun 2014). https://doi.org/10.3115/v1/W14-3302, https://aclanthology.org/W14-3302", + "5. Carratino, L., Cisse, M., Jenatton, R., Vert, J.P.: On mixup regularization. The Journal of Machine Learning Research 23(1), 14632-14662 (2022)", + "6. Chai, S., Chen, J.: One-shot neural backdoor erasing via adversarial weight masking. arXiv preprint arXiv:2207.04497 (2022)", + "7. Chen, H., Fu, C., Zhao, J., Koushanfar, F.: Deepinspect: A black-box trojan detection and mitigation framework for deep neural networks. In: IJCAI. p. 8 (2019)", + "8. Chen, K., Lou, X., Xu, G., Li, J., Zhang, T.: Clean-image backdoor: Attacking multi-label models with poisoned labels only. In: The Eleventh International Conference on Learning Representations (2023)", + "9. Chen, X., Liu, C., Li, B., Lu, K., Song, D.: Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526 (2017)", + "10. Cheng, S., Liu, Y., Ma, S., Zhang, X.: Deep feature space trojan attack of neural networks by controlled detoxification. In: AAAI. pp. 1148-1156 (2021)", + "11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255. IEEE (2009)", + "12. Doan, K., Lao, Y., Zhao, W., Li, P.: Lira: Learnable, imperceptible and robust backdoor attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11966-11976 (2021)", + "13. Edraki, M., Karim, N., Rahnavard, N., Mian, A., Shah, M.: Odyssey: Creation, analysis and detection of trojan models. IEEE Transactions on Information Forensics and Security 16, 4521-4533 (2021)", + "14. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)", + "15. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=6Tm1mposlrM" + ], + "bbox": [ + 79, + 201, + 920, + 947 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", + "bbox": [ + 179, + 14, + 837, + 30 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Gao, Y., Xu, C., Wang, D., Chen, S., Ranasinghe, D.C., Nepal, S.: Strip: A defence against trojan attacks on deep neural networks. In: Proceedings of the 35th Annual Computer Security Applications Conference. pp. 113-125 (2019)", + "17. Gu, T., Liu, K., Dolan-Gavitt, B., Garg, S.: Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access 7, 47230-47244 (2019)", + "18. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: European conference on computer vision. pp. 630-645. Springer (2016)", + "19. Hong, S., Chandrasekaran, V., Kaya, Y., Dumitras, T., Papernot, N.: On the effectiveness of mitigating data poisoning attacks with gradient shaping. arXiv preprint arXiv:2002.11497 (2020)", + "20. Janai, J., Güney, F., Behl, A., Geiger, A., et al.: Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends in Computer Graphics and Vision 12(1-3), 1-308 (2020)", + "21. Karim, N., Arafat, A.A., Rakin, A.S., Guo, Z., Rahnavard, N.: Fisher information guided purification against backdoor attacks. In: Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (CCS) (2024)", + "22. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "23. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)", + "24. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)", + "25. Levine, A., Feizi, S.: Deep partition aggregation: Provable defense against general poisoning attacks. arXiv preprint arXiv:2006.14768 (2020)", + "26. Li, X., Chen, Z., Zhao, Y., Tong, Z., Zhao, Y., Lim, A., Zhou, J.T.: Pointba: Towards backdoor attacks in 3d point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16492-16501 (2021)", + "27. Li, Y., Lyu, X., Koren, N., Lyu, L., Li, B., Ma, X.: Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems 34 (2021)", + "28. Li, Y., Lyu, X., Koren, N., Lyu, L., Li, B., Ma, X.: Neural attention distillation: Erasing backdoor triggers from deep neural networks. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=910K40M-oXE", + "29. Li, Y., Lyu, X., Ma, X., Koren, N., Lyu, L., Li, B., Jiang, Y.G.: Reconstructive neuron pruning for backdoor defense. arXiv preprint arXiv:2305.14876 (2023)", + "30. Li, Y., Wu, B., Jiang, Y., Li, Z., Xia, S.T.: Backdoor learning: A survey. arXiv preprint arXiv:2007.08745 (2020)", + "31. Li, Y., Li, Y., Wu, B., Li, L., He, R., Lyu, S.: Invisible backdoor attack with sample-specific triggers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16463-16472 (2021)", + "32. Lin, J., Xu, L., Liu, Y., Zhang, X.: Composite backdoor attack for deep neural network by mixing existing benign features. In: CCS. pp. 113-131 (2020)", + "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "34. Liu, K., Dolan-Gavitt, B., Garg, S.: Fine-pruning: Defending against backdooring attacks on deep neural networks. In: International Symposium on Research in Attacks, Intrusions, and Defenses. pp. 273-294. Springer (2018)" + ], + "bbox": [ + 79, + 56, + 917, + 947 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "N. Karim et al.", + "bbox": [ + 157, + 14, + 312, + 30 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Liu, Y., Ma, S., Afer, Y., Lee, W.C., Zhai, J., Wang, W., Zhang, X.: Trojaning attack on neural networks (2017)", + "36. Liu, Y., Xie, Y., Srivastava, A.: Neural trojans. In: 2017 IEEE International Conference on Computer Design (ICCD). pp. 45-48. IEEE (2017)", + "37. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)", + "38. Manoj, N., Blum, A.: Excess capacity and backdoor poisoning. Advances in Neural Information Processing Systems 34, 20373-20384 (2021)", + "39. Nguyen, A., Tran, A.: Wanet-imperceptible warping-based backdoor attack. arXiv preprint arXiv:2102.10369 (2021)", + "40. Nguyen, T.A., Tran, A.: Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems 33, 3454-3464 (2020)", + "41. Ning, R., Li, J., Xin, C., Wu, H.: Invisible poison: A blackbox clean label backdoor attack to deep neural networks. In: IEEE INFOCOM 2021-IEEE Conference on Computer Communications. pp. 1-10. IEEE (2021)", + "42. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "43. Qiu, H., Zeng, Y., Guo, S., Zhang, T., Qiu, M., Thuraisingham, B.: Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. In: Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security. pp. 363-377 (2021)", + "44. Ridnik, T., Sharir, G., Ben-Cohen, A., Ben-Baruch, E., Noy, A.: ML-decoder: Scalable and versatile classification head. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 32-41 (2023)", + "45. Sherstinsky, A.: Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. Physica D: Nonlinear Phenomena 404, 132306 (2020)", + "46. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)", + "47. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: The german traffic sign recognition benchmark: a multi-class classification competition. In: The 2011 international joint conference on neural networks. pp. 1453-1460. IEEE (2011)", + "48. Tiedemann, J.: Parallel data, tools and interfaces in opus. In: Lrec. vol. 2012, pp. 2214-2218. Citeseer (2012)", + "49. Tran, B., Li, J., Madry, A.: Spectral signatures in backdoor attacks. Advances in neural information processing systems 31 (2018)", + "50. Turner, A., Tsipras, D., Madry, A.: Clean-label backdoor attacks (2019), https://openreview.net/forum?id=HJg6e2CcK7", + "51. Wang, B., Yao, Y., Shan, S., Li, H., Viswanath, B., Zheng, H., Zhao, B.Y.: Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In: 2019 IEEE Symposium on Security and Privacy (SP). pp. 707-723. IEEE (2019)", + "52. Wang, R., Zhang, G., Liu, S., Chen, P.Y., Xiong, J., Wang, M.: Practical detection of trojan neural networks: Data-limited and data-free cases. In: European Conference on Computer Vision. pp. 222-238. Springer (2020)", + "53. Wang, Z., Zhai, J., Ma, S.: Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15074-15084 (2022)", + "54. Wu, D., Wang, Y.: Adversarial neuron pruning purifies backdoored deep models. In: NeurIPS (2021)" + ], + "bbox": [ + 75, + 56, + 917, + 947 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification", + "bbox": [ + 179, + 14, + 837, + 32 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems 33, 2958-2969 (2020)", + "56. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)", + "57. Xian, X., Wang, G., Srinivasa, J., Kundu, A., Bi, X., Hong, M., Ding, J.: Understanding backdoor attacks through the adaptability hypothesis. In: Proc. International Conference on Machine Learning (2023)", + "58. Xiang, Z., Miller, D.J., Chen, S., Li, X., Kesidis, G.: A backdoor attack against 3d point cloud classifiers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7597-7607 (2021)", + "59. Zeng, Y., Chen, S., Park, W., Mao, Z.M., Jin, M., Jia, R.: Adversarial unlearning of backdoors via implicit hypergradient. arXiv preprint arXiv:2110.03735 (2021)", + "60. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)", + "61. Zhang, L., Deng, Z., Kawaguchi, K., Ghorbani, A., Zou, J.: How does mixup help with robustness and generalization? arXiv preprint arXiv:2010.04819 (2020)", + "62. Zhang, Z., Liu, Q., Wang, Z., Lu, Z., Hu, Q.: Backdoor defense via deconfounded representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12228-12238 (2023)", + "63. Zhao, P., Chen, P.Y., Das, P., Ramamurthy, K.N., Lin, X.: Bridging mode connectivity in loss landscapes and adversarial robustness. arXiv preprint arXiv:2005.00060 (2020)", + "64. Zhao, S., Ma, X., Zheng, X., Bailey, J., Chen, J., Jiang, Y.G.: Clean-label backdoor attacks on video recognition models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14443-14452 (2020)", + "65. Zheng, R., Tang, R., Li, J., Liu, L.: Data-free backdoor removal based on channel lipschitzness. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V. pp. 175-191. Springer (2022)", + "66. Zhu, M., Wei, S., Shen, L., Fan, Y., Wu, B.: Enhancing fine-tuning based backdoor defense with sharpness-aware minimization. arXiv preprint arXiv:2304.11823 (2023)" + ], + "bbox": [ + 77, + 56, + 917, + 663 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "N. Karim et al.", + "bbox": [ + 157, + 14, + 312, + 30 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_model.json b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..14ebbad421bd8f8eb2d4551fb783687d1a8e5cb5 --- /dev/null +++ b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_model.json @@ -0,0 +1,2447 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.128, + 0.049, + 0.873, + 0.105 + ], + "angle": 0, + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.14, + 0.916, + 0.179 + ], + "angle": 0, + "content": "Nazmul Karim\\*,1, Abdullah Al Arafat\\*,2, Umar Khalid\\* Zhishan Guo\\* and Nazanin Rahnavard\\*" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.195, + 0.825, + 0.214 + ], + "angle": 0, + "content": "1University of Central Florida 2North Carolina State University" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.262, + 0.851, + 0.744 + ], + "angle": 0, + "content": "Abstract. Recent studies have revealed the vulnerability of deep neural networks (DNNs) to various backdoor attacks, where the behavior of DNNs can be compromised by utilizing certain types of triggers or poisoning mechanisms. State-of-the-art (SOTA) defenses employ too-sophisticated mechanisms that require either a computationally expensive adversarial search module for reverse-engineering the trigger distribution or an over-sensitive hyper-parameter selection module. Moreover, they offer sub-par performance in challenging scenarios, e.g., limited validation data and strong attacks. In this paper, we propose—Neural mask Fine-Tuning (NFT)—with an aim to optimally re-organize the neuron activities in a way that the effect of the backdoor is removed. Utilizing a simple data augmentation like MixUp, NFT relaxes the trigger synthesis process and eliminates the requirement of the adversarial search module. Our study further reveals that direct weight fine-tuning under limited validation data results in poor post-purification clean test accuracy, primarily due to overfitting issue. To overcome this, we propose to fine-tune neural masks instead of model weights. In addition, a mask regularizer has been devised to further mitigate the model drift during the purification process. The distinct characteristics of NFT render it highly efficient in both runtime and sample usage, as it can remove the backdoor even when a single sample is available from each class. We validate the effectiveness of NFT through extensive experiments covering the tasks of image classification, object detection, video action recognition, 3D point cloud, and natural language processing. We evaluate our method against 14 different attacks (LIRA, WaNet, etc.) on 11 benchmark data sets (ImageNet, UCF101, Pascal VOC, ModelNet, OpenSubtitles2012, etc.). Our code is available online in this GitHub Repository." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.78, + 0.31, + 0.8 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.821, + 0.919, + 0.918 + ], + "angle": 0, + "content": "Machine learning and computer vision algorithms are increasingly common in safety-critical applications [3, 20], necessitating the design of secure and robust learning algorithms. Backdoor attack [9,17] on deep neural network (DNN) models is one of the heavily studied branches of AI safety and robustness. Backdoor defenses can generally be categorized into two major groups based on whether" + }, + { + "type": "page_footnote", + "bbox": [ + 0.086, + 0.931, + 0.295, + 0.948 + ], + "angle": 0, + "content": "* Equal Contribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.095, + 0.031 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.311, + 0.032 + ], + "angle": 0, + "content": "N. Karim et al." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.056, + 0.923, + 0.504 + ], + "angle": 0, + "content": "the defense is done during training or test. Training time defense (e.g., [25,27,49]) focuses on training a benign model on the poisonous data, while test time defense (e.g., [6,29,34,54]) deals with purifying backdoor model after it has already been trained. In this work, our goal is to develop an efficient test-time defense. Some test-time defenses focus on synthesizing trigger patterns [7,51,52] followed by vanilla weight fine-tuning. Most of these defenses aim to synthesize class-specific triggers independently or use additional models to generate the triggers simultaneously. Recent state-of-the-art (SOTA) backdoor defense methods, e.g., ANP [54], I-BAU [59], AWM [6], employ very similar techniques and do not work well without an expensive adversarial search module. For example, ANP [54] performs an adversarial search to find vulnerable neurons responsible for backdoor behavior. Identifying and pruning vulnerable neurons require an exhaustive adversarial search, resulting in high computational costs. Similar to ANP, AWM and I-BAU also resort to trigger synthesizing with a modified adversarial search process. Another very recent technique, FT-SAM [66] uses sharpness-aware minimization (SAM) [15] to fine-tune model weights. SAM is a recently proposed optimizer that utilizes Stochastic Gradient Descent (SGD), which penalizes sudden changes in the loss surface by constraining the search area to a compact region. Since SAM performs a double forward pass to compute the loss gradient twice, it results in a notable runtime increase for FT-SAM. In our work, we aim to develop an effective backdoor defense system that neither requires an expensive adversarial search process to recover the trigger nor a special type of optimizer with a runtime bottleneck." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.923, + 0.95 + ], + "angle": 0, + "content": "To achieve this goal, we propose a simple yet effective approach Neural mask Fine-Tuning (NFT), to remove backdoor through augmented fine-tuning of cost-efficient neural masks. We start with replacing the expensive adversarial search-based trigger synthesis process with a simple data augmentation technique—MixUp [60]. In general, the backdoor is inserted by forcing the model to memorize the trigger distribution. Intuitively, synthesizing and unlearning that trigger distribution would effectively remove the backdoor. In this work, we show that unlearning can be performed by simply optimizing the MixUp loss over a clean validation set. Our theoretical analysis suggests that MixUp loss is an upper bound on the standard loss obtained from triggered (synthesized or already known) validation data, termed as ideal purification loss. As the minimization of ideal loss guarantees backdoor purification, minimizing the MixUp loss would effectively remove the backdoor (Sec. 4.1). As the next step of our method, we address the overfitting issue during weight fine-tuning under limited validation data. In general, the outcome of such overfitting is poor post-purification test accuracy, which is not desirable for any backdoor defense. To this end, we propose to fine-tune a set of neural masks instead of the model weights, as this type of soft-masking enables us to reprogram the neurons affected by the backdoor without significantly altering the original backdoor model. As an added step to this, a mask regularizer has been introduced to further mitigate model drift during the purification process. In addition, we deploy a mask scheduling function to have better control over the purification process. Our experimental" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.182, + 0.015, + 0.839, + 0.035 + ], + "angle": 0, + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.055, + 0.921, + 0.115 + ], + "angle": 0, + "content": "results indicate that these straightforward yet intuitive steps significantly improve the post-purification test accuracy as compared to previous SOTA. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.125, + 0.92, + 0.202 + ], + "angle": 0, + "content": "- We propose a novel backdoor removal framework utilizing simple MixUp-based model fine-tuning. Our thorough analysis shows how minimizing the MixUp loss eliminates the requirement of an expensive trigger synthesis process while effectively removing the backdoor (Sec. 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.203, + 0.922, + 0.357 + ], + "angle": 0, + "content": "- To preserve the post-purification test accuracy, we propose to fine-tune soft neural masks (instead of weights) as it prevents any drastic change in the original backdoor model (Sec. 4.2). Additionally, a novel mask regularizer has been introduced that further encourages the purified model to retain the class separability of the original model. In addition to being computationally efficient, our proposed method shows significant improvement in sample efficiency as it can purify backdoor even with one-shot fine-tuning, i.e., only a single sample is available from each class (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.922, + 0.436 + ], + "angle": 0, + "content": "- To show the effectiveness of NFT, we perform an extensive evaluation with 11 different datasets. Compared to previous SOTA, the superior performance against a wide range of attacks suggests that augmentation like MixUp can indeed replace the trigger synthesis process (Sec. 5)." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.463, + 0.332, + 0.484 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.502, + 0.922, + 0.832 + ], + "angle": 0, + "content": "Backdoor Attack. Neural networks are intrinsically vulnerable to backdoor attacks [38, 57]. A substantial number of studies have investigated the possibility of backdoor attacks after the initial studies [9, 17, 36] found the existence of backdoors in DNNs. Generally, backdoor attacks are categorized into two types: clean-label attacks and poison-label attacks. A clean-label backdoor attack does not alter the label [41, 50, 64], while a poison-label attack aims at specific target classes such that the DNN misclassifies to those classes in the presence of a trigger [28]. As for trigger types, researchers have studied numerous types of triggering patterns in their respective attacks [9, 13, 17, 30]. Such triggers can exist in the form of dynamic patterns [30] or as simple as a single pixel [49]. Some of the more complex backdoor triggers that have been proposed in the literature are sinusoidal strips [2], adversarial patterns [64], and blending backgrounds [9]. Besides, backdoor attacks exist for many different tasks, e.g., multi-label clean image attack [8] has been proposed that alters the label distribution to insert triggers into the model, which works well in multi-label (e.g., detection) settings; domain adaptation [1] setting while adversary source can successfully insert backdoor to the target domains, etc." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.833, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Backdoor Defense. Generally, the backdoor defense methods are categorized into two types: Training Time Defense, and Test Time Defense. Regarding training time defense techniques (a few to mention [16, 19, 36, 49]), the researchers have proposed numerous defense methods through input pre-processing [36], poison-suppression [19], model diagnosis [30], network pruning [36, 55], and model reconstruction [63], etc. Notably, DeepSweep [43] explores different augmentations to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.095, + 0.03 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.31, + 0.032 + ], + "angle": 0, + "content": "N. Karim et al." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.055, + 0.923, + 0.329 + ], + "angle": 0, + "content": "purify a backdoor model and rectify the triggered samples. Although DeepSweep revealed that different augmentation functions could be leveraged to invert the backdoor effect of a model or erase the trigger from a trigger-embedded image, our work re-purposes the usage of augmentation differently to cover the approximate (unknown) trigger distribution during the purification phase. Moreover, DeepSweep assumes that backdoor triggers are known to the defender, which is hardly a practical assumption. In the case of test time defenses, besides the works mentioned in the introduction related to reverse-engineering of backdoor triggers in the input samples [7, 51, 52], several recent works explored the model vulnerability/sensitivity towards adversarially perturbed neurons [54], weights [6], or network channels [65]. However, these approaches require expensive adversarial search processes to be effective. A concurrent work FIP [21] studied the loss-surface smoothness of the backdoor model and developed a purification method by regularizing the spectral norm of the model." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.359, + 0.331, + 0.38 + ], + "angle": 0, + "content": "3 Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.401, + 0.923, + 0.558 + ], + "angle": 0, + "content": "Attack Model. Our work considers the most commonly used data poisoning attacks. Consider \\(\\mathbb{D}_{\\mathrm{train}} = \\{x_i, y_i\\}_{i=1}^N\\) as the training data where \\(x_i \\in \\mathbb{R}^d\\) is an input sample labeled as \\(y_i \\in \\{0, \\dots, c-1\\}\\) sampled from unknown distribution \\(\\mathcal{D}\\) of the task to be learned. Here, \\(N\\) is the total number of samples, and \\(d\\) is the dimension of the input sample. In addition, we assume there are \\(c\\) number of classes in the input data. Let \\(f_{\\theta^*}: \\mathbb{R}^d \\to \\mathbb{R}^c\\) be a benign (ideal) DNN trained with \\(\\mathbb{D}_{\\mathrm{train}} \\sim \\mathcal{D}\\). Here, \\(\\theta^*\\) is the DNN parameters that is to be optimized using a suitable loss function \\(\\ell(.,.)\\). The total empirical loss can be defined as," + }, + { + "type": "equation", + "bbox": [ + 0.309, + 0.576, + 0.916, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\theta^ {*}, \\mathbb {D} _ {\\text {t r a i n}}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ \\ell \\left(y _ {i}, f _ {\\theta^ {*}} \\left(x _ {i}\\right)\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.64, + 0.923, + 0.78 + ], + "angle": 0, + "content": "Now, consider an adversary inserts backdoor to a model \\( f_{\\theta}(.) \\) through modifying a small subset of \\( \\mathbb{D}_{\\mathrm{train}} \\) as \\( \\{\\hat{x}_i,\\hat{y}_i\\} \\) such that \\( \\hat{y}_i = \\arg \\max f_\\theta (\\hat{x}_i) \\) preserving \\( y_{i} = \\arg \\max f_{\\theta}(x_{i}),\\forall (x_{i},y_{i})\\in \\mathbb{D}_{\\mathrm{train}}. \\) Here, \\( \\hat{x}_i = x_i + \\delta \\) is the triggered input with adversary set target label \\( \\hat{y}_i\\neq y_i \\), where \\( \\delta \\in \\mathbb{R}^d \\) represents trigger pattern. Defense Objective. Consider a defense model where defender removes backdoor from \\( f_{\\theta}(.) \\) using a small validation data \\( \\mathbb{D}_{\\mathrm{val}} = \\{x_i,y_i\\}_{i = 1}^{N_{val}} \\) such that \\( y_{i} = \\arg \\max f_{\\theta_{c}}(\\hat{x}_{i}) \\), where \\( y_{i}\\neq \\hat{y}_{i} \\)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.808, + 0.515, + 0.833 + ], + "angle": 0, + "content": "4 Neural Fine-Tuning (NFT)" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.851, + 0.923, + 0.949 + ], + "angle": 0, + "content": "Let us consider a fully-connected DNN, \\( f_{\\theta}:\\mathbb{R}^{d}\\to \\mathbb{R}^{c} \\), that receives a datapoint \\( x\\in \\mathbb{R}^d \\) and predicts a probability distribution \\( p\\in \\mathbb{R}^c \\); where \\( c \\) is the number of classes. In general, \\( x \\) goes through a multi-layer DNN architecture before the DNN model predicts an output class \\( i = \\arg \\max p \\). Let us consider a multi-layer DNN architecture with \\( L \\) layers in which the \\( l \\)-th layer contains \\( k_{l} \\) neurons." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.18, + 0.015, + 0.839, + 0.035 + ], + "angle": 0, + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.056, + 0.92, + 0.096 + ], + "angle": 0, + "content": "The neurons of each layer produce activations \\(\\psi_{l}\\in \\mathbb{R}^{k_{l}}\\) based on the output activations of previous layer \\(\\psi_{l - 1}\\in \\mathbb{R}^{k_{l - 1}}\\). To be specific," + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.106, + 0.919, + 0.128 + ], + "angle": 0, + "content": "\\[\n\\psi_ {l} := \\sigma \\left(\\Theta_ {l} ^ {T} \\cdot \\psi_ {l - 1} + b _ {l}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.138, + 0.923, + 0.281 + ], + "angle": 0, + "content": "where \\(\\sigma(.)\\) is a non-linear activation function, matrix \\(\\Theta_l = [\\Theta_l^{(1)}\\dots \\Theta_l^{(k_l)}]\\in \\mathbb{R}^{k_{l - 1}\\times k_l}\\), for \\(l = 1,2,\\ldots ,L\\), includes the weights of the \\(l\\)-th layer, and \\(b_{l}\\in \\mathbb{R}^{k_{l}}\\) is the bias vector. Here, \\(\\Theta_l^{(j)}\\in \\mathbb{R}^{k_{l - 1}}\\) denotes the weights vector corresponding to the activation of the \\(j\\)-th neuron of the \\(l\\)-th layer. Model parameters \\(\\theta\\) can be expressed as \\(\\theta = \\{\\Theta_1,\\dots ,\\Theta_L\\}\\). This type of multi-layer DNN architecture is also valid for convolutional neural networks, where we use multiple 2-D arrays of neurons (i.e., filters) instead of a 1-D array." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.305, + 0.405, + 0.326 + ], + "angle": 0, + "content": "4.1 Backdoor Suppressor" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.333, + 0.924, + 0.626 + ], + "angle": 0, + "content": "Our objective is to make the backdoor model forget about poison distribution while retaining the knowledge of clean distribution. To understand how we achieve this objective, let us first revisit the process of generating triggered data \\((\\hat{x})\\). In general, \\(\\hat{x}\\) is created by adding minor modifications (i.e., adding triggers \\(\\delta\\)) to clean data \\((x)\\). Note that the backdoor is inserted by forcing the model to learn the mapping, \\(\\hat{x} \\rightarrow \\hat{y}\\). Here, \\((\\hat{x}, \\hat{y})\\) is the poison data. If we were to change the mapping from \\((\\hat{x} \\rightarrow \\hat{y})\\) to \\((\\hat{x} \\rightarrow y)\\), we would have a robust clean model instead of a backdoor model. This is because, in this case, the model would treat \\(\\delta\\) as one type of augmentation and \\(\\hat{x}\\) as the augmented clean data. In summary, the backdoor insertion process functions as an augmentation process if we simply use \\(y\\) instead of \\(\\hat{y}\\). Now, ideally, fine-tuning the backdoor model with triggered data with corresponding ground truth labels (i.e., \\(\\{\\hat{x}, y\\}\\)) would remove the backdoor effect from the model. Let us consider this ideal scenario where we have access to the trigger \\(\\delta\\) during purification. We define the ideal purification loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.638, + 0.918, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {i d e a l}} (\\theta , \\mathbb {D} _ {\\text {v a l}}) = \\frac {1}{N _ {\\text {v a l}}} \\sum_ {i = 1} ^ {N _ {\\text {v a l}}} \\ell \\left(y _ {i}, f _ {\\theta} \\left(\\hat {x} _ {i}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.697, + 0.924, + 0.836 + ], + "angle": 0, + "content": "where \\(\\hat{x} = x + \\delta\\) and \\(y\\) is the ground truth label of \\(x\\). In our work, as we do not have access to \\(\\delta\\) or adversarially reverse engineer it [6,54,59], we relax this process by strongly augmenting the clean validation data, i.e., creating augmented \\(\\mathbb{D}_{\\mathrm{val}}\\) with already known augmentation technique such as MixUp [60]. For MixUp, we can easily perform \\(\\tilde{x}_{i,j} = \\lambda x_i + (1 - \\lambda)x_j\\) and \\(\\tilde{y}_{i,j} = \\lambda y_i + (1 - \\lambda)y_j\\) for \\(\\lambda \\in [0,1]\\); here \\(\\tilde{y}_{i,j}\\) represents the linear combination of one-hot vectors corresponding to \\(y_i\\) and \\(y_j\\). The loss after the MixUp becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.262, + 0.847, + 0.918, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {m i x}} (\\theta , \\mathbb {D} _ {\\text {v a l}}) = \\frac {1}{N _ {\\text {v a l}} ^ {2}} \\sum_ {i, j = 1} ^ {N _ {\\text {v a l}}} \\underset {\\lambda \\sim \\mathcal {D} _ {\\lambda}} {\\mathbb {E}} \\ell \\left(\\tilde {y} _ {i, j}, f _ {\\theta} \\left(\\tilde {x} _ {i, j}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.91, + 0.924, + 0.951 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{\\lambda}\\) is a distribution supported on \\([0,1]\\). In our work, we consider the widely used \\(\\mathcal{D}_{\\lambda}\\) - Beta distribution \\(Beta(\\alpha, \\beta)\\) for \\(\\alpha, \\beta > 0\\). We provide both" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.078, + 0.017, + 0.096, + 0.032 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.313, + 0.032 + ], + "angle": 0, + "content": "N. Karim et al." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.92, + 0.096 + ], + "angle": 0, + "content": "empirical (Sec. 5) and theoretical proof for a binary classification problem on why minimizing Eq. (4) would effectively remove the backdoor." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.096, + 0.923, + 0.212 + ], + "angle": 0, + "content": "Theoretical Justifications. For a fully-connected neural network (NN) with logistic loss \\(\\ell(y, f_{\\theta}(x)) = \\log(1 + \\exp(f_{\\theta}(x))) - yf_{\\theta}(x)\\) with \\(y \\in \\{0, 1\\}\\), it can be shown that \\(\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})\\) is an upper-bound of the second order Taylor expansion of the ideal loss \\(\\mathcal{L}^{\\mathrm{ideal}}(\\theta, \\mathbb{D}_{\\mathrm{val}})\\). With the nonlinearity \\(\\sigma\\) for ReLU and max-pooling in NN, the function \\(f_{\\theta}\\) satisfies that \\(f_{\\theta}(x) = \\nabla f_{\\theta}(x)^T x\\) and \\(\\nabla^2 f_{\\theta}(x) = 0\\) almost everywhere, where the gradient is taken with respect to the input \\(x\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.212, + 0.923, + 0.252 + ], + "angle": 0, + "content": "We first rewrite the \\(\\mathcal{L}^{\\mathrm{ideal}}(\\theta, \\mathbb{D}_{\\mathrm{val}})\\) using Taylor series approximation. The second-order Taylor expansion of \\(\\ell(y, f_{\\theta}(x + \\delta))\\) is given by," + }, + { + "type": "equation", + "bbox": [ + 0.076, + 0.262, + 0.938, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\ell (y, f _ {\\theta} (x + \\delta)) = \\ell (y, f _ {\\theta} (x)) + (g (f _ {\\theta} (x)) - y) (f _ {\\theta} (\\delta)) + \\frac {1}{2} g (f _ {\\theta} (x)) (1 - g (f _ {\\theta} (x))) (f _ {\\theta} (\\delta)) ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.309, + 0.923, + 0.372 + ], + "angle": 0, + "content": "where \\( g(x) = \\frac{e^x}{1 + e^x} \\) is the logistic function. Based on the MixUp related analysis in prior works [5, 61], the following can be derived for \\( \\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}}) \\) using the second-order Taylor series expansion," + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.924, + 0.436 + ], + "angle": 0, + "content": "Lemma 1. Assuming \\( f_{\\theta}(x) = \\nabla f_{\\theta}(x)^{T}x \\) and \\( \\nabla^2 f_{\\theta}(x) = 0 \\) (which are satisfied by ReLU and max-pooling activation functions), \\( \\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}}) \\) can be expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.437, + 0.92, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {m i x}} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) = \\mathcal {L} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) + \\mathcal {R} _ {1} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) + \\mathcal {R} _ {2} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.466, + 0.152, + 0.484 + ], + "angle": 0, + "content": "where," + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.495, + 0.905, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {R} _ {1} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\geq \\frac {R c _ {x} \\mathbb {E} _ {\\lambda} [ (1 - \\lambda) ] \\sqrt {d}}{N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} | g (f _ {\\theta} (x _ {i})) - y _ {i} | \\cdot | | \\nabla f _ {\\theta} (x _ {i}) | | _ {2} \\\\ \\mathcal {R} _ {2} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\geq \\frac {R ^ {2} c _ {x} ^ {2} \\mathbb {E} _ {\\lambda} [ (1 - \\lambda) ] ^ {2} d}{2 N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} | g (f _ {\\theta} (x _ {i})) (1 - g (f _ {\\theta} (x _ {i}))) | \\cdot | | \\nabla f _ {\\theta} (x _ {i}) | | _ {2} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.618, + 0.907, + 0.64 + ], + "angle": 0, + "content": "where \\(R = \\min_{i\\in [N_{\\mathrm{val}}]}\\langle \\nabla f_{\\theta}(x_i),x_i\\rangle /||\\nabla f_{\\theta}(x_i)||\\cdot ||x_i||\\) and \\(c_{x} > 0\\) is a constant." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.646, + 0.92, + 0.684 + ], + "angle": 0, + "content": "By comparing \\(\\ell(y, f_{\\theta}(x + \\delta))\\) and \\(\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})\\) for a fully connected NN, we can prove the following." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.924, + 0.755 + ], + "angle": 0, + "content": "Theorem 1. Suppose that \\( f_{\\theta}(x) = \\nabla f_{\\theta}(x)^{T}x \\), \\( \\nabla^2 f_{\\theta}(x) = 0 \\) and there exists a constant \\( c_{x} > 0 \\) such that \\( \\| x_{i}\\|_{2}\\geq c_{x}\\sqrt{d} \\) for all \\( i\\in \\{1,\\ldots ,N_{\\mathrm{val}}\\} \\). Then, for any \\( f_{\\theta} \\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.764, + 0.868, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {m i x}} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\geq \\frac {1}{N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} \\ell \\left(y _ {i}, f _ {\\theta} (x _ {i} + \\varepsilon_ {i})\\right) \\geq \\frac {1}{N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} \\ell \\left(y _ {i}, f _ {\\theta} (x _ {i} + \\varepsilon)\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.829, + 0.924, + 0.872 + ], + "angle": 0, + "content": "where \\(\\varepsilon_{i} = R_{i}c_{x}\\mathbb{E}_{\\lambda \\sim \\mathcal{D}_{\\lambda}}[1 - \\lambda ]\\sqrt{d}\\) with \\(R_{i} = \\langle \\nabla f_{\\theta}(x_{i}),x_{i}\\rangle /\\| \\nabla f_{\\theta}(x_{i})\\| \\cdot \\| x_{i}\\|\\) and \\(\\varepsilon = \\min \\{\\varepsilon_i\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.88, + 0.473, + 0.9 + ], + "angle": 0, + "content": "Proof. is provided in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.909, + 0.924, + 0.951 + ], + "angle": 0, + "content": "Theorem 1 implies that as long as \\(||\\delta ||\\leq \\varepsilon\\) holds, the MixUp loss \\(\\mathcal{L}^{\\mathrm{mix}}(\\theta ,\\mathbb{D}_{\\mathrm{val}})\\) can be considered as an upper-bound of \\(\\mathcal{L}^{\\mathrm{ideal}}(\\theta ,\\mathbb{D}_{\\mathrm{val}})\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.18, + 0.015, + 0.839, + 0.035 + ], + "angle": 0, + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.057, + 0.452, + 0.076 + ], + "angle": 0, + "content": "4.2 Clean Accuracy Retainer" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.089, + 0.923, + 0.168 + ], + "angle": 0, + "content": "In practice, it is desirable for a backdoor defense technique to be highly runtime efficient and retain the clean test accuracy of the original model. For better runtime efficiency and to retain clean accuracy, we choose to apply neural mask fine-tuning instead of fine-tuning the entire model, which can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.178, + 0.92, + 0.203 + ], + "angle": 0, + "content": "\\[\n\\widehat {M} = \\quad \\arg \\min \\quad \\mathcal {L} ^ {\\operatorname {m i x}} (\\theta \\odot M, \\mathbb {D} _ {\\text {v a l}}) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.203, + 0.534, + 0.22 + ], + "angle": 0, + "content": "\\[\nM \\mid m _ {l} ^ {(i)} \\in [ \\mu (l), 1 ], \\forall l, i\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.826, + 0.255 + ], + "angle": 0, + "content": "We only optimize for neural masks \\(M\\) using \\(\\mathbb{D}_{\\mathrm{val}}\\) and define \\(\\theta \\odot M\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.249, + 0.27, + 0.919, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\theta \\odot M := \\left\\{\\Theta_ {1} \\odot M _ {1}, \\Theta_ {2} \\odot M _ {2}, \\dots , \\Theta_ {L} \\odot M _ {L} \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.307, + 0.919, + 0.345 + ], + "angle": 0, + "content": "where \\(\\theta \\coloneqq \\{\\Theta_1,\\dots ,\\Theta_L\\}\\), \\(M\\coloneqq \\{M_1,\\ldots ,M_L\\}\\), and \\(M_{l} = [m_{l}^{(1)}\\dots m_{l}^{k_{l}}]^{T}\\in \\mathbb{R}^{k_{l}}\\). We have" + }, + { + "type": "equation", + "bbox": [ + 0.249, + 0.345, + 0.919, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\Theta_ {l} \\odot M _ {l} := \\left[ m _ {l} ^ {(1)} \\Theta_ {l} ^ {(1)} \\dotsm m _ {l} ^ {\\left(k _ {l}\\right)} \\Theta_ {l} ^ {\\left(k _ {l}\\right)} \\right] \\in \\mathbb {R} ^ {k _ {l - 1} \\times k _ {l}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.38, + 0.922, + 0.733 + ], + "angle": 0, + "content": "Note, in Eq. (8), a scalar mask \\( m_{l}^{(i)} \\) is applied to the weight vector \\( \\Theta_{l}^{(i)} \\) corresponding to the \\( i^{th} \\) neuron of \\( l^{th} \\) layer. In our work, we formulate a constraint optimization problem where \\( M \\) depends on a mask scheduling function \\( \\mu(l): [1, L] \\to [0, 1] \\). Notice that \\( \\mu(l) \\) provides the lower limit of possible \\( M \\)'s for the \\( l^{th} \\) layer's neurons' mask. We find the suitable function for \\( \\mu(l) \\) by analyzing commonly used mathematical functions (e.g., cosine, logarithmic, cubic, etc.). We analyze the impact of these functions in Section 5.3 and choose an exponential formulation (e.g., \\( \\alpha \\cdot e^{-\\beta \\cdot l} \\)) for \\( \\mu(l) \\) since it produces the best possible outcome in terms of backdoor removal performance. Such formulation significantly reduces the mask search space, which leads to reduced runtime. Furthermore, the overall formulation for \\( M \\) encourages relatively small changes to the original backdoor model's parameters. This helps us retain on-par clean test accuracy after backdoor removal, which is highly desirable for a defense technique. To this end, we aim to purify the backdoor model by optimizing for the best possible mask \\( \\widehat{M} \\) that suppresses backdoor-affected neurons and bolsters the neurons responsible for clean test accuracy. Here, \\( \\widehat{M} \\) should give us a purified model as, \\( f_{\\theta_c}(.) \\), where \\( \\theta_c = \\theta \\odot \\widehat{M} \\). Noteworthy, we do not change the bias as it may harm the classification accuracy." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.761, + 0.46, + 0.781 + ], + "angle": 0, + "content": "4.3 Sample Efficiency of NFT" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.793, + 0.922, + 0.95 + ], + "angle": 0, + "content": "In this section, we discuss how careful modifications to the optimization scenario can make NFT highly sample-efficient. For this analysis, we first train a backdoor model (PreActResNet-18 [18]) on a poisoned CIFAR10 dataset for 200 epochs. For poisoning the dataset, we use TrojanNet [35] backdoor attack with a poison rate of \\(10\\%\\). In Fig. 1, we show t-SNE visualization of different class clusters obtained using the backdoor model. Instead of 10 clusters, we have an additional cluster (red color, labeled \"11\") that sits closely with the original target class cluster (in this case, the target class is \"0\"). We name this cluster \"poison cluster\"," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.095, + 0.03 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.31, + 0.032 + ], + "angle": 0, + "content": "N. Karim et al." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.053, + 0.36, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.202, + 0.312, + 0.217 + ], + "angle": 0, + "content": "(a) Backdoor Model" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.053, + 0.635, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.202, + 0.609, + 0.217 + ], + "angle": 0, + "content": "(b) Purification w/o Reg." + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.052, + 0.911, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.202, + 0.884, + 0.217 + ], + "angle": 0, + "content": "(c) Purification with Reg." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.923, + 0.412 + ], + "angle": 0, + "content": "Fig. 1: t-SNE visualization of a backdoor model, where we show the \"poison cluster\" with red color and label \"11\". Since the attack target class is \"0\", cluster \"0\" and the poison cluster sit closely with each other (Fig. 1a). After purification, the cluster should break, and all triggered samples should be classified according to their original label. In Fig. 1b, we perform one-shot NFT without employing the regularizer. Due to the overfitting issue, the clean clusters lose their separability that can be established with Mask regularizer, which tackles this issue (larger cluster gaps as compared to scenarios in Fig. 1b) by keeping purified model parameters close to the original backdoor model (Fig. 1c). This, in turn, produces better clean test accuracy. For evaluation, we train a PreActResNet18 [18] on CIFAR10 dataset with a poison rate of \\(10\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.45, + 0.921, + 0.528 + ], + "angle": 0, + "content": "whereas other clusters are \"clean clusters\". This cluster contains the embeddings of attacked or triggered samples from all other classes. The goal of any defense system is to break the formation of the poison cluster so that poison samples return to their original clusters." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.528, + 0.922, + 0.685 + ], + "angle": 0, + "content": "One-Shot NFT. Let us consider that there is only 1 sample (per class) available for the validation set \\(\\mathbb{D}_{\\mathrm{val}}\\). Applying NFT with this \\(\\mathbb{D}_{\\mathrm{val}}\\) forms the clusters shown in Fig. 1b. Notice that the poison cluster breaks even with one-shot fine-tuning, indicating the backdoor's effect is removed successfully. However, since only one sample is available per class, the model easily overfits \\(\\mathbb{D}_{\\mathrm{val}}\\), reducing margins between clean clusters. Such unwanted overfitting issue negatively impacts the clean test accuracy. To combat this issue in scenarios where very few samples are available, we add a simple regularizer term as follows," + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.7, + 0.918, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\underset {M \\mid m _ {l} ^ {(i)} \\in [ \\mu (l), 1 ], \\forall l, i} {\\arg \\min } \\mathcal {L} ^ {\\operatorname {m i x}} \\left(\\theta \\odot M, \\mathbb {D} _ {\\text {v a l}}\\right) + \\eta_ {c} \\| M _ {0} - M \\| _ {1} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.922, + 0.95 + ], + "angle": 0, + "content": "where \\(M_0\\) are the initial mask values (initialized as 1's) and \\(\\eta_c\\) is the regularizer coefficient. By minimizing the \\(\\ell_1\\)-norm of the mask differences, we try to keep the purified model parameters \\((\\theta \\odot M)\\) close to the original backdoor model parameters \\((\\theta \\odot M_0)\\). We hope to preserve the original decision boundary between clean clusters by keeping these parameters as close as possible. Note that the overfitting issue is not as prominent whenever we have a reasonably sized \\(\\mathbb{D}_{\\mathrm{val}}\\) (e.g., \\(1\\%\\) of \\(\\mathbb{D}_{\\mathrm{train}}\\)). Therefore, we choose the value of \\(\\eta_c\\) to be \\(5e^{-4} / n_c\\), which dynamically changes based on the number of samples available per class \\((n_c)\\). As the number of samples increases, the impact of the regularizer reduces. Note that, Eq. (9) represents the final optimization function for our proposed method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.181, + 0.015, + 0.839, + 0.035 + ], + "angle": 0, + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.054, + 0.921, + 0.127 + ], + "angle": 0, + "content": "Table 1: Comparison of different defense methods for CIFAR10 and ImageNet. Average drop \\((\\downarrow)\\) indicates the \\(\\%\\) changes in ASR/ACC compared to the baseline, i.e., ASR/ACC of No Defense. A good defense should have a large ASR drop with a small ACC drop. Attacks are implemented with a poison rate of \\(10\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.142, + 0.923, + 0.564 + ], + "angle": 0, + "content": "
DatasetMethodNo DefenseANPI-BAUAWMFT-SAMRNPNFT (Ours)
AttacksASRACCASRACCASRACCASRACCASRACCASRACCASRACC
CIFAR-10Benign095.21092.28093.98093.56093.80093.16094.10
Badnets10092.964.8785.922.8485.965.7287.854.3486.172.7588.461.7490.82
Blend10094.114.7787.613.8189.105.5390.842.1388.930.9191.530.3193.17
Troj-one10089.573.7882.185.4786.206.9187.245.4186.453.8487.391.6487.71
Troj-all10088.333.9181.955.5384.896.8285.944.4284.604.0285.801.7987.10
SIG10088.641.0481.920.3783.604.1283.570.9085.380.5186.460.1287.16
Dyn-one10092.524.7388.611.7887.267.4891.163.3588.418.6190.051.3790.81
Dyn-all10092.614.2888.322.1984.517.3089.742.4687.7210.5790.281.4291.53
CLB10092.780.8387.411.4185.075.7886.701.8984.186.1290.381.0490.37
CBA93.2090.1727.8083.7945.1185.6336.1285.0538.8185.5817.7286.4021.6087.97
FBA10090.787.9582.9066.7087.4210.6687.3522.3187.069.4887.636.2188.56
WaNet98.6492.295.8186.703.1889.247.7286.942.9688.458.1090.262.3889.65
ISSBA99.8092.786.7685.423.8289.2012.4890.034.5789.597.5888.624.2490.18
LIRA99.2592.157.3487.414.5189.616.1388.503.8689.2211.8387.591.5390.57
BPPA99.7093.829.9490.2310.4690.579.9490.6810.6090.889.7491.375.0491.78
Avg. Drop--92.63↓ 5.94↓ 88.10↓ 4.66↓ 91.21↓ 3.71↓ 92.61↓ 4.26↓ 92.06↓ 2.95↓ 95.56↓ 1.81
ImageNetBenign077.06073.52071.85074.21071.63075.20075.51
Badnets99.2474.535.9169.376.3166.282.8769.464.1869.447.5870.493.6170.96
Troj-one99.2174.027.6369.157.7367.145.7469.352.8670.622.9472.173.1672.37
Troj-all97.5874.459.1869.867.5468.206.0269.643.2769.854.8171.452.6872.13
Blend10074.426.4370.207.7968.517.4568.618.1568.915.6970.243.8371.52
SIG94.6674.691.2369.824.2866.085.3770.023.4769.744.3670.732.9472.36
CLB95.0874.146.7169.194.3766.417.6469.703.5069.329.4471.523.0572.25
Dyn-one98.2474.806.6869.658.3269.618.6270.174.4270.0512.5670.392.6271.91
Dyn-all98.5675.0813.4970.189.8268.9212.6870.244.8169.9014.1869.473.7771.62
LIRA96.0474.6112.8669.2212.0869.8013.2769.353.1612.3170.5071.382.6270.73
WaNet97.6074.486.3468.345.6767.236.3170.024.4266.827.7871.624.7171.63
ISSBA98.2374.387.6168.424.5067.928.2169.513.3568.029.7470.812.0670.67
Avg. Drop--90.08↓ 5.17↓ 88.90↓ 7.41↓ 90.01↓ 4.72↓ 92.24↓ 5.61↓ 89.37↓ 3.66↓ 94.03↓ 2.84
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.593, + 0.441, + 0.614 + ], + "angle": 0, + "content": "5 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.629, + 0.383, + 0.648 + ], + "angle": 0, + "content": "5.1 Evaluation Settings" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.658, + 0.92, + 0.891 + ], + "angle": 0, + "content": "Datasets. We evaluate the proposed method through a range of experiments on two widely used datasets for backdoor attack study: CIFAR10 [22] with 10 classes, GTSRB [47] with 43 classes. For the scalability test of our method, we also consider Tiny-ImageNet [24] with 100,000 images distributed among 200 classes and ImageNet [11] with 1.28M images distributed among 1000 classes. For multi-label clean-image backdoor attacks, we use object detection datasets Pascal VOC [14] and MS-COCO [33]. UCF-101 [46] and HMDB51 [23] have been used for evaluating in action recognition task. The ModelNet [56] dataset was also utilized to assess the performance of a 3D point cloud classifier. In addition to vision, we also consider attacks on natural language generation and use WMT2014 En-De [4] machine translation and OpenSubtitles2012 [48] dialogue generation datasets (Results are in the Supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.892, + 0.919, + 0.951 + ], + "angle": 0, + "content": "Attacks Configurations. Here, we first briefly overview the attack configurations on single-label image recognition datasets. We consider 14 state-of-the-art backdoor attacks: 1) Badnets [17], 2) Blend attack [9], 3 & 4) TrojanNet" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.313, + 0.032 + ], + "angle": 0, + "content": "N. Karim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.054, + 0.921, + 0.127 + ], + "angle": 0, + "content": "Table 2: Performance analysis for the multi-label backdoor attack [8]. We choose 3 object detection datasets [14,33] and ML-decoder [44] network architecture for this evaluation. Mean average precision (mAP) and ASR of the model, with and without defenses, have been shown." + }, + { + "type": "table", + "bbox": [ + 0.163, + 0.142, + 0.835, + 0.24 + ], + "angle": 0, + "content": "
DatasetNo defenseANPAWMRNPFT-SAMNFT (Ours)
ASRmAPASRmAPASRmAPASRmAPASRmAPASRmAP
VOC0786.492.521.786.926.687.319.287.619.386.817.389.1
VOC1284.891.918.685.319.085.913.886.414.687.114.288.4
MS-COCO85.688.019.784.122.683.417.184.319.283.816.685.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.264, + 0.921, + 0.318 + ], + "angle": 0, + "content": "Table 3: Performance analysis for action recognition tasks where we choose 2 video datasets for evaluation. We consider a clean-label attack [64], where we need to generate adversarial perturbations for each input frame." + }, + { + "type": "table", + "bbox": [ + 0.184, + 0.334, + 0.816, + 0.416 + ], + "angle": 0, + "content": "
DatasetNo defenseI-BAUAWMRNPFT-SAMNFT (Ours)
ASRACCASRACCASRACCASRACCASRACCASRACC
UCF-10181.375.620.470.620.870.117.070.315.971.613.371.2
HMDB-5180.245.017.541.115.240.912.640.410.841.79.440.8
" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.459, + 0.921, + 0.771 + ], + "angle": 0, + "content": "(Troj-one & Troj-all) [35], 5) Sinusoidal signal attack (SIG) [2], 6 & 7) Input-Aware Attack (Dyn-one and Dyn-all) [40], 8) Clean-label attack (CLB) [50], 9) Composite backdoor (CBA) [32], 10) Deep feature space attack (FBA) [10], 11) Warping-based backdoor attack (WaNet) [39], 12) Invisible triggers based backdoor attack (ISSBA) [31], 13) Imperceptible backdoor attack (LIRA) [12], and 14) Quantization and contrastive learning-based attack (BPPA) [53]. In order to facilitate a fair comparison, we adopt trigger patterns and settings similar to those used in the original papers. Specifically, for both Trojan-one and Dyn-one attacks, we set all triggered images to have the same target label (i.e., all2one), whereas, for Trojan-all and Dyn-all attacks, we have uniformly distributed the target labels across all classes (i.e., all2all). Details on the hyper-parameters and overall training settings can be found in the Supplementary. We measure the success of an attack using two metrics: clean test accuracy (ACC) defined as the percentage of clean samples that are classified to their original target label and attack success rate (ASR) defined as the percentage of poison test samples (\\(\\hat{x}\\)) that are classified to the target label (\\(\\hat{y}\\))." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.775, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Defenses Configurations. We compare our approach with 10 existing backdoor mitigation methods: 1) \\(FT-SAM\\) [66]; 2) Adversarial Neural Pruning (ANP) [54]; 3) Implicit Backdoor Adversarial Unlearning (\\(I-BAU\\)) [59]; 4) Adversarial Weight Masking (\\(AWM\\)) [6]; 5) Reconstructive Neuron Pruning (RNP) [29]; 6) Fine-Pruning (\\(FP\\)) [36]; 7) Mode Connectivity Repair (\\(MCR\\)) [63]; 8) Neural Attention Distillation (\\(NAD\\)) [28], 9) Causality-inspired Backdoor Defense (\\(CBD\\)) [62]), 10) Anti Backdoor Learning (\\(ABL\\)) [27]. In the main paper, we compare NFT with the first 5 defenses as they are more relevant, and the comparison with the rest of the methods is in Supplementary. To apply NFT, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.182, + 0.015, + 0.839, + 0.035 + ], + "angle": 0, + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.917, + 0.032 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.054, + 0.921, + 0.145 + ], + "angle": 0, + "content": "Table 4: Removal performance (\\%) of NFT against backdoor attacks on 3D point cloud classifiers. The attack methods [26], namely Poison-Label Backdoor Attack (PointPBA) with interaction trigger (PointPBA-I), PointPBA with orientation trigger (PointPBA-O), and Clean-Label Backdoor Attack (PointCBA) were considered, as well as the \"backdoor points\" based attack (3DPC-BA) outlined in prior work [58]." + }, + { + "type": "table", + "bbox": [ + 0.178, + 0.16, + 0.821, + 0.27 + ], + "angle": 0, + "content": "
AttackNo DefenseANPAWMRNPFT-SAMNFT (Ours)
ASRACCASRACCASRACCASRACCASRACCASRACC
PointBA-I98.689.113.682.615.483.98.184.08.884.59.685.7
PointBA-O94.789.814.882.013.182.49.483.88.285.07.585.3
PointCBA66.088.721.283.321.583.818.684.620.384.719.486.1
3DPC-BA93.891.216.884.715.685.913.985.713.186.312.687.7
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.309, + 0.92, + 0.389 + ], + "angle": 0, + "content": "take \\(1\\%\\) clean validation data (set aside from the training set) and fine-tune the model for 100 epochs. An SGD optimizer has been employed with a learning rate of 0.05 and a momentum of 0.95. The rest of the experimental details for NFT and other defense methods are in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.423, + 0.553, + 0.444 + ], + "angle": 0, + "content": "5.2 Performance Comparison of NFT" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.462, + 0.92, + 0.52 + ], + "angle": 0, + "content": "In this section, we compare the performance of NFT with other defenses in various scenarios: single-label (i.e., image classification), multi-label (i.e., object detection), video action recognition, and 3D point cloud." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.522, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Single-Label Backdoor attack. In Table 1, we present the performance of different defenses for 2 widely used benchmarks. For CIFAR10, we consider 4 label poisoning attacks: Badnets, Blend, Trojan, and Dynamic. For all of these attacks, NFT shows significant performance improvement over other baseline methods. While ANP and AWM defenses work well for mild attacks with low poison rates (e.g., \\(5\\%\\)), the performance deteriorates for attacks with high poison rates (e.g., \\(\\geq 10\\%\\)). It is observable that NFT performs well across all attack scenarios, e.g., obtaining a \\(99.69\\%\\) drop in ASR for blend attack while also performing well for clean data (only \\(0.94\\%\\) of ACC drop). For Trojan and Dynamic attacks, we consider two different versions of attacks based on label-mapping criteria (all2one and all2all). The drop in attack success rate shows the effectiveness of NFT against such attacks. Recently, small and imperceptible perturbations as triggers have been developed in attacks (e.g., WaNet, LIRA, etc.) to fool the defense systems. While AWM generates perturbations as a proxy for these imperceptible triggers, NFT does a better job in this regard. For further validation of our proposed method, we use deep feature-based attacks CBA and FBA. Both of these attacks manipulate features to insert backdoor behavior. Overall, we achieve an average drop of \\(95.56\\%\\) in ASR while sacrificing an ACC of \\(1.81\\%\\). For the scalability test, we consider a large and widely used dataset in vision tasks, ImageNet. In consistency with other datasets, NFT also obtains SOTA performance in this particular dataset. Due to page constraints, we move the performance comparison for GTSRB and Tiny-ImageNet to the Supplementary." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.313, + 0.032 + ], + "angle": 0, + "content": "N. Karim et al." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.055, + 0.271, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.051, + 0.482, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.167, + 0.482, + 0.204 + ], + "angle": 0, + "content": "Fig.2: Ablation with different Mask Scheduling Function \\((\\mu)\\)" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.051, + 0.71, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.053, + 0.921, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.166, + 0.92, + 0.203 + ], + "angle": 0, + "content": "Fig.3: Mask Distribution of AWM (left) and NFT (right)." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.255, + 0.921, + 0.509 + ], + "angle": 0, + "content": "Multi-label Backdoor Attack. In Table 2, we also evaluate our proposed method on multi-label clean-image backdoor attack [8]. In general, we put a trigger on the image and change the corresponding ground truth of that image. However, a certain combination of objects has been used as a trigger pattern. For example, if a combination of car, person, and truck is present in the image, it will fool the model into misclassifying it. Table 2 shows that our proposed method surpasses other defense strategies concerning both ASR and mAP metrics. Notably, ANP and AWM, which rely on adversarial search limited applicability in multi-label scenarios. This limitation arises from the less accurate process of approximating triggers for object detection. Conversely, FT-SAM's optimization driven by sharpness proves effective in removing backdoors, yet it achieves a lower mAP post-purification. This outcome is not ideal since the goal is to eliminate backdoors without significantly compromising clean accuracy." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.514, + 0.92, + 0.846 + ], + "angle": 0, + "content": "Action Recognition Model. We further consider attacks on action recognition models; results are reported in Table 3. We use two widely used datasets, UCF-101 [46] and HMDB51 [23], with a CNN+LSTM network architecture. An ImageNet pre-trained ResNet50 network has been used for the CNN, and a sequential input-based Long Short Term Memory (LSTM) [45] network has been put on top of it. We subsample the input video by keeping one out of every 5 frames and use a fixed frame resolution of \\(224 \\times 224\\). We choose a trigger size of \\(20 \\times 20\\). Following [64], we create the required perturbation for clean-label attack by running projected gradient descent (PGD) [37] for 2000 steps with a perturbation norm of \\(\\epsilon = 16\\). Note that our proposed augmentation strategies for image classification are directly applicable to video recognition. During training, we keep \\(5\\%\\) samples from each class to use them later as the clean validation set. Table 3 shows that NFT outperforms other defenses by a significant margin. In the case of a high number of classes and multiple image frames in the same input, it is a challenging task to optimize for the proper trigger pattern through the adversarial search described in I-BAU and AWM. Without a good approximation of the trigger, these methods seem to underperform in most cases." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.852, + 0.922, + 0.95 + ], + "angle": 0, + "content": "3D Point Cloud. In this phase of our work, we assess NFT's resilience against attacks on 3D point cloud classifiers [26,58]. To evaluate, we utilize the ModelNet dataset [56] and the PointNet++ architecture [42]. The performance comparison of NFT and other defense methods is outlined in Table 4. NFT outperforms other defenses due to its unique formulations of the objective function." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.18, + 0.015, + 0.838, + 0.035 + ], + "angle": 0, + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.056, + 0.487, + 0.11 + ], + "angle": 0, + "content": "Table 5: Average runtime of different defense methods against all 14 attacks. An NVIDIA RTX3090 GPU is used." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.127, + 0.494, + 0.166 + ], + "angle": 0, + "content": "
Method|ANP I-BAU AWM FT-SAM RNP NFT (Ours)
Runtime (sec.)118.192.5112.598.7102.628.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.054, + 0.92, + 0.09 + ], + "angle": 0, + "content": "Table 6: Sensitivity analysis of \\( \\alpha \\) and \\( \\beta \\) for LIRA on CIFAR10." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.106, + 0.905, + 0.171 + ], + "angle": 0, + "content": "
α0.90.90.90.80.80.80.70.70.7
β0.250.500.750.250.500.750.250.500.75
ASR3.653.424.872.731.533.764.184.925.23
ACC88.5288.3489.1289.9790.5789.3487.6487.7888.10
" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.209, + 0.334, + 0.23 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.246, + 0.7, + 0.265 + ], + "angle": 0, + "content": "For all ablation studies, we consider the CIFAR10 dataset." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.267, + 0.921, + 0.363 + ], + "angle": 0, + "content": "Runtime Analysis. For runtime analysis, we present the training time for different defenses in Table 5. ANP and AWM both employ computationally expensive adversarial search procedures to prune neurons, which makes them almost \\(4\\mathrm{x}\\) slower than our method. However, NFT offers a computationally less expensive defense with SOTA performance in major benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.365, + 0.922, + 0.773 + ], + "angle": 0, + "content": "Choice of Scheduling Function, \\(\\mu\\). For choosing the suitable function for \\(\\mu\\), we conduct a detailed study with commonly used mathematical functions. Note that, we only consider a family of functions that decreases over the depth of the network (shown in Fig. 2). This allows more variations for deeper layer weights, making sense as they are more responsible for DNN decision-making. In our work, we choose an exponential formulation for \\(\\mu\\) as it offers superior performance. We also perform sensitivity analysis of scheduling parameters \\(\\alpha\\) and \\(\\beta\\) in Table 6. In Fig. 3, we show the generated mask distributions of AWM and NFT. Compared to AWM, NFT produces more uniformly distributed masks that seem helpful for backdoor purification. We show the impact of \\(\\mu\\) in Table 7. Nature of Optimization. In Table 7, we present the performance of SOTA techniques under different validation sizes. Even with 10 samples (single-shot), NFT performs reasonably well and offers better performance as compared to AWM. This again shows that the trigger generation process is less accurate and effective for a very small validation set. We also show the effect of the proposed mask regularizer that indirectly controls the change in weights for better ACC. Although AWM employs a similar \\(\\ell_1\\) regularizer for masks, our proposed regularizer is more intuitive and specifically designed for better ACC preservation. While AWM encourages sparse solutions for \\(M\\) (shown in the left subfigure of Fig. 3), it helps with ASR but heavily compromises ACC. We also show the performance of NFT without augmentations." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.774, + 0.922, + 0.949 + ], + "angle": 0, + "content": "Label Correction Rate. In the standard backdoor removal metric, it is sufficient for backdoored images to be classified as a non-target class (any class other than \\(\\hat{y}\\)). However, we also consider another metric, label correction rate (LCR), for quantifying the success of a defense. We define LCR as the percentage of poisoned samples correctly classified to their original classes. Any method with the highest value of LCR is considered to be the best defense method. For this evaluation, we use CIFAR10 dataset and 6 backdoor attacks. Initially, the correction rate is \\(0\\%\\) with no defense as the ASR is close to \\(100\\%\\). Table 8 shows that NFT obtains better performance in terms of LCR." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.313, + 0.034 + ], + "angle": 0, + "content": "N. Karim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.054, + 0.923, + 0.182 + ], + "angle": 0, + "content": "Table 7: Purification performance (\\%) for various validation data sizes. NFT performs reasonably well even with as few as 10 samples, i.e., one sample (shot) per class for CIFAR10. We also show the impact of the mask regularizer, mask scheduling function \\(\\mu\\), and augmentations on performance, which resonates with Fig. 1. Mask regularizer has the most impact on the clean test accuracy (around \\(7\\%\\) worse without the regularizer). Without strong augmentations, we have a better ACC with a slightly worse ASR (around \\(6\\%\\) drop)." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.195, + 0.885, + 0.387 + ], + "angle": 0, + "content": "
AttackDynamicWaNetLIRA
Samples101001010010100
MethodASRACCASRACCASRACCASRACCASRACCASRACC
No Defense10092.5210092.5298.6492.2998.6492.2999.2592.1599.2592.15
AWM86.7455.739.1685.3383.0162.217.2384.3891.4566.6410.8385.87
FT-SAM8.3573.495.7284.709.3575.985.5686.6311.8372.404.8588.82
NFT w/o Reg.5.6776.741.3682.214.1876.723.0283.314.8374.582.3283.61
NFT w/o Aug.11.9181.8610.5989.5310.3683.107.8189.6812.2381.059.1688.74
NFT w/o μ(l)5.1180.323.0488.585.8582.464.6488.026.4881.944.3388.75
NFT4.8380.511.7290.084.4183.582.9689.155.1882.722.0489.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.403, + 0.924, + 0.458 + ], + "angle": 0, + "content": "Table 8: Label Correction Rate (%) for different defense techniques, defined as the percentage of backdoor samples that are correctly classified to their original ground truth label." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.474, + 0.744, + 0.613 + ], + "angle": 0, + "content": "
DefenseBadnetsTrojanBlendSIGCLBWaNet
No Defense000000
ANP84.7480.5281.3853.3582.7280.23
I-BAU78.4177.1277.5639.4678.0780.65
AWM79.3778.2479.8144.5179.8679.18
FT-SAM85.5680.6984.4957.6482.0483.62
NFT (Ours)86.8281.1585.6155.1886.2385.70
" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.647, + 0.291, + 0.669 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.688, + 0.925, + 0.942 + ], + "angle": 0, + "content": "We proposed a backdoor purification framework, NFT, utilizing an augmentation-based neural mask fine-tuning approach. NFT can change the backdoor model weights in a computationally efficient manner while ensuring SOTA purification performance. Our proposed method showed that the addition of MixUp during fine-tuning replaces the need for a computationally expensive trigger synthesis process. Furthermore, we proposed a novel mask regularizer that helps us preserve the cluster separability of the original backdoor model. By preserving this separability, the proposed method offers better clean test accuracy compared to SOTA methods. Furthermore, we suggested using a mask scheduling function that reduces the mask search space and improves the computational efficiency further. Our extensive experiments on 5 different tasks validate the efficiency and efficacy of the proposed backdoor purification method. We also conducted a detailed ablation study to explain the reasoning behind our design choices." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.18, + 0.015, + 0.838, + 0.032 + ], + "angle": 0, + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.054, + 0.352, + 0.076 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.095, + 0.921, + 0.133 + ], + "angle": 0, + "content": "This work was supported in part by the National Science Foundation under Grant ECCS-1810256 and CMMI-2246672." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.163, + 0.236, + 0.183 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.202, + 0.92, + 0.257 + ], + "angle": 0, + "content": "1. Ahmed, S., Al Arafat, A., Rizve, M.N., Hossain, R., Guo, Z., Rakin, A.S.: Ssda: Secure source-free domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19180-19190 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.257, + 0.921, + 0.31 + ], + "angle": 0, + "content": "2. Barni, M., Kallas, K., Tondi, B.: A new backdoor attack in cnns by training set corruption without label poisoning. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 101-105. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.31, + 0.921, + 0.363 + ], + "angle": 0, + "content": "3. Bian, J., Al Arafat, A., Xiong, H., Li, J., Li, L., Chen, H., Wang, J., Dou, D., Guo, Z.: Machine learning in real-time internet of things (iot) systems: A survey. IEEE Internet of Things Journal 9(11), 8364-8386 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.363, + 0.921, + 0.47 + ], + "angle": 0, + "content": "4. Bojar, O., Buck, C., Federmann, C., Haddow, B., Koehn, P., Leveling, J., Monz, C., Pecina, P., Post, M., Saint-Amand, H., Soricut, R., Specia, L., Tamchyna, A.: Findings of the 2014 workshop on statistical machine translation. In: Proceedings of the Ninth Workshop on Statistical Machine Translation. pp. 12-58. Association for Computational Linguistics, Baltimore, Maryland, USA (Jun 2014). https://doi.org/10.3115/v1/W14-3302, https://aclanthology.org/W14-3302" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.47, + 0.918, + 0.505 + ], + "angle": 0, + "content": "5. Carratino, L., Cisse, M., Jenatton, R., Vert, J.P.: On mixup regularization. The Journal of Machine Learning Research 23(1), 14632-14662 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.505, + 0.918, + 0.54 + ], + "angle": 0, + "content": "6. Chai, S., Chen, J.: One-shot neural backdoor erasing via adversarial weight masking. arXiv preprint arXiv:2207.04497 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.541, + 0.918, + 0.576 + ], + "angle": 0, + "content": "7. Chen, H., Fu, C., Zhao, J., Koushanfar, F.: Deepinspect: A black-box trojan detection and mitigation framework for deep neural networks. In: IJCAI. p. 8 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.577, + 0.918, + 0.629 + ], + "angle": 0, + "content": "8. Chen, K., Lou, X., Xu, G., Li, J., Zhang, T.: Clean-image backdoor: Attacking multi-label models with poisoned labels only. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.63, + 0.918, + 0.665 + ], + "angle": 0, + "content": "9. Chen, X., Liu, C., Li, B., Lu, K., Song, D.: Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.665, + 0.918, + 0.699 + ], + "angle": 0, + "content": "10. Cheng, S., Liu, Y., Ma, S., Zhang, X.: Deep feature space trojan attack of neural networks by controlled detoxification. In: AAAI. pp. 1148-1156 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.7, + 0.918, + 0.735 + ], + "angle": 0, + "content": "11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.736, + 0.918, + 0.789 + ], + "angle": 0, + "content": "12. Doan, K., Lao, Y., Zhao, W., Li, P.: Lira: Learnable, imperceptible and robust backdoor attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11966-11976 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.918, + 0.842 + ], + "angle": 0, + "content": "13. Edraki, M., Karim, N., Rahnavard, N., Mian, A., Shah, M.: Odyssey: Creation, analysis and detection of trojan models. IEEE Transactions on Information Forensics and Security 16, 4521-4533 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.843, + 0.918, + 0.895 + ], + "angle": 0, + "content": "14. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.896, + 0.918, + 0.949 + ], + "angle": 0, + "content": "15. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=6Tm1mposlrM" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.202, + 0.921, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.031 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.313, + 0.032 + ], + "angle": 0, + "content": "N. Karim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.92, + 0.111 + ], + "angle": 0, + "content": "16. Gao, Y., Xu, C., Wang, D., Chen, S., Ranasinghe, D.C., Nepal, S.: Strip: A defence against trojan attacks on deep neural networks. In: Proceedings of the 35th Annual Computer Security Applications Conference. pp. 113-125 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.111, + 0.918, + 0.145 + ], + "angle": 0, + "content": "17. Gu, T., Liu, K., Dolan-Gavitt, B., Garg, S.: Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access 7, 47230-47244 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.145, + 0.918, + 0.18 + ], + "angle": 0, + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: European conference on computer vision. pp. 630-645. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.918, + 0.232 + ], + "angle": 0, + "content": "19. Hong, S., Chandrasekaran, V., Kaya, Y., Dumitras, T., Papernot, N.: On the effectiveness of mitigating data poisoning attacks with gradient shaping. arXiv preprint arXiv:2002.11497 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.918, + 0.285 + ], + "angle": 0, + "content": "20. Janai, J., Güney, F., Behl, A., Geiger, A., et al.: Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends in Computer Graphics and Vision 12(1-3), 1-308 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.285, + 0.918, + 0.338 + ], + "angle": 0, + "content": "21. Karim, N., Arafat, A.A., Rakin, A.S., Guo, Z., Rahnavard, N.: Fisher information guided purification against backdoor attacks. In: Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (CCS) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.918, + 0.372 + ], + "angle": 0, + "content": "22. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.918, + 0.424 + ], + "angle": 0, + "content": "23. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.424, + 0.918, + 0.459 + ], + "angle": 0, + "content": "24. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.918, + 0.494 + ], + "angle": 0, + "content": "25. Levine, A., Feizi, S.: Deep partition aggregation: Provable defense against general poisoning attacks. arXiv preprint arXiv:2006.14768 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.494, + 0.918, + 0.547 + ], + "angle": 0, + "content": "26. Li, X., Chen, Z., Zhao, Y., Tong, Z., Zhao, Y., Lim, A., Zhou, J.T.: Pointba: Towards backdoor attacks in 3d point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16492-16501 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.918, + 0.599 + ], + "angle": 0, + "content": "27. Li, Y., Lyu, X., Koren, N., Lyu, L., Li, B., Ma, X.: Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems 34 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.918, + 0.668 + ], + "angle": 0, + "content": "28. Li, Y., Lyu, X., Koren, N., Lyu, L., Li, B., Ma, X.: Neural attention distillation: Erasing backdoor triggers from deep neural networks. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=910K40M-oXE" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.668, + 0.918, + 0.704 + ], + "angle": 0, + "content": "29. Li, Y., Lyu, X., Ma, X., Koren, N., Lyu, L., Li, B., Jiang, Y.G.: Reconstructive neuron pruning for backdoor defense. arXiv preprint arXiv:2305.14876 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.918, + 0.739 + ], + "angle": 0, + "content": "30. Li, Y., Wu, B., Jiang, Y., Li, Z., Xia, S.T.: Backdoor learning: A survey. arXiv preprint arXiv:2007.08745 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.739, + 0.918, + 0.791 + ], + "angle": 0, + "content": "31. Li, Y., Li, Y., Wu, B., Li, L., He, R., Lyu, S.: Invisible backdoor attack with sample-specific triggers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16463-16472 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.791, + 0.918, + 0.826 + ], + "angle": 0, + "content": "32. Lin, J., Xu, L., Liu, Y., Zhang, X.: Composite backdoor attack for deep neural network by mixing existing benign features. In: CCS. pp. 113-131 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.826, + 0.918, + 0.897 + ], + "angle": 0, + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.897, + 0.918, + 0.949 + ], + "angle": 0, + "content": "34. Liu, K., Dolan-Gavitt, B., Garg, S.: Fine-pruning: Defending against backdooring attacks on deep neural networks. In: International Symposium on Research in Attacks, Intrusions, and Defenses. pp. 273-294. Springer (2018)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.18, + 0.015, + 0.839, + 0.033 + ], + "angle": 0, + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.093 + ], + "angle": 0, + "content": "35. Liu, Y., Ma, S., Afer, Y., Lee, W.C., Zhai, J., Wang, W., Zhang, X.: Trojaning attack on neural networks (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.093, + 0.92, + 0.129 + ], + "angle": 0, + "content": "36. Liu, Y., Xie, Y., Srivastava, A.: Neural trojans. In: 2017 IEEE International Conference on Computer Design (ICCD). pp. 45-48. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.129, + 0.92, + 0.164 + ], + "angle": 0, + "content": "37. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.164, + 0.918, + 0.198 + ], + "angle": 0, + "content": "38. Manoj, N., Blum, A.: Excess capacity and backdoor poisoning. Advances in Neural Information Processing Systems 34, 20373-20384 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.198, + 0.918, + 0.232 + ], + "angle": 0, + "content": "39. Nguyen, A., Tran, A.: Wanet-imperceptible warping-based backdoor attack. arXiv preprint arXiv:2102.10369 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.232, + 0.918, + 0.267 + ], + "angle": 0, + "content": "40. Nguyen, T.A., Tran, A.: Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems 33, 3454-3464 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.267, + 0.918, + 0.32 + ], + "angle": 0, + "content": "41. Ning, R., Li, J., Xin, C., Wu, H.: Invisible poison: A blackbox clean label backdoor attack to deep neural networks. In: IEEE INFOCOM 2021-IEEE Conference on Computer Communications. pp. 1-10. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.32, + 0.918, + 0.372 + ], + "angle": 0, + "content": "42. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.372, + 0.918, + 0.442 + ], + "angle": 0, + "content": "43. Qiu, H., Zeng, Y., Guo, S., Zhang, T., Qiu, M., Thuraisingham, B.: Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. In: Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security. pp. 363-377 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.442, + 0.918, + 0.495 + ], + "angle": 0, + "content": "44. Ridnik, T., Sharir, G., Ben-Cohen, A., Ben-Baruch, E., Noy, A.: ML-decoder: Scalable and versatile classification head. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 32-41 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.495, + 0.918, + 0.547 + ], + "angle": 0, + "content": "45. Sherstinsky, A.: Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. Physica D: Nonlinear Phenomena 404, 132306 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.547, + 0.918, + 0.582 + ], + "angle": 0, + "content": "46. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.582, + 0.918, + 0.635 + ], + "angle": 0, + "content": "47. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: The german traffic sign recognition benchmark: a multi-class classification competition. In: The 2011 international joint conference on neural networks. pp. 1453-1460. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.634, + 0.918, + 0.669 + ], + "angle": 0, + "content": "48. Tiedemann, J.: Parallel data, tools and interfaces in opus. In: Lrec. vol. 2012, pp. 2214-2218. Citeseer (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.669, + 0.918, + 0.704 + ], + "angle": 0, + "content": "49. Tran, B., Li, J., Madry, A.: Spectral signatures in backdoor attacks. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.704, + 0.918, + 0.739 + ], + "angle": 0, + "content": "50. Turner, A., Tsipras, D., Madry, A.: Clean-label backdoor attacks (2019), https://openreview.net/forum?id=HJg6e2CcK7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.739, + 0.918, + 0.791 + ], + "angle": 0, + "content": "51. Wang, B., Yao, Y., Shan, S., Li, H., Viswanath, B., Zheng, H., Zhao, B.Y.: Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In: 2019 IEEE Symposium on Security and Privacy (SP). pp. 707-723. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.791, + 0.918, + 0.844 + ], + "angle": 0, + "content": "52. Wang, R., Zhang, G., Liu, S., Chen, P.Y., Xiong, J., Wang, M.: Practical detection of trojan neural networks: Data-limited and data-free cases. In: European Conference on Computer Vision. pp. 222-238. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.844, + 0.918, + 0.914 + ], + "angle": 0, + "content": "53. Wang, Z., Zhai, J., Ma, S.: Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15074-15084 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.914, + 0.918, + 0.949 + ], + "angle": 0, + "content": "54. Wu, D., Wang, Y.: Adversarial neuron pruning purifies backdoored deep models. In: NeurIPS (2021)" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.313, + 0.032 + ], + "angle": 0, + "content": "N. Karim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.057, + 0.92, + 0.111 + ], + "angle": 0, + "content": "55. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems 33, 2958-2969 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.112, + 0.92, + 0.164 + ], + "angle": 0, + "content": "56. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.918, + 0.219 + ], + "angle": 0, + "content": "57. Xian, X., Wang, G., Srinivasa, J., Kundu, A., Bi, X., Hong, M., Ding, J.: Understanding backdoor attacks through the adaptability hypothesis. In: Proc. International Conference on Machine Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.918, + 0.272 + ], + "angle": 0, + "content": "58. Xiang, Z., Miller, D.J., Chen, S., Li, X., Kesidis, G.: A backdoor attack against 3d point cloud classifiers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7597-7607 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.272, + 0.918, + 0.308 + ], + "angle": 0, + "content": "59. Zeng, Y., Chen, S., Park, W., Mao, Z.M., Jin, M., Jia, R.: Adversarial unlearning of backdoors via implicit hypergradient. arXiv preprint arXiv:2110.03735 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.918, + 0.343 + ], + "angle": 0, + "content": "60. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.918, + 0.379 + ], + "angle": 0, + "content": "61. Zhang, L., Deng, Z., Kawaguchi, K., Ghorbani, A., Zou, J.: How does mixup help with robustness and generalization? arXiv preprint arXiv:2010.04819 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.918, + 0.432 + ], + "angle": 0, + "content": "62. Zhang, Z., Liu, Q., Wang, Z., Lu, Z., Hu, Q.: Backdoor defense via deconfounded representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12228-12238 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.918, + 0.485 + ], + "angle": 0, + "content": "63. Zhao, P., Chen, P.Y., Das, P., Ramamurthy, K.N., Lin, X.: Bridging mode connectivity in loss landscapes and adversarial robustness. arXiv preprint arXiv:2005.00060 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.485, + 0.918, + 0.539 + ], + "angle": 0, + "content": "64. Zhao, S., Ma, X., Zheng, X., Bailey, J., Chen, J., Jiang, Y.G.: Clean-label backdoor attacks on video recognition models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14443-14452 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.918, + 0.61 + ], + "angle": 0, + "content": "65. Zheng, R., Tang, R., Li, J., Liu, L.: Data-free backdoor removal based on channel lipschitzness. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V. pp. 175-191. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.918, + 0.664 + ], + "angle": 0, + "content": "66. Zhu, M., Wei, S., Shen, L., Fan, Y., Wu, B.: Enhancing fine-tuning based backdoor defense with sharpness-aware minimization. arXiv preprint arXiv:2304.11823 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.057, + 0.92, + 0.664 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_origin.pdf b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..46cb7631f4d85a8558dd74839843fbc0ae123e7f --- /dev/null +++ b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/e2c8d233-a520-4d13-93d3-ddd88d0cec9c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17fc0c3caea0b6a53f4b748fe88dc025ba970813ae0367b561119a15bbb225aa +size 950845 diff --git a/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/full.md b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1306bb4511731d26946d667fbc2d6881aed34de9 --- /dev/null +++ b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/full.md @@ -0,0 +1,317 @@ +# Augmented Neural Fine-Tuning for Efficient Backdoor Purification + +Nazmul Karim\*,1, Abdullah Al Arafat\*,2, Umar Khalid\* Zhishan Guo\* and Nazanin Rahnavard\* + +1University of Central Florida 2North Carolina State University + +Abstract. Recent studies have revealed the vulnerability of deep neural networks (DNNs) to various backdoor attacks, where the behavior of DNNs can be compromised by utilizing certain types of triggers or poisoning mechanisms. State-of-the-art (SOTA) defenses employ too-sophisticated mechanisms that require either a computationally expensive adversarial search module for reverse-engineering the trigger distribution or an over-sensitive hyper-parameter selection module. Moreover, they offer sub-par performance in challenging scenarios, e.g., limited validation data and strong attacks. In this paper, we propose—Neural mask Fine-Tuning (NFT)—with an aim to optimally re-organize the neuron activities in a way that the effect of the backdoor is removed. Utilizing a simple data augmentation like MixUp, NFT relaxes the trigger synthesis process and eliminates the requirement of the adversarial search module. Our study further reveals that direct weight fine-tuning under limited validation data results in poor post-purification clean test accuracy, primarily due to overfitting issue. To overcome this, we propose to fine-tune neural masks instead of model weights. In addition, a mask regularizer has been devised to further mitigate the model drift during the purification process. The distinct characteristics of NFT render it highly efficient in both runtime and sample usage, as it can remove the backdoor even when a single sample is available from each class. We validate the effectiveness of NFT through extensive experiments covering the tasks of image classification, object detection, video action recognition, 3D point cloud, and natural language processing. We evaluate our method against 14 different attacks (LIRA, WaNet, etc.) on 11 benchmark data sets (ImageNet, UCF101, Pascal VOC, ModelNet, OpenSubtitles2012, etc.). Our code is available online in this GitHub Repository. + +# 1 Introduction + +Machine learning and computer vision algorithms are increasingly common in safety-critical applications [3, 20], necessitating the design of secure and robust learning algorithms. Backdoor attack [9,17] on deep neural network (DNN) models is one of the heavily studied branches of AI safety and robustness. Backdoor defenses can generally be categorized into two major groups based on whether + +the defense is done during training or test. Training time defense (e.g., [25,27,49]) focuses on training a benign model on the poisonous data, while test time defense (e.g., [6,29,34,54]) deals with purifying backdoor model after it has already been trained. In this work, our goal is to develop an efficient test-time defense. Some test-time defenses focus on synthesizing trigger patterns [7,51,52] followed by vanilla weight fine-tuning. Most of these defenses aim to synthesize class-specific triggers independently or use additional models to generate the triggers simultaneously. Recent state-of-the-art (SOTA) backdoor defense methods, e.g., ANP [54], I-BAU [59], AWM [6], employ very similar techniques and do not work well without an expensive adversarial search module. For example, ANP [54] performs an adversarial search to find vulnerable neurons responsible for backdoor behavior. Identifying and pruning vulnerable neurons require an exhaustive adversarial search, resulting in high computational costs. Similar to ANP, AWM and I-BAU also resort to trigger synthesizing with a modified adversarial search process. Another very recent technique, FT-SAM [66] uses sharpness-aware minimization (SAM) [15] to fine-tune model weights. SAM is a recently proposed optimizer that utilizes Stochastic Gradient Descent (SGD), which penalizes sudden changes in the loss surface by constraining the search area to a compact region. Since SAM performs a double forward pass to compute the loss gradient twice, it results in a notable runtime increase for FT-SAM. In our work, we aim to develop an effective backdoor defense system that neither requires an expensive adversarial search process to recover the trigger nor a special type of optimizer with a runtime bottleneck. + +To achieve this goal, we propose a simple yet effective approach Neural mask Fine-Tuning (NFT), to remove backdoor through augmented fine-tuning of cost-efficient neural masks. We start with replacing the expensive adversarial search-based trigger synthesis process with a simple data augmentation technique—MixUp [60]. In general, the backdoor is inserted by forcing the model to memorize the trigger distribution. Intuitively, synthesizing and unlearning that trigger distribution would effectively remove the backdoor. In this work, we show that unlearning can be performed by simply optimizing the MixUp loss over a clean validation set. Our theoretical analysis suggests that MixUp loss is an upper bound on the standard loss obtained from triggered (synthesized or already known) validation data, termed as ideal purification loss. As the minimization of ideal loss guarantees backdoor purification, minimizing the MixUp loss would effectively remove the backdoor (Sec. 4.1). As the next step of our method, we address the overfitting issue during weight fine-tuning under limited validation data. In general, the outcome of such overfitting is poor post-purification test accuracy, which is not desirable for any backdoor defense. To this end, we propose to fine-tune a set of neural masks instead of the model weights, as this type of soft-masking enables us to reprogram the neurons affected by the backdoor without significantly altering the original backdoor model. As an added step to this, a mask regularizer has been introduced to further mitigate model drift during the purification process. In addition, we deploy a mask scheduling function to have better control over the purification process. Our experimental + +results indicate that these straightforward yet intuitive steps significantly improve the post-purification test accuracy as compared to previous SOTA. Our contributions can be summarized as follows: + +- We propose a novel backdoor removal framework utilizing simple MixUp-based model fine-tuning. Our thorough analysis shows how minimizing the MixUp loss eliminates the requirement of an expensive trigger synthesis process while effectively removing the backdoor (Sec. 4.1). + +- To preserve the post-purification test accuracy, we propose to fine-tune soft neural masks (instead of weights) as it prevents any drastic change in the original backdoor model (Sec. 4.2). Additionally, a novel mask regularizer has been introduced that further encourages the purified model to retain the class separability of the original model. In addition to being computationally efficient, our proposed method shows significant improvement in sample efficiency as it can purify backdoor even with one-shot fine-tuning, i.e., only a single sample is available from each class (Sec. 4.3). + +- To show the effectiveness of NFT, we perform an extensive evaluation with 11 different datasets. Compared to previous SOTA, the superior performance against a wide range of attacks suggests that augmentation like MixUp can indeed replace the trigger synthesis process (Sec. 5). + +# 2 Related Work + +Backdoor Attack. Neural networks are intrinsically vulnerable to backdoor attacks [38, 57]. A substantial number of studies have investigated the possibility of backdoor attacks after the initial studies [9, 17, 36] found the existence of backdoors in DNNs. Generally, backdoor attacks are categorized into two types: clean-label attacks and poison-label attacks. A clean-label backdoor attack does not alter the label [41, 50, 64], while a poison-label attack aims at specific target classes such that the DNN misclassifies to those classes in the presence of a trigger [28]. As for trigger types, researchers have studied numerous types of triggering patterns in their respective attacks [9, 13, 17, 30]. Such triggers can exist in the form of dynamic patterns [30] or as simple as a single pixel [49]. Some of the more complex backdoor triggers that have been proposed in the literature are sinusoidal strips [2], adversarial patterns [64], and blending backgrounds [9]. Besides, backdoor attacks exist for many different tasks, e.g., multi-label clean image attack [8] has been proposed that alters the label distribution to insert triggers into the model, which works well in multi-label (e.g., detection) settings; domain adaptation [1] setting while adversary source can successfully insert backdoor to the target domains, etc. + +Backdoor Defense. Generally, the backdoor defense methods are categorized into two types: Training Time Defense, and Test Time Defense. Regarding training time defense techniques (a few to mention [16, 19, 36, 49]), the researchers have proposed numerous defense methods through input pre-processing [36], poison-suppression [19], model diagnosis [30], network pruning [36, 55], and model reconstruction [63], etc. Notably, DeepSweep [43] explores different augmentations to + +purify a backdoor model and rectify the triggered samples. Although DeepSweep revealed that different augmentation functions could be leveraged to invert the backdoor effect of a model or erase the trigger from a trigger-embedded image, our work re-purposes the usage of augmentation differently to cover the approximate (unknown) trigger distribution during the purification phase. Moreover, DeepSweep assumes that backdoor triggers are known to the defender, which is hardly a practical assumption. In the case of test time defenses, besides the works mentioned in the introduction related to reverse-engineering of backdoor triggers in the input samples [7, 51, 52], several recent works explored the model vulnerability/sensitivity towards adversarially perturbed neurons [54], weights [6], or network channels [65]. However, these approaches require expensive adversarial search processes to be effective. A concurrent work FIP [21] studied the loss-surface smoothness of the backdoor model and developed a purification method by regularizing the spectral norm of the model. + +# 3 Threat Model + +Attack Model. Our work considers the most commonly used data poisoning attacks. Consider $\mathbb{D}_{\mathrm{train}} = \{x_i, y_i\}_{i=1}^N$ as the training data where $x_i \in \mathbb{R}^d$ is an input sample labeled as $y_i \in \{0, \dots, c-1\}$ sampled from unknown distribution $\mathcal{D}$ of the task to be learned. Here, $N$ is the total number of samples, and $d$ is the dimension of the input sample. In addition, we assume there are $c$ number of classes in the input data. Let $f_{\theta^*}: \mathbb{R}^d \to \mathbb{R}^c$ be a benign (ideal) DNN trained with $\mathbb{D}_{\mathrm{train}} \sim \mathcal{D}$ . Here, $\theta^*$ is the DNN parameters that is to be optimized using a suitable loss function $\ell(.,.)$ . The total empirical loss can be defined as, + +$$ +\mathcal {L} \left(\theta^ {*}, \mathbb {D} _ {\text {t r a i n}}\right) = \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ \ell \left(y _ {i}, f _ {\theta^ {*}} \left(x _ {i}\right)\right) \right], \tag {1} +$$ + +Now, consider an adversary inserts backdoor to a model $f_{\theta}(.)$ through modifying a small subset of $\mathbb{D}_{\mathrm{train}}$ as $\{\hat{x}_i,\hat{y}_i\}$ such that $\hat{y}_i = \arg \max f_\theta (\hat{x}_i)$ preserving $y_{i} = \arg \max f_{\theta}(x_{i}),\forall (x_{i},y_{i})\in \mathbb{D}_{\mathrm{train}}.$ Here, $\hat{x}_i = x_i + \delta$ is the triggered input with adversary set target label $\hat{y}_i\neq y_i$ , where $\delta \in \mathbb{R}^d$ represents trigger pattern. Defense Objective. Consider a defense model where defender removes backdoor from $f_{\theta}(.)$ using a small validation data $\mathbb{D}_{\mathrm{val}} = \{x_i,y_i\}_{i = 1}^{N_{val}}$ such that $y_{i} = \arg \max f_{\theta_{c}}(\hat{x}_{i})$ , where $y_{i}\neq \hat{y}_{i}$ . + +# 4 Neural Fine-Tuning (NFT) + +Let us consider a fully-connected DNN, $f_{\theta}:\mathbb{R}^{d}\to \mathbb{R}^{c}$ , that receives a datapoint $x\in \mathbb{R}^d$ and predicts a probability distribution $p\in \mathbb{R}^c$ ; where $c$ is the number of classes. In general, $x$ goes through a multi-layer DNN architecture before the DNN model predicts an output class $i = \arg \max p$ . Let us consider a multi-layer DNN architecture with $L$ layers in which the $l$ -th layer contains $k_{l}$ neurons. + +The neurons of each layer produce activations $\psi_{l}\in \mathbb{R}^{k_{l}}$ based on the output activations of previous layer $\psi_{l - 1}\in \mathbb{R}^{k_{l - 1}}$ . To be specific, + +$$ +\psi_ {l} := \sigma \left(\Theta_ {l} ^ {T} \cdot \psi_ {l - 1} + b _ {l}\right), \tag {2} +$$ + +where $\sigma(.)$ is a non-linear activation function, matrix $\Theta_l = [\Theta_l^{(1)}\dots \Theta_l^{(k_l)}]\in \mathbb{R}^{k_{l - 1}\times k_l}$ , for $l = 1,2,\ldots ,L$ , includes the weights of the $l$ -th layer, and $b_{l}\in \mathbb{R}^{k_{l}}$ is the bias vector. Here, $\Theta_l^{(j)}\in \mathbb{R}^{k_{l - 1}}$ denotes the weights vector corresponding to the activation of the $j$ -th neuron of the $l$ -th layer. Model parameters $\theta$ can be expressed as $\theta = \{\Theta_1,\dots ,\Theta_L\}$ . This type of multi-layer DNN architecture is also valid for convolutional neural networks, where we use multiple 2-D arrays of neurons (i.e., filters) instead of a 1-D array. + +# 4.1 Backdoor Suppressor + +Our objective is to make the backdoor model forget about poison distribution while retaining the knowledge of clean distribution. To understand how we achieve this objective, let us first revisit the process of generating triggered data $(\hat{x})$ . In general, $\hat{x}$ is created by adding minor modifications (i.e., adding triggers $\delta$ ) to clean data $(x)$ . Note that the backdoor is inserted by forcing the model to learn the mapping, $\hat{x} \rightarrow \hat{y}$ . Here, $(\hat{x}, \hat{y})$ is the poison data. If we were to change the mapping from $(\hat{x} \rightarrow \hat{y})$ to $(\hat{x} \rightarrow y)$ , we would have a robust clean model instead of a backdoor model. This is because, in this case, the model would treat $\delta$ as one type of augmentation and $\hat{x}$ as the augmented clean data. In summary, the backdoor insertion process functions as an augmentation process if we simply use $y$ instead of $\hat{y}$ . Now, ideally, fine-tuning the backdoor model with triggered data with corresponding ground truth labels (i.e., $\{\hat{x}, y\}$ ) would remove the backdoor effect from the model. Let us consider this ideal scenario where we have access to the trigger $\delta$ during purification. We define the ideal purification loss as: + +$$ +\mathcal {L} ^ {\text {i d e a l}} (\theta , \mathbb {D} _ {\text {v a l}}) = \frac {1}{N _ {\text {v a l}}} \sum_ {i = 1} ^ {N _ {\text {v a l}}} \ell \left(y _ {i}, f _ {\theta} \left(\hat {x} _ {i}\right)\right), \tag {3} +$$ + +where $\hat{x} = x + \delta$ and $y$ is the ground truth label of $x$ . In our work, as we do not have access to $\delta$ or adversarially reverse engineer it [6,54,59], we relax this process by strongly augmenting the clean validation data, i.e., creating augmented $\mathbb{D}_{\mathrm{val}}$ with already known augmentation technique such as MixUp [60]. For MixUp, we can easily perform $\tilde{x}_{i,j} = \lambda x_i + (1 - \lambda)x_j$ and $\tilde{y}_{i,j} = \lambda y_i + (1 - \lambda)y_j$ for $\lambda \in [0,1]$ ; here $\tilde{y}_{i,j}$ represents the linear combination of one-hot vectors corresponding to $y_i$ and $y_j$ . The loss after the MixUp becomes: + +$$ +\mathcal {L} ^ {\text {m i x}} (\theta , \mathbb {D} _ {\text {v a l}}) = \frac {1}{N _ {\text {v a l}} ^ {2}} \sum_ {i, j = 1} ^ {N _ {\text {v a l}}} \underset {\lambda \sim \mathcal {D} _ {\lambda}} {\mathbb {E}} \ell \left(\tilde {y} _ {i, j}, f _ {\theta} \left(\tilde {x} _ {i, j}\right)\right), \tag {4} +$$ + +where $\mathcal{D}_{\lambda}$ is a distribution supported on $[0,1]$ . In our work, we consider the widely used $\mathcal{D}_{\lambda}$ - Beta distribution $Beta(\alpha, \beta)$ for $\alpha, \beta > 0$ . We provide both + +empirical (Sec. 5) and theoretical proof for a binary classification problem on why minimizing Eq. (4) would effectively remove the backdoor. + +Theoretical Justifications. For a fully-connected neural network (NN) with logistic loss $\ell(y, f_{\theta}(x)) = \log(1 + \exp(f_{\theta}(x))) - yf_{\theta}(x)$ with $y \in \{0, 1\}$ , it can be shown that $\mathcal{L}^{\mathrm{mix}}(\theta, \mathbb{D}_{\mathrm{val}})$ is an upper-bound of the second order Taylor expansion of the ideal loss $\mathcal{L}^{\mathrm{ideal}}(\theta, \mathbb{D}_{\mathrm{val}})$ . With the nonlinearity $\sigma$ for ReLU and max-pooling in NN, the function $f_{\theta}$ satisfies that $f_{\theta}(x) = \nabla f_{\theta}(x)^T x$ and $\nabla^2 f_{\theta}(x) = 0$ almost everywhere, where the gradient is taken with respect to the input $x$ . + +We first rewrite the $\mathcal{L}^{\mathrm{ideal}}(\theta, \mathbb{D}_{\mathrm{val}})$ using Taylor series approximation. The second-order Taylor expansion of $\ell(y, f_{\theta}(x + \delta))$ is given by, + +$$ +\ell (y, f _ {\theta} (x + \delta)) = \ell (y, f _ {\theta} (x)) + (g (f _ {\theta} (x)) - y) (f _ {\theta} (\delta)) + \frac {1}{2} g (f _ {\theta} (x)) (1 - g (f _ {\theta} (x))) (f _ {\theta} (\delta)) ^ {2}, +$$ + +where $g(x) = \frac{e^x}{1 + e^x}$ is the logistic function. Based on the MixUp related analysis in prior works [5, 61], the following can be derived for $\mathcal{L}^{\mathrm{mix}}(\theta, \mathbb{D}_{\mathrm{val}})$ using the second-order Taylor series expansion, + +Lemma 1. Assuming $f_{\theta}(x) = \nabla f_{\theta}(x)^{T}x$ and $\nabla^2 f_{\theta}(x) = 0$ (which are satisfied by ReLU and max-pooling activation functions), $\mathcal{L}^{\mathrm{mix}}(\theta, \mathbb{D}_{\mathrm{val}})$ can be expressed as, + +$$ +\mathcal {L} ^ {\mathrm {m i x}} (\theta , \mathbb {D} _ {\mathrm {v a l}}) = \mathcal {L} (\theta , \mathbb {D} _ {\mathrm {v a l}}) + \mathcal {R} _ {1} (\theta , \mathbb {D} _ {\mathrm {v a l}}) + \mathcal {R} _ {2} (\theta , \mathbb {D} _ {\mathrm {v a l}}) \qquad (5) +$$ + +where, + +$$ +\begin{array}{l} \mathcal {R} _ {1} (\theta , \mathbb {D} _ {\mathrm {v a l}}) \geq \frac {R c _ {x} \mathbb {E} _ {\lambda} [ (1 - \lambda) ] \sqrt {d}}{N _ {\mathrm {v a l}}} \sum_ {i = 1} ^ {N _ {\mathrm {v a l}}} | g (f _ {\theta} (x _ {i})) - y _ {i} | \cdot | | \nabla f _ {\theta} (x _ {i}) | | _ {2} \\ \mathcal {R} _ {2} (\theta , \mathbb {D} _ {\mathrm {v a l}}) \geq \frac {R ^ {2} c _ {x} ^ {2} \mathbb {E} _ {\lambda} [ (1 - \lambda) ] ^ {2} d}{2 N _ {\mathrm {v a l}}} \sum_ {i = 1} ^ {N _ {\mathrm {v a l}}} | g (f _ {\theta} (x _ {i})) (1 - g (f _ {\theta} (x _ {i}))) | \cdot | | \nabla f _ {\theta} (x _ {i}) | | _ {2} ^ {2}, \\ \end{array} +$$ + +where $R = \min_{i\in [N_{\mathrm{val}}]}\langle \nabla f_{\theta}(x_i),x_i\rangle /||\nabla f_{\theta}(x_i)||\cdot ||x_i||$ and $c_{x} > 0$ is a constant. + +By comparing $\ell(y, f_{\theta}(x + \delta))$ and $\mathcal{L}^{\mathrm{mix}}(\theta, \mathbb{D}_{\mathrm{val}})$ for a fully connected NN, we can prove the following. + +Theorem 1. Suppose that $f_{\theta}(x) = \nabla f_{\theta}(x)^{T}x$ , $\nabla^2 f_{\theta}(x) = 0$ and there exists a constant $c_{x} > 0$ such that $\| x_{i}\|_{2}\geq c_{x}\sqrt{d}$ for all $i\in \{1,\ldots ,N_{\mathrm{val}}\}$ . Then, for any $f_{\theta}$ , we have + +$$ +\mathcal {L} ^ {\mathrm {m i x}} (\theta , \mathbb {D} _ {\mathrm {v a l}}) \geq \frac {1}{N _ {\mathrm {v a l}}} \sum_ {i = 1} ^ {N _ {\mathrm {v a l}}} \ell \left(y _ {i}, f _ {\theta} (x _ {i} + \varepsilon_ {i})\right) \geq \frac {1}{N _ {\mathrm {v a l}}} \sum_ {i = 1} ^ {N _ {\mathrm {v a l}}} \ell \left(y _ {i}, f _ {\theta} (x _ {i} + \varepsilon)\right) +$$ + +where $\varepsilon_{i} = R_{i}c_{x}\mathbb{E}_{\lambda \sim \mathcal{D}_{\lambda}}[1 - \lambda ]\sqrt{d}$ with $R_{i} = \langle \nabla f_{\theta}(x_{i}),x_{i}\rangle /\| \nabla f_{\theta}(x_{i})\| \cdot \| x_{i}\|$ and $\varepsilon = \min \{\varepsilon_i\}$ . + +Proof. is provided in Supplementary. + +Theorem 1 implies that as long as $||\delta ||\leq \varepsilon$ holds, the MixUp loss $\mathcal{L}^{\mathrm{mix}}(\theta ,\mathbb{D}_{\mathrm{val}})$ can be considered as an upper-bound of $\mathcal{L}^{\mathrm{ideal}}(\theta ,\mathbb{D}_{\mathrm{val}})$ . + +# 4.2 Clean Accuracy Retainer + +In practice, it is desirable for a backdoor defense technique to be highly runtime efficient and retain the clean test accuracy of the original model. For better runtime efficiency and to retain clean accuracy, we choose to apply neural mask fine-tuning instead of fine-tuning the entire model, which can be formulated as, + +$$ +\widehat {M} = \quad \arg \min \quad \mathcal {L} ^ {\operatorname {m i x}} (\theta \odot M, \mathbb {D} _ {\text {v a l}}) \tag {6} +$$ + +$$ +M \mid m _ {l} ^ {(i)} \in [ \mu (l), 1 ], \forall l, i +$$ + +We only optimize for neural masks $M$ using $\mathbb{D}_{\mathrm{val}}$ and define $\theta \odot M$ as, + +$$ +\theta \odot M := \left\{\Theta_ {1} \odot M _ {1}, \Theta_ {2} \odot M _ {2}, \dots , \Theta_ {L} \odot M _ {L} \right\}, \tag {7} +$$ + +where $\theta \coloneqq \{\Theta_1,\dots ,\Theta_L\}$ , $M\coloneqq \{M_1,\ldots ,M_L\}$ , and $M_{l} = [m_{l}^{(1)}\dots m_{l}^{k_{l}}]^{T}\in \mathbb{R}^{k_{l}}$ . We have + +$$ +\Theta_ {l} \odot M _ {l} := \left[ m _ {l} ^ {(1)} \Theta_ {l} ^ {(1)} \dotsm m _ {l} ^ {\left(k _ {l}\right)} \Theta_ {l} ^ {\left(k _ {l}\right)} \right] \in \mathbb {R} ^ {k _ {l - 1} \times k _ {l}}. \tag {8} +$$ + +Note, in Eq. (8), a scalar mask $m_{l}^{(i)}$ is applied to the weight vector $\Theta_{l}^{(i)}$ corresponding to the $i^{th}$ neuron of $l^{th}$ layer. In our work, we formulate a constraint optimization problem where $M$ depends on a mask scheduling function $\mu(l): [1, L] \to [0, 1]$ . Notice that $\mu(l)$ provides the lower limit of possible $M$ 's for the $l^{th}$ layer's neurons' mask. We find the suitable function for $\mu(l)$ by analyzing commonly used mathematical functions (e.g., cosine, logarithmic, cubic, etc.). We analyze the impact of these functions in Section 5.3 and choose an exponential formulation (e.g., $\alpha \cdot e^{-\beta \cdot l}$ ) for $\mu(l)$ since it produces the best possible outcome in terms of backdoor removal performance. Such formulation significantly reduces the mask search space, which leads to reduced runtime. Furthermore, the overall formulation for $M$ encourages relatively small changes to the original backdoor model's parameters. This helps us retain on-par clean test accuracy after backdoor removal, which is highly desirable for a defense technique. To this end, we aim to purify the backdoor model by optimizing for the best possible mask $\widehat{M}$ that suppresses backdoor-affected neurons and bolsters the neurons responsible for clean test accuracy. Here, $\widehat{M}$ should give us a purified model as, $f_{\theta_c}(.)$ , where $\theta_c = \theta \odot \widehat{M}$ . Noteworthy, we do not change the bias as it may harm the classification accuracy. + +# 4.3 Sample Efficiency of NFT + +In this section, we discuss how careful modifications to the optimization scenario can make NFT highly sample-efficient. For this analysis, we first train a backdoor model (PreActResNet-18 [18]) on a poisoned CIFAR10 dataset for 200 epochs. For poisoning the dataset, we use TrojanNet [35] backdoor attack with a poison rate of $10\%$ . In Fig. 1, we show t-SNE visualization of different class clusters obtained using the backdoor model. Instead of 10 clusters, we have an additional cluster (red color, labeled "11") that sits closely with the original target class cluster (in this case, the target class is "0"). We name this cluster "poison cluster", + +![](images/661d214f21bf33f101b31e1d7c849ca255d3f6bf4039d4385dea1cfe3859d108.jpg) +(a) Backdoor Model + +![](images/5ea7a8cde66c1e057de8cc687d9d6c2923d24fe44c2d8f7c205b9aeaf214366e.jpg) +(b) Purification w/o Reg. +Fig. 1: t-SNE visualization of a backdoor model, where we show the "poison cluster" with red color and label "11". Since the attack target class is "0", cluster "0" and the poison cluster sit closely with each other (Fig. 1a). After purification, the cluster should break, and all triggered samples should be classified according to their original label. In Fig. 1b, we perform one-shot NFT without employing the regularizer. Due to the overfitting issue, the clean clusters lose their separability that can be established with Mask regularizer, which tackles this issue (larger cluster gaps as compared to scenarios in Fig. 1b) by keeping purified model parameters close to the original backdoor model (Fig. 1c). This, in turn, produces better clean test accuracy. For evaluation, we train a PreActResNet18 [18] on CIFAR10 dataset with a poison rate of $10\%$ . + +![](images/0143fd2bc8fd83fcc35e7784a9c45b38fcf4f56f9b330947f453d39d70d5fad1.jpg) +(c) Purification with Reg. + +whereas other clusters are "clean clusters". This cluster contains the embeddings of attacked or triggered samples from all other classes. The goal of any defense system is to break the formation of the poison cluster so that poison samples return to their original clusters. + +One-Shot NFT. Let us consider that there is only 1 sample (per class) available for the validation set $\mathbb{D}_{\mathrm{val}}$ . Applying NFT with this $\mathbb{D}_{\mathrm{val}}$ forms the clusters shown in Fig. 1b. Notice that the poison cluster breaks even with one-shot fine-tuning, indicating the backdoor's effect is removed successfully. However, since only one sample is available per class, the model easily overfits $\mathbb{D}_{\mathrm{val}}$ , reducing margins between clean clusters. Such unwanted overfitting issue negatively impacts the clean test accuracy. To combat this issue in scenarios where very few samples are available, we add a simple regularizer term as follows, + +$$ +\underset {M \mid m _ {l} ^ {(i)} \in [ \mu (l), 1 ], \forall l, i} {\arg \min } \mathcal {L} ^ {\operatorname {m i x}} \left(\theta \odot M, \mathbb {D} _ {\text {v a l}}\right) + \eta_ {c} \| M _ {0} - M \| _ {1} \tag {9} +$$ + +where $M_0$ are the initial mask values (initialized as 1's) and $\eta_c$ is the regularizer coefficient. By minimizing the $\ell_1$ -norm of the mask differences, we try to keep the purified model parameters $(\theta \odot M)$ close to the original backdoor model parameters $(\theta \odot M_0)$ . We hope to preserve the original decision boundary between clean clusters by keeping these parameters as close as possible. Note that the overfitting issue is not as prominent whenever we have a reasonably sized $\mathbb{D}_{\mathrm{val}}$ (e.g., $1\%$ of $\mathbb{D}_{\mathrm{train}}$ ). Therefore, we choose the value of $\eta_c$ to be $5e^{-4} / n_c$ , which dynamically changes based on the number of samples available per class $(n_c)$ . As the number of samples increases, the impact of the regularizer reduces. Note that, Eq. (9) represents the final optimization function for our proposed method. + +Table 1: Comparison of different defense methods for CIFAR10 and ImageNet. Average drop $(\downarrow)$ indicates the $\%$ changes in ASR/ACC compared to the baseline, i.e., ASR/ACC of No Defense. A good defense should have a large ASR drop with a small ACC drop. Attacks are implemented with a poison rate of $10\%$ . + +
DatasetMethodNo DefenseANPI-BAUAWMFT-SAMRNPNFT (Ours)
AttacksASRACCASRACCASRACCASRACCASRACCASRACCASRACC
CIFAR-10Benign095.21092.28093.98093.56093.80093.16094.10
Badnets10092.964.8785.922.8485.965.7287.854.3486.172.7588.461.7490.82
Blend10094.114.7787.613.8189.105.5390.842.1388.930.9191.530.3193.17
Troj-one10089.573.7882.185.4786.206.9187.245.4186.453.8487.391.6487.71
Troj-all10088.333.9181.955.5384.896.8285.944.4284.604.0285.801.7987.10
SIG10088.641.0481.920.3783.604.1283.570.9085.380.5186.460.1287.16
Dyn-one10092.524.7388.611.7887.267.4891.163.3588.418.6190.051.3790.81
Dyn-all10092.614.2888.322.1984.517.3089.742.4687.7210.5790.281.4291.53
CLB10092.780.8387.411.4185.075.7886.701.8984.186.1290.381.0490.37
CBA93.2090.1727.8083.7945.1185.6336.1285.0538.8185.5817.7286.4021.6087.97
FBA10090.787.9582.9066.7087.4210.6687.3522.3187.069.4887.636.2188.56
WaNet98.6492.295.8186.703.1889.247.7286.942.9688.458.1090.262.3889.65
ISSBA99.8092.786.7685.423.8289.2012.4890.034.5789.597.5888.624.2490.18
LIRA99.2592.157.3487.414.5189.616.1388.503.8689.2211.8387.591.5390.57
BPPA99.7093.829.9490.2310.4690.579.9490.6810.6090.889.7491.375.0491.78
Avg. Drop--92.63↓ 5.94↓ 88.10↓ 4.66↓ 91.21↓ 3.71↓ 92.61↓ 4.26↓ 92.06↓ 2.95↓ 95.56↓ 1.81
ImageNetBenign077.06073.52071.85074.21071.63075.20075.51
Badnets99.2474.535.9169.376.3166.282.8769.464.1869.447.5870.493.6170.96
Troj-one99.2174.027.6369.157.7367.145.7469.352.8670.622.9472.173.1672.37
Troj-all97.5874.459.1869.867.5468.206.0269.643.2769.854.8171.452.6872.13
Blend10074.426.4370.207.7968.517.4568.618.1568.915.6970.243.8371.52
SIG94.6674.691.2369.824.2866.085.3770.023.4769.744.3670.732.9472.36
CLB95.0874.146.7169.194.3766.417.6469.703.5069.329.4471.523.0572.25
Dyn-one98.2474.806.6869.658.3269.618.6270.174.4270.0512.5670.392.6271.91
Dyn-all98.5675.0813.4970.189.8268.9212.6870.244.8169.9014.1869.473.7771.62
LIRA96.0474.6112.8669.2212.0869.8013.2769.353.1612.3170.5071.382.6270.73
WaNet97.6074.486.3468.345.6767.236.3170.024.4266.827.7871.624.7171.63
ISSBA98.2374.387.6168.424.5067.928.2169.513.3568.029.7470.812.0670.67
Avg. Drop--90.08↓ 5.17↓ 88.90↓ 7.41↓ 90.01↓ 4.72↓ 92.24↓ 5.61↓ 89.37↓ 3.66↓ 94.03↓ 2.84
+ +# 5 Experimental Results + +# 5.1 Evaluation Settings + +Datasets. We evaluate the proposed method through a range of experiments on two widely used datasets for backdoor attack study: CIFAR10 [22] with 10 classes, GTSRB [47] with 43 classes. For the scalability test of our method, we also consider Tiny-ImageNet [24] with 100,000 images distributed among 200 classes and ImageNet [11] with 1.28M images distributed among 1000 classes. For multi-label clean-image backdoor attacks, we use object detection datasets Pascal VOC [14] and MS-COCO [33]. UCF-101 [46] and HMDB51 [23] have been used for evaluating in action recognition task. The ModelNet [56] dataset was also utilized to assess the performance of a 3D point cloud classifier. In addition to vision, we also consider attacks on natural language generation and use WMT2014 En-De [4] machine translation and OpenSubtitles2012 [48] dialogue generation datasets (Results are in the Supplementary). + +Attacks Configurations. Here, we first briefly overview the attack configurations on single-label image recognition datasets. We consider 14 state-of-the-art backdoor attacks: 1) Badnets [17], 2) Blend attack [9], 3 & 4) TrojanNet + +Table 2: Performance analysis for the multi-label backdoor attack [8]. We choose 3 object detection datasets [14,33] and ML-decoder [44] network architecture for this evaluation. Mean average precision (mAP) and ASR of the model, with and without defenses, have been shown. + +
DatasetNo defenseANPAWMRNPFT-SAMNFT (Ours)
ASRmAPASRmAPASRmAPASRmAPASRmAPASRmAP
VOC0786.492.521.786.926.687.319.287.619.386.817.389.1
VOC1284.891.918.685.319.085.913.886.414.687.114.288.4
MS-COCO85.688.019.784.122.683.417.184.319.283.816.685.8
+ +Table 3: Performance analysis for action recognition tasks where we choose 2 video datasets for evaluation. We consider a clean-label attack [64], where we need to generate adversarial perturbations for each input frame. + +
DatasetNo defenseI-BAUAWMRNPFT-SAMNFT (Ours)
ASRACCASRACCASRACCASRACCASRACCASRACC
UCF-10181.375.620.470.620.870.117.070.315.971.613.371.2
HMDB-5180.245.017.541.115.240.912.640.410.841.79.440.8
+ +(Troj-one & Troj-all) [35], 5) Sinusoidal signal attack (SIG) [2], 6 & 7) Input-Aware Attack (Dyn-one and Dyn-all) [40], 8) Clean-label attack (CLB) [50], 9) Composite backdoor (CBA) [32], 10) Deep feature space attack (FBA) [10], 11) Warping-based backdoor attack (WaNet) [39], 12) Invisible triggers based backdoor attack (ISSBA) [31], 13) Imperceptible backdoor attack (LIRA) [12], and 14) Quantization and contrastive learning-based attack (BPPA) [53]. In order to facilitate a fair comparison, we adopt trigger patterns and settings similar to those used in the original papers. Specifically, for both Trojan-one and Dyn-one attacks, we set all triggered images to have the same target label (i.e., all2one), whereas, for Trojan-all and Dyn-all attacks, we have uniformly distributed the target labels across all classes (i.e., all2all). Details on the hyper-parameters and overall training settings can be found in the Supplementary. We measure the success of an attack using two metrics: clean test accuracy (ACC) defined as the percentage of clean samples that are classified to their original target label and attack success rate (ASR) defined as the percentage of poison test samples ( $\hat{x}$ ) that are classified to the target label ( $\hat{y}$ ). + +Defenses Configurations. We compare our approach with 10 existing backdoor mitigation methods: 1) $FT-SAM$ [66]; 2) Adversarial Neural Pruning (ANP) [54]; 3) Implicit Backdoor Adversarial Unlearning ( $I-BAU$ ) [59]; 4) Adversarial Weight Masking ( $AWM$ ) [6]; 5) Reconstructive Neuron Pruning (RNP) [29]; 6) Fine-Pruning ( $FP$ ) [36]; 7) Mode Connectivity Repair ( $MCR$ ) [63]; 8) Neural Attention Distillation ( $NAD$ ) [28], 9) Causality-inspired Backdoor Defense ( $CBD$ ) [62]), 10) Anti Backdoor Learning ( $ABL$ ) [27]. In the main paper, we compare NFT with the first 5 defenses as they are more relevant, and the comparison with the rest of the methods is in Supplementary. To apply NFT, we + +Table 4: Removal performance (\%) of NFT against backdoor attacks on 3D point cloud classifiers. The attack methods [26], namely Poison-Label Backdoor Attack (PointPBA) with interaction trigger (PointPBA-I), PointPBA with orientation trigger (PointPBA-O), and Clean-Label Backdoor Attack (PointCBA) were considered, as well as the "backdoor points" based attack (3DPC-BA) outlined in prior work [58]. + +
AttackNo DefenseANPAWMRNPFT-SAMNFT (Ours)
ASRACCASRACCASRACCASRACCASRACCASRACC
PointBA-I98.689.113.682.615.483.98.184.08.884.59.685.7
PointBA-O94.789.814.882.013.182.49.483.88.285.07.585.3
PointCBA66.088.721.283.321.583.818.684.620.384.719.486.1
3DPC-BA93.891.216.884.715.685.913.985.713.186.312.687.7
+ +take $1\%$ clean validation data (set aside from the training set) and fine-tune the model for 100 epochs. An SGD optimizer has been employed with a learning rate of 0.05 and a momentum of 0.95. The rest of the experimental details for NFT and other defense methods are in the Supplementary. + +# 5.2 Performance Comparison of NFT + +In this section, we compare the performance of NFT with other defenses in various scenarios: single-label (i.e., image classification), multi-label (i.e., object detection), video action recognition, and 3D point cloud. + +Single-Label Backdoor attack. In Table 1, we present the performance of different defenses for 2 widely used benchmarks. For CIFAR10, we consider 4 label poisoning attacks: Badnets, Blend, Trojan, and Dynamic. For all of these attacks, NFT shows significant performance improvement over other baseline methods. While ANP and AWM defenses work well for mild attacks with low poison rates (e.g., $5\%$ ), the performance deteriorates for attacks with high poison rates (e.g., $\geq 10\%$ ). It is observable that NFT performs well across all attack scenarios, e.g., obtaining a $99.69\%$ drop in ASR for blend attack while also performing well for clean data (only $0.94\%$ of ACC drop). For Trojan and Dynamic attacks, we consider two different versions of attacks based on label-mapping criteria (all2one and all2all). The drop in attack success rate shows the effectiveness of NFT against such attacks. Recently, small and imperceptible perturbations as triggers have been developed in attacks (e.g., WaNet, LIRA, etc.) to fool the defense systems. While AWM generates perturbations as a proxy for these imperceptible triggers, NFT does a better job in this regard. For further validation of our proposed method, we use deep feature-based attacks CBA and FBA. Both of these attacks manipulate features to insert backdoor behavior. Overall, we achieve an average drop of $95.56\%$ in ASR while sacrificing an ACC of $1.81\%$ . For the scalability test, we consider a large and widely used dataset in vision tasks, ImageNet. In consistency with other datasets, NFT also obtains SOTA performance in this particular dataset. Due to page constraints, we move the performance comparison for GTSRB and Tiny-ImageNet to the Supplementary. + +![](images/739caefc6cc6f472a3e3fa20536eceebcdcc00859c38125ba0e0bdc9ce58bd04.jpg) +Fig.2: Ablation with different Mask Scheduling Function $(\mu)$ + +![](images/0a630b1ad7cd024747d42e8992d218f11393921131a29a5c2041a784a595ed3a.jpg) +Fig.3: Mask Distribution of AWM (left) and NFT (right). + +![](images/33c18e501e4c85b3e148d7a67d4a97b2dcb5ee0ed683ec77f0ed7c64e534259a.jpg) + +![](images/75b5989a31239d4f1ac6b0417f212a1d00d58ffa8377f9dcaa9aded1fd44feba.jpg) + +Multi-label Backdoor Attack. In Table 2, we also evaluate our proposed method on multi-label clean-image backdoor attack [8]. In general, we put a trigger on the image and change the corresponding ground truth of that image. However, a certain combination of objects has been used as a trigger pattern. For example, if a combination of car, person, and truck is present in the image, it will fool the model into misclassifying it. Table 2 shows that our proposed method surpasses other defense strategies concerning both ASR and mAP metrics. Notably, ANP and AWM, which rely on adversarial search limited applicability in multi-label scenarios. This limitation arises from the less accurate process of approximating triggers for object detection. Conversely, FT-SAM's optimization driven by sharpness proves effective in removing backdoors, yet it achieves a lower mAP post-purification. This outcome is not ideal since the goal is to eliminate backdoors without significantly compromising clean accuracy. + +Action Recognition Model. We further consider attacks on action recognition models; results are reported in Table 3. We use two widely used datasets, UCF-101 [46] and HMDB51 [23], with a CNN+LSTM network architecture. An ImageNet pre-trained ResNet50 network has been used for the CNN, and a sequential input-based Long Short Term Memory (LSTM) [45] network has been put on top of it. We subsample the input video by keeping one out of every 5 frames and use a fixed frame resolution of $224 \times 224$ . We choose a trigger size of $20 \times 20$ . Following [64], we create the required perturbation for clean-label attack by running projected gradient descent (PGD) [37] for 2000 steps with a perturbation norm of $\epsilon = 16$ . Note that our proposed augmentation strategies for image classification are directly applicable to video recognition. During training, we keep $5\%$ samples from each class to use them later as the clean validation set. Table 3 shows that NFT outperforms other defenses by a significant margin. In the case of a high number of classes and multiple image frames in the same input, it is a challenging task to optimize for the proper trigger pattern through the adversarial search described in I-BAU and AWM. Without a good approximation of the trigger, these methods seem to underperform in most cases. + +3D Point Cloud. In this phase of our work, we assess NFT's resilience against attacks on 3D point cloud classifiers [26,58]. To evaluate, we utilize the ModelNet dataset [56] and the PointNet++ architecture [42]. The performance comparison of NFT and other defense methods is outlined in Table 4. NFT outperforms other defenses due to its unique formulations of the objective function. + +Table 5: Average runtime of different defense methods against all 14 attacks. An NVIDIA RTX3090 GPU is used. + +
Method|ANP I-BAU AWM FT-SAM RNP NFT (Ours)
Runtime (sec.)118.192.5112.598.7102.628.4
+ +Table 6: Sensitivity analysis of $\alpha$ and $\beta$ for LIRA on CIFAR10. + +
α0.90.90.90.80.80.80.70.70.7
β0.250.500.750.250.500.750.250.500.75
ASR3.653.424.872.731.533.764.184.925.23
ACC88.5288.3489.1289.9790.5789.3487.6487.7888.10
+ +# 5.3 Ablation Study + +For all ablation studies, we consider the CIFAR10 dataset. + +Runtime Analysis. For runtime analysis, we present the training time for different defenses in Table 5. ANP and AWM both employ computationally expensive adversarial search procedures to prune neurons, which makes them almost $4\mathrm{x}$ slower than our method. However, NFT offers a computationally less expensive defense with SOTA performance in major benchmarks. + +Choice of Scheduling Function, $\mu$ . For choosing the suitable function for $\mu$ , we conduct a detailed study with commonly used mathematical functions. Note that, we only consider a family of functions that decreases over the depth of the network (shown in Fig. 2). This allows more variations for deeper layer weights, making sense as they are more responsible for DNN decision-making. In our work, we choose an exponential formulation for $\mu$ as it offers superior performance. We also perform sensitivity analysis of scheduling parameters $\alpha$ and $\beta$ in Table 6. In Fig. 3, we show the generated mask distributions of AWM and NFT. Compared to AWM, NFT produces more uniformly distributed masks that seem helpful for backdoor purification. We show the impact of $\mu$ in Table 7. Nature of Optimization. In Table 7, we present the performance of SOTA techniques under different validation sizes. Even with 10 samples (single-shot), NFT performs reasonably well and offers better performance as compared to AWM. This again shows that the trigger generation process is less accurate and effective for a very small validation set. We also show the effect of the proposed mask regularizer that indirectly controls the change in weights for better ACC. Although AWM employs a similar $\ell_1$ regularizer for masks, our proposed regularizer is more intuitive and specifically designed for better ACC preservation. While AWM encourages sparse solutions for $M$ (shown in the left subfigure of Fig. 3), it helps with ASR but heavily compromises ACC. We also show the performance of NFT without augmentations. + +Label Correction Rate. In the standard backdoor removal metric, it is sufficient for backdoored images to be classified as a non-target class (any class other than $\hat{y}$ ). However, we also consider another metric, label correction rate (LCR), for quantifying the success of a defense. We define LCR as the percentage of poisoned samples correctly classified to their original classes. Any method with the highest value of LCR is considered to be the best defense method. For this evaluation, we use CIFAR10 dataset and 6 backdoor attacks. Initially, the correction rate is $0\%$ with no defense as the ASR is close to $100\%$ . Table 8 shows that NFT obtains better performance in terms of LCR. + +Table 7: Purification performance (\%) for various validation data sizes. NFT performs reasonably well even with as few as 10 samples, i.e., one sample (shot) per class for CIFAR10. We also show the impact of the mask regularizer, mask scheduling function $\mu$ , and augmentations on performance, which resonates with Fig. 1. Mask regularizer has the most impact on the clean test accuracy (around $7\%$ worse without the regularizer). Without strong augmentations, we have a better ACC with a slightly worse ASR (around $6\%$ drop). + +
AttackDynamicWaNetLIRA
Samples101001010010100
MethodASRACCASRACCASRACCASRACCASRACCASRACC
No Defense10092.5210092.5298.6492.2998.6492.2999.2592.1599.2592.15
AWM86.7455.739.1685.3383.0162.217.2384.3891.4566.6410.8385.87
FT-SAM8.3573.495.7284.709.3575.985.5686.6311.8372.404.8588.82
NFT w/o Reg.5.6776.741.3682.214.1876.723.0283.314.8374.582.3283.61
NFT w/o Aug.11.9181.8610.5989.5310.3683.107.8189.6812.2381.059.1688.74
NFT w/o μ(l)5.1180.323.0488.585.8582.464.6488.026.4881.944.3388.75
NFT4.8380.511.7290.084.4183.582.9689.155.1882.722.0489.34
+ +Table 8: Label Correction Rate (%) for different defense techniques, defined as the percentage of backdoor samples that are correctly classified to their original ground truth label. + +
DefenseBadnetsTrojanBlendSIGCLBWaNet
No Defense000000
ANP84.7480.5281.3853.3582.7280.23
I-BAU78.4177.1277.5639.4678.0780.65
AWM79.3778.2479.8144.5179.8679.18
FT-SAM85.5680.6984.4957.6482.0483.62
NFT (Ours)86.8281.1585.6155.1886.2385.70
+ +# 6 Conclusion + +We proposed a backdoor purification framework, NFT, utilizing an augmentation-based neural mask fine-tuning approach. NFT can change the backdoor model weights in a computationally efficient manner while ensuring SOTA purification performance. Our proposed method showed that the addition of MixUp during fine-tuning replaces the need for a computationally expensive trigger synthesis process. Furthermore, we proposed a novel mask regularizer that helps us preserve the cluster separability of the original backdoor model. By preserving this separability, the proposed method offers better clean test accuracy compared to SOTA methods. Furthermore, we suggested using a mask scheduling function that reduces the mask search space and improves the computational efficiency further. Our extensive experiments on 5 different tasks validate the efficiency and efficacy of the proposed backdoor purification method. We also conducted a detailed ablation study to explain the reasoning behind our design choices. + +# Acknowledgements + +This work was supported in part by the National Science Foundation under Grant ECCS-1810256 and CMMI-2246672. + +# References + +1. Ahmed, S., Al Arafat, A., Rizve, M.N., Hossain, R., Guo, Z., Rakin, A.S.: Ssda: Secure source-free domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19180-19190 (2023) +2. Barni, M., Kallas, K., Tondi, B.: A new backdoor attack in cnns by training set corruption without label poisoning. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 101-105. IEEE (2019) +3. Bian, J., Al Arafat, A., Xiong, H., Li, J., Li, L., Chen, H., Wang, J., Dou, D., Guo, Z.: Machine learning in real-time internet of things (iot) systems: A survey. IEEE Internet of Things Journal 9(11), 8364-8386 (2022) +4. Bojar, O., Buck, C., Federmann, C., Haddow, B., Koehn, P., Leveling, J., Monz, C., Pecina, P., Post, M., Saint-Amand, H., Soricut, R., Specia, L., Tamchyna, A.: Findings of the 2014 workshop on statistical machine translation. In: Proceedings of the Ninth Workshop on Statistical Machine Translation. pp. 12-58. Association for Computational Linguistics, Baltimore, Maryland, USA (Jun 2014). https://doi.org/10.3115/v1/W14-3302, https://aclanthology.org/W14-3302 +5. Carratino, L., Cisse, M., Jenatton, R., Vert, J.P.: On mixup regularization. The Journal of Machine Learning Research 23(1), 14632-14662 (2022) +6. Chai, S., Chen, J.: One-shot neural backdoor erasing via adversarial weight masking. arXiv preprint arXiv:2207.04497 (2022) +7. Chen, H., Fu, C., Zhao, J., Koushanfar, F.: Deepinspect: A black-box trojan detection and mitigation framework for deep neural networks. In: IJCAI. p. 8 (2019) +8. Chen, K., Lou, X., Xu, G., Li, J., Zhang, T.: Clean-image backdoor: Attacking multi-label models with poisoned labels only. In: The Eleventh International Conference on Learning Representations (2023) +9. Chen, X., Liu, C., Li, B., Lu, K., Song, D.: Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526 (2017) +10. Cheng, S., Liu, Y., Ma, S., Zhang, X.: Deep feature space trojan attack of neural networks by controlled detoxification. In: AAAI. pp. 1148-1156 (2021) +11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255. IEEE (2009) +12. Doan, K., Lao, Y., Zhao, W., Li, P.: Lira: Learnable, imperceptible and robust backdoor attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11966-11976 (2021) +13. Edraki, M., Karim, N., Rahnavard, N., Mian, A., Shah, M.: Odyssey: Creation, analysis and detection of trojan models. IEEE Transactions on Information Forensics and Security 16, 4521-4533 (2021) +14. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010) +15. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=6Tm1mposlrM + +16. Gao, Y., Xu, C., Wang, D., Chen, S., Ranasinghe, D.C., Nepal, S.: Strip: A defence against trojan attacks on deep neural networks. In: Proceedings of the 35th Annual Computer Security Applications Conference. pp. 113-125 (2019) +17. Gu, T., Liu, K., Dolan-Gavitt, B., Garg, S.: Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access 7, 47230-47244 (2019) +18. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: European conference on computer vision. pp. 630-645. Springer (2016) +19. Hong, S., Chandrasekaran, V., Kaya, Y., Dumitras, T., Papernot, N.: On the effectiveness of mitigating data poisoning attacks with gradient shaping. arXiv preprint arXiv:2002.11497 (2020) +20. Janai, J., Güney, F., Behl, A., Geiger, A., et al.: Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends in Computer Graphics and Vision 12(1-3), 1-308 (2020) +21. Karim, N., Arafat, A.A., Rakin, A.S., Guo, Z., Rahnavard, N.: Fisher information guided purification against backdoor attacks. In: Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (CCS) (2024) +22. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +23. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011) +24. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) +25. Levine, A., Feizi, S.: Deep partition aggregation: Provable defense against general poisoning attacks. arXiv preprint arXiv:2006.14768 (2020) +26. Li, X., Chen, Z., Zhao, Y., Tong, Z., Zhao, Y., Lim, A., Zhou, J.T.: Pointba: Towards backdoor attacks in 3d point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16492-16501 (2021) +27. Li, Y., Lyu, X., Koren, N., Lyu, L., Li, B., Ma, X.: Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems 34 (2021) +28. Li, Y., Lyu, X., Koren, N., Lyu, L., Li, B., Ma, X.: Neural attention distillation: Erasing backdoor triggers from deep neural networks. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=910K40M-oXE +29. Li, Y., Lyu, X., Ma, X., Koren, N., Lyu, L., Li, B., Jiang, Y.G.: Reconstructive neuron pruning for backdoor defense. arXiv preprint arXiv:2305.14876 (2023) +30. Li, Y., Wu, B., Jiang, Y., Li, Z., Xia, S.T.: Backdoor learning: A survey. arXiv preprint arXiv:2007.08745 (2020) +31. Li, Y., Li, Y., Wu, B., Li, L., He, R., Lyu, S.: Invisible backdoor attack with sample-specific triggers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16463-16472 (2021) +32. Lin, J., Xu, L., Liu, Y., Zhang, X.: Composite backdoor attack for deep neural network by mixing existing benign features. In: CCS. pp. 113-131 (2020) +33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +34. Liu, K., Dolan-Gavitt, B., Garg, S.: Fine-pruning: Defending against backdooring attacks on deep neural networks. In: International Symposium on Research in Attacks, Intrusions, and Defenses. pp. 273-294. Springer (2018) + +35. Liu, Y., Ma, S., Afer, Y., Lee, W.C., Zhai, J., Wang, W., Zhang, X.: Trojaning attack on neural networks (2017) +36. Liu, Y., Xie, Y., Srivastava, A.: Neural trojans. In: 2017 IEEE International Conference on Computer Design (ICCD). pp. 45-48. IEEE (2017) +37. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) +38. Manoj, N., Blum, A.: Excess capacity and backdoor poisoning. Advances in Neural Information Processing Systems 34, 20373-20384 (2021) +39. Nguyen, A., Tran, A.: Wanet-imperceptible warping-based backdoor attack. arXiv preprint arXiv:2102.10369 (2021) +40. Nguyen, T.A., Tran, A.: Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems 33, 3454-3464 (2020) +41. Ning, R., Li, J., Xin, C., Wu, H.: Invisible poison: A blackbox clean label backdoor attack to deep neural networks. In: IEEE INFOCOM 2021-IEEE Conference on Computer Communications. pp. 1-10. IEEE (2021) +42. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +43. Qiu, H., Zeng, Y., Guo, S., Zhang, T., Qiu, M., Thuraisingham, B.: Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. In: Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security. pp. 363-377 (2021) +44. Ridnik, T., Sharir, G., Ben-Cohen, A., Ben-Baruch, E., Noy, A.: ML-decoder: Scalable and versatile classification head. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 32-41 (2023) +45. Sherstinsky, A.: Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. Physica D: Nonlinear Phenomena 404, 132306 (2020) +46. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) +47. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: The german traffic sign recognition benchmark: a multi-class classification competition. In: The 2011 international joint conference on neural networks. pp. 1453-1460. IEEE (2011) +48. Tiedemann, J.: Parallel data, tools and interfaces in opus. In: Lrec. vol. 2012, pp. 2214-2218. Citeseer (2012) +49. Tran, B., Li, J., Madry, A.: Spectral signatures in backdoor attacks. Advances in neural information processing systems 31 (2018) +50. Turner, A., Tsipras, D., Madry, A.: Clean-label backdoor attacks (2019), https://openreview.net/forum?id=HJg6e2CcK7 +51. Wang, B., Yao, Y., Shan, S., Li, H., Viswanath, B., Zheng, H., Zhao, B.Y.: Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In: 2019 IEEE Symposium on Security and Privacy (SP). pp. 707-723. IEEE (2019) +52. Wang, R., Zhang, G., Liu, S., Chen, P.Y., Xiong, J., Wang, M.: Practical detection of trojan neural networks: Data-limited and data-free cases. In: European Conference on Computer Vision. pp. 222-238. Springer (2020) +53. Wang, Z., Zhai, J., Ma, S.: Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15074-15084 (2022) +54. Wu, D., Wang, Y.: Adversarial neuron pruning purifies backdoored deep models. In: NeurIPS (2021) + +55. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems 33, 2958-2969 (2020) +56. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015) +57. Xian, X., Wang, G., Srinivasa, J., Kundu, A., Bi, X., Hong, M., Ding, J.: Understanding backdoor attacks through the adaptability hypothesis. In: Proc. International Conference on Machine Learning (2023) +58. Xiang, Z., Miller, D.J., Chen, S., Li, X., Kesidis, G.: A backdoor attack against 3d point cloud classifiers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7597-7607 (2021) +59. Zeng, Y., Chen, S., Park, W., Mao, Z.M., Jin, M., Jia, R.: Adversarial unlearning of backdoors via implicit hypergradient. arXiv preprint arXiv:2110.03735 (2021) +60. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) +61. Zhang, L., Deng, Z., Kawaguchi, K., Ghorbani, A., Zou, J.: How does mixup help with robustness and generalization? arXiv preprint arXiv:2010.04819 (2020) +62. Zhang, Z., Liu, Q., Wang, Z., Lu, Z., Hu, Q.: Backdoor defense via deconfounded representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12228-12238 (2023) +63. Zhao, P., Chen, P.Y., Das, P., Ramamurthy, K.N., Lin, X.: Bridging mode connectivity in loss landscapes and adversarial robustness. arXiv preprint arXiv:2005.00060 (2020) +64. Zhao, S., Ma, X., Zheng, X., Bailey, J., Chen, J., Jiang, Y.G.: Clean-label backdoor attacks on video recognition models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14443-14452 (2020) +65. Zheng, R., Tang, R., Li, J., Liu, L.: Data-free backdoor removal based on channel lipschitzness. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V. pp. 175-191. Springer (2022) +66. Zhu, M., Wei, S., Shen, L., Fan, Y., Wu, B.: Enhancing fine-tuning based backdoor defense with sharpness-aware minimization. arXiv preprint arXiv:2304.11823 (2023) \ No newline at end of file diff --git a/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/images.zip b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1d245ae96a71f5e0b21cfc3ce14a438c3c9d6dc0 --- /dev/null +++ b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02cbbda18a89b76b22427d009ac635332c55b3f811ceeb8865f2d4d02e184be5 +size 572548 diff --git a/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/layout.json b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ffbcc18a682435fe7fe35ba5e5b165e174c85f6e --- /dev/null +++ b/2024/Augmented Neural Fine-tuning for Efficient Backdoor Purification/layout.json @@ -0,0 +1,11286 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 30, + 360, + 64 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 30, + 360, + 64 + ], + "spans": [ + { + "bbox": [ + 52, + 30, + 360, + 64 + ], + "type": "text", + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 33, + 86, + 378, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 86, + 378, + 110 + ], + "spans": [ + { + "bbox": [ + 33, + 86, + 378, + 110 + ], + "type": "text", + "content": "Nazmul Karim\\*,1, Abdullah Al Arafat\\*,2, Umar Khalid\\* Zhishan Guo\\* and Nazanin Rahnavard\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 71, + 119, + 340, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 119, + 340, + 131 + ], + "spans": [ + { + "bbox": [ + 71, + 119, + 340, + 131 + ], + "type": "text", + "content": "1University of Central Florida 2North Carolina State University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 161, + 351, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 351, + 457 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 351, + 457 + ], + "type": "text", + "content": "Abstract. Recent studies have revealed the vulnerability of deep neural networks (DNNs) to various backdoor attacks, where the behavior of DNNs can be compromised by utilizing certain types of triggers or poisoning mechanisms. State-of-the-art (SOTA) defenses employ too-sophisticated mechanisms that require either a computationally expensive adversarial search module for reverse-engineering the trigger distribution or an over-sensitive hyper-parameter selection module. Moreover, they offer sub-par performance in challenging scenarios, e.g., limited validation data and strong attacks. In this paper, we propose—Neural mask Fine-Tuning (NFT)—with an aim to optimally re-organize the neuron activities in a way that the effect of the backdoor is removed. Utilizing a simple data augmentation like MixUp, NFT relaxes the trigger synthesis process and eliminates the requirement of the adversarial search module. Our study further reveals that direct weight fine-tuning under limited validation data results in poor post-purification clean test accuracy, primarily due to overfitting issue. To overcome this, we propose to fine-tune neural masks instead of model weights. In addition, a mask regularizer has been devised to further mitigate the model drift during the purification process. The distinct characteristics of NFT render it highly efficient in both runtime and sample usage, as it can remove the backdoor even when a single sample is available from each class. We validate the effectiveness of NFT through extensive experiments covering the tasks of image classification, object detection, video action recognition, 3D point cloud, and natural language processing. We evaluate our method against 14 different attacks (LIRA, WaNet, etc.) on 11 benchmark data sets (ImageNet, UCF101, Pascal VOC, ModelNet, OpenSubtitles2012, etc.). Our code is available online in this GitHub Repository." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 479, + 128, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 479, + 128, + 492 + ], + "spans": [ + { + "bbox": [ + 33, + 479, + 128, + 492 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 504, + 379, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 504, + 379, + 564 + ], + "spans": [ + { + "bbox": [ + 33, + 504, + 379, + 564 + ], + "type": "text", + "content": "Machine learning and computer vision algorithms are increasingly common in safety-critical applications [3, 20], necessitating the design of secure and robust learning algorithms. Backdoor attack [9,17] on deep neural network (DNN) models is one of the heavily studied branches of AI safety and robustness. Backdoor defenses can generally be categorized into two major groups based on whether" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 35, + 572, + 121, + 583 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 572, + 121, + 583 + ], + "spans": [ + { + "bbox": [ + 35, + 572, + 121, + 583 + ], + "type": "text", + "content": "* Equal Contribution" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 34, + 381, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 34, + 381, + 309 + ], + "spans": [ + { + "bbox": [ + 31, + 34, + 381, + 309 + ], + "type": "text", + "content": "the defense is done during training or test. Training time defense (e.g., [25,27,49]) focuses on training a benign model on the poisonous data, while test time defense (e.g., [6,29,34,54]) deals with purifying backdoor model after it has already been trained. In this work, our goal is to develop an efficient test-time defense. Some test-time defenses focus on synthesizing trigger patterns [7,51,52] followed by vanilla weight fine-tuning. Most of these defenses aim to synthesize class-specific triggers independently or use additional models to generate the triggers simultaneously. Recent state-of-the-art (SOTA) backdoor defense methods, e.g., ANP [54], I-BAU [59], AWM [6], employ very similar techniques and do not work well without an expensive adversarial search module. For example, ANP [54] performs an adversarial search to find vulnerable neurons responsible for backdoor behavior. Identifying and pruning vulnerable neurons require an exhaustive adversarial search, resulting in high computational costs. Similar to ANP, AWM and I-BAU also resort to trigger synthesizing with a modified adversarial search process. Another very recent technique, FT-SAM [66] uses sharpness-aware minimization (SAM) [15] to fine-tune model weights. SAM is a recently proposed optimizer that utilizes Stochastic Gradient Descent (SGD), which penalizes sudden changes in the loss surface by constraining the search area to a compact region. Since SAM performs a double forward pass to compute the loss gradient twice, it results in a notable runtime increase for FT-SAM. In our work, we aim to develop an effective backdoor defense system that neither requires an expensive adversarial search process to recover the trigger nor a special type of optimizer with a runtime bottleneck." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 321, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 321, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 321, + 381, + 584 + ], + "type": "text", + "content": "To achieve this goal, we propose a simple yet effective approach Neural mask Fine-Tuning (NFT), to remove backdoor through augmented fine-tuning of cost-efficient neural masks. We start with replacing the expensive adversarial search-based trigger synthesis process with a simple data augmentation technique—MixUp [60]. In general, the backdoor is inserted by forcing the model to memorize the trigger distribution. Intuitively, synthesizing and unlearning that trigger distribution would effectively remove the backdoor. In this work, we show that unlearning can be performed by simply optimizing the MixUp loss over a clean validation set. Our theoretical analysis suggests that MixUp loss is an upper bound on the standard loss obtained from triggered (synthesized or already known) validation data, termed as ideal purification loss. As the minimization of ideal loss guarantees backdoor purification, minimizing the MixUp loss would effectively remove the backdoor (Sec. 4.1). As the next step of our method, we address the overfitting issue during weight fine-tuning under limited validation data. In general, the outcome of such overfitting is poor post-purification test accuracy, which is not desirable for any backdoor defense. To this end, we propose to fine-tune a set of neural masks instead of the model weights, as this type of soft-masking enables us to reprogram the neurons affected by the backdoor without significantly altering the original backdoor model. As an added step to this, a mask regularizer has been introduced to further mitigate model drift during the purification process. In addition, we deploy a mask scheduling function to have better control over the purification process. Our experimental" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 39, + 19 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 128, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 128, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 128, + 19 + ], + "type": "text", + "content": "N. Karim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 380, + 70 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 380, + 70 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 380, + 70 + ], + "type": "text", + "content": "results indicate that these straightforward yet intuitive steps significantly improve the post-purification test accuracy as compared to previous SOTA. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 37, + 76, + 379, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 76, + 379, + 124 + ], + "spans": [ + { + "bbox": [ + 37, + 76, + 379, + 124 + ], + "type": "text", + "content": "- We propose a novel backdoor removal framework utilizing simple MixUp-based model fine-tuning. Our thorough analysis shows how minimizing the MixUp loss eliminates the requirement of an expensive trigger synthesis process while effectively removing the backdoor (Sec. 4.1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 37, + 124, + 380, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 124, + 380, + 219 + ], + "spans": [ + { + "bbox": [ + 37, + 124, + 380, + 219 + ], + "type": "text", + "content": "- To preserve the post-purification test accuracy, we propose to fine-tune soft neural masks (instead of weights) as it prevents any drastic change in the original backdoor model (Sec. 4.2). Additionally, a novel mask regularizer has been introduced that further encourages the purified model to retain the class separability of the original model. In addition to being computationally efficient, our proposed method shows significant improvement in sample efficiency as it can purify backdoor even with one-shot fine-tuning, i.e., only a single sample is available from each class (Sec. 4.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 220, + 380, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 220, + 380, + 268 + ], + "spans": [ + { + "bbox": [ + 37, + 220, + 380, + 268 + ], + "type": "text", + "content": "- To show the effectiveness of NFT, we perform an extensive evaluation with 11 different datasets. Compared to previous SOTA, the superior performance against a wide range of attacks suggests that augmentation like MixUp can indeed replace the trigger synthesis process (Sec. 5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 284, + 137, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 284, + 137, + 297 + ], + "spans": [ + { + "bbox": [ + 32, + 284, + 137, + 297 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 308, + 380, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 308, + 380, + 511 + ], + "spans": [ + { + "bbox": [ + 32, + 308, + 380, + 511 + ], + "type": "text", + "content": "Backdoor Attack. Neural networks are intrinsically vulnerable to backdoor attacks [38, 57]. A substantial number of studies have investigated the possibility of backdoor attacks after the initial studies [9, 17, 36] found the existence of backdoors in DNNs. Generally, backdoor attacks are categorized into two types: clean-label attacks and poison-label attacks. A clean-label backdoor attack does not alter the label [41, 50, 64], while a poison-label attack aims at specific target classes such that the DNN misclassifies to those classes in the presence of a trigger [28]. As for trigger types, researchers have studied numerous types of triggering patterns in their respective attacks [9, 13, 17, 30]. Such triggers can exist in the form of dynamic patterns [30] or as simple as a single pixel [49]. Some of the more complex backdoor triggers that have been proposed in the literature are sinusoidal strips [2], adversarial patterns [64], and blending backgrounds [9]. Besides, backdoor attacks exist for many different tasks, e.g., multi-label clean image attack [8] has been proposed that alters the label distribution to insert triggers into the model, which works well in multi-label (e.g., detection) settings; domain adaptation [1] setting while adversary source can successfully insert backdoor to the target domains, etc." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 512, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 512, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 32, + 512, + 380, + 584 + ], + "type": "text", + "content": "Backdoor Defense. Generally, the backdoor defense methods are categorized into two types: Training Time Defense, and Test Time Defense. Regarding training time defense techniques (a few to mention [16, 19, 36, 49]), the researchers have proposed numerous defense methods through input pre-processing [36], poison-suppression [19], model diagnosis [30], network pruning [36, 55], and model reconstruction [63], etc. Notably, DeepSweep [43] explores different augmentations to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 381, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 381, + 202 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 381, + 202 + ], + "type": "text", + "content": "purify a backdoor model and rectify the triggered samples. Although DeepSweep revealed that different augmentation functions could be leveraged to invert the backdoor effect of a model or erase the trigger from a trigger-embedded image, our work re-purposes the usage of augmentation differently to cover the approximate (unknown) trigger distribution during the purification phase. Moreover, DeepSweep assumes that backdoor triggers are known to the defender, which is hardly a practical assumption. In the case of test time defenses, besides the works mentioned in the introduction related to reverse-engineering of backdoor triggers in the input samples [7, 51, 52], several recent works explored the model vulnerability/sensitivity towards adversarially perturbed neurons [54], weights [6], or network channels [65]. However, these approaches require expensive adversarial search processes to be effective. A concurrent work FIP [21] studied the loss-surface smoothness of the backdoor model and developed a purification method by regularizing the spectral norm of the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 220, + 136, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 220, + 136, + 233 + ], + "spans": [ + { + "bbox": [ + 33, + 220, + 136, + 233 + ], + "type": "text", + "content": "3 Threat Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "spans": [ + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": "Attack Model. Our work considers the most commonly used data poisoning attacks. Consider " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{train}} = \\{x_i, y_i\\}_{i=1}^N" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": " as the training data where " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": " is an input sample labeled as " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "y_i \\in \\{0, \\dots, c-1\\}" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": " sampled from unknown distribution " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": " of the task to be learned. Here, " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": " is the total number of samples, and " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": " is the dimension of the input sample. In addition, we assume there are " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": " number of classes in the input data. Let " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "f_{\\theta^*}: \\mathbb{R}^d \\to \\mathbb{R}^c" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": " be a benign (ideal) DNN trained with " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{train}} \\sim \\mathcal{D}" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "\\theta^*" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": " is the DNN parameters that is to be optimized using a suitable loss function " + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "inline_equation", + "content": "\\ell(.,.)" + }, + { + "bbox": [ + 32, + 246, + 381, + 343 + ], + "type": "text", + "content": ". The total empirical loss can be defined as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 354, + 378, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 354, + 378, + 384 + ], + "spans": [ + { + "bbox": [ + 127, + 354, + 378, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\theta^ {*}, \\mathbb {D} _ {\\text {t r a i n}}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ \\ell \\left(y _ {i}, f _ {\\theta^ {*}} \\left(x _ {i}\\right)\\right) \\right], \\tag {1}", + "image_path": "1ef6aeaa2e9abbc376da091291ea79f0e6fead12e1a23aef68e485670f7ea471.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "spans": [ + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": "Now, consider an adversary inserts backdoor to a model " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "f_{\\theta}(.)" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": " through modifying a small subset of " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{train}}" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_i,\\hat{y}_i\\}" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "\\hat{y}_i = \\arg \\max f_\\theta (\\hat{x}_i)" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": " preserving " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "y_{i} = \\arg \\max f_{\\theta}(x_{i}),\\forall (x_{i},y_{i})\\in \\mathbb{D}_{\\mathrm{train}}." + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": " Here, " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "\\hat{x}_i = x_i + \\delta" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": " is the triggered input with adversary set target label " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "\\hat{y}_i\\neq y_i" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": " represents trigger pattern. Defense Objective. Consider a defense model where defender removes backdoor from " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "f_{\\theta}(.)" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": " using a small validation data " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{val}} = \\{x_i,y_i\\}_{i = 1}^{N_{val}}" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "y_{i} = \\arg \\max f_{\\theta_{c}}(\\hat{x}_{i})" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "inline_equation", + "content": "y_{i}\\neq \\hat{y}_{i}" + }, + { + "bbox": [ + 32, + 393, + 381, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 496, + 212, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 496, + 212, + 512 + ], + "spans": [ + { + "bbox": [ + 33, + 496, + 212, + 512 + ], + "type": "text", + "content": "4 Neural Fine-Tuning (NFT)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": "Let us consider a fully-connected DNN, " + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "inline_equation", + "content": "f_{\\theta}:\\mathbb{R}^{d}\\to \\mathbb{R}^{c}" + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": ", that receives a datapoint " + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "inline_equation", + "content": "x\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": " and predicts a probability distribution " + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "inline_equation", + "content": "p\\in \\mathbb{R}^c" + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": "; where " + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": " is the number of classes. In general, " + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": " goes through a multi-layer DNN architecture before the DNN model predicts an output class " + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "inline_equation", + "content": "i = \\arg \\max p" + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": ". Let us consider a multi-layer DNN architecture with " + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": " layers in which the " + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": "-th layer contains " + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "inline_equation", + "content": "k_{l}" + }, + { + "bbox": [ + 32, + 523, + 381, + 583 + ], + "type": "text", + "content": " neurons." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 39, + 18 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 128, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 128, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 128, + 19 + ], + "type": "text", + "content": "N. Karim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "spans": [ + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "text", + "content": "The neurons of each layer produce activations " + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "inline_equation", + "content": "\\psi_{l}\\in \\mathbb{R}^{k_{l}}" + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "text", + "content": " based on the output activations of previous layer " + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "inline_equation", + "content": "\\psi_{l - 1}\\in \\mathbb{R}^{k_{l - 1}}" + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "text", + "content": ". To be specific," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 65, + 379, + 78 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 65, + 379, + 78 + ], + "spans": [ + { + "bbox": [ + 154, + 65, + 379, + 78 + ], + "type": "interline_equation", + "content": "\\psi_ {l} := \\sigma \\left(\\Theta_ {l} ^ {T} \\cdot \\psi_ {l - 1} + b _ {l}\\right), \\tag {2}", + "image_path": "43da5d98ff1527d7edcfb46fb967babfb26c0be2a330932e012a3d40ace88ae4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "spans": [ + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "\\sigma(.)" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": " is a non-linear activation function, matrix " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "\\Theta_l = [\\Theta_l^{(1)}\\dots \\Theta_l^{(k_l)}]\\in \\mathbb{R}^{k_{l - 1}\\times k_l}" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "l = 1,2,\\ldots ,L" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": ", includes the weights of the " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": "-th layer, and " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "b_{l}\\in \\mathbb{R}^{k_{l}}" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": " is the bias vector. Here, " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "\\Theta_l^{(j)}\\in \\mathbb{R}^{k_{l - 1}}" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": " denotes the weights vector corresponding to the activation of the " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": "-th neuron of the " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": "-th layer. Model parameters " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": " can be expressed as " + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\Theta_1,\\dots ,\\Theta_L\\}" + }, + { + "bbox": [ + 30, + 84, + 381, + 172 + ], + "type": "text", + "content": ". This type of multi-layer DNN architecture is also valid for convolutional neural networks, where we use multiple 2-D arrays of neurons (i.e., filters) instead of a 1-D array." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 187, + 167, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 187, + 167, + 200 + ], + "spans": [ + { + "bbox": [ + 31, + 187, + 167, + 200 + ], + "type": "text", + "content": "4.1 Backdoor Suppressor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "spans": [ + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": "Our objective is to make the backdoor model forget about poison distribution while retaining the knowledge of clean distribution. To understand how we achieve this objective, let us first revisit the process of generating triggered data " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "(\\hat{x})" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": ". In general, " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": " is created by adding minor modifications (i.e., adding triggers " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": ") to clean data " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "(x)" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": ". Note that the backdoor is inserted by forcing the model to learn the mapping, " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "\\hat{x} \\rightarrow \\hat{y}" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "(\\hat{x}, \\hat{y})" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": " is the poison data. If we were to change the mapping from " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "(\\hat{x} \\rightarrow \\hat{y})" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "(\\hat{x} \\rightarrow y)" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": ", we would have a robust clean model instead of a backdoor model. This is because, in this case, the model would treat " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": " as one type of augmentation and " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": " as the augmented clean data. In summary, the backdoor insertion process functions as an augmentation process if we simply use " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": ". Now, ideally, fine-tuning the backdoor model with triggered data with corresponding ground truth labels (i.e., " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}, y\\}" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": ") would remove the backdoor effect from the model. Let us consider this ideal scenario where we have access to the trigger " + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 30, + 204, + 381, + 384 + ], + "type": "text", + "content": " during purification. We define the ideal purification loss as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 124, + 392, + 379, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 392, + 379, + 422 + ], + "spans": [ + { + "bbox": [ + 124, + 392, + 379, + 422 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {i d e a l}} (\\theta , \\mathbb {D} _ {\\text {v a l}}) = \\frac {1}{N _ {\\text {v a l}}} \\sum_ {i = 1} ^ {N _ {\\text {v a l}}} \\ell \\left(y _ {i}, f _ {\\theta} \\left(\\hat {x} _ {i}\\right)\\right), \\tag {3}", + "image_path": "4a4942016ec63f9ed124efe685c893e047bde6898f2dc4f7424d9afb63e5ec6f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "spans": [ + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "\\hat{x} = x + \\delta" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": " is the ground truth label of " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": ". In our work, as we do not have access to " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": " or adversarially reverse engineer it [6,54,59], we relax this process by strongly augmenting the clean validation data, i.e., creating augmented " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{val}}" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": " with already known augmentation technique such as MixUp [60]. For MixUp, we can easily perform " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{i,j} = \\lambda x_i + (1 - \\lambda)x_j" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{i,j} = \\lambda y_i + (1 - \\lambda)y_j" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": "; here " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{i,j}" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": " represents the linear combination of one-hot vectors corresponding to " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "inline_equation", + "content": "y_j" + }, + { + "bbox": [ + 30, + 428, + 381, + 514 + ], + "type": "text", + "content": ". The loss after the MixUp becomes:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 520, + 379, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 520, + 379, + 552 + ], + "spans": [ + { + "bbox": [ + 108, + 520, + 379, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {m i x}} (\\theta , \\mathbb {D} _ {\\text {v a l}}) = \\frac {1}{N _ {\\text {v a l}} ^ {2}} \\sum_ {i, j = 1} ^ {N _ {\\text {v a l}}} \\underset {\\lambda \\sim \\mathcal {D} _ {\\lambda}} {\\mathbb {E}} \\ell \\left(\\tilde {y} _ {i, j}, f _ {\\theta} \\left(\\tilde {x} _ {i, j}\\right)\\right), \\tag {4}", + "image_path": "9cf89985ec911089dd5b6bb17fcf579f2c8bc1fc02440e4c507fc4e56baaf3b1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\lambda}" + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "text", + "content": " is a distribution supported on " + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "text", + "content": ". In our work, we consider the widely used " + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\lambda}" + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "text", + "content": " - Beta distribution " + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "inline_equation", + "content": "Beta(\\alpha, \\beta)" + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta > 0" + }, + { + "bbox": [ + 31, + 559, + 381, + 584 + ], + "type": "text", + "content": ". We provide both" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "type": "text", + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 379, + 59 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 379, + 59 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 379, + 59 + ], + "type": "text", + "content": "empirical (Sec. 5) and theoretical proof for a binary classification problem on why minimizing Eq. (4) would effectively remove the backdoor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "spans": [ + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": "Theoretical Justifications. For a fully-connected neural network (NN) with logistic loss " + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\ell(y, f_{\\theta}(x)) = \\log(1 + \\exp(f_{\\theta}(x))) - yf_{\\theta}(x)" + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "inline_equation", + "content": "y \\in \\{0, 1\\}" + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": ", it can be shown that " + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})" + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": " is an upper-bound of the second order Taylor expansion of the ideal loss " + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{ideal}}(\\theta, \\mathbb{D}_{\\mathrm{val}})" + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": ". With the nonlinearity " + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": " for ReLU and max-pooling in NN, the function " + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": " satisfies that " + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x) = \\nabla f_{\\theta}(x)^T x" + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\nabla^2 f_{\\theta}(x) = 0" + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": " almost everywhere, where the gradient is taken with respect to the input " + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 30, + 59, + 381, + 130 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 130, + 381, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 130, + 381, + 154 + ], + "spans": [ + { + "bbox": [ + 30, + 130, + 381, + 154 + ], + "type": "text", + "content": "We first rewrite the " + }, + { + "bbox": [ + 30, + 130, + 381, + 154 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{ideal}}(\\theta, \\mathbb{D}_{\\mathrm{val}})" + }, + { + "bbox": [ + 30, + 130, + 381, + 154 + ], + "type": "text", + "content": " using Taylor series approximation. The second-order Taylor expansion of " + }, + { + "bbox": [ + 30, + 130, + 381, + 154 + ], + "type": "inline_equation", + "content": "\\ell(y, f_{\\theta}(x + \\delta))" + }, + { + "bbox": [ + 30, + 130, + 381, + 154 + ], + "type": "text", + "content": " is given by," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 161, + 387, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 161, + 387, + 184 + ], + "spans": [ + { + "bbox": [ + 31, + 161, + 387, + 184 + ], + "type": "interline_equation", + "content": "\\ell (y, f _ {\\theta} (x + \\delta)) = \\ell (y, f _ {\\theta} (x)) + (g (f _ {\\theta} (x)) - y) (f _ {\\theta} (\\delta)) + \\frac {1}{2} g (f _ {\\theta} (x)) (1 - g (f _ {\\theta} (x))) (f _ {\\theta} (\\delta)) ^ {2},", + "image_path": "9837cbc5a85eb080d14e29f6ecf7460297780e5971bf27d57f1d082ed8065936.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 190, + 381, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 190, + 381, + 228 + ], + "spans": [ + { + "bbox": [ + 30, + 190, + 381, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 190, + 381, + 228 + ], + "type": "inline_equation", + "content": "g(x) = \\frac{e^x}{1 + e^x}" + }, + { + "bbox": [ + 30, + 190, + 381, + 228 + ], + "type": "text", + "content": " is the logistic function. Based on the MixUp related analysis in prior works [5, 61], the following can be derived for " + }, + { + "bbox": [ + 30, + 190, + 381, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})" + }, + { + "bbox": [ + 30, + 190, + 381, + 228 + ], + "type": "text", + "content": " using the second-order Taylor series expansion," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 233, + 381, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 233, + 381, + 268 + ], + "spans": [ + { + "bbox": [ + 31, + 233, + 381, + 268 + ], + "type": "text", + "content": "Lemma 1. Assuming " + }, + { + "bbox": [ + 31, + 233, + 381, + 268 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x) = \\nabla f_{\\theta}(x)^{T}x" + }, + { + "bbox": [ + 31, + 233, + 381, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 233, + 381, + 268 + ], + "type": "inline_equation", + "content": "\\nabla^2 f_{\\theta}(x) = 0" + }, + { + "bbox": [ + 31, + 233, + 381, + 268 + ], + "type": "text", + "content": " (which are satisfied by ReLU and max-pooling activation functions), " + }, + { + "bbox": [ + 31, + 233, + 381, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})" + }, + { + "bbox": [ + 31, + 233, + 381, + 268 + ], + "type": "text", + "content": " can be expressed as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 268, + 379, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 268, + 379, + 284 + ], + "spans": [ + { + "bbox": [ + 89, + 268, + 379, + 284 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {m i x}} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) = \\mathcal {L} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) + \\mathcal {R} _ {1} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) + \\mathcal {R} _ {2} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\qquad (5)", + "image_path": "a38dde0dd1694adb0b8d19f2397444a75fa7acef6db5d7bb56dd44daedde358d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 286, + 62, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 286, + 62, + 297 + ], + "spans": [ + { + "bbox": [ + 32, + 286, + 62, + 297 + ], + "type": "text", + "content": "where," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 38, + 304, + 373, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 304, + 373, + 376 + ], + "spans": [ + { + "bbox": [ + 38, + 304, + 373, + 376 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {R} _ {1} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\geq \\frac {R c _ {x} \\mathbb {E} _ {\\lambda} [ (1 - \\lambda) ] \\sqrt {d}}{N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} | g (f _ {\\theta} (x _ {i})) - y _ {i} | \\cdot | | \\nabla f _ {\\theta} (x _ {i}) | | _ {2} \\\\ \\mathcal {R} _ {2} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\geq \\frac {R ^ {2} c _ {x} ^ {2} \\mathbb {E} _ {\\lambda} [ (1 - \\lambda) ] ^ {2} d}{2 N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} | g (f _ {\\theta} (x _ {i})) (1 - g (f _ {\\theta} (x _ {i}))) | \\cdot | | \\nabla f _ {\\theta} (x _ {i}) | | _ {2} ^ {2}, \\\\ \\end{array}", + "image_path": "82801ce9cbacc2ae3fe9dc30642afde0c1a29597b705964dc16d2b69314a437d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 32, + 380, + 374, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 380, + 374, + 393 + ], + "spans": [ + { + "bbox": [ + 32, + 380, + 374, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 32, + 380, + 374, + 393 + ], + "type": "inline_equation", + "content": "R = \\min_{i\\in [N_{\\mathrm{val}}]}\\langle \\nabla f_{\\theta}(x_i),x_i\\rangle /||\\nabla f_{\\theta}(x_i)||\\cdot ||x_i||" + }, + { + "bbox": [ + 32, + 380, + 374, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 32, + 380, + 374, + 393 + ], + "type": "inline_equation", + "content": "c_{x} > 0" + }, + { + "bbox": [ + 32, + 380, + 374, + 393 + ], + "type": "text", + "content": " is a constant." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 397, + 379, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 397, + 379, + 420 + ], + "spans": [ + { + "bbox": [ + 30, + 397, + 379, + 420 + ], + "type": "text", + "content": "By comparing " + }, + { + "bbox": [ + 30, + 397, + 379, + 420 + ], + "type": "inline_equation", + "content": "\\ell(y, f_{\\theta}(x + \\delta))" + }, + { + "bbox": [ + 30, + 397, + 379, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 397, + 379, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{mix}}(\\theta, \\mathbb{D}_{\\mathrm{val}})" + }, + { + "bbox": [ + 30, + 397, + 379, + 420 + ], + "type": "text", + "content": " for a fully connected NN, we can prove the following." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "spans": [ + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "text", + "content": "Theorem 1. Suppose that " + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x) = \\nabla f_{\\theta}(x)^{T}x" + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "inline_equation", + "content": "\\nabla^2 f_{\\theta}(x) = 0" + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "text", + "content": " and there exists a constant " + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "inline_equation", + "content": "c_{x} > 0" + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "inline_equation", + "content": "\\| x_{i}\\|_{2}\\geq c_{x}\\sqrt{d}" + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\ldots ,N_{\\mathrm{val}}\\}" + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "text", + "content": ". Then, for any " + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 31, + 427, + 381, + 464 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 52, + 469, + 358, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 469, + 358, + 503 + ], + "spans": [ + { + "bbox": [ + 52, + 469, + 358, + 503 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {m i x}} (\\theta , \\mathbb {D} _ {\\mathrm {v a l}}) \\geq \\frac {1}{N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} \\ell \\left(y _ {i}, f _ {\\theta} (x _ {i} + \\varepsilon_ {i})\\right) \\geq \\frac {1}{N _ {\\mathrm {v a l}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v a l}}} \\ell \\left(y _ {i}, f _ {\\theta} (x _ {i} + \\varepsilon)\\right)", + "image_path": "1198dcdf9db441b242d148aa2935adb387147399a82d4904e42174b7c97ff8ae.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 31, + 509, + 381, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 509, + 381, + 536 + ], + "spans": [ + { + "bbox": [ + 31, + 509, + 381, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 509, + 381, + 536 + ], + "type": "inline_equation", + "content": "\\varepsilon_{i} = R_{i}c_{x}\\mathbb{E}_{\\lambda \\sim \\mathcal{D}_{\\lambda}}[1 - \\lambda ]\\sqrt{d}" + }, + { + "bbox": [ + 31, + 509, + 381, + 536 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 31, + 509, + 381, + 536 + ], + "type": "inline_equation", + "content": "R_{i} = \\langle \\nabla f_{\\theta}(x_{i}),x_{i}\\rangle /\\| \\nabla f_{\\theta}(x_{i})\\| \\cdot \\| x_{i}\\|" + }, + { + "bbox": [ + 31, + 509, + 381, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 509, + 381, + 536 + ], + "type": "inline_equation", + "content": "\\varepsilon = \\min \\{\\varepsilon_i\\}" + }, + { + "bbox": [ + 31, + 509, + 381, + 536 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 32, + 541, + 195, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 541, + 195, + 553 + ], + "spans": [ + { + "bbox": [ + 32, + 541, + 195, + 553 + ], + "type": "text", + "content": "Proof. is provided in Supplementary." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 30, + 559, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 559, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 559, + 381, + 584 + ], + "type": "text", + "content": "Theorem 1 implies that as long as " + }, + { + "bbox": [ + 30, + 559, + 381, + 584 + ], + "type": "inline_equation", + "content": "||\\delta ||\\leq \\varepsilon" + }, + { + "bbox": [ + 30, + 559, + 381, + 584 + ], + "type": "text", + "content": " holds, the MixUp loss " + }, + { + "bbox": [ + 30, + 559, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{mix}}(\\theta ,\\mathbb{D}_{\\mathrm{val}})" + }, + { + "bbox": [ + 30, + 559, + 381, + 584 + ], + "type": "text", + "content": " can be considered as an upper-bound of " + }, + { + "bbox": [ + 30, + 559, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{ideal}}(\\theta ,\\mathbb{D}_{\\mathrm{val}})" + }, + { + "bbox": [ + 30, + 559, + 381, + 584 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "text", + "content": "N. Karim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 186, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 186, + 46 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 186, + 46 + ], + "type": "text", + "content": "4.2 Clean Accuracy Retainer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 54, + 381, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 54, + 381, + 103 + ], + "spans": [ + { + "bbox": [ + 30, + 54, + 381, + 103 + ], + "type": "text", + "content": "In practice, it is desirable for a backdoor defense technique to be highly runtime efficient and retain the clean test accuracy of the original model. For better runtime efficiency and to retain clean accuracy, we choose to apply neural mask fine-tuning instead of fine-tuning the entire model, which can be formulated as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 109, + 379, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 109, + 379, + 124 + ], + "spans": [ + { + "bbox": [ + 111, + 109, + 379, + 124 + ], + "type": "interline_equation", + "content": "\\widehat {M} = \\quad \\arg \\min \\quad \\mathcal {L} ^ {\\operatorname {m i x}} (\\theta \\odot M, \\mathbb {D} _ {\\text {v a l}}) \\tag {6}", + "image_path": "b81b0346022de488fc72aa07d4ee381e041ae36a0f053fbf46ba09d96bca82f4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 124, + 220, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 124, + 220, + 135 + ], + "spans": [ + { + "bbox": [ + 137, + 124, + 220, + 135 + ], + "type": "interline_equation", + "content": "M \\mid m _ {l} ^ {(i)} \\in [ \\mu (l), 1 ], \\forall l, i", + "image_path": "8cfae5fb42dba007734b5297d8c3b43957c5216543425088f5d7535403eb9809.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 144, + 341, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 144, + 341, + 156 + ], + "spans": [ + { + "bbox": [ + 31, + 144, + 341, + 156 + ], + "type": "text", + "content": "We only optimize for neural masks " + }, + { + "bbox": [ + 31, + 144, + 341, + 156 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 31, + 144, + 341, + 156 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 31, + 144, + 341, + 156 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{val}}" + }, + { + "bbox": [ + 31, + 144, + 341, + 156 + ], + "type": "text", + "content": " and define " + }, + { + "bbox": [ + 31, + 144, + 341, + 156 + ], + "type": "inline_equation", + "content": "\\theta \\odot M" + }, + { + "bbox": [ + 31, + 144, + 341, + 156 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 166, + 379, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 166, + 379, + 180 + ], + "spans": [ + { + "bbox": [ + 102, + 166, + 379, + 180 + ], + "type": "interline_equation", + "content": "\\theta \\odot M := \\left\\{\\Theta_ {1} \\odot M _ {1}, \\Theta_ {2} \\odot M _ {2}, \\dots , \\Theta_ {L} \\odot M _ {L} \\right\\}, \\tag {7}", + "image_path": "d91dd59321359811825fc2b77383546d66d05f8ac415b1bd1df910c03db8de23.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 188, + 379, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 188, + 379, + 212 + ], + "spans": [ + { + "bbox": [ + 30, + 188, + 379, + 212 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 188, + 379, + 212 + ], + "type": "inline_equation", + "content": "\\theta \\coloneqq \\{\\Theta_1,\\dots ,\\Theta_L\\}" + }, + { + "bbox": [ + 30, + 188, + 379, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 30, + 188, + 379, + 212 + ], + "type": "inline_equation", + "content": "M\\coloneqq \\{M_1,\\ldots ,M_L\\}" + }, + { + "bbox": [ + 30, + 188, + 379, + 212 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 30, + 188, + 379, + 212 + ], + "type": "inline_equation", + "content": "M_{l} = [m_{l}^{(1)}\\dots m_{l}^{k_{l}}]^{T}\\in \\mathbb{R}^{k_{l}}" + }, + { + "bbox": [ + 30, + 188, + 379, + 212 + ], + "type": "text", + "content": ". We have" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 212, + 379, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 212, + 379, + 227 + ], + "spans": [ + { + "bbox": [ + 102, + 212, + 379, + 227 + ], + "type": "interline_equation", + "content": "\\Theta_ {l} \\odot M _ {l} := \\left[ m _ {l} ^ {(1)} \\Theta_ {l} ^ {(1)} \\dotsm m _ {l} ^ {\\left(k _ {l}\\right)} \\Theta_ {l} ^ {\\left(k _ {l}\\right)} \\right] \\in \\mathbb {R} ^ {k _ {l - 1} \\times k _ {l}}. \\tag {8}", + "image_path": "af572d7ab5ff623be365d1e34a5e450ef831422bdf65dee20c9dda1559e4794c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "spans": [ + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": "Note, in Eq. (8), a scalar mask " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "m_{l}^{(i)}" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " is applied to the weight vector " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "\\Theta_{l}^{(i)}" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " neuron of " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " layer. In our work, we formulate a constraint optimization problem where " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " depends on a mask scheduling function " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "\\mu(l): [1, L] \\to [0, 1]" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": ". Notice that " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "\\mu(l)" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " provides the lower limit of possible " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": "'s for the " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " layer's neurons' mask. We find the suitable function for " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "\\mu(l)" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " by analyzing commonly used mathematical functions (e.g., cosine, logarithmic, cubic, etc.). We analyze the impact of these functions in Section 5.3 and choose an exponential formulation (e.g., " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "\\alpha \\cdot e^{-\\beta \\cdot l}" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": ") for " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "\\mu(l)" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " since it produces the best possible outcome in terms of backdoor removal performance. Such formulation significantly reduces the mask search space, which leads to reduced runtime. Furthermore, the overall formulation for " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " encourages relatively small changes to the original backdoor model's parameters. This helps us retain on-par clean test accuracy after backdoor removal, which is highly desirable for a defense technique. To this end, we aim to purify the backdoor model by optimizing for the best possible mask " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "\\widehat{M}" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " that suppresses backdoor-affected neurons and bolsters the neurons responsible for clean test accuracy. Here, " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "\\widehat{M}" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": " should give us a purified model as, " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "f_{\\theta_c}(.)" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "inline_equation", + "content": "\\theta_c = \\theta \\odot \\widehat{M}" + }, + { + "bbox": [ + 30, + 233, + 380, + 450 + ], + "type": "text", + "content": ". Noteworthy, we do not change the bias as it may harm the classification accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 468, + 189, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 468, + 189, + 480 + ], + "spans": [ + { + "bbox": [ + 31, + 468, + 189, + 480 + ], + "type": "text", + "content": "4.3 Sample Efficiency of NFT" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "type": "text", + "content": "In this section, we discuss how careful modifications to the optimization scenario can make NFT highly sample-efficient. For this analysis, we first train a backdoor model (PreActResNet-18 [18]) on a poisoned CIFAR10 dataset for 200 epochs. For poisoning the dataset, we use TrojanNet [35] backdoor attack with a poison rate of " + }, + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "type": "text", + "content": ". In Fig. 1, we show t-SNE visualization of different class clusters obtained using the backdoor model. Instead of 10 clusters, we have an additional cluster (red color, labeled \"11\") that sits closely with the original target class cluster (in this case, the target class is \"0\"). We name this cluster \"poison cluster\"," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "type": "text", + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 35, + 32, + 148, + 123 + ], + "blocks": [ + { + "bbox": [ + 35, + 32, + 148, + 123 + ], + "lines": [ + { + "bbox": [ + 35, + 32, + 148, + 123 + ], + "spans": [ + { + "bbox": [ + 35, + 32, + 148, + 123 + ], + "type": "image", + "image_path": "661d214f21bf33f101b31e1d7c849ca255d3f6bf4039d4385dea1cfe3859d108.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 124, + 128, + 133 + ], + "lines": [ + { + "bbox": [ + 56, + 124, + 128, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 124, + 128, + 133 + ], + "type": "text", + "content": "(a) Backdoor Model" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 150, + 32, + 262, + 123 + ], + "blocks": [ + { + "bbox": [ + 150, + 32, + 262, + 123 + ], + "lines": [ + { + "bbox": [ + 150, + 32, + 262, + 123 + ], + "spans": [ + { + "bbox": [ + 150, + 32, + 262, + 123 + ], + "type": "image", + "image_path": "5ea7a8cde66c1e057de8cc687d9d6c2923d24fe44c2d8f7c205b9aeaf214366e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 124, + 251, + 133 + ], + "lines": [ + { + "bbox": [ + 160, + 124, + 251, + 133 + ], + "spans": [ + { + "bbox": [ + 160, + 124, + 251, + 133 + ], + "type": "text", + "content": "(b) Purification w/o Reg." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 31, + 142, + 381, + 253 + ], + "lines": [ + { + "bbox": [ + 31, + 142, + 381, + 253 + ], + "spans": [ + { + "bbox": [ + 31, + 142, + 381, + 253 + ], + "type": "text", + "content": "Fig. 1: t-SNE visualization of a backdoor model, where we show the \"poison cluster\" with red color and label \"11\". Since the attack target class is \"0\", cluster \"0\" and the poison cluster sit closely with each other (Fig. 1a). After purification, the cluster should break, and all triggered samples should be classified according to their original label. In Fig. 1b, we perform one-shot NFT without employing the regularizer. Due to the overfitting issue, the clean clusters lose their separability that can be established with Mask regularizer, which tackles this issue (larger cluster gaps as compared to scenarios in Fig. 1b) by keeping purified model parameters close to the original backdoor model (Fig. 1c). This, in turn, produces better clean test accuracy. For evaluation, we train a PreActResNet18 [18] on CIFAR10 dataset with a poison rate of " + }, + { + "bbox": [ + 31, + 142, + 381, + 253 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 31, + 142, + 381, + 253 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 262, + 31, + 376, + 123 + ], + "blocks": [ + { + "bbox": [ + 262, + 31, + 376, + 123 + ], + "lines": [ + { + "bbox": [ + 262, + 31, + 376, + 123 + ], + "spans": [ + { + "bbox": [ + 262, + 31, + 376, + 123 + ], + "type": "image", + "image_path": "0143fd2bc8fd83fcc35e7784a9c45b38fcf4f56f9b330947f453d39d70d5fad1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 124, + 365, + 133 + ], + "lines": [ + { + "bbox": [ + 274, + 124, + 365, + 133 + ], + "spans": [ + { + "bbox": [ + 274, + 124, + 365, + 133 + ], + "type": "text", + "content": "(c) Purification with Reg." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 276, + 380, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 276, + 380, + 324 + ], + "spans": [ + { + "bbox": [ + 30, + 276, + 380, + 324 + ], + "type": "text", + "content": "whereas other clusters are \"clean clusters\". This cluster contains the embeddings of attacked or triggered samples from all other classes. The goal of any defense system is to break the formation of the poison cluster so that poison samples return to their original clusters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 324, + 380, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 324, + 380, + 421 + ], + "spans": [ + { + "bbox": [ + 30, + 324, + 380, + 421 + ], + "type": "text", + "content": "One-Shot NFT. Let us consider that there is only 1 sample (per class) available for the validation set " + }, + { + "bbox": [ + 30, + 324, + 380, + 421 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{val}}" + }, + { + "bbox": [ + 30, + 324, + 380, + 421 + ], + "type": "text", + "content": ". Applying NFT with this " + }, + { + "bbox": [ + 30, + 324, + 380, + 421 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{val}}" + }, + { + "bbox": [ + 30, + 324, + 380, + 421 + ], + "type": "text", + "content": " forms the clusters shown in Fig. 1b. Notice that the poison cluster breaks even with one-shot fine-tuning, indicating the backdoor's effect is removed successfully. However, since only one sample is available per class, the model easily overfits " + }, + { + "bbox": [ + 30, + 324, + 380, + 421 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{val}}" + }, + { + "bbox": [ + 30, + 324, + 380, + 421 + ], + "type": "text", + "content": ", reducing margins between clean clusters. Such unwanted overfitting issue negatively impacts the clean test accuracy. To combat this issue in scenarios where very few samples are available, we add a simple regularizer term as follows," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 87, + 430, + 379, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 430, + 379, + 455 + ], + "spans": [ + { + "bbox": [ + 87, + 430, + 379, + 455 + ], + "type": "interline_equation", + "content": "\\underset {M \\mid m _ {l} ^ {(i)} \\in [ \\mu (l), 1 ], \\forall l, i} {\\arg \\min } \\mathcal {L} ^ {\\operatorname {m i x}} \\left(\\theta \\odot M, \\mathbb {D} _ {\\text {v a l}}\\right) + \\eta_ {c} \\| M _ {0} - M \\| _ {1} \\tag {9}", + "image_path": "3442801e79cebed4d1b898e11112c5edaf843dca0f2c074144c65b805fa87437.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": " are the initial mask values (initialized as 1's) and " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\eta_c" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": " is the regularizer coefficient. By minimizing the " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": "-norm of the mask differences, we try to keep the purified model parameters " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "(\\theta \\odot M)" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": " close to the original backdoor model parameters " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "(\\theta \\odot M_0)" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": ". We hope to preserve the original decision boundary between clean clusters by keeping these parameters as close as possible. Note that the overfitting issue is not as prominent whenever we have a reasonably sized " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{val}}" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{train}}" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": "). Therefore, we choose the value of " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\eta_c" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": " to be " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "5e^{-4} / n_c" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": ", which dynamically changes based on the number of samples available per class " + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "inline_equation", + "content": "(n_c)" + }, + { + "bbox": [ + 31, + 464, + 380, + 584 + ], + "type": "text", + "content": ". As the number of samples increases, the impact of the regularizer reduces. Note that, Eq. (9) represents the final optimization function for our proposed method." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 128, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 128, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 128, + 19 + ], + "type": "text", + "content": "N. Karim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 35, + 87, + 381, + 346 + ], + "blocks": [ + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "lines": [ + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "spans": [ + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "text", + "content": "Table 1: Comparison of different defense methods for CIFAR10 and ImageNet. Average drop " + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "text", + "content": " changes in ASR/ACC compared to the baseline, i.e., ASR/ACC of No Defense. A good defense should have a large ASR drop with a small ACC drop. Attacks are implemented with a poison rate of " + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 35, + 87, + 381, + 346 + ], + "lines": [ + { + "bbox": [ + 35, + 87, + 381, + 346 + ], + "spans": [ + { + "bbox": [ + 35, + 87, + 381, + 346 + ], + "type": "table", + "html": "
DatasetMethodNo DefenseANPI-BAUAWMFT-SAMRNPNFT (Ours)
AttacksASRACCASRACCASRACCASRACCASRACCASRACCASRACC
CIFAR-10Benign095.21092.28093.98093.56093.80093.16094.10
Badnets10092.964.8785.922.8485.965.7287.854.3486.172.7588.461.7490.82
Blend10094.114.7787.613.8189.105.5390.842.1388.930.9191.530.3193.17
Troj-one10089.573.7882.185.4786.206.9187.245.4186.453.8487.391.6487.71
Troj-all10088.333.9181.955.5384.896.8285.944.4284.604.0285.801.7987.10
SIG10088.641.0481.920.3783.604.1283.570.9085.380.5186.460.1287.16
Dyn-one10092.524.7388.611.7887.267.4891.163.3588.418.6190.051.3790.81
Dyn-all10092.614.2888.322.1984.517.3089.742.4687.7210.5790.281.4291.53
CLB10092.780.8387.411.4185.075.7886.701.8984.186.1290.381.0490.37
CBA93.2090.1727.8083.7945.1185.6336.1285.0538.8185.5817.7286.4021.6087.97
FBA10090.787.9582.9066.7087.4210.6687.3522.3187.069.4887.636.2188.56
WaNet98.6492.295.8186.703.1889.247.7286.942.9688.458.1090.262.3889.65
ISSBA99.8092.786.7685.423.8289.2012.4890.034.5789.597.5888.624.2490.18
LIRA99.2592.157.3487.414.5189.616.1388.503.8689.2211.8387.591.5390.57
BPPA99.7093.829.9490.2310.4690.579.9490.6810.6090.889.7491.375.0491.78
Avg. Drop--92.63↓ 5.94↓ 88.10↓ 4.66↓ 91.21↓ 3.71↓ 92.61↓ 4.26↓ 92.06↓ 2.95↓ 95.56↓ 1.81
ImageNetBenign077.06073.52071.85074.21071.63075.20075.51
Badnets99.2474.535.9169.376.3166.282.8769.464.1869.447.5870.493.6170.96
Troj-one99.2174.027.6369.157.7367.145.7469.352.8670.622.9472.173.1672.37
Troj-all97.5874.459.1869.867.5468.206.0269.643.2769.854.8171.452.6872.13
Blend10074.426.4370.207.7968.517.4568.618.1568.915.6970.243.8371.52
SIG94.6674.691.2369.824.2866.085.3770.023.4769.744.3670.732.9472.36
CLB95.0874.146.7169.194.3766.417.6469.703.5069.329.4471.523.0572.25
Dyn-one98.2474.806.6869.658.3269.618.6270.174.4270.0512.5670.392.6271.91
Dyn-all98.5675.0813.4970.189.8268.9212.6870.244.8169.9014.1869.473.7771.62
LIRA96.0474.6112.8669.2212.0869.8013.2769.353.1612.3170.5071.382.6270.73
WaNet97.6074.486.3468.345.6767.236.3170.024.4266.827.7871.624.7171.63
ISSBA98.2374.387.6168.424.5067.928.2169.513.3568.029.7470.812.0670.67
Avg. Drop--90.08↓ 5.17↓ 88.90↓ 7.41↓ 90.01↓ 4.72↓ 92.24↓ 5.61↓ 89.37↓ 3.66↓ 94.03↓ 2.84
", + "image_path": "a34836711a5956bf4dc669f1b4c06d9a6d0865ab075d6cd032ac36345e50d895.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 364, + 182, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 364, + 182, + 377 + ], + "spans": [ + { + "bbox": [ + 31, + 364, + 182, + 377 + ], + "type": "text", + "content": "5 Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 386, + 158, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 386, + 158, + 398 + ], + "spans": [ + { + "bbox": [ + 31, + 386, + 158, + 398 + ], + "type": "text", + "content": "5.1 Evaluation Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 404, + 379, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 404, + 379, + 547 + ], + "spans": [ + { + "bbox": [ + 30, + 404, + 379, + 547 + ], + "type": "text", + "content": "Datasets. We evaluate the proposed method through a range of experiments on two widely used datasets for backdoor attack study: CIFAR10 [22] with 10 classes, GTSRB [47] with 43 classes. For the scalability test of our method, we also consider Tiny-ImageNet [24] with 100,000 images distributed among 200 classes and ImageNet [11] with 1.28M images distributed among 1000 classes. For multi-label clean-image backdoor attacks, we use object detection datasets Pascal VOC [14] and MS-COCO [33]. UCF-101 [46] and HMDB51 [23] have been used for evaluating in action recognition task. The ModelNet [56] dataset was also utilized to assess the performance of a 3D point cloud classifier. In addition to vision, we also consider attacks on natural language generation and use WMT2014 En-De [4] machine translation and OpenSubtitles2012 [48] dialogue generation datasets (Results are in the Supplementary)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 548, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 548, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 548, + 379, + 584 + ], + "type": "text", + "content": "Attacks Configurations. Here, we first briefly overview the attack configurations on single-label image recognition datasets. We consider 14 state-of-the-art backdoor attacks: 1) Badnets [17], 2) Blend attack [9], 3 & 4) TrojanNet" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "type": "text", + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 87, + 344, + 147 + ], + "blocks": [ + { + "bbox": [ + 30, + 33, + 380, + 78 + ], + "lines": [ + { + "bbox": [ + 30, + 33, + 380, + 78 + ], + "spans": [ + { + "bbox": [ + 30, + 33, + 380, + 78 + ], + "type": "text", + "content": "Table 2: Performance analysis for the multi-label backdoor attack [8]. We choose 3 object detection datasets [14,33] and ML-decoder [44] network architecture for this evaluation. Mean average precision (mAP) and ASR of the model, with and without defenses, have been shown." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 87, + 344, + 147 + ], + "lines": [ + { + "bbox": [ + 67, + 87, + 344, + 147 + ], + "spans": [ + { + "bbox": [ + 67, + 87, + 344, + 147 + ], + "type": "table", + "html": "
DatasetNo defenseANPAWMRNPFT-SAMNFT (Ours)
ASRmAPASRmAPASRmAPASRmAPASRmAPASRmAP
VOC0786.492.521.786.926.687.319.287.619.386.817.389.1
VOC1284.891.918.685.319.085.913.886.414.687.114.288.4
MS-COCO85.688.019.784.122.683.417.184.319.283.816.685.8
", + "image_path": "1d3a12e3b1916942cb0a610e3de458dbe0f40a0fd376629a8bc65dde13629c0d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 75, + 205, + 337, + 255 + ], + "blocks": [ + { + "bbox": [ + 30, + 162, + 380, + 195 + ], + "lines": [ + { + "bbox": [ + 30, + 162, + 380, + 195 + ], + "spans": [ + { + "bbox": [ + 30, + 162, + 380, + 195 + ], + "type": "text", + "content": "Table 3: Performance analysis for action recognition tasks where we choose 2 video datasets for evaluation. We consider a clean-label attack [64], where we need to generate adversarial perturbations for each input frame." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 205, + 337, + 255 + ], + "lines": [ + { + "bbox": [ + 75, + 205, + 337, + 255 + ], + "spans": [ + { + "bbox": [ + 75, + 205, + 337, + 255 + ], + "type": "table", + "html": "
DatasetNo defenseI-BAUAWMRNPFT-SAMNFT (Ours)
ASRACCASRACCASRACCASRACCASRACCASRACC
UCF-10181.375.620.470.620.870.117.070.315.971.613.371.2
HMDB-5180.245.017.541.115.240.912.640.410.841.79.440.8
", + "image_path": "899fcd9db2b7eda0d8bcb04ae8a9655fb304b306f0712328d8f8e8c4f0aee82a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 282, + 380, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 282, + 380, + 474 + ], + "spans": [ + { + "bbox": [ + 30, + 282, + 380, + 474 + ], + "type": "text", + "content": "(Troj-one & Troj-all) [35], 5) Sinusoidal signal attack (SIG) [2], 6 & 7) Input-Aware Attack (Dyn-one and Dyn-all) [40], 8) Clean-label attack (CLB) [50], 9) Composite backdoor (CBA) [32], 10) Deep feature space attack (FBA) [10], 11) Warping-based backdoor attack (WaNet) [39], 12) Invisible triggers based backdoor attack (ISSBA) [31], 13) Imperceptible backdoor attack (LIRA) [12], and 14) Quantization and contrastive learning-based attack (BPPA) [53]. In order to facilitate a fair comparison, we adopt trigger patterns and settings similar to those used in the original papers. Specifically, for both Trojan-one and Dyn-one attacks, we set all triggered images to have the same target label (i.e., all2one), whereas, for Trojan-all and Dyn-all attacks, we have uniformly distributed the target labels across all classes (i.e., all2all). Details on the hyper-parameters and overall training settings can be found in the Supplementary. We measure the success of an attack using two metrics: clean test accuracy (ACC) defined as the percentage of clean samples that are classified to their original target label and attack success rate (ASR) defined as the percentage of poison test samples (" + }, + { + "bbox": [ + 30, + 282, + 380, + 474 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 30, + 282, + 380, + 474 + ], + "type": "text", + "content": ") that are classified to the target label (" + }, + { + "bbox": [ + 30, + 282, + 380, + 474 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 30, + 282, + 380, + 474 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": "Defenses Configurations. We compare our approach with 10 existing backdoor mitigation methods: 1) " + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "inline_equation", + "content": "FT-SAM" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": " [66]; 2) Adversarial Neural Pruning (ANP) [54]; 3) Implicit Backdoor Adversarial Unlearning (" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "inline_equation", + "content": "I-BAU" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": ") [59]; 4) Adversarial Weight Masking (" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "inline_equation", + "content": "AWM" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": ") [6]; 5) Reconstructive Neuron Pruning (RNP) [29]; 6) Fine-Pruning (" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": ") [36]; 7) Mode Connectivity Repair (" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "inline_equation", + "content": "MCR" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": ") [63]; 8) Neural Attention Distillation (" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "inline_equation", + "content": "NAD" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": ") [28], 9) Causality-inspired Backdoor Defense (" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "inline_equation", + "content": "CBD" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": ") [62]), 10) Anti Backdoor Learning (" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "inline_equation", + "content": "ABL" + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": ") [27]. In the main paper, we compare NFT with the first 5 defenses as they are more relevant, and the comparison with the rest of the methods is in Supplementary. To apply NFT, we" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "text", + "content": "N. Karim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 98, + 339, + 166 + ], + "blocks": [ + { + "bbox": [ + 31, + 33, + 380, + 89 + ], + "lines": [ + { + "bbox": [ + 31, + 33, + 380, + 89 + ], + "spans": [ + { + "bbox": [ + 31, + 33, + 380, + 89 + ], + "type": "text", + "content": "Table 4: Removal performance (\\%) of NFT against backdoor attacks on 3D point cloud classifiers. The attack methods [26], namely Poison-Label Backdoor Attack (PointPBA) with interaction trigger (PointPBA-I), PointPBA with orientation trigger (PointPBA-O), and Clean-Label Backdoor Attack (PointCBA) were considered, as well as the \"backdoor points\" based attack (3DPC-BA) outlined in prior work [58]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 98, + 339, + 166 + ], + "lines": [ + { + "bbox": [ + 73, + 98, + 339, + 166 + ], + "spans": [ + { + "bbox": [ + 73, + 98, + 339, + 166 + ], + "type": "table", + "html": "
AttackNo DefenseANPAWMRNPFT-SAMNFT (Ours)
ASRACCASRACCASRACCASRACCASRACCASRACC
PointBA-I98.689.113.682.615.483.98.184.08.884.59.685.7
PointBA-O94.789.814.882.013.182.49.483.88.285.07.585.3
PointCBA66.088.721.283.321.583.818.684.620.384.719.486.1
3DPC-BA93.891.216.884.715.685.913.985.713.186.312.687.7
", + "image_path": "fb3e8caf6671e7a1f3bffe231407c4a16f93e71b24f163ddef478eeee0a5717d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 190, + 379, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 190, + 379, + 239 + ], + "spans": [ + { + "bbox": [ + 30, + 190, + 379, + 239 + ], + "type": "text", + "content": "take " + }, + { + "bbox": [ + 30, + 190, + 379, + 239 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 30, + 190, + 379, + 239 + ], + "type": "text", + "content": " clean validation data (set aside from the training set) and fine-tune the model for 100 epochs. An SGD optimizer has been employed with a learning rate of 0.05 and a momentum of 0.95. The rest of the experimental details for NFT and other defense methods are in the Supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 260, + 228, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 260, + 228, + 273 + ], + "spans": [ + { + "bbox": [ + 31, + 260, + 228, + 273 + ], + "type": "text", + "content": "5.2 Performance Comparison of NFT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 284, + 379, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 284, + 379, + 319 + ], + "spans": [ + { + "bbox": [ + 30, + 284, + 379, + 319 + ], + "type": "text", + "content": "In this section, we compare the performance of NFT with other defenses in various scenarios: single-label (i.e., image classification), multi-label (i.e., object detection), video action recognition, and 3D point cloud." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "text", + "content": "Single-Label Backdoor attack. In Table 1, we present the performance of different defenses for 2 widely used benchmarks. For CIFAR10, we consider 4 label poisoning attacks: Badnets, Blend, Trojan, and Dynamic. For all of these attacks, NFT shows significant performance improvement over other baseline methods. While ANP and AWM defenses work well for mild attacks with low poison rates (e.g., " + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "text", + "content": "), the performance deteriorates for attacks with high poison rates (e.g., " + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\geq 10\\%" + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "text", + "content": "). It is observable that NFT performs well across all attack scenarios, e.g., obtaining a " + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "inline_equation", + "content": "99.69\\%" + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "text", + "content": " drop in ASR for blend attack while also performing well for clean data (only " + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "inline_equation", + "content": "0.94\\%" + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "text", + "content": " of ACC drop). For Trojan and Dynamic attacks, we consider two different versions of attacks based on label-mapping criteria (all2one and all2all). The drop in attack success rate shows the effectiveness of NFT against such attacks. Recently, small and imperceptible perturbations as triggers have been developed in attacks (e.g., WaNet, LIRA, etc.) to fool the defense systems. While AWM generates perturbations as a proxy for these imperceptible triggers, NFT does a better job in this regard. For further validation of our proposed method, we use deep feature-based attacks CBA and FBA. Both of these attacks manipulate features to insert backdoor behavior. Overall, we achieve an average drop of " + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "inline_equation", + "content": "95.56\\%" + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "text", + "content": " in ASR while sacrificing an ACC of " + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "inline_equation", + "content": "1.81\\%" + }, + { + "bbox": [ + 30, + 321, + 380, + 584 + ], + "type": "text", + "content": ". For the scalability test, we consider a large and widely used dataset in vision tasks, ImageNet. In consistency with other datasets, NFT also obtains SOTA performance in this particular dataset. Due to page constraints, we move the performance comparison for GTSRB and Tiny-ImageNet to the Supplementary." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 32, + 33, + 111, + 95 + ], + "blocks": [ + { + "bbox": [ + 32, + 33, + 111, + 95 + ], + "lines": [ + { + "bbox": [ + 32, + 33, + 111, + 95 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 111, + 95 + ], + "type": "image", + "image_path": "739caefc6cc6f472a3e3fa20536eceebcdcc00859c38125ba0e0bdc9ce58bd04.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 102, + 199, + 125 + ], + "lines": [ + { + "bbox": [ + 31, + 102, + 199, + 125 + ], + "spans": [ + { + "bbox": [ + 31, + 102, + 199, + 125 + ], + "type": "text", + "content": "Fig.2: Ablation with different Mask Scheduling Function " + }, + { + "bbox": [ + 31, + 102, + 199, + 125 + ], + "type": "inline_equation", + "content": "(\\mu)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 119, + 31, + 199, + 94 + ], + "blocks": [ + { + "bbox": [ + 119, + 31, + 199, + 94 + ], + "lines": [ + { + "bbox": [ + 119, + 31, + 199, + 94 + ], + "spans": [ + { + "bbox": [ + 119, + 31, + 199, + 94 + ], + "type": "image", + "image_path": "0a630b1ad7cd024747d42e8992d218f11393921131a29a5c2041a784a595ed3a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 102, + 379, + 124 + ], + "lines": [ + { + "bbox": [ + 213, + 102, + 379, + 124 + ], + "spans": [ + { + "bbox": [ + 213, + 102, + 379, + 124 + ], + "type": "text", + "content": "Fig.3: Mask Distribution of AWM (left) and NFT (right)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 213, + 31, + 293, + 94 + ], + "blocks": [ + { + "bbox": [ + 213, + 31, + 293, + 94 + ], + "lines": [ + { + "bbox": [ + 213, + 31, + 293, + 94 + ], + "spans": [ + { + "bbox": [ + 213, + 31, + 293, + 94 + ], + "type": "image", + "image_path": "33c18e501e4c85b3e148d7a67d4a97b2dcb5ee0ed683ec77f0ed7c64e534259a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 300, + 32, + 380, + 94 + ], + "blocks": [ + { + "bbox": [ + 300, + 32, + 380, + 94 + ], + "lines": [ + { + "bbox": [ + 300, + 32, + 380, + 94 + ], + "spans": [ + { + "bbox": [ + 300, + 32, + 380, + 94 + ], + "type": "image", + "image_path": "75b5989a31239d4f1ac6b0417f212a1d00d58ffa8377f9dcaa9aded1fd44feba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 156, + 380, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 156, + 380, + 313 + ], + "spans": [ + { + "bbox": [ + 30, + 156, + 380, + 313 + ], + "type": "text", + "content": "Multi-label Backdoor Attack. In Table 2, we also evaluate our proposed method on multi-label clean-image backdoor attack [8]. In general, we put a trigger on the image and change the corresponding ground truth of that image. However, a certain combination of objects has been used as a trigger pattern. For example, if a combination of car, person, and truck is present in the image, it will fool the model into misclassifying it. Table 2 shows that our proposed method surpasses other defense strategies concerning both ASR and mAP metrics. Notably, ANP and AWM, which rely on adversarial search limited applicability in multi-label scenarios. This limitation arises from the less accurate process of approximating triggers for object detection. Conversely, FT-SAM's optimization driven by sharpness proves effective in removing backdoors, yet it achieves a lower mAP post-purification. This outcome is not ideal since the goal is to eliminate backdoors without significantly compromising clean accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "spans": [ + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "text", + "content": "Action Recognition Model. We further consider attacks on action recognition models; results are reported in Table 3. We use two widely used datasets, UCF-101 [46] and HMDB51 [23], with a CNN+LSTM network architecture. An ImageNet pre-trained ResNet50 network has been used for the CNN, and a sequential input-based Long Short Term Memory (LSTM) [45] network has been put on top of it. We subsample the input video by keeping one out of every 5 frames and use a fixed frame resolution of " + }, + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "text", + "content": ". We choose a trigger size of " + }, + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "inline_equation", + "content": "20 \\times 20" + }, + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "text", + "content": ". Following [64], we create the required perturbation for clean-label attack by running projected gradient descent (PGD) [37] for 2000 steps with a perturbation norm of " + }, + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "inline_equation", + "content": "\\epsilon = 16" + }, + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "text", + "content": ". Note that our proposed augmentation strategies for image classification are directly applicable to video recognition. During training, we keep " + }, + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 30, + 316, + 379, + 520 + ], + "type": "text", + "content": " samples from each class to use them later as the clean validation set. Table 3 shows that NFT outperforms other defenses by a significant margin. In the case of a high number of classes and multiple image frames in the same input, it is a challenging task to optimize for the proper trigger pattern through the adversarial search described in I-BAU and AWM. Without a good approximation of the trigger, these methods seem to underperform in most cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "content": "3D Point Cloud. In this phase of our work, we assess NFT's resilience against attacks on 3D point cloud classifiers [26,58]. To evaluate, we utilize the ModelNet dataset [56] and the PointNet++ architecture [42]. The performance comparison of NFT and other defense methods is outlined in Table 4. NFT outperforms other defenses due to its unique formulations of the objective function." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "text", + "content": "N. Karim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 34, + 78, + 204, + 102 + ], + "blocks": [ + { + "bbox": [ + 31, + 34, + 201, + 67 + ], + "lines": [ + { + "bbox": [ + 31, + 34, + 201, + 67 + ], + "spans": [ + { + "bbox": [ + 31, + 34, + 201, + 67 + ], + "type": "text", + "content": "Table 5: Average runtime of different defense methods against all 14 attacks. An NVIDIA RTX3090 GPU is used." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 34, + 78, + 204, + 102 + ], + "lines": [ + { + "bbox": [ + 34, + 78, + 204, + 102 + ], + "spans": [ + { + "bbox": [ + 34, + 78, + 204, + 102 + ], + "type": "table", + "html": "
Method|ANP I-BAU AWM FT-SAM RNP NFT (Ours)
Runtime (sec.)118.192.5112.598.7102.628.4
", + "image_path": "32f052f28caedc9430265e0477f4fdb9dc23ba00839e742fb0a38011776b51fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 218, + 65, + 373, + 105 + ], + "blocks": [ + { + "bbox": [ + 211, + 33, + 379, + 55 + ], + "lines": [ + { + "bbox": [ + 211, + 33, + 379, + 55 + ], + "spans": [ + { + "bbox": [ + 211, + 33, + 379, + 55 + ], + "type": "text", + "content": "Table 6: Sensitivity analysis of " + }, + { + "bbox": [ + 211, + 33, + 379, + 55 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 211, + 33, + 379, + 55 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 211, + 33, + 379, + 55 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 211, + 33, + 379, + 55 + ], + "type": "text", + "content": " for LIRA on CIFAR10." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 218, + 65, + 373, + 105 + ], + "lines": [ + { + "bbox": [ + 218, + 65, + 373, + 105 + ], + "spans": [ + { + "bbox": [ + 218, + 65, + 373, + 105 + ], + "type": "table", + "html": "
α0.90.90.90.80.80.80.70.70.7
β0.250.500.750.250.500.750.250.500.75
ASR3.653.424.872.731.533.764.184.925.23
ACC88.5288.3489.1289.9790.5789.3487.6487.7888.10
", + "image_path": "2bec722c56030f5539df398af269e04255d514e0e2db51e05e2a258f48cb8f7f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 128, + 137, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 128, + 137, + 141 + ], + "spans": [ + { + "bbox": [ + 31, + 128, + 137, + 141 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 151, + 289, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 151, + 289, + 162 + ], + "spans": [ + { + "bbox": [ + 30, + 151, + 289, + 162 + ], + "type": "text", + "content": "For all ablation studies, we consider the CIFAR10 dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 164, + 380, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 164, + 380, + 223 + ], + "spans": [ + { + "bbox": [ + 30, + 164, + 380, + 223 + ], + "type": "text", + "content": "Runtime Analysis. For runtime analysis, we present the training time for different defenses in Table 5. ANP and AWM both employ computationally expensive adversarial search procedures to prune neurons, which makes them almost " + }, + { + "bbox": [ + 30, + 164, + 380, + 223 + ], + "type": "inline_equation", + "content": "4\\mathrm{x}" + }, + { + "bbox": [ + 30, + 164, + 380, + 223 + ], + "type": "text", + "content": " slower than our method. However, NFT offers a computationally less expensive defense with SOTA performance in major benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "spans": [ + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "content": "Choice of Scheduling Function, " + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "content": ". For choosing the suitable function for " + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "content": ", we conduct a detailed study with commonly used mathematical functions. Note that, we only consider a family of functions that decreases over the depth of the network (shown in Fig. 2). This allows more variations for deeper layer weights, making sense as they are more responsible for DNN decision-making. In our work, we choose an exponential formulation for " + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "content": " as it offers superior performance. We also perform sensitivity analysis of scheduling parameters " + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "content": " in Table 6. In Fig. 3, we show the generated mask distributions of AWM and NFT. Compared to AWM, NFT produces more uniformly distributed masks that seem helpful for backdoor purification. We show the impact of " + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "content": " in Table 7. Nature of Optimization. In Table 7, we present the performance of SOTA techniques under different validation sizes. Even with 10 samples (single-shot), NFT performs reasonably well and offers better performance as compared to AWM. This again shows that the trigger generation process is less accurate and effective for a very small validation set. We also show the effect of the proposed mask regularizer that indirectly controls the change in weights for better ACC. Although AWM employs a similar " + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "content": " regularizer for masks, our proposed regularizer is more intuitive and specifically designed for better ACC preservation. While AWM encourages sparse solutions for " + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 30, + 224, + 380, + 475 + ], + "type": "text", + "content": " (shown in the left subfigure of Fig. 3), it helps with ASR but heavily compromises ACC. We also show the performance of NFT without augmentations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 476, + 380, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 476, + 380, + 583 + ], + "spans": [ + { + "bbox": [ + 30, + 476, + 380, + 583 + ], + "type": "text", + "content": "Label Correction Rate. In the standard backdoor removal metric, it is sufficient for backdoored images to be classified as a non-target class (any class other than " + }, + { + "bbox": [ + 30, + 476, + 380, + 583 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 30, + 476, + 380, + 583 + ], + "type": "text", + "content": "). However, we also consider another metric, label correction rate (LCR), for quantifying the success of a defense. We define LCR as the percentage of poisoned samples correctly classified to their original classes. Any method with the highest value of LCR is considered to be the best defense method. For this evaluation, we use CIFAR10 dataset and 6 backdoor attacks. Initially, the correction rate is " + }, + { + "bbox": [ + 30, + 476, + 380, + 583 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 30, + 476, + 380, + 583 + ], + "type": "text", + "content": " with no defense as the ASR is close to " + }, + { + "bbox": [ + 30, + 476, + 380, + 583 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 30, + 476, + 380, + 583 + ], + "type": "text", + "content": ". Table 8 shows that NFT obtains better performance in terms of LCR." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 74, + 9, + 346, + 21 + ], + "type": "text", + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 119, + 365, + 238 + ], + "blocks": [ + { + "bbox": [ + 30, + 33, + 381, + 111 + ], + "lines": [ + { + "bbox": [ + 30, + 33, + 381, + 111 + ], + "spans": [ + { + "bbox": [ + 30, + 33, + 381, + 111 + ], + "type": "text", + "content": "Table 7: Purification performance (\\%) for various validation data sizes. NFT performs reasonably well even with as few as 10 samples, i.e., one sample (shot) per class for CIFAR10. We also show the impact of the mask regularizer, mask scheduling function " + }, + { + "bbox": [ + 30, + 33, + 381, + 111 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 30, + 33, + 381, + 111 + ], + "type": "text", + "content": ", and augmentations on performance, which resonates with Fig. 1. Mask regularizer has the most impact on the clean test accuracy (around " + }, + { + "bbox": [ + 30, + 33, + 381, + 111 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 30, + 33, + 381, + 111 + ], + "type": "text", + "content": " worse without the regularizer). Without strong augmentations, we have a better ACC with a slightly worse ASR (around " + }, + { + "bbox": [ + 30, + 33, + 381, + 111 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 30, + 33, + 381, + 111 + ], + "type": "text", + "content": " drop)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 119, + 365, + 238 + ], + "lines": [ + { + "bbox": [ + 47, + 119, + 365, + 238 + ], + "spans": [ + { + "bbox": [ + 47, + 119, + 365, + 238 + ], + "type": "table", + "html": "
AttackDynamicWaNetLIRA
Samples101001010010100
MethodASRACCASRACCASRACCASRACCASRACCASRACC
No Defense10092.5210092.5298.6492.2998.6492.2999.2592.1599.2592.15
AWM86.7455.739.1685.3383.0162.217.2384.3891.4566.6410.8385.87
FT-SAM8.3573.495.7284.709.3575.985.5686.6311.8372.404.8588.82
NFT w/o Reg.5.6776.741.3682.214.1876.723.0283.314.8374.582.3283.61
NFT w/o Aug.11.9181.8610.5989.5310.3683.107.8189.6812.2381.059.1688.74
NFT w/o μ(l)5.1180.323.0488.585.8582.464.6488.026.4881.944.3388.75
NFT4.8380.511.7290.084.4183.582.9689.155.1882.722.0489.34
", + "image_path": "b605384c58315c47c95786660a4f8b428bbfdfc5499453ba738e237a02b92779.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 106, + 291, + 307, + 376 + ], + "blocks": [ + { + "bbox": [ + 30, + 247, + 381, + 281 + ], + "lines": [ + { + "bbox": [ + 30, + 247, + 381, + 281 + ], + "spans": [ + { + "bbox": [ + 30, + 247, + 381, + 281 + ], + "type": "text", + "content": "Table 8: Label Correction Rate (%) for different defense techniques, defined as the percentage of backdoor samples that are correctly classified to their original ground truth label." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 291, + 307, + 376 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 307, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 307, + 376 + ], + "type": "table", + "html": "
DefenseBadnetsTrojanBlendSIGCLBWaNet
No Defense000000
ANP84.7480.5281.3853.3582.7280.23
I-BAU78.4177.1277.5639.4678.0780.65
AWM79.3778.2479.8144.5179.8679.18
FT-SAM85.5680.6984.4957.6482.0483.62
NFT (Ours)86.8281.1585.6155.1886.2385.70
", + "image_path": "c284aa376c7fa0ebec1a8f7d62c6585f5869ed55645823f49678f6447177e32b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 397, + 120, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 397, + 120, + 411 + ], + "spans": [ + { + "bbox": [ + 31, + 397, + 120, + 411 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 423, + 382, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 423, + 382, + 579 + ], + "spans": [ + { + "bbox": [ + 30, + 423, + 382, + 579 + ], + "type": "text", + "content": "We proposed a backdoor purification framework, NFT, utilizing an augmentation-based neural mask fine-tuning approach. NFT can change the backdoor model weights in a computationally efficient manner while ensuring SOTA purification performance. Our proposed method showed that the addition of MixUp during fine-tuning replaces the need for a computationally expensive trigger synthesis process. Furthermore, we proposed a novel mask regularizer that helps us preserve the cluster separability of the original backdoor model. By preserving this separability, the proposed method offers better clean test accuracy compared to SOTA methods. Furthermore, we suggested using a mask scheduling function that reduces the mask search space and improves the computational efficiency further. Our extensive experiments on 5 different tasks validate the efficiency and efficacy of the proposed backdoor purification method. We also conducted a detailed ablation study to explain the reasoning behind our design choices." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 129, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 129, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 129, + 20 + ], + "type": "text", + "content": "N. Karim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 145, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 145, + 46 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 145, + 46 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 58, + 380, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 58, + 380, + 81 + ], + "spans": [ + { + "bbox": [ + 31, + 58, + 380, + 81 + ], + "type": "text", + "content": "This work was supported in part by the National Science Foundation under Grant ECCS-1810256 and CMMI-2246672." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 100, + 97, + 112 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 100, + 97, + 112 + ], + "spans": [ + { + "bbox": [ + 32, + 100, + 97, + 112 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 124, + 380, + 583 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 36, + 124, + 379, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 124, + 379, + 158 + ], + "spans": [ + { + "bbox": [ + 36, + 124, + 379, + 158 + ], + "type": "text", + "content": "1. Ahmed, S., Al Arafat, A., Rizve, M.N., Hossain, R., Guo, Z., Rakin, A.S.: Ssda: Secure source-free domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19180-19190 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 36, + 158, + 380, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 158, + 380, + 190 + ], + "spans": [ + { + "bbox": [ + 36, + 158, + 380, + 190 + ], + "type": "text", + "content": "2. Barni, M., Kallas, K., Tondi, B.: A new backdoor attack in cnns by training set corruption without label poisoning. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 101-105. IEEE (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 36, + 190, + 380, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 190, + 380, + 223 + ], + "spans": [ + { + "bbox": [ + 36, + 190, + 380, + 223 + ], + "type": "text", + "content": "3. Bian, J., Al Arafat, A., Xiong, H., Li, J., Li, L., Chen, H., Wang, J., Dou, D., Guo, Z.: Machine learning in real-time internet of things (iot) systems: A survey. IEEE Internet of Things Journal 9(11), 8364-8386 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 223, + 380, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 223, + 380, + 289 + ], + "spans": [ + { + "bbox": [ + 37, + 223, + 380, + 289 + ], + "type": "text", + "content": "4. Bojar, O., Buck, C., Federmann, C., Haddow, B., Koehn, P., Leveling, J., Monz, C., Pecina, P., Post, M., Saint-Amand, H., Soricut, R., Specia, L., Tamchyna, A.: Findings of the 2014 workshop on statistical machine translation. In: Proceedings of the Ninth Workshop on Statistical Machine Translation. pp. 12-58. Association for Computational Linguistics, Baltimore, Maryland, USA (Jun 2014). https://doi.org/10.3115/v1/W14-3302, https://aclanthology.org/W14-3302" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 289, + 379, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 289, + 379, + 310 + ], + "spans": [ + { + "bbox": [ + 37, + 289, + 379, + 310 + ], + "type": "text", + "content": "5. Carratino, L., Cisse, M., Jenatton, R., Vert, J.P.: On mixup regularization. The Journal of Machine Learning Research 23(1), 14632-14662 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 310, + 379, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 310, + 379, + 332 + ], + "spans": [ + { + "bbox": [ + 37, + 310, + 379, + 332 + ], + "type": "text", + "content": "6. Chai, S., Chen, J.: One-shot neural backdoor erasing via adversarial weight masking. arXiv preprint arXiv:2207.04497 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 332, + 379, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 332, + 379, + 354 + ], + "spans": [ + { + "bbox": [ + 37, + 332, + 379, + 354 + ], + "type": "text", + "content": "7. Chen, H., Fu, C., Zhao, J., Koushanfar, F.: Deepinspect: A black-box trojan detection and mitigation framework for deep neural networks. In: IJCAI. p. 8 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 37, + 354, + 379, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 354, + 379, + 386 + ], + "spans": [ + { + "bbox": [ + 37, + 354, + 379, + 386 + ], + "type": "text", + "content": "8. Chen, K., Lou, X., Xu, G., Li, J., Zhang, T.: Clean-image backdoor: Attacking multi-label models with poisoned labels only. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 37, + 387, + 379, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 387, + 379, + 408 + ], + "spans": [ + { + "bbox": [ + 37, + 387, + 379, + 408 + ], + "type": "text", + "content": "9. Chen, X., Liu, C., Li, B., Lu, K., Song, D.: Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 408, + 379, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 408, + 379, + 429 + ], + "spans": [ + { + "bbox": [ + 33, + 408, + 379, + 429 + ], + "type": "text", + "content": "10. Cheng, S., Liu, Y., Ma, S., Zhang, X.: Deep feature space trojan attack of neural networks by controlled detoxification. In: AAAI. pp. 1148-1156 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 430, + 379, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 430, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 33, + 430, + 379, + 452 + ], + "type": "text", + "content": "11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 452, + 379, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 452, + 379, + 485 + ], + "spans": [ + { + "bbox": [ + 33, + 452, + 379, + 485 + ], + "type": "text", + "content": "12. Doan, K., Lao, Y., Zhao, W., Li, P.: Lira: Learnable, imperceptible and robust backdoor attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11966-11976 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "spans": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "text", + "content": "13. Edraki, M., Karim, N., Rahnavard, N., Mian, A., Shah, M.: Odyssey: Creation, analysis and detection of trojan models. IEEE Transactions on Information Forensics and Security 16, 4521-4533 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 518, + 379, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 518, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 33, + 518, + 379, + 550 + ], + "type": "text", + "content": "14. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "text", + "content": "15. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=6Tm1mposlrM" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 74, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 74, + 9, + 346, + 19 + ], + "type": "text", + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "text", + "content": "16. Gao, Y., Xu, C., Wang, D., Chen, S., Ranasinghe, D.C., Nepal, S.: Strip: A defence against trojan attacks on deep neural networks. In: Proceedings of the 35th Annual Computer Security Applications Conference. pp. 113-125 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 68, + 379, + 89 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 68, + 379, + 89 + ], + "spans": [ + { + "bbox": [ + 33, + 68, + 379, + 89 + ], + "type": "text", + "content": "17. Gu, T., Liu, K., Dolan-Gavitt, B., Garg, S.: Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access 7, 47230-47244 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 89, + 379, + 110 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 89, + 379, + 110 + ], + "spans": [ + { + "bbox": [ + 33, + 89, + 379, + 110 + ], + "type": "text", + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: European conference on computer vision. pp. 630-645. Springer (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 110, + 379, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 110, + 379, + 142 + ], + "spans": [ + { + "bbox": [ + 33, + 110, + 379, + 142 + ], + "type": "text", + "content": "19. Hong, S., Chandrasekaran, V., Kaya, Y., Dumitras, T., Papernot, N.: On the effectiveness of mitigating data poisoning attacks with gradient shaping. arXiv preprint arXiv:2002.11497 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 142, + 379, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 142, + 379, + 175 + ], + "spans": [ + { + "bbox": [ + 33, + 142, + 379, + 175 + ], + "type": "text", + "content": "20. Janai, J., Güney, F., Behl, A., Geiger, A., et al.: Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends in Computer Graphics and Vision 12(1-3), 1-308 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 175, + 379, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 175, + 379, + 207 + ], + "spans": [ + { + "bbox": [ + 33, + 175, + 379, + 207 + ], + "type": "text", + "content": "21. Karim, N., Arafat, A.A., Rakin, A.S., Guo, Z., Rahnavard, N.: Fisher information guided purification against backdoor attacks. In: Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (CCS) (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 207, + 379, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 207, + 379, + 228 + ], + "spans": [ + { + "bbox": [ + 33, + 207, + 379, + 228 + ], + "type": "text", + "content": "22. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 228, + 379, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 228, + 379, + 260 + ], + "spans": [ + { + "bbox": [ + 33, + 228, + 379, + 260 + ], + "type": "text", + "content": "23. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 260, + 379, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 260, + 379, + 282 + ], + "spans": [ + { + "bbox": [ + 33, + 260, + 379, + 282 + ], + "type": "text", + "content": "24. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 282, + 379, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 282, + 379, + 303 + ], + "spans": [ + { + "bbox": [ + 33, + 282, + 379, + 303 + ], + "type": "text", + "content": "25. Levine, A., Feizi, S.: Deep partition aggregation: Provable defense against general poisoning attacks. arXiv preprint arXiv:2006.14768 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 303, + 379, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 303, + 379, + 336 + ], + "spans": [ + { + "bbox": [ + 33, + 303, + 379, + 336 + ], + "type": "text", + "content": "26. Li, X., Chen, Z., Zhao, Y., Tong, Z., Zhao, Y., Lim, A., Zhou, J.T.: Pointba: Towards backdoor attacks in 3d point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16492-16501 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 336, + 379, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 336, + 379, + 368 + ], + "spans": [ + { + "bbox": [ + 33, + 336, + 379, + 368 + ], + "type": "text", + "content": "27. Li, Y., Lyu, X., Koren, N., Lyu, L., Li, B., Ma, X.: Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems 34 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 368, + 379, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 368, + 379, + 410 + ], + "spans": [ + { + "bbox": [ + 33, + 368, + 379, + 410 + ], + "type": "text", + "content": "28. Li, Y., Lyu, X., Koren, N., Lyu, L., Li, B., Ma, X.: Neural attention distillation: Erasing backdoor triggers from deep neural networks. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=910K40M-oXE" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 410, + 379, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 410, + 379, + 432 + ], + "spans": [ + { + "bbox": [ + 33, + 410, + 379, + 432 + ], + "type": "text", + "content": "29. Li, Y., Lyu, X., Ma, X., Koren, N., Lyu, L., Li, B., Jiang, Y.G.: Reconstructive neuron pruning for backdoor defense. arXiv preprint arXiv:2305.14876 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 432, + 379, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 432, + 379, + 454 + ], + "spans": [ + { + "bbox": [ + 33, + 432, + 379, + 454 + ], + "type": "text", + "content": "30. Li, Y., Wu, B., Jiang, Y., Li, Z., Xia, S.T.: Backdoor learning: A survey. arXiv preprint arXiv:2007.08745 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 454, + 379, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 454, + 379, + 486 + ], + "spans": [ + { + "bbox": [ + 33, + 454, + 379, + 486 + ], + "type": "text", + "content": "31. Li, Y., Li, Y., Wu, B., Li, L., He, R., Lyu, S.: Invisible backdoor attack with sample-specific triggers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16463-16472 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 486, + 379, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 486, + 379, + 507 + ], + "spans": [ + { + "bbox": [ + 33, + 486, + 379, + 507 + ], + "type": "text", + "content": "32. Lin, J., Xu, L., Liu, Y., Zhang, X.: Composite backdoor attack for deep neural network by mixing existing benign features. In: CCS. pp. 113-131 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 507, + 379, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 507, + 379, + 551 + ], + "spans": [ + { + "bbox": [ + 33, + 507, + 379, + 551 + ], + "type": "text", + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "text", + "content": "34. Liu, K., Dolan-Gavitt, B., Garg, S.: Fine-pruning: Defending against backdooring attacks on deep neural networks. In: International Symposium on Research in Attacks, Intrusions, and Defenses. pp. 273-294. Springer (2018)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "text", + "content": "N. Karim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 57 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 379, + 57 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 379, + 57 + ], + "type": "text", + "content": "35. Liu, Y., Ma, S., Afer, Y., Lee, W.C., Zhai, J., Wang, W., Zhang, X.: Trojaning attack on neural networks (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 57, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 57, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 31, + 57, + 379, + 79 + ], + "type": "text", + "content": "36. Liu, Y., Xie, Y., Srivastava, A.: Neural trojans. In: 2017 IEEE International Conference on Computer Design (ICCD). pp. 45-48. IEEE (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 79, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 79, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 32, + 79, + 379, + 100 + ], + "type": "text", + "content": "37. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 100, + 379, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 100, + 379, + 121 + ], + "spans": [ + { + "bbox": [ + 32, + 100, + 379, + 121 + ], + "type": "text", + "content": "38. Manoj, N., Blum, A.: Excess capacity and backdoor poisoning. Advances in Neural Information Processing Systems 34, 20373-20384 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 121, + 379, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 121, + 379, + 142 + ], + "spans": [ + { + "bbox": [ + 32, + 121, + 379, + 142 + ], + "type": "text", + "content": "39. Nguyen, A., Tran, A.: Wanet-imperceptible warping-based backdoor attack. arXiv preprint arXiv:2102.10369 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 142, + 379, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 142, + 379, + 164 + ], + "spans": [ + { + "bbox": [ + 32, + 142, + 379, + 164 + ], + "type": "text", + "content": "40. Nguyen, T.A., Tran, A.: Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems 33, 3454-3464 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 164, + 379, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 164, + 379, + 196 + ], + "spans": [ + { + "bbox": [ + 32, + 164, + 379, + 196 + ], + "type": "text", + "content": "41. Ning, R., Li, J., Xin, C., Wu, H.: Invisible poison: A blackbox clean label backdoor attack to deep neural networks. In: IEEE INFOCOM 2021-IEEE Conference on Computer Communications. pp. 1-10. IEEE (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 196, + 379, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 196, + 379, + 228 + ], + "spans": [ + { + "bbox": [ + 32, + 196, + 379, + 228 + ], + "type": "text", + "content": "42. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 32, + 228, + 379, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 228, + 379, + 271 + ], + "spans": [ + { + "bbox": [ + 32, + 228, + 379, + 271 + ], + "type": "text", + "content": "43. Qiu, H., Zeng, Y., Guo, S., Zhang, T., Qiu, M., Thuraisingham, B.: Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. In: Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security. pp. 363-377 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 32, + 271, + 379, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 271, + 379, + 304 + ], + "spans": [ + { + "bbox": [ + 32, + 271, + 379, + 304 + ], + "type": "text", + "content": "44. Ridnik, T., Sharir, G., Ben-Cohen, A., Ben-Baruch, E., Noy, A.: ML-decoder: Scalable and versatile classification head. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 32-41 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 32, + 304, + 379, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 304, + 379, + 336 + ], + "spans": [ + { + "bbox": [ + 32, + 304, + 379, + 336 + ], + "type": "text", + "content": "45. Sherstinsky, A.: Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. Physica D: Nonlinear Phenomena 404, 132306 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 32, + 336, + 379, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 336, + 379, + 357 + ], + "spans": [ + { + "bbox": [ + 32, + 336, + 379, + 357 + ], + "type": "text", + "content": "46. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 32, + 357, + 379, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 357, + 379, + 390 + ], + "spans": [ + { + "bbox": [ + 32, + 357, + 379, + 390 + ], + "type": "text", + "content": "47. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: The german traffic sign recognition benchmark: a multi-class classification competition. In: The 2011 international joint conference on neural networks. pp. 1453-1460. IEEE (2011)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 32, + 389, + 379, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 389, + 379, + 411 + ], + "spans": [ + { + "bbox": [ + 32, + 389, + 379, + 411 + ], + "type": "text", + "content": "48. Tiedemann, J.: Parallel data, tools and interfaces in opus. In: Lrec. vol. 2012, pp. 2214-2218. Citeseer (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 32, + 411, + 379, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 411, + 379, + 432 + ], + "spans": [ + { + "bbox": [ + 32, + 411, + 379, + 432 + ], + "type": "text", + "content": "49. Tran, B., Li, J., Madry, A.: Spectral signatures in backdoor attacks. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 32, + 432, + 379, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 432, + 379, + 454 + ], + "spans": [ + { + "bbox": [ + 32, + 432, + 379, + 454 + ], + "type": "text", + "content": "50. Turner, A., Tsipras, D., Madry, A.: Clean-label backdoor attacks (2019), https://openreview.net/forum?id=HJg6e2CcK7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 32, + 454, + 379, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 454, + 379, + 486 + ], + "spans": [ + { + "bbox": [ + 32, + 454, + 379, + 486 + ], + "type": "text", + "content": "51. Wang, B., Yao, Y., Shan, S., Li, H., Viswanath, B., Zheng, H., Zhao, B.Y.: Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In: 2019 IEEE Symposium on Security and Privacy (SP). pp. 707-723. IEEE (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 32, + 486, + 379, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 486, + 379, + 519 + ], + "spans": [ + { + "bbox": [ + 32, + 486, + 379, + 519 + ], + "type": "text", + "content": "52. Wang, R., Zhang, G., Liu, S., Chen, P.Y., Xiong, J., Wang, M.: Practical detection of trojan neural networks: Data-limited and data-free cases. In: European Conference on Computer Vision. pp. 222-238. Springer (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 32, + 519, + 379, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 519, + 379, + 562 + ], + "spans": [ + { + "bbox": [ + 32, + 519, + 379, + 562 + ], + "type": "text", + "content": "53. Wang, Z., Zhai, J., Ma, S.: Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15074-15084 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 32, + 562, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 562, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 562, + 379, + 583 + ], + "type": "text", + "content": "54. Wu, D., Wang, Y.: Adversarial neuron pruning purifies backdoored deep models. In: NeurIPS (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 74, + 9, + 346, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 9, + 346, + 20 + ], + "spans": [ + { + "bbox": [ + 74, + 9, + 346, + 20 + ], + "type": "text", + "content": "Augmented Neural Fine-Tuning for Efficient Backdoor Purification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 408 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "text", + "content": "55. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems 33, 2958-2969 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "type": "text", + "content": "56. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 101, + 379, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 101, + 379, + 134 + ], + "spans": [ + { + "bbox": [ + 33, + 101, + 379, + 134 + ], + "type": "text", + "content": "57. Xian, X., Wang, G., Srinivasa, J., Kundu, A., Bi, X., Hong, M., Ding, J.: Understanding backdoor attacks through the adaptability hypothesis. In: Proc. International Conference on Machine Learning (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 134, + 379, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 134, + 379, + 167 + ], + "spans": [ + { + "bbox": [ + 33, + 134, + 379, + 167 + ], + "type": "text", + "content": "58. Xiang, Z., Miller, D.J., Chen, S., Li, X., Kesidis, G.: A backdoor attack against 3d point cloud classifiers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7597-7607 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 167, + 379, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 167, + 379, + 189 + ], + "spans": [ + { + "bbox": [ + 33, + 167, + 379, + 189 + ], + "type": "text", + "content": "59. Zeng, Y., Chen, S., Park, W., Mao, Z.M., Jin, M., Jia, R.: Adversarial unlearning of backdoors via implicit hypergradient. arXiv preprint arXiv:2110.03735 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 189, + 379, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 189, + 379, + 210 + ], + "spans": [ + { + "bbox": [ + 33, + 189, + 379, + 210 + ], + "type": "text", + "content": "60. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 210, + 379, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 210, + 379, + 233 + ], + "spans": [ + { + "bbox": [ + 33, + 210, + 379, + 233 + ], + "type": "text", + "content": "61. Zhang, L., Deng, Z., Kawaguchi, K., Ghorbani, A., Zou, J.: How does mixup help with robustness and generalization? arXiv preprint arXiv:2010.04819 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "spans": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "text", + "content": "62. Zhang, Z., Liu, Q., Wang, Z., Lu, Z., Hu, Q.: Backdoor defense via deconfounded representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12228-12238 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 265, + 379, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 265, + 379, + 298 + ], + "spans": [ + { + "bbox": [ + 33, + 265, + 379, + 298 + ], + "type": "text", + "content": "63. Zhao, P., Chen, P.Y., Das, P., Ramamurthy, K.N., Lin, X.: Bridging mode connectivity in loss landscapes and adversarial robustness. arXiv preprint arXiv:2005.00060 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 298, + 379, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 298, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 33, + 298, + 379, + 331 + ], + "type": "text", + "content": "64. Zhao, S., Ma, X., Zheng, X., Bailey, J., Chen, J., Jiang, Y.G.: Clean-label backdoor attacks on video recognition models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14443-14452 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 331, + 379, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 331, + 379, + 375 + ], + "spans": [ + { + "bbox": [ + 33, + 331, + 379, + 375 + ], + "type": "text", + "content": "65. Zheng, R., Tang, R., Li, J., Liu, L.: Data-free backdoor removal based on channel lipschitzness. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V. pp. 175-191. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 375, + 379, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 375, + 379, + 408 + ], + "spans": [ + { + "bbox": [ + 33, + 375, + 379, + 408 + ], + "type": "text", + "content": "66. Zhu, M., Wei, S., Shen, L., Fan, Y., Wu, B.: Enhancing fine-tuning based backdoor defense with sharpness-aware minimization. arXiv preprint arXiv:2304.11823 (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 129, + 19 + ], + "type": "text", + "content": "N. Karim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_content_list.json b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7dc6f58d5c313375d1c35877665035378af66200 --- /dev/null +++ b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "Auto-DAS: Automated Proxy Discovery for Training-free Distillation-aware Architecture Search", + "text_level": 1, + "bbox": [ + 248, + 140, + 754, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haosen Sun $^{1}$ , Lujun Li $^{1\\boxtimes}$ , Peijie Dong $^{2}$ , Zimian Wei $^{3}$ , and Shitong Shao $^{4}$", + "bbox": [ + 236, + 233, + 764, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ The Hong Kong University of Science and Technology $^{2}$ The Hong Kong University of Science and Technology (Guangzhou)", + "bbox": [ + 241, + 261, + 733, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ National University of Defense Technology", + "bbox": [ + 352, + 290, + 648, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Southeast University \nhsunas@connect.ust.hk;lilujunai@gmail.com;pdong212@connect.hkust-gz.edu.cn; weizimian16@nudt.edu.cn; shaoshitong@seu.edu.cn", + "bbox": [ + 215, + 304, + 795, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract Distillation-aware Architecture Search (DAS) seeks to discover the ideal student architecture that delivers superior performance by distilling knowledge from a given teacher model. Previous DAS methods involve time-consuming training-based search processes. Recently, the training-free DAS method (i.e., DisWOT) proposes KD-based proxies and achieves significant search acceleration. However, we observe that DisWOT suffers from limitations such as the need for manual design and poor generalization to diverse architectures, such as the Vision Transformer (ViT). To address these issues, we present Auto-DAS, an automatic proxy discovery framework using an Evolutionary Algorithm (EA) for training-free DAS. Specifically, we empirically find that proxies conditioned on student instinct statistics and teacher-student interaction statistics can effectively predict distillation accuracy. Then, we represent the proxy with computation graphs and construct the proxy search space using instinct and interaction statistics as inputs. To identify promising proxies, our search space incorporates various types of basic transformations and network distance operators inspired by previous proxy and KD-loss designs. Next, our EA initializes populations, evaluates, performs crossover and mutation operations, and selects the best correlation candidate with distillation accuracy. We introduce an adaptive-elite selection strategy to enhance search efficiency and strive for a balance between exploitation and exploration. Finally, we conduct training-free DAS with discovered proxy before the optimal student distillation phase. In this way, our auto-discovery framework eliminates the need for manual design and tuning, while also adapting to different search spaces through direct correlation optimization. Extensive experiments demonstrate that Auto-DAS generalizes well to various architectures and search spaces (e.g., ResNet, ViT, NAS-Bench-101, and NAS-Bench-201), achieving state-of-the-art results in both ranking correlation and final searched accuracy. Code at: https://github.com/lliai/Auto-DAS.", + "bbox": [ + 261, + 377, + 738, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Knowledge Distillation $\\cdot$ Architecture Search $\\cdot$ AutoML", + "bbox": [ + 261, + 805, + 720, + 819 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author & contributed equally to this work.", + "bbox": [ + 228, + 825, + 640, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation (KD) transfers logits [17, 52], features [26, 32], or relation [40, 47] knowledge of heavy teacher models as additional supervised signals to augment lightweight student models. KD originally served as model compression technique for large models and has recently been used as a common strategy for performance improvement in different scenarios. For example, distilling a CNN teacher for a Vision Transformer (ViT) [49] can significantly enhance its performance, particularly when training from scratch in data-limited scenarios. However, the architecture gap between the teacher and student network creates a mismatch between the receptive field and the prediction reliability [38, 56]. These disparities contribute to distillation gains, i.e., students with a smaller architecture gap with teachers will enjoy better final results.", + "bbox": [ + 212, + 174, + 787, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate these problems, a new task, Distillation-aware Architecture Search (DAS), is presented to search for an optimal student architecture from the given teacher, which obtains the best performance with the help of neural architecture search. In contrast to the traditional NAS methods, the DAS approach needs to consider the effect of the teacher and employ distillation accuracy as the search evaluation metric. Some observations [7,36] indicate that the models searched by traditional NAS methods [9,20,35] are suboptimal in distillation, and there is a common performance ranking gap for obtained students between vanilla training and distillation. For the DAS task, some training-based methods leverage Reinforcement Learning (RL) [36] or gradient method [14] as search strategies to find the ideal student. However, these methods bring in high search costs (e.g., 5 days with 200 TPUv2 in AKD [36]), which causes heavy burdens for application. Recently, DisWOT [7] first presents the training-free DAS framework with some KD-based proxies, exploiting the connection between training-free proxies and knowledge distillation loss. It achieves significant search acceleration by using statistics between random networks for predicting distillation accuracy without candidate training. However, some drawbacks still limit its extensive application, which could highlight from two aspects: (1) Expert design and tuning costs. These KD-based proxies are inspired by some distillation studies with extensive expert intuition and require extensive tuning. It is hard to state if these proxies are optimal for DAS as limited by the designer's experience. (2) Generalization issue. Handcrafted ones only perform well on some specific settings but are hard to adapt to different scenarios without changing/modification. As shown in Figure 1, existing handcrafted proxies perform well on the CNNs but do not generalize well to ViT models. Thus, two questions arise: (1) How to efficiently discover the proxies without expert knowledge? (2) How to find the optimal proxies for different search spaces?", + "bbox": [ + 212, + 340, + 787, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the first problem, we observe that previous proxy expressions are usually constructed for typical inputs and common primitive operations. Inspired by Auto-Zero [42], we explore a from-scratch proxy discover framework alternative to handcrafted for DAS task. For the second problem, we can evolve various proxy candidates by directly taking their predictability for distillation accuracy as feedback. In addition, we find that proxies with inputs from student", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Haosen Sun et al.", + "bbox": [ + 271, + 114, + 388, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d12551264457751c84f472b52274f3cecd4b9be8d16ed1feec256c248236d827.jpg", + "image_caption": [ + "Figure 1: Spearman correlation $(\\%)$ on CNN models(left) and ViT models(right)." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 504, + 239 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/583e57acfb12c7a207318434e031723c5fc7098778de63e64f4d9b30c71a3b05.jpg", + "image_caption": [ + "Figure 2: Illustration of Auto-DAS." + ], + "image_footnote": [], + "bbox": [ + 519, + 142, + 767, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distinct statistics and teacher-student interaction statistics consistently achieve better results. For example, DisWOT+Synflow in Figure 1 achieves stable improvements on multiple search spaces. Based on these observations, we present Auto-DAS, a from-scratch framework for discovering proxies for DAS tasks are proposed as an alternative to manually designed ones. As shown in Figure 2, we first build the proxy search space, consisting of both instinct and interaction proxy. The instinct proxy is used with student's weights and gradient statistics as input and different unary and binary mathematical operations as candidate components. For interaction proxy, we select teacher-student features, embeddings, logits as inputs, different basic transformation operations, and distance functions to form the computation graph. Then, we randomly sample various proxies as the population and then perform mutation and crossover operations according to the ranking ability of different proxies. To improve generality, we directly treat the Kendall correlation between the prediction of the candidate proxy and the ground truth under distillation as the fitting objective of the evolutionary algorithm. To speed up the process, we employ an adaptive-elite selection strategy during the evolution search. With the automatic search framework, our Auto-DAS surpasses existing training-free NAS approaches by a large margin without involving too much prior knowledge. Finally, after getting the Auto-DAS proxy in a given search space, we search for the best student architectures with the highest Auto-DAS score and apply the searched students to distill the knowledge from the given teacher.", + "bbox": [ + 212, + 309, + 787, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In comparison to manual DAS techniques, our Auto-DASframework offers several advantages: (1) Autonomy. It facilitates the spontaneous discovery of more potent and efficient proxies that may elude human experts, thus mitigating bias and ensuring that the resultant architectures are tailored for the intended problem or dataset. (2) Versatility. The search mechanism of AutoDASenhances adaptability by uncovering proxies that perform well across diverse search spaces. Our framework can evolve flexible proxies to generalize across architectures, based on the ultimate distillation outcomes for the search objectives. (3) Enlightening. Our approach enables the automatic identification of more expressive, efficient, and effective proxies for optimizing and customizing student architectures, paving the way for novel research and applications in DAS and KD tasks. (4) Efficacy. We have conducted comprehensive experiments on ResNet, NAS-Bench-101, NAS-Bench-201, and ViT-Bench search spaces to confirm the", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Auto-DAS", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "superiority of our Auto-DAS. The results demonstrate that our Auto-DAScan surpass other zero-shot proxies in distillation accuracy and achieve state-of-the-art results in rank consistency, such as a $40\\%$ improvement in Kendall correlation over DisWOT. To summarize, our key contributions to Auto-DAS are:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the pioneering framework for discovering proxies from scratch, Auto-DAS, to address handcrafted limitations and enhance generalizability in predicting distillation accuracy across various students.", + "- We devise a well-structured proxy search space, incorporating instinct and interaction statistics as inputs, along with diverse unary, binary, transforms, and distance functions as operations. We develop evolutionary algorithms with adaptive-elite selection for effective and efficient proxy search.", + "- We provide empirical evidence that Auto-DASachieves cutting-edge performance across ResNet, NAS-Bench-101, NAS-Bench-201, ViT-Bench search spaces and multiple datasets." + ], + "bbox": [ + 225, + 220, + 784, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 398, + 382, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distillation-aware architecture search. The increasing demand for practical applications of large-scale models [5, 12] has led to impressive advancements in Knowledge distillation (KD) across various domains [26-30, 32, 43], including image classification [17], object detection [51, 55], and NLP [21]. However, discrepancies between teacher and student architectures can substantially impact distillation outcomes. Consequently, distillation-aware student architecture search for a given teacher model has emerged as a crucial new area of research. Traditional reinforcement learning-based search methods [36] utilize KD-loss as feedback for DAS but incur significant additional costs during the search process. Subsequent research employs efficient gradient-based methods [14], albeit introducing new optimization challenges. Recently, DisWOT [7] introduced the first training-free DAS framework. However, DisWOT relies on a manually designed proxy, which lacks generalization across different search spaces. To address this limitation, we propose the first automated proxy discovery for training-free DAS tasks. Our method not only retains the advantages of DisWOT, such as efficiency and flexibility, but also automatically optimizes proxies without expert intervention. Auto-DAS paves the way for novel DAS research and offers valuable insights based on search result analysis for the research community.", + "bbox": [ + 212, + 431, + 787, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training-free architecture search. Conventional NAS [18, 60] involves designing search spaces, algorithms, and evaluations to automatically discover optimal architectures under given constraints. Training-based methods employ train-then-search or weight sharing processes [19, 41, 54, 60]. Training-free NAS, also known as zero-shot NAS [6, 8, 31, 33, 34, 57, 59], enables faster search by predicting performance without training. Zero-cost proxies are categorized as parameter-level or architecture-level. Parameter-level approaches rely on pruning and saliency values [23, 45, 50], while architecture-level methods evaluate expressiveness [34] or linear map correlations [37]. DisWOT proposed KD-based", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Haosen Sun et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e97d29716cc13def24a2a57360e1a9f3f2e92f997add73ccd837dafed587af9.jpg", + "image_caption": [ + "Figure 3: The overall illustration of Auto-DAS. Our search space includes unary, binary, transform, and distance options with different inputs. Then, we construct our intrinsic & interaction proxies and use crossover & mutation to search for the best proxy." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 454, + 340 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3b122400afaf5e511e6128e4655bcb0395e563f90d364110258e92d40e1704ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 142, + 756, + 340 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/162eb9b9fbd2f004362805105d6c1cf357d745b0a94663998fb870436edc57c6.jpg", + "table_caption": [ + "Table 1: Operations of Auto-DAS. More details are available in the Appendix." + ], + "table_footnote": [], + "table_body": "
ProxyInputTypePrimitive Operations
instinctActivationsUnaryexp, mish, leaky, relu, tanh, softmax, sigmoid, pow2, pow4, log, abslog, abssqrt
GradientsUnaryllnorm, revert, min - maxto - mean, to - std, invert, lnNorm, normalizedSum, no - op
WeightsBinarysum, subtract, multiply, divide, dot.
InteractionFeatureTransformtanh, log, sqrt, batchnorm, min - max, norm, softmax, exp, mish, leaky, relu
EmbedTransformmask, no, bmm, mm, scale, multi - scale, scale, local, batch, channel, drop, satt, natt, catt
LogitsDistancel1, l2, ℓKL, ℓhard, ℓCosine, ℓPearson
", + "bbox": [ + 222, + 448, + 777, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "proxies, but their fixed formulations limit generalization. To overcome these limitations, we consider student internal statistics and teacher-student similarities as inputs and evolve superior proxies from scratch for training-free DAS tasks. Inspired by Auto-Zero, our innovations include a novel proxy search space and a specialized search process, promoting automatic proxy design for distillation. Our approach differs from EZNAS [2] in terms of tasks (ours is customized for CNNs and ViTs), search algorithms (ours employs adaptive-elite selection), and search space (ours incorporates teacher-student knowledge). Unlike EZNAS and Auto-KD, which focus solely on vanilla classification, we directly search on the distillation task and include teacher-student knowledge and additional options in the search space.", + "bbox": [ + 212, + 558, + 787, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 747, + 380, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section delves into the intricacies of existing proxies, our proposed search space, the search algorithm, and an analysis of the results. Subsequently, we outline a training-free student search process utilizing the discovered zero-cost proxy. Figure 3 illustrates the workflow of our approach.", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Auto-DAS", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Examination of existing proxies", + "text_level": 1, + "bbox": [ + 215, + 146, + 522, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To provide context for our training-free method, we first examine current zero-cost proxies: SNIP [23] calculates a saliency metric at initialization using a single mini-batch of data to estimate the loss change when a specific parameter is eliminated. Synflow [45] introduces a modified version of synaptic saliency scores that prevents layer collapse during parameter pruning, while Fisher [46] computes the sum over all activation gradients in the network, applicable for channel pruning. Their detailed formulations are as follows:", + "bbox": [ + 212, + 169, + 787, + 275 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {s n i p} = \\left| \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W} \\right|, \\rho_ {s y n f l o w} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W}, \\rho_ {f i s h e r} = \\left(\\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {A}} \\mathcal {A}\\right) ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 286, + 785, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L},\\mathcal{W},\\mathcal{A}$ are loss function, weight and activation. $H$ is the Hessian matrix. These proxies consist of various unary, binary operations with statistics input from the candidate model, named instinct proxies in our paper. For the distillation scenario, DisWOT presents some KD-based proxies based on KD-loss design, as follows:", + "bbox": [ + 212, + 324, + 787, + 400 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {D i s W O T} = \\mathcal {D} _ {L 2} (\\mathcal {G} ([ \\mathcal {A} S, \\mathcal {A} T ])) + \\mathcal {D} _ {L 2} (\\mathcal {G} ([ \\mathcal {F} S, \\mathcal {F} T ])), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 411, + 785, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{A}$ , $\\mathcal{F}$ , $\\mathcal{E}$ , $\\mathcal{P}$ are activation maps, feature maps, embeddings, and logits prediction. $[\\cdot, \\cdot]$ means to perform the same operation on two primitives. $\\mathcal{G}$ is the gram-matrix. $\\mathcal{D}_{L2}$ , $\\mathcal{D}_{KL}$ are $\\ell_2$ and $\\ell_{KL}$ distance. These KD-based proxies use teacher-student pair-wise statistics as input and different transformations and distance functions to construct formations that we denote as interaction proxies. Based on these analyses, we build search spaces for instinct and interactive proxies in the next part.", + "bbox": [ + 212, + 434, + 787, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Search Space for Proxy Discovery", + "text_level": 1, + "bbox": [ + 214, + 560, + 540, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our search space structure comprises three primary components: input choices, primitive operations, and proxy representation. As illustrated in Table 1, for input choices in constructing zero-cost proxies, the instinct proxy extracts activations, gradients, and weights from the student model, while the interaction proxy selects features, embeddings, and logits of the teacher-student network. These inputs represent the most informative options common across traditional and KD-based proxies. For instance, features provide insights into data distribution and internal network representations, while gradients highlight weight sensitivity to loss functions. Following this observation, we register each student layer's activation, weights, and corresponding gradients as potential inputs for our zero-cost proxy. Regarding primitive operations, we consider unary and binary operations for instinct proxies, and transform and distance operations for interaction proxies. We include mathematical, normalization, and scaling operations for unary operations; matrix sums and products for binary operations; and activation, normalization, scaling, and other transformations, along with typical distance functions for interaction proxies. Our proxy is represented as a computation graph, where input nodes are instinct and interaction statistics, and", + "bbox": [ + 212, + 583, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Haosen Sun et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Evolutionary Strategy for Auto-DAS Proxy Discovery" + ], + "code_body": "Input: Architecture space $S$ , candidate set $\\mathcal{P}$ , iteration limit $\\mathcal{T}$ , selection proportion $r$ , temporary set $\\mathcal{R}$ , elite count $k$ . \nOutput: Highest-scoring Auto-DAS configuration. \n1: Initialize candidate set $\\mathcal{P}0 := \\text{GenerateInitialCandidates}(P_i)$ ; \n2: Empty temporary set $\\mathcal{R} := \\emptyset$ ; \n3: for $i = 1,2,\\ldots,\\mathcal{T}$ do \n4: Reset temporary set $\\mathcal{R} := \\emptyset$ ; \n5: Populate $\\mathcal{R}$ with random selections from $\\mathcal{P}$ ; \n6: Elite group $G_i k := \\text{SelectTopPerformers}(\\mathcal{R}, k)$ ; \n7: Choose parent $G_i^p := \\text{RandomlyPick}(G_i k)$ ; \n8: Generate offspring $G_i^m := \\text{MUTATE}(G_i^p)$ ; \n9: // Adaptive Elite Preservation \n10: Randomly generate baseline $E$ ; \n11: if $\\rho(G_i^m) > \\rho(E)$ then \n12: Include $G_i^m$ in $P$ ; \n13: else \n14: Include $E$ in $P$ ; \n15: end if \n16: Eliminate lowest-scoring proxy from population. \n17: end for", + "bbox": [ + 215, + 162, + 785, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "intermediate nodes are primitive operations. The graph's output is the proxy score used for distillation ranking. Our comprehensive search space encompasses most zero-cost proxies, KD-based losses, and model similarity formulations, constituting our core innovation and contribution.", + "bbox": [ + 212, + 473, + 784, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Evolution Procedure and Objective", + "text_level": 1, + "bbox": [ + 214, + 555, + 552, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Utilizing our proxy search space, we employ an Evolutionary Algorithm (EA) to identify optimal proxy expressions (refer to Figure 3 and Algorithm 1). Our EA initiates with a population of candidate proxies and iteratively evolves it over generations using genetic operators such as selection, crossover, and mutation to generate superior solutions. For objective fitting, each proxy is evaluated based on the ranking correlation between its output proxy $\\mathcal{Q}$ scores and actual performances $\\mathcal{D}$ to efficiently discover the optimal proxy $\\rho^{*}$ from search space $\\mathcal{S}$ , as follows:", + "bbox": [ + 212, + 579, + 784, + 698 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {A u t o - D A S} ^ {*} = \\arg \\max \\rho \\in \\mathcal {S} (\\tau (\\mathcal {D}, \\mathcal {Q})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 700, + 784, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where Kendall's Tau $\\tau$ serves as the correlation coefficient. Each evolution generation selects the top-k candidates with the highest Auto-DAS scores and randomly chooses a parent from these candidates for mutation. During mutation, a node in the computation graph is randomly selected and mutated with newly generated primitive operations.", + "bbox": [ + 212, + 719, + 782, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficient proxy search with existing benchmark. Obtaining pre-trained results $\\mathcal{D}$ can be resource-intensive. To address this, we leverage pre-computed architecture-performance pairs from existing DisWOT benchmarks [7]. These", + "bbox": [ + 212, + 795, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Auto-DAS", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1574db6e0a5d877c676ddfcae74fa4b96f79bbe72cec7416dcdf1fc68b316557.jpg", + "image_caption": [ + "Figure 4: SearchFigure 5: Correlation visualization of NWOT (left), DisWOT curve on ResNet. (middle), Auto-DAS (right) on ResNet search space." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 359, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fc68fb86f5381bfe872a5a8275673afc92045b793d2fe30ba3e4e9f9bdf4933.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 143, + 496, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61c3ce3061e85c0eba9a00b7d38a2643303b3c5e614c2e15f28b45621882ba63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 143, + 638, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab24aa712070466b8d6158c7f20bca7c44d9e974e18a9e304b02bf5ecad163c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 143, + 779, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "benchmarks provide distillation results for various ResNet and NAS-Bench-101/201 search space models. Utilizing this benchmark eliminates the need to retrain performance ground-truths and significantly reduces the proxy search process time to merely 0.05 GPU-day. Due to its excellent generality (demonstrated in our extensive experiments), our searched proxy $\\rho_{Auto - DAS}$ transfers well to similar search spaces and can be directly applied in most scenarios without additional search budget.", + "bbox": [ + 212, + 294, + 782, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive-elite selection strategy. To mitigate population deterioration and premature convergence, we introduce an adaptive-elite selection strategy. This approach involves comparing the performance of mutation-generated offspring with that of a randomly generated individual. The individual with a higher Auto-DAS score is then added to the population. By maintaining or enhancing the overall population performance over time, this strategy accelerates convergence towards a high-performing solution (see Figure 4).", + "bbox": [ + 212, + 401, + 784, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Searched Zero-cost Proxy", + "text_level": 1, + "bbox": [ + 215, + 531, + 473, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present formulas of the searched proxies on $\\rho_{CNN}$ for CNN Search space (e.g., ResNet, NAS-Bench) and $\\rho_{ViT}$ ViT search space as follows:", + "bbox": [ + 214, + 558, + 787, + 589 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {C N N} = \\mathcal {D} _ {K L} (\\operatorname {s i g m o i d} ([ \\mathcal {F} _ {S}, \\mathcal {F} _ {T} ])) + \\frac {1}{M} \\sum \\left| \\log \\operatorname {s o f t m a x} \\left(\\frac {\\partial L}{\\partial W}\\right) \\right| _ {F}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 599, + 785, + 627 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {V I T} = \\mathcal {D} _ {K L} (\\operatorname {m a s k} ([ \\mathcal {F} _ {S}, \\mathcal {F} _ {T} ])) + \\frac {M}{\\sum \\operatorname {s i g m o i d} \\left( \\right.\\left| \\right. \\log \\left| \\frac {\\partial L}{\\partial W} \\right|\\left. \\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 638, + 785, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Where $W$ is the weight parameter, $\\partial \\mathcal{L} / \\partial W$ is the corresponding gradient. $\\| \\cdot \\| _F$ means the Frobenius-norm. $M$ denotes mean operation.", + "bbox": [ + 212, + 672, + 782, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of searched proxy. The discovered formulations indicate that (1) mathematical operations like softmax, logsoftmax, sigmoid, and mask operations benefit proxy predictability. (2) Gradient and features are vital input statistics for instinct and interaction proxies. As shown in Figure 5, the discovered proxies of Auto-DAS achieve significant ranking improvement.", + "bbox": [ + 212, + 703, + 782, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Understanding why our Auto-DAS surpasses other NAS methods. (1) Many training-based distillation NAS methods use weight sharing and proxies, often leading to inaccurate performance estimates. Our Auto-DAS searches proxies based on real performance, yielding one with strong ranking ability. (2)", + "bbox": [ + 214, + 779, + 784, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Haosen Sun et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9d0bae625201bc26f364ce6eeb17005bc7c6fe8adc139ddd0618e11cca59faea.jpg", + "table_caption": [ + "Table 2: Spearman correlation (\\%) results on the NAS-Bench-101 and NAS-Bench-201 (NB-101/201) datasets. In this context, NB-101/201-KD denotes the task of distilling the accuracies of the architectures on the respective NAS-Bench-101/201 datasets." + ], + "table_footnote": [], + "table_body": "
MethodResNetNB-101NB-201NB-101-KDNB-201-KD
FLOPs72.9230.8163.3815.5664.55
Fisher81.37-38.8135.91-33.924.45
Grad_Norm82.35-39.2358.70-39.16-10.01
SNIP85.07-29.0158.17-21.7816.91
Synflow88.3043.6974.6120.3674.63
NWOT45.6632.8464.4122.9735.27
DisWOT(Ms)77.2449.6165.7450.1653.88
DisWOT(Mr)49.3830.7456.4642.9445.27
DisWOT91.3846.5772.3652.4564.90
Auto-DAS98.4471.8487.7385.9574.66
", + "bbox": [ + 313, + 198, + 687, + 351 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Other train-free NAS methods primarily target vanilla training, not distillation. They rely on weak correlation proxies. Ours exhibits a stronger correlation, ensuring more efficient model discovery. While DisWOT considers distillation in its proxy design, it is manually crafted and lacks direct optimization and feedback from distillation results. In contrast, our Auto-DAS method not only includes all distillation designs in the proxy search space settings but also optimizes our proxies directly based on the distillation results. Our approach naturally outperforms other methods in terms of distillation performance.", + "bbox": [ + 212, + 378, + 787, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Training-free Student Search", + "text_level": 1, + "bbox": [ + 215, + 518, + 501, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The training-free student search enables us to efficiently explore numerous candidate architectures without the need for costly and time-consuming training. After obtaining an effective proxy with the aid of an evolutionary search, we employ it to conduct a training-free student search. Utilizing randomly initialized weights, denoted as $\\mathcal{W}$ , we seek the optimal student, represented as $\\alpha^{*}$ , from the search space $\\mathcal{A}$ , as follows:", + "bbox": [ + 212, + 542, + 787, + 633 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha^ {*} = \\arg \\max _ {\\alpha \\in \\mathcal {A}} \\rho_ {A u t o - D A S} ^ {*} (\\alpha , \\mathcal {W}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 641, + 785, + 660 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After the search phase, we use the pre-trained teacher model to distill the optimal student network $\\alpha^{*}$ . We adopt the original knowledge distillation method [17] as the auxiliary loss function.", + "bbox": [ + 212, + 667, + 787, + 714 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experimental results", + "text_level": 1, + "bbox": [ + 215, + 734, + 455, + 752 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we present the ranking correlation and search accuracy results of our Auto-DAS on CNN, NAS-Bench, and ViT search space. For fair comparisons, we employ the same search, distillation settings, KD accuracy ground truths in DisWOT as the ground truth for correlation validation. We report mean results based on more than 3 repeated trials.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Auto-DAS", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f029a75cf83afc897dd7eba7eb3d92b26312ec16d63401144655cd6fcc6e15fe.jpg", + "table_caption": [ + "Table 3: Accuracy $(\\%)$ of NDS-ResNet space on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
ConstraintsNWOTSynflowDisWOTAuto-DAS
5M-FLOPs63.1964.2865.9866.52
100M-FLOPs70.3872.1272.8973.25
0.5M-Param.70.3871.5872.8973.32
1M-Param.72.5773.5674.2374.66
MethodFitNetsATSPRKDCRD
Random NAS70.1270.1670.4671.1971.59
DisWOT74.8574.574.9574.6275.25
Auto-GAS75.2874.9875.5575.3175.86
", + "bbox": [ + 230, + 184, + 741, + 243 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7a020f43c9e0a26fbe85647328ccdb77be33fdc5a972c75c47b37d0ae6b19659.jpg", + "table_caption": [ + "Table 4: Results $(\\%)$ of other KD methods under 1M parameters.", + "Table 5: Top-1 accuracy (%) results on ImageNet." + ], + "table_footnote": [], + "table_body": "
ModelsBaselineBaseline+WSLDBaseline+DISTRandomNAS+KDZen-NAS+KDDisWOTAuto-DAS
ResNet1869.772.0472.0771.6871.8872.0872.58
ResNet5077.1NA78.4577.4877.8078.6279.25
", + "bbox": [ + 222, + 289, + 779, + 334 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experiments on CNN models", + "text_level": 1, + "bbox": [ + 215, + 369, + 504, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Architecture space and methodology. For CNN validation, we focus on ResNet variants, encompassing both CIFar-ResNet and NDS-ResNet domains. Our CIFar-ResNet space, inspired by He et al. [15], comprises three phases with 1,3,5,7 block options, tailored for CIFAR-100. We also explore the NDS-ResNet realm for both CIFAR-100 and ImageNet-1k search trials. Across all models, we employ our discovered $\\rho_{CNN}$ without additional proxy exploration. During knowledge transfer, all identified student networks undergo training using CRD protocols [48], with ResNet56 serving as the mentor model. For ImageNet-1k, we investigate ResNet18/50-caliber students under equivalent parameter restrictions.", + "bbox": [ + 212, + 410, + 785, + 559 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Findings on compact datasets. Our ranking analysis is rooted in individual knowledge transfer outcomes within the cifar-ResNet domain, as provided by DisWOT. Table 2 demonstrates that Auto-DAS achieves a notably higher Spearman coefficient compared to Fisher, SNIP, FLOPs, and NWOT. Our exploration experiments, conducted in the NDS-ResNet realm, pit our approach against Synflow, NWOT, and DisWOT, under 0.5M and 1M parameter limits, as well as 50M and 100M FLOPs thresholds. As illustrated in Table 3, Auto-DAS surpasses previous state-of-the-art methods by $0.9\\% \\sim 2.0\\% \\uparrow$ under identical constraints. Furthermore, we evaluate Auto-DAS across various KD techniques (refer to Table 4). The results affirm our method's efficacy, consistently outperforming DisWOT across different KD approaches.", + "bbox": [ + 212, + 565, + 787, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ImageNet performance analysis. Table 5 presents the effectiveness of student models identified by Auto-DAS on ImageNet, with ResNet34/101 acting as teacher networks. For the ResNet18 variant, Auto-DAS attains $72.58\\%$ accuracy, outperforming both DisWOT and other KD+NAS methodologies. Similarly, with the ResNet50 model, Auto-DAS achieves a leading accuracy of $79.25\\%$ . These findings underscore the superior capability of our Auto-DAS approach in enhancing model precision across various architectures.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Haosen Sun et al.", + "bbox": [ + 271, + 114, + 388, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/efb7b5ce7cb4c52748031166f76cef525f43ec9110097f3aa585b2a6600a9088.jpg", + "table_caption": [ + "Table 6: Evaluation of Architectures on CIFAR-10, CIFAR-100, and ImageNet-16 in NAS-Bench-201 [11]. Distilled Accuracy (\\%) represents the classification accuracy of the discovered architecture after distillation training. Search Time (s) denotes the computational cost (in GPU-seconds) during the search phase. The performances of NWOT and TE-NAS are obtained from their respective publications. Our proposed Auto-DASapproach demonstrates superior accuracy with rapid search speed." + ], + "table_footnote": [], + "table_body": "
TypeModelCIFAR-10CIFAR-100ImageNet-16-120
Dis. Acc(%)Time (s)Speed-upDis. Acc(%)Time (s)Speed-upDis. Acc(%)Time (s)Speed-up
Multi-trialRS93.63216K1.0×71.28460K1.0×44.881M1.0×
RL [3]92.83216K1.0×71.71460K1.0×44.351M1.0×
BOHB [13]93.49216K1.0×70.84460K1.0×44.331M1.0×
RSPS [25]91.6710K21.6×57.9946K21.6×36.87104K9.6×
One-shotGDAS [10]93.3922K12.0×70.7039K11.7×42.35130K7.7×
DARTS [35]89.2223K9.4×66.2480K5.8×43.18110K9.1×
Zero-shotNWOT [37]93.732.2K100×73.314.6K100×45.4310K100×
EZNAS [2]93.632.2K100×69.824.6K100×43.4710K100×
Zero-shot DASDisWOT93.551.2K180×74.219.2K180×47.3020K180×
DisWOT(Mr)93.490.72K300×73.6218.4K300×45.6340K300×
Auto-DAS93.961.8K180×74.7313.8K180×47.5030K180×
", + "bbox": [ + 220, + 239, + 782, + 387 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Experiments on NAS-Bench", + "text_level": 1, + "bbox": [ + 215, + 417, + 495, + 431 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experimental setup and search space. The NAS-Bench-101 [53] and NAS-Bench-201 [11] datasets serve as crucial benchmarks for assessing the efficacy of various NAS algorithms. DisWOT offers knowledge distillation outcomes for architectures within the NAS-Bench-101/201 search spaces (referred to as NB-101-KD and NB-201-KD). Adhering to these benchmark protocols, we apply the identified $\\rho_{CNN}$ to model search and validate distillation rankings without additional proxy searches. Comprehensive implementation details are available in the Appendix.", + "bbox": [ + 212, + 441, + 787, + 563 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Performance analysis. Table 2 illustrates that our Auto-DAS method achieves superior ranking performance, significantly outperforming DisWOT. Furthermore, we evaluate the standard correlation in NAS-Bench-101/201, with the ranking outcomes highlighting Auto-DAS's versatility through substantial improvements over existing techniques. Regarding search trials presented in Table 6, Auto-KD demonstrates remarkable gains in the accuracy-efficiency tradeoff compared to multi-trial NAS approaches (such as Random Search (RS) and Reinforcement Learning (RL)) and Gradient-based One-shot NAS methods (like GDAS and DARTS), which typically incur substantial search costs. Our method also consistently enhances distillation accuracy when compared to other Zero-shot NAS techniques (including NWOT, TE-NAS, and DisWOT). These exceptional ranking and search results underscore the robust generalization capabilities of our approach across challenging NAS-Bench datasets.", + "bbox": [ + 212, + 564, + 787, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Experiments on Vision Transformer", + "text_level": 1, + "bbox": [ + 215, + 782, + 558, + 799 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Architecture space and methodology. To assess the ranking consistency of zero-cost proxies for ViT structures, we develop a benchmark called ViT-Bench.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Auto-DAS", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/69f58a1f8143f32790026ed48514caf5e5993ddda657533a833ac46e9bd118cd.jpg", + "table_caption": [ + "Table 7: Results (%) of Ranking Correlation on CIFAR-100, Flowers, and Chaoyang." + ], + "table_footnote": [], + "table_body": "
Search SpaceProxyCIFAR-100FlowersChaoyang
KendallSpearmanPearsonKendallSpearmanPearsonKendallSpearmanPearson
PiTGraSP-42.02-58.71-31.53-50.66-69.64-40.90-16.00-22.94-19.07
SynFlow69.7987.0570.8062.2279.9871.6630.9642.6639.24
TENAS-2.13-3.21-1.68-2.86-4.23-3.33-3.34-5.04-3.55
NWOT-2.61-4.13-0.522.673.690.714.736.874.43
TF-TAS63.8382.2058.3764.4882.9167.2337.9952.9242.68
DisWOT-19.05-28.83-28.56-19.57-30.33-18.01-3.362.4015.22
Auto-DAS72.81±1.6890.32±1.7977.76±2.5876.42±1.1293.09±1.8587.63±1.3648.36±2.4666.13±1.5866.58±1.86
", + "bbox": [ + 218, + 171, + 784, + 258 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ab5640ead45ebb45bb8dee5145a82ded86849ddb8f55b116bea874e5c2947b80.jpg", + "table_caption": [ + "Table 8: Comparing the performance of our proposed Auto-DAS approach across three diverse datasets - CIFAR-100, flowers, and Chaoyang - using the PiT search space, we have achieved competitive results while incurring the lowest search cost, measured in time cost on a single GPU." + ], + "table_footnote": [], + "table_body": "
Search SpaceProxyCIFAR-100FlowersChaoyang
Param(M)Dis.Acc(%)Search CostParam(M)Dis.Acc(%)Search CostParam(M)Dis.Acc(%)Search Cost
PiTRandom5.3375.84N/A4.8865.30N/A5.2482.94N/A
GraSP4.5376.031.24 h3.7266.581.85 h4.6383.870.86 h
SynFlow11.0577.131.08 h5.2368.120.99 h4.9383.730.70 h
TENAS6.9376.095.14 h4.2668.035.14 h6.7683.645.07 h
NWOT5.2176.643.02 h10.7767.723.09 h6.3783.313.08 h
TF-TAS16.0777.061.21 h10.3068.210.95 h4.3284.340.71 h
DisWOT10.3875.820.88 h8.7667.580.88 h5.8583.420.88 h
Auto-DAS12.4277.881.28 h10.5269.321.28 h4.8884.831.28 h
Auto-DAS (G-free)11.3577.720.82h10.8869.450.82h5.2584.980.82h
", + "bbox": [ + 220, + 351, + 785, + 449 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This benchmark provides ground-truth accuracy for ViTs on compact datasets (CIFAR-100 [22], Flowers [39], and Chaoyang [58]). Inspired by [24], which shows ViTs gain significantly on small datasets when distilled from an efficient CNN teacher, we concentrate on the distillation accuracy of student ViTs on these datasets with a predetermined teacher. We randomly sample ViTs from PiT [16] search spaces and train them individually using the distillation settings from [24]. ResNet56 serves as the CNN teacher. ViT-Bench encompasses 5k KD accuracy ground truths per dataset. We divide the benchmark into validation and test sets at a 4:3 ratio. For the proxy search phase, we configure EA parameters $(\\mathcal{P}, \\mathcal{T}, r, k)$ in Alg. 1 as $(20, 100, 0.9, 5)$ and evaluate proxy candidates' ranking performance in the validation set each iteration. Post-search, we equitably compare the Auto-DAS proxy with conventional ones by random sampling in the test set. Our validation and test sets are distinct to ensure unbiased proxy comparisons.", + "bbox": [ + 215, + 487, + 785, + 683 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Findings on compact datasets. Table 7 presents the rank consistency results of existing proxies across three datasets. Our proposed Auto-DAS surpasses other notable proxies in metrics like Kendall's tau [1], Spearman's rho [44], and Pearson's correlation coefficient [4], showcasing its efficacy and superiority in achieving consistent and precise ViT architecture search. To further demonstrate Auto-DAS's effectiveness, Table 8 displays search trial results. Our approach achieves superior distillation outcomes and faster search speeds across all three datasets. These empirical findings highlight the importance of effective zero-cost proxies and the role of our proposed Auto-DAS in optimizing ViT architecture search performance.", + "bbox": [ + 215, + 688, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Haosen Sun et al.", + "bbox": [ + 271, + 114, + 388, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3b25f953580cd6df6711f0e52598a96b2b5953ebaabf22cd4e858659b5db14f9.jpg", + "table_caption": [ + "Table 9: Results of different ViT models on ImageNet." + ], + "table_footnote": [], + "table_body": "
ModelMethodParam.Top-1 Acc.ModelMethodParam.Top-1 Acc.
DeiT-TinyBaseline+KD5.7M74.5%Swin-TinyBaseline+KD29.0M81.7%
TF-TAS+KD5.7M74.7%TF-TAS+KD29.2M82.0%
DisWOT5.8M75.3%DisWOT29.0M82.2%
Auto-DAS6.1M75.9%Auto-DAS29.2M82.9%
PiT-TinyBaseline+KD4.9M73.2%PiT-SmallBaseline+KD23.5M79.9%
TF-TAS+KD4.9M73.5%TF-TAS+KD24.2M80.1%
DisWOT4.6M73.6%DisWOT23.8M80.3%
Auto-DAS4.8M74.2%Auto-DAS24.5M81.0%
", + "bbox": [ + 269, + 171, + 733, + 287 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8f2e562ae5d47ef094a7a364bbe3a90c16f8fbba00923adf7879b2ce9cd0eecf.jpg", + "table_caption": [ + "Table 10: Spearman correlation (\\%) on optimal proxy in our search space with vanilla inputs or operations." + ], + "table_footnote": [], + "table_body": "
MethodsSettingsResNetNB-101NB-201NB-101-KDNB-201-KDPiT-CIFAR-100PiT-FlowersPiT-Chaoyang
Auto-DASAuto-DAS 98.4471.8487.7385.9574.6690.3293.0966.13
Auto-DAS (instinct)Activations70.2865.8978.9666.3560.8574.8676.8552.86
Gradients77.5372.6886.9273.5566.3880.5682.7556.67
weights60.3552.7570.2558.6755.2560.8965.8241.25
Unary78.8574.2588.2577.4266.5382.8384.3858.69
Binary62.5258.8772.8964.3558.9665.4570.2548.46
All80.5275.8689.6278.5668.5583.2586.5260.12
Auto-DAS (interaction)Feature90.3547.9662.6876.3366.3881.9585.3959.68
Embed88.6546.5560.7574.8563.2880.2583.2756.88
Logits85.4742.5855.8670.5358.6276.8378.9652.45
Transform92.3548.8864.6678.9568.6587.2576.5561.35
Distance89.6346.5363.8575.8265.8682.6984.5658.75
All92.8250.8865.8880.5670.4885.4288.7363.38
", + "bbox": [ + 220, + 362, + 779, + 507 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Computational cost analysis. (1) Like most train-free NAS methods, Auto-DAS involves minor additional gradient computations compared to DisWOT. To address this, we explore a gradient-free version of Auto-DAS in Table 8 (labeled Auto-DAS (G-free)). Results indicate that our gradient-free Auto-DAS achieves better accuracy-efficiency trade-offs than DisWOT. (2) While DisWOT is efficient and yields good results on certain datasets, it lacks generalizability across various models and tasks. In contrast, Auto-DAS is automated, versatile, and effective, better suited diverse application needs. Moreover, given the low cost of training-free approaches, the small additional time investment in searching is justified by the consistent performance gains achieved.", + "bbox": [ + 212, + 556, + 787, + 709 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ImageNet performance analysis. Examining the results in Table 9, it's clear that Auto-DAS outperforms other methods across various models. For DeiT-Tiny, Auto-DAS achieves $75.9\\%$ accuracy, surpassing Baseline+KD $(74.5\\%)$ , TF-TAS+KD $(74.7\\%)$ , and DisWOT $(75.3\\%)$ . Similarly, Auto-DAS excels with the Swin-Tiny model. This trend persists for PiT-Tiny and PiT-Small models, where Auto-DAS consistently outperforms other methods. These findings underscore the effectiveness and adaptability of our Auto-DAS method in enhancing accuracy across different ViT architectures.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Auto-DAS", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9736a737a53a422dcaa4f5de4cf7ff6a38e21b12d826d6d6a6b047362536a128.jpg", + "image_caption": [ + "Figure 6: Correlation visualization for Auto-DAS on NAS-Bench-101, NAS-Bench-201 and ViT-Bench (from left to right)." + ], + "image_footnote": [], + "bbox": [ + 222, + 142, + 408, + 251 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/82aa422130857dbfa25b0857318cda29db1312bdd47caca9495564c368fe004a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 142, + 593, + 250 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1da0ebc9194a28aff3c1d5b979a994f9d3d224093998cd2d6713dfebf1dd147c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 142, + 781, + 251 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 320, + 401, + 334 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proxy search space. We conduct experiments in Table 10 to analyze different components in our proxy search space. The results show that Auto-DAS (interaction) outperforms Auto-DAS (instinct) overall. The feature input and transform operations are particularly important for Auto-DAS (interaction), while weight input and binary operations contribute weakly. On the other hand, gradient input and unary operations show promising results for Auto-DAS (instinct). Different ops have varying impacts on the proxies' accuracy, and combining them enhances the performance of Auto-DAS approaches. These findings highlight the significance of considering different ops and their interactions in the search space for optimal performance.", + "bbox": [ + 212, + 345, + 785, + 496 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Search algorithm. We use the EA with adaptive-elite selection for proxy search. As shown in Figure 4, the EA adaptive-elite selection obtains better final search results than EA and random search in the proxy search process.", + "bbox": [ + 212, + 496, + 785, + 542 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Correlation visualization.. To intuitively observe the proxy's predictability, we visualize the score of Auto-DAS and the ground-truth performance in Figure 6. These visualizations demonstrate that Auto-DAS effectively detects the true distillation results.", + "bbox": [ + 212, + 542, + 785, + 602 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 626, + 359, + 642 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present Auto-DAS, a novel training-free framework for distillation-aware student architecture search. It includes proxy search and training-free student search. With discovered proxies, Auto-DAS enables efficient search without training costs. This significantly outperforms handcrafted proxies in distillation performance. Comprehensive results on NAS benchmarks, ViT, and CNN spaces over multiple datasets illustrate our method's superior ranking and distillation abilities. We hope our novel investigations will give more insight and new directions for the knowledge distillation and NAS research communities.", + "bbox": [ + 212, + 657, + 785, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Limitation.** We mainly evaluate the same benchmarks as other NAS methods for fair comparison. We will continue to expand Auto-DAS for more tasks in future work. **Social impact.** Our Auto-DAS focuses on technology improvement without social and ethical implications.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Haosen Sun et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdi, H.: The Kendall rank correlation coefficient. Encyclopedia of measurement and statistics 2, 508-510 (2007)", + "2. Akhauri, Y., Munoz, J.P., Jain, N., Iyer, R.: EZNAS: Evolving zero-cost proxies for neural architecture scoring. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) NeurIPS (2022), https://openreview.net/forum?id=1SqaDG4dvdt", + "3. Baker, B., Gupta, O., Naik, N., Raskar, R.: Designing neural network architectures using reinforcement learning. In: ICLR (2017)", + "4. Bowley, A.: The standard deviation of the correlation coefficient. Journal of the American Statistical Association 23(161), 31-34 (1928)", + "5. Brown, T.B., Mann, B., Subbiah, N.R.M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., M., D., Ziegler, Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. arXiv preprint, arXiv:2005.14165 (2020)", + "6. Dong, P., Li, L., Tang, Z., Liu, X., Pan, X., Wang, Q., Chu, X.: Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. In: ICML (2024)", + "7. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023)", + "8. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV. pp. 17076-17086 (2023)", + "9. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022)", + "0. Dong, X., Yang, Y.: Searching for a robust neural architecture in fourgpu hours. CVPR (2019)", + "1. Dong, X., Yang, Y.: Nas-bench-201: Extending the scope of reproducible neural architecture search. In: ICLR (2019)", + "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)", + "3. Falkner, S., Klein, A., Hutter, F.: Bohb: Robust and efficient hyperparameter optimization at scale. In: ICML (2018)", + "4. Gu, J., Tresp, V.: Search for better students to learn distilled knowledge. arXiv preprint arXiv:2001.11612 (2020)", + "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "6. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021)", + "7. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "8. Hu, S., Xie, S., Zheng, H., Liu, C., Shi, J., Liu, X., Lin, D.: Dsnas: Direct neural architecture search without parameter retraining. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "9. Hu, Y., Liang, Y., Guo, Z., Wan, R., Zhang, X., Wei, Y., Gu, Q., Sun, J.: Angle-based search space shrinking for neural architecture search. In: ECCV (2020)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Auto-DAS", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021)", + "21. Kim, Y., Rush, A.M.: Sequence-level knowledge distillation. In: EMNLP (2016)", + "22. Krizhevsky, A.: Learning multiple layers of features from tiny images (2009)", + "23. Lee, N., Ajanthan, T., Torr, P.: Snip: Single-shot network pruning based on connection sensitivity. In: ICLR (2018)", + "24. Li, K., Yu, R., Wang, Z., Yuan, L., Song, G., Chen, J.: Locality guidance for improving vision transformers on tiny datasets. In: European Conference on Computer Vision. pp. 110-127. Springer (2022)", + "25. Li, L., Talwalkar, A.S.: Random search and reproducibility for neural architecture search. ArXiv (2019)", + "26. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022)", + "27. Li, L., Bao, Y., Dong, P., Yang, C., Li, A., Luo, W., Liu, Q., Xue, W., Guo, Y.: Detkds: Knowledge distillation search for object detectors. In: ICML (2024)", + "28. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024)", + "29. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023)", + "30. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022)", + "31. Li, L., Sun, H., Li, S., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Auto-gas: Automated proxy discovery for training-free generative architecture search. In: ECCV (2024)", + "32. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation. In: ICLR (2020)", + "33. Li, L., Wei, Z., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Attnzero: Efficient attention discovery for vision transformers. In: ECCV (2024)", + "34. Lin, M., Wang, P., Sun, Z., Chen, H., Sun, X., Qian, Q., Li, H., Jin, R.: Zen-nas: A zero-shot nas for high-performance image recognition. ICCV (2021)", + "35. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019)", + "36. Liu, Y., Jia, X., Tan, M., Vemulapalli, R., Zhu, Y., Green, B., Wang, X.: Search to distill: Pearls are everywhere but not the eyes. In: CVPR (2020)", + "37. Mellor, J., Turner, J., Storkey, A., Crowley, E.J.: Neural architecture search without training. In: ICML (2021)", + "38. Mirzadeh, S.I., Farajtabar, M., Li, A., Levine, N., Matsukawa, A., Ghasemzadeh, H.: Improved knowledge distillation via teacher assistant. In: AAAI (2020)", + "39. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008)", + "40. Park, W., Lu, Y., Cho, M., Kim, D.: Relational knowledge distillation. In: CVPR (2019)", + "41. Pham, H., Guan, M.Y., Zoph, B., Le, Q.V., Dean, J.: Efficient neural architecture search via parameter sharing. In: ICML. pp. 4092-4101 (2018)", + "42. Real, E., Liang, C., So, D.R., Le, Q.V.: Automl-zero: Evolving machine learning algorithms from scratch (2020)", + "43. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Haosen Sun et al.", + "bbox": [ + 271, + 114, + 388, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Stephanou, M., Varughese, M.: Sequential estimation of spearman rank correlation using hermite series estimators. Journal of Multivariate Analysis 186, 104783 (2021)", + "45. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS (2020)", + "46. Theis, L., Korshunova, I., Tejani, A., Huszár, F.: Faster gaze prediction with dense networks and fisher pruning. ArXiv abs/1801.05787 (2018)", + "47. Tian, Y., Krishnan, D., Isola, P.: Contrastive representation distillation. In: ICLR (2020)", + "48. Tian, Y., Krishnan, D., Isola, P.: Contrastive representation distillation. In: ICLR (2020)", + "49. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers amp; distillation through attention. In: ICML (2021)", + "50. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020)", + "51. Wang, T., Yuan, L., Zhang, X., Feng, J.: Distilling object detectors with fine-grained feature imitation. In: CVPR (2019)", + "52. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching (2022)", + "53. Ying, C., Klein, A., Christiansen, E., Real, E., Murphy, K., Hutter, F.: Nas-bench-101: Towards reproducible neural architecture search. In: ICML (2019)", + "54. You, S., Huang, T., Yang, M., Wang, F., Qian, C., Zhang, C.: Greedynas: Towards fast one-shot nas with greedy supernet. In: CVPR (2020)", + "55. Zhang, L., Ma, K.: Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In: ICLR (2020)", + "56. Zhou, H., Song, L., Chen, J., Zhou, Y., Wang, G., Yuan, J., Zhang, Q.: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective (2021)", + "57. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024)", + "58. Zhu, C., Chen, W., Peng, T., Wang, Y., Jin, M.: Hard sample aware noise robust learning for histopathology image classification. TMI (2021)", + "59. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024)", + "60. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 659 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Auto-DAS", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_model.json b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5c374a42624e611584df13328ad4c03b3d99c60b --- /dev/null +++ b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_model.json @@ -0,0 +1,2324 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.141, + 0.756, + 0.207 + ], + "angle": 0, + "content": "Auto-DAS: Automated Proxy Discovery for Training-free Distillation-aware Architecture Search" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.234, + 0.765, + 0.252 + ], + "angle": 0, + "content": "Haosen Sun\\(^{1}\\), Lujun Li\\(^{1\\boxtimes}\\), Peijie Dong\\(^{2}\\), Zimian Wei\\(^{3}\\), and Shitong Shao\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.262, + 0.734, + 0.291 + ], + "angle": 0, + "content": "\\(^{1}\\) The Hong Kong University of Science and Technology \\(^{2}\\) The Hong Kong University of Science and Technology (Guangzhou)" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.291, + 0.65, + 0.305 + ], + "angle": 0, + "content": "\\(^{3}\\) National University of Defense Technology" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.305, + 0.796, + 0.345 + ], + "angle": 0, + "content": "\\(^{4}\\) Southeast University \nhsunas@connect.ust.hk;lilujunai@gmail.com;pdong212@connect.hkust-gz.edu.cn; weizimian16@nudt.edu.cn; shaoshitong@seu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.378, + 0.74, + 0.793 + ], + "angle": 0, + "content": "Abstract Distillation-aware Architecture Search (DAS) seeks to discover the ideal student architecture that delivers superior performance by distilling knowledge from a given teacher model. Previous DAS methods involve time-consuming training-based search processes. Recently, the training-free DAS method (i.e., DisWOT) proposes KD-based proxies and achieves significant search acceleration. However, we observe that DisWOT suffers from limitations such as the need for manual design and poor generalization to diverse architectures, such as the Vision Transformer (ViT). To address these issues, we present Auto-DAS, an automatic proxy discovery framework using an Evolutionary Algorithm (EA) for training-free DAS. Specifically, we empirically find that proxies conditioned on student instinct statistics and teacher-student interaction statistics can effectively predict distillation accuracy. Then, we represent the proxy with computation graphs and construct the proxy search space using instinct and interaction statistics as inputs. To identify promising proxies, our search space incorporates various types of basic transformations and network distance operators inspired by previous proxy and KD-loss designs. Next, our EA initializes populations, evaluates, performs crossover and mutation operations, and selects the best correlation candidate with distillation accuracy. We introduce an adaptive-elite selection strategy to enhance search efficiency and strive for a balance between exploitation and exploration. Finally, we conduct training-free DAS with discovered proxy before the optimal student distillation phase. In this way, our auto-discovery framework eliminates the need for manual design and tuning, while also adapting to different search spaces through direct correlation optimization. Extensive experiments demonstrate that Auto-DAS generalizes well to various architectures and search spaces (e.g., ResNet, ViT, NAS-Bench-101, and NAS-Bench-201), achieving state-of-the-art results in both ranking correlation and final searched accuracy. Code at: https://github.com/lliai/Auto-DAS." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.806, + 0.721, + 0.82 + ], + "angle": 0, + "content": "Keywords: Knowledge Distillation \\(\\cdot\\) Architecture Search \\(\\cdot\\) AutoML" + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.826, + 0.641, + 0.84 + ], + "angle": 0, + "content": "Corresponding author & contributed equally to this work." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.127 + ], + "angle": 0, + "content": "Haosen Sun et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.341 + ], + "angle": 0, + "content": "Knowledge Distillation (KD) transfers logits [17, 52], features [26, 32], or relation [40, 47] knowledge of heavy teacher models as additional supervised signals to augment lightweight student models. KD originally served as model compression technique for large models and has recently been used as a common strategy for performance improvement in different scenarios. For example, distilling a CNN teacher for a Vision Transformer (ViT) [49] can significantly enhance its performance, particularly when training from scratch in data-limited scenarios. However, the architecture gap between the teacher and student network creates a mismatch between the receptive field and the prediction reliability [38, 56]. These disparities contribute to distillation gains, i.e., students with a smaller architecture gap with teachers will enjoy better final results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.789, + 0.75 + ], + "angle": 0, + "content": "To alleviate these problems, a new task, Distillation-aware Architecture Search (DAS), is presented to search for an optimal student architecture from the given teacher, which obtains the best performance with the help of neural architecture search. In contrast to the traditional NAS methods, the DAS approach needs to consider the effect of the teacher and employ distillation accuracy as the search evaluation metric. Some observations [7,36] indicate that the models searched by traditional NAS methods [9,20,35] are suboptimal in distillation, and there is a common performance ranking gap for obtained students between vanilla training and distillation. For the DAS task, some training-based methods leverage Reinforcement Learning (RL) [36] or gradient method [14] as search strategies to find the ideal student. However, these methods bring in high search costs (e.g., 5 days with 200 TPUv2 in AKD [36]), which causes heavy burdens for application. Recently, DisWOT [7] first presents the training-free DAS framework with some KD-based proxies, exploiting the connection between training-free proxies and knowledge distillation loss. It achieves significant search acceleration by using statistics between random networks for predicting distillation accuracy without candidate training. However, some drawbacks still limit its extensive application, which could highlight from two aspects: (1) Expert design and tuning costs. These KD-based proxies are inspired by some distillation studies with extensive expert intuition and require extensive tuning. It is hard to state if these proxies are optimal for DAS as limited by the designer's experience. (2) Generalization issue. Handcrafted ones only perform well on some specific settings but are hard to adapt to different scenarios without changing/modification. As shown in Figure 1, existing handcrafted proxies perform well on the CNNs but do not generalize well to ViT models. Thus, two questions arise: (1) How to efficiently discover the proxies without expert knowledge? (2) How to find the optimal proxies for different search spaces?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "For the first problem, we observe that previous proxy expressions are usually constructed for typical inputs and common primitive operations. Inspired by Auto-Zero [42], we explore a from-scratch proxy discover framework alternative to handcrafted for DAS task. For the second problem, we can evolve various proxy candidates by directly taking their predictability for distillation accuracy as feedback. In addition, we find that proxies with inputs from student" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Auto-DAS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.506, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.25, + 0.498, + 0.278 + ], + "angle": 0, + "content": "Figure 1: Spearman correlation \\((\\%)\\) on CNN models(left) and ViT models(right)." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.143, + 0.768, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.25, + 0.765, + 0.264 + ], + "angle": 0, + "content": "Figure 2: Illustration of Auto-DAS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.788, + 0.643 + ], + "angle": 0, + "content": "distinct statistics and teacher-student interaction statistics consistently achieve better results. For example, DisWOT+Synflow in Figure 1 achieves stable improvements on multiple search spaces. Based on these observations, we present Auto-DAS, a from-scratch framework for discovering proxies for DAS tasks are proposed as an alternative to manually designed ones. As shown in Figure 2, we first build the proxy search space, consisting of both instinct and interaction proxy. The instinct proxy is used with student's weights and gradient statistics as input and different unary and binary mathematical operations as candidate components. For interaction proxy, we select teacher-student features, embeddings, logits as inputs, different basic transformation operations, and distance functions to form the computation graph. Then, we randomly sample various proxies as the population and then perform mutation and crossover operations according to the ranking ability of different proxies. To improve generality, we directly treat the Kendall correlation between the prediction of the candidate proxy and the ground truth under distillation as the fitting objective of the evolutionary algorithm. To speed up the process, we employ an adaptive-elite selection strategy during the evolution search. With the automatic search framework, our Auto-DAS surpasses existing training-free NAS approaches by a large margin without involving too much prior knowledge. Finally, after getting the Auto-DAS proxy in a given search space, we search for the best student architectures with the highest Auto-DAS score and apply the searched students to distill the knowledge from the given teacher." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In comparison to manual DAS techniques, our Auto-DASframework offers several advantages: (1) Autonomy. It facilitates the spontaneous discovery of more potent and efficient proxies that may elude human experts, thus mitigating bias and ensuring that the resultant architectures are tailored for the intended problem or dataset. (2) Versatility. The search mechanism of AutoDASenhances adaptability by uncovering proxies that perform well across diverse search spaces. Our framework can evolve flexible proxies to generalize across architectures, based on the ultimate distillation outcomes for the search objectives. (3) Enlightening. Our approach enables the automatic identification of more expressive, efficient, and effective proxies for optimizing and customizing student architectures, paving the way for novel research and applications in DAS and KD tasks. (4) Efficacy. We have conducted comprehensive experiments on ResNet, NAS-Bench-101, NAS-Bench-201, and ViT-Bench search spaces to confirm the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "Haosen Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "superiority of our Auto-DAS. The results demonstrate that our Auto-DAScan surpass other zero-shot proxies in distillation accuracy and achieve state-of-the-art results in rank consistency, such as a \\(40\\%\\) improvement in Kendall correlation over DisWOT. To summarize, our key contributions to Auto-DAS are:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.222, + 0.784, + 0.265 + ], + "angle": 0, + "content": "- We introduce the pioneering framework for discovering proxies from scratch, Auto-DAS, to address handcrafted limitations and enhance generalizability in predicting distillation accuracy across various students." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.268, + 0.785, + 0.326 + ], + "angle": 0, + "content": "- We devise a well-structured proxy search space, incorporating instinct and interaction statistics as inputs, along with diverse unary, binary, transforms, and distance functions as operations. We develop evolutionary algorithms with adaptive-elite selection for effective and efficient proxy search." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.329, + 0.784, + 0.372 + ], + "angle": 0, + "content": "- We provide empirical evidence that Auto-DASachieves cutting-edge performance across ResNet, NAS-Bench-101, NAS-Bench-201, ViT-Bench search spaces and multiple datasets." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.222, + 0.785, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.399, + 0.383, + 0.415 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Distillation-aware architecture search. The increasing demand for practical applications of large-scale models [5, 12] has led to impressive advancements in Knowledge distillation (KD) across various domains [26-30, 32, 43], including image classification [17], object detection [51, 55], and NLP [21]. However, discrepancies between teacher and student architectures can substantially impact distillation outcomes. Consequently, distillation-aware student architecture search for a given teacher model has emerged as a crucial new area of research. Traditional reinforcement learning-based search methods [36] utilize KD-loss as feedback for DAS but incur significant additional costs during the search process. Subsequent research employs efficient gradient-based methods [14], albeit introducing new optimization challenges. Recently, DisWOT [7] introduced the first training-free DAS framework. However, DisWOT relies on a manually designed proxy, which lacks generalization across different search spaces. To address this limitation, we propose the first automated proxy discovery for training-free DAS tasks. Our method not only retains the advantages of DisWOT, such as efficiency and flexibility, but also automatically optimizes proxies without expert intervention. Auto-DAS paves the way for novel DAS research and offers valuable insights based on search result analysis for the research community." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Training-free architecture search. Conventional NAS [18, 60] involves designing search spaces, algorithms, and evaluations to automatically discover optimal architectures under given constraints. Training-based methods employ train-then-search or weight sharing processes [19, 41, 54, 60]. Training-free NAS, also known as zero-shot NAS [6, 8, 31, 33, 34, 57, 59], enables faster search by predicting performance without training. Zero-cost proxies are categorized as parameter-level or architecture-level. Parameter-level approaches rely on pruning and saliency values [23, 45, 50], while architecture-level methods evaluate expressiveness [34] or linear map correlations [37]. DisWOT proposed KD-based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Auto-DAS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.144, + 0.455, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.143, + 0.758, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.351, + 0.785, + 0.407 + ], + "angle": 0, + "content": "Figure 3: The overall illustration of Auto-DAS. Our search space includes unary, binary, transform, and distance options with different inputs. Then, we construct our intrinsic & interaction proxies and use crossover & mutation to search for the best proxy." + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.422, + 0.763, + 0.436 + ], + "angle": 0, + "content": "Table 1: Operations of Auto-DAS. More details are available in the Appendix." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.449, + 0.779, + 0.531 + ], + "angle": 0, + "content": "
ProxyInputTypePrimitive Operations
instinctActivationsUnaryexp, mish, leaky, relu, tanh, softmax, sigmoid, pow2, pow4, log, abslog, abssqrt
GradientsUnaryllnorm, revert, min - maxto - mean, to - std, invert, lnNorm, normalizedSum, no - op
WeightsBinarysum, subtract, multiply, divide, dot.
InteractionFeatureTransformtanh, log, sqrt, batchnorm, min - max, norm, softmax, exp, mish, leaky, relu
EmbedTransformmask, no, bmm, mm, scale, multi - scale, scale, local, batch, channel, drop, satt, natt, catt
LogitsDistancel1, l2, ℓKL, ℓhard, ℓCosine, ℓPearson
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.727 + ], + "angle": 0, + "content": "proxies, but their fixed formulations limit generalization. To overcome these limitations, we consider student internal statistics and teacher-student similarities as inputs and evolve superior proxies from scratch for training-free DAS tasks. Inspired by Auto-Zero, our innovations include a novel proxy search space and a specialized search process, promoting automatic proxy design for distillation. Our approach differs from EZNAS [2] in terms of tasks (ours is customized for CNNs and ViTs), search algorithms (ours employs adaptive-elite selection), and search space (ours incorporates teacher-student knowledge). Unlike EZNAS and Auto-KD, which focus solely on vanilla classification, we directly search on the distillation task and include teacher-student knowledge and additional options in the search space." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.381, + 0.766 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "This section delves into the intricacies of existing proxies, our proposed search space, the search algorithm, and an analysis of the results. Subsequently, we outline a training-free student search process utilizing the discovered zero-cost proxy. Figure 3 illustrates the workflow of our approach." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "Haosen Sun et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.524, + 0.163 + ], + "angle": 0, + "content": "3.1 Examination of existing proxies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.276 + ], + "angle": 0, + "content": "To provide context for our training-free method, we first examine current zero-cost proxies: SNIP [23] calculates a saliency metric at initialization using a single mini-batch of data to estimate the loss change when a specific parameter is eliminated. Synflow [45] introduces a modified version of synaptic saliency scores that prevents layer collapse during parameter pruning, while Fisher [46] computes the sum over all activation gradients in the network, applicable for channel pruning. Their detailed formulations are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.287, + 0.786, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\rho_ {s n i p} = \\left| \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W} \\right|, \\rho_ {s y n f l o w} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W}, \\rho_ {f i s h e r} = \\left(\\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {A}} \\mathcal {A}\\right) ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.401 + ], + "angle": 0, + "content": "where \\(\\mathcal{L},\\mathcal{W},\\mathcal{A}\\) are loss function, weight and activation. \\(H\\) is the Hessian matrix. These proxies consist of various unary, binary operations with statistics input from the candidate model, named instinct proxies in our paper. For the distillation scenario, DisWOT presents some KD-based proxies based on KD-loss design, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.412, + 0.786, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\rho_ {D i s W O T} = \\mathcal {D} _ {L 2} (\\mathcal {G} ([ \\mathcal {A} S, \\mathcal {A} T ])) + \\mathcal {D} _ {L 2} (\\mathcal {G} ([ \\mathcal {F} S, \\mathcal {F} T ])), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.542 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}\\), \\(\\mathcal{F}\\), \\(\\mathcal{E}\\), \\(\\mathcal{P}\\) are activation maps, feature maps, embeddings, and logits prediction. \\([\\cdot, \\cdot]\\) means to perform the same operation on two primitives. \\(\\mathcal{G}\\) is the gram-matrix. \\(\\mathcal{D}_{L2}\\), \\(\\mathcal{D}_{KL}\\) are \\(\\ell_2\\) and \\(\\ell_{KL}\\) distance. These KD-based proxies use teacher-student pair-wise statistics as input and different transformations and distance functions to construct formations that we denote as interaction proxies. Based on these analyses, we build search spaces for instinct and interactive proxies in the next part." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.561, + 0.542, + 0.577 + ], + "angle": 0, + "content": "3.2 Search Space for Proxy Discovery" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Our search space structure comprises three primary components: input choices, primitive operations, and proxy representation. As illustrated in Table 1, for input choices in constructing zero-cost proxies, the instinct proxy extracts activations, gradients, and weights from the student model, while the interaction proxy selects features, embeddings, and logits of the teacher-student network. These inputs represent the most informative options common across traditional and KD-based proxies. For instance, features provide insights into data distribution and internal network representations, while gradients highlight weight sensitivity to loss functions. Following this observation, we register each student layer's activation, weights, and corresponding gradients as potential inputs for our zero-cost proxy. Regarding primitive operations, we consider unary and binary operations for instinct proxies, and transform and distance operations for interaction proxies. We include mathematical, normalization, and scaling operations for unary operations; matrix sums and products for binary operations; and activation, normalization, scaling, and other transformations, along with typical distance functions for interaction proxies. Our proxy is represented as a computation graph, where input nodes are instinct and interaction statistics, and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Auto-DAS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.146, + 0.71, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Evolutionary Strategy for Auto-DAS Proxy Discovery" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.163, + 0.787, + 0.445 + ], + "angle": 0, + "content": "Input: Architecture space \\(S\\), candidate set \\(\\mathcal{P}\\), iteration limit \\(\\mathcal{T}\\), selection proportion \\(r\\), temporary set \\(\\mathcal{R}\\), elite count \\(k\\). \nOutput: Highest-scoring Auto-DAS configuration. \n1: Initialize candidate set \\(\\mathcal{P}0 := \\text{GenerateInitialCandidates}(P_i)\\); \n2: Empty temporary set \\(\\mathcal{R} := \\emptyset\\); \n3: for \\(i = 1,2,\\ldots,\\mathcal{T}\\) do \n4: Reset temporary set \\(\\mathcal{R} := \\emptyset\\); \n5: Populate \\(\\mathcal{R}\\) with random selections from \\(\\mathcal{P}\\); \n6: Elite group \\(G_i k := \\text{SelectTopPerformers}(\\mathcal{R}, k)\\); \n7: Choose parent \\(G_i^p := \\text{RandomlyPick}(G_i k)\\); \n8: Generate offspring \\(G_i^m := \\text{MUTATE}(G_i^p)\\); \n9: // Adaptive Elite Preservation \n10: Randomly generate baseline \\(E\\); \n11: if \\(\\rho(G_i^m) > \\rho(E)\\) then \n12: Include \\(G_i^m\\) in \\(P\\); \n13: else \n14: Include \\(E\\) in \\(P\\); \n15: end if \n16: Eliminate lowest-scoring proxy from population. \n17: end for" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.534 + ], + "angle": 0, + "content": "intermediate nodes are primitive operations. The graph's output is the proxy score used for distillation ranking. Our comprehensive search space encompasses most zero-cost proxies, KD-based losses, and model similarity formulations, constituting our core innovation and contribution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.556, + 0.553, + 0.571 + ], + "angle": 0, + "content": "3.3 Evolution Procedure and Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.785, + 0.699 + ], + "angle": 0, + "content": "Utilizing our proxy search space, we employ an Evolutionary Algorithm (EA) to identify optimal proxy expressions (refer to Figure 3 and Algorithm 1). Our EA initiates with a population of candidate proxies and iteratively evolves it over generations using genetic operators such as selection, crossover, and mutation to generate superior solutions. For objective fitting, each proxy is evaluated based on the ranking correlation between its output proxy \\(\\mathcal{Q}\\) scores and actual performances \\(\\mathcal{D}\\) to efficiently discover the optimal proxy \\(\\rho^{*}\\) from search space \\(\\mathcal{S}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.701, + 0.785, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\rho_ {A u t o - D A S} ^ {*} = \\arg \\max \\rho \\in \\mathcal {S} (\\tau (\\mathcal {D}, \\mathcal {Q})), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.795 + ], + "angle": 0, + "content": "where Kendall's Tau \\(\\tau\\) serves as the correlation coefficient. Each evolution generation selects the top-k candidates with the highest Auto-DAS scores and randomly chooses a parent from these candidates for mutation. During mutation, a node in the computation graph is randomly selected and mutated with newly generated primitive operations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Efficient proxy search with existing benchmark. Obtaining pre-trained results \\(\\mathcal{D}\\) can be resource-intensive. To address this, we leverage pre-computed architecture-performance pairs from existing DisWOT benchmarks [7]. These" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "Haosen Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.36, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.144, + 0.498, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.144, + 0.639, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.144, + 0.781, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.236, + 0.788, + 0.265 + ], + "angle": 0, + "content": "Figure 4: SearchFigure 5: Correlation visualization of NWOT (left), DisWOT curve on ResNet. (middle), Auto-DAS (right) on ResNet search space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.784, + 0.401 + ], + "angle": 0, + "content": "benchmarks provide distillation results for various ResNet and NAS-Bench-101/201 search space models. Utilizing this benchmark eliminates the need to retrain performance ground-truths and significantly reduces the proxy search process time to merely 0.05 GPU-day. Due to its excellent generality (demonstrated in our extensive experiments), our searched proxy \\(\\rho_{Auto - DAS}\\) transfers well to similar search spaces and can be directly applied in most scenarios without additional search budget." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.507 + ], + "angle": 0, + "content": "Adaptive-elite selection strategy. To mitigate population deterioration and premature convergence, we introduce an adaptive-elite selection strategy. This approach involves comparing the performance of mutation-generated offspring with that of a randomly generated individual. The individual with a higher Auto-DAS score is then added to the population. By maintaining or enhancing the overall population performance over time, this strategy accelerates convergence towards a high-performing solution (see Figure 4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.532, + 0.475, + 0.547 + ], + "angle": 0, + "content": "3.4 Searched Zero-cost Proxy" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.559, + 0.788, + 0.59 + ], + "angle": 0, + "content": "We present formulas of the searched proxies on \\(\\rho_{CNN}\\) for CNN Search space (e.g., ResNet, NAS-Bench) and \\(\\rho_{ViT}\\) ViT search space as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.288, + 0.601, + 0.787, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\rho_ {C N N} = \\mathcal {D} _ {K L} (\\operatorname {s i g m o i d} ([ \\mathcal {F} _ {S}, \\mathcal {F} _ {T} ])) + \\frac {1}{M} \\sum \\left| \\log \\operatorname {s o f t m a x} \\left(\\frac {\\partial L}{\\partial W}\\right) \\right| _ {F}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.639, + 0.786, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\rho_ {V I T} = \\mathcal {D} _ {K L} (\\operatorname {m a s k} ([ \\mathcal {F} _ {S}, \\mathcal {F} _ {T} ])) + \\frac {M}{\\sum \\operatorname {s i g m o i d} \\left( \\right.\\left| \\right. \\log \\left| \\frac {\\partial L}{\\partial W} \\right|\\left. \\right)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.784, + 0.703 + ], + "angle": 0, + "content": "Where \\(W\\) is the weight parameter, \\(\\partial \\mathcal{L} / \\partial W\\) is the corresponding gradient. \\(\\| \\cdot \\| _F\\) means the Frobenius-norm. \\(M\\) denotes mean operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.784, + 0.779 + ], + "angle": 0, + "content": "Analysis of searched proxy. The discovered formulations indicate that (1) mathematical operations like softmax, logsoftmax, sigmoid, and mask operations benefit proxy predictability. (2) Gradient and features are vital input statistics for instinct and interaction proxies. As shown in Figure 5, the discovered proxies of Auto-DAS achieve significant ranking improvement." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Understanding why our Auto-DAS surpasses other NAS methods. (1) Many training-based distillation NAS methods use weight sharing and proxies, often leading to inaccurate performance estimates. Our Auto-DAS searches proxies based on real performance, yielding one with strong ranking ability. (2)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Auto-DAS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Spearman correlation (\\%) results on the NAS-Bench-101 and NAS-Bench-201 (NB-101/201) datasets. In this context, NB-101/201-KD denotes the task of distilling the accuracies of the architectures on the respective NAS-Bench-101/201 datasets." + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.199, + 0.689, + 0.352 + ], + "angle": 0, + "content": "
MethodResNetNB-101NB-201NB-101-KDNB-201-KD
FLOPs72.9230.8163.3815.5664.55
Fisher81.37-38.8135.91-33.924.45
Grad_Norm82.35-39.2358.70-39.16-10.01
SNIP85.07-29.0158.17-21.7816.91
Synflow88.3043.6974.6120.3674.63
NWOT45.6632.8464.4122.9735.27
DisWOT(Ms)77.2449.6165.7450.1653.88
DisWOT(Mr)49.3830.7456.4642.9445.27
DisWOT91.3846.5772.3652.4564.90
Auto-DAS98.4471.8487.7385.9574.66
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Other train-free NAS methods primarily target vanilla training, not distillation. They rely on weak correlation proxies. Ours exhibits a stronger correlation, ensuring more efficient model discovery. While DisWOT considers distillation in its proxy design, it is manually crafted and lacks direct optimization and feedback from distillation results. In contrast, our Auto-DAS method not only includes all distillation designs in the proxy search space settings but also optimizes our proxies directly based on the distillation results. Our approach naturally outperforms other methods in terms of distillation performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.52, + 0.502, + 0.536 + ], + "angle": 0, + "content": "3.5 Training-free Student Search" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.634 + ], + "angle": 0, + "content": "The training-free student search enables us to efficiently explore numerous candidate architectures without the need for costly and time-consuming training. After obtaining an effective proxy with the aid of an evolutionary search, we employ it to conduct a training-free student search. Utilizing randomly initialized weights, denoted as \\(\\mathcal{W}\\), we seek the optimal student, represented as \\(\\alpha^{*}\\), from the search space \\(\\mathcal{A}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.642, + 0.786, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\alpha^ {*} = \\arg \\max _ {\\alpha \\in \\mathcal {A}} \\rho_ {A u t o - D A S} ^ {*} (\\alpha , \\mathcal {W}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.788, + 0.715 + ], + "angle": 0, + "content": "After the search phase, we use the pre-trained teacher model to distill the optimal student network \\(\\alpha^{*}\\). We adopt the original knowledge distillation method [17] as the auxiliary loss function." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.735, + 0.456, + 0.753 + ], + "angle": 0, + "content": "4 Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this section, we present the ranking correlation and search accuracy results of our Auto-DAS on CNN, NAS-Bench, and ViT search space. For fair comparisons, we employ the same search, distillation settings, KD accuracy ground truths in DisWOT as the ground truth for correlation validation. We report mean results based on more than 3 repeated trials." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.127 + ], + "angle": 0, + "content": "Haosen Sun et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.475, + 0.173 + ], + "angle": 0, + "content": "Table 3: Accuracy \\((\\%)\\) of NDS-ResNet space on CIFAR-100." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.185, + 0.742, + 0.244 + ], + "angle": 0, + "content": "
ConstraintsNWOTSynflowDisWOTAuto-DAS
5M-FLOPs63.1964.2865.9866.52
100M-FLOPs70.3872.1272.8973.25
0.5M-Param.70.3871.5872.8973.32
1M-Param.72.5773.5674.2374.66
MethodFitNetsATSPRKDCRD
Random NAS70.1270.1670.4671.1971.59
DisWOT74.8574.574.9574.6275.25
Auto-GAS75.2874.9875.5575.3175.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.481, + 0.145, + 0.739, + 0.173 + ], + "angle": 0, + "content": "Table 4: Results \\((\\%)\\) of other KD methods under 1M parameters." + }, + { + "type": "table_caption", + "bbox": [ + 0.328, + 0.263, + 0.674, + 0.277 + ], + "angle": 0, + "content": "Table 5: Top-1 accuracy (%) results on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.29, + 0.78, + 0.335 + ], + "angle": 0, + "content": "
ModelsBaselineBaseline+WSLDBaseline+DISTRandomNAS+KDZen-NAS+KDDisWOTAuto-DAS
ResNet1869.772.0472.0771.6871.8872.0872.58
ResNet5077.1NA78.4577.4877.8078.6279.25
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.37, + 0.505, + 0.385 + ], + "angle": 0, + "content": "4.1 Experiments on CNN models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.56 + ], + "angle": 0, + "content": "Architecture space and methodology. For CNN validation, we focus on ResNet variants, encompassing both CIFar-ResNet and NDS-ResNet domains. Our CIFar-ResNet space, inspired by He et al. [15], comprises three phases with 1,3,5,7 block options, tailored for CIFAR-100. We also explore the NDS-ResNet realm for both CIFAR-100 and ImageNet-1k search trials. Across all models, we employ our discovered \\(\\rho_{CNN}\\) without additional proxy exploration. During knowledge transfer, all identified student networks undergo training using CRD protocols [48], with ResNet56 serving as the mentor model. For ImageNet-1k, we investigate ResNet18/50-caliber students under equivalent parameter restrictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Findings on compact datasets. Our ranking analysis is rooted in individual knowledge transfer outcomes within the cifar-ResNet domain, as provided by DisWOT. Table 2 demonstrates that Auto-DAS achieves a notably higher Spearman coefficient compared to Fisher, SNIP, FLOPs, and NWOT. Our exploration experiments, conducted in the NDS-ResNet realm, pit our approach against Synflow, NWOT, and DisWOT, under 0.5M and 1M parameter limits, as well as 50M and 100M FLOPs thresholds. As illustrated in Table 3, Auto-DAS surpasses previous state-of-the-art methods by \\(0.9\\% \\sim 2.0\\% \\uparrow\\) under identical constraints. Furthermore, we evaluate Auto-DAS across various KD techniques (refer to Table 4). The results affirm our method's efficacy, consistently outperforming DisWOT across different KD approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "ImageNet performance analysis. Table 5 presents the effectiveness of student models identified by Auto-DAS on ImageNet, with ResNet34/101 acting as teacher networks. For the ResNet18 variant, Auto-DAS attains \\(72.58\\%\\) accuracy, outperforming both DisWOT and other KD+NAS methodologies. Similarly, with the ResNet50 model, Auto-DAS achieves a leading accuracy of \\(79.25\\%\\). These findings underscore the superior capability of our Auto-DAS approach in enhancing model precision across various architectures." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Auto-DAS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.229 + ], + "angle": 0, + "content": "Table 6: Evaluation of Architectures on CIFAR-10, CIFAR-100, and ImageNet-16 in NAS-Bench-201 [11]. Distilled Accuracy (\\%) represents the classification accuracy of the discovered architecture after distillation training. Search Time (s) denotes the computational cost (in GPU-seconds) during the search phase. The performances of NWOT and TE-NAS are obtained from their respective publications. Our proposed Auto-DASapproach demonstrates superior accuracy with rapid search speed." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.241, + 0.784, + 0.388 + ], + "angle": 0, + "content": "
TypeModelCIFAR-10CIFAR-100ImageNet-16-120
Dis. Acc(%)Time (s)Speed-upDis. Acc(%)Time (s)Speed-upDis. Acc(%)Time (s)Speed-up
Multi-trialRS93.63216K1.0×71.28460K1.0×44.881M1.0×
RL [3]92.83216K1.0×71.71460K1.0×44.351M1.0×
BOHB [13]93.49216K1.0×70.84460K1.0×44.331M1.0×
RSPS [25]91.6710K21.6×57.9946K21.6×36.87104K9.6×
One-shotGDAS [10]93.3922K12.0×70.7039K11.7×42.35130K7.7×
DARTS [35]89.2223K9.4×66.2480K5.8×43.18110K9.1×
Zero-shotNWOT [37]93.732.2K100×73.314.6K100×45.4310K100×
EZNAS [2]93.632.2K100×69.824.6K100×43.4710K100×
Zero-shot DASDisWOT93.551.2K180×74.219.2K180×47.3020K180×
DisWOT(Mr)93.490.72K300×73.6218.4K300×45.6340K300×
Auto-DAS93.961.8K180×74.7313.8K180×47.5030K180×
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.418, + 0.496, + 0.433 + ], + "angle": 0, + "content": "4.2 Experiments on NAS-Bench" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.564 + ], + "angle": 0, + "content": "Experimental setup and search space. The NAS-Bench-101 [53] and NAS-Bench-201 [11] datasets serve as crucial benchmarks for assessing the efficacy of various NAS algorithms. DisWOT offers knowledge distillation outcomes for architectures within the NAS-Bench-101/201 search spaces (referred to as NB-101-KD and NB-201-KD). Adhering to these benchmark protocols, we apply the identified \\(\\rho_{CNN}\\) to model search and validate distillation rankings without additional proxy searches. Comprehensive implementation details are available in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Performance analysis. Table 2 illustrates that our Auto-DAS method achieves superior ranking performance, significantly outperforming DisWOT. Furthermore, we evaluate the standard correlation in NAS-Bench-101/201, with the ranking outcomes highlighting Auto-DAS's versatility through substantial improvements over existing techniques. Regarding search trials presented in Table 6, Auto-KD demonstrates remarkable gains in the accuracy-efficiency tradeoff compared to multi-trial NAS approaches (such as Random Search (RS) and Reinforcement Learning (RL)) and Gradient-based One-shot NAS methods (like GDAS and DARTS), which typically incur substantial search costs. Our method also consistently enhances distillation accuracy when compared to other Zero-shot NAS techniques (including NWOT, TE-NAS, and DisWOT). These exceptional ranking and search results underscore the robust generalization capabilities of our approach across challenging NAS-Bench datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.559, + 0.8 + ], + "angle": 0, + "content": "4.3 Experiments on Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Architecture space and methodology. To assess the ranking consistency of zero-cost proxies for ViT structures, we develop a benchmark called ViT-Bench." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.127 + ], + "angle": 0, + "content": "Haosen Sun et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 7: Results (%) of Ranking Correlation on CIFAR-100, Flowers, and Chaoyang." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.172, + 0.785, + 0.259 + ], + "angle": 0, + "content": "
Search SpaceProxyCIFAR-100FlowersChaoyang
KendallSpearmanPearsonKendallSpearmanPearsonKendallSpearmanPearson
PiTGraSP-42.02-58.71-31.53-50.66-69.64-40.90-16.00-22.94-19.07
SynFlow69.7987.0570.8062.2279.9871.6630.9642.6639.24
TENAS-2.13-3.21-1.68-2.86-4.23-3.33-3.34-5.04-3.55
NWOT-2.61-4.13-0.522.673.690.714.736.874.43
TF-TAS63.8382.2058.3764.4882.9167.2337.9952.9242.68
DisWOT-19.05-28.83-28.56-19.57-30.33-18.01-3.362.4015.22
Auto-DAS72.81±1.6890.32±1.7977.76±2.5876.42±1.1293.09±1.8587.63±1.3648.36±2.4666.13±1.5866.58±1.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.282, + 0.787, + 0.338 + ], + "angle": 0, + "content": "Table 8: Comparing the performance of our proposed Auto-DAS approach across three diverse datasets - CIFAR-100, flowers, and Chaoyang - using the PiT search space, we have achieved competitive results while incurring the lowest search cost, measured in time cost on a single GPU." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.352, + 0.787, + 0.45 + ], + "angle": 0, + "content": "
Search SpaceProxyCIFAR-100FlowersChaoyang
Param(M)Dis.Acc(%)Search CostParam(M)Dis.Acc(%)Search CostParam(M)Dis.Acc(%)Search Cost
PiTRandom5.3375.84N/A4.8865.30N/A5.2482.94N/A
GraSP4.5376.031.24 h3.7266.581.85 h4.6383.870.86 h
SynFlow11.0577.131.08 h5.2368.120.99 h4.9383.730.70 h
TENAS6.9376.095.14 h4.2668.035.14 h6.7683.645.07 h
NWOT5.2176.643.02 h10.7767.723.09 h6.3783.313.08 h
TF-TAS16.0777.061.21 h10.3068.210.95 h4.3284.340.71 h
DisWOT10.3875.820.88 h8.7667.580.88 h5.8583.420.88 h
Auto-DAS12.4277.881.28 h10.5269.321.28 h4.8884.831.28 h
Auto-DAS (G-free)11.3577.720.82h10.8869.450.82h5.2584.980.82h
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.488, + 0.787, + 0.684 + ], + "angle": 0, + "content": "This benchmark provides ground-truth accuracy for ViTs on compact datasets (CIFAR-100 [22], Flowers [39], and Chaoyang [58]). Inspired by [24], which shows ViTs gain significantly on small datasets when distilled from an efficient CNN teacher, we concentrate on the distillation accuracy of student ViTs on these datasets with a predetermined teacher. We randomly sample ViTs from PiT [16] search spaces and train them individually using the distillation settings from [24]. ResNet56 serves as the CNN teacher. ViT-Bench encompasses 5k KD accuracy ground truths per dataset. We divide the benchmark into validation and test sets at a 4:3 ratio. For the proxy search phase, we configure EA parameters \\((\\mathcal{P}, \\mathcal{T}, r, k)\\) in Alg. 1 as \\((20, 100, 0.9, 5)\\) and evaluate proxy candidates' ranking performance in the validation set each iteration. Post-search, we equitably compare the Auto-DAS proxy with conventional ones by random sampling in the test set. Our validation and test sets are distinct to ensure unbiased proxy comparisons." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Findings on compact datasets. Table 7 presents the rank consistency results of existing proxies across three datasets. Our proposed Auto-DAS surpasses other notable proxies in metrics like Kendall's tau [1], Spearman's rho [44], and Pearson's correlation coefficient [4], showcasing its efficacy and superiority in achieving consistent and precise ViT architecture search. To further demonstrate Auto-DAS's effectiveness, Table 8 displays search trial results. Our approach achieves superior distillation outcomes and faster search speeds across all three datasets. These empirical findings highlight the importance of effective zero-cost proxies and the role of our proposed Auto-DAS in optimizing ViT architecture search performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Auto-DAS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.316, + 0.145, + 0.687, + 0.16 + ], + "angle": 0, + "content": "Table 9: Results of different ViT models on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.27, + 0.172, + 0.734, + 0.288 + ], + "angle": 0, + "content": "
ModelMethodParam.Top-1 Acc.ModelMethodParam.Top-1 Acc.
DeiT-TinyBaseline+KD5.7M74.5%Swin-TinyBaseline+KD29.0M81.7%
TF-TAS+KD5.7M74.7%TF-TAS+KD29.2M82.0%
DisWOT5.8M75.3%DisWOT29.0M82.2%
Auto-DAS6.1M75.9%Auto-DAS29.2M82.9%
PiT-TinyBaseline+KD4.9M73.2%PiT-SmallBaseline+KD23.5M79.9%
TF-TAS+KD4.9M73.5%TF-TAS+KD24.2M80.1%
DisWOT4.6M73.6%DisWOT23.8M80.3%
Auto-DAS4.8M74.2%Auto-DAS24.5M81.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.322, + 0.787, + 0.351 + ], + "angle": 0, + "content": "Table 10: Spearman correlation (\\%) on optimal proxy in our search space with vanilla inputs or operations." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.363, + 0.78, + 0.508 + ], + "angle": 0, + "content": "
MethodsSettingsResNetNB-101NB-201NB-101-KDNB-201-KDPiT-CIFAR-100PiT-FlowersPiT-Chaoyang
Auto-DASAuto-DAS 98.4471.8487.7385.9574.6690.3293.0966.13
Auto-DAS (instinct)Activations70.2865.8978.9666.3560.8574.8676.8552.86
Gradients77.5372.6886.9273.5566.3880.5682.7556.67
weights60.3552.7570.2558.6755.2560.8965.8241.25
Unary78.8574.2588.2577.4266.5382.8384.3858.69
Binary62.5258.8772.8964.3558.9665.4570.2548.46
All80.5275.8689.6278.5668.5583.2586.5260.12
Auto-DAS (interaction)Feature90.3547.9662.6876.3366.3881.9585.3959.68
Embed88.6546.5560.7574.8563.2880.2583.2756.88
Logits85.4742.5855.8670.5358.6276.8378.9652.45
Transform92.3548.8864.6678.9568.6587.2576.5561.35
Distance89.6346.5363.8575.8265.8682.6984.5658.75
All92.8250.8865.8880.5670.4885.4288.7363.38
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Computational cost analysis. (1) Like most train-free NAS methods, Auto-DAS involves minor additional gradient computations compared to DisWOT. To address this, we explore a gradient-free version of Auto-DAS in Table 8 (labeled Auto-DAS (G-free)). Results indicate that our gradient-free Auto-DAS achieves better accuracy-efficiency trade-offs than DisWOT. (2) While DisWOT is efficient and yields good results on certain datasets, it lacks generalizability across various models and tasks. In contrast, Auto-DAS is automated, versatile, and effective, better suited diverse application needs. Moreover, given the low cost of training-free approaches, the small additional time investment in searching is justified by the consistent performance gains achieved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "ImageNet performance analysis. Examining the results in Table 9, it's clear that Auto-DAS outperforms other methods across various models. For DeiT-Tiny, Auto-DAS achieves \\(75.9\\%\\) accuracy, surpassing Baseline+KD \\((74.5\\%)\\), TF-TAS+KD \\((74.7\\%)\\), and DisWOT \\((75.3\\%)\\). Similarly, Auto-DAS excels with the Swin-Tiny model. This trend persists for PiT-Tiny and PiT-Small models, where Auto-DAS consistently outperforms other methods. These findings underscore the effectiveness and adaptability of our Auto-DAS method in enhancing accuracy across different ViT architectures." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "Haosen Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.143, + 0.409, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.143, + 0.594, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.143, + 0.782, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.262, + 0.784, + 0.292 + ], + "angle": 0, + "content": "Figure 6: Correlation visualization for Auto-DAS on NAS-Bench-101, NAS-Bench-201 and ViT-Bench (from left to right)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.321, + 0.402, + 0.335 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.787, + 0.497 + ], + "angle": 0, + "content": "Proxy search space. We conduct experiments in Table 10 to analyze different components in our proxy search space. The results show that Auto-DAS (interaction) outperforms Auto-DAS (instinct) overall. The feature input and transform operations are particularly important for Auto-DAS (interaction), while weight input and binary operations contribute weakly. On the other hand, gradient input and unary operations show promising results for Auto-DAS (instinct). Different ops have varying impacts on the proxies' accuracy, and combining them enhances the performance of Auto-DAS approaches. These findings highlight the significance of considering different ops and their interactions in the search space for optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.787, + 0.543 + ], + "angle": 0, + "content": "Search algorithm. We use the EA with adaptive-elite selection for proxy search. As shown in Figure 4, the EA adaptive-elite selection obtains better final search results than EA and random search in the proxy search process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.787, + 0.603 + ], + "angle": 0, + "content": "Correlation visualization.. To intuitively observe the proxy's predictability, we visualize the score of Auto-DAS and the ground-truth performance in Figure 6. These visualizations demonstrate that Auto-DAS effectively detects the true distillation results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.627, + 0.36, + 0.643 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.779 + ], + "angle": 0, + "content": "We present Auto-DAS, a novel training-free framework for distillation-aware student architecture search. It includes proxy search and training-free student search. With discovered proxies, Auto-DAS enables efficient search without training costs. This significantly outperforms handcrafted proxies in distillation performance. Comprehensive results on NAS benchmarks, ViT, and CNN spaces over multiple datasets illustrate our method's superior ranking and distillation abilities. We hope our novel investigations will give more insight and new directions for the knowledge distillation and NAS research communities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "**Limitation.** We mainly evaluate the same benchmarks as other NAS methods for fair comparison. We will continue to expand Auto-DAS for more tasks in future work. **Social impact.** Our Auto-DAS focuses on technology improvement without social and ethical implications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Auto-DAS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.207 + ], + "angle": 0, + "content": "1. Abdi, H.: The Kendall rank correlation coefficient. Encyclopedia of measurement and statistics 2, 508-510 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.249 + ], + "angle": 0, + "content": "2. Akhauri, Y., Munoz, J.P., Jain, N., Iyer, R.: EZNAS: Evolving zero-cost proxies for neural architecture scoring. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) NeurIPS (2022), https://openreview.net/forum?id=1SqaDG4dvdt" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.278 + ], + "angle": 0, + "content": "3. Baker, B., Gupta, O., Naik, N., Raskar, R.: Designing neural network architectures using reinforcement learning. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.785, + 0.306 + ], + "angle": 0, + "content": "4. Bowley, A.: The standard deviation of the correlation coefficient. Journal of the American Statistical Association 23(161), 31-34 (1928)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.389 + ], + "angle": 0, + "content": "5. Brown, T.B., Mann, B., Subbiah, N.R.M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., M., D., Ziegler, Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. arXiv preprint, arXiv:2005.14165 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.432 + ], + "angle": 0, + "content": "6. Dong, P., Li, L., Tang, Z., Liu, X., Pan, X., Wang, Q., Chu, X.: Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. In: ICML (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.459 + ], + "angle": 0, + "content": "7. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.502 + ], + "angle": 0, + "content": "8. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV. pp. 17076-17086 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.531 + ], + "angle": 0, + "content": "9. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.559 + ], + "angle": 0, + "content": "0. Dong, X., Yang, Y.: Searching for a robust neural architecture in fourgpu hours. CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.785, + 0.586 + ], + "angle": 0, + "content": "1. Dong, X., Yang, Y.: Nas-bench-201: Extending the scope of reproducible neural architecture search. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.629 + ], + "angle": 0, + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.656 + ], + "angle": 0, + "content": "3. Falkner, S., Klein, A., Hutter, F.: Bohb: Robust and efficient hyperparameter optimization at scale. In: ICML (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.685 + ], + "angle": 0, + "content": "4. Gu, J., Tresp, V.: Search for better students to learn distilled knowledge. arXiv preprint arXiv:2001.11612 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.713 + ], + "angle": 0, + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.742 + ], + "angle": 0, + "content": "6. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.77 + ], + "angle": 0, + "content": "7. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "8. Hu, S., Xie, S., Zheng, H., Liu, C., Shi, J., Liu, X., Lin, D.: Dsnas: Direct neural architecture search without parameter retraining. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Hu, Y., Liang, Y., Guo, Z., Wan, R., Zhang, X., Wei, Y., Gu, Q., Sun, J.: Angle-based search space shrinking for neural architecture search. In: ECCV (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.127 + ], + "angle": 0, + "content": "Haosen Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.771, + 0.189 + ], + "angle": 0, + "content": "21. Kim, Y., Rush, A.M.: Sequence-level knowledge distillation. In: EMNLP (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.749, + 0.202 + ], + "angle": 0, + "content": "22. Krizhevsky, A.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "23. Lee, N., Ajanthan, T., Torr, P.: Snip: Single-shot network pruning based on connection sensitivity. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.27 + ], + "angle": 0, + "content": "24. Li, K., Yu, R., Wang, Z., Yuan, L., Song, G., Chen, J.: Locality guidance for improving vision transformers on tiny datasets. In: European Conference on Computer Vision. pp. 110-127. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "25. Li, L., Talwalkar, A.S.: Random search and reproducibility for neural architecture search. ArXiv (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.785, + 0.324 + ], + "angle": 0, + "content": "26. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "27. Li, L., Bao, Y., Dong, P., Yang, C., Li, A., Luo, W., Liu, Q., Xue, W., Guo, Y.: Detkds: Knowledge distillation search for object detectors. In: ICML (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.379 + ], + "angle": 0, + "content": "28. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "29. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.432 + ], + "angle": 0, + "content": "30. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.472 + ], + "angle": 0, + "content": "31. Li, L., Sun, H., Li, S., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Auto-gas: Automated proxy discovery for training-free generative architecture search. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.5 + ], + "angle": 0, + "content": "32. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "33. Li, L., Wei, Z., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Attnzero: Efficient attention discovery for vision transformers. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.555 + ], + "angle": 0, + "content": "34. Lin, M., Wang, P., Sun, Z., Chen, H., Sun, X., Qian, Q., Li, H., Jin, R.: Zen-nas: A zero-shot nas for high-performance image recognition. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.595 + ], + "angle": 0, + "content": "35. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.623 + ], + "angle": 0, + "content": "36. Liu, Y., Jia, X., Tan, M., Vemulapalli, R., Zhu, Y., Green, B., Wang, X.: Search to distill: Pearls are everywhere but not the eyes. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.65 + ], + "angle": 0, + "content": "37. Mellor, J., Turner, J., Storkey, A., Crowley, E.J.: Neural architecture search without training. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.677 + ], + "angle": 0, + "content": "38. Mirzadeh, S.I., Farajtabar, M., Li, A., Levine, N., Matsukawa, A., Ghasemzadeh, H.: Improved knowledge distillation via teacher assistant. In: AAAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "39. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.745 + ], + "angle": 0, + "content": "40. Park, W., Lu, Y., Cho, M., Kim, D.: Relational knowledge distillation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "41. Pham, H., Guan, M.Y., Zoph, B., Le, Q.V., Dean, J.: Efficient neural architecture search via parameter sharing. In: ICML. pp. 4092-4101 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "42. Real, E., Liang, C., So, D.R., Le, Q.V.: Automl-zero: Evolving machine learning algorithms from scratch (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "43. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Auto-DAS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "44. Stephanou, M., Varughese, M.: Sequential estimation of spearman rank correlation using hermite series estimators. Journal of Multivariate Analysis 186, 104783 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "45. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "46. Theis, L., Korshunova, I., Tejani, A., Huszár, F.: Faster gaze prediction with dense networks and fisher pruning. ArXiv abs/1801.05787 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.787, + 0.272 + ], + "angle": 0, + "content": "47. Tian, Y., Krishnan, D., Isola, P.: Contrastive representation distillation. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.787, + 0.3 + ], + "angle": 0, + "content": "48. Tian, Y., Krishnan, D., Isola, P.: Contrastive representation distillation. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "49. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers amp; distillation through attention. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.787, + 0.37 + ], + "angle": 0, + "content": "50. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.787, + 0.397 + ], + "angle": 0, + "content": "51. Wang, T., Yuan, L., Zhang, X., Feng, J.: Distilling object detectors with fine-grained feature imitation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.787, + 0.425 + ], + "angle": 0, + "content": "52. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.787, + 0.452 + ], + "angle": 0, + "content": "53. Ying, C., Klein, A., Christiansen, E., Real, E., Murphy, K., Hutter, F.: Nas-bench-101: Towards reproducible neural architecture search. In: ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.787, + 0.48 + ], + "angle": 0, + "content": "54. You, S., Huang, T., Yang, M., Wang, F., Qian, C., Zhang, C.: Greedynas: Towards fast one-shot nas with greedy supernet. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.787, + 0.508 + ], + "angle": 0, + "content": "55. Zhang, L., Ma, K.: Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.536 + ], + "angle": 0, + "content": "56. Zhou, H., Song, L., Chen, J., Zhou, Y., Wang, G., Yuan, J., Zhang, Q.: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.563 + ], + "angle": 0, + "content": "57. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.591 + ], + "angle": 0, + "content": "58. Zhu, C., Chen, W., Peng, T., Wang, Y., Jin, M.: Hard sample aware noise robust learning for histopathology image classification. TMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.632 + ], + "angle": 0, + "content": "59. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.632, + 0.787, + 0.66 + ], + "angle": 0, + "content": "60. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.66 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_origin.pdf b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e4a3f74f58827174f04b552e7b75b97baba13e79 --- /dev/null +++ b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/2d2e968b-d0b7-491f-99e1-797a65b3cfb4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d0869fb332e3431d61105c0053aed03974cbf107126233278d5ee25f3164cd1 +size 1355101 diff --git a/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/full.md b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0967ef9c2684baf0fa154359876f9184042fea07 --- /dev/null +++ b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/full.md @@ -0,0 +1,306 @@ +# Auto-DAS: Automated Proxy Discovery for Training-free Distillation-aware Architecture Search + +Haosen Sun $^{1}$ , Lujun Li $^{1\boxtimes}$ , Peijie Dong $^{2}$ , Zimian Wei $^{3}$ , and Shitong Shao $^{4}$ + +$^{1}$ The Hong Kong University of Science and Technology $^{2}$ The Hong Kong University of Science and Technology (Guangzhou) + +$^{3}$ National University of Defense Technology + +$^{4}$ Southeast University +hsunas@connect.ust.hk;lilujunai@gmail.com;pdong212@connect.hkust-gz.edu.cn; weizimian16@nudt.edu.cn; shaoshitong@seu.edu.cn + +Abstract Distillation-aware Architecture Search (DAS) seeks to discover the ideal student architecture that delivers superior performance by distilling knowledge from a given teacher model. Previous DAS methods involve time-consuming training-based search processes. Recently, the training-free DAS method (i.e., DisWOT) proposes KD-based proxies and achieves significant search acceleration. However, we observe that DisWOT suffers from limitations such as the need for manual design and poor generalization to diverse architectures, such as the Vision Transformer (ViT). To address these issues, we present Auto-DAS, an automatic proxy discovery framework using an Evolutionary Algorithm (EA) for training-free DAS. Specifically, we empirically find that proxies conditioned on student instinct statistics and teacher-student interaction statistics can effectively predict distillation accuracy. Then, we represent the proxy with computation graphs and construct the proxy search space using instinct and interaction statistics as inputs. To identify promising proxies, our search space incorporates various types of basic transformations and network distance operators inspired by previous proxy and KD-loss designs. Next, our EA initializes populations, evaluates, performs crossover and mutation operations, and selects the best correlation candidate with distillation accuracy. We introduce an adaptive-elite selection strategy to enhance search efficiency and strive for a balance between exploitation and exploration. Finally, we conduct training-free DAS with discovered proxy before the optimal student distillation phase. In this way, our auto-discovery framework eliminates the need for manual design and tuning, while also adapting to different search spaces through direct correlation optimization. Extensive experiments demonstrate that Auto-DAS generalizes well to various architectures and search spaces (e.g., ResNet, ViT, NAS-Bench-101, and NAS-Bench-201), achieving state-of-the-art results in both ranking correlation and final searched accuracy. Code at: https://github.com/lliai/Auto-DAS. + +Keywords: Knowledge Distillation $\cdot$ Architecture Search $\cdot$ AutoML + +# 1 Introduction + +Knowledge Distillation (KD) transfers logits [17, 52], features [26, 32], or relation [40, 47] knowledge of heavy teacher models as additional supervised signals to augment lightweight student models. KD originally served as model compression technique for large models and has recently been used as a common strategy for performance improvement in different scenarios. For example, distilling a CNN teacher for a Vision Transformer (ViT) [49] can significantly enhance its performance, particularly when training from scratch in data-limited scenarios. However, the architecture gap between the teacher and student network creates a mismatch between the receptive field and the prediction reliability [38, 56]. These disparities contribute to distillation gains, i.e., students with a smaller architecture gap with teachers will enjoy better final results. + +To alleviate these problems, a new task, Distillation-aware Architecture Search (DAS), is presented to search for an optimal student architecture from the given teacher, which obtains the best performance with the help of neural architecture search. In contrast to the traditional NAS methods, the DAS approach needs to consider the effect of the teacher and employ distillation accuracy as the search evaluation metric. Some observations [7,36] indicate that the models searched by traditional NAS methods [9,20,35] are suboptimal in distillation, and there is a common performance ranking gap for obtained students between vanilla training and distillation. For the DAS task, some training-based methods leverage Reinforcement Learning (RL) [36] or gradient method [14] as search strategies to find the ideal student. However, these methods bring in high search costs (e.g., 5 days with 200 TPUv2 in AKD [36]), which causes heavy burdens for application. Recently, DisWOT [7] first presents the training-free DAS framework with some KD-based proxies, exploiting the connection between training-free proxies and knowledge distillation loss. It achieves significant search acceleration by using statistics between random networks for predicting distillation accuracy without candidate training. However, some drawbacks still limit its extensive application, which could highlight from two aspects: (1) Expert design and tuning costs. These KD-based proxies are inspired by some distillation studies with extensive expert intuition and require extensive tuning. It is hard to state if these proxies are optimal for DAS as limited by the designer's experience. (2) Generalization issue. Handcrafted ones only perform well on some specific settings but are hard to adapt to different scenarios without changing/modification. As shown in Figure 1, existing handcrafted proxies perform well on the CNNs but do not generalize well to ViT models. Thus, two questions arise: (1) How to efficiently discover the proxies without expert knowledge? (2) How to find the optimal proxies for different search spaces? + +For the first problem, we observe that previous proxy expressions are usually constructed for typical inputs and common primitive operations. Inspired by Auto-Zero [42], we explore a from-scratch proxy discover framework alternative to handcrafted for DAS task. For the second problem, we can evolve various proxy candidates by directly taking their predictability for distillation accuracy as feedback. In addition, we find that proxies with inputs from student + +![](images/d12551264457751c84f472b52274f3cecd4b9be8d16ed1feec256c248236d827.jpg) +Figure 1: Spearman correlation $(\%)$ on CNN models(left) and ViT models(right). + +![](images/583e57acfb12c7a207318434e031723c5fc7098778de63e64f4d9b30c71a3b05.jpg) +Figure 2: Illustration of Auto-DAS. + +distinct statistics and teacher-student interaction statistics consistently achieve better results. For example, DisWOT+Synflow in Figure 1 achieves stable improvements on multiple search spaces. Based on these observations, we present Auto-DAS, a from-scratch framework for discovering proxies for DAS tasks are proposed as an alternative to manually designed ones. As shown in Figure 2, we first build the proxy search space, consisting of both instinct and interaction proxy. The instinct proxy is used with student's weights and gradient statistics as input and different unary and binary mathematical operations as candidate components. For interaction proxy, we select teacher-student features, embeddings, logits as inputs, different basic transformation operations, and distance functions to form the computation graph. Then, we randomly sample various proxies as the population and then perform mutation and crossover operations according to the ranking ability of different proxies. To improve generality, we directly treat the Kendall correlation between the prediction of the candidate proxy and the ground truth under distillation as the fitting objective of the evolutionary algorithm. To speed up the process, we employ an adaptive-elite selection strategy during the evolution search. With the automatic search framework, our Auto-DAS surpasses existing training-free NAS approaches by a large margin without involving too much prior knowledge. Finally, after getting the Auto-DAS proxy in a given search space, we search for the best student architectures with the highest Auto-DAS score and apply the searched students to distill the knowledge from the given teacher. + +In comparison to manual DAS techniques, our Auto-DASframework offers several advantages: (1) Autonomy. It facilitates the spontaneous discovery of more potent and efficient proxies that may elude human experts, thus mitigating bias and ensuring that the resultant architectures are tailored for the intended problem or dataset. (2) Versatility. The search mechanism of AutoDASenhances adaptability by uncovering proxies that perform well across diverse search spaces. Our framework can evolve flexible proxies to generalize across architectures, based on the ultimate distillation outcomes for the search objectives. (3) Enlightening. Our approach enables the automatic identification of more expressive, efficient, and effective proxies for optimizing and customizing student architectures, paving the way for novel research and applications in DAS and KD tasks. (4) Efficacy. We have conducted comprehensive experiments on ResNet, NAS-Bench-101, NAS-Bench-201, and ViT-Bench search spaces to confirm the + +superiority of our Auto-DAS. The results demonstrate that our Auto-DAScan surpass other zero-shot proxies in distillation accuracy and achieve state-of-the-art results in rank consistency, such as a $40\%$ improvement in Kendall correlation over DisWOT. To summarize, our key contributions to Auto-DAS are: + +- We introduce the pioneering framework for discovering proxies from scratch, Auto-DAS, to address handcrafted limitations and enhance generalizability in predicting distillation accuracy across various students. +- We devise a well-structured proxy search space, incorporating instinct and interaction statistics as inputs, along with diverse unary, binary, transforms, and distance functions as operations. We develop evolutionary algorithms with adaptive-elite selection for effective and efficient proxy search. +- We provide empirical evidence that Auto-DASachieves cutting-edge performance across ResNet, NAS-Bench-101, NAS-Bench-201, ViT-Bench search spaces and multiple datasets. + +# 2 Related work + +Distillation-aware architecture search. The increasing demand for practical applications of large-scale models [5, 12] has led to impressive advancements in Knowledge distillation (KD) across various domains [26-30, 32, 43], including image classification [17], object detection [51, 55], and NLP [21]. However, discrepancies between teacher and student architectures can substantially impact distillation outcomes. Consequently, distillation-aware student architecture search for a given teacher model has emerged as a crucial new area of research. Traditional reinforcement learning-based search methods [36] utilize KD-loss as feedback for DAS but incur significant additional costs during the search process. Subsequent research employs efficient gradient-based methods [14], albeit introducing new optimization challenges. Recently, DisWOT [7] introduced the first training-free DAS framework. However, DisWOT relies on a manually designed proxy, which lacks generalization across different search spaces. To address this limitation, we propose the first automated proxy discovery for training-free DAS tasks. Our method not only retains the advantages of DisWOT, such as efficiency and flexibility, but also automatically optimizes proxies without expert intervention. Auto-DAS paves the way for novel DAS research and offers valuable insights based on search result analysis for the research community. + +Training-free architecture search. Conventional NAS [18, 60] involves designing search spaces, algorithms, and evaluations to automatically discover optimal architectures under given constraints. Training-based methods employ train-then-search or weight sharing processes [19, 41, 54, 60]. Training-free NAS, also known as zero-shot NAS [6, 8, 31, 33, 34, 57, 59], enables faster search by predicting performance without training. Zero-cost proxies are categorized as parameter-level or architecture-level. Parameter-level approaches rely on pruning and saliency values [23, 45, 50], while architecture-level methods evaluate expressiveness [34] or linear map correlations [37]. DisWOT proposed KD-based + +![](images/1e97d29716cc13def24a2a57360e1a9f3f2e92f997add73ccd837dafed587af9.jpg) +Figure 3: The overall illustration of Auto-DAS. Our search space includes unary, binary, transform, and distance options with different inputs. Then, we construct our intrinsic & interaction proxies and use crossover & mutation to search for the best proxy. + +![](images/3b122400afaf5e511e6128e4655bcb0395e563f90d364110258e92d40e1704ea.jpg) + +Table 1: Operations of Auto-DAS. More details are available in the Appendix. + +
ProxyInputTypePrimitive Operations
instinctActivationsUnaryexp, mish, leaky, relu, tanh, softmax, sigmoid, pow2, pow4, log, abslog, abssqrt
GradientsUnaryllnorm, revert, min - maxto - mean, to - std, invert, lnNorm, normalizedSum, no - op
WeightsBinarysum, subtract, multiply, divide, dot.
InteractionFeatureTransformtanh, log, sqrt, batchnorm, min - max, norm, softmax, exp, mish, leaky, relu
EmbedTransformmask, no, bmm, mm, scale, multi - scale, scale, local, batch, channel, drop, satt, natt, catt
LogitsDistancel1, l2, ℓKL, ℓhard, ℓCosine, ℓPearson
+ +proxies, but their fixed formulations limit generalization. To overcome these limitations, we consider student internal statistics and teacher-student similarities as inputs and evolve superior proxies from scratch for training-free DAS tasks. Inspired by Auto-Zero, our innovations include a novel proxy search space and a specialized search process, promoting automatic proxy design for distillation. Our approach differs from EZNAS [2] in terms of tasks (ours is customized for CNNs and ViTs), search algorithms (ours employs adaptive-elite selection), and search space (ours incorporates teacher-student knowledge). Unlike EZNAS and Auto-KD, which focus solely on vanilla classification, we directly search on the distillation task and include teacher-student knowledge and additional options in the search space. + +# 3 Methodology + +This section delves into the intricacies of existing proxies, our proposed search space, the search algorithm, and an analysis of the results. Subsequently, we outline a training-free student search process utilizing the discovered zero-cost proxy. Figure 3 illustrates the workflow of our approach. + +# 3.1 Examination of existing proxies + +To provide context for our training-free method, we first examine current zero-cost proxies: SNIP [23] calculates a saliency metric at initialization using a single mini-batch of data to estimate the loss change when a specific parameter is eliminated. Synflow [45] introduces a modified version of synaptic saliency scores that prevents layer collapse during parameter pruning, while Fisher [46] computes the sum over all activation gradients in the network, applicable for channel pruning. Their detailed formulations are as follows: + +$$ +\rho_ {s n i p} = \left| \frac {\partial \mathcal {L}}{\partial \mathcal {W}} \odot \mathcal {W} \right|, \rho_ {s y n f l o w} = \frac {\partial \mathcal {L}}{\partial \mathcal {W}} \odot \mathcal {W}, \rho_ {f i s h e r} = \left(\frac {\partial \mathcal {L}}{\partial \mathcal {A}} \mathcal {A}\right) ^ {2}, \tag {1} +$$ + +where $\mathcal{L},\mathcal{W},\mathcal{A}$ are loss function, weight and activation. $H$ is the Hessian matrix. These proxies consist of various unary, binary operations with statistics input from the candidate model, named instinct proxies in our paper. For the distillation scenario, DisWOT presents some KD-based proxies based on KD-loss design, as follows: + +$$ +\rho_ {D i s W O T} = \mathcal {D} _ {L 2} (\mathcal {G} ([ \mathcal {A} S, \mathcal {A} T ])) + \mathcal {D} _ {L 2} (\mathcal {G} ([ \mathcal {F} S, \mathcal {F} T ])), \tag {2} +$$ + +where $\mathcal{A}$ , $\mathcal{F}$ , $\mathcal{E}$ , $\mathcal{P}$ are activation maps, feature maps, embeddings, and logits prediction. $[\cdot, \cdot]$ means to perform the same operation on two primitives. $\mathcal{G}$ is the gram-matrix. $\mathcal{D}_{L2}$ , $\mathcal{D}_{KL}$ are $\ell_2$ and $\ell_{KL}$ distance. These KD-based proxies use teacher-student pair-wise statistics as input and different transformations and distance functions to construct formations that we denote as interaction proxies. Based on these analyses, we build search spaces for instinct and interactive proxies in the next part. + +# 3.2 Search Space for Proxy Discovery + +Our search space structure comprises three primary components: input choices, primitive operations, and proxy representation. As illustrated in Table 1, for input choices in constructing zero-cost proxies, the instinct proxy extracts activations, gradients, and weights from the student model, while the interaction proxy selects features, embeddings, and logits of the teacher-student network. These inputs represent the most informative options common across traditional and KD-based proxies. For instance, features provide insights into data distribution and internal network representations, while gradients highlight weight sensitivity to loss functions. Following this observation, we register each student layer's activation, weights, and corresponding gradients as potential inputs for our zero-cost proxy. Regarding primitive operations, we consider unary and binary operations for instinct proxies, and transform and distance operations for interaction proxies. We include mathematical, normalization, and scaling operations for unary operations; matrix sums and products for binary operations; and activation, normalization, scaling, and other transformations, along with typical distance functions for interaction proxies. Our proxy is represented as a computation graph, where input nodes are instinct and interaction statistics, and + +Algorithm 1 Evolutionary Strategy for Auto-DAS Proxy Discovery +Input: Architecture space $S$ , candidate set $\mathcal{P}$ , iteration limit $\mathcal{T}$ , selection proportion $r$ , temporary set $\mathcal{R}$ , elite count $k$ . +Output: Highest-scoring Auto-DAS configuration. +1: Initialize candidate set $\mathcal{P}0 := \text{GenerateInitialCandidates}(P_i)$ ; +2: Empty temporary set $\mathcal{R} := \emptyset$ ; +3: for $i = 1,2,\ldots,\mathcal{T}$ do +4: Reset temporary set $\mathcal{R} := \emptyset$ ; +5: Populate $\mathcal{R}$ with random selections from $\mathcal{P}$ ; +6: Elite group $G_i k := \text{SelectTopPerformers}(\mathcal{R}, k)$ ; +7: Choose parent $G_i^p := \text{RandomlyPick}(G_i k)$ ; +8: Generate offspring $G_i^m := \text{MUTATE}(G_i^p)$ ; +9: // Adaptive Elite Preservation +10: Randomly generate baseline $E$ ; +11: if $\rho(G_i^m) > \rho(E)$ then +12: Include $G_i^m$ in $P$ ; +13: else +14: Include $E$ in $P$ ; +15: end if +16: Eliminate lowest-scoring proxy from population. +17: end for + +intermediate nodes are primitive operations. The graph's output is the proxy score used for distillation ranking. Our comprehensive search space encompasses most zero-cost proxies, KD-based losses, and model similarity formulations, constituting our core innovation and contribution. + +# 3.3 Evolution Procedure and Objective + +Utilizing our proxy search space, we employ an Evolutionary Algorithm (EA) to identify optimal proxy expressions (refer to Figure 3 and Algorithm 1). Our EA initiates with a population of candidate proxies and iteratively evolves it over generations using genetic operators such as selection, crossover, and mutation to generate superior solutions. For objective fitting, each proxy is evaluated based on the ranking correlation between its output proxy $\mathcal{Q}$ scores and actual performances $\mathcal{D}$ to efficiently discover the optimal proxy $\rho^{*}$ from search space $\mathcal{S}$ , as follows: + +$$ +\rho_ {A u t o - D A S} ^ {*} = \arg \max \rho \in \mathcal {S} (\tau (\mathcal {D}, \mathcal {Q})), \tag {3} +$$ + +where Kendall's Tau $\tau$ serves as the correlation coefficient. Each evolution generation selects the top-k candidates with the highest Auto-DAS scores and randomly chooses a parent from these candidates for mutation. During mutation, a node in the computation graph is randomly selected and mutated with newly generated primitive operations. + +Efficient proxy search with existing benchmark. Obtaining pre-trained results $\mathcal{D}$ can be resource-intensive. To address this, we leverage pre-computed architecture-performance pairs from existing DisWOT benchmarks [7]. These + +![](images/1574db6e0a5d877c676ddfcae74fa4b96f79bbe72cec7416dcdf1fc68b316557.jpg) +Figure 4: SearchFigure 5: Correlation visualization of NWOT (left), DisWOT curve on ResNet. (middle), Auto-DAS (right) on ResNet search space. + +![](images/3fc68fb86f5381bfe872a5a8275673afc92045b793d2fe30ba3e4e9f9bdf4933.jpg) + +![](images/61c3ce3061e85c0eba9a00b7d38a2643303b3c5e614c2e15f28b45621882ba63.jpg) + +![](images/ab24aa712070466b8d6158c7f20bca7c44d9e974e18a9e304b02bf5ecad163c5.jpg) + +benchmarks provide distillation results for various ResNet and NAS-Bench-101/201 search space models. Utilizing this benchmark eliminates the need to retrain performance ground-truths and significantly reduces the proxy search process time to merely 0.05 GPU-day. Due to its excellent generality (demonstrated in our extensive experiments), our searched proxy $\rho_{Auto - DAS}$ transfers well to similar search spaces and can be directly applied in most scenarios without additional search budget. + +Adaptive-elite selection strategy. To mitigate population deterioration and premature convergence, we introduce an adaptive-elite selection strategy. This approach involves comparing the performance of mutation-generated offspring with that of a randomly generated individual. The individual with a higher Auto-DAS score is then added to the population. By maintaining or enhancing the overall population performance over time, this strategy accelerates convergence towards a high-performing solution (see Figure 4). + +# 3.4 Searched Zero-cost Proxy + +We present formulas of the searched proxies on $\rho_{CNN}$ for CNN Search space (e.g., ResNet, NAS-Bench) and $\rho_{ViT}$ ViT search space as follows: + +$$ +\rho_ {C N N} = \mathcal {D} _ {K L} (\operatorname {s i g m o i d} ([ \mathcal {F} _ {S}, \mathcal {F} _ {T} ])) + \frac {1}{M} \sum \left| \log \operatorname {s o f t m a x} \left(\frac {\partial L}{\partial W}\right) \right| _ {F}, \tag {4} +$$ + +$$ +\rho_ {V I T} = \mathcal {D} _ {K L} (\operatorname {m a s k} ([ \mathcal {F} _ {S}, \mathcal {F} _ {T} ])) + \frac {M}{\sum \operatorname {s i g m o i d} \left( \right.\left| \right. \log \left| \frac {\partial L}{\partial W} \right|\left. \right)} \tag {5} +$$ + +Where $W$ is the weight parameter, $\partial \mathcal{L} / \partial W$ is the corresponding gradient. $\| \cdot \| _F$ means the Frobenius-norm. $M$ denotes mean operation. + +Analysis of searched proxy. The discovered formulations indicate that (1) mathematical operations like softmax, logsoftmax, sigmoid, and mask operations benefit proxy predictability. (2) Gradient and features are vital input statistics for instinct and interaction proxies. As shown in Figure 5, the discovered proxies of Auto-DAS achieve significant ranking improvement. + +Understanding why our Auto-DAS surpasses other NAS methods. (1) Many training-based distillation NAS methods use weight sharing and proxies, often leading to inaccurate performance estimates. Our Auto-DAS searches proxies based on real performance, yielding one with strong ranking ability. (2) + +Table 2: Spearman correlation (\%) results on the NAS-Bench-101 and NAS-Bench-201 (NB-101/201) datasets. In this context, NB-101/201-KD denotes the task of distilling the accuracies of the architectures on the respective NAS-Bench-101/201 datasets. + +
MethodResNetNB-101NB-201NB-101-KDNB-201-KD
FLOPs72.9230.8163.3815.5664.55
Fisher81.37-38.8135.91-33.924.45
Grad_Norm82.35-39.2358.70-39.16-10.01
SNIP85.07-29.0158.17-21.7816.91
Synflow88.3043.6974.6120.3674.63
NWOT45.6632.8464.4122.9735.27
DisWOT(Ms)77.2449.6165.7450.1653.88
DisWOT(Mr)49.3830.7456.4642.9445.27
DisWOT91.3846.5772.3652.4564.90
Auto-DAS98.4471.8487.7385.9574.66
+ +Other train-free NAS methods primarily target vanilla training, not distillation. They rely on weak correlation proxies. Ours exhibits a stronger correlation, ensuring more efficient model discovery. While DisWOT considers distillation in its proxy design, it is manually crafted and lacks direct optimization and feedback from distillation results. In contrast, our Auto-DAS method not only includes all distillation designs in the proxy search space settings but also optimizes our proxies directly based on the distillation results. Our approach naturally outperforms other methods in terms of distillation performance. + +# 3.5 Training-free Student Search + +The training-free student search enables us to efficiently explore numerous candidate architectures without the need for costly and time-consuming training. After obtaining an effective proxy with the aid of an evolutionary search, we employ it to conduct a training-free student search. Utilizing randomly initialized weights, denoted as $\mathcal{W}$ , we seek the optimal student, represented as $\alpha^{*}$ , from the search space $\mathcal{A}$ , as follows: + +$$ +\alpha^ {*} = \arg \max _ {\alpha \in \mathcal {A}} \rho_ {A u t o - D A S} ^ {*} (\alpha , \mathcal {W}). \tag {6} +$$ + +After the search phase, we use the pre-trained teacher model to distill the optimal student network $\alpha^{*}$ . We adopt the original knowledge distillation method [17] as the auxiliary loss function. + +# 4 Experimental results + +In this section, we present the ranking correlation and search accuracy results of our Auto-DAS on CNN, NAS-Bench, and ViT search space. For fair comparisons, we employ the same search, distillation settings, KD accuracy ground truths in DisWOT as the ground truth for correlation validation. We report mean results based on more than 3 repeated trials. + +Table 3: Accuracy $(\%)$ of NDS-ResNet space on CIFAR-100. + +
ConstraintsNWOTSynflowDisWOTAuto-DAS
5M-FLOPs63.1964.2865.9866.52
100M-FLOPs70.3872.1272.8973.25
0.5M-Param.70.3871.5872.8973.32
1M-Param.72.5773.5674.2374.66
MethodFitNetsATSPRKDCRD
Random NAS70.1270.1670.4671.1971.59
DisWOT74.8574.574.9574.6275.25
Auto-GAS75.2874.9875.5575.3175.86
+ +Table 4: Results $(\%)$ of other KD methods under 1M parameters. +Table 5: Top-1 accuracy (%) results on ImageNet. + +
ModelsBaselineBaseline+WSLDBaseline+DISTRandomNAS+KDZen-NAS+KDDisWOTAuto-DAS
ResNet1869.772.0472.0771.6871.8872.0872.58
ResNet5077.1NA78.4577.4877.8078.6279.25
+ +# 4.1 Experiments on CNN models + +Architecture space and methodology. For CNN validation, we focus on ResNet variants, encompassing both CIFar-ResNet and NDS-ResNet domains. Our CIFar-ResNet space, inspired by He et al. [15], comprises three phases with 1,3,5,7 block options, tailored for CIFAR-100. We also explore the NDS-ResNet realm for both CIFAR-100 and ImageNet-1k search trials. Across all models, we employ our discovered $\rho_{CNN}$ without additional proxy exploration. During knowledge transfer, all identified student networks undergo training using CRD protocols [48], with ResNet56 serving as the mentor model. For ImageNet-1k, we investigate ResNet18/50-caliber students under equivalent parameter restrictions. + +Findings on compact datasets. Our ranking analysis is rooted in individual knowledge transfer outcomes within the cifar-ResNet domain, as provided by DisWOT. Table 2 demonstrates that Auto-DAS achieves a notably higher Spearman coefficient compared to Fisher, SNIP, FLOPs, and NWOT. Our exploration experiments, conducted in the NDS-ResNet realm, pit our approach against Synflow, NWOT, and DisWOT, under 0.5M and 1M parameter limits, as well as 50M and 100M FLOPs thresholds. As illustrated in Table 3, Auto-DAS surpasses previous state-of-the-art methods by $0.9\% \sim 2.0\% \uparrow$ under identical constraints. Furthermore, we evaluate Auto-DAS across various KD techniques (refer to Table 4). The results affirm our method's efficacy, consistently outperforming DisWOT across different KD approaches. + +ImageNet performance analysis. Table 5 presents the effectiveness of student models identified by Auto-DAS on ImageNet, with ResNet34/101 acting as teacher networks. For the ResNet18 variant, Auto-DAS attains $72.58\%$ accuracy, outperforming both DisWOT and other KD+NAS methodologies. Similarly, with the ResNet50 model, Auto-DAS achieves a leading accuracy of $79.25\%$ . These findings underscore the superior capability of our Auto-DAS approach in enhancing model precision across various architectures. + +Table 6: Evaluation of Architectures on CIFAR-10, CIFAR-100, and ImageNet-16 in NAS-Bench-201 [11]. Distilled Accuracy (\%) represents the classification accuracy of the discovered architecture after distillation training. Search Time (s) denotes the computational cost (in GPU-seconds) during the search phase. The performances of NWOT and TE-NAS are obtained from their respective publications. Our proposed Auto-DASapproach demonstrates superior accuracy with rapid search speed. + +
TypeModelCIFAR-10CIFAR-100ImageNet-16-120
Dis. Acc(%)Time (s)Speed-upDis. Acc(%)Time (s)Speed-upDis. Acc(%)Time (s)Speed-up
Multi-trialRS93.63216K1.0×71.28460K1.0×44.881M1.0×
RL [3]92.83216K1.0×71.71460K1.0×44.351M1.0×
BOHB [13]93.49216K1.0×70.84460K1.0×44.331M1.0×
RSPS [25]91.6710K21.6×57.9946K21.6×36.87104K9.6×
One-shotGDAS [10]93.3922K12.0×70.7039K11.7×42.35130K7.7×
DARTS [35]89.2223K9.4×66.2480K5.8×43.18110K9.1×
Zero-shotNWOT [37]93.732.2K100×73.314.6K100×45.4310K100×
EZNAS [2]93.632.2K100×69.824.6K100×43.4710K100×
Zero-shot DASDisWOT93.551.2K180×74.219.2K180×47.3020K180×
DisWOT(Mr)93.490.72K300×73.6218.4K300×45.6340K300×
Auto-DAS93.961.8K180×74.7313.8K180×47.5030K180×
+ +# 4.2 Experiments on NAS-Bench + +Experimental setup and search space. The NAS-Bench-101 [53] and NAS-Bench-201 [11] datasets serve as crucial benchmarks for assessing the efficacy of various NAS algorithms. DisWOT offers knowledge distillation outcomes for architectures within the NAS-Bench-101/201 search spaces (referred to as NB-101-KD and NB-201-KD). Adhering to these benchmark protocols, we apply the identified $\rho_{CNN}$ to model search and validate distillation rankings without additional proxy searches. Comprehensive implementation details are available in the Appendix. + +Performance analysis. Table 2 illustrates that our Auto-DAS method achieves superior ranking performance, significantly outperforming DisWOT. Furthermore, we evaluate the standard correlation in NAS-Bench-101/201, with the ranking outcomes highlighting Auto-DAS's versatility through substantial improvements over existing techniques. Regarding search trials presented in Table 6, Auto-KD demonstrates remarkable gains in the accuracy-efficiency tradeoff compared to multi-trial NAS approaches (such as Random Search (RS) and Reinforcement Learning (RL)) and Gradient-based One-shot NAS methods (like GDAS and DARTS), which typically incur substantial search costs. Our method also consistently enhances distillation accuracy when compared to other Zero-shot NAS techniques (including NWOT, TE-NAS, and DisWOT). These exceptional ranking and search results underscore the robust generalization capabilities of our approach across challenging NAS-Bench datasets. + +# 4.3 Experiments on Vision Transformer + +Architecture space and methodology. To assess the ranking consistency of zero-cost proxies for ViT structures, we develop a benchmark called ViT-Bench. + +Table 7: Results (%) of Ranking Correlation on CIFAR-100, Flowers, and Chaoyang. + +
Search SpaceProxyCIFAR-100FlowersChaoyang
KendallSpearmanPearsonKendallSpearmanPearsonKendallSpearmanPearson
PiTGraSP-42.02-58.71-31.53-50.66-69.64-40.90-16.00-22.94-19.07
SynFlow69.7987.0570.8062.2279.9871.6630.9642.6639.24
TENAS-2.13-3.21-1.68-2.86-4.23-3.33-3.34-5.04-3.55
NWOT-2.61-4.13-0.522.673.690.714.736.874.43
TF-TAS63.8382.2058.3764.4882.9167.2337.9952.9242.68
DisWOT-19.05-28.83-28.56-19.57-30.33-18.01-3.362.4015.22
Auto-DAS72.81±1.6890.32±1.7977.76±2.5876.42±1.1293.09±1.8587.63±1.3648.36±2.4666.13±1.5866.58±1.86
+ +Table 8: Comparing the performance of our proposed Auto-DAS approach across three diverse datasets - CIFAR-100, flowers, and Chaoyang - using the PiT search space, we have achieved competitive results while incurring the lowest search cost, measured in time cost on a single GPU. + +
Search SpaceProxyCIFAR-100FlowersChaoyang
Param(M)Dis.Acc(%)Search CostParam(M)Dis.Acc(%)Search CostParam(M)Dis.Acc(%)Search Cost
PiTRandom5.3375.84N/A4.8865.30N/A5.2482.94N/A
GraSP4.5376.031.24 h3.7266.581.85 h4.6383.870.86 h
SynFlow11.0577.131.08 h5.2368.120.99 h4.9383.730.70 h
TENAS6.9376.095.14 h4.2668.035.14 h6.7683.645.07 h
NWOT5.2176.643.02 h10.7767.723.09 h6.3783.313.08 h
TF-TAS16.0777.061.21 h10.3068.210.95 h4.3284.340.71 h
DisWOT10.3875.820.88 h8.7667.580.88 h5.8583.420.88 h
Auto-DAS12.4277.881.28 h10.5269.321.28 h4.8884.831.28 h
Auto-DAS (G-free)11.3577.720.82h10.8869.450.82h5.2584.980.82h
+ +This benchmark provides ground-truth accuracy for ViTs on compact datasets (CIFAR-100 [22], Flowers [39], and Chaoyang [58]). Inspired by [24], which shows ViTs gain significantly on small datasets when distilled from an efficient CNN teacher, we concentrate on the distillation accuracy of student ViTs on these datasets with a predetermined teacher. We randomly sample ViTs from PiT [16] search spaces and train them individually using the distillation settings from [24]. ResNet56 serves as the CNN teacher. ViT-Bench encompasses 5k KD accuracy ground truths per dataset. We divide the benchmark into validation and test sets at a 4:3 ratio. For the proxy search phase, we configure EA parameters $(\mathcal{P}, \mathcal{T}, r, k)$ in Alg. 1 as $(20, 100, 0.9, 5)$ and evaluate proxy candidates' ranking performance in the validation set each iteration. Post-search, we equitably compare the Auto-DAS proxy with conventional ones by random sampling in the test set. Our validation and test sets are distinct to ensure unbiased proxy comparisons. + +Findings on compact datasets. Table 7 presents the rank consistency results of existing proxies across three datasets. Our proposed Auto-DAS surpasses other notable proxies in metrics like Kendall's tau [1], Spearman's rho [44], and Pearson's correlation coefficient [4], showcasing its efficacy and superiority in achieving consistent and precise ViT architecture search. To further demonstrate Auto-DAS's effectiveness, Table 8 displays search trial results. Our approach achieves superior distillation outcomes and faster search speeds across all three datasets. These empirical findings highlight the importance of effective zero-cost proxies and the role of our proposed Auto-DAS in optimizing ViT architecture search performance. + +Table 9: Results of different ViT models on ImageNet. + +
ModelMethodParam.Top-1 Acc.ModelMethodParam.Top-1 Acc.
DeiT-TinyBaseline+KD5.7M74.5%Swin-TinyBaseline+KD29.0M81.7%
TF-TAS+KD5.7M74.7%TF-TAS+KD29.2M82.0%
DisWOT5.8M75.3%DisWOT29.0M82.2%
Auto-DAS6.1M75.9%Auto-DAS29.2M82.9%
PiT-TinyBaseline+KD4.9M73.2%PiT-SmallBaseline+KD23.5M79.9%
TF-TAS+KD4.9M73.5%TF-TAS+KD24.2M80.1%
DisWOT4.6M73.6%DisWOT23.8M80.3%
Auto-DAS4.8M74.2%Auto-DAS24.5M81.0%
+ +Table 10: Spearman correlation (\%) on optimal proxy in our search space with vanilla inputs or operations. + +
MethodsSettingsResNetNB-101NB-201NB-101-KDNB-201-KDPiT-CIFAR-100PiT-FlowersPiT-Chaoyang
Auto-DASAuto-DAS 98.4471.8487.7385.9574.6690.3293.0966.13
Auto-DAS (instinct)Activations70.2865.8978.9666.3560.8574.8676.8552.86
Gradients77.5372.6886.9273.5566.3880.5682.7556.67
weights60.3552.7570.2558.6755.2560.8965.8241.25
Unary78.8574.2588.2577.4266.5382.8384.3858.69
Binary62.5258.8772.8964.3558.9665.4570.2548.46
All80.5275.8689.6278.5668.5583.2586.5260.12
Auto-DAS (interaction)Feature90.3547.9662.6876.3366.3881.9585.3959.68
Embed88.6546.5560.7574.8563.2880.2583.2756.88
Logits85.4742.5855.8670.5358.6276.8378.9652.45
Transform92.3548.8864.6678.9568.6587.2576.5561.35
Distance89.6346.5363.8575.8265.8682.6984.5658.75
All92.8250.8865.8880.5670.4885.4288.7363.38
+ +Computational cost analysis. (1) Like most train-free NAS methods, Auto-DAS involves minor additional gradient computations compared to DisWOT. To address this, we explore a gradient-free version of Auto-DAS in Table 8 (labeled Auto-DAS (G-free)). Results indicate that our gradient-free Auto-DAS achieves better accuracy-efficiency trade-offs than DisWOT. (2) While DisWOT is efficient and yields good results on certain datasets, it lacks generalizability across various models and tasks. In contrast, Auto-DAS is automated, versatile, and effective, better suited diverse application needs. Moreover, given the low cost of training-free approaches, the small additional time investment in searching is justified by the consistent performance gains achieved. + +ImageNet performance analysis. Examining the results in Table 9, it's clear that Auto-DAS outperforms other methods across various models. For DeiT-Tiny, Auto-DAS achieves $75.9\%$ accuracy, surpassing Baseline+KD $(74.5\%)$ , TF-TAS+KD $(74.7\%)$ , and DisWOT $(75.3\%)$ . Similarly, Auto-DAS excels with the Swin-Tiny model. This trend persists for PiT-Tiny and PiT-Small models, where Auto-DAS consistently outperforms other methods. These findings underscore the effectiveness and adaptability of our Auto-DAS method in enhancing accuracy across different ViT architectures. + +![](images/9736a737a53a422dcaa4f5de4cf7ff6a38e21b12d826d6d6a6b047362536a128.jpg) +Figure 6: Correlation visualization for Auto-DAS on NAS-Bench-101, NAS-Bench-201 and ViT-Bench (from left to right). + +![](images/82aa422130857dbfa25b0857318cda29db1312bdd47caca9495564c368fe004a.jpg) + +![](images/1da0ebc9194a28aff3c1d5b979a994f9d3d224093998cd2d6713dfebf1dd147c.jpg) + +# 4.4 Ablation Studies + +Proxy search space. We conduct experiments in Table 10 to analyze different components in our proxy search space. The results show that Auto-DAS (interaction) outperforms Auto-DAS (instinct) overall. The feature input and transform operations are particularly important for Auto-DAS (interaction), while weight input and binary operations contribute weakly. On the other hand, gradient input and unary operations show promising results for Auto-DAS (instinct). Different ops have varying impacts on the proxies' accuracy, and combining them enhances the performance of Auto-DAS approaches. These findings highlight the significance of considering different ops and their interactions in the search space for optimal performance. + +Search algorithm. We use the EA with adaptive-elite selection for proxy search. As shown in Figure 4, the EA adaptive-elite selection obtains better final search results than EA and random search in the proxy search process. + +Correlation visualization.. To intuitively observe the proxy's predictability, we visualize the score of Auto-DAS and the ground-truth performance in Figure 6. These visualizations demonstrate that Auto-DAS effectively detects the true distillation results. + +# 5 Conclusion + +We present Auto-DAS, a novel training-free framework for distillation-aware student architecture search. It includes proxy search and training-free student search. With discovered proxies, Auto-DAS enables efficient search without training costs. This significantly outperforms handcrafted proxies in distillation performance. Comprehensive results on NAS benchmarks, ViT, and CNN spaces over multiple datasets illustrate our method's superior ranking and distillation abilities. We hope our novel investigations will give more insight and new directions for the knowledge distillation and NAS research communities. + +**Limitation.** We mainly evaluate the same benchmarks as other NAS methods for fair comparison. We will continue to expand Auto-DAS for more tasks in future work. **Social impact.** Our Auto-DAS focuses on technology improvement without social and ethical implications. + +# References + +1. Abdi, H.: The Kendall rank correlation coefficient. Encyclopedia of measurement and statistics 2, 508-510 (2007) +2. Akhauri, Y., Munoz, J.P., Jain, N., Iyer, R.: EZNAS: Evolving zero-cost proxies for neural architecture scoring. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) NeurIPS (2022), https://openreview.net/forum?id=1SqaDG4dvdt +3. Baker, B., Gupta, O., Naik, N., Raskar, R.: Designing neural network architectures using reinforcement learning. In: ICLR (2017) +4. Bowley, A.: The standard deviation of the correlation coefficient. Journal of the American Statistical Association 23(161), 31-34 (1928) +5. Brown, T.B., Mann, B., Subbiah, N.R.M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., M., D., Ziegler, Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. arXiv preprint, arXiv:2005.14165 (2020) +6. Dong, P., Li, L., Tang, Z., Liu, X., Pan, X., Wang, Q., Chu, X.: Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. In: ICML (2024) +7. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023) +8. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV. pp. 17076-17086 (2023) +9. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022) +0. Dong, X., Yang, Y.: Searching for a robust neural architecture in fourgpu hours. CVPR (2019) +1. Dong, X., Yang, Y.: Nas-bench-201: Extending the scope of reproducible neural architecture search. In: ICLR (2019) +2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) +3. Falkner, S., Klein, A., Hutter, F.: Bohb: Robust and efficient hyperparameter optimization at scale. In: ICML (2018) +4. Gu, J., Tresp, V.: Search for better students to learn distilled knowledge. arXiv preprint arXiv:2001.11612 (2020) +5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +6. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021) +7. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +8. Hu, S., Xie, S., Zheng, H., Liu, C., Shi, J., Liu, X., Lin, D.: Dsnas: Direct neural architecture search without parameter retraining. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +9. Hu, Y., Liang, Y., Guo, Z., Wan, R., Zhang, X., Wei, Y., Gu, Q., Sun, J.: Angle-based search space shrinking for neural architecture search. In: ECCV (2020) + +20. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021) +21. Kim, Y., Rush, A.M.: Sequence-level knowledge distillation. In: EMNLP (2016) +22. Krizhevsky, A.: Learning multiple layers of features from tiny images (2009) +23. Lee, N., Ajanthan, T., Torr, P.: Snip: Single-shot network pruning based on connection sensitivity. In: ICLR (2018) +24. Li, K., Yu, R., Wang, Z., Yuan, L., Song, G., Chen, J.: Locality guidance for improving vision transformers on tiny datasets. In: European Conference on Computer Vision. pp. 110-127. Springer (2022) +25. Li, L., Talwalkar, A.S.: Random search and reproducibility for neural architecture search. ArXiv (2019) +26. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022) +27. Li, L., Bao, Y., Dong, P., Yang, C., Li, A., Luo, W., Liu, Q., Xue, W., Guo, Y.: Detkds: Knowledge distillation search for object detectors. In: ICML (2024) +28. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024) +29. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023) +30. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022) +31. Li, L., Sun, H., Li, S., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Auto-gas: Automated proxy discovery for training-free generative architecture search. In: ECCV (2024) +32. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation. In: ICLR (2020) +33. Li, L., Wei, Z., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Attnzero: Efficient attention discovery for vision transformers. In: ECCV (2024) +34. Lin, M., Wang, P., Sun, Z., Chen, H., Sun, X., Qian, Q., Li, H., Jin, R.: Zen-nas: A zero-shot nas for high-performance image recognition. ICCV (2021) +35. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019) +36. Liu, Y., Jia, X., Tan, M., Vemulapalli, R., Zhu, Y., Green, B., Wang, X.: Search to distill: Pearls are everywhere but not the eyes. In: CVPR (2020) +37. Mellor, J., Turner, J., Storkey, A., Crowley, E.J.: Neural architecture search without training. In: ICML (2021) +38. Mirzadeh, S.I., Farajtabar, M., Li, A., Levine, N., Matsukawa, A., Ghasemzadeh, H.: Improved knowledge distillation via teacher assistant. In: AAAI (2020) +39. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008) +40. Park, W., Lu, Y., Cho, M., Kim, D.: Relational knowledge distillation. In: CVPR (2019) +41. Pham, H., Guan, M.Y., Zoph, B., Le, Q.V., Dean, J.: Efficient neural architecture search via parameter sharing. In: ICML. pp. 4092-4101 (2018) +42. Real, E., Liang, C., So, D.R., Le, Q.V.: Automl-zero: Evolving machine learning algorithms from scratch (2020) +43. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023) + +44. Stephanou, M., Varughese, M.: Sequential estimation of spearman rank correlation using hermite series estimators. Journal of Multivariate Analysis 186, 104783 (2021) +45. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS (2020) +46. Theis, L., Korshunova, I., Tejani, A., Huszár, F.: Faster gaze prediction with dense networks and fisher pruning. ArXiv abs/1801.05787 (2018) +47. Tian, Y., Krishnan, D., Isola, P.: Contrastive representation distillation. In: ICLR (2020) +48. Tian, Y., Krishnan, D., Isola, P.: Contrastive representation distillation. In: ICLR (2020) +49. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers amp; distillation through attention. In: ICML (2021) +50. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020) +51. Wang, T., Yuan, L., Zhang, X., Feng, J.: Distilling object detectors with fine-grained feature imitation. In: CVPR (2019) +52. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching (2022) +53. Ying, C., Klein, A., Christiansen, E., Real, E., Murphy, K., Hutter, F.: Nas-bench-101: Towards reproducible neural architecture search. In: ICML (2019) +54. You, S., Huang, T., Yang, M., Wang, F., Qian, C., Zhang, C.: Greedynas: Towards fast one-shot nas with greedy supernet. In: CVPR (2020) +55. Zhang, L., Ma, K.: Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In: ICLR (2020) +56. Zhou, H., Song, L., Chen, J., Zhou, Y., Wang, G., Yuan, J., Zhang, Q.: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective (2021) +57. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024) +58. Zhu, C., Chen, W., Peng, T., Wang, Y., Jin, M.: Hard sample aware noise robust learning for histopathology image classification. TMI (2021) +59. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024) +60. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018) \ No newline at end of file diff --git a/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/images.zip b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c76e2f35b5de376dc1eba2f66bda3cfeba9b77e5 --- /dev/null +++ b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bc7380286e3a43d58756286f1fb1cfd551d66a228ffd4d228c63ba874f32d2f +size 566585 diff --git a/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/layout.json b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5619fe67afa68f8d9a63f9bc00a816f8f3030fb --- /dev/null +++ b/2024/Auto-DAS_ Automated Proxy Discovery for Training-free Distillation-aware Architecture Search/layout.json @@ -0,0 +1,8729 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 111, + 462, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 111, + 462, + 163 + ], + "spans": [ + { + "bbox": [ + 152, + 111, + 462, + 163 + ], + "type": "text", + "content": "Auto-DAS: Automated Proxy Discovery for Training-free Distillation-aware Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "spans": [ + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "content": "Haosen Sun" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "content": ", Lujun Li" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "content": ", Peijie Dong" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "content": ", Zimian Wei" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "content": ", and Shitong Shao" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 207, + 449, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 207, + 449, + 230 + ], + "spans": [ + { + "bbox": [ + 148, + 207, + 449, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 207, + 449, + 230 + ], + "type": "text", + "content": " The Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 148, + 207, + 449, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 207, + 449, + 230 + ], + "type": "text", + "content": " The Hong Kong University of Science and Technology (Guangzhou)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 230, + 397, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 230, + 397, + 241 + ], + "spans": [ + { + "bbox": [ + 216, + 230, + 397, + 241 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 216, + 230, + 397, + 241 + ], + "type": "text", + "content": " National University of Defense Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 241, + 487, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 487, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 487, + 273 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 241, + 487, + 273 + ], + "type": "text", + "content": " Southeast University \nhsunas@connect.ust.hk;lilujunai@gmail.com;pdong212@connect.hkust-gz.edu.cn; weizimian16@nudt.edu.cn; shaoshitong@seu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 299, + 452, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 299, + 452, + 628 + ], + "spans": [ + { + "bbox": [ + 160, + 299, + 452, + 628 + ], + "type": "text", + "content": "Abstract Distillation-aware Architecture Search (DAS) seeks to discover the ideal student architecture that delivers superior performance by distilling knowledge from a given teacher model. Previous DAS methods involve time-consuming training-based search processes. Recently, the training-free DAS method (i.e., DisWOT) proposes KD-based proxies and achieves significant search acceleration. However, we observe that DisWOT suffers from limitations such as the need for manual design and poor generalization to diverse architectures, such as the Vision Transformer (ViT). To address these issues, we present Auto-DAS, an automatic proxy discovery framework using an Evolutionary Algorithm (EA) for training-free DAS. Specifically, we empirically find that proxies conditioned on student instinct statistics and teacher-student interaction statistics can effectively predict distillation accuracy. Then, we represent the proxy with computation graphs and construct the proxy search space using instinct and interaction statistics as inputs. To identify promising proxies, our search space incorporates various types of basic transformations and network distance operators inspired by previous proxy and KD-loss designs. Next, our EA initializes populations, evaluates, performs crossover and mutation operations, and selects the best correlation candidate with distillation accuracy. We introduce an adaptive-elite selection strategy to enhance search efficiency and strive for a balance between exploitation and exploration. Finally, we conduct training-free DAS with discovered proxy before the optimal student distillation phase. In this way, our auto-discovery framework eliminates the need for manual design and tuning, while also adapting to different search spaces through direct correlation optimization. Extensive experiments demonstrate that Auto-DAS generalizes well to various architectures and search spaces (e.g., ResNet, ViT, NAS-Bench-101, and NAS-Bench-201), achieving state-of-the-art results in both ranking correlation and final searched accuracy. Code at: https://github.com/lliai/Auto-DAS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 638, + 441, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 638, + 441, + 649 + ], + "spans": [ + { + "bbox": [ + 160, + 638, + 441, + 649 + ], + "type": "text", + "content": "Keywords: Knowledge Distillation " + }, + { + "bbox": [ + 160, + 638, + 441, + 649 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 441, + 649 + ], + "type": "text", + "content": " Architecture Search " + }, + { + "bbox": [ + 160, + 638, + 441, + 649 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 441, + 649 + ], + "type": "text", + "content": " AutoML" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 654, + 392, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 654, + 392, + 665 + ], + "spans": [ + { + "bbox": [ + 140, + 654, + 392, + 665 + ], + "type": "text", + "content": "Corresponding author & contributed equally to this work." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "content": "Knowledge Distillation (KD) transfers logits [17, 52], features [26, 32], or relation [40, 47] knowledge of heavy teacher models as additional supervised signals to augment lightweight student models. KD originally served as model compression technique for large models and has recently been used as a common strategy for performance improvement in different scenarios. For example, distilling a CNN teacher for a Vision Transformer (ViT) [49] can significantly enhance its performance, particularly when training from scratch in data-limited scenarios. However, the architecture gap between the teacher and student network creates a mismatch between the receptive field and the prediction reliability [38, 56]. These disparities contribute to distillation gains, i.e., students with a smaller architecture gap with teachers will enjoy better final results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 594 + ], + "type": "text", + "content": "To alleviate these problems, a new task, Distillation-aware Architecture Search (DAS), is presented to search for an optimal student architecture from the given teacher, which obtains the best performance with the help of neural architecture search. In contrast to the traditional NAS methods, the DAS approach needs to consider the effect of the teacher and employ distillation accuracy as the search evaluation metric. Some observations [7,36] indicate that the models searched by traditional NAS methods [9,20,35] are suboptimal in distillation, and there is a common performance ranking gap for obtained students between vanilla training and distillation. For the DAS task, some training-based methods leverage Reinforcement Learning (RL) [36] or gradient method [14] as search strategies to find the ideal student. However, these methods bring in high search costs (e.g., 5 days with 200 TPUv2 in AKD [36]), which causes heavy burdens for application. Recently, DisWOT [7] first presents the training-free DAS framework with some KD-based proxies, exploiting the connection between training-free proxies and knowledge distillation loss. It achieves significant search acceleration by using statistics between random networks for predicting distillation accuracy without candidate training. However, some drawbacks still limit its extensive application, which could highlight from two aspects: (1) Expert design and tuning costs. These KD-based proxies are inspired by some distillation studies with extensive expert intuition and require extensive tuning. It is hard to state if these proxies are optimal for DAS as limited by the designer's experience. (2) Generalization issue. Handcrafted ones only perform well on some specific settings but are hard to adapt to different scenarios without changing/modification. As shown in Figure 1, existing handcrafted proxies perform well on the CNNs but do not generalize well to ViT models. Thus, two questions arise: (1) How to efficiently discover the proxies without expert knowledge? (2) How to find the optimal proxies for different search spaces?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "For the first problem, we observe that previous proxy expressions are usually constructed for typical inputs and common primitive operations. Inspired by Auto-Zero [42], we explore a from-scratch proxy discover framework alternative to handcrafted for DAS task. For the second problem, we can evolve various proxy candidates by directly taking their predictability for distillation accuracy as feedback. In addition, we find that proxies with inputs from student" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "text", + "content": "Haosen Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 309, + 190 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 309, + 190 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 309, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 309, + 190 + ], + "type": "image", + "image_path": "d12551264457751c84f472b52274f3cecd4b9be8d16ed1feec256c248236d827.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 198, + 304, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 198, + 304, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 304, + 220 + ], + "type": "text", + "content": "Figure 1: Spearman correlation " + }, + { + "bbox": [ + 132, + 198, + 304, + 220 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 132, + 198, + 304, + 220 + ], + "type": "text", + "content": " on CNN models(left) and ViT models(right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 318, + 113, + 470, + 190 + ], + "blocks": [ + { + "bbox": [ + 318, + 113, + 470, + 190 + ], + "lines": [ + { + "bbox": [ + 318, + 113, + 470, + 190 + ], + "spans": [ + { + "bbox": [ + 318, + 113, + 470, + 190 + ], + "type": "image", + "image_path": "583e57acfb12c7a207318434e031723c5fc7098778de63e64f4d9b30c71a3b05.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 198, + 468, + 209 + ], + "lines": [ + { + "bbox": [ + 318, + 198, + 468, + 209 + ], + "spans": [ + { + "bbox": [ + 318, + 198, + 468, + 209 + ], + "type": "text", + "content": "Figure 2: Illustration of Auto-DAS." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 245, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 509 + ], + "type": "text", + "content": "distinct statistics and teacher-student interaction statistics consistently achieve better results. For example, DisWOT+Synflow in Figure 1 achieves stable improvements on multiple search spaces. Based on these observations, we present Auto-DAS, a from-scratch framework for discovering proxies for DAS tasks are proposed as an alternative to manually designed ones. As shown in Figure 2, we first build the proxy search space, consisting of both instinct and interaction proxy. The instinct proxy is used with student's weights and gradient statistics as input and different unary and binary mathematical operations as candidate components. For interaction proxy, we select teacher-student features, embeddings, logits as inputs, different basic transformation operations, and distance functions to form the computation graph. Then, we randomly sample various proxies as the population and then perform mutation and crossover operations according to the ranking ability of different proxies. To improve generality, we directly treat the Kendall correlation between the prediction of the candidate proxy and the ground truth under distillation as the fitting objective of the evolutionary algorithm. To speed up the process, we employ an adaptive-elite selection strategy during the evolution search. With the automatic search framework, our Auto-DAS surpasses existing training-free NAS approaches by a large margin without involving too much prior knowledge. Finally, after getting the Auto-DAS proxy in a given search space, we search for the best student architectures with the highest Auto-DAS score and apply the searched students to distill the knowledge from the given teacher." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In comparison to manual DAS techniques, our Auto-DASframework offers several advantages: (1) Autonomy. It facilitates the spontaneous discovery of more potent and efficient proxies that may elude human experts, thus mitigating bias and ensuring that the resultant architectures are tailored for the intended problem or dataset. (2) Versatility. The search mechanism of AutoDASenhances adaptability by uncovering proxies that perform well across diverse search spaces. Our framework can evolve flexible proxies to generalize across architectures, based on the ultimate distillation outcomes for the search objectives. (3) Enlightening. Our approach enables the automatic identification of more expressive, efficient, and effective proxies for optimizing and customizing student architectures, paving the way for novel research and applications in DAS and KD tasks. (4) Efficacy. We have conducted comprehensive experiments on ResNet, NAS-Bench-101, NAS-Bench-201, and ViT-Bench search spaces to confirm the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Auto-DAS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "superiority of our Auto-DAS. The results demonstrate that our Auto-DAScan surpass other zero-shot proxies in distillation accuracy and achieve state-of-the-art results in rank consistency, such as a " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " improvement in Kendall correlation over DisWOT. To summarize, our key contributions to Auto-DAS are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 175, + 480, + 294 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 175, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 175, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 175, + 479, + 209 + ], + "type": "text", + "content": "- We introduce the pioneering framework for discovering proxies from scratch, Auto-DAS, to address handcrafted limitations and enhance generalizability in predicting distillation accuracy across various students." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 212, + 480, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 212, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 139, + 212, + 480, + 258 + ], + "type": "text", + "content": "- We devise a well-structured proxy search space, incorporating instinct and interaction statistics as inputs, along with diverse unary, binary, transforms, and distance functions as operations. We develop evolutionary algorithms with adaptive-elite selection for effective and efficient proxy search." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 260, + 479, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 260, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 139, + 260, + 479, + 294 + ], + "type": "text", + "content": "- We provide empirical evidence that Auto-DASachieves cutting-edge performance across ResNet, NAS-Bench-101, NAS-Bench-201, ViT-Bench search spaces and multiple datasets." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 316, + 234, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 234, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 234, + 328 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 556 + ], + "type": "text", + "content": "Distillation-aware architecture search. The increasing demand for practical applications of large-scale models [5, 12] has led to impressive advancements in Knowledge distillation (KD) across various domains [26-30, 32, 43], including image classification [17], object detection [51, 55], and NLP [21]. However, discrepancies between teacher and student architectures can substantially impact distillation outcomes. Consequently, distillation-aware student architecture search for a given teacher model has emerged as a crucial new area of research. Traditional reinforcement learning-based search methods [36] utilize KD-loss as feedback for DAS but incur significant additional costs during the search process. Subsequent research employs efficient gradient-based methods [14], albeit introducing new optimization challenges. Recently, DisWOT [7] introduced the first training-free DAS framework. However, DisWOT relies on a manually designed proxy, which lacks generalization across different search spaces. To address this limitation, we propose the first automated proxy discovery for training-free DAS tasks. Our method not only retains the advantages of DisWOT, such as efficiency and flexibility, but also automatically optimizes proxies without expert intervention. Auto-DAS paves the way for novel DAS research and offers valuable insights based on search result analysis for the research community." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Training-free architecture search. Conventional NAS [18, 60] involves designing search spaces, algorithms, and evaluations to automatically discover optimal architectures under given constraints. Training-based methods employ train-then-search or weight sharing processes [19, 41, 54, 60]. Training-free NAS, also known as zero-shot NAS [6, 8, 31, 33, 34, 57, 59], enables faster search by predicting performance without training. Zero-cost proxies are categorized as parameter-level or architecture-level. Parameter-level approaches rely on pruning and saliency values [23, 45, 50], while architecture-level methods evaluate expressiveness [34] or linear map correlations [37]. DisWOT proposed KD-based" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "Haosen Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 278, + 270 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 278, + 270 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 278, + 270 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 278, + 270 + ], + "type": "image", + "image_path": "1e97d29716cc13def24a2a57360e1a9f3f2e92f997add73ccd837dafed587af9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 277, + 480, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 322 + ], + "type": "text", + "content": "Figure 3: The overall illustration of Auto-DAS. Our search space includes unary, binary, transform, and distance options with different inputs. Then, we construct our intrinsic & interaction proxies and use crossover & mutation to search for the best proxy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 285, + 113, + 463, + 270 + ], + "blocks": [ + { + "bbox": [ + 285, + 113, + 463, + 270 + ], + "lines": [ + { + "bbox": [ + 285, + 113, + 463, + 270 + ], + "spans": [ + { + "bbox": [ + 285, + 113, + 463, + 270 + ], + "type": "image", + "image_path": "3b122400afaf5e511e6128e4655bcb0395e563f90d364110258e92d40e1704ea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 355, + 476, + 420 + ], + "blocks": [ + { + "bbox": [ + 144, + 334, + 466, + 345 + ], + "lines": [ + { + "bbox": [ + 144, + 334, + 466, + 345 + ], + "spans": [ + { + "bbox": [ + 144, + 334, + 466, + 345 + ], + "type": "text", + "content": "Table 1: Operations of Auto-DAS. More details are available in the Appendix." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 355, + 476, + 420 + ], + "lines": [ + { + "bbox": [ + 136, + 355, + 476, + 420 + ], + "spans": [ + { + "bbox": [ + 136, + 355, + 476, + 420 + ], + "type": "table", + "html": "
ProxyInputTypePrimitive Operations
instinctActivationsUnaryexp, mish, leaky, relu, tanh, softmax, sigmoid, pow2, pow4, log, abslog, abssqrt
GradientsUnaryllnorm, revert, min - maxto - mean, to - std, invert, lnNorm, normalizedSum, no - op
WeightsBinarysum, subtract, multiply, divide, dot.
InteractionFeatureTransformtanh, log, sqrt, batchnorm, min - max, norm, softmax, exp, mish, leaky, relu
EmbedTransformmask, no, bmm, mm, scale, multi - scale, scale, local, batch, channel, drop, satt, natt, catt
LogitsDistancel1, l2, ℓKL, ℓhard, ℓCosine, ℓPearson
", + "image_path": "162eb9b9fbd2f004362805105d6c1cf357d745b0a94663998fb870436edc57c6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 442, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 575 + ], + "type": "text", + "content": "proxies, but their fixed formulations limit generalization. To overcome these limitations, we consider student internal statistics and teacher-student similarities as inputs and evolve superior proxies from scratch for training-free DAS tasks. Inspired by Auto-Zero, our innovations include a novel proxy search space and a specialized search process, promoting automatic proxy design for distillation. Our approach differs from EZNAS [2] in terms of tasks (ours is customized for CNNs and ViTs), search algorithms (ours employs adaptive-elite selection), and search space (ours incorporates teacher-student knowledge). Unlike EZNAS and Auto-KD, which focus solely on vanilla classification, we directly search on the distillation task and include teacher-student knowledge and additional options in the search space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 592, + 233, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 233, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 233, + 606 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "This section delves into the intricacies of existing proxies, our proposed search space, the search algorithm, and an analysis of the results. Subsequently, we outline a training-free student search process utilizing the discovered zero-cost proxy. Figure 3 illustrates the workflow of our approach." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "Auto-DAS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 320, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 320, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 320, + 129 + ], + "type": "text", + "content": "3.1 Examination of existing proxies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 218 + ], + "type": "text", + "content": "To provide context for our training-free method, we first examine current zero-cost proxies: SNIP [23] calculates a saliency metric at initialization using a single mini-batch of data to estimate the loss change when a specific parameter is eliminated. Synflow [45] introduces a modified version of synaptic saliency scores that prevents layer collapse during parameter pruning, while Fisher [46] computes the sum over all activation gradients in the network, applicable for channel pruning. Their detailed formulations are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 227, + 481, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 227, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 169, + 227, + 481, + 255 + ], + "type": "interline_equation", + "content": "\\rho_ {s n i p} = \\left| \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W} \\right|, \\rho_ {s y n f l o w} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W}, \\rho_ {f i s h e r} = \\left(\\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {A}} \\mathcal {A}\\right) ^ {2}, \\tag {1}", + "image_path": "c5ce37c0da9c186146ce298a16de497ef9f7ab3b613f713c417fa283017f0fc2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 257, + 482, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 257, + 482, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{L},\\mathcal{W},\\mathcal{A}" + }, + { + "bbox": [ + 130, + 257, + 482, + 317 + ], + "type": "text", + "content": " are loss function, weight and activation. " + }, + { + "bbox": [ + 130, + 257, + 482, + 317 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 257, + 482, + 317 + ], + "type": "text", + "content": " is the Hessian matrix. These proxies consist of various unary, binary operations with statistics input from the candidate model, named instinct proxies in our paper. For the distillation scenario, DisWOT presents some KD-based proxies based on KD-loss design, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 326, + 481, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 326, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 192, + 326, + 481, + 338 + ], + "type": "interline_equation", + "content": "\\rho_ {D i s W O T} = \\mathcal {D} _ {L 2} (\\mathcal {G} ([ \\mathcal {A} S, \\mathcal {A} T ])) + \\mathcal {D} _ {L 2} (\\mathcal {G} ([ \\mathcal {F} S, \\mathcal {F} T ])), \\tag {2}", + "image_path": "1f5157d5b7b271fdb8cc7b2689e43e78f5137f953bed2c347d2057d781fce6c3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": " are activation maps, feature maps, embeddings, and logits prediction. " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": " means to perform the same operation on two primitives. " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": " is the gram-matrix. " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L2}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{KL}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\ell_{KL}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": " distance. These KD-based proxies use teacher-student pair-wise statistics as input and different transformations and distance functions to construct formations that we denote as interaction proxies. Based on these analyses, we build search spaces for instinct and interactive proxies in the next part." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 444, + 331, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 331, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 331, + 456 + ], + "type": "text", + "content": "3.2 Search Space for Proxy Discovery" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": "Our search space structure comprises three primary components: input choices, primitive operations, and proxy representation. As illustrated in Table 1, for input choices in constructing zero-cost proxies, the instinct proxy extracts activations, gradients, and weights from the student model, while the interaction proxy selects features, embeddings, and logits of the teacher-student network. These inputs represent the most informative options common across traditional and KD-based proxies. For instance, features provide insights into data distribution and internal network representations, while gradients highlight weight sensitivity to loss functions. Following this observation, we register each student layer's activation, weights, and corresponding gradients as potential inputs for our zero-cost proxy. Regarding primitive operations, we consider unary and binary operations for instinct proxies, and transform and distance operations for interaction proxies. We include mathematical, normalization, and scaling operations for unary operations; matrix sums and products for binary operations; and activation, normalization, scaling, and other transformations, along with typical distance functions for interaction proxies. Our proxy is represented as a computation graph, where input nodes are instinct and interaction statistics, and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "Haosen Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 129, + 481, + 352 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 434, + 128 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 434, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 434, + 128 + ], + "type": "text", + "content": "Algorithm 1 Evolutionary Strategy for Auto-DAS Proxy Discovery" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "Input: Architecture space " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", candidate set " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", iteration limit " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", selection proportion " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", temporary set " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", elite count " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ". \nOutput: Highest-scoring Auto-DAS configuration. \n1: Initialize candidate set " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{P}0 := \\text{GenerateInitialCandidates}(P_i)" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n2: Empty temporary set " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{R} := \\emptyset" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n3: for " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "i = 1,2,\\ldots,\\mathcal{T}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " do \n4: Reset temporary set " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{R} := \\emptyset" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n5: Populate " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " with random selections from " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n6: Elite group " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "G_i k := \\text{SelectTopPerformers}(\\mathcal{R}, k)" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n7: Choose parent " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "G_i^p := \\text{RandomlyPick}(G_i k)" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n8: Generate offspring " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "G_i^m := \\text{MUTATE}(G_i^p)" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n9: // Adaptive Elite Preservation \n10: Randomly generate baseline " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n11: if " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\rho(G_i^m) > \\rho(E)" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " then \n12: Include " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "G_i^m" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n13: else \n14: Include " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n15: end if \n16: Eliminate lowest-scoring proxy from population. \n17: end for" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 375, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 422 + ], + "type": "text", + "content": "intermediate nodes are primitive operations. The graph's output is the proxy score used for distillation ranking. Our comprehensive search space encompasses most zero-cost proxies, KD-based losses, and model similarity formulations, constituting our core innovation and contribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 440, + 338, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 440, + 338, + 452 + ], + "spans": [ + { + "bbox": [ + 131, + 440, + 338, + 452 + ], + "type": "text", + "content": "3.3 Evolution Procedure and Objective" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "text", + "content": "Utilizing our proxy search space, we employ an Evolutionary Algorithm (EA) to identify optimal proxy expressions (refer to Figure 3 and Algorithm 1). Our EA initiates with a population of candidate proxies and iteratively evolves it over generations using genetic operators such as selection, crossover, and mutation to generate superior solutions. For objective fitting, each proxy is evaluated based on the ranking correlation between its output proxy " + }, + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "text", + "content": " scores and actual performances " + }, + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "text", + "content": " to efficiently discover the optimal proxy " + }, + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "inline_equation", + "content": "\\rho^{*}" + }, + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "text", + "content": " from search space " + }, + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 459, + 480, + 553 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 555, + 480, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 555, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 228, + 555, + 480, + 566 + ], + "type": "interline_equation", + "content": "\\rho_ {A u t o - D A S} ^ {*} = \\arg \\max \\rho \\in \\mathcal {S} (\\tau (\\mathcal {D}, \\mathcal {Q})), \\tag {3}", + "image_path": "f924cad0a69dacfee48af29c8cfbc468ffd600766f91180389d560eff0e2bee1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "content": "where Kendall's Tau " + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "content": " serves as the correlation coefficient. Each evolution generation selects the top-k candidates with the highest Auto-DAS scores and randomly chooses a parent from these candidates for mutation. During mutation, a node in the computation graph is randomly selected and mutated with newly generated primitive operations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "content": "Efficient proxy search with existing benchmark. Obtaining pre-trained results " + }, + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "content": " can be resource-intensive. To address this, we leverage pre-computed architecture-performance pairs from existing DisWOT benchmarks [7]. These" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Auto-DAS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 220, + 178 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 220, + 178 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 220, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 220, + 178 + ], + "type": "image", + "image_path": "1574db6e0a5d877c676ddfcae74fa4b96f79bbe72cec7416dcdf1fc68b316557.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 186, + 482, + 209 + ], + "lines": [ + { + "bbox": [ + 131, + 186, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 186, + 482, + 209 + ], + "type": "text", + "content": "Figure 4: SearchFigure 5: Correlation visualization of NWOT (left), DisWOT curve on ResNet. (middle), Auto-DAS (right) on ResNet search space." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 304, + 178 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 304, + 178 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 304, + 178 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 304, + 178 + ], + "type": "image", + "image_path": "3fc68fb86f5381bfe872a5a8275673afc92045b793d2fe30ba3e4e9f9bdf4933.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 391, + 178 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 391, + 178 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 391, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 391, + 178 + ], + "type": "image", + "image_path": "61c3ce3061e85c0eba9a00b7d38a2643303b3c5e614c2e15f28b45621882ba63.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 391, + 114, + 477, + 178 + ], + "blocks": [ + { + "bbox": [ + 391, + 114, + 477, + 178 + ], + "lines": [ + { + "bbox": [ + 391, + 114, + 477, + 178 + ], + "spans": [ + { + "bbox": [ + 391, + 114, + 477, + 178 + ], + "type": "image", + "image_path": "ab24aa712070466b8d6158c7f20bca7c44d9e974e18a9e304b02bf5ecad163c5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 233, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 479, + 317 + ], + "type": "text", + "content": "benchmarks provide distillation results for various ResNet and NAS-Bench-101/201 search space models. Utilizing this benchmark eliminates the need to retrain performance ground-truths and significantly reduces the proxy search process time to merely 0.05 GPU-day. Due to its excellent generality (demonstrated in our extensive experiments), our searched proxy " + }, + { + "bbox": [ + 130, + 233, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\rho_{Auto - DAS}" + }, + { + "bbox": [ + 130, + 233, + 479, + 317 + ], + "type": "text", + "content": " transfers well to similar search spaces and can be directly applied in most scenarios without additional search budget." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 318, + 480, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 401 + ], + "type": "text", + "content": "Adaptive-elite selection strategy. To mitigate population deterioration and premature convergence, we introduce an adaptive-elite selection strategy. This approach involves comparing the performance of mutation-generated offspring with that of a randomly generated individual. The individual with a higher Auto-DAS score is then added to the population. By maintaining or enhancing the overall population performance over time, this strategy accelerates convergence towards a high-performing solution (see Figure 4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 421, + 290, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 421, + 290, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 290, + 433 + ], + "type": "text", + "content": "3.4 Searched Zero-cost Proxy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 442, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 482, + 467 + ], + "type": "text", + "content": "We present formulas of the searched proxies on " + }, + { + "bbox": [ + 131, + 442, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\rho_{CNN}" + }, + { + "bbox": [ + 131, + 442, + 482, + 467 + ], + "type": "text", + "content": " for CNN Search space (e.g., ResNet, NAS-Bench) and " + }, + { + "bbox": [ + 131, + 442, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\rho_{ViT}" + }, + { + "bbox": [ + 131, + 442, + 482, + 467 + ], + "type": "text", + "content": " ViT search space as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 176, + 475, + 481, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 475, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 176, + 475, + 481, + 497 + ], + "type": "interline_equation", + "content": "\\rho_ {C N N} = \\mathcal {D} _ {K L} (\\operatorname {s i g m o i d} ([ \\mathcal {F} _ {S}, \\mathcal {F} _ {T} ])) + \\frac {1}{M} \\sum \\left| \\log \\operatorname {s o f t m a x} \\left(\\frac {\\partial L}{\\partial W}\\right) \\right| _ {F}, \\tag {4}", + "image_path": "8dd14adde9c6afc4226ea97294e03a86adf28f955903a17878103deeb5fe95db.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 188, + 506, + 481, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 506, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 188, + 506, + 481, + 530 + ], + "type": "interline_equation", + "content": "\\rho_ {V I T} = \\mathcal {D} _ {K L} (\\operatorname {m a s k} ([ \\mathcal {F} _ {S}, \\mathcal {F} _ {T} ])) + \\frac {M}{\\sum \\operatorname {s i g m o i d} \\left( \\right.\\left| \\right. \\log \\left| \\frac {\\partial L}{\\partial W} \\right|\\left. \\right)} \\tag {5}", + "image_path": "c0b72f1021f0bca313a9712238b15fcda1320dfad04a1292c078c5e4af6420c5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "text", + "content": " is the weight parameter, " + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{L} / \\partial W" + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "text", + "content": " is the corresponding gradient. " + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _F" + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "text", + "content": " means the Frobenius-norm. " + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "text", + "content": " denotes mean operation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 557, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 479, + 616 + ], + "type": "text", + "content": "Analysis of searched proxy. The discovered formulations indicate that (1) mathematical operations like softmax, logsoftmax, sigmoid, and mask operations benefit proxy predictability. (2) Gradient and features are vital input statistics for instinct and interaction proxies. As shown in Figure 5, the discovered proxies of Auto-DAS achieve significant ranking improvement." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 480, + 666 + ], + "type": "text", + "content": "Understanding why our Auto-DAS surpasses other NAS methods. (1) Many training-based distillation NAS methods use weight sharing and proxies, often leading to inaccurate performance estimates. Our Auto-DAS searches proxies based on real performance, yielding one with strong ranking ability. (2)" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "Haosen Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 157, + 421, + 278 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Spearman correlation (\\%) results on the NAS-Bench-101 and NAS-Bench-201 (NB-101/201) datasets. In this context, NB-101/201-KD denotes the task of distilling the accuracies of the architectures on the respective NAS-Bench-101/201 datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 157, + 421, + 278 + ], + "lines": [ + { + "bbox": [ + 192, + 157, + 421, + 278 + ], + "spans": [ + { + "bbox": [ + 192, + 157, + 421, + 278 + ], + "type": "table", + "html": "
MethodResNetNB-101NB-201NB-101-KDNB-201-KD
FLOPs72.9230.8163.3815.5664.55
Fisher81.37-38.8135.91-33.924.45
Grad_Norm82.35-39.2358.70-39.16-10.01
SNIP85.07-29.0158.17-21.7816.91
Synflow88.3043.6974.6120.3674.63
NWOT45.6632.8464.4122.9735.27
DisWOT(Ms)77.2449.6165.7450.1653.88
DisWOT(Mr)49.3830.7456.4642.9445.27
DisWOT91.3846.5772.3652.4564.90
Auto-DAS98.4471.8487.7385.9574.66
", + "image_path": "9d0bae625201bc26f364ce6eeb17005bc7c6fe8adc139ddd0618e11cca59faea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 396 + ], + "type": "text", + "content": "Other train-free NAS methods primarily target vanilla training, not distillation. They rely on weak correlation proxies. Ours exhibits a stronger correlation, ensuring more efficient model discovery. While DisWOT considers distillation in its proxy design, it is manually crafted and lacks direct optimization and feedback from distillation results. In contrast, our Auto-DAS method not only includes all distillation designs in the proxy search space settings but also optimizes our proxies directly based on the distillation results. Our approach naturally outperforms other methods in terms of distillation performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 411, + 307, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 307, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 307, + 424 + ], + "type": "text", + "content": "3.5 Training-free Student Search" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "content": "The training-free student search enables us to efficiently explore numerous candidate architectures without the need for costly and time-consuming training. After obtaining an effective proxy with the aid of an evolutionary search, we employ it to conduct a training-free student search. Utilizing randomly initialized weights, denoted as " + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "content": ", we seek the optimal student, represented as " + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\alpha^{*}" + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "content": ", from the search space " + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 508, + 481, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 508, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 242, + 508, + 481, + 523 + ], + "type": "interline_equation", + "content": "\\alpha^ {*} = \\arg \\max _ {\\alpha \\in \\mathcal {A}} \\rho_ {A u t o - D A S} ^ {*} (\\alpha , \\mathcal {W}). \\tag {6}", + "image_path": "6ef3343780ed0e48ec01a71f0050b6f4fd6acf36f078bd56c561939a2a6d7dfa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 529, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 566 + ], + "type": "text", + "content": "After the search phase, we use the pre-trained teacher model to distill the optimal student network " + }, + { + "bbox": [ + 130, + 529, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\alpha^{*}" + }, + { + "bbox": [ + 130, + 529, + 482, + 566 + ], + "type": "text", + "content": ". We adopt the original knowledge distillation method [17] as the auxiliary loss function." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 582, + 279, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 279, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 279, + 596 + ], + "type": "text", + "content": "4 Experimental results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this section, we present the ranking correlation and search accuracy results of our Auto-DAS on CNN, NAS-Bench, and ViT search space. For fair comparisons, we employ the same search, distillation settings, KD accuracy ground truths in DisWOT as the ground truth for correlation validation. We report mean results based on more than 3 repeated trials." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "Auto-DAS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 146, + 454, + 193 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 290, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 290, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 290, + 137 + ], + "type": "text", + "content": "Table 3: Accuracy " + }, + { + "bbox": [ + 132, + 114, + 290, + 137 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 132, + 114, + 290, + 137 + ], + "type": "text", + "content": " of NDS-ResNet space on CIFAR-100." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 146, + 454, + 193 + ], + "lines": [ + { + "bbox": [ + 141, + 146, + 454, + 193 + ], + "spans": [ + { + "bbox": [ + 141, + 146, + 454, + 193 + ], + "type": "table", + "html": "
ConstraintsNWOTSynflowDisWOTAuto-DAS
5M-FLOPs63.1964.2865.9866.52
100M-FLOPs70.3872.1272.8973.25
0.5M-Param.70.3871.5872.8973.32
1M-Param.72.5773.5674.2374.66
MethodFitNetsATSPRKDCRD
Random NAS70.1270.1670.4671.1971.59
DisWOT74.8574.574.9574.6275.25
Auto-GAS75.2874.9875.5575.3175.86
", + "image_path": "f029a75cf83afc897dd7eba7eb3d92b26312ec16d63401144655cd6fcc6e15fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 229, + 477, + 265 + ], + "blocks": [ + { + "bbox": [ + 294, + 114, + 452, + 137 + ], + "lines": [ + { + "bbox": [ + 294, + 114, + 452, + 137 + ], + "spans": [ + { + "bbox": [ + 294, + 114, + 452, + 137 + ], + "type": "text", + "content": "Table 4: Results " + }, + { + "bbox": [ + 294, + 114, + 452, + 137 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 294, + 114, + 452, + 137 + ], + "type": "text", + "content": " of other KD methods under 1M parameters." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 208, + 412, + 219 + ], + "lines": [ + { + "bbox": [ + 200, + 208, + 412, + 219 + ], + "spans": [ + { + "bbox": [ + 200, + 208, + 412, + 219 + ], + "type": "text", + "content": "Table 5: Top-1 accuracy (%) results on ImageNet." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 229, + 477, + 265 + ], + "lines": [ + { + "bbox": [ + 136, + 229, + 477, + 265 + ], + "spans": [ + { + "bbox": [ + 136, + 229, + 477, + 265 + ], + "type": "table", + "html": "
ModelsBaselineBaseline+WSLDBaseline+DISTRandomNAS+KDZen-NAS+KDDisWOTAuto-DAS
ResNet1869.772.0472.0771.6871.8872.0872.58
ResNet5077.1NA78.4577.4877.8078.6279.25
", + "image_path": "7a020f43c9e0a26fbe85647328ccdb77be33fdc5a972c75c47b37d0ae6b19659.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 309, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 309, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 309, + 304 + ], + "type": "text", + "content": "4.1 Experiments on CNN models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 325, + 481, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 443 + ], + "type": "text", + "content": "Architecture space and methodology. For CNN validation, we focus on ResNet variants, encompassing both CIFar-ResNet and NDS-ResNet domains. Our CIFar-ResNet space, inspired by He et al. [15], comprises three phases with 1,3,5,7 block options, tailored for CIFAR-100. We also explore the NDS-ResNet realm for both CIFAR-100 and ImageNet-1k search trials. Across all models, we employ our discovered " + }, + { + "bbox": [ + 130, + 325, + 481, + 443 + ], + "type": "inline_equation", + "content": "\\rho_{CNN}" + }, + { + "bbox": [ + 130, + 325, + 481, + 443 + ], + "type": "text", + "content": " without additional proxy exploration. During knowledge transfer, all identified student networks undergo training using CRD protocols [48], with ResNet56 serving as the mentor model. For ImageNet-1k, we investigate ResNet18/50-caliber students under equivalent parameter restrictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": "Findings on compact datasets. Our ranking analysis is rooted in individual knowledge transfer outcomes within the cifar-ResNet domain, as provided by DisWOT. Table 2 demonstrates that Auto-DAS achieves a notably higher Spearman coefficient compared to Fisher, SNIP, FLOPs, and NWOT. Our exploration experiments, conducted in the NDS-ResNet realm, pit our approach against Synflow, NWOT, and DisWOT, under 0.5M and 1M parameter limits, as well as 50M and 100M FLOPs thresholds. As illustrated in Table 3, Auto-DAS surpasses previous state-of-the-art methods by " + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "inline_equation", + "content": "0.9\\% \\sim 2.0\\% \\uparrow" + }, + { + "bbox": [ + 130, + 448, + 482, + 579 + ], + "type": "text", + "content": " under identical constraints. Furthermore, we evaluate Auto-DAS across various KD techniques (refer to Table 4). The results affirm our method's efficacy, consistently outperforming DisWOT across different KD approaches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "ImageNet performance analysis. Table 5 presents the effectiveness of student models identified by Auto-DAS on ImageNet, with ResNet34/101 acting as teacher networks. For the ResNet18 variant, Auto-DAS attains " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "72.58\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " accuracy, outperforming both DisWOT and other KD+NAS methodologies. Similarly, with the ResNet50 model, Auto-DAS achieves a leading accuracy of " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "79.25\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ". These findings underscore the superior capability of our Auto-DAS approach in enhancing model precision across various architectures." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "text", + "content": "Haosen Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 190, + 479, + 307 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": "Table 6: Evaluation of Architectures on CIFAR-10, CIFAR-100, and ImageNet-16 in NAS-Bench-201 [11]. Distilled Accuracy (\\%) represents the classification accuracy of the discovered architecture after distillation training. Search Time (s) denotes the computational cost (in GPU-seconds) during the search phase. The performances of NWOT and TE-NAS are obtained from their respective publications. Our proposed Auto-DASapproach demonstrates superior accuracy with rapid search speed." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 190, + 479, + 307 + ], + "lines": [ + { + "bbox": [ + 135, + 190, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 479, + 307 + ], + "type": "table", + "html": "
TypeModelCIFAR-10CIFAR-100ImageNet-16-120
Dis. Acc(%)Time (s)Speed-upDis. Acc(%)Time (s)Speed-upDis. Acc(%)Time (s)Speed-up
Multi-trialRS93.63216K1.0×71.28460K1.0×44.881M1.0×
RL [3]92.83216K1.0×71.71460K1.0×44.351M1.0×
BOHB [13]93.49216K1.0×70.84460K1.0×44.331M1.0×
RSPS [25]91.6710K21.6×57.9946K21.6×36.87104K9.6×
One-shotGDAS [10]93.3922K12.0×70.7039K11.7×42.35130K7.7×
DARTS [35]89.2223K9.4×66.2480K5.8×43.18110K9.1×
Zero-shotNWOT [37]93.732.2K100×73.314.6K100×45.4310K100×
EZNAS [2]93.632.2K100×69.824.6K100×43.4710K100×
Zero-shot DASDisWOT93.551.2K180×74.219.2K180×47.3020K180×
DisWOT(Mr)93.490.72K300×73.6218.4K300×45.6340K300×
Auto-DAS93.961.8K180×74.7313.8K180×47.5030K180×
", + "image_path": "efb7b5ce7cb4c52748031166f76cef525f43ec9110097f3aa585b2a6600a9088.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 331, + 303, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 303, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 303, + 342 + ], + "type": "text", + "content": "4.2 Experiments on NAS-Bench" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 446 + ], + "type": "text", + "content": "Experimental setup and search space. The NAS-Bench-101 [53] and NAS-Bench-201 [11] datasets serve as crucial benchmarks for assessing the efficacy of various NAS algorithms. DisWOT offers knowledge distillation outcomes for architectures within the NAS-Bench-101/201 search spaces (referred to as NB-101-KD and NB-201-KD). Adhering to these benchmark protocols, we apply the identified " + }, + { + "bbox": [ + 130, + 350, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\rho_{CNN}" + }, + { + "bbox": [ + 130, + 350, + 482, + 446 + ], + "type": "text", + "content": " to model search and validate distillation rankings without additional proxy searches. Comprehensive implementation details are available in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 604 + ], + "type": "text", + "content": "Performance analysis. Table 2 illustrates that our Auto-DAS method achieves superior ranking performance, significantly outperforming DisWOT. Furthermore, we evaluate the standard correlation in NAS-Bench-101/201, with the ranking outcomes highlighting Auto-DAS's versatility through substantial improvements over existing techniques. Regarding search trials presented in Table 6, Auto-KD demonstrates remarkable gains in the accuracy-efficiency tradeoff compared to multi-trial NAS approaches (such as Random Search (RS) and Reinforcement Learning (RL)) and Gradient-based One-shot NAS methods (like GDAS and DARTS), which typically incur substantial search costs. Our method also consistently enhances distillation accuracy when compared to other Zero-shot NAS techniques (including NWOT, TE-NAS, and DisWOT). These exceptional ranking and search results underscore the robust generalization capabilities of our approach across challenging NAS-Bench datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 620, + 342, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 342, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 342, + 633 + ], + "type": "text", + "content": "4.3 Experiments on Vision Transformer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Architecture space and methodology. To assess the ranking consistency of zero-cost proxies for ViT structures, we develop a benchmark called ViT-Bench." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "Auto-DAS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 136, + 480, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 7: Results (%) of Ranking Correlation on CIFAR-100, Flowers, and Chaoyang." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 136, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 134, + 136, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 134, + 136, + 480, + 205 + ], + "type": "table", + "html": "
Search SpaceProxyCIFAR-100FlowersChaoyang
KendallSpearmanPearsonKendallSpearmanPearsonKendallSpearmanPearson
PiTGraSP-42.02-58.71-31.53-50.66-69.64-40.90-16.00-22.94-19.07
SynFlow69.7987.0570.8062.2279.9871.6630.9642.6639.24
TENAS-2.13-3.21-1.68-2.86-4.23-3.33-3.34-5.04-3.55
NWOT-2.61-4.13-0.522.673.690.714.736.874.43
TF-TAS63.8382.2058.3764.4882.9167.2337.9952.9242.68
DisWOT-19.05-28.83-28.56-19.57-30.33-18.01-3.362.4015.22
Auto-DAS72.81±1.6890.32±1.7977.76±2.5876.42±1.1293.09±1.8587.63±1.3648.36±2.4666.13±1.5866.58±1.86
", + "image_path": "69f58a1f8143f32790026ed48514caf5e5993ddda657533a833ac46e9bd118cd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 278, + 481, + 356 + ], + "blocks": [ + { + "bbox": [ + 132, + 223, + 481, + 267 + ], + "lines": [ + { + "bbox": [ + 132, + 223, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 223, + 481, + 267 + ], + "type": "text", + "content": "Table 8: Comparing the performance of our proposed Auto-DAS approach across three diverse datasets - CIFAR-100, flowers, and Chaoyang - using the PiT search space, we have achieved competitive results while incurring the lowest search cost, measured in time cost on a single GPU." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 278, + 481, + 356 + ], + "lines": [ + { + "bbox": [ + 135, + 278, + 481, + 356 + ], + "spans": [ + { + "bbox": [ + 135, + 278, + 481, + 356 + ], + "type": "table", + "html": "
Search SpaceProxyCIFAR-100FlowersChaoyang
Param(M)Dis.Acc(%)Search CostParam(M)Dis.Acc(%)Search CostParam(M)Dis.Acc(%)Search Cost
PiTRandom5.3375.84N/A4.8865.30N/A5.2482.94N/A
GraSP4.5376.031.24 h3.7266.581.85 h4.6383.870.86 h
SynFlow11.0577.131.08 h5.2368.120.99 h4.9383.730.70 h
TENAS6.9376.095.14 h4.2668.035.14 h6.7683.645.07 h
NWOT5.2176.643.02 h10.7767.723.09 h6.3783.313.08 h
TF-TAS16.0777.061.21 h10.3068.210.95 h4.3284.340.71 h
DisWOT10.3875.820.88 h8.7667.580.88 h5.8583.420.88 h
Auto-DAS12.4277.881.28 h10.5269.321.28 h4.8884.831.28 h
Auto-DAS (G-free)11.3577.720.82h10.8869.450.82h5.2584.980.82h
", + "image_path": "ab5640ead45ebb45bb8dee5145a82ded86849ddb8f55b116bea874e5c2947b80.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 386, + 481, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 541 + ], + "type": "text", + "content": "This benchmark provides ground-truth accuracy for ViTs on compact datasets (CIFAR-100 [22], Flowers [39], and Chaoyang [58]). Inspired by [24], which shows ViTs gain significantly on small datasets when distilled from an efficient CNN teacher, we concentrate on the distillation accuracy of student ViTs on these datasets with a predetermined teacher. We randomly sample ViTs from PiT [16] search spaces and train them individually using the distillation settings from [24]. ResNet56 serves as the CNN teacher. ViT-Bench encompasses 5k KD accuracy ground truths per dataset. We divide the benchmark into validation and test sets at a 4:3 ratio. For the proxy search phase, we configure EA parameters " + }, + { + "bbox": [ + 132, + 386, + 481, + 541 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}, \\mathcal{T}, r, k)" + }, + { + "bbox": [ + 132, + 386, + 481, + 541 + ], + "type": "text", + "content": " in Alg. 1 as " + }, + { + "bbox": [ + 132, + 386, + 481, + 541 + ], + "type": "inline_equation", + "content": "(20, 100, 0.9, 5)" + }, + { + "bbox": [ + 132, + 386, + 481, + 541 + ], + "type": "text", + "content": " and evaluate proxy candidates' ranking performance in the validation set each iteration. Post-search, we equitably compare the Auto-DAS proxy with conventional ones by random sampling in the test set. Our validation and test sets are distinct to ensure unbiased proxy comparisons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": "Findings on compact datasets. Table 7 presents the rank consistency results of existing proxies across three datasets. Our proposed Auto-DAS surpasses other notable proxies in metrics like Kendall's tau [1], Spearman's rho [44], and Pearson's correlation coefficient [4], showcasing its efficacy and superiority in achieving consistent and precise ViT architecture search. To further demonstrate Auto-DAS's effectiveness, Table 8 displays search trial results. Our approach achieves superior distillation outcomes and faster search speeds across all three datasets. These empirical findings highlight the importance of effective zero-cost proxies and the role of our proposed Auto-DAS in optimizing ViT architecture search performance." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "text", + "content": "Haosen Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 165, + 136, + 449, + 228 + ], + "blocks": [ + { + "bbox": [ + 193, + 114, + 420, + 126 + ], + "lines": [ + { + "bbox": [ + 193, + 114, + 420, + 126 + ], + "spans": [ + { + "bbox": [ + 193, + 114, + 420, + 126 + ], + "type": "text", + "content": "Table 9: Results of different ViT models on ImageNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 136, + 449, + 228 + ], + "lines": [ + { + "bbox": [ + 165, + 136, + 449, + 228 + ], + "spans": [ + { + "bbox": [ + 165, + 136, + 449, + 228 + ], + "type": "table", + "html": "
ModelMethodParam.Top-1 Acc.ModelMethodParam.Top-1 Acc.
DeiT-TinyBaseline+KD5.7M74.5%Swin-TinyBaseline+KD29.0M81.7%
TF-TAS+KD5.7M74.7%TF-TAS+KD29.2M82.0%
DisWOT5.8M75.3%DisWOT29.0M82.2%
Auto-DAS6.1M75.9%Auto-DAS29.2M82.9%
PiT-TinyBaseline+KD4.9M73.2%PiT-SmallBaseline+KD23.5M79.9%
TF-TAS+KD4.9M73.5%TF-TAS+KD24.2M80.1%
DisWOT4.6M73.6%DisWOT23.8M80.3%
Auto-DAS4.8M74.2%Auto-DAS24.5M81.0%
", + "image_path": "3b25f953580cd6df6711f0e52598a96b2b5953ebaabf22cd4e858659b5db14f9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 287, + 477, + 402 + ], + "blocks": [ + { + "bbox": [ + 132, + 255, + 481, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 255, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 481, + 277 + ], + "type": "text", + "content": "Table 10: Spearman correlation (\\%) on optimal proxy in our search space with vanilla inputs or operations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 287, + 477, + 402 + ], + "lines": [ + { + "bbox": [ + 135, + 287, + 477, + 402 + ], + "spans": [ + { + "bbox": [ + 135, + 287, + 477, + 402 + ], + "type": "table", + "html": "
MethodsSettingsResNetNB-101NB-201NB-101-KDNB-201-KDPiT-CIFAR-100PiT-FlowersPiT-Chaoyang
Auto-DASAuto-DAS 98.4471.8487.7385.9574.6690.3293.0966.13
Auto-DAS (instinct)Activations70.2865.8978.9666.3560.8574.8676.8552.86
Gradients77.5372.6886.9273.5566.3880.5682.7556.67
weights60.3552.7570.2558.6755.2560.8965.8241.25
Unary78.8574.2588.2577.4266.5382.8384.3858.69
Binary62.5258.8772.8964.3558.9665.4570.2548.46
All80.5275.8689.6278.5668.5583.2586.5260.12
Auto-DAS (interaction)Feature90.3547.9662.6876.3366.3881.9585.3959.68
Embed88.6546.5560.7574.8563.2880.2583.2756.88
Logits85.4742.5855.8670.5358.6276.8378.9652.45
Transform92.3548.8864.6678.9568.6587.2576.5561.35
Distance89.6346.5363.8575.8265.8682.6984.5658.75
All92.8250.8865.8880.5670.4885.4288.7363.38
", + "image_path": "8f2e562ae5d47ef094a7a364bbe3a90c16f8fbba00923adf7879b2ce9cd0eecf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 441, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 562 + ], + "type": "text", + "content": "Computational cost analysis. (1) Like most train-free NAS methods, Auto-DAS involves minor additional gradient computations compared to DisWOT. To address this, we explore a gradient-free version of Auto-DAS in Table 8 (labeled Auto-DAS (G-free)). Results indicate that our gradient-free Auto-DAS achieves better accuracy-efficiency trade-offs than DisWOT. (2) While DisWOT is efficient and yields good results on certain datasets, it lacks generalizability across various models and tasks. In contrast, Auto-DAS is automated, versatile, and effective, better suited diverse application needs. Moreover, given the low cost of training-free approaches, the small additional time investment in searching is justified by the consistent performance gains achieved." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "ImageNet performance analysis. Examining the results in Table 9, it's clear that Auto-DAS outperforms other methods across various models. For DeiT-Tiny, Auto-DAS achieves " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "75.9\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " accuracy, surpassing Baseline+KD " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "(74.5\\%)" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", TF-TAS+KD " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "(74.7\\%)" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", and DisWOT " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "(75.3\\%)" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". Similarly, Auto-DAS excels with the Swin-Tiny model. This trend persists for PiT-Tiny and PiT-Small models, where Auto-DAS consistently outperforms other methods. These findings underscore the effectiveness and adaptability of our Auto-DAS method in enhancing accuracy across different ViT architectures." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "Auto-DAS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 113, + 250, + 199 + ], + "blocks": [ + { + "bbox": [ + 136, + 113, + 250, + 199 + ], + "lines": [ + { + "bbox": [ + 136, + 113, + 250, + 199 + ], + "spans": [ + { + "bbox": [ + 136, + 113, + 250, + 199 + ], + "type": "image", + "image_path": "9736a737a53a422dcaa4f5de4cf7ff6a38e21b12d826d6d6a6b047362536a128.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 207, + 479, + 231 + ], + "lines": [ + { + "bbox": [ + 131, + 207, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 207, + 479, + 231 + ], + "type": "text", + "content": "Figure 6: Correlation visualization for Auto-DAS on NAS-Bench-101, NAS-Bench-201 and ViT-Bench (from left to right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 251, + 113, + 363, + 198 + ], + "blocks": [ + { + "bbox": [ + 251, + 113, + 363, + 198 + ], + "lines": [ + { + "bbox": [ + 251, + 113, + 363, + 198 + ], + "spans": [ + { + "bbox": [ + 251, + 113, + 363, + 198 + ], + "type": "image", + "image_path": "82aa422130857dbfa25b0857318cda29db1312bdd47caca9495564c368fe004a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 364, + 113, + 478, + 199 + ], + "blocks": [ + { + "bbox": [ + 364, + 113, + 478, + 199 + ], + "lines": [ + { + "bbox": [ + 364, + 113, + 478, + 199 + ], + "spans": [ + { + "bbox": [ + 364, + 113, + 478, + 199 + ], + "type": "image", + "image_path": "1da0ebc9194a28aff3c1d5b979a994f9d3d224093998cd2d6713dfebf1dd147c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 254, + 246, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 246, + 265 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 246, + 265 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 274, + 481, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 481, + 393 + ], + "type": "text", + "content": "Proxy search space. We conduct experiments in Table 10 to analyze different components in our proxy search space. The results show that Auto-DAS (interaction) outperforms Auto-DAS (instinct) overall. The feature input and transform operations are particularly important for Auto-DAS (interaction), while weight input and binary operations contribute weakly. On the other hand, gradient input and unary operations show promising results for Auto-DAS (instinct). Different ops have varying impacts on the proxies' accuracy, and combining them enhances the performance of Auto-DAS approaches. These findings highlight the significance of considering different ops and their interactions in the search space for optimal performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 393, + 481, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 481, + 430 + ], + "type": "text", + "content": "Search algorithm. We use the EA with adaptive-elite selection for proxy search. As shown in Figure 4, the EA adaptive-elite selection obtains better final search results than EA and random search in the proxy search process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 430, + 481, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 477 + ], + "type": "text", + "content": "Correlation visualization.. To intuitively observe the proxy's predictability, we visualize the score of Auto-DAS and the ground-truth performance in Figure 6. These visualizations demonstrate that Auto-DAS effectively detects the true distillation results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 496, + 220, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 496, + 220, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 496, + 220, + 509 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 616 + ], + "type": "text", + "content": "We present Auto-DAS, a novel training-free framework for distillation-aware student architecture search. It includes proxy search and training-free student search. With discovered proxies, Auto-DAS enables efficient search without training costs. This significantly outperforms handcrafted proxies in distillation performance. Comprehensive results on NAS benchmarks, ViT, and CNN spaces over multiple datasets illustrate our method's superior ranking and distillation abilities. We hope our novel investigations will give more insight and new directions for the knowledge distillation and NAS research communities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "**Limitation.** We mainly evaluate the same benchmarks as other NAS methods for fair comparison. We will continue to expand Auto-DAS for more tasks in future work. **Social impact.** Our Auto-DAS focuses on technology improvement without social and ethical implications." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "Haosen Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "text", + "content": "1. Abdi, H.: The Kendall rank correlation coefficient. Encyclopedia of measurement and statistics 2, 508-510 (2007)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 197 + ], + "type": "text", + "content": "2. Akhauri, Y., Munoz, J.P., Jain, N., Iyer, R.: EZNAS: Evolving zero-cost proxies for neural architecture scoring. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) NeurIPS (2022), https://openreview.net/forum?id=1SqaDG4dvdt" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "text", + "content": "3. Baker, B., Gupta, O., Naik, N., Raskar, R.: Designing neural network architectures using reinforcement learning. In: ICLR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 220, + 480, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 242 + ], + "type": "text", + "content": "4. Bowley, A.: The standard deviation of the correlation coefficient. Journal of the American Statistical Association 23(161), 31-34 (1928)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 308 + ], + "type": "text", + "content": "5. Brown, T.B., Mann, B., Subbiah, N.R.M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., M., D., Ziegler, Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. arXiv preprint, arXiv:2005.14165 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "6. Dong, P., Li, L., Tang, Z., Liu, X., Pan, X., Wang, Q., Chu, X.: Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. In: ICML (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 363 + ], + "type": "text", + "content": "7. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 397 + ], + "type": "text", + "content": "8. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV. pp. 17076-17086 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 480, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 420 + ], + "type": "text", + "content": "9. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 421, + 480, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 442 + ], + "type": "text", + "content": "0. Dong, X., Yang, Y.: Searching for a robust neural architecture in fourgpu hours. CVPR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 444, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 480, + 464 + ], + "type": "text", + "content": "1. Dong, X., Yang, Y.: Nas-bench-201: Extending the scope of reproducible neural architecture search. In: ICLR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 465, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 498 + ], + "type": "text", + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 498, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 519 + ], + "type": "text", + "content": "3. Falkner, S., Klein, A., Hutter, F.: Bohb: Robust and efficient hyperparameter optimization at scale. In: ICML (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "type": "text", + "content": "4. Gu, J., Tresp, V.: Search for better students to learn distilled knowledge. arXiv preprint arXiv:2001.11612 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 544, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 564 + ], + "type": "text", + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "text", + "content": "6. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "type": "text", + "content": "7. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "8. Hu, S., Xie, S., Zheng, H., Liu, C., Shi, J., Liu, X., Lin, D.: Dsnas: Direct neural architecture search without parameter retraining. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "9. Hu, Y., Liang, Y., Guo, Z., Wan, R., Zhang, X., Wei, Y., Gu, Q., Sun, J.: Angle-based search space shrinking for neural architecture search. In: ECCV (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Auto-DAS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "20. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 471, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 471, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 471, + 149 + ], + "type": "text", + "content": "21. Kim, Y., Rush, A.M.: Sequence-level knowledge distillation. In: EMNLP (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 149, + 458, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 458, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 458, + 159 + ], + "type": "text", + "content": "22. Krizhevsky, A.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "23. Lee, N., Ajanthan, T., Torr, P.: Snip: Single-shot network pruning based on connection sensitivity. In: ICLR (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "type": "text", + "content": "24. Li, K., Yu, R., Wang, Z., Yuan, L., Song, G., Chen, J.: Locality guidance for improving vision transformers on tiny datasets. In: European Conference on Computer Vision. pp. 110-127. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "text", + "content": "25. Li, L., Talwalkar, A.S.: Random search and reproducibility for neural architecture search. ArXiv (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "type": "text", + "content": "26. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "text", + "content": "27. Li, L., Bao, Y., Dong, P., Yang, C., Li, A., Luo, W., Liu, Q., Xue, W., Guo, Y.: Detkds: Knowledge distillation search for object detectors. In: ICML (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "text", + "content": "28. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "text", + "content": "29. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "30. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "text", + "content": "31. Li, L., Sun, H., Li, S., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Auto-gas: Automated proxy discovery for training-free generative architecture search. In: ECCV (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "32. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation. In: ICLR (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "33. Li, L., Wei, Z., Dong, P., Luo, W., Xue, W., Liu, Q., Guo, Y.: Attnzero: Efficient attention discovery for vision transformers. In: ECCV (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "34. Lin, M., Wang, P., Sun, Z., Chen, H., Sun, X., Qian, Q., Li, H., Jin, R.: Zen-nas: A zero-shot nas for high-performance image recognition. ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 440, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 471 + ], + "type": "text", + "content": "35. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "36. Liu, Y., Jia, X., Tan, M., Vemulapalli, R., Zhu, Y., Green, B., Wang, X.: Search to distill: Pearls are everywhere but not the eyes. In: CVPR (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "text", + "content": "37. Mellor, J., Turner, J., Storkey, A., Crowley, E.J.: Neural architecture search without training. In: ICML (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "type": "text", + "content": "38. Mirzadeh, S.I., Farajtabar, M., Li, A., Levine, N., Matsukawa, A., Ghasemzadeh, H.: Improved knowledge distillation via teacher assistant. In: AAAI (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "39. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "40. Park, W., Lu, Y., Cho, M., Kim, D.: Relational knowledge distillation. In: CVPR (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "41. Pham, H., Guan, M.Y., Zoph, B., Le, Q.V., Dean, J.: Efficient neural architecture search via parameter sharing. In: ICML. pp. 4092-4101 (2018)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "text", + "content": "42. Real, E., Liang, C., So, D.R., Le, Q.V.: Automl-zero: Evolving machine learning algorithms from scratch (2020)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "43. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "text", + "content": "Haosen Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 522 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "44. Stephanou, M., Varughese, M.: Sequential estimation of spearman rank correlation using hermite series estimators. Journal of Multivariate Analysis 186, 104783 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "45. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "type": "text", + "content": "46. Theis, L., Korshunova, I., Tejani, A., Huszár, F.: Faster gaze prediction with dense networks and fisher pruning. ArXiv abs/1801.05787 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 215 + ], + "type": "text", + "content": "47. Tian, Y., Krishnan, D., Isola, P.: Contrastive representation distillation. In: ICLR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 215, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 481, + 237 + ], + "type": "text", + "content": "48. Tian, Y., Krishnan, D., Isola, P.: Contrastive representation distillation. In: ICLR (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "type": "text", + "content": "49. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers amp; distillation through attention. In: ICML (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 293 + ], + "type": "text", + "content": "50. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 314 + ], + "type": "text", + "content": "51. Wang, T., Yuan, L., Zhang, X., Feng, J.: Distilling object detectors with fine-grained feature imitation. In: CVPR (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 314, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 481, + 336 + ], + "type": "text", + "content": "52. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 481, + 357 + ], + "type": "text", + "content": "53. Ying, C., Klein, A., Christiansen, E., Real, E., Murphy, K., Hutter, F.: Nas-bench-101: Towards reproducible neural architecture search. In: ICML (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 357, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 481, + 380 + ], + "type": "text", + "content": "54. You, S., Huang, T., Yang, M., Wang, F., Qian, C., Zhang, C.: Greedynas: Towards fast one-shot nas with greedy supernet. In: CVPR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 402 + ], + "type": "text", + "content": "55. Zhang, L., Ma, K.: Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In: ICLR (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 402, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 424 + ], + "type": "text", + "content": "56. Zhou, H., Song, L., Chen, J., Zhou, Y., Wang, G., Yuan, J., Zhang, Q.: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 445 + ], + "type": "text", + "content": "57. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "type": "text", + "content": "58. Zhu, C., Chen, W., Peng, T., Wang, Y., Jin, M.: Hard sample aware noise robust learning for histopathology image classification. TMI (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 500 + ], + "type": "text", + "content": "59. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 500, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 481, + 522 + ], + "type": "text", + "content": "60. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Auto-DAS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/d262e2a7-fb7b-49f5-8763-ff8884371602_content_list.json b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/d262e2a7-fb7b-49f5-8763-ff8884371602_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4a6289e4cfd9b6c972984f9f12b7930946ebe44a --- /dev/null +++ b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/d262e2a7-fb7b-49f5-8763-ff8884371602_content_list.json @@ -0,0 +1,1676 @@ +[ + { + "type": "text", + "text": "Auto-GAS: Automated Proxy Discovery for Training-free Generative Architecture Search", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lujun Li $^{1}$ , Haosen Sun $^{1}$ , Shiwen Li $^{2}$ , Peijie Dong $^{3}$ , Wenhan Luo $^{1}$ , Wei Xue $^{1}$ , Qifeng Liu $^{1\\boxtimes}$ , and Yike Guo $^{1\\boxtimes}$", + "bbox": [ + 223, + 214, + 779, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong University of Science and Technology", + "bbox": [ + 310, + 257, + 689, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Xidian University", + "bbox": [ + 433, + 272, + 566, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 The Hong Kong University of Science and Technology (Guangzhou) lilujunai@gmail.com;hsunas@connect.ust.hk;shiwenlisw@gmail.com; pdong212@connect.hkust-gz.edu.cn;whluo.china@gmail.com;weixue@ust.hk; liuqifeng@ust.hk;yikeguo@ust.hk", + "bbox": [ + 232, + 286, + 766, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract In this paper, we introduce Auto-GAS, the first training-free Generative Architecture Search (GAS) framework enabled by an autodiscovered proxy. Generative models like Generative Adversarial Networks (GANs) are now widely used in many real-time applications. Previous GAS methods use differentiable or evolutionary search to find optimal GAN generators for fast inference and memory efficiency. However, the high computational overhead of these training-based GAS techniques limits their adoption. To improve search efficiency, we explore training-free GAS but find existing zero-cost proxies designed for classification tasks underperform on generation benchmarks. To address this challenge, we develop a custom proxy search framework tailored for GAS tasks to enhance predictive power. Specifically, we construct an information-aware proxy that takes feature statistics as inputs and utilizes advanced transform, encoding, reduction, and augment operations to represent candidate proxies. Then, we employ an evolutionary algorithm to perform crossover and mutation on superior candidates within the population based on correlation evaluation. Finally, we perform generator search without training using the optimized proxy. Thus, Auto-GAS enables automated proxy discovery for GAS while significantly accelerating the search before training stage. Extensive experiments on image generation and image-to-image translation tasks demonstrate that Auto-GAS strikes superior accuracy-speed tradeoffs over state-of-the-art methods. Remarkably, Auto-GAS achieves competitive scores with $110 \\times$ faster search than GAN Compression. Code at: https://github.com/lliai/Auto-GAS.", + "bbox": [ + 259, + 411, + 743, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Generative Model $\\cdot$ Generative Architecture Search $\\cdot$ AutoML", + "bbox": [ + 259, + 773, + 743, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors.", + "bbox": [ + 230, + 825, + 410, + 839 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/13997cbe8f92b03bf62c64e1aa614da74dc4de13312dd10e9698ecc61bf1718d.jpg", + "table_caption": [ + "Table 1: Comparison of search cost for generative architecture search methods." + ], + "table_footnote": [], + "table_body": "
MethodAlgorithm Cost (GPU-days)MethodAlgorithm Cost (GPU-days)
AGAN [65]RL1200DGGAN [42]Heuristic580
AutoGAN [18]RL2EAS-GAN [40]EA1
AlphaGAN [59]Gradient1.2EGAN [64]EA1.3
EAGAN [70]EA0.8Auto-GAS (ours)Training-free0.09
", + "bbox": [ + 222, + 170, + 779, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 273, + 375, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative models like Generative Adversarial Networks (GANs) [19] have made tremendous strides in visual generation tasks such as image and video synthesis [5,50]. However, their demanding computational footprint presents barriers to deploying these models on edge devices with strict resource constraints [25,47]. To overcome this, there is growing interest in model compression techniques [7, 12, 14, 49, 67]. Among them, Neural Architecture Search (NAS) [16] automates architecture design, finding compact generative models for edge computing, and enabling flexible structures without pre-trained weights. In contrast, other compression methods are limited to fixed, handcrafted architectures and rely on pre-trained models. The integration of NAS in generative models enables real-time visual creation on edge devices, unlocking new interactive experiences in domains like augmented reality [5,24,50].", + "bbox": [ + 212, + 308, + 785, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conventional Generative Architecture Search (GAS) approaches typically rely on expensive training-based training routes to assess candidate architectures, such as maintaining weight-sharing supernets [13, 22] or conducting multiple training trials from scratch [56, 76]. However, training generative models presents unique obstacles that greatly increase the costs and difficulties of applying these techniques for GAS. The training processes for generative models are generally more complex than discriminative tasks because inherent instabilities (e.g., mode collapse [20]) can occur during optimization. In mode collapse, the model generates only a limited variety of examples from the underlying distribution. These instabilities make it challenging to reliably evaluate architectural candidates with standard NAS training-based evaluations. As shown in Table 1, gradient-based GAS methods like AdversarialNAS [17] and AlphaGAN [59] amplify these problems by jointly searching the generator and discriminator architecture spaces. This coupled search increases the overall complexity and fragility of the optimization procedure. To address these difficulties, DGGAN [42] introduced a progressive search approach where candidates are trained incrementally. However, its process still requires a massive 580 GPU days to complete due to the computationally intensive nature of generative model training.", + "bbox": [ + 212, + 491, + 785, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, training-free NAS approaches have demonstrated remarkable efficiency using zero-cost proxies to predict network performance without training network parameters. These methods are mainly present for encoder-only models for classification tasks. Inspired by their computational efficiency, we comprehensively evaluate existing proxies when applied to generative tasks. As shown", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9e5ba8747072f801b0fc63fbbb0149f87a792a36743ff05d30f7ef192733802b.jpg", + "image_caption": [ + "Figure 1: Correlation visualization of proxies (from left to right: Param., Synflow, NWOT, and Auto-GAS) in Generation benchmarks [15]." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 362, + 237 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/15e402b01e4469643860bc640273e7639ca896db14da9a9f94e5b5f04ad58517.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 145, + 504, + 237 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/80a0b71a52514bc63f8fafad1f12f208c2b8be34a69383e654b60e369ba31a3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 146, + 643, + 237 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2bd49668b642029ed754f2efa63851b39b2931d4e1271f953417e0de625ba727.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 146, + 784, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in Figure 1, we can make some interesting observations: (1) We observe that parameter-level proxies (e.g., Synflow [57], SNIP [28], and GraSP [62]), which estimate performance based on weight importance, generally perform poorer when ranking generator candidates. This indicates their assumptions regarding how architecture impacts learning do not directly translate to generative objectives and training dynamics. (2) Architecture-level proxies (e.g., NWOT [45] and Zen [39]) featuring topological encodings demonstrate stronger ranking ability. However, even the best architecture proxies leave room for improvement to enable performing generative NAS at scale. These findings motivate developing customized proxies that more precisely capture the unique characteristics of visual generation. Thus, a question naturally arises: how to design customized proxies for generation tasks?", + "bbox": [ + 212, + 309, + 784, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To answer this question, we present Auto-GAS, the first automated generative architecture search framework that aims to efficiently utilize custom training-free proxies. Firstly, we derive that the mutual information of latent features within generative models is strongly coupled with task optimization objectives based on information bottleneck theory [3,54,61]. Then, we develop proxy search strategies to better estimate the mutual information of these features. Specifically, we build comprehensive operation search spaces and evolutionary algorithms to search optimal information-aware proxies. To better process feature statistics, our search space involves transform-based, encoding-based, reduction-based, and augment-based operators. The search process in Auto-GAS begins with initializing the population of proxies and then evaluates, crosses, and varies them to improve the quality of the proxies. Throughout the search, we directly utilize the ranking correlation with accuracy results in the benchmark as fitting objectives, aiming to find a more suitable proxy for the generation tasks. With this evolutionary process, Auto-GAS outperforms existing training-free NAS approaches by significant margins without manual tuning. Additionally, we search for generator architectures using Auto-GAS and subsequently implement a full training process on the searched generator.", + "bbox": [ + 212, + 491, + 785, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We perform extensive experiments to validate the performance and efficiency strengths of our Auto-GAS in generation tasks like image generation and imaged-to-image translation. The experiments demonstrate that our Auto-GAS can achieve better-generating performance than other zero-shot proxies and state-of-the-art results on rank consistency. In contrast to handcrafted GAS methods,", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Automated Proxy Discovery for Training-free Generative Architecture Search", + "bbox": [ + 217, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "our Auto-GAS framework enjoys many benefits: (1) Automatic. It enables the automatic discovery of more expressive and efficient proxies that human expert might ignore, which can reduce human bias and ensures that the resulting architectures are optimized for the target problem or dataset. (2) General. The search procedure of Auto-GAS improves generalization by discovering proxies that generalize well on different search spaces. Based on the final results for the search objectives, our framework can evolve adaptive proxies to generalize across architectures. (3) Insightful. Our framework allows for the automatic discovery of more expressive, efficient, and effective proxies for optimizing and customizing generation architectures, opening up new research directions for automated GAN architecture search. Our contributions can be summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present Auto-GAS, the first training-free architecture search framework for generation tasks. Auto-GAS achieves an ultra-efficient and elegant search process with our auto-searched customized proxy.", + "- We propose a well-organized proxy search space, including latent features as input and various transforms, encoding, reduction, and augment operations as options. We develop evolutionary algorithms for effective & efficient proxy search.", + "- We conduct extensive experiments on image generation and image-to-image translation tasks. Auto-GAS obtains significant search acceleration and competitive performance on multiple datasets." + ], + "bbox": [ + 225, + 320, + 782, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 492, + 387, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generative architecture search. The field of Neural Architecture Search (NAS) traditionally encompasses the design of exploration spaces, search methodologies, and assessment techniques to autonomously uncover optimal architectural configurations within specified parameters. Empirical approaches utilize an iterative process of training followed by exploration, either through multiple iterations [4,73] or by employing weight-sharing strategies [41,72]. NAS has also been extensively applied to the domain of GANs Architecture Search (GAS). The primary goal in GAS is to automatically identify superior GAN structures that enhance generative performance. Pioneering reinforcement learning-based GAS techniques, such as AGAN [65] and AutoGAN [18], leverage RNN controllers to construct GAN architectures and employ the Inception Score (IS) as their evaluation criterion. E2GAN [60] and AdversarialNAS [17] implement alternative reward functions. EAGAN [70] conducts discrete searches and enhances the robustness of the GAN exploration process. Nevertheless, these training-dependent search methodologies continue to be hampered by inefficient exploration and substantial computational demands. To tackle this challenge, we propose a training-free approach, grounded in architecture decoupling, to enhance the discovery of competitive GAN structures in a more computationally efficient manner.", + "bbox": [ + 212, + 522, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training-free architecture search. Zero-shot architecture exploration [10, 11, 29, 39], alternatively referred to as training-free NAS, offers a swifter alternative to conventional NAS methods. This approach's evaluation process en", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tails only a handful of forward or backward passes with a small batch of input data, rendering it highly efficient. Zero-cost indicators can be categorized into parameter-centric and architecture-centric proxies. Parameter-centric zero-cost proxies (e.g., Synflow [57], SNIP [28], and GraSP [62]) rely on pruning techniques and aggregate the saliency values of each layer's weights as a proxy. For instance, Synflow [58] introduces a modified version of synaptic saliency scores to prevent layer collapse during parameter pruning, while Fisher [1] calculates the sum of all activation gradients in the network, applicable to channel pruning. Grasp [63] ensures gradient flow preservation throughout the network for efficient training. Their detailed formulations are as follows:", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {s n i p} = \\left| \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W} \\right|, \\rho_ {s y n f l o w} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W}, \\rho_ {f i s h e r} = \\left(\\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {A}} \\mathcal {A}\\right) ^ {2}, \\rho_ {g r a s p} = - (H \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}}) \\odot \\mathcal {W}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 306, + 789, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ , $\\mathcal{W}$ , $\\mathcal{A}$ are loss function, weight and activation. $H$ is the Hessian matrix. However, the complexity of loss function $\\mathcal{L}$ in generation tasks renders these gradient-based methods ineffective. Regarding architecture-centric proxies, ZenNAS [39] introduced the novel Zen-Score to assess network expressivity, while NWOT [45] examined correlations between linear maps induced by data points for untrained architectures. Nevertheless, these approaches employ hand-crafted fixed proxies for classification tasks and encoder-only CNN models. In generation tasks, the encoder-decoder generator typically features intricate branches and multi-scale features, posing significant challenges for traditional proxies in predicting final accuracy. To address this issue, we automatically optimize the first customized proxies for GAS. Furthermore, our framework differs from EZNAS-like methods [2, 36, 74, 75] in several aspects: (1) Tasks. EZNAS is exclusively designed for classification, whereas our Auto-GAS is tailored for generation tasks. (2) Search algorithms. EZNAS merely employs genetic programming, while we propose a proxy evolution search strategy that substantially enhances search efficiency. (3) Model statistic inputs. Our method solely utilizes features as input, as we observed poor performance of parameter-level proxies. Conversely, EZNAS selects parameters and gradients as inputs. (4) Search space. Our Auto-GAS incorporates numerous feature-related operations absent in EZNAS.", + "bbox": [ + 212, + 349, + 787, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 657, + 380, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first review existing proxies. Then, we present formulations and insights of our proxy design (see Figure 2). Finally, we illustrate the detailed search process.", + "bbox": [ + 212, + 689, + 787, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Proxy Search Space", + "text_level": 1, + "bbox": [ + 215, + 755, + 426, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivation of our customized proxy design. Generative tasks are relevant to the Information Bottleneck (IB) theory [3, 54, 61], which guides generative models to learn compact representations that retain important characteristics for accurate sample generation. In generative tasks, the IB theory helps understand", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Automated Proxy Discovery for Training-free Generative Architecture Search", + "bbox": [ + 215, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3412afae11f3dd45eb17afde1ccc179e4ae3df5c2d14571ee9aa87a9ac4c6ef6.jpg", + "image_caption": [ + "Figure 2: Overall proxy search of Auto-GAS. We extract the latent features of the generator with input data forward-propagation. Then we select various transform, encoding, reduction, and augment operators from the proxy search space to build candidate proxies. Finally, we perform a ranking evaluation to cross and mutate to get the best proxy." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "how generative models can balance compression (capturing essential features) and reconstruction (faithful generation of samples) [55, 68]. IB theory suggests maximizing the Mutual Information (MI) $I(\\mathbf{X}; \\mathbf{F}_i)$ between the input data $\\mathbf{X}$ and the corresponding latent representation $\\mathbf{F}_i$ can be expressed as:", + "bbox": [ + 212, + 411, + 787, + 472 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI (\\mathbf {X}; \\mathbf {F} _ {i}) = H (\\mathbf {F} _ {i}) - H (\\mathbf {F} _ {i} | \\mathbf {X}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 488, + 785, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $H(\\mathbf{F}_i)$ stands for the entropy of the latent representation $\\mathbf{F}_i$ and $H(\\mathbf{F}_i|\\mathbf{X})$ denotes the conditional entropy. Considering that neural networks are deterministic, and the dataset has a finite number of instances, the conditional entropy becomes zero, simplifying the mutual information (MI) estimation as $I(\\mathbf{X};\\mathbf{F}_i) = H(\\mathbf{F}_i)$ . Consequently, some information coding and compression studies [26, 55] propose some operations (e.g., Entropy and Gram-matrix) to approximate the mutual information on latent representation. These inspired us to develop our train-free method design thought chain: assess generative model $\\rightarrow$ estimate MI $\\rightarrow$ design expressions via Entropy/Gram-matrix ops $\\rightarrow$ our train-free proxy. However, how about designing this potential expressions? Recent Auto-Zero approaches [30, 51] present a promising solution: searching for the best machine learning solution from scratch. In addition, several studies [6] have shown that normalize and activation operations can enhance the formulation. Therefore, we automatically search for optimal proxies based on these operators.", + "bbox": [ + 212, + 520, + 787, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proxy primitive operations. Our approach evolves the customized proxy in the generation benchmark [15], which consists of U-Net-like encoder-decoder models with various operators and their training results. We extract the latent features of each layer in both the encoder and decoder modules as inputs to the search process and represent the zero-cost proxy as a computation graph. This graph is constructed as transform $\\rightarrow$ encoding $\\rightarrow$ reduction $\\rightarrow$ augment", + "bbox": [ + 212, + 750, + 789, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "parts in processing generator statistics. The available primitive operations can be summarized as follows:", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Transform operations: These operations involve different activation functions and feature processing techniques to enhance raw features and represent architectural potentials. Examples of transform operations include relu, swish, mish, scale,norm, exp, and softmax.", + "- Encoding operations: Inspired by coding algorithms and feature information estimation, we employ binary and matrix-based operations in the encoding part. These operators help us better estimate the mutual correlation information of samples. Some encoding operations include sign, above-zero, above-median, gram-matrix, person, and hamming.", + "- Reduction operations: Reduction operators bridge the information matrix and the proxy values. They are valuable for extracting useful matrix information from different mathematical perspectives. Reduction operations include sloqdet, abslog, trace, and eigenvalue.", + "- Augment operations: Augment operators are applied to enhance the proxy values' ability to predict performance results. These operations include arithmetic functions such as abs, pow, l2-norm, log, sum, and mean." + ], + "bbox": [ + 230, + 189, + 784, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By utilizing these primitive operations, our approach enables the evolution of the customized proxy, facilitating the search for optimal solutions in the generation benchmark. These mathematical operations are essential building blocks to construct a diverse and expressive search space of zero-cost proxy. We provide more details of their formulas and properties in the supplementary material.", + "bbox": [ + 212, + 443, + 784, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Proxy Evolution Search", + "text_level": 1, + "bbox": [ + 214, + 541, + 455, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evolutionary process and goal. Utilizing our proxy search space and graph representation, we employ an evolutionary algorithm (EA) outlined in Alg. 1 to discover optimal proxy expressions. Our EA commences with an initial set of candidate proxies and progressively refines the population across generations, employing genetic operators such as selection, crossover, and mutation to yield superior solutions. For fitness evaluation, each proxy is assessed based on the ranking correlation between its output proxy $\\mathcal{Q}$ scores and pre-determined performance benchmarks $\\mathcal{D}$ , enabling efficient identification of the optimal proxy $\\rho^{*}$ within the search space $\\mathcal{P}$ , as follows:", + "bbox": [ + 212, + 566, + 784, + 702 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {A u t o - G A S} ^ {*} = \\underset {\\rho \\in \\mathcal {P}} {\\arg \\max } (\\tau (\\mathcal {D}, \\mathcal {Q})). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 715, + 784, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where Kendall's Tau $\\tau$ is formulated as the correlation coefficient. Each evolution picks top-k candidates with the highest evaluation score and randomly selects a parent from the candidates for mutation.", + "bbox": [ + 212, + 750, + 782, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficient proxy search with existing generation benchmark. It is expensive to derive pre-trained results $\\mathcal{D}$ , we adopt an alternative solution to achieve proxy search using pre-computed architecture-performance pairs provided in", + "bbox": [ + 212, + 795, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Automated Proxy Discovery for Training-free Generative Architecture Search", + "bbox": [ + 215, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Evolutionary Search for Auto-GAS Proxy" + ], + "code_body": "Input: Exploration domain $S$ , candidate set $\\mathcal{P}$ , maximum iterations $\\mathcal{T}$ , sampling ratio $r$ , sampled pool $\\mathcal{R}$ , top $k$ selections. \nOutput: Auto-GAS proxy with highest Kendall's Tau $\\tau$ coefficient. \n1: $\\mathcal{P}0 := \\text{Initialize candidate set}(P_i)$ ; \n2: Sampled pool $\\mathcal{R} := \\emptyset$ ; \n3: for $i = 1, 2, \\ldots, T$ do \n4: Select $G_i^s := \\text{SelectTopk}(\\mathcal{R}, k)$ ; \n5: // greedy evolutionary strategy \n6: Empty sampled pool $\\mathcal{R} := \\emptyset$ ; \n7: Mutate := MUTATE( $G_i^s$ ); \n8: Crossover $G_i^c := \\text{CROSSOVER}(\\mathcal{G}_i^s)$ ; \n9: Incorporate $G_i^m$ into $R$ ; \n10: Incorporate $G_i^c$ into $R$ ; \n11: if $\\mathcal{R} < \\mathcal{P}$ then \n12: Introduce random samples \n13: else \n14: Proceed to line 4; \n15: end if \n16: end for", + "bbox": [ + 215, + 162, + 785, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the existing TransNAS-Bench-101 [15] benchmarks. Specifically, this benchmark involves Autoencoding task with 7k different generators in macro-level search space and their evaluation metric (i.e., SSIM). Each generator network follows an encoder-decoder structure in Pix2Pix, where the encoders are the searched backbones, and the decoders contain 14 layers of convolution and deconvolution. Thanks to this existing benchmark, we do not need to train performance groundtruths again, and it only takes 1.5 hours in our proxy search process. In light of excellent generality (proven in our extensive experiments), our searched proxy $\\rho_{Auto-GAS}$ can be well transferred to similar search spaces and can be directly used in most scenarios without additional search budget.", + "bbox": [ + 212, + 472, + 787, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Discovered Proxy Analysis", + "text_level": 1, + "bbox": [ + 215, + 645, + 483, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Searched Zero-cost Proxy. We present formulas of the searched proxies on the generation benchmark as follows:", + "bbox": [ + 212, + 669, + 785, + 700 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {A u t o - G A S} ^ {*} = \\left(\\operatorname {s l o g d e t} \\left([ \\mathbf {F} ^ {\\prime}, \\mathbf {F} ^ {\\prime T} ]\\right)\\right) ^ {2}, \\mathbf {F} ^ {\\prime} = \\operatorname {a b o v e - m e a n} (\\operatorname {s w i s h} (\\mathbf {F})) \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 708, + 785, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{F}$ represents the original features, and transformed features $\\mathbf{F}'$ are obtained by applying the swish operation followed by the above-mean operation to the original features $\\mathbf{F}$ . The $[\\mathbf{F}',\\mathbf{F}'^T]$ indicates the $N\\times N$ gram-matrix of the transformed features $\\mathbf{F}'$ . The proxy value $\\rho^{*}$ is calculated by taking the square of the slogdet of the concatenated gram-matrix $[\\mathbf{F}',\\mathbf{F}'^T]$ .", + "bbox": [ + 212, + 733, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of searched proxy. The discovered formulations indicate that (1) Mathematical operations, such as swish activation and slogdet, significantly con-", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa3176a361b7f39924f01568c4edbc04ecdfe2460484be07faeb0920e3902b75.jpg", + "table_caption": [ + "Table 2: Ranking results on TransNAS-Bench-101 [15]." + ], + "table_footnote": [], + "table_body": "
MethodKendallSpearmanPearsonMethodKendallSpearmanPearson
SNIP [1]-2.04%-3.55%-0.26%ZiCo [29]18.63%15.53%21.13%
EZNAS [2]23.30%25.91%32.03%Synflow [57]-50.91%-65.01%-49.17%
NWOT [45]51.28%63.94%48.88%Params.-12.97%-15.24%-5.80%
SNIP [28]30.81%23.31%29.34%FLOPs48.91%64.01%52.36%
Zen [39]-2.27%2.58%5.39%Auto-GAS61.14%79.32%81.25%
", + "bbox": [ + 230, + 171, + 772, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tribute to the predictability of the proxy. These operations enhance the proxy's ability to capture important features and characteristics of the generator. (2) Gram-Matrix and Mutual Information: The expression utilizes a sample-wise gram-matrix, which accurately estimates the latent mutual information of the generator. The gram-matrix captures the relationships and dependencies among the transformed features, providing valuable insights into the generator's behaviour. The discovered proxy of Auto-GAS, as illustrated in Figure 1, showcases a substantial improvement in ranking. This improvement suggests that the searched proxy effectively captures essential characteristics of the generator, leading to enhanced performance in the benchmark.", + "bbox": [ + 212, + 301, + 784, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Understanding why our Auto-GAS surpasses other NAS methods. (1) Many train-based NAS methods employ weight-sharing and other proxy settings, which often lead to inaccurate estimations of architecture performance. In contrast, our Auto-GAS method searches for proxies based on actual performance, resulting in a proxy that demonstrates strong ranking capability. (2) Other train-free NAS methods are primarily designed for classification rather than generative tasks. Consequently, these methods search for architectures using weak correlation proxies. In contrast, our proxy exhibits a stronger correlation, ensuring the discovery of more efficient models.", + "bbox": [ + 212, + 452, + 785, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Training-free Generative Models Search", + "text_level": 1, + "bbox": [ + 214, + 613, + 589, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Auto-GAS has made efforts to search generative Models in a non-training way to ease the instability of GAN training and improve search efficiency. In this stage, we directly search the generator in $\\rho_{Auto - GAS}$ . In each iteration, we generate $P$ candidate generators to make up the population $\\mathcal{A}$ , where all these candidates use the randomly initialized weights $W$ . Each candidate is characterized by both architecture and weight variables. In each iteration, we evaluate $P$ candidates without training. We employ the information-aware proxy to assign scores to all these candidates. The formulation of the search stage can be succinctly stated as follows:", + "bbox": [ + 212, + 642, + 787, + 775 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {A u t o - G A S} ^ {*} = \\arg \\max _ {\\alpha \\in \\mathcal {A}} \\rho_ {A u t o - G A S} ^ {*} (\\alpha , \\mathcal {W}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 780, + 784, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that in scenarios involving discriminator searches, we also use $\\rho_{Auto - GAS}$ for searching the discriminator architectures with real sample inputs.", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Automated Proxy Discovery for Training-free Generative Architecture Search", + "bbox": [ + 217, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3ed1a8b0ab2127f4c96e66386dea0cb44b5fd543a335036b8e195715172affa9.jpg", + "table_caption": [ + "Table 3: Comparison results on the CIFAR-10 and STL-10 datasets. We use FID (the lower the better) and the mIoU metric (the higher the better) for evaluations." + ], + "table_footnote": [], + "table_body": "
MethodAlgorithmGPU daysCIFAR-10 (32×32)STL-10 (48×48)
IS↑FID↓IS↑FID↓
WGAN-GP [20]7.8629.3--
SN-GAN [46]8.2221.709.1640.1
ProbGAN [21]Manual-7.7524.608.8746.74
Improv MMD GAN [66]8.2916.219.3437.63
BigGAN [5]9.2214.73--
AGAN [65]12008.2930.509.2352.70
AutoGAN [18]RL2.008.5512.429.1631.01
AlphaGAN [59]1.28.7015.56--
AdversarialNAS [17]1.007.8624.048.5238.85
EGAN [64]1.256.90---
EAS-GAN [40]EA1.007.4533.20-38.84
EAGAN [70]0.88.4112.839.6923.82
Fisher [1]0.128.0317.669.3729.82
Zen [39]0.118.0215.999.4528.66
SNIP [28]Training-free0.127.9616.679.5829.25
NWOT [45]0.098.0516.329.2327.46
Auto-GAS0.098.4412.219.7223.19
", + "bbox": [ + 274, + 184, + 723, + 395 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.5 Generative Models Training", + "text_level": 1, + "bbox": [ + 215, + 422, + 493, + 439 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "After the architecture searches, we train generative models with randomly selected data samples. Simultaneously, a batch of noise samples is generated from a noise distribution. The generator tries to generate data from random noise $z$ as $G(z)$ , which should be as close as possible to the real data distribution. The discriminator tries to distinguish the generated data from the real data. The value function $V(D,G)$ for a minimax game is defined as follows:", + "bbox": [ + 212, + 448, + 787, + 541 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {G} \\max _ {D} V (D, G) = \\mathbb {E} _ {\\mathbf {x} \\sim p _ {d a t a} (\\mathbf {x})} [ \\log D (\\mathbf {x}) ] \\\\ + \\mathbb {E} _ {\\mathbf {z} \\sim p _ {z} (\\mathbf {z})} [ \\log (1 - D (G (\\mathbf {z})) ] \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 561, + 784, + 606 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathbf{x} \\sim p_{data}(\\mathbf{x})$ denotes the distribution of real data, $\\mathbf{z} \\sim p_z(\\mathbf{z})$ denotes the input noise to the generator.", + "bbox": [ + 212, + 616, + 787, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 669, + 375, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we first perform proxy search experiments on the generation benchmark, and then we conduct architecture search experiments on image generation and image-to-image translation tasks by translocating the searched proxies. Finally, we provide a detailed analysis of our approach.", + "bbox": [ + 212, + 702, + 785, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experiments on Generation Benchmark", + "text_level": 1, + "bbox": [ + 215, + 784, + 589, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset and implementation. We adopt the search space and performance ground truth provided by the Autoencoding task in TransNAS-Bench-101 [15] as", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/712476a5a6d6d7106f04f3db0b10a8be41fead19bfe73ca06da7ca54ced3c628.jpg", + "image_caption": [ + "Figure 3: Searched model on CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 498, + 180 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4176af95007f0863eee020d9a23fc5e5e3b9b360d20a7e1a9e97d5e3081ee716.jpg", + "image_caption": [ + "Figure 4: Searched model on STL-10." + ], + "image_footnote": [], + "bbox": [ + 504, + 152, + 781, + 180 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the benchmark for searching and evaluating GAS proxies. We split the 4:3 ratio of the benchmark into the validation and test sets. During the proxy search phase, we set $(\\mathcal{P},\\mathcal{T},r,k)$ in Alg. 1 as (20,100,0.9,5) and evaluate with 50 architectures in the validation set. After the search, we fairly compare the Auto-GAS proxy with the traditional ones by randomly sampling 200 architectures in the test set. Note that our validation and test sets do not overlap to ensure fair proxy comparisons. Detailed implementations are in the Appendix.", + "bbox": [ + 212, + 236, + 784, + 342 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison results. Table 5 demonstrates the ranking capabilities of different methods on Generation Benchmark. Auto-GAS (ours) achieves the best rankings across all three measures: Kendall's Tau $(61.14\\%)$ , Spearman $(79.32\\%)$ , and Pearson $(81.25\\%)$ . These results indicate that Auto-GAS performed well in preserving the rankings compared to other methods. These findings highlight the effectiveness of Auto-GAS for generating meaningful and accurate rankings.", + "bbox": [ + 212, + 342, + 785, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Experiments on Image Generation", + "text_level": 1, + "bbox": [ + 215, + 454, + 545, + 469 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Datasets and implementation. We perform evaluations using CIFAR-10 [27] and STL-10 [8] datasets, aligning with other GAS methodologies [17, 18, 70]. CIFAR-10 comprises 50k training and 10k test $32 \\times 32$ images, while STL-10 contains 100,500 images at a higher $48 \\times 48$ resolution. We directly apply searched proxies to identify optimal architectures within EAGAN's [70] search space. The exploration process spans 150 epochs with a $P = 32$ individual population. Generator and discriminator batch sizes are 80 and 40, respectively. We utilize 16 randomly generated Gaussian noise images for proxy score computation to expedite evaluation. Post-search, we conduct full training of the best-scoring generator for 500 epochs. For evaluation, following previous GAS works [17, 18, 70], we generate 50,000 images to calculate IS and FID metrics. Figures 3 and 4 illustrate the discovered architectures for CIFAR-10 and STL-10.", + "bbox": [ + 212, + 477, + 787, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training-based NAS method comparison. We benchmark Auto-GAS against various training-based NAS approaches: BigGAN [5], E2GAN [44], Adversarial-NAS [17], and EAGAN [70]. As shown in Table 3, Auto-GAS demonstrates superior efficiency. For CIFAR-10, Auto-GAS achieves lower FID than BigGAN and comparable overall results to E2GAN, while outperforming AdversarialNAS in both FID and IS. On STL-10, Auto-GAS significantly outperforms E2GAN and AdversarialNAS in both metrics, rivaling EAGAN in IS and surpassing it in FID. In summary, Auto-GAS emerges as an efficient and competitive approach, either surpassing or matching several state-of-the-art training-based NAS methods across diverse evaluation metrics and datasets.", + "bbox": [ + 212, + 657, + 790, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zero-shot proxy comparison. Table 3 also juxtaposes the Auto-GAS proxy with four alternative zero-shot proxies. Fisher and SNIP, relying on gradient-", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Automated Proxy Discovery for Training-free Generative Architecture Search", + "bbox": [ + 217, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/72d08bd9a031a6004d57eccc91283f988d5ba46becb7da42c0514489d3f40eb5.jpg", + "table_caption": [ + "Table 4: Image translation evaluation on horse→zebra and Cityscapes datasets." + ], + "table_footnote": [], + "table_body": "
ModelDatasetMethod#ParametersGPU daysFID (↓)mIoU (↑)
ValueRatioValueRatioValueIncreaseValueDrop
CycleGANhorse→zebra(512×512)Original11.4M---61.53--
COEV [53]----96.1534.6-
GAN Comp. [38]0.34M33.3×11.601.00×71.8110.3-
Fast GAN Comp. [38]0.36M32.1×3.123.72×65.193.66-
Auto-GAS0.39M29.3×0.11105.45×63.662.13-
GauGANCityscapes(512×512)Original93.0M---57.60-62.18-
GAN Comp. [38]20.4M4.6×17.701.00×55.192.4161.220.96
Fast GAN Comp. [38]20.2M4.6×10.401.70×56.251.3561.171.01
Auto-GAS21.0M4.5×0.16110.63×56.041.5661.210.97
", + "bbox": [ + 220, + 170, + 784, + 316 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "based computations, yield similar scores. Zen-Score and NWOT exhibit positive correlations with model accuracy, outperforming gradient-based methods. These results validate the Auto-GAS proxy's efficacy, positioning it as a formidable contender among zero-shot proxy methods.", + "bbox": [ + 212, + 345, + 787, + 407 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Experiments on Image-to-image Translation.", + "text_level": 1, + "bbox": [ + 214, + 430, + 629, + 446 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Datasets and implementation. We comprehensively evaluate our approach on image translation tasks using high-resolution datasets, including [9] and horse $\\rightarrow$ zebra [71] at $512\\times 512$ resolutions. We apply the proxy for architecture exploration, adopting GAN Compression's [38] search space for fair comparisons, focusing on channel and layer number configurations. We evolve to compress Pix2Pix [23] on horse $\\rightarrow$ zebra and GauGAN [48] on cityscapes. The search process involves 500 iterations, followed by model retraining with a 0.0002 learning rate and 16 batch size, adhering to the original paper [23].", + "bbox": [ + 212, + 455, + 787, + 577 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "GAN compression method comparison. Table 4 demonstrates that Auto-GAS achieves the highest FID score of 63.66 for CycleGAN with a superior compression ratio $(29.3\\times)$ compared to alternatives. This indicates Auto-GAS effectively reduces model size while preserving image quality. For GauGAN, Auto-GAS exhibits a marginally higher FID score and a competitive mIoU score of 61.22.", + "bbox": [ + 212, + 578, + 787, + 666 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Analysis. Figure 5 illustrates Auto-GAS's proficiency in translation tasks. In contrast, FastGAN Compression not only yields a lower FID score but also fails to eliminate artifacts like brown discoloration in zebra models. Conversely, Auto-GAS-generated zebra images accurately capture authentic black-and-white zebra patterns.", + "bbox": [ + 212, + 667, + 787, + 744 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 768, + 390, + 784 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Search efficiency benefits. For image generation (see Table 3), our Auto-GAS method exhibits significantly superior search efficiency to training-based NAS GANs and has a definite advantage over other zero-cost proxies. The Fisher and", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b5866431c6aa2827cf85ca6c855adf63100b90989d9a48ff6cd067154a3d236d.jpg", + "image_caption": [ + "Figure 5: Visualization results on image translation tasks using high-resolution datasets, e.g., horse→zebra and Cityscapes with $512 \\times 512$ resolution." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 784, + 258 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ba38058d59aca62d1cd734bde25b95717dd5bcc762e1fd5c8fe97dc277805674.jpg", + "table_caption": [ + "Table 5: Ablations of our proxy search." + ], + "table_footnote": [], + "table_body": "
Exp. MethodKendallSpearmanFID↓@CIFAR-10FID↓@STL-10
IVanilla Gram-matrix24.28%35.75%17.8829.62
Auto-GAS61.14%79.32%12.2123.19
IIAuto-GAS without Transform55.78%73.60%14.2225.58
Auto-GAS without Encoding51.64%68.62%15.2626.49
Auto-GAS without Reduction52.78%71.95%14.6325.74
Auto-GAS without Augment58.40%76.06%13.8625.12
IIIAuto-GAS (Gram-matrix)61.14%79.32%12.2123.19
Auto-GAS (Pearson)60.07%77.53%12.8524.62
Auto-GAS (Hamming)60.58%78.57%12.3523.88
IVAuto-GAS (Proxy Random)29.49%38.25%16.6528.12
Auto-GAS (Proxy Evolution)61.14%79.32%12.2123.19
", + "bbox": [ + 267, + 339, + 733, + 500 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SNIP require gradient backpropagation, leading to increased time consumption during the search process. In contrast, our Auto-GAS only requires a single forward pass, making it faster than methods that necessitate multiple forwards or gradient backpropagation. For image translation, as presented in Table 4, Auto-GAS achieves a $110 \\times$ enhancement compared to GAN Compression [38].", + "bbox": [ + 212, + 535, + 784, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of proxy search. (1) Our proxy search space includes transform, encoding, reduction, and augmentation operations. Ablations in Table 5 (II) indicate that all these operators contribute, with encoding operators having a greater influence. (3) Ablations for different encoding operators in Table 5 (III) show that the gram-matrix is important, but other operators (e.g., Pearson correlation matrix) also achieve similar performance. (2) Our contribution stems from search space design and proxy search processes. Ablations in Table 5 demonstrate that the various components of our search space contribute. In addition, our proxy evolution search plays an important role in exploring extensive search space and it significantly outperforms random search in Table 5 (IV).", + "bbox": [ + 212, + 611, + 784, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Operations for mutual information estimation. (1) Gram-matrix calculates matrix correlation for mutual information, which measures the salient features and relevant information captured by generative models. The comparison in Table 5 (I) shows that the vanilla Gram-matrix can be used as a proxy for NAS but is weaker than our Auto-GAS. (2) We also compare other operations for fea-", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Automated Proxy Discovery for Training-free Generative Architecture Search", + "bbox": [ + 217, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f246ef192f40563d4877328f69b429fb3eeb9e4d98bf50673c0b3eb44283e4f9.jpg", + "table_caption": [ + "Table 6: Results of different correlation methods on the CIFAR-10 dataset. analysis on the CIFAR-10 dataset." + ], + "table_footnote": [], + "table_body": "
Correlation MethodsCIFAR-10HyperparameterCIFAR-10
IS↑FID↓IS↑FID↓
Hessian Trace8.3615.47Batch Size (20)8.5212.28
L2norm7.5624.62Batch Size (40)8.0216.52
Entropy7.7025.43Number of Individuals (16)8.4412.21
Variance7.6723.75Number of Individuals (64)8.1116.32
Gram-matrix8.1217.88EA iteration interval (5)7.7818.25
Auto-GAS8.4412.21EA iteration interval (10)8.3813.69
", + "bbox": [ + 217, + 183, + 741, + 300 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ture information estimation in Table. 6. The results of $L_{2}$ norm|Entropy|Variance are worse than Auto-GAS, because they overlooks higher-order feature statistics. Hessian Trace [43] yields relatively promising performance but its computational complexity is excessive.", + "bbox": [ + 212, + 329, + 782, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Hyperparameter analysis. The results in Table 7 indicate that a larger batch size may speed up convergence but also risk overfitting. An appropriate population size ensures our more efficient architecture exploration. Larger iteration intervals allow for a more comprehensive exploration in our architecture space.", + "bbox": [ + 212, + 390, + 784, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 474, + 357, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present Auto-GAS, a novel training-free generative architecture search framework by evolving customized proxy. We derive from the information bottleneck theory using the latent layer feature as input and building candidate proxies with various transform, encoding, reduction, and augment options. We search for the best proxy using evolutionary algorithms. With customized proxies, our Auto-GAS allows an efficient search for promising architectures. Comprehensive experiment results on image generation and image translation illustrate the promising performance of Auto-GAS. For future work, we will augment our Auto-GAS with advanced methods [31-35,37,52,69]. We hope our novel investigations will provide some insight for the generation and NAS research community.", + "bbox": [ + 212, + 508, + 784, + 674 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Limitation.** We mainly evaluate the same benchmarks as other NAS methods for fair comparison. We will extend Auto-GAS to more generative models in future work. **Social impact.** Our Auto-GAS focuses on technology improvement without social and ethical implications.", + "bbox": [ + 212, + 675, + 784, + 736 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 761, + 401, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The research was supported by Theme-based Research Scheme (T45-205/21-N) from Hong Kong RGC, and Generative AI Research and Development Centre from InnoHK.", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdelfattah, M.S., Mehrotra, A., Dudziak, L., Lane, N.D.: Zero-cost proxies for lightweight nas. arXiv preprint arXiv:2101.08134 (2021)", + "2. Akhauri, Y., Munoz, J.P., Jain, N., Iyer, R.: EZNAS: Evolving zero-cost proxies for neural architecture scoring. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) NeurIPS (2022), https://openreview.net/forum?id=1SqaDG4dvdt", + "3. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. ICLR (2017)", + "4. Baker, B., Gupta, O., Naik, N., Raskar, R.: Designing neural network architectures using reinforcement learning. In: ICLR (2017)", + "5. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)", + "6. Cavagnero, N., Robbiano, L., Pistilli, F., Caputo, B., Averta, G.: Entropic score metric: Decoupling topology and size in training-free nas. In: ICCV workshop (2023)", + "7. Chen, K., Yang, L., Chen, Y., Chen, K., Xu, Y., Li, L.: Gp-nas-ensemble: a model for the nas performance prediction. In: CVPRW (2022)", + "8. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Fourteenth International Conference on Artificial Intelligence and Statistics (2011)", + "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. CVPR (2016)", + "0. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023)", + "1. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV (2023)", + "2. Dong, P., Niu, X., Li, L., Tian, Z., Wang, X., Wei, Z., Pan, H., Li, D.: Rd-nas: Enhancing one-shot supernet ranking ability via ranking distillation from zero-cost proxies. arXiv preprint arXiv:2301.09850 (2023)", + "3. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022)", + "4. Dong, P., Niu, X., Tian, Z., Li, L., Wang, X., Wei, Z., Pan, H., Li, D.: Progressive meta-pooling learning for lightweight image classification model. In: ICASSP (2023)", + "5. Duan, Y., Chen, X., Xu, H., Chen, Z., Liang, X., Zhang, T., Li, Z.: Transnas-bench-101: Improving transferability and generalizability of cross-task neural architecture search. In: CVPR. pp. 5251-5260 (2021)", + "6. Elsken, T., Metzen, J.H., Hutter, F.: Neural architecture search: A survey. arXiv preprint arXiv:1808.05377 (2018)", + "7. Gao, C., Chen, Y., Liu, S., Tan, Z., Yan, S.: Adversarialnas: Adversarial neural architecture search for gans. In: CVPR (2020)", + "8. Gong, X., Chang, S., Jiang, Y., Wang, Z.: Autogan: Neural architecture search for generative adversarial networks. ICCV (2019)", + "9. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)", + "20. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. NeurIPS (2017)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Automated Proxy Discovery for Training-free Generative Architecture Search", + "bbox": [ + 217, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. He, H., Wang, H., Lee, G.H., Tian, Y.: Probgan: Towards probabilistic gan with theoretical guarantees. In: ICLR (2018)", + "22. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021)", + "23. Isola, P., Zhu, J., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017). https://doi.org/10.1109/CVPR.2017.632", + "24. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017)", + "25. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (June 2019)", + "26. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.E.: Similarity of neural network representations revisited. In: ICML (2019)", + "27. Krizhevsky, A., Nair, V., Hinton, G.: The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html (2014)", + "28. Lee, N., Ajanthan, T., Torr, P.H.: Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340 (2018)", + "29. Li, G., Yang, Y., Bhardwaj, K., Marculescu, R.: Zico: Zero-shot NAS via inverse coefficient of variation on gradients. In: The Eleventh ICLR (2023), https://openreview.net/forum?id=rwo-1s5GqGn", + "30. Li, H., Fu, T., Dai, J., Li, H., Huang, G., Zhu, X.: Autoloss-zero: Searching loss functions from scratch for generic tasks. In: CVPR (2022)", + "31. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022)", + "32. Li, L., Bao, Y., Dong, P., Yang, C., Li, A., Luo, W., Liu, Q., Xue, W., Guo, Y.: Detkds: Knowledge distillation search for object detectors. In: ICML (2024)", + "33. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024)", + "34. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023)", + "35. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022)", + "36. Li, L., Sun, H., Dong, P., Wei, Z., Shao, S.: Auto-das: Automated proxy discovery for training-free distillation-aware architecture search. In: ECCV (2024)", + "37. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation. In: ICLR (2020)", + "38. Li, M., Lin, J., Ding, Y., Liu, Z., Zhu, J.Y., Han, S.: Gan compression: Efficient architectures for interactive conditional gans. CVPR (2020)", + "39. Lin, M., Wang, P., Sun, Z., Chen, H., Sun, X., Qian, Q., Li, H., Jin, R.: Zen-nas: A zero-shot nas for high-performance image recognition. ICCV (2021)", + "40. Lin, Q., Fang, Z., Chen, Y., Tan, K.C., Li, Y.: Evolutionary architectural search for generative adversarial networks. IEEE Transactions on Emerging Topics in Computational Intelligence (2022)", + "41. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019)", + "42. Liu, L., Zhang, Y., Deng, J., Soatto, S.: Dynamically grown generative adversarial networks. AAAI (2021)", + "43. Liu, Y., Yu, S., Lin, T.: Hessian regularization of deep neural networks: A novel approach based on stochastic estimators of hessian trace. Neurocomputing 536, 13-20 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Luo, Y., Zhang, Y., Cai, X., Yuan, X.: E2gan: End-to-end generative adversarial network for multivariate time series imputation. In: 28th international joint conference on artificial intelligence. pp. 3094-3100. AAAI Press Palo Alto, CA, USA (2019)", + "45. Mellor, J., Turner, J., Storkey, A., Crowley, E.J.: Neural architecture search without training. In: ICML (2021)", + "46. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. ArXiv abs/1802.05957 (2018)", + "47. Odena, A., Olah, C., Shlens, J.: Conditional image synthesis with auxiliary classifier gans. In: ICML. PMLR (2017)", + "48. Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: CVPR (2019)", + "49. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI (2022)", + "50. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR abs/1511.06434 (2016)", + "51. Real, E., Liang, C., So, D.R., Le, Q.V.: Automl-zero: Evolving machine learning algorithms from scratch (2020)", + "52. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023)", + "53. Shu, H., Wang, Y., Jia, X., Han, K., Chen, H., Xu, C., Tian, Q., Xu, C.: Coevolutionary compression for unpaired image translation. In: ICCV (2019)", + "54. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information. arXiv:1703.00810 (2017)", + "55. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information. arXiv preprint, arXiv:1703.00810 (2017)", + "56. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)", + "57. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS (2020)", + "58. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS 33, 6377-6389 (2020)", + "59. Tian, Y., Shen, L., Su, G., Li, Z., Liu, W.: Alphagan: Fully differentiable architecture search for generative adversarial networks. arXiv preprint arXiv:2006.09134 (2020)", + "60. Tian, Y., Wang, Q., Huang, Z., Li, W., Dai, D., Yang, M., Wang, J., Fink, O.: Off-policy reinforcement learning for efficient and effective gan architecture search. In: ECCV (2020)", + "61. Tishby, N., Zaslavsky, N.: Deep learning and the information bottleneck principle. In: ITW. pp. 1-5. IEEE (2015)", + "62. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020)", + "63. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020)", + "64. Wang, C., Xu, C., Yao, X., Tao, D.: Evolutionary generative adversarial networks. IEEE Transactions on Evolutionary Computation 23(6), 921-934 (2019)", + "65. Wang, H., Huan, J.: Agan: Towards automated design of generative adversarial networks. arXiv preprint arXiv:1906.11080 (2019)", + "66. Wang, W., Sun, Y., Halgamuge, S.: Improving MMD-GAN training with repulsive loss function. In: ICLR (2019)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Automated Proxy Discovery for Training-free Generative Architecture Search", + "bbox": [ + 217, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "67. Wei, Z., Pan, H., Li, L.L., Lu, M., Niu, X., Dong, P., Li, D.: Convformer: Closing the gap between cnn and vision transformers. arXiv preprint arXiv:2209.07738 (2022)", + "68. Wolchover, N., Reading, L.: New theory cracks open the black box of deep learning. Quanta Magazine (2017)", + "69. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching. In: ICLR (2023)", + "70. Ying, G., He, X., Gao, B., Han, B., Chu, X.: Eagan: Efficient two-stage evolutionary architecture search for gans. In: ECCV. Springer (2022)", + "71. Yu, A., Grauman, K.: Fine-grained visual comparisons with local learning. CVPR (2014)", + "72. Zhang, M., Li, H., Pan, S., Chang, X., Su, S.: Overcoming multi-model forgetting in one-shot nas with diversity maximization. In: CVPR. pp. 7809-7818 (2020)", + "73. Zhong, Z., Yang, Z., Deng, B., Yan, J., Wu, W., Shao, J., Liu, C.L.: Blockqnn: Efficient block-wise neural network architecture generation. IEEE transactions on pattern analysis and machine intelligence (2020)", + "74. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024)", + "75. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024)", + "76. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018)" + ], + "bbox": [ + 215, + 146, + 787, + 465 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Lujun Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/d262e2a7-fb7b-49f5-8763-ff8884371602_model.json b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/d262e2a7-fb7b-49f5-8763-ff8884371602_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6deac90e1c53f4da3f3904b32c8a9ba6337092ef --- /dev/null +++ b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/d262e2a7-fb7b-49f5-8763-ff8884371602_model.json @@ -0,0 +1,2579 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "Auto-GAS: Automated Proxy Discovery for Training-free Generative Architecture Search" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.215, + 0.78, + 0.248 + ], + "angle": 0, + "content": "Lujun Li\\(^{1}\\), Haosen Sun\\(^{1}\\), Shiwen Li\\(^{2}\\), Peijie Dong\\(^{3}\\), Wenhan Luo\\(^{1}\\), Wei Xue\\(^{1}\\), Qifeng Liu\\(^{1\\boxtimes}\\), and Yike Guo\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.258, + 0.69, + 0.273 + ], + "angle": 0, + "content": "1 The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.273, + 0.568, + 0.287 + ], + "angle": 0, + "content": "\\(^{2}\\) Xidian University" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.287, + 0.767, + 0.342 + ], + "angle": 0, + "content": "3 The Hong Kong University of Science and Technology (Guangzhou) lilujunai@gmail.com;hsunas@connect.ust.hk;shiwenlisw@gmail.com; pdong212@connect.hkust-gz.edu.cn;whluo.china@gmail.com;weixue@ust.hk; liuqifeng@ust.hk;yikeguo@ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.412, + 0.744, + 0.759 + ], + "angle": 0, + "content": "Abstract In this paper, we introduce Auto-GAS, the first training-free Generative Architecture Search (GAS) framework enabled by an autodiscovered proxy. Generative models like Generative Adversarial Networks (GANs) are now widely used in many real-time applications. Previous GAS methods use differentiable or evolutionary search to find optimal GAN generators for fast inference and memory efficiency. However, the high computational overhead of these training-based GAS techniques limits their adoption. To improve search efficiency, we explore training-free GAS but find existing zero-cost proxies designed for classification tasks underperform on generation benchmarks. To address this challenge, we develop a custom proxy search framework tailored for GAS tasks to enhance predictive power. Specifically, we construct an information-aware proxy that takes feature statistics as inputs and utilizes advanced transform, encoding, reduction, and augment operations to represent candidate proxies. Then, we employ an evolutionary algorithm to perform crossover and mutation on superior candidates within the population based on correlation evaluation. Finally, we perform generator search without training using the optimized proxy. Thus, Auto-GAS enables automated proxy discovery for GAS while significantly accelerating the search before training stage. Extensive experiments on image generation and image-to-image translation tasks demonstrate that Auto-GAS strikes superior accuracy-speed tradeoffs over state-of-the-art methods. Remarkably, Auto-GAS achieves competitive scores with \\(110 \\times\\) faster search than GAN Compression. Code at: https://github.com/lliai/Auto-GAS." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.775, + 0.744, + 0.803 + ], + "angle": 0, + "content": "Keywords: Generative Model \\(\\cdot\\) Generative Architecture Search \\(\\cdot\\) AutoML" + }, + { + "type": "page_footnote", + "bbox": [ + 0.231, + 0.826, + 0.411, + 0.84 + ], + "angle": 0, + "content": "Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.145, + 0.768, + 0.16 + ], + "angle": 0, + "content": "Table 1: Comparison of search cost for generative architecture search methods." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.171, + 0.781, + 0.245 + ], + "angle": 0, + "content": "
MethodAlgorithm Cost (GPU-days)MethodAlgorithm Cost (GPU-days)
AGAN [65]RL1200DGGAN [42]Heuristic580
AutoGAN [18]RL2EAS-GAN [40]EA1
AlphaGAN [59]Gradient1.2EGAN [64]EA1.3
EAGAN [70]EA0.8Auto-GAS (ours)Training-free0.09
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.274, + 0.377, + 0.29 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.787, + 0.491 + ], + "angle": 0, + "content": "Generative models like Generative Adversarial Networks (GANs) [19] have made tremendous strides in visual generation tasks such as image and video synthesis [5,50]. However, their demanding computational footprint presents barriers to deploying these models on edge devices with strict resource constraints [25,47]. To overcome this, there is growing interest in model compression techniques [7, 12, 14, 49, 67]. Among them, Neural Architecture Search (NAS) [16] automates architecture design, finding compact generative models for edge computing, and enabling flexible structures without pre-trained weights. In contrast, other compression methods are limited to fixed, handcrafted architectures and rely on pre-trained models. The integration of NAS in generative models enables real-time visual creation on edge devices, unlocking new interactive experiences in domains like augmented reality [5,24,50]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.765 + ], + "angle": 0, + "content": "Conventional Generative Architecture Search (GAS) approaches typically rely on expensive training-based training routes to assess candidate architectures, such as maintaining weight-sharing supernets [13, 22] or conducting multiple training trials from scratch [56, 76]. However, training generative models presents unique obstacles that greatly increase the costs and difficulties of applying these techniques for GAS. The training processes for generative models are generally more complex than discriminative tasks because inherent instabilities (e.g., mode collapse [20]) can occur during optimization. In mode collapse, the model generates only a limited variety of examples from the underlying distribution. These instabilities make it challenging to reliably evaluate architectural candidates with standard NAS training-based evaluations. As shown in Table 1, gradient-based GAS methods like AdversarialNAS [17] and AlphaGAN [59] amplify these problems by jointly searching the generator and discriminator architecture spaces. This coupled search increases the overall complexity and fragility of the optimization procedure. To address these difficulties, DGGAN [42] introduced a progressive search approach where candidates are trained incrementally. However, its process still requires a massive 580 GPU days to complete due to the computationally intensive nature of generative model training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Recently, training-free NAS approaches have demonstrated remarkable efficiency using zero-cost proxies to predict network performance without training network parameters. These methods are mainly present for encoder-only models for classification tasks. Inspired by their computational efficiency, we comprehensively evaluate existing proxies when applied to generative tasks. As shown" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.218, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.363, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.146, + 0.505, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.147, + 0.644, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.147, + 0.785, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.249, + 0.784, + 0.278 + ], + "angle": 0, + "content": "Figure 1: Correlation visualization of proxies (from left to right: Param., Synflow, NWOT, and Auto-GAS) in Generation benchmarks [15]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.49 + ], + "angle": 0, + "content": "in Figure 1, we can make some interesting observations: (1) We observe that parameter-level proxies (e.g., Synflow [57], SNIP [28], and GraSP [62]), which estimate performance based on weight importance, generally perform poorer when ranking generator candidates. This indicates their assumptions regarding how architecture impacts learning do not directly translate to generative objectives and training dynamics. (2) Architecture-level proxies (e.g., NWOT [45] and Zen [39]) featuring topological encodings demonstrate stronger ranking ability. However, even the best architecture proxies leave room for improvement to enable performing generative NAS at scale. These findings motivate developing customized proxies that more precisely capture the unique characteristics of visual generation. Thus, a question naturally arises: how to design customized proxies for generation tasks?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.763 + ], + "angle": 0, + "content": "To answer this question, we present Auto-GAS, the first automated generative architecture search framework that aims to efficiently utilize custom training-free proxies. Firstly, we derive that the mutual information of latent features within generative models is strongly coupled with task optimization objectives based on information bottleneck theory [3,54,61]. Then, we develop proxy search strategies to better estimate the mutual information of these features. Specifically, we build comprehensive operation search spaces and evolutionary algorithms to search optimal information-aware proxies. To better process feature statistics, our search space involves transform-based, encoding-based, reduction-based, and augment-based operators. The search process in Auto-GAS begins with initializing the population of proxies and then evaluates, crosses, and varies them to improve the quality of the proxies. Throughout the search, we directly utilize the ranking correlation with accuracy results in the benchmark as fitting objectives, aiming to find a more suitable proxy for the generation tasks. With this evolutionary process, Auto-GAS outperforms existing training-free NAS approaches by significant margins without manual tuning. Additionally, we search for generator architectures using Auto-GAS and subsequently implement a full training process on the searched generator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We perform extensive experiments to validate the performance and efficiency strengths of our Auto-GAS in generation tasks like image generation and imaged-to-image translation. The experiments demonstrate that our Auto-GAS can achieve better-generating performance than other zero-shot proxies and state-of-the-art results on rank consistency. In contrast to handcrafted GAS methods," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "our Auto-GAS framework enjoys many benefits: (1) Automatic. It enables the automatic discovery of more expressive and efficient proxies that human expert might ignore, which can reduce human bias and ensures that the resulting architectures are optimized for the target problem or dataset. (2) General. The search procedure of Auto-GAS improves generalization by discovering proxies that generalize well on different search spaces. Based on the final results for the search objectives, our framework can evolve adaptive proxies to generalize across architectures. (3) Insightful. Our framework allows for the automatic discovery of more expressive, efficient, and effective proxies for optimizing and customizing generation architectures, opening up new research directions for automated GAN architecture search. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.321, + 0.784, + 0.365 + ], + "angle": 0, + "content": "- We present Auto-GAS, the first training-free architecture search framework for generation tasks. Auto-GAS achieves an ultra-efficient and elegant search process with our auto-searched customized proxy." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.366, + 0.784, + 0.424 + ], + "angle": 0, + "content": "- We propose a well-organized proxy search space, including latent features as input and various transforms, encoding, reduction, and augment operations as options. We develop evolutionary algorithms for effective & efficient proxy search." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.426, + 0.784, + 0.47 + ], + "angle": 0, + "content": "- We conduct extensive experiments on image generation and image-to-image translation tasks. Auto-GAS obtains significant search acceleration and competitive performance on multiple datasets." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.321, + 0.784, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.493, + 0.388, + 0.509 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Generative architecture search. The field of Neural Architecture Search (NAS) traditionally encompasses the design of exploration spaces, search methodologies, and assessment techniques to autonomously uncover optimal architectural configurations within specified parameters. Empirical approaches utilize an iterative process of training followed by exploration, either through multiple iterations [4,73] or by employing weight-sharing strategies [41,72]. NAS has also been extensively applied to the domain of GANs Architecture Search (GAS). The primary goal in GAS is to automatically identify superior GAN structures that enhance generative performance. Pioneering reinforcement learning-based GAS techniques, such as AGAN [65] and AutoGAN [18], leverage RNN controllers to construct GAN architectures and employ the Inception Score (IS) as their evaluation criterion. E2GAN [60] and AdversarialNAS [17] implement alternative reward functions. EAGAN [70] conducts discrete searches and enhances the robustness of the GAN exploration process. Nevertheless, these training-dependent search methodologies continue to be hampered by inefficient exploration and substantial computational demands. To tackle this challenge, we propose a training-free approach, grounded in architecture decoupling, to enhance the discovery of competitive GAN structures in a more computationally efficient manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Training-free architecture search. Zero-shot architecture exploration [10, 11, 29, 39], alternatively referred to as training-free NAS, offers a swifter alternative to conventional NAS methods. This approach's evaluation process en" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "tails only a handful of forward or backward passes with a small batch of input data, rendering it highly efficient. Zero-cost indicators can be categorized into parameter-centric and architecture-centric proxies. Parameter-centric zero-cost proxies (e.g., Synflow [57], SNIP [28], and GraSP [62]) rely on pruning techniques and aggregate the saliency values of each layer's weights as a proxy. For instance, Synflow [58] introduces a modified version of synaptic saliency scores to prevent layer collapse during parameter pruning, while Fisher [1] calculates the sum of all activation gradients in the network, applicable to channel pruning. Grasp [63] ensures gradient flow preservation throughout the network for efficient training. Their detailed formulations are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.307, + 0.79, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\rho_ {s n i p} = \\left| \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W} \\right|, \\rho_ {s y n f l o w} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W}, \\rho_ {f i s h e r} = \\left(\\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {A}} \\mathcal {A}\\right) ^ {2}, \\rho_ {g r a s p} = - (H \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}}) \\odot \\mathcal {W}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.789, + 0.637 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\), \\(\\mathcal{W}\\), \\(\\mathcal{A}\\) are loss function, weight and activation. \\(H\\) is the Hessian matrix. However, the complexity of loss function \\(\\mathcal{L}\\) in generation tasks renders these gradient-based methods ineffective. Regarding architecture-centric proxies, ZenNAS [39] introduced the novel Zen-Score to assess network expressivity, while NWOT [45] examined correlations between linear maps induced by data points for untrained architectures. Nevertheless, these approaches employ hand-crafted fixed proxies for classification tasks and encoder-only CNN models. In generation tasks, the encoder-decoder generator typically features intricate branches and multi-scale features, posing significant challenges for traditional proxies in predicting final accuracy. To address this issue, we automatically optimize the first customized proxies for GAS. Furthermore, our framework differs from EZNAS-like methods [2, 36, 74, 75] in several aspects: (1) Tasks. EZNAS is exclusively designed for classification, whereas our Auto-GAS is tailored for generation tasks. (2) Search algorithms. EZNAS merely employs genetic programming, while we propose a proxy evolution search strategy that substantially enhances search efficiency. (3) Model statistic inputs. Our method solely utilizes features as input, as we observed poor performance of parameter-level proxies. Conversely, EZNAS selects parameters and gradients as inputs. (4) Search space. Our Auto-GAS incorporates numerous feature-related operations absent in EZNAS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.381, + 0.677 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.737 + ], + "angle": 0, + "content": "In this section, we first review existing proxies. Then, we present formulations and insights of our proxy design (see Figure 2). Finally, we illustrate the detailed search process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.427, + 0.772 + ], + "angle": 0, + "content": "3.1 Proxy Search Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Motivation of our customized proxy design. Generative tasks are relevant to the Information Bottleneck (IB) theory [3, 54, 61], which guides generative models to learn compact representations that retain important characteristics for accurate sample generation. In generative tasks, the IB theory helps understand" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.788, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.789, + 0.379 + ], + "angle": 0, + "content": "Figure 2: Overall proxy search of Auto-GAS. We extract the latent features of the generator with input data forward-propagation. Then we select various transform, encoding, reduction, and augment operators from the proxy search space to build candidate proxies. Finally, we perform a ranking evaluation to cross and mutate to get the best proxy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.789, + 0.473 + ], + "angle": 0, + "content": "how generative models can balance compression (capturing essential features) and reconstruction (faithful generation of samples) [55, 68]. IB theory suggests maximizing the Mutual Information (MI) \\( I(\\mathbf{X}; \\mathbf{F}_i) \\) between the input data \\( \\mathbf{X} \\) and the corresponding latent representation \\( \\mathbf{F}_i \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.489, + 0.787, + 0.507 + ], + "angle": 0, + "content": "\\[\nI (\\mathbf {X}; \\mathbf {F} _ {i}) = H (\\mathbf {F} _ {i}) - H (\\mathbf {F} _ {i} | \\mathbf {X}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.789, + 0.748 + ], + "angle": 0, + "content": "where \\( H(\\mathbf{F}_i) \\) stands for the entropy of the latent representation \\( \\mathbf{F}_i \\) and \\( H(\\mathbf{F}_i|\\mathbf{X}) \\) denotes the conditional entropy. Considering that neural networks are deterministic, and the dataset has a finite number of instances, the conditional entropy becomes zero, simplifying the mutual information (MI) estimation as \\( I(\\mathbf{X};\\mathbf{F}_i) = H(\\mathbf{F}_i) \\). Consequently, some information coding and compression studies [26, 55] propose some operations (e.g., Entropy and Gram-matrix) to approximate the mutual information on latent representation. These inspired us to develop our train-free method design thought chain: assess generative model \\(\\rightarrow\\) estimate MI \\(\\rightarrow\\) design expressions via Entropy/Gram-matrix ops \\(\\rightarrow\\) our train-free proxy. However, how about designing this potential expressions? Recent Auto-Zero approaches [30, 51] present a promising solution: searching for the best machine learning solution from scratch. In addition, several studies [6] have shown that normalize and activation operations can enhance the formulation. Therefore, we automatically search for optimal proxies based on these operators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Proxy primitive operations. Our approach evolves the customized proxy in the generation benchmark [15], which consists of U-Net-like encoder-decoder models with various operators and their training results. We extract the latent features of each layer in both the encoder and decoder modules as inputs to the search process and represent the zero-cost proxy as a computation graph. This graph is constructed as transform \\(\\rightarrow\\) encoding \\(\\rightarrow\\) reduction \\(\\rightarrow\\) augment" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "parts in processing generator statistics. The available primitive operations can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.19, + 0.784, + 0.249 + ], + "angle": 0, + "content": "- Transform operations: These operations involve different activation functions and feature processing techniques to enhance raw features and represent architectural potentials. Examples of transform operations include relu, swish, mish, scale,norm, exp, and softmax." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.251, + 0.785, + 0.325 + ], + "angle": 0, + "content": "- Encoding operations: Inspired by coding algorithms and feature information estimation, we employ binary and matrix-based operations in the encoding part. These operators help us better estimate the mutual correlation information of samples. Some encoding operations include sign, above-zero, above-median, gram-matrix, person, and hamming." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.326, + 0.784, + 0.384 + ], + "angle": 0, + "content": "- Reduction operations: Reduction operators bridge the information matrix and the proxy values. They are valuable for extracting useful matrix information from different mathematical perspectives. Reduction operations include sloqdet, abslog, trace, and eigenvalue." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.387, + 0.784, + 0.43 + ], + "angle": 0, + "content": "- Augment operations: Augment operators are applied to enhance the proxy values' ability to predict performance results. These operations include arithmetic functions such as abs, pow, l2-norm, log, sum, and mean." + }, + { + "type": "list", + "bbox": [ + 0.232, + 0.19, + 0.785, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.785, + 0.52 + ], + "angle": 0, + "content": "By utilizing these primitive operations, our approach enables the evolution of the customized proxy, facilitating the search for optimal solutions in the generation benchmark. These mathematical operations are essential building blocks to construct a diverse and expressive search space of zero-cost proxy. We provide more details of their formulas and properties in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.542, + 0.457, + 0.557 + ], + "angle": 0, + "content": "3.2 Proxy Evolution Search" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.703 + ], + "angle": 0, + "content": "Evolutionary process and goal. Utilizing our proxy search space and graph representation, we employ an evolutionary algorithm (EA) outlined in Alg. 1 to discover optimal proxy expressions. Our EA commences with an initial set of candidate proxies and progressively refines the population across generations, employing genetic operators such as selection, crossover, and mutation to yield superior solutions. For fitness evaluation, each proxy is assessed based on the ranking correlation between its output proxy \\(\\mathcal{Q}\\) scores and pre-determined performance benchmarks \\(\\mathcal{D}\\), enabling efficient identification of the optimal proxy \\(\\rho^{*}\\) within the search space \\(\\mathcal{P}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.717, + 0.785, + 0.738 + ], + "angle": 0, + "content": "\\[\n\\rho_ {A u t o - G A S} ^ {*} = \\underset {\\rho \\in \\mathcal {P}} {\\arg \\max } (\\tau (\\mathcal {D}, \\mathcal {Q})). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.794 + ], + "angle": 0, + "content": "where Kendall's Tau \\(\\tau\\) is formulated as the correlation coefficient. Each evolution picks top-k candidates with the highest evaluation score and randomly selects a parent from the candidates for mutation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Efficient proxy search with existing generation benchmark. It is expensive to derive pre-trained results \\(\\mathcal{D}\\), we adopt an alternative solution to achieve proxy search using pre-computed architecture-performance pairs provided in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.146, + 0.624, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Evolutionary Search for Auto-GAS Proxy" + }, + { + "type": "algorithm", + "bbox": [ + 0.216, + 0.163, + 0.787, + 0.445 + ], + "angle": 0, + "content": "Input: Exploration domain \\(S\\), candidate set \\(\\mathcal{P}\\), maximum iterations \\(\\mathcal{T}\\), sampling ratio \\(r\\), sampled pool \\(\\mathcal{R}\\), top \\(k\\) selections. \nOutput: Auto-GAS proxy with highest Kendall's Tau \\(\\tau\\) coefficient. \n1: \\(\\mathcal{P}0 := \\text{Initialize candidate set}(P_i)\\); \n2: Sampled pool \\(\\mathcal{R} := \\emptyset\\); \n3: for \\(i = 1, 2, \\ldots, T\\) do \n4: Select \\(G_i^s := \\text{SelectTopk}(\\mathcal{R}, k)\\); \n5: // greedy evolutionary strategy \n6: Empty sampled pool \\(\\mathcal{R} := \\emptyset\\); \n7: Mutate := MUTATE(\\(G_i^s\\)); \n8: Crossover \\(G_i^c := \\text{CROSSOVER}(\\mathcal{G}_i^s)\\); \n9: Incorporate \\(G_i^m\\) into \\(R\\); \n10: Incorporate \\(G_i^c\\) into \\(R\\); \n11: if \\(\\mathcal{R} < \\mathcal{P}\\) then \n12: Introduce random samples \n13: else \n14: Proceed to line 4; \n15: end if \n16: end for" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.625 + ], + "angle": 0, + "content": "the existing TransNAS-Bench-101 [15] benchmarks. Specifically, this benchmark involves Autoencoding task with 7k different generators in macro-level search space and their evaluation metric (i.e., SSIM). Each generator network follows an encoder-decoder structure in Pix2Pix, where the encoders are the searched backbones, and the decoders contain 14 layers of convolution and deconvolution. Thanks to this existing benchmark, we do not need to train performance groundtruths again, and it only takes 1.5 hours in our proxy search process. In light of excellent generality (proven in our extensive experiments), our searched proxy \\(\\rho_{Auto-GAS}\\) can be well transferred to similar search spaces and can be directly used in most scenarios without additional search budget." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.484, + 0.661 + ], + "angle": 0, + "content": "3.3 Discovered Proxy Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.787, + 0.701 + ], + "angle": 0, + "content": "Searched Zero-cost Proxy. We present formulas of the searched proxies on the generation benchmark as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.291, + 0.709, + 0.787, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\rho_ {A u t o - G A S} ^ {*} = \\left(\\operatorname {s l o g d e t} \\left([ \\mathbf {F} ^ {\\prime}, \\mathbf {F} ^ {\\prime T} ]\\right)\\right) ^ {2}, \\mathbf {F} ^ {\\prime} = \\operatorname {a b o v e - m e a n} (\\operatorname {s w i s h} (\\mathbf {F})) \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}\\) represents the original features, and transformed features \\(\\mathbf{F}'\\) are obtained by applying the swish operation followed by the above-mean operation to the original features \\(\\mathbf{F}\\). The \\([\\mathbf{F}',\\mathbf{F}'^T]\\) indicates the \\(N\\times N\\) gram-matrix of the transformed features \\(\\mathbf{F}'\\). The proxy value \\(\\rho^{*}\\) is calculated by taking the square of the slogdet of the concatenated gram-matrix \\([\\mathbf{F}',\\mathbf{F}'^T]\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Analysis of searched proxy. The discovered formulations indicate that (1) Mathematical operations, such as swish activation and slogdet, significantly con-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.218, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.315, + 0.145, + 0.687, + 0.16 + ], + "angle": 0, + "content": "Table 2: Ranking results on TransNAS-Bench-101 [15]." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.172, + 0.773, + 0.273 + ], + "angle": 0, + "content": "
MethodKendallSpearmanPearsonMethodKendallSpearmanPearson
SNIP [1]-2.04%-3.55%-0.26%ZiCo [29]18.63%15.53%21.13%
EZNAS [2]23.30%25.91%32.03%Synflow [57]-50.91%-65.01%-49.17%
NWOT [45]51.28%63.94%48.88%Params.-12.97%-15.24%-5.80%
SNIP [28]30.81%23.31%29.34%FLOPs48.91%64.01%52.36%
Zen [39]-2.27%2.58%5.39%Auto-GAS61.14%79.32%81.25%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.785, + 0.451 + ], + "angle": 0, + "content": "tribute to the predictability of the proxy. These operations enhance the proxy's ability to capture important features and characteristics of the generator. (2) Gram-Matrix and Mutual Information: The expression utilizes a sample-wise gram-matrix, which accurately estimates the latent mutual information of the generator. The gram-matrix captures the relationships and dependencies among the transformed features, providing valuable insights into the generator's behaviour. The discovered proxy of Auto-GAS, as illustrated in Figure 1, showcases a substantial improvement in ranking. This improvement suggests that the searched proxy effectively captures essential characteristics of the generator, leading to enhanced performance in the benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.787, + 0.59 + ], + "angle": 0, + "content": "Understanding why our Auto-GAS surpasses other NAS methods. (1) Many train-based NAS methods employ weight-sharing and other proxy settings, which often lead to inaccurate estimations of architecture performance. In contrast, our Auto-GAS method searches for proxies based on actual performance, resulting in a proxy that demonstrates strong ranking capability. (2) Other train-free NAS methods are primarily designed for classification rather than generative tasks. Consequently, these methods search for architectures using weak correlation proxies. In contrast, our proxy exhibits a stronger correlation, ensuring the discovery of more efficient models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.614, + 0.591, + 0.63 + ], + "angle": 0, + "content": "3.4 Training-free Generative Models Search" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.776 + ], + "angle": 0, + "content": "Auto-GAS has made efforts to search generative Models in a non-training way to ease the instability of GAN training and improve search efficiency. In this stage, we directly search the generator in \\(\\rho_{Auto - GAS}\\). In each iteration, we generate \\(P\\) candidate generators to make up the population \\(\\mathcal{A}\\), where all these candidates use the randomly initialized weights \\(W\\). Each candidate is characterized by both architecture and weight variables. In each iteration, we evaluate \\(P\\) candidates without training. We employ the information-aware proxy to assign scores to all these candidates. The formulation of the search stage can be succinctly stated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.781, + 0.785, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {A u t o - G A S} ^ {*} = \\arg \\max _ {\\alpha \\in \\mathcal {A}} \\rho_ {A u t o - G A S} ^ {*} (\\alpha , \\mathcal {W}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Note that in scenarios involving discriminator searches, we also use \\(\\rho_{Auto - GAS}\\) for searching the discriminator architectures with real sample inputs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Comparison results on the CIFAR-10 and STL-10 datasets. We use FID (the lower the better) and the mIoU metric (the higher the better) for evaluations." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.185, + 0.725, + 0.396 + ], + "angle": 0, + "content": "
MethodAlgorithmGPU daysCIFAR-10 (32×32)STL-10 (48×48)
IS↑FID↓IS↑FID↓
WGAN-GP [20]7.8629.3--
SN-GAN [46]8.2221.709.1640.1
ProbGAN [21]Manual-7.7524.608.8746.74
Improv MMD GAN [66]8.2916.219.3437.63
BigGAN [5]9.2214.73--
AGAN [65]12008.2930.509.2352.70
AutoGAN [18]RL2.008.5512.429.1631.01
AlphaGAN [59]1.28.7015.56--
AdversarialNAS [17]1.007.8624.048.5238.85
EGAN [64]1.256.90---
EAS-GAN [40]EA1.007.4533.20-38.84
EAGAN [70]0.88.4112.839.6923.82
Fisher [1]0.128.0317.669.3729.82
Zen [39]0.118.0215.999.4528.66
SNIP [28]Training-free0.127.9616.679.5829.25
NWOT [45]0.098.0516.329.2327.46
Auto-GAS0.098.4412.219.7223.19
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.424, + 0.495, + 0.44 + ], + "angle": 0, + "content": "3.5 Generative Models Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.542 + ], + "angle": 0, + "content": "After the architecture searches, we train generative models with randomly selected data samples. Simultaneously, a batch of noise samples is generated from a noise distribution. The generator tries to generate data from random noise \\( z \\) as \\( G(z) \\), which should be as close as possible to the real data distribution. The discriminator tries to distinguish the generated data from the real data. The value function \\( V(D,G) \\) for a minimax game is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.563, + 0.785, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {G} \\max _ {D} V (D, G) = \\mathbb {E} _ {\\mathbf {x} \\sim p _ {d a t a} (\\mathbf {x})} [ \\log D (\\mathbf {x}) ] \\\\ + \\mathbb {E} _ {\\mathbf {z} \\sim p _ {z} (\\mathbf {z})} [ \\log (1 - D (G (\\mathbf {z})) ] \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.648 + ], + "angle": 0, + "content": "where \\(\\mathbf{x} \\sim p_{data}(\\mathbf{x})\\) denotes the distribution of real data, \\(\\mathbf{z} \\sim p_z(\\mathbf{z})\\) denotes the input noise to the generator." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.67, + 0.377, + 0.687 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.787, + 0.764 + ], + "angle": 0, + "content": "In this section, we first perform proxy search experiments on the generation benchmark, and then we conduct architecture search experiments on image generation and image-to-image translation tasks by translocating the searched proxies. Finally, we provide a detailed analysis of our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.591, + 0.801 + ], + "angle": 0, + "content": "4.1 Experiments on Generation Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Dataset and implementation. We adopt the search space and performance ground truth provided by the Autoencoding task in TransNAS-Bench-101 [15] as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.218, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.15, + 0.499, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.153, + 0.782, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.194, + 0.498, + 0.209 + ], + "angle": 0, + "content": "Figure 3: Searched model on CIFAR-10." + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.194, + 0.771, + 0.209 + ], + "angle": 0, + "content": "Figure 4: Searched model on STL-10." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.237, + 0.785, + 0.343 + ], + "angle": 0, + "content": "the benchmark for searching and evaluating GAS proxies. We split the 4:3 ratio of the benchmark into the validation and test sets. During the proxy search phase, we set \\((\\mathcal{P},\\mathcal{T},r,k)\\) in Alg. 1 as (20,100,0.9,5) and evaluate with 50 architectures in the validation set. After the search, we fairly compare the Auto-GAS proxy with the traditional ones by randomly sampling 200 architectures in the test set. Note that our validation and test sets do not overlap to ensure fair proxy comparisons. Detailed implementations are in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.787, + 0.435 + ], + "angle": 0, + "content": "Comparison results. Table 5 demonstrates the ranking capabilities of different methods on Generation Benchmark. Auto-GAS (ours) achieves the best rankings across all three measures: Kendall's Tau \\((61.14\\%)\\), Spearman \\((79.32\\%)\\), and Pearson \\((81.25\\%)\\). These results indicate that Auto-GAS performed well in preserving the rankings compared to other methods. These findings highlight the effectiveness of Auto-GAS for generating meaningful and accurate rankings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.547, + 0.47 + ], + "angle": 0, + "content": "4.2 Experiments on Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Datasets and implementation. We perform evaluations using CIFAR-10 [27] and STL-10 [8] datasets, aligning with other GAS methodologies [17, 18, 70]. CIFAR-10 comprises 50k training and 10k test \\(32 \\times 32\\) images, while STL-10 contains 100,500 images at a higher \\(48 \\times 48\\) resolution. We directly apply searched proxies to identify optimal architectures within EAGAN's [70] search space. The exploration process spans 150 epochs with a \\(P = 32\\) individual population. Generator and discriminator batch sizes are 80 and 40, respectively. We utilize 16 randomly generated Gaussian noise images for proxy score computation to expedite evaluation. Post-search, we conduct full training of the best-scoring generator for 500 epochs. For evaluation, following previous GAS works [17, 18, 70], we generate 50,000 images to calculate IS and FID metrics. Figures 3 and 4 illustrate the discovered architectures for CIFAR-10 and STL-10." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.792, + 0.809 + ], + "angle": 0, + "content": "Training-based NAS method comparison. We benchmark Auto-GAS against various training-based NAS approaches: BigGAN [5], E2GAN [44], Adversarial-NAS [17], and EAGAN [70]. As shown in Table 3, Auto-GAS demonstrates superior efficiency. For CIFAR-10, Auto-GAS achieves lower FID than BigGAN and comparable overall results to E2GAN, while outperforming AdversarialNAS in both FID and IS. On STL-10, Auto-GAS significantly outperforms E2GAN and AdversarialNAS in both metrics, rivaling EAGAN in IS and surpassing it in FID. In summary, Auto-GAS emerges as an efficient and competitive approach, either surpassing or matching several state-of-the-art training-based NAS methods across diverse evaluation metrics and datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Zero-shot proxy comparison. Table 3 also juxtaposes the Auto-GAS proxy with four alternative zero-shot proxies. Fisher and SNIP, relying on gradient-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.232, + 0.145, + 0.77, + 0.16 + ], + "angle": 0, + "content": "Table 4: Image translation evaluation on horse→zebra and Cityscapes datasets." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.171, + 0.785, + 0.318 + ], + "angle": 0, + "content": "
ModelDatasetMethod#ParametersGPU daysFID (↓)mIoU (↑)
ValueRatioValueRatioValueIncreaseValueDrop
CycleGANhorse→zebra(512×512)Original11.4M---61.53--
COEV [53]----96.1534.6-
GAN Comp. [38]0.34M33.3×11.601.00×71.8110.3-
Fast GAN Comp. [38]0.36M32.1×3.123.72×65.193.66-
Auto-GAS0.39M29.3×0.11105.45×63.662.13-
GauGANCityscapes(512×512)Original93.0M---57.60-62.18-
GAN Comp. [38]20.4M4.6×17.701.00×55.192.4161.220.96
Fast GAN Comp. [38]20.2M4.6×10.401.70×56.251.3561.171.01
Auto-GAS21.0M4.5×0.16110.63×56.041.5661.210.97
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.409 + ], + "angle": 0, + "content": "based computations, yield similar scores. Zen-Score and NWOT exhibit positive correlations with model accuracy, outperforming gradient-based methods. These results validate the Auto-GAS proxy's efficacy, positioning it as a formidable contender among zero-shot proxy methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.431, + 0.63, + 0.447 + ], + "angle": 0, + "content": "4.3 Experiments on Image-to-image Translation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.578 + ], + "angle": 0, + "content": "Datasets and implementation. We comprehensively evaluate our approach on image translation tasks using high-resolution datasets, including [9] and horse \\(\\rightarrow\\) zebra [71] at \\(512\\times 512\\) resolutions. We apply the proxy for architecture exploration, adopting GAN Compression's [38] search space for fair comparisons, focusing on channel and layer number configurations. We evolve to compress Pix2Pix [23] on horse \\(\\rightarrow\\) zebra and GauGAN [48] on cityscapes. The search process involves 500 iterations, followed by model retraining with a 0.0002 learning rate and 16 batch size, adhering to the original paper [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.667 + ], + "angle": 0, + "content": "GAN compression method comparison. Table 4 demonstrates that Auto-GAS achieves the highest FID score of 63.66 for CycleGAN with a superior compression ratio \\((29.3\\times)\\) compared to alternatives. This indicates Auto-GAS effectively reduces model size while preserving image quality. For GauGAN, Auto-GAS exhibits a marginally higher FID score and a competitive mIoU score of 61.22." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.788, + 0.746 + ], + "angle": 0, + "content": "Qualitative Analysis. Figure 5 illustrates Auto-GAS's proficiency in translation tasks. In contrast, FastGAN Compression not only yields a lower FID score but also fails to eliminate artifacts like brown discoloration in zebra models. Conversely, Auto-GAS-generated zebra images accurately capture authentic black-and-white zebra patterns." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.391, + 0.785 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Search efficiency benefits. For image generation (see Table 3), our Auto-GAS method exhibits significantly superior search efficiency to training-based NAS GANs and has a definite advantage over other zero-cost proxies. The Fisher and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.218, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.785, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.268, + 0.784, + 0.297 + ], + "angle": 0, + "content": "Figure 5: Visualization results on image translation tasks using high-resolution datasets, e.g., horse→zebra and Cityscapes with \\(512 \\times 512\\) resolution." + }, + { + "type": "table_caption", + "bbox": [ + 0.366, + 0.315, + 0.634, + 0.328 + ], + "angle": 0, + "content": "Table 5: Ablations of our proxy search." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.34, + 0.734, + 0.501 + ], + "angle": 0, + "content": "
Exp. MethodKendallSpearmanFID↓@CIFAR-10FID↓@STL-10
IVanilla Gram-matrix24.28%35.75%17.8829.62
Auto-GAS61.14%79.32%12.2123.19
IIAuto-GAS without Transform55.78%73.60%14.2225.58
Auto-GAS without Encoding51.64%68.62%15.2626.49
Auto-GAS without Reduction52.78%71.95%14.6325.74
Auto-GAS without Augment58.40%76.06%13.8625.12
IIIAuto-GAS (Gram-matrix)61.14%79.32%12.2123.19
Auto-GAS (Pearson)60.07%77.53%12.8524.62
Auto-GAS (Hamming)60.58%78.57%12.3523.88
IVAuto-GAS (Proxy Random)29.49%38.25%16.6528.12
Auto-GAS (Proxy Evolution)61.14%79.32%12.2123.19
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.785, + 0.611 + ], + "angle": 0, + "content": "SNIP require gradient backpropagation, leading to increased time consumption during the search process. In contrast, our Auto-GAS only requires a single forward pass, making it faster than methods that necessitate multiple forwards or gradient backpropagation. For image translation, as presented in Table 4, Auto-GAS achieves a \\(110 \\times\\) enhancement compared to GAN Compression [38]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Ablation of proxy search. (1) Our proxy search space includes transform, encoding, reduction, and augmentation operations. Ablations in Table 5 (II) indicate that all these operators contribute, with encoding operators having a greater influence. (3) Ablations for different encoding operators in Table 5 (III) show that the gram-matrix is important, but other operators (e.g., Pearson correlation matrix) also achieve similar performance. (2) Our contribution stems from search space design and proxy search processes. Ablations in Table 5 demonstrate that the various components of our search space contribute. In addition, our proxy evolution search plays an important role in exploring extensive search space and it significantly outperforms random search in Table 5 (IV)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Operations for mutual information estimation. (1) Gram-matrix calculates matrix correlation for mutual information, which measures the salient features and relevant information captured by generative models. The comparison in Table 5 (I) shows that the vanilla Gram-matrix can be used as a proxy for NAS but is weaker than our Auto-GAS. (2) We also compare other operations for fea-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.74, + 0.173 + ], + "angle": 0, + "content": "Table 6: Results of different correlation methods on the CIFAR-10 dataset. analysis on the CIFAR-10 dataset." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.184, + 0.742, + 0.301 + ], + "angle": 0, + "content": "
Correlation MethodsCIFAR-10HyperparameterCIFAR-10
IS↑FID↓IS↑FID↓
Hessian Trace8.3615.47Batch Size (20)8.5212.28
L2norm7.5624.62Batch Size (40)8.0216.52
Entropy7.7025.43Number of Individuals (16)8.4412.21
Variance7.6723.75Number of Individuals (64)8.1116.32
Gram-matrix8.1217.88EA iteration interval (5)7.7818.25
Auto-GAS8.4412.21EA iteration interval (10)8.3813.69
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.784, + 0.39 + ], + "angle": 0, + "content": "ture information estimation in Table. 6. The results of \\( L_{2} \\) norm|Entropy|Variance are worse than Auto-GAS, because they overlooks higher-order feature statistics. Hessian Trace [43] yields relatively promising performance but its computational complexity is excessive." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.452 + ], + "angle": 0, + "content": "Hyperparameter analysis. The results in Table 7 indicate that a larger batch size may speed up convergence but also risk overfitting. An appropriate population size ensures our more efficient architecture exploration. Larger iteration intervals allow for a more comprehensive exploration in our architecture space." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.476, + 0.359, + 0.492 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.785, + 0.675 + ], + "angle": 0, + "content": "In this paper, we present Auto-GAS, a novel training-free generative architecture search framework by evolving customized proxy. We derive from the information bottleneck theory using the latent layer feature as input and building candidate proxies with various transform, encoding, reduction, and augment options. We search for the best proxy using evolutionary algorithms. With customized proxies, our Auto-GAS allows an efficient search for promising architectures. Comprehensive experiment results on image generation and image translation illustrate the promising performance of Auto-GAS. For future work, we will augment our Auto-GAS with advanced methods [31-35,37,52,69]. We hope our novel investigations will provide some insight for the generation and NAS research community." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.676, + 0.785, + 0.737 + ], + "angle": 0, + "content": "**Limitation.** We mainly evaluate the same benchmarks as other NAS methods for fair comparison. We will extend Auto-GAS to more generative models in future work. **Social impact.** Our Auto-GAS focuses on technology improvement without social and ethical implications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.762, + 0.403, + 0.779 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The research was supported by Theme-based Research Scheme (T45-205/21-N) from Hong Kong RGC, and Generative AI Research and Development Centre from InnoHK." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.218, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.207 + ], + "angle": 0, + "content": "1. Abdelfattah, M.S., Mehrotra, A., Dudziak, L., Lane, N.D.: Zero-cost proxies for lightweight nas. arXiv preprint arXiv:2101.08134 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.248 + ], + "angle": 0, + "content": "2. Akhauri, Y., Munoz, J.P., Jain, N., Iyer, R.: EZNAS: Evolving zero-cost proxies for neural architecture scoring. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) NeurIPS (2022), https://openreview.net/forum?id=1SqaDG4dvdt" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.277 + ], + "angle": 0, + "content": "3. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.785, + 0.306 + ], + "angle": 0, + "content": "4. Baker, B., Gupta, O., Naik, N., Raskar, R.: Designing neural network architectures using reinforcement learning. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.334 + ], + "angle": 0, + "content": "5. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.785, + 0.375 + ], + "angle": 0, + "content": "6. Cavagnero, N., Robbiano, L., Pistilli, F., Caputo, B., Averta, G.: Entropic score metric: Decoupling topology and size in training-free nas. In: ICCV workshop (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.404 + ], + "angle": 0, + "content": "7. Chen, K., Yang, L., Chen, Y., Chen, K., Xu, Y., Li, L.: Gp-nas-ensemble: a model for the nas performance prediction. In: CVPRW (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.445 + ], + "angle": 0, + "content": "8. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Fourteenth International Conference on Artificial Intelligence and Statistics (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.488 + ], + "angle": 0, + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.516 + ], + "angle": 0, + "content": "0. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.544 + ], + "angle": 0, + "content": "1. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.586 + ], + "angle": 0, + "content": "2. Dong, P., Niu, X., Li, L., Tian, Z., Wang, X., Wei, Z., Pan, H., Li, D.: Rd-nas: Enhancing one-shot supernet ranking ability via ranking distillation from zero-cost proxies. arXiv preprint arXiv:2301.09850 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.614 + ], + "angle": 0, + "content": "3. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "4. Dong, P., Niu, X., Tian, Z., Li, L., Wang, X., Wei, Z., Pan, H., Li, D.: Progressive meta-pooling learning for lightweight image classification model. In: ICASSP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.698 + ], + "angle": 0, + "content": "5. Duan, Y., Chen, X., Xu, H., Chen, Z., Liang, X., Zhang, T., Li, Z.: Transnas-bench-101: Improving transferability and generalizability of cross-task neural architecture search. In: CVPR. pp. 5251-5260 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.727 + ], + "angle": 0, + "content": "6. Elsken, T., Metzen, J.H., Hutter, F.: Neural architecture search: A survey. arXiv preprint arXiv:1808.05377 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.755 + ], + "angle": 0, + "content": "7. Gao, C., Chen, Y., Liu, S., Tan, Z., Yan, S.: Adversarialnas: Adversarial neural architecture search for gans. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.783 + ], + "angle": 0, + "content": "8. Gong, X., Chang, S., Jiang, Y., Wang, Z.: Autogan: Neural architecture search for generative adversarial networks. ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.811 + ], + "angle": 0, + "content": "9. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. NeurIPS (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. He, H., Wang, H., Lee, G.H., Tian, Y.: Probgan: Towards probabilistic gan with theoretical guarantees. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "22. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "23. Isola, P., Zhu, J., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017). https://doi.org/10.1109/CVPR.2017.632" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.271 + ], + "angle": 0, + "content": "24. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "25. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.785, + 0.325 + ], + "angle": 0, + "content": "26. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.E.: Similarity of neural network representations revisited. In: ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "27. Krizhevsky, A., Nair, V., Hinton, G.: The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.379 + ], + "angle": 0, + "content": "28. Lee, N., Ajanthan, T., Torr, P.H.: Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "29. Li, G., Yang, Y., Bhardwaj, K., Marculescu, R.: Zico: Zero-shot NAS via inverse coefficient of variation on gradients. In: The Eleventh ICLR (2023), https://openreview.net/forum?id=rwo-1s5GqGn" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.447 + ], + "angle": 0, + "content": "30. Li, H., Fu, T., Dai, J., Li, H., Huang, G., Zhu, X.: Autoloss-zero: Searching loss functions from scratch for generic tasks. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "31. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.501 + ], + "angle": 0, + "content": "32. Li, L., Bao, Y., Dong, P., Yang, C., Li, A., Luo, W., Liu, Q., Xue, W., Guo, Y.: Detkds: Knowledge distillation search for object detectors. In: ICML (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "33. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.555 + ], + "angle": 0, + "content": "34. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.582 + ], + "angle": 0, + "content": "35. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.609 + ], + "angle": 0, + "content": "36. Li, L., Sun, H., Dong, P., Wei, Z., Shao, S.: Auto-das: Automated proxy discovery for training-free distillation-aware architecture search. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.636 + ], + "angle": 0, + "content": "37. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.785, + 0.664 + ], + "angle": 0, + "content": "38. Li, M., Lin, J., Ding, Y., Liu, Z., Zhu, J.Y., Han, S.: Gan compression: Efficient architectures for interactive conditional gans. CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.691 + ], + "angle": 0, + "content": "39. Lin, M., Wang, P., Sun, Z., Chen, H., Sun, X., Qian, Q., Li, H., Jin, R.: Zen-nas: A zero-shot nas for high-performance image recognition. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "40. Lin, Q., Fang, Z., Chen, Y., Tan, K.C., Li, Y.: Evolutionary architectural search for generative adversarial networks. IEEE Transactions on Emerging Topics in Computational Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.772 + ], + "angle": 0, + "content": "41. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "42. Liu, L., Zhang, Y., Deng, J., Soatto, S.: Dynamically grown generative adversarial networks. AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "43. Liu, Y., Yu, S., Lin, T.: Hessian regularization of deep neural networks: A novel approach based on stochastic estimators of hessian trace. Neurocomputing 536, 13-20 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.218, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "44. Luo, Y., Zhang, Y., Cai, X., Yuan, X.: E2gan: End-to-end generative adversarial network for multivariate time series imputation. In: 28th international joint conference on artificial intelligence. pp. 3094-3100. AAAI Press Palo Alto, CA, USA (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.23 + ], + "angle": 0, + "content": "45. Mellor, J., Turner, J., Storkey, A., Crowley, E.J.: Neural architecture search without training. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.231, + 0.786, + 0.257 + ], + "angle": 0, + "content": "46. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. ArXiv abs/1802.05957 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.258, + 0.786, + 0.284 + ], + "angle": 0, + "content": "47. Odena, A., Olah, C., Shlens, J.: Conditional image synthesis with auxiliary classifier gans. In: ICML. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.786, + 0.311 + ], + "angle": 0, + "content": "48. Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.312, + 0.786, + 0.338 + ], + "angle": 0, + "content": "49. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.339, + 0.786, + 0.366 + ], + "angle": 0, + "content": "50. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR abs/1511.06434 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.367, + 0.786, + 0.392 + ], + "angle": 0, + "content": "51. Real, E., Liang, C., So, D.R., Le, Q.V.: Automl-zero: Evolving machine learning algorithms from scratch (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.393, + 0.786, + 0.433 + ], + "angle": 0, + "content": "52. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.434, + 0.786, + 0.46 + ], + "angle": 0, + "content": "53. Shu, H., Wang, Y., Jia, X., Han, K., Chen, H., Xu, C., Tian, Q., Xu, C.: Coevolutionary compression for unpaired image translation. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.786, + 0.487 + ], + "angle": 0, + "content": "54. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information. arXiv:1703.00810 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.786, + 0.514 + ], + "angle": 0, + "content": "55. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information. arXiv preprint, arXiv:1703.00810 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.515, + 0.786, + 0.541 + ], + "angle": 0, + "content": "56. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.542, + 0.786, + 0.568 + ], + "angle": 0, + "content": "57. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.786, + 0.595 + ], + "angle": 0, + "content": "58. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS 33, 6377-6389 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.597, + 0.786, + 0.636 + ], + "angle": 0, + "content": "59. Tian, Y., Shen, L., Su, G., Li, Z., Liu, W.: Alphagan: Fully differentiable architecture search for generative adversarial networks. arXiv preprint arXiv:2006.09134 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.637, + 0.786, + 0.677 + ], + "angle": 0, + "content": "60. Tian, Y., Wang, Q., Huang, Z., Li, W., Dai, D., Yang, M., Wang, J., Fink, O.: Off-policy reinforcement learning for efficient and effective gan architecture search. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.786, + 0.704 + ], + "angle": 0, + "content": "61. Tishby, N., Zaslavsky, N.: Deep learning and the information bottleneck principle. In: ITW. pp. 1-5. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.705, + 0.786, + 0.731 + ], + "angle": 0, + "content": "62. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.732, + 0.786, + 0.758 + ], + "angle": 0, + "content": "63. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.786, + 0.785 + ], + "angle": 0, + "content": "64. Wang, C., Xu, C., Yao, X., Tao, D.: Evolutionary generative adversarial networks. IEEE Transactions on Evolutionary Computation 23(6), 921-934 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.786, + 0.812 + ], + "angle": 0, + "content": "65. Wang, H., Huan, J.: Agan: Towards automated design of generative adversarial networks. arXiv preprint arXiv:1906.11080 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "66. Wang, W., Sun, Y., Halgamuge, S.: Improving MMD-GAN training with repulsive loss function. In: ICLR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Lujun Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "67. Wei, Z., Pan, H., Li, L.L., Lu, M., Niu, X., Dong, P., Li, D.: Convformer: Closing the gap between cnn and vision transformers. arXiv preprint arXiv:2209.07738 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "68. Wolchover, N., Reading, L.: New theory cracks open the black box of deep learning. Quanta Magazine (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "69. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.273 + ], + "angle": 0, + "content": "70. Ying, G., He, X., Gao, B., Han, B., Chu, X.: Eagan: Efficient two-stage evolutionary architecture search for gans. In: ECCV. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "71. Yu, A., Grauman, K.: Fine-grained visual comparisons with local learning. CVPR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "72. Zhang, M., Li, H., Pan, S., Chang, X., Su, S.: Overcoming multi-model forgetting in one-shot nas with diversity maximization. In: CVPR. pp. 7809-7818 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "73. Zhong, Z., Yang, Z., Deng, B., Yan, J., Wu, W., Shao, J., Liu, C.L.: Blockqnn: Efficient block-wise neural network architecture generation. IEEE transactions on pattern analysis and machine intelligence (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.398 + ], + "angle": 0, + "content": "74. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "75. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.467 + ], + "angle": 0, + "content": "76. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.467 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/d262e2a7-fb7b-49f5-8763-ff8884371602_origin.pdf b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/d262e2a7-fb7b-49f5-8763-ff8884371602_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bf7f9c104d4a0270012f3dbda9d354d472b3ab65 --- /dev/null +++ b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/d262e2a7-fb7b-49f5-8763-ff8884371602_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3311987634a6c5fda286fd005575e26502f2675834079e17f9c89d0df4bb879 +size 5658251 diff --git a/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/full.md b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0ead61b711c415604879f6253490b77d5166d361 --- /dev/null +++ b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/full.md @@ -0,0 +1,325 @@ +# Auto-GAS: Automated Proxy Discovery for Training-free Generative Architecture Search + +Lujun Li $^{1}$ , Haosen Sun $^{1}$ , Shiwen Li $^{2}$ , Peijie Dong $^{3}$ , Wenhan Luo $^{1}$ , Wei Xue $^{1}$ , Qifeng Liu $^{1\boxtimes}$ , and Yike Guo $^{1\boxtimes}$ + +1 The Hong Kong University of Science and Technology + +$^{2}$ Xidian University + +3 The Hong Kong University of Science and Technology (Guangzhou) lilujunai@gmail.com;hsunas@connect.ust.hk;shiwenlisw@gmail.com; pdong212@connect.hkust-gz.edu.cn;whluo.china@gmail.com;weixue@ust.hk; liuqifeng@ust.hk;yikeguo@ust.hk + +Abstract In this paper, we introduce Auto-GAS, the first training-free Generative Architecture Search (GAS) framework enabled by an autodiscovered proxy. Generative models like Generative Adversarial Networks (GANs) are now widely used in many real-time applications. Previous GAS methods use differentiable or evolutionary search to find optimal GAN generators for fast inference and memory efficiency. However, the high computational overhead of these training-based GAS techniques limits their adoption. To improve search efficiency, we explore training-free GAS but find existing zero-cost proxies designed for classification tasks underperform on generation benchmarks. To address this challenge, we develop a custom proxy search framework tailored for GAS tasks to enhance predictive power. Specifically, we construct an information-aware proxy that takes feature statistics as inputs and utilizes advanced transform, encoding, reduction, and augment operations to represent candidate proxies. Then, we employ an evolutionary algorithm to perform crossover and mutation on superior candidates within the population based on correlation evaluation. Finally, we perform generator search without training using the optimized proxy. Thus, Auto-GAS enables automated proxy discovery for GAS while significantly accelerating the search before training stage. Extensive experiments on image generation and image-to-image translation tasks demonstrate that Auto-GAS strikes superior accuracy-speed tradeoffs over state-of-the-art methods. Remarkably, Auto-GAS achieves competitive scores with $110 \times$ faster search than GAN Compression. Code at: https://github.com/lliai/Auto-GAS. + +Keywords: Generative Model $\cdot$ Generative Architecture Search $\cdot$ AutoML + +Table 1: Comparison of search cost for generative architecture search methods. + +
MethodAlgorithm Cost (GPU-days)MethodAlgorithm Cost (GPU-days)
AGAN [65]RL1200DGGAN [42]Heuristic580
AutoGAN [18]RL2EAS-GAN [40]EA1
AlphaGAN [59]Gradient1.2EGAN [64]EA1.3
EAGAN [70]EA0.8Auto-GAS (ours)Training-free0.09
+ +# 1 Introduction + +Generative models like Generative Adversarial Networks (GANs) [19] have made tremendous strides in visual generation tasks such as image and video synthesis [5,50]. However, their demanding computational footprint presents barriers to deploying these models on edge devices with strict resource constraints [25,47]. To overcome this, there is growing interest in model compression techniques [7, 12, 14, 49, 67]. Among them, Neural Architecture Search (NAS) [16] automates architecture design, finding compact generative models for edge computing, and enabling flexible structures without pre-trained weights. In contrast, other compression methods are limited to fixed, handcrafted architectures and rely on pre-trained models. The integration of NAS in generative models enables real-time visual creation on edge devices, unlocking new interactive experiences in domains like augmented reality [5,24,50]. + +Conventional Generative Architecture Search (GAS) approaches typically rely on expensive training-based training routes to assess candidate architectures, such as maintaining weight-sharing supernets [13, 22] or conducting multiple training trials from scratch [56, 76]. However, training generative models presents unique obstacles that greatly increase the costs and difficulties of applying these techniques for GAS. The training processes for generative models are generally more complex than discriminative tasks because inherent instabilities (e.g., mode collapse [20]) can occur during optimization. In mode collapse, the model generates only a limited variety of examples from the underlying distribution. These instabilities make it challenging to reliably evaluate architectural candidates with standard NAS training-based evaluations. As shown in Table 1, gradient-based GAS methods like AdversarialNAS [17] and AlphaGAN [59] amplify these problems by jointly searching the generator and discriminator architecture spaces. This coupled search increases the overall complexity and fragility of the optimization procedure. To address these difficulties, DGGAN [42] introduced a progressive search approach where candidates are trained incrementally. However, its process still requires a massive 580 GPU days to complete due to the computationally intensive nature of generative model training. + +Recently, training-free NAS approaches have demonstrated remarkable efficiency using zero-cost proxies to predict network performance without training network parameters. These methods are mainly present for encoder-only models for classification tasks. Inspired by their computational efficiency, we comprehensively evaluate existing proxies when applied to generative tasks. As shown + +![](images/9e5ba8747072f801b0fc63fbbb0149f87a792a36743ff05d30f7ef192733802b.jpg) +Figure 1: Correlation visualization of proxies (from left to right: Param., Synflow, NWOT, and Auto-GAS) in Generation benchmarks [15]. + +![](images/15e402b01e4469643860bc640273e7639ca896db14da9a9f94e5b5f04ad58517.jpg) + +![](images/80a0b71a52514bc63f8fafad1f12f208c2b8be34a69383e654b60e369ba31a3d.jpg) + +![](images/2bd49668b642029ed754f2efa63851b39b2931d4e1271f953417e0de625ba727.jpg) + +in Figure 1, we can make some interesting observations: (1) We observe that parameter-level proxies (e.g., Synflow [57], SNIP [28], and GraSP [62]), which estimate performance based on weight importance, generally perform poorer when ranking generator candidates. This indicates their assumptions regarding how architecture impacts learning do not directly translate to generative objectives and training dynamics. (2) Architecture-level proxies (e.g., NWOT [45] and Zen [39]) featuring topological encodings demonstrate stronger ranking ability. However, even the best architecture proxies leave room for improvement to enable performing generative NAS at scale. These findings motivate developing customized proxies that more precisely capture the unique characteristics of visual generation. Thus, a question naturally arises: how to design customized proxies for generation tasks? + +To answer this question, we present Auto-GAS, the first automated generative architecture search framework that aims to efficiently utilize custom training-free proxies. Firstly, we derive that the mutual information of latent features within generative models is strongly coupled with task optimization objectives based on information bottleneck theory [3,54,61]. Then, we develop proxy search strategies to better estimate the mutual information of these features. Specifically, we build comprehensive operation search spaces and evolutionary algorithms to search optimal information-aware proxies. To better process feature statistics, our search space involves transform-based, encoding-based, reduction-based, and augment-based operators. The search process in Auto-GAS begins with initializing the population of proxies and then evaluates, crosses, and varies them to improve the quality of the proxies. Throughout the search, we directly utilize the ranking correlation with accuracy results in the benchmark as fitting objectives, aiming to find a more suitable proxy for the generation tasks. With this evolutionary process, Auto-GAS outperforms existing training-free NAS approaches by significant margins without manual tuning. Additionally, we search for generator architectures using Auto-GAS and subsequently implement a full training process on the searched generator. + +We perform extensive experiments to validate the performance and efficiency strengths of our Auto-GAS in generation tasks like image generation and imaged-to-image translation. The experiments demonstrate that our Auto-GAS can achieve better-generating performance than other zero-shot proxies and state-of-the-art results on rank consistency. In contrast to handcrafted GAS methods, + +our Auto-GAS framework enjoys many benefits: (1) Automatic. It enables the automatic discovery of more expressive and efficient proxies that human expert might ignore, which can reduce human bias and ensures that the resulting architectures are optimized for the target problem or dataset. (2) General. The search procedure of Auto-GAS improves generalization by discovering proxies that generalize well on different search spaces. Based on the final results for the search objectives, our framework can evolve adaptive proxies to generalize across architectures. (3) Insightful. Our framework allows for the automatic discovery of more expressive, efficient, and effective proxies for optimizing and customizing generation architectures, opening up new research directions for automated GAN architecture search. Our contributions can be summarized as follows: + +- We present Auto-GAS, the first training-free architecture search framework for generation tasks. Auto-GAS achieves an ultra-efficient and elegant search process with our auto-searched customized proxy. +- We propose a well-organized proxy search space, including latent features as input and various transforms, encoding, reduction, and augment operations as options. We develop evolutionary algorithms for effective & efficient proxy search. +- We conduct extensive experiments on image generation and image-to-image translation tasks. Auto-GAS obtains significant search acceleration and competitive performance on multiple datasets. + +# 2 Related Work + +Generative architecture search. The field of Neural Architecture Search (NAS) traditionally encompasses the design of exploration spaces, search methodologies, and assessment techniques to autonomously uncover optimal architectural configurations within specified parameters. Empirical approaches utilize an iterative process of training followed by exploration, either through multiple iterations [4,73] or by employing weight-sharing strategies [41,72]. NAS has also been extensively applied to the domain of GANs Architecture Search (GAS). The primary goal in GAS is to automatically identify superior GAN structures that enhance generative performance. Pioneering reinforcement learning-based GAS techniques, such as AGAN [65] and AutoGAN [18], leverage RNN controllers to construct GAN architectures and employ the Inception Score (IS) as their evaluation criterion. E2GAN [60] and AdversarialNAS [17] implement alternative reward functions. EAGAN [70] conducts discrete searches and enhances the robustness of the GAN exploration process. Nevertheless, these training-dependent search methodologies continue to be hampered by inefficient exploration and substantial computational demands. To tackle this challenge, we propose a training-free approach, grounded in architecture decoupling, to enhance the discovery of competitive GAN structures in a more computationally efficient manner. + +Training-free architecture search. Zero-shot architecture exploration [10, 11, 29, 39], alternatively referred to as training-free NAS, offers a swifter alternative to conventional NAS methods. This approach's evaluation process en + +tails only a handful of forward or backward passes with a small batch of input data, rendering it highly efficient. Zero-cost indicators can be categorized into parameter-centric and architecture-centric proxies. Parameter-centric zero-cost proxies (e.g., Synflow [57], SNIP [28], and GraSP [62]) rely on pruning techniques and aggregate the saliency values of each layer's weights as a proxy. For instance, Synflow [58] introduces a modified version of synaptic saliency scores to prevent layer collapse during parameter pruning, while Fisher [1] calculates the sum of all activation gradients in the network, applicable to channel pruning. Grasp [63] ensures gradient flow preservation throughout the network for efficient training. Their detailed formulations are as follows: + +$$ +\rho_ {s n i p} = \left| \frac {\partial \mathcal {L}}{\partial \mathcal {W}} \odot \mathcal {W} \right|, \rho_ {s y n f l o w} = \frac {\partial \mathcal {L}}{\partial \mathcal {W}} \odot \mathcal {W}, \rho_ {f i s h e r} = \left(\frac {\partial \mathcal {L}}{\partial \mathcal {A}} \mathcal {A}\right) ^ {2}, \rho_ {g r a s p} = - (H \frac {\partial \mathcal {L}}{\partial \mathcal {W}}) \odot \mathcal {W}, \tag {1} +$$ + +where $\mathcal{L}$ , $\mathcal{W}$ , $\mathcal{A}$ are loss function, weight and activation. $H$ is the Hessian matrix. However, the complexity of loss function $\mathcal{L}$ in generation tasks renders these gradient-based methods ineffective. Regarding architecture-centric proxies, ZenNAS [39] introduced the novel Zen-Score to assess network expressivity, while NWOT [45] examined correlations between linear maps induced by data points for untrained architectures. Nevertheless, these approaches employ hand-crafted fixed proxies for classification tasks and encoder-only CNN models. In generation tasks, the encoder-decoder generator typically features intricate branches and multi-scale features, posing significant challenges for traditional proxies in predicting final accuracy. To address this issue, we automatically optimize the first customized proxies for GAS. Furthermore, our framework differs from EZNAS-like methods [2, 36, 74, 75] in several aspects: (1) Tasks. EZNAS is exclusively designed for classification, whereas our Auto-GAS is tailored for generation tasks. (2) Search algorithms. EZNAS merely employs genetic programming, while we propose a proxy evolution search strategy that substantially enhances search efficiency. (3) Model statistic inputs. Our method solely utilizes features as input, as we observed poor performance of parameter-level proxies. Conversely, EZNAS selects parameters and gradients as inputs. (4) Search space. Our Auto-GAS incorporates numerous feature-related operations absent in EZNAS. + +# 3 Methodology + +In this section, we first review existing proxies. Then, we present formulations and insights of our proxy design (see Figure 2). Finally, we illustrate the detailed search process. + +# 3.1 Proxy Search Space + +Motivation of our customized proxy design. Generative tasks are relevant to the Information Bottleneck (IB) theory [3, 54, 61], which guides generative models to learn compact representations that retain important characteristics for accurate sample generation. In generative tasks, the IB theory helps understand + +![](images/3412afae11f3dd45eb17afde1ccc179e4ae3df5c2d14571ee9aa87a9ac4c6ef6.jpg) +Figure 2: Overall proxy search of Auto-GAS. We extract the latent features of the generator with input data forward-propagation. Then we select various transform, encoding, reduction, and augment operators from the proxy search space to build candidate proxies. Finally, we perform a ranking evaluation to cross and mutate to get the best proxy. + +how generative models can balance compression (capturing essential features) and reconstruction (faithful generation of samples) [55, 68]. IB theory suggests maximizing the Mutual Information (MI) $I(\mathbf{X}; \mathbf{F}_i)$ between the input data $\mathbf{X}$ and the corresponding latent representation $\mathbf{F}_i$ can be expressed as: + +$$ +I (\mathbf {X}; \mathbf {F} _ {i}) = H (\mathbf {F} _ {i}) - H (\mathbf {F} _ {i} | \mathbf {X}) \tag {2} +$$ + +where $H(\mathbf{F}_i)$ stands for the entropy of the latent representation $\mathbf{F}_i$ and $H(\mathbf{F}_i|\mathbf{X})$ denotes the conditional entropy. Considering that neural networks are deterministic, and the dataset has a finite number of instances, the conditional entropy becomes zero, simplifying the mutual information (MI) estimation as $I(\mathbf{X};\mathbf{F}_i) = H(\mathbf{F}_i)$ . Consequently, some information coding and compression studies [26, 55] propose some operations (e.g., Entropy and Gram-matrix) to approximate the mutual information on latent representation. These inspired us to develop our train-free method design thought chain: assess generative model $\rightarrow$ estimate MI $\rightarrow$ design expressions via Entropy/Gram-matrix ops $\rightarrow$ our train-free proxy. However, how about designing this potential expressions? Recent Auto-Zero approaches [30, 51] present a promising solution: searching for the best machine learning solution from scratch. In addition, several studies [6] have shown that normalize and activation operations can enhance the formulation. Therefore, we automatically search for optimal proxies based on these operators. + +Proxy primitive operations. Our approach evolves the customized proxy in the generation benchmark [15], which consists of U-Net-like encoder-decoder models with various operators and their training results. We extract the latent features of each layer in both the encoder and decoder modules as inputs to the search process and represent the zero-cost proxy as a computation graph. This graph is constructed as transform $\rightarrow$ encoding $\rightarrow$ reduction $\rightarrow$ augment + +parts in processing generator statistics. The available primitive operations can be summarized as follows: + +- Transform operations: These operations involve different activation functions and feature processing techniques to enhance raw features and represent architectural potentials. Examples of transform operations include relu, swish, mish, scale,norm, exp, and softmax. +- Encoding operations: Inspired by coding algorithms and feature information estimation, we employ binary and matrix-based operations in the encoding part. These operators help us better estimate the mutual correlation information of samples. Some encoding operations include sign, above-zero, above-median, gram-matrix, person, and hamming. +- Reduction operations: Reduction operators bridge the information matrix and the proxy values. They are valuable for extracting useful matrix information from different mathematical perspectives. Reduction operations include sloqdet, abslog, trace, and eigenvalue. +- Augment operations: Augment operators are applied to enhance the proxy values' ability to predict performance results. These operations include arithmetic functions such as abs, pow, l2-norm, log, sum, and mean. + +By utilizing these primitive operations, our approach enables the evolution of the customized proxy, facilitating the search for optimal solutions in the generation benchmark. These mathematical operations are essential building blocks to construct a diverse and expressive search space of zero-cost proxy. We provide more details of their formulas and properties in the supplementary material. + +# 3.2 Proxy Evolution Search + +Evolutionary process and goal. Utilizing our proxy search space and graph representation, we employ an evolutionary algorithm (EA) outlined in Alg. 1 to discover optimal proxy expressions. Our EA commences with an initial set of candidate proxies and progressively refines the population across generations, employing genetic operators such as selection, crossover, and mutation to yield superior solutions. For fitness evaluation, each proxy is assessed based on the ranking correlation between its output proxy $\mathcal{Q}$ scores and pre-determined performance benchmarks $\mathcal{D}$ , enabling efficient identification of the optimal proxy $\rho^{*}$ within the search space $\mathcal{P}$ , as follows: + +$$ +\rho_ {A u t o - G A S} ^ {*} = \underset {\rho \in \mathcal {P}} {\arg \max } (\tau (\mathcal {D}, \mathcal {Q})). \tag {3} +$$ + +where Kendall's Tau $\tau$ is formulated as the correlation coefficient. Each evolution picks top-k candidates with the highest evaluation score and randomly selects a parent from the candidates for mutation. + +Efficient proxy search with existing generation benchmark. It is expensive to derive pre-trained results $\mathcal{D}$ , we adopt an alternative solution to achieve proxy search using pre-computed architecture-performance pairs provided in + +Algorithm 1 Evolutionary Search for Auto-GAS Proxy +Input: Exploration domain $S$ , candidate set $\mathcal{P}$ , maximum iterations $\mathcal{T}$ , sampling ratio $r$ , sampled pool $\mathcal{R}$ , top $k$ selections. +Output: Auto-GAS proxy with highest Kendall's Tau $\tau$ coefficient. +1: $\mathcal{P}0 := \text{Initialize candidate set}(P_i)$ ; +2: Sampled pool $\mathcal{R} := \emptyset$ ; +3: for $i = 1, 2, \ldots, T$ do +4: Select $G_i^s := \text{SelectTopk}(\mathcal{R}, k)$ ; +5: // greedy evolutionary strategy +6: Empty sampled pool $\mathcal{R} := \emptyset$ ; +7: Mutate := MUTATE( $G_i^s$ ); +8: Crossover $G_i^c := \text{CROSSOVER}(\mathcal{G}_i^s)$ ; +9: Incorporate $G_i^m$ into $R$ ; +10: Incorporate $G_i^c$ into $R$ ; +11: if $\mathcal{R} < \mathcal{P}$ then +12: Introduce random samples +13: else +14: Proceed to line 4; +15: end if +16: end for + +the existing TransNAS-Bench-101 [15] benchmarks. Specifically, this benchmark involves Autoencoding task with 7k different generators in macro-level search space and their evaluation metric (i.e., SSIM). Each generator network follows an encoder-decoder structure in Pix2Pix, where the encoders are the searched backbones, and the decoders contain 14 layers of convolution and deconvolution. Thanks to this existing benchmark, we do not need to train performance groundtruths again, and it only takes 1.5 hours in our proxy search process. In light of excellent generality (proven in our extensive experiments), our searched proxy $\rho_{Auto-GAS}$ can be well transferred to similar search spaces and can be directly used in most scenarios without additional search budget. + +# 3.3 Discovered Proxy Analysis + +Searched Zero-cost Proxy. We present formulas of the searched proxies on the generation benchmark as follows: + +$$ +\rho_ {A u t o - G A S} ^ {*} = \left(\operatorname {s l o g d e t} \left([ \mathbf {F} ^ {\prime}, \mathbf {F} ^ {\prime T} ]\right)\right) ^ {2}, \mathbf {F} ^ {\prime} = \operatorname {a b o v e - m e a n} (\operatorname {s w i s h} (\mathbf {F})) \qquad (4) +$$ + +where $\mathbf{F}$ represents the original features, and transformed features $\mathbf{F}'$ are obtained by applying the swish operation followed by the above-mean operation to the original features $\mathbf{F}$ . The $[\mathbf{F}',\mathbf{F}'^T]$ indicates the $N\times N$ gram-matrix of the transformed features $\mathbf{F}'$ . The proxy value $\rho^{*}$ is calculated by taking the square of the slogdet of the concatenated gram-matrix $[\mathbf{F}',\mathbf{F}'^T]$ . + +Analysis of searched proxy. The discovered formulations indicate that (1) Mathematical operations, such as swish activation and slogdet, significantly con- + +Table 2: Ranking results on TransNAS-Bench-101 [15]. + +
MethodKendallSpearmanPearsonMethodKendallSpearmanPearson
SNIP [1]-2.04%-3.55%-0.26%ZiCo [29]18.63%15.53%21.13%
EZNAS [2]23.30%25.91%32.03%Synflow [57]-50.91%-65.01%-49.17%
NWOT [45]51.28%63.94%48.88%Params.-12.97%-15.24%-5.80%
SNIP [28]30.81%23.31%29.34%FLOPs48.91%64.01%52.36%
Zen [39]-2.27%2.58%5.39%Auto-GAS61.14%79.32%81.25%
+ +tribute to the predictability of the proxy. These operations enhance the proxy's ability to capture important features and characteristics of the generator. (2) Gram-Matrix and Mutual Information: The expression utilizes a sample-wise gram-matrix, which accurately estimates the latent mutual information of the generator. The gram-matrix captures the relationships and dependencies among the transformed features, providing valuable insights into the generator's behaviour. The discovered proxy of Auto-GAS, as illustrated in Figure 1, showcases a substantial improvement in ranking. This improvement suggests that the searched proxy effectively captures essential characteristics of the generator, leading to enhanced performance in the benchmark. + +Understanding why our Auto-GAS surpasses other NAS methods. (1) Many train-based NAS methods employ weight-sharing and other proxy settings, which often lead to inaccurate estimations of architecture performance. In contrast, our Auto-GAS method searches for proxies based on actual performance, resulting in a proxy that demonstrates strong ranking capability. (2) Other train-free NAS methods are primarily designed for classification rather than generative tasks. Consequently, these methods search for architectures using weak correlation proxies. In contrast, our proxy exhibits a stronger correlation, ensuring the discovery of more efficient models. + +# 3.4 Training-free Generative Models Search + +Auto-GAS has made efforts to search generative Models in a non-training way to ease the instability of GAN training and improve search efficiency. In this stage, we directly search the generator in $\rho_{Auto - GAS}$ . In each iteration, we generate $P$ candidate generators to make up the population $\mathcal{A}$ , where all these candidates use the randomly initialized weights $W$ . Each candidate is characterized by both architecture and weight variables. In each iteration, we evaluate $P$ candidates without training. We employ the information-aware proxy to assign scores to all these candidates. The formulation of the search stage can be succinctly stated as follows: + +$$ +\alpha_ {A u t o - G A S} ^ {*} = \arg \max _ {\alpha \in \mathcal {A}} \rho_ {A u t o - G A S} ^ {*} (\alpha , \mathcal {W}). \tag {5} +$$ + +Note that in scenarios involving discriminator searches, we also use $\rho_{Auto - GAS}$ for searching the discriminator architectures with real sample inputs. + +Table 3: Comparison results on the CIFAR-10 and STL-10 datasets. We use FID (the lower the better) and the mIoU metric (the higher the better) for evaluations. + +
MethodAlgorithmGPU daysCIFAR-10 (32×32)STL-10 (48×48)
IS↑FID↓IS↑FID↓
WGAN-GP [20]7.8629.3--
SN-GAN [46]8.2221.709.1640.1
ProbGAN [21]Manual-7.7524.608.8746.74
Improv MMD GAN [66]8.2916.219.3437.63
BigGAN [5]9.2214.73--
AGAN [65]12008.2930.509.2352.70
AutoGAN [18]RL2.008.5512.429.1631.01
AlphaGAN [59]1.28.7015.56--
AdversarialNAS [17]1.007.8624.048.5238.85
EGAN [64]1.256.90---
EAS-GAN [40]EA1.007.4533.20-38.84
EAGAN [70]0.88.4112.839.6923.82
Fisher [1]0.128.0317.669.3729.82
Zen [39]0.118.0215.999.4528.66
SNIP [28]Training-free0.127.9616.679.5829.25
NWOT [45]0.098.0516.329.2327.46
Auto-GAS0.098.4412.219.7223.19
+ +# 3.5 Generative Models Training + +After the architecture searches, we train generative models with randomly selected data samples. Simultaneously, a batch of noise samples is generated from a noise distribution. The generator tries to generate data from random noise $z$ as $G(z)$ , which should be as close as possible to the real data distribution. The discriminator tries to distinguish the generated data from the real data. The value function $V(D,G)$ for a minimax game is defined as follows: + +$$ +\begin{array}{l} \min _ {G} \max _ {D} V (D, G) = \mathbb {E} _ {\mathbf {x} \sim p _ {d a t a} (\mathbf {x})} [ \log D (\mathbf {x}) ] \\ + \mathbb {E} _ {\mathbf {z} \sim p _ {z} (\mathbf {z})} [ \log (1 - D (G (\mathbf {z})) ] \tag {6} \\ \end{array} +$$ + +where $\mathbf{x} \sim p_{data}(\mathbf{x})$ denotes the distribution of real data, $\mathbf{z} \sim p_z(\mathbf{z})$ denotes the input noise to the generator. + +# 4 Experiments + +In this section, we first perform proxy search experiments on the generation benchmark, and then we conduct architecture search experiments on image generation and image-to-image translation tasks by translocating the searched proxies. Finally, we provide a detailed analysis of our approach. + +# 4.1 Experiments on Generation Benchmark + +Dataset and implementation. We adopt the search space and performance ground truth provided by the Autoencoding task in TransNAS-Bench-101 [15] as + +![](images/712476a5a6d6d7106f04f3db0b10a8be41fead19bfe73ca06da7ca54ced3c628.jpg) +Figure 3: Searched model on CIFAR-10. + +![](images/4176af95007f0863eee020d9a23fc5e5e3b9b360d20a7e1a9e97d5e3081ee716.jpg) +Figure 4: Searched model on STL-10. + +the benchmark for searching and evaluating GAS proxies. We split the 4:3 ratio of the benchmark into the validation and test sets. During the proxy search phase, we set $(\mathcal{P},\mathcal{T},r,k)$ in Alg. 1 as (20,100,0.9,5) and evaluate with 50 architectures in the validation set. After the search, we fairly compare the Auto-GAS proxy with the traditional ones by randomly sampling 200 architectures in the test set. Note that our validation and test sets do not overlap to ensure fair proxy comparisons. Detailed implementations are in the Appendix. + +Comparison results. Table 5 demonstrates the ranking capabilities of different methods on Generation Benchmark. Auto-GAS (ours) achieves the best rankings across all three measures: Kendall's Tau $(61.14\%)$ , Spearman $(79.32\%)$ , and Pearson $(81.25\%)$ . These results indicate that Auto-GAS performed well in preserving the rankings compared to other methods. These findings highlight the effectiveness of Auto-GAS for generating meaningful and accurate rankings. + +# 4.2 Experiments on Image Generation + +Datasets and implementation. We perform evaluations using CIFAR-10 [27] and STL-10 [8] datasets, aligning with other GAS methodologies [17, 18, 70]. CIFAR-10 comprises 50k training and 10k test $32 \times 32$ images, while STL-10 contains 100,500 images at a higher $48 \times 48$ resolution. We directly apply searched proxies to identify optimal architectures within EAGAN's [70] search space. The exploration process spans 150 epochs with a $P = 32$ individual population. Generator and discriminator batch sizes are 80 and 40, respectively. We utilize 16 randomly generated Gaussian noise images for proxy score computation to expedite evaluation. Post-search, we conduct full training of the best-scoring generator for 500 epochs. For evaluation, following previous GAS works [17, 18, 70], we generate 50,000 images to calculate IS and FID metrics. Figures 3 and 4 illustrate the discovered architectures for CIFAR-10 and STL-10. + +Training-based NAS method comparison. We benchmark Auto-GAS against various training-based NAS approaches: BigGAN [5], E2GAN [44], Adversarial-NAS [17], and EAGAN [70]. As shown in Table 3, Auto-GAS demonstrates superior efficiency. For CIFAR-10, Auto-GAS achieves lower FID than BigGAN and comparable overall results to E2GAN, while outperforming AdversarialNAS in both FID and IS. On STL-10, Auto-GAS significantly outperforms E2GAN and AdversarialNAS in both metrics, rivaling EAGAN in IS and surpassing it in FID. In summary, Auto-GAS emerges as an efficient and competitive approach, either surpassing or matching several state-of-the-art training-based NAS methods across diverse evaluation metrics and datasets. + +Zero-shot proxy comparison. Table 3 also juxtaposes the Auto-GAS proxy with four alternative zero-shot proxies. Fisher and SNIP, relying on gradient- + +Table 4: Image translation evaluation on horse→zebra and Cityscapes datasets. + +
ModelDatasetMethod#ParametersGPU daysFID (↓)mIoU (↑)
ValueRatioValueRatioValueIncreaseValueDrop
CycleGANhorse→zebra(512×512)Original11.4M---61.53--
COEV [53]----96.1534.6-
GAN Comp. [38]0.34M33.3×11.601.00×71.8110.3-
Fast GAN Comp. [38]0.36M32.1×3.123.72×65.193.66-
Auto-GAS0.39M29.3×0.11105.45×63.662.13-
GauGANCityscapes(512×512)Original93.0M---57.60-62.18-
GAN Comp. [38]20.4M4.6×17.701.00×55.192.4161.220.96
Fast GAN Comp. [38]20.2M4.6×10.401.70×56.251.3561.171.01
Auto-GAS21.0M4.5×0.16110.63×56.041.5661.210.97
+ +based computations, yield similar scores. Zen-Score and NWOT exhibit positive correlations with model accuracy, outperforming gradient-based methods. These results validate the Auto-GAS proxy's efficacy, positioning it as a formidable contender among zero-shot proxy methods. + +# 4.3 Experiments on Image-to-image Translation. + +Datasets and implementation. We comprehensively evaluate our approach on image translation tasks using high-resolution datasets, including [9] and horse $\rightarrow$ zebra [71] at $512\times 512$ resolutions. We apply the proxy for architecture exploration, adopting GAN Compression's [38] search space for fair comparisons, focusing on channel and layer number configurations. We evolve to compress Pix2Pix [23] on horse $\rightarrow$ zebra and GauGAN [48] on cityscapes. The search process involves 500 iterations, followed by model retraining with a 0.0002 learning rate and 16 batch size, adhering to the original paper [23]. + +GAN compression method comparison. Table 4 demonstrates that Auto-GAS achieves the highest FID score of 63.66 for CycleGAN with a superior compression ratio $(29.3\times)$ compared to alternatives. This indicates Auto-GAS effectively reduces model size while preserving image quality. For GauGAN, Auto-GAS exhibits a marginally higher FID score and a competitive mIoU score of 61.22. + +Qualitative Analysis. Figure 5 illustrates Auto-GAS's proficiency in translation tasks. In contrast, FastGAN Compression not only yields a lower FID score but also fails to eliminate artifacts like brown discoloration in zebra models. Conversely, Auto-GAS-generated zebra images accurately capture authentic black-and-white zebra patterns. + +# 4.4 Ablation Study + +Search efficiency benefits. For image generation (see Table 3), our Auto-GAS method exhibits significantly superior search efficiency to training-based NAS GANs and has a definite advantage over other zero-cost proxies. The Fisher and + +![](images/b5866431c6aa2827cf85ca6c855adf63100b90989d9a48ff6cd067154a3d236d.jpg) +Figure 5: Visualization results on image translation tasks using high-resolution datasets, e.g., horse→zebra and Cityscapes with $512 \times 512$ resolution. + +Table 5: Ablations of our proxy search. + +
Exp. MethodKendallSpearmanFID↓@CIFAR-10FID↓@STL-10
IVanilla Gram-matrix24.28%35.75%17.8829.62
Auto-GAS61.14%79.32%12.2123.19
IIAuto-GAS without Transform55.78%73.60%14.2225.58
Auto-GAS without Encoding51.64%68.62%15.2626.49
Auto-GAS without Reduction52.78%71.95%14.6325.74
Auto-GAS without Augment58.40%76.06%13.8625.12
IIIAuto-GAS (Gram-matrix)61.14%79.32%12.2123.19
Auto-GAS (Pearson)60.07%77.53%12.8524.62
Auto-GAS (Hamming)60.58%78.57%12.3523.88
IVAuto-GAS (Proxy Random)29.49%38.25%16.6528.12
Auto-GAS (Proxy Evolution)61.14%79.32%12.2123.19
+ +SNIP require gradient backpropagation, leading to increased time consumption during the search process. In contrast, our Auto-GAS only requires a single forward pass, making it faster than methods that necessitate multiple forwards or gradient backpropagation. For image translation, as presented in Table 4, Auto-GAS achieves a $110 \times$ enhancement compared to GAN Compression [38]. + +Ablation of proxy search. (1) Our proxy search space includes transform, encoding, reduction, and augmentation operations. Ablations in Table 5 (II) indicate that all these operators contribute, with encoding operators having a greater influence. (3) Ablations for different encoding operators in Table 5 (III) show that the gram-matrix is important, but other operators (e.g., Pearson correlation matrix) also achieve similar performance. (2) Our contribution stems from search space design and proxy search processes. Ablations in Table 5 demonstrate that the various components of our search space contribute. In addition, our proxy evolution search plays an important role in exploring extensive search space and it significantly outperforms random search in Table 5 (IV). + +Operations for mutual information estimation. (1) Gram-matrix calculates matrix correlation for mutual information, which measures the salient features and relevant information captured by generative models. The comparison in Table 5 (I) shows that the vanilla Gram-matrix can be used as a proxy for NAS but is weaker than our Auto-GAS. (2) We also compare other operations for fea- + +Table 6: Results of different correlation methods on the CIFAR-10 dataset. analysis on the CIFAR-10 dataset. + +
Correlation MethodsCIFAR-10HyperparameterCIFAR-10
IS↑FID↓IS↑FID↓
Hessian Trace8.3615.47Batch Size (20)8.5212.28
L2norm7.5624.62Batch Size (40)8.0216.52
Entropy7.7025.43Number of Individuals (16)8.4412.21
Variance7.6723.75Number of Individuals (64)8.1116.32
Gram-matrix8.1217.88EA iteration interval (5)7.7818.25
Auto-GAS8.4412.21EA iteration interval (10)8.3813.69
+ +ture information estimation in Table. 6. The results of $L_{2}$ norm|Entropy|Variance are worse than Auto-GAS, because they overlooks higher-order feature statistics. Hessian Trace [43] yields relatively promising performance but its computational complexity is excessive. + +Hyperparameter analysis. The results in Table 7 indicate that a larger batch size may speed up convergence but also risk overfitting. An appropriate population size ensures our more efficient architecture exploration. Larger iteration intervals allow for a more comprehensive exploration in our architecture space. + +# 5 Conclusion + +In this paper, we present Auto-GAS, a novel training-free generative architecture search framework by evolving customized proxy. We derive from the information bottleneck theory using the latent layer feature as input and building candidate proxies with various transform, encoding, reduction, and augment options. We search for the best proxy using evolutionary algorithms. With customized proxies, our Auto-GAS allows an efficient search for promising architectures. Comprehensive experiment results on image generation and image translation illustrate the promising performance of Auto-GAS. For future work, we will augment our Auto-GAS with advanced methods [31-35,37,52,69]. We hope our novel investigations will provide some insight for the generation and NAS research community. + +**Limitation.** We mainly evaluate the same benchmarks as other NAS methods for fair comparison. We will extend Auto-GAS to more generative models in future work. **Social impact.** Our Auto-GAS focuses on technology improvement without social and ethical implications. + +# Acknowledgements + +The research was supported by Theme-based Research Scheme (T45-205/21-N) from Hong Kong RGC, and Generative AI Research and Development Centre from InnoHK. + +# References + +1. Abdelfattah, M.S., Mehrotra, A., Dudziak, L., Lane, N.D.: Zero-cost proxies for lightweight nas. arXiv preprint arXiv:2101.08134 (2021) +2. Akhauri, Y., Munoz, J.P., Jain, N., Iyer, R.: EZNAS: Evolving zero-cost proxies for neural architecture scoring. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) NeurIPS (2022), https://openreview.net/forum?id=1SqaDG4dvdt +3. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. ICLR (2017) +4. Baker, B., Gupta, O., Naik, N., Raskar, R.: Designing neural network architectures using reinforcement learning. In: ICLR (2017) +5. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019) +6. Cavagnero, N., Robbiano, L., Pistilli, F., Caputo, B., Averta, G.: Entropic score metric: Decoupling topology and size in training-free nas. In: ICCV workshop (2023) +7. Chen, K., Yang, L., Chen, Y., Chen, K., Xu, Y., Li, L.: Gp-nas-ensemble: a model for the nas performance prediction. In: CVPRW (2022) +8. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Fourteenth International Conference on Artificial Intelligence and Statistics (2011) +9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. CVPR (2016) +0. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023) +1. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV (2023) +2. Dong, P., Niu, X., Li, L., Tian, Z., Wang, X., Wei, Z., Pan, H., Li, D.: Rd-nas: Enhancing one-shot supernet ranking ability via ranking distillation from zero-cost proxies. arXiv preprint arXiv:2301.09850 (2023) +3. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022) +4. Dong, P., Niu, X., Tian, Z., Li, L., Wang, X., Wei, Z., Pan, H., Li, D.: Progressive meta-pooling learning for lightweight image classification model. In: ICASSP (2023) +5. Duan, Y., Chen, X., Xu, H., Chen, Z., Liang, X., Zhang, T., Li, Z.: Transnas-bench-101: Improving transferability and generalizability of cross-task neural architecture search. In: CVPR. pp. 5251-5260 (2021) +6. Elsken, T., Metzen, J.H., Hutter, F.: Neural architecture search: A survey. arXiv preprint arXiv:1808.05377 (2018) +7. Gao, C., Chen, Y., Liu, S., Tan, Z., Yan, S.: Adversarialnas: Adversarial neural architecture search for gans. In: CVPR (2020) +8. Gong, X., Chang, S., Jiang, Y., Wang, Z.: Autogan: Neural architecture search for generative adversarial networks. ICCV (2019) +9. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014) +20. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. NeurIPS (2017) + +21. He, H., Wang, H., Lee, G.H., Tian, Y.: Probgan: Towards probabilistic gan with theoretical guarantees. In: ICLR (2018) +22. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021) +23. Isola, P., Zhu, J., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017). https://doi.org/10.1109/CVPR.2017.632 +24. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017) +25. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (June 2019) +26. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.E.: Similarity of neural network representations revisited. In: ICML (2019) +27. Krizhevsky, A., Nair, V., Hinton, G.: The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html (2014) +28. Lee, N., Ajanthan, T., Torr, P.H.: Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340 (2018) +29. Li, G., Yang, Y., Bhardwaj, K., Marculescu, R.: Zico: Zero-shot NAS via inverse coefficient of variation on gradients. In: The Eleventh ICLR (2023), https://openreview.net/forum?id=rwo-1s5GqGn +30. Li, H., Fu, T., Dai, J., Li, H., Huang, G., Zhu, X.: Autoloss-zero: Searching loss functions from scratch for generic tasks. In: CVPR (2022) +31. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022) +32. Li, L., Bao, Y., Dong, P., Yang, C., Li, A., Luo, W., Liu, Q., Xue, W., Guo, Y.: Detkds: Knowledge distillation search for object detectors. In: ICML (2024) +33. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024) +34. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023) +35. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022) +36. Li, L., Sun, H., Dong, P., Wei, Z., Shao, S.: Auto-das: Automated proxy discovery for training-free distillation-aware architecture search. In: ECCV (2024) +37. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation. In: ICLR (2020) +38. Li, M., Lin, J., Ding, Y., Liu, Z., Zhu, J.Y., Han, S.: Gan compression: Efficient architectures for interactive conditional gans. CVPR (2020) +39. Lin, M., Wang, P., Sun, Z., Chen, H., Sun, X., Qian, Q., Li, H., Jin, R.: Zen-nas: A zero-shot nas for high-performance image recognition. ICCV (2021) +40. Lin, Q., Fang, Z., Chen, Y., Tan, K.C., Li, Y.: Evolutionary architectural search for generative adversarial networks. IEEE Transactions on Emerging Topics in Computational Intelligence (2022) +41. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019) +42. Liu, L., Zhang, Y., Deng, J., Soatto, S.: Dynamically grown generative adversarial networks. AAAI (2021) +43. Liu, Y., Yu, S., Lin, T.: Hessian regularization of deep neural networks: A novel approach based on stochastic estimators of hessian trace. Neurocomputing 536, 13-20 (2023) + +44. Luo, Y., Zhang, Y., Cai, X., Yuan, X.: E2gan: End-to-end generative adversarial network for multivariate time series imputation. In: 28th international joint conference on artificial intelligence. pp. 3094-3100. AAAI Press Palo Alto, CA, USA (2019) +45. Mellor, J., Turner, J., Storkey, A., Crowley, E.J.: Neural architecture search without training. In: ICML (2021) +46. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. ArXiv abs/1802.05957 (2018) +47. Odena, A., Olah, C., Shlens, J.: Conditional image synthesis with auxiliary classifier gans. In: ICML. PMLR (2017) +48. Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: CVPR (2019) +49. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI (2022) +50. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR abs/1511.06434 (2016) +51. Real, E., Liang, C., So, D.R., Le, Q.V.: Automl-zero: Evolving machine learning algorithms from scratch (2020) +52. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023) +53. Shu, H., Wang, Y., Jia, X., Han, K., Chen, H., Xu, C., Tian, Q., Xu, C.: Coevolutionary compression for unpaired image translation. In: ICCV (2019) +54. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information. arXiv:1703.00810 (2017) +55. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information. arXiv preprint, arXiv:1703.00810 (2017) +56. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019) +57. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS (2020) +58. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS 33, 6377-6389 (2020) +59. Tian, Y., Shen, L., Su, G., Li, Z., Liu, W.: Alphagan: Fully differentiable architecture search for generative adversarial networks. arXiv preprint arXiv:2006.09134 (2020) +60. Tian, Y., Wang, Q., Huang, Z., Li, W., Dai, D., Yang, M., Wang, J., Fink, O.: Off-policy reinforcement learning for efficient and effective gan architecture search. In: ECCV (2020) +61. Tishby, N., Zaslavsky, N.: Deep learning and the information bottleneck principle. In: ITW. pp. 1-5. IEEE (2015) +62. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020) +63. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020) +64. Wang, C., Xu, C., Yao, X., Tao, D.: Evolutionary generative adversarial networks. IEEE Transactions on Evolutionary Computation 23(6), 921-934 (2019) +65. Wang, H., Huan, J.: Agan: Towards automated design of generative adversarial networks. arXiv preprint arXiv:1906.11080 (2019) +66. Wang, W., Sun, Y., Halgamuge, S.: Improving MMD-GAN training with repulsive loss function. In: ICLR (2019) + +67. Wei, Z., Pan, H., Li, L.L., Lu, M., Niu, X., Dong, P., Li, D.: Convformer: Closing the gap between cnn and vision transformers. arXiv preprint arXiv:2209.07738 (2022) +68. Wolchover, N., Reading, L.: New theory cracks open the black box of deep learning. Quanta Magazine (2017) +69. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching. In: ICLR (2023) +70. Ying, G., He, X., Gao, B., Han, B., Chu, X.: Eagan: Efficient two-stage evolutionary architecture search for gans. In: ECCV. Springer (2022) +71. Yu, A., Grauman, K.: Fine-grained visual comparisons with local learning. CVPR (2014) +72. Zhang, M., Li, H., Pan, S., Chang, X., Su, S.: Overcoming multi-model forgetting in one-shot nas with diversity maximization. In: CVPR. pp. 7809-7818 (2020) +73. Zhong, Z., Yang, Z., Deng, B., Yan, J., Wu, W., Shao, J., Liu, C.L.: Blockqnn: Efficient block-wise neural network architecture generation. IEEE transactions on pattern analysis and machine intelligence (2020) +74. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024) +75. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024) +76. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018) \ No newline at end of file diff --git a/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/images.zip b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..154e848563c0e317e3a742b7a0512afd50c0c152 --- /dev/null +++ b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d94e9691ba574938e6fb5474cee3362822423dba0e96e41e7e3848241a547a40 +size 525172 diff --git a/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/layout.json b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4863f1743e030143bad8b65c55aa9b89dc93291b --- /dev/null +++ b/2024/Auto-GAS_ Automated Proxy Discovery for Training-free Generative Architecture Search/layout.json @@ -0,0 +1,9666 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "text", + "content": "Auto-GAS: Automated Proxy Discovery for Training-free Generative Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "spans": [ + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "text", + "content": "Lujun Li" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "text", + "content": ", Haosen Sun" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "text", + "content": ", Shiwen Li" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "text", + "content": ", Peijie Dong" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "text", + "content": ", Wenhan Luo" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "text", + "content": ", Wei Xue" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "text", + "content": ", Qifeng Liu" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "text", + "content": ", and Yike Guo" + }, + { + "bbox": [ + 137, + 170, + 477, + 196 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 204, + 422, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 204, + 422, + 216 + ], + "spans": [ + { + "bbox": [ + 190, + 204, + 422, + 216 + ], + "type": "text", + "content": "1 The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 265, + 216, + 347, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 216, + 347, + 227 + ], + "spans": [ + { + "bbox": [ + 265, + 216, + 347, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 265, + 216, + 347, + 227 + ], + "type": "text", + "content": " Xidian University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 227, + 469, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 227, + 469, + 270 + ], + "spans": [ + { + "bbox": [ + 142, + 227, + 469, + 270 + ], + "type": "text", + "content": "3 The Hong Kong University of Science and Technology (Guangzhou) lilujunai@gmail.com;hsunas@connect.ust.hk;shiwenlisw@gmail.com; pdong212@connect.hkust-gz.edu.cn;whluo.china@gmail.com;weixue@ust.hk; liuqifeng@ust.hk;yikeguo@ust.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 326, + 455, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 326, + 455, + 601 + ], + "spans": [ + { + "bbox": [ + 159, + 326, + 455, + 601 + ], + "type": "text", + "content": "Abstract In this paper, we introduce Auto-GAS, the first training-free Generative Architecture Search (GAS) framework enabled by an autodiscovered proxy. Generative models like Generative Adversarial Networks (GANs) are now widely used in many real-time applications. Previous GAS methods use differentiable or evolutionary search to find optimal GAN generators for fast inference and memory efficiency. However, the high computational overhead of these training-based GAS techniques limits their adoption. To improve search efficiency, we explore training-free GAS but find existing zero-cost proxies designed for classification tasks underperform on generation benchmarks. To address this challenge, we develop a custom proxy search framework tailored for GAS tasks to enhance predictive power. Specifically, we construct an information-aware proxy that takes feature statistics as inputs and utilizes advanced transform, encoding, reduction, and augment operations to represent candidate proxies. Then, we employ an evolutionary algorithm to perform crossover and mutation on superior candidates within the population based on correlation evaluation. Finally, we perform generator search without training using the optimized proxy. Thus, Auto-GAS enables automated proxy discovery for GAS while significantly accelerating the search before training stage. Extensive experiments on image generation and image-to-image translation tasks demonstrate that Auto-GAS strikes superior accuracy-speed tradeoffs over state-of-the-art methods. Remarkably, Auto-GAS achieves competitive scores with " + }, + { + "bbox": [ + 159, + 326, + 455, + 601 + ], + "type": "inline_equation", + "content": "110 \\times" + }, + { + "bbox": [ + 159, + 326, + 455, + 601 + ], + "type": "text", + "content": " faster search than GAN Compression. Code at: https://github.com/lliai/Auto-GAS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 613, + 455, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 613, + 455, + 635 + ], + "spans": [ + { + "bbox": [ + 159, + 613, + 455, + 635 + ], + "type": "text", + "content": "Keywords: Generative Model " + }, + { + "bbox": [ + 159, + 613, + 455, + 635 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 613, + 455, + 635 + ], + "type": "text", + "content": " Generative Architecture Search " + }, + { + "bbox": [ + 159, + 613, + 455, + 635 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 613, + 455, + 635 + ], + "type": "text", + "content": " AutoML" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 251, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 251, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 251, + 665 + ], + "type": "text", + "content": "Corresponding authors." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 135, + 477, + 194 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "type": "text", + "content": "Table 1: Comparison of search cost for generative architecture search methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 135, + 477, + 194 + ], + "lines": [ + { + "bbox": [ + 136, + 135, + 477, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 135, + 477, + 194 + ], + "type": "table", + "html": "
MethodAlgorithm Cost (GPU-days)MethodAlgorithm Cost (GPU-days)
AGAN [65]RL1200DGGAN [42]Heuristic580
AutoGAN [18]RL2EAS-GAN [40]EA1
AlphaGAN [59]Gradient1.2EGAN [64]EA1.3
EAGAN [70]EA0.8Auto-GAS (ours)Training-free0.09
", + "image_path": "13997cbe8f92b03bf62c64e1aa614da74dc4de13312dd10e9698ecc61bf1718d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 230, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 230, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 230, + 229 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 244, + 481, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 481, + 388 + ], + "type": "text", + "content": "Generative models like Generative Adversarial Networks (GANs) [19] have made tremendous strides in visual generation tasks such as image and video synthesis [5,50]. However, their demanding computational footprint presents barriers to deploying these models on edge devices with strict resource constraints [25,47]. To overcome this, there is growing interest in model compression techniques [7, 12, 14, 49, 67]. Among them, Neural Architecture Search (NAS) [16] automates architecture design, finding compact generative models for edge computing, and enabling flexible structures without pre-trained weights. In contrast, other compression methods are limited to fixed, handcrafted architectures and rely on pre-trained models. The integration of NAS in generative models enables real-time visual creation on edge devices, unlocking new interactive experiences in domains like augmented reality [5,24,50]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 605 + ], + "type": "text", + "content": "Conventional Generative Architecture Search (GAS) approaches typically rely on expensive training-based training routes to assess candidate architectures, such as maintaining weight-sharing supernets [13, 22] or conducting multiple training trials from scratch [56, 76]. However, training generative models presents unique obstacles that greatly increase the costs and difficulties of applying these techniques for GAS. The training processes for generative models are generally more complex than discriminative tasks because inherent instabilities (e.g., mode collapse [20]) can occur during optimization. In mode collapse, the model generates only a limited variety of examples from the underlying distribution. These instabilities make it challenging to reliably evaluate architectural candidates with standard NAS training-based evaluations. As shown in Table 1, gradient-based GAS methods like AdversarialNAS [17] and AlphaGAN [59] amplify these problems by jointly searching the generator and discriminator architecture spaces. This coupled search increases the overall complexity and fragility of the optimization procedure. To address these difficulties, DGGAN [42] introduced a progressive search approach where candidates are trained incrementally. However, its process still requires a massive 580 GPU days to complete due to the computationally intensive nature of generative model training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Recently, training-free NAS approaches have demonstrated remarkable efficiency using zero-cost proxies to predict network performance without training network parameters. These methods are mainly present for encoder-only models for classification tasks. Inspired by their computational efficiency, we comprehensively evaluate existing proxies when applied to generative tasks. As shown" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 222, + 188 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 222, + 188 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 222, + 188 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 222, + 188 + ], + "type": "image", + "image_path": "9e5ba8747072f801b0fc63fbbb0149f87a792a36743ff05d30f7ef192733802b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 197, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 130, + 197, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 479, + 220 + ], + "type": "text", + "content": "Figure 1: Correlation visualization of proxies (from left to right: Param., Synflow, NWOT, and Auto-GAS) in Generation benchmarks [15]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 115, + 309, + 188 + ], + "blocks": [ + { + "bbox": [ + 222, + 115, + 309, + 188 + ], + "lines": [ + { + "bbox": [ + 222, + 115, + 309, + 188 + ], + "spans": [ + { + "bbox": [ + 222, + 115, + 309, + 188 + ], + "type": "image", + "image_path": "15e402b01e4469643860bc640273e7639ca896db14da9a9f94e5b5f04ad58517.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 116, + 394, + 188 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 394, + 188 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 394, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 394, + 188 + ], + "type": "image", + "image_path": "80a0b71a52514bc63f8fafad1f12f208c2b8be34a69383e654b60e369ba31a3d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 394, + 116, + 480, + 188 + ], + "blocks": [ + { + "bbox": [ + 394, + 116, + 480, + 188 + ], + "lines": [ + { + "bbox": [ + 394, + 116, + 480, + 188 + ], + "spans": [ + { + "bbox": [ + 394, + 116, + 480, + 188 + ], + "type": "image", + "image_path": "2bd49668b642029ed754f2efa63851b39b2931d4e1271f953417e0de625ba727.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 245, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 388 + ], + "type": "text", + "content": "in Figure 1, we can make some interesting observations: (1) We observe that parameter-level proxies (e.g., Synflow [57], SNIP [28], and GraSP [62]), which estimate performance based on weight importance, generally perform poorer when ranking generator candidates. This indicates their assumptions regarding how architecture impacts learning do not directly translate to generative objectives and training dynamics. (2) Architecture-level proxies (e.g., NWOT [45] and Zen [39]) featuring topological encodings demonstrate stronger ranking ability. However, even the best architecture proxies leave room for improvement to enable performing generative NAS at scale. These findings motivate developing customized proxies that more precisely capture the unique characteristics of visual generation. Thus, a question naturally arises: how to design customized proxies for generation tasks?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 389, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 604 + ], + "type": "text", + "content": "To answer this question, we present Auto-GAS, the first automated generative architecture search framework that aims to efficiently utilize custom training-free proxies. Firstly, we derive that the mutual information of latent features within generative models is strongly coupled with task optimization objectives based on information bottleneck theory [3,54,61]. Then, we develop proxy search strategies to better estimate the mutual information of these features. Specifically, we build comprehensive operation search spaces and evolutionary algorithms to search optimal information-aware proxies. To better process feature statistics, our search space involves transform-based, encoding-based, reduction-based, and augment-based operators. The search process in Auto-GAS begins with initializing the population of proxies and then evaluates, crosses, and varies them to improve the quality of the proxies. Throughout the search, we directly utilize the ranking correlation with accuracy results in the benchmark as fitting objectives, aiming to find a more suitable proxy for the generation tasks. With this evolutionary process, Auto-GAS outperforms existing training-free NAS approaches by significant margins without manual tuning. Additionally, we search for generator architectures using Auto-GAS and subsequently implement a full training process on the searched generator." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "We perform extensive experiments to validate the performance and efficiency strengths of our Auto-GAS in generation tasks like image generation and imaged-to-image translation. The experiments demonstrate that our Auto-GAS can achieve better-generating performance than other zero-shot proxies and state-of-the-art results on rank consistency. In contrast to handcrafted GAS methods," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "text", + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "our Auto-GAS framework enjoys many benefits: (1) Automatic. It enables the automatic discovery of more expressive and efficient proxies that human expert might ignore, which can reduce human bias and ensures that the resulting architectures are optimized for the target problem or dataset. (2) General. The search procedure of Auto-GAS improves generalization by discovering proxies that generalize well on different search spaces. Based on the final results for the search objectives, our framework can evolve adaptive proxies to generalize across architectures. (3) Insightful. Our framework allows for the automatic discovery of more expressive, efficient, and effective proxies for optimizing and customizing generation architectures, opening up new research directions for automated GAN architecture search. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 254, + 479, + 372 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 254, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 479, + 289 + ], + "type": "text", + "content": "- We present Auto-GAS, the first training-free architecture search framework for generation tasks. Auto-GAS achieves an ultra-efficient and elegant search process with our auto-searched customized proxy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 289, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 289, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 139, + 289, + 479, + 335 + ], + "type": "text", + "content": "- We propose a well-organized proxy search space, including latent features as input and various transforms, encoding, reduction, and augment operations as options. We develop evolutionary algorithms for effective & efficient proxy search." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 337, + 479, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 337, + 479, + 372 + ], + "spans": [ + { + "bbox": [ + 139, + 337, + 479, + 372 + ], + "type": "text", + "content": "- We conduct extensive experiments on image generation and image-to-image translation tasks. Auto-GAS obtains significant search acceleration and competitive performance on multiple datasets." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 390, + 237, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 237, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 237, + 403 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 414, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 628 + ], + "type": "text", + "content": "Generative architecture search. The field of Neural Architecture Search (NAS) traditionally encompasses the design of exploration spaces, search methodologies, and assessment techniques to autonomously uncover optimal architectural configurations within specified parameters. Empirical approaches utilize an iterative process of training followed by exploration, either through multiple iterations [4,73] or by employing weight-sharing strategies [41,72]. NAS has also been extensively applied to the domain of GANs Architecture Search (GAS). The primary goal in GAS is to automatically identify superior GAN structures that enhance generative performance. Pioneering reinforcement learning-based GAS techniques, such as AGAN [65] and AutoGAN [18], leverage RNN controllers to construct GAN architectures and employ the Inception Score (IS) as their evaluation criterion. E2GAN [60] and AdversarialNAS [17] implement alternative reward functions. EAGAN [70] conducts discrete searches and enhances the robustness of the GAN exploration process. Nevertheless, these training-dependent search methodologies continue to be hampered by inefficient exploration and substantial computational demands. To tackle this challenge, we propose a training-free approach, grounded in architecture decoupling, to enhance the discovery of competitive GAN structures in a more computationally efficient manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Training-free architecture search. Zero-shot architecture exploration [10, 11, 29, 39], alternatively referred to as training-free NAS, offers a swifter alternative to conventional NAS methods. This approach's evaluation process en" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "tails only a handful of forward or backward passes with a small batch of input data, rendering it highly efficient. Zero-cost indicators can be categorized into parameter-centric and architecture-centric proxies. Parameter-centric zero-cost proxies (e.g., Synflow [57], SNIP [28], and GraSP [62]) rely on pruning techniques and aggregate the saliency values of each layer's weights as a proxy. For instance, Synflow [58] introduces a modified version of synaptic saliency scores to prevent layer collapse during parameter pruning, while Fisher [1] calculates the sum of all activation gradients in the network, applicable to channel pruning. Grasp [63] ensures gradient flow preservation throughout the network for efficient training. Their detailed formulations are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 243, + 483, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 243, + 483, + 277 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 483, + 277 + ], + "type": "interline_equation", + "content": "\\rho_ {s n i p} = \\left| \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W} \\right|, \\rho_ {s y n f l o w} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}} \\odot \\mathcal {W}, \\rho_ {f i s h e r} = \\left(\\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {A}} \\mathcal {A}\\right) ^ {2}, \\rho_ {g r a s p} = - (H \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W}}) \\odot \\mathcal {W}, \\tag {1}", + "image_path": "19fb7c66669cd76260e5ebbf4970a2f56d05d9f7d04df78ebe29ec0c4ed65084.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "text", + "content": " are loss function, weight and activation. " + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "text", + "content": " is the Hessian matrix. However, the complexity of loss function " + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 277, + 482, + 504 + ], + "type": "text", + "content": " in generation tasks renders these gradient-based methods ineffective. Regarding architecture-centric proxies, ZenNAS [39] introduced the novel Zen-Score to assess network expressivity, while NWOT [45] examined correlations between linear maps induced by data points for untrained architectures. Nevertheless, these approaches employ hand-crafted fixed proxies for classification tasks and encoder-only CNN models. In generation tasks, the encoder-decoder generator typically features intricate branches and multi-scale features, posing significant challenges for traditional proxies in predicting final accuracy. To address this issue, we automatically optimize the first customized proxies for GAS. Furthermore, our framework differs from EZNAS-like methods [2, 36, 74, 75] in several aspects: (1) Tasks. EZNAS is exclusively designed for classification, whereas our Auto-GAS is tailored for generation tasks. (2) Search algorithms. EZNAS merely employs genetic programming, while we propose a proxy evolution search strategy that substantially enhances search efficiency. (3) Model statistic inputs. Our method solely utilizes features as input, as we observed poor performance of parameter-level proxies. Conversely, EZNAS selects parameters and gradients as inputs. (4) Search space. Our Auto-GAS incorporates numerous feature-related operations absent in EZNAS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 521, + 233, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 233, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 233, + 536 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 583 + ], + "type": "text", + "content": "In this section, we first review existing proxies. Then, we present formulations and insights of our proxy design (see Figure 2). Finally, we illustrate the detailed search process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 598, + 261, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 261, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 261, + 611 + ], + "type": "text", + "content": "3.1 Proxy Search Space" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Motivation of our customized proxy design. Generative tasks are relevant to the Information Bottleneck (IB) theory [3, 54, 61], which guides generative models to learn compact representations that retain important characteristics for accurate sample generation. In generative tasks, the IB theory helps understand" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 448, + 102 + ], + "type": "text", + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 234 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 234 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 234 + ], + "type": "image", + "image_path": "3412afae11f3dd45eb17afde1ccc179e4ae3df5c2d14571ee9aa87a9ac4c6ef6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 300 + ], + "type": "text", + "content": "Figure 2: Overall proxy search of Auto-GAS. We extract the latent features of the generator with input data forward-propagation. Then we select various transform, encoding, reduction, and augment operators from the proxy search space to build candidate proxies. Finally, we perform a ranking evaluation to cross and mutate to get the best proxy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 326, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 374 + ], + "type": "text", + "content": "how generative models can balance compression (capturing essential features) and reconstruction (faithful generation of samples) [55, 68]. IB theory suggests maximizing the Mutual Information (MI) " + }, + { + "bbox": [ + 130, + 326, + 482, + 374 + ], + "type": "inline_equation", + "content": "I(\\mathbf{X}; \\mathbf{F}_i)" + }, + { + "bbox": [ + 130, + 326, + 482, + 374 + ], + "type": "text", + "content": " between the input data " + }, + { + "bbox": [ + 130, + 326, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 326, + 482, + 374 + ], + "type": "text", + "content": " and the corresponding latent representation " + }, + { + "bbox": [ + 130, + 326, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 130, + 326, + 482, + 374 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 241, + 387, + 481, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 387, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 241, + 387, + 481, + 401 + ], + "type": "interline_equation", + "content": "I (\\mathbf {X}; \\mathbf {F} _ {i}) = H (\\mathbf {F} _ {i}) - H (\\mathbf {F} _ {i} | \\mathbf {X}) \\tag {2}", + "image_path": "32ab701c315027a38a8677778283ecaaf4525c718604aae4c86c03831363b19d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "inline_equation", + "content": "H(\\mathbf{F}_i)" + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "text", + "content": " stands for the entropy of the latent representation " + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "inline_equation", + "content": "H(\\mathbf{F}_i|\\mathbf{X})" + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "text", + "content": " denotes the conditional entropy. Considering that neural networks are deterministic, and the dataset has a finite number of instances, the conditional entropy becomes zero, simplifying the mutual information (MI) estimation as " + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "inline_equation", + "content": "I(\\mathbf{X};\\mathbf{F}_i) = H(\\mathbf{F}_i)" + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "text", + "content": ". Consequently, some information coding and compression studies [26, 55] propose some operations (e.g., Entropy and Gram-matrix) to approximate the mutual information on latent representation. These inspired us to develop our train-free method design thought chain: assess generative model " + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "text", + "content": " estimate MI " + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "text", + "content": " design expressions via Entropy/Gram-matrix ops " + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 412, + 482, + 592 + ], + "type": "text", + "content": " our train-free proxy. However, how about designing this potential expressions? Recent Auto-Zero approaches [30, 51] present a promising solution: searching for the best machine learning solution from scratch. In addition, several studies [6] have shown that normalize and activation operations can enhance the formulation. Therefore, we automatically search for optimal proxies based on these operators." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "content": "Proxy primitive operations. Our approach evolves the customized proxy in the generation benchmark [15], which consists of U-Net-like encoder-decoder models with various operators and their training results. We extract the latent features of each layer in both the encoder and decoder modules as inputs to the search process and represent the zero-cost proxy as a computation graph. This graph is constructed as transform " + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "content": " encoding " + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "content": " reduction " + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "content": " augment" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "parts in processing generator statistics. The available primitive operations can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 150, + 480, + 340 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 141, + 150, + 479, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 150, + 479, + 197 + ], + "spans": [ + { + "bbox": [ + 141, + 150, + 479, + 197 + ], + "type": "text", + "content": "- Transform operations: These operations involve different activation functions and feature processing techniques to enhance raw features and represent architectural potentials. Examples of transform operations include relu, swish, mish, scale,norm, exp, and softmax." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 198, + 480, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 198, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 198, + 480, + 257 + ], + "type": "text", + "content": "- Encoding operations: Inspired by coding algorithms and feature information estimation, we employ binary and matrix-based operations in the encoding part. These operators help us better estimate the mutual correlation information of samples. Some encoding operations include sign, above-zero, above-median, gram-matrix, person, and hamming." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 258, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 258, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 141, + 258, + 479, + 304 + ], + "type": "text", + "content": "- Reduction operations: Reduction operators bridge the information matrix and the proxy values. They are valuable for extracting useful matrix information from different mathematical perspectives. Reduction operations include sloqdet, abslog, trace, and eigenvalue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 306, + 479, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 306, + 479, + 340 + ], + "spans": [ + { + "bbox": [ + 141, + 306, + 479, + 340 + ], + "type": "text", + "content": "- Augment operations: Augment operators are applied to enhance the proxy values' ability to predict performance results. These operations include arithmetic functions such as abs, pow, l2-norm, log, sum, and mean." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 351, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 480, + 411 + ], + "type": "text", + "content": "By utilizing these primitive operations, our approach enables the evolution of the customized proxy, facilitating the search for optimal solutions in the generation benchmark. These mathematical operations are essential building blocks to construct a diverse and expressive search space of zero-cost proxy. We provide more details of their formulas and properties in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 429, + 279, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 429, + 279, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 429, + 279, + 441 + ], + "type": "text", + "content": "3.2 Proxy Evolution Search" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": "Evolutionary process and goal. Utilizing our proxy search space and graph representation, we employ an evolutionary algorithm (EA) outlined in Alg. 1 to discover optimal proxy expressions. Our EA commences with an initial set of candidate proxies and progressively refines the population across generations, employing genetic operators such as selection, crossover, and mutation to yield superior solutions. For fitness evaluation, each proxy is assessed based on the ranking correlation between its output proxy " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": " scores and pre-determined performance benchmarks " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": ", enabling efficient identification of the optimal proxy " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\rho^{*}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": " within the search space " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 237, + 567, + 480, + 584 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 567, + 480, + 584 + ], + "spans": [ + { + "bbox": [ + 237, + 567, + 480, + 584 + ], + "type": "interline_equation", + "content": "\\rho_ {A u t o - G A S} ^ {*} = \\underset {\\rho \\in \\mathcal {P}} {\\arg \\max } (\\tau (\\mathcal {D}, \\mathcal {Q})). \\tag {3}", + "image_path": "7c02e4be686499b0e6f08f4688ecf7ba50bb1883994eaecc3d87a17ec89620fb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "text", + "content": "where Kendall's Tau " + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "text", + "content": " is formulated as the correlation coefficient. Each evolution picks top-k candidates with the highest evaluation score and randomly selects a parent from the candidates for mutation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "content": "Efficient proxy search with existing generation benchmark. It is expensive to derive pre-trained results " + }, + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "content": ", we adopt an alternative solution to achieve proxy search using pre-computed architecture-performance pairs provided in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 448, + 102 + ], + "type": "text", + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 129, + 481, + 352 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 381, + 128 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 381, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 381, + 128 + ], + "type": "text", + "content": "Algorithm 1 Evolutionary Search for Auto-GAS Proxy" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "Input: Exploration domain " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", candidate set " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", maximum iterations " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", sampling ratio " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", sampled pool " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": ", top " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " selections. \nOutput: Auto-GAS proxy with highest Kendall's Tau " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " coefficient. \n1: " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{P}0 := \\text{Initialize candidate set}(P_i)" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n2: Sampled pool " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{R} := \\emptyset" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n3: for " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "i = 1, 2, \\ldots, T" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " do \n4: Select " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "G_i^s := \\text{SelectTopk}(\\mathcal{R}, k)" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n5: // greedy evolutionary strategy \n6: Empty sampled pool " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{R} := \\emptyset" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n7: Mutate := MUTATE(" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "G_i^s" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "); \n8: Crossover " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "G_i^c := \\text{CROSSOVER}(\\mathcal{G}_i^s)" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n9: Incorporate " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "G_i^m" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n10: Incorporate " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "G_i^c" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": "; \n11: if " + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{R} < \\mathcal{P}" + }, + { + "bbox": [ + 132, + 129, + 481, + 352 + ], + "type": "text", + "content": " then \n12: Introduce random samples \n13: else \n14: Proceed to line 4; \n15: end if \n16: end for" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "content": "the existing TransNAS-Bench-101 [15] benchmarks. Specifically, this benchmark involves Autoencoding task with 7k different generators in macro-level search space and their evaluation metric (i.e., SSIM). Each generator network follows an encoder-decoder structure in Pix2Pix, where the encoders are the searched backbones, and the decoders contain 14 layers of convolution and deconvolution. Thanks to this existing benchmark, we do not need to train performance groundtruths again, and it only takes 1.5 hours in our proxy search process. In light of excellent generality (proven in our extensive experiments), our searched proxy " + }, + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\rho_{Auto-GAS}" + }, + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "content": " can be well transferred to similar search spaces and can be directly used in most scenarios without additional search budget." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 511, + 296, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 296, + 523 + ], + "type": "text", + "content": "3.3 Discovered Proxy Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 530, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 481, + 555 + ], + "type": "text", + "content": "Searched Zero-cost Proxy. We present formulas of the searched proxies on the generation benchmark as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 178, + 561, + 481, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 561, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 178, + 561, + 481, + 576 + ], + "type": "interline_equation", + "content": "\\rho_ {A u t o - G A S} ^ {*} = \\left(\\operatorname {s l o g d e t} \\left([ \\mathbf {F} ^ {\\prime}, \\mathbf {F} ^ {\\prime T} ]\\right)\\right) ^ {2}, \\mathbf {F} ^ {\\prime} = \\operatorname {a b o v e - m e a n} (\\operatorname {s w i s h} (\\mathbf {F})) \\qquad (4)", + "image_path": "9d0c9cc7b65b6770a1bed62c0f987bc7817542ee91e6135f8379ae3fdd2a759d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": " represents the original features, and transformed features " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{F}'" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": " are obtained by applying the swish operation followed by the above-mean operation to the original features " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "[\\mathbf{F}',\\mathbf{F}'^T]" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "N\\times N" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": " gram-matrix of the transformed features " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{F}'" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": ". The proxy value " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\rho^{*}" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": " is calculated by taking the square of the slogdet of the concatenated gram-matrix " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "[\\mathbf{F}',\\mathbf{F}'^T]" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Analysis of searched proxy. The discovered formulations indicate that (1) Mathematical operations, such as swish activation and slogdet, significantly con-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 136, + 473, + 216 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 420, + 126 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 420, + 126 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 420, + 126 + ], + "type": "text", + "content": "Table 2: Ranking results on TransNAS-Bench-101 [15]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 136, + 473, + 216 + ], + "lines": [ + { + "bbox": [ + 141, + 136, + 473, + 216 + ], + "spans": [ + { + "bbox": [ + 141, + 136, + 473, + 216 + ], + "type": "table", + "html": "
MethodKendallSpearmanPearsonMethodKendallSpearmanPearson
SNIP [1]-2.04%-3.55%-0.26%ZiCo [29]18.63%15.53%21.13%
EZNAS [2]23.30%25.91%32.03%Synflow [57]-50.91%-65.01%-49.17%
NWOT [45]51.28%63.94%48.88%Params.-12.97%-15.24%-5.80%
SNIP [28]30.81%23.31%29.34%FLOPs48.91%64.01%52.36%
Zen [39]-2.27%2.58%5.39%Auto-GAS61.14%79.32%81.25%
", + "image_path": "fa3176a361b7f39924f01568c4edbc04ecdfe2460484be07faeb0920e3902b75.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 357 + ], + "type": "text", + "content": "tribute to the predictability of the proxy. These operations enhance the proxy's ability to capture important features and characteristics of the generator. (2) Gram-Matrix and Mutual Information: The expression utilizes a sample-wise gram-matrix, which accurately estimates the latent mutual information of the generator. The gram-matrix captures the relationships and dependencies among the transformed features, providing valuable insights into the generator's behaviour. The discovered proxy of Auto-GAS, as illustrated in Figure 1, showcases a substantial improvement in ranking. This improvement suggests that the searched proxy effectively captures essential characteristics of the generator, leading to enhanced performance in the benchmark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 358, + 481, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 481, + 467 + ], + "type": "text", + "content": "Understanding why our Auto-GAS surpasses other NAS methods. (1) Many train-based NAS methods employ weight-sharing and other proxy settings, which often lead to inaccurate estimations of architecture performance. In contrast, our Auto-GAS method searches for proxies based on actual performance, resulting in a proxy that demonstrates strong ranking capability. (2) Other train-free NAS methods are primarily designed for classification rather than generative tasks. Consequently, these methods search for architectures using weak correlation proxies. In contrast, our proxy exhibits a stronger correlation, ensuring the discovery of more efficient models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 486, + 361, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 361, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 361, + 498 + ], + "type": "text", + "content": "3.4 Training-free Generative Models Search" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "text", + "content": "Auto-GAS has made efforts to search generative Models in a non-training way to ease the instability of GAN training and improve search efficiency. In this stage, we directly search the generator in " + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "inline_equation", + "content": "\\rho_{Auto - GAS}" + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "text", + "content": ". In each iteration, we generate " + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "text", + "content": " candidate generators to make up the population " + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "text", + "content": ", where all these candidates use the randomly initialized weights " + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "text", + "content": ". Each candidate is characterized by both architecture and weight variables. In each iteration, we evaluate " + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 509, + 482, + 614 + ], + "type": "text", + "content": " candidates without training. We employ the information-aware proxy to assign scores to all these candidates. The formulation of the search stage can be succinctly stated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 618, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 618, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 219, + 618, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\alpha_ {A u t o - G A S} ^ {*} = \\arg \\max _ {\\alpha \\in \\mathcal {A}} \\rho_ {A u t o - G A S} ^ {*} (\\alpha , \\mathcal {W}). \\tag {5}", + "image_path": "b8857c3222fa1ad337b1a82ca78adfcdc6dacf35e085dc0e9956f149ca027b0c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "Note that in scenarios involving discriminator searches, we also use " + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\rho_{Auto - GAS}" + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": " for searching the discriminator architectures with real sample inputs." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "text", + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 146, + 443, + 313 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Comparison results on the CIFAR-10 and STL-10 datasets. We use FID (the lower the better) and the mIoU metric (the higher the better) for evaluations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 146, + 443, + 313 + ], + "lines": [ + { + "bbox": [ + 168, + 146, + 443, + 313 + ], + "spans": [ + { + "bbox": [ + 168, + 146, + 443, + 313 + ], + "type": "table", + "html": "
MethodAlgorithmGPU daysCIFAR-10 (32×32)STL-10 (48×48)
IS↑FID↓IS↑FID↓
WGAN-GP [20]7.8629.3--
SN-GAN [46]8.2221.709.1640.1
ProbGAN [21]Manual-7.7524.608.8746.74
Improv MMD GAN [66]8.2916.219.3437.63
BigGAN [5]9.2214.73--
AGAN [65]12008.2930.509.2352.70
AutoGAN [18]RL2.008.5512.429.1631.01
AlphaGAN [59]1.28.7015.56--
AdversarialNAS [17]1.007.8624.048.5238.85
EGAN [64]1.256.90---
EAS-GAN [40]EA1.007.4533.20-38.84
EAGAN [70]0.88.4112.839.6923.82
Fisher [1]0.128.0317.669.3729.82
Zen [39]0.118.0215.999.4528.66
SNIP [28]Training-free0.127.9616.679.5829.25
NWOT [45]0.098.0516.329.2327.46
Auto-GAS0.098.4412.219.7223.19
", + "image_path": "3ed1a8b0ab2127f4c96e66386dea0cb44b5fd543a335036b8e195715172affa9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 335, + 302, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 302, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 302, + 348 + ], + "type": "text", + "content": "3.5 Generative Models Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 355, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 429 + ], + "type": "text", + "content": "After the architecture searches, we train generative models with randomly selected data samples. Simultaneously, a batch of noise samples is generated from a noise distribution. The generator tries to generate data from random noise " + }, + { + "bbox": [ + 130, + 355, + 482, + 429 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 355, + 482, + 429 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 355, + 482, + 429 + ], + "type": "inline_equation", + "content": "G(z)" + }, + { + "bbox": [ + 130, + 355, + 482, + 429 + ], + "type": "text", + "content": ", which should be as close as possible to the real data distribution. The discriminator tries to distinguish the generated data from the real data. The value function " + }, + { + "bbox": [ + 130, + 355, + 482, + 429 + ], + "type": "inline_equation", + "content": "V(D,G)" + }, + { + "bbox": [ + 130, + 355, + 482, + 429 + ], + "type": "text", + "content": " for a minimax game is defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 445, + 480, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 445, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 214, + 445, + 480, + 480 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {G} \\max _ {D} V (D, G) = \\mathbb {E} _ {\\mathbf {x} \\sim p _ {d a t a} (\\mathbf {x})} [ \\log D (\\mathbf {x}) ] \\\\ + \\mathbb {E} _ {\\mathbf {z} \\sim p _ {z} (\\mathbf {z})} [ \\log (1 - D (G (\\mathbf {z})) ] \\tag {6} \\\\ \\end{array}", + "image_path": "cdc5539a5a4d3a34f4e0d034f1406771b7e076aecf52ace916158b370639df14.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 488, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 488, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim p_{data}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 488, + 482, + 513 + ], + "type": "text", + "content": " denotes the distribution of real data, " + }, + { + "bbox": [ + 130, + 488, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim p_z(\\mathbf{z})" + }, + { + "bbox": [ + 130, + 488, + 482, + 513 + ], + "type": "text", + "content": " denotes the input noise to the generator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 530, + 230, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 530, + 230, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 530, + 230, + 544 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 556, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 481, + 605 + ], + "type": "text", + "content": "In this section, we first perform proxy search experiments on the generation benchmark, and then we conduct architecture search experiments on image generation and image-to-image translation tasks by translocating the searched proxies. Finally, we provide a detailed analysis of our approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 621, + 361, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 361, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 361, + 634 + ], + "type": "text", + "content": "4.1 Experiments on Generation Benchmark" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Dataset and implementation. We adopt the search space and performance ground truth provided by the Autoencoding task in TransNAS-Bench-101 [15] as" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 305, + 143 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 305, + 143 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 305, + 143 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 305, + 143 + ], + "type": "image", + "image_path": "712476a5a6d6d7106f04f3db0b10a8be41fead19bfe73ca06da7ca54ced3c628.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 153, + 304, + 165 + ], + "lines": [ + { + "bbox": [ + 135, + 153, + 304, + 165 + ], + "spans": [ + { + "bbox": [ + 135, + 153, + 304, + 165 + ], + "type": "text", + "content": "Figure 3: Searched model on CIFAR-10." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 121, + 478, + 143 + ], + "blocks": [ + { + "bbox": [ + 309, + 121, + 478, + 143 + ], + "lines": [ + { + "bbox": [ + 309, + 121, + 478, + 143 + ], + "spans": [ + { + "bbox": [ + 309, + 121, + 478, + 143 + ], + "type": "image", + "image_path": "4176af95007f0863eee020d9a23fc5e5e3b9b360d20a7e1a9e97d5e3081ee716.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 153, + 471, + 165 + ], + "lines": [ + { + "bbox": [ + 314, + 153, + 471, + 165 + ], + "spans": [ + { + "bbox": [ + 314, + 153, + 471, + 165 + ], + "type": "text", + "content": "Figure 4: Searched model on STL-10." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 187, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 187, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 480, + 271 + ], + "type": "text", + "content": "the benchmark for searching and evaluating GAS proxies. We split the 4:3 ratio of the benchmark into the validation and test sets. During the proxy search phase, we set " + }, + { + "bbox": [ + 130, + 187, + 480, + 271 + ], + "type": "inline_equation", + "content": "(\\mathcal{P},\\mathcal{T},r,k)" + }, + { + "bbox": [ + 130, + 187, + 480, + 271 + ], + "type": "text", + "content": " in Alg. 1 as (20,100,0.9,5) and evaluate with 50 architectures in the validation set. After the search, we fairly compare the Auto-GAS proxy with the traditional ones by randomly sampling 200 architectures in the test set. Note that our validation and test sets do not overlap to ensure fair proxy comparisons. Detailed implementations are in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 271, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 481, + 344 + ], + "type": "text", + "content": "Comparison results. Table 5 demonstrates the ranking capabilities of different methods on Generation Benchmark. Auto-GAS (ours) achieves the best rankings across all three measures: Kendall's Tau " + }, + { + "bbox": [ + 130, + 271, + 481, + 344 + ], + "type": "inline_equation", + "content": "(61.14\\%)" + }, + { + "bbox": [ + 130, + 271, + 481, + 344 + ], + "type": "text", + "content": ", Spearman " + }, + { + "bbox": [ + 130, + 271, + 481, + 344 + ], + "type": "inline_equation", + "content": "(79.32\\%)" + }, + { + "bbox": [ + 130, + 271, + 481, + 344 + ], + "type": "text", + "content": ", and Pearson " + }, + { + "bbox": [ + 130, + 271, + 481, + 344 + ], + "type": "inline_equation", + "content": "(81.25\\%)" + }, + { + "bbox": [ + 130, + 271, + 481, + 344 + ], + "type": "text", + "content": ". These results indicate that Auto-GAS performed well in preserving the rankings compared to other methods. These findings highlight the effectiveness of Auto-GAS for generating meaningful and accurate rankings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 360, + 334, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 334, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 334, + 372 + ], + "type": "text", + "content": "4.2 Experiments on Image Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "text", + "content": "Datasets and implementation. We perform evaluations using CIFAR-10 [27] and STL-10 [8] datasets, aligning with other GAS methodologies [17, 18, 70]. CIFAR-10 comprises 50k training and 10k test " + }, + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "text", + "content": " images, while STL-10 contains 100,500 images at a higher " + }, + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "text", + "content": " resolution. We directly apply searched proxies to identify optimal architectures within EAGAN's [70] search space. The exploration process spans 150 epochs with a " + }, + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "inline_equation", + "content": "P = 32" + }, + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "text", + "content": " individual population. Generator and discriminator batch sizes are 80 and 40, respectively. We utilize 16 randomly generated Gaussian noise images for proxy score computation to expedite evaluation. Post-search, we conduct full training of the best-scoring generator for 500 epochs. For evaluation, following previous GAS works [17, 18, 70], we generate 50,000 images to calculate IS and FID metrics. Figures 3 and 4 illustrate the discovered architectures for CIFAR-10 and STL-10." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 484, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 484, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 484, + 640 + ], + "type": "text", + "content": "Training-based NAS method comparison. We benchmark Auto-GAS against various training-based NAS approaches: BigGAN [5], E2GAN [44], Adversarial-NAS [17], and EAGAN [70]. As shown in Table 3, Auto-GAS demonstrates superior efficiency. For CIFAR-10, Auto-GAS achieves lower FID than BigGAN and comparable overall results to E2GAN, while outperforming AdversarialNAS in both FID and IS. On STL-10, Auto-GAS significantly outperforms E2GAN and AdversarialNAS in both metrics, rivaling EAGAN in IS and surpassing it in FID. In summary, Auto-GAS emerges as an efficient and competitive approach, either surpassing or matching several state-of-the-art training-based NAS methods across diverse evaluation metrics and datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Zero-shot proxy comparison. Table 3 also juxtaposes the Auto-GAS proxy with four alternative zero-shot proxies. Fisher and SNIP, relying on gradient-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "text", + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 135, + 480, + 251 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 471, + 126 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 471, + 126 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 471, + 126 + ], + "type": "text", + "content": "Table 4: Image translation evaluation on horse→zebra and Cityscapes datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 135, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 135, + 135, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 135, + 135, + 480, + 251 + ], + "type": "table", + "html": "
ModelDatasetMethod#ParametersGPU daysFID (↓)mIoU (↑)
ValueRatioValueRatioValueIncreaseValueDrop
CycleGANhorse→zebra(512×512)Original11.4M---61.53--
COEV [53]----96.1534.6-
GAN Comp. [38]0.34M33.3×11.601.00×71.8110.3-
Fast GAN Comp. [38]0.36M32.1×3.123.72×65.193.66-
Auto-GAS0.39M29.3×0.11105.45×63.662.13-
GauGANCityscapes(512×512)Original93.0M---57.60-62.18-
GAN Comp. [38]20.4M4.6×17.701.00×55.192.4161.220.96
Fast GAN Comp. [38]20.2M4.6×10.401.70×56.251.3561.171.01
Auto-GAS21.0M4.5×0.16110.63×56.041.5661.210.97
", + "image_path": "72d08bd9a031a6004d57eccc91283f988d5ba46becb7da42c0514489d3f40eb5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 323 + ], + "type": "text", + "content": "based computations, yield similar scores. Zen-Score and NWOT exhibit positive correlations with model accuracy, outperforming gradient-based methods. These results validate the Auto-GAS proxy's efficacy, positioning it as a formidable contender among zero-shot proxy methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 341, + 385, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 385, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 385, + 354 + ], + "type": "text", + "content": "4.3 Experiments on Image-to-image Translation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": "Datasets and implementation. We comprehensively evaluate our approach on image translation tasks using high-resolution datasets, including [9] and horse " + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": " zebra [71] at " + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": " resolutions. We apply the proxy for architecture exploration, adopting GAN Compression's [38] search space for fair comparisons, focusing on channel and layer number configurations. We evolve to compress Pix2Pix [23] on horse " + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": " zebra and GauGAN [48] on cityscapes. The search process involves 500 iterations, followed by model retraining with a 0.0002 learning rate and 16 batch size, adhering to the original paper [23]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 458, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 528 + ], + "type": "text", + "content": "GAN compression method comparison. Table 4 demonstrates that Auto-GAS achieves the highest FID score of 63.66 for CycleGAN with a superior compression ratio " + }, + { + "bbox": [ + 130, + 458, + 482, + 528 + ], + "type": "inline_equation", + "content": "(29.3\\times)" + }, + { + "bbox": [ + 130, + 458, + 482, + 528 + ], + "type": "text", + "content": " compared to alternatives. This indicates Auto-GAS effectively reduces model size while preserving image quality. For GauGAN, Auto-GAS exhibits a marginally higher FID score and a competitive mIoU score of 61.22." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 529, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 590 + ], + "type": "text", + "content": "Qualitative Analysis. Figure 5 illustrates Auto-GAS's proficiency in translation tasks. In contrast, FastGAN Compression not only yields a lower FID score but also fails to eliminate artifacts like brown discoloration in zebra models. Conversely, Auto-GAS-generated zebra images accurately capture authentic black-and-white zebra patterns." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 609, + 239, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 239, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 239, + 621 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Search efficiency benefits. For image generation (see Table 3), our Auto-GAS method exhibits significantly superior search efficiency to training-based NAS GANs and has a definite advantage over other zero-cost proxies. The Fisher and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 480, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 480, + 205 + ], + "type": "image", + "image_path": "b5866431c6aa2827cf85ca6c855adf63100b90989d9a48ff6cd067154a3d236d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 212, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 132, + 212, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 479, + 235 + ], + "type": "text", + "content": "Figure 5: Visualization results on image translation tasks using high-resolution datasets, e.g., horse→zebra and Cityscapes with " + }, + { + "bbox": [ + 132, + 212, + 479, + 235 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 132, + 212, + 479, + 235 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 164, + 269, + 449, + 396 + ], + "blocks": [ + { + "bbox": [ + 223, + 249, + 388, + 259 + ], + "lines": [ + { + "bbox": [ + 223, + 249, + 388, + 259 + ], + "spans": [ + { + "bbox": [ + 223, + 249, + 388, + 259 + ], + "type": "text", + "content": "Table 5: Ablations of our proxy search." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 269, + 449, + 396 + ], + "lines": [ + { + "bbox": [ + 164, + 269, + 449, + 396 + ], + "spans": [ + { + "bbox": [ + 164, + 269, + 449, + 396 + ], + "type": "table", + "html": "
Exp. MethodKendallSpearmanFID↓@CIFAR-10FID↓@STL-10
IVanilla Gram-matrix24.28%35.75%17.8829.62
Auto-GAS61.14%79.32%12.2123.19
IIAuto-GAS without Transform55.78%73.60%14.2225.58
Auto-GAS without Encoding51.64%68.62%15.2626.49
Auto-GAS without Reduction52.78%71.95%14.6325.74
Auto-GAS without Augment58.40%76.06%13.8625.12
IIIAuto-GAS (Gram-matrix)61.14%79.32%12.2123.19
Auto-GAS (Pearson)60.07%77.53%12.8524.62
Auto-GAS (Hamming)60.58%78.57%12.3523.88
IVAuto-GAS (Proxy Random)29.49%38.25%16.6528.12
Auto-GAS (Proxy Evolution)61.14%79.32%12.2123.19
", + "image_path": "ba38058d59aca62d1cd734bde25b95717dd5bcc762e1fd5c8fe97dc277805674.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 424, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 480, + 483 + ], + "type": "text", + "content": "SNIP require gradient backpropagation, leading to increased time consumption during the search process. In contrast, our Auto-GAS only requires a single forward pass, making it faster than methods that necessitate multiple forwards or gradient backpropagation. For image translation, as presented in Table 4, Auto-GAS achieves a " + }, + { + "bbox": [ + 130, + 424, + 480, + 483 + ], + "type": "inline_equation", + "content": "110 \\times" + }, + { + "bbox": [ + 130, + 424, + 480, + 483 + ], + "type": "text", + "content": " enhancement compared to GAN Compression [38]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 604 + ], + "type": "text", + "content": "Ablation of proxy search. (1) Our proxy search space includes transform, encoding, reduction, and augmentation operations. Ablations in Table 5 (II) indicate that all these operators contribute, with encoding operators having a greater influence. (3) Ablations for different encoding operators in Table 5 (III) show that the gram-matrix is important, but other operators (e.g., Pearson correlation matrix) also achieve similar performance. (2) Our contribution stems from search space design and proxy search processes. Ablations in Table 5 demonstrate that the various components of our search space contribute. In addition, our proxy evolution search plays an important role in exploring extensive search space and it significantly outperforms random search in Table 5 (IV)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Operations for mutual information estimation. (1) Gram-matrix calculates matrix correlation for mutual information, which measures the salient features and relevant information captured by generative models. The comparison in Table 5 (I) shows that the vanilla Gram-matrix can be used as a proxy for NAS but is weaker than our Auto-GAS. (2) We also compare other operations for fea-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "text", + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 145, + 454, + 238 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 452, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 452, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 452, + 137 + ], + "type": "text", + "content": "Table 6: Results of different correlation methods on the CIFAR-10 dataset. analysis on the CIFAR-10 dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 145, + 454, + 238 + ], + "lines": [ + { + "bbox": [ + 133, + 145, + 454, + 238 + ], + "spans": [ + { + "bbox": [ + 133, + 145, + 454, + 238 + ], + "type": "table", + "html": "
Correlation MethodsCIFAR-10HyperparameterCIFAR-10
IS↑FID↓IS↑FID↓
Hessian Trace8.3615.47Batch Size (20)8.5212.28
L2norm7.5624.62Batch Size (40)8.0216.52
Entropy7.7025.43Number of Individuals (16)8.4412.21
Variance7.6723.75Number of Individuals (64)8.1116.32
Gram-matrix8.1217.88EA iteration interval (5)7.7818.25
Auto-GAS8.4412.21EA iteration interval (10)8.3813.69
", + "image_path": "f246ef192f40563d4877328f69b429fb3eeb9e4d98bf50673c0b3eb44283e4f9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 479, + 308 + ], + "type": "text", + "content": "ture information estimation in Table. 6. The results of " + }, + { + "bbox": [ + 130, + 261, + 479, + 308 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 261, + 479, + 308 + ], + "type": "text", + "content": " norm|Entropy|Variance are worse than Auto-GAS, because they overlooks higher-order feature statistics. Hessian Trace [43] yields relatively promising performance but its computational complexity is excessive." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 357 + ], + "type": "text", + "content": "Hyperparameter analysis. The results in Table 7 indicate that a larger batch size may speed up convergence but also risk overfitting. An appropriate population size ensures our more efficient architecture exploration. Larger iteration intervals allow for a more comprehensive exploration in our architecture space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 376, + 219, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 219, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 219, + 389 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 480, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 534 + ], + "type": "text", + "content": "In this paper, we present Auto-GAS, a novel training-free generative architecture search framework by evolving customized proxy. We derive from the information bottleneck theory using the latent layer feature as input and building candidate proxies with various transform, encoding, reduction, and augment options. We search for the best proxy using evolutionary algorithms. With customized proxies, our Auto-GAS allows an efficient search for promising architectures. Comprehensive experiment results on image generation and image translation illustrate the promising performance of Auto-GAS. For future work, we will augment our Auto-GAS with advanced methods [31-35,37,52,69]. We hope our novel investigations will provide some insight for the generation and NAS research community." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 535, + 480, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 535, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 480, + 583 + ], + "type": "text", + "content": "**Limitation.** We mainly evaluate the same benchmarks as other NAS methods for fair comparison. We will extend Auto-GAS to more generative models in future work. **Social impact.** Our Auto-GAS focuses on technology improvement without social and ethical implications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 603, + 246, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 246, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 246, + 616 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "The research was supported by Theme-based Research Scheme (T45-205/21-N) from Hong Kong RGC, and Generative AI Research and Development Centre from InnoHK." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "text", + "content": "1. Abdelfattah, M.S., Mehrotra, A., Dudziak, L., Lane, N.D.: Zero-cost proxies for lightweight nas. arXiv preprint arXiv:2101.08134 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 196 + ], + "type": "text", + "content": "2. Akhauri, Y., Munoz, J.P., Jain, N., Iyer, R.: EZNAS: Evolving zero-cost proxies for neural architecture scoring. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) NeurIPS (2022), https://openreview.net/forum?id=1SqaDG4dvdt" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 219 + ], + "type": "text", + "content": "3. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. ICLR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 220, + 480, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 242 + ], + "type": "text", + "content": "4. Baker, B., Gupta, O., Naik, N., Raskar, R.: Designing neural network architectures using reinforcement learning. In: ICLR (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 480, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 264 + ], + "type": "text", + "content": "5. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "type": "text", + "content": "6. Cavagnero, N., Robbiano, L., Pistilli, F., Caputo, B., Averta, G.: Entropic score metric: Decoupling topology and size in training-free nas. In: ICCV workshop (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "text", + "content": "7. Chen, K., Yang, L., Chen, Y., Chen, K., Xu, Y., Li, L.: Gp-nas-ensemble: a model for the nas performance prediction. In: CVPRW (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 321, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 352 + ], + "type": "text", + "content": "8. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Fourteenth International Conference on Artificial Intelligence and Statistics (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "type": "text", + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. CVPR (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 388, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 408 + ], + "type": "text", + "content": "0. Dong, P., Li, L., Wei, Z.: Diswot: Student architecture search for distillation without training. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "text", + "content": "1. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: ICCV (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "text", + "content": "2. Dong, P., Niu, X., Li, L., Tian, Z., Wang, X., Wei, Z., Pan, H., Li, D.: Rd-nas: Enhancing one-shot supernet ranking ability via ranking distillation from zero-cost proxies. arXiv preprint arXiv:2301.09850 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "type": "text", + "content": "3. Dong, P., Niu, X., Li, L., Xie, L., Zou, W., Ye, T., Wei, Z., Pan, H.: Prior-guided one-shot neural architecture search. arXiv preprint arXiv:2206.13329 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "text", + "content": "4. Dong, P., Niu, X., Tian, Z., Li, L., Wang, X., Wei, Z., Pan, H., Li, D.: Progressive meta-pooling learning for lightweight image classification model. In: ICASSP (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 521, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 552 + ], + "type": "text", + "content": "5. Duan, Y., Chen, X., Xu, H., Chen, Z., Liang, X., Zhang, T., Li, Z.: Transnas-bench-101: Improving transferability and generalizability of cross-task neural architecture search. In: CVPR. pp. 5251-5260 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 555, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 575 + ], + "type": "text", + "content": "6. Elsken, T., Metzen, J.H., Hutter, F.: Neural architecture search: A survey. arXiv preprint arXiv:1808.05377 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "type": "text", + "content": "7. Gao, C., Chen, Y., Liu, S., Tan, Z., Yan, S.: Adversarialnas: Adversarial neural architecture search for gans. In: CVPR (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "8. Gong, X., Chang, S., Jiang, Y., Wang, Z.: Autogan: Neural architecture search for generative adversarial networks. ICCV (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 622, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 642 + ], + "type": "text", + "content": "9. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "20. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. NeurIPS (2017)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "text", + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "21. He, H., Wang, H., Lee, G.H., Tian, Y.: Probgan: Towards probabilistic gan with theoretical guarantees. In: ICLR (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "22. Hu, Y., Wang, X., Li, L., Gu, Q.: Improving one-shot nas with shrinking-and-expanding supernet. Pattern Recognition (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "text", + "content": "23. Isola, P., Zhu, J., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017). https://doi.org/10.1109/CVPR.2017.632" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "text", + "content": "24. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "text", + "content": "25. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (June 2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 235, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 257 + ], + "type": "text", + "content": "26. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.E.: Similarity of neural network representations revisited. In: ICML (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "text", + "content": "27. Krizhevsky, A., Nair, V., Hinton, G.: The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "text", + "content": "28. Lee, N., Ajanthan, T., Torr, P.H.: Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "29. Li, G., Yang, Y., Bhardwaj, K., Marculescu, R.: Zico: Zero-shot NAS via inverse coefficient of variation on gradients. In: The Eleventh ICLR (2023), https://openreview.net/forum?id=rwo-1s5GqGn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "text", + "content": "30. Li, H., Fu, T., Dai, J., Li, H., Huang, G., Zhu, X.: Autoloss-zero: Searching loss functions from scratch for generic tasks. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "31. Li, L.: Self-regulated feature learning via teacher-free feature distillation. In: ECCV (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "32. Li, L., Bao, Y., Dong, P., Yang, C., Li, A., Luo, W., Liu, Q., Xue, W., Guo, Y.: Detkds: Knowledge distillation search for object detectors. In: ICML (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "33. Li, L., Dong, P., Li, A., Wei, Z., Yang, Y.: Kd-zero: Evolving knowledge distiller for any teacher-student pairs. NeuIPS (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "34. Li, L., Dong, P., Wei, Z., Yang, Y.: Automated knowledge distillation via monte carlo tree search. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "type": "text", + "content": "35. Li, L., Jin, Z.: Shadow knowledge distillation: Bridging offline and online knowledge transfer. In: NeuIPS (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "type": "text", + "content": "36. Li, L., Sun, H., Dong, P., Wei, Z., Shao, S.: Auto-das: Automated proxy discovery for training-free distillation-aware architecture search. In: ECCV (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 482, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 503 + ], + "type": "text", + "content": "37. Li, L., Wang, Y., Yao, A., Qian, Y., Zhou, X., He, K.: Explicit connection distillation. In: ICLR (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 503, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 480, + 525 + ], + "type": "text", + "content": "38. Li, M., Lin, J., Ding, Y., Liu, Z., Zhu, J.Y., Han, S.: Gan compression: Efficient architectures for interactive conditional gans. CVPR (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "text", + "content": "39. Lin, M., Wang, P., Sun, Z., Chen, H., Sun, X., Qian, Q., Li, H., Jin, R.: Zen-nas: A zero-shot nas for high-performance image recognition. ICCV (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "text", + "content": "40. Lin, Q., Fang, Z., Chen, Y., Tan, K.C., Li, Y.: Evolutionary architectural search for generative adversarial networks. IEEE Transactions on Emerging Topics in Computational Intelligence (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 578, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 611 + ], + "type": "text", + "content": "41. Liu, H., Simonyan, K., Yang, Y.: DARTS: differentiable architecture search. In: 7th ICLR, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. vol. abs/1806.09055 (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "42. Liu, L., Zhang, Y., Deng, J., Soatto, S.: Dynamically grown generative adversarial networks. AAAI (2021)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "43. Liu, Y., Yu, S., Lin, T.: Hessian regularization of deep neural networks: A novel approach based on stochastic estimators of hessian trace. Neurocomputing 536, 13-20 (2023)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "44. Luo, Y., Zhang, Y., Cai, X., Yuan, X.: E2gan: End-to-end generative adversarial network for multivariate time series imputation. In: 28th international joint conference on artificial intelligence. pp. 3094-3100. AAAI Press Palo Alto, CA, USA (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "text", + "content": "45. Mellor, J., Turner, J., Storkey, A., Crowley, E.J.: Neural architecture search without training. In: ICML (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 203 + ], + "type": "text", + "content": "46. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. ArXiv abs/1802.05957 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 204, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 204, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 204, + 481, + 224 + ], + "type": "text", + "content": "47. Odena, A., Olah, C., Shlens, J.: Conditional image synthesis with auxiliary classifier gans. In: ICML. PMLR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 225, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 481, + 246 + ], + "type": "text", + "content": "48. Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: CVPR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 247, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 247, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 481, + 267 + ], + "type": "text", + "content": "49. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 268, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 268, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 268, + 481, + 289 + ], + "type": "text", + "content": "50. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR abs/1511.06434 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 290, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 290, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 133, + 290, + 481, + 310 + ], + "type": "text", + "content": "51. Real, E., Liang, C., So, D.R., Le, Q.V.: Automl-zero: Evolving machine learning algorithms from scratch (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 311, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 311, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 481, + 342 + ], + "type": "text", + "content": "52. Shao, S., Dai, X., Yin, S., Li, L., Chen, H., Hu, Y.: Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 343, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 343, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 343, + 481, + 364 + ], + "type": "text", + "content": "53. Shu, H., Wang, Y., Jia, X., Han, K., Chen, H., Xu, C., Tian, Q., Xu, C.: Coevolutionary compression for unpaired image translation. In: ICCV (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 481, + 385 + ], + "type": "text", + "content": "54. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information. arXiv:1703.00810 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "type": "text", + "content": "55. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information. arXiv preprint, arXiv:1703.00810 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 407, + 481, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 481, + 428 + ], + "type": "text", + "content": "56. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 429, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 429, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 429, + 481, + 449 + ], + "type": "text", + "content": "57. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 450, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 481, + 471 + ], + "type": "text", + "content": "58. Tanaka, H., Kunin, D., Yamins, D.L., Ganguli, S.: Pruning neural networks without any data by iteratively conserving synaptic flow. NeurIPS 33, 6377-6389 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 472, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 481, + 503 + ], + "type": "text", + "content": "59. Tian, Y., Shen, L., Su, G., Li, Z., Liu, W.: Alphagan: Fully differentiable architecture search for generative adversarial networks. arXiv preprint arXiv:2006.09134 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 504, + 481, + 536 + ], + "type": "text", + "content": "60. Tian, Y., Wang, Q., Huang, Z., Li, W., Dai, D., Yang, M., Wang, J., Fink, O.: Off-policy reinforcement learning for efficient and effective gan architecture search. In: ECCV (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 481, + 557 + ], + "type": "text", + "content": "61. Tishby, N., Zaslavsky, N.: Deep learning and the information bottleneck principle. In: ITW. pp. 1-5. IEEE (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 558, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 558, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 558, + 481, + 578 + ], + "type": "text", + "content": "62. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 579, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 579, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 133, + 579, + 481, + 600 + ], + "type": "text", + "content": "63. Wang, C., Zhang, G., Grosse, R.: Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376 (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 601, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 481, + 621 + ], + "type": "text", + "content": "64. Wang, C., Xu, C., Yao, X., Tao, D.: Evolutionary generative adversarial networks. IEEE Transactions on Evolutionary Computation 23(6), 921-934 (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 133, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 481, + 643 + ], + "type": "text", + "content": "65. Wang, H., Huan, J.: Agan: Towards automated design of generative adversarial networks. arXiv preprint arXiv:1906.11080 (2019)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "text", + "content": "66. Wang, W., Sun, Y., Halgamuge, S.: Improving MMD-GAN training with repulsive loss function. In: ICLR (2019)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 447, + 102 + ], + "type": "text", + "content": "Automated Proxy Discovery for Training-free Generative Architecture Search" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 369 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "67. Wei, Z., Pan, H., Li, L.L., Lu, M., Niu, X., Dong, P., Li, D.: Convformer: Closing the gap between cnn and vision transformers. arXiv preprint arXiv:2209.07738 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "68. Wolchover, N., Reading, L.: New theory cracks open the black box of deep learning. Quanta Magazine (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "69. Xiaolong, L., Lujun, L., Chao, L., Yao, A.: Norm: Knowledge distillation via n-to-one representation matching. In: ICLR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "70. Ying, G., He, X., Gao, B., Han, B., Chu, X.: Eagan: Efficient two-stage evolutionary architecture search for gans. In: ECCV. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "71. Yu, A., Grauman, K.: Fine-grained visual comparisons with local learning. CVPR (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "72. Zhang, M., Li, H., Pan, S., Chang, X., Su, S.: Overcoming multi-model forgetting in one-shot nas with diversity maximization. In: CVPR. pp. 7809-7818 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "73. Zhong, Z., Yang, Z., Deng, B., Yan, J., Wu, W., Shao, J., Liu, C.L.: Blockqnn: Efficient block-wise neural network architecture generation. IEEE transactions on pattern analysis and machine intelligence (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "text", + "content": "74. Zhu, C., Li, L., Wu, Y., Sun, Z.: Saswot: Real-time semantic segmentation architecture search without training. In: AAAI (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "75. Zimian Wei, Z., Li, L.L., Dong, P., Hui, Z., Li, A., Lu, M., Pan, H., Li, D.: Autoprox: Training-free vision transformer architecture search via automatic proxy discovery. In: AAAI (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "76. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures for scalable image recognition. In: CVPR (2018)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Lujun Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_content_list.json b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e81444c6d240eec16d1583a24f5a166cfbd23101 --- /dev/null +++ b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_content_list.json @@ -0,0 +1,1878 @@ +[ + { + "type": "text", + "text": "AutoDIR: Automatic All-in-One Image Restoration with Latent Diffusion", + "text_level": 1, + "bbox": [ + 277, + 140, + 727, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 254, + 210, + 746, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Chinese University of Hong Kong 2Shanghai AI Laboratory {ytjiang@link, zhaoyangzhang@link, tfxue@ie, jwgu@cse}.cuhk.edu.hk", + "bbox": [ + 243, + 239, + 756, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present AutoDIR, an innovative all-in-one image restoration system incorporating latent diffusion. AutoDIR excels in its ability to automatically identify and restore images suffering from a range of unknown degradations. AutoDIR offers intuitive open-vocabulary image editing, empowering users to customize and enhance images according to their preferences. AutoDIR consists of two key stages: a Blind Image Quality Assessment (BIQA) stage based on a semantic-agnostic vision-language model which automatically detects unknown image degradations for input images, an All-in-One Image Restoration (AIR) stage utilizes structural-corrected latent diffusion which handles multiple types of image degradations. Extensive experimental evaluation demonstrates that AutoDIR outperforms state-of-the-art approaches for a wider range of image restoration tasks. The design of AutoDIR also enables flexible user control (via text prompt) and generalization to new tasks as a foundation model of image restoration. Project is available at: https://jiangyitong.github.io/AutoDIR_webpage/.", + "bbox": [ + 261, + 306, + 743, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 556, + 375, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Restoring scene details and improving image quality is often the very first step of any computer vision system, which significantly determines the overall system performance. Depending on the operating environments and the signal processing pipelines, images captured by real-world computer vision systems often undergo multiple unknown degradations, such as noise, resolution loss, motion blur, defocus, color imbalance, chromatic aberration, haze, flare, distortion, etc. Previous image restoration methods often address these basic image degradations separately, including deblurring [23,63], denoising [45,76] deraining [24,48], super-resolution [7,60], low-light enhancement [37,75], deraindrop [43], and dehaze [44,53], by using specific single-task models. While these single-task approaches yield promising results within their respective tasks, they encounter difficulties when applied to complex real-world situations that involve multiple unknown degradations or require multiple steps of enhancement.", + "bbox": [ + 212, + 588, + 787, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore a more general model capable of handling multiple unknown degradations for a single image. To achieve this goal, the corresponding", + "bbox": [ + 215, + 785, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 232, + 825, + 377, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a5c11c6de9c5bac36bcab16b7d0996d6005eccd2bf5ec845dd9b541c5bc33c4.jpg", + "image_caption": [ + "Fig. 1: We propose AutoDIR, an automatic all-in-one model for image restoration capable of handling multiple types of image degradations. Left: For images with multiple unknown degradations, AutoDIR automatically decomposes the task into multiple steps and supports user interaction via an intuitive text prompt. Right: AutoDIR effectively restores clean images from different degradations and can handle images with unknown degradations in unseen tasks. (Please zoom in for details)" + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 588, + 320 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5c50eff13c1b3380bde7eb038e1e144a7dd77521512b2423832b8217b75165eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 143, + 784, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model should possess the following abilities: (1) decomposing and distinguishing unknown degradations, (2) restoring various degradations in a task-agnostic framework, and (3) allowing users to freely tune restoration results according to their visual preferences. Numerous related efforts have been made to solve similar problems, but none of them can simultaneously satisfy all three points.", + "bbox": [ + 212, + 444, + 787, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [13, 66, 74] have explored assessing unknown degradations by training an image classifier and using this information to explicitly select a specific image restoration model [8, 33, 69]. However, these methods require training a specific model for each task and primarily focus on a small set of tasks (e.g., 3 tasks). Experiments demonstrate that on large datasets with a wide range of degradations, training an accurate image degradation classifier solely based on image information is still challenging (See Tab. 1).", + "bbox": [ + 212, + 521, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other works attempt to unify multiple image restoration methods [26, 62] in a single model by utilizing generative model priors. These methods, however, either require known degradation formulations or can only handle images with a single degradation from a predefined small set [31, 42], thus unable to implicitly distinguish multiple unknown degradations.", + "bbox": [ + 212, + 627, + 787, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a pipeline named AutoDIR, which satisfies all three aforementioned abilities and can automatically detect and restore images with multiple unknown degradations. AutoDIR consists of two stages: the Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA) stage and the All-in-One Image Restoration (AIR) stage guided by text prompts generated in SA-BIQA.", + "bbox": [ + 212, + 703, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the quality assessment stage, we enable accurate identification of each degradation in cases of unknown artifacts in an open-vocabulary manner. This is achieved through our proposed Semantic-Agnostic CLIP (SA-CLIP) model, which employs a Semantic-Agnostic regularization term to transfer the origi", + "bbox": [ + 212, + 780, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 284, + 113, + 738, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nal Semantic-Recognitive CLIP to a semantic-agnostic form, focusing on image structural quality rather than semantic content. Additionally, we can utilize the text embeddings generated by the SA-BIQA stage as instructions to guide the further restoration model. This approach can not only enable effective restoration but also allows for flexible user control and editing at runtime by providing open-vocabulary instructions, as shown in Fig. 1.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The restoration stage is to handle degradations using a multitasking image restoration model jointly trained on a wide range of tasks. Given the diverse nature of different tasks (e.g., some tasks like super-resolution need to hallucinate textures, but other tasks like low-light enhancement need to preserve everything except brightness), we propose a hybrid approach that maximizes the hallucinating ability of diffusion models while preserving the image structural consistency by involving extra structural inductive bias.", + "bbox": [ + 212, + 237, + 784, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate the effectiveness and generalization ability of AutoDIR, we have conducted a comprehensive set of experiments encompassing seven image restoration tasks, including denoising, motion deblurring, low-light enhancement, dehazing, deraining, deraindrop, and super-resolution. Experimental results show that AutoDIR consistently outperforms state-of-the-art methods. AutoDIR is also evaluated for the restoration of images captured by under-display cameras and underwater cameras, which are examples of imaging systems with multiple unknown degradations.", + "bbox": [ + 212, + 343, + 785, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 488, + 387, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unified All-in-One Image Restoration Previous approaches for unified all-in-one image restoration can be categorized into two main groups: unsupervised generative prior-based methods [2,5,15,21,26,36,40,62,65] and end-to-end supervised learning-based methods [12,31,32,41,57,71]. Generative priors for image restoration have been widely exploited for image restoration, in the form of generative adversarial networks (GANs) inversion and filters [1,2,5,17,21,36,40], Vector Quantization codebooks [7,34,58,78], latent diffusion models [1,15,26,62]. Nevertheless, these methods rely on specific known image degradation functions, which constrain the types of degradations they can handle. End-to-end supervised learning-based methods typically learn image embeddings extracted by an auxiliary degradation predictor to guide the image restoration model [31,41,57,71]. For example, Valanarasu et al. [57] leverage a transformer encoder to capture hierarchical features of haze, while Li et al. [31] employ a degradation classifier trained using contrastive learning. Park et al. [41] design a degradation classifier for multiple types of degradation to select appropriate Adaptive Discriminant filters, altering network parameters based on specific degradations. However, these methods are constrained to handle limited image degradation categories and cannot restore images with multiple unknown degradations. In contrast, with an SA-BIQA stage, the proposed AutoDIR is able to unify complex degradations without relying on assumptions about the specific degradation functions.", + "bbox": [ + 212, + 522, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "AutoDIR", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-based Image Synthesis and Editing Early works [55,64,67,70,80] on text-to-image synthesis and editing primarily relied on generative adversarial networks (GANs) [16]. Recently, developments in diffusion models [19, 49] and language-vision models [46] have led to significant progress in text-based image synthesis and editing [6,39,47,50]. The emergence of diffusion-based methods has offered new avenues for text-to-image editing, which can be broadly categorized into data-driven [4] and no-extra-data-required approaches [18,27,77]. In the data-driven category, Brooks et al. [4] employ a stable diffusion model fine-tuned with a large dataset of text prompt and image pairs. However, it is primarily designed for semantic image editing and struggles to deliver satisfactory results in tasks related to image quality enhancement. On the other hand, some methods focus on image editing without extra training data. For instance, recent works interpolate the text embedding [27] or score estimate [77] of the input and desired images. However, it requires individual finetuning of the diffusion model for each input image, leading to time-consuming operations. Hertz et al. [18] improve editing efficiency by directly manipulating cross-attention maps without the need for per-image finetuning. However, it requires interior maps in the reversing process and is thus not applicable to real image editing. Unlike the above methods, AutoDIR supports real image enhancement without per-image finetuning.", + "bbox": [ + 212, + 145, + 787, + 449 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3846f06a49da9827c19d7fae7d4e2f056c7d781fc8be5b354f86e20c6bdd906.jpg", + "image_caption": [ + "Fig. 2: Diagram of the proposed All-in-One Image Restoration with Latent Diffusion (AutoDIR). Refer to Sec. 3 for more details." + ], + "image_footnote": [], + "bbox": [ + 215, + 503, + 787, + 739 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 282, + 112, + 738, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Our Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2 shows the overall flowchart of the proposed AutoDIR (Automatic All-in-One Image Restoration with Latent Diffusion), a unified model that can automatically detect and address multiple unknown degradations in images. AutoDIR comprises two main stages:", + "bbox": [ + 212, + 176, + 787, + 238 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA): This stage automatically identifies the predominant degradations present in the input image, such as noise, blur, haze, and generates a corresponding text prompt, denoted as $e_{auto}$ , which is subsequently used in the image restoration process.", + "- All-in-one Image Restoration (AIR): This stage utilizes a Structural-Correction Latent Diffusion Model(SC-LDM) to produce a restored image $I_{res}$ , guided by the text embedding $e_{auto}$ from SA-BIQA or a user-defined open-vocabulary instruction $e_{user}$ ." + ], + "bbox": [ + 223, + 247, + 795, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA)", + "text_level": 1, + "bbox": [ + 212, + 390, + 785, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As demonstrated by previous works [13, 66, 74], a common approach to assess image degradation is to train an image classifier specifically for that purpose. However, this naive approach faces challenges when dealing with large datasets containing a wide range of artifacts. Even with the use of a heavy Vision Transformer(ViT) encoder, the accuracy of classifying image degradation solely based on image information is limited to $77.65\\%$ , as shown in Tab. 1.", + "bbox": [ + 212, + 416, + 785, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this limitation, we propose leveraging human language knowledge to enhance the detection of image degradation. We introduce a Semantic-Agnostic CLIP (SA-CLIP) model as our backbone for Blind Image Quality Assessment (BIQA). SA-CLIP builds upon the CLIP model, which establishes a connection between human language knowledge and image quality. However, we observed that directly applying CLIP or naively fine-tuning it for BIQA tasks does not yield reliable results, as we will later show in Tab. 1. We delved into this problem and identified the reason behind this issue. The pre-trained CLIP model is primarily trained for visual recognition tasks, which prioritize semantic information rather than image quality. Consequently, it delivers low accuracy in BIQA tasks. For instance, the model may struggle to differentiate between a low-light dog image and a noisy dog image because it focuses more on the \"dog\" aspect rather than the presence of noise or illumination.", + "bbox": [ + 212, + 507, + 785, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To overcome this, we tackle the problem in two steps: (i) We construct a new image quality assessment task for fine-tuning CLIP. (ii) We propose a new regularization term for semantic-agnostic and image-quality-aware training to derive SA-CLIP model.", + "bbox": [ + 212, + 704, + 785, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, let $\\mathcal{C}$ denotes the set of image distortions, i.e. $\\mathcal{C} = \\{c_1, c_2, \\dots, c_{K-1}, c_K\\}$ , where $c_i$ is an image distortion type, e.g. \"noise\" and $K-1$ is the number of distortions we consider. We also add a special type $c_K = \"\\mathrm{no}\"$ to indicate the end of the automatic multiple-step restoration. The textural description prompt set is $\\mathcal{T} = \\{T | T = \"\\text{A photo needs } \\{c_i\\} \\text{ artifact reduction.}\"\\}$ .", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "AutoDIR", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/273b652ad63ce3fd963634df577ac5892a5b22d5efd20da325af629726f6394f.jpg", + "image_caption": [ + "a) Original CLIP" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 372, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89f83f10074e8d1e8455214b5bb9b4ae0cd89108760557b78b2443736a59dadd.jpg", + "image_caption": [ + "b) Finetuned CLIP", + "Fig. 3: t-SNE visualization of image embeddings $\\mathcal{E}_{\\mathcal{I}}(I)$ of CLIP of Blind Image Quality Assessment (BIQA) on SOTs dehazing dataset [30]. Image embeddings of foggy images and their ground-truth clean images are extracted by a) original CLIP. b) finetuned CLIP. c) SA-CLIP finetuned with semantic-agnostic constraint. This illustrates that semantic-agnostic constraint can separate the embeddings of the degraded images from the clean images, while original CLIP and finetuned CLIP features cannot." + ], + "image_footnote": [], + "bbox": [ + 372, + 143, + 524, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d2556025d7a9319662fe8fdbe3d24fcbac81eac7c485758ceaf9e311c7909f4.jpg", + "image_caption": [ + "c) Finetuned CLIP with Semantic-Agnostic constraint" + ], + "image_footnote": [], + "bbox": [ + 527, + 143, + 785, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$c \\in \\mathcal{C}$ . Given a corrupted image $I$ which undergoes several unknown artifacts, our Semantic-Agnostic CLIP aims to identify the dominant degradation of $I$ and extract the corresponding text embedding. SA-CLIP consists of an image encoder $\\mathcal{E}_{\\mathcal{I}}$ and a text encoder $\\mathcal{E}_{\\mathcal{T}}$ . We first obtain the image embedding $\\mathcal{E}_{\\mathcal{I}}(I) \\in \\mathbb{R}^d$ and the text embedding $\\mathcal{E}_{\\mathcal{T}}(T) \\in \\mathbb{R}^{K \\times d}$ , then we compute the cosine similarity between image embedding $\\mathcal{E}_{\\mathcal{I}}(I)$ and each prompt embedding $\\mathcal{E}_{\\mathcal{T}}(T_i) \\in \\mathbb{R}^d$ by:", + "bbox": [ + 214, + 369, + 787, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o g i t} \\left(c _ {i} \\mid I\\right) = \\frac {\\mathcal {E} _ {\\mathcal {I}} (I) \\cdot \\mathcal {E} _ {\\mathcal {T}} \\left(T _ {i}\\right)}{\\| \\mathcal {E} _ {\\mathcal {I}} (I) \\| _ {2} \\| \\mathcal {E} _ {\\mathcal {T}} \\left(T _ {i}\\right) \\| _ {2}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 470, + 785, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T_{i}$ is the $i$ -th element of $\\mathcal{T}$ . Next, we calculate the probability $\\hat{p}(c_i|I)$ and obtain the output text embedding of BIQA $e_{auto}$ by:", + "bbox": [ + 214, + 515, + 787, + 546 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} \\left(c _ {i} \\mid I\\right) = \\frac {\\exp \\left(\\operatorname {l o g i t} \\left(c _ {i} \\mid I\\right)\\right)}{\\sum_ {i = 1} ^ {K} \\exp \\left(\\operatorname {l o g i t} \\left(c _ {i} \\mid I\\right)\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 556, + 785, + 592 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ne _ {a u t o} = \\sum_ {i = 1} ^ {K} \\hat {p} (c _ {i} | I) \\mathcal {E} _ {\\mathcal {T}} (T _ {i}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 604, + 785, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Naive fine-tuning on image quality assessment During the optimization of CLIP model, we freeze the parameters of the text encoder $\\mathcal{E}_{\\mathcal{T}}$ and finetune the image encoder $\\mathcal{E}_{\\mathcal{I}}$ using the multi-class fidelity loss [56]. The fidelity loss can be denoted as:", + "bbox": [ + 214, + 659, + 787, + 717 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {F I D} = 1 - \\sum_ {i = 1} ^ {K} \\sqrt {y \\left(c _ {i} | I\\right) \\hat {p} \\left(c _ {i} | I\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 715, + 785, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $y(c_{i}|I)$ is a binary variable that equals 1 if $c_{i}$ is the dominant degradation in image $I$ , and 0 otherwise.", + "bbox": [ + 214, + 763, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic-Agnostic Constraint fine-tuning on image quality assessment Since the original CLIP model is pre-trained on tasks such as image", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 282, + 112, + 738, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "classification, its corresponding $\\mathcal{E}_{\\mathcal{I}}(I)$ encoders tend to encode images based on their semantic information (e.g., cat or dog) rather than their image quality (e.g., noisy or clean). When we fine-tune the CLIP model for generating texts for BIQA according to image quality, this becomes a significant limitation. As shown in Fig. 3 a) and b), the image embeddings extracted by the original CLIP and the finetuned CLIP on foggy, as well as their corresponding ground-truth clean images, cannot be separated, indicating a focus on semantic information rather than image quality differences.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address this issue, we propose a novel approach called the semantic-agnostic constraint loss $(L_{SA})$ to regulate the fine-tuning process and prevent the model from relying solely on semantic information rather than image quality. The semantic-agnostic loss $(L_{SA})$ applies a penalty when the CLIP model suggests that the artifact $c_{i}$ is present in the ground-truth clean image $I_{gt}$ which corresponds to the degraded image $I$ . This penalty forces the CLIP model to distinguish between $I_{gt}$ and $I$ based on image quality, encouraging the CLIP image encoder $(\\mathcal{E}_{\\mathcal{I}}(I))$ to focus on extracting image quality information rather than semantic information. This constraint loss can be derived using the following equation:", + "bbox": [ + 212, + 268, + 784, + 416 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {S A} = \\sum_ {i = 1} ^ {K} \\sqrt {y \\left(c _ {i} | I\\right) \\hat {p} \\left(c _ {i} \\mid I _ {g t}\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 419, + 784, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $y(c_{i}|I)$ is a binary variable that is equal to 1 if $c_{i}$ represents the dominant degradation in image $I$ , and 0 otherwise, and $\\hat{p}(c_{i}|I_{gt})$ denotes the probability that the dominant degradation in $I_{gt}$ is represented by $c_{i}$ .", + "bbox": [ + 212, + 465, + 782, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We incorporate the semantic-agnostic constraint $L_{SA}$ with the fidelity loss $L_{F1D}$ , resulting in the total loss $L_{BIQA}$ for SA-CLIP model:", + "bbox": [ + 212, + 511, + 782, + 542 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {B I Q A} = L _ {F I D} + \\lambda L _ {S A}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 556, + 784, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where we set $\\lambda = 1$ in our experiments. As shown in Fig. 3 c), CLIP fine-tuned with a semantic-agnostic constraint can separate the embeddings of the degraded images from the clean images as expected.", + "bbox": [ + 212, + 583, + 782, + 628 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/285bcdbf809575f1085a9305cc6b14b4622de653c6e44321a81c786f8bc84c82.jpg", + "table_caption": [ + "Table 1: Average accuracy of image degradation detection on seven degradation tasks and clean image. For details, please refer to the supplementary." + ], + "table_footnote": [], + "table_body": "
MethodsViT-classifierOriginal CLIPFine-tuned CLIPSA-CLIP (ours)
Average Accuracy77.65%61.96%83.31%97.94%
", + "bbox": [ + 263, + 684, + 743, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Effectiveness of Semantic-Agnostic CLIP", + "text_level": 1, + "bbox": [ + 212, + 782, + 598, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our SA-CLIP with a ViT-based [14] classifier, original CLIP, and naive fine-tuned CLIP to demonstrate the effectiveness of the Semantic-Agnostic", + "bbox": [ + 212, + 808, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "AutoDIR", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df162e78af1b30ba7ad3190ce0d9498d12736c65101fd09a82cfa52a67470a8b.jpg", + "image_caption": [ + "Fig. 4: Structural-Correction Latent Diffusion Model (SC-LDM) can maintain complex structures of the original images, while Latent Diffusion Model (LDM) fails to." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training scheme. Tab. 1 presents the effectiveness of our Semantic-Agnostic training scheme. The results demonstrate that our SA-CLIP, which incorporates human language information, achieves higher accuracy compared to the ViT-based classifier which considers only image information. Moreover, when CLIP is finetuned with Semantic-Agnostic regularization, the model focuses more on image quality information rather than semantic information, resulting in superior performance compared to the original pre-trained CLIP and CLIP fine-tuned without Semantic-Agnostic regularization. For a visual demonstration of the effectiveness of the Semantic-Agnostic constraint, please refer to the supplementary materials.", + "bbox": [ + 212, + 268, + 787, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 All-in-one Image Restoration (AIR)", + "text_level": 1, + "bbox": [ + 215, + 446, + 555, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Naively applying LDM on multi-task image restoration The All-in-One Image Restoration (AIR) stage aims to handle multiple degradations in a single shared framework. Recent advancements in diffusion-based generative models [4, 49] have demonstrated their remarkable ability to generate diverse images, making them suitable for multi-task image restoration. Previous studies [7, 60] have shown that generative models possess a remarkable ability to generate missing or distorted details, especially for tasks that require hallucination, such as super-resolution. Based on these insights, we conduct the AIR stage based on the Latent Diffusion Model (LDM) [49]. LDM incorporates both text and image embedding conditions to restore images $I_{sd}$ using generative priors. The text embedding condition $e = \\{e_{auto}, e_{user}\\}$ aims to separate different types of image degradations, while the latent image embedding condition $z_I = \\mathcal{E}_{ldm}(I)$ from the image encoder $\\mathcal{E}_{ldm}$ of the LDM provides structural information.", + "bbox": [ + 212, + 476, + 787, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, although LDM-based generative models can provide a foundation for multi-task image restoration, they have limitations in reconstructing images with complex and small structures due to the compression-reconstruction process with Variational Autoencoder (VAE) [29], as discussed in [9, 20, 72]. [20] attempted to reduce distortion caused by the compression-reconstruction process by retraining Variational Autoencoder (VAE) networks on category-specific images (e.g., human faces) to learn specialized probability distributions. However, this approach is not suitable for image restoration tasks due to the diverse image content. To address these limitations, we introduce a lightweight plug-in structural-correction module to the LDM, enhancing its ability to handle complex and small structures during image restoration.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 284, + 113, + 738, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d55cabfa1fa34f83a49d6defe0d3007f40186590d6bea43c105304fa4b38f268.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 349, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3d59953ed90b4723c4590c13a66ddb9279141a63dc7f017db8e06e51f70fd7d3.jpg", + "image_caption": [ + "...haze ..." + ], + "image_footnote": [], + "bbox": [ + 351, + 143, + 483, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5edc56324eeb178111d7e0db15c7ab0f413a74d0fee5711ee57468a9a22715c1.jpg", + "image_caption": [ + "...low-resolution..." + ], + "image_footnote": [], + "bbox": [ + 485, + 143, + 617, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f0dca7fbfdb041b400b9e3a0fe6195eb62c1d17458ec314235d00cc496744247.jpg", + "image_caption": [ + "... raindrop ..." + ], + "image_footnote": [], + "bbox": [ + 620, + 143, + 751, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7e9a69be4fcca3a1f9dd045795aca623f1e877140cf00771b48145e9a940729a.jpg", + "image_caption": [ + "Fig. 5: Cross-attention maps of text-conditioned diffusion image restoration. The top-left shows an input image with raindrops on the left half. The remaining plots show the cross-attention masks for the keywords \"haze\", \"low-resolution\", and \"raindrop\" in the text prompt used for restoration. While the cross-attention maps for \"haze\" and \"low-resolution\" are more uniformly distributed over the entire image, the map for \"raindrop\" correctly focuses on the actual image artifacts, as expected." + ], + "image_footnote": [], + "bbox": [ + 764, + 146, + 785, + 213 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Structural-Correction Latent Diffusion Model (SC-LDM) While the LDM-based generative models can provide a foundation for multi-tasking image restoration, it is widely noticed that they may fail to maintain original image structures such as human faces and texts [9, 20, 72], as shown in Fig. 4. To address the structure distortion, we employ an efficient structural-correction module (SCM) denoted as $\\mathcal{F}$ . The purpose of SCM is to extract the contextual information $\\mathcal{R}$ from the original image and incorporate it with the intermediate image restoration result $I_{sd}$ , decoded by image decoder $\\mathcal{D}$ of LDM, in a residual manner. This is achieved through the following equation:", + "bbox": [ + 212, + 344, + 784, + 481 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nI _ {r e s} = I _ {s d} + w \\cdot \\mathcal {F} ([ I _ {s d}, I ]), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 494, + 784, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $[\\cdot, \\cdot]$ denotes concatenation, and $w$ is an adjustable coefficient that ranges between 0 and 1. The value of $w$ determines the extent to which contextual information is utilized to recover the final result. A larger value of $w$ emphasizes the use of contextual information, which is beneficial for tasks that require structure consistency, such as low light enhancement. Conversely, a smaller value of $w$ is often employed to maintain the generation capability of latent diffusion model for tasks like super-resolution. By integrating SCM, AutoDIR effectively restores the distorted context of the original image, seamlessly incorporating the enhancements made during the editing phase as shown in Fig. 4.", + "bbox": [ + 212, + 525, + 782, + 661 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During training, we fine-tune the UNet backbone $\\epsilon_{\\theta}(e,[z_t,z_I],t)$ of the Latent Diffusion Model (LDM) for image restoration tasks. The objective function is defined as:", + "bbox": [ + 212, + 662, + 782, + 705 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L D} = \\mathbb {E} _ {\\mathcal {E} _ {l d m} (x), z _ {I}, e, \\epsilon \\sim \\mathcal {N} (0, 1), t} [ \\| \\epsilon - \\epsilon_ {\\theta} (e, [ z _ {t}, z _ {I} ], t) \\| _ {2} ^ {2} ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 720, + 784, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $z_{t}$ is the noisy latent. For the SC-LDM, instead of generating the edited latent $\\hat{z}_{t}$ with the complete reversing sampling process, which is time-consuming, we utilize the estimated edited latent $\\tilde{z}$ , which is calculated by:", + "bbox": [ + 212, + 751, + 782, + 797 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} = \\frac {z _ {t}}{\\sqrt {\\bar {\\alpha}}} - \\frac {\\sqrt {1 - \\bar {\\alpha}} \\epsilon_ {\\theta} (e , [ z _ {t} , z _ {I} ] , t)}{\\sqrt {\\bar {\\alpha}}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 809, + 784, + 844 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "AutoDIR", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0d5f5f0b10595e76b7a79643c4c9ca9777641d646e68c664668f55824ce3d80e.jpg", + "image_caption": [ + "Fig. 6: AutoDIR deals with multiple unknown image degradations. Iterative image restoration on unseen Under-Display Camera (TOLED) dataset [79] and unseen Enhancing Underwater Visual Perception (EUVP) dataset [22]. Left: AutoDIR decomposes and solves the unseen artifact in several steps. Right: all-in-one AirNet [31], PromptIR [42] and zero-shot method GDP [15] for comparison." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 251 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}$ represents the noise scheduler introduced in [49]. The loss function for Structural-Correction Latent Diffusion Model (SC-LDM) is further defined as:", + "bbox": [ + 212, + 359, + 784, + 390 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nL _ {A I R} = \\left\\| I _ {g t} - \\left(\\mathcal {F} (\\mathcal {D} (\\tilde {z}), I) + \\mathcal {D} (\\tilde {z})\\right) \\right\\| _ {2} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 398, + 784, + 416 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mechanism of handling multi-task image restoration Fig. 5 shows our experiment of exploring the mechanism of the text conditions disentangling different image restoration tasks during the reverse diffusion process. We find that different text conditions yield different cross-attention maps. As shown in Fig. 5, changing the text prompt leads to significant changes in the cross-attention map. The map closely aligns with the text prompt, which either uniformly distributes attention across the entire image for the \"dehazing\" prompt, or concentrates on the part with edges or textures for the \"low-resolution\" prompt, or focuses on specific areas such as raindrops for the \"deraindrop\" prompt. This shows that AutoDIR can guide the diffusion attention toward regions that are more likely to have image artifacts.", + "bbox": [ + 212, + 428, + 787, + 594 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 614, + 375, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To fully validate the feasibility and effectiveness of unifying multiple restoration and enhancement tasks, we compare AutoDIR with both CNN-based [10,31,42], and diffusion-based multitasking [49] backbones on two tasks with unknown image degradations as well as seven representative image restoration tasks. Specifically, we highlight the advantages of our generative model-based pipeline over discriminative models for tasks that require hallucinating details such as superresolution. Please refer to the appendix for implementation details and datasets. The source code and trained models will be released upon publication.", + "bbox": [ + 212, + 645, + 787, + 767 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Results on Multiple Unknown Degradations", + "text_level": 1, + "bbox": [ + 212, + 786, + 625, + 801 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To show that AutoDIR can handle multiple unknown image degradations, we conducted experiments on two additional datasets: the Under-Display-Camera", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 284, + 113, + 738, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5e90bb7cf4472c124e6aed89c31b8587906bff4846463e8a452cbf374470456b.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 356, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b1dfb610abf6620351953d43dd405406be9d6ac6d96a1fe02e06ccd1727ab1c3.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 356, + 143, + 493, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f7e022fe3877da94d8a60c0d1d79e8dd7e0e93d4f8d0c415b1f0f61cf87da2fe.jpg", + "image_caption": [ + "AirNet" + ], + "image_footnote": [], + "bbox": [ + 493, + 143, + 640, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3899a14490feef55417010a0d105b856de2dfbe7a140f5c6963caad7890d5d2e.jpg", + "image_caption": [ + "PromptR" + ], + "image_footnote": [], + "bbox": [ + 640, + 143, + 784, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/656da66404c1876f084626bf334520c6092714e45e5d6c88a64da6c57b9843a7.jpg", + "image_caption": [ + "All-in-One NAFNet", + "Fig.7: Super-resolution qualitative comparisons with AirNet [31], PromptIR [42], NAFNet [10] trained on all-in-one tasks, NAFNet [10] trained on Super-Res, and Latent Diffusion Model (LDM) [49]. See the Appendix for more results (Zoom in for better view)." + ], + "image_footnote": [], + "bbox": [ + 217, + 226, + 356, + 297 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/97f7a3f1b65fb880ebfc4d0b79fdb32c9a2e13d03b764944606fb804980b7566.jpg", + "image_caption": [ + "Super-res NAFNet" + ], + "image_footnote": [], + "bbox": [ + 356, + 226, + 493, + 297 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7310c367a93b8cbb7156ed2d5446092d8f162bb12e48755dee57fe1dadcbbefb.jpg", + "image_caption": [ + "LD" + ], + "image_footnote": [], + "bbox": [ + 493, + 226, + 640, + 297 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/199ac676e73619f2c6c60a34fc9fd9b8c666957e539e0df4b620cc5f220f9301.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 640, + 226, + 784, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(TOLED) dataset [79] and the Enhancing Underwater Visual Perception (EUVP) dataset [22]. These two datasets consist of multiple unknown image degradations and are completely excluded for training, allowing us to evaluate the performance of AutoDIR in handling unknown, diverse image distortions. Fig. 6 demonstrates the iterative process of AutoDIR to handle images from these two datasets. As shown, AutoDIR can automatically decompose the UDC restoration in two steps (\"brightening\" and \"deblurring\"), and the Underwater restoration in two steps (\"deblurring\" and \"dehazing\"), enabling the restoration of input images. For comparison, we also show the results of a recent all-in-one AirNet [31], PromptIR [42], diffusion-based zero-shot image restoration method GDP [15]. Furthermore, as shown in Tab. 3, AutoDIR outperforms all the other black-box and classifier-based all-in-one methods on quantitative perceptual results. We also conducted a user study on 28 images of TOLED and real-captured images by OPPO Reno6 device. We collected 22 forms and 616 responses. And AutoDIR receives $96.9\\%$ of the votes as the best result, please refer to the appendix for details.", + "bbox": [ + 212, + 405, + 787, + 632 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1df20a506b7fa236f3d0ccbe9a7154957a020b53a7e596973990f4da01ef8519.jpg", + "table_caption": [ + "Table 2: Quantitative comparison on SR" + ], + "table_footnote": [], + "table_body": "
MethodSuper Resolution
MUSIQ ↑CLIP-IQA ↑NIQE ↓NIMA ↑LPIPS ↓PSNR ↑SSIM ↑
NAFNet40.770.2928.454.450.27828.520.754
NAFNet-SR45.490.2998.234.560.33227.950.726
AirNet20.210.27611.074.030.42528.140.626
PromptIR22.890.30110.044.160.30929.620.792
LD50.710.3756.444.900.37824.110.649
Ours63.020.5964.625.250.28223.490.599
", + "bbox": [ + 243, + 659, + 563, + 758 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b0743c42d37401fb39b2c2309706b34283d45eaa8b69e7472a05c64763fc34bd.jpg", + "table_caption": [ + "Table 3: Quantitative comparison on Zero-shot UDC task." + ], + "table_footnote": [], + "table_body": "
MethodUnder-displayCamera
MUSIQ ↑CLIP-IQA ↑NIQE ↓NIMA ↑
NAFNet37.160.2787.4024.258
AirNet41.200.2657.6814.048
PromptIR45.400.3038.2944.344
LD43.410.2887.9834.371
Ours49.270.3116.3924.427
", + "bbox": [ + 578, + 672, + 772, + 760 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Results on Multi-task Image Restoration", + "text_level": 1, + "bbox": [ + 215, + 781, + 599, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare with state-of-the-art methods: the CNN-based single-task image restoration pipeline NAFNet [10], CNN-based all-in-one AirNet [31], PromptIR", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "AutoDIR", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9a5b900d554556b63b95f38df8dfbc259a08f7f0422314e38f9bbf3fa9fe50d1.jpg", + "image_caption": [ + "Fig. 8: Qualitative comparisons with NAFNet [10] and Latent Diffusion Model (LDM) [49], AirNet [31] and PromptIR [42] on deraining, dehazing, low light enhancement, and deraindrop, denoise and deblur tasks. See the Appendix for more results. (Zoom in for better view)" + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 785, + 468 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "[42], and the generative backbone Latent Diffusion [49]. They are all fine-tuned for all seven tasks, employing an identical training configuration to AutoDIR. Tab. 2 and Tab. 4 present quantitative results of AutoDIR and other methods in all seven tasks. As demonstrated, AutoDIR achieves competitive performance across all seven tasks. Specifically, to further evaluate the ability of discriminative and generative methods on super-resolution tasks, we include task-specific discriminative NAFNet training on SR. As shown in Tab. 2 and Fig. 7, AutoDIR quantitatively and qualitatively outperforms all the all-in-one discriminative, task-specific discriminative methods and the black-box latent diffusion model. Please be noted that, as discussed in previous works [3,25,68], full-reference metrics like PSNR, SSIM, and LPIPS [73] often do not align with human evaluation on super-resolution tasks. Therefore, following [60,68], we include non-reference metrics MUSIQ [28], CLIP-IQA [59] NIQE [38] and NIMA [54] to better evaluate the results. We also show the qualitative comparison results on the other six image restoration tasks in Fig. 8, which demonstrate that AutoDIR can achieve more visually appealing results than the other baselines. A comprehensive and detailed analysis of the quantitative and qualitative results for each specific task can be found in the appendix.", + "bbox": [ + 212, + 566, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 282, + 112, + 738, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/62924e3e7d90510a80645dc8d31f7c814b61e130b337f91dac72b0e5ef8d62b3.jpg", + "table_caption": [ + "Table 4: Quantitative comparison on denoising, deraining, dehazing, deraindrop, and low light enhancement tasks. The best results are marked in boldface and the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodDenoiseDerainDehazeDeraindropLow lightDeblur
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
NAFNet0.80528.100.1950.92630.460.1000.96827.750.0280.87225.040.1380.82721.150.2400.80526.670.208
LD0.62522.580.1860.65123.210.1470.76323.490.0910.73824.840.1040.77018.970.1590.69522.530.241
AirNet0.80329.100.2130.92930.990.0550.94426.520.0310.89227.130.0940.81821.260.2370.80026.500.201
PromptIR0.82429.890.1810.93833.970.0490.97129.130.0220.90027.410.0810.83122.420.1330.81926.820.198
Ours0.83229.680.1670.96535.090.0470.97329.340.0200.92430.100.0630.88822.370.1290.82827.070.157
", + "bbox": [ + 218, + 186, + 787, + 266 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a1c2ce9d065d8a641fe4c2fd848ec33a7e75c7f7d5acb6288a43099f8966b4ec.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 217, + 277, + 388, + 388 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1005250dba3b5903da76d2eac5cd6fc772c1d3b666c52a9b7da3b9d8b82f694a.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 388, + 277, + 519, + 388 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/780ce63fcdf2988697ebadc33d3f13b30a0b062e1487ac6ef4b0a99cc5d4f005.jpg", + "image_caption": [ + "AIR-v0", + "Fig. 9: Qualitative comparisons on NAFNet [10], AIR finetuned i) with latent diffusion weight (AIR-v0) and ii) with weight pre-trained on seven tasks (AIR-v1) on unseen de-snowing and JPEG task." + ], + "image_footnote": [], + "bbox": [ + 519, + 277, + 651, + 388 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4c7b0811ea0605ae7c16f1290a10d7896326b150ff543cedee5049256553de10.jpg", + "image_caption": [ + "AIR-v1" + ], + "image_footnote": [], + "bbox": [ + 651, + 277, + 784, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 483, + 400, + 497 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The first ablation study is to validate the importance of the Semantic-Agnostic BIQA and Structural Correction Module (SCM), as shown in Tab. 5. The presence of proper guidance from the SA-BIQA is crucial for enabling accurate image restoration or enhancement in complex tasks. As shown in Tab. 5, incorporating the SA-BIQA into the AutoDIR pipeline consistently improves performance in multitasking scenarios. Similarly, the Structural-Correction latent diffusion can maintain fine image structures compared with original latent-diffusion-based image restoration. Tab. 5 and Fig. 4 also showed that incorporating the SCM consistently achieves performance gains.", + "bbox": [ + 212, + 507, + 784, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our second ablation study is to investigate the possibility of using the all-in-one image restoration backbone (AIR) as a foundation model for general image restoration, which is based on the hypothesis that joint learning of multiple image restoration tasks may result in basis operators that are useful for other image restoration tasks. Inspired by prior work on vision foundation models [52,61,81], we conduct experiments on the unseen de-snowing and JPEG tasks to evaluate this hypothesis. For our de-snowing experiments, we utilize 90 snowy-clean image pairs of the Comprehensive Snow Dataset (CSD) [11] to fine-tune the AIR which is initialized with the pre-trained AIR weights obtained from training on the seven tasks. For the JPEG compression task, we use 40 compressed-clean image pairs of LIVE dataset [51]. Fig. 5 demonstrates the fine-tuned AIR with the initialization of pre-trained weight adapts to the de-snowing task more quickly compared to the module without the pre-trained weights. This shows the new", + "bbox": [ + 212, + 643, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "AutoDIR", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8296901db263fd09d912baa2ff39cb65dad325b4612d6ee7f18a4274d8b99eec.jpg", + "table_caption": [ + "Table 5: Ablation study on BIQA and SCM on deraindrop and dehazing tasks." + ], + "table_footnote": [], + "table_body": "
BIQASCMDeraindropDehaze
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
XX0.73824.840.1040.76323.490.091
X✓0.90328.840.0680.92527.530.038
✓X0.89625.580.0920.80424.470.074
✓✓0.92430.100.0630.97329.340.020
", + "bbox": [ + 339, + 170, + 663, + 272 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "image restoration task can benefit from the knowledge acquired from other image restoration tasks, indicating the implicit correlations between them. Moreover, it also outperforms the NAFNet specifically trained for the de-snowing task, which further shows the feasibility of constructing a foundational image restoration model by exploring the underlying correlations.", + "bbox": [ + 212, + 299, + 787, + 377 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions and Discussions", + "text_level": 1, + "bbox": [ + 214, + 398, + 527, + 416 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present AutoDIR, an all-in-one image restoration system with latent diffusion, which is capable of automatically restoring images with multiple unknown degradations and also supports open-vocabulary image editing user instructions. By jointly training from multiple individual image restoration datasets, AutoDIR can be viewed as a foundation model for image restoration, which can automatically detect and address multiple image degradations (with SA-BIQA) and enables effective few-shot learning for novel tasks.", + "bbox": [ + 212, + 430, + 787, + 536 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To our knowledge, AutoDIR is the most general comprehensive framework for automatic image restoration in the context of complex real-scenario images with multiple unknown degradations. We hope that this framework can provide a solid foundation for further works on complex real-scenario images and has the potential to be extended to various practical applications, such as local image enhancement and multi-task video restoration.", + "bbox": [ + 212, + 537, + 787, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "There are several limitations of AutoDIR that we plan to address in the future. First, the computational cost is similar to Latent Diffusion [49] and other diffusion-based image restoration methods [60], which is often significantly higher than non-generative networks such as NAFNet [10]. Some recent work on accelerating diffusion networks such as one-step diffusion [35] provides possible solutions. Second, AutoDIR currently focuses more on global image restoration, such as deblurring, denoising, and low-light enhancement, rather than local region-based image editing, such as flare removal and inpainting. Extending AutoDIR for local image editing by incorporating semantic segmentation may be an interesting direction to explore.", + "bbox": [ + 212, + 627, + 787, + 779 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 282, + 112, + 738, + 130 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abu-Hussein, S., Tirer, T., Giryes, R.: DIR: Adaptive diffusion for image reconstruction. arXiv preprint arXiv:2212.03221 (2022)", + "2. Bau, D., Strobelt, H., Peebles, W., Wulff, J., Zhou, B., Zhu, J.Y., Torralba, A.: Semantic photo manipulation with a generative image prior. ACM Transactions on Graphics (TOG) 38(4), 59 (2019)", + "3. Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "4. Brooks, T., Holynski, A., Efros, A.A.: InstructPix2Pix: Learning to follow image editing instructions. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "5. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "6. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG) 42(4), 1-10 (2023)", + "7. Chen, C., Shi, X., Qin, Y., Li, X., Han, X., Yang, T., Guo, S.: Real-world blind super-resolution via feature matching with implicit high-resolution priors. In: IEEE Transactions on Multimedia (TMM) (2022)", + "8. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "9. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2024)", + "0. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of European Conferences on Computer Vision (ECCV) (2022)", + "1. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2021)", + "2. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multicontrostatic regularization: Toward a unified model. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "3. Cheon, M., Yoon, S.J., Kang, B., Lee, J.: Perceptual image quality assessment with transformers. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of International Conference on Learning Representations (ICLR) (2020)", + "5. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2014)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "AutoDIR", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code GAN prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "18. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-Prompt image editing with cross attention control. Proceedings of International Conference on Learning Representations (ICLR) (2023)", + "19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "20. Huang, Z., Chan, K.C., Jiang, Y., Liu, Z.: Collaborative diffusion for multi-modal face generation and editing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "21. Hussein, S.A., Tirer, T., Giryes, R.: Image-adaptive GAN based reconstruction. In: Proceedings of the AAAI Conference on Artificial Intelligence (2020)", + "22. Islam, M.J., Xia, Y., Sattar, J.: Fast underwater image enhancement for improved visual perception. IEEE Robotics and Automation Letters (RA-L) (2020)", + "23. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: XYDeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "24. Jiang, K., Wang, Z., Yi, P., Chen, C., Huang, B., Luo, Y., Ma, J., Jiang, J.: Multiscale progressive fusion network for single image deraining. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "25. Jinjin, G., Haoming, C., Haoyu, C., Xiaoxing, Y., Ren, J.S., Chao, D.: Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In: Proceedings of European Conferences on Computer Vision (ECCV) (2020)", + "26. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "27. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "28. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: MUSIQ: Multi-scale image quality transformer. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV)", + "29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2014)", + "30. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing (TIP) (2018)", + "31. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-In-One image restoration for unknown corruption. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "32. Li, R., Tan, R.T., Cheong, L.F.: All-in-One bad weather removal using architectural search. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "33. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "34. Liu, K., Jiang, Y., Choi, I., Gu, J.: Learning image-adaptive codebooks for class-agnostic image restoration. Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 284, + 112, + 738, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Liu, X., Zhang, X., Ma, J., Peng, J., Liu, Q.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380 (2023)", + "36. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: PULSE: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "37. Mingde, Y., Jie, H., Xin, J., Ruikang, X., Shenglong, Z., Man, Z., Zhiwei, X.: Generalized lightness adaptation with channel selective normalization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)", + "38. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters (3), 209-212 (2012)", + "39. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models (2021)", + "40. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 7474-7489 (2021)", + "41. Park, D., Lee, B.H., Chun, S.Y.: All-in-One image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "42. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "43. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "44. Qin, X., Wang, Z., Bai, Y., Xie, X., Jia, H.: FFA-Net: Feature fusion attention network for single image dehazing. In: Proceedings of the AAAI Conference on Artificial Intelligence (2020)", + "45. Quan, Y., Chen, M., Pang, T., Ji, H.: Self2Self with dropout: Learning self-supervised denoising from single image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning (ICML) (2021)", + "47. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "48. Ren, D., Zuo, W., Hu, Q., Zhu, P., Meng, D.: Progressive image deraining networks: A better and simpler baseline. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "50. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "AutoDIR", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Sheikh, H.R., Sabir, M.F., Bovik, A.C.: A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing (TIP) (2006)", + "52. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "53. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE Transactions on Image Processing (TIP) (2023)", + "54. Talebi, H., Milanfar, P.: Nima: Neural image assessment. IEEE Transactions on Image Processing (TIP) 27(8), 3998-4011 (2018)", + "55. Tao, M., Tang, H., Wu, F., Jing, X.Y., Bao, B.K., Xu, C.: DF-GAN: A simple and effective baseline for text-to-image synthesis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "56. Tsai, M.F., Liu, T.Y., Qin, T., Chen, H.H., Ma, W.Y.: FRank: a ranking method with fidelity loss. In: ACM SIGGRAPH Conference Proceedings (2007)", + "57. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "58. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2017)", + "59. Wang, J., Chan, K.C., Loy, C.C.: Exploring CLIP for assessing the look and feel of images. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)", + "60. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. In: arXiv preprint arXiv:2305.07015 (2023)", + "61. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: Proceedings of International Conference on Machine Learning (ICML) (2022)", + "62. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. In: Proceedings of International Conference on Learning Representations (ICLR) (2023)", + "63. Whang, J., Delbracio, M., Talebi, H., Sahara, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "64. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "65. Yang, T., Ren, P., Xie, X., Zhang, L.: GAN prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "66. Yao, X., She, D., Zhao, S., Liang, J., Lai, Y.K., Yang, J.: Attention-aware polarity sensitive embedding for affective image retrieval. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "67. Ye, H., Yang, X., Takac, M., Sunderraman, R., Ji, S.: Improving text-to-image synthesis using contrastive learning. Proceedings of The British Machine Vision Conference (BMVC) (2021)", + "68. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1", + "bbox": [ + 284, + 112, + 738, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "69. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "70. Zhang, H., Koh, J.Y., Baldridge, J., Lee, H., Yang, Y.: Cross-modal contrastive learning for text-to-image generation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "71. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "72. Zhang, L., Chen, X., Wang, Y., Lu, Y., Qiao, Y.: Brush your text: Synthesize any scene text on images via diffusion model. arXiv preprint arXiv:2312.12232 (2023)", + "73. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "74. Zhang, W., Ma, K., Yan, J., Deng, D., Wang, Z.: Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology 30(1), 36-47 (2020)", + "75. Zhang, Z., Jiang, Y., Jiang, J., Wang, X., Luo, P., Gu, J.: STAR: A structure-aware lightweight transformer for real-time image enhancement. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2021)", + "76. Zhang, Z., Jiang, Y., Shao, W., Wang, X., Luo, P., Lin, K., Gu, J.: Real-time controllable denoising for image and video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "77. Zhang, Z., Han, L., Ghosh, A., Metaxas, D.N., Ren, J.: SINE: Single image editing with text-to-image diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "78. Zhou, S., Chan, K., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "79. Zhou, Y., Ren, D., Emerton, N., Lim, S., Large, T.: Image restoration for underdisplay camera. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "80. Zhu, M., Pan, P., Chen, W., Yang, Y.: DM-GAN: Dynamic memory generative adversarial networks for text-to-image synthesis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "81. Zhu, X., Zhu, J., Li, H., Wu, X., Li, H., Wang, X., Dai, J.: Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 686 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "AutoDIR", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_model.json b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a6e09e2a61d1ef8c61d581f83257f3c0a85bebb1 --- /dev/null +++ b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_model.json @@ -0,0 +1,2944 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.279, + 0.141, + 0.728, + 0.185 + ], + "angle": 0, + "content": "AutoDIR: Automatic All-in-One Image Restoration with Latent Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.212, + 0.747, + 0.229 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.24, + 0.758, + 0.27 + ], + "angle": 0, + "content": "1The Chinese University of Hong Kong 2Shanghai AI Laboratory {ytjiang@link, zhaoyangzhang@link, tfxue@ie, jwgu@cse}.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.308, + 0.744, + 0.53 + ], + "angle": 0, + "content": "Abstract. We present AutoDIR, an innovative all-in-one image restoration system incorporating latent diffusion. AutoDIR excels in its ability to automatically identify and restore images suffering from a range of unknown degradations. AutoDIR offers intuitive open-vocabulary image editing, empowering users to customize and enhance images according to their preferences. AutoDIR consists of two key stages: a Blind Image Quality Assessment (BIQA) stage based on a semantic-agnostic vision-language model which automatically detects unknown image degradations for input images, an All-in-One Image Restoration (AIR) stage utilizes structural-corrected latent diffusion which handles multiple types of image degradations. Extensive experimental evaluation demonstrates that AutoDIR outperforms state-of-the-art approaches for a wider range of image restoration tasks. The design of AutoDIR also enables flexible user control (via text prompt) and generalization to new tasks as a foundation model of image restoration. Project is available at: https://jiangyitong.github.io/AutoDIR_webpage/." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.557, + 0.377, + 0.573 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.788, + 0.785 + ], + "angle": 0, + "content": "Restoring scene details and improving image quality is often the very first step of any computer vision system, which significantly determines the overall system performance. Depending on the operating environments and the signal processing pipelines, images captured by real-world computer vision systems often undergo multiple unknown degradations, such as noise, resolution loss, motion blur, defocus, color imbalance, chromatic aberration, haze, flare, distortion, etc. Previous image restoration methods often address these basic image degradations separately, including deblurring [23,63], denoising [45,76] deraining [24,48], super-resolution [7,60], low-light enhancement [37,75], deraindrop [43], and dehaze [44,53], by using specific single-task models. While these single-task approaches yield promising results within their respective tasks, they encounter difficulties when applied to complex real-world situations that involve multiple unknown degradations or require multiple steps of enhancement." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.786, + 0.788, + 0.817 + ], + "angle": 0, + "content": "In this paper, we explore a more general model capable of handling multiple unknown degradations for a single image. To achieve this goal, the corresponding" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.378, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.114, + 0.739, + 0.129 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.589, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.145, + 0.785, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Fig. 1: We propose AutoDIR, an automatic all-in-one model for image restoration capable of handling multiple types of image degradations. Left: For images with multiple unknown degradations, AutoDIR automatically decomposes the task into multiple steps and supports user interaction via an intuitive text prompt. Right: AutoDIR effectively restores clean images from different degradations and can handle images with unknown degradations in unseen tasks. (Please zoom in for details)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.521 + ], + "angle": 0, + "content": "model should possess the following abilities: (1) decomposing and distinguishing unknown degradations, (2) restoring various degradations in a task-agnostic framework, and (3) allowing users to freely tune restoration results according to their visual preferences. Numerous related efforts have been made to solve similar problems, but none of them can simultaneously satisfy all three points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Recent works [13, 66, 74] have explored assessing unknown degradations by training an image classifier and using this information to explicitly select a specific image restoration model [8, 33, 69]. However, these methods require training a specific model for each task and primarily focus on a small set of tasks (e.g., 3 tasks). Experiments demonstrate that on large datasets with a wide range of degradations, training an accurate image degradation classifier solely based on image information is still challenging (See Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Other works attempt to unify multiple image restoration methods [26, 62] in a single model by utilizing generative model priors. These methods, however, either require known degradation formulations or can only handle images with a single degradation from a predefined small set [31, 42], thus unable to implicitly distinguish multiple unknown degradations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.78 + ], + "angle": 0, + "content": "In this work, we propose a pipeline named AutoDIR, which satisfies all three aforementioned abilities and can automatically detect and restore images with multiple unknown degradations. AutoDIR consists of two stages: the Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA) stage and the All-in-One Image Restoration (AIR) stage guided by text prompts generated in SA-BIQA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the quality assessment stage, we enable accurate identification of each degradation in cases of unknown artifacts in an open-vocabulary manner. This is achieved through our proposed Semantic-Agnostic CLIP (SA-CLIP) model, which employs a Semantic-Agnostic regularization term to transfer the origi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AutoDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "nal Semantic-Recognitive CLIP to a semantic-agnostic form, focusing on image structural quality rather than semantic content. Additionally, we can utilize the text embeddings generated by the SA-BIQA stage as instructions to guide the further restoration model. This approach can not only enable effective restoration but also allows for flexible user control and editing at runtime by providing open-vocabulary instructions, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.344 + ], + "angle": 0, + "content": "The restoration stage is to handle degradations using a multitasking image restoration model jointly trained on a wide range of tasks. Given the diverse nature of different tasks (e.g., some tasks like super-resolution need to hallucinate textures, but other tasks like low-light enhancement need to preserve everything except brightness), we propose a hybrid approach that maximizes the hallucinating ability of diffusion models while preserving the image structural consistency by involving extra structural inductive bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.787, + 0.465 + ], + "angle": 0, + "content": "To evaluate the effectiveness and generalization ability of AutoDIR, we have conducted a comprehensive set of experiments encompassing seven image restoration tasks, including denoising, motion deblurring, low-light enhancement, dehazing, deraining, deraindrop, and super-resolution. Experimental results show that AutoDIR consistently outperforms state-of-the-art methods. AutoDIR is also evaluated for the restoration of images captured by under-display cameras and underwater cameras, which are examples of imaging systems with multiple unknown degradations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.388, + 0.506 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Unified All-in-One Image Restoration Previous approaches for unified all-in-one image restoration can be categorized into two main groups: unsupervised generative prior-based methods [2,5,15,21,26,36,40,62,65] and end-to-end supervised learning-based methods [12,31,32,41,57,71]. Generative priors for image restoration have been widely exploited for image restoration, in the form of generative adversarial networks (GANs) inversion and filters [1,2,5,17,21,36,40], Vector Quantization codebooks [7,34,58,78], latent diffusion models [1,15,26,62]. Nevertheless, these methods rely on specific known image degradation functions, which constrain the types of degradations they can handle. End-to-end supervised learning-based methods typically learn image embeddings extracted by an auxiliary degradation predictor to guide the image restoration model [31,41,57,71]. For example, Valanarasu et al. [57] leverage a transformer encoder to capture hierarchical features of haze, while Li et al. [31] employ a degradation classifier trained using contrastive learning. Park et al. [41] design a degradation classifier for multiple types of degradation to select appropriate Adaptive Discriminant filters, altering network parameters based on specific degradations. However, these methods are constrained to handle limited image degradation categories and cannot restore images with multiple unknown degradations. In contrast, with an SA-BIQA stage, the proposed AutoDIR is able to unify complex degradations without relying on assumptions about the specific degradation functions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.284, + 0.113, + 0.739, + 0.129 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.45 + ], + "angle": 0, + "content": "Text-based Image Synthesis and Editing Early works [55,64,67,70,80] on text-to-image synthesis and editing primarily relied on generative adversarial networks (GANs) [16]. Recently, developments in diffusion models [19, 49] and language-vision models [46] have led to significant progress in text-based image synthesis and editing [6,39,47,50]. The emergence of diffusion-based methods has offered new avenues for text-to-image editing, which can be broadly categorized into data-driven [4] and no-extra-data-required approaches [18,27,77]. In the data-driven category, Brooks et al. [4] employ a stable diffusion model fine-tuned with a large dataset of text prompt and image pairs. However, it is primarily designed for semantic image editing and struggles to deliver satisfactory results in tasks related to image quality enhancement. On the other hand, some methods focus on image editing without extra training data. For instance, recent works interpolate the text embedding [27] or score estimate [77] of the input and desired images. However, it requires individual finetuning of the diffusion model for each input image, leading to time-consuming operations. Hertz et al. [18] improve editing efficiency by directly manipulating cross-attention maps without the need for per-image finetuning. However, it requires interior maps in the reversing process and is thus not applicable to real image editing. Unlike the above methods, AutoDIR supports real image enhancement without per-image finetuning." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.504, + 0.789, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.748, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Fig. 2: Diagram of the proposed All-in-One Image Restoration with Latent Diffusion (AutoDIR). Refer to Sec. 3 for more details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AutoDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.377, + 0.161 + ], + "angle": 0, + "content": "3 Our Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Fig. 2 shows the overall flowchart of the proposed AutoDIR (Automatic All-in-One Image Restoration with Latent Diffusion), a unified model that can automatically detect and address multiple unknown degradations in images. AutoDIR comprises two main stages:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.248, + 0.786, + 0.308 + ], + "angle": 0, + "content": "- Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA): This stage automatically identifies the predominant degradations present in the input image, such as noise, blur, haze, and generates a corresponding text prompt, denoted as \\( e_{auto} \\), which is subsequently used in the image restoration process." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.309, + 0.797, + 0.369 + ], + "angle": 0, + "content": "- All-in-one Image Restoration (AIR): This stage utilizes a Structural-Correction Latent Diffusion Model(SC-LDM) to produce a restored image \\( I_{res} \\), guided by the text embedding \\( e_{auto} \\) from SA-BIQA or a user-defined open-vocabulary instruction \\( e_{user} \\)." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.248, + 0.797, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.391, + 0.786, + 0.408 + ], + "angle": 0, + "content": "3.1 Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.786, + 0.507 + ], + "angle": 0, + "content": "As demonstrated by previous works [13, 66, 74], a common approach to assess image degradation is to train an image classifier specifically for that purpose. However, this naive approach faces challenges when dealing with large datasets containing a wide range of artifacts. Even with the use of a heavy Vision Transformer(ViT) encoder, the accuracy of classifying image degradation solely based on image information is limited to \\(77.65\\%\\), as shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.703 + ], + "angle": 0, + "content": "To address this limitation, we propose leveraging human language knowledge to enhance the detection of image degradation. We introduce a Semantic-Agnostic CLIP (SA-CLIP) model as our backbone for Blind Image Quality Assessment (BIQA). SA-CLIP builds upon the CLIP model, which establishes a connection between human language knowledge and image quality. However, we observed that directly applying CLIP or naively fine-tuning it for BIQA tasks does not yield reliable results, as we will later show in Tab. 1. We delved into this problem and identified the reason behind this issue. The pre-trained CLIP model is primarily trained for visual recognition tasks, which prioritize semantic information rather than image quality. Consequently, it delivers low accuracy in BIQA tasks. For instance, the model may struggle to differentiate between a low-light dog image and a noisy dog image because it focuses more on the \"dog\" aspect rather than the presence of noise or illumination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.763 + ], + "angle": 0, + "content": "To overcome this, we tackle the problem in two steps: (i) We construct a new image quality assessment task for fine-tuning CLIP. (ii) We propose a new regularization term for semantic-agnostic and image-quality-aware training to derive SA-CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, let \\(\\mathcal{C}\\) denotes the set of image distortions, i.e. \\(\\mathcal{C} = \\{c_1, c_2, \\dots, c_{K-1}, c_K\\}\\), where \\(c_i\\) is an image distortion type, e.g. \"noise\" and \\(K-1\\) is the number of distortions we consider. We also add a special type \\(c_K = \"\\mathrm{no}\"\\) to indicate the end of the automatic multiple-step restoration. The textural description prompt set is \\(\\mathcal{T} = \\{T | T = \"\\text{A photo needs } \\{c_i\\} \\text{ artifact reduction.}\"\\}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.284, + 0.113, + 0.739, + 0.129 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.373, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.235, + 0.328, + 0.245 + ], + "angle": 0, + "content": "a) Original CLIP" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.145, + 0.526, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.235, + 0.493, + 0.245 + ], + "angle": 0, + "content": "b) Finetuned CLIP" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.145, + 0.787, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.235, + 0.771, + 0.245 + ], + "angle": 0, + "content": "c) Finetuned CLIP with Semantic-Agnostic constraint" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.256, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 3: t-SNE visualization of image embeddings \\(\\mathcal{E}_{\\mathcal{I}}(I)\\) of CLIP of Blind Image Quality Assessment (BIQA) on SOTs dehazing dataset [30]. Image embeddings of foggy images and their ground-truth clean images are extracted by a) original CLIP. b) finetuned CLIP. c) SA-CLIP finetuned with semantic-agnostic constraint. This illustrates that semantic-agnostic constraint can separate the embeddings of the degraded images from the clean images, while original CLIP and finetuned CLIP features cannot." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.37, + 0.789, + 0.462 + ], + "angle": 0, + "content": "\\(c \\in \\mathcal{C}\\). Given a corrupted image \\(I\\) which undergoes several unknown artifacts, our Semantic-Agnostic CLIP aims to identify the dominant degradation of \\(I\\) and extract the corresponding text embedding. SA-CLIP consists of an image encoder \\(\\mathcal{E}_{\\mathcal{I}}\\) and a text encoder \\(\\mathcal{E}_{\\mathcal{T}}\\). We first obtain the image embedding \\(\\mathcal{E}_{\\mathcal{I}}(I) \\in \\mathbb{R}^d\\) and the text embedding \\(\\mathcal{E}_{\\mathcal{T}}(T) \\in \\mathbb{R}^{K \\times d}\\), then we compute the cosine similarity between image embedding \\(\\mathcal{E}_{\\mathcal{I}}(I)\\) and each prompt embedding \\(\\mathcal{E}_{\\mathcal{T}}(T_i) \\in \\mathbb{R}^d\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.472, + 0.787, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o g i t} \\left(c _ {i} \\mid I\\right) = \\frac {\\mathcal {E} _ {\\mathcal {I}} (I) \\cdot \\mathcal {E} _ {\\mathcal {T}} \\left(T _ {i}\\right)}{\\| \\mathcal {E} _ {\\mathcal {I}} (I) \\| _ {2} \\| \\mathcal {E} _ {\\mathcal {T}} \\left(T _ {i}\\right) \\| _ {2}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.516, + 0.788, + 0.547 + ], + "angle": 0, + "content": "where \\( T_{i} \\) is the \\( i \\)-th element of \\( \\mathcal{T} \\). Next, we calculate the probability \\( \\hat{p}(c_i|I) \\) and obtain the output text embedding of BIQA \\( e_{auto} \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.557, + 0.787, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\hat {p} \\left(c _ {i} \\mid I\\right) = \\frac {\\exp \\left(\\operatorname {l o g i t} \\left(c _ {i} \\mid I\\right)\\right)}{\\sum_ {i = 1} ^ {K} \\exp \\left(\\operatorname {l o g i t} \\left(c _ {i} \\mid I\\right)\\right)}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.605, + 0.787, + 0.647 + ], + "angle": 0, + "content": "\\[\ne _ {a u t o} = \\sum_ {i = 1} ^ {K} \\hat {p} (c _ {i} | I) \\mathcal {E} _ {\\mathcal {T}} (T _ {i}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.66, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Naive fine-tuning on image quality assessment During the optimization of CLIP model, we freeze the parameters of the text encoder \\(\\mathcal{E}_{\\mathcal{T}}\\) and finetune the image encoder \\(\\mathcal{E}_{\\mathcal{I}}\\) using the multi-class fidelity loss [56]. The fidelity loss can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.717, + 0.787, + 0.758 + ], + "angle": 0, + "content": "\\[\nL _ {F I D} = 1 - \\sum_ {i = 1} ^ {K} \\sqrt {y \\left(c _ {i} | I\\right) \\hat {p} \\left(c _ {i} | I\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.764, + 0.788, + 0.794 + ], + "angle": 0, + "content": "where \\( y(c_{i}|I) \\) is a binary variable that equals 1 if \\( c_{i} \\) is the dominant degradation in image \\( I \\), and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Semantic-Agnostic Constraint fine-tuning on image quality assessment Since the original CLIP model is pre-trained on tasks such as image" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AutoDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "classification, its corresponding \\(\\mathcal{E}_{\\mathcal{I}}(I)\\) encoders tend to encode images based on their semantic information (e.g., cat or dog) rather than their image quality (e.g., noisy or clean). When we fine-tune the CLIP model for generating texts for BIQA according to image quality, this becomes a significant limitation. As shown in Fig. 3 a) and b), the image embeddings extracted by the original CLIP and the finetuned CLIP on foggy, as well as their corresponding ground-truth clean images, cannot be separated, indicating a focus on semantic information rather than image quality differences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.785, + 0.417 + ], + "angle": 0, + "content": "To address this issue, we propose a novel approach called the semantic-agnostic constraint loss \\((L_{SA})\\) to regulate the fine-tuning process and prevent the model from relying solely on semantic information rather than image quality. The semantic-agnostic loss \\((L_{SA})\\) applies a penalty when the CLIP model suggests that the artifact \\(c_{i}\\) is present in the ground-truth clean image \\(I_{gt}\\) which corresponds to the degraded image \\(I\\). This penalty forces the CLIP model to distinguish between \\(I_{gt}\\) and \\(I\\) based on image quality, encouraging the CLIP image encoder \\((\\mathcal{E}_{\\mathcal{I}}(I))\\) to focus on extracting image quality information rather than semantic information. This constraint loss can be derived using the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.42, + 0.785, + 0.459 + ], + "angle": 0, + "content": "\\[\nL _ {S A} = \\sum_ {i = 1} ^ {K} \\sqrt {y \\left(c _ {i} | I\\right) \\hat {p} \\left(c _ {i} \\mid I _ {g t}\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.784, + 0.512 + ], + "angle": 0, + "content": "where \\( y(c_{i}|I) \\) is a binary variable that is equal to 1 if \\( c_{i} \\) represents the dominant degradation in image \\( I \\), and 0 otherwise, and \\( \\hat{p}(c_{i}|I_{gt}) \\) denotes the probability that the dominant degradation in \\( I_{gt} \\) is represented by \\( c_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.784, + 0.543 + ], + "angle": 0, + "content": "We incorporate the semantic-agnostic constraint \\(L_{SA}\\) with the fidelity loss \\(L_{F1D}\\), resulting in the total loss \\(L_{BIQA}\\) for SA-CLIP model:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.557, + 0.785, + 0.572 + ], + "angle": 0, + "content": "\\[\nL _ {B I Q A} = L _ {F I D} + \\lambda L _ {S A}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.784, + 0.629 + ], + "angle": 0, + "content": "where we set \\(\\lambda = 1\\) in our experiments. As shown in Fig. 3 c), CLIP fine-tuned with a semantic-agnostic constraint can separate the embeddings of the degraded images from the clean images as expected." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.656, + 0.784, + 0.684 + ], + "angle": 0, + "content": "Table 1: Average accuracy of image degradation detection on seven degradation tasks and clean image. For details, please refer to the supplementary." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.685, + 0.744, + 0.732 + ], + "angle": 0, + "content": "
MethodsViT-classifierOriginal CLIPFine-tuned CLIPSA-CLIP (ours)
Average Accuracy77.65%61.96%83.31%97.94%
" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.783, + 0.599, + 0.799 + ], + "angle": 0, + "content": "3.2 Effectiveness of Semantic-Agnostic CLIP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We compare our SA-CLIP with a ViT-based [14] classifier, original CLIP, and naive fine-tuned CLIP to demonstrate the effectiveness of the Semantic-Agnostic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.114, + 0.739, + 0.129 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.209, + 0.785, + 0.238 + ], + "angle": 0, + "content": "Fig. 4: Structural-Correction Latent Diffusion Model (SC-LDM) can maintain complex structures of the original images, while Latent Diffusion Model (LDM) fails to." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.42 + ], + "angle": 0, + "content": "training scheme. Tab. 1 presents the effectiveness of our Semantic-Agnostic training scheme. The results demonstrate that our SA-CLIP, which incorporates human language information, achieves higher accuracy compared to the ViT-based classifier which considers only image information. Moreover, when CLIP is finetuned with Semantic-Agnostic regularization, the model focuses more on image quality information rather than semantic information, resulting in superior performance compared to the original pre-trained CLIP and CLIP fine-tuned without Semantic-Agnostic regularization. For a visual demonstration of the effectiveness of the Semantic-Agnostic constraint, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.447, + 0.557, + 0.464 + ], + "angle": 0, + "content": "3.3 All-in-one Image Restoration (AIR)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Naively applying LDM on multi-task image restoration The All-in-One Image Restoration (AIR) stage aims to handle multiple degradations in a single shared framework. Recent advancements in diffusion-based generative models [4, 49] have demonstrated their remarkable ability to generate diverse images, making them suitable for multi-task image restoration. Previous studies [7, 60] have shown that generative models possess a remarkable ability to generate missing or distorted details, especially for tasks that require hallucination, such as super-resolution. Based on these insights, we conduct the AIR stage based on the Latent Diffusion Model (LDM) [49]. LDM incorporates both text and image embedding conditions to restore images \\( I_{sd} \\) using generative priors. The text embedding condition \\( e = \\{e_{auto}, e_{user}\\} \\) aims to separate different types of image degradations, while the latent image embedding condition \\( z_I = \\mathcal{E}_{ldm}(I) \\) from the image encoder \\( \\mathcal{E}_{ldm} \\) of the LDM provides structural information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "However, although LDM-based generative models can provide a foundation for multi-task image restoration, they have limitations in reconstructing images with complex and small structures due to the compression-reconstruction process with Variational Autoencoder (VAE) [29], as discussed in [9, 20, 72]. [20] attempted to reduce distortion caused by the compression-reconstruction process by retraining Variational Autoencoder (VAE) networks on category-specific images (e.g., human faces) to learn specialized probability distributions. However, this approach is not suitable for image restoration tasks due to the diverse image content. To address these limitations, we introduce a lightweight plug-in structural-correction module to the LDM, enhancing its ability to handle complex and small structures during image restoration." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AutoDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.351, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.207, + 0.298, + 0.218 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.145, + 0.484, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.207, + 0.453, + 0.217 + ], + "angle": 0, + "content": "...haze ..." + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.145, + 0.619, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.207, + 0.607, + 0.218 + ], + "angle": 0, + "content": "...low-resolution..." + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.145, + 0.753, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.207, + 0.73, + 0.218 + ], + "angle": 0, + "content": "... raindrop ..." + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.147, + 0.787, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.23, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Fig. 5: Cross-attention maps of text-conditioned diffusion image restoration. The top-left shows an input image with raindrops on the left half. The remaining plots show the cross-attention masks for the keywords \"haze\", \"low-resolution\", and \"raindrop\" in the text prompt used for restoration. While the cross-attention maps for \"haze\" and \"low-resolution\" are more uniformly distributed over the entire image, the map for \"raindrop\" correctly focuses on the actual image artifacts, as expected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.785, + 0.482 + ], + "angle": 0, + "content": "Structural-Correction Latent Diffusion Model (SC-LDM) While the LDM-based generative models can provide a foundation for multi-tasking image restoration, it is widely noticed that they may fail to maintain original image structures such as human faces and texts [9, 20, 72], as shown in Fig. 4. To address the structure distortion, we employ an efficient structural-correction module (SCM) denoted as \\(\\mathcal{F}\\). The purpose of SCM is to extract the contextual information \\(\\mathcal{R}\\) from the original image and incorporate it with the intermediate image restoration result \\(I_{sd}\\), decoded by image decoder \\(\\mathcal{D}\\) of LDM, in a residual manner. This is achieved through the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.496, + 0.785, + 0.513 + ], + "angle": 0, + "content": "\\[\nI _ {r e s} = I _ {s d} + w \\cdot \\mathcal {F} ([ I _ {s d}, I ]), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.784, + 0.662 + ], + "angle": 0, + "content": "where \\([\\cdot, \\cdot]\\) denotes concatenation, and \\(w\\) is an adjustable coefficient that ranges between 0 and 1. The value of \\(w\\) determines the extent to which contextual information is utilized to recover the final result. A larger value of \\(w\\) emphasizes the use of contextual information, which is beneficial for tasks that require structure consistency, such as low light enhancement. Conversely, a smaller value of \\(w\\) is often employed to maintain the generation capability of latent diffusion model for tasks like super-resolution. By integrating SCM, AutoDIR effectively restores the distorted context of the original image, seamlessly incorporating the enhancements made during the editing phase as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.784, + 0.707 + ], + "angle": 0, + "content": "During training, we fine-tune the UNet backbone \\(\\epsilon_{\\theta}(e,[z_t,z_I],t)\\) of the Latent Diffusion Model (LDM) for image restoration tasks. The objective function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.721, + 0.785, + 0.74 + ], + "angle": 0, + "content": "\\[\nL _ {L D} = \\mathbb {E} _ {\\mathcal {E} _ {l d m} (x), z _ {I}, e, \\epsilon \\sim \\mathcal {N} (0, 1), t} [ \\| \\epsilon - \\epsilon_ {\\theta} (e, [ z _ {t}, z _ {I} ], t) \\| _ {2} ^ {2} ], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.752, + 0.784, + 0.798 + ], + "angle": 0, + "content": "where \\( z_{t} \\) is the noisy latent. For the SC-LDM, instead of generating the edited latent \\( \\hat{z}_{t} \\) with the complete reversing sampling process, which is time-consuming, we utilize the estimated edited latent \\( \\tilde{z} \\), which is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.81, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} = \\frac {z _ {t}}{\\sqrt {\\bar {\\alpha}}} - \\frac {\\sqrt {1 - \\bar {\\alpha}} \\epsilon_ {\\theta} (e , [ z _ {t} , z _ {I} ] , t)}{\\sqrt {\\bar {\\alpha}}}, \\tag {9}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.114, + 0.739, + 0.129 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Fig. 6: AutoDIR deals with multiple unknown image degradations. Iterative image restoration on unseen Under-Display Camera (TOLED) dataset [79] and unseen Enhancing Underwater Visual Perception (EUVP) dataset [22]. Left: AutoDIR decomposes and solves the unseen artifact in several steps. Right: all-in-one AirNet [31], PromptIR [42] and zero-shot method GDP [15] for comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.785, + 0.391 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}\\) represents the noise scheduler introduced in [49]. The loss function for Structural-Correction Latent Diffusion Model (SC-LDM) is further defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.4, + 0.785, + 0.417 + ], + "angle": 0, + "content": "\\[\nL _ {A I R} = \\left\\| I _ {g t} - \\left(\\mathcal {F} (\\mathcal {D} (\\tilde {z}), I) + \\mathcal {D} (\\tilde {z})\\right) \\right\\| _ {2} ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Mechanism of handling multi-task image restoration Fig. 5 shows our experiment of exploring the mechanism of the text conditions disentangling different image restoration tasks during the reverse diffusion process. We find that different text conditions yield different cross-attention maps. As shown in Fig. 5, changing the text prompt leads to significant changes in the cross-attention map. The map closely aligns with the text prompt, which either uniformly distributes attention across the entire image for the \"dehazing\" prompt, or concentrates on the part with edges or textures for the \"low-resolution\" prompt, or focuses on specific areas such as raindrops for the \"deraindrop\" prompt. This shows that AutoDIR can guide the diffusion attention toward regions that are more likely to have image artifacts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.376, + 0.634 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.768 + ], + "angle": 0, + "content": "To fully validate the feasibility and effectiveness of unifying multiple restoration and enhancement tasks, we compare AutoDIR with both CNN-based [10,31,42], and diffusion-based multitasking [49] backbones on two tasks with unknown image degradations as well as seven representative image restoration tasks. Specifically, we highlight the advantages of our generative model-based pipeline over discriminative models for tasks that require hallucinating details such as superresolution. Please refer to the appendix for implementation details and datasets. The source code and trained models will be released upon publication." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.787, + 0.627, + 0.803 + ], + "angle": 0, + "content": "4.1 Results on Multiple Unknown Degradations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To show that AutoDIR can handle multiple unknown image degradations, we conducted experiments on two additional datasets: the Under-Display-Camera" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AutoDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.357, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.221, + 0.295, + 0.226 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.145, + 0.495, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.221, + 0.436, + 0.226 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.145, + 0.641, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.221, + 0.582, + 0.226 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.145, + 0.785, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.221, + 0.73, + 0.226 + ], + "angle": 0, + "content": "PromptR" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.227, + 0.357, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.3, + 0.322, + 0.306 + ], + "angle": 0, + "content": "All-in-One NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.227, + 0.495, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.3, + 0.459, + 0.307 + ], + "angle": 0, + "content": "Super-res NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.227, + 0.641, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.3, + 0.581, + 0.306 + ], + "angle": 0, + "content": "LD" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.227, + 0.785, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.3, + 0.721, + 0.306 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.319, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Fig.7: Super-resolution qualitative comparisons with AirNet [31], PromptIR [42], NAFNet [10] trained on all-in-one tasks, NAFNet [10] trained on Super-Res, and Latent Diffusion Model (LDM) [49]. See the Appendix for more results (Zoom in for better view)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.633 + ], + "angle": 0, + "content": "(TOLED) dataset [79] and the Enhancing Underwater Visual Perception (EUVP) dataset [22]. These two datasets consist of multiple unknown image degradations and are completely excluded for training, allowing us to evaluate the performance of AutoDIR in handling unknown, diverse image distortions. Fig. 6 demonstrates the iterative process of AutoDIR to handle images from these two datasets. As shown, AutoDIR can automatically decompose the UDC restoration in two steps (\"brightening\" and \"deblurring\"), and the Underwater restoration in two steps (\"deblurring\" and \"dehazing\"), enabling the restoration of input images. For comparison, we also show the results of a recent all-in-one AirNet [31], PromptIR [42], diffusion-based zero-shot image restoration method GDP [15]. Furthermore, as shown in Tab. 3, AutoDIR outperforms all the other black-box and classifier-based all-in-one methods on quantitative perceptual results. We also conducted a user study on 28 images of TOLED and real-captured images by OPPO Reno6 device. We collected 22 forms and 616 responses. And AutoDIR receives \\(96.9\\%\\) of the votes as the best result, please refer to the appendix for details." + }, + { + "type": "table_caption", + "bbox": [ + 0.262, + 0.646, + 0.54, + 0.659 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on SR" + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.66, + 0.564, + 0.759 + ], + "angle": 0, + "content": "
MethodSuper Resolution
MUSIQ ↑CLIP-IQA ↑NIQE ↓NIMA ↑LPIPS ↓PSNR ↑SSIM ↑
NAFNet40.770.2928.454.450.27828.520.754
NAFNet-SR45.490.2998.234.560.33227.950.726
AirNet20.210.27611.074.030.42528.140.626
PromptIR22.890.30110.044.160.30929.620.792
LD50.710.3756.444.900.37824.110.649
Ours63.020.5964.625.250.28223.490.599
" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.646, + 0.792, + 0.671 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison on Zero-shot UDC task." + }, + { + "type": "table", + "bbox": [ + 0.58, + 0.673, + 0.774, + 0.761 + ], + "angle": 0, + "content": "
MethodUnder-displayCamera
MUSIQ ↑CLIP-IQA ↑NIQE ↓NIMA ↑
NAFNet37.160.2787.4024.258
AirNet41.200.2657.6814.048
PromptIR45.400.3038.2944.344
LD43.410.2887.9834.371
Ours49.270.3116.3924.427
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.782, + 0.6, + 0.798 + ], + "angle": 0, + "content": "4.2 Results on Multi-task Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We compare with state-of-the-art methods: the CNN-based single-task image restoration pipeline NAFNet [10], CNN-based all-in-one AirNet [31], PromptIR" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.284, + 0.113, + 0.739, + 0.129 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.787, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.536 + ], + "angle": 0, + "content": "Fig. 8: Qualitative comparisons with NAFNet [10] and Latent Diffusion Model (LDM) [49], AirNet [31] and PromptIR [42] on deraining, dehazing, low light enhancement, and deraindrop, denoise and deblur tasks. See the Appendix for more results. (Zoom in for better view)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.844 + ], + "angle": 0, + "content": "[42], and the generative backbone Latent Diffusion [49]. They are all fine-tuned for all seven tasks, employing an identical training configuration to AutoDIR. Tab. 2 and Tab. 4 present quantitative results of AutoDIR and other methods in all seven tasks. As demonstrated, AutoDIR achieves competitive performance across all seven tasks. Specifically, to further evaluate the ability of discriminative and generative methods on super-resolution tasks, we include task-specific discriminative NAFNet training on SR. As shown in Tab. 2 and Fig. 7, AutoDIR quantitatively and qualitatively outperforms all the all-in-one discriminative, task-specific discriminative methods and the black-box latent diffusion model. Please be noted that, as discussed in previous works [3,25,68], full-reference metrics like PSNR, SSIM, and LPIPS [73] often do not align with human evaluation on super-resolution tasks. Therefore, following [60,68], we include non-reference metrics MUSIQ [28], CLIP-IQA [59] NIQE [38] and NIMA [54] to better evaluate the results. We also show the qualitative comparison results on the other six image restoration tasks in Fig. 8, which demonstrate that AutoDIR can achieve more visually appealing results than the other baselines. A comprehensive and detailed analysis of the quantitative and qualitative results for each specific task can be found in the appendix." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AutoDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.185 + ], + "angle": 0, + "content": "Table 4: Quantitative comparison on denoising, deraining, dehazing, deraindrop, and low light enhancement tasks. The best results are marked in boldface and the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.187, + 0.789, + 0.267 + ], + "angle": 0, + "content": "
MethodDenoiseDerainDehazeDeraindropLow lightDeblur
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
NAFNet0.80528.100.1950.92630.460.1000.96827.750.0280.87225.040.1380.82721.150.2400.80526.670.208
LD0.62522.580.1860.65123.210.1470.76323.490.0910.73824.840.1040.77018.970.1590.69522.530.241
AirNet0.80329.100.2130.92930.990.0550.94426.520.0310.89227.130.0940.81821.260.2370.80026.500.201
PromptIR0.82429.890.1810.93833.970.0490.97129.130.0220.90027.410.0810.83122.420.1330.81926.820.198
Ours0.83229.680.1670.96535.090.0470.97329.340.0200.92430.100.0630.88822.370.1290.82827.070.157
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.279, + 0.389, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.392, + 0.337, + 0.402 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.279, + 0.521, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.392, + 0.478, + 0.402 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.279, + 0.653, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.392, + 0.608, + 0.402 + ], + "angle": 0, + "content": "AIR-v0" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.279, + 0.785, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.392, + 0.735, + 0.402 + ], + "angle": 0, + "content": "AIR-v1" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.413, + 0.785, + 0.455 + ], + "angle": 0, + "content": "Fig. 9: Qualitative comparisons on NAFNet [10], AIR finetuned i) with latent diffusion weight (AIR-v0) and ii) with weight pre-trained on seven tasks (AIR-v1) on unseen de-snowing and JPEG task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.401, + 0.498 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.644 + ], + "angle": 0, + "content": "The first ablation study is to validate the importance of the Semantic-Agnostic BIQA and Structural Correction Module (SCM), as shown in Tab. 5. The presence of proper guidance from the SA-BIQA is crucial for enabling accurate image restoration or enhancement in complex tasks. As shown in Tab. 5, incorporating the SA-BIQA into the AutoDIR pipeline consistently improves performance in multitasking scenarios. Similarly, the Structural-Correction latent diffusion can maintain fine image structures compared with original latent-diffusion-based image restoration. Tab. 5 and Fig. 4 also showed that incorporating the SCM consistently achieves performance gains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our second ablation study is to investigate the possibility of using the all-in-one image restoration backbone (AIR) as a foundation model for general image restoration, which is based on the hypothesis that joint learning of multiple image restoration tasks may result in basis operators that are useful for other image restoration tasks. Inspired by prior work on vision foundation models [52,61,81], we conduct experiments on the unseen de-snowing and JPEG tasks to evaluate this hypothesis. For our de-snowing experiments, we utilize 90 snowy-clean image pairs of the Comprehensive Snow Dataset (CSD) [11] to fine-tune the AIR which is initialized with the pre-trained AIR weights obtained from training on the seven tasks. For the JPEG compression task, we use 40 compressed-clean image pairs of LIVE dataset [51]. Fig. 5 demonstrates the fine-tuned AIR with the initialization of pre-trained weight adapts to the de-snowing task more quickly compared to the module without the pre-trained weights. This shows the new" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.284, + 0.113, + 0.739, + 0.131 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "table_caption", + "bbox": [ + 0.234, + 0.144, + 0.768, + 0.16 + ], + "angle": 0, + "content": "Table 5: Ablation study on BIQA and SCM on deraindrop and dehazing tasks." + }, + { + "type": "table", + "bbox": [ + 0.341, + 0.171, + 0.664, + 0.273 + ], + "angle": 0, + "content": "
BIQASCMDeraindropDehaze
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
XX0.73824.840.1040.76323.490.091
X✓0.90328.840.0680.92527.530.038
✓X0.89625.580.0920.80424.470.074
✓✓0.92430.100.0630.97329.340.020
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.378 + ], + "angle": 0, + "content": "image restoration task can benefit from the knowledge acquired from other image restoration tasks, indicating the implicit correlations between them. Moreover, it also outperforms the NAFNet specifically trained for the de-snowing task, which further shows the feasibility of constructing a foundational image restoration model by exploring the underlying correlations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.399, + 0.528, + 0.417 + ], + "angle": 0, + "content": "5 Conclusions and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.537 + ], + "angle": 0, + "content": "In this paper, we present AutoDIR, an all-in-one image restoration system with latent diffusion, which is capable of automatically restoring images with multiple unknown degradations and also supports open-vocabulary image editing user instructions. By jointly training from multiple individual image restoration datasets, AutoDIR can be viewed as a foundation model for image restoration, which can automatically detect and address multiple image degradations (with SA-BIQA) and enables effective few-shot learning for novel tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.627 + ], + "angle": 0, + "content": "To our knowledge, AutoDIR is the most general comprehensive framework for automatic image restoration in the context of complex real-scenario images with multiple unknown degradations. We hope that this framework can provide a solid foundation for further works on complex real-scenario images and has the potential to be extended to various practical applications, such as local image enhancement and multi-task video restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.78 + ], + "angle": 0, + "content": "There are several limitations of AutoDIR that we plan to address in the future. First, the computational cost is similar to Latent Diffusion [49] and other diffusion-based image restoration methods [60], which is often significantly higher than non-generative networks such as NAFNet [10]. Some recent work on accelerating diffusion networks such as one-step diffusion [35] provides possible solutions. Second, AutoDIR currently focuses more on global image restoration, such as deblurring, denoising, and low-light enhancement, rather than local region-based image editing, such as flare removal and inpainting. Extending AutoDIR for local image editing by incorporating semantic segmentation may be an interesting direction to explore." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AutoDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Abu-Hussein, S., Tirer, T., Giryes, R.: DIR: Adaptive diffusion for image reconstruction. arXiv preprint arXiv:2212.03221 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.2, + 0.785, + 0.242 + ], + "angle": 0, + "content": "2. Bau, D., Strobelt, H., Peebles, W., Wulff, J., Zhou, B., Zhu, J.Y., Torralba, A.: Semantic photo manipulation with a generative image prior. ACM Transactions on Graphics (TOG) 38(4), 59 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.268 + ], + "angle": 0, + "content": "3. Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.785, + 0.309 + ], + "angle": 0, + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: InstructPix2Pix: Learning to follow image editing instructions. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.309, + 0.785, + 0.35 + ], + "angle": 0, + "content": "5. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.785, + 0.391 + ], + "angle": 0, + "content": "6. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG) 42(4), 1-10 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.432 + ], + "angle": 0, + "content": "7. Chen, C., Shi, X., Qin, Y., Li, X., Han, X., Yang, T., Guo, S.: Real-world blind super-resolution via feature matching with implicit high-resolution priors. In: IEEE Transactions on Multimedia (TMM) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.785, + 0.472 + ], + "angle": 0, + "content": "8. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.472, + 0.785, + 0.513 + ], + "angle": 0, + "content": "9. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.513, + 0.785, + 0.54 + ], + "angle": 0, + "content": "0. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of European Conferences on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.54, + 0.785, + 0.595 + ], + "angle": 0, + "content": "1. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.65 + ], + "angle": 0, + "content": "2. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multicontrostatic regularization: Toward a unified model. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "3. Cheon, M., Yoon, S.J., Kang, B., Lee, J.: Perceptual image quality assessment with transformers. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.745 + ], + "angle": 0, + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of International Conference on Learning Representations (ICLR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "5. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2014)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.113, + 0.739, + 0.129 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code GAN prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "18. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-Prompt image editing with cross attention control. Proceedings of International Conference on Learning Representations (ICLR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "20. Huang, Z., Chan, K.C., Jiang, Y., Liu, Z.: Collaborative diffusion for multi-modal face generation and editing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "21. Hussein, S.A., Tirer, T., Giryes, R.: Image-adaptive GAN based reconstruction. In: Proceedings of the AAAI Conference on Artificial Intelligence (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "22. Islam, M.J., Xia, Y., Sattar, J.: Fast underwater image enhancement for improved visual perception. IEEE Robotics and Automation Letters (RA-L) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.398 + ], + "angle": 0, + "content": "23. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: XYDeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.439 + ], + "angle": 0, + "content": "24. Jiang, K., Wang, Z., Yi, P., Chen, C., Huang, B., Luo, Y., Ma, J., Jiang, J.: Multiscale progressive fusion network for single image deraining. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "25. Jinjin, G., Haoming, C., Haoyu, C., Xiaoxing, Y., Ren, J.S., Chao, D.: Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In: Proceedings of European Conferences on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "26. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.577 + ], + "angle": 0, + "content": "27. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.619 + ], + "angle": 0, + "content": "28. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: MUSIQ: Multi-scale image quality transformer. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.691, + 0.633 + ], + "angle": 0, + "content": "29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "30. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing (TIP) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "31. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-In-One image restoration for unknown corruption. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.785, + 0.758 + ], + "angle": 0, + "content": "32. Li, R., Tan, R.T., Cheong, L.F.: All-in-One bad weather removal using architectural search. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "34. Liu, K., Jiang, Y., Choi, I., Gu, J.: Learning image-adaptive codebooks for class-agnostic image restoration. Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AutoDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Liu, X., Zhang, X., Ma, J., Peng, J., Liu, Q.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "36. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: PULSE: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "37. Mingde, Y., Jie, H., Xin, J., Ruikang, X., Shenglong, Z., Man, Z., Zhiwei, X.: Generalized lightness adaptation with channel selective normalization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "38. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters (3), 209-212 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "39. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.787, + 0.384 + ], + "angle": 0, + "content": "40. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 7474-7489 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.787, + 0.438 + ], + "angle": 0, + "content": "41. Park, D., Lee, B.H., Chun, S.Y.: All-in-One image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "42. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "43. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "44. Qin, X., Wang, Z., Bai, Y., Xie, X., Jia, H.: FFA-Net: Feature fusion attention network for single image dehazing. In: Proceedings of the AAAI Conference on Artificial Intelligence (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "45. Quan, Y., Chen, M., Pang, T., Ji, H.: Self2Self with dropout: Learning self-supervised denoising from single image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.787, + 0.66 + ], + "angle": 0, + "content": "46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning (ICML) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.702 + ], + "angle": 0, + "content": "47. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.744 + ], + "angle": 0, + "content": "48. Ren, D., Zuo, W., Hu, Q., Zhu, P., Meng, D.: Progressive image deraining networks: A better and simpler baseline. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "50. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.113, + 0.739, + 0.129 + ], + "angle": 0, + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "51. Sheikh, H.R., Sabir, M.F., Bovik, A.C.: A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing (TIP) (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.243 + ], + "angle": 0, + "content": "52. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.271 + ], + "angle": 0, + "content": "53. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE Transactions on Image Processing (TIP) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.298 + ], + "angle": 0, + "content": "54. Talebi, H., Milanfar, P.: Nima: Neural image assessment. IEEE Transactions on Image Processing (TIP) 27(8), 3998-4011 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "55. Tao, M., Tang, H., Wu, F., Jing, X.Y., Bao, B.K., Xu, C.: DF-GAN: A simple and effective baseline for text-to-image synthesis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "56. Tsai, M.F., Liu, T.Y., Qin, T., Chen, H.H., Ma, W.Y.: FRank: a ranking method with fidelity loss. In: ACM SIGGRAPH Conference Proceedings (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.406 + ], + "angle": 0, + "content": "57. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "58. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.459 + ], + "angle": 0, + "content": "59. Wang, J., Chan, K.C., Loy, C.C.: Exploring CLIP for assessing the look and feel of images. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.486 + ], + "angle": 0, + "content": "60. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. In: arXiv preprint arXiv:2305.07015 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.486, + 0.785, + 0.541 + ], + "angle": 0, + "content": "61. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: Proceedings of International Conference on Machine Learning (ICML) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.582 + ], + "angle": 0, + "content": "62. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. In: Proceedings of International Conference on Learning Representations (ICLR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.622 + ], + "angle": 0, + "content": "63. Whang, J., Delbracio, M., Talebi, H., Sahara, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.785, + 0.677 + ], + "angle": 0, + "content": "64. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "65. Yang, T., Ren, P., Xie, X., Zhang, L.: GAN prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "66. Yao, X., She, D., Zhao, S., Liang, J., Lai, Y.K., Yang, J.: Attention-aware polarity sensitive embedding for affective image retrieval. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "67. Ye, H., Yang, X., Takac, M., Sunderraman, R., Ji, S.: Improving text-to-image synthesis using contrastive learning. Proceedings of The British Machine Vision Conference (BMVC) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "68. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AutoDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "69. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "70. Zhang, H., Koh, J.Y., Baldridge, J., Lee, H., Yang, Y.: Cross-modal contrastive learning for text-to-image generation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "71. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.3 + ], + "angle": 0, + "content": "72. Zhang, L., Chen, X., Wang, Y., Lu, Y., Qiao, Y.: Brush your text: Synthesize any scene text on images via diffusion model. arXiv preprint arXiv:2312.12232 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "73. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.384 + ], + "angle": 0, + "content": "74. Zhang, W., Ma, K., Yan, J., Deng, D., Wang, Z.: Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology 30(1), 36-47 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.424 + ], + "angle": 0, + "content": "75. Zhang, Z., Jiang, Y., Jiang, J., Wang, X., Luo, P., Gu, J.: STAR: A structure-aware lightweight transformer for real-time image enhancement. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.467 + ], + "angle": 0, + "content": "76. Zhang, Z., Jiang, Y., Shao, W., Wang, X., Luo, P., Lin, K., Gu, J.: Real-time controllable denoising for image and video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.508 + ], + "angle": 0, + "content": "77. Zhang, Z., Han, L., Ghosh, A., Metaxas, D.N., Ren, J.: SINE: Single image editing with text-to-image diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.55 + ], + "angle": 0, + "content": "78. Zhou, S., Chan, K., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "79. Zhou, Y., Ren, D., Emerton, N., Lim, S., Large, T.: Image restoration for underdisplay camera. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.633 + ], + "angle": 0, + "content": "80. Zhu, M., Pan, P., Chen, W., Yang, Y.: DM-GAN: Dynamic memory generative adversarial networks for text-to-image synthesis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.688 + ], + "angle": 0, + "content": "81. Zhu, X., Zhu, J., Li, H., Wu, X., Li, H., Wang, X., Dai, J.: Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_origin.pdf b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..056ff60c7ff4e3ec8710ac52207c1a0eebdcfc9e --- /dev/null +++ b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/04591189-82ea-4136-9449-5ecbc2d97a0a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f701dd3183082de989bb3a582dc2ed5bcb1eb83cac3e4e728909718184d4e81e +size 6223502 diff --git a/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/full.md b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bdde46a05d0700086febecf379b31da168b675c3 --- /dev/null +++ b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/full.md @@ -0,0 +1,356 @@ +# AutoDIR: Automatic All-in-One Image Restoration with Latent Diffusion + +Yitong Jiang\*1,2 Zhaoyang Zhang\*1 Tianfan Xue1 Jinwei Gu1 + +1The Chinese University of Hong Kong 2Shanghai AI Laboratory {ytjiang@link, zhaoyangzhang@link, tfxue@ie, jwgu@cse}.cuhk.edu.hk + +Abstract. We present AutoDIR, an innovative all-in-one image restoration system incorporating latent diffusion. AutoDIR excels in its ability to automatically identify and restore images suffering from a range of unknown degradations. AutoDIR offers intuitive open-vocabulary image editing, empowering users to customize and enhance images according to their preferences. AutoDIR consists of two key stages: a Blind Image Quality Assessment (BIQA) stage based on a semantic-agnostic vision-language model which automatically detects unknown image degradations for input images, an All-in-One Image Restoration (AIR) stage utilizes structural-corrected latent diffusion which handles multiple types of image degradations. Extensive experimental evaluation demonstrates that AutoDIR outperforms state-of-the-art approaches for a wider range of image restoration tasks. The design of AutoDIR also enables flexible user control (via text prompt) and generalization to new tasks as a foundation model of image restoration. Project is available at: https://jiangyitong.github.io/AutoDIR_webpage/. + +# 1 Introduction + +Restoring scene details and improving image quality is often the very first step of any computer vision system, which significantly determines the overall system performance. Depending on the operating environments and the signal processing pipelines, images captured by real-world computer vision systems often undergo multiple unknown degradations, such as noise, resolution loss, motion blur, defocus, color imbalance, chromatic aberration, haze, flare, distortion, etc. Previous image restoration methods often address these basic image degradations separately, including deblurring [23,63], denoising [45,76] deraining [24,48], super-resolution [7,60], low-light enhancement [37,75], deraindrop [43], and dehaze [44,53], by using specific single-task models. While these single-task approaches yield promising results within their respective tasks, they encounter difficulties when applied to complex real-world situations that involve multiple unknown degradations or require multiple steps of enhancement. + +In this paper, we explore a more general model capable of handling multiple unknown degradations for a single image. To achieve this goal, the corresponding + +![](images/2a5c11c6de9c5bac36bcab16b7d0996d6005eccd2bf5ec845dd9b541c5bc33c4.jpg) +Fig. 1: We propose AutoDIR, an automatic all-in-one model for image restoration capable of handling multiple types of image degradations. Left: For images with multiple unknown degradations, AutoDIR automatically decomposes the task into multiple steps and supports user interaction via an intuitive text prompt. Right: AutoDIR effectively restores clean images from different degradations and can handle images with unknown degradations in unseen tasks. (Please zoom in for details) + +![](images/5c50eff13c1b3380bde7eb038e1e144a7dd77521512b2423832b8217b75165eb.jpg) + +model should possess the following abilities: (1) decomposing and distinguishing unknown degradations, (2) restoring various degradations in a task-agnostic framework, and (3) allowing users to freely tune restoration results according to their visual preferences. Numerous related efforts have been made to solve similar problems, but none of them can simultaneously satisfy all three points. + +Recent works [13, 66, 74] have explored assessing unknown degradations by training an image classifier and using this information to explicitly select a specific image restoration model [8, 33, 69]. However, these methods require training a specific model for each task and primarily focus on a small set of tasks (e.g., 3 tasks). Experiments demonstrate that on large datasets with a wide range of degradations, training an accurate image degradation classifier solely based on image information is still challenging (See Tab. 1). + +Other works attempt to unify multiple image restoration methods [26, 62] in a single model by utilizing generative model priors. These methods, however, either require known degradation formulations or can only handle images with a single degradation from a predefined small set [31, 42], thus unable to implicitly distinguish multiple unknown degradations. + +In this work, we propose a pipeline named AutoDIR, which satisfies all three aforementioned abilities and can automatically detect and restore images with multiple unknown degradations. AutoDIR consists of two stages: the Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA) stage and the All-in-One Image Restoration (AIR) stage guided by text prompts generated in SA-BIQA. + +In the quality assessment stage, we enable accurate identification of each degradation in cases of unknown artifacts in an open-vocabulary manner. This is achieved through our proposed Semantic-Agnostic CLIP (SA-CLIP) model, which employs a Semantic-Agnostic regularization term to transfer the origi + +nal Semantic-Recognitive CLIP to a semantic-agnostic form, focusing on image structural quality rather than semantic content. Additionally, we can utilize the text embeddings generated by the SA-BIQA stage as instructions to guide the further restoration model. This approach can not only enable effective restoration but also allows for flexible user control and editing at runtime by providing open-vocabulary instructions, as shown in Fig. 1. + +The restoration stage is to handle degradations using a multitasking image restoration model jointly trained on a wide range of tasks. Given the diverse nature of different tasks (e.g., some tasks like super-resolution need to hallucinate textures, but other tasks like low-light enhancement need to preserve everything except brightness), we propose a hybrid approach that maximizes the hallucinating ability of diffusion models while preserving the image structural consistency by involving extra structural inductive bias. + +To evaluate the effectiveness and generalization ability of AutoDIR, we have conducted a comprehensive set of experiments encompassing seven image restoration tasks, including denoising, motion deblurring, low-light enhancement, dehazing, deraining, deraindrop, and super-resolution. Experimental results show that AutoDIR consistently outperforms state-of-the-art methods. AutoDIR is also evaluated for the restoration of images captured by under-display cameras and underwater cameras, which are examples of imaging systems with multiple unknown degradations. + +# 2 Related Work + +Unified All-in-One Image Restoration Previous approaches for unified all-in-one image restoration can be categorized into two main groups: unsupervised generative prior-based methods [2,5,15,21,26,36,40,62,65] and end-to-end supervised learning-based methods [12,31,32,41,57,71]. Generative priors for image restoration have been widely exploited for image restoration, in the form of generative adversarial networks (GANs) inversion and filters [1,2,5,17,21,36,40], Vector Quantization codebooks [7,34,58,78], latent diffusion models [1,15,26,62]. Nevertheless, these methods rely on specific known image degradation functions, which constrain the types of degradations they can handle. End-to-end supervised learning-based methods typically learn image embeddings extracted by an auxiliary degradation predictor to guide the image restoration model [31,41,57,71]. For example, Valanarasu et al. [57] leverage a transformer encoder to capture hierarchical features of haze, while Li et al. [31] employ a degradation classifier trained using contrastive learning. Park et al. [41] design a degradation classifier for multiple types of degradation to select appropriate Adaptive Discriminant filters, altering network parameters based on specific degradations. However, these methods are constrained to handle limited image degradation categories and cannot restore images with multiple unknown degradations. In contrast, with an SA-BIQA stage, the proposed AutoDIR is able to unify complex degradations without relying on assumptions about the specific degradation functions. + +Text-based Image Synthesis and Editing Early works [55,64,67,70,80] on text-to-image synthesis and editing primarily relied on generative adversarial networks (GANs) [16]. Recently, developments in diffusion models [19, 49] and language-vision models [46] have led to significant progress in text-based image synthesis and editing [6,39,47,50]. The emergence of diffusion-based methods has offered new avenues for text-to-image editing, which can be broadly categorized into data-driven [4] and no-extra-data-required approaches [18,27,77]. In the data-driven category, Brooks et al. [4] employ a stable diffusion model fine-tuned with a large dataset of text prompt and image pairs. However, it is primarily designed for semantic image editing and struggles to deliver satisfactory results in tasks related to image quality enhancement. On the other hand, some methods focus on image editing without extra training data. For instance, recent works interpolate the text embedding [27] or score estimate [77] of the input and desired images. However, it requires individual finetuning of the diffusion model for each input image, leading to time-consuming operations. Hertz et al. [18] improve editing efficiency by directly manipulating cross-attention maps without the need for per-image finetuning. However, it requires interior maps in the reversing process and is thus not applicable to real image editing. Unlike the above methods, AutoDIR supports real image enhancement without per-image finetuning. + +![](images/b3846f06a49da9827c19d7fae7d4e2f056c7d781fc8be5b354f86e20c6bdd906.jpg) +Fig. 2: Diagram of the proposed All-in-One Image Restoration with Latent Diffusion (AutoDIR). Refer to Sec. 3 for more details. + +# 3 Our Method + +Fig. 2 shows the overall flowchart of the proposed AutoDIR (Automatic All-in-One Image Restoration with Latent Diffusion), a unified model that can automatically detect and address multiple unknown degradations in images. AutoDIR comprises two main stages: + +- Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA): This stage automatically identifies the predominant degradations present in the input image, such as noise, blur, haze, and generates a corresponding text prompt, denoted as $e_{auto}$ , which is subsequently used in the image restoration process. +- All-in-one Image Restoration (AIR): This stage utilizes a Structural-Correction Latent Diffusion Model(SC-LDM) to produce a restored image $I_{res}$ , guided by the text embedding $e_{auto}$ from SA-BIQA or a user-defined open-vocabulary instruction $e_{user}$ . + +# 3.1 Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA) + +As demonstrated by previous works [13, 66, 74], a common approach to assess image degradation is to train an image classifier specifically for that purpose. However, this naive approach faces challenges when dealing with large datasets containing a wide range of artifacts. Even with the use of a heavy Vision Transformer(ViT) encoder, the accuracy of classifying image degradation solely based on image information is limited to $77.65\%$ , as shown in Tab. 1. + +To address this limitation, we propose leveraging human language knowledge to enhance the detection of image degradation. We introduce a Semantic-Agnostic CLIP (SA-CLIP) model as our backbone for Blind Image Quality Assessment (BIQA). SA-CLIP builds upon the CLIP model, which establishes a connection between human language knowledge and image quality. However, we observed that directly applying CLIP or naively fine-tuning it for BIQA tasks does not yield reliable results, as we will later show in Tab. 1. We delved into this problem and identified the reason behind this issue. The pre-trained CLIP model is primarily trained for visual recognition tasks, which prioritize semantic information rather than image quality. Consequently, it delivers low accuracy in BIQA tasks. For instance, the model may struggle to differentiate between a low-light dog image and a noisy dog image because it focuses more on the "dog" aspect rather than the presence of noise or illumination. + +To overcome this, we tackle the problem in two steps: (i) We construct a new image quality assessment task for fine-tuning CLIP. (ii) We propose a new regularization term for semantic-agnostic and image-quality-aware training to derive SA-CLIP model. + +As illustrated in Fig. 2, let $\mathcal{C}$ denotes the set of image distortions, i.e. $\mathcal{C} = \{c_1, c_2, \dots, c_{K-1}, c_K\}$ , where $c_i$ is an image distortion type, e.g. "noise" and $K-1$ is the number of distortions we consider. We also add a special type $c_K = "\mathrm{no}"$ to indicate the end of the automatic multiple-step restoration. The textural description prompt set is $\mathcal{T} = \{T | T = "\text{A photo needs } \{c_i\} \text{ artifact reduction.}"\}$ . + +![](images/273b652ad63ce3fd963634df577ac5892a5b22d5efd20da325af629726f6394f.jpg) +a) Original CLIP + +![](images/89f83f10074e8d1e8455214b5bb9b4ae0cd89108760557b78b2443736a59dadd.jpg) +b) Finetuned CLIP +Fig. 3: t-SNE visualization of image embeddings $\mathcal{E}_{\mathcal{I}}(I)$ of CLIP of Blind Image Quality Assessment (BIQA) on SOTs dehazing dataset [30]. Image embeddings of foggy images and their ground-truth clean images are extracted by a) original CLIP. b) finetuned CLIP. c) SA-CLIP finetuned with semantic-agnostic constraint. This illustrates that semantic-agnostic constraint can separate the embeddings of the degraded images from the clean images, while original CLIP and finetuned CLIP features cannot. + +![](images/3d2556025d7a9319662fe8fdbe3d24fcbac81eac7c485758ceaf9e311c7909f4.jpg) +c) Finetuned CLIP with Semantic-Agnostic constraint + +$c \in \mathcal{C}$ . Given a corrupted image $I$ which undergoes several unknown artifacts, our Semantic-Agnostic CLIP aims to identify the dominant degradation of $I$ and extract the corresponding text embedding. SA-CLIP consists of an image encoder $\mathcal{E}_{\mathcal{I}}$ and a text encoder $\mathcal{E}_{\mathcal{T}}$ . We first obtain the image embedding $\mathcal{E}_{\mathcal{I}}(I) \in \mathbb{R}^d$ and the text embedding $\mathcal{E}_{\mathcal{T}}(T) \in \mathbb{R}^{K \times d}$ , then we compute the cosine similarity between image embedding $\mathcal{E}_{\mathcal{I}}(I)$ and each prompt embedding $\mathcal{E}_{\mathcal{T}}(T_i) \in \mathbb{R}^d$ by: + +$$ +\operatorname {l o g i t} \left(c _ {i} \mid I\right) = \frac {\mathcal {E} _ {\mathcal {I}} (I) \cdot \mathcal {E} _ {\mathcal {T}} \left(T _ {i}\right)}{\| \mathcal {E} _ {\mathcal {I}} (I) \| _ {2} \| \mathcal {E} _ {\mathcal {T}} \left(T _ {i}\right) \| _ {2}}, \tag {1} +$$ + +where $T_{i}$ is the $i$ -th element of $\mathcal{T}$ . Next, we calculate the probability $\hat{p}(c_i|I)$ and obtain the output text embedding of BIQA $e_{auto}$ by: + +$$ +\hat {p} \left(c _ {i} \mid I\right) = \frac {\exp \left(\operatorname {l o g i t} \left(c _ {i} \mid I\right)\right)}{\sum_ {i = 1} ^ {K} \exp \left(\operatorname {l o g i t} \left(c _ {i} \mid I\right)\right)}, \tag {2} +$$ + +$$ +e _ {a u t o} = \sum_ {i = 1} ^ {K} \hat {p} (c _ {i} | I) \mathcal {E} _ {\mathcal {T}} (T _ {i}). \tag {3} +$$ + +Naive fine-tuning on image quality assessment During the optimization of CLIP model, we freeze the parameters of the text encoder $\mathcal{E}_{\mathcal{T}}$ and finetune the image encoder $\mathcal{E}_{\mathcal{I}}$ using the multi-class fidelity loss [56]. The fidelity loss can be denoted as: + +$$ +L _ {F I D} = 1 - \sum_ {i = 1} ^ {K} \sqrt {y \left(c _ {i} | I\right) \hat {p} \left(c _ {i} | I\right)}, \tag {4} +$$ + +where $y(c_{i}|I)$ is a binary variable that equals 1 if $c_{i}$ is the dominant degradation in image $I$ , and 0 otherwise. + +Semantic-Agnostic Constraint fine-tuning on image quality assessment Since the original CLIP model is pre-trained on tasks such as image + +classification, its corresponding $\mathcal{E}_{\mathcal{I}}(I)$ encoders tend to encode images based on their semantic information (e.g., cat or dog) rather than their image quality (e.g., noisy or clean). When we fine-tune the CLIP model for generating texts for BIQA according to image quality, this becomes a significant limitation. As shown in Fig. 3 a) and b), the image embeddings extracted by the original CLIP and the finetuned CLIP on foggy, as well as their corresponding ground-truth clean images, cannot be separated, indicating a focus on semantic information rather than image quality differences. + +To address this issue, we propose a novel approach called the semantic-agnostic constraint loss $(L_{SA})$ to regulate the fine-tuning process and prevent the model from relying solely on semantic information rather than image quality. The semantic-agnostic loss $(L_{SA})$ applies a penalty when the CLIP model suggests that the artifact $c_{i}$ is present in the ground-truth clean image $I_{gt}$ which corresponds to the degraded image $I$ . This penalty forces the CLIP model to distinguish between $I_{gt}$ and $I$ based on image quality, encouraging the CLIP image encoder $(\mathcal{E}_{\mathcal{I}}(I))$ to focus on extracting image quality information rather than semantic information. This constraint loss can be derived using the following equation: + +$$ +L _ {S A} = \sum_ {i = 1} ^ {K} \sqrt {y \left(c _ {i} | I\right) \hat {p} \left(c _ {i} \mid I _ {g t}\right)}, \tag {5} +$$ + +where $y(c_{i}|I)$ is a binary variable that is equal to 1 if $c_{i}$ represents the dominant degradation in image $I$ , and 0 otherwise, and $\hat{p}(c_{i}|I_{gt})$ denotes the probability that the dominant degradation in $I_{gt}$ is represented by $c_{i}$ . + +We incorporate the semantic-agnostic constraint $L_{SA}$ with the fidelity loss $L_{F1D}$ , resulting in the total loss $L_{BIQA}$ for SA-CLIP model: + +$$ +L _ {B I Q A} = L _ {F I D} + \lambda L _ {S A}, \tag {6} +$$ + +where we set $\lambda = 1$ in our experiments. As shown in Fig. 3 c), CLIP fine-tuned with a semantic-agnostic constraint can separate the embeddings of the degraded images from the clean images as expected. + +Table 1: Average accuracy of image degradation detection on seven degradation tasks and clean image. For details, please refer to the supplementary. + +
MethodsViT-classifierOriginal CLIPFine-tuned CLIPSA-CLIP (ours)
Average Accuracy77.65%61.96%83.31%97.94%
+ +# 3.2 Effectiveness of Semantic-Agnostic CLIP + +We compare our SA-CLIP with a ViT-based [14] classifier, original CLIP, and naive fine-tuned CLIP to demonstrate the effectiveness of the Semantic-Agnostic + +![](images/df162e78af1b30ba7ad3190ce0d9498d12736c65101fd09a82cfa52a67470a8b.jpg) +Fig. 4: Structural-Correction Latent Diffusion Model (SC-LDM) can maintain complex structures of the original images, while Latent Diffusion Model (LDM) fails to. + +training scheme. Tab. 1 presents the effectiveness of our Semantic-Agnostic training scheme. The results demonstrate that our SA-CLIP, which incorporates human language information, achieves higher accuracy compared to the ViT-based classifier which considers only image information. Moreover, when CLIP is finetuned with Semantic-Agnostic regularization, the model focuses more on image quality information rather than semantic information, resulting in superior performance compared to the original pre-trained CLIP and CLIP fine-tuned without Semantic-Agnostic regularization. For a visual demonstration of the effectiveness of the Semantic-Agnostic constraint, please refer to the supplementary materials. + +# 3.3 All-in-one Image Restoration (AIR) + +Naively applying LDM on multi-task image restoration The All-in-One Image Restoration (AIR) stage aims to handle multiple degradations in a single shared framework. Recent advancements in diffusion-based generative models [4, 49] have demonstrated their remarkable ability to generate diverse images, making them suitable for multi-task image restoration. Previous studies [7, 60] have shown that generative models possess a remarkable ability to generate missing or distorted details, especially for tasks that require hallucination, such as super-resolution. Based on these insights, we conduct the AIR stage based on the Latent Diffusion Model (LDM) [49]. LDM incorporates both text and image embedding conditions to restore images $I_{sd}$ using generative priors. The text embedding condition $e = \{e_{auto}, e_{user}\}$ aims to separate different types of image degradations, while the latent image embedding condition $z_I = \mathcal{E}_{ldm}(I)$ from the image encoder $\mathcal{E}_{ldm}$ of the LDM provides structural information. + +However, although LDM-based generative models can provide a foundation for multi-task image restoration, they have limitations in reconstructing images with complex and small structures due to the compression-reconstruction process with Variational Autoencoder (VAE) [29], as discussed in [9, 20, 72]. [20] attempted to reduce distortion caused by the compression-reconstruction process by retraining Variational Autoencoder (VAE) networks on category-specific images (e.g., human faces) to learn specialized probability distributions. However, this approach is not suitable for image restoration tasks due to the diverse image content. To address these limitations, we introduce a lightweight plug-in structural-correction module to the LDM, enhancing its ability to handle complex and small structures during image restoration. + +![](images/d55cabfa1fa34f83a49d6defe0d3007f40186590d6bea43c105304fa4b38f268.jpg) +Input + +![](images/3d59953ed90b4723c4590c13a66ddb9279141a63dc7f017db8e06e51f70fd7d3.jpg) +...haze ... + +![](images/5edc56324eeb178111d7e0db15c7ab0f413a74d0fee5711ee57468a9a22715c1.jpg) +...low-resolution... + +![](images/f0dca7fbfdb041b400b9e3a0fe6195eb62c1d17458ec314235d00cc496744247.jpg) +... raindrop ... + +![](images/7e9a69be4fcca3a1f9dd045795aca623f1e877140cf00771b48145e9a940729a.jpg) +Fig. 5: Cross-attention maps of text-conditioned diffusion image restoration. The top-left shows an input image with raindrops on the left half. The remaining plots show the cross-attention masks for the keywords "haze", "low-resolution", and "raindrop" in the text prompt used for restoration. While the cross-attention maps for "haze" and "low-resolution" are more uniformly distributed over the entire image, the map for "raindrop" correctly focuses on the actual image artifacts, as expected. + +Structural-Correction Latent Diffusion Model (SC-LDM) While the LDM-based generative models can provide a foundation for multi-tasking image restoration, it is widely noticed that they may fail to maintain original image structures such as human faces and texts [9, 20, 72], as shown in Fig. 4. To address the structure distortion, we employ an efficient structural-correction module (SCM) denoted as $\mathcal{F}$ . The purpose of SCM is to extract the contextual information $\mathcal{R}$ from the original image and incorporate it with the intermediate image restoration result $I_{sd}$ , decoded by image decoder $\mathcal{D}$ of LDM, in a residual manner. This is achieved through the following equation: + +$$ +I _ {r e s} = I _ {s d} + w \cdot \mathcal {F} ([ I _ {s d}, I ]), \tag {7} +$$ + +where $[\cdot, \cdot]$ denotes concatenation, and $w$ is an adjustable coefficient that ranges between 0 and 1. The value of $w$ determines the extent to which contextual information is utilized to recover the final result. A larger value of $w$ emphasizes the use of contextual information, which is beneficial for tasks that require structure consistency, such as low light enhancement. Conversely, a smaller value of $w$ is often employed to maintain the generation capability of latent diffusion model for tasks like super-resolution. By integrating SCM, AutoDIR effectively restores the distorted context of the original image, seamlessly incorporating the enhancements made during the editing phase as shown in Fig. 4. + +During training, we fine-tune the UNet backbone $\epsilon_{\theta}(e,[z_t,z_I],t)$ of the Latent Diffusion Model (LDM) for image restoration tasks. The objective function is defined as: + +$$ +L _ {L D} = \mathbb {E} _ {\mathcal {E} _ {l d m} (x), z _ {I}, e, \epsilon \sim \mathcal {N} (0, 1), t} [ \| \epsilon - \epsilon_ {\theta} (e, [ z _ {t}, z _ {I} ], t) \| _ {2} ^ {2} ], \tag {8} +$$ + +where $z_{t}$ is the noisy latent. For the SC-LDM, instead of generating the edited latent $\hat{z}_{t}$ with the complete reversing sampling process, which is time-consuming, we utilize the estimated edited latent $\tilde{z}$ , which is calculated by: + +$$ +\tilde {z} = \frac {z _ {t}}{\sqrt {\bar {\alpha}}} - \frac {\sqrt {1 - \bar {\alpha}} \epsilon_ {\theta} (e , [ z _ {t} , z _ {I} ] , t)}{\sqrt {\bar {\alpha}}}, \tag {9} +$$ + +![](images/0d5f5f0b10595e76b7a79643c4c9ca9777641d646e68c664668f55824ce3d80e.jpg) +Fig. 6: AutoDIR deals with multiple unknown image degradations. Iterative image restoration on unseen Under-Display Camera (TOLED) dataset [79] and unseen Enhancing Underwater Visual Perception (EUVP) dataset [22]. Left: AutoDIR decomposes and solves the unseen artifact in several steps. Right: all-in-one AirNet [31], PromptIR [42] and zero-shot method GDP [15] for comparison. + +where $\bar{\alpha}$ represents the noise scheduler introduced in [49]. The loss function for Structural-Correction Latent Diffusion Model (SC-LDM) is further defined as: + +$$ +L _ {A I R} = \left\| I _ {g t} - \left(\mathcal {F} (\mathcal {D} (\tilde {z}), I) + \mathcal {D} (\tilde {z})\right) \right\| _ {2} ^ {2}. \tag {10} +$$ + +Mechanism of handling multi-task image restoration Fig. 5 shows our experiment of exploring the mechanism of the text conditions disentangling different image restoration tasks during the reverse diffusion process. We find that different text conditions yield different cross-attention maps. As shown in Fig. 5, changing the text prompt leads to significant changes in the cross-attention map. The map closely aligns with the text prompt, which either uniformly distributes attention across the entire image for the "dehazing" prompt, or concentrates on the part with edges or textures for the "low-resolution" prompt, or focuses on specific areas such as raindrops for the "deraindrop" prompt. This shows that AutoDIR can guide the diffusion attention toward regions that are more likely to have image artifacts. + +# 4 Experiments + +To fully validate the feasibility and effectiveness of unifying multiple restoration and enhancement tasks, we compare AutoDIR with both CNN-based [10,31,42], and diffusion-based multitasking [49] backbones on two tasks with unknown image degradations as well as seven representative image restoration tasks. Specifically, we highlight the advantages of our generative model-based pipeline over discriminative models for tasks that require hallucinating details such as superresolution. Please refer to the appendix for implementation details and datasets. The source code and trained models will be released upon publication. + +# 4.1 Results on Multiple Unknown Degradations + +To show that AutoDIR can handle multiple unknown image degradations, we conducted experiments on two additional datasets: the Under-Display-Camera + +![](images/5e90bb7cf4472c124e6aed89c31b8587906bff4846463e8a452cbf374470456b.jpg) +GT + +![](images/b1dfb610abf6620351953d43dd405406be9d6ac6d96a1fe02e06ccd1727ab1c3.jpg) +Input + +![](images/f7e022fe3877da94d8a60c0d1d79e8dd7e0e93d4f8d0c415b1f0f61cf87da2fe.jpg) +AirNet + +![](images/3899a14490feef55417010a0d105b856de2dfbe7a140f5c6963caad7890d5d2e.jpg) +PromptR + +![](images/656da66404c1876f084626bf334520c6092714e45e5d6c88a64da6c57b9843a7.jpg) +All-in-One NAFNet +Fig.7: Super-resolution qualitative comparisons with AirNet [31], PromptIR [42], NAFNet [10] trained on all-in-one tasks, NAFNet [10] trained on Super-Res, and Latent Diffusion Model (LDM) [49]. See the Appendix for more results (Zoom in for better view). + +![](images/97f7a3f1b65fb880ebfc4d0b79fdb32c9a2e13d03b764944606fb804980b7566.jpg) +Super-res NAFNet + +![](images/7310c367a93b8cbb7156ed2d5446092d8f162bb12e48755dee57fe1dadcbbefb.jpg) +LD + +![](images/199ac676e73619f2c6c60a34fc9fd9b8c666957e539e0df4b620cc5f220f9301.jpg) +Ours + +(TOLED) dataset [79] and the Enhancing Underwater Visual Perception (EUVP) dataset [22]. These two datasets consist of multiple unknown image degradations and are completely excluded for training, allowing us to evaluate the performance of AutoDIR in handling unknown, diverse image distortions. Fig. 6 demonstrates the iterative process of AutoDIR to handle images from these two datasets. As shown, AutoDIR can automatically decompose the UDC restoration in two steps ("brightening" and "deblurring"), and the Underwater restoration in two steps ("deblurring" and "dehazing"), enabling the restoration of input images. For comparison, we also show the results of a recent all-in-one AirNet [31], PromptIR [42], diffusion-based zero-shot image restoration method GDP [15]. Furthermore, as shown in Tab. 3, AutoDIR outperforms all the other black-box and classifier-based all-in-one methods on quantitative perceptual results. We also conducted a user study on 28 images of TOLED and real-captured images by OPPO Reno6 device. We collected 22 forms and 616 responses. And AutoDIR receives $96.9\%$ of the votes as the best result, please refer to the appendix for details. + +Table 2: Quantitative comparison on SR + +
MethodSuper Resolution
MUSIQ ↑CLIP-IQA ↑NIQE ↓NIMA ↑LPIPS ↓PSNR ↑SSIM ↑
NAFNet40.770.2928.454.450.27828.520.754
NAFNet-SR45.490.2998.234.560.33227.950.726
AirNet20.210.27611.074.030.42528.140.626
PromptIR22.890.30110.044.160.30929.620.792
LD50.710.3756.444.900.37824.110.649
Ours63.020.5964.625.250.28223.490.599
+ +Table 3: Quantitative comparison on Zero-shot UDC task. + +
MethodUnder-displayCamera
MUSIQ ↑CLIP-IQA ↑NIQE ↓NIMA ↑
NAFNet37.160.2787.4024.258
AirNet41.200.2657.6814.048
PromptIR45.400.3038.2944.344
LD43.410.2887.9834.371
Ours49.270.3116.3924.427
+ +# 4.2 Results on Multi-task Image Restoration + +We compare with state-of-the-art methods: the CNN-based single-task image restoration pipeline NAFNet [10], CNN-based all-in-one AirNet [31], PromptIR + +![](images/9a5b900d554556b63b95f38df8dfbc259a08f7f0422314e38f9bbf3fa9fe50d1.jpg) +Fig. 8: Qualitative comparisons with NAFNet [10] and Latent Diffusion Model (LDM) [49], AirNet [31] and PromptIR [42] on deraining, dehazing, low light enhancement, and deraindrop, denoise and deblur tasks. See the Appendix for more results. (Zoom in for better view) + +[42], and the generative backbone Latent Diffusion [49]. They are all fine-tuned for all seven tasks, employing an identical training configuration to AutoDIR. Tab. 2 and Tab. 4 present quantitative results of AutoDIR and other methods in all seven tasks. As demonstrated, AutoDIR achieves competitive performance across all seven tasks. Specifically, to further evaluate the ability of discriminative and generative methods on super-resolution tasks, we include task-specific discriminative NAFNet training on SR. As shown in Tab. 2 and Fig. 7, AutoDIR quantitatively and qualitatively outperforms all the all-in-one discriminative, task-specific discriminative methods and the black-box latent diffusion model. Please be noted that, as discussed in previous works [3,25,68], full-reference metrics like PSNR, SSIM, and LPIPS [73] often do not align with human evaluation on super-resolution tasks. Therefore, following [60,68], we include non-reference metrics MUSIQ [28], CLIP-IQA [59] NIQE [38] and NIMA [54] to better evaluate the results. We also show the qualitative comparison results on the other six image restoration tasks in Fig. 8, which demonstrate that AutoDIR can achieve more visually appealing results than the other baselines. A comprehensive and detailed analysis of the quantitative and qualitative results for each specific task can be found in the appendix. + +Table 4: Quantitative comparison on denoising, deraining, dehazing, deraindrop, and low light enhancement tasks. The best results are marked in boldface and the second-best results are underlined. + +
MethodDenoiseDerainDehazeDeraindropLow lightDeblur
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
NAFNet0.80528.100.1950.92630.460.1000.96827.750.0280.87225.040.1380.82721.150.2400.80526.670.208
LD0.62522.580.1860.65123.210.1470.76323.490.0910.73824.840.1040.77018.970.1590.69522.530.241
AirNet0.80329.100.2130.92930.990.0550.94426.520.0310.89227.130.0940.81821.260.2370.80026.500.201
PromptIR0.82429.890.1810.93833.970.0490.97129.130.0220.90027.410.0810.83122.420.1330.81926.820.198
Ours0.83229.680.1670.96535.090.0470.97329.340.0200.92430.100.0630.88822.370.1290.82827.070.157
+ +![](images/a1c2ce9d065d8a641fe4c2fd848ec33a7e75c7f7d5acb6288a43099f8966b4ec.jpg) +Input + +![](images/1005250dba3b5903da76d2eac5cd6fc772c1d3b666c52a9b7da3b9d8b82f694a.jpg) +NAFNet + +![](images/780ce63fcdf2988697ebadc33d3f13b30a0b062e1487ac6ef4b0a99cc5d4f005.jpg) +AIR-v0 +Fig. 9: Qualitative comparisons on NAFNet [10], AIR finetuned i) with latent diffusion weight (AIR-v0) and ii) with weight pre-trained on seven tasks (AIR-v1) on unseen de-snowing and JPEG task. + +![](images/4c7b0811ea0605ae7c16f1290a10d7896326b150ff543cedee5049256553de10.jpg) +AIR-v1 + +# 4.3 Ablation Studies + +The first ablation study is to validate the importance of the Semantic-Agnostic BIQA and Structural Correction Module (SCM), as shown in Tab. 5. The presence of proper guidance from the SA-BIQA is crucial for enabling accurate image restoration or enhancement in complex tasks. As shown in Tab. 5, incorporating the SA-BIQA into the AutoDIR pipeline consistently improves performance in multitasking scenarios. Similarly, the Structural-Correction latent diffusion can maintain fine image structures compared with original latent-diffusion-based image restoration. Tab. 5 and Fig. 4 also showed that incorporating the SCM consistently achieves performance gains. + +Our second ablation study is to investigate the possibility of using the all-in-one image restoration backbone (AIR) as a foundation model for general image restoration, which is based on the hypothesis that joint learning of multiple image restoration tasks may result in basis operators that are useful for other image restoration tasks. Inspired by prior work on vision foundation models [52,61,81], we conduct experiments on the unseen de-snowing and JPEG tasks to evaluate this hypothesis. For our de-snowing experiments, we utilize 90 snowy-clean image pairs of the Comprehensive Snow Dataset (CSD) [11] to fine-tune the AIR which is initialized with the pre-trained AIR weights obtained from training on the seven tasks. For the JPEG compression task, we use 40 compressed-clean image pairs of LIVE dataset [51]. Fig. 5 demonstrates the fine-tuned AIR with the initialization of pre-trained weight adapts to the de-snowing task more quickly compared to the module without the pre-trained weights. This shows the new + +Table 5: Ablation study on BIQA and SCM on deraindrop and dehazing tasks. + +
BIQASCMDeraindropDehaze
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
XX0.73824.840.1040.76323.490.091
X✓0.90328.840.0680.92527.530.038
✓X0.89625.580.0920.80424.470.074
✓✓0.92430.100.0630.97329.340.020
+ +image restoration task can benefit from the knowledge acquired from other image restoration tasks, indicating the implicit correlations between them. Moreover, it also outperforms the NAFNet specifically trained for the de-snowing task, which further shows the feasibility of constructing a foundational image restoration model by exploring the underlying correlations. + +# 5 Conclusions and Discussions + +In this paper, we present AutoDIR, an all-in-one image restoration system with latent diffusion, which is capable of automatically restoring images with multiple unknown degradations and also supports open-vocabulary image editing user instructions. By jointly training from multiple individual image restoration datasets, AutoDIR can be viewed as a foundation model for image restoration, which can automatically detect and address multiple image degradations (with SA-BIQA) and enables effective few-shot learning for novel tasks. + +To our knowledge, AutoDIR is the most general comprehensive framework for automatic image restoration in the context of complex real-scenario images with multiple unknown degradations. We hope that this framework can provide a solid foundation for further works on complex real-scenario images and has the potential to be extended to various practical applications, such as local image enhancement and multi-task video restoration. + +There are several limitations of AutoDIR that we plan to address in the future. First, the computational cost is similar to Latent Diffusion [49] and other diffusion-based image restoration methods [60], which is often significantly higher than non-generative networks such as NAFNet [10]. Some recent work on accelerating diffusion networks such as one-step diffusion [35] provides possible solutions. Second, AutoDIR currently focuses more on global image restoration, such as deblurring, denoising, and low-light enhancement, rather than local region-based image editing, such as flare removal and inpainting. Extending AutoDIR for local image editing by incorporating semantic segmentation may be an interesting direction to explore. + +# References + +1. Abu-Hussein, S., Tirer, T., Giryes, R.: DIR: Adaptive diffusion for image reconstruction. arXiv preprint arXiv:2212.03221 (2022) +2. Bau, D., Strobelt, H., Peebles, W., Wulff, J., Zhou, B., Zhu, J.Y., Torralba, A.: Semantic photo manipulation with a generative image prior. ACM Transactions on Graphics (TOG) 38(4), 59 (2019) +3. Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +4. Brooks, T., Holynski, A., Efros, A.A.: InstructPix2Pix: Learning to follow image editing instructions. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023) +5. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +6. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG) 42(4), 1-10 (2023) +7. Chen, C., Shi, X., Qin, Y., Li, X., Han, X., Yang, T., Guo, S.: Real-world blind super-resolution via feature matching with implicit high-resolution priors. In: IEEE Transactions on Multimedia (TMM) (2022) +8. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +9. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2024) +0. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of European Conferences on Computer Vision (ECCV) (2022) +1. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2021) +2. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multicontrostatic regularization: Toward a unified model. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +3. Cheon, M., Yoon, S.J., Kang, B., Lee, J.: Perceptual image quality assessment with transformers. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of International Conference on Learning Representations (ICLR) (2020) +5. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2014) + +17. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code GAN prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +18. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-Prompt image editing with cross attention control. Proceedings of International Conference on Learning Representations (ICLR) (2023) +19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2020) +20. Huang, Z., Chan, K.C., Jiang, Y., Liu, Z.: Collaborative diffusion for multi-modal face generation and editing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +21. Hussein, S.A., Tirer, T., Giryes, R.: Image-adaptive GAN based reconstruction. In: Proceedings of the AAAI Conference on Artificial Intelligence (2020) +22. Islam, M.J., Xia, Y., Sattar, J.: Fast underwater image enhancement for improved visual perception. IEEE Robotics and Automation Letters (RA-L) (2020) +23. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: XYDeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +24. Jiang, K., Wang, Z., Yi, P., Chen, C., Huang, B., Luo, Y., Ma, J., Jiang, J.: Multiscale progressive fusion network for single image deraining. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +25. Jinjin, G., Haoming, C., Haoyu, C., Xiaoxing, Y., Ren, J.S., Chao, D.: Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In: Proceedings of European Conferences on Computer Vision (ECCV) (2020) +26. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +27. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +28. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: MUSIQ: Multi-scale image quality transformer. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) +29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2014) +30. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing (TIP) (2018) +31. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-In-One image restoration for unknown corruption. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +32. Li, R., Tan, R.T., Cheong, L.F.: All-in-One bad weather removal using architectural search. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +33. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +34. Liu, K., Jiang, Y., Choi, I., Gu, J.: Learning image-adaptive codebooks for class-agnostic image restoration. Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023) + +35. Liu, X., Zhang, X., Ma, J., Peng, J., Liu, Q.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380 (2023) +36. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: PULSE: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +37. Mingde, Y., Jie, H., Xin, J., Ruikang, X., Shenglong, Z., Man, Z., Zhiwei, X.: Generalized lightness adaptation with channel selective normalization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023) +38. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a "completely blind" image quality analyzer. IEEE Signal processing letters (3), 209-212 (2012) +39. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models (2021) +40. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 7474-7489 (2021) +41. Park, D., Lee, B.H., Chun, S.Y.: All-in-One image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +42. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023) +43. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +44. Qin, X., Wang, Z., Bai, Y., Xie, X., Jia, H.: FFA-Net: Feature fusion attention network for single image dehazing. In: Proceedings of the AAAI Conference on Artificial Intelligence (2020) +45. Quan, Y., Chen, M., Pang, T., Ji, H.: Self2Self with dropout: Learning self-supervised denoising from single image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning (ICML) (2021) +47. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +48. Ren, D., Zuo, W., Hu, Q., Zhu, P., Meng, D.: Progressive image deraining networks: A better and simpler baseline. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +50. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) + +51. Sheikh, H.R., Sabir, M.F., Bovik, A.C.: A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing (TIP) (2006) +52. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +53. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE Transactions on Image Processing (TIP) (2023) +54. Talebi, H., Milanfar, P.: Nima: Neural image assessment. IEEE Transactions on Image Processing (TIP) 27(8), 3998-4011 (2018) +55. Tao, M., Tang, H., Wu, F., Jing, X.Y., Bao, B.K., Xu, C.: DF-GAN: A simple and effective baseline for text-to-image synthesis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +56. Tsai, M.F., Liu, T.Y., Qin, T., Chen, H.H., Ma, W.Y.: FRank: a ranking method with fidelity loss. In: ACM SIGGRAPH Conference Proceedings (2007) +57. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +58. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2017) +59. Wang, J., Chan, K.C., Loy, C.C.: Exploring CLIP for assessing the look and feel of images. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023) +60. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. In: arXiv preprint arXiv:2305.07015 (2023) +61. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: Proceedings of International Conference on Machine Learning (ICML) (2022) +62. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. In: Proceedings of International Conference on Learning Representations (ICLR) (2023) +63. Whang, J., Delbracio, M., Talebi, H., Sahara, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +64. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +65. Yang, T., Ren, P., Xie, X., Zhang, L.: GAN prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +66. Yao, X., She, D., Zhao, S., Liang, J., Lai, Y.K., Yang, J.: Attention-aware polarity sensitive embedding for affective image retrieval. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +67. Ye, H., Yang, X., Takac, M., Sunderraman, R., Ji, S.: Improving text-to-image synthesis using contrastive learning. Proceedings of The British Machine Vision Conference (BMVC) (2021) +68. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild (2024) + +69. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +70. Zhang, H., Koh, J.Y., Baldridge, J., Lee, H., Yang, Y.: Cross-modal contrastive learning for text-to-image generation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +71. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +72. Zhang, L., Chen, X., Wang, Y., Lu, Y., Qiao, Y.: Brush your text: Synthesize any scene text on images via diffusion model. arXiv preprint arXiv:2312.12232 (2023) +73. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +74. Zhang, W., Ma, K., Yan, J., Deng, D., Wang, Z.: Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology 30(1), 36-47 (2020) +75. Zhang, Z., Jiang, Y., Jiang, J., Wang, X., Luo, P., Gu, J.: STAR: A structure-aware lightweight transformer for real-time image enhancement. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2021) +76. Zhang, Z., Jiang, Y., Shao, W., Wang, X., Luo, P., Lin, K., Gu, J.: Real-time controllable denoising for image and video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +77. Zhang, Z., Han, L., Ghosh, A., Metaxas, D.N., Ren, J.: SINE: Single image editing with text-to-image diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +78. Zhou, S., Chan, K., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022) +79. Zhou, Y., Ren, D., Emerton, N., Lim, S., Large, T.: Image restoration for underdisplay camera. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +80. Zhu, M., Pan, P., Chen, W., Yang, Y.: DM-GAN: Dynamic memory generative adversarial networks for text-to-image synthesis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +81. Zhu, X., Zhu, J., Li, H., Wu, X., Li, H., Wang, X., Dai, J.: Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) \ No newline at end of file diff --git a/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/images.zip b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c63ab43304bf8deb1196550f29ee2a6f9e4b8f45 --- /dev/null +++ b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a877cc80d085d92e92576a545798e0ac01506fb0a2abbdc23f1bce43d535c83 +size 696647 diff --git a/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/layout.json b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f623dc2352b5a2931127a82bb8b20e72fde5a0fe --- /dev/null +++ b/2024/AutoDIR_ Automatic All-in-One Image Restoration with Latent Diffusion/layout.json @@ -0,0 +1,10570 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 111, + 445, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 111, + 445, + 146 + ], + "spans": [ + { + "bbox": [ + 170, + 111, + 445, + 146 + ], + "type": "text", + "content": "AutoDIR: Automatic All-in-One Image Restoration with Latent Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "spans": [ + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 190, + 463, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 190, + 463, + 213 + ], + "spans": [ + { + "bbox": [ + 149, + 190, + 463, + 213 + ], + "type": "text", + "content": "1The Chinese University of Hong Kong 2Shanghai AI Laboratory {ytjiang@link, zhaoyangzhang@link, tfxue@ie, jwgu@cse}.cuhk.edu.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 243, + 455, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 455, + 419 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 455, + 419 + ], + "type": "text", + "content": "Abstract. We present AutoDIR, an innovative all-in-one image restoration system incorporating latent diffusion. AutoDIR excels in its ability to automatically identify and restore images suffering from a range of unknown degradations. AutoDIR offers intuitive open-vocabulary image editing, empowering users to customize and enhance images according to their preferences. AutoDIR consists of two key stages: a Blind Image Quality Assessment (BIQA) stage based on a semantic-agnostic vision-language model which automatically detects unknown image degradations for input images, an All-in-One Image Restoration (AIR) stage utilizes structural-corrected latent diffusion which handles multiple types of image degradations. Extensive experimental evaluation demonstrates that AutoDIR outperforms state-of-the-art approaches for a wider range of image restoration tasks. The design of AutoDIR also enables flexible user control (via text prompt) and generalization to new tasks as a foundation model of image restoration. Project is available at: https://jiangyitong.github.io/AutoDIR_webpage/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 441, + 230, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 441, + 230, + 453 + ], + "spans": [ + { + "bbox": [ + 133, + 441, + 230, + 453 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 466, + 482, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 621 + ], + "type": "text", + "content": "Restoring scene details and improving image quality is often the very first step of any computer vision system, which significantly determines the overall system performance. Depending on the operating environments and the signal processing pipelines, images captured by real-world computer vision systems often undergo multiple unknown degradations, such as noise, resolution loss, motion blur, defocus, color imbalance, chromatic aberration, haze, flare, distortion, etc. Previous image restoration methods often address these basic image degradations separately, including deblurring [23,63], denoising [45,76] deraining [24,48], super-resolution [7,60], low-light enhancement [37,75], deraindrop [43], and dehaze [44,53], by using specific single-task models. While these single-task approaches yield promising results within their respective tasks, they encounter difficulties when applied to complex real-world situations that involve multiple unknown degradations or require multiple steps of enhancement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 622, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 482, + 647 + ], + "type": "text", + "content": "In this paper, we explore a more general model capable of handling multiple unknown degradations for a single image. To achieve this goal, the corresponding" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 231, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 231, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 231, + 665 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 360, + 254 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 360, + 254 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 360, + 254 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 360, + 254 + ], + "type": "image", + "image_path": "2a5c11c6de9c5bac36bcab16b7d0996d6005eccd2bf5ec845dd9b541c5bc33c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 328 + ], + "type": "text", + "content": "Fig. 1: We propose AutoDIR, an automatic all-in-one model for image restoration capable of handling multiple types of image degradations. Left: For images with multiple unknown degradations, AutoDIR automatically decomposes the task into multiple steps and supports user interaction via an intuitive text prompt. Right: AutoDIR effectively restores clean images from different degradations and can handle images with unknown degradations in unseen tasks. (Please zoom in for details)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 367, + 114, + 480, + 254 + ], + "blocks": [ + { + "bbox": [ + 367, + 114, + 480, + 254 + ], + "lines": [ + { + "bbox": [ + 367, + 114, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 367, + 114, + 480, + 254 + ], + "type": "image", + "image_path": "5c50eff13c1b3380bde7eb038e1e144a7dd77521512b2423832b8217b75165eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 352, + 482, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 412 + ], + "type": "text", + "content": "model should possess the following abilities: (1) decomposing and distinguishing unknown degradations, (2) restoring various degradations in a task-agnostic framework, and (3) allowing users to freely tune restoration results according to their visual preferences. Numerous related efforts have been made to solve similar problems, but none of them can simultaneously satisfy all three points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 496 + ], + "type": "text", + "content": "Recent works [13, 66, 74] have explored assessing unknown degradations by training an image classifier and using this information to explicitly select a specific image restoration model [8, 33, 69]. However, these methods require training a specific model for each task and primarily focus on a small set of tasks (e.g., 3 tasks). Experiments demonstrate that on large datasets with a wide range of degradations, training an accurate image degradation classifier solely based on image information is still challenging (See Tab. 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 497, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 556 + ], + "type": "text", + "content": "Other works attempt to unify multiple image restoration methods [26, 62] in a single model by utilizing generative model priors. These methods, however, either require known degradation formulations or can only handle images with a single degradation from a predefined small set [31, 42], thus unable to implicitly distinguish multiple unknown degradations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "content": "In this work, we propose a pipeline named AutoDIR, which satisfies all three aforementioned abilities and can automatically detect and restore images with multiple unknown degradations. AutoDIR consists of two stages: the Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA) stage and the All-in-One Image Restoration (AIR) stage guided by text prompts generated in SA-BIQA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": "In the quality assessment stage, we enable accurate identification of each degradation in cases of unknown artifacts in an open-vocabulary manner. This is achieved through our proposed Semantic-Agnostic CLIP (SA-CLIP) model, which employs a Semantic-Agnostic regularization term to transfer the origi" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 174, + 90, + 452, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 90, + 452, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 90, + 452, + 102 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "nal Semantic-Recognitive CLIP to a semantic-agnostic form, focusing on image structural quality rather than semantic content. Additionally, we can utilize the text embeddings generated by the SA-BIQA stage as instructions to guide the further restoration model. This approach can not only enable effective restoration but also allows for flexible user control and editing at runtime by providing open-vocabulary instructions, as shown in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 272 + ], + "type": "text", + "content": "The restoration stage is to handle degradations using a multitasking image restoration model jointly trained on a wide range of tasks. Given the diverse nature of different tasks (e.g., some tasks like super-resolution need to hallucinate textures, but other tasks like low-light enhancement need to preserve everything except brightness), we propose a hybrid approach that maximizes the hallucinating ability of diffusion models while preserving the image structural consistency by involving extra structural inductive bias." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 481, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 481, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 481, + 368 + ], + "type": "text", + "content": "To evaluate the effectiveness and generalization ability of AutoDIR, we have conducted a comprehensive set of experiments encompassing seven image restoration tasks, including denoising, motion deblurring, low-light enhancement, dehazing, deraining, deraindrop, and super-resolution. Experimental results show that AutoDIR consistently outperforms state-of-the-art methods. AutoDIR is also evaluated for the restoration of images captured by under-display cameras and underwater cameras, which are examples of imaging systems with multiple unknown degradations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 387, + 237, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 237, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 237, + 400 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 665 + ], + "type": "text", + "content": "Unified All-in-One Image Restoration Previous approaches for unified all-in-one image restoration can be categorized into two main groups: unsupervised generative prior-based methods [2,5,15,21,26,36,40,62,65] and end-to-end supervised learning-based methods [12,31,32,41,57,71]. Generative priors for image restoration have been widely exploited for image restoration, in the form of generative adversarial networks (GANs) inversion and filters [1,2,5,17,21,36,40], Vector Quantization codebooks [7,34,58,78], latent diffusion models [1,15,26,62]. Nevertheless, these methods rely on specific known image degradation functions, which constrain the types of degradations they can handle. End-to-end supervised learning-based methods typically learn image embeddings extracted by an auxiliary degradation predictor to guide the image restoration model [31,41,57,71]. For example, Valanarasu et al. [57] leverage a transformer encoder to capture hierarchical features of haze, while Li et al. [31] employ a degradation classifier trained using contrastive learning. Park et al. [41] design a degradation classifier for multiple types of degradation to select appropriate Adaptive Discriminant filters, altering network parameters based on specific degradations. However, these methods are constrained to handle limited image degradation categories and cannot restore images with multiple unknown degradations. In contrast, with an SA-BIQA stage, the proposed AutoDIR is able to unify complex degradations without relying on assumptions about the specific degradation functions." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "AutoDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 356 + ], + "type": "text", + "content": "Text-based Image Synthesis and Editing Early works [55,64,67,70,80] on text-to-image synthesis and editing primarily relied on generative adversarial networks (GANs) [16]. Recently, developments in diffusion models [19, 49] and language-vision models [46] have led to significant progress in text-based image synthesis and editing [6,39,47,50]. The emergence of diffusion-based methods has offered new avenues for text-to-image editing, which can be broadly categorized into data-driven [4] and no-extra-data-required approaches [18,27,77]. In the data-driven category, Brooks et al. [4] employ a stable diffusion model fine-tuned with a large dataset of text prompt and image pairs. However, it is primarily designed for semantic image editing and struggles to deliver satisfactory results in tasks related to image quality enhancement. On the other hand, some methods focus on image editing without extra training data. For instance, recent works interpolate the text embedding [27] or score estimate [77] of the input and desired images. However, it requires individual finetuning of the diffusion model for each input image, leading to time-consuming operations. Hertz et al. [18] improve editing efficiency by directly manipulating cross-attention maps without the need for per-image finetuning. However, it requires interior maps in the reversing process and is thus not applicable to real image editing. Unlike the above methods, AutoDIR supports real image enhancement without per-image finetuning." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 132, + 399, + 482, + 586 + ], + "blocks": [ + { + "bbox": [ + 132, + 399, + 482, + 586 + ], + "lines": [ + { + "bbox": [ + 132, + 399, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 482, + 586 + ], + "type": "image", + "image_path": "b3846f06a49da9827c19d7fae7d4e2f056c7d781fc8be5b354f86e20c6bdd906.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 592, + 482, + 615 + ], + "lines": [ + { + "bbox": [ + 130, + 592, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 482, + 615 + ], + "type": "text", + "content": "Fig. 2: Diagram of the proposed All-in-One Image Restoration with Latent Diffusion (AutoDIR). Refer to Sec. 3 for more details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 89, + 452, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 89, + 452, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 89, + 452, + 102 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "text", + "content": "3 Our Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "content": "Fig. 2 shows the overall flowchart of the proposed AutoDIR (Automatic All-in-One Image Restoration with Latent Diffusion), a unified model that can automatically detect and address multiple unknown degradations in images. AutoDIR comprises two main stages:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 196, + 487, + 292 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 137, + 196, + 481, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 196, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 137, + 196, + 481, + 243 + ], + "type": "text", + "content": "- Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA): This stage automatically identifies the predominant degradations present in the input image, such as noise, blur, haze, and generates a corresponding text prompt, denoted as " + }, + { + "bbox": [ + 137, + 196, + 481, + 243 + ], + "type": "inline_equation", + "content": "e_{auto}" + }, + { + "bbox": [ + 137, + 196, + 481, + 243 + ], + "type": "text", + "content": ", which is subsequently used in the image restoration process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 244, + 487, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 244, + 487, + 292 + ], + "spans": [ + { + "bbox": [ + 137, + 244, + 487, + 292 + ], + "type": "text", + "content": "- All-in-one Image Restoration (AIR): This stage utilizes a Structural-Correction Latent Diffusion Model(SC-LDM) to produce a restored image " + }, + { + "bbox": [ + 137, + 244, + 487, + 292 + ], + "type": "inline_equation", + "content": "I_{res}" + }, + { + "bbox": [ + 137, + 244, + 487, + 292 + ], + "type": "text", + "content": ", guided by the text embedding " + }, + { + "bbox": [ + 137, + 244, + 487, + 292 + ], + "type": "inline_equation", + "content": "e_{auto}" + }, + { + "bbox": [ + 137, + 244, + 487, + 292 + ], + "type": "text", + "content": " from SA-BIQA or a user-defined open-vocabulary instruction " + }, + { + "bbox": [ + 137, + 244, + 487, + 292 + ], + "type": "inline_equation", + "content": "e_{user}" + }, + { + "bbox": [ + 137, + 244, + 487, + 292 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 309, + 481, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 481, + 323 + ], + "type": "text", + "content": "3.1 Semantic-Agnostic Blind Image Quality Assessment (SA-BIQA)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 481, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 481, + 401 + ], + "type": "text", + "content": "As demonstrated by previous works [13, 66, 74], a common approach to assess image degradation is to train an image classifier specifically for that purpose. However, this naive approach faces challenges when dealing with large datasets containing a wide range of artifacts. Even with the use of a heavy Vision Transformer(ViT) encoder, the accuracy of classifying image degradation solely based on image information is limited to " + }, + { + "bbox": [ + 130, + 330, + 481, + 401 + ], + "type": "inline_equation", + "content": "77.65\\%" + }, + { + "bbox": [ + 130, + 330, + 481, + 401 + ], + "type": "text", + "content": ", as shown in Tab. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 556 + ], + "type": "text", + "content": "To address this limitation, we propose leveraging human language knowledge to enhance the detection of image degradation. We introduce a Semantic-Agnostic CLIP (SA-CLIP) model as our backbone for Blind Image Quality Assessment (BIQA). SA-CLIP builds upon the CLIP model, which establishes a connection between human language knowledge and image quality. However, we observed that directly applying CLIP or naively fine-tuning it for BIQA tasks does not yield reliable results, as we will later show in Tab. 1. We delved into this problem and identified the reason behind this issue. The pre-trained CLIP model is primarily trained for visual recognition tasks, which prioritize semantic information rather than image quality. Consequently, it delivers low accuracy in BIQA tasks. For instance, the model may struggle to differentiate between a low-light dog image and a noisy dog image because it focuses more on the \"dog\" aspect rather than the presence of noise or illumination." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 604 + ], + "type": "text", + "content": "To overcome this, we tackle the problem in two steps: (i) We construct a new image quality assessment task for fine-tuning CLIP. (ii) We propose a new regularization term for semantic-agnostic and image-quality-aware training to derive SA-CLIP model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "As illustrated in Fig. 2, let " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " denotes the set of image distortions, i.e. " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_1, c_2, \\dots, c_{K-1}, c_K\\}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " is an image distortion type, e.g. \"noise\" and " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "K-1" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " is the number of distortions we consider. We also add a special type " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "c_K = \"\\mathrm{no}\"" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " to indicate the end of the automatic multiple-step restoration. The textural description prompt set is " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{T | T = \"\\text{A photo needs } \\{c_i\\} \\text{ artifact reduction.}\"\\}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "AutoDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 228, + 185 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 228, + 185 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 228, + 185 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 228, + 185 + ], + "type": "image", + "image_path": "273b652ad63ce3fd963634df577ac5892a5b22d5efd20da325af629726f6394f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 186, + 200, + 194 + ], + "lines": [ + { + "bbox": [ + 156, + 186, + 200, + 194 + ], + "spans": [ + { + "bbox": [ + 156, + 186, + 200, + 194 + ], + "type": "text", + "content": "a) Original CLIP" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 114, + 321, + 185 + ], + "blocks": [ + { + "bbox": [ + 228, + 114, + 321, + 185 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 321, + 185 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 321, + 185 + ], + "type": "image", + "image_path": "89f83f10074e8d1e8455214b5bb9b4ae0cd89108760557b78b2443736a59dadd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 186, + 301, + 194 + ], + "lines": [ + { + "bbox": [ + 251, + 186, + 301, + 194 + ], + "spans": [ + { + "bbox": [ + 251, + 186, + 301, + 194 + ], + "type": "text", + "content": "b) Finetuned CLIP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 202, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 131, + 202, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 202, + 482, + 269 + ], + "type": "text", + "content": "Fig. 3: t-SNE visualization of image embeddings " + }, + { + "bbox": [ + 131, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{I}}(I)" + }, + { + "bbox": [ + 131, + 202, + 482, + 269 + ], + "type": "text", + "content": " of CLIP of Blind Image Quality Assessment (BIQA) on SOTs dehazing dataset [30]. Image embeddings of foggy images and their ground-truth clean images are extracted by a) original CLIP. b) finetuned CLIP. c) SA-CLIP finetuned with semantic-agnostic constraint. This illustrates that semantic-agnostic constraint can separate the embeddings of the degraded images from the clean images, while original CLIP and finetuned CLIP features cannot." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 323, + 114, + 481, + 185 + ], + "blocks": [ + { + "bbox": [ + 323, + 114, + 481, + 185 + ], + "lines": [ + { + "bbox": [ + 323, + 114, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 323, + 114, + 481, + 185 + ], + "type": "image", + "image_path": "3d2556025d7a9319662fe8fdbe3d24fcbac81eac7c485758ceaf9e311c7909f4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 186, + 471, + 194 + ], + "lines": [ + { + "bbox": [ + 331, + 186, + 471, + 194 + ], + "spans": [ + { + "bbox": [ + 331, + 186, + 471, + 194 + ], + "type": "text", + "content": "c) Finetuned CLIP with Semantic-Agnostic constraint" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}" + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "content": ". Given a corrupted image " + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "content": " which undergoes several unknown artifacts, our Semantic-Agnostic CLIP aims to identify the dominant degradation of " + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "content": " and extract the corresponding text embedding. SA-CLIP consists of an image encoder " + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{I}}" + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{T}}" + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "content": ". We first obtain the image embedding " + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{I}}(I) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "content": " and the text embedding " + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{T}}(T) \\in \\mathbb{R}^{K \\times d}" + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "content": ", then we compute the cosine similarity between image embedding " + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{I}}(I)" + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "content": " and each prompt embedding " + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{T}}(T_i) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 293, + 482, + 365 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 373, + 481, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 373, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 233, + 373, + 481, + 401 + ], + "type": "interline_equation", + "content": "\\operatorname {l o g i t} \\left(c _ {i} \\mid I\\right) = \\frac {\\mathcal {E} _ {\\mathcal {I}} (I) \\cdot \\mathcal {E} _ {\\mathcal {T}} \\left(T _ {i}\\right)}{\\| \\mathcal {E} _ {\\mathcal {I}} (I) \\| _ {2} \\| \\mathcal {E} _ {\\mathcal {T}} \\left(T _ {i}\\right) \\| _ {2}}, \\tag {1}", + "image_path": "455809ceeae26ecdb1c522459ed26adbbaebb1ab206a8565d2312d5f35a5286a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "text", + "content": ". Next, we calculate the probability " + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\hat{p}(c_i|I)" + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "text", + "content": " and obtain the output text embedding of BIQA " + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "inline_equation", + "content": "e_{auto}" + }, + { + "bbox": [ + 131, + 408, + 482, + 433 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 235, + 441, + 481, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 441, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 235, + 441, + 481, + 469 + ], + "type": "interline_equation", + "content": "\\hat {p} \\left(c _ {i} \\mid I\\right) = \\frac {\\exp \\left(\\operatorname {l o g i t} \\left(c _ {i} \\mid I\\right)\\right)}{\\sum_ {i = 1} ^ {K} \\exp \\left(\\operatorname {l o g i t} \\left(c _ {i} \\mid I\\right)\\right)}, \\tag {2}", + "image_path": "b74ff32ed694fb74de33ea3781071ebd3722da60e5bf4740b773531d0477ea86.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 249, + 479, + 481, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 249, + 479, + 481, + 512 + ], + "type": "interline_equation", + "content": "e _ {a u t o} = \\sum_ {i = 1} ^ {K} \\hat {p} (c _ {i} | I) \\mathcal {E} _ {\\mathcal {T}} (T _ {i}). \\tag {3}", + "image_path": "db0901011426c6aaeb38263eac6af0ee792137f42e531da8fadce0121fd9dc77.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 522, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 482, + 568 + ], + "type": "text", + "content": "Naive fine-tuning on image quality assessment During the optimization of CLIP model, we freeze the parameters of the text encoder " + }, + { + "bbox": [ + 131, + 522, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{T}}" + }, + { + "bbox": [ + 131, + 522, + 482, + 568 + ], + "type": "text", + "content": " and finetune the image encoder " + }, + { + "bbox": [ + 131, + 522, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{I}}" + }, + { + "bbox": [ + 131, + 522, + 482, + 568 + ], + "type": "text", + "content": " using the multi-class fidelity loss [56]. The fidelity loss can be denoted as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 235, + 567, + 481, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 235, + 567, + 481, + 600 + ], + "type": "interline_equation", + "content": "L _ {F I D} = 1 - \\sum_ {i = 1} ^ {K} \\sqrt {y \\left(c _ {i} | I\\right) \\hat {p} \\left(c _ {i} | I\\right)}, \\tag {4}", + "image_path": "aa7be2afaf29bfbdaefcca72266c152d25f137746048c78e1f89cf5c3473d733.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 605, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 605, + 482, + 628 + ], + "type": "inline_equation", + "content": "y(c_{i}|I)" + }, + { + "bbox": [ + 131, + 605, + 482, + 628 + ], + "type": "text", + "content": " is a binary variable that equals 1 if " + }, + { + "bbox": [ + 131, + 605, + 482, + 628 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 605, + 482, + 628 + ], + "type": "text", + "content": " is the dominant degradation in image " + }, + { + "bbox": [ + 131, + 605, + 482, + 628 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 605, + 482, + 628 + ], + "type": "text", + "content": ", and 0 otherwise." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "Semantic-Agnostic Constraint fine-tuning on image quality assessment Since the original CLIP model is pre-trained on tasks such as image" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 89, + 452, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 89, + 452, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 89, + 452, + 102 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "classification, its corresponding " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{I}}(I)" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " encoders tend to encode images based on their semantic information (e.g., cat or dog) rather than their image quality (e.g., noisy or clean). When we fine-tune the CLIP model for generating texts for BIQA according to image quality, this becomes a significant limitation. As shown in Fig. 3 a) and b), the image embeddings extracted by the original CLIP and the finetuned CLIP on foggy, as well as their corresponding ground-truth clean images, cannot be separated, indicating a focus on semantic information rather than image quality differences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "content": "To address this issue, we propose a novel approach called the semantic-agnostic constraint loss " + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "inline_equation", + "content": "(L_{SA})" + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "content": " to regulate the fine-tuning process and prevent the model from relying solely on semantic information rather than image quality. The semantic-agnostic loss " + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "inline_equation", + "content": "(L_{SA})" + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "content": " applies a penalty when the CLIP model suggests that the artifact " + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "content": " is present in the ground-truth clean image " + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "content": " which corresponds to the degraded image " + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "content": ". This penalty forces the CLIP model to distinguish between " + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "content": " based on image quality, encouraging the CLIP image encoder " + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}_{\\mathcal{I}}(I))" + }, + { + "bbox": [ + 130, + 213, + 480, + 330 + ], + "type": "text", + "content": " to focus on extracting image quality information rather than semantic information. This constraint loss can be derived using the following equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 245, + 332, + 480, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 332, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 245, + 332, + 480, + 363 + ], + "type": "interline_equation", + "content": "L _ {S A} = \\sum_ {i = 1} ^ {K} \\sqrt {y \\left(c _ {i} | I\\right) \\hat {p} \\left(c _ {i} \\mid I _ {g t}\\right)}, \\tag {5}", + "image_path": "0fdf1143ba9c60eb180ad14aba02488a6b3d9d3b62f255bd4637b7e036d37a5e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "inline_equation", + "content": "y(c_{i}|I)" + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "text", + "content": " is a binary variable that is equal to 1 if " + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "text", + "content": " represents the dominant degradation in image " + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "text", + "content": ", and 0 otherwise, and " + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "inline_equation", + "content": "\\hat{p}(c_{i}|I_{gt})" + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "text", + "content": " denotes the probability that the dominant degradation in " + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "text", + "content": " is represented by " + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 369, + 479, + 405 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 405, + 479, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 479, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 479, + 430 + ], + "type": "text", + "content": "We incorporate the semantic-agnostic constraint " + }, + { + "bbox": [ + 130, + 405, + 479, + 430 + ], + "type": "inline_equation", + "content": "L_{SA}" + }, + { + "bbox": [ + 130, + 405, + 479, + 430 + ], + "type": "text", + "content": " with the fidelity loss " + }, + { + "bbox": [ + 130, + 405, + 479, + 430 + ], + "type": "inline_equation", + "content": "L_{F1D}" + }, + { + "bbox": [ + 130, + 405, + 479, + 430 + ], + "type": "text", + "content": ", resulting in the total loss " + }, + { + "bbox": [ + 130, + 405, + 479, + 430 + ], + "type": "inline_equation", + "content": "L_{BIQA}" + }, + { + "bbox": [ + 130, + 405, + 479, + 430 + ], + "type": "text", + "content": " for SA-CLIP model:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 254, + 441, + 480, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 441, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 254, + 441, + 480, + 453 + ], + "type": "interline_equation", + "content": "L _ {B I Q A} = L _ {F I D} + \\lambda L _ {S A}, \\tag {6}", + "image_path": "8ba3d00276d500bfb98973821e1f347e8b1b5d3421a2d3545935f8d6c5cdc218.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "text", + "content": " in our experiments. As shown in Fig. 3 c), CLIP fine-tuned with a semantic-agnostic constraint can separate the embeddings of the degraded images from the clean images as expected." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 161, + 542, + 455, + 579 + ], + "blocks": [ + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "lines": [ + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 479, + 541 + ], + "type": "text", + "content": "Table 1: Average accuracy of image degradation detection on seven degradation tasks and clean image. For details, please refer to the supplementary." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 542, + 455, + 579 + ], + "lines": [ + { + "bbox": [ + 161, + 542, + 455, + 579 + ], + "spans": [ + { + "bbox": [ + 161, + 542, + 455, + 579 + ], + "type": "table", + "html": "
MethodsViT-classifierOriginal CLIPFine-tuned CLIPSA-CLIP (ours)
Average Accuracy77.65%61.96%83.31%97.94%
", + "image_path": "285bcdbf809575f1085a9305cc6b14b4622de653c6e44321a81c786f8bc84c82.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 620, + 366, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 620, + 366, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 366, + 632 + ], + "type": "text", + "content": "3.2 Effectiveness of Semantic-Agnostic CLIP" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 640, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 479, + 665 + ], + "type": "text", + "content": "We compare our SA-CLIP with a ViT-based [14] classifier, original CLIP, and naive fine-tuned CLIP to demonstrate the effectiveness of the Semantic-Agnostic" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "AutoDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 159 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 159 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 159 + ], + "type": "image", + "image_path": "df162e78af1b30ba7ad3190ce0d9498d12736c65101fd09a82cfa52a67470a8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 165, + 480, + 188 + ], + "lines": [ + { + "bbox": [ + 130, + 165, + 480, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 480, + 188 + ], + "type": "text", + "content": "Fig. 4: Structural-Correction Latent Diffusion Model (SC-LDM) can maintain complex structures of the original images, while Latent Diffusion Model (LDM) fails to." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "text", + "content": "training scheme. Tab. 1 presents the effectiveness of our Semantic-Agnostic training scheme. The results demonstrate that our SA-CLIP, which incorporates human language information, achieves higher accuracy compared to the ViT-based classifier which considers only image information. Moreover, when CLIP is finetuned with Semantic-Agnostic regularization, the model focuses more on image quality information rather than semantic information, resulting in superior performance compared to the original pre-trained CLIP and CLIP fine-tuned without Semantic-Agnostic regularization. For a visual demonstration of the effectiveness of the Semantic-Agnostic constraint, please refer to the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 354, + 340, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 340, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 340, + 367 + ], + "type": "text", + "content": "3.3 All-in-one Image Restoration (AIR)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "text", + "content": "Naively applying LDM on multi-task image restoration The All-in-One Image Restoration (AIR) stage aims to handle multiple degradations in a single shared framework. Recent advancements in diffusion-based generative models [4, 49] have demonstrated their remarkable ability to generate diverse images, making them suitable for multi-task image restoration. Previous studies [7, 60] have shown that generative models possess a remarkable ability to generate missing or distorted details, especially for tasks that require hallucination, such as super-resolution. Based on these insights, we conduct the AIR stage based on the Latent Diffusion Model (LDM) [49]. LDM incorporates both text and image embedding conditions to restore images " + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "inline_equation", + "content": "I_{sd}" + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "text", + "content": " using generative priors. The text embedding condition " + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "inline_equation", + "content": "e = \\{e_{auto}, e_{user}\\}" + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "text", + "content": " aims to separate different types of image degradations, while the latent image embedding condition " + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "inline_equation", + "content": "z_I = \\mathcal{E}_{ldm}(I)" + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "text", + "content": " from the image encoder " + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{ldm}" + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "text", + "content": " of the LDM provides structural information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "However, although LDM-based generative models can provide a foundation for multi-task image restoration, they have limitations in reconstructing images with complex and small structures due to the compression-reconstruction process with Variational Autoencoder (VAE) [29], as discussed in [9, 20, 72]. [20] attempted to reduce distortion caused by the compression-reconstruction process by retraining Variational Autoencoder (VAE) networks on category-specific images (e.g., human faces) to learn specialized probability distributions. However, this approach is not suitable for image restoration tasks due to the diverse image content. To address these limitations, we introduce a lightweight plug-in structural-correction module to the LDM, enhancing its ability to handle complex and small structures during image restoration." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 174, + 90, + 452, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 90, + 452, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 90, + 452, + 102 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 214, + 162 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 214, + 162 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 214, + 162 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 214, + 162 + ], + "type": "image", + "image_path": "d55cabfa1fa34f83a49d6defe0d3007f40186590d6bea43c105304fa4b38f268.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 163, + 182, + 172 + ], + "lines": [ + { + "bbox": [ + 165, + 163, + 182, + 172 + ], + "spans": [ + { + "bbox": [ + 165, + 163, + 182, + 172 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 114, + 296, + 162 + ], + "blocks": [ + { + "bbox": [ + 215, + 114, + 296, + 162 + ], + "lines": [ + { + "bbox": [ + 215, + 114, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 215, + 114, + 296, + 162 + ], + "type": "image", + "image_path": "3d59953ed90b4723c4590c13a66ddb9279141a63dc7f017db8e06e51f70fd7d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 163, + 277, + 171 + ], + "lines": [ + { + "bbox": [ + 236, + 163, + 277, + 171 + ], + "spans": [ + { + "bbox": [ + 236, + 163, + 277, + 171 + ], + "type": "text", + "content": "...haze ..." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 297, + 114, + 378, + 162 + ], + "blocks": [ + { + "bbox": [ + 297, + 114, + 378, + 162 + ], + "lines": [ + { + "bbox": [ + 297, + 114, + 378, + 162 + ], + "spans": [ + { + "bbox": [ + 297, + 114, + 378, + 162 + ], + "type": "image", + "image_path": "5edc56324eeb178111d7e0db15c7ab0f413a74d0fee5711ee57468a9a22715c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 163, + 371, + 172 + ], + "lines": [ + { + "bbox": [ + 304, + 163, + 371, + 172 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 371, + 172 + ], + "type": "text", + "content": "...low-resolution..." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 114, + 460, + 162 + ], + "blocks": [ + { + "bbox": [ + 380, + 114, + 460, + 162 + ], + "lines": [ + { + "bbox": [ + 380, + 114, + 460, + 162 + ], + "spans": [ + { + "bbox": [ + 380, + 114, + 460, + 162 + ], + "type": "image", + "image_path": "f0dca7fbfdb041b400b9e3a0fe6195eb62c1d17458ec314235d00cc496744247.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 163, + 446, + 172 + ], + "lines": [ + { + "bbox": [ + 393, + 163, + 446, + 172 + ], + "spans": [ + { + "bbox": [ + 393, + 163, + 446, + 172 + ], + "type": "text", + "content": "... raindrop ..." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 468, + 116, + 481, + 169 + ], + "blocks": [ + { + "bbox": [ + 468, + 116, + 481, + 169 + ], + "lines": [ + { + "bbox": [ + 468, + 116, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 468, + 116, + 481, + 169 + ], + "type": "image", + "image_path": "7e9a69be4fcca3a1f9dd045795aca623f1e877140cf00771b48145e9a940729a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "type": "text", + "content": "Fig. 5: Cross-attention maps of text-conditioned diffusion image restoration. The top-left shows an input image with raindrops on the left half. The remaining plots show the cross-attention masks for the keywords \"haze\", \"low-resolution\", and \"raindrop\" in the text prompt used for restoration. While the cross-attention maps for \"haze\" and \"low-resolution\" are more uniformly distributed over the entire image, the map for \"raindrop\" correctly focuses on the actual image artifacts, as expected." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "text", + "content": "Structural-Correction Latent Diffusion Model (SC-LDM) While the LDM-based generative models can provide a foundation for multi-tasking image restoration, it is widely noticed that they may fail to maintain original image structures such as human faces and texts [9, 20, 72], as shown in Fig. 4. To address the structure distortion, we employ an efficient structural-correction module (SCM) denoted as " + }, + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "text", + "content": ". The purpose of SCM is to extract the contextual information " + }, + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "text", + "content": " from the original image and incorporate it with the intermediate image restoration result " + }, + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "inline_equation", + "content": "I_{sd}" + }, + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "text", + "content": ", decoded by image decoder " + }, + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 273, + 480, + 381 + ], + "type": "text", + "content": " of LDM, in a residual manner. This is achieved through the following equation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 247, + 392, + 480, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 392, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 247, + 392, + 480, + 406 + ], + "type": "interline_equation", + "content": "I _ {r e s} = I _ {s d} + w \\cdot \\mathcal {F} ([ I _ {s d}, I ]), \\tag {7}", + "image_path": "eef889ec2b5bfe106157393ff29c6b6b1abcc8e5d4dc7a4493be8b4a9150280b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "text", + "content": " denotes concatenation, and " + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "text", + "content": " is an adjustable coefficient that ranges between 0 and 1. The value of " + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "text", + "content": " determines the extent to which contextual information is utilized to recover the final result. A larger value of " + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "text", + "content": " emphasizes the use of contextual information, which is beneficial for tasks that require structure consistency, such as low light enhancement. Conversely, a smaller value of " + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 416, + 479, + 524 + ], + "type": "text", + "content": " is often employed to maintain the generation capability of latent diffusion model for tasks like super-resolution. By integrating SCM, AutoDIR effectively restores the distorted context of the original image, seamlessly incorporating the enhancements made during the editing phase as shown in Fig. 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 525, + 479, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 479, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 479, + 559 + ], + "type": "text", + "content": "During training, we fine-tune the UNet backbone " + }, + { + "bbox": [ + 130, + 525, + 479, + 559 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(e,[z_t,z_I],t)" + }, + { + "bbox": [ + 130, + 525, + 479, + 559 + ], + "type": "text", + "content": " of the Latent Diffusion Model (LDM) for image restoration tasks. The objective function is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 190, + 571, + 480, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 571, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 190, + 571, + 480, + 586 + ], + "type": "interline_equation", + "content": "L _ {L D} = \\mathbb {E} _ {\\mathcal {E} _ {l d m} (x), z _ {I}, e, \\epsilon \\sim \\mathcal {N} (0, 1), t} [ \\| \\epsilon - \\epsilon_ {\\theta} (e, [ z _ {t}, z _ {I} ], t) \\| _ {2} ^ {2} ], \\tag {8}", + "image_path": "829dc68e02af7946ba7e588d0eb7affc0cc462c86fedf92030ddb1ddb3d6b9dc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 595, + 479, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 595, + 479, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 595, + 479, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 595, + 479, + 632 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 595, + 479, + 632 + ], + "type": "text", + "content": " is the noisy latent. For the SC-LDM, instead of generating the edited latent " + }, + { + "bbox": [ + 130, + 595, + 479, + 632 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t}" + }, + { + "bbox": [ + 130, + 595, + 479, + 632 + ], + "type": "text", + "content": " with the complete reversing sampling process, which is time-consuming, we utilize the estimated edited latent " + }, + { + "bbox": [ + 130, + 595, + 479, + 632 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 130, + 595, + 479, + 632 + ], + "type": "text", + "content": ", which is calculated by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 231, + 641, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 641, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 231, + 641, + 480, + 669 + ], + "type": "interline_equation", + "content": "\\tilde {z} = \\frac {z _ {t}}{\\sqrt {\\bar {\\alpha}}} - \\frac {\\sqrt {1 - \\bar {\\alpha}} \\epsilon_ {\\theta} (e , [ z _ {t} , z _ {I} ] , t)}{\\sqrt {\\bar {\\alpha}}}, \\tag {9}", + "image_path": "5a6f94017dda91a596b91a95b017c01e3490d77679d765636d9b9e1de8a112bc.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "AutoDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 199 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 199 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 199 + ], + "type": "image", + "image_path": "0d5f5f0b10595e76b7a79643c4c9ca9777641d646e68c664668f55824ce3d80e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 208, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 264 + ], + "type": "text", + "content": "Fig. 6: AutoDIR deals with multiple unknown image degradations. Iterative image restoration on unseen Under-Display Camera (TOLED) dataset [79] and unseen Enhancing Underwater Visual Perception (EUVP) dataset [22]. Left: AutoDIR decomposes and solves the unseen artifact in several steps. Right: all-in-one AirNet [31], PromptIR [42] and zero-shot method GDP [15] for comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}" + }, + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "text", + "content": " represents the noise scheduler introduced in [49]. The loss function for Structural-Correction Latent Diffusion Model (SC-LDM) is further defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 221, + 316, + 480, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 316, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 221, + 316, + 480, + 330 + ], + "type": "interline_equation", + "content": "L _ {A I R} = \\left\\| I _ {g t} - \\left(\\mathcal {F} (\\mathcal {D} (\\tilde {z}), I) + \\mathcal {D} (\\tilde {z})\\right) \\right\\| _ {2} ^ {2}. \\tag {10}", + "image_path": "8a357b0ddc5ad902f4b07846c9e24abe8c9dbd4148f353f5434045c59e872993.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 339, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 471 + ], + "type": "text", + "content": "Mechanism of handling multi-task image restoration Fig. 5 shows our experiment of exploring the mechanism of the text conditions disentangling different image restoration tasks during the reverse diffusion process. We find that different text conditions yield different cross-attention maps. As shown in Fig. 5, changing the text prompt leads to significant changes in the cross-attention map. The map closely aligns with the text prompt, which either uniformly distributes attention across the entire image for the \"dehazing\" prompt, or concentrates on the part with edges or textures for the \"low-resolution\" prompt, or focuses on specific areas such as raindrops for the \"deraindrop\" prompt. This shows that AutoDIR can guide the diffusion attention toward regions that are more likely to have image artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 487, + 230, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 230, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 230, + 502 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": "To fully validate the feasibility and effectiveness of unifying multiple restoration and enhancement tasks, we compare AutoDIR with both CNN-based [10,31,42], and diffusion-based multitasking [49] backbones on two tasks with unknown image degradations as well as seven representative image restoration tasks. Specifically, we highlight the advantages of our generative model-based pipeline over discriminative models for tasks that require hallucinating details such as superresolution. Please refer to the appendix for implementation details and datasets. The source code and trained models will be released upon publication." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 623, + 383, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 383, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 383, + 635 + ], + "type": "text", + "content": "4.1 Results on Multiple Unknown Degradations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "To show that AutoDIR can handle multiple unknown image degradations, we conducted experiments on two additional datasets: the Under-Display-Camera" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 174, + 90, + 452, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 90, + 452, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 90, + 452, + 102 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 218, + 175 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 218, + 175 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 218, + 175 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 218, + 175 + ], + "type": "image", + "image_path": "5e90bb7cf4472c124e6aed89c31b8587906bff4846463e8a452cbf374470456b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 175, + 180, + 178 + ], + "lines": [ + { + "bbox": [ + 172, + 175, + 180, + 178 + ], + "spans": [ + { + "bbox": [ + 172, + 175, + 180, + 178 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 218, + 114, + 302, + 175 + ], + "blocks": [ + { + "bbox": [ + 218, + 114, + 302, + 175 + ], + "lines": [ + { + "bbox": [ + 218, + 114, + 302, + 175 + ], + "spans": [ + { + "bbox": [ + 218, + 114, + 302, + 175 + ], + "type": "image", + "image_path": "b1dfb610abf6620351953d43dd405406be9d6ac6d96a1fe02e06ccd1727ab1c3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 175, + 266, + 178 + ], + "lines": [ + { + "bbox": [ + 254, + 175, + 266, + 178 + ], + "spans": [ + { + "bbox": [ + 254, + 175, + 266, + 178 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 302, + 114, + 392, + 175 + ], + "blocks": [ + { + "bbox": [ + 302, + 114, + 392, + 175 + ], + "lines": [ + { + "bbox": [ + 302, + 114, + 392, + 175 + ], + "spans": [ + { + "bbox": [ + 302, + 114, + 392, + 175 + ], + "type": "image", + "image_path": "f7e022fe3877da94d8a60c0d1d79e8dd7e0e93d4f8d0c415b1f0f61cf87da2fe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 175, + 356, + 178 + ], + "lines": [ + { + "bbox": [ + 340, + 175, + 356, + 178 + ], + "spans": [ + { + "bbox": [ + 340, + 175, + 356, + 178 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 114, + 480, + 175 + ], + "blocks": [ + { + "bbox": [ + 392, + 114, + 480, + 175 + ], + "lines": [ + { + "bbox": [ + 392, + 114, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 392, + 114, + 480, + 175 + ], + "type": "image", + "image_path": "3899a14490feef55417010a0d105b856de2dfbe7a140f5c6963caad7890d5d2e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 175, + 446, + 178 + ], + "lines": [ + { + "bbox": [ + 425, + 175, + 446, + 178 + ], + "spans": [ + { + "bbox": [ + 425, + 175, + 446, + 178 + ], + "type": "text", + "content": "PromptR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 179, + 218, + 236 + ], + "blocks": [ + { + "bbox": [ + 133, + 179, + 218, + 236 + ], + "lines": [ + { + "bbox": [ + 133, + 179, + 218, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 179, + 218, + 236 + ], + "type": "image", + "image_path": "656da66404c1876f084626bf334520c6092714e45e5d6c88a64da6c57b9843a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 237, + 197, + 242 + ], + "lines": [ + { + "bbox": [ + 156, + 237, + 197, + 242 + ], + "spans": [ + { + "bbox": [ + 156, + 237, + 197, + 242 + ], + "type": "text", + "content": "All-in-One NAFNet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 252, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 252, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 480, + 297 + ], + "type": "text", + "content": "Fig.7: Super-resolution qualitative comparisons with AirNet [31], PromptIR [42], NAFNet [10] trained on all-in-one tasks, NAFNet [10] trained on Super-Res, and Latent Diffusion Model (LDM) [49]. See the Appendix for more results (Zoom in for better view)." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 218, + 179, + 302, + 236 + ], + "blocks": [ + { + "bbox": [ + 218, + 179, + 302, + 236 + ], + "lines": [ + { + "bbox": [ + 218, + 179, + 302, + 236 + ], + "spans": [ + { + "bbox": [ + 218, + 179, + 302, + 236 + ], + "type": "image", + "image_path": "97f7a3f1b65fb880ebfc4d0b79fdb32c9a2e13d03b764944606fb804980b7566.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 237, + 280, + 243 + ], + "lines": [ + { + "bbox": [ + 241, + 237, + 280, + 243 + ], + "spans": [ + { + "bbox": [ + 241, + 237, + 280, + 243 + ], + "type": "text", + "content": "Super-res NAFNet" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 302, + 179, + 392, + 236 + ], + "blocks": [ + { + "bbox": [ + 302, + 179, + 392, + 236 + ], + "lines": [ + { + "bbox": [ + 302, + 179, + 392, + 236 + ], + "spans": [ + { + "bbox": [ + 302, + 179, + 392, + 236 + ], + "type": "image", + "image_path": "7310c367a93b8cbb7156ed2d5446092d8f162bb12e48755dee57fe1dadcbbefb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 237, + 355, + 242 + ], + "lines": [ + { + "bbox": [ + 347, + 237, + 355, + 242 + ], + "spans": [ + { + "bbox": [ + 347, + 237, + 355, + 242 + ], + "type": "text", + "content": "LD" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 392, + 179, + 480, + 236 + ], + "blocks": [ + { + "bbox": [ + 392, + 179, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 392, + 179, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 392, + 179, + 480, + 236 + ], + "type": "image", + "image_path": "199ac676e73619f2c6c60a34fc9fd9b8c666957e539e0df4b620cc5f220f9301.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 237, + 441, + 242 + ], + "lines": [ + { + "bbox": [ + 430, + 237, + 441, + 242 + ], + "spans": [ + { + "bbox": [ + 430, + 237, + 441, + 242 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 321, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 501 + ], + "type": "text", + "content": "(TOLED) dataset [79] and the Enhancing Underwater Visual Perception (EUVP) dataset [22]. These two datasets consist of multiple unknown image degradations and are completely excluded for training, allowing us to evaluate the performance of AutoDIR in handling unknown, diverse image distortions. Fig. 6 demonstrates the iterative process of AutoDIR to handle images from these two datasets. As shown, AutoDIR can automatically decompose the UDC restoration in two steps (\"brightening\" and \"deblurring\"), and the Underwater restoration in two steps (\"deblurring\" and \"dehazing\"), enabling the restoration of input images. For comparison, we also show the results of a recent all-in-one AirNet [31], PromptIR [42], diffusion-based zero-shot image restoration method GDP [15]. Furthermore, as shown in Tab. 3, AutoDIR outperforms all the other black-box and classifier-based all-in-one methods on quantitative perceptual results. We also conducted a user study on 28 images of TOLED and real-captured images by OPPO Reno6 device. We collected 22 forms and 616 responses. And AutoDIR receives " + }, + { + "bbox": [ + 130, + 321, + 482, + 501 + ], + "type": "inline_equation", + "content": "96.9\\%" + }, + { + "bbox": [ + 130, + 321, + 482, + 501 + ], + "type": "text", + "content": " of the votes as the best result, please refer to the appendix for details." + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 149, + 522, + 345, + 601 + ], + "blocks": [ + { + "bbox": [ + 160, + 511, + 330, + 521 + ], + "lines": [ + { + "bbox": [ + 160, + 511, + 330, + 521 + ], + "spans": [ + { + "bbox": [ + 160, + 511, + 330, + 521 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on SR" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 522, + 345, + 601 + ], + "lines": [ + { + "bbox": [ + 149, + 522, + 345, + 601 + ], + "spans": [ + { + "bbox": [ + 149, + 522, + 345, + 601 + ], + "type": "table", + "html": "
MethodSuper Resolution
MUSIQ ↑CLIP-IQA ↑NIQE ↓NIMA ↑LPIPS ↓PSNR ↑SSIM ↑
NAFNet40.770.2928.454.450.27828.520.754
NAFNet-SR45.490.2998.234.560.33227.950.726
AirNet20.210.27611.074.030.42528.140.626
PromptIR22.890.30110.044.160.30929.620.792
LD50.710.3756.444.900.37824.110.649
Ours63.020.5964.625.250.28223.490.599
", + "image_path": "1df20a506b7fa236f3d0ccbe9a7154957a020b53a7e596973990f4da01ef8519.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 354, + 533, + 473, + 602 + ], + "blocks": [ + { + "bbox": [ + 343, + 511, + 484, + 531 + ], + "lines": [ + { + "bbox": [ + 343, + 511, + 484, + 531 + ], + "spans": [ + { + "bbox": [ + 343, + 511, + 484, + 531 + ], + "type": "text", + "content": "Table 3: Quantitative comparison on Zero-shot UDC task." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 354, + 533, + 473, + 602 + ], + "lines": [ + { + "bbox": [ + 354, + 533, + 473, + 602 + ], + "spans": [ + { + "bbox": [ + 354, + 533, + 473, + 602 + ], + "type": "table", + "html": "
MethodUnder-displayCamera
MUSIQ ↑CLIP-IQA ↑NIQE ↓NIMA ↑
NAFNet37.160.2787.4024.258
AirNet41.200.2657.6814.048
PromptIR45.400.3038.2944.344
LD43.410.2887.9834.371
Ours49.270.3116.3924.427
", + "image_path": "b0743c42d37401fb39b2c2309706b34283d45eaa8b69e7472a05c64763fc34bd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 619, + 367, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 367, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 367, + 632 + ], + "type": "text", + "content": "4.2 Results on Multi-task Image Restoration" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We compare with state-of-the-art methods: the CNN-based single-task image restoration pipeline NAFNet [10], CNN-based all-in-one AirNet [31], PromptIR" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "AutoDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 481, + 371 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 481, + 371 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 481, + 371 + ], + "type": "image", + "image_path": "9a5b900d554556b63b95f38df8dfbc259a08f7f0422314e38f9bbf3fa9fe50d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 378, + 482, + 424 + ], + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 424 + ], + "type": "text", + "content": "Fig. 8: Qualitative comparisons with NAFNet [10] and Latent Diffusion Model (LDM) [49], AirNet [31] and PromptIR [42] on deraining, dehazing, low light enhancement, and deraindrop, denoise and deblur tasks. See the Appendix for more results. (Zoom in for better view)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": "[42], and the generative backbone Latent Diffusion [49]. They are all fine-tuned for all seven tasks, employing an identical training configuration to AutoDIR. Tab. 2 and Tab. 4 present quantitative results of AutoDIR and other methods in all seven tasks. As demonstrated, AutoDIR achieves competitive performance across all seven tasks. Specifically, to further evaluate the ability of discriminative and generative methods on super-resolution tasks, we include task-specific discriminative NAFNet training on SR. As shown in Tab. 2 and Fig. 7, AutoDIR quantitatively and qualitatively outperforms all the all-in-one discriminative, task-specific discriminative methods and the black-box latent diffusion model. Please be noted that, as discussed in previous works [3,25,68], full-reference metrics like PSNR, SSIM, and LPIPS [73] often do not align with human evaluation on super-resolution tasks. Therefore, following [60,68], we include non-reference metrics MUSIQ [28], CLIP-IQA [59] NIQE [38] and NIMA [54] to better evaluate the results. We also show the qualitative comparison results on the other six image restoration tasks in Fig. 8, which demonstrate that AutoDIR can achieve more visually appealing results than the other baselines. A comprehensive and detailed analysis of the quantitative and qualitative results for each specific task can be found in the appendix." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 89, + 452, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 89, + 452, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 89, + 452, + 102 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 148, + 482, + 211 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 146 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 146 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 146 + ], + "type": "text", + "content": "Table 4: Quantitative comparison on denoising, deraining, dehazing, deraindrop, and low light enhancement tasks. The best results are marked in boldface and the second-best results are underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 148, + 482, + 211 + ], + "lines": [ + { + "bbox": [ + 134, + 148, + 482, + 211 + ], + "spans": [ + { + "bbox": [ + 134, + 148, + 482, + 211 + ], + "type": "table", + "html": "
MethodDenoiseDerainDehazeDeraindropLow lightDeblur
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
NAFNet0.80528.100.1950.92630.460.1000.96827.750.0280.87225.040.1380.82721.150.2400.80526.670.208
LD0.62522.580.1860.65123.210.1470.76323.490.0910.73824.840.1040.77018.970.1590.69522.530.241
AirNet0.80329.100.2130.92930.990.0550.94426.520.0310.89227.130.0940.81821.260.2370.80026.500.201
PromptIR0.82429.890.1810.93833.970.0490.97129.130.0220.90027.410.0810.83122.420.1330.81926.820.198
Ours0.83229.680.1670.96535.090.0470.97329.340.0200.92430.100.0630.88822.370.1290.82827.070.157
", + "image_path": "62924e3e7d90510a80645dc8d31f7c814b61e130b337f91dac72b0e5ef8d62b3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 220, + 238, + 308 + ], + "blocks": [ + { + "bbox": [ + 133, + 220, + 238, + 308 + ], + "lines": [ + { + "bbox": [ + 133, + 220, + 238, + 308 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 238, + 308 + ], + "type": "image", + "image_path": "a1c2ce9d065d8a641fe4c2fd848ec33a7e75c7f7d5acb6288a43099f8966b4ec.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 310, + 206, + 318 + ], + "lines": [ + { + "bbox": [ + 189, + 310, + 206, + 318 + ], + "spans": [ + { + "bbox": [ + 189, + 310, + 206, + 318 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 238, + 220, + 318, + 308 + ], + "blocks": [ + { + "bbox": [ + 238, + 220, + 318, + 308 + ], + "lines": [ + { + "bbox": [ + 238, + 220, + 318, + 308 + ], + "spans": [ + { + "bbox": [ + 238, + 220, + 318, + 308 + ], + "type": "image", + "image_path": "1005250dba3b5903da76d2eac5cd6fc772c1d3b666c52a9b7da3b9d8b82f694a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 310, + 292, + 318 + ], + "lines": [ + { + "bbox": [ + 266, + 310, + 292, + 318 + ], + "spans": [ + { + "bbox": [ + 266, + 310, + 292, + 318 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 220, + 399, + 308 + ], + "blocks": [ + { + "bbox": [ + 318, + 220, + 399, + 308 + ], + "lines": [ + { + "bbox": [ + 318, + 220, + 399, + 308 + ], + "spans": [ + { + "bbox": [ + 318, + 220, + 399, + 308 + ], + "type": "image", + "image_path": "780ce63fcdf2988697ebadc33d3f13b30a0b062e1487ac6ef4b0a99cc5d4f005.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 310, + 372, + 318 + ], + "lines": [ + { + "bbox": [ + 348, + 310, + 372, + 318 + ], + "spans": [ + { + "bbox": [ + 348, + 310, + 372, + 318 + ], + "type": "text", + "content": "AIR-v0" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 327, + 480, + 360 + ], + "lines": [ + { + "bbox": [ + 131, + 327, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 327, + 480, + 360 + ], + "type": "text", + "content": "Fig. 9: Qualitative comparisons on NAFNet [10], AIR finetuned i) with latent diffusion weight (AIR-v0) and ii) with weight pre-trained on seven tasks (AIR-v1) on unseen de-snowing and JPEG task." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 399, + 220, + 480, + 308 + ], + "blocks": [ + { + "bbox": [ + 399, + 220, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 399, + 220, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 399, + 220, + 480, + 308 + ], + "type": "image", + "image_path": "4c7b0811ea0605ae7c16f1290a10d7896326b150ff543cedee5049256553de10.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 310, + 449, + 318 + ], + "lines": [ + { + "bbox": [ + 427, + 310, + 449, + 318 + ], + "spans": [ + { + "bbox": [ + 427, + 310, + 449, + 318 + ], + "type": "text", + "content": "AIR-v1" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 383, + 245, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 245, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 245, + 394 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 402, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 510 + ], + "type": "text", + "content": "The first ablation study is to validate the importance of the Semantic-Agnostic BIQA and Structural Correction Module (SCM), as shown in Tab. 5. The presence of proper guidance from the SA-BIQA is crucial for enabling accurate image restoration or enhancement in complex tasks. As shown in Tab. 5, incorporating the SA-BIQA into the AutoDIR pipeline consistently improves performance in multitasking scenarios. Similarly, the Structural-Correction latent diffusion can maintain fine image structures compared with original latent-diffusion-based image restoration. Tab. 5 and Fig. 4 also showed that incorporating the SCM consistently achieves performance gains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "content": "Our second ablation study is to investigate the possibility of using the all-in-one image restoration backbone (AIR) as a foundation model for general image restoration, which is based on the hypothesis that joint learning of multiple image restoration tasks may result in basis operators that are useful for other image restoration tasks. Inspired by prior work on vision foundation models [52,61,81], we conduct experiments on the unseen de-snowing and JPEG tasks to evaluate this hypothesis. For our de-snowing experiments, we utilize 90 snowy-clean image pairs of the Comprehensive Snow Dataset (CSD) [11] to fine-tune the AIR which is initialized with the pre-trained AIR weights obtained from training on the seven tasks. For the JPEG compression task, we use 40 compressed-clean image pairs of LIVE dataset [51]. Fig. 5 demonstrates the fine-tuned AIR with the initialization of pre-trained weight adapts to the de-snowing task more quickly compared to the module without the pre-trained weights. This shows the new" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "AutoDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 208, + 135, + 406, + 216 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "type": "text", + "content": "Table 5: Ablation study on BIQA and SCM on deraindrop and dehazing tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 208, + 135, + 406, + 216 + ], + "lines": [ + { + "bbox": [ + 208, + 135, + 406, + 216 + ], + "spans": [ + { + "bbox": [ + 208, + 135, + 406, + 216 + ], + "type": "table", + "html": "
BIQASCMDeraindropDehaze
SSIM↑PSNR↑LPIPS↓SSIM↑PSNR↑LPIPS↓
XX0.73824.840.1040.76323.490.091
X✓0.90328.840.0680.92527.530.038
✓X0.89625.580.0920.80424.470.074
✓✓0.92430.100.0630.97329.340.020
", + "image_path": "8296901db263fd09d912baa2ff39cb65dad325b4612d6ee7f18a4274d8b99eec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 237, + 482, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 299 + ], + "type": "text", + "content": "image restoration task can benefit from the knowledge acquired from other image restoration tasks, indicating the implicit correlations between them. Moreover, it also outperforms the NAFNet specifically trained for the de-snowing task, which further shows the feasibility of constructing a foundational image restoration model by exploring the underlying correlations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 316, + 323, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 323, + 330 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 323, + 330 + ], + "type": "text", + "content": "5 Conclusions and Discussions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 341, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 425 + ], + "type": "text", + "content": "In this paper, we present AutoDIR, an all-in-one image restoration system with latent diffusion, which is capable of automatically restoring images with multiple unknown degradations and also supports open-vocabulary image editing user instructions. By jointly training from multiple individual image restoration datasets, AutoDIR can be viewed as a foundation model for image restoration, which can automatically detect and address multiple image degradations (with SA-BIQA) and enables effective few-shot learning for novel tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 496 + ], + "type": "text", + "content": "To our knowledge, AutoDIR is the most general comprehensive framework for automatic image restoration in the context of complex real-scenario images with multiple unknown degradations. We hope that this framework can provide a solid foundation for further works on complex real-scenario images and has the potential to be extended to various practical applications, such as local image enhancement and multi-task video restoration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": "There are several limitations of AutoDIR that we plan to address in the future. First, the computational cost is similar to Latent Diffusion [49] and other diffusion-based image restoration methods [60], which is often significantly higher than non-generative networks such as NAFNet [10]. Some recent work on accelerating diffusion networks such as one-step diffusion [35] provides possible solutions. Second, AutoDIR currently focuses more on global image restoration, such as deblurring, denoising, and low-light enhancement, rather than local region-based image editing, such as flare removal and inpainting. Extending AutoDIR for local image editing by incorporating semantic segmentation may be an interesting direction to explore." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 89, + 452, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 89, + 452, + 103 + ], + "spans": [ + { + "bbox": [ + 173, + 89, + 452, + 103 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Abu-Hussein, S., Tirer, T., Giryes, R.: DIR: Adaptive diffusion for image reconstruction. arXiv preprint arXiv:2212.03221 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 158, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 158, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 138, + 158, + 480, + 191 + ], + "type": "text", + "content": "2. Bau, D., Strobelt, H., Peebles, W., Wulff, J., Zhou, B., Zhu, J.Y., Torralba, A.: Semantic photo manipulation with a generative image prior. ACM Transactions on Graphics (TOG) 38(4), 59 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 212 + ], + "type": "text", + "content": "3. Blau, Y., Michaeli, T.: The perception-distortion tradeoff. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 212, + 480, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 480, + 244 + ], + "type": "text", + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: InstructPix2Pix: Learning to follow image editing instructions. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 244, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 244, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 244, + 480, + 277 + ], + "type": "text", + "content": "5. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 480, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 309 + ], + "type": "text", + "content": "6. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG) 42(4), 1-10 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "7. Chen, C., Shi, X., Qin, Y., Li, X., Han, X., Yang, T., Guo, S.: Real-world blind super-resolution via feature matching with implicit high-resolution priors. In: IEEE Transactions on Multimedia (TMM) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "text", + "content": "8. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 373, + 480, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 480, + 406 + ], + "type": "text", + "content": "9. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 406, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 406, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 406, + 480, + 427 + ], + "type": "text", + "content": "0. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of European Conferences on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 427, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 480, + 471 + ], + "type": "text", + "content": "1. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "type": "text", + "content": "2. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multicontrostatic regularization: Toward a unified model. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "text", + "content": "3. Cheon, M., Yoon, S.J., Kang, B., Lee, J.: Perceptual image quality assessment with transformers. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 546, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 590 + ], + "type": "text", + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of International Conference on Learning Representations (ICLR) (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "5. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2014)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "AutoDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "17. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code GAN prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "18. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-Prompt image editing with cross attention control. Proceedings of International Conference on Learning Representations (ICLR) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "20. Huang, Z., Chan, K.C., Jiang, Y., Liu, Z.: Collaborative diffusion for multi-modal face generation and editing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "21. Hussein, S.A., Tirer, T., Giryes, R.: Image-adaptive GAN based reconstruction. In: Proceedings of the AAAI Conference on Artificial Intelligence (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "22. Islam, M.J., Xia, Y., Sattar, J.: Fast underwater image enhancement for improved visual perception. IEEE Robotics and Automation Letters (RA-L) (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "text", + "content": "23. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: XYDeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "text", + "content": "24. Jiang, K., Wang, Z., Yi, P., Chen, C., Huang, B., Luo, Y., Ma, J., Jiang, J.: Multiscale progressive fusion network for single image deraining. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "25. Jinjin, G., Haoming, C., Haoyu, C., Xiaoxing, Y., Ren, J.S., Chao, D.: Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In: Proceedings of European Conferences on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "26. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 456 + ], + "type": "text", + "content": "27. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "text", + "content": "28. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: MUSIQ: Multi-scale image quality transformer. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 422, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 422, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 422, + 501 + ], + "type": "text", + "content": "29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "30. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing (TIP) (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "text", + "content": "31. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-In-One image restoration for unknown corruption. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 566, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 480, + 600 + ], + "type": "text", + "content": "32. Li, R., Tan, R.T., Cheong, L.F.: All-in-One bad weather removal using architectural search. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "33. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "34. Liu, K., Jiang, Y., Choi, I., Gu, J.: Learning image-adaptive codebooks for class-agnostic image restoration. Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 174, + 89, + 452, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 89, + 452, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 89, + 452, + 102 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Liu, X., Zhang, X., Ma, J., Peng, J., Liu, Q.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "36. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: PULSE: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "type": "text", + "content": "37. Mingde, Y., Jie, H., Xin, J., Ruikang, X., Shenglong, Z., Man, Z., Zhiwei, X.: Generalized lightness adaptation with channel selective normalization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 237 + ], + "type": "text", + "content": "38. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters (3), 209-212 (2012)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "type": "text", + "content": "39. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 304 + ], + "type": "text", + "content": "40. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 7474-7489 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 346 + ], + "type": "text", + "content": "41. Park, D., Lee, B.H., Chun, S.Y.: All-in-One image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "type": "text", + "content": "42. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "text", + "content": "43. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "type": "text", + "content": "44. Qin, X., Wang, Z., Bai, Y., Xie, X., Jia, H.: FFA-Net: Feature fusion attention network for single image dehazing. In: Proceedings of the AAAI Conference on Artificial Intelligence (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 479 + ], + "type": "text", + "content": "45. Quan, Y., Chen, M., Pang, T., Ji, H.: Self2Self with dropout: Learning self-supervised denoising from single image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 522 + ], + "type": "text", + "content": "46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning (ICML) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 555 + ], + "type": "text", + "content": "47. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "type": "text", + "content": "48. Ren, D., Zuo, W., Hu, Q., Zhu, P., Meng, D.: Progressive image deraining networks: A better and simpler baseline. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 481, + 621 + ], + "type": "text", + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "type": "text", + "content": "50. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "AutoDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "51. Sheikh, H.R., Sabir, M.F., Bovik, A.C.: A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing (TIP) (2006)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "type": "text", + "content": "52. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "text", + "content": "53. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE Transactions on Image Processing (TIP) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "text", + "content": "54. Talebi, H., Milanfar, P.: Nima: Neural image assessment. IEEE Transactions on Image Processing (TIP) 27(8), 3998-4011 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "55. Tao, M., Tang, H., Wu, F., Jing, X.Y., Bao, B.K., Xu, C.: DF-GAN: A simple and effective baseline for text-to-image synthesis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "text", + "content": "56. Tsai, M.F., Liu, T.Y., Qin, T., Chen, H.H., Ma, W.Y.: FRank: a ranking method with fidelity loss. In: ACM SIGGRAPH Conference Proceedings (2007)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "text", + "content": "57. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "58. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "text", + "content": "59. Wang, J., Chan, K.C., Loy, C.C.: Exploring CLIP for assessing the look and feel of images. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 363, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 384 + ], + "type": "text", + "content": "60. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. In: arXiv preprint arXiv:2305.07015 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 384, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 480, + 428 + ], + "type": "text", + "content": "61. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: Proceedings of International Conference on Machine Learning (ICML) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 460 + ], + "type": "text", + "content": "62. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. In: Proceedings of International Conference on Learning Representations (ICLR) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 460, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 492 + ], + "type": "text", + "content": "63. Whang, J., Delbracio, M., Talebi, H., Sahara, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 492, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 480, + 536 + ], + "type": "text", + "content": "64. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "text", + "content": "65. Yang, T., Ren, P., Xie, X., Zhang, L.: GAN prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "66. Yao, X., She, D., Zhao, S., Liang, J., Lai, Y.K., Yang, J.: Attention-aware polarity sensitive embedding for affective image retrieval. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "67. Ye, H., Yang, X., Takac, M., Sunderraman, R., Ji, S.: Improving text-to-image synthesis using contrastive learning. Proceedings of The British Machine Vision Conference (BMVC) (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "68. Yu, F., Gu, J., Li, Z., Hu, J., Kong, X., Wang, X., He, J., Qiao, Y., Dong, C.: Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild (2024)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 174, + 89, + 452, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 89, + 452, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 89, + 452, + 102 + ], + "type": "text", + "content": "Yitong Jiang\\*1,2 Zhaoyang Zhang\\*1 Tianfan Xue1 Jinwei Gu1" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 544 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "69. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "70. Zhang, H., Koh, J.Y., Baldridge, J., Lee, H., Yang, Y.: Cross-modal contrastive learning for text-to-image generation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "71. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "text", + "content": "72. Zhang, L., Chen, X., Wang, Y., Lu, Y., Qiao, Y.: Brush your text: Synthesize any scene text on images via diffusion model. arXiv preprint arXiv:2312.12232 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "73. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "74. Zhang, W., Ma, K., Yan, J., Deng, D., Wang, Z.: Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology 30(1), 36-47 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "text", + "content": "75. Zhang, Z., Jiang, Y., Jiang, J., Wang, X., Luo, P., Gu, J.: STAR: A structure-aware lightweight transformer for real-time image enhancement. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "text", + "content": "76. Zhang, Z., Jiang, Y., Shao, W., Wang, X., Luo, P., Lin, K., Gu, J.: Real-time controllable denoising for image and video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "text", + "content": "77. Zhang, Z., Han, L., Ghosh, A., Metaxas, D.N., Ren, J.: SINE: Single image editing with text-to-image diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "text", + "content": "78. Zhou, S., Chan, K., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "79. Zhou, Y., Ren, D., Emerton, N., Lim, S., Large, T.: Image restoration for underdisplay camera. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "80. Zhu, M., Pan, P., Chen, W., Yang, Y.: DM-GAN: Dynamic memory generative adversarial networks for text-to-image synthesis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 501, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 544 + ], + "type": "text", + "content": "81. Zhu, X., Zhu, J., Li, H., Wu, X., Li, H., Wang, X., Dai, J.: Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "AutoDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_content_list.json b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bddfe3cf22f26f1620cd2a67825e23c13b9009e7 --- /dev/null +++ b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_content_list.json @@ -0,0 +1,3498 @@ +[ + { + "type": "text", + "text": "AutoEval-Video: An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiuyuan Chen $^{1,*}$ , Yuan Lin $^{2,\\dagger}$ , Yuchen Zhang $^{2,\\dagger}$ , and Weiran Huang $^{1,\\dagger}$", + "bbox": [ + 240, + 237, + 761, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 MIFA Lab, Qing Yuan Research Institute, SEIEE, Shanghai Jiao Tong University \n2 ByteDance Research", + "bbox": [ + 217, + 263, + 777, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We propose a novel and challenging benchmark, AutoEval-Video, to comprehensively evaluate large vision-language models in open-ended video question answering. The comprehensiveness of AutoEval-Video is demonstrated in two aspects: 1) AutoEval-Video constructs open-ended video-questions across 9 skill dimensions, addressing capabilities of perception, comprehension, and generation; 2) AutoEval-Video contains newly collected videos that cover over 40 distinct themes. To efficiently evaluate responses to the open-ended questions, we employ an LLM-based evaluation approach, but instead of merely providing a reference answer, we annotate unique evaluation rules for every single instance (video-question pair). To maximize the robustness of these rules, we develop a novel adversarial annotation mechanism. By using instance-specific rules as prompt, GPT-4, as an automatic evaluator, can achieve a stable evaluation accuracy of around $97.0\\%$ , comparable to the $94.9\\% - 97.5\\%$ accuracy of a human evaluator. Furthermore, we assess the performance of eleven large vision-language models on AutoEval-Video. Among them, GPT-4V(ison) significantly outperforms other models, achieving an accuracy of $32.2\\%$ . However, there is still substantial room for improvement compared to human accuracy of $72.8\\%$ . By conducting an extensive case study, we uncover several drawbacks of GPT-4V, such as limited temporal and dynamic comprehension, and overly general responses. Code is available at https://github.com/Xiuyuan-Chen/AutoEval-Video.", + "bbox": [ + 259, + 325, + 743, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 667, + 375, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative large vision-language models [2, 5, 21, 22, 29, 30, 32, 48] have exhibited the considerable potential to address fundamental challenges in video understanding, temporal and causal comprehension [3]. Furthermore, these models can deal with more complex open-ended video tasks, surpassing traditional specialized video tasks, like video captioning [17, 40, 43], action classification [8, 12, 36], and object tracking [6, 42].", + "bbox": [ + 212, + 698, + 787, + 790 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during an internship at ByteDance Research.", + "bbox": [ + 217, + 797, + 609, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Correspondence to Yuan Lin, Yuchen Zhang, and Weiran Huang (weiran.huang@outlook.com).", + "bbox": [ + 218, + 811, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/239d7d0839f6cf035935b13fbf3c4e0fb73a199a89b685e31fa1b2fe98f657c2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SkillExample of Video-Question
Dynamic PerceptionThe video depicts various ways of relaxing and de-stress.\nQuestion: Please List the three relaxation techniques from the video in order.
State Transitions PerceptionThe video captures the process of a chemistry experiment.\nQuestion: What is the change in the color of the liquid after blowing?
Camera Movement PerceptionThe camera's position in the video undergoes a series of adjustments.\nQuestion: What changes have been made to the position of the camera used for filming this video?
Explanatory ReasoningThe camera records a basketball player sinking a shot and flashing a V sign to the camera.\nQuestion: Why is the individual wearing blue jeans in the video making that particular gesture?
Counterfactual ReasoningA purple animal character in the cartoon follows the guidance of another character and falls into a trap.\nQuestion: What action could the purple animal take to avoid the consequences shown in the video?
Predictive ReasoningThe video records a person putting groceries into a cupboard and preparing to close the cupboard door (but without any depiction of the door-closing action).\nQuestion: What the person will do next?
Comparison ReasoningThe video sequentially depicts several individuals drinking water, where one person exhibits discomfort while the others appear normal.\nQuestion: How does the clothing of the uncomfortable person differ from that of the others?
Reasoning with External KnowledgeThe video records a basketball game, in which one team scored two three-pointers and a two-pointer.\nQuestion: In the video footage, what is the score of the team wearing black?
DescriptionThe video highlights two additional features of a vacuum cleaner product.\nQuestion: Sequentially describe the two extra features of the vacuum cleaner.
", + "bbox": [ + 215, + 143, + 785, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: Skill Dimensions and Corresponding Examples for Video-Questions in AutoEval-Video.", + "bbox": [ + 215, + 405, + 785, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Effectively evaluating large vision-language models is crucial for improving their performance, as it identifies areas for improvement and enables the comparison and investigation of various models. Several benchmarks [6,8,10,12,17, 35,36,40,42,43] have been developed to evaluate capabilities of specific tasks in video understanding. Nonetheless, they fail to assess model performance in a comprehensive manner. To close this gap, recent studies [10,20,35] aim to build benchmarks covering a wide range of skills. These benchmarks primarily consist of multiple-choice question-answers to facilitate effective evaluation. However, in comparison to open-ended question answering, multiple-choice question answering presents two key limitations: 1) they simplify the tasks, and 2) they do not accurately reflect real-world scenarios. For instance, in a video-based dialogue, responses should be generated directly, rather than making a selection from a set of options.", + "bbox": [ + 212, + 450, + 785, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce AutoEval-Video, a comprehensive and challenging benchmark designed to evaluate the video understanding capabilities of large vision-language models. AutoEval-Video comprises 327 challenging open-ended video-question instances that span across nine skill dimensions in Tab. 1, which address video-specific perception, comprehension, and generation skills. All videos in AutoEval-Video are newly collected from YouTube, including over 40 genres such as daily life, gaming, sports, and more. We also establish rigorous annotation standards to guarantee the difficulty and quality of our benchmark. These standards include minimizing single-frame bias [3, 15, 19] and language bias [9, 27], and setting objective questions.", + "bbox": [ + 212, + 650, + 787, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assessing responses to open-ended questions can be difficult, especially when examining outputs generated by LLMs. These responses may involve additional", + "bbox": [ + 214, + 803, + 785, + 834 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Chen, Y. Lin, Y. Zhang, and W. Huang", + "bbox": [ + 271, + 114, + 558, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "information such as video description and chain-of-thought [41] analysis, diminishing the reliability of traditional automatic evaluation metrics, which typically involve comparing a given response to a reference answer, such as BLEU [34], ROUGE [25] and CIDEr [39]. To address this issue, we employ an LLM-based evaluation approach [7, 31, 45, 49]. An LLM evaluator comprises a strong LLM and its prompting strategy. Existing works [26, 46, 50] design a unified prompt for specific tasks. However, due to the diversity of videos and questions in AutoEval-Video, a single unified prompt is insufficient for accurate evaluation. To tackle this challenge, we meticulously annotate distinct evaluation rules for each instance instead of solely providing a reference answer. To improve the precision of these rules, we develop a novel adversarial annotation mechanism. A red team is engaged to devise unaddressed corner cases that the rules may not cover, assisting the annotation team in refining them. The final rule is ultimate when the red team cannot discover any further corner cases that break the rules. After intensive annotation, using instance-specific rules as prompt, GPT-4 can serve as an automatic evaluator, achieving an evaluation accuracy of $97.0\\%$ , comparable to the single human accuracy of $97.5\\%$ .", + "bbox": [ + 217, + 145, + 785, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Leveraging AutoEval-Video, we have conducted a comprehensive evaluation of eleven prominent large vision-language models, which fall into two categories: ImageLLM and VideoLLM. Our assessment reveals that GPT-4V(ison) significantly surpasses other models. Utilizing 8 randomly sampled frames as input, GPT-4V achieves an accuracy of $22.2\\%$ , compared to $72.8\\%$ accuracy of humans, indicating considerable potential for improvement. Moreover, we discover that by increasing the input to 16 frames, GPT-4V's performance markedly improves to $32.2\\%$ . Finally, we perform an in-depth case study for GPT-4V, identifying several critical challenges that demand further investigation, such as limited temporal and dynamic understanding, and overly general responses.", + "bbox": [ + 217, + 405, + 785, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 585, + 385, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Vision-Language Models. The remarkable success of LLMs has been accompanied by significant progress in large vision-language models. Recent works [2, 5, 21, 29] improve the visual perception and cognition capabilities by utilizing the strong understanding and generation abilities of LLMs. In addition, some research [22, 32, 48] further incorporates video inputs and leverages the extensive potential of LLMs for video understanding. The considerable progress necessitates more challenging and comprehensive benchmarks to evaluate the video comprehension capabilities of large vision-language models.", + "bbox": [ + 217, + 625, + 785, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Benchmarks for Video Understanding. Several video benchmarks have been employed to assess model performance in various areas, such as action recognition [11, 12, 47], video captioning [13, 14, 43], and action classification [6, 8, 16, 18, 36, 42]. However, these benchmarks do not provide a comprehensive assessment of a model's capabilities in video understanding. To address this issue, some general benchmarks [20, 35] have been established, developing", + "bbox": [ + 217, + 748, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2dfc754396e327471d28713cc48f2c538ae1e0cac0ecd76b2555559c3f0aa801.jpg", + "image_caption": [ + "Q: What are the steps to prepare this dish as shown in the video?" + ], + "image_footnote": [], + "bbox": [ + 225, + 155, + 777, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rules", + "text_level": 1, + "bbox": [ + 488, + 205, + 514, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[Video Description]: In the video, a person first uses a knife to cut tofu into cubes on a cutting board. Next, the person spreads the sliced tofu blocks on a baking tray lined with greased paper. Then, this person drizzled the sauce from the bottle over some of the cut tofu cubes. Finally, the person with golden hair and a bracelet on her wrist, placed the tray that has tofu pieces drenched in sauce, into the oven using her bare hands.", + "bbox": [ + 225, + 215, + 774, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[Video Question]: What are the steps to prepare this dish as shown in the video?", + "bbox": [ + 227, + 237, + 473, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules:", + "bbox": [ + 227, + 244, + 584, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1. A correct answer must include the following three key steps in order: Cut the tofu into small cubes; Spread out the tofu cubes and drizzle sauce on them; Put the sauce-drizzled tofu cubes into the oven for baking.", + "bbox": [ + 227, + 252, + 774, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. The answer to the description of the shape of tofu blocks can be anything similar to 'block'. It doesn't necessarily have to be described as a perfectly shaped cube.", + "bbox": [ + 227, + 265, + 728, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Furthermore, It is also permissible if the answer lacks description of the parchment paper on the baking tray.", + "bbox": [ + 227, + 273, + 566, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. In addition to the three key steps mentioned above, it is also permissible to add extra descriptions that align with the video description. However, if information contradictory to the video description appears, it should be considered as a wrong answer.", + "bbox": [ + 227, + 280, + 774, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. The response should contain correctly spelled words. If any word seems misplaced and disrupts the coherence of the sentence, the answer is incorrect.", + "bbox": [ + 227, + 292, + 692, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[Answer to be judged]: ", + "bbox": [ + 227, + 300, + 344, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[Form of the result]:", + "bbox": [ + 227, + 306, + 295, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above.", + "bbox": [ + 227, + 314, + 774, + 328 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a1230a3fb782e2e315ae6b93ea9f67b033ca5ccef49d7e27d1c6d861ab8f5f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 333, + 334, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "text_level": 1, + "bbox": [ + 292, + 348, + 336, + 354 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Cut the tofu into pieces and then place it on a wooden board.", + "2. Drizzle the sauce over some tofu slices.", + "3. Put on gloves and place the baking tray with tofu pieces into the oven." + ], + "bbox": [ + 227, + 356, + 401, + 391 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d0327b2a3132818e7f86b10f351db6626d04bdd5b320df59eb6776aa858adc33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 349, + 522, + 356 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Cut the tofu into small cubes and spread them out. \n2. Place the tofu cubes on a baking tray and drizzle with sauce.", + "3. Put the sauce-drizzled tofu cubes into the oven to bake." + ], + "bbox": [ + 418, + 356, + 581, + 392 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ce1d568fa0900963a23c2bab841f309f9fcd171755974ff06f96d795c50f4bfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 349, + 707, + 354 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Cut the white tofu into pieces, then place them in a baking tray.", + "2. Drizzle sauce over some of the tofu pieces.", + "3. Place the baking tray with the tofu pieces in the oven." + ], + "bbox": [ + 599, + 356, + 774, + 383 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b45f7bbfc1fcb8f40bc799d967014199ea9ef5dc3c00f37b239e3b12ceef4646.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 395, + 326, + 417 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "The answer provided is incorrect for the following reasons:", + "1. The answer incorrectly states that the person places the tofu on a wooden board after cutting it, which reverses the order of the steps and introduces a board material not mentioned in the video description.", + "2. The sauce is misspelled as \"sauce,\" which is a spelling error.", + "3. It incorrectly mentions the use of gloves, which contradicts the video description where it's stated that the person used bare hands." + ], + "bbox": [ + 227, + 425, + 400, + 503 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/78a20e3ac12970a7ef7539bd85f8e6ea7aec210f305c7b1b577d53ca816a8879.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 395, + 511, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The [Answer to be judged] correctly includes the three key steps required by the rules: cutting tofu into cubes, spreading the tofu out and drizzling sauce on them, and putting the sauce-drizzled tofu into the oven to bake. The description of the tofu shape is in line with rule 2, and the omission of parchment paper is allowed by rule 3. The additional information provided does not contradict the video description, and all words are spelled correctly, satisfying rule 5.", + "bbox": [ + 416, + 425, + 583, + 498 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39f2b8bb76ab92cb8042a2b382e4741d15fe1c11e75a440949ce04188896f5bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 395, + 699, + 416 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "The answer provided correctly states the three key steps in the order they appear in the video description: cutting the tofu into pieces, drizzling sauce on the tofu, and placing the tray in the oven.", + "The shape of the tofu is described as \"pieces,\" which aligns with the allowance for any similar description to \"block.\" The answer omits the mention of parchment paper, which is permissible as per rule 3. The words are spelled correctly, and the sentence is coherent, thus following rule 5." + ], + "bbox": [ + 599, + 425, + 774, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 1: Example Instance in AutoEval-Video and Automatic Evaluation Process. Each instance has three components: video, question, and rules. For each video and question, we establish specific and detailed evaluation rules. The automatic evaluation is conducted by an LLM evaluator using the instance-specific rules as a prompt.", + "bbox": [ + 215, + 525, + 785, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "multiple-choice VQA to assess models' abilities across different aspects. Nevertheless, compared to open-ended question answering, multiple-choice question answering oversimplify tasks and may not reflect realistic instructions.", + "bbox": [ + 215, + 587, + 785, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, existing benchmarks mainly employ two methods for video collection: 1) sourcing videos related to a specific task from web [4,16] or large-scale video datasets [1,11,47]; 2) recording new videos based on pre-defined scripts [35] or themes [10]. However, both approaches suffer from limited visual diversity, leading to a video distribution that inadequate represents real-world scenarios.", + "bbox": [ + 215, + 633, + 785, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 AutoEval-Video", + "text_level": 1, + "bbox": [ + 215, + 739, + 408, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 777, + 339, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In AutoEval-Video, each instance comprises three components: video, question, and rules, as illustrated in Fig. 1. The construction involves three steps.", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Chen, Y. Lin, Y. Zhang, and W. Huang", + "bbox": [ + 269, + 114, + 558, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, to address the essential aspects of real-world multi-modal understanding, the uncovered areas of existing benchmarks, and the limitations of current models, we identify nine skill dimensions, as illustrated in Tab. 1. These dimensions encompass diverse skill areas (including perception, cognition, and generation) and various reasoning types (such as descriptive, explanatory, predictive, comparative, and counterfactual).", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, for each skill dimension, an annotation team gathers videos from YouTube1 and crafts relevant open-ended questions. It is critical to prevent potential single-frame bias and language bias while formulating these questions. To ensure the assessability of responses, the questions must retain objectivity.", + "bbox": [ + 212, + 239, + 782, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, after acquiring the video-question pairs, the annotation group develops evaluation rules for each instance. These rules establish precise conditions for considering whether a response is correct or wrong, involving precise criteria for recognizing seemingly accurate but flawed corner case responses, and vice versa. To ensure thoroughness, we employ a second annotation team, known as the \"red team\", to devise unaddressed corner cases. The final rule is ultimate when the red team cannot discover any further corner cases that break the rules.", + "bbox": [ + 212, + 301, + 784, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As demonstrated above, the annotation of AutoEval-Video is sophisticated and labor-intensive. We construct a total of 327 instances. The average cost per instance amounts to $28.", + "bbox": [ + 212, + 410, + 784, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Skill Dimensions", + "text_level": 1, + "bbox": [ + 214, + 481, + 398, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video understanding encompasses two fundamental skills: perception and cognition. Different from image perception, video perception emphasizes temporal information, such as activity, state transitions, and camera movement. With regard to cognition, our benchmark considers reasoning based on temporal or spatial information, simultaneously spanning causal reasoning, comparison, and reasoning with external knowledge. In addition, considering that hallucination [24,28,37,51] is a prominent challenge for contemporary large vision-language models, we incorporate tasks requiring descriptive ability in our benchmark. In summary, we identify nine skill dimensions, as shown in Tab. 1, which will address the aforementioned capabilities.", + "bbox": [ + 212, + 508, + 787, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Video and Question Collection", + "text_level": 1, + "bbox": [ + 214, + 685, + 514, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Collection. We gather all videos from YouTube, taking into account the following considerations:", + "bbox": [ + 212, + 713, + 782, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- To minimize the probability of utilizing videos that have been employed in the model training process, we collect new videos from YouTube rather than employing well-known open-sourced video datasets.", + "bbox": [ + 225, + 756, + 782, + 801 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA", + "bbox": [ + 217, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 The videos collected from YouTube comply with Creative Commons License (https://support.google.com/youtube/answer/2797468).", + "bbox": [ + 217, + 810, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- We aim to construct a benchmark that incorporates real-world visual content so as to directly assess the capability of models in addressing realistic challenges. Thus, we choose to select videos from YouTube instead of filming new videos based on scripts.", + "bbox": [ + 223, + 146, + 787, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we examine the video categories of our benchmark, which covers mainstream video genres in real-world. All video sources are guaranteed to possess a resolution of 720P or higher. More statistics of the videos can be found in Fig. 2.", + "bbox": [ + 212, + 236, + 787, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1df86a7d356967af43059472221e271cbc70ba10814001fed209baee5a500ecd.jpg", + "image_caption": [ + "(a) Genre distribution of videos in AutoEval-Video", + "Fig. 2: Statistics of AutoEval-Video. (a) Genre distribution of videos. (b) Statistics on question and rule lengths, number of scenes, and video duration." + ], + "image_footnote": [], + "bbox": [ + 256, + 299, + 485, + 474 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/64c4836cfff2b367fc0cca68d0b52f2ee44f73559b9034e406b6abceb3fb7d6e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StatisticsMeanMedian
Question Length (#words)10.9410.00
Rule Length (#words)302.95296.00
Video Shots (#Scenes)3.162.00
Video Duration (#Seconds)14.5810.00
", + "bbox": [ + 524, + 314, + 750, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(b) Statistics of AutoEval-Video", + "bbox": [ + 522, + 470, + 743, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Question Collection. The questions of AutoEval-Video are meticulously designed by human annotators, avoiding templates or paradigms, to guarantee a diverse range of questions. Annotators may either initiate a question and then search for an appropriate video, or select a video first and craft a question afterward. To maximize dataset diversity, we annotate only one question per video. To obtain high-quality video-question pairs that can effectively diagnose models' video comprehension abilities, we propose three principles for constructing the benchmark: avoiding single-frame bias, reducing language bias, and formulating objective questions, as discussed below.", + "bbox": [ + 212, + 547, + 787, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Many video benchmarks suffer from single-frame bias [3, 15, 19], indicating that the information in an oracle frame is almost sufficient to deduce the answer. To address this issue, we instruct annotators to devise questions that require the consideration of multiple frames for successful completion. Questions that can be answered based on only one frame are deemed invalid.", + "bbox": [ + 212, + 686, + 787, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Existing video-language benchmarks are affected by language bias [9, 27], suggesting that models can obtain the correct answers without seeing video content. To tackle this issue, we involve three additional human annotators who try their best to answer questions without viewing the videos. If any of the annotators can accurately answer the question, the instance will be discarded.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Chen, Y. Lin, Y. Zhang, and W. Huang", + "bbox": [ + 269, + 114, + 558, + 130 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f904da36f262dad975fc90f565726b0b89b99c87aab70e9b5b050e606b349922.jpg", + "image_caption": [ + "Q: Why did the race car driver exit the vehicle?" + ], + "image_footnote": [], + "bbox": [ + 228, + 159, + 295, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6fe5b1c963e28a609ed7ebcd54c1cc0d592b60b97fafc16a8b773523b3caa0cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 159, + 364, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2fed1de0ca7ee75019c78d66e8857be285c51f89a0f51dcbc4f62b1ec4065e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 159, + 433, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0409356616cd15bdc3eab3b965a8caf7dea77599b43dc5b1a3f33058d0ec0fb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 159, + 501, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c14c30632fde6bf9878099fdb7576b18544a5491910757a7f7f4c75262148ce8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 159, + 570, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60bb6dd6c6544299f90b79ae95f491f8e877c66806429e24296d5b36ea2c7ece.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 159, + 638, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08c81c8a6ead6f58c7b0f5d7676a22fabe1b1d16d9345dfef3a2360bbdd94198.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 159, + 705, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11e927e71edb90ad76cad873b1204f02003c95810da8b7484a8da17296bc1ef2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 159, + 784, + 199 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initial Rules", + "text_level": 1, + "bbox": [ + 362, + 213, + 416, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[Video description]: In a racetrack, a white race car flipped over and came to a halt, then caught fire. Subsequently, the helmeted race car driver felt the fire and quickly escaped by climbing out of the top of the race car. Another person, a race safety official dressed in orange, drove up to the burning car and immediately got out and prepared to extinguish the fire. [Video Question]: Why did the race car driver exit the vehicle?", + "bbox": [ + 228, + 222, + 549, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules:", + "bbox": [ + 228, + 261, + 547, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1. The main key point in the answer lies in \"race car caught on fire\". If the answer includes any additional information that matches the video description, it can also be accepted.", + "bbox": [ + 228, + 276, + 547, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2. The words in the answer must be spelled correctly. If a word does not make sense in the sentence, the answer is incorrect.", + "bbox": [ + 228, + 292, + 547, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3. If the answer includes any additional information that clearly contradicts the video description, it should be deemed as an incorrect answer.", + "bbox": [ + 228, + 308, + 547, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[Answer to be judged]: (Adversarial Answer)", + "bbox": [ + 228, + 323, + 406, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[Form of the Result]:", + "bbox": [ + 228, + 330, + 308, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above.", + "bbox": [ + 228, + 339, + 547, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "text_level": 1, + "bbox": [ + 633, + 212, + 725, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Because the race car driver's car caught fire, which might have been caused by the driver's improper operation.", + "bbox": [ + 584, + 220, + 772, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fffe01af2769336d17c414d6fbd5392b475a68c0f05401ea2101ba83f68377e.jpg", + "image_caption": [ + "Fig. 3: Adversarial Annotation for Rules. A red team develops adversarial answers that break the initial rules. Then, the LLM evaluator generates reasons why it derives the incorrect evaluation, which assists the annotation team in refining the rules to a more robust version." + ], + "image_footnote": [], + "bbox": [ + 669, + 247, + 691, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The reason provided in the answer, \"improper operation by the driver\", adds speculative information not present in the video description. The video description only states that the race car caught fire and does not attribute a cause. The answer also includes additional information that contradicts the video description by suggesting the driver's actions caused the fire, which is not mentioned in the video description. Therefore, the answer should be deemed incorrect.", + "bbox": [ + 583, + 277, + 774, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Updated Rules", + "text_level": 1, + "bbox": [ + 467, + 392, + 534, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[Video description]: In a racetrack, a white race car flipped over and came to a halt, then caught fire. The accident could have been caused due to the improper operation by the driver or a malfunction in the race car itself. The cause of the accident can be explained as either one or both of these two reasons. Subsequently, the helmed race car driver felt the fire and quickly escaped by climbing out of the top of the race car. Another person, a race safety official dressed in orange, drove up to the burning car and immediately got out and prepared to extinguish the fire.", + "bbox": [ + 228, + 402, + 772, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[Video Question]: Why did the race car driver exit the vehicle?", + "bbox": [ + 228, + 434, + 446, + 440 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules:", + "bbox": [ + 228, + 441, + 625, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1. The main key point in the answer lies in \"race car caught on fire\". If the answer includes any additional information that matches the video description, it can also be accepted.", + "bbox": [ + 228, + 449, + 772, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2. The words in the answer must be spelled correctly. If a word does not make sense in the sentence, the answer is incorrect.", + "bbox": [ + 228, + 465, + 647, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3. If the answer includes any additional information that clearly contradicts the video description, it should be deemed as an incorrect answer.", + "bbox": [ + 228, + 473, + 705, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. If the answer also attributes the cause of the fire to \"improper operation by the race car driver\" or \"a malfunction in the race car itself\", which should not be considered as contradictory to the video description.", + "bbox": [ + 228, + 479, + 772, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[Answer to be judged]: (Adversarial Answer)", + "bbox": [ + 228, + 496, + 406, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[Form of the Result]:", + "bbox": [ + 228, + 503, + 305, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above.", + "bbox": [ + 228, + 511, + 772, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our benchmark, all question-answer pairs should be assessed by a set of rules, requiring objective questions since evaluating subjective responses is challenging. For instance, a video depicting a man with a neutral expression, accompanied with a question, \"What is the man thinking about?\" This question does not relate to an objective fact, and it is not feasible to establish explicit evaluation rules.", + "bbox": [ + 212, + 608, + 785, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Annotation of Rules", + "text_level": 1, + "bbox": [ + 215, + 719, + 429, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each instance in AutoEval-Video, we annotate a set of rules describing the criteria for determining the correctness of an answer. The rationale behind this approach is as follows:", + "bbox": [ + 212, + 742, + 785, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- All questions in AutoEval-Video are open-ended, rendering the enumeration of all reference answers infeasible. Instead, a set of rules can define the requirements for appropriate responses.", + "bbox": [ + 223, + 794, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Modern LLMs, such as GPT-4, possess remarkable language understanding capabilities, allowing them to comprehend the provided rules and act as automatic evaluators. Unlike other works that use a single, unified prompt to assess all video-question responses [26, 46, 50], we annotate specific rules for each video-question pair, ensuring an in-depth and specific understanding of the criteria.", + "bbox": [ + 225, + 146, + 787, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The valid evaluation rules include a detailed description of the video content, questions about the video, key points of the correct answer, and the required output format, as illustrated in Fig. 1. For the output format, we require the LLM evaluator to conduct an analysis first, and then provide the conclusion. Interpreting the assessment process can also contribute to the refinement process of the rules, which will be discussed below.", + "bbox": [ + 212, + 244, + 782, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To guarantee the thoroughness of the rules, we employ an adversarial annotation process (shown in Fig. 3) conducted by a separate group of annotators, known as the red team. Initially, the annotation team establishes initial rules denoted as $r_0$ . For each $k$ -th round of attack, the red team will endeavor to construct an answer, $a_k$ , causing erroneous judgments from an LLM evaluator (e.g. GPT-4) with the preceding rules $r_{k-1}$ . Upon obtaining an adversarial answer $a_k$ , the annotation team must update their rules to a refined version, $r_k$ . The iterative attack-refine process will continue until the red team cannot identify any adversarial answer. The red team may construct attack answers through various methods, such as: 1) directly writing one based on their understanding of the rules; and 2) prompting LLMs to generate adversarial answers. During the refinement process, the annotators can utilize the interpretations of the LLM evaluator, which will help to quickly figure out why the evaluator gets confused and facilitate the production of more accurate rules. On average, there are 3 attack-refine iterations in the construction of rules in AutoEval-Video.", + "bbox": [ + 212, + 338, + 787, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, to enhance the motivation and engagement of both the annotation team and the red team in confrontation, we devised a reward mechanism. For each instance, the annotation team initially has a full reward, while the red team starts with zero reward. If the red team successfully discovers an adversarial answer, they will earn a reward, while the annotation team will lose a portion of their reward. As the number of iterations increases, there will be a corresponding increase in rewards for successful attacks, as well as higher penalties for the annotation team. This approach not only encourages the red team to target mature rules but also urges the annotation team to consistently refine their rules throughout every iteration, ensuring optimal results.", + "bbox": [ + 212, + 568, + 787, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 739, + 375, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 770, + 428, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Models. Utilizing AutoEval-Video, we assess the performance of eleven models, including six ImageLLMs (i.e., GPT-4V [44], Gemini-Ultra-Vision [38], LLaVA1.5 [29], Qwen-VL [2], InstructBLIP [5], and BLIP-2 [21]) and five VideoLLMs", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Chen, Y. Lin, Y. Zhang, and W. Huang", + "bbox": [ + 271, + 114, + 558, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(i.e., Video-ChatGPT [32], VideoChat, VideoChat2 [22], Video-LLaMA [48], and LLaMA-VID [23]). For each video, we uniformly sample 8 or 16 frames, resulting in two different input configurations. In particular, for ImageLLMs, we concatenate the 8 or 16 frames to create a single image as visual input (refer to supplementary material for the detailed concatenation method).", + "bbox": [ + 212, + 145, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "LLM as an automatic evaluator. We adopt GPT-4 [33] as an automatic evaluator for AutoEval-Video. For each instance, GPT-4 generates an evaluation of the correctness of a given answer to be judged using the rules as prompts, as demonstrated in Fig. 1.", + "bbox": [ + 212, + 224, + 782, + 285 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Human baseline. We recruit three human participants to answer questions in AutoEval-Video. The overall average accuracy is $72.8\\%$ . Through closely examining the mistakes made by humans, we identify two major challenges they face: 1) Questions related to external knowledge, such as unfamiliar games, may lead to incorrect answers from humans; 2) Questions requiring precise memory skills, such as recalling multi-step processes in a video, may cause a person to overlook one or more steps. More examples can be found in supplementary material.", + "bbox": [ + 212, + 287, + 784, + 393 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f388c51b99a739d3b7ba0982b1d3285f4fcf8b29358438c4933317ce0cf14ab7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model TypeModelAccuracy
8 Frames16 Frames
ImageLLMBLIP-2 [21]0.31.2
InstructBLIP [5]7.914.3
Qwen-VL [2]7.010.0
LLaVA-1.5 [29]8.58.2
GPT-4V [44]22.232.2
Gemini-Ultra-Vision [38]17.920.9
VideoLLMVideo-ChatGPT [32]6.74.9
VideoChat [22]13.410.9
Video-LLaMA [48]11.28.5
LLaMA-VID [23]9.88.0
VideoChat2 [22]0.30.6
Human (full videos)72.8
", + "bbox": [ + 284, + 396, + 712, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Evaluation Results of Different Models on AutoEval-Video. We also provide human results after viewing full videos.", + "bbox": [ + 212, + 604, + 784, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Performances on AutoEval-Video", + "text_level": 1, + "bbox": [ + 214, + 674, + 535, + 690 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare 11 large vision-language models on Tab. 2. Findings show: 1) GPT-4V is much stronger than other models; 2) Among open-sourced models, VideoLLM generally outperforms ImageLLM; 3) GPT-4V falls significantly short of human-level, reflecting the challenge of our benchmark and a big headroom for further advancements.", + "bbox": [ + 212, + 700, + 784, + 775 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We find that open-sourced models suffer from three typical categories of bad cases. 1) Hallucination: the described objects or events are not present in the video, leading to wrong answers; 2) Alignment: since AutoEval-Video includes a diverse range of queries, it is often observed that the models' responses do", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA", + "bbox": [ + 217, + 113, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "not align with the question; 3) Weak LLM: many models cannot get to the correct answer due to their limited language generation capabilities. We have provided examples for each category in the supplementary material. GPT-4V shows significant progress in these directions, but it still has fundamental issues, as discussed in Sec. 4.4.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Among open-sourced models, VideoChat achieves the highest score. Interestingly, GPT-4V correctly answers only $32.7\\%$ instances where VideoChat is correct, indicating that VideoChat is stronger in certain domains. These domains are 1) traffic-related tasks; 2) comparison between dynamic objects; and 3) giving video-specific responses rather than general vague answers. In supplementary material, we provide examples for each of these three domains.", + "bbox": [ + 212, + 224, + 787, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, we notice that 8 frames are insufficient for answering certain questions. Therefore, we evaluate models using 16 frames as input. Under this setting, GPT-4V significantly improves, reaching $32.2\\%$ accuracy (examples provided in supplementary material). While longer inputs usually benefit ImageLLMs, VideoLLMs don't exhibit the same trend, possibly because they are explicitly trained in the 8-frame setting.", + "bbox": [ + 212, + 316, + 787, + 407 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ddf774dcf0cc36735fc7093d854851902be8e60d2279f383f67121d090d9c3e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GPT-4 PromptAccuracy of Evaluators
Models/Human ResponsesGPT-4V Responses
Instance-specific rules97.096.6
W/o adversarial annotation91.588.1
W/o rules (unified prompt)87.069.5
Human97.594.9
", + "bbox": [ + 297, + 411, + 699, + 489 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 3: Ablation Study on Automatic Evaluators. We assess the evaluation accuracy of GPT-4 with instance-specific rules, initial rules, and unified prompts on responses generated by models/humans, as well as responses exclusively generated by GPT-4V. In addition, we present the evaluation accuracy of a single human.", + "bbox": [ + 212, + 500, + 787, + 555 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Accuracy of GPT-4 as Automatic Evaluator", + "text_level": 1, + "bbox": [ + 214, + 595, + 624, + 612 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To validate GPT-4 as an automatic evaluator, we create a test dataset consisting of 200 randomly sampled (video, question, response) triplets, either model-generated or human-generated. Then we employed three human experts to evaluate each response, utilizing their collective decision as the ground truth. As shown in Tab. 3, with the evaluation rules, GPT-4 achieves $97\\%$ agreement with the ground truth, comparable to $97.5\\%$ of a single human expert.", + "bbox": [ + 212, + 623, + 787, + 714 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We perform an ablation study to investigate the necessity of our proposed way of annotation. As reported in Tab. 3, if we use the initial version of rules without adversarial annotation, the accuracy drops to $91.5\\%$ ; using a unified prompt (refer to supplementary material) instead of instance-specific rules makes the accuracy further drop to $87\\%$ . It implies that both \"instance-specific rules\" and \"adversarial annotation\" are important.", + "bbox": [ + 212, + 715, + 787, + 806 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As the model gets stronger, it makes fewer simple mistakes (e.g., grammatical errors or obvious hallucinations) and generates longer chain-of-thoughts, thus", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Chen, Y. Lin, Y. Zhang, and W. Huang", + "bbox": [ + 269, + 113, + 557, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the wrong answer becomes harder to distinguish from the correct ones. Tab. 3 shows that the accuracy of a single human, initial rules, and unified prompt all significantly drop on a more challenging subset of responses (50 responses exclusively generated by GPT-4V), while instance-specific rules maintain a relatively stable accuracy at $96.6\\%$ , highlighting its robustness for evaluating strong models.", + "bbox": [ + 212, + 146, + 787, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A key concern when employing GPT-4 as an automated evaluator is its potential bias towards outputs produced by GPT-4V. To address this concern, we created a test set composed of instances GPT-4V generated incorrectly. On this test dataset, the GPT-4 evaluator with instance-specific rules achieves an accuracy of $97.9\\%$ , surpassing the human evaluator's accuracy of $93.6\\%$ . It suggests that GPT-4 with instance-specific rules demonstrates superior accuracy over human evaluators in pinpointing inaccuracies in GPT-4V's responses.", + "bbox": [ + 212, + 239, + 787, + 347 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Case Study for GPT-4V", + "text_level": 1, + "bbox": [ + 214, + 366, + 464, + 382 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Since GPT-4V is the current state-of-the-art, we report a detailed error analysis for it on AutoEval-Video. Here is a list of typical hard cases.", + "bbox": [ + 212, + 388, + 785, + 420 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Understanding Temporal Features. In many cases, simply looking at a couple of isolated frames isn't enough to answer the question. Take instance 200 in Tab. 4, with its moving camera revealing the layout of a house. The model needs to understand the temporal ordering of frames to infer the house's spatial design to answer \"When I'm at the entrance, how do I get to the balcony that has three decorations on the table?\" Moreover, in instance 305 in supplementary material, the speed at which things change (like \"in an instant\") is a temporal feature. Most vision-language models, including GPT-4V, struggle to extract temporal features from many continuous frames.", + "bbox": [ + 212, + 422, + 787, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Connecting Multiple Frames. Many incorrect responses are due to GPT-4V tending to reason on a single frame, ignoring the connection between frames. In Tab. 4, instance 43 demonstrates this fact. Another interesting example is instance 116 in supplementary material. In describing a sequence of actions, GPT-4V mentions a certain action many times because it appears in many frames, although the correct answer should do it just once.", + "bbox": [ + 212, + 560, + 787, + 651 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "GPT-4V also struggles to reason between scenes. For example, in instance 112 of Tab. 4, the first scene features a deer in third-person view gradually opening its eyes. The second scene shows an approaching train from the deer's first-person view. To answer \"Why does this orange animal have its eyes wide open?\", the model must connect these two events, but GPT-4V failed. There are more examples in supplementary material supporting this observation.", + "bbox": [ + 212, + 652, + 787, + 744 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moving Objects. Another group of hard cases involves moving objects, such as comparing the speed of two cars. As shown in Tab. 4, instance 191 records a police car overtaking a firefighting truck. However, GPT-4V perceives all frames as semantically similar and is unable to identify the faster vehicle.", + "bbox": [ + 212, + 746, + 787, + 806 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Overly General Responses. GPT-4V sometimes answers questions in a general context rather than addressing the specific content of given videos. For", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA", + "bbox": [ + 217, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Question: When I am at the entrance, how can I reach the balcony that has three decorations on the table? [ID:200]", + "bbox": [ + 218, + 176, + 777, + 188 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/de694994e0badecad62fd255f0f716c58d338507bc6bac40d3ba88cc3b838f2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 189, + 287, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/16545f4e842034deeb652a9ceaec13a59f268bf3dcb7d1b0899ab513f22e8dc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 189, + 357, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/90b60928e53313c8d0619fe373c657667973d12255d7fc0d63a285017ea7666e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 189, + 426, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/693cbdc5aa6edcfcb3c9cc79b8fcc85b68848fe9abf51d5c3e13c11043d9ae4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 189, + 496, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/76b9dc500c776af4f313bf377db02a44377894051bd5d50603b9ffef4a2372c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 189, + 568, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/62c56d4401c1627b4d3299b5e67cdaeb590381516be9a9ed18c5a06f307e1438.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 189, + 638, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4cd5cdb4c8a9d94a7c9281a9c9ce262491165e780b5e2399f6fe0c0244f10249.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 189, + 709, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9c4517d63728cc72a83b0d9f477867d410f1d32a0b423e66d1d458d0b1b7f8e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 189, + 717, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e6aa00eb7bf05b0644e761b55d9dbeac352304791139424ca1c28ebec90fb757.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 189, + 787, + 229 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sample Answer: Upon entering through the door, you can first turn left into the kitchen, and then turn right. At this point, you are facing the target balcony. All you need to do is continue straight, open the balcony door, and step outside.", + "bbox": [ + 217, + 236, + 777, + 266 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "GPT-4V: To reach the balcony with three decorations on the table, enter the house through the main entrance, proceed straight ahead through the living room, and then turn left to go through the double doors leading to the outdoor covered patio area.", + "bbox": [ + 215, + 270, + 777, + 301 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Question: What action did the third parrot from the left perform?", + "bbox": [ + 218, + 306, + 547, + 318 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "[ID:43]", + "bbox": [ + 740, + 308, + 777, + 318 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c18c133a1fd93831c7bb94d7611875a9ec8b5c3cc148bb567e880e983c8ad83a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 321, + 287, + 361 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d1e5ef6631b8f54a2084a449a94983f1b408f0f5dde36aae7512fbe06b79890c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 321, + 357, + 361 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/825c4162f43e1aa204b7a102844f6176590716da20a135b03cf4b4827adaa37b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 321, + 426, + 361 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c3d61df45c7df4d93b31d40974183a134ab5053d11c36750b49d2081ac417741.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 321, + 496, + 361 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d40e0f4e8153f40eb80eacb1198afef9aa55b5fcf937d117e8d9366199f2f3d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 321, + 568, + 361 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a6e61392e7a0347ddd977398cfe318ca0d55ff5333ec457827d2fc438a397037.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 321, + 637, + 361 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c1ff13fce68bef9f0d5300610a6a3118c0aa8fa231ed08f2cef227c65630d15a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 321, + 707, + 361 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/db818d7fee07e35b22a3ac1ed3e117d3ccd9902663381d02d1a73875e2e4f820.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 321, + 779, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sample Answer: It flew into the person's hand.", + "bbox": [ + 217, + 367, + 457, + 378 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "GPT-4V: The third parrot from the left appears to be perched on a hand and interacting with a person. It looks like the parrot might be engaged in some sort of play or social interaction with the individual.", + "bbox": [ + 215, + 381, + 777, + 402 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Question: Why are the eyes of this orange animal wide open?", + "bbox": [ + 218, + 409, + 524, + 421 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "[ID:112]", + "bbox": [ + 735, + 410, + 777, + 420 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ddccb5b3fcd7a42577a4e7fb6d289ec0567dc2fe6957646cd0e5c903e2d42b22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 422, + 287, + 460 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b3fce0b1c6e040f5fccb774f4cf55d765c6d0984fc33589df1ff9e56c61bc5c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 422, + 357, + 460 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/76378cdbae2f59b234c91ae3b573031ac71049d8d63fbdce26fbda8fab9a39e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 422, + 426, + 460 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1e44f5dd026a6cbac11f78e79a3e736a598da582b1f1708c54daae0958a9b003.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 422, + 496, + 460 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b1212bcec6ee662d652a401314bb0718a6ab042d1c73427517f901b1bbcc6164.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 422, + 568, + 462 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b77219fc07699635591815e78e02f785c8a66b5d685a3aed264d0ca71ae107f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 422, + 638, + 462 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/36d21ad8274bcf3b4f6d2193456e8d16ce1886626781beba4fdc3119ba1c8097.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 422, + 707, + 462 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c759aa760f4a9e3bcb6e354b2bd8c6b93f27b60384b8a61bc610006026add116.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 422, + 779, + 462 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sample Answer: Because the animal saw a train approaching, the white creature on the tracks was in danger.", + "bbox": [ + 217, + 468, + 759, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "GPT-4V: The orange animal appears to be a stylized or animated character, possibly showing surprise, alarm, or excitement. Wide open eyes in animation are often used to convey strong emotions or reactions. This could be in response to something happening in the scene or to depict a particular personality trait of the character. Without additional context, it's hard to specify exactly why, but wide eyes are a common visual cue for heightened emotions in animation.", + "bbox": [ + 215, + 482, + 777, + 534 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Question: In the video, which car is moving at a slower pace?", + "bbox": [ + 218, + 540, + 524, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "[ID:191]", + "bbox": [ + 735, + 541, + 777, + 551 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5325daf7ff6016eb239556c6dd2dbb01f9dea11817078bb9200cf8cc629f062a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 553, + 287, + 593 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2e757ee2c41f97a227a6fa2ba2fd6687e0c4d7a40b2f8e1d2a8f73e976af8d52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 553, + 357, + 593 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c7ac256d89a3fc030ae40e5549c9cafa61ffca8e5fb32a9a54eda8dc881c0a31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 553, + 426, + 593 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7c08fa2d1c1010e5c5f151f4826a16b733d3d49a4aeb2e0ab855897968caab59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 553, + 496, + 593 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9e470c6a282ed81326b2d70a69719f9137e5d970749efabb7af2d7d4608a75eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 553, + 568, + 593 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/603bcd71f7bce2763da91383ce7773468b2591b0df6663bc9fd00b8228671190.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 553, + 638, + 593 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a4da0c08eab688ab9b636ce88130955c164b007488dfe57b157508519f9fd1f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 553, + 707, + 593 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6b3a95950591d1df9ae4f8653502c0748403e2c976042048110133c88475cbec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 553, + 779, + 593 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sample Answer: The police car.", + "bbox": [ + 217, + 599, + 383, + 609 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "GPT-4V: I'm unable to view videos or discern the speed of objects within them. The image you've provided is static, so it doesn't convey the speed of the cars. If you have a specific question about the details within the image, feel free to ask!", + "bbox": [ + 215, + 614, + 777, + 643 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Question: Why are the audience members standing up?", + "bbox": [ + 218, + 652, + 495, + 662 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "[ID:97]", + "bbox": [ + 740, + 652, + 777, + 662 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/88c421d77bf1c152a7f1d9a52bc1a3e9317747abdf27384daad21b1cf97396c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 664, + 287, + 705 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fd30662b93930e5b501518d617b6c64d1d6e05e3f9b7bf3d6920fd0ad5cd1697.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 664, + 357, + 705 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/534329cc39713f0d03e1015aa93a11d8130eefd89844af917a57a2c8f3bc4d7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 664, + 426, + 705 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/44cb96792b1f03b84cf1bf26bf8e73de1df72c13be6944f7f52b999ac0fc5744.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 664, + 496, + 705 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e6a0b04593e19a1aa404b9075180cefa2956a3812f40ec16e38da64b67926430.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 664, + 568, + 705 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/75541d58c1efc202c91601f86d38c575e44fb7fa91e7163ba30e6491bf151a14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 664, + 638, + 705 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/92356d737e3ebf918b514d9da2d2c21eeb19626b8c4163680cffd29c4e080d79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 664, + 707, + 705 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3c828c0869d0b90ef6c555aa21326cb83e4de67c644d6a47d4b56fcd1faa3878.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 664, + 779, + 705 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sample Answer: The player in the blue team uniform scored.", + "bbox": [ + 217, + 710, + 524, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "GPT-4V: The audiences in sports video games, like the one depicted in the images, often stand up in response to exciting moments or plays. This is a way to simulate real-life reactions of fans during intense or critical moments in a game, such as a close score, a game-winning shot, or a spectacular play. The game developers include these reactions to enhance the realism and immersion of the game experience for players.", + "bbox": [ + 215, + 724, + 777, + 767 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Examples of Incorrect Responses from GPT-4V.", + "bbox": [ + 279, + 779, + 720, + 792 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Chen, Y. Lin, Y. Zhang, and W. Huang", + "bbox": [ + 269, + 114, + 558, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Question: What was the score of the team in black in the video footage?", + "bbox": [ + 218, + 146, + 547, + 156 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/15541733488d8b987a2402afff3a6bf9356d809a24ae1a3848816607640d439f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 156, + 285, + 196 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f0bfaa4faa2fcf376899fde15fbf44c6d6264ba9a27ae627dfe91cb1cbfbce3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 156, + 357, + 196 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e2452a077ddd57bba54c4c11e913b0f6eb3e1dde6cab383ea7a484e653031f93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 156, + 428, + 196 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/72e6a56aa63819866f901b36b6fffec58663dca3fb9f5bfd161528625097c6e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 156, + 498, + 196 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eb71f5d6acead61ed1d57dcea26a0a1c1e89add7d339dc5d0eea3a197a647ac4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 156, + 568, + 196 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ddfb573517526e7564a72120435e5d29ca778dc91d8535fb68574fc14856e4b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 156, + 638, + 196 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/50020e358c933046a906b8d9a7e7971767b62401fdab3620e750c4fb4d6a9f1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 156, + 707, + 196 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/34314a702dfa8f485a94238713240575cccd454cc7fa8e61cc9167064d7b5a08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 156, + 779, + 196 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sample Answer: The team in black scored eight points.", + "bbox": [ + 218, + 200, + 483, + 212 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Human: The black team scores 3 points in the video.", + "bbox": [ + 218, + 215, + 465, + 224 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Question: How did the character detect the enemy?", + "bbox": [ + 218, + 231, + 455, + 239 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f2a604066ecb847ecd92970a7b806752da3315c439359e848db7983de9e95499.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 239, + 287, + 281 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/43e7b3734abcdf9fb359487fcf8a686a6f486c5fa707dfc3f3b0ea80ca8dc84e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 239, + 357, + 281 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/aa2884480050e99c2bc0d93097873a8b05625c6e882e64e532f4ff000fe02a45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 239, + 428, + 281 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c2951fa477b119e619251b2d1be3922daefc9668daea3137d9268d158c24185d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 239, + 498, + 281 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/35df2db282d514bcee5cb10211f11a5eebe24a5226266127d760653cb86a46da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 239, + 568, + 281 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7d881cb3e4a1e39ebd0425df0799fef23c6c5426817587ad6c3541cb1c9ec22c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 239, + 638, + 281 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a4ae9af5c50fb3b72be868af5d1d6dde9f1878a5cc3ccc69cff4bb09b02b9f8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 239, + 707, + 281 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/555e7f9ed738d6baaa749dfb0762f167189e1d37f79dd1a06066cc07887cda4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 239, + 779, + 281 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sample Answer: The enemy shot the character so the enemy's location could be known.", + "bbox": [ + 218, + 286, + 638, + 297 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Human: The character used the scope on the sniper rifle to clearly observe distant enemies.", + "bbox": [ + 218, + 299, + 633, + 309 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Question: Why did this person initially appear stunned and then suddenly burst into laughter?", + "bbox": [ + 218, + 315, + 645, + 325 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "[ID:133]", + "bbox": [ + 738, + 315, + 777, + 325 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/03b370152bbf7663cb168b3f51b4f2cec54e87530bd949c0c167c97396e56243.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 325, + 287, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/23f7380e3094b2e9bc213cbeca3e70a9a638267b6dd00bb2dfd9d6d4ec216581.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 325, + 357, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7d45efc847118d000543d1fc8cd0a73fd9207a519017acf1091cabb392f0ecb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 325, + 428, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dfc49fe608f1f191956eda529d9628383199b8186aad87641423131facd8747b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 325, + 498, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5ecfe81cf6be51d2725dc3fd27d6006a2137b793276dcac198fd4b3485f078fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 325, + 568, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/06290bf9bd3a6e29b93e2ed0bbb1069707f7e96ef51accb20eea7db060085d06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 325, + 638, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e3b7ff46583e97b391adee657884a3b31759501db2038d711486e461af036a03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 325, + 707, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/26923a9240379e716d67fa87e264ba5cf447ba304d461daf1e75998860c23a4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 325, + 779, + 364 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sample Answer: Because the person threw grenades to kill the enemy but knocked a teammate down accidentally.", + "bbox": [ + 218, + 369, + 759, + 381 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Human: Because she didn't expect that the bomb she threw would explode and harm someone.", + "bbox": [ + 218, + 383, + 656, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Hard cases for humans due to insufficient external knowledge.", + "text_level": 1, + "bbox": [ + 228, + 406, + 769, + 420 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "instance, in instance 97 shown in Tab. 4, GPT-4V answers \"Why are the audience members standing up?\" in a general way, even though there is an obvious and concrete reason. Similarly, instance 197 gives a general guide for operating a coffee machine even though the video demonstrates a specific one. It implies that GPT-4V tends to rely on its strong language model to produce a safe answer, sometimes ignoring key details in the video.", + "bbox": [ + 212, + 438, + 784, + 529 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Hard Cases for Humans", + "text_level": 1, + "bbox": [ + 215, + 551, + 459, + 565 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Humans achieve an accuracy rate of $72.8\\%$ on AutoEval-Video. We conduct an in-depth analysis of these cases and discover that there are two major hard cases for humans: questions related to external knowledge and questions requiring accurate description and summarizing ability.", + "bbox": [ + 212, + 578, + 784, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Questions Related to External Knowledge. Some questions in AutoEval-Video require a relatively high level of mastery of external knowledge in specific domains such as sports and games, resulting in human difficulties. For example, in Tab. 5, instance 38 shows a misunderstanding due to human participants not knowing basketball scoring rules. Similarly, instance 133 requires explaining a game streamer's laughter, an easy task for someone familiar with the game but confusing for others. However, considering the cognitive level of the general public, these shortcomings are reasonable and expected.", + "bbox": [ + 212, + 641, + 784, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Accurate Description and Summarizing Ability. Questions requiring detailed, step-by-step video descriptions are more challenging for humans compared to semantic reasoning problems due to their complexity and precision. For instance, as shown in Tab. 6, in instance 173, human makes careless mistakes by ignoring necessary CPU installation steps shown in the video. In instance", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "143, errors occur due to insufficient observation, misidentifying corn kernels as soybeans. This phenomenon is common in many step-by-step descriptive tasks.", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Question: How can you prepare a dish using these ingredients and tools as demonstrated in the video? [ID:143]", + "bbox": [ + 218, + 183, + 777, + 193 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2f3a7d63bc41154e6550ba9cc5383769134b66aa16c2eb85f9fbdeeb3dad4da6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 194, + 777, + 234 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Sample Answer: Cut the potatoes, beets, purple cabbage, carrots, mushrooms, and bell peppers, put them in the steamer with the corn kernels, cut the gluten into strips, heat the steamer on the gas stove, and put the gluten into the air fryer.", + "bbox": [ + 215, + 239, + 777, + 271 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Human: The kitchen utensils needed for preparation include a cutting board, a chef's knife, a small steamer pot, and an air fryer. The preparation process for the ingredients is as follows: 1. Cut the potatoes into small chunks. 2. Slice the leafy greens into segments. 3. Chop the purple cabbage. 4. Slice the carrots into thick pieces. 5. Halve the white mushrooms. 6. Slice the bell peppers. 7. Prepare a small bowl of soaked soybeans. 8. Cut the blanched beef pieces into strips. Finally, Place the vegetables from steps 1-7 into the steamer pot, and place the beef strips from step 8 into the air fryer.", + "bbox": [ + 215, + 275, + 777, + 335 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Question: How to install a CPU on a motherboard as demonstrated in the video? [ID:173]", + "bbox": [ + 218, + 342, + 777, + 352 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4a8dbd0856ce054046a3c41e3d8e554ab80bfa799f73e02ec3179b989064722a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 352, + 777, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Sample Answer: According to the clip, here is how to install a CPU on a motherboard: 1. Lift the metal arm next to the CPU socket. 2. Look for a small triangle indicator on the corner of the CPU and the socket and align them properly. 3. Hold the CPU by the sides and insert the CPU into the CPU socket on the motherboard. 4. Lower the arm down completely into its original position.", + "bbox": [ + 215, + 398, + 777, + 439 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Human: 1. Open the latch at the CPU location on the motherboard. 2. Place the side of the CPU with pins onto the corresponding location on the motherboard, aligning with the notches. 3. Close the latch.", + "bbox": [ + 215, + 443, + 777, + 464 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/300c393ea160633bb5c97d40f917e1928e8ff235ede155f73ada9d5d268f0c1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 470, + 777, + 522 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Sample Answer: First, put the pillow at the head of the bed; Second, put the clothes that need to be washed on the floor, and put the clothes that do not need to be washed on the chair; Third, pile up all the debris except for sheets quilts, and pillows on the floor; Fourth, after arranging the sheets and quilts, place the doll bear on the side near the head of the bed.", + "bbox": [ + 215, + 527, + 777, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Human: The order in which this person organizes the room is as follows: 1. Neatly arrange multiple pillows at the head of the bed. 2. Gather multiple items left on the bed onto a chair. 3. Remove other items from the bed, such as a large teddy bear and a white cushion. 4. Straighten the bedsheet to make it look even. 5. Place the teddy bear back at the head of the bed.", + "bbox": [ + 215, + 571, + 777, + 612 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 6: Hard cases for humans related to descriptive ability.", + "bbox": [ + 266, + 626, + 733, + 640 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 215, + 684, + 531, + 700 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce AutoEval-Video, a novel benchmark for comprehensively assessing large VLMs in open-ended video question answering. Utilizing an LLM evaluator, models can be automatically and accurately evaluated on AutoEval-Video. In particular, GPT-4V achieves the highest accuracy of $32.2\\%$ , still far from human accuracy of $72.8\\%$ . Through an extensive case study, we identify several limitations of current models' capabilities in video understanding, such as temporal and dynamic comprehension, and overly general responses. These findings will shape our future research direction.", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Chen, Y. Lin, Y. Zhang, and W. Huang", + "bbox": [ + 269, + 114, + 557, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 217, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We would like to express our sincere gratitude to the reviewers of ECCV 2024 for their insightful and constructive feedback. Their valuable comments have greatly contributed to improving the quality of our work.", + "bbox": [ + 215, + 179, + 784, + 224 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 251, + 323, + 268 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675 (2016)", + "2. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, A., Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z., Zhang, J., Zhang, X., Zhang, Y., Zhang, Z., Zhou, C., Zhou, J., Zhou, X., Zhu, T.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)", + "3. Buch, S., Eyzaguirre, C., Gaidon, A., Wu, J., Fei-Fei, L., Niebles, J.C.: Revisiting the \"video\" in video-language understanding. In: CVPR. pp. 2917-2927 (2022)", + "4. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR. pp. 961-970 (2015)", + "5. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)", + "6. Fan, H., Lin, L., Yang, F., Chu, P., Deng, G., Yu, S., Bai, H., Xu, Y., Liao, C., Ling, H.: Lasot: A high-quality benchmark for large-scale single object tracking. In: CVPR. pp. 5374-5383 (2019)", + "7. Fu, J., Ng, S.K., Jiang, Z., Liu, P.: Gptscore: Evaluate as you desire. arXiv preprint arXiv:2302.04166 (2023)", + "8. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The \"something something\" video database for learning and evaluating visual common sense. In: ICCV. pp. 5842-5850 (2017)", + "9. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR. pp. 6904-6913 (2017)", + "0. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022)", + "1. Grunde-McLaughlin, M., Krishna, R., Agrawala, M.: Agqa: A benchmark for compositional spatio-temporal reasoning. In: CVPR. pp. 11287-11297 (2021)", + "2. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., et al.: Ava: A video dataset of spatiotemporally localized atomic visual actions. In: CVPR. pp. 6047-6056 (2018)", + "3. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + ], + "bbox": [ + 225, + 282, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA", + "bbox": [ + 218, + 113, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Hu, A., Chen, S., Zhang, L., Jin, Q.: Infomatic: An informative metric for reference-free image caption evaluation. arXiv preprint arXiv:2305.06002 (2023)", + "15. Huang, D.A., Ramanathan, V., Mahajan, D., Torresani, L., Paluri, M., Fei-Fei, L., Niebles, J.C.: What makes a video a video: Analyzing temporal information in video understanding models and datasets. In: CVPR. pp. 7366-7375 (2018)", + "16. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)", + "17. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV. pp. 706-715 (2017)", + "18. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: ICCV. pp. 2556-2563. IEEE (2011)", + "19. Lei, J., Berg, T.L., Bansal, M.: Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428 (2022)", + "20. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)", + "21. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "22. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)", + "23. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023)", + "24. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)", + "25. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL (2004)", + "26. Lin, Y.T., Chen, Y.N.: Llm-eval: Unified multi-dimensional automatic evaluation for open-domain conversations with large language models. arXiv preprint arXiv:2305.13711 (2023)", + "27. Liu, D., Qu, X., Hu, W.: Reducing the vision and language bias for temporal sentence grounding. ACM MM (2022), https://api_semanticscholar.org/CorpusID:251104851", + "28. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565 (2023)", + "29. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)", + "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "31. Liu, Y., Iter, D., Xu, Y., Wang, S., Xu, R., Zhu, C.: G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634 (2023)", + "32. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023)", + "33. OpenAI: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL. pp. 311-318 (2002)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Chen, Y. Lin, Y. Zhang, and W. Huang", + "bbox": [ + 269, + 114, + 557, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Pătrăucean, V., Smaira, L., Gupta, A., Continente, A.R., Markeeva, L., Banarse, D., Koppula, S., Heyward, J., Malinowski, M., Yang, Y., et al.: Perception test: A diagnostic benchmark for multimodal video models. arXiv preprint arXiv:2305.13786 (2023)", + "36. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)", + "37. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)", + "38. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)", + "39. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR. pp. 4566-4575 (2015)", + "40. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y.F., Wang, W.Y.: Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In: ICCV. pp. 4581-4591 (2019)", + "41. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. NeurIPS 35, 24824-24837 (2022)", + "42. Wu, Y., Lim, J., Yang, M.H.: Online object tracking: A benchmark. In: CVPR. pp. 2411-2418 (2013)", + "43. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR. pp. 5288-5296 (2016)", + "44. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023)", + "45. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)", + "46. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)", + "47. Yu, Z., Zheng, L., Zhao, Z., Wu, F., Fan, J., Ren, K., Yu, J.: Anetqa: A large-scale benchmark for fine-grained compositional reasoning over untrimmed videos. In: CVPR. pp. 23191-23200 (2023)", + "48. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)", + "49. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)", + "50. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)", + "51. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754 (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 797 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA", + "bbox": [ + 218, + 113, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_model.json b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..33c8f1d339eef72cb88d463add1042bc64c3842e --- /dev/null +++ b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_model.json @@ -0,0 +1,4018 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.773, + 0.21 + ], + "angle": 0, + "content": "AutoEval-Video: An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.238, + 0.763, + 0.255 + ], + "angle": 0, + "content": "Xiuyuan Chen\\(^{1,*}\\), Yuan Lin\\(^{2,\\dagger}\\), Yuchen Zhang\\(^{2,\\dagger}\\), and Weiran Huang\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.264, + 0.778, + 0.297 + ], + "angle": 0, + "content": "1 MIFA Lab, Qing Yuan Research Institute, SEIEE, Shanghai Jiao Tong University \n2 ByteDance Research" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.327, + 0.744, + 0.646 + ], + "angle": 0, + "content": "Abstract. We propose a novel and challenging benchmark, AutoEval-Video, to comprehensively evaluate large vision-language models in open-ended video question answering. The comprehensiveness of AutoEval-Video is demonstrated in two aspects: 1) AutoEval-Video constructs open-ended video-questions across 9 skill dimensions, addressing capabilities of perception, comprehension, and generation; 2) AutoEval-Video contains newly collected videos that cover over 40 distinct themes. To efficiently evaluate responses to the open-ended questions, we employ an LLM-based evaluation approach, but instead of merely providing a reference answer, we annotate unique evaluation rules for every single instance (video-question pair). To maximize the robustness of these rules, we develop a novel adversarial annotation mechanism. By using instance-specific rules as prompt, GPT-4, as an automatic evaluator, can achieve a stable evaluation accuracy of around \\(97.0\\%\\), comparable to the \\(94.9\\% - 97.5\\%\\) accuracy of a human evaluator. Furthermore, we assess the performance of eleven large vision-language models on AutoEval-Video. Among them, GPT-4V(ison) significantly outperforms other models, achieving an accuracy of \\(32.2\\%\\). However, there is still substantial room for improvement compared to human accuracy of \\(72.8\\%\\). By conducting an extensive case study, we uncover several drawbacks of GPT-4V, such as limited temporal and dynamic comprehension, and overly general responses. Code is available at https://github.com/Xiuyuan-Chen/AutoEval-Video." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.669, + 0.377, + 0.685 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.791 + ], + "angle": 0, + "content": "Generative large vision-language models [2, 5, 21, 22, 29, 30, 32, 48] have exhibited the considerable potential to address fundamental challenges in video understanding, temporal and causal comprehension [3]. Furthermore, these models can deal with more complex open-ended video tasks, surpassing traditional specialized video tasks, like video captioning [17, 40, 43], action classification [8, 12, 36], and object tracking [6, 42]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.798, + 0.611, + 0.812 + ], + "angle": 0, + "content": "* Work done during an internship at ByteDance Research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "† Correspondence to Yuan Lin, Yuchen Zhang, and Weiran Huang (weiran.huang@outlook.com)." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.798, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.56, + 0.129 + ], + "angle": 0, + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.396 + ], + "angle": 0, + "content": "
SkillExample of Video-Question
Dynamic PerceptionThe video depicts various ways of relaxing and de-stress.\nQuestion: Please List the three relaxation techniques from the video in order.
State Transitions PerceptionThe video captures the process of a chemistry experiment.\nQuestion: What is the change in the color of the liquid after blowing?
Camera Movement PerceptionThe camera's position in the video undergoes a series of adjustments.\nQuestion: What changes have been made to the position of the camera used for filming this video?
Explanatory ReasoningThe camera records a basketball player sinking a shot and flashing a V sign to the camera.\nQuestion: Why is the individual wearing blue jeans in the video making that particular gesture?
Counterfactual ReasoningA purple animal character in the cartoon follows the guidance of another character and falls into a trap.\nQuestion: What action could the purple animal take to avoid the consequences shown in the video?
Predictive ReasoningThe video records a person putting groceries into a cupboard and preparing to close the cupboard door (but without any depiction of the door-closing action).\nQuestion: What the person will do next?
Comparison ReasoningThe video sequentially depicts several individuals drinking water, where one person exhibits discomfort while the others appear normal.\nQuestion: How does the clothing of the uncomfortable person differ from that of the others?
Reasoning with External KnowledgeThe video records a basketball game, in which one team scored two three-pointers and a two-pointer.\nQuestion: In the video footage, what is the score of the team wearing black?
DescriptionThe video highlights two additional features of a vacuum cleaner product.\nQuestion: Sequentially describe the two extra features of the vacuum cleaner.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.406, + 0.787, + 0.433 + ], + "angle": 0, + "content": "Table 1: Skill Dimensions and Corresponding Examples for Video-Questions in AutoEval-Video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.787, + 0.649 + ], + "angle": 0, + "content": "Effectively evaluating large vision-language models is crucial for improving their performance, as it identifies areas for improvement and enables the comparison and investigation of various models. Several benchmarks [6,8,10,12,17, 35,36,40,42,43] have been developed to evaluate capabilities of specific tasks in video understanding. Nonetheless, they fail to assess model performance in a comprehensive manner. To close this gap, recent studies [10,20,35] aim to build benchmarks covering a wide range of skills. These benchmarks primarily consist of multiple-choice question-answers to facilitate effective evaluation. However, in comparison to open-ended question answering, multiple-choice question answering presents two key limitations: 1) they simplify the tasks, and 2) they do not accurately reflect real-world scenarios. For instance, in a video-based dialogue, responses should be generated directly, rather than making a selection from a set of options." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.803 + ], + "angle": 0, + "content": "In this paper, we introduce AutoEval-Video, a comprehensive and challenging benchmark designed to evaluate the video understanding capabilities of large vision-language models. AutoEval-Video comprises 327 challenging open-ended video-question instances that span across nine skill dimensions in Tab. 1, which address video-specific perception, comprehension, and generation skills. All videos in AutoEval-Video are newly collected from YouTube, including over 40 genres such as daily life, gaming, sports, and more. We also establish rigorous annotation standards to guarantee the difficulty and quality of our benchmark. These standards include minimizing single-frame bias [3, 15, 19] and language bias [9, 27], and setting objective questions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.804, + 0.787, + 0.835 + ], + "angle": 0, + "content": "Assessing responses to open-ended questions can be difficult, especially when examining outputs generated by LLMs. These responses may involve additional" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.146, + 0.786, + 0.403 + ], + "angle": 0, + "content": "information such as video description and chain-of-thought [41] analysis, diminishing the reliability of traditional automatic evaluation metrics, which typically involve comparing a given response to a reference answer, such as BLEU [34], ROUGE [25] and CIDEr [39]. To address this issue, we employ an LLM-based evaluation approach [7, 31, 45, 49]. An LLM evaluator comprises a strong LLM and its prompting strategy. Existing works [26, 46, 50] design a unified prompt for specific tasks. However, due to the diversity of videos and questions in AutoEval-Video, a single unified prompt is insufficient for accurate evaluation. To tackle this challenge, we meticulously annotate distinct evaluation rules for each instance instead of solely providing a reference answer. To improve the precision of these rules, we develop a novel adversarial annotation mechanism. A red team is engaged to devise unaddressed corner cases that the rules may not cover, assisting the annotation team in refining them. The final rule is ultimate when the red team cannot discover any further corner cases that break the rules. After intensive annotation, using instance-specific rules as prompt, GPT-4 can serve as an automatic evaluator, achieving an evaluation accuracy of \\(97.0\\%\\), comparable to the single human accuracy of \\(97.5\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.406, + 0.786, + 0.556 + ], + "angle": 0, + "content": "Leveraging AutoEval-Video, we have conducted a comprehensive evaluation of eleven prominent large vision-language models, which fall into two categories: ImageLLM and VideoLLM. Our assessment reveals that GPT-4V(ison) significantly surpasses other models. Utilizing 8 randomly sampled frames as input, GPT-4V achieves an accuracy of \\(22.2\\%\\), compared to \\(72.8\\%\\) accuracy of humans, indicating considerable potential for improvement. Moreover, we discover that by increasing the input to 16 frames, GPT-4V's performance markedly improves to \\(32.2\\%\\). Finally, we perform an in-depth case study for GPT-4V, identifying several critical challenges that demand further investigation, such as limited temporal and dynamic understanding, and overly general responses." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.587, + 0.386, + 0.603 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.626, + 0.786, + 0.747 + ], + "angle": 0, + "content": "Large Vision-Language Models. The remarkable success of LLMs has been accompanied by significant progress in large vision-language models. Recent works [2, 5, 21, 29] improve the visual perception and cognition capabilities by utilizing the strong understanding and generation abilities of LLMs. In addition, some research [22, 32, 48] further incorporates video inputs and leverages the extensive potential of LLMs for video understanding. The considerable progress necessitates more challenging and comprehensive benchmarks to evaluate the video comprehension capabilities of large vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.749, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Benchmarks for Video Understanding. Several video benchmarks have been employed to assess model performance in various areas, such as action recognition [11, 12, 47], video captioning [13, 14, 43], and action classification [6, 8, 16, 18, 36, 42]. However, these benchmarks do not provide a comprehensive assessment of a model's capabilities in video understanding. To address this issue, some general benchmarks [20, 35] have been established, developing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.56, + 0.13 + ], + "angle": 0, + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.146, + 0.512, + 0.156 + ], + "angle": 0, + "content": "Q: What are the steps to prepare this dish as shown in the video?" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.156, + 0.779, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.49, + 0.207, + 0.515, + 0.215 + ], + "angle": 0, + "content": "Rules" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.217, + 0.776, + 0.238 + ], + "angle": 0, + "content": "[Video Description]: In the video, a person first uses a knife to cut tofu into cubes on a cutting board. Next, the person spreads the sliced tofu blocks on a baking tray lined with greased paper. Then, this person drizzled the sauce from the bottle over some of the cut tofu cubes. Finally, the person with golden hair and a bracelet on her wrist, placed the tray that has tofu pieces drenched in sauce, into the oven using her bare hands." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.238, + 0.475, + 0.245 + ], + "angle": 0, + "content": "[Video Question]: What are the steps to prepare this dish as shown in the video?" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.245, + 0.586, + 0.252 + ], + "angle": 0, + "content": "Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.253, + 0.775, + 0.266 + ], + "angle": 0, + "content": "1. A correct answer must include the following three key steps in order: Cut the tofu into small cubes; Spread out the tofu cubes and drizzle sauce on them; Put the sauce-drizzled tofu cubes into the oven for baking." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.266, + 0.729, + 0.274 + ], + "angle": 0, + "content": "2. The answer to the description of the shape of tofu blocks can be anything similar to 'block'. It doesn't necessarily have to be described as a perfectly shaped cube." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.274, + 0.567, + 0.281 + ], + "angle": 0, + "content": "3. Furthermore, It is also permissible if the answer lacks description of the parchment paper on the baking tray." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.281, + 0.775, + 0.294 + ], + "angle": 0, + "content": "4. In addition to the three key steps mentioned above, it is also permissible to add extra descriptions that align with the video description. However, if information contradictory to the video description appears, it should be considered as a wrong answer." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.294, + 0.693, + 0.301 + ], + "angle": 0, + "content": "5. The response should contain correctly spelled words. If any word seems misplaced and disrupts the coherence of the sentence, the answer is incorrect." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.301, + 0.345, + 0.308 + ], + "angle": 0, + "content": "[Answer to be judged]: " + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.308, + 0.297, + 0.315 + ], + "angle": 0, + "content": "[Form of the result]:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.315, + 0.775, + 0.329 + ], + "angle": 0, + "content": "Please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above." + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.334, + 0.335, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.294, + 0.349, + 0.338, + 0.356 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.357, + 0.402, + 0.369 + ], + "angle": 0, + "content": "1. Cut the tofu into pieces and then place it on a wooden board." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.37, + 0.357, + 0.377 + ], + "angle": 0, + "content": "2. Drizzle the sauce over some tofu slices." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.377, + 0.402, + 0.392 + ], + "angle": 0, + "content": "3. Put on gloves and place the baking tray with tofu pieces into the oven." + }, + { + "type": "list", + "bbox": [ + 0.228, + 0.357, + 0.402, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.351, + 0.523, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.357, + 0.582, + 0.376 + ], + "angle": 0, + "content": "1. Cut the tofu into small cubes and spread them out. \n2. Place the tofu cubes on a baking tray and drizzle with sauce." + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.377, + 0.581, + 0.393 + ], + "angle": 0, + "content": "3. Put the sauce-drizzled tofu cubes into the oven to bake." + }, + { + "type": "list", + "bbox": [ + 0.419, + 0.357, + 0.582, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.35, + 0.709, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.357, + 0.775, + 0.371 + ], + "angle": 0, + "content": "1. Cut the white tofu into pieces, then place them in a baking tray." + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.371, + 0.741, + 0.378 + ], + "angle": 0, + "content": "2. Drizzle sauce over some of the tofu pieces." + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.378, + 0.771, + 0.385 + ], + "angle": 0, + "content": "3. Place the baking tray with the tofu pieces in the oven." + }, + { + "type": "list", + "bbox": [ + 0.6, + 0.357, + 0.775, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.396, + 0.327, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.426, + 0.4, + 0.44 + ], + "angle": 0, + "content": "The answer provided is incorrect for the following reasons:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.441, + 0.401, + 0.469 + ], + "angle": 0, + "content": "1. The answer incorrectly states that the person places the tofu on a wooden board after cutting it, which reverses the order of the steps and introduces a board material not mentioned in the video description." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.469, + 0.401, + 0.483 + ], + "angle": 0, + "content": "2. The sauce is misspelled as \"sauce,\" which is a spelling error." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.483, + 0.401, + 0.504 + ], + "angle": 0, + "content": "3. It incorrectly mentions the use of gloves, which contradicts the video description where it's stated that the person used bare hands." + }, + { + "type": "list", + "bbox": [ + 0.228, + 0.426, + 0.401, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.396, + 0.512, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.426, + 0.584, + 0.499 + ], + "angle": 0, + "content": "The [Answer to be judged] correctly includes the three key steps required by the rules: cutting tofu into cubes, spreading the tofu out and drizzling sauce on them, and putting the sauce-drizzled tofu into the oven to bake. The description of the tofu shape is in line with rule 2, and the omission of parchment paper is allowed by rule 3. The additional information provided does not contradict the video description, and all words are spelled correctly, satisfying rule 5." + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.396, + 0.7, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.426, + 0.776, + 0.455 + ], + "angle": 0, + "content": "The answer provided correctly states the three key steps in the order they appear in the video description: cutting the tofu into pieces, drizzling sauce on the tofu, and placing the tray in the oven." + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.455, + 0.776, + 0.498 + ], + "angle": 0, + "content": "The shape of the tofu is described as \"pieces,\" which aligns with the allowance for any similar description to \"block.\" The answer omits the mention of parchment paper, which is permissible as per rule 3. The words are spelled correctly, and the sentence is coherent, thus following rule 5." + }, + { + "type": "list", + "bbox": [ + 0.6, + 0.426, + 0.776, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.526, + 0.787, + 0.582 + ], + "angle": 0, + "content": "Fig. 1: Example Instance in AutoEval-Video and Automatic Evaluation Process. Each instance has three components: video, question, and rules. For each video and question, we establish specific and detailed evaluation rules. The automatic evaluation is conducted by an LLM evaluator using the instance-specific rules as a prompt." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.588, + 0.787, + 0.633 + ], + "angle": 0, + "content": "multiple-choice VQA to assess models' abilities across different aspects. Nevertheless, compared to open-ended question answering, multiple-choice question answering oversimplify tasks and may not reflect realistic instructions." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.635, + 0.787, + 0.711 + ], + "angle": 0, + "content": "In addition, existing benchmarks mainly employ two methods for video collection: 1) sourcing videos related to a specific task from web [4,16] or large-scale video datasets [1,11,47]; 2) recording new videos based on pre-defined scripts [35] or themes [10]. However, both approaches suffer from limited visual diversity, leading to a video distribution that inadequate represents real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.409, + 0.756 + ], + "angle": 0, + "content": "3 AutoEval-Video" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.779, + 0.341, + 0.792 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In AutoEval-Video, each instance comprises three components: video, question, and rules, as illustrated in Fig. 1. The construction involves three steps." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.218, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.237 + ], + "angle": 0, + "content": "First, to address the essential aspects of real-world multi-modal understanding, the uncovered areas of existing benchmarks, and the limitations of current models, we identify nine skill dimensions, as illustrated in Tab. 1. These dimensions encompass diverse skill areas (including perception, cognition, and generation) and various reasoning types (such as descriptive, explanatory, predictive, comparative, and counterfactual)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.784, + 0.3 + ], + "angle": 0, + "content": "Second, for each skill dimension, an annotation team gathers videos from YouTube1 and crafts relevant open-ended questions. It is critical to prevent potential single-frame bias and language bias while formulating these questions. To ensure the assessability of responses, the questions must retain objectivity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.785, + 0.408 + ], + "angle": 0, + "content": "Finally, after acquiring the video-question pairs, the annotation group develops evaluation rules for each instance. These rules establish precise conditions for considering whether a response is correct or wrong, involving precise criteria for recognizing seemingly accurate but flawed corner case responses, and vice versa. To ensure thoroughness, we employ a second annotation team, known as the \"red team\", to devise unaddressed corner cases. The final rule is ultimate when the red team cannot discover any further corner cases that break the rules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.785, + 0.456 + ], + "angle": 0, + "content": "As demonstrated above, the annotation of AutoEval-Video is sophisticated and labor-intensive. We construct a total of 327 instances. The average cost per instance amounts to $28." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.482, + 0.4, + 0.495 + ], + "angle": 0, + "content": "3.2 Skill Dimensions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.661 + ], + "angle": 0, + "content": "Video understanding encompasses two fundamental skills: perception and cognition. Different from image perception, video perception emphasizes temporal information, such as activity, state transitions, and camera movement. With regard to cognition, our benchmark considers reasoning based on temporal or spatial information, simultaneously spanning causal reasoning, comparison, and reasoning with external knowledge. In addition, considering that hallucination [24,28,37,51] is a prominent challenge for contemporary large vision-language models, we incorporate tasks requiring descriptive ability in our benchmark. In summary, we identify nine skill dimensions, as shown in Tab. 1, which will address the aforementioned capabilities." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.686, + 0.515, + 0.701 + ], + "angle": 0, + "content": "3.3 Video and Question Collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.714, + 0.784, + 0.744 + ], + "angle": 0, + "content": "Video Collection. We gather all videos from YouTube, taking into account the following considerations:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.757, + 0.784, + 0.803 + ], + "angle": 0, + "content": "- To minimize the probability of utilizing videos that have been employed in the model training process, we collect new videos from YouTube rather than employing well-known open-sourced video datasets." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "1 The videos collected from YouTube comply with Creative Commons License (https://support.google.com/youtube/answer/2797468)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.56, + 0.131 + ], + "angle": 0, + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "- We aim to construct a benchmark that incorporates real-world visual content so as to directly assess the capability of models in addressing realistic challenges. Thus, we choose to select videos from YouTube instead of filming new videos based on scripts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.237, + 0.788, + 0.297 + ], + "angle": 0, + "content": "In addition, we examine the video categories of our benchmark, which covers mainstream video genres in real-world. All video sources are guaranteed to possess a resolution of 720P or higher. More statistics of the videos can be found in Fig. 2." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.3, + 0.486, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.476, + 0.48, + 0.503 + ], + "angle": 0, + "content": "(a) Genre distribution of videos in AutoEval-Video" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.315, + 0.75, + 0.466 + ], + "angle": 0, + "content": "
StatisticsMeanMedian
Question Length (#words)10.9410.00
Rule Length (#words)302.95296.00
Video Shots (#Scenes)3.162.00
Video Duration (#Seconds)14.5810.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.472, + 0.744, + 0.489 + ], + "angle": 0, + "content": "(b) Statistics of AutoEval-Video" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Fig. 2: Statistics of AutoEval-Video. (a) Genre distribution of videos. (b) Statistics on question and rule lengths, number of scenes, and video duration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Question Collection. The questions of AutoEval-Video are meticulously designed by human annotators, avoiding templates or paradigms, to guarantee a diverse range of questions. Annotators may either initiate a question and then search for an appropriate video, or select a video first and craft a question afterward. To maximize dataset diversity, we annotate only one question per video. To obtain high-quality video-question pairs that can effectively diagnose models' video comprehension abilities, we propose three principles for constructing the benchmark: avoiding single-frame bias, reducing language bias, and formulating objective questions, as discussed below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Many video benchmarks suffer from single-frame bias [3, 15, 19], indicating that the information in an oracle frame is almost sufficient to deduce the answer. To address this issue, we instruct annotators to devise questions that require the consideration of multiple frames for successful completion. Questions that can be answered based on only one frame are deemed invalid." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Existing video-language benchmarks are affected by language bias [9, 27], suggesting that models can obtain the correct answers without seeing video content. To tackle this issue, we involve three additional human annotators who try their best to answer questions without viewing the videos. If any of the annotators can accurately answer the question, the instance will be discarded." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.15, + 0.438, + 0.16 + ], + "angle": 0, + "content": "Q: Why did the race car driver exit the vehicle?" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.16, + 0.297, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.16, + 0.365, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.16, + 0.434, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.16, + 0.503, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.16, + 0.571, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.16, + 0.639, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.16, + 0.707, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.16, + 0.785, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.363, + 0.214, + 0.417, + 0.222 + ], + "angle": 0, + "content": "Initial Rules" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.223, + 0.55, + 0.261 + ], + "angle": 0, + "content": "[Video description]: In a racetrack, a white race car flipped over and came to a halt, then caught fire. Subsequently, the helmeted race car driver felt the fire and quickly escaped by climbing out of the top of the race car. Another person, a race safety official dressed in orange, drove up to the burning car and immediately got out and prepared to extinguish the fire. [Video Question]: Why did the race car driver exit the vehicle?" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.262, + 0.549, + 0.277 + ], + "angle": 0, + "content": "Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules:" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.277, + 0.549, + 0.292 + ], + "angle": 0, + "content": "1. The main key point in the answer lies in \"race car caught on fire\". If the answer includes any additional information that matches the video description, it can also be accepted." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.293, + 0.549, + 0.308 + ], + "angle": 0, + "content": "2. The words in the answer must be spelled correctly. If a word does not make sense in the sentence, the answer is incorrect." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.309, + 0.549, + 0.323 + ], + "angle": 0, + "content": "3. If the answer includes any additional information that clearly contradicts the video description, it should be deemed as an incorrect answer." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.324, + 0.408, + 0.332 + ], + "angle": 0, + "content": "[Answer to be judged]: (Adversarial Answer)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.332, + 0.309, + 0.339 + ], + "angle": 0, + "content": "[Form of the Result]:" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.34, + 0.549, + 0.364 + ], + "angle": 0, + "content": "please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above." + }, + { + "type": "title", + "bbox": [ + 0.634, + 0.213, + 0.726, + 0.221 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.585, + 0.221, + 0.774, + 0.244 + ], + "angle": 0, + "content": "Because the race car driver's car caught fire, which might have been caused by the driver's improper operation." + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.248, + 0.692, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.584, + 0.278, + 0.775, + 0.357 + ], + "angle": 0, + "content": "The reason provided in the answer, \"improper operation by the driver\", adds speculative information not present in the video description. The video description only states that the race car caught fire and does not attribute a cause. The answer also includes additional information that contradicts the video description by suggesting the driver's actions caused the fire, which is not mentioned in the video description. Therefore, the answer should be deemed incorrect." + }, + { + "type": "title", + "bbox": [ + 0.468, + 0.393, + 0.535, + 0.402 + ], + "angle": 0, + "content": "Updated Rules" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.403, + 0.774, + 0.434 + ], + "angle": 0, + "content": "[Video description]: In a racetrack, a white race car flipped over and came to a halt, then caught fire. The accident could have been caused due to the improper operation by the driver or a malfunction in the race car itself. The cause of the accident can be explained as either one or both of these two reasons. Subsequently, the helmed race car driver felt the fire and quickly escaped by climbing out of the top of the race car. Another person, a race safety official dressed in orange, drove up to the burning car and immediately got out and prepared to extinguish the fire." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.435, + 0.447, + 0.441 + ], + "angle": 0, + "content": "[Video Question]: Why did the race car driver exit the vehicle?" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.442, + 0.626, + 0.449 + ], + "angle": 0, + "content": "Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules:" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.45, + 0.774, + 0.465 + ], + "angle": 0, + "content": "1. The main key point in the answer lies in \"race car caught on fire\". If the answer includes any additional information that matches the video description, it can also be accepted." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.466, + 0.648, + 0.473 + ], + "angle": 0, + "content": "2. The words in the answer must be spelled correctly. If a word does not make sense in the sentence, the answer is incorrect." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.474, + 0.707, + 0.48 + ], + "angle": 0, + "content": "3. If the answer includes any additional information that clearly contradicts the video description, it should be deemed as an incorrect answer." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.481, + 0.774, + 0.496 + ], + "angle": 0, + "content": "4. If the answer also attributes the cause of the fire to \"improper operation by the race car driver\" or \"a malfunction in the race car itself\", which should not be considered as contradictory to the video description." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.497, + 0.407, + 0.504 + ], + "angle": 0, + "content": "[Answer to be judged]: (Adversarial Answer)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.505, + 0.307, + 0.511 + ], + "angle": 0, + "content": "[Form of the Result]:" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.512, + 0.774, + 0.528 + ], + "angle": 0, + "content": "please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.548, + 0.787, + 0.602 + ], + "angle": 0, + "content": "Fig. 3: Adversarial Annotation for Rules. A red team develops adversarial answers that break the initial rules. Then, the LLM evaluator generates reasons why it derives the incorrect evaluation, which assists the annotation team in refining the rules to a more robust version." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.787, + 0.699 + ], + "angle": 0, + "content": "In our benchmark, all question-answer pairs should be assessed by a set of rules, requiring objective questions since evaluating subjective responses is challenging. For instance, a video depicting a man with a neutral expression, accompanied with a question, \"What is the man thinking about?\" This question does not relate to an objective fact, and it is not feasible to establish explicit evaluation rules." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.72, + 0.43, + 0.734 + ], + "angle": 0, + "content": "3.4 Annotation of Rules" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.743, + 0.787, + 0.789 + ], + "angle": 0, + "content": "For each instance in AutoEval-Video, we annotate a set of rules describing the criteria for determining the correctness of an answer. The rationale behind this approach is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "- All questions in AutoEval-Video are open-ended, rendering the enumeration of all reference answers infeasible. Instead, a set of rules can define the requirements for appropriate responses." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.56, + 0.129 + ], + "angle": 0, + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.147, + 0.788, + 0.236 + ], + "angle": 0, + "content": "- Modern LLMs, such as GPT-4, possess remarkable language understanding capabilities, allowing them to comprehend the provided rules and act as automatic evaluators. Unlike other works that use a single, unified prompt to assess all video-question responses [26, 46, 50], we annotate specific rules for each video-question pair, ensuring an in-depth and specific understanding of the criteria." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.784, + 0.336 + ], + "angle": 0, + "content": "The valid evaluation rules include a detailed description of the video content, questions about the video, key points of the correct answer, and the required output format, as illustrated in Fig. 1. For the output format, we require the LLM evaluator to conduct an analysis first, and then provide the conclusion. Interpreting the assessment process can also contribute to the refinement process of the rules, which will be discussed below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.565 + ], + "angle": 0, + "content": "To guarantee the thoroughness of the rules, we employ an adversarial annotation process (shown in Fig. 3) conducted by a separate group of annotators, known as the red team. Initially, the annotation team establishes initial rules denoted as \\( r_0 \\). For each \\( k \\)-th round of attack, the red team will endeavor to construct an answer, \\( a_k \\), causing erroneous judgments from an LLM evaluator (e.g. GPT-4) with the preceding rules \\( r_{k-1} \\). Upon obtaining an adversarial answer \\( a_k \\), the annotation team must update their rules to a refined version, \\( r_k \\). The iterative attack-refine process will continue until the red team cannot identify any adversarial answer. The red team may construct attack answers through various methods, such as: 1) directly writing one based on their understanding of the rules; and 2) prompting LLMs to generate adversarial answers. During the refinement process, the annotators can utilize the interpretations of the LLM evaluator, which will help to quickly figure out why the evaluator gets confused and facilitate the production of more accurate rules. On average, there are 3 attack-refine iterations in the construction of rules in AutoEval-Video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Furthermore, to enhance the motivation and engagement of both the annotation team and the red team in confrontation, we devised a reward mechanism. For each instance, the annotation team initially has a full reward, while the red team starts with zero reward. If the red team successfully discovers an adversarial answer, they will earn a reward, while the annotation team will lose a portion of their reward. As the number of iterations increases, there will be a corresponding increase in rewards for successful attacks, as well as higher penalties for the annotation team. This approach not only encourages the red team to target mature rules but also urges the annotation team to consistently refine their rules throughout every iteration, ensuring optimal results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.377, + 0.758 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.429, + 0.788 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Models. Utilizing AutoEval-Video, we assess the performance of eleven models, including six ImageLLMs (i.e., GPT-4V [44], Gemini-Ultra-Vision [38], LLaVA1.5 [29], Qwen-VL [2], InstructBLIP [5], and BLIP-2 [21]) and five VideoLLMs" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.218, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.783, + 0.223 + ], + "angle": 0, + "content": "(i.e., Video-ChatGPT [32], VideoChat, VideoChat2 [22], Video-LLaMA [48], and LLaMA-VID [23]). For each video, we uniformly sample 8 or 16 frames, resulting in two different input configurations. In particular, for ImageLLMs, we concatenate the 8 or 16 frames to create a single image as visual input (refer to supplementary material for the detailed concatenation method)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.784, + 0.286 + ], + "angle": 0, + "content": "LLM as an automatic evaluator. We adopt GPT-4 [33] as an automatic evaluator for AutoEval-Video. For each instance, GPT-4 generates an evaluation of the correctness of a given answer to be judged using the rules as prompts, as demonstrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.785, + 0.395 + ], + "angle": 0, + "content": "Human baseline. We recruit three human participants to answer questions in AutoEval-Video. The overall average accuracy is \\(72.8\\%\\). Through closely examining the mistakes made by humans, we identify two major challenges they face: 1) Questions related to external knowledge, such as unfamiliar games, may lead to incorrect answers from humans; 2) Questions requiring precise memory skills, such as recalling multi-step processes in a video, may cause a person to overlook one or more steps. More examples can be found in supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.285, + 0.397, + 0.714, + 0.596 + ], + "angle": 0, + "content": "
Model TypeModelAccuracy
8 Frames16 Frames
ImageLLMBLIP-2 [21]0.31.2
InstructBLIP [5]7.914.3
Qwen-VL [2]7.010.0
LLaVA-1.5 [29]8.58.2
GPT-4V [44]22.232.2
Gemini-Ultra-Vision [38]17.920.9
VideoLLMVideo-ChatGPT [32]6.74.9
VideoChat [22]13.410.9
Video-LLaMA [48]11.28.5
LLaMA-VID [23]9.88.0
VideoChat2 [22]0.30.6
Human (full videos)72.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.605, + 0.785, + 0.634 + ], + "angle": 0, + "content": "Table 2: Evaluation Results of Different Models on AutoEval-Video. We also provide human results after viewing full videos." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.675, + 0.536, + 0.691 + ], + "angle": 0, + "content": "4.2 Performances on AutoEval-Video" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.785, + 0.776 + ], + "angle": 0, + "content": "We compare 11 large vision-language models on Tab. 2. Findings show: 1) GPT-4V is much stronger than other models; 2) Among open-sourced models, VideoLLM generally outperforms ImageLLM; 3) GPT-4V falls significantly short of human-level, reflecting the challenge of our benchmark and a big headroom for further advancements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We find that open-sourced models suffer from three typical categories of bad cases. 1) Hallucination: the described objects or events are not present in the video, leading to wrong answers; 2) Alignment: since AutoEval-Video includes a diverse range of queries, it is often observed that the models' responses do" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.558, + 0.129 + ], + "angle": 0, + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.221 + ], + "angle": 0, + "content": "not align with the question; 3) Weak LLM: many models cannot get to the correct answer due to their limited language generation capabilities. We have provided examples for each category in the supplementary material. GPT-4V shows significant progress in these directions, but it still has fundamental issues, as discussed in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.788, + 0.317 + ], + "angle": 0, + "content": "Among open-sourced models, VideoChat achieves the highest score. Interestingly, GPT-4V correctly answers only \\(32.7\\%\\) instances where VideoChat is correct, indicating that VideoChat is stronger in certain domains. These domains are 1) traffic-related tasks; 2) comparison between dynamic objects; and 3) giving video-specific responses rather than general vague answers. In supplementary material, we provide examples for each of these three domains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Finally, we notice that 8 frames are insufficient for answering certain questions. Therefore, we evaluate models using 16 frames as input. Under this setting, GPT-4V significantly improves, reaching \\(32.2\\%\\) accuracy (examples provided in supplementary material). While longer inputs usually benefit ImageLLMs, VideoLLMs don't exhibit the same trend, possibly because they are explicitly trained in the 8-frame setting." + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.412, + 0.7, + 0.491 + ], + "angle": 0, + "content": "
GPT-4 PromptAccuracy of Evaluators
Models/Human ResponsesGPT-4V Responses
Instance-specific rules97.096.6
W/o adversarial annotation91.588.1
W/o rules (unified prompt)87.069.5
Human97.594.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.5, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Table 3: Ablation Study on Automatic Evaluators. We assess the evaluation accuracy of GPT-4 with instance-specific rules, initial rules, and unified prompts on responses generated by models/humans, as well as responses exclusively generated by GPT-4V. In addition, we present the evaluation accuracy of a single human." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.597, + 0.625, + 0.613 + ], + "angle": 0, + "content": "4.3 Accuracy of GPT-4 as Automatic Evaluator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.715 + ], + "angle": 0, + "content": "To validate GPT-4 as an automatic evaluator, we create a test dataset consisting of 200 randomly sampled (video, question, response) triplets, either model-generated or human-generated. Then we employed three human experts to evaluate each response, utilizing their collective decision as the ground truth. As shown in Tab. 3, with the evaluation rules, GPT-4 achieves \\(97\\%\\) agreement with the ground truth, comparable to \\(97.5\\%\\) of a single human expert." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.808 + ], + "angle": 0, + "content": "We perform an ablation study to investigate the necessity of our proposed way of annotation. As reported in Tab. 3, if we use the initial version of rules without adversarial annotation, the accuracy drops to \\(91.5\\%\\); using a unified prompt (refer to supplementary material) instead of instance-specific rules makes the accuracy further drop to \\(87\\%\\). It implies that both \"instance-specific rules\" and \"adversarial annotation\" are important." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As the model gets stronger, it makes fewer simple mistakes (e.g., grammatical errors or obvious hallucinations) and generates longer chain-of-thoughts, thus" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.218, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.236 + ], + "angle": 0, + "content": "the wrong answer becomes harder to distinguish from the correct ones. Tab. 3 shows that the accuracy of a single human, initial rules, and unified prompt all significantly drop on a more challenging subset of responses (50 responses exclusively generated by GPT-4V), while instance-specific rules maintain a relatively stable accuracy at \\(96.6\\%\\), highlighting its robustness for evaluating strong models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.348 + ], + "angle": 0, + "content": "A key concern when employing GPT-4 as an automated evaluator is its potential bias towards outputs produced by GPT-4V. To address this concern, we created a test set composed of instances GPT-4V generated incorrectly. On this test dataset, the GPT-4 evaluator with instance-specific rules achieves an accuracy of \\(97.9\\%\\), surpassing the human evaluator's accuracy of \\(93.6\\%\\). It suggests that GPT-4 with instance-specific rules demonstrates superior accuracy over human evaluators in pinpointing inaccuracies in GPT-4V's responses." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.367, + 0.465, + 0.383 + ], + "angle": 0, + "content": "4.4 Case Study for GPT-4V" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.787, + 0.421 + ], + "angle": 0, + "content": "Since GPT-4V is the current state-of-the-art, we report a detailed error analysis for it on AutoEval-Video. Here is a list of typical hard cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Understanding Temporal Features. In many cases, simply looking at a couple of isolated frames isn't enough to answer the question. Take instance 200 in Tab. 4, with its moving camera revealing the layout of a house. The model needs to understand the temporal ordering of frames to infer the house's spatial design to answer \"When I'm at the entrance, how do I get to the balcony that has three decorations on the table?\" Moreover, in instance 305 in supplementary material, the speed at which things change (like \"in an instant\") is a temporal feature. Most vision-language models, including GPT-4V, struggle to extract temporal features from many continuous frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.652 + ], + "angle": 0, + "content": "Connecting Multiple Frames. Many incorrect responses are due to GPT-4V tending to reason on a single frame, ignoring the connection between frames. In Tab. 4, instance 43 demonstrates this fact. Another interesting example is instance 116 in supplementary material. In describing a sequence of actions, GPT-4V mentions a certain action many times because it appears in many frames, although the correct answer should do it just once." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.745 + ], + "angle": 0, + "content": "GPT-4V also struggles to reason between scenes. For example, in instance 112 of Tab. 4, the first scene features a deer in third-person view gradually opening its eyes. The second scene shows an approaching train from the deer's first-person view. To answer \"Why does this orange animal have its eyes wide open?\", the model must connect these two events, but GPT-4V failed. There are more examples in supplementary material supporting this observation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.788, + 0.808 + ], + "angle": 0, + "content": "Moving Objects. Another group of hard cases involves moving objects, such as comparing the speed of two cars. As shown in Tab. 4, instance 191 records a police car overtaking a firefighting truck. However, GPT-4V perceives all frames as semantically similar and is unable to identify the faster vehicle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Overly General Responses. GPT-4V sometimes answers questions in a general context rather than addressing the specific content of given videos. For" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.559, + 0.129 + ], + "angle": 0, + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.177, + 0.779, + 0.189 + ], + "angle": 0, + "content": "Question: When I am at the entrance, how can I reach the balcony that has three decorations on the table? [ID:200]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.19, + 0.288, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.19, + 0.358, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.19, + 0.428, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.19, + 0.498, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.19, + 0.569, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.19, + 0.64, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.19, + 0.71, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.19, + 0.718, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.19, + 0.788, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.237, + 0.779, + 0.267 + ], + "angle": 0, + "content": "Sample Answer: Upon entering through the door, you can first turn left into the kitchen, and then turn right. At this point, you are facing the target balcony. All you need to do is continue straight, open the balcony door, and step outside." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.271, + 0.779, + 0.303 + ], + "angle": 0, + "content": "GPT-4V: To reach the balcony with three decorations on the table, enter the house through the main entrance, proceed straight ahead through the living room, and then turn left to go through the double doors leading to the outdoor covered patio area." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.308, + 0.548, + 0.319 + ], + "angle": 0, + "content": "Question: What action did the third parrot from the left perform?" + }, + { + "type": "text", + "bbox": [ + 0.741, + 0.309, + 0.778, + 0.319 + ], + "angle": 0, + "content": "[ID:43]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.322, + 0.288, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.322, + 0.358, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.322, + 0.428, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.322, + 0.498, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.322, + 0.569, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.322, + 0.638, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.322, + 0.709, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.322, + 0.78, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.368, + 0.459, + 0.379 + ], + "angle": 0, + "content": "Sample Answer: It flew into the person's hand." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.382, + 0.779, + 0.403 + ], + "angle": 0, + "content": "GPT-4V: The third parrot from the left appears to be perched on a hand and interacting with a person. It looks like the parrot might be engaged in some sort of play or social interaction with the individual." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.41, + 0.525, + 0.422 + ], + "angle": 0, + "content": "Question: Why are the eyes of this orange animal wide open?" + }, + { + "type": "text", + "bbox": [ + 0.736, + 0.411, + 0.778, + 0.421 + ], + "angle": 0, + "content": "[ID:112]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.423, + 0.288, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.423, + 0.358, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.423, + 0.428, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.423, + 0.498, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.423, + 0.569, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.423, + 0.639, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.423, + 0.709, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.423, + 0.78, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.469, + 0.76, + 0.48 + ], + "angle": 0, + "content": "Sample Answer: Because the animal saw a train approaching, the white creature on the tracks was in danger." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.483, + 0.779, + 0.535 + ], + "angle": 0, + "content": "GPT-4V: The orange animal appears to be a stylized or animated character, possibly showing surprise, alarm, or excitement. Wide open eyes in animation are often used to convey strong emotions or reactions. This could be in response to something happening in the scene or to depict a particular personality trait of the character. Without additional context, it's hard to specify exactly why, but wide eyes are a common visual cue for heightened emotions in animation." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.541, + 0.525, + 0.552 + ], + "angle": 0, + "content": "Question: In the video, which car is moving at a slower pace?" + }, + { + "type": "text", + "bbox": [ + 0.736, + 0.542, + 0.778, + 0.552 + ], + "angle": 0, + "content": "[ID:191]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.554, + 0.288, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.554, + 0.358, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.554, + 0.428, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.554, + 0.498, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.554, + 0.569, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.554, + 0.639, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.554, + 0.709, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.554, + 0.78, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.6, + 0.385, + 0.611 + ], + "angle": 0, + "content": "Sample Answer: The police car." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.615, + 0.779, + 0.645 + ], + "angle": 0, + "content": "GPT-4V: I'm unable to view videos or discern the speed of objects within them. The image you've provided is static, so it doesn't convey the speed of the cars. If you have a specific question about the details within the image, feel free to ask!" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.653, + 0.496, + 0.664 + ], + "angle": 0, + "content": "Question: Why are the audience members standing up?" + }, + { + "type": "text", + "bbox": [ + 0.741, + 0.653, + 0.778, + 0.663 + ], + "angle": 0, + "content": "[ID:97]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.665, + 0.288, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.665, + 0.358, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.665, + 0.428, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.665, + 0.498, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.665, + 0.569, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.665, + 0.639, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.665, + 0.709, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.665, + 0.78, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.712, + 0.526, + 0.723 + ], + "angle": 0, + "content": "Sample Answer: The player in the blue team uniform scored." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.726, + 0.779, + 0.768 + ], + "angle": 0, + "content": "GPT-4V: The audiences in sports video games, like the one depicted in the images, often stand up in response to exciting moments or plays. This is a way to simulate real-life reactions of fans during intense or critical moments in a game, such as a close score, a game-winning shot, or a spectacular play. The game developers include these reactions to enhance the realism and immersion of the game experience for players." + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.78, + 0.721, + 0.794 + ], + "angle": 0, + "content": "Table 4: Examples of Incorrect Responses from GPT-4V." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.549, + 0.157 + ], + "angle": 0, + "content": "Question: What was the score of the team in black in the video footage?" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.157, + 0.287, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.157, + 0.359, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.157, + 0.429, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.157, + 0.499, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.157, + 0.569, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.157, + 0.64, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.157, + 0.709, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.157, + 0.78, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.202, + 0.484, + 0.213 + ], + "angle": 0, + "content": "Sample Answer: The team in black scored eight points." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.216, + 0.466, + 0.226 + ], + "angle": 0, + "content": "Human: The black team scores 3 points in the video." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.232, + 0.456, + 0.241 + ], + "angle": 0, + "content": "Question: How did the character detect the enemy?" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.241, + 0.288, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.241, + 0.359, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.241, + 0.429, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.241, + 0.499, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.241, + 0.57, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.241, + 0.64, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.241, + 0.709, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.241, + 0.78, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.287, + 0.64, + 0.298 + ], + "angle": 0, + "content": "Sample Answer: The enemy shot the character so the enemy's location could be known." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.3, + 0.634, + 0.31 + ], + "angle": 0, + "content": "Human: The character used the scope on the sniper rifle to clearly observe distant enemies." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.316, + 0.647, + 0.326 + ], + "angle": 0, + "content": "Question: Why did this person initially appear stunned and then suddenly burst into laughter?" + }, + { + "type": "text", + "bbox": [ + 0.74, + 0.316, + 0.778, + 0.326 + ], + "angle": 0, + "content": "[ID:133]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.326, + 0.288, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.326, + 0.359, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.326, + 0.429, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.326, + 0.499, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.326, + 0.57, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.326, + 0.64, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.326, + 0.709, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.326, + 0.78, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.371, + 0.761, + 0.382 + ], + "angle": 0, + "content": "Sample Answer: Because the person threw grenades to kill the enemy but knocked a teammate down accidentally." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.385, + 0.657, + 0.395 + ], + "angle": 0, + "content": "Human: Because she didn't expect that the bomb she threw would explode and harm someone." + }, + { + "type": "title", + "bbox": [ + 0.23, + 0.407, + 0.771, + 0.421 + ], + "angle": 0, + "content": "Table 5: Hard cases for humans due to insufficient external knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.785, + 0.53 + ], + "angle": 0, + "content": "instance, in instance 97 shown in Tab. 4, GPT-4V answers \"Why are the audience members standing up?\" in a general way, even though there is an obvious and concrete reason. Similarly, instance 197 gives a general guide for operating a coffee machine even though the video demonstrates a specific one. It implies that GPT-4V tends to rely on its strong language model to produce a safe answer, sometimes ignoring key details in the video." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.46, + 0.566 + ], + "angle": 0, + "content": "4.5 Hard Cases for Humans" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.785, + 0.64 + ], + "angle": 0, + "content": "Humans achieve an accuracy rate of \\(72.8\\%\\) on AutoEval-Video. We conduct an in-depth analysis of these cases and discover that there are two major hard cases for humans: questions related to external knowledge and questions requiring accurate description and summarizing ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Questions Related to External Knowledge. Some questions in AutoEval-Video require a relatively high level of mastery of external knowledge in specific domains such as sports and games, resulting in human difficulties. For example, in Tab. 5, instance 38 shows a misunderstanding due to human participants not knowing basketball scoring rules. Similarly, instance 133 requires explaining a game streamer's laughter, an easy task for someone familiar with the game but confusing for others. However, considering the cognitive level of the general public, these shortcomings are reasonable and expected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Accurate Description and Summarizing Ability. Questions requiring detailed, step-by-step video descriptions are more challenging for humans compared to semantic reasoning problems due to their complexity and precision. For instance, as shown in Tab. 6, in instance 173, human makes careless mistakes by ignoring necessary CPU installation steps shown in the video. In instance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.558, + 0.129 + ], + "angle": 0, + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "143, errors occur due to insufficient observation, misidentifying corn kernels as soybeans. This phenomenon is common in many step-by-step descriptive tasks." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.184, + 0.779, + 0.194 + ], + "angle": 0, + "content": "Question: How can you prepare a dish using these ingredients and tools as demonstrated in the video? [ID:143]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.195, + 0.778, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.24, + 0.778, + 0.272 + ], + "angle": 0, + "content": "Sample Answer: Cut the potatoes, beets, purple cabbage, carrots, mushrooms, and bell peppers, put them in the steamer with the corn kernels, cut the gluten into strips, heat the steamer on the gas stove, and put the gluten into the air fryer." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.276, + 0.779, + 0.337 + ], + "angle": 0, + "content": "Human: The kitchen utensils needed for preparation include a cutting board, a chef's knife, a small steamer pot, and an air fryer. The preparation process for the ingredients is as follows: 1. Cut the potatoes into small chunks. 2. Slice the leafy greens into segments. 3. Chop the purple cabbage. 4. Slice the carrots into thick pieces. 5. Halve the white mushrooms. 6. Slice the bell peppers. 7. Prepare a small bowl of soaked soybeans. 8. Cut the blanched beef pieces into strips. Finally, Place the vegetables from steps 1-7 into the steamer pot, and place the beef strips from step 8 into the air fryer." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.343, + 0.778, + 0.353 + ], + "angle": 0, + "content": "Question: How to install a CPU on a motherboard as demonstrated in the video? [ID:173]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.353, + 0.778, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.4, + 0.778, + 0.44 + ], + "angle": 0, + "content": "Sample Answer: According to the clip, here is how to install a CPU on a motherboard: 1. Lift the metal arm next to the CPU socket. 2. Look for a small triangle indicator on the corner of the CPU and the socket and align them properly. 3. Hold the CPU by the sides and insert the CPU into the CPU socket on the motherboard. 4. Lower the arm down completely into its original position." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.444, + 0.778, + 0.465 + ], + "angle": 0, + "content": "Human: 1. Open the latch at the CPU location on the motherboard. 2. Place the side of the CPU with pins onto the corresponding location on the motherboard, aligning with the notches. 3. Close the latch." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.472, + 0.778, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.528, + 0.778, + 0.568 + ], + "angle": 0, + "content": "Sample Answer: First, put the pillow at the head of the bed; Second, put the clothes that need to be washed on the floor, and put the clothes that do not need to be washed on the chair; Third, pile up all the debris except for sheets quilts, and pillows on the floor; Fourth, after arranging the sheets and quilts, place the doll bear on the side near the head of the bed." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.573, + 0.778, + 0.613 + ], + "angle": 0, + "content": "Human: The order in which this person organizes the room is as follows: 1. Neatly arrange multiple pillows at the head of the bed. 2. Gather multiple items left on the bed onto a chair. 3. Remove other items from the bed, such as a large teddy bear and a white cushion. 4. Straighten the bedsheet to make it look even. 5. Place the teddy bear back at the head of the bed." + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.627, + 0.734, + 0.641 + ], + "angle": 0, + "content": "Table 6: Hard cases for humans related to descriptive ability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.685, + 0.532, + 0.701 + ], + "angle": 0, + "content": "5 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this paper, we introduce AutoEval-Video, a novel benchmark for comprehensively assessing large VLMs in open-ended video question answering. Utilizing an LLM evaluator, models can be automatically and accurately evaluated on AutoEval-Video. In particular, GPT-4V achieves the highest accuracy of \\(32.2\\%\\), still far from human accuracy of \\(72.8\\%\\). Through an extensive case study, we identify several limitations of current models' capabilities in video understanding, such as temporal and dynamic comprehension, and overly general responses. These findings will shape our future research direction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.18, + 0.785, + 0.226 + ], + "angle": 0, + "content": "We would like to express our sincere gratitude to the reviewers of ECCV 2024 for their insightful and constructive feedback. Their valuable comments have greatly contributed to improving the quality of our work." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.252, + 0.324, + 0.269 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.284, + 0.785, + 0.326 + ], + "angle": 0, + "content": "1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.327, + 0.785, + 0.422 + ], + "angle": 0, + "content": "2. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, A., Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z., Zhang, J., Zhang, X., Zhang, Y., Zhang, Z., Zhou, C., Zhou, J., Zhou, X., Zhu, T.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.785, + 0.451 + ], + "angle": 0, + "content": "3. Buch, S., Eyzaguirre, C., Gaidon, A., Wu, J., Fei-Fei, L., Niebles, J.C.: Revisiting the \"video\" in video-language understanding. In: CVPR. pp. 2917-2927 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.785, + 0.492 + ], + "angle": 0, + "content": "4. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR. pp. 961-970 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.534 + ], + "angle": 0, + "content": "5. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.576 + ], + "angle": 0, + "content": "6. Fan, H., Lin, L., Yang, F., Chu, P., Deng, G., Yu, S., Bai, H., Xu, Y., Liao, C., Ling, H.: Lasot: A high-quality benchmark for large-scale single object tracking. In: CVPR. pp. 5374-5383 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.604 + ], + "angle": 0, + "content": "7. Fu, J., Ng, S.K., Jiang, Z., Liu, P.: Gptscore: Evaluate as you desire. arXiv preprint arXiv:2302.04166 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.785, + 0.659 + ], + "angle": 0, + "content": "8. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The \"something something\" video database for learning and evaluating visual common sense. In: ICCV. pp. 5842-5850 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.701 + ], + "angle": 0, + "content": "9. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR. pp. 6904-6913 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "0. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "1. Grunde-McLaughlin, M., Krishna, R., Agrawala, M.: Agqa: A benchmark for compositional spatio-temporal reasoning. In: CVPR. pp. 11287-11297 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "2. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., et al.: Ava: A video dataset of spatiotemporally localized atomic visual actions. In: CVPR. pp. 6047-6056 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.284, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.558, + 0.129 + ], + "angle": 0, + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "14. Hu, A., Chen, S., Zhang, L., Jin, Q.: Infomatic: An informative metric for reference-free image caption evaluation. arXiv preprint arXiv:2305.06002 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.219 + ], + "angle": 0, + "content": "15. Huang, D.A., Ramanathan, V., Mahajan, D., Torresani, L., Paluri, M., Fei-Fei, L., Niebles, J.C.: What makes a video a video: Analyzing temporal information in video understanding models and datasets. In: CVPR. pp. 7366-7375 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.22, + 0.785, + 0.262 + ], + "angle": 0, + "content": "16. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.263, + 0.785, + 0.292 + ], + "angle": 0, + "content": "17. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV. pp. 706-715 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.293, + 0.785, + 0.321 + ], + "angle": 0, + "content": "18. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: ICCV. pp. 2556-2563. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.322, + 0.785, + 0.349 + ], + "angle": 0, + "content": "19. Lei, J., Berg, T.L., Bansal, M.: Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.35, + 0.785, + 0.392 + ], + "angle": 0, + "content": "20. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.435 + ], + "angle": 0, + "content": "21. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.436, + 0.785, + 0.478 + ], + "angle": 0, + "content": "22. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.479, + 0.785, + 0.508 + ], + "angle": 0, + "content": "23. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.537 + ], + "angle": 0, + "content": "24. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.551 + ], + "angle": 0, + "content": "25. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.552, + 0.785, + 0.594 + ], + "angle": 0, + "content": "26. Lin, Y.T., Chen, Y.N.: Llm-eval: Unified multi-dimensional automatic evaluation for open-domain conversations with large language models. arXiv preprint arXiv:2305.13711 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.637 + ], + "angle": 0, + "content": "27. Liu, D., Qu, X., Hu, W.: Reducing the vision and language bias for temporal sentence grounding. ACM MM (2022), https://api_semanticscholar.org/CorpusID:251104851" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.638, + 0.785, + 0.667 + ], + "angle": 0, + "content": "28. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.667, + 0.785, + 0.695 + ], + "angle": 0, + "content": "29. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.696, + 0.785, + 0.724 + ], + "angle": 0, + "content": "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.725, + 0.785, + 0.754 + ], + "angle": 0, + "content": "31. Liu, Y., Iter, D., Xu, Y., Wang, S., Xu, R., Zhu, C.: G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.755, + 0.785, + 0.796 + ], + "angle": 0, + "content": "32. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.797, + 0.725, + 0.812 + ], + "angle": 0, + "content": "33. OpenAI: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.841 + ], + "angle": 0, + "content": "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL. pp. 311-318 (2002)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.203 + ], + "angle": 0, + "content": "35. Pătrăucean, V., Smaira, L., Gupta, A., Continente, A.R., Markeeva, L., Banarse, D., Koppula, S., Heyward, J., Malinowski, M., Yang, Y., et al.: Perception test: A diagnostic benchmark for multimodal video models. arXiv preprint arXiv:2305.13786 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "36. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.272 + ], + "angle": 0, + "content": "37. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.314 + ], + "angle": 0, + "content": "38. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.342 + ], + "angle": 0, + "content": "39. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR. pp. 4566-4575 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "40. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y.F., Wang, W.Y.: Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In: ICCV. pp. 4581-4591 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.424 + ], + "angle": 0, + "content": "41. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. NeurIPS 35, 24824-24837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.452 + ], + "angle": 0, + "content": "42. Wu, Y., Lim, J., Yang, M.H.: Online object tracking: A benchmark. In: CVPR. pp. 2411-2418 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.479 + ], + "angle": 0, + "content": "43. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR. pp. 5288-5296 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.786, + 0.522 + ], + "angle": 0, + "content": "44. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.786, + 0.563 + ], + "angle": 0, + "content": "45. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.605 + ], + "angle": 0, + "content": "46. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.646 + ], + "angle": 0, + "content": "47. Yu, Z., Zheng, L., Zhao, Z., Wu, F., Fan, J., Ren, K., Yu, J.: Anetqa: A large-scale benchmark for fine-grained compositional reasoning over untrimmed videos. In: CVPR. pp. 23191-23200 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "48. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.715 + ], + "angle": 0, + "content": "49. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.786, + 0.757 + ], + "angle": 0, + "content": "50. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "51. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.799 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_origin.pdf b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5cea8f910e06a4c87681c7d6cc8a9e19a166df7 --- /dev/null +++ b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/4a8f6c0b-f38f-47eb-af35-072ebcf3736a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07f93869dab4ee9bf14234129f57a06d2c80b89d22eedae0935ae03d3ec1eb31 +size 16222420 diff --git a/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/full.md b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8a1eeff63c23d29b4423dbad9e22bedcab84297 --- /dev/null +++ b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/full.md @@ -0,0 +1,584 @@ +# AutoEval-Video: An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering + +Xiuyuan Chen $^{1,*}$ , Yuan Lin $^{2,\dagger}$ , Yuchen Zhang $^{2,\dagger}$ , and Weiran Huang $^{1,\dagger}$ + +1 MIFA Lab, Qing Yuan Research Institute, SEIEE, Shanghai Jiao Tong University +2 ByteDance Research + +Abstract. We propose a novel and challenging benchmark, AutoEval-Video, to comprehensively evaluate large vision-language models in open-ended video question answering. The comprehensiveness of AutoEval-Video is demonstrated in two aspects: 1) AutoEval-Video constructs open-ended video-questions across 9 skill dimensions, addressing capabilities of perception, comprehension, and generation; 2) AutoEval-Video contains newly collected videos that cover over 40 distinct themes. To efficiently evaluate responses to the open-ended questions, we employ an LLM-based evaluation approach, but instead of merely providing a reference answer, we annotate unique evaluation rules for every single instance (video-question pair). To maximize the robustness of these rules, we develop a novel adversarial annotation mechanism. By using instance-specific rules as prompt, GPT-4, as an automatic evaluator, can achieve a stable evaluation accuracy of around $97.0\%$ , comparable to the $94.9\% - 97.5\%$ accuracy of a human evaluator. Furthermore, we assess the performance of eleven large vision-language models on AutoEval-Video. Among them, GPT-4V(ison) significantly outperforms other models, achieving an accuracy of $32.2\%$ . However, there is still substantial room for improvement compared to human accuracy of $72.8\%$ . By conducting an extensive case study, we uncover several drawbacks of GPT-4V, such as limited temporal and dynamic comprehension, and overly general responses. Code is available at https://github.com/Xiuyuan-Chen/AutoEval-Video. + +# 1 Introduction + +Generative large vision-language models [2, 5, 21, 22, 29, 30, 32, 48] have exhibited the considerable potential to address fundamental challenges in video understanding, temporal and causal comprehension [3]. Furthermore, these models can deal with more complex open-ended video tasks, surpassing traditional specialized video tasks, like video captioning [17, 40, 43], action classification [8, 12, 36], and object tracking [6, 42]. + +
SkillExample of Video-Question
Dynamic PerceptionThe video depicts various ways of relaxing and de-stress. +Question: Please List the three relaxation techniques from the video in order.
State Transitions PerceptionThe video captures the process of a chemistry experiment. +Question: What is the change in the color of the liquid after blowing?
Camera Movement PerceptionThe camera's position in the video undergoes a series of adjustments. +Question: What changes have been made to the position of the camera used for filming this video?
Explanatory ReasoningThe camera records a basketball player sinking a shot and flashing a V sign to the camera. +Question: Why is the individual wearing blue jeans in the video making that particular gesture?
Counterfactual ReasoningA purple animal character in the cartoon follows the guidance of another character and falls into a trap. +Question: What action could the purple animal take to avoid the consequences shown in the video?
Predictive ReasoningThe video records a person putting groceries into a cupboard and preparing to close the cupboard door (but without any depiction of the door-closing action). +Question: What the person will do next?
Comparison ReasoningThe video sequentially depicts several individuals drinking water, where one person exhibits discomfort while the others appear normal. +Question: How does the clothing of the uncomfortable person differ from that of the others?
Reasoning with External KnowledgeThe video records a basketball game, in which one team scored two three-pointers and a two-pointer. +Question: In the video footage, what is the score of the team wearing black?
DescriptionThe video highlights two additional features of a vacuum cleaner product. +Question: Sequentially describe the two extra features of the vacuum cleaner.
+ +Table 1: Skill Dimensions and Corresponding Examples for Video-Questions in AutoEval-Video. + +Effectively evaluating large vision-language models is crucial for improving their performance, as it identifies areas for improvement and enables the comparison and investigation of various models. Several benchmarks [6,8,10,12,17, 35,36,40,42,43] have been developed to evaluate capabilities of specific tasks in video understanding. Nonetheless, they fail to assess model performance in a comprehensive manner. To close this gap, recent studies [10,20,35] aim to build benchmarks covering a wide range of skills. These benchmarks primarily consist of multiple-choice question-answers to facilitate effective evaluation. However, in comparison to open-ended question answering, multiple-choice question answering presents two key limitations: 1) they simplify the tasks, and 2) they do not accurately reflect real-world scenarios. For instance, in a video-based dialogue, responses should be generated directly, rather than making a selection from a set of options. + +In this paper, we introduce AutoEval-Video, a comprehensive and challenging benchmark designed to evaluate the video understanding capabilities of large vision-language models. AutoEval-Video comprises 327 challenging open-ended video-question instances that span across nine skill dimensions in Tab. 1, which address video-specific perception, comprehension, and generation skills. All videos in AutoEval-Video are newly collected from YouTube, including over 40 genres such as daily life, gaming, sports, and more. We also establish rigorous annotation standards to guarantee the difficulty and quality of our benchmark. These standards include minimizing single-frame bias [3, 15, 19] and language bias [9, 27], and setting objective questions. + +Assessing responses to open-ended questions can be difficult, especially when examining outputs generated by LLMs. These responses may involve additional + +information such as video description and chain-of-thought [41] analysis, diminishing the reliability of traditional automatic evaluation metrics, which typically involve comparing a given response to a reference answer, such as BLEU [34], ROUGE [25] and CIDEr [39]. To address this issue, we employ an LLM-based evaluation approach [7, 31, 45, 49]. An LLM evaluator comprises a strong LLM and its prompting strategy. Existing works [26, 46, 50] design a unified prompt for specific tasks. However, due to the diversity of videos and questions in AutoEval-Video, a single unified prompt is insufficient for accurate evaluation. To tackle this challenge, we meticulously annotate distinct evaluation rules for each instance instead of solely providing a reference answer. To improve the precision of these rules, we develop a novel adversarial annotation mechanism. A red team is engaged to devise unaddressed corner cases that the rules may not cover, assisting the annotation team in refining them. The final rule is ultimate when the red team cannot discover any further corner cases that break the rules. After intensive annotation, using instance-specific rules as prompt, GPT-4 can serve as an automatic evaluator, achieving an evaluation accuracy of $97.0\%$ , comparable to the single human accuracy of $97.5\%$ . + +Leveraging AutoEval-Video, we have conducted a comprehensive evaluation of eleven prominent large vision-language models, which fall into two categories: ImageLLM and VideoLLM. Our assessment reveals that GPT-4V(ison) significantly surpasses other models. Utilizing 8 randomly sampled frames as input, GPT-4V achieves an accuracy of $22.2\%$ , compared to $72.8\%$ accuracy of humans, indicating considerable potential for improvement. Moreover, we discover that by increasing the input to 16 frames, GPT-4V's performance markedly improves to $32.2\%$ . Finally, we perform an in-depth case study for GPT-4V, identifying several critical challenges that demand further investigation, such as limited temporal and dynamic understanding, and overly general responses. + +# 2 Related Work + +Large Vision-Language Models. The remarkable success of LLMs has been accompanied by significant progress in large vision-language models. Recent works [2, 5, 21, 29] improve the visual perception and cognition capabilities by utilizing the strong understanding and generation abilities of LLMs. In addition, some research [22, 32, 48] further incorporates video inputs and leverages the extensive potential of LLMs for video understanding. The considerable progress necessitates more challenging and comprehensive benchmarks to evaluate the video comprehension capabilities of large vision-language models. + +Benchmarks for Video Understanding. Several video benchmarks have been employed to assess model performance in various areas, such as action recognition [11, 12, 47], video captioning [13, 14, 43], and action classification [6, 8, 16, 18, 36, 42]. However, these benchmarks do not provide a comprehensive assessment of a model's capabilities in video understanding. To address this issue, some general benchmarks [20, 35] have been established, developing + +![](images/2dfc754396e327471d28713cc48f2c538ae1e0cac0ecd76b2555559c3f0aa801.jpg) +Q: What are the steps to prepare this dish as shown in the video? + +# Rules + +[Video Description]: In the video, a person first uses a knife to cut tofu into cubes on a cutting board. Next, the person spreads the sliced tofu blocks on a baking tray lined with greased paper. Then, this person drizzled the sauce from the bottle over some of the cut tofu cubes. Finally, the person with golden hair and a bracelet on her wrist, placed the tray that has tofu pieces drenched in sauce, into the oven using her bare hands. + +[Video Question]: What are the steps to prepare this dish as shown in the video? + +Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules: + +1. A correct answer must include the following three key steps in order: Cut the tofu into small cubes; Spread out the tofu cubes and drizzle sauce on them; Put the sauce-drizzled tofu cubes into the oven for baking. + +2. The answer to the description of the shape of tofu blocks can be anything similar to 'block'. It doesn't necessarily have to be described as a perfectly shaped cube. + +3. Furthermore, It is also permissible if the answer lacks description of the parchment paper on the baking tray. + +4. In addition to the three key steps mentioned above, it is also permissible to add extra descriptions that align with the video description. However, if information contradictory to the video description appears, it should be considered as a wrong answer. + +5. The response should contain correctly spelled words. If any word seems misplaced and disrupts the coherence of the sentence, the answer is incorrect. + +[Answer to be judged]: + +[Form of the result]: + +Please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above. + +![](images/3a1230a3fb782e2e315ae6b93ea9f67b033ca5ccef49d7e27d1c6d861ab8f5f6.jpg) + +# + +1. Cut the tofu into pieces and then place it on a wooden board. +2. Drizzle the sauce over some tofu slices. +3. Put on gloves and place the baking tray with tofu pieces into the oven. + +![](images/d0327b2a3132818e7f86b10f351db6626d04bdd5b320df59eb6776aa858adc33.jpg) + +1. Cut the tofu into small cubes and spread them out. +2. Place the tofu cubes on a baking tray and drizzle with sauce. +3. Put the sauce-drizzled tofu cubes into the oven to bake. + +![](images/ce1d568fa0900963a23c2bab841f309f9fcd171755974ff06f96d795c50f4bfd.jpg) + +1. Cut the white tofu into pieces, then place them in a baking tray. +2. Drizzle sauce over some of the tofu pieces. +3. Place the baking tray with the tofu pieces in the oven. + +![](images/b45f7bbfc1fcb8f40bc799d967014199ea9ef5dc3c00f37b239e3b12ceef4646.jpg) + +The answer provided is incorrect for the following reasons: +1. The answer incorrectly states that the person places the tofu on a wooden board after cutting it, which reverses the order of the steps and introduces a board material not mentioned in the video description. +2. The sauce is misspelled as "sauce," which is a spelling error. +3. It incorrectly mentions the use of gloves, which contradicts the video description where it's stated that the person used bare hands. + +![](images/78a20e3ac12970a7ef7539bd85f8e6ea7aec210f305c7b1b577d53ca816a8879.jpg) + +The [Answer to be judged] correctly includes the three key steps required by the rules: cutting tofu into cubes, spreading the tofu out and drizzling sauce on them, and putting the sauce-drizzled tofu into the oven to bake. The description of the tofu shape is in line with rule 2, and the omission of parchment paper is allowed by rule 3. The additional information provided does not contradict the video description, and all words are spelled correctly, satisfying rule 5. + +![](images/39f2b8bb76ab92cb8042a2b382e4741d15fe1c11e75a440949ce04188896f5bf.jpg) + +The answer provided correctly states the three key steps in the order they appear in the video description: cutting the tofu into pieces, drizzling sauce on the tofu, and placing the tray in the oven. +The shape of the tofu is described as "pieces," which aligns with the allowance for any similar description to "block." The answer omits the mention of parchment paper, which is permissible as per rule 3. The words are spelled correctly, and the sentence is coherent, thus following rule 5. + +Fig. 1: Example Instance in AutoEval-Video and Automatic Evaluation Process. Each instance has three components: video, question, and rules. For each video and question, we establish specific and detailed evaluation rules. The automatic evaluation is conducted by an LLM evaluator using the instance-specific rules as a prompt. + +multiple-choice VQA to assess models' abilities across different aspects. Nevertheless, compared to open-ended question answering, multiple-choice question answering oversimplify tasks and may not reflect realistic instructions. + +In addition, existing benchmarks mainly employ two methods for video collection: 1) sourcing videos related to a specific task from web [4,16] or large-scale video datasets [1,11,47]; 2) recording new videos based on pre-defined scripts [35] or themes [10]. However, both approaches suffer from limited visual diversity, leading to a video distribution that inadequate represents real-world scenarios. + +# 3 AutoEval-Video + +# 3.1 Overview + +In AutoEval-Video, each instance comprises three components: video, question, and rules, as illustrated in Fig. 1. The construction involves three steps. + +First, to address the essential aspects of real-world multi-modal understanding, the uncovered areas of existing benchmarks, and the limitations of current models, we identify nine skill dimensions, as illustrated in Tab. 1. These dimensions encompass diverse skill areas (including perception, cognition, and generation) and various reasoning types (such as descriptive, explanatory, predictive, comparative, and counterfactual). + +Second, for each skill dimension, an annotation team gathers videos from YouTube1 and crafts relevant open-ended questions. It is critical to prevent potential single-frame bias and language bias while formulating these questions. To ensure the assessability of responses, the questions must retain objectivity. + +Finally, after acquiring the video-question pairs, the annotation group develops evaluation rules for each instance. These rules establish precise conditions for considering whether a response is correct or wrong, involving precise criteria for recognizing seemingly accurate but flawed corner case responses, and vice versa. To ensure thoroughness, we employ a second annotation team, known as the "red team", to devise unaddressed corner cases. The final rule is ultimate when the red team cannot discover any further corner cases that break the rules. + +As demonstrated above, the annotation of AutoEval-Video is sophisticated and labor-intensive. We construct a total of 327 instances. The average cost per instance amounts to $28. + +# 3.2 Skill Dimensions + +Video understanding encompasses two fundamental skills: perception and cognition. Different from image perception, video perception emphasizes temporal information, such as activity, state transitions, and camera movement. With regard to cognition, our benchmark considers reasoning based on temporal or spatial information, simultaneously spanning causal reasoning, comparison, and reasoning with external knowledge. In addition, considering that hallucination [24,28,37,51] is a prominent challenge for contemporary large vision-language models, we incorporate tasks requiring descriptive ability in our benchmark. In summary, we identify nine skill dimensions, as shown in Tab. 1, which will address the aforementioned capabilities. + +# 3.3 Video and Question Collection + +Video Collection. We gather all videos from YouTube, taking into account the following considerations: + +- To minimize the probability of utilizing videos that have been employed in the model training process, we collect new videos from YouTube rather than employing well-known open-sourced video datasets. + +- We aim to construct a benchmark that incorporates real-world visual content so as to directly assess the capability of models in addressing realistic challenges. Thus, we choose to select videos from YouTube instead of filming new videos based on scripts. + +In addition, we examine the video categories of our benchmark, which covers mainstream video genres in real-world. All video sources are guaranteed to possess a resolution of 720P or higher. More statistics of the videos can be found in Fig. 2. + +![](images/1df86a7d356967af43059472221e271cbc70ba10814001fed209baee5a500ecd.jpg) +(a) Genre distribution of videos in AutoEval-Video +Fig. 2: Statistics of AutoEval-Video. (a) Genre distribution of videos. (b) Statistics on question and rule lengths, number of scenes, and video duration. + +
StatisticsMeanMedian
Question Length (#words)10.9410.00
Rule Length (#words)302.95296.00
Video Shots (#Scenes)3.162.00
Video Duration (#Seconds)14.5810.00
+ +(b) Statistics of AutoEval-Video + +Question Collection. The questions of AutoEval-Video are meticulously designed by human annotators, avoiding templates or paradigms, to guarantee a diverse range of questions. Annotators may either initiate a question and then search for an appropriate video, or select a video first and craft a question afterward. To maximize dataset diversity, we annotate only one question per video. To obtain high-quality video-question pairs that can effectively diagnose models' video comprehension abilities, we propose three principles for constructing the benchmark: avoiding single-frame bias, reducing language bias, and formulating objective questions, as discussed below. + +Many video benchmarks suffer from single-frame bias [3, 15, 19], indicating that the information in an oracle frame is almost sufficient to deduce the answer. To address this issue, we instruct annotators to devise questions that require the consideration of multiple frames for successful completion. Questions that can be answered based on only one frame are deemed invalid. + +Existing video-language benchmarks are affected by language bias [9, 27], suggesting that models can obtain the correct answers without seeing video content. To tackle this issue, we involve three additional human annotators who try their best to answer questions without viewing the videos. If any of the annotators can accurately answer the question, the instance will be discarded. + +![](images/f904da36f262dad975fc90f565726b0b89b99c87aab70e9b5b050e606b349922.jpg) +Q: Why did the race car driver exit the vehicle? + +![](images/6fe5b1c963e28a609ed7ebcd54c1cc0d592b60b97fafc16a8b773523b3caa0cd.jpg) + +![](images/a2fed1de0ca7ee75019c78d66e8857be285c51f89a0f51dcbc4f62b1ec4065e7.jpg) + +![](images/0409356616cd15bdc3eab3b965a8caf7dea77599b43dc5b1a3f33058d0ec0fb6.jpg) + +![](images/c14c30632fde6bf9878099fdb7576b18544a5491910757a7f7f4c75262148ce8.jpg) + +![](images/60bb6dd6c6544299f90b79ae95f491f8e877c66806429e24296d5b36ea2c7ece.jpg) + +![](images/08c81c8a6ead6f58c7b0f5d7676a22fabe1b1d16d9345dfef3a2360bbdd94198.jpg) + +![](images/11e927e71edb90ad76cad873b1204f02003c95810da8b7484a8da17296bc1ef2.jpg) + +# Initial Rules + +[Video description]: In a racetrack, a white race car flipped over and came to a halt, then caught fire. Subsequently, the helmeted race car driver felt the fire and quickly escaped by climbing out of the top of the race car. Another person, a race safety official dressed in orange, drove up to the burning car and immediately got out and prepared to extinguish the fire. [Video Question]: Why did the race car driver exit the vehicle? + +Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules: + +1. The main key point in the answer lies in "race car caught on fire". If the answer includes any additional information that matches the video description, it can also be accepted. + +2. The words in the answer must be spelled correctly. If a word does not make sense in the sentence, the answer is incorrect. + +3. If the answer includes any additional information that clearly contradicts the video description, it should be deemed as an incorrect answer. + +[Answer to be judged]: (Adversarial Answer) + +[Form of the Result]: + +please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above. + +# + +Because the race car driver's car caught fire, which might have been caused by the driver's improper operation. + +![](images/9fffe01af2769336d17c414d6fbd5392b475a68c0f05401ea2101ba83f68377e.jpg) +Fig. 3: Adversarial Annotation for Rules. A red team develops adversarial answers that break the initial rules. Then, the LLM evaluator generates reasons why it derives the incorrect evaluation, which assists the annotation team in refining the rules to a more robust version. + +The reason provided in the answer, "improper operation by the driver", adds speculative information not present in the video description. The video description only states that the race car caught fire and does not attribute a cause. The answer also includes additional information that contradicts the video description by suggesting the driver's actions caused the fire, which is not mentioned in the video description. Therefore, the answer should be deemed incorrect. + +# Updated Rules + +[Video description]: In a racetrack, a white race car flipped over and came to a halt, then caught fire. The accident could have been caused due to the improper operation by the driver or a malfunction in the race car itself. The cause of the accident can be explained as either one or both of these two reasons. Subsequently, the helmed race car driver felt the fire and quickly escaped by climbing out of the top of the race car. Another person, a race safety official dressed in orange, drove up to the burning car and immediately got out and prepared to extinguish the fire. + +[Video Question]: Why did the race car driver exit the vehicle? + +Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules: + +1. The main key point in the answer lies in "race car caught on fire". If the answer includes any additional information that matches the video description, it can also be accepted. + +2. The words in the answer must be spelled correctly. If a word does not make sense in the sentence, the answer is incorrect. + +3. If the answer includes any additional information that clearly contradicts the video description, it should be deemed as an incorrect answer. + +4. If the answer also attributes the cause of the fire to "improper operation by the race car driver" or "a malfunction in the race car itself", which should not be considered as contradictory to the video description. + +[Answer to be judged]: (Adversarial Answer) + +[Form of the Result]: + +please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above. + +In our benchmark, all question-answer pairs should be assessed by a set of rules, requiring objective questions since evaluating subjective responses is challenging. For instance, a video depicting a man with a neutral expression, accompanied with a question, "What is the man thinking about?" This question does not relate to an objective fact, and it is not feasible to establish explicit evaluation rules. + +# 3.4 Annotation of Rules + +For each instance in AutoEval-Video, we annotate a set of rules describing the criteria for determining the correctness of an answer. The rationale behind this approach is as follows: + +- All questions in AutoEval-Video are open-ended, rendering the enumeration of all reference answers infeasible. Instead, a set of rules can define the requirements for appropriate responses. + +- Modern LLMs, such as GPT-4, possess remarkable language understanding capabilities, allowing them to comprehend the provided rules and act as automatic evaluators. Unlike other works that use a single, unified prompt to assess all video-question responses [26, 46, 50], we annotate specific rules for each video-question pair, ensuring an in-depth and specific understanding of the criteria. + +The valid evaluation rules include a detailed description of the video content, questions about the video, key points of the correct answer, and the required output format, as illustrated in Fig. 1. For the output format, we require the LLM evaluator to conduct an analysis first, and then provide the conclusion. Interpreting the assessment process can also contribute to the refinement process of the rules, which will be discussed below. + +To guarantee the thoroughness of the rules, we employ an adversarial annotation process (shown in Fig. 3) conducted by a separate group of annotators, known as the red team. Initially, the annotation team establishes initial rules denoted as $r_0$ . For each $k$ -th round of attack, the red team will endeavor to construct an answer, $a_k$ , causing erroneous judgments from an LLM evaluator (e.g. GPT-4) with the preceding rules $r_{k-1}$ . Upon obtaining an adversarial answer $a_k$ , the annotation team must update their rules to a refined version, $r_k$ . The iterative attack-refine process will continue until the red team cannot identify any adversarial answer. The red team may construct attack answers through various methods, such as: 1) directly writing one based on their understanding of the rules; and 2) prompting LLMs to generate adversarial answers. During the refinement process, the annotators can utilize the interpretations of the LLM evaluator, which will help to quickly figure out why the evaluator gets confused and facilitate the production of more accurate rules. On average, there are 3 attack-refine iterations in the construction of rules in AutoEval-Video. + +Furthermore, to enhance the motivation and engagement of both the annotation team and the red team in confrontation, we devised a reward mechanism. For each instance, the annotation team initially has a full reward, while the red team starts with zero reward. If the red team successfully discovers an adversarial answer, they will earn a reward, while the annotation team will lose a portion of their reward. As the number of iterations increases, there will be a corresponding increase in rewards for successful attacks, as well as higher penalties for the annotation team. This approach not only encourages the red team to target mature rules but also urges the annotation team to consistently refine their rules throughout every iteration, ensuring optimal results. + +# 4 Experiments + +# 4.1 Experimental Setup + +Models. Utilizing AutoEval-Video, we assess the performance of eleven models, including six ImageLLMs (i.e., GPT-4V [44], Gemini-Ultra-Vision [38], LLaVA1.5 [29], Qwen-VL [2], InstructBLIP [5], and BLIP-2 [21]) and five VideoLLMs + +(i.e., Video-ChatGPT [32], VideoChat, VideoChat2 [22], Video-LLaMA [48], and LLaMA-VID [23]). For each video, we uniformly sample 8 or 16 frames, resulting in two different input configurations. In particular, for ImageLLMs, we concatenate the 8 or 16 frames to create a single image as visual input (refer to supplementary material for the detailed concatenation method). + +LLM as an automatic evaluator. We adopt GPT-4 [33] as an automatic evaluator for AutoEval-Video. For each instance, GPT-4 generates an evaluation of the correctness of a given answer to be judged using the rules as prompts, as demonstrated in Fig. 1. + +Human baseline. We recruit three human participants to answer questions in AutoEval-Video. The overall average accuracy is $72.8\%$ . Through closely examining the mistakes made by humans, we identify two major challenges they face: 1) Questions related to external knowledge, such as unfamiliar games, may lead to incorrect answers from humans; 2) Questions requiring precise memory skills, such as recalling multi-step processes in a video, may cause a person to overlook one or more steps. More examples can be found in supplementary material. + +
Model TypeModelAccuracy
8 Frames16 Frames
ImageLLMBLIP-2 [21]0.31.2
InstructBLIP [5]7.914.3
Qwen-VL [2]7.010.0
LLaVA-1.5 [29]8.58.2
GPT-4V [44]22.232.2
Gemini-Ultra-Vision [38]17.920.9
VideoLLMVideo-ChatGPT [32]6.74.9
VideoChat [22]13.410.9
Video-LLaMA [48]11.28.5
LLaMA-VID [23]9.88.0
VideoChat2 [22]0.30.6
Human (full videos)72.8
+ +Table 2: Evaluation Results of Different Models on AutoEval-Video. We also provide human results after viewing full videos. + +# 4.2 Performances on AutoEval-Video + +We compare 11 large vision-language models on Tab. 2. Findings show: 1) GPT-4V is much stronger than other models; 2) Among open-sourced models, VideoLLM generally outperforms ImageLLM; 3) GPT-4V falls significantly short of human-level, reflecting the challenge of our benchmark and a big headroom for further advancements. + +We find that open-sourced models suffer from three typical categories of bad cases. 1) Hallucination: the described objects or events are not present in the video, leading to wrong answers; 2) Alignment: since AutoEval-Video includes a diverse range of queries, it is often observed that the models' responses do + +not align with the question; 3) Weak LLM: many models cannot get to the correct answer due to their limited language generation capabilities. We have provided examples for each category in the supplementary material. GPT-4V shows significant progress in these directions, but it still has fundamental issues, as discussed in Sec. 4.4. + +Among open-sourced models, VideoChat achieves the highest score. Interestingly, GPT-4V correctly answers only $32.7\%$ instances where VideoChat is correct, indicating that VideoChat is stronger in certain domains. These domains are 1) traffic-related tasks; 2) comparison between dynamic objects; and 3) giving video-specific responses rather than general vague answers. In supplementary material, we provide examples for each of these three domains. + +Finally, we notice that 8 frames are insufficient for answering certain questions. Therefore, we evaluate models using 16 frames as input. Under this setting, GPT-4V significantly improves, reaching $32.2\%$ accuracy (examples provided in supplementary material). While longer inputs usually benefit ImageLLMs, VideoLLMs don't exhibit the same trend, possibly because they are explicitly trained in the 8-frame setting. + +
GPT-4 PromptAccuracy of Evaluators
Models/Human ResponsesGPT-4V Responses
Instance-specific rules97.096.6
W/o adversarial annotation91.588.1
W/o rules (unified prompt)87.069.5
Human97.594.9
+ +Table 3: Ablation Study on Automatic Evaluators. We assess the evaluation accuracy of GPT-4 with instance-specific rules, initial rules, and unified prompts on responses generated by models/humans, as well as responses exclusively generated by GPT-4V. In addition, we present the evaluation accuracy of a single human. + +# 4.3 Accuracy of GPT-4 as Automatic Evaluator + +To validate GPT-4 as an automatic evaluator, we create a test dataset consisting of 200 randomly sampled (video, question, response) triplets, either model-generated or human-generated. Then we employed three human experts to evaluate each response, utilizing their collective decision as the ground truth. As shown in Tab. 3, with the evaluation rules, GPT-4 achieves $97\%$ agreement with the ground truth, comparable to $97.5\%$ of a single human expert. + +We perform an ablation study to investigate the necessity of our proposed way of annotation. As reported in Tab. 3, if we use the initial version of rules without adversarial annotation, the accuracy drops to $91.5\%$ ; using a unified prompt (refer to supplementary material) instead of instance-specific rules makes the accuracy further drop to $87\%$ . It implies that both "instance-specific rules" and "adversarial annotation" are important. + +As the model gets stronger, it makes fewer simple mistakes (e.g., grammatical errors or obvious hallucinations) and generates longer chain-of-thoughts, thus + +the wrong answer becomes harder to distinguish from the correct ones. Tab. 3 shows that the accuracy of a single human, initial rules, and unified prompt all significantly drop on a more challenging subset of responses (50 responses exclusively generated by GPT-4V), while instance-specific rules maintain a relatively stable accuracy at $96.6\%$ , highlighting its robustness for evaluating strong models. + +A key concern when employing GPT-4 as an automated evaluator is its potential bias towards outputs produced by GPT-4V. To address this concern, we created a test set composed of instances GPT-4V generated incorrectly. On this test dataset, the GPT-4 evaluator with instance-specific rules achieves an accuracy of $97.9\%$ , surpassing the human evaluator's accuracy of $93.6\%$ . It suggests that GPT-4 with instance-specific rules demonstrates superior accuracy over human evaluators in pinpointing inaccuracies in GPT-4V's responses. + +# 4.4 Case Study for GPT-4V + +Since GPT-4V is the current state-of-the-art, we report a detailed error analysis for it on AutoEval-Video. Here is a list of typical hard cases. + +Understanding Temporal Features. In many cases, simply looking at a couple of isolated frames isn't enough to answer the question. Take instance 200 in Tab. 4, with its moving camera revealing the layout of a house. The model needs to understand the temporal ordering of frames to infer the house's spatial design to answer "When I'm at the entrance, how do I get to the balcony that has three decorations on the table?" Moreover, in instance 305 in supplementary material, the speed at which things change (like "in an instant") is a temporal feature. Most vision-language models, including GPT-4V, struggle to extract temporal features from many continuous frames. + +Connecting Multiple Frames. Many incorrect responses are due to GPT-4V tending to reason on a single frame, ignoring the connection between frames. In Tab. 4, instance 43 demonstrates this fact. Another interesting example is instance 116 in supplementary material. In describing a sequence of actions, GPT-4V mentions a certain action many times because it appears in many frames, although the correct answer should do it just once. + +GPT-4V also struggles to reason between scenes. For example, in instance 112 of Tab. 4, the first scene features a deer in third-person view gradually opening its eyes. The second scene shows an approaching train from the deer's first-person view. To answer "Why does this orange animal have its eyes wide open?", the model must connect these two events, but GPT-4V failed. There are more examples in supplementary material supporting this observation. + +Moving Objects. Another group of hard cases involves moving objects, such as comparing the speed of two cars. As shown in Tab. 4, instance 191 records a police car overtaking a firefighting truck. However, GPT-4V perceives all frames as semantically similar and is unable to identify the faster vehicle. + +Overly General Responses. GPT-4V sometimes answers questions in a general context rather than addressing the specific content of given videos. For + +Question: When I am at the entrance, how can I reach the balcony that has three decorations on the table? [ID:200] + +![](images/de694994e0badecad62fd255f0f716c58d338507bc6bac40d3ba88cc3b838f2c.jpg) + +![](images/16545f4e842034deeb652a9ceaec13a59f268bf3dcb7d1b0899ab513f22e8dc4.jpg) + +![](images/90b60928e53313c8d0619fe373c657667973d12255d7fc0d63a285017ea7666e.jpg) + +![](images/693cbdc5aa6edcfcb3c9cc79b8fcc85b68848fe9abf51d5c3e13c11043d9ae4b.jpg) + +![](images/76b9dc500c776af4f313bf377db02a44377894051bd5d50603b9ffef4a2372c1.jpg) + +![](images/62c56d4401c1627b4d3299b5e67cdaeb590381516be9a9ed18c5a06f307e1438.jpg) + +![](images/4cd5cdb4c8a9d94a7c9281a9c9ce262491165e780b5e2399f6fe0c0244f10249.jpg) + +![](images/9c4517d63728cc72a83b0d9f477867d410f1d32a0b423e66d1d458d0b1b7f8e0.jpg) + +![](images/e6aa00eb7bf05b0644e761b55d9dbeac352304791139424ca1c28ebec90fb757.jpg) + +Sample Answer: Upon entering through the door, you can first turn left into the kitchen, and then turn right. At this point, you are facing the target balcony. All you need to do is continue straight, open the balcony door, and step outside. + +GPT-4V: To reach the balcony with three decorations on the table, enter the house through the main entrance, proceed straight ahead through the living room, and then turn left to go through the double doors leading to the outdoor covered patio area. + +Question: What action did the third parrot from the left perform? + +[ID:43] + +![](images/c18c133a1fd93831c7bb94d7611875a9ec8b5c3cc148bb567e880e983c8ad83a.jpg) + +![](images/d1e5ef6631b8f54a2084a449a94983f1b408f0f5dde36aae7512fbe06b79890c.jpg) + +![](images/825c4162f43e1aa204b7a102844f6176590716da20a135b03cf4b4827adaa37b.jpg) + +![](images/c3d61df45c7df4d93b31d40974183a134ab5053d11c36750b49d2081ac417741.jpg) + +![](images/d40e0f4e8153f40eb80eacb1198afef9aa55b5fcf937d117e8d9366199f2f3d2.jpg) + +![](images/a6e61392e7a0347ddd977398cfe318ca0d55ff5333ec457827d2fc438a397037.jpg) + +![](images/c1ff13fce68bef9f0d5300610a6a3118c0aa8fa231ed08f2cef227c65630d15a.jpg) + +![](images/db818d7fee07e35b22a3ac1ed3e117d3ccd9902663381d02d1a73875e2e4f820.jpg) + +Sample Answer: It flew into the person's hand. + +GPT-4V: The third parrot from the left appears to be perched on a hand and interacting with a person. It looks like the parrot might be engaged in some sort of play or social interaction with the individual. + +Question: Why are the eyes of this orange animal wide open? + +[ID:112] + +![](images/ddccb5b3fcd7a42577a4e7fb6d289ec0567dc2fe6957646cd0e5c903e2d42b22.jpg) + +![](images/b3fce0b1c6e040f5fccb774f4cf55d765c6d0984fc33589df1ff9e56c61bc5c7.jpg) + +![](images/76378cdbae2f59b234c91ae3b573031ac71049d8d63fbdce26fbda8fab9a39e8.jpg) + +![](images/1e44f5dd026a6cbac11f78e79a3e736a598da582b1f1708c54daae0958a9b003.jpg) + +![](images/b1212bcec6ee662d652a401314bb0718a6ab042d1c73427517f901b1bbcc6164.jpg) + +![](images/b77219fc07699635591815e78e02f785c8a66b5d685a3aed264d0ca71ae107f2.jpg) + +![](images/36d21ad8274bcf3b4f6d2193456e8d16ce1886626781beba4fdc3119ba1c8097.jpg) + +![](images/c759aa760f4a9e3bcb6e354b2bd8c6b93f27b60384b8a61bc610006026add116.jpg) + +Sample Answer: Because the animal saw a train approaching, the white creature on the tracks was in danger. + +GPT-4V: The orange animal appears to be a stylized or animated character, possibly showing surprise, alarm, or excitement. Wide open eyes in animation are often used to convey strong emotions or reactions. This could be in response to something happening in the scene or to depict a particular personality trait of the character. Without additional context, it's hard to specify exactly why, but wide eyes are a common visual cue for heightened emotions in animation. + +Question: In the video, which car is moving at a slower pace? + +[ID:191] + +![](images/5325daf7ff6016eb239556c6dd2dbb01f9dea11817078bb9200cf8cc629f062a.jpg) + +![](images/2e757ee2c41f97a227a6fa2ba2fd6687e0c4d7a40b2f8e1d2a8f73e976af8d52.jpg) + +![](images/c7ac256d89a3fc030ae40e5549c9cafa61ffca8e5fb32a9a54eda8dc881c0a31.jpg) + +![](images/7c08fa2d1c1010e5c5f151f4826a16b733d3d49a4aeb2e0ab855897968caab59.jpg) + +![](images/9e470c6a282ed81326b2d70a69719f9137e5d970749efabb7af2d7d4608a75eb.jpg) + +![](images/603bcd71f7bce2763da91383ce7773468b2591b0df6663bc9fd00b8228671190.jpg) + +![](images/a4da0c08eab688ab9b636ce88130955c164b007488dfe57b157508519f9fd1f8.jpg) + +![](images/6b3a95950591d1df9ae4f8653502c0748403e2c976042048110133c88475cbec.jpg) + +Sample Answer: The police car. + +GPT-4V: I'm unable to view videos or discern the speed of objects within them. The image you've provided is static, so it doesn't convey the speed of the cars. If you have a specific question about the details within the image, feel free to ask! + +Question: Why are the audience members standing up? + +[ID:97] + +![](images/88c421d77bf1c152a7f1d9a52bc1a3e9317747abdf27384daad21b1cf97396c3.jpg) + +![](images/fd30662b93930e5b501518d617b6c64d1d6e05e3f9b7bf3d6920fd0ad5cd1697.jpg) + +![](images/534329cc39713f0d03e1015aa93a11d8130eefd89844af917a57a2c8f3bc4d7f.jpg) + +![](images/44cb96792b1f03b84cf1bf26bf8e73de1df72c13be6944f7f52b999ac0fc5744.jpg) + +![](images/e6a0b04593e19a1aa404b9075180cefa2956a3812f40ec16e38da64b67926430.jpg) + +![](images/75541d58c1efc202c91601f86d38c575e44fb7fa91e7163ba30e6491bf151a14.jpg) + +![](images/92356d737e3ebf918b514d9da2d2c21eeb19626b8c4163680cffd29c4e080d79.jpg) + +![](images/3c828c0869d0b90ef6c555aa21326cb83e4de67c644d6a47d4b56fcd1faa3878.jpg) + +Sample Answer: The player in the blue team uniform scored. + +GPT-4V: The audiences in sports video games, like the one depicted in the images, often stand up in response to exciting moments or plays. This is a way to simulate real-life reactions of fans during intense or critical moments in a game, such as a close score, a game-winning shot, or a spectacular play. The game developers include these reactions to enhance the realism and immersion of the game experience for players. + +Table 4: Examples of Incorrect Responses from GPT-4V. + +Question: What was the score of the team in black in the video footage? + +![](images/15541733488d8b987a2402afff3a6bf9356d809a24ae1a3848816607640d439f.jpg) + +![](images/f0bfaa4faa2fcf376899fde15fbf44c6d6264ba9a27ae627dfe91cb1cbfbce3e.jpg) + +![](images/e2452a077ddd57bba54c4c11e913b0f6eb3e1dde6cab383ea7a484e653031f93.jpg) + +![](images/72e6a56aa63819866f901b36b6fffec58663dca3fb9f5bfd161528625097c6e0.jpg) + +![](images/eb71f5d6acead61ed1d57dcea26a0a1c1e89add7d339dc5d0eea3a197a647ac4.jpg) + +![](images/ddfb573517526e7564a72120435e5d29ca778dc91d8535fb68574fc14856e4b4.jpg) + +![](images/50020e358c933046a906b8d9a7e7971767b62401fdab3620e750c4fb4d6a9f1d.jpg) + +![](images/34314a702dfa8f485a94238713240575cccd454cc7fa8e61cc9167064d7b5a08.jpg) + +Sample Answer: The team in black scored eight points. + +Human: The black team scores 3 points in the video. + +Question: How did the character detect the enemy? + +![](images/f2a604066ecb847ecd92970a7b806752da3315c439359e848db7983de9e95499.jpg) + +![](images/43e7b3734abcdf9fb359487fcf8a686a6f486c5fa707dfc3f3b0ea80ca8dc84e.jpg) + +![](images/aa2884480050e99c2bc0d93097873a8b05625c6e882e64e532f4ff000fe02a45.jpg) + +![](images/c2951fa477b119e619251b2d1be3922daefc9668daea3137d9268d158c24185d.jpg) + +![](images/35df2db282d514bcee5cb10211f11a5eebe24a5226266127d760653cb86a46da.jpg) + +![](images/7d881cb3e4a1e39ebd0425df0799fef23c6c5426817587ad6c3541cb1c9ec22c.jpg) + +![](images/a4ae9af5c50fb3b72be868af5d1d6dde9f1878a5cc3ccc69cff4bb09b02b9f8b.jpg) + +![](images/555e7f9ed738d6baaa749dfb0762f167189e1d37f79dd1a06066cc07887cda4b.jpg) + +Sample Answer: The enemy shot the character so the enemy's location could be known. + +Human: The character used the scope on the sniper rifle to clearly observe distant enemies. + +Question: Why did this person initially appear stunned and then suddenly burst into laughter? + +[ID:133] + +![](images/03b370152bbf7663cb168b3f51b4f2cec54e87530bd949c0c167c97396e56243.jpg) + +![](images/23f7380e3094b2e9bc213cbeca3e70a9a638267b6dd00bb2dfd9d6d4ec216581.jpg) + +![](images/7d45efc847118d000543d1fc8cd0a73fd9207a519017acf1091cabb392f0ecb1.jpg) + +![](images/dfc49fe608f1f191956eda529d9628383199b8186aad87641423131facd8747b.jpg) + +![](images/5ecfe81cf6be51d2725dc3fd27d6006a2137b793276dcac198fd4b3485f078fd.jpg) + +![](images/06290bf9bd3a6e29b93e2ed0bbb1069707f7e96ef51accb20eea7db060085d06.jpg) + +![](images/e3b7ff46583e97b391adee657884a3b31759501db2038d711486e461af036a03.jpg) + +![](images/26923a9240379e716d67fa87e264ba5cf447ba304d461daf1e75998860c23a4a.jpg) + +Sample Answer: Because the person threw grenades to kill the enemy but knocked a teammate down accidentally. + +Human: Because she didn't expect that the bomb she threw would explode and harm someone. + +# Table 5: Hard cases for humans due to insufficient external knowledge. + +instance, in instance 97 shown in Tab. 4, GPT-4V answers "Why are the audience members standing up?" in a general way, even though there is an obvious and concrete reason. Similarly, instance 197 gives a general guide for operating a coffee machine even though the video demonstrates a specific one. It implies that GPT-4V tends to rely on its strong language model to produce a safe answer, sometimes ignoring key details in the video. + +# 4.5 Hard Cases for Humans + +Humans achieve an accuracy rate of $72.8\%$ on AutoEval-Video. We conduct an in-depth analysis of these cases and discover that there are two major hard cases for humans: questions related to external knowledge and questions requiring accurate description and summarizing ability. + +Questions Related to External Knowledge. Some questions in AutoEval-Video require a relatively high level of mastery of external knowledge in specific domains such as sports and games, resulting in human difficulties. For example, in Tab. 5, instance 38 shows a misunderstanding due to human participants not knowing basketball scoring rules. Similarly, instance 133 requires explaining a game streamer's laughter, an easy task for someone familiar with the game but confusing for others. However, considering the cognitive level of the general public, these shortcomings are reasonable and expected. + +Accurate Description and Summarizing Ability. Questions requiring detailed, step-by-step video descriptions are more challenging for humans compared to semantic reasoning problems due to their complexity and precision. For instance, as shown in Tab. 6, in instance 173, human makes careless mistakes by ignoring necessary CPU installation steps shown in the video. In instance + +143, errors occur due to insufficient observation, misidentifying corn kernels as soybeans. This phenomenon is common in many step-by-step descriptive tasks. + +Question: How can you prepare a dish using these ingredients and tools as demonstrated in the video? [ID:143] + +![](images/2f3a7d63bc41154e6550ba9cc5383769134b66aa16c2eb85f9fbdeeb3dad4da6.jpg) + +Sample Answer: Cut the potatoes, beets, purple cabbage, carrots, mushrooms, and bell peppers, put them in the steamer with the corn kernels, cut the gluten into strips, heat the steamer on the gas stove, and put the gluten into the air fryer. + +Human: The kitchen utensils needed for preparation include a cutting board, a chef's knife, a small steamer pot, and an air fryer. The preparation process for the ingredients is as follows: 1. Cut the potatoes into small chunks. 2. Slice the leafy greens into segments. 3. Chop the purple cabbage. 4. Slice the carrots into thick pieces. 5. Halve the white mushrooms. 6. Slice the bell peppers. 7. Prepare a small bowl of soaked soybeans. 8. Cut the blanched beef pieces into strips. Finally, Place the vegetables from steps 1-7 into the steamer pot, and place the beef strips from step 8 into the air fryer. + +Question: How to install a CPU on a motherboard as demonstrated in the video? [ID:173] + +![](images/4a8dbd0856ce054046a3c41e3d8e554ab80bfa799f73e02ec3179b989064722a.jpg) + +Sample Answer: According to the clip, here is how to install a CPU on a motherboard: 1. Lift the metal arm next to the CPU socket. 2. Look for a small triangle indicator on the corner of the CPU and the socket and align them properly. 3. Hold the CPU by the sides and insert the CPU into the CPU socket on the motherboard. 4. Lower the arm down completely into its original position. + +Human: 1. Open the latch at the CPU location on the motherboard. 2. Place the side of the CPU with pins onto the corresponding location on the motherboard, aligning with the notches. 3. Close the latch. + +![](images/300c393ea160633bb5c97d40f917e1928e8ff235ede155f73ada9d5d268f0c1a.jpg) + +Sample Answer: First, put the pillow at the head of the bed; Second, put the clothes that need to be washed on the floor, and put the clothes that do not need to be washed on the chair; Third, pile up all the debris except for sheets quilts, and pillows on the floor; Fourth, after arranging the sheets and quilts, place the doll bear on the side near the head of the bed. + +Human: The order in which this person organizes the room is as follows: 1. Neatly arrange multiple pillows at the head of the bed. 2. Gather multiple items left on the bed onto a chair. 3. Remove other items from the bed, such as a large teddy bear and a white cushion. 4. Straighten the bedsheet to make it look even. 5. Place the teddy bear back at the head of the bed. + +Table 6: Hard cases for humans related to descriptive ability. + +# 5 Conclusion and Future Work + +In this paper, we introduce AutoEval-Video, a novel benchmark for comprehensively assessing large VLMs in open-ended video question answering. Utilizing an LLM evaluator, models can be automatically and accurately evaluated on AutoEval-Video. In particular, GPT-4V achieves the highest accuracy of $32.2\%$ , still far from human accuracy of $72.8\%$ . Through an extensive case study, we identify several limitations of current models' capabilities in video understanding, such as temporal and dynamic comprehension, and overly general responses. These findings will shape our future research direction. + +# Acknowledgment + +We would like to express our sincere gratitude to the reviewers of ECCV 2024 for their insightful and constructive feedback. Their valuable comments have greatly contributed to improving the quality of our work. + +# References + +1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675 (2016) +2. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, A., Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z., Zhang, J., Zhang, X., Zhang, Y., Zhang, Z., Zhou, C., Zhou, J., Zhou, X., Zhu, T.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023) +3. Buch, S., Eyzaguirre, C., Gaidon, A., Wu, J., Fei-Fei, L., Niebles, J.C.: Revisiting the "video" in video-language understanding. In: CVPR. pp. 2917-2927 (2022) +4. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR. pp. 961-970 (2015) +5. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) +6. Fan, H., Lin, L., Yang, F., Chu, P., Deng, G., Yu, S., Bai, H., Xu, Y., Liao, C., Ling, H.: Lasot: A high-quality benchmark for large-scale single object tracking. In: CVPR. pp. 5374-5383 (2019) +7. Fu, J., Ng, S.K., Jiang, Z., Liu, P.: Gptscore: Evaluate as you desire. arXiv preprint arXiv:2302.04166 (2023) +8. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The "something something" video database for learning and evaluating visual common sense. In: ICCV. pp. 5842-5850 (2017) +9. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR. pp. 6904-6913 (2017) +0. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022) +1. Grunde-McLaughlin, M., Krishna, R., Agrawala, M.: Agqa: A benchmark for compositional spatio-temporal reasoning. In: CVPR. pp. 11287-11297 (2021) +2. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., et al.: Ava: A video dataset of spatiotemporally localized atomic visual actions. In: CVPR. pp. 6047-6056 (2018) +3. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) + +14. Hu, A., Chen, S., Zhang, L., Jin, Q.: Infomatic: An informative metric for reference-free image caption evaluation. arXiv preprint arXiv:2305.06002 (2023) +15. Huang, D.A., Ramanathan, V., Mahajan, D., Torresani, L., Paluri, M., Fei-Fei, L., Niebles, J.C.: What makes a video a video: Analyzing temporal information in video understanding models and datasets. In: CVPR. pp. 7366-7375 (2018) +16. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) +17. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV. pp. 706-715 (2017) +18. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: ICCV. pp. 2556-2563. IEEE (2011) +19. Lei, J., Berg, T.L., Bansal, M.: Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428 (2022) +20. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023) +21. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +22. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023) +23. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023) +24. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023) +25. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL (2004) +26. Lin, Y.T., Chen, Y.N.: Llm-eval: Unified multi-dimensional automatic evaluation for open-domain conversations with large language models. arXiv preprint arXiv:2305.13711 (2023) +27. Liu, D., Qu, X., Hu, W.: Reducing the vision and language bias for temporal sentence grounding. ACM MM (2022), https://api_semanticscholar.org/CorpusID:251104851 +28. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565 (2023) +29. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) +30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +31. Liu, Y., Iter, D., Xu, Y., Wang, S., Xu, R., Zhu, C.: G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634 (2023) +32. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023) +33. OpenAI: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL. pp. 311-318 (2002) + +35. Pătrăucean, V., Smaira, L., Gupta, A., Continente, A.R., Markeeva, L., Banarse, D., Koppula, S., Heyward, J., Malinowski, M., Yang, Y., et al.: Perception test: A diagnostic benchmark for multimodal video models. arXiv preprint arXiv:2305.13786 (2023) +36. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) +37. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023) +38. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) +39. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR. pp. 4566-4575 (2015) +40. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y.F., Wang, W.Y.: Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In: ICCV. pp. 4581-4591 (2019) +41. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. NeurIPS 35, 24824-24837 (2022) +42. Wu, Y., Lim, J., Yang, M.H.: Online object tracking: A benchmark. In: CVPR. pp. 2411-2418 (2013) +43. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR. pp. 5288-5296 (2016) +44. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023) +45. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023) +46. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023) +47. Yu, Z., Zheng, L., Zhao, Z., Wu, F., Fan, J., Ren, K., Yu, J.: Anetqa: A large-scale benchmark for fine-grained compositional reasoning over untrimmed videos. In: CVPR. pp. 23191-23200 (2023) +48. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023) +49. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023) +50. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023) +51. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754 (2023) \ No newline at end of file diff --git a/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/images.zip b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b9a0613ffce25bd56d9f8eed19fc19dc85d7bfa4 --- /dev/null +++ b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0648b1eb1d0fe3ee7b0d9b78d1ecb627f94ff607a86c725557b8bd688e566bfd +size 558617 diff --git a/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/layout.json b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f19d2280c0555f693fc5b8b8389a014c60e7085f --- /dev/null +++ b/2024/AutoEval-Video_ An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering/layout.json @@ -0,0 +1,13656 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 166 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 166 + ], + "type": "text", + "content": "AutoEval-Video: An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "spans": [ + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "type": "text", + "content": "Xiuyuan Chen" + }, + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "type": "text", + "content": ", Yuan Lin" + }, + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "type": "text", + "content": ", Yuchen Zhang" + }, + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "type": "text", + "content": ", and Weiran Huang" + }, + { + "bbox": [ + 147, + 188, + 466, + 201 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 209, + 476, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 209, + 476, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 209, + 476, + 235 + ], + "type": "text", + "content": "1 MIFA Lab, Qing Yuan Research Institute, SEIEE, Shanghai Jiao Tong University \n2 ByteDance Research" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "spans": [ + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "text", + "content": "Abstract. We propose a novel and challenging benchmark, AutoEval-Video, to comprehensively evaluate large vision-language models in open-ended video question answering. The comprehensiveness of AutoEval-Video is demonstrated in two aspects: 1) AutoEval-Video constructs open-ended video-questions across 9 skill dimensions, addressing capabilities of perception, comprehension, and generation; 2) AutoEval-Video contains newly collected videos that cover over 40 distinct themes. To efficiently evaluate responses to the open-ended questions, we employ an LLM-based evaluation approach, but instead of merely providing a reference answer, we annotate unique evaluation rules for every single instance (video-question pair). To maximize the robustness of these rules, we develop a novel adversarial annotation mechanism. By using instance-specific rules as prompt, GPT-4, as an automatic evaluator, can achieve a stable evaluation accuracy of around " + }, + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "inline_equation", + "content": "97.0\\%" + }, + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "text", + "content": ", comparable to the " + }, + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "inline_equation", + "content": "94.9\\% - 97.5\\%" + }, + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "text", + "content": " accuracy of a human evaluator. Furthermore, we assess the performance of eleven large vision-language models on AutoEval-Video. Among them, GPT-4V(ison) significantly outperforms other models, achieving an accuracy of " + }, + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "inline_equation", + "content": "32.2\\%" + }, + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "text", + "content": ". However, there is still substantial room for improvement compared to human accuracy of " + }, + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "inline_equation", + "content": "72.8\\%" + }, + { + "bbox": [ + 159, + 258, + 455, + 511 + ], + "type": "text", + "content": ". By conducting an extensive case study, we uncover several drawbacks of GPT-4V, such as limited temporal and dynamic comprehension, and overly general responses. Code is available at https://github.com/Xiuyuan-Chen/AutoEval-Video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 529, + 230, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 230, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 230, + 542 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 553, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 626 + ], + "type": "text", + "content": "Generative large vision-language models [2, 5, 21, 22, 29, 30, 32, 48] have exhibited the considerable potential to address fundamental challenges in video understanding, temporal and causal comprehension [3]. Furthermore, these models can deal with more complex open-ended video tasks, surpassing traditional specialized video tasks, like video captioning [17, 40, 43], action classification [8, 12, 36], and object tracking [6, 42]." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 632, + 373, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 373, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 373, + 643 + ], + "type": "text", + "content": "* Work done during an internship at ByteDance Research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "† Correspondence to Yuan Lin, Yuchen Zhang, and Weiran Huang (weiran.huang@outlook.com)." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 313 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 313 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 313 + ], + "type": "table", + "html": "
SkillExample of Video-Question
Dynamic PerceptionThe video depicts various ways of relaxing and de-stress.\nQuestion: Please List the three relaxation techniques from the video in order.
State Transitions PerceptionThe video captures the process of a chemistry experiment.\nQuestion: What is the change in the color of the liquid after blowing?
Camera Movement PerceptionThe camera's position in the video undergoes a series of adjustments.\nQuestion: What changes have been made to the position of the camera used for filming this video?
Explanatory ReasoningThe camera records a basketball player sinking a shot and flashing a V sign to the camera.\nQuestion: Why is the individual wearing blue jeans in the video making that particular gesture?
Counterfactual ReasoningA purple animal character in the cartoon follows the guidance of another character and falls into a trap.\nQuestion: What action could the purple animal take to avoid the consequences shown in the video?
Predictive ReasoningThe video records a person putting groceries into a cupboard and preparing to close the cupboard door (but without any depiction of the door-closing action).\nQuestion: What the person will do next?
Comparison ReasoningThe video sequentially depicts several individuals drinking water, where one person exhibits discomfort while the others appear normal.\nQuestion: How does the clothing of the uncomfortable person differ from that of the others?
Reasoning with External KnowledgeThe video records a basketball game, in which one team scored two three-pointers and a two-pointer.\nQuestion: In the video footage, what is the score of the team wearing black?
DescriptionThe video highlights two additional features of a vacuum cleaner product.\nQuestion: Sequentially describe the two extra features of the vacuum cleaner.
", + "image_path": "239d7d0839f6cf035935b13fbf3c4e0fb73a199a89b685e31fa1b2fe98f657c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 321, + 481, + 342 + ], + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 342 + ], + "type": "text", + "content": "Table 1: Skill Dimensions and Corresponding Examples for Video-Questions in AutoEval-Video." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 357, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 481, + 514 + ], + "type": "text", + "content": "Effectively evaluating large vision-language models is crucial for improving their performance, as it identifies areas for improvement and enables the comparison and investigation of various models. Several benchmarks [6,8,10,12,17, 35,36,40,42,43] have been developed to evaluate capabilities of specific tasks in video understanding. Nonetheless, they fail to assess model performance in a comprehensive manner. To close this gap, recent studies [10,20,35] aim to build benchmarks covering a wide range of skills. These benchmarks primarily consist of multiple-choice question-answers to facilitate effective evaluation. However, in comparison to open-ended question answering, multiple-choice question answering presents two key limitations: 1) they simplify the tasks, and 2) they do not accurately reflect real-world scenarios. For instance, in a video-based dialogue, responses should be generated directly, rather than making a selection from a set of options." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "content": "In this paper, we introduce AutoEval-Video, a comprehensive and challenging benchmark designed to evaluate the video understanding capabilities of large vision-language models. AutoEval-Video comprises 327 challenging open-ended video-question instances that span across nine skill dimensions in Tab. 1, which address video-specific perception, comprehension, and generation skills. All videos in AutoEval-Video are newly collected from YouTube, including over 40 genres such as daily life, gaming, sports, and more. We also establish rigorous annotation standards to guarantee the difficulty and quality of our benchmark. These standards include minimizing single-frame bias [3, 15, 19] and language bias [9, 27], and setting objective questions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 636, + 481, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 636, + 481, + 661 + ], + "spans": [ + { + "bbox": [ + 131, + 636, + 481, + 661 + ], + "type": "text", + "content": "Assessing responses to open-ended questions can be difficult, especially when examining outputs generated by LLMs. These responses may involve additional" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 342, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 342, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 342, + 102 + ], + "type": "text", + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "text", + "content": "information such as video description and chain-of-thought [41] analysis, diminishing the reliability of traditional automatic evaluation metrics, which typically involve comparing a given response to a reference answer, such as BLEU [34], ROUGE [25] and CIDEr [39]. To address this issue, we employ an LLM-based evaluation approach [7, 31, 45, 49]. An LLM evaluator comprises a strong LLM and its prompting strategy. Existing works [26, 46, 50] design a unified prompt for specific tasks. However, due to the diversity of videos and questions in AutoEval-Video, a single unified prompt is insufficient for accurate evaluation. To tackle this challenge, we meticulously annotate distinct evaluation rules for each instance instead of solely providing a reference answer. To improve the precision of these rules, we develop a novel adversarial annotation mechanism. A red team is engaged to devise unaddressed corner cases that the rules may not cover, assisting the annotation team in refining them. The final rule is ultimate when the red team cannot discover any further corner cases that break the rules. After intensive annotation, using instance-specific rules as prompt, GPT-4 can serve as an automatic evaluator, achieving an evaluation accuracy of " + }, + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "inline_equation", + "content": "97.0\\%" + }, + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "text", + "content": ", comparable to the single human accuracy of " + }, + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "inline_equation", + "content": "97.5\\%" + }, + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 321, + 481, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 321, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 133, + 321, + 481, + 440 + ], + "type": "text", + "content": "Leveraging AutoEval-Video, we have conducted a comprehensive evaluation of eleven prominent large vision-language models, which fall into two categories: ImageLLM and VideoLLM. Our assessment reveals that GPT-4V(ison) significantly surpasses other models. Utilizing 8 randomly sampled frames as input, GPT-4V achieves an accuracy of " + }, + { + "bbox": [ + 133, + 321, + 481, + 440 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 133, + 321, + 481, + 440 + ], + "type": "text", + "content": ", compared to " + }, + { + "bbox": [ + 133, + 321, + 481, + 440 + ], + "type": "inline_equation", + "content": "72.8\\%" + }, + { + "bbox": [ + 133, + 321, + 481, + 440 + ], + "type": "text", + "content": " accuracy of humans, indicating considerable potential for improvement. Moreover, we discover that by increasing the input to 16 frames, GPT-4V's performance markedly improves to " + }, + { + "bbox": [ + 133, + 321, + 481, + 440 + ], + "type": "inline_equation", + "content": "32.2\\%" + }, + { + "bbox": [ + 133, + 321, + 481, + 440 + ], + "type": "text", + "content": ". Finally, we perform an in-depth case study for GPT-4V, identifying several critical challenges that demand further investigation, such as limited temporal and dynamic understanding, and overly general responses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 464, + 236, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 236, + 477 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 236, + 477 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 495, + 481, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 495, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 133, + 495, + 481, + 591 + ], + "type": "text", + "content": "Large Vision-Language Models. The remarkable success of LLMs has been accompanied by significant progress in large vision-language models. Recent works [2, 5, 21, 29] improve the visual perception and cognition capabilities by utilizing the strong understanding and generation abilities of LLMs. In addition, some research [22, 32, 48] further incorporates video inputs and leverages the extensive potential of LLMs for video understanding. The considerable progress necessitates more challenging and comprehensive benchmarks to evaluate the video comprehension capabilities of large vision-language models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 593, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 593, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 593, + 481, + 665 + ], + "type": "text", + "content": "Benchmarks for Video Understanding. Several video benchmarks have been employed to assess model performance in various areas, such as action recognition [11, 12, 47], video captioning [13, 14, 43], and action classification [6, 8, 16, 18, 36, 42]. However, these benchmarks do not provide a comprehensive assessment of a model's capabilities in video understanding. To address this issue, some general benchmarks [20, 35] have been established, developing" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 123, + 476, + 160 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 313, + 123 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 313, + 123 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 313, + 123 + ], + "type": "text", + "content": "Q: What are the steps to prepare this dish as shown in the video?" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 123, + 476, + 160 + ], + "lines": [ + { + "bbox": [ + 138, + 123, + 476, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 123, + 476, + 160 + ], + "type": "image", + "image_path": "2dfc754396e327471d28713cc48f2c538ae1e0cac0ecd76b2555559c3f0aa801.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 299, + 163, + 315, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 163, + 315, + 170 + ], + "spans": [ + { + "bbox": [ + 299, + 163, + 315, + 170 + ], + "type": "text", + "content": "Rules" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 171, + 474, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 474, + 188 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 474, + 188 + ], + "type": "text", + "content": "[Video Description]: In the video, a person first uses a knife to cut tofu into cubes on a cutting board. Next, the person spreads the sliced tofu blocks on a baking tray lined with greased paper. Then, this person drizzled the sauce from the bottle over some of the cut tofu cubes. Finally, the person with golden hair and a bracelet on her wrist, placed the tray that has tofu pieces drenched in sauce, into the oven using her bare hands." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 188, + 290, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 188, + 290, + 194 + ], + "spans": [ + { + "bbox": [ + 139, + 188, + 290, + 194 + ], + "type": "text", + "content": "[Video Question]: What are the steps to prepare this dish as shown in the video?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 194, + 358, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 194, + 358, + 199 + ], + "spans": [ + { + "bbox": [ + 139, + 194, + 358, + 199 + ], + "type": "text", + "content": "Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 200, + 474, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 200, + 474, + 210 + ], + "spans": [ + { + "bbox": [ + 139, + 200, + 474, + 210 + ], + "type": "text", + "content": "1. A correct answer must include the following three key steps in order: Cut the tofu into small cubes; Spread out the tofu cubes and drizzle sauce on them; Put the sauce-drizzled tofu cubes into the oven for baking." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 210, + 446, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 210, + 446, + 217 + ], + "spans": [ + { + "bbox": [ + 139, + 210, + 446, + 217 + ], + "type": "text", + "content": "2. The answer to the description of the shape of tofu blocks can be anything similar to 'block'. It doesn't necessarily have to be described as a perfectly shaped cube." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 217, + 347, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 217, + 347, + 222 + ], + "spans": [ + { + "bbox": [ + 139, + 217, + 347, + 222 + ], + "type": "text", + "content": "3. Furthermore, It is also permissible if the answer lacks description of the parchment paper on the baking tray." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 222, + 474, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 222, + 474, + 232 + ], + "spans": [ + { + "bbox": [ + 139, + 222, + 474, + 232 + ], + "type": "text", + "content": "4. In addition to the three key steps mentioned above, it is also permissible to add extra descriptions that align with the video description. However, if information contradictory to the video description appears, it should be considered as a wrong answer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 232, + 424, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 232, + 424, + 238 + ], + "spans": [ + { + "bbox": [ + 139, + 232, + 424, + 238 + ], + "type": "text", + "content": "5. The response should contain correctly spelled words. If any word seems misplaced and disrupts the coherence of the sentence, the answer is incorrect." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 238, + 211, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 238, + 211, + 243 + ], + "spans": [ + { + "bbox": [ + 139, + 238, + 211, + 243 + ], + "type": "text", + "content": "[Answer to be judged]: " + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 243, + 181, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 243, + 181, + 249 + ], + "spans": [ + { + "bbox": [ + 139, + 243, + 181, + 249 + ], + "type": "text", + "content": "[Form of the result]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 139, + 249, + 474, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 249, + 474, + 260 + ], + "spans": [ + { + "bbox": [ + 139, + 249, + 474, + 260 + ], + "type": "text", + "content": "Please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 192, + 264, + 205, + 272 + ], + "blocks": [ + { + "bbox": [ + 192, + 264, + 205, + 272 + ], + "lines": [ + { + "bbox": [ + 192, + 264, + 205, + 272 + ], + "spans": [ + { + "bbox": [ + 192, + 264, + 205, + 272 + ], + "type": "image", + "image_path": "3a1230a3fb782e2e315ae6b93ea9f67b033ca5ccef49d7e27d1c6d861ab8f5f6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 179, + 276, + 206, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 276, + 206, + 281 + ], + "spans": [ + { + "bbox": [ + 179, + 276, + 206, + 281 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 139, + 282, + 246, + 310 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 139, + 282, + 246, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 282, + 246, + 292 + ], + "spans": [ + { + "bbox": [ + 139, + 282, + 246, + 292 + ], + "type": "text", + "content": "1. Cut the tofu into pieces and then place it on a wooden board." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 139, + 293, + 218, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 293, + 218, + 298 + ], + "spans": [ + { + "bbox": [ + 139, + 293, + 218, + 298 + ], + "type": "text", + "content": "2. Drizzle the sauce over some tofu slices." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 139, + 298, + 246, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 298, + 246, + 310 + ], + "spans": [ + { + "bbox": [ + 139, + 298, + 246, + 310 + ], + "type": "text", + "content": "3. Put on gloves and place the baking tray with tofu pieces into the oven." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 309, + 277, + 320, + 282 + ], + "blocks": [ + { + "bbox": [ + 309, + 277, + 320, + 282 + ], + "lines": [ + { + "bbox": [ + 309, + 277, + 320, + 282 + ], + "spans": [ + { + "bbox": [ + 309, + 277, + 320, + 282 + ], + "type": "image", + "image_path": "d0327b2a3132818e7f86b10f351db6626d04bdd5b320df59eb6776aa858adc33.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 256, + 282, + 356, + 311 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 256, + 282, + 356, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 282, + 356, + 297 + ], + "spans": [ + { + "bbox": [ + 256, + 282, + 356, + 297 + ], + "type": "text", + "content": "1. Cut the tofu into small cubes and spread them out. \n2. Place the tofu cubes on a baking tray and drizzle with sauce." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 256, + 298, + 355, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 298, + 355, + 311 + ], + "spans": [ + { + "bbox": [ + 256, + 298, + 355, + 311 + ], + "type": "text", + "content": "3. Put the sauce-drizzled tofu cubes into the oven to bake." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 406, + 277, + 433, + 281 + ], + "blocks": [ + { + "bbox": [ + 406, + 277, + 433, + 281 + ], + "lines": [ + { + "bbox": [ + 406, + 277, + 433, + 281 + ], + "spans": [ + { + "bbox": [ + 406, + 277, + 433, + 281 + ], + "type": "image", + "image_path": "ce1d568fa0900963a23c2bab841f309f9fcd171755974ff06f96d795c50f4bfd.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 367, + 282, + 474, + 304 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 367, + 282, + 474, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 282, + 474, + 293 + ], + "spans": [ + { + "bbox": [ + 367, + 282, + 474, + 293 + ], + "type": "text", + "content": "1. Cut the white tofu into pieces, then place them in a baking tray." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 367, + 293, + 453, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 293, + 453, + 299 + ], + "spans": [ + { + "bbox": [ + 367, + 293, + 453, + 299 + ], + "type": "text", + "content": "2. Drizzle sauce over some of the tofu pieces." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 367, + 299, + 471, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 299, + 471, + 304 + ], + "spans": [ + { + "bbox": [ + 367, + 299, + 471, + 304 + ], + "type": "text", + "content": "3. Place the baking tray with the tofu pieces in the oven." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 187, + 313, + 200, + 331 + ], + "blocks": [ + { + "bbox": [ + 187, + 313, + 200, + 331 + ], + "lines": [ + { + "bbox": [ + 187, + 313, + 200, + 331 + ], + "spans": [ + { + "bbox": [ + 187, + 313, + 200, + 331 + ], + "type": "image", + "image_path": "b45f7bbfc1fcb8f40bc799d967014199ea9ef5dc3c00f37b239e3b12ceef4646.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 139, + 337, + 245, + 399 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 139, + 337, + 244, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 337, + 244, + 348 + ], + "spans": [ + { + "bbox": [ + 139, + 337, + 244, + 348 + ], + "type": "text", + "content": "The answer provided is incorrect for the following reasons:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 139, + 349, + 245, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 349, + 245, + 371 + ], + "spans": [ + { + "bbox": [ + 139, + 349, + 245, + 371 + ], + "type": "text", + "content": "1. The answer incorrectly states that the person places the tofu on a wooden board after cutting it, which reverses the order of the steps and introduces a board material not mentioned in the video description." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 139, + 371, + 245, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 371, + 245, + 382 + ], + "spans": [ + { + "bbox": [ + 139, + 371, + 245, + 382 + ], + "type": "text", + "content": "2. The sauce is misspelled as \"sauce,\" which is a spelling error." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 139, + 382, + 245, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 382, + 245, + 399 + ], + "spans": [ + { + "bbox": [ + 139, + 382, + 245, + 399 + ], + "type": "text", + "content": "3. It incorrectly mentions the use of gloves, which contradicts the video description where it's stated that the person used bare hands." + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 301, + 313, + 313, + 330 + ], + "blocks": [ + { + "bbox": [ + 301, + 313, + 313, + 330 + ], + "lines": [ + { + "bbox": [ + 301, + 313, + 313, + 330 + ], + "spans": [ + { + "bbox": [ + 301, + 313, + 313, + 330 + ], + "type": "image", + "image_path": "78a20e3ac12970a7ef7539bd85f8e6ea7aec210f305c7b1b577d53ca816a8879.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 255, + 337, + 357, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 337, + 357, + 395 + ], + "spans": [ + { + "bbox": [ + 255, + 337, + 357, + 395 + ], + "type": "text", + "content": "The [Answer to be judged] correctly includes the three key steps required by the rules: cutting tofu into cubes, spreading the tofu out and drizzling sauce on them, and putting the sauce-drizzled tofu into the oven to bake. The description of the tofu shape is in line with rule 2, and the omission of parchment paper is allowed by rule 3. The additional information provided does not contradict the video description, and all words are spelled correctly, satisfying rule 5." + } + ] + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 415, + 313, + 428, + 330 + ], + "blocks": [ + { + "bbox": [ + 415, + 313, + 428, + 330 + ], + "lines": [ + { + "bbox": [ + 415, + 313, + 428, + 330 + ], + "spans": [ + { + "bbox": [ + 415, + 313, + 428, + 330 + ], + "type": "image", + "image_path": "39f2b8bb76ab92cb8042a2b382e4741d15fe1c11e75a440949ce04188896f5bf.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 367, + 337, + 474, + 394 + ], + "type": "list", + "angle": 0, + "index": 42, + "blocks": [ + { + "bbox": [ + 367, + 337, + 474, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 337, + 474, + 360 + ], + "spans": [ + { + "bbox": [ + 367, + 337, + 474, + 360 + ], + "type": "text", + "content": "The answer provided correctly states the three key steps in the order they appear in the video description: cutting the tofu into pieces, drizzling sauce on the tofu, and placing the tray in the oven." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 367, + 360, + 474, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 360, + 474, + 394 + ], + "spans": [ + { + "bbox": [ + 367, + 360, + 474, + 394 + ], + "type": "text", + "content": "The shape of the tofu is described as \"pieces,\" which aligns with the allowance for any similar description to \"block.\" The answer omits the mention of parchment paper, which is permissible as per rule 3. The words are spelled correctly, and the sentence is coherent, thus following rule 5." + } + ] + } + ], + "index": 41 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 416, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 481, + 460 + ], + "type": "text", + "content": "Fig. 1: Example Instance in AutoEval-Video and Automatic Evaluation Process. Each instance has three components: video, question, and rules. For each video and question, we establish specific and detailed evaluation rules. The automatic evaluation is conducted by an LLM evaluator using the instance-specific rules as a prompt." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 132, + 465, + 481, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 481, + 501 + ], + "type": "text", + "content": "multiple-choice VQA to assess models' abilities across different aspects. Nevertheless, compared to open-ended question answering, multiple-choice question answering oversimplify tasks and may not reflect realistic instructions." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 132, + 502, + 481, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 563 + ], + "type": "text", + "content": "In addition, existing benchmarks mainly employ two methods for video collection: 1) sourcing videos related to a specific task from web [4,16] or large-scale video datasets [1,11,47]; 2) recording new videos based on pre-defined scripts [35] or themes [10]. However, both approaches suffer from limited visual diversity, leading to a video distribution that inadequate represents real-world scenarios." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 132, + 586, + 250, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 250, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 250, + 598 + ], + "type": "text", + "content": "3 AutoEval-Video" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 132, + 616, + 208, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 208, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 208, + 627 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "In AutoEval-Video, each instance comprises three components: video, question, and rules, as illustrated in Fig. 1. The construction involves three steps." + } + ] + } + ], + "index": 48 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 342, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 342, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 342, + 102 + ], + "type": "text", + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "First, to address the essential aspects of real-world multi-modal understanding, the uncovered areas of existing benchmarks, and the limitations of current models, we identify nine skill dimensions, as illustrated in Tab. 1. These dimensions encompass diverse skill areas (including perception, cognition, and generation) and various reasoning types (such as descriptive, explanatory, predictive, comparative, and counterfactual)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 479, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 479, + 237 + ], + "type": "text", + "content": "Second, for each skill dimension, an annotation team gathers videos from YouTube1 and crafts relevant open-ended questions. It is critical to prevent potential single-frame bias and language bias while formulating these questions. To ensure the assessability of responses, the questions must retain objectivity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 323 + ], + "type": "text", + "content": "Finally, after acquiring the video-question pairs, the annotation group develops evaluation rules for each instance. These rules establish precise conditions for considering whether a response is correct or wrong, involving precise criteria for recognizing seemingly accurate but flawed corner case responses, and vice versa. To ensure thoroughness, we employ a second annotation team, known as the \"red team\", to devise unaddressed corner cases. The final rule is ultimate when the red team cannot discover any further corner cases that break the rules." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 325, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 480, + 361 + ], + "type": "text", + "content": "As demonstrated above, the annotation of AutoEval-Video is sophisticated and labor-intensive. We construct a total of 327 instances. The average cost per instance amounts to $28." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 381, + 244, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 244, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 244, + 392 + ], + "type": "text", + "content": "3.2 Skill Dimensions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "content": "Video understanding encompasses two fundamental skills: perception and cognition. Different from image perception, video perception emphasizes temporal information, such as activity, state transitions, and camera movement. With regard to cognition, our benchmark considers reasoning based on temporal or spatial information, simultaneously spanning causal reasoning, comparison, and reasoning with external knowledge. In addition, considering that hallucination [24,28,37,51] is a prominent challenge for contemporary large vision-language models, we incorporate tasks requiring descriptive ability in our benchmark. In summary, we identify nine skill dimensions, as shown in Tab. 1, which will address the aforementioned capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 543, + 315, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 543, + 315, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 543, + 315, + 555 + ], + "type": "text", + "content": "3.3 Video and Question Collection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 565, + 479, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 479, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 479, + 589 + ], + "type": "text", + "content": "Video Collection. We gather all videos from YouTube, taking into account the following considerations:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 599, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 479, + 635 + ], + "type": "text", + "content": "- To minimize the probability of utilizing videos that have been employed in the model training process, we collect new videos from YouTube rather than employing well-known open-sourced video datasets." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 133, + 90, + 447, + 102 + ], + "type": "text", + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 480, + 665 + ], + "type": "text", + "content": "1 The videos collected from YouTube comply with Creative Commons License (https://support.google.com/youtube/answer/2797468)." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 482, + 163 + ], + "type": "text", + "content": "- We aim to construct a benchmark that incorporates real-world visual content so as to directly assess the capability of models in addressing realistic challenges. Thus, we choose to select videos from YouTube instead of filming new videos based on scripts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 187, + 482, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 187, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 482, + 235 + ], + "type": "text", + "content": "In addition, we examine the video categories of our benchmark, which covers mainstream video genres in real-world. All video sources are guaranteed to possess a resolution of 720P or higher. More statistics of the videos can be found in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 157, + 237, + 297, + 376 + ], + "blocks": [ + { + "bbox": [ + 157, + 237, + 297, + 376 + ], + "lines": [ + { + "bbox": [ + 157, + 237, + 297, + 376 + ], + "spans": [ + { + "bbox": [ + 157, + 237, + 297, + 376 + ], + "type": "image", + "image_path": "1df86a7d356967af43059472221e271cbc70ba10814001fed209baee5a500ecd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 376, + 293, + 398 + ], + "lines": [ + { + "bbox": [ + 162, + 376, + 293, + 398 + ], + "spans": [ + { + "bbox": [ + 162, + 376, + 293, + 398 + ], + "type": "text", + "content": "(a) Genre distribution of videos in AutoEval-Video" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 405, + 482, + 429 + ], + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 429 + ], + "type": "text", + "content": "Fig. 2: Statistics of AutoEval-Video. (a) Genre distribution of videos. (b) Statistics on question and rule lengths, number of scenes, and video duration." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 321, + 249, + 459, + 369 + ], + "blocks": [ + { + "bbox": [ + 321, + 249, + 459, + 369 + ], + "lines": [ + { + "bbox": [ + 321, + 249, + 459, + 369 + ], + "spans": [ + { + "bbox": [ + 321, + 249, + 459, + 369 + ], + "type": "table", + "html": "
StatisticsMeanMedian
Question Length (#words)10.9410.00
Rule Length (#words)302.95296.00
Video Shots (#Scenes)3.162.00
Video Duration (#Seconds)14.5810.00
", + "image_path": "64c4836cfff2b367fc0cca68d0b52f2ee44f73559b9034e406b6abceb3fb7d6e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 373, + 455, + 387 + ], + "lines": [ + { + "bbox": [ + 320, + 373, + 455, + 387 + ], + "spans": [ + { + "bbox": [ + 320, + 373, + 455, + 387 + ], + "type": "text", + "content": "(b) Statistics of AutoEval-Video" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "content": "Question Collection. The questions of AutoEval-Video are meticulously designed by human annotators, avoiding templates or paradigms, to guarantee a diverse range of questions. Annotators may either initiate a question and then search for an appropriate video, or select a video first and craft a question afterward. To maximize dataset diversity, we annotate only one question per video. To obtain high-quality video-question pairs that can effectively diagnose models' video comprehension abilities, we propose three principles for constructing the benchmark: avoiding single-frame bias, reducing language bias, and formulating objective questions, as discussed below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 544, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 604 + ], + "type": "text", + "content": "Many video benchmarks suffer from single-frame bias [3, 15, 19], indicating that the information in an oracle frame is almost sufficient to deduce the answer. To address this issue, we instruct annotators to devise questions that require the consideration of multiple frames for successful completion. Questions that can be answered based on only one frame are deemed invalid." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Existing video-language benchmarks are affected by language bias [9, 27], suggesting that models can obtain the correct answers without seeing video content. To tackle this issue, we involve three additional human annotators who try their best to answer questions without viewing the videos. If any of the annotators can accurately answer the question, the instance will be discarded." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 342, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 342, + 103 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 342, + 103 + ], + "type": "text", + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 126, + 181, + 158 + ], + "blocks": [ + { + "bbox": [ + 139, + 118, + 268, + 126 + ], + "lines": [ + { + "bbox": [ + 139, + 118, + 268, + 126 + ], + "spans": [ + { + "bbox": [ + 139, + 118, + 268, + 126 + ], + "type": "text", + "content": "Q: Why did the race car driver exit the vehicle?" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 126, + 181, + 158 + ], + "lines": [ + { + "bbox": [ + 140, + 126, + 181, + 158 + ], + "spans": [ + { + "bbox": [ + 140, + 126, + 181, + 158 + ], + "type": "image", + "image_path": "f904da36f262dad975fc90f565726b0b89b99c87aab70e9b5b050e606b349922.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 182, + 126, + 223, + 158 + ], + "blocks": [ + { + "bbox": [ + 182, + 126, + 223, + 158 + ], + "lines": [ + { + "bbox": [ + 182, + 126, + 223, + 158 + ], + "spans": [ + { + "bbox": [ + 182, + 126, + 223, + 158 + ], + "type": "image", + "image_path": "6fe5b1c963e28a609ed7ebcd54c1cc0d592b60b97fafc16a8b773523b3caa0cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 224, + 126, + 265, + 158 + ], + "blocks": [ + { + "bbox": [ + 224, + 126, + 265, + 158 + ], + "lines": [ + { + "bbox": [ + 224, + 126, + 265, + 158 + ], + "spans": [ + { + "bbox": [ + 224, + 126, + 265, + 158 + ], + "type": "image", + "image_path": "a2fed1de0ca7ee75019c78d66e8857be285c51f89a0f51dcbc4f62b1ec4065e7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 266, + 126, + 307, + 158 + ], + "blocks": [ + { + "bbox": [ + 266, + 126, + 307, + 158 + ], + "lines": [ + { + "bbox": [ + 266, + 126, + 307, + 158 + ], + "spans": [ + { + "bbox": [ + 266, + 126, + 307, + 158 + ], + "type": "image", + "image_path": "0409356616cd15bdc3eab3b965a8caf7dea77599b43dc5b1a3f33058d0ec0fb6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 126, + 349, + 158 + ], + "blocks": [ + { + "bbox": [ + 308, + 126, + 349, + 158 + ], + "lines": [ + { + "bbox": [ + 308, + 126, + 349, + 158 + ], + "spans": [ + { + "bbox": [ + 308, + 126, + 349, + 158 + ], + "type": "image", + "image_path": "c14c30632fde6bf9878099fdb7576b18544a5491910757a7f7f4c75262148ce8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 350, + 126, + 391, + 158 + ], + "blocks": [ + { + "bbox": [ + 350, + 126, + 391, + 158 + ], + "lines": [ + { + "bbox": [ + 350, + 126, + 391, + 158 + ], + "spans": [ + { + "bbox": [ + 350, + 126, + 391, + 158 + ], + "type": "image", + "image_path": "60bb6dd6c6544299f90b79ae95f491f8e877c66806429e24296d5b36ea2c7ece.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 391, + 126, + 432, + 158 + ], + "blocks": [ + { + "bbox": [ + 391, + 126, + 432, + 158 + ], + "lines": [ + { + "bbox": [ + 391, + 126, + 432, + 158 + ], + "spans": [ + { + "bbox": [ + 391, + 126, + 432, + 158 + ], + "type": "image", + "image_path": "08c81c8a6ead6f58c7b0f5d7676a22fabe1b1d16d9345dfef3a2360bbdd94198.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 126, + 480, + 158 + ], + "blocks": [ + { + "bbox": [ + 434, + 126, + 480, + 158 + ], + "lines": [ + { + "bbox": [ + 434, + 126, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 434, + 126, + 480, + 158 + ], + "type": "image", + "image_path": "11e927e71edb90ad76cad873b1204f02003c95810da8b7484a8da17296bc1ef2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 222, + 169, + 255, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 169, + 255, + 175 + ], + "spans": [ + { + "bbox": [ + 222, + 169, + 255, + 175 + ], + "type": "text", + "content": "Initial Rules" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 140, + 176, + 336, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 176, + 336, + 206 + ], + "spans": [ + { + "bbox": [ + 140, + 176, + 336, + 206 + ], + "type": "text", + "content": "[Video description]: In a racetrack, a white race car flipped over and came to a halt, then caught fire. Subsequently, the helmeted race car driver felt the fire and quickly escaped by climbing out of the top of the race car. Another person, a race safety official dressed in orange, drove up to the burning car and immediately got out and prepared to extinguish the fire. [Video Question]: Why did the race car driver exit the vehicle?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 140, + 207, + 335, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 207, + 335, + 219 + ], + "spans": [ + { + "bbox": [ + 140, + 207, + 335, + 219 + ], + "type": "text", + "content": "Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 140, + 219, + 335, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 219, + 335, + 231 + ], + "spans": [ + { + "bbox": [ + 140, + 219, + 335, + 231 + ], + "type": "text", + "content": "1. The main key point in the answer lies in \"race car caught on fire\". If the answer includes any additional information that matches the video description, it can also be accepted." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 140, + 232, + 335, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 232, + 335, + 243 + ], + "spans": [ + { + "bbox": [ + 140, + 232, + 335, + 243 + ], + "type": "text", + "content": "2. The words in the answer must be spelled correctly. If a word does not make sense in the sentence, the answer is incorrect." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 140, + 244, + 335, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 244, + 335, + 255 + ], + "spans": [ + { + "bbox": [ + 140, + 244, + 335, + 255 + ], + "type": "text", + "content": "3. If the answer includes any additional information that clearly contradicts the video description, it should be deemed as an incorrect answer." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 140, + 256, + 249, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 256, + 249, + 262 + ], + "spans": [ + { + "bbox": [ + 140, + 256, + 249, + 262 + ], + "type": "text", + "content": "[Answer to be judged]: (Adversarial Answer)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 262, + 189, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 262, + 189, + 268 + ], + "spans": [ + { + "bbox": [ + 140, + 262, + 189, + 268 + ], + "type": "text", + "content": "[Form of the Result]:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 140, + 269, + 335, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 269, + 335, + 288 + ], + "spans": [ + { + "bbox": [ + 140, + 269, + 335, + 288 + ], + "type": "text", + "content": "please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 388, + 168, + 444, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 168, + 444, + 175 + ], + "spans": [ + { + "bbox": [ + 388, + 168, + 444, + 175 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 358, + 175, + 473, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 175, + 473, + 193 + ], + "spans": [ + { + "bbox": [ + 358, + 175, + 473, + 193 + ], + "type": "text", + "content": "Because the race car driver's car caught fire, which might have been caused by the driver's improper operation." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 410, + 196, + 423, + 216 + ], + "blocks": [ + { + "bbox": [ + 410, + 196, + 423, + 216 + ], + "lines": [ + { + "bbox": [ + 410, + 196, + 423, + 216 + ], + "spans": [ + { + "bbox": [ + 410, + 196, + 423, + 216 + ], + "type": "image", + "image_path": "9fffe01af2769336d17c414d6fbd5392b475a68c0f05401ea2101ba83f68377e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 434, + 481, + 476 + ], + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 476 + ], + "type": "text", + "content": "Fig. 3: Adversarial Annotation for Rules. A red team develops adversarial answers that break the initial rules. Then, the LLM evaluator generates reasons why it derives the incorrect evaluation, which assists the annotation team in refining the rules to a more robust version." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 357, + 220, + 474, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 220, + 474, + 282 + ], + "spans": [ + { + "bbox": [ + 357, + 220, + 474, + 282 + ], + "type": "text", + "content": "The reason provided in the answer, \"improper operation by the driver\", adds speculative information not present in the video description. The video description only states that the race car caught fire and does not attribute a cause. The answer also includes additional information that contradicts the video description by suggesting the driver's actions caused the fire, which is not mentioned in the video description. Therefore, the answer should be deemed incorrect." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 286, + 311, + 327, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 311, + 327, + 318 + ], + "spans": [ + { + "bbox": [ + 286, + 311, + 327, + 318 + ], + "type": "text", + "content": "Updated Rules" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 140, + 319, + 473, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 319, + 473, + 343 + ], + "spans": [ + { + "bbox": [ + 140, + 319, + 473, + 343 + ], + "type": "text", + "content": "[Video description]: In a racetrack, a white race car flipped over and came to a halt, then caught fire. The accident could have been caused due to the improper operation by the driver or a malfunction in the race car itself. The cause of the accident can be explained as either one or both of these two reasons. Subsequently, the helmed race car driver felt the fire and quickly escaped by climbing out of the top of the race car. Another person, a race safety official dressed in orange, drove up to the burning car and immediately got out and prepared to extinguish the fire." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 140, + 344, + 273, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 344, + 273, + 349 + ], + "spans": [ + { + "bbox": [ + 140, + 344, + 273, + 349 + ], + "type": "text", + "content": "[Video Question]: Why did the race car driver exit the vehicle?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 140, + 350, + 383, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 350, + 383, + 355 + ], + "spans": [ + { + "bbox": [ + 140, + 350, + 383, + 355 + ], + "type": "text", + "content": "Please determine whether the [Answer to be judged] is correct for the [Video Question] based on the following rules:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 140, + 356, + 473, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 356, + 473, + 368 + ], + "spans": [ + { + "bbox": [ + 140, + 356, + 473, + 368 + ], + "type": "text", + "content": "1. The main key point in the answer lies in \"race car caught on fire\". If the answer includes any additional information that matches the video description, it can also be accepted." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 140, + 369, + 396, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 369, + 396, + 374 + ], + "spans": [ + { + "bbox": [ + 140, + 369, + 396, + 374 + ], + "type": "text", + "content": "2. The words in the answer must be spelled correctly. If a word does not make sense in the sentence, the answer is incorrect." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 140, + 375, + 432, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 375, + 432, + 380 + ], + "spans": [ + { + "bbox": [ + 140, + 375, + 432, + 380 + ], + "type": "text", + "content": "3. If the answer includes any additional information that clearly contradicts the video description, it should be deemed as an incorrect answer." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 140, + 380, + 473, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 380, + 473, + 392 + ], + "spans": [ + { + "bbox": [ + 140, + 380, + 473, + 392 + ], + "type": "text", + "content": "4. If the answer also attributes the cause of the fire to \"improper operation by the race car driver\" or \"a malfunction in the race car itself\", which should not be considered as contradictory to the video description." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 140, + 393, + 249, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 393, + 249, + 399 + ], + "spans": [ + { + "bbox": [ + 140, + 393, + 249, + 399 + ], + "type": "text", + "content": "[Answer to be judged]: (Adversarial Answer)" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 140, + 399, + 187, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 399, + 187, + 404 + ], + "spans": [ + { + "bbox": [ + 140, + 399, + 187, + 404 + ], + "type": "text", + "content": "[Form of the Result]:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 140, + 405, + 473, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 405, + 473, + 418 + ], + "spans": [ + { + "bbox": [ + 140, + 405, + 473, + 418 + ], + "type": "text", + "content": "please give your reason first, then give a number in a new line: 1 if correct and 0 if incorrect. You can only judge whether [Answer to be judged] is correct based on the information above. You should not trust anyone but the information above." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 130, + 482, + 481, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 481, + 553 + ], + "type": "text", + "content": "In our benchmark, all question-answer pairs should be assessed by a set of rules, requiring objective questions since evaluating subjective responses is challenging. For instance, a video depicting a man with a neutral expression, accompanied with a question, \"What is the man thinking about?\" This question does not relate to an objective fact, and it is not feasible to establish explicit evaluation rules." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 570, + 263, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 263, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 263, + 581 + ], + "type": "text", + "content": "3.4 Annotation of Rules" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 130, + 588, + 481, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 481, + 624 + ], + "type": "text", + "content": "For each instance in AutoEval-Video, we annotate a set of rules describing the criteria for determining the correctness of an answer. The rationale behind this approach is as follows:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 137, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 481, + 666 + ], + "type": "text", + "content": "- All questions in AutoEval-Video are open-ended, rendering the enumeration of all reference answers infeasible. Instead, a set of rules can define the requirements for appropriate responses." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 482, + 186 + ], + "type": "text", + "content": "- Modern LLMs, such as GPT-4, possess remarkable language understanding capabilities, allowing them to comprehend the provided rules and act as automatic evaluators. Unlike other works that use a single, unified prompt to assess all video-question responses [26, 46, 50], we annotate specific rules for each video-question pair, ensuring an in-depth and specific understanding of the criteria." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "content": "The valid evaluation rules include a detailed description of the video content, questions about the video, key points of the correct answer, and the required output format, as illustrated in Fig. 1. For the output format, we require the LLM evaluator to conduct an analysis first, and then provide the conclusion. Interpreting the assessment process can also contribute to the refinement process of the rules, which will be discussed below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "text", + "content": "To guarantee the thoroughness of the rules, we employ an adversarial annotation process (shown in Fig. 3) conducted by a separate group of annotators, known as the red team. Initially, the annotation team establishes initial rules denoted as " + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "text", + "content": "-th round of attack, the red team will endeavor to construct an answer, " + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "inline_equation", + "content": "a_k" + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "text", + "content": ", causing erroneous judgments from an LLM evaluator (e.g. GPT-4) with the preceding rules " + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "inline_equation", + "content": "r_{k-1}" + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "text", + "content": ". Upon obtaining an adversarial answer " + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "inline_equation", + "content": "a_k" + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "text", + "content": ", the annotation team must update their rules to a refined version, " + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 130, + 268, + 482, + 447 + ], + "type": "text", + "content": ". The iterative attack-refine process will continue until the red team cannot identify any adversarial answer. The red team may construct attack answers through various methods, such as: 1) directly writing one based on their understanding of the rules; and 2) prompting LLMs to generate adversarial answers. During the refinement process, the annotators can utilize the interpretations of the LLM evaluator, which will help to quickly figure out why the evaluator gets confused and facilitate the production of more accurate rules. On average, there are 3 attack-refine iterations in the construction of rules in AutoEval-Video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 570 + ], + "type": "text", + "content": "Furthermore, to enhance the motivation and engagement of both the annotation team and the red team in confrontation, we devised a reward mechanism. For each instance, the annotation team initially has a full reward, while the red team starts with zero reward. If the red team successfully discovers an adversarial answer, they will earn a reward, while the annotation team will lose a portion of their reward. As the number of iterations increases, there will be a corresponding increase in rewards for successful attacks, as well as higher penalties for the annotation team. This approach not only encourages the red team to target mature rules but also urges the annotation team to consistently refine their rules throughout every iteration, ensuring optimal results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 586, + 230, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 230, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 230, + 600 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 610, + 262, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 262, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 262, + 624 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Models. Utilizing AutoEval-Video, we assess the performance of eleven models, including six ImageLLMs (i.e., GPT-4V [44], Gemini-Ultra-Vision [38], LLaVA1.5 [29], Qwen-VL [2], InstructBLIP [5], and BLIP-2 [21]) and five VideoLLMs" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 342, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 342, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 342, + 102 + ], + "type": "text", + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": "(i.e., Video-ChatGPT [32], VideoChat, VideoChat2 [22], Video-LLaMA [48], and LLaMA-VID [23]). For each video, we uniformly sample 8 or 16 frames, resulting in two different input configurations. In particular, for ImageLLMs, we concatenate the 8 or 16 frames to create a single image as visual input (refer to supplementary material for the detailed concatenation method)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 178, + 479, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 479, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 479, + 226 + ], + "type": "text", + "content": "LLM as an automatic evaluator. We adopt GPT-4 [33] as an automatic evaluator for AutoEval-Video. For each instance, GPT-4 generates an evaluation of the correctness of a given answer to be judged using the rules as prompts, as demonstrated in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "text", + "content": "Human baseline. We recruit three human participants to answer questions in AutoEval-Video. The overall average accuracy is " + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "inline_equation", + "content": "72.8\\%" + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "text", + "content": ". Through closely examining the mistakes made by humans, we identify two major challenges they face: 1) Questions related to external knowledge, such as unfamiliar games, may lead to incorrect answers from humans; 2) Questions requiring precise memory skills, such as recalling multi-step processes in a video, may cause a person to overlook one or more steps. More examples can be found in supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 174, + 314, + 436, + 472 + ], + "blocks": [ + { + "bbox": [ + 174, + 314, + 436, + 472 + ], + "lines": [ + { + "bbox": [ + 174, + 314, + 436, + 472 + ], + "spans": [ + { + "bbox": [ + 174, + 314, + 436, + 472 + ], + "type": "table", + "html": "
Model TypeModelAccuracy
8 Frames16 Frames
ImageLLMBLIP-2 [21]0.31.2
InstructBLIP [5]7.914.3
Qwen-VL [2]7.010.0
LLaVA-1.5 [29]8.58.2
GPT-4V [44]22.232.2
Gemini-Ultra-Vision [38]17.920.9
VideoLLMVideo-ChatGPT [32]6.74.9
VideoChat [22]13.410.9
Video-LLaMA [48]11.28.5
LLaMA-VID [23]9.88.0
VideoChat2 [22]0.30.6
Human (full videos)72.8
", + "image_path": "f388c51b99a739d3b7ba0982b1d3285f4fcf8b29358438c4933317ce0cf14ab7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 479, + 480, + 502 + ], + "lines": [ + { + "bbox": [ + 130, + 479, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 480, + 502 + ], + "type": "text", + "content": "Table 2: Evaluation Results of Different Models on AutoEval-Video. We also provide human results after viewing full videos." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 534, + 328, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 534, + 328, + 547 + ], + "spans": [ + { + "bbox": [ + 131, + 534, + 328, + 547 + ], + "type": "text", + "content": "4.2 Performances on AutoEval-Video" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 555, + 480, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 480, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 480, + 614 + ], + "type": "text", + "content": "We compare 11 large vision-language models on Tab. 2. Findings show: 1) GPT-4V is much stronger than other models; 2) Among open-sourced models, VideoLLM generally outperforms ImageLLM; 3) GPT-4V falls significantly short of human-level, reflecting the challenge of our benchmark and a big headroom for further advancements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "We find that open-sourced models suffer from three typical categories of bad cases. 1) Hallucination: the described objects or events are not present in the video, leading to wrong answers; 2) Alignment: since AutoEval-Video includes a diverse range of queries, it is often observed that the models' responses do" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 133, + 90, + 447, + 103 + ], + "type": "text", + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "not align with the question; 3) Weak LLM: many models cannot get to the correct answer due to their limited language generation capabilities. We have provided examples for each category in the supplementary material. GPT-4V shows significant progress in these directions, but it still has fundamental issues, as discussed in Sec. 4.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 178, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 482, + 251 + ], + "type": "text", + "content": "Among open-sourced models, VideoChat achieves the highest score. Interestingly, GPT-4V correctly answers only " + }, + { + "bbox": [ + 130, + 178, + 482, + 251 + ], + "type": "inline_equation", + "content": "32.7\\%" + }, + { + "bbox": [ + 130, + 178, + 482, + 251 + ], + "type": "text", + "content": " instances where VideoChat is correct, indicating that VideoChat is stronger in certain domains. These domains are 1) traffic-related tasks; 2) comparison between dynamic objects; and 3) giving video-specific responses rather than general vague answers. In supplementary material, we provide examples for each of these three domains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 323 + ], + "type": "text", + "content": "Finally, we notice that 8 frames are insufficient for answering certain questions. Therefore, we evaluate models using 16 frames as input. Under this setting, GPT-4V significantly improves, reaching " + }, + { + "bbox": [ + 130, + 251, + 482, + 323 + ], + "type": "inline_equation", + "content": "32.2\\%" + }, + { + "bbox": [ + 130, + 251, + 482, + 323 + ], + "type": "text", + "content": " accuracy (examples provided in supplementary material). While longer inputs usually benefit ImageLLMs, VideoLLMs don't exhibit the same trend, possibly because they are explicitly trained in the 8-frame setting." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 182, + 326, + 428, + 388 + ], + "blocks": [ + { + "bbox": [ + 182, + 326, + 428, + 388 + ], + "lines": [ + { + "bbox": [ + 182, + 326, + 428, + 388 + ], + "spans": [ + { + "bbox": [ + 182, + 326, + 428, + 388 + ], + "type": "table", + "html": "
GPT-4 PromptAccuracy of Evaluators
Models/Human ResponsesGPT-4V Responses
Instance-specific rules97.096.6
W/o adversarial annotation91.588.1
W/o rules (unified prompt)87.069.5
Human97.594.9
", + "image_path": "ddf774dcf0cc36735fc7093d854851902be8e60d2279f383f67121d090d9c3e7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 482, + 440 + ], + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 440 + ], + "type": "text", + "content": "Table 3: Ablation Study on Automatic Evaluators. We assess the evaluation accuracy of GPT-4 with instance-specific rules, initial rules, and unified prompts on responses generated by models/humans, as well as responses exclusively generated by GPT-4V. In addition, we present the evaluation accuracy of a single human." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 472, + 382, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 382, + 485 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 382, + 485 + ], + "type": "text", + "content": "4.3 Accuracy of GPT-4 as Automatic Evaluator" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 494, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 566 + ], + "type": "text", + "content": "To validate GPT-4 as an automatic evaluator, we create a test dataset consisting of 200 randomly sampled (video, question, response) triplets, either model-generated or human-generated. Then we employed three human experts to evaluate each response, utilizing their collective decision as the ground truth. As shown in Tab. 3, with the evaluation rules, GPT-4 achieves " + }, + { + "bbox": [ + 130, + 494, + 482, + 566 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 130, + 494, + 482, + 566 + ], + "type": "text", + "content": " agreement with the ground truth, comparable to " + }, + { + "bbox": [ + 130, + 494, + 482, + 566 + ], + "type": "inline_equation", + "content": "97.5\\%" + }, + { + "bbox": [ + 130, + 494, + 482, + 566 + ], + "type": "text", + "content": " of a single human expert." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": "We perform an ablation study to investigate the necessity of our proposed way of annotation. As reported in Tab. 3, if we use the initial version of rules without adversarial annotation, the accuracy drops to " + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "inline_equation", + "content": "91.5\\%" + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": "; using a unified prompt (refer to supplementary material) instead of instance-specific rules makes the accuracy further drop to " + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": ". It implies that both \"instance-specific rules\" and \"adversarial annotation\" are important." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "As the model gets stronger, it makes fewer simple mistakes (e.g., grammatical errors or obvious hallucinations) and generates longer chain-of-thoughts, thus" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 341, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 341, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 341, + 102 + ], + "type": "text", + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": "the wrong answer becomes harder to distinguish from the correct ones. Tab. 3 shows that the accuracy of a single human, initial rules, and unified prompt all significantly drop on a more challenging subset of responses (50 responses exclusively generated by GPT-4V), while instance-specific rules maintain a relatively stable accuracy at " + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "inline_equation", + "content": "96.6\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": ", highlighting its robustness for evaluating strong models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": "A key concern when employing GPT-4 as an automated evaluator is its potential bias towards outputs produced by GPT-4V. To address this concern, we created a test set composed of instances GPT-4V generated incorrectly. On this test dataset, the GPT-4 evaluator with instance-specific rules achieves an accuracy of " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "97.9\\%" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": ", surpassing the human evaluator's accuracy of " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "93.6\\%" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": ". It suggests that GPT-4 with instance-specific rules demonstrates superior accuracy over human evaluators in pinpointing inaccuracies in GPT-4V's responses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 290, + 284, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 290, + 284, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 284, + 303 + ], + "type": "text", + "content": "4.4 Case Study for GPT-4V" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 481, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 481, + 333 + ], + "type": "text", + "content": "Since GPT-4V is the current state-of-the-art, we report a detailed error analysis for it on AutoEval-Video. Here is a list of typical hard cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 335, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 441 + ], + "type": "text", + "content": "Understanding Temporal Features. In many cases, simply looking at a couple of isolated frames isn't enough to answer the question. Take instance 200 in Tab. 4, with its moving camera revealing the layout of a house. The model needs to understand the temporal ordering of frames to infer the house's spatial design to answer \"When I'm at the entrance, how do I get to the balcony that has three decorations on the table?\" Moreover, in instance 305 in supplementary material, the speed at which things change (like \"in an instant\") is a temporal feature. Most vision-language models, including GPT-4V, struggle to extract temporal features from many continuous frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 444, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 516 + ], + "type": "text", + "content": "Connecting Multiple Frames. Many incorrect responses are due to GPT-4V tending to reason on a single frame, ignoring the connection between frames. In Tab. 4, instance 43 demonstrates this fact. Another interesting example is instance 116 in supplementary material. In describing a sequence of actions, GPT-4V mentions a certain action many times because it appears in many frames, although the correct answer should do it just once." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "content": "GPT-4V also struggles to reason between scenes. For example, in instance 112 of Tab. 4, the first scene features a deer in third-person view gradually opening its eyes. The second scene shows an approaching train from the deer's first-person view. To answer \"Why does this orange animal have its eyes wide open?\", the model must connect these two events, but GPT-4V failed. There are more examples in supplementary material supporting this observation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "content": "Moving Objects. Another group of hard cases involves moving objects, such as comparing the speed of two cars. As shown in Tab. 4, instance 191 records a police car overtaking a firefighting truck. However, GPT-4V perceives all frames as semantically similar and is unable to identify the faster vehicle." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Overly General Responses. GPT-4V sometimes answers questions in a general context rather than addressing the specific content of given videos. For" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 133, + 90, + 447, + 102 + ], + "type": "text", + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 140, + 476, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 140, + 476, + 149 + ], + "spans": [ + { + "bbox": [ + 134, + 140, + 476, + 149 + ], + "type": "text", + "content": "Question: When I am at the entrance, how can I reach the balcony that has three decorations on the table? [ID:200]" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 150, + 176, + 182 + ], + "blocks": [ + { + "bbox": [ + 134, + 150, + 176, + 182 + ], + "lines": [ + { + "bbox": [ + 134, + 150, + 176, + 182 + ], + "spans": [ + { + "bbox": [ + 134, + 150, + 176, + 182 + ], + "type": "image", + "image_path": "de694994e0badecad62fd255f0f716c58d338507bc6bac40d3ba88cc3b838f2c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 176, + 150, + 219, + 182 + ], + "blocks": [ + { + "bbox": [ + 176, + 150, + 219, + 182 + ], + "lines": [ + { + "bbox": [ + 176, + 150, + 219, + 182 + ], + "spans": [ + { + "bbox": [ + 176, + 150, + 219, + 182 + ], + "type": "image", + "image_path": "16545f4e842034deeb652a9ceaec13a59f268bf3dcb7d1b0899ab513f22e8dc4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 150, + 261, + 182 + ], + "blocks": [ + { + "bbox": [ + 220, + 150, + 261, + 182 + ], + "lines": [ + { + "bbox": [ + 220, + 150, + 261, + 182 + ], + "spans": [ + { + "bbox": [ + 220, + 150, + 261, + 182 + ], + "type": "image", + "image_path": "90b60928e53313c8d0619fe373c657667973d12255d7fc0d63a285017ea7666e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 263, + 150, + 304, + 182 + ], + "blocks": [ + { + "bbox": [ + 263, + 150, + 304, + 182 + ], + "lines": [ + { + "bbox": [ + 263, + 150, + 304, + 182 + ], + "spans": [ + { + "bbox": [ + 263, + 150, + 304, + 182 + ], + "type": "image", + "image_path": "693cbdc5aa6edcfcb3c9cc79b8fcc85b68848fe9abf51d5c3e13c11043d9ae4b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 150, + 348, + 182 + ], + "blocks": [ + { + "bbox": [ + 306, + 150, + 348, + 182 + ], + "lines": [ + { + "bbox": [ + 306, + 150, + 348, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 150, + 348, + 182 + ], + "type": "image", + "image_path": "76b9dc500c776af4f313bf377db02a44377894051bd5d50603b9ffef4a2372c1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 348, + 150, + 391, + 182 + ], + "blocks": [ + { + "bbox": [ + 348, + 150, + 391, + 182 + ], + "lines": [ + { + "bbox": [ + 348, + 150, + 391, + 182 + ], + "spans": [ + { + "bbox": [ + 348, + 150, + 391, + 182 + ], + "type": "image", + "image_path": "62c56d4401c1627b4d3299b5e67cdaeb590381516be9a9ed18c5a06f307e1438.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 392, + 150, + 434, + 182 + ], + "blocks": [ + { + "bbox": [ + 392, + 150, + 434, + 182 + ], + "lines": [ + { + "bbox": [ + 392, + 150, + 434, + 182 + ], + "spans": [ + { + "bbox": [ + 392, + 150, + 434, + 182 + ], + "type": "image", + "image_path": "4cd5cdb4c8a9d94a7c9281a9c9ce262491165e780b5e2399f6fe0c0244f10249.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 150, + 439, + 182 + ], + "blocks": [ + { + "bbox": [ + 435, + 150, + 439, + 182 + ], + "lines": [ + { + "bbox": [ + 435, + 150, + 439, + 182 + ], + "spans": [ + { + "bbox": [ + 435, + 150, + 439, + 182 + ], + "type": "image", + "image_path": "9c4517d63728cc72a83b0d9f477867d410f1d32a0b423e66d1d458d0b1b7f8e0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 440, + 150, + 482, + 182 + ], + "blocks": [ + { + "bbox": [ + 440, + 150, + 482, + 182 + ], + "lines": [ + { + "bbox": [ + 440, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 440, + 150, + 482, + 182 + ], + "type": "image", + "image_path": "e6aa00eb7bf05b0644e761b55d9dbeac352304791139424ca1c28ebec90fb757.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 187, + 476, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 187, + 476, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 187, + 476, + 211 + ], + "type": "text", + "content": "Sample Answer: Upon entering through the door, you can first turn left into the kitchen, and then turn right. At this point, you are facing the target balcony. All you need to do is continue straight, open the balcony door, and step outside." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 214, + 476, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 476, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 476, + 239 + ], + "type": "text", + "content": "GPT-4V: To reach the balcony with three decorations on the table, enter the house through the main entrance, proceed straight ahead through the living room, and then turn left to go through the double doors leading to the outdoor covered patio area." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 243, + 335, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 243, + 335, + 252 + ], + "spans": [ + { + "bbox": [ + 134, + 243, + 335, + 252 + ], + "type": "text", + "content": "Question: What action did the third parrot from the left perform?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 453, + 244, + 476, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 453, + 244, + 476, + 252 + ], + "spans": [ + { + "bbox": [ + 453, + 244, + 476, + 252 + ], + "type": "text", + "content": "[ID:43]" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 134, + 255, + 176, + 286 + ], + "blocks": [ + { + "bbox": [ + 134, + 255, + 176, + 286 + ], + "lines": [ + { + "bbox": [ + 134, + 255, + 176, + 286 + ], + "spans": [ + { + "bbox": [ + 134, + 255, + 176, + 286 + ], + "type": "image", + "image_path": "c18c133a1fd93831c7bb94d7611875a9ec8b5c3cc148bb567e880e983c8ad83a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 176, + 255, + 219, + 286 + ], + "blocks": [ + { + "bbox": [ + 176, + 255, + 219, + 286 + ], + "lines": [ + { + "bbox": [ + 176, + 255, + 219, + 286 + ], + "spans": [ + { + "bbox": [ + 176, + 255, + 219, + 286 + ], + "type": "image", + "image_path": "d1e5ef6631b8f54a2084a449a94983f1b408f0f5dde36aae7512fbe06b79890c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 220, + 255, + 261, + 286 + ], + "blocks": [ + { + "bbox": [ + 220, + 255, + 261, + 286 + ], + "lines": [ + { + "bbox": [ + 220, + 255, + 261, + 286 + ], + "spans": [ + { + "bbox": [ + 220, + 255, + 261, + 286 + ], + "type": "image", + "image_path": "825c4162f43e1aa204b7a102844f6176590716da20a135b03cf4b4827adaa37b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 263, + 255, + 304, + 286 + ], + "blocks": [ + { + "bbox": [ + 263, + 255, + 304, + 286 + ], + "lines": [ + { + "bbox": [ + 263, + 255, + 304, + 286 + ], + "spans": [ + { + "bbox": [ + 263, + 255, + 304, + 286 + ], + "type": "image", + "image_path": "c3d61df45c7df4d93b31d40974183a134ab5053d11c36750b49d2081ac417741.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 306, + 255, + 348, + 286 + ], + "blocks": [ + { + "bbox": [ + 306, + 255, + 348, + 286 + ], + "lines": [ + { + "bbox": [ + 306, + 255, + 348, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 255, + 348, + 286 + ], + "type": "image", + "image_path": "d40e0f4e8153f40eb80eacb1198afef9aa55b5fcf937d117e8d9366199f2f3d2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 349, + 255, + 390, + 286 + ], + "blocks": [ + { + "bbox": [ + 349, + 255, + 390, + 286 + ], + "lines": [ + { + "bbox": [ + 349, + 255, + 390, + 286 + ], + "spans": [ + { + "bbox": [ + 349, + 255, + 390, + 286 + ], + "type": "image", + "image_path": "a6e61392e7a0347ddd977398cfe318ca0d55ff5333ec457827d2fc438a397037.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 391, + 255, + 433, + 286 + ], + "blocks": [ + { + "bbox": [ + 391, + 255, + 433, + 286 + ], + "lines": [ + { + "bbox": [ + 391, + 255, + 433, + 286 + ], + "spans": [ + { + "bbox": [ + 391, + 255, + 433, + 286 + ], + "type": "image", + "image_path": "c1ff13fce68bef9f0d5300610a6a3118c0aa8fa231ed08f2cef227c65630d15a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 435, + 255, + 477, + 286 + ], + "blocks": [ + { + "bbox": [ + 435, + 255, + 477, + 286 + ], + "lines": [ + { + "bbox": [ + 435, + 255, + 477, + 286 + ], + "spans": [ + { + "bbox": [ + 435, + 255, + 477, + 286 + ], + "type": "image", + "image_path": "db818d7fee07e35b22a3ac1ed3e117d3ccd9902663381d02d1a73875e2e4f820.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 133, + 291, + 280, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 291, + 280, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 291, + 280, + 300 + ], + "type": "text", + "content": "Sample Answer: It flew into the person's hand." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 302, + 476, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 302, + 476, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 476, + 319 + ], + "type": "text", + "content": "GPT-4V: The third parrot from the left appears to be perched on a hand and interacting with a person. It looks like the parrot might be engaged in some sort of play or social interaction with the individual." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 134, + 324, + 321, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 324, + 321, + 334 + ], + "spans": [ + { + "bbox": [ + 134, + 324, + 321, + 334 + ], + "type": "text", + "content": "Question: Why are the eyes of this orange animal wide open?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 450, + 325, + 476, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 450, + 325, + 476, + 333 + ], + "spans": [ + { + "bbox": [ + 450, + 325, + 476, + 333 + ], + "type": "text", + "content": "[ID:112]" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 134, + 335, + 176, + 365 + ], + "blocks": [ + { + "bbox": [ + 134, + 335, + 176, + 365 + ], + "lines": [ + { + "bbox": [ + 134, + 335, + 176, + 365 + ], + "spans": [ + { + "bbox": [ + 134, + 335, + 176, + 365 + ], + "type": "image", + "image_path": "ddccb5b3fcd7a42577a4e7fb6d289ec0567dc2fe6957646cd0e5c903e2d42b22.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 176, + 335, + 219, + 365 + ], + "blocks": [ + { + "bbox": [ + 176, + 335, + 219, + 365 + ], + "lines": [ + { + "bbox": [ + 176, + 335, + 219, + 365 + ], + "spans": [ + { + "bbox": [ + 176, + 335, + 219, + 365 + ], + "type": "image", + "image_path": "b3fce0b1c6e040f5fccb774f4cf55d765c6d0984fc33589df1ff9e56c61bc5c7.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 220, + 335, + 261, + 365 + ], + "blocks": [ + { + "bbox": [ + 220, + 335, + 261, + 365 + ], + "lines": [ + { + "bbox": [ + 220, + 335, + 261, + 365 + ], + "spans": [ + { + "bbox": [ + 220, + 335, + 261, + 365 + ], + "type": "image", + "image_path": "76378cdbae2f59b234c91ae3b573031ac71049d8d63fbdce26fbda8fab9a39e8.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 263, + 335, + 304, + 365 + ], + "blocks": [ + { + "bbox": [ + 263, + 335, + 304, + 365 + ], + "lines": [ + { + "bbox": [ + 263, + 335, + 304, + 365 + ], + "spans": [ + { + "bbox": [ + 263, + 335, + 304, + 365 + ], + "type": "image", + "image_path": "1e44f5dd026a6cbac11f78e79a3e736a598da582b1f1708c54daae0958a9b003.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 306, + 335, + 348, + 366 + ], + "blocks": [ + { + "bbox": [ + 306, + 335, + 348, + 366 + ], + "lines": [ + { + "bbox": [ + 306, + 335, + 348, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 348, + 366 + ], + "type": "image", + "image_path": "b1212bcec6ee662d652a401314bb0718a6ab042d1c73427517f901b1bbcc6164.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 349, + 335, + 391, + 366 + ], + "blocks": [ + { + "bbox": [ + 349, + 335, + 391, + 366 + ], + "lines": [ + { + "bbox": [ + 349, + 335, + 391, + 366 + ], + "spans": [ + { + "bbox": [ + 349, + 335, + 391, + 366 + ], + "type": "image", + "image_path": "b77219fc07699635591815e78e02f785c8a66b5d685a3aed264d0ca71ae107f2.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 391, + 335, + 433, + 366 + ], + "blocks": [ + { + "bbox": [ + 391, + 335, + 433, + 366 + ], + "lines": [ + { + "bbox": [ + 391, + 335, + 433, + 366 + ], + "spans": [ + { + "bbox": [ + 391, + 335, + 433, + 366 + ], + "type": "image", + "image_path": "36d21ad8274bcf3b4f6d2193456e8d16ce1886626781beba4fdc3119ba1c8097.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 435, + 335, + 477, + 366 + ], + "blocks": [ + { + "bbox": [ + 435, + 335, + 477, + 366 + ], + "lines": [ + { + "bbox": [ + 435, + 335, + 477, + 366 + ], + "spans": [ + { + "bbox": [ + 435, + 335, + 477, + 366 + ], + "type": "image", + "image_path": "c759aa760f4a9e3bcb6e354b2bd8c6b93f27b60384b8a61bc610006026add116.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 133, + 371, + 465, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 371, + 465, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 371, + 465, + 380 + ], + "type": "text", + "content": "Sample Answer: Because the animal saw a train approaching, the white creature on the tracks was in danger." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 382, + 476, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 476, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 476, + 423 + ], + "type": "text", + "content": "GPT-4V: The orange animal appears to be a stylized or animated character, possibly showing surprise, alarm, or excitement. Wide open eyes in animation are often used to convey strong emotions or reactions. This could be in response to something happening in the scene or to depict a particular personality trait of the character. Without additional context, it's hard to specify exactly why, but wide eyes are a common visual cue for heightened emotions in animation." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 134, + 428, + 321, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 428, + 321, + 437 + ], + "spans": [ + { + "bbox": [ + 134, + 428, + 321, + 437 + ], + "type": "text", + "content": "Question: In the video, which car is moving at a slower pace?" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 450, + 429, + 476, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 450, + 429, + 476, + 437 + ], + "spans": [ + { + "bbox": [ + 450, + 429, + 476, + 437 + ], + "type": "text", + "content": "[ID:191]" + } + ] + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 134, + 438, + 176, + 470 + ], + "blocks": [ + { + "bbox": [ + 134, + 438, + 176, + 470 + ], + "lines": [ + { + "bbox": [ + 134, + 438, + 176, + 470 + ], + "spans": [ + { + "bbox": [ + 134, + 438, + 176, + 470 + ], + "type": "image", + "image_path": "5325daf7ff6016eb239556c6dd2dbb01f9dea11817078bb9200cf8cc629f062a.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 176, + 438, + 219, + 470 + ], + "blocks": [ + { + "bbox": [ + 176, + 438, + 219, + 470 + ], + "lines": [ + { + "bbox": [ + 176, + 438, + 219, + 470 + ], + "spans": [ + { + "bbox": [ + 176, + 438, + 219, + 470 + ], + "type": "image", + "image_path": "2e757ee2c41f97a227a6fa2ba2fd6687e0c4d7a40b2f8e1d2a8f73e976af8d52.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 220, + 438, + 261, + 470 + ], + "blocks": [ + { + "bbox": [ + 220, + 438, + 261, + 470 + ], + "lines": [ + { + "bbox": [ + 220, + 438, + 261, + 470 + ], + "spans": [ + { + "bbox": [ + 220, + 438, + 261, + 470 + ], + "type": "image", + "image_path": "c7ac256d89a3fc030ae40e5549c9cafa61ffca8e5fb32a9a54eda8dc881c0a31.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 263, + 438, + 304, + 470 + ], + "blocks": [ + { + "bbox": [ + 263, + 438, + 304, + 470 + ], + "lines": [ + { + "bbox": [ + 263, + 438, + 304, + 470 + ], + "spans": [ + { + "bbox": [ + 263, + 438, + 304, + 470 + ], + "type": "image", + "image_path": "7c08fa2d1c1010e5c5f151f4826a16b733d3d49a4aeb2e0ab855897968caab59.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 306, + 438, + 348, + 470 + ], + "blocks": [ + { + "bbox": [ + 306, + 438, + 348, + 470 + ], + "lines": [ + { + "bbox": [ + 306, + 438, + 348, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 348, + 470 + ], + "type": "image", + "image_path": "9e470c6a282ed81326b2d70a69719f9137e5d970749efabb7af2d7d4608a75eb.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 349, + 438, + 391, + 470 + ], + "blocks": [ + { + "bbox": [ + 349, + 438, + 391, + 470 + ], + "lines": [ + { + "bbox": [ + 349, + 438, + 391, + 470 + ], + "spans": [ + { + "bbox": [ + 349, + 438, + 391, + 470 + ], + "type": "image", + "image_path": "603bcd71f7bce2763da91383ce7773468b2591b0df6663bc9fd00b8228671190.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 391, + 438, + 433, + 470 + ], + "blocks": [ + { + "bbox": [ + 391, + 438, + 433, + 470 + ], + "lines": [ + { + "bbox": [ + 391, + 438, + 433, + 470 + ], + "spans": [ + { + "bbox": [ + 391, + 438, + 433, + 470 + ], + "type": "image", + "image_path": "a4da0c08eab688ab9b636ce88130955c164b007488dfe57b157508519f9fd1f8.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 435, + 438, + 477, + 470 + ], + "blocks": [ + { + "bbox": [ + 435, + 438, + 477, + 470 + ], + "lines": [ + { + "bbox": [ + 435, + 438, + 477, + 470 + ], + "spans": [ + { + "bbox": [ + 435, + 438, + 477, + 470 + ], + "type": "image", + "image_path": "6b3a95950591d1df9ae4f8653502c0748403e2c976042048110133c88475cbec.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 133, + 475, + 235, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 235, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 235, + 483 + ], + "type": "text", + "content": "Sample Answer: The police car." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 132, + 487, + 476, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 476, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 476, + 510 + ], + "type": "text", + "content": "GPT-4V: I'm unable to view videos or discern the speed of objects within them. The image you've provided is static, so it doesn't convey the speed of the cars. If you have a specific question about the details within the image, feel free to ask!" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 134, + 517, + 303, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 517, + 303, + 525 + ], + "spans": [ + { + "bbox": [ + 134, + 517, + 303, + 525 + ], + "type": "text", + "content": "Question: Why are the audience members standing up?" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 453, + 517, + 476, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 453, + 517, + 476, + 525 + ], + "spans": [ + { + "bbox": [ + 453, + 517, + 476, + 525 + ], + "type": "text", + "content": "[ID:97]" + } + ] + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 134, + 526, + 176, + 559 + ], + "blocks": [ + { + "bbox": [ + 134, + 526, + 176, + 559 + ], + "lines": [ + { + "bbox": [ + 134, + 526, + 176, + 559 + ], + "spans": [ + { + "bbox": [ + 134, + 526, + 176, + 559 + ], + "type": "image", + "image_path": "88c421d77bf1c152a7f1d9a52bc1a3e9317747abdf27384daad21b1cf97396c3.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 176, + 526, + 219, + 559 + ], + "blocks": [ + { + "bbox": [ + 176, + 526, + 219, + 559 + ], + "lines": [ + { + "bbox": [ + 176, + 526, + 219, + 559 + ], + "spans": [ + { + "bbox": [ + 176, + 526, + 219, + 559 + ], + "type": "image", + "image_path": "fd30662b93930e5b501518d617b6c64d1d6e05e3f9b7bf3d6920fd0ad5cd1697.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 220, + 526, + 261, + 559 + ], + "blocks": [ + { + "bbox": [ + 220, + 526, + 261, + 559 + ], + "lines": [ + { + "bbox": [ + 220, + 526, + 261, + 559 + ], + "spans": [ + { + "bbox": [ + 220, + 526, + 261, + 559 + ], + "type": "image", + "image_path": "534329cc39713f0d03e1015aa93a11d8130eefd89844af917a57a2c8f3bc4d7f.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 263, + 526, + 304, + 559 + ], + "blocks": [ + { + "bbox": [ + 263, + 526, + 304, + 559 + ], + "lines": [ + { + "bbox": [ + 263, + 526, + 304, + 559 + ], + "spans": [ + { + "bbox": [ + 263, + 526, + 304, + 559 + ], + "type": "image", + "image_path": "44cb96792b1f03b84cf1bf26bf8e73de1df72c13be6944f7f52b999ac0fc5744.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 306, + 526, + 348, + 559 + ], + "blocks": [ + { + "bbox": [ + 306, + 526, + 348, + 559 + ], + "lines": [ + { + "bbox": [ + 306, + 526, + 348, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 348, + 559 + ], + "type": "image", + "image_path": "e6a0b04593e19a1aa404b9075180cefa2956a3812f40ec16e38da64b67926430.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 349, + 526, + 391, + 559 + ], + "blocks": [ + { + "bbox": [ + 349, + 526, + 391, + 559 + ], + "lines": [ + { + "bbox": [ + 349, + 526, + 391, + 559 + ], + "spans": [ + { + "bbox": [ + 349, + 526, + 391, + 559 + ], + "type": "image", + "image_path": "75541d58c1efc202c91601f86d38c575e44fb7fa91e7163ba30e6491bf151a14.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 391, + 526, + 433, + 559 + ], + "blocks": [ + { + "bbox": [ + 391, + 526, + 433, + 559 + ], + "lines": [ + { + "bbox": [ + 391, + 526, + 433, + 559 + ], + "spans": [ + { + "bbox": [ + 391, + 526, + 433, + 559 + ], + "type": "image", + "image_path": "92356d737e3ebf918b514d9da2d2c21eeb19626b8c4163680cffd29c4e080d79.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 435, + 526, + 477, + 559 + ], + "blocks": [ + { + "bbox": [ + 435, + 526, + 477, + 559 + ], + "lines": [ + { + "bbox": [ + 435, + 526, + 477, + 559 + ], + "spans": [ + { + "bbox": [ + 435, + 526, + 477, + 559 + ], + "type": "image", + "image_path": "3c828c0869d0b90ef6c555aa21326cb83e4de67c644d6a47d4b56fcd1faa3878.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "bbox": [ + 133, + 563, + 321, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 563, + 321, + 572 + ], + "spans": [ + { + "bbox": [ + 133, + 563, + 321, + 572 + ], + "type": "text", + "content": "Sample Answer: The player in the blue team uniform scored." + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 132, + 574, + 476, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 476, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 476, + 608 + ], + "type": "text", + "content": "GPT-4V: The audiences in sports video games, like the one depicted in the images, often stand up in response to exciting moments or plays. This is a way to simulate real-life reactions of fans during intense or critical moments in a game, such as a close score, a game-winning shot, or a spectacular play. The game developers include these reactions to enhance the realism and immersion of the game experience for players." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 171, + 617, + 441, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 617, + 441, + 628 + ], + "spans": [ + { + "bbox": [ + 171, + 617, + 441, + 628 + ], + "type": "text", + "content": "Table 4: Examples of Incorrect Responses from GPT-4V." + } + ] + } + ], + "index": 62 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 342, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 342, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 342, + 102 + ], + "type": "text", + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 335, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 335, + 124 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 335, + 124 + ], + "type": "text", + "content": "Question: What was the score of the team in black in the video footage?" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 124, + 175, + 156 + ], + "blocks": [ + { + "bbox": [ + 134, + 124, + 175, + 156 + ], + "lines": [ + { + "bbox": [ + 134, + 124, + 175, + 156 + ], + "spans": [ + { + "bbox": [ + 134, + 124, + 175, + 156 + ], + "type": "image", + "image_path": "15541733488d8b987a2402afff3a6bf9356d809a24ae1a3848816607640d439f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 176, + 124, + 219, + 156 + ], + "blocks": [ + { + "bbox": [ + 176, + 124, + 219, + 156 + ], + "lines": [ + { + "bbox": [ + 176, + 124, + 219, + 156 + ], + "spans": [ + { + "bbox": [ + 176, + 124, + 219, + 156 + ], + "type": "image", + "image_path": "f0bfaa4faa2fcf376899fde15fbf44c6d6264ba9a27ae627dfe91cb1cbfbce3e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 124, + 262, + 156 + ], + "blocks": [ + { + "bbox": [ + 220, + 124, + 262, + 156 + ], + "lines": [ + { + "bbox": [ + 220, + 124, + 262, + 156 + ], + "spans": [ + { + "bbox": [ + 220, + 124, + 262, + 156 + ], + "type": "image", + "image_path": "e2452a077ddd57bba54c4c11e913b0f6eb3e1dde6cab383ea7a484e653031f93.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 263, + 124, + 305, + 156 + ], + "blocks": [ + { + "bbox": [ + 263, + 124, + 305, + 156 + ], + "lines": [ + { + "bbox": [ + 263, + 124, + 305, + 156 + ], + "spans": [ + { + "bbox": [ + 263, + 124, + 305, + 156 + ], + "type": "image", + "image_path": "72e6a56aa63819866f901b36b6fffec58663dca3fb9f5bfd161528625097c6e0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 124, + 348, + 156 + ], + "blocks": [ + { + "bbox": [ + 306, + 124, + 348, + 156 + ], + "lines": [ + { + "bbox": [ + 306, + 124, + 348, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 124, + 348, + 156 + ], + "type": "image", + "image_path": "eb71f5d6acead61ed1d57dcea26a0a1c1e89add7d339dc5d0eea3a197a647ac4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 349, + 124, + 391, + 156 + ], + "blocks": [ + { + "bbox": [ + 349, + 124, + 391, + 156 + ], + "lines": [ + { + "bbox": [ + 349, + 124, + 391, + 156 + ], + "spans": [ + { + "bbox": [ + 349, + 124, + 391, + 156 + ], + "type": "image", + "image_path": "ddfb573517526e7564a72120435e5d29ca778dc91d8535fb68574fc14856e4b4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 392, + 124, + 433, + 156 + ], + "blocks": [ + { + "bbox": [ + 392, + 124, + 433, + 156 + ], + "lines": [ + { + "bbox": [ + 392, + 124, + 433, + 156 + ], + "spans": [ + { + "bbox": [ + 392, + 124, + 433, + 156 + ], + "type": "image", + "image_path": "50020e358c933046a906b8d9a7e7971767b62401fdab3620e750c4fb4d6a9f1d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 124, + 477, + 156 + ], + "blocks": [ + { + "bbox": [ + 435, + 124, + 477, + 156 + ], + "lines": [ + { + "bbox": [ + 435, + 124, + 477, + 156 + ], + "spans": [ + { + "bbox": [ + 435, + 124, + 477, + 156 + ], + "type": "image", + "image_path": "34314a702dfa8f485a94238713240575cccd454cc7fa8e61cc9167064d7b5a08.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 159, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 159, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 134, + 159, + 296, + 168 + ], + "type": "text", + "content": "Sample Answer: The team in black scored eight points." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 171, + 285, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 171, + 285, + 178 + ], + "spans": [ + { + "bbox": [ + 134, + 171, + 285, + 178 + ], + "type": "text", + "content": "Human: The black team scores 3 points in the video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 183, + 279, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 183, + 279, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 183, + 279, + 190 + ], + "type": "text", + "content": "Question: How did the character detect the enemy?" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 134, + 190, + 176, + 223 + ], + "blocks": [ + { + "bbox": [ + 134, + 190, + 176, + 223 + ], + "lines": [ + { + "bbox": [ + 134, + 190, + 176, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 190, + 176, + 223 + ], + "type": "image", + "image_path": "f2a604066ecb847ecd92970a7b806752da3315c439359e848db7983de9e95499.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 176, + 190, + 219, + 223 + ], + "blocks": [ + { + "bbox": [ + 176, + 190, + 219, + 223 + ], + "lines": [ + { + "bbox": [ + 176, + 190, + 219, + 223 + ], + "spans": [ + { + "bbox": [ + 176, + 190, + 219, + 223 + ], + "type": "image", + "image_path": "43e7b3734abcdf9fb359487fcf8a686a6f486c5fa707dfc3f3b0ea80ca8dc84e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 220, + 190, + 262, + 223 + ], + "blocks": [ + { + "bbox": [ + 220, + 190, + 262, + 223 + ], + "lines": [ + { + "bbox": [ + 220, + 190, + 262, + 223 + ], + "spans": [ + { + "bbox": [ + 220, + 190, + 262, + 223 + ], + "type": "image", + "image_path": "aa2884480050e99c2bc0d93097873a8b05625c6e882e64e532f4ff000fe02a45.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 263, + 190, + 305, + 223 + ], + "blocks": [ + { + "bbox": [ + 263, + 190, + 305, + 223 + ], + "lines": [ + { + "bbox": [ + 263, + 190, + 305, + 223 + ], + "spans": [ + { + "bbox": [ + 263, + 190, + 305, + 223 + ], + "type": "image", + "image_path": "c2951fa477b119e619251b2d1be3922daefc9668daea3137d9268d158c24185d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 306, + 190, + 348, + 223 + ], + "blocks": [ + { + "bbox": [ + 306, + 190, + 348, + 223 + ], + "lines": [ + { + "bbox": [ + 306, + 190, + 348, + 223 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 348, + 223 + ], + "type": "image", + "image_path": "35df2db282d514bcee5cb10211f11a5eebe24a5226266127d760653cb86a46da.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 349, + 190, + 391, + 223 + ], + "blocks": [ + { + "bbox": [ + 349, + 190, + 391, + 223 + ], + "lines": [ + { + "bbox": [ + 349, + 190, + 391, + 223 + ], + "spans": [ + { + "bbox": [ + 349, + 190, + 391, + 223 + ], + "type": "image", + "image_path": "7d881cb3e4a1e39ebd0425df0799fef23c6c5426817587ad6c3541cb1c9ec22c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 392, + 190, + 433, + 223 + ], + "blocks": [ + { + "bbox": [ + 392, + 190, + 433, + 223 + ], + "lines": [ + { + "bbox": [ + 392, + 190, + 433, + 223 + ], + "spans": [ + { + "bbox": [ + 392, + 190, + 433, + 223 + ], + "type": "image", + "image_path": "a4ae9af5c50fb3b72be868af5d1d6dde9f1878a5cc3ccc69cff4bb09b02b9f8b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 435, + 190, + 477, + 223 + ], + "blocks": [ + { + "bbox": [ + 435, + 190, + 477, + 223 + ], + "lines": [ + { + "bbox": [ + 435, + 190, + 477, + 223 + ], + "spans": [ + { + "bbox": [ + 435, + 190, + 477, + 223 + ], + "type": "image", + "image_path": "555e7f9ed738d6baaa749dfb0762f167189e1d37f79dd1a06066cc07887cda4b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 134, + 227, + 391, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 227, + 391, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 227, + 391, + 236 + ], + "type": "text", + "content": "Sample Answer: The enemy shot the character so the enemy's location could be known." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 134, + 237, + 388, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 237, + 388, + 245 + ], + "spans": [ + { + "bbox": [ + 134, + 237, + 388, + 245 + ], + "type": "text", + "content": "Human: The character used the scope on the sniper rifle to clearly observe distant enemies." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 134, + 250, + 395, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 250, + 395, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 250, + 395, + 258 + ], + "type": "text", + "content": "Question: Why did this person initially appear stunned and then suddenly burst into laughter?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 452, + 250, + 476, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 250, + 476, + 258 + ], + "spans": [ + { + "bbox": [ + 452, + 250, + 476, + 258 + ], + "type": "text", + "content": "[ID:133]" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 134, + 258, + 176, + 289 + ], + "blocks": [ + { + "bbox": [ + 134, + 258, + 176, + 289 + ], + "lines": [ + { + "bbox": [ + 134, + 258, + 176, + 289 + ], + "spans": [ + { + "bbox": [ + 134, + 258, + 176, + 289 + ], + "type": "image", + "image_path": "03b370152bbf7663cb168b3f51b4f2cec54e87530bd949c0c167c97396e56243.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 176, + 258, + 219, + 289 + ], + "blocks": [ + { + "bbox": [ + 176, + 258, + 219, + 289 + ], + "lines": [ + { + "bbox": [ + 176, + 258, + 219, + 289 + ], + "spans": [ + { + "bbox": [ + 176, + 258, + 219, + 289 + ], + "type": "image", + "image_path": "23f7380e3094b2e9bc213cbeca3e70a9a638267b6dd00bb2dfd9d6d4ec216581.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 220, + 258, + 262, + 289 + ], + "blocks": [ + { + "bbox": [ + 220, + 258, + 262, + 289 + ], + "lines": [ + { + "bbox": [ + 220, + 258, + 262, + 289 + ], + "spans": [ + { + "bbox": [ + 220, + 258, + 262, + 289 + ], + "type": "image", + "image_path": "7d45efc847118d000543d1fc8cd0a73fd9207a519017acf1091cabb392f0ecb1.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 263, + 258, + 305, + 289 + ], + "blocks": [ + { + "bbox": [ + 263, + 258, + 305, + 289 + ], + "lines": [ + { + "bbox": [ + 263, + 258, + 305, + 289 + ], + "spans": [ + { + "bbox": [ + 263, + 258, + 305, + 289 + ], + "type": "image", + "image_path": "dfc49fe608f1f191956eda529d9628383199b8186aad87641423131facd8747b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 306, + 258, + 348, + 289 + ], + "blocks": [ + { + "bbox": [ + 306, + 258, + 348, + 289 + ], + "lines": [ + { + "bbox": [ + 306, + 258, + 348, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 348, + 289 + ], + "type": "image", + "image_path": "5ecfe81cf6be51d2725dc3fd27d6006a2137b793276dcac198fd4b3485f078fd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 349, + 258, + 391, + 289 + ], + "blocks": [ + { + "bbox": [ + 349, + 258, + 391, + 289 + ], + "lines": [ + { + "bbox": [ + 349, + 258, + 391, + 289 + ], + "spans": [ + { + "bbox": [ + 349, + 258, + 391, + 289 + ], + "type": "image", + "image_path": "06290bf9bd3a6e29b93e2ed0bbb1069707f7e96ef51accb20eea7db060085d06.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 392, + 258, + 433, + 289 + ], + "blocks": [ + { + "bbox": [ + 392, + 258, + 433, + 289 + ], + "lines": [ + { + "bbox": [ + 392, + 258, + 433, + 289 + ], + "spans": [ + { + "bbox": [ + 392, + 258, + 433, + 289 + ], + "type": "image", + "image_path": "e3b7ff46583e97b391adee657884a3b31759501db2038d711486e461af036a03.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 435, + 258, + 477, + 289 + ], + "blocks": [ + { + "bbox": [ + 435, + 258, + 477, + 289 + ], + "lines": [ + { + "bbox": [ + 435, + 258, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 435, + 258, + 477, + 289 + ], + "type": "image", + "image_path": "26923a9240379e716d67fa87e264ba5cf447ba304d461daf1e75998860c23a4a.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 134, + 293, + 465, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 293, + 465, + 302 + ], + "spans": [ + { + "bbox": [ + 134, + 293, + 465, + 302 + ], + "type": "text", + "content": "Sample Answer: Because the person threw grenades to kill the enemy but knocked a teammate down accidentally." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 134, + 304, + 402, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 304, + 402, + 312 + ], + "spans": [ + { + "bbox": [ + 134, + 304, + 402, + 312 + ], + "type": "text", + "content": "Human: Because she didn't expect that the bomb she threw would explode and harm someone." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 140, + 322, + 471, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 322, + 471, + 333 + ], + "spans": [ + { + "bbox": [ + 140, + 322, + 471, + 333 + ], + "type": "text", + "content": "Table 5: Hard cases for humans due to insufficient external knowledge." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 130, + 347, + 480, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 480, + 419 + ], + "type": "text", + "content": "instance, in instance 97 shown in Tab. 4, GPT-4V answers \"Why are the audience members standing up?\" in a general way, even though there is an obvious and concrete reason. Similarly, instance 197 gives a general guide for operating a coffee machine even though the video demonstrates a specific one. It implies that GPT-4V tends to rely on its strong language model to produce a safe answer, sometimes ignoring key details in the video." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 437, + 281, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 281, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 281, + 448 + ], + "type": "text", + "content": "4.5 Hard Cases for Humans" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 130, + 458, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 480, + 506 + ], + "type": "text", + "content": "Humans achieve an accuracy rate of " + }, + { + "bbox": [ + 130, + 458, + 480, + 506 + ], + "type": "inline_equation", + "content": "72.8\\%" + }, + { + "bbox": [ + 130, + 458, + 480, + 506 + ], + "type": "text", + "content": " on AutoEval-Video. We conduct an in-depth analysis of these cases and discover that there are two major hard cases for humans: questions related to external knowledge and questions requiring accurate description and summarizing ability." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 130, + 508, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 480, + 604 + ], + "type": "text", + "content": "Questions Related to External Knowledge. Some questions in AutoEval-Video require a relatively high level of mastery of external knowledge in specific domains such as sports and games, resulting in human difficulties. For example, in Tab. 5, instance 38 shows a misunderstanding due to human participants not knowing basketball scoring rules. Similarly, instance 133 requires explaining a game streamer's laughter, an easy task for someone familiar with the game but confusing for others. However, considering the cognitive level of the general public, these shortcomings are reasonable and expected." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Accurate Description and Summarizing Ability. Questions requiring detailed, step-by-step video descriptions are more challenging for humans compared to semantic reasoning problems due to their complexity and precision. For instance, as shown in Tab. 6, in instance 173, human makes careless mistakes by ignoring necessary CPU installation steps shown in the video. In instance" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "143, errors occur due to insufficient observation, misidentifying corn kernels as soybeans. This phenomenon is common in many step-by-step descriptive tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 145, + 476, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 145, + 476, + 153 + ], + "spans": [ + { + "bbox": [ + 134, + 145, + 476, + 153 + ], + "type": "text", + "content": "Question: How can you prepare a dish using these ingredients and tools as demonstrated in the video? [ID:143]" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 154, + 476, + 186 + ], + "blocks": [ + { + "bbox": [ + 134, + 154, + 476, + 186 + ], + "lines": [ + { + "bbox": [ + 134, + 154, + 476, + 186 + ], + "spans": [ + { + "bbox": [ + 134, + 154, + 476, + 186 + ], + "type": "image", + "image_path": "2f3a7d63bc41154e6550ba9cc5383769134b66aa16c2eb85f9fbdeeb3dad4da6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 190, + 476, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 190, + 476, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 476, + 215 + ], + "type": "text", + "content": "Sample Answer: Cut the potatoes, beets, purple cabbage, carrots, mushrooms, and bell peppers, put them in the steamer with the corn kernels, cut the gluten into strips, heat the steamer on the gas stove, and put the gluten into the air fryer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 218, + 476, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 476, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 476, + 266 + ], + "type": "text", + "content": "Human: The kitchen utensils needed for preparation include a cutting board, a chef's knife, a small steamer pot, and an air fryer. The preparation process for the ingredients is as follows: 1. Cut the potatoes into small chunks. 2. Slice the leafy greens into segments. 3. Chop the purple cabbage. 4. Slice the carrots into thick pieces. 5. Halve the white mushrooms. 6. Slice the bell peppers. 7. Prepare a small bowl of soaked soybeans. 8. Cut the blanched beef pieces into strips. Finally, Place the vegetables from steps 1-7 into the steamer pot, and place the beef strips from step 8 into the air fryer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 271, + 476, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 271, + 476, + 279 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 476, + 279 + ], + "type": "text", + "content": "Question: How to install a CPU on a motherboard as demonstrated in the video? [ID:173]" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 134, + 279, + 476, + 312 + ], + "blocks": [ + { + "bbox": [ + 134, + 279, + 476, + 312 + ], + "lines": [ + { + "bbox": [ + 134, + 279, + 476, + 312 + ], + "spans": [ + { + "bbox": [ + 134, + 279, + 476, + 312 + ], + "type": "image", + "image_path": "4a8dbd0856ce054046a3c41e3d8e554ab80bfa799f73e02ec3179b989064722a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 316, + 476, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 476, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 476, + 348 + ], + "type": "text", + "content": "Sample Answer: According to the clip, here is how to install a CPU on a motherboard: 1. Lift the metal arm next to the CPU socket. 2. Look for a small triangle indicator on the corner of the CPU and the socket and align them properly. 3. Hold the CPU by the sides and insert the CPU into the CPU socket on the motherboard. 4. Lower the arm down completely into its original position." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 351, + 476, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 476, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 476, + 368 + ], + "type": "text", + "content": "Human: 1. Open the latch at the CPU location on the motherboard. 2. Place the side of the CPU with pins onto the corresponding location on the motherboard, aligning with the notches. 3. Close the latch." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 134, + 373, + 476, + 414 + ], + "blocks": [ + { + "bbox": [ + 134, + 373, + 476, + 414 + ], + "lines": [ + { + "bbox": [ + 134, + 373, + 476, + 414 + ], + "spans": [ + { + "bbox": [ + 134, + 373, + 476, + 414 + ], + "type": "image", + "image_path": "300c393ea160633bb5c97d40f917e1928e8ff235ede155f73ada9d5d268f0c1a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 476, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 476, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 476, + 449 + ], + "type": "text", + "content": "Sample Answer: First, put the pillow at the head of the bed; Second, put the clothes that need to be washed on the floor, and put the clothes that do not need to be washed on the chair; Third, pile up all the debris except for sheets quilts, and pillows on the floor; Fourth, after arranging the sheets and quilts, place the doll bear on the side near the head of the bed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 476, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 476, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 476, + 485 + ], + "type": "text", + "content": "Human: The order in which this person organizes the room is as follows: 1. Neatly arrange multiple pillows at the head of the bed. 2. Gather multiple items left on the bed onto a chair. 3. Remove other items from the bed, such as a large teddy bear and a white cushion. 4. Straighten the bedsheet to make it look even. 5. Place the teddy bear back at the head of the bed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 163, + 496, + 449, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 496, + 449, + 507 + ], + "spans": [ + { + "bbox": [ + 163, + 496, + 449, + 507 + ], + "type": "text", + "content": "Table 6: Hard cases for humans related to descriptive ability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 542, + 325, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 325, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 325, + 555 + ], + "type": "text", + "content": "5 Conclusion and Future Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "In this paper, we introduce AutoEval-Video, a novel benchmark for comprehensively assessing large VLMs in open-ended video question answering. Utilizing an LLM evaluator, models can be automatically and accurately evaluated on AutoEval-Video. In particular, GPT-4V achieves the highest accuracy of " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "32.2\\%" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ", still far from human accuracy of " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "72.8\\%" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ". Through an extensive case study, we identify several limitations of current models' capabilities in video understanding, such as temporal and dynamic comprehension, and overly general responses. These findings will shape our future research direction." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 341, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 341, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 341, + 102 + ], + "type": "text", + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 142, + 480, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 142, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 142, + 480, + 178 + ], + "type": "text", + "content": "We would like to express our sincere gratitude to the reviewers of ECCV 2024 for their insightful and constructive feedback. Their valuable comments have greatly contributed to improving the quality of our work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 199, + 198, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 199, + 198, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 199, + 198, + 213 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 224, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 258 + ], + "type": "text", + "content": "1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 258, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 334 + ], + "type": "text", + "content": "2. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, A., Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z., Zhang, J., Zhang, X., Zhang, Y., Zhang, Z., Zhou, C., Zhou, J., Zhou, X., Zhu, T.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 335, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 357 + ], + "type": "text", + "content": "3. Buch, S., Eyzaguirre, C., Gaidon, A., Wu, J., Fei-Fei, L., Niebles, J.C.: Revisiting the \"video\" in video-language understanding. In: CVPR. pp. 2917-2927 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "type": "text", + "content": "4. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR. pp. 961-970 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "type": "text", + "content": "5. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "text", + "content": "6. Fan, H., Lin, L., Yang, F., Chu, P., Deng, G., Yu, S., Bai, H., Xu, Y., Liao, C., Ling, H.: Lasot: A high-quality benchmark for large-scale single object tracking. In: CVPR. pp. 5374-5383 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 456, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 478 + ], + "type": "text", + "content": "7. Fu, J., Ng, S.K., Jiang, Z., Liu, P.: Gptscore: Evaluate as you desire. arXiv preprint arXiv:2302.04166 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 479, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 521 + ], + "type": "text", + "content": "8. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The \"something something\" video database for learning and evaluating visual common sense. In: ICCV. pp. 5842-5850 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "text", + "content": "9. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR. pp. 6904-6913 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "text", + "content": "0. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "type": "text", + "content": "1. Grunde-McLaughlin, M., Krishna, R., Agrawala, M.: Agqa: A benchmark for compositional spatio-temporal reasoning. In: CVPR. pp. 11287-11297 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "2. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., et al.: Ava: A video dataset of spatiotemporally localized atomic visual actions. In: CVPR. pp. 6047-6056 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "3. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "text", + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 139 + ], + "type": "text", + "content": "14. Hu, A., Chen, S., Zhang, L., Jin, Q.: Infomatic: An informative metric for reference-free image caption evaluation. arXiv preprint arXiv:2305.06002 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 173 + ], + "type": "text", + "content": "15. Huang, D.A., Ramanathan, V., Mahajan, D., Torresani, L., Paluri, M., Fei-Fei, L., Niebles, J.C.: What makes a video a video: Analyzing temporal information in video understanding models and datasets. In: CVPR. pp. 7366-7375 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 174, + 480, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 480, + 207 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 480, + 207 + ], + "type": "text", + "content": "16. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 208, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 208, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 480, + 231 + ], + "type": "text", + "content": "17. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV. pp. 706-715 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 232, + 480, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 480, + 254 + ], + "type": "text", + "content": "18. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: ICCV. pp. 2556-2563. IEEE (2011)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 255, + 480, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 480, + 276 + ], + "type": "text", + "content": "19. Lei, J., Berg, T.L., Bansal, M.: Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "type": "text", + "content": "20. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 344 + ], + "type": "text", + "content": "21. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 345, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 480, + 378 + ], + "type": "text", + "content": "22. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 379, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 379, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 379, + 480, + 402 + ], + "type": "text", + "content": "23. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 425 + ], + "type": "text", + "content": "24. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 425, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 436 + ], + "type": "text", + "content": "25. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL (2004)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 437, + 480, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 480, + 470 + ], + "type": "text", + "content": "26. Lin, Y.T., Chen, Y.N.: Llm-eval: Unified multi-dimensional automatic evaluation for open-domain conversations with large language models. arXiv preprint arXiv:2305.13711 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 471, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 504 + ], + "type": "text", + "content": "27. Liu, D., Qu, X., Hu, W.: Reducing the vision and language bias for temporal sentence grounding. ACM MM (2022), https://api_semanticscholar.org/CorpusID:251104851" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 505, + 480, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 480, + 528 + ], + "type": "text", + "content": "28. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 528, + 480, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 480, + 550 + ], + "type": "text", + "content": "29. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 551, + 480, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 480, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 480, + 573 + ], + "type": "text", + "content": "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 574, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 480, + 597 + ], + "type": "text", + "content": "31. Liu, Y., Iter, D., Xu, Y., Wang, S., Xu, R., Zhu, C.: G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 597, + 480, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 480, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 480, + 630 + ], + "type": "text", + "content": "32. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 631, + 443, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 631, + 443, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 631, + 443, + 643 + ], + "type": "text", + "content": "33. OpenAI: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL. pp. 311-318 (2002)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 341, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 341, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 341, + 102 + ], + "type": "text", + "content": "X. Chen, Y. Lin, Y. Zhang, and W. Huang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 632 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 481, + 160 + ], + "type": "text", + "content": "35. Pătrăucean, V., Smaira, L., Gupta, A., Continente, A.R., Markeeva, L., Banarse, D., Koppula, S., Heyward, J., Malinowski, M., Yang, Y., et al.: Perception test: A diagnostic benchmark for multimodal video models. arXiv preprint arXiv:2305.13786 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "text", + "content": "36. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "37. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "38. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "39. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR. pp. 4566-4575 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "40. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y.F., Wang, W.Y.: Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In: ICCV. pp. 4581-4591 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "41. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. NeurIPS 35, 24824-24837 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "42. Wu, Y., Lim, J., Yang, M.H.: Online object tracking: A benchmark. In: CVPR. pp. 2411-2418 (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "43. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR. pp. 5288-5296 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "44. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "45. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "46. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "47. Yu, Z., Zheng, L., Zhao, Z., Wu, F., Fan, J., Ren, K., Yu, J.: Anetqa: A large-scale benchmark for fine-grained compositional reasoning over untrimmed videos. In: CVPR. pp. 23191-23200 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "48. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "49. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "50. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "51. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "text", + "content": "AutoEval-Video: Benchmark for Large Vision Language Models in Video QA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_content_list.json b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..003aa0b6e07ff5d28f4493dfc293a31b2e7ffe5f --- /dev/null +++ b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_content_list.json @@ -0,0 +1,1833 @@ +[ + { + "type": "text", + "text": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos", + "text_level": 1, + "bbox": [ + 277, + 140, + 727, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ekta Prashnani, Koki Nagano, Shalini De Mello, David Luebke, and Orazio Gallo", + "bbox": [ + 230, + 210, + 771, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NVIDIA", + "bbox": [ + 470, + 237, + 529, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{eprashnani, knagano, shalinig, dluebke, ogallo}@nvidia.com", + "bbox": [ + 240, + 252, + 764, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Modern avatar generators allow anyone to synthesize photorealistic real-time talking avatars, ushering in a new era of avatar-based human communication, such as with immersive AR/VR interactions or videoconferencing with limited bandwidths. Their safe adoption, however, requires a mechanism to verify if the rendered avatar is trustworthy: does it use the appearance of an individual without their consent? We term this task avatar fingerprinting. To tackle it, we first introduce a large-scale dataset of real and synthetic videos of people interacting on a video call, where the synthetic videos are generated using the facial appearance of one person and the expressions of another. We verify the identity driving the expressions in a synthetic video, by learning motion signatures that are independent of the facial appearance shown. Our solution, the first in this space, achieves an average AUC of 0.85. Critical to its practical use, it also generalizes to new generators never seen in training (average AUC of 0.83). The proposed dataset and other resources can be found at: https://research.nvidia.com/labs/nxp/avatar-fingerprinting/.", + "bbox": [ + 259, + 296, + 743, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Synthetic Media Verification $\\cdot$ Avatar Generators $\\cdot$ Avatar Attribution", + "bbox": [ + 261, + 506, + 740, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 547, + 356, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent digital avatar generators have fueled a myriad of computer vision and graphics applications, allowing anyone to synthesize real-time photorealistic personas. Major companies are now supporting avatar-driven remote interactions over immersive AR/VR (e.g. Meta's Pixel Codec Avatar [48], Apple's Vision Pro Persona [1]) or video conferencing (e.g. NVIDIA's MAXINE [5], Microsoft's Teams [3]), and selfie filters for altering and enhancing appearance (e.g. by Snap and Tiktok). While the avatar generation technology today is still young, the legitimate use of synthetic avatars will be ubiquitous in the future. Without proper guardrails, this poses a real risk of unauthorized use and large-scale spread of visual disinformation. To ensure the safe use of such a technology, the relevant question is no longer whether the content is \"real\" or not—since, by design, the videos and avatars are all synthetic—but rather, whether the synthetically-generated videos and avatars are \"trustworthy\" or not.", + "bbox": [ + 212, + 580, + 787, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When video conferencing, for instance, a synthetic video portrait generator can be used to save valuable bandwidth by reconstructing a synthetic avatar of the sender at the receiver's end, using only a frame of the target identity and a compact representation of the speaker's facial motion. To ensure the authorized use of such synthetically-generated videos, we want to verify if the driving identity behind a synthetic video $(\\mathrm{ID}_1$ or $\\mathrm{ID}_2$ in Figure 1) is authorized to control the likeness, or the appearance, of the", + "bbox": [ + 212, + 753, + 789, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4eae27ba7b367c955894eebea718d85ee1012e9264b5c7b84c556e11c7b33fba.jpg", + "image_caption": [ + "Fig. 1: Talking-head avatar generators can synthesize realistic videos of a target identity from driving videos of different identities. Our method extracts appearance-agnostic temporal facial features and learns an embedding in which the synthetic videos driven by one identity fall close to each other and far from those driven by other identities, regardless of the appearance of the synthetic video. By comparing distances in the embedding space, we evaluate whether an avatar is driven by an authorized identity or not. During evaluation, we only rely on the synthetic videos as input, without requiring any prior knowledge about the driving identity." + ], + "image_footnote": [], + "bbox": [ + 292, + 143, + 707, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "synthetic video (target identity in Figure 1). Crucially, we want to only leverage the synthetic video avatars to do so. We call this novel task avatar fingerprinting.", + "bbox": [ + 212, + 433, + 785, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We leverage a simple but fundamental observation: facial motions tend to be idiosyncratic, that is, individuals talk and emote in unique ways. For instance, someone may raise one of her eyebrows more than the other, or smile more while talking. These \"dynamic identity signatures\" [51] have been shown to carry enough information for humans to recognize other individuals, even when the physical appearance of their face is altered [31, 41, 51]. This makes them attractive for our task, as they can be derived solely from the driving identity of a talking-head video, regardless of the appearance.", + "bbox": [ + 212, + 463, + 787, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fortunately, modern avatar generators are becoming increasingly accurate at capturing the facial motion of a person and rendering it onto a target identity. As a first solution to avatar fingerprinting, we then propose to estimate a dynamic identity signature for an identity from the synthesized videos they drove—regardless of the target identities shown. Specifically, we extract facial landmarks and their temporal dynamics from the video. We then introduce a novel contrastive loss to learn a dynamic identity embedding, a space where the dynamic identity signatures of a driving identity across multiple videos and target identities are close to each other, and far from those of other driving identities (see Figure 1). We show that this approach, albeit straightforward, is robust and generalizes to generators not seen in training.", + "bbox": [ + 212, + 565, + 787, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Avatar fingerprinting is a new task, and no existing datasets serve its training and validation requirements. There are two key requirements from a dataset. First, we need videos of multiple subjects delivering both scripted and free-form monologues, captured under realistic conditions, such as varying video quality and gaze direction. This allows us to assess if a model leverages talking styles, and not the specific choice of words. Second, to evaluate if the model learns to extract dynamic identity signatures effectively from synthetic videos, we need synthetic talking-head videos for the case in which the driving and target identities are different (cross-reenactment), and that in which they match (self-reenactment). Unfortunately, existing datasets of real videos", + "bbox": [ + 212, + 710, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "E. Prashnani et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "only fulfill a subset of these requirements (Table 1 and Section 4). Further, no existing dataset of synthetic videos contain both self- and cross-reenactments per subject or use the state-of-the-art talking-head generation technology. To foster research in this new domain, then, we introduce the NVIDIA Facial Reenactment (NVFAIR) dataset, containing real and synthetic face reenactment talking-head videos (Figure 2). Our dataset, which includes ethnically diverse subjects, provides $10 \\times$ more synthetic facial reenactments than the second largest dataset, for a total of over 650,000 synthetic videos. It is also the only one using multiple state-of-the-art face reenactment generators [34,60,62], or to provide cross-reenactments driven by all identities, which is critical for training and evaluating avatar fingerprinting algorithms.", + "bbox": [ + 218, + 146, + 787, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary:", + "text_level": 1, + "bbox": [ + 215, + 299, + 303, + 313 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the novel task of avatar fingerprinting, which focuses on verifying the driving identity of synthetic talking-head avatars, rather than classifying them as real or synthetic (by design, all inputs to our model are synthetic).", + "- We release the first large-scale dataset of subjects delivering scripted and natural monologues, complete with self- and cross-reenactment videos synthesized with multiple state-of-the-art generators.", + "- We propose a solution for this novel task in the context of video conferencing by extracting person-specific motion signatures, and demonstrate its robustness to various distortions and avatar generators not seen in training." + ], + "bbox": [ + 225, + 321, + 782, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 470, + 366, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed avatar fingerprinting task aims to verify authorized use of synthetic avatars: this is fundamentally a different problem than traditional forensics research where one aims to detect synthetic media (e.g. deepfake detection) or actively mark synthetic content. In our case, by design, the content being evaluated is always synthetic: we aim to evaluate its authorized use. Since this is a novel task, no methods exist to directly address it. No methods currently exist to directly address this novel task. Here we discuss the related areas of research.", + "bbox": [ + 218, + 500, + 789, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based Attribution of Synthetic Media. Learning-based approaches have been used to identify the origin of synthetic media, or to determine if it has been manipulated or altered in some way. Previous work used a pre-trained GAN generator to attribute a synthesized image to its generator via GAN inversion by leveraging the fact that a real image is less invertible [11, 36, 63]. Yet other works focus on attributing other forms of synthetic media, such as text [50]. In contrast, our focus is to attribute a talking-head avatar to the identity driving it, regardless of the appearance of the avatars. Some existing works learn fingerprints associated with cameras to determine whether a video is manipulated [19], or embed watermarks into images and videos [12, 25, 47, 58], which are also shown to transfer to GAN-generated images [64]. Subsequent research introduced a watermark-based conditional GANs for scalable fingerprinting [65]. Our method, in contrast, is a passive technique that does not rely on active watermarks.", + "bbox": [ + 218, + 604, + 787, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deepfake Detection Based on Identity-Specific Features. Deepfake detection (\"is a video synthetic?\") and avatar fingerprinting (\"whose identity is used to generate this synthetic video?\") are fundamentally different tasks. Most existing solutions for deepfake detection train a real-vs-synthetic classifier [26, 28, 43, 52, 57, 68], and therefore", + "bbox": [ + 218, + 781, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos", + "bbox": [ + 281, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cannot be adapted to avatar fingerprinting (where all inputs are synthetic). However, a specific class of detectors leverage identity-specific features to detect synthetic videos by posing the detection problem as an identity-recognition problem. In our experiments, we evaluate some of these methods as baselines. Specifically, Agarwal et al. exploit person-specific patterns in facial expressions to detect fake videos [9]. ID-Reveal used facial shapes and motions encoded in a low-dimensional space of a 3D morphable model [13] to handle both face-swapping and face-reenactment deepfakes [20]. Other works explored soft-biometric approaches such as leveraging vocal mannerisms [15], phoneme-viseme consistencies [8], word-facial expression consistencies [6,10], and dynamics of ears [7]. While many of these works need person-specific training, previous works [6, 18, 20, 45] extended this idea to train a CNN-based detector using a large-scale in-the-wild video data [17] and variants of contrastive learning [27, 38, 55, 61]. Agarwal et al. combined static facial appearance using a facial recognition model and dynamic facial behaviors using a CNN, and showed that this approach is effective for detecting face-swap deepfakes [6]. Yet another line of research has explored temporal inconsistencies of face identities within a video [45], and identity inconsistencies of inner and outer face regions [22]. Our experiments show that for avatar fingerprinting, such features that are designed to distinguish real from synthetic videos are not reliable, since we only have access to synthetic videos.", + "bbox": [ + 217, + 146, + 785, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Facial Identity Signatures. Cognitive scientists have studied the impact of \"dynamic facial identity signatures\" (i.e., characteristic or identity-specific movements of a face) for identity recognition for humans [51]. In one experiment, scientists projected facial animations generated by human actors onto an average head and found that subjects discriminated between individuals based solely on facial motion [31]. In another, subjects correctly attributed animations of synthetic faces to their morphed versions [41]. While these studies point to the existence of \"dynamic facial identity signatures\" that humans rely on, ours is the first method that isolates these from videos. Talking-Head Datasets and Generators. Existing talking-head datasets include those that contain only real videos showing a variety of emotions [16, 46, 59], and others that also contain synthetic videos [21, 24, 30, 35, 42, 44, 53]. These datasets cater to traditional forensics and facial analysis tasks. Therefore, they do not contain self-and cross-reenactments driven by multiple identities, as well as scripted and free-form monologues across diverse capture settings. There novel requirements posed by the avatar fingerprinting task motivate us to design our own dataset. We focus specifically on face-reenactment talking-head avatar generators for synthetic video generation—this class of generators are the most relevant to AR/VR interactions, video conferencing, and several other applications [2-5, 48]—and combine various modes of human expression for capturing real videos. Given a target facial image and a driving video, these generators reenact the target image using the facial expressions and head pose from the driving video [23, 33, 34, 37, 49, 56, 60, 62, 66, 67]. Another class of talking-head generators use person-specific models [40] and some models aim to preserve the style of the target identity in the synthesized video [39]. However, these models require person-specific training, making them difficult to scale.", + "bbox": [ + 217, + 421, + 785, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Terminology", + "text_level": 1, + "bbox": [ + 217, + 784, + 354, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We seek to verify the trustworthiness of a synthesized talking-head video, termed target video. We assume that an avatar-generation tool (e.g., [60]) created it by animating", + "bbox": [ + 217, + 811, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "E. Prashnani et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8a04554487271f3aa6ec5972c84632c6fe7d5ba5b7d1cfea152b0541cc4272d8.jpg", + "image_caption": [ + "Fig. 2: We introduce the NVFAIR dataset, containing real and synthetic talking-head videos. We capture subjects talking in both scripted and free-form settings. To encourage natural performance, we record the subjects while videoconferencing with each other (left). We then synthesize more than 650,000 talking-head videos—the largest collection till date—using three state-of-the-art face-reenactment talking-head generators. On the right, each row corresponds to a driving identity $(\\mathrm{ID}_i \\rightarrow (\\cdot))$ and each column corresponds to a different target identity $((\\cdot) \\rightarrow \\mathrm{ID}_i)$ . The videos in which driving and target identity match are self-reenactments, the rest are cross-reenactments." + ], + "image_footnote": [], + "bbox": [ + 217, + 141, + 787, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "an image (target image) using the expressions and head poses obtained from another video, the driving video. We call driving identity the identity of the person in the driving video, and target identity the identity of the person in the target image. When driving and target identities match, the target video is a self-reenactment, while the case of a driving identity used to animate a different target identity is cross-reenactment. In both cases, the appearance of the synthesized video is derived from the target identity. This terminology allows us to formally state our goal: we want to verify that a target video is a self-reenactment. With this terminology in mind, we introduce our dataset, which includes real videos as well as self- and cross-reenactment videos.", + "bbox": [ + 212, + 369, + 787, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 The NVIDIA Facial Reenactment (NVFAIR) Dataset", + "text_level": 1, + "bbox": [ + 212, + 516, + 689, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall that avatar fingerprinting is not about detection of synthetic media. Rather, we already know a video to be synthesized, and seek to verify that the driving identity is authorized. This new task dictates a set of requirements for the dataset to be effective for training and evaluation. Specifically, we need a dataset that contains", + "bbox": [ + 212, + 540, + 787, + 595 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. multiple real videos per identity, with scripted and free-form conversations, and with both natural and prescribed emotions,", + "2. self- and cross-reenactments of target identities, with cross-reenactments driven by all subjects to allow for a variety of driving facial structures, and", + "3. multiple face-reenactment generators." + ], + "bbox": [ + 220, + 598, + 784, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All relevant existing datasets only capture a subset of these requirements (see Table 1; the Supplement contains further discussion). Moreover they do not use the state-of-the-art talking-head generation technology to synthesize the self- and cross-reenactment videos. We introduce the NVFAIR dataset that features all of the above properties. With over 650,000 synthetic videos, it provides $10 \\times$ as many videos as the next largest dataset, and uses three different state-of-the-art generators for the self- and cross-reenactment videos. Figure 2 shows an overview of data capture and synthesis.", + "bbox": [ + 212, + 669, + 818, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Real Data Capture", + "text_level": 1, + "bbox": [ + 214, + 789, + 388, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Capturing videos of monologues delivered by different subjects for the purpose of identity verification introduces two conflicting goals. On the one hand a controlled evalua-", + "bbox": [ + 212, + 810, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos", + "bbox": [ + 281, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9063d62166d5ead3ef8c3e401f6bc6c5c467c5c5f98c47bfbaed2f9a0e2af457.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset# Subjects (source)(F)ree / (S)scripted?Emotion: (N)atural / (P)rescribed(R)eel / (S)elf / (C)ross-reenact.?#Face Reenact.Avg. Videos per Subject# Face- reenact. GeneratorsEthnic Diversity
RAVDESS [46]24 (new)(S)(P)(R) onlyN/A120 (R)N/A✓
MEAD [59]60 (new)(S)(P)(R) onlyN/A720 (R)N/A✓
CREMA-D [16]91 (new)(S)(P)(R) onlyN/A81 (R)N/A✓
VFHQ [24]36 ([46])(S)(P)+(N)(R)+(S)1,7371201✓
FF++ [54]1000 (YT)(F)(N)(R)+(C)2,0001(R)+2(C)2✓
KoDF [42]403 (new)(F)+(S)(N)(R)+(C)61,000150(R)+151(C)1X
NVFAIR (Ours)161 (46 new + [46], [16])(F)+(S)(P)+(N)(R)+(S)+(C)654,72676(R)+228(S) +3,840(C)3✓
", + "bbox": [ + 215, + 143, + 785, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: The existing talking-head video datasets were designed for tasks such as deepfake detection or facial emotion analysis. Avatar fingerprinting is a fundamentally different task. As a result, no existing dataset satisfies the requirements for training and evaluating models for avatar fingerprinting. To overcome this limitation, we introduce the NVFAIR dataset, which is the first dataset that offers the complete set of monologue modalities, and features the largest collection of facial reenactments to date. Specifically, it provides scripted and free-form monologues, with natural and prescribed emotions, and self- and cross-reenactments (driven by all remaining subjects) synthesized using three generators, alongside original videos for newly-recorded subjects.", + "bbox": [ + 212, + 277, + 785, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion of the trained models requires predictability of what is spoken to prevent identification algorithms from latching onto the spoken content itself. On the other, we want the subjects to act as they would in a casual conversation, rather than reciting memorized text, to capture their uniquely identifying mannerisms. We address this trade-off by recording the subjects while videoconferencing in pairs, which creates the impression of being in a natural conversation. This differs from existing datasets, in which the subjects look at the camera, but are not interacting with another person during the recording [16,42,46]. We also design two distinct recording strategies: a free-form stage where the subjects are given only general guidance on the topics, and a more controlled scripted stage in which subjects speak short, memorized monologues of 2-3 sentences each, see Figure 2(a). To capture the variability of real-life scenarios, we provided minimal instructions on how to setup the video call, allowing for diverse face, scale, and lightning, bandwidth stability, and background scene clutter. In total we record 46 subjects of diverse genders, ages, and ethnicities, while strictly abiding by a pre-approved IRB protocol (see Supplement for details and privacy considerations).", + "bbox": [ + 212, + 404, + 785, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage I: Free-Form Monologues. In this first stage, the two subjects on the call alternate between asking and answering seven pre-defined questions. The questions are designed to avoid sensitive or potentially inflammatory topics. This is critical because we later use sentences spoken by one individual to animate the video of a second individual, quite literally putting words in their mouths. The complete list of questions is in the Supplement. To further create a natural interaction, the subject listening is encouraged to actively but silently engage with the one speaking (e.g., by nodding or smiling).", + "bbox": [ + 212, + 636, + 785, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage II: Scripted Monologues. For this stage, we prepared thirty short utterances consisting of two or three sentences each. We chose this length to allow for memorization, while still providing enough content to trigger facial expressions. However, to avoid inducing unnatural expressions, we do not prescribe specific emotions for each utterance. For instance, we do not ask to express anger for a sentence, but we do choose sentences that may naturally evoke it, and used punctuation to encourage it, e.g., \"Will you please", + "bbox": [ + 212, + 753, + 785, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "E. Prashnani et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "answer the darn phone? The constant ringing is driving me insane!\" We instruct the subjects to split their screens to show both this list and video call and encourage them to speak to their recording partner when reciting, see Stage II in Figure 2(a). More details, including the full list of utterances can be found in the Supplement.", + "bbox": [ + 215, + 146, + 785, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Synthetic Talking-Head Videos", + "text_level": 1, + "bbox": [ + 215, + 226, + 470, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using the videos described in Section 4.1, as well as the original videos from the CREMA-D [16] and RAVDESS [46] datasets, we generate synthetic talking-head videos to train and evaluate our avatar fingerprinting algorithm. Specifically, we pool the 91 identities from the original videos of CREMA-D [16], the 24 identities from those of RAVDESS [46], and the 46 from our own video-conferencing data capture, for a total of 161 unique identities $\\mathcal{I}$ . Recall that we have several real videos for each identity $\\mathrm{ID}_i\\in \\mathcal{I}$ . To avoid a combinatorial explosion of synthetic videos, for all pairs of identities $\\mathrm{ID}_i$ and $\\mathrm{ID}_j$ , we use $\\mathrm{ID}_j$ as the target identity and we randomly select 8 of the videos of $\\mathrm{ID}_i$ to generate 8 cross-reenactment videos, $\\{\\mathrm{ID}_i^k\\to \\mathrm{ID}_j\\}_{k = \\{1,\\dots ,8\\}}$ (all 8 share the same target image). We also generate self-reenactment videos for each of the target identities, by animating their neutral-face images derived from captured videos with each of their real videos.", + "bbox": [ + 215, + 251, + 787, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use three different generators for synthesizing the videos for all the 161 identities: face-vid2vid [60], LIA [62], and TPS [34]. This allows us to test if our model generalizes across generators. We chose these talking-head generators because they are the state of the art and they preserve the identity-specific facial motion dynamics well. Nevertheless, the reconstruction is not perfect; for instance, in the third row of Figure 2(b) the person in the driving video $(\\mathrm{ID}_5)$ is squinting, but the eyes are shut in all the synthetic videos, including the self-reenactment video. In total we generate more than 650,000 synthetic videos, which required more than 2,500 RTX 3090 GPU hours. More details are in the Supplement.", + "bbox": [ + 215, + 426, + 787, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Method", + "text_level": 1, + "bbox": [ + 215, + 577, + 313, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overview. We seek to verify the driving identity of a synthetic video, independently of the target identity. We leverage the finding from cognitive science research that each person emotes in unique ways when communicating, and that this signal is sufficient for recognition, even when the actual appearance is artificially corrupted [31, 41, 51]. We note that our method does not latch onto artifacts introduced by the generators—a property that we demonstrate by showing generalization to new generators not seen during training. Rather, our features capture the dynamics of the expressions, like the way a person frowns, or the way she smiles. Notably, they are distinct from the temporal artifacts introduced by the generator, and that existing algorithms use to detect whether a video is synthetic or real [29].", + "bbox": [ + 215, + 609, + 787, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "An overview of our algorithm for avatar fingerprinting is shown in Figure 3. To capture expressions, we extract the relative position of facial landmarks over time from the input video, as shown in Figure 3 (Section 5.1). We learn to project these temporal signatures onto a dynamic identity embedding in which features belonging to the same driving identity are close to each other regardless of the target identity, i.e., independently of appearance (Figure 1). To learn this embedding we train a 3DCNN (where the", + "bbox": [ + 215, + 753, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos", + "bbox": [ + 281, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b93e19b071ac1ecc9fc948c81a089ccf1213e326648c1c47d17fbe0d1e9245e9.jpg", + "image_caption": [ + "Fig. 3: We extract landmarks from the frames of a talking-head clip, compute their normalized pairwise distances, and concatenate the frame-wise features. We then learn an identity embedding using a loss that pulls closer features of videos driven by the same identity and pushes away those driven by others. $\\mathrm{ID}_i\\rightarrow \\mathrm{ID}_j$ indicates a video that looks like identity $j$ (the \"target\" identity), and is driven by identity $i$ ." + ], + "image_footnote": [], + "bbox": [ + 292, + 143, + 707, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "third dimension spans video frames—a temporal CNN) with a novel contrastive loss that pulls together all embedding vectors of synthetic videos driven by an individual, while pushing away the embedding vectors of videos driven by all others individuals (Section 5.2). More implementation details are in Section 5.3.", + "bbox": [ + 212, + 378, + 787, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Dynamic Facial Identity Features", + "text_level": 1, + "bbox": [ + 214, + 457, + 488, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our first step is to extract temporal features that summarize short segments of the video we wish to fingerprint. We identify the following guiding principles for the extracted features. We would like features that:", + "bbox": [ + 212, + 479, + 785, + 523 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. have minimal dependency on the appearance of the face in the video (that is, the target identity),", + "2. reflect the dynamics of the expressions, and", + "3. capture subtle expressions." + ], + "bbox": [ + 220, + 523, + 782, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One choice could be per-frame 3DMM features [14]: a strategy also used by Cozzolino et al. to detect synthetic videos [20]. However, we empirically observe that 3DMM features are not sufficiently expressive, and do not satisfy desideratum 3 (see ablation experiments in Section 6). We observe a similar behavior for action units [9]. Facial landmarks [32] address this issue, but are sensitive to the shape of the face in the video, and thus to the target identity.", + "bbox": [ + 212, + 580, + 785, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To leverage the expressiveness of facial landmarks while abstracting from the underlying facial shape, we compute the pairwise normalized Euclidean distances between each pair of landmarks of a frame. We concatenate these distances into a single vector for the frame, $\\mathbf{d}_f$ . A subset of the facial landmarks and distances are shown in Figure 3.", + "bbox": [ + 212, + 667, + 785, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then break the input video into clips, which are sequences consisting of F frames and offset by one frame (e.g., [1,F], [2,F+1], etc.), and concatenate the vectors from all the frames in each clip. Using the change in the relative position of the landmarks over a short period of time (the length of a clip) allows us to capture temporal dynamics with minimal dependence on the absolute position of each landmark, i.e., independently of the shape of the face.", + "bbox": [ + 212, + 724, + 785, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show empirically that our features are a good representation for our task, by comparing against alternative choices for input features such as 3DMM (Section 6).", + "bbox": [ + 212, + 810, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "E. Prashnani et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 Dynamic Identity Embedding Contrastive Loss", + "text_level": 1, + "bbox": [ + 215, + 146, + 584, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While the features described in Section 5.1 are designed to extract low-level motion dynamics, they cannot be used directly to disambiguate two target videos based on their driving identities. We tackle this problem by learning a dynamic identity embedding, a space where videos driven by one subject map to points that are close to each other and far from the videos driven by anybody else.", + "bbox": [ + 212, + 167, + 782, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, we use a temporal 3DCNN to extract an embedding vector from a clip, which, as described before, is a short segment of an input video. To train the network we use a dataset of synthetic videos driven by different identities. We denote as $\\mathcal{C}_{\\mathrm{ID}_1\\to \\mathrm{ID}_2}^k (t)$ the embedding produced by the network for the clip starting at time $t$ in the $k$ -th video, of a target identity $\\mathrm{ID}_2$ driven by identity $\\mathrm{ID}_1$ . As stated above, we have two main objectives, which we capture with the following terms in our proposed loss function.", + "bbox": [ + 212, + 239, + 784, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We Want to Pull Together All the Videos Driven by $ID_{1}$ . To achieve this, we define the following term:", + "bbox": [ + 214, + 339, + 782, + 368 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {N} _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\sum_ {\\mathrm {I D} _ {l}, k} \\max _ {n} s \\left(\\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {2}} ^ {j} (t), \\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {l}} ^ {k} (n)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 376, + 785, + 407 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $s(\\cdot, \\cdot) = e^{-\\|\\cdot\\|^{2}}$ is a similarity metric. Intuitively, Equation 1 takes two videos, $j$ and $k$ , both driven by $\\mathrm{ID}_1$ . Given a clip starting at time $t$ in the first video, it looks for the most similar clip in the second video. Since the driving identity is the same for both videos, Equation 1 encourages an embedding where clips that capture a similar expression are closer to each other. Equation 1 is high even if only one clip from video $k$ has a similar temporal signature to $\\mathcal{C}_{\\mathrm{ID}_1 \\to \\mathrm{ID}_2}^j(t)$ . That is because even just one occurrence of the same expression is evidence that the driving identity may be the same. Of course, other driving identities may use similar expressions and we address that with the loss term described below. Additionally, we note that $k$ spans the set of all videos driven by $\\mathrm{ID}_1$ , and $\\mathrm{ID}_l$ spans all identities, including $\\mathrm{ID}_l = \\mathrm{ID}_1$ and $\\mathrm{ID}_l = \\mathrm{ID}_2$ .", + "bbox": [ + 214, + 419, + 787, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We Want to Push Away Videos not Driven by $ID_{1}$ . We define the following term:", + "bbox": [ + 215, + 578, + 746, + 593 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {Q} _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\sum_ {\\mathrm {I D} _ {l} \\neq \\mathrm {I D} _ {1}, k} \\max _ {n} s \\left(\\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {2}} ^ {j} (t), \\mathcal {C} _ {\\mathrm {I D} _ {l} \\rightarrow \\mathrm {I D} _ {2}} ^ {k} (n)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 601, + 785, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where, similarly to Equation 1, we take a clip from video $j$ , and look for the most similar clip in video $k$ . This time the two videos share the same target identity, but are driven by different identities: we want all the videos driven by identities different from $\\mathrm{ID}_1$ to be pushed away from those driven by $\\mathrm{ID}_1$ , including videos where $\\mathrm{ID}_1$ is the target identity. Note that $\\mathrm{ID}_2$ spans all identities, including $\\mathrm{ID}_2 = \\mathrm{ID}_1$ and $\\mathrm{ID}_2 = \\mathrm{ID}_l$ .", + "bbox": [ + 214, + 642, + 782, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We Want to Rely on the Temporal Dynamics of the Videos Driven by $ID_{1}$ . Although we use a temporal CNN, the model could still learn to rely on static expressions, such as a snapshot of the person smiling, rather than the temporal progression of expression leading to, or following, the smile. To further encourage the model to learn from the temporal dynamics, we introduce an additional term:", + "bbox": [ + 214, + 726, + 782, + 799 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R} _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\sum_ {\\mathrm {I D} _ {l}, k} \\max _ {n} s \\left(\\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {2}} ^ {j} (t), \\tilde {\\mathcal {C}} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {l}} ^ {k} (n)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 806, + 785, + 844 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos", + "bbox": [ + 279, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathcal{C}}_{\\mathrm{ID}_1\\to \\mathrm{ID}_l}^k$ denotes a version of the clip $\\mathcal{C}_{\\mathrm{ID}_1\\to \\mathrm{ID}_l}^k$ from Equation 1 with randomly shuffled frame ordering. We want such time-shuffled versions of the clips driven by $\\mathrm{ID}_1$ to be pushed away from the pristine clips driven of $\\mathrm{ID}_1$ . Effectively, this means that the driving identity of the time-shuffled clips is regarded as different from $\\mathrm{ID}_1$ . In other words, we want to pull together video clips in the learned embedding space only when the temporal facial dynamics are characteristic of $\\mathrm{ID}_1$ . We further show the importance of this term in Section 6.3.", + "bbox": [ + 212, + 143, + 787, + 253 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Combining Equations 1, 2, and 3, we write the probability that the embedding vector $\\mathcal{C}_{\\mathrm{ID}_1\\to \\mathrm{ID}_2}^j (t)$ lies close to the embedding vectors for all video clips driven by $\\mathrm{ID}_1$ and far from all the videos driven by others as", + "bbox": [ + 212, + 253, + 787, + 301 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\np _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\frac {\\mathrm {N} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t)}{\\mathrm {N} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t) + \\mathrm {Q} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t) + \\mathrm {R} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 313, + 785, + 349 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and the complete loss term as", + "bbox": [ + 214, + 359, + 415, + 375 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}, t} - \\log \\left(p _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 388, + 785, + 424 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Implementation", + "text_level": 1, + "bbox": [ + 214, + 450, + 370, + 465 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Parameter Choices. To extract the per-frame dynamic facial identity features $\\mathbf{d}_f$ , we detect 126 facial landmarks for each frame [32], and compute the per-frame normalized pairwise Euclidean distances between these landmarks. The clip duration is set to 71 frames. We find that this is sufficient to capture the facial dynamics that are meaningful for avatar fingerprinting, while also maintaining a good trade-off between speed and accuracy. We also experiment with shorter-duration video clips (see Supplement). The input tensor to the temporal CNN is obtained by concatenating $\\mathbf{d}_f$ across 71 frames. In each batch, we include 8 unique identities. For each identity $\\mathrm{ID}_i$ , the pull term (Equation 1) comprises 16 clips: 8 are self-reenactments, randomly sampled from the full set, and the remaining are cross-reenactments with $\\mathrm{ID}_i$ as the driving identity. This allows the neural network to pull together videos based purely on the facial motion, regardless of the appearance of the video. The push term (Equation 2) for $\\mathrm{ID}_i$ consists of clips with the remaining 7 identities in the batch serving as driving identities (8 clips per driving identity), as well as the time-shuffled self-reenactments of $\\mathrm{ID}_i$ (Equation 3). Additional training details can be found in the Supplement.", + "bbox": [ + 212, + 481, + 787, + 698 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training, Validation, and Testing Datasets. Of the 161 total identities (pooling together the identities from our dataset, RAVDESS, and CREMA-D, see Section 4.2), we reserve 35 for testing, 14 for validation, and 112 for training. We ensure that there are no cross-set cross-reenactments: that is, identities in the training set only drive other training-set identities (and similarly for the validation and test sets). This allows us to test the models on facial appearances and expressions that were not seen during training. To evaluate the generalization of trained models to new generators, we train our network on videos generated for the training set identities using one generator, and evaluate on the synthetic videos of test-set identities synthesized using remaining two generators.", + "bbox": [ + 212, + 709, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "E. Prashnani et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2856cf1c7e156858dd5e3054c4b3c1d882ad6ad1da1715cf87f1feae9ff52e88.jpg", + "image_caption": [ + "$\\mathrm{ID}_6 \\rightarrow \\mathrm{ID}_6$ d=0.84" + ], + "image_footnote": [], + "bbox": [ + 272, + 142, + 385, + 241 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ec4805d1f008fd55e310f48309f9773b78166d3cc9be3f108d15495835f6f0ab.jpg", + "image_caption": [ + "$\\mathrm{ID}_6 \\rightarrow \\mathrm{ID}_1$ d=0.95" + ], + "image_footnote": [], + "bbox": [ + 388, + 142, + 500, + 241 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7d52bc44a8a860f9b99c6870a2f92ababe5ce647a9e786f82496f24f66de7aaf.jpg", + "image_caption": [ + "$\\mathrm{ID}_1 \\rightarrow \\mathrm{ID}_6$ d=1.58" + ], + "image_footnote": [], + "bbox": [ + 501, + 142, + 614, + 241 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7f4e3e07804a221975f738c25106a4f6cd1d17add076c2b0a06eb57bb735375f.jpg", + "image_caption": [ + "$\\mathrm{ID}_3 \\rightarrow \\mathrm{ID}_6$ d=2.11" + ], + "image_footnote": [], + "bbox": [ + 617, + 154, + 728, + 241 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f78e140b5a9a6085818b9e3867cf826e6984a2bbe7bacd59fda9f585ba8ca1bc.jpg", + "image_caption": [ + "$\\mathrm{ID}_2 \\rightarrow \\mathrm{ID}_2$ d=0.589", + "Fig. 4: Animated figure. Open in a media-enabled viewer like Adobe Reader and click on the inset. Our embeddings capture the dynamics of an expression, rather than the appearance of the face. For each row, we pick a reference identity. The green box indicates reenactments driven by the reference identity, the red and blue are cross-reenactments of the reference identity. We compute the average distance of each clip shown here against all other clips driven by the reference identity. The average distance to the other clips of the reference identity is consistent for a given motion, and lower (better) when the reference identity is driving as compared to the cross-reenactments that use the reference identity as target. Here, we show videos generated by face-vid2vid [60], and use the embedding vectors predicted by the model trained on the same generator (see Figure 6 generalization to new generators not seen during training)." + ], + "image_footnote": [], + "bbox": [ + 272, + 253, + 385, + 351 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0aa8dcad2786fb3228b0e2ba6b47264d30154270e9a53ba93090fb1fcbfefb82.jpg", + "image_caption": [ + "$\\mathrm{ID}_2 \\rightarrow \\mathrm{ID}_7$ d=0.961" + ], + "image_footnote": [], + "bbox": [ + 390, + 253, + 500, + 351 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/981c1d7f1581ec1c50eff2f0ba39936413500444ef06494e776cff9242966e9e.jpg", + "image_caption": [ + "$\\mathrm{ID}_7 \\rightarrow \\mathrm{ID}_2$ d=2.042" + ], + "image_footnote": [], + "bbox": [ + 501, + 265, + 614, + 351 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d089d3e5d94a267d087796094d566c27637b70e4d8ebca94c70d5da4f051a52a.jpg", + "image_caption": [ + "$\\mathrm{ID}_8 \\rightarrow \\mathrm{ID}_2$ d=1.787" + ], + "image_footnote": [], + "bbox": [ + 617, + 265, + 728, + 351 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 521, + 341, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We thoroughly evaluate our algorithm both qualitatively and quantitatively. Our algorithm outperforms reasonable baselines (Section 6.1), it generalizes to generators not seen in training (Section 6.2), and it is robust to video compression (Supplement). We also perform a number of ablation studies to analyze our design choices: our input features, the importance of the time-shuffling term in the loss function (Section 6.3), and the impact of clip duration (Supplement).", + "bbox": [ + 212, + 564, + 787, + 651 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We begin by evaluating qualitatively our method's ability to extract embedding vectors based on the driving identity. Figure 4 shows a set of self- and cross-reenacted clips (please view the animation in a media-enabled viewer, such as Adobe Acrobat). For each row, we take one identity as reference and we compute the embedding vectors of clips that use it both as the driving and the target identity. We then compute the average Euclidean distance of the resulting embedding vectors against those of the self-reenactments by the same reference identity. We note that the average distance $d$ is lower when the driving identity matches the reference identity (first two columns). We also note that the distances between the clips in the first two columns are similar: this confirms that the distance is a function of the facial motion, rather than the facial appearance. When the driving identity changes, the average distance increases, even if the target identity matches the reference identity, which is precisely our goal. More results are in the Supplement.", + "bbox": [ + 212, + 652, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos", + "bbox": [ + 281, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/478bf99b11578ecc7a6ae33c0f8d5184a9db71068e5815d85d685da73eba7339.jpg", + "image_caption": [ + "Fig.5: ROC curves and AUC values for our method and two baselines: Agarwal et al. [9] and ID-Reveal [20]. Each sub-plot shows the results on our test set for each of the three talking-head generators: face-vid2vid [60], LIA [62], and TPS [34]." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 750, + 258 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To evaluate our approach more formally, we use the 35 unique test-set identities that are not used as driving or target identities in the training set (Section 5.3). One at a time, we treat each identity $\\mathrm{ID}_i$ as target and synthesize cross-reenactments using all the remaining identities as drivers. This is the set of \"unauthorized\" synthetic videos for $\\mathrm{ID}_i$ . The self-reenacted samples for $\\mathrm{ID}_i$ form the \"authorized\" set. Note that there are several self-reenacted videos of $\\mathrm{ID}_i$ , one per original video of $\\mathrm{ID}_i$ .", + "bbox": [ + 212, + 318, + 782, + 404 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For each target identity $\\mathrm{ID}_i$ , we extract the dynamic identity embedding vector of all the clips in the pool of its self- and cross-reenacted videos, and compute their Euclidean distances. That is, for clip $k$ we compute", + "bbox": [ + 212, + 405, + 784, + 448 + ], + "page_idx": 11 + }, + { + "type": "equation", + "text": "\n$$\nd \\left(\\mathcal {C} _ {\\mathrm {I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}} ^ {k}, \\mathcal {C} _ {\\mathrm {I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}} ^ {l}\\right), \\forall l \\neq k, \\text {a n d}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 458, + 609, + 477 + ], + "page_idx": 11 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l}u \\left(\\mathrm {I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}, \\mathrm {C I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}\\right), \\quad \\forall i, \\quad \\forall i \\neq j.\\end{array}\\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 470, + 784, + 498 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We threshold these distances for each target identity to get an ROC curve, and average across the target identities to get the overall area under the curve (AUC). We note that this AUC measures one model's ability to classify a synthetic video as self-reenactment or as cross-reenacted. We conduct further analysis of our model's ability to classify other categories of videos—such as, evaluating AUC on same-utterance self- vs. cross-reenactments or on scripted vs. free-form monologues—in the Supplement.", + "bbox": [ + 212, + 507, + 787, + 594 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.1 Comparisons with Existing Methods", + "text_level": 1, + "bbox": [ + 214, + 614, + 511, + 630 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Avatar fingerprinting is a novel task, and no existing methods directly address it. The closest related works aim at detecting real versus synthetic media. As discussed in Section 2, some of these detectors learn identity-specific features such as facial expressions and head poses [9], or facial shapes and motion [20] and can serve as baselines for the task of avatar fingerprinting with some adaptation. The work by Agarwal et al. trains a model to detect synthetic videos of a specific identity [9]. To adapt it to our task, we train 35 different models, one for each identity in the evaluation, by splitting the corresponding original videos into two subsets. We then test each model on the self- and cross-reenactment videos of the corresponding identity. ID-Reveal, trained on a large-scale dataset, learns an embedding space where real videos of a specific identity are grouped together [20]. Since it shows good generalization to new identities, for the task of synthetic media detection, we directly use the pre-trained model on our data to detect, once again, self- versus cross-reenactment. Figure 5 shows the ROC curves for our method compared to these baselines, on three face-reenactment generators (face-vid2vid [60],", + "bbox": [ + 212, + 638, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "E. Prashnani et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6134ed937975c61d817d7a7c19087cca5cc2543f7f5fb83ff84fd9586f142441.jpg", + "image_caption": [ + "Fig. 6: Generalization to new generators. To study the robustness to new talking-head generators, we train three versions of our model on three different generators and test on all three." + ], + "image_footnote": [], + "bbox": [ + 250, + 143, + 415, + 260 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f8d69fd2100d41683e7984a489e0ff7557bb23b487ec3845844064087ebd3962.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 143, + 578, + 260 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1432e080d0917739ae6197e7975d0ced0d0f909e9a0f7a8ff6b4e9afd15b4c9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 143, + 745, + 261 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "LIA [62], and TPS [34]). Our method (AUC=0.868 on face-vid2vid) outperforms by a wide margin both ID-Reveal (AUC=0.720 on face-vid2vid), and the method by Agarwal et al. (AUC=0.726 on face-vid2vid). We also note that, unlike ID-Reveal and our method, Agarwal et al. uses a different model per identity.", + "bbox": [ + 212, + 303, + 784, + 361 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.2 Generalization to New Generators", + "text_level": 1, + "bbox": [ + 214, + 378, + 495, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For an avatar fingerprinting model algorithm to be broadly applicable, generalization to new talking-head generators that are not seen in training is crucial. Since our dataset contains videos synthesized by three different generators, we can train three models, one with each generator, and test these models on all three generators. Figure 6 shows the resulting ROC curves and AUC values: the overlap of the curves and similar AUC values in each subplot confirms that our method generalizes well to new generators.", + "bbox": [ + 212, + 401, + 784, + 488 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 506, + 362, + 520 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our method outperforms existing baselines by introducing two novel components: the dynamic facial identity features, which capture the facial dynamics in a compact and expressive way, and the loss function, which defines the shape of the identity embedding. Here we study the contribution of each, using the face-vid2vid generator for training and testing. We evaluate the contribution of our dynamic facial identity features by swapping them with 3DMM features [13], a popular choice to capture facial dynamics. Since we use a temporal CNN backbone similar to the one from ID-Reveal, for this ablation we use the loss function proposed in their original paper [20]. We re-train the same network using our features and observe a jump from 0.718 to 0.754 in terms of AUC. Upon inspection we notice that the 3DMM features tend to over-smooth the facial motion, and are unable to capture subtle dynamics that prove critical to avatar fingerprinting, and which our features capture. We also evaluate the contribution of our dynamic identity embedding loss and observe a further improvement (AUC 0.868). With our loss formulation, the advantage of $\\mathbf{R}_{j,\\mathrm{ID}_1,\\mathrm{ID}_2}(t)$ in Eq. 5 is also evident when compared against a model trained without this term (AUC 0.850). Table 2 summarizes this ablation study. Additional experiments in the Supplement show the impact of $F$ , performance on scripted vs. free-from monologues, and robustness to video distortions.", + "bbox": [ + 212, + 527, + 787, + 772 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.4 Limitations", + "text_level": 1, + "bbox": [ + 214, + 789, + 338, + 803 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our algorithm is less discriminative of subjects that are less emotive and more neutral. In the future, relying on more granular dynamic signatures that can extract micro", + "bbox": [ + 212, + 810, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos", + "bbox": [ + 281, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/afaf7bda5b651afac0aa06a213e2f81807e1dcae11f95fce1ed51335526f1fe0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Input FeaturesLossAUC
3DMMID-Reveal rec. loss [20]0.718
Landmark distancesID-Reveal rec. loss [20]0.754
Landmark distancesOur loss without Rj, ID1, ID2(t)0.850
Landmark distancesOur loss with Rj, ID1, ID2(t)0.868
", + "bbox": [ + 321, + 143, + 679, + 215 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 2: Ablation study showing the importance of our input features and loss function design.", + "bbox": [ + 218, + 222, + 781, + 236 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "expressions can help alleviate this. The performance of our method degrades when expressions that are critical to verifying the driving identity are not captured by the synthetic portrait generator. Lastly, our dataset currently features only one style of interaction: one-on-one conversations. We plan to expand to other conversation styles, such as one-way speeches, in future.", + "bbox": [ + 212, + 252, + 785, + 325 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Societal Impact", + "text_level": 1, + "bbox": [ + 215, + 344, + 380, + 362 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We acknowledge the societal importance of introducing guardrails when it comes to the use of talking-head generation technology. We present this work as a step towards trustworthy use of such technologies. Nevertheless, our work could be misconstrued as having solved the problem and inadvertently accelerate the unhindered adoption of synthetic talking-head technology. We do not advocate for this. Instead, we stress that this is the first work on this topic and underscore the importance of further research.", + "bbox": [ + 212, + 371, + 785, + 459 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 469, + 351, + 484 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Highly photo-real synthetic talking-head generators are becoming increasingly beneficial to applications such as video conferencing and AR/VR-based remote interactions. This trend raises the important new research question of how best to also ensure their safe use in such scenarios. To this end, we investigate the new problem of avatar fingerprinting, to authenticate legitimate talking-heads created by authorized users. We leverage the fact that driving individuals have uniquely identifying dynamic motion signatures, which are also preserved in the synthetic videos that they drive. Since none exists, we contribute a new large-scale dataset carefully designed for avatar fingerprinting and related tasks. We hope that our work lays the foundation for further research.", + "bbox": [ + 212, + 497, + 785, + 627 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. We would like to thank the data-capture participants, and Desiree Luong, Woody Luong, and Josh Holland for their help with Figure 2. We acknowledge David Taubenheim for the voiceover in the demo video, and Abhishek Badki for help with the training infrastructure. We thank Joohwan Kim, Rachel Brown, Anjul Patney, Ben Boudaoud, Josef Spjut, Saori Kaji, Nikki Pope, and Kai Pong for their help with putting together the data capture protocols, informed consent form, photo release form, and agreements for data governance and third-party data sharing. Koki Nagano, Ekta Prashnani, and David Luebke were partially supported by DARPA's Semantic Forensics (SemaFor) contract (HR0011-20-3-0005). This research was funded, in part, by DARPA's Semantic Forensics (SemaFor) contract HR0011-20-3-0005. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government.", + "bbox": [ + 212, + 638, + 787, + 823 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Distribution Statement \"A\" (Approved for Public Release, Distribution Unlimited).", + "bbox": [ + 215, + 824, + 759, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "E. Prashnani et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 310, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Apple's vision pro. https://www.apple.com/apple-vision-pro/, accessed: 2024-03-06", + "2. Heygen. https://www.heygen.com, accessed: 2023-11-16", + "3. Microsoft teams mesh. https://www.microsoft.com/en-us/microsoft-teams/microsoft-mesh, accessed: 2024-03-06", + "4. Myheritage. https://www.myheritage.com, accessed: 2023-11-16", + "5. Nvidia's maxine. https://developer.nvidia.com/maxine, accessed: 2024-03-06", + "6. Agarwal, S., El-Gaaly, T., Farid, H., Lim, S.N.: Detecting deep-fake videos from appearance and behavior. 2020 IEEE International Workshop on Information Forensics and Security (WIFS) pp. 1-6 (2020)", + "7. Agarwal, S., Farid, H.: Detecting deep-fake videos from aural and oral dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2021)", + "8. Agarwal, S., Farid, H., Fried, O., Agrawala, M.: Detecting deep-fake videos from phonemeviseme mismatches. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2020)", + "9. Agarwal, S., Farid, H., Gu, Y., He, M., Nagano, K., Li, H.: Protecting world leaders against deep fakes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2019)", + "0. Agarwal, S., Hu, L., Ng, E., Darrell, T., Li, H., Rohrbach, A.: Watch those words: Video falsification detection using word-conditioned facial motion. In: IEEE Winter Conference on Applications of Computer Vision (WACV) (2023)", + "1. Albright, M., McCloskey, S.: Source generator attribution via inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2019)", + "2. Baluja, S.: Hiding images in plain sight: Deep steganography. In: Advances in Neural Information Processing Systems (NeurIPS) (2017)", + "3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of SIGGRAPH (1999)", + "4. Blanz, V., Vetter, T.: Face recognition based on fitting a 3D morphable model. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2003)", + "5. Boháček, M., Farid, H.: Protecting world leaders against deep fakes using facial, gestural, and vocal mannerisms. Proceedings of the national academy of Sciences (2022)", + "6. Cao, H., Cooper, D.G., Keutmann, M.K., Gur, R.C., Nenkova, A., Verma, R.: CREMA-D: Crowd-sourced emotional multimodal actors dataset. IEEE Transactions on Affective Computing (2014)", + "7. Chung, J.S., Nagrani, A., Zisserman, A.: Voxceleb2: Deep speaker recognition. In: INTER-SPEECH (2018)", + "8. Cozzolino, D., Nießner, M., Verdoliva, L.: Audio-visual person-of-interest deepfake detection (2022)", + "9. Cozzolino, D., Poggi, G., Verdoliva, L.: Extracting camera-based fingerprints for video forensics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019)", + "20. Cozzolino, D., Rössler, A., Thies, J., Nießner, M., Verdoliva, L.: ID-Reveal: Identity-aware DeepFake video detection. In: IEEE International Conference on Computer Vision (ICCV) (2021)", + "21. Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., Canton Ferrer, C.: The deepfake detection challenge dataset. arXiv preprint arXiv:2006.07397 (2020)" + ], + "bbox": [ + 225, + 179, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos", + "bbox": [ + 281, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Dong, X., Bao, J., Chen, D., Zhang, T., Zhang, W., Yu, N., Chen, D., Wen, F., Guo, B.: Protecting celebrities from deepfake with identity consistency transformer. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "23. Drobyshev, N., Chelishev, J., Khakhulin, T., Ivakhnenko, A., Lempitsky, V., Zakharov, E.: MegaPortraits: One-shot megapixel neural head avatars (2022)", + "24. Fox, G., Liu, W., Kim, H., Seidel, H.P., Elgharib, M., Theobalt, C.: VideoForensicsHQ: Detecting high-quality manipulated face videos. In: IEEE International Conference on Multimedia and Expo (2021)", + "25. Fridrich, J.: Steganography in Digital Media: Principles, Algorithms, and Applications. Cambridge University Press (2009)", + "26. Ge, S., Lin, F., Li, C., Zhang, D., Wang, W., Zeng, D.: Deepfake video detection via predictive representation learning. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) (2022)", + "27. Hadsell, R., Chopra, S., LeCun, Y.: Dimensionality reduction by learning an invariant mapping. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2006)", + "28. Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M.: Lips don't lie: A generalisable and robust approach to face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "29. Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M.: Lips don't lie: A generalisable and robust approach to face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "30. He, Y., Gan, B., Chen, S., Zhou, Y., Yin, G., Song, L., Sheng, L., Shao, J., Liu, Z.: Forgerynet: A versatile benchmark for comprehensive forgery analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4360-4369 (2021)", + "31. Hill, H., Johnston, A.: Categorizing sex and identity from the biological motion of faces. Current Biology (2001)", + "32. Honari, S., Molchanov, P., Tyree, S., Vincent, P., Pal, C., Kautz, J.: Improving landmark localization with semi-supervised learning. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "33. Hong, F.T., Zhang, L., Shen, L., Xu, D.: Depth-aware generative adversarial network for talking head video generation (2022)", + "34. Jian Zhao, H.Z.: Thin-plate spline motion model for image animation (2022)", + "35. Jiang, L., Li, R., Wu, W., Qian, C., Loy, C.C.: Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "36. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "37. Khakhulin, T., Sklyarova, V., Lempitsky, V., Zakharov, E.: Realistic one-shot mesh-based head avatars. In: European Conference on Computer Vision (ECCV) (2022)", + "38. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "39. Kim, H., Elgharib, M., Zollhöfer, M., Seidel, H.P., Beeler, T., Richardt, C., Theobalt, C.: Neural style-preserving visual dubbing. ACM Transactions on Graphics (ToG) (2019)", + "40. Kim, H., Garrido, P., Tewari, A., Xu, W., Thies, J., Niessner, M., Pérez, P., Richardt, C., Zollhöfer, M., Theobalt, C.: Deep video portraits. ACM Transactions on Graphics (ToG) (2018)", + "41. Knappmeyer, B., Thornton, I., Bulthoff, H.: Facial motion can determine facial identity. Journal of Vision (2001)", + "42. Kwon, P., You, J., Nam, G., Park, S., Chae, G.: Kodf: A large-scale korean deepfake detection dataset. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10744-10753 (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 838 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "E. Prashnani et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Li, J., Xie, H., Yu, L., Zhang, Y.: Wavelet-enhanced weakly supervised local feature learning for face forgery detection. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)", + "44. Li, Y., Sun, P., Qi, H., Lyu, S.: Celeb-DF: A large-scale challenging dataset for DeepFake forensics. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "45. Liu, B., Liu, B., Ding, M., Zhu, T., Yu, X.: Ti2net: Temporal identity inconsistency network for deepfake detection. In: IEEE Winter Conference on Applications of Computer Vision (WACV) (2023)", + "46. Livingstone, S.R., Russo, F.A.: The ryerson audio-visual database of emotional speech and song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one (2018)", + "47. Luo, X., Zhan, R., Chang, H., Yang, F., Milanfar, P.: Distortion agnostic deep watermarking. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "48. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "49. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit Warping for Animation with Image Sets. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "50. Munir, S., Batool, B., Shafiq, Z., Srinivasan, P., Zaffar, F.: Through the looking glass: Learning to attribute synthetic text generated by language models. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume (2021)", + "51. O'Toole, A.J., Roark, D.A., Abdi, H.: Recognizing moving faces: A psychological and neural synthesis. Trends in Cognitive Sciences (2002)", + "52. Passos, L.A., Jodas, D., da Costa, K.A., Junior, L.A.S., Rodrigues, D., Del Ser, J., Camacho, D., Papa, J.P.: A review of deep learning-based approaches for deepfake content detection. arXiv preprint arXiv:2202.06095 (2022)", + "53. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: FaceForensics++: Learning to detect manipulated facial images. In: IEEE International Conference on Computer Vision (ICCV) (2019)", + "54. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: Faceforensics++: Learning to detect manipulated facial images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1-11 (2019)", + "55. Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: A unified embedding for face recognition and clustering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015)", + "56. Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)", + "57. Sun, K., Yao, T., Chen, S., Ding, S., Li, J., Ji, R.: Dual contrastive learning for general face forgery detection (2022)", + "58. Tancik, M., Mildenhall, B., Ng, R.: Stegastamp: Invisible hyperlinks in physical photographs. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "59. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: European Conference on Computer Vision (ECCV) (2020)", + "60. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "61. Wang, X., Han, X., Huang, W., Dong, D., Scott, M.R.: Multi-similarity loss with general pair weighting for deep metric learning. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "62. Wang, Y., Yang, D., Bremond, F., Dantcheva, A.: Latent image animator: Learning to animate images via latent space navigation. In: International Conference on Learning Representations (ICLR) (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos", + "bbox": [ + 281, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Yacoob, Y.: Gan-scanner: A detector for faces of stylegan+ (2021), https://github.com/yaseryacoob/GAN-Scanner", + "64. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: IEEE International Conference on Computer Vision (ICCV) (2021)", + "65. Yu, N., Skripniuk, V., Chen, D., Davis, L., Fritz, M.: Responsible disclosure of generative models using scalable fingerprinting. In: International Conference on Learning Representations (ICLR) (2022)", + "66. Zakharov, E., Ivakhnenko, A., Shysheya, A., Lempitsky, V.: Fast bi-layer neural synthesis of one-shot realistic head avatars. In: European Conference on Computer Vision (ECCV) (2020)", + "67. Zakharov, E., Shysheya, A., Burkov, E., Lempitsky, V.: Few-shot adversarial learning of realistic neural talking head models. In: IEEE International Conference on Computer Vision (ICCV) (2019)", + "68. Zheng, Y., Bao, J., Chen, D., Zeng, M., Wen, F.: Exploring temporal coherence for more general video face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 372 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "E. Prashnani et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_model.json b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..972648786a4d8684893c2975c4a1d996f69bc46a --- /dev/null +++ b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_model.json @@ -0,0 +1,2689 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.279, + 0.141, + 0.728, + 0.187 + ], + "angle": 0, + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.212, + 0.772, + 0.227 + ], + "angle": 0, + "content": "Ekta Prashnani, Koki Nagano, Shalini De Mello, David Luebke, and Orazio Gallo" + }, + { + "type": "text", + "bbox": [ + 0.472, + 0.238, + 0.531, + 0.251 + ], + "angle": 0, + "content": "NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.253, + 0.765, + 0.266 + ], + "angle": 0, + "content": "{eprashnani, knagano, shalinig, dluebke, ogallo}@nvidia.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.297, + 0.744, + 0.496 + ], + "angle": 0, + "content": "Abstract. Modern avatar generators allow anyone to synthesize photorealistic real-time talking avatars, ushering in a new era of avatar-based human communication, such as with immersive AR/VR interactions or videoconferencing with limited bandwidths. Their safe adoption, however, requires a mechanism to verify if the rendered avatar is trustworthy: does it use the appearance of an individual without their consent? We term this task avatar fingerprinting. To tackle it, we first introduce a large-scale dataset of real and synthetic videos of people interacting on a video call, where the synthetic videos are generated using the facial appearance of one person and the expressions of another. We verify the identity driving the expressions in a synthetic video, by learning motion signatures that are independent of the facial appearance shown. Our solution, the first in this space, achieves an average AUC of 0.85. Critical to its practical use, it also generalizes to new generators never seen in training (average AUC of 0.83). The proposed dataset and other resources can be found at: https://research.nvidia.com/labs/nxp/avatar-fingerprinting/." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.507, + 0.741, + 0.521 + ], + "angle": 0, + "content": "Keywords: Synthetic Media Verification \\(\\cdot\\) Avatar Generators \\(\\cdot\\) Avatar Attribution" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.548, + 0.357, + 0.564 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.789, + 0.753 + ], + "angle": 0, + "content": "Recent digital avatar generators have fueled a myriad of computer vision and graphics applications, allowing anyone to synthesize real-time photorealistic personas. Major companies are now supporting avatar-driven remote interactions over immersive AR/VR (e.g. Meta's Pixel Codec Avatar [48], Apple's Vision Pro Persona [1]) or video conferencing (e.g. NVIDIA's MAXINE [5], Microsoft's Teams [3]), and selfie filters for altering and enhancing appearance (e.g. by Snap and Tiktok). While the avatar generation technology today is still young, the legitimate use of synthetic avatars will be ubiquitous in the future. Without proper guardrails, this poses a real risk of unauthorized use and large-scale spread of visual disinformation. To ensure the safe use of such a technology, the relevant question is no longer whether the content is \"real\" or not—since, by design, the videos and avatars are all synthetic—but rather, whether the synthetically-generated videos and avatars are \"trustworthy\" or not." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.754, + 0.79, + 0.841 + ], + "angle": 0, + "content": "When video conferencing, for instance, a synthetic video portrait generator can be used to save valuable bandwidth by reconstructing a synthetic avatar of the sender at the receiver's end, using only a frame of the target identity and a compact representation of the speaker's facial motion. To ensure the authorized use of such synthetically-generated videos, we want to verify if the driving identity behind a synthetic video \\((\\mathrm{ID}_1\\) or \\(\\mathrm{ID}_2\\) in Figure 1) is authorized to control the likeness, or the appearance, of the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.381, + 0.127 + ], + "angle": 0, + "content": "E. Prashnani et al." + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.144, + 0.709, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Fig. 1: Talking-head avatar generators can synthesize realistic videos of a target identity from driving videos of different identities. Our method extracts appearance-agnostic temporal facial features and learns an embedding in which the synthetic videos driven by one identity fall close to each other and far from those driven by other identities, regardless of the appearance of the synthetic video. By comparing distances in the embedding space, we evaluate whether an avatar is driven by an authorized identity or not. During evaluation, we only rely on the synthetic videos as input, without requiring any prior knowledge about the driving identity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.787, + 0.463 + ], + "angle": 0, + "content": "synthetic video (target identity in Figure 1). Crucially, we want to only leverage the synthetic video avatars to do so. We call this novel task avatar fingerprinting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.565 + ], + "angle": 0, + "content": "We leverage a simple but fundamental observation: facial motions tend to be idiosyncratic, that is, individuals talk and emote in unique ways. For instance, someone may raise one of her eyebrows more than the other, or smile more while talking. These \"dynamic identity signatures\" [51] have been shown to carry enough information for humans to recognize other individuals, even when the physical appearance of their face is altered [31, 41, 51]. This makes them attractive for our task, as they can be derived solely from the driving identity of a talking-head video, regardless of the appearance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Fortunately, modern avatar generators are becoming increasingly accurate at capturing the facial motion of a person and rendering it onto a target identity. As a first solution to avatar fingerprinting, we then propose to estimate a dynamic identity signature for an identity from the synthesized videos they drove—regardless of the target identities shown. Specifically, we extract facial landmarks and their temporal dynamics from the video. We then introduce a novel contrastive loss to learn a dynamic identity embedding, a space where the dynamic identity signatures of a driving identity across multiple videos and target identities are close to each other, and far from those of other driving identities (see Figure 1). We show that this approach, albeit straightforward, is robust and generalizes to generators not seen in training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Avatar fingerprinting is a new task, and no existing datasets serve its training and validation requirements. There are two key requirements from a dataset. First, we need videos of multiple subjects delivering both scripted and free-form monologues, captured under realistic conditions, such as varying video quality and gaze direction. This allows us to assess if a model leverages talking styles, and not the specific choice of words. Second, to evaluate if the model learns to extract dynamic identity signatures effectively from synthetic videos, we need synthetic talking-head videos for the case in which the driving and target identities are different (cross-reenactment), and that in which they match (self-reenactment). Unfortunately, existing datasets of real videos" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.788, + 0.292 + ], + "angle": 0, + "content": "only fulfill a subset of these requirements (Table 1 and Section 4). Further, no existing dataset of synthetic videos contain both self- and cross-reenactments per subject or use the state-of-the-art talking-head generation technology. To foster research in this new domain, then, we introduce the NVIDIA Facial Reenactment (NVFAIR) dataset, containing real and synthetic face reenactment talking-head videos (Figure 2). Our dataset, which includes ethnically diverse subjects, provides \\(10 \\times\\) more synthetic facial reenactments than the second largest dataset, for a total of over 650,000 synthetic videos. It is also the only one using multiple state-of-the-art face reenactment generators [34,60,62], or to provide cross-reenactments driven by all identities, which is critical for training and evaluating avatar fingerprinting algorithms." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.3, + 0.304, + 0.314 + ], + "angle": 0, + "content": "In summary:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.322, + 0.784, + 0.363 + ], + "angle": 0, + "content": "- We introduce the novel task of avatar fingerprinting, which focuses on verifying the driving identity of synthetic talking-head avatars, rather than classifying them as real or synthetic (by design, all inputs to our model are synthetic)." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.365, + 0.784, + 0.405 + ], + "angle": 0, + "content": "- We release the first large-scale dataset of subjects delivering scripted and natural monologues, complete with self- and cross-reenactment videos synthesized with multiple state-of-the-art generators." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.408, + 0.784, + 0.449 + ], + "angle": 0, + "content": "- We propose a solution for this novel task in the context of video conferencing by extracting person-specific motion signatures, and demonstrate its robustness to various distortions and avatar generators not seen in training." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.322, + 0.784, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.471, + 0.367, + 0.487 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.501, + 0.79, + 0.601 + ], + "angle": 0, + "content": "Our proposed avatar fingerprinting task aims to verify authorized use of synthetic avatars: this is fundamentally a different problem than traditional forensics research where one aims to detect synthetic media (e.g. deepfake detection) or actively mark synthetic content. In our case, by design, the content being evaluated is always synthetic: we aim to evaluate its authorized use. Since this is a novel task, no methods exist to directly address it. No methods currently exist to directly address this novel task. Here we discuss the related areas of research." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.606, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Learning-based Attribution of Synthetic Media. Learning-based approaches have been used to identify the origin of synthetic media, or to determine if it has been manipulated or altered in some way. Previous work used a pre-trained GAN generator to attribute a synthesized image to its generator via GAN inversion by leveraging the fact that a real image is less invertible [11, 36, 63]. Yet other works focus on attributing other forms of synthetic media, such as text [50]. In contrast, our focus is to attribute a talking-head avatar to the identity driving it, regardless of the appearance of the avatars. Some existing works learn fingerprints associated with cameras to determine whether a video is manipulated [19], or embed watermarks into images and videos [12, 25, 47, 58], which are also shown to transfer to GAN-generated images [64]. Subsequent research introduced a watermark-based conditional GANs for scalable fingerprinting [65]. Our method, in contrast, is a passive technique that does not rely on active watermarks." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.782, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Deepfake Detection Based on Identity-Specific Features. Deepfake detection (\"is a video synthetic?\") and avatar fingerprinting (\"whose identity is used to generate this synthetic video?\") are fundamentally different tasks. Most existing solutions for deepfake detection train a real-vs-synthetic classifier [26, 28, 43, 52, 57, 68], and therefore" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.229, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.381, + 0.127 + ], + "angle": 0, + "content": "E. Prashnani et al." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.42 + ], + "angle": 0, + "content": "cannot be adapted to avatar fingerprinting (where all inputs are synthetic). However, a specific class of detectors leverage identity-specific features to detect synthetic videos by posing the detection problem as an identity-recognition problem. In our experiments, we evaluate some of these methods as baselines. Specifically, Agarwal et al. exploit person-specific patterns in facial expressions to detect fake videos [9]. ID-Reveal used facial shapes and motions encoded in a low-dimensional space of a 3D morphable model [13] to handle both face-swapping and face-reenactment deepfakes [20]. Other works explored soft-biometric approaches such as leveraging vocal mannerisms [15], phoneme-viseme consistencies [8], word-facial expression consistencies [6,10], and dynamics of ears [7]. While many of these works need person-specific training, previous works [6, 18, 20, 45] extended this idea to train a CNN-based detector using a large-scale in-the-wild video data [17] and variants of contrastive learning [27, 38, 55, 61]. Agarwal et al. combined static facial appearance using a facial recognition model and dynamic facial behaviors using a CNN, and showed that this approach is effective for detecting face-swap deepfakes [6]. Yet another line of research has explored temporal inconsistencies of face identities within a video [45], and identity inconsistencies of inner and outer face regions [22]. Our experiments show that for avatar fingerprinting, such features that are designed to distinguish real from synthetic videos are not reliable, since we only have access to synthetic videos." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.422, + 0.786, + 0.766 + ], + "angle": 0, + "content": "Dynamic Facial Identity Signatures. Cognitive scientists have studied the impact of \"dynamic facial identity signatures\" (i.e., characteristic or identity-specific movements of a face) for identity recognition for humans [51]. In one experiment, scientists projected facial animations generated by human actors onto an average head and found that subjects discriminated between individuals based solely on facial motion [31]. In another, subjects correctly attributed animations of synthetic faces to their morphed versions [41]. While these studies point to the existence of \"dynamic facial identity signatures\" that humans rely on, ours is the first method that isolates these from videos. Talking-Head Datasets and Generators. Existing talking-head datasets include those that contain only real videos showing a variety of emotions [16, 46, 59], and others that also contain synthetic videos [21, 24, 30, 35, 42, 44, 53]. These datasets cater to traditional forensics and facial analysis tasks. Therefore, they do not contain self-and cross-reenactments driven by multiple identities, as well as scripted and free-form monologues across diverse capture settings. There novel requirements posed by the avatar fingerprinting task motivate us to design our own dataset. We focus specifically on face-reenactment talking-head avatar generators for synthetic video generation—this class of generators are the most relevant to AR/VR interactions, video conferencing, and several other applications [2-5, 48]—and combine various modes of human expression for capturing real videos. Given a target facial image and a driving video, these generators reenact the target image using the facial expressions and head pose from the driving video [23, 33, 34, 37, 49, 56, 60, 62, 66, 67]. Another class of talking-head generators use person-specific models [40] and some models aim to preserve the style of the target identity in the synthesized video [39]. However, these models require person-specific training, making them difficult to scale." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.785, + 0.355, + 0.803 + ], + "angle": 0, + "content": "3 Terminology" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "We seek to verify the trustworthiness of a synthesized talking-head video, termed target video. We assume that an avatar-generation tool (e.g., [60]) created it by animating" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.142, + 0.788, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 2: We introduce the NVFAIR dataset, containing real and synthetic talking-head videos. We capture subjects talking in both scripted and free-form settings. To encourage natural performance, we record the subjects while videoconferencing with each other (left). We then synthesize more than 650,000 talking-head videos—the largest collection till date—using three state-of-the-art face-reenactment talking-head generators. On the right, each row corresponds to a driving identity \\((\\mathrm{ID}_i \\rightarrow (\\cdot))\\) and each column corresponds to a different target identity \\(((\\cdot) \\rightarrow \\mathrm{ID}_i)\\). The videos in which driving and target identity match are self-reenactments, the rest are cross-reenactments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.789, + 0.501 + ], + "angle": 0, + "content": "an image (target image) using the expressions and head poses obtained from another video, the driving video. We call driving identity the identity of the person in the driving video, and target identity the identity of the person in the target image. When driving and target identities match, the target video is a self-reenactment, while the case of a driving identity used to animate a different target identity is cross-reenactment. In both cases, the appearance of the synthesized video is derived from the target identity. This terminology allows us to formally state our goal: we want to verify that a target video is a self-reenactment. With this terminology in mind, we introduce our dataset, which includes real videos as well as self- and cross-reenactment videos." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.517, + 0.691, + 0.534 + ], + "angle": 0, + "content": "4 The NVIDIA Facial Reenactment (NVFAIR) Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.597 + ], + "angle": 0, + "content": "Recall that avatar fingerprinting is not about detection of synthetic media. Rather, we already know a video to be synthesized, and seek to verify that the driving identity is authorized. This new task dictates a set of requirements for the dataset to be effective for training and evaluation. Specifically, we need a dataset that contains" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.599, + 0.785, + 0.626 + ], + "angle": 0, + "content": "1. multiple real videos per identity, with scripted and free-form conversations, and with both natural and prescribed emotions," + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.627, + 0.785, + 0.655 + ], + "angle": 0, + "content": "2. self- and cross-reenactments of target identities, with cross-reenactments driven by all subjects to allow for a variety of driving facial structures, and" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.656, + 0.495, + 0.669 + ], + "angle": 0, + "content": "3. multiple face-reenactment generators." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.599, + 0.785, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.819, + 0.771 + ], + "angle": 0, + "content": "All relevant existing datasets only capture a subset of these requirements (see Table 1; the Supplement contains further discussion). Moreover they do not use the state-of-the-art talking-head generation technology to synthesize the self- and cross-reenactment videos. We introduce the NVFAIR dataset that features all of the above properties. With over 650,000 synthetic videos, it provides \\(10 \\times\\) as many videos as the next largest dataset, and uses three different state-of-the-art generators for the self- and cross-reenactment videos. Figure 2 shows an overview of data capture and synthesis." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.79, + 0.389, + 0.805 + ], + "angle": 0, + "content": "4.1 Real Data Capture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Capturing videos of monologues delivered by different subjects for the purpose of identity verification introduces two conflicting goals. On the one hand a controlled evalua-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.381, + 0.127 + ], + "angle": 0, + "content": "E. Prashnani et al." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.272 + ], + "angle": 0, + "content": "
Dataset# Subjects (source)(F)ree / (S)scripted?Emotion: (N)atural / (P)rescribed(R)eel / (S)elf / (C)ross-reenact.?#Face Reenact.Avg. Videos per Subject# Face- reenact. GeneratorsEthnic Diversity
RAVDESS [46]24 (new)(S)(P)(R) onlyN/A120 (R)N/A✓
MEAD [59]60 (new)(S)(P)(R) onlyN/A720 (R)N/A✓
CREMA-D [16]91 (new)(S)(P)(R) onlyN/A81 (R)N/A✓
VFHQ [24]36 ([46])(S)(P)+(N)(R)+(S)1,7371201✓
FF++ [54]1000 (YT)(F)(N)(R)+(C)2,0001(R)+2(C)2✓
KoDF [42]403 (new)(F)+(S)(N)(R)+(C)61,000150(R)+151(C)1X
NVFAIR (Ours)161 (46 new + [46], [16])(F)+(S)(P)+(N)(R)+(S)+(C)654,72676(R)+228(S) +3,840(C)3✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.278, + 0.787, + 0.385 + ], + "angle": 0, + "content": "Table 1: The existing talking-head video datasets were designed for tasks such as deepfake detection or facial emotion analysis. Avatar fingerprinting is a fundamentally different task. As a result, no existing dataset satisfies the requirements for training and evaluating models for avatar fingerprinting. To overcome this limitation, we introduce the NVFAIR dataset, which is the first dataset that offers the complete set of monologue modalities, and features the largest collection of facial reenactments to date. Specifically, it provides scripted and free-form monologues, with natural and prescribed emotions, and self- and cross-reenactments (driven by all remaining subjects) synthesized using three generators, alongside original videos for newly-recorded subjects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.787, + 0.621 + ], + "angle": 0, + "content": "tion of the trained models requires predictability of what is spoken to prevent identification algorithms from latching onto the spoken content itself. On the other, we want the subjects to act as they would in a casual conversation, rather than reciting memorized text, to capture their uniquely identifying mannerisms. We address this trade-off by recording the subjects while videoconferencing in pairs, which creates the impression of being in a natural conversation. This differs from existing datasets, in which the subjects look at the camera, but are not interacting with another person during the recording [16,42,46]. We also design two distinct recording strategies: a free-form stage where the subjects are given only general guidance on the topics, and a more controlled scripted stage in which subjects speak short, memorized monologues of 2-3 sentences each, see Figure 2(a). To capture the variability of real-life scenarios, we provided minimal instructions on how to setup the video call, allowing for diverse face, scale, and lightning, bandwidth stability, and background scene clutter. In total we record 46 subjects of diverse genders, ages, and ethnicities, while strictly abiding by a pre-approved IRB protocol (see Supplement for details and privacy considerations)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.787, + 0.74 + ], + "angle": 0, + "content": "Stage I: Free-Form Monologues. In this first stage, the two subjects on the call alternate between asking and answering seven pre-defined questions. The questions are designed to avoid sensitive or potentially inflammatory topics. This is critical because we later use sentences spoken by one individual to animate the video of a second individual, quite literally putting words in their mouths. The complete list of questions is in the Supplement. To further create a natural interaction, the subject listening is encouraged to actively but silently engage with the one speaking (e.g., by nodding or smiling)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.754, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Stage II: Scripted Monologues. For this stage, we prepared thirty short utterances consisting of two or three sentences each. We chose this length to allow for memorization, while still providing enough content to trigger facial expressions. However, to avoid inducing unnatural expressions, we do not prescribe specific emotions for each utterance. For instance, we do not ask to express anger for a sentence, but we do choose sentences that may naturally evoke it, and used punctuation to encourage it, e.g., \"Will you please" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.205 + ], + "angle": 0, + "content": "answer the darn phone? The constant ringing is driving me insane!\" We instruct the subjects to split their screens to show both this list and video call and encourage them to speak to their recording partner when reciting, see Stage II in Figure 2(a). More details, including the full list of utterances can be found in the Supplement." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.227, + 0.471, + 0.243 + ], + "angle": 0, + "content": "4.2 Synthetic Talking-Head Videos" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.252, + 0.788, + 0.426 + ], + "angle": 0, + "content": "Using the videos described in Section 4.1, as well as the original videos from the CREMA-D [16] and RAVDESS [46] datasets, we generate synthetic talking-head videos to train and evaluate our avatar fingerprinting algorithm. Specifically, we pool the 91 identities from the original videos of CREMA-D [16], the 24 identities from those of RAVDESS [46], and the 46 from our own video-conferencing data capture, for a total of 161 unique identities \\(\\mathcal{I}\\). Recall that we have several real videos for each identity \\(\\mathrm{ID}_i\\in \\mathcal{I}\\). To avoid a combinatorial explosion of synthetic videos, for all pairs of identities \\(\\mathrm{ID}_i\\) and \\(\\mathrm{ID}_j\\), we use \\(\\mathrm{ID}_j\\) as the target identity and we randomly select 8 of the videos of \\(\\mathrm{ID}_i\\) to generate 8 cross-reenactment videos, \\(\\{\\mathrm{ID}_i^k\\to \\mathrm{ID}_j\\}_{k = \\{1,\\dots ,8\\}}\\) (all 8 share the same target image). We also generate self-reenactment videos for each of the target identities, by animating their neutral-face images derived from captured videos with each of their real videos." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.427, + 0.788, + 0.555 + ], + "angle": 0, + "content": "We use three different generators for synthesizing the videos for all the 161 identities: face-vid2vid [60], LIA [62], and TPS [34]. This allows us to test if our model generalizes across generators. We chose these talking-head generators because they are the state of the art and they preserve the identity-specific facial motion dynamics well. Nevertheless, the reconstruction is not perfect; for instance, in the third row of Figure 2(b) the person in the driving video \\((\\mathrm{ID}_5)\\) is squinting, but the eyes are shut in all the synthetic videos, including the self-reenactment video. In total we generate more than 650,000 synthetic videos, which required more than 2,500 RTX 3090 GPU hours. More details are in the Supplement." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.578, + 0.315, + 0.594 + ], + "angle": 0, + "content": "5 Method" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.61, + 0.788, + 0.754 + ], + "angle": 0, + "content": "Overview. We seek to verify the driving identity of a synthetic video, independently of the target identity. We leverage the finding from cognitive science research that each person emotes in unique ways when communicating, and that this signal is sufficient for recognition, even when the actual appearance is artificially corrupted [31, 41, 51]. We note that our method does not latch onto artifacts introduced by the generators—a property that we demonstrate by showing generalization to new generators not seen during training. Rather, our features capture the dynamics of the expressions, like the way a person frowns, or the way she smiles. Notably, they are distinct from the temporal artifacts introduced by the generator, and that existing algorithms use to detect whether a video is synthetic or real [29]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.755, + 0.788, + 0.84 + ], + "angle": 0, + "content": "An overview of our algorithm for avatar fingerprinting is shown in Figure 3. To capture expressions, we extract the relative position of facial landmarks over time from the input video, as shown in Figure 3 (Section 5.1). We learn to project these temporal signatures onto a dynamic identity embedding in which features belonging to the same driving identity are close to each other regardless of the target identity, i.e., independently of appearance (Figure 1). To learn this embedding we train a 3DCNN (where the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.229, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "E. Prashnani et al." + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.144, + 0.709, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.788, + 0.37 + ], + "angle": 0, + "content": "Fig. 3: We extract landmarks from the frames of a talking-head clip, compute their normalized pairwise distances, and concatenate the frame-wise features. We then learn an identity embedding using a loss that pulls closer features of videos driven by the same identity and pushes away those driven by others. \\(\\mathrm{ID}_i\\rightarrow \\mathrm{ID}_j\\) indicates a video that looks like identity \\(j\\) (the \"target\" identity), and is driven by identity \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.439 + ], + "angle": 0, + "content": "third dimension spans video frames—a temporal CNN) with a novel contrastive loss that pulls together all embedding vectors of synthetic videos driven by an individual, while pushing away the embedding vectors of videos driven by all others individuals (Section 5.2). More implementation details are in Section 5.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.458, + 0.49, + 0.473 + ], + "angle": 0, + "content": "5.1 Dynamic Facial Identity Features" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.524 + ], + "angle": 0, + "content": "Our first step is to extract temporal features that summarize short segments of the video we wish to fingerprint. We identify the following guiding principles for the extracted features. We would like features that:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.524, + 0.784, + 0.551 + ], + "angle": 0, + "content": "1. have minimal dependency on the appearance of the face in the video (that is, the target identity)," + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.553, + 0.533, + 0.566 + ], + "angle": 0, + "content": "2. reflect the dynamics of the expressions, and" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.567, + 0.422, + 0.581 + ], + "angle": 0, + "content": "3. capture subtle expressions." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.524, + 0.784, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.667 + ], + "angle": 0, + "content": "One choice could be per-frame 3DMM features [14]: a strategy also used by Cozzolino et al. to detect synthetic videos [20]. However, we empirically observe that 3DMM features are not sufficiently expressive, and do not satisfy desideratum 3 (see ablation experiments in Section 6). We observe a similar behavior for action units [9]. Facial landmarks [32] address this issue, but are sensitive to the shape of the face in the video, and thus to the target identity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.787, + 0.725 + ], + "angle": 0, + "content": "To leverage the expressiveness of facial landmarks while abstracting from the underlying facial shape, we compute the pairwise normalized Euclidean distances between each pair of landmarks of a frame. We concatenate these distances into a single vector for the frame, \\(\\mathbf{d}_f\\). A subset of the facial landmarks and distances are shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.787, + 0.81 + ], + "angle": 0, + "content": "We then break the input video into clips, which are sequences consisting of F frames and offset by one frame (e.g., [1,F], [2,F+1], etc.), and concatenate the vectors from all the frames in each clip. Using the change in the relative position of the landmarks over a short period of time (the length of a clip) allows us to capture temporal dynamics with minimal dependence on the absolute position of each landmark, i.e., independently of the shape of the face." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We show empirically that our features are a good representation for our task, by comparing against alternative choices for input features such as 3DMM (Section 6)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.586, + 0.162 + ], + "angle": 0, + "content": "5.2 Dynamic Identity Embedding Contrastive Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.784, + 0.241 + ], + "angle": 0, + "content": "While the features described in Section 5.1 are designed to extract low-level motion dynamics, they cannot be used directly to disambiguate two target videos based on their driving identities. We tackle this problem by learning a dynamic identity embedding, a space where videos driven by one subject map to points that are close to each other and far from the videos driven by anybody else." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.785, + 0.327 + ], + "angle": 0, + "content": "Specifically, we use a temporal 3DCNN to extract an embedding vector from a clip, which, as described before, is a short segment of an input video. To train the network we use a dataset of synthetic videos driven by different identities. We denote as \\(\\mathcal{C}_{\\mathrm{ID}_1\\to \\mathrm{ID}_2}^k (t)\\) the embedding produced by the network for the clip starting at time \\(t\\) in the \\(k\\)-th video, of a target identity \\(\\mathrm{ID}_2\\) driven by identity \\(\\mathrm{ID}_1\\). As stated above, we have two main objectives, which we capture with the following terms in our proposed loss function." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.34, + 0.784, + 0.369 + ], + "angle": 0, + "content": "We Want to Pull Together All the Videos Driven by \\( ID_{1} \\). To achieve this, we define the following term:" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.377, + 0.786, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\mathrm {N} _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\sum_ {\\mathrm {I D} _ {l}, k} \\max _ {n} s \\left(\\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {2}} ^ {j} (t), \\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {l}} ^ {k} (n)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.42, + 0.788, + 0.568 + ], + "angle": 0, + "content": "where \\( s(\\cdot, \\cdot) = e^{-\\|\\cdot\\|^{2}} \\) is a similarity metric. Intuitively, Equation 1 takes two videos, \\( j \\) and \\( k \\), both driven by \\( \\mathrm{ID}_1 \\). Given a clip starting at time \\( t \\) in the first video, it looks for the most similar clip in the second video. Since the driving identity is the same for both videos, Equation 1 encourages an embedding where clips that capture a similar expression are closer to each other. Equation 1 is high even if only one clip from video \\( k \\) has a similar temporal signature to \\( \\mathcal{C}_{\\mathrm{ID}_1 \\to \\mathrm{ID}_2}^j(t) \\). That is because even just one occurrence of the same expression is evidence that the driving identity may be the same. Of course, other driving identities may use similar expressions and we address that with the loss term described below. Additionally, we note that \\( k \\) spans the set of all videos driven by \\( \\mathrm{ID}_1 \\), and \\( \\mathrm{ID}_l \\) spans all identities, including \\( \\mathrm{ID}_l = \\mathrm{ID}_1 \\) and \\( \\mathrm{ID}_l = \\mathrm{ID}_2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.579, + 0.747, + 0.594 + ], + "angle": 0, + "content": "We Want to Push Away Videos not Driven by \\( ID_{1} \\). We define the following term:" + }, + { + "type": "equation", + "bbox": [ + 0.305, + 0.602, + 0.787, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\mathrm {Q} _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\sum_ {\\mathrm {I D} _ {l} \\neq \\mathrm {I D} _ {1}, k} \\max _ {n} s \\left(\\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {2}} ^ {j} (t), \\mathcal {C} _ {\\mathrm {I D} _ {l} \\rightarrow \\mathrm {I D} _ {2}} ^ {k} (n)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.643, + 0.784, + 0.716 + ], + "angle": 0, + "content": "where, similarly to Equation 1, we take a clip from video \\(j\\), and look for the most similar clip in video \\(k\\). This time the two videos share the same target identity, but are driven by different identities: we want all the videos driven by identities different from \\(\\mathrm{ID}_1\\) to be pushed away from those driven by \\(\\mathrm{ID}_1\\), including videos where \\(\\mathrm{ID}_1\\) is the target identity. Note that \\(\\mathrm{ID}_2\\) spans all identities, including \\(\\mathrm{ID}_2 = \\mathrm{ID}_1\\) and \\(\\mathrm{ID}_2 = \\mathrm{ID}_l\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.727, + 0.784, + 0.8 + ], + "angle": 0, + "content": "We Want to Rely on the Temporal Dynamics of the Videos Driven by \\( ID_{1} \\). Although we use a temporal CNN, the model could still learn to rely on static expressions, such as a snapshot of the person smiling, rather than the temporal progression of expression leading to, or following, the smile. To further encourage the model to learn from the temporal dynamics, we introduce an additional term:" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.807, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R} _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\sum_ {\\mathrm {I D} _ {l}, k} \\max _ {n} s \\left(\\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {2}} ^ {j} (t), \\tilde {\\mathcal {C}} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {l}} ^ {k} (n)\\right), \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "E. Prashnani et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.254 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathcal{C}}_{\\mathrm{ID}_1\\to \\mathrm{ID}_l}^k\\) denotes a version of the clip \\(\\mathcal{C}_{\\mathrm{ID}_1\\to \\mathrm{ID}_l}^k\\) from Equation 1 with randomly shuffled frame ordering. We want such time-shuffled versions of the clips driven by \\(\\mathrm{ID}_1\\) to be pushed away from the pristine clips driven of \\(\\mathrm{ID}_1\\). Effectively, this means that the driving identity of the time-shuffled clips is regarded as different from \\(\\mathrm{ID}_1\\). In other words, we want to pull together video clips in the learned embedding space only when the temporal facial dynamics are characteristic of \\(\\mathrm{ID}_1\\). We further show the importance of this term in Section 6.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.302 + ], + "angle": 0, + "content": "Combining Equations 1, 2, and 3, we write the probability that the embedding vector \\(\\mathcal{C}_{\\mathrm{ID}_1\\to \\mathrm{ID}_2}^j (t)\\) lies close to the embedding vectors for all video clips driven by \\(\\mathrm{ID}_1\\) and far from all the videos driven by others as" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.314, + 0.787, + 0.35 + ], + "angle": 0, + "content": "\\[\np _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\frac {\\mathrm {N} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t)}{\\mathrm {N} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t) + \\mathrm {Q} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t) + \\mathrm {R} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.361, + 0.416, + 0.376 + ], + "angle": 0, + "content": "and the complete loss term as" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.39, + 0.787, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}, t} - \\log \\left(p _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t)\\right). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.451, + 0.371, + 0.467 + ], + "angle": 0, + "content": "5.3 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.789, + 0.699 + ], + "angle": 0, + "content": "Parameter Choices. To extract the per-frame dynamic facial identity features \\(\\mathbf{d}_f\\), we detect 126 facial landmarks for each frame [32], and compute the per-frame normalized pairwise Euclidean distances between these landmarks. The clip duration is set to 71 frames. We find that this is sufficient to capture the facial dynamics that are meaningful for avatar fingerprinting, while also maintaining a good trade-off between speed and accuracy. We also experiment with shorter-duration video clips (see Supplement). The input tensor to the temporal CNN is obtained by concatenating \\(\\mathbf{d}_f\\) across 71 frames. In each batch, we include 8 unique identities. For each identity \\(\\mathrm{ID}_i\\), the pull term (Equation 1) comprises 16 clips: 8 are self-reenactments, randomly sampled from the full set, and the remaining are cross-reenactments with \\(\\mathrm{ID}_i\\) as the driving identity. This allows the neural network to pull together videos based purely on the facial motion, regardless of the appearance of the video. The push term (Equation 2) for \\(\\mathrm{ID}_i\\) consists of clips with the remaining 7 identities in the batch serving as driving identities (8 clips per driving identity), as well as the time-shuffled self-reenactments of \\(\\mathrm{ID}_i\\) (Equation 3). Additional training details can be found in the Supplement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Training, Validation, and Testing Datasets. Of the 161 total identities (pooling together the identities from our dataset, RAVDESS, and CREMA-D, see Section 4.2), we reserve 35 for testing, 14 for validation, and 112 for training. We ensure that there are no cross-set cross-reenactments: that is, identities in the training set only drive other training-set identities (and similarly for the validation and test sets). This allows us to test the models on facial appearances and expressions that were not seen during training. To evaluate the generalization of trained models to new generators, we train our network on videos generated for the training set identities using one generator, and evaluate on the synthetic videos of test-set identities synthesized using remaining two generators." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.143, + 0.387, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.242, + 0.366, + 0.251 + ], + "angle": 0, + "content": "\\(\\mathrm{ID}_6 \\rightarrow \\mathrm{ID}_6\\) d=0.84" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.143, + 0.501, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.243, + 0.48, + 0.251 + ], + "angle": 0, + "content": "\\(\\mathrm{ID}_6 \\rightarrow \\mathrm{ID}_1\\) d=0.95" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.143, + 0.616, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.243, + 0.594, + 0.251 + ], + "angle": 0, + "content": "\\(\\mathrm{ID}_1 \\rightarrow \\mathrm{ID}_6\\) d=1.58" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.155, + 0.73, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.243, + 0.708, + 0.251 + ], + "angle": 0, + "content": "\\(\\mathrm{ID}_3 \\rightarrow \\mathrm{ID}_6\\) d=2.11" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.254, + 0.386, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.354, + 0.376, + 0.363 + ], + "angle": 0, + "content": "\\(\\mathrm{ID}_2 \\rightarrow \\mathrm{ID}_2\\) d=0.589" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.254, + 0.5, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.354, + 0.489, + 0.363 + ], + "angle": 0, + "content": "\\(\\mathrm{ID}_2 \\rightarrow \\mathrm{ID}_7\\) d=0.961" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.266, + 0.616, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.354, + 0.605, + 0.363 + ], + "angle": 0, + "content": "\\(\\mathrm{ID}_7 \\rightarrow \\mathrm{ID}_2\\) d=2.042" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.266, + 0.73, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.354, + 0.72, + 0.363 + ], + "angle": 0, + "content": "\\(\\mathrm{ID}_8 \\rightarrow \\mathrm{ID}_2\\) d=1.787" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Fig. 4: Animated figure. Open in a media-enabled viewer like Adobe Reader and click on the inset. Our embeddings capture the dynamics of an expression, rather than the appearance of the face. For each row, we pick a reference identity. The green box indicates reenactments driven by the reference identity, the red and blue are cross-reenactments of the reference identity. We compute the average distance of each clip shown here against all other clips driven by the reference identity. The average distance to the other clips of the reference identity is consistent for a given motion, and lower (better) when the reference identity is driving as compared to the cross-reenactments that use the reference identity as target. Here, we show videos generated by face-vid2vid [60], and use the embedding vectors predicted by the model trained on the same generator (see Figure 6 generalization to new generators not seen during training)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.342, + 0.538 + ], + "angle": 0, + "content": "6 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.652 + ], + "angle": 0, + "content": "We thoroughly evaluate our algorithm both qualitatively and quantitatively. Our algorithm outperforms reasonable baselines (Section 6.1), it generalizes to generators not seen in training (Section 6.2), and it is robust to video compression (Supplement). We also perform a number of ablation studies to analyze our design choices: our input features, the importance of the time-shuffling term in the loss function (Section 6.3), and the impact of clip duration (Supplement)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We begin by evaluating qualitatively our method's ability to extract embedding vectors based on the driving identity. Figure 4 shows a set of self- and cross-reenacted clips (please view the animation in a media-enabled viewer, such as Adobe Acrobat). For each row, we take one identity as reference and we compute the embedding vectors of clips that use it both as the driving and the target identity. We then compute the average Euclidean distance of the resulting embedding vectors against those of the self-reenactments by the same reference identity. We note that the average distance \\( d \\) is lower when the driving identity matches the reference identity (first two columns). We also note that the distances between the clips in the first two columns are similar: this confirms that the distance is a function of the facial motion, rather than the facial appearance. When the driving identity changes, the average distance increases, even if the target identity matches the reference identity, which is precisely our goal. More results are in the Supplement." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "E. Prashnani et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.144, + 0.75, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Fig.5: ROC curves and AUC values for our method and two baselines: Agarwal et al. [9] and ID-Reveal [20]. Each sub-plot shows the results on our test set for each of the three talking-head generators: face-vid2vid [60], LIA [62], and TPS [34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.784, + 0.405 + ], + "angle": 0, + "content": "To evaluate our approach more formally, we use the 35 unique test-set identities that are not used as driving or target identities in the training set (Section 5.3). One at a time, we treat each identity \\(\\mathrm{ID}_i\\) as target and synthesize cross-reenactments using all the remaining identities as drivers. This is the set of \"unauthorized\" synthetic videos for \\(\\mathrm{ID}_i\\). The self-reenacted samples for \\(\\mathrm{ID}_i\\) form the \"authorized\" set. Note that there are several self-reenacted videos of \\(\\mathrm{ID}_i\\), one per original video of \\(\\mathrm{ID}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.785, + 0.449 + ], + "angle": 0, + "content": "For each target identity \\(\\mathrm{ID}_i\\), we extract the dynamic identity embedding vector of all the clips in the pool of its self- and cross-reenacted videos, and compute their Euclidean distances. That is, for clip \\(k\\) we compute" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.459, + 0.61, + 0.478 + ], + "angle": 0, + "content": "\\[\nd \\left(\\mathcal {C} _ {\\mathrm {I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}} ^ {k}, \\mathcal {C} _ {\\mathrm {I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}} ^ {l}\\right), \\forall l \\neq k, \\text {a n d}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.472, + 0.785, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l}u \\left(\\mathrm {I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}, \\mathrm {C I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}\\right), \\quad \\forall i, \\quad \\forall i \\neq j.\\end{array}\\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.595 + ], + "angle": 0, + "content": "We threshold these distances for each target identity to get an ROC curve, and average across the target identities to get the overall area under the curve (AUC). We note that this AUC measures one model's ability to classify a synthetic video as self-reenactment or as cross-reenacted. We conduct further analysis of our model's ability to classify other categories of videos—such as, evaluating AUC on same-utterance self- vs. cross-reenactments or on scripted vs. free-form monologues—in the Supplement." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.615, + 0.512, + 0.631 + ], + "angle": 0, + "content": "6.1 Comparisons with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Avatar fingerprinting is a novel task, and no existing methods directly address it. The closest related works aim at detecting real versus synthetic media. As discussed in Section 2, some of these detectors learn identity-specific features such as facial expressions and head poses [9], or facial shapes and motion [20] and can serve as baselines for the task of avatar fingerprinting with some adaptation. The work by Agarwal et al. trains a model to detect synthetic videos of a specific identity [9]. To adapt it to our task, we train 35 different models, one for each identity in the evaluation, by splitting the corresponding original videos into two subsets. We then test each model on the self- and cross-reenactment videos of the corresponding identity. ID-Reveal, trained on a large-scale dataset, learns an embedding space where real videos of a specific identity are grouped together [20]. Since it shows good generalization to new identities, for the task of synthetic media detection, we directly use the pre-trained model on our data to detect, once again, self- versus cross-reenactment. Figure 5 shows the ROC curves for our method compared to these baselines, on three face-reenactment generators (face-vid2vid [60]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.144, + 0.416, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.144, + 0.58, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.144, + 0.746, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.264, + 0.784, + 0.292 + ], + "angle": 0, + "content": "Fig. 6: Generalization to new generators. To study the robustness to new talking-head generators, we train three versions of our model on three different generators and test on all three." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.785, + 0.362 + ], + "angle": 0, + "content": "LIA [62], and TPS [34]). Our method (AUC=0.868 on face-vid2vid) outperforms by a wide margin both ID-Reveal (AUC=0.720 on face-vid2vid), and the method by Agarwal et al. (AUC=0.726 on face-vid2vid). We also note that, unlike ID-Reveal and our method, Agarwal et al. uses a different model per identity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.38, + 0.496, + 0.394 + ], + "angle": 0, + "content": "6.2 Generalization to New Generators" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.489 + ], + "angle": 0, + "content": "For an avatar fingerprinting model algorithm to be broadly applicable, generalization to new talking-head generators that are not seen in training is crucial. Since our dataset contains videos synthesized by three different generators, we can train three models, one with each generator, and test these models on all three generators. Figure 6 shows the resulting ROC curves and AUC values: the overlap of the curves and similar AUC values in each subplot confirms that our method generalizes well to new generators." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.363, + 0.521 + ], + "angle": 0, + "content": "6.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Our method outperforms existing baselines by introducing two novel components: the dynamic facial identity features, which capture the facial dynamics in a compact and expressive way, and the loss function, which defines the shape of the identity embedding. Here we study the contribution of each, using the face-vid2vid generator for training and testing. We evaluate the contribution of our dynamic facial identity features by swapping them with 3DMM features [13], a popular choice to capture facial dynamics. Since we use a temporal CNN backbone similar to the one from ID-Reveal, for this ablation we use the loss function proposed in their original paper [20]. We re-train the same network using our features and observe a jump from 0.718 to 0.754 in terms of AUC. Upon inspection we notice that the 3DMM features tend to over-smooth the facial motion, and are unable to capture subtle dynamics that prove critical to avatar fingerprinting, and which our features capture. We also evaluate the contribution of our dynamic identity embedding loss and observe a further improvement (AUC 0.868). With our loss formulation, the advantage of \\(\\mathbf{R}_{j,\\mathrm{ID}_1,\\mathrm{ID}_2}(t)\\) in Eq. 5 is also evident when compared against a model trained without this term (AUC 0.850). Table 2 summarizes this ablation study. Additional experiments in the Supplement show the impact of \\(F\\), performance on scripted vs. free-from monologues, and robustness to video distortions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.79, + 0.339, + 0.804 + ], + "angle": 0, + "content": "6.4 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Our algorithm is less discriminative of subjects that are less emotive and more neutral. In the future, relying on more granular dynamic signatures that can extract micro" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "E. Prashnani et al." + }, + { + "type": "table", + "bbox": [ + 0.323, + 0.145, + 0.681, + 0.216 + ], + "angle": 0, + "content": "
Input FeaturesLossAUC
3DMMID-Reveal rec. loss [20]0.718
Landmark distancesID-Reveal rec. loss [20]0.754
Landmark distancesOur loss without Rj, ID1, ID2(t)0.850
Landmark distancesOur loss with Rj, ID1, ID2(t)0.868
" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.223, + 0.782, + 0.237 + ], + "angle": 0, + "content": "Table 2: Ablation study showing the importance of our input features and loss function design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.787, + 0.326 + ], + "angle": 0, + "content": "expressions can help alleviate this. The performance of our method degrades when expressions that are critical to verifying the driving identity are not captured by the synthetic portrait generator. Lastly, our dataset currently features only one style of interaction: one-on-one conversations. We plan to expand to other conversation styles, such as one-way speeches, in future." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.345, + 0.381, + 0.363 + ], + "angle": 0, + "content": "7 Societal Impact" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.787, + 0.46 + ], + "angle": 0, + "content": "We acknowledge the societal importance of introducing guardrails when it comes to the use of talking-head generation technology. We present this work as a step towards trustworthy use of such technologies. Nevertheless, our work could be misconstrued as having solved the problem and inadvertently accelerate the unhindered adoption of synthetic talking-head technology. We do not advocate for this. Instead, we stress that this is the first work on this topic and underscore the importance of further research." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.352, + 0.485 + ], + "angle": 0, + "content": "8 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.787, + 0.628 + ], + "angle": 0, + "content": "Highly photo-real synthetic talking-head generators are becoming increasingly beneficial to applications such as video conferencing and AR/VR-based remote interactions. This trend raises the important new research question of how best to also ensure their safe use in such scenarios. To this end, we investigate the new problem of avatar fingerprinting, to authenticate legitimate talking-heads created by authorized users. We leverage the fact that driving individuals have uniquely identifying dynamic motion signatures, which are also preserved in the synthetic videos that they drive. Since none exists, we contribute a new large-scale dataset carefully designed for avatar fingerprinting and related tasks. We hope that our work lays the foundation for further research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.824 + ], + "angle": 0, + "content": "Acknowledgements. We would like to thank the data-capture participants, and Desiree Luong, Woody Luong, and Josh Holland for their help with Figure 2. We acknowledge David Taubenheim for the voiceover in the demo video, and Abhishek Badki for help with the training infrastructure. We thank Joohwan Kim, Rachel Brown, Anjul Patney, Ben Boudaoud, Josef Spjut, Saori Kaji, Nikki Pope, and Kai Pong for their help with putting together the data capture protocols, informed consent form, photo release form, and agreements for data governance and third-party data sharing. Koki Nagano, Ekta Prashnani, and David Luebke were partially supported by DARPA's Semantic Forensics (SemaFor) contract (HR0011-20-3-0005). This research was funded, in part, by DARPA's Semantic Forensics (SemaFor) contract HR0011-20-3-0005. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.761, + 0.84 + ], + "angle": 0, + "content": "Distribution Statement \"A\" (Approved for Public Release, Distribution Unlimited)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.206 + ], + "angle": 0, + "content": "1. Apple's vision pro. https://www.apple.com/apple-vision-pro/, accessed: 2024-03-06" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.207, + 0.628, + 0.221 + ], + "angle": 0, + "content": "2. Heygen. https://www.heygen.com, accessed: 2023-11-16" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.222, + 0.785, + 0.247 + ], + "angle": 0, + "content": "3. Microsoft teams mesh. https://www.microsoft.com/en-us/microsoft-teams/microsoft-mesh, accessed: 2024-03-06" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.249, + 0.686, + 0.261 + ], + "angle": 0, + "content": "4. Myheritage. https://www.myheritage.com, accessed: 2023-11-16" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.262, + 0.785, + 0.286 + ], + "angle": 0, + "content": "5. Nvidia's maxine. https://developer.nvidia.com/maxine, accessed: 2024-03-06" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.328 + ], + "angle": 0, + "content": "6. Agarwal, S., El-Gaaly, T., Farid, H., Lim, S.N.: Detecting deep-fake videos from appearance and behavior. 2020 IEEE International Workshop on Information Forensics and Security (WIFS) pp. 1-6 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.369 + ], + "angle": 0, + "content": "7. Agarwal, S., Farid, H.: Detecting deep-fake videos from aural and oral dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.37, + 0.785, + 0.409 + ], + "angle": 0, + "content": "8. Agarwal, S., Farid, H., Fried, O., Agrawala, M.: Detecting deep-fake videos from phonemeviseme mismatches. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.785, + 0.45 + ], + "angle": 0, + "content": "9. Agarwal, S., Farid, H., Gu, Y., He, M., Nagano, K., Li, H.: Protecting world leaders against deep fakes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.785, + 0.49 + ], + "angle": 0, + "content": "0. Agarwal, S., Hu, L., Ng, E., Darrell, T., Li, H., Rohrbach, A.: Watch those words: Video falsification detection using word-conditioned facial motion. In: IEEE Winter Conference on Applications of Computer Vision (WACV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.529 + ], + "angle": 0, + "content": "1. Albright, M., McCloskey, S.: Source generator attribution via inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.53, + 0.785, + 0.556 + ], + "angle": 0, + "content": "2. Baluja, S.: Hiding images in plain sight: Deep steganography. In: Advances in Neural Information Processing Systems (NeurIPS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.558, + 0.785, + 0.583 + ], + "angle": 0, + "content": "3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of SIGGRAPH (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.584, + 0.785, + 0.611 + ], + "angle": 0, + "content": "4. Blanz, V., Vetter, T.: Face recognition based on fitting a 3D morphable model. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.612, + 0.785, + 0.638 + ], + "angle": 0, + "content": "5. Boháček, M., Farid, H.: Protecting world leaders against deep fakes using facial, gestural, and vocal mannerisms. Proceedings of the national academy of Sciences (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.639, + 0.785, + 0.678 + ], + "angle": 0, + "content": "6. Cao, H., Cooper, D.G., Keutmann, M.K., Gur, R.C., Nenkova, A., Verma, R.: CREMA-D: Crowd-sourced emotional multimodal actors dataset. IEEE Transactions on Affective Computing (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.679, + 0.785, + 0.704 + ], + "angle": 0, + "content": "7. Chung, J.S., Nagrani, A., Zisserman, A.: Voxceleb2: Deep speaker recognition. In: INTER-SPEECH (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.706, + 0.785, + 0.731 + ], + "angle": 0, + "content": "8. Cozzolino, D., Nießner, M., Verdoliva, L.: Audio-visual person-of-interest deepfake detection (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.733, + 0.785, + 0.772 + ], + "angle": 0, + "content": "9. Cozzolino, D., Poggi, G., Verdoliva, L.: Extracting camera-based fingerprints for video forensics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.773, + 0.785, + 0.812 + ], + "angle": 0, + "content": "20. Cozzolino, D., Rössler, A., Thies, J., Nießner, M., Verdoliva, L.: ID-Reveal: Identity-aware DeepFake video detection. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., Canton Ferrer, C.: The deepfake detection challenge dataset. arXiv preprint arXiv:2006.07397 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "E. Prashnani et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.187 + ], + "angle": 0, + "content": "22. Dong, X., Bao, J., Chen, D., Zhang, T., Zhang, W., Yu, N., Chen, D., Wen, F., Guo, B.: Protecting celebrities from deepfake with identity consistency transformer. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.188, + 0.786, + 0.214 + ], + "angle": 0, + "content": "23. Drobyshev, N., Chelishev, J., Khakhulin, T., Ivakhnenko, A., Lempitsky, V., Zakharov, E.: MegaPortraits: One-shot megapixel neural head avatars (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.215, + 0.786, + 0.253 + ], + "angle": 0, + "content": "24. Fox, G., Liu, W., Kim, H., Seidel, H.P., Elgharib, M., Theobalt, C.: VideoForensicsHQ: Detecting high-quality manipulated face videos. In: IEEE International Conference on Multimedia and Expo (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.253, + 0.786, + 0.278 + ], + "angle": 0, + "content": "25. Fridrich, J.: Steganography in Digital Media: Principles, Algorithms, and Applications. Cambridge University Press (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.279, + 0.786, + 0.318 + ], + "angle": 0, + "content": "26. Ge, S., Lin, F., Li, C., Zhang, D., Wang, W., Zeng, D.: Deepfake video detection via predictive representation learning. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.786, + 0.344 + ], + "angle": 0, + "content": "27. Hadsell, R., Chopra, S., LeCun, Y.: Dimensionality reduction by learning an invariant mapping. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.345, + 0.786, + 0.383 + ], + "angle": 0, + "content": "28. Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M.: Lips don't lie: A generalisable and robust approach to face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.422 + ], + "angle": 0, + "content": "29. Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M.: Lips don't lie: A generalisable and robust approach to face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.423, + 0.786, + 0.462 + ], + "angle": 0, + "content": "30. He, Y., Gan, B., Chen, S., Zhou, Y., Yin, G., Song, L., Sheng, L., Shao, J., Liu, Z.: Forgerynet: A versatile benchmark for comprehensive forgery analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4360-4369 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.463, + 0.786, + 0.487 + ], + "angle": 0, + "content": "31. Hill, H., Johnston, A.: Categorizing sex and identity from the biological motion of faces. Current Biology (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.786, + 0.527 + ], + "angle": 0, + "content": "32. Honari, S., Molchanov, P., Tyree, S., Vincent, P., Pal, C., Kautz, J.: Improving landmark localization with semi-supervised learning. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.786, + 0.553 + ], + "angle": 0, + "content": "33. Hong, F.T., Zhang, L., Shen, L., Xu, D.: Depth-aware generative adversarial network for talking head video generation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.554, + 0.697, + 0.566 + ], + "angle": 0, + "content": "34. Jian Zhao, H.Z.: Thin-plate spline motion model for image animation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.567, + 0.786, + 0.605 + ], + "angle": 0, + "content": "35. Jiang, L., Li, R., Wu, W., Qian, C., Loy, C.C.: Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.644 + ], + "angle": 0, + "content": "36. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.645, + 0.786, + 0.669 + ], + "angle": 0, + "content": "37. Khakhulin, T., Sklyarova, V., Lempitsky, V., Zakharov, E.: Realistic one-shot mesh-based head avatars. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.67, + 0.786, + 0.709 + ], + "angle": 0, + "content": "38. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.71, + 0.786, + 0.735 + ], + "angle": 0, + "content": "39. Kim, H., Elgharib, M., Zollhöfer, M., Seidel, H.P., Beeler, T., Richardt, C., Theobalt, C.: Neural style-preserving visual dubbing. ACM Transactions on Graphics (ToG) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.736, + 0.786, + 0.773 + ], + "angle": 0, + "content": "40. Kim, H., Garrido, P., Tewari, A., Xu, W., Thies, J., Niessner, M., Pérez, P., Richardt, C., Zollhöfer, M., Theobalt, C.: Deep video portraits. ACM Transactions on Graphics (ToG) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.775, + 0.786, + 0.8 + ], + "angle": 0, + "content": "41. Knappmeyer, B., Thornton, I., Bulthoff, H.: Facial motion can determine facial identity. Journal of Vision (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.801, + 0.786, + 0.839 + ], + "angle": 0, + "content": "42. Kwon, P., You, J., Nam, G., Park, S., Chae, G.: Kodf: A large-scale korean deepfake detection dataset. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10744-10753 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.187 + ], + "angle": 0, + "content": "43. Li, J., Xie, H., Yu, L., Zhang, Y.: Wavelet-enhanced weakly supervised local feature learning for face forgery detection. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.188, + 0.785, + 0.214 + ], + "angle": 0, + "content": "44. Li, Y., Sun, P., Qi, H., Lyu, S.: Celeb-DF: A large-scale challenging dataset for DeepFake forensics. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.215, + 0.785, + 0.252 + ], + "angle": 0, + "content": "45. Liu, B., Liu, B., Ding, M., Zhu, T., Yu, X.: Ti2net: Temporal identity inconsistency network for deepfake detection. In: IEEE Winter Conference on Applications of Computer Vision (WACV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.254, + 0.785, + 0.292 + ], + "angle": 0, + "content": "46. Livingstone, S.R., Russo, F.A.: The ryerson audio-visual database of emotional speech and song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.293, + 0.785, + 0.318 + ], + "angle": 0, + "content": "47. Luo, X., Zhan, R., Chang, H., Yang, F., Milanfar, P.: Distortion agnostic deep watermarking. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.785, + 0.344 + ], + "angle": 0, + "content": "48. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.345, + 0.785, + 0.37 + ], + "angle": 0, + "content": "49. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit Warping for Animation with Image Sets. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.422 + ], + "angle": 0, + "content": "50. Munir, S., Batool, B., Shafiq, Z., Srinivasan, P., Zaffar, F.: Through the looking glass: Learning to attribute synthetic text generated by language models. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.423, + 0.785, + 0.448 + ], + "angle": 0, + "content": "51. O'Toole, A.J., Roark, D.A., Abdi, H.: Recognizing moving faces: A psychological and neural synthesis. Trends in Cognitive Sciences (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.449, + 0.785, + 0.487 + ], + "angle": 0, + "content": "52. Passos, L.A., Jodas, D., da Costa, K.A., Junior, L.A.S., Rodrigues, D., Del Ser, J., Camacho, D., Papa, J.P.: A review of deep learning-based approaches for deepfake content detection. arXiv preprint arXiv:2202.06095 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.527 + ], + "angle": 0, + "content": "53. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: FaceForensics++: Learning to detect manipulated facial images. In: IEEE International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.566 + ], + "angle": 0, + "content": "54. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: Faceforensics++: Learning to detect manipulated facial images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1-11 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.567, + 0.785, + 0.605 + ], + "angle": 0, + "content": "55. Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: A unified embedding for face recognition and clustering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.631 + ], + "angle": 0, + "content": "56. Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.632, + 0.785, + 0.657 + ], + "angle": 0, + "content": "57. Sun, K., Yao, T., Chen, S., Ding, S., Li, J., Ji, R.: Dual contrastive learning for general face forgery detection (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.683 + ], + "angle": 0, + "content": "58. Tancik, M., Mildenhall, B., Ng, R.: Stegastamp: Invisible hyperlinks in physical photographs. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.684, + 0.785, + 0.722 + ], + "angle": 0, + "content": "59. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.723, + 0.785, + 0.761 + ], + "angle": 0, + "content": "60. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.762, + 0.785, + 0.8 + ], + "angle": 0, + "content": "61. Wang, X., Han, X., Huang, W., Dong, D., Scott, M.R.: Multi-similarity loss with general pair weighting for deep metric learning. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.801, + 0.785, + 0.84 + ], + "angle": 0, + "content": "62. Wang, Y., Yang, D., Bremond, F., Dantcheva, A.: Latent image animator: Learning to animate images via latent space navigation. In: International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "E. Prashnani et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.174 + ], + "angle": 0, + "content": "63. Yacoob, Y.: Gan-scanner: A detector for faces of stylegan+ (2021), https://github.com/yaseryacoob/GAN-Scanner" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.787, + 0.214 + ], + "angle": 0, + "content": "64. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.215, + 0.787, + 0.254 + ], + "angle": 0, + "content": "65. Yu, N., Skripniuk, V., Chen, D., Davis, L., Fritz, M.: Responsible disclosure of generative models using scalable fingerprinting. In: International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.255, + 0.787, + 0.293 + ], + "angle": 0, + "content": "66. Zakharov, E., Ivakhnenko, A., Shysheya, A., Lempitsky, V.: Fast bi-layer neural synthesis of one-shot realistic head avatars. In: European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.294, + 0.787, + 0.332 + ], + "angle": 0, + "content": "67. Zakharov, E., Shysheya, A., Burkov, E., Lempitsky, V.: Few-shot adversarial learning of realistic neural talking head models. In: IEEE International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.787, + 0.373 + ], + "angle": 0, + "content": "68. Zheng, Y., Bao, J., Chen, D., Zeng, M., Wen, F.: Exploring temporal coherence for more general video face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.373 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_origin.pdf b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c98fd9c53d6ed0937f1388f5c5bf30576f033b6e --- /dev/null +++ b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/b4dea277-ef85-45d4-9e38-90aa487a616d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e01cea3bd0c7f462e670c47ed8b02eb3f8d1cd770a736f2ac5b7c3579fc1076 +size 25301446 diff --git a/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/full.md b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7c88d8eb5a4e14f11696b61b0c995a49d52108db --- /dev/null +++ b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/full.md @@ -0,0 +1,329 @@ +# Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos + +Ekta Prashnani, Koki Nagano, Shalini De Mello, David Luebke, and Orazio Gallo + +NVIDIA + +{eprashnani, knagano, shalinig, dluebke, ogallo}@nvidia.com + +Abstract. Modern avatar generators allow anyone to synthesize photorealistic real-time talking avatars, ushering in a new era of avatar-based human communication, such as with immersive AR/VR interactions or videoconferencing with limited bandwidths. Their safe adoption, however, requires a mechanism to verify if the rendered avatar is trustworthy: does it use the appearance of an individual without their consent? We term this task avatar fingerprinting. To tackle it, we first introduce a large-scale dataset of real and synthetic videos of people interacting on a video call, where the synthetic videos are generated using the facial appearance of one person and the expressions of another. We verify the identity driving the expressions in a synthetic video, by learning motion signatures that are independent of the facial appearance shown. Our solution, the first in this space, achieves an average AUC of 0.85. Critical to its practical use, it also generalizes to new generators never seen in training (average AUC of 0.83). The proposed dataset and other resources can be found at: https://research.nvidia.com/labs/nxp/avatar-fingerprinting/. + +Keywords: Synthetic Media Verification $\cdot$ Avatar Generators $\cdot$ Avatar Attribution + +# 1 Introduction + +Recent digital avatar generators have fueled a myriad of computer vision and graphics applications, allowing anyone to synthesize real-time photorealistic personas. Major companies are now supporting avatar-driven remote interactions over immersive AR/VR (e.g. Meta's Pixel Codec Avatar [48], Apple's Vision Pro Persona [1]) or video conferencing (e.g. NVIDIA's MAXINE [5], Microsoft's Teams [3]), and selfie filters for altering and enhancing appearance (e.g. by Snap and Tiktok). While the avatar generation technology today is still young, the legitimate use of synthetic avatars will be ubiquitous in the future. Without proper guardrails, this poses a real risk of unauthorized use and large-scale spread of visual disinformation. To ensure the safe use of such a technology, the relevant question is no longer whether the content is "real" or not—since, by design, the videos and avatars are all synthetic—but rather, whether the synthetically-generated videos and avatars are "trustworthy" or not. + +When video conferencing, for instance, a synthetic video portrait generator can be used to save valuable bandwidth by reconstructing a synthetic avatar of the sender at the receiver's end, using only a frame of the target identity and a compact representation of the speaker's facial motion. To ensure the authorized use of such synthetically-generated videos, we want to verify if the driving identity behind a synthetic video $(\mathrm{ID}_1$ or $\mathrm{ID}_2$ in Figure 1) is authorized to control the likeness, or the appearance, of the + +![](images/4eae27ba7b367c955894eebea718d85ee1012e9264b5c7b84c556e11c7b33fba.jpg) +Fig. 1: Talking-head avatar generators can synthesize realistic videos of a target identity from driving videos of different identities. Our method extracts appearance-agnostic temporal facial features and learns an embedding in which the synthetic videos driven by one identity fall close to each other and far from those driven by other identities, regardless of the appearance of the synthetic video. By comparing distances in the embedding space, we evaluate whether an avatar is driven by an authorized identity or not. During evaluation, we only rely on the synthetic videos as input, without requiring any prior knowledge about the driving identity. + +synthetic video (target identity in Figure 1). Crucially, we want to only leverage the synthetic video avatars to do so. We call this novel task avatar fingerprinting. + +We leverage a simple but fundamental observation: facial motions tend to be idiosyncratic, that is, individuals talk and emote in unique ways. For instance, someone may raise one of her eyebrows more than the other, or smile more while talking. These "dynamic identity signatures" [51] have been shown to carry enough information for humans to recognize other individuals, even when the physical appearance of their face is altered [31, 41, 51]. This makes them attractive for our task, as they can be derived solely from the driving identity of a talking-head video, regardless of the appearance. + +Fortunately, modern avatar generators are becoming increasingly accurate at capturing the facial motion of a person and rendering it onto a target identity. As a first solution to avatar fingerprinting, we then propose to estimate a dynamic identity signature for an identity from the synthesized videos they drove—regardless of the target identities shown. Specifically, we extract facial landmarks and their temporal dynamics from the video. We then introduce a novel contrastive loss to learn a dynamic identity embedding, a space where the dynamic identity signatures of a driving identity across multiple videos and target identities are close to each other, and far from those of other driving identities (see Figure 1). We show that this approach, albeit straightforward, is robust and generalizes to generators not seen in training. + +Avatar fingerprinting is a new task, and no existing datasets serve its training and validation requirements. There are two key requirements from a dataset. First, we need videos of multiple subjects delivering both scripted and free-form monologues, captured under realistic conditions, such as varying video quality and gaze direction. This allows us to assess if a model leverages talking styles, and not the specific choice of words. Second, to evaluate if the model learns to extract dynamic identity signatures effectively from synthetic videos, we need synthetic talking-head videos for the case in which the driving and target identities are different (cross-reenactment), and that in which they match (self-reenactment). Unfortunately, existing datasets of real videos + +only fulfill a subset of these requirements (Table 1 and Section 4). Further, no existing dataset of synthetic videos contain both self- and cross-reenactments per subject or use the state-of-the-art talking-head generation technology. To foster research in this new domain, then, we introduce the NVIDIA Facial Reenactment (NVFAIR) dataset, containing real and synthetic face reenactment talking-head videos (Figure 2). Our dataset, which includes ethnically diverse subjects, provides $10 \times$ more synthetic facial reenactments than the second largest dataset, for a total of over 650,000 synthetic videos. It is also the only one using multiple state-of-the-art face reenactment generators [34,60,62], or to provide cross-reenactments driven by all identities, which is critical for training and evaluating avatar fingerprinting algorithms. + +# In summary: + +- We introduce the novel task of avatar fingerprinting, which focuses on verifying the driving identity of synthetic talking-head avatars, rather than classifying them as real or synthetic (by design, all inputs to our model are synthetic). +- We release the first large-scale dataset of subjects delivering scripted and natural monologues, complete with self- and cross-reenactment videos synthesized with multiple state-of-the-art generators. +- We propose a solution for this novel task in the context of video conferencing by extracting person-specific motion signatures, and demonstrate its robustness to various distortions and avatar generators not seen in training. + +# 2 Related Work + +Our proposed avatar fingerprinting task aims to verify authorized use of synthetic avatars: this is fundamentally a different problem than traditional forensics research where one aims to detect synthetic media (e.g. deepfake detection) or actively mark synthetic content. In our case, by design, the content being evaluated is always synthetic: we aim to evaluate its authorized use. Since this is a novel task, no methods exist to directly address it. No methods currently exist to directly address this novel task. Here we discuss the related areas of research. + +Learning-based Attribution of Synthetic Media. Learning-based approaches have been used to identify the origin of synthetic media, or to determine if it has been manipulated or altered in some way. Previous work used a pre-trained GAN generator to attribute a synthesized image to its generator via GAN inversion by leveraging the fact that a real image is less invertible [11, 36, 63]. Yet other works focus on attributing other forms of synthetic media, such as text [50]. In contrast, our focus is to attribute a talking-head avatar to the identity driving it, regardless of the appearance of the avatars. Some existing works learn fingerprints associated with cameras to determine whether a video is manipulated [19], or embed watermarks into images and videos [12, 25, 47, 58], which are also shown to transfer to GAN-generated images [64]. Subsequent research introduced a watermark-based conditional GANs for scalable fingerprinting [65]. Our method, in contrast, is a passive technique that does not rely on active watermarks. + +Deepfake Detection Based on Identity-Specific Features. Deepfake detection ("is a video synthetic?") and avatar fingerprinting ("whose identity is used to generate this synthetic video?") are fundamentally different tasks. Most existing solutions for deepfake detection train a real-vs-synthetic classifier [26, 28, 43, 52, 57, 68], and therefore + +cannot be adapted to avatar fingerprinting (where all inputs are synthetic). However, a specific class of detectors leverage identity-specific features to detect synthetic videos by posing the detection problem as an identity-recognition problem. In our experiments, we evaluate some of these methods as baselines. Specifically, Agarwal et al. exploit person-specific patterns in facial expressions to detect fake videos [9]. ID-Reveal used facial shapes and motions encoded in a low-dimensional space of a 3D morphable model [13] to handle both face-swapping and face-reenactment deepfakes [20]. Other works explored soft-biometric approaches such as leveraging vocal mannerisms [15], phoneme-viseme consistencies [8], word-facial expression consistencies [6,10], and dynamics of ears [7]. While many of these works need person-specific training, previous works [6, 18, 20, 45] extended this idea to train a CNN-based detector using a large-scale in-the-wild video data [17] and variants of contrastive learning [27, 38, 55, 61]. Agarwal et al. combined static facial appearance using a facial recognition model and dynamic facial behaviors using a CNN, and showed that this approach is effective for detecting face-swap deepfakes [6]. Yet another line of research has explored temporal inconsistencies of face identities within a video [45], and identity inconsistencies of inner and outer face regions [22]. Our experiments show that for avatar fingerprinting, such features that are designed to distinguish real from synthetic videos are not reliable, since we only have access to synthetic videos. + +Dynamic Facial Identity Signatures. Cognitive scientists have studied the impact of "dynamic facial identity signatures" (i.e., characteristic or identity-specific movements of a face) for identity recognition for humans [51]. In one experiment, scientists projected facial animations generated by human actors onto an average head and found that subjects discriminated between individuals based solely on facial motion [31]. In another, subjects correctly attributed animations of synthetic faces to their morphed versions [41]. While these studies point to the existence of "dynamic facial identity signatures" that humans rely on, ours is the first method that isolates these from videos. Talking-Head Datasets and Generators. Existing talking-head datasets include those that contain only real videos showing a variety of emotions [16, 46, 59], and others that also contain synthetic videos [21, 24, 30, 35, 42, 44, 53]. These datasets cater to traditional forensics and facial analysis tasks. Therefore, they do not contain self-and cross-reenactments driven by multiple identities, as well as scripted and free-form monologues across diverse capture settings. There novel requirements posed by the avatar fingerprinting task motivate us to design our own dataset. We focus specifically on face-reenactment talking-head avatar generators for synthetic video generation—this class of generators are the most relevant to AR/VR interactions, video conferencing, and several other applications [2-5, 48]—and combine various modes of human expression for capturing real videos. Given a target facial image and a driving video, these generators reenact the target image using the facial expressions and head pose from the driving video [23, 33, 34, 37, 49, 56, 60, 62, 66, 67]. Another class of talking-head generators use person-specific models [40] and some models aim to preserve the style of the target identity in the synthesized video [39]. However, these models require person-specific training, making them difficult to scale. + +# 3 Terminology + +We seek to verify the trustworthiness of a synthesized talking-head video, termed target video. We assume that an avatar-generation tool (e.g., [60]) created it by animating + +![](images/8a04554487271f3aa6ec5972c84632c6fe7d5ba5b7d1cfea152b0541cc4272d8.jpg) +Fig. 2: We introduce the NVFAIR dataset, containing real and synthetic talking-head videos. We capture subjects talking in both scripted and free-form settings. To encourage natural performance, we record the subjects while videoconferencing with each other (left). We then synthesize more than 650,000 talking-head videos—the largest collection till date—using three state-of-the-art face-reenactment talking-head generators. On the right, each row corresponds to a driving identity $(\mathrm{ID}_i \rightarrow (\cdot))$ and each column corresponds to a different target identity $((\cdot) \rightarrow \mathrm{ID}_i)$ . The videos in which driving and target identity match are self-reenactments, the rest are cross-reenactments. + +an image (target image) using the expressions and head poses obtained from another video, the driving video. We call driving identity the identity of the person in the driving video, and target identity the identity of the person in the target image. When driving and target identities match, the target video is a self-reenactment, while the case of a driving identity used to animate a different target identity is cross-reenactment. In both cases, the appearance of the synthesized video is derived from the target identity. This terminology allows us to formally state our goal: we want to verify that a target video is a self-reenactment. With this terminology in mind, we introduce our dataset, which includes real videos as well as self- and cross-reenactment videos. + +# 4 The NVIDIA Facial Reenactment (NVFAIR) Dataset + +Recall that avatar fingerprinting is not about detection of synthetic media. Rather, we already know a video to be synthesized, and seek to verify that the driving identity is authorized. This new task dictates a set of requirements for the dataset to be effective for training and evaluation. Specifically, we need a dataset that contains + +1. multiple real videos per identity, with scripted and free-form conversations, and with both natural and prescribed emotions, +2. self- and cross-reenactments of target identities, with cross-reenactments driven by all subjects to allow for a variety of driving facial structures, and +3. multiple face-reenactment generators. + +All relevant existing datasets only capture a subset of these requirements (see Table 1; the Supplement contains further discussion). Moreover they do not use the state-of-the-art talking-head generation technology to synthesize the self- and cross-reenactment videos. We introduce the NVFAIR dataset that features all of the above properties. With over 650,000 synthetic videos, it provides $10 \times$ as many videos as the next largest dataset, and uses three different state-of-the-art generators for the self- and cross-reenactment videos. Figure 2 shows an overview of data capture and synthesis. + +# 4.1 Real Data Capture + +Capturing videos of monologues delivered by different subjects for the purpose of identity verification introduces two conflicting goals. On the one hand a controlled evalua- + +
Dataset# Subjects (source)(F)ree / (S)scripted?Emotion: (N)atural / (P)rescribed(R)eel / (S)elf / (C)ross-reenact.?#Face Reenact.Avg. Videos per Subject# Face- reenact. GeneratorsEthnic Diversity
RAVDESS [46]24 (new)(S)(P)(R) onlyN/A120 (R)N/A✓
MEAD [59]60 (new)(S)(P)(R) onlyN/A720 (R)N/A✓
CREMA-D [16]91 (new)(S)(P)(R) onlyN/A81 (R)N/A✓
VFHQ [24]36 ([46])(S)(P)+(N)(R)+(S)1,7371201✓
FF++ [54]1000 (YT)(F)(N)(R)+(C)2,0001(R)+2(C)2✓
KoDF [42]403 (new)(F)+(S)(N)(R)+(C)61,000150(R)+151(C)1X
NVFAIR (Ours)161 (46 new + [46], [16])(F)+(S)(P)+(N)(R)+(S)+(C)654,72676(R)+228(S) +3,840(C)3✓
+ +Table 1: The existing talking-head video datasets were designed for tasks such as deepfake detection or facial emotion analysis. Avatar fingerprinting is a fundamentally different task. As a result, no existing dataset satisfies the requirements for training and evaluating models for avatar fingerprinting. To overcome this limitation, we introduce the NVFAIR dataset, which is the first dataset that offers the complete set of monologue modalities, and features the largest collection of facial reenactments to date. Specifically, it provides scripted and free-form monologues, with natural and prescribed emotions, and self- and cross-reenactments (driven by all remaining subjects) synthesized using three generators, alongside original videos for newly-recorded subjects. + +tion of the trained models requires predictability of what is spoken to prevent identification algorithms from latching onto the spoken content itself. On the other, we want the subjects to act as they would in a casual conversation, rather than reciting memorized text, to capture their uniquely identifying mannerisms. We address this trade-off by recording the subjects while videoconferencing in pairs, which creates the impression of being in a natural conversation. This differs from existing datasets, in which the subjects look at the camera, but are not interacting with another person during the recording [16,42,46]. We also design two distinct recording strategies: a free-form stage where the subjects are given only general guidance on the topics, and a more controlled scripted stage in which subjects speak short, memorized monologues of 2-3 sentences each, see Figure 2(a). To capture the variability of real-life scenarios, we provided minimal instructions on how to setup the video call, allowing for diverse face, scale, and lightning, bandwidth stability, and background scene clutter. In total we record 46 subjects of diverse genders, ages, and ethnicities, while strictly abiding by a pre-approved IRB protocol (see Supplement for details and privacy considerations). + +Stage I: Free-Form Monologues. In this first stage, the two subjects on the call alternate between asking and answering seven pre-defined questions. The questions are designed to avoid sensitive or potentially inflammatory topics. This is critical because we later use sentences spoken by one individual to animate the video of a second individual, quite literally putting words in their mouths. The complete list of questions is in the Supplement. To further create a natural interaction, the subject listening is encouraged to actively but silently engage with the one speaking (e.g., by nodding or smiling). + +Stage II: Scripted Monologues. For this stage, we prepared thirty short utterances consisting of two or three sentences each. We chose this length to allow for memorization, while still providing enough content to trigger facial expressions. However, to avoid inducing unnatural expressions, we do not prescribe specific emotions for each utterance. For instance, we do not ask to express anger for a sentence, but we do choose sentences that may naturally evoke it, and used punctuation to encourage it, e.g., "Will you please + +answer the darn phone? The constant ringing is driving me insane!" We instruct the subjects to split their screens to show both this list and video call and encourage them to speak to their recording partner when reciting, see Stage II in Figure 2(a). More details, including the full list of utterances can be found in the Supplement. + +# 4.2 Synthetic Talking-Head Videos + +Using the videos described in Section 4.1, as well as the original videos from the CREMA-D [16] and RAVDESS [46] datasets, we generate synthetic talking-head videos to train and evaluate our avatar fingerprinting algorithm. Specifically, we pool the 91 identities from the original videos of CREMA-D [16], the 24 identities from those of RAVDESS [46], and the 46 from our own video-conferencing data capture, for a total of 161 unique identities $\mathcal{I}$ . Recall that we have several real videos for each identity $\mathrm{ID}_i\in \mathcal{I}$ . To avoid a combinatorial explosion of synthetic videos, for all pairs of identities $\mathrm{ID}_i$ and $\mathrm{ID}_j$ , we use $\mathrm{ID}_j$ as the target identity and we randomly select 8 of the videos of $\mathrm{ID}_i$ to generate 8 cross-reenactment videos, $\{\mathrm{ID}_i^k\to \mathrm{ID}_j\}_{k = \{1,\dots ,8\}}$ (all 8 share the same target image). We also generate self-reenactment videos for each of the target identities, by animating their neutral-face images derived from captured videos with each of their real videos. + +We use three different generators for synthesizing the videos for all the 161 identities: face-vid2vid [60], LIA [62], and TPS [34]. This allows us to test if our model generalizes across generators. We chose these talking-head generators because they are the state of the art and they preserve the identity-specific facial motion dynamics well. Nevertheless, the reconstruction is not perfect; for instance, in the third row of Figure 2(b) the person in the driving video $(\mathrm{ID}_5)$ is squinting, but the eyes are shut in all the synthetic videos, including the self-reenactment video. In total we generate more than 650,000 synthetic videos, which required more than 2,500 RTX 3090 GPU hours. More details are in the Supplement. + +# 5 Method + +Overview. We seek to verify the driving identity of a synthetic video, independently of the target identity. We leverage the finding from cognitive science research that each person emotes in unique ways when communicating, and that this signal is sufficient for recognition, even when the actual appearance is artificially corrupted [31, 41, 51]. We note that our method does not latch onto artifacts introduced by the generators—a property that we demonstrate by showing generalization to new generators not seen during training. Rather, our features capture the dynamics of the expressions, like the way a person frowns, or the way she smiles. Notably, they are distinct from the temporal artifacts introduced by the generator, and that existing algorithms use to detect whether a video is synthetic or real [29]. + +An overview of our algorithm for avatar fingerprinting is shown in Figure 3. To capture expressions, we extract the relative position of facial landmarks over time from the input video, as shown in Figure 3 (Section 5.1). We learn to project these temporal signatures onto a dynamic identity embedding in which features belonging to the same driving identity are close to each other regardless of the target identity, i.e., independently of appearance (Figure 1). To learn this embedding we train a 3DCNN (where the + +![](images/b93e19b071ac1ecc9fc948c81a089ccf1213e326648c1c47d17fbe0d1e9245e9.jpg) +Fig. 3: We extract landmarks from the frames of a talking-head clip, compute their normalized pairwise distances, and concatenate the frame-wise features. We then learn an identity embedding using a loss that pulls closer features of videos driven by the same identity and pushes away those driven by others. $\mathrm{ID}_i\rightarrow \mathrm{ID}_j$ indicates a video that looks like identity $j$ (the "target" identity), and is driven by identity $i$ . + +third dimension spans video frames—a temporal CNN) with a novel contrastive loss that pulls together all embedding vectors of synthetic videos driven by an individual, while pushing away the embedding vectors of videos driven by all others individuals (Section 5.2). More implementation details are in Section 5.3. + +# 5.1 Dynamic Facial Identity Features + +Our first step is to extract temporal features that summarize short segments of the video we wish to fingerprint. We identify the following guiding principles for the extracted features. We would like features that: + +1. have minimal dependency on the appearance of the face in the video (that is, the target identity), +2. reflect the dynamics of the expressions, and +3. capture subtle expressions. + +One choice could be per-frame 3DMM features [14]: a strategy also used by Cozzolino et al. to detect synthetic videos [20]. However, we empirically observe that 3DMM features are not sufficiently expressive, and do not satisfy desideratum 3 (see ablation experiments in Section 6). We observe a similar behavior for action units [9]. Facial landmarks [32] address this issue, but are sensitive to the shape of the face in the video, and thus to the target identity. + +To leverage the expressiveness of facial landmarks while abstracting from the underlying facial shape, we compute the pairwise normalized Euclidean distances between each pair of landmarks of a frame. We concatenate these distances into a single vector for the frame, $\mathbf{d}_f$ . A subset of the facial landmarks and distances are shown in Figure 3. + +We then break the input video into clips, which are sequences consisting of F frames and offset by one frame (e.g., [1,F], [2,F+1], etc.), and concatenate the vectors from all the frames in each clip. Using the change in the relative position of the landmarks over a short period of time (the length of a clip) allows us to capture temporal dynamics with minimal dependence on the absolute position of each landmark, i.e., independently of the shape of the face. + +We show empirically that our features are a good representation for our task, by comparing against alternative choices for input features such as 3DMM (Section 6). + +# 5.2 Dynamic Identity Embedding Contrastive Loss + +While the features described in Section 5.1 are designed to extract low-level motion dynamics, they cannot be used directly to disambiguate two target videos based on their driving identities. We tackle this problem by learning a dynamic identity embedding, a space where videos driven by one subject map to points that are close to each other and far from the videos driven by anybody else. + +Specifically, we use a temporal 3DCNN to extract an embedding vector from a clip, which, as described before, is a short segment of an input video. To train the network we use a dataset of synthetic videos driven by different identities. We denote as $\mathcal{C}_{\mathrm{ID}_1\to \mathrm{ID}_2}^k (t)$ the embedding produced by the network for the clip starting at time $t$ in the $k$ -th video, of a target identity $\mathrm{ID}_2$ driven by identity $\mathrm{ID}_1$ . As stated above, we have two main objectives, which we capture with the following terms in our proposed loss function. + +We Want to Pull Together All the Videos Driven by $ID_{1}$ . To achieve this, we define the following term: + +$$ +\mathrm {N} _ {j, \mathrm {I D} _ {1}, \mathrm {I D} _ {2}} (t) = \sum_ {\mathrm {I D} _ {l}, k} \max _ {n} s \left(\mathcal {C} _ {\mathrm {I D} _ {1} \rightarrow \mathrm {I D} _ {2}} ^ {j} (t), \mathcal {C} _ {\mathrm {I D} _ {1} \rightarrow \mathrm {I D} _ {l}} ^ {k} (n)\right), \tag {1} +$$ + +where $s(\cdot, \cdot) = e^{-\|\cdot\|^{2}}$ is a similarity metric. Intuitively, Equation 1 takes two videos, $j$ and $k$ , both driven by $\mathrm{ID}_1$ . Given a clip starting at time $t$ in the first video, it looks for the most similar clip in the second video. Since the driving identity is the same for both videos, Equation 1 encourages an embedding where clips that capture a similar expression are closer to each other. Equation 1 is high even if only one clip from video $k$ has a similar temporal signature to $\mathcal{C}_{\mathrm{ID}_1 \to \mathrm{ID}_2}^j(t)$ . That is because even just one occurrence of the same expression is evidence that the driving identity may be the same. Of course, other driving identities may use similar expressions and we address that with the loss term described below. Additionally, we note that $k$ spans the set of all videos driven by $\mathrm{ID}_1$ , and $\mathrm{ID}_l$ spans all identities, including $\mathrm{ID}_l = \mathrm{ID}_1$ and $\mathrm{ID}_l = \mathrm{ID}_2$ . + +We Want to Push Away Videos not Driven by $ID_{1}$ . We define the following term: + +$$ +\mathrm {Q} _ {j, \mathrm {I D} _ {1}, \mathrm {I D} _ {2}} (t) = \sum_ {\mathrm {I D} _ {l} \neq \mathrm {I D} _ {1}, k} \max _ {n} s \left(\mathcal {C} _ {\mathrm {I D} _ {1} \rightarrow \mathrm {I D} _ {2}} ^ {j} (t), \mathcal {C} _ {\mathrm {I D} _ {l} \rightarrow \mathrm {I D} _ {2}} ^ {k} (n)\right), \tag {2} +$$ + +where, similarly to Equation 1, we take a clip from video $j$ , and look for the most similar clip in video $k$ . This time the two videos share the same target identity, but are driven by different identities: we want all the videos driven by identities different from $\mathrm{ID}_1$ to be pushed away from those driven by $\mathrm{ID}_1$ , including videos where $\mathrm{ID}_1$ is the target identity. Note that $\mathrm{ID}_2$ spans all identities, including $\mathrm{ID}_2 = \mathrm{ID}_1$ and $\mathrm{ID}_2 = \mathrm{ID}_l$ . + +We Want to Rely on the Temporal Dynamics of the Videos Driven by $ID_{1}$ . Although we use a temporal CNN, the model could still learn to rely on static expressions, such as a snapshot of the person smiling, rather than the temporal progression of expression leading to, or following, the smile. To further encourage the model to learn from the temporal dynamics, we introduce an additional term: + +$$ +\mathrm {R} _ {j, \mathrm {I D} _ {1}, \mathrm {I D} _ {2}} (t) = \sum_ {\mathrm {I D} _ {l}, k} \max _ {n} s \left(\mathcal {C} _ {\mathrm {I D} _ {1} \rightarrow \mathrm {I D} _ {2}} ^ {j} (t), \tilde {\mathcal {C}} _ {\mathrm {I D} _ {1} \rightarrow \mathrm {I D} _ {l}} ^ {k} (n)\right), \tag {3} +$$ + +where $\widetilde{\mathcal{C}}_{\mathrm{ID}_1\to \mathrm{ID}_l}^k$ denotes a version of the clip $\mathcal{C}_{\mathrm{ID}_1\to \mathrm{ID}_l}^k$ from Equation 1 with randomly shuffled frame ordering. We want such time-shuffled versions of the clips driven by $\mathrm{ID}_1$ to be pushed away from the pristine clips driven of $\mathrm{ID}_1$ . Effectively, this means that the driving identity of the time-shuffled clips is regarded as different from $\mathrm{ID}_1$ . In other words, we want to pull together video clips in the learned embedding space only when the temporal facial dynamics are characteristic of $\mathrm{ID}_1$ . We further show the importance of this term in Section 6.3. + +Combining Equations 1, 2, and 3, we write the probability that the embedding vector $\mathcal{C}_{\mathrm{ID}_1\to \mathrm{ID}_2}^j (t)$ lies close to the embedding vectors for all video clips driven by $\mathrm{ID}_1$ and far from all the videos driven by others as + +$$ +p _ {j, \mathrm {I D} _ {1}, \mathrm {I D} _ {2}} (t) = \frac {\mathrm {N} _ {j , \mathrm {I D} _ {1} , \mathrm {I D} _ {2}} (t)}{\mathrm {N} _ {j , \mathrm {I D} _ {1} , \mathrm {I D} _ {2}} (t) + \mathrm {Q} _ {j , \mathrm {I D} _ {1} , \mathrm {I D} _ {2}} (t) + \mathrm {R} _ {j , \mathrm {I D} _ {1} , \mathrm {I D} _ {2}} (t)}, \tag {4} +$$ + +and the complete loss term as + +$$ +\mathcal {L} = \sum_ {j, \mathrm {I D} _ {1}, \mathrm {I D} _ {2}, t} - \log \left(p _ {j, \mathrm {I D} _ {1}, \mathrm {I D} _ {2}} (t)\right). \tag {5} +$$ + +# 5.3 Implementation + +Parameter Choices. To extract the per-frame dynamic facial identity features $\mathbf{d}_f$ , we detect 126 facial landmarks for each frame [32], and compute the per-frame normalized pairwise Euclidean distances between these landmarks. The clip duration is set to 71 frames. We find that this is sufficient to capture the facial dynamics that are meaningful for avatar fingerprinting, while also maintaining a good trade-off between speed and accuracy. We also experiment with shorter-duration video clips (see Supplement). The input tensor to the temporal CNN is obtained by concatenating $\mathbf{d}_f$ across 71 frames. In each batch, we include 8 unique identities. For each identity $\mathrm{ID}_i$ , the pull term (Equation 1) comprises 16 clips: 8 are self-reenactments, randomly sampled from the full set, and the remaining are cross-reenactments with $\mathrm{ID}_i$ as the driving identity. This allows the neural network to pull together videos based purely on the facial motion, regardless of the appearance of the video. The push term (Equation 2) for $\mathrm{ID}_i$ consists of clips with the remaining 7 identities in the batch serving as driving identities (8 clips per driving identity), as well as the time-shuffled self-reenactments of $\mathrm{ID}_i$ (Equation 3). Additional training details can be found in the Supplement. + +Training, Validation, and Testing Datasets. Of the 161 total identities (pooling together the identities from our dataset, RAVDESS, and CREMA-D, see Section 4.2), we reserve 35 for testing, 14 for validation, and 112 for training. We ensure that there are no cross-set cross-reenactments: that is, identities in the training set only drive other training-set identities (and similarly for the validation and test sets). This allows us to test the models on facial appearances and expressions that were not seen during training. To evaluate the generalization of trained models to new generators, we train our network on videos generated for the training set identities using one generator, and evaluate on the synthetic videos of test-set identities synthesized using remaining two generators. + +![](images/2856cf1c7e156858dd5e3054c4b3c1d882ad6ad1da1715cf87f1feae9ff52e88.jpg) +$\mathrm{ID}_6 \rightarrow \mathrm{ID}_6$ d=0.84 + +![](images/ec4805d1f008fd55e310f48309f9773b78166d3cc9be3f108d15495835f6f0ab.jpg) +$\mathrm{ID}_6 \rightarrow \mathrm{ID}_1$ d=0.95 + +![](images/7d52bc44a8a860f9b99c6870a2f92ababe5ce647a9e786f82496f24f66de7aaf.jpg) +$\mathrm{ID}_1 \rightarrow \mathrm{ID}_6$ d=1.58 + +![](images/7f4e3e07804a221975f738c25106a4f6cd1d17add076c2b0a06eb57bb735375f.jpg) +$\mathrm{ID}_3 \rightarrow \mathrm{ID}_6$ d=2.11 + +![](images/f78e140b5a9a6085818b9e3867cf826e6984a2bbe7bacd59fda9f585ba8ca1bc.jpg) +$\mathrm{ID}_2 \rightarrow \mathrm{ID}_2$ d=0.589 +Fig. 4: Animated figure. Open in a media-enabled viewer like Adobe Reader and click on the inset. Our embeddings capture the dynamics of an expression, rather than the appearance of the face. For each row, we pick a reference identity. The green box indicates reenactments driven by the reference identity, the red and blue are cross-reenactments of the reference identity. We compute the average distance of each clip shown here against all other clips driven by the reference identity. The average distance to the other clips of the reference identity is consistent for a given motion, and lower (better) when the reference identity is driving as compared to the cross-reenactments that use the reference identity as target. Here, we show videos generated by face-vid2vid [60], and use the embedding vectors predicted by the model trained on the same generator (see Figure 6 generalization to new generators not seen during training). + +![](images/0aa8dcad2786fb3228b0e2ba6b47264d30154270e9a53ba93090fb1fcbfefb82.jpg) +$\mathrm{ID}_2 \rightarrow \mathrm{ID}_7$ d=0.961 + +![](images/981c1d7f1581ec1c50eff2f0ba39936413500444ef06494e776cff9242966e9e.jpg) +$\mathrm{ID}_7 \rightarrow \mathrm{ID}_2$ d=2.042 + +![](images/d089d3e5d94a267d087796094d566c27637b70e4d8ebca94c70d5da4f051a52a.jpg) +$\mathrm{ID}_8 \rightarrow \mathrm{ID}_2$ d=1.787 + +# 6 Evaluation + +We thoroughly evaluate our algorithm both qualitatively and quantitatively. Our algorithm outperforms reasonable baselines (Section 6.1), it generalizes to generators not seen in training (Section 6.2), and it is robust to video compression (Supplement). We also perform a number of ablation studies to analyze our design choices: our input features, the importance of the time-shuffling term in the loss function (Section 6.3), and the impact of clip duration (Supplement). + +We begin by evaluating qualitatively our method's ability to extract embedding vectors based on the driving identity. Figure 4 shows a set of self- and cross-reenacted clips (please view the animation in a media-enabled viewer, such as Adobe Acrobat). For each row, we take one identity as reference and we compute the embedding vectors of clips that use it both as the driving and the target identity. We then compute the average Euclidean distance of the resulting embedding vectors against those of the self-reenactments by the same reference identity. We note that the average distance $d$ is lower when the driving identity matches the reference identity (first two columns). We also note that the distances between the clips in the first two columns are similar: this confirms that the distance is a function of the facial motion, rather than the facial appearance. When the driving identity changes, the average distance increases, even if the target identity matches the reference identity, which is precisely our goal. More results are in the Supplement. + +![](images/478bf99b11578ecc7a6ae33c0f8d5184a9db71068e5815d85d685da73eba7339.jpg) +Fig.5: ROC curves and AUC values for our method and two baselines: Agarwal et al. [9] and ID-Reveal [20]. Each sub-plot shows the results on our test set for each of the three talking-head generators: face-vid2vid [60], LIA [62], and TPS [34]. + +To evaluate our approach more formally, we use the 35 unique test-set identities that are not used as driving or target identities in the training set (Section 5.3). One at a time, we treat each identity $\mathrm{ID}_i$ as target and synthesize cross-reenactments using all the remaining identities as drivers. This is the set of "unauthorized" synthetic videos for $\mathrm{ID}_i$ . The self-reenacted samples for $\mathrm{ID}_i$ form the "authorized" set. Note that there are several self-reenacted videos of $\mathrm{ID}_i$ , one per original video of $\mathrm{ID}_i$ . + +For each target identity $\mathrm{ID}_i$ , we extract the dynamic identity embedding vector of all the clips in the pool of its self- and cross-reenacted videos, and compute their Euclidean distances. That is, for clip $k$ we compute + +$$ +d \left(\mathcal {C} _ {\mathrm {I D} _ {i} \rightarrow \mathrm {I D} _ {i}} ^ {k}, \mathcal {C} _ {\mathrm {I D} _ {i} \rightarrow \mathrm {I D} _ {i}} ^ {l}\right), \forall l \neq k, \text {a n d} +$$ + +$$ +\begin{array}{l}u \left(\mathrm {I D} _ {i} \rightarrow \mathrm {I D} _ {i}, \mathrm {C I D} _ {i} \rightarrow \mathrm {I D} _ {i}\right), \quad \forall i, \quad \forall i \neq j.\end{array}\tag {6} +$$ + +We threshold these distances for each target identity to get an ROC curve, and average across the target identities to get the overall area under the curve (AUC). We note that this AUC measures one model's ability to classify a synthetic video as self-reenactment or as cross-reenacted. We conduct further analysis of our model's ability to classify other categories of videos—such as, evaluating AUC on same-utterance self- vs. cross-reenactments or on scripted vs. free-form monologues—in the Supplement. + +# 6.1 Comparisons with Existing Methods + +Avatar fingerprinting is a novel task, and no existing methods directly address it. The closest related works aim at detecting real versus synthetic media. As discussed in Section 2, some of these detectors learn identity-specific features such as facial expressions and head poses [9], or facial shapes and motion [20] and can serve as baselines for the task of avatar fingerprinting with some adaptation. The work by Agarwal et al. trains a model to detect synthetic videos of a specific identity [9]. To adapt it to our task, we train 35 different models, one for each identity in the evaluation, by splitting the corresponding original videos into two subsets. We then test each model on the self- and cross-reenactment videos of the corresponding identity. ID-Reveal, trained on a large-scale dataset, learns an embedding space where real videos of a specific identity are grouped together [20]. Since it shows good generalization to new identities, for the task of synthetic media detection, we directly use the pre-trained model on our data to detect, once again, self- versus cross-reenactment. Figure 5 shows the ROC curves for our method compared to these baselines, on three face-reenactment generators (face-vid2vid [60], + +![](images/6134ed937975c61d817d7a7c19087cca5cc2543f7f5fb83ff84fd9586f142441.jpg) +Fig. 6: Generalization to new generators. To study the robustness to new talking-head generators, we train three versions of our model on three different generators and test on all three. + +![](images/f8d69fd2100d41683e7984a489e0ff7557bb23b487ec3845844064087ebd3962.jpg) + +![](images/1432e080d0917739ae6197e7975d0ced0d0f909e9a0f7a8ff6b4e9afd15b4c9b.jpg) + +LIA [62], and TPS [34]). Our method (AUC=0.868 on face-vid2vid) outperforms by a wide margin both ID-Reveal (AUC=0.720 on face-vid2vid), and the method by Agarwal et al. (AUC=0.726 on face-vid2vid). We also note that, unlike ID-Reveal and our method, Agarwal et al. uses a different model per identity. + +# 6.2 Generalization to New Generators + +For an avatar fingerprinting model algorithm to be broadly applicable, generalization to new talking-head generators that are not seen in training is crucial. Since our dataset contains videos synthesized by three different generators, we can train three models, one with each generator, and test these models on all three generators. Figure 6 shows the resulting ROC curves and AUC values: the overlap of the curves and similar AUC values in each subplot confirms that our method generalizes well to new generators. + +# 6.3 Ablation Study + +Our method outperforms existing baselines by introducing two novel components: the dynamic facial identity features, which capture the facial dynamics in a compact and expressive way, and the loss function, which defines the shape of the identity embedding. Here we study the contribution of each, using the face-vid2vid generator for training and testing. We evaluate the contribution of our dynamic facial identity features by swapping them with 3DMM features [13], a popular choice to capture facial dynamics. Since we use a temporal CNN backbone similar to the one from ID-Reveal, for this ablation we use the loss function proposed in their original paper [20]. We re-train the same network using our features and observe a jump from 0.718 to 0.754 in terms of AUC. Upon inspection we notice that the 3DMM features tend to over-smooth the facial motion, and are unable to capture subtle dynamics that prove critical to avatar fingerprinting, and which our features capture. We also evaluate the contribution of our dynamic identity embedding loss and observe a further improvement (AUC 0.868). With our loss formulation, the advantage of $\mathbf{R}_{j,\mathrm{ID}_1,\mathrm{ID}_2}(t)$ in Eq. 5 is also evident when compared against a model trained without this term (AUC 0.850). Table 2 summarizes this ablation study. Additional experiments in the Supplement show the impact of $F$ , performance on scripted vs. free-from monologues, and robustness to video distortions. + +# 6.4 Limitations + +Our algorithm is less discriminative of subjects that are less emotive and more neutral. In the future, relying on more granular dynamic signatures that can extract micro + +
Input FeaturesLossAUC
3DMMID-Reveal rec. loss [20]0.718
Landmark distancesID-Reveal rec. loss [20]0.754
Landmark distancesOur loss without Rj, ID1, ID2(t)0.850
Landmark distancesOur loss with Rj, ID1, ID2(t)0.868
+ +Table 2: Ablation study showing the importance of our input features and loss function design. + +expressions can help alleviate this. The performance of our method degrades when expressions that are critical to verifying the driving identity are not captured by the synthetic portrait generator. Lastly, our dataset currently features only one style of interaction: one-on-one conversations. We plan to expand to other conversation styles, such as one-way speeches, in future. + +# 7 Societal Impact + +We acknowledge the societal importance of introducing guardrails when it comes to the use of talking-head generation technology. We present this work as a step towards trustworthy use of such technologies. Nevertheless, our work could be misconstrued as having solved the problem and inadvertently accelerate the unhindered adoption of synthetic talking-head technology. We do not advocate for this. Instead, we stress that this is the first work on this topic and underscore the importance of further research. + +# 8 Conclusions + +Highly photo-real synthetic talking-head generators are becoming increasingly beneficial to applications such as video conferencing and AR/VR-based remote interactions. This trend raises the important new research question of how best to also ensure their safe use in such scenarios. To this end, we investigate the new problem of avatar fingerprinting, to authenticate legitimate talking-heads created by authorized users. We leverage the fact that driving individuals have uniquely identifying dynamic motion signatures, which are also preserved in the synthetic videos that they drive. Since none exists, we contribute a new large-scale dataset carefully designed for avatar fingerprinting and related tasks. We hope that our work lays the foundation for further research. + +Acknowledgements. We would like to thank the data-capture participants, and Desiree Luong, Woody Luong, and Josh Holland for their help with Figure 2. We acknowledge David Taubenheim for the voiceover in the demo video, and Abhishek Badki for help with the training infrastructure. We thank Joohwan Kim, Rachel Brown, Anjul Patney, Ben Boudaoud, Josef Spjut, Saori Kaji, Nikki Pope, and Kai Pong for their help with putting together the data capture protocols, informed consent form, photo release form, and agreements for data governance and third-party data sharing. Koki Nagano, Ekta Prashnani, and David Luebke were partially supported by DARPA's Semantic Forensics (SemaFor) contract (HR0011-20-3-0005). This research was funded, in part, by DARPA's Semantic Forensics (SemaFor) contract HR0011-20-3-0005. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government. + +Distribution Statement "A" (Approved for Public Release, Distribution Unlimited). + +# References + +1. Apple's vision pro. https://www.apple.com/apple-vision-pro/, accessed: 2024-03-06 +2. Heygen. https://www.heygen.com, accessed: 2023-11-16 +3. Microsoft teams mesh. https://www.microsoft.com/en-us/microsoft-teams/microsoft-mesh, accessed: 2024-03-06 +4. Myheritage. https://www.myheritage.com, accessed: 2023-11-16 +5. Nvidia's maxine. https://developer.nvidia.com/maxine, accessed: 2024-03-06 +6. Agarwal, S., El-Gaaly, T., Farid, H., Lim, S.N.: Detecting deep-fake videos from appearance and behavior. 2020 IEEE International Workshop on Information Forensics and Security (WIFS) pp. 1-6 (2020) +7. Agarwal, S., Farid, H.: Detecting deep-fake videos from aural and oral dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2021) +8. Agarwal, S., Farid, H., Fried, O., Agrawala, M.: Detecting deep-fake videos from phonemeviseme mismatches. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2020) +9. Agarwal, S., Farid, H., Gu, Y., He, M., Nagano, K., Li, H.: Protecting world leaders against deep fakes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2019) +0. Agarwal, S., Hu, L., Ng, E., Darrell, T., Li, H., Rohrbach, A.: Watch those words: Video falsification detection using word-conditioned facial motion. In: IEEE Winter Conference on Applications of Computer Vision (WACV) (2023) +1. Albright, M., McCloskey, S.: Source generator attribution via inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2019) +2. Baluja, S.: Hiding images in plain sight: Deep steganography. In: Advances in Neural Information Processing Systems (NeurIPS) (2017) +3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of SIGGRAPH (1999) +4. Blanz, V., Vetter, T.: Face recognition based on fitting a 3D morphable model. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2003) +5. Boháček, M., Farid, H.: Protecting world leaders against deep fakes using facial, gestural, and vocal mannerisms. Proceedings of the national academy of Sciences (2022) +6. Cao, H., Cooper, D.G., Keutmann, M.K., Gur, R.C., Nenkova, A., Verma, R.: CREMA-D: Crowd-sourced emotional multimodal actors dataset. IEEE Transactions on Affective Computing (2014) +7. Chung, J.S., Nagrani, A., Zisserman, A.: Voxceleb2: Deep speaker recognition. In: INTER-SPEECH (2018) +8. Cozzolino, D., Nießner, M., Verdoliva, L.: Audio-visual person-of-interest deepfake detection (2022) +9. Cozzolino, D., Poggi, G., Verdoliva, L.: Extracting camera-based fingerprints for video forensics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019) +20. Cozzolino, D., Rössler, A., Thies, J., Nießner, M., Verdoliva, L.: ID-Reveal: Identity-aware DeepFake video detection. In: IEEE International Conference on Computer Vision (ICCV) (2021) +21. Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., Canton Ferrer, C.: The deepfake detection challenge dataset. arXiv preprint arXiv:2006.07397 (2020) + +22. Dong, X., Bao, J., Chen, D., Zhang, T., Zhang, W., Yu, N., Chen, D., Wen, F., Guo, B.: Protecting celebrities from deepfake with identity consistency transformer. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +23. Drobyshev, N., Chelishev, J., Khakhulin, T., Ivakhnenko, A., Lempitsky, V., Zakharov, E.: MegaPortraits: One-shot megapixel neural head avatars (2022) +24. Fox, G., Liu, W., Kim, H., Seidel, H.P., Elgharib, M., Theobalt, C.: VideoForensicsHQ: Detecting high-quality manipulated face videos. In: IEEE International Conference on Multimedia and Expo (2021) +25. Fridrich, J.: Steganography in Digital Media: Principles, Algorithms, and Applications. Cambridge University Press (2009) +26. Ge, S., Lin, F., Li, C., Zhang, D., Wang, W., Zeng, D.: Deepfake video detection via predictive representation learning. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) (2022) +27. Hadsell, R., Chopra, S., LeCun, Y.: Dimensionality reduction by learning an invariant mapping. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2006) +28. Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M.: Lips don't lie: A generalisable and robust approach to face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +29. Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M.: Lips don't lie: A generalisable and robust approach to face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +30. He, Y., Gan, B., Chen, S., Zhou, Y., Yin, G., Song, L., Sheng, L., Shao, J., Liu, Z.: Forgerynet: A versatile benchmark for comprehensive forgery analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4360-4369 (2021) +31. Hill, H., Johnston, A.: Categorizing sex and identity from the biological motion of faces. Current Biology (2001) +32. Honari, S., Molchanov, P., Tyree, S., Vincent, P., Pal, C., Kautz, J.: Improving landmark localization with semi-supervised learning. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +33. Hong, F.T., Zhang, L., Shen, L., Xu, D.: Depth-aware generative adversarial network for talking head video generation (2022) +34. Jian Zhao, H.Z.: Thin-plate spline motion model for image animation (2022) +35. Jiang, L., Li, R., Wu, W., Qian, C., Loy, C.C.: Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +36. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +37. Khakhulin, T., Sklyarova, V., Lempitsky, V., Zakharov, E.: Realistic one-shot mesh-based head avatars. In: European Conference on Computer Vision (ECCV) (2022) +38. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +39. Kim, H., Elgharib, M., Zollhöfer, M., Seidel, H.P., Beeler, T., Richardt, C., Theobalt, C.: Neural style-preserving visual dubbing. ACM Transactions on Graphics (ToG) (2019) +40. Kim, H., Garrido, P., Tewari, A., Xu, W., Thies, J., Niessner, M., Pérez, P., Richardt, C., Zollhöfer, M., Theobalt, C.: Deep video portraits. ACM Transactions on Graphics (ToG) (2018) +41. Knappmeyer, B., Thornton, I., Bulthoff, H.: Facial motion can determine facial identity. Journal of Vision (2001) +42. Kwon, P., You, J., Nam, G., Park, S., Chae, G.: Kodf: A large-scale korean deepfake detection dataset. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10744-10753 (2021) + +43. Li, J., Xie, H., Yu, L., Zhang, Y.: Wavelet-enhanced weakly supervised local feature learning for face forgery detection. In: Proceedings of the 30th ACM International Conference on Multimedia (2022) +44. Li, Y., Sun, P., Qi, H., Lyu, S.: Celeb-DF: A large-scale challenging dataset for DeepFake forensics. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +45. Liu, B., Liu, B., Ding, M., Zhu, T., Yu, X.: Ti2net: Temporal identity inconsistency network for deepfake detection. In: IEEE Winter Conference on Applications of Computer Vision (WACV) (2023) +46. Livingstone, S.R., Russo, F.A.: The ryerson audio-visual database of emotional speech and song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one (2018) +47. Luo, X., Zhan, R., Chang, H., Yang, F., Milanfar, P.: Distortion agnostic deep watermarking. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +48. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +49. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit Warping for Animation with Image Sets. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) +50. Munir, S., Batool, B., Shafiq, Z., Srinivasan, P., Zaffar, F.: Through the looking glass: Learning to attribute synthetic text generated by language models. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume (2021) +51. O'Toole, A.J., Roark, D.A., Abdi, H.: Recognizing moving faces: A psychological and neural synthesis. Trends in Cognitive Sciences (2002) +52. Passos, L.A., Jodas, D., da Costa, K.A., Junior, L.A.S., Rodrigues, D., Del Ser, J., Camacho, D., Papa, J.P.: A review of deep learning-based approaches for deepfake content detection. arXiv preprint arXiv:2202.06095 (2022) +53. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: FaceForensics++: Learning to detect manipulated facial images. In: IEEE International Conference on Computer Vision (ICCV) (2019) +54. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: Faceforensics++: Learning to detect manipulated facial images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1-11 (2019) +55. Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: A unified embedding for face recognition and clustering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015) +56. Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) +57. Sun, K., Yao, T., Chen, S., Ding, S., Li, J., Ji, R.: Dual contrastive learning for general face forgery detection (2022) +58. Tancik, M., Mildenhall, B., Ng, R.: Stegastamp: Invisible hyperlinks in physical photographs. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +59. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: European Conference on Computer Vision (ECCV) (2020) +60. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +61. Wang, X., Han, X., Huang, W., Dong, D., Scott, M.R.: Multi-similarity loss with general pair weighting for deep metric learning. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +62. Wang, Y., Yang, D., Bremond, F., Dantcheva, A.: Latent image animator: Learning to animate images via latent space navigation. In: International Conference on Learning Representations (ICLR) (2022) + +63. Yacoob, Y.: Gan-scanner: A detector for faces of stylegan+ (2021), https://github.com/yaseryacoob/GAN-Scanner +64. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: IEEE International Conference on Computer Vision (ICCV) (2021) +65. Yu, N., Skripniuk, V., Chen, D., Davis, L., Fritz, M.: Responsible disclosure of generative models using scalable fingerprinting. In: International Conference on Learning Representations (ICLR) (2022) +66. Zakharov, E., Ivakhnenko, A., Shysheya, A., Lempitsky, V.: Fast bi-layer neural synthesis of one-shot realistic head avatars. In: European Conference on Computer Vision (ECCV) (2020) +67. Zakharov, E., Shysheya, A., Burkov, E., Lempitsky, V.: Few-shot adversarial learning of realistic neural talking head models. In: IEEE International Conference on Computer Vision (ICCV) (2019) +68. Zheng, Y., Bao, J., Chen, D., Zeng, M., Wen, F.: Exploring temporal coherence for more general video face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) \ No newline at end of file diff --git a/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/images.zip b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..743b45250a73ecc57aa8cb23fe1fd8f17d421836 --- /dev/null +++ b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07ac21a21f601798f55a325cedfe8a8ab3c5d708d50eedef672c4c1212466122 +size 324558 diff --git a/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/layout.json b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e94267108a0fa8a553b7ccc8d3614192660908cf --- /dev/null +++ b/2024/Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos/layout.json @@ -0,0 +1,9756 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 111, + 445, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 111, + 445, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 111, + 445, + 148 + ], + "type": "text", + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 167, + 472, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 167, + 472, + 179 + ], + "spans": [ + { + "bbox": [ + 141, + 167, + 472, + 179 + ], + "type": "text", + "content": "Ekta Prashnani, Koki Nagano, Shalini De Mello, David Luebke, and Orazio Gallo" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 288, + 188, + 324, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 188, + 324, + 198 + ], + "spans": [ + { + "bbox": [ + 288, + 188, + 324, + 198 + ], + "type": "text", + "content": "NVIDIA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 200, + 468, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 200, + 468, + 210 + ], + "spans": [ + { + "bbox": [ + 147, + 200, + 468, + 210 + ], + "type": "text", + "content": "{eprashnani, knagano, shalinig, dluebke, ogallo}@nvidia.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 235, + 455, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 235, + 455, + 392 + ], + "spans": [ + { + "bbox": [ + 159, + 235, + 455, + 392 + ], + "type": "text", + "content": "Abstract. Modern avatar generators allow anyone to synthesize photorealistic real-time talking avatars, ushering in a new era of avatar-based human communication, such as with immersive AR/VR interactions or videoconferencing with limited bandwidths. Their safe adoption, however, requires a mechanism to verify if the rendered avatar is trustworthy: does it use the appearance of an individual without their consent? We term this task avatar fingerprinting. To tackle it, we first introduce a large-scale dataset of real and synthetic videos of people interacting on a video call, where the synthetic videos are generated using the facial appearance of one person and the expressions of another. We verify the identity driving the expressions in a synthetic video, by learning motion signatures that are independent of the facial appearance shown. Our solution, the first in this space, achieves an average AUC of 0.85. Critical to its practical use, it also generalizes to new generators never seen in training (average AUC of 0.83). The proposed dataset and other resources can be found at: https://research.nvidia.com/labs/nxp/avatar-fingerprinting/." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 401, + 453, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 401, + 453, + 412 + ], + "spans": [ + { + "bbox": [ + 160, + 401, + 453, + 412 + ], + "type": "text", + "content": "Keywords: Synthetic Media Verification " + }, + { + "bbox": [ + 160, + 401, + 453, + 412 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 401, + 453, + 412 + ], + "type": "text", + "content": " Avatar Generators " + }, + { + "bbox": [ + 160, + 401, + 453, + 412 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 401, + 453, + 412 + ], + "type": "text", + "content": " Avatar Attribution" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 434, + 218, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 218, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 218, + 446 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 596 + ], + "type": "text", + "content": "Recent digital avatar generators have fueled a myriad of computer vision and graphics applications, allowing anyone to synthesize real-time photorealistic personas. Major companies are now supporting avatar-driven remote interactions over immersive AR/VR (e.g. Meta's Pixel Codec Avatar [48], Apple's Vision Pro Persona [1]) or video conferencing (e.g. NVIDIA's MAXINE [5], Microsoft's Teams [3]), and selfie filters for altering and enhancing appearance (e.g. by Snap and Tiktok). While the avatar generation technology today is still young, the legitimate use of synthetic avatars will be ubiquitous in the future. Without proper guardrails, this poses a real risk of unauthorized use and large-scale spread of visual disinformation. To ensure the safe use of such a technology, the relevant question is no longer whether the content is \"real\" or not—since, by design, the videos and avatars are all synthetic—but rather, whether the synthetically-generated videos and avatars are \"trustworthy\" or not." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 597, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 483, + 666 + ], + "type": "text", + "content": "When video conferencing, for instance, a synthetic video portrait generator can be used to save valuable bandwidth by reconstructing a synthetic avatar of the sender at the receiver's end, using only a frame of the target identity and a compact representation of the speaker's facial motion. To ensure the authorized use of such synthetically-generated videos, we want to verify if the driving identity behind a synthetic video " + }, + { + "bbox": [ + 130, + 597, + 483, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{ID}_1" + }, + { + "bbox": [ + 130, + 597, + 483, + 666 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 597, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_2" + }, + { + "bbox": [ + 130, + 597, + 483, + 666 + ], + "type": "text", + "content": " in Figure 1) is authorized to control the likeness, or the appearance, of the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 114, + 433, + 252 + ], + "blocks": [ + { + "bbox": [ + 179, + 114, + 433, + 252 + ], + "lines": [ + { + "bbox": [ + 179, + 114, + 433, + 252 + ], + "spans": [ + { + "bbox": [ + 179, + 114, + 433, + 252 + ], + "type": "image", + "image_path": "4eae27ba7b367c955894eebea718d85ee1012e9264b5c7b84c556e11c7b33fba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 328 + ], + "type": "text", + "content": "Fig. 1: Talking-head avatar generators can synthesize realistic videos of a target identity from driving videos of different identities. Our method extracts appearance-agnostic temporal facial features and learns an embedding in which the synthetic videos driven by one identity fall close to each other and far from those driven by other identities, regardless of the appearance of the synthetic video. By comparing distances in the embedding space, we evaluate whether an avatar is driven by an authorized identity or not. During evaluation, we only rely on the synthetic videos as input, without requiring any prior knowledge about the driving identity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 481, + 366 + ], + "type": "text", + "content": "synthetic video (target identity in Figure 1). Crucially, we want to only leverage the synthetic video avatars to do so. We call this novel task avatar fingerprinting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 367, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 447 + ], + "type": "text", + "content": "We leverage a simple but fundamental observation: facial motions tend to be idiosyncratic, that is, individuals talk and emote in unique ways. For instance, someone may raise one of her eyebrows more than the other, or smile more while talking. These \"dynamic identity signatures\" [51] have been shown to carry enough information for humans to recognize other individuals, even when the physical appearance of their face is altered [31, 41, 51]. This makes them attractive for our task, as they can be derived solely from the driving identity of a talking-head video, regardless of the appearance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 562 + ], + "type": "text", + "content": "Fortunately, modern avatar generators are becoming increasingly accurate at capturing the facial motion of a person and rendering it onto a target identity. As a first solution to avatar fingerprinting, we then propose to estimate a dynamic identity signature for an identity from the synthesized videos they drove—regardless of the target identities shown. Specifically, we extract facial landmarks and their temporal dynamics from the video. We then introduce a novel contrastive loss to learn a dynamic identity embedding, a space where the dynamic identity signatures of a driving identity across multiple videos and target identities are close to each other, and far from those of other driving identities (see Figure 1). We show that this approach, albeit straightforward, is robust and generalizes to generators not seen in training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 563, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 665 + ], + "type": "text", + "content": "Avatar fingerprinting is a new task, and no existing datasets serve its training and validation requirements. There are two key requirements from a dataset. First, we need videos of multiple subjects delivering both scripted and free-form monologues, captured under realistic conditions, such as varying video quality and gaze direction. This allows us to assess if a model leverages talking styles, and not the specific choice of words. Second, to evaluate if the model learns to extract dynamic identity signatures effectively from synthetic videos, we need synthetic talking-head videos for the case in which the driving and target identities are different (cross-reenactment), and that in which they match (self-reenactment). Unfortunately, existing datasets of real videos" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "E. Prashnani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 482, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 482, + 231 + ], + "type": "text", + "content": "only fulfill a subset of these requirements (Table 1 and Section 4). Further, no existing dataset of synthetic videos contain both self- and cross-reenactments per subject or use the state-of-the-art talking-head generation technology. To foster research in this new domain, then, we introduce the NVIDIA Facial Reenactment (NVFAIR) dataset, containing real and synthetic face reenactment talking-head videos (Figure 2). Our dataset, which includes ethnically diverse subjects, provides " + }, + { + "bbox": [ + 134, + 116, + 482, + 231 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 134, + 116, + 482, + 231 + ], + "type": "text", + "content": " more synthetic facial reenactments than the second largest dataset, for a total of over 650,000 synthetic videos. It is also the only one using multiple state-of-the-art face reenactment generators [34,60,62], or to provide cross-reenactments driven by all identities, which is critical for training and evaluating avatar fingerprinting algorithms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 237, + 186, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 186, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 186, + 248 + ], + "type": "text", + "content": "In summary:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 255, + 479, + 355 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 255, + 479, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 479, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 479, + 287 + ], + "type": "text", + "content": "- We introduce the novel task of avatar fingerprinting, which focuses on verifying the driving identity of synthetic talking-head avatars, rather than classifying them as real or synthetic (by design, all inputs to our model are synthetic)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 289, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 479, + 320 + ], + "type": "text", + "content": "- We release the first large-scale dataset of subjects delivering scripted and natural monologues, complete with self- and cross-reenactment videos synthesized with multiple state-of-the-art generators." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 323, + 479, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 479, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 479, + 355 + ], + "type": "text", + "content": "- We propose a solution for this novel task in the context of video conferencing by extracting person-specific motion signatures, and demonstrate its robustness to various distortions and avatar generators not seen in training." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 373, + 224, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 224, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 224, + 385 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 396, + 483, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 396, + 483, + 475 + ], + "spans": [ + { + "bbox": [ + 134, + 396, + 483, + 475 + ], + "type": "text", + "content": "Our proposed avatar fingerprinting task aims to verify authorized use of synthetic avatars: this is fundamentally a different problem than traditional forensics research where one aims to detect synthetic media (e.g. deepfake detection) or actively mark synthetic content. In our case, by design, the content being evaluated is always synthetic: we aim to evaluate its authorized use. Since this is a novel task, no methods exist to directly address it. No methods currently exist to directly address this novel task. Here we discuss the related areas of research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 479, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 479, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 134, + 479, + 482, + 616 + ], + "type": "text", + "content": "Learning-based Attribution of Synthetic Media. Learning-based approaches have been used to identify the origin of synthetic media, or to determine if it has been manipulated or altered in some way. Previous work used a pre-trained GAN generator to attribute a synthesized image to its generator via GAN inversion by leveraging the fact that a real image is less invertible [11, 36, 63]. Yet other works focus on attributing other forms of synthetic media, such as text [50]. In contrast, our focus is to attribute a talking-head avatar to the identity driving it, regardless of the appearance of the avatars. Some existing works learn fingerprints associated with cameras to determine whether a video is manipulated [19], or embed watermarks into images and videos [12, 25, 47, 58], which are also shown to transfer to GAN-generated images [64]. Subsequent research introduced a watermark-based conditional GANs for scalable fingerprinting [65]. Our method, in contrast, is a passive technique that does not rely on active watermarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 619, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 619, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 619, + 482, + 665 + ], + "type": "text", + "content": "Deepfake Detection Based on Identity-Specific Features. Deepfake detection (\"is a video synthetic?\") and avatar fingerprinting (\"whose identity is used to generate this synthetic video?\") are fundamentally different tasks. Most existing solutions for deepfake detection train a real-vs-synthetic classifier [26, 28, 43, 52, 57, 68], and therefore" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 332 + ], + "type": "text", + "content": "cannot be adapted to avatar fingerprinting (where all inputs are synthetic). However, a specific class of detectors leverage identity-specific features to detect synthetic videos by posing the detection problem as an identity-recognition problem. In our experiments, we evaluate some of these methods as baselines. Specifically, Agarwal et al. exploit person-specific patterns in facial expressions to detect fake videos [9]. ID-Reveal used facial shapes and motions encoded in a low-dimensional space of a 3D morphable model [13] to handle both face-swapping and face-reenactment deepfakes [20]. Other works explored soft-biometric approaches such as leveraging vocal mannerisms [15], phoneme-viseme consistencies [8], word-facial expression consistencies [6,10], and dynamics of ears [7]. While many of these works need person-specific training, previous works [6, 18, 20, 45] extended this idea to train a CNN-based detector using a large-scale in-the-wild video data [17] and variants of contrastive learning [27, 38, 55, 61]. Agarwal et al. combined static facial appearance using a facial recognition model and dynamic facial behaviors using a CNN, and showed that this approach is effective for detecting face-swap deepfakes [6]. Yet another line of research has explored temporal inconsistencies of face identities within a video [45], and identity inconsistencies of inner and outer face regions [22]. Our experiments show that for avatar fingerprinting, such features that are designed to distinguish real from synthetic videos are not reliable, since we only have access to synthetic videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 334, + 481, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 334, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 133, + 334, + 481, + 606 + ], + "type": "text", + "content": "Dynamic Facial Identity Signatures. Cognitive scientists have studied the impact of \"dynamic facial identity signatures\" (i.e., characteristic or identity-specific movements of a face) for identity recognition for humans [51]. In one experiment, scientists projected facial animations generated by human actors onto an average head and found that subjects discriminated between individuals based solely on facial motion [31]. In another, subjects correctly attributed animations of synthetic faces to their morphed versions [41]. While these studies point to the existence of \"dynamic facial identity signatures\" that humans rely on, ours is the first method that isolates these from videos. Talking-Head Datasets and Generators. Existing talking-head datasets include those that contain only real videos showing a variety of emotions [16, 46, 59], and others that also contain synthetic videos [21, 24, 30, 35, 42, 44, 53]. These datasets cater to traditional forensics and facial analysis tasks. Therefore, they do not contain self-and cross-reenactments driven by multiple identities, as well as scripted and free-form monologues across diverse capture settings. There novel requirements posed by the avatar fingerprinting task motivate us to design our own dataset. We focus specifically on face-reenactment talking-head avatar generators for synthetic video generation—this class of generators are the most relevant to AR/VR interactions, video conferencing, and several other applications [2-5, 48]—and combine various modes of human expression for capturing real videos. Given a target facial image and a driving video, these generators reenact the target image using the facial expressions and head pose from the driving video [23, 33, 34, 37, 49, 56, 60, 62, 66, 67]. Another class of talking-head generators use person-specific models [40] and some models aim to preserve the style of the target identity in the synthesized video [39]. However, these models require person-specific training, making them difficult to scale." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 621, + 217, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 217, + 635 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 217, + 635 + ], + "type": "text", + "content": "3 Terminology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "text", + "content": "We seek to verify the trustworthiness of a synthesized talking-head video, termed target video. We assume that an avatar-generation tool (e.g., [60]) created it by animating" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "E. Prashnani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 112, + 482, + 198 + ], + "blocks": [ + { + "bbox": [ + 133, + 112, + 482, + 198 + ], + "lines": [ + { + "bbox": [ + 133, + 112, + 482, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 482, + 198 + ], + "type": "image", + "image_path": "8a04554487271f3aa6ec5972c84632c6fe7d5ba5b7d1cfea152b0541cc4272d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 278 + ], + "type": "text", + "content": "Fig. 2: We introduce the NVFAIR dataset, containing real and synthetic talking-head videos. We capture subjects talking in both scripted and free-form settings. To encourage natural performance, we record the subjects while videoconferencing with each other (left). We then synthesize more than 650,000 talking-head videos—the largest collection till date—using three state-of-the-art face-reenactment talking-head generators. On the right, each row corresponds to a driving identity " + }, + { + "bbox": [ + 130, + 205, + 482, + 278 + ], + "type": "inline_equation", + "content": "(\\mathrm{ID}_i \\rightarrow (\\cdot))" + }, + { + "bbox": [ + 130, + 205, + 482, + 278 + ], + "type": "text", + "content": " and each column corresponds to a different target identity " + }, + { + "bbox": [ + 130, + 205, + 482, + 278 + ], + "type": "inline_equation", + "content": "((\\cdot) \\rightarrow \\mathrm{ID}_i)" + }, + { + "bbox": [ + 130, + 205, + 482, + 278 + ], + "type": "text", + "content": ". The videos in which driving and target identity match are self-reenactments, the rest are cross-reenactments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 396 + ], + "type": "text", + "content": "an image (target image) using the expressions and head poses obtained from another video, the driving video. We call driving identity the identity of the person in the driving video, and target identity the identity of the person in the target image. When driving and target identities match, the target video is a self-reenactment, while the case of a driving identity used to animate a different target identity is cross-reenactment. In both cases, the appearance of the synthesized video is derived from the target identity. This terminology allows us to formally state our goal: we want to verify that a target video is a self-reenactment. With this terminology in mind, we introduce our dataset, which includes real videos as well as self- and cross-reenactment videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 409, + 422, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 422, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 422, + 422 + ], + "type": "text", + "content": "4 The NVIDIA Facial Reenactment (NVFAIR) Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 428, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 472 + ], + "type": "text", + "content": "Recall that avatar fingerprinting is not about detection of synthetic media. Rather, we already know a video to be synthesized, and seek to verify that the driving identity is authorized. This new task dictates a set of requirements for the dataset to be effective for training and evaluation. Specifically, we need a dataset that contains" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 474, + 480, + 529 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 474, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 474, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 136, + 474, + 480, + 495 + ], + "type": "text", + "content": "1. multiple real videos per identity, with scripted and free-form conversations, and with both natural and prescribed emotions," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 496, + 480, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 496, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 135, + 496, + 480, + 518 + ], + "type": "text", + "content": "2. self- and cross-reenactments of target identities, with cross-reenactments driven by all subjects to allow for a variety of driving facial structures, and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 519, + 302, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 519, + 302, + 529 + ], + "spans": [ + { + "bbox": [ + 135, + 519, + 302, + 529 + ], + "type": "text", + "content": "3. multiple face-reenactment generators." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 530, + 501, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 501, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 501, + 610 + ], + "type": "text", + "content": "All relevant existing datasets only capture a subset of these requirements (see Table 1; the Supplement contains further discussion). Moreover they do not use the state-of-the-art talking-head generation technology to synthesize the self- and cross-reenactment videos. We introduce the NVFAIR dataset that features all of the above properties. With over 650,000 synthetic videos, it provides " + }, + { + "bbox": [ + 130, + 530, + 501, + 610 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 130, + 530, + 501, + 610 + ], + "type": "text", + "content": " as many videos as the next largest dataset, and uses three different state-of-the-art generators for the self- and cross-reenactment videos. Figure 2 shows an overview of data capture and synthesis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 625, + 238, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 625, + 238, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 625, + 238, + 637 + ], + "type": "text", + "content": "4.1 Real Data Capture" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": "Capturing videos of monologues delivered by different subjects for the purpose of identity verification introduces two conflicting goals. On the one hand a controlled evalua-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "text", + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 215 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 215 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 215 + ], + "type": "table", + "html": "
Dataset# Subjects (source)(F)ree / (S)scripted?Emotion: (N)atural / (P)rescribed(R)eel / (S)elf / (C)ross-reenact.?#Face Reenact.Avg. Videos per Subject# Face- reenact. GeneratorsEthnic Diversity
RAVDESS [46]24 (new)(S)(P)(R) onlyN/A120 (R)N/A✓
MEAD [59]60 (new)(S)(P)(R) onlyN/A720 (R)N/A✓
CREMA-D [16]91 (new)(S)(P)(R) onlyN/A81 (R)N/A✓
VFHQ [24]36 ([46])(S)(P)+(N)(R)+(S)1,7371201✓
FF++ [54]1000 (YT)(F)(N)(R)+(C)2,0001(R)+2(C)2✓
KoDF [42]403 (new)(F)+(S)(N)(R)+(C)61,000150(R)+151(C)1X
NVFAIR (Ours)161 (46 new + [46], [16])(F)+(S)(P)+(N)(R)+(S)+(C)654,72676(R)+228(S) +3,840(C)3✓
", + "image_path": "9063d62166d5ead3ef8c3e401f6bc6c5c467c5c5f98c47bfbaed2f9a0e2af457.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 220, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 481, + 304 + ], + "type": "text", + "content": "Table 1: The existing talking-head video datasets were designed for tasks such as deepfake detection or facial emotion analysis. Avatar fingerprinting is a fundamentally different task. As a result, no existing dataset satisfies the requirements for training and evaluating models for avatar fingerprinting. To overcome this limitation, we introduce the NVFAIR dataset, which is the first dataset that offers the complete set of monologue modalities, and features the largest collection of facial reenactments to date. Specifically, it provides scripted and free-form monologues, with natural and prescribed emotions, and self- and cross-reenactments (driven by all remaining subjects) synthesized using three generators, alongside original videos for newly-recorded subjects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 320, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 481, + 491 + ], + "type": "text", + "content": "tion of the trained models requires predictability of what is spoken to prevent identification algorithms from latching onto the spoken content itself. On the other, we want the subjects to act as they would in a casual conversation, rather than reciting memorized text, to capture their uniquely identifying mannerisms. We address this trade-off by recording the subjects while videoconferencing in pairs, which creates the impression of being in a natural conversation. This differs from existing datasets, in which the subjects look at the camera, but are not interacting with another person during the recording [16,42,46]. We also design two distinct recording strategies: a free-form stage where the subjects are given only general guidance on the topics, and a more controlled scripted stage in which subjects speak short, memorized monologues of 2-3 sentences each, see Figure 2(a). To capture the variability of real-life scenarios, we provided minimal instructions on how to setup the video call, allowing for diverse face, scale, and lightning, bandwidth stability, and background scene clutter. In total we record 46 subjects of diverse genders, ages, and ethnicities, while strictly abiding by a pre-approved IRB protocol (see Supplement for details and privacy considerations)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 504, + 481, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 481, + 586 + ], + "type": "text", + "content": "Stage I: Free-Form Monologues. In this first stage, the two subjects on the call alternate between asking and answering seven pre-defined questions. The questions are designed to avoid sensitive or potentially inflammatory topics. This is critical because we later use sentences spoken by one individual to animate the video of a second individual, quite literally putting words in their mouths. The complete list of questions is in the Supplement. To further create a natural interaction, the subject listening is encouraged to actively but silently engage with the one speaking (e.g., by nodding or smiling)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 597, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 481, + 667 + ], + "type": "text", + "content": "Stage II: Scripted Monologues. For this stage, we prepared thirty short utterances consisting of two or three sentences each. We chose this length to allow for memorization, while still providing enough content to trigger facial expressions. However, to avoid inducing unnatural expressions, we do not prescribe specific emotions for each utterance. For instance, we do not ask to express anger for a sentence, but we do choose sentences that may naturally evoke it, and used punctuation to encourage it, e.g., \"Will you please" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "E. Prashnani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 162 + ], + "type": "text", + "content": "answer the darn phone? The constant ringing is driving me insane!\" We instruct the subjects to split their screens to show both this list and video call and encourage them to speak to their recording partner when reciting, see Stage II in Figure 2(a). More details, including the full list of utterances can be found in the Supplement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 179, + 288, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 179, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 288, + 192 + ], + "type": "text", + "content": "4.2 Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "text", + "content": "Using the videos described in Section 4.1, as well as the original videos from the CREMA-D [16] and RAVDESS [46] datasets, we generate synthetic talking-head videos to train and evaluate our avatar fingerprinting algorithm. Specifically, we pool the 91 identities from the original videos of CREMA-D [16], the 24 identities from those of RAVDESS [46], and the 46 from our own video-conferencing data capture, for a total of 161 unique identities " + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "text", + "content": ". Recall that we have several real videos for each identity " + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i\\in \\mathcal{I}" + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "text", + "content": ". To avoid a combinatorial explosion of synthetic videos, for all pairs of identities " + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_j" + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_j" + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "text", + "content": " as the target identity and we randomly select 8 of the videos of " + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "text", + "content": " to generate 8 cross-reenactment videos, " + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{ID}_i^k\\to \\mathrm{ID}_j\\}_{k = \\{1,\\dots ,8\\}}" + }, + { + "bbox": [ + 132, + 199, + 482, + 337 + ], + "type": "text", + "content": " (all 8 share the same target image). We also generate self-reenactment videos for each of the target identities, by animating their neutral-face images derived from captured videos with each of their real videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 338, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 482, + 439 + ], + "type": "text", + "content": "We use three different generators for synthesizing the videos for all the 161 identities: face-vid2vid [60], LIA [62], and TPS [34]. This allows us to test if our model generalizes across generators. We chose these talking-head generators because they are the state of the art and they preserve the identity-specific facial motion dynamics well. Nevertheless, the reconstruction is not perfect; for instance, in the third row of Figure 2(b) the person in the driving video " + }, + { + "bbox": [ + 132, + 338, + 482, + 439 + ], + "type": "inline_equation", + "content": "(\\mathrm{ID}_5)" + }, + { + "bbox": [ + 132, + 338, + 482, + 439 + ], + "type": "text", + "content": " is squinting, but the eyes are shut in all the synthetic videos, including the self-reenactment video. In total we generate more than 650,000 synthetic videos, which required more than 2,500 RTX 3090 GPU hours. More details are in the Supplement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 457, + 192, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 192, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 192, + 470 + ], + "type": "text", + "content": "5 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 483, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 482, + 597 + ], + "type": "text", + "content": "Overview. We seek to verify the driving identity of a synthetic video, independently of the target identity. We leverage the finding from cognitive science research that each person emotes in unique ways when communicating, and that this signal is sufficient for recognition, even when the actual appearance is artificially corrupted [31, 41, 51]. We note that our method does not latch onto artifacts introduced by the generators—a property that we demonstrate by showing generalization to new generators not seen during training. Rather, our features capture the dynamics of the expressions, like the way a person frowns, or the way she smiles. Notably, they are distinct from the temporal artifacts introduced by the generator, and that existing algorithms use to detect whether a video is synthetic or real [29]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 597, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 482, + 665 + ], + "type": "text", + "content": "An overview of our algorithm for avatar fingerprinting is shown in Figure 3. To capture expressions, we extract the relative position of facial landmarks over time from the input video, as shown in Figure 3 (Section 5.1). We learn to project these temporal signatures onto a dynamic identity embedding in which features belonging to the same driving identity are close to each other regardless of the target identity, i.e., independently of appearance (Figure 1). To learn this embedding we train a 3DCNN (where the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 114, + 433, + 237 + ], + "blocks": [ + { + "bbox": [ + 179, + 114, + 433, + 237 + ], + "lines": [ + { + "bbox": [ + 179, + 114, + 433, + 237 + ], + "spans": [ + { + "bbox": [ + 179, + 114, + 433, + 237 + ], + "type": "image", + "image_path": "b93e19b071ac1ecc9fc948c81a089ccf1213e326648c1c47d17fbe0d1e9245e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 293 + ], + "type": "text", + "content": "Fig. 3: We extract landmarks from the frames of a talking-head clip, compute their normalized pairwise distances, and concatenate the frame-wise features. We then learn an identity embedding using a loss that pulls closer features of videos driven by the same identity and pushes away those driven by others. " + }, + { + "bbox": [ + 130, + 239, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i\\rightarrow \\mathrm{ID}_j" + }, + { + "bbox": [ + 130, + 239, + 482, + 293 + ], + "type": "text", + "content": " indicates a video that looks like identity " + }, + { + "bbox": [ + 130, + 239, + 482, + 293 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 239, + 482, + 293 + ], + "type": "text", + "content": " (the \"target\" identity), and is driven by identity " + }, + { + "bbox": [ + 130, + 239, + 482, + 293 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 239, + 482, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 482, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 347 + ], + "type": "text", + "content": "third dimension spans video frames—a temporal CNN) with a novel contrastive loss that pulls together all embedding vectors of synthetic videos driven by an individual, while pushing away the embedding vectors of videos driven by all others individuals (Section 5.2). More implementation details are in Section 5.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 362, + 299, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 362, + 299, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 362, + 299, + 374 + ], + "type": "text", + "content": "5.1 Dynamic Facial Identity Features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "type": "text", + "content": "Our first step is to extract temporal features that summarize short segments of the video we wish to fingerprint. We identify the following guiding principles for the extracted features. We would like features that:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 415, + 479, + 460 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 135, + 415, + 479, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 415, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 135, + 415, + 479, + 436 + ], + "type": "text", + "content": "1. have minimal dependency on the appearance of the face in the video (that is, the target identity)," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 437, + 326, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 437, + 326, + 448 + ], + "spans": [ + { + "bbox": [ + 135, + 437, + 326, + 448 + ], + "type": "text", + "content": "2. reflect the dynamics of the expressions, and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 449, + 258, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 449, + 258, + 460 + ], + "spans": [ + { + "bbox": [ + 136, + 449, + 258, + 460 + ], + "type": "text", + "content": "3. capture subtle expressions." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 460, + 481, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 528 + ], + "type": "text", + "content": "One choice could be per-frame 3DMM features [14]: a strategy also used by Cozzolino et al. to detect synthetic videos [20]. However, we empirically observe that 3DMM features are not sufficiently expressive, and do not satisfy desideratum 3 (see ablation experiments in Section 6). We observe a similar behavior for action units [9]. Facial landmarks [32] address this issue, but are sensitive to the shape of the face in the video, and thus to the target identity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 529, + 481, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 481, + 574 + ], + "type": "text", + "content": "To leverage the expressiveness of facial landmarks while abstracting from the underlying facial shape, we compute the pairwise normalized Euclidean distances between each pair of landmarks of a frame. We concatenate these distances into a single vector for the frame, " + }, + { + "bbox": [ + 130, + 529, + 481, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_f" + }, + { + "bbox": [ + 130, + 529, + 481, + 574 + ], + "type": "text", + "content": ". A subset of the facial landmarks and distances are shown in Figure 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 574, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 481, + 641 + ], + "type": "text", + "content": "We then break the input video into clips, which are sequences consisting of F frames and offset by one frame (e.g., [1,F], [2,F+1], etc.), and concatenate the vectors from all the frames in each clip. Using the change in the relative position of the landmarks over a short period of time (the length of a clip) allows us to capture temporal dynamics with minimal dependence on the absolute position of each landmark, i.e., independently of the shape of the face." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": "We show empirically that our features are a good representation for our task, by comparing against alternative choices for input features such as 3DMM (Section 6)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "E. Prashnani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "type": "text", + "content": "5.2 Dynamic Identity Embedding Contrastive Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 479, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 479, + 190 + ], + "type": "text", + "content": "While the features described in Section 5.1 are designed to extract low-level motion dynamics, they cannot be used directly to disambiguate two target videos based on their driving identities. We tackle this problem by learning a dynamic identity embedding, a space where videos driven by one subject map to points that are close to each other and far from the videos driven by anybody else." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "text", + "content": "Specifically, we use a temporal 3DCNN to extract an embedding vector from a clip, which, as described before, is a short segment of an input video. To train the network we use a dataset of synthetic videos driven by different identities. We denote as " + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{ID}_1\\to \\mathrm{ID}_2}^k (t)" + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "text", + "content": " the embedding produced by the network for the clip starting at time " + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "text", + "content": "-th video, of a target identity " + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_2" + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "text", + "content": " driven by identity " + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 130, + 190, + 480, + 258 + ], + "type": "text", + "content": ". As stated above, we have two main objectives, which we capture with the following terms in our proposed loss function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 269, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 269, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 131, + 269, + 479, + 292 + ], + "type": "text", + "content": "We Want to Pull Together All the Videos Driven by " + }, + { + "bbox": [ + 131, + 269, + 479, + 292 + ], + "type": "inline_equation", + "content": "ID_{1}" + }, + { + "bbox": [ + 131, + 269, + 479, + 292 + ], + "type": "text", + "content": ". To achieve this, we define the following term:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 298, + 481, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 298, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 195, + 298, + 481, + 323 + ], + "type": "interline_equation", + "content": "\\mathrm {N} _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\sum_ {\\mathrm {I D} _ {l}, k} \\max _ {n} s \\left(\\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {2}} ^ {j} (t), \\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {l}} ^ {k} (n)\\right), \\tag {1}", + "image_path": "cc4e5bbbef4dbe06b681a64bd33c5547c9e8d9e02b052b0cc5e7ec616c54528e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "s(\\cdot, \\cdot) = e^{-\\|\\cdot\\|^{2}}" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": " is a similarity metric. Intuitively, Equation 1 takes two videos, " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": ", both driven by " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": ". Given a clip starting at time " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": " in the first video, it looks for the most similar clip in the second video. Since the driving identity is the same for both videos, Equation 1 encourages an embedding where clips that capture a similar expression are closer to each other. Equation 1 is high even if only one clip from video " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": " has a similar temporal signature to " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{ID}_1 \\to \\mathrm{ID}_2}^j(t)" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": ". That is because even just one occurrence of the same expression is evidence that the driving identity may be the same. Of course, other driving identities may use similar expressions and we address that with the loss term described below. Additionally, we note that " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": " spans the set of all videos driven by " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_l" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": " spans all identities, including " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_l = \\mathrm{ID}_1" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_l = \\mathrm{ID}_2" + }, + { + "bbox": [ + 131, + 332, + 482, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 458, + 457, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 457, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 457, + 470 + ], + "type": "text", + "content": "We Want to Push Away Videos not Driven by " + }, + { + "bbox": [ + 132, + 458, + 457, + 470 + ], + "type": "inline_equation", + "content": "ID_{1}" + }, + { + "bbox": [ + 132, + 458, + 457, + 470 + ], + "type": "text", + "content": ". We define the following term:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 186, + 476, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 476, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 186, + 476, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\mathrm {Q} _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\sum_ {\\mathrm {I D} _ {l} \\neq \\mathrm {I D} _ {1}, k} \\max _ {n} s \\left(\\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {2}} ^ {j} (t), \\mathcal {C} _ {\\mathrm {I D} _ {l} \\rightarrow \\mathrm {I D} _ {2}} ^ {k} (n)\\right), \\tag {2}", + "image_path": "ae92d120e07681539f43f21c664cc330ac8634b952a40bb495a92c4084d8e6be.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "content": "where, similarly to Equation 1, we take a clip from video " + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "content": ", and look for the most similar clip in video " + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "content": ". This time the two videos share the same target identity, but are driven by different identities: we want all the videos driven by identities different from " + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "content": " to be pushed away from those driven by " + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "content": ", including videos where " + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "content": " is the target identity. Note that " + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_2" + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "content": " spans all identities, including " + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_2 = \\mathrm{ID}_1" + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_2 = \\mathrm{ID}_l" + }, + { + "bbox": [ + 131, + 509, + 479, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 575, + 479, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 479, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 479, + 633 + ], + "type": "text", + "content": "We Want to Rely on the Temporal Dynamics of the Videos Driven by " + }, + { + "bbox": [ + 131, + 575, + 479, + 633 + ], + "type": "inline_equation", + "content": "ID_{1}" + }, + { + "bbox": [ + 131, + 575, + 479, + 633 + ], + "type": "text", + "content": ". Although we use a temporal CNN, the model could still learn to rely on static expressions, such as a snapshot of the person smiling, rather than the temporal progression of expression leading to, or following, the smile. To further encourage the model to learn from the temporal dynamics, we introduce an additional term:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 195, + 639, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 639, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 195, + 639, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathrm {R} _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\sum_ {\\mathrm {I D} _ {l}, k} \\max _ {n} s \\left(\\mathcal {C} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {2}} ^ {j} (t), \\tilde {\\mathcal {C}} _ {\\mathrm {I D} _ {1} \\rightarrow \\mathrm {I D} _ {l}} ^ {k} (n)\\right), \\tag {3}", + "image_path": "4885a2ef142f6ce66e29bfe46a46981ab05d8359f402b83a8fb96b35b0af7b66.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "text", + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{C}}_{\\mathrm{ID}_1\\to \\mathrm{ID}_l}^k" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " denotes a version of the clip " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{ID}_1\\to \\mathrm{ID}_l}^k" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " from Equation 1 with randomly shuffled frame ordering. We want such time-shuffled versions of the clips driven by " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " to be pushed away from the pristine clips driven of " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ". Effectively, this means that the driving identity of the time-shuffled clips is regarded as different from " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ". In other words, we want to pull together video clips in the learned embedding space only when the temporal facial dynamics are characteristic of " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ". We further show the importance of this term in Section 6.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 482, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 239 + ], + "type": "text", + "content": "Combining Equations 1, 2, and 3, we write the probability that the embedding vector " + }, + { + "bbox": [ + 130, + 201, + 482, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{ID}_1\\to \\mathrm{ID}_2}^j (t)" + }, + { + "bbox": [ + 130, + 201, + 482, + 239 + ], + "type": "text", + "content": " lies close to the embedding vectors for all video clips driven by " + }, + { + "bbox": [ + 130, + 201, + 482, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1" + }, + { + "bbox": [ + 130, + 201, + 482, + 239 + ], + "type": "text", + "content": " and far from all the videos driven by others as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 248, + 481, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 248, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 184, + 248, + 481, + 277 + ], + "type": "interline_equation", + "content": "p _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t) = \\frac {\\mathrm {N} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t)}{\\mathrm {N} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t) + \\mathrm {Q} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t) + \\mathrm {R} _ {j , \\mathrm {I D} _ {1} , \\mathrm {I D} _ {2}} (t)}, \\tag {4}", + "image_path": "87c7dbc8f3ff9c5d1964da03a28f66d3c9b74ea30b8c565e68d053ddc45337a2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 285, + 254, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 254, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 254, + 297 + ], + "type": "text", + "content": "and the complete loss term as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 236, + 308, + 481, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 308, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 236, + 308, + 481, + 336 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}, t} - \\log \\left(p _ {j, \\mathrm {I D} _ {1}, \\mathrm {I D} _ {2}} (t)\\right). \\tag {5}", + "image_path": "35b6b4395299722cb3c38a4c19f3b9844984422907faeb249b10e77558fa0f8e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 357, + 227, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 227, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 227, + 369 + ], + "type": "text", + "content": "5.3 Implementation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "text", + "content": "Parameter Choices. To extract the per-frame dynamic facial identity features " + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_f" + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "text", + "content": ", we detect 126 facial landmarks for each frame [32], and compute the per-frame normalized pairwise Euclidean distances between these landmarks. The clip duration is set to 71 frames. We find that this is sufficient to capture the facial dynamics that are meaningful for avatar fingerprinting, while also maintaining a good trade-off between speed and accuracy. We also experiment with shorter-duration video clips (see Supplement). The input tensor to the temporal CNN is obtained by concatenating " + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_f" + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "text", + "content": " across 71 frames. In each batch, we include 8 unique identities. For each identity " + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "text", + "content": ", the pull term (Equation 1) comprises 16 clips: 8 are self-reenactments, randomly sampled from the full set, and the remaining are cross-reenactments with " + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "text", + "content": " as the driving identity. This allows the neural network to pull together videos based purely on the facial motion, regardless of the appearance of the video. The push term (Equation 2) for " + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "text", + "content": " consists of clips with the remaining 7 identities in the batch serving as driving identities (8 clips per driving identity), as well as the time-shuffled self-reenactments of " + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 381, + 482, + 553 + ], + "type": "text", + "content": " (Equation 3). Additional training details can be found in the Supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 562, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 482, + 666 + ], + "type": "text", + "content": "Training, Validation, and Testing Datasets. Of the 161 total identities (pooling together the identities from our dataset, RAVDESS, and CREMA-D, see Section 4.2), we reserve 35 for testing, 14 for validation, and 112 for training. We ensure that there are no cross-set cross-reenactments: that is, identities in the training set only drive other training-set identities (and similarly for the validation and test sets). This allows us to test the models on facial appearances and expressions that were not seen during training. To evaluate the generalization of trained models to new generators, we train our network on videos generated for the training set identities using one generator, and evaluate on the synthetic videos of test-set identities synthesized using remaining two generators." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "E. Prashnani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 113, + 236, + 191 + ], + "blocks": [ + { + "bbox": [ + 167, + 113, + 236, + 191 + ], + "lines": [ + { + "bbox": [ + 167, + 113, + 236, + 191 + ], + "spans": [ + { + "bbox": [ + 167, + 113, + 236, + 191 + ], + "type": "image", + "image_path": "2856cf1c7e156858dd5e3054c4b3c1d882ad6ad1da1715cf87f1feae9ff52e88.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 191, + 223, + 198 + ], + "lines": [ + { + "bbox": [ + 180, + 191, + 223, + 198 + ], + "spans": [ + { + "bbox": [ + 180, + 191, + 223, + 198 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_6 \\rightarrow \\mathrm{ID}_6" + }, + { + "bbox": [ + 180, + 191, + 223, + 198 + ], + "type": "text", + "content": " d=0.84" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 113, + 306, + 191 + ], + "blocks": [ + { + "bbox": [ + 238, + 113, + 306, + 191 + ], + "lines": [ + { + "bbox": [ + 238, + 113, + 306, + 191 + ], + "spans": [ + { + "bbox": [ + 238, + 113, + 306, + 191 + ], + "type": "image", + "image_path": "ec4805d1f008fd55e310f48309f9773b78166d3cc9be3f108d15495835f6f0ab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 192, + 293, + 198 + ], + "lines": [ + { + "bbox": [ + 250, + 192, + 293, + 198 + ], + "spans": [ + { + "bbox": [ + 250, + 192, + 293, + 198 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_6 \\rightarrow \\mathrm{ID}_1" + }, + { + "bbox": [ + 250, + 192, + 293, + 198 + ], + "type": "text", + "content": " d=0.95" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 113, + 376, + 191 + ], + "blocks": [ + { + "bbox": [ + 307, + 113, + 376, + 191 + ], + "lines": [ + { + "bbox": [ + 307, + 113, + 376, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 376, + 191 + ], + "type": "image", + "image_path": "7d52bc44a8a860f9b99c6870a2f92ababe5ce647a9e786f82496f24f66de7aaf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 192, + 363, + 198 + ], + "lines": [ + { + "bbox": [ + 320, + 192, + 363, + 198 + ], + "spans": [ + { + "bbox": [ + 320, + 192, + 363, + 198 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_1 \\rightarrow \\mathrm{ID}_6" + }, + { + "bbox": [ + 320, + 192, + 363, + 198 + ], + "type": "text", + "content": " d=1.58" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 378, + 122, + 446, + 191 + ], + "blocks": [ + { + "bbox": [ + 378, + 122, + 446, + 191 + ], + "lines": [ + { + "bbox": [ + 378, + 122, + 446, + 191 + ], + "spans": [ + { + "bbox": [ + 378, + 122, + 446, + 191 + ], + "type": "image", + "image_path": "7f4e3e07804a221975f738c25106a4f6cd1d17add076c2b0a06eb57bb735375f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 192, + 433, + 198 + ], + "lines": [ + { + "bbox": [ + 391, + 192, + 433, + 198 + ], + "spans": [ + { + "bbox": [ + 391, + 192, + 433, + 198 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_3 \\rightarrow \\mathrm{ID}_6" + }, + { + "bbox": [ + 391, + 192, + 433, + 198 + ], + "type": "text", + "content": " d=2.11" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 167, + 201, + 236, + 278 + ], + "blocks": [ + { + "bbox": [ + 167, + 201, + 236, + 278 + ], + "lines": [ + { + "bbox": [ + 167, + 201, + 236, + 278 + ], + "spans": [ + { + "bbox": [ + 167, + 201, + 236, + 278 + ], + "type": "image", + "image_path": "f78e140b5a9a6085818b9e3867cf826e6984a2bbe7bacd59fda9f585ba8ca1bc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 280, + 230, + 287 + ], + "lines": [ + { + "bbox": [ + 174, + 280, + 230, + 287 + ], + "spans": [ + { + "bbox": [ + 174, + 280, + 230, + 287 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_2 \\rightarrow \\mathrm{ID}_2" + }, + { + "bbox": [ + 174, + 280, + 230, + 287 + ], + "type": "text", + "content": " d=0.589" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 289, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 395 + ], + "type": "text", + "content": "Fig. 4: Animated figure. Open in a media-enabled viewer like Adobe Reader and click on the inset. Our embeddings capture the dynamics of an expression, rather than the appearance of the face. For each row, we pick a reference identity. The green box indicates reenactments driven by the reference identity, the red and blue are cross-reenactments of the reference identity. We compute the average distance of each clip shown here against all other clips driven by the reference identity. The average distance to the other clips of the reference identity is consistent for a given motion, and lower (better) when the reference identity is driving as compared to the cross-reenactments that use the reference identity as target. Here, we show videos generated by face-vid2vid [60], and use the embedding vectors predicted by the model trained on the same generator (see Figure 6 generalization to new generators not seen during training)." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 239, + 201, + 306, + 278 + ], + "blocks": [ + { + "bbox": [ + 239, + 201, + 306, + 278 + ], + "lines": [ + { + "bbox": [ + 239, + 201, + 306, + 278 + ], + "spans": [ + { + "bbox": [ + 239, + 201, + 306, + 278 + ], + "type": "image", + "image_path": "0aa8dcad2786fb3228b0e2ba6b47264d30154270e9a53ba93090fb1fcbfefb82.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 280, + 299, + 287 + ], + "lines": [ + { + "bbox": [ + 244, + 280, + 299, + 287 + ], + "spans": [ + { + "bbox": [ + 244, + 280, + 299, + 287 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_2 \\rightarrow \\mathrm{ID}_7" + }, + { + "bbox": [ + 244, + 280, + 299, + 287 + ], + "type": "text", + "content": " d=0.961" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 210, + 376, + 278 + ], + "blocks": [ + { + "bbox": [ + 307, + 210, + 376, + 278 + ], + "lines": [ + { + "bbox": [ + 307, + 210, + 376, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 210, + 376, + 278 + ], + "type": "image", + "image_path": "981c1d7f1581ec1c50eff2f0ba39936413500444ef06494e776cff9242966e9e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 280, + 370, + 287 + ], + "lines": [ + { + "bbox": [ + 314, + 280, + 370, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 280, + 370, + 287 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_7 \\rightarrow \\mathrm{ID}_2" + }, + { + "bbox": [ + 314, + 280, + 370, + 287 + ], + "type": "text", + "content": " d=2.042" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 378, + 210, + 446, + 278 + ], + "blocks": [ + { + "bbox": [ + 378, + 210, + 446, + 278 + ], + "lines": [ + { + "bbox": [ + 378, + 210, + 446, + 278 + ], + "spans": [ + { + "bbox": [ + 378, + 210, + 446, + 278 + ], + "type": "image", + "image_path": "d089d3e5d94a267d087796094d566c27637b70e4d8ebca94c70d5da4f051a52a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 280, + 440, + 287 + ], + "lines": [ + { + "bbox": [ + 384, + 280, + 440, + 287 + ], + "spans": [ + { + "bbox": [ + 384, + 280, + 440, + 287 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_8 \\rightarrow \\mathrm{ID}_2" + }, + { + "bbox": [ + 384, + 280, + 440, + 287 + ], + "type": "text", + "content": " d=1.787" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 413, + 209, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 209, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 209, + 426 + ], + "type": "text", + "content": "6 Evaluation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 447, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 516 + ], + "type": "text", + "content": "We thoroughly evaluate our algorithm both qualitatively and quantitatively. Our algorithm outperforms reasonable baselines (Section 6.1), it generalizes to generators not seen in training (Section 6.2), and it is robust to video compression (Supplement). We also perform a number of ablation studies to analyze our design choices: our input features, the importance of the time-shuffling term in the loss function (Section 6.3), and the impact of clip duration (Supplement)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 517, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 666 + ], + "type": "text", + "content": "We begin by evaluating qualitatively our method's ability to extract embedding vectors based on the driving identity. Figure 4 shows a set of self- and cross-reenacted clips (please view the animation in a media-enabled viewer, such as Adobe Acrobat). For each row, we take one identity as reference and we compute the embedding vectors of clips that use it both as the driving and the target identity. We then compute the average Euclidean distance of the resulting embedding vectors against those of the self-reenactments by the same reference identity. We note that the average distance " + }, + { + "bbox": [ + 130, + 517, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 517, + 482, + 666 + ], + "type": "text", + "content": " is lower when the driving identity matches the reference identity (first two columns). We also note that the distances between the clips in the first two columns are similar: this confirms that the distance is a function of the facial motion, rather than the facial appearance. When the driving identity changes, the average distance increases, even if the target identity matches the reference identity, which is precisely our goal. More results are in the Supplement." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "text", + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 459, + 205 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 459, + 205 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 459, + 205 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 459, + 205 + ], + "type": "image", + "image_path": "478bf99b11578ecc7a6ae33c0f8d5184a9db71068e5815d85d685da73eba7339.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 208, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 240 + ], + "type": "text", + "content": "Fig.5: ROC curves and AUC values for our method and two baselines: Agarwal et al. [9] and ID-Reveal [20]. Each sub-plot shows the results on our test set for each of the three talking-head generators: face-vid2vid [60], LIA [62], and TPS [34]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "text", + "content": "To evaluate our approach more formally, we use the 35 unique test-set identities that are not used as driving or target identities in the training set (Section 5.3). One at a time, we treat each identity " + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "text", + "content": " as target and synthesize cross-reenactments using all the remaining identities as drivers. This is the set of \"unauthorized\" synthetic videos for " + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "text", + "content": ". The self-reenacted samples for " + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "text", + "content": " form the \"authorized\" set. Note that there are several self-reenacted videos of " + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "text", + "content": ", one per original video of " + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 252, + 479, + 320 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 321, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 480, + 355 + ], + "type": "text", + "content": "For each target identity " + }, + { + "bbox": [ + 130, + 321, + 480, + 355 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_i" + }, + { + "bbox": [ + 130, + 321, + 480, + 355 + ], + "type": "text", + "content": ", we extract the dynamic identity embedding vector of all the clips in the pool of its self- and cross-reenacted videos, and compute their Euclidean distances. That is, for clip " + }, + { + "bbox": [ + 130, + 321, + 480, + 355 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 321, + 480, + 355 + ], + "type": "text", + "content": " we compute" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 221, + 363, + 373, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 363, + 373, + 378 + ], + "spans": [ + { + "bbox": [ + 221, + 363, + 373, + 378 + ], + "type": "interline_equation", + "content": "d \\left(\\mathcal {C} _ {\\mathrm {I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}} ^ {k}, \\mathcal {C} _ {\\mathrm {I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}} ^ {l}\\right), \\forall l \\neq k, \\text {a n d}", + "image_path": "ff44fe4e9a6d61dfcf4459615da68ea96131bc34c081ed145bcfc45e055d5136.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 221, + 373, + 480, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 373, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 221, + 373, + 480, + 395 + ], + "type": "interline_equation", + "content": "\\begin{array}{l}u \\left(\\mathrm {I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}, \\mathrm {C I D} _ {i} \\rightarrow \\mathrm {I D} _ {i}\\right), \\quad \\forall i, \\quad \\forall i \\neq j.\\end{array}\\tag {6}", + "image_path": "99b5f3a1dfb7a047db8afc3a4974682ace04f2326321984f71116c13048cfa07.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 471 + ], + "type": "text", + "content": "We threshold these distances for each target identity to get an ROC curve, and average across the target identities to get the overall area under the curve (AUC). We note that this AUC measures one model's ability to classify a synthetic video as self-reenactment or as cross-reenacted. We conduct further analysis of our model's ability to classify other categories of videos—such as, evaluating AUC on same-utterance self- vs. cross-reenactments or on scripted vs. free-form monologues—in the Supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 487, + 313, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 313, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 313, + 499 + ], + "type": "text", + "content": "6.1 Comparisons with Existing Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 506, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 666 + ], + "type": "text", + "content": "Avatar fingerprinting is a novel task, and no existing methods directly address it. The closest related works aim at detecting real versus synthetic media. As discussed in Section 2, some of these detectors learn identity-specific features such as facial expressions and head poses [9], or facial shapes and motion [20] and can serve as baselines for the task of avatar fingerprinting with some adaptation. The work by Agarwal et al. trains a model to detect synthetic videos of a specific identity [9]. To adapt it to our task, we train 35 different models, one for each identity in the evaluation, by splitting the corresponding original videos into two subsets. We then test each model on the self- and cross-reenactment videos of the corresponding identity. ID-Reveal, trained on a large-scale dataset, learns an embedding space where real videos of a specific identity are grouped together [20]. Since it shows good generalization to new identities, for the task of synthetic media detection, we directly use the pre-trained model on our data to detect, once again, self- versus cross-reenactment. Figure 5 shows the ROC curves for our method compared to these baselines, on three face-reenactment generators (face-vid2vid [60]," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "E. Prashnani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 114, + 254, + 206 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 254, + 206 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 254, + 206 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 254, + 206 + ], + "type": "image", + "image_path": "6134ed937975c61d817d7a7c19087cca5cc2543f7f5fb83ff84fd9586f142441.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 209, + 479, + 231 + ], + "lines": [ + { + "bbox": [ + 130, + 209, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 479, + 231 + ], + "type": "text", + "content": "Fig. 6: Generalization to new generators. To study the robustness to new talking-head generators, we train three versions of our model on three different generators and test on all three." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 114, + 354, + 206 + ], + "blocks": [ + { + "bbox": [ + 256, + 114, + 354, + 206 + ], + "lines": [ + { + "bbox": [ + 256, + 114, + 354, + 206 + ], + "spans": [ + { + "bbox": [ + 256, + 114, + 354, + 206 + ], + "type": "image", + "image_path": "f8d69fd2100d41683e7984a489e0ff7557bb23b487ec3845844064087ebd3962.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 355, + 114, + 456, + 207 + ], + "blocks": [ + { + "bbox": [ + 355, + 114, + 456, + 207 + ], + "lines": [ + { + "bbox": [ + 355, + 114, + 456, + 207 + ], + "spans": [ + { + "bbox": [ + 355, + 114, + 456, + 207 + ], + "type": "image", + "image_path": "1432e080d0917739ae6197e7975d0ced0d0f909e9a0f7a8ff6b4e9afd15b4c9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 240, + 480, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 480, + 286 + ], + "type": "text", + "content": "LIA [62], and TPS [34]). Our method (AUC=0.868 on face-vid2vid) outperforms by a wide margin both ID-Reveal (AUC=0.720 on face-vid2vid), and the method by Agarwal et al. (AUC=0.726 on face-vid2vid). We also note that, unlike ID-Reveal and our method, Agarwal et al. uses a different model per identity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 300, + 303, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 303, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 303, + 312 + ], + "type": "text", + "content": "6.2 Generalization to New Generators" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 318, + 480, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 387 + ], + "type": "text", + "content": "For an avatar fingerprinting model algorithm to be broadly applicable, generalization to new talking-head generators that are not seen in training is crucial. Since our dataset contains videos synthesized by three different generators, we can train three models, one with each generator, and test these models on all three generators. Figure 6 shows the resulting ROC curves and AUC values: the overlap of the curves and similar AUC values in each subplot confirms that our method generalizes well to new generators." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 401, + 222, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 222, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 222, + 412 + ], + "type": "text", + "content": "6.3 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 418, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 612 + ], + "type": "text", + "content": "Our method outperforms existing baselines by introducing two novel components: the dynamic facial identity features, which capture the facial dynamics in a compact and expressive way, and the loss function, which defines the shape of the identity embedding. Here we study the contribution of each, using the face-vid2vid generator for training and testing. We evaluate the contribution of our dynamic facial identity features by swapping them with 3DMM features [13], a popular choice to capture facial dynamics. Since we use a temporal CNN backbone similar to the one from ID-Reveal, for this ablation we use the loss function proposed in their original paper [20]. We re-train the same network using our features and observe a jump from 0.718 to 0.754 in terms of AUC. Upon inspection we notice that the 3DMM features tend to over-smooth the facial motion, and are unable to capture subtle dynamics that prove critical to avatar fingerprinting, and which our features capture. We also evaluate the contribution of our dynamic identity embedding loss and observe a further improvement (AUC 0.868). With our loss formulation, the advantage of " + }, + { + "bbox": [ + 130, + 418, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{j,\\mathrm{ID}_1,\\mathrm{ID}_2}(t)" + }, + { + "bbox": [ + 130, + 418, + 482, + 612 + ], + "type": "text", + "content": " in Eq. 5 is also evident when compared against a model trained without this term (AUC 0.850). Table 2 summarizes this ablation study. Additional experiments in the Supplement show the impact of " + }, + { + "bbox": [ + 130, + 418, + 482, + 612 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 418, + 482, + 612 + ], + "type": "text", + "content": ", performance on scripted vs. free-from monologues, and robustness to video distortions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 625, + 207, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 625, + 207, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 625, + 207, + 636 + ], + "type": "text", + "content": "6.4 Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "type": "text", + "content": "Our algorithm is less discriminative of subjects that are less emotive and more neutral. In the future, relying on more granular dynamic signatures that can extract micro" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 448, + 102 + ], + "type": "text", + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 197, + 114, + 416, + 171 + ], + "blocks": [ + { + "bbox": [ + 197, + 114, + 416, + 171 + ], + "lines": [ + { + "bbox": [ + 197, + 114, + 416, + 171 + ], + "spans": [ + { + "bbox": [ + 197, + 114, + 416, + 171 + ], + "type": "table", + "html": "
Input FeaturesLossAUC
3DMMID-Reveal rec. loss [20]0.718
Landmark distancesID-Reveal rec. loss [20]0.754
Landmark distancesOur loss without Rj, ID1, ID2(t)0.850
Landmark distancesOur loss with Rj, ID1, ID2(t)0.868
", + "image_path": "afaf7bda5b651afac0aa06a213e2f81807e1dcae11f95fce1ed51335526f1fe0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 176, + 478, + 187 + ], + "lines": [ + { + "bbox": [ + 134, + 176, + 478, + 187 + ], + "spans": [ + { + "bbox": [ + 134, + 176, + 478, + 187 + ], + "type": "text", + "content": "Table 2: Ablation study showing the importance of our input features and loss function design." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 200, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 258 + ], + "type": "text", + "content": "expressions can help alleviate this. The performance of our method degrades when expressions that are critical to verifying the driving identity are not captured by the synthetic portrait generator. Lastly, our dataset currently features only one style of interaction: one-on-one conversations. We plan to expand to other conversation styles, such as one-way speeches, in future." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 273, + 233, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 233, + 287 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 233, + 287 + ], + "type": "text", + "content": "7 Societal Impact" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 294, + 481, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 481, + 364 + ], + "type": "text", + "content": "We acknowledge the societal importance of introducing guardrails when it comes to the use of talking-head generation technology. We present this work as a step towards trustworthy use of such technologies. Nevertheless, our work could be misconstrued as having solved the problem and inadvertently accelerate the unhindered adoption of synthetic talking-head technology. We do not advocate for this. Instead, we stress that this is the first work on this topic and underscore the importance of further research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 372, + 215, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 215, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 215, + 384 + ], + "type": "text", + "content": "8 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 394, + 481, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 481, + 497 + ], + "type": "text", + "content": "Highly photo-real synthetic talking-head generators are becoming increasingly beneficial to applications such as video conferencing and AR/VR-based remote interactions. This trend raises the important new research question of how best to also ensure their safe use in such scenarios. To this end, we investigate the new problem of avatar fingerprinting, to authenticate legitimate talking-heads created by authorized users. We leverage the fact that driving individuals have uniquely identifying dynamic motion signatures, which are also preserved in the synthetic videos that they drive. Since none exists, we contribute a new large-scale dataset carefully designed for avatar fingerprinting and related tasks. We hope that our work lays the foundation for further research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 506, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 652 + ], + "type": "text", + "content": "Acknowledgements. We would like to thank the data-capture participants, and Desiree Luong, Woody Luong, and Josh Holland for their help with Figure 2. We acknowledge David Taubenheim for the voiceover in the demo video, and Abhishek Badki for help with the training infrastructure. We thank Joohwan Kim, Rachel Brown, Anjul Patney, Ben Boudaoud, Josef Spjut, Saori Kaji, Nikki Pope, and Kai Pong for their help with putting together the data capture protocols, informed consent form, photo release form, and agreements for data governance and third-party data sharing. Koki Nagano, Ekta Prashnani, and David Luebke were partially supported by DARPA's Semantic Forensics (SemaFor) contract (HR0011-20-3-0005). This research was funded, in part, by DARPA's Semantic Forensics (SemaFor) contract HR0011-20-3-0005. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 653, + 465, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 465, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 465, + 665 + ], + "type": "text", + "content": "Distribution Statement \"A\" (Approved for Public Release, Distribution Unlimited)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "E. Prashnani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 142, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 480, + 163 + ], + "type": "text", + "content": "1. Apple's vision pro. https://www.apple.com/apple-vision-pro/, accessed: 2024-03-06" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 163, + 384, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 384, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 384, + 175 + ], + "type": "text", + "content": "2. Heygen. https://www.heygen.com, accessed: 2023-11-16" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 175, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 175, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 175, + 480, + 195 + ], + "type": "text", + "content": "3. Microsoft teams mesh. https://www.microsoft.com/en-us/microsoft-teams/microsoft-mesh, accessed: 2024-03-06" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 197, + 419, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 197, + 419, + 206 + ], + "spans": [ + { + "bbox": [ + 138, + 197, + 419, + 206 + ], + "type": "text", + "content": "4. Myheritage. https://www.myheritage.com, accessed: 2023-11-16" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 207, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 207, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 480, + 226 + ], + "type": "text", + "content": "5. Nvidia's maxine. https://developer.nvidia.com/maxine, accessed: 2024-03-06" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "6. Agarwal, S., El-Gaaly, T., Farid, H., Lim, S.N.: Detecting deep-fake videos from appearance and behavior. 2020 IEEE International Workshop on Information Forensics and Security (WIFS) pp. 1-6 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 261, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 292 + ], + "type": "text", + "content": "7. Agarwal, S., Farid, H.: Detecting deep-fake videos from aural and oral dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 293, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 323 + ], + "type": "text", + "content": "8. Agarwal, S., Farid, H., Fried, O., Agrawala, M.: Detecting deep-fake videos from phonemeviseme mismatches. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "type": "text", + "content": "9. Agarwal, S., Farid, H., Gu, Y., He, M., Nagano, K., Li, H.: Protecting world leaders against deep fakes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 356, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 480, + 388 + ], + "type": "text", + "content": "0. Agarwal, S., Hu, L., Ng, E., Darrell, T., Li, H., Rohrbach, A.: Watch those words: Video falsification detection using word-conditioned facial motion. In: IEEE Winter Conference on Applications of Computer Vision (WACV) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 388, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 418 + ], + "type": "text", + "content": "1. Albright, M., McCloskey, S.: Source generator attribution via inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 419, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 480, + 440 + ], + "type": "text", + "content": "2. Baluja, S.: Hiding images in plain sight: Deep steganography. In: Advances in Neural Information Processing Systems (NeurIPS) (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 441, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 441, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 138, + 441, + 480, + 461 + ], + "type": "text", + "content": "3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of SIGGRAPH (1999)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 462, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 462, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 462, + 480, + 483 + ], + "type": "text", + "content": "4. Blanz, V., Vetter, T.: Face recognition based on fitting a 3D morphable model. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2003)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 484, + 480, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 484, + 480, + 505 + ], + "spans": [ + { + "bbox": [ + 138, + 484, + 480, + 505 + ], + "type": "text", + "content": "5. Boháček, M., Farid, H.: Protecting world leaders against deep fakes using facial, gestural, and vocal mannerisms. Proceedings of the national academy of Sciences (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 506, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 506, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 506, + 480, + 536 + ], + "type": "text", + "content": "6. Cao, H., Cooper, D.G., Keutmann, M.K., Gur, R.C., Nenkova, A., Verma, R.: CREMA-D: Crowd-sourced emotional multimodal actors dataset. IEEE Transactions on Affective Computing (2014)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 537, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 537, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 537, + 480, + 557 + ], + "type": "text", + "content": "7. Chung, J.S., Nagrani, A., Zisserman, A.: Voxceleb2: Deep speaker recognition. In: INTER-SPEECH (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 559, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 559, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 559, + 480, + 578 + ], + "type": "text", + "content": "8. Cozzolino, D., Nießner, M., Verdoliva, L.: Audio-visual person-of-interest deepfake detection (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 580, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 580, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 580, + 480, + 611 + ], + "type": "text", + "content": "9. Cozzolino, D., Poggi, G., Verdoliva, L.: Extracting camera-based fingerprints for video forensics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 612, + 480, + 643 + ], + "type": "text", + "content": "20. Cozzolino, D., Rössler, A., Thies, J., Nießner, M., Verdoliva, L.: ID-Reveal: Identity-aware DeepFake video detection. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "21. Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., Canton Ferrer, C.: The deepfake detection challenge dataset. arXiv preprint arXiv:2006.07397 (2020)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 481, + 148 + ], + "type": "text", + "content": "22. Dong, X., Bao, J., Chen, D., Zhang, T., Zhang, W., Yu, N., Chen, D., Wen, F., Guo, B.: Protecting celebrities from deepfake with identity consistency transformer. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 148, + 481, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 148, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 481, + 169 + ], + "type": "text", + "content": "23. Drobyshev, N., Chelishev, J., Khakhulin, T., Ivakhnenko, A., Lempitsky, V., Zakharov, E.: MegaPortraits: One-shot megapixel neural head avatars (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 170, + 481, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 481, + 200 + ], + "type": "text", + "content": "24. Fox, G., Liu, W., Kim, H., Seidel, H.P., Elgharib, M., Theobalt, C.: VideoForensicsHQ: Detecting high-quality manipulated face videos. In: IEEE International Conference on Multimedia and Expo (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 200, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 200, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 200, + 481, + 220 + ], + "type": "text", + "content": "25. Fridrich, J.: Steganography in Digital Media: Principles, Algorithms, and Applications. Cambridge University Press (2009)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 220, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 481, + 251 + ], + "type": "text", + "content": "26. Ge, S., Lin, F., Li, C., Zhang, D., Wang, W., Zeng, D.: Deepfake video detection via predictive representation learning. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 252, + 481, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 481, + 272 + ], + "type": "text", + "content": "27. Hadsell, R., Chopra, S., LeCun, Y.: Dimensionality reduction by learning an invariant mapping. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2006)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 273, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 481, + 303 + ], + "type": "text", + "content": "28. Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M.: Lips don't lie: A generalisable and robust approach to face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 334 + ], + "type": "text", + "content": "29. Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M.: Lips don't lie: A generalisable and robust approach to face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 335, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 481, + 365 + ], + "type": "text", + "content": "30. He, Y., Gan, B., Chen, S., Zhou, Y., Yin, G., Song, L., Sheng, L., Shao, J., Liu, Z.: Forgerynet: A versatile benchmark for comprehensive forgery analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4360-4369 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 366, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 481, + 385 + ], + "type": "text", + "content": "31. Hill, H., Johnston, A.: Categorizing sex and identity from the biological motion of faces. Current Biology (2001)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 481, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 417 + ], + "type": "text", + "content": "32. Honari, S., Molchanov, P., Tyree, S., Vincent, P., Pal, C., Kautz, J.: Improving landmark localization with semi-supervised learning. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 481, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 437 + ], + "type": "text", + "content": "33. Hong, F.T., Zhang, L., Shen, L., Xu, D.: Depth-aware generative adversarial network for talking head video generation (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 438, + 426, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 426, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 426, + 448 + ], + "type": "text", + "content": "34. Jian Zhao, H.Z.: Thin-plate spline motion model for image animation (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 449, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 481, + 479 + ], + "type": "text", + "content": "35. Jiang, L., Li, R., Wu, W., Qian, C., Loy, C.C.: Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 510 + ], + "type": "text", + "content": "36. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 510, + 481, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 529 + ], + "type": "text", + "content": "37. Khakhulin, T., Sklyarova, V., Lempitsky, V., Zakharov, E.: Realistic one-shot mesh-based head avatars. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 530, + 481, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 530, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 530, + 481, + 561 + ], + "type": "text", + "content": "38. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 562, + 481, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 481, + 582 + ], + "type": "text", + "content": "39. Kim, H., Elgharib, M., Zollhöfer, M., Seidel, H.P., Beeler, T., Richardt, C., Theobalt, C.: Neural style-preserving visual dubbing. ACM Transactions on Graphics (ToG) (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 582, + 481, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 481, + 612 + ], + "type": "text", + "content": "40. Kim, H., Garrido, P., Tewari, A., Xu, W., Thies, J., Niessner, M., Pérez, P., Richardt, C., Zollhöfer, M., Theobalt, C.: Deep video portraits. ACM Transactions on Graphics (ToG) (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 613, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 613, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 613, + 481, + 633 + ], + "type": "text", + "content": "41. Knappmeyer, B., Thornton, I., Bulthoff, H.: Facial motion can determine facial identity. Journal of Vision (2001)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 634, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 634, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 481, + 664 + ], + "type": "text", + "content": "42. Kwon, P., You, J., Nam, G., Park, S., Chae, G.: Kodf: A large-scale korean deepfake detection dataset. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10744-10753 (2021)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "E. Prashnani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 148 + ], + "type": "text", + "content": "43. Li, J., Xie, H., Yu, L., Zhang, Y.: Wavelet-enhanced weakly supervised local feature learning for face forgery detection. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 148, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 148, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 480, + 169 + ], + "type": "text", + "content": "44. Li, Y., Sun, P., Qi, H., Lyu, S.: Celeb-DF: A large-scale challenging dataset for DeepFake forensics. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 170, + 480, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 480, + 199 + ], + "type": "text", + "content": "45. Liu, B., Liu, B., Ding, M., Zhu, T., Yu, X.: Ti2net: Temporal identity inconsistency network for deepfake detection. In: IEEE Winter Conference on Applications of Computer Vision (WACV) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 201, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 480, + 231 + ], + "type": "text", + "content": "46. Livingstone, S.R., Russo, F.A.: The ryerson audio-visual database of emotional speech and song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 232, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 480, + 251 + ], + "type": "text", + "content": "47. Luo, X., Zhan, R., Chang, H., Yang, F., Milanfar, P.: Distortion agnostic deep watermarking. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 252, + 480, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 480, + 272 + ], + "type": "text", + "content": "48. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 273, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 480, + 293 + ], + "type": "text", + "content": "49. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit Warping for Animation with Image Sets. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 334 + ], + "type": "text", + "content": "50. Munir, S., Batool, B., Shafiq, Z., Srinivasan, P., Zaffar, F.: Through the looking glass: Learning to attribute synthetic text generated by language models. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 335, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 480, + 354 + ], + "type": "text", + "content": "51. O'Toole, A.J., Roark, D.A., Abdi, H.: Recognizing moving faces: A psychological and neural synthesis. Trends in Cognitive Sciences (2002)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 355, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 355, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 355, + 480, + 385 + ], + "type": "text", + "content": "52. Passos, L.A., Jodas, D., da Costa, K.A., Junior, L.A.S., Rodrigues, D., Del Ser, J., Camacho, D., Papa, J.P.: A review of deep learning-based approaches for deepfake content detection. arXiv preprint arXiv:2202.06095 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 417 + ], + "type": "text", + "content": "53. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: FaceForensics++: Learning to detect manipulated facial images. In: IEEE International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 480, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 448 + ], + "type": "text", + "content": "54. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: Faceforensics++: Learning to detect manipulated facial images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1-11 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 449, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 479 + ], + "type": "text", + "content": "55. Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: A unified embedding for face recognition and clustering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 499 + ], + "type": "text", + "content": "56. Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 500, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 480, + 520 + ], + "type": "text", + "content": "57. Sun, K., Yao, T., Chen, S., Ding, S., Li, J., Ji, R.: Dual contrastive learning for general face forgery detection (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 521, + 480, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 540 + ], + "type": "text", + "content": "58. Tancik, M., Mildenhall, B., Ng, R.: Stegastamp: Invisible hyperlinks in physical photographs. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 541, + 480, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 480, + 571 + ], + "type": "text", + "content": "59. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 572, + 480, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 480, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 480, + 602 + ], + "type": "text", + "content": "60. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 603, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 480, + 633 + ], + "type": "text", + "content": "61. Wang, X., Han, X., Huang, W., Dong, D., Scott, M.R.: Multi-similarity loss with general pair weighting for deep metric learning. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 634, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 634, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 480, + 665 + ], + "type": "text", + "content": "62. Wang, Y., Yang, D., Bremond, F., Dantcheva, A.: Latent image animator: Learning to animate images via latent space navigation. In: International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": "Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 295 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 137 + ], + "type": "text", + "content": "63. Yacoob, Y.: Gan-scanner: A detector for faces of stylegan+ (2021), https://github.com/yaseryacoob/GAN-Scanner" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 481, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 481, + 169 + ], + "type": "text", + "content": "64. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 170, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 481, + 201 + ], + "type": "text", + "content": "65. Yu, N., Skripniuk, V., Chen, D., Davis, L., Fritz, M.: Responsible disclosure of generative models using scalable fingerprinting. In: International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 201, + 481, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 481, + 232 + ], + "type": "text", + "content": "66. Zakharov, E., Ivakhnenko, A., Shysheya, A., Lempitsky, V.: Fast bi-layer neural synthesis of one-shot realistic head avatars. In: European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 232, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 481, + 262 + ], + "type": "text", + "content": "67. Zakharov, E., Shysheya, A., Burkov, E., Lempitsky, V.: Few-shot adversarial learning of realistic neural talking head models. In: IEEE International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 295 + ], + "type": "text", + "content": "68. Zheng, Y., Bao, J., Chen, D., Zeng, M., Wen, F.: Exploring temporal coherence for more general video face forgery detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "E. Prashnani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_content_list.json b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3ebb203ae02226dd54d6517f58f10057838b4f04 --- /dev/null +++ b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_content_list.json @@ -0,0 +1,1964 @@ +[ + { + "type": "text", + "text": "AvatarPose: Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feichi Lu $^{1*}$ , Zijian Dong $^{1,2*}$ , Jie Song $^{1\\dagger}$ , and Otmar Hilliges $^{1}$", + "bbox": [ + 256, + 233, + 745, + 251 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Department of Computer Science, ETH Zürich, Switzerland,", + "2 Max Planck Institute for Intelligent Systems, Germany" + ], + "bbox": [ + 290, + 261, + 709, + 290 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6c607c5062696b75767920ec8fc75be6e627631d2f90d51e1d5f16262ff6430.jpg", + "image_caption": [ + "Fig. 1: We propose AvatarPose, a method for estimating the 3D poses and shapes of multiple closely interacting people from multi-view videos. To this end, we first reconstruct the avatar of each individual and leverage the learned personalized avatars as priors to refine poses via color and silhouette rendering loss. We alternate between avatar refinement and pose optimization to obtain the final pose estimation." + ], + "image_footnote": [], + "bbox": [ + 232, + 304, + 750, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Despite progress in human motion capture, existing multiview methods often face challenges in estimating the 3D pose and shape of multiple closely interacting people. This difficulty arises from reliance on accurate 2D joint estimations, which are hard to obtain due to occlusions and body contact when people are in close interaction. To address this, we propose a novel method leveraging the personalized implicit neural avatar of each individual as a prior, which significantly improves the robustness and precision of this challenging pose estimation task. Concretely, the avatars are efficiently reconstructed via layered volume rendering from sparse multi-view videos. The reconstructed avatar prior allows for the direct optimization of 3D poses based on color and silhouette rendering loss, bypassing the issues associated with noisy 2D detections. To handle interpenetration, we propose a collision loss on the overlapping shape regions of avatars to add penetration constraints. Moreover, both 3D poses and avatars are optimized in an alternating manner. Our experimental results demonstrate state-of-the-art performance on several public datasets.", + "bbox": [ + 259, + 529, + 740, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: human pose estimation $\\cdot$ human close interaction $\\cdot$ multiview pose estimation $\\cdot$ avatar prior", + "bbox": [ + 259, + 777, + 740, + 806 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 220, + 810, + 362, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Now at HKUST(GZ)&HKUST", + "bbox": [ + 220, + 824, + 441, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "People frequently engage in interactions with each other in daily life to offer physical support or convey emotions. For AI systems to interpret 3D human interactions, a foundational step is to reconstruct 3D human poses and shapes using consumer-grade sensors, such as cameras. However, in closely interacting scenarios, such as hugging or kissing, 3D pose estimations face challenges due to strong occlusions. The problem of depth ambiguity worsens when close contact is involved, making it more difficult to predict 3D poses from 2D images or videos. Consequently, a multi-camera setup becomes essential to provide additional observations and to address depth ambiguity in pose estimation.", + "bbox": [ + 212, + 175, + 787, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the ubiquity of close human interactions, the study of estimating such human motions is scarce. Most previous multi-human methods with a multi-view setup [2-4, 19, 22, 30, 58, 66, 72] focus on scenarios where people are at a distance from each other. Some methods [2, 4, 19, 22, 72] typically formulate the problem as a cross-view matching problem, relying heavily on 2D joint estimations for subsequent 3D triangulation. These methods demonstrate high sensitivity to noisy or missing 2D joint estimations, particularly when occlusion occurs. Another group of learning-based methods [58, 61, 63, 66] attempts to integrate 2D features from each view into a 3D voxel space and predict 3D human poses from identified 3D subvolumes of each individual. These methods are more robust to occlusion, but they encounter challenges with generalization and are sensitive to distribution shifts due to the lack of annotated 3D data. To correct abnormal or missing pose estimations, some methods [18, 21] leverage parametric body models like SMPL [37] as full-body priors and fit these 3D models to 2D joint estimations. Although this approach alleviates the issue of abnormal or missing joint estimations, it remains constrained by noisy 2D detections.", + "bbox": [ + 212, + 311, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Embracing the challenging problem of pose estimation for close interactions, we propose a novel method to estimate the 3D poses and shapes of multiple people observed from a sparse set of cameras. Our goal is to ensure that even in close contact, the estimated human poses and shapes are accurate and free from interpenetrations. The key to our method is reconstructing implicit textured avatars of each individual in the scene and leveraging them as a strong personalized prior for pose optimization (Figure 1). In contrast to relying solely on noisy 2D joint detections, this textured avatar prior enables us to leverage pixel-wise color and silhouette information for pose refinement, significantly increasing the robustness and accuracy of our method (Table 3). Meanwhile, the reconstructed avatar provides crucial geometric and appearance information to avoid collisions between individuals. Compared to methods using SMPL body shape [25,42] to penalize collisions, our implicit avatar model contains a more detailed clothed shape and additional appearance cues and enables efficient computation of penetration loss. Due to the mutually beneficial relationship between avatar and pose, we alternate between pose optimization and avatar refinement.", + "bbox": [ + 212, + 553, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, to accelerate the learning of avatars, we model each human individual in canonical space using an efficient neural radiance field variant in Instant NGP [43] and combine it with an efficient SMPL-based deformation", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "module [37]. To learn and render avatar models of multiple people, we adapt layered volume rendering [53, 70] to our avatar model. This adaptation allows us to jointly optimize all avatar models through a straightforward rendering loss. Once learned, the avatar can be animated and rendered based on pose parameters at interactive rates, thus naturally serving as an efficient personalized textured prior for pose optimization. With the learned personalized prior, we optimize pose via a novel objective function. Different from previous methods based on 2D reprojection error of joints, we directly optimize pose parameters via minimizing color and silhouette rendering losses while keeping the learned avatar model fixed. To prevent interpenetration between human individuals, a collision loss is introduced by penalizing the situation when a 3D point is occupied by multiple avatars. To remove artifacts in the initial avatar due to imperfect pose initialization, we further refine avatars based on optimized poses. Throughout the optimization process, both initial personalized avatar models and SMPL parameters are optimized in an alternating manner, motivated by the insight that accurate 3D pose estimations improve avatar learning, and improved avatar models, in turn, increase the precision of overall pose estimations.", + "bbox": [ + 212, + 145, + 787, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We experimentally demonstrate that our method significantly outperforms previous state-of-the-art methods on several public datasets both quantitatively (Table 1 and Table 2) and qualitatively (Figure 3) especially when people are in close interaction. In summary, our contributions are:", + "bbox": [ + 212, + 402, + 787, + 463 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a pipeline that efficiently creates implicit neural avatars of closely interacting people and leverages the learned avatars as priors to optimize poses.", + "- The avatar prior enables us to design a novel objective function that leverages color and mask rendering loss for pose optimization. We show the superiority of this loss function compared to the 2D reprojection error of 3D joints used by most of the previous methods.", + "- Based on the learned avatar, a collision loss is introduced to avoid penetration when individuals are in contact." + ], + "bbox": [ + 223, + 473, + 782, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 632, + 387, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Multi-Person 3D Pose Estimation", + "text_level": 1, + "bbox": [ + 215, + 664, + 539, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite significant progress in multi-human 2D pose estimation [9,13,15,24,34, 49] and 3D pose estimation from monocular images or video [7, 31-33, 38-40, 45, 47, 55, 56, 59, 68, 69], the reconstruction accuracy is still limited due to depth ambiguity and strong occlusions when humans are in close contact. A multi-view setting [4, 5, 19, 22, 23, 35, 36, 54, 61, 63, 66, 72] helps to alleviate these challenges. One straightforward idea of most methods [2, 4, 19, 22, 72, 76] is to formulate the problem into cross-view matching and association problems. MVPose [19] performs 2D person parsing in each image and leverages cross-view person matching to infer 3D pose. 4DAssociation [72] additionally adds tracking into this process to form a unified graph for associating 4D information. However, these methods", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "AvatarPose", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are sensitive to the noisy estimation of 2D pose. In contrast to these matching-based methods, some recent methods [5,14,17,23,35,50,58,61,63,66,71,75] directly learn deep neural networks to regress poses. Faster Voxelpose [66] employs the feature volume proposed by Voxelpose [58] and enhances computational efficiency. Graph [63] designs three graph neural network models for human center detection and pose estimation. MvP [61] simplifies the multi-person pose estimation by direct regression using the transformer model. A concurrent method CloseMocap [54] proposes to learn a model from a synthetic dataset simulating occlusion situations. However, heavily relying on the 2D or 3D features as input during training, these learning-based methods suffer from generalization issues when subjects, motions, and camera configurations change.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further improve the robustness, a statistical parametric body model such as SMPL [37] is explored in [21] as a regularization prior for 3D joint refinement. Some follow-ups [18,73] show that parametric models help in correcting implausible 3D pose estimates and filling in missing joints. However, this coarse body prior highly relies on aligning the 3D joints to 2D pose estimations, which are inaccurate when occlusion happens. Different from all of these previous methods, we explore the usage of personalized textured avatar models as priors to refine human poses. This prior enables us to leverage color and silhouette information from multi-view observation for refining poses of closely interacting humans.", + "bbox": [ + 212, + 313, + 787, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 3D Human Modeling", + "text_level": 1, + "bbox": [ + 214, + 469, + 437, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parametric human body models [1,28,37,44,65] can represent minimally clothed human shapes by deforming a template mesh. It is challenging to extend this explicit representation for modeling clothed humans due to the fixed topology and resolution. To overcome this limitation, methods such as SNARF [12] and SCANimate [52] propose to model articulated human shapes based on 3D implicit representations. Many works [6,8,10,20,26,29,48,57,60,62] fit implicit neural fields to RGB or RGB-D videos by neural rendering to reconstruct the shape and appearance of a single human body. However, when applied to a multi-human scene, these methods are not able to achieve good fidelity due to strong occlusions. Recent methods including ST-NeRF [70] and [53] leverage layered neural representation to model multiple humans with sparse multi-view videos and thus can generate novel view synthesis of dynamic multiple humans. The main problem of all aforementioned approaches, however, is their high reliance on 3D human pose estimation [70,74]: the deformation of the human model requires accurate human poses, which is hard to obtain when people are in close interaction. To address this challenge, our method is orthogonal to others, aiming at leveraging the learned avatar as priors for pose estimation in close interaction.", + "bbox": [ + 212, + 492, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 3D Human Datasets for Close Interaction", + "text_level": 1, + "bbox": [ + 214, + 770, + 607, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most of the existing datasets [3,11,30] like Shelf and Panoptic studying multiperson pose estimation focus on the scene where people are at a distance from each other and rarely involved in close interactions. To study close interactions", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "among people, MultiHuman [74] dataset captures multi-person interaction with some close interactions and occlusions. ExPI [27] creates a multi-person extreme motion dataset with close interactions, but it focuses mainly on motion prediction for future frames instead of pose estimation from sparse views. CHI3D [25] captures two-person interaction datasets and proposes to learn a contact estimation module from annotations to improve the precision of pose estimation. The most recent work Hi4D [67] creates a challenging dataset of physically close human interaction and proposes a method to disentangle human bodies and estimate poses. However, this method relies on the 3D ground truth of clothed human meshes captured with expensive 3D body scanners. Due to the reliance on the limited annotated 3D data, this method faces challenges in generalization with different people and camera configurations. In contrast to all the works before, we intend to solve the problem of pose estimation in close human interaction without requiring accurate 3D scans or other training data.", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 382, + 330, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a dynamic scene captured by a sparse set of RGB cameras, our goal is to estimate the 3D pose and shape of multiple people even if they interact closely. To address this challenging task, our key idea is to first reconstruct the personalized avatar of each individual in the scene and leverage them as a strong prior to refine the appearance and pose in an alternating manner. An overview of our method is shown in Figure 2.", + "bbox": [ + 212, + 414, + 787, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first introduce an efficient pipeline to create avatars of multiple people in a scene (Section 3.1 and Figure 2(a)). Specifically, we leverage an accelerated neural radiance field to represent the shape and appearance of each individual in canonical space and deform it at an interactive rate. We then adapt layered volume rendering to our pipeline, which composites the rendering of avatars into one image, thus enabling direct learning from multi-view video inputs.", + "bbox": [ + 212, + 506, + 787, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thanks to the learned avatar prior for each individual, we can enhance 3D pose optimization via a combination of RGB and silhouette rendering loss (Section 3.2 and Figure 2(b)). While previous work heavily relies on noisy 2D joint detection, we show that employing such pixel-wise color and silhouette information can largely increase precision and robustness. Moreover, a collision loss is introduced to avoid interpenetration. Finally, we alternate between avatar learning and pose optimization to get complete and accurate 3D human poses.", + "bbox": [ + 212, + 597, + 787, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Multi-Avatar Prior Learning", + "text_level": 1, + "bbox": [ + 215, + 724, + 496, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Avatar Model We represent each human individual in canonical space using an accelerated neural radiance field [43] and model shape-aware articulated deformation based on SMPL [37].", + "bbox": [ + 212, + 751, + 785, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Canonical Appearance Representation: To model human shape and appearance, we create canonical radiance field $\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}$ for each human instance", + "bbox": [ + 222, + 806, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "AvatarPose", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9b4547620a3b3a8d47af20c8bfe126a4cdda69f8bc7a9752685987b415f28558.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 145, + 759, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8091daf5024b3007c79f566d9dfa3b8cbe004e3914731505176196066d7cd96c.jpg", + "image_caption": [ + "Fig. 2: Method Overview: Our method consists of two modules: (a) Multi-Avatar Prior Learning: Given the input multi-view images and estimated poses $\\Theta^{(l)}$ , we sample points $\\mathbf{x}^{(l)}$ for each individual $l$ along the rays in posed space and warp these points into canonical space and calculate their color $\\mathbf{c}^{(l)}$ and density $\\sigma^{(l)}$ via the canonical appearance network $\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}$ . We leverage layered volume rendering [70] to attain the final pixel color and compare it with the original input image to optimize the parameters of avatars. (b) Avatar-guided Pose Optimization: Given learned avatar model $\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}$ and initial poses $\\Theta^{(l)}$ of each individual $l$ , we deform all of the avatars based on SMPL-based deformer and render them jointly via layered volume rendering. We compare the composite rendering with input observation and minimize the RGB and mask rendering loss to optimize poses. A collision loss is additionally introduced to avoid interpenetration. Finally, we alternate between two modules to obtain the final result. For clarity, the parameters to be optimized are marked as red in each module." + ], + "image_footnote": [], + "bbox": [ + 233, + 275, + 769, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$l \\in [1, L]$ , where $L$ is the number of humans in the scene. $\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}$ takes a 3D point $\\bar{\\mathbf{x}}^{(l)}$ as input and predicts its density $\\sigma^{(l)}$ and color $\\mathbf{c}^{(l)}$ . Following Instant-NGP [43] and InstantAvatar [29] to accelerate the rendering, $\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}$ is parameterized via using a hash table to store feature grids at different scales.", + "bbox": [ + 240, + 618, + 787, + 684 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pose Representation: We represent the 3D pose and underlying body shape for all human instances by SMPL parameters $\\Theta = \\{\\Theta^{(l)}\\}_{l\\in [1,L]}$ . For each human $l$ , $\\Theta^{(l)} = \\{\\beta^{(l)},\\theta^{(l)},t^{(l)}\\}$ contains shape parameters $\\beta^{(l)}\\in \\mathbb{R}^{10}$ , pose parameters $\\theta^{(l)}\\in \\mathbb{R}^{72}$ and translation $t^{(l)}\\in \\mathbb{R}^3$ of SMPL.", + "- Deformer: To enable animation given targeted poses $\\Theta^{(l)}$ , we require the radiance field in the posed space. Given a point $\\mathbf{x}^{(l)}$ in deformed space of human $l$ , we determine the corresponding canonical point $\\bar{\\mathbf{x}}^{(l)}$ by inverse linear blend skinning(LBS) [37]: $\\bar{\\mathbf{x}}^{(l)}(\\mathbf{x}^{(l)},\\Theta^{(l)}) = (\\sum_{i = 1}^{n_b}w_i(\\Theta^{(l)})\\mathbf{B}_i(\\Theta^{(l)}))^{-1}\\mathbf{x}^{(l)}$ , where $\\mathbf{B}_i$ is the rigid bone transformation matrix for joint $i\\in \\{1,\\dots,n_b\\}$ under pose $\\Theta^{(l)}$ . $w_{i}$ is the skinning weights of the nearest neighbor of $\\mathbf{x}^{(l)}$ in" + ], + "bbox": [ + 225, + 686, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the deformed SMPL vertices. We obtain the radiance field at the point $\\mathbf{x}^{(l)}$ by evaluating the canonical appearance field at the corresponding point $\\bar{\\mathbf{x}}^{(l)}$ .", + "bbox": [ + 241, + 145, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Layered Volume Rendering To obtain the pixel value for a ray $\\mathbf{r} \\in \\mathcal{R}$ , we raycast every human instance separately with a layered rendering strategy similar to ST-NeRF [70]. Specifically, we first calculate the intersection points between the ray and the 3D bounding box of each human instance and uniformly sample points in each bounding box. To distinguish different identities, we assign each sampled point $\\mathbf{x}_i$ a one-hot representation $\\mathbf{m}_i = [m_i^{(1)}, \\dots, m_i^{(L)}]$ to indicate which human identity it belongs to. After sorting all sampled points by their depth values and calculating their corresponding color $\\mathbf{c}_i$ and density $\\sigma_i$ from the avatar model, if $m_i^{(l)} = 1$ , we compute", + "bbox": [ + 215, + 190, + 787, + 330 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {i}, \\sigma_ {i} = \\bar {\\mathbf {F}} _ {\\sigma_ {f}} ^ {(l)} \\left(\\bar {\\mathbf {x}} _ {i} \\left(\\mathbf {x} _ {i}, \\boldsymbol {\\Theta} ^ {(l)}\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 340, + 785, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The color of each ray is computed via numerical integration [41].", + "bbox": [ + 238, + 369, + 705, + 386 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {C}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\prod_ {j < i} (1 - \\alpha_ {j}) \\mathbf {c} _ {i} \\quad \\alpha_ {i} = 1 - \\exp (- \\sigma_ {i} \\delta_ {i}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 398, + 784, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\delta_{i}$ is the distance between samples. The accumulated alpha value, which represents ray opacity, can be computed via:", + "bbox": [ + 215, + 448, + 782, + 478 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\prod_ {j < i} (1 - \\alpha_ {j}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 489, + 784, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each human identity $l \\in [1, L]$ , the corresponding instance ray opacity can be calculated via:", + "bbox": [ + 215, + 541, + 782, + 570 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha^ {(l)} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\prod_ {j < i} (1 - \\alpha_ {j}) m _ {i} ^ {(l)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 580, + 784, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training The overall training process is shown in Figure 2(a). For training avatar layers, we minimize the Huber loss $\\rho$ between the predicted pixel color $\\hat{\\mathbf{C}}(\\mathbf{r})$ and the ground truth pixel color $\\mathbf{C}_{gt}(\\mathbf{r})$ :", + "bbox": [ + 215, + 633, + 782, + 680 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R G B} = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} \\rho \\left(\\left\\| \\hat {\\mathbf {C}} (\\mathbf {r}) - \\mathbf {C} _ {g t} (\\mathbf {r}) \\right\\|\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 691, + 784, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since instance segmentation of human performers is hard to obtain and is not accurate, we choose foreground segmentation as our mask supervision, which is obtained via SAM-Track [16]. We apply a loss for optimizing the rendered alpha values $\\alpha$ to reduce the artifacts in the floating area:", + "bbox": [ + 215, + 736, + 782, + 796 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l p h a}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} \\in \\mathcal {R}} (\\alpha (\\mathbf {r}) - \\alpha_ {S A M} (\\mathbf {r})) ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 808, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "AvatarPose", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following [53], we add a regularization loss for instance alpha values to make sure every pixel can only be rendered from one human layer:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l a y e r}} = - \\frac {1}{L \\mid \\mathcal {R} \\mid} \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\sum_ {l = 1} ^ {L} \\alpha^ {(l)} (\\mathbf {r}) \\log \\left(\\alpha^ {(l)} (\\mathbf {r})\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 189, + 785, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similar to [29], we also add hard surface and density regularization terms in the learning process. More training details can be found in the Supp Mat.", + "bbox": [ + 214, + 236, + 782, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Avatar-guided Pose Optimization", + "text_level": 1, + "bbox": [ + 214, + 289, + 535, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Equipped with learned avatars obtained in Section 3.1, we aim to estimate the 3D shape and pose of multiple humans with close physical contact. To achieve this, we leverage the avatars as priors to handle the challenging pose ambiguities caused by contact. Specifically, we first initialize pose parameters using an off-the-shelf 3D pose estimator [72] and refine the pose via a rendering loss between rendered posed avatars and 2D observations (Figure 2(b)). Since the initial imperfect pose estimations may cause artifacts in avatar reconstruction, we further refine the weights of the avatar model using the optimized pose. This process is formulated as an alternating optimization to refine both poses and avatars.", + "bbox": [ + 212, + 314, + 784, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initialization The initial 3D pose proposals are estimated by leveraging the off-the-shelf 3D human pose estimator [72]. After that, we register a SMPL model to the estimated 3D joints to obtain the initial pose parameters $\\Theta_0$ . Given these estimated poses, the initial avatar model of each individual is further learned from multi-view videos.", + "bbox": [ + 212, + 465, + 784, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Objective To leverage avatars as priors to tackle challenges caused by contact, we optimize the following objective:", + "bbox": [ + 214, + 556, + 782, + 588 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\boldsymbol {\\Theta}) = \\lambda_ {R G B} \\mathcal {L} _ {R G B} (\\boldsymbol {\\Theta}) + \\lambda_ {\\alpha} \\mathcal {L} _ {\\alpha} (\\boldsymbol {\\Theta}) \\tag {8} \\\\ + \\lambda_ {r e g} \\mathcal {L} _ {r e g} (\\Theta) + \\lambda_ {p a} \\mathcal {L} _ {p a} (\\Theta). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 599, + 785, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\Theta$ is the SMPL parameters to be optimized for all avatars, which is also consistent with the pose parameters to deform the avatar model (Section 3.1). Following [62], we represent $\\Theta$ as $\\Theta_0 + \\mathrm{MLP}(\\Theta_0)$ and refine poses by changing the parameters of the neural network. This representation empirically shows more robust results compared to directly optimizing SMPL parameters.", + "bbox": [ + 212, + 643, + 784, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different from previous methods [18,21,58,72], the personalized multi-ayar prior allows us to leverage appearance and silhouette information to refine initial poses. Specifically, we calculate $\\mathcal{L}_{RGB}$ to ensure the color consistency between the rendered pixel $\\hat{\\mathbf{C}} (\\mathbf{r},\\boldsymbol {\\Theta})$ (Equation (1) and Equation (2)) of deformed avatar with poses $\\Theta$ and the corresponding ground-truth pixel color $\\mathbf{C}_{gt}(\\mathbf{r})$ .", + "bbox": [ + 212, + 720, + 784, + 797 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R G B} (\\boldsymbol {\\Theta}) = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} \\rho \\left(\\left\\| \\hat {\\mathbf {C}} (\\mathbf {r}, \\boldsymbol {\\Theta}) - \\mathbf {C} _ {g t} (\\mathbf {r}) \\right\\|\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 806, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, a cross-entropy loss $\\mathcal{L}_{\\alpha}$ is introduced to ensure that the rendered mask $\\alpha (\\mathbf{r},\\Theta)$ (Equation (3)) of the reposed avatar is aligned with the estimated SAM-Track mask $\\alpha_{SAM}(\\mathbf{r})$ by:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\alpha} (\\boldsymbol {\\Theta}) = - \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\alpha_ {S A M} (\\mathbf {r}) \\log (\\alpha (\\mathbf {r}, \\boldsymbol {\\Theta})) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 204, + 785, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To penalize the unnatural poses and avoid elbows and knees bending in the wrong direction, we add an L2 regularization term and combine it with the pose prior in SMPLify [7] to constrain physically implausible joint rotation:", + "bbox": [ + 212, + 239, + 782, + 285 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} (\\boldsymbol {\\Theta}) = \\| \\boldsymbol {\\Theta} \\| _ {2} + \\lambda \\sum_ {i \\in I} \\exp (\\boldsymbol {\\Theta} _ {i}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 294, + 785, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $I$ is the set of pose indices corresponding to elbows and knees.", + "bbox": [ + 238, + 332, + 733, + 347 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A key challenge to correctly estimate poses in close interaction is to handle interpenetration. Since every avatar is modeled separately, the surfaces tend to intersect when they are in contact. To handle this, we first select sampled points inside multiple instances as $\\mathcal{S} = \\{\\mathbf{x}_i\\mid \\alpha_i^{(p)} > 0,\\alpha_i^{(q)} > 0,p,q\\in [1,L],p\\neq q\\}$ ( $\\alpha_{i}^{(p)}$ is calculated from Equation (1) and Equation (2) corresponding with a point $\\mathbf{x}_i$ with $m_i^{(p)} = 1$ ). We then propose a collision loss $\\mathcal{L}_{pa}$ for penalizing penetration:", + "bbox": [ + 212, + 348, + 784, + 443 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p a} (\\boldsymbol {\\Theta}) = \\frac {1}{| \\mathcal {S} |} \\sum_ {\\boldsymbol {x} _ {i} \\in \\mathcal {S}} \\alpha_ {i} ^ {(p)} (\\boldsymbol {\\Theta}) \\alpha_ {i} ^ {(q)} (\\boldsymbol {\\Theta}) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 454, + 785, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Intuitively, this loss guarantees every sample point in 3D space can not be occupied by multiple avatar models simultaneously, which guarantees better pose estimation in close contact.", + "bbox": [ + 212, + 493, + 782, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alternating Optimization Since artifacts sometimes appear on initial avatars due to imperfect pose initialization, we further refine avatars based on optimized poses via minimizing the loss function in Section 3.1. Finally, the optimization of poses and avatars is formulated in an alternating fashion for $N$ steps. More details of optimization can be found in the Supp Mat.", + "bbox": [ + 212, + 550, + 784, + 626 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 647, + 375, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We mainly evaluate our proposed method on Hi4D [67] and CHI3D [25], which are challenging datasets of two humans in close interaction. To demonstrate the generalization ability of our method for more than two people, we also evaluate on Shelf [3] and MultiHuman [74] including three or four people. More details are shown in the Supp Mat.", + "bbox": [ + 212, + 676, + 785, + 752 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics. We use the Mean Per Joint Position Error (MPJPE) [58] to measure the distance between the ground truth 3D poses and the estimated poses. We also choose the Percentage of Correct Parts (PCP3D) metric [19] to calculate the percentage of correct parts. $\\mathrm{AP}_K$ [58] and Recall [21] are also leveraged to evaluate performance. More details are shown in the Supp Mat.", + "bbox": [ + 212, + 763, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "AvatarPose", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a33cfa3ca61ced7bb2f4e29ba22b051cfa9400ab5079e3e8eeeb0722bc712573.jpg", + "image_caption": [ + "Fig. 3: Qualitative Comparison with SotA methods [19,61,64,66,72] on Hi4D and CHI3D. We show two examples from the Hi4D and CHI3D datasets compared with Graph, MvP, Faster VoxelPose, MVPose, and 4DAssociation. For each example, we show 2D projections on two sampled views." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 785, + 391 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We compare our method on the Hi4D dataset with state-of-the-art methods in three categories discussed in Section 2.1. For learning-based methods, we choose Graph [64], MvP [61], and Faster VoxelPose [66] and fine-tune these models on a subset of the evaluating datasets. For pure association-based methods, we choose 4DAssociation [72] and MVPose* [19]. MVPose [18] adds temporal tracking and SMPL prior to MVPose* [19] and is regarded as a SMPL-guided method. More details about baselines can be found in the Supp Mat.", + "bbox": [ + 212, + 489, + 787, + 597 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Comparison to SotA", + "text_level": 1, + "bbox": [ + 215, + 623, + 431, + 638 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1 and Table 2 summarizes our quantitative comparisons on Hi4D and CHI3D with SotA (State-of-the-art) methods. Our method largely outperforms the other SotA methods in all of the metrics including MPJPE, PCP3D, and $\\mathrm{AP}_K$ . More comparisons on Shelf and MultiHuman are shown in the Supp Mat.", + "bbox": [ + 212, + 652, + 787, + 715 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison with Graph [63], MvP [61], Faster VoxelPose [66]. Our method outperforms Graph, MvP, and Faster VoxelPose on Hi4D and CHI3D. These methods are prone to overfitting to training pose distributions, thus struggling with challenging poses of interacting actors. Specifically, Faster VoxelPose sometimes misses actors in close contact, leading to a relatively low recall. While Graph and MvP have better recall, Graph fails to consistently track actors across frames, and MvP results in many misaligned joints between actors.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e68c2567b5663aa2dcaf27d0d178a74cc997632be7f96c3ccb888ad62225fdf7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑Recall(%) ↑
MvP [61]92.7774.1441.5963.8693.84
Graph [63]89.6271.5544.7567.3393.31
Faster VoxelPose [66]68.4073.6744.0568.7083.55
MVPose* [19]53.0587.5767.9780.2893.80
MVPose [18]42.6390.7671.7990.1993.30
4DAssociation [72]41.2988.6280.8797.2798.78
Ours32.1096.9091.4897.3398.78
", + "bbox": [ + 228, + 143, + 772, + 273 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7d252e6b6a06cb16418533ccfdc937da198205d55d6dd6544df50df423eb4d50.jpg", + "table_caption": [ + "Table 1: Quantitative Comparison with SotA on the Hi4D [67] Dataset (8 views). We compare our method with MvP [61], Graph [63], Faster VoxelPose [66], MVPose* [19], MVPose [18] and 4DAssociation [72]. We report MPJPE, PCP, $\\mathrm{AP}_K$ , and Recall metric for all methods." + ], + "table_footnote": [], + "table_body": "
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑Recall(%) ↑
MvP [61]55.3889.4763.5892.5399.06
Graph [63]45.3392.4674.0295.2599.17
Faster VoxelPose [66]67.8178.4129.2882.8893.34
MVPose* [19]50.4290.3969.1375.7288.72
MVPose [18]34.0593.3579.9486.9188.18
4DAssociation [72]37.4799.3089.6698.6799.85
Ours32.9899.7993.2099.7999.85
", + "bbox": [ + 218, + 368, + 785, + 505 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Quantitative Comparison with SotA on the CHI3D [25] Dataset (4 views). We compare our method with Faster VoxelPose [66], MVPose* [19], MV-Pose [18] and 4DAssociation [72]. We report MPJPE, PCP, $\\mathrm{AP}_K$ , and Recall metric for all methods.", + "bbox": [ + 215, + 506, + 787, + 560 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with MVPose [18]. Our method achieves much better MPJPE and precision than MVPose. MVPose relies heavily on noisy 2D joint detected in close interactions. In contrast, our method leverages color and silhouette rendering loss to optimize poses, leading to robustness to occlusions. Additionally, this top-down method cannot detect actors correctly with strong occlusions, as shown in Figure 3.", + "bbox": [ + 212, + 614, + 785, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with 4DAssociation [72]. Finally, we compare our method with the bottom-up association method. As shown in Table 1 and Table 2, our method outperforms 4DAssociation in most metrics. Figure 3 shows that when actors are close, this bottom-up method is inclined to associate joints with the wrong human instances. This is because they solve the joint association with a greedy algorithm, which is sensitive to missing and inaccurate 2D joint detections.", + "bbox": [ + 212, + 748, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "AvatarPose", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ca7495f30530c603fa0431f97c957acb56b029ff68e8cbd66cc4a9a64aa534b3.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 401, + 239 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5751a58588ccc2b09ebe946de64833b59a2e54620d3e5d31683ecd3549ab6732.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 410, + 146, + 589, + 241 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/79063d68ec8bb730e85d5cffc5f86bf63ba694a472853abd67b6b2a8c7293429.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 596, + 147, + 781, + 242 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c8e92be45992602255faae1f8d413a6aa830184d4e3b849b1f2277cb47cc8100.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 220, + 258, + 401, + 351 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1544a0211dc4dd190cfc86513394a251616c8df67a7cdd94ca58d32cf047e78b.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 410, + 258, + 589, + 351 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/37ec93390bfdf6f5cdcfc231f4204692236f2c29be2a0812955b9cb8f5bf1bd5.jpg", + "image_caption": [ + "(f)", + "Fig. 4: Qualitative Results of our method on Hi4D (a)(b), CHI3D (c)(d), MultiHuman Real-Cap (e), and Shelf (f). The left and middle columns in each sub-figure show the 2D projections of the estimated 3D skeletons and SMPL body meshes on two views. The right column in each sub-figure demonstrates skeletons and SMPL bodies in 3D scenes." + ], + "image_footnote": [], + "bbox": [ + 596, + 258, + 779, + 351 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Additional Qualitative Samples", + "text_level": 1, + "bbox": [ + 215, + 479, + 519, + 494 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Figure 4 demonstrates more qualitative results of our method on Hi4D, CHI3D, and MultiHuman Real-Cap with challenging and close interactions among 2 or 3 people. We also demonstrate results on the Shelf [3] dataset, which contains 4 people without close contact. More results can be found in the Supp Mat.", + "bbox": [ + 212, + 503, + 787, + 566 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 588, + 388, + 603 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To validate the effectiveness of our method, we conduct a detailed analysis of different design choices of our algorithm. All the experiments are conducted on the Hi4D Dataset.", + "bbox": [ + 212, + 613, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with SMPL Body Prior. To validate the effectiveness of our personalized avatar prior, we compare with a baseline that optimizes SMPL parameters to align 3D joint reprojections to 2D observations. Our method significantly outperforms the SMPL prior baseline (Table 3) by leveraging color and silhouette rendering loss to refine poses, reducing reliance on inaccurate joint detections when occlusion happens. Figure 5 shows that while the baseline method incorrectly estimates the arm pose and even leads to penetration, our method reconstructs the poses accurately.", + "bbox": [ + 214, + 672, + 787, + 796 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Color and Mask Loss. To demonstrate the effectiveness of the color and silhouette rendering loss in our optimization process, we design baselines without RGB", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d3dd3bb0fb33cc515cf527d829bd5e0599372cd5876b0d1b73478989efb6619e.jpg", + "image_caption": [ + "Fig.5: Comparison with SMPL Body Prior. Only fitting SMPL to 2D observations, some joints in close contact such as arms are incorrectly estimated and even cause intersections between body surfaces. In contrast, our personalized prior enables accurate estimation of poses." + ], + "image_footnote": [], + "bbox": [ + 357, + 145, + 640, + 243 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e122d0cdce8f8952bf9b8458a6098bdf0bda3db112652379436061494496a72b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑
Ours (SMPL fitted)40.4195.1084.0494.37
Ours (w/o RGB loss)78.4083.2617.8474.55
Ours (w/o Mask loss)31.0097.3390.2899.06
Ours (Joint Optimization)66.0484.5622.5576.70
Ours29.3798.0296.7999.06
", + "bbox": [ + 254, + 327, + 746, + 444 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Quantitative Ablation Results. Ablations to evaluate our method with only the SMPL fitted method, our method without RGB loss and without Silhouette loss, and our method without alternating optimization.", + "bbox": [ + 212, + 444, + 787, + 484 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "loss and mask loss respectively for comparison. Table 3 shows that the RGB loss significantly improves pose optimization, and the optimization process will completely deviate from the correct trajectory in the absence of RGB loss. Mask loss is proven to slightly increase the accuracy by adding additional constraints on the rendered human silhouette.", + "bbox": [ + 212, + 529, + 787, + 604 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Alternating Optimization. We choose a joint optimization method widely used in avatar reconstructions [20, 26] to compare with our alternating optimization. Table 3 shows that our method significantly outperforms the chosen baseline. In Figure 6, the avatars from the baseline suffer from artifacts around contact body parts, leading to wrong 3D pose estimations. In contrast, our method can faithfully reconstruct avatars and poses with challenging initializations.", + "bbox": [ + 212, + 625, + 787, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Penetration Loss. Our penetration loss serves an important role in avoiding interpenetration. Comparing our method to the ablated version where we remove this loss, Figure 7 shows that the SMPL body of one person partially intersects with the other person in the contact area. This is also confirmed by the rendering result of the avatar. By penalizing the collision of density fields of avatars in 3D space, our method largely reduces the penetration and thus achieves more accurate pose estimation.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "AvatarPose", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d915227c893543113339fdb599925d70d6a87c5bcd1557216f2da5c3ecbd32da.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 388, + 143, + 457, + 227 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0df18bae9cc532ba8ad51ca35942a4aed5a706b0046f7284bdbbc8030b6f8175.jpg", + "image_caption": [ + "Joint Opt" + ], + "image_footnote": [], + "bbox": [ + 473, + 148, + 529, + 223 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8685c5c8784522317a84ddd4d1a089aab591005576eb640e17849a71fd03977d.jpg", + "image_caption": [ + "Alternating Opt" + ], + "image_footnote": [], + "bbox": [ + 549, + 148, + 606, + 223 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/bba9dfc4bb1ad3117e08a5e98679235867f5f088d6c17c473898d81ada6a75c3.jpg", + "image_caption": [ + "Fig. 6: Ablation of Alternating Optimization. We show the results of rendered avatars and projections of the estimated 3D poses. Joint optimization suffers from artifacts around the contact part and in turn causes wrong pose estimations. In contrast, ours reconstructs both avatars and poses correctly." + ], + "image_footnote": [], + "bbox": [ + 359, + 319, + 418, + 417 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ec8179686a182c6612bd46c4ca0af2eeff0589110c6949c0416b5904be9a2a1d.jpg", + "image_caption": [ + "Ours(w/o penetration loss)" + ], + "image_footnote": [], + "bbox": [ + 421, + 330, + 468, + 416 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d9326f997f7193d82750934f74f4a00b77c352b465a732e3698b4424a77ca6e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 330, + 526, + 415 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ce07464fb9f4705079d24143265206ab9f1d4637c7985beedcdd768668281084.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 527, + 330, + 584, + 415 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b784b2eb3781b8a547ea8005a2c557ac3e23d6b906255c0d595f072e04ad57f3.jpg", + "image_caption": [ + "Fig. 7: Ablation of Penetration Loss. Without penetration loss, both the avatar and underlying SMPL body tend to have collisions on surfaces." + ], + "image_footnote": [], + "bbox": [ + 586, + 330, + 633, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 214, + 501, + 504, + 515 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Despite greatly improving $98\\%$ of the sequences, our method shows minor improvements when the pose initialization is in the opposite direction or most joints are severely misaligned, leading to local minima in optimization. Moreover, we do not model hands in our avatar model, it will be a promising direction to integrate hand models [46,51] into the personalized avatar. More discussions about limitations and future work can be found in the Supp Mat.", + "bbox": [ + 212, + 527, + 787, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 641, + 359, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose AvatarPose, a novel method to estimate the 3D poses of multiple people in close interaction from sparse multi-view videos. Unlike previous methods leveraging SMPL body prior, we leverage the reconstructed avatars as personalized priors to guide pose optimization. The avatar prior enables us to use color and silhouette observations and to introduce a collision loss in close contact. Our method outperforms SotA methods significantly on public datasets of close human interactions.", + "bbox": [ + 212, + 672, + 787, + 779 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by the Swiss SERI Consolidation Grant \"AI-PERCEIVE\". We thank Wenbo Wang, Tianjian Jiang and Juan Zarate for providing us with helpful guidance and feedback.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: Scape: shape completion and animation of people. In: ACM SIGGRAPH 2005 Papers, pp. 408-416 (2005)", + "2. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures for multiple human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1669-1676 (2014)", + "3. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures for multiple human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1669-1676 (2014)", + "4. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures revisited: Multiple human pose estimation. IEEE transactions on pattern analysis and machine intelligence 38(10), 1929-1942 (2015)", + "5. Benzine, A., Chabot, F., Luvison, B., Pham, Q.C., Achard, C.: Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6856-6865 (2020)", + "6. Bhatnagar, B.L., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Combining implicit function learning and parametric models for 3d human reconstruction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 311-329. Springer (2020)", + "7. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 561-578. Springer (2016)", + "8. Burov, A., Nießner, M., Thies, J.: Dynamic surface function networks for clothed human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10754-10764 (2021)", + "9. Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7291-7299 (2017)", + "0. Chen, J., Zhang, Y., Kang, D., Zhe, X., Bao, L., Jia, X., Lu, H.: Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629 (2021)", + "1. Chen, L., Ai, H., Chen, R., Zhuang, Z., Liu, S.: Cross-view tracking for multi-human 3d pose estimation at over 100 fps. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3279-3288 (2020)", + "2. Chen, X., Zheng, Y., Black, M.J., Hilliges, O., Geiger, A.: Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11594-11604 (2021)", + "3. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7103-7112 (2018)", + "4. Chen, Y., Gu, R., Huang, O., Jia, G.: Vtp: volumetric transformer for multi-view multi-person 3d pose estimation. Applied Intelligence pp. 1-12 (2023)", + "5. Cheng, B., Xiao, B., Wang, J., Shi, H., Huang, T.S., Zhang, L.: Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5386-5395 (2020)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "AvatarPose", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Cheng, Y., Li, L., Xu, Y., Li, X., Yang, Z., Wang, W., Yang, Y.: Segment and track anything. arXiv preprint arXiv:2305.06558 (2023)", + "17. Choudhury, R., Kitani, K.M., Jeni, L.A.: Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14750-14760 (2023)", + "18. Dong, J., Fang, Q., Jiang, W., Yang, Y., Bao, H., Zhou, X.: Fast and robust multiperson 3d pose estimation and tracking from multiple views. In: T-PAMI (2021)", + "19. Dong, J., Jiang, W., Huang, Q., Bao, H., Zhou, X.: Fast and robust multi-person 3d pose estimation from multiple views. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7792-7801 (2019)", + "20. Dong, Z., Guo, C., Song, J., Chen, X., Geiger, A., Hilliges, O.: Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20470-20480 (2022)", + "21. Dong, Z., Song, J., Chen, X., Guo, C., Hilliges, O.: Shape-aware multi-person pose estimation from multi-view images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11158-11168 (2021)", + "22. Ershadi-Nasab, S., Noury, E., Kasaei, S., Sanaei, E.: Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications 77, 15573-15601 (2018)", + "23. Fabbri, M., Lanzi, F., Calderara, S., Alletto, S., Cucchiara, R.: Compressed volumetric heatmaps for multi-person 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7204-7213 (2020)", + "24. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2334-2343 (2017)", + "25. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7214-7223 (2020)", + "26. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12858-12868 (2023)", + "27. Guo, W., Bie, X., Alameda-Pineda, X., Moreno-Noguer, F.: Multi-person extreme motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13053-13064 (2022)", + "28. Hasler, N., Stoll, C., Sunkel, M., Rosenhahn, B., Seidel, H.P.: A statistical model of human pose and body shape. In: Computer graphics forum. vol. 28, pp. 337-346. Wiley Online Library (2009)", + "29. Jiang, T., Chen, X., Song, J., Hilliges, O.: Instantavatar: Learning avatars from monocular video in 60 seconds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16922-16932 (2023)", + "30. Joo, H., Liu, H., Tan, L., Gui, L., Nabbe, B., Matthews, I., Kanade, T., Nobuhara, S., Sheikh, Y.: Panoptic studio: A massively multiview system for social motion capture. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3334-3342 (2015)", + "31. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7122-7131 (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Kocabas, M., Huang, C.H.P., Hilliges, O., Black, M.J.: Pare: Part attention regressor for 3d human body estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11127-11137 (2021)", + "33. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: Spec: Seeing people in the wild with an estimated camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11035-11045 (2021)", + "34. Li, Y., Zhang, S., Wang, Z., Yang, S., Yang, W., Xia, S.T., Zhou, E.: Tokenpose: Learning keypoint tokens for human pose estimation. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 11313-11322 (2021)", + "35. Lin, J., Lee, G.H.: Multi-view multi-person 3d pose estimation with plane sweep stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11886-11895 (2021)", + "36. Liu, Y., Gall, J., Stoll, C., Dai, Q., Seidel, H.P., Theobalt, C.: Markerless motion capture of multiple characters using multiview image segmentation. IEEE transactions on pattern analysis and machine intelligence 35(11), 2720-2735 (2013)", + "37. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)", + "38. Luvizon, D.C., Habermann, M., Golyanik, V., Kortylewski, A., Theobalt, C.: Scene-aware 3d multi-human motion capture from a single camera. In: Computer Graphics Forum. vol. 42, pp. 371-383. Wiley Online Library (2023)", + "39. Martinez, J., Hossain, R., Romero, J., Little, J.J.: A simple yet effective baseline for 3d human pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2640-2649 (2017)", + "40. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H.P., Xu, W., Casas, D., Theobalt, C.: Vnect: Real-time 3d human pose estimation with a single rgb camera. Acm transactions on graphics (tog) 36(4), 1-14 (2017)", + "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "42. Muller, L., Osman, A.A.A., Tang, S., Huang, C.H.P., Black, M.J.: On self-contact and human pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9990-9999 (June 2021)", + "43. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "44. Osman, A.A., Bolkart, T., Black, M.J.: Star: Sparse trained articulated human body regressor. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16. pp. 598-613. Springer (2020)", + "45. Patel, P., Huang, C.H.P., Tesch, J., Hoffmann, D.T., Tripathi, S., Black, M.J.: Agora: Avatars in geography optimized for regression analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13468-13478 (2021)", + "46. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10975-10985 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "AvatarPose", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3d human pose and shape from a single color image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 459-468 (2018)", + "48. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021)", + "49. Pishchulin, L., Insafutdinov, E., Tang, S., Andres, B., Andriluka, M., Gehler, P.V., Schiele, B.: Deepcut: Joint subset partition and labeling for multi person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4929-4937 (2016)", + "50. Reddy, N.D., Guigues, L., Pishchulin, L., Eledath, J., Narasimhan, S.G.: Tessen-track: End-to-end learnable multi-person articulated 3d pose tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15190-15200 (2021)", + "51. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)", + "52. Saito, S., Yang, J., Ma, Q., Black, M.J.: Scanimate: Weakly supervised learning of skinned clothed avatar networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2886-2897 (2021)", + "53. Shuai, Q., Geng, C., Fang, Q., Peng, S., Shen, W., Zhou, X., Bao, H.: Novel view synthesis of human interactions from sparse multi-view videos. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22, Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10.1145/3528233.3530704, https://doi.org/10.1145/3528233.3530704", + "54. Shuai, Q., Yu, Z., Zhou, Z., Fan, L., Yang, H., Yang, C., Zhou, X.: Reconstructing close human interactions from multiple views. ACM Transactions on Graphics (dec 2023). https://doi.org/10.1145/3618336", + "55. Sun, Y., Bao, Q., Liu, W., Fu, Y., Black, M.J., Mei, T.: Monocular, one-stage, regression of multiple 3d people. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11179-11188 (2021)", + "56. Sun, Y., Liu, W., Bao, Q., Fu, Y., Mei, T., Black, M.J.: Putting people in their place: Monocular regression of 3d people in depth. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13243-13252 (2022)", + "57. Te, G., Li, X., Li, X., Wang, J., Hu, W., Lu, Y.: Neural capture of animatable 3d human from monocular video. In: European Conference on Computer Vision. pp. 275-291. Springer (2022)", + "58. Tu, H., Wang, C., Zeng, W.: Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 197-212. Springer (2020)", + "59. Wang, C., Li, J., Liu, W., Qian, C., Lu, C.: Hmor: Hierarchical multi-person ordinal relations for monocular multi-person 3d pose estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 242-259. Springer (2020)", + "60. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Wang, T., Zhang, J., Cai, Y., Yan, S., Feng, J.: Direct multi-view multi-person 3d pose estimation. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 13153-13164. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/6da9003b743b65f4c0ccd295cc484e57-Paper.pdf", + "62. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022)", + "63. Wu, S., Jin, S., Liu, W., Bai, L., Qian, C., Liu, D., Ouyang, W.: Graph-based 3d multi-person pose estimation using multi-view images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11148-11157 (2021)", + "64. Wu, S., Jin, S., Liu, W., Bai, L., Qian, C., Liu, D., Ouyang, W.: Graph-based 3d multi-person pose estimation using multi-view images. In: ICCV (2021)", + "65. Xu, H., Bazavan, E.G., Zanfir, A., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Ghum & ghuml: Generative 3d human shape and articulated pose models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6184-6193 (2020)", + "66. Ye, H., Zhu, W., Wang, C., Wu, R., Wang, Y.: Faster voxelpose: Real-time 3d human pose estimation by orthographic projection. In: European Conference on Computer Vision. pp. 142-159. Springer (2022)", + "67. Yin, Y., Guo, C., Kaufmann, M., Zarate, J.J., Song, J., Hilliges, O.: Hi4d: 4d instance segmentation of close human interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17016-17027 (2023)", + "68. Zanfir, A., Marinoiu, E., Sminchisescu, C.: Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2148-2157 (2018)", + "69. Zhang, H., Tian, Y., Zhou, X., Ouyang, W., Liu, Y., Wang, L., Sun, Z.: Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11446-11456 (2021)", + "70. Zhang, J., Liu, X., Ye, X., Zhao, F., Zhang, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG) 40(4), 1-18 (2021)", + "71. Zhang, Y., Wang, C., Wang, X., Liu, W., Zeng, W.: Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 2613-2626 (2022)", + "72. Zhang, Y., An, L., Yu, T., Li, X., Li, K., Liu, Y.: 4d association graph for realtime multi-person motion capture using multiple video cameras. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1324-1333 (2020)", + "73. Zhang, Y., Li, Z., An, L., Li, M., Yu, T., Liu, Y.: Lightweight multi-person total motion capture using sparse multi-view cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5560-5569 (2021)", + "74. Zheng, Y., Shao, R., Zhang, Y., Yu, T., Zheng, Z., Dai, Q., Liu, Y.: Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6239-6249 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "AvatarPose", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "75. Zhou, H., Hong, C., Han, Y., Huang, P., Zhuang, Y.: Mh pose: 3d human pose estimation based on high-quality heatmap. In: 2021 IEEE International Conference on Big Data (Big Data). pp. 3215-3222. IEEE (2021)", + "76. Zhou, Z., Shuai, Q., Wang, Y., Fang, Q., Ji, X., Li, F., Bao, H., Zhou, X.: Quickpose: Real-time multi-view multi-person pose estimation in crowded scenes. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 231 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "F. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_model.json b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b001acab5fc4feade95bc8cf2bc59bf54b46af77 --- /dev/null +++ b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_model.json @@ -0,0 +1,2924 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.234, + 0.141, + 0.773, + 0.21 + ], + "angle": 0, + "content": "AvatarPose: Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.234, + 0.746, + 0.252 + ], + "angle": 0, + "content": "Feichi Lu\\(^{1*}\\), Zijian Dong\\(^{1,2*}\\), Jie Song\\(^{1\\dagger}\\), and Otmar Hilliges\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.262, + 0.71, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Computer Science, ETH Zürich, Switzerland," + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.277, + 0.692, + 0.291 + ], + "angle": 0, + "content": "2 Max Planck Institute for Intelligent Systems, Germany" + }, + { + "type": "list", + "bbox": [ + 0.292, + 0.262, + 0.71, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.305, + 0.75, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.513 + ], + "angle": 0, + "content": "Fig. 1: We propose AvatarPose, a method for estimating the 3D poses and shapes of multiple closely interacting people from multi-view videos. To this end, we first reconstruct the avatar of each individual and leverage the learned personalized avatars as priors to refine poses via color and silhouette rendering loss. We alternate between avatar refinement and pose optimization to obtain the final pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.53, + 0.741, + 0.766 + ], + "angle": 0, + "content": "Abstract. Despite progress in human motion capture, existing multiview methods often face challenges in estimating the 3D pose and shape of multiple closely interacting people. This difficulty arises from reliance on accurate 2D joint estimations, which are hard to obtain due to occlusions and body contact when people are in close interaction. To address this, we propose a novel method leveraging the personalized implicit neural avatar of each individual as a prior, which significantly improves the robustness and precision of this challenging pose estimation task. Concretely, the avatars are efficiently reconstructed via layered volume rendering from sparse multi-view videos. The reconstructed avatar prior allows for the direct optimization of 3D poses based on color and silhouette rendering loss, bypassing the issues associated with noisy 2D detections. To handle interpenetration, we propose a collision loss on the overlapping shape regions of avatars to add penetration constraints. Moreover, both 3D poses and avatars are optimized in an alternating manner. Our experimental results demonstrate state-of-the-art performance on several public datasets." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.778, + 0.741, + 0.807 + ], + "angle": 0, + "content": "Keywords: human pose estimation \\(\\cdot\\) human close interaction \\(\\cdot\\) multiview pose estimation \\(\\cdot\\) avatar prior" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.811, + 0.363, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.442, + 0.84 + ], + "angle": 0, + "content": "† Now at HKUST(GZ)&HKUST" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.811, + 0.442, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.311 + ], + "angle": 0, + "content": "People frequently engage in interactions with each other in daily life to offer physical support or convey emotions. For AI systems to interpret 3D human interactions, a foundational step is to reconstruct 3D human poses and shapes using consumer-grade sensors, such as cameras. However, in closely interacting scenarios, such as hugging or kissing, 3D pose estimations face challenges due to strong occlusions. The problem of depth ambiguity worsens when close contact is involved, making it more difficult to predict 3D poses from 2D images or videos. Consequently, a multi-camera setup becomes essential to provide additional observations and to address depth ambiguity in pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Despite the ubiquity of close human interactions, the study of estimating such human motions is scarce. Most previous multi-human methods with a multi-view setup [2-4, 19, 22, 30, 58, 66, 72] focus on scenarios where people are at a distance from each other. Some methods [2, 4, 19, 22, 72] typically formulate the problem as a cross-view matching problem, relying heavily on 2D joint estimations for subsequent 3D triangulation. These methods demonstrate high sensitivity to noisy or missing 2D joint estimations, particularly when occlusion occurs. Another group of learning-based methods [58, 61, 63, 66] attempts to integrate 2D features from each view into a 3D voxel space and predict 3D human poses from identified 3D subvolumes of each individual. These methods are more robust to occlusion, but they encounter challenges with generalization and are sensitive to distribution shifts due to the lack of annotated 3D data. To correct abnormal or missing pose estimations, some methods [18, 21] leverage parametric body models like SMPL [37] as full-body priors and fit these 3D models to 2D joint estimations. Although this approach alleviates the issue of abnormal or missing joint estimations, it remains constrained by noisy 2D detections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Embracing the challenging problem of pose estimation for close interactions, we propose a novel method to estimate the 3D poses and shapes of multiple people observed from a sparse set of cameras. Our goal is to ensure that even in close contact, the estimated human poses and shapes are accurate and free from interpenetrations. The key to our method is reconstructing implicit textured avatars of each individual in the scene and leveraging them as a strong personalized prior for pose optimization (Figure 1). In contrast to relying solely on noisy 2D joint detections, this textured avatar prior enables us to leverage pixel-wise color and silhouette information for pose refinement, significantly increasing the robustness and accuracy of our method (Table 3). Meanwhile, the reconstructed avatar provides crucial geometric and appearance information to avoid collisions between individuals. Compared to methods using SMPL body shape [25,42] to penalize collisions, our implicit avatar model contains a more detailed clothed shape and additional appearance cues and enables efficient computation of penetration loss. Due to the mutually beneficial relationship between avatar and pose, we alternate between pose optimization and avatar refinement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "More specifically, to accelerate the learning of avatars, we model each human individual in canonical space using an efficient neural radiance field variant in Instant NGP [43] and combine it with an efficient SMPL-based deformation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AvatarPose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.403 + ], + "angle": 0, + "content": "module [37]. To learn and render avatar models of multiple people, we adapt layered volume rendering [53, 70] to our avatar model. This adaptation allows us to jointly optimize all avatar models through a straightforward rendering loss. Once learned, the avatar can be animated and rendered based on pose parameters at interactive rates, thus naturally serving as an efficient personalized textured prior for pose optimization. With the learned personalized prior, we optimize pose via a novel objective function. Different from previous methods based on 2D reprojection error of joints, we directly optimize pose parameters via minimizing color and silhouette rendering losses while keeping the learned avatar model fixed. To prevent interpenetration between human individuals, a collision loss is introduced by penalizing the situation when a 3D point is occupied by multiple avatars. To remove artifacts in the initial avatar due to imperfect pose initialization, we further refine avatars based on optimized poses. Throughout the optimization process, both initial personalized avatar models and SMPL parameters are optimized in an alternating manner, motivated by the insight that accurate 3D pose estimations improve avatar learning, and improved avatar models, in turn, increase the precision of overall pose estimations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.464 + ], + "angle": 0, + "content": "We experimentally demonstrate that our method significantly outperforms previous state-of-the-art methods on several public datasets both quantitatively (Table 1 and Table 2) and qualitatively (Figure 3) especially when people are in close interaction. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.474, + 0.784, + 0.517 + ], + "angle": 0, + "content": "- We propose a pipeline that efficiently creates implicit neural avatars of closely interacting people and leverages the learned avatars as priors to optimize poses." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.519, + 0.784, + 0.575 + ], + "angle": 0, + "content": "- The avatar prior enables us to design a novel objective function that leverages color and mask rendering loss for pose optimization. We show the superiority of this loss function compared to the 2D reprojection error of 3D joints used by most of the previous methods." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.58, + 0.784, + 0.605 + ], + "angle": 0, + "content": "- Based on the learned avatar, a collision loss is introduced to avoid penetration when individuals are in contact." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.474, + 0.784, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.388, + 0.649 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.54, + 0.679 + ], + "angle": 0, + "content": "2.1 Multi-Person 3D Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Despite significant progress in multi-human 2D pose estimation [9,13,15,24,34, 49] and 3D pose estimation from monocular images or video [7, 31-33, 38-40, 45, 47, 55, 56, 59, 68, 69], the reconstruction accuracy is still limited due to depth ambiguity and strong occlusions when humans are in close contact. A multi-view setting [4, 5, 19, 22, 23, 35, 36, 54, 61, 63, 66, 72] helps to alleviate these challenges. One straightforward idea of most methods [2, 4, 19, 22, 72, 76] is to formulate the problem into cross-view matching and association problems. MVPose [19] performs 2D person parsing in each image and leverages cross-view person matching to infer 3D pose. 4DAssociation [72] additionally adds tracking into this process to form a unified graph for associating 4D information. However, these methods" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "are sensitive to the noisy estimation of 2D pose. In contrast to these matching-based methods, some recent methods [5,14,17,23,35,50,58,61,63,66,71,75] directly learn deep neural networks to regress poses. Faster Voxelpose [66] employs the feature volume proposed by Voxelpose [58] and enhances computational efficiency. Graph [63] designs three graph neural network models for human center detection and pose estimation. MvP [61] simplifies the multi-person pose estimation by direct regression using the transformer model. A concurrent method CloseMocap [54] proposes to learn a model from a synthetic dataset simulating occlusion situations. However, heavily relying on the 2D or 3D features as input during training, these learning-based methods suffer from generalization issues when subjects, motions, and camera configurations change." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.45 + ], + "angle": 0, + "content": "To further improve the robustness, a statistical parametric body model such as SMPL [37] is explored in [21] as a regularization prior for 3D joint refinement. Some follow-ups [18,73] show that parametric models help in correcting implausible 3D pose estimates and filling in missing joints. However, this coarse body prior highly relies on aligning the 3D joints to 2D pose estimations, which are inaccurate when occlusion happens. Different from all of these previous methods, we explore the usage of personalized textured avatar models as priors to refine human poses. This prior enables us to leverage color and silhouette information from multi-view observation for refining poses of closely interacting humans." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.47, + 0.439, + 0.485 + ], + "angle": 0, + "content": "2.2 3D Human Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.751 + ], + "angle": 0, + "content": "Parametric human body models [1,28,37,44,65] can represent minimally clothed human shapes by deforming a template mesh. It is challenging to extend this explicit representation for modeling clothed humans due to the fixed topology and resolution. To overcome this limitation, methods such as SNARF [12] and SCANimate [52] propose to model articulated human shapes based on 3D implicit representations. Many works [6,8,10,20,26,29,48,57,60,62] fit implicit neural fields to RGB or RGB-D videos by neural rendering to reconstruct the shape and appearance of a single human body. However, when applied to a multi-human scene, these methods are not able to achieve good fidelity due to strong occlusions. Recent methods including ST-NeRF [70] and [53] leverage layered neural representation to model multiple humans with sparse multi-view videos and thus can generate novel view synthesis of dynamic multiple humans. The main problem of all aforementioned approaches, however, is their high reliance on 3D human pose estimation [70,74]: the deformation of the human model requires accurate human poses, which is hard to obtain when people are in close interaction. To address this challenge, our method is orthogonal to others, aiming at leveraging the learned avatar as priors for pose estimation in close interaction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.608, + 0.787 + ], + "angle": 0, + "content": "2.3 3D Human Datasets for Close Interaction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Most of the existing datasets [3,11,30] like Shelf and Panoptic studying multiperson pose estimation focus on the scene where people are at a distance from each other and rarely involved in close interactions. To study close interactions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AvatarPose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "among people, MultiHuman [74] dataset captures multi-person interaction with some close interactions and occlusions. ExPI [27] creates a multi-person extreme motion dataset with close interactions, but it focuses mainly on motion prediction for future frames instead of pose estimation from sparse views. CHI3D [25] captures two-person interaction datasets and proposes to learn a contact estimation module from annotations to improve the precision of pose estimation. The most recent work Hi4D [67] creates a challenging dataset of physically close human interaction and proposes a method to disentangle human bodies and estimate poses. However, this method relies on the 3D ground truth of clothed human meshes captured with expensive 3D body scanners. Due to the reliance on the limited annotated 3D data, this method faces challenges in generalization with different people and camera configurations. In contrast to all the works before, we intend to solve the problem of pose estimation in close human interaction without requiring accurate 3D scans or other training data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.331, + 0.399 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.506 + ], + "angle": 0, + "content": "Given a dynamic scene captured by a sparse set of RGB cameras, our goal is to estimate the 3D pose and shape of multiple people even if they interact closely. To address this challenging task, our key idea is to first reconstruct the personalized avatar of each individual in the scene and leverage them as a strong prior to refine the appearance and pose in an alternating manner. An overview of our method is shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.597 + ], + "angle": 0, + "content": "We first introduce an efficient pipeline to create avatars of multiple people in a scene (Section 3.1 and Figure 2(a)). Specifically, we leverage an accelerated neural radiance field to represent the shape and appearance of each individual in canonical space and deform it at an interactive rate. We then adapt layered volume rendering to our pipeline, which composites the rendering of avatars into one image, thus enabling direct learning from multi-view video inputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Thanks to the learned avatar prior for each individual, we can enhance 3D pose optimization via a combination of RGB and silhouette rendering loss (Section 3.2 and Figure 2(b)). While previous work heavily relies on noisy 2D joint detection, we show that employing such pixel-wise color and silhouette information can largely increase precision and robustness. Moreover, a collision loss is introduced to avoid interpenetration. Finally, we alternate between avatar learning and pose optimization to get complete and accurate 3D human poses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.498, + 0.741 + ], + "angle": 0, + "content": "3.1 Multi-Avatar Prior Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.752, + 0.787, + 0.799 + ], + "angle": 0, + "content": "Avatar Model We represent each human individual in canonical space using an accelerated neural radiance field [43] and model shape-aware articulated deformation based on SMPL [37]." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.808, + 0.788, + 0.843 + ], + "angle": 0, + "content": "- Canonical Appearance Representation: To model human shape and appearance, we create canonical radiance field \\(\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}\\) for each human instance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.146, + 0.761, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.276, + 0.77, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.405, + 0.789, + 0.588 + ], + "angle": 0, + "content": "Fig. 2: Method Overview: Our method consists of two modules: (a) Multi-Avatar Prior Learning: Given the input multi-view images and estimated poses \\(\\Theta^{(l)}\\), we sample points \\(\\mathbf{x}^{(l)}\\) for each individual \\(l\\) along the rays in posed space and warp these points into canonical space and calculate their color \\(\\mathbf{c}^{(l)}\\) and density \\(\\sigma^{(l)}\\) via the canonical appearance network \\(\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}\\). We leverage layered volume rendering [70] to attain the final pixel color and compare it with the original input image to optimize the parameters of avatars. (b) Avatar-guided Pose Optimization: Given learned avatar model \\(\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}\\) and initial poses \\(\\Theta^{(l)}\\) of each individual \\(l\\), we deform all of the avatars based on SMPL-based deformer and render them jointly via layered volume rendering. We compare the composite rendering with input observation and minimize the RGB and mask rendering loss to optimize poses. A collision loss is additionally introduced to avoid interpenetration. Finally, we alternate between two modules to obtain the final result. For clarity, the parameters to be optimized are marked as red in each module." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.619, + 0.788, + 0.685 + ], + "angle": 0, + "content": "\\(l \\in [1, L]\\), where \\(L\\) is the number of humans in the scene. \\(\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}\\) takes a 3D point \\(\\bar{\\mathbf{x}}^{(l)}\\) as input and predicts its density \\(\\sigma^{(l)}\\) and color \\(\\mathbf{c}^{(l)}\\). Following Instant-NGP [43] and InstantAvatar [29] to accelerate the rendering, \\(\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}\\) is parameterized via using a hash table to store feature grids at different scales." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.687, + 0.788, + 0.748 + ], + "angle": 0, + "content": "- Pose Representation: We represent the 3D pose and underlying body shape for all human instances by SMPL parameters \\(\\Theta = \\{\\Theta^{(l)}\\}_{l\\in [1,L]}\\). For each human \\(l\\), \\(\\Theta^{(l)} = \\{\\beta^{(l)},\\theta^{(l)},t^{(l)}\\}\\) contains shape parameters \\(\\beta^{(l)}\\in \\mathbb{R}^{10}\\), pose parameters \\(\\theta^{(l)}\\in \\mathbb{R}^{72}\\) and translation \\(t^{(l)}\\in \\mathbb{R}^3\\) of SMPL." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "- Deformer: To enable animation given targeted poses \\(\\Theta^{(l)}\\), we require the radiance field in the posed space. Given a point \\(\\mathbf{x}^{(l)}\\) in deformed space of human \\(l\\), we determine the corresponding canonical point \\(\\bar{\\mathbf{x}}^{(l)}\\) by inverse linear blend skinning(LBS) [37]: \\(\\bar{\\mathbf{x}}^{(l)}(\\mathbf{x}^{(l)},\\Theta^{(l)}) = (\\sum_{i = 1}^{n_b}w_i(\\Theta^{(l)})\\mathbf{B}_i(\\Theta^{(l)}))^{-1}\\mathbf{x}^{(l)}\\), where \\(\\mathbf{B}_i\\) is the rigid bone transformation matrix for joint \\(i\\in \\{1,\\dots,n_b\\}\\) under pose \\(\\Theta^{(l)}\\). \\(w_{i}\\) is the skinning weights of the nearest neighbor of \\(\\mathbf{x}^{(l)}\\) in" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.687, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AvatarPose" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "the deformed SMPL vertices. We obtain the radiance field at the point \\(\\mathbf{x}^{(l)}\\) by evaluating the canonical appearance field at the corresponding point \\(\\bar{\\mathbf{x}}^{(l)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.191, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Layered Volume Rendering To obtain the pixel value for a ray \\(\\mathbf{r} \\in \\mathcal{R}\\), we raycast every human instance separately with a layered rendering strategy similar to ST-NeRF [70]. Specifically, we first calculate the intersection points between the ray and the 3D bounding box of each human instance and uniformly sample points in each bounding box. To distinguish different identities, we assign each sampled point \\(\\mathbf{x}_i\\) a one-hot representation \\(\\mathbf{m}_i = [m_i^{(1)}, \\dots, m_i^{(L)}]\\) to indicate which human identity it belongs to. After sorting all sampled points by their depth values and calculating their corresponding color \\(\\mathbf{c}_i\\) and density \\(\\sigma_i\\) from the avatar model, if \\(m_i^{(l)} = 1\\), we compute" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.342, + 0.786, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {i}, \\sigma_ {i} = \\bar {\\mathbf {F}} _ {\\sigma_ {f}} ^ {(l)} \\left(\\bar {\\mathbf {x}} _ {i} \\left(\\mathbf {x} _ {i}, \\boldsymbol {\\Theta} ^ {(l)}\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.371, + 0.706, + 0.387 + ], + "angle": 0, + "content": "The color of each ray is computed via numerical integration [41]." + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.399, + 0.785, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {C}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\prod_ {j < i} (1 - \\alpha_ {j}) \\mathbf {c} _ {i} \\quad \\alpha_ {i} = 1 - \\exp (- \\sigma_ {i} \\delta_ {i}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.449, + 0.784, + 0.479 + ], + "angle": 0, + "content": "where \\(\\delta_{i}\\) is the distance between samples. The accumulated alpha value, which represents ray opacity, can be computed via:" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.491, + 0.785, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\alpha (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\prod_ {j < i} (1 - \\alpha_ {j}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.542, + 0.784, + 0.571 + ], + "angle": 0, + "content": "For each human identity \\( l \\in [1, L] \\), the corresponding instance ray opacity can be calculated via:" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.581, + 0.785, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\alpha^ {(l)} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\prod_ {j < i} (1 - \\alpha_ {j}) m _ {i} ^ {(l)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.635, + 0.784, + 0.681 + ], + "angle": 0, + "content": "Training The overall training process is shown in Figure 2(a). For training avatar layers, we minimize the Huber loss \\(\\rho\\) between the predicted pixel color \\(\\hat{\\mathbf{C}}(\\mathbf{r})\\) and the ground truth pixel color \\(\\mathbf{C}_{gt}(\\mathbf{r})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.693, + 0.785, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R G B} = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} \\rho \\left(\\left\\| \\hat {\\mathbf {C}} (\\mathbf {r}) - \\mathbf {C} _ {g t} (\\mathbf {r}) \\right\\|\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.737, + 0.784, + 0.797 + ], + "angle": 0, + "content": "Since instance segmentation of human performers is hard to obtain and is not accurate, we choose foreground segmentation as our mask supervision, which is obtained via SAM-Track [16]. We apply a loss for optimizing the rendered alpha values \\(\\alpha\\) to reduce the artifacts in the floating area:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.809, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l p h a}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} \\in \\mathcal {R}} (\\alpha (\\mathbf {r}) - \\alpha_ {S A M} (\\mathbf {r})) ^ {2}. \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Following [53], we add a regularization loss for instance alpha values to make sure every pixel can only be rendered from one human layer:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.19, + 0.786, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l a y e r}} = - \\frac {1}{L \\mid \\mathcal {R} \\mid} \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\sum_ {l = 1} ^ {L} \\alpha^ {(l)} (\\mathbf {r}) \\log \\left(\\alpha^ {(l)} (\\mathbf {r})\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.237, + 0.784, + 0.268 + ], + "angle": 0, + "content": "Similar to [29], we also add hard surface and density regularization terms in the learning process. More training details can be found in the Supp Mat." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.29, + 0.536, + 0.306 + ], + "angle": 0, + "content": "3.2 Avatar-guided Pose Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.785, + 0.452 + ], + "angle": 0, + "content": "Equipped with learned avatars obtained in Section 3.1, we aim to estimate the 3D shape and pose of multiple humans with close physical contact. To achieve this, we leverage the avatars as priors to handle the challenging pose ambiguities caused by contact. Specifically, we first initialize pose parameters using an off-the-shelf 3D pose estimator [72] and refine the pose via a rendering loss between rendered posed avatars and 2D observations (Figure 2(b)). Since the initial imperfect pose estimations may cause artifacts in avatar reconstruction, we further refine the weights of the avatar model using the optimized pose. This process is formulated as an alternating optimization to refine both poses and avatars." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.785, + 0.542 + ], + "angle": 0, + "content": "Initialization The initial 3D pose proposals are estimated by leveraging the off-the-shelf 3D human pose estimator [72]. After that, we register a SMPL model to the estimated 3D joints to obtain the initial pose parameters \\(\\Theta_0\\). Given these estimated poses, the initial avatar model of each individual is further learned from multi-view videos." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.557, + 0.784, + 0.589 + ], + "angle": 0, + "content": "Objective To leverage avatars as priors to tackle challenges caused by contact, we optimize the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.6, + 0.786, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\boldsymbol {\\Theta}) = \\lambda_ {R G B} \\mathcal {L} _ {R G B} (\\boldsymbol {\\Theta}) + \\lambda_ {\\alpha} \\mathcal {L} _ {\\alpha} (\\boldsymbol {\\Theta}) \\tag {8} \\\\ + \\lambda_ {r e g} \\mathcal {L} _ {r e g} (\\Theta) + \\lambda_ {p a} \\mathcal {L} _ {p a} (\\Theta). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.785, + 0.72 + ], + "angle": 0, + "content": "Here, \\(\\Theta\\) is the SMPL parameters to be optimized for all avatars, which is also consistent with the pose parameters to deform the avatar model (Section 3.1). Following [62], we represent \\(\\Theta\\) as \\(\\Theta_0 + \\mathrm{MLP}(\\Theta_0)\\) and refine poses by changing the parameters of the neural network. This representation empirically shows more robust results compared to directly optimizing SMPL parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.721, + 0.785, + 0.798 + ], + "angle": 0, + "content": "Different from previous methods [18,21,58,72], the personalized multi-ayar prior allows us to leverage appearance and silhouette information to refine initial poses. Specifically, we calculate \\(\\mathcal{L}_{RGB}\\) to ensure the color consistency between the rendered pixel \\(\\hat{\\mathbf{C}} (\\mathbf{r},\\boldsymbol {\\Theta})\\) (Equation (1) and Equation (2)) of deformed avatar with poses \\(\\Theta\\) and the corresponding ground-truth pixel color \\(\\mathbf{C}_{gt}(\\mathbf{r})\\)." + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.808, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R G B} (\\boldsymbol {\\Theta}) = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} \\rho \\left(\\left\\| \\hat {\\mathbf {C}} (\\mathbf {r}, \\boldsymbol {\\Theta}) - \\mathbf {C} _ {g t} (\\mathbf {r}) \\right\\|\\right). \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AvatarPose" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "Additionally, a cross-entropy loss \\(\\mathcal{L}_{\\alpha}\\) is introduced to ensure that the rendered mask \\(\\alpha (\\mathbf{r},\\Theta)\\) (Equation (3)) of the reposed avatar is aligned with the estimated SAM-Track mask \\(\\alpha_{SAM}(\\mathbf{r})\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.205, + 0.786, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\alpha} (\\boldsymbol {\\Theta}) = - \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\alpha_ {S A M} (\\mathbf {r}) \\log (\\alpha (\\mathbf {r}, \\boldsymbol {\\Theta})) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.784, + 0.286 + ], + "angle": 0, + "content": "To penalize the unnatural poses and avoid elbows and knees bending in the wrong direction, we add an L2 regularization term and combine it with the pose prior in SMPLify [7] to constrain physically implausible joint rotation:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.295, + 0.786, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} (\\boldsymbol {\\Theta}) = \\| \\boldsymbol {\\Theta} \\| _ {2} + \\lambda \\sum_ {i \\in I} \\exp (\\boldsymbol {\\Theta} _ {i}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.333, + 0.735, + 0.348 + ], + "angle": 0, + "content": "where \\(I\\) is the set of pose indices corresponding to elbows and knees." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.785, + 0.444 + ], + "angle": 0, + "content": "A key challenge to correctly estimate poses in close interaction is to handle interpenetration. Since every avatar is modeled separately, the surfaces tend to intersect when they are in contact. To handle this, we first select sampled points inside multiple instances as \\(\\mathcal{S} = \\{\\mathbf{x}_i\\mid \\alpha_i^{(p)} > 0,\\alpha_i^{(q)} > 0,p,q\\in [1,L],p\\neq q\\}\\) (\\(\\alpha_{i}^{(p)}\\) is calculated from Equation (1) and Equation (2) corresponding with a point \\(\\mathbf{x}_i\\) with \\(m_i^{(p)} = 1\\)). We then propose a collision loss \\(\\mathcal{L}_{pa}\\) for penalizing penetration:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.455, + 0.786, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p a} (\\boldsymbol {\\Theta}) = \\frac {1}{| \\mathcal {S} |} \\sum_ {\\boldsymbol {x} _ {i} \\in \\mathcal {S}} \\alpha_ {i} ^ {(p)} (\\boldsymbol {\\Theta}) \\alpha_ {i} ^ {(q)} (\\boldsymbol {\\Theta}) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.784, + 0.539 + ], + "angle": 0, + "content": "Intuitively, this loss guarantees every sample point in 3D space can not be occupied by multiple avatar models simultaneously, which guarantees better pose estimation in close contact." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.785, + 0.627 + ], + "angle": 0, + "content": "Alternating Optimization Since artifacts sometimes appear on initial avatars due to imperfect pose initialization, we further refine avatars based on optimized poses via minimizing the loss function in Section 3.1. Finally, the optimization of poses and avatars is formulated in an alternating fashion for \\( N \\) steps. More details of optimization can be found in the Supp Mat." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.376, + 0.667 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.787, + 0.753 + ], + "angle": 0, + "content": "Datasets. We mainly evaluate our proposed method on Hi4D [67] and CHI3D [25], which are challenging datasets of two humans in close interaction. To demonstrate the generalization ability of our method for more than two people, we also evaluate on Shelf [3] and MultiHuman [74] including three or four people. More details are shown in the Supp Mat." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Metrics. We use the Mean Per Joint Position Error (MPJPE) [58] to measure the distance between the ground truth 3D poses and the estimated poses. We also choose the Percentage of Correct Parts (PCP3D) metric [19] to calculate the percentage of correct parts. \\(\\mathrm{AP}_K\\) [58] and Recall [21] are also leveraged to evaluate performance. More details are shown in the Supp Mat." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.787, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.459 + ], + "angle": 0, + "content": "Fig. 3: Qualitative Comparison with SotA methods [19,61,64,66,72] on Hi4D and CHI3D. We show two examples from the Hi4D and CHI3D datasets compared with Graph, MvP, Faster VoxelPose, MVPose, and 4DAssociation. For each example, we show 2D projections on two sampled views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Baselines. We compare our method on the Hi4D dataset with state-of-the-art methods in three categories discussed in Section 2.1. For learning-based methods, we choose Graph [64], MvP [61], and Faster VoxelPose [66] and fine-tune these models on a subset of the evaluating datasets. For pure association-based methods, we choose 4DAssociation [72] and MVPose* [19]. MVPose [18] adds temporal tracking and SMPL prior to MVPose* [19] and is regarded as a SMPL-guided method. More details about baselines can be found in the Supp Mat." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.433, + 0.64 + ], + "angle": 0, + "content": "4.1 Comparison to SotA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Table 1 and Table 2 summarizes our quantitative comparisons on Hi4D and CHI3D with SotA (State-of-the-art) methods. Our method largely outperforms the other SotA methods in all of the metrics including MPJPE, PCP3D, and \\(\\mathrm{AP}_K\\). More comparisons on Shelf and MultiHuman are shown in the Supp Mat." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Comparison with Graph [63], MvP [61], Faster VoxelPose [66]. Our method outperforms Graph, MvP, and Faster VoxelPose on Hi4D and CHI3D. These methods are prone to overfitting to training pose distributions, thus struggling with challenging poses of interacting actors. Specifically, Faster VoxelPose sometimes misses actors in close contact, leading to a relatively low recall. While Graph and MvP have better recall, Graph fails to consistently track actors across frames, and MvP results in many misaligned joints between actors." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AvatarPose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.144, + 0.773, + 0.274 + ], + "angle": 0, + "content": "
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑Recall(%) ↑
MvP [61]92.7774.1441.5963.8693.84
Graph [63]89.6271.5544.7567.3393.31
Faster VoxelPose [66]68.4073.6744.0568.7083.55
MVPose* [19]53.0587.5767.9780.2893.80
MVPose [18]42.6390.7671.7990.1993.30
4DAssociation [72]41.2988.6280.8797.2798.78
Ours32.1096.9091.4897.3398.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.276, + 0.787, + 0.332 + ], + "angle": 0, + "content": "Table 1: Quantitative Comparison with SotA on the Hi4D [67] Dataset (8 views). We compare our method with MvP [61], Graph [63], Faster VoxelPose [66], MVPose* [19], MVPose [18] and 4DAssociation [72]. We report MPJPE, PCP, \\(\\mathrm{AP}_K\\), and Recall metric for all methods." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.369, + 0.787, + 0.506 + ], + "angle": 0, + "content": "
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑Recall(%) ↑
MvP [61]55.3889.4763.5892.5399.06
Graph [63]45.3392.4674.0295.2599.17
Faster VoxelPose [66]67.8178.4129.2882.8893.34
MVPose* [19]50.4290.3969.1375.7288.72
MVPose [18]34.0593.3579.9486.9188.18
4DAssociation [72]37.4799.3089.6698.6799.85
Ours32.9899.7993.2099.7999.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.507, + 0.788, + 0.561 + ], + "angle": 0, + "content": "Table 2: Quantitative Comparison with SotA on the CHI3D [25] Dataset (4 views). We compare our method with Faster VoxelPose [66], MVPose* [19], MV-Pose [18] and 4DAssociation [72]. We report MPJPE, PCP, \\(\\mathrm{AP}_K\\), and Recall metric for all methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.787, + 0.707 + ], + "angle": 0, + "content": "Comparison with MVPose [18]. Our method achieves much better MPJPE and precision than MVPose. MVPose relies heavily on noisy 2D joint detected in close interactions. In contrast, our method leverages color and silhouette rendering loss to optimize poses, leading to robustness to occlusions. Additionally, this top-down method cannot detect actors correctly with strong occlusions, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Comparison with 4DAssociation [72]. Finally, we compare our method with the bottom-up association method. As shown in Table 1 and Table 2, our method outperforms 4DAssociation in most metrics. Figure 3 shows that when actors are close, this bottom-up method is inclined to associate joints with the wrong human instances. This is because they solve the joint association with a greedy algorithm, which is sensitive to missing and inaccurate 2D joint detections." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.403, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.243, + 0.321, + 0.256 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.147, + 0.59, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.243, + 0.511, + 0.256 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.148, + 0.782, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.244, + 0.699, + 0.257 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.259, + 0.403, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.354, + 0.321, + 0.368 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.259, + 0.59, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.354, + 0.512, + 0.368 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.259, + 0.781, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.354, + 0.701, + 0.368 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.38, + 0.788, + 0.45 + ], + "angle": 0, + "content": "Fig. 4: Qualitative Results of our method on Hi4D (a)(b), CHI3D (c)(d), MultiHuman Real-Cap (e), and Shelf (f). The left and middle columns in each sub-figure show the 2D projections of the estimated 3D skeletons and SMPL body meshes on two views. The right column in each sub-figure demonstrates skeletons and SMPL bodies in 3D scenes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.52, + 0.496 + ], + "angle": 0, + "content": "4.2 Additional Qualitative Samples" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Figure 4 demonstrates more qualitative results of our method on Hi4D, CHI3D, and MultiHuman Real-Cap with challenging and close interactions among 2 or 3 people. We also demonstrate results on the Shelf [3] dataset, which contains 4 people without close contact. More results can be found in the Supp Mat." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.39, + 0.604 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.659 + ], + "angle": 0, + "content": "To validate the effectiveness of our method, we conduct a detailed analysis of different design choices of our algorithm. All the experiments are conducted on the Hi4D Dataset." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.788, + 0.797 + ], + "angle": 0, + "content": "Comparison with SMPL Body Prior. To validate the effectiveness of our personalized avatar prior, we compare with a baseline that optimizes SMPL parameters to align 3D joint reprojections to 2D observations. Our method significantly outperforms the SMPL prior baseline (Table 3) by leveraging color and silhouette rendering loss to refine poses, reducing reliance on inaccurate joint detections when occlusion happens. Figure 5 shows that while the baseline method incorrectly estimates the arm pose and even leads to penetration, our method reconstructs the poses accurately." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Color and Mask Loss. To demonstrate the effectiveness of the color and silhouette rendering loss in our optimization process, we design baselines without RGB" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AvatarPose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.146, + 0.642, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.312 + ], + "angle": 0, + "content": "Fig.5: Comparison with SMPL Body Prior. Only fitting SMPL to 2D observations, some joints in close contact such as arms are incorrectly estimated and even cause intersections between body surfaces. In contrast, our personalized prior enables accurate estimation of poses." + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.328, + 0.747, + 0.445 + ], + "angle": 0, + "content": "
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑
Ours (SMPL fitted)40.4195.1084.0494.37
Ours (w/o RGB loss)78.4083.2617.8474.55
Ours (w/o Mask loss)31.0097.3390.2899.06
Ours (Joint Optimization)66.0484.5622.5576.70
Ours29.3798.0296.7999.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Table 3: Quantitative Ablation Results. Ablations to evaluate our method with only the SMPL fitted method, our method without RGB loss and without Silhouette loss, and our method without alternating optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.605 + ], + "angle": 0, + "content": "loss and mask loss respectively for comparison. Table 3 shows that the RGB loss significantly improves pose optimization, and the optimization process will completely deviate from the correct trajectory in the absence of RGB loss. Mask loss is proven to slightly increase the accuracy by adding additional constraints on the rendered human silhouette." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Alternating Optimization. We choose a joint optimization method widely used in avatar reconstructions [20, 26] to compare with our alternating optimization. Table 3 shows that our method significantly outperforms the chosen baseline. In Figure 6, the avatars from the baseline suffer from artifacts around contact body parts, leading to wrong 3D pose estimations. In contrast, our method can faithfully reconstruct avatars and poses with challenging initializations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Penetration Loss. Our penetration loss serves an important role in avoiding interpenetration. Comparing our method to the ablated version where we remove this loss, Figure 7 shows that the SMPL body of one person partially intersects with the other person in the contact area. This is also confirmed by the rendering result of the avatar. By penalizing the collision of density fields of avatars in 3D space, our method largely reduces the penetration and thus achieves more accurate pose estimation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.145, + 0.459, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.229, + 0.43, + 0.237 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.474, + 0.15, + 0.531, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.228, + 0.525, + 0.237 + ], + "angle": 0, + "content": "Joint Opt" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.15, + 0.607, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.228, + 0.612, + 0.238 + ], + "angle": 0, + "content": "Alternating Opt" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Fig. 6: Ablation of Alternating Optimization. We show the results of rendered avatars and projections of the estimated 3D poses. Joint optimization suffers from artifacts around the contact part and in turn causes wrong pose estimations. In contrast, ours reconstructs both avatars and poses correctly." + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.32, + 0.419, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.332, + 0.47, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.421, + 0.537, + 0.432 + ], + "angle": 0, + "content": "Ours(w/o penetration loss)" + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.331, + 0.527, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.332, + 0.585, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.331, + 0.634, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.421, + 0.601, + 0.431 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.473 + ], + "angle": 0, + "content": "Fig. 7: Ablation of Penetration Loss. Without penetration loss, both the avatar and underlying SMPL body tend to have collisions on surfaces." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.502, + 0.506, + 0.516 + ], + "angle": 0, + "content": "4.4 Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.62 + ], + "angle": 0, + "content": "Despite greatly improving \\(98\\%\\) of the sequences, our method shows minor improvements when the pose initialization is in the opposite direction or most joints are severely misaligned, leading to local minima in optimization. Moreover, we do not model hands in our avatar model, it will be a promising direction to integrate hand models [46,51] into the personalized avatar. More discussions about limitations and future work can be found in the Supp Mat." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.642, + 0.36, + 0.658 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.78 + ], + "angle": 0, + "content": "In this paper, we propose AvatarPose, a novel method to estimate the 3D poses of multiple people in close interaction from sparse multi-view videos. Unlike previous methods leveraging SMPL body prior, we leverage the reconstructed avatars as personalized priors to guide pose optimization. The avatar prior enables us to use color and silhouette observations and to introduce a collision loss in close contact. Our method outperforms SotA methods significantly on public datasets of close human interactions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by the Swiss SERI Consolidation Grant \"AI-PERCEIVE\". We thank Wenbo Wang, Tianjian Jiang and Juan Zarate for providing us with helpful guidance and feedback." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AvatarPose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: Scape: shape completion and animation of people. In: ACM SIGGRAPH 2005 Papers, pp. 408-416 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.255 + ], + "angle": 0, + "content": "2. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures for multiple human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1669-1676 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.256, + 0.785, + 0.296 + ], + "angle": 0, + "content": "3. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures for multiple human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1669-1676 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.785, + 0.337 + ], + "angle": 0, + "content": "4. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures revisited: Multiple human pose estimation. IEEE transactions on pattern analysis and machine intelligence 38(10), 1929-1942 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Benzine, A., Chabot, F., Luvison, B., Pham, Q.C., Achard, C.: Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6856-6865 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.445 + ], + "angle": 0, + "content": "6. Bhatnagar, B.L., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Combining implicit function learning and parametric models for 3d human reconstruction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 311-329. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.514 + ], + "angle": 0, + "content": "7. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 561-578. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.555 + ], + "angle": 0, + "content": "8. Burov, A., Nießner, M., Thies, J.: Dynamic surface function networks for clothed human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10754-10764 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "9. Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7291-7299 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.785, + 0.622 + ], + "angle": 0, + "content": "0. Chen, J., Zhang, Y., Kang, D., Zhe, X., Bao, L., Jia, X., Lu, H.: Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "1. Chen, L., Ai, H., Chen, R., Zhuang, Z., Liu, S.: Cross-view tracking for multi-human 3d pose estimation at over 100 fps. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3279-3288 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.718 + ], + "angle": 0, + "content": "2. Chen, X., Zheng, Y., Black, M.J., Hilliges, O., Geiger, A.: Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11594-11604 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "3. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7103-7112 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "4. Chen, Y., Gu, R., Huang, O., Jia, G.: Vtp: volumetric transformer for multi-view multi-person 3d pose estimation. Applied Intelligence pp. 1-12 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Cheng, B., Xiao, B., Wang, J., Shi, H., Huang, T.S., Zhang, L.: Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5386-5395 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Cheng, Y., Li, L., Xu, Y., Li, X., Yang, Z., Wang, W., Yang, Y.: Segment and track anything. arXiv preprint arXiv:2305.06558 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.216 + ], + "angle": 0, + "content": "17. Choudhury, R., Kitani, K.M., Jeni, L.A.: Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14750-14760 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.242 + ], + "angle": 0, + "content": "18. Dong, J., Fang, Q., Jiang, W., Yang, Y., Bao, H., Zhou, X.: Fast and robust multiperson 3d pose estimation and tracking from multiple views. In: T-PAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.242, + 0.785, + 0.283 + ], + "angle": 0, + "content": "19. Dong, J., Jiang, W., Huang, Q., Bao, H., Zhou, X.: Fast and robust multi-person 3d pose estimation from multiple views. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7792-7801 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.283, + 0.785, + 0.338 + ], + "angle": 0, + "content": "20. Dong, Z., Guo, C., Song, J., Chen, X., Geiger, A., Hilliges, O.: Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20470-20480 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "21. Dong, Z., Song, J., Chen, X., Guo, C., Hilliges, O.: Shape-aware multi-person pose estimation from multi-view images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11158-11168 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "22. Ershadi-Nasab, S., Noury, E., Kasaei, S., Sanaei, E.: Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications 77, 15573-15601 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.785, + 0.473 + ], + "angle": 0, + "content": "23. Fabbri, M., Lanzi, F., Calderara, S., Alletto, S., Cucchiara, R.: Compressed volumetric heatmaps for multi-person 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7204-7213 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "24. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2334-2343 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.568 + ], + "angle": 0, + "content": "25. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7214-7223 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.623 + ], + "angle": 0, + "content": "26. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12858-12868 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "27. Guo, W., Bie, X., Alameda-Pineda, X., Moreno-Noguer, F.: Multi-person extreme motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13053-13064 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.704 + ], + "angle": 0, + "content": "28. Hasler, N., Stoll, C., Sunkel, M., Rosenhahn, B., Seidel, H.P.: A statistical model of human pose and body shape. In: Computer graphics forum. vol. 28, pp. 337-346. Wiley Online Library (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "29. Jiang, T., Chen, X., Song, J., Hilliges, O.: Instantavatar: Learning avatars from monocular video in 60 seconds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16922-16932 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "30. Joo, H., Liu, H., Tan, L., Gui, L., Nabbe, B., Matthews, I., Kanade, T., Nobuhara, S., Sheikh, Y.: Panoptic studio: A massively multiview system for social motion capture. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3334-3342 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7122-7131 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AvatarPose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Kocabas, M., Huang, C.H.P., Hilliges, O., Black, M.J.: Pare: Part attention regressor for 3d human body estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11127-11137 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "33. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: Spec: Seeing people in the wild with an estimated camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11035-11045 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.247, + 0.786, + 0.288 + ], + "angle": 0, + "content": "34. Li, Y., Zhang, S., Wang, Z., Yang, S., Yang, W., Xia, S.T., Zhou, E.: Tokenpose: Learning keypoint tokens for human pose estimation. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 11313-11322 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "35. Lin, J., Lee, G.H.: Multi-view multi-person 3d pose estimation with plane sweep stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11886-11895 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.333, + 0.786, + 0.374 + ], + "angle": 0, + "content": "36. Liu, Y., Gall, J., Stoll, C., Dai, Q., Seidel, H.P., Theobalt, C.: Markerless motion capture of multiple characters using multiview image segmentation. IEEE transactions on pattern analysis and machine intelligence 35(11), 2720-2735 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.376, + 0.787, + 0.415 + ], + "angle": 0, + "content": "37. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.417, + 0.787, + 0.459 + ], + "angle": 0, + "content": "38. Luvizon, D.C., Habermann, M., Golyanik, V., Kortylewski, A., Theobalt, C.: Scene-aware 3d multi-human motion capture from a single camera. In: Computer Graphics Forum. vol. 42, pp. 371-383. Wiley Online Library (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "39. Martinez, J., Hossain, R., Romero, J., Little, J.J.: A simple yet effective baseline for 3d human pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2640-2649 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.502, + 0.787, + 0.543 + ], + "angle": 0, + "content": "40. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H.P., Xu, W., Casas, D., Theobalt, C.: Vnect: Real-time 3d human pose estimation with a single rgb camera. Acm transactions on graphics (tog) 36(4), 1-14 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.545, + 0.787, + 0.585 + ], + "angle": 0, + "content": "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.587, + 0.787, + 0.628 + ], + "angle": 0, + "content": "42. Muller, L., Osman, A.A.A., Tang, S., Huang, C.H.P., Black, M.J.: On self-contact and human pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9990-9999 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.63, + 0.787, + 0.67 + ], + "angle": 0, + "content": "43. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.672, + 0.787, + 0.727 + ], + "angle": 0, + "content": "44. Osman, A.A., Bolkart, T., Black, M.J.: Star: Sparse trained articulated human body regressor. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16. pp. 598-613. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.729, + 0.787, + 0.783 + ], + "angle": 0, + "content": "45. Patel, P., Huang, C.H.P., Tesch, J., Hoffmann, D.T., Tripathi, S., Black, M.J.: Agora: Avatars in geography optimized for regression analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13468-13478 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "46. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10975-10985 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "47. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3d human pose and shape from a single color image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 459-468 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.247 + ], + "angle": 0, + "content": "48. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.303 + ], + "angle": 0, + "content": "49. Pishchulin, L., Insafutdinov, E., Tang, S., Andres, B., Andriluka, M., Gehler, P.V., Schiele, B.: Deepcut: Joint subset partition and labeling for multi person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4929-4937 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.787, + 0.359 + ], + "angle": 0, + "content": "50. Reddy, N.D., Guigues, L., Pishchulin, L., Eledath, J., Narasimhan, S.G.: Tessen-track: End-to-end learnable multi-person articulated 3d pose tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15190-15200 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.388 + ], + "angle": 0, + "content": "51. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.431 + ], + "angle": 0, + "content": "52. Saito, S., Yang, J., Ma, Q., Black, M.J.: Scanimate: Weakly supervised learning of skinned clothed avatar networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2886-2897 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.5 + ], + "angle": 0, + "content": "53. Shuai, Q., Geng, C., Fang, Q., Peng, S., Shen, W., Zhou, X., Bao, H.: Novel view synthesis of human interactions from sparse multi-view videos. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22, Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10.1145/3528233.3530704, https://doi.org/10.1145/3528233.3530704" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.543 + ], + "angle": 0, + "content": "54. Shuai, Q., Yu, Z., Zhou, Z., Fan, L., Yang, H., Yang, C., Zhou, X.: Reconstructing close human interactions from multiple views. ACM Transactions on Graphics (dec 2023). https://doi.org/10.1145/3618336" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.586 + ], + "angle": 0, + "content": "55. Sun, Y., Bao, Q., Liu, W., Fu, Y., Black, M.J., Mei, T.: Monocular, one-stage, regression of multiple 3d people. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11179-11188 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.642 + ], + "angle": 0, + "content": "56. Sun, Y., Liu, W., Bao, Q., Fu, Y., Mei, T., Black, M.J.: Putting people in their place: Monocular regression of 3d people in depth. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13243-13252 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "57. Te, G., Li, X., Li, X., Wang, J., Hu, W., Lu, Y.: Neural capture of animatable 3d human from monocular video. In: European Conference on Computer Vision. pp. 275-291. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.741 + ], + "angle": 0, + "content": "58. Tu, H., Wang, C., Zeng, W.: Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 197-212. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.798 + ], + "angle": 0, + "content": "59. Wang, C., Li, J., Liu, W., Qian, C., Lu, C.: Hmor: Hierarchical multi-person ordinal relations for monocular multi-person 3d pose estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 242-259. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "60. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "AvatarPose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.231 + ], + "angle": 0, + "content": "61. Wang, T., Zhang, J., Cai, Y., Yan, S., Feng, J.: Direct multi-view multi-person 3d pose estimation. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 13153-13164. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/6da9003b743b65f4c0ccd295cc484e57-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "62. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.288, + 0.787, + 0.327 + ], + "angle": 0, + "content": "63. Wu, S., Jin, S., Liu, W., Bai, L., Qian, C., Liu, D., Ouyang, W.: Graph-based 3d multi-person pose estimation using multi-view images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11148-11157 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.786, + 0.356 + ], + "angle": 0, + "content": "64. Wu, S., Jin, S., Liu, W., Bai, L., Qian, C., Liu, D., Ouyang, W.: Graph-based 3d multi-person pose estimation using multi-view images. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.412 + ], + "angle": 0, + "content": "65. Xu, H., Bazavan, E.G., Zanfir, A., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Ghum & ghuml: Generative 3d human shape and articulated pose models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6184-6193 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "66. Ye, H., Zhu, W., Wang, C., Wu, R., Wang, Y.: Faster voxelpose: Real-time 3d human pose estimation by orthographic projection. In: European Conference on Computer Vision. pp. 142-159. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "67. Yin, Y., Guo, C., Kaufmann, M., Zarate, J.J., Song, J., Hilliges, O.: Hi4d: 4d instance segmentation of close human interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17016-17027 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.55 + ], + "angle": 0, + "content": "68. Zanfir, A., Marinoiu, E., Sminchisescu, C.: Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2148-2157 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "69. Zhang, H., Tian, Y., Zhou, X., Ouyang, W., Liu, Y., Wang, L., Sun, Z.: Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11446-11456 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.647 + ], + "angle": 0, + "content": "70. Zhang, J., Liu, X., Ye, X., Zhao, F., Zhang, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG) 40(4), 1-18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "71. Zhang, Y., Wang, C., Wang, X., Liu, W., Zeng, W.: Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 2613-2626 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "72. Zhang, Y., An, L., Yu, T., Li, X., Li, K., Liu, Y.: 4d association graph for realtime multi-person motion capture using multiple video cameras. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1324-1333 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "73. Zhang, Y., Li, Z., An, L., Li, M., Yu, T., Liu, Y.: Lightweight multi-person total motion capture using sparse multi-view cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5560-5569 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "74. Zheng, Y., Shao, R., Zhang, Y., Yu, T., Zheng, Z., Dai, Q., Liu, Y.: Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6239-6249 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.35, + 0.128 + ], + "angle": 0, + "content": "F. Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "75. Zhou, H., Hong, C., Han, Y., Huang, P., Zhuang, Y.: Mh pose: 3d human pose estimation based on high-quality heatmap. In: 2021 IEEE International Conference on Big Data (Big Data). pp. 3215-3222. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "76. Zhou, Z., Shuai, Q., Wang, Y., Fang, Q., Ji, X., Li, F., Bao, H., Zhou, X.: Quickpose: Real-time multi-view multi-person pose estimation in crowded scenes. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.232 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_origin.pdf b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ad4823f443984987ba64ef50b5bbe677108f413c --- /dev/null +++ b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/ef51f60f-e009-41a2-9e0c-f2ae729e6483_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adfaecb3ee37e060c44873d773145c074ca19cfc0fc2fa5771b76aad70ffc9ef +size 10033668 diff --git a/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/full.md b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3de3abee84dff0eb305242b105af063a40bdd590 --- /dev/null +++ b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/full.md @@ -0,0 +1,355 @@ +# AvatarPose: Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos + +Feichi Lu $^{1*}$ , Zijian Dong $^{1,2*}$ , Jie Song $^{1\dagger}$ , and Otmar Hilliges $^{1}$ + +$^{1}$ Department of Computer Science, ETH Zürich, Switzerland, +2 Max Planck Institute for Intelligent Systems, Germany + +![](images/d6c607c5062696b75767920ec8fc75be6e627631d2f90d51e1d5f16262ff6430.jpg) +Fig. 1: We propose AvatarPose, a method for estimating the 3D poses and shapes of multiple closely interacting people from multi-view videos. To this end, we first reconstruct the avatar of each individual and leverage the learned personalized avatars as priors to refine poses via color and silhouette rendering loss. We alternate between avatar refinement and pose optimization to obtain the final pose estimation. + +Abstract. Despite progress in human motion capture, existing multiview methods often face challenges in estimating the 3D pose and shape of multiple closely interacting people. This difficulty arises from reliance on accurate 2D joint estimations, which are hard to obtain due to occlusions and body contact when people are in close interaction. To address this, we propose a novel method leveraging the personalized implicit neural avatar of each individual as a prior, which significantly improves the robustness and precision of this challenging pose estimation task. Concretely, the avatars are efficiently reconstructed via layered volume rendering from sparse multi-view videos. The reconstructed avatar prior allows for the direct optimization of 3D poses based on color and silhouette rendering loss, bypassing the issues associated with noisy 2D detections. To handle interpenetration, we propose a collision loss on the overlapping shape regions of avatars to add penetration constraints. Moreover, both 3D poses and avatars are optimized in an alternating manner. Our experimental results demonstrate state-of-the-art performance on several public datasets. + +Keywords: human pose estimation $\cdot$ human close interaction $\cdot$ multiview pose estimation $\cdot$ avatar prior + +# 1 Introduction + +People frequently engage in interactions with each other in daily life to offer physical support or convey emotions. For AI systems to interpret 3D human interactions, a foundational step is to reconstruct 3D human poses and shapes using consumer-grade sensors, such as cameras. However, in closely interacting scenarios, such as hugging or kissing, 3D pose estimations face challenges due to strong occlusions. The problem of depth ambiguity worsens when close contact is involved, making it more difficult to predict 3D poses from 2D images or videos. Consequently, a multi-camera setup becomes essential to provide additional observations and to address depth ambiguity in pose estimation. + +Despite the ubiquity of close human interactions, the study of estimating such human motions is scarce. Most previous multi-human methods with a multi-view setup [2-4, 19, 22, 30, 58, 66, 72] focus on scenarios where people are at a distance from each other. Some methods [2, 4, 19, 22, 72] typically formulate the problem as a cross-view matching problem, relying heavily on 2D joint estimations for subsequent 3D triangulation. These methods demonstrate high sensitivity to noisy or missing 2D joint estimations, particularly when occlusion occurs. Another group of learning-based methods [58, 61, 63, 66] attempts to integrate 2D features from each view into a 3D voxel space and predict 3D human poses from identified 3D subvolumes of each individual. These methods are more robust to occlusion, but they encounter challenges with generalization and are sensitive to distribution shifts due to the lack of annotated 3D data. To correct abnormal or missing pose estimations, some methods [18, 21] leverage parametric body models like SMPL [37] as full-body priors and fit these 3D models to 2D joint estimations. Although this approach alleviates the issue of abnormal or missing joint estimations, it remains constrained by noisy 2D detections. + +Embracing the challenging problem of pose estimation for close interactions, we propose a novel method to estimate the 3D poses and shapes of multiple people observed from a sparse set of cameras. Our goal is to ensure that even in close contact, the estimated human poses and shapes are accurate and free from interpenetrations. The key to our method is reconstructing implicit textured avatars of each individual in the scene and leveraging them as a strong personalized prior for pose optimization (Figure 1). In contrast to relying solely on noisy 2D joint detections, this textured avatar prior enables us to leverage pixel-wise color and silhouette information for pose refinement, significantly increasing the robustness and accuracy of our method (Table 3). Meanwhile, the reconstructed avatar provides crucial geometric and appearance information to avoid collisions between individuals. Compared to methods using SMPL body shape [25,42] to penalize collisions, our implicit avatar model contains a more detailed clothed shape and additional appearance cues and enables efficient computation of penetration loss. Due to the mutually beneficial relationship between avatar and pose, we alternate between pose optimization and avatar refinement. + +More specifically, to accelerate the learning of avatars, we model each human individual in canonical space using an efficient neural radiance field variant in Instant NGP [43] and combine it with an efficient SMPL-based deformation + +module [37]. To learn and render avatar models of multiple people, we adapt layered volume rendering [53, 70] to our avatar model. This adaptation allows us to jointly optimize all avatar models through a straightforward rendering loss. Once learned, the avatar can be animated and rendered based on pose parameters at interactive rates, thus naturally serving as an efficient personalized textured prior for pose optimization. With the learned personalized prior, we optimize pose via a novel objective function. Different from previous methods based on 2D reprojection error of joints, we directly optimize pose parameters via minimizing color and silhouette rendering losses while keeping the learned avatar model fixed. To prevent interpenetration between human individuals, a collision loss is introduced by penalizing the situation when a 3D point is occupied by multiple avatars. To remove artifacts in the initial avatar due to imperfect pose initialization, we further refine avatars based on optimized poses. Throughout the optimization process, both initial personalized avatar models and SMPL parameters are optimized in an alternating manner, motivated by the insight that accurate 3D pose estimations improve avatar learning, and improved avatar models, in turn, increase the precision of overall pose estimations. + +We experimentally demonstrate that our method significantly outperforms previous state-of-the-art methods on several public datasets both quantitatively (Table 1 and Table 2) and qualitatively (Figure 3) especially when people are in close interaction. In summary, our contributions are: + +- We propose a pipeline that efficiently creates implicit neural avatars of closely interacting people and leverages the learned avatars as priors to optimize poses. +- The avatar prior enables us to design a novel objective function that leverages color and mask rendering loss for pose optimization. We show the superiority of this loss function compared to the 2D reprojection error of 3D joints used by most of the previous methods. +- Based on the learned avatar, a collision loss is introduced to avoid penetration when individuals are in contact. + +# 2 Related Work + +# 2.1 Multi-Person 3D Pose Estimation + +Despite significant progress in multi-human 2D pose estimation [9,13,15,24,34, 49] and 3D pose estimation from monocular images or video [7, 31-33, 38-40, 45, 47, 55, 56, 59, 68, 69], the reconstruction accuracy is still limited due to depth ambiguity and strong occlusions when humans are in close contact. A multi-view setting [4, 5, 19, 22, 23, 35, 36, 54, 61, 63, 66, 72] helps to alleviate these challenges. One straightforward idea of most methods [2, 4, 19, 22, 72, 76] is to formulate the problem into cross-view matching and association problems. MVPose [19] performs 2D person parsing in each image and leverages cross-view person matching to infer 3D pose. 4DAssociation [72] additionally adds tracking into this process to form a unified graph for associating 4D information. However, these methods + +are sensitive to the noisy estimation of 2D pose. In contrast to these matching-based methods, some recent methods [5,14,17,23,35,50,58,61,63,66,71,75] directly learn deep neural networks to regress poses. Faster Voxelpose [66] employs the feature volume proposed by Voxelpose [58] and enhances computational efficiency. Graph [63] designs three graph neural network models for human center detection and pose estimation. MvP [61] simplifies the multi-person pose estimation by direct regression using the transformer model. A concurrent method CloseMocap [54] proposes to learn a model from a synthetic dataset simulating occlusion situations. However, heavily relying on the 2D or 3D features as input during training, these learning-based methods suffer from generalization issues when subjects, motions, and camera configurations change. + +To further improve the robustness, a statistical parametric body model such as SMPL [37] is explored in [21] as a regularization prior for 3D joint refinement. Some follow-ups [18,73] show that parametric models help in correcting implausible 3D pose estimates and filling in missing joints. However, this coarse body prior highly relies on aligning the 3D joints to 2D pose estimations, which are inaccurate when occlusion happens. Different from all of these previous methods, we explore the usage of personalized textured avatar models as priors to refine human poses. This prior enables us to leverage color and silhouette information from multi-view observation for refining poses of closely interacting humans. + +# 2.2 3D Human Modeling + +Parametric human body models [1,28,37,44,65] can represent minimally clothed human shapes by deforming a template mesh. It is challenging to extend this explicit representation for modeling clothed humans due to the fixed topology and resolution. To overcome this limitation, methods such as SNARF [12] and SCANimate [52] propose to model articulated human shapes based on 3D implicit representations. Many works [6,8,10,20,26,29,48,57,60,62] fit implicit neural fields to RGB or RGB-D videos by neural rendering to reconstruct the shape and appearance of a single human body. However, when applied to a multi-human scene, these methods are not able to achieve good fidelity due to strong occlusions. Recent methods including ST-NeRF [70] and [53] leverage layered neural representation to model multiple humans with sparse multi-view videos and thus can generate novel view synthesis of dynamic multiple humans. The main problem of all aforementioned approaches, however, is their high reliance on 3D human pose estimation [70,74]: the deformation of the human model requires accurate human poses, which is hard to obtain when people are in close interaction. To address this challenge, our method is orthogonal to others, aiming at leveraging the learned avatar as priors for pose estimation in close interaction. + +# 2.3 3D Human Datasets for Close Interaction + +Most of the existing datasets [3,11,30] like Shelf and Panoptic studying multiperson pose estimation focus on the scene where people are at a distance from each other and rarely involved in close interactions. To study close interactions + +among people, MultiHuman [74] dataset captures multi-person interaction with some close interactions and occlusions. ExPI [27] creates a multi-person extreme motion dataset with close interactions, but it focuses mainly on motion prediction for future frames instead of pose estimation from sparse views. CHI3D [25] captures two-person interaction datasets and proposes to learn a contact estimation module from annotations to improve the precision of pose estimation. The most recent work Hi4D [67] creates a challenging dataset of physically close human interaction and proposes a method to disentangle human bodies and estimate poses. However, this method relies on the 3D ground truth of clothed human meshes captured with expensive 3D body scanners. Due to the reliance on the limited annotated 3D data, this method faces challenges in generalization with different people and camera configurations. In contrast to all the works before, we intend to solve the problem of pose estimation in close human interaction without requiring accurate 3D scans or other training data. + +# 3 Method + +Given a dynamic scene captured by a sparse set of RGB cameras, our goal is to estimate the 3D pose and shape of multiple people even if they interact closely. To address this challenging task, our key idea is to first reconstruct the personalized avatar of each individual in the scene and leverage them as a strong prior to refine the appearance and pose in an alternating manner. An overview of our method is shown in Figure 2. + +We first introduce an efficient pipeline to create avatars of multiple people in a scene (Section 3.1 and Figure 2(a)). Specifically, we leverage an accelerated neural radiance field to represent the shape and appearance of each individual in canonical space and deform it at an interactive rate. We then adapt layered volume rendering to our pipeline, which composites the rendering of avatars into one image, thus enabling direct learning from multi-view video inputs. + +Thanks to the learned avatar prior for each individual, we can enhance 3D pose optimization via a combination of RGB and silhouette rendering loss (Section 3.2 and Figure 2(b)). While previous work heavily relies on noisy 2D joint detection, we show that employing such pixel-wise color and silhouette information can largely increase precision and robustness. Moreover, a collision loss is introduced to avoid interpenetration. Finally, we alternate between avatar learning and pose optimization to get complete and accurate 3D human poses. + +# 3.1 Multi-Avatar Prior Learning + +Avatar Model We represent each human individual in canonical space using an accelerated neural radiance field [43] and model shape-aware articulated deformation based on SMPL [37]. + +- Canonical Appearance Representation: To model human shape and appearance, we create canonical radiance field $\bar{\mathbf{F}}_{\sigma_f}^{(l)}$ for each human instance + +![](images/9b4547620a3b3a8d47af20c8bfe126a4cdda69f8bc7a9752685987b415f28558.jpg) + +![](images/8091daf5024b3007c79f566d9dfa3b8cbe004e3914731505176196066d7cd96c.jpg) +Fig. 2: Method Overview: Our method consists of two modules: (a) Multi-Avatar Prior Learning: Given the input multi-view images and estimated poses $\Theta^{(l)}$ , we sample points $\mathbf{x}^{(l)}$ for each individual $l$ along the rays in posed space and warp these points into canonical space and calculate their color $\mathbf{c}^{(l)}$ and density $\sigma^{(l)}$ via the canonical appearance network $\bar{\mathbf{F}}_{\sigma_f}^{(l)}$ . We leverage layered volume rendering [70] to attain the final pixel color and compare it with the original input image to optimize the parameters of avatars. (b) Avatar-guided Pose Optimization: Given learned avatar model $\bar{\mathbf{F}}_{\sigma_f}^{(l)}$ and initial poses $\Theta^{(l)}$ of each individual $l$ , we deform all of the avatars based on SMPL-based deformer and render them jointly via layered volume rendering. We compare the composite rendering with input observation and minimize the RGB and mask rendering loss to optimize poses. A collision loss is additionally introduced to avoid interpenetration. Finally, we alternate between two modules to obtain the final result. For clarity, the parameters to be optimized are marked as red in each module. + +$l \in [1, L]$ , where $L$ is the number of humans in the scene. $\bar{\mathbf{F}}_{\sigma_f}^{(l)}$ takes a 3D point $\bar{\mathbf{x}}^{(l)}$ as input and predicts its density $\sigma^{(l)}$ and color $\mathbf{c}^{(l)}$ . Following Instant-NGP [43] and InstantAvatar [29] to accelerate the rendering, $\bar{\mathbf{F}}_{\sigma_f}^{(l)}$ is parameterized via using a hash table to store feature grids at different scales. + +- Pose Representation: We represent the 3D pose and underlying body shape for all human instances by SMPL parameters $\Theta = \{\Theta^{(l)}\}_{l\in [1,L]}$ . For each human $l$ , $\Theta^{(l)} = \{\beta^{(l)},\theta^{(l)},t^{(l)}\}$ contains shape parameters $\beta^{(l)}\in \mathbb{R}^{10}$ , pose parameters $\theta^{(l)}\in \mathbb{R}^{72}$ and translation $t^{(l)}\in \mathbb{R}^3$ of SMPL. +- Deformer: To enable animation given targeted poses $\Theta^{(l)}$ , we require the radiance field in the posed space. Given a point $\mathbf{x}^{(l)}$ in deformed space of human $l$ , we determine the corresponding canonical point $\bar{\mathbf{x}}^{(l)}$ by inverse linear blend skinning(LBS) [37]: $\bar{\mathbf{x}}^{(l)}(\mathbf{x}^{(l)},\Theta^{(l)}) = (\sum_{i = 1}^{n_b}w_i(\Theta^{(l)})\mathbf{B}_i(\Theta^{(l)}))^{-1}\mathbf{x}^{(l)}$ , where $\mathbf{B}_i$ is the rigid bone transformation matrix for joint $i\in \{1,\dots,n_b\}$ under pose $\Theta^{(l)}$ . $w_{i}$ is the skinning weights of the nearest neighbor of $\mathbf{x}^{(l)}$ in + +the deformed SMPL vertices. We obtain the radiance field at the point $\mathbf{x}^{(l)}$ by evaluating the canonical appearance field at the corresponding point $\bar{\mathbf{x}}^{(l)}$ . + +Layered Volume Rendering To obtain the pixel value for a ray $\mathbf{r} \in \mathcal{R}$ , we raycast every human instance separately with a layered rendering strategy similar to ST-NeRF [70]. Specifically, we first calculate the intersection points between the ray and the 3D bounding box of each human instance and uniformly sample points in each bounding box. To distinguish different identities, we assign each sampled point $\mathbf{x}_i$ a one-hot representation $\mathbf{m}_i = [m_i^{(1)}, \dots, m_i^{(L)}]$ to indicate which human identity it belongs to. After sorting all sampled points by their depth values and calculating their corresponding color $\mathbf{c}_i$ and density $\sigma_i$ from the avatar model, if $m_i^{(l)} = 1$ , we compute + +$$ +\mathbf {c} _ {i}, \sigma_ {i} = \bar {\mathbf {F}} _ {\sigma_ {f}} ^ {(l)} \left(\bar {\mathbf {x}} _ {i} \left(\mathbf {x} _ {i}, \boldsymbol {\Theta} ^ {(l)}\right)\right). \tag {1} +$$ + +The color of each ray is computed via numerical integration [41]. + +$$ +\hat {\mathbf {C}} (\mathbf {r}) = \sum_ {i = 1} ^ {N} \alpha_ {i} \prod_ {j < i} (1 - \alpha_ {j}) \mathbf {c} _ {i} \quad \alpha_ {i} = 1 - \exp (- \sigma_ {i} \delta_ {i}), \tag {2} +$$ + +where $\delta_{i}$ is the distance between samples. The accumulated alpha value, which represents ray opacity, can be computed via: + +$$ +\alpha (\mathbf {r}) = \sum_ {i = 1} ^ {N} \alpha_ {i} \prod_ {j < i} (1 - \alpha_ {j}). \tag {3} +$$ + +For each human identity $l \in [1, L]$ , the corresponding instance ray opacity can be calculated via: + +$$ +\alpha^ {(l)} (\mathbf {r}) = \sum_ {i = 1} ^ {N} \alpha_ {i} \prod_ {j < i} (1 - \alpha_ {j}) m _ {i} ^ {(l)}. \tag {4} +$$ + +Training The overall training process is shown in Figure 2(a). For training avatar layers, we minimize the Huber loss $\rho$ between the predicted pixel color $\hat{\mathbf{C}}(\mathbf{r})$ and the ground truth pixel color $\mathbf{C}_{gt}(\mathbf{r})$ : + +$$ +\mathcal {L} _ {R G B} = \frac {1}{| \mathcal {R} |} \sum_ {r \in \mathcal {R}} \rho \left(\left\| \hat {\mathbf {C}} (\mathbf {r}) - \mathbf {C} _ {g t} (\mathbf {r}) \right\|\right). \tag {5} +$$ + +Since instance segmentation of human performers is hard to obtain and is not accurate, we choose foreground segmentation as our mask supervision, which is obtained via SAM-Track [16]. We apply a loss for optimizing the rendered alpha values $\alpha$ to reduce the artifacts in the floating area: + +$$ +\mathcal {L} _ {\text {a l p h a}} = \frac {1}{| \mathcal {R} |} \sum_ {\mathbf {r} \in \mathcal {R}} (\alpha (\mathbf {r}) - \alpha_ {S A M} (\mathbf {r})) ^ {2}. \tag {6} +$$ + +Following [53], we add a regularization loss for instance alpha values to make sure every pixel can only be rendered from one human layer: + +$$ +\mathcal {L} _ {\text {l a y e r}} = - \frac {1}{L \mid \mathcal {R} \mid} \sum_ {\mathbf {r} \in \mathcal {R}} \sum_ {l = 1} ^ {L} \alpha^ {(l)} (\mathbf {r}) \log \left(\alpha^ {(l)} (\mathbf {r})\right). \tag {7} +$$ + +Similar to [29], we also add hard surface and density regularization terms in the learning process. More training details can be found in the Supp Mat. + +# 3.2 Avatar-guided Pose Optimization + +Equipped with learned avatars obtained in Section 3.1, we aim to estimate the 3D shape and pose of multiple humans with close physical contact. To achieve this, we leverage the avatars as priors to handle the challenging pose ambiguities caused by contact. Specifically, we first initialize pose parameters using an off-the-shelf 3D pose estimator [72] and refine the pose via a rendering loss between rendered posed avatars and 2D observations (Figure 2(b)). Since the initial imperfect pose estimations may cause artifacts in avatar reconstruction, we further refine the weights of the avatar model using the optimized pose. This process is formulated as an alternating optimization to refine both poses and avatars. + +Initialization The initial 3D pose proposals are estimated by leveraging the off-the-shelf 3D human pose estimator [72]. After that, we register a SMPL model to the estimated 3D joints to obtain the initial pose parameters $\Theta_0$ . Given these estimated poses, the initial avatar model of each individual is further learned from multi-view videos. + +Objective To leverage avatars as priors to tackle challenges caused by contact, we optimize the following objective: + +$$ +\begin{array}{l} \mathcal {L} (\boldsymbol {\Theta}) = \lambda_ {R G B} \mathcal {L} _ {R G B} (\boldsymbol {\Theta}) + \lambda_ {\alpha} \mathcal {L} _ {\alpha} (\boldsymbol {\Theta}) \tag {8} \\ + \lambda_ {r e g} \mathcal {L} _ {r e g} (\Theta) + \lambda_ {p a} \mathcal {L} _ {p a} (\Theta). \\ \end{array} +$$ + +Here, $\Theta$ is the SMPL parameters to be optimized for all avatars, which is also consistent with the pose parameters to deform the avatar model (Section 3.1). Following [62], we represent $\Theta$ as $\Theta_0 + \mathrm{MLP}(\Theta_0)$ and refine poses by changing the parameters of the neural network. This representation empirically shows more robust results compared to directly optimizing SMPL parameters. + +Different from previous methods [18,21,58,72], the personalized multi-ayar prior allows us to leverage appearance and silhouette information to refine initial poses. Specifically, we calculate $\mathcal{L}_{RGB}$ to ensure the color consistency between the rendered pixel $\hat{\mathbf{C}} (\mathbf{r},\boldsymbol {\Theta})$ (Equation (1) and Equation (2)) of deformed avatar with poses $\Theta$ and the corresponding ground-truth pixel color $\mathbf{C}_{gt}(\mathbf{r})$ . + +$$ +\mathcal {L} _ {R G B} (\boldsymbol {\Theta}) = \frac {1}{| \mathcal {R} |} \sum_ {r \in \mathcal {R}} \rho \left(\left\| \hat {\mathbf {C}} (\mathbf {r}, \boldsymbol {\Theta}) - \mathbf {C} _ {g t} (\mathbf {r}) \right\|\right). \tag {9} +$$ + +Additionally, a cross-entropy loss $\mathcal{L}_{\alpha}$ is introduced to ensure that the rendered mask $\alpha (\mathbf{r},\Theta)$ (Equation (3)) of the reposed avatar is aligned with the estimated SAM-Track mask $\alpha_{SAM}(\mathbf{r})$ by: + +$$ +\mathcal {L} _ {\alpha} (\boldsymbol {\Theta}) = - \sum_ {\mathbf {r} \in \mathcal {R}} \alpha_ {S A M} (\mathbf {r}) \log (\alpha (\mathbf {r}, \boldsymbol {\Theta})) \tag {10} +$$ + +To penalize the unnatural poses and avoid elbows and knees bending in the wrong direction, we add an L2 regularization term and combine it with the pose prior in SMPLify [7] to constrain physically implausible joint rotation: + +$$ +\mathcal {L} _ {r e g} (\boldsymbol {\Theta}) = \| \boldsymbol {\Theta} \| _ {2} + \lambda \sum_ {i \in I} \exp (\boldsymbol {\Theta} _ {i}), \tag {11} +$$ + +where $I$ is the set of pose indices corresponding to elbows and knees. + +A key challenge to correctly estimate poses in close interaction is to handle interpenetration. Since every avatar is modeled separately, the surfaces tend to intersect when they are in contact. To handle this, we first select sampled points inside multiple instances as $\mathcal{S} = \{\mathbf{x}_i\mid \alpha_i^{(p)} > 0,\alpha_i^{(q)} > 0,p,q\in [1,L],p\neq q\}$ ( $\alpha_{i}^{(p)}$ is calculated from Equation (1) and Equation (2) corresponding with a point $\mathbf{x}_i$ with $m_i^{(p)} = 1$ ). We then propose a collision loss $\mathcal{L}_{pa}$ for penalizing penetration: + +$$ +\mathcal {L} _ {p a} (\boldsymbol {\Theta}) = \frac {1}{| \mathcal {S} |} \sum_ {\boldsymbol {x} _ {i} \in \mathcal {S}} \alpha_ {i} ^ {(p)} (\boldsymbol {\Theta}) \alpha_ {i} ^ {(q)} (\boldsymbol {\Theta}) \tag {12} +$$ + +Intuitively, this loss guarantees every sample point in 3D space can not be occupied by multiple avatar models simultaneously, which guarantees better pose estimation in close contact. + +Alternating Optimization Since artifacts sometimes appear on initial avatars due to imperfect pose initialization, we further refine avatars based on optimized poses via minimizing the loss function in Section 3.1. Finally, the optimization of poses and avatars is formulated in an alternating fashion for $N$ steps. More details of optimization can be found in the Supp Mat. + +# 4 Experiments + +Datasets. We mainly evaluate our proposed method on Hi4D [67] and CHI3D [25], which are challenging datasets of two humans in close interaction. To demonstrate the generalization ability of our method for more than two people, we also evaluate on Shelf [3] and MultiHuman [74] including three or four people. More details are shown in the Supp Mat. + +Metrics. We use the Mean Per Joint Position Error (MPJPE) [58] to measure the distance between the ground truth 3D poses and the estimated poses. We also choose the Percentage of Correct Parts (PCP3D) metric [19] to calculate the percentage of correct parts. $\mathrm{AP}_K$ [58] and Recall [21] are also leveraged to evaluate performance. More details are shown in the Supp Mat. + +![](images/a33cfa3ca61ced7bb2f4e29ba22b051cfa9400ab5079e3e8eeeb0722bc712573.jpg) +Fig. 3: Qualitative Comparison with SotA methods [19,61,64,66,72] on Hi4D and CHI3D. We show two examples from the Hi4D and CHI3D datasets compared with Graph, MvP, Faster VoxelPose, MVPose, and 4DAssociation. For each example, we show 2D projections on two sampled views. + +Baselines. We compare our method on the Hi4D dataset with state-of-the-art methods in three categories discussed in Section 2.1. For learning-based methods, we choose Graph [64], MvP [61], and Faster VoxelPose [66] and fine-tune these models on a subset of the evaluating datasets. For pure association-based methods, we choose 4DAssociation [72] and MVPose* [19]. MVPose [18] adds temporal tracking and SMPL prior to MVPose* [19] and is regarded as a SMPL-guided method. More details about baselines can be found in the Supp Mat. + +# 4.1 Comparison to SotA + +Table 1 and Table 2 summarizes our quantitative comparisons on Hi4D and CHI3D with SotA (State-of-the-art) methods. Our method largely outperforms the other SotA methods in all of the metrics including MPJPE, PCP3D, and $\mathrm{AP}_K$ . More comparisons on Shelf and MultiHuman are shown in the Supp Mat. + +Comparison with Graph [63], MvP [61], Faster VoxelPose [66]. Our method outperforms Graph, MvP, and Faster VoxelPose on Hi4D and CHI3D. These methods are prone to overfitting to training pose distributions, thus struggling with challenging poses of interacting actors. Specifically, Faster VoxelPose sometimes misses actors in close contact, leading to a relatively low recall. While Graph and MvP have better recall, Graph fails to consistently track actors across frames, and MvP results in many misaligned joints between actors. + +
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑Recall(%) ↑
MvP [61]92.7774.1441.5963.8693.84
Graph [63]89.6271.5544.7567.3393.31
Faster VoxelPose [66]68.4073.6744.0568.7083.55
MVPose* [19]53.0587.5767.9780.2893.80
MVPose [18]42.6390.7671.7990.1993.30
4DAssociation [72]41.2988.6280.8797.2798.78
Ours32.1096.9091.4897.3398.78
+ +Table 1: Quantitative Comparison with SotA on the Hi4D [67] Dataset (8 views). We compare our method with MvP [61], Graph [63], Faster VoxelPose [66], MVPose* [19], MVPose [18] and 4DAssociation [72]. We report MPJPE, PCP, $\mathrm{AP}_K$ , and Recall metric for all methods. + +
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑Recall(%) ↑
MvP [61]55.3889.4763.5892.5399.06
Graph [63]45.3392.4674.0295.2599.17
Faster VoxelPose [66]67.8178.4129.2882.8893.34
MVPose* [19]50.4290.3969.1375.7288.72
MVPose [18]34.0593.3579.9486.9188.18
4DAssociation [72]37.4799.3089.6698.6799.85
Ours32.9899.7993.2099.7999.85
+ +Table 2: Quantitative Comparison with SotA on the CHI3D [25] Dataset (4 views). We compare our method with Faster VoxelPose [66], MVPose* [19], MV-Pose [18] and 4DAssociation [72]. We report MPJPE, PCP, $\mathrm{AP}_K$ , and Recall metric for all methods. + +Comparison with MVPose [18]. Our method achieves much better MPJPE and precision than MVPose. MVPose relies heavily on noisy 2D joint detected in close interactions. In contrast, our method leverages color and silhouette rendering loss to optimize poses, leading to robustness to occlusions. Additionally, this top-down method cannot detect actors correctly with strong occlusions, as shown in Figure 3. + +Comparison with 4DAssociation [72]. Finally, we compare our method with the bottom-up association method. As shown in Table 1 and Table 2, our method outperforms 4DAssociation in most metrics. Figure 3 shows that when actors are close, this bottom-up method is inclined to associate joints with the wrong human instances. This is because they solve the joint association with a greedy algorithm, which is sensitive to missing and inaccurate 2D joint detections. + +![](images/ca7495f30530c603fa0431f97c957acb56b029ff68e8cbd66cc4a9a64aa534b3.jpg) +(a) + +![](images/5751a58588ccc2b09ebe946de64833b59a2e54620d3e5d31683ecd3549ab6732.jpg) +(c) + +![](images/79063d68ec8bb730e85d5cffc5f86bf63ba694a472853abd67b6b2a8c7293429.jpg) +(e) + +![](images/c8e92be45992602255faae1f8d413a6aa830184d4e3b849b1f2277cb47cc8100.jpg) +(b) + +![](images/1544a0211dc4dd190cfc86513394a251616c8df67a7cdd94ca58d32cf047e78b.jpg) +(d) + +![](images/37ec93390bfdf6f5cdcfc231f4204692236f2c29be2a0812955b9cb8f5bf1bd5.jpg) +(f) +Fig. 4: Qualitative Results of our method on Hi4D (a)(b), CHI3D (c)(d), MultiHuman Real-Cap (e), and Shelf (f). The left and middle columns in each sub-figure show the 2D projections of the estimated 3D skeletons and SMPL body meshes on two views. The right column in each sub-figure demonstrates skeletons and SMPL bodies in 3D scenes. + +# 4.2 Additional Qualitative Samples + +Figure 4 demonstrates more qualitative results of our method on Hi4D, CHI3D, and MultiHuman Real-Cap with challenging and close interactions among 2 or 3 people. We also demonstrate results on the Shelf [3] dataset, which contains 4 people without close contact. More results can be found in the Supp Mat. + +# 4.3 Ablation Study + +To validate the effectiveness of our method, we conduct a detailed analysis of different design choices of our algorithm. All the experiments are conducted on the Hi4D Dataset. + +Comparison with SMPL Body Prior. To validate the effectiveness of our personalized avatar prior, we compare with a baseline that optimizes SMPL parameters to align 3D joint reprojections to 2D observations. Our method significantly outperforms the SMPL prior baseline (Table 3) by leveraging color and silhouette rendering loss to refine poses, reducing reliance on inaccurate joint detections when occlusion happens. Figure 5 shows that while the baseline method incorrectly estimates the arm pose and even leads to penetration, our method reconstructs the poses accurately. + +Color and Mask Loss. To demonstrate the effectiveness of the color and silhouette rendering loss in our optimization process, we design baselines without RGB + +![](images/d3dd3bb0fb33cc515cf527d829bd5e0599372cd5876b0d1b73478989efb6619e.jpg) +Fig.5: Comparison with SMPL Body Prior. Only fitting SMPL to 2D observations, some joints in close contact such as arms are incorrectly estimated and even cause intersections between body surfaces. In contrast, our personalized prior enables accurate estimation of poses. + +
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑
Ours (SMPL fitted)40.4195.1084.0494.37
Ours (w/o RGB loss)78.4083.2617.8474.55
Ours (w/o Mask loss)31.0097.3390.2899.06
Ours (Joint Optimization)66.0484.5622.5576.70
Ours29.3798.0296.7999.06
+ +Table 3: Quantitative Ablation Results. Ablations to evaluate our method with only the SMPL fitted method, our method without RGB loss and without Silhouette loss, and our method without alternating optimization. + +loss and mask loss respectively for comparison. Table 3 shows that the RGB loss significantly improves pose optimization, and the optimization process will completely deviate from the correct trajectory in the absence of RGB loss. Mask loss is proven to slightly increase the accuracy by adding additional constraints on the rendered human silhouette. + +Alternating Optimization. We choose a joint optimization method widely used in avatar reconstructions [20, 26] to compare with our alternating optimization. Table 3 shows that our method significantly outperforms the chosen baseline. In Figure 6, the avatars from the baseline suffer from artifacts around contact body parts, leading to wrong 3D pose estimations. In contrast, our method can faithfully reconstruct avatars and poses with challenging initializations. + +Penetration Loss. Our penetration loss serves an important role in avoiding interpenetration. Comparing our method to the ablated version where we remove this loss, Figure 7 shows that the SMPL body of one person partially intersects with the other person in the contact area. This is also confirmed by the rendering result of the avatar. By penalizing the collision of density fields of avatars in 3D space, our method largely reduces the penetration and thus achieves more accurate pose estimation. + +![](images/d915227c893543113339fdb599925d70d6a87c5bcd1557216f2da5c3ecbd32da.jpg) +GT + +![](images/0df18bae9cc532ba8ad51ca35942a4aed5a706b0046f7284bdbbc8030b6f8175.jpg) +Joint Opt + +![](images/8685c5c8784522317a84ddd4d1a089aab591005576eb640e17849a71fd03977d.jpg) +Alternating Opt + +![](images/bba9dfc4bb1ad3117e08a5e98679235867f5f088d6c17c473898d81ada6a75c3.jpg) +Fig. 6: Ablation of Alternating Optimization. We show the results of rendered avatars and projections of the estimated 3D poses. Joint optimization suffers from artifacts around the contact part and in turn causes wrong pose estimations. In contrast, ours reconstructs both avatars and poses correctly. + +![](images/ec8179686a182c6612bd46c4ca0af2eeff0589110c6949c0416b5904be9a2a1d.jpg) +Ours(w/o penetration loss) + +![](images/d9326f997f7193d82750934f74f4a00b77c352b465a732e3698b4424a77ca6e7.jpg) + +![](images/ce07464fb9f4705079d24143265206ab9f1d4637c7985beedcdd768668281084.jpg) +Ours + +![](images/b784b2eb3781b8a547ea8005a2c557ac3e23d6b906255c0d595f072e04ad57f3.jpg) +Fig. 7: Ablation of Penetration Loss. Without penetration loss, both the avatar and underlying SMPL body tend to have collisions on surfaces. + +# 4.4 Limitations and Future Work + +Despite greatly improving $98\%$ of the sequences, our method shows minor improvements when the pose initialization is in the opposite direction or most joints are severely misaligned, leading to local minima in optimization. Moreover, we do not model hands in our avatar model, it will be a promising direction to integrate hand models [46,51] into the personalized avatar. More discussions about limitations and future work can be found in the Supp Mat. + +# 5 Conclusion + +In this paper, we propose AvatarPose, a novel method to estimate the 3D poses of multiple people in close interaction from sparse multi-view videos. Unlike previous methods leveraging SMPL body prior, we leverage the reconstructed avatars as personalized priors to guide pose optimization. The avatar prior enables us to use color and silhouette observations and to introduce a collision loss in close contact. Our method outperforms SotA methods significantly on public datasets of close human interactions. + +Acknowledgements This work was supported by the Swiss SERI Consolidation Grant "AI-PERCEIVE". We thank Wenbo Wang, Tianjian Jiang and Juan Zarate for providing us with helpful guidance and feedback. + +# References + +1. Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: Scape: shape completion and animation of people. In: ACM SIGGRAPH 2005 Papers, pp. 408-416 (2005) +2. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures for multiple human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1669-1676 (2014) +3. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures for multiple human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1669-1676 (2014) +4. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures revisited: Multiple human pose estimation. IEEE transactions on pattern analysis and machine intelligence 38(10), 1929-1942 (2015) +5. Benzine, A., Chabot, F., Luvison, B., Pham, Q.C., Achard, C.: Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6856-6865 (2020) +6. Bhatnagar, B.L., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Combining implicit function learning and parametric models for 3d human reconstruction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 311-329. Springer (2020) +7. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 561-578. Springer (2016) +8. Burov, A., Nießner, M., Thies, J.: Dynamic surface function networks for clothed human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10754-10764 (2021) +9. Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7291-7299 (2017) +0. Chen, J., Zhang, Y., Kang, D., Zhe, X., Bao, L., Jia, X., Lu, H.: Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629 (2021) +1. Chen, L., Ai, H., Chen, R., Zhuang, Z., Liu, S.: Cross-view tracking for multi-human 3d pose estimation at over 100 fps. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3279-3288 (2020) +2. Chen, X., Zheng, Y., Black, M.J., Hilliges, O., Geiger, A.: Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11594-11604 (2021) +3. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7103-7112 (2018) +4. Chen, Y., Gu, R., Huang, O., Jia, G.: Vtp: volumetric transformer for multi-view multi-person 3d pose estimation. Applied Intelligence pp. 1-12 (2023) +5. Cheng, B., Xiao, B., Wang, J., Shi, H., Huang, T.S., Zhang, L.: Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5386-5395 (2020) + +16. Cheng, Y., Li, L., Xu, Y., Li, X., Yang, Z., Wang, W., Yang, Y.: Segment and track anything. arXiv preprint arXiv:2305.06558 (2023) +17. Choudhury, R., Kitani, K.M., Jeni, L.A.: Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14750-14760 (2023) +18. Dong, J., Fang, Q., Jiang, W., Yang, Y., Bao, H., Zhou, X.: Fast and robust multiperson 3d pose estimation and tracking from multiple views. In: T-PAMI (2021) +19. Dong, J., Jiang, W., Huang, Q., Bao, H., Zhou, X.: Fast and robust multi-person 3d pose estimation from multiple views. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7792-7801 (2019) +20. Dong, Z., Guo, C., Song, J., Chen, X., Geiger, A., Hilliges, O.: Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20470-20480 (2022) +21. Dong, Z., Song, J., Chen, X., Guo, C., Hilliges, O.: Shape-aware multi-person pose estimation from multi-view images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11158-11168 (2021) +22. Ershadi-Nasab, S., Noury, E., Kasaei, S., Sanaei, E.: Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications 77, 15573-15601 (2018) +23. Fabbri, M., Lanzi, F., Calderara, S., Alletto, S., Cucchiara, R.: Compressed volumetric heatmaps for multi-person 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7204-7213 (2020) +24. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2334-2343 (2017) +25. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7214-7223 (2020) +26. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12858-12868 (2023) +27. Guo, W., Bie, X., Alameda-Pineda, X., Moreno-Noguer, F.: Multi-person extreme motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13053-13064 (2022) +28. Hasler, N., Stoll, C., Sunkel, M., Rosenhahn, B., Seidel, H.P.: A statistical model of human pose and body shape. In: Computer graphics forum. vol. 28, pp. 337-346. Wiley Online Library (2009) +29. Jiang, T., Chen, X., Song, J., Hilliges, O.: Instantavatar: Learning avatars from monocular video in 60 seconds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16922-16932 (2023) +30. Joo, H., Liu, H., Tan, L., Gui, L., Nabbe, B., Matthews, I., Kanade, T., Nobuhara, S., Sheikh, Y.: Panoptic studio: A massively multiview system for social motion capture. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3334-3342 (2015) +31. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7122-7131 (2018) + +32. Kocabas, M., Huang, C.H.P., Hilliges, O., Black, M.J.: Pare: Part attention regressor for 3d human body estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11127-11137 (2021) +33. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: Spec: Seeing people in the wild with an estimated camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11035-11045 (2021) +34. Li, Y., Zhang, S., Wang, Z., Yang, S., Yang, W., Xia, S.T., Zhou, E.: Tokenpose: Learning keypoint tokens for human pose estimation. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 11313-11322 (2021) +35. Lin, J., Lee, G.H.: Multi-view multi-person 3d pose estimation with plane sweep stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11886-11895 (2021) +36. Liu, Y., Gall, J., Stoll, C., Dai, Q., Seidel, H.P., Theobalt, C.: Markerless motion capture of multiple characters using multiview image segmentation. IEEE transactions on pattern analysis and machine intelligence 35(11), 2720-2735 (2013) +37. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015) +38. Luvizon, D.C., Habermann, M., Golyanik, V., Kortylewski, A., Theobalt, C.: Scene-aware 3d multi-human motion capture from a single camera. In: Computer Graphics Forum. vol. 42, pp. 371-383. Wiley Online Library (2023) +39. Martinez, J., Hossain, R., Romero, J., Little, J.J.: A simple yet effective baseline for 3d human pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2640-2649 (2017) +40. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H.P., Xu, W., Casas, D., Theobalt, C.: Vnect: Real-time 3d human pose estimation with a single rgb camera. Acm transactions on graphics (tog) 36(4), 1-14 (2017) +41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +42. Muller, L., Osman, A.A.A., Tang, S., Huang, C.H.P., Black, M.J.: On self-contact and human pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9990-9999 (June 2021) +43. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +44. Osman, A.A., Bolkart, T., Black, M.J.: Star: Sparse trained articulated human body regressor. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16. pp. 598-613. Springer (2020) +45. Patel, P., Huang, C.H.P., Tesch, J., Hoffmann, D.T., Tripathi, S., Black, M.J.: Agora: Avatars in geography optimized for regression analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13468-13478 (2021) +46. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10975-10985 (2019) + +47. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3d human pose and shape from a single color image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 459-468 (2018) +48. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021) +49. Pishchulin, L., Insafutdinov, E., Tang, S., Andres, B., Andriluka, M., Gehler, P.V., Schiele, B.: Deepcut: Joint subset partition and labeling for multi person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4929-4937 (2016) +50. Reddy, N.D., Guigues, L., Pishchulin, L., Eledath, J., Narasimhan, S.G.: Tessen-track: End-to-end learnable multi-person articulated 3d pose tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15190-15200 (2021) +51. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022) +52. Saito, S., Yang, J., Ma, Q., Black, M.J.: Scanimate: Weakly supervised learning of skinned clothed avatar networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2886-2897 (2021) +53. Shuai, Q., Geng, C., Fang, Q., Peng, S., Shen, W., Zhou, X., Bao, H.: Novel view synthesis of human interactions from sparse multi-view videos. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22, Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10.1145/3528233.3530704, https://doi.org/10.1145/3528233.3530704 +54. Shuai, Q., Yu, Z., Zhou, Z., Fan, L., Yang, H., Yang, C., Zhou, X.: Reconstructing close human interactions from multiple views. ACM Transactions on Graphics (dec 2023). https://doi.org/10.1145/3618336 +55. Sun, Y., Bao, Q., Liu, W., Fu, Y., Black, M.J., Mei, T.: Monocular, one-stage, regression of multiple 3d people. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11179-11188 (2021) +56. Sun, Y., Liu, W., Bao, Q., Fu, Y., Mei, T., Black, M.J.: Putting people in their place: Monocular regression of 3d people in depth. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13243-13252 (2022) +57. Te, G., Li, X., Li, X., Wang, J., Hu, W., Lu, Y.: Neural capture of animatable 3d human from monocular video. In: European Conference on Computer Vision. pp. 275-291. Springer (2022) +58. Tu, H., Wang, C., Zeng, W.: Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 197-212. Springer (2020) +59. Wang, C., Li, J., Liu, W., Qian, C., Lu, C.: Hmor: Hierarchical multi-person ordinal relations for monocular multi-person 3d pose estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 242-259. Springer (2020) +60. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022) + +61. Wang, T., Zhang, J., Cai, Y., Yan, S., Feng, J.: Direct multi-view multi-person 3d pose estimation. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 13153-13164. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/6da9003b743b65f4c0ccd295cc484e57-Paper.pdf +62. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022) +63. Wu, S., Jin, S., Liu, W., Bai, L., Qian, C., Liu, D., Ouyang, W.: Graph-based 3d multi-person pose estimation using multi-view images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11148-11157 (2021) +64. Wu, S., Jin, S., Liu, W., Bai, L., Qian, C., Liu, D., Ouyang, W.: Graph-based 3d multi-person pose estimation using multi-view images. In: ICCV (2021) +65. Xu, H., Bazavan, E.G., Zanfir, A., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Ghum & ghuml: Generative 3d human shape and articulated pose models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6184-6193 (2020) +66. Ye, H., Zhu, W., Wang, C., Wu, R., Wang, Y.: Faster voxelpose: Real-time 3d human pose estimation by orthographic projection. In: European Conference on Computer Vision. pp. 142-159. Springer (2022) +67. Yin, Y., Guo, C., Kaufmann, M., Zarate, J.J., Song, J., Hilliges, O.: Hi4d: 4d instance segmentation of close human interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17016-17027 (2023) +68. Zanfir, A., Marinoiu, E., Sminchisescu, C.: Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2148-2157 (2018) +69. Zhang, H., Tian, Y., Zhou, X., Ouyang, W., Liu, Y., Wang, L., Sun, Z.: Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11446-11456 (2021) +70. Zhang, J., Liu, X., Ye, X., Zhao, F., Zhang, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG) 40(4), 1-18 (2021) +71. Zhang, Y., Wang, C., Wang, X., Liu, W., Zeng, W.: Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 2613-2626 (2022) +72. Zhang, Y., An, L., Yu, T., Li, X., Li, K., Liu, Y.: 4d association graph for realtime multi-person motion capture using multiple video cameras. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1324-1333 (2020) +73. Zhang, Y., Li, Z., An, L., Li, M., Yu, T., Liu, Y.: Lightweight multi-person total motion capture using sparse multi-view cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5560-5569 (2021) +74. Zheng, Y., Shao, R., Zhang, Y., Yu, T., Zheng, Z., Dai, Q., Liu, Y.: Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6239-6249 (2021) + +75. Zhou, H., Hong, C., Han, Y., Huang, P., Zhuang, Y.: Mh pose: 3d human pose estimation based on high-quality heatmap. In: 2021 IEEE International Conference on Big Data (Big Data). pp. 3215-3222. IEEE (2021) +76. Zhou, Z., Shuai, Q., Wang, Y., Fang, Q., Ji, X., Li, F., Bao, H., Zhou, X.: Quickpose: Real-time multi-view multi-person pose estimation in crowded scenes. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022) \ No newline at end of file diff --git a/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/images.zip b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..42a15059f26be14b5017a42912d2802bd9bca613 --- /dev/null +++ b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6204c61608bf974c8d1d24a7123c18d04e3d0fb0a2774e8b68f20117552c693f +size 570371 diff --git a/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/layout.json b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..48abc927c120b10ff01bc472c0cabff205fc3b9f --- /dev/null +++ b/2024/AvatarPose_ Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos/layout.json @@ -0,0 +1,10477 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 166 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 166 + ], + "type": "text", + "content": "AvatarPose: Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "spans": [ + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "type": "text", + "content": "Feichi Lu" + }, + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "type": "text", + "content": ", Zijian Dong" + }, + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "type": "text", + "content": ", Jie Song" + }, + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "type": "text", + "content": ", and Otmar Hilliges" + }, + { + "bbox": [ + 157, + 185, + 456, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 178, + 207, + 434, + 230 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 178, + 207, + 434, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 207, + 434, + 219 + ], + "spans": [ + { + "bbox": [ + 178, + 207, + 434, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 178, + 207, + 434, + 219 + ], + "type": "text", + "content": " Department of Computer Science, ETH Zürich, Switzerland," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 219, + 423, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 219, + 423, + 230 + ], + "spans": [ + { + "bbox": [ + 189, + 219, + 423, + 230 + ], + "type": "text", + "content": "2 Max Planck Institute for Intelligent Systems, Germany" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 142, + 241, + 459, + 340 + ], + "blocks": [ + { + "bbox": [ + 142, + 241, + 459, + 340 + ], + "lines": [ + { + "bbox": [ + 142, + 241, + 459, + 340 + ], + "spans": [ + { + "bbox": [ + 142, + 241, + 459, + 340 + ], + "type": "image", + "image_path": "d6c607c5062696b75767920ec8fc75be6e627631d2f90d51e1d5f16262ff6430.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 350, + 482, + 406 + ], + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 406 + ], + "type": "text", + "content": "Fig. 1: We propose AvatarPose, a method for estimating the 3D poses and shapes of multiple closely interacting people from multi-view videos. To this end, we first reconstruct the avatar of each individual and leverage the learned personalized avatars as priors to refine poses via color and silhouette rendering loss. We alternate between avatar refinement and pose optimization to obtain the final pose estimation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 419, + 453, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 419, + 453, + 606 + ], + "spans": [ + { + "bbox": [ + 159, + 419, + 453, + 606 + ], + "type": "text", + "content": "Abstract. Despite progress in human motion capture, existing multiview methods often face challenges in estimating the 3D pose and shape of multiple closely interacting people. This difficulty arises from reliance on accurate 2D joint estimations, which are hard to obtain due to occlusions and body contact when people are in close interaction. To address this, we propose a novel method leveraging the personalized implicit neural avatar of each individual as a prior, which significantly improves the robustness and precision of this challenging pose estimation task. Concretely, the avatars are efficiently reconstructed via layered volume rendering from sparse multi-view videos. The reconstructed avatar prior allows for the direct optimization of 3D poses based on color and silhouette rendering loss, bypassing the issues associated with noisy 2D detections. To handle interpenetration, we propose a collision loss on the overlapping shape regions of avatars to add penetration constraints. Moreover, both 3D poses and avatars are optimized in an alternating manner. Our experimental results demonstrate state-of-the-art performance on several public datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "spans": [ + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": "Keywords: human pose estimation " + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": " human close interaction " + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": " multiview pose estimation " + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": " avatar prior" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 642, + 222, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 642, + 222, + 653 + ], + "spans": [ + { + "bbox": [ + 135, + 642, + 222, + 653 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 653, + 270, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 270, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 270, + 665 + ], + "type": "text", + "content": "† Now at HKUST(GZ)&HKUST" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 246 + ], + "type": "text", + "content": "People frequently engage in interactions with each other in daily life to offer physical support or convey emotions. For AI systems to interpret 3D human interactions, a foundational step is to reconstruct 3D human poses and shapes using consumer-grade sensors, such as cameras. However, in closely interacting scenarios, such as hugging or kissing, 3D pose estimations face challenges due to strong occlusions. The problem of depth ambiguity worsens when close contact is involved, making it more difficult to predict 3D poses from 2D images or videos. Consequently, a multi-camera setup becomes essential to provide additional observations and to address depth ambiguity in pose estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 437 + ], + "type": "text", + "content": "Despite the ubiquity of close human interactions, the study of estimating such human motions is scarce. Most previous multi-human methods with a multi-view setup [2-4, 19, 22, 30, 58, 66, 72] focus on scenarios where people are at a distance from each other. Some methods [2, 4, 19, 22, 72] typically formulate the problem as a cross-view matching problem, relying heavily on 2D joint estimations for subsequent 3D triangulation. These methods demonstrate high sensitivity to noisy or missing 2D joint estimations, particularly when occlusion occurs. Another group of learning-based methods [58, 61, 63, 66] attempts to integrate 2D features from each view into a 3D voxel space and predict 3D human poses from identified 3D subvolumes of each individual. These methods are more robust to occlusion, but they encounter challenges with generalization and are sensitive to distribution shifts due to the lack of annotated 3D data. To correct abnormal or missing pose estimations, some methods [18, 21] leverage parametric body models like SMPL [37] as full-body priors and fit these 3D models to 2D joint estimations. Although this approach alleviates the issue of abnormal or missing joint estimations, it remains constrained by noisy 2D detections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 438, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 628 + ], + "type": "text", + "content": "Embracing the challenging problem of pose estimation for close interactions, we propose a novel method to estimate the 3D poses and shapes of multiple people observed from a sparse set of cameras. Our goal is to ensure that even in close contact, the estimated human poses and shapes are accurate and free from interpenetrations. The key to our method is reconstructing implicit textured avatars of each individual in the scene and leveraging them as a strong personalized prior for pose optimization (Figure 1). In contrast to relying solely on noisy 2D joint detections, this textured avatar prior enables us to leverage pixel-wise color and silhouette information for pose refinement, significantly increasing the robustness and accuracy of our method (Table 3). Meanwhile, the reconstructed avatar provides crucial geometric and appearance information to avoid collisions between individuals. Compared to methods using SMPL body shape [25,42] to penalize collisions, our implicit avatar model contains a more detailed clothed shape and additional appearance cues and enables efficient computation of penetration loss. Due to the mutually beneficial relationship between avatar and pose, we alternate between pose optimization and avatar refinement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "More specifically, to accelerate the learning of avatars, we model each human individual in canonical space using an efficient neural radiance field variant in Instant NGP [43] and combine it with an efficient SMPL-based deformation" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": "module [37]. To learn and render avatar models of multiple people, we adapt layered volume rendering [53, 70] to our avatar model. This adaptation allows us to jointly optimize all avatar models through a straightforward rendering loss. Once learned, the avatar can be animated and rendered based on pose parameters at interactive rates, thus naturally serving as an efficient personalized textured prior for pose optimization. With the learned personalized prior, we optimize pose via a novel objective function. Different from previous methods based on 2D reprojection error of joints, we directly optimize pose parameters via minimizing color and silhouette rendering losses while keeping the learned avatar model fixed. To prevent interpenetration between human individuals, a collision loss is introduced by penalizing the situation when a 3D point is occupied by multiple avatars. To remove artifacts in the initial avatar due to imperfect pose initialization, we further refine avatars based on optimized poses. Throughout the optimization process, both initial personalized avatar models and SMPL parameters are optimized in an alternating manner, motivated by the insight that accurate 3D pose estimations improve avatar learning, and improved avatar models, in turn, increase the precision of overall pose estimations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 367 + ], + "type": "text", + "content": "We experimentally demonstrate that our method significantly outperforms previous state-of-the-art methods on several public datasets both quantitatively (Table 1 and Table 2) and qualitatively (Figure 3) especially when people are in close interaction. In summary, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 375, + 479, + 479 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 375, + 479, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 479, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 479, + 409 + ], + "type": "text", + "content": "- We propose a pipeline that efficiently creates implicit neural avatars of closely interacting people and leverages the learned avatars as priors to optimize poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 411, + 479, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 479, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 479, + 455 + ], + "type": "text", + "content": "- The avatar prior enables us to design a novel objective function that leverages color and mask rendering loss for pose optimization. We show the superiority of this loss function compared to the 2D reprojection error of 3D joints used by most of the previous methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 459, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 459, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 137, + 459, + 479, + 479 + ], + "type": "text", + "content": "- Based on the learned avatar, a collision loss is introduced to avoid penetration when individuals are in contact." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 501, + 237, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 237, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 237, + 514 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 526, + 330, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 330, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 330, + 537 + ], + "type": "text", + "content": "2.1 Multi-Person 3D Pose Estimation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Despite significant progress in multi-human 2D pose estimation [9,13,15,24,34, 49] and 3D pose estimation from monocular images or video [7, 31-33, 38-40, 45, 47, 55, 56, 59, 68, 69], the reconstruction accuracy is still limited due to depth ambiguity and strong occlusions when humans are in close contact. A multi-view setting [4, 5, 19, 22, 23, 35, 36, 54, 61, 63, 66, 72] helps to alleviate these challenges. One straightforward idea of most methods [2, 4, 19, 22, 72, 76] is to formulate the problem into cross-view matching and association problems. MVPose [19] performs 2D person parsing in each image and leverages cross-view person matching to infer 3D pose. 4DAssociation [72] additionally adds tracking into this process to form a unified graph for associating 4D information. However, these methods" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "AvatarPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "are sensitive to the noisy estimation of 2D pose. In contrast to these matching-based methods, some recent methods [5,14,17,23,35,50,58,61,63,66,71,75] directly learn deep neural networks to regress poses. Faster Voxelpose [66] employs the feature volume proposed by Voxelpose [58] and enhances computational efficiency. Graph [63] designs three graph neural network models for human center detection and pose estimation. MvP [61] simplifies the multi-person pose estimation by direct regression using the transformer model. A concurrent method CloseMocap [54] proposes to learn a model from a synthetic dataset simulating occlusion situations. However, heavily relying on the 2D or 3D features as input during training, these learning-based methods suffer from generalization issues when subjects, motions, and camera configurations change." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 356 + ], + "type": "text", + "content": "To further improve the robustness, a statistical parametric body model such as SMPL [37] is explored in [21] as a regularization prior for 3D joint refinement. Some follow-ups [18,73] show that parametric models help in correcting implausible 3D pose estimates and filling in missing joints. However, this coarse body prior highly relies on aligning the 3D joints to 2D pose estimations, which are inaccurate when occlusion happens. Different from all of these previous methods, we explore the usage of personalized textured avatar models as priors to refine human poses. This prior enables us to leverage color and silhouette information from multi-view observation for refining poses of closely interacting humans." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 372, + 268, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 268, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 268, + 384 + ], + "type": "text", + "content": "2.2 3D Human Modeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 594 + ], + "type": "text", + "content": "Parametric human body models [1,28,37,44,65] can represent minimally clothed human shapes by deforming a template mesh. It is challenging to extend this explicit representation for modeling clothed humans due to the fixed topology and resolution. To overcome this limitation, methods such as SNARF [12] and SCANimate [52] propose to model articulated human shapes based on 3D implicit representations. Many works [6,8,10,20,26,29,48,57,60,62] fit implicit neural fields to RGB or RGB-D videos by neural rendering to reconstruct the shape and appearance of a single human body. However, when applied to a multi-human scene, these methods are not able to achieve good fidelity due to strong occlusions. Recent methods including ST-NeRF [70] and [53] leverage layered neural representation to model multiple humans with sparse multi-view videos and thus can generate novel view synthesis of dynamic multiple humans. The main problem of all aforementioned approaches, however, is their high reliance on 3D human pose estimation [70,74]: the deformation of the human model requires accurate human poses, which is hard to obtain when people are in close interaction. To address this challenge, our method is orthogonal to others, aiming at leveraging the learned avatar as priors for pose estimation in close interaction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 610, + 372, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 372, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 372, + 623 + ], + "type": "text", + "content": "2.3 3D Human Datasets for Close Interaction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Most of the existing datasets [3,11,30] like Shelf and Panoptic studying multiperson pose estimation focus on the scene where people are at a distance from each other and rarely involved in close interactions. To study close interactions" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "among people, MultiHuman [74] dataset captures multi-person interaction with some close interactions and occlusions. ExPI [27] creates a multi-person extreme motion dataset with close interactions, but it focuses mainly on motion prediction for future frames instead of pose estimation from sparse views. CHI3D [25] captures two-person interaction datasets and proposes to learn a contact estimation module from annotations to improve the precision of pose estimation. The most recent work Hi4D [67] creates a challenging dataset of physically close human interaction and proposes a method to disentangle human bodies and estimate poses. However, this method relies on the 3D ground truth of clothed human meshes captured with expensive 3D body scanners. Due to the reliance on the limited annotated 3D data, this method faces challenges in generalization with different people and camera configurations. In contrast to all the works before, we intend to solve the problem of pose estimation in close human interaction without requiring accurate 3D scans or other training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 303, + 202, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 202, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 202, + 316 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 328, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 400 + ], + "type": "text", + "content": "Given a dynamic scene captured by a sparse set of RGB cameras, our goal is to estimate the 3D pose and shape of multiple people even if they interact closely. To address this challenging task, our key idea is to first reconstruct the personalized avatar of each individual in the scene and leverage them as a strong prior to refine the appearance and pose in an alternating manner. An overview of our method is shown in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 472 + ], + "type": "text", + "content": "We first introduce an efficient pipeline to create avatars of multiple people in a scene (Section 3.1 and Figure 2(a)). Specifically, we leverage an accelerated neural radiance field to represent the shape and appearance of each individual in canonical space and deform it at an interactive rate. We then adapt layered volume rendering to our pipeline, which composites the rendering of avatars into one image, thus enabling direct learning from multi-view video inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 473, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 557 + ], + "type": "text", + "content": "Thanks to the learned avatar prior for each individual, we can enhance 3D pose optimization via a combination of RGB and silhouette rendering loss (Section 3.2 and Figure 2(b)). While previous work heavily relies on noisy 2D joint detection, we show that employing such pixel-wise color and silhouette information can largely increase precision and robustness. Moreover, a collision loss is introduced to avoid interpenetration. Finally, we alternate between avatar learning and pose optimization to get complete and accurate 3D human poses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 574, + 304, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 304, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 304, + 586 + ], + "type": "text", + "content": "3.1 Multi-Avatar Prior Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 595, + 481, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 595, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 595, + 481, + 632 + ], + "type": "text", + "content": "Avatar Model We represent each human individual in canonical space using an accelerated neural radiance field [43] and model shape-aware articulated deformation based on SMPL [37]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 639, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 639, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 136, + 639, + 482, + 667 + ], + "type": "text", + "content": "- Canonical Appearance Representation: To model human shape and appearance, we create canonical radiance field " + }, + { + "bbox": [ + 136, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}" + }, + { + "bbox": [ + 136, + 639, + 482, + 667 + ], + "type": "text", + "content": " for each human instance" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "AvatarPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 115, + 465, + 221 + ], + "blocks": [ + { + "bbox": [ + 143, + 115, + 465, + 221 + ], + "lines": [ + { + "bbox": [ + 143, + 115, + 465, + 221 + ], + "spans": [ + { + "bbox": [ + 143, + 115, + 465, + 221 + ], + "type": "image", + "image_path": "9b4547620a3b3a8d47af20c8bfe126a4cdda69f8bc7a9752685987b415f28558.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 218, + 471, + 310 + ], + "blocks": [ + { + "bbox": [ + 143, + 218, + 471, + 310 + ], + "lines": [ + { + "bbox": [ + 143, + 218, + 471, + 310 + ], + "spans": [ + { + "bbox": [ + 143, + 218, + 471, + 310 + ], + "type": "image", + "image_path": "8091daf5024b3007c79f566d9dfa3b8cbe004e3914731505176196066d7cd96c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": "Fig. 2: Method Overview: Our method consists of two modules: (a) Multi-Avatar Prior Learning: Given the input multi-view images and estimated poses " + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\Theta^{(l)}" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": ", we sample points " + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(l)}" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": " for each individual " + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": " along the rays in posed space and warp these points into canonical space and calculate their color " + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{(l)}" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\sigma^{(l)}" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": " via the canonical appearance network " + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": ". We leverage layered volume rendering [70] to attain the final pixel color and compare it with the original input image to optimize the parameters of avatars. (b) Avatar-guided Pose Optimization: Given learned avatar model " + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": " and initial poses " + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\Theta^{(l)}" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": " of each individual " + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 320, + 482, + 465 + ], + "type": "text", + "content": ", we deform all of the avatars based on SMPL-based deformer and render them jointly via layered volume rendering. We compare the composite rendering with input observation and minimize the RGB and mask rendering loss to optimize poses. A collision loss is additionally introduced to avoid interpenetration. Finally, we alternate between two modules to obtain the final result. For clarity, the parameters to be optimized are marked as red in each module." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "inline_equation", + "content": "l \\in [1, L]" + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "text", + "content": " is the number of humans in the scene. " + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}" + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "text", + "content": " takes a 3D point " + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^{(l)}" + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "text", + "content": " as input and predicts its density " + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\sigma^{(l)}" + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{(l)}" + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "text", + "content": ". Following Instant-NGP [43] and InstantAvatar [29] to accelerate the rendering, " + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{\\sigma_f}^{(l)}" + }, + { + "bbox": [ + 147, + 490, + 482, + 542 + ], + "type": "text", + "content": " is parameterized via using a hash table to store feature grids at different scales." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 544, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "text", + "content": "- Pose Representation: We represent the 3D pose and underlying body shape for all human instances by SMPL parameters " + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\Theta^{(l)}\\}_{l\\in [1,L]}" + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "text", + "content": ". For each human " + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\Theta^{(l)} = \\{\\beta^{(l)},\\theta^{(l)},t^{(l)}\\}" + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "text", + "content": " contains shape parameters " + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\beta^{(l)}\\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "text", + "content": ", pose parameters " + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\theta^{(l)}\\in \\mathbb{R}^{72}" + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "inline_equation", + "content": "t^{(l)}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 138, + 544, + 482, + 592 + ], + "type": "text", + "content": " of SMPL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": "- Deformer: To enable animation given targeted poses " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{(l)}" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": ", we require the radiance field in the posed space. Given a point " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(l)}" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": " in deformed space of human " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": ", we determine the corresponding canonical point " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^{(l)}" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": " by inverse linear blend skinning(LBS) [37]: " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^{(l)}(\\mathbf{x}^{(l)},\\Theta^{(l)}) = (\\sum_{i = 1}^{n_b}w_i(\\Theta^{(l)})\\mathbf{B}_i(\\Theta^{(l)}))^{-1}\\mathbf{x}^{(l)}" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": " is the rigid bone transformation matrix for joint " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots,n_b\\}" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": " under pose " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{(l)}" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": " is the skinning weights of the nearest neighbor of " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(l)}" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": " in" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 115, + 479, + 140 + ], + "type": "text", + "content": "the deformed SMPL vertices. We obtain the radiance field at the point " + }, + { + "bbox": [ + 148, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(l)}" + }, + { + "bbox": [ + 148, + 115, + 479, + 140 + ], + "type": "text", + "content": " by evaluating the canonical appearance field at the corresponding point " + }, + { + "bbox": [ + 148, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^{(l)}" + }, + { + "bbox": [ + 148, + 115, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "text", + "content": "Layered Volume Rendering To obtain the pixel value for a ray " + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in \\mathcal{R}" + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "text", + "content": ", we raycast every human instance separately with a layered rendering strategy similar to ST-NeRF [70]. Specifically, we first calculate the intersection points between the ray and the 3D bounding box of each human instance and uniformly sample points in each bounding box. To distinguish different identities, we assign each sampled point " + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "text", + "content": " a one-hot representation " + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i = [m_i^{(1)}, \\dots, m_i^{(L)}]" + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "text", + "content": " to indicate which human identity it belongs to. After sorting all sampled points by their depth values and calculating their corresponding color " + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\sigma_i" + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "text", + "content": " from the avatar model, if " + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "inline_equation", + "content": "m_i^{(l)} = 1" + }, + { + "bbox": [ + 132, + 151, + 482, + 262 + ], + "type": "text", + "content": ", we compute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 251, + 270, + 481, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 270, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 251, + 270, + 481, + 285 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {i}, \\sigma_ {i} = \\bar {\\mathbf {F}} _ {\\sigma_ {f}} ^ {(l)} \\left(\\bar {\\mathbf {x}} _ {i} \\left(\\mathbf {x} _ {i}, \\boldsymbol {\\Theta} ^ {(l)}\\right)\\right). \\tag {1}", + "image_path": "2096953163a64d4319426848fd003284a7bf2c12377a2ad06381999615524d81.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 293, + 432, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 293, + 432, + 306 + ], + "spans": [ + { + "bbox": [ + 146, + 293, + 432, + 306 + ], + "type": "text", + "content": "The color of each ray is computed via numerical integration [41]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 198, + 316, + 480, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 316, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 198, + 316, + 480, + 347 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {C}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\prod_ {j < i} (1 - \\alpha_ {j}) \\mathbf {c} _ {i} \\quad \\alpha_ {i} = 1 - \\exp (- \\sigma_ {i} \\delta_ {i}), \\tag {2}", + "image_path": "e32675990ddb9c42749ee4258e9ba0c63ca6440fa0723fd74736ce3f49733f3c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 355, + 479, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 355, + 479, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 355, + 479, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 355, + 479, + 379 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 132, + 355, + 479, + 379 + ], + "type": "text", + "content": " is the distance between samples. The accumulated alpha value, which represents ray opacity, can be computed via:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 252, + 388, + 480, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 388, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 252, + 388, + 480, + 420 + ], + "type": "interline_equation", + "content": "\\alpha (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\prod_ {j < i} (1 - \\alpha_ {j}). \\tag {3}", + "image_path": "4119d8fdd8b92f6a1df6e4dfba34aa246763d4a79eb432e55f68148548a12dbe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 429, + 479, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 479, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 479, + 452 + ], + "type": "text", + "content": "For each human identity " + }, + { + "bbox": [ + 132, + 429, + 479, + 452 + ], + "type": "inline_equation", + "content": "l \\in [1, L]" + }, + { + "bbox": [ + 132, + 429, + 479, + 452 + ], + "type": "text", + "content": ", the corresponding instance ray opacity can be calculated via:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 238, + 460, + 480, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 238, + 460, + 480, + 492 + ], + "type": "interline_equation", + "content": "\\alpha^ {(l)} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\prod_ {j < i} (1 - \\alpha_ {j}) m _ {i} ^ {(l)}. \\tag {4}", + "image_path": "62697083118a63f2c3089d376ad81591496509595a73b8961b5f4a94fb1199b7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 502, + 479, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 479, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 479, + 539 + ], + "type": "text", + "content": "Training The overall training process is shown in Figure 2(a). For training avatar layers, we minimize the Huber loss " + }, + { + "bbox": [ + 132, + 502, + 479, + 539 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 132, + 502, + 479, + 539 + ], + "type": "text", + "content": " between the predicted pixel color " + }, + { + "bbox": [ + 132, + 502, + 479, + 539 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}(\\mathbf{r})" + }, + { + "bbox": [ + 132, + 502, + 479, + 539 + ], + "type": "text", + "content": " and the ground truth pixel color " + }, + { + "bbox": [ + 132, + 502, + 479, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{gt}(\\mathbf{r})" + }, + { + "bbox": [ + 132, + 502, + 479, + 539 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 223, + 548, + 480, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 548, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 223, + 548, + 480, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R G B} = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} \\rho \\left(\\left\\| \\hat {\\mathbf {C}} (\\mathbf {r}) - \\mathbf {C} _ {g t} (\\mathbf {r}) \\right\\|\\right). \\tag {5}", + "image_path": "6b3cdc433adb7cbeaec21181125d87bfd03954275a07f6429ea09bf9b85bcff4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 583, + 479, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 583, + 479, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 479, + 631 + ], + "type": "text", + "content": "Since instance segmentation of human performers is hard to obtain and is not accurate, we choose foreground segmentation as our mask supervision, which is obtained via SAM-Track [16]. We apply a loss for optimizing the rendered alpha values " + }, + { + "bbox": [ + 132, + 583, + 479, + 631 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 583, + 479, + 631 + ], + "type": "text", + "content": " to reduce the artifacts in the floating area:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 225, + 640, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 640, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 225, + 640, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l p h a}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} \\in \\mathcal {R}} (\\alpha (\\mathbf {r}) - \\alpha_ {S A M} (\\mathbf {r})) ^ {2}. \\tag {6}", + "image_path": "0ba8b97a73fa053ffe90f8bca23577a9b042fa203aa5c946d4a9aef4524dbd32.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "AvatarPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Following [53], we add a regularization loss for instance alpha values to make sure every pixel can only be rendered from one human layer:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 150, + 481, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 209, + 150, + 481, + 183 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l a y e r}} = - \\frac {1}{L \\mid \\mathcal {R} \\mid} \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\sum_ {l = 1} ^ {L} \\alpha^ {(l)} (\\mathbf {r}) \\log \\left(\\alpha^ {(l)} (\\mathbf {r})\\right). \\tag {7}", + "image_path": "b30e4e0ed1ec9e09b79345785723953071554bf68b16001ebe7fa7cbec075b98.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 187, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 187, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 131, + 187, + 479, + 212 + ], + "type": "text", + "content": "Similar to [29], we also add hard surface and density regularization terms in the learning process. More training details can be found in the Supp Mat." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 229, + 328, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 328, + 242 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 328, + 242 + ], + "type": "text", + "content": "3.2 Avatar-guided Pose Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 480, + 357 + ], + "type": "text", + "content": "Equipped with learned avatars obtained in Section 3.1, we aim to estimate the 3D shape and pose of multiple humans with close physical contact. To achieve this, we leverage the avatars as priors to handle the challenging pose ambiguities caused by contact. Specifically, we first initialize pose parameters using an off-the-shelf 3D pose estimator [72] and refine the pose via a rendering loss between rendered posed avatars and 2D observations (Figure 2(b)). Since the initial imperfect pose estimations may cause artifacts in avatar reconstruction, we further refine the weights of the avatar model using the optimized pose. This process is formulated as an alternating optimization to refine both poses and avatars." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "type": "text", + "content": "Initialization The initial 3D pose proposals are estimated by leveraging the off-the-shelf 3D human pose estimator [72]. After that, we register a SMPL model to the estimated 3D joints to obtain the initial pose parameters " + }, + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "type": "inline_equation", + "content": "\\Theta_0" + }, + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "type": "text", + "content": ". Given these estimated poses, the initial avatar model of each individual is further learned from multi-view videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 441, + 479, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 479, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 479, + 466 + ], + "type": "text", + "content": "Objective To leverage avatars as priors to tackle challenges caused by contact, we optimize the following objective:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 475, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 475, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 228, + 475, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\boldsymbol {\\Theta}) = \\lambda_ {R G B} \\mathcal {L} _ {R G B} (\\boldsymbol {\\Theta}) + \\lambda_ {\\alpha} \\mathcal {L} _ {\\alpha} (\\boldsymbol {\\Theta}) \\tag {8} \\\\ + \\lambda_ {r e g} \\mathcal {L} _ {r e g} (\\Theta) + \\lambda_ {p a} \\mathcal {L} _ {p a} (\\Theta). \\\\ \\end{array}", + "image_path": "f486e87c976285b026bd6d437c90cfedf9a0da8abb7e1a173db6a8b0d4b8a333.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 480, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 570 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 510, + 480, + 570 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 510, + 480, + 570 + ], + "type": "text", + "content": " is the SMPL parameters to be optimized for all avatars, which is also consistent with the pose parameters to deform the avatar model (Section 3.1). Following [62], we represent " + }, + { + "bbox": [ + 130, + 510, + 480, + 570 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 510, + 480, + 570 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 510, + 480, + 570 + ], + "type": "inline_equation", + "content": "\\Theta_0 + \\mathrm{MLP}(\\Theta_0)" + }, + { + "bbox": [ + 130, + 510, + 480, + 570 + ], + "type": "text", + "content": " and refine poses by changing the parameters of the neural network. This representation empirically shows more robust results compared to directly optimizing SMPL parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "text", + "content": "Different from previous methods [18,21,58,72], the personalized multi-ayar prior allows us to leverage appearance and silhouette information to refine initial poses. Specifically, we calculate " + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{RGB}" + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "text", + "content": " to ensure the color consistency between the rendered pixel " + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}} (\\mathbf{r},\\boldsymbol {\\Theta})" + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "text", + "content": " (Equation (1) and Equation (2)) of deformed avatar with poses " + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "text", + "content": " and the corresponding ground-truth pixel color " + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{gt}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 208, + 639, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 639, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 208, + 639, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R G B} (\\boldsymbol {\\Theta}) = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} \\rho \\left(\\left\\| \\hat {\\mathbf {C}} (\\mathbf {r}, \\boldsymbol {\\Theta}) - \\mathbf {C} _ {g t} (\\mathbf {r}) \\right\\|\\right). \\tag {9}", + "image_path": "71e8004ccc3210557f2bcf390f31c074d38864dab64816c63afdd8e2bf244d40.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Additionally, a cross-entropy loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\alpha}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " is introduced to ensure that the rendered mask " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\alpha (\\mathbf{r},\\Theta)" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " (Equation (3)) of the reposed avatar is aligned with the estimated SAM-Track mask " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\alpha_{SAM}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 162, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 162, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 225, + 162, + 481, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\alpha} (\\boldsymbol {\\Theta}) = - \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\alpha_ {S A M} (\\mathbf {r}) \\log (\\alpha (\\mathbf {r}, \\boldsymbol {\\Theta})) \\tag {10}", + "image_path": "265af6f5ff13b1fb4ab8f3e33568cc28f460cff711978bf812eeefb4331ba408.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 479, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 479, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 479, + 226 + ], + "type": "text", + "content": "To penalize the unnatural poses and avoid elbows and knees bending in the wrong direction, we add an L2 regularization term and combine it with the pose prior in SMPLify [7] to constrain physically implausible joint rotation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 233, + 233, + 481, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 233, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 233, + 233, + 481, + 258 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} (\\boldsymbol {\\Theta}) = \\| \\boldsymbol {\\Theta} \\| _ {2} + \\lambda \\sum_ {i \\in I} \\exp (\\boldsymbol {\\Theta} _ {i}), \\tag {11}", + "image_path": "b868e9caf3de50c3ae554a7807e38419b4c478d42043cb926a4add16232e7042.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 263, + 449, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 263, + 449, + 275 + ], + "spans": [ + { + "bbox": [ + 146, + 263, + 449, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 146, + 263, + 449, + 275 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 146, + 263, + 449, + 275 + ], + "type": "text", + "content": " is the set of pose indices corresponding to elbows and knees." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "text", + "content": "A key challenge to correctly estimate poses in close interaction is to handle interpenetration. Since every avatar is modeled separately, the surfaces tend to intersect when they are in contact. To handle this, we first select sampled points inside multiple instances as " + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{\\mathbf{x}_i\\mid \\alpha_i^{(p)} > 0,\\alpha_i^{(q)} > 0,p,q\\in [1,L],p\\neq q\\}" + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{(p)}" + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "text", + "content": " is calculated from Equation (1) and Equation (2) corresponding with a point " + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "inline_equation", + "content": "m_i^{(p)} = 1" + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "text", + "content": "). We then propose a collision loss " + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pa}" + }, + { + "bbox": [ + 130, + 276, + 480, + 351 + ], + "type": "text", + "content": " for penalizing penetration:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 360, + 481, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 360, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 228, + 360, + 481, + 387 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p a} (\\boldsymbol {\\Theta}) = \\frac {1}{| \\mathcal {S} |} \\sum_ {\\boldsymbol {x} _ {i} \\in \\mathcal {S}} \\alpha_ {i} ^ {(p)} (\\boldsymbol {\\Theta}) \\alpha_ {i} ^ {(q)} (\\boldsymbol {\\Theta}) \\tag {12}", + "image_path": "139b26c1625e52a278df705915488a7055c6ee585131adb63b42738f1516938a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 391, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 479, + 426 + ], + "type": "text", + "content": "Intuitively, this loss guarantees every sample point in 3D space can not be occupied by multiple avatar models simultaneously, which guarantees better pose estimation in close contact." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 436, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 480, + 496 + ], + "type": "text", + "content": "Alternating Optimization Since artifacts sometimes appear on initial avatars due to imperfect pose initialization, we further refine avatars based on optimized poses via minimizing the loss function in Section 3.1. Finally, the optimization of poses and avatars is formulated in an alternating fashion for " + }, + { + "bbox": [ + 130, + 436, + 480, + 496 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 436, + 480, + 496 + ], + "type": "text", + "content": " steps. More details of optimization can be found in the Supp Mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 513, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 230, + 528 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 536, + 481, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 596 + ], + "type": "text", + "content": "Datasets. We mainly evaluate our proposed method on Hi4D [67] and CHI3D [25], which are challenging datasets of two humans in close interaction. To demonstrate the generalization ability of our method for more than two people, we also evaluate on Shelf [3] and MultiHuman [74] including three or four people. More details are shown in the Supp Mat." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "Metrics. We use the Mean Per Joint Position Error (MPJPE) [58] to measure the distance between the ground truth 3D poses and the estimated poses. We also choose the Percentage of Correct Parts (PCP3D) metric [19] to calculate the percentage of correct parts. " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_K" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " [58] and Recall [21] are also leveraged to evaluate performance. More details are shown in the Supp Mat." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "AvatarPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 481, + 310 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 481, + 310 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 481, + 310 + ], + "type": "image", + "image_path": "a33cfa3ca61ced7bb2f4e29ba22b051cfa9400ab5079e3e8eeeb0722bc712573.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 318, + 482, + 363 + ], + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 363 + ], + "type": "text", + "content": "Fig. 3: Qualitative Comparison with SotA methods [19,61,64,66,72] on Hi4D and CHI3D. We show two examples from the Hi4D and CHI3D datasets compared with Graph, MvP, Faster VoxelPose, MVPose, and 4DAssociation. For each example, we show 2D projections on two sampled views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 388, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 473 + ], + "type": "text", + "content": "Baselines. We compare our method on the Hi4D dataset with state-of-the-art methods in three categories discussed in Section 2.1. For learning-based methods, we choose Graph [64], MvP [61], and Faster VoxelPose [66] and fine-tune these models on a subset of the evaluating datasets. For pure association-based methods, we choose 4DAssociation [72] and MVPose* [19]. MVPose [18] adds temporal tracking and SMPL prior to MVPose* [19] and is regarded as a SMPL-guided method. More details about baselines can be found in the Supp Mat." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 494, + 264, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 264, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 264, + 506 + ], + "type": "text", + "content": "4.1 Comparison to SotA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 517, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 567 + ], + "type": "text", + "content": "Table 1 and Table 2 summarizes our quantitative comparisons on Hi4D and CHI3D with SotA (State-of-the-art) methods. Our method largely outperforms the other SotA methods in all of the metrics including MPJPE, PCP3D, and " + }, + { + "bbox": [ + 130, + 517, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_K" + }, + { + "bbox": [ + 130, + 517, + 482, + 567 + ], + "type": "text", + "content": ". More comparisons on Shelf and MultiHuman are shown in the Supp Mat." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Comparison with Graph [63], MvP [61], Faster VoxelPose [66]. Our method outperforms Graph, MvP, and Faster VoxelPose on Hi4D and CHI3D. These methods are prone to overfitting to training pose distributions, thus struggling with challenging poses of interacting actors. Specifically, Faster VoxelPose sometimes misses actors in close contact, leading to a relatively low recall. While Graph and MvP have better recall, Graph fails to consistently track actors across frames, and MvP results in many misaligned joints between actors." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 114, + 473, + 217 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 473, + 217 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 473, + 217 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 473, + 217 + ], + "type": "table", + "html": "
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑Recall(%) ↑
MvP [61]92.7774.1441.5963.8693.84
Graph [63]89.6271.5544.7567.3393.31
Faster VoxelPose [66]68.4073.6744.0568.7083.55
MVPose* [19]53.0587.5767.9780.2893.80
MVPose [18]42.6390.7671.7990.1993.30
4DAssociation [72]41.2988.6280.8797.2798.78
Ours32.1096.9091.4897.3398.78
", + "image_path": "e68c2567b5663aa2dcaf27d0d178a74cc997632be7f96c3ccb888ad62225fdf7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 292, + 481, + 400 + ], + "blocks": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "text", + "content": "Table 1: Quantitative Comparison with SotA on the Hi4D [67] Dataset (8 views). We compare our method with MvP [61], Graph [63], Faster VoxelPose [66], MVPose* [19], MVPose [18] and 4DAssociation [72]. We report MPJPE, PCP, " + }, + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_K" + }, + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "text", + "content": ", and Recall metric for all methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 292, + 481, + 400 + ], + "lines": [ + { + "bbox": [ + 134, + 292, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 134, + 292, + 481, + 400 + ], + "type": "table", + "html": "
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑Recall(%) ↑
MvP [61]55.3889.4763.5892.5399.06
Graph [63]45.3392.4674.0295.2599.17
Faster VoxelPose [66]67.8178.4129.2882.8893.34
MVPose* [19]50.4290.3969.1375.7288.72
MVPose [18]34.0593.3579.9486.9188.18
4DAssociation [72]37.4799.3089.6698.6799.85
Ours32.9899.7993.2099.7999.85
", + "image_path": "7d252e6b6a06cb16418533ccfdc937da198205d55d6dd6544df50df423eb4d50.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 401, + 482, + 444 + ], + "lines": [ + { + "bbox": [ + 132, + 401, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 482, + 444 + ], + "type": "text", + "content": "Table 2: Quantitative Comparison with SotA on the CHI3D [25] Dataset (4 views). We compare our method with Faster VoxelPose [66], MVPose* [19], MV-Pose [18] and 4DAssociation [72]. We report MPJPE, PCP, " + }, + { + "bbox": [ + 132, + 401, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_K" + }, + { + "bbox": [ + 132, + 401, + 482, + 444 + ], + "type": "text", + "content": ", and Recall metric for all methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 487, + 481, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 559 + ], + "type": "text", + "content": "Comparison with MVPose [18]. Our method achieves much better MPJPE and precision than MVPose. MVPose relies heavily on noisy 2D joint detected in close interactions. In contrast, our method leverages color and silhouette rendering loss to optimize poses, leading to robustness to occlusions. Additionally, this top-down method cannot detect actors correctly with strong occlusions, as shown in Figure 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 593, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 481, + 666 + ], + "type": "text", + "content": "Comparison with 4DAssociation [72]. Finally, we compare our method with the bottom-up association method. As shown in Table 1 and Table 2, our method outperforms 4DAssociation in most metrics. Figure 3 shows that when actors are close, this bottom-up method is inclined to associate joints with the wrong human instances. This is because they solve the joint association with a greedy algorithm, which is sensitive to missing and inaccurate 2D joint detections." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "AvatarPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 246, + 190 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 246, + 190 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 246, + 190 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 246, + 190 + ], + "type": "image", + "image_path": "ca7495f30530c603fa0431f97c957acb56b029ff68e8cbd66cc4a9a64aa534b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 192, + 196, + 202 + ], + "lines": [ + { + "bbox": [ + 184, + 192, + 196, + 202 + ], + "spans": [ + { + "bbox": [ + 184, + 192, + 196, + 202 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 251, + 116, + 361, + 191 + ], + "blocks": [ + { + "bbox": [ + 251, + 116, + 361, + 191 + ], + "lines": [ + { + "bbox": [ + 251, + 116, + 361, + 191 + ], + "spans": [ + { + "bbox": [ + 251, + 116, + 361, + 191 + ], + "type": "image", + "image_path": "5751a58588ccc2b09ebe946de64833b59a2e54620d3e5d31683ecd3549ab6732.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 192, + 312, + 202 + ], + "lines": [ + { + "bbox": [ + 301, + 192, + 312, + 202 + ], + "spans": [ + { + "bbox": [ + 301, + 192, + 312, + 202 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 117, + 478, + 192 + ], + "blocks": [ + { + "bbox": [ + 365, + 117, + 478, + 192 + ], + "lines": [ + { + "bbox": [ + 365, + 117, + 478, + 192 + ], + "spans": [ + { + "bbox": [ + 365, + 117, + 478, + 192 + ], + "type": "image", + "image_path": "79063d68ec8bb730e85d5cffc5f86bf63ba694a472853abd67b6b2a8c7293429.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 193, + 427, + 203 + ], + "lines": [ + { + "bbox": [ + 415, + 193, + 427, + 203 + ], + "spans": [ + { + "bbox": [ + 415, + 193, + 427, + 203 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 135, + 205, + 246, + 278 + ], + "blocks": [ + { + "bbox": [ + 135, + 205, + 246, + 278 + ], + "lines": [ + { + "bbox": [ + 135, + 205, + 246, + 278 + ], + "spans": [ + { + "bbox": [ + 135, + 205, + 246, + 278 + ], + "type": "image", + "image_path": "c8e92be45992602255faae1f8d413a6aa830184d4e3b849b1f2277cb47cc8100.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 280, + 196, + 291 + ], + "lines": [ + { + "bbox": [ + 183, + 280, + 196, + 291 + ], + "spans": [ + { + "bbox": [ + 183, + 280, + 196, + 291 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 251, + 205, + 361, + 278 + ], + "blocks": [ + { + "bbox": [ + 251, + 205, + 361, + 278 + ], + "lines": [ + { + "bbox": [ + 251, + 205, + 361, + 278 + ], + "spans": [ + { + "bbox": [ + 251, + 205, + 361, + 278 + ], + "type": "image", + "image_path": "1544a0211dc4dd190cfc86513394a251616c8df67a7cdd94ca58d32cf047e78b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 280, + 313, + 291 + ], + "lines": [ + { + "bbox": [ + 299, + 280, + 313, + 291 + ], + "spans": [ + { + "bbox": [ + 299, + 280, + 313, + 291 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 205, + 477, + 278 + ], + "blocks": [ + { + "bbox": [ + 365, + 205, + 477, + 278 + ], + "lines": [ + { + "bbox": [ + 365, + 205, + 477, + 278 + ], + "spans": [ + { + "bbox": [ + 365, + 205, + 477, + 278 + ], + "type": "image", + "image_path": "37ec93390bfdf6f5cdcfc231f4204692236f2c29be2a0812955b9cb8f5bf1bd5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 280, + 429, + 291 + ], + "lines": [ + { + "bbox": [ + 417, + 280, + 429, + 291 + ], + "spans": [ + { + "bbox": [ + 417, + 280, + 429, + 291 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 300, + 482, + 356 + ], + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 356 + ], + "type": "text", + "content": "Fig. 4: Qualitative Results of our method on Hi4D (a)(b), CHI3D (c)(d), MultiHuman Real-Cap (e), and Shelf (f). The left and middle columns in each sub-figure show the 2D projections of the estimated 3D skeletons and SMPL body meshes on two views. The right column in each sub-figure demonstrates skeletons and SMPL bodies in 3D scenes." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 318, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 318, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 318, + 392 + ], + "type": "text", + "content": "4.2 Additional Qualitative Samples" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 399, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 449 + ], + "type": "text", + "content": "Figure 4 demonstrates more qualitative results of our method on Hi4D, CHI3D, and MultiHuman Real-Cap with challenging and close interactions among 2 or 3 people. We also demonstrate results on the Shelf [3] dataset, which contains 4 people without close contact. More results can be found in the Supp Mat." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 466, + 238, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 238, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 238, + 478 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 486, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 521 + ], + "type": "text", + "content": "To validate the effectiveness of our method, we conduct a detailed analysis of different design choices of our algorithm. All the experiments are conducted on the Hi4D Dataset." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 533, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 482, + 631 + ], + "type": "text", + "content": "Comparison with SMPL Body Prior. To validate the effectiveness of our personalized avatar prior, we compare with a baseline that optimizes SMPL parameters to align 3D joint reprojections to 2D observations. Our method significantly outperforms the SMPL prior baseline (Table 3) by leveraging color and silhouette rendering loss to refine poses, reducing reliance on inaccurate joint detections when occlusion happens. Figure 5 shows that while the baseline method incorrectly estimates the arm pose and even leads to penetration, our method reconstructs the poses accurately." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Color and Mask Loss. To demonstrate the effectiveness of the color and silhouette rendering loss in our optimization process, we design baselines without RGB" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 219, + 115, + 392, + 193 + ], + "blocks": [ + { + "bbox": [ + 219, + 115, + 392, + 193 + ], + "lines": [ + { + "bbox": [ + 219, + 115, + 392, + 193 + ], + "spans": [ + { + "bbox": [ + 219, + 115, + 392, + 193 + ], + "type": "image", + "image_path": "d3dd3bb0fb33cc515cf527d829bd5e0599372cd5876b0d1b73478989efb6619e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 202, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 247 + ], + "type": "text", + "content": "Fig.5: Comparison with SMPL Body Prior. Only fitting SMPL to 2D observations, some joints in close contact such as arms are incorrectly estimated and even cause intersections between body surfaces. In contrast, our personalized prior enables accurate estimation of poses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 156, + 259, + 457, + 352 + ], + "blocks": [ + { + "bbox": [ + 156, + 259, + 457, + 352 + ], + "lines": [ + { + "bbox": [ + 156, + 259, + 457, + 352 + ], + "spans": [ + { + "bbox": [ + 156, + 259, + 457, + 352 + ], + "type": "table", + "html": "
MethodMPJPE(mm) ↓PCP(%) ↑AP50 ↑AP100 ↑
Ours (SMPL fitted)40.4195.1084.0494.37
Ours (w/o RGB loss)78.4083.2617.8474.55
Ours (w/o Mask loss)31.0097.3390.2899.06
Ours (Joint Optimization)66.0484.5622.5576.70
Ours29.3798.0296.7999.06
", + "image_path": "e122d0cdce8f8952bf9b8458a6098bdf0bda3db112652379436061494496a72b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 384 + ], + "type": "text", + "content": "Table 3: Quantitative Ablation Results. Ablations to evaluate our method with only the SMPL fitted method, our method without RGB loss and without Silhouette loss, and our method without alternating optimization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": "loss and mask loss respectively for comparison. Table 3 shows that the RGB loss significantly improves pose optimization, and the optimization process will completely deviate from the correct trajectory in the absence of RGB loss. Mask loss is proven to slightly increase the accuracy by adding additional constraints on the rendered human silhouette." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "content": "Alternating Optimization. We choose a joint optimization method widely used in avatar reconstructions [20, 26] to compare with our alternating optimization. Table 3 shows that our method significantly outperforms the chosen baseline. In Figure 6, the avatars from the baseline suffer from artifacts around contact body parts, leading to wrong 3D pose estimations. In contrast, our method can faithfully reconstruct avatars and poses with challenging initializations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Penetration Loss. Our penetration loss serves an important role in avoiding interpenetration. Comparing our method to the ablated version where we remove this loss, Figure 7 shows that the SMPL body of one person partially intersects with the other person in the contact area. This is also confirmed by the rendering result of the avatar. By penalizing the collision of density fields of avatars in 3D space, our method largely reduces the penetration and thus achieves more accurate pose estimation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "AvatarPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 238, + 114, + 280, + 180 + ], + "blocks": [ + { + "bbox": [ + 238, + 114, + 280, + 180 + ], + "lines": [ + { + "bbox": [ + 238, + 114, + 280, + 180 + ], + "spans": [ + { + "bbox": [ + 238, + 114, + 280, + 180 + ], + "type": "image", + "image_path": "d915227c893543113339fdb599925d70d6a87c5bcd1557216f2da5c3ecbd32da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 181, + 263, + 187 + ], + "lines": [ + { + "bbox": [ + 252, + 181, + 263, + 187 + ], + "spans": [ + { + "bbox": [ + 252, + 181, + 263, + 187 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 290, + 118, + 324, + 177 + ], + "blocks": [ + { + "bbox": [ + 290, + 118, + 324, + 177 + ], + "lines": [ + { + "bbox": [ + 290, + 118, + 324, + 177 + ], + "spans": [ + { + "bbox": [ + 290, + 118, + 324, + 177 + ], + "type": "image", + "image_path": "0df18bae9cc532ba8ad51ca35942a4aed5a706b0046f7284bdbbc8030b6f8175.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 180, + 321, + 187 + ], + "lines": [ + { + "bbox": [ + 296, + 180, + 321, + 187 + ], + "spans": [ + { + "bbox": [ + 296, + 180, + 321, + 187 + ], + "type": "text", + "content": "Joint Opt" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 336, + 118, + 371, + 177 + ], + "blocks": [ + { + "bbox": [ + 336, + 118, + 371, + 177 + ], + "lines": [ + { + "bbox": [ + 336, + 118, + 371, + 177 + ], + "spans": [ + { + "bbox": [ + 336, + 118, + 371, + 177 + ], + "type": "image", + "image_path": "8685c5c8784522317a84ddd4d1a089aab591005576eb640e17849a71fd03977d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 180, + 374, + 188 + ], + "lines": [ + { + "bbox": [ + 335, + 180, + 374, + 188 + ], + "spans": [ + { + "bbox": [ + 335, + 180, + 374, + 188 + ], + "type": "text", + "content": "Alternating Opt" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 253, + 256, + 331 + ], + "blocks": [ + { + "bbox": [ + 130, + 197, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 242 + ], + "type": "text", + "content": "Fig. 6: Ablation of Alternating Optimization. We show the results of rendered avatars and projections of the estimated 3D poses. Joint optimization suffers from artifacts around the contact part and in turn causes wrong pose estimations. In contrast, ours reconstructs both avatars and poses correctly." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 253, + 256, + 331 + ], + "lines": [ + { + "bbox": [ + 220, + 253, + 256, + 331 + ], + "spans": [ + { + "bbox": [ + 220, + 253, + 256, + 331 + ], + "type": "image", + "image_path": "bba9dfc4bb1ad3117e08a5e98679235867f5f088d6c17c473898d81ada6a75c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 258, + 262, + 287, + 330 + ], + "blocks": [ + { + "bbox": [ + 258, + 262, + 287, + 330 + ], + "lines": [ + { + "bbox": [ + 258, + 262, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 258, + 262, + 287, + 330 + ], + "type": "image", + "image_path": "ec8179686a182c6612bd46c4ca0af2eeff0589110c6949c0416b5904be9a2a1d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 333, + 328, + 342 + ], + "lines": [ + { + "bbox": [ + 253, + 333, + 328, + 342 + ], + "spans": [ + { + "bbox": [ + 253, + 333, + 328, + 342 + ], + "type": "text", + "content": "Ours(w/o penetration loss)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 293, + 262, + 322, + 329 + ], + "blocks": [ + { + "bbox": [ + 293, + 262, + 322, + 329 + ], + "lines": [ + { + "bbox": [ + 293, + 262, + 322, + 329 + ], + "spans": [ + { + "bbox": [ + 293, + 262, + 322, + 329 + ], + "type": "image", + "image_path": "d9326f997f7193d82750934f74f4a00b77c352b465a732e3698b4424a77ca6e7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 323, + 262, + 358, + 329 + ], + "blocks": [ + { + "bbox": [ + 323, + 262, + 358, + 329 + ], + "lines": [ + { + "bbox": [ + 323, + 262, + 358, + 329 + ], + "spans": [ + { + "bbox": [ + 323, + 262, + 358, + 329 + ], + "type": "image", + "image_path": "ce07464fb9f4705079d24143265206ab9f1d4637c7985beedcdd768668281084.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 333, + 367, + 341 + ], + "lines": [ + { + "bbox": [ + 352, + 333, + 367, + 341 + ], + "spans": [ + { + "bbox": [ + 352, + 333, + 367, + 341 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 359, + 262, + 388, + 331 + ], + "blocks": [ + { + "bbox": [ + 359, + 262, + 388, + 331 + ], + "lines": [ + { + "bbox": [ + 359, + 262, + 388, + 331 + ], + "spans": [ + { + "bbox": [ + 359, + 262, + 388, + 331 + ], + "type": "image", + "image_path": "b784b2eb3781b8a547ea8005a2c557ac3e23d6b906255c0d595f072e04ad57f3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 351, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 374 + ], + "type": "text", + "content": "Fig. 7: Ablation of Penetration Loss. Without penetration loss, both the avatar and underlying SMPL body tend to have collisions on surfaces." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 397, + 309, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 397, + 309, + 408 + ], + "spans": [ + { + "bbox": [ + 131, + 397, + 309, + 408 + ], + "type": "text", + "content": "4.4 Limitations and Future Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 418, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 491 + ], + "type": "text", + "content": "Despite greatly improving " + }, + { + "bbox": [ + 130, + 418, + 482, + 491 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 130, + 418, + 482, + 491 + ], + "type": "text", + "content": " of the sequences, our method shows minor improvements when the pose initialization is in the opposite direction or most joints are severely misaligned, leading to local minima in optimization. Moreover, we do not model hands in our avatar model, it will be a promising direction to integrate hand models [46,51] into the personalized avatar. More discussions about limitations and future work can be found in the Supp Mat." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 508, + 220, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 220, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 220, + 521 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "content": "In this paper, we propose AvatarPose, a novel method to estimate the 3D poses of multiple people in close interaction from sparse multi-view videos. Unlike previous methods leveraging SMPL body prior, we leverage the reconstructed avatars as personalized priors to guide pose optimization. The avatar prior enables us to use color and silhouette observations and to introduce a collision loss in close contact. Our method outperforms SotA methods significantly on public datasets of close human interactions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Acknowledgements This work was supported by the Swiss SERI Consolidation Grant \"AI-PERCEIVE\". We thank Wenbo Wang, Tianjian Jiang and Juan Zarate for providing us with helpful guidance and feedback." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: Scape: shape completion and animation of people. In: ACM SIGGRAPH 2005 Papers, pp. 408-416 (2005)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "text", + "content": "2. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures for multiple human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1669-1676 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 202, + 480, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 202, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 480, + 234 + ], + "type": "text", + "content": "3. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures for multiple human pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1669-1676 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "text", + "content": "4. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab, N., Ilic, S.: 3d pictorial structures revisited: Multiple human pose estimation. IEEE transactions on pattern analysis and machine intelligence 38(10), 1929-1942 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 308 + ], + "type": "text", + "content": "5. Benzine, A., Chabot, F., Luvison, B., Pham, Q.C., Achard, C.: Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6856-6865 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 352 + ], + "type": "text", + "content": "6. Bhatnagar, B.L., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Combining implicit function learning and parametric models for 3d human reconstruction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 311-329. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 407 + ], + "type": "text", + "content": "7. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 561-578. Springer (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "type": "text", + "content": "8. Burov, A., Nießner, M., Thies, J.: Dynamic surface function networks for clothed human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10754-10764 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 472 + ], + "type": "text", + "content": "9. Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7291-7299 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 472, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 492 + ], + "type": "text", + "content": "0. Chen, J., Zhang, Y., Kang, D., Zhe, X., Bao, L., Jia, X., Lu, H.: Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "type": "text", + "content": "1. Chen, L., Ai, H., Chen, R., Zhuang, Z., Liu, S.: Cross-view tracking for multi-human 3d pose estimation at over 100 fps. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3279-3288 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 568 + ], + "type": "text", + "content": "2. Chen, X., Zheng, Y., Black, M.J., Hilliges, O., Geiger, A.: Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11594-11604 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "text", + "content": "3. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7103-7112 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "text", + "content": "4. Chen, Y., Gu, R., Huang, O., Jia, G.: Vtp: volumetric transformer for multi-view multi-person 3d pose estimation. Applied Intelligence pp. 1-12 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "5. Cheng, B., Xiao, B., Wang, J., Shi, H., Huang, T.S., Zhang, L.: Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5386-5395 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "AvatarPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Cheng, Y., Li, L., Xu, Y., Li, X., Yang, Z., Wang, W., Yang, Y.: Segment and track anything. arXiv preprint arXiv:2305.06558 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "text", + "content": "17. Choudhury, R., Kitani, K.M., Jeni, L.A.: Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14750-14760 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "text", + "content": "18. Dong, J., Fang, Q., Jiang, W., Yang, Y., Bao, H., Zhou, X.: Fast and robust multiperson 3d pose estimation and tracking from multiple views. In: T-PAMI (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 191, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 191, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 191, + 480, + 224 + ], + "type": "text", + "content": "19. Dong, J., Jiang, W., Huang, Q., Bao, H., Zhou, X.: Fast and robust multi-person 3d pose estimation from multiple views. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7792-7801 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 267 + ], + "type": "text", + "content": "20. Dong, Z., Guo, C., Song, J., Chen, X., Geiger, A., Hilliges, O.: Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20470-20480 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "21. Dong, Z., Song, J., Chen, X., Guo, C., Hilliges, O.: Shape-aware multi-person pose estimation from multi-view images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11158-11168 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "text", + "content": "22. Ershadi-Nasab, S., Noury, E., Kasaei, S., Sanaei, E.: Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications 77, 15573-15601 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 374 + ], + "type": "text", + "content": "23. Fabbri, M., Lanzi, F., Calderara, S., Alletto, S., Cucchiara, R.: Compressed volumetric heatmaps for multi-person 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7204-7213 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "24. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2334-2343 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 449 + ], + "type": "text", + "content": "25. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7214-7223 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 449, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 493 + ], + "type": "text", + "content": "26. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12858-12868 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "text", + "content": "27. Guo, W., Bie, X., Alameda-Pineda, X., Moreno-Noguer, F.: Multi-person extreme motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13053-13064 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "28. Hasler, N., Stoll, C., Sunkel, M., Rosenhahn, B., Seidel, H.P.: A statistical model of human pose and body shape. In: Computer graphics forum. vol. 28, pp. 337-346. Wiley Online Library (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "text", + "content": "29. Jiang, T., Chen, X., Song, J., Hilliges, O.: Instantavatar: Learning avatars from monocular video in 60 seconds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16922-16932 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "text", + "content": "30. Joo, H., Liu, H., Tan, L., Gui, L., Nabbe, B., Matthews, I., Kanade, T., Nobuhara, S., Sheikh, Y.: Panoptic studio: A massively multiview system for social motion capture. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3334-3342 (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7122-7131 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Kocabas, M., Huang, C.H.P., Hilliges, O., Black, M.J.: Pare: Part attention regressor for 3d human body estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11127-11137 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "33. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: Spec: Seeing people in the wild with an estimated camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11035-11045 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "34. Li, Y., Zhang, S., Wang, Z., Yang, S., Yang, W., Xia, S.T., Zhou, E.: Tokenpose: Learning keypoint tokens for human pose estimation. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 11313-11322 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "35. Lin, J., Lee, G.H.: Multi-view multi-person 3d pose estimation with plane sweep stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11886-11895 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "text", + "content": "36. Liu, Y., Gall, J., Stoll, C., Dai, Q., Seidel, H.P., Theobalt, C.: Markerless motion capture of multiple characters using multiview image segmentation. IEEE transactions on pattern analysis and machine intelligence 35(11), 2720-2735 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 328 + ], + "type": "text", + "content": "37. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "text", + "content": "38. Luvizon, D.C., Habermann, M., Golyanik, V., Kortylewski, A., Theobalt, C.: Scene-aware 3d multi-human motion capture from a single camera. In: Computer Graphics Forum. vol. 42, pp. 371-383. Wiley Online Library (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "text", + "content": "39. Martinez, J., Hossain, R., Romero, J., Little, J.J.: A simple yet effective baseline for 3d human pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2640-2649 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "text", + "content": "40. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H.P., Xu, W., Casas, D., Theobalt, C.: Vnect: Real-time 3d human pose estimation with a single rgb camera. Acm transactions on graphics (tog) 36(4), 1-14 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "text", + "content": "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "text", + "content": "42. Muller, L., Osman, A.A.A., Tang, S., Huang, C.H.P., Black, M.J.: On self-contact and human pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9990-9999 (June 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "text", + "content": "43. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 532, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 575 + ], + "type": "text", + "content": "44. Osman, A.A., Bolkart, T., Black, M.J.: Star: Sparse trained articulated human body regressor. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16. pp. 598-613. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "text", + "content": "45. Patel, P., Huang, C.H.P., Tesch, J., Hoffmann, D.T., Tripathi, S., Black, M.J.: Agora: Avatars in geography optimized for regression analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13468-13478 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "46. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10975-10985 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "AvatarPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "47. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3d human pose and shape from a single color image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 459-468 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 195 + ], + "type": "text", + "content": "48. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "type": "text", + "content": "49. Pishchulin, L., Insafutdinov, E., Tang, S., Andres, B., Andriluka, M., Gehler, P.V., Schiele, B.: Deepcut: Joint subset partition and labeling for multi person pose estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4929-4937 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 284 + ], + "type": "text", + "content": "50. Reddy, N.D., Guigues, L., Pishchulin, L., Eledath, J., Narasimhan, S.G.: Tessen-track: End-to-end learnable multi-person articulated 3d pose tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15190-15200 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "text", + "content": "51. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "52. Saito, S., Yang, J., Ma, Q., Black, M.J.: Scanimate: Weakly supervised learning of skinned clothed avatar networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2886-2897 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 396 + ], + "type": "text", + "content": "53. Shuai, Q., Geng, C., Fang, Q., Peng, S., Shen, W., Zhou, X., Bao, H.: Novel view synthesis of human interactions from sparse multi-view videos. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22, Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10.1145/3528233.3530704, https://doi.org/10.1145/3528233.3530704" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "text", + "content": "54. Shuai, Q., Yu, Z., Zhou, Z., Fan, L., Yang, H., Yang, C., Zhou, X.: Reconstructing close human interactions from multiple views. ACM Transactions on Graphics (dec 2023). https://doi.org/10.1145/3618336" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "text", + "content": "55. Sun, Y., Bao, Q., Liu, W., Fu, Y., Black, M.J., Mei, T.: Monocular, one-stage, regression of multiple 3d people. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11179-11188 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 508 + ], + "type": "text", + "content": "56. Sun, Y., Liu, W., Bao, Q., Fu, Y., Mei, T., Black, M.J.: Putting people in their place: Monocular regression of 3d people in depth. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13243-13252 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "57. Te, G., Li, X., Li, X., Wang, J., Hu, W., Lu, Y.: Neural capture of animatable 3d human from monocular video. In: European Conference on Computer Vision. pp. 275-291. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "text", + "content": "58. Tu, H., Wang, C., Zeng, W.: Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 197-212. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "text", + "content": "59. Wang, C., Li, J., Liu, W., Qian, C., Lu, C.: Hmor: Hierarchical multi-person ordinal relations for monocular multi-person 3d pose estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 242-259. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "60. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 182 + ], + "type": "text", + "content": "61. Wang, T., Zhang, J., Cai, Y., Yan, S., Feng, J.: Direct multi-view multi-person 3d pose estimation. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 13153-13164. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/6da9003b743b65f4c0ccd295cc484e57-Paper.pdf" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "type": "text", + "content": "62. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 481, + 258 + ], + "type": "text", + "content": "63. Wu, S., Jin, S., Liu, W., Bai, L., Qian, C., Liu, D., Ouyang, W.: Graph-based 3d multi-person pose estimation using multi-view images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11148-11157 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "64. Wu, S., Jin, S., Liu, W., Bai, L., Qian, C., Liu, D., Ouyang, W.: Graph-based 3d multi-person pose estimation using multi-view images. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "text", + "content": "65. Xu, H., Bazavan, E.G., Zanfir, A., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Ghum & ghuml: Generative 3d human shape and articulated pose models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6184-6193 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "66. Ye, H., Zhu, W., Wang, C., Wu, R., Wang, Y.: Faster voxelpose: Real-time 3d human pose estimation by orthographic projection. In: European Conference on Computer Vision. pp. 142-159. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "67. Yin, Y., Guo, C., Kaufmann, M., Zarate, J.J., Song, J., Hilliges, O.: Hi4d: 4d instance segmentation of close human interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17016-17027 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "text", + "content": "68. Zanfir, A., Marinoiu, E., Sminchisescu, C.: Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2148-2157 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "69. Zhang, H., Tian, Y., Zhou, X., Ouyang, W., Liu, Y., Wang, L., Sun, Z.: Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11446-11456 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "70. Zhang, J., Liu, X., Ye, X., Zhao, F., Zhang, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG) 40(4), 1-18 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "71. Zhang, Y., Wang, C., Wang, X., Liu, W., Zeng, W.: Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 2613-2626 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "72. Zhang, Y., An, L., Yu, T., Li, X., Li, K., Liu, Y.: 4d association graph for realtime multi-person motion capture using multiple video cameras. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1324-1333 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "73. Zhang, Y., Li, Z., An, L., Li, M., Yu, T., Liu, Y.: Lightweight multi-person total motion capture using sparse multi-view cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5560-5569 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "74. Zheng, Y., Shao, R., Zhang, Y., Yu, T., Zheng, Z., Dai, Q., Liu, Y.: Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6239-6249 (2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "AvatarPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 183 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "75. Zhou, H., Hong, C., Han, Y., Huang, P., Zhuang, Y.: Mh pose: 3d human pose estimation based on high-quality heatmap. In: 2021 IEEE International Conference on Big Data (Big Data). pp. 3215-3222. IEEE (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "76. Zhou, Z., Shuai, Q., Wang, Y., Fang, Q., Ji, X., Li, F., Bao, H., Zhou, X.: Quickpose: Real-time multi-view multi-person pose estimation in crowded scenes. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "F. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_content_list.json b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f67e651b0c78e20eeb366dd49745107653c0e051 --- /dev/null +++ b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_content_list.json @@ -0,0 +1,1666 @@ +[ + { + "type": "text", + "text": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting", + "text_level": 1, + "bbox": [ + 264, + 140, + 743, + 188 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lingzhe Zhao $^{1*}$ , Peng Wang $^{1,2*}$ , and Peidong Liu $^{1\\dagger}$", + "bbox": [ + 297, + 210, + 702, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Westlake University $^{2}$ Zhejiang University $\\{$ zhaolingzhe, wangpeng, liupeidong\\}@westlake.edu.cn https://github.com/wu-cvgl/BAD-Gaussians/", + "bbox": [ + 310, + 238, + 691, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. While neural rendering has demonstrated impressive capabilities in 3D scene reconstruction and novel view synthesis, it heavily relies on high-quality sharp images and accurate camera poses. Numerous approaches have been proposed to train Neural Radiance Fields (NeRF) with motion-blurred images, commonly encountered in real-world scenarios such as low-light or long-exposure conditions. However, the implicit representation of NeRF struggles to accurately recover intricate details from severely motion-blurred images and cannot achieve real-time rendering. In contrast, recent advancements in 3D Gaussian Splatting achieve high-quality 3D scene reconstruction and real-time rendering by explicitly optimizing point clouds into 3D Gaussians. In this paper, we introduce a novel approach, named BAD-Gaussians (Bundle Adjusted Deblur Gaussian Splatting), which leverages explicit Gaussian representation and handles severe motion-blurred images with inaccurate camera poses to achieve high-quality scene reconstruction. Our method models the physical image formation process of motion-blurred images and jointly learns the parameters of Gaussians while recovering camera motion trajectories during exposure time. In our experiments, we demonstrate that BAD-Gaussians not only achieves superior rendering quality compared to previous state-of-the-art deblur neural rendering methods on both synthetic and real datasets but also enables real-time rendering capabilities.", + "bbox": [ + 261, + 320, + 743, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Gaussian Splatting $\\cdot$ Deblurring $\\cdot$ Bundle Adjustment $\\cdot$ Differentiable Rendering", + "bbox": [ + 261, + 638, + 740, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 693, + 375, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Acquiring accurate 3D scene representations from 2D images has long been a challenging problem in computer vision. Serving as a fundamental component in various applications such as virtual/augmented reality and robotics navigation, substantial efforts have been dedicated to addressing this challenge over the last few decades. Among those pioneering works, Neural Radiance Fields", + "bbox": [ + 212, + 726, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 217, + 810, + 364, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 222, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(NeRF) [30], parameterized by Multi-layer Perceptrons (MLP), stands out for its utilization of differentiable volume rendering technique [22, 28] and has garnered significant attention due to its capability to recover high-quality 3D scene representation from 2D images.", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Numerous works have focused on enhancing the performance of NeRF, particularly in terms of training [5,31,37] and rendering efficiency [12,51]. A recent advancement, 3D Gaussian Splatting (3D-GS) [14], extends the implicit neural rendering [30] to explicit point clouds. By projecting these optimized point clouds (Gaussians) onto the image plane, 3D-GS [14] achieves real-time rendering while enhancing the efficiency of NeRF in both training and rendering, and also improves rendering quality. However, both NeRF-based [1,30,31] methods and 3D-GS [14] heavily rely on well-captured sharp images and accurately precomputed camera poses, typically obtained from COLMAP [38]. Motion-blurred images, a common form of image degradation, often encountered in low-light or long-exposure conditions, can notably impair the performance of both NeRF and 3D-GS. The challenges posed by motion-blurred images to NeRF and 3D-GS can be attributed to three primary factors: (a) NeRF and 3D-GS rely on high-quality sharp images for supervision. However, motion-blurred images violate this assumption and exhibit notably inaccurate corresponding geometry between multi-view frames, thus presenting significant difficulties in achieving accurate 3D scene representation for both NeRF and 3D-GS; (b) Accurate camera poses are essential for training NeRF and 3D-GS. However, recovering accurate poses from multi-view motion-blurred images using COLMAP [38] is challenging. (c) 3D-GS necessitates sparse cloud points from COLMAP as the initialization of Gaussians. The mismatched features between multi-view blurred images and the inaccuracies in pose calibration further exacerbate the issue, leading to COLMAP producing fewer cloud points. This introduces an additional initialization issue for 3D-GS. Therefore, these factors result in a notable drop in performance for 3D-GS when dealing with motion-blurred images.", + "bbox": [ + 212, + 210, + 787, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit neural representations, i.e. NeRF [30], have been employed to reconstruct sharp 3D scenes from motion-blurred images [19,27,47]. For example, Deblur-NeRF [27] introduces a deformable sparse kernel that alters a canonical kernel at spatial locations to simulate the blurring process. DP-NeRF [19] integrates physical priors derived from the motion-blurred image acquisition process into Deblur-NeRF [27] to construct a clean NeRF representation. In contrast, BAD-NeRF [47] models the physical process of capturing motion-blurred images and jointly optimizes NeRF while recovering the camera trajectory within the exposure time. However, these implicit deblur rendering methods encounter significant challenges in achieving real-time rendering and producing high-quality outputs with intricate details. Additionally, the implicit representation introduces extra difficulties in optimizing the neural parameters and camera poses as mentioned in [11].", + "bbox": [ + 212, + 593, + 787, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to address these challenges, we propose Bundle Adjusted Deblur Gaussian Splatting, the first motion deblur framework based on 3D-GS, which we refer to as BAD-Gaussians. We incorporate the physical process of motion", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Lingzhe Zhao, Peng Wang, Peidong Liu", + "bbox": [ + 271, + 114, + 535, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "blur into the training of 3D-GS, employing a spline function to characterize the trajectory within the camera's exposure time. In the training of BAD-Gaussians, the camera trajectory within exposure time is optimized using gradients derived from the Gaussians of the scene, while jointly optimizing the Gaussians themselves. Specifically, the trajectory of each motion-blurred image is represented by the initial and final poses at the beginning and end of the exposure time, respectively. By assuming the exposure time is typically short, we can interpolate between the initial and final poses to obtain every camera pose along the trajectory. From this trajectory, we generate a sequence of virtual sharp images by projecting the scene's Gaussians onto the image plane. These virtual sharp images are then averaged to synthesize the blurred images, following the physical blur process. Finally, the Gaussians along the trajectory are optimized by minimizing the photometric error between the synthesized blurred images and the input blurred images through differentiable Gaussian rasterization.", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate BAD-Gaussians using both synthetic and real datasets. The experimental results demonstrate that BAD-Gaussians outperforms prior state-of-the-art implicit neural rendering methods by explicitly incorporating the image formation process of motion-blurred images into the training of 3D-GS, achieving better rendering performance in terms of real-time rendering speed and superior rendering quality. In summary, our contributions can be outlined as follows:", + "bbox": [ + 212, + 357, + 787, + 448 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a photometric bundle adjustment formulation specifically designed for motion-blurred images, achieving the first real-time rendering performance from motion-blurred images within the framework of 3D Gaussian Splatting;", + "- We demonstrate how this formulation enables the acquisition of high-quality 3D scene representation from a set of motion-blurred images;", + "- Our approach successfully deblurs severe motion-blurred images, synthesizes higher-quality novel view images, and achieves real-time rendering, surpassing previous state-of-the-art implicit deblurring rendering methods." + ], + "bbox": [ + 225, + 452, + 785, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 606, + 388, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 215, + 633, + 455, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF [30], employing implicit MLP, exhibits remarkable performance in synthesizing high-quality novel view images and accurately representing 3D scenes. Numerous extension works have been proposed to enhance NeRF's performance, including improvements in training [5,31,37,42] and rendering [12,18,36,49,51] efficiency, as well as anti-alias capabilities [1-3]. Additionally, various methods aim to bolster NeRF's robustness against imperfect inputs, such as inaccurate camera poses [13, 25, 35, 48], few-shot images [8, 15, 33], and low-quality images [19, 23, 24, 27, 29, 47]. In the following section, we will primarily focus on reviewing methods closely related to our work.", + "bbox": [ + 212, + 654, + 787, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fast Neural Rendering. Numerous approaches inspired by NeRF [30] have sought to enhance its rendering efficiency by employing advanced data structures to reconstruct radiance fields, thereby minimizing the computational cost", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "associated with implicit MLPs used in NeRF. These approaches are primarily categorized into grid-based [4, 5, 10, 37] and hash-based [31] methods. Despite these efforts, achieving real-time rendering for unbounded and complete scenes, as well as high-resolution images, remains challenging. In contrast to methods based on volume rendering and implicit representations, which may hinder fast rendering, recent advancements like 3D-GS [14] achieve real-time high-quality rendering through pure explicit point scene representation and the differentiable Gaussian rasterization. Nevertheless, these 3D scene representations heavily rely on accurately posed high-quality images.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF for Camera Optimization. The pioneering work, BARF [25], was the first to propose simultaneous optimization of camera parameters alongside NeRF. They employed a coarse-to-fine bundle adjustment strategy to enhance camera pose recovery. Concurrently, SC-NeRF [13] introduced a method to learn various camera models, encompassing both extrinsic and intrinsic parameters, along with scene representation. In the latest development, CamP [35] introduced a preconditioner to mitigate correlations between camera parameters, aiming for improved joint optimization of camera and NeRF parameters. In contrast to the aforementioned works [13,25,35], which focus solely on optimizing camera poses for sharp images, our approach goes further by recovering the trajectory of each blurred image within the exposure time.", + "bbox": [ + 212, + 286, + 787, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF for Deblurring. Several scene deblurring methods based on NeRF have been proposed, including Deblur-NeRF [27] and DP-NeRF [19], which reconstruct sharp scene representations from sets of motion-blurred images. However, these methods fix inaccurate camera poses recovered from blurred images during training, leading to degradation in reconstruction performance. BAD-NeRF [47] and ExBluRF [20] jointly learns camera motion trajectories within exposure time and radiance fields, following the physical blur process. Despite these advancements, existing deblurring neural rendering strategies struggle to achieve real-time rendering and reconstruct intricate scene details due to the implicit MLP structure, necessitating significant improvements in rendering efficiency and quality. Moreover, optimizing the 3D scenes jointly with camera poses faces additional challenges due to NeRF's implicit representations. To address these issues, we propose achieving deblurring capability within the framework of Gaussian Splatting [14].", + "bbox": [ + 212, + 455, + 787, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Image Deblurring", + "text_level": 1, + "bbox": [ + 214, + 691, + 410, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Two primary categories typically classify existing techniques for addressing the motion deblurring problem: the first involves formulating the issue as an optimization task, wherein gradient descent is employed during inference to jointly refine the blur kernel and the latent sharp image [6, 9, 16, 21, 34, 40, 50]. Another pipeline phrases the deblurring task as an end-to-end learning, particularly leveraging deep learning methodologies. With the support of substantial datasets and deep neural networks, superior results have been achieved for both single image deblurring [17, 32, 44] and video deblurring [41]. However, these 2D deblurring", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Lingzhe Zhao, Peng Wang, Peidong Liu", + "bbox": [ + 271, + 114, + 535, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/15396aca0f0f42d1d32043761dc5782fcbb89522fdb0fd0ff89603e9a3434fc8.jpg", + "image_caption": [ + "Fig. 1: The pipeline of BAD-Gaussians. Our approach utilizes Gaussian representations to depict sharp 3D scenes derived from a series of motion-blurred images, along with their inaccurate poses and sparse point clouds from COLMAP. Employing forward projection and differentiable Gaussian rasterization, we jointly optimize the Gaussians in the scene and the camera trajectory within exposure time, by backpropagating gradients from Gaussians to camera poses. Following the physical process of motion blur, we model motion-blurred images by averaging the virtual sharp images captured during the exposure time. These virtual camera poses are represented and interpolated using a continuous spline within the SE(3) space. The joint optimization of Gaussians and camera trajectories is achieved by minimizing the photometric loss between synthesized and actual blurry images." + ], + "image_footnote": [], + "bbox": [ + 215, + 146, + 784, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "methods cannot exploit the 3D scene geometry between multi-view images, thus failing to ensure the view consistency of the scene from different viewpoints. In contrast, our approach focuses on leveraging the geometry within multi-view blurry images to reconstruct a high-quality 3D scene.", + "bbox": [ + 212, + 502, + 787, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 582, + 330, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BAD-Gaussians aims to recover a sharp 3D scene representation by jointly learning the camera motion trajectories and Gaussians parameters, given a sequence of motion-blurred images, along with their inaccurate poses and sparse point clouds estimated from COLMAP [38], as shown in Fig. 1. This is achieved by minimizing the photometric error between the input blurred images and the synthesized blurred images generated based on the physical motion blur image formation model. We will deliver each content in the following sections.", + "bbox": [ + 212, + 609, + 787, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary: 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 214, + 734, + 555, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following 3D-GS [14], the scene is represented by a series of 3D Gaussians. Each Gaussian, denoted as $\\mathbf{G}$ , is parameterized by its mean position $\\pmb{\\mu} \\in \\mathbb{R}^3$ , 3D covariance $\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}$ , opacity $\\mathbf{o} \\in \\mathbb{R}$ and color $\\mathbf{c} \\in \\mathbb{R}^3$ . The distribution of each scaled Gaussian is defined as:", + "bbox": [ + 212, + 756, + 787, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} (\\mathbf {x}) = e ^ {- \\frac {1}{2} (\\mathbf {x} - \\boldsymbol {\\mu}) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {x} - \\boldsymbol {\\mu})}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 820, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure that the 3D covariance $\\pmb{\\Sigma}$ remains positive semi-definite, which is physically meaningful, and to reduce the optimization difficulty, 3D-GS represents $\\pmb{\\Sigma}$ using a scale $\\mathbf{S} \\in \\mathbb{R}^3$ and rotation matrix $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ stored by a quaternion $\\mathbf{q} \\in \\mathbb{R}^4$ :", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 205, + 784, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to enable differentiable Gaussian rasterization, 3D-GS projects 3D Gaussians to 2D from a given camera pose $\\mathbf{T}_c = \\{\\mathbf{R}_c\\in \\mathbb{R}^{3\\times 3},\\mathbf{t}_c\\in \\mathbb{R}^3\\}$ for rasterizing and rendering using the following equation, as described in [53]:", + "bbox": [ + 212, + 229, + 782, + 275 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {R} _ {c} \\boldsymbol {\\Sigma} \\mathbf {R} _ {c} ^ {T} \\mathbf {J} ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 285, + 784, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{\\Sigma}^{\\prime}\\in \\mathbb{R}^{2\\times 2}$ is the 2D covariance matrix, $\\mathbf{J}\\in \\mathbb{R}^{2\\times 3}$ is the Jacobian of the affine approximation of the projective transformation. Afterward, each pixel color is rendered by rasterizing these $N$ sorted 2D Gaussians based on their depths, following the formulation:", + "bbox": [ + 212, + 311, + 782, + 373 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} = \\sum_ {i} ^ {N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 386, + 784, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_i$ is the learnable color of each Gaussian, and $\\alpha_{i}$ is the alpha value computed by evaluating a 2D covariance $\\pmb{\\Sigma}^{\\prime}$ multiplied with the learned Gaussian opacity $\\mathbf{o}$ :", + "bbox": [ + 212, + 436, + 782, + 483 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\mathbf {o} _ {i} \\cdot \\exp (- \\sigma_ {i}), \\quad \\sigma_ {i} = \\frac {1}{2} \\Delta_ {i} ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime - 1} \\Delta_ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 494, + 784, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Delta \\in \\mathbb{R}^2$ is the offset between the pixel center and the 2D Gaussian center.", + "bbox": [ + 212, + 532, + 782, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The derivations presented above demonstrate that the rendered pixel color, denoted as $\\mathbf{C}$ in Eq. (4), is a function that is differentiable with respect to all of the learnable Gaussians $\\mathbf{G}$ , and the camera poses $\\mathbf{T}_c$ . This facilitates our bundle adjustment formulation, accommodating a set of motion-blurred images and inaccurate camera poses within the framework of 3D-GS.", + "bbox": [ + 212, + 547, + 782, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Physical Motion Blur Image Formation Model", + "text_level": 1, + "bbox": [ + 214, + 645, + 643, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The physical process of image formation in a digital camera encompasses the gathering of photons during the exposure period, followed by their conversion into measurable electric charges. Mathematically representing this phenomenon involves integrating across a sequence of simulated virtual latent sharp images, as follows:", + "bbox": [ + 212, + 669, + 782, + 742 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} (\\mathbf {u}) = \\phi \\int_ {0} ^ {\\tau} \\mathbf {C} _ {\\mathrm {t}} (\\mathbf {u}) \\mathrm {d t}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 742, + 784, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{B}(\\mathbf{u})\\in \\mathbb{R}^{\\mathrm{H}\\times \\mathrm{W}\\times 3}$ denotes the real captured motion-blurred image, $\\mathbf{u}\\in \\mathbb{R}^2$ represents the pixel location in the image with height H and width W, $\\phi$ serves as a normalization factor, $\\tau$ is the camera exposure time, $\\mathbf{C}_{\\mathrm{t}}(\\mathbf{u})\\in \\mathbb{R}^{\\mathrm{H}\\times \\mathrm{W}\\times 3}$ is the virtual latent sharp image captured at timestamp t $\\in [0,\\tau ]$ within the", + "bbox": [ + 212, + 777, + 782, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Lingzhe Zhao, Peng Wang, Peidong Liu", + "bbox": [ + 271, + 114, + 535, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "exposure time. The blurred image $\\mathbf{B}(\\mathbf{u})$ , resulting from camera motion during the exposure time, is calculated by averaging all the virtual images $\\mathbf{C}_{\\mathrm{t}}(\\mathbf{u})$ across different timestamps $t$ . This discrete approximated of this model is dependent on the number $n$ of discrete samples, as denoted:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} (\\mathbf {u}) \\approx \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\mathbf {C} _ {i} (\\mathbf {u}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 219, + 785, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The level of motion blur within an image is contingent upon the movement of the camera during the exposure time. For instance, a swiftly moving camera results in minimal relative motion, particularly with shorter exposure times, while a slowly moving camera yields motion-blurred images, especially in low-light scenarios with prolonged exposure times. Additionally, it can be deduced that $\\mathbf{B}(\\mathbf{u})$ demonstrates differentiability concerning each virtual sharp image $\\mathbf{C}_i(\\mathbf{u})$ .", + "bbox": [ + 212, + 268, + 784, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Camera Motion Trajectory Modeling in 3D-GS", + "text_level": 1, + "bbox": [ + 214, + 397, + 651, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on Eq. (7), a straightforward approach to address motion-blurred images involves determining each virtual sharp image $\\mathbf{C}_i$ , which serves as the dependent variable of the motion-blur image $\\mathbf{B}$ . Given that a sharp image $\\mathbf{C}_i$ can be rendered from a specified camera pose $\\mathbf{T}_i$ within the framework of 3D-GS (i.e. $\\mathbf{G}_{\\theta}$ ) [14], establishing a one-to-one correspondence between poses and virtual sharp images is feasible. Consequently, we formulate the corresponding poses of each latent sharp image within the exposure time $\\tau$ by employing a camera motion trajectory represented through linear interpolation between two camera poses, one at the beginning of the exposure $\\mathbf{T}_{\\mathrm{start}} \\in \\mathbf{SE}(3)$ and the other at the end $\\mathbf{T}_{\\mathrm{end}} \\in \\mathbf{SE}(3)$ . The virtual camera pose at time $t \\in [0, \\tau]$ can thus be expressed as follows:", + "bbox": [ + 212, + 422, + 782, + 589 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {t} = \\mathbf {T} _ {\\text {s t a r t}} \\cdot \\exp \\left(\\frac {t}{\\tau} \\cdot \\log \\left(\\mathbf {T} _ {\\text {s t a r t}} ^ {- 1} \\cdot \\mathbf {T} _ {\\text {e n d}}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 599, + 784, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tau$ represents the exposure time. $\\frac{t}{\\tau}$ can be further discretized and derived as $\\frac{i}{n-1}$ for the $i^{th}$ sampled virtual sharp image (i.e. $C_i$ ) in Eq. (7) (i.e. with pose denoted as $\\mathbf{T}_i$ ), when there are $n$ sampled images in total. It follows that $\\mathbf{T}_i$ is differentiable with respect to both $\\mathbf{T}_{\\mathrm{start}}$ and $\\mathbf{T}_{\\mathrm{end}}$ . We refer to prior works [26, 47] for a comprehensive understanding of the interpolation and derivations of the related Jacobian. [26]. The objective of BAD-Gaussians is to estimate both $\\mathbf{T}_{\\mathrm{start}}$ and $\\mathbf{T}_{\\mathrm{end}}$ for each frame, along with the learnable parameters of Gaussians $\\mathbf{G}_{\\theta}$ .", + "bbox": [ + 212, + 638, + 787, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Loss Function", + "text_level": 1, + "bbox": [ + 214, + 784, + 377, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From a collection of $K$ motion-blurred images, we can proceed to estimate both the learnable parameters $\\pmb{\\theta}$ (i.e. mean position $\\pmb{\\mu}$ , 3D covariance $\\pmb{\\Sigma}$ , opacity $\\mathbf{o}$ )", + "bbox": [ + 212, + 809, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0324598dcf4cd206fc1b176fd8eb40c856f723519e2463b7b401e4ac94a7a21b.jpg", + "table_caption": [ + "Table 1: Ablation studies on the number of virtual camera poses $n$ . The results indicate that performance reaches a saturation point with increasing number $n$ ." + ], + "table_footnote": [], + "table_body": "
nCozy2roomTanabata
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
331.63.9389.081025.43.8215.2287
432.99.9552.055827.08.8704.1764
533.68.9595.045828.39.9011.1400
633.86.9627.036829.57.9214.1101
834.48.9654.031531.15.9440.0696
1034.68.9521.025832.12.9481.0464
1534.78.9679.025733.58.9652.0198
2034.63.9646.030733.87.9668.0151
", + "bbox": [ + 359, + 184, + 638, + 285 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/db5831506a084222f3c2e47c9875b2ef5effb317bac08fb77e48aea855a920d7.jpg", + "table_caption": [ + "Table 2: Ablation studies on the effect of trajectory representations. We denote Deblur-NeRF-S and Deblur-NeRF-R as the synthetic and real data from Deblur-NeRF, respectively. The results demonstrate that cubic interpolation improves performance in scenes with complex camera trajectories (i.e. $MBA - VO$ and Deblur-NeRF-R)." + ], + "table_footnote": [], + "table_body": "
Deblur-NeRF-SMBA-VODeblur-NeRF-R
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Linear Interpolation33.92.9467.042230.96.8868.149124.77.7778.1291
Cubic B-spline33.24.9395.047331.06.8883.145426.11.8157.1067
", + "bbox": [ + 271, + 352, + 720, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and color $\\mathbf{c}$ ) of 3D-GS and the camera motion trajectories (i.e. $\\mathbf{T}_{\\mathrm{start}}$ and $\\mathbf{T}_{\\mathrm{end}}$ ) for each image. This estimation is accomplished by minimizing the following loss function, which includes an $\\mathcal{L}_1$ loss and a D-SSIM term between $\\mathbf{B}_k(\\mathbf{u})$ , the $k^{th}$ blurry image synthesized from 3D-GS using the aforementioned image formation model (i.e. Eq. (7)), and $\\mathbf{B}_k^{gt}(\\mathbf{u})$ , the corresponding real captured blurry image:", + "bbox": [ + 215, + 435, + 787, + 512 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 531, + 785, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To optimize the learnable Gaussians parameter $\\pmb{\\theta}$ , and camera pose $\\mathbf{T}$ (i.e. $\\mathbf{T}_{\\mathrm{start}}$ and $\\mathbf{T}_{\\mathrm{end}}$ in our model) for each image, it is necessary to derive the corresponding Jacobians to complete the gradient flow:", + "bbox": [ + 215, + 569, + 787, + 614 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {\\theta}} = \\sum_ {k = 0} ^ {K - 1} \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {B} _ {k}} \\cdot \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\frac {\\partial \\mathbf {B} _ {k}}{\\partial \\mathbf {C} _ {i}} \\frac {\\partial \\mathbf {C} _ {i}}{\\partial \\boldsymbol {\\theta}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 633, + 785, + 674 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {T}} = \\sum_ {k = 0} ^ {K - 1} \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {B} _ {k}} \\cdot \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\frac {\\partial \\mathbf {B} _ {k}}{\\partial \\mathbf {C} _ {i}} \\frac {\\partial \\mathbf {C} _ {i}}{\\partial \\boldsymbol {\\theta}} \\frac {\\partial \\boldsymbol {\\theta}}{\\partial \\mathbf {T}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 705, + 785, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{B}_k(\\mathbf{u})$ and $\\mathbf{C}_i(\\mathbf{u})$ are denoted as $\\mathbf{B}_k$ and $\\mathbf{C}_i$ for simplification, and we parameterize both $\\mathbf{T}_{\\mathrm{start}}$ and $\\mathbf{T}_{\\mathrm{end}}$ with their corresponding Lie algebras of $\\mathbf{SE}(3)$ , which can be represented by a 6D vector respectively. For further details on the Jacobian of Gaussians to camera pose, $\\frac{\\partial\\pmb{\\theta}}{\\partial\\mathbf{T}}$ , please refer to our supplemental materials.", + "bbox": [ + 215, + 763, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Lingzhe Zhao, Peng Wang, Peidong Liu", + "bbox": [ + 271, + 114, + 535, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Quantitative comparisons for deblurring on the synthetic dataset of Deblur-NeRF [27], referred to as DB-NeRF in the table due to space constraints. Notably, DB-NeRF* and DP-NeRF* are trained with ground-truth poses, while the others are trained with poses estimated by COLMAP [38]. The experiments highlight the superior performance of our method over previous approaches. Additionally, the results demonstrate the sensitivity of Deblur-NeRF and DP-NeRF to pose accuracy. Regarding rendering efficiency, our method achieves over 200 FPS, whereas Deblur-NeRF, DP-NeRF, BAD-NeRF and ExBluRF fall below 1 FPS. Our method takes about 30 minutes to train, while other methods take more than 10 hours. Each color shading indicates the best and second-best result, respectively.", + "bbox": [ + 212, + 143, + 787, + 285 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/91698fd506a4fffc9a26bed17c9c81933e32eac03770f763ab0185a4e59f6f1a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CozyroomFactoryPoolTanabataTrolley
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
NeRF [30]26.13.7886.248422.02.5581.455029.90.7901.259520.57.5653.458421.71.6413.3814
3D-GS [14]25.86.7908.227021.73.5623.450329.37.7816.256120.51.5773.427821.65.6587.3614
MPR [52]29.90.8862.091525.07.6994.240933.28.8938.129022.60.7203.250726.24.8356.1762
SRN [44]29.47.8759.095026.54.7604.240432.94.8847.104523.19.7274.243825.36.8119.1618
DB-NeRF [27]29.53.8786.087925.85.7651.234031.07.8477.139723.20.7243.247125.68.8111.1683
DB-NeRF*30.26.8933.079126.40.7991.219132.30.8755.134524.56.7749.216626.24.8254.1671
DP-NeRF [19]29.02.8761.077325.42.7559.213930.48.8402.112723.93.7540.203826.17.8269.1359
DP-NeRF*30.77.9020.058427.69.8328.184733.22.8922.095425.27.7973.177926.99.8413.1312
BAD-NeRF [47]32.11.9137.051432.18.9105.118932.22.8680.090825.80.8047.156329.68.8954.0752
ExBluRF [20]28.460.87070.111329.380.85110.164631.660.86120.171925.430.80060.22725.70.81120.234
Ours34.68.9521.025831.88.9270.095236.95.9434.022532.12.9481.046433.97.9628.0209
", + "bbox": [ + 220, + 295, + 782, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Quantitative comparisons for novel view synthesis on the synthetic dataset of Deblur-NeRF [27], referred to as DB-NeRF in the table due to space constraints. The results demonstrate Our methods outperform previous state-of-the-art approaches, delivering the best performance across the board.", + "bbox": [ + 212, + 440, + 787, + 494 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a883d2cea534a366e2d74fd0d24ac28c3cef4364ef4a07556c23aae407c0db65.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CozyroomFactoryPoolTanabataTrolley
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [30]25.45.7734.251422.51.5672.419929.23.7711.275620.74.5774.434121.09.6157.3924
3D-GS [14]25.24.7789.208421.34.5532.437128.61.7562.261020.37.5755.395820.86.6167.3701
MPR [52]+ [14]29.00.8685.089423.38.6848.237429.04.8170.176122.53.7349.216226.27.8352.1694
SRN [44]+ [14]28.64.8662.090325.45.7740.204529.97.8207.165023.09.7456.214124.89.8103.1666
DB-NeRF [27]29.09.8718.093725.19.7359.231530.97.8401.160023.71.7483.236924.46.7758.2215
DB-NeRF*29.88.8901.074726.06.8023.210630.94.8399.169424.82.7861.204525.78.8122.1797
DP-NeRF [19]29.56.8831.068327.91.8317.183330.98.8454.126524.73.7867.186525.62.8123.1606
DP-NeRF*30.16.8958.060828.43.8420.177131.78.8623.130625.37.8046.170126.44.8307.1424
BAD-NeRF [47]31.07.9027.055131.71.9038.120430.95.8389.110725.42.8011.156628.60.8806.0843
Ours32.35.9278.034028.18.9164.099933.30.8964.055131.00.9427.047831.44.9443.0345
", + "bbox": [ + 220, + 507, + 782, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 660, + 375, + 678 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 698, + 446, + 714 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We implemented our method within the Nerfstudio [43] framework based on their Splatifto implementation. Camera pose interpolation and optimization on the SE(3) manifold is implemented with pypose [45]. Both the optimization of Gaussians and camera poses are performed using the Adam optimizer. The number of virtual camera poses (i.e., $n$ in Equation (7)) is set to 10, considering the trade-off between performance and efficiency. We initialize camera poses and Gaussians using estimations obtained", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7011ab042154068401cb65683bf7c08cd83afed5810c9bd9f9daad89427271c5.jpg", + "table_caption": [ + "Table 5: Quantitative comparisons for deblurring on the synthetic dataset of MBA-VO [27]. The results demonstrate that our method achieves the best performance even when the camera undergoes acceleration" + ], + "table_footnote": [], + "table_body": "
ArchViz-lowArchViz-highAverage
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
NeRF [30]26.26.7887.377523.29.7098.491024.78.7493.4343
3D-GS [14]26.54.8113.338523.43.7243.461424.99.7678.4000
MPR [52]29.60.8757.210325.04.7711.357627.32.8234.2840
SRN [44]30.15.8814.170327.07.8190.279628.61.8502.2250
DB-NeRF [27]28.38.8484.179225.71.7762.322027.05.8123.2506
DB-NeRF*29.65.8744.176426.44.8010.317228.05.8377.2468
DP-NeRF [19]28.24.8490.161525.74.7839.291326.99.8165.2264
DP-NeRF*28.57.8481.151126.80.8081.273627.69.8281.2124
BAD-NeRF [47]30.51.8749.165427.54.8109.267929.03.8429.2167
ExBluRF [20]28.4.8372.251828.43.8518.172128.42.8445.2120
Ours32.28.9167.113429.64.8568.184730.96.8868.1491
", + "bbox": [ + 330, + 198, + 668, + 319 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "from COLMAP [38]. All experiments are conducted on an NVIDIA RTX 4090 GPU.", + "bbox": [ + 212, + 347, + 782, + 377 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmark Datasets. We evaluate the performance of our method using both synthetic and real datasets provided by Deblur-NeRF [27] and BAD-NeRF [47]. Deblur-NeRF [47] contains five scenes, where the blurred images are generated via Blender [7] by averaging sharp virtual images captured during the exposure time, under the assumption of consistent velocity camera motion. For our evaluation, we utilize the dataset from BAD-NeRF [47], which expands upon the number of virtual sharp images to create more realistic motion-blurred images while keeping other settings consistent with [27]. Additionally, Deblur-NeRF [27] captured a real motion-blurred dataset by intentionally shaking a handheld camera during exposure.", + "bbox": [ + 212, + 381, + 784, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To enhance the evaluation of our method on severely motion-blurred images, we incorporate a dataset tailored for motion blur-aware visual odometry benchmarking (i.e. MBA-VO [26]). Different from the assumption of constant velocity in the Deblur-NeRF dataset [27], the blur images synthesized from MBA-VO [26] are based on real camera motion trajectories obtained from the ETH3D dataset [39], which do not exhibit constant velocity and include accelerations, thus presenting notable challenges.", + "bbox": [ + 212, + 534, + 784, + 638 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines and Evaluation Metrics. We conduct a comparative deblurring analysis between our method and state-of-the-art learning-based single image deblurring algorithms including SRN [44] and MPR [52]. Additionally, we include evaluations against closely related approaches such as Deblur-NeRF [27], DP-NeRF [19], BAD-NeRF [47] and ExBluRF [20]. To evaluate deblurring performance, we render high-quality images corresponding to the midpoint (i.e. $\\mathbf{C}_{\\frac{\\tau}{2}}$ in Eq. (6)) of exposure time of each training image from the optimized Gaussians.", + "bbox": [ + 212, + 643, + 784, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To facilitate novel view synthesis using SRN [44] and MPR [52], we use the deblurred images obtained from pre-trained models as inputs for training 3D-GS, as these single deblurring methods are not optimized for synthesizing novel images. During the training stage, all poses are estimated using COLMAP [38]. Additionally, we train Deblur-NeRF [27] and DP-NeRF [19] with ground-truth camera poses from Blender to assess the impact of inaccurate pose estimation on", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Lingzhe Zhao, Peng Wang, Peidong Liu", + "bbox": [ + 271, + 114, + 535, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d53905ffe7f5b311e3e36b30c4669e13d288c016d4d132e72dd32a34da914a34.jpg", + "image_caption": [ + "Fig. 2: Qualitative novel view synthesis results of different methods with synthetic datasets. Despite being trained with ground truth poses $(^{*})$ , BAD-Gaussians outperforms Deblur-NeRF* and DP-NeRF* in recovering high-quality scenes from motion-blurred images with inaccurate camera poses, showcasing its superior performance." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 795, + 441 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "these two methods. The quality of the rendered sharp image is evaluated using PSNR, SSIM, and LPIPS metrics.", + "bbox": [ + 212, + 553, + 784, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For pose accuracy evaluation, we compare the Absolute Trajectory Error (ATE) metric, against classical structure-from-motion framework COLMAP [38] and BAD-NeRF [47].", + "bbox": [ + 212, + 584, + 785, + 631 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 656, + 388, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Virtual Camera Poses. We experimented to investigate the effect of the number of interpolated virtual camera poses within the exposure time, as described by $n$ in Eq. (7). Two of the five synthetic scenes provided by Deblur-NeRF [27] were selected, representing sequences with minor and severe motion blur (i.e. Cozy2room and Tanabata), respectively. We varied the number, $n$ , from 3 to 20 in our study, and the rendering metric results are presented in Table 1. Our experimental results demonstrate that increasing the number of interpolated virtual camera poses helps in addressing severe motion blur. However, marginal improvements are observed for minor blur instances. Based on our experiments, we choose $n = 10$ interpolated virtual images to strike a balance between ren", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "dering performance and training efficiency (larger $n$ means more computational resources).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Trajectory Representations. To assess the impact of different trajectory representations, we conduct two experiments: one involves optimizing $\\mathbf{T}_{\\mathrm{start}}$ and $\\mathbf{T}_{\\mathrm{end}}$ to depict a linear trajectory, while the other utilizes a higher-order spline (i.e., cubic B-spline) that jointly optimizes four control knots $\\mathbf{T}_1$ , $\\mathbf{T}_2$ , $\\mathbf{T}_3$ , and $\\mathbf{T}_4$ to capture more complex camera motions. We refer [46] for more details about cubic B-spline. The average quantitative results on the synthetic datasets of Deblur-NeRF-S [27] (Cozy2room, Factory, Pool, Tanabata and Trolley) and MBA-VO [26] (ArchViz-low and ArchViz-high), along with real datasets, Deblur-NeRF-S, which includes 10 real captured scenes from Deblur-NeRF, are presented in Table 2. In synthetic scenes, linear interpolation demonstrates comparable performance to cubic B-spline interpolation. However, in real captured scenes, cubic B-spline interpolation outperforms linear interpolation, particularly due to the longer exposure time during image capture. The effectiveness of cubic B-spline interpolation in real scenes can be attributed to its ability to better model the nuances of camera motion over longer time intervals. Conversely, linear interpolation is sufficient to accurately represent camera motion trajectories within shorter time intervals, as observed in synthetic scenes. Combining the training efficiency and rendering quality, we employ linear interpolation in synthetic datasets and cubic B-spline in real data.", + "bbox": [ + 212, + 180, + 787, + 467 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/58c8b77b6f1cc002d2c71b2c870d0c5a10b8b18d849cf61dded6598022c18314.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS↓Ttrain(h)
NeRF [30]22.69.6347.36872.5
3DGS [14]22.72.6390.28870.3
DB-NeRF [27]25.63.7675.1820>5
DP-NeRF [19]25.91.7734.1589>5
BAD-NeRF [47]24.33.7003.2196>5
ExBluRF [20]25.31.7293.20844.0
Ours-Cubic26.11.8157.10640.5
", + "bbox": [ + 375, + 491, + 627, + 593 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6: Quantitative comparisons for novel view synthesis on the real captured dataset of Deblur-NeRF [27]. Full results can be found in our supplementary material.", + "bbox": [ + 212, + 593, + 782, + 621 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Results", + "text_level": 1, + "bbox": [ + 214, + 679, + 323, + 693 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on Synthetic Data. We evaluate our approach against baseline methods using synthetic datasets obtained from both Deblur-NeRF [27] and MBAVO [26]. The quantitative evaluation results on deblurring and novel view synthesis with the Deblur-NeRF dataset are presented in Table 3 and Table 4 respectively. All metrics (i.e. PSNR, SSIM, LPIPS) demonstrate substantial improvements over prior state-of-the-art methods (i.e. 3.6 and $1.7\\mathrm{dB}$ higher than the second best method on average in Table 3 and Table 4, respectively).", + "bbox": [ + 212, + 703, + 782, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results demonstrate that both NeRF [30] and 3D-GS [14] are suffering from motion-blurred images, which motivates the necessity of our method.", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Lingzhe Zhao, Peng Wang, Peidong Liu", + "bbox": [ + 271, + 114, + 535, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/43b01f0e8fcc0343e2736b6dce4c20139c2f7d63ce030a18bd2fc21398670e3e.jpg", + "image_caption": [ + "Fig. 3: Qualitative novel view synthesis results of different methods with the real datasets. The experimental results demonstrate that our method achieves superior performance over prior methods on the real dataset as well. In contrast, BAD-NeRF yields poorer results when applied to real data and exhibits satisfactory performance only within synthetic datasets." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 428 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Single-stage methods such as MPR [52] and SRN [44] fall short of matching the performance of our approach due to their limited utilization of geometric information across multi-view images. Additionally, our method surpasses Deblur-NeRF [27] and DP-NeRF [19], partly because they fail to optimize the inaccurate camera poses estimated from COLMAP during training. We additionally trained Deblur-NeRF and DP-NeRF with ground truth poses (denoted as *), and the results demonstrate improved performance, reflecting the sensitivity to the pose accuracies of both Deblur-NeRF and DP-NeRF. A notable result from both Table 3 and Table 4 is that our method performs worse than BAD-NeRF on the Factory sequence. We find that it is caused by the inferior capability to represent the sky by Gaussian splatting compared to NeRF. Nevertheless, our method still performs better than all prior methods with a large margin on average.", + "bbox": [ + 212, + 547, + 787, + 731 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative results are presented in Fig. 2. It demonstrates that our method effectively attains high-quality scene representations with intricate details, being trained from a series of blurred images. In Fig. 2, it is evident that Deblur-NeRF [27] and DP-NeRF [19] encounter difficulties in modeling regions with significant depth variations, attributable to their method of synthesizing blurred images through convolution with a blur kernel applied to rendered images. The physically based method BAD-NeRF [47] exhibits superior performance overall,", + "bbox": [ + 212, + 734, + 787, + 843 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "yet it still presents some deficiencies, particularly noticeable around areas with significant color and depth variation. It demonstrates the effectiveness of the explicit Gaussian splatting representation over the implicit neural representation.", + "bbox": [ + 212, + 146, + 782, + 189 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We further assess the performance of our methods using a dataset characterized by severe motion blur and camera movements with varying velocities, sourced from MBA-VO [26]. The results presented in Table 5 demonstrate that our method achieves the best performance even when the camera undergoes acceleration.", + "bbox": [ + 212, + 191, + 782, + 263 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results on Real Data. The performance of novel view synthesis on real data sourced from Deblur-NeRF [27] are also evaluated. The quantitative results, as presented in Table 6, demonstrate the superior performance of our method compared to other approaches. Furthermore, the qualitative results illustrated in Fig. 3 vividly demonstrate our method's ability to deliver intricate details.", + "bbox": [ + 212, + 270, + 782, + 344 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Pose Estimation. Due to the unknown metric scale, we align the estimated trajectories against their corresponding ground truth for the computation of the absolute trajectory error metric. The experimental results presented in Table 7 demonstrate the effectiveness of our approach in recovering precise camera poses.", + "bbox": [ + 212, + 349, + 782, + 407 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/b650d06ad24b8a9be752fc4042659a1632922fca64ddb1163ffbe51effc15fdd.jpg", + "table_caption": [ + "Table 7: Pose estimation performance of BAD-Gaussians on various blur sequences. The results are in the absolute trajectory error metric in centimeters (ATE/cm). The COLMAP-sharp / COLMAP-blur represents the result of COLMAP with sharp/blurry images respectively." + ], + "table_footnote": [], + "table_body": "
Cozy2roomFactoryPoolTanabataTrolleyArchViz-lowArchViz-highAverage
COLMAP-sharp [38].133±.079.101±.052.172±.047.277±.103.173±.079.148±.078.115±.066.1599
COLMAP-blur [38].287±.1611.69±.1.24.633±.2361.59±.1.06.912±.559.536±.309.766±.391.9163
BAD-NeRF [47].316±.101.600±.208.095±.0461.34±.543.122±.0701.24±.6352.18±1.54.8419
Ours.165±.057.395±.426.264±.0981.20±.621.134±.070.520±.5911.41±1.11.5840
", + "bbox": [ + 259, + 506, + 736, + 559 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 612, + 357, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented the first pipeline to learn Gaussian splittings from a set of motion-blurred images with inaccurate camera poses. Our pipeline can jointly optimize the 3D scene representation and camera motion trajectories. Extensive experimental evaluations demonstrate that our method can deliver high-quality novel view images, and achieve real-time rendering compared to prior state-of-the-art works.", + "bbox": [ + 212, + 643, + 782, + 732 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 758, + 401, + 776 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation.", + "bbox": [ + 212, + 790, + 782, + 835 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Lingzhe Zhao, Peng Wang, Peidong Liu", + "bbox": [ + 271, + 114, + 535, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://jonbarron.info/mipnerf/2, 3", + "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://jonbarron.info/mipnerf360/3", + "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2023), https://jonbarron.info/zipnerf/ 3", + "4. Cao, A., Johnson, J.: HexPlane: A Fast Representation for Dynamic Scenes. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://caoang327.github.io/HexPlane/ 4", + "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensorRF: Tensorial Radiance Fields. In: European Conference on Computer Vision (ECCV) (2022), https://apchenstu.github.io/TensoRF/ 2, 3, 4", + "6. Cho, S., Lee, S.: Fast Motion Deblurring. ACM Transactions on Graphics (TOG) 28(5) (2009) 4", + "7. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org 10", + "8. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised NeRF: Fewer Views and Faster Training for Free. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://www.cs.cmu.edu/~dsnerf/3", + "9. Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T.: Removing camera shake from a single photograph. In: ACM Transactions on Graphics (TOG) (2006) 4", + "0. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://sarafridov.github.io/K-Planes/4", + "1. Fu, Y., Liu, S., Kulkarni, A., Kautz, J., Efros, A.A., Wang, X.: Colmap-Free 3D Gaussian Splatting. arXiv preprint arXiv:2312.07504 (2023), https://oasisyang.github.io/colmap-free-3dgs/2", + "2. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: FastNeRF: High-Fidelity Neural Rendering at 200FPS. In: International Conference on Computer Vision (ICCV) (2021), https://microsoft.github.io/FastNeRF/2, 3", + "3. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-Calibrating Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://postech-cvlab.github.io/SCNeRF/ 3, 4", + "4. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (TOG) 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ 2, 4, 5, 7, 9, 10, 12, 13", + "5. Kim, M., Seo, S., Han, B.: InfoNeRF: Ray Entropy Minimization for Few-Shot Neural Volume Rendering. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://cv.snu.ac.kr/research/InfoNeRF/3" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Krishnan, D., Fergus, R.: Fast image deconvolution using Hyper-Laplacian priors. In: Neural Information Processing Systems (NIPS) (2009) 4", + "17. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better. In: International Conference on Computer Vision (ICCV) (2019), https://github.com/VITA-Group/DeblurGANv2 4", + "18. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://github.com/facebookresearch/pytorch3d/ 3", + "19. Lee, D., Lee, M., Shin, C., Lee, S.: DP-NeRF: Deblurred Neural Radiance Field With Physical Scene Priors. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://dogyoonlee.github.io/dpnerf/2, 3, 4, 9, 10, 11, 12, 13", + "20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: ExBluRF: Efficient Radiance Fields for Extreme Motion Blurred Images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023) 4, 9, 10, 12", + "21. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithm. In: Computer Vision and Pattern Recognition (CVPR) (2009) 4", + "22. Levoy, M.: Efficient Ray Tracing of Volume Data. ACM Transactions on Graphics (TOG) 9(3), 245-261 (1990) 2", + "23. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields. In: International Conference on Learning Representations (ICLR) (2024), https://moyangli00.github.io/usb_nerf/ 3", + "24. Li, W., Wan, P., Wang, P., Li, J., Zhou, Y., Liu, P.: BeNeRF: Neural Radiance Fields from a Single Blurry Image and Event Stream. In: European Conference on Computer Vision (ECCV). Springer (2024), https://github.com/WU-CVGL/BeNeRF 3", + "25. Lin, Chen-Hsuan and Ma, Wei-Chiu and Torralba, Antonio and Lucey, Simon: BARF: Bundle-Adjusting Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://chenhsuanlin.bitbucket.io/bundle-adjusting-NeRF 3, 4", + "26. Liu, P., Zuo, X., Larsson, V., Pollefeys, M.: MBA-VO: Motion Blur Aware Visual Odometry. In: International Conference on Computer Vision (ICCV) (2021), https://github.com/ethliup/MBA-VO 7, 10, 12, 14", + "27. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural Radiance Fields from Blurry Images. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://limacv.github.io/deblurnerf/ 2, 3, 4, 9, 10, 11, 12, 13, 14", + "28. Max, N.: Optical Models for Direct Volume Rendering. ACM Transactions on Graphics (TOG) 1(2), 99-108 (1995) 2", + "29. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://bmild.github.io/rawnerf/3", + "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: European Conference on Computer Vision (ECCV) (2020), https://www.matthewtancik.com/nerf 2, 3, 9, 10, 12", + "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics (TOG) 41(4), 102:1-102:15 (Jul 2022), https://nvlabs.github.io/instant-ngp/2, 3, 4" + ], + "bbox": [ + 217, + 147, + 784, + 838 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Lingzhe Zhao, Peng Wang, Peidong Liu", + "bbox": [ + 271, + 114, + 535, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Nah, S., Kim, T.H., Lee, K.M.: Deep Multi-Scale Convolutional Neural Network for Dynamic Scene Deblurring. In: Computer Vision and Pattern Recognition (CVPR) (2017), https://github.com/SeungjunNah/DeepDeblur_release 4", + "33. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse Inputs. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://m-niemeyer.github.io/regnerf/3", + "34. Park, H., Lee, K.M.: Joint Estimation of Camera Pose, Depth, Deblurring, and Super-Resolution from a Blurred Image Sequence. In: International Conference on Computer Vision (ICCV). vol. 2017-Octob (2017). https://doi.org/10.1109/ICCV.2017.494 4", + "35. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: CamP: Camera Preconditioning for Neural Radiance Fields. ACM Transactions on Graphics (TOG) (2023), https://camp-nerf.github.io/3, 4", + "36. Piala, M., Clark, R.: TermiNeRF: Ray Termination Prediction for Efficient Neural Rendering. In: International Conference on 3D Vision (3DV) (2021), https://projects.mackopes.com/terminerf/3", + "37. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance Fields without Neural Networks. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://alexyu.net/plenoxels/ 2, 3, 4", + "38. Schonberger, J.L., Frahm, J.M.: Structure-from-motion Revisited. In: Computer Vision and Pattern Recognition (CVPR) (2016), https://github.com/colmap/ colmap 2, 5, 9, 10, 11, 14", + "39. Schops, T., Sattler, T., Pollefeys, M.: BAD SLAM: Bundle Adjusted Direct RGB-D SLAM. In: Computer Vision and Pattern Recognition (CVPR) (2019), https://github.com/ETH3D/badslam 10", + "40. Shan, Q., Jia, J., A., A.: High-quality Motion Deblurring from a Single Image. ACM Transactions on Graphics (TOG) 27(3) (2008) 4", + "41. Su, S., Delbracio, M., Wang, J.: Deep Video Deblurring for Hand-held Cameras. In: Computer Vision and Pattern Recognition (CVPR) (2017), https://github.com/shuochsu/DeepVideoDeblurring 4", + "42. Sun, C., Sun, M., Chen, H.: Direct Voxel Grid Optimization: Super-fast Convergence for Radiance Fields Reconstruction. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://sunset1995.github.io/dvgo/3", + "43. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A Modular Framework for Neural Radiance Field Development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023), https://github.com/nerfstudio-project/nerfstudio/9", + "44. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Computer Vision and Pattern Recognition (CVPR) (2018), https://github.com/jiangsutx/SRN-Deblur 4, 9, 10, 13", + "45. Wang, C., Gao, D., Xu, K., Geng, J., Hu, Y., Qiu, Y., Li, B., Yang, F., Moon, B., Pandey, A., Aryan, Xu, J., Wu, T., He, H., Huang, D., Ren, Z., Zhao, S., Fu, T., Reddy, P., Lin, X., Wang, W., Shi, J., Talak, R., Cao, K., Du, Y., Wang, H., Yu, H., Wang, S., Chen, S., Kashyap, A., Bandaru, R., Dantu, K., Wu, J., Xie, L., Carlone, L., Hutter, M., Scherer, S.: PyPose: A library for robot learning with physics-based optimization. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023), https://github.com/pypyse/pypyse 9" + ], + "bbox": [ + 212, + 146, + 787, + 842 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Wang, P., Zhao, L., Ma, R., Liu, P.: Supplementary Material for BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields https://openaccess.thecvf.com/content/CVPR2023/supplemental/Wang_BAD-NeRF_Bundle_Adjusted_CVPR_2023SUPPLEMENTAL.pdf 12", + "47. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://wangpeng000.github.io/BAD-NeRF/2, 3, 4, 7, 9, 10, 11, 12, 13, 14", + "48. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021), https://nerfmm.active.vision/3", + "49. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: NeX: Real-time View Synthesis with Neural Basis Expansion. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://nex-mpi.github.io/3", + "50. Xu, L., Jia, J.: Two-Phase Kernel Estimation for robust motion deblurring. In: European Conference on Computer Vision (ECCV) (2010) 4", + "51. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for Real-time Rendering of Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://alexyu.net/plenoctrees/ 2, 3", + "52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage Progressive Image Restoration. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://github.com/swz30/MPRNet 9, 10, 13", + "53. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: EWA Volume Splatting. In: Proceedings Visualization, 2001. VIS'01. pp. 29-538. IEEE (2001), https://cgl.ethz.ch/research/past_projects/surfels/ewavolumesplatting/index.html 6" + ], + "bbox": [ + 215, + 146, + 785, + 481 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Lingzhe Zhao, Peng Wang, Peidong Liu", + "bbox": [ + 271, + 114, + 537, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_model.json b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d85c6160ae3a86d54cfe01273f2e362c0cd9408 --- /dev/null +++ b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_model.json @@ -0,0 +1,2271 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.265, + 0.141, + 0.744, + 0.189 + ], + "angle": 0, + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.212, + 0.704, + 0.229 + ], + "angle": 0, + "content": "Lingzhe Zhao\\(^{1*}\\), Peng Wang\\(^{1,2*}\\), and Peidong Liu\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.239, + 0.692, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\)Westlake University \\(^{2}\\)Zhejiang University \\(\\{\\)zhaolingzhe, wangpeng, liupeidong\\}@westlake.edu.cn https://github.com/wu-cvgl/BAD-Gaussians/" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.321, + 0.744, + 0.626 + ], + "angle": 0, + "content": "Abstract. While neural rendering has demonstrated impressive capabilities in 3D scene reconstruction and novel view synthesis, it heavily relies on high-quality sharp images and accurate camera poses. Numerous approaches have been proposed to train Neural Radiance Fields (NeRF) with motion-blurred images, commonly encountered in real-world scenarios such as low-light or long-exposure conditions. However, the implicit representation of NeRF struggles to accurately recover intricate details from severely motion-blurred images and cannot achieve real-time rendering. In contrast, recent advancements in 3D Gaussian Splatting achieve high-quality 3D scene reconstruction and real-time rendering by explicitly optimizing point clouds into 3D Gaussians. In this paper, we introduce a novel approach, named BAD-Gaussians (Bundle Adjusted Deblur Gaussian Splatting), which leverages explicit Gaussian representation and handles severe motion-blurred images with inaccurate camera poses to achieve high-quality scene reconstruction. Our method models the physical image formation process of motion-blurred images and jointly learns the parameters of Gaussians while recovering camera motion trajectories during exposure time. In our experiments, we demonstrate that BAD-Gaussians not only achieves superior rendering quality compared to previous state-of-the-art deblur neural rendering methods on both synthetic and real datasets but also enables real-time rendering capabilities." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.639, + 0.741, + 0.668 + ], + "angle": 0, + "content": "Keywords: 3D Gaussian Splatting \\(\\cdot\\) Deblurring \\(\\cdot\\) Bundle Adjustment \\(\\cdot\\) Differentiable Rendering" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.694, + 0.377, + 0.71 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Acquiring accurate 3D scene representations from 2D images has long been a challenging problem in computer vision. Serving as a fundamental component in various applications such as virtual/augmented reality and robotics navigation, substantial efforts have been dedicated to addressing this challenge over the last few decades. Among those pioneering works, Neural Radiance Fields" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.365, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.383, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.537, + 0.129 + ], + "angle": 0, + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "(NeRF) [30], parameterized by Multi-layer Perceptrons (MLP), stands out for its utilization of differentiable volume rendering technique [22, 28] and has garnered significant attention due to its capability to recover high-quality 3D scene representation from 2D images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.212, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Numerous works have focused on enhancing the performance of NeRF, particularly in terms of training [5,31,37] and rendering efficiency [12,51]. A recent advancement, 3D Gaussian Splatting (3D-GS) [14], extends the implicit neural rendering [30] to explicit point clouds. By projecting these optimized point clouds (Gaussians) onto the image plane, 3D-GS [14] achieves real-time rendering while enhancing the efficiency of NeRF in both training and rendering, and also improves rendering quality. However, both NeRF-based [1,30,31] methods and 3D-GS [14] heavily rely on well-captured sharp images and accurately precomputed camera poses, typically obtained from COLMAP [38]. Motion-blurred images, a common form of image degradation, often encountered in low-light or long-exposure conditions, can notably impair the performance of both NeRF and 3D-GS. The challenges posed by motion-blurred images to NeRF and 3D-GS can be attributed to three primary factors: (a) NeRF and 3D-GS rely on high-quality sharp images for supervision. However, motion-blurred images violate this assumption and exhibit notably inaccurate corresponding geometry between multi-view frames, thus presenting significant difficulties in achieving accurate 3D scene representation for both NeRF and 3D-GS; (b) Accurate camera poses are essential for training NeRF and 3D-GS. However, recovering accurate poses from multi-view motion-blurred images using COLMAP [38] is challenging. (c) 3D-GS necessitates sparse cloud points from COLMAP as the initialization of Gaussians. The mismatched features between multi-view blurred images and the inaccuracies in pose calibration further exacerbate the issue, leading to COLMAP producing fewer cloud points. This introduces an additional initialization issue for 3D-GS. Therefore, these factors result in a notable drop in performance for 3D-GS when dealing with motion-blurred images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.79 + ], + "angle": 0, + "content": "Implicit neural representations, i.e. NeRF [30], have been employed to reconstruct sharp 3D scenes from motion-blurred images [19,27,47]. For example, Deblur-NeRF [27] introduces a deformable sparse kernel that alters a canonical kernel at spatial locations to simulate the blurring process. DP-NeRF [19] integrates physical priors derived from the motion-blurred image acquisition process into Deblur-NeRF [27] to construct a clean NeRF representation. In contrast, BAD-NeRF [47] models the physical process of capturing motion-blurred images and jointly optimizes NeRF while recovering the camera trajectory within the exposure time. However, these implicit deblur rendering methods encounter significant challenges in achieving real-time rendering and producing high-quality outputs with intricate details. Additionally, the implicit representation introduces extra difficulties in optimizing the neural parameters and camera poses as mentioned in [11]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In order to address these challenges, we propose Bundle Adjusted Deblur Gaussian Splatting, the first motion deblur framework based on 3D-GS, which we refer to as BAD-Gaussians. We incorporate the physical process of motion" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "blur into the training of 3D-GS, employing a spline function to characterize the trajectory within the camera's exposure time. In the training of BAD-Gaussians, the camera trajectory within exposure time is optimized using gradients derived from the Gaussians of the scene, while jointly optimizing the Gaussians themselves. Specifically, the trajectory of each motion-blurred image is represented by the initial and final poses at the beginning and end of the exposure time, respectively. By assuming the exposure time is typically short, we can interpolate between the initial and final poses to obtain every camera pose along the trajectory. From this trajectory, we generate a sequence of virtual sharp images by projecting the scene's Gaussians onto the image plane. These virtual sharp images are then averaged to synthesize the blurred images, following the physical blur process. Finally, the Gaussians along the trajectory are optimized by minimizing the photometric error between the synthesized blurred images and the input blurred images through differentiable Gaussian rasterization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.449 + ], + "angle": 0, + "content": "We evaluate BAD-Gaussians using both synthetic and real datasets. The experimental results demonstrate that BAD-Gaussians outperforms prior state-of-the-art implicit neural rendering methods by explicitly incorporating the image formation process of motion-blurred images into the training of 3D-GS, achieving better rendering performance in terms of real-time rendering speed and superior rendering quality. In summary, our contributions can be outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.453, + 0.787, + 0.514 + ], + "angle": 0, + "content": "- We introduce a photometric bundle adjustment formulation specifically designed for motion-blurred images, achieving the first real-time rendering performance from motion-blurred images within the framework of 3D Gaussian Splatting;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.514, + 0.787, + 0.543 + ], + "angle": 0, + "content": "- We demonstrate how this formulation enables the acquisition of high-quality 3D scene representation from a set of motion-blurred images;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.543, + 0.787, + 0.588 + ], + "angle": 0, + "content": "- Our approach successfully deblurs severe motion-blurred images, synthesizes higher-quality novel view images, and achieves real-time rendering, surpassing previous state-of-the-art implicit deblurring rendering methods." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.453, + 0.787, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.389, + 0.623 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.457, + 0.649 + ], + "angle": 0, + "content": "2.1 Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.791 + ], + "angle": 0, + "content": "NeRF [30], employing implicit MLP, exhibits remarkable performance in synthesizing high-quality novel view images and accurately representing 3D scenes. Numerous extension works have been proposed to enhance NeRF's performance, including improvements in training [5,31,37,42] and rendering [12,18,36,49,51] efficiency, as well as anti-alias capabilities [1-3]. Additionally, various methods aim to bolster NeRF's robustness against imperfect inputs, such as inaccurate camera poses [13, 25, 35, 48], few-shot images [8, 15, 33], and low-quality images [19, 23, 24, 27, 29, 47]. In the following section, we will primarily focus on reviewing methods closely related to our work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Fast Neural Rendering. Numerous approaches inspired by NeRF [30] have sought to enhance its rendering efficiency by employing advanced data structures to reconstruct radiance fields, thereby minimizing the computational cost" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.537, + 0.129 + ], + "angle": 0, + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "associated with implicit MLPs used in NeRF. These approaches are primarily categorized into grid-based [4, 5, 10, 37] and hash-based [31] methods. Despite these efforts, achieving real-time rendering for unbounded and complete scenes, as well as high-resolution images, remains challenging. In contrast to methods based on volume rendering and implicit representations, which may hinder fast rendering, recent advancements like 3D-GS [14] achieve real-time high-quality rendering through pure explicit point scene representation and the differentiable Gaussian rasterization. Nevertheless, these 3D scene representations heavily rely on accurately posed high-quality images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.452 + ], + "angle": 0, + "content": "NeRF for Camera Optimization. The pioneering work, BARF [25], was the first to propose simultaneous optimization of camera parameters alongside NeRF. They employed a coarse-to-fine bundle adjustment strategy to enhance camera pose recovery. Concurrently, SC-NeRF [13] introduced a method to learn various camera models, encompassing both extrinsic and intrinsic parameters, along with scene representation. In the latest development, CamP [35] introduced a preconditioner to mitigate correlations between camera parameters, aiming for improved joint optimization of camera and NeRF parameters. In contrast to the aforementioned works [13,25,35], which focus solely on optimizing camera poses for sharp images, our approach goes further by recovering the trajectory of each blurred image within the exposure time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.671 + ], + "angle": 0, + "content": "NeRF for Deblurring. Several scene deblurring methods based on NeRF have been proposed, including Deblur-NeRF [27] and DP-NeRF [19], which reconstruct sharp scene representations from sets of motion-blurred images. However, these methods fix inaccurate camera poses recovered from blurred images during training, leading to degradation in reconstruction performance. BAD-NeRF [47] and ExBluRF [20] jointly learns camera motion trajectories within exposure time and radiance fields, following the physical blur process. Despite these advancements, existing deblurring neural rendering strategies struggle to achieve real-time rendering and reconstruct intricate scene details due to the implicit MLP structure, necessitating significant improvements in rendering efficiency and quality. Moreover, optimizing the 3D scenes jointly with camera poses faces additional challenges due to NeRF's implicit representations. To address these issues, we propose achieving deblurring capability within the framework of Gaussian Splatting [14]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.411, + 0.709 + ], + "angle": 0, + "content": "2.2 Image Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Two primary categories typically classify existing techniques for addressing the motion deblurring problem: the first involves formulating the issue as an optimization task, wherein gradient descent is employed during inference to jointly refine the blur kernel and the latent sharp image [6, 9, 16, 21, 34, 40, 50]. Another pipeline phrases the deblurring task as an end-to-end learning, particularly leveraging deep learning methodologies. With the support of substantial datasets and deep neural networks, superior results have been achieved for both single image deblurring [17, 32, 44] and video deblurring [41]. However, these 2D deblurring" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Fig. 1: The pipeline of BAD-Gaussians. Our approach utilizes Gaussian representations to depict sharp 3D scenes derived from a series of motion-blurred images, along with their inaccurate poses and sparse point clouds from COLMAP. Employing forward projection and differentiable Gaussian rasterization, we jointly optimize the Gaussians in the scene and the camera trajectory within exposure time, by backpropagating gradients from Gaussians to camera poses. Following the physical process of motion blur, we model motion-blurred images by averaging the virtual sharp images captured during the exposure time. These virtual camera poses are represented and interpolated using a continuous spline within the SE(3) space. The joint optimization of Gaussians and camera trajectories is achieved by minimizing the photometric loss between synthesized and actual blurry images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.565 + ], + "angle": 0, + "content": "methods cannot exploit the 3D scene geometry between multi-view images, thus failing to ensure the view consistency of the scene from different viewpoints. In contrast, our approach focuses on leveraging the geometry within multi-view blurry images to reconstruct a high-quality 3D scene." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.583, + 0.331, + 0.599 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.718 + ], + "angle": 0, + "content": "BAD-Gaussians aims to recover a sharp 3D scene representation by jointly learning the camera motion trajectories and Gaussians parameters, given a sequence of motion-blurred images, along with their inaccurate poses and sparse point clouds estimated from COLMAP [38], as shown in Fig. 1. This is achieved by minimizing the photometric error between the input blurred images and the synthesized blurred images generated based on the physical motion blur image formation model. We will deliver each content in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.736, + 0.557, + 0.751 + ], + "angle": 0, + "content": "3.1 Preliminary: 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Following 3D-GS [14], the scene is represented by a series of 3D Gaussians. Each Gaussian, denoted as \\(\\mathbf{G}\\), is parameterized by its mean position \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\), 3D covariance \\(\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}\\), opacity \\(\\mathbf{o} \\in \\mathbb{R}\\) and color \\(\\mathbf{c} \\in \\mathbb{R}^3\\). The distribution of each scaled Gaussian is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.821, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} (\\mathbf {x}) = e ^ {- \\frac {1}{2} (\\mathbf {x} - \\boldsymbol {\\mu}) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {x} - \\boldsymbol {\\mu})}. \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.536, + 0.129 + ], + "angle": 0, + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": "To ensure that the 3D covariance \\(\\pmb{\\Sigma}\\) remains positive semi-definite, which is physically meaningful, and to reduce the optimization difficulty, 3D-GS represents \\(\\pmb{\\Sigma}\\) using a scale \\(\\mathbf{S} \\in \\mathbb{R}^3\\) and rotation matrix \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\) stored by a quaternion \\(\\mathbf{q} \\in \\mathbb{R}^4\\):" + }, + { + "type": "equation", + "bbox": [ + 0.448, + 0.207, + 0.785, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.23, + 0.784, + 0.276 + ], + "angle": 0, + "content": "In order to enable differentiable Gaussian rasterization, 3D-GS projects 3D Gaussians to 2D from a given camera pose \\(\\mathbf{T}_c = \\{\\mathbf{R}_c\\in \\mathbb{R}^{3\\times 3},\\mathbf{t}_c\\in \\mathbb{R}^3\\}\\) for rasterizing and rendering using the following equation, as described in [53]:" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.286, + 0.785, + 0.303 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {R} _ {c} \\boldsymbol {\\Sigma} \\mathbf {R} _ {c} ^ {T} \\mathbf {J} ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.784, + 0.374 + ], + "angle": 0, + "content": "where \\(\\pmb{\\Sigma}^{\\prime}\\in \\mathbb{R}^{2\\times 2}\\) is the 2D covariance matrix, \\(\\mathbf{J}\\in \\mathbb{R}^{2\\times 3}\\) is the Jacobian of the affine approximation of the projective transformation. Afterward, each pixel color is rendered by rasterizing these \\(N\\) sorted 2D Gaussians based on their depths, following the formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.387, + 0.785, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} = \\sum_ {i} ^ {N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.784, + 0.484 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_i\\) is the learnable color of each Gaussian, and \\(\\alpha_{i}\\) is the alpha value computed by evaluating a 2D covariance \\(\\pmb{\\Sigma}^{\\prime}\\) multiplied with the learned Gaussian opacity \\(\\mathbf{o}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\mathbf {o} _ {i} \\cdot \\exp (- \\sigma_ {i}), \\quad \\sigma_ {i} = \\frac {1}{2} \\Delta_ {i} ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime - 1} \\Delta_ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.784, + 0.547 + ], + "angle": 0, + "content": "where \\(\\Delta \\in \\mathbb{R}^2\\) is the offset between the pixel center and the 2D Gaussian center." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.784, + 0.623 + ], + "angle": 0, + "content": "The derivations presented above demonstrate that the rendered pixel color, denoted as \\(\\mathbf{C}\\) in Eq. (4), is a function that is differentiable with respect to all of the learnable Gaussians \\(\\mathbf{G}\\), and the camera poses \\(\\mathbf{T}_c\\). This facilitates our bundle adjustment formulation, accommodating a set of motion-blurred images and inaccurate camera poses within the framework of 3D-GS." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.646, + 0.644, + 0.661 + ], + "angle": 0, + "content": "3.2 Physical Motion Blur Image Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.784, + 0.743 + ], + "angle": 0, + "content": "The physical process of image formation in a digital camera encompasses the gathering of photons during the exposure period, followed by their conversion into measurable electric charges. Mathematically representing this phenomenon involves integrating across a sequence of simulated virtual latent sharp images, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.743, + 0.785, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} (\\mathbf {u}) = \\phi \\int_ {0} ^ {\\tau} \\mathbf {C} _ {\\mathrm {t}} (\\mathbf {u}) \\mathrm {d t}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.784, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}(\\mathbf{u})\\in \\mathbb{R}^{\\mathrm{H}\\times \\mathrm{W}\\times 3}\\) denotes the real captured motion-blurred image, \\(\\mathbf{u}\\in \\mathbb{R}^2\\) represents the pixel location in the image with height H and width W, \\(\\phi\\) serves as a normalization factor, \\(\\tau\\) is the camera exposure time, \\(\\mathbf{C}_{\\mathrm{t}}(\\mathbf{u})\\in \\mathbb{R}^{\\mathrm{H}\\times \\mathrm{W}\\times 3}\\) is the virtual latent sharp image captured at timestamp t \\(\\in [0,\\tau ]\\) within the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "exposure time. The blurred image \\(\\mathbf{B}(\\mathbf{u})\\), resulting from camera motion during the exposure time, is calculated by averaging all the virtual images \\(\\mathbf{C}_{\\mathrm{t}}(\\mathbf{u})\\) across different timestamps \\(t\\). This discrete approximated of this model is dependent on the number \\(n\\) of discrete samples, as denoted:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.22, + 0.786, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} (\\mathbf {u}) \\approx \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\mathbf {C} _ {i} (\\mathbf {u}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.785, + 0.377 + ], + "angle": 0, + "content": "The level of motion blur within an image is contingent upon the movement of the camera during the exposure time. For instance, a swiftly moving camera results in minimal relative motion, particularly with shorter exposure times, while a slowly moving camera yields motion-blurred images, especially in low-light scenarios with prolonged exposure times. Additionally, it can be deduced that \\(\\mathbf{B}(\\mathbf{u})\\) demonstrates differentiability concerning each virtual sharp image \\(\\mathbf{C}_i(\\mathbf{u})\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.398, + 0.652, + 0.414 + ], + "angle": 0, + "content": "3.3 Camera Motion Trajectory Modeling in 3D-GS" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.784, + 0.59 + ], + "angle": 0, + "content": "Based on Eq. (7), a straightforward approach to address motion-blurred images involves determining each virtual sharp image \\(\\mathbf{C}_i\\), which serves as the dependent variable of the motion-blur image \\(\\mathbf{B}\\). Given that a sharp image \\(\\mathbf{C}_i\\) can be rendered from a specified camera pose \\(\\mathbf{T}_i\\) within the framework of 3D-GS (i.e. \\(\\mathbf{G}_{\\theta}\\)) [14], establishing a one-to-one correspondence between poses and virtual sharp images is feasible. Consequently, we formulate the corresponding poses of each latent sharp image within the exposure time \\(\\tau\\) by employing a camera motion trajectory represented through linear interpolation between two camera poses, one at the beginning of the exposure \\(\\mathbf{T}_{\\mathrm{start}} \\in \\mathbf{SE}(3)\\) and the other at the end \\(\\mathbf{T}_{\\mathrm{end}} \\in \\mathbf{SE}(3)\\). The virtual camera pose at time \\(t \\in [0, \\tau]\\) can thus be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.601, + 0.785, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {t} = \\mathbf {T} _ {\\text {s t a r t}} \\cdot \\exp \\left(\\frac {t}{\\tau} \\cdot \\log \\left(\\mathbf {T} _ {\\text {s t a r t}} ^ {- 1} \\cdot \\mathbf {T} _ {\\text {e n d}}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.763 + ], + "angle": 0, + "content": "where \\(\\tau\\) represents the exposure time. \\(\\frac{t}{\\tau}\\) can be further discretized and derived as \\(\\frac{i}{n-1}\\) for the \\(i^{th}\\) sampled virtual sharp image (i.e. \\(C_i\\)) in Eq. (7) (i.e. with pose denoted as \\(\\mathbf{T}_i\\)), when there are \\(n\\) sampled images in total. It follows that \\(\\mathbf{T}_i\\) is differentiable with respect to both \\(\\mathbf{T}_{\\mathrm{start}}\\) and \\(\\mathbf{T}_{\\mathrm{end}}\\). We refer to prior works [26, 47] for a comprehensive understanding of the interpolation and derivations of the related Jacobian. [26]. The objective of BAD-Gaussians is to estimate both \\(\\mathbf{T}_{\\mathrm{start}}\\) and \\(\\mathbf{T}_{\\mathrm{end}}\\) for each frame, along with the learnable parameters of Gaussians \\(\\mathbf{G}_{\\theta}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.785, + 0.378, + 0.799 + ], + "angle": 0, + "content": "3.4 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.841 + ], + "angle": 0, + "content": "From a collection of \\(K\\) motion-blurred images, we can proceed to estimate both the learnable parameters \\(\\pmb{\\theta}\\) (i.e. mean position \\(\\pmb{\\mu}\\), 3D covariance \\(\\pmb{\\Sigma}\\), opacity \\(\\mathbf{o}\\))" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.537, + 0.129 + ], + "angle": 0, + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Ablation studies on the number of virtual camera poses \\( n \\). The results indicate that performance reaches a saturation point with increasing number \\( n \\)." + }, + { + "type": "table", + "bbox": [ + 0.361, + 0.185, + 0.64, + 0.286 + ], + "angle": 0, + "content": "
nCozy2roomTanabata
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
331.63.9389.081025.43.8215.2287
432.99.9552.055827.08.8704.1764
533.68.9595.045828.39.9011.1400
633.86.9627.036829.57.9214.1101
834.48.9654.031531.15.9440.0696
1034.68.9521.025832.12.9481.0464
1534.78.9679.025733.58.9652.0198
2034.63.9646.030733.87.9668.0151
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.286, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Table 2: Ablation studies on the effect of trajectory representations. We denote Deblur-NeRF-S and Deblur-NeRF-R as the synthetic and real data from Deblur-NeRF, respectively. The results demonstrate that cubic interpolation improves performance in scenes with complex camera trajectories (i.e. \\(MBA - VO\\) and Deblur-NeRF-R)." + }, + { + "type": "table", + "bbox": [ + 0.272, + 0.353, + 0.722, + 0.401 + ], + "angle": 0, + "content": "
Deblur-NeRF-SMBA-VODeblur-NeRF-R
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Linear Interpolation33.92.9467.042230.96.8868.149124.77.7778.1291
Cubic B-spline33.24.9395.047331.06.8883.145426.11.8157.1067
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.436, + 0.788, + 0.513 + ], + "angle": 0, + "content": "and color \\(\\mathbf{c}\\)) of 3D-GS and the camera motion trajectories (i.e. \\(\\mathbf{T}_{\\mathrm{start}}\\) and \\(\\mathbf{T}_{\\mathrm{end}}\\)) for each image. This estimation is accomplished by minimizing the following loss function, which includes an \\(\\mathcal{L}_1\\) loss and a D-SSIM term between \\(\\mathbf{B}_k(\\mathbf{u})\\), the \\(k^{th}\\) blurry image synthesized from 3D-GS using the aforementioned image formation model (i.e. Eq. (7)), and \\(\\mathbf{B}_k^{gt}(\\mathbf{u})\\), the corresponding real captured blurry image:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.532, + 0.786, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.57, + 0.788, + 0.616 + ], + "angle": 0, + "content": "To optimize the learnable Gaussians parameter \\(\\pmb{\\theta}\\), and camera pose \\(\\mathbf{T}\\) (i.e. \\(\\mathbf{T}_{\\mathrm{start}}\\) and \\(\\mathbf{T}_{\\mathrm{end}}\\) in our model) for each image, it is necessary to derive the corresponding Jacobians to complete the gradient flow:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.634, + 0.786, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {\\theta}} = \\sum_ {k = 0} ^ {K - 1} \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {B} _ {k}} \\cdot \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\frac {\\partial \\mathbf {B} _ {k}}{\\partial \\mathbf {C} _ {i}} \\frac {\\partial \\mathbf {C} _ {i}}{\\partial \\boldsymbol {\\theta}}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.707, + 0.786, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {T}} = \\sum_ {k = 0} ^ {K - 1} \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {B} _ {k}} \\cdot \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\frac {\\partial \\mathbf {B} _ {k}}{\\partial \\mathbf {C} _ {i}} \\frac {\\partial \\mathbf {C} _ {i}}{\\partial \\boldsymbol {\\theta}} \\frac {\\partial \\boldsymbol {\\theta}}{\\partial \\mathbf {T}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}_k(\\mathbf{u})\\) and \\(\\mathbf{C}_i(\\mathbf{u})\\) are denoted as \\(\\mathbf{B}_k\\) and \\(\\mathbf{C}_i\\) for simplification, and we parameterize both \\(\\mathbf{T}_{\\mathrm{start}}\\) and \\(\\mathbf{T}_{\\mathrm{end}}\\) with their corresponding Lie algebras of \\(\\mathbf{SE}(3)\\), which can be represented by a 6D vector respectively. For further details on the Jacobian of Gaussians to camera pose, \\(\\frac{\\partial\\pmb{\\theta}}{\\partial\\mathbf{T}}\\), please refer to our supplemental materials." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.286 + ], + "angle": 0, + "content": "Table 3: Quantitative comparisons for deblurring on the synthetic dataset of Deblur-NeRF [27], referred to as DB-NeRF in the table due to space constraints. Notably, DB-NeRF* and DP-NeRF* are trained with ground-truth poses, while the others are trained with poses estimated by COLMAP [38]. The experiments highlight the superior performance of our method over previous approaches. Additionally, the results demonstrate the sensitivity of Deblur-NeRF and DP-NeRF to pose accuracy. Regarding rendering efficiency, our method achieves over 200 FPS, whereas Deblur-NeRF, DP-NeRF, BAD-NeRF and ExBluRF fall below 1 FPS. Our method takes about 30 minutes to train, while other methods take more than 10 hours. Each color shading indicates the best and second-best result, respectively." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.296, + 0.784, + 0.424 + ], + "angle": 0, + "content": "
CozyroomFactoryPoolTanabataTrolley
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
NeRF [30]26.13.7886.248422.02.5581.455029.90.7901.259520.57.5653.458421.71.6413.3814
3D-GS [14]25.86.7908.227021.73.5623.450329.37.7816.256120.51.5773.427821.65.6587.3614
MPR [52]29.90.8862.091525.07.6994.240933.28.8938.129022.60.7203.250726.24.8356.1762
SRN [44]29.47.8759.095026.54.7604.240432.94.8847.104523.19.7274.243825.36.8119.1618
DB-NeRF [27]29.53.8786.087925.85.7651.234031.07.8477.139723.20.7243.247125.68.8111.1683
DB-NeRF*30.26.8933.079126.40.7991.219132.30.8755.134524.56.7749.216626.24.8254.1671
DP-NeRF [19]29.02.8761.077325.42.7559.213930.48.8402.112723.93.7540.203826.17.8269.1359
DP-NeRF*30.77.9020.058427.69.8328.184733.22.8922.095425.27.7973.177926.99.8413.1312
BAD-NeRF [47]32.11.9137.051432.18.9105.118932.22.8680.090825.80.8047.156329.68.8954.0752
ExBluRF [20]28.460.87070.111329.380.85110.164631.660.86120.171925.430.80060.22725.70.81120.234
Ours34.68.9521.025831.88.9270.095236.95.9434.022532.12.9481.046433.97.9628.0209
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Table 4: Quantitative comparisons for novel view synthesis on the synthetic dataset of Deblur-NeRF [27], referred to as DB-NeRF in the table due to space constraints. The results demonstrate Our methods outperform previous state-of-the-art approaches, delivering the best performance across the board." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.508, + 0.784, + 0.631 + ], + "angle": 0, + "content": "
CozyroomFactoryPoolTanabataTrolley
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [30]25.45.7734.251422.51.5672.419929.23.7711.275620.74.5774.434121.09.6157.3924
3D-GS [14]25.24.7789.208421.34.5532.437128.61.7562.261020.37.5755.395820.86.6167.3701
MPR [52]+ [14]29.00.8685.089423.38.6848.237429.04.8170.176122.53.7349.216226.27.8352.1694
SRN [44]+ [14]28.64.8662.090325.45.7740.204529.97.8207.165023.09.7456.214124.89.8103.1666
DB-NeRF [27]29.09.8718.093725.19.7359.231530.97.8401.160023.71.7483.236924.46.7758.2215
DB-NeRF*29.88.8901.074726.06.8023.210630.94.8399.169424.82.7861.204525.78.8122.1797
DP-NeRF [19]29.56.8831.068327.91.8317.183330.98.8454.126524.73.7867.186525.62.8123.1606
DP-NeRF*30.16.8958.060828.43.8420.177131.78.8623.130625.37.8046.170126.44.8307.1424
BAD-NeRF [47]31.07.9027.055131.71.9038.120430.95.8389.110725.42.8011.156628.60.8806.0843
Ours32.35.9278.034028.18.9164.099933.30.8964.055131.00.9427.047831.44.9443.0345
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.377, + 0.679 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.699, + 0.447, + 0.715 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation Details. We implemented our method within the Nerfstudio [43] framework based on their Splatifto implementation. Camera pose interpolation and optimization on the SE(3) manifold is implemented with pypose [45]. Both the optimization of Gaussians and camera poses are performed using the Adam optimizer. The number of virtual camera poses (i.e., \\( n \\) in Equation (7)) is set to 10, considering the trade-off between performance and efficiency. We initialize camera poses and Gaussians using estimations obtained" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.537, + 0.129 + ], + "angle": 0, + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Quantitative comparisons for deblurring on the synthetic dataset of MBA-VO [27]. The results demonstrate that our method achieves the best performance even when the camera undergoes acceleration" + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.199, + 0.669, + 0.32 + ], + "angle": 0, + "content": "
ArchViz-lowArchViz-highAverage
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
NeRF [30]26.26.7887.377523.29.7098.491024.78.7493.4343
3D-GS [14]26.54.8113.338523.43.7243.461424.99.7678.4000
MPR [52]29.60.8757.210325.04.7711.357627.32.8234.2840
SRN [44]30.15.8814.170327.07.8190.279628.61.8502.2250
DB-NeRF [27]28.38.8484.179225.71.7762.322027.05.8123.2506
DB-NeRF*29.65.8744.176426.44.8010.317228.05.8377.2468
DP-NeRF [19]28.24.8490.161525.74.7839.291326.99.8165.2264
DP-NeRF*28.57.8481.151126.80.8081.273627.69.8281.2124
BAD-NeRF [47]30.51.8749.165427.54.8109.267929.03.8429.2167
ExBluRF [20]28.4.8372.251828.43.8518.172128.42.8445.2120
Ours32.28.9167.113429.64.8568.184730.96.8868.1491
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.784, + 0.378 + ], + "angle": 0, + "content": "from COLMAP [38]. All experiments are conducted on an NVIDIA RTX 4090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.534 + ], + "angle": 0, + "content": "Benchmark Datasets. We evaluate the performance of our method using both synthetic and real datasets provided by Deblur-NeRF [27] and BAD-NeRF [47]. Deblur-NeRF [47] contains five scenes, where the blurred images are generated via Blender [7] by averaging sharp virtual images captured during the exposure time, under the assumption of consistent velocity camera motion. For our evaluation, we utilize the dataset from BAD-NeRF [47], which expands upon the number of virtual sharp images to create more realistic motion-blurred images while keeping other settings consistent with [27]. Additionally, Deblur-NeRF [27] captured a real motion-blurred dataset by intentionally shaking a handheld camera during exposure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.785, + 0.64 + ], + "angle": 0, + "content": "To enhance the evaluation of our method on severely motion-blurred images, we incorporate a dataset tailored for motion blur-aware visual odometry benchmarking (i.e. MBA-VO [26]). Different from the assumption of constant velocity in the Deblur-NeRF dataset [27], the blur images synthesized from MBA-VO [26] are based on real camera motion trajectories obtained from the ETH3D dataset [39], which do not exhibit constant velocity and include accelerations, thus presenting notable challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Baselines and Evaluation Metrics. We conduct a comparative deblurring analysis between our method and state-of-the-art learning-based single image deblurring algorithms including SRN [44] and MPR [52]. Additionally, we include evaluations against closely related approaches such as Deblur-NeRF [27], DP-NeRF [19], BAD-NeRF [47] and ExBluRF [20]. To evaluate deblurring performance, we render high-quality images corresponding to the midpoint (i.e. \\(\\mathbf{C}_{\\frac{\\tau}{2}}\\) in Eq. (6)) of exposure time of each training image from the optimized Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To facilitate novel view synthesis using SRN [44] and MPR [52], we use the deblurred images obtained from pre-trained models as inputs for training 3D-GS, as these single deblurring methods are not optimized for synthesizing novel images. During the training stage, all poses are estimated using COLMAP [38]. Additionally, we train Deblur-NeRF [27] and DP-NeRF [19] with ground-truth camera poses from Blender to assess the impact of inaccurate pose estimation on" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.797, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig. 2: Qualitative novel view synthesis results of different methods with synthetic datasets. Despite being trained with ground truth poses \\((^{*})\\), BAD-Gaussians outperforms Deblur-NeRF* and DP-NeRF* in recovering high-quality scenes from motion-blurred images with inaccurate camera poses, showcasing its superior performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.785, + 0.584 + ], + "angle": 0, + "content": "these two methods. The quality of the rendered sharp image is evaluated using PSNR, SSIM, and LPIPS metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.787, + 0.632 + ], + "angle": 0, + "content": "For pose accuracy evaluation, we compare the Absolute Trajectory Error (ATE) metric, against classical structure-from-motion framework COLMAP [38] and BAD-NeRF [47]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.39, + 0.673 + ], + "angle": 0, + "content": "4.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Virtual Camera Poses. We experimented to investigate the effect of the number of interpolated virtual camera poses within the exposure time, as described by \\( n \\) in Eq. (7). Two of the five synthetic scenes provided by Deblur-NeRF [27] were selected, representing sequences with minor and severe motion blur (i.e. Cozy2room and Tanabata), respectively. We varied the number, \\( n \\), from 3 to 20 in our study, and the rendering metric results are presented in Table 1. Our experimental results demonstrate that increasing the number of interpolated virtual camera poses helps in addressing severe motion blur. However, marginal improvements are observed for minor blur instances. Based on our experiments, we choose \\( n = 10 \\) interpolated virtual images to strike a balance between ren" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.537, + 0.129 + ], + "angle": 0, + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "dering performance and training efficiency (larger \\( n \\) means more computational resources)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Trajectory Representations. To assess the impact of different trajectory representations, we conduct two experiments: one involves optimizing \\(\\mathbf{T}_{\\mathrm{start}}\\) and \\(\\mathbf{T}_{\\mathrm{end}}\\) to depict a linear trajectory, while the other utilizes a higher-order spline (i.e., cubic B-spline) that jointly optimizes four control knots \\(\\mathbf{T}_1\\), \\(\\mathbf{T}_2\\), \\(\\mathbf{T}_3\\), and \\(\\mathbf{T}_4\\) to capture more complex camera motions. We refer [46] for more details about cubic B-spline. The average quantitative results on the synthetic datasets of Deblur-NeRF-S [27] (Cozy2room, Factory, Pool, Tanabata and Trolley) and MBA-VO [26] (ArchViz-low and ArchViz-high), along with real datasets, Deblur-NeRF-S, which includes 10 real captured scenes from Deblur-NeRF, are presented in Table 2. In synthetic scenes, linear interpolation demonstrates comparable performance to cubic B-spline interpolation. However, in real captured scenes, cubic B-spline interpolation outperforms linear interpolation, particularly due to the longer exposure time during image capture. The effectiveness of cubic B-spline interpolation in real scenes can be attributed to its ability to better model the nuances of camera motion over longer time intervals. Conversely, linear interpolation is sufficient to accurately represent camera motion trajectories within shorter time intervals, as observed in synthetic scenes. Combining the training efficiency and rendering quality, we employ linear interpolation in synthetic datasets and cubic B-spline in real data." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.492, + 0.629, + 0.594 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS↓Ttrain(h)
NeRF [30]22.69.6347.36872.5
3DGS [14]22.72.6390.28870.3
DB-NeRF [27]25.63.7675.1820>5
DP-NeRF [19]25.91.7734.1589>5
BAD-NeRF [47]24.33.7003.2196>5
ExBluRF [20]25.31.7293.20844.0
Ours-Cubic26.11.8157.10640.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.594, + 0.784, + 0.622 + ], + "angle": 0, + "content": "Table 6: Quantitative comparisons for novel view synthesis on the real captured dataset of Deblur-NeRF [27]. Full results can be found in our supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.68, + 0.325, + 0.694 + ], + "angle": 0, + "content": "4.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.784, + 0.81 + ], + "angle": 0, + "content": "Results on Synthetic Data. We evaluate our approach against baseline methods using synthetic datasets obtained from both Deblur-NeRF [27] and MBAVO [26]. The quantitative evaluation results on deblurring and novel view synthesis with the Deblur-NeRF dataset are presented in Table 3 and Table 4 respectively. All metrics (i.e. PSNR, SSIM, LPIPS) demonstrate substantial improvements over prior state-of-the-art methods (i.e. 3.6 and \\(1.7\\mathrm{dB}\\) higher than the second best method on average in Table 3 and Table 4, respectively)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The results demonstrate that both NeRF [30] and 3D-GS [14] are suffering from motion-blurred images, which motivates the necessity of our method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Fig. 3: Qualitative novel view synthesis results of different methods with the real datasets. The experimental results demonstrate that our method achieves superior performance over prior methods on the real dataset as well. In contrast, BAD-NeRF yields poorer results when applied to real data and exhibits satisfactory performance only within synthetic datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Single-stage methods such as MPR [52] and SRN [44] fall short of matching the performance of our approach due to their limited utilization of geometric information across multi-view images. Additionally, our method surpasses Deblur-NeRF [27] and DP-NeRF [19], partly because they fail to optimize the inaccurate camera poses estimated from COLMAP during training. We additionally trained Deblur-NeRF and DP-NeRF with ground truth poses (denoted as *), and the results demonstrate improved performance, reflecting the sensitivity to the pose accuracies of both Deblur-NeRF and DP-NeRF. A notable result from both Table 3 and Table 4 is that our method performs worse than BAD-NeRF on the Factory sequence. We find that it is caused by the inferior capability to represent the sky by Gaussian splatting compared to NeRF. Nevertheless, our method still performs better than all prior methods with a large margin on average." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Qualitative results are presented in Fig. 2. It demonstrates that our method effectively attains high-quality scene representations with intricate details, being trained from a series of blurred images. In Fig. 2, it is evident that Deblur-NeRF [27] and DP-NeRF [19] encounter difficulties in modeling regions with significant depth variations, attributable to their method of synthesizing blurred images through convolution with a blur kernel applied to rendered images. The physically based method BAD-NeRF [47] exhibits superior performance overall," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.537, + 0.129 + ], + "angle": 0, + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.19 + ], + "angle": 0, + "content": "yet it still presents some deficiencies, particularly noticeable around areas with significant color and depth variation. It demonstrates the effectiveness of the explicit Gaussian splatting representation over the implicit neural representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.784, + 0.265 + ], + "angle": 0, + "content": "We further assess the performance of our methods using a dataset characterized by severe motion blur and camera movements with varying velocities, sourced from MBA-VO [26]. The results presented in Table 5 demonstrate that our method achieves the best performance even when the camera undergoes acceleration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.784, + 0.345 + ], + "angle": 0, + "content": "Results on Real Data. The performance of novel view synthesis on real data sourced from Deblur-NeRF [27] are also evaluated. The quantitative results, as presented in Table 6, demonstrate the superior performance of our method compared to other approaches. Furthermore, the qualitative results illustrated in Fig. 3 vividly demonstrate our method's ability to deliver intricate details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.784, + 0.409 + ], + "angle": 0, + "content": "Pose Estimation. Due to the unknown metric scale, we align the estimated trajectories against their corresponding ground truth for the computation of the absolute trajectory error metric. The experimental results presented in Table 7 demonstrate the effectiveness of our approach in recovering precise camera poses." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.437, + 0.784, + 0.491 + ], + "angle": 0, + "content": "Table 7: Pose estimation performance of BAD-Gaussians on various blur sequences. The results are in the absolute trajectory error metric in centimeters (ATE/cm). The COLMAP-sharp / COLMAP-blur represents the result of COLMAP with sharp/blurry images respectively." + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.507, + 0.737, + 0.56 + ], + "angle": 0, + "content": "
Cozy2roomFactoryPoolTanabataTrolleyArchViz-lowArchViz-highAverage
COLMAP-sharp [38].133±.079.101±.052.172±.047.277±.103.173±.079.148±.078.115±.066.1599
COLMAP-blur [38].287±.1611.69±.1.24.633±.2361.59±.1.06.912±.559.536±.309.766±.391.9163
BAD-NeRF [47].316±.101.600±.208.095±.0461.34±.543.122±.0701.24±.6352.18±1.54.8419
Ours.165±.057.395±.426.264±.0981.20±.621.134±.070.520±.5911.41±1.11.5840
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.613, + 0.358, + 0.629 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.784, + 0.733 + ], + "angle": 0, + "content": "We presented the first pipeline to learn Gaussian splittings from a set of motion-blurred images with inaccurate camera poses. Our pipeline can jointly optimize the 3D scene representation and camera motion trajectories. Extensive experimental evaluations demonstrate that our method can deliver high-quality novel view images, and achieve real-time rendering compared to prior state-of-the-art works." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.759, + 0.403, + 0.777 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.791, + 0.784, + 0.836 + ], + "angle": 0, + "content": "This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.235 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://jonbarron.info/mipnerf/2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.236, + 0.786, + 0.278 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://jonbarron.info/mipnerf360/3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.786, + 0.32 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2023), https://jonbarron.info/zipnerf/ 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.786, + 0.362 + ], + "angle": 0, + "content": "4. Cao, A., Johnson, J.: HexPlane: A Fast Representation for Dynamic Scenes. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://caoang327.github.io/HexPlane/ 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.786, + 0.404 + ], + "angle": 0, + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensorRF: Tensorial Radiance Fields. In: European Conference on Computer Vision (ECCV) (2022), https://apchenstu.github.io/TensoRF/ 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.433 + ], + "angle": 0, + "content": "6. Cho, S., Lee, S.: Fast Motion Deblurring. ACM Transactions on Graphics (TOG) 28(5) (2009) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.786, + 0.474 + ], + "angle": 0, + "content": "7. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.786, + 0.517 + ], + "angle": 0, + "content": "8. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised NeRF: Fewer Views and Faster Training for Free. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://www.cs.cmu.edu/~dsnerf/3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.786, + 0.559 + ], + "angle": 0, + "content": "9. Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T.: Removing camera shake from a single photograph. In: ACM Transactions on Graphics (TOG) (2006) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.786, + 0.613 + ], + "angle": 0, + "content": "0. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://sarafridov.github.io/K-Planes/4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.615, + 0.786, + 0.657 + ], + "angle": 0, + "content": "1. Fu, Y., Liu, S., Kulkarni, A., Kautz, J., Efros, A.A., Wang, X.: Colmap-Free 3D Gaussian Splatting. arXiv preprint arXiv:2312.07504 (2023), https://oasisyang.github.io/colmap-free-3dgs/2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.786, + 0.7 + ], + "angle": 0, + "content": "2. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: FastNeRF: High-Fidelity Neural Rendering at 200FPS. In: International Conference on Computer Vision (ICCV) (2021), https://microsoft.github.io/FastNeRF/2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.786, + 0.742 + ], + "angle": 0, + "content": "3. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-Calibrating Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://postech-cvlab.github.io/SCNeRF/ 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.786, + 0.798 + ], + "angle": 0, + "content": "4. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (TOG) 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ 2, 4, 5, 7, 9, 10, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Kim, M., Seo, S., Han, B.: InfoNeRF: Ray Entropy Minimization for Few-Shot Neural Volume Rendering. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://cv.snu.ac.kr/research/InfoNeRF/3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.537, + 0.129 + ], + "angle": 0, + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Krishnan, D., Fergus, R.: Fast image deconvolution using Hyper-Laplacian priors. In: Neural Information Processing Systems (NIPS) (2009) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better. In: International Conference on Computer Vision (ICCV) (2019), https://github.com/VITA-Group/DeblurGANv2 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.257 + ], + "angle": 0, + "content": "18. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://github.com/facebookresearch/pytorch3d/ 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.785, + 0.298 + ], + "angle": 0, + "content": "19. Lee, D., Lee, M., Shin, C., Lee, S.: DP-NeRF: Deblurred Neural Radiance Field With Physical Scene Priors. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://dogyoonlee.github.io/dpnerf/2, 3, 4, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: ExBluRF: Efficient Radiance Fields for Extreme Motion Blurred Images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023) 4, 9, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "21. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithm. In: Computer Vision and Pattern Recognition (CVPR) (2009) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "22. Levoy, M.: Efficient Ray Tracing of Volume Data. ACM Transactions on Graphics (TOG) 9(3), 245-261 (1990) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.405, + 0.785, + 0.446 + ], + "angle": 0, + "content": "23. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields. In: International Conference on Learning Representations (ICLR) (2024), https://moyangli00.github.io/usb_nerf/ 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.446, + 0.785, + 0.499 + ], + "angle": 0, + "content": "24. Li, W., Wan, P., Wang, P., Li, J., Zhou, Y., Liu, P.: BeNeRF: Neural Radiance Fields from a Single Blurry Image and Event Stream. In: European Conference on Computer Vision (ECCV). Springer (2024), https://github.com/WU-CVGL/BeNeRF 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.5, + 0.785, + 0.555 + ], + "angle": 0, + "content": "25. Lin, Chen-Hsuan and Ma, Wei-Chiu and Torralba, Antonio and Lucey, Simon: BARF: Bundle-Adjusting Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://chenhsuanlin.bitbucket.io/bundle-adjusting-NeRF 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "26. Liu, P., Zuo, X., Larsson, V., Pollefeys, M.: MBA-VO: Motion Blur Aware Visual Odometry. In: International Conference on Computer Vision (ICCV) (2021), https://github.com/ethliup/MBA-VO 7, 10, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.596, + 0.785, + 0.65 + ], + "angle": 0, + "content": "27. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural Radiance Fields from Blurry Images. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://limacv.github.io/deblurnerf/ 2, 3, 4, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.65, + 0.785, + 0.677 + ], + "angle": 0, + "content": "28. Max, N.: Optical Models for Direct Volume Rendering. ACM Transactions on Graphics (TOG) 1(2), 99-108 (1995) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.677, + 0.785, + 0.731 + ], + "angle": 0, + "content": "29. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://bmild.github.io/rawnerf/3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.731, + 0.785, + 0.785 + ], + "angle": 0, + "content": "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: European Conference on Computer Vision (ECCV) (2020), https://www.matthewtancik.com/nerf 2, 3, 9, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.785, + 0.839 + ], + "angle": 0, + "content": "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics (TOG) 41(4), 102:1-102:15 (Jul 2022), https://nvlabs.github.io/instant-ngp/2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "32. Nah, S., Kim, T.H., Lee, K.M.: Deep Multi-Scale Convolutional Neural Network for Dynamic Scene Deblurring. In: Computer Vision and Pattern Recognition (CVPR) (2017), https://github.com/SeungjunNah/DeepDeblur_release 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.247 + ], + "angle": 0, + "content": "33. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse Inputs. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://m-niemeyer.github.io/regnerf/3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.247, + 0.788, + 0.302 + ], + "angle": 0, + "content": "34. Park, H., Lee, K.M.: Joint Estimation of Camera Pose, Depth, Deblurring, and Super-Resolution from a Blurred Image Sequence. In: International Conference on Computer Vision (ICCV). vol. 2017-Octob (2017). https://doi.org/10.1109/ICCV.2017.494 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.302, + 0.788, + 0.346 + ], + "angle": 0, + "content": "35. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: CamP: Camera Preconditioning for Neural Radiance Fields. ACM Transactions on Graphics (TOG) (2023), https://camp-nerf.github.io/3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.389 + ], + "angle": 0, + "content": "36. Piala, M., Clark, R.: TermiNeRF: Ray Termination Prediction for Efficient Neural Rendering. In: International Conference on 3D Vision (3DV) (2021), https://projects.mackopes.com/terminerf/3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.432 + ], + "angle": 0, + "content": "37. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance Fields without Neural Networks. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://alexyu.net/plenoxels/ 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.473 + ], + "angle": 0, + "content": "38. Schonberger, J.L., Frahm, J.M.: Structure-from-motion Revisited. In: Computer Vision and Pattern Recognition (CVPR) (2016), https://github.com/colmap/ colmap 2, 5, 9, 10, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.516 + ], + "angle": 0, + "content": "39. Schops, T., Sattler, T., Pollefeys, M.: BAD SLAM: Bundle Adjusted Direct RGB-D SLAM. In: Computer Vision and Pattern Recognition (CVPR) (2019), https://github.com/ETH3D/badslam 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.546 + ], + "angle": 0, + "content": "40. Shan, Q., Jia, J., A., A.: High-quality Motion Deblurring from a Single Image. ACM Transactions on Graphics (TOG) 27(3) (2008) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.588 + ], + "angle": 0, + "content": "41. Su, S., Delbracio, M., Wang, J.: Deep Video Deblurring for Hand-held Cameras. In: Computer Vision and Pattern Recognition (CVPR) (2017), https://github.com/shuochsu/DeepVideoDeblurring 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.63 + ], + "angle": 0, + "content": "42. Sun, C., Sun, M., Chen, H.: Direct Voxel Grid Optimization: Super-fast Convergence for Radiance Fields Reconstruction. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://sunset1995.github.io/dvgo/3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.699 + ], + "angle": 0, + "content": "43. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A Modular Framework for Neural Radiance Field Development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023), https://github.com/nerfstudio-project/nerfstudio/9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.743 + ], + "angle": 0, + "content": "44. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Computer Vision and Pattern Recognition (CVPR) (2018), https://github.com/jiangsutx/SRN-Deblur 4, 9, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.843 + ], + "angle": 0, + "content": "45. Wang, C., Gao, D., Xu, K., Geng, J., Hu, Y., Qiu, Y., Li, B., Yang, F., Moon, B., Pandey, A., Aryan, Xu, J., Wu, T., He, H., Huang, D., Ren, Z., Zhao, S., Fu, T., Reddy, P., Lin, X., Wang, W., Shi, J., Talak, R., Cao, K., Du, Y., Wang, H., Yu, H., Wang, S., Chen, S., Kashyap, A., Bandaru, R., Dantu, K., Wu, J., Xie, L., Carlone, L., Hutter, M., Scherer, S.: PyPose: A library for robot learning with physics-based optimization. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023), https://github.com/pypyse/pypyse 9" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.843 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.538, + 0.129 + ], + "angle": 0, + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "46. Wang, P., Zhao, L., Ma, R., Liu, P.: Supplementary Material for BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields https://openaccess.thecvf.com/content/CVPR2023/supplemental/Wang_BAD-NeRF_Bundle_Adjusted_CVPR_2023SUPPLEMENTAL.pdf 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "47. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://wangpeng000.github.io/BAD-NeRF/2, 3, 4, 7, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.285 + ], + "angle": 0, + "content": "48. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021), https://nerfmm.active.vision/3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "49. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: NeX: Real-time View Synthesis with Neural Basis Expansion. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://nex-mpi.github.io/3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.787, + 0.357 + ], + "angle": 0, + "content": "50. Xu, L., Jia, J.: Two-Phase Kernel Estimation for robust motion deblurring. In: European Conference on Computer Vision (ECCV) (2010) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "51. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for Real-time Rendering of Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://alexyu.net/plenoctrees/ 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage Progressive Image Restoration. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://github.com/swz30/MPRNet 9, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.482 + ], + "angle": 0, + "content": "53. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: EWA Volume Splatting. In: Proceedings Visualization, 2001. VIS'01. pp. 29-538. IEEE (2001), https://cgl.ethz.ch/research/past_projects/surfels/ewavolumesplatting/index.html 6" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.482 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_origin.pdf b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6af059ac3ae2e79e404e269b58f95e9f88626203 --- /dev/null +++ b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/4f2c6ae0-3ef6-48eb-ba10-7da18a4a4629_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc61b3d864cadd56ca4f705376bf55e9773ccf2f41d0bd12c1a1aab004688ff3 +size 8922650 diff --git a/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/full.md b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..62a376589d4b97cbb1c77ae33ff87b01b50a86fd --- /dev/null +++ b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/full.md @@ -0,0 +1,289 @@ +# BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting + +Lingzhe Zhao $^{1*}$ , Peng Wang $^{1,2*}$ , and Peidong Liu $^{1\dagger}$ + +$^{1}$ Westlake University $^{2}$ Zhejiang University $\{$ zhaolingzhe, wangpeng, liupeidong\}@westlake.edu.cn https://github.com/wu-cvgl/BAD-Gaussians/ + +Abstract. While neural rendering has demonstrated impressive capabilities in 3D scene reconstruction and novel view synthesis, it heavily relies on high-quality sharp images and accurate camera poses. Numerous approaches have been proposed to train Neural Radiance Fields (NeRF) with motion-blurred images, commonly encountered in real-world scenarios such as low-light or long-exposure conditions. However, the implicit representation of NeRF struggles to accurately recover intricate details from severely motion-blurred images and cannot achieve real-time rendering. In contrast, recent advancements in 3D Gaussian Splatting achieve high-quality 3D scene reconstruction and real-time rendering by explicitly optimizing point clouds into 3D Gaussians. In this paper, we introduce a novel approach, named BAD-Gaussians (Bundle Adjusted Deblur Gaussian Splatting), which leverages explicit Gaussian representation and handles severe motion-blurred images with inaccurate camera poses to achieve high-quality scene reconstruction. Our method models the physical image formation process of motion-blurred images and jointly learns the parameters of Gaussians while recovering camera motion trajectories during exposure time. In our experiments, we demonstrate that BAD-Gaussians not only achieves superior rendering quality compared to previous state-of-the-art deblur neural rendering methods on both synthetic and real datasets but also enables real-time rendering capabilities. + +Keywords: 3D Gaussian Splatting $\cdot$ Deblurring $\cdot$ Bundle Adjustment $\cdot$ Differentiable Rendering + +# 1 Introduction + +Acquiring accurate 3D scene representations from 2D images has long been a challenging problem in computer vision. Serving as a fundamental component in various applications such as virtual/augmented reality and robotics navigation, substantial efforts have been dedicated to addressing this challenge over the last few decades. Among those pioneering works, Neural Radiance Fields + +(NeRF) [30], parameterized by Multi-layer Perceptrons (MLP), stands out for its utilization of differentiable volume rendering technique [22, 28] and has garnered significant attention due to its capability to recover high-quality 3D scene representation from 2D images. + +Numerous works have focused on enhancing the performance of NeRF, particularly in terms of training [5,31,37] and rendering efficiency [12,51]. A recent advancement, 3D Gaussian Splatting (3D-GS) [14], extends the implicit neural rendering [30] to explicit point clouds. By projecting these optimized point clouds (Gaussians) onto the image plane, 3D-GS [14] achieves real-time rendering while enhancing the efficiency of NeRF in both training and rendering, and also improves rendering quality. However, both NeRF-based [1,30,31] methods and 3D-GS [14] heavily rely on well-captured sharp images and accurately precomputed camera poses, typically obtained from COLMAP [38]. Motion-blurred images, a common form of image degradation, often encountered in low-light or long-exposure conditions, can notably impair the performance of both NeRF and 3D-GS. The challenges posed by motion-blurred images to NeRF and 3D-GS can be attributed to three primary factors: (a) NeRF and 3D-GS rely on high-quality sharp images for supervision. However, motion-blurred images violate this assumption and exhibit notably inaccurate corresponding geometry between multi-view frames, thus presenting significant difficulties in achieving accurate 3D scene representation for both NeRF and 3D-GS; (b) Accurate camera poses are essential for training NeRF and 3D-GS. However, recovering accurate poses from multi-view motion-blurred images using COLMAP [38] is challenging. (c) 3D-GS necessitates sparse cloud points from COLMAP as the initialization of Gaussians. The mismatched features between multi-view blurred images and the inaccuracies in pose calibration further exacerbate the issue, leading to COLMAP producing fewer cloud points. This introduces an additional initialization issue for 3D-GS. Therefore, these factors result in a notable drop in performance for 3D-GS when dealing with motion-blurred images. + +Implicit neural representations, i.e. NeRF [30], have been employed to reconstruct sharp 3D scenes from motion-blurred images [19,27,47]. For example, Deblur-NeRF [27] introduces a deformable sparse kernel that alters a canonical kernel at spatial locations to simulate the blurring process. DP-NeRF [19] integrates physical priors derived from the motion-blurred image acquisition process into Deblur-NeRF [27] to construct a clean NeRF representation. In contrast, BAD-NeRF [47] models the physical process of capturing motion-blurred images and jointly optimizes NeRF while recovering the camera trajectory within the exposure time. However, these implicit deblur rendering methods encounter significant challenges in achieving real-time rendering and producing high-quality outputs with intricate details. Additionally, the implicit representation introduces extra difficulties in optimizing the neural parameters and camera poses as mentioned in [11]. + +In order to address these challenges, we propose Bundle Adjusted Deblur Gaussian Splatting, the first motion deblur framework based on 3D-GS, which we refer to as BAD-Gaussians. We incorporate the physical process of motion + +blur into the training of 3D-GS, employing a spline function to characterize the trajectory within the camera's exposure time. In the training of BAD-Gaussians, the camera trajectory within exposure time is optimized using gradients derived from the Gaussians of the scene, while jointly optimizing the Gaussians themselves. Specifically, the trajectory of each motion-blurred image is represented by the initial and final poses at the beginning and end of the exposure time, respectively. By assuming the exposure time is typically short, we can interpolate between the initial and final poses to obtain every camera pose along the trajectory. From this trajectory, we generate a sequence of virtual sharp images by projecting the scene's Gaussians onto the image plane. These virtual sharp images are then averaged to synthesize the blurred images, following the physical blur process. Finally, the Gaussians along the trajectory are optimized by minimizing the photometric error between the synthesized blurred images and the input blurred images through differentiable Gaussian rasterization. + +We evaluate BAD-Gaussians using both synthetic and real datasets. The experimental results demonstrate that BAD-Gaussians outperforms prior state-of-the-art implicit neural rendering methods by explicitly incorporating the image formation process of motion-blurred images into the training of 3D-GS, achieving better rendering performance in terms of real-time rendering speed and superior rendering quality. In summary, our contributions can be outlined as follows: + +- We introduce a photometric bundle adjustment formulation specifically designed for motion-blurred images, achieving the first real-time rendering performance from motion-blurred images within the framework of 3D Gaussian Splatting; +- We demonstrate how this formulation enables the acquisition of high-quality 3D scene representation from a set of motion-blurred images; +- Our approach successfully deblurs severe motion-blurred images, synthesizes higher-quality novel view images, and achieves real-time rendering, surpassing previous state-of-the-art implicit deblurring rendering methods. + +# 2 Related Work + +# 2.1 Neural Radiance Fields + +NeRF [30], employing implicit MLP, exhibits remarkable performance in synthesizing high-quality novel view images and accurately representing 3D scenes. Numerous extension works have been proposed to enhance NeRF's performance, including improvements in training [5,31,37,42] and rendering [12,18,36,49,51] efficiency, as well as anti-alias capabilities [1-3]. Additionally, various methods aim to bolster NeRF's robustness against imperfect inputs, such as inaccurate camera poses [13, 25, 35, 48], few-shot images [8, 15, 33], and low-quality images [19, 23, 24, 27, 29, 47]. In the following section, we will primarily focus on reviewing methods closely related to our work. + +Fast Neural Rendering. Numerous approaches inspired by NeRF [30] have sought to enhance its rendering efficiency by employing advanced data structures to reconstruct radiance fields, thereby minimizing the computational cost + +associated with implicit MLPs used in NeRF. These approaches are primarily categorized into grid-based [4, 5, 10, 37] and hash-based [31] methods. Despite these efforts, achieving real-time rendering for unbounded and complete scenes, as well as high-resolution images, remains challenging. In contrast to methods based on volume rendering and implicit representations, which may hinder fast rendering, recent advancements like 3D-GS [14] achieve real-time high-quality rendering through pure explicit point scene representation and the differentiable Gaussian rasterization. Nevertheless, these 3D scene representations heavily rely on accurately posed high-quality images. + +NeRF for Camera Optimization. The pioneering work, BARF [25], was the first to propose simultaneous optimization of camera parameters alongside NeRF. They employed a coarse-to-fine bundle adjustment strategy to enhance camera pose recovery. Concurrently, SC-NeRF [13] introduced a method to learn various camera models, encompassing both extrinsic and intrinsic parameters, along with scene representation. In the latest development, CamP [35] introduced a preconditioner to mitigate correlations between camera parameters, aiming for improved joint optimization of camera and NeRF parameters. In contrast to the aforementioned works [13,25,35], which focus solely on optimizing camera poses for sharp images, our approach goes further by recovering the trajectory of each blurred image within the exposure time. + +NeRF for Deblurring. Several scene deblurring methods based on NeRF have been proposed, including Deblur-NeRF [27] and DP-NeRF [19], which reconstruct sharp scene representations from sets of motion-blurred images. However, these methods fix inaccurate camera poses recovered from blurred images during training, leading to degradation in reconstruction performance. BAD-NeRF [47] and ExBluRF [20] jointly learns camera motion trajectories within exposure time and radiance fields, following the physical blur process. Despite these advancements, existing deblurring neural rendering strategies struggle to achieve real-time rendering and reconstruct intricate scene details due to the implicit MLP structure, necessitating significant improvements in rendering efficiency and quality. Moreover, optimizing the 3D scenes jointly with camera poses faces additional challenges due to NeRF's implicit representations. To address these issues, we propose achieving deblurring capability within the framework of Gaussian Splatting [14]. + +# 2.2 Image Deblurring + +Two primary categories typically classify existing techniques for addressing the motion deblurring problem: the first involves formulating the issue as an optimization task, wherein gradient descent is employed during inference to jointly refine the blur kernel and the latent sharp image [6, 9, 16, 21, 34, 40, 50]. Another pipeline phrases the deblurring task as an end-to-end learning, particularly leveraging deep learning methodologies. With the support of substantial datasets and deep neural networks, superior results have been achieved for both single image deblurring [17, 32, 44] and video deblurring [41]. However, these 2D deblurring + +![](images/15396aca0f0f42d1d32043761dc5782fcbb89522fdb0fd0ff89603e9a3434fc8.jpg) +Fig. 1: The pipeline of BAD-Gaussians. Our approach utilizes Gaussian representations to depict sharp 3D scenes derived from a series of motion-blurred images, along with their inaccurate poses and sparse point clouds from COLMAP. Employing forward projection and differentiable Gaussian rasterization, we jointly optimize the Gaussians in the scene and the camera trajectory within exposure time, by backpropagating gradients from Gaussians to camera poses. Following the physical process of motion blur, we model motion-blurred images by averaging the virtual sharp images captured during the exposure time. These virtual camera poses are represented and interpolated using a continuous spline within the SE(3) space. The joint optimization of Gaussians and camera trajectories is achieved by minimizing the photometric loss between synthesized and actual blurry images. + +methods cannot exploit the 3D scene geometry between multi-view images, thus failing to ensure the view consistency of the scene from different viewpoints. In contrast, our approach focuses on leveraging the geometry within multi-view blurry images to reconstruct a high-quality 3D scene. + +# 3 Method + +BAD-Gaussians aims to recover a sharp 3D scene representation by jointly learning the camera motion trajectories and Gaussians parameters, given a sequence of motion-blurred images, along with their inaccurate poses and sparse point clouds estimated from COLMAP [38], as shown in Fig. 1. This is achieved by minimizing the photometric error between the input blurred images and the synthesized blurred images generated based on the physical motion blur image formation model. We will deliver each content in the following sections. + +# 3.1 Preliminary: 3D Gaussian Splitting + +Following 3D-GS [14], the scene is represented by a series of 3D Gaussians. Each Gaussian, denoted as $\mathbf{G}$ , is parameterized by its mean position $\pmb{\mu} \in \mathbb{R}^3$ , 3D covariance $\pmb{\Sigma} \in \mathbb{R}^{3 \times 3}$ , opacity $\mathbf{o} \in \mathbb{R}$ and color $\mathbf{c} \in \mathbb{R}^3$ . The distribution of each scaled Gaussian is defined as: + +$$ +\mathbf {G} (\mathbf {x}) = e ^ {- \frac {1}{2} (\mathbf {x} - \boldsymbol {\mu}) ^ {\top} \boldsymbol {\Sigma} ^ {- 1} (\mathbf {x} - \boldsymbol {\mu})}. \tag {1} +$$ + +To ensure that the 3D covariance $\pmb{\Sigma}$ remains positive semi-definite, which is physically meaningful, and to reduce the optimization difficulty, 3D-GS represents $\pmb{\Sigma}$ using a scale $\mathbf{S} \in \mathbb{R}^3$ and rotation matrix $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ stored by a quaternion $\mathbf{q} \in \mathbb{R}^4$ : + +$$ +\boldsymbol {\Sigma} = \mathbf {R} \mathbf {S} \mathbf {S} ^ {T} \mathbf {R} ^ {T}. \tag {2} +$$ + +In order to enable differentiable Gaussian rasterization, 3D-GS projects 3D Gaussians to 2D from a given camera pose $\mathbf{T}_c = \{\mathbf{R}_c\in \mathbb{R}^{3\times 3},\mathbf{t}_c\in \mathbb{R}^3\}$ for rasterizing and rendering using the following equation, as described in [53]: + +$$ +\boldsymbol {\Sigma} ^ {\prime} = \mathbf {J} \mathbf {R} _ {c} \boldsymbol {\Sigma} \mathbf {R} _ {c} ^ {T} \mathbf {J} ^ {T}, \tag {3} +$$ + +where $\pmb{\Sigma}^{\prime}\in \mathbb{R}^{2\times 2}$ is the 2D covariance matrix, $\mathbf{J}\in \mathbb{R}^{2\times 3}$ is the Jacobian of the affine approximation of the projective transformation. Afterward, each pixel color is rendered by rasterizing these $N$ sorted 2D Gaussians based on their depths, following the formulation: + +$$ +\mathbf {C} = \sum_ {i} ^ {N} \mathbf {c} _ {i} \alpha_ {i} \prod_ {j} ^ {i - 1} (1 - \alpha_ {j}), \tag {4} +$$ + +where $\mathbf{c}_i$ is the learnable color of each Gaussian, and $\alpha_{i}$ is the alpha value computed by evaluating a 2D covariance $\pmb{\Sigma}^{\prime}$ multiplied with the learned Gaussian opacity $\mathbf{o}$ : + +$$ +\alpha_ {i} = \mathbf {o} _ {i} \cdot \exp (- \sigma_ {i}), \quad \sigma_ {i} = \frac {1}{2} \Delta_ {i} ^ {T} \boldsymbol {\Sigma} ^ {\prime - 1} \Delta_ {i}, \tag {5} +$$ + +where $\Delta \in \mathbb{R}^2$ is the offset between the pixel center and the 2D Gaussian center. + +The derivations presented above demonstrate that the rendered pixel color, denoted as $\mathbf{C}$ in Eq. (4), is a function that is differentiable with respect to all of the learnable Gaussians $\mathbf{G}$ , and the camera poses $\mathbf{T}_c$ . This facilitates our bundle adjustment formulation, accommodating a set of motion-blurred images and inaccurate camera poses within the framework of 3D-GS. + +# 3.2 Physical Motion Blur Image Formation Model + +The physical process of image formation in a digital camera encompasses the gathering of photons during the exposure period, followed by their conversion into measurable electric charges. Mathematically representing this phenomenon involves integrating across a sequence of simulated virtual latent sharp images, as follows: + +$$ +\mathbf {B} (\mathbf {u}) = \phi \int_ {0} ^ {\tau} \mathbf {C} _ {\mathrm {t}} (\mathbf {u}) \mathrm {d t}, \tag {6} +$$ + +where $\mathbf{B}(\mathbf{u})\in \mathbb{R}^{\mathrm{H}\times \mathrm{W}\times 3}$ denotes the real captured motion-blurred image, $\mathbf{u}\in \mathbb{R}^2$ represents the pixel location in the image with height H and width W, $\phi$ serves as a normalization factor, $\tau$ is the camera exposure time, $\mathbf{C}_{\mathrm{t}}(\mathbf{u})\in \mathbb{R}^{\mathrm{H}\times \mathrm{W}\times 3}$ is the virtual latent sharp image captured at timestamp t $\in [0,\tau ]$ within the + +exposure time. The blurred image $\mathbf{B}(\mathbf{u})$ , resulting from camera motion during the exposure time, is calculated by averaging all the virtual images $\mathbf{C}_{\mathrm{t}}(\mathbf{u})$ across different timestamps $t$ . This discrete approximated of this model is dependent on the number $n$ of discrete samples, as denoted: + +$$ +\mathbf {B} (\mathbf {u}) \approx \frac {1}{n} \sum_ {i = 0} ^ {n - 1} \mathbf {C} _ {i} (\mathbf {u}). \tag {7} +$$ + +The level of motion blur within an image is contingent upon the movement of the camera during the exposure time. For instance, a swiftly moving camera results in minimal relative motion, particularly with shorter exposure times, while a slowly moving camera yields motion-blurred images, especially in low-light scenarios with prolonged exposure times. Additionally, it can be deduced that $\mathbf{B}(\mathbf{u})$ demonstrates differentiability concerning each virtual sharp image $\mathbf{C}_i(\mathbf{u})$ . + +# 3.3 Camera Motion Trajectory Modeling in 3D-GS + +Based on Eq. (7), a straightforward approach to address motion-blurred images involves determining each virtual sharp image $\mathbf{C}_i$ , which serves as the dependent variable of the motion-blur image $\mathbf{B}$ . Given that a sharp image $\mathbf{C}_i$ can be rendered from a specified camera pose $\mathbf{T}_i$ within the framework of 3D-GS (i.e. $\mathbf{G}_{\theta}$ ) [14], establishing a one-to-one correspondence between poses and virtual sharp images is feasible. Consequently, we formulate the corresponding poses of each latent sharp image within the exposure time $\tau$ by employing a camera motion trajectory represented through linear interpolation between two camera poses, one at the beginning of the exposure $\mathbf{T}_{\mathrm{start}} \in \mathbf{SE}(3)$ and the other at the end $\mathbf{T}_{\mathrm{end}} \in \mathbf{SE}(3)$ . The virtual camera pose at time $t \in [0, \tau]$ can thus be expressed as follows: + +$$ +\mathbf {T} _ {t} = \mathbf {T} _ {\text {s t a r t}} \cdot \exp \left(\frac {t}{\tau} \cdot \log \left(\mathbf {T} _ {\text {s t a r t}} ^ {- 1} \cdot \mathbf {T} _ {\text {e n d}}\right)\right), \tag {8} +$$ + +where $\tau$ represents the exposure time. $\frac{t}{\tau}$ can be further discretized and derived as $\frac{i}{n-1}$ for the $i^{th}$ sampled virtual sharp image (i.e. $C_i$ ) in Eq. (7) (i.e. with pose denoted as $\mathbf{T}_i$ ), when there are $n$ sampled images in total. It follows that $\mathbf{T}_i$ is differentiable with respect to both $\mathbf{T}_{\mathrm{start}}$ and $\mathbf{T}_{\mathrm{end}}$ . We refer to prior works [26, 47] for a comprehensive understanding of the interpolation and derivations of the related Jacobian. [26]. The objective of BAD-Gaussians is to estimate both $\mathbf{T}_{\mathrm{start}}$ and $\mathbf{T}_{\mathrm{end}}$ for each frame, along with the learnable parameters of Gaussians $\mathbf{G}_{\theta}$ . + +# 3.4 Loss Function + +From a collection of $K$ motion-blurred images, we can proceed to estimate both the learnable parameters $\pmb{\theta}$ (i.e. mean position $\pmb{\mu}$ , 3D covariance $\pmb{\Sigma}$ , opacity $\mathbf{o}$ ) + +Table 1: Ablation studies on the number of virtual camera poses $n$ . The results indicate that performance reaches a saturation point with increasing number $n$ . + +
nCozy2roomTanabata
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
331.63.9389.081025.43.8215.2287
432.99.9552.055827.08.8704.1764
533.68.9595.045828.39.9011.1400
633.86.9627.036829.57.9214.1101
834.48.9654.031531.15.9440.0696
1034.68.9521.025832.12.9481.0464
1534.78.9679.025733.58.9652.0198
2034.63.9646.030733.87.9668.0151
+ +Table 2: Ablation studies on the effect of trajectory representations. We denote Deblur-NeRF-S and Deblur-NeRF-R as the synthetic and real data from Deblur-NeRF, respectively. The results demonstrate that cubic interpolation improves performance in scenes with complex camera trajectories (i.e. $MBA - VO$ and Deblur-NeRF-R). + +
Deblur-NeRF-SMBA-VODeblur-NeRF-R
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Linear Interpolation33.92.9467.042230.96.8868.149124.77.7778.1291
Cubic B-spline33.24.9395.047331.06.8883.145426.11.8157.1067
+ +and color $\mathbf{c}$ ) of 3D-GS and the camera motion trajectories (i.e. $\mathbf{T}_{\mathrm{start}}$ and $\mathbf{T}_{\mathrm{end}}$ ) for each image. This estimation is accomplished by minimizing the following loss function, which includes an $\mathcal{L}_1$ loss and a D-SSIM term between $\mathbf{B}_k(\mathbf{u})$ , the $k^{th}$ blurry image synthesized from 3D-GS using the aforementioned image formation model (i.e. Eq. (7)), and $\mathbf{B}_k^{gt}(\mathbf{u})$ , the corresponding real captured blurry image: + +$$ +\mathcal {L} = (1 - \lambda) \mathcal {L} _ {1} + \lambda \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}}. \tag {9} +$$ + +To optimize the learnable Gaussians parameter $\pmb{\theta}$ , and camera pose $\mathbf{T}$ (i.e. $\mathbf{T}_{\mathrm{start}}$ and $\mathbf{T}_{\mathrm{end}}$ in our model) for each image, it is necessary to derive the corresponding Jacobians to complete the gradient flow: + +$$ +\frac {\partial \mathcal {L}}{\partial \boldsymbol {\theta}} = \sum_ {k = 0} ^ {K - 1} \frac {\partial \mathcal {L}}{\partial \mathbf {B} _ {k}} \cdot \frac {1}{n} \sum_ {i = 0} ^ {n - 1} \frac {\partial \mathbf {B} _ {k}}{\partial \mathbf {C} _ {i}} \frac {\partial \mathbf {C} _ {i}}{\partial \boldsymbol {\theta}}, \tag {10} +$$ + +$$ +\frac {\partial \mathcal {L}}{\partial \mathbf {T}} = \sum_ {k = 0} ^ {K - 1} \frac {\partial \mathcal {L}}{\partial \mathbf {B} _ {k}} \cdot \frac {1}{n} \sum_ {i = 0} ^ {n - 1} \frac {\partial \mathbf {B} _ {k}}{\partial \mathbf {C} _ {i}} \frac {\partial \mathbf {C} _ {i}}{\partial \boldsymbol {\theta}} \frac {\partial \boldsymbol {\theta}}{\partial \mathbf {T}}, \tag {11} +$$ + +where $\mathbf{B}_k(\mathbf{u})$ and $\mathbf{C}_i(\mathbf{u})$ are denoted as $\mathbf{B}_k$ and $\mathbf{C}_i$ for simplification, and we parameterize both $\mathbf{T}_{\mathrm{start}}$ and $\mathbf{T}_{\mathrm{end}}$ with their corresponding Lie algebras of $\mathbf{SE}(3)$ , which can be represented by a 6D vector respectively. For further details on the Jacobian of Gaussians to camera pose, $\frac{\partial\pmb{\theta}}{\partial\mathbf{T}}$ , please refer to our supplemental materials. + +Table 3: Quantitative comparisons for deblurring on the synthetic dataset of Deblur-NeRF [27], referred to as DB-NeRF in the table due to space constraints. Notably, DB-NeRF* and DP-NeRF* are trained with ground-truth poses, while the others are trained with poses estimated by COLMAP [38]. The experiments highlight the superior performance of our method over previous approaches. Additionally, the results demonstrate the sensitivity of Deblur-NeRF and DP-NeRF to pose accuracy. Regarding rendering efficiency, our method achieves over 200 FPS, whereas Deblur-NeRF, DP-NeRF, BAD-NeRF and ExBluRF fall below 1 FPS. Our method takes about 30 minutes to train, while other methods take more than 10 hours. Each color shading indicates the best and second-best result, respectively. + +
CozyroomFactoryPoolTanabataTrolley
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
NeRF [30]26.13.7886.248422.02.5581.455029.90.7901.259520.57.5653.458421.71.6413.3814
3D-GS [14]25.86.7908.227021.73.5623.450329.37.7816.256120.51.5773.427821.65.6587.3614
MPR [52]29.90.8862.091525.07.6994.240933.28.8938.129022.60.7203.250726.24.8356.1762
SRN [44]29.47.8759.095026.54.7604.240432.94.8847.104523.19.7274.243825.36.8119.1618
DB-NeRF [27]29.53.8786.087925.85.7651.234031.07.8477.139723.20.7243.247125.68.8111.1683
DB-NeRF*30.26.8933.079126.40.7991.219132.30.8755.134524.56.7749.216626.24.8254.1671
DP-NeRF [19]29.02.8761.077325.42.7559.213930.48.8402.112723.93.7540.203826.17.8269.1359
DP-NeRF*30.77.9020.058427.69.8328.184733.22.8922.095425.27.7973.177926.99.8413.1312
BAD-NeRF [47]32.11.9137.051432.18.9105.118932.22.8680.090825.80.8047.156329.68.8954.0752
ExBluRF [20]28.460.87070.111329.380.85110.164631.660.86120.171925.430.80060.22725.70.81120.234
Ours34.68.9521.025831.88.9270.095236.95.9434.022532.12.9481.046433.97.9628.0209
+ +Table 4: Quantitative comparisons for novel view synthesis on the synthetic dataset of Deblur-NeRF [27], referred to as DB-NeRF in the table due to space constraints. The results demonstrate Our methods outperform previous state-of-the-art approaches, delivering the best performance across the board. + +
CozyroomFactoryPoolTanabataTrolley
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [30]25.45.7734.251422.51.5672.419929.23.7711.275620.74.5774.434121.09.6157.3924
3D-GS [14]25.24.7789.208421.34.5532.437128.61.7562.261020.37.5755.395820.86.6167.3701
MPR [52]+ [14]29.00.8685.089423.38.6848.237429.04.8170.176122.53.7349.216226.27.8352.1694
SRN [44]+ [14]28.64.8662.090325.45.7740.204529.97.8207.165023.09.7456.214124.89.8103.1666
DB-NeRF [27]29.09.8718.093725.19.7359.231530.97.8401.160023.71.7483.236924.46.7758.2215
DB-NeRF*29.88.8901.074726.06.8023.210630.94.8399.169424.82.7861.204525.78.8122.1797
DP-NeRF [19]29.56.8831.068327.91.8317.183330.98.8454.126524.73.7867.186525.62.8123.1606
DP-NeRF*30.16.8958.060828.43.8420.177131.78.8623.130625.37.8046.170126.44.8307.1424
BAD-NeRF [47]31.07.9027.055131.71.9038.120430.95.8389.110725.42.8011.156628.60.8806.0843
Ours32.35.9278.034028.18.9164.099933.30.8964.055131.00.9427.047831.44.9443.0345
+ +# 4 Experiments + +# 4.1 Experimental Settings + +Implementation Details. We implemented our method within the Nerfstudio [43] framework based on their Splatifto implementation. Camera pose interpolation and optimization on the SE(3) manifold is implemented with pypose [45]. Both the optimization of Gaussians and camera poses are performed using the Adam optimizer. The number of virtual camera poses (i.e., $n$ in Equation (7)) is set to 10, considering the trade-off between performance and efficiency. We initialize camera poses and Gaussians using estimations obtained + +Table 5: Quantitative comparisons for deblurring on the synthetic dataset of MBA-VO [27]. The results demonstrate that our method achieves the best performance even when the camera undergoes acceleration + +
ArchViz-lowArchViz-highAverage
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
NeRF [30]26.26.7887.377523.29.7098.491024.78.7493.4343
3D-GS [14]26.54.8113.338523.43.7243.461424.99.7678.4000
MPR [52]29.60.8757.210325.04.7711.357627.32.8234.2840
SRN [44]30.15.8814.170327.07.8190.279628.61.8502.2250
DB-NeRF [27]28.38.8484.179225.71.7762.322027.05.8123.2506
DB-NeRF*29.65.8744.176426.44.8010.317228.05.8377.2468
DP-NeRF [19]28.24.8490.161525.74.7839.291326.99.8165.2264
DP-NeRF*28.57.8481.151126.80.8081.273627.69.8281.2124
BAD-NeRF [47]30.51.8749.165427.54.8109.267929.03.8429.2167
ExBluRF [20]28.4.8372.251828.43.8518.172128.42.8445.2120
Ours32.28.9167.113429.64.8568.184730.96.8868.1491
+ +from COLMAP [38]. All experiments are conducted on an NVIDIA RTX 4090 GPU. + +Benchmark Datasets. We evaluate the performance of our method using both synthetic and real datasets provided by Deblur-NeRF [27] and BAD-NeRF [47]. Deblur-NeRF [47] contains five scenes, where the blurred images are generated via Blender [7] by averaging sharp virtual images captured during the exposure time, under the assumption of consistent velocity camera motion. For our evaluation, we utilize the dataset from BAD-NeRF [47], which expands upon the number of virtual sharp images to create more realistic motion-blurred images while keeping other settings consistent with [27]. Additionally, Deblur-NeRF [27] captured a real motion-blurred dataset by intentionally shaking a handheld camera during exposure. + +To enhance the evaluation of our method on severely motion-blurred images, we incorporate a dataset tailored for motion blur-aware visual odometry benchmarking (i.e. MBA-VO [26]). Different from the assumption of constant velocity in the Deblur-NeRF dataset [27], the blur images synthesized from MBA-VO [26] are based on real camera motion trajectories obtained from the ETH3D dataset [39], which do not exhibit constant velocity and include accelerations, thus presenting notable challenges. + +Baselines and Evaluation Metrics. We conduct a comparative deblurring analysis between our method and state-of-the-art learning-based single image deblurring algorithms including SRN [44] and MPR [52]. Additionally, we include evaluations against closely related approaches such as Deblur-NeRF [27], DP-NeRF [19], BAD-NeRF [47] and ExBluRF [20]. To evaluate deblurring performance, we render high-quality images corresponding to the midpoint (i.e. $\mathbf{C}_{\frac{\tau}{2}}$ in Eq. (6)) of exposure time of each training image from the optimized Gaussians. + +To facilitate novel view synthesis using SRN [44] and MPR [52], we use the deblurred images obtained from pre-trained models as inputs for training 3D-GS, as these single deblurring methods are not optimized for synthesizing novel images. During the training stage, all poses are estimated using COLMAP [38]. Additionally, we train Deblur-NeRF [27] and DP-NeRF [19] with ground-truth camera poses from Blender to assess the impact of inaccurate pose estimation on + +![](images/d53905ffe7f5b311e3e36b30c4669e13d288c016d4d132e72dd32a34da914a34.jpg) +Fig. 2: Qualitative novel view synthesis results of different methods with synthetic datasets. Despite being trained with ground truth poses $(^{*})$ , BAD-Gaussians outperforms Deblur-NeRF* and DP-NeRF* in recovering high-quality scenes from motion-blurred images with inaccurate camera poses, showcasing its superior performance. + +these two methods. The quality of the rendered sharp image is evaluated using PSNR, SSIM, and LPIPS metrics. + +For pose accuracy evaluation, we compare the Absolute Trajectory Error (ATE) metric, against classical structure-from-motion framework COLMAP [38] and BAD-NeRF [47]. + +# 4.2 Ablation Study + +Virtual Camera Poses. We experimented to investigate the effect of the number of interpolated virtual camera poses within the exposure time, as described by $n$ in Eq. (7). Two of the five synthetic scenes provided by Deblur-NeRF [27] were selected, representing sequences with minor and severe motion blur (i.e. Cozy2room and Tanabata), respectively. We varied the number, $n$ , from 3 to 20 in our study, and the rendering metric results are presented in Table 1. Our experimental results demonstrate that increasing the number of interpolated virtual camera poses helps in addressing severe motion blur. However, marginal improvements are observed for minor blur instances. Based on our experiments, we choose $n = 10$ interpolated virtual images to strike a balance between ren + +dering performance and training efficiency (larger $n$ means more computational resources). + +Trajectory Representations. To assess the impact of different trajectory representations, we conduct two experiments: one involves optimizing $\mathbf{T}_{\mathrm{start}}$ and $\mathbf{T}_{\mathrm{end}}$ to depict a linear trajectory, while the other utilizes a higher-order spline (i.e., cubic B-spline) that jointly optimizes four control knots $\mathbf{T}_1$ , $\mathbf{T}_2$ , $\mathbf{T}_3$ , and $\mathbf{T}_4$ to capture more complex camera motions. We refer [46] for more details about cubic B-spline. The average quantitative results on the synthetic datasets of Deblur-NeRF-S [27] (Cozy2room, Factory, Pool, Tanabata and Trolley) and MBA-VO [26] (ArchViz-low and ArchViz-high), along with real datasets, Deblur-NeRF-S, which includes 10 real captured scenes from Deblur-NeRF, are presented in Table 2. In synthetic scenes, linear interpolation demonstrates comparable performance to cubic B-spline interpolation. However, in real captured scenes, cubic B-spline interpolation outperforms linear interpolation, particularly due to the longer exposure time during image capture. The effectiveness of cubic B-spline interpolation in real scenes can be attributed to its ability to better model the nuances of camera motion over longer time intervals. Conversely, linear interpolation is sufficient to accurately represent camera motion trajectories within shorter time intervals, as observed in synthetic scenes. Combining the training efficiency and rendering quality, we employ linear interpolation in synthetic datasets and cubic B-spline in real data. + +
PSNR↑SSIM↑LPIPS↓Ttrain(h)
NeRF [30]22.69.6347.36872.5
3DGS [14]22.72.6390.28870.3
DB-NeRF [27]25.63.7675.1820>5
DP-NeRF [19]25.91.7734.1589>5
BAD-NeRF [47]24.33.7003.2196>5
ExBluRF [20]25.31.7293.20844.0
Ours-Cubic26.11.8157.10640.5
+ +Table 6: Quantitative comparisons for novel view synthesis on the real captured dataset of Deblur-NeRF [27]. Full results can be found in our supplementary material. + +# 4.3 Results + +Results on Synthetic Data. We evaluate our approach against baseline methods using synthetic datasets obtained from both Deblur-NeRF [27] and MBAVO [26]. The quantitative evaluation results on deblurring and novel view synthesis with the Deblur-NeRF dataset are presented in Table 3 and Table 4 respectively. All metrics (i.e. PSNR, SSIM, LPIPS) demonstrate substantial improvements over prior state-of-the-art methods (i.e. 3.6 and $1.7\mathrm{dB}$ higher than the second best method on average in Table 3 and Table 4, respectively). + +The results demonstrate that both NeRF [30] and 3D-GS [14] are suffering from motion-blurred images, which motivates the necessity of our method. + +![](images/43b01f0e8fcc0343e2736b6dce4c20139c2f7d63ce030a18bd2fc21398670e3e.jpg) +Fig. 3: Qualitative novel view synthesis results of different methods with the real datasets. The experimental results demonstrate that our method achieves superior performance over prior methods on the real dataset as well. In contrast, BAD-NeRF yields poorer results when applied to real data and exhibits satisfactory performance only within synthetic datasets. + +Single-stage methods such as MPR [52] and SRN [44] fall short of matching the performance of our approach due to their limited utilization of geometric information across multi-view images. Additionally, our method surpasses Deblur-NeRF [27] and DP-NeRF [19], partly because they fail to optimize the inaccurate camera poses estimated from COLMAP during training. We additionally trained Deblur-NeRF and DP-NeRF with ground truth poses (denoted as *), and the results demonstrate improved performance, reflecting the sensitivity to the pose accuracies of both Deblur-NeRF and DP-NeRF. A notable result from both Table 3 and Table 4 is that our method performs worse than BAD-NeRF on the Factory sequence. We find that it is caused by the inferior capability to represent the sky by Gaussian splatting compared to NeRF. Nevertheless, our method still performs better than all prior methods with a large margin on average. + +Qualitative results are presented in Fig. 2. It demonstrates that our method effectively attains high-quality scene representations with intricate details, being trained from a series of blurred images. In Fig. 2, it is evident that Deblur-NeRF [27] and DP-NeRF [19] encounter difficulties in modeling regions with significant depth variations, attributable to their method of synthesizing blurred images through convolution with a blur kernel applied to rendered images. The physically based method BAD-NeRF [47] exhibits superior performance overall, + +yet it still presents some deficiencies, particularly noticeable around areas with significant color and depth variation. It demonstrates the effectiveness of the explicit Gaussian splatting representation over the implicit neural representation. + +We further assess the performance of our methods using a dataset characterized by severe motion blur and camera movements with varying velocities, sourced from MBA-VO [26]. The results presented in Table 5 demonstrate that our method achieves the best performance even when the camera undergoes acceleration. + +Results on Real Data. The performance of novel view synthesis on real data sourced from Deblur-NeRF [27] are also evaluated. The quantitative results, as presented in Table 6, demonstrate the superior performance of our method compared to other approaches. Furthermore, the qualitative results illustrated in Fig. 3 vividly demonstrate our method's ability to deliver intricate details. + +Pose Estimation. Due to the unknown metric scale, we align the estimated trajectories against their corresponding ground truth for the computation of the absolute trajectory error metric. The experimental results presented in Table 7 demonstrate the effectiveness of our approach in recovering precise camera poses. + +Table 7: Pose estimation performance of BAD-Gaussians on various blur sequences. The results are in the absolute trajectory error metric in centimeters (ATE/cm). The COLMAP-sharp / COLMAP-blur represents the result of COLMAP with sharp/blurry images respectively. + +
Cozy2roomFactoryPoolTanabataTrolleyArchViz-lowArchViz-highAverage
COLMAP-sharp [38].133±.079.101±.052.172±.047.277±.103.173±.079.148±.078.115±.066.1599
COLMAP-blur [38].287±.1611.69±.1.24.633±.2361.59±.1.06.912±.559.536±.309.766±.391.9163
BAD-NeRF [47].316±.101.600±.208.095±.0461.34±.543.122±.0701.24±.6352.18±1.54.8419
Ours.165±.057.395±.426.264±.0981.20±.621.134±.070.520±.5911.41±1.11.5840
+ +# 5 Conclusion + +We presented the first pipeline to learn Gaussian splittings from a set of motion-blurred images with inaccurate camera poses. Our pipeline can jointly optimize the 3D scene representation and camera motion trajectories. Extensive experimental evaluations demonstrate that our method can deliver high-quality novel view images, and achieve real-time rendering compared to prior state-of-the-art works. + +# Acknowledgements + +This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation. + +# References + +1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://jonbarron.info/mipnerf/2, 3 +2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://jonbarron.info/mipnerf360/3 +3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2023), https://jonbarron.info/zipnerf/ 3 +4. Cao, A., Johnson, J.: HexPlane: A Fast Representation for Dynamic Scenes. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://caoang327.github.io/HexPlane/ 4 +5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensorRF: Tensorial Radiance Fields. In: European Conference on Computer Vision (ECCV) (2022), https://apchenstu.github.io/TensoRF/ 2, 3, 4 +6. Cho, S., Lee, S.: Fast Motion Deblurring. ACM Transactions on Graphics (TOG) 28(5) (2009) 4 +7. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org 10 +8. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised NeRF: Fewer Views and Faster Training for Free. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://www.cs.cmu.edu/~dsnerf/3 +9. Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T.: Removing camera shake from a single photograph. In: ACM Transactions on Graphics (TOG) (2006) 4 +0. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://sarafridov.github.io/K-Planes/4 +1. Fu, Y., Liu, S., Kulkarni, A., Kautz, J., Efros, A.A., Wang, X.: Colmap-Free 3D Gaussian Splatting. arXiv preprint arXiv:2312.07504 (2023), https://oasisyang.github.io/colmap-free-3dgs/2 +2. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: FastNeRF: High-Fidelity Neural Rendering at 200FPS. In: International Conference on Computer Vision (ICCV) (2021), https://microsoft.github.io/FastNeRF/2, 3 +3. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-Calibrating Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://postech-cvlab.github.io/SCNeRF/ 3, 4 +4. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (TOG) 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ 2, 4, 5, 7, 9, 10, 12, 13 +5. Kim, M., Seo, S., Han, B.: InfoNeRF: Ray Entropy Minimization for Few-Shot Neural Volume Rendering. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://cv.snu.ac.kr/research/InfoNeRF/3 + +16. Krishnan, D., Fergus, R.: Fast image deconvolution using Hyper-Laplacian priors. In: Neural Information Processing Systems (NIPS) (2009) 4 +17. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better. In: International Conference on Computer Vision (ICCV) (2019), https://github.com/VITA-Group/DeblurGANv2 4 +18. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://github.com/facebookresearch/pytorch3d/ 3 +19. Lee, D., Lee, M., Shin, C., Lee, S.: DP-NeRF: Deblurred Neural Radiance Field With Physical Scene Priors. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://dogyoonlee.github.io/dpnerf/2, 3, 4, 9, 10, 11, 12, 13 +20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: ExBluRF: Efficient Radiance Fields for Extreme Motion Blurred Images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023) 4, 9, 10, 12 +21. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithm. In: Computer Vision and Pattern Recognition (CVPR) (2009) 4 +22. Levoy, M.: Efficient Ray Tracing of Volume Data. ACM Transactions on Graphics (TOG) 9(3), 245-261 (1990) 2 +23. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields. In: International Conference on Learning Representations (ICLR) (2024), https://moyangli00.github.io/usb_nerf/ 3 +24. Li, W., Wan, P., Wang, P., Li, J., Zhou, Y., Liu, P.: BeNeRF: Neural Radiance Fields from a Single Blurry Image and Event Stream. In: European Conference on Computer Vision (ECCV). Springer (2024), https://github.com/WU-CVGL/BeNeRF 3 +25. Lin, Chen-Hsuan and Ma, Wei-Chiu and Torralba, Antonio and Lucey, Simon: BARF: Bundle-Adjusting Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://chenhsuanlin.bitbucket.io/bundle-adjusting-NeRF 3, 4 +26. Liu, P., Zuo, X., Larsson, V., Pollefeys, M.: MBA-VO: Motion Blur Aware Visual Odometry. In: International Conference on Computer Vision (ICCV) (2021), https://github.com/ethliup/MBA-VO 7, 10, 12, 14 +27. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural Radiance Fields from Blurry Images. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://limacv.github.io/deblurnerf/ 2, 3, 4, 9, 10, 11, 12, 13, 14 +28. Max, N.: Optical Models for Direct Volume Rendering. ACM Transactions on Graphics (TOG) 1(2), 99-108 (1995) 2 +29. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://bmild.github.io/rawnerf/3 +30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: European Conference on Computer Vision (ECCV) (2020), https://www.matthewtancik.com/nerf 2, 3, 9, 10, 12 +31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics (TOG) 41(4), 102:1-102:15 (Jul 2022), https://nvlabs.github.io/instant-ngp/2, 3, 4 + +32. Nah, S., Kim, T.H., Lee, K.M.: Deep Multi-Scale Convolutional Neural Network for Dynamic Scene Deblurring. In: Computer Vision and Pattern Recognition (CVPR) (2017), https://github.com/SeungjunNah/DeepDeblur_release 4 +33. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse Inputs. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://m-niemeyer.github.io/regnerf/3 +34. Park, H., Lee, K.M.: Joint Estimation of Camera Pose, Depth, Deblurring, and Super-Resolution from a Blurred Image Sequence. In: International Conference on Computer Vision (ICCV). vol. 2017-Octob (2017). https://doi.org/10.1109/ICCV.2017.494 4 +35. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: CamP: Camera Preconditioning for Neural Radiance Fields. ACM Transactions on Graphics (TOG) (2023), https://camp-nerf.github.io/3, 4 +36. Piala, M., Clark, R.: TermiNeRF: Ray Termination Prediction for Efficient Neural Rendering. In: International Conference on 3D Vision (3DV) (2021), https://projects.mackopes.com/terminerf/3 +37. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance Fields without Neural Networks. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://alexyu.net/plenoxels/ 2, 3, 4 +38. Schonberger, J.L., Frahm, J.M.: Structure-from-motion Revisited. In: Computer Vision and Pattern Recognition (CVPR) (2016), https://github.com/colmap/ colmap 2, 5, 9, 10, 11, 14 +39. Schops, T., Sattler, T., Pollefeys, M.: BAD SLAM: Bundle Adjusted Direct RGB-D SLAM. In: Computer Vision and Pattern Recognition (CVPR) (2019), https://github.com/ETH3D/badslam 10 +40. Shan, Q., Jia, J., A., A.: High-quality Motion Deblurring from a Single Image. ACM Transactions on Graphics (TOG) 27(3) (2008) 4 +41. Su, S., Delbracio, M., Wang, J.: Deep Video Deblurring for Hand-held Cameras. In: Computer Vision and Pattern Recognition (CVPR) (2017), https://github.com/shuochsu/DeepVideoDeblurring 4 +42. Sun, C., Sun, M., Chen, H.: Direct Voxel Grid Optimization: Super-fast Convergence for Radiance Fields Reconstruction. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://sunset1995.github.io/dvgo/3 +43. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A Modular Framework for Neural Radiance Field Development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023), https://github.com/nerfstudio-project/nerfstudio/9 +44. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Computer Vision and Pattern Recognition (CVPR) (2018), https://github.com/jiangsutx/SRN-Deblur 4, 9, 10, 13 +45. Wang, C., Gao, D., Xu, K., Geng, J., Hu, Y., Qiu, Y., Li, B., Yang, F., Moon, B., Pandey, A., Aryan, Xu, J., Wu, T., He, H., Huang, D., Ren, Z., Zhao, S., Fu, T., Reddy, P., Lin, X., Wang, W., Shi, J., Talak, R., Cao, K., Du, Y., Wang, H., Yu, H., Wang, S., Chen, S., Kashyap, A., Bandaru, R., Dantu, K., Wu, J., Xie, L., Carlone, L., Hutter, M., Scherer, S.: PyPose: A library for robot learning with physics-based optimization. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023), https://github.com/pypyse/pypyse 9 + +46. Wang, P., Zhao, L., Ma, R., Liu, P.: Supplementary Material for BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields https://openaccess.thecvf.com/content/CVPR2023/supplemental/Wang_BAD-NeRF_Bundle_Adjusted_CVPR_2023SUPPLEMENTAL.pdf 12 +47. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://wangpeng000.github.io/BAD-NeRF/2, 3, 4, 7, 9, 10, 11, 12, 13, 14 +48. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021), https://nerfmm.active.vision/3 +49. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: NeX: Real-time View Synthesis with Neural Basis Expansion. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://nex-mpi.github.io/3 +50. Xu, L., Jia, J.: Two-Phase Kernel Estimation for robust motion deblurring. In: European Conference on Computer Vision (ECCV) (2010) 4 +51. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for Real-time Rendering of Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://alexyu.net/plenoctrees/ 2, 3 +52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage Progressive Image Restoration. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://github.com/swz30/MPRNet 9, 10, 13 +53. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: EWA Volume Splatting. In: Proceedings Visualization, 2001. VIS'01. pp. 29-538. IEEE (2001), https://cgl.ethz.ch/research/past_projects/surfels/ewavolumesplatting/index.html 6 \ No newline at end of file diff --git a/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/images.zip b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fe947a4886f8e928fdefb49a35e75a79555bfd22 --- /dev/null +++ b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d3c817ea0ab8059b8b6deb18f7b892d2edeb19e88cc9c4252903c8cf6eeb1d5 +size 743532 diff --git a/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/layout.json b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b945a370b6e74a0a4714eda9707bd716314d10f1 --- /dev/null +++ b/2024/BAD-Gaussians_ Bundle Adjusted Deblur Gaussian Splatting/layout.json @@ -0,0 +1,9013 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 111, + 455, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 111, + 455, + 149 + ], + "spans": [ + { + "bbox": [ + 162, + 111, + 455, + 149 + ], + "type": "text", + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 182, + 167, + 430, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 167, + 430, + 181 + ], + "spans": [ + { + "bbox": [ + 182, + 167, + 430, + 181 + ], + "type": "text", + "content": "Lingzhe Zhao" + }, + { + "bbox": [ + 182, + 167, + 430, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 182, + 167, + 430, + 181 + ], + "type": "text", + "content": ", Peng Wang" + }, + { + "bbox": [ + 182, + 167, + 430, + 181 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 182, + 167, + 430, + 181 + ], + "type": "text", + "content": ", and Peidong Liu" + }, + { + "bbox": [ + 182, + 167, + 430, + 181 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 189, + 423, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 189, + 423, + 223 + ], + "spans": [ + { + "bbox": [ + 190, + 189, + 423, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 190, + 189, + 423, + 223 + ], + "type": "text", + "content": "Westlake University " + }, + { + "bbox": [ + 190, + 189, + 423, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 190, + 189, + 423, + 223 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 190, + 189, + 423, + 223 + ], + "type": "inline_equation", + "content": "\\{" + }, + { + "bbox": [ + 190, + 189, + 423, + 223 + ], + "type": "text", + "content": "zhaolingzhe, wangpeng, liupeidong\\}@westlake.edu.cn https://github.com/wu-cvgl/BAD-Gaussians/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 254, + 455, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 254, + 455, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 254, + 455, + 495 + ], + "type": "text", + "content": "Abstract. While neural rendering has demonstrated impressive capabilities in 3D scene reconstruction and novel view synthesis, it heavily relies on high-quality sharp images and accurate camera poses. Numerous approaches have been proposed to train Neural Radiance Fields (NeRF) with motion-blurred images, commonly encountered in real-world scenarios such as low-light or long-exposure conditions. However, the implicit representation of NeRF struggles to accurately recover intricate details from severely motion-blurred images and cannot achieve real-time rendering. In contrast, recent advancements in 3D Gaussian Splatting achieve high-quality 3D scene reconstruction and real-time rendering by explicitly optimizing point clouds into 3D Gaussians. In this paper, we introduce a novel approach, named BAD-Gaussians (Bundle Adjusted Deblur Gaussian Splatting), which leverages explicit Gaussian representation and handles severe motion-blurred images with inaccurate camera poses to achieve high-quality scene reconstruction. Our method models the physical image formation process of motion-blurred images and jointly learns the parameters of Gaussians while recovering camera motion trajectories during exposure time. In our experiments, we demonstrate that BAD-Gaussians not only achieves superior rendering quality compared to previous state-of-the-art deblur neural rendering methods on both synthetic and real datasets but also enables real-time rendering capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 506, + 453, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 506, + 453, + 529 + ], + "spans": [ + { + "bbox": [ + 160, + 506, + 453, + 529 + ], + "type": "text", + "content": "Keywords: 3D Gaussian Splatting " + }, + { + "bbox": [ + 160, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 453, + 529 + ], + "type": "text", + "content": " Deblurring " + }, + { + "bbox": [ + 160, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 453, + 529 + ], + "type": "text", + "content": " Bundle Adjustment " + }, + { + "bbox": [ + 160, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 453, + 529 + ], + "type": "text", + "content": " Differentiable Rendering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 549, + 230, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 230, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 230, + 562 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "text", + "content": "Acquiring accurate 3D scene representations from 2D images has long been a challenging problem in computer vision. Serving as a fundamental component in various applications such as virtual/augmented reality and robotics navigation, substantial efforts have been dedicated to addressing this challenge over the last few decades. Among those pioneering works, Neural Radiance Fields" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 223, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 223, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 223, + 653 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 653, + 234, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "(NeRF) [30], parameterized by Multi-layer Perceptrons (MLP), stands out for its utilization of differentiable volume rendering technique [22, 28] and has garnered significant attention due to its capability to recover high-quality 3D scene representation from 2D images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 482, + 467 + ], + "type": "text", + "content": "Numerous works have focused on enhancing the performance of NeRF, particularly in terms of training [5,31,37] and rendering efficiency [12,51]. A recent advancement, 3D Gaussian Splatting (3D-GS) [14], extends the implicit neural rendering [30] to explicit point clouds. By projecting these optimized point clouds (Gaussians) onto the image plane, 3D-GS [14] achieves real-time rendering while enhancing the efficiency of NeRF in both training and rendering, and also improves rendering quality. However, both NeRF-based [1,30,31] methods and 3D-GS [14] heavily rely on well-captured sharp images and accurately precomputed camera poses, typically obtained from COLMAP [38]. Motion-blurred images, a common form of image degradation, often encountered in low-light or long-exposure conditions, can notably impair the performance of both NeRF and 3D-GS. The challenges posed by motion-blurred images to NeRF and 3D-GS can be attributed to three primary factors: (a) NeRF and 3D-GS rely on high-quality sharp images for supervision. However, motion-blurred images violate this assumption and exhibit notably inaccurate corresponding geometry between multi-view frames, thus presenting significant difficulties in achieving accurate 3D scene representation for both NeRF and 3D-GS; (b) Accurate camera poses are essential for training NeRF and 3D-GS. However, recovering accurate poses from multi-view motion-blurred images using COLMAP [38] is challenging. (c) 3D-GS necessitates sparse cloud points from COLMAP as the initialization of Gaussians. The mismatched features between multi-view blurred images and the inaccuracies in pose calibration further exacerbate the issue, leading to COLMAP producing fewer cloud points. This introduces an additional initialization issue for 3D-GS. Therefore, these factors result in a notable drop in performance for 3D-GS when dealing with motion-blurred images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 470, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 625 + ], + "type": "text", + "content": "Implicit neural representations, i.e. NeRF [30], have been employed to reconstruct sharp 3D scenes from motion-blurred images [19,27,47]. For example, Deblur-NeRF [27] introduces a deformable sparse kernel that alters a canonical kernel at spatial locations to simulate the blurring process. DP-NeRF [19] integrates physical priors derived from the motion-blurred image acquisition process into Deblur-NeRF [27] to construct a clean NeRF representation. In contrast, BAD-NeRF [47] models the physical process of capturing motion-blurred images and jointly optimizes NeRF while recovering the camera trajectory within the exposure time. However, these implicit deblur rendering methods encounter significant challenges in achieving real-time rendering and producing high-quality outputs with intricate details. Additionally, the implicit representation introduces extra difficulties in optimizing the neural parameters and camera poses as mentioned in [11]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "In order to address these challenges, we propose Bundle Adjusted Deblur Gaussian Splatting, the first motion deblur framework based on 3D-GS, which we refer to as BAD-Gaussians. We incorporate the physical process of motion" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "text", + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "blur into the training of 3D-GS, employing a spline function to characterize the trajectory within the camera's exposure time. In the training of BAD-Gaussians, the camera trajectory within exposure time is optimized using gradients derived from the Gaussians of the scene, while jointly optimizing the Gaussians themselves. Specifically, the trajectory of each motion-blurred image is represented by the initial and final poses at the beginning and end of the exposure time, respectively. By assuming the exposure time is typically short, we can interpolate between the initial and final poses to obtain every camera pose along the trajectory. From this trajectory, we generate a sequence of virtual sharp images by projecting the scene's Gaussians onto the image plane. These virtual sharp images are then averaged to synthesize the blurred images, following the physical blur process. Finally, the Gaussians along the trajectory are optimized by minimizing the photometric error between the synthesized blurred images and the input blurred images through differentiable Gaussian rasterization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 283, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 355 + ], + "type": "text", + "content": "We evaluate BAD-Gaussians using both synthetic and real datasets. The experimental results demonstrate that BAD-Gaussians outperforms prior state-of-the-art implicit neural rendering methods by explicitly incorporating the image formation process of motion-blurred images into the training of 3D-GS, achieving better rendering performance in terms of real-time rendering speed and superior rendering quality. In summary, our contributions can be outlined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 358, + 481, + 465 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 358, + 481, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 481, + 407 + ], + "type": "text", + "content": "- We introduce a photometric bundle adjustment formulation specifically designed for motion-blurred images, achieving the first real-time rendering performance from motion-blurred images within the framework of 3D Gaussian Splatting;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 407, + 481, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 430 + ], + "type": "text", + "content": "- We demonstrate how this formulation enables the acquisition of high-quality 3D scene representation from a set of motion-blurred images;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 430, + 481, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 430, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 138, + 430, + 481, + 465 + ], + "type": "text", + "content": "- Our approach successfully deblurs severe motion-blurred images, synthesizes higher-quality novel view images, and achieves real-time rendering, surpassing previous state-of-the-art implicit deblurring rendering methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 480, + 238, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 238, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 238, + 493 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 502, + 279, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 279, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 279, + 514 + ], + "type": "text", + "content": "2.1 Neural Radiance Fields" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 518, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 626 + ], + "type": "text", + "content": "NeRF [30], employing implicit MLP, exhibits remarkable performance in synthesizing high-quality novel view images and accurately representing 3D scenes. Numerous extension works have been proposed to enhance NeRF's performance, including improvements in training [5,31,37,42] and rendering [12,18,36,49,51] efficiency, as well as anti-alias capabilities [1-3]. Additionally, various methods aim to bolster NeRF's robustness against imperfect inputs, such as inaccurate camera poses [13, 25, 35, 48], few-shot images [8, 15, 33], and low-quality images [19, 23, 24, 27, 29, 47]. In the following section, we will primarily focus on reviewing methods closely related to our work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Fast Neural Rendering. Numerous approaches inspired by NeRF [30] have sought to enhance its rendering efficiency by employing advanced data structures to reconstruct radiance fields, thereby minimizing the computational cost" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "associated with implicit MLPs used in NeRF. These approaches are primarily categorized into grid-based [4, 5, 10, 37] and hash-based [31] methods. Despite these efforts, achieving real-time rendering for unbounded and complete scenes, as well as high-resolution images, remains challenging. In contrast to methods based on volume rendering and implicit representations, which may hinder fast rendering, recent advancements like 3D-GS [14] achieve real-time high-quality rendering through pure explicit point scene representation and the differentiable Gaussian rasterization. Nevertheless, these 3D scene representations heavily rely on accurately posed high-quality images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 227, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 357 + ], + "type": "text", + "content": "NeRF for Camera Optimization. The pioneering work, BARF [25], was the first to propose simultaneous optimization of camera parameters alongside NeRF. They employed a coarse-to-fine bundle adjustment strategy to enhance camera pose recovery. Concurrently, SC-NeRF [13] introduced a method to learn various camera models, encompassing both extrinsic and intrinsic parameters, along with scene representation. In the latest development, CamP [35] introduced a preconditioner to mitigate correlations between camera parameters, aiming for improved joint optimization of camera and NeRF parameters. In contrast to the aforementioned works [13,25,35], which focus solely on optimizing camera poses for sharp images, our approach goes further by recovering the trajectory of each blurred image within the exposure time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 482, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 531 + ], + "type": "text", + "content": "NeRF for Deblurring. Several scene deblurring methods based on NeRF have been proposed, including Deblur-NeRF [27] and DP-NeRF [19], which reconstruct sharp scene representations from sets of motion-blurred images. However, these methods fix inaccurate camera poses recovered from blurred images during training, leading to degradation in reconstruction performance. BAD-NeRF [47] and ExBluRF [20] jointly learns camera motion trajectories within exposure time and radiance fields, following the physical blur process. Despite these advancements, existing deblurring neural rendering strategies struggle to achieve real-time rendering and reconstruct intricate scene details due to the implicit MLP structure, necessitating significant improvements in rendering efficiency and quality. Moreover, optimizing the 3D scenes jointly with camera poses faces additional challenges due to NeRF's implicit representations. To address these issues, we propose achieving deblurring capability within the framework of Gaussian Splatting [14]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 548, + 251, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 251, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 251, + 561 + ], + "type": "text", + "content": "2.2 Image Deblurring" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Two primary categories typically classify existing techniques for addressing the motion deblurring problem: the first involves formulating the issue as an optimization task, wherein gradient descent is employed during inference to jointly refine the blur kernel and the latent sharp image [6, 9, 16, 21, 34, 40, 50]. Another pipeline phrases the deblurring task as an end-to-end learning, particularly leveraging deep learning methodologies. With the support of substantial datasets and deep neural networks, superior results have been achieved for both single image deblurring [17, 32, 44] and video deblurring [41]. However, these 2D deblurring" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "text", + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 116, + 480, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 247 + ], + "type": "image", + "image_path": "15396aca0f0f42d1d32043761dc5782fcbb89522fdb0fd0ff89603e9a3434fc8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 376 + ], + "type": "text", + "content": "Fig. 1: The pipeline of BAD-Gaussians. Our approach utilizes Gaussian representations to depict sharp 3D scenes derived from a series of motion-blurred images, along with their inaccurate poses and sparse point clouds from COLMAP. Employing forward projection and differentiable Gaussian rasterization, we jointly optimize the Gaussians in the scene and the camera trajectory within exposure time, by backpropagating gradients from Gaussians to camera poses. Following the physical process of motion blur, we model motion-blurred images by averaging the virtual sharp images captured during the exposure time. These virtual camera poses are represented and interpolated using a continuous spline within the SE(3) space. The joint optimization of Gaussians and camera trajectories is achieved by minimizing the photometric loss between synthesized and actual blurry images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 398, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 447 + ], + "type": "text", + "content": "methods cannot exploit the 3D scene geometry between multi-view images, thus failing to ensure the view consistency of the scene from different viewpoints. In contrast, our approach focuses on leveraging the geometry within multi-view blurry images to reconstruct a high-quality 3D scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 461, + 202, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 202, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 202, + 474 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 483, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 568 + ], + "type": "text", + "content": "BAD-Gaussians aims to recover a sharp 3D scene representation by jointly learning the camera motion trajectories and Gaussians parameters, given a sequence of motion-blurred images, along with their inaccurate poses and sparse point clouds estimated from COLMAP [38], as shown in Fig. 1. This is achieved by minimizing the photometric error between the input blurred images and the synthesized blurred images generated based on the physical motion blur image formation model. We will deliver each content in the following sections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 582, + 340, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 340, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 340, + 594 + ], + "type": "text", + "content": "3.1 Preliminary: 3D Gaussian Splitting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "content": "Following 3D-GS [14], the scene is represented by a series of 3D Gaussians. Each Gaussian, denoted as " + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "content": ", is parameterized by its mean position " + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "content": ", 3D covariance " + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "content": ". The distribution of each scaled Gaussian is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 650, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 650, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 242, + 650, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {G} (\\mathbf {x}) = e ^ {- \\frac {1}{2} (\\mathbf {x} - \\boldsymbol {\\mu}) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {x} - \\boldsymbol {\\mu})}. \\tag {1}", + "image_path": "b3ffbee15dd85070dc985c3d502ef44953ce2aee2f17ad8d34c00323cb846b83.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "To ensure that the 3D covariance " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " remains positive semi-definite, which is physically meaningful, and to reduce the optimization difficulty, 3D-GS represents " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " using a scale " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and rotation matrix " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " stored by a quaternion " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 274, + 163, + 480, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 163, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 274, + 163, + 480, + 176 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}. \\tag {2}", + "image_path": "bc97eac1ded28a6d7792bac6f213979632b122bd2816dadb334def8e18f9323b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 479, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 479, + 218 + ], + "type": "text", + "content": "In order to enable differentiable Gaussian rasterization, 3D-GS projects 3D Gaussians to 2D from a given camera pose " + }, + { + "bbox": [ + 130, + 182, + 479, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_c = \\{\\mathbf{R}_c\\in \\mathbb{R}^{3\\times 3},\\mathbf{t}_c\\in \\mathbb{R}^3\\}" + }, + { + "bbox": [ + 130, + 182, + 479, + 218 + ], + "type": "text", + "content": " for rasterizing and rendering using the following equation, as described in [53]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 266, + 226, + 480, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 226, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 266, + 226, + 480, + 239 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {R} _ {c} \\boldsymbol {\\Sigma} \\mathbf {R} _ {c} ^ {T} \\mathbf {J} ^ {T}, \\tag {3}", + "image_path": "cdb10158e90cefaa5a65fbefa3f544c230b30080028b51b845672c4a189d7e86.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 479, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 296 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 247, + 479, + 296 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{\\prime}\\in \\mathbb{R}^{2\\times 2}" + }, + { + "bbox": [ + 130, + 247, + 479, + 296 + ], + "type": "text", + "content": " is the 2D covariance matrix, " + }, + { + "bbox": [ + 130, + 247, + 479, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{J}\\in \\mathbb{R}^{2\\times 3}" + }, + { + "bbox": [ + 130, + 247, + 479, + 296 + ], + "type": "text", + "content": " is the Jacobian of the affine approximation of the projective transformation. Afterward, each pixel color is rendered by rasterizing these " + }, + { + "bbox": [ + 130, + 247, + 479, + 296 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 247, + 479, + 296 + ], + "type": "text", + "content": " sorted 2D Gaussians based on their depths, following the formulation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 306, + 480, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 306, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 252, + 306, + 480, + 338 + ], + "type": "interline_equation", + "content": "\\mathbf {C} = \\sum_ {i} ^ {N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {4}", + "image_path": "77a9eac65e27d845729671bec6ef7f0666a2f38dd5e1f833217a4b83f40d514c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "text", + "content": " is the learnable color of each Gaussian, and " + }, + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "text", + "content": " is the alpha value computed by evaluating a 2D covariance " + }, + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{\\prime}" + }, + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "text", + "content": " multiplied with the learned Gaussian opacity " + }, + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 346, + 479, + 383 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 220, + 392, + 480, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 220, + 392, + 480, + 413 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\mathbf {o} _ {i} \\cdot \\exp (- \\sigma_ {i}), \\quad \\sigma_ {i} = \\frac {1}{2} \\Delta_ {i} ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime - 1} \\Delta_ {i}, \\tag {5}", + "image_path": "90b5732fe645331bdb72a60a9fc4f3172eebfdd06a5b8ebd4af6df2a006917fd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 422, + 479, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 479, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 479, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 422, + 479, + 433 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 422, + 479, + 433 + ], + "type": "text", + "content": " is the offset between the pixel center and the 2D Gaussian center." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 434, + 479, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 479, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 479, + 493 + ], + "type": "text", + "content": "The derivations presented above demonstrate that the rendered pixel color, denoted as " + }, + { + "bbox": [ + 130, + 434, + 479, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 434, + 479, + 493 + ], + "type": "text", + "content": " in Eq. (4), is a function that is differentiable with respect to all of the learnable Gaussians " + }, + { + "bbox": [ + 130, + 434, + 479, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 434, + 479, + 493 + ], + "type": "text", + "content": ", and the camera poses " + }, + { + "bbox": [ + 130, + 434, + 479, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_c" + }, + { + "bbox": [ + 130, + 434, + 479, + 493 + ], + "type": "text", + "content": ". This facilitates our bundle adjustment formulation, accommodating a set of motion-blurred images and inaccurate camera poses within the framework of 3D-GS." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 511, + 394, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 511, + 394, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 511, + 394, + 523 + ], + "type": "text", + "content": "3.2 Physical Motion Blur Image Formation Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 530, + 479, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 479, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 479, + 588 + ], + "type": "text", + "content": "The physical process of image formation in a digital camera encompasses the gathering of photons during the exposure period, followed by their conversion into measurable electric charges. Mathematically representing this phenomenon involves integrating across a sequence of simulated virtual latent sharp images, as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 252, + 588, + 480, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 588, + 480, + 612 + ], + "spans": [ + { + "bbox": [ + 252, + 588, + 480, + 612 + ], + "type": "interline_equation", + "content": "\\mathbf {B} (\\mathbf {u}) = \\phi \\int_ {0} ^ {\\tau} \\mathbf {C} _ {\\mathrm {t}} (\\mathbf {u}) \\mathrm {d t}, \\tag {6}", + "image_path": "cfd6aa98a85fb45eb38fad154b69e7af1dc7fae64814e928e39e09cd12b18d1d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{B}(\\mathbf{u})\\in \\mathbb{R}^{\\mathrm{H}\\times \\mathrm{W}\\times 3}" + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "text", + "content": " denotes the real captured motion-blurred image, " + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{u}\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "text", + "content": " represents the pixel location in the image with height H and width W, " + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "text", + "content": " serves as a normalization factor, " + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "text", + "content": " is the camera exposure time, " + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{t}}(\\mathbf{u})\\in \\mathbb{R}^{\\mathrm{H}\\times \\mathrm{W}\\times 3}" + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "text", + "content": " is the virtual latent sharp image captured at timestamp t " + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\in [0,\\tau ]" + }, + { + "bbox": [ + 130, + 616, + 479, + 666 + ], + "type": "text", + "content": " within the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "text", + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "exposure time. The blurred image " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{B}(\\mathbf{u})" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ", resulting from camera motion during the exposure time, is calculated by averaging all the virtual images " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{t}}(\\mathbf{u})" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " across different timestamps " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ". This discrete approximated of this model is dependent on the number " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " of discrete samples, as denoted:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 174, + 481, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 174, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 261, + 174, + 481, + 205 + ], + "type": "interline_equation", + "content": "\\mathbf {B} (\\mathbf {u}) \\approx \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\mathbf {C} _ {i} (\\mathbf {u}). \\tag {7}", + "image_path": "1c69e29f34b703ae20041b956ea2f98f4c0c7130e9251d6a70c7779c4b0ed686.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 480, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 298 + ], + "type": "text", + "content": "The level of motion blur within an image is contingent upon the movement of the camera during the exposure time. For instance, a swiftly moving camera results in minimal relative motion, particularly with shorter exposure times, while a slowly moving camera yields motion-blurred images, especially in low-light scenarios with prolonged exposure times. Additionally, it can be deduced that " + }, + { + "bbox": [ + 130, + 213, + 480, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{B}(\\mathbf{u})" + }, + { + "bbox": [ + 130, + 213, + 480, + 298 + ], + "type": "text", + "content": " demonstrates differentiability concerning each virtual sharp image " + }, + { + "bbox": [ + 130, + 213, + 480, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(\\mathbf{u})" + }, + { + "bbox": [ + 130, + 213, + 480, + 298 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 315, + 399, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 315, + 399, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 315, + 399, + 327 + ], + "type": "text", + "content": "3.3 Camera Motion Trajectory Modeling in 3D-GS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": "Based on Eq. (7), a straightforward approach to address motion-blurred images involves determining each virtual sharp image " + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i" + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": ", which serves as the dependent variable of the motion-blur image " + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": ". Given that a sharp image " + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i" + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": " can be rendered from a specified camera pose " + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i" + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": " within the framework of 3D-GS (i.e. " + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\theta}" + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": ") [14], establishing a one-to-one correspondence between poses and virtual sharp images is feasible. Consequently, we formulate the corresponding poses of each latent sharp image within the exposure time " + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": " by employing a camera motion trajectory represented through linear interpolation between two camera poses, one at the beginning of the exposure " + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}} \\in \\mathbf{SE}(3)" + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": " and the other at the end " + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}} \\in \\mathbf{SE}(3)" + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": ". The virtual camera pose at time " + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "inline_equation", + "content": "t \\in [0, \\tau]" + }, + { + "bbox": [ + 130, + 335, + 479, + 467 + ], + "type": "text", + "content": " can thus be expressed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 221, + 475, + 480, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 475, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 221, + 475, + 480, + 498 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {t} = \\mathbf {T} _ {\\text {s t a r t}} \\cdot \\exp \\left(\\frac {t}{\\tau} \\cdot \\log \\left(\\mathbf {T} _ {\\text {s t a r t}} ^ {- 1} \\cdot \\mathbf {T} _ {\\text {e n d}}\\right)\\right), \\tag {8}", + "image_path": "02fbe34a972d2c5953354fbbd70d2a7defcd68d7132ec63110da2e88069e130f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": " represents the exposure time. " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\frac{t}{\\tau}" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": " can be further discretized and derived as " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\frac{i}{n-1}" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": " sampled virtual sharp image (i.e. " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": ") in Eq. (7) (i.e. with pose denoted as " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": "), when there are " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": " sampled images in total. It follows that " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": " is differentiable with respect to both " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}}" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}}" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": ". We refer to prior works [26, 47] for a comprehensive understanding of the interpolation and derivations of the related Jacobian. [26]. The objective of BAD-Gaussians is to estimate both " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}}" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}}" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": " for each frame, along with the learnable parameters of Gaussians " + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\theta}" + }, + { + "bbox": [ + 130, + 506, + 482, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 621, + 231, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 231, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 231, + 632 + ], + "type": "text", + "content": "3.4 Loss Function" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": "From a collection of " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": " motion-blurred images, we can proceed to estimate both the learnable parameters " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": " (i.e. mean position " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": ", 3D covariance " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 220, + 146, + 391, + 226 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Ablation studies on the number of virtual camera poses " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ". The results indicate that performance reaches a saturation point with increasing number " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 220, + 146, + 391, + 226 + ], + "lines": [ + { + "bbox": [ + 220, + 146, + 391, + 226 + ], + "spans": [ + { + "bbox": [ + 220, + 146, + 391, + 226 + ], + "type": "table", + "html": "
nCozy2roomTanabata
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
331.63.9389.081025.43.8215.2287
432.99.9552.055827.08.8704.1764
533.68.9595.045828.39.9011.1400
633.86.9627.036829.57.9214.1101
834.48.9654.031531.15.9440.0696
1034.68.9521.025832.12.9481.0464
1534.78.9679.025733.58.9652.0198
2034.63.9646.030733.87.9668.0151
", + "image_path": "0324598dcf4cd206fc1b176fd8eb40c856f723519e2463b7b401e4ac94a7a21b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 166, + 279, + 441, + 317 + ], + "blocks": [ + { + "bbox": [ + 132, + 226, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 132, + 226, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 482, + 270 + ], + "type": "text", + "content": "Table 2: Ablation studies on the effect of trajectory representations. We denote Deblur-NeRF-S and Deblur-NeRF-R as the synthetic and real data from Deblur-NeRF, respectively. The results demonstrate that cubic interpolation improves performance in scenes with complex camera trajectories (i.e. " + }, + { + "bbox": [ + 132, + 226, + 482, + 270 + ], + "type": "inline_equation", + "content": "MBA - VO" + }, + { + "bbox": [ + 132, + 226, + 482, + 270 + ], + "type": "text", + "content": " and Deblur-NeRF-R)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 166, + 279, + 441, + 317 + ], + "lines": [ + { + "bbox": [ + 166, + 279, + 441, + 317 + ], + "spans": [ + { + "bbox": [ + 166, + 279, + 441, + 317 + ], + "type": "table", + "html": "
Deblur-NeRF-SMBA-VODeblur-NeRF-R
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Linear Interpolation33.92.9467.042230.96.8868.149124.77.7778.1291
Cubic B-spline33.24.9395.047331.06.8883.145426.11.8157.1067
", + "image_path": "db5831506a084222f3c2e47c9875b2ef5effb317bac08fb77e48aea855a920d7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "text", + "content": "and color " + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "text", + "content": ") of 3D-GS and the camera motion trajectories (i.e. " + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}}" + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}}" + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "text", + "content": ") for each image. This estimation is accomplished by minimizing the following loss function, which includes an " + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "text", + "content": " loss and a D-SSIM term between " + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_k(\\mathbf{u})" + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "text", + "content": " blurry image synthesized from 3D-GS using the aforementioned image formation model (i.e. Eq. (7)), and " + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_k^{gt}(\\mathbf{u})" + }, + { + "bbox": [ + 132, + 345, + 482, + 406 + ], + "type": "text", + "content": ", the corresponding real captured blurry image:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 421, + 481, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 421, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 246, + 421, + 481, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}}. \\tag {9}", + "image_path": "9975b0f0e7824c10929a7346dfa5ca5ae64fe4af908f4d818c3d517687612518.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "text", + "content": "To optimize the learnable Gaussians parameter " + }, + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "text", + "content": ", and camera pose " + }, + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "text", + "content": " (i.e. " + }, + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}}" + }, + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}}" + }, + { + "bbox": [ + 132, + 451, + 482, + 487 + ], + "type": "text", + "content": " in our model) for each image, it is necessary to derive the corresponding Jacobians to complete the gradient flow:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 232, + 502, + 481, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 502, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 232, + 502, + 481, + 534 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {\\theta}} = \\sum_ {k = 0} ^ {K - 1} \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {B} _ {k}} \\cdot \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\frac {\\partial \\mathbf {B} _ {k}}{\\partial \\mathbf {C} _ {i}} \\frac {\\partial \\mathbf {C} _ {i}}{\\partial \\boldsymbol {\\theta}}, \\tag {10}", + "image_path": "46809bb4350a9349207a65070d5c19b8ce2dacf183b84e4fc0113a9861a00fa2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 223, + 559, + 481, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 559, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 223, + 559, + 481, + 593 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {T}} = \\sum_ {k = 0} ^ {K - 1} \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {B} _ {k}} \\cdot \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\frac {\\partial \\mathbf {B} _ {k}}{\\partial \\mathbf {C} _ {i}} \\frac {\\partial \\mathbf {C} _ {i}}{\\partial \\boldsymbol {\\theta}} \\frac {\\partial \\boldsymbol {\\theta}}{\\partial \\mathbf {T}}, \\tag {11}", + "image_path": "aecab9e2fd3757a4d29740dce62b1598580d5d2ff9b7a0941a5fc5f84257d71d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_k(\\mathbf{u})" + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(\\mathbf{u})" + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": " are denoted as " + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_k" + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i" + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": " for simplification, and we parameterize both " + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}}" + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}}" + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": " with their corresponding Lie algebras of " + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{SE}(3)" + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": ", which can be represented by a 6D vector respectively. For further details on the Jacobian of Gaussians to camera pose, " + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\pmb{\\theta}}{\\partial\\mathbf{T}}" + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": ", please refer to our supplemental materials." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "text", + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 226 + ], + "type": "text", + "content": "Table 3: Quantitative comparisons for deblurring on the synthetic dataset of Deblur-NeRF [27], referred to as DB-NeRF in the table due to space constraints. Notably, DB-NeRF* and DP-NeRF* are trained with ground-truth poses, while the others are trained with poses estimated by COLMAP [38]. The experiments highlight the superior performance of our method over previous approaches. Additionally, the results demonstrate the sensitivity of Deblur-NeRF and DP-NeRF to pose accuracy. Regarding rendering efficiency, our method achieves over 200 FPS, whereas Deblur-NeRF, DP-NeRF, BAD-NeRF and ExBluRF fall below 1 FPS. Our method takes about 30 minutes to train, while other methods take more than 10 hours. Each color shading indicates the best and second-best result, respectively." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 234, + 479, + 335 + ], + "blocks": [ + { + "bbox": [ + 135, + 234, + 479, + 335 + ], + "lines": [ + { + "bbox": [ + 135, + 234, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 135, + 234, + 479, + 335 + ], + "type": "table", + "html": "
CozyroomFactoryPoolTanabataTrolley
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
NeRF [30]26.13.7886.248422.02.5581.455029.90.7901.259520.57.5653.458421.71.6413.3814
3D-GS [14]25.86.7908.227021.73.5623.450329.37.7816.256120.51.5773.427821.65.6587.3614
MPR [52]29.90.8862.091525.07.6994.240933.28.8938.129022.60.7203.250726.24.8356.1762
SRN [44]29.47.8759.095026.54.7604.240432.94.8847.104523.19.7274.243825.36.8119.1618
DB-NeRF [27]29.53.8786.087925.85.7651.234031.07.8477.139723.20.7243.247125.68.8111.1683
DB-NeRF*30.26.8933.079126.40.7991.219132.30.8755.134524.56.7749.216626.24.8254.1671
DP-NeRF [19]29.02.8761.077325.42.7559.213930.48.8402.112723.93.7540.203826.17.8269.1359
DP-NeRF*30.77.9020.058427.69.8328.184733.22.8922.095425.27.7973.177926.99.8413.1312
BAD-NeRF [47]32.11.9137.051432.18.9105.118932.22.8680.090825.80.8047.156329.68.8954.0752
ExBluRF [20]28.460.87070.111329.380.85110.164631.660.86120.171925.430.80060.22725.70.81120.234
Ours34.68.9521.025831.88.9270.095236.95.9434.022532.12.9481.046433.97.9628.0209
", + "image_path": "91698fd506a4fffc9a26bed17c9c81933e32eac03770f763ab0185a4e59f6f1a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 349, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 392 + ], + "type": "text", + "content": "Table 4: Quantitative comparisons for novel view synthesis on the synthetic dataset of Deblur-NeRF [27], referred to as DB-NeRF in the table due to space constraints. The results demonstrate Our methods outperform previous state-of-the-art approaches, delivering the best performance across the board." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 135, + 402, + 479, + 499 + ], + "blocks": [ + { + "bbox": [ + 135, + 402, + 479, + 499 + ], + "lines": [ + { + "bbox": [ + 135, + 402, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 135, + 402, + 479, + 499 + ], + "type": "table", + "html": "
CozyroomFactoryPoolTanabataTrolley
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [30]25.45.7734.251422.51.5672.419929.23.7711.275620.74.5774.434121.09.6157.3924
3D-GS [14]25.24.7789.208421.34.5532.437128.61.7562.261020.37.5755.395820.86.6167.3701
MPR [52]+ [14]29.00.8685.089423.38.6848.237429.04.8170.176122.53.7349.216226.27.8352.1694
SRN [44]+ [14]28.64.8662.090325.45.7740.204529.97.8207.165023.09.7456.214124.89.8103.1666
DB-NeRF [27]29.09.8718.093725.19.7359.231530.97.8401.160023.71.7483.236924.46.7758.2215
DB-NeRF*29.88.8901.074726.06.8023.210630.94.8399.169424.82.7861.204525.78.8122.1797
DP-NeRF [19]29.56.8831.068327.91.8317.183330.98.8454.126524.73.7867.186525.62.8123.1606
DP-NeRF*30.16.8958.060828.43.8420.177131.78.8623.130625.37.8046.170126.44.8307.1424
BAD-NeRF [47]31.07.9027.055131.71.9038.120430.95.8389.110725.42.8011.156628.60.8806.0843
Ours32.35.9278.034028.18.9164.099933.30.8964.055131.00.9427.047831.44.9443.0345
", + "image_path": "a883d2cea534a366e2d74fd0d24ac28c3cef4364ef4a07556c23aae407c0db65.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 523, + 230, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 230, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 230, + 537 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 553, + 273, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 273, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 273, + 566 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. We implemented our method within the Nerfstudio [43] framework based on their Splatifto implementation. Camera pose interpolation and optimization on the SE(3) manifold is implemented with pypose [45]. Both the optimization of Gaussians and camera poses are performed using the Adam optimizer. The number of virtual camera poses (i.e., " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " in Equation (7)) is set to 10, considering the trade-off between performance and efficiency. We initialize camera poses and Gaussians using estimations obtained" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 157, + 409, + 253 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Quantitative comparisons for deblurring on the synthetic dataset of MBA-VO [27]. The results demonstrate that our method achieves the best performance even when the camera undergoes acceleration" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 202, + 157, + 409, + 253 + ], + "lines": [ + { + "bbox": [ + 202, + 157, + 409, + 253 + ], + "spans": [ + { + "bbox": [ + 202, + 157, + 409, + 253 + ], + "type": "table", + "html": "
ArchViz-lowArchViz-highAverage
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
NeRF [30]26.26.7887.377523.29.7098.491024.78.7493.4343
3D-GS [14]26.54.8113.338523.43.7243.461424.99.7678.4000
MPR [52]29.60.8757.210325.04.7711.357627.32.8234.2840
SRN [44]30.15.8814.170327.07.8190.279628.61.8502.2250
DB-NeRF [27]28.38.8484.179225.71.7762.322027.05.8123.2506
DB-NeRF*29.65.8744.176426.44.8010.317228.05.8377.2468
DP-NeRF [19]28.24.8490.161525.74.7839.291326.99.8165.2264
DP-NeRF*28.57.8481.151126.80.8081.273627.69.8281.2124
BAD-NeRF [47]30.51.8749.165427.54.8109.267929.03.8429.2167
ExBluRF [20]28.4.8372.251828.43.8518.172128.42.8445.2120
Ours32.28.9167.113429.64.8568.184730.96.8868.1491
", + "image_path": "7011ab042154068401cb65683bf7c08cd83afed5810c9bd9f9daad89427271c5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 275, + 479, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 479, + 299 + ], + "type": "text", + "content": "from COLMAP [38]. All experiments are conducted on an NVIDIA RTX 4090 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 422 + ], + "type": "text", + "content": "Benchmark Datasets. We evaluate the performance of our method using both synthetic and real datasets provided by Deblur-NeRF [27] and BAD-NeRF [47]. Deblur-NeRF [47] contains five scenes, where the blurred images are generated via Blender [7] by averaging sharp virtual images captured during the exposure time, under the assumption of consistent velocity camera motion. For our evaluation, we utilize the dataset from BAD-NeRF [47], which expands upon the number of virtual sharp images to create more realistic motion-blurred images while keeping other settings consistent with [27]. Additionally, Deblur-NeRF [27] captured a real motion-blurred dataset by intentionally shaking a handheld camera during exposure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 480, + 506 + ], + "type": "text", + "content": "To enhance the evaluation of our method on severely motion-blurred images, we incorporate a dataset tailored for motion blur-aware visual odometry benchmarking (i.e. MBA-VO [26]). Different from the assumption of constant velocity in the Deblur-NeRF dataset [27], the blur images synthesized from MBA-VO [26] are based on real camera motion trajectories obtained from the ETH3D dataset [39], which do not exhibit constant velocity and include accelerations, thus presenting notable challenges." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 593 + ], + "type": "text", + "content": "Baselines and Evaluation Metrics. We conduct a comparative deblurring analysis between our method and state-of-the-art learning-based single image deblurring algorithms including SRN [44] and MPR [52]. Additionally, we include evaluations against closely related approaches such as Deblur-NeRF [27], DP-NeRF [19], BAD-NeRF [47] and ExBluRF [20]. To evaluate deblurring performance, we render high-quality images corresponding to the midpoint (i.e. " + }, + { + "bbox": [ + 130, + 510, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\frac{\\tau}{2}}" + }, + { + "bbox": [ + 130, + 510, + 480, + 593 + ], + "type": "text", + "content": " in Eq. (6)) of exposure time of each training image from the optimized Gaussians." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "To facilitate novel view synthesis using SRN [44] and MPR [52], we use the deblurred images obtained from pre-trained models as inputs for training 3D-GS, as these single deblurring methods are not optimized for synthesizing novel images. During the training stage, all poses are estimated using COLMAP [38]. Additionally, we train Deblur-NeRF [27] and DP-NeRF [19] with ground-truth camera poses from Blender to assess the impact of inaccurate pose estimation on" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "text", + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 487, + 350 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 487, + 350 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 487, + 350 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 487, + 350 + ], + "type": "image", + "image_path": "d53905ffe7f5b311e3e36b30c4669e13d288c016d4d132e72dd32a34da914a34.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 358, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 413 + ], + "type": "text", + "content": "Fig. 2: Qualitative novel view synthesis results of different methods with synthetic datasets. Despite being trained with ground truth poses " + }, + { + "bbox": [ + 130, + 358, + 482, + 413 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 130, + 358, + 482, + 413 + ], + "type": "text", + "content": ", BAD-Gaussians outperforms Deblur-NeRF* and DP-NeRF* in recovering high-quality scenes from motion-blurred images with inaccurate camera poses, showcasing its superior performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 438, + 480, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 480, + 462 + ], + "type": "text", + "content": "these two methods. The quality of the rendered sharp image is evaluated using PSNR, SSIM, and LPIPS metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 463, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 500 + ], + "type": "text", + "content": "For pose accuracy evaluation, we compare the Absolute Trajectory Error (ATE) metric, against classical structure-from-motion framework COLMAP [38] and BAD-NeRF [47]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 520, + 238, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 238, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 238, + 533 + ], + "type": "text", + "content": "4.2 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Virtual Camera Poses. We experimented to investigate the effect of the number of interpolated virtual camera poses within the exposure time, as described by " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " in Eq. (7). Two of the five synthetic scenes provided by Deblur-NeRF [27] were selected, representing sequences with minor and severe motion blur (i.e. Cozy2room and Tanabata), respectively. We varied the number, " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", from 3 to 20 in our study, and the rendering metric results are presented in Table 1. Our experimental results demonstrate that increasing the number of interpolated virtual camera poses helps in addressing severe motion blur. However, marginal improvements are observed for minor blur instances. Based on our experiments, we choose " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "n = 10" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " interpolated virtual images to strike a balance between ren" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "dering performance and training efficiency (larger " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " means more computational resources)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "text", + "content": "Trajectory Representations. To assess the impact of different trajectory representations, we conduct two experiments: one involves optimizing " + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}}" + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}}" + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "text", + "content": " to depict a linear trajectory, while the other utilizes a higher-order spline (i.e., cubic B-spline) that jointly optimizes four control knots " + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_1" + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_2" + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_3" + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_4" + }, + { + "bbox": [ + 130, + 143, + 482, + 370 + ], + "type": "text", + "content": " to capture more complex camera motions. We refer [46] for more details about cubic B-spline. The average quantitative results on the synthetic datasets of Deblur-NeRF-S [27] (Cozy2room, Factory, Pool, Tanabata and Trolley) and MBA-VO [26] (ArchViz-low and ArchViz-high), along with real datasets, Deblur-NeRF-S, which includes 10 real captured scenes from Deblur-NeRF, are presented in Table 2. In synthetic scenes, linear interpolation demonstrates comparable performance to cubic B-spline interpolation. However, in real captured scenes, cubic B-spline interpolation outperforms linear interpolation, particularly due to the longer exposure time during image capture. The effectiveness of cubic B-spline interpolation in real scenes can be attributed to its ability to better model the nuances of camera motion over longer time intervals. Conversely, linear interpolation is sufficient to accurately represent camera motion trajectories within shorter time intervals, as observed in synthetic scenes. Combining the training efficiency and rendering quality, we employ linear interpolation in synthetic datasets and cubic B-spline in real data." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 230, + 389, + 384, + 470 + ], + "blocks": [ + { + "bbox": [ + 230, + 389, + 384, + 470 + ], + "lines": [ + { + "bbox": [ + 230, + 389, + 384, + 470 + ], + "spans": [ + { + "bbox": [ + 230, + 389, + 384, + 470 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS↓Ttrain(h)
NeRF [30]22.69.6347.36872.5
3DGS [14]22.72.6390.28870.3
DB-NeRF [27]25.63.7675.1820>5
DP-NeRF [19]25.91.7734.1589>5
BAD-NeRF [47]24.33.7003.2196>5
ExBluRF [20]25.31.7293.20844.0
Ours-Cubic26.11.8157.10640.5
", + "image_path": "58c8b77b6f1cc002d2c71b2c870d0c5a10b8b18d849cf61dded6598022c18314.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 470, + 479, + 492 + ], + "lines": [ + { + "bbox": [ + 130, + 470, + 479, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 479, + 492 + ], + "type": "text", + "content": "Table 6: Quantitative comparisons for novel view synthesis on the real captured dataset of Deblur-NeRF [27]. Full results can be found in our supplementary material." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 538, + 198, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 538, + 198, + 549 + ], + "spans": [ + { + "bbox": [ + 131, + 538, + 198, + 549 + ], + "type": "text", + "content": "4.3 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 479, + 641 + ], + "type": "text", + "content": "Results on Synthetic Data. We evaluate our approach against baseline methods using synthetic datasets obtained from both Deblur-NeRF [27] and MBAVO [26]. The quantitative evaluation results on deblurring and novel view synthesis with the Deblur-NeRF dataset are presented in Table 3 and Table 4 respectively. All metrics (i.e. PSNR, SSIM, LPIPS) demonstrate substantial improvements over prior state-of-the-art methods (i.e. 3.6 and " + }, + { + "bbox": [ + 130, + 557, + 479, + 641 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 557, + 479, + 641 + ], + "type": "text", + "content": " higher than the second best method on average in Table 3 and Table 4, respectively)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "The results demonstrate that both NeRF [30] and 3D-GS [14] are suffering from motion-blurred images, which motivates the necessity of our method." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "text", + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 339 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 339 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 339 + ], + "type": "image", + "image_path": "43b01f0e8fcc0343e2736b6dce4c20139c2f7d63ce030a18bd2fc21398670e3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 347, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 403 + ], + "type": "text", + "content": "Fig. 3: Qualitative novel view synthesis results of different methods with the real datasets. The experimental results demonstrate that our method achieves superior performance over prior methods on the real dataset as well. In contrast, BAD-NeRF yields poorer results when applied to real data and exhibits satisfactory performance only within synthetic datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 434, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 579 + ], + "type": "text", + "content": "Single-stage methods such as MPR [52] and SRN [44] fall short of matching the performance of our approach due to their limited utilization of geometric information across multi-view images. Additionally, our method surpasses Deblur-NeRF [27] and DP-NeRF [19], partly because they fail to optimize the inaccurate camera poses estimated from COLMAP during training. We additionally trained Deblur-NeRF and DP-NeRF with ground truth poses (denoted as *), and the results demonstrate improved performance, reflecting the sensitivity to the pose accuracies of both Deblur-NeRF and DP-NeRF. A notable result from both Table 3 and Table 4 is that our method performs worse than BAD-NeRF on the Factory sequence. We find that it is caused by the inferior capability to represent the sky by Gaussian splatting compared to NeRF. Nevertheless, our method still performs better than all prior methods with a large margin on average." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 668 + ], + "type": "text", + "content": "Qualitative results are presented in Fig. 2. It demonstrates that our method effectively attains high-quality scene representations with intricate details, being trained from a series of blurred images. In Fig. 2, it is evident that Deblur-NeRF [27] and DP-NeRF [19] encounter difficulties in modeling regions with significant depth variations, attributable to their method of synthesizing blurred images through convolution with a blur kernel applied to rendered images. The physically based method BAD-NeRF [47] exhibits superior performance overall," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "text", + "content": "yet it still presents some deficiencies, particularly noticeable around areas with significant color and depth variation. It demonstrates the effectiveness of the explicit Gaussian splatting representation over the implicit neural representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 479, + 209 + ], + "type": "text", + "content": "We further assess the performance of our methods using a dataset characterized by severe motion blur and camera movements with varying velocities, sourced from MBA-VO [26]. The results presented in Table 5 demonstrate that our method achieves the best performance even when the camera undergoes acceleration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 214, + 479, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 479, + 273 + ], + "type": "text", + "content": "Results on Real Data. The performance of novel view synthesis on real data sourced from Deblur-NeRF [27] are also evaluated. The quantitative results, as presented in Table 6, demonstrate the superior performance of our method compared to other approaches. Furthermore, the qualitative results illustrated in Fig. 3 vividly demonstrate our method's ability to deliver intricate details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 479, + 323 + ], + "type": "text", + "content": "Pose Estimation. Due to the unknown metric scale, we align the estimated trajectories against their corresponding ground truth for the computation of the absolute trajectory error metric. The experimental results presented in Table 7 demonstrate the effectiveness of our approach in recovering precise camera poses." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 159, + 401, + 451, + 443 + ], + "blocks": [ + { + "bbox": [ + 130, + 346, + 479, + 388 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 479, + 388 + ], + "type": "text", + "content": "Table 7: Pose estimation performance of BAD-Gaussians on various blur sequences. The results are in the absolute trajectory error metric in centimeters (ATE/cm). The COLMAP-sharp / COLMAP-blur represents the result of COLMAP with sharp/blurry images respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 401, + 451, + 443 + ], + "lines": [ + { + "bbox": [ + 159, + 401, + 451, + 443 + ], + "spans": [ + { + "bbox": [ + 159, + 401, + 451, + 443 + ], + "type": "table", + "html": "
Cozy2roomFactoryPoolTanabataTrolleyArchViz-lowArchViz-highAverage
COLMAP-sharp [38].133±.079.101±.052.172±.047.277±.103.173±.079.148±.078.115±.066.1599
COLMAP-blur [38].287±.1611.69±.1.24.633±.2361.59±.1.06.912±.559.536±.309.766±.391.9163
BAD-NeRF [47].316±.101.600±.208.095±.0461.34±.543.122±.0701.24±.6352.18±1.54.8419
Ours.165±.057.395±.426.264±.0981.20±.621.134±.070.520±.5911.41±1.11.5840
", + "image_path": "b650d06ad24b8a9be752fc4042659a1632922fca64ddb1163ffbe51effc15fdd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 485, + 219, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 485, + 219, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 485, + 219, + 498 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 479, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 479, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 479, + 580 + ], + "type": "text", + "content": "We presented the first pipeline to learn Gaussian splittings from a set of motion-blurred images with inaccurate camera poses. Our pipeline can jointly optimize the 3D scene representation and camera motion trajectories. Extensive experimental evaluations demonstrate that our method can deliver high-quality novel view images, and achieve real-time rendering compared to prior state-of-the-art works." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 601, + 246, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 246, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 246, + 615 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 626, + 479, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 626, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 626, + 479, + 662 + ], + "type": "text", + "content": "This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "text", + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://jonbarron.info/mipnerf/2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 186, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 481, + 220 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://jonbarron.info/mipnerf360/3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2023), https://jonbarron.info/zipnerf/ 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "text", + "content": "4. Cao, A., Johnson, J.: HexPlane: A Fast Representation for Dynamic Scenes. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://caoang327.github.io/HexPlane/ 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "type": "text", + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensorRF: Tensorial Radiance Fields. In: European Conference on Computer Vision (ECCV) (2022), https://apchenstu.github.io/TensoRF/ 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "text", + "content": "6. Cho, S., Lee, S.: Fast Motion Deblurring. ACM Transactions on Graphics (TOG) 28(5) (2009) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 375 + ], + "type": "text", + "content": "7. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "type": "text", + "content": "8. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised NeRF: Fewer Views and Faster Training for Free. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://www.cs.cmu.edu/~dsnerf/3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 410, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 481, + 442 + ], + "type": "text", + "content": "9. Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T.: Removing camera shake from a single photograph. In: ACM Transactions on Graphics (TOG) (2006) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 443, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 481, + 485 + ], + "type": "text", + "content": "0. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://sarafridov.github.io/K-Planes/4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "text", + "content": "1. Fu, Y., Liu, S., Kulkarni, A., Kautz, J., Efros, A.A., Wang, X.: Colmap-Free 3D Gaussian Splatting. arXiv preprint arXiv:2312.07504 (2023), https://oasisyang.github.io/colmap-free-3dgs/2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 481, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 554 + ], + "type": "text", + "content": "2. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: FastNeRF: High-Fidelity Neural Rendering at 200FPS. In: International Conference on Computer Vision (ICCV) (2021), https://microsoft.github.io/FastNeRF/2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "type": "text", + "content": "3. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-Calibrating Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://postech-cvlab.github.io/SCNeRF/ 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 481, + 632 + ], + "type": "text", + "content": "4. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (TOG) 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ 2, 4, 5, 7, 9, 10, 12, 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "5. Kim, M., Seo, S., Han, B.: InfoNeRF: Ray Entropy Minimization for Few-Shot Neural Volume Rendering. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://cv.snu.ac.kr/research/InfoNeRF/3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Krishnan, D., Fergus, R.: Fast image deconvolution using Hyper-Laplacian priors. In: Neural Information Processing Systems (NIPS) (2009) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "text", + "content": "17. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better. In: International Conference on Computer Vision (ICCV) (2019), https://github.com/VITA-Group/DeblurGANv2 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "text", + "content": "18. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://github.com/facebookresearch/pytorch3d/ 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "text", + "content": "19. Lee, D., Lee, M., Shin, C., Lee, S.: DP-NeRF: Deblurred Neural Radiance Field With Physical Scene Priors. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://dogyoonlee.github.io/dpnerf/2, 3, 4, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "text", + "content": "20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: ExBluRF: Efficient Radiance Fields for Extreme Motion Blurred Images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023) 4, 9, 10, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 480, + 300 + ], + "type": "text", + "content": "21. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithm. In: Computer Vision and Pattern Recognition (CVPR) (2009) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 480, + 320 + ], + "type": "text", + "content": "22. Levoy, M.: Efficient Ray Tracing of Volume Data. ACM Transactions on Graphics (TOG) 9(3), 245-261 (1990) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 320, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 320, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 133, + 320, + 480, + 353 + ], + "type": "text", + "content": "23. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields. In: International Conference on Learning Representations (ICLR) (2024), https://moyangli00.github.io/usb_nerf/ 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 353, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 353, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 133, + 353, + 480, + 395 + ], + "type": "text", + "content": "24. Li, W., Wan, P., Wang, P., Li, J., Zhou, Y., Liu, P.: BeNeRF: Neural Radiance Fields from a Single Blurry Image and Event Stream. In: European Conference on Computer Vision (ECCV). Springer (2024), https://github.com/WU-CVGL/BeNeRF 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 396, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 480, + 439 + ], + "type": "text", + "content": "25. Lin, Chen-Hsuan and Ma, Wei-Chiu and Torralba, Antonio and Lucey, Simon: BARF: Bundle-Adjusting Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://chenhsuanlin.bitbucket.io/bundle-adjusting-NeRF 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 439, + 480, + 472 + ], + "type": "text", + "content": "26. Liu, P., Zuo, X., Larsson, V., Pollefeys, M.: MBA-VO: Motion Blur Aware Visual Odometry. In: International Conference on Computer Vision (ICCV) (2021), https://github.com/ethliup/MBA-VO 7, 10, 12, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 472, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 480, + 514 + ], + "type": "text", + "content": "27. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural Radiance Fields from Blurry Images. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://limacv.github.io/deblurnerf/ 2, 3, 4, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 514, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 480, + 536 + ], + "type": "text", + "content": "28. Max, N.: Optical Models for Direct Volume Rendering. ACM Transactions on Graphics (TOG) 1(2), 99-108 (1995) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 536, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 578 + ], + "type": "text", + "content": "29. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://bmild.github.io/rawnerf/3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 578, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 480, + 621 + ], + "type": "text", + "content": "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: European Conference on Computer Vision (ECCV) (2020), https://www.matthewtancik.com/nerf 2, 3, 9, 10, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 621, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 480, + 664 + ], + "type": "text", + "content": "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics (TOG) 41(4), 102:1-102:15 (Jul 2022), https://nvlabs.github.io/instant-ngp/2, 3, 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 328, + 102 + ], + "type": "text", + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 667 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "32. Nah, S., Kim, T.H., Lee, K.M.: Deep Multi-Scale Convolutional Neural Network for Dynamic Scene Deblurring. In: Computer Vision and Pattern Recognition (CVPR) (2017), https://github.com/SeungjunNah/DeepDeblur_release 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 195 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 195 + ], + "type": "text", + "content": "33. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse Inputs. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://m-niemeyer.github.io/regnerf/3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 195, + 482, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 195, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 482, + 239 + ], + "type": "text", + "content": "34. Park, H., Lee, K.M.: Joint Estimation of Camera Pose, Depth, Deblurring, and Super-Resolution from a Blurred Image Sequence. In: International Conference on Computer Vision (ICCV). vol. 2017-Octob (2017). https://doi.org/10.1109/ICCV.2017.494 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "type": "text", + "content": "35. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: CamP: Camera Preconditioning for Neural Radiance Fields. ACM Transactions on Graphics (TOG) (2023), https://camp-nerf.github.io/3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "type": "text", + "content": "36. Piala, M., Clark, R.: TermiNeRF: Ray Termination Prediction for Efficient Neural Rendering. In: International Conference on 3D Vision (3DV) (2021), https://projects.mackopes.com/terminerf/3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 308, + 482, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 342 + ], + "type": "text", + "content": "37. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance Fields without Neural Networks. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://alexyu.net/plenoxels/ 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 374 + ], + "type": "text", + "content": "38. Schonberger, J.L., Frahm, J.M.: Structure-from-motion Revisited. In: Computer Vision and Pattern Recognition (CVPR) (2016), https://github.com/colmap/ colmap 2, 5, 9, 10, 11, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 374, + 482, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 408 + ], + "type": "text", + "content": "39. Schops, T., Sattler, T., Pollefeys, M.: BAD SLAM: Bundle Adjusted Direct RGB-D SLAM. In: Computer Vision and Pattern Recognition (CVPR) (2019), https://github.com/ETH3D/badslam 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 408, + 482, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 432 + ], + "type": "text", + "content": "40. Shan, Q., Jia, J., A., A.: High-quality Motion Deblurring from a Single Image. ACM Transactions on Graphics (TOG) 27(3) (2008) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 432, + 482, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 465 + ], + "type": "text", + "content": "41. Su, S., Delbracio, M., Wang, J.: Deep Video Deblurring for Hand-held Cameras. In: Computer Vision and Pattern Recognition (CVPR) (2017), https://github.com/shuochsu/DeepVideoDeblurring 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 465, + 482, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 498 + ], + "type": "text", + "content": "42. Sun, C., Sun, M., Chen, H.: Direct Voxel Grid Optimization: Super-fast Convergence for Radiance Fields Reconstruction. In: Computer Vision and Pattern Recognition (CVPR) (2022), https://sunset1995.github.io/dvgo/3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 498, + 482, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 553 + ], + "type": "text", + "content": "43. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A Modular Framework for Neural Radiance Field Development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023), https://github.com/nerfstudio-project/nerfstudio/9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 554, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 588 + ], + "type": "text", + "content": "44. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Computer Vision and Pattern Recognition (CVPR) (2018), https://github.com/jiangsutx/SRN-Deblur 4, 9, 10, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 588, + 482, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 667 + ], + "type": "text", + "content": "45. Wang, C., Gao, D., Xu, K., Geng, J., Hu, Y., Qiu, Y., Li, B., Yang, F., Moon, B., Pandey, A., Aryan, Xu, J., Wu, T., He, H., Huang, D., Ren, Z., Zhao, S., Fu, T., Reddy, P., Lin, X., Wang, W., Shi, J., Talak, R., Cao, K., Du, Y., Wang, H., Yu, H., Wang, S., Chen, S., Kashyap, A., Bandaru, R., Dantu, K., Wu, J., Xie, L., Carlone, L., Hutter, M., Scherer, S.: PyPose: A library for robot learning with physics-based optimization. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023), https://github.com/pypyse/pypyse 9" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 381 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "text", + "content": "46. Wang, P., Zhao, L., Ma, R., Liu, P.: Supplementary Material for BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields https://openaccess.thecvf.com/content/CVPR2023/supplemental/Wang_BAD-NeRF_Bundle_Adjusted_CVPR_2023SUPPLEMENTAL.pdf 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "47. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Computer Vision and Pattern Recognition (CVPR) (2023), https://wangpeng000.github.io/BAD-NeRF/2, 3, 4, 7, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "text", + "content": "48. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021), https://nerfmm.active.vision/3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "49. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: NeX: Real-time View Synthesis with Neural Basis Expansion. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://nex-mpi.github.io/3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 260, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 481, + 282 + ], + "type": "text", + "content": "50. Xu, L., Jia, J.: Two-Phase Kernel Estimation for robust motion deblurring. In: European Conference on Computer Vision (ECCV) (2010) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "text", + "content": "51. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for Real-time Rendering of Neural Radiance Fields. In: International Conference on Computer Vision (ICCV) (2021), https://alexyu.net/plenoctrees/ 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage Progressive Image Restoration. In: Computer Vision and Pattern Recognition (CVPR) (2021), https://github.com/swz30/MPRNet 9, 10, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 481, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 381 + ], + "type": "text", + "content": "53. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: EWA Volume Splatting. In: Proceedings Visualization, 2001. VIS'01. pp. 29-538. IEEE (2001), https://cgl.ethz.ch/research/past_projects/surfels/ewavolumesplatting/index.html 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 329, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 329, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 329, + 102 + ], + "type": "text", + "content": "Lingzhe Zhao, Peng Wang, Peidong Liu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_content_list.json b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a27f3c70d3782b145b90a4494d3c009bb463bcc6 --- /dev/null +++ b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_content_list.json @@ -0,0 +1,2003 @@ +[ + { + "type": "text", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "text_level": 1, + "bbox": [ + 108, + 48, + 891, + 107 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haozhe Feng $^{1,3\\star}$ , Tianyu Pang $^{2\\star\\star}$ , Chao Du $^{2}$ , Wei Chen $^{1\\star\\star}$ , Shuicheng Yan $^{2}$ , and Min Lin $^{2}$", + "bbox": [ + 181, + 139, + 818, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Sea AI Lab, Singapore $^{3}$ Tencent Inc. {fenghz, chenvis} @zju.edu.cn {tianyupang, duchao, yansc, linmin} @sea.com", + "bbox": [ + 244, + 195, + 750, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Federated learning (FL) is a general principle for decentralized clients to train a server model collectively without sharing local data. FL is a promising framework with practical applications, but its standard training paradigm requires the clients to backpropagate through the model to compute gradients. Since these clients are typically edge devices and not fully trusted, executing backpropagation on them incurs computational and storage overhead as well as white-box vulnerability. In light of this, we develop backpropagation-free federated learning, dubbed BAFFLE, in which backpropagation is replaced by multiple forward processes to estimate gradients. BAFFLE is 1) memory-efficient and easily fits uploading bandwidth; 2) compatible with inference-only hardware optimization and model quantization or pruning; and 3) well-suited to trusted execution environments, because the clients in BAFFLE only execute forward propagation and return a set of scalars to the server. Empirically we use BAFFLE to train deep models from scratch or to finetune pretrained models, achieving acceptable results.", + "bbox": [ + 147, + 339, + 849, + 624 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Federated Learning $\\cdot$ Finite Difference $\\cdot$ Backpropagation-Free Learning", + "bbox": [ + 147, + 643, + 849, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 79, + 715, + 309, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) allows decentralized clients to collaboratively train a server model [62]. In each training round, the selected clients compute model gradients or updates on their local private datasets, without explicitly exchanging sample points to the server. While FL describes a promising blueprint and has several applications [32, 51, 96], the mainstream training paradigm of FL is still gradient-based that requires the clients to locally execute backpropagation, which leads to two practical limitations:", + "bbox": [ + 77, + 760, + 917, + 897 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during an internship at Sea AI Lab.", + "bbox": [ + 79, + 913, + 566, + 930 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Corresponding author.", + "bbox": [ + 77, + 931, + 326, + 947 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) Overhead for edge devices. The clients in FL are usually edge devices, such as mobile phones and IoT sensors, whose hardware is primarily optimized for inference-only purposes [77,86], rather than for backpropagation. Due to the limited resources, computationally affordable models running on edge devices are typically quantized and pruned [91], making exact backpropagation difficult. In addition, standard implementations of backpropagation rely on either forward-mode or reverse-mode auto-differentiation in contemporary machine learning packages [14, 73], which increases storage requirements.", + "(ii) White-box vulnerability. To facilitate gradient computing, the server regularly distributes its model status to the clients, but this white-box exposure of the model renders the server vulnerable to, e.g., poisoning or inversion attacks from malicious clients [25,78,95,101]. With that, recent attempts are made to exploit trusted execution environments (TEEs) in FL, which can isolate the model status within a black-box secure area and significantly reduce the success rate of malicious evasion [19,64,65]. However, TEEs are highly memory-constrained [85], while backpropagation is memory-consuming to restore intermediate states." + ], + "bbox": [ + 72, + 55, + 922, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While numerous solutions have been proposed to alleviate these limitations (related work discussed in Section 5), we raise an essential question: how to perform backpropagation-free $FL$ ? Inspired by the literature on zero-order optimization [80], we intend to substitute backpropagation with multiple forward or inference processes to estimate the gradients. Technically speaking, we propose the framework of BAckpropagation-Free Federated LEarning (BAFFLE). As illustrated in Figure 1, BAFFLE consists of three conceptual steps: (1) each client locally perturbs the model parameters $2K$ times as $\\mathbf{W} \\pm \\boldsymbol{\\delta}_k$ (the server sends the random seed to clients for generating $\\{\\pmb{\\delta}_k\\}_{k=1}^K$ ); (2) each client executes forward processes on the perturbed models using its private dataset $\\mathbb{D}_c$ and obtains $K$ loss differences $\\{\\Delta \\mathcal{L}(\\mathbf{W}, \\pmb{\\delta}_k; \\mathbb{D}_c)\\}_{k=1}^K$ ; (3) the server aggregates loss differences to estimate gradients.", + "bbox": [ + 72, + 385, + 922, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "BAFFLE's defining characteristic is that it only utilizes forward propagation, which is memory-efficient and does not require auto-differentiation. It is well-adapted to model quantization and pruning as well as inference-only hardware optimization on edge devices. Compared to backpropagation, the computation graph of BAFFLE is more easily optimized, such as by slicing it into per-layer calculation [44]. Since each loss difference $\\Delta \\mathcal{L}(\\mathbf{W},\\delta_k;\\mathbb{D}_c)$ is a scalar, BAFFLE can easily accommodate the uploading bandwidth of clients by adjusting the value of $K$ as opposed to using, e.g., gradient compression [82]. BAFFLE is also compatible with recent advances in inference approaches for TEE [83, 85], providing an efficient solution for combining TEE into FL and preventing white-box evasion.", + "bbox": [ + 72, + 627, + 922, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We adapt secure aggregation [10] to zero-order optimization and investigate ways to improve gradient estimation in BAFFLE. Empirically, BAFFLE is used to train models from scratch on MNIST [49] and CIFAR-10/100 [48], and transfer ImageNet-pretrained models to OfficeHome [87]. Compared to conventional FL, it achieves suboptimal but acceptable performance. These results shed light on the potential of BAFFLE and general backpropagation-free methods in FL.", + "bbox": [ + 72, + 830, + 922, + 949 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 77, + 16, + 94, + 30 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ef0025f69632bbca34234f0731ec79e5ed18a0e337e927ce34f80ffddb92f639.jpg", + "image_caption": [ + "Fig. 1: A sketch map of BAFFLE. In addition to the global parameters update $\\Delta \\mathbf{W}$ , each client downloads random seeds to locally generate perturbations $\\pm \\delta_{1:K}$ and perform $2K$ times of forward propagation (i.e., inference) to compute loss differences. The server can recover these perturbations using the same random seeds and obtain $\\Delta \\mathcal{L}(\\mathbf{W},\\pmb{\\delta}_k)$ by secure aggregation. Each loss difference $\\Delta \\mathcal{L}(\\mathbf{W},\\pmb{\\delta}_k;\\mathbb{D}_c)$ is a floating-point number, so $K$ can be easily adjusted to fit the uploading bandwidth." + ], + "image_footnote": [], + "bbox": [ + 84, + 58, + 912, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Preliminaries", + "text_level": 1, + "bbox": [ + 75, + 463, + 319, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finite difference. Gradient-based optimization techniques (either first-order or higher-order) are the most frequently used tools to train deep networks [27]. Nevertheless, recent progress demonstrates promising applications of zero-order optimization methods for training, particularly when exact derivatives cannot be obtained [23, 55, 69] or backward processes are computationally prohibitive [34, 70]. Zero-order approaches require only multiple forward processes that may be executed in parallel. Along this routine, finite difference stems from the definition of derivatives and can be generalized to higher-order and multivariate cases by Taylor's expansion. For any differentiable loss $\\mathcal{L}(\\mathbf{W};\\mathbb{D})$ and a small perturbation $\\delta \\in \\mathbb{R}^n$ , finite difference employs the forward difference scheme", + "bbox": [ + 72, + 508, + 922, + 702 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {W} + \\boldsymbol {\\delta}; \\mathbb {D}) - \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) = \\boldsymbol {\\delta} ^ {\\top} \\nabla_ {\\mathbf {W}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) + o (\\| \\boldsymbol {\\delta} \\| _ {2}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 717, + 917, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta^{\\top}\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})$ is a scaled directional derivative along $\\pmb{\\delta}$ . Furthermore, we can use the central difference scheme to obtain higher-order residuals as", + "bbox": [ + 75, + 756, + 917, + 796 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {W} + \\boldsymbol {\\delta}; \\mathbb {D}) - \\mathcal {L} (\\mathbf {W} - \\boldsymbol {\\delta}; \\mathbb {D}) = 2 \\boldsymbol {\\delta} ^ {\\top} \\nabla_ {\\mathbf {W}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) + o (\\| \\boldsymbol {\\delta} \\| _ {2} ^ {2}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 811, + 917, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated learning. Suppose we have $C$ clients, and the $c$ -th client's private dataset is defined as $\\mathbb{D}_c \\coloneqq \\{(\\mathbf{X}_i^c, \\mathbf{y}_i^c)\\}_{i=1}^{N_c}$ with $N_c$ input-label pairs. Let $\\mathcal{L}(\\mathbf{W}; \\mathbb{D}_c)$ represent the loss function for the dataset $\\mathbb{D}_c$ , where $\\mathbf{W} \\in \\mathbb{R}^n$ denotes the server model's global parameters. The training objective of FL is to find $\\mathbf{W}$ that minimizes the total loss function as", + "bbox": [ + 75, + 850, + 920, + 947 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 840, + 34 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Backpropagation-free federated learning (BAFFLE)", + "bbox": [ + 79, + 56, + 782, + 74 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Notations: Se denotes the operations done on servers; Cl denotes the operations done on clients; TEE for the TEE module; and $\\Rightarrow$ denotes the communication process.", + "2: Inputs: $C$ clients with local dataset $\\{\\mathbb{D}_c\\}_{c = 1}^C$ containing $N_{c}$ input-label pairs, $N = \\sum_{c = 1}^{C}N_{c}$ ; learning rate $\\eta$ , training iterations $T$ , perturbation number $K$ , noise scale $\\sigma$ .", + "3: Se: initializing model parameters $\\mathbf{W} \\gets \\mathbf{W}_0$ ;", + "4: // optional" + ], + "bbox": [ + 84, + 79, + 920, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Se: encoding the computing paradigm into TEE as $\\mathrm{TEE} \\circ \\Delta \\mathcal{L}(\\mathbf{W}, \\delta; \\mathbb{D})$ ;", + "bbox": [ + 116, + 224, + 823, + 240 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "5: for $t = 0$ to $T - 1$ do", + "6: $\\mathrm{Se} \\Rightarrow$ all Cl: downloading model parameters $\\mathbf{W}_t$ and the computing paradigm;", + "7: // 4 Bytes" + ], + "bbox": [ + 87, + 243, + 910, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathrm{Se} \\Rightarrow$ all Cl: downloading the random seed $s_t$ ;", + "bbox": [ + 140, + 295, + 590, + 312 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "8: Se: sampling $K$ perturbations $\\{\\pmb{\\delta}_k\\}_{k=1}^K$ from $\\mathcal{N}(0, \\sigma^2\\mathbf{I})$ using the random seed $s_t$ ;", + "9: all Cl: negotiating a group of zero-sum noises $\\{\\pmb{\\epsilon}_c\\}_{c=1}^C$ for secure aggregation;" + ], + "bbox": [ + 87, + 313, + 910, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "10: for $c = 1$ to $C$ do", + "bbox": [ + 79, + 349, + 324, + 364 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "11: Cl: sampling $K$ perturbations $\\{\\pmb {\\delta}_k\\}_{k = 1}^K$ from $\\mathcal{N}(0,\\sigma^2\\mathbf{I})$ using the random seed $s_t$ ;", + "12: Cl: computing TEE $\\circ \\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k;\\mathbb{D}_c)$ via forward propagation for each $k$" + ], + "bbox": [ + 79, + 365, + 917, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "13: // $4\\times K$ Bytes", + "bbox": [ + 79, + 421, + 319, + 437 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathrm{Cl} \\Rightarrow \\mathrm{Se}$ : uploading $K$ outputs $\\left\\{\\mathrm{TEE} \\circ \\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k; \\mathbb{D}_c) + \\frac{N}{N_c} \\pmb{\\epsilon}_c\\right\\}_{k=1}^K$ ;", + "bbox": [ + 164, + 437, + 840, + 469 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "14: end for", + "15: Se: aggregating $\\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k)\\gets \\sum_{c = 1}^{C}\\frac{N_c}{N}\\left[\\mathrm{TEE}\\circ \\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k;\\mathbb{D}_c) + \\frac{N}{N_c}\\pmb {\\epsilon}_c\\right]$ for each $k$", + "16: Se: computing $\\widehat{\\nabla_{\\mathbf{W}_t}}\\mathcal{L}(\\mathbf{W}_t)\\gets \\frac{1}{K}\\sum_{k = 1}^{K}\\frac{\\delta_k}{2\\sigma^2}\\varDelta\\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k);$", + "17: Se: $\\mathbf{W}_{t + 1}\\gets \\mathbf{W}_t - \\eta \\widehat{\\nabla\\mathbf{w}_t}\\mathcal{L}(\\mathbf{W}_t)$", + "18: end for", + "19: Return: final model parameters $\\mathbf{W}_T$ ." + ], + "bbox": [ + 79, + 469, + 917, + 627 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {W}) := \\sum_ {c = 1} ^ {C} \\frac {N _ {c}}{N} \\mathcal {L} (\\mathbf {W}; \\mathbb {D} _ {c}), \\text {w h e r e} N = \\sum_ {c = 1} ^ {C} N _ {c}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 687, + 917, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the conventional FL framework, clients compute gradients $\\{\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}_c)\\}_{c = 1}^C$ or model updates locally through backpropagation and then upload them to the server. Federated average [62] performs global aggregation using $\\varDelta\\mathbf{W}:= \\sum_{i=1}^{C}\\frac{N_c}{N}\\varDelta\\mathbf{W}_c$ , where $\\varDelta\\mathbf{W}_c$ is the local update obtained via executing $\\mathbf{W}_c\\gets \\mathbf{W}_c - \\eta \\nabla_{\\mathbf{W}_c}\\mathcal{L}(\\mathbf{W}_c;\\mathbb{D}_c)$ multiple times and $\\eta$ is learning rate.", + "bbox": [ + 72, + 749, + 920, + 850 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zeroth-order FL. Similar to our work, DLZO [54] and FedZO [22] present zeroth-order optimization methods for FL independently in batch-level and epoch-level communications. However, they concentrate primarily on basic linear models with softmax regression problems and ignore deep models. Besides, they also do not account for server security aggregation in conjunction with zero-order", + "bbox": [ + 72, + 852, + 924, + 949 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 77, + 17, + 94, + 29 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "optimization. In comparison, BAFFLE enables security aggregation, can train deep models such as WideResNet from scratch and achieves reasonable results, e.g. $95.17\\%$ accuracy on MNIST with 20 communication rounds versus $83.58\\%$ for FedZO with 1,000 rounds.", + "bbox": [ + 72, + 55, + 920, + 134 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Backpropagation-Free Federated Learning", + "text_level": 1, + "bbox": [ + 75, + 162, + 728, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce zero-order optimization into FL and develop BAF-FLE, a backpropagation-free federated learning framework that uses multiple forward processes in place of backpropagation. An initial attempt is to apply finite difference as the gradient estimator. To estimate the full gradients, we need to perturb each parameter $w \\in \\mathbf{W}$ once to approximate the partial derivative $\\frac{\\partial \\mathcal{L}(\\mathbf{W}; \\mathbb{D})}{\\partial w}$ , causing the forward computations to grow with $n$ (recall that $\\mathbf{W} \\in \\mathbb{R}^n$ ) and making it difficult to scale to large models. In light of this, we resort to Stein's identity [80] to obtain an unbiased estimation of gradients from loss differences calculated on various perturbations. As depicted in Figure 1, BAFFLE clients need only download random seeds and global parameters update, generate perturbations locally, execute multiple forward propagations and upload loss differences back to the server. Furthermore, we also present convergence analyses of BAFFLE, providing guidelines for model design and acceleration of training.", + "bbox": [ + 72, + 201, + 922, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Unbiased Gradient Estimation with Stein's Identity", + "text_level": 1, + "bbox": [ + 75, + 486, + 774, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous work on sign-based optimization [66] demonstrates that deep networks can be effectively trained if the majority of gradients have proper signs. Thus, we propose performing forward propagation multiple times on perturbed parameters, in order to obtain a stochastic estimation of gradients without backpropagation. Specifically, assuming that the loss function $\\mathcal{L}(\\mathbf{W};\\mathbb{D})$ is continuously differentiable w.r.t. $\\mathbf{W}$ given any dataset $\\mathbb{D}$ , which is true (almost everywhere) for deep networks using non-linear activation functions, we define a smoothed loss function as:", + "bbox": [ + 72, + 518, + 920, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\sigma} (\\mathbf {W}; \\mathbb {D}) := \\mathbb {E} _ {\\boldsymbol {\\delta} \\sim \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I})} \\mathcal {L} (\\mathbf {W} + \\boldsymbol {\\delta}; \\mathbb {D}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 689, + 912, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the perturbation $\\delta$ follows a Gaussian distribution with mean 0 and covariance $\\sigma^2\\mathbf{I}$ . Stein [80] proves the Stein's identity (proof recapped in Appendix A):", + "bbox": [ + 75, + 723, + 917, + 764 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {W}} \\mathcal {L} _ {\\sigma} (\\mathbf {W}; \\mathbb {D}) = \\mathbb {E} _ {\\boldsymbol {\\delta} \\sim \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I})} \\left[ \\frac {\\boldsymbol {\\delta}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W}, \\boldsymbol {\\delta}; \\mathbb {D}) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 237, + 777, + 912, + 817 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varDelta\\mathcal{L}(\\mathbf{W},\\boldsymbol {\\delta};\\mathbb{D}):= \\mathcal{L}(\\mathbf{W} + \\boldsymbol {\\delta};\\mathbb{D}) - \\mathcal{L}(\\mathbf{W} - \\boldsymbol {\\delta};\\mathbb{D})$ is the loss difference. Note that computing a loss difference only requires the execution of two forwards $\\mathcal{L}(\\mathbf{W}+$ $\\pmb{\\delta};\\mathbb{D})$ and $\\mathcal{L}(\\mathbf{W} - \\pmb {\\delta};\\mathbb{D})$ without backpropagation. It is trivial that $\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})$ is continuously differentiable for any $\\sigma \\geq 0$ and $\\nabla_{\\mathbf{W}}\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})$ converges uniformly as $\\sigma \\rightarrow 0$ ; hence, it follows that $\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}) = \\lim_{\\sigma \\to 0}\\nabla_{\\mathbf{W}}\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})$ . Therefore, we can obtain a stochastic estimation of gradients using Monte Carlo", + "bbox": [ + 75, + 830, + 920, + 949 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by 1) selecting a small value of $\\sigma$ ; 2) randomly sampling $K$ perturbations from $\\mathcal{N}(0, \\sigma^2 \\mathbf{I})$ as $\\{\\pmb{\\delta}_k\\}_{k=1}^K$ ; and 3) utilizing the Stein's identity in Eq. (5) to calculate", + "bbox": [ + 72, + 55, + 920, + 95 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\nabla_ {\\mathbf {W}}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) := \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left[ \\frac {\\delta_ {k}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W}, \\delta_ {k}; \\mathbb {D}) \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 110, + 917, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Operating Flow of BAFFLE", + "text_level": 1, + "bbox": [ + 75, + 188, + 489, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the forward-only gradient estimator $\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})$ derived in Eq. (6), we outline the basic operating flow of our BAFFLE system (Algorithm 1) as follows:", + "bbox": [ + 72, + 219, + 917, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model initialization. (Lines $3\\sim 4$ , done by server) The server initializes the model parameters to $\\mathbf{W}_0$ and optionally encodes the computing paradigm of loss differences $\\Delta \\mathcal{L}(\\mathbf{W},\\delta ;\\mathbb{D})$ into the TEE module (see Appendix B for more information on TEE);", + "bbox": [ + 72, + 258, + 920, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Loading paradigm:** (Lines 6~7, server $\\Rightarrow$ all clients) In round $t$ , the server distributes the most recent model parameters $\\mathbf{W}_t$ (or the model update $\\Delta \\mathbf{W}_t$ ) and the computing paradigm to all the $C$ clients. In addition, in BAFFLE, the server sends a random seed $s_t$ (rather than directly sending the perturbations to reduce communication burden);", + "bbox": [ + 72, + 336, + 917, + 432 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Local computation. (Lines 11~12, done by clients) Each client generates $K$ perturbations $\\{\\delta_k\\}_{k=1}^K$ locally from $\\mathcal{N}(0, \\sigma^2\\mathbf{I})$ using random seed $s_t$ , and executes the computing paradigm to obtain loss differences. $K$ is chosen adaptively based on clients' computation capability;", + "bbox": [ + 72, + 432, + 917, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Uploading loss differences. (Line 13, all clients $\\Rightarrow$ server) Each client uploads $K$ noisy outputs $\\{\\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k;\\mathbb{D}_c) + \\frac{N}{N_c}\\pmb {\\epsilon}_c\\}_{k = 1}^K$ to the server, where each output is a floating-point number and the noise $\\pmb{\\epsilon}_{c}$ is negotiated by all clients to be zero-sum. The total uploaded Bytes is $4\\times K$", + "bbox": [ + 72, + 510, + 917, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Secure aggregation. (Lines 15~16, done by server) In order to prevent the server from recovering the exact loss differences and causing privacy leakage [25], we adopt the secure aggregation method [10] that was originally proposed for conventional FL and apply it to BAFFLE. Specifically, all clients negotiate a group of noises $\\{\\pmb{\\epsilon}_c\\}_{c=1}^C$ satisfying $\\sum_{c=1}^{C} \\pmb{\\epsilon}_c = 0$ . Then we can reorganize our gradient estimator as", + "bbox": [ + 72, + 588, + 917, + 705 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\nabla_ {\\mathbf {W} _ {t}}} \\mathcal {L} (\\mathbf {W} _ {t}) = \\frac {1}{K} \\sum_ {c = 1} ^ {C} \\frac {N _ {c}}{N} \\sum_ {k = 1} ^ {K} \\left[ \\frac {\\boldsymbol {\\delta} _ {k}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}; \\mathbb {D} _ {c}) \\right] = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {\\boldsymbol {\\delta} _ {k}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}),\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 720, + 910, + 770 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathcal {L} \\left(\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}\\right) = \\sum_ {c = 1} ^ {C} \\frac {N _ {c}}{N} \\left[ \\Delta \\mathcal {L} \\left(\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}; \\mathbb {D} _ {c}\\right) + \\frac {N}{N _ {c}} \\boldsymbol {\\epsilon} _ {c} \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 82, + 777, + 917, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since $\\{\\pmb{\\epsilon}_c\\}_{c=1}^C$ are zero-sum, there is $\\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k) = \\sum_{c=1}^C \\frac{N_c}{N} \\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k; \\mathbb{D}_c)$ and Eq. (7) holds. Therefore, the server can correctly aggregate $\\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k)$ and protect client privacy against recovering $\\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k; \\mathbb{D}_c)$ .", + "bbox": [ + 72, + 845, + 917, + 905 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark on communication cost. After getting the gradient estimation $\\widehat{\\nabla_{\\mathbf{W}_t}}\\mathcal{L}(\\mathbf{W}_t)$ , the server updates the parameters to $\\mathbf{W}_{t + 1}$ using techniques such", + "bbox": [ + 75, + 907, + 917, + 949 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 77, + 16, + 94, + 30 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as gradient descent with learning rate $\\eta$ . Similar to the discussion in McMahan et al. [62], the BAFFLE form presented in Algorithm 1 corresponds to the batch-level communication (also named FedSGD) where Lines $11\\sim 12$ execute once for each round $t$ . In batch-level settings, we reduce the uploaded Bytes from $4\\times |\\mathbf{W}|$ to $4\\times K$ . We can generalize BAFFLE to an analog of epoch-level communication (also named FedAvg), in which each client updates its local parameters multiple steps using the gradient estimator $\\widehat{\\nabla_{\\mathbf{W}_t}}\\mathcal{L}(\\mathbf{W}_t,\\mathbb{D}_c)$ derived from $\\Delta \\mathcal{L}(\\mathbf{W}_t,\\delta_k;\\mathbb{D}_c)$ via Eq. (6), and upload model updates to the server after several local epochs. In epoch-level settings, the uploaded Bytes are the same as FedAvg. In experiments, we analyze both batch-level and epoch-level settings for BAFFLE and report the results.", + "bbox": [ + 72, + 55, + 922, + 274 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Convergence Analyses", + "text_level": 1, + "bbox": [ + 75, + 295, + 416, + 317 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Now we analyze the convergence rate of our gradient estimation method. For continuously differentiable loss functions, we have $\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}) = \\lim_{\\sigma \\to 0}\\nabla_{\\mathbf{W}}\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})$ so we choose a relatively small value for $\\sigma$ . The convergence guarantee can be derived as follows:", + "bbox": [ + 72, + 323, + 929, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 1. (Proof in Appendix A) Suppose $\\sigma$ is a small value and the central difference scheme in Eq. (2) holds. For perturbations $\\{\\pmb{\\delta}_k\\}_{k=1}^K \\stackrel{\\mathrm{iid}}{\\sim} \\mathcal{N}(0, \\sigma^2 \\mathbf{I})$ , the empirical covariance matrix is $\\widehat{\\pmb{\\Sigma}} := \\frac{1}{K \\sigma^2} \\sum_{k=1}^K \\pmb{\\delta}_k \\pmb{\\delta}_k^T$ and mean is $\\widehat{\\pmb{\\delta}} := \\frac{1}{K} \\sum_{k=1}^K \\pmb{\\delta}_k$ . Then for any $\\mathbf{W} \\in \\mathbb{R}^n$ , the relation between $\\widehat{\\nabla_{\\mathbf{W}}} \\mathcal{L}(\\mathbf{W}; \\mathbb{D})$ and the true gradient $\\nabla_{\\mathbf{W}} \\mathcal{L}(\\mathbf{W}; \\mathbb{D})$ can be written as", + "bbox": [ + 75, + 413, + 922, + 523 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\nabla_ {\\mathbf {W}}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) = \\widehat {\\boldsymbol {\\Sigma}} \\nabla_ {\\mathbf {W}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) + o (\\widehat {\\boldsymbol {\\delta}}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 531, + 917, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbb{E}[\\widehat{\\pmb{\\Sigma}}] = \\mathbf{I}$ , $\\mathbb{E}[\\widehat{\\boldsymbol{\\delta}}] = \\mathbf{0}$ .", + "bbox": [ + 77, + 562, + 355, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Taking the expectation on both sides of Eq. (8), we obtain $\\mathbb{E}[\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})] = \\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})$ , which degrades to Stein's identity. To determine the convergence rate w.r.t. the value of $K$ , we have:", + "bbox": [ + 75, + 593, + 917, + 653 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 2. (Adamczak et al. [3]) With overwhelming probability, the empirical covariance matrix satisfies the inequality $\\| \\widehat{\\mathbf{\\Sigma}} -\\mathbf{I}\\| _2\\leq C_0\\sqrt{\\frac{n}{K}}$ , where $\\| \\cdot \\| _2$ denotes the 2-norm for matrix and $C_0$ is an absolute positive constant.", + "bbox": [ + 75, + 658, + 917, + 721 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that in the finetuning setting, $n$ represents the number of trainable parameters, excluding frozen parameters. As concluded, $\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})$ provides an unbiased estimation for the true gradients with convergence rate of $\\mathcal{O}\\left(\\sqrt{\\frac{n}{K}}\\right)$ . Empirically, $\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})$ is used as a noisy gradient to train models, the generalization of which has been analyzed in previous work [50, 103].", + "bbox": [ + 72, + 726, + 917, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 75, + 853, + 309, + 876 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate BAFFLE on 4 benchmark datasets: MNIST [49], CIFAR-10/100 [48] and OfficeHome [87]. We consider three models: 1) LeNet [49] with two convolutional layers as the shallow model ( $2.7 \\times 10^{4}$ parameters); 2) WideResNet [98]", + "bbox": [ + 72, + 889, + 917, + 949 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: The classification accuracy (\\%) of BAFFLE in iid scenarios ( $C = 10$ ) and epoch-level communication settings with different $K$ values ( $K_{1} / K_{2}$ annotations mean using $K_{1}$ for MNIST and $K_{2}$ for CIFAR-10/100). In this configuration, each client updates its local model based on BAFFLE estimated gradients and uploads model updates to the server after an entire epoch on the local dataset. The four guidelines work well under epoch-level settings with total communication rounds 20/40 for MNIST and CIFAR-10/100.", + "bbox": [ + 72, + 52, + 922, + 180 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/272dd137ddef27dfa351b0f255b725dbe4aad0f48d579d99e59ba8e12453bfb4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsLeNetWRN-10-2
MNISTCIFAR-10CIFAR-100MNISTCIFAR-10CIFAR-100
K100/20087.2748.7841.5488.3552.2746.61
200/50089.4851.8245.6889.5755.5951.65
500/100092.1853.6248.7295.1758.6353.15
Ablation Study (100/200)w/o EMA85.0647.9736.8185.8950.0145.86
ReLU81.5544.9939.4979.0849.7644.44
SELU86.1848.6537.3476.4443.3741.79
Central76.0245.9736.5377.4542.8939.62
BP Baselines94.3158.7554.6797.1162.2960.08
", + "bbox": [ + 77, + 195, + 922, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: The accuracy $(\\%)$ of BAFFLE in label non-iid scenarios $(C = 100)$ and epoch-level settings with total communication rounds 40 and different $K$ values. We employ Dirichlet dist. with $\\alpha = 0.3$ to ensure each client has a unique label distribution.", + "bbox": [ + 72, + 435, + 922, + 492 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/255d089263d364a516fe2eb7029b8761782ba890565c9e78716249d99b44bd95.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsLeNet CIFAR-10 | CIFAR-100WRN-10-2 CIFAR-10 | CIFAR-100
K20035.21 | 28.1239.53 | 30.44
50038.14 | 30.9241.69 | 32.89
100039.71 | 33.3543.42 | 34.08
BP Baselines44.41 | 38.4351.18 | 40.85
", + "bbox": [ + 121, + 505, + 876, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with depth $= 10$ and width $= 2$ (WRN-10-2) as the light weight deep model $(3.0 \\times 10^{5}$ parameters) and 3) MobileNet [38] as the deep neural networks $(1.3 \\times 10^{7}$ parameters) that works on ImageNet.", + "bbox": [ + 75, + 671, + 917, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Participation and communication settings. To perform a comprehensive evaluation of BAFFLE, we simulate three popular FL scenarios [17] with the FedLab [99] participations: iid participations, label non-iid participations and feature non-iid participations. For iid participations, we set the client number $C = 10$ and use uniform distribution to build local datasets. Then we evaluate our BAFFLE on MNIST and CIFAR-10/100 under both batch-level (FedSGD) and epoch-level (FedAvg) communication settings. For label non-iid participations, we set client number $C = 100$ , use Dirichlet distribution with $\\alpha = 0.3$ to build clients. For feature non-iid participations, we build clients from the prevailing domain adaptation dataset OfficeHome, which contains 65 categories from 4 different domains, i.e. Art, Clipart, Product and Real-world. We set the", + "bbox": [ + 72, + 734, + 922, + 949 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 77, + 16, + 94, + 29 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1ce11cb3cbf01404dd6cb86b7f30597656d75f14cfd81cdecf3519eff06f653.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 79, + 60, + 503, + 227 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8da81ede411fd39381a7acb952ac0bf7110674536c8ea0ba8fe804dc5802ba59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 518, + 58, + 910, + 227 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3fe921cd0a5dade0c69db3f6c4176ee080b9b34a98e2fbb6056d6beac6680870.jpg", + "image_caption": [ + "LeNet" + ], + "image_footnote": [], + "bbox": [ + 77, + 235, + 510, + 403 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2568a114b196847cb3766fa2fbf239c6d630b1572d1e8cf6d0cd4dac877203f8.jpg", + "image_caption": [ + "WRN-10-2", + "Fig. 2: The classification accuracy (\\%) of BAFFLE in iid scenarios ( $C = 10$ ) and batch-level communication settings with various $K$ values. We treat the models trained by exact gradients on conventional FL systems as the backpropagation (BP) baselines. On different datasets and architectures, our BAFFLE achieves comparable performance to the exact gradient results with a reasonable $K$ ." + ], + "image_footnote": [], + "bbox": [ + 515, + 235, + 910, + 403 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "total client number to $C = 40$ and generate 10 clients from each domain. As results, we report Top-1 accuracy for MNIST, CIFAR-10 and OfficeHome and Top-5 accuracy for OfficeHome and CIFAR-100.", + "bbox": [ + 72, + 559, + 920, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hyperparameters. Following the settings in Section 2, we use FedAVG to aggregate gradients from multiple clients and use SGD-based optimizer to update global parameters. Specifically, we use Adam [46] to train a random initialized model with $\\beta = (0.9,0.99)$ , learning rate 0.01 and epochs 20/40 for MNIST and CIFAR-10/100. For OfficeHome, we adapt the transfer learning [40] by loading the ImageNet-pretrained model and finetuning the final layers with Adam, but setting learning rate 0.005 and epochs 40. In BAFFLE, the perturbation scale $\\sigma$ and number $K$ are the most important hyperparameters. As shown in Theorem 1, with less noise and more samples, the BAFFLE will obtain more accurate gradients, leading to improved performance. However, there exists a trade-off between accuracy and computational efficiency: an extremely small $\\sigma$ will cause the underflow problem [27] and a large $K$ will increase computational cost. In practice, we empirically set $\\sigma = 10^{-4}$ because it is the smallest value that does not cause numerical problems in all experiments, and works well on edge devices with half-precision floating-point numbers. We also evaluate the impact of $K$ across a broad range from 100 to 5000.", + "bbox": [ + 72, + 637, + 922, + 949 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 840, + 34 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Four Guidelines for BAFFLE", + "text_level": 1, + "bbox": [ + 77, + 53, + 501, + 74 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For a general family of continuously differentiable models, we analyze their convergence rate of BAFFLE in Section 3.3. Since deep networks are usually stacked with multiple linear layers and non-linear activation, this layer linearity can be utilized to improve the accuracy-efficiency trade-off. Combining the linearity property and the unique conditions in edge devices (e.g., small data size and half-precision format), we present four guidelines for model design and training that can increase accuracy without introducing extra computation (Appendix C shows the details of linearity analysis):", + "bbox": [ + 77, + 92, + 920, + 248 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Using twice forward difference (twice-FD) scheme rather than central scheme. Combining difference scheme Eq. (1) and Eq. (2), we find that by executing twice as many forward inferences (i.e. $\\mathbf{W} \\pm \\boldsymbol{\\delta}$ ), the central scheme achieves lower residuals than twice-FD, despite the fact that twice-FD can benefit from additional sample times. With the same forward times (e.g., $2K$ ), determining which scheme performs better is a practical issue. As shown in Appendix C, we find that twice-FD performs better in all experiments, in part because the linearity reduces the benefit from second-order residuals.", + "bbox": [ + 77, + 250, + 917, + 404 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Using Hardswish in BAFFLE. ReLU is effective when the middle features $(h(\\cdot)$ denotes the feature mapping) have the same sign before and after perturbations, i.e. $h(\\mathbf{W} + \\boldsymbol{\\delta})\\cdot h(\\mathbf{W}) > 0$ . Since ReLU is not differentiable at zero, the value jump occurs when the sign of features changes after perturbations, i.e. $h(\\mathbf{W} + \\boldsymbol{\\delta})\\cdot h(\\mathbf{W}) < 0$ . We use Hardswish [37] to overcome this problem as it is continuously differentiable at zero and easy to implement on edge devices.", + "bbox": [ + 77, + 406, + 917, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Using exponential moving average (EMA) to reduce oscillations. As shown in Theorem 1, there exists an zero-mean white-noise $\\widehat{\\delta}$ between the true gradient and our estimation. To smooth out the oscillations caused by white noise, we apply EMA strategies from BYOL [28] to the global parameters, with a smoothing coefficient of 0.995.", + "bbox": [ + 77, + 523, + 917, + 622 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Using GroupNorm as opposed to BatchNorm. On edge devices, the dataset size is typically small, which leads to inaccurate batch statistics estimation and degrades performance when using BatchNorm. Thus we employ GroupNorm [94] to solve this issue.", + "bbox": [ + 77, + 624, + 917, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Performance on IID Clients", + "text_level": 1, + "bbox": [ + 77, + 734, + 481, + 754 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Following the settings in Section 4.1, we evaluate the performance of BAFFLE in the iid scenarios. We reproduce all experiments on the BP-based FL systems with the same settings and use them as the baselines. We refer to the baseline results as exact gradients and report the training process of BAFFLE in Figure 2. The value of $K$ (e.g., 200 for LeNet and 500 for WRN-10-2) is much less than the dimensions of parameter space (e.g., $2.7 \\times 10^4$ for LeNet and $3 \\times 10^5$ for WRN-10-2). Since the convergence rate to the exact gradient is $\\mathcal{O}\\left(\\sqrt{\\frac{n}{K}}\\right)$ , the marginal benefit of increasing $K$ decreases. For instance, increasing $K$ from 2000 to 5000 on CIFAR-10 with WRN-10-2 barely improves accuracy by $2\\%$ . Given that the", + "bbox": [ + 77, + 772, + 920, + 947 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ca68b037c69f060f586106a9393a3e95e4fec742e47881cf4b49cbd035d3419c.jpg", + "table_caption": [ + "Table 3: The Top-1|Top-5 accuracy (\\%) of BAFFLE on OfficeHome with feature non-iid participations ( $C = 40$ ) and epoch-level settings with 40 comm. rounds. We use the pretrained MobileNet, freeze the backbone and finetune the FC layers." + ], + "table_footnote": [], + "table_body": "
SettingsDomainsAvg.
ArtClipartProductReal World
2044.75|69.4652.48|73.8866.63|89.7763.78|87.8356.91|80.24
5047.87|71.3253.43|76.8371.28|91.7467.02|89.9559.90|82.46
K 10050.32|74.4257.43|80.7374.19|93.0269.53|90.4362.87|84.65
20051.42|76.6460.98|86.4176.05|94.4271.51|93.1464.98|87.65
50053.33|77.8562.58|86.6478.84|95.2373.17|93.8566.98|88.40
BP Baselines55.71|80.4365.13|88.6582.44|96.0577.19|95.0470.12|90.04
", + "bbox": [ + 77, + 123, + 905, + 295 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "convergence rate of Gaussian perturbations is $\\mathcal{O}\\left(\\sqrt{\\frac{n}{K}}\\right)$ , the sampling efficiency may be improved by choosing an alternative distribution for perturbations.", + "bbox": [ + 72, + 307, + 917, + 346 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation studies. As depicted in Figure 3, we conduct ablation studies for BAFFLE to evaluate the aforementioned guidelines. In general, twice-FD, Hardswish and EMA can all improve the accuracy. For two difference schemes, we compare the twice-FD to central scheme with the same computation cost and show that the former outperforms the later, demonstrating that linearity reduces the gain from second-order residuals. As to activation functions, Hardswish is superior to ReLU and SELU because it is differentiable at zero and vanishes to zero in the negative part. Moreover, EMA enhances the performance of training strategies by reducing the effect of white noise.", + "bbox": [ + 72, + 347, + 920, + 521 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Communication efficiency. Compared to the batch-level communication settings (FedSGD) in a BP-based FL system, BAFFLE requires each client to upload a $K$ -dimensional vector to the server and downloads the updated global parameters in each communication round. Since $K$ is significantly less than the parameter amounts (e.g., 500 versus 0.3 million), BAFFLE reduces data transfer by approximately half. To reduce communication costs, the prevalent FL system requires each client to perform model optimization on the local training dataset and upload the model updates to the server after a specified number of local epochs. BAFFLE can also perform epoch-level communications by employing an $\\mathcal{O}(n)$ additional memory to store the perturbation in each forward and estimate the local gradient using Eq. (6). Then each client optimizes the local model with SGD and uploads local updates after several epochs. As shown in Table 1, we evaluate the performance of BAFFLE under one-epoch communication settings. As epoch-level communication is more prevalent in the real-world FL, all the following experiments will be conducted in this context. In brief, BAFFLE uploads the same Bytes as BP-based FL in epoch-level communication while the total communication rounds are much less than FedZO [22], e.g. 20 versus 1000 on MNIST.", + "bbox": [ + 72, + 521, + 917, + 853 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Performance on Non-IID Clients", + "text_level": 1, + "bbox": [ + 75, + 879, + 542, + 897 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Following Section 4.1, we evaluate the performance of BAFFLE in both label non-iid and feature non-iid scenarios. For label non-iid scenarios, we use the", + "bbox": [ + 72, + 908, + 920, + 949 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 891, + 16, + 915, + 30 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dd059e2c8e099f1a1b5c7263d248d88bb6a848b6c87d126b74995e61d751f6ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 79, + 60, + 508, + 227 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/25c4bb93dfc6f715f6f970d6fbf027483f6ccd621d92bdbb2ff1c79d5515921f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 518, + 58, + 910, + 227 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/23ebf4cff94ea2c9dd73d7d5bb532ed93041385e3b494647897946d91f7d3668.jpg", + "image_caption": [ + "LeNet" + ], + "image_footnote": [], + "bbox": [ + 77, + 235, + 508, + 403 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6dfb402337254ea1a4544638c5f36452f0047e51115a9cdce482f0671e6fc905.jpg", + "image_caption": [ + "WRN-10-2", + "Fig. 3: The ablation study of BAFFLE guidelines, with $K = 100$ on MNIST and $K = 500$ on CIFAR-10. As seen, twice-FD, Hardswish, and EMA all improve performance without extra computation. EMA reduces oscillations by lessening Gaussian noise." + ], + "image_footnote": [], + "bbox": [ + 515, + 235, + 910, + 403 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "CIFAR-10/100 datasets and employ Dirichlet distribution to ensure that each client has a unique label distribution. We evaluate the performance of BAFFLE with 100 clients and various K values. As seen in Table 2, the model suffers a significant drop in accuracy (e.g., $14\\%$ in CIFAR-10 and $16\\%$ in CIFAR-100) due to the label non-iid effect. For feature non-iid scenarios, we construct clients using the OfficeHome dataset and use MobileNet as the deep model. As seen in Table 3, we use the transfer learning strategy to train MobileNet, i.e., we load the parameters pretrained on ImageNet, freeze the backbone parameters, and retrain the classification layers. The accuracy decrease is approximately $3\\% \\sim 5\\%$ .", + "bbox": [ + 72, + 523, + 920, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Computation Efficiency, Memory and Robustness", + "text_level": 1, + "bbox": [ + 75, + 725, + 748, + 744 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "BAFFLE uses $K$ times forward passes instead of backward. Since the backward pass is about as expensive as two normal forward passes [35] and five single-precision accelerated forward passes [67], BAFFLE results in approximately $\\frac{K}{5}$ times the computation expense of BP-based FL. Although BAFFLE results in $\\frac{K}{5} - 1$ times extra computation cost, we show the cost can be reduced with proper training strategies, e.g., the transfer learning in Table 3 can reduce $K$ to 20 on the MobileNet and the $224 \\times 224$ sized OfficeHome dataset.", + "bbox": [ + 72, + 754, + 917, + 887 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Moreover, BAFFLE can reduce huge memory cost on edge devices with the efficiency in static memory and dynamic memory. The auto-differential framework is used to run BP on deep networks, which requires extra static memory", + "bbox": [ + 72, + 889, + 920, + 949 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d4f5d60b7a6f870d548d06c72cade95d7970fc350285dc2d7bb2ef97e668cad3.jpg", + "table_caption": [ + "Table 4: The GPU memory cost (MB) of vanilla BP and BAFFLE, respectively. 'min~max' denotes the minimum and maximum dynamic memory for BAFFLE. We also report the ratio (\\%) of vanilla BP to BAFFLE's max memory cost." + ], + "table_footnote": [], + "table_body": "
BackboneCIFAR-10/100OfficeHome/ImageNet
BPBAFFLERatioBPBAFFLERatio
LeNet168067~17410.35252786~2017.95
WRN-10-2187875~19610.43342594~2517.32
MobileNet2041102~21710.635271121~2895.48
", + "bbox": [ + 79, + 123, + 915, + 232 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(e.g., 200MB for Caffe [41] and 1GB for Pytorch [72]) and imposes a considerable burden on edge devices such as IoT sensors. Due to the necessity of restoring intermediate states, BP also requires enormous amounts of dynamic memory ( $\\geq$ 5GB for MobileNet [24]). Since BAFFLE only requires inference, we can slice the computation graph and execute the forwards per layer [44]. As shown in Table 4, BAFFLE reduces the memory cost to $5\\% \\sim 10\\%$ by executing layer-by-layer inference. By applying kernel-wise computations, we can further reduce the memory cost to approximately $1\\%$ (e.g., 64MB for MobileNet [85]), which is suitable for scenarios with extremely limited storage resources, such as TEE.", + "bbox": [ + 72, + 258, + 920, + 434 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Recent works exploit TEE to protect models from white-box attacks by preventing model exposure [44]. However, due to the security guarantee, the usable memory of TEE is usually small [85] (e.g., 90MB on Intel SGX for Skylake CPU [61]), which is typically far less than what a backpropagation-based FL system requires. In contrast, BAFFLE can execute in TEE due to its little memory cost (more details are in Appendix B). Membership inference attacks and model inversion attacks need to repeatedly perform model inference and obtain confidence values or classification scores [78, 101]. Given that BAFFLE provides stochastic loss differences $\\Delta \\mathcal{L}(\\mathbf{W},\\delta ;\\mathbb{D})$ associated with the random perturbation $\\delta$ , the off-the-shelf inference attacks may not perform on BAFFLE directly (while adaptively designed attacking strategies are possible to evade BAFFLE). Motivated by differential privacy [1], we further design heuristic experiments to study the information leakage from $\\Delta \\mathcal{L}$ (details in Appendix D). As shown in Figure 5, the $\\Delta \\mathcal{L}$ between real data and random noise is hard to distinguish, indicating it is difficult for attackers to obtain useful information from BAFFLE's outputs.", + "bbox": [ + 72, + 435, + 920, + 728 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Related Work", + "text_level": 1, + "bbox": [ + 75, + 762, + 331, + 783 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Along the research routine of FL, many efforts have been devoted to, e.g., dealing with non-IID distributions [21, 53, 75, 90, 102], multi-task learning [60, 79], and preserving privacy of clients [11, 12, 26, 31, 56, 58, 63, 84]. Below we introduce the work on efficiency and vulnerability in FL following the survey of Kairouz et al. [42], which is more related to this paper.", + "bbox": [ + 72, + 811, + 920, + 908 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Efficiency in FL. It is widely understood that the communication and computational efficiency is a primary bottleneck for deploying FL in practice [6,18,74,", + "bbox": [ + 75, + 908, + 917, + 949 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "89, 92]. Specifically, communicating between the server and clients could be potentially expensive and unreliable. The seminal work of Konečný et al. [47] introduces sparsification and quantization to reduce the communication cost, where several theoretical works investigate the optimal trade-off between the communication cost and model accuracy [2, 7, 15, 30, 100]. Since practical clients usually have slower upload than download bandwidth, much research interest focuses on gradient compression [4, 8, 36, 82]. On the other hand, different methods have been proposed to reduce the computational burden of local clients [16, 29, 33], since these clients are usually edge devices with limited resources. Training paradigms exploiting tensor factorization in FL can also achieve promising performance [45, 59].", + "bbox": [ + 72, + 55, + 922, + 269 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Vulnerability in FL. The characteristic of decentralization in FL is beneficial to protecting data privacy of clients, but in the meanwhile, providing white-box accessibility of model status leaves flexibility for malicious clients to perform poisoning/backdoor attacks [5, 9, 71, 88, 95], model/gradient inversion attacks [25, 39, 101], and membership inference attacks [57, 68, 78]. To alleviate the vulnerability in FL, several defense strategies have been proposed via selecting reliable clients [43], data augmentation [13], update clipping [81], robust training [52], model perturbation [97], detection methods [20, 76], and methods based on differential privacy [93].", + "bbox": [ + 72, + 273, + 922, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 75, + 495, + 510, + 517 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Backpropagation is the gold standard for training deep networks, and it is also utilized by traditional FL systems. However, backpropagation is unsuited for edge devices due to their limited resources and possible lack of reliability. Using zero-order optimization techniques, we explore the possibility of BAFFLE in this paper. We need to specify that there are scenarios in which clients are fully trusted and have sufficient computing and storage resources. In these situations, traditional FL with backpropagation is preferred.", + "bbox": [ + 72, + 556, + 922, + 692 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While our preliminary studies on BAFFLE have generated encouraging results, there are still a number of tough topics to investigate: (i) Compared to the models trained using exact gradients, the accuracy of models trained using BAFFLE is inferior. One reason is that we select small values of $K$ (e.g., 500) relative to the number of model parameters (e.g., $3.0 \\times 10^{5}$ ); another reason is that gradient descent is designed for exact gradients, whereas our noisy gradient estimation may require advanced learning algorithms. (ii) The empirical variance of zero-order gradient estimators affects training convergence in BAFFLE. It is crucial to research variance reduction approaches, such as control variates and non-Gaussian sampling distributions. (iii) Stein's identity is proposed for loss functions with Gaussian noises imposed on model parameters. Intuitively, this smoothness is related to differential privacy in FL, but determining their relationship requires theoretical derivations.", + "bbox": [ + 72, + 695, + 922, + 949 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 77, + 53, + 336, + 74 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This paper is supported by the National Science Foundation of China (62132017), Zhejiang Provincial Natural Science Foundation of China (LD24F020011).", + "bbox": [ + 75, + 89, + 920, + 131 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 77, + 156, + 237, + 177 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abadi, M., Chu, A., Goodfellow, I.J., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, Vienna, Austria, October 24-28, 2016 (2016)", + "2. Acharya, J., Canonne, C.L., Tyagi, H.: Inference under information constraints i: Lower bounds from chi-square contraction. IEEE Transactions on Information Theory (2020)", + "3. Adamczak, R., Litvak, A.E., Pajor, A., Tomczak-Jaegermann, N.: Sharp bounds on the rate of convergence of the empirical covariance matrix. Comptes Rendus Mathematique 349(3-4), 195-200 (2011)", + "4. Alistarh, D., Grubic, D., Li, J., Tomioka, R., Vojnovic, M.: Qsgd: Communication-efficient sgd via gradient quantization and encoding. Advances in Neural Information Processing Systems (NeurIPS) (2017)", + "5. Bagdasaryan, E., Veit, A., Hua, Y., Estrin, D., Shmatikov, V.: How to back-door federated learning. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2020)", + "6. Balakrishnan, R., Li, T., Zhou, T., Himayat, N., Smith, V., Bilmes, J.: Diverse client selection for federated learning via submodular maximization. In: International Conference on Learning Representations (ICLR) (2022)", + "7. Barnes, L.P., Han, Y., Özgür, A.: Lower bounds for learning distributions under communication constraints via fisher information. Journal of Machine Learning Research (JMLR) (2020)", + "8. Basu, D., Data, D., Karakus, C., Diggavi, S.: Qsparse-local-sgd: Distributed sgd with quantization, sparsification and local computations. Advances in Neural Information Processing Systems (NeurIPS) (2019)", + "9. Bhagoji, A.N., Chakraborty, S., Mittal, P., Calo, S.: Analyzing federated learning through an adversarial lens. In: International Conference on Machine Learning (ICML) (2019)", + "10. Bonawitz, K.A., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for privacy-preserving machine learning. In: ACM Conference on Computer and Communications Security (CCS) (2017)", + "11. Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for federated learning on user-held data. arXiv preprint arXiv:1611.04482 (2016)", + "12. Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for privacy-preserving machine learning. In: ACM SIGSAC Conference on Computer and Communications Security (2017)", + "13. Borgnia, E., Geiping, J., Cherepanova, V., Fowl, L., Gupta, A., Ghiasi, A., Huang, F., Goldblum, M., Goldstein, T.: Dp-instahide: Provably defusing poisoning and backdoor attacks with differentially private data augmentations. arXiv preprint arXiv:2103.02079 (2021)" + ], + "bbox": [ + 89, + 191, + 920, + 947 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Bradbury, J., Frostig, R., Hawkins, P., Johnson, M.J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., Zhang, Q.: JAX: composable transformations of Python+NumPy programs (2018), http://github.com/google/jax", + "15. Braverman, M., Garg, A., Ma, T., Nguyen, H.L., Woodruff, D.P.: Communication lower bounds for statistical estimation problems via a distributed data processing inequality. In: ACM Symposium on Theory of Computing (2016)", + "16. Caldas, S., Konečny, J., McMahan, H.B., Talwalkar, A.: Expanding the reach of federated learning by reducing client resource requirements. arXiv preprint arXiv:1812.07210 (2018)", + "17. Caldas, S., Wu, P., Li, T., Konečny, J., McMahan, H.B., Smith, V., Talwalkar, A.: LEAF: A benchmark for federated settings. CoRR (2018)", + "18. Chen, M., Shlezinger, N., Poor, H.V., Eldar, Y.C., Cui, S.: Communication-efficient federated learning. Proceedings of the National Academy of Sciences (2021)", + "19. Chen, Y., Luo, F., Li, T., Xiang, T., Liu, Z., Li, J.: A training-integrity privacy-preserving federated learning scheme with trusted execution environment. Information Sciences (2020)", + "20. Dong, Y., Yang, X., Deng, Z., Pang, T., Xiao, Z., Su, H., Zhu, J.: Black-box detection of backdoor attacks with limited information and data. In: IEEE International Conference on Computer Vision (ICCV) (2021)", + "21. Eichner, H., Koren, T., McMahan, B., Srebro, N., Talwar, K.: Semi-cyclic stochastic gradient descent. In: International Conference on Machine Learning (ICML) (2019)", + "22. Fang, W., Yu, Z., Jiang, Y., Shi, Y., Jones, C.N., Zhou, Y.: Communication-efficient stochastic zeroth-order optimization for federated learning. IEEE Trans. Signal Process. (2022)", + "23. Flaxman, A.D., Kalai, A.T., McMahan, H.B.: Online convex optimization in the bandit setting: gradient descent without a gradient. arXiv preprint cs/0408007 (2004)", + "24. Gao, Y., Liu, Y., Zhang, H., Li, Z., Zhu, Y., Lin, H., Yang, M.: Estimatinggpu memory consumption of deep learning models. In: ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering (2020)", + "25. Geiping, J., Bauermeister, H., Dröge, H., Moeller, M.: Inverting gradients-how easy is it to break privacy in federated learning? In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "26. Ghazi, B., Kumar, R., Manurangsi, P., Pagh, R.: Private counting from anonymous messages: Near-optimal accuracy with vanishing communication overhead. In: International Conference on Machine Learning (ICML) (2020)", + "27. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press (2016)", + "28. Grill, J., Strub, F., Altché, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Doersch, C., Pires, B.A., Guo, Z., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent - A new approach to self-supervised learning. In: NeurIPS 2020 (2020)", + "29. Hamer, J., Mohri, M., Suresh, A.T.: Fedboost: A communication-efficient algorithm for federated learning. In: International Conference on Machine Learning (ICML) (2020)", + "30. Han, Y., Ozgür, A., Weissman, T.: Geometric lower bounds for distributed parameter estimation under communication constraints. In: Conference On Learning Theory (COLT) (2018)" + ], + "bbox": [ + 87, + 56, + 917, + 947 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Hao, M., Li, H., Luo, X., Xu, G., Yang, H., Liu, S.: Efficient and privacy-enhanced federated learning for industrial artificial intelligence. IEEE Transactions on Industrial Informatics (2019)", + "32. Hard, A., Rao, K., Mathews, R., Ramaswamy, S., Beaufays, F., Augenstein, S., Eichner, H., Kiddon, C., Ramage, D.: Federated learning for mobile keyboard prediction. arXiv preprint arXiv:1811.03604 (2018)", + "33. He, C., Annavaram, M., Avestimehr, S.: Group knowledge transfer: Federated learning of large cnns at the edge. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "34. He, D., Shi, W., Li, S., Gao, X., Zhang, J., Bian, J., Wang, L., Liu, T.Y.: Learning physics-informed neural networks without stacked back-propagation. arXiv preprint arXiv:2202.09340 (2022)", + "35. Hinton, G., Srivastava, N.: Csc321: Introduction to neural networks and machine learning. Lecture 10 (2010)", + "36. Horváth, S., Ho, C.Y., Horvath, L., Sahu, A.N., Canini, M., Richtárik, P.: Natural compression for distributed deep learning. arXiv preprint arXiv:1905.10988 (2019)", + "37. Howard, A., Pang, R., Adam, H., Le, Q.V., Sandler, M., Chen, B., Wang, W., Chen, L., Tan, M., Chu, G., Vasudevan, V., Zhu, Y.: Searching for mobilenetv3. In: IEEE International Conference on Computer Vision (ICCV) (2019)", + "38. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR (2017)", + "39. Huang, Y., Gupta, S., Song, Z., Li, K., Arora, S.: Evaluating gradient inversion attacks and defenses in federated learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "40. Huh, M., Agrawal, P., Efros, A.A.: What makes imagenet good for transfer learning? CoRR (2016)", + "41. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R.B., Guadarrama, S., Darrell, T.: Caffe: Convolutional architecture for fast feature embedding. In: Hua, K.A., Rui, Y., Steinmetz, R., Hanjalic, A., Natsev, A., Zhu, W. (eds.) Proceedings of the ACM International Conference on Multimedia (2014)", + "42. Kairouz, P., McMahan, H.B., Avent, B., Bellet, A., Dennis, M., Bhagoji, A.N., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R., et al.: Advances and open problems in federated learning. Foundations and Trends in Machine Learning 14(1-2), 1-210 (2021)", + "43. Kang, J., Xiong, Z., Niyato, D., Zou, Y., Zhang, Y., Guizani, M.: Reliable federated learning for mobile networks. IEEE Wireless Communications (2020)", + "44. Kim, K., Kim, C.H., Rhee, J.J., Yu, X., Chen, H., Tian, D., Lee, B.: Vessels: Efficient and scalable deep learning prediction on trusted processors. In: ACM Symposium on Cloud Computing (2020)", + "45. Kim, Y., Sun, J., Yu, H., Jiang, X.: Federated tensor factorization for computational phenotyping. In: ACM International Conference on Knowledge Discovery and Data Mining (SIGKDD) (2017)", + "46. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: 3rd International Conference on Learning Representations, ICLR 2015 (2015)", + "47. Konečny, J., McMahan, H.B., Yu, F.X., Richtárik, P., Suresh, A.T., Bacon, D.: Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492 (2016)", + "48. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images (2009)" + ], + "bbox": [ + 87, + 56, + 917, + 947 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 891, + 16, + 917, + 30 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of IEEE (1998)", + "50. Li, J., Luo, X., Qiao, M.: On generalization error bounds of noisy gradient methods for non-convex learning. In: International Conference on Learning Representations (ICLR) (2020)", + "51. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering (2020)", + "52. Li, T., Hu, S., Beirami, A., Smith, V.: Ditto: Fair and robust federated learning through personalization. In: International Conference on Machine Learning (ICML) (2021)", + "53. Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data. In: International Conference on Learning Representations (ICLR) (2020)", + "54. Li, Z., Chen, L.: Communication-efficient decentralized zeroth-order method on heterogeneous data. In: WCSP 2021 (2021)", + "55. Liu, S., Chen, P.Y., Kailkhura, B., Zhang, G., Hero III, A.O., Varshney, P.K.: A primer on zeroth-order optimization in signal processing and machine learning: Principals, recent advances, and applications. IEEE Signal Processing Magazine (2020)", + "56. Liu, Y., Suresh, A.T., Yu, F.X.X., Kumar, S., Riley, M.: Learning discrete distributions: user vs item-level privacy. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "57. Luo, X., Wu, Y., Xiao, X., Ooi, B.C.: Feature inference attack on model predictions in vertical federated learning. In: IEEE International Conference on Data Engineering (ICDE) (2021)", + "58. Lyu, L., Yu, H., Ma, X., Sun, L., Zhao, J., Yang, Q., Yu, P.S.: Privacy and robustness in federated learning: Attacks and defenses. IEEE Transactions on Neural Networks and Learning Systems pp. 1-21 (2022)", + "59. Ma, J., Zhang, Q., Lou, J., Ho, J.C., Xiong, L., Jiang, X.: Privacy-preserving tensor factorization for collaborative health data analysis. In: ACM International Conference on Information and Knowledge Management (CIKM) (2019)", + "60. Marfoq, O., Neglia, G., Bellet, A., Kameni, L., Vidal, R.: Federated multi-task learning under a mixture of distributions. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "61. McKeen, F., Alexandrovich, I., Anati, I., Caspi, D., Johnson, S., Leslie-Hurd, R., Rozas, C.: Intel® software guard extensions (intel® sgx) support for dynamic memory management inside an enclave. In: Proceedings of the Hardware and Architectural Support for Security and Privacy (2016)", + "62. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2017)", + "63. McMahan, B., Ramage, D., Talwar, K., Zhang, L.: Learning differentially private recurrent language models. In: International Conference on Learning Representations (ICLR) (2018)", + "64. Mo, F., Haddadi, H., Katevas, K., Marin, E., Perino, D., Courtellis, N.: Ppf: privacy-preserving federated learning with trusted execution environments. In: Annual International Conference on Mobile Systems, Applications, and Services (2021)", + "65. Mondal, A., More, Y., Rooparaghunath, R.H., Gupta, D.: Flatee: Federated learning across trusted execution environments. arXiv preprint arXiv:2111.06867 (2021)" + ], + "bbox": [ + 87, + 56, + 920, + 947 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Moulay, E., Léchappé, V., Plestan, F.: Properties of the sign gradient descent algorithms. Inf. Sci. 492, 29-39 (2019)", + "67. Nakandala, S., Nagrecha, K., Kumar, A., Papakonstantinou, Y.: Incremental and approximate computations for accelerating deep CNN inference. ACM Trans. Database Syst. 45 (2020)", + "68. Nasr, M., Shokri, R., Houmansadr, A.: Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In: IEEE symposium on security and privacy (S&P). IEEE (2019)", + "69. Nesterov, Y., Spokoiny, V.: Random gradient-free minimization of convex functions. Foundations of Computational Mathematics (2017)", + "70. Pang, T., Xu, K., Li, C., Song, Y., Ermon, S., Zhu, J.: Efficient learning of generative models via finite-difference score matching. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "71. Pang, T., Yang, X., Dong, Y., Su, H., Zhu, J.: Accumulative poisoning attacks on real-time data. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "72. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E.Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems 32 (2019)", + "73. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)", + "74. Rothchild, D., Panda, A., Ullah, E., Ivkin, N., Stoica, I., Braverman, V., Gonzalez, J., Arora, R.: Fetchsgd: Communication-efficient federated learning with sketching. In: International Conference on Machine Learning (ICML) (2020)", + "75. Sattler, F., Wiedemann, S., Müller, K.R., Samek, W.: Robust and communication-efficient federated learning from non-iid data. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2019)", + "76. Seetharaman, S., Malaviya, S., KV, R., Shukla, M., Lodha, S.: Influence based defense against data poisoning attacks in online learning. arXiv preprint arXiv:2104.13230 (2021)", + "77. Sharma, H., Park, J., Suda, N., Lai, L., Chau, B., Chandra, V., Esmaeilzadeh, H.: Bit fusion: Bit-level dynamically composable architecture for accelerating deep neural network. In: IEEE Annual International Symposium on Computer Architecture (ISCA) (2018)", + "78. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. In: IEEE symposium on security and privacy (S&P). IEEE (2017)", + "79. Smith, V., Chiang, C.K., Sanjabi, M., Talwalkar, A.S.: Federated multi-task learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2017)", + "80. Stein, C.M.: Estimation of the mean of a multivariate normal distribution. The annals of Statistics (1981)", + "81. Sun, Z., Kairouz, P., Suresh, A.T., McMahan, H.B.: Can you really backdoor federated learning? arXiv preprint arXiv:1911.07963 (2019)", + "82. Suresh, A.T., Felix, X.Y., Kumar, S., McMahan, H.B.: Distributed mean estimation with limited communication. In: International Conference on Machine Learning (ICML) (2017)" + ], + "bbox": [ + 87, + 56, + 920, + 947 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "83. Tramer, F., Boneh, D.: Slalom: Fast, verifiable and private execution of neural networks in trusted hardware. In: International Conference on Learning Representations (ICLR) (2019)", + "84. Truex, S., Baracaldo, N., Anwar, A., Steinke, T., Ludwig, H., Zhang, R., Zhou, Y.: A hybrid approach to privacy-preserving federated learning. In: ACM workshop on Artificial Intelligence and Security (2019)", + "85. Truong, J.B., Gallagher, W., Guo, T., Walls, R.J.: Memory-efficient deep learning inference in trusted execution environments. In: IEEE International Conference on Cloud Engineering (IC2E) (2021)", + "86. Umuroglu, Y., Rasnayake, L., Själander, M.: Bismo: A scalable bit-serial matrix multiplication overlay for reconfigurable computing. In: International Conference on Field Programmable Logic and Applications (FPL) (2018)", + "87. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: CVPR (2017)", + "88. Wang, H., Sreenivasan, K., Rajput, S., Vishwakarma, H., Agarwal, S., Sohn, J.y., Lee, K., Papailiopoulos, D.: Attack of the tails: Yes, you really can backdoor federated learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "89. Wang, H.P., Stich, S., He, Y., Fritz, M.: Prog fed: effective, communication, and computation efficient federated learning by progressive training. In: International Conference on Machine Learning (ICML) (2022)", + "90. Wang, J., Liu, Q., Liang, H., Joshi, G., Poor, H.V.: Tackling the objective inconsistency problem in heterogeneous federated optimization. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "91. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "92. Wang, L., Wang, W., Li, B.: Cmfl: Mitigating communication overhead for federated learning. In: IEEE International Conference on Distributed Computing Systems (ICDCS) (2019)", + "93. Wei, K., Li, J., Ding, M., Ma, C., Yang, H.H., Farokhi, F., Jin, S., Quek, T.Q., Poor, H.V.: Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security (2020)", + "94. Wu, Y., He, K.: Group normalization. Int. J. Comput. Vis. 128(3), 742-755 (2020)", + "95. Xie, C., Huang, K., Chen, P.Y., Li, B.: Distributed backdoor attacks against federated learning. In: International Conference on Learning Representations (ICLR) (2020)", + "96. Yang, T., Andrew, G., Eichner, H., Sun, H., Li, W., Kong, N., Ramage, D., Beaufays, F.: Applied federated learning: Improving google keyboard query suggestions. arXiv preprint arXiv:1812.02903 (2018)", + "97. Yang, X., Feng, Y., Fang, W., Shao, J., Tang, X., Xia, S.T., Lu, R.: An accuracy-lossless perturbation method for defending privacy attacks in federated learning. In: ACM Web Conference (WWW) (2022)", + "98. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: Proceedings of the British Machine Vision Conference 2016 (2016)", + "99. Zeng, D., Liang, S., Hu, X., Wang, H., Xu, Z.: Fedlab: A flexible federated learning framework. arXiv preprint arXiv:2107.11621 (2021)", + "100. Zhang, Y., Duchi, J., Jordan, M.I., Wainwright, M.J.: Information-theoretic lower bounds for distributed statistical estimation with communication constraints. Advances in Neural Information Processing Systems (NeurIPS) (2013)" + ], + "bbox": [ + 79, + 56, + 920, + 949 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Haozhe Feng, Tianyu Pang et al.", + "bbox": [ + 157, + 14, + 481, + 32 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "101. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "102. Zhao, Y., Li, M., Lai, L., Suda, N., Civin, D., Chandra, V.: Federated learning with non-iid data. arXiv preprint arXiv:1806.00582 (2018)", + "103. Zhu, Z., Wu, J., Yu, B., Wu, L., Ma, J.: The anisotropic noise in stochastic gradient descent: Its behavior of escaping from sharp minima and regularization effects. In: International Conference on Machine Learning (ICML) (2019)" + ], + "bbox": [ + 79, + 55, + 917, + 201 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", + "bbox": [ + 181, + 14, + 837, + 34 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 891, + 16, + 915, + 30 + ], + "page_idx": 20 + } +] \ No newline at end of file diff --git a/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_model.json b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..644fb24b225853c8ad963085ee5b1f77000346bb --- /dev/null +++ b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_model.json @@ -0,0 +1,3278 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.11, + 0.049, + 0.892, + 0.108 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.14, + 0.819, + 0.182 + ], + "angle": 0, + "content": "Haozhe Feng\\(^{1,3\\star}\\), Tianyu Pang\\(^{2\\star\\star}\\), Chao Du\\(^{2}\\), Wei Chen\\(^{1\\star\\star}\\), Shuicheng Yan\\(^{2}\\), and Min Lin\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.196, + 0.751, + 0.288 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Lab of CAD&CG, Zhejiang University \\(^{2}\\) Sea AI Lab, Singapore \\(^{3}\\) Tencent Inc. {fenghz, chenvis} @zju.edu.cn {tianyupang, duchao, yansc, linmin} @sea.com" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.34, + 0.851, + 0.626 + ], + "angle": 0, + "content": "Abstract. Federated learning (FL) is a general principle for decentralized clients to train a server model collectively without sharing local data. FL is a promising framework with practical applications, but its standard training paradigm requires the clients to backpropagate through the model to compute gradients. Since these clients are typically edge devices and not fully trusted, executing backpropagation on them incurs computational and storage overhead as well as white-box vulnerability. In light of this, we develop backpropagation-free federated learning, dubbed BAFFLE, in which backpropagation is replaced by multiple forward processes to estimate gradients. BAFFLE is 1) memory-efficient and easily fits uploading bandwidth; 2) compatible with inference-only hardware optimization and model quantization or pruning; and 3) well-suited to trusted execution environments, because the clients in BAFFLE only execute forward propagation and return a set of scalars to the server. Empirically we use BAFFLE to train deep models from scratch or to finetune pretrained models, achieving acceptable results." + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.644, + 0.851, + 0.681 + ], + "angle": 0, + "content": "Keywords: Federated Learning \\(\\cdot\\) Finite Difference \\(\\cdot\\) Backpropagation-Free Learning" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.717, + 0.31, + 0.736 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.761, + 0.92, + 0.898 + ], + "angle": 0, + "content": "Federated learning (FL) allows decentralized clients to collaboratively train a server model [62]. In each training round, the selected clients compute model gradients or updates on their local private datasets, without explicitly exchanging sample points to the server. While FL describes a promising blueprint and has several applications [32, 51, 96], the mainstream training paradigm of FL is still gradient-based that requires the clients to locally execute backpropagation, which leads to two practical limitations:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.914, + 0.568, + 0.931 + ], + "angle": 0, + "content": "* Work done during an internship at Sea AI Lab." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.932, + 0.327, + 0.949 + ], + "angle": 0, + "content": "\\*\\* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.914, + 0.568, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.031 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.484, + 0.034 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.923, + 0.213 + ], + "angle": 0, + "content": "(i) Overhead for edge devices. The clients in FL are usually edge devices, such as mobile phones and IoT sensors, whose hardware is primarily optimized for inference-only purposes [77,86], rather than for backpropagation. Due to the limited resources, computationally affordable models running on edge devices are typically quantized and pruned [91], making exact backpropagation difficult. In addition, standard implementations of backpropagation rely on either forward-mode or reverse-mode auto-differentiation in contemporary machine learning packages [14, 73], which increases storage requirements." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.221, + 0.923, + 0.378 + ], + "angle": 0, + "content": "(ii) White-box vulnerability. To facilitate gradient computing, the server regularly distributes its model status to the clients, but this white-box exposure of the model renders the server vulnerable to, e.g., poisoning or inversion attacks from malicious clients [25,78,95,101]. With that, recent attempts are made to exploit trusted execution environments (TEEs) in FL, which can isolate the model status within a black-box secure area and significantly reduce the success rate of malicious evasion [19,64,65]. However, TEEs are highly memory-constrained [85], while backpropagation is memory-consuming to restore intermediate states." + }, + { + "type": "list", + "bbox": [ + 0.075, + 0.056, + 0.923, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.386, + 0.924, + 0.621 + ], + "angle": 0, + "content": "While numerous solutions have been proposed to alleviate these limitations (related work discussed in Section 5), we raise an essential question: how to perform backpropagation-free \\( FL \\)? Inspired by the literature on zero-order optimization [80], we intend to substitute backpropagation with multiple forward or inference processes to estimate the gradients. Technically speaking, we propose the framework of BAckpropagation-Free Federated LEarning (BAFFLE). As illustrated in Figure 1, BAFFLE consists of three conceptual steps: (1) each client locally perturbs the model parameters \\( 2K \\) times as \\( \\mathbf{W} \\pm \\boldsymbol{\\delta}_k \\) (the server sends the random seed to clients for generating \\( \\{\\pmb{\\delta}_k\\}_{k=1}^K \\)); (2) each client executes forward processes on the perturbed models using its private dataset \\( \\mathbb{D}_c \\) and obtains \\( K \\) loss differences \\( \\{\\Delta \\mathcal{L}(\\mathbf{W}, \\pmb{\\delta}_k; \\mathbb{D}_c)\\}_{k=1}^K \\); (3) the server aggregates loss differences to estimate gradients." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.924, + 0.825 + ], + "angle": 0, + "content": "BAFFLE's defining characteristic is that it only utilizes forward propagation, which is memory-efficient and does not require auto-differentiation. It is well-adapted to model quantization and pruning as well as inference-only hardware optimization on edge devices. Compared to backpropagation, the computation graph of BAFFLE is more easily optimized, such as by slicing it into per-layer calculation [44]. Since each loss difference \\(\\Delta \\mathcal{L}(\\mathbf{W},\\delta_k;\\mathbb{D}_c)\\) is a scalar, BAFFLE can easily accommodate the uploading bandwidth of clients by adjusting the value of \\(K\\) as opposed to using, e.g., gradient compression [82]. BAFFLE is also compatible with recent advances in inference approaches for TEE [83, 85], providing an efficient solution for combining TEE into FL and preventing white-box evasion." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.832, + 0.924, + 0.95 + ], + "angle": 0, + "content": "We adapt secure aggregation [10] to zero-order optimization and investigate ways to improve gradient estimation in BAFFLE. Empirically, BAFFLE is used to train models from scratch on MNIST [49] and CIFAR-10/100 [48], and transfer ImageNet-pretrained models to OfficeHome [87]. Compared to conventional FL, it achieves suboptimal but acceptable performance. These results shed light on the potential of BAFFLE and general backpropagation-free methods in FL." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.059, + 0.913, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.92, + 0.427 + ], + "angle": 0, + "content": "Fig. 1: A sketch map of BAFFLE. In addition to the global parameters update \\(\\Delta \\mathbf{W}\\), each client downloads random seeds to locally generate perturbations \\(\\pm \\delta_{1:K}\\) and perform \\(2K\\) times of forward propagation (i.e., inference) to compute loss differences. The server can recover these perturbations using the same random seeds and obtain \\(\\Delta \\mathcal{L}(\\mathbf{W},\\pmb{\\delta}_k)\\) by secure aggregation. Each loss difference \\(\\Delta \\mathcal{L}(\\mathbf{W},\\pmb{\\delta}_k;\\mathbb{D}_c)\\) is a floating-point number, so \\(K\\) can be easily adjusted to fit the uploading bandwidth." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.464, + 0.322, + 0.484 + ], + "angle": 0, + "content": "2 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.509, + 0.923, + 0.704 + ], + "angle": 0, + "content": "Finite difference. Gradient-based optimization techniques (either first-order or higher-order) are the most frequently used tools to train deep networks [27]. Nevertheless, recent progress demonstrates promising applications of zero-order optimization methods for training, particularly when exact derivatives cannot be obtained [23, 55, 69] or backward processes are computationally prohibitive [34, 70]. Zero-order approaches require only multiple forward processes that may be executed in parallel. Along this routine, finite difference stems from the definition of derivatives and can be generalized to higher-order and multivariate cases by Taylor's expansion. For any differentiable loss \\(\\mathcal{L}(\\mathbf{W};\\mathbb{D})\\) and a small perturbation \\(\\delta \\in \\mathbb{R}^n\\), finite difference employs the forward difference scheme" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.718, + 0.919, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {W} + \\boldsymbol {\\delta}; \\mathbb {D}) - \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) = \\boldsymbol {\\delta} ^ {\\top} \\nabla_ {\\mathbf {W}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) + o (\\| \\boldsymbol {\\delta} \\| _ {2}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.919, + 0.798 + ], + "angle": 0, + "content": "where \\(\\delta^{\\top}\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})\\) is a scaled directional derivative along \\(\\pmb{\\delta}\\). Furthermore, we can use the central difference scheme to obtain higher-order residuals as" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.812, + 0.918, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {W} + \\boldsymbol {\\delta}; \\mathbb {D}) - \\mathcal {L} (\\mathbf {W} - \\boldsymbol {\\delta}; \\mathbb {D}) = 2 \\boldsymbol {\\delta} ^ {\\top} \\nabla_ {\\mathbf {W}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) + o (\\| \\boldsymbol {\\delta} \\| _ {2} ^ {2}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.852, + 0.922, + 0.949 + ], + "angle": 0, + "content": "Federated learning. Suppose we have \\(C\\) clients, and the \\(c\\)-th client's private dataset is defined as \\(\\mathbb{D}_c \\coloneqq \\{(\\mathbf{X}_i^c, \\mathbf{y}_i^c)\\}_{i=1}^{N_c}\\) with \\(N_c\\) input-label pairs. Let \\(\\mathcal{L}(\\mathbf{W}; \\mathbb{D}_c)\\) represent the loss function for the dataset \\(\\mathbb{D}_c\\), where \\(\\mathbf{W} \\in \\mathbb{R}^n\\) denotes the server model's global parameters. The training objective of FL is to find \\(\\mathbf{W}\\) that minimizes the total loss function as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.018, + 0.096, + 0.03 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.483, + 0.033 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.081, + 0.057, + 0.784, + 0.076 + ], + "angle": 0, + "content": "Algorithm 1 Backpropagation-free federated learning (BAFFLE)" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.081, + 0.921, + 0.135 + ], + "angle": 0, + "content": "1: Notations: Se denotes the operations done on servers; Cl denotes the operations done on clients; TEE for the TEE module; and \\(\\Rightarrow\\) denotes the communication process." + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.136, + 0.921, + 0.19 + ], + "angle": 0, + "content": "2: Inputs: \\(C\\) clients with local dataset \\(\\{\\mathbb{D}_c\\}_{c = 1}^C\\) containing \\(N_{c}\\) input-label pairs, \\(N = \\sum_{c = 1}^{C}N_{c}\\); learning rate \\(\\eta\\), training iterations \\(T\\), perturbation number \\(K\\), noise scale \\(\\sigma\\)." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.19, + 0.553, + 0.207 + ], + "angle": 0, + "content": "3: Se: initializing model parameters \\(\\mathbf{W} \\gets \\mathbf{W}_0\\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.207, + 0.231, + 0.224 + ], + "angle": 0, + "content": "4: // optional" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.081, + 0.921, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.225, + 0.824, + 0.242 + ], + "angle": 0, + "content": "Se: encoding the computing paradigm into TEE as \\(\\mathrm{TEE} \\circ \\Delta \\mathcal{L}(\\mathbf{W}, \\delta; \\mathbb{D})\\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.244, + 0.336, + 0.259 + ], + "angle": 0, + "content": "5: for \\( t = 0 \\) to \\( T - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.261, + 0.911, + 0.278 + ], + "angle": 0, + "content": "6: \\(\\mathrm{Se} \\Rightarrow\\) all Cl: downloading model parameters \\(\\mathbf{W}_t\\) and the computing paradigm;" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.279, + 0.248, + 0.296 + ], + "angle": 0, + "content": "7: // 4 Bytes" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.244, + 0.911, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.297, + 0.593, + 0.313 + ], + "angle": 0, + "content": "\\(\\mathrm{Se} \\Rightarrow\\) all Cl: downloading the random seed \\(s_t\\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.314, + 0.912, + 0.332 + ], + "angle": 0, + "content": "8: Se: sampling \\(K\\) perturbations \\(\\{\\pmb{\\delta}_k\\}_{k=1}^K\\) from \\(\\mathcal{N}(0, \\sigma^2\\mathbf{I})\\) using the random seed \\(s_t\\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.333, + 0.895, + 0.349 + ], + "angle": 0, + "content": "9: all Cl: negotiating a group of zero-sum noises \\(\\{\\pmb{\\epsilon}_c\\}_{c=1}^C\\) for secure aggregation;" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.314, + 0.912, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.35, + 0.326, + 0.365 + ], + "angle": 0, + "content": "10: for \\(c = 1\\) to \\(C\\) do" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.367, + 0.918, + 0.401 + ], + "angle": 0, + "content": "11: Cl: sampling \\(K\\) perturbations \\(\\{\\pmb {\\delta}_k\\}_{k = 1}^K\\) from \\(\\mathcal{N}(0,\\sigma^2\\mathbf{I})\\) using the random seed \\(s_t\\);" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.403, + 0.889, + 0.421 + ], + "angle": 0, + "content": "12: Cl: computing TEE \\(\\circ \\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k;\\mathbb{D}_c)\\) via forward propagation for each \\(k\\)" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.367, + 0.918, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.422, + 0.321, + 0.438 + ], + "angle": 0, + "content": "13: // \\(4\\times K\\) Bytes" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.439, + 0.842, + 0.47 + ], + "angle": 0, + "content": "\\(\\mathrm{Cl} \\Rightarrow \\mathrm{Se}\\): uploading \\(K\\) outputs \\(\\left\\{\\mathrm{TEE} \\circ \\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k; \\mathbb{D}_c) + \\frac{N}{N_c} \\pmb{\\epsilon}_c\\right\\}_{k=1}^K\\);" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.47, + 0.226, + 0.482 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.484, + 0.918, + 0.534 + ], + "angle": 0, + "content": "15: Se: aggregating \\(\\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k)\\gets \\sum_{c = 1}^{C}\\frac{N_c}{N}\\left[\\mathrm{TEE}\\circ \\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k;\\mathbb{D}_c) + \\frac{N}{N_c}\\pmb {\\epsilon}_c\\right]\\) for each \\(k\\)" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.535, + 0.662, + 0.571 + ], + "angle": 0, + "content": "16: Se: computing \\(\\widehat{\\nabla_{\\mathbf{W}_t}}\\mathcal{L}(\\mathbf{W}_t)\\gets \\frac{1}{K}\\sum_{k = 1}^{K}\\frac{\\delta_k}{2\\sigma^2}\\varDelta\\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k);\\)" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.572, + 0.469, + 0.593 + ], + "angle": 0, + "content": "17: Se: \\(\\mathbf{W}_{t + 1}\\gets \\mathbf{W}_t - \\eta \\widehat{\\nabla\\mathbf{w}_t}\\mathcal{L}(\\mathbf{W}_t)\\)" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.594, + 0.201, + 0.608 + ], + "angle": 0, + "content": "18: end for" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.611, + 0.493, + 0.628 + ], + "angle": 0, + "content": "19: Return: final model parameters \\(\\mathbf{W}_T\\)." + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.47, + 0.918, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.248, + 0.688, + 0.918, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {W}) := \\sum_ {c = 1} ^ {C} \\frac {N _ {c}}{N} \\mathcal {L} (\\mathbf {W}; \\mathbb {D} _ {c}), \\text {w h e r e} N = \\sum_ {c = 1} ^ {C} N _ {c}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.751, + 0.922, + 0.852 + ], + "angle": 0, + "content": "In the conventional FL framework, clients compute gradients \\(\\{\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}_c)\\}_{c = 1}^C\\) or model updates locally through backpropagation and then upload them to the server. Federated average [62] performs global aggregation using \\(\\varDelta\\mathbf{W}:= \\sum_{i=1}^{C}\\frac{N_c}{N}\\varDelta\\mathbf{W}_c\\), where \\(\\varDelta\\mathbf{W}_c\\) is the local update obtained via executing \\(\\mathbf{W}_c\\gets \\mathbf{W}_c - \\eta \\nabla_{\\mathbf{W}_c}\\mathcal{L}(\\mathbf{W}_c;\\mathbb{D}_c)\\) multiple times and \\(\\eta\\) is learning rate." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.853, + 0.927, + 0.95 + ], + "angle": 0, + "content": "Zeroth-order FL. Similar to our work, DLZO [54] and FedZO [22] present zeroth-order optimization methods for FL independently in batch-level and epoch-level communications. However, they concentrate primarily on basic linear models with softmax regression problems and ignore deep models. Besides, they also do not account for server security aggregation in conjunction with zero-order" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.015, + 0.839, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.921, + 0.136 + ], + "angle": 0, + "content": "optimization. In comparison, BAFFLE enables security aggregation, can train deep models such as WideResNet from scratch and achieves reasonable results, e.g. \\(95.17\\%\\) accuracy on MNIST with 20 communication rounds versus \\(83.58\\%\\) for FedZO with 1,000 rounds." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.164, + 0.731, + 0.187 + ], + "angle": 0, + "content": "3 Backpropagation-Free Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.203, + 0.923, + 0.461 + ], + "angle": 0, + "content": "In this section, we introduce zero-order optimization into FL and develop BAF-FLE, a backpropagation-free federated learning framework that uses multiple forward processes in place of backpropagation. An initial attempt is to apply finite difference as the gradient estimator. To estimate the full gradients, we need to perturb each parameter \\( w \\in \\mathbf{W} \\) once to approximate the partial derivative \\( \\frac{\\partial \\mathcal{L}(\\mathbf{W}; \\mathbb{D})}{\\partial w} \\), causing the forward computations to grow with \\( n \\) (recall that \\( \\mathbf{W} \\in \\mathbb{R}^n \\)) and making it difficult to scale to large models. In light of this, we resort to Stein's identity [80] to obtain an unbiased estimation of gradients from loss differences calculated on various perturbations. As depicted in Figure 1, BAFFLE clients need only download random seeds and global parameters update, generate perturbations locally, execute multiple forward propagations and upload loss differences back to the server. Furthermore, we also present convergence analyses of BAFFLE, providing guidelines for model design and acceleration of training." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.487, + 0.776, + 0.508 + ], + "angle": 0, + "content": "3.1 Unbiased Gradient Estimation with Stein's Identity" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.519, + 0.921, + 0.675 + ], + "angle": 0, + "content": "Previous work on sign-based optimization [66] demonstrates that deep networks can be effectively trained if the majority of gradients have proper signs. Thus, we propose performing forward propagation multiple times on perturbed parameters, in order to obtain a stochastic estimation of gradients without backpropagation. Specifically, assuming that the loss function \\(\\mathcal{L}(\\mathbf{W};\\mathbb{D})\\) is continuously differentiable w.r.t. \\(\\mathbf{W}\\) given any dataset \\(\\mathbb{D}\\), which is true (almost everywhere) for deep networks using non-linear activation functions, we define a smoothed loss function as:" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.691, + 0.915, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\sigma} (\\mathbf {W}; \\mathbb {D}) := \\mathbb {E} _ {\\boldsymbol {\\delta} \\sim \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I})} \\mathcal {L} (\\mathbf {W} + \\boldsymbol {\\delta}; \\mathbb {D}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.92, + 0.765 + ], + "angle": 0, + "content": "where the perturbation \\(\\delta\\) follows a Gaussian distribution with mean 0 and covariance \\(\\sigma^2\\mathbf{I}\\). Stein [80] proves the Stein's identity (proof recapped in Appendix A):" + }, + { + "type": "equation", + "bbox": [ + 0.239, + 0.778, + 0.915, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {W}} \\mathcal {L} _ {\\sigma} (\\mathbf {W}; \\mathbb {D}) = \\mathbb {E} _ {\\boldsymbol {\\delta} \\sim \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I})} \\left[ \\frac {\\boldsymbol {\\delta}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W}, \\boldsymbol {\\delta}; \\mathbb {D}) \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.832, + 0.922, + 0.95 + ], + "angle": 0, + "content": "where \\(\\varDelta\\mathcal{L}(\\mathbf{W},\\boldsymbol {\\delta};\\mathbb{D}):= \\mathcal{L}(\\mathbf{W} + \\boldsymbol {\\delta};\\mathbb{D}) - \\mathcal{L}(\\mathbf{W} - \\boldsymbol {\\delta};\\mathbb{D})\\) is the loss difference. Note that computing a loss difference only requires the execution of two forwards \\(\\mathcal{L}(\\mathbf{W}+\\) \\(\\pmb{\\delta};\\mathbb{D})\\) and \\(\\mathcal{L}(\\mathbf{W} - \\pmb {\\delta};\\mathbb{D})\\) without backpropagation. It is trivial that \\(\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})\\) is continuously differentiable for any \\(\\sigma \\geq 0\\) and \\(\\nabla_{\\mathbf{W}}\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})\\) converges uniformly as \\(\\sigma \\rightarrow 0\\); hence, it follows that \\(\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}) = \\lim_{\\sigma \\to 0}\\nabla_{\\mathbf{W}}\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})\\). Therefore, we can obtain a stochastic estimation of gradients using Monte Carlo" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.078, + 0.017, + 0.096, + 0.031 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.484, + 0.033 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.921, + 0.096 + ], + "angle": 0, + "content": "by 1) selecting a small value of \\(\\sigma\\); 2) randomly sampling \\(K\\) perturbations from \\(\\mathcal{N}(0, \\sigma^2 \\mathbf{I})\\) as \\(\\{\\pmb{\\delta}_k\\}_{k=1}^K\\); and 3) utilizing the Stein's identity in Eq. (5) to calculate" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.112, + 0.92, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\nabla_ {\\mathbf {W}}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) := \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left[ \\frac {\\delta_ {k}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W}, \\delta_ {k}; \\mathbb {D}) \\right]. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.189, + 0.49, + 0.209 + ], + "angle": 0, + "content": "3.2 Operating Flow of BAFFLE" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.22, + 0.92, + 0.259 + ], + "angle": 0, + "content": "Based on the forward-only gradient estimator \\(\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})\\) derived in Eq. (6), we outline the basic operating flow of our BAFFLE system (Algorithm 1) as follows:" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.26, + 0.922, + 0.336 + ], + "angle": 0, + "content": "Model initialization. (Lines \\(3\\sim 4\\), done by server) The server initializes the model parameters to \\(\\mathbf{W}_0\\) and optionally encodes the computing paradigm of loss differences \\(\\Delta \\mathcal{L}(\\mathbf{W},\\delta ;\\mathbb{D})\\) into the TEE module (see Appendix B for more information on TEE);" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.337, + 0.92, + 0.433 + ], + "angle": 0, + "content": "**Loading paradigm:** (Lines 6~7, server \\(\\Rightarrow\\) all clients) In round \\(t\\), the server distributes the most recent model parameters \\(\\mathbf{W}_t\\) (or the model update \\(\\Delta \\mathbf{W}_t\\)) and the computing paradigm to all the \\(C\\) clients. In addition, in BAFFLE, the server sends a random seed \\(s_t\\) (rather than directly sending the perturbations to reduce communication burden);" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.434, + 0.92, + 0.512 + ], + "angle": 0, + "content": "Local computation. (Lines 11~12, done by clients) Each client generates \\(K\\) perturbations \\(\\{\\delta_k\\}_{k=1}^K\\) locally from \\(\\mathcal{N}(0, \\sigma^2\\mathbf{I})\\) using random seed \\(s_t\\), and executes the computing paradigm to obtain loss differences. \\(K\\) is chosen adaptively based on clients' computation capability;" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.512, + 0.92, + 0.589 + ], + "angle": 0, + "content": "Uploading loss differences. (Line 13, all clients \\(\\Rightarrow\\) server) Each client uploads \\(K\\) noisy outputs \\(\\{\\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k;\\mathbb{D}_c) + \\frac{N}{N_c}\\pmb {\\epsilon}_c\\}_{k = 1}^K\\) to the server, where each output is a floating-point number and the noise \\(\\pmb{\\epsilon}_{c}\\) is negotiated by all clients to be zero-sum. The total uploaded Bytes is \\(4\\times K\\)" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.59, + 0.92, + 0.707 + ], + "angle": 0, + "content": "Secure aggregation. (Lines 15~16, done by server) In order to prevent the server from recovering the exact loss differences and causing privacy leakage [25], we adopt the secure aggregation method [10] that was originally proposed for conventional FL and apply it to BAFFLE. Specifically, all clients negotiate a group of noises \\(\\{\\pmb{\\epsilon}_c\\}_{c=1}^C\\) satisfying \\(\\sum_{c=1}^{C} \\pmb{\\epsilon}_c = 0\\). Then we can reorganize our gradient estimator as" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.721, + 0.912, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\nabla_ {\\mathbf {W} _ {t}}} \\mathcal {L} (\\mathbf {W} _ {t}) = \\frac {1}{K} \\sum_ {c = 1} ^ {C} \\frac {N _ {c}}{N} \\sum_ {k = 1} ^ {K} \\left[ \\frac {\\boldsymbol {\\delta} _ {k}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}; \\mathbb {D} _ {c}) \\right] = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {\\boldsymbol {\\delta} _ {k}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.778, + 0.919, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathcal {L} \\left(\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}\\right) = \\sum_ {c = 1} ^ {C} \\frac {N _ {c}}{N} \\left[ \\Delta \\mathcal {L} \\left(\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}; \\mathbb {D} _ {c}\\right) + \\frac {N}{N _ {c}} \\boldsymbol {\\epsilon} _ {c} \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.847, + 0.919, + 0.907 + ], + "angle": 0, + "content": "Since \\(\\{\\pmb{\\epsilon}_c\\}_{c=1}^C\\) are zero-sum, there is \\(\\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k) = \\sum_{c=1}^C \\frac{N_c}{N} \\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k; \\mathbb{D}_c)\\) and Eq. (7) holds. Therefore, the server can correctly aggregate \\(\\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k)\\) and protect client privacy against recovering \\(\\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k; \\mathbb{D}_c)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.908, + 0.92, + 0.95 + ], + "angle": 0, + "content": "Remark on communication cost. After getting the gradient estimation \\(\\widehat{\\nabla_{\\mathbf{W}_t}}\\mathcal{L}(\\mathbf{W}_t)\\), the server updates the parameters to \\(\\mathbf{W}_{t + 1}\\) using techniques such" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.015, + 0.839, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.923, + 0.275 + ], + "angle": 0, + "content": "as gradient descent with learning rate \\(\\eta\\). Similar to the discussion in McMahan et al. [62], the BAFFLE form presented in Algorithm 1 corresponds to the batch-level communication (also named FedSGD) where Lines \\(11\\sim 12\\) execute once for each round \\(t\\). In batch-level settings, we reduce the uploaded Bytes from \\(4\\times |\\mathbf{W}|\\) to \\(4\\times K\\). We can generalize BAFFLE to an analog of epoch-level communication (also named FedAvg), in which each client updates its local parameters multiple steps using the gradient estimator \\(\\widehat{\\nabla_{\\mathbf{W}_t}}\\mathcal{L}(\\mathbf{W}_t,\\mathbb{D}_c)\\) derived from \\(\\Delta \\mathcal{L}(\\mathbf{W}_t,\\delta_k;\\mathbb{D}_c)\\) via Eq. (6), and upload model updates to the server after several local epochs. In epoch-level settings, the uploaded Bytes are the same as FedAvg. In experiments, we analyze both batch-level and epoch-level settings for BAFFLE and report the results." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.297, + 0.417, + 0.318 + ], + "angle": 0, + "content": "3.3 Convergence Analyses" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.325, + 0.931, + 0.407 + ], + "angle": 0, + "content": "Now we analyze the convergence rate of our gradient estimation method. For continuously differentiable loss functions, we have \\(\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}) = \\lim_{\\sigma \\to 0}\\nabla_{\\mathbf{W}}\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})\\) so we choose a relatively small value for \\(\\sigma\\). The convergence guarantee can be derived as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.924, + 0.525 + ], + "angle": 0, + "content": "Theorem 1. (Proof in Appendix A) Suppose \\(\\sigma\\) is a small value and the central difference scheme in Eq. (2) holds. For perturbations \\(\\{\\pmb{\\delta}_k\\}_{k=1}^K \\stackrel{\\mathrm{iid}}{\\sim} \\mathcal{N}(0, \\sigma^2 \\mathbf{I})\\), the empirical covariance matrix is \\(\\widehat{\\pmb{\\Sigma}} := \\frac{1}{K \\sigma^2} \\sum_{k=1}^K \\pmb{\\delta}_k \\pmb{\\delta}_k^T\\) and mean is \\(\\widehat{\\pmb{\\delta}} := \\frac{1}{K} \\sum_{k=1}^K \\pmb{\\delta}_k\\). Then for any \\(\\mathbf{W} \\in \\mathbb{R}^n\\), the relation between \\(\\widehat{\\nabla_{\\mathbf{W}}} \\mathcal{L}(\\mathbf{W}; \\mathbb{D})\\) and the true gradient \\(\\nabla_{\\mathbf{W}} \\mathcal{L}(\\mathbf{W}; \\mathbb{D})\\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.292, + 0.532, + 0.919, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\nabla_ {\\mathbf {W}}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) = \\widehat {\\boldsymbol {\\Sigma}} \\nabla_ {\\mathbf {W}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) + o (\\widehat {\\boldsymbol {\\delta}}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.564, + 0.356, + 0.589 + ], + "angle": 0, + "content": "where \\(\\mathbb{E}[\\widehat{\\pmb{\\Sigma}}] = \\mathbf{I}\\), \\(\\mathbb{E}[\\widehat{\\boldsymbol{\\delta}}] = \\mathbf{0}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.92, + 0.655 + ], + "angle": 0, + "content": "Taking the expectation on both sides of Eq. (8), we obtain \\(\\mathbb{E}[\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})] = \\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})\\), which degrades to Stein's identity. To determine the convergence rate w.r.t. the value of \\(K\\), we have:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.92, + 0.723 + ], + "angle": 0, + "content": "Theorem 2. (Adamczak et al. [3]) With overwhelming probability, the empirical covariance matrix satisfies the inequality \\(\\| \\widehat{\\mathbf{\\Sigma}} -\\mathbf{I}\\| _2\\leq C_0\\sqrt{\\frac{n}{K}}\\), where \\(\\| \\cdot \\| _2\\) denotes the 2-norm for matrix and \\(C_0\\) is an absolute positive constant." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.727, + 0.92, + 0.831 + ], + "angle": 0, + "content": "Note that in the finetuning setting, \\( n \\) represents the number of trainable parameters, excluding frozen parameters. As concluded, \\( \\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}) \\) provides an unbiased estimation for the true gradients with convergence rate of \\( \\mathcal{O}\\left(\\sqrt{\\frac{n}{K}}\\right) \\). Empirically, \\( \\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}) \\) is used as a noisy gradient to train models, the generalization of which has been analyzed in previous work [50, 103]." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.855, + 0.312, + 0.878 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.89, + 0.92, + 0.951 + ], + "angle": 0, + "content": "We evaluate BAFFLE on 4 benchmark datasets: MNIST [49], CIFAR-10/100 [48] and OfficeHome [87]. We consider three models: 1) LeNet [49] with two convolutional layers as the shallow model (\\(2.7 \\times 10^{4}\\) parameters); 2) WideResNet [98]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.03 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.015, + 0.484, + 0.033 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.053, + 0.923, + 0.181 + ], + "angle": 0, + "content": "Table 1: The classification accuracy (\\%) of BAFFLE in iid scenarios (\\( C = 10 \\)) and epoch-level communication settings with different \\( K \\) values (\\( K_{1} / K_{2} \\) annotations mean using \\( K_{1} \\) for MNIST and \\( K_{2} \\) for CIFAR-10/100). In this configuration, each client updates its local model based on BAFFLE estimated gradients and uploads model updates to the server after an entire epoch on the local dataset. The four guidelines work well under epoch-level settings with total communication rounds 20/40 for MNIST and CIFAR-10/100." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.196, + 0.923, + 0.413 + ], + "angle": 0, + "content": "
SettingsLeNetWRN-10-2
MNISTCIFAR-10CIFAR-100MNISTCIFAR-10CIFAR-100
K100/20087.2748.7841.5488.3552.2746.61
200/50089.4851.8245.6889.5755.5951.65
500/100092.1853.6248.7295.1758.6353.15
Ablation Study (100/200)w/o EMA85.0647.9736.8185.8950.0145.86
ReLU81.5544.9939.4979.0849.7644.44
SELU86.1848.6537.3476.4443.3741.79
Central76.0245.9736.5377.4542.8939.62
BP Baselines94.3158.7554.6797.1162.2960.08
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.436, + 0.924, + 0.493 + ], + "angle": 0, + "content": "Table 2: The accuracy \\((\\%)\\) of BAFFLE in label non-iid scenarios \\((C = 100)\\) and epoch-level settings with total communication rounds 40 and different \\(K\\) values. We employ Dirichlet dist. with \\(\\alpha = 0.3\\) to ensure each client has a unique label distribution." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.506, + 0.878, + 0.645 + ], + "angle": 0, + "content": "
SettingsLeNet CIFAR-10 | CIFAR-100WRN-10-2 CIFAR-10 | CIFAR-100
K20035.21 | 28.1239.53 | 30.44
50038.14 | 30.9241.69 | 32.89
100039.71 | 33.3543.42 | 34.08
BP Baselines44.41 | 38.4351.18 | 40.85
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.92, + 0.735 + ], + "angle": 0, + "content": "with depth \\( = 10 \\) and width \\( = 2 \\) (WRN-10-2) as the light weight deep model \\( (3.0 \\times 10^{5} \\) parameters) and 3) MobileNet [38] as the deep neural networks \\( (1.3 \\times 10^{7} \\) parameters) that works on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.736, + 0.924, + 0.951 + ], + "angle": 0, + "content": "Participation and communication settings. To perform a comprehensive evaluation of BAFFLE, we simulate three popular FL scenarios [17] with the FedLab [99] participations: iid participations, label non-iid participations and feature non-iid participations. For iid participations, we set the client number \\( C = 10 \\) and use uniform distribution to build local datasets. Then we evaluate our BAFFLE on MNIST and CIFAR-10/100 under both batch-level (FedSGD) and epoch-level (FedAvg) communication settings. For label non-iid participations, we set client number \\( C = 100 \\), use Dirichlet distribution with \\( \\alpha = 0.3 \\) to build clients. For feature non-iid participations, we build clients from the prevailing domain adaptation dataset OfficeHome, which contains 65 categories from 4 different domains, i.e. Art, Clipart, Product and Real-world. We set the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.184, + 0.015, + 0.841, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.061, + 0.506, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.06, + 0.911, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.236, + 0.511, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.407, + 0.362, + 0.421 + ], + "angle": 0, + "content": "LeNet" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.236, + 0.911, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.407, + 0.783, + 0.421 + ], + "angle": 0, + "content": "WRN-10-2" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.435, + 0.921, + 0.526 + ], + "angle": 0, + "content": "Fig. 2: The classification accuracy (\\%) of BAFFLE in iid scenarios (\\(C = 10\\)) and batch-level communication settings with various \\(K\\) values. We treat the models trained by exact gradients on conventional FL systems as the backpropagation (BP) baselines. On different datasets and architectures, our BAFFLE achieves comparable performance to the exact gradient results with a reasonable \\(K\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.56, + 0.922, + 0.618 + ], + "angle": 0, + "content": "total client number to \\( C = 40 \\) and generate 10 clients from each domain. As results, we report Top-1 accuracy for MNIST, CIFAR-10 and OfficeHome and Top-5 accuracy for OfficeHome and CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.638, + 0.924, + 0.95 + ], + "angle": 0, + "content": "Hyperparameters. Following the settings in Section 2, we use FedAVG to aggregate gradients from multiple clients and use SGD-based optimizer to update global parameters. Specifically, we use Adam [46] to train a random initialized model with \\(\\beta = (0.9,0.99)\\), learning rate 0.01 and epochs 20/40 for MNIST and CIFAR-10/100. For OfficeHome, we adapt the transfer learning [40] by loading the ImageNet-pretrained model and finetuning the final layers with Adam, but setting learning rate 0.005 and epochs 40. In BAFFLE, the perturbation scale \\(\\sigma\\) and number \\(K\\) are the most important hyperparameters. As shown in Theorem 1, with less noise and more samples, the BAFFLE will obtain more accurate gradients, leading to improved performance. However, there exists a trade-off between accuracy and computational efficiency: an extremely small \\(\\sigma\\) will cause the underflow problem [27] and a large \\(K\\) will increase computational cost. In practice, we empirically set \\(\\sigma = 10^{-4}\\) because it is the smallest value that does not cause numerical problems in all experiments, and works well on edge devices with half-precision floating-point numbers. We also evaluate the impact of \\(K\\) across a broad range from 100 to 5000." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.484, + 0.034 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.055, + 0.502, + 0.075 + ], + "angle": 0, + "content": "4.1 Four Guidelines for BAFFLE" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.922, + 0.249 + ], + "angle": 0, + "content": "For a general family of continuously differentiable models, we analyze their convergence rate of BAFFLE in Section 3.3. Since deep networks are usually stacked with multiple linear layers and non-linear activation, this layer linearity can be utilized to improve the accuracy-efficiency trade-off. Combining the linearity property and the unique conditions in edge devices (e.g., small data size and half-precision format), we present four guidelines for model design and training that can increase accuracy without introducing extra computation (Appendix C shows the details of linearity analysis):" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.251, + 0.92, + 0.405 + ], + "angle": 0, + "content": "Using twice forward difference (twice-FD) scheme rather than central scheme. Combining difference scheme Eq. (1) and Eq. (2), we find that by executing twice as many forward inferences (i.e. \\(\\mathbf{W} \\pm \\boldsymbol{\\delta}\\)), the central scheme achieves lower residuals than twice-FD, despite the fact that twice-FD can benefit from additional sample times. With the same forward times (e.g., \\(2K\\)), determining which scheme performs better is a practical issue. As shown in Appendix C, we find that twice-FD performs better in all experiments, in part because the linearity reduces the benefit from second-order residuals." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.407, + 0.92, + 0.523 + ], + "angle": 0, + "content": "Using Hardswish in BAFFLE. ReLU is effective when the middle features \\((h(\\cdot)\\) denotes the feature mapping) have the same sign before and after perturbations, i.e. \\(h(\\mathbf{W} + \\boldsymbol{\\delta})\\cdot h(\\mathbf{W}) > 0\\). Since ReLU is not differentiable at zero, the value jump occurs when the sign of features changes after perturbations, i.e. \\(h(\\mathbf{W} + \\boldsymbol{\\delta})\\cdot h(\\mathbf{W}) < 0\\). We use Hardswish [37] to overcome this problem as it is continuously differentiable at zero and easy to implement on edge devices." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.525, + 0.92, + 0.624 + ], + "angle": 0, + "content": "Using exponential moving average (EMA) to reduce oscillations. As shown in Theorem 1, there exists an zero-mean white-noise \\(\\widehat{\\delta}\\) between the true gradient and our estimation. To smooth out the oscillations caused by white noise, we apply EMA strategies from BYOL [28] to the global parameters, with a smoothing coefficient of 0.995." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.625, + 0.92, + 0.703 + ], + "angle": 0, + "content": "Using GroupNorm as opposed to BatchNorm. On edge devices, the dataset size is typically small, which leads to inaccurate batch statistics estimation and degrades performance when using BatchNorm. Thus we employ GroupNorm [94] to solve this issue." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.736, + 0.484, + 0.755 + ], + "angle": 0, + "content": "4.2 Performance on IID Clients" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.773, + 0.922, + 0.949 + ], + "angle": 0, + "content": "Following the settings in Section 4.1, we evaluate the performance of BAFFLE in the iid scenarios. We reproduce all experiments on the BP-based FL systems with the same settings and use them as the baselines. We refer to the baseline results as exact gradients and report the training process of BAFFLE in Figure 2. The value of \\( K \\) (e.g., 200 for LeNet and 500 for WRN-10-2) is much less than the dimensions of parameter space (e.g., \\( 2.7 \\times 10^4 \\) for LeNet and \\( 3 \\times 10^5 \\) for WRN-10-2). Since the convergence rate to the exact gradient is \\( \\mathcal{O}\\left(\\sqrt{\\frac{n}{K}}\\right) \\), the marginal benefit of increasing \\( K \\) decreases. For instance, increasing \\( K \\) from 2000 to 5000 on CIFAR-10 with WRN-10-2 barely improves accuracy by \\( 2\\% \\). Given that the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.917, + 0.032 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.053, + 0.921, + 0.109 + ], + "angle": 0, + "content": "Table 3: The Top-1|Top-5 accuracy (\\%) of BAFFLE on OfficeHome with feature non-iid participations (\\( C = 40 \\)) and epoch-level settings with 40 comm. rounds. We use the pretrained MobileNet, freeze the backbone and finetune the FC layers." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.124, + 0.907, + 0.297 + ], + "angle": 0, + "content": "
SettingsDomainsAvg.
ArtClipartProductReal World
2044.75|69.4652.48|73.8866.63|89.7763.78|87.8356.91|80.24
5047.87|71.3253.43|76.8371.28|91.7467.02|89.9559.90|82.46
K 10050.32|74.4257.43|80.7374.19|93.0269.53|90.4362.87|84.65
20051.42|76.6460.98|86.4176.05|94.4271.51|93.1464.98|87.65
50053.33|77.8562.58|86.6478.84|95.2373.17|93.8566.98|88.40
BP Baselines55.71|80.4365.13|88.6582.44|96.0577.19|95.0470.12|90.04
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.308, + 0.918, + 0.347 + ], + "angle": 0, + "content": "convergence rate of Gaussian perturbations is \\(\\mathcal{O}\\left(\\sqrt{\\frac{n}{K}}\\right)\\), the sampling efficiency may be improved by choosing an alternative distribution for perturbations." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.348, + 0.921, + 0.522 + ], + "angle": 0, + "content": "Ablation studies. As depicted in Figure 3, we conduct ablation studies for BAFFLE to evaluate the aforementioned guidelines. In general, twice-FD, Hardswish and EMA can all improve the accuracy. For two difference schemes, we compare the twice-FD to central scheme with the same computation cost and show that the former outperforms the later, demonstrating that linearity reduces the gain from second-order residuals. As to activation functions, Hardswish is superior to ReLU and SELU because it is differentiable at zero and vanishes to zero in the negative part. Moreover, EMA enhances the performance of training strategies by reducing the effect of white noise." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.523, + 0.92, + 0.854 + ], + "angle": 0, + "content": "Communication efficiency. Compared to the batch-level communication settings (FedSGD) in a BP-based FL system, BAFFLE requires each client to upload a \\(K\\)-dimensional vector to the server and downloads the updated global parameters in each communication round. Since \\(K\\) is significantly less than the parameter amounts (e.g., 500 versus 0.3 million), BAFFLE reduces data transfer by approximately half. To reduce communication costs, the prevalent FL system requires each client to perform model optimization on the local training dataset and upload the model updates to the server after a specified number of local epochs. BAFFLE can also perform epoch-level communications by employing an \\(\\mathcal{O}(n)\\) additional memory to store the perturbation in each forward and estimate the local gradient using Eq. (6). Then each client optimizes the local model with SGD and uploads local updates after several epochs. As shown in Table 1, we evaluate the performance of BAFFLE under one-epoch communication settings. As epoch-level communication is more prevalent in the real-world FL, all the following experiments will be conducted in this context. In brief, BAFFLE uploads the same Bytes as BP-based FL in epoch-level communication while the total communication rounds are much less than FedZO [22], e.g. 20 versus 1000 on MNIST." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.88, + 0.544, + 0.898 + ], + "angle": 0, + "content": "4.3 Performance on Non-IID Clients" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.91, + 0.922, + 0.951 + ], + "angle": 0, + "content": "Following Section 4.1, we evaluate the performance of BAFFLE in both label non-iid and feature non-iid scenarios. For label non-iid scenarios, we use the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.484, + 0.034 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.061, + 0.51, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.06, + 0.911, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.237, + 0.51, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.407, + 0.362, + 0.421 + ], + "angle": 0, + "content": "LeNet" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.237, + 0.911, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.407, + 0.783, + 0.421 + ], + "angle": 0, + "content": "WRN-10-2" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.435, + 0.921, + 0.492 + ], + "angle": 0, + "content": "Fig. 3: The ablation study of BAFFLE guidelines, with \\( K = 100 \\) on MNIST and \\( K = 500 \\) on CIFAR-10. As seen, twice-FD, Hardswish, and EMA all improve performance without extra computation. EMA reduces oscillations by lessening Gaussian noise." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.525, + 0.922, + 0.701 + ], + "angle": 0, + "content": "CIFAR-10/100 datasets and employ Dirichlet distribution to ensure that each client has a unique label distribution. We evaluate the performance of BAFFLE with 100 clients and various K values. As seen in Table 2, the model suffers a significant drop in accuracy (e.g., \\(14\\%\\) in CIFAR-10 and \\(16\\%\\) in CIFAR-100) due to the label non-iid effect. For feature non-iid scenarios, we construct clients using the OfficeHome dataset and use MobileNet as the deep model. As seen in Table 3, we use the transfer learning strategy to train MobileNet, i.e., we load the parameters pretrained on ImageNet, freeze the backbone parameters, and retrain the classification layers. The accuracy decrease is approximately \\(3\\% \\sim 5\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.726, + 0.75, + 0.746 + ], + "angle": 0, + "content": "4.4 Computation Efficiency, Memory and Robustness" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.755, + 0.92, + 0.889 + ], + "angle": 0, + "content": "BAFFLE uses \\( K \\) times forward passes instead of backward. Since the backward pass is about as expensive as two normal forward passes [35] and five single-precision accelerated forward passes [67], BAFFLE results in approximately \\( \\frac{K}{5} \\) times the computation expense of BP-based FL. Although BAFFLE results in \\( \\frac{K}{5} - 1 \\) times extra computation cost, we show the cost can be reduced with proper training strategies, e.g., the transfer learning in Table 3 can reduce \\( K \\) to 20 on the MobileNet and the \\( 224 \\times 224 \\) sized OfficeHome dataset." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.891, + 0.922, + 0.951 + ], + "angle": 0, + "content": "Moreover, BAFFLE can reduce huge memory cost on edge devices with the efficiency in static memory and dynamic memory. The auto-differential framework is used to run BP on deep networks, which requires extra static memory" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.918, + 0.031 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.053, + 0.921, + 0.109 + ], + "angle": 0, + "content": "Table 4: The GPU memory cost (MB) of vanilla BP and BAFFLE, respectively. 'min~max' denotes the minimum and maximum dynamic memory for BAFFLE. We also report the ratio (\\%) of vanilla BP to BAFFLE's max memory cost." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.125, + 0.917, + 0.234 + ], + "angle": 0, + "content": "
BackboneCIFAR-10/100OfficeHome/ImageNet
BPBAFFLERatioBPBAFFLERatio
LeNet168067~17410.35252786~2017.95
WRN-10-2187875~19610.43342594~2517.32
MobileNet2041102~21710.635271121~2895.48
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.26, + 0.921, + 0.435 + ], + "angle": 0, + "content": "(e.g., 200MB for Caffe [41] and 1GB for Pytorch [72]) and imposes a considerable burden on edge devices such as IoT sensors. Due to the necessity of restoring intermediate states, BP also requires enormous amounts of dynamic memory (\\(\\geq\\) 5GB for MobileNet [24]). Since BAFFLE only requires inference, we can slice the computation graph and execute the forwards per layer [44]. As shown in Table 4, BAFFLE reduces the memory cost to \\(5\\% \\sim 10\\%\\) by executing layer-by-layer inference. By applying kernel-wise computations, we can further reduce the memory cost to approximately \\(1\\%\\) (e.g., 64MB for MobileNet [85]), which is suitable for scenarios with extremely limited storage resources, such as TEE." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.437, + 0.921, + 0.73 + ], + "angle": 0, + "content": "Recent works exploit TEE to protect models from white-box attacks by preventing model exposure [44]. However, due to the security guarantee, the usable memory of TEE is usually small [85] (e.g., 90MB on Intel SGX for Skylake CPU [61]), which is typically far less than what a backpropagation-based FL system requires. In contrast, BAFFLE can execute in TEE due to its little memory cost (more details are in Appendix B). Membership inference attacks and model inversion attacks need to repeatedly perform model inference and obtain confidence values or classification scores [78, 101]. Given that BAFFLE provides stochastic loss differences \\(\\Delta \\mathcal{L}(\\mathbf{W},\\delta ;\\mathbb{D})\\) associated with the random perturbation \\(\\delta\\), the off-the-shelf inference attacks may not perform on BAFFLE directly (while adaptively designed attacking strategies are possible to evade BAFFLE). Motivated by differential privacy [1], we further design heuristic experiments to study the information leakage from \\(\\Delta \\mathcal{L}\\) (details in Appendix D). As shown in Figure 5, the \\(\\Delta \\mathcal{L}\\) between real data and random noise is hard to distinguish, indicating it is difficult for attackers to obtain useful information from BAFFLE's outputs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.764, + 0.332, + 0.785 + ], + "angle": 0, + "content": "5 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.812, + 0.922, + 0.91 + ], + "angle": 0, + "content": "Along the research routine of FL, many efforts have been devoted to, e.g., dealing with non-IID distributions [21, 53, 75, 90, 102], multi-task learning [60, 79], and preserving privacy of clients [11, 12, 26, 31, 56, 58, 63, 84]. Below we introduce the work on efficiency and vulnerability in FL following the survey of Kairouz et al. [42], which is more related to this paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.91, + 0.92, + 0.951 + ], + "angle": 0, + "content": "Efficiency in FL. It is widely understood that the communication and computational efficiency is a primary bottleneck for deploying FL in practice [6,18,74," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.484, + 0.034 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.923, + 0.271 + ], + "angle": 0, + "content": "89, 92]. Specifically, communicating between the server and clients could be potentially expensive and unreliable. The seminal work of Konečný et al. [47] introduces sparsification and quantization to reduce the communication cost, where several theoretical works investigate the optimal trade-off between the communication cost and model accuracy [2, 7, 15, 30, 100]. Since practical clients usually have slower upload than download bandwidth, much research interest focuses on gradient compression [4, 8, 36, 82]. On the other hand, different methods have been proposed to reduce the computational burden of local clients [16, 29, 33], since these clients are usually edge devices with limited resources. Training paradigms exploiting tensor factorization in FL can also achieve promising performance [45, 59]." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.274, + 0.923, + 0.451 + ], + "angle": 0, + "content": "Vulnerability in FL. The characteristic of decentralization in FL is beneficial to protecting data privacy of clients, but in the meanwhile, providing white-box accessibility of model status leaves flexibility for malicious clients to perform poisoning/backdoor attacks [5, 9, 71, 88, 95], model/gradient inversion attacks [25, 39, 101], and membership inference attacks [57, 68, 78]. To alleviate the vulnerability in FL, several defense strategies have been proposed via selecting reliable clients [43], data augmentation [13], update clipping [81], robust training [52], model perturbation [97], detection methods [20, 76], and methods based on differential privacy [93]." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.497, + 0.511, + 0.518 + ], + "angle": 0, + "content": "6 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.557, + 0.923, + 0.693 + ], + "angle": 0, + "content": "Backpropagation is the gold standard for training deep networks, and it is also utilized by traditional FL systems. However, backpropagation is unsuited for edge devices due to their limited resources and possible lack of reliability. Using zero-order optimization techniques, we explore the possibility of BAFFLE in this paper. We need to specify that there are scenarios in which clients are fully trusted and have sufficient computing and storage resources. In these situations, traditional FL with backpropagation is preferred." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.697, + 0.923, + 0.95 + ], + "angle": 0, + "content": "While our preliminary studies on BAFFLE have generated encouraging results, there are still a number of tough topics to investigate: (i) Compared to the models trained using exact gradients, the accuracy of models trained using BAFFLE is inferior. One reason is that we select small values of \\( K \\) (e.g., 500) relative to the number of model parameters (e.g., \\( 3.0 \\times 10^{5} \\)); another reason is that gradient descent is designed for exact gradients, whereas our noisy gradient estimation may require advanced learning algorithms. (ii) The empirical variance of zero-order gradient estimators affects training convergence in BAFFLE. It is crucial to research variance reduction approaches, such as control variates and non-Gaussian sampling distributions. (iii) Stein's identity is proposed for loss functions with Gaussian noises imposed on model parameters. Intuitively, this smoothness is related to differential privacy in FL, but determining their relationship requires theoretical derivations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.054, + 0.338, + 0.076 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.921, + 0.132 + ], + "angle": 0, + "content": "This paper is supported by the National Science Foundation of China (62132017), Zhejiang Provincial Natural Science Foundation of China (LD24F020011)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.157, + 0.238, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.193, + 0.92, + 0.264 + ], + "angle": 0, + "content": "1. Abadi, M., Chu, A., Goodfellow, I.J., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, Vienna, Austria, October 24-28, 2016 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.264, + 0.921, + 0.317 + ], + "angle": 0, + "content": "2. Acharya, J., Canonne, C.L., Tyagi, H.: Inference under information constraints i: Lower bounds from chi-square contraction. IEEE Transactions on Information Theory (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.317, + 0.921, + 0.37 + ], + "angle": 0, + "content": "3. Adamczak, R., Litvak, A.E., Pajor, A., Tomczak-Jaegermann, N.: Sharp bounds on the rate of convergence of the empirical covariance matrix. Comptes Rendus Mathematique 349(3-4), 195-200 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.37, + 0.921, + 0.422 + ], + "angle": 0, + "content": "4. Alistarh, D., Grubic, D., Li, J., Tomioka, R., Vojnovic, M.: Qsgd: Communication-efficient sgd via gradient quantization and encoding. Advances in Neural Information Processing Systems (NeurIPS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.422, + 0.921, + 0.475 + ], + "angle": 0, + "content": "5. Bagdasaryan, E., Veit, A., Hua, Y., Estrin, D., Shmatikov, V.: How to back-door federated learning. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.476, + 0.921, + 0.528 + ], + "angle": 0, + "content": "6. Balakrishnan, R., Li, T., Zhou, T., Himayat, N., Smith, V., Bilmes, J.: Diverse client selection for federated learning via submodular maximization. In: International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.528, + 0.921, + 0.58 + ], + "angle": 0, + "content": "7. Barnes, L.P., Han, Y., Özgür, A.: Lower bounds for learning distributions under communication constraints via fisher information. Journal of Machine Learning Research (JMLR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.58, + 0.921, + 0.632 + ], + "angle": 0, + "content": "8. Basu, D., Data, D., Karakus, C., Diggavi, S.: Qsparse-local-sgd: Distributed sgd with quantization, sparsification and local computations. Advances in Neural Information Processing Systems (NeurIPS) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.632, + 0.921, + 0.685 + ], + "angle": 0, + "content": "9. Bhagoji, A.N., Chakraborty, S., Mittal, P., Calo, S.: Analyzing federated learning through an adversarial lens. In: International Conference on Machine Learning (ICML) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.686, + 0.92, + 0.756 + ], + "angle": 0, + "content": "10. Bonawitz, K.A., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for privacy-preserving machine learning. In: ACM Conference on Computer and Communications Security (CCS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.756, + 0.92, + 0.808 + ], + "angle": 0, + "content": "11. Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for federated learning on user-held data. arXiv preprint arXiv:1611.04482 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.808, + 0.92, + 0.878 + ], + "angle": 0, + "content": "12. Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for privacy-preserving machine learning. In: ACM SIGSAC Conference on Computer and Communications Security (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.878, + 0.92, + 0.949 + ], + "angle": 0, + "content": "13. Borgnia, E., Geiping, J., Cherepanova, V., Fowl, L., Gupta, A., Ghiasi, A., Huang, F., Goldblum, M., Goldstein, T.: Dp-instahide: Provably defusing poisoning and backdoor attacks with differentially private data augmentations. arXiv preprint arXiv:2103.02079 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.193, + 0.921, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.484, + 0.034 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.057, + 0.92, + 0.129 + ], + "angle": 0, + "content": "14. Bradbury, J., Frostig, R., Hawkins, P., Johnson, M.J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., Zhang, Q.: JAX: composable transformations of Python+NumPy programs (2018), http://github.com/google/jax" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.129, + 0.92, + 0.182 + ], + "angle": 0, + "content": "15. Braverman, M., Garg, A., Ma, T., Nguyen, H.L., Woodruff, D.P.: Communication lower bounds for statistical estimation problems via a distributed data processing inequality. In: ACM Symposium on Theory of Computing (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.182, + 0.92, + 0.233 + ], + "angle": 0, + "content": "16. Caldas, S., Konečny, J., McMahan, H.B., Talwalkar, A.: Expanding the reach of federated learning by reducing client resource requirements. arXiv preprint arXiv:1812.07210 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.233, + 0.918, + 0.268 + ], + "angle": 0, + "content": "17. Caldas, S., Wu, P., Li, T., Konečny, J., McMahan, H.B., Smith, V., Talwalkar, A.: LEAF: A benchmark for federated settings. CoRR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.268, + 0.92, + 0.32 + ], + "angle": 0, + "content": "18. Chen, M., Shlezinger, N., Poor, H.V., Eldar, Y.C., Cui, S.: Communication-efficient federated learning. Proceedings of the National Academy of Sciences (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.32, + 0.92, + 0.373 + ], + "angle": 0, + "content": "19. Chen, Y., Luo, F., Li, T., Xiang, T., Liu, Z., Li, J.: A training-integrity privacy-preserving federated learning scheme with trusted execution environment. Information Sciences (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.373, + 0.92, + 0.425 + ], + "angle": 0, + "content": "20. Dong, Y., Yang, X., Deng, Z., Pang, T., Xiao, Z., Su, H., Zhu, J.: Black-box detection of backdoor attacks with limited information and data. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.425, + 0.918, + 0.478 + ], + "angle": 0, + "content": "21. Eichner, H., Koren, T., McMahan, B., Srebro, N., Talwar, K.: Semi-cyclic stochastic gradient descent. In: International Conference on Machine Learning (ICML) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.478, + 0.918, + 0.53 + ], + "angle": 0, + "content": "22. Fang, W., Yu, Z., Jiang, Y., Shi, Y., Jones, C.N., Zhou, Y.: Communication-efficient stochastic zeroth-order optimization for federated learning. IEEE Trans. Signal Process. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.53, + 0.918, + 0.582 + ], + "angle": 0, + "content": "23. Flaxman, A.D., Kalai, A.T., McMahan, H.B.: Online convex optimization in the bandit setting: gradient descent without a gradient. arXiv preprint cs/0408007 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.582, + 0.92, + 0.653 + ], + "angle": 0, + "content": "24. Gao, Y., Liu, Y., Zhang, H., Li, Z., Zhu, Y., Lin, H., Yang, M.: Estimatinggpu memory consumption of deep learning models. In: ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.653, + 0.918, + 0.705 + ], + "angle": 0, + "content": "25. Geiping, J., Bauermeister, H., Dröge, H., Moeller, M.: Inverting gradients-how easy is it to break privacy in federated learning? In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.705, + 0.918, + 0.758 + ], + "angle": 0, + "content": "26. Ghazi, B., Kumar, R., Manurangsi, P., Pagh, R.: Private counting from anonymous messages: Near-optimal accuracy with vanishing communication overhead. In: International Conference on Machine Learning (ICML) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.758, + 0.871, + 0.774 + ], + "angle": 0, + "content": "27. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.774, + 0.92, + 0.844 + ], + "angle": 0, + "content": "28. Grill, J., Strub, F., Altché, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Doersch, C., Pires, B.A., Guo, Z., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent - A new approach to self-supervised learning. In: NeurIPS 2020 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.844, + 0.918, + 0.897 + ], + "angle": 0, + "content": "29. Hamer, J., Mohri, M., Suresh, A.T.: Fedboost: A communication-efficient algorithm for federated learning. In: International Conference on Machine Learning (ICML) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.897, + 0.918, + 0.949 + ], + "angle": 0, + "content": "30. Han, Y., Ozgür, A., Weissman, T.: Geometric lower bounds for distributed parameter estimation under communication constraints. In: Conference On Learning Theory (COLT) (2018)" + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.893, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.057, + 0.918, + 0.111 + ], + "angle": 0, + "content": "31. Hao, M., Li, H., Luo, X., Xu, G., Yang, H., Liu, S.: Efficient and privacy-enhanced federated learning for industrial artificial intelligence. IEEE Transactions on Industrial Informatics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.112, + 0.92, + 0.164 + ], + "angle": 0, + "content": "32. Hard, A., Rao, K., Mathews, R., Ramaswamy, S., Beaufays, F., Augenstein, S., Eichner, H., Kiddon, C., Ramage, D.: Federated learning for mobile keyboard prediction. arXiv preprint arXiv:1811.03604 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.165, + 0.92, + 0.218 + ], + "angle": 0, + "content": "33. He, C., Annavaram, M., Avestimehr, S.: Group knowledge transfer: Federated learning of large cnns at the edge. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.218, + 0.918, + 0.272 + ], + "angle": 0, + "content": "34. He, D., Shi, W., Li, S., Gao, X., Zhang, J., Bian, J., Wang, L., Liu, T.Y.: Learning physics-informed neural networks without stacked back-propagation. arXiv preprint arXiv:2202.09340 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.272, + 0.918, + 0.307 + ], + "angle": 0, + "content": "35. Hinton, G., Srivastava, N.: Csc321: Introduction to neural networks and machine learning. Lecture 10 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.308, + 0.918, + 0.343 + ], + "angle": 0, + "content": "36. Horváth, S., Ho, C.Y., Horvath, L., Sahu, A.N., Canini, M., Richtárik, P.: Natural compression for distributed deep learning. arXiv preprint arXiv:1905.10988 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.344, + 0.917, + 0.396 + ], + "angle": 0, + "content": "37. Howard, A., Pang, R., Adam, H., Le, Q.V., Sandler, M., Chen, B., Wang, W., Chen, L., Tan, M., Chu, G., Vasudevan, V., Zhu, Y.: Searching for mobilenetv3. In: IEEE International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.397, + 0.918, + 0.45 + ], + "angle": 0, + "content": "38. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.451, + 0.918, + 0.503 + ], + "angle": 0, + "content": "39. Huang, Y., Gupta, S., Song, Z., Li, K., Arora, S.: Evaluating gradient inversion attacks and defenses in federated learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.504, + 0.918, + 0.539 + ], + "angle": 0, + "content": "40. Huh, M., Agrawal, P., Efros, A.A.: What makes imagenet good for transfer learning? CoRR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.539, + 0.918, + 0.61 + ], + "angle": 0, + "content": "41. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R.B., Guadarrama, S., Darrell, T.: Caffe: Convolutional architecture for fast feature embedding. In: Hua, K.A., Rui, Y., Steinmetz, R., Hanjalic, A., Natsev, A., Zhu, W. (eds.) Proceedings of the ACM International Conference on Multimedia (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.611, + 0.918, + 0.682 + ], + "angle": 0, + "content": "42. Kairouz, P., McMahan, H.B., Avent, B., Bellet, A., Dennis, M., Bhagoji, A.N., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R., et al.: Advances and open problems in federated learning. Foundations and Trends in Machine Learning 14(1-2), 1-210 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.682, + 0.918, + 0.717 + ], + "angle": 0, + "content": "43. Kang, J., Xiong, Z., Niyato, D., Zou, Y., Zhang, Y., Guizani, M.: Reliable federated learning for mobile networks. IEEE Wireless Communications (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.718, + 0.918, + 0.77 + ], + "angle": 0, + "content": "44. Kim, K., Kim, C.H., Rhee, J.J., Yu, X., Chen, H., Tian, D., Lee, B.: Vessels: Efficient and scalable deep learning prediction on trusted processors. In: ACM Symposium on Cloud Computing (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.771, + 0.918, + 0.824 + ], + "angle": 0, + "content": "45. Kim, Y., Sun, J., Yu, H., Jiang, X.: Federated tensor factorization for computational phenotyping. In: ACM International Conference on Knowledge Discovery and Data Mining (SIGKDD) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.825, + 0.918, + 0.859 + ], + "angle": 0, + "content": "46. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: 3rd International Conference on Learning Representations, ICLR 2015 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.86, + 0.918, + 0.913 + ], + "angle": 0, + "content": "47. Konečny, J., McMahan, H.B., Yu, F.X., Richtárik, P., Suresh, A.T., Bacon, D.: Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.914, + 0.918, + 0.949 + ], + "angle": 0, + "content": "48. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.484, + 0.034 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.057, + 0.918, + 0.093 + ], + "angle": 0, + "content": "49. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of IEEE (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.093, + 0.92, + 0.145 + ], + "angle": 0, + "content": "50. Li, J., Luo, X., Qiao, M.: On generalization error bounds of noisy gradient methods for non-convex learning. In: International Conference on Learning Representations (ICLR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.146, + 0.917, + 0.18 + ], + "angle": 0, + "content": "51. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.18, + 0.921, + 0.232 + ], + "angle": 0, + "content": "52. Li, T., Hu, S., Beirami, A., Smith, V.: Ditto: Fair and robust federated learning through personalization. In: International Conference on Machine Learning (ICML) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.233, + 0.92, + 0.285 + ], + "angle": 0, + "content": "53. Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data. In: International Conference on Learning Representations (ICLR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.285, + 0.918, + 0.32 + ], + "angle": 0, + "content": "54. Li, Z., Chen, L.: Communication-efficient decentralized zeroth-order method on heterogeneous data. In: WCSP 2021 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.32, + 0.92, + 0.389 + ], + "angle": 0, + "content": "55. Liu, S., Chen, P.Y., Kailkhura, B., Zhang, G., Hero III, A.O., Varshney, P.K.: A primer on zeroth-order optimization in signal processing and machine learning: Principals, recent advances, and applications. IEEE Signal Processing Magazine (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.39, + 0.918, + 0.442 + ], + "angle": 0, + "content": "56. Liu, Y., Suresh, A.T., Yu, F.X.X., Kumar, S., Riley, M.: Learning discrete distributions: user vs item-level privacy. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.442, + 0.918, + 0.495 + ], + "angle": 0, + "content": "57. Luo, X., Wu, Y., Xiao, X., Ooi, B.C.: Feature inference attack on model predictions in vertical federated learning. In: IEEE International Conference on Data Engineering (ICDE) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.495, + 0.918, + 0.547 + ], + "angle": 0, + "content": "58. Lyu, L., Yu, H., Ma, X., Sun, L., Zhao, J., Yang, Q., Yu, P.S.: Privacy and robustness in federated learning: Attacks and defenses. IEEE Transactions on Neural Networks and Learning Systems pp. 1-21 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.547, + 0.918, + 0.599 + ], + "angle": 0, + "content": "59. Ma, J., Zhang, Q., Lou, J., Ho, J.C., Xiong, L., Jiang, X.: Privacy-preserving tensor factorization for collaborative health data analysis. In: ACM International Conference on Information and Knowledge Management (CIKM) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.599, + 0.918, + 0.651 + ], + "angle": 0, + "content": "60. Marfoq, O., Neglia, G., Bellet, A., Kameni, L., Vidal, R.: Federated multi-task learning under a mixture of distributions. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.652, + 0.918, + 0.722 + ], + "angle": 0, + "content": "61. McKeen, F., Alexandrovich, I., Anati, I., Caspi, D., Johnson, S., Leslie-Hurd, R., Rozas, C.: Intel® software guard extensions (intel® sgx) support for dynamic memory management inside an enclave. In: Proceedings of the Hardware and Architectural Support for Security and Privacy (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.722, + 0.918, + 0.774 + ], + "angle": 0, + "content": "62. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.774, + 0.918, + 0.826 + ], + "angle": 0, + "content": "63. McMahan, B., Ramage, D., Talwar, K., Zhang, L.: Learning differentially private recurrent language models. In: International Conference on Learning Representations (ICLR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.826, + 0.918, + 0.896 + ], + "angle": 0, + "content": "64. Mo, F., Haddadi, H., Katevas, K., Marin, E., Perino, D., Courtellis, N.: Ppf: privacy-preserving federated learning with trusted execution environments. In: Annual International Conference on Mobile Systems, Applications, and Services (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.896, + 0.918, + 0.949 + ], + "angle": 0, + "content": "65. Mondal, A., More, Y., Rooparaghunath, R.H., Gupta, D.: Flatee: Federated learning across trusted execution environments. arXiv preprint arXiv:2111.06867 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.057, + 0.921, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.058, + 0.918, + 0.093 + ], + "angle": 0, + "content": "66. Moulay, E., Léchappé, V., Plestan, F.: Properties of the sign gradient descent algorithms. Inf. Sci. 492, 29-39 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.093, + 0.92, + 0.145 + ], + "angle": 0, + "content": "67. Nakandala, S., Nagrecha, K., Kumar, A., Papakonstantinou, Y.: Incremental and approximate computations for accelerating deep CNN inference. ACM Trans. Database Syst. 45 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.146, + 0.921, + 0.214 + ], + "angle": 0, + "content": "68. Nasr, M., Shokri, R., Houmansadr, A.: Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In: IEEE symposium on security and privacy (S&P). IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.216, + 0.92, + 0.251 + ], + "angle": 0, + "content": "69. Nesterov, Y., Spokoiny, V.: Random gradient-free minimization of convex functions. Foundations of Computational Mathematics (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.251, + 0.92, + 0.302 + ], + "angle": 0, + "content": "70. Pang, T., Xu, K., Li, C., Song, Y., Ermon, S., Zhu, J.: Efficient learning of generative models via finite-difference score matching. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.303, + 0.92, + 0.354 + ], + "angle": 0, + "content": "71. Pang, T., Yang, X., Dong, Y., Su, H., Zhu, J.: Accumulative poisoning attacks on real-time data. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.355, + 0.92, + 0.442 + ], + "angle": 0, + "content": "72. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E.Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.444, + 0.92, + 0.512 + ], + "angle": 0, + "content": "73. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.513, + 0.92, + 0.565 + ], + "angle": 0, + "content": "74. Rothchild, D., Panda, A., Ullah, E., Ivkin, N., Stoica, I., Braverman, V., Gonzalez, J., Arora, R.: Fetchsgd: Communication-efficient federated learning with sketching. In: International Conference on Machine Learning (ICML) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.566, + 0.92, + 0.618 + ], + "angle": 0, + "content": "75. Sattler, F., Wiedemann, S., Müller, K.R., Samek, W.: Robust and communication-efficient federated learning from non-iid data. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.619, + 0.92, + 0.669 + ], + "angle": 0, + "content": "76. Seetharaman, S., Malaviya, S., KV, R., Shukla, M., Lodha, S.: Influence based defense against data poisoning attacks in online learning. arXiv preprint arXiv:2104.13230 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.67, + 0.92, + 0.74 + ], + "angle": 0, + "content": "77. Sharma, H., Park, J., Suda, N., Lai, L., Chau, B., Chandra, V., Esmaeilzadeh, H.: Bit fusion: Bit-level dynamically composable architecture for accelerating deep neural network. In: IEEE Annual International Symposium on Computer Architecture (ISCA) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.741, + 0.92, + 0.793 + ], + "angle": 0, + "content": "78. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. In: IEEE symposium on security and privacy (S&P). IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.794, + 0.92, + 0.828 + ], + "angle": 0, + "content": "79. Smith, V., Chiang, C.K., Sanjabi, M., Talwalkar, A.S.: Federated multi-task learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.828, + 0.92, + 0.862 + ], + "angle": 0, + "content": "80. Stein, C.M.: Estimation of the mean of a multivariate normal distribution. The annals of Statistics (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.863, + 0.92, + 0.896 + ], + "angle": 0, + "content": "81. Sun, Z., Kairouz, P., Suresh, A.T., McMahan, H.B.: Can you really backdoor federated learning? arXiv preprint arXiv:1911.07963 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.897, + 0.92, + 0.949 + ], + "angle": 0, + "content": "82. Suresh, A.T., Felix, X.Y., Kumar, S., McMahan, H.B.: Distributed mean estimation with limited communication. In: International Conference on Machine Learning (ICML) (2017)" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.058, + 0.921, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.484, + 0.034 + ], + "angle": 0, + "content": "Haozhe Feng, Tianyu Pang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.057, + 0.92, + 0.111 + ], + "angle": 0, + "content": "83. Tramer, F., Boneh, D.: Slalom: Fast, verifiable and private execution of neural networks in trusted hardware. In: International Conference on Learning Representations (ICLR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.112, + 0.92, + 0.164 + ], + "angle": 0, + "content": "84. Truex, S., Baracaldo, N., Anwar, A., Steinke, T., Ludwig, H., Zhang, R., Zhou, Y.: A hybrid approach to privacy-preserving federated learning. In: ACM workshop on Artificial Intelligence and Security (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.165, + 0.921, + 0.219 + ], + "angle": 0, + "content": "85. Truong, J.B., Gallagher, W., Guo, T., Walls, R.J.: Memory-efficient deep learning inference in trusted execution environments. In: IEEE International Conference on Cloud Engineering (IC2E) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.219, + 0.92, + 0.272 + ], + "angle": 0, + "content": "86. Umuroglu, Y., Rasnayake, L., Själander, M.: Bismo: A scalable bit-serial matrix multiplication overlay for reconfigurable computing. In: International Conference on Field Programmable Logic and Applications (FPL) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.272, + 0.918, + 0.308 + ], + "angle": 0, + "content": "87. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.308, + 0.921, + 0.378 + ], + "angle": 0, + "content": "88. Wang, H., Sreenivasan, K., Rajput, S., Vishwakarma, H., Agarwal, S., Sohn, J.y., Lee, K., Papailiopoulos, D.: Attack of the tails: Yes, you really can backdoor federated learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.379, + 0.921, + 0.432 + ], + "angle": 0, + "content": "89. Wang, H.P., Stich, S., He, Y., Fritz, M.: Prog fed: effective, communication, and computation efficient federated learning by progressive training. In: International Conference on Machine Learning (ICML) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.433, + 0.92, + 0.487 + ], + "angle": 0, + "content": "90. Wang, J., Liu, Q., Liang, H., Joshi, G., Poor, H.V.: Tackling the objective inconsistency problem in heterogeneous federated optimization. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.487, + 0.92, + 0.539 + ], + "angle": 0, + "content": "91. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.539, + 0.92, + 0.592 + ], + "angle": 0, + "content": "92. Wang, L., Wang, W., Li, B.: Cmfl: Mitigating communication overhead for federated learning. In: IEEE International Conference on Distributed Computing Systems (ICDCS) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.592, + 0.92, + 0.647 + ], + "angle": 0, + "content": "93. Wei, K., Li, J., Ding, M., Ma, C., Yang, H.H., Farokhi, F., Jin, S., Quek, T.Q., Poor, H.V.: Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.647, + 0.918, + 0.664 + ], + "angle": 0, + "content": "94. Wu, Y., He, K.: Group normalization. Int. J. Comput. Vis. 128(3), 742-755 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.665, + 0.92, + 0.717 + ], + "angle": 0, + "content": "95. Xie, C., Huang, K., Chen, P.Y., Li, B.: Distributed backdoor attacks against federated learning. In: International Conference on Learning Representations (ICLR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.718, + 0.92, + 0.77 + ], + "angle": 0, + "content": "96. Yang, T., Andrew, G., Eichner, H., Sun, H., Li, W., Kong, N., Ramage, D., Beaufays, F.: Applied federated learning: Improving google keyboard query suggestions. arXiv preprint arXiv:1812.02903 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.771, + 0.92, + 0.824 + ], + "angle": 0, + "content": "97. Yang, X., Feng, Y., Fang, W., Shao, J., Tang, X., Xia, S.T., Lu, R.: An accuracy-lossless perturbation method for defending privacy attacks in federated learning. In: ACM Web Conference (WWW) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.825, + 0.92, + 0.86 + ], + "angle": 0, + "content": "98. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: Proceedings of the British Machine Vision Conference 2016 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.861, + 0.92, + 0.896 + ], + "angle": 0, + "content": "99. Zeng, D., Liang, S., Hu, X., Wang, H., Xu, Z.: Fedlab: A flexible federated learning framework. arXiv preprint arXiv:2107.11621 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.896, + 0.92, + 0.95 + ], + "angle": 0, + "content": "100. Zhang, Y., Duchi, J., Jordan, M.I., Wainwright, M.J.: Information-theoretic lower bounds for distributed statistical estimation with communication constraints. Advances in Neural Information Processing Systems (NeurIPS) (2013)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.921, + 0.95 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.892, + 0.017, + 0.917, + 0.032 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.056, + 0.918, + 0.111 + ], + "angle": 0, + "content": "101. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.112, + 0.92, + 0.147 + ], + "angle": 0, + "content": "102. Zhao, Y., Li, M., Lai, L., Suda, N., Civin, D., Chandra, V.: Federated learning with non-iid data. arXiv preprint arXiv:1806.00582 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.147, + 0.92, + 0.202 + ], + "angle": 0, + "content": "103. Zhu, Z., Wu, J., Yu, B., Wu, L., Ma, J.: The anisotropic noise in stochastic gradient descent: Its behavior of escaping from sharp minima and regularization effects. In: International Conference on Machine Learning (ICML) (2019)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.056, + 0.92, + 0.202 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_origin.pdf b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cc6b877f7b77410786ee427f40f0a9d744ec9688 --- /dev/null +++ b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/9626a0c5-5cdd-4a37-9478-04758a13b5ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b8c8cc95f63d8c6cc5da66a9602278feffc23ec78c2a94e9e8c1ffdd9eef81f +size 889906 diff --git a/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/full.md b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a40dd2bdd05a272d8c2e1dc3f635e8ae34add69e --- /dev/null +++ b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/full.md @@ -0,0 +1,377 @@ +# BAFFLE: A Baseline of Backpropagation-Free Federated Learning + +Haozhe Feng $^{1,3\star}$ , Tianyu Pang $^{2\star\star}$ , Chao Du $^{2}$ , Wei Chen $^{1\star\star}$ , Shuicheng Yan $^{2}$ , and Min Lin $^{2}$ + +$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Sea AI Lab, Singapore $^{3}$ Tencent Inc. {fenghz, chenvis} @zju.edu.cn {tianyupang, duchao, yansc, linmin} @sea.com + +Abstract. Federated learning (FL) is a general principle for decentralized clients to train a server model collectively without sharing local data. FL is a promising framework with practical applications, but its standard training paradigm requires the clients to backpropagate through the model to compute gradients. Since these clients are typically edge devices and not fully trusted, executing backpropagation on them incurs computational and storage overhead as well as white-box vulnerability. In light of this, we develop backpropagation-free federated learning, dubbed BAFFLE, in which backpropagation is replaced by multiple forward processes to estimate gradients. BAFFLE is 1) memory-efficient and easily fits uploading bandwidth; 2) compatible with inference-only hardware optimization and model quantization or pruning; and 3) well-suited to trusted execution environments, because the clients in BAFFLE only execute forward propagation and return a set of scalars to the server. Empirically we use BAFFLE to train deep models from scratch or to finetune pretrained models, achieving acceptable results. + +Keywords: Federated Learning $\cdot$ Finite Difference $\cdot$ Backpropagation-Free Learning + +# 1 Introduction + +Federated learning (FL) allows decentralized clients to collaboratively train a server model [62]. In each training round, the selected clients compute model gradients or updates on their local private datasets, without explicitly exchanging sample points to the server. While FL describes a promising blueprint and has several applications [32, 51, 96], the mainstream training paradigm of FL is still gradient-based that requires the clients to locally execute backpropagation, which leads to two practical limitations: + +(i) Overhead for edge devices. The clients in FL are usually edge devices, such as mobile phones and IoT sensors, whose hardware is primarily optimized for inference-only purposes [77,86], rather than for backpropagation. Due to the limited resources, computationally affordable models running on edge devices are typically quantized and pruned [91], making exact backpropagation difficult. In addition, standard implementations of backpropagation rely on either forward-mode or reverse-mode auto-differentiation in contemporary machine learning packages [14, 73], which increases storage requirements. +(ii) White-box vulnerability. To facilitate gradient computing, the server regularly distributes its model status to the clients, but this white-box exposure of the model renders the server vulnerable to, e.g., poisoning or inversion attacks from malicious clients [25,78,95,101]. With that, recent attempts are made to exploit trusted execution environments (TEEs) in FL, which can isolate the model status within a black-box secure area and significantly reduce the success rate of malicious evasion [19,64,65]. However, TEEs are highly memory-constrained [85], while backpropagation is memory-consuming to restore intermediate states. + +While numerous solutions have been proposed to alleviate these limitations (related work discussed in Section 5), we raise an essential question: how to perform backpropagation-free $FL$ ? Inspired by the literature on zero-order optimization [80], we intend to substitute backpropagation with multiple forward or inference processes to estimate the gradients. Technically speaking, we propose the framework of BAckpropagation-Free Federated LEarning (BAFFLE). As illustrated in Figure 1, BAFFLE consists of three conceptual steps: (1) each client locally perturbs the model parameters $2K$ times as $\mathbf{W} \pm \boldsymbol{\delta}_k$ (the server sends the random seed to clients for generating $\{\pmb{\delta}_k\}_{k=1}^K$ ); (2) each client executes forward processes on the perturbed models using its private dataset $\mathbb{D}_c$ and obtains $K$ loss differences $\{\Delta \mathcal{L}(\mathbf{W}, \pmb{\delta}_k; \mathbb{D}_c)\}_{k=1}^K$ ; (3) the server aggregates loss differences to estimate gradients. + +BAFFLE's defining characteristic is that it only utilizes forward propagation, which is memory-efficient and does not require auto-differentiation. It is well-adapted to model quantization and pruning as well as inference-only hardware optimization on edge devices. Compared to backpropagation, the computation graph of BAFFLE is more easily optimized, such as by slicing it into per-layer calculation [44]. Since each loss difference $\Delta \mathcal{L}(\mathbf{W},\delta_k;\mathbb{D}_c)$ is a scalar, BAFFLE can easily accommodate the uploading bandwidth of clients by adjusting the value of $K$ as opposed to using, e.g., gradient compression [82]. BAFFLE is also compatible with recent advances in inference approaches for TEE [83, 85], providing an efficient solution for combining TEE into FL and preventing white-box evasion. + +We adapt secure aggregation [10] to zero-order optimization and investigate ways to improve gradient estimation in BAFFLE. Empirically, BAFFLE is used to train models from scratch on MNIST [49] and CIFAR-10/100 [48], and transfer ImageNet-pretrained models to OfficeHome [87]. Compared to conventional FL, it achieves suboptimal but acceptable performance. These results shed light on the potential of BAFFLE and general backpropagation-free methods in FL. + +![](images/ef0025f69632bbca34234f0731ec79e5ed18a0e337e927ce34f80ffddb92f639.jpg) +Fig. 1: A sketch map of BAFFLE. In addition to the global parameters update $\Delta \mathbf{W}$ , each client downloads random seeds to locally generate perturbations $\pm \delta_{1:K}$ and perform $2K$ times of forward propagation (i.e., inference) to compute loss differences. The server can recover these perturbations using the same random seeds and obtain $\Delta \mathcal{L}(\mathbf{W},\pmb{\delta}_k)$ by secure aggregation. Each loss difference $\Delta \mathcal{L}(\mathbf{W},\pmb{\delta}_k;\mathbb{D}_c)$ is a floating-point number, so $K$ can be easily adjusted to fit the uploading bandwidth. + +# 2 Preliminaries + +Finite difference. Gradient-based optimization techniques (either first-order or higher-order) are the most frequently used tools to train deep networks [27]. Nevertheless, recent progress demonstrates promising applications of zero-order optimization methods for training, particularly when exact derivatives cannot be obtained [23, 55, 69] or backward processes are computationally prohibitive [34, 70]. Zero-order approaches require only multiple forward processes that may be executed in parallel. Along this routine, finite difference stems from the definition of derivatives and can be generalized to higher-order and multivariate cases by Taylor's expansion. For any differentiable loss $\mathcal{L}(\mathbf{W};\mathbb{D})$ and a small perturbation $\delta \in \mathbb{R}^n$ , finite difference employs the forward difference scheme + +$$ +\mathcal {L} (\mathbf {W} + \boldsymbol {\delta}; \mathbb {D}) - \mathcal {L} (\mathbf {W}; \mathbb {D}) = \boldsymbol {\delta} ^ {\top} \nabla_ {\mathbf {W}} \mathcal {L} (\mathbf {W}; \mathbb {D}) + o (\| \boldsymbol {\delta} \| _ {2}), \tag {1} +$$ + +where $\delta^{\top}\nabla_{\mathbf{W}}\mathcal{L}(\mathbf{W};\mathbb{D})$ is a scaled directional derivative along $\pmb{\delta}$ . Furthermore, we can use the central difference scheme to obtain higher-order residuals as + +$$ +\mathcal {L} (\mathbf {W} + \boldsymbol {\delta}; \mathbb {D}) - \mathcal {L} (\mathbf {W} - \boldsymbol {\delta}; \mathbb {D}) = 2 \boldsymbol {\delta} ^ {\top} \nabla_ {\mathbf {W}} \mathcal {L} (\mathbf {W}; \mathbb {D}) + o (\| \boldsymbol {\delta} \| _ {2} ^ {2}). \tag {2} +$$ + +Federated learning. Suppose we have $C$ clients, and the $c$ -th client's private dataset is defined as $\mathbb{D}_c \coloneqq \{(\mathbf{X}_i^c, \mathbf{y}_i^c)\}_{i=1}^{N_c}$ with $N_c$ input-label pairs. Let $\mathcal{L}(\mathbf{W}; \mathbb{D}_c)$ represent the loss function for the dataset $\mathbb{D}_c$ , where $\mathbf{W} \in \mathbb{R}^n$ denotes the server model's global parameters. The training objective of FL is to find $\mathbf{W}$ that minimizes the total loss function as + +Algorithm 1 Backpropagation-free federated learning (BAFFLE) + +1: Notations: Se denotes the operations done on servers; Cl denotes the operations done on clients; TEE for the TEE module; and $\Rightarrow$ denotes the communication process. +2: Inputs: $C$ clients with local dataset $\{\mathbb{D}_c\}_{c = 1}^C$ containing $N_{c}$ input-label pairs, $N = \sum_{c = 1}^{C}N_{c}$ ; learning rate $\eta$ , training iterations $T$ , perturbation number $K$ , noise scale $\sigma$ . +3: Se: initializing model parameters $\mathbf{W} \gets \mathbf{W}_0$ ; +4: // optional + +Se: encoding the computing paradigm into TEE as $\mathrm{TEE} \circ \Delta \mathcal{L}(\mathbf{W}, \delta; \mathbb{D})$ ; + +5: for $t = 0$ to $T - 1$ do +6: $\mathrm{Se} \Rightarrow$ all Cl: downloading model parameters $\mathbf{W}_t$ and the computing paradigm; +7: // 4 Bytes + +$\mathrm{Se} \Rightarrow$ all Cl: downloading the random seed $s_t$ ; + +8: Se: sampling $K$ perturbations $\{\pmb{\delta}_k\}_{k=1}^K$ from $\mathcal{N}(0, \sigma^2\mathbf{I})$ using the random seed $s_t$ ; +9: all Cl: negotiating a group of zero-sum noises $\{\pmb{\epsilon}_c\}_{c=1}^C$ for secure aggregation; + +10: for $c = 1$ to $C$ do + +11: Cl: sampling $K$ perturbations $\{\pmb {\delta}_k\}_{k = 1}^K$ from $\mathcal{N}(0,\sigma^2\mathbf{I})$ using the random seed $s_t$ ; +12: Cl: computing TEE $\circ \Delta \mathcal{L}(\mathbf{W}_t,\pmb {\delta}_k;\mathbb{D}_c)$ via forward propagation for each $k$ + +13: // $4\times K$ Bytes + +$\mathrm{Cl} \Rightarrow \mathrm{Se}$ : uploading $K$ outputs $\left\{\mathrm{TEE} \circ \Delta \mathcal{L}(\mathbf{W}_t, \pmb{\delta}_k; \mathbb{D}_c) + \frac{N}{N_c} \pmb{\epsilon}_c\right\}_{k=1}^K$ ; + +14: end for +15: Se: aggregating $\Delta \mathcal{L}(\mathbf{W}_t,\pmb {\delta}_k)\gets \sum_{c = 1}^{C}\frac{N_c}{N}\left[\mathrm{TEE}\circ \Delta \mathcal{L}(\mathbf{W}_t,\pmb {\delta}_k;\mathbb{D}_c) + \frac{N}{N_c}\pmb {\epsilon}_c\right]$ for each $k$ +16: Se: computing $\widehat{\nabla_{\mathbf{W}_t}}\mathcal{L}(\mathbf{W}_t)\gets \frac{1}{K}\sum_{k = 1}^{K}\frac{\delta_k}{2\sigma^2}\varDelta\mathcal{L}(\mathbf{W}_t,\pmb {\delta}_k);$ +17: Se: $\mathbf{W}_{t + 1}\gets \mathbf{W}_t - \eta \widehat{\nabla\mathbf{w}_t}\mathcal{L}(\mathbf{W}_t)$ +18: end for +19: Return: final model parameters $\mathbf{W}_T$ . + +$$ +\mathcal {L} (\mathbf {W}) := \sum_ {c = 1} ^ {C} \frac {N _ {c}}{N} \mathcal {L} (\mathbf {W}; \mathbb {D} _ {c}), \text {w h e r e} N = \sum_ {c = 1} ^ {C} N _ {c}. \tag {3} +$$ + +In the conventional FL framework, clients compute gradients $\{\nabla_{\mathbf{W}}\mathcal{L}(\mathbf{W};\mathbb{D}_c)\}_{c = 1}^C$ or model updates locally through backpropagation and then upload them to the server. Federated average [62] performs global aggregation using $\varDelta\mathbf{W}:= \sum_{i=1}^{C}\frac{N_c}{N}\varDelta\mathbf{W}_c$ , where $\varDelta\mathbf{W}_c$ is the local update obtained via executing $\mathbf{W}_c\gets \mathbf{W}_c - \eta \nabla_{\mathbf{W}_c}\mathcal{L}(\mathbf{W}_c;\mathbb{D}_c)$ multiple times and $\eta$ is learning rate. + +Zeroth-order FL. Similar to our work, DLZO [54] and FedZO [22] present zeroth-order optimization methods for FL independently in batch-level and epoch-level communications. However, they concentrate primarily on basic linear models with softmax regression problems and ignore deep models. Besides, they also do not account for server security aggregation in conjunction with zero-order + +optimization. In comparison, BAFFLE enables security aggregation, can train deep models such as WideResNet from scratch and achieves reasonable results, e.g. $95.17\%$ accuracy on MNIST with 20 communication rounds versus $83.58\%$ for FedZO with 1,000 rounds. + +# 3 Backpropagation-Free Federated Learning + +In this section, we introduce zero-order optimization into FL and develop BAF-FLE, a backpropagation-free federated learning framework that uses multiple forward processes in place of backpropagation. An initial attempt is to apply finite difference as the gradient estimator. To estimate the full gradients, we need to perturb each parameter $w \in \mathbf{W}$ once to approximate the partial derivative $\frac{\partial \mathcal{L}(\mathbf{W}; \mathbb{D})}{\partial w}$ , causing the forward computations to grow with $n$ (recall that $\mathbf{W} \in \mathbb{R}^n$ ) and making it difficult to scale to large models. In light of this, we resort to Stein's identity [80] to obtain an unbiased estimation of gradients from loss differences calculated on various perturbations. As depicted in Figure 1, BAFFLE clients need only download random seeds and global parameters update, generate perturbations locally, execute multiple forward propagations and upload loss differences back to the server. Furthermore, we also present convergence analyses of BAFFLE, providing guidelines for model design and acceleration of training. + +# 3.1 Unbiased Gradient Estimation with Stein's Identity + +Previous work on sign-based optimization [66] demonstrates that deep networks can be effectively trained if the majority of gradients have proper signs. Thus, we propose performing forward propagation multiple times on perturbed parameters, in order to obtain a stochastic estimation of gradients without backpropagation. Specifically, assuming that the loss function $\mathcal{L}(\mathbf{W};\mathbb{D})$ is continuously differentiable w.r.t. $\mathbf{W}$ given any dataset $\mathbb{D}$ , which is true (almost everywhere) for deep networks using non-linear activation functions, we define a smoothed loss function as: + +$$ +\mathcal {L} _ {\sigma} (\mathbf {W}; \mathbb {D}) := \mathbb {E} _ {\boldsymbol {\delta} \sim \mathcal {N} (0, \sigma^ {2} \mathbf {I})} \mathcal {L} (\mathbf {W} + \boldsymbol {\delta}; \mathbb {D}), \tag {4} +$$ + +where the perturbation $\delta$ follows a Gaussian distribution with mean 0 and covariance $\sigma^2\mathbf{I}$ . Stein [80] proves the Stein's identity (proof recapped in Appendix A): + +$$ +\nabla_ {\mathbf {W}} \mathcal {L} _ {\sigma} (\mathbf {W}; \mathbb {D}) = \mathbb {E} _ {\boldsymbol {\delta} \sim \mathcal {N} (0, \sigma^ {2} \mathbf {I})} \left[ \frac {\boldsymbol {\delta}}{2 \sigma^ {2}} \Delta \mathcal {L} (\mathbf {W}, \boldsymbol {\delta}; \mathbb {D}) \right], \tag {5} +$$ + +where $\varDelta\mathcal{L}(\mathbf{W},\boldsymbol {\delta};\mathbb{D}):= \mathcal{L}(\mathbf{W} + \boldsymbol {\delta};\mathbb{D}) - \mathcal{L}(\mathbf{W} - \boldsymbol {\delta};\mathbb{D})$ is the loss difference. Note that computing a loss difference only requires the execution of two forwards $\mathcal{L}(\mathbf{W}+$ $\pmb{\delta};\mathbb{D})$ and $\mathcal{L}(\mathbf{W} - \pmb {\delta};\mathbb{D})$ without backpropagation. It is trivial that $\mathcal{L}_{\sigma}(\mathbf{W};\mathbb{D})$ is continuously differentiable for any $\sigma \geq 0$ and $\nabla_{\mathbf{W}}\mathcal{L}_{\sigma}(\mathbf{W};\mathbb{D})$ converges uniformly as $\sigma \rightarrow 0$ ; hence, it follows that $\nabla_{\mathbf{W}}\mathcal{L}(\mathbf{W};\mathbb{D}) = \lim_{\sigma \to 0}\nabla_{\mathbf{W}}\mathcal{L}_{\sigma}(\mathbf{W};\mathbb{D})$ . Therefore, we can obtain a stochastic estimation of gradients using Monte Carlo + +by 1) selecting a small value of $\sigma$ ; 2) randomly sampling $K$ perturbations from $\mathcal{N}(0, \sigma^2 \mathbf{I})$ as $\{\pmb{\delta}_k\}_{k=1}^K$ ; and 3) utilizing the Stein's identity in Eq. (5) to calculate + +$$ +\widehat {\nabla_ {\mathbf {W}}} \mathcal {L} (\mathbf {W}; \mathbb {D}) := \frac {1}{K} \sum_ {k = 1} ^ {K} \left[ \frac {\delta_ {k}}{2 \sigma^ {2}} \Delta \mathcal {L} (\mathbf {W}, \delta_ {k}; \mathbb {D}) \right]. \tag {6} +$$ + +# 3.2 Operating Flow of BAFFLE + +Based on the forward-only gradient estimator $\widehat{\nabla_{\mathbf{W}}}\mathcal{L}(\mathbf{W};\mathbb{D})$ derived in Eq. (6), we outline the basic operating flow of our BAFFLE system (Algorithm 1) as follows: + +Model initialization. (Lines $3\sim 4$ , done by server) The server initializes the model parameters to $\mathbf{W}_0$ and optionally encodes the computing paradigm of loss differences $\Delta \mathcal{L}(\mathbf{W},\delta ;\mathbb{D})$ into the TEE module (see Appendix B for more information on TEE); + +**Loading paradigm:** (Lines 6~7, server $\Rightarrow$ all clients) In round $t$ , the server distributes the most recent model parameters $\mathbf{W}_t$ (or the model update $\Delta \mathbf{W}_t$ ) and the computing paradigm to all the $C$ clients. In addition, in BAFFLE, the server sends a random seed $s_t$ (rather than directly sending the perturbations to reduce communication burden); + +Local computation. (Lines 11~12, done by clients) Each client generates $K$ perturbations $\{\delta_k\}_{k=1}^K$ locally from $\mathcal{N}(0, \sigma^2\mathbf{I})$ using random seed $s_t$ , and executes the computing paradigm to obtain loss differences. $K$ is chosen adaptively based on clients' computation capability; + +Uploading loss differences. (Line 13, all clients $\Rightarrow$ server) Each client uploads $K$ noisy outputs $\{\Delta \mathcal{L}(\mathbf{W}_t,\pmb {\delta}_k;\mathbb{D}_c) + \frac{N}{N_c}\pmb {\epsilon}_c\}_{k = 1}^K$ to the server, where each output is a floating-point number and the noise $\pmb{\epsilon}_{c}$ is negotiated by all clients to be zero-sum. The total uploaded Bytes is $4\times K$ + +Secure aggregation. (Lines 15~16, done by server) In order to prevent the server from recovering the exact loss differences and causing privacy leakage [25], we adopt the secure aggregation method [10] that was originally proposed for conventional FL and apply it to BAFFLE. Specifically, all clients negotiate a group of noises $\{\pmb{\epsilon}_c\}_{c=1}^C$ satisfying $\sum_{c=1}^{C} \pmb{\epsilon}_c = 0$ . Then we can reorganize our gradient estimator as + +$$ +\widehat {\nabla_ {\mathbf {W} _ {t}}} \mathcal {L} (\mathbf {W} _ {t}) = \frac {1}{K} \sum_ {c = 1} ^ {C} \frac {N _ {c}}{N} \sum_ {k = 1} ^ {K} \left[ \frac {\boldsymbol {\delta} _ {k}}{2 \sigma^ {2}} \Delta \mathcal {L} (\mathbf {W} _ {t}, \boldsymbol {\delta} _ {k}; \mathbb {D} _ {c}) \right] = \frac {1}{K} \sum_ {k = 1} ^ {K} \frac {\boldsymbol {\delta} _ {k}}{2 \sigma^ {2}} \Delta \mathcal {L} (\mathbf {W} _ {t}, \boldsymbol {\delta} _ {k}), +$$ + +$$ +\Delta \mathcal {L} \left(\mathbf {W} _ {t}, \boldsymbol {\delta} _ {k}\right) = \sum_ {c = 1} ^ {C} \frac {N _ {c}}{N} \left[ \Delta \mathcal {L} \left(\mathbf {W} _ {t}, \boldsymbol {\delta} _ {k}; \mathbb {D} _ {c}\right) + \frac {N}{N _ {c}} \boldsymbol {\epsilon} _ {c} \right]. \tag {7} +$$ + +Since $\{\pmb{\epsilon}_c\}_{c=1}^C$ are zero-sum, there is $\Delta \mathcal{L}(\mathbf{W}_t, \pmb{\delta}_k) = \sum_{c=1}^C \frac{N_c}{N} \Delta \mathcal{L}(\mathbf{W}_t, \pmb{\delta}_k; \mathbb{D}_c)$ and Eq. (7) holds. Therefore, the server can correctly aggregate $\Delta \mathcal{L}(\mathbf{W}_t, \pmb{\delta}_k)$ and protect client privacy against recovering $\Delta \mathcal{L}(\mathbf{W}_t, \pmb{\delta}_k; \mathbb{D}_c)$ . + +Remark on communication cost. After getting the gradient estimation $\widehat{\nabla_{\mathbf{W}_t}}\mathcal{L}(\mathbf{W}_t)$ , the server updates the parameters to $\mathbf{W}_{t + 1}$ using techniques such + +as gradient descent with learning rate $\eta$ . Similar to the discussion in McMahan et al. [62], the BAFFLE form presented in Algorithm 1 corresponds to the batch-level communication (also named FedSGD) where Lines $11\sim 12$ execute once for each round $t$ . In batch-level settings, we reduce the uploaded Bytes from $4\times |\mathbf{W}|$ to $4\times K$ . We can generalize BAFFLE to an analog of epoch-level communication (also named FedAvg), in which each client updates its local parameters multiple steps using the gradient estimator $\widehat{\nabla_{\mathbf{W}_t}}\mathcal{L}(\mathbf{W}_t,\mathbb{D}_c)$ derived from $\Delta \mathcal{L}(\mathbf{W}_t,\delta_k;\mathbb{D}_c)$ via Eq. (6), and upload model updates to the server after several local epochs. In epoch-level settings, the uploaded Bytes are the same as FedAvg. In experiments, we analyze both batch-level and epoch-level settings for BAFFLE and report the results. + +# 3.3 Convergence Analyses + +Now we analyze the convergence rate of our gradient estimation method. For continuously differentiable loss functions, we have $\nabla_{\mathbf{W}}\mathcal{L}(\mathbf{W};\mathbb{D}) = \lim_{\sigma \to 0}\nabla_{\mathbf{W}}\mathcal{L}_{\sigma}(\mathbf{W};\mathbb{D})$ so we choose a relatively small value for $\sigma$ . The convergence guarantee can be derived as follows: + +Theorem 1. (Proof in Appendix A) Suppose $\sigma$ is a small value and the central difference scheme in Eq. (2) holds. For perturbations $\{\pmb{\delta}_k\}_{k=1}^K \stackrel{\mathrm{iid}}{\sim} \mathcal{N}(0, \sigma^2 \mathbf{I})$ , the empirical covariance matrix is $\widehat{\pmb{\Sigma}} := \frac{1}{K \sigma^2} \sum_{k=1}^K \pmb{\delta}_k \pmb{\delta}_k^T$ and mean is $\widehat{\pmb{\delta}} := \frac{1}{K} \sum_{k=1}^K \pmb{\delta}_k$ . Then for any $\mathbf{W} \in \mathbb{R}^n$ , the relation between $\widehat{\nabla_{\mathbf{W}}} \mathcal{L}(\mathbf{W}; \mathbb{D})$ and the true gradient $\nabla_{\mathbf{W}} \mathcal{L}(\mathbf{W}; \mathbb{D})$ can be written as + +$$ +\widehat {\nabla_ {\mathbf {W}}} \mathcal {L} (\mathbf {W}; \mathbb {D}) = \widehat {\boldsymbol {\Sigma}} \nabla_ {\mathbf {W}} \mathcal {L} (\mathbf {W}; \mathbb {D}) + o (\widehat {\boldsymbol {\delta}}), \tag {8} +$$ + +where $\mathbb{E}[\widehat{\pmb{\Sigma}}] = \mathbf{I}$ , $\mathbb{E}[\widehat{\boldsymbol{\delta}}] = \mathbf{0}$ . + +Taking the expectation on both sides of Eq. (8), we obtain $\mathbb{E}[\widehat{\nabla_{\mathbf{W}}}\mathcal{L}(\mathbf{W};\mathbb{D})] = \nabla_{\mathbf{W}}\mathcal{L}(\mathbf{W};\mathbb{D})$ , which degrades to Stein's identity. To determine the convergence rate w.r.t. the value of $K$ , we have: + +Theorem 2. (Adamczak et al. [3]) With overwhelming probability, the empirical covariance matrix satisfies the inequality $\| \widehat{\mathbf{\Sigma}} -\mathbf{I}\| _2\leq C_0\sqrt{\frac{n}{K}}$ , where $\| \cdot \| _2$ denotes the 2-norm for matrix and $C_0$ is an absolute positive constant. + +Note that in the finetuning setting, $n$ represents the number of trainable parameters, excluding frozen parameters. As concluded, $\widehat{\nabla_{\mathbf{W}}}\mathcal{L}(\mathbf{W};\mathbb{D})$ provides an unbiased estimation for the true gradients with convergence rate of $\mathcal{O}\left(\sqrt{\frac{n}{K}}\right)$ . Empirically, $\widehat{\nabla_{\mathbf{W}}}\mathcal{L}(\mathbf{W};\mathbb{D})$ is used as a noisy gradient to train models, the generalization of which has been analyzed in previous work [50, 103]. + +# 4 Experiments + +We evaluate BAFFLE on 4 benchmark datasets: MNIST [49], CIFAR-10/100 [48] and OfficeHome [87]. We consider three models: 1) LeNet [49] with two convolutional layers as the shallow model ( $2.7 \times 10^{4}$ parameters); 2) WideResNet [98] + +Table 1: The classification accuracy (\%) of BAFFLE in iid scenarios ( $C = 10$ ) and epoch-level communication settings with different $K$ values ( $K_{1} / K_{2}$ annotations mean using $K_{1}$ for MNIST and $K_{2}$ for CIFAR-10/100). In this configuration, each client updates its local model based on BAFFLE estimated gradients and uploads model updates to the server after an entire epoch on the local dataset. The four guidelines work well under epoch-level settings with total communication rounds 20/40 for MNIST and CIFAR-10/100. + +
SettingsLeNetWRN-10-2
MNISTCIFAR-10CIFAR-100MNISTCIFAR-10CIFAR-100
K100/20087.2748.7841.5488.3552.2746.61
200/50089.4851.8245.6889.5755.5951.65
500/100092.1853.6248.7295.1758.6353.15
Ablation Study (100/200)w/o EMA85.0647.9736.8185.8950.0145.86
ReLU81.5544.9939.4979.0849.7644.44
SELU86.1848.6537.3476.4443.3741.79
Central76.0245.9736.5377.4542.8939.62
BP Baselines94.3158.7554.6797.1162.2960.08
+ +Table 2: The accuracy $(\%)$ of BAFFLE in label non-iid scenarios $(C = 100)$ and epoch-level settings with total communication rounds 40 and different $K$ values. We employ Dirichlet dist. with $\alpha = 0.3$ to ensure each client has a unique label distribution. + +
SettingsLeNet CIFAR-10 | CIFAR-100WRN-10-2 CIFAR-10 | CIFAR-100
K20035.21 | 28.1239.53 | 30.44
50038.14 | 30.9241.69 | 32.89
100039.71 | 33.3543.42 | 34.08
BP Baselines44.41 | 38.4351.18 | 40.85
+ +with depth $= 10$ and width $= 2$ (WRN-10-2) as the light weight deep model $(3.0 \times 10^{5}$ parameters) and 3) MobileNet [38] as the deep neural networks $(1.3 \times 10^{7}$ parameters) that works on ImageNet. + +Participation and communication settings. To perform a comprehensive evaluation of BAFFLE, we simulate three popular FL scenarios [17] with the FedLab [99] participations: iid participations, label non-iid participations and feature non-iid participations. For iid participations, we set the client number $C = 10$ and use uniform distribution to build local datasets. Then we evaluate our BAFFLE on MNIST and CIFAR-10/100 under both batch-level (FedSGD) and epoch-level (FedAvg) communication settings. For label non-iid participations, we set client number $C = 100$ , use Dirichlet distribution with $\alpha = 0.3$ to build clients. For feature non-iid participations, we build clients from the prevailing domain adaptation dataset OfficeHome, which contains 65 categories from 4 different domains, i.e. Art, Clipart, Product and Real-world. We set the + +![](images/a1ce11cb3cbf01404dd6cb86b7f30597656d75f14cfd81cdecf3519eff06f653.jpg) + +![](images/8da81ede411fd39381a7acb952ac0bf7110674536c8ea0ba8fe804dc5802ba59.jpg) + +![](images/3fe921cd0a5dade0c69db3f6c4176ee080b9b34a98e2fbb6056d6beac6680870.jpg) +LeNet + +![](images/2568a114b196847cb3766fa2fbf239c6d630b1572d1e8cf6d0cd4dac877203f8.jpg) +WRN-10-2 +Fig. 2: The classification accuracy (\%) of BAFFLE in iid scenarios ( $C = 10$ ) and batch-level communication settings with various $K$ values. We treat the models trained by exact gradients on conventional FL systems as the backpropagation (BP) baselines. On different datasets and architectures, our BAFFLE achieves comparable performance to the exact gradient results with a reasonable $K$ . + +total client number to $C = 40$ and generate 10 clients from each domain. As results, we report Top-1 accuracy for MNIST, CIFAR-10 and OfficeHome and Top-5 accuracy for OfficeHome and CIFAR-100. + +Hyperparameters. Following the settings in Section 2, we use FedAVG to aggregate gradients from multiple clients and use SGD-based optimizer to update global parameters. Specifically, we use Adam [46] to train a random initialized model with $\beta = (0.9,0.99)$ , learning rate 0.01 and epochs 20/40 for MNIST and CIFAR-10/100. For OfficeHome, we adapt the transfer learning [40] by loading the ImageNet-pretrained model and finetuning the final layers with Adam, but setting learning rate 0.005 and epochs 40. In BAFFLE, the perturbation scale $\sigma$ and number $K$ are the most important hyperparameters. As shown in Theorem 1, with less noise and more samples, the BAFFLE will obtain more accurate gradients, leading to improved performance. However, there exists a trade-off between accuracy and computational efficiency: an extremely small $\sigma$ will cause the underflow problem [27] and a large $K$ will increase computational cost. In practice, we empirically set $\sigma = 10^{-4}$ because it is the smallest value that does not cause numerical problems in all experiments, and works well on edge devices with half-precision floating-point numbers. We also evaluate the impact of $K$ across a broad range from 100 to 5000. + +# 4.1 Four Guidelines for BAFFLE + +For a general family of continuously differentiable models, we analyze their convergence rate of BAFFLE in Section 3.3. Since deep networks are usually stacked with multiple linear layers and non-linear activation, this layer linearity can be utilized to improve the accuracy-efficiency trade-off. Combining the linearity property and the unique conditions in edge devices (e.g., small data size and half-precision format), we present four guidelines for model design and training that can increase accuracy without introducing extra computation (Appendix C shows the details of linearity analysis): + +Using twice forward difference (twice-FD) scheme rather than central scheme. Combining difference scheme Eq. (1) and Eq. (2), we find that by executing twice as many forward inferences (i.e. $\mathbf{W} \pm \boldsymbol{\delta}$ ), the central scheme achieves lower residuals than twice-FD, despite the fact that twice-FD can benefit from additional sample times. With the same forward times (e.g., $2K$ ), determining which scheme performs better is a practical issue. As shown in Appendix C, we find that twice-FD performs better in all experiments, in part because the linearity reduces the benefit from second-order residuals. + +Using Hardswish in BAFFLE. ReLU is effective when the middle features $(h(\cdot)$ denotes the feature mapping) have the same sign before and after perturbations, i.e. $h(\mathbf{W} + \boldsymbol{\delta})\cdot h(\mathbf{W}) > 0$ . Since ReLU is not differentiable at zero, the value jump occurs when the sign of features changes after perturbations, i.e. $h(\mathbf{W} + \boldsymbol{\delta})\cdot h(\mathbf{W}) < 0$ . We use Hardswish [37] to overcome this problem as it is continuously differentiable at zero and easy to implement on edge devices. + +Using exponential moving average (EMA) to reduce oscillations. As shown in Theorem 1, there exists an zero-mean white-noise $\widehat{\delta}$ between the true gradient and our estimation. To smooth out the oscillations caused by white noise, we apply EMA strategies from BYOL [28] to the global parameters, with a smoothing coefficient of 0.995. + +Using GroupNorm as opposed to BatchNorm. On edge devices, the dataset size is typically small, which leads to inaccurate batch statistics estimation and degrades performance when using BatchNorm. Thus we employ GroupNorm [94] to solve this issue. + +# 4.2 Performance on IID Clients + +Following the settings in Section 4.1, we evaluate the performance of BAFFLE in the iid scenarios. We reproduce all experiments on the BP-based FL systems with the same settings and use them as the baselines. We refer to the baseline results as exact gradients and report the training process of BAFFLE in Figure 2. The value of $K$ (e.g., 200 for LeNet and 500 for WRN-10-2) is much less than the dimensions of parameter space (e.g., $2.7 \times 10^4$ for LeNet and $3 \times 10^5$ for WRN-10-2). Since the convergence rate to the exact gradient is $\mathcal{O}\left(\sqrt{\frac{n}{K}}\right)$ , the marginal benefit of increasing $K$ decreases. For instance, increasing $K$ from 2000 to 5000 on CIFAR-10 with WRN-10-2 barely improves accuracy by $2\%$ . Given that the + +Table 3: The Top-1|Top-5 accuracy (\%) of BAFFLE on OfficeHome with feature non-iid participations ( $C = 40$ ) and epoch-level settings with 40 comm. rounds. We use the pretrained MobileNet, freeze the backbone and finetune the FC layers. + +
SettingsDomainsAvg.
ArtClipartProductReal World
2044.75|69.4652.48|73.8866.63|89.7763.78|87.8356.91|80.24
5047.87|71.3253.43|76.8371.28|91.7467.02|89.9559.90|82.46
K 10050.32|74.4257.43|80.7374.19|93.0269.53|90.4362.87|84.65
20051.42|76.6460.98|86.4176.05|94.4271.51|93.1464.98|87.65
50053.33|77.8562.58|86.6478.84|95.2373.17|93.8566.98|88.40
BP Baselines55.71|80.4365.13|88.6582.44|96.0577.19|95.0470.12|90.04
+ +convergence rate of Gaussian perturbations is $\mathcal{O}\left(\sqrt{\frac{n}{K}}\right)$ , the sampling efficiency may be improved by choosing an alternative distribution for perturbations. + +Ablation studies. As depicted in Figure 3, we conduct ablation studies for BAFFLE to evaluate the aforementioned guidelines. In general, twice-FD, Hardswish and EMA can all improve the accuracy. For two difference schemes, we compare the twice-FD to central scheme with the same computation cost and show that the former outperforms the later, demonstrating that linearity reduces the gain from second-order residuals. As to activation functions, Hardswish is superior to ReLU and SELU because it is differentiable at zero and vanishes to zero in the negative part. Moreover, EMA enhances the performance of training strategies by reducing the effect of white noise. + +Communication efficiency. Compared to the batch-level communication settings (FedSGD) in a BP-based FL system, BAFFLE requires each client to upload a $K$ -dimensional vector to the server and downloads the updated global parameters in each communication round. Since $K$ is significantly less than the parameter amounts (e.g., 500 versus 0.3 million), BAFFLE reduces data transfer by approximately half. To reduce communication costs, the prevalent FL system requires each client to perform model optimization on the local training dataset and upload the model updates to the server after a specified number of local epochs. BAFFLE can also perform epoch-level communications by employing an $\mathcal{O}(n)$ additional memory to store the perturbation in each forward and estimate the local gradient using Eq. (6). Then each client optimizes the local model with SGD and uploads local updates after several epochs. As shown in Table 1, we evaluate the performance of BAFFLE under one-epoch communication settings. As epoch-level communication is more prevalent in the real-world FL, all the following experiments will be conducted in this context. In brief, BAFFLE uploads the same Bytes as BP-based FL in epoch-level communication while the total communication rounds are much less than FedZO [22], e.g. 20 versus 1000 on MNIST. + +# 4.3 Performance on Non-IID Clients + +Following Section 4.1, we evaluate the performance of BAFFLE in both label non-iid and feature non-iid scenarios. For label non-iid scenarios, we use the + +![](images/dd059e2c8e099f1a1b5c7263d248d88bb6a848b6c87d126b74995e61d751f6ca.jpg) + +![](images/25c4bb93dfc6f715f6f970d6fbf027483f6ccd621d92bdbb2ff1c79d5515921f.jpg) + +![](images/23ebf4cff94ea2c9dd73d7d5bb532ed93041385e3b494647897946d91f7d3668.jpg) +LeNet + +![](images/6dfb402337254ea1a4544638c5f36452f0047e51115a9cdce482f0671e6fc905.jpg) +WRN-10-2 +Fig. 3: The ablation study of BAFFLE guidelines, with $K = 100$ on MNIST and $K = 500$ on CIFAR-10. As seen, twice-FD, Hardswish, and EMA all improve performance without extra computation. EMA reduces oscillations by lessening Gaussian noise. + +CIFAR-10/100 datasets and employ Dirichlet distribution to ensure that each client has a unique label distribution. We evaluate the performance of BAFFLE with 100 clients and various K values. As seen in Table 2, the model suffers a significant drop in accuracy (e.g., $14\%$ in CIFAR-10 and $16\%$ in CIFAR-100) due to the label non-iid effect. For feature non-iid scenarios, we construct clients using the OfficeHome dataset and use MobileNet as the deep model. As seen in Table 3, we use the transfer learning strategy to train MobileNet, i.e., we load the parameters pretrained on ImageNet, freeze the backbone parameters, and retrain the classification layers. The accuracy decrease is approximately $3\% \sim 5\%$ . + +# 4.4 Computation Efficiency, Memory and Robustness + +BAFFLE uses $K$ times forward passes instead of backward. Since the backward pass is about as expensive as two normal forward passes [35] and five single-precision accelerated forward passes [67], BAFFLE results in approximately $\frac{K}{5}$ times the computation expense of BP-based FL. Although BAFFLE results in $\frac{K}{5} - 1$ times extra computation cost, we show the cost can be reduced with proper training strategies, e.g., the transfer learning in Table 3 can reduce $K$ to 20 on the MobileNet and the $224 \times 224$ sized OfficeHome dataset. + +Moreover, BAFFLE can reduce huge memory cost on edge devices with the efficiency in static memory and dynamic memory. The auto-differential framework is used to run BP on deep networks, which requires extra static memory + +Table 4: The GPU memory cost (MB) of vanilla BP and BAFFLE, respectively. 'min~max' denotes the minimum and maximum dynamic memory for BAFFLE. We also report the ratio (\%) of vanilla BP to BAFFLE's max memory cost. + +
BackboneCIFAR-10/100OfficeHome/ImageNet
BPBAFFLERatioBPBAFFLERatio
LeNet168067~17410.35252786~2017.95
WRN-10-2187875~19610.43342594~2517.32
MobileNet2041102~21710.635271121~2895.48
+ +(e.g., 200MB for Caffe [41] and 1GB for Pytorch [72]) and imposes a considerable burden on edge devices such as IoT sensors. Due to the necessity of restoring intermediate states, BP also requires enormous amounts of dynamic memory ( $\geq$ 5GB for MobileNet [24]). Since BAFFLE only requires inference, we can slice the computation graph and execute the forwards per layer [44]. As shown in Table 4, BAFFLE reduces the memory cost to $5\% \sim 10\%$ by executing layer-by-layer inference. By applying kernel-wise computations, we can further reduce the memory cost to approximately $1\%$ (e.g., 64MB for MobileNet [85]), which is suitable for scenarios with extremely limited storage resources, such as TEE. + +Recent works exploit TEE to protect models from white-box attacks by preventing model exposure [44]. However, due to the security guarantee, the usable memory of TEE is usually small [85] (e.g., 90MB on Intel SGX for Skylake CPU [61]), which is typically far less than what a backpropagation-based FL system requires. In contrast, BAFFLE can execute in TEE due to its little memory cost (more details are in Appendix B). Membership inference attacks and model inversion attacks need to repeatedly perform model inference and obtain confidence values or classification scores [78, 101]. Given that BAFFLE provides stochastic loss differences $\Delta \mathcal{L}(\mathbf{W},\delta ;\mathbb{D})$ associated with the random perturbation $\delta$ , the off-the-shelf inference attacks may not perform on BAFFLE directly (while adaptively designed attacking strategies are possible to evade BAFFLE). Motivated by differential privacy [1], we further design heuristic experiments to study the information leakage from $\Delta \mathcal{L}$ (details in Appendix D). As shown in Figure 5, the $\Delta \mathcal{L}$ between real data and random noise is hard to distinguish, indicating it is difficult for attackers to obtain useful information from BAFFLE's outputs. + +# 5 Related Work + +Along the research routine of FL, many efforts have been devoted to, e.g., dealing with non-IID distributions [21, 53, 75, 90, 102], multi-task learning [60, 79], and preserving privacy of clients [11, 12, 26, 31, 56, 58, 63, 84]. Below we introduce the work on efficiency and vulnerability in FL following the survey of Kairouz et al. [42], which is more related to this paper. + +Efficiency in FL. It is widely understood that the communication and computational efficiency is a primary bottleneck for deploying FL in practice [6,18,74, + +89, 92]. Specifically, communicating between the server and clients could be potentially expensive and unreliable. The seminal work of Konečný et al. [47] introduces sparsification and quantization to reduce the communication cost, where several theoretical works investigate the optimal trade-off between the communication cost and model accuracy [2, 7, 15, 30, 100]. Since practical clients usually have slower upload than download bandwidth, much research interest focuses on gradient compression [4, 8, 36, 82]. On the other hand, different methods have been proposed to reduce the computational burden of local clients [16, 29, 33], since these clients are usually edge devices with limited resources. Training paradigms exploiting tensor factorization in FL can also achieve promising performance [45, 59]. + +Vulnerability in FL. The characteristic of decentralization in FL is beneficial to protecting data privacy of clients, but in the meanwhile, providing white-box accessibility of model status leaves flexibility for malicious clients to perform poisoning/backdoor attacks [5, 9, 71, 88, 95], model/gradient inversion attacks [25, 39, 101], and membership inference attacks [57, 68, 78]. To alleviate the vulnerability in FL, several defense strategies have been proposed via selecting reliable clients [43], data augmentation [13], update clipping [81], robust training [52], model perturbation [97], detection methods [20, 76], and methods based on differential privacy [93]. + +# 6 Conclusion and Discussion + +Backpropagation is the gold standard for training deep networks, and it is also utilized by traditional FL systems. However, backpropagation is unsuited for edge devices due to their limited resources and possible lack of reliability. Using zero-order optimization techniques, we explore the possibility of BAFFLE in this paper. We need to specify that there are scenarios in which clients are fully trusted and have sufficient computing and storage resources. In these situations, traditional FL with backpropagation is preferred. + +While our preliminary studies on BAFFLE have generated encouraging results, there are still a number of tough topics to investigate: (i) Compared to the models trained using exact gradients, the accuracy of models trained using BAFFLE is inferior. One reason is that we select small values of $K$ (e.g., 500) relative to the number of model parameters (e.g., $3.0 \times 10^{5}$ ); another reason is that gradient descent is designed for exact gradients, whereas our noisy gradient estimation may require advanced learning algorithms. (ii) The empirical variance of zero-order gradient estimators affects training convergence in BAFFLE. It is crucial to research variance reduction approaches, such as control variates and non-Gaussian sampling distributions. (iii) Stein's identity is proposed for loss functions with Gaussian noises imposed on model parameters. Intuitively, this smoothness is related to differential privacy in FL, but determining their relationship requires theoretical derivations. + +# Acknowledgments + +This paper is supported by the National Science Foundation of China (62132017), Zhejiang Provincial Natural Science Foundation of China (LD24F020011). + +# References + +1. Abadi, M., Chu, A., Goodfellow, I.J., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, Vienna, Austria, October 24-28, 2016 (2016) +2. Acharya, J., Canonne, C.L., Tyagi, H.: Inference under information constraints i: Lower bounds from chi-square contraction. IEEE Transactions on Information Theory (2020) +3. Adamczak, R., Litvak, A.E., Pajor, A., Tomczak-Jaegermann, N.: Sharp bounds on the rate of convergence of the empirical covariance matrix. Comptes Rendus Mathematique 349(3-4), 195-200 (2011) +4. Alistarh, D., Grubic, D., Li, J., Tomioka, R., Vojnovic, M.: Qsgd: Communication-efficient sgd via gradient quantization and encoding. Advances in Neural Information Processing Systems (NeurIPS) (2017) +5. Bagdasaryan, E., Veit, A., Hua, Y., Estrin, D., Shmatikov, V.: How to back-door federated learning. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2020) +6. Balakrishnan, R., Li, T., Zhou, T., Himayat, N., Smith, V., Bilmes, J.: Diverse client selection for federated learning via submodular maximization. In: International Conference on Learning Representations (ICLR) (2022) +7. Barnes, L.P., Han, Y., Özgür, A.: Lower bounds for learning distributions under communication constraints via fisher information. Journal of Machine Learning Research (JMLR) (2020) +8. Basu, D., Data, D., Karakus, C., Diggavi, S.: Qsparse-local-sgd: Distributed sgd with quantization, sparsification and local computations. Advances in Neural Information Processing Systems (NeurIPS) (2019) +9. Bhagoji, A.N., Chakraborty, S., Mittal, P., Calo, S.: Analyzing federated learning through an adversarial lens. In: International Conference on Machine Learning (ICML) (2019) +10. Bonawitz, K.A., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for privacy-preserving machine learning. In: ACM Conference on Computer and Communications Security (CCS) (2017) +11. Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for federated learning on user-held data. arXiv preprint arXiv:1611.04482 (2016) +12. Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for privacy-preserving machine learning. In: ACM SIGSAC Conference on Computer and Communications Security (2017) +13. Borgnia, E., Geiping, J., Cherepanova, V., Fowl, L., Gupta, A., Ghiasi, A., Huang, F., Goldblum, M., Goldstein, T.: Dp-instahide: Provably defusing poisoning and backdoor attacks with differentially private data augmentations. arXiv preprint arXiv:2103.02079 (2021) + +14. Bradbury, J., Frostig, R., Hawkins, P., Johnson, M.J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., Zhang, Q.: JAX: composable transformations of Python+NumPy programs (2018), http://github.com/google/jax +15. Braverman, M., Garg, A., Ma, T., Nguyen, H.L., Woodruff, D.P.: Communication lower bounds for statistical estimation problems via a distributed data processing inequality. In: ACM Symposium on Theory of Computing (2016) +16. Caldas, S., Konečny, J., McMahan, H.B., Talwalkar, A.: Expanding the reach of federated learning by reducing client resource requirements. arXiv preprint arXiv:1812.07210 (2018) +17. Caldas, S., Wu, P., Li, T., Konečny, J., McMahan, H.B., Smith, V., Talwalkar, A.: LEAF: A benchmark for federated settings. CoRR (2018) +18. Chen, M., Shlezinger, N., Poor, H.V., Eldar, Y.C., Cui, S.: Communication-efficient federated learning. Proceedings of the National Academy of Sciences (2021) +19. Chen, Y., Luo, F., Li, T., Xiang, T., Liu, Z., Li, J.: A training-integrity privacy-preserving federated learning scheme with trusted execution environment. Information Sciences (2020) +20. Dong, Y., Yang, X., Deng, Z., Pang, T., Xiao, Z., Su, H., Zhu, J.: Black-box detection of backdoor attacks with limited information and data. In: IEEE International Conference on Computer Vision (ICCV) (2021) +21. Eichner, H., Koren, T., McMahan, B., Srebro, N., Talwar, K.: Semi-cyclic stochastic gradient descent. In: International Conference on Machine Learning (ICML) (2019) +22. Fang, W., Yu, Z., Jiang, Y., Shi, Y., Jones, C.N., Zhou, Y.: Communication-efficient stochastic zeroth-order optimization for federated learning. IEEE Trans. Signal Process. (2022) +23. Flaxman, A.D., Kalai, A.T., McMahan, H.B.: Online convex optimization in the bandit setting: gradient descent without a gradient. arXiv preprint cs/0408007 (2004) +24. Gao, Y., Liu, Y., Zhang, H., Li, Z., Zhu, Y., Lin, H., Yang, M.: Estimatinggpu memory consumption of deep learning models. In: ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering (2020) +25. Geiping, J., Bauermeister, H., Dröge, H., Moeller, M.: Inverting gradients-how easy is it to break privacy in federated learning? In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +26. Ghazi, B., Kumar, R., Manurangsi, P., Pagh, R.: Private counting from anonymous messages: Near-optimal accuracy with vanishing communication overhead. In: International Conference on Machine Learning (ICML) (2020) +27. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press (2016) +28. Grill, J., Strub, F., Altché, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Doersch, C., Pires, B.A., Guo, Z., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent - A new approach to self-supervised learning. In: NeurIPS 2020 (2020) +29. Hamer, J., Mohri, M., Suresh, A.T.: Fedboost: A communication-efficient algorithm for federated learning. In: International Conference on Machine Learning (ICML) (2020) +30. Han, Y., Ozgür, A., Weissman, T.: Geometric lower bounds for distributed parameter estimation under communication constraints. In: Conference On Learning Theory (COLT) (2018) + +31. Hao, M., Li, H., Luo, X., Xu, G., Yang, H., Liu, S.: Efficient and privacy-enhanced federated learning for industrial artificial intelligence. IEEE Transactions on Industrial Informatics (2019) +32. Hard, A., Rao, K., Mathews, R., Ramaswamy, S., Beaufays, F., Augenstein, S., Eichner, H., Kiddon, C., Ramage, D.: Federated learning for mobile keyboard prediction. arXiv preprint arXiv:1811.03604 (2018) +33. He, C., Annavaram, M., Avestimehr, S.: Group knowledge transfer: Federated learning of large cnns at the edge. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +34. He, D., Shi, W., Li, S., Gao, X., Zhang, J., Bian, J., Wang, L., Liu, T.Y.: Learning physics-informed neural networks without stacked back-propagation. arXiv preprint arXiv:2202.09340 (2022) +35. Hinton, G., Srivastava, N.: Csc321: Introduction to neural networks and machine learning. Lecture 10 (2010) +36. Horváth, S., Ho, C.Y., Horvath, L., Sahu, A.N., Canini, M., Richtárik, P.: Natural compression for distributed deep learning. arXiv preprint arXiv:1905.10988 (2019) +37. Howard, A., Pang, R., Adam, H., Le, Q.V., Sandler, M., Chen, B., Wang, W., Chen, L., Tan, M., Chu, G., Vasudevan, V., Zhu, Y.: Searching for mobilenetv3. In: IEEE International Conference on Computer Vision (ICCV) (2019) +38. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR (2017) +39. Huang, Y., Gupta, S., Song, Z., Li, K., Arora, S.: Evaluating gradient inversion attacks and defenses in federated learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +40. Huh, M., Agrawal, P., Efros, A.A.: What makes imagenet good for transfer learning? CoRR (2016) +41. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R.B., Guadarrama, S., Darrell, T.: Caffe: Convolutional architecture for fast feature embedding. In: Hua, K.A., Rui, Y., Steinmetz, R., Hanjalic, A., Natsev, A., Zhu, W. (eds.) Proceedings of the ACM International Conference on Multimedia (2014) +42. Kairouz, P., McMahan, H.B., Avent, B., Bellet, A., Dennis, M., Bhagoji, A.N., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R., et al.: Advances and open problems in federated learning. Foundations and Trends in Machine Learning 14(1-2), 1-210 (2021) +43. Kang, J., Xiong, Z., Niyato, D., Zou, Y., Zhang, Y., Guizani, M.: Reliable federated learning for mobile networks. IEEE Wireless Communications (2020) +44. Kim, K., Kim, C.H., Rhee, J.J., Yu, X., Chen, H., Tian, D., Lee, B.: Vessels: Efficient and scalable deep learning prediction on trusted processors. In: ACM Symposium on Cloud Computing (2020) +45. Kim, Y., Sun, J., Yu, H., Jiang, X.: Federated tensor factorization for computational phenotyping. In: ACM International Conference on Knowledge Discovery and Data Mining (SIGKDD) (2017) +46. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: 3rd International Conference on Learning Representations, ICLR 2015 (2015) +47. Konečny, J., McMahan, H.B., Yu, F.X., Richtárik, P., Suresh, A.T., Bacon, D.: Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492 (2016) +48. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images (2009) + +49. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of IEEE (1998) +50. Li, J., Luo, X., Qiao, M.: On generalization error bounds of noisy gradient methods for non-convex learning. In: International Conference on Learning Representations (ICLR) (2020) +51. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering (2020) +52. Li, T., Hu, S., Beirami, A., Smith, V.: Ditto: Fair and robust federated learning through personalization. In: International Conference on Machine Learning (ICML) (2021) +53. Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data. In: International Conference on Learning Representations (ICLR) (2020) +54. Li, Z., Chen, L.: Communication-efficient decentralized zeroth-order method on heterogeneous data. In: WCSP 2021 (2021) +55. Liu, S., Chen, P.Y., Kailkhura, B., Zhang, G., Hero III, A.O., Varshney, P.K.: A primer on zeroth-order optimization in signal processing and machine learning: Principals, recent advances, and applications. IEEE Signal Processing Magazine (2020) +56. Liu, Y., Suresh, A.T., Yu, F.X.X., Kumar, S., Riley, M.: Learning discrete distributions: user vs item-level privacy. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +57. Luo, X., Wu, Y., Xiao, X., Ooi, B.C.: Feature inference attack on model predictions in vertical federated learning. In: IEEE International Conference on Data Engineering (ICDE) (2021) +58. Lyu, L., Yu, H., Ma, X., Sun, L., Zhao, J., Yang, Q., Yu, P.S.: Privacy and robustness in federated learning: Attacks and defenses. IEEE Transactions on Neural Networks and Learning Systems pp. 1-21 (2022) +59. Ma, J., Zhang, Q., Lou, J., Ho, J.C., Xiong, L., Jiang, X.: Privacy-preserving tensor factorization for collaborative health data analysis. In: ACM International Conference on Information and Knowledge Management (CIKM) (2019) +60. Marfoq, O., Neglia, G., Bellet, A., Kameni, L., Vidal, R.: Federated multi-task learning under a mixture of distributions. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +61. McKeen, F., Alexandrovich, I., Anati, I., Caspi, D., Johnson, S., Leslie-Hurd, R., Rozas, C.: Intel® software guard extensions (intel® sgx) support for dynamic memory management inside an enclave. In: Proceedings of the Hardware and Architectural Support for Security and Privacy (2016) +62. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2017) +63. McMahan, B., Ramage, D., Talwar, K., Zhang, L.: Learning differentially private recurrent language models. In: International Conference on Learning Representations (ICLR) (2018) +64. Mo, F., Haddadi, H., Katevas, K., Marin, E., Perino, D., Courtellis, N.: Ppf: privacy-preserving federated learning with trusted execution environments. In: Annual International Conference on Mobile Systems, Applications, and Services (2021) +65. Mondal, A., More, Y., Rooparaghunath, R.H., Gupta, D.: Flatee: Federated learning across trusted execution environments. arXiv preprint arXiv:2111.06867 (2021) + +66. Moulay, E., Léchappé, V., Plestan, F.: Properties of the sign gradient descent algorithms. Inf. Sci. 492, 29-39 (2019) +67. Nakandala, S., Nagrecha, K., Kumar, A., Papakonstantinou, Y.: Incremental and approximate computations for accelerating deep CNN inference. ACM Trans. Database Syst. 45 (2020) +68. Nasr, M., Shokri, R., Houmansadr, A.: Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In: IEEE symposium on security and privacy (S&P). IEEE (2019) +69. Nesterov, Y., Spokoiny, V.: Random gradient-free minimization of convex functions. Foundations of Computational Mathematics (2017) +70. Pang, T., Xu, K., Li, C., Song, Y., Ermon, S., Zhu, J.: Efficient learning of generative models via finite-difference score matching. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +71. Pang, T., Yang, X., Dong, Y., Su, H., Zhu, J.: Accumulative poisoning attacks on real-time data. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +72. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E.Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems 32 (2019) +73. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) +74. Rothchild, D., Panda, A., Ullah, E., Ivkin, N., Stoica, I., Braverman, V., Gonzalez, J., Arora, R.: Fetchsgd: Communication-efficient federated learning with sketching. In: International Conference on Machine Learning (ICML) (2020) +75. Sattler, F., Wiedemann, S., Müller, K.R., Samek, W.: Robust and communication-efficient federated learning from non-iid data. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2019) +76. Seetharaman, S., Malaviya, S., KV, R., Shukla, M., Lodha, S.: Influence based defense against data poisoning attacks in online learning. arXiv preprint arXiv:2104.13230 (2021) +77. Sharma, H., Park, J., Suda, N., Lai, L., Chau, B., Chandra, V., Esmaeilzadeh, H.: Bit fusion: Bit-level dynamically composable architecture for accelerating deep neural network. In: IEEE Annual International Symposium on Computer Architecture (ISCA) (2018) +78. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. In: IEEE symposium on security and privacy (S&P). IEEE (2017) +79. Smith, V., Chiang, C.K., Sanjabi, M., Talwalkar, A.S.: Federated multi-task learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2017) +80. Stein, C.M.: Estimation of the mean of a multivariate normal distribution. The annals of Statistics (1981) +81. Sun, Z., Kairouz, P., Suresh, A.T., McMahan, H.B.: Can you really backdoor federated learning? arXiv preprint arXiv:1911.07963 (2019) +82. Suresh, A.T., Felix, X.Y., Kumar, S., McMahan, H.B.: Distributed mean estimation with limited communication. In: International Conference on Machine Learning (ICML) (2017) + +83. Tramer, F., Boneh, D.: Slalom: Fast, verifiable and private execution of neural networks in trusted hardware. In: International Conference on Learning Representations (ICLR) (2019) +84. Truex, S., Baracaldo, N., Anwar, A., Steinke, T., Ludwig, H., Zhang, R., Zhou, Y.: A hybrid approach to privacy-preserving federated learning. In: ACM workshop on Artificial Intelligence and Security (2019) +85. Truong, J.B., Gallagher, W., Guo, T., Walls, R.J.: Memory-efficient deep learning inference in trusted execution environments. In: IEEE International Conference on Cloud Engineering (IC2E) (2021) +86. Umuroglu, Y., Rasnayake, L., Själander, M.: Bismo: A scalable bit-serial matrix multiplication overlay for reconfigurable computing. In: International Conference on Field Programmable Logic and Applications (FPL) (2018) +87. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: CVPR (2017) +88. Wang, H., Sreenivasan, K., Rajput, S., Vishwakarma, H., Agarwal, S., Sohn, J.y., Lee, K., Papailiopoulos, D.: Attack of the tails: Yes, you really can backdoor federated learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +89. Wang, H.P., Stich, S., He, Y., Fritz, M.: Prog fed: effective, communication, and computation efficient federated learning by progressive training. In: International Conference on Machine Learning (ICML) (2022) +90. Wang, J., Liu, Q., Liang, H., Joshi, G., Poor, H.V.: Tackling the objective inconsistency problem in heterogeneous federated optimization. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +91. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +92. Wang, L., Wang, W., Li, B.: Cmfl: Mitigating communication overhead for federated learning. In: IEEE International Conference on Distributed Computing Systems (ICDCS) (2019) +93. Wei, K., Li, J., Ding, M., Ma, C., Yang, H.H., Farokhi, F., Jin, S., Quek, T.Q., Poor, H.V.: Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security (2020) +94. Wu, Y., He, K.: Group normalization. Int. J. Comput. Vis. 128(3), 742-755 (2020) +95. Xie, C., Huang, K., Chen, P.Y., Li, B.: Distributed backdoor attacks against federated learning. In: International Conference on Learning Representations (ICLR) (2020) +96. Yang, T., Andrew, G., Eichner, H., Sun, H., Li, W., Kong, N., Ramage, D., Beaufays, F.: Applied federated learning: Improving google keyboard query suggestions. arXiv preprint arXiv:1812.02903 (2018) +97. Yang, X., Feng, Y., Fang, W., Shao, J., Tang, X., Xia, S.T., Lu, R.: An accuracy-lossless perturbation method for defending privacy attacks in federated learning. In: ACM Web Conference (WWW) (2022) +98. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: Proceedings of the British Machine Vision Conference 2016 (2016) +99. Zeng, D., Liang, S., Hu, X., Wang, H., Xu, Z.: Fedlab: A flexible federated learning framework. arXiv preprint arXiv:2107.11621 (2021) +100. Zhang, Y., Duchi, J., Jordan, M.I., Wainwright, M.J.: Information-theoretic lower bounds for distributed statistical estimation with communication constraints. Advances in Neural Information Processing Systems (NeurIPS) (2013) + +101. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +102. Zhao, Y., Li, M., Lai, L., Suda, N., Civin, D., Chandra, V.: Federated learning with non-iid data. arXiv preprint arXiv:1806.00582 (2018) +103. Zhu, Z., Wu, J., Yu, B., Wu, L., Ma, J.: The anisotropic noise in stochastic gradient descent: Its behavior of escaping from sharp minima and regularization effects. In: International Conference on Machine Learning (ICML) (2019) \ No newline at end of file diff --git a/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/images.zip b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..036080e15e5c7aec269ef6682d2a1f964aa99377 --- /dev/null +++ b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca41d9c625cfaa816b7ac58e9f61552af4e64334dc474bbe96265bb8c967910 +size 473230 diff --git a/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/layout.json b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f37ff403b9b7418582a72fb5a0624e4a9e0d1f54 --- /dev/null +++ b/2024/BAFFLE_ A Baseline of Backpropagation-Free Federated Learning/layout.json @@ -0,0 +1,14248 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 45, + 30, + 368, + 66 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 30, + 368, + 66 + ], + "spans": [ + { + "bbox": [ + 45, + 30, + 368, + 66 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "spans": [ + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "text", + "content": "Haozhe Feng" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "inline_equation", + "content": "^{1,3\\star}" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "text", + "content": ", Tianyu Pang" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "inline_equation", + "content": "^{2\\star\\star}" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "text", + "content": ", Chao Du" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "text", + "content": ", Wei Chen" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "inline_equation", + "content": "^{1\\star\\star}" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "text", + "content": ", Shuicheng Yan" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "text", + "content": ", and Min Lin" + }, + { + "bbox": [ + 75, + 86, + 338, + 111 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 101, + 120, + 310, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 120, + 310, + 177 + ], + "spans": [ + { + "bbox": [ + 101, + 120, + 310, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 120, + 310, + 177 + ], + "type": "text", + "content": " State Key Lab of CAD&CG, Zhejiang University " + }, + { + "bbox": [ + 101, + 120, + 310, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 101, + 120, + 310, + 177 + ], + "type": "text", + "content": " Sea AI Lab, Singapore " + }, + { + "bbox": [ + 101, + 120, + 310, + 177 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 101, + 120, + 310, + 177 + ], + "type": "text", + "content": " Tencent Inc. {fenghz, chenvis} @zju.edu.cn {tianyupang, duchao, yansc, linmin} @sea.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 209, + 351, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 351, + 384 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 351, + 384 + ], + "type": "text", + "content": "Abstract. Federated learning (FL) is a general principle for decentralized clients to train a server model collectively without sharing local data. FL is a promising framework with practical applications, but its standard training paradigm requires the clients to backpropagate through the model to compute gradients. Since these clients are typically edge devices and not fully trusted, executing backpropagation on them incurs computational and storage overhead as well as white-box vulnerability. In light of this, we develop backpropagation-free federated learning, dubbed BAFFLE, in which backpropagation is replaced by multiple forward processes to estimate gradients. BAFFLE is 1) memory-efficient and easily fits uploading bandwidth; 2) compatible with inference-only hardware optimization and model quantization or pruning; and 3) well-suited to trusted execution environments, because the clients in BAFFLE only execute forward propagation and return a set of scalars to the server. Empirically we use BAFFLE to train deep models from scratch or to finetune pretrained models, achieving acceptable results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 396, + 351, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 396, + 351, + 418 + ], + "spans": [ + { + "bbox": [ + 61, + 396, + 351, + 418 + ], + "type": "text", + "content": "Keywords: Federated Learning " + }, + { + "bbox": [ + 61, + 396, + 351, + 418 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 61, + 396, + 351, + 418 + ], + "type": "text", + "content": " Finite Difference " + }, + { + "bbox": [ + 61, + 396, + 351, + 418 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 61, + 396, + 351, + 418 + ], + "type": "text", + "content": " Backpropagation-Free Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 440, + 128, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 440, + 128, + 452 + ], + "spans": [ + { + "bbox": [ + 33, + 440, + 128, + 452 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 468, + 379, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 468, + 379, + 552 + ], + "spans": [ + { + "bbox": [ + 32, + 468, + 379, + 552 + ], + "type": "text", + "content": "Federated learning (FL) allows decentralized clients to collaboratively train a server model [62]. In each training round, the selected clients compute model gradients or updates on their local private datasets, without explicitly exchanging sample points to the server. While FL describes a promising blueprint and has several applications [32, 51, 96], the mainstream training paradigm of FL is still gradient-based that requires the clients to locally execute backpropagation, which leads to two practical limitations:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 562, + 234, + 572 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 562, + 234, + 572 + ], + "spans": [ + { + "bbox": [ + 33, + 562, + 234, + 572 + ], + "type": "text", + "content": "* Work done during an internship at Sea AI Lab." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 573, + 135, + 583 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 573, + 135, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 573, + 135, + 583 + ], + "type": "text", + "content": "\\*\\* Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 381, + 232 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 30, + 34, + 381, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 381, + 130 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 381, + 130 + ], + "type": "text", + "content": "(i) Overhead for edge devices. The clients in FL are usually edge devices, such as mobile phones and IoT sensors, whose hardware is primarily optimized for inference-only purposes [77,86], rather than for backpropagation. Due to the limited resources, computationally affordable models running on edge devices are typically quantized and pruned [91], making exact backpropagation difficult. In addition, standard implementations of backpropagation rely on either forward-mode or reverse-mode auto-differentiation in contemporary machine learning packages [14, 73], which increases storage requirements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 135, + 381, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 135, + 381, + 232 + ], + "spans": [ + { + "bbox": [ + 30, + 135, + 381, + 232 + ], + "type": "text", + "content": "(ii) White-box vulnerability. To facilitate gradient computing, the server regularly distributes its model status to the clients, but this white-box exposure of the model renders the server vulnerable to, e.g., poisoning or inversion attacks from malicious clients [25,78,95,101]. With that, recent attempts are made to exploit trusted execution environments (TEEs) in FL, which can isolate the model status within a black-box secure area and significantly reduce the success rate of malicious evasion [19,64,65]. However, TEEs are highly memory-constrained [85], while backpropagation is memory-consuming to restore intermediate states." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "spans": [ + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "text", + "content": "While numerous solutions have been proposed to alleviate these limitations (related work discussed in Section 5), we raise an essential question: how to perform backpropagation-free " + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "inline_equation", + "content": "FL" + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "text", + "content": "? Inspired by the literature on zero-order optimization [80], we intend to substitute backpropagation with multiple forward or inference processes to estimate the gradients. Technically speaking, we propose the framework of BAckpropagation-Free Federated LEarning (BAFFLE). As illustrated in Figure 1, BAFFLE consists of three conceptual steps: (1) each client locally perturbs the model parameters " + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "text", + "content": " times as " + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\pm \\boldsymbol{\\delta}_k" + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "text", + "content": " (the server sends the random seed to clients for generating " + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\delta}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "text", + "content": "); (2) each client executes forward processes on the perturbed models using its private dataset " + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c" + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "text", + "content": " and obtains " + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "text", + "content": " loss differences " + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "inline_equation", + "content": "\\{\\Delta \\mathcal{L}(\\mathbf{W}, \\pmb{\\delta}_k; \\mathbb{D}_c)\\}_{k=1}^K" + }, + { + "bbox": [ + 30, + 237, + 381, + 381 + ], + "type": "text", + "content": "; (3) the server aggregates loss differences to estimate gradients." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 386, + 381, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 386, + 381, + 507 + ], + "spans": [ + { + "bbox": [ + 30, + 386, + 381, + 507 + ], + "type": "text", + "content": "BAFFLE's defining characteristic is that it only utilizes forward propagation, which is memory-efficient and does not require auto-differentiation. It is well-adapted to model quantization and pruning as well as inference-only hardware optimization on edge devices. Compared to backpropagation, the computation graph of BAFFLE is more easily optimized, such as by slicing it into per-layer calculation [44]. Since each loss difference " + }, + { + "bbox": [ + 30, + 386, + 381, + 507 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W},\\delta_k;\\mathbb{D}_c)" + }, + { + "bbox": [ + 30, + 386, + 381, + 507 + ], + "type": "text", + "content": " is a scalar, BAFFLE can easily accommodate the uploading bandwidth of clients by adjusting the value of " + }, + { + "bbox": [ + 30, + 386, + 381, + 507 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 386, + 381, + 507 + ], + "type": "text", + "content": " as opposed to using, e.g., gradient compression [82]. BAFFLE is also compatible with recent advances in inference approaches for TEE [83, 85], providing an efficient solution for combining TEE into FL and preventing white-box evasion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 511, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 511, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 511, + 381, + 584 + ], + "type": "text", + "content": "We adapt secure aggregation [10] to zero-order optimization and investigate ways to improve gradient estimation in BAFFLE. Empirically, BAFFLE is used to train models from scratch on MNIST [49] and CIFAR-10/100 [48], and transfer ImageNet-pretrained models to OfficeHome [87]. Compared to conventional FL, it achieves suboptimal but acceptable performance. These results shed light on the potential of BAFFLE and general backpropagation-free methods in FL." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 35, + 36, + 377, + 187 + ], + "blocks": [ + { + "bbox": [ + 35, + 36, + 377, + 187 + ], + "lines": [ + { + "bbox": [ + 35, + 36, + 377, + 187 + ], + "spans": [ + { + "bbox": [ + 35, + 36, + 377, + 187 + ], + "type": "image", + "image_path": "ef0025f69632bbca34234f0731ec79e5ed18a0e337e927ce34f80ffddb92f639.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "lines": [ + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "spans": [ + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "text", + "content": "Fig. 1: A sketch map of BAFFLE. In addition to the global parameters update " + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "text", + "content": ", each client downloads random seeds to locally generate perturbations " + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "inline_equation", + "content": "\\pm \\delta_{1:K}" + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "text", + "content": " and perform " + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "text", + "content": " times of forward propagation (i.e., inference) to compute loss differences. The server can recover these perturbations using the same random seeds and obtain " + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W},\\pmb{\\delta}_k)" + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "text", + "content": " by secure aggregation. Each loss difference " + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W},\\pmb{\\delta}_k;\\mathbb{D}_c)" + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "text", + "content": " is a floating-point number, so " + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 31, + 196, + 379, + 262 + ], + "type": "text", + "content": " can be easily adjusted to fit the uploading bandwidth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 285, + 132, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 285, + 132, + 297 + ], + "spans": [ + { + "bbox": [ + 31, + 285, + 132, + 297 + ], + "type": "text", + "content": "2 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 313, + 381, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 313, + 381, + 432 + ], + "spans": [ + { + "bbox": [ + 30, + 313, + 381, + 432 + ], + "type": "text", + "content": "Finite difference. Gradient-based optimization techniques (either first-order or higher-order) are the most frequently used tools to train deep networks [27]. Nevertheless, recent progress demonstrates promising applications of zero-order optimization methods for training, particularly when exact derivatives cannot be obtained [23, 55, 69] or backward processes are computationally prohibitive [34, 70]. Zero-order approaches require only multiple forward processes that may be executed in parallel. Along this routine, finite difference stems from the definition of derivatives and can be generalized to higher-order and multivariate cases by Taylor's expansion. For any differentiable loss " + }, + { + "bbox": [ + 30, + 313, + 381, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 30, + 313, + 381, + 432 + ], + "type": "text", + "content": " and a small perturbation " + }, + { + "bbox": [ + 30, + 313, + 381, + 432 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 30, + 313, + 381, + 432 + ], + "type": "text", + "content": ", finite difference employs the forward difference scheme" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 441, + 379, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 441, + 379, + 456 + ], + "spans": [ + { + "bbox": [ + 85, + 441, + 379, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {W} + \\boldsymbol {\\delta}; \\mathbb {D}) - \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) = \\boldsymbol {\\delta} ^ {\\top} \\nabla_ {\\mathbf {W}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) + o (\\| \\boldsymbol {\\delta} \\| _ {2}), \\tag {1}", + "image_path": "e43c099713a83d07ba96869974abb85b7fbf4b97774521c323a6c6120d6e4d0c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 465, + 379, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 465, + 379, + 490 + ], + "spans": [ + { + "bbox": [ + 31, + 465, + 379, + 490 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 465, + 379, + 490 + ], + "type": "inline_equation", + "content": "\\delta^{\\top}\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 31, + 465, + 379, + 490 + ], + "type": "text", + "content": " is a scaled directional derivative along " + }, + { + "bbox": [ + 31, + 465, + 379, + 490 + ], + "type": "inline_equation", + "content": "\\pmb{\\delta}" + }, + { + "bbox": [ + 31, + 465, + 379, + 490 + ], + "type": "text", + "content": ". Furthermore, we can use the central difference scheme to obtain higher-order residuals as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 499, + 379, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 499, + 379, + 514 + ], + "spans": [ + { + "bbox": [ + 74, + 499, + 379, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {W} + \\boldsymbol {\\delta}; \\mathbb {D}) - \\mathcal {L} (\\mathbf {W} - \\boldsymbol {\\delta}; \\mathbb {D}) = 2 \\boldsymbol {\\delta} ^ {\\top} \\nabla_ {\\mathbf {W}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) + o (\\| \\boldsymbol {\\delta} \\| _ {2} ^ {2}). \\tag {2}", + "image_path": "7449fa4897f815f40959bf0e5657a1a2a4eb5e8e5bb2674a84e0fb968cee176f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "spans": [ + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "content": "Federated learning. Suppose we have " + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "content": " clients, and the " + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "content": "-th client's private dataset is defined as " + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c \\coloneqq \\{(\\mathbf{X}_i^c, \\mathbf{y}_i^c)\\}_{i=1}^{N_c}" + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "content": " input-label pairs. Let " + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{W}; \\mathbb{D}_c)" + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "content": " represent the loss function for the dataset " + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c" + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "content": " denotes the server model's global parameters. The training objective of FL is to find " + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 31, + 523, + 380, + 583 + ], + "type": "text", + "content": " that minimizes the total loss function as" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 347, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 323, + 46 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 323, + 46 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 323, + 46 + ], + "type": "text", + "content": "Algorithm 1 Backpropagation-free federated learning (BAFFLE)" + } + ] + } + ], + "index": 2, + "type": "text" + }, + { + "bbox": [ + 35, + 49, + 380, + 137 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 35, + 49, + 380, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 49, + 380, + 83 + ], + "spans": [ + { + "bbox": [ + 35, + 49, + 380, + 83 + ], + "type": "text", + "content": "1: Notations: Se denotes the operations done on servers; Cl denotes the operations done on clients; TEE for the TEE module; and " + }, + { + "bbox": [ + 35, + 49, + 380, + 83 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 35, + 49, + 380, + 83 + ], + "type": "text", + "content": " denotes the communication process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "spans": [ + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "content": "2: Inputs: " + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "content": " clients with local dataset " + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{D}_c\\}_{c = 1}^C" + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "content": " input-label pairs, " + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "inline_equation", + "content": "N = \\sum_{c = 1}^{C}N_{c}" + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "content": "; learning rate " + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "content": ", training iterations " + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "content": ", perturbation number " + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "content": ", noise scale " + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 35, + 83, + 380, + 116 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 36, + 116, + 228, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 116, + 228, + 127 + ], + "spans": [ + { + "bbox": [ + 36, + 116, + 228, + 127 + ], + "type": "text", + "content": "3: Se: initializing model parameters " + }, + { + "bbox": [ + 36, + 116, + 228, + 127 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\gets \\mathbf{W}_0" + }, + { + "bbox": [ + 36, + 116, + 228, + 127 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 36, + 127, + 95, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 127, + 95, + 137 + ], + "spans": [ + { + "bbox": [ + 36, + 127, + 95, + 137 + ], + "type": "text", + "content": "4: // optional" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 48, + 138, + 340, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 138, + 340, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 138, + 340, + 148 + ], + "type": "text", + "content": "Se: encoding the computing paradigm into TEE as " + }, + { + "bbox": [ + 48, + 138, + 340, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{TEE} \\circ \\Delta \\mathcal{L}(\\mathbf{W}, \\delta; \\mathbb{D})" + }, + { + "bbox": [ + 48, + 138, + 340, + 148 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 36, + 150, + 376, + 182 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 36, + 150, + 138, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 150, + 138, + 159 + ], + "spans": [ + { + "bbox": [ + 36, + 150, + 138, + 159 + ], + "type": "text", + "content": "5: for " + }, + { + "bbox": [ + 36, + 150, + 138, + 159 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 36, + 150, + 138, + 159 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 36, + 150, + 138, + 159 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 36, + 150, + 138, + 159 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 36, + 160, + 376, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 160, + 376, + 170 + ], + "spans": [ + { + "bbox": [ + 36, + 160, + 376, + 170 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 36, + 160, + 376, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{Se} \\Rightarrow" + }, + { + "bbox": [ + 36, + 160, + 376, + 170 + ], + "type": "text", + "content": " all Cl: downloading model parameters " + }, + { + "bbox": [ + 36, + 160, + 376, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_t" + }, + { + "bbox": [ + 36, + 160, + 376, + 170 + ], + "type": "text", + "content": " and the computing paradigm;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 36, + 171, + 102, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 171, + 102, + 182 + ], + "spans": [ + { + "bbox": [ + 36, + 171, + 102, + 182 + ], + "type": "text", + "content": "7: // 4 Bytes" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 58, + 182, + 244, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 182, + 244, + 192 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 244, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{Se} \\Rightarrow" + }, + { + "bbox": [ + 58, + 182, + 244, + 192 + ], + "type": "text", + "content": " all Cl: downloading the random seed " + }, + { + "bbox": [ + 58, + 182, + 244, + 192 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 58, + 182, + 244, + 192 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 36, + 193, + 376, + 214 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "spans": [ + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "text", + "content": "8: Se: sampling " + }, + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "text", + "content": " perturbations " + }, + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\delta}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, \\sigma^2\\mathbf{I})" + }, + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "text", + "content": " using the random seed " + }, + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 36, + 193, + 376, + 204 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 36, + 204, + 369, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 204, + 369, + 214 + ], + "spans": [ + { + "bbox": [ + 36, + 204, + 369, + 214 + ], + "type": "text", + "content": "9: all Cl: negotiating a group of zero-sum noises " + }, + { + "bbox": [ + 36, + 204, + 369, + 214 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\epsilon}_c\\}_{c=1}^C" + }, + { + "bbox": [ + 36, + 204, + 369, + 214 + ], + "type": "text", + "content": " for secure aggregation;" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 33, + 215, + 134, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 215, + 134, + 224 + ], + "spans": [ + { + "bbox": [ + 33, + 215, + 134, + 224 + ], + "type": "text", + "content": "10: for " + }, + { + "bbox": [ + 33, + 215, + 134, + 224 + ], + "type": "inline_equation", + "content": "c = 1" + }, + { + "bbox": [ + 33, + 215, + 134, + 224 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 33, + 215, + 134, + 224 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 33, + 215, + 134, + 224 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 225, + 379, + 258 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "spans": [ + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "text", + "content": "11: Cl: sampling " + }, + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "text", + "content": " perturbations " + }, + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "inline_equation", + "content": "\\{\\pmb {\\delta}_k\\}_{k = 1}^K" + }, + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\sigma^2\\mathbf{I})" + }, + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "text", + "content": " using the random seed " + }, + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 33, + 225, + 379, + 246 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 247, + 367, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 247, + 367, + 258 + ], + "spans": [ + { + "bbox": [ + 33, + 247, + 367, + 258 + ], + "type": "text", + "content": "12: Cl: computing TEE " + }, + { + "bbox": [ + 33, + 247, + 367, + 258 + ], + "type": "inline_equation", + "content": "\\circ \\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k;\\mathbb{D}_c)" + }, + { + "bbox": [ + 33, + 247, + 367, + 258 + ], + "type": "text", + "content": " via forward propagation for each " + }, + { + "bbox": [ + 33, + 247, + 367, + 258 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 33, + 259, + 132, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 259, + 132, + 269 + ], + "spans": [ + { + "bbox": [ + 33, + 259, + 132, + 269 + ], + "type": "text", + "content": "13: // " + }, + { + "bbox": [ + 33, + 259, + 132, + 269 + ], + "type": "inline_equation", + "content": "4\\times K" + }, + { + "bbox": [ + 33, + 259, + 132, + 269 + ], + "type": "text", + "content": " Bytes" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 269, + 347, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 269, + 347, + 289 + ], + "spans": [ + { + "bbox": [ + 68, + 269, + 347, + 289 + ], + "type": "inline_equation", + "content": "\\mathrm{Cl} \\Rightarrow \\mathrm{Se}" + }, + { + "bbox": [ + 68, + 269, + 347, + 289 + ], + "type": "text", + "content": ": uploading " + }, + { + "bbox": [ + 68, + 269, + 347, + 289 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 269, + 347, + 289 + ], + "type": "text", + "content": " outputs " + }, + { + "bbox": [ + 68, + 269, + 347, + 289 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathrm{TEE} \\circ \\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k; \\mathbb{D}_c) + \\frac{N}{N_c} \\pmb{\\epsilon}_c\\right\\}_{k=1}^K" + }, + { + "bbox": [ + 68, + 269, + 347, + 289 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 33, + 289, + 379, + 386 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 33, + 289, + 93, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 289, + 93, + 296 + ], + "spans": [ + { + "bbox": [ + 33, + 289, + 93, + 296 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 33, + 297, + 379, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 297, + 379, + 328 + ], + "spans": [ + { + "bbox": [ + 33, + 297, + 379, + 328 + ], + "type": "text", + "content": "15: Se: aggregating " + }, + { + "bbox": [ + 33, + 297, + 379, + 328 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k)\\gets \\sum_{c = 1}^{C}\\frac{N_c}{N}\\left[\\mathrm{TEE}\\circ \\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k;\\mathbb{D}_c) + \\frac{N}{N_c}\\pmb {\\epsilon}_c\\right]" + }, + { + "bbox": [ + 33, + 297, + 379, + 328 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 33, + 297, + 379, + 328 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 33, + 329, + 273, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 329, + 273, + 351 + ], + "spans": [ + { + "bbox": [ + 33, + 329, + 273, + 351 + ], + "type": "text", + "content": "16: Se: computing " + }, + { + "bbox": [ + 33, + 329, + 273, + 351 + ], + "type": "inline_equation", + "content": "\\widehat{\\nabla_{\\mathbf{W}_t}}\\mathcal{L}(\\mathbf{W}_t)\\gets \\frac{1}{K}\\sum_{k = 1}^{K}\\frac{\\delta_k}{2\\sigma^2}\\varDelta\\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k);" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 33, + 351, + 193, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 351, + 193, + 364 + ], + "spans": [ + { + "bbox": [ + 33, + 351, + 193, + 364 + ], + "type": "text", + "content": "17: Se: " + }, + { + "bbox": [ + 33, + 351, + 193, + 364 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{t + 1}\\gets \\mathbf{W}_t - \\eta \\widehat{\\nabla\\mathbf{w}_t}\\mathcal{L}(\\mathbf{W}_t)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 33, + 365, + 83, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 365, + 83, + 373 + ], + "spans": [ + { + "bbox": [ + 33, + 365, + 83, + 373 + ], + "type": "text", + "content": "18: end for" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 33, + 375, + 203, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 375, + 203, + 386 + ], + "spans": [ + { + "bbox": [ + 33, + 375, + 203, + 386 + ], + "type": "text", + "content": "19: Return: final model parameters " + }, + { + "bbox": [ + 33, + 375, + 203, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_T" + }, + { + "bbox": [ + 33, + 375, + 203, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 102, + 423, + 379, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 423, + 379, + 455 + ], + "spans": [ + { + "bbox": [ + 102, + 423, + 379, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {W}) := \\sum_ {c = 1} ^ {C} \\frac {N _ {c}}{N} \\mathcal {L} (\\mathbf {W}; \\mathbb {D} _ {c}), \\text {w h e r e} N = \\sum_ {c = 1} ^ {C} N _ {c}. \\tag {3}", + "image_path": "0afc44fb45b67e09090f741e15403f63bde452c3aa6db9931b34700edb85132d.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "spans": [ + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "text", + "content": "In the conventional FL framework, clients compute gradients " + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "inline_equation", + "content": "\\{\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}_c)\\}_{c = 1}^C" + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "text", + "content": " or model updates locally through backpropagation and then upload them to the server. Federated average [62] performs global aggregation using " + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "inline_equation", + "content": "\\varDelta\\mathbf{W}:= \\sum_{i=1}^{C}\\frac{N_c}{N}\\varDelta\\mathbf{W}_c" + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "inline_equation", + "content": "\\varDelta\\mathbf{W}_c" + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "text", + "content": " is the local update obtained via executing " + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_c\\gets \\mathbf{W}_c - \\eta \\nabla_{\\mathbf{W}_c}\\mathcal{L}(\\mathbf{W}_c;\\mathbb{D}_c)" + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "text", + "content": " multiple times and " + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 30, + 461, + 380, + 523 + ], + "type": "text", + "content": " is learning rate." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 30, + 524, + 382, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 524, + 382, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 524, + 382, + 584 + ], + "type": "text", + "content": "Zeroth-order FL. Similar to our work, DLZO [54] and FedZO [22] present zeroth-order optimization methods for FL independently in batch-level and epoch-level communications. However, they concentrate primarily on basic linear models with softmax regression problems and ignore deep models. Besides, they also do not account for server security aggregation in conjunction with zero-order" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": "optimization. In comparison, BAFFLE enables security aggregation, can train deep models such as WideResNet from scratch and achieves reasonable results, e.g. " + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "95.17\\%" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": " accuracy on MNIST with 20 communication rounds versus " + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "83.58\\%" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": " for FedZO with 1,000 rounds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 100, + 301, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 100, + 301, + 115 + ], + "spans": [ + { + "bbox": [ + 31, + 100, + 301, + 115 + ], + "type": "text", + "content": "3 Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "spans": [ + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "text", + "content": "In this section, we introduce zero-order optimization into FL and develop BAF-FLE, a backpropagation-free federated learning framework that uses multiple forward processes in place of backpropagation. An initial attempt is to apply finite difference as the gradient estimator. To estimate the full gradients, we need to perturb each parameter " + }, + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "inline_equation", + "content": "w \\in \\mathbf{W}" + }, + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "text", + "content": " once to approximate the partial derivative " + }, + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\mathcal{L}(\\mathbf{W}; \\mathbb{D})}{\\partial w}" + }, + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "text", + "content": ", causing the forward computations to grow with " + }, + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "text", + "content": " (recall that " + }, + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 30, + 124, + 381, + 283 + ], + "type": "text", + "content": ") and making it difficult to scale to large models. In light of this, we resort to Stein's identity [80] to obtain an unbiased estimation of gradients from loss differences calculated on various perturbations. As depicted in Figure 1, BAFFLE clients need only download random seeds and global parameters update, generate perturbations locally, execute multiple forward propagations and upload loss differences back to the server. Furthermore, we also present convergence analyses of BAFFLE, providing guidelines for model design and acceleration of training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 299, + 320, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 299, + 320, + 312 + ], + "spans": [ + { + "bbox": [ + 31, + 299, + 320, + 312 + ], + "type": "text", + "content": "3.1 Unbiased Gradient Estimation with Stein's Identity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 319, + 380, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 319, + 380, + 415 + ], + "spans": [ + { + "bbox": [ + 30, + 319, + 380, + 415 + ], + "type": "text", + "content": "Previous work on sign-based optimization [66] demonstrates that deep networks can be effectively trained if the majority of gradients have proper signs. Thus, we propose performing forward propagation multiple times on perturbed parameters, in order to obtain a stochastic estimation of gradients without backpropagation. Specifically, assuming that the loss function " + }, + { + "bbox": [ + 30, + 319, + 380, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 30, + 319, + 380, + 415 + ], + "type": "text", + "content": " is continuously differentiable w.r.t. " + }, + { + "bbox": [ + 30, + 319, + 380, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 30, + 319, + 380, + 415 + ], + "type": "text", + "content": " given any dataset " + }, + { + "bbox": [ + 30, + 319, + 380, + 415 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 30, + 319, + 380, + 415 + ], + "type": "text", + "content": ", which is true (almost everywhere) for deep networks using non-linear activation functions, we define a smoothed loss function as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 122, + 424, + 377, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 424, + 377, + 436 + ], + "spans": [ + { + "bbox": [ + 122, + 424, + 377, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\sigma} (\\mathbf {W}; \\mathbb {D}) := \\mathbb {E} _ {\\boldsymbol {\\delta} \\sim \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I})} \\mathcal {L} (\\mathbf {W} + \\boldsymbol {\\delta}; \\mathbb {D}), \\tag {4}", + "image_path": "fd5c7fca2b15ff8fb78ca44914e19f1bf215130fb8d9260038e71e62ee69203c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 445, + 379, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 445, + 379, + 470 + ], + "spans": [ + { + "bbox": [ + 31, + 445, + 379, + 470 + ], + "type": "text", + "content": "where the perturbation " + }, + { + "bbox": [ + 31, + 445, + 379, + 470 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 31, + 445, + 379, + 470 + ], + "type": "text", + "content": " follows a Gaussian distribution with mean 0 and covariance " + }, + { + "bbox": [ + 31, + 445, + 379, + 470 + ], + "type": "inline_equation", + "content": "\\sigma^2\\mathbf{I}" + }, + { + "bbox": [ + 31, + 445, + 379, + 470 + ], + "type": "text", + "content": ". Stein [80] proves the Stein's identity (proof recapped in Appendix A):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 98, + 478, + 377, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 478, + 377, + 503 + ], + "spans": [ + { + "bbox": [ + 98, + 478, + 377, + 503 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {W}} \\mathcal {L} _ {\\sigma} (\\mathbf {W}; \\mathbb {D}) = \\mathbb {E} _ {\\boldsymbol {\\delta} \\sim \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I})} \\left[ \\frac {\\boldsymbol {\\delta}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W}, \\boldsymbol {\\delta}; \\mathbb {D}) \\right], \\tag {5}", + "image_path": "1b1c53b0d1848266047df2570190e90c06315fb4dd7a25caed5619d00105166c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\varDelta\\mathcal{L}(\\mathbf{W},\\boldsymbol {\\delta};\\mathbb{D}):= \\mathcal{L}(\\mathbf{W} + \\boldsymbol {\\delta};\\mathbb{D}) - \\mathcal{L}(\\mathbf{W} - \\boldsymbol {\\delta};\\mathbb{D})" + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "content": " is the loss difference. Note that computing a loss difference only requires the execution of two forwards " + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{W}+" + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\pmb{\\delta};\\mathbb{D})" + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{W} - \\pmb {\\delta};\\mathbb{D})" + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "content": " without backpropagation. It is trivial that " + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "content": " is continuously differentiable for any " + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\sigma \\geq 0" + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{W}}\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "content": " converges uniformly as " + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\sigma \\rightarrow 0" + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "content": "; hence, it follows that " + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}) = \\lim_{\\sigma \\to 0}\\nabla_{\\mathbf{W}}\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 31, + 511, + 380, + 584 + ], + "type": "text", + "content": ". Therefore, we can obtain a stochastic estimation of gradients using Monte Carlo" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "text", + "content": "by 1) selecting a small value of " + }, + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "text", + "content": "; 2) randomly sampling " + }, + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "text", + "content": " perturbations from " + }, + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, \\sigma^2 \\mathbf{I})" + }, + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\delta}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 30, + 34, + 380, + 59 + ], + "type": "text", + "content": "; and 3) utilizing the Stein's identity in Eq. (5) to calculate" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 68, + 379, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 68, + 379, + 99 + ], + "spans": [ + { + "bbox": [ + 106, + 68, + 379, + 99 + ], + "type": "interline_equation", + "content": "\\widehat {\\nabla_ {\\mathbf {W}}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) := \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left[ \\frac {\\delta_ {k}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W}, \\delta_ {k}; \\mathbb {D}) \\right]. \\tag {6}", + "image_path": "28af052a4eeeec4f55d492c45f0c8d787b79d6331926d17b877eaad221730055.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 116, + 202, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 116, + 202, + 128 + ], + "spans": [ + { + "bbox": [ + 31, + 116, + 202, + 128 + ], + "type": "text", + "content": "3.2 Operating Flow of BAFFLE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 135, + 379, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 135, + 379, + 159 + ], + "spans": [ + { + "bbox": [ + 30, + 135, + 379, + 159 + ], + "type": "text", + "content": "Based on the forward-only gradient estimator " + }, + { + "bbox": [ + 30, + 135, + 379, + 159 + ], + "type": "inline_equation", + "content": "\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 30, + 135, + 379, + 159 + ], + "type": "text", + "content": " derived in Eq. (6), we outline the basic operating flow of our BAFFLE system (Algorithm 1) as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 159, + 380, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 159, + 380, + 206 + ], + "spans": [ + { + "bbox": [ + 30, + 159, + 380, + 206 + ], + "type": "text", + "content": "Model initialization. (Lines " + }, + { + "bbox": [ + 30, + 159, + 380, + 206 + ], + "type": "inline_equation", + "content": "3\\sim 4" + }, + { + "bbox": [ + 30, + 159, + 380, + 206 + ], + "type": "text", + "content": ", done by server) The server initializes the model parameters to " + }, + { + "bbox": [ + 30, + 159, + 380, + 206 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_0" + }, + { + "bbox": [ + 30, + 159, + 380, + 206 + ], + "type": "text", + "content": " and optionally encodes the computing paradigm of loss differences " + }, + { + "bbox": [ + 30, + 159, + 380, + 206 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W},\\delta ;\\mathbb{D})" + }, + { + "bbox": [ + 30, + 159, + 380, + 206 + ], + "type": "text", + "content": " into the TEE module (see Appendix B for more information on TEE);" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "spans": [ + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "text", + "content": "**Loading paradigm:** (Lines 6~7, server " + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "text", + "content": " all clients) In round " + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "text", + "content": ", the server distributes the most recent model parameters " + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_t" + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "text", + "content": " (or the model update " + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}_t" + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "text", + "content": ") and the computing paradigm to all the " + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "text", + "content": " clients. In addition, in BAFFLE, the server sends a random seed " + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 30, + 207, + 379, + 266 + ], + "type": "text", + "content": " (rather than directly sending the perturbations to reduce communication burden);" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "spans": [ + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "text", + "content": "Local computation. (Lines 11~12, done by clients) Each client generates " + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "text", + "content": " perturbations " + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "inline_equation", + "content": "\\{\\delta_k\\}_{k=1}^K" + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "text", + "content": " locally from " + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, \\sigma^2\\mathbf{I})" + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "text", + "content": " using random seed " + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "text", + "content": ", and executes the computing paradigm to obtain loss differences. " + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 266, + 379, + 314 + ], + "type": "text", + "content": " is chosen adaptively based on clients' computation capability;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "spans": [ + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "text", + "content": "Uploading loss differences. (Line 13, all clients " + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "text", + "content": " server) Each client uploads " + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "text", + "content": " noisy outputs " + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "inline_equation", + "content": "\\{\\Delta \\mathcal{L}(\\mathbf{W}_t,\\pmb {\\delta}_k;\\mathbb{D}_c) + \\frac{N}{N_c}\\pmb {\\epsilon}_c\\}_{k = 1}^K" + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "text", + "content": " to the server, where each output is a floating-point number and the noise " + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{c}" + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "text", + "content": " is negotiated by all clients to be zero-sum. The total uploaded Bytes is " + }, + { + "bbox": [ + 30, + 314, + 379, + 362 + ], + "type": "inline_equation", + "content": "4\\times K" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 362, + 379, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 362, + 379, + 434 + ], + "spans": [ + { + "bbox": [ + 30, + 362, + 379, + 434 + ], + "type": "text", + "content": "Secure aggregation. (Lines 15~16, done by server) In order to prevent the server from recovering the exact loss differences and causing privacy leakage [25], we adopt the secure aggregation method [10] that was originally proposed for conventional FL and apply it to BAFFLE. Specifically, all clients negotiate a group of noises " + }, + { + "bbox": [ + 30, + 362, + 379, + 434 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\epsilon}_c\\}_{c=1}^C" + }, + { + "bbox": [ + 30, + 362, + 379, + 434 + ], + "type": "text", + "content": " satisfying " + }, + { + "bbox": [ + 30, + 362, + 379, + 434 + ], + "type": "inline_equation", + "content": "\\sum_{c=1}^{C} \\pmb{\\epsilon}_c = 0" + }, + { + "bbox": [ + 30, + 362, + 379, + 434 + ], + "type": "text", + "content": ". Then we can reorganize our gradient estimator as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 35, + 443, + 376, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 443, + 376, + 474 + ], + "spans": [ + { + "bbox": [ + 35, + 443, + 376, + 474 + ], + "type": "interline_equation", + "content": "\\widehat {\\nabla_ {\\mathbf {W} _ {t}}} \\mathcal {L} (\\mathbf {W} _ {t}) = \\frac {1}{K} \\sum_ {c = 1} ^ {C} \\frac {N _ {c}}{N} \\sum_ {k = 1} ^ {K} \\left[ \\frac {\\boldsymbol {\\delta} _ {k}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}; \\mathbb {D} _ {c}) \\right] = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {\\boldsymbol {\\delta} _ {k}}{2 \\sigma^ {2}} \\Delta \\mathcal {L} (\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}),", + "image_path": "a21025e6840ee6cb41c2bdc6097e32b824c328d588bf25a2e326fd45c861347e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 34, + 478, + 379, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 34, + 478, + 379, + 520 + ], + "spans": [ + { + "bbox": [ + 34, + 478, + 379, + 520 + ], + "type": "interline_equation", + "content": "\\Delta \\mathcal {L} \\left(\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}\\right) = \\sum_ {c = 1} ^ {C} \\frac {N _ {c}}{N} \\left[ \\Delta \\mathcal {L} \\left(\\mathbf {W} _ {t}, \\boldsymbol {\\delta} _ {k}; \\mathbb {D} _ {c}\\right) + \\frac {N}{N _ {c}} \\boldsymbol {\\epsilon} _ {c} \\right]. \\tag {7}", + "image_path": "df3a7fbf389176ff3e87d5fba7e43d89d80f9c182930a7b5fcf18b62971de207.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "spans": [ + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\epsilon}_c\\}_{c=1}^C" + }, + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "text", + "content": " are zero-sum, there is " + }, + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k) = \\sum_{c=1}^C \\frac{N_c}{N} \\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k; \\mathbb{D}_c)" + }, + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "text", + "content": " and Eq. (7) holds. Therefore, the server can correctly aggregate " + }, + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k)" + }, + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "text", + "content": " and protect client privacy against recovering " + }, + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W}_t, \\pmb{\\delta}_k; \\mathbb{D}_c)" + }, + { + "bbox": [ + 30, + 520, + 379, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 31, + 558, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 558, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 558, + 379, + 584 + ], + "type": "text", + "content": "Remark on communication cost. After getting the gradient estimation " + }, + { + "bbox": [ + 31, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "\\widehat{\\nabla_{\\mathbf{W}_t}}\\mathcal{L}(\\mathbf{W}_t)" + }, + { + "bbox": [ + 31, + 558, + 379, + 584 + ], + "type": "text", + "content": ", the server updates the parameters to " + }, + { + "bbox": [ + 31, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{t + 1}" + }, + { + "bbox": [ + 31, + 558, + 379, + 584 + ], + "type": "text", + "content": " using techniques such" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "text", + "content": "as gradient descent with learning rate " + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "text", + "content": ". Similar to the discussion in McMahan et al. [62], the BAFFLE form presented in Algorithm 1 corresponds to the batch-level communication (also named FedSGD) where Lines " + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "inline_equation", + "content": "11\\sim 12" + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "text", + "content": " execute once for each round " + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "text", + "content": ". In batch-level settings, we reduce the uploaded Bytes from " + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "inline_equation", + "content": "4\\times |\\mathbf{W}|" + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "inline_equation", + "content": "4\\times K" + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "text", + "content": ". We can generalize BAFFLE to an analog of epoch-level communication (also named FedAvg), in which each client updates its local parameters multiple steps using the gradient estimator " + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "inline_equation", + "content": "\\widehat{\\nabla_{\\mathbf{W}_t}}\\mathcal{L}(\\mathbf{W}_t,\\mathbb{D}_c)" + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "text", + "content": " derived from " + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W}_t,\\delta_k;\\mathbb{D}_c)" + }, + { + "bbox": [ + 30, + 34, + 381, + 169 + ], + "type": "text", + "content": " via Eq. (6), and upload model updates to the server after several local epochs. In epoch-level settings, the uploaded Bytes are the same as FedAvg. In experiments, we analyze both batch-level and epoch-level settings for BAFFLE and report the results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 182, + 172, + 195 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 182, + 172, + 195 + ], + "spans": [ + { + "bbox": [ + 31, + 182, + 172, + 195 + ], + "type": "text", + "content": "3.3 Convergence Analyses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 199, + 384, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 199, + 384, + 250 + ], + "spans": [ + { + "bbox": [ + 30, + 199, + 384, + 250 + ], + "type": "text", + "content": "Now we analyze the convergence rate of our gradient estimation method. For continuously differentiable loss functions, we have " + }, + { + "bbox": [ + 30, + 199, + 384, + 250 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D}) = \\lim_{\\sigma \\to 0}\\nabla_{\\mathbf{W}}\\mathcal{L}_{\\sigma}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 30, + 199, + 384, + 250 + ], + "type": "text", + "content": " so we choose a relatively small value for " + }, + { + "bbox": [ + 30, + 199, + 384, + 250 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 30, + 199, + 384, + 250 + ], + "type": "text", + "content": ". The convergence guarantee can be derived as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "spans": [ + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "text", + "content": "Theorem 1. (Proof in Appendix A) Suppose " + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "text", + "content": " is a small value and the central difference scheme in Eq. (2) holds. For perturbations " + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\delta}_k\\}_{k=1}^K \\stackrel{\\mathrm{iid}}{\\sim} \\mathcal{N}(0, \\sigma^2 \\mathbf{I})" + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "text", + "content": ", the empirical covariance matrix is " + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{\\Sigma}} := \\frac{1}{K \\sigma^2} \\sum_{k=1}^K \\pmb{\\delta}_k \\pmb{\\delta}_k^T" + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "text", + "content": " and mean is " + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{\\delta}} := \\frac{1}{K} \\sum_{k=1}^K \\pmb{\\delta}_k" + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "text", + "content": ". Then for any " + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "text", + "content": ", the relation between " + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "inline_equation", + "content": "\\widehat{\\nabla_{\\mathbf{W}}} \\mathcal{L}(\\mathbf{W}; \\mathbb{D})" + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "text", + "content": " and the true gradient " + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{W}} \\mathcal{L}(\\mathbf{W}; \\mathbb{D})" + }, + { + "bbox": [ + 31, + 254, + 381, + 322 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 327, + 379, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 327, + 379, + 342 + ], + "spans": [ + { + "bbox": [ + 120, + 327, + 379, + 342 + ], + "type": "interline_equation", + "content": "\\widehat {\\nabla_ {\\mathbf {W}}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) = \\widehat {\\boldsymbol {\\Sigma}} \\nabla_ {\\mathbf {W}} \\mathcal {L} (\\mathbf {W}; \\mathbb {D}) + o (\\widehat {\\boldsymbol {\\delta}}), \\tag {8}", + "image_path": "624f757fcc9f916cdff1acfe9d5fc3dee0b124eca45c6d7d24147749b596f4ef.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 346, + 147, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 346, + 147, + 362 + ], + "spans": [ + { + "bbox": [ + 32, + 346, + 147, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 32, + 346, + 147, + 362 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\widehat{\\pmb{\\Sigma}}] = \\mathbf{I}" + }, + { + "bbox": [ + 32, + 346, + 147, + 362 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 32, + 346, + 147, + 362 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\widehat{\\boldsymbol{\\delta}}] = \\mathbf{0}" + }, + { + "bbox": [ + 32, + 346, + 147, + 362 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 365, + 379, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 365, + 379, + 402 + ], + "spans": [ + { + "bbox": [ + 31, + 365, + 379, + 402 + ], + "type": "text", + "content": "Taking the expectation on both sides of Eq. (8), we obtain " + }, + { + "bbox": [ + 31, + 365, + 379, + 402 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})] = \\nabla_{\\mathbf{W}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 31, + 365, + 379, + 402 + ], + "type": "text", + "content": ", which degrades to Stein's identity. To determine the convergence rate w.r.t. the value of " + }, + { + "bbox": [ + 31, + 365, + 379, + 402 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 31, + 365, + 379, + 402 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 31, + 405, + 379, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 405, + 379, + 444 + ], + "spans": [ + { + "bbox": [ + 31, + 405, + 379, + 444 + ], + "type": "text", + "content": "Theorem 2. (Adamczak et al. [3]) With overwhelming probability, the empirical covariance matrix satisfies the inequality " + }, + { + "bbox": [ + 31, + 405, + 379, + 444 + ], + "type": "inline_equation", + "content": "\\| \\widehat{\\mathbf{\\Sigma}} -\\mathbf{I}\\| _2\\leq C_0\\sqrt{\\frac{n}{K}}" + }, + { + "bbox": [ + 31, + 405, + 379, + 444 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 31, + 405, + 379, + 444 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 31, + 405, + 379, + 444 + ], + "type": "text", + "content": " denotes the 2-norm for matrix and " + }, + { + "bbox": [ + 31, + 405, + 379, + 444 + ], + "type": "inline_equation", + "content": "C_0" + }, + { + "bbox": [ + 31, + 405, + 379, + 444 + ], + "type": "text", + "content": " is an absolute positive constant." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "spans": [ + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "text", + "content": "Note that in the finetuning setting, " + }, + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "text", + "content": " represents the number of trainable parameters, excluding frozen parameters. As concluded, " + }, + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "inline_equation", + "content": "\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "text", + "content": " provides an unbiased estimation for the true gradients with convergence rate of " + }, + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\left(\\sqrt{\\frac{n}{K}}\\right)" + }, + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "text", + "content": ". Empirically, " + }, + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "inline_equation", + "content": "\\widehat{\\nabla_{\\mathbf{W}}}\\mathcal{L}(\\mathbf{W};\\mathbb{D})" + }, + { + "bbox": [ + 30, + 447, + 379, + 511 + ], + "type": "text", + "content": " is used as a noisy gradient to train models, the generalization of which has been analyzed in previous work [50, 103]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 525, + 128, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 525, + 128, + 539 + ], + "spans": [ + { + "bbox": [ + 31, + 525, + 128, + 539 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 547, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 547, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 547, + 379, + 584 + ], + "type": "text", + "content": "We evaluate BAFFLE on 4 benchmark datasets: MNIST [49], CIFAR-10/100 [48] and OfficeHome [87]. We consider three models: 1) LeNet [49] with two convolutional layers as the shallow model (" + }, + { + "bbox": [ + 30, + 547, + 379, + 584 + ], + "type": "inline_equation", + "content": "2.7 \\times 10^{4}" + }, + { + "bbox": [ + 30, + 547, + 379, + 584 + ], + "type": "text", + "content": " parameters); 2) WideResNet [98]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "spans": [ + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "text", + "content": "Table 1: The classification accuracy (\\%) of BAFFLE in iid scenarios (" + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "inline_equation", + "content": "C = 10" + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "text", + "content": ") and epoch-level communication settings with different " + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "text", + "content": " values (" + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "inline_equation", + "content": "K_{1} / K_{2}" + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "text", + "content": " annotations mean using " + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "inline_equation", + "content": "K_{1}" + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "text", + "content": " for MNIST and " + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "inline_equation", + "content": "K_{2}" + }, + { + "bbox": [ + 30, + 32, + 381, + 111 + ], + "type": "text", + "content": " for CIFAR-10/100). In this configuration, each client updates its local model based on BAFFLE estimated gradients and uploads model updates to the server after an entire epoch on the local dataset. The four guidelines work well under epoch-level settings with total communication rounds 20/40 for MNIST and CIFAR-10/100." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 32, + 120, + 381, + 253 + ], + "blocks": [ + { + "bbox": [ + 32, + 120, + 381, + 253 + ], + "lines": [ + { + "bbox": [ + 32, + 120, + 381, + 253 + ], + "spans": [ + { + "bbox": [ + 32, + 120, + 381, + 253 + ], + "type": "table", + "html": "
SettingsLeNetWRN-10-2
MNISTCIFAR-10CIFAR-100MNISTCIFAR-10CIFAR-100
K100/20087.2748.7841.5488.3552.2746.61
200/50089.4851.8245.6889.5755.5951.65
500/100092.1853.6248.7295.1758.6353.15
Ablation Study (100/200)w/o EMA85.0647.9736.8185.8950.0145.86
ReLU81.5544.9939.4979.0849.7644.44
SELU86.1848.6537.3476.4443.3741.79
Central76.0245.9736.5377.4542.8939.62
BP Baselines94.3158.7554.6797.1162.2960.08
", + "image_path": "272dd137ddef27dfa351b0f255b725dbe4aad0f48d579d99e59ba8e12453bfb4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "spans": [ + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "text", + "content": "Table 2: The accuracy " + }, + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "text", + "content": " of BAFFLE in label non-iid scenarios " + }, + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "inline_equation", + "content": "(C = 100)" + }, + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "text", + "content": " and epoch-level settings with total communication rounds 40 and different " + }, + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "text", + "content": " values. We employ Dirichlet dist. with " + }, + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "inline_equation", + "content": "\\alpha = 0.3" + }, + { + "bbox": [ + 30, + 268, + 381, + 303 + ], + "type": "text", + "content": " to ensure each client has a unique label distribution." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 311, + 362, + 396 + ], + "blocks": [ + { + "bbox": [ + 50, + 311, + 362, + 396 + ], + "lines": [ + { + "bbox": [ + 50, + 311, + 362, + 396 + ], + "spans": [ + { + "bbox": [ + 50, + 311, + 362, + 396 + ], + "type": "table", + "html": "
SettingsLeNet CIFAR-10 | CIFAR-100WRN-10-2 CIFAR-10 | CIFAR-100
K20035.21 | 28.1239.53 | 30.44
50038.14 | 30.9241.69 | 32.89
100039.71 | 33.3543.42 | 34.08
BP Baselines44.41 | 38.4351.18 | 40.85
", + "image_path": "255d089263d364a516fe2eb7029b8761782ba890565c9e78716249d99b44bd95.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "text", + "content": "with depth " + }, + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "inline_equation", + "content": "= 10" + }, + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "inline_equation", + "content": "= 2" + }, + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "text", + "content": " (WRN-10-2) as the light weight deep model " + }, + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "inline_equation", + "content": "(3.0 \\times 10^{5}" + }, + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "text", + "content": " parameters) and 3) MobileNet [38] as the deep neural networks " + }, + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "inline_equation", + "content": "(1.3 \\times 10^{7}" + }, + { + "bbox": [ + 31, + 413, + 379, + 452 + ], + "type": "text", + "content": " parameters) that works on ImageNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "text", + "content": "Participation and communication settings. To perform a comprehensive evaluation of BAFFLE, we simulate three popular FL scenarios [17] with the FedLab [99] participations: iid participations, label non-iid participations and feature non-iid participations. For iid participations, we set the client number " + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "inline_equation", + "content": "C = 10" + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "text", + "content": " and use uniform distribution to build local datasets. Then we evaluate our BAFFLE on MNIST and CIFAR-10/100 under both batch-level (FedSGD) and epoch-level (FedAvg) communication settings. For label non-iid participations, we set client number " + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "inline_equation", + "content": "C = 100" + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "text", + "content": ", use Dirichlet distribution with " + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\alpha = 0.3" + }, + { + "bbox": [ + 30, + 452, + 381, + 584 + ], + "type": "text", + "content": " to build clients. For feature non-iid participations, we build clients from the prevailing domain adaptation dataset OfficeHome, which contains 65 categories from 4 different domains, i.e. Art, Clipart, Product and Real-world. We set the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 37, + 208, + 140 + ], + "blocks": [ + { + "bbox": [ + 33, + 37, + 208, + 140 + ], + "lines": [ + { + "bbox": [ + 33, + 37, + 208, + 140 + ], + "spans": [ + { + "bbox": [ + 33, + 37, + 208, + 140 + ], + "type": "image", + "image_path": "a1ce11cb3cbf01404dd6cb86b7f30597656d75f14cfd81cdecf3519eff06f653.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 214, + 36, + 376, + 140 + ], + "blocks": [ + { + "bbox": [ + 214, + 36, + 376, + 140 + ], + "lines": [ + { + "bbox": [ + 214, + 36, + 376, + 140 + ], + "spans": [ + { + "bbox": [ + 214, + 36, + 376, + 140 + ], + "type": "image", + "image_path": "8da81ede411fd39381a7acb952ac0bf7110674536c8ea0ba8fe804dc5802ba59.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 32, + 145, + 211, + 248 + ], + "blocks": [ + { + "bbox": [ + 32, + 145, + 211, + 248 + ], + "lines": [ + { + "bbox": [ + 32, + 145, + 211, + 248 + ], + "spans": [ + { + "bbox": [ + 32, + 145, + 211, + 248 + ], + "type": "image", + "image_path": "3fe921cd0a5dade0c69db3f6c4176ee080b9b34a98e2fbb6056d6beac6680870.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 250, + 149, + 258 + ], + "lines": [ + { + "bbox": [ + 127, + 250, + 149, + 258 + ], + "spans": [ + { + "bbox": [ + 127, + 250, + 149, + 258 + ], + "type": "text", + "content": "LeNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 213, + 145, + 376, + 248 + ], + "blocks": [ + { + "bbox": [ + 213, + 145, + 376, + 248 + ], + "lines": [ + { + "bbox": [ + 213, + 145, + 376, + 248 + ], + "spans": [ + { + "bbox": [ + 213, + 145, + 376, + 248 + ], + "type": "image", + "image_path": "2568a114b196847cb3766fa2fbf239c6d630b1572d1e8cf6d0cd4dac877203f8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 250, + 323, + 258 + ], + "lines": [ + { + "bbox": [ + 286, + 250, + 323, + 258 + ], + "spans": [ + { + "bbox": [ + 286, + 250, + 323, + 258 + ], + "type": "text", + "content": "WRN-10-2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 30, + 267, + 380, + 323 + ], + "lines": [ + { + "bbox": [ + 30, + 267, + 380, + 323 + ], + "spans": [ + { + "bbox": [ + 30, + 267, + 380, + 323 + ], + "type": "text", + "content": "Fig. 2: The classification accuracy (\\%) of BAFFLE in iid scenarios (" + }, + { + "bbox": [ + 30, + 267, + 380, + 323 + ], + "type": "inline_equation", + "content": "C = 10" + }, + { + "bbox": [ + 30, + 267, + 380, + 323 + ], + "type": "text", + "content": ") and batch-level communication settings with various " + }, + { + "bbox": [ + 30, + 267, + 380, + 323 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 267, + 380, + 323 + ], + "type": "text", + "content": " values. We treat the models trained by exact gradients on conventional FL systems as the backpropagation (BP) baselines. On different datasets and architectures, our BAFFLE achieves comparable performance to the exact gradient results with a reasonable " + }, + { + "bbox": [ + 30, + 267, + 380, + 323 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 267, + 380, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 344, + 380, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 344, + 380, + 380 + ], + "spans": [ + { + "bbox": [ + 30, + 344, + 380, + 380 + ], + "type": "text", + "content": "total client number to " + }, + { + "bbox": [ + 30, + 344, + 380, + 380 + ], + "type": "inline_equation", + "content": "C = 40" + }, + { + "bbox": [ + 30, + 344, + 380, + 380 + ], + "type": "text", + "content": " and generate 10 clients from each domain. As results, we report Top-1 accuracy for MNIST, CIFAR-10 and OfficeHome and Top-5 accuracy for OfficeHome and CIFAR-100." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "text", + "content": "Hyperparameters. Following the settings in Section 2, we use FedAVG to aggregate gradients from multiple clients and use SGD-based optimizer to update global parameters. Specifically, we use Adam [46] to train a random initialized model with " + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\beta = (0.9,0.99)" + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "text", + "content": ", learning rate 0.01 and epochs 20/40 for MNIST and CIFAR-10/100. For OfficeHome, we adapt the transfer learning [40] by loading the ImageNet-pretrained model and finetuning the final layers with Adam, but setting learning rate 0.005 and epochs 40. In BAFFLE, the perturbation scale " + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "text", + "content": " and number " + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "text", + "content": " are the most important hyperparameters. As shown in Theorem 1, with less noise and more samples, the BAFFLE will obtain more accurate gradients, leading to improved performance. However, there exists a trade-off between accuracy and computational efficiency: an extremely small " + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "text", + "content": " will cause the underflow problem [27] and a large " + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "text", + "content": " will increase computational cost. In practice, we empirically set " + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\sigma = 10^{-4}" + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "text", + "content": " because it is the smallest value that does not cause numerical problems in all experiments, and works well on edge devices with half-precision floating-point numbers. We also evaluate the impact of " + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 392, + 381, + 584 + ], + "type": "text", + "content": " across a broad range from 100 to 5000." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 347, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 207, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 207, + 46 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 207, + 46 + ], + "type": "text", + "content": "4.1 Four Guidelines for BAFFLE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 57, + 380, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 57, + 380, + 153 + ], + "spans": [ + { + "bbox": [ + 32, + 57, + 380, + 153 + ], + "type": "text", + "content": "For a general family of continuously differentiable models, we analyze their convergence rate of BAFFLE in Section 3.3. Since deep networks are usually stacked with multiple linear layers and non-linear activation, this layer linearity can be utilized to improve the accuracy-efficiency trade-off. Combining the linearity property and the unique conditions in edge devices (e.g., small data size and half-precision format), we present four guidelines for model design and training that can increase accuracy without introducing extra computation (Appendix C shows the details of linearity analysis):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 154, + 379, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 154, + 379, + 249 + ], + "spans": [ + { + "bbox": [ + 32, + 154, + 379, + 249 + ], + "type": "text", + "content": "Using twice forward difference (twice-FD) scheme rather than central scheme. Combining difference scheme Eq. (1) and Eq. (2), we find that by executing twice as many forward inferences (i.e. " + }, + { + "bbox": [ + 32, + 154, + 379, + 249 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\pm \\boldsymbol{\\delta}" + }, + { + "bbox": [ + 32, + 154, + 379, + 249 + ], + "type": "text", + "content": "), the central scheme achieves lower residuals than twice-FD, despite the fact that twice-FD can benefit from additional sample times. With the same forward times (e.g., " + }, + { + "bbox": [ + 32, + 154, + 379, + 249 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 32, + 154, + 379, + 249 + ], + "type": "text", + "content": "), determining which scheme performs better is a practical issue. As shown in Appendix C, we find that twice-FD performs better in all experiments, in part because the linearity reduces the benefit from second-order residuals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 250, + 379, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 250, + 379, + 321 + ], + "spans": [ + { + "bbox": [ + 32, + 250, + 379, + 321 + ], + "type": "text", + "content": "Using Hardswish in BAFFLE. ReLU is effective when the middle features " + }, + { + "bbox": [ + 32, + 250, + 379, + 321 + ], + "type": "inline_equation", + "content": "(h(\\cdot)" + }, + { + "bbox": [ + 32, + 250, + 379, + 321 + ], + "type": "text", + "content": " denotes the feature mapping) have the same sign before and after perturbations, i.e. " + }, + { + "bbox": [ + 32, + 250, + 379, + 321 + ], + "type": "inline_equation", + "content": "h(\\mathbf{W} + \\boldsymbol{\\delta})\\cdot h(\\mathbf{W}) > 0" + }, + { + "bbox": [ + 32, + 250, + 379, + 321 + ], + "type": "text", + "content": ". Since ReLU is not differentiable at zero, the value jump occurs when the sign of features changes after perturbations, i.e. " + }, + { + "bbox": [ + 32, + 250, + 379, + 321 + ], + "type": "inline_equation", + "content": "h(\\mathbf{W} + \\boldsymbol{\\delta})\\cdot h(\\mathbf{W}) < 0" + }, + { + "bbox": [ + 32, + 250, + 379, + 321 + ], + "type": "text", + "content": ". We use Hardswish [37] to overcome this problem as it is continuously differentiable at zero and easy to implement on edge devices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 322, + 379, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 322, + 379, + 383 + ], + "spans": [ + { + "bbox": [ + 32, + 322, + 379, + 383 + ], + "type": "text", + "content": "Using exponential moving average (EMA) to reduce oscillations. As shown in Theorem 1, there exists an zero-mean white-noise " + }, + { + "bbox": [ + 32, + 322, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\widehat{\\delta}" + }, + { + "bbox": [ + 32, + 322, + 379, + 383 + ], + "type": "text", + "content": " between the true gradient and our estimation. To smooth out the oscillations caused by white noise, we apply EMA strategies from BYOL [28] to the global parameters, with a smoothing coefficient of 0.995." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 384, + 379, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 384, + 379, + 432 + ], + "spans": [ + { + "bbox": [ + 32, + 384, + 379, + 432 + ], + "type": "text", + "content": "Using GroupNorm as opposed to BatchNorm. On edge devices, the dataset size is typically small, which leads to inaccurate batch statistics estimation and degrades performance when using BatchNorm. Thus we employ GroupNorm [94] to solve this issue." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 452, + 199, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 452, + 199, + 464 + ], + "spans": [ + { + "bbox": [ + 32, + 452, + 199, + 464 + ], + "type": "text", + "content": "4.2 Performance on IID Clients" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "text", + "content": "Following the settings in Section 4.1, we evaluate the performance of BAFFLE in the iid scenarios. We reproduce all experiments on the BP-based FL systems with the same settings and use them as the baselines. We refer to the baseline results as exact gradients and report the training process of BAFFLE in Figure 2. The value of " + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "text", + "content": " (e.g., 200 for LeNet and 500 for WRN-10-2) is much less than the dimensions of parameter space (e.g., " + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "inline_equation", + "content": "2.7 \\times 10^4" + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "text", + "content": " for LeNet and " + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "inline_equation", + "content": "3 \\times 10^5" + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "text", + "content": " for WRN-10-2). Since the convergence rate to the exact gradient is " + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\left(\\sqrt{\\frac{n}{K}}\\right)" + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "text", + "content": ", the marginal benefit of increasing " + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "text", + "content": " decreases. For instance, increasing " + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "text", + "content": " from 2000 to 5000 on CIFAR-10 with WRN-10-2 barely improves accuracy by " + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 32, + 475, + 380, + 583 + ], + "type": "text", + "content": ". Given that the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 32, + 76, + 374, + 182 + ], + "blocks": [ + { + "bbox": [ + 31, + 32, + 380, + 67 + ], + "lines": [ + { + "bbox": [ + 31, + 32, + 380, + 67 + ], + "spans": [ + { + "bbox": [ + 31, + 32, + 380, + 67 + ], + "type": "text", + "content": "Table 3: The Top-1|Top-5 accuracy (\\%) of BAFFLE on OfficeHome with feature non-iid participations (" + }, + { + "bbox": [ + 31, + 32, + 380, + 67 + ], + "type": "inline_equation", + "content": "C = 40" + }, + { + "bbox": [ + 31, + 32, + 380, + 67 + ], + "type": "text", + "content": ") and epoch-level settings with 40 comm. rounds. We use the pretrained MobileNet, freeze the backbone and finetune the FC layers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 32, + 76, + 374, + 182 + ], + "lines": [ + { + "bbox": [ + 32, + 76, + 374, + 182 + ], + "spans": [ + { + "bbox": [ + 32, + 76, + 374, + 182 + ], + "type": "table", + "html": "
SettingsDomainsAvg.
ArtClipartProductReal World
2044.75|69.4652.48|73.8866.63|89.7763.78|87.8356.91|80.24
5047.87|71.3253.43|76.8371.28|91.7467.02|89.9559.90|82.46
K 10050.32|74.4257.43|80.7374.19|93.0269.53|90.4362.87|84.65
20051.42|76.6460.98|86.4176.05|94.4271.51|93.1464.98|87.65
50053.33|77.8562.58|86.6478.84|95.2373.17|93.8566.98|88.40
BP Baselines55.71|80.4365.13|88.6582.44|96.0577.19|95.0470.12|90.04
", + "image_path": "ca68b037c69f060f586106a9393a3e95e4fec742e47881cf4b49cbd035d3419c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 189, + 379, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 189, + 379, + 213 + ], + "spans": [ + { + "bbox": [ + 30, + 189, + 379, + 213 + ], + "type": "text", + "content": "convergence rate of Gaussian perturbations is " + }, + { + "bbox": [ + 30, + 189, + 379, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\left(\\sqrt{\\frac{n}{K}}\\right)" + }, + { + "bbox": [ + 30, + 189, + 379, + 213 + ], + "type": "text", + "content": ", the sampling efficiency may be improved by choosing an alternative distribution for perturbations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 214, + 380, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 214, + 380, + 321 + ], + "spans": [ + { + "bbox": [ + 30, + 214, + 380, + 321 + ], + "type": "text", + "content": "Ablation studies. As depicted in Figure 3, we conduct ablation studies for BAFFLE to evaluate the aforementioned guidelines. In general, twice-FD, Hardswish and EMA can all improve the accuracy. For two difference schemes, we compare the twice-FD to central scheme with the same computation cost and show that the former outperforms the later, demonstrating that linearity reduces the gain from second-order residuals. As to activation functions, Hardswish is superior to ReLU and SELU because it is differentiable at zero and vanishes to zero in the negative part. Moreover, EMA enhances the performance of training strategies by reducing the effect of white noise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 321, + 379, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 321, + 379, + 525 + ], + "spans": [ + { + "bbox": [ + 30, + 321, + 379, + 525 + ], + "type": "text", + "content": "Communication efficiency. Compared to the batch-level communication settings (FedSGD) in a BP-based FL system, BAFFLE requires each client to upload a " + }, + { + "bbox": [ + 30, + 321, + 379, + 525 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 321, + 379, + 525 + ], + "type": "text", + "content": "-dimensional vector to the server and downloads the updated global parameters in each communication round. Since " + }, + { + "bbox": [ + 30, + 321, + 379, + 525 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 321, + 379, + 525 + ], + "type": "text", + "content": " is significantly less than the parameter amounts (e.g., 500 versus 0.3 million), BAFFLE reduces data transfer by approximately half. To reduce communication costs, the prevalent FL system requires each client to perform model optimization on the local training dataset and upload the model updates to the server after a specified number of local epochs. BAFFLE can also perform epoch-level communications by employing an " + }, + { + "bbox": [ + 30, + 321, + 379, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n)" + }, + { + "bbox": [ + 30, + 321, + 379, + 525 + ], + "type": "text", + "content": " additional memory to store the perturbation in each forward and estimate the local gradient using Eq. (6). Then each client optimizes the local model with SGD and uploads local updates after several epochs. As shown in Table 1, we evaluate the performance of BAFFLE under one-epoch communication settings. As epoch-level communication is more prevalent in the real-world FL, all the following experiments will be conducted in this context. In brief, BAFFLE uploads the same Bytes as BP-based FL in epoch-level communication while the total communication rounds are much less than FedZO [22], e.g. 20 versus 1000 on MNIST." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 541, + 224, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 541, + 224, + 552 + ], + "spans": [ + { + "bbox": [ + 31, + 541, + 224, + 552 + ], + "type": "text", + "content": "4.3 Performance on Non-IID Clients" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "text", + "content": "Following Section 4.1, we evaluate the performance of BAFFLE in both label non-iid and feature non-iid scenarios. For label non-iid scenarios, we use the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 37, + 210, + 140 + ], + "blocks": [ + { + "bbox": [ + 33, + 37, + 210, + 140 + ], + "lines": [ + { + "bbox": [ + 33, + 37, + 210, + 140 + ], + "spans": [ + { + "bbox": [ + 33, + 37, + 210, + 140 + ], + "type": "image", + "image_path": "dd059e2c8e099f1a1b5c7263d248d88bb6a848b6c87d126b74995e61d751f6ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 214, + 36, + 376, + 140 + ], + "blocks": [ + { + "bbox": [ + 214, + 36, + 376, + 140 + ], + "lines": [ + { + "bbox": [ + 214, + 36, + 376, + 140 + ], + "spans": [ + { + "bbox": [ + 214, + 36, + 376, + 140 + ], + "type": "image", + "image_path": "25c4bb93dfc6f715f6f970d6fbf027483f6ccd621d92bdbb2ff1c79d5515921f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 32, + 145, + 210, + 248 + ], + "blocks": [ + { + "bbox": [ + 32, + 145, + 210, + 248 + ], + "lines": [ + { + "bbox": [ + 32, + 145, + 210, + 248 + ], + "spans": [ + { + "bbox": [ + 32, + 145, + 210, + 248 + ], + "type": "image", + "image_path": "23ebf4cff94ea2c9dd73d7d5bb532ed93041385e3b494647897946d91f7d3668.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 250, + 149, + 258 + ], + "lines": [ + { + "bbox": [ + 127, + 250, + 149, + 258 + ], + "spans": [ + { + "bbox": [ + 127, + 250, + 149, + 258 + ], + "type": "text", + "content": "LeNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 213, + 145, + 376, + 248 + ], + "blocks": [ + { + "bbox": [ + 213, + 145, + 376, + 248 + ], + "lines": [ + { + "bbox": [ + 213, + 145, + 376, + 248 + ], + "spans": [ + { + "bbox": [ + 213, + 145, + 376, + 248 + ], + "type": "image", + "image_path": "6dfb402337254ea1a4544638c5f36452f0047e51115a9cdce482f0671e6fc905.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 250, + 323, + 258 + ], + "lines": [ + { + "bbox": [ + 286, + 250, + 323, + 258 + ], + "spans": [ + { + "bbox": [ + 286, + 250, + 323, + 258 + ], + "type": "text", + "content": "WRN-10-2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 30, + 267, + 380, + 302 + ], + "lines": [ + { + "bbox": [ + 30, + 267, + 380, + 302 + ], + "spans": [ + { + "bbox": [ + 30, + 267, + 380, + 302 + ], + "type": "text", + "content": "Fig. 3: The ablation study of BAFFLE guidelines, with " + }, + { + "bbox": [ + 30, + 267, + 380, + 302 + ], + "type": "inline_equation", + "content": "K = 100" + }, + { + "bbox": [ + 30, + 267, + 380, + 302 + ], + "type": "text", + "content": " on MNIST and " + }, + { + "bbox": [ + 30, + 267, + 380, + 302 + ], + "type": "inline_equation", + "content": "K = 500" + }, + { + "bbox": [ + 30, + 267, + 380, + 302 + ], + "type": "text", + "content": " on CIFAR-10. As seen, twice-FD, Hardswish, and EMA all improve performance without extra computation. EMA reduces oscillations by lessening Gaussian noise." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 322, + 380, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 322, + 380, + 431 + ], + "spans": [ + { + "bbox": [ + 30, + 322, + 380, + 431 + ], + "type": "text", + "content": "CIFAR-10/100 datasets and employ Dirichlet distribution to ensure that each client has a unique label distribution. We evaluate the performance of BAFFLE with 100 clients and various K values. As seen in Table 2, the model suffers a significant drop in accuracy (e.g., " + }, + { + "bbox": [ + 30, + 322, + 380, + 431 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 30, + 322, + 380, + 431 + ], + "type": "text", + "content": " in CIFAR-10 and " + }, + { + "bbox": [ + 30, + 322, + 380, + 431 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 30, + 322, + 380, + 431 + ], + "type": "text", + "content": " in CIFAR-100) due to the label non-iid effect. For feature non-iid scenarios, we construct clients using the OfficeHome dataset and use MobileNet as the deep model. As seen in Table 3, we use the transfer learning strategy to train MobileNet, i.e., we load the parameters pretrained on ImageNet, freeze the backbone parameters, and retrain the classification layers. The accuracy decrease is approximately " + }, + { + "bbox": [ + 30, + 322, + 380, + 431 + ], + "type": "inline_equation", + "content": "3\\% \\sim 5\\%" + }, + { + "bbox": [ + 30, + 322, + 380, + 431 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 31, + 446, + 309, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 446, + 309, + 458 + ], + "spans": [ + { + "bbox": [ + 31, + 446, + 309, + 458 + ], + "type": "text", + "content": "4.4 Computation Efficiency, Memory and Robustness" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "spans": [ + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "text", + "content": "BAFFLE uses " + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "text", + "content": " times forward passes instead of backward. Since the backward pass is about as expensive as two normal forward passes [35] and five single-precision accelerated forward passes [67], BAFFLE results in approximately " + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "inline_equation", + "content": "\\frac{K}{5}" + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "text", + "content": " times the computation expense of BP-based FL. Although BAFFLE results in " + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "inline_equation", + "content": "\\frac{K}{5} - 1" + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "text", + "content": " times extra computation cost, we show the cost can be reduced with proper training strategies, e.g., the transfer learning in Table 3 can reduce " + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "text", + "content": " to 20 on the MobileNet and the " + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 30, + 464, + 379, + 546 + ], + "type": "text", + "content": " sized OfficeHome dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 547, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 547, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 547, + 380, + 584 + ], + "type": "text", + "content": "Moreover, BAFFLE can reduce huge memory cost on edge devices with the efficiency in static memory and dynamic memory. The auto-differential framework is used to run BP on deep networks, which requires extra static memory" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 33, + 76, + 378, + 143 + ], + "blocks": [ + { + "bbox": [ + 31, + 32, + 380, + 67 + ], + "lines": [ + { + "bbox": [ + 31, + 32, + 380, + 67 + ], + "spans": [ + { + "bbox": [ + 31, + 32, + 380, + 67 + ], + "type": "text", + "content": "Table 4: The GPU memory cost (MB) of vanilla BP and BAFFLE, respectively. 'min~max' denotes the minimum and maximum dynamic memory for BAFFLE. We also report the ratio (\\%) of vanilla BP to BAFFLE's max memory cost." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 33, + 76, + 378, + 143 + ], + "lines": [ + { + "bbox": [ + 33, + 76, + 378, + 143 + ], + "spans": [ + { + "bbox": [ + 33, + 76, + 378, + 143 + ], + "type": "table", + "html": "
BackboneCIFAR-10/100OfficeHome/ImageNet
BPBAFFLERatioBPBAFFLERatio
LeNet168067~17410.35252786~2017.95
WRN-10-2187875~19610.43342594~2517.32
MobileNet2041102~21710.635271121~2895.48
", + "image_path": "d4f5d60b7a6f870d548d06c72cade95d7970fc350285dc2d7bb2ef97e668cad3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 159, + 380, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 159, + 380, + 267 + ], + "spans": [ + { + "bbox": [ + 30, + 159, + 380, + 267 + ], + "type": "text", + "content": "(e.g., 200MB for Caffe [41] and 1GB for Pytorch [72]) and imposes a considerable burden on edge devices such as IoT sensors. Due to the necessity of restoring intermediate states, BP also requires enormous amounts of dynamic memory (" + }, + { + "bbox": [ + 30, + 159, + 380, + 267 + ], + "type": "inline_equation", + "content": "\\geq" + }, + { + "bbox": [ + 30, + 159, + 380, + 267 + ], + "type": "text", + "content": " 5GB for MobileNet [24]). Since BAFFLE only requires inference, we can slice the computation graph and execute the forwards per layer [44]. As shown in Table 4, BAFFLE reduces the memory cost to " + }, + { + "bbox": [ + 30, + 159, + 380, + 267 + ], + "type": "inline_equation", + "content": "5\\% \\sim 10\\%" + }, + { + "bbox": [ + 30, + 159, + 380, + 267 + ], + "type": "text", + "content": " by executing layer-by-layer inference. By applying kernel-wise computations, we can further reduce the memory cost to approximately " + }, + { + "bbox": [ + 30, + 159, + 380, + 267 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 30, + 159, + 380, + 267 + ], + "type": "text", + "content": " (e.g., 64MB for MobileNet [85]), which is suitable for scenarios with extremely limited storage resources, such as TEE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "spans": [ + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "text", + "content": "Recent works exploit TEE to protect models from white-box attacks by preventing model exposure [44]. However, due to the security guarantee, the usable memory of TEE is usually small [85] (e.g., 90MB on Intel SGX for Skylake CPU [61]), which is typically far less than what a backpropagation-based FL system requires. In contrast, BAFFLE can execute in TEE due to its little memory cost (more details are in Appendix B). Membership inference attacks and model inversion attacks need to repeatedly perform model inference and obtain confidence values or classification scores [78, 101]. Given that BAFFLE provides stochastic loss differences " + }, + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}(\\mathbf{W},\\delta ;\\mathbb{D})" + }, + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "text", + "content": " associated with the random perturbation " + }, + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "text", + "content": ", the off-the-shelf inference attacks may not perform on BAFFLE directly (while adaptively designed attacking strategies are possible to evade BAFFLE). Motivated by differential privacy [1], we further design heuristic experiments to study the information leakage from " + }, + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}" + }, + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "text", + "content": " (details in Appendix D). As shown in Figure 5, the " + }, + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}" + }, + { + "bbox": [ + 30, + 268, + 380, + 448 + ], + "type": "text", + "content": " between real data and random noise is hard to distinguish, indicating it is difficult for attackers to obtain useful information from BAFFLE's outputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 469, + 137, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 469, + 137, + 482 + ], + "spans": [ + { + "bbox": [ + 31, + 469, + 137, + 482 + ], + "type": "text", + "content": "5 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 499, + 380, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 499, + 380, + 559 + ], + "spans": [ + { + "bbox": [ + 30, + 499, + 380, + 559 + ], + "type": "text", + "content": "Along the research routine of FL, many efforts have been devoted to, e.g., dealing with non-IID distributions [21, 53, 75, 90, 102], multi-task learning [60, 79], and preserving privacy of clients [11, 12, 26, 31, 56, 58, 63, 84]. Below we introduce the work on efficiency and vulnerability in FL following the survey of Kairouz et al. [42], which is more related to this paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 559, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 559, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 559, + 379, + 584 + ], + "type": "text", + "content": "Efficiency in FL. It is widely understood that the communication and computational efficiency is a primary bottleneck for deploying FL in practice [6,18,74," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 381, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 381, + 166 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 381, + 166 + ], + "type": "text", + "content": "89, 92]. Specifically, communicating between the server and clients could be potentially expensive and unreliable. The seminal work of Konečný et al. [47] introduces sparsification and quantization to reduce the communication cost, where several theoretical works investigate the optimal trade-off between the communication cost and model accuracy [2, 7, 15, 30, 100]. Since practical clients usually have slower upload than download bandwidth, much research interest focuses on gradient compression [4, 8, 36, 82]. On the other hand, different methods have been proposed to reduce the computational burden of local clients [16, 29, 33], since these clients are usually edge devices with limited resources. Training paradigms exploiting tensor factorization in FL can also achieve promising performance [45, 59]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 168, + 381, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 168, + 381, + 277 + ], + "spans": [ + { + "bbox": [ + 30, + 168, + 381, + 277 + ], + "type": "text", + "content": "Vulnerability in FL. The characteristic of decentralization in FL is beneficial to protecting data privacy of clients, but in the meanwhile, providing white-box accessibility of model status leaves flexibility for malicious clients to perform poisoning/backdoor attacks [5, 9, 71, 88, 95], model/gradient inversion attacks [25, 39, 101], and membership inference attacks [57, 68, 78]. To alleviate the vulnerability in FL, several defense strategies have been proposed via selecting reliable clients [43], data augmentation [13], update clipping [81], robust training [52], model perturbation [97], detection methods [20, 76], and methods based on differential privacy [93]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 305, + 211, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 305, + 211, + 318 + ], + "spans": [ + { + "bbox": [ + 31, + 305, + 211, + 318 + ], + "type": "text", + "content": "6 Conclusion and Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 342, + 381, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 342, + 381, + 426 + ], + "spans": [ + { + "bbox": [ + 30, + 342, + 381, + 426 + ], + "type": "text", + "content": "Backpropagation is the gold standard for training deep networks, and it is also utilized by traditional FL systems. However, backpropagation is unsuited for edge devices due to their limited resources and possible lack of reliability. Using zero-order optimization techniques, we explore the possibility of BAFFLE in this paper. We need to specify that there are scenarios in which clients are fully trusted and have sufficient computing and storage resources. In these situations, traditional FL with backpropagation is preferred." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 428, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 428, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 428, + 381, + 584 + ], + "type": "text", + "content": "While our preliminary studies on BAFFLE have generated encouraging results, there are still a number of tough topics to investigate: (i) Compared to the models trained using exact gradients, the accuracy of models trained using BAFFLE is inferior. One reason is that we select small values of " + }, + { + "bbox": [ + 30, + 428, + 381, + 584 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 428, + 381, + 584 + ], + "type": "text", + "content": " (e.g., 500) relative to the number of model parameters (e.g., " + }, + { + "bbox": [ + 30, + 428, + 381, + 584 + ], + "type": "inline_equation", + "content": "3.0 \\times 10^{5}" + }, + { + "bbox": [ + 30, + 428, + 381, + 584 + ], + "type": "text", + "content": "); another reason is that gradient descent is designed for exact gradients, whereas our noisy gradient estimation may require advanced learning algorithms. (ii) The empirical variance of zero-order gradient estimators affects training convergence in BAFFLE. It is crucial to research variance reduction approaches, such as control variates and non-Gaussian sampling distributions. (iii) Stein's identity is proposed for loss functions with Gaussian noises imposed on model parameters. Intuitively, this smoothness is related to differential privacy in FL, but determining their relationship requires theoretical derivations." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 139, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 139, + 46 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 139, + 46 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 55, + 380, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 55, + 380, + 81 + ], + "spans": [ + { + "bbox": [ + 31, + 55, + 380, + 81 + ], + "type": "text", + "content": "This paper is supported by the National Science Foundation of China (62132017), Zhejiang Provincial Natural Science Foundation of China (LD24F020011)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 96, + 98, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 96, + 98, + 109 + ], + "spans": [ + { + "bbox": [ + 32, + 96, + 98, + 109 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 118, + 380, + 583 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 41, + 118, + 379, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 118, + 379, + 162 + ], + "spans": [ + { + "bbox": [ + 41, + 118, + 379, + 162 + ], + "type": "text", + "content": "1. Abadi, M., Chu, A., Goodfellow, I.J., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, Vienna, Austria, October 24-28, 2016 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 40, + 162, + 380, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 162, + 380, + 194 + ], + "spans": [ + { + "bbox": [ + 40, + 162, + 380, + 194 + ], + "type": "text", + "content": "2. Acharya, J., Canonne, C.L., Tyagi, H.: Inference under information constraints i: Lower bounds from chi-square contraction. IEEE Transactions on Information Theory (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 40, + 194, + 380, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 194, + 380, + 227 + ], + "spans": [ + { + "bbox": [ + 40, + 194, + 380, + 227 + ], + "type": "text", + "content": "3. Adamczak, R., Litvak, A.E., Pajor, A., Tomczak-Jaegermann, N.: Sharp bounds on the rate of convergence of the empirical covariance matrix. Comptes Rendus Mathematique 349(3-4), 195-200 (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 40, + 227, + 380, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 227, + 380, + 259 + ], + "spans": [ + { + "bbox": [ + 40, + 227, + 380, + 259 + ], + "type": "text", + "content": "4. Alistarh, D., Grubic, D., Li, J., Tomioka, R., Vojnovic, M.: Qsgd: Communication-efficient sgd via gradient quantization and encoding. Advances in Neural Information Processing Systems (NeurIPS) (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 40, + 259, + 380, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 259, + 380, + 292 + ], + "spans": [ + { + "bbox": [ + 40, + 259, + 380, + 292 + ], + "type": "text", + "content": "5. Bagdasaryan, E., Veit, A., Hua, Y., Estrin, D., Shmatikov, V.: How to back-door federated learning. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 40, + 292, + 380, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 292, + 380, + 324 + ], + "spans": [ + { + "bbox": [ + 40, + 292, + 380, + 324 + ], + "type": "text", + "content": "6. Balakrishnan, R., Li, T., Zhou, T., Himayat, N., Smith, V., Bilmes, J.: Diverse client selection for federated learning via submodular maximization. In: International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 40, + 324, + 380, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 324, + 380, + 356 + ], + "spans": [ + { + "bbox": [ + 40, + 324, + 380, + 356 + ], + "type": "text", + "content": "7. Barnes, L.P., Han, Y., Özgür, A.: Lower bounds for learning distributions under communication constraints via fisher information. Journal of Machine Learning Research (JMLR) (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 40, + 356, + 380, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 356, + 380, + 388 + ], + "spans": [ + { + "bbox": [ + 40, + 356, + 380, + 388 + ], + "type": "text", + "content": "8. Basu, D., Data, D., Karakus, C., Diggavi, S.: Qsparse-local-sgd: Distributed sgd with quantization, sparsification and local computations. Advances in Neural Information Processing Systems (NeurIPS) (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 40, + 388, + 380, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 388, + 380, + 421 + ], + "spans": [ + { + "bbox": [ + 40, + 388, + 380, + 421 + ], + "type": "text", + "content": "9. Bhagoji, A.N., Chakraborty, S., Mittal, P., Calo, S.: Analyzing federated learning through an adversarial lens. In: International Conference on Machine Learning (ICML) (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 37, + 421, + 379, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 421, + 379, + 464 + ], + "spans": [ + { + "bbox": [ + 37, + 421, + 379, + 464 + ], + "type": "text", + "content": "10. Bonawitz, K.A., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for privacy-preserving machine learning. In: ACM Conference on Computer and Communications Security (CCS) (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 37, + 464, + 379, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 464, + 379, + 496 + ], + "spans": [ + { + "bbox": [ + 37, + 464, + 379, + 496 + ], + "type": "text", + "content": "11. Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for federated learning on user-held data. arXiv preprint arXiv:1611.04482 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 37, + 496, + 379, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 496, + 379, + 539 + ], + "spans": [ + { + "bbox": [ + 37, + 496, + 379, + 539 + ], + "type": "text", + "content": "12. Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K.: Practical secure aggregation for privacy-preserving machine learning. In: ACM SIGSAC Conference on Computer and Communications Security (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 37, + 539, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 539, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 37, + 539, + 379, + 583 + ], + "type": "text", + "content": "13. Borgnia, E., Geiping, J., Cherepanova, V., Fowl, L., Gupta, A., Ghiasi, A., Huang, F., Goldblum, M., Goldstein, T.: Dp-instahide: Provably defusing poisoning and backdoor attacks with differentially private data augmentations. arXiv preprint arXiv:2103.02079 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 36, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 36, + 35, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 35, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 36, + 35, + 379, + 79 + ], + "type": "text", + "content": "14. Bradbury, J., Frostig, R., Hawkins, P., Johnson, M.J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., Zhang, Q.: JAX: composable transformations of Python+NumPy programs (2018), http://github.com/google/jax" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 37, + 79, + 379, + 111 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 79, + 379, + 111 + ], + "spans": [ + { + "bbox": [ + 37, + 79, + 379, + 111 + ], + "type": "text", + "content": "15. Braverman, M., Garg, A., Ma, T., Nguyen, H.L., Woodruff, D.P.: Communication lower bounds for statistical estimation problems via a distributed data processing inequality. In: ACM Symposium on Theory of Computing (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 37, + 111, + 379, + 143 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 111, + 379, + 143 + ], + "spans": [ + { + "bbox": [ + 37, + 111, + 379, + 143 + ], + "type": "text", + "content": "16. Caldas, S., Konečny, J., McMahan, H.B., Talwalkar, A.: Expanding the reach of federated learning by reducing client resource requirements. arXiv preprint arXiv:1812.07210 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 143, + 379, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 143, + 379, + 164 + ], + "spans": [ + { + "bbox": [ + 37, + 143, + 379, + 164 + ], + "type": "text", + "content": "17. Caldas, S., Wu, P., Li, T., Konečny, J., McMahan, H.B., Smith, V., Talwalkar, A.: LEAF: A benchmark for federated settings. CoRR (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 164, + 379, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 164, + 379, + 196 + ], + "spans": [ + { + "bbox": [ + 37, + 164, + 379, + 196 + ], + "type": "text", + "content": "18. Chen, M., Shlezinger, N., Poor, H.V., Eldar, Y.C., Cui, S.: Communication-efficient federated learning. Proceedings of the National Academy of Sciences (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 196, + 379, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 196, + 379, + 229 + ], + "spans": [ + { + "bbox": [ + 37, + 196, + 379, + 229 + ], + "type": "text", + "content": "19. Chen, Y., Luo, F., Li, T., Xiang, T., Liu, Z., Li, J.: A training-integrity privacy-preserving federated learning scheme with trusted execution environment. Information Sciences (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 229, + 379, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 229, + 379, + 261 + ], + "spans": [ + { + "bbox": [ + 37, + 229, + 379, + 261 + ], + "type": "text", + "content": "20. Dong, Y., Yang, X., Deng, Z., Pang, T., Xiao, Z., Su, H., Zhu, J.: Black-box detection of backdoor attacks with limited information and data. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 261, + 379, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 261, + 379, + 293 + ], + "spans": [ + { + "bbox": [ + 37, + 261, + 379, + 293 + ], + "type": "text", + "content": "21. Eichner, H., Koren, T., McMahan, B., Srebro, N., Talwar, K.: Semi-cyclic stochastic gradient descent. In: International Conference on Machine Learning (ICML) (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 293, + 379, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 293, + 379, + 325 + ], + "spans": [ + { + "bbox": [ + 37, + 293, + 379, + 325 + ], + "type": "text", + "content": "22. Fang, W., Yu, Z., Jiang, Y., Shi, Y., Jones, C.N., Zhou, Y.: Communication-efficient stochastic zeroth-order optimization for federated learning. IEEE Trans. Signal Process. (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 325, + 379, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 325, + 379, + 357 + ], + "spans": [ + { + "bbox": [ + 37, + 325, + 379, + 357 + ], + "type": "text", + "content": "23. Flaxman, A.D., Kalai, A.T., McMahan, H.B.: Online convex optimization in the bandit setting: gradient descent without a gradient. arXiv preprint cs/0408007 (2004)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 37, + 357, + 379, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 357, + 379, + 401 + ], + "spans": [ + { + "bbox": [ + 37, + 357, + 379, + 401 + ], + "type": "text", + "content": "24. Gao, Y., Liu, Y., Zhang, H., Li, Z., Zhu, Y., Lin, H., Yang, M.: Estimatinggpu memory consumption of deep learning models. In: ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 37, + 401, + 379, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 401, + 379, + 433 + ], + "spans": [ + { + "bbox": [ + 37, + 401, + 379, + 433 + ], + "type": "text", + "content": "25. Geiping, J., Bauermeister, H., Dröge, H., Moeller, M.: Inverting gradients-how easy is it to break privacy in federated learning? In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 37, + 433, + 379, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 433, + 379, + 466 + ], + "spans": [ + { + "bbox": [ + 37, + 433, + 379, + 466 + ], + "type": "text", + "content": "26. Ghazi, B., Kumar, R., Manurangsi, P., Pagh, R.: Private counting from anonymous messages: Near-optimal accuracy with vanishing communication overhead. In: International Conference on Machine Learning (ICML) (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 37, + 466, + 359, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 466, + 359, + 476 + ], + "spans": [ + { + "bbox": [ + 37, + 466, + 359, + 476 + ], + "type": "text", + "content": "27. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 37, + 476, + 379, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 476, + 379, + 519 + ], + "spans": [ + { + "bbox": [ + 37, + 476, + 379, + 519 + ], + "type": "text", + "content": "28. Grill, J., Strub, F., Altché, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Doersch, C., Pires, B.A., Guo, Z., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent - A new approach to self-supervised learning. In: NeurIPS 2020 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 37, + 519, + 379, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 519, + 379, + 551 + ], + "spans": [ + { + "bbox": [ + 37, + 519, + 379, + 551 + ], + "type": "text", + "content": "29. Hamer, J., Mohri, M., Suresh, A.T.: Fedboost: A communication-efficient algorithm for federated learning. In: International Conference on Machine Learning (ICML) (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 37, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 37, + 551, + 379, + 583 + ], + "type": "text", + "content": "30. Han, Y., Ozgür, A., Weissman, T.: Geometric lower bounds for distributed parameter estimation under communication constraints. In: Conference On Learning Theory (COLT) (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 36, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 36, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 36, + 35, + 379, + 68 + ], + "type": "text", + "content": "31. Hao, M., Li, H., Luo, X., Xu, G., Yang, H., Liu, S.: Efficient and privacy-enhanced federated learning for industrial artificial intelligence. IEEE Transactions on Industrial Informatics (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 36, + 68, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 68, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 36, + 68, + 379, + 100 + ], + "type": "text", + "content": "32. Hard, A., Rao, K., Mathews, R., Ramaswamy, S., Beaufays, F., Augenstein, S., Eichner, H., Kiddon, C., Ramage, D.: Federated learning for mobile keyboard prediction. arXiv preprint arXiv:1811.03604 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 36, + 101, + 379, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 101, + 379, + 134 + ], + "spans": [ + { + "bbox": [ + 36, + 101, + 379, + 134 + ], + "type": "text", + "content": "33. He, C., Annavaram, M., Avestimehr, S.: Group knowledge transfer: Federated learning of large cnns at the edge. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 134, + 379, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 134, + 379, + 167 + ], + "spans": [ + { + "bbox": [ + 37, + 134, + 379, + 167 + ], + "type": "text", + "content": "34. He, D., Shi, W., Li, S., Gao, X., Zhang, J., Bian, J., Wang, L., Liu, T.Y.: Learning physics-informed neural networks without stacked back-propagation. arXiv preprint arXiv:2202.09340 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 167, + 379, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 167, + 379, + 188 + ], + "spans": [ + { + "bbox": [ + 37, + 167, + 379, + 188 + ], + "type": "text", + "content": "35. Hinton, G., Srivastava, N.: Csc321: Introduction to neural networks and machine learning. Lecture 10 (2010)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 189, + 379, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 189, + 379, + 210 + ], + "spans": [ + { + "bbox": [ + 37, + 189, + 379, + 210 + ], + "type": "text", + "content": "36. Horváth, S., Ho, C.Y., Horvath, L., Sahu, A.N., Canini, M., Richtárik, P.: Natural compression for distributed deep learning. arXiv preprint arXiv:1905.10988 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 211, + 378, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 211, + 378, + 243 + ], + "spans": [ + { + "bbox": [ + 37, + 211, + 378, + 243 + ], + "type": "text", + "content": "37. Howard, A., Pang, R., Adam, H., Le, Q.V., Sandler, M., Chen, B., Wang, W., Chen, L., Tan, M., Chu, G., Vasudevan, V., Zhu, Y.: Searching for mobilenetv3. In: IEEE International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 244, + 379, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 244, + 379, + 276 + ], + "spans": [ + { + "bbox": [ + 37, + 244, + 379, + 276 + ], + "type": "text", + "content": "38. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 277, + 379, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 277, + 379, + 309 + ], + "spans": [ + { + "bbox": [ + 37, + 277, + 379, + 309 + ], + "type": "text", + "content": "39. Huang, Y., Gupta, S., Song, Z., Li, K., Arora, S.: Evaluating gradient inversion attacks and defenses in federated learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 309, + 379, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 309, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 37, + 309, + 379, + 331 + ], + "type": "text", + "content": "40. Huh, M., Agrawal, P., Efros, A.A.: What makes imagenet good for transfer learning? CoRR (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 37, + 331, + 379, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 331, + 379, + 375 + ], + "spans": [ + { + "bbox": [ + 37, + 331, + 379, + 375 + ], + "type": "text", + "content": "41. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R.B., Guadarrama, S., Darrell, T.: Caffe: Convolutional architecture for fast feature embedding. In: Hua, K.A., Rui, Y., Steinmetz, R., Hanjalic, A., Natsev, A., Zhu, W. (eds.) Proceedings of the ACM International Conference on Multimedia (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 37, + 375, + 379, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 375, + 379, + 419 + ], + "spans": [ + { + "bbox": [ + 37, + 375, + 379, + 419 + ], + "type": "text", + "content": "42. Kairouz, P., McMahan, H.B., Avent, B., Bellet, A., Dennis, M., Bhagoji, A.N., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R., et al.: Advances and open problems in federated learning. Foundations and Trends in Machine Learning 14(1-2), 1-210 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 37, + 419, + 379, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 419, + 379, + 440 + ], + "spans": [ + { + "bbox": [ + 37, + 419, + 379, + 440 + ], + "type": "text", + "content": "43. Kang, J., Xiong, Z., Niyato, D., Zou, Y., Zhang, Y., Guizani, M.: Reliable federated learning for mobile networks. IEEE Wireless Communications (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 37, + 441, + 379, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 441, + 379, + 473 + ], + "spans": [ + { + "bbox": [ + 37, + 441, + 379, + 473 + ], + "type": "text", + "content": "44. Kim, K., Kim, C.H., Rhee, J.J., Yu, X., Chen, H., Tian, D., Lee, B.: Vessels: Efficient and scalable deep learning prediction on trusted processors. In: ACM Symposium on Cloud Computing (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 37, + 474, + 379, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 474, + 379, + 506 + ], + "spans": [ + { + "bbox": [ + 37, + 474, + 379, + 506 + ], + "type": "text", + "content": "45. Kim, Y., Sun, J., Yu, H., Jiang, X.: Federated tensor factorization for computational phenotyping. In: ACM International Conference on Knowledge Discovery and Data Mining (SIGKDD) (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 37, + 507, + 379, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 507, + 379, + 528 + ], + "spans": [ + { + "bbox": [ + 37, + 507, + 379, + 528 + ], + "type": "text", + "content": "46. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: 3rd International Conference on Learning Representations, ICLR 2015 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 37, + 528, + 379, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 528, + 379, + 561 + ], + "spans": [ + { + "bbox": [ + 37, + 528, + 379, + 561 + ], + "type": "text", + "content": "47. Konečny, J., McMahan, H.B., Yu, F.X., Richtárik, P., Suresh, A.T., Bacon, D.: Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 37, + 562, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 562, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 37, + 562, + 379, + 583 + ], + "type": "text", + "content": "48. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 379, + 19 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 36, + 35, + 380, + 583 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 36, + 35, + 379, + 57 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 35, + 379, + 57 + ], + "spans": [ + { + "bbox": [ + 36, + 35, + 379, + 57 + ], + "type": "text", + "content": "49. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of IEEE (1998)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 36, + 57, + 379, + 89 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 57, + 379, + 89 + ], + "spans": [ + { + "bbox": [ + 36, + 57, + 379, + 89 + ], + "type": "text", + "content": "50. Li, J., Luo, X., Qiao, M.: On generalization error bounds of noisy gradient methods for non-convex learning. In: International Conference on Learning Representations (ICLR) (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 37, + 89, + 378, + 110 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 89, + 378, + 110 + ], + "spans": [ + { + "bbox": [ + 37, + 89, + 378, + 110 + ], + "type": "text", + "content": "51. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 110, + 380, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 110, + 380, + 142 + ], + "spans": [ + { + "bbox": [ + 37, + 110, + 380, + 142 + ], + "type": "text", + "content": "52. Li, T., Hu, S., Beirami, A., Smith, V.: Ditto: Fair and robust federated learning through personalization. In: International Conference on Machine Learning (ICML) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 143, + 379, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 143, + 379, + 175 + ], + "spans": [ + { + "bbox": [ + 37, + 143, + 379, + 175 + ], + "type": "text", + "content": "53. Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data. In: International Conference on Learning Representations (ICLR) (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 175, + 379, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 175, + 379, + 196 + ], + "spans": [ + { + "bbox": [ + 37, + 175, + 379, + 196 + ], + "type": "text", + "content": "54. Li, Z., Chen, L.: Communication-efficient decentralized zeroth-order method on heterogeneous data. In: WCSP 2021 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 196, + 379, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 196, + 379, + 239 + ], + "spans": [ + { + "bbox": [ + 37, + 196, + 379, + 239 + ], + "type": "text", + "content": "55. Liu, S., Chen, P.Y., Kailkhura, B., Zhang, G., Hero III, A.O., Varshney, P.K.: A primer on zeroth-order optimization in signal processing and machine learning: Principals, recent advances, and applications. IEEE Signal Processing Magazine (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 239, + 379, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 239, + 379, + 271 + ], + "spans": [ + { + "bbox": [ + 37, + 239, + 379, + 271 + ], + "type": "text", + "content": "56. Liu, Y., Suresh, A.T., Yu, F.X.X., Kumar, S., Riley, M.: Learning discrete distributions: user vs item-level privacy. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 271, + 379, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 271, + 379, + 304 + ], + "spans": [ + { + "bbox": [ + 37, + 271, + 379, + 304 + ], + "type": "text", + "content": "57. Luo, X., Wu, Y., Xiao, X., Ooi, B.C.: Feature inference attack on model predictions in vertical federated learning. In: IEEE International Conference on Data Engineering (ICDE) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 304, + 379, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 304, + 379, + 336 + ], + "spans": [ + { + "bbox": [ + 37, + 304, + 379, + 336 + ], + "type": "text", + "content": "58. Lyu, L., Yu, H., Ma, X., Sun, L., Zhao, J., Yang, Q., Yu, P.S.: Privacy and robustness in federated learning: Attacks and defenses. IEEE Transactions on Neural Networks and Learning Systems pp. 1-21 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 37, + 336, + 379, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 336, + 379, + 368 + ], + "spans": [ + { + "bbox": [ + 37, + 336, + 379, + 368 + ], + "type": "text", + "content": "59. Ma, J., Zhang, Q., Lou, J., Ho, J.C., Xiong, L., Jiang, X.: Privacy-preserving tensor factorization for collaborative health data analysis. In: ACM International Conference on Information and Knowledge Management (CIKM) (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 37, + 368, + 379, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 368, + 379, + 400 + ], + "spans": [ + { + "bbox": [ + 37, + 368, + 379, + 400 + ], + "type": "text", + "content": "60. Marfoq, O., Neglia, G., Bellet, A., Kameni, L., Vidal, R.: Federated multi-task learning under a mixture of distributions. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 37, + 400, + 379, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 400, + 379, + 444 + ], + "spans": [ + { + "bbox": [ + 37, + 400, + 379, + 444 + ], + "type": "text", + "content": "61. McKeen, F., Alexandrovich, I., Anati, I., Caspi, D., Johnson, S., Leslie-Hurd, R., Rozas, C.: Intel® software guard extensions (intel® sgx) support for dynamic memory management inside an enclave. In: Proceedings of the Hardware and Architectural Support for Security and Privacy (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 37, + 444, + 379, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 444, + 379, + 476 + ], + "spans": [ + { + "bbox": [ + 37, + 444, + 379, + 476 + ], + "type": "text", + "content": "62. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 37, + 476, + 379, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 476, + 379, + 507 + ], + "spans": [ + { + "bbox": [ + 37, + 476, + 379, + 507 + ], + "type": "text", + "content": "63. McMahan, B., Ramage, D., Talwar, K., Zhang, L.: Learning differentially private recurrent language models. In: International Conference on Learning Representations (ICLR) (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 37, + 507, + 379, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 507, + 379, + 551 + ], + "spans": [ + { + "bbox": [ + 37, + 507, + 379, + 551 + ], + "type": "text", + "content": "64. Mo, F., Haddadi, H., Katevas, K., Marin, E., Perino, D., Courtellis, N.: Ppf: privacy-preserving federated learning with trusted execution environments. In: Annual International Conference on Mobile Systems, Applications, and Services (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 37, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 37, + 551, + 379, + 583 + ], + "type": "text", + "content": "65. Mondal, A., More, Y., Rooparaghunath, R.H., Gupta, D.: Flatee: Federated learning across trusted execution environments. arXiv preprint arXiv:2111.06867 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 36, + 35, + 380, + 583 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 36, + 35, + 379, + 57 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 35, + 379, + 57 + ], + "spans": [ + { + "bbox": [ + 36, + 35, + 379, + 57 + ], + "type": "text", + "content": "66. Moulay, E., Léchappé, V., Plestan, F.: Properties of the sign gradient descent algorithms. Inf. Sci. 492, 29-39 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 36, + 57, + 379, + 89 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 57, + 379, + 89 + ], + "spans": [ + { + "bbox": [ + 36, + 57, + 379, + 89 + ], + "type": "text", + "content": "67. Nakandala, S., Nagrecha, K., Kumar, A., Papakonstantinou, Y.: Incremental and approximate computations for accelerating deep CNN inference. ACM Trans. Database Syst. 45 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 36, + 89, + 380, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 89, + 380, + 131 + ], + "spans": [ + { + "bbox": [ + 36, + 89, + 380, + 131 + ], + "type": "text", + "content": "68. Nasr, M., Shokri, R., Houmansadr, A.: Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In: IEEE symposium on security and privacy (S&P). IEEE (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 132, + 379, + 154 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 132, + 379, + 154 + ], + "spans": [ + { + "bbox": [ + 37, + 132, + 379, + 154 + ], + "type": "text", + "content": "69. Nesterov, Y., Spokoiny, V.: Random gradient-free minimization of convex functions. Foundations of Computational Mathematics (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 154, + 379, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 154, + 379, + 185 + ], + "spans": [ + { + "bbox": [ + 37, + 154, + 379, + 185 + ], + "type": "text", + "content": "70. Pang, T., Xu, K., Li, C., Song, Y., Ermon, S., Zhu, J.: Efficient learning of generative models via finite-difference score matching. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 186, + 379, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 186, + 379, + 217 + ], + "spans": [ + { + "bbox": [ + 37, + 186, + 379, + 217 + ], + "type": "text", + "content": "71. Pang, T., Yang, X., Dong, Y., Su, H., Zhu, J.: Accumulative poisoning attacks on real-time data. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 218, + 379, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 218, + 379, + 271 + ], + "spans": [ + { + "bbox": [ + 37, + 218, + 379, + 271 + ], + "type": "text", + "content": "72. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E.Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 273, + 379, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 273, + 379, + 314 + ], + "spans": [ + { + "bbox": [ + 37, + 273, + 379, + 314 + ], + "type": "text", + "content": "73. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 315, + 379, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 315, + 379, + 347 + ], + "spans": [ + { + "bbox": [ + 37, + 315, + 379, + 347 + ], + "type": "text", + "content": "74. Rothchild, D., Panda, A., Ullah, E., Ivkin, N., Stoica, I., Braverman, V., Gonzalez, J., Arora, R.: Fetchsgd: Communication-efficient federated learning with sketching. In: International Conference on Machine Learning (ICML) (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 348, + 379, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 348, + 379, + 380 + ], + "spans": [ + { + "bbox": [ + 37, + 348, + 379, + 380 + ], + "type": "text", + "content": "75. Sattler, F., Wiedemann, S., Müller, K.R., Samek, W.: Robust and communication-efficient federated learning from non-iid data. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 37, + 380, + 379, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 380, + 379, + 411 + ], + "spans": [ + { + "bbox": [ + 37, + 380, + 379, + 411 + ], + "type": "text", + "content": "76. Seetharaman, S., Malaviya, S., KV, R., Shukla, M., Lodha, S.: Influence based defense against data poisoning attacks in online learning. arXiv preprint arXiv:2104.13230 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 37, + 412, + 379, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 412, + 379, + 455 + ], + "spans": [ + { + "bbox": [ + 37, + 412, + 379, + 455 + ], + "type": "text", + "content": "77. Sharma, H., Park, J., Suda, N., Lai, L., Chau, B., Chandra, V., Esmaeilzadeh, H.: Bit fusion: Bit-level dynamically composable architecture for accelerating deep neural network. In: IEEE Annual International Symposium on Computer Architecture (ISCA) (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 37, + 455, + 379, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 455, + 379, + 487 + ], + "spans": [ + { + "bbox": [ + 37, + 455, + 379, + 487 + ], + "type": "text", + "content": "78. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. In: IEEE symposium on security and privacy (S&P). IEEE (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 37, + 488, + 379, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 488, + 379, + 509 + ], + "spans": [ + { + "bbox": [ + 37, + 488, + 379, + 509 + ], + "type": "text", + "content": "79. Smith, V., Chiang, C.K., Sanjabi, M., Talwalkar, A.S.: Federated multi-task learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 37, + 509, + 379, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 509, + 379, + 530 + ], + "spans": [ + { + "bbox": [ + 37, + 509, + 379, + 530 + ], + "type": "text", + "content": "80. Stein, C.M.: Estimation of the mean of a multivariate normal distribution. The annals of Statistics (1981)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 37, + 530, + 379, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 530, + 379, + 551 + ], + "spans": [ + { + "bbox": [ + 37, + 530, + 379, + 551 + ], + "type": "text", + "content": "81. Sun, Z., Kairouz, P., Suresh, A.T., McMahan, H.B.: Can you really backdoor federated learning? arXiv preprint arXiv:1911.07963 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 37, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 37, + 551, + 379, + 583 + ], + "type": "text", + "content": "82. Suresh, A.T., Felix, X.Y., Kumar, S., McMahan, H.B.: Distributed mean estimation with limited communication. In: International Conference on Machine Learning (ICML) (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 380, + 584 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 36, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 36, + 35, + 379, + 68 + ], + "type": "text", + "content": "83. Tramer, F., Boneh, D.: Slalom: Fast, verifiable and private execution of neural networks in trusted hardware. In: International Conference on Learning Representations (ICLR) (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 36, + 68, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 68, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 36, + 68, + 379, + 100 + ], + "type": "text", + "content": "84. Truex, S., Baracaldo, N., Anwar, A., Steinke, T., Ludwig, H., Zhang, R., Zhou, Y.: A hybrid approach to privacy-preserving federated learning. In: ACM workshop on Artificial Intelligence and Security (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 36, + 101, + 380, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 101, + 380, + 134 + ], + "spans": [ + { + "bbox": [ + 36, + 101, + 380, + 134 + ], + "type": "text", + "content": "85. Truong, J.B., Gallagher, W., Guo, T., Walls, R.J.: Memory-efficient deep learning inference in trusted execution environments. In: IEEE International Conference on Cloud Engineering (IC2E) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 134, + 379, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 134, + 379, + 167 + ], + "spans": [ + { + "bbox": [ + 37, + 134, + 379, + 167 + ], + "type": "text", + "content": "86. Umuroglu, Y., Rasnayake, L., Själander, M.: Bismo: A scalable bit-serial matrix multiplication overlay for reconfigurable computing. In: International Conference on Field Programmable Logic and Applications (FPL) (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 167, + 379, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 167, + 379, + 189 + ], + "spans": [ + { + "bbox": [ + 37, + 167, + 379, + 189 + ], + "type": "text", + "content": "87. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: CVPR (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 189, + 380, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 189, + 380, + 232 + ], + "spans": [ + { + "bbox": [ + 37, + 189, + 380, + 232 + ], + "type": "text", + "content": "88. Wang, H., Sreenivasan, K., Rajput, S., Vishwakarma, H., Agarwal, S., Sohn, J.y., Lee, K., Papailiopoulos, D.: Attack of the tails: Yes, you really can backdoor federated learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 233, + 380, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 233, + 380, + 265 + ], + "spans": [ + { + "bbox": [ + 37, + 233, + 380, + 265 + ], + "type": "text", + "content": "89. Wang, H.P., Stich, S., He, Y., Fritz, M.: Prog fed: effective, communication, and computation efficient federated learning by progressive training. In: International Conference on Machine Learning (ICML) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 266, + 379, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 266, + 379, + 299 + ], + "spans": [ + { + "bbox": [ + 37, + 266, + 379, + 299 + ], + "type": "text", + "content": "90. Wang, J., Liu, Q., Liang, H., Joshi, G., Poor, H.V.: Tackling the objective inconsistency problem in heterogeneous federated optimization. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 299, + 379, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 299, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 37, + 299, + 379, + 331 + ], + "type": "text", + "content": "91. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 331, + 379, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 331, + 379, + 364 + ], + "spans": [ + { + "bbox": [ + 37, + 331, + 379, + 364 + ], + "type": "text", + "content": "92. Wang, L., Wang, W., Li, B.: Cmfl: Mitigating communication overhead for federated learning. In: IEEE International Conference on Distributed Computing Systems (ICDCS) (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 37, + 364, + 379, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 364, + 379, + 397 + ], + "spans": [ + { + "bbox": [ + 37, + 364, + 379, + 397 + ], + "type": "text", + "content": "93. Wei, K., Li, J., Ding, M., Ma, C., Yang, H.H., Farokhi, F., Jin, S., Quek, T.Q., Poor, H.V.: Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 37, + 397, + 379, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 397, + 379, + 408 + ], + "spans": [ + { + "bbox": [ + 37, + 397, + 379, + 408 + ], + "type": "text", + "content": "94. Wu, Y., He, K.: Group normalization. Int. J. Comput. Vis. 128(3), 742-755 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 37, + 408, + 379, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 408, + 379, + 440 + ], + "spans": [ + { + "bbox": [ + 37, + 408, + 379, + 440 + ], + "type": "text", + "content": "95. Xie, C., Huang, K., Chen, P.Y., Li, B.: Distributed backdoor attacks against federated learning. In: International Conference on Learning Representations (ICLR) (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 37, + 441, + 379, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 441, + 379, + 473 + ], + "spans": [ + { + "bbox": [ + 37, + 441, + 379, + 473 + ], + "type": "text", + "content": "96. Yang, T., Andrew, G., Eichner, H., Sun, H., Li, W., Kong, N., Ramage, D., Beaufays, F.: Applied federated learning: Improving google keyboard query suggestions. arXiv preprint arXiv:1812.02903 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 37, + 474, + 379, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 474, + 379, + 506 + ], + "spans": [ + { + "bbox": [ + 37, + 474, + 379, + 506 + ], + "type": "text", + "content": "97. Yang, X., Feng, Y., Fang, W., Shao, J., Tang, X., Xia, S.T., Lu, R.: An accuracy-lossless perturbation method for defending privacy attacks in federated learning. In: ACM Web Conference (WWW) (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 37, + 507, + 379, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 507, + 379, + 528 + ], + "spans": [ + { + "bbox": [ + 37, + 507, + 379, + 528 + ], + "type": "text", + "content": "98. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: Proceedings of the British Machine Vision Conference 2016 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 37, + 529, + 379, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 529, + 379, + 551 + ], + "spans": [ + { + "bbox": [ + 37, + 529, + 379, + 551 + ], + "type": "text", + "content": "99. Zeng, D., Liang, S., Hu, X., Wang, H., Xu, Z.: Fedlab: A flexible federated learning framework. arXiv preprint arXiv:2107.11621 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "text", + "content": "100. Zhang, Y., Duchi, J., Jordan, M.I., Wainwright, M.J.: Information-theoretic lower bounds for distributed statistical estimation with communication constraints. Advances in Neural Information Processing Systems (NeurIPS) (2013)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 199, + 20 + ], + "type": "text", + "content": "Haozhe Feng, Tianyu Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 34, + 379, + 124 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 33, + 34, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 34, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 33, + 34, + 379, + 68 + ], + "type": "text", + "content": "101. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 68, + 379, + 90 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 68, + 379, + 90 + ], + "spans": [ + { + "bbox": [ + 33, + 68, + 379, + 90 + ], + "type": "text", + "content": "102. Zhao, Y., Li, M., Lai, L., Suda, N., Civin, D., Chandra, V.: Federated learning with non-iid data. arXiv preprint arXiv:1806.00582 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 90, + 379, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 90, + 379, + 124 + ], + "spans": [ + { + "bbox": [ + 33, + 90, + 379, + 124 + ], + "type": "text", + "content": "103. Zhu, Z., Wu, J., Yu, B., Wu, L., Ma, J.: The anisotropic noise in stochastic gradient descent: Its behavior of escaping from sharp minima and regularization effects. In: International Conference on Machine Learning (ICML) (2019)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 75, + 9, + 346, + 21 + ], + "type": "text", + "content": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "spans": [ + { + "bbox": [ + 368, + 10, + 378, + 19 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 20 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_content_list.json b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f5da4a3700d8c9b0a3309fa817738bef40bca509 --- /dev/null +++ b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_content_list.json @@ -0,0 +1,1749 @@ +[ + { + "type": "text", + "text": "BAGS: Blur Agnostic Gaussian Splitting through Multi-Scale Kernel Modeling", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Peng*, Yutao Tang*, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa", + "bbox": [ + 217, + 210, + 785, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Johns Hopkins University, Baltimore MD 21218, USA {cpeng26, ytang67, yzhou223, nwang43, xliu253, dlil90, rchella4}@jhu.edu", + "bbox": [ + 250, + 255, + 754, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent efforts in using 3D Gaussians for scene reconstruction and novel view synthesis can achieve impressive results on curated benchmarks; however, images captured in real life are often blurry. In this work, we analyze the robustness of Gaussian-Splatting-based methods against various image blur, such as motion blur, defocus blur, downscaling blur, etc. Under these degradations, Gaussian-Splatting-based methods tend to overfit and produce worse results than Neural-Radiance-Field-based methods. To address this issue, we propose Blur Agnostic Gaussian Splatting (BAGS). BAGS introduces additional 2D modeling capacities such that a 3D-consistent and high quality scene can be reconstructed despite image-wise blur. Specifically, we model blur by estimating per-pixel convolution kernels from a Blur Proposal Network (BPN). BPN is designed to consider spatial, color, and depth variations of the scene to maximize modeling capacity. Additionally, BPN also proposes a quality-assessing mask, which indicates regions where blur occur. Finally, we introduce a coarse-to-fine kernel optimization scheme; this optimization scheme is fast and avoids sub-optimal solutions due to a sparse point cloud initialization, which often occurs when we apply Structure-from-Motion on blurry images. We demonstrate that BAGS achieves photorealistic renderings under various challenging blur conditions and imaging geometry, while significantly improving upon existing approaches.", + "bbox": [ + 261, + 316, + 743, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 633, + 375, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High quality scene reconstruction and novel view synthesis from 2D images is a long-standing research problem with extensive applications in robotics, virtual reality, e-commerce, cinematography, etc. Significant advancements in this field have been made in recent years with the introduction of Neural Radiance Field (NeRF) [24]. NeRF can achieve photorealistic view synthesis by implicitly representing the scene with a Multi-Layer Perceptron (MLP) and optimizing the MLP with differentiable ray-tracing. Many subsequent works have since been proposed to improve various aspects of NeRF, including optimization acceleration [6,9,25,37], anti-aliasing [2,3], dynamic and generative modeling [19,31,32,40], etc. Recently, 3D Gaussian Splatting (3DGS) [12] has been", + "bbox": [ + 212, + 664, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 232, + 825, + 372, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "introduced as an alternative to NeRF-based approaches. 3DGS uses Gaussians as explicit 3D representations and a differentiable rasterization technique, which enables high quality scene reconstruction in a short time.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8d3fcd2a975c5ff47d3896df15332accc4b114e259f8f13e0234835833cd5d68.jpg", + "image_caption": [ + "Fig. 1: We introduce BAGS, which can reconstruct high quality scenes even from blurry training images. Moreover, BAGS can provide kernels and masks that indicate the types and regions of the blur, as shown by the highlighted regions under Mask." + ], + "image_footnote": [], + "bbox": [ + 228, + 218, + 429, + 340 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6d6a7a545446cf20e33a0dd494818bb705972490b3261080cf73a7031fee036d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 219, + 555, + 339 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/adf1ed441c893461dd93777535ee991d9a7d6817c6c3ade3ed33c852faaf8490.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 218, + 787, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Practical applications of 3DGS, however, can be challenging. Despite its impressive performance on curated datasets, 3DGS requires high quality images and a good point cloud initialization to work well. Without these conditions, 3DGS often generates undesired Gaussians to overfit observation noise, leading to worse renderings, e.g., compared to NeRF. For instance, images acquired in real life or from the internet are often in non-ideal conditions. The camera may be unstable or out of focus, leading to motion or defocus blur. Images may also have different resolutions depending on the sensor types or post processes.", + "bbox": [ + 212, + 402, + 787, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There often exists sufficient information within a multi-view image set to generate a high quality reconstruction, even if the individual images are degraded. This is because the same region is often observed repeatedly from different views; such redundancy can then be explored to better recover the underlying scene. Works like Deblur-NeRF [21] and its variants [15, 16, 30, 41] are designed based on this insight. Formally, a simple yet powerful forward model can be used to describe the potentially blurry pixel value $\\tilde{C}(x)$ at coordinate $x$ :", + "bbox": [ + 212, + 523, + 787, + 628 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {C} (x) = \\sum_ {x _ {k} \\in \\mathcal {N} (x)} C \\left(x _ {k}\\right) h \\left(x _ {k}\\right) \\text {s . t} \\sum_ {x _ {k} \\in \\mathcal {N} (x)} h \\left(x _ {k}\\right) = 1, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 635, + 785, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $C(x_{k})$ is the clear pixel value at pixel coordinates $x_{k}$ around the neighborhood of $x$ , and $h(x_{k})$ is the blur kernel weights. In this formulation, $C(x_{k})$ is multi-view consistent, and $\\tilde{C}(x)$ is multi-view inconsistent due to the blur from $h$ . Previous methods [15, 16, 21, 30, 41] aim to simultaneously optimize a multi-view consistent $C(x_{k})$ through NeRF and the multi-view inconsistent $\\tilde{C}(x)$ through a view-specific network that predicts $h(x_{k})$ . As such, we render only the clear $C(x_{k})$ after training. In comparison, vanilla NeRF [24] treats $C(x_{k})$ and $\\tilde{C}(x)$ as one variable, forcing multi-view inconsistencies to be baked into the scene.", + "bbox": [ + 212, + 672, + 787, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we build upon the blur formulation in Eq. (1), and explore two key differences between NeRF and 3DGS in robust reconstruction. NeRF-based approaches suffer from slow training speed, i.e. only limited number of rays can", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Peng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "be rendered at once. This is exacerbated by kernel modeling, which requires the rendering of a patch to model a single pixel. To mitigate this issue, current approaches model the kernel $h$ sparsely [15, 21, 30] or only for specific types of blur [16, 41]. Since 3DGS can rasterize entire images efficiently, it potentially has great synergy with Eq. (1). On the flip side, NeRF requires only calibrated camera poses for reconstruction, whereas 3DGS also requires a good point cloud initialization. For noisy images, camera poses can often be estimated based on just a few correct matches, but a dense point cloud is challenging to obtain.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig 1, we introduce Blur Agnostic Gaussian Splatting, or BAGS, which is a Splatting-based method that is robust against various types of blur. At the core of BAGS is a Blur Proposal Network (BPN). Contrary to previous methods [15,21,30] that sparsely estimate $h$ , BPN directly estimates dense kernels on full images during training. We find that estimating a dense $h$ is more effective and no less efficient than estimating a sparse one, which also has to predict kernel positions $x_{k}$ . BPN considers the position and view embedding of pixel $x$ , accounting for the spatial variance in image blur. Furthermore, a three-layer Convolution Neural Network (CNN) is used to compute the RGBD features of the rendered image. These extracted features allow BPN to consider the color and depth variation of the scene, e.g. around edges and corners, more effectively. Finally, BPN also computes a per-pixel mask, which blends the estimated clear and blurry pixels together. We make the observation that not necessarily all pixels are blurry; therefore, by constraining on its sparsity, such a mask can visualize regions where blur occurs. This is particularly useful as a way to evaluate image quality.", + "bbox": [ + 212, + 268, + 787, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to BPN, BAGS proposes a coarse-to-fine kernel optimization scheme. In some cases, BAGS cannot perform reconstruction well if it jointly optimizes BPN with a very sparse point cloud, due to the inherent ambiguity of separating blur from the observation. Furthermore, optimizing a dense kernel on full resolution directly is computationally expensive and time-consuming. To solve these conundrums, we first optimize the scene and BPN at lower scales with a smaller kernel dimension. This allows the optimization process to address a less challenging and ambiguous problem initially. We then gradually increase the resolution of the training images and the estimated kernels; since the blur problem has been partially addressed at a small scale, optimization at larger scales is more stable and effective. Under this scheme, we not only observe improved performance but also faster optimization process.", + "bbox": [ + 212, + 513, + 787, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions can be summarized in three parts:", + "bbox": [ + 238, + 696, + 712, + 712 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a Blur Proposal Network, which considers spatial, depth, and color variations of the scene to model image blur; BPN can also produce a mask that indicates blurry regions in an input image.", + "2. We introduce a coarse-to-fine optimization scheme, which gradually increases the training image resolution and the estimated kernel size with additional neural network layers; this improves the stability of the joint optimization process given a sparse point cloud." + ], + "bbox": [ + 220, + 731, + 782, + 835 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BAGS", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. We evaluate the overall method, BAGS, on three image blur scenarios and find significant quantitative and visual improvements compared to current SoTA methods.", + "bbox": [ + 220, + 146, + 784, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work and Background", + "text_level": 1, + "bbox": [ + 215, + 205, + 550, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single Image Restoration. Restoring images from degradation, such as blur, has a long history of research. In general, this is a heavily ill-posed problem as many potential high quality images can lead to the same degraded observations. Traditionally, various priors are used to constrain the solution space [5, 14, 18, 34, 38, 43]. More recently, deep-learning-based methods have achieved great performances by learning a better prior in a data driven way, e.g. as a function that directly maps blurry images to clean images. In this paradigm, we train a powerful restoration model with millions of training pairs and a large neural network. Ideally, images can be restored by such a model [1, 4, 8, 13, 17, 42, 45, 47]. This paradigm has many issues in practice. For example, real training pairs are difficult to acquire in large scale; therefore, synthetic data is often used, which leads to significant domain shift problems. Every type of degradation also requires its own training sets and models, which exacerbate the data and computation costs. Even for degradations such as low resolution, for which is easy to synthesize training pairs, the restoration is unlikely to be multi-view consistent.", + "bbox": [ + 212, + 239, + 787, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF in Non-Ideal Conditions. NeRF [24] works well when images are clean and well-calibrated; however, images acquired in the wild are often less ideal. Factors such as low light, camera motion, object motion, and incorrect focus can degrade image quality or affect multi-view consistency. Many works introduce domain knowledge to model the non-ideal contributors, such that the radiance field only models a canonical, multi-view consistent scene. RawNeRF [23] uses an approximated tonemapped loss and variable exposure to account for low light observations. NeRF-W [22] uses an image-specific embedding to model inconsistent appearances and transient objects. For object motion, many works [28,29,33,40] use an implicit function to describe scene deformation, such that the traced rays are first deformed. For image degradation such as blur, similar ideas have been used [11, 15, 21, 30, 41] to model the forward image degradation process; i.e. the 3D consistent scene and 2D inconsistent degradation will be separated based on the multi-view photometric loss.", + "bbox": [ + 212, + 470, + 787, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. Recently, 3DGS [12] introduces a novel, point-based scene reconstruction method that improves upon NeRF in many aspects. 3DGS utilizes a set of 3D Gaussian primitives [48] $\\{\\mathcal{G}_n|n = 1,\\dots ,N\\}$ to represent the scene. These Gaussians are parameterized as follows:", + "bbox": [ + 212, + 689, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {n} (\\mathbf {v}) = e ^ {- \\frac {1}{2} (\\mathbf {v} - \\mathbf {p} _ {n}) ^ {T} \\Sigma_ {n} ^ {- 1} (\\mathbf {v} - \\mathbf {p} _ {n})}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 753, + 784, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the center $\\mathbf{p}_n\\in \\mathbb{R}^{3\\times 1}$ and covariance matrix $\\Sigma_{n}\\in \\mathbb{R}^{3\\times 3}$ define the position, scale, and rotation of the Gaussian at point $\\mathbf{v}\\in \\mathbb{R}^{3\\times 1}$ in space. Additionally, each Gaussian also contains opacity $\\alpha$ and color $c$ . For rendering, the Gaussians are projected onto the image plane as $\\mathcal{G}_n^{\\mathrm{2D}}$ , where the center and covariance ma", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Peng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trices are transformed based on the camera rotation $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ and translation $\\mathbf{t} \\in \\mathbb{R}^{3 \\times 1}$ :", + "bbox": [ + 212, + 146, + 782, + 172 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {n} ^ {\\mathrm {2 D}} = \\mathbf {R} \\mathbf {p} _ {n} + \\mathbf {t}, \\Sigma_ {n} ^ {\\mathrm {2 D}} = \\mathbf {R} \\Sigma_ {n} \\mathbf {R} ^ {T}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 170, + 784, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can then aggregate 2D Gaussians according to their depth order $D$ over a pixel/tile, based on alpha blending:", + "bbox": [ + 212, + 191, + 782, + 222 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC (x) = \\sum_ {i = 1} ^ {D} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i} \\mathcal {G} _ {i} ^ {\\mathrm {2 D}}, T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} \\mathcal {G} _ {j} ^ {\\mathrm {2 D}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 231, + 785, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With well-designed tiling, this splatting-based rasterization technique is significantly faster than ray-tracing in NeRF-based methods without affecting rendering quality; various work have ensued to further improve the Gaussian representation [10, 26, 27, 44] and the rasterization process [20, 46].", + "bbox": [ + 212, + 275, + 782, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Blur Agnostic Gaussian Splitting", + "text_level": 1, + "bbox": [ + 215, + 349, + 576, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BAGS is designed to robustly optimize 3D Gaussians by introducing additional modeling capacities in 2D, which allows BAGS to model away 3D-inconsistent blur from the scene. As shown in Fig. 2, BAGS consists of two parts: a Blur Proposal Network (BPN) and a coarse-to-fine optimization scheme. For clarity, we first describe the design of BPN on a single scale, and then describe modifications on BPN and the training loss on scale $s$ for multi-scale training.", + "bbox": [ + 212, + 378, + 784, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Blur Proposal Network", + "text_level": 1, + "bbox": [ + 215, + 482, + 454, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the forward blur model described in Eq. (1), BPN models a per-pixel convolution kernel $h(x) \\in \\mathbb{R}^{K \\times K}$ for every pixel $x$ , where $K$ is the kernel size. The rasterization efficiency in Gaussian Splatting allows us to model $h(x)$ as a full convolution kernel. Additionally, BPN estimates a per-pixel scalar mask $m(x) \\in [0, 1]$ , controlling the areas where blur modeling takes place.", + "bbox": [ + 212, + 508, + 782, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Modal Features. In scene reconstruction, a large neural network will impact training time. As such, we use a four-layer MLP $\\mathcal{F}_{\\mathrm{kernel}}$ to model blur. To maximize the capacity of $\\mathcal{F}_{\\mathrm{kernel}}$ , we design Multi-Modal Features (MMF) that can help distinguish the scene from blur. Real image blur typically varies between different observed images and different pixels within an image; therefore, we provide $\\mathcal{F}_{\\mathrm{kernel}}$ with a learnable view embedding $l(i)$ and a positional embedding $p(x)$ for training view $i$ and coordinate $x$ . Additionally, we provide $\\mathcal{F}_{\\mathrm{kernel}}$ with a set of features $f_{\\mathrm{RGBD}}$ extracted from the rendered image $C$ and depth $D$ from rasterization, using a small, three-layer CNN $\\mathcal{F}_{\\mathrm{feat}}$ . These features are concatenated with $l(i)$ and $p(x)$ to form the MMF for $\\mathcal{F}_{\\mathrm{kernel}}$ . Consequently, the forward process to estimate the kernel $h(x,i)$ and the mask $m(x,i)$ given a specific pixel $x$ and training view $i$ can be noted as:", + "bbox": [ + 212, + 590, + 785, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh (x, i), m (x, i) = \\mathcal {F} _ {\\mathrm {k e r n e l}} (l (i) \\oplus p (x) \\oplus f _ {\\mathrm {R G B D}} (x, i)),\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 773, + 683, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} f _ {\\mathrm {R G B D}} (x, i) = \\mathcal {F} _ {\\text {f e a t}} (C (x, i) \\oplus D (x, i)), 0 \\leq m \\leq 1.\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 787, + 702, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Incorporating RGBD features enbales BPN to explore the patch statistics around the current pixel, especially at edges and corners where blur is prominent. Scene", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BAGS", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b016e80466ffca5bd1f2417ee88c805e96dacf91e99ab72f453121fbfb55acce.jpg", + "image_caption": [ + "Fig. 2: BAGS is optimized by training a Blur Proposal Network on top of the scene $\\mathcal{G}(\\mathbf{v})$ over multiple scales. Top: We extract $f_{\\mathrm{RGBD}}^s$ from color and depth, which are concatenated with the position and view embedding $p(x), l(i)$ to form the Multi-Modal Feature (MMF). The kernel MLP then estimates the per-pixel kernel $h^s$ and mask $m^s$ . We use $h^s$ to model the blur image $\\tilde{C}^s$ and employ $m^s$ to blend the rendered image $C^s$ and the blur-modeled image $\\tilde{C}^s$ , yielding $C_{\\mathrm{out}}^s$ . Bottom: After $N_s$ steps, we upscale image resolution and modify the kernel MLP to produce $h^{s - 1}$ with a larger kernel size." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 787, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "depth also strongly correlates with the magnitude of blur. For example, defocus blur affects pixels that are outside of a specific focus plane; scene content that is further away from this focus plane becomes more blurry. For camera motion blur, pixels in the near plane may shift more than pixels in the far plane. We find the inclusion of $f_{\\mathrm{RGBD}}$ notably improves BPN's modeling capacity and the quality of novel view synthesis.", + "bbox": [ + 212, + 488, + 787, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Residual Mask Bottleneck. As described in Eq. (5), $\\mathcal{F}_{\\mathrm{kern}}$ also estimates a per-pixel scalar mask $m(x,i)$ . Specifically, this mask is used to blend the Gaussian-splatted image $C$ with the blur-modeled image $\\tilde{C}$ as follows:", + "bbox": [ + 212, + 606, + 787, + 654 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC _ {\\mathrm {o u t}} (x, i) = \\left(1 - m (x, i)\\right) C (x, i) + m (x, i) \\tilde {C} (x, i), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 662, + 785, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $C_{\\mathrm{out}}$ is the final output at training. We note that not necessarily all pixels are blurry in an observed image. The blending $m$ provides an additional degree of freedom to allow $C$ , the clear pixel, to be directly compared with $C_{\\mathrm{obs}}$ ; i.e., if $C_{\\mathrm{obs}}(x,i)$ suffers from blur, then $m(x,i)$ should be large, giving a higher weight to $\\hat{C}(x,i)$ ; if $C_{\\mathrm{obs}}(x,i)$ if free of blur, $m(x,i)$ should be small, favoring $C(x,i)$ .", + "bbox": [ + 212, + 691, + 787, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This design is similar to residual learning by providing the output with a skip connection from $C$ as initialization. We apply a small sparsity constraint on $m$ to encourage a larger weight towards $C$ by default. After training, we can extract $m$ to automatically identify the degraded regions in an input image.", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Peng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59011e96dc7eeaa4cb97086bdd07afed2a6c8f42bac595790d90bc3099cd39f5.jpg", + "image_caption": [ + "(a) Without C2F" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 359, + 217 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9c49ff277167511f89fcbd1b36f4d882ea98fdfd77582fd3ef9ed1c53dc386a.jpg", + "image_caption": [ + "(b) BAGS", + "Fig. 3: Gaussians may get stuck at a local minimum without proper initialization. As indicated in 3a, optimizing $h$ directly can lead to noisy surfaces. Naively densifying the scene before adding $h$ can also lead to noisy Gaussians; as shown in 3c, the noisy Gasussians are not well removed even after adding $h$ . By using a coarse-to-fine training schedule, we achieve better results in 3b and 3d." + ], + "image_footnote": [], + "bbox": [ + 361, + 143, + 500, + 217 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ecc8a9b246519d1859d7ebace6e8b20c303e3c1bc3b803e0be6e9eee12c38f0.jpg", + "image_caption": [ + "(c) Naive densification" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 643, + 217 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8cde32d8cbbf6de12e1f5119189dc5016ef47e701cf81aba73020f40f491aaf.jpg", + "image_caption": [ + "(d) BAGS" + ], + "image_footnote": [], + "bbox": [ + 643, + 143, + 785, + 217 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Coarse-to-Fine Kernel Optimization", + "text_level": 1, + "bbox": [ + 215, + 321, + 558, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A high quality 3DGS reconstruction requires a good point cloud estimation from Structure from Motion (SfM). Camera calibration parameters, which can be estimated from just a few matches, are relatively robust against image blur. On the other hand, obtaining a sufficiently dense point cloud can be challenging. Unfortunately, 3DGS is prone to over-fitting without a good initialization. This problem can worsen when a blur model is introduced and the kernels $h$ provide additional degrees of freedom in scene optimization, as shown in Fig. 3. Furthermore, estimating dense per-pixel kernels on a full resolution image can be computationally expensive. A sparse estimation of the kernel does not meaningfully improve efficiency, as rendering is cheap in rasterization while the neural network to predict any type of kernel is still costly. Fundamentally, the complexity to compute per-pixel kernel scales quadratically with resolution. To address these problems, we propose a coarse-to-fine kernel optimization scheme, which initially optimizes Eq. (5) at a low resolution and gradually increases render resolution, achieving a more stable training process and better efficiency.", + "bbox": [ + 212, + 349, + 787, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Scale BPN. For multi-scale optimization, BPN should generate kernels with a consistent field of view. We define scale based on the downscaling ratio $2^{s - 1}$ in resolution, where the original resolution is defined as $s = 1$ . Additionally, if the image resolution is downscaled by a factor of two, i.e. $s = 2$ , the proposed kernel size $K^s$ should also be similarly downscaled. As shown in Fig. 2, a proper $K^s$ ensures that the kernel is always modeling blur from the same regions.", + "bbox": [ + 212, + 587, + 787, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To enable multi-scale training, we separate $\\mathcal{F}_{\\mathrm{kernel}}^s$ into two parts: a two-layer base MLP $\\mathcal{F}_{\\mathrm{base}}^s$ and two one-layer MLP heads $\\mathcal{F}_{\\mathrm{blur}}^s$ and $\\mathcal{F}_{\\mathrm{mask}}^s$ . The base MLP extracts features from MMF, such features are then provided to the two heads for generating the kernels $h^s$ and masks $m^s$ at scale $s$ . In practice, only $\\mathcal{F}_{\\mathrm{blur}}^s$ has a scale-dependent parameter size corresponding to the kernel size $K^s$ ; therefore, a new MLP layer with output dimension $K^s \\times K^s$ is added for every scale. The updated multi-scale BPN can be described as follows:", + "bbox": [ + 212, + 681, + 787, + 786 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {i n t e r}} (x, i) = \\mathcal {F} _ {\\text {b a s e}} ^ {s} (l (i) \\oplus p (x) \\oplus f _ {\\text {R G B D}} (x, i))\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 787, + 661, + 805 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {s} (x, i) = \\operatorname {s o f t m a x} \\left(\\mathcal {F} _ {\\text {b l u r}} ^ {s} \\left(f _ {\\text {i n t e r}} (x, i)\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 806, + 785, + 824 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nm ^ {s} (x, i) = \\operatorname {s i g m o i d} \\left(\\mathcal {F} _ {\\text {m a s k}} ^ {s} \\left(f _ {\\text {i n t e r}} (x, i)\\right)\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 825, + 640, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BAGS", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Coarse-to-Fine Optimization. We split training into multiple stages based on scale. Starting from the coarsest scale $s = S$ to the finest scale $s = 1$ , we optimize BAGS with a downsampling factor of $2^{s - 1}$ applied to $C, \\tilde{C}, C_{\\mathrm{obs}}$ , and $h$ . The optimization for each scale lasts for $N_{s}$ steps, we then increase the resolution in $C, \\tilde{C}, C_{\\mathrm{obs}}$ , and $h$ . The scale-dependent training loss can be described as:", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\lambda_ {\\mathrm {p h o t o}} \\| C _ {\\mathrm {o u t}} ^ {s} - C _ {\\mathrm {o b s}} ^ {s} \\| + \\lambda_ {\\mathrm {D S}} \\mathcal {L} _ {\\mathrm {D - S S I M}} \\left(C _ {\\mathrm {o u t}} ^ {s}, C _ {\\mathrm {o b s}} ^ {s}\\right) + \\lambda_ {\\mathrm {m a s k}} \\| m ^ {s} \\|, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 223, + 785, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $C_{\\mathrm{obs}}^s$ and $C_{\\mathrm{out}}^s$ are the observed and predicted image at scale $s$ , $\\mathcal{L}_{\\mathrm{D - SSIM}}$ is the structural similarity loss; $\\| m^{s}\\|$ is the mask sparsity loss. We find this training scheme achieves the best of both worlds: activating BPN at low resolution generates less overfitting artifacts during densifying the point cloud given the reduced visibility of blur. Moreover, it also serves as an efficient warm-up for BPN at higher resolution, as computing per-pixel kernels at low resolution is much cheaper.", + "bbox": [ + 212, + 242, + 787, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 361, + 375, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate BAGS under three degradation scenarios: camera motion blur, defocus blur, and mix resolution. All of these degradations are commonly observed in images acquired in real life or online. We further introduce a new unbounded 360 drone dataset, which is collected at sunset or night and experiences degradation due to a combination of motion blur and low light condition.", + "bbox": [ + 212, + 388, + 787, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Camera Motion and Defocus Blur. Camera motion and defocus blur are often observed in real acquisitions, e.g. due to low light, incorrect auto-focus, or limitation in the finite aperture of real cameras. Deblur-NeRF [21] provides real scene acquisitions that contain camera motion and defocus blur; both blur categories consist of ten scenes. While real world scenes have blur-free references, the camera settings, e.g., exposure, may not be consistent with source views. Experiments on synthetic blur are reported in supplemental material.", + "bbox": [ + 212, + 468, + 787, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mix Resolution. To analyze the scenario where the input images are of inconsistent resolution, we apply various downsampling operations on the Mip-NeRF 360 dataset [2]. Specifically, we split the training images into four equal parts and apply 4X, 3X, 2X, and no downscaling to each of the parts. The image splits are sampled in uniform randomness to ensure no downscaling covers a consecutive interval of the camera trajectory. We then use NeRFStudio's data processing pipeline [39], which converts all images to the highest resolution and calibrates with COLMAP [35]. Similar to previous works [2,3,46], we perform the entire process on downscaled data instead of the original 4K resolution.", + "bbox": [ + 212, + 579, + 787, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unbounded Drone Motion Blur. The dataset introduced by DeblurNeRF [21] serves as a good benchmark for performance comparison; however, it is collected in a forward bounded format. These forward bounded scenes focus on a small area and have significant overlaps within the images; as such, it is unclear if the methodology [15, 16, 21, 30, 41] based on these sets can generalize to other scenarios. In fact, none of the deblurring methods can work for unbounded scenes. To supplement DeblurNeRF's dataset and understand how BAGS performs in realistic settings, we collect five sets of drone footage on different buildings in 360", + "bbox": [ + 212, + 719, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Peng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/037523c94d0a7112955b58b87388cbcb0db060bfa5e00227677f55c2271bbe27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 143, + 498, + 246 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ea7ccc980b07eef8189fdc13f9551f029dba350f8a929945f62ee895438e041e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 776, + 247 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/50167ccbb2b1ac244c6d0f66fcbdae44a4c4ad0c964f51ed80ee18a555e19d12.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 225, + 250, + 498, + 354 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4bfb76c4ba18ea9b7865c6aec613cbb509ec68b02498a756cd7cc766489dcb90.jpg", + "image_caption": [ + "(d)", + "Fig. 4: Ablation study of BAGS with different sub-modules. In 4a and 4b, introducing RGBD features and coarse-to-fine optimization improve novel view synthesis quality, and alternative sparse deformable kernel suffers in performance; 4d demonstrates our speed improvement against $\\mathrm{BAGS}^{\\mathrm{noC2F}}$ . In 4c, larger kernel generally leads to better performance, other than in the mix resolution scenario." + ], + "image_footnote": [], + "bbox": [ + 503, + 250, + 776, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "unbounded format. These footages are collected at sunset or night, and suffer from camera motion blur and low light noise. For each scene, we also capture clear frames from a stable drone position for testing, resulting in around 100 images overall. We use NeRFStudio [39] to calibrate images from these footages. More details on this dataset can be found in supplemental material.", + "bbox": [ + 212, + 455, + 784, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. BPN's $\\mathcal{F}_{\\mathrm{feat}}$ is constructed as a three-layer CNN; the intermediate channel size is 64, the kernel size for each layer is 5, 5, and 3. $\\mathcal{F}_{\\mathrm{kernel}}$ is constructed with four linear layers; the intermediate channel size is 64. For multi-scale training, we use three scales starting with scale $s = 3$ , where $K_{3} = 5$ , $K_{2} = 9$ , and $K_{1} = 17$ . BAGS implements its Gaussian Splatting and rasterization based on Mip-Splatting [46], which has been shown to yield good results across multiple resolutions. More implementation details can be found in supplemental material.", + "bbox": [ + 212, + 537, + 784, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 671, + 401, + 684 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare three designs in BAGS to demonstrate the effectiveness of different components:", + "bbox": [ + 212, + 696, + 782, + 727 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- BAGS $^{\\text{noRGBD}}$ : We implement BPN without the RGBD features $f_{\\text{RGBD}}$ .", + "- BAGS $^{\\text{deform}}$ : We implement BPN with a sparse deformable kernel, similar to DeblurNeRF [21].", + "- BAGS $^{\\mathrm{noC2F}}$ : We implement BAGS without a coarse-to-fine kernel optimization scheme." + ], + "bbox": [ + 223, + 729, + 782, + 803 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We further investigate the effect of the final kernel size $K_{1}$ on reconstruction quality across all degradation scenarios.", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BAGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Fig. 4, the introduction of RGBD features to BPN greatly improves its ability to disentangle image blur from the underlying scene. The addition of $f_{\\mathrm{RGBD}}$ increases the average performance by $1.57 / 0.92$ dB in PSNR and $0.03 / 0.03$ in SSIM for camera motion and defocus blur, respectively. This agrees with our intuition that blur is highly correlated with depth and image features like edges. Our coarse-to-fine optimization scheme also notably boosts robustness against defocus blur by $1.64$ dB and $0.08$ in PSNR and SSIM on average. As visualized in Fig. 3a, directly optimizing 3DGS on a sparse point cloud in $\\mathrm{BAGS}^{\\mathrm{noC2F}}$ leaves many ambiguities; while training views have low loss, test views show a lack of the correct surface structure and many artifact Gaussians. Densifying Gaussians without BPN first improves performance for defocus blur; however, this strategy is bad for motion blur, as shown in Fig. 3c. In Fig. 4d, we show that our coarse-to-fine strategy is not only performant, but also fast; on average, BAGS can reach high performance in less than fifteen minutes, which is much faster than the two hours required by PDRF [30]. This strategy is also faster compared to single-scale training, as full-scale training in our scheme uses less iterations, and lower scale training does not sacrifice performance. Both multi-scale and single-scale training take $40 \\mathrm{K}$ iterations overall.", + "bbox": [ + 212, + 146, + 787, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We also implement a deformable version of BPN in $\\mathrm{BAGS}^{\\mathrm{deform}}$ , where BPN estimates the kernel position and kernel weights for ten points within a $20 \\times 20$ -pixel patch. We show its performances in Fig. 4a, where $\\mathrm{BAGS}^{\\mathrm{deform}}$ is tested on camera motion blur. Camera motion blur often has a sparse representation, i.e. the camera motion trajectory [16, 41]; therefore, a ten-point estimation of the kernel may perform well. However, we find that $\\mathrm{BAGS}^{\\mathrm{deform}}$ still underperforms compared to a full kernel estimation in BAGS. This may be due to the difference between NeRF and 3DGS; i.e., a ray-based deformation can be better expressed in NeRF, while a patch-based convolution can be better expressed in 3DGS. We also note that $\\mathrm{BAGS}^{\\mathrm{deform}}$ is not more efficient than BAGS despite its sparse estimation. This is because approaches like [15, 21, 30] require one forward pass to estimate one kernel position and weight; therefore, we need to forward multiple times for every pixel to estimate the sparse kernel. In comparison, BAGS only requires a single forward pass to obtain the full kernel.", + "bbox": [ + 212, + 417, + 787, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, we ablate on the kernel dimension estimated by BPN. In Fig. 4c, we show that larger kernels in general lead to higher performances; interestingly, this is not true in the mixed resolution scenario. This is likely because downscaling is uniform across all pixels in an image, i.e. a downscaling factor of four means a kernel size of five can fully explain the degradation. As such, increasing the kernel size does not improve performance and introduces noise instead.", + "bbox": [ + 212, + 630, + 787, + 723 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 734, + 460, + 750 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Camera Motion and Defocus Blur. As shown in Table 1, we observe that Gaussian Splatting [46] performs markedly worse than vanilla NeRF in the presence of blurry images, despite its strong performance in other curated benchmarks. Since NeRF models a continuous neural volume with an MLP, individual degradation can be smoothly explained by spreading the error into empty space.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Peng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/df5ca05d0a6e89289d13f1c4dee31b237031eaf4f509313749c57ac656d2d3b6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Camera MotionNeRF [24]Mip-Splatting [46]Deblur-NeRF [21]DP-NeRF [15]PDRF [30]BAGS
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Ball24.080.6240.39923.220.6190.34027.360.7660.22327.200.7650.209
Basket23.720.7090.32223.240.6880.28827.670.8450.14827.740.8460.129
Buick21.590.6330.35021.460.6580.26624.770.7700.17525.700.7920.141
Coffee26.480.8060.29024.730.7490.28830.930.8980.12432.440.9150.101
Decoration22.390.6610.36320.550.6410.29924.190.7710.18623.510.7400.275
Girl20.070.7080.32019.870.7140.27822.270.7980.16921.800.7700.205
Heron20.500.5220.41319.430.5050.33222.630.6870.21022.520.6760.281
Parterre23.140.6200.40522.280.5900.32125.820.7600.21624.970.7680.227
Puppet22.090.6090.33922.050.6310.26725.240.7510.15825.120.7580.181
Stair22.870.4560.48721.910.4740.38725.390.6300.21025.780.6470.229
Average22.690.6350.36921.870.6270.30625.630.7670.18225.910.7750.160
Defocus
Cake24.420.7210.22522.170.6450.29026.270.7800.12826.160.7780.127
Caps22.730.6310.28021.000.5030.44223.870.7130.16123.950.7120.143
Cisco20.720.7220.12620.080.7090.16320.830.7270.08720.730.7260.084
Coral19.810.5660.26919.600.5550.32319.850.6000.12122.800.7410.096
Cupcake21.880.6810.21621.550.6810.21322.260.7220.11620.110.6110.118
Cups25.020.7580.23220.930.6460.31426.210.7990.12726.750.8140.104
Daisy22.740.6200.26221.590.6140.27323.520.6870.12123.790.6970.108
Sausage17.790.4830.27917.780.4730.28418.010.5000.18018.350.5440.147
Seal22.790.6270.26822.130.5920.31026.040.7770.10525.950.7780.103
Tools26.080.8520.15523.980.8120.19627.810.8950.06128.070.8980.054
Average22.400.6660.23121.080.6230.28123.470.7200.12123.670.7300.108
", + "bbox": [ + 218, + 143, + 782, + 386 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Quantitative comparisons on the camera motion and defocus blur dataset. We color code the best PSNR, SSIM, and LPIPS performances.", + "bbox": [ + 215, + 388, + 782, + 416 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "On the other hand, 3DGS is based on a discrete representation and tends to produce concentrated artifacts. Our BAGS can achieve new SoTA performances compared to previous NeRF-based deblurring methods despite the unfavorable performances in 3DGS. Specifically, we find significant improvements on camera motion blur across all metrics and achieve a fifty percent improvement in LPIPS on average. As visualized in Fig. 5, BAGS achieves much sharper results compared to previous methods [15, 16, 21, 30], to the point where it is hard to distinguish the groundtruth image from our rendered results.", + "bbox": [ + 212, + 436, + 784, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "While improvements on defocus blur is less drastic in Table 1, we still observe great visual improvements. Several reasons contribute to the less apparent quantitative improvements. First, the collected groundtruth images have different exposure compared to training images; as such, image-wise similarity metrics like PSNR are less reliable. Secondly, many scenes contain specular surfaces, which all methods do a bad job at reconstructing. As shown in Fig. 5, much better details can be recovered through BAGS than other methods.", + "bbox": [ + 212, + 559, + 784, + 665 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mix Resolution and Low Light Motion Blur. We investigate the robustness of several SoTA reconstruction methods on downscaling perturbation and motion blur observed in low light condition. The experiments are done on unbounded scenes and are more challenging and realistic than the forward bounded scenes in DeblurNeRF [21]. We note that previous approaches [15, 21, 30, 41] cannot handle unbounded scenes; therefore, we do not include them in Table 2.", + "bbox": [ + 212, + 672, + 784, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Table 2, Mip Splating performs on par with Mip-NeRF 360; both are vastly better than NeRFacto [39], which stores hash features [25] to accelerate the training process and uses a very small MLP represent the scene. We also employ a Super-Resolution (SR) model HAT [7] on downscaled images in the hope that single-image SR can improve performances. However, pre-trained", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BAGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1359a9896b09a52e78e39c9fcd739a2e436e7770f845c14d15e16fb6c440ce1e.jpg", + "image_caption": [ + "Fig. 5: Visualizations of test views on camera motion and defocus blur dataset. Mip-Sp and Db-NeRF are short for Mip-Splatting [46] and Deblur-NeRF [21]." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 458 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SR models are limited by domain shifts, dedicated upscaling factors, and the lack of multi-view consistency. As visualized in Fig. 6, while the render is somewhat sharper after SR, the color becomes unfaithful. Compared with all other methods, our rendered results are much sharper in details. This robust performance comes from BPN's ability to model Gaussian blur on high resolution before comparing them with the downscaled observations.", + "bbox": [ + 212, + 513, + 782, + 604 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Motion blur is often observed in low light condition. To this end, we perform experiments on scenes acquired near sunset or at night. As shown in Table 2, BAGS again improves significantly over other reconstruction approaches, which shows that our method can be generalized to different imaging geometry. As shown in Fig. 6, we employ MPR [47], a single image deblurring model, on the motion-blurred images before reconstructing with Mip-Splatting [46]. Similar to previous attempts, the deblurring prior leads to a slightly clearer rendering but is not sufficient to remove all blur. BAGS reconstructs the scene remarkably well compared to other methods, where even the window details can be recovered correctly. Please refer to the supplemental material for more visualizations.", + "bbox": [ + 212, + 607, + 784, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Kernel and Mask Visualization. BAGS models image degradation through convolution kernels $h$ and masks $m$ , which we visualize in Fig. 7. By constraining on its sparsity, we can see that $m$ meaningfully highlights the regions of blur in a given training image. On the other hand, low-value regions indicate either high quality observation or textureless regions. By visualizing the kernels modeled", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Peng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3e45467730ec4141014c4cf905dbdee57131eaa03d517fc13a61ae4fc6f40fb4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MixedResNeRFacto [39]Mip-NeRF 360 [2]Mip-Sp [46]Mip-Sp-HAT [7]BAGS
PSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPS
Bicycle19.250.3490.49623.820.6100.37524.030.6280.36423.20
Bonsai21.800.6190.25929.910.8710.19029.270.8760.19328.38
Counter24.070.7140.28028.090.8480.22027.790.8530.21627.02
Flowers19.630.4180.45921.980.5660.38921.960.5700.39720.81
Garden21.610.4890.35725.340.6670.31725.280.6820.30523.96
Kitchen22.730.5840.30427.820.7790.23827.490.7780.24126.10
Room24.680.7910.21430.380.9020.19830.430.9030.20329.61
Stump23.190.5550.37725.750.6880.32025.870.7030.29924.09
Treehill20.110.4160.50122.520.5580.43722.860.5500.45721.76
Average21.900.5480.36126.180.7210.29826.110.7270.29724.99
Lowlight-droneNeRFacto [39]Mip-NeRF 360 [2]Mip-Sp [46]Mip-Sp-MPR [47]BAGS
PSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPS
Math-night21.330.5820.32629.510.8820.19129.610.8840.19929.15
Library-night24.480.7680.38126.380.8210.34926.170.8200.37025.89
BioMed-sunset23.760.6500.43924.480.6880.40224.850.7120.42624.78
Admin-sunset21.960.6210.40025.470.7570.35525.540.7640.36925.19
Aud-sunset24.670.7090.39825.900.7500.36826.330.7700.38626.77
Average23.240.6660.38926.350.7800.33326.500.7900.35026.36
", + "bbox": [ + 220, + 143, + 782, + 383 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a150f63ba1c28b6ed69c90ce742a8964fa304dd8763c209ed722bd9fe0081097.jpg", + "table_caption": [ + "Table 2: Quantitative comparisons on mixed resolution and low light motion blur in unbounded 360 geometry. Mip-Sp is short for Mip-Splatting [46]. Mip-Sp-HAT and Mip-Sp-HAT-MPR use prior models to restore images first before Mip-Splatting." + ], + "table_footnote": [], + "table_body": "
Novel ViewNeRFactoMipNeRF-360Mip-SpMip-Sp-HATOursG.T.
PSNR/SSIM21.61/.57624.09/.67025.11/.74624.41/.72228.46/.899MixedRes
PSNR/SSIM23.08/.41224.98/.56224.93/.59024.68/.58731.79/.949MixedRes
Novel ViewNeRFactoMipNeRF-360Mip-SpMip-Sp-MPROursG.T.
PSNR/SSIM17.86/.60818.33/.70418.57/.71418.51/.71824.47/.850LowLight
PSNR/SSIM22.40/.68325.36/.81325.51/.81727.03/.86731.75/.938LowLight
", + "bbox": [ + 220, + 465, + 782, + 712 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 6: Visualizations of test views on mix resolution and low light motion blur dataset.", + "bbox": [ + 215, + 724, + 782, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "at different pixel, we can also easily characterize the types of observed blur. Specifically, the estimated kernels in camera motion blur exhibit clear patterns of camera movement, while those for defocus blur show Gaussian-like distributions based on the pixel's distance from the focus plane. These self-emerged properties", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BAGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "in BAGS provide ways for us to automatically and precisely evaluate the quality of training images.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/41c5d2ff34a34ecc86c2bdd8b8d8e61aaddeb70ddeed3818c36b9dd2b10b876b.jpg", + "image_caption": [ + "Camera Motion" + ], + "image_footnote": [], + "bbox": [ + 217, + 194, + 359, + 268 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d108c2bd4038b2113bb4c1b36309890bbcb6224036c84d5389aba32cc486fcdb.jpg", + "image_caption": [ + "Kernel & mask", + "Fig. 7: Visualization of estimated convolution kernels and masks in different blur." + ], + "image_footnote": [], + "bbox": [ + 359, + 195, + 500, + 268 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f135d8835104fbccc2059f2138b0de441611621c5d191256ff298c631592f938.jpg", + "image_caption": [ + "Defocus" + ], + "image_footnote": [], + "bbox": [ + 501, + 195, + 643, + 268 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/71e0538c850d11aa89879fe1add6c250e005f017734edcf0c392a007149f40b9.jpg", + "image_caption": [ + "Kernel & mask" + ], + "image_footnote": [], + "bbox": [ + 650, + 195, + 785, + 268 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Discussion. BAGS can achieve great results; however, there also exists many potential improvements. The added neural networks and convolution operations require additional computation. While BAGS is a generalist approach and addresses multiple types of blur, it may become very expensive to estimate per-pixel convolution in high resolution. To this end, we have explored utilizing pixel shuffling [36], which converts spatial resolution into channels, before we estimate blur kernel; i.e., this assumes that within a small region, blur kernel $h$ stays roughly consistent. As shown in the supplemental material, we can scale to great visual results at 2K resolution on the DeblurNeRF dataset. Many other potential directions can also address this complexity issue, including leveraging the mask $m$ to identify important regions to focus on, a better designed low rank kernel estimation, degradation-specific optimization, etc. As such, while BAGS is already more efficient than comparable NeRF-based methods, we anticipate that its computation cost can be further optimized in the future.", + "bbox": [ + 212, + 319, + 787, + 531 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 551, + 357, + 568 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present BAGS, a novel scene reconstruction method that can handle various noise in input images. We demonstrate that a vanilla Gaussian Splatting is particularly susceptible to degraded images due to its discrete representation, even compared with NeRF. To address this issue, BAGS implements a 2D degradation model, BPN, which estimates convolution kernels jointly with 3D scene optimization. BPN allows for additional freedom to address 3D inconsistencies, and produces interpretable kernels and masks to indicate the degradation types and regions within an image. Additionally, BAGS leverages a coarse-to-fine kernel optimization scheme, which gradually models the 3D scene with higher resolution images and kernels. This approach successfully addresses the 2D-3D ambiguity in optimizing convolution kernels with a very sparse point cloud. We perform extensive experiments on a variety of noise in training images, including camera motion blur, defocus blur, and low resolution. Under these conditions, BAGS can render high fidelity scenes and significantly improves upon previous approaches. In the future, we hope to further improve the computational efficiency in BAGS and dynamically adjust kernel capacity based on how degraded the training views are.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Peng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is based upon work supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. The authors also thank the support and feedbacks from researchers, Kaiwen Jiang, Dr. Ravi Ramamoorthi, at UCSD.", + "bbox": [ + 212, + 179, + 787, + 330 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 356, + 321, + 372 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abuolaim, A., Afifi, M., Brown, M.S.: Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1231-1239 (2022)", + "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)", + "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)", + "4. Cao, Y., Liu, M., Liu, S., Wang, X., Lei, L., Zuo, W.: Physics-guided iso-dependent sensor noise modeling for extreme low-light photography. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5744-5753 (2023)", + "5. Chan, T.F., Wong, C.: Total variation blind deconvolution. IEEE Trans. Image Process. 7(3), 370-375 (1998)", + "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)", + "7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22367-22377 (June 2023)", + "8. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4641-4650 (2021)", + "9. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)", + "10. Gao, J., Gu, C., Lin, Y., Zhu, H., Cao, X., Zhang, L., Yao, Y.: Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043 (2023)", + "11. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-nerf: High dynamic range neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18398-18408 (2022)" + ], + "bbox": [ + 218, + 390, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BAGS", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "13. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5886-5895 (2023)", + "14. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011. pp. 233-240. IEEE Computer Society (2011)", + "15. Lee, D., Lee, M., Shin, C., Lee, S.: Dp-nerf: Deblurred neural radiance field with physical scene priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12386-12396 (2023)", + "16. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023)", + "17. Lee, J., Son, H., Rim, J., Cho, S., Lee, S.: Iterative filter adaptive network for single image defocus deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2034-2042 (2021)", + "18. Levin, A.: Blind motion deblurring using image statistics. In: Scholkopf, B., Platt, J.C., Hofmann, T. (eds.) Advances in Neural Information Processing Systems 19, Proceedings of the Twentieth Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada, December 4-7, 2006. pp. 841-848. MIT Press (2006)", + "19. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "20. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109 (2023)", + "21. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)", + "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7210-7219. Computer Vision Foundation / IEEE (2021)", + "23. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16190–16199 (June 2022)", + "24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I. Lecture Notes in Computer Science, vol. 12346, pp. 405-421. Springer (2020)", + "25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Peng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "26. Navaneet, K., Meibodi, K.P., Koohpayegani, S.A., Piriyawash, H.: Compact3d: Compressing gaussian splat radiance field models with vector quantization. arXiv preprint arXiv:2311.18159 (2023)", + "27. Niedermayr, S., Stumpfegger, J., Westermann, R.: Compressed 3d gaussian splatt-ting for accelerated novel view synthesis. arXiv preprint arXiv:2401.02436 (2023)", + "28. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)", + "29. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., MartinBrualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. 40(6) (dec 2021)", + "30. Peng, C., Chellappa, R.: Pdrf: progressively deblurring radiance field for fast scene reconstruction from blurry images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2029-2037 (2023)", + "31. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "32. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961 (2020)", + "33. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 10318-10327. Computer Vision Foundation / IEEE (2021)", + "34. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: Nonlinear Phenomena 60(1), 259-268 (1992)", + "35. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "36. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016)", + "37. Sun, C., et al.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "38. Tai, Y.W., et al.: Richardson-lucy deblurring for scenes under a projective motion path. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(8), 1603-1618 (2011)", + "39. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023)", + "40. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: IEEE International Conference on Computer Vision (ICCV). IEEE (2021)", + "41. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)", + "42. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490 (2022)", + "43. Xu, L., Zheng, S., Jia, J.: Unnatural L0 sparse representation for natural image deblurring. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition," + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BAGS", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Portland, OR, USA, June 23-28, 2013. pp. 1107-1114. IEEE Computer Society (2013)", + "44. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)", + "45. Yu, X., Dai, P., Li, W., Ma, L., Shen, J., Li, J., Qi, X.: Towards efficient and scale-robust ultra-high-definition image demoiring. In: European Conference on Computer Vision. pp. 646-662. Springer (2022)", + "46. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. arXiv preprint arXiv:2311.16493 (2023)", + "47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)", + "48. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371-378 (2001)" + ], + "bbox": [ + 215, + 146, + 785, + 354 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Peng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_model.json b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..74ec741dfd46e115d4f7ebaef4ee30a1c371262b --- /dev/null +++ b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_model.json @@ -0,0 +1,2403 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "BAGS: Blur Agnostic Gaussian Splitting through Multi-Scale Kernel Modeling" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.212, + 0.787, + 0.246 + ], + "angle": 0, + "content": "Cheng Peng*, Yutao Tang*, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.256, + 0.756, + 0.285 + ], + "angle": 0, + "content": "Johns Hopkins University, Baltimore MD 21218, USA {cpeng26, ytang67, yzhou223, nwang43, xliu253, dlil90, rchella4}@jhu.edu" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.318, + 0.744, + 0.61 + ], + "angle": 0, + "content": "Abstract. Recent efforts in using 3D Gaussians for scene reconstruction and novel view synthesis can achieve impressive results on curated benchmarks; however, images captured in real life are often blurry. In this work, we analyze the robustness of Gaussian-Splatting-based methods against various image blur, such as motion blur, defocus blur, downscaling blur, etc. Under these degradations, Gaussian-Splatting-based methods tend to overfit and produce worse results than Neural-Radiance-Field-based methods. To address this issue, we propose Blur Agnostic Gaussian Splatting (BAGS). BAGS introduces additional 2D modeling capacities such that a 3D-consistent and high quality scene can be reconstructed despite image-wise blur. Specifically, we model blur by estimating per-pixel convolution kernels from a Blur Proposal Network (BPN). BPN is designed to consider spatial, color, and depth variations of the scene to maximize modeling capacity. Additionally, BPN also proposes a quality-assessing mask, which indicates regions where blur occur. Finally, we introduce a coarse-to-fine kernel optimization scheme; this optimization scheme is fast and avoids sub-optimal solutions due to a sparse point cloud initialization, which often occurs when we apply Structure-from-Motion on blurry images. We demonstrate that BAGS achieves photorealistic renderings under various challenging blur conditions and imaging geometry, while significantly improving upon existing approaches." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.635, + 0.377, + 0.651 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.817 + ], + "angle": 0, + "content": "High quality scene reconstruction and novel view synthesis from 2D images is a long-standing research problem with extensive applications in robotics, virtual reality, e-commerce, cinematography, etc. Significant advancements in this field have been made in recent years with the introduction of Neural Radiance Field (NeRF) [24]. NeRF can achieve photorealistic view synthesis by implicitly representing the scene with a Multi-Layer Perceptron (MLP) and optimizing the MLP with differentiable ray-tracing. Many subsequent works have since been proposed to improve various aspects of NeRF, including optimization acceleration [6,9,25,37], anti-aliasing [2,3], dynamic and generative modeling [19,31,32,40], etc. Recently, 3D Gaussian Splatting (3DGS) [12] has been" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.374, + 0.841 + ], + "angle": 0, + "content": "*Equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Peng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "introduced as an alternative to NeRF-based approaches. 3DGS uses Gaussians as explicit 3D representations and a differentiable rasterization technique, which enables high quality scene reconstruction in a short time." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.219, + 0.431, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.22, + 0.556, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.219, + 0.788, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Fig. 1: We introduce BAGS, which can reconstruct high quality scenes even from blurry training images. Moreover, BAGS can provide kernels and masks that indicate the types and regions of the blur, as shown by the highlighted regions under Mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.524 + ], + "angle": 0, + "content": "Practical applications of 3DGS, however, can be challenging. Despite its impressive performance on curated datasets, 3DGS requires high quality images and a good point cloud initialization to work well. Without these conditions, 3DGS often generates undesired Gaussians to overfit observation noise, leading to worse renderings, e.g., compared to NeRF. For instance, images acquired in real life or from the internet are often in non-ideal conditions. The camera may be unstable or out of focus, leading to motion or defocus blur. Images may also have different resolutions depending on the sensor types or post processes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.63 + ], + "angle": 0, + "content": "There often exists sufficient information within a multi-view image set to generate a high quality reconstruction, even if the individual images are degraded. This is because the same region is often observed repeatedly from different views; such redundancy can then be explored to better recover the underlying scene. Works like Deblur-NeRF [21] and its variants [15, 16, 30, 41] are designed based on this insight. Formally, a simple yet powerful forward model can be used to describe the potentially blurry pixel value \\(\\tilde{C}(x)\\) at coordinate \\(x\\):" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.636, + 0.787, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\tilde {C} (x) = \\sum_ {x _ {k} \\in \\mathcal {N} (x)} C \\left(x _ {k}\\right) h \\left(x _ {k}\\right) \\text {s . t} \\sum_ {x _ {k} \\in \\mathcal {N} (x)} h \\left(x _ {k}\\right) = 1, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.795 + ], + "angle": 0, + "content": "where \\( C(x_{k}) \\) is the clear pixel value at pixel coordinates \\( x_{k} \\) around the neighborhood of \\( x \\), and \\( h(x_{k}) \\) is the blur kernel weights. In this formulation, \\( C(x_{k}) \\) is multi-view consistent, and \\( \\tilde{C}(x) \\) is multi-view inconsistent due to the blur from \\( h \\). Previous methods [15, 16, 21, 30, 41] aim to simultaneously optimize a multi-view consistent \\( C(x_{k}) \\) through NeRF and the multi-view inconsistent \\( \\tilde{C}(x) \\) through a view-specific network that predicts \\( h(x_{k}) \\). As such, we render only the clear \\( C(x_{k}) \\) after training. In comparison, vanilla NeRF [24] treats \\( C(x_{k}) \\) and \\( \\tilde{C}(x) \\) as one variable, forcing multi-view inconsistencies to be baked into the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this work, we build upon the blur formulation in Eq. (1), and explore two key differences between NeRF and 3DGS in robust reconstruction. NeRF-based approaches suffer from slow training speed, i.e. only limited number of rays can" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "be rendered at once. This is exacerbated by kernel modeling, which requires the rendering of a patch to model a single pixel. To mitigate this issue, current approaches model the kernel \\( h \\) sparsely [15, 21, 30] or only for specific types of blur [16, 41]. Since 3DGS can rasterize entire images efficiently, it potentially has great synergy with Eq. (1). On the flip side, NeRF requires only calibrated camera poses for reconstruction, whereas 3DGS also requires a good point cloud initialization. For noisy images, camera poses can often be estimated based on just a few correct matches, but a dense point cloud is challenging to obtain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.789, + 0.513 + ], + "angle": 0, + "content": "As shown in Fig 1, we introduce Blur Agnostic Gaussian Splatting, or BAGS, which is a Splatting-based method that is robust against various types of blur. At the core of BAGS is a Blur Proposal Network (BPN). Contrary to previous methods [15,21,30] that sparsely estimate \\( h \\), BPN directly estimates dense kernels on full images during training. We find that estimating a dense \\( h \\) is more effective and no less efficient than estimating a sparse one, which also has to predict kernel positions \\( x_{k} \\). BPN considers the position and view embedding of pixel \\( x \\), accounting for the spatial variance in image blur. Furthermore, a three-layer Convolution Neural Network (CNN) is used to compute the RGBD features of the rendered image. These extracted features allow BPN to consider the color and depth variation of the scene, e.g. around edges and corners, more effectively. Finally, BPN also computes a per-pixel mask, which blends the estimated clear and blurry pixels together. We make the observation that not necessarily all pixels are blurry; therefore, by constraining on its sparsity, such a mask can visualize regions where blur occurs. This is particularly useful as a way to evaluate image quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.789, + 0.695 + ], + "angle": 0, + "content": "In addition to BPN, BAGS proposes a coarse-to-fine kernel optimization scheme. In some cases, BAGS cannot perform reconstruction well if it jointly optimizes BPN with a very sparse point cloud, due to the inherent ambiguity of separating blur from the observation. Furthermore, optimizing a dense kernel on full resolution directly is computationally expensive and time-consuming. To solve these conundrums, we first optimize the scene and BPN at lower scales with a smaller kernel dimension. This allows the optimization process to address a less challenging and ambiguous problem initially. We then gradually increase the resolution of the training images and the estimated kernels; since the blur problem has been partially addressed at a small scale, optimization at larger scales is more stable and effective. Under this scheme, we not only observe improved performance but also faster optimization process." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.698, + 0.714, + 0.713 + ], + "angle": 0, + "content": "In summary, our contributions can be summarized in three parts:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.732, + 0.784, + 0.775 + ], + "angle": 0, + "content": "1. We introduce a Blur Proposal Network, which considers spatial, depth, and color variations of the scene to model image blur; BPN can also produce a mask that indicates blurry regions in an input image." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.78, + 0.784, + 0.837 + ], + "angle": 0, + "content": "2. We introduce a coarse-to-fine optimization scheme, which gradually increases the training image resolution and the estimated kernel size with additional neural network layers; this improves the stability of the joint optimization process given a sparse point cloud." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.732, + 0.784, + 0.837 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Peng et al." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "3. We evaluate the overall method, BAGS, on three image blur scenarios and find significant quantitative and visual improvements compared to current SoTA methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.207, + 0.551, + 0.225 + ], + "angle": 0, + "content": "2 Related Work and Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Single Image Restoration. Restoring images from degradation, such as blur, has a long history of research. In general, this is a heavily ill-posed problem as many potential high quality images can lead to the same degraded observations. Traditionally, various priors are used to constrain the solution space [5, 14, 18, 34, 38, 43]. More recently, deep-learning-based methods have achieved great performances by learning a better prior in a data driven way, e.g. as a function that directly maps blurry images to clean images. In this paradigm, we train a powerful restoration model with millions of training pairs and a large neural network. Ideally, images can be restored by such a model [1, 4, 8, 13, 17, 42, 45, 47]. This paradigm has many issues in practice. For example, real training pairs are difficult to acquire in large scale; therefore, synthetic data is often used, which leads to significant domain shift problems. Every type of degradation also requires its own training sets and models, which exacerbate the data and computation costs. Even for degradations such as low resolution, for which is easy to synthesize training pairs, the restoration is unlikely to be multi-view consistent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.684 + ], + "angle": 0, + "content": "NeRF in Non-Ideal Conditions. NeRF [24] works well when images are clean and well-calibrated; however, images acquired in the wild are often less ideal. Factors such as low light, camera motion, object motion, and incorrect focus can degrade image quality or affect multi-view consistency. Many works introduce domain knowledge to model the non-ideal contributors, such that the radiance field only models a canonical, multi-view consistent scene. RawNeRF [23] uses an approximated tonemapped loss and variable exposure to account for low light observations. NeRF-W [22] uses an image-specific embedding to model inconsistent appearances and transient objects. For object motion, many works [28,29,33,40] use an implicit function to describe scene deformation, such that the traced rays are first deformed. For image degradation such as blur, similar ideas have been used [11, 15, 21, 30, 41] to model the forward image degradation process; i.e. the 3D consistent scene and 2D inconsistent degradation will be separated based on the multi-view photometric loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.751 + ], + "angle": 0, + "content": "3D Gaussian Splatting. Recently, 3DGS [12] introduces a novel, point-based scene reconstruction method that improves upon NeRF in many aspects. 3DGS utilizes a set of 3D Gaussian primitives [48] \\(\\{\\mathcal{G}_n|n = 1,\\dots ,N\\}\\) to represent the scene. These Gaussians are parameterized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.755, + 0.785, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {n} (\\mathbf {v}) = e ^ {- \\frac {1}{2} (\\mathbf {v} - \\mathbf {p} _ {n}) ^ {T} \\Sigma_ {n} ^ {- 1} (\\mathbf {v} - \\mathbf {p} _ {n})}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where the center \\(\\mathbf{p}_n\\in \\mathbb{R}^{3\\times 1}\\) and covariance matrix \\(\\Sigma_{n}\\in \\mathbb{R}^{3\\times 3}\\) define the position, scale, and rotation of the Gaussian at point \\(\\mathbf{v}\\in \\mathbb{R}^{3\\times 1}\\) in space. Additionally, each Gaussian also contains opacity \\(\\alpha\\) and color \\(c\\). For rendering, the Gaussians are projected onto the image plane as \\(\\mathcal{G}_n^{\\mathrm{2D}}\\), where the center and covariance ma" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.174 + ], + "angle": 0, + "content": "trices are transformed based on the camera rotation \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\) and translation \\(\\mathbf{t} \\in \\mathbb{R}^{3 \\times 1}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.171, + 0.785, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {n} ^ {\\mathrm {2 D}} = \\mathbf {R} \\mathbf {p} _ {n} + \\mathbf {t}, \\Sigma_ {n} ^ {\\mathrm {2 D}} = \\mathbf {R} \\Sigma_ {n} \\mathbf {R} ^ {T}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.784, + 0.223 + ], + "angle": 0, + "content": "We can then aggregate 2D Gaussians according to their depth order \\( D \\) over a pixel/tile, based on alpha blending:" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.232, + 0.786, + 0.272 + ], + "angle": 0, + "content": "\\[\nC (x) = \\sum_ {i = 1} ^ {D} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i} \\mathcal {G} _ {i} ^ {\\mathrm {2 D}}, T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} \\mathcal {G} _ {j} ^ {\\mathrm {2 D}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.784, + 0.337 + ], + "angle": 0, + "content": "With well-designed tiling, this splatting-based rasterization technique is significantly faster than ray-tracing in NeRF-based methods without affecting rendering quality; various work have ensued to further improve the Gaussian representation [10, 26, 27, 44] and the rasterization process [20, 46]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.578, + 0.369 + ], + "angle": 0, + "content": "3 Blur Agnostic Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.785, + 0.47 + ], + "angle": 0, + "content": "BAGS is designed to robustly optimize 3D Gaussians by introducing additional modeling capacities in 2D, which allows BAGS to model away 3D-inconsistent blur from the scene. As shown in Fig. 2, BAGS consists of two parts: a Blur Proposal Network (BPN) and a coarse-to-fine optimization scheme. For clarity, we first describe the design of BPN on a single scale, and then describe modifications on BPN and the training loss on scale \\( s \\) for multi-scale training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.455, + 0.499 + ], + "angle": 0, + "content": "3.1 Blur Proposal Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.784, + 0.587 + ], + "angle": 0, + "content": "Following the forward blur model described in Eq. (1), BPN models a per-pixel convolution kernel \\( h(x) \\in \\mathbb{R}^{K \\times K} \\) for every pixel \\( x \\), where \\( K \\) is the kernel size. The rasterization efficiency in Gaussian Splatting allows us to model \\( h(x) \\) as a full convolution kernel. Additionally, BPN estimates a per-pixel scalar mask \\( m(x) \\in [0, 1] \\), controlling the areas where blur modeling takes place." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.786, + 0.772 + ], + "angle": 0, + "content": "Multi-Modal Features. In scene reconstruction, a large neural network will impact training time. As such, we use a four-layer MLP \\(\\mathcal{F}_{\\mathrm{kernel}}\\) to model blur. To maximize the capacity of \\(\\mathcal{F}_{\\mathrm{kernel}}\\), we design Multi-Modal Features (MMF) that can help distinguish the scene from blur. Real image blur typically varies between different observed images and different pixels within an image; therefore, we provide \\(\\mathcal{F}_{\\mathrm{kernel}}\\) with a learnable view embedding \\(l(i)\\) and a positional embedding \\(p(x)\\) for training view \\(i\\) and coordinate \\(x\\). Additionally, we provide \\(\\mathcal{F}_{\\mathrm{kernel}}\\) with a set of features \\(f_{\\mathrm{RGBD}}\\) extracted from the rendered image \\(C\\) and depth \\(D\\) from rasterization, using a small, three-layer CNN \\(\\mathcal{F}_{\\mathrm{feat}}\\). These features are concatenated with \\(l(i)\\) and \\(p(x)\\) to form the MMF for \\(\\mathcal{F}_{\\mathrm{kernel}}\\). Consequently, the forward process to estimate the kernel \\(h(x,i)\\) and the mask \\(m(x,i)\\) given a specific pixel \\(x\\) and training view \\(i\\) can be noted as:" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.775, + 0.684, + 0.791 + ], + "angle": 0, + "content": "\\[\nh (x, i), m (x, i) = \\mathcal {F} _ {\\mathrm {k e r n e l}} (l (i) \\oplus p (x) \\oplus f _ {\\mathrm {R G B D}} (x, i)),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.789, + 0.703, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} f _ {\\mathrm {R G B D}} (x, i) = \\mathcal {F} _ {\\text {f e a t}} (C (x, i) \\oplus D (x, i)), 0 \\leq m \\leq 1.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Incorporating RGBD features enbales BPN to explore the patch statistics around the current pixel, especially at edges and corners where blur is prominent. Scene" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Peng et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.789, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.789, + 0.462 + ], + "angle": 0, + "content": "Fig. 2: BAGS is optimized by training a Blur Proposal Network on top of the scene \\(\\mathcal{G}(\\mathbf{v})\\) over multiple scales. Top: We extract \\(f_{\\mathrm{RGBD}}^s\\) from color and depth, which are concatenated with the position and view embedding \\(p(x), l(i)\\) to form the Multi-Modal Feature (MMF). The kernel MLP then estimates the per-pixel kernel \\(h^s\\) and mask \\(m^s\\). We use \\(h^s\\) to model the blur image \\(\\tilde{C}^s\\) and employ \\(m^s\\) to blend the rendered image \\(C^s\\) and the blur-modeled image \\(\\tilde{C}^s\\), yielding \\(C_{\\mathrm{out}}^s\\). Bottom: After \\(N_s\\) steps, we upscale image resolution and modify the kernel MLP to produce \\(h^{s - 1}\\) with a larger kernel size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.789, + 0.583 + ], + "angle": 0, + "content": "depth also strongly correlates with the magnitude of blur. For example, defocus blur affects pixels that are outside of a specific focus plane; scene content that is further away from this focus plane becomes more blurry. For camera motion blur, pixels in the near plane may shift more than pixels in the far plane. We find the inclusion of \\( f_{\\mathrm{RGBD}} \\) notably improves BPN's modeling capacity and the quality of novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.789, + 0.655 + ], + "angle": 0, + "content": "Residual Mask Bottleneck. As described in Eq. (5), \\(\\mathcal{F}_{\\mathrm{kern}}\\) also estimates a per-pixel scalar mask \\(m(x,i)\\). Specifically, this mask is used to blend the Gaussian-splatted image \\(C\\) with the blur-modeled image \\(\\tilde{C}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.664, + 0.787, + 0.683 + ], + "angle": 0, + "content": "\\[\nC _ {\\mathrm {o u t}} (x, i) = \\left(1 - m (x, i)\\right) C (x, i) + m (x, i) \\tilde {C} (x, i), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.789, + 0.772 + ], + "angle": 0, + "content": "where \\( C_{\\mathrm{out}} \\) is the final output at training. We note that not necessarily all pixels are blurry in an observed image. The blending \\( m \\) provides an additional degree of freedom to allow \\( C \\), the clear pixel, to be directly compared with \\( C_{\\mathrm{obs}} \\); i.e., if \\( C_{\\mathrm{obs}}(x,i) \\) suffers from blur, then \\( m(x,i) \\) should be large, giving a higher weight to \\( \\hat{C}(x,i) \\); if \\( C_{\\mathrm{obs}}(x,i) \\) if free of blur, \\( m(x,i) \\) should be small, favoring \\( C(x,i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.843 + ], + "angle": 0, + "content": "This design is similar to residual learning by providing the output with a skip connection from \\( C \\) as initialization. We apply a small sparsity constraint on \\( m \\) to encourage a larger weight towards \\( C \\) by default. After training, we can extract \\( m \\) to automatically identify the degraded regions in an input image." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.36, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.22, + 0.341, + 0.231 + ], + "angle": 0, + "content": "(a) Without C2F" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.144, + 0.501, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.22, + 0.464, + 0.231 + ], + "angle": 0, + "content": "(b) BAGS" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.645, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.22, + 0.64, + 0.231 + ], + "angle": 0, + "content": "(c) Naive densification" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.144, + 0.787, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.22, + 0.747, + 0.231 + ], + "angle": 0, + "content": "(d) BAGS" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.234, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Fig. 3: Gaussians may get stuck at a local minimum without proper initialization. As indicated in 3a, optimizing \\( h \\) directly can lead to noisy surfaces. Naively densifying the scene before adding \\( h \\) can also lead to noisy Gaussians; as shown in 3c, the noisy Gasussians are not well removed even after adding \\( h \\). By using a coarse-to-fine training schedule, we achieve better results in 3b and 3d." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.56, + 0.339 + ], + "angle": 0, + "content": "3.2 Coarse-to-Fine Kernel Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.579 + ], + "angle": 0, + "content": "A high quality 3DGS reconstruction requires a good point cloud estimation from Structure from Motion (SfM). Camera calibration parameters, which can be estimated from just a few matches, are relatively robust against image blur. On the other hand, obtaining a sufficiently dense point cloud can be challenging. Unfortunately, 3DGS is prone to over-fitting without a good initialization. This problem can worsen when a blur model is introduced and the kernels \\( h \\) provide additional degrees of freedom in scene optimization, as shown in Fig. 3. Furthermore, estimating dense per-pixel kernels on a full resolution image can be computationally expensive. A sparse estimation of the kernel does not meaningfully improve efficiency, as rendering is cheap in rasterization while the neural network to predict any type of kernel is still costly. Fundamentally, the complexity to compute per-pixel kernel scales quadratically with resolution. To address these problems, we propose a coarse-to-fine kernel optimization scheme, which initially optimizes Eq. (5) at a low resolution and gradually increases render resolution, achieving a more stable training process and better efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Multi-Scale BPN. For multi-scale optimization, BPN should generate kernels with a consistent field of view. We define scale based on the downscaling ratio \\( 2^{s - 1} \\) in resolution, where the original resolution is defined as \\( s = 1 \\). Additionally, if the image resolution is downscaled by a factor of two, i.e. \\( s = 2 \\), the proposed kernel size \\( K^s \\) should also be similarly downscaled. As shown in Fig. 2, a proper \\( K^s \\) ensures that the kernel is always modeling blur from the same regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.787 + ], + "angle": 0, + "content": "To enable multi-scale training, we separate \\(\\mathcal{F}_{\\mathrm{kernel}}^s\\) into two parts: a two-layer base MLP \\(\\mathcal{F}_{\\mathrm{base}}^s\\) and two one-layer MLP heads \\(\\mathcal{F}_{\\mathrm{blur}}^s\\) and \\(\\mathcal{F}_{\\mathrm{mask}}^s\\). The base MLP extracts features from MMF, such features are then provided to the two heads for generating the kernels \\(h^s\\) and masks \\(m^s\\) at scale \\(s\\). In practice, only \\(\\mathcal{F}_{\\mathrm{blur}}^s\\) has a scale-dependent parameter size corresponding to the kernel size \\(K^s\\); therefore, a new MLP layer with output dimension \\(K^s \\times K^s\\) is added for every scale. The updated multi-scale BPN can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.789, + 0.662, + 0.806 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {i n t e r}} (x, i) = \\mathcal {F} _ {\\text {b a s e}} ^ {s} (l (i) \\oplus p (x) \\oplus f _ {\\text {R G B D}} (x, i))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.808, + 0.786, + 0.825 + ], + "angle": 0, + "content": "\\[\nh ^ {s} (x, i) = \\operatorname {s o f t m a x} \\left(\\mathcal {F} _ {\\text {b l u r}} ^ {s} \\left(f _ {\\text {i n t e r}} (x, i)\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.827, + 0.641, + 0.843 + ], + "angle": 0, + "content": "\\[\nm ^ {s} (x, i) = \\operatorname {s i g m o i d} \\left(\\mathcal {F} _ {\\text {m a s k}} ^ {s} \\left(f _ {\\text {i n t e r}} (x, i)\\right)\\right).\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Peng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "Coarse-to-Fine Optimization. We split training into multiple stages based on scale. Starting from the coarsest scale \\( s = S \\) to the finest scale \\( s = 1 \\), we optimize BAGS with a downsampling factor of \\( 2^{s - 1} \\) applied to \\( C, \\tilde{C}, C_{\\mathrm{obs}} \\), and \\( h \\). The optimization for each scale lasts for \\( N_{s} \\) steps, we then increase the resolution in \\( C, \\tilde{C}, C_{\\mathrm{obs}} \\), and \\( h \\). The scale-dependent training loss can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.224, + 0.786, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\lambda_ {\\mathrm {p h o t o}} \\| C _ {\\mathrm {o u t}} ^ {s} - C _ {\\mathrm {o b s}} ^ {s} \\| + \\lambda_ {\\mathrm {D S}} \\mathcal {L} _ {\\mathrm {D - S S I M}} \\left(C _ {\\mathrm {o u t}} ^ {s}, C _ {\\mathrm {o b s}} ^ {s}\\right) + \\lambda_ {\\mathrm {m a s k}} \\| m ^ {s} \\|, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.349 + ], + "angle": 0, + "content": "where \\( C_{\\mathrm{obs}}^s \\) and \\( C_{\\mathrm{out}}^s \\) are the observed and predicted image at scale \\( s \\), \\( \\mathcal{L}_{\\mathrm{D - SSIM}} \\) is the structural similarity loss; \\( \\| m^{s}\\| \\) is the mask sparsity loss. We find this training scheme achieves the best of both worlds: activating BPN at low resolution generates less overfitting artifacts during densifying the point cloud given the reduced visibility of blur. Moreover, it also serves as an efficient warm-up for BPN at higher resolution, as computing per-pixel kernels at low resolution is much cheaper." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.362, + 0.376, + 0.379 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.466 + ], + "angle": 0, + "content": "We evaluate BAGS under three degradation scenarios: camera motion blur, defocus blur, and mix resolution. All of these degradations are commonly observed in images acquired in real life or online. We further introduce a new unbounded 360 drone dataset, which is collected at sunset or night and experiences degradation due to a combination of motion blur and low light condition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Camera Motion and Defocus Blur. Camera motion and defocus blur are often observed in real acquisitions, e.g. due to low light, incorrect auto-focus, or limitation in the finite aperture of real cameras. Deblur-NeRF [21] provides real scene acquisitions that contain camera motion and defocus blur; both blur categories consist of ten scenes. While real world scenes have blur-free references, the camera settings, e.g., exposure, may not be consistent with source views. Experiments on synthetic blur are reported in supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Mix Resolution. To analyze the scenario where the input images are of inconsistent resolution, we apply various downsampling operations on the Mip-NeRF 360 dataset [2]. Specifically, we split the training images into four equal parts and apply 4X, 3X, 2X, and no downscaling to each of the parts. The image splits are sampled in uniform randomness to ensure no downscaling covers a consecutive interval of the camera trajectory. We then use NeRFStudio's data processing pipeline [39], which converts all images to the highest resolution and calibrates with COLMAP [35]. Similar to previous works [2,3,46], we perform the entire process on downscaled data instead of the original 4K resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Unbounded Drone Motion Blur. The dataset introduced by DeblurNeRF [21] serves as a good benchmark for performance comparison; however, it is collected in a forward bounded format. These forward bounded scenes focus on a small area and have significant overlaps within the images; as such, it is unclear if the methodology [15, 16, 21, 30, 41] based on these sets can generalize to other scenarios. In fact, none of the deblurring methods can work for unbounded scenes. To supplement DeblurNeRF's dataset and understand how BAGS performs in realistic settings, we collect five sets of drone footage on different buildings in 360" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.688, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.499, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.145, + 0.777, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.25, + 0.499, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.359, + 0.373, + 0.37 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.25, + 0.777, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.359, + 0.651, + 0.37 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.373, + 0.785, + 0.442 + ], + "angle": 0, + "content": "Fig. 4: Ablation study of BAGS with different sub-modules. In 4a and 4b, introducing RGBD features and coarse-to-fine optimization improve novel view synthesis quality, and alternative sparse deformable kernel suffers in performance; 4d demonstrates our speed improvement against \\( \\mathrm{BAGS}^{\\mathrm{noC2F}} \\). In 4c, larger kernel generally leads to better performance, other than in the mix resolution scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.785, + 0.532 + ], + "angle": 0, + "content": "unbounded format. These footages are collected at sunset or night, and suffer from camera motion blur and low light noise. For each scene, we also capture clear frames from a stable drone position for testing, resulting in around 100 images overall. We use NeRFStudio [39] to calibrate images from these footages. More details on this dataset can be found in supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.658 + ], + "angle": 0, + "content": "Implementation Details. BPN's \\(\\mathcal{F}_{\\mathrm{feat}}\\) is constructed as a three-layer CNN; the intermediate channel size is 64, the kernel size for each layer is 5, 5, and 3. \\(\\mathcal{F}_{\\mathrm{kernel}}\\) is constructed with four linear layers; the intermediate channel size is 64. For multi-scale training, we use three scales starting with scale \\(s = 3\\), where \\(K_{3} = 5\\), \\(K_{2} = 9\\), and \\(K_{1} = 17\\). BAGS implements its Gaussian Splatting and rasterization based on Mip-Splatting [46], which has been shown to yield good results across multiple resolutions. More implementation details can be found in supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.402, + 0.685 + ], + "angle": 0, + "content": "4.1 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.784, + 0.728 + ], + "angle": 0, + "content": "We compare three designs in BAGS to demonstrate the effectiveness of different components:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.73, + 0.758, + 0.746 + ], + "angle": 0, + "content": "- BAGS\\(^{\\text{noRGBD}}\\): We implement BPN without the RGBD features \\( f_{\\text{RGBD}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.746, + 0.784, + 0.775 + ], + "angle": 0, + "content": "- BAGS\\(^{\\text{deform}}\\): We implement BPN with a sparse deformable kernel, similar to DeblurNeRF [21]." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.776, + 0.784, + 0.804 + ], + "angle": 0, + "content": "- BAGS\\(^{\\mathrm{noC2F}}\\): We implement BAGS without a coarse-to-fine kernel optimization scheme." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.73, + 0.784, + 0.804 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We further investigate the effect of the final kernel size \\( K_{1} \\) on reconstruction quality across all degradation scenarios." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Peng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.418 + ], + "angle": 0, + "content": "As shown in Fig. 4, the introduction of RGBD features to BPN greatly improves its ability to disentangle image blur from the underlying scene. The addition of \\( f_{\\mathrm{RGBD}} \\) increases the average performance by \\( 1.57 / 0.92 \\) dB in PSNR and \\( 0.03 / 0.03 \\) in SSIM for camera motion and defocus blur, respectively. This agrees with our intuition that blur is highly correlated with depth and image features like edges. Our coarse-to-fine optimization scheme also notably boosts robustness against defocus blur by \\( 1.64 \\) dB and \\( 0.08 \\) in PSNR and SSIM on average. As visualized in Fig. 3a, directly optimizing 3DGS on a sparse point cloud in \\( \\mathrm{BAGS}^{\\mathrm{noC2F}} \\) leaves many ambiguities; while training views have low loss, test views show a lack of the correct surface structure and many artifact Gaussians. Densifying Gaussians without BPN first improves performance for defocus blur; however, this strategy is bad for motion blur, as shown in Fig. 3c. In Fig. 4d, we show that our coarse-to-fine strategy is not only performant, but also fast; on average, BAGS can reach high performance in less than fifteen minutes, which is much faster than the two hours required by PDRF [30]. This strategy is also faster compared to single-scale training, as full-scale training in our scheme uses less iterations, and lower scale training does not sacrifice performance. Both multi-scale and single-scale training take \\( 40 \\mathrm{K} \\) iterations overall." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.63 + ], + "angle": 0, + "content": "We also implement a deformable version of BPN in \\( \\mathrm{BAGS}^{\\mathrm{deform}} \\), where BPN estimates the kernel position and kernel weights for ten points within a \\( 20 \\times 20 \\)-pixel patch. We show its performances in Fig. 4a, where \\( \\mathrm{BAGS}^{\\mathrm{deform}} \\) is tested on camera motion blur. Camera motion blur often has a sparse representation, i.e. the camera motion trajectory [16, 41]; therefore, a ten-point estimation of the kernel may perform well. However, we find that \\( \\mathrm{BAGS}^{\\mathrm{deform}} \\) still underperforms compared to a full kernel estimation in BAGS. This may be due to the difference between NeRF and 3DGS; i.e., a ray-based deformation can be better expressed in NeRF, while a patch-based convolution can be better expressed in 3DGS. We also note that \\( \\mathrm{BAGS}^{\\mathrm{deform}} \\) is not more efficient than BAGS despite its sparse estimation. This is because approaches like [15, 21, 30] require one forward pass to estimate one kernel position and weight; therefore, we need to forward multiple times for every pixel to estimate the sparse kernel. In comparison, BAGS only requires a single forward pass to obtain the full kernel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Finally, we ablate on the kernel dimension estimated by BPN. In Fig. 4c, we show that larger kernels in general lead to higher performances; interestingly, this is not true in the mixed resolution scenario. This is likely because downscaling is uniform across all pixels in an image, i.e. a downscaling factor of four means a kernel size of five can fully explain the degradation. As such, increasing the kernel size does not improve performance and introduces noise instead." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.735, + 0.461, + 0.751 + ], + "angle": 0, + "content": "4.2 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Camera Motion and Defocus Blur. As shown in Table 1, we observe that Gaussian Splatting [46] performs markedly worse than vanilla NeRF in the presence of blurry images, despite its strong performance in other curated benchmarks. Since NeRF models a continuous neural volume with an MLP, individual degradation can be smoothly explained by spreading the error into empty space." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.688, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.783, + 0.387 + ], + "angle": 0, + "content": "
Camera MotionNeRF [24]Mip-Splatting [46]Deblur-NeRF [21]DP-NeRF [15]PDRF [30]BAGS
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Ball24.080.6240.39923.220.6190.34027.360.7660.22327.200.7650.209
Basket23.720.7090.32223.240.6880.28827.670.8450.14827.740.8460.129
Buick21.590.6330.35021.460.6580.26624.770.7700.17525.700.7920.141
Coffee26.480.8060.29024.730.7490.28830.930.8980.12432.440.9150.101
Decoration22.390.6610.36320.550.6410.29924.190.7710.18623.510.7400.275
Girl20.070.7080.32019.870.7140.27822.270.7980.16921.800.7700.205
Heron20.500.5220.41319.430.5050.33222.630.6870.21022.520.6760.281
Parterre23.140.6200.40522.280.5900.32125.820.7600.21624.970.7680.227
Puppet22.090.6090.33922.050.6310.26725.240.7510.15825.120.7580.181
Stair22.870.4560.48721.910.4740.38725.390.6300.21025.780.6470.229
Average22.690.6350.36921.870.6270.30625.630.7670.18225.910.7750.160
Defocus
Cake24.420.7210.22522.170.6450.29026.270.7800.12826.160.7780.127
Caps22.730.6310.28021.000.5030.44223.870.7130.16123.950.7120.143
Cisco20.720.7220.12620.080.7090.16320.830.7270.08720.730.7260.084
Coral19.810.5660.26919.600.5550.32319.850.6000.12122.800.7410.096
Cupcake21.880.6810.21621.550.6810.21322.260.7220.11620.110.6110.118
Cups25.020.7580.23220.930.6460.31426.210.7990.12726.750.8140.104
Daisy22.740.6200.26221.590.6140.27323.520.6870.12123.790.6970.108
Sausage17.790.4830.27917.780.4730.28418.010.5000.18018.350.5440.147
Seal22.790.6270.26822.130.5920.31026.040.7770.10525.950.7780.103
Tools26.080.8520.15523.980.8120.19627.810.8950.06128.070.8980.054
Average22.400.6660.23121.080.6230.28123.470.7200.12123.670.7300.108
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.39, + 0.784, + 0.417 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons on the camera motion and defocus blur dataset. We color code the best PSNR, SSIM, and LPIPS performances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.785, + 0.558 + ], + "angle": 0, + "content": "On the other hand, 3DGS is based on a discrete representation and tends to produce concentrated artifacts. Our BAGS can achieve new SoTA performances compared to previous NeRF-based deblurring methods despite the unfavorable performances in 3DGS. Specifically, we find significant improvements on camera motion blur across all metrics and achieve a fifty percent improvement in LPIPS on average. As visualized in Fig. 5, BAGS achieves much sharper results compared to previous methods [15, 16, 21, 30], to the point where it is hard to distinguish the groundtruth image from our rendered results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.785, + 0.666 + ], + "angle": 0, + "content": "While improvements on defocus blur is less drastic in Table 1, we still observe great visual improvements. Several reasons contribute to the less apparent quantitative improvements. First, the collected groundtruth images have different exposure compared to training images; as such, image-wise similarity metrics like PSNR are less reliable. Secondly, many scenes contain specular surfaces, which all methods do a bad job at reconstructing. As shown in Fig. 5, much better details can be recovered through BAGS than other methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Mix Resolution and Low Light Motion Blur. We investigate the robustness of several SoTA reconstruction methods on downscaling perturbation and motion blur observed in low light condition. The experiments are done on unbounded scenes and are more challenging and realistic than the forward bounded scenes in DeblurNeRF [21]. We note that previous approaches [15, 21, 30, 41] cannot handle unbounded scenes; therefore, we do not include them in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "As shown in Table 2, Mip Splating performs on par with Mip-NeRF 360; both are vastly better than NeRFacto [39], which stores hash features [25] to accelerate the training process and uses a very small MLP represent the scene. We also employ a Super-Resolution (SR) model HAT [7] on downscaled images in the hope that single-image SR can improve performances. However, pre-trained" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Peng et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.784, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.471, + 0.785, + 0.501 + ], + "angle": 0, + "content": "Fig. 5: Visualizations of test views on camera motion and defocus blur dataset. Mip-Sp and Db-NeRF are short for Mip-Splatting [46] and Deblur-NeRF [21]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.784, + 0.605 + ], + "angle": 0, + "content": "SR models are limited by domain shifts, dedicated upscaling factors, and the lack of multi-view consistency. As visualized in Fig. 6, while the render is somewhat sharper after SR, the color becomes unfaithful. Compared with all other methods, our rendered results are much sharper in details. This robust performance comes from BPN's ability to model Gaussian blur on high resolution before comparing them with the downscaled observations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.785, + 0.759 + ], + "angle": 0, + "content": "Motion blur is often observed in low light condition. To this end, we perform experiments on scenes acquired near sunset or at night. As shown in Table 2, BAGS again improves significantly over other reconstruction approaches, which shows that our method can be generalized to different imaging geometry. As shown in Fig. 6, we employ MPR [47], a single image deblurring model, on the motion-blurred images before reconstructing with Mip-Splatting [46]. Similar to previous attempts, the deblurring prior leads to a slightly clearer rendering but is not sufficient to remove all blur. BAGS reconstructs the scene remarkably well compared to other methods, where even the window details can be recovered correctly. Please refer to the supplemental material for more visualizations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Kernel and Mask Visualization. BAGS models image degradation through convolution kernels \\( h \\) and masks \\( m \\), which we visualize in Fig. 7. By constraining on its sparsity, we can see that \\( m \\) meaningfully highlights the regions of blur in a given training image. On the other hand, low-value regions indicate either high quality observation or textureless regions. By visualizing the kernels modeled" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.688, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.384 + ], + "angle": 0, + "content": "
MixedResNeRFacto [39]Mip-NeRF 360 [2]Mip-Sp [46]Mip-Sp-HAT [7]BAGS
PSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPS
Bicycle19.250.3490.49623.820.6100.37524.030.6280.36423.20
Bonsai21.800.6190.25929.910.8710.19029.270.8760.19328.38
Counter24.070.7140.28028.090.8480.22027.790.8530.21627.02
Flowers19.630.4180.45921.980.5660.38921.960.5700.39720.81
Garden21.610.4890.35725.340.6670.31725.280.6820.30523.96
Kitchen22.730.5840.30427.820.7790.23827.490.7780.24126.10
Room24.680.7910.21430.380.9020.19830.430.9030.20329.61
Stump23.190.5550.37725.750.6880.32025.870.7030.29924.09
Treehill20.110.4160.50122.520.5580.43722.860.5500.45721.76
Average21.900.5480.36126.180.7210.29826.110.7270.29724.99
Lowlight-droneNeRFacto [39]Mip-NeRF 360 [2]Mip-Sp [46]Mip-Sp-MPR [47]BAGS
PSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPS
Math-night21.330.5820.32629.510.8820.19129.610.8840.19929.15
Library-night24.480.7680.38126.380.8210.34926.170.8200.37025.89
BioMed-sunset23.760.6500.43924.480.6880.40224.850.7120.42624.78
Admin-sunset21.960.6210.40025.470.7570.35525.540.7640.36925.19
Aud-sunset24.670.7090.39825.900.7500.36826.330.7700.38626.77
Average23.240.6660.38926.350.7800.33326.500.7900.35026.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.387, + 0.785, + 0.43 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons on mixed resolution and low light motion blur in unbounded 360 geometry. Mip-Sp is short for Mip-Splatting [46]. Mip-Sp-HAT and Mip-Sp-HAT-MPR use prior models to restore images first before Mip-Splatting." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.467, + 0.784, + 0.713 + ], + "angle": 0, + "content": "
Novel ViewNeRFactoMipNeRF-360Mip-SpMip-Sp-HATOursG.T.
PSNR/SSIM21.61/.57624.09/.67025.11/.74624.41/.72228.46/.899MixedRes
PSNR/SSIM23.08/.41224.98/.56224.93/.59024.68/.58731.79/.949MixedRes
Novel ViewNeRFactoMipNeRF-360Mip-SpMip-Sp-MPROursG.T.
PSNR/SSIM17.86/.60818.33/.70418.57/.71418.51/.71824.47/.850LowLight
PSNR/SSIM22.40/.68325.36/.81325.51/.81727.03/.86731.75/.938LowLight
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.726, + 0.784, + 0.74 + ], + "angle": 0, + "content": "Fig. 6: Visualizations of test views on mix resolution and low light motion blur dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "at different pixel, we can also easily characterize the types of observed blur. Specifically, the estimated kernels in camera motion blur exhibit clear patterns of camera movement, while those for defocus blur show Gaussian-like distributions based on the pixel's distance from the focus plane. These self-emerged properties" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Peng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "in BAGS provide ways for us to automatically and precisely evaluate the quality of training images." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.195, + 0.36, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.272, + 0.335, + 0.282 + ], + "angle": 0, + "content": "Camera Motion" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.196, + 0.501, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.271, + 0.476, + 0.282 + ], + "angle": 0, + "content": "Kernel & mask" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.196, + 0.645, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.271, + 0.598, + 0.282 + ], + "angle": 0, + "content": "Defocus" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.196, + 0.787, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.271, + 0.759, + 0.282 + ], + "angle": 0, + "content": "Kernel & mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.29, + 0.773, + 0.305 + ], + "angle": 0, + "content": "Fig. 7: Visualization of estimated convolution kernels and masks in different blur." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Discussion. BAGS can achieve great results; however, there also exists many potential improvements. The added neural networks and convolution operations require additional computation. While BAGS is a generalist approach and addresses multiple types of blur, it may become very expensive to estimate per-pixel convolution in high resolution. To this end, we have explored utilizing pixel shuffling [36], which converts spatial resolution into channels, before we estimate blur kernel; i.e., this assumes that within a small region, blur kernel \\( h \\) stays roughly consistent. As shown in the supplemental material, we can scale to great visual results at 2K resolution on the DeblurNeRF dataset. Many other potential directions can also address this complexity issue, including leveraging the mask \\( m \\) to identify important regions to focus on, a better designed low rank kernel estimation, degradation-specific optimization, etc. As such, while BAGS is already more efficient than comparable NeRF-based methods, we anticipate that its computation cost can be further optimized in the future." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.359, + 0.569 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We present BAGS, a novel scene reconstruction method that can handle various noise in input images. We demonstrate that a vanilla Gaussian Splatting is particularly susceptible to degraded images due to its discrete representation, even compared with NeRF. To address this issue, BAGS implements a 2D degradation model, BPN, which estimates convolution kernels jointly with 3D scene optimization. BPN allows for additional freedom to address 3D inconsistencies, and produces interpretable kernels and masks to indicate the degradation types and regions within an image. Additionally, BAGS leverages a coarse-to-fine kernel optimization scheme, which gradually models the 3D scene with higher resolution images and kernels. This approach successfully addresses the 2D-3D ambiguity in optimizing convolution kernels with a very sparse point cloud. We perform extensive experiments on a variety of noise in training images, including camera motion blur, defocus blur, and low resolution. Under these conditions, BAGS can render high fidelity scenes and significantly improves upon previous approaches. In the future, we hope to further improve the computational efficiency in BAGS and dynamically adjust kernel capacity based on how degraded the training views are." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.332 + ], + "angle": 0, + "content": "This research is based upon work supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. The authors also thank the support and feedbacks from researchers, Kaiwen Jiang, Dr. Ravi Ramamoorthi, at UCSD." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.357, + 0.323, + 0.373 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.787, + 0.446 + ], + "angle": 0, + "content": "1. Abuolaim, A., Afifi, M., Brown, M.S.: Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1231-1239 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.787, + 0.475 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.477, + 0.787, + 0.517 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.787, + 0.572 + ], + "angle": 0, + "content": "4. Cao, Y., Liu, M., Liu, S., Wang, X., Lei, L., Zuo, W.: Physics-guided iso-dependent sensor noise modeling for extreme low-light photography. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5744-5753 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.787, + 0.602 + ], + "angle": 0, + "content": "5. Chan, T.F., Wong, C.: Total variation blind deconvolution. IEEE Trans. Image Process. 7(3), 370-375 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.603, + 0.787, + 0.629 + ], + "angle": 0, + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.787, + 0.672 + ], + "angle": 0, + "content": "7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22367-22377 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.713 + ], + "angle": 0, + "content": "8. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4641-4650 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.787, + 0.756 + ], + "angle": 0, + "content": "9. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.786, + 0.799 + ], + "angle": 0, + "content": "10. Gao, J., Gu, C., Lin, Y., Zhu, H., Cao, X., Zhang, L., Yao, Y.: Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "11. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-nerf: High dynamic range neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18398-18408 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.391, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Peng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "12. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.178, + 0.785, + 0.218 + ], + "angle": 0, + "content": "13. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5886-5895 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.785, + 0.274 + ], + "angle": 0, + "content": "14. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011. pp. 233-240. IEEE Computer Society (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.277, + 0.785, + 0.317 + ], + "angle": 0, + "content": "15. Lee, D., Lee, M., Shin, C., Lee, S.: Dp-nerf: Deblurred neural radiance field with physical scene priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12386-12396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.319, + 0.785, + 0.359 + ], + "angle": 0, + "content": "16. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.362, + 0.785, + 0.402 + ], + "angle": 0, + "content": "17. Lee, J., Son, H., Rim, J., Cho, S., Lee, S.: Iterative filter adaptive network for single image defocus deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2034-2042 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.404, + 0.785, + 0.471 + ], + "angle": 0, + "content": "18. Levin, A.: Blind motion deblurring using image statistics. In: Scholkopf, B., Platt, J.C., Hofmann, T. (eds.) Advances in Neural Information Processing Systems 19, Proceedings of the Twentieth Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada, December 4-7, 2006. pp. 841-848. MIT Press (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "19. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.517, + 0.785, + 0.557 + ], + "angle": 0, + "content": "20. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.56, + 0.785, + 0.6 + ], + "angle": 0, + "content": "21. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.602, + 0.785, + 0.669 + ], + "angle": 0, + "content": "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7210-7219. Computer Vision Foundation / IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.672, + 0.785, + 0.727 + ], + "angle": 0, + "content": "23. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16190–16199 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.729, + 0.785, + 0.797 + ], + "angle": 0, + "content": "24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I. Lecture Notes in Computer Science, vol. 12346, pp. 405-421. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "26. Navaneet, K., Meibodi, K.P., Koohpayegani, S.A., Piriyawash, H.: Compact3d: Compressing gaussian splat radiance field models with vector quantization. arXiv preprint arXiv:2311.18159 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "27. Niedermayr, S., Stumpfegger, J., Westermann, R.: Compressed 3d gaussian splatt-ting for accelerated novel view synthesis. arXiv preprint arXiv:2401.02436 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "28. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "29. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., MartinBrualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. 40(6) (dec 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "30. Peng, C., Chellappa, R.: Pdrf: progressively deblurring radiance field for fast scene reconstruction from blurry images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2029-2037 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.355 + ], + "angle": 0, + "content": "31. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.383 + ], + "angle": 0, + "content": "32. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.438 + ], + "angle": 0, + "content": "33. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 10318-10327. Computer Vision Foundation / IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.466 + ], + "angle": 0, + "content": "34. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: Nonlinear Phenomena 60(1), 259-268 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.494 + ], + "angle": 0, + "content": "35. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.549 + ], + "angle": 0, + "content": "36. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "37. Sun, C., et al.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.632 + ], + "angle": 0, + "content": "38. Tai, Y.W., et al.: Richardson-lucy deblurring for scenes under a projective motion path. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(8), 1603-1618 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.688 + ], + "angle": 0, + "content": "39. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "40. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: IEEE International Conference on Computer Vision (ICCV). IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "41. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "42. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "43. Xu, L., Zheng, S., Jia, J.: Unnatural L0 sparse representation for natural image deblurring. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition," + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Peng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "Portland, OR, USA, June 23-28, 2013. pp. 1107-1114. IEEE Computer Society (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "44. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "45. Yu, X., Dai, P., Li, W., Ma, L., Shen, J., Li, J., Qi, X.: Towards efficient and scale-robust ultra-high-definition image demoiring. In: European Conference on Computer Vision. pp. 646-662. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.786, + 0.286 + ], + "angle": 0, + "content": "46. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. arXiv preprint arXiv:2311.16493 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.356 + ], + "angle": 0, + "content": "48. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371-378 (2001)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_origin.pdf b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30b31b5b6708b2d30cf8b60615ecdf17fe6ed9a0 --- /dev/null +++ b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/6df548a5-1a2f-4b34-87d4-0c97d9a0b9d3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b54fd96c9268b129715f990fc2a224f91ea51fd6dc2f12e84a20d3b8b9c0b5c6 +size 8515133 diff --git a/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/full.md b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3b68d5a117d35a102be758738b7964fafb3d7098 --- /dev/null +++ b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/full.md @@ -0,0 +1,305 @@ +# BAGS: Blur Agnostic Gaussian Splitting through Multi-Scale Kernel Modeling + +Cheng Peng*, Yutao Tang*, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa + +Johns Hopkins University, Baltimore MD 21218, USA {cpeng26, ytang67, yzhou223, nwang43, xliu253, dlil90, rchella4}@jhu.edu + +Abstract. Recent efforts in using 3D Gaussians for scene reconstruction and novel view synthesis can achieve impressive results on curated benchmarks; however, images captured in real life are often blurry. In this work, we analyze the robustness of Gaussian-Splatting-based methods against various image blur, such as motion blur, defocus blur, downscaling blur, etc. Under these degradations, Gaussian-Splatting-based methods tend to overfit and produce worse results than Neural-Radiance-Field-based methods. To address this issue, we propose Blur Agnostic Gaussian Splatting (BAGS). BAGS introduces additional 2D modeling capacities such that a 3D-consistent and high quality scene can be reconstructed despite image-wise blur. Specifically, we model blur by estimating per-pixel convolution kernels from a Blur Proposal Network (BPN). BPN is designed to consider spatial, color, and depth variations of the scene to maximize modeling capacity. Additionally, BPN also proposes a quality-assessing mask, which indicates regions where blur occur. Finally, we introduce a coarse-to-fine kernel optimization scheme; this optimization scheme is fast and avoids sub-optimal solutions due to a sparse point cloud initialization, which often occurs when we apply Structure-from-Motion on blurry images. We demonstrate that BAGS achieves photorealistic renderings under various challenging blur conditions and imaging geometry, while significantly improving upon existing approaches. + +# 1 Introduction + +High quality scene reconstruction and novel view synthesis from 2D images is a long-standing research problem with extensive applications in robotics, virtual reality, e-commerce, cinematography, etc. Significant advancements in this field have been made in recent years with the introduction of Neural Radiance Field (NeRF) [24]. NeRF can achieve photorealistic view synthesis by implicitly representing the scene with a Multi-Layer Perceptron (MLP) and optimizing the MLP with differentiable ray-tracing. Many subsequent works have since been proposed to improve various aspects of NeRF, including optimization acceleration [6,9,25,37], anti-aliasing [2,3], dynamic and generative modeling [19,31,32,40], etc. Recently, 3D Gaussian Splatting (3DGS) [12] has been + +introduced as an alternative to NeRF-based approaches. 3DGS uses Gaussians as explicit 3D representations and a differentiable rasterization technique, which enables high quality scene reconstruction in a short time. + +![](images/8d3fcd2a975c5ff47d3896df15332accc4b114e259f8f13e0234835833cd5d68.jpg) +Fig. 1: We introduce BAGS, which can reconstruct high quality scenes even from blurry training images. Moreover, BAGS can provide kernels and masks that indicate the types and regions of the blur, as shown by the highlighted regions under Mask. + +![](images/6d6a7a545446cf20e33a0dd494818bb705972490b3261080cf73a7031fee036d.jpg) + +![](images/adf1ed441c893461dd93777535ee991d9a7d6817c6c3ade3ed33c852faaf8490.jpg) + +Practical applications of 3DGS, however, can be challenging. Despite its impressive performance on curated datasets, 3DGS requires high quality images and a good point cloud initialization to work well. Without these conditions, 3DGS often generates undesired Gaussians to overfit observation noise, leading to worse renderings, e.g., compared to NeRF. For instance, images acquired in real life or from the internet are often in non-ideal conditions. The camera may be unstable or out of focus, leading to motion or defocus blur. Images may also have different resolutions depending on the sensor types or post processes. + +There often exists sufficient information within a multi-view image set to generate a high quality reconstruction, even if the individual images are degraded. This is because the same region is often observed repeatedly from different views; such redundancy can then be explored to better recover the underlying scene. Works like Deblur-NeRF [21] and its variants [15, 16, 30, 41] are designed based on this insight. Formally, a simple yet powerful forward model can be used to describe the potentially blurry pixel value $\tilde{C}(x)$ at coordinate $x$ : + +$$ +\tilde {C} (x) = \sum_ {x _ {k} \in \mathcal {N} (x)} C \left(x _ {k}\right) h \left(x _ {k}\right) \text {s . t} \sum_ {x _ {k} \in \mathcal {N} (x)} h \left(x _ {k}\right) = 1, \tag {1} +$$ + +where $C(x_{k})$ is the clear pixel value at pixel coordinates $x_{k}$ around the neighborhood of $x$ , and $h(x_{k})$ is the blur kernel weights. In this formulation, $C(x_{k})$ is multi-view consistent, and $\tilde{C}(x)$ is multi-view inconsistent due to the blur from $h$ . Previous methods [15, 16, 21, 30, 41] aim to simultaneously optimize a multi-view consistent $C(x_{k})$ through NeRF and the multi-view inconsistent $\tilde{C}(x)$ through a view-specific network that predicts $h(x_{k})$ . As such, we render only the clear $C(x_{k})$ after training. In comparison, vanilla NeRF [24] treats $C(x_{k})$ and $\tilde{C}(x)$ as one variable, forcing multi-view inconsistencies to be baked into the scene. + +In this work, we build upon the blur formulation in Eq. (1), and explore two key differences between NeRF and 3DGS in robust reconstruction. NeRF-based approaches suffer from slow training speed, i.e. only limited number of rays can + +be rendered at once. This is exacerbated by kernel modeling, which requires the rendering of a patch to model a single pixel. To mitigate this issue, current approaches model the kernel $h$ sparsely [15, 21, 30] or only for specific types of blur [16, 41]. Since 3DGS can rasterize entire images efficiently, it potentially has great synergy with Eq. (1). On the flip side, NeRF requires only calibrated camera poses for reconstruction, whereas 3DGS also requires a good point cloud initialization. For noisy images, camera poses can often be estimated based on just a few correct matches, but a dense point cloud is challenging to obtain. + +As shown in Fig 1, we introduce Blur Agnostic Gaussian Splatting, or BAGS, which is a Splatting-based method that is robust against various types of blur. At the core of BAGS is a Blur Proposal Network (BPN). Contrary to previous methods [15,21,30] that sparsely estimate $h$ , BPN directly estimates dense kernels on full images during training. We find that estimating a dense $h$ is more effective and no less efficient than estimating a sparse one, which also has to predict kernel positions $x_{k}$ . BPN considers the position and view embedding of pixel $x$ , accounting for the spatial variance in image blur. Furthermore, a three-layer Convolution Neural Network (CNN) is used to compute the RGBD features of the rendered image. These extracted features allow BPN to consider the color and depth variation of the scene, e.g. around edges and corners, more effectively. Finally, BPN also computes a per-pixel mask, which blends the estimated clear and blurry pixels together. We make the observation that not necessarily all pixels are blurry; therefore, by constraining on its sparsity, such a mask can visualize regions where blur occurs. This is particularly useful as a way to evaluate image quality. + +In addition to BPN, BAGS proposes a coarse-to-fine kernel optimization scheme. In some cases, BAGS cannot perform reconstruction well if it jointly optimizes BPN with a very sparse point cloud, due to the inherent ambiguity of separating blur from the observation. Furthermore, optimizing a dense kernel on full resolution directly is computationally expensive and time-consuming. To solve these conundrums, we first optimize the scene and BPN at lower scales with a smaller kernel dimension. This allows the optimization process to address a less challenging and ambiguous problem initially. We then gradually increase the resolution of the training images and the estimated kernels; since the blur problem has been partially addressed at a small scale, optimization at larger scales is more stable and effective. Under this scheme, we not only observe improved performance but also faster optimization process. + +In summary, our contributions can be summarized in three parts: + +1. We introduce a Blur Proposal Network, which considers spatial, depth, and color variations of the scene to model image blur; BPN can also produce a mask that indicates blurry regions in an input image. +2. We introduce a coarse-to-fine optimization scheme, which gradually increases the training image resolution and the estimated kernel size with additional neural network layers; this improves the stability of the joint optimization process given a sparse point cloud. + +3. We evaluate the overall method, BAGS, on three image blur scenarios and find significant quantitative and visual improvements compared to current SoTA methods. + +# 2 Related Work and Background + +Single Image Restoration. Restoring images from degradation, such as blur, has a long history of research. In general, this is a heavily ill-posed problem as many potential high quality images can lead to the same degraded observations. Traditionally, various priors are used to constrain the solution space [5, 14, 18, 34, 38, 43]. More recently, deep-learning-based methods have achieved great performances by learning a better prior in a data driven way, e.g. as a function that directly maps blurry images to clean images. In this paradigm, we train a powerful restoration model with millions of training pairs and a large neural network. Ideally, images can be restored by such a model [1, 4, 8, 13, 17, 42, 45, 47]. This paradigm has many issues in practice. For example, real training pairs are difficult to acquire in large scale; therefore, synthetic data is often used, which leads to significant domain shift problems. Every type of degradation also requires its own training sets and models, which exacerbate the data and computation costs. Even for degradations such as low resolution, for which is easy to synthesize training pairs, the restoration is unlikely to be multi-view consistent. + +NeRF in Non-Ideal Conditions. NeRF [24] works well when images are clean and well-calibrated; however, images acquired in the wild are often less ideal. Factors such as low light, camera motion, object motion, and incorrect focus can degrade image quality or affect multi-view consistency. Many works introduce domain knowledge to model the non-ideal contributors, such that the radiance field only models a canonical, multi-view consistent scene. RawNeRF [23] uses an approximated tonemapped loss and variable exposure to account for low light observations. NeRF-W [22] uses an image-specific embedding to model inconsistent appearances and transient objects. For object motion, many works [28,29,33,40] use an implicit function to describe scene deformation, such that the traced rays are first deformed. For image degradation such as blur, similar ideas have been used [11, 15, 21, 30, 41] to model the forward image degradation process; i.e. the 3D consistent scene and 2D inconsistent degradation will be separated based on the multi-view photometric loss. + +3D Gaussian Splatting. Recently, 3DGS [12] introduces a novel, point-based scene reconstruction method that improves upon NeRF in many aspects. 3DGS utilizes a set of 3D Gaussian primitives [48] $\{\mathcal{G}_n|n = 1,\dots ,N\}$ to represent the scene. These Gaussians are parameterized as follows: + +$$ +\mathcal {G} _ {n} (\mathbf {v}) = e ^ {- \frac {1}{2} (\mathbf {v} - \mathbf {p} _ {n}) ^ {T} \Sigma_ {n} ^ {- 1} (\mathbf {v} - \mathbf {p} _ {n})}, \tag {2} +$$ + +where the center $\mathbf{p}_n\in \mathbb{R}^{3\times 1}$ and covariance matrix $\Sigma_{n}\in \mathbb{R}^{3\times 3}$ define the position, scale, and rotation of the Gaussian at point $\mathbf{v}\in \mathbb{R}^{3\times 1}$ in space. Additionally, each Gaussian also contains opacity $\alpha$ and color $c$ . For rendering, the Gaussians are projected onto the image plane as $\mathcal{G}_n^{\mathrm{2D}}$ , where the center and covariance ma + +trices are transformed based on the camera rotation $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ and translation $\mathbf{t} \in \mathbb{R}^{3 \times 1}$ : + +$$ +\mathbf {p} _ {n} ^ {\mathrm {2 D}} = \mathbf {R} \mathbf {p} _ {n} + \mathbf {t}, \Sigma_ {n} ^ {\mathrm {2 D}} = \mathbf {R} \Sigma_ {n} \mathbf {R} ^ {T}. \tag {3} +$$ + +We can then aggregate 2D Gaussians according to their depth order $D$ over a pixel/tile, based on alpha blending: + +$$ +C (x) = \sum_ {i = 1} ^ {D} T _ {i} \alpha_ {i} \mathbf {c} _ {i} \mathcal {G} _ {i} ^ {\mathrm {2 D}}, T _ {i} = \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j} \mathcal {G} _ {j} ^ {\mathrm {2 D}}\right). \tag {4} +$$ + +With well-designed tiling, this splatting-based rasterization technique is significantly faster than ray-tracing in NeRF-based methods without affecting rendering quality; various work have ensued to further improve the Gaussian representation [10, 26, 27, 44] and the rasterization process [20, 46]. + +# 3 Blur Agnostic Gaussian Splitting + +BAGS is designed to robustly optimize 3D Gaussians by introducing additional modeling capacities in 2D, which allows BAGS to model away 3D-inconsistent blur from the scene. As shown in Fig. 2, BAGS consists of two parts: a Blur Proposal Network (BPN) and a coarse-to-fine optimization scheme. For clarity, we first describe the design of BPN on a single scale, and then describe modifications on BPN and the training loss on scale $s$ for multi-scale training. + +# 3.1 Blur Proposal Network + +Following the forward blur model described in Eq. (1), BPN models a per-pixel convolution kernel $h(x) \in \mathbb{R}^{K \times K}$ for every pixel $x$ , where $K$ is the kernel size. The rasterization efficiency in Gaussian Splatting allows us to model $h(x)$ as a full convolution kernel. Additionally, BPN estimates a per-pixel scalar mask $m(x) \in [0, 1]$ , controlling the areas where blur modeling takes place. + +Multi-Modal Features. In scene reconstruction, a large neural network will impact training time. As such, we use a four-layer MLP $\mathcal{F}_{\mathrm{kernel}}$ to model blur. To maximize the capacity of $\mathcal{F}_{\mathrm{kernel}}$ , we design Multi-Modal Features (MMF) that can help distinguish the scene from blur. Real image blur typically varies between different observed images and different pixels within an image; therefore, we provide $\mathcal{F}_{\mathrm{kernel}}$ with a learnable view embedding $l(i)$ and a positional embedding $p(x)$ for training view $i$ and coordinate $x$ . Additionally, we provide $\mathcal{F}_{\mathrm{kernel}}$ with a set of features $f_{\mathrm{RGBD}}$ extracted from the rendered image $C$ and depth $D$ from rasterization, using a small, three-layer CNN $\mathcal{F}_{\mathrm{feat}}$ . These features are concatenated with $l(i)$ and $p(x)$ to form the MMF for $\mathcal{F}_{\mathrm{kernel}}$ . Consequently, the forward process to estimate the kernel $h(x,i)$ and the mask $m(x,i)$ given a specific pixel $x$ and training view $i$ can be noted as: + +$$ +h (x, i), m (x, i) = \mathcal {F} _ {\mathrm {k e r n e l}} (l (i) \oplus p (x) \oplus f _ {\mathrm {R G B D}} (x, i)), +$$ + +$$ +\text {w h e r e} f _ {\mathrm {R G B D}} (x, i) = \mathcal {F} _ {\text {f e a t}} (C (x, i) \oplus D (x, i)), 0 \leq m \leq 1. +$$ + +Incorporating RGBD features enbales BPN to explore the patch statistics around the current pixel, especially at edges and corners where blur is prominent. Scene + +![](images/b016e80466ffca5bd1f2417ee88c805e96dacf91e99ab72f453121fbfb55acce.jpg) +Fig. 2: BAGS is optimized by training a Blur Proposal Network on top of the scene $\mathcal{G}(\mathbf{v})$ over multiple scales. Top: We extract $f_{\mathrm{RGBD}}^s$ from color and depth, which are concatenated with the position and view embedding $p(x), l(i)$ to form the Multi-Modal Feature (MMF). The kernel MLP then estimates the per-pixel kernel $h^s$ and mask $m^s$ . We use $h^s$ to model the blur image $\tilde{C}^s$ and employ $m^s$ to blend the rendered image $C^s$ and the blur-modeled image $\tilde{C}^s$ , yielding $C_{\mathrm{out}}^s$ . Bottom: After $N_s$ steps, we upscale image resolution and modify the kernel MLP to produce $h^{s - 1}$ with a larger kernel size. + +depth also strongly correlates with the magnitude of blur. For example, defocus blur affects pixels that are outside of a specific focus plane; scene content that is further away from this focus plane becomes more blurry. For camera motion blur, pixels in the near plane may shift more than pixels in the far plane. We find the inclusion of $f_{\mathrm{RGBD}}$ notably improves BPN's modeling capacity and the quality of novel view synthesis. + +Residual Mask Bottleneck. As described in Eq. (5), $\mathcal{F}_{\mathrm{kern}}$ also estimates a per-pixel scalar mask $m(x,i)$ . Specifically, this mask is used to blend the Gaussian-splatted image $C$ with the blur-modeled image $\tilde{C}$ as follows: + +$$ +C _ {\mathrm {o u t}} (x, i) = \left(1 - m (x, i)\right) C (x, i) + m (x, i) \tilde {C} (x, i), \tag {6} +$$ + +where $C_{\mathrm{out}}$ is the final output at training. We note that not necessarily all pixels are blurry in an observed image. The blending $m$ provides an additional degree of freedom to allow $C$ , the clear pixel, to be directly compared with $C_{\mathrm{obs}}$ ; i.e., if $C_{\mathrm{obs}}(x,i)$ suffers from blur, then $m(x,i)$ should be large, giving a higher weight to $\hat{C}(x,i)$ ; if $C_{\mathrm{obs}}(x,i)$ if free of blur, $m(x,i)$ should be small, favoring $C(x,i)$ . + +This design is similar to residual learning by providing the output with a skip connection from $C$ as initialization. We apply a small sparsity constraint on $m$ to encourage a larger weight towards $C$ by default. After training, we can extract $m$ to automatically identify the degraded regions in an input image. + +![](images/59011e96dc7eeaa4cb97086bdd07afed2a6c8f42bac595790d90bc3099cd39f5.jpg) +(a) Without C2F + +![](images/e9c49ff277167511f89fcbd1b36f4d882ea98fdfd77582fd3ef9ed1c53dc386a.jpg) +(b) BAGS +Fig. 3: Gaussians may get stuck at a local minimum without proper initialization. As indicated in 3a, optimizing $h$ directly can lead to noisy surfaces. Naively densifying the scene before adding $h$ can also lead to noisy Gaussians; as shown in 3c, the noisy Gasussians are not well removed even after adding $h$ . By using a coarse-to-fine training schedule, we achieve better results in 3b and 3d. + +![](images/7ecc8a9b246519d1859d7ebace6e8b20c303e3c1bc3b803e0be6e9eee12c38f0.jpg) +(c) Naive densification + +![](images/f8cde32d8cbbf6de12e1f5119189dc5016ef47e701cf81aba73020f40f491aaf.jpg) +(d) BAGS + +# 3.2 Coarse-to-Fine Kernel Optimization + +A high quality 3DGS reconstruction requires a good point cloud estimation from Structure from Motion (SfM). Camera calibration parameters, which can be estimated from just a few matches, are relatively robust against image blur. On the other hand, obtaining a sufficiently dense point cloud can be challenging. Unfortunately, 3DGS is prone to over-fitting without a good initialization. This problem can worsen when a blur model is introduced and the kernels $h$ provide additional degrees of freedom in scene optimization, as shown in Fig. 3. Furthermore, estimating dense per-pixel kernels on a full resolution image can be computationally expensive. A sparse estimation of the kernel does not meaningfully improve efficiency, as rendering is cheap in rasterization while the neural network to predict any type of kernel is still costly. Fundamentally, the complexity to compute per-pixel kernel scales quadratically with resolution. To address these problems, we propose a coarse-to-fine kernel optimization scheme, which initially optimizes Eq. (5) at a low resolution and gradually increases render resolution, achieving a more stable training process and better efficiency. + +Multi-Scale BPN. For multi-scale optimization, BPN should generate kernels with a consistent field of view. We define scale based on the downscaling ratio $2^{s - 1}$ in resolution, where the original resolution is defined as $s = 1$ . Additionally, if the image resolution is downscaled by a factor of two, i.e. $s = 2$ , the proposed kernel size $K^s$ should also be similarly downscaled. As shown in Fig. 2, a proper $K^s$ ensures that the kernel is always modeling blur from the same regions. + +To enable multi-scale training, we separate $\mathcal{F}_{\mathrm{kernel}}^s$ into two parts: a two-layer base MLP $\mathcal{F}_{\mathrm{base}}^s$ and two one-layer MLP heads $\mathcal{F}_{\mathrm{blur}}^s$ and $\mathcal{F}_{\mathrm{mask}}^s$ . The base MLP extracts features from MMF, such features are then provided to the two heads for generating the kernels $h^s$ and masks $m^s$ at scale $s$ . In practice, only $\mathcal{F}_{\mathrm{blur}}^s$ has a scale-dependent parameter size corresponding to the kernel size $K^s$ ; therefore, a new MLP layer with output dimension $K^s \times K^s$ is added for every scale. The updated multi-scale BPN can be described as follows: + +$$ +f _ {\text {i n t e r}} (x, i) = \mathcal {F} _ {\text {b a s e}} ^ {s} (l (i) \oplus p (x) \oplus f _ {\text {R G B D}} (x, i)) +$$ + +$$ +h ^ {s} (x, i) = \operatorname {s o f t m a x} \left(\mathcal {F} _ {\text {b l u r}} ^ {s} \left(f _ {\text {i n t e r}} (x, i)\right)\right), \tag {7} +$$ + +$$ +m ^ {s} (x, i) = \operatorname {s i g m o i d} \left(\mathcal {F} _ {\text {m a s k}} ^ {s} \left(f _ {\text {i n t e r}} (x, i)\right)\right). +$$ + +Coarse-to-Fine Optimization. We split training into multiple stages based on scale. Starting from the coarsest scale $s = S$ to the finest scale $s = 1$ , we optimize BAGS with a downsampling factor of $2^{s - 1}$ applied to $C, \tilde{C}, C_{\mathrm{obs}}$ , and $h$ . The optimization for each scale lasts for $N_{s}$ steps, we then increase the resolution in $C, \tilde{C}, C_{\mathrm{obs}}$ , and $h$ . The scale-dependent training loss can be described as: + +$$ +\mathcal {L} _ {s} = \lambda_ {\mathrm {p h o t o}} \| C _ {\mathrm {o u t}} ^ {s} - C _ {\mathrm {o b s}} ^ {s} \| + \lambda_ {\mathrm {D S}} \mathcal {L} _ {\mathrm {D - S S I M}} \left(C _ {\mathrm {o u t}} ^ {s}, C _ {\mathrm {o b s}} ^ {s}\right) + \lambda_ {\mathrm {m a s k}} \| m ^ {s} \|, \tag {8} +$$ + +where $C_{\mathrm{obs}}^s$ and $C_{\mathrm{out}}^s$ are the observed and predicted image at scale $s$ , $\mathcal{L}_{\mathrm{D - SSIM}}$ is the structural similarity loss; $\| m^{s}\|$ is the mask sparsity loss. We find this training scheme achieves the best of both worlds: activating BPN at low resolution generates less overfitting artifacts during densifying the point cloud given the reduced visibility of blur. Moreover, it also serves as an efficient warm-up for BPN at higher resolution, as computing per-pixel kernels at low resolution is much cheaper. + +# 4 Experiments + +We evaluate BAGS under three degradation scenarios: camera motion blur, defocus blur, and mix resolution. All of these degradations are commonly observed in images acquired in real life or online. We further introduce a new unbounded 360 drone dataset, which is collected at sunset or night and experiences degradation due to a combination of motion blur and low light condition. + +Camera Motion and Defocus Blur. Camera motion and defocus blur are often observed in real acquisitions, e.g. due to low light, incorrect auto-focus, or limitation in the finite aperture of real cameras. Deblur-NeRF [21] provides real scene acquisitions that contain camera motion and defocus blur; both blur categories consist of ten scenes. While real world scenes have blur-free references, the camera settings, e.g., exposure, may not be consistent with source views. Experiments on synthetic blur are reported in supplemental material. + +Mix Resolution. To analyze the scenario where the input images are of inconsistent resolution, we apply various downsampling operations on the Mip-NeRF 360 dataset [2]. Specifically, we split the training images into four equal parts and apply 4X, 3X, 2X, and no downscaling to each of the parts. The image splits are sampled in uniform randomness to ensure no downscaling covers a consecutive interval of the camera trajectory. We then use NeRFStudio's data processing pipeline [39], which converts all images to the highest resolution and calibrates with COLMAP [35]. Similar to previous works [2,3,46], we perform the entire process on downscaled data instead of the original 4K resolution. + +Unbounded Drone Motion Blur. The dataset introduced by DeblurNeRF [21] serves as a good benchmark for performance comparison; however, it is collected in a forward bounded format. These forward bounded scenes focus on a small area and have significant overlaps within the images; as such, it is unclear if the methodology [15, 16, 21, 30, 41] based on these sets can generalize to other scenarios. In fact, none of the deblurring methods can work for unbounded scenes. To supplement DeblurNeRF's dataset and understand how BAGS performs in realistic settings, we collect five sets of drone footage on different buildings in 360 + +![](images/037523c94d0a7112955b58b87388cbcb0db060bfa5e00227677f55c2271bbe27.jpg) + +![](images/ea7ccc980b07eef8189fdc13f9551f029dba350f8a929945f62ee895438e041e.jpg) + +![](images/50167ccbb2b1ac244c6d0f66fcbdae44a4c4ad0c964f51ed80ee18a555e19d12.jpg) +(c) + +![](images/4bfb76c4ba18ea9b7865c6aec613cbb509ec68b02498a756cd7cc766489dcb90.jpg) +(d) +Fig. 4: Ablation study of BAGS with different sub-modules. In 4a and 4b, introducing RGBD features and coarse-to-fine optimization improve novel view synthesis quality, and alternative sparse deformable kernel suffers in performance; 4d demonstrates our speed improvement against $\mathrm{BAGS}^{\mathrm{noC2F}}$ . In 4c, larger kernel generally leads to better performance, other than in the mix resolution scenario. + +unbounded format. These footages are collected at sunset or night, and suffer from camera motion blur and low light noise. For each scene, we also capture clear frames from a stable drone position for testing, resulting in around 100 images overall. We use NeRFStudio [39] to calibrate images from these footages. More details on this dataset can be found in supplemental material. + +Implementation Details. BPN's $\mathcal{F}_{\mathrm{feat}}$ is constructed as a three-layer CNN; the intermediate channel size is 64, the kernel size for each layer is 5, 5, and 3. $\mathcal{F}_{\mathrm{kernel}}$ is constructed with four linear layers; the intermediate channel size is 64. For multi-scale training, we use three scales starting with scale $s = 3$ , where $K_{3} = 5$ , $K_{2} = 9$ , and $K_{1} = 17$ . BAGS implements its Gaussian Splatting and rasterization based on Mip-Splatting [46], which has been shown to yield good results across multiple resolutions. More implementation details can be found in supplemental material. + +# 4.1 Ablation Studies + +We compare three designs in BAGS to demonstrate the effectiveness of different components: + +- BAGS $^{\text{noRGBD}}$ : We implement BPN without the RGBD features $f_{\text{RGBD}}$ . +- BAGS $^{\text{deform}}$ : We implement BPN with a sparse deformable kernel, similar to DeblurNeRF [21]. +- BAGS $^{\mathrm{noC2F}}$ : We implement BAGS without a coarse-to-fine kernel optimization scheme. + +We further investigate the effect of the final kernel size $K_{1}$ on reconstruction quality across all degradation scenarios. + +As shown in Fig. 4, the introduction of RGBD features to BPN greatly improves its ability to disentangle image blur from the underlying scene. The addition of $f_{\mathrm{RGBD}}$ increases the average performance by $1.57 / 0.92$ dB in PSNR and $0.03 / 0.03$ in SSIM for camera motion and defocus blur, respectively. This agrees with our intuition that blur is highly correlated with depth and image features like edges. Our coarse-to-fine optimization scheme also notably boosts robustness against defocus blur by $1.64$ dB and $0.08$ in PSNR and SSIM on average. As visualized in Fig. 3a, directly optimizing 3DGS on a sparse point cloud in $\mathrm{BAGS}^{\mathrm{noC2F}}$ leaves many ambiguities; while training views have low loss, test views show a lack of the correct surface structure and many artifact Gaussians. Densifying Gaussians without BPN first improves performance for defocus blur; however, this strategy is bad for motion blur, as shown in Fig. 3c. In Fig. 4d, we show that our coarse-to-fine strategy is not only performant, but also fast; on average, BAGS can reach high performance in less than fifteen minutes, which is much faster than the two hours required by PDRF [30]. This strategy is also faster compared to single-scale training, as full-scale training in our scheme uses less iterations, and lower scale training does not sacrifice performance. Both multi-scale and single-scale training take $40 \mathrm{K}$ iterations overall. + +We also implement a deformable version of BPN in $\mathrm{BAGS}^{\mathrm{deform}}$ , where BPN estimates the kernel position and kernel weights for ten points within a $20 \times 20$ -pixel patch. We show its performances in Fig. 4a, where $\mathrm{BAGS}^{\mathrm{deform}}$ is tested on camera motion blur. Camera motion blur often has a sparse representation, i.e. the camera motion trajectory [16, 41]; therefore, a ten-point estimation of the kernel may perform well. However, we find that $\mathrm{BAGS}^{\mathrm{deform}}$ still underperforms compared to a full kernel estimation in BAGS. This may be due to the difference between NeRF and 3DGS; i.e., a ray-based deformation can be better expressed in NeRF, while a patch-based convolution can be better expressed in 3DGS. We also note that $\mathrm{BAGS}^{\mathrm{deform}}$ is not more efficient than BAGS despite its sparse estimation. This is because approaches like [15, 21, 30] require one forward pass to estimate one kernel position and weight; therefore, we need to forward multiple times for every pixel to estimate the sparse kernel. In comparison, BAGS only requires a single forward pass to obtain the full kernel. + +Finally, we ablate on the kernel dimension estimated by BPN. In Fig. 4c, we show that larger kernels in general lead to higher performances; interestingly, this is not true in the mixed resolution scenario. This is likely because downscaling is uniform across all pixels in an image, i.e. a downscaling factor of four means a kernel size of five can fully explain the degradation. As such, increasing the kernel size does not improve performance and introduces noise instead. + +# 4.2 Quantitative Evaluation + +Camera Motion and Defocus Blur. As shown in Table 1, we observe that Gaussian Splatting [46] performs markedly worse than vanilla NeRF in the presence of blurry images, despite its strong performance in other curated benchmarks. Since NeRF models a continuous neural volume with an MLP, individual degradation can be smoothly explained by spreading the error into empty space. + +
Camera MotionNeRF [24]Mip-Splatting [46]Deblur-NeRF [21]DP-NeRF [15]PDRF [30]BAGS
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Ball24.080.6240.39923.220.6190.34027.360.7660.22327.200.7650.209
Basket23.720.7090.32223.240.6880.28827.670.8450.14827.740.8460.129
Buick21.590.6330.35021.460.6580.26624.770.7700.17525.700.7920.141
Coffee26.480.8060.29024.730.7490.28830.930.8980.12432.440.9150.101
Decoration22.390.6610.36320.550.6410.29924.190.7710.18623.510.7400.275
Girl20.070.7080.32019.870.7140.27822.270.7980.16921.800.7700.205
Heron20.500.5220.41319.430.5050.33222.630.6870.21022.520.6760.281
Parterre23.140.6200.40522.280.5900.32125.820.7600.21624.970.7680.227
Puppet22.090.6090.33922.050.6310.26725.240.7510.15825.120.7580.181
Stair22.870.4560.48721.910.4740.38725.390.6300.21025.780.6470.229
Average22.690.6350.36921.870.6270.30625.630.7670.18225.910.7750.160
Defocus
Cake24.420.7210.22522.170.6450.29026.270.7800.12826.160.7780.127
Caps22.730.6310.28021.000.5030.44223.870.7130.16123.950.7120.143
Cisco20.720.7220.12620.080.7090.16320.830.7270.08720.730.7260.084
Coral19.810.5660.26919.600.5550.32319.850.6000.12122.800.7410.096
Cupcake21.880.6810.21621.550.6810.21322.260.7220.11620.110.6110.118
Cups25.020.7580.23220.930.6460.31426.210.7990.12726.750.8140.104
Daisy22.740.6200.26221.590.6140.27323.520.6870.12123.790.6970.108
Sausage17.790.4830.27917.780.4730.28418.010.5000.18018.350.5440.147
Seal22.790.6270.26822.130.5920.31026.040.7770.10525.950.7780.103
Tools26.080.8520.15523.980.8120.19627.810.8950.06128.070.8980.054
Average22.400.6660.23121.080.6230.28123.470.7200.12123.670.7300.108
+ +Table 1: Quantitative comparisons on the camera motion and defocus blur dataset. We color code the best PSNR, SSIM, and LPIPS performances. + +On the other hand, 3DGS is based on a discrete representation and tends to produce concentrated artifacts. Our BAGS can achieve new SoTA performances compared to previous NeRF-based deblurring methods despite the unfavorable performances in 3DGS. Specifically, we find significant improvements on camera motion blur across all metrics and achieve a fifty percent improvement in LPIPS on average. As visualized in Fig. 5, BAGS achieves much sharper results compared to previous methods [15, 16, 21, 30], to the point where it is hard to distinguish the groundtruth image from our rendered results. + +While improvements on defocus blur is less drastic in Table 1, we still observe great visual improvements. Several reasons contribute to the less apparent quantitative improvements. First, the collected groundtruth images have different exposure compared to training images; as such, image-wise similarity metrics like PSNR are less reliable. Secondly, many scenes contain specular surfaces, which all methods do a bad job at reconstructing. As shown in Fig. 5, much better details can be recovered through BAGS than other methods. + +Mix Resolution and Low Light Motion Blur. We investigate the robustness of several SoTA reconstruction methods on downscaling perturbation and motion blur observed in low light condition. The experiments are done on unbounded scenes and are more challenging and realistic than the forward bounded scenes in DeblurNeRF [21]. We note that previous approaches [15, 21, 30, 41] cannot handle unbounded scenes; therefore, we do not include them in Table 2. + +As shown in Table 2, Mip Splating performs on par with Mip-NeRF 360; both are vastly better than NeRFacto [39], which stores hash features [25] to accelerate the training process and uses a very small MLP represent the scene. We also employ a Super-Resolution (SR) model HAT [7] on downscaled images in the hope that single-image SR can improve performances. However, pre-trained + +![](images/1359a9896b09a52e78e39c9fcd739a2e436e7770f845c14d15e16fb6c440ce1e.jpg) +Fig. 5: Visualizations of test views on camera motion and defocus blur dataset. Mip-Sp and Db-NeRF are short for Mip-Splatting [46] and Deblur-NeRF [21]. + +SR models are limited by domain shifts, dedicated upscaling factors, and the lack of multi-view consistency. As visualized in Fig. 6, while the render is somewhat sharper after SR, the color becomes unfaithful. Compared with all other methods, our rendered results are much sharper in details. This robust performance comes from BPN's ability to model Gaussian blur on high resolution before comparing them with the downscaled observations. + +Motion blur is often observed in low light condition. To this end, we perform experiments on scenes acquired near sunset or at night. As shown in Table 2, BAGS again improves significantly over other reconstruction approaches, which shows that our method can be generalized to different imaging geometry. As shown in Fig. 6, we employ MPR [47], a single image deblurring model, on the motion-blurred images before reconstructing with Mip-Splatting [46]. Similar to previous attempts, the deblurring prior leads to a slightly clearer rendering but is not sufficient to remove all blur. BAGS reconstructs the scene remarkably well compared to other methods, where even the window details can be recovered correctly. Please refer to the supplemental material for more visualizations. + +Kernel and Mask Visualization. BAGS models image degradation through convolution kernels $h$ and masks $m$ , which we visualize in Fig. 7. By constraining on its sparsity, we can see that $m$ meaningfully highlights the regions of blur in a given training image. On the other hand, low-value regions indicate either high quality observation or textureless regions. By visualizing the kernels modeled + +
MixedResNeRFacto [39]Mip-NeRF 360 [2]Mip-Sp [46]Mip-Sp-HAT [7]BAGS
PSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPS
Bicycle19.250.3490.49623.820.6100.37524.030.6280.36423.20
Bonsai21.800.6190.25929.910.8710.19029.270.8760.19328.38
Counter24.070.7140.28028.090.8480.22027.790.8530.21627.02
Flowers19.630.4180.45921.980.5660.38921.960.5700.39720.81
Garden21.610.4890.35725.340.6670.31725.280.6820.30523.96
Kitchen22.730.5840.30427.820.7790.23827.490.7780.24126.10
Room24.680.7910.21430.380.9020.19830.430.9030.20329.61
Stump23.190.5550.37725.750.6880.32025.870.7030.29924.09
Treehill20.110.4160.50122.520.5580.43722.860.5500.45721.76
Average21.900.5480.36126.180.7210.29826.110.7270.29724.99
Lowlight-droneNeRFacto [39]Mip-NeRF 360 [2]Mip-Sp [46]Mip-Sp-MPR [47]BAGS
PSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPS
Math-night21.330.5820.32629.510.8820.19129.610.8840.19929.15
Library-night24.480.7680.38126.380.8210.34926.170.8200.37025.89
BioMed-sunset23.760.6500.43924.480.6880.40224.850.7120.42624.78
Admin-sunset21.960.6210.40025.470.7570.35525.540.7640.36925.19
Aud-sunset24.670.7090.39825.900.7500.36826.330.7700.38626.77
Average23.240.6660.38926.350.7800.33326.500.7900.35026.36
+ +Table 2: Quantitative comparisons on mixed resolution and low light motion blur in unbounded 360 geometry. Mip-Sp is short for Mip-Splatting [46]. Mip-Sp-HAT and Mip-Sp-HAT-MPR use prior models to restore images first before Mip-Splatting. + +
Novel ViewNeRFactoMipNeRF-360Mip-SpMip-Sp-HATOursG.T.
PSNR/SSIM21.61/.57624.09/.67025.11/.74624.41/.72228.46/.899MixedRes
PSNR/SSIM23.08/.41224.98/.56224.93/.59024.68/.58731.79/.949MixedRes
Novel ViewNeRFactoMipNeRF-360Mip-SpMip-Sp-MPROursG.T.
PSNR/SSIM17.86/.60818.33/.70418.57/.71418.51/.71824.47/.850LowLight
PSNR/SSIM22.40/.68325.36/.81325.51/.81727.03/.86731.75/.938LowLight
+ +Fig. 6: Visualizations of test views on mix resolution and low light motion blur dataset. + +at different pixel, we can also easily characterize the types of observed blur. Specifically, the estimated kernels in camera motion blur exhibit clear patterns of camera movement, while those for defocus blur show Gaussian-like distributions based on the pixel's distance from the focus plane. These self-emerged properties + +in BAGS provide ways for us to automatically and precisely evaluate the quality of training images. + +![](images/41c5d2ff34a34ecc86c2bdd8b8d8e61aaddeb70ddeed3818c36b9dd2b10b876b.jpg) +Camera Motion + +![](images/d108c2bd4038b2113bb4c1b36309890bbcb6224036c84d5389aba32cc486fcdb.jpg) +Kernel & mask +Fig. 7: Visualization of estimated convolution kernels and masks in different blur. + +![](images/f135d8835104fbccc2059f2138b0de441611621c5d191256ff298c631592f938.jpg) +Defocus + +![](images/71e0538c850d11aa89879fe1add6c250e005f017734edcf0c392a007149f40b9.jpg) +Kernel & mask + +Discussion. BAGS can achieve great results; however, there also exists many potential improvements. The added neural networks and convolution operations require additional computation. While BAGS is a generalist approach and addresses multiple types of blur, it may become very expensive to estimate per-pixel convolution in high resolution. To this end, we have explored utilizing pixel shuffling [36], which converts spatial resolution into channels, before we estimate blur kernel; i.e., this assumes that within a small region, blur kernel $h$ stays roughly consistent. As shown in the supplemental material, we can scale to great visual results at 2K resolution on the DeblurNeRF dataset. Many other potential directions can also address this complexity issue, including leveraging the mask $m$ to identify important regions to focus on, a better designed low rank kernel estimation, degradation-specific optimization, etc. As such, while BAGS is already more efficient than comparable NeRF-based methods, we anticipate that its computation cost can be further optimized in the future. + +# 5 Conclusion + +We present BAGS, a novel scene reconstruction method that can handle various noise in input images. We demonstrate that a vanilla Gaussian Splatting is particularly susceptible to degraded images due to its discrete representation, even compared with NeRF. To address this issue, BAGS implements a 2D degradation model, BPN, which estimates convolution kernels jointly with 3D scene optimization. BPN allows for additional freedom to address 3D inconsistencies, and produces interpretable kernels and masks to indicate the degradation types and regions within an image. Additionally, BAGS leverages a coarse-to-fine kernel optimization scheme, which gradually models the 3D scene with higher resolution images and kernels. This approach successfully addresses the 2D-3D ambiguity in optimizing convolution kernels with a very sparse point cloud. We perform extensive experiments on a variety of noise in training images, including camera motion blur, defocus blur, and low resolution. Under these conditions, BAGS can render high fidelity scenes and significantly improves upon previous approaches. In the future, we hope to further improve the computational efficiency in BAGS and dynamically adjust kernel capacity based on how degraded the training views are. + +# Acknowledgement + +This research is based upon work supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. The authors also thank the support and feedbacks from researchers, Kaiwen Jiang, Dr. Ravi Ramamoorthi, at UCSD. + +# References + +1. Abuolaim, A., Afifi, M., Brown, M.S.: Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1231-1239 (2022) +2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022) +3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023) +4. Cao, Y., Liu, M., Liu, S., Wang, X., Lei, L., Zuo, W.: Physics-guided iso-dependent sensor noise modeling for extreme low-light photography. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5744-5753 (2023) +5. Chan, T.F., Wong, C.: Total variation blind deconvolution. IEEE Trans. Image Process. 7(3), 370-375 (1998) +6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022) +7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22367-22377 (June 2023) +8. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4641-4650 (2021) +9. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022) +10. Gao, J., Gu, C., Lin, Y., Zhu, H., Cao, X., Zhang, L., Yao, Y.: Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043 (2023) +11. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-nerf: High dynamic range neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18398-18408 (2022) + +12. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +13. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5886-5895 (2023) +14. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011. pp. 233-240. IEEE Computer Society (2011) +15. Lee, D., Lee, M., Shin, C., Lee, S.: Dp-nerf: Deblurred neural radiance field with physical scene priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12386-12396 (2023) +16. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023) +17. Lee, J., Son, H., Rim, J., Cho, S., Lee, S.: Iterative filter adaptive network for single image defocus deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2034-2042 (2021) +18. Levin, A.: Blind motion deblurring using image statistics. In: Scholkopf, B., Platt, J.C., Hofmann, T. (eds.) Advances in Neural Information Processing Systems 19, Proceedings of the Twentieth Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada, December 4-7, 2006. pp. 841-848. MIT Press (2006) +19. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +20. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109 (2023) +21. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022) +22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7210-7219. Computer Vision Foundation / IEEE (2021) +23. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16190–16199 (June 2022) +24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I. Lecture Notes in Computer Science, vol. 12346, pp. 405-421. Springer (2020) +25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022) + +26. Navaneet, K., Meibodi, K.P., Koohpayegani, S.A., Piriyawash, H.: Compact3d: Compressing gaussian splat radiance field models with vector quantization. arXiv preprint arXiv:2311.18159 (2023) +27. Niedermayr, S., Stumpfegger, J., Westermann, R.: Compressed 3d gaussian splatt-ting for accelerated novel view synthesis. arXiv preprint arXiv:2401.02436 (2023) +28. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021) +29. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., MartinBrualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. 40(6) (dec 2021) +30. Peng, C., Chellappa, R.: Pdrf: progressively deblurring radiance field for fast scene reconstruction from blurry images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2029-2037 (2023) +31. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +32. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961 (2020) +33. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 10318-10327. Computer Vision Foundation / IEEE (2021) +34. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: Nonlinear Phenomena 60(1), 259-268 (1992) +35. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +36. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016) +37. Sun, C., et al.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +38. Tai, Y.W., et al.: Richardson-lucy deblurring for scenes under a projective motion path. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(8), 1603-1618 (2011) +39. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023) +40. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: IEEE International Conference on Computer Vision (ICCV). IEEE (2021) +41. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023) +42. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490 (2022) +43. Xu, L., Zheng, S., Jia, J.: Unnatural L0 sparse representation for natural image deblurring. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition, + +Portland, OR, USA, June 23-28, 2013. pp. 1107-1114. IEEE Computer Society (2013) +44. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023) +45. Yu, X., Dai, P., Li, W., Ma, L., Shen, J., Li, J., Qi, X.: Towards efficient and scale-robust ultra-high-definition image demoiring. In: European Conference on Computer Vision. pp. 646-662. Springer (2022) +46. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. arXiv preprint arXiv:2311.16493 (2023) +47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021) +48. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371-378 (2001) \ No newline at end of file diff --git a/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/images.zip b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..454cb4c44c3f4c9e65257cd16f45c9dc460f5882 --- /dev/null +++ b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb7ccedc1ce6f83667b3c820288eececebbddc83734aff5503051632b050c908 +size 861086 diff --git a/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/layout.json b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f99edd71405803fa73d068e02fad385eb45c45f0 --- /dev/null +++ b/2024/BAGS_ Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling/layout.json @@ -0,0 +1,10399 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "BAGS: Blur Agnostic Gaussian Splitting through Multi-Scale Kernel Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": "Cheng Peng*, Yutao Tang*, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 153, + 202, + 462, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 202, + 462, + 225 + ], + "spans": [ + { + "bbox": [ + 153, + 202, + 462, + 225 + ], + "type": "text", + "content": "Johns Hopkins University, Baltimore MD 21218, USA {cpeng26, ytang67, yzhou223, nwang43, xliu253, dlil90, rchella4}@jhu.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 251, + 455, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 251, + 455, + 483 + ], + "spans": [ + { + "bbox": [ + 160, + 251, + 455, + 483 + ], + "type": "text", + "content": "Abstract. Recent efforts in using 3D Gaussians for scene reconstruction and novel view synthesis can achieve impressive results on curated benchmarks; however, images captured in real life are often blurry. In this work, we analyze the robustness of Gaussian-Splatting-based methods against various image blur, such as motion blur, defocus blur, downscaling blur, etc. Under these degradations, Gaussian-Splatting-based methods tend to overfit and produce worse results than Neural-Radiance-Field-based methods. To address this issue, we propose Blur Agnostic Gaussian Splatting (BAGS). BAGS introduces additional 2D modeling capacities such that a 3D-consistent and high quality scene can be reconstructed despite image-wise blur. Specifically, we model blur by estimating per-pixel convolution kernels from a Blur Proposal Network (BPN). BPN is designed to consider spatial, color, and depth variations of the scene to maximize modeling capacity. Additionally, BPN also proposes a quality-assessing mask, which indicates regions where blur occur. Finally, we introduce a coarse-to-fine kernel optimization scheme; this optimization scheme is fast and avoids sub-optimal solutions due to a sparse point cloud initialization, which often occurs when we apply Structure-from-Motion on blurry images. We demonstrate that BAGS achieves photorealistic renderings under various challenging blur conditions and imaging geometry, while significantly improving upon existing approaches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 502, + 230, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 502, + 230, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 502, + 230, + 515 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": "High quality scene reconstruction and novel view synthesis from 2D images is a long-standing research problem with extensive applications in robotics, virtual reality, e-commerce, cinematography, etc. Significant advancements in this field have been made in recent years with the introduction of Neural Radiance Field (NeRF) [24]. NeRF can achieve photorealistic view synthesis by implicitly representing the scene with a Multi-Layer Perceptron (MLP) and optimizing the MLP with differentiable ray-tracing. Many subsequent works have since been proposed to improve various aspects of NeRF, including optimization acceleration [6,9,25,37], anti-aliasing [2,3], dynamic and generative modeling [19,31,32,40], etc. Recently, 3D Gaussian Splatting (3DGS) [12] has been" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 228, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 228, + 666 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 228, + 666 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "introduced as an alternative to NeRF-based approaches. 3DGS uses Gaussians as explicit 3D representations and a differentiable rasterization technique, which enables high quality scene reconstruction in a short time." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 173, + 263, + 270 + ], + "blocks": [ + { + "bbox": [ + 140, + 173, + 263, + 270 + ], + "lines": [ + { + "bbox": [ + 140, + 173, + 263, + 270 + ], + "spans": [ + { + "bbox": [ + 140, + 173, + 263, + 270 + ], + "type": "image", + "image_path": "8d3fcd2a975c5ff47d3896df15332accc4b114e259f8f13e0234835833cd5d68.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "text", + "content": "Fig. 1: We introduce BAGS, which can reconstruct high quality scenes even from blurry training images. Moreover, BAGS can provide kernels and masks that indicate the types and regions of the blur, as shown by the highlighted regions under Mask." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 272, + 174, + 340, + 269 + ], + "blocks": [ + { + "bbox": [ + 272, + 174, + 340, + 269 + ], + "lines": [ + { + "bbox": [ + 272, + 174, + 340, + 269 + ], + "spans": [ + { + "bbox": [ + 272, + 174, + 340, + 269 + ], + "type": "image", + "image_path": "6d6a7a545446cf20e33a0dd494818bb705972490b3261080cf73a7031fee036d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 352, + 173, + 482, + 269 + ], + "blocks": [ + { + "bbox": [ + 352, + 173, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 352, + 173, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 352, + 173, + 482, + 269 + ], + "type": "image", + "image_path": "adf1ed441c893461dd93777535ee991d9a7d6817c6c3ade3ed33c852faaf8490.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 415 + ], + "type": "text", + "content": "Practical applications of 3DGS, however, can be challenging. Despite its impressive performance on curated datasets, 3DGS requires high quality images and a good point cloud initialization to work well. Without these conditions, 3DGS often generates undesired Gaussians to overfit observation noise, leading to worse renderings, e.g., compared to NeRF. For instance, images acquired in real life or from the internet are often in non-ideal conditions. The camera may be unstable or out of focus, leading to motion or defocus blur. Images may also have different resolutions depending on the sensor types or post processes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "content": "There often exists sufficient information within a multi-view image set to generate a high quality reconstruction, even if the individual images are degraded. This is because the same region is often observed repeatedly from different views; such redundancy can then be explored to better recover the underlying scene. Works like Deblur-NeRF [21] and its variants [15, 16, 30, 41] are designed based on this insight. Formally, a simple yet powerful forward model can be used to describe the potentially blurry pixel value " + }, + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{C}(x)" + }, + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "content": " at coordinate " + }, + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 195, + 503, + 481, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 503, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 195, + 503, + 481, + 531 + ], + "type": "interline_equation", + "content": "\\tilde {C} (x) = \\sum_ {x _ {k} \\in \\mathcal {N} (x)} C \\left(x _ {k}\\right) h \\left(x _ {k}\\right) \\text {s . t} \\sum_ {x _ {k} \\in \\mathcal {N} (x)} h \\left(x _ {k}\\right) = 1, \\tag {1}", + "image_path": "811d7b6023e7787991a3da4a003898ec6dbeac2a230484318a753e85ab66b226.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "C(x_{k})" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " is the clear pixel value at pixel coordinates " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " around the neighborhood of " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "h(x_{k})" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " is the blur kernel weights. In this formulation, " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "C(x_{k})" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " is multi-view consistent, and " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{C}(x)" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " is multi-view inconsistent due to the blur from " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ". Previous methods [15, 16, 21, 30, 41] aim to simultaneously optimize a multi-view consistent " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "C(x_{k})" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " through NeRF and the multi-view inconsistent " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{C}(x)" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " through a view-specific network that predicts " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "h(x_{k})" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ". As such, we render only the clear " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "C(x_{k})" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " after training. In comparison, vanilla NeRF [24] treats " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "C(x_{k})" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{C}(x)" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " as one variable, forcing multi-view inconsistencies to be baked into the scene." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "In this work, we build upon the blur formulation in Eq. (1), and explore two key differences between NeRF and 3DGS in robust reconstruction. NeRF-based approaches suffer from slow training speed, i.e. only limited number of rays can" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Peng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "be rendered at once. This is exacerbated by kernel modeling, which requires the rendering of a patch to model a single pixel. To mitigate this issue, current approaches model the kernel " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " sparsely [15, 21, 30] or only for specific types of blur [16, 41]. Since 3DGS can rasterize entire images efficiently, it potentially has great synergy with Eq. (1). On the flip side, NeRF requires only calibrated camera poses for reconstruction, whereas 3DGS also requires a good point cloud initialization. For noisy images, camera poses can often be estimated based on just a few correct matches, but a dense point cloud is challenging to obtain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "text", + "content": "As shown in Fig 1, we introduce Blur Agnostic Gaussian Splatting, or BAGS, which is a Splatting-based method that is robust against various types of blur. At the core of BAGS is a Blur Proposal Network (BPN). Contrary to previous methods [15,21,30] that sparsely estimate " + }, + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "text", + "content": ", BPN directly estimates dense kernels on full images during training. We find that estimating a dense " + }, + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "text", + "content": " is more effective and no less efficient than estimating a sparse one, which also has to predict kernel positions " + }, + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "text", + "content": ". BPN considers the position and view embedding of pixel " + }, + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 213, + 482, + 406 + ], + "type": "text", + "content": ", accounting for the spatial variance in image blur. Furthermore, a three-layer Convolution Neural Network (CNN) is used to compute the RGBD features of the rendered image. These extracted features allow BPN to consider the color and depth variation of the scene, e.g. around edges and corners, more effectively. Finally, BPN also computes a per-pixel mask, which blends the estimated clear and blurry pixels together. We make the observation that not necessarily all pixels are blurry; therefore, by constraining on its sparsity, such a mask can visualize regions where blur occurs. This is particularly useful as a way to evaluate image quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 407, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 550 + ], + "type": "text", + "content": "In addition to BPN, BAGS proposes a coarse-to-fine kernel optimization scheme. In some cases, BAGS cannot perform reconstruction well if it jointly optimizes BPN with a very sparse point cloud, due to the inherent ambiguity of separating blur from the observation. Furthermore, optimizing a dense kernel on full resolution directly is computationally expensive and time-consuming. To solve these conundrums, we first optimize the scene and BPN at lower scales with a smaller kernel dimension. This allows the optimization process to address a less challenging and ambiguous problem initially. We then gradually increase the resolution of the training images and the estimated kernels; since the blur problem has been partially addressed at a small scale, optimization at larger scales is more stable and effective. Under this scheme, we not only observe improved performance but also faster optimization process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 552, + 436, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 552, + 436, + 564 + ], + "spans": [ + { + "bbox": [ + 146, + 552, + 436, + 564 + ], + "type": "text", + "content": "In summary, our contributions can be summarized in three parts:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 579, + 479, + 662 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 579, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 579, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 136, + 579, + 479, + 613 + ], + "type": "text", + "content": "1. We introduce a Blur Proposal Network, which considers spatial, depth, and color variations of the scene to model image blur; BPN can also produce a mask that indicates blurry regions in an input image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 617, + 479, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 617, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 135, + 617, + 479, + 662 + ], + "type": "text", + "content": "2. We introduce a coarse-to-fine optimization scheme, which gradually increases the training image resolution and the estimated kernel size with additional neural network layers; this improves the stability of the joint optimization process given a sparse point cloud." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 135, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 480, + 152 + ], + "type": "text", + "content": "3. We evaluate the overall method, BAGS, on three image blur scenarios and find significant quantitative and visual improvements compared to current SoTA methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 163, + 337, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 337, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 337, + 178 + ], + "type": "text", + "content": "2 Related Work and Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 370 + ], + "type": "text", + "content": "Single Image Restoration. Restoring images from degradation, such as blur, has a long history of research. In general, this is a heavily ill-posed problem as many potential high quality images can lead to the same degraded observations. Traditionally, various priors are used to constrain the solution space [5, 14, 18, 34, 38, 43]. More recently, deep-learning-based methods have achieved great performances by learning a better prior in a data driven way, e.g. as a function that directly maps blurry images to clean images. In this paradigm, we train a powerful restoration model with millions of training pairs and a large neural network. Ideally, images can be restored by such a model [1, 4, 8, 13, 17, 42, 45, 47]. This paradigm has many issues in practice. For example, real training pairs are difficult to acquire in large scale; therefore, synthetic data is often used, which leads to significant domain shift problems. Every type of degradation also requires its own training sets and models, which exacerbate the data and computation costs. Even for degradations such as low resolution, for which is easy to synthesize training pairs, the restoration is unlikely to be multi-view consistent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 373, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 541 + ], + "type": "text", + "content": "NeRF in Non-Ideal Conditions. NeRF [24] works well when images are clean and well-calibrated; however, images acquired in the wild are often less ideal. Factors such as low light, camera motion, object motion, and incorrect focus can degrade image quality or affect multi-view consistency. Many works introduce domain knowledge to model the non-ideal contributors, such that the radiance field only models a canonical, multi-view consistent scene. RawNeRF [23] uses an approximated tonemapped loss and variable exposure to account for low light observations. NeRF-W [22] uses an image-specific embedding to model inconsistent appearances and transient objects. For object motion, many works [28,29,33,40] use an implicit function to describe scene deformation, such that the traced rays are first deformed. For image degradation such as blur, similar ideas have been used [11, 15, 21, 30, 41] to model the forward image degradation process; i.e. the 3D consistent scene and 2D inconsistent degradation will be separated based on the multi-view photometric loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 594 + ], + "type": "text", + "content": "3D Gaussian Splatting. Recently, 3DGS [12] introduces a novel, point-based scene reconstruction method that improves upon NeRF in many aspects. 3DGS utilizes a set of 3D Gaussian primitives [48] " + }, + { + "bbox": [ + 130, + 546, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_n|n = 1,\\dots ,N\\}" + }, + { + "bbox": [ + 130, + 546, + 482, + 594 + ], + "type": "text", + "content": " to represent the scene. These Gaussians are parameterized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 239, + 597, + 480, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 597, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 239, + 597, + 480, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {n} (\\mathbf {v}) = e ^ {- \\frac {1}{2} (\\mathbf {v} - \\mathbf {p} _ {n}) ^ {T} \\Sigma_ {n} ^ {- 1} (\\mathbf {v} - \\mathbf {p} _ {n})}, \\tag {2}", + "image_path": "cd9e8c4d8f507bb60f2b19d14f7c306cc63df6d58b26dd55745700eeb8e8abdc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "where the center " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_n\\in \\mathbb{R}^{3\\times 1}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Sigma_{n}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " define the position, scale, and rotation of the Gaussian at point " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}\\in \\mathbb{R}^{3\\times 1}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " in space. Additionally, each Gaussian also contains opacity " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ". For rendering, the Gaussians are projected onto the image plane as " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_n^{\\mathrm{2D}}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", where the center and covariance ma" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Peng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 137 + ], + "type": "text", + "content": "trices are transformed based on the camera rotation " + }, + { + "bbox": [ + 130, + 116, + 479, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 116, + 479, + 137 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 130, + 116, + 479, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^{3 \\times 1}" + }, + { + "bbox": [ + 130, + 116, + 479, + 137 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 135, + 480, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 135, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 235, + 135, + 480, + 149 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {n} ^ {\\mathrm {2 D}} = \\mathbf {R} \\mathbf {p} _ {n} + \\mathbf {t}, \\Sigma_ {n} ^ {\\mathrm {2 D}} = \\mathbf {R} \\Sigma_ {n} \\mathbf {R} ^ {T}. \\tag {3}", + "image_path": "936b086055b00dc4cbcc1ff208a043fb9d57779ef77b0fca1f8bd76f05b68a3c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "type": "text", + "content": "We can then aggregate 2D Gaussians according to their depth order " + }, + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "type": "text", + "content": " over a pixel/tile, based on alpha blending:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 183, + 481, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 211, + 183, + 481, + 215 + ], + "type": "interline_equation", + "content": "C (x) = \\sum_ {i = 1} ^ {D} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i} \\mathcal {G} _ {i} ^ {\\mathrm {2 D}}, T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} \\mathcal {G} _ {j} ^ {\\mathrm {2 D}}\\right). \\tag {4}", + "image_path": "54ffc7250ef04373c1fc6a5f600278d8f363936259a7cb9c30163e0293d9e202.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 218, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 479, + 266 + ], + "type": "text", + "content": "With well-designed tiling, this splatting-based rasterization technique is significantly faster than ray-tracing in NeRF-based methods without affecting rendering quality; various work have ensued to further improve the Gaussian representation [10, 26, 27, 44] and the rasterization process [20, 46]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 277, + 353, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 353, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 353, + 292 + ], + "type": "text", + "content": "3 Blur Agnostic Gaussian Splitting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "text", + "content": "BAGS is designed to robustly optimize 3D Gaussians by introducing additional modeling capacities in 2D, which allows BAGS to model away 3D-inconsistent blur from the scene. As shown in Fig. 2, BAGS consists of two parts: a Blur Proposal Network (BPN) and a coarse-to-fine optimization scheme. For clarity, we first describe the design of BPN on a single scale, and then describe modifications on BPN and the training loss on scale " + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "text", + "content": " for multi-scale training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 382, + 278, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 278, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 278, + 395 + ], + "type": "text", + "content": "3.1 Blur Proposal Network" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "text", + "content": "Following the forward blur model described in Eq. (1), BPN models a per-pixel convolution kernel " + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "inline_equation", + "content": "h(x) \\in \\mathbb{R}^{K \\times K}" + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "text", + "content": " for every pixel " + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "text", + "content": " is the kernel size. The rasterization efficiency in Gaussian Splatting allows us to model " + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "inline_equation", + "content": "h(x)" + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "text", + "content": " as a full convolution kernel. Additionally, BPN estimates a per-pixel scalar mask " + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "inline_equation", + "content": "m(x) \\in [0, 1]" + }, + { + "bbox": [ + 130, + 403, + 479, + 464 + ], + "type": "text", + "content": ", controlling the areas where blur modeling takes place." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": "Multi-Modal Features. In scene reconstruction, a large neural network will impact training time. As such, we use a four-layer MLP " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{kernel}}" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " to model blur. To maximize the capacity of " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{kernel}}" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": ", we design Multi-Modal Features (MMF) that can help distinguish the scene from blur. Real image blur typically varies between different observed images and different pixels within an image; therefore, we provide " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{kernel}}" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " with a learnable view embedding " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "l(i)" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " and a positional embedding " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " for training view " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " and coordinate " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": ". Additionally, we provide " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{kernel}}" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " with a set of features " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{RGBD}}" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " extracted from the rendered image " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " from rasterization, using a small, three-layer CNN " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": ". These features are concatenated with " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "l(i)" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " to form the MMF for " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{kernel}}" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": ". Consequently, the forward process to estimate the kernel " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "h(x,i)" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " and the mask " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "m(x,i)" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " given a specific pixel " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " and training view " + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 468, + 481, + 611 + ], + "type": "text", + "content": " can be noted as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 196, + 613, + 418, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 613, + 418, + 626 + ], + "spans": [ + { + "bbox": [ + 196, + 613, + 418, + 626 + ], + "type": "interline_equation", + "content": "h (x, i), m (x, i) = \\mathcal {F} _ {\\mathrm {k e r n e l}} (l (i) \\oplus p (x) \\oplus f _ {\\mathrm {R G B D}} (x, i)),", + "image_path": "b4c8ea9827ce0da08fba8c9a0dd92133f20cd570479bc1cc3cb6c1ea9add539f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 184, + 624, + 430, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 624, + 430, + 639 + ], + "spans": [ + { + "bbox": [ + 184, + 624, + 430, + 639 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} f _ {\\mathrm {R G B D}} (x, i) = \\mathcal {F} _ {\\text {f e a t}} (C (x, i) \\oplus D (x, i)), 0 \\leq m \\leq 1.", + "image_path": "239346e7ea7bd133e0635323888eb37cf9fd2a26e150934051bfb260ee319e28.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Incorporating RGBD features enbales BPN to explore the patch statistics around the current pixel, especially at edges and corners where blur is prominent. Scene" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 482, + 277 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 277 + ], + "type": "image", + "image_path": "b016e80466ffca5bd1f2417ee88c805e96dacf91e99ab72f453121fbfb55acce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": "Fig. 2: BAGS is optimized by training a Blur Proposal Network on top of the scene " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\mathbf{v})" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " over multiple scales. Top: We extract " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{RGBD}}^s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " from color and depth, which are concatenated with the position and view embedding " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "p(x), l(i)" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " to form the Multi-Modal Feature (MMF). The kernel MLP then estimates the per-pixel kernel " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "h^s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " and mask " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "m^s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "h^s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " to model the blur image " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\tilde{C}^s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " and employ " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "m^s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " to blend the rendered image " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "C^s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " and the blur-modeled image " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\tilde{C}^s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{out}}^s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": ". Bottom: After " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " steps, we upscale image resolution and modify the kernel MLP to produce " + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "inline_equation", + "content": "h^{s - 1}" + }, + { + "bbox": [ + 130, + 286, + 482, + 365 + ], + "type": "text", + "content": " with a larger kernel size." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 387, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 461 + ], + "type": "text", + "content": "depth also strongly correlates with the magnitude of blur. For example, defocus blur affects pixels that are outside of a specific focus plane; scene content that is further away from this focus plane becomes more blurry. For camera motion blur, pixels in the near plane may shift more than pixels in the far plane. We find the inclusion of " + }, + { + "bbox": [ + 130, + 387, + 482, + 461 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{RGBD}}" + }, + { + "bbox": [ + 130, + 387, + 482, + 461 + ], + "type": "text", + "content": " notably improves BPN's modeling capacity and the quality of novel view synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "text", + "content": "Residual Mask Bottleneck. As described in Eq. (5), " + }, + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{kern}}" + }, + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "text", + "content": " also estimates a per-pixel scalar mask " + }, + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "inline_equation", + "content": "m(x,i)" + }, + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "text", + "content": ". Specifically, this mask is used to blend the Gaussian-splatted image " + }, + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "text", + "content": " with the blur-modeled image " + }, + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\tilde{C}" + }, + { + "bbox": [ + 130, + 480, + 482, + 518 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 197, + 525, + 481, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 525, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 197, + 525, + 481, + 540 + ], + "type": "interline_equation", + "content": "C _ {\\mathrm {o u t}} (x, i) = \\left(1 - m (x, i)\\right) C (x, i) + m (x, i) \\tilde {C} (x, i), \\tag {6}", + "image_path": "2923de4470a7afc009f503188b266680cc4f3d9d502b5bcc6d785731c4fe4386.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{out}}" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": " is the final output at training. We note that not necessarily all pixels are blurry in an observed image. The blending " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": " provides an additional degree of freedom to allow " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": ", the clear pixel, to be directly compared with " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{obs}}" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": "; i.e., if " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{obs}}(x,i)" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": " suffers from blur, then " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "m(x,i)" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": " should be large, giving a higher weight to " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\hat{C}(x,i)" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": "; if " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{obs}}(x,i)" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": " if free of blur, " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "m(x,i)" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": " should be small, favoring " + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "inline_equation", + "content": "C(x,i)" + }, + { + "bbox": [ + 130, + 548, + 482, + 611 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "This design is similar to residual learning by providing the output with a skip connection from " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " as initialization. We apply a small sparsity constraint on " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " to encourage a larger weight towards " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " by default. After training, we can extract " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " to automatically identify the degraded regions in an input image." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Peng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 220, + 172 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 220, + 172 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 220, + 172 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 220, + 172 + ], + "type": "image", + "image_path": "59011e96dc7eeaa4cb97086bdd07afed2a6c8f42bac595790d90bc3099cd39f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 174, + 208, + 182 + ], + "lines": [ + { + "bbox": [ + 146, + 174, + 208, + 182 + ], + "spans": [ + { + "bbox": [ + 146, + 174, + 208, + 182 + ], + "type": "text", + "content": "(a) Without C2F" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 114, + 306, + 172 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 306, + 172 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 306, + 172 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 306, + 172 + ], + "type": "image", + "image_path": "e9c49ff277167511f89fcbd1b36f4d882ea98fdfd77582fd3ef9ed1c53dc386a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 174, + 283, + 182 + ], + "lines": [ + { + "bbox": [ + 246, + 174, + 283, + 182 + ], + "spans": [ + { + "bbox": [ + 246, + 174, + 283, + 182 + ], + "type": "text", + "content": "(b) BAGS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 185, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 185, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 482, + 240 + ], + "type": "text", + "content": "Fig. 3: Gaussians may get stuck at a local minimum without proper initialization. As indicated in 3a, optimizing " + }, + { + "bbox": [ + 130, + 185, + 482, + 240 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 185, + 482, + 240 + ], + "type": "text", + "content": " directly can lead to noisy surfaces. Naively densifying the scene before adding " + }, + { + "bbox": [ + 130, + 185, + 482, + 240 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 185, + 482, + 240 + ], + "type": "text", + "content": " can also lead to noisy Gaussians; as shown in 3c, the noisy Gasussians are not well removed even after adding " + }, + { + "bbox": [ + 130, + 185, + 482, + 240 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 185, + 482, + 240 + ], + "type": "text", + "content": ". By using a coarse-to-fine training schedule, we achieve better results in 3b and 3d." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 394, + 172 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 394, + 172 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 394, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 394, + 172 + ], + "type": "image", + "image_path": "7ecc8a9b246519d1859d7ebace6e8b20c303e3c1bc3b803e0be6e9eee12c38f0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 174, + 391, + 182 + ], + "lines": [ + { + "bbox": [ + 310, + 174, + 391, + 182 + ], + "spans": [ + { + "bbox": [ + 310, + 174, + 391, + 182 + ], + "type": "text", + "content": "(c) Naive densification" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 394, + 114, + 481, + 172 + ], + "blocks": [ + { + "bbox": [ + 394, + 114, + 481, + 172 + ], + "lines": [ + { + "bbox": [ + 394, + 114, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 394, + 114, + 481, + 172 + ], + "type": "image", + "image_path": "f8cde32d8cbbf6de12e1f5119189dc5016ef47e701cf81aba73020f40f491aaf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 174, + 457, + 182 + ], + "lines": [ + { + "bbox": [ + 419, + 174, + 457, + 182 + ], + "spans": [ + { + "bbox": [ + 419, + 174, + 457, + 182 + ], + "type": "text", + "content": "(d) BAGS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 255, + 342, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 342, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 342, + 268 + ], + "type": "text", + "content": "3.2 Coarse-to-Fine Kernel Optimization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": "A high quality 3DGS reconstruction requires a good point cloud estimation from Structure from Motion (SfM). Camera calibration parameters, which can be estimated from just a few matches, are relatively robust against image blur. On the other hand, obtaining a sufficiently dense point cloud can be challenging. Unfortunately, 3DGS is prone to over-fitting without a good initialization. This problem can worsen when a blur model is introduced and the kernels " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " provide additional degrees of freedom in scene optimization, as shown in Fig. 3. Furthermore, estimating dense per-pixel kernels on a full resolution image can be computationally expensive. A sparse estimation of the kernel does not meaningfully improve efficiency, as rendering is cheap in rasterization while the neural network to predict any type of kernel is still costly. Fundamentally, the complexity to compute per-pixel kernel scales quadratically with resolution. To address these problems, we propose a coarse-to-fine kernel optimization scheme, which initially optimizes Eq. (5) at a low resolution and gradually increases render resolution, achieving a more stable training process and better efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "text", + "content": "Multi-Scale BPN. For multi-scale optimization, BPN should generate kernels with a consistent field of view. We define scale based on the downscaling ratio " + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "inline_equation", + "content": "2^{s - 1}" + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "text", + "content": " in resolution, where the original resolution is defined as " + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "text", + "content": ". Additionally, if the image resolution is downscaled by a factor of two, i.e. " + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "inline_equation", + "content": "s = 2" + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "text", + "content": ", the proposed kernel size " + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "inline_equation", + "content": "K^s" + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "text", + "content": " should also be similarly downscaled. As shown in Fig. 2, a proper " + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "inline_equation", + "content": "K^s" + }, + { + "bbox": [ + 130, + 465, + 482, + 537 + ], + "type": "text", + "content": " ensures that the kernel is always modeling blur from the same regions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": "To enable multi-scale training, we separate " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{kernel}}^s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": " into two parts: a two-layer base MLP " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{base}}^s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": " and two one-layer MLP heads " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{blur}}^s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{mask}}^s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": ". The base MLP extracts features from MMF, such features are then provided to the two heads for generating the kernels " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "h^s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": " and masks " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "m^s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": " at scale " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": ". In practice, only " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{blur}}^s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": " has a scale-dependent parameter size corresponding to the kernel size " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "K^s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": "; therefore, a new MLP layer with output dimension " + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "inline_equation", + "content": "K^s \\times K^s" + }, + { + "bbox": [ + 130, + 540, + 482, + 623 + ], + "type": "text", + "content": " is added for every scale. The updated multi-scale BPN can be described as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 208, + 624, + 405, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 624, + 405, + 638 + ], + "spans": [ + { + "bbox": [ + 208, + 624, + 405, + 638 + ], + "type": "interline_equation", + "content": "f _ {\\text {i n t e r}} (x, i) = \\mathcal {F} _ {\\text {b a s e}} ^ {s} (l (i) \\oplus p (x) \\oplus f _ {\\text {R G B D}} (x, i))", + "image_path": "8413222dcaa4fd929f06fef8e24749a6cd9e737340e3a30884f9e08d8275b478.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 225, + 639, + 481, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 639, + 481, + 653 + ], + "spans": [ + { + "bbox": [ + 225, + 639, + 481, + 653 + ], + "type": "interline_equation", + "content": "h ^ {s} (x, i) = \\operatorname {s o f t m a x} \\left(\\mathcal {F} _ {\\text {b l u r}} ^ {s} \\left(f _ {\\text {i n t e r}} (x, i)\\right)\\right), \\tag {7}", + "image_path": "ae58b45f67628557776933441c2f238e74504d8cbc7043b713c8f8121b90575c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 221, + 654, + 392, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 654, + 392, + 667 + ], + "spans": [ + { + "bbox": [ + 221, + 654, + 392, + 667 + ], + "type": "interline_equation", + "content": "m ^ {s} (x, i) = \\operatorname {s i g m o i d} \\left(\\mathcal {F} _ {\\text {m a s k}} ^ {s} \\left(f _ {\\text {i n t e r}} (x, i)\\right)\\right).", + "image_path": "4f99a6df1bddad1abcb2dbdedee35ba3b334a72197272ab9ef033973bd2b0dfe.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "Coarse-to-Fine Optimization. We split training into multiple stages based on scale. Starting from the coarsest scale " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "s = S" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " to the finest scale " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ", we optimize BAGS with a downsampling factor of " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "2^{s - 1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " applied to " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "C, \\tilde{C}, C_{\\mathrm{obs}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ". The optimization for each scale lasts for " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " steps, we then increase the resolution in " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "C, \\tilde{C}, C_{\\mathrm{obs}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ". The scale-dependent training loss can be described as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 177, + 481, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 177, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 164, + 177, + 481, + 190 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\lambda_ {\\mathrm {p h o t o}} \\| C _ {\\mathrm {o u t}} ^ {s} - C _ {\\mathrm {o b s}} ^ {s} \\| + \\lambda_ {\\mathrm {D S}} \\mathcal {L} _ {\\mathrm {D - S S I M}} \\left(C _ {\\mathrm {o u t}} ^ {s}, C _ {\\mathrm {o b s}} ^ {s}\\right) + \\lambda_ {\\mathrm {m a s k}} \\| m ^ {s} \\|, \\tag {8}", + "image_path": "7a6aeae7d1c8ba3d4fabc20607dd624f13749f318574b905ac20226344bf672c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{obs}}^s" + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{out}}^s" + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "text", + "content": " are the observed and predicted image at scale " + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{D - SSIM}}" + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "text", + "content": " is the structural similarity loss; " + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\| m^{s}\\|" + }, + { + "bbox": [ + 130, + 192, + 482, + 276 + ], + "type": "text", + "content": " is the mask sparsity loss. We find this training scheme achieves the best of both worlds: activating BPN at low resolution generates less overfitting artifacts during densifying the point cloud given the reduced visibility of blur. Moreover, it also serves as an efficient warm-up for BPN at higher resolution, as computing per-pixel kernels at low resolution is much cheaper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 286, + 230, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 230, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 230, + 300 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": "We evaluate BAGS under three degradation scenarios: camera motion blur, defocus blur, and mix resolution. All of these degradations are commonly observed in images acquired in real life or online. We further introduce a new unbounded 360 drone dataset, which is collected at sunset or night and experiences degradation due to a combination of motion blur and low light condition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 371, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 456 + ], + "type": "text", + "content": "Camera Motion and Defocus Blur. Camera motion and defocus blur are often observed in real acquisitions, e.g. due to low light, incorrect auto-focus, or limitation in the finite aperture of real cameras. Deblur-NeRF [21] provides real scene acquisitions that contain camera motion and defocus blur; both blur categories consist of ten scenes. While real world scenes have blur-free references, the camera settings, e.g., exposure, may not be consistent with source views. Experiments on synthetic blur are reported in supplemental material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 459, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 567 + ], + "type": "text", + "content": "Mix Resolution. To analyze the scenario where the input images are of inconsistent resolution, we apply various downsampling operations on the Mip-NeRF 360 dataset [2]. Specifically, we split the training images into four equal parts and apply 4X, 3X, 2X, and no downscaling to each of the parts. The image splits are sampled in uniform randomness to ensure no downscaling covers a consecutive interval of the camera trajectory. We then use NeRFStudio's data processing pipeline [39], which converts all images to the highest resolution and calibrates with COLMAP [35]. Similar to previous works [2,3,46], we perform the entire process on downscaled data instead of the original 4K resolution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": "Unbounded Drone Motion Blur. The dataset introduced by DeblurNeRF [21] serves as a good benchmark for performance comparison; however, it is collected in a forward bounded format. These forward bounded scenes focus on a small area and have significant overlaps within the images; as such, it is unclear if the methodology [15, 16, 21, 30, 41] based on these sets can generalize to other scenarios. In fact, none of the deblurring methods can work for unbounded scenes. To supplement DeblurNeRF's dataset and understand how BAGS performs in realistic settings, we collect five sets of drone footage on different buildings in 360" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Peng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 305, + 195 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 305, + 195 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 305, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 305, + 195 + ], + "type": "image", + "image_path": "037523c94d0a7112955b58b87388cbcb0db060bfa5e00227677f55c2271bbe27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 475, + 196 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 475, + 196 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 475, + 196 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 475, + 196 + ], + "type": "image", + "image_path": "ea7ccc980b07eef8189fdc13f9551f029dba350f8a929945f62ee895438e041e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 198, + 305, + 281 + ], + "blocks": [ + { + "bbox": [ + 138, + 198, + 305, + 281 + ], + "lines": [ + { + "bbox": [ + 138, + 198, + 305, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 305, + 281 + ], + "type": "image", + "image_path": "50167ccbb2b1ac244c6d0f66fcbdae44a4c4ad0c964f51ed80ee18a555e19d12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 284, + 228, + 293 + ], + "lines": [ + { + "bbox": [ + 217, + 284, + 228, + 293 + ], + "spans": [ + { + "bbox": [ + 217, + 284, + 228, + 293 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 198, + 475, + 281 + ], + "blocks": [ + { + "bbox": [ + 308, + 198, + 475, + 281 + ], + "lines": [ + { + "bbox": [ + 308, + 198, + 475, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 475, + 281 + ], + "type": "image", + "image_path": "4bfb76c4ba18ea9b7865c6aec613cbb509ec68b02498a756cd7cc766489dcb90.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 284, + 398, + 293 + ], + "lines": [ + { + "bbox": [ + 386, + 284, + 398, + 293 + ], + "spans": [ + { + "bbox": [ + 386, + 284, + 398, + 293 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 295, + 480, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 295, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 480, + 350 + ], + "type": "text", + "content": "Fig. 4: Ablation study of BAGS with different sub-modules. In 4a and 4b, introducing RGBD features and coarse-to-fine optimization improve novel view synthesis quality, and alternative sparse deformable kernel suffers in performance; 4d demonstrates our speed improvement against " + }, + { + "bbox": [ + 130, + 295, + 480, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{BAGS}^{\\mathrm{noC2F}}" + }, + { + "bbox": [ + 130, + 295, + 480, + 350 + ], + "type": "text", + "content": ". In 4c, larger kernel generally leads to better performance, other than in the mix resolution scenario." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 361, + 480, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 421 + ], + "type": "text", + "content": "unbounded format. These footages are collected at sunset or night, and suffer from camera motion blur and low light noise. For each scene, we also capture clear frames from a stable drone position for testing, resulting in around 100 images overall. We use NeRFStudio [39] to calibrate images from these footages. More details on this dataset can be found in supplemental material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "text", + "content": "Implementation Details. BPN's " + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "text", + "content": " is constructed as a three-layer CNN; the intermediate channel size is 64, the kernel size for each layer is 5, 5, and 3. " + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{kernel}}" + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "text", + "content": " is constructed with four linear layers; the intermediate channel size is 64. For multi-scale training, we use three scales starting with scale " + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "inline_equation", + "content": "s = 3" + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "inline_equation", + "content": "K_{3} = 5" + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "inline_equation", + "content": "K_{2} = 9" + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "inline_equation", + "content": "K_{1} = 17" + }, + { + "bbox": [ + 130, + 426, + 480, + 521 + ], + "type": "text", + "content": ". BAGS implements its Gaussian Splatting and rasterization based on Mip-Splatting [46], which has been shown to yield good results across multiple resolutions. More implementation details can be found in supplemental material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 532, + 246, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 246, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 246, + 542 + ], + "type": "text", + "content": "4.1 Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": "We compare three designs in BAGS to demonstrate the effectiveness of different components:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 578, + 479, + 636 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 578, + 463, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 463, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 463, + 590 + ], + "type": "text", + "content": "- BAGS" + }, + { + "bbox": [ + 138, + 578, + 463, + 590 + ], + "type": "inline_equation", + "content": "^{\\text{noRGBD}}" + }, + { + "bbox": [ + 138, + 578, + 463, + 590 + ], + "type": "text", + "content": ": We implement BPN without the RGBD features " + }, + { + "bbox": [ + 138, + 578, + 463, + 590 + ], + "type": "inline_equation", + "content": "f_{\\text{RGBD}}" + }, + { + "bbox": [ + 138, + 578, + 463, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 590, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 479, + 613 + ], + "type": "text", + "content": "- BAGS" + }, + { + "bbox": [ + 138, + 590, + 479, + 613 + ], + "type": "inline_equation", + "content": "^{\\text{deform}}" + }, + { + "bbox": [ + 138, + 590, + 479, + 613 + ], + "type": "text", + "content": ": We implement BPN with a sparse deformable kernel, similar to DeblurNeRF [21]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 137, + 614, + 479, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 614, + 479, + 636 + ], + "spans": [ + { + "bbox": [ + 137, + 614, + 479, + 636 + ], + "type": "text", + "content": "- BAGS" + }, + { + "bbox": [ + 137, + 614, + 479, + 636 + ], + "type": "inline_equation", + "content": "^{\\mathrm{noC2F}}" + }, + { + "bbox": [ + 137, + 614, + 479, + 636 + ], + "type": "text", + "content": ": We implement BAGS without a coarse-to-fine kernel optimization scheme." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "We further investigate the effect of the final kernel size " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "K_{1}" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " on reconstruction quality across all degradation scenarios." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": "As shown in Fig. 4, the introduction of RGBD features to BPN greatly improves its ability to disentangle image blur from the underlying scene. The addition of " + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{RGBD}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": " increases the average performance by " + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "1.57 / 0.92" + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": " dB in PSNR and " + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "0.03 / 0.03" + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": " in SSIM for camera motion and defocus blur, respectively. This agrees with our intuition that blur is highly correlated with depth and image features like edges. Our coarse-to-fine optimization scheme also notably boosts robustness against defocus blur by " + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "1.64" + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": " dB and " + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "0.08" + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": " in PSNR and SSIM on average. As visualized in Fig. 3a, directly optimizing 3DGS on a sparse point cloud in " + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{BAGS}^{\\mathrm{noC2F}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": " leaves many ambiguities; while training views have low loss, test views show a lack of the correct surface structure and many artifact Gaussians. Densifying Gaussians without BPN first improves performance for defocus blur; however, this strategy is bad for motion blur, as shown in Fig. 3c. In Fig. 4d, we show that our coarse-to-fine strategy is not only performant, but also fast; on average, BAGS can reach high performance in less than fifteen minutes, which is much faster than the two hours required by PDRF [30]. This strategy is also faster compared to single-scale training, as full-scale training in our scheme uses less iterations, and lower scale training does not sacrifice performance. Both multi-scale and single-scale training take " + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "40 \\mathrm{K}" + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": " iterations overall." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": "We also implement a deformable version of BPN in " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{BAGS}^{\\mathrm{deform}}" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": ", where BPN estimates the kernel position and kernel weights for ten points within a " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "20 \\times 20" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": "-pixel patch. We show its performances in Fig. 4a, where " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{BAGS}^{\\mathrm{deform}}" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " is tested on camera motion blur. Camera motion blur often has a sparse representation, i.e. the camera motion trajectory [16, 41]; therefore, a ten-point estimation of the kernel may perform well. However, we find that " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{BAGS}^{\\mathrm{deform}}" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " still underperforms compared to a full kernel estimation in BAGS. This may be due to the difference between NeRF and 3DGS; i.e., a ray-based deformation can be better expressed in NeRF, while a patch-based convolution can be better expressed in 3DGS. We also note that " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{BAGS}^{\\mathrm{deform}}" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " is not more efficient than BAGS despite its sparse estimation. This is because approaches like [15, 21, 30] require one forward pass to estimate one kernel position and weight; therefore, we need to forward multiple times for every pixel to estimate the sparse kernel. In comparison, BAGS only requires a single forward pass to obtain the full kernel." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 499, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 573 + ], + "type": "text", + "content": "Finally, we ablate on the kernel dimension estimated by BPN. In Fig. 4c, we show that larger kernels in general lead to higher performances; interestingly, this is not true in the mixed resolution scenario. This is likely because downscaling is uniform across all pixels in an image, i.e. a downscaling factor of four means a kernel size of five can fully explain the degradation. As such, increasing the kernel size does not improve performance and introduces noise instead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 582, + 282, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 282, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 282, + 594 + ], + "type": "text", + "content": "4.2 Quantitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Camera Motion and Defocus Blur. As shown in Table 1, we observe that Gaussian Splatting [46] performs markedly worse than vanilla NeRF in the presence of blurry images, despite its strong performance in other curated benchmarks. Since NeRF models a continuous neural volume with an MLP, individual degradation can be smoothly explained by spreading the error into empty space." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Peng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 479, + 306 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 306 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 306 + ], + "type": "table", + "html": "
Camera MotionNeRF [24]Mip-Splatting [46]Deblur-NeRF [21]DP-NeRF [15]PDRF [30]BAGS
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Ball24.080.6240.39923.220.6190.34027.360.7660.22327.200.7650.209
Basket23.720.7090.32223.240.6880.28827.670.8450.14827.740.8460.129
Buick21.590.6330.35021.460.6580.26624.770.7700.17525.700.7920.141
Coffee26.480.8060.29024.730.7490.28830.930.8980.12432.440.9150.101
Decoration22.390.6610.36320.550.6410.29924.190.7710.18623.510.7400.275
Girl20.070.7080.32019.870.7140.27822.270.7980.16921.800.7700.205
Heron20.500.5220.41319.430.5050.33222.630.6870.21022.520.6760.281
Parterre23.140.6200.40522.280.5900.32125.820.7600.21624.970.7680.227
Puppet22.090.6090.33922.050.6310.26725.240.7510.15825.120.7580.181
Stair22.870.4560.48721.910.4740.38725.390.6300.21025.780.6470.229
Average22.690.6350.36921.870.6270.30625.630.7670.18225.910.7750.160
Defocus
Cake24.420.7210.22522.170.6450.29026.270.7800.12826.160.7780.127
Caps22.730.6310.28021.000.5030.44223.870.7130.16123.950.7120.143
Cisco20.720.7220.12620.080.7090.16320.830.7270.08720.730.7260.084
Coral19.810.5660.26919.600.5550.32319.850.6000.12122.800.7410.096
Cupcake21.880.6810.21621.550.6810.21322.260.7220.11620.110.6110.118
Cups25.020.7580.23220.930.6460.31426.210.7990.12726.750.8140.104
Daisy22.740.6200.26221.590.6140.27323.520.6870.12123.790.6970.108
Sausage17.790.4830.27917.780.4730.28418.010.5000.18018.350.5440.147
Seal22.790.6270.26822.130.5920.31026.040.7770.10525.950.7780.103
Tools26.080.8520.15523.980.8120.19627.810.8950.06128.070.8980.054
Average22.400.6660.23121.080.6230.28123.470.7200.12123.670.7300.108
", + "image_path": "df5ca05d0a6e89289d13f1c4dee31b237031eaf4f509313749c57ac656d2d3b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 308, + 479, + 330 + ], + "lines": [ + { + "bbox": [ + 132, + 308, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 479, + 330 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons on the camera motion and defocus blur dataset. We color code the best PSNR, SSIM, and LPIPS performances." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 346, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 441 + ], + "type": "text", + "content": "On the other hand, 3DGS is based on a discrete representation and tends to produce concentrated artifacts. Our BAGS can achieve new SoTA performances compared to previous NeRF-based deblurring methods despite the unfavorable performances in 3DGS. Specifically, we find significant improvements on camera motion blur across all metrics and achieve a fifty percent improvement in LPIPS on average. As visualized in Fig. 5, BAGS achieves much sharper results compared to previous methods [15, 16, 21, 30], to the point where it is hard to distinguish the groundtruth image from our rendered results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 443, + 480, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 480, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 480, + 527 + ], + "type": "text", + "content": "While improvements on defocus blur is less drastic in Table 1, we still observe great visual improvements. Several reasons contribute to the less apparent quantitative improvements. First, the collected groundtruth images have different exposure compared to training images; as such, image-wise similarity metrics like PSNR are less reliable. Secondly, many scenes contain specular surfaces, which all methods do a bad job at reconstructing. As shown in Fig. 5, much better details can be recovered through BAGS than other methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 604 + ], + "type": "text", + "content": "Mix Resolution and Low Light Motion Blur. We investigate the robustness of several SoTA reconstruction methods on downscaling perturbation and motion blur observed in low light condition. The experiments are done on unbounded scenes and are more challenging and realistic than the forward bounded scenes in DeblurNeRF [21]. We note that previous approaches [15, 21, 30, 41] cannot handle unbounded scenes; therefore, we do not include them in Table 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "As shown in Table 2, Mip Splating performs on par with Mip-NeRF 360; both are vastly better than NeRFacto [39], which stores hash features [25] to accelerate the training process and uses a very small MLP represent the scene. We also employ a Super-Resolution (SR) model HAT [7] on downscaled images in the hope that single-image SR can improve performances. However, pre-trained" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 363 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 363 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 363 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 363 + ], + "type": "image", + "image_path": "1359a9896b09a52e78e39c9fcd739a2e436e7770f845c14d15e16fb6c440ce1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 373, + 480, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 373, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 480, + 396 + ], + "type": "text", + "content": "Fig. 5: Visualizations of test views on camera motion and defocus blur dataset. Mip-Sp and Db-NeRF are short for Mip-Splatting [46] and Deblur-NeRF [21]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 407, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 479, + 479 + ], + "type": "text", + "content": "SR models are limited by domain shifts, dedicated upscaling factors, and the lack of multi-view consistency. As visualized in Fig. 6, while the render is somewhat sharper after SR, the color becomes unfaithful. Compared with all other methods, our rendered results are much sharper in details. This robust performance comes from BPN's ability to model Gaussian blur on high resolution before comparing them with the downscaled observations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 481, + 480, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 480, + 601 + ], + "type": "text", + "content": "Motion blur is often observed in low light condition. To this end, we perform experiments on scenes acquired near sunset or at night. As shown in Table 2, BAGS again improves significantly over other reconstruction approaches, which shows that our method can be generalized to different imaging geometry. As shown in Fig. 6, we employ MPR [47], a single image deblurring model, on the motion-blurred images before reconstructing with Mip-Splatting [46]. Similar to previous attempts, the deblurring prior leads to a slightly clearer rendering but is not sufficient to remove all blur. BAGS reconstructs the scene remarkably well compared to other methods, where even the window details can be recovered correctly. Please refer to the supplemental material for more visualizations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Kernel and Mask Visualization. BAGS models image degradation through convolution kernels " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " and masks " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": ", which we visualize in Fig. 7. By constraining on its sparsity, we can see that " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " meaningfully highlights the regions of blur in a given training image. On the other hand, low-value regions indicate either high quality observation or textureless regions. By visualizing the kernels modeled" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Peng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 479, + 304 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 304 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 304 + ], + "type": "table", + "html": "
MixedResNeRFacto [39]Mip-NeRF 360 [2]Mip-Sp [46]Mip-Sp-HAT [7]BAGS
PSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPS
Bicycle19.250.3490.49623.820.6100.37524.030.6280.36423.20
Bonsai21.800.6190.25929.910.8710.19029.270.8760.19328.38
Counter24.070.7140.28028.090.8480.22027.790.8530.21627.02
Flowers19.630.4180.45921.980.5660.38921.960.5700.39720.81
Garden21.610.4890.35725.340.6670.31725.280.6820.30523.96
Kitchen22.730.5840.30427.820.7790.23827.490.7780.24126.10
Room24.680.7910.21430.380.9020.19830.430.9030.20329.61
Stump23.190.5550.37725.750.6880.32025.870.7030.29924.09
Treehill20.110.4160.50122.520.5580.43722.860.5500.45721.76
Average21.900.5480.36126.180.7210.29826.110.7270.29724.99
Lowlight-droneNeRFacto [39]Mip-NeRF 360 [2]Mip-Sp [46]Mip-Sp-MPR [47]BAGS
PSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPSPSNRSSIM LPIPS
Math-night21.330.5820.32629.510.8820.19129.610.8840.19929.15
Library-night24.480.7680.38126.380.8210.34926.170.8200.37025.89
BioMed-sunset23.760.6500.43924.480.6880.40224.850.7120.42624.78
Admin-sunset21.960.6210.40025.470.7570.35525.540.7640.36925.19
Aud-sunset24.670.7090.39825.900.7500.36826.330.7700.38626.77
Average23.240.6660.38926.350.7800.33326.500.7900.35026.36
", + "image_path": "3e45467730ec4141014c4cf905dbdee57131eaa03d517fc13a61ae4fc6f40fb4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 369, + 479, + 564 + ], + "blocks": [ + { + "bbox": [ + 132, + 306, + 480, + 340 + ], + "lines": [ + { + "bbox": [ + 132, + 306, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 480, + 340 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons on mixed resolution and low light motion blur in unbounded 360 geometry. Mip-Sp is short for Mip-Splatting [46]. Mip-Sp-HAT and Mip-Sp-HAT-MPR use prior models to restore images first before Mip-Splatting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 369, + 479, + 564 + ], + "lines": [ + { + "bbox": [ + 135, + 369, + 479, + 564 + ], + "spans": [ + { + "bbox": [ + 135, + 369, + 479, + 564 + ], + "type": "table", + "html": "
Novel ViewNeRFactoMipNeRF-360Mip-SpMip-Sp-HATOursG.T.
PSNR/SSIM21.61/.57624.09/.67025.11/.74624.41/.72228.46/.899MixedRes
PSNR/SSIM23.08/.41224.98/.56224.93/.59024.68/.58731.79/.949MixedRes
Novel ViewNeRFactoMipNeRF-360Mip-SpMip-Sp-MPROursG.T.
PSNR/SSIM17.86/.60818.33/.70418.57/.71418.51/.71824.47/.850LowLight
PSNR/SSIM22.40/.68325.36/.81325.51/.81727.03/.86731.75/.938LowLight
", + "image_path": "a150f63ba1c28b6ed69c90ce742a8964fa304dd8763c209ed722bd9fe0081097.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 574, + 479, + 586 + ], + "lines": [ + { + "bbox": [ + 132, + 574, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 479, + 586 + ], + "type": "text", + "content": "Fig. 6: Visualizations of test views on mix resolution and low light motion blur dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "at different pixel, we can also easily characterize the types of observed blur. Specifically, the estimated kernels in camera motion blur exhibit clear patterns of camera movement, while those for defocus blur show Gaussian-like distributions based on the pixel's distance from the focus plane. These self-emerged properties" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "in BAGS provide ways for us to automatically and precisely evaluate the quality of training images." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 154, + 220, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 154, + 220, + 213 + ], + "lines": [ + { + "bbox": [ + 133, + 154, + 220, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 154, + 220, + 213 + ], + "type": "image", + "image_path": "41c5d2ff34a34ecc86c2bdd8b8d8e61aaddeb70ddeed3818c36b9dd2b10b876b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 215, + 205, + 223 + ], + "lines": [ + { + "bbox": [ + 149, + 215, + 205, + 223 + ], + "spans": [ + { + "bbox": [ + 149, + 215, + 205, + 223 + ], + "type": "text", + "content": "Camera Motion" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 155, + 306, + 213 + ], + "blocks": [ + { + "bbox": [ + 220, + 155, + 306, + 213 + ], + "lines": [ + { + "bbox": [ + 220, + 155, + 306, + 213 + ], + "spans": [ + { + "bbox": [ + 220, + 155, + 306, + 213 + ], + "type": "image", + "image_path": "d108c2bd4038b2113bb4c1b36309890bbcb6224036c84d5389aba32cc486fcdb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 214, + 291, + 223 + ], + "lines": [ + { + "bbox": [ + 237, + 214, + 291, + 223 + ], + "spans": [ + { + "bbox": [ + 237, + 214, + 291, + 223 + ], + "type": "text", + "content": "Kernel & mask" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 229, + 473, + 241 + ], + "lines": [ + { + "bbox": [ + 138, + 229, + 473, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 229, + 473, + 241 + ], + "type": "text", + "content": "Fig. 7: Visualization of estimated convolution kernels and masks in different blur." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 155, + 394, + 213 + ], + "blocks": [ + { + "bbox": [ + 307, + 155, + 394, + 213 + ], + "lines": [ + { + "bbox": [ + 307, + 155, + 394, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 155, + 394, + 213 + ], + "type": "image", + "image_path": "f135d8835104fbccc2059f2138b0de441611621c5d191256ff298c631592f938.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 214, + 365, + 223 + ], + "lines": [ + { + "bbox": [ + 336, + 214, + 365, + 223 + ], + "spans": [ + { + "bbox": [ + 336, + 214, + 365, + 223 + ], + "type": "text", + "content": "Defocus" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 398, + 155, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 398, + 155, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 398, + 155, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 398, + 155, + 481, + 213 + ], + "type": "image", + "image_path": "71e0538c850d11aa89879fe1add6c250e005f017734edcf0c392a007149f40b9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 214, + 464, + 223 + ], + "lines": [ + { + "bbox": [ + 410, + 214, + 464, + 223 + ], + "spans": [ + { + "bbox": [ + 410, + 214, + 464, + 223 + ], + "type": "text", + "content": "Kernel & mask" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": "Discussion. BAGS can achieve great results; however, there also exists many potential improvements. The added neural networks and convolution operations require additional computation. While BAGS is a generalist approach and addresses multiple types of blur, it may become very expensive to estimate per-pixel convolution in high resolution. To this end, we have explored utilizing pixel shuffling [36], which converts spatial resolution into channels, before we estimate blur kernel; i.e., this assumes that within a small region, blur kernel " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": " stays roughly consistent. As shown in the supplemental material, we can scale to great visual results at 2K resolution on the DeblurNeRF dataset. Many other potential directions can also address this complexity issue, including leveraging the mask " + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 253, + 482, + 421 + ], + "type": "text", + "content": " to identify important regions to focus on, a better designed low rank kernel estimation, degradation-specific optimization, etc. As such, while BAGS is already more efficient than comparable NeRF-based methods, we anticipate that its computation cost can be further optimized in the future." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 437, + 219, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 219, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 219, + 450 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "We present BAGS, a novel scene reconstruction method that can handle various noise in input images. We demonstrate that a vanilla Gaussian Splatting is particularly susceptible to degraded images due to its discrete representation, even compared with NeRF. To address this issue, BAGS implements a 2D degradation model, BPN, which estimates convolution kernels jointly with 3D scene optimization. BPN allows for additional freedom to address 3D inconsistencies, and produces interpretable kernels and masks to indicate the degradation types and regions within an image. Additionally, BAGS leverages a coarse-to-fine kernel optimization scheme, which gradually models the 3D scene with higher resolution images and kernels. This approach successfully addresses the 2D-3D ambiguity in optimizing convolution kernels with a very sparse point cloud. We perform extensive experiments on a variety of noise in training images, including camera motion blur, defocus blur, and low resolution. Under these conditions, BAGS can render high fidelity scenes and significantly improves upon previous approaches. In the future, we hope to further improve the computational efficiency in BAGS and dynamically adjust kernel capacity based on how degraded the training views are." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Peng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 262 + ], + "type": "text", + "content": "This research is based upon work supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. The authors also thank the support and feedbacks from researchers, Kaiwen Jiang, Dr. Ravi Ramamoorthi, at UCSD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 282, + 197, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 197, + 295 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 197, + 295 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 309, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 309, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 481, + 353 + ], + "type": "text", + "content": "1. Abuolaim, A., Afifi, M., Brown, M.S.: Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1231-1239 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 354, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 376 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 377, + 481, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 377, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 377, + 481, + 409 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "type": "text", + "content": "4. Cao, Y., Liu, M., Liu, S., Wang, X., Lei, L., Zuo, W.: Physics-guided iso-dependent sensor noise modeling for extreme low-light photography. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5744-5753 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 454, + 481, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 476 + ], + "type": "text", + "content": "5. Chan, T.F., Wong, C.: Total variation blind deconvolution. IEEE Trans. Image Process. 7(3), 370-375 (1998)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 477, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 477, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 477, + 481, + 498 + ], + "type": "text", + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 499, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 481, + 532 + ], + "type": "text", + "content": "7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22367-22377 (June 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 533, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 564 + ], + "type": "text", + "content": "8. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4641-4650 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 566, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 481, + 598 + ], + "type": "text", + "content": "9. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 481, + 632 + ], + "type": "text", + "content": "10. Gao, J., Gu, C., Lin, Y., Zhu, H., Cao, X., Zhang, L., Yao, Y.: Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "11. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-nerf: High dynamic range neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18398-18408 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "12. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "13. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5886-5895 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 480, + 217 + ], + "type": "text", + "content": "14. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011. pp. 233-240. IEEE Computer Society (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 219, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 219, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 219, + 480, + 251 + ], + "type": "text", + "content": "15. Lee, D., Lee, M., Shin, C., Lee, S.: Dp-nerf: Deblurred neural radiance field with physical scene priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12386-12396 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 252, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 252, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 133, + 252, + 480, + 284 + ], + "type": "text", + "content": "16. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 286, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 286, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 133, + 286, + 480, + 318 + ], + "type": "text", + "content": "17. Lee, J., Son, H., Rim, J., Cho, S., Lee, S.: Iterative filter adaptive network for single image defocus deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2034-2042 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 319, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 319, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 133, + 319, + 480, + 373 + ], + "type": "text", + "content": "18. Levin, A.: Blind motion deblurring using image statistics. In: Scholkopf, B., Platt, J.C., Hofmann, T. (eds.) Advances in Neural Information Processing Systems 19, Proceedings of the Twentieth Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada, December 4-7, 2006. pp. 841-848. MIT Press (2006)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 375, + 480, + 407 + ], + "type": "text", + "content": "19. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "text", + "content": "20. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 443, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 443, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 443, + 480, + 475 + ], + "type": "text", + "content": "21. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 476, + 480, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 476, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 133, + 476, + 480, + 529 + ], + "type": "text", + "content": "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7210-7219. Computer Vision Foundation / IEEE (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 532, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 532, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 133, + 532, + 480, + 575 + ], + "type": "text", + "content": "23. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16190–16199 (June 2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 577, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 577, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 133, + 577, + 480, + 631 + ], + "type": "text", + "content": "24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I. Lecture Notes in Computer Science, vol. 12346, pp. 405-421. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Peng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "26. Navaneet, K., Meibodi, K.P., Koohpayegani, S.A., Piriyawash, H.: Compact3d: Compressing gaussian splat radiance field models with vector quantization. arXiv preprint arXiv:2311.18159 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "27. Niedermayr, S., Stumpfegger, J., Westermann, R.: Compressed 3d gaussian splatt-ting for accelerated novel view synthesis. arXiv preprint arXiv:2401.02436 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "28. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "29. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., MartinBrualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. 40(6) (dec 2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "30. Peng, C., Chellappa, R.: Pdrf: progressively deblurring radiance field for fast scene reconstruction from blurry images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2029-2037 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "31. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "text", + "content": "32. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "text", + "content": "33. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 10318-10327. Computer Vision Foundation / IEEE (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "34. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: Nonlinear Phenomena 60(1), 259-268 (1992)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "35. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "type": "text", + "content": "36. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "37. Sun, C., et al.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "38. Tai, Y.W., et al.: Richardson-lucy deblurring for scenes under a projective motion path. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(8), 1603-1618 (2011)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "text", + "content": "39. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Kerr, J., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., McAllister, D., Kanazawa, A.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "40. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: IEEE International Conference on Computer Vision (ICCV). IEEE (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "41. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "42. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "43. Xu, L., Zheng, S., Jia, J.: Unnatural L0 sparse representation for natural image deblurring. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition," + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 281 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "Portland, OR, USA, June 23-28, 2013. pp. 1107-1114. IEEE Computer Society (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "44. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "45. Yu, X., Dai, P., Li, W., Ma, L., Shen, J., Li, J., Qi, X.: Towards efficient and scale-robust ultra-high-definition image demoiring. In: European Conference on Computer Vision. pp. 646-662. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "46. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. arXiv preprint arXiv:2311.16493 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "48. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371-378 (2001)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Peng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_content_list.json b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b96b7505f4050e7cc53ac7317389eecee95cb7b --- /dev/null +++ b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_content_list.json @@ -0,0 +1,1980 @@ +[ + { + "type": "text", + "text": "BAM-DETR: Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos", + "text_level": 1, + "bbox": [ + 240, + 140, + 763, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pilhyeon Lee $^{1\\star}$ and Hyeran Byun $^{2}$", + "bbox": [ + 364, + 234, + 637, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Artificial Intelligence, Inha University", + "bbox": [ + 316, + 261, + 684, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, Yonsei University", + "bbox": [ + 316, + 277, + 681, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Temporal sentence grounding aims to localize moments relevant to a language description. Recently, DETR-like approaches achieved notable progress by predicting the center and length of a target moment. However, they suffer from the issue of center misalignment raised by the inherent ambiguity of moment centers, leading to inaccurate predictions. To remedy this problem, we propose a novel boundary-oriented moment formulation. In our paradigm, the model no longer needs to find the precise center but instead suffices to predict any anchor point within the interval, from which the boundaries are directly estimated. Based on this idea, we design a boundary-aligned moment detection transformer, equipped with a dual-pathway decoding process. Specifically, it refines the anchor and boundaries within parallel pathways using global and boundary-focused attention, respectively. This separate design allows the model to focus on desirable regions, enabling precise refinement of moment predictions. Further, we propose a quality-based ranking method, ensuring that proposals with high localization qualities are prioritized over incomplete ones. Experiments on three benchmarks validate the effectiveness of the proposed methods. The code is available here.", + "bbox": [ + 261, + 321, + 743, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Temporal sentence grounding $\\cdot$ Detection transformer", + "bbox": [ + 261, + 584, + 707, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 621, + 375, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed a notable surge in the popularity of short-form video content on social media platforms like TikTok, YouTube Shorts, and Instagram Reels. As such, users prefer to selectively engage with short moments of interest rather than passively watch an entire long video. This trend highlights the importance of localizing desired moments. As a result, moment localization tasks have emerged as pivotal research topics in video understanding, including temporal action detection [20, 29, 75], video summarization [50, 71], and highlight detection [48, 58]. Within this context, we tackle temporal sentence grounding [1, 72], aiming to retrieve moments corresponding to free-form language descriptions.", + "bbox": [ + 212, + 650, + 787, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address temporal sentence grounding, numerous efforts have been undertaken in the last decade [9, 11, 31, 49, 64, 72]. Especially, taking inspiration", + "bbox": [ + 212, + 787, + 785, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Correspondence to: Pilhyeon Lee ", + "bbox": [ + 217, + 824, + 660, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd3dbc8d5ee3eec6576c86a0597040e9692a12aa97821ba3cf81a4369a1ed038.jpg", + "image_caption": [ + "(a) Conventional moment modeling" + ], + "image_footnote": [], + "bbox": [ + 330, + 143, + 669, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/98965eb2948b6a3bd61e7f7cd5ca7f366452887af5193e63ad134e5ce0069035.jpg", + "image_caption": [ + "(b) Boundary-oriented moment modeling (Ours)", + "Fig. 1: Comparison of moment modeling approaches under the scenario of an ambiguous center from QVHighlights. (a) The conventional method formulates a moment with a tuple of $(c, l)$ . (b) In contrast, we propose to model it with a triplet of $(p, d_s, d_e)$ ." + ], + "image_footnote": [], + "bbox": [ + 341, + 258, + 669, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from DETR [3], query-based approaches have become a promising research direction owing to the architectural simplicity [17,21,26,37,41]. By decoding temporal spans (i.e., moments) from a handful set of learnable queries, they achieve promising grounding performance while maintaining high inference speed.", + "bbox": [ + 212, + 356, + 784, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing query-based models typically predict a moment using its center and width, i.e., $(c, l)$ , under the assumption that the boundaries are equidistant from the center. However, such a formulation can be problematic as the center of a moment might be ambiguous. An illustrative example is presented in Fig. 1 (top), where the frame at the center of the ground-truth moment is less relevant to the given sentence, as demonstrated by the low saliency label. Indeed, a moment center does not necessarily serve as the best representative of the sentence. This ambiguity can challenge the model's ability to precisely locate the centers, and the misaligned centers lead to low-quality predictions (Fig. 1a). To probe the impact of center misalignment, we conducted a diagnostic experiment in Table 1. The results reveal that (i) existing methods struggle with detecting accurate center points and (ii) they suffer from significant performance drops when the predicted centers deviate from the ground-truth centers (i.e., large center errors).", + "bbox": [ + 212, + 417, + 784, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we present a novel boundary-oriented formulation for moments, as illustrated in Fig. 1b, where each moment is represented by a triplet consisting of an anchor point and its distances to the boundaries, i.e., $(p,d_s,d_e)$ . This asymmetric formulation liberates the model from the stringent requirement of predicting the center. Instead, it is sufficient for the model to predict any salient anchor within the target moment, and the distances from the anchor to the onset and offset are predicted subsequently. By directly locating the boundaries based on the anchor point, the model can achieve improved boundary alignment, even when the anchor point does not coincide with the actual center.", + "bbox": [ + 212, + 613, + 784, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon the proposed moment modeling, we introduce a new framework equipped with a dedicated decoder design, dubbed Boundary-Aligned Moment Detection Transformer (BAM-DETR). The design of its decoding layers originates from the intuition that the refining processes for an anchor and boundaries should be distinct from each other. That is to say, a model needs to scan over the whole video to find a potential anchor point that enables estimating the", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Lee and H. Byun", + "bbox": [ + 271, + 114, + 405, + 127 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/9d5b6a342cc2190314f02b3d5c04c5aa7b6800056568a888d752e41da9c99cc8.jpg", + "table_caption": [ + "Table 1: Impact of misaligned centers on QVHighlights. The top-1 predictions are grouped based on their center errors normalized by ground-truth lengths. For each group, we present the mean IoU (\\%) and the proportion (in parentheses). Only the predictions whose centers fall within the ground-truth moments are considered here." + ], + "table_footnote": [ + "$^\\dagger$ anchor points are utilized for grouping." + ], + "table_body": "
Method[0, 0.1)[0.1, 0.2)[0.2, 0.3)[0.3, 0.4)[0.4, 0.5)All
Moment-DETR [21]83.2065.0054.8243.7734.9867.84
(46 %)(24 %)(14 %)(9 %)(7 %)(100 %)
QD-DETR [41]87.7967.9555.4044.9336.4574.09
(56 %)(18 %)(11 %)(9 %)(5 %)(100 %)
BAM-DETR†77.6278.2077.5278.9671.0877.21
(Ours)(24 %)(21 %)(21 %)(22 %)(12 %)(100 %)
", + "bbox": [ + 243, + 210, + 754, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rough location of the target moment. On the other hand, it is required to focus on fine-grained details in the vicinity to refine the boundaries to be aligned with those of the target moments. From this motivation, in contrast to existing methods, our model adopts a dual-pathway decoding pipeline to predict an anchor point and boundaries in a parallel way. To be specific, we leverage two different types of queries respectively for anchor and boundary refinement. The former aggregates global information with standard attention, while the latter concentrates on the sparse local neighborhood of the boundaries using the proposed boundary-focused attention. These distinct designs of two pathways allow for effective moment localization with minimal computational overhead increase.", + "bbox": [ + 212, + 320, + 785, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, we identify the problem of the conventional scoring method, where binary classification (or matching) scores are used for proposal ranking. This leads to suboptimal results for the grounding task since a fractional moment may have high matching scores with the given sentence. To handle this issue, we propose to rank proposals based on their localization qualities. Accordingly, we modify the typical matching function and training objectives of the query-based model to be localization-oriented by discarding the role of classification scores. In this way, our model can prioritize the moment proposals exhibiting high overlap with ground truths at inference, leading to improved grounding performance.", + "bbox": [ + 212, + 470, + 785, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The advantage of our BAM-DETR is showcased in Table 1. It can be observed that the anchor points predicted by our model are evenly distributed within the ground-truth intervals. Importantly, our model consistently produces precise moments with high IoUs ( $>0.7$ ) across different groups, indicating that it does not depend on accurate center prediction. On average, our model shows superior grounding performance over the existing methods with the help of our moment formulation and quality-based scoring. In a later section, we validate the efficacy of the proposed methods through extensive experiments. Notably, our model outperforms previous methods by large margins on three public benchmarks.", + "bbox": [ + 212, + 607, + 785, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 761, + 397, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Temporal Sentence Grounding in Videos", + "text_level": 1, + "bbox": [ + 214, + 787, + 596, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal sentence grounding requires seeking temporal spans semantically relevant to the given sentence in a video. Proposal-based approaches adopt the", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BAM-DETR", + "bbox": [ + 642, + 113, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "two-stage pipeline, i.e., proposal generation and ranking. They generate moment proposals by relying on sliding windows [11,13,34,73] or utilizing pre-defined anchors [4,59,64,67,74]. Several works process all possible candidates at once with 2D maps [25,32,56,57,72]. Meanwhile, proposal-free methods are developed for efficient grounding by directly predicting the moments [42,65] or estimating the probabilities of each frame being starting and ending positions [15,70]. Some approaches perform dense regression by predicting the boundaries from individual frames [6,39,66]. Recently, query-based models streamline the complicated sentence grounding pipeline by removing handcrafted techniques [2,17,26,37,60]. There are also attempts to unify temporal sentence grounding with other video understanding tasks into a single framework [28,61].", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method belongs to the query-based group [21, 41], inheriting the benefit of architectural simplicity. In contrast to others, we employ a boundary-oriented formulation of moments to relieve the heavy reliance on center predictions, leading to better boundary alignment. Our method also relates to dense regression methods [6, 28, 66] that predict boundaries from each frame as an anchor. Comparatively, our model leverages dynamic anchors that are gradually adjusted through decoding, enabling precise grounding using a small set of predictions.", + "bbox": [ + 212, + 313, + 787, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Detection Transformers", + "text_level": 1, + "bbox": [ + 214, + 436, + 457, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query-based temporal sentence grounding models by design are closely related to the family of detection transformers (DETR) [3]. Since the advent of DETR, a number of variants have been introduced to improve it from various perspectives [30, 33, 36, 53]. Some works focus on reducing the excessive computational costs of vanilla transformers in order to leverage multi-scale features [7, 23, 47, 76, 78]. On the other hand, several methods attempt to speed up the model convergence by manipulating the attention operations [12, 35, 62, 68] or incorporating the denoising process during the model training [24, 69].", + "bbox": [ + 212, + 458, + 787, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The most relevant works to ours are those which propose explicit anchor modeling for object queries using center points [40, 55] or boxes (center, width, and height) [35, 78]. In comparison to an object, a moment in temporal sentence grounding has its own challenges such as center ambiguity and indistinct boundaries. To accommodate the discrepancies, we propose a novel boundary-oriented modeling of moments to replace the conventional center-based 1D box modeling. The advantages of our approach are clearly verified in the experiments.", + "bbox": [ + 212, + 580, + 787, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 705, + 330, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an untrimmed video and a sentence, the goal of temporal sentence grounding is to localize relevant moments $\\{\\varphi_{n} = (t_{s_{n}}, t_{e_{n}})\\}_{n=1}^{N}$ , where $N$ denotes the number of ground truths in the video and $\\varphi_{n}$ indicates the temporal interval of the $n$ -th moment. Note that a video may have multiple moments that match the sentence, i.e., $N \\geq 1$ . During the test time, the model is expected to produce a total of $M$ predictions, $\\{(\\hat{\\varphi}_{m}, q_{m})\\}_{m=1}^{M}$ , where $\\hat{\\varphi}_{m} = (\\hat{t}_{s_{m}}, \\hat{t}_{e_{m}})$ is the $m$ -th prediction, while $q_{m}$ is its score for ranking.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Lee and H. Byun", + "bbox": [ + 271, + 114, + 405, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2903c8ae1d069fab841fdb58bbd2833f062e86bbe46ab08b5242bbd0b0b33af6.jpg", + "image_caption": [ + "(a) Overall framework of BAM-DETR" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 524, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f141d3e762f7ff75e61dd3db547f118e69356d1ff696dc37eb7aaa8b94a2349.jpg", + "image_caption": [ + "(b) Dual-pathway decoding layer", + "Fig. 2: (a) Overview of the proposed BAM-DETR. (b) Details of the proposed dual-pathway decoding layer. It consists of two parallel pathways respectively for anchor and boundary updates, which refine previous moment predictions in a sequential manner." + ], + "image_footnote": [], + "bbox": [ + 524, + 143, + 782, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivation. We tackle two main problems in the current prediction process. On one hand, existing query-based approaches adopt the symmetric design of $(c,l)$ to predict a moment, i.e., $\\hat{\\varphi} = (c - 0.5l, c + 0.5l)$ . As we discuss in Sec. 1, this strategy suffers from the issue of over-reliance on the center prediction, leading to unstable performance. To address this, we propose a boundary-oriented modeling with a triplet of $(p,d_s,d_e)$ , where a moment is represented by $\\hat{\\varphi} = (p - d_s,p + d_e)$ . This asymmetric design enables direct alignment of boundaries without relying on precise center prediction. On the other hand, previous models use classification scores as $q_{m}$ . This scoring is prone to sub-optimal solutions since only a fraction of the moment may well match the sentence, leading incomplete predictions to be highly ranked. To handle this, we propose quality-based scoring to sort the proposals based on their localization qualities rather than the degree of matching.", + "bbox": [ + 212, + 335, + 784, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 539, + 339, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2a, our BAM-DETR follows the encoder-decoder pipeline. Briefly, it first extracts multimodal video features using the encoders. Taking them as memory, the dual-pathway decoder predicts temporal spans and their quality scores by progressively refining learnable initial spans.", + "bbox": [ + 212, + 564, + 784, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Feature Extraction", + "text_level": 1, + "bbox": [ + 215, + 646, + 418, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the convention [42, 70, 72], we employ pre-trained encoders for token-level unimodal feature extraction. It is worth noting that all unimodal encoders are kept frozen during training to avoid memory exploding. In this stage, we obtain $D_v$ -dim video features $E_v \\in \\mathbb{R}^{N_v \\times D_v}$ and $D_t$ -dim text features $E_t \\in \\mathbb{R}^{N_t \\times D_t}$ , where $N_v$ and $N_t$ are the numbers of clips and words, respectively.", + "bbox": [ + 212, + 671, + 784, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Multimodal Encoder", + "text_level": 1, + "bbox": [ + 215, + 768, + 433, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We feed the unimodal features of the video and text to the multimodal encoder, so as to fuse them into text-aware video representations for temporal sentence grounding. While various multimodal encoders are explored [21,28,37], we", + "bbox": [ + 212, + 794, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BAM-DETR", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "adopt the text-to-video encoder design [41] consisting of cross- and self-attention blocks [54]. Before getting into the encoder, unimodal features are projected into a shared space to facilitate cross-modal interaction, i.e., $\\mathcal{V} = f_v(E_v) \\in \\mathbb{R}^{N_v \\times D}$ , $\\mathcal{T} = f_t(E_t) \\in \\mathbb{R}^{N_t \\times D}$ , where $D$ is the embedding dimension. Afterward, multi-head cross-attention blocks are used to inject textual information into the clip-level video representations. In specific, we project $\\mathcal{V}$ to $\\mathbf{Q}_{\\mathcal{V}}$ (query) while $\\mathcal{T}$ to $\\mathbf{K}_{\\mathcal{T}}$ (key) and $\\mathbf{V}_{\\mathcal{T}}$ (value). Then a cross-attention block can be formulated as:", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} ^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\mathcal {V}} \\mathbf {K} _ {\\mathcal {T}} ^ {\\top}}{\\sqrt {D}}\\right) \\mathbf {V} _ {\\mathcal {T}} + \\mathcal {V}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 256, + 785, + 289 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} ^ {\\prime \\prime} = \\operatorname {F F N} (\\mathcal {V} ^ {\\prime}) + \\mathcal {V} ^ {\\prime},\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 291, + 529, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{FFN}(\\cdot)$ is a feed-forward network. Although we here present the singlehead attention block, it can readily generalize to a multi-headed version [54]. We denote the resulting multimodal representations obtained after $L_{E}$ multi-head cross-attention blocks by $\\tilde{\\mathcal{V}}\\in \\mathbb{R}^{N_v\\times D}$ .", + "bbox": [ + 212, + 310, + 782, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, self-attention blocks are leveraged to enhance the representations by allowing the inter-clip interaction. Here we project $\\tilde{\\mathcal{V}}$ to $\\mathbf{Q}_{\\tilde{\\mathcal{V}}}$ , $\\mathbf{K}_{\\tilde{\\mathcal{V}}}$ , and $\\mathbf{V}_{\\tilde{\\mathcal{V}}}$ . We note that the query and the key are supplemented with fixed sinusoidal positional encoding [3, 54] for temporal awareness. Then the self-attention block is defined in a similar way to Eq. (1) but with different inputs. The enhanced clip-level representations after $L_{E}$ multi-head self-attention blocks are denoted by $\\hat{\\mathcal{V}} \\in \\mathbb{R}^{N_v \\times D}$ , which will serve as the memory for the decoder.", + "bbox": [ + 212, + 371, + 784, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is widely known that providing saliency guidance to the memory features helps the model to better understand the semantic relationship between the video and text [21, 28]. As in previous works [17, 26], we impose saliency score constraints on the memory features. Specifically, we leverage a saliency predictor $S(\\cdot)$ and train the model with the following margin-based training objective.", + "bbox": [ + 212, + 477, + 784, + 553 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a r g i n}} = \\max (0, \\alpha + S (\\hat {v} ^ {\\text {l o w}}) - S (\\hat {v} ^ {\\text {h i g h}})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 564, + 784, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ is a margin and $(\\hat{v}^{\\mathrm{low}},\\hat{v}^{\\mathrm{high}})$ is the sampled feature pair satisfying that the saliency label of $\\hat{v}^{\\mathrm{low}}$ is lower than that of $\\hat{v}^{\\mathrm{high}}$ . In case of the absence of saliency labels, we collect clips within and outside the ground-truth moment intervals to build a pair. In addition, we employ the rank-aware contrastive loss $\\mathcal{L}_{\\mathrm{cont}}$ and the negative relation loss $\\mathcal{L}_{\\mathrm{neg}}$ , following Moon et al. [41]. Due to space limits, we refer the readers to Appendix for the loss formulations. In summary, the overall saliency loss is defined as $\\mathcal{L}_{\\mathrm{sal}} = \\mathcal{L}_{\\mathrm{margin}} + \\mathcal{L}_{\\mathrm{cont}} + \\mathcal{L}_{\\mathrm{neg}}$ .", + "bbox": [ + 212, + 592, + 787, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Dual-pathway Decoder", + "text_level": 1, + "bbox": [ + 214, + 723, + 450, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the multimodal representations $\\hat{\\mathcal{V}}$ as memory features, we aim to localize temporal spans corresponding to the sentence. We adopt the new boundary-oriented formulation for moment prediction, where each prediction is represented by a triplet of $(p,d_s,d_e)$ , where $p$ is the anchor point, while $d_{s}$ and $d_{e}$ are the distances from the anchor to the starting and ending points, respectively. To make full use of the proposed formulation, we design a dual-pathway decoding", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Lee and H. Byun", + "bbox": [ + 271, + 114, + 405, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer with two parallel pathways (Fig. 2b). Formally, the inputs of the $l$ -th layer are anchor queries $\\mathbf{C}_p^l \\in \\mathbb{R}^{M \\times D}$ , boundary queries $\\mathbf{C}_s^l$ , $\\mathbf{C}_e^l \\in \\mathbb{R}^{M \\times D}$ , and previous moment predictions $\\mathbf{A}^l = [\\mathbf{p}^l; \\mathbf{d}_s^l; \\mathbf{d}_e^l] \\in \\mathbb{R}^{M \\times 3}$ , where $M$ is the number of queries (predictions). Note that the initial queries and spans, i.e., $\\{\\mathbf{C}_p^0, \\mathbf{C}_s^0, \\mathbf{C}_e^0, \\mathbf{A}^0\\}$ , are learnable parameters. We elaborate on the two pathways in the following.", + "bbox": [ + 214, + 146, + 787, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Anchor updating pathway. Given the predictions and the anchor queries from the preceding layer, the goal of this pathway is to adjust the anchor position so that the boundaries can be predicted based on it. Intuitively, in order to obtain a valuable anchor point without redundancy, an anchor query should communicate with other queries as well as the memory features. To this goal, the anchor updating pathway consists of a self-attention layer, a cross-attention layer, and a feed-forward network. In the self-attention layer, the anchor queries $\\mathbf{C}_p^l$ are first projected into $\\mathbf{Q}_{\\mathbf{C}_p^l}$ , $\\mathbf{K}_{\\mathbf{C}_p^l}$ , and $\\mathbf{V}_{\\mathbf{C}_p^l}$ . Since an anchor query itself lacks positional information, we build positional encoding. Following the previous works [17,41], we extend the current spans $\\mathbf{A}^l$ to the positional information of the queries, i.e., $\\mathbf{P}_{\\mathbf{A}^l} = \\mathrm{MLP}(\\mathrm{PE}(\\mathbf{A}^l)) \\in \\mathbb{R}^{M \\times D}$ , where $\\mathrm{PE}(\\cdot)$ denotes the point-wise mapping from a position to the corresponding sinusoidal encoding while $\\mathrm{MLP}(\\cdot)$ is multi-layer perceptron. The self-attention for anchor queries is defined as:", + "bbox": [ + 214, + 250, + 787, + 445 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {C}} _ {p} ^ {l} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {Q} _ {\\mathbf {C} _ {p} ^ {l}} + \\mathbf {P} _ {\\mathbf {A} ^ {l}}\\right) \\left(\\mathbf {K} _ {\\mathbf {C} _ {p} ^ {l}} + \\mathbf {P} _ {\\mathbf {A} ^ {l}}\\right) ^ {\\top}}{\\sqrt {D}}\\right) \\mathbf {V} _ {\\mathbf {C} _ {p} ^ {l}} + \\mathbf {C} _ {p} ^ {l}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 449, + 785, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After inter-query interaction, we employ a global cross-attention layer to aggregate multi-modal features from the memory. The anchor queries $\\tilde{\\mathbf{C}}_p^l$ is projected into $\\mathbf{Q}_{\\tilde{\\mathbf{C}}_p^l}$ while the memory $\\hat{\\nu}$ is projected to $\\mathbf{K}_{\\hat{\\nu}}$ and $\\mathbf{V}_{\\hat{\\nu}}$ . To make the query location-aware, we leverage the sinusoidal encoding of current anchor positions, i.e., $\\mathbf{P_{p^l}} = \\mathrm{PE}(\\mathbf{p}^l)\\in \\mathbb{R}^{M\\times D}$ . Similarly, the memory leverages the positional encoding, i.e., $\\mathbf{P}_{\\hat{\\nu}} = \\mathrm{PE}(\\hat{\\nu})\\in \\mathbb{R}^{N_v\\times D}$ . We use concatenation instead of summation to separate the roles of features and positional encoding [35, 40]. The cross-attention between anchor queries and the memory can be expressed as:", + "bbox": [ + 214, + 492, + 787, + 621 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {C}} _ {p} ^ {l} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {Q} _ {\\tilde {\\mathbf {C}} _ {p} ^ {l}} \\parallel \\mathbf {P} _ {\\mathbf {p} ^ {l}}\\right)\\left(\\mathbf {K} _ {\\hat {\\mathcal {V}}} \\parallel \\mathbf {P} _ {\\hat {\\mathcal {V}}}\\right)\\right) ^ {\\top}}{\\sqrt {2 D}}\\left. \\right) \\mathbf {V} _ {\\hat {\\mathcal {V}}} + \\tilde {\\mathbf {C}} _ {p} ^ {l}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 625, + 785, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After all, the anchor queries are updated with a feed-forward network, i.e., $\\mathbf{C}_p^{(l + 1)} = \\mathrm{FFN}(\\hat{\\mathbf{C}}_p^l) + \\hat{\\mathbf{C}}_p^l$ . Lastly, we adjust the anchor positions using sigmoid-based refinement [78]: $\\hat{\\mathbf{A}}^l = [\\mathbf{p}^{(l + 1)};\\mathbf{d}_s^l;\\mathbf{d}_e^l]$ where $\\mathbf{p}^{(l + 1)} = \\sigma (\\sigma^{-1}(\\mathbf{p}^l) + \\varDelta \\mathbf{p}^l)$ with $\\varDelta\\mathbf{p}^l=\\mathrm{MLP}(\\mathbf{C}_p^{(l+1)})\\in\\mathbb{R}^M$ and the sigmoid function $\\sigma(\\cdot)$ .", + "bbox": [ + 214, + 670, + 787, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Boundary updating pathway. After the anchor update, we refine the boundaries of the predictions. It is widely perceived that a model needs to focus on fine-grained features in the neighborhood rather than far ones to adjust temporal boundaries [27, 29, 51]. Inspired by this, we devise a boundary-focused attention layer (Fig. 3). For brevity, we explain the process of starting boundary update.", + "bbox": [ + 214, + 763, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BAM-DETR", + "bbox": [ + 642, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/037939b1153ca7e602e2d4e2373f4bc11f4f85ee6c8db0801185ac7a7dace92d.jpg", + "image_caption": [ + "Fig. 3: Boundary-focused attention layer for starting queries." + ], + "image_footnote": [], + "bbox": [ + 323, + 143, + 679, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The plain memory $\\hat{\\mathcal{V}}$ lacks the inductive bias due to the property of attentional layers [8]. Thus we first build locality-enhanced memory features for effective boundary refinement. To this goal, we obtain boundary-sensitive features with several 1D convolutional layers, i.e., $\\hat{\\mathcal{V}}_{\\mathrm{s}} = f_{\\mathrm{s}}(\\hat{\\mathcal{V}})$ . Then we encourage them to highly activate around the starting position of target moments. In detail, we impose regularization on the clip-wise activation scores obtained by channel-mean, i.e., $\\hat{g}^s = \\mathrm{mean}(\\sigma (\\hat{\\mathcal{V}}_s))\\in \\mathbb{R}^{N_v}$ . The regularization loss is defined as:", + "bbox": [ + 212, + 366, + 787, + 472 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g u l}} ^ {s} = - \\frac {1}{N _ {v}} \\sum_ {i = 1} ^ {N _ {v}} \\left(g _ {i} ^ {s} \\log \\left(\\hat {g} _ {i} ^ {s}\\right) + \\left(1 - g _ {i} ^ {s}\\right) \\log \\left(1 - \\hat {g} _ {i} ^ {s}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 315, + 476, + 785, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $g_{i}^{s}$ is the binary label obtained as $g_{i}^{s} = \\mathbb{1}[i\\in \\mathcal{B}^{s}]$ , where $\\mathcal{B}^s$ is the neighbor clip set around starting points with a radius of $r^s$ (set to 1/10 of the moment length). We merge boundary-sensitive features with the plain ones to form locality-enhanced memory, i.e., $\\hat{\\mathcal{V}}_s' = [\\hat{\\mathcal{V}}\\parallel \\hat{\\mathcal{V}}_s]\\in \\mathbb{R}^{N_v\\times 2D}$ . We can obtain $\\mathcal{L}_{\\mathrm{regul}}^e$ and $\\hat{\\mathcal{V}}_e^\\prime$ in the same way. The total regularization term is $\\mathcal{L}_{\\mathrm{regul}} = \\mathcal{L}_{\\mathrm{regul}}^s +\\mathcal{L}_{\\mathrm{regul}}^e$ .", + "bbox": [ + 212, + 518, + 787, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the locality-enhanced memory, the boundary queries need to capture fine-grained details around the boundaries for refinement. To efficiently aggregate useful features near the boundaries, we employ deformable attention [77, 78]. Regarding the starting boundary, i.e., $\\mathbf{p}^{(l + 1)} - \\mathbf{d}_s^l$ , as the origin, we predict offsets and weights to select $K$ neighbors, i.e., $\\mathbf{o}^s = \\phi_o(\\mathbf{C}_s^l)\\in \\mathbb{R}^{M\\times K}$ , $\\mathbf{w}^s = \\mathrm{softmax}(\\phi_w(\\mathbf{C}_s^l))\\in \\mathbb{R}^{M\\times K}$ , where $\\phi_{*}$ are fully-connected layers. Features from the sampled neighbors are then aggregated into the starting queries as:", + "bbox": [ + 212, + 606, + 787, + 712 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {C}} _ {s} ^ {l} = \\sum_ {k = 1} ^ {K} \\left[ \\mathbf {w} _ {k} ^ {s} \\cdot \\hat {\\mathcal {V}} _ {s} ^ {\\prime} \\left[ \\mathbf {p} ^ {(l + 1)} - \\mathbf {d} _ {s} ^ {l} + \\mathbf {o} _ {k} ^ {s} \\right] \\right] + \\mathbf {C} _ {s} ^ {l}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 713, + 785, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where we denote the sampling process from memory by $\\hat{\\mathcal{V}}_s^{\\prime}[\\cdot]$ . Lastly, we adopt a feed-forward network to obtain the updated starting queries, i.e., $\\mathbf{C}_s^{(l + 1)} = \\mathrm{FFN}(\\hat{\\mathbf{C}}_s^l)$ . We also obtain $\\mathbf{C}_e^{(l + 1)}$ in the same way. With the updated queries, we refine the boundaries to be better aligned with those of ground truths using sigmoid-based refinement as similar in the anchor update, leading to the refined", + "bbox": [ + 212, + 758, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Lee and H. Byun", + "bbox": [ + 271, + 114, + 405, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "predictions $\\mathbf{A}^{(l + 1)} = [\\mathbf{p}^{(l + 1)};\\mathbf{d}_s^{(l + 1)};\\mathbf{d}_e^{(l + 1)}]$ . Note that we utilize the deformable attention for the purpose of local feature aggregation, which sharply differs from the original purpose of efficient multi-scale global operation [78]. We provide comparison experiments regarding their roles in Appendix.", + "bbox": [ + 214, + 143, + 782, + 208 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Moment prediction. We repeatedly update the predictions through a total of $L_{D}$ dual-pathway decoding layers. We denote the resulting predictions by $\\mathbf{A} = \\left[\\mathbf{p};\\mathbf{d}_s;\\mathbf{d}_e\\right]$ , the anchor queries by $\\mathbf{C}_p$ , and the boundary queries by $\\mathbf{C}_s$ and $\\mathbf{C}_e$ . Then we cast the predictions in the form of starting and ending timestamps, i.e., $\\hat{\\varphi} = \\left[\\mathbf{p} - \\mathbf{d}_s;\\mathbf{p} + \\mathbf{d}_e\\right] \\in \\mathbb{R}^{M \\times 2}$ , which serves as the final results.", + "bbox": [ + 215, + 220, + 782, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Quality-based Scoring", + "text_level": 1, + "bbox": [ + 215, + 316, + 444, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After producing the moment predictions, we opt to rank them for evaluation. In the convention, query-based models utilize classification scores as the measure, which exhibits how well the proposals semantically match the sentence. However, it does not necessarily represent the localization qualities of proposals. Hence we propose to estimate the localization quality of each moment prediction. Formally, the quality score can be derived as $\\mathbf{q} = \\sigma (\\mathrm{MLP}([\\mathbf{C}_p\\parallel \\mathbf{C}_s\\parallel \\mathbf{C}_e]))\\in \\mathbb{R}^M$ , where $\\sigma$ is sigmoid activation. Then the quality loss is defined as follows.", + "bbox": [ + 214, + 339, + 782, + 445 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {q u a l}} = \\sum_ {m = 1} ^ {M} \\left| q _ {m} - \\max _ {\\forall n} \\left(\\frac {\\left| \\hat {\\varphi} _ {m} \\cap \\varphi_ {n} \\right|}{\\left| \\hat {\\varphi} _ {m} \\cup \\varphi_ {n} \\right|}\\right) \\right|, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 454, + 784, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the objective of the quality head is to predict the maximum IoUs of the proposals with ground-truth moments.", + "bbox": [ + 214, + 502, + 782, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.6 Matching", + "text_level": 1, + "bbox": [ + 215, + 553, + 341, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As in the standard of query-based models [3,21], we perform Hungarian matching [19] between predictions and ground truths. The optimal matching results $\\psi^{*}$ can be derived as follows.", + "bbox": [ + 214, + 575, + 782, + 621 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\psi^ {*} = \\arg \\min _ {\\psi \\in \\mathfrak {G} _ {N}} \\sum_ {n = 1} ^ {N} \\mathcal {C} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 623, + 784, + 664 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right) = \\lambda_ {l _ {1}} \\mathcal {L} _ {l _ {1}} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right) + \\lambda_ {\\mathrm {i o u}} \\mathcal {L} _ {\\mathrm {i o u}} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 667, + 697, + 685 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathfrak{G}_N$ denotes the combination pool and $\\psi (n)$ is the index of the prediction matched by the $n$ -th ground truth. $\\mathcal{L}_{l_1}$ and $\\mathcal{L}_{\\mathrm{iou}}$ respectively represent the $L_{1}$ distance and the generalized IoU [46] between the moments, while $\\lambda_{l_1}$ and $\\lambda_{\\mathrm{iou}}$ are their weights. Note that in contrast to other works, the classification term is not involved in the matching process, leading to localization-oriented matching.", + "bbox": [ + 214, + 686, + 782, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Once the matching is completed, we minimize the matching cost between each pair of the matching results $\\psi^{*}$ . The localization loss is defined as:", + "bbox": [ + 214, + 763, + 782, + 792 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l o c}} = \\sum_ {n = 1} ^ {N} \\mathcal {C} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi^ {*} (n)}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 801, + 784, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BAM-DETR", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Overall training objectives. Our model is trained in an end-to-end fashion and the overall training objective is defined as follows.", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {l o c}} + \\lambda_ {\\text {q u a l}} \\mathcal {L} _ {\\text {q u a l}} + \\lambda_ {\\text {s a l}} \\mathcal {L} _ {\\text {s a l}} + \\lambda_ {\\text {r e g u l}} \\mathcal {L} _ {\\text {r e g u l}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 180, + 785, + 196 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda_{*}$ are the balancing parameters.", + "bbox": [ + 215, + 199, + 501, + 215 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 237, + 375, + 255 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 268, + 444, + 285 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. QVHighlights [21] is a recently built dataset, containing a total of 10,148 videos and 10,310 sentences from vlog and news domains. In addition to moments, it provides segment-level saliency score annotations within each moment. Importantly, it allows a single sentence corresponding to multiple disjoint moments (1.8 on average). Due to this practical setup, we utilize QVHighlights as the main benchmark. Charades-STA [11] includes 16,128 sentence-moment pairs with 9,848 indoor videos. The average duration of videos and moments is 30.6 and 8.1 seconds, respectively. TACoS [45] contains 127 cooking videos encompassing a total of 18,818 sentence-moment pairs. This dataset is known to be challenging since the moments occupy only a small portion (6.1 sec on average) within considerably long videos (4.8 min on average).", + "bbox": [ + 212, + 294, + 785, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. Following the standard protocol, we measure the Recall@1 (R1) under the IoU thresholds of 0.3, 0.5, and 0.7 by default. Since QVHighlights contains multiple ground-truth moments per sentence, we report the mean average precision (mAP) with IoU thresholds of 0.5 and 0.75 as well as the average mAP over a set of IoU thresholds [0.5:0.05:0.95]. Meanwhile, we compute the mean IoU of top-1 predictions on Charades-STA and TACoS. Note that the performances at high IoU thresholds (e.g., 0.7) exhibit how well the predictions align with the ground truths.", + "bbox": [ + 212, + 460, + 785, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 603, + 455, + 618 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For a fair comparison, we adopt the same feature extraction strategy with the competitors [21,26,41]. Specifically, we adopt the CLIP [44] text features for text and the concatenation of Slowfast [10] (ResNet-50) and CLIP [44] (ViT-B/32) features for videos unless otherwise specified. The video features are extracted every 1 second for Charades-STA and 2 seconds for QVHighlihtgs and TACoS. To compare with audio-augmented models [37], we optionally employ audio features extracted by PANNs [18] pre-trained on AudioSet [14]. Due to the long video duration, we uniformly sample 200 feature vectors from each video for TACoS.", + "bbox": [ + 212, + 628, + 785, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We set the embedding dimension $D$ to 256, the number of attention heads to 8, the number of queries $M$ to 10, the number of boundary points $K$ to 3, and the margin $\\alpha$ to 0.2. We determine the numbers of encoding and decoding layers as same with the prior work [41], i.e., $L_{E} = L_{D} = 2$ . The balancing parameters are set as: $\\lambda_{l1} = 10$ , $\\lambda_{\\mathrm{iou}} = \\lambda_{\\mathrm{sal}} = \\lambda_{\\mathrm{regul}} = 1$ , $\\lambda_{\\mathrm{qual}} = 2$ . As in previous works [26,41], we increase $\\lambda_{\\mathrm{sal}}$ to 4 when saliency labels are unavailable, i.e., for Charades-STA", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Lee and H. Byun", + "bbox": [ + 271, + 114, + 405, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/93207d62e389d9eed764733d63b49577816eb983f6dcef3dc8e1784e891fc688.jpg", + "table_caption": [ + "Table 2: Results on the QVHighlights test split." + ], + "table_footnote": [ + "$\\dagger$ additional use of audio modality $\\ddagger$ reproduced by official checkpoint" + ], + "table_body": "
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
w/o pre-training
MCN [1]11.412.7224.948.2210.67
CAL [9]25.4911.5423.407.659.89
CLIP [44]16.885.1918.117.007.67
XML [22]46.6933.4647.8934.6734.90
Moment-DETR [21]52.8933.0254.8229.4030.73
UMT† [37]56.2341.1853.8337.0136.12
MH-DETR [60]60.0542.4860.7538.1338.38
QD-DETR [41]62.4044.9862.5239.8839.86
QD-DETR† [41]63.0645.1063.0440.1040.19
UniVTG [28]58.8640.8657.6035.5935.47
EaTR‡ [17]57.9842.4159.9539.2939.00
MomentDiff [26]57.4239.6654.0235.7335.95
BAM-DETR62.7148.6464.5746.3345.36
BAM-DETR†64.0748.1265.6147.5146.91
w/ pre-training on 4.2M data labeled by CLIP
UniVTG [28]65.4350.0664.0645.0243.63
w/ pre-training on 236K ASR captions
Moment-DETR [21]59.7840.3360.5135.3636.14
UMT† [37]60.8343.2657.3339.1238.08
QD-DETR [41]63.1845.1963.3740.3539.96
BAM-DETR63.8847.9266.3348.2246.67
", + "bbox": [ + 282, + 169, + 722, + 463 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "and TACoS. Our model is trained from scratch for 200 epochs on QVHighlights and 100 epochs on the other datasets using the AdamW optimizer [38] with a learning rate of 1e-4 and a batch size of 32.", + "bbox": [ + 212, + 484, + 787, + 532 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 214, + 556, + 545, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on QVHighlights. We compare our model with existing state-of-the-arts including recent query-based approaches [17, 21, 26, 37, 41, 60] on the test split. As shown in Table 2, our BAM-DETR consistently outperforms the comparative models under various settings. In detail, without pre-training, our model surpasses the previous state-of-the-art model [41] by large margins, e.g., $3.66\\%$ in R1@0.7 and $6.45\\%$ in mAP@0.75. These improvements under strict IoU thresholds verify the superior localization ability of our method. When leveraging auxiliary audio features, the performance further boosts especially in R1@0.5 and mAPs, enlarging the gap between the competitors including those with audio features [37, 41]. To compare with the methods with pretraining, we pretrain our model on middle-scale ASR caption data [21]. Again, our BAM-DETR achieves state-of-the-art results, while showing the least gap with the method [28] that leverages large-scale data for pre-training. Notably, our model even outperforms it in terms of mAPs with much fewer (about $18\\times$ ) pre-training data, manifesting the effectiveness of the proposed methods.", + "bbox": [ + 212, + 582, + 787, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on Charades-STA. Experimental results on the test split are shown in Table 3. Not only does our BAM-DETR outperform the query-based com", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BAM-DETR", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/07fb522f5ce489018a86bff3bf088735c4d3fa27ffcbb9433a67cc8687f7975e.jpg", + "table_caption": [ + "Table 3: Results on the Charades-STA and TACoS test splits." + ], + "table_footnote": [], + "table_body": "
MethodCharades-STATACoS
R1@0.3R1@0.5R1@0.7mIoUR1@0.3R1@0.5R1@0.7mIoU
2D-TAN [72]58.7646.0227.4041.2540.0127.9912.9227.22
VSLNet [70]60.3042.6924.1441.5835.5423.5413.1524.99
Moment-DETR [21]65.8352.0730.5945.5437.9724.6711.9725.49
QD-DETR [41]-57.3132.55-52.3936.7721.0735.76
UniVTG [28]70.8158.0135.6550.1051.4434.9717.3533.60
MomentDiff [26]-55.5732.42-46.6428.9212.3730.36
BAM-DETR72.9359.9539.3852.3356.6941.5426.7739.31
", + "bbox": [ + 222, + 169, + 781, + 291 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a880065b250b21766432508c9956ba74da706f939c0db24a10566ba0f2ccbfbe.jpg", + "table_caption": [ + "Table 4: Results on the anti-biased Charades-STA test split against the moment location and length. VGG [52] and Glove [43] features are employed for all models." + ], + "table_footnote": [ + " reproduced by official codebase" + ], + "table_body": "
Methodw.r.t. moment locationw.r.t. moment length
R1@0.3R1@0.5R1@0.7\\( \\mathrm{mAP}_{avg} \\)R1@0.3R1@0.5R1@0.7\\( \\mathrm{mAP}_{avg} \\)
2D-TAN [72]27.8120.4410.8417.2339.6828.6817.7222.79
MMN [56]33.5827.2014.1219.1843.5834.3119.9426.85
Moment-DETR [21]29.9422.1611.5618.6642.7334.3916.1224.02
\\( QD-DETR^‡ \\)[41]56.1746.8228.1330.7067.3954.4432.8736.99
MomentDiff [26]48.3933.5915.7121.3751.2538.3223.3828.19
BAM-DETR59.8350.0032.0831.6868.4055.4640.7443.21
", + "bbox": [ + 222, + 335, + 781, + 446 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "petitors [21, 26, 41], but it achieves a new state-of-the-art by surpassing the best performing anchor-free model [28] for all metrics. Notably, a large gap of $3.73\\%$ is observed in R1@0.7, which confirms the strong localization ability of our model. Results on TACoS. We present the comparison results on the test set in Table 3. It can be observed that our BAM-DETR achieves a new state-of-the-art with pronounced performances under the strict IoU thresholds, which is consistent with the above results on other datasets. On this challenging benchmark, our model surpasses the previous best model [41] by $5.7\\%$ (relatively $27\\%$ ) in R1@0.7. These results clearly exhibit the superiority of the proposed model.", + "bbox": [ + 212, + 468, + 787, + 604 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Robustness Evaluation", + "text_level": 1, + "bbox": [ + 215, + 625, + 449, + 638 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Query-based models potentially have a temporal bias [16, 63] against the locations and lengths of moments. To measure robustness, we evaluate our model on the anti-biased Charades-STA [26] with distribution shifts of the moment location and length between training and test sets. Table 4 summarizes the results, where our model outperforms all competitors under both anti-biased settings. Especially, it shows significant performance gaps under the moment length bias. This can be expected since our model directly localizes boundaries instead of predicting lengths, which lessens the effect of bias. This robustness test corroborates the advantage of our boundary-oriented moment modeling.", + "bbox": [ + 212, + 646, + 787, + 782 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Analysis", + "text_level": 1, + "bbox": [ + 215, + 801, + 333, + 816 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct analytical experiments on the QVHighlights validation split.", + "bbox": [ + 215, + 824, + 740, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Lee and H. Byun", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/dc8e96552aa8987ff809aa7a69ec5eab6f1dc392042cec87c8a756cd28c850f0.jpg", + "table_caption": [ + "Table 5: Ablation study of components on QVHighlights." + ], + "table_footnote": [], + "table_body": "
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
Baseline62.3947.8762.6441.5441.75
+ boundary-oriented modeling63.4249.2362.8643.2442.42
+ dual-pathway decoder63.6150.2663.0144.9844.16
+ quality-based scoring65.1051.6165.4148.5647.61
", + "bbox": [ + 276, + 169, + 727, + 253 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/234e5cb5642634cffd27e99598d79a0cdbc0f6c470d0f86cb13723faa9f2a8c1.jpg", + "table_caption": [ + "Table 6: Ablative experiments on QVHighlights. Default settings are marked gray.", + "(a) Number of predictions" + ], + "table_footnote": [], + "table_body": "
MR1mAP
@0.5@0.7Avg.
562.1949.0345.18
1065.1051.6147.61
1565.4850.3248.23
2065.2351.6148.01
", + "bbox": [ + 225, + 303, + 382, + 386 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/10002ff49de3faa01d89f91c605ba683bf6a32a4ea96ea0a2d790248f1f6494b.jpg", + "table_caption": [ + "(b) Choice of memory features" + ], + "table_footnote": [], + "table_body": "
MemoryR1mAP Avg.
@0.5@0.7
plain63.7449.6846.21
sensitive63.8149.8746.34
merged65.1051.6147.61
", + "bbox": [ + 397, + 303, + 583, + 375 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9dc75ae1b163f66d079f699c997a35da42920a65438b87d3eb10f30467e9ad52.jpg", + "table_caption": [ + "(c) Number of boundary points" + ], + "table_footnote": [], + "table_body": "
KR1mAP
@0.5@0.7Avg.
1 (fixed)63.2348.3245.17
164.3249.8747.14
365.1051.6147.61
564.9050.4546.70
", + "bbox": [ + 598, + 303, + 782, + 386 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9f735c3269b61d11fb96d4e56382a4bab13832c8a4d5b91bc96412dbf6c13532.jpg", + "table_caption": [ + "(d) Query choices for quality prediction (e) Combinations of attention layers for updating pathways" + ], + "table_footnote": [], + "table_body": "
CpCsCeR1mAP
@0.5@0.7Avg.
✓62.9749.0346.26
✓64.3950.7146.66
✓64.5849.2346.71
✓✓✓65.1051.6147.61
", + "bbox": [ + 228, + 405, + 444, + 487 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bf269cb49dee9ba5148519c8fa1a15fd9183910a1a7bd35d0a9ed542f0aad8de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Anchor queryBoundary queryR1 @0.5@0.7mAP Avg.FLOPsParams
Global (shared)62.2649.0344.180.71G8.2M
GlobalGlobal63.6149.8144.540.72G11.5M
GlobalFocused63.7449.6846.210.65G9.5M
FocusedFocused62.4849.1645.870.62G7.6M
", + "bbox": [ + 457, + 405, + 789, + 487 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of each component. We analyze the effect of each component in Table 5. The direct adoption of our moment modeling solely improves the performance, particularly at strict thresholds, showcasing its advantage in boundary alignment. Employing the dual-pathway decoder leads to further gains, which suggests the essential role of separate pathways. Lastly, the quality-based scoring considerably elevates the scores, especially in terms of mAPs.", + "bbox": [ + 212, + 503, + 785, + 594 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Number of predictions. We experiment with varying numbers of predictions in Table 6a, where the model achieves robust results when $M$ is sufficiently large. By default, we set $M$ to 10 for a fair comparison with the previous works [21,41].", + "bbox": [ + 212, + 595, + 785, + 642 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Locality-enhanced features. We analyze the effect of the choice of input features for boundary-focused attention in Table 6b. The results indicate that boundary-sensitive features are slightly more helpful than plain ones in precise localization. In addition, the merged features achieve the best performance.", + "bbox": [ + 212, + 642, + 785, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Number of sampled boundary points. We analyze the effect of the number of sampled boundary points $K$ in Table 6c. For comparisons, we report the case where the fixed boundary features are sampled without using offsets ( $1^{\\mathrm{st}}$ row). As shown in the table, the dynamic selection of neighborhoods rather than fixed boundaries is important for accurate moment localization. The performance improves when sampling multiple points, while it saturates at $K = 3$ .", + "bbox": [ + 212, + 703, + 785, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Query choices for quality prediction. Our model leverages both anchor and boundary queries for quality prediction. We investigate the effect of query choices in Table 6d. As a result, utilizing all the queries shows better localization", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BAM-DETR", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a3183881a6d67efc4689850fe951014885c2be812c4cc1e918f170c9e79e6fe6.jpg", + "image_caption": [ + "(a) Boundary hit rate", + "Fig. 4: Analytical experiments on QVHighlights." + ], + "image_footnote": [], + "bbox": [ + 215, + 145, + 390, + 237 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f84c2f8966cf19bf98b821dd9769ed1120c79d299a12ce45c3faec02403e29e0.jpg", + "image_caption": [ + "(b) Visualization results" + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 787, + 236 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "performance than using one of them. This might be trivial as different queries contain complementary information for quality prediction.", + "bbox": [ + 212, + 287, + 782, + 316 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Attention layers. Table 6e compares different combinations of attention layers for updating pathways in terms of performances and costs. For an apple-to/apple comparison, we leverage the plain memory features for both attention layers. First of all, employing separate global cross-attention for different queries leads to a huge parameter increase yet limited score gains. Replacing the global one for boundary queries with our focused attention layer substantially reduces the cost, while achieving the best performance. Meanwhile, the use of focused attention for both types of queries leads to inferior results.", + "bbox": [ + 212, + 318, + 784, + 436 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Boundary alignment. Standard IoU-based metrics are indirect measures for boundary alignment. To precisely diagnose the ability, inspired by Deeplab [5], we compute the boundary hit rate of predictions with varying band widths. We expand ground-truth boundary points with a band width to form starting/ending zones and regard a prediction as correct if both of its boundaries fall within the corresponding zones. More details can be found in Appendix. To disentangle the effect of ranking, we mark a video as correct if at least one prediction is correct and measure the video-level hit rate. As shown in Fig. 4a, our model greatly outperforms the recent competitors. The sharp increase in low band widths clearly validates the superiority of our model in boundary alignment.", + "bbox": [ + 212, + 439, + 784, + 589 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative results. As shown in Fig. 4b, previous models fail to localize accurate moments with misleading center predictions. Especially, QD-DETR [41] accurately predicts the moment length but suffers from the misaligned center, leading to the limited IoU. In contrast, thanks to the novel moment modeling, our model predicts well-aligned boundaries without relying on center prediction.", + "bbox": [ + 212, + 589, + 784, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 686, + 359, + 704 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we identified the center misalignment issue of existing query-based models for sentence grounding. To address it, we presented boundary-oriented moment modeling where boundaries are directly predicted without relying on centers. Based on the modeling, we designed the boundary-aligned moment detection transformer characterized by dual-pathway decoding. Further, we proposed localization quality-based scoring of predictions. The efficacy of the proposed methods is validated by thorough examinations. We hope this work sheds light on the issue of center-based moment modeling in detection transformers.", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Lee and H. Byun", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project was supported by the National Research Foundation of Korea grant funded by the Korea government (MSIT) (No. 2022R1A2B5B02001467; RS-2024-00346364).", + "bbox": [ + 212, + 172, + 787, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017) 1, 11", + "2. Cao, M., Chen, L., Shou, M.Z., Zhang, C., Zou, Y.: On pursuit of designing multimodal transformer for video grounding. In: EMNLP. pp. 9810-9823 (2021) 4", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229. Springer (2020) 2, 4, 6, 9", + "4. Chen, J., Chen, X., Ma, L., Jie, Z., Chua, T.S.: Temporally grounding natural sentence in video. In: EMNLP. pp. 162-171 (2018) 4", + "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. PAMI 40(4), 834-848 (2017) 14", + "6. Chen, L., Lu, C., Tang, S., Xiao, J., Zhang, D., Tan, C., Li, X.: Rethinking the bottom-up framework for query-based video localization. In: AAAI. vol. 34, pp. 10551-10558 (2020) 4", + "7. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: ICCV. pp. 2988-2997 (2021) 4", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 8", + "9. Escorcia, V., Soldan, M., Sivic, J., Ghanem, B., Russell, B.: Temporal localization of moments in video collections with natural language. arXiv preprint arXiv:1907.12763 (2019) 1, 11", + "0. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6202-6211 (2019) 10", + "1. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: CVPR. pp. 5267-5275 (2017) 1, 4, 10", + "2. Gao, P., Zheng, M., Wang, X., Dai, J., Li, H.: Fast convergence of detr with spatially modulated co-attention. In: ICCV. pp. 3621-3630 (2021) 4", + "3. Ge, R., Gao, J., Chen, K., Nevatia, R.: Mac: Mining activity concepts for language-based temporal localization. In: WACV. pp. 245-253. IEEE (2019) 4", + "4. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: ICASSP. pp. 776-780. IEEE (2017) 10", + "5. Ghosh, S., Agarwal, A., Parekh, Z., Hauptmann, A.G.: Excl: Extractive clip localization using natural language descriptions. In: NAACL. pp. 1984-1990 (2019) 4", + "6. Hao, J., Sun, H., Ren, P., Wang, J., Qi, Q., Liao, J.: Can shuffling video benefit temporal bias problem: A novel training framework for temporal grounding. In: ECCV. pp. 130-147. Springer (2022) 12" + ], + "bbox": [ + 225, + 267, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BAM-DETR", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: ICCV. pp. 13846-13856 (2023) 2, 4, 6, 7, 11", + "18. Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Trans. on Audio, Speech, and Language Processing 28, 2880-2894 (2020) 10", + "19. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) 9", + "20. Lee, P., Kim, T., Shim, M., Wee, D., Byun, H.: Decomposed cross-modal distillation for rgb-based temporal action detection. In: CVPR. pp. 2373-2383 (2023) 1", + "21. Lei, J., Berg, T.L., Bansal, M.: Detecting moments and highlights in videos via natural language queries. In: Neurips. vol. 34, pp. 11846-11858 (2021) 2, 3, 4, 5, 6, 9, 10, 11, 12, 13", + "22. Lei, J., Yu, L., Berg, T.L., Bansal, M.: Tv: A large-scale dataset for video-subtitle moment retrieval. In: ECCV. pp. 447-463. Springer (2020) 11", + "23. Li, F., Zeng, A., Liu, S., Zhang, H., Li, H., Zhang, L., Ni, L.M.: Lite detr: An interleaved multi-scale encoder for efficient detr. In: CVPR. pp. 18558-18567 (2023) 4", + "24. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: CVPR. pp. 13619-13627 (2022) 4", + "25. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In: ICCV. pp. 12032-12042 (2023) 4", + "26. Li, P., Xie, C.W., Xie, H., Zhao, L., Zhang, L., Zheng, Y., Zhao, D., Zhang, Y.: Momentdiff: Generative video moment retrieval from random to real. In: Neurips (2023) 2, 4, 6, 10, 11, 12", + "27. Lin, C., Xu, C., Luo, D., Wang, Y., Tai, Y., Wang, C., Li, J., Huang, F., Fu, Y.: Learning salient boundary feature for anchor-free temporal action localization. In: CVPR. pp. 3320-3329 (2021) 7", + "28. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: ICCV. pp. 2794-2804 (2023) 4, 5, 6, 11, 12", + "29. Lin, T., Zhao, X., Su, H., Wang, C., Yang, M.: Bsn: Boundary sensitive network for temporal action proposal generation. In: ECCV. pp. 3-19 (2018) 1, 7", + "30. Lin, Y., Yuan, Y., Zhang, Z., Li, C., Zheng, N., Hu, H.: Detr does not need multiscale or locality design. In: ICCV. pp. 6545-6554 (2023) 4", + "31. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: CVPR. pp. 11235-11244 (2021) 1", + "32. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: CVPR. pp. 11235-11244 (2021) 4", + "33. Liu, F., Wei, H., Zhao, W., Li, G., Peng, J., Li, Z.: Wb-detr: transformer-based detector without backbone. In: ICCV. pp. 2979-2987 (2021) 4", + "34. Liu, M., Wang, X., Nie, L., Tian, Q., Chen, B., Chua, T.S.: Cross-modal moment localization in videos. In: ACM MM. pp. 843-851 (2018) 4", + "35. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. In: ICLR (2022) 4, 7", + "36. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., Zhang, L.: Detection transformer with stable matching. In: ICCV. pp. 6491-6500 (2023) 4" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Lee and H. Byun", + "bbox": [ + 271, + 114, + 405, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: CVPR. pp. 3042-3051 (2022) 2, 4, 5, 10, 11", + "38. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 11", + "39. Lu, C., Chen, L., Tan, C., Li, X., Xiao, J.: Debug: A dense bottom-up grounding approach for natural language video localization. In: EMNLP-IJCNLP. pp. 5144-5153 (2019) 4", + "40. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: ICCV. pp. 3651-3660 (2021) 4, 7", + "41. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: CVPR. pp. 23023-23033 (2023) 2, 3, 4, 6, 7, 10, 11, 12, 13, 14", + "42. Mun, J., Cho, M., Han, B.: Local-global video-text interactions for temporal grounding. In: CVPR. pp. 10810-10819 (2020) 4, 5", + "43. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP. pp. 1532-1543 (2014) 12", + "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021) 10, 11", + "45. Regneri, M., Rohrbach, M., Wetzel, D., Thater, S., Schiele, B., Pinkal, M.: Grounding action descriptions in videos. Trans. Assoc. Comput. Linguistics 1, 25-36 (2013) 10", + "46. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: CVPR. pp. 658-666 (2019) 9", + "47. Roh, B., Shin, J., Shin, W., Kim, S.: Sparse detr: Efficient end-to-end object detection with learnable sparsity. In: ICLR (2022) 4", + "48. Rui, Y., Gupta, A., Acero, A.: Automatically extracting highlights for tv baseball programs. In: ACM MM. pp. 105-115 (2000) 1", + "49. Shao, D., Xiong, Y., Zhao, Y., Huang, Q., Qiao, Y., Lin, D.: Find and focus: Retrieve and localize video events with natural language queries. In: ECCV. pp. 200-216 (2018) 1", + "50. Sharghi, A., Laurel, J.S., Gong, B.: Query-focused video summarization: Dataset, evaluation, and a memory network based approach. In: CVPR. pp. 4788-4797 (2017) 1", + "51. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: CVPR. pp. 18857-18866 (2023) 7", + "52. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) 12", + "53. Sun, Z., Cao, S., Yang, Y., Kitani, K.M.: Rethinking transformer-based set prediction for object detection. In: ICCV. pp. 3611-3620 (2021) 4", + "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Neurips. vol. 30 (2017) 6", + "55. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: AAAI. vol. 36, pp. 2567-2575 (2022) 4", + "56. Wang, Z., Wang, L., Wu, T., Li, T., Wu, G.: Negative sample matters: A renaissance of metric learning for temporal grounding. In: AAAI. vol. 36, pp. 2613-2623 (2022) 4, 12" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BAM-DETR", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "57. Xiao, S., Chen, L., Zhang, S., Ji, W., Shao, J., Ye, L., Xiao, J.: Boundary proposal network for two-stage natural language video localization. In: AAAI. vol. 35, pp. 2986-2994 (2021) 4", + "58. Xiong, B., Kalantidis, Y., Ghadiyaram, D., Grauman, K.: Less is more: Learning highlight detection from video duration. In: CVPR. pp. 1258-1267 (2019) 1", + "59. Xu, H., He, K., Plummer, B.A., Sigal, L., Sclaroff, S., Saenko, K.: Multilevel language and vision integration for text-to-clip retrieval. In: AAAI. vol. 33, pp. 9062–9069 (2019) 4", + "60. Xu, Y., Sun, Y., Li, Y., Shi, Y., Zhu, X., Du, S.: Mh-detr: Video moment and highlight detection with cross-modal transformer. In: ACM MM (2023) 4, 11", + "61. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: Unloc: A unified framework for video localization tasks. In: ICCV. pp. 13623-13633 (2023) 4", + "62. Ye, M., Ke, L., Li, S., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Cascade-detr: Delving into high-quality universal object detection. In: ICCV. pp. 6704-6714 (2023) 4", + "63. Yuan, Y., Lan, X., Wang, X., Chen, L., Wang, Z., Zhu, W.: A closer look at temporal sentence grounding in videos: Dataset and metric. In: Proc. 2nd Int. Workshop on Human-Centric Multimedia Analysis. pp. 13-21 (2021) 12", + "64. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In: Neurips. vol. 32 (2019) 1, 4", + "65. Yuan, Y., Mei, T., Zhu, W.: To find where you talk: Temporal sentence localization in video with attention based location regression. In: AAAI. vol. 33, pp. 9159-9166 (2019) 4", + "66. Zeng, R., Xu, H., Huang, W., Chen, P., Tan, M., Gan, C.: Dense regression network for video grounding. In: CVPR. pp. 10287-10296 (2020) 4", + "67. Zhang, D., Dai, X., Wang, X., Wang, Y.F., Davis, L.S.: Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In: CVPR. pp. 1247-1257 (2019) 4", + "68. Zhang, G., Luo, Z., Yu, Y., Cui, K., Lu, S.: Accelerating detr convergence via semantic-aligned matching. In: CVPR. pp. 949-958 (2022) 4", + "69. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: ICLR (2023) 4", + "70. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. In: ACL (2020) 4, 5, 12", + "71. Zhang, K., Chao, W.L., Sha, F., Grauman, K.: Video summarization with long short-term memory. In: ECCV. pp. 766-782. Springer (2016) 1", + "72. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. vol. 34, pp. 12870-12877 (2020) 1, 4, 5, 12", + "73. Zhang, S., Su, J., Luo, J.: Exploiting temporal relationships in video moment localization with natural language. In: ACM MM. pp. 1230-1238 (2019) 4", + "74. Zhang, Z., Lin, Z., Zhao, Z., Xiao, Z.: Cross-modal interaction networks for query-based moment retrieval in videos. In: ACM SIGIR. pp. 655-664 (2019) 4", + "75. Zhao, Y., Xiong, Y., Wang, L., Wu, Z., Tang, X., Lin, D.: Temporal action detection with structured segment networks. In: ICCV. pp. 2914-2923 (2017) 1", + "76. Zheng, D., Dong, W., Hu, H., Chen, X., Wang, Y.: Less is more: Focus attention for efficient detr. In: ICCV. pp. 6674-6683 (2023) 4" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P. Lee and H. Byun", + "bbox": [ + 271, + 114, + 405, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "77. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: CVPR. pp. 9308-9316 (2019) 8", + "78. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021) 4, 7, 8, 9" + ], + "bbox": [ + 215, + 146, + 784, + 204 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BAM-DETR", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_model.json b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b85289d9da34617f512dca1806431235a4ac414 --- /dev/null +++ b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_model.json @@ -0,0 +1,2922 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.241, + 0.141, + 0.764, + 0.208 + ], + "angle": 0, + "content": "BAM-DETR: Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.235, + 0.638, + 0.251 + ], + "angle": 0, + "content": "Pilhyeon Lee\\(^{1\\star}\\) and Hyeran Byun\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.262, + 0.686, + 0.278 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Artificial Intelligence, Inha University" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.278, + 0.682, + 0.291 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Computer Science, Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.322, + 0.744, + 0.572 + ], + "angle": 0, + "content": "Abstract. Temporal sentence grounding aims to localize moments relevant to a language description. Recently, DETR-like approaches achieved notable progress by predicting the center and length of a target moment. However, they suffer from the issue of center misalignment raised by the inherent ambiguity of moment centers, leading to inaccurate predictions. To remedy this problem, we propose a novel boundary-oriented moment formulation. In our paradigm, the model no longer needs to find the precise center but instead suffices to predict any anchor point within the interval, from which the boundaries are directly estimated. Based on this idea, we design a boundary-aligned moment detection transformer, equipped with a dual-pathway decoding process. Specifically, it refines the anchor and boundaries within parallel pathways using global and boundary-focused attention, respectively. This separate design allows the model to focus on desirable regions, enabling precise refinement of moment predictions. Further, we propose a quality-based ranking method, ensuring that proposals with high localization qualities are prioritized over incomplete ones. Experiments on three benchmarks validate the effectiveness of the proposed methods. The code is available here." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.585, + 0.708, + 0.599 + ], + "angle": 0, + "content": "Keywords: Temporal sentence grounding \\(\\cdot\\) Detection transformer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.376, + 0.637 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.787 + ], + "angle": 0, + "content": "Recent years have witnessed a notable surge in the popularity of short-form video content on social media platforms like TikTok, YouTube Shorts, and Instagram Reels. As such, users prefer to selectively engage with short moments of interest rather than passively watch an entire long video. This trend highlights the importance of localizing desired moments. As a result, moment localization tasks have emerged as pivotal research topics in video understanding, including temporal action detection [20, 29, 75], video summarization [50, 71], and highlight detection [48, 58]. Within this context, we tackle temporal sentence grounding [1, 72], aiming to retrieve moments corresponding to free-form language descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.787, + 0.818 + ], + "angle": 0, + "content": "To address temporal sentence grounding, numerous efforts have been undertaken in the last decade [9, 11, 31, 49, 64, 72]. Especially, taking inspiration" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.661, + 0.841 + ], + "angle": 0, + "content": "* Correspondence to: Pilhyeon Lee " + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.128 + ], + "angle": 0, + "content": "P. Lee and H. Byun" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.144, + 0.67, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.247, + 0.594, + 0.258 + ], + "angle": 0, + "content": "(a) Conventional moment modeling" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.26, + 0.671, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.281, + 0.625, + 0.292 + ], + "angle": 0, + "content": "(b) Boundary-oriented moment modeling (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.303, + 0.784, + 0.346 + ], + "angle": 0, + "content": "Fig. 1: Comparison of moment modeling approaches under the scenario of an ambiguous center from QVHighlights. (a) The conventional method formulates a moment with a tuple of \\((c, l)\\). (b) In contrast, we propose to model it with a triplet of \\((p, d_s, d_e)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.785, + 0.417 + ], + "angle": 0, + "content": "from DETR [3], query-based approaches have become a promising research direction owing to the architectural simplicity [17,21,26,37,41]. By decoding temporal spans (i.e., moments) from a handful set of learnable queries, they achieve promising grounding performance while maintaining high inference speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.613 + ], + "angle": 0, + "content": "Existing query-based models typically predict a moment using its center and width, i.e., \\((c, l)\\), under the assumption that the boundaries are equidistant from the center. However, such a formulation can be problematic as the center of a moment might be ambiguous. An illustrative example is presented in Fig. 1 (top), where the frame at the center of the ground-truth moment is less relevant to the given sentence, as demonstrated by the low saliency label. Indeed, a moment center does not necessarily serve as the best representative of the sentence. This ambiguity can challenge the model's ability to precisely locate the centers, and the misaligned centers lead to low-quality predictions (Fig. 1a). To probe the impact of center misalignment, we conducted a diagnostic experiment in Table 1. The results reveal that (i) existing methods struggle with detecting accurate center points and (ii) they suffer from significant performance drops when the predicted centers deviate from the ground-truth centers (i.e., large center errors)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.749 + ], + "angle": 0, + "content": "To address this challenge, we present a novel boundary-oriented formulation for moments, as illustrated in Fig. 1b, where each moment is represented by a triplet consisting of an anchor point and its distances to the boundaries, i.e., \\((p,d_s,d_e)\\). This asymmetric formulation liberates the model from the stringent requirement of predicting the center. Instead, it is sufficient for the model to predict any salient anchor within the target moment, and the distances from the anchor to the onset and offset are predicted subsequently. By directly locating the boundaries based on the anchor point, the model can achieve improved boundary alignment, even when the anchor point does not coincide with the actual center." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Building upon the proposed moment modeling, we introduce a new framework equipped with a dedicated decoder design, dubbed Boundary-Aligned Moment Detection Transformer (BAM-DETR). The design of its decoding layers originates from the intuition that the refining processes for an anchor and boundaries should be distinct from each other. That is to say, a model needs to scan over the whole video to find a potential anchor point that enables estimating the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAM-DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: Impact of misaligned centers on QVHighlights. The top-1 predictions are grouped based on their center errors normalized by ground-truth lengths. For each group, we present the mean IoU (\\%) and the proportion (in parentheses). Only the predictions whose centers fall within the ground-truth moments are considered here." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.212, + 0.756, + 0.3 + ], + "angle": 0, + "content": "
Method[0, 0.1)[0.1, 0.2)[0.2, 0.3)[0.3, 0.4)[0.4, 0.5)All
Moment-DETR [21]83.2065.0054.8243.7734.9867.84
(46 %)(24 %)(14 %)(9 %)(7 %)(100 %)
QD-DETR [41]87.7967.9555.4044.9336.4574.09
(56 %)(18 %)(11 %)(9 %)(5 %)(100 %)
BAM-DETR†77.6278.2077.5278.9671.0877.21
(Ours)(24 %)(21 %)(21 %)(22 %)(12 %)(100 %)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.243, + 0.3, + 0.474, + 0.312 + ], + "angle": 0, + "content": "\\(^\\dagger\\) anchor points are utilized for grouping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.787, + 0.47 + ], + "angle": 0, + "content": "rough location of the target moment. On the other hand, it is required to focus on fine-grained details in the vicinity to refine the boundaries to be aligned with those of the target moments. From this motivation, in contrast to existing methods, our model adopts a dual-pathway decoding pipeline to predict an anchor point and boundaries in a parallel way. To be specific, we leverage two different types of queries respectively for anchor and boundary refinement. The former aggregates global information with standard attention, while the latter concentrates on the sparse local neighborhood of the boundaries using the proposed boundary-focused attention. These distinct designs of two pathways allow for effective moment localization with minimal computational overhead increase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.787, + 0.607 + ], + "angle": 0, + "content": "In addition, we identify the problem of the conventional scoring method, where binary classification (or matching) scores are used for proposal ranking. This leads to suboptimal results for the grounding task since a fractional moment may have high matching scores with the given sentence. To handle this issue, we propose to rank proposals based on their localization qualities. Accordingly, we modify the typical matching function and training objectives of the query-based model to be localization-oriented by discarding the role of classification scores. In this way, our model can prioritize the moment proposals exhibiting high overlap with ground truths at inference, leading to improved grounding performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.787, + 0.743 + ], + "angle": 0, + "content": "The advantage of our BAM-DETR is showcased in Table 1. It can be observed that the anchor points predicted by our model are evenly distributed within the ground-truth intervals. Importantly, our model consistently produces precise moments with high IoUs (\\(>0.7\\)) across different groups, indicating that it does not depend on accurate center prediction. On average, our model shows superior grounding performance over the existing methods with the help of our moment formulation and quality-based scoring. In a later section, we validate the efficacy of the proposed methods through extensive experiments. Notably, our model outperforms previous methods by large margins on three public benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.762, + 0.398, + 0.779 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.598, + 0.805 + ], + "angle": 0, + "content": "2.1 Temporal Sentence Grounding in Videos" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Temporal sentence grounding requires seeking temporal spans semantically relevant to the given sentence in a video. Proposal-based approaches adopt the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.129 + ], + "angle": 0, + "content": "P. Lee and H. Byun" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "two-stage pipeline, i.e., proposal generation and ranking. They generate moment proposals by relying on sliding windows [11,13,34,73] or utilizing pre-defined anchors [4,59,64,67,74]. Several works process all possible candidates at once with 2D maps [25,32,56,57,72]. Meanwhile, proposal-free methods are developed for efficient grounding by directly predicting the moments [42,65] or estimating the probabilities of each frame being starting and ending positions [15,70]. Some approaches perform dense regression by predicting the boundaries from individual frames [6,39,66]. Recently, query-based models streamline the complicated sentence grounding pipeline by removing handcrafted techniques [2,17,26,37,60]. There are also attempts to unify temporal sentence grounding with other video understanding tasks into a single framework [28,61]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Our method belongs to the query-based group [21, 41], inheriting the benefit of architectural simplicity. In contrast to others, we employ a boundary-oriented formulation of moments to relieve the heavy reliance on center predictions, leading to better boundary alignment. Our method also relates to dense regression methods [6, 28, 66] that predict boundaries from each frame as an anchor. Comparatively, our model leverages dynamic anchors that are gradually adjusted through decoding, enabling precise grounding using a small set of predictions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.438, + 0.458, + 0.452 + ], + "angle": 0, + "content": "2.2 Detection Transformers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Query-based temporal sentence grounding models by design are closely related to the family of detection transformers (DETR) [3]. Since the advent of DETR, a number of variants have been introduced to improve it from various perspectives [30, 33, 36, 53]. Some works focus on reducing the excessive computational costs of vanilla transformers in order to leverage multi-scale features [7, 23, 47, 76, 78]. On the other hand, several methods attempt to speed up the model convergence by manipulating the attention operations [12, 35, 62, 68] or incorporating the denoising process during the model training [24, 69]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.687 + ], + "angle": 0, + "content": "The most relevant works to ours are those which propose explicit anchor modeling for object queries using center points [40, 55] or boxes (center, width, and height) [35, 78]. In comparison to an object, a moment in temporal sentence grounding has its own challenges such as center ambiguity and indistinct boundaries. To accommodate the discrepancies, we propose a novel boundary-oriented modeling of moments to replace the conventional center-based 1D box modeling. The advantages of our approach are clearly verified in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.706, + 0.331, + 0.722 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Given an untrimmed video and a sentence, the goal of temporal sentence grounding is to localize relevant moments \\(\\{\\varphi_{n} = (t_{s_{n}}, t_{e_{n}})\\}_{n=1}^{N}\\), where \\(N\\) denotes the number of ground truths in the video and \\(\\varphi_{n}\\) indicates the temporal interval of the \\(n\\)-th moment. Note that a video may have multiple moments that match the sentence, i.e., \\(N \\geq 1\\). During the test time, the model is expected to produce a total of \\(M\\) predictions, \\(\\{(\\hat{\\varphi}_{m}, q_{m})\\}_{m=1}^{M}\\), where \\(\\hat{\\varphi}_{m} = (\\hat{t}_{s_{m}}, \\hat{t}_{e_{m}})\\) is the \\(m\\)-th prediction, while \\(q_{m}\\) is its score for ranking." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.644, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAM-DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.525, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.261, + 0.46, + 0.271 + ], + "angle": 0, + "content": "(a) Overall framework of BAM-DETR" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.144, + 0.784, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.261, + 0.735, + 0.272 + ], + "angle": 0, + "content": "(b) Dual-pathway decoding layer" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.282, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Fig. 2: (a) Overview of the proposed BAM-DETR. (b) Details of the proposed dual-pathway decoding layer. It consists of two parallel pathways respectively for anchor and boundary updates, which refine previous moment predictions in a sequential manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.785, + 0.518 + ], + "angle": 0, + "content": "Motivation. We tackle two main problems in the current prediction process. On one hand, existing query-based approaches adopt the symmetric design of \\((c,l)\\) to predict a moment, i.e., \\(\\hat{\\varphi} = (c - 0.5l, c + 0.5l)\\). As we discuss in Sec. 1, this strategy suffers from the issue of over-reliance on the center prediction, leading to unstable performance. To address this, we propose a boundary-oriented modeling with a triplet of \\((p,d_s,d_e)\\), where a moment is represented by \\(\\hat{\\varphi} = (p - d_s,p + d_e)\\). This asymmetric design enables direct alignment of boundaries without relying on precise center prediction. On the other hand, previous models use classification scores as \\(q_{m}\\). This scoring is prone to sub-optimal solutions since only a fraction of the moment may well match the sentence, leading incomplete predictions to be highly ranked. To handle this, we propose quality-based scoring to sort the proposals based on their localization qualities rather than the degree of matching." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.54, + 0.341, + 0.554 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.785, + 0.626 + ], + "angle": 0, + "content": "As shown in Fig. 2a, our BAM-DETR follows the encoder-decoder pipeline. Briefly, it first extracts multimodal video features using the encoders. Taking them as memory, the dual-pathway decoder predicts temporal spans and their quality scores by progressively refining learnable initial spans." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.419, + 0.66 + ], + "angle": 0, + "content": "3.2 Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Following the convention [42, 70, 72], we employ pre-trained encoders for token-level unimodal feature extraction. It is worth noting that all unimodal encoders are kept frozen during training to avoid memory exploding. In this stage, we obtain \\( D_v \\)-dim video features \\( E_v \\in \\mathbb{R}^{N_v \\times D_v} \\) and \\( D_t \\)-dim text features \\( E_t \\in \\mathbb{R}^{N_t \\times D_t} \\), where \\( N_v \\) and \\( N_t \\) are the numbers of clips and words, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.434, + 0.784 + ], + "angle": 0, + "content": "3.3 Multimodal Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.841 + ], + "angle": 0, + "content": "We feed the unimodal features of the video and text to the multimodal encoder, so as to fuse them into text-aware video representations for temporal sentence grounding. While various multimodal encoders are explored [21,28,37], we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.129 + ], + "angle": 0, + "content": "P. Lee and H. Byun" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "adopt the text-to-video encoder design [41] consisting of cross- and self-attention blocks [54]. Before getting into the encoder, unimodal features are projected into a shared space to facilitate cross-modal interaction, i.e., \\(\\mathcal{V} = f_v(E_v) \\in \\mathbb{R}^{N_v \\times D}\\), \\(\\mathcal{T} = f_t(E_t) \\in \\mathbb{R}^{N_t \\times D}\\), where \\(D\\) is the embedding dimension. Afterward, multi-head cross-attention blocks are used to inject textual information into the clip-level video representations. In specific, we project \\(\\mathcal{V}\\) to \\(\\mathbf{Q}_{\\mathcal{V}}\\) (query) while \\(\\mathcal{T}\\) to \\(\\mathbf{K}_{\\mathcal{T}}\\) (key) and \\(\\mathbf{V}_{\\mathcal{T}}\\) (value). Then a cross-attention block can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.257, + 0.786, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} ^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\mathcal {V}} \\mathbf {K} _ {\\mathcal {T}} ^ {\\top}}{\\sqrt {D}}\\right) \\mathbf {V} _ {\\mathcal {T}} + \\mathcal {V}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.292, + 0.531, + 0.308 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} ^ {\\prime \\prime} = \\operatorname {F F N} (\\mathcal {V} ^ {\\prime}) + \\mathcal {V} ^ {\\prime},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.784, + 0.37 + ], + "angle": 0, + "content": "where \\(\\mathrm{FFN}(\\cdot)\\) is a feed-forward network. Although we here present the singlehead attention block, it can readily generalize to a multi-headed version [54]. We denote the resulting multimodal representations obtained after \\(L_{E}\\) multi-head cross-attention blocks by \\(\\tilde{\\mathcal{V}}\\in \\mathbb{R}^{N_v\\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.477 + ], + "angle": 0, + "content": "Subsequently, self-attention blocks are leveraged to enhance the representations by allowing the inter-clip interaction. Here we project \\(\\tilde{\\mathcal{V}}\\) to \\(\\mathbf{Q}_{\\tilde{\\mathcal{V}}}\\), \\(\\mathbf{K}_{\\tilde{\\mathcal{V}}}\\), and \\(\\mathbf{V}_{\\tilde{\\mathcal{V}}}\\). We note that the query and the key are supplemented with fixed sinusoidal positional encoding [3, 54] for temporal awareness. Then the self-attention block is defined in a similar way to Eq. (1) but with different inputs. The enhanced clip-level representations after \\(L_{E}\\) multi-head self-attention blocks are denoted by \\(\\hat{\\mathcal{V}} \\in \\mathbb{R}^{N_v \\times D}\\), which will serve as the memory for the decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.785, + 0.554 + ], + "angle": 0, + "content": "It is widely known that providing saliency guidance to the memory features helps the model to better understand the semantic relationship between the video and text [21, 28]. As in previous works [17, 26], we impose saliency score constraints on the memory features. Specifically, we leverage a saliency predictor \\( S(\\cdot) \\) and train the model with the following margin-based training objective." + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.565, + 0.785, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a r g i n}} = \\max (0, \\alpha + S (\\hat {v} ^ {\\text {l o w}}) - S (\\hat {v} ^ {\\text {h i g h}})), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.703 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a margin and \\((\\hat{v}^{\\mathrm{low}},\\hat{v}^{\\mathrm{high}})\\) is the sampled feature pair satisfying that the saliency label of \\(\\hat{v}^{\\mathrm{low}}\\) is lower than that of \\(\\hat{v}^{\\mathrm{high}}\\). In case of the absence of saliency labels, we collect clips within and outside the ground-truth moment intervals to build a pair. In addition, we employ the rank-aware contrastive loss \\(\\mathcal{L}_{\\mathrm{cont}}\\) and the negative relation loss \\(\\mathcal{L}_{\\mathrm{neg}}\\), following Moon et al. [41]. Due to space limits, we refer the readers to Appendix for the loss formulations. In summary, the overall saliency loss is defined as \\(\\mathcal{L}_{\\mathrm{sal}} = \\mathcal{L}_{\\mathrm{margin}} + \\mathcal{L}_{\\mathrm{cont}} + \\mathcal{L}_{\\mathrm{neg}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.452, + 0.739 + ], + "angle": 0, + "content": "3.4 Dual-pathway Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.843 + ], + "angle": 0, + "content": "With the multimodal representations \\(\\hat{\\mathcal{V}}\\) as memory features, we aim to localize temporal spans corresponding to the sentence. We adopt the new boundary-oriented formulation for moment prediction, where each prediction is represented by a triplet of \\((p,d_s,d_e)\\), where \\(p\\) is the anchor point, while \\(d_{s}\\) and \\(d_{e}\\) are the distances from the anchor to the starting and ending points, respectively. To make full use of the proposed formulation, we design a dual-pathway decoding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAM-DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.225 + ], + "angle": 0, + "content": "layer with two parallel pathways (Fig. 2b). Formally, the inputs of the \\(l\\)-th layer are anchor queries \\(\\mathbf{C}_p^l \\in \\mathbb{R}^{M \\times D}\\), boundary queries \\(\\mathbf{C}_s^l\\), \\(\\mathbf{C}_e^l \\in \\mathbb{R}^{M \\times D}\\), and previous moment predictions \\(\\mathbf{A}^l = [\\mathbf{p}^l; \\mathbf{d}_s^l; \\mathbf{d}_e^l] \\in \\mathbb{R}^{M \\times 3}\\), where \\(M\\) is the number of queries (predictions). Note that the initial queries and spans, i.e., \\(\\{\\mathbf{C}_p^0, \\mathbf{C}_s^0, \\mathbf{C}_e^0, \\mathbf{A}^0\\}\\), are learnable parameters. We elaborate on the two pathways in the following." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.25, + 0.789, + 0.446 + ], + "angle": 0, + "content": "Anchor updating pathway. Given the predictions and the anchor queries from the preceding layer, the goal of this pathway is to adjust the anchor position so that the boundaries can be predicted based on it. Intuitively, in order to obtain a valuable anchor point without redundancy, an anchor query should communicate with other queries as well as the memory features. To this goal, the anchor updating pathway consists of a self-attention layer, a cross-attention layer, and a feed-forward network. In the self-attention layer, the anchor queries \\(\\mathbf{C}_p^l\\) are first projected into \\(\\mathbf{Q}_{\\mathbf{C}_p^l}\\), \\(\\mathbf{K}_{\\mathbf{C}_p^l}\\), and \\(\\mathbf{V}_{\\mathbf{C}_p^l}\\). Since an anchor query itself lacks positional information, we build positional encoding. Following the previous works [17,41], we extend the current spans \\(\\mathbf{A}^l\\) to the positional information of the queries, i.e., \\(\\mathbf{P}_{\\mathbf{A}^l} = \\mathrm{MLP}(\\mathrm{PE}(\\mathbf{A}^l)) \\in \\mathbb{R}^{M \\times D}\\), where \\(\\mathrm{PE}(\\cdot)\\) denotes the point-wise mapping from a position to the corresponding sinusoidal encoding while \\(\\mathrm{MLP}(\\cdot)\\) is multi-layer perceptron. The self-attention for anchor queries is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.45, + 0.787, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {C}} _ {p} ^ {l} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {Q} _ {\\mathbf {C} _ {p} ^ {l}} + \\mathbf {P} _ {\\mathbf {A} ^ {l}}\\right) \\left(\\mathbf {K} _ {\\mathbf {C} _ {p} ^ {l}} + \\mathbf {P} _ {\\mathbf {A} ^ {l}}\\right) ^ {\\top}}{\\sqrt {D}}\\right) \\mathbf {V} _ {\\mathbf {C} _ {p} ^ {l}} + \\mathbf {C} _ {p} ^ {l}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.493, + 0.788, + 0.622 + ], + "angle": 0, + "content": "After inter-query interaction, we employ a global cross-attention layer to aggregate multi-modal features from the memory. The anchor queries \\(\\tilde{\\mathbf{C}}_p^l\\) is projected into \\(\\mathbf{Q}_{\\tilde{\\mathbf{C}}_p^l}\\) while the memory \\(\\hat{\\nu}\\) is projected to \\(\\mathbf{K}_{\\hat{\\nu}}\\) and \\(\\mathbf{V}_{\\hat{\\nu}}\\). To make the query location-aware, we leverage the sinusoidal encoding of current anchor positions, i.e., \\(\\mathbf{P_{p^l}} = \\mathrm{PE}(\\mathbf{p}^l)\\in \\mathbb{R}^{M\\times D}\\). Similarly, the memory leverages the positional encoding, i.e., \\(\\mathbf{P}_{\\hat{\\nu}} = \\mathrm{PE}(\\hat{\\nu})\\in \\mathbb{R}^{N_v\\times D}\\). We use concatenation instead of summation to separate the roles of features and positional encoding [35, 40]. The cross-attention between anchor queries and the memory can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.309, + 0.626, + 0.787, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {C}} _ {p} ^ {l} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {Q} _ {\\tilde {\\mathbf {C}} _ {p} ^ {l}} \\parallel \\mathbf {P} _ {\\mathbf {p} ^ {l}}\\right)\\left(\\mathbf {K} _ {\\hat {\\mathcal {V}}} \\parallel \\mathbf {P} _ {\\hat {\\mathcal {V}}}\\right)\\right) ^ {\\top}}{\\sqrt {2 D}}\\left. \\right) \\mathbf {V} _ {\\hat {\\mathcal {V}}} + \\tilde {\\mathbf {C}} _ {p} ^ {l}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.671, + 0.788, + 0.741 + ], + "angle": 0, + "content": "After all, the anchor queries are updated with a feed-forward network, i.e., \\(\\mathbf{C}_p^{(l + 1)} = \\mathrm{FFN}(\\hat{\\mathbf{C}}_p^l) + \\hat{\\mathbf{C}}_p^l\\). Lastly, we adjust the anchor positions using sigmoid-based refinement [78]: \\(\\hat{\\mathbf{A}}^l = [\\mathbf{p}^{(l + 1)};\\mathbf{d}_s^l;\\mathbf{d}_e^l]\\) where \\(\\mathbf{p}^{(l + 1)} = \\sigma (\\sigma^{-1}(\\mathbf{p}^l) + \\varDelta \\mathbf{p}^l)\\) with \\(\\varDelta\\mathbf{p}^l=\\mathrm{MLP}(\\mathbf{C}_p^{(l+1)})\\in\\mathbb{R}^M\\) and the sigmoid function \\(\\sigma(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Boundary updating pathway. After the anchor update, we refine the boundaries of the predictions. It is widely perceived that a model needs to focus on fine-grained features in the neighborhood rather than far ones to adjust temporal boundaries [27, 29, 51]. Inspired by this, we devise a boundary-focused attention layer (Fig. 3). For brevity, we explain the process of starting boundary update." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.129 + ], + "angle": 0, + "content": "P. Lee and H. Byun" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.145, + 0.681, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.334, + 0.704, + 0.348 + ], + "angle": 0, + "content": "Fig. 3: Boundary-focused attention layer for starting queries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.473 + ], + "angle": 0, + "content": "The plain memory \\(\\hat{\\mathcal{V}}\\) lacks the inductive bias due to the property of attentional layers [8]. Thus we first build locality-enhanced memory features for effective boundary refinement. To this goal, we obtain boundary-sensitive features with several 1D convolutional layers, i.e., \\(\\hat{\\mathcal{V}}_{\\mathrm{s}} = f_{\\mathrm{s}}(\\hat{\\mathcal{V}})\\). Then we encourage them to highly activate around the starting position of target moments. In detail, we impose regularization on the clip-wise activation scores obtained by channel-mean, i.e., \\(\\hat{g}^s = \\mathrm{mean}(\\sigma (\\hat{\\mathcal{V}}_s))\\in \\mathbb{R}^{N_v}\\). The regularization loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.316, + 0.477, + 0.786, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g u l}} ^ {s} = - \\frac {1}{N _ {v}} \\sum_ {i = 1} ^ {N _ {v}} \\left(g _ {i} ^ {s} \\log \\left(\\hat {g} _ {i} ^ {s}\\right) + \\left(1 - g _ {i} ^ {s}\\right) \\log \\left(1 - \\hat {g} _ {i} ^ {s}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.602 + ], + "angle": 0, + "content": "where \\( g_{i}^{s} \\) is the binary label obtained as \\( g_{i}^{s} = \\mathbb{1}[i\\in \\mathcal{B}^{s}] \\), where \\( \\mathcal{B}^s \\) is the neighbor clip set around starting points with a radius of \\( r^s \\) (set to 1/10 of the moment length). We merge boundary-sensitive features with the plain ones to form locality-enhanced memory, i.e., \\( \\hat{\\mathcal{V}}_s' = [\\hat{\\mathcal{V}}\\parallel \\hat{\\mathcal{V}}_s]\\in \\mathbb{R}^{N_v\\times 2D} \\). We can obtain \\( \\mathcal{L}_{\\mathrm{regul}}^e \\) and \\( \\hat{\\mathcal{V}}_e^\\prime \\) in the same way. The total regularization term is \\( \\mathcal{L}_{\\mathrm{regul}} = \\mathcal{L}_{\\mathrm{regul}}^s +\\mathcal{L}_{\\mathrm{regul}}^e \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Given the locality-enhanced memory, the boundary queries need to capture fine-grained details around the boundaries for refinement. To efficiently aggregate useful features near the boundaries, we employ deformable attention [77, 78]. Regarding the starting boundary, i.e., \\(\\mathbf{p}^{(l + 1)} - \\mathbf{d}_s^l\\), as the origin, we predict offsets and weights to select \\(K\\) neighbors, i.e., \\(\\mathbf{o}^s = \\phi_o(\\mathbf{C}_s^l)\\in \\mathbb{R}^{M\\times K}\\), \\(\\mathbf{w}^s = \\mathrm{softmax}(\\phi_w(\\mathbf{C}_s^l))\\in \\mathbb{R}^{M\\times K}\\), where \\(\\phi_{*}\\) are fully-connected layers. Features from the sampled neighbors are then aggregated into the starting queries as:" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.714, + 0.786, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {C}} _ {s} ^ {l} = \\sum_ {k = 1} ^ {K} \\left[ \\mathbf {w} _ {k} ^ {s} \\cdot \\hat {\\mathcal {V}} _ {s} ^ {\\prime} \\left[ \\mathbf {p} ^ {(l + 1)} - \\mathbf {d} _ {s} ^ {l} + \\mathbf {o} _ {k} ^ {s} \\right] \\right] + \\mathbf {C} _ {s} ^ {l}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.76, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where we denote the sampling process from memory by \\(\\hat{\\mathcal{V}}_s^{\\prime}[\\cdot]\\). Lastly, we adopt a feed-forward network to obtain the updated starting queries, i.e., \\(\\mathbf{C}_s^{(l + 1)} = \\mathrm{FFN}(\\hat{\\mathbf{C}}_s^l)\\). We also obtain \\(\\mathbf{C}_e^{(l + 1)}\\) in the same way. With the updated queries, we refine the boundaries to be better aligned with those of ground truths using sigmoid-based refinement as similar in the anchor update, leading to the refined" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAM-DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.209 + ], + "angle": 0, + "content": "predictions \\(\\mathbf{A}^{(l + 1)} = [\\mathbf{p}^{(l + 1)};\\mathbf{d}_s^{(l + 1)};\\mathbf{d}_e^{(l + 1)}]\\). Note that we utilize the deformable attention for the purpose of local feature aggregation, which sharply differs from the original purpose of efficient multi-scale global operation [78]. We provide comparison experiments regarding their roles in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.221, + 0.784, + 0.298 + ], + "angle": 0, + "content": "Moment prediction. We repeatedly update the predictions through a total of \\(L_{D}\\) dual-pathway decoding layers. We denote the resulting predictions by \\(\\mathbf{A} = \\left[\\mathbf{p};\\mathbf{d}_s;\\mathbf{d}_e\\right]\\), the anchor queries by \\(\\mathbf{C}_p\\), and the boundary queries by \\(\\mathbf{C}_s\\) and \\(\\mathbf{C}_e\\). Then we cast the predictions in the form of starting and ending timestamps, i.e., \\(\\hat{\\varphi} = \\left[\\mathbf{p} - \\mathbf{d}_s;\\mathbf{p} + \\mathbf{d}_e\\right] \\in \\mathbb{R}^{M \\times 2}\\), which serves as the final results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.317, + 0.445, + 0.332 + ], + "angle": 0, + "content": "3.5 Quality-based Scoring" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.34, + 0.784, + 0.446 + ], + "angle": 0, + "content": "After producing the moment predictions, we opt to rank them for evaluation. In the convention, query-based models utilize classification scores as the measure, which exhibits how well the proposals semantically match the sentence. However, it does not necessarily represent the localization qualities of proposals. Hence we propose to estimate the localization quality of each moment prediction. Formally, the quality score can be derived as \\(\\mathbf{q} = \\sigma (\\mathrm{MLP}([\\mathbf{C}_p\\parallel \\mathbf{C}_s\\parallel \\mathbf{C}_e]))\\in \\mathbb{R}^M\\), where \\(\\sigma\\) is sigmoid activation. Then the quality loss is defined as follows." + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.455, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {q u a l}} = \\sum_ {m = 1} ^ {M} \\left| q _ {m} - \\max _ {\\forall n} \\left(\\frac {\\left| \\hat {\\varphi} _ {m} \\cap \\varphi_ {n} \\right|}{\\left| \\hat {\\varphi} _ {m} \\cup \\varphi_ {n} \\right|}\\right) \\right|, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.503, + 0.784, + 0.534 + ], + "angle": 0, + "content": "where the objective of the quality head is to predict the maximum IoUs of the proposals with ground-truth moments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.342, + 0.569 + ], + "angle": 0, + "content": "3.6 Matching" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.577, + 0.784, + 0.622 + ], + "angle": 0, + "content": "As in the standard of query-based models [3,21], we perform Hungarian matching [19] between predictions and ground truths. The optimal matching results \\(\\psi^{*}\\) can be derived as follows." + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.624, + 0.785, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\psi^ {*} = \\arg \\min _ {\\psi \\in \\mathfrak {G} _ {N}} \\sum_ {n = 1} ^ {N} \\mathcal {C} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.669, + 0.698, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right) = \\lambda_ {l _ {1}} \\mathcal {L} _ {l _ {1}} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right) + \\lambda_ {\\mathrm {i o u}} \\mathcal {L} _ {\\mathrm {i o u}} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.688, + 0.784, + 0.763 + ], + "angle": 0, + "content": "where \\(\\mathfrak{G}_N\\) denotes the combination pool and \\(\\psi (n)\\) is the index of the prediction matched by the \\(n\\)-th ground truth. \\(\\mathcal{L}_{l_1}\\) and \\(\\mathcal{L}_{\\mathrm{iou}}\\) respectively represent the \\(L_{1}\\) distance and the generalized IoU [46] between the moments, while \\(\\lambda_{l_1}\\) and \\(\\lambda_{\\mathrm{iou}}\\) are their weights. Note that in contrast to other works, the classification term is not involved in the matching process, leading to localization-oriented matching." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.764, + 0.784, + 0.794 + ], + "angle": 0, + "content": "Once the matching is completed, we minimize the matching cost between each pair of the matching results \\(\\psi^{*}\\). The localization loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.803, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {l o c}} = \\sum_ {n = 1} ^ {N} \\mathcal {C} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi^ {*} (n)}\\right). \\tag {9}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.129 + ], + "angle": 0, + "content": "P. Lee and H. Byun" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Overall training objectives. Our model is trained in an end-to-end fashion and the overall training objective is defined as follows." + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.181, + 0.786, + 0.198 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {l o c}} + \\lambda_ {\\text {q u a l}} \\mathcal {L} _ {\\text {q u a l}} + \\lambda_ {\\text {s a l}} \\mathcal {L} _ {\\text {s a l}} + \\lambda_ {\\text {r e g u l}} \\mathcal {L} _ {\\text {r e g u l}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.2, + 0.502, + 0.216 + ], + "angle": 0, + "content": "where \\(\\lambda_{*}\\) are the balancing parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.238, + 0.376, + 0.256 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.27, + 0.446, + 0.286 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.787, + 0.461 + ], + "angle": 0, + "content": "Datasets. QVHighlights [21] is a recently built dataset, containing a total of 10,148 videos and 10,310 sentences from vlog and news domains. In addition to moments, it provides segment-level saliency score annotations within each moment. Importantly, it allows a single sentence corresponding to multiple disjoint moments (1.8 on average). Due to this practical setup, we utilize QVHighlights as the main benchmark. Charades-STA [11] includes 16,128 sentence-moment pairs with 9,848 indoor videos. The average duration of videos and moments is 30.6 and 8.1 seconds, respectively. TACoS [45] contains 127 cooking videos encompassing a total of 18,818 sentence-moment pairs. This dataset is known to be challenging since the moments occupy only a small portion (6.1 sec on average) within considerably long videos (4.8 min on average)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.787, + 0.581 + ], + "angle": 0, + "content": "Evaluation metrics. Following the standard protocol, we measure the Recall@1 (R1) under the IoU thresholds of 0.3, 0.5, and 0.7 by default. Since QVHighlights contains multiple ground-truth moments per sentence, we report the mean average precision (mAP) with IoU thresholds of 0.5 and 0.75 as well as the average mAP over a set of IoU thresholds [0.5:0.05:0.95]. Meanwhile, we compute the mean IoU of top-1 predictions on Charades-STA and TACoS. Note that the performances at high IoU thresholds (e.g., 0.7) exhibit how well the predictions align with the ground truths." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.457, + 0.619 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.749 + ], + "angle": 0, + "content": "For a fair comparison, we adopt the same feature extraction strategy with the competitors [21,26,41]. Specifically, we adopt the CLIP [44] text features for text and the concatenation of Slowfast [10] (ResNet-50) and CLIP [44] (ViT-B/32) features for videos unless otherwise specified. The video features are extracted every 1 second for Charades-STA and 2 seconds for QVHighlihtgs and TACoS. To compare with audio-augmented models [37], we optionally employ audio features extracted by PANNs [18] pre-trained on AudioSet [14]. Due to the long video duration, we uniformly sample 200 feature vectors from each video for TACoS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We set the embedding dimension \\(D\\) to 256, the number of attention heads to 8, the number of queries \\(M\\) to 10, the number of boundary points \\(K\\) to 3, and the margin \\(\\alpha\\) to 0.2. We determine the numbers of encoding and decoding layers as same with the prior work [41], i.e., \\(L_{E} = L_{D} = 2\\). The balancing parameters are set as: \\(\\lambda_{l1} = 10\\), \\(\\lambda_{\\mathrm{iou}} = \\lambda_{\\mathrm{sal}} = \\lambda_{\\mathrm{regul}} = 1\\), \\(\\lambda_{\\mathrm{qual}} = 2\\). As in previous works [26,41], we increase \\(\\lambda_{\\mathrm{sal}}\\) to 4 when saliency labels are unavailable, i.e., for Charades-STA" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAM-DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.338, + 0.145, + 0.665, + 0.16 + ], + "angle": 0, + "content": "Table 2: Results on the QVHighlights test split." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.17, + 0.723, + 0.464 + ], + "angle": 0, + "content": "
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
w/o pre-training
MCN [1]11.412.7224.948.2210.67
CAL [9]25.4911.5423.407.659.89
CLIP [44]16.885.1918.117.007.67
XML [22]46.6933.4647.8934.6734.90
Moment-DETR [21]52.8933.0254.8229.4030.73
UMT† [37]56.2341.1853.8337.0136.12
MH-DETR [60]60.0542.4860.7538.1338.38
QD-DETR [41]62.4044.9862.5239.8839.86
QD-DETR† [41]63.0645.1063.0440.1040.19
UniVTG [28]58.8640.8657.6035.5935.47
EaTR‡ [17]57.9842.4159.9539.2939.00
MomentDiff [26]57.4239.6654.0235.7335.95
BAM-DETR62.7148.6464.5746.3345.36
BAM-DETR†64.0748.1265.6147.5146.91
w/ pre-training on 4.2M data labeled by CLIP
UniVTG [28]65.4350.0664.0645.0243.63
w/ pre-training on 236K ASR captions
Moment-DETR [21]59.7840.3360.5135.3636.14
UMT† [37]60.8343.2657.3339.1238.08
QD-DETR [41]63.1845.1963.3740.3539.96
BAM-DETR63.8847.9266.3348.2246.67
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.283, + 0.465, + 0.678, + 0.477 + ], + "angle": 0, + "content": "\\(\\dagger\\)additional use of audio modality \\(\\ddagger\\)reproduced by official checkpoint" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.534 + ], + "angle": 0, + "content": "and TACoS. Our model is trained from scratch for 200 epochs on QVHighlights and 100 epochs on the other datasets using the AdamW optimizer [38] with a learning rate of 1e-4 and a batch size of 32." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.557, + 0.546, + 0.572 + ], + "angle": 0, + "content": "4.3 Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Results on QVHighlights. We compare our model with existing state-of-the-arts including recent query-based approaches [17, 21, 26, 37, 41, 60] on the test split. As shown in Table 2, our BAM-DETR consistently outperforms the comparative models under various settings. In detail, without pre-training, our model surpasses the previous state-of-the-art model [41] by large margins, e.g., \\(3.66\\%\\) in R1@0.7 and \\(6.45\\%\\) in mAP@0.75. These improvements under strict IoU thresholds verify the superior localization ability of our method. When leveraging auxiliary audio features, the performance further boosts especially in R1@0.5 and mAPs, enlarging the gap between the competitors including those with audio features [37, 41]. To compare with the methods with pretraining, we pretrain our model on middle-scale ASR caption data [21]. Again, our BAM-DETR achieves state-of-the-art results, while showing the least gap with the method [28] that leverages large-scale data for pre-training. Notably, our model even outperforms it in terms of mAPs with much fewer (about \\(18\\times\\)) pre-training data, manifesting the effectiveness of the proposed methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Results on Charades-STA. Experimental results on the test split are shown in Table 3. Not only does our BAM-DETR outperform the query-based com" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "P. Lee and H. Byun" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.145, + 0.71, + 0.16 + ], + "angle": 0, + "content": "Table 3: Results on the Charades-STA and TACoS test splits." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.17, + 0.782, + 0.292 + ], + "angle": 0, + "content": "
MethodCharades-STATACoS
R1@0.3R1@0.5R1@0.7mIoUR1@0.3R1@0.5R1@0.7mIoU
2D-TAN [72]58.7646.0227.4041.2540.0127.9912.9227.22
VSLNet [70]60.3042.6924.1441.5835.5423.5413.1524.99
Moment-DETR [21]65.8352.0730.5945.5437.9724.6711.9725.49
QD-DETR [41]-57.3132.55-52.3936.7721.0735.76
UniVTG [28]70.8158.0135.6550.1051.4434.9717.3533.60
MomentDiff [26]-55.5732.42-46.6428.9212.3730.36
BAM-DETR72.9359.9539.3852.3356.6941.5426.7739.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.298, + 0.785, + 0.327 + ], + "angle": 0, + "content": "Table 4: Results on the anti-biased Charades-STA test split against the moment location and length. VGG [52] and Glove [43] features are employed for all models." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.337, + 0.782, + 0.447 + ], + "angle": 0, + "content": "
Methodw.r.t. moment locationw.r.t. moment length
R1@0.3R1@0.5R1@0.7\\( \\mathrm{mAP}_{avg} \\)R1@0.3R1@0.5R1@0.7\\( \\mathrm{mAP}_{avg} \\)
2D-TAN [72]27.8120.4410.8417.2339.6828.6817.7222.79
MMN [56]33.5827.2014.1219.1843.5834.3119.9426.85
Moment-DETR [21]29.9422.1611.5618.6642.7334.3916.1224.02
\\( QD-DETR^‡ \\)[41]56.1746.8228.1330.7067.3954.4432.8736.99
MomentDiff [26]48.3933.5915.7121.3751.2538.3223.3828.19
BAM-DETR59.8350.0032.0831.6868.4055.4640.7443.21
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.223, + 0.448, + 0.411, + 0.459 + ], + "angle": 0, + "content": " reproduced by official codebase" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.606 + ], + "angle": 0, + "content": "petitors [21, 26, 41], but it achieves a new state-of-the-art by surpassing the best performing anchor-free model [28] for all metrics. Notably, a large gap of \\(3.73\\%\\) is observed in R1@0.7, which confirms the strong localization ability of our model. Results on TACoS. We present the comparison results on the test set in Table 3. It can be observed that our BAM-DETR achieves a new state-of-the-art with pronounced performances under the strict IoU thresholds, which is consistent with the above results on other datasets. On this challenging benchmark, our model surpasses the previous best model [41] by \\(5.7\\%\\) (relatively \\(27\\%\\)) in R1@0.7. These results clearly exhibit the superiority of the proposed model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.45, + 0.639 + ], + "angle": 0, + "content": "4.4 Robustness Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.784 + ], + "angle": 0, + "content": "Query-based models potentially have a temporal bias [16, 63] against the locations and lengths of moments. To measure robustness, we evaluate our model on the anti-biased Charades-STA [26] with distribution shifts of the moment location and length between training and test sets. Table 4 summarizes the results, where our model outperforms all competitors under both anti-biased settings. Especially, it shows significant performance gaps under the moment length bias. This can be expected since our model directly localizes boundaries instead of predicting lengths, which lessens the effect of bias. This robustness test corroborates the advantage of our boundary-oriented moment modeling." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.803, + 0.334, + 0.818 + ], + "angle": 0, + "content": "4.5 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.741, + 0.841 + ], + "angle": 0, + "content": "We conduct analytical experiments on the QVHighlights validation split." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAM-DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.306, + 0.145, + 0.696, + 0.16 + ], + "angle": 0, + "content": "Table 5: Ablation study of components on QVHighlights." + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.17, + 0.728, + 0.254 + ], + "angle": 0, + "content": "
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
Baseline62.3947.8762.6441.5441.75
+ boundary-oriented modeling63.4249.2362.8643.2442.42
+ dual-pathway decoder63.6150.2663.0144.9844.16
+ quality-based scoring65.1051.6165.4148.5647.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.263, + 0.782, + 0.278 + ], + "angle": 0, + "content": "Table 6: Ablative experiments on QVHighlights. Default settings are marked gray." + }, + { + "type": "table_caption", + "bbox": [ + 0.225, + 0.29, + 0.38, + 0.303 + ], + "angle": 0, + "content": "(a) Number of predictions" + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.304, + 0.383, + 0.387 + ], + "angle": 0, + "content": "
MR1mAP
@0.5@0.7Avg.
562.1949.0345.18
1065.1051.6147.61
1565.4850.3248.23
2065.2351.6148.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.398, + 0.29, + 0.577, + 0.302 + ], + "angle": 0, + "content": "(b) Choice of memory features" + }, + { + "type": "table", + "bbox": [ + 0.398, + 0.304, + 0.584, + 0.376 + ], + "angle": 0, + "content": "
MemoryR1mAP Avg.
@0.5@0.7
plain63.7449.6846.21
sensitive63.8149.8746.34
merged65.1051.6147.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.594, + 0.29, + 0.778, + 0.303 + ], + "angle": 0, + "content": "(c) Number of boundary points" + }, + { + "type": "table", + "bbox": [ + 0.599, + 0.304, + 0.783, + 0.387 + ], + "angle": 0, + "content": "
KR1mAP
@0.5@0.7Avg.
1 (fixed)63.2348.3245.17
164.3249.8747.14
365.1051.6147.61
564.9050.4546.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.391, + 0.791, + 0.404 + ], + "angle": 0, + "content": "(d) Query choices for quality prediction (e) Combinations of attention layers for updating pathways" + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.406, + 0.445, + 0.488 + ], + "angle": 0, + "content": "
CpCsCeR1mAP
@0.5@0.7Avg.
✓62.9749.0346.26
✓64.3950.7146.66
✓64.5849.2346.71
✓✓✓65.1051.6147.61
" + }, + { + "type": "table", + "bbox": [ + 0.458, + 0.406, + 0.79, + 0.488 + ], + "angle": 0, + "content": "
Anchor queryBoundary queryR1 @0.5@0.7mAP Avg.FLOPsParams
Global (shared)62.2649.0344.180.71G8.2M
GlobalGlobal63.6149.8144.540.72G11.5M
GlobalFocused63.7449.6846.210.65G9.5M
FocusedFocused62.4849.1645.870.62G7.6M
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.787, + 0.595 + ], + "angle": 0, + "content": "Effect of each component. We analyze the effect of each component in Table 5. The direct adoption of our moment modeling solely improves the performance, particularly at strict thresholds, showcasing its advantage in boundary alignment. Employing the dual-pathway decoder leads to further gains, which suggests the essential role of separate pathways. Lastly, the quality-based scoring considerably elevates the scores, especially in terms of mAPs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.643 + ], + "angle": 0, + "content": "Number of predictions. We experiment with varying numbers of predictions in Table 6a, where the model achieves robust results when \\( M \\) is sufficiently large. By default, we set \\( M \\) to 10 for a fair comparison with the previous works [21,41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.703 + ], + "angle": 0, + "content": "Locality-enhanced features. We analyze the effect of the choice of input features for boundary-focused attention in Table 6b. The results indicate that boundary-sensitive features are slightly more helpful than plain ones in precise localization. In addition, the merged features achieve the best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.794 + ], + "angle": 0, + "content": "Number of sampled boundary points. We analyze the effect of the number of sampled boundary points \\( K \\) in Table 6c. For comparisons, we report the case where the fixed boundary features are sampled without using offsets (\\( 1^{\\mathrm{st}} \\) row). As shown in the table, the dynamic selection of neighborhoods rather than fixed boundaries is important for accurate moment localization. The performance improves when sampling multiple points, while it saturates at \\( K = 3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Query choices for quality prediction. Our model leverages both anchor and boundary queries for quality prediction. We investigate the effect of query choices in Table 6d. As a result, utilizing all the queries shows better localization" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "P. Lee and H. Byun" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.146, + 0.391, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.24, + 0.37, + 0.251 + ], + "angle": 0, + "content": "(a) Boundary hit rate" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.144, + 0.788, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.24, + 0.671, + 0.251 + ], + "angle": 0, + "content": "(b) Visualization results" + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.263, + 0.665, + 0.277 + ], + "angle": 0, + "content": "Fig. 4: Analytical experiments on QVHighlights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.784, + 0.318 + ], + "angle": 0, + "content": "performance than using one of them. This might be trivial as different queries contain complementary information for quality prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.785, + 0.438 + ], + "angle": 0, + "content": "Attention layers. Table 6e compares different combinations of attention layers for updating pathways in terms of performances and costs. For an apple-to/apple comparison, we leverage the plain memory features for both attention layers. First of all, employing separate global cross-attention for different queries leads to a huge parameter increase yet limited score gains. Replacing the global one for boundary queries with our focused attention layer substantially reduces the cost, while achieving the best performance. Meanwhile, the use of focused attention for both types of queries leads to inferior results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.59 + ], + "angle": 0, + "content": "Boundary alignment. Standard IoU-based metrics are indirect measures for boundary alignment. To precisely diagnose the ability, inspired by Deeplab [5], we compute the boundary hit rate of predictions with varying band widths. We expand ground-truth boundary points with a band width to form starting/ending zones and regard a prediction as correct if both of its boundaries fall within the corresponding zones. More details can be found in Appendix. To disentangle the effect of ranking, we mark a video as correct if at least one prediction is correct and measure the video-level hit rate. As shown in Fig. 4a, our model greatly outperforms the recent competitors. The sharp increase in low band widths clearly validates the superiority of our model in boundary alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.785, + 0.667 + ], + "angle": 0, + "content": "Qualitative results. As shown in Fig. 4b, previous models fail to localize accurate moments with misleading center predictions. Especially, QD-DETR [41] accurately predicts the moment length but suffers from the misaligned center, leading to the limited IoU. In contrast, thanks to the novel moment modeling, our model predicts well-aligned boundaries without relying on center prediction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.36, + 0.705 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this paper, we identified the center misalignment issue of existing query-based models for sentence grounding. To address it, we presented boundary-oriented moment modeling where boundaries are directly predicted without relying on centers. Based on the modeling, we designed the boundary-aligned moment detection transformer characterized by dual-pathway decoding. Further, we proposed localization quality-based scoring of predictions. The efficacy of the proposed methods is validated by thorough examinations. We hope this work sheds light on the issue of center-based moment modeling in detection transformers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAM-DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.22 + ], + "angle": 0, + "content": "This project was supported by the National Research Foundation of Korea grant funded by the Korea government (MSIT) (No. 2022R1A2B5B02001467; RS-2024-00346364)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.241, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.785, + 0.308 + ], + "angle": 0, + "content": "1. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017) 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.338 + ], + "angle": 0, + "content": "2. Cao, M., Chen, L., Shou, M.Z., Zhang, C., Zou, Y.: On pursuit of designing multimodal transformer for video grounding. In: EMNLP. pp. 9810-9823 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.377 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229. Springer (2020) 2, 4, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "4. Chen, J., Chen, X., Ma, L., Jie, Z., Chua, T.S.: Temporally grounding natural sentence in video. In: EMNLP. pp. 162-171 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.446 + ], + "angle": 0, + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. PAMI 40(4), 834-848 (2017) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "6. Chen, L., Lu, C., Tang, S., Xiao, J., Zhang, D., Tan, C., Li, X.: Rethinking the bottom-up framework for query-based video localization. In: AAAI. vol. 34, pp. 10551-10558 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.526 + ], + "angle": 0, + "content": "7. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: ICCV. pp. 2988-2997 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "9. Escorcia, V., Soldan, M., Sivic, J., Ghanem, B., Russell, B.: Temporal localization of moments in video collections with natural language. arXiv preprint arXiv:1907.12763 (2019) 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "0. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6202-6211 (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.664 + ], + "angle": 0, + "content": "1. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: CVPR. pp. 5267-5275 (2017) 1, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "2. Gao, P., Zheng, M., Wang, X., Dai, J., Li, H.: Fast convergence of detr with spatially modulated co-attention. In: ICCV. pp. 3621-3630 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "3. Ge, R., Gao, J., Chen, K., Nevatia, R.: Mac: Mining activity concepts for language-based temporal localization. In: WACV. pp. 245-253. IEEE (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "4. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: ICASSP. pp. 776-780. IEEE (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.797 + ], + "angle": 0, + "content": "5. Ghosh, S., Agarwal, A., Parekh, Z., Hauptmann, A.G.: Excl: Extractive clip localization using natural language descriptions. In: NAACL. pp. 1984-1990 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Hao, J., Sun, H., Ren, P., Wang, J., Qi, Q., Liao, J.: Can shuffling video benefit temporal bias problem: A novel training framework for temporal grounding. In: ECCV. pp. 130-147. Springer (2022) 12" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.128 + ], + "angle": 0, + "content": "P. Lee and H. Byun" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "17. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: ICCV. pp. 13846-13856 (2023) 2, 4, 6, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "18. Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Trans. on Audio, Speech, and Language Processing 28, 2880-2894 (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.257 + ], + "angle": 0, + "content": "19. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "20. Lee, P., Kim, T., Shim, M., Wee, D., Byun, H.: Decomposed cross-modal distillation for rgb-based temporal action detection. In: CVPR. pp. 2373-2383 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.324 + ], + "angle": 0, + "content": "21. Lei, J., Berg, T.L., Bansal, M.: Detecting moments and highlights in videos via natural language queries. In: Neurips. vol. 34, pp. 11846-11858 (2021) 2, 3, 4, 5, 6, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "22. Lei, J., Yu, L., Berg, T.L., Bansal, M.: Tv: A large-scale dataset for video-subtitle moment retrieval. In: ECCV. pp. 447-463. Springer (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.39 + ], + "angle": 0, + "content": "23. Li, F., Zeng, A., Liu, S., Zhang, H., Li, H., Zhang, L., Ni, L.M.: Lite detr: An interleaved multi-scale encoder for efficient detr. In: CVPR. pp. 18558-18567 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.42 + ], + "angle": 0, + "content": "24. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: CVPR. pp. 13619-13627 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "25. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In: ICCV. pp. 12032-12042 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "26. Li, P., Xie, C.W., Xie, H., Zhao, L., Zhang, L., Zheng, Y., Zhao, D., Zhang, Y.: Momentdiff: Generative video moment retrieval from random to real. In: Neurips (2023) 2, 4, 6, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "27. Lin, C., Xu, C., Luo, D., Wang, Y., Tai, Y., Wang, C., Li, J., Huang, F., Fu, Y.: Learning salient boundary feature for anchor-free temporal action localization. In: CVPR. pp. 3320-3329 (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "28. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: ICCV. pp. 2794-2804 (2023) 4, 5, 6, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.61 + ], + "angle": 0, + "content": "29. Lin, T., Zhao, X., Su, H., Wang, C., Yang, M.: Bsn: Boundary sensitive network for temporal action proposal generation. In: ECCV. pp. 3-19 (2018) 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "30. Lin, Y., Yuan, Y., Zhang, Z., Li, C., Zheng, N., Hu, H.: Detr does not need multiscale or locality design. In: ICCV. pp. 6545-6554 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "31. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: CVPR. pp. 11235-11244 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "32. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: CVPR. pp. 11235-11244 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.746 + ], + "angle": 0, + "content": "33. Liu, F., Wei, H., Zhao, W., Li, G., Peng, J., Li, Z.: Wb-detr: transformer-based detector without backbone. In: ICCV. pp. 2979-2987 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "34. Liu, M., Wang, X., Nie, L., Tian, Q., Chen, B., Chua, T.S.: Cross-modal moment localization in videos. In: ACM MM. pp. 843-851 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "35. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. In: ICLR (2022) 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., Zhang, L.: Detection transformer with stable matching. In: ICCV. pp. 6491-6500 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAM-DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: CVPR. pp. 3042-3051 (2022) 2, 4, 5, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.216 + ], + "angle": 0, + "content": "38. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.217, + 0.787, + 0.258 + ], + "angle": 0, + "content": "39. Lu, C., Chen, L., Tan, C., Li, X., Xiao, J.: Debug: A dense bottom-up grounding approach for natural language video localization. In: EMNLP-IJCNLP. pp. 5144-5153 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.298 + ], + "angle": 0, + "content": "40. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: ICCV. pp. 3651-3660 (2021) 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "41. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: CVPR. pp. 23023-23033 (2023) 2, 3, 4, 6, 7, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "42. Mun, J., Cho, M., Han, B.: Local-global video-text interactions for temporal grounding. In: CVPR. pp. 10810-10819 (2020) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.397 + ], + "angle": 0, + "content": "43. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP. pp. 1532-1543 (2014) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.478 + ], + "angle": 0, + "content": "45. Regneri, M., Rohrbach, M., Wetzel, D., Thater, S., Schiele, B., Pinkal, M.: Grounding action descriptions in videos. Trans. Assoc. Comput. Linguistics 1, 25-36 (2013) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.522 + ], + "angle": 0, + "content": "46. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: CVPR. pp. 658-666 (2019) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.55 + ], + "angle": 0, + "content": "47. Roh, B., Shin, J., Shin, W., Kim, S.: Sparse detr: Efficient end-to-end object detection with learnable sparsity. In: ICLR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "48. Rui, Y., Gupta, A., Acero, A.: Automatically extracting highlights for tv baseball programs. In: ACM MM. pp. 105-115 (2000) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "49. Shao, D., Xiong, Y., Zhao, Y., Huang, Q., Qiao, Y., Lin, D.: Find and focus: Retrieve and localize video events with natural language queries. In: ECCV. pp. 200-216 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "50. Sharghi, A., Laurel, J.S., Gong, B.: Query-focused video summarization: Dataset, evaluation, and a memory network based approach. In: CVPR. pp. 4788-4797 (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "51. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: CVPR. pp. 18857-18866 (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "52. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.787, + 0.744 + ], + "angle": 0, + "content": "53. Sun, Z., Cao, S., Yang, Y., Kitani, K.M.: Rethinking transformer-based set prediction for object detection. In: ICCV. pp. 3611-3620 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.771 + ], + "angle": 0, + "content": "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Neurips. vol. 30 (2017) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "55. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: AAAI. vol. 36, pp. 2567-2575 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "56. Wang, Z., Wang, L., Wu, T., Li, T., Wu, G.: Negative sample matters: A renaissance of metric learning for temporal grounding. In: AAAI. vol. 36, pp. 2613-2623 (2022) 4, 12" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.129 + ], + "angle": 0, + "content": "P. Lee and H. Byun" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "57. Xiao, S., Chen, L., Zhang, S., Ji, W., Shao, J., Ye, L., Xiao, J.: Boundary proposal network for two-stage natural language video localization. In: AAAI. vol. 35, pp. 2986-2994 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "58. Xiong, B., Kalantidis, Y., Ghadiyaram, D., Grauman, K.: Less is more: Learning highlight detection from video duration. In: CVPR. pp. 1258-1267 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "59. Xu, H., He, K., Plummer, B.A., Sigal, L., Sclaroff, S., Saenko, K.: Multilevel language and vision integration for text-to-clip retrieval. In: AAAI. vol. 33, pp. 9062–9069 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "60. Xu, Y., Sun, Y., Li, Y., Shi, Y., Zhu, X., Du, S.: Mh-detr: Video moment and highlight detection with cross-modal transformer. In: ACM MM (2023) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "61. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: Unloc: A unified framework for video localization tasks. In: ICCV. pp. 13623-13633 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "62. Ye, M., Ke, L., Li, S., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Cascade-detr: Delving into high-quality universal object detection. In: ICCV. pp. 6704-6714 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "63. Yuan, Y., Lan, X., Wang, X., Chen, L., Wang, Z., Zhu, W.: A closer look at temporal sentence grounding in videos: Dataset and metric. In: Proc. 2nd Int. Workshop on Human-Centric Multimedia Analysis. pp. 13-21 (2021) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.451 + ], + "angle": 0, + "content": "64. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In: Neurips. vol. 32 (2019) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.452, + 0.785, + 0.493 + ], + "angle": 0, + "content": "65. Yuan, Y., Mei, T., Zhu, W.: To find where you talk: Temporal sentence localization in video with attention based location regression. In: AAAI. vol. 33, pp. 9159-9166 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "66. Zeng, R., Xu, H., Huang, W., Chen, P., Tan, M., Gan, C.: Dense regression network for video grounding. In: CVPR. pp. 10287-10296 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "67. Zhang, D., Dai, X., Wang, X., Wang, Y.F., Davis, L.S.: Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In: CVPR. pp. 1247-1257 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "68. Zhang, G., Luo, Z., Yu, Y., Cui, K., Lu, S.: Accelerating detr convergence via semantic-aligned matching. In: CVPR. pp. 949-958 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "69. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: ICLR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "70. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. In: ACL (2020) 4, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "71. Zhang, K., Chao, W.L., Sha, F., Grauman, K.: Video summarization with long short-term memory. In: ECCV. pp. 766-782. Springer (2016) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "72. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. vol. 34, pp. 12870-12877 (2020) 1, 4, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.757 + ], + "angle": 0, + "content": "73. Zhang, S., Su, J., Luo, J.: Exploiting temporal relationships in video moment localization with natural language. In: ACM MM. pp. 1230-1238 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "74. Zhang, Z., Lin, Z., Zhao, Z., Xiao, Z.: Cross-modal interaction networks for query-based moment retrieval in videos. In: ACM SIGIR. pp. 655-664 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "75. Zhao, Y., Xiong, Y., Wang, L., Wu, Z., Tang, X., Lin, D.: Temporal action detection with structured segment networks. In: ICCV. pp. 2914-2923 (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "76. Zheng, D., Dong, W., Hu, H., Chen, X., Wang, Y.: Less is more: Focus attention for efficient detr. In: ICCV. pp. 6674-6683 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAM-DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "77. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: CVPR. pp. 9308-9316 (2019) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.205 + ], + "angle": 0, + "content": "78. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021) 4, 7, 8, 9" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.205 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_origin.pdf b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..66ba61dc752cf353deb70482faee3526c6c65e18 --- /dev/null +++ b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/cc01728d-bb7f-4681-9e77-2025b1c8109a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df6c51126d29682f4286532309ce8f038ae0662cf66eb83edcd40a643f9c8ca5 +size 2013997 diff --git a/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/full.md b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d7809193bbf9644ba7e6f135c57c8be3bed31b3e --- /dev/null +++ b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/full.md @@ -0,0 +1,379 @@ +# BAM-DETR: Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos + +Pilhyeon Lee $^{1\star}$ and Hyeran Byun $^{2}$ + +$^{1}$ Department of Artificial Intelligence, Inha University + +$^{2}$ Department of Computer Science, Yonsei University + +Abstract. Temporal sentence grounding aims to localize moments relevant to a language description. Recently, DETR-like approaches achieved notable progress by predicting the center and length of a target moment. However, they suffer from the issue of center misalignment raised by the inherent ambiguity of moment centers, leading to inaccurate predictions. To remedy this problem, we propose a novel boundary-oriented moment formulation. In our paradigm, the model no longer needs to find the precise center but instead suffices to predict any anchor point within the interval, from which the boundaries are directly estimated. Based on this idea, we design a boundary-aligned moment detection transformer, equipped with a dual-pathway decoding process. Specifically, it refines the anchor and boundaries within parallel pathways using global and boundary-focused attention, respectively. This separate design allows the model to focus on desirable regions, enabling precise refinement of moment predictions. Further, we propose a quality-based ranking method, ensuring that proposals with high localization qualities are prioritized over incomplete ones. Experiments on three benchmarks validate the effectiveness of the proposed methods. The code is available here. + +Keywords: Temporal sentence grounding $\cdot$ Detection transformer + +# 1 Introduction + +Recent years have witnessed a notable surge in the popularity of short-form video content on social media platforms like TikTok, YouTube Shorts, and Instagram Reels. As such, users prefer to selectively engage with short moments of interest rather than passively watch an entire long video. This trend highlights the importance of localizing desired moments. As a result, moment localization tasks have emerged as pivotal research topics in video understanding, including temporal action detection [20, 29, 75], video summarization [50, 71], and highlight detection [48, 58]. Within this context, we tackle temporal sentence grounding [1, 72], aiming to retrieve moments corresponding to free-form language descriptions. + +To address temporal sentence grounding, numerous efforts have been undertaken in the last decade [9, 11, 31, 49, 64, 72]. Especially, taking inspiration + +![](images/cd3dbc8d5ee3eec6576c86a0597040e9692a12aa97821ba3cf81a4369a1ed038.jpg) +(a) Conventional moment modeling + +![](images/98965eb2948b6a3bd61e7f7cd5ca7f366452887af5193e63ad134e5ce0069035.jpg) +(b) Boundary-oriented moment modeling (Ours) +Fig. 1: Comparison of moment modeling approaches under the scenario of an ambiguous center from QVHighlights. (a) The conventional method formulates a moment with a tuple of $(c, l)$ . (b) In contrast, we propose to model it with a triplet of $(p, d_s, d_e)$ . + +from DETR [3], query-based approaches have become a promising research direction owing to the architectural simplicity [17,21,26,37,41]. By decoding temporal spans (i.e., moments) from a handful set of learnable queries, they achieve promising grounding performance while maintaining high inference speed. + +Existing query-based models typically predict a moment using its center and width, i.e., $(c, l)$ , under the assumption that the boundaries are equidistant from the center. However, such a formulation can be problematic as the center of a moment might be ambiguous. An illustrative example is presented in Fig. 1 (top), where the frame at the center of the ground-truth moment is less relevant to the given sentence, as demonstrated by the low saliency label. Indeed, a moment center does not necessarily serve as the best representative of the sentence. This ambiguity can challenge the model's ability to precisely locate the centers, and the misaligned centers lead to low-quality predictions (Fig. 1a). To probe the impact of center misalignment, we conducted a diagnostic experiment in Table 1. The results reveal that (i) existing methods struggle with detecting accurate center points and (ii) they suffer from significant performance drops when the predicted centers deviate from the ground-truth centers (i.e., large center errors). + +To address this challenge, we present a novel boundary-oriented formulation for moments, as illustrated in Fig. 1b, where each moment is represented by a triplet consisting of an anchor point and its distances to the boundaries, i.e., $(p,d_s,d_e)$ . This asymmetric formulation liberates the model from the stringent requirement of predicting the center. Instead, it is sufficient for the model to predict any salient anchor within the target moment, and the distances from the anchor to the onset and offset are predicted subsequently. By directly locating the boundaries based on the anchor point, the model can achieve improved boundary alignment, even when the anchor point does not coincide with the actual center. + +Building upon the proposed moment modeling, we introduce a new framework equipped with a dedicated decoder design, dubbed Boundary-Aligned Moment Detection Transformer (BAM-DETR). The design of its decoding layers originates from the intuition that the refining processes for an anchor and boundaries should be distinct from each other. That is to say, a model needs to scan over the whole video to find a potential anchor point that enables estimating the + +Table 1: Impact of misaligned centers on QVHighlights. The top-1 predictions are grouped based on their center errors normalized by ground-truth lengths. For each group, we present the mean IoU (\%) and the proportion (in parentheses). Only the predictions whose centers fall within the ground-truth moments are considered here. + +
Method[0, 0.1)[0.1, 0.2)[0.2, 0.3)[0.3, 0.4)[0.4, 0.5)All
Moment-DETR [21]83.2065.0054.8243.7734.9867.84
(46 %)(24 %)(14 %)(9 %)(7 %)(100 %)
QD-DETR [41]87.7967.9555.4044.9336.4574.09
(56 %)(18 %)(11 %)(9 %)(5 %)(100 %)
BAM-DETR†77.6278.2077.5278.9671.0877.21
(Ours)(24 %)(21 %)(21 %)(22 %)(12 %)(100 %)
+ +$^\dagger$ anchor points are utilized for grouping. + +rough location of the target moment. On the other hand, it is required to focus on fine-grained details in the vicinity to refine the boundaries to be aligned with those of the target moments. From this motivation, in contrast to existing methods, our model adopts a dual-pathway decoding pipeline to predict an anchor point and boundaries in a parallel way. To be specific, we leverage two different types of queries respectively for anchor and boundary refinement. The former aggregates global information with standard attention, while the latter concentrates on the sparse local neighborhood of the boundaries using the proposed boundary-focused attention. These distinct designs of two pathways allow for effective moment localization with minimal computational overhead increase. + +In addition, we identify the problem of the conventional scoring method, where binary classification (or matching) scores are used for proposal ranking. This leads to suboptimal results for the grounding task since a fractional moment may have high matching scores with the given sentence. To handle this issue, we propose to rank proposals based on their localization qualities. Accordingly, we modify the typical matching function and training objectives of the query-based model to be localization-oriented by discarding the role of classification scores. In this way, our model can prioritize the moment proposals exhibiting high overlap with ground truths at inference, leading to improved grounding performance. + +The advantage of our BAM-DETR is showcased in Table 1. It can be observed that the anchor points predicted by our model are evenly distributed within the ground-truth intervals. Importantly, our model consistently produces precise moments with high IoUs ( $>0.7$ ) across different groups, indicating that it does not depend on accurate center prediction. On average, our model shows superior grounding performance over the existing methods with the help of our moment formulation and quality-based scoring. In a later section, we validate the efficacy of the proposed methods through extensive experiments. Notably, our model outperforms previous methods by large margins on three public benchmarks. + +# 2 Related Works + +# 2.1 Temporal Sentence Grounding in Videos + +Temporal sentence grounding requires seeking temporal spans semantically relevant to the given sentence in a video. Proposal-based approaches adopt the + +two-stage pipeline, i.e., proposal generation and ranking. They generate moment proposals by relying on sliding windows [11,13,34,73] or utilizing pre-defined anchors [4,59,64,67,74]. Several works process all possible candidates at once with 2D maps [25,32,56,57,72]. Meanwhile, proposal-free methods are developed for efficient grounding by directly predicting the moments [42,65] or estimating the probabilities of each frame being starting and ending positions [15,70]. Some approaches perform dense regression by predicting the boundaries from individual frames [6,39,66]. Recently, query-based models streamline the complicated sentence grounding pipeline by removing handcrafted techniques [2,17,26,37,60]. There are also attempts to unify temporal sentence grounding with other video understanding tasks into a single framework [28,61]. + +Our method belongs to the query-based group [21, 41], inheriting the benefit of architectural simplicity. In contrast to others, we employ a boundary-oriented formulation of moments to relieve the heavy reliance on center predictions, leading to better boundary alignment. Our method also relates to dense regression methods [6, 28, 66] that predict boundaries from each frame as an anchor. Comparatively, our model leverages dynamic anchors that are gradually adjusted through decoding, enabling precise grounding using a small set of predictions. + +# 2.2 Detection Transformers + +Query-based temporal sentence grounding models by design are closely related to the family of detection transformers (DETR) [3]. Since the advent of DETR, a number of variants have been introduced to improve it from various perspectives [30, 33, 36, 53]. Some works focus on reducing the excessive computational costs of vanilla transformers in order to leverage multi-scale features [7, 23, 47, 76, 78]. On the other hand, several methods attempt to speed up the model convergence by manipulating the attention operations [12, 35, 62, 68] or incorporating the denoising process during the model training [24, 69]. + +The most relevant works to ours are those which propose explicit anchor modeling for object queries using center points [40, 55] or boxes (center, width, and height) [35, 78]. In comparison to an object, a moment in temporal sentence grounding has its own challenges such as center ambiguity and indistinct boundaries. To accommodate the discrepancies, we propose a novel boundary-oriented modeling of moments to replace the conventional center-based 1D box modeling. The advantages of our approach are clearly verified in the experiments. + +# 3 Method + +Given an untrimmed video and a sentence, the goal of temporal sentence grounding is to localize relevant moments $\{\varphi_{n} = (t_{s_{n}}, t_{e_{n}})\}_{n=1}^{N}$ , where $N$ denotes the number of ground truths in the video and $\varphi_{n}$ indicates the temporal interval of the $n$ -th moment. Note that a video may have multiple moments that match the sentence, i.e., $N \geq 1$ . During the test time, the model is expected to produce a total of $M$ predictions, $\{(\hat{\varphi}_{m}, q_{m})\}_{m=1}^{M}$ , where $\hat{\varphi}_{m} = (\hat{t}_{s_{m}}, \hat{t}_{e_{m}})$ is the $m$ -th prediction, while $q_{m}$ is its score for ranking. + +![](images/2903c8ae1d069fab841fdb58bbd2833f062e86bbe46ab08b5242bbd0b0b33af6.jpg) +(a) Overall framework of BAM-DETR + +![](images/7f141d3e762f7ff75e61dd3db547f118e69356d1ff696dc37eb7aaa8b94a2349.jpg) +(b) Dual-pathway decoding layer +Fig. 2: (a) Overview of the proposed BAM-DETR. (b) Details of the proposed dual-pathway decoding layer. It consists of two parallel pathways respectively for anchor and boundary updates, which refine previous moment predictions in a sequential manner. + +Motivation. We tackle two main problems in the current prediction process. On one hand, existing query-based approaches adopt the symmetric design of $(c,l)$ to predict a moment, i.e., $\hat{\varphi} = (c - 0.5l, c + 0.5l)$ . As we discuss in Sec. 1, this strategy suffers from the issue of over-reliance on the center prediction, leading to unstable performance. To address this, we propose a boundary-oriented modeling with a triplet of $(p,d_s,d_e)$ , where a moment is represented by $\hat{\varphi} = (p - d_s,p + d_e)$ . This asymmetric design enables direct alignment of boundaries without relying on precise center prediction. On the other hand, previous models use classification scores as $q_{m}$ . This scoring is prone to sub-optimal solutions since only a fraction of the moment may well match the sentence, leading incomplete predictions to be highly ranked. To handle this, we propose quality-based scoring to sort the proposals based on their localization qualities rather than the degree of matching. + +# 3.1 Overview + +As shown in Fig. 2a, our BAM-DETR follows the encoder-decoder pipeline. Briefly, it first extracts multimodal video features using the encoders. Taking them as memory, the dual-pathway decoder predicts temporal spans and their quality scores by progressively refining learnable initial spans. + +# 3.2 Feature Extraction + +Following the convention [42, 70, 72], we employ pre-trained encoders for token-level unimodal feature extraction. It is worth noting that all unimodal encoders are kept frozen during training to avoid memory exploding. In this stage, we obtain $D_v$ -dim video features $E_v \in \mathbb{R}^{N_v \times D_v}$ and $D_t$ -dim text features $E_t \in \mathbb{R}^{N_t \times D_t}$ , where $N_v$ and $N_t$ are the numbers of clips and words, respectively. + +# 3.3 Multimodal Encoder + +We feed the unimodal features of the video and text to the multimodal encoder, so as to fuse them into text-aware video representations for temporal sentence grounding. While various multimodal encoders are explored [21,28,37], we + +adopt the text-to-video encoder design [41] consisting of cross- and self-attention blocks [54]. Before getting into the encoder, unimodal features are projected into a shared space to facilitate cross-modal interaction, i.e., $\mathcal{V} = f_v(E_v) \in \mathbb{R}^{N_v \times D}$ , $\mathcal{T} = f_t(E_t) \in \mathbb{R}^{N_t \times D}$ , where $D$ is the embedding dimension. Afterward, multi-head cross-attention blocks are used to inject textual information into the clip-level video representations. In specific, we project $\mathcal{V}$ to $\mathbf{Q}_{\mathcal{V}}$ (query) while $\mathcal{T}$ to $\mathbf{K}_{\mathcal{T}}$ (key) and $\mathbf{V}_{\mathcal{T}}$ (value). Then a cross-attention block can be formulated as: + +$$ +\mathcal {V} ^ {\prime} = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q} _ {\mathcal {V}} \mathbf {K} _ {\mathcal {T}} ^ {\top}}{\sqrt {D}}\right) \mathbf {V} _ {\mathcal {T}} + \mathcal {V}, \tag {1} +$$ + +$$ +\mathcal {V} ^ {\prime \prime} = \operatorname {F F N} (\mathcal {V} ^ {\prime}) + \mathcal {V} ^ {\prime}, +$$ + +where $\mathrm{FFN}(\cdot)$ is a feed-forward network. Although we here present the singlehead attention block, it can readily generalize to a multi-headed version [54]. We denote the resulting multimodal representations obtained after $L_{E}$ multi-head cross-attention blocks by $\tilde{\mathcal{V}}\in \mathbb{R}^{N_v\times D}$ . + +Subsequently, self-attention blocks are leveraged to enhance the representations by allowing the inter-clip interaction. Here we project $\tilde{\mathcal{V}}$ to $\mathbf{Q}_{\tilde{\mathcal{V}}}$ , $\mathbf{K}_{\tilde{\mathcal{V}}}$ , and $\mathbf{V}_{\tilde{\mathcal{V}}}$ . We note that the query and the key are supplemented with fixed sinusoidal positional encoding [3, 54] for temporal awareness. Then the self-attention block is defined in a similar way to Eq. (1) but with different inputs. The enhanced clip-level representations after $L_{E}$ multi-head self-attention blocks are denoted by $\hat{\mathcal{V}} \in \mathbb{R}^{N_v \times D}$ , which will serve as the memory for the decoder. + +It is widely known that providing saliency guidance to the memory features helps the model to better understand the semantic relationship between the video and text [21, 28]. As in previous works [17, 26], we impose saliency score constraints on the memory features. Specifically, we leverage a saliency predictor $S(\cdot)$ and train the model with the following margin-based training objective. + +$$ +\mathcal {L} _ {\text {m a r g i n}} = \max (0, \alpha + S (\hat {v} ^ {\text {l o w}}) - S (\hat {v} ^ {\text {h i g h}})), \tag {2} +$$ + +where $\alpha$ is a margin and $(\hat{v}^{\mathrm{low}},\hat{v}^{\mathrm{high}})$ is the sampled feature pair satisfying that the saliency label of $\hat{v}^{\mathrm{low}}$ is lower than that of $\hat{v}^{\mathrm{high}}$ . In case of the absence of saliency labels, we collect clips within and outside the ground-truth moment intervals to build a pair. In addition, we employ the rank-aware contrastive loss $\mathcal{L}_{\mathrm{cont}}$ and the negative relation loss $\mathcal{L}_{\mathrm{neg}}$ , following Moon et al. [41]. Due to space limits, we refer the readers to Appendix for the loss formulations. In summary, the overall saliency loss is defined as $\mathcal{L}_{\mathrm{sal}} = \mathcal{L}_{\mathrm{margin}} + \mathcal{L}_{\mathrm{cont}} + \mathcal{L}_{\mathrm{neg}}$ . + +# 3.4 Dual-pathway Decoder + +With the multimodal representations $\hat{\mathcal{V}}$ as memory features, we aim to localize temporal spans corresponding to the sentence. We adopt the new boundary-oriented formulation for moment prediction, where each prediction is represented by a triplet of $(p,d_s,d_e)$ , where $p$ is the anchor point, while $d_{s}$ and $d_{e}$ are the distances from the anchor to the starting and ending points, respectively. To make full use of the proposed formulation, we design a dual-pathway decoding + +layer with two parallel pathways (Fig. 2b). Formally, the inputs of the $l$ -th layer are anchor queries $\mathbf{C}_p^l \in \mathbb{R}^{M \times D}$ , boundary queries $\mathbf{C}_s^l$ , $\mathbf{C}_e^l \in \mathbb{R}^{M \times D}$ , and previous moment predictions $\mathbf{A}^l = [\mathbf{p}^l; \mathbf{d}_s^l; \mathbf{d}_e^l] \in \mathbb{R}^{M \times 3}$ , where $M$ is the number of queries (predictions). Note that the initial queries and spans, i.e., $\{\mathbf{C}_p^0, \mathbf{C}_s^0, \mathbf{C}_e^0, \mathbf{A}^0\}$ , are learnable parameters. We elaborate on the two pathways in the following. + +Anchor updating pathway. Given the predictions and the anchor queries from the preceding layer, the goal of this pathway is to adjust the anchor position so that the boundaries can be predicted based on it. Intuitively, in order to obtain a valuable anchor point without redundancy, an anchor query should communicate with other queries as well as the memory features. To this goal, the anchor updating pathway consists of a self-attention layer, a cross-attention layer, and a feed-forward network. In the self-attention layer, the anchor queries $\mathbf{C}_p^l$ are first projected into $\mathbf{Q}_{\mathbf{C}_p^l}$ , $\mathbf{K}_{\mathbf{C}_p^l}$ , and $\mathbf{V}_{\mathbf{C}_p^l}$ . Since an anchor query itself lacks positional information, we build positional encoding. Following the previous works [17,41], we extend the current spans $\mathbf{A}^l$ to the positional information of the queries, i.e., $\mathbf{P}_{\mathbf{A}^l} = \mathrm{MLP}(\mathrm{PE}(\mathbf{A}^l)) \in \mathbb{R}^{M \times D}$ , where $\mathrm{PE}(\cdot)$ denotes the point-wise mapping from a position to the corresponding sinusoidal encoding while $\mathrm{MLP}(\cdot)$ is multi-layer perceptron. The self-attention for anchor queries is defined as: + +$$ +\tilde {\mathbf {C}} _ {p} ^ {l} = \operatorname {s o f t m a x} \left(\frac {\left(\mathbf {Q} _ {\mathbf {C} _ {p} ^ {l}} + \mathbf {P} _ {\mathbf {A} ^ {l}}\right) \left(\mathbf {K} _ {\mathbf {C} _ {p} ^ {l}} + \mathbf {P} _ {\mathbf {A} ^ {l}}\right) ^ {\top}}{\sqrt {D}}\right) \mathbf {V} _ {\mathbf {C} _ {p} ^ {l}} + \mathbf {C} _ {p} ^ {l}. \tag {3} +$$ + +After inter-query interaction, we employ a global cross-attention layer to aggregate multi-modal features from the memory. The anchor queries $\tilde{\mathbf{C}}_p^l$ is projected into $\mathbf{Q}_{\tilde{\mathbf{C}}_p^l}$ while the memory $\hat{\nu}$ is projected to $\mathbf{K}_{\hat{\nu}}$ and $\mathbf{V}_{\hat{\nu}}$ . To make the query location-aware, we leverage the sinusoidal encoding of current anchor positions, i.e., $\mathbf{P_{p^l}} = \mathrm{PE}(\mathbf{p}^l)\in \mathbb{R}^{M\times D}$ . Similarly, the memory leverages the positional encoding, i.e., $\mathbf{P}_{\hat{\nu}} = \mathrm{PE}(\hat{\nu})\in \mathbb{R}^{N_v\times D}$ . We use concatenation instead of summation to separate the roles of features and positional encoding [35, 40]. The cross-attention between anchor queries and the memory can be expressed as: + +$$ +\hat {\mathbf {C}} _ {p} ^ {l} = \operatorname {s o f t m a x} \left(\frac {\left(\mathbf {Q} _ {\tilde {\mathbf {C}} _ {p} ^ {l}} \parallel \mathbf {P} _ {\mathbf {p} ^ {l}}\right)\left(\mathbf {K} _ {\hat {\mathcal {V}}} \parallel \mathbf {P} _ {\hat {\mathcal {V}}}\right)\right) ^ {\top}}{\sqrt {2 D}}\left. \right) \mathbf {V} _ {\hat {\mathcal {V}}} + \tilde {\mathbf {C}} _ {p} ^ {l}. \tag {4} +$$ + +After all, the anchor queries are updated with a feed-forward network, i.e., $\mathbf{C}_p^{(l + 1)} = \mathrm{FFN}(\hat{\mathbf{C}}_p^l) + \hat{\mathbf{C}}_p^l$ . Lastly, we adjust the anchor positions using sigmoid-based refinement [78]: $\hat{\mathbf{A}}^l = [\mathbf{p}^{(l + 1)};\mathbf{d}_s^l;\mathbf{d}_e^l]$ where $\mathbf{p}^{(l + 1)} = \sigma (\sigma^{-1}(\mathbf{p}^l) + \varDelta \mathbf{p}^l)$ with $\varDelta\mathbf{p}^l=\mathrm{MLP}(\mathbf{C}_p^{(l+1)})\in\mathbb{R}^M$ and the sigmoid function $\sigma(\cdot)$ . + +Boundary updating pathway. After the anchor update, we refine the boundaries of the predictions. It is widely perceived that a model needs to focus on fine-grained features in the neighborhood rather than far ones to adjust temporal boundaries [27, 29, 51]. Inspired by this, we devise a boundary-focused attention layer (Fig. 3). For brevity, we explain the process of starting boundary update. + +![](images/037939b1153ca7e602e2d4e2373f4bc11f4f85ee6c8db0801185ac7a7dace92d.jpg) +Fig. 3: Boundary-focused attention layer for starting queries. + +The plain memory $\hat{\mathcal{V}}$ lacks the inductive bias due to the property of attentional layers [8]. Thus we first build locality-enhanced memory features for effective boundary refinement. To this goal, we obtain boundary-sensitive features with several 1D convolutional layers, i.e., $\hat{\mathcal{V}}_{\mathrm{s}} = f_{\mathrm{s}}(\hat{\mathcal{V}})$ . Then we encourage them to highly activate around the starting position of target moments. In detail, we impose regularization on the clip-wise activation scores obtained by channel-mean, i.e., $\hat{g}^s = \mathrm{mean}(\sigma (\hat{\mathcal{V}}_s))\in \mathbb{R}^{N_v}$ . The regularization loss is defined as: + +$$ +\mathcal {L} _ {\text {r e g u l}} ^ {s} = - \frac {1}{N _ {v}} \sum_ {i = 1} ^ {N _ {v}} \left(g _ {i} ^ {s} \log \left(\hat {g} _ {i} ^ {s}\right) + \left(1 - g _ {i} ^ {s}\right) \log \left(1 - \hat {g} _ {i} ^ {s}\right)\right), \tag {5} +$$ + +where $g_{i}^{s}$ is the binary label obtained as $g_{i}^{s} = \mathbb{1}[i\in \mathcal{B}^{s}]$ , where $\mathcal{B}^s$ is the neighbor clip set around starting points with a radius of $r^s$ (set to 1/10 of the moment length). We merge boundary-sensitive features with the plain ones to form locality-enhanced memory, i.e., $\hat{\mathcal{V}}_s' = [\hat{\mathcal{V}}\parallel \hat{\mathcal{V}}_s]\in \mathbb{R}^{N_v\times 2D}$ . We can obtain $\mathcal{L}_{\mathrm{regul}}^e$ and $\hat{\mathcal{V}}_e^\prime$ in the same way. The total regularization term is $\mathcal{L}_{\mathrm{regul}} = \mathcal{L}_{\mathrm{regul}}^s +\mathcal{L}_{\mathrm{regul}}^e$ . + +Given the locality-enhanced memory, the boundary queries need to capture fine-grained details around the boundaries for refinement. To efficiently aggregate useful features near the boundaries, we employ deformable attention [77, 78]. Regarding the starting boundary, i.e., $\mathbf{p}^{(l + 1)} - \mathbf{d}_s^l$ , as the origin, we predict offsets and weights to select $K$ neighbors, i.e., $\mathbf{o}^s = \phi_o(\mathbf{C}_s^l)\in \mathbb{R}^{M\times K}$ , $\mathbf{w}^s = \mathrm{softmax}(\phi_w(\mathbf{C}_s^l))\in \mathbb{R}^{M\times K}$ , where $\phi_{*}$ are fully-connected layers. Features from the sampled neighbors are then aggregated into the starting queries as: + +$$ +\hat {\mathbf {C}} _ {s} ^ {l} = \sum_ {k = 1} ^ {K} \left[ \mathbf {w} _ {k} ^ {s} \cdot \hat {\mathcal {V}} _ {s} ^ {\prime} \left[ \mathbf {p} ^ {(l + 1)} - \mathbf {d} _ {s} ^ {l} + \mathbf {o} _ {k} ^ {s} \right] \right] + \mathbf {C} _ {s} ^ {l}, \tag {6} +$$ + +where we denote the sampling process from memory by $\hat{\mathcal{V}}_s^{\prime}[\cdot]$ . Lastly, we adopt a feed-forward network to obtain the updated starting queries, i.e., $\mathbf{C}_s^{(l + 1)} = \mathrm{FFN}(\hat{\mathbf{C}}_s^l)$ . We also obtain $\mathbf{C}_e^{(l + 1)}$ in the same way. With the updated queries, we refine the boundaries to be better aligned with those of ground truths using sigmoid-based refinement as similar in the anchor update, leading to the refined + +predictions $\mathbf{A}^{(l + 1)} = [\mathbf{p}^{(l + 1)};\mathbf{d}_s^{(l + 1)};\mathbf{d}_e^{(l + 1)}]$ . Note that we utilize the deformable attention for the purpose of local feature aggregation, which sharply differs from the original purpose of efficient multi-scale global operation [78]. We provide comparison experiments regarding their roles in Appendix. + +Moment prediction. We repeatedly update the predictions through a total of $L_{D}$ dual-pathway decoding layers. We denote the resulting predictions by $\mathbf{A} = \left[\mathbf{p};\mathbf{d}_s;\mathbf{d}_e\right]$ , the anchor queries by $\mathbf{C}_p$ , and the boundary queries by $\mathbf{C}_s$ and $\mathbf{C}_e$ . Then we cast the predictions in the form of starting and ending timestamps, i.e., $\hat{\varphi} = \left[\mathbf{p} - \mathbf{d}_s;\mathbf{p} + \mathbf{d}_e\right] \in \mathbb{R}^{M \times 2}$ , which serves as the final results. + +# 3.5 Quality-based Scoring + +After producing the moment predictions, we opt to rank them for evaluation. In the convention, query-based models utilize classification scores as the measure, which exhibits how well the proposals semantically match the sentence. However, it does not necessarily represent the localization qualities of proposals. Hence we propose to estimate the localization quality of each moment prediction. Formally, the quality score can be derived as $\mathbf{q} = \sigma (\mathrm{MLP}([\mathbf{C}_p\parallel \mathbf{C}_s\parallel \mathbf{C}_e]))\in \mathbb{R}^M$ , where $\sigma$ is sigmoid activation. Then the quality loss is defined as follows. + +$$ +\mathcal {L} _ {\text {q u a l}} = \sum_ {m = 1} ^ {M} \left| q _ {m} - \max _ {\forall n} \left(\frac {\left| \hat {\varphi} _ {m} \cap \varphi_ {n} \right|}{\left| \hat {\varphi} _ {m} \cup \varphi_ {n} \right|}\right) \right|, \tag {7} +$$ + +where the objective of the quality head is to predict the maximum IoUs of the proposals with ground-truth moments. + +# 3.6 Matching + +As in the standard of query-based models [3,21], we perform Hungarian matching [19] between predictions and ground truths. The optimal matching results $\psi^{*}$ can be derived as follows. + +$$ +\psi^ {*} = \arg \min _ {\psi \in \mathfrak {G} _ {N}} \sum_ {n = 1} ^ {N} \mathcal {C} \left(\varphi_ {n}, \hat {\varphi} _ {\psi (n)}\right), \tag {8} +$$ + +$$ +\mathcal {C} \left(\varphi_ {n}, \hat {\varphi} _ {\psi (n)}\right) = \lambda_ {l _ {1}} \mathcal {L} _ {l _ {1}} \left(\varphi_ {n}, \hat {\varphi} _ {\psi (n)}\right) + \lambda_ {\mathrm {i o u}} \mathcal {L} _ {\mathrm {i o u}} \left(\varphi_ {n}, \hat {\varphi} _ {\psi (n)}\right), +$$ + +where $\mathfrak{G}_N$ denotes the combination pool and $\psi (n)$ is the index of the prediction matched by the $n$ -th ground truth. $\mathcal{L}_{l_1}$ and $\mathcal{L}_{\mathrm{iou}}$ respectively represent the $L_{1}$ distance and the generalized IoU [46] between the moments, while $\lambda_{l_1}$ and $\lambda_{\mathrm{iou}}$ are their weights. Note that in contrast to other works, the classification term is not involved in the matching process, leading to localization-oriented matching. + +Once the matching is completed, we minimize the matching cost between each pair of the matching results $\psi^{*}$ . The localization loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {l o c}} = \sum_ {n = 1} ^ {N} \mathcal {C} \left(\varphi_ {n}, \hat {\varphi} _ {\psi^ {*} (n)}\right). \tag {9} +$$ + +Overall training objectives. Our model is trained in an end-to-end fashion and the overall training objective is defined as follows. + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {l o c}} + \lambda_ {\text {q u a l}} \mathcal {L} _ {\text {q u a l}} + \lambda_ {\text {s a l}} \mathcal {L} _ {\text {s a l}} + \lambda_ {\text {r e g u l}} \mathcal {L} _ {\text {r e g u l}}, \tag {10} +$$ + +where $\lambda_{*}$ are the balancing parameters. + +# 4 Experiments + +# 4.1 Experimental Settings + +Datasets. QVHighlights [21] is a recently built dataset, containing a total of 10,148 videos and 10,310 sentences from vlog and news domains. In addition to moments, it provides segment-level saliency score annotations within each moment. Importantly, it allows a single sentence corresponding to multiple disjoint moments (1.8 on average). Due to this practical setup, we utilize QVHighlights as the main benchmark. Charades-STA [11] includes 16,128 sentence-moment pairs with 9,848 indoor videos. The average duration of videos and moments is 30.6 and 8.1 seconds, respectively. TACoS [45] contains 127 cooking videos encompassing a total of 18,818 sentence-moment pairs. This dataset is known to be challenging since the moments occupy only a small portion (6.1 sec on average) within considerably long videos (4.8 min on average). + +Evaluation metrics. Following the standard protocol, we measure the Recall@1 (R1) under the IoU thresholds of 0.3, 0.5, and 0.7 by default. Since QVHighlights contains multiple ground-truth moments per sentence, we report the mean average precision (mAP) with IoU thresholds of 0.5 and 0.75 as well as the average mAP over a set of IoU thresholds [0.5:0.05:0.95]. Meanwhile, we compute the mean IoU of top-1 predictions on Charades-STA and TACoS. Note that the performances at high IoU thresholds (e.g., 0.7) exhibit how well the predictions align with the ground truths. + +# 4.2 Implementation Details + +For a fair comparison, we adopt the same feature extraction strategy with the competitors [21,26,41]. Specifically, we adopt the CLIP [44] text features for text and the concatenation of Slowfast [10] (ResNet-50) and CLIP [44] (ViT-B/32) features for videos unless otherwise specified. The video features are extracted every 1 second for Charades-STA and 2 seconds for QVHighlihtgs and TACoS. To compare with audio-augmented models [37], we optionally employ audio features extracted by PANNs [18] pre-trained on AudioSet [14]. Due to the long video duration, we uniformly sample 200 feature vectors from each video for TACoS. + +We set the embedding dimension $D$ to 256, the number of attention heads to 8, the number of queries $M$ to 10, the number of boundary points $K$ to 3, and the margin $\alpha$ to 0.2. We determine the numbers of encoding and decoding layers as same with the prior work [41], i.e., $L_{E} = L_{D} = 2$ . The balancing parameters are set as: $\lambda_{l1} = 10$ , $\lambda_{\mathrm{iou}} = \lambda_{\mathrm{sal}} = \lambda_{\mathrm{regul}} = 1$ , $\lambda_{\mathrm{qual}} = 2$ . As in previous works [26,41], we increase $\lambda_{\mathrm{sal}}$ to 4 when saliency labels are unavailable, i.e., for Charades-STA + +Table 2: Results on the QVHighlights test split. + +
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
w/o pre-training
MCN [1]11.412.7224.948.2210.67
CAL [9]25.4911.5423.407.659.89
CLIP [44]16.885.1918.117.007.67
XML [22]46.6933.4647.8934.6734.90
Moment-DETR [21]52.8933.0254.8229.4030.73
UMT† [37]56.2341.1853.8337.0136.12
MH-DETR [60]60.0542.4860.7538.1338.38
QD-DETR [41]62.4044.9862.5239.8839.86
QD-DETR† [41]63.0645.1063.0440.1040.19
UniVTG [28]58.8640.8657.6035.5935.47
EaTR‡ [17]57.9842.4159.9539.2939.00
MomentDiff [26]57.4239.6654.0235.7335.95
BAM-DETR62.7148.6464.5746.3345.36
BAM-DETR†64.0748.1265.6147.5146.91
w/ pre-training on 4.2M data labeled by CLIP
UniVTG [28]65.4350.0664.0645.0243.63
w/ pre-training on 236K ASR captions
Moment-DETR [21]59.7840.3360.5135.3636.14
UMT† [37]60.8343.2657.3339.1238.08
QD-DETR [41]63.1845.1963.3740.3539.96
BAM-DETR63.8847.9266.3348.2246.67
+ +$\dagger$ additional use of audio modality $\ddagger$ reproduced by official checkpoint + +and TACoS. Our model is trained from scratch for 200 epochs on QVHighlights and 100 epochs on the other datasets using the AdamW optimizer [38] with a learning rate of 1e-4 and a batch size of 32. + +# 4.3 Comparison with State-of-the-arts + +Results on QVHighlights. We compare our model with existing state-of-the-arts including recent query-based approaches [17, 21, 26, 37, 41, 60] on the test split. As shown in Table 2, our BAM-DETR consistently outperforms the comparative models under various settings. In detail, without pre-training, our model surpasses the previous state-of-the-art model [41] by large margins, e.g., $3.66\%$ in R1@0.7 and $6.45\%$ in mAP@0.75. These improvements under strict IoU thresholds verify the superior localization ability of our method. When leveraging auxiliary audio features, the performance further boosts especially in R1@0.5 and mAPs, enlarging the gap between the competitors including those with audio features [37, 41]. To compare with the methods with pretraining, we pretrain our model on middle-scale ASR caption data [21]. Again, our BAM-DETR achieves state-of-the-art results, while showing the least gap with the method [28] that leverages large-scale data for pre-training. Notably, our model even outperforms it in terms of mAPs with much fewer (about $18\times$ ) pre-training data, manifesting the effectiveness of the proposed methods. + +Results on Charades-STA. Experimental results on the test split are shown in Table 3. Not only does our BAM-DETR outperform the query-based com + +Table 3: Results on the Charades-STA and TACoS test splits. + +
MethodCharades-STATACoS
R1@0.3R1@0.5R1@0.7mIoUR1@0.3R1@0.5R1@0.7mIoU
2D-TAN [72]58.7646.0227.4041.2540.0127.9912.9227.22
VSLNet [70]60.3042.6924.1441.5835.5423.5413.1524.99
Moment-DETR [21]65.8352.0730.5945.5437.9724.6711.9725.49
QD-DETR [41]-57.3132.55-52.3936.7721.0735.76
UniVTG [28]70.8158.0135.6550.1051.4434.9717.3533.60
MomentDiff [26]-55.5732.42-46.6428.9212.3730.36
BAM-DETR72.9359.9539.3852.3356.6941.5426.7739.31
+ +Table 4: Results on the anti-biased Charades-STA test split against the moment location and length. VGG [52] and Glove [43] features are employed for all models. + +
Methodw.r.t. moment locationw.r.t. moment length
R1@0.3R1@0.5R1@0.7\( \mathrm{mAP}_{avg} \)R1@0.3R1@0.5R1@0.7\( \mathrm{mAP}_{avg} \)
2D-TAN [72]27.8120.4410.8417.2339.6828.6817.7222.79
MMN [56]33.5827.2014.1219.1843.5834.3119.9426.85
Moment-DETR [21]29.9422.1611.5618.6642.7334.3916.1224.02
\( QD-DETR^‡ \)[41]56.1746.8228.1330.7067.3954.4432.8736.99
MomentDiff [26]48.3933.5915.7121.3751.2538.3223.3828.19
BAM-DETR59.8350.0032.0831.6868.4055.4640.7443.21
+ + reproduced by official codebase + +petitors [21, 26, 41], but it achieves a new state-of-the-art by surpassing the best performing anchor-free model [28] for all metrics. Notably, a large gap of $3.73\%$ is observed in R1@0.7, which confirms the strong localization ability of our model. Results on TACoS. We present the comparison results on the test set in Table 3. It can be observed that our BAM-DETR achieves a new state-of-the-art with pronounced performances under the strict IoU thresholds, which is consistent with the above results on other datasets. On this challenging benchmark, our model surpasses the previous best model [41] by $5.7\%$ (relatively $27\%$ ) in R1@0.7. These results clearly exhibit the superiority of the proposed model. + +# 4.4 Robustness Evaluation + +Query-based models potentially have a temporal bias [16, 63] against the locations and lengths of moments. To measure robustness, we evaluate our model on the anti-biased Charades-STA [26] with distribution shifts of the moment location and length between training and test sets. Table 4 summarizes the results, where our model outperforms all competitors under both anti-biased settings. Especially, it shows significant performance gaps under the moment length bias. This can be expected since our model directly localizes boundaries instead of predicting lengths, which lessens the effect of bias. This robustness test corroborates the advantage of our boundary-oriented moment modeling. + +# 4.5 Analysis + +We conduct analytical experiments on the QVHighlights validation split. + +Table 5: Ablation study of components on QVHighlights. + +
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
Baseline62.3947.8762.6441.5441.75
+ boundary-oriented modeling63.4249.2362.8643.2442.42
+ dual-pathway decoder63.6150.2663.0144.9844.16
+ quality-based scoring65.1051.6165.4148.5647.61
+ +Table 6: Ablative experiments on QVHighlights. Default settings are marked gray. +(a) Number of predictions + +
MR1mAP
@0.5@0.7Avg.
562.1949.0345.18
1065.1051.6147.61
1565.4850.3248.23
2065.2351.6148.01
+ +(b) Choice of memory features + +
MemoryR1mAP Avg.
@0.5@0.7
plain63.7449.6846.21
sensitive63.8149.8746.34
merged65.1051.6147.61
+ +(c) Number of boundary points + +
KR1mAP
@0.5@0.7Avg.
1 (fixed)63.2348.3245.17
164.3249.8747.14
365.1051.6147.61
564.9050.4546.70
+ +(d) Query choices for quality prediction (e) Combinations of attention layers for updating pathways + +
CpCsCeR1mAP
@0.5@0.7Avg.
✓62.9749.0346.26
✓64.3950.7146.66
✓64.5849.2346.71
✓✓✓65.1051.6147.61
+ +
Anchor queryBoundary queryR1 @0.5@0.7mAP Avg.FLOPsParams
Global (shared)62.2649.0344.180.71G8.2M
GlobalGlobal63.6149.8144.540.72G11.5M
GlobalFocused63.7449.6846.210.65G9.5M
FocusedFocused62.4849.1645.870.62G7.6M
+ +Effect of each component. We analyze the effect of each component in Table 5. The direct adoption of our moment modeling solely improves the performance, particularly at strict thresholds, showcasing its advantage in boundary alignment. Employing the dual-pathway decoder leads to further gains, which suggests the essential role of separate pathways. Lastly, the quality-based scoring considerably elevates the scores, especially in terms of mAPs. + +Number of predictions. We experiment with varying numbers of predictions in Table 6a, where the model achieves robust results when $M$ is sufficiently large. By default, we set $M$ to 10 for a fair comparison with the previous works [21,41]. + +Locality-enhanced features. We analyze the effect of the choice of input features for boundary-focused attention in Table 6b. The results indicate that boundary-sensitive features are slightly more helpful than plain ones in precise localization. In addition, the merged features achieve the best performance. + +Number of sampled boundary points. We analyze the effect of the number of sampled boundary points $K$ in Table 6c. For comparisons, we report the case where the fixed boundary features are sampled without using offsets ( $1^{\mathrm{st}}$ row). As shown in the table, the dynamic selection of neighborhoods rather than fixed boundaries is important for accurate moment localization. The performance improves when sampling multiple points, while it saturates at $K = 3$ . + +Query choices for quality prediction. Our model leverages both anchor and boundary queries for quality prediction. We investigate the effect of query choices in Table 6d. As a result, utilizing all the queries shows better localization + +![](images/a3183881a6d67efc4689850fe951014885c2be812c4cc1e918f170c9e79e6fe6.jpg) +(a) Boundary hit rate +Fig. 4: Analytical experiments on QVHighlights. + +![](images/f84c2f8966cf19bf98b821dd9769ed1120c79d299a12ce45c3faec02403e29e0.jpg) +(b) Visualization results + +performance than using one of them. This might be trivial as different queries contain complementary information for quality prediction. + +Attention layers. Table 6e compares different combinations of attention layers for updating pathways in terms of performances and costs. For an apple-to/apple comparison, we leverage the plain memory features for both attention layers. First of all, employing separate global cross-attention for different queries leads to a huge parameter increase yet limited score gains. Replacing the global one for boundary queries with our focused attention layer substantially reduces the cost, while achieving the best performance. Meanwhile, the use of focused attention for both types of queries leads to inferior results. + +Boundary alignment. Standard IoU-based metrics are indirect measures for boundary alignment. To precisely diagnose the ability, inspired by Deeplab [5], we compute the boundary hit rate of predictions with varying band widths. We expand ground-truth boundary points with a band width to form starting/ending zones and regard a prediction as correct if both of its boundaries fall within the corresponding zones. More details can be found in Appendix. To disentangle the effect of ranking, we mark a video as correct if at least one prediction is correct and measure the video-level hit rate. As shown in Fig. 4a, our model greatly outperforms the recent competitors. The sharp increase in low band widths clearly validates the superiority of our model in boundary alignment. + +Qualitative results. As shown in Fig. 4b, previous models fail to localize accurate moments with misleading center predictions. Especially, QD-DETR [41] accurately predicts the moment length but suffers from the misaligned center, leading to the limited IoU. In contrast, thanks to the novel moment modeling, our model predicts well-aligned boundaries without relying on center prediction. + +# 5 Conclusion + +In this paper, we identified the center misalignment issue of existing query-based models for sentence grounding. To address it, we presented boundary-oriented moment modeling where boundaries are directly predicted without relying on centers. Based on the modeling, we designed the boundary-aligned moment detection transformer characterized by dual-pathway decoding. Further, we proposed localization quality-based scoring of predictions. The efficacy of the proposed methods is validated by thorough examinations. We hope this work sheds light on the issue of center-based moment modeling in detection transformers. + +# Acknowledgements + +This project was supported by the National Research Foundation of Korea grant funded by the Korea government (MSIT) (No. 2022R1A2B5B02001467; RS-2024-00346364). + +# References + +1. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017) 1, 11 +2. Cao, M., Chen, L., Shou, M.Z., Zhang, C., Zou, Y.: On pursuit of designing multimodal transformer for video grounding. In: EMNLP. pp. 9810-9823 (2021) 4 +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229. Springer (2020) 2, 4, 6, 9 +4. Chen, J., Chen, X., Ma, L., Jie, Z., Chua, T.S.: Temporally grounding natural sentence in video. In: EMNLP. pp. 162-171 (2018) 4 +5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. PAMI 40(4), 834-848 (2017) 14 +6. Chen, L., Lu, C., Tang, S., Xiao, J., Zhang, D., Tan, C., Li, X.: Rethinking the bottom-up framework for query-based video localization. In: AAAI. vol. 34, pp. 10551-10558 (2020) 4 +7. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: ICCV. pp. 2988-2997 (2021) 4 +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 8 +9. Escorcia, V., Soldan, M., Sivic, J., Ghanem, B., Russell, B.: Temporal localization of moments in video collections with natural language. arXiv preprint arXiv:1907.12763 (2019) 1, 11 +0. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6202-6211 (2019) 10 +1. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: CVPR. pp. 5267-5275 (2017) 1, 4, 10 +2. Gao, P., Zheng, M., Wang, X., Dai, J., Li, H.: Fast convergence of detr with spatially modulated co-attention. In: ICCV. pp. 3621-3630 (2021) 4 +3. Ge, R., Gao, J., Chen, K., Nevatia, R.: Mac: Mining activity concepts for language-based temporal localization. In: WACV. pp. 245-253. IEEE (2019) 4 +4. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: ICASSP. pp. 776-780. IEEE (2017) 10 +5. Ghosh, S., Agarwal, A., Parekh, Z., Hauptmann, A.G.: Excl: Extractive clip localization using natural language descriptions. In: NAACL. pp. 1984-1990 (2019) 4 +6. Hao, J., Sun, H., Ren, P., Wang, J., Qi, Q., Liao, J.: Can shuffling video benefit temporal bias problem: A novel training framework for temporal grounding. In: ECCV. pp. 130-147. Springer (2022) 12 + +17. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: ICCV. pp. 13846-13856 (2023) 2, 4, 6, 7, 11 +18. Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Trans. on Audio, Speech, and Language Processing 28, 2880-2894 (2020) 10 +19. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) 9 +20. Lee, P., Kim, T., Shim, M., Wee, D., Byun, H.: Decomposed cross-modal distillation for rgb-based temporal action detection. In: CVPR. pp. 2373-2383 (2023) 1 +21. Lei, J., Berg, T.L., Bansal, M.: Detecting moments and highlights in videos via natural language queries. In: Neurips. vol. 34, pp. 11846-11858 (2021) 2, 3, 4, 5, 6, 9, 10, 11, 12, 13 +22. Lei, J., Yu, L., Berg, T.L., Bansal, M.: Tv: A large-scale dataset for video-subtitle moment retrieval. In: ECCV. pp. 447-463. Springer (2020) 11 +23. Li, F., Zeng, A., Liu, S., Zhang, H., Li, H., Zhang, L., Ni, L.M.: Lite detr: An interleaved multi-scale encoder for efficient detr. In: CVPR. pp. 18558-18567 (2023) 4 +24. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: CVPR. pp. 13619-13627 (2022) 4 +25. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In: ICCV. pp. 12032-12042 (2023) 4 +26. Li, P., Xie, C.W., Xie, H., Zhao, L., Zhang, L., Zheng, Y., Zhao, D., Zhang, Y.: Momentdiff: Generative video moment retrieval from random to real. In: Neurips (2023) 2, 4, 6, 10, 11, 12 +27. Lin, C., Xu, C., Luo, D., Wang, Y., Tai, Y., Wang, C., Li, J., Huang, F., Fu, Y.: Learning salient boundary feature for anchor-free temporal action localization. In: CVPR. pp. 3320-3329 (2021) 7 +28. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: ICCV. pp. 2794-2804 (2023) 4, 5, 6, 11, 12 +29. Lin, T., Zhao, X., Su, H., Wang, C., Yang, M.: Bsn: Boundary sensitive network for temporal action proposal generation. In: ECCV. pp. 3-19 (2018) 1, 7 +30. Lin, Y., Yuan, Y., Zhang, Z., Li, C., Zheng, N., Hu, H.: Detr does not need multiscale or locality design. In: ICCV. pp. 6545-6554 (2023) 4 +31. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: CVPR. pp. 11235-11244 (2021) 1 +32. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: CVPR. pp. 11235-11244 (2021) 4 +33. Liu, F., Wei, H., Zhao, W., Li, G., Peng, J., Li, Z.: Wb-detr: transformer-based detector without backbone. In: ICCV. pp. 2979-2987 (2021) 4 +34. Liu, M., Wang, X., Nie, L., Tian, Q., Chen, B., Chua, T.S.: Cross-modal moment localization in videos. In: ACM MM. pp. 843-851 (2018) 4 +35. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. In: ICLR (2022) 4, 7 +36. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., Zhang, L.: Detection transformer with stable matching. In: ICCV. pp. 6491-6500 (2023) 4 + +37. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: CVPR. pp. 3042-3051 (2022) 2, 4, 5, 10, 11 +38. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 11 +39. Lu, C., Chen, L., Tan, C., Li, X., Xiao, J.: Debug: A dense bottom-up grounding approach for natural language video localization. In: EMNLP-IJCNLP. pp. 5144-5153 (2019) 4 +40. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: ICCV. pp. 3651-3660 (2021) 4, 7 +41. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: CVPR. pp. 23023-23033 (2023) 2, 3, 4, 6, 7, 10, 11, 12, 13, 14 +42. Mun, J., Cho, M., Han, B.: Local-global video-text interactions for temporal grounding. In: CVPR. pp. 10810-10819 (2020) 4, 5 +43. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP. pp. 1532-1543 (2014) 12 +44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021) 10, 11 +45. Regneri, M., Rohrbach, M., Wetzel, D., Thater, S., Schiele, B., Pinkal, M.: Grounding action descriptions in videos. Trans. Assoc. Comput. Linguistics 1, 25-36 (2013) 10 +46. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: CVPR. pp. 658-666 (2019) 9 +47. Roh, B., Shin, J., Shin, W., Kim, S.: Sparse detr: Efficient end-to-end object detection with learnable sparsity. In: ICLR (2022) 4 +48. Rui, Y., Gupta, A., Acero, A.: Automatically extracting highlights for tv baseball programs. In: ACM MM. pp. 105-115 (2000) 1 +49. Shao, D., Xiong, Y., Zhao, Y., Huang, Q., Qiao, Y., Lin, D.: Find and focus: Retrieve and localize video events with natural language queries. In: ECCV. pp. 200-216 (2018) 1 +50. Sharghi, A., Laurel, J.S., Gong, B.: Query-focused video summarization: Dataset, evaluation, and a memory network based approach. In: CVPR. pp. 4788-4797 (2017) 1 +51. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: CVPR. pp. 18857-18866 (2023) 7 +52. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) 12 +53. Sun, Z., Cao, S., Yang, Y., Kitani, K.M.: Rethinking transformer-based set prediction for object detection. In: ICCV. pp. 3611-3620 (2021) 4 +54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Neurips. vol. 30 (2017) 6 +55. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: AAAI. vol. 36, pp. 2567-2575 (2022) 4 +56. Wang, Z., Wang, L., Wu, T., Li, T., Wu, G.: Negative sample matters: A renaissance of metric learning for temporal grounding. In: AAAI. vol. 36, pp. 2613-2623 (2022) 4, 12 + +57. Xiao, S., Chen, L., Zhang, S., Ji, W., Shao, J., Ye, L., Xiao, J.: Boundary proposal network for two-stage natural language video localization. In: AAAI. vol. 35, pp. 2986-2994 (2021) 4 +58. Xiong, B., Kalantidis, Y., Ghadiyaram, D., Grauman, K.: Less is more: Learning highlight detection from video duration. In: CVPR. pp. 1258-1267 (2019) 1 +59. Xu, H., He, K., Plummer, B.A., Sigal, L., Sclaroff, S., Saenko, K.: Multilevel language and vision integration for text-to-clip retrieval. In: AAAI. vol. 33, pp. 9062–9069 (2019) 4 +60. Xu, Y., Sun, Y., Li, Y., Shi, Y., Zhu, X., Du, S.: Mh-detr: Video moment and highlight detection with cross-modal transformer. In: ACM MM (2023) 4, 11 +61. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: Unloc: A unified framework for video localization tasks. In: ICCV. pp. 13623-13633 (2023) 4 +62. Ye, M., Ke, L., Li, S., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Cascade-detr: Delving into high-quality universal object detection. In: ICCV. pp. 6704-6714 (2023) 4 +63. Yuan, Y., Lan, X., Wang, X., Chen, L., Wang, Z., Zhu, W.: A closer look at temporal sentence grounding in videos: Dataset and metric. In: Proc. 2nd Int. Workshop on Human-Centric Multimedia Analysis. pp. 13-21 (2021) 12 +64. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In: Neurips. vol. 32 (2019) 1, 4 +65. Yuan, Y., Mei, T., Zhu, W.: To find where you talk: Temporal sentence localization in video with attention based location regression. In: AAAI. vol. 33, pp. 9159-9166 (2019) 4 +66. Zeng, R., Xu, H., Huang, W., Chen, P., Tan, M., Gan, C.: Dense regression network for video grounding. In: CVPR. pp. 10287-10296 (2020) 4 +67. Zhang, D., Dai, X., Wang, X., Wang, Y.F., Davis, L.S.: Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In: CVPR. pp. 1247-1257 (2019) 4 +68. Zhang, G., Luo, Z., Yu, Y., Cui, K., Lu, S.: Accelerating detr convergence via semantic-aligned matching. In: CVPR. pp. 949-958 (2022) 4 +69. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: ICLR (2023) 4 +70. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. In: ACL (2020) 4, 5, 12 +71. Zhang, K., Chao, W.L., Sha, F., Grauman, K.: Video summarization with long short-term memory. In: ECCV. pp. 766-782. Springer (2016) 1 +72. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. vol. 34, pp. 12870-12877 (2020) 1, 4, 5, 12 +73. Zhang, S., Su, J., Luo, J.: Exploiting temporal relationships in video moment localization with natural language. In: ACM MM. pp. 1230-1238 (2019) 4 +74. Zhang, Z., Lin, Z., Zhao, Z., Xiao, Z.: Cross-modal interaction networks for query-based moment retrieval in videos. In: ACM SIGIR. pp. 655-664 (2019) 4 +75. Zhao, Y., Xiong, Y., Wang, L., Wu, Z., Tang, X., Lin, D.: Temporal action detection with structured segment networks. In: ICCV. pp. 2914-2923 (2017) 1 +76. Zheng, D., Dong, W., Hu, H., Chen, X., Wang, Y.: Less is more: Focus attention for efficient detr. In: ICCV. pp. 6674-6683 (2023) 4 + +77. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: CVPR. pp. 9308-9316 (2019) 8 +78. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021) 4, 7, 8, 9 \ No newline at end of file diff --git a/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/images.zip b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fafb30c54007978cfce612bb93b0ef224f4c63c7 --- /dev/null +++ b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b33f8bf4ac4b0d542a90a3965254ddee41dc5acc7ca185591a5785e272d4d4a9 +size 510894 diff --git a/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/layout.json b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..18b2a2714fbdbf4b085e37406c0ddf94ae668d47 --- /dev/null +++ b/2024/BAM-DETR_ Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos/layout.json @@ -0,0 +1,11942 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 147, + 111, + 467, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 111, + 467, + 164 + ], + "spans": [ + { + "bbox": [ + 147, + 111, + 467, + 164 + ], + "type": "text", + "content": "BAM-DETR: Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 223, + 186, + 390, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 186, + 390, + 198 + ], + "spans": [ + { + "bbox": [ + 223, + 186, + 390, + 198 + ], + "type": "text", + "content": "Pilhyeon Lee" + }, + { + "bbox": [ + 223, + 186, + 390, + 198 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 223, + 186, + 390, + 198 + ], + "type": "text", + "content": " and Hyeran Byun" + }, + { + "bbox": [ + 223, + 186, + 390, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 207, + 419, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 207, + 419, + 220 + ], + "spans": [ + { + "bbox": [ + 194, + 207, + 419, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 207, + 419, + 220 + ], + "type": "text", + "content": " Department of Artificial Intelligence, Inha University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 220, + 417, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 220, + 417, + 230 + ], + "spans": [ + { + "bbox": [ + 194, + 220, + 417, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 194, + 220, + 417, + 230 + ], + "type": "text", + "content": " Department of Computer Science, Yonsei University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 255, + 455, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 455, + 453 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 455, + 453 + ], + "type": "text", + "content": "Abstract. Temporal sentence grounding aims to localize moments relevant to a language description. Recently, DETR-like approaches achieved notable progress by predicting the center and length of a target moment. However, they suffer from the issue of center misalignment raised by the inherent ambiguity of moment centers, leading to inaccurate predictions. To remedy this problem, we propose a novel boundary-oriented moment formulation. In our paradigm, the model no longer needs to find the precise center but instead suffices to predict any anchor point within the interval, from which the boundaries are directly estimated. Based on this idea, we design a boundary-aligned moment detection transformer, equipped with a dual-pathway decoding process. Specifically, it refines the anchor and boundaries within parallel pathways using global and boundary-focused attention, respectively. This separate design allows the model to focus on desirable regions, enabling precise refinement of moment predictions. Further, we propose a quality-based ranking method, ensuring that proposals with high localization qualities are prioritized over incomplete ones. Experiments on three benchmarks validate the effectiveness of the proposed methods. The code is available here." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 463, + 433, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 463, + 433, + 474 + ], + "spans": [ + { + "bbox": [ + 160, + 463, + 433, + 474 + ], + "type": "text", + "content": "Keywords: Temporal sentence grounding " + }, + { + "bbox": [ + 160, + 463, + 433, + 474 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 463, + 433, + 474 + ], + "type": "text", + "content": " Detection transformer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 492, + 230, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 230, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 230, + 504 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 515, + 482, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 623 + ], + "type": "text", + "content": "Recent years have witnessed a notable surge in the popularity of short-form video content on social media platforms like TikTok, YouTube Shorts, and Instagram Reels. As such, users prefer to selectively engage with short moments of interest rather than passively watch an entire long video. This trend highlights the importance of localizing desired moments. As a result, moment localization tasks have emerged as pivotal research topics in video understanding, including temporal action detection [20, 29, 75], video summarization [50, 71], and highlight detection [48, 58]. Within this context, we tackle temporal sentence grounding [1, 72], aiming to retrieve moments corresponding to free-form language descriptions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 624, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 481, + 647 + ], + "type": "text", + "content": "To address temporal sentence grounding, numerous efforts have been undertaken in the last decade [9, 11, 31, 49, 64, 72]. Especially, taking inspiration" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 404, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 404, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 404, + 666 + ], + "type": "text", + "content": "* Correspondence to: Pilhyeon Lee " + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 202, + 114, + 410, + 194 + ], + "blocks": [ + { + "bbox": [ + 202, + 114, + 410, + 194 + ], + "lines": [ + { + "bbox": [ + 202, + 114, + 410, + 194 + ], + "spans": [ + { + "bbox": [ + 202, + 114, + 410, + 194 + ], + "type": "image", + "image_path": "cd3dbc8d5ee3eec6576c86a0597040e9692a12aa97821ba3cf81a4369a1ed038.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 195, + 363, + 204 + ], + "lines": [ + { + "bbox": [ + 256, + 195, + 363, + 204 + ], + "spans": [ + { + "bbox": [ + 256, + 195, + 363, + 204 + ], + "type": "text", + "content": "(a) Conventional moment modeling" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 205, + 410, + 221 + ], + "blocks": [ + { + "bbox": [ + 209, + 205, + 410, + 221 + ], + "lines": [ + { + "bbox": [ + 209, + 205, + 410, + 221 + ], + "spans": [ + { + "bbox": [ + 209, + 205, + 410, + 221 + ], + "type": "image", + "image_path": "98965eb2948b6a3bd61e7f7cd5ca7f366452887af5193e63ad134e5ce0069035.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 222, + 382, + 231 + ], + "lines": [ + { + "bbox": [ + 238, + 222, + 382, + 231 + ], + "spans": [ + { + "bbox": [ + 238, + 222, + 382, + 231 + ], + "type": "text", + "content": "(b) Boundary-oriented moment modeling (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 239, + 479, + 274 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 479, + 274 + ], + "type": "text", + "content": "Fig. 1: Comparison of moment modeling approaches under the scenario of an ambiguous center from QVHighlights. (a) The conventional method formulates a moment with a tuple of " + }, + { + "bbox": [ + 131, + 239, + 479, + 274 + ], + "type": "inline_equation", + "content": "(c, l)" + }, + { + "bbox": [ + 131, + 239, + 479, + 274 + ], + "type": "text", + "content": ". (b) In contrast, we propose to model it with a triplet of " + }, + { + "bbox": [ + 131, + 239, + 479, + 274 + ], + "type": "inline_equation", + "content": "(p, d_s, d_e)" + }, + { + "bbox": [ + 131, + 239, + 479, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 480, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 480, + 330 + ], + "type": "text", + "content": "from DETR [3], query-based approaches have become a promising research direction owing to the architectural simplicity [17,21,26,37,41]. By decoding temporal spans (i.e., moments) from a handful set of learnable queries, they achieve promising grounding performance while maintaining high inference speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 331, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 485 + ], + "type": "text", + "content": "Existing query-based models typically predict a moment using its center and width, i.e., " + }, + { + "bbox": [ + 130, + 331, + 480, + 485 + ], + "type": "inline_equation", + "content": "(c, l)" + }, + { + "bbox": [ + 130, + 331, + 480, + 485 + ], + "type": "text", + "content": ", under the assumption that the boundaries are equidistant from the center. However, such a formulation can be problematic as the center of a moment might be ambiguous. An illustrative example is presented in Fig. 1 (top), where the frame at the center of the ground-truth moment is less relevant to the given sentence, as demonstrated by the low saliency label. Indeed, a moment center does not necessarily serve as the best representative of the sentence. This ambiguity can challenge the model's ability to precisely locate the centers, and the misaligned centers lead to low-quality predictions (Fig. 1a). To probe the impact of center misalignment, we conducted a diagnostic experiment in Table 1. The results reveal that (i) existing methods struggle with detecting accurate center points and (ii) they suffer from significant performance drops when the predicted centers deviate from the ground-truth centers (i.e., large center errors)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "type": "text", + "content": "To address this challenge, we present a novel boundary-oriented formulation for moments, as illustrated in Fig. 1b, where each moment is represented by a triplet consisting of an anchor point and its distances to the boundaries, i.e., " + }, + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "type": "inline_equation", + "content": "(p,d_s,d_e)" + }, + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "type": "text", + "content": ". This asymmetric formulation liberates the model from the stringent requirement of predicting the center. Instead, it is sufficient for the model to predict any salient anchor within the target moment, and the distances from the anchor to the onset and offset are predicted subsequently. By directly locating the boundaries based on the anchor point, the model can achieve improved boundary alignment, even when the anchor point does not coincide with the actual center." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Building upon the proposed moment modeling, we introduce a new framework equipped with a dedicated decoder design, dubbed Boundary-Aligned Moment Detection Transformer (BAM-DETR). The design of its decoding layers originates from the intuition that the refining processes for an anchor and boundaries should be distinct from each other. That is to say, a model needs to scan over the whole video to find a potential anchor point that enables estimating the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "type": "text", + "content": "P. Lee and H. Byun" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 167, + 462, + 237 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: Impact of misaligned centers on QVHighlights. The top-1 predictions are grouped based on their center errors normalized by ground-truth lengths. For each group, we present the mean IoU (\\%) and the proportion (in parentheses). Only the predictions whose centers fall within the ground-truth moments are considered here." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 167, + 462, + 237 + ], + "lines": [ + { + "bbox": [ + 149, + 167, + 462, + 237 + ], + "spans": [ + { + "bbox": [ + 149, + 167, + 462, + 237 + ], + "type": "table", + "html": "
Method[0, 0.1)[0.1, 0.2)[0.2, 0.3)[0.3, 0.4)[0.4, 0.5)All
Moment-DETR [21]83.2065.0054.8243.7734.9867.84
(46 %)(24 %)(14 %)(9 %)(7 %)(100 %)
QD-DETR [41]87.7967.9555.4044.9336.4574.09
(56 %)(18 %)(11 %)(9 %)(5 %)(100 %)
BAM-DETR†77.6278.2077.5278.9671.0877.21
(Ours)(24 %)(21 %)(21 %)(22 %)(12 %)(100 %)
", + "image_path": "9d5b6a342cc2190314f02b3d5c04c5aa7b6800056568a888d752e41da9c99cc8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 148, + 237, + 290, + 247 + ], + "lines": [ + { + "bbox": [ + 148, + 237, + 290, + 247 + ], + "spans": [ + { + "bbox": [ + 148, + 237, + 290, + 247 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 148, + 237, + 290, + 247 + ], + "type": "text", + "content": " anchor points are utilized for grouping." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 254, + 481, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 481, + 372 + ], + "type": "text", + "content": "rough location of the target moment. On the other hand, it is required to focus on fine-grained details in the vicinity to refine the boundaries to be aligned with those of the target moments. From this motivation, in contrast to existing methods, our model adopts a dual-pathway decoding pipeline to predict an anchor point and boundaries in a parallel way. To be specific, we leverage two different types of queries respectively for anchor and boundary refinement. The former aggregates global information with standard attention, while the latter concentrates on the sparse local neighborhood of the boundaries using the proposed boundary-focused attention. These distinct designs of two pathways allow for effective moment localization with minimal computational overhead increase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 481, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 481, + 480 + ], + "type": "text", + "content": "In addition, we identify the problem of the conventional scoring method, where binary classification (or matching) scores are used for proposal ranking. This leads to suboptimal results for the grounding task since a fractional moment may have high matching scores with the given sentence. To handle this issue, we propose to rank proposals based on their localization qualities. Accordingly, we modify the typical matching function and training objectives of the query-based model to be localization-oriented by discarding the role of classification scores. In this way, our model can prioritize the moment proposals exhibiting high overlap with ground truths at inference, leading to improved grounding performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 481, + 481, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 481, + 588 + ], + "type": "text", + "content": "The advantage of our BAM-DETR is showcased in Table 1. It can be observed that the anchor points predicted by our model are evenly distributed within the ground-truth intervals. Importantly, our model consistently produces precise moments with high IoUs (" + }, + { + "bbox": [ + 130, + 481, + 481, + 588 + ], + "type": "inline_equation", + "content": ">0.7" + }, + { + "bbox": [ + 130, + 481, + 481, + 588 + ], + "type": "text", + "content": ") across different groups, indicating that it does not depend on accurate center prediction. On average, our model shows superior grounding performance over the existing methods with the help of our moment formulation and quality-based scoring. In a later section, we validate the efficacy of the proposed methods through extensive experiments. Notably, our model outperforms previous methods by large margins on three public benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 603, + 243, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 603, + 243, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 603, + 243, + 616 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 624, + 365, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 365, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 365, + 637 + ], + "type": "text", + "content": "2.1 Temporal Sentence Grounding in Videos" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Temporal sentence grounding requires seeking temporal spans semantically relevant to the given sentence in a video. Proposal-based approaches adopt the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 90, + 447, + 100 + ], + "type": "text", + "content": "BAM-DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "two-stage pipeline, i.e., proposal generation and ranking. They generate moment proposals by relying on sliding windows [11,13,34,73] or utilizing pre-defined anchors [4,59,64,67,74]. Several works process all possible candidates at once with 2D maps [25,32,56,57,72]. Meanwhile, proposal-free methods are developed for efficient grounding by directly predicting the moments [42,65] or estimating the probabilities of each frame being starting and ending positions [15,70]. Some approaches perform dense regression by predicting the boundaries from individual frames [6,39,66]. Recently, query-based models streamline the complicated sentence grounding pipeline by removing handcrafted techniques [2,17,26,37,60]. There are also attempts to unify temporal sentence grounding with other video understanding tasks into a single framework [28,61]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "type": "text", + "content": "Our method belongs to the query-based group [21, 41], inheriting the benefit of architectural simplicity. In contrast to others, we employ a boundary-oriented formulation of moments to relieve the heavy reliance on center predictions, leading to better boundary alignment. Our method also relates to dense regression methods [6, 28, 66] that predict boundaries from each frame as an anchor. Comparatively, our model leverages dynamic anchors that are gradually adjusted through decoding, enabling precise grounding using a small set of predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 346, + 280, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 280, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 280, + 357 + ], + "type": "text", + "content": "2.2 Detection Transformers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 460 + ], + "type": "text", + "content": "Query-based temporal sentence grounding models by design are closely related to the family of detection transformers (DETR) [3]. Since the advent of DETR, a number of variants have been introduced to improve it from various perspectives [30, 33, 36, 53]. Some works focus on reducing the excessive computational costs of vanilla transformers in order to leverage multi-scale features [7, 23, 47, 76, 78]. On the other hand, several methods attempt to speed up the model convergence by manipulating the attention operations [12, 35, 62, 68] or incorporating the denoising process during the model training [24, 69]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "content": "The most relevant works to ours are those which propose explicit anchor modeling for object queries using center points [40, 55] or boxes (center, width, and height) [35, 78]. In comparison to an object, a moment in temporal sentence grounding has its own challenges such as center ambiguity and indistinct boundaries. To accommodate the discrepancies, we propose a novel boundary-oriented modeling of moments to replace the conventional center-based 1D box modeling. The advantages of our approach are clearly verified in the experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 559, + 202, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 202, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 202, + 571 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Given an untrimmed video and a sentence, the goal of temporal sentence grounding is to localize relevant moments " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\varphi_{n} = (t_{s_{n}}, t_{e_{n}})\\}_{n=1}^{N}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " denotes the number of ground truths in the video and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varphi_{n}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " indicates the temporal interval of the " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "-th moment. Note that a video may have multiple moments that match the sentence, i.e., " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "N \\geq 1" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". During the test time, the model is expected to produce a total of " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " predictions, " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{(\\hat{\\varphi}_{m}, q_{m})\\}_{m=1}^{M}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\varphi}_{m} = (\\hat{t}_{s_{m}}, \\hat{t}_{e_{m}})" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "-th prediction, while " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "q_{m}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " is its score for ranking." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "text", + "content": "P. Lee and H. Byun" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 321, + 202 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 321, + 202 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 321, + 202 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 321, + 202 + ], + "type": "image", + "image_path": "2903c8ae1d069fab841fdb58bbd2833f062e86bbe46ab08b5242bbd0b0b33af6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 206, + 281, + 214 + ], + "lines": [ + { + "bbox": [ + 172, + 206, + 281, + 214 + ], + "spans": [ + { + "bbox": [ + 172, + 206, + 281, + 214 + ], + "type": "text", + "content": "(a) Overall framework of BAM-DETR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 321, + 114, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 321, + 114, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 321, + 114, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 321, + 114, + 479, + 205 + ], + "type": "image", + "image_path": "7f141d3e762f7ff75e61dd3db547f118e69356d1ff696dc37eb7aaa8b94a2349.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 206, + 449, + 215 + ], + "lines": [ + { + "bbox": [ + 356, + 206, + 449, + 215 + ], + "spans": [ + { + "bbox": [ + 356, + 206, + 449, + 215 + ], + "type": "text", + "content": "(b) Dual-pathway decoding layer" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 223, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 223, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 480, + 257 + ], + "type": "text", + "content": "Fig. 2: (a) Overview of the proposed BAM-DETR. (b) Details of the proposed dual-pathway decoding layer. It consists of two parallel pathways respectively for anchor and boundary updates, which refine previous moment predictions in a sequential manner." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "text", + "content": "Motivation. We tackle two main problems in the current prediction process. On one hand, existing query-based approaches adopt the symmetric design of " + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "inline_equation", + "content": "(c,l)" + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "text", + "content": " to predict a moment, i.e., " + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "inline_equation", + "content": "\\hat{\\varphi} = (c - 0.5l, c + 0.5l)" + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "text", + "content": ". As we discuss in Sec. 1, this strategy suffers from the issue of over-reliance on the center prediction, leading to unstable performance. To address this, we propose a boundary-oriented modeling with a triplet of " + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "inline_equation", + "content": "(p,d_s,d_e)" + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "text", + "content": ", where a moment is represented by " + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "inline_equation", + "content": "\\hat{\\varphi} = (p - d_s,p + d_e)" + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "text", + "content": ". This asymmetric design enables direct alignment of boundaries without relying on precise center prediction. On the other hand, previous models use classification scores as " + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "inline_equation", + "content": "q_{m}" + }, + { + "bbox": [ + 130, + 266, + 480, + 410 + ], + "type": "text", + "content": ". This scoring is prone to sub-optimal solutions since only a fraction of the moment may well match the sentence, leading incomplete predictions to be highly ranked. To handle this, we propose quality-based scoring to sort the proposals based on their localization qualities rather than the degree of matching." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 427, + 208, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 427, + 208, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 208, + 438 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 447, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 480, + 495 + ], + "type": "text", + "content": "As shown in Fig. 2a, our BAM-DETR follows the encoder-decoder pipeline. Briefly, it first extracts multimodal video features using the encoders. Taking them as memory, the dual-pathway decoder predicts temporal spans and their quality scores by progressively refining learnable initial spans." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 512, + 256, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 256, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 256, + 522 + ], + "type": "text", + "content": "3.2 Feature Extraction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "text", + "content": "Following the convention [42, 70, 72], we employ pre-trained encoders for token-level unimodal feature extraction. It is worth noting that all unimodal encoders are kept frozen during training to avoid memory exploding. In this stage, we obtain " + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "text", + "content": "-dim video features " + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "inline_equation", + "content": "E_v \\in \\mathbb{R}^{N_v \\times D_v}" + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "inline_equation", + "content": "D_t" + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "text", + "content": "-dim text features " + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "inline_equation", + "content": "E_t \\in \\mathbb{R}^{N_t \\times D_t}" + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 130, + 532, + 480, + 593 + ], + "type": "text", + "content": " are the numbers of clips and words, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 609, + 265, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 265, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 265, + 620 + ], + "type": "text", + "content": "3.3 Multimodal Encoder" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "content": "We feed the unimodal features of the video and text to the multimodal encoder, so as to fuse them into text-aware video representations for temporal sentence grounding. While various multimodal encoders are explored [21,28,37], we" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAM-DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "adopt the text-to-video encoder design [41] consisting of cross- and self-attention blocks [54]. Before getting into the encoder, unimodal features are projected into a shared space to facilitate cross-modal interaction, i.e., " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = f_v(E_v) \\in \\mathbb{R}^{N_v \\times D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = f_t(E_t) \\in \\mathbb{R}^{N_t \\times D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " is the embedding dimension. Afterward, multi-head cross-attention blocks are used to inject textual information into the clip-level video representations. In specific, we project " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " (query) while " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " (key) and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " (value). Then a cross-attention block can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 203, + 481, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 203, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 237, + 203, + 481, + 229 + ], + "type": "interline_equation", + "content": "\\mathcal {V} ^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\mathcal {V}} \\mathbf {K} _ {\\mathcal {T}} ^ {\\top}}{\\sqrt {D}}\\right) \\mathbf {V} _ {\\mathcal {T}} + \\mathcal {V}, \\tag {1}", + "image_path": "85444161fa6a2d3e207c44c5daae452b49c28eeb27131239aa1831a50689f03e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 231, + 324, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 231, + 324, + 243 + ], + "spans": [ + { + "bbox": [ + 235, + 231, + 324, + 243 + ], + "type": "interline_equation", + "content": "\\mathcal {V} ^ {\\prime \\prime} = \\operatorname {F F N} (\\mathcal {V} ^ {\\prime}) + \\mathcal {V} ^ {\\prime},", + "image_path": "def807d7a8e2224da7ee55e301a0e316443d9b93a5d934a751975177773ca7d0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 246, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 479, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 246, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}(\\cdot)" + }, + { + "bbox": [ + 130, + 246, + 479, + 293 + ], + "type": "text", + "content": " is a feed-forward network. Although we here present the singlehead attention block, it can readily generalize to a multi-headed version [54]. We denote the resulting multimodal representations obtained after " + }, + { + "bbox": [ + 130, + 246, + 479, + 293 + ], + "type": "inline_equation", + "content": "L_{E}" + }, + { + "bbox": [ + 130, + 246, + 479, + 293 + ], + "type": "text", + "content": " multi-head cross-attention blocks by " + }, + { + "bbox": [ + 130, + 246, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{V}}\\in \\mathbb{R}^{N_v\\times D}" + }, + { + "bbox": [ + 130, + 246, + 479, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "text", + "content": "Subsequently, self-attention blocks are leveraged to enhance the representations by allowing the inter-clip interaction. Here we project " + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\tilde{\\mathcal{V}}}" + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\tilde{\\mathcal{V}}}" + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\tilde{\\mathcal{V}}}" + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "text", + "content": ". We note that the query and the key are supplemented with fixed sinusoidal positional encoding [3, 54] for temporal awareness. Then the self-attention block is defined in a similar way to Eq. (1) but with different inputs. The enhanced clip-level representations after " + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "inline_equation", + "content": "L_{E}" + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "text", + "content": " multi-head self-attention blocks are denoted by " + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}} \\in \\mathbb{R}^{N_v \\times D}" + }, + { + "bbox": [ + 130, + 294, + 480, + 377 + ], + "type": "text", + "content": ", which will serve as the memory for the decoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 378, + 480, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 480, + 438 + ], + "type": "text", + "content": "It is widely known that providing saliency guidance to the memory features helps the model to better understand the semantic relationship between the video and text [21, 28]. As in previous works [17, 26], we impose saliency score constraints on the memory features. Specifically, we leverage a saliency predictor " + }, + { + "bbox": [ + 130, + 378, + 480, + 438 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 130, + 378, + 480, + 438 + ], + "type": "text", + "content": " and train the model with the following margin-based training objective." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 447, + 480, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 447, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 216, + 447, + 480, + 461 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a r g i n}} = \\max (0, \\alpha + S (\\hat {v} ^ {\\text {l o w}}) - S (\\hat {v} ^ {\\text {h i g h}})), \\tag {2}", + "image_path": "d5d4c3da27d4ef137545e08a9a1eded4393a640c39ec3c8d0721cdd5c124e222.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "text", + "content": " is a margin and " + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "inline_equation", + "content": "(\\hat{v}^{\\mathrm{low}},\\hat{v}^{\\mathrm{high}})" + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "text", + "content": " is the sampled feature pair satisfying that the saliency label of " + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\hat{v}^{\\mathrm{low}}" + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "text", + "content": " is lower than that of " + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\hat{v}^{\\mathrm{high}}" + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "text", + "content": ". In case of the absence of saliency labels, we collect clips within and outside the ground-truth moment intervals to build a pair. In addition, we employ the rank-aware contrastive loss " + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cont}}" + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "text", + "content": " and the negative relation loss " + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{neg}}" + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "text", + "content": ", following Moon et al. [41]. Due to space limits, we refer the readers to Appendix for the loss formulations. In summary, the overall saliency loss is defined as " + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sal}} = \\mathcal{L}_{\\mathrm{margin}} + \\mathcal{L}_{\\mathrm{cont}} + \\mathcal{L}_{\\mathrm{neg}}" + }, + { + "bbox": [ + 130, + 469, + 482, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 573, + 276, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 276, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 276, + 585 + ], + "type": "text", + "content": "3.4 Dual-pathway Decoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "With the multimodal representations " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " as memory features, we aim to localize temporal spans corresponding to the sentence. We adopt the new boundary-oriented formulation for moment prediction, where each prediction is represented by a triplet of " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "(p,d_s,d_e)" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " is the anchor point, while " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "d_{s}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "d_{e}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " are the distances from the anchor to the starting and ending points, respectively. To make full use of the proposed formulation, we design a dual-pathway decoding" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "text", + "content": "P. Lee and H. Byun" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "text", + "content": "layer with two parallel pathways (Fig. 2b). Formally, the inputs of the " + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "text", + "content": "-th layer are anchor queries " + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_p^l \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "text", + "content": ", boundary queries " + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_s^l" + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_e^l \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "text", + "content": ", and previous moment predictions " + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^l = [\\mathbf{p}^l; \\mathbf{d}_s^l; \\mathbf{d}_e^l] \\in \\mathbb{R}^{M \\times 3}" + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "text", + "content": " is the number of queries (predictions). Note that the initial queries and spans, i.e., " + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}_p^0, \\mathbf{C}_s^0, \\mathbf{C}_e^0, \\mathbf{A}^0\\}" + }, + { + "bbox": [ + 131, + 116, + 482, + 178 + ], + "type": "text", + "content": ", are learnable parameters. We elaborate on the two pathways in the following." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "content": "Anchor updating pathway. Given the predictions and the anchor queries from the preceding layer, the goal of this pathway is to adjust the anchor position so that the boundaries can be predicted based on it. Intuitively, in order to obtain a valuable anchor point without redundancy, an anchor query should communicate with other queries as well as the memory features. To this goal, the anchor updating pathway consists of a self-attention layer, a cross-attention layer, and a feed-forward network. In the self-attention layer, the anchor queries " + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_p^l" + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "content": " are first projected into " + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{C}_p^l}" + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathbf{C}_p^l}" + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{C}_p^l}" + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "content": ". Since an anchor query itself lacks positional information, we build positional encoding. Following the previous works [17,41], we extend the current spans " + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^l" + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "content": " to the positional information of the queries, i.e., " + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{A}^l} = \\mathrm{MLP}(\\mathrm{PE}(\\mathbf{A}^l)) \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{PE}(\\cdot)" + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "content": " denotes the point-wise mapping from a position to the corresponding sinusoidal encoding while " + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}(\\cdot)" + }, + { + "bbox": [ + 131, + 198, + 482, + 353 + ], + "type": "text", + "content": " is multi-layer perceptron. The self-attention for anchor queries is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 356, + 481, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 356, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 180, + 356, + 481, + 387 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {C}} _ {p} ^ {l} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {Q} _ {\\mathbf {C} _ {p} ^ {l}} + \\mathbf {P} _ {\\mathbf {A} ^ {l}}\\right) \\left(\\mathbf {K} _ {\\mathbf {C} _ {p} ^ {l}} + \\mathbf {P} _ {\\mathbf {A} ^ {l}}\\right) ^ {\\top}}{\\sqrt {D}}\\right) \\mathbf {V} _ {\\mathbf {C} _ {p} ^ {l}} + \\mathbf {C} _ {p} ^ {l}. \\tag {3}", + "image_path": "577cbc2103cd26ac89b12afab916c55f2bf23e4d935417e777d85c909806c5ee.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "text", + "content": "After inter-query interaction, we employ a global cross-attention layer to aggregate multi-modal features from the memory. The anchor queries " + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{C}}_p^l" + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "text", + "content": " is projected into " + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\tilde{\\mathbf{C}}_p^l}" + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "text", + "content": " while the memory " + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\hat{\\nu}" + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "text", + "content": " is projected to " + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\hat{\\nu}}" + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\hat{\\nu}}" + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "text", + "content": ". To make the query location-aware, we leverage the sinusoidal encoding of current anchor positions, i.e., " + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{P_{p^l}} = \\mathrm{PE}(\\mathbf{p}^l)\\in \\mathbb{R}^{M\\times D}" + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "text", + "content": ". Similarly, the memory leverages the positional encoding, i.e., " + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\hat{\\nu}} = \\mathrm{PE}(\\hat{\\nu})\\in \\mathbb{R}^{N_v\\times D}" + }, + { + "bbox": [ + 131, + 390, + 482, + 492 + ], + "type": "text", + "content": ". We use concatenation instead of summation to separate the roles of features and positional encoding [35, 40]. The cross-attention between anchor queries and the memory can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 189, + 495, + 481, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 495, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 189, + 495, + 481, + 528 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {C}} _ {p} ^ {l} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {Q} _ {\\tilde {\\mathbf {C}} _ {p} ^ {l}} \\parallel \\mathbf {P} _ {\\mathbf {p} ^ {l}}\\right)\\left(\\mathbf {K} _ {\\hat {\\mathcal {V}}} \\parallel \\mathbf {P} _ {\\hat {\\mathcal {V}}}\\right)\\right) ^ {\\top}}{\\sqrt {2 D}}\\left. \\right) \\mathbf {V} _ {\\hat {\\mathcal {V}}} + \\tilde {\\mathbf {C}} _ {p} ^ {l}. \\tag {4}", + "image_path": "2181d675ebb4f0f31389e3be79e2209a28d402b41d6653697739d2fda245590e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "text", + "content": "After all, the anchor queries are updated with a feed-forward network, i.e., " + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_p^{(l + 1)} = \\mathrm{FFN}(\\hat{\\mathbf{C}}_p^l) + \\hat{\\mathbf{C}}_p^l" + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "text", + "content": ". Lastly, we adjust the anchor positions using sigmoid-based refinement [78]: " + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}^l = [\\mathbf{p}^{(l + 1)};\\mathbf{d}_s^l;\\mathbf{d}_e^l]" + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{(l + 1)} = \\sigma (\\sigma^{-1}(\\mathbf{p}^l) + \\varDelta \\mathbf{p}^l)" + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\varDelta\\mathbf{p}^l=\\mathrm{MLP}(\\mathbf{C}_p^{(l+1)})\\in\\mathbb{R}^M" + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "text", + "content": " and the sigmoid function " + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 131, + 531, + 482, + 586 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": "Boundary updating pathway. After the anchor update, we refine the boundaries of the predictions. It is widely perceived that a model needs to focus on fine-grained features in the neighborhood rather than far ones to adjust temporal boundaries [27, 29, 51]. Inspired by this, we devise a boundary-focused attention layer (Fig. 3). For brevity, we explain the process of starting boundary update." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 90, + 447, + 100 + ], + "type": "text", + "content": "BAM-DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 198, + 114, + 416, + 257 + ], + "blocks": [ + { + "bbox": [ + 198, + 114, + 416, + 257 + ], + "lines": [ + { + "bbox": [ + 198, + 114, + 416, + 257 + ], + "spans": [ + { + "bbox": [ + 198, + 114, + 416, + 257 + ], + "type": "image", + "image_path": "037939b1153ca7e602e2d4e2373f4bc11f4f85ee6c8db0801185ac7a7dace92d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 264, + 430, + 275 + ], + "lines": [ + { + "bbox": [ + 181, + 264, + 430, + 275 + ], + "spans": [ + { + "bbox": [ + 181, + 264, + 430, + 275 + ], + "type": "text", + "content": "Fig. 3: Boundary-focused attention layer for starting queries." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "text", + "content": "The plain memory " + }, + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "text", + "content": " lacks the inductive bias due to the property of attentional layers [8]. Thus we first build locality-enhanced memory features for effective boundary refinement. To this goal, we obtain boundary-sensitive features with several 1D convolutional layers, i.e., " + }, + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_{\\mathrm{s}} = f_{\\mathrm{s}}(\\hat{\\mathcal{V}})" + }, + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "text", + "content": ". Then we encourage them to highly activate around the starting position of target moments. In detail, we impose regularization on the clip-wise activation scores obtained by channel-mean, i.e., " + }, + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\hat{g}^s = \\mathrm{mean}(\\sigma (\\hat{\\mathcal{V}}_s))\\in \\mathbb{R}^{N_v}" + }, + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "text", + "content": ". The regularization loss is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 193, + 377, + 481, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 377, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 193, + 377, + 481, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g u l}} ^ {s} = - \\frac {1}{N _ {v}} \\sum_ {i = 1} ^ {N _ {v}} \\left(g _ {i} ^ {s} \\log \\left(\\hat {g} _ {i} ^ {s}\\right) + \\left(1 - g _ {i} ^ {s}\\right) \\log \\left(1 - \\hat {g} _ {i} ^ {s}\\right)\\right), \\tag {5}", + "image_path": "570f76d11626c87e222c54b0d469bc54c259d4dad7b7b49fda04a40b121e4b72.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "inline_equation", + "content": "g_{i}^{s}" + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "content": " is the binary label obtained as " + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "inline_equation", + "content": "g_{i}^{s} = \\mathbb{1}[i\\in \\mathcal{B}^{s}]" + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^s" + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "content": " is the neighbor clip set around starting points with a radius of " + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "inline_equation", + "content": "r^s" + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "content": " (set to 1/10 of the moment length). We merge boundary-sensitive features with the plain ones to form locality-enhanced memory, i.e., " + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_s' = [\\hat{\\mathcal{V}}\\parallel \\hat{\\mathcal{V}}_s]\\in \\mathbb{R}^{N_v\\times 2D}" + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "content": ". We can obtain " + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{regul}}^e" + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_e^\\prime" + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "content": " in the same way. The total regularization term is " + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{regul}} = \\mathcal{L}_{\\mathrm{regul}}^s +\\mathcal{L}_{\\mathrm{regul}}^e" + }, + { + "bbox": [ + 130, + 411, + 482, + 476 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "content": "Given the locality-enhanced memory, the boundary queries need to capture fine-grained details around the boundaries for refinement. To efficiently aggregate useful features near the boundaries, we employ deformable attention [77, 78]. Regarding the starting boundary, i.e., " + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{(l + 1)} - \\mathbf{d}_s^l" + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "content": ", as the origin, we predict offsets and weights to select " + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "content": " neighbors, i.e., " + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^s = \\phi_o(\\mathbf{C}_s^l)\\in \\mathbb{R}^{M\\times K}" + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^s = \\mathrm{softmax}(\\phi_w(\\mathbf{C}_s^l))\\in \\mathbb{R}^{M\\times K}" + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\phi_{*}" + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "content": " are fully-connected layers. Features from the sampled neighbors are then aggregated into the starting queries as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 211, + 565, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 211, + 565, + 481, + 597 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {C}} _ {s} ^ {l} = \\sum_ {k = 1} ^ {K} \\left[ \\mathbf {w} _ {k} ^ {s} \\cdot \\hat {\\mathcal {V}} _ {s} ^ {\\prime} \\left[ \\mathbf {p} ^ {(l + 1)} - \\mathbf {d} _ {s} ^ {l} + \\mathbf {o} _ {k} ^ {s} \\right] \\right] + \\mathbf {C} _ {s} ^ {l}, \\tag {6}", + "image_path": "d872ebba7082a629757d42d25c7879b34528229c0a1196a9dcdcb732982761cb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 601, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 666 + ], + "type": "text", + "content": "where we denote the sampling process from memory by " + }, + { + "bbox": [ + 130, + 601, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_s^{\\prime}[\\cdot]" + }, + { + "bbox": [ + 130, + 601, + 482, + 666 + ], + "type": "text", + "content": ". Lastly, we adopt a feed-forward network to obtain the updated starting queries, i.e., " + }, + { + "bbox": [ + 130, + 601, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_s^{(l + 1)} = \\mathrm{FFN}(\\hat{\\mathbf{C}}_s^l)" + }, + { + "bbox": [ + 130, + 601, + 482, + 666 + ], + "type": "text", + "content": ". We also obtain " + }, + { + "bbox": [ + 130, + 601, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_e^{(l + 1)}" + }, + { + "bbox": [ + 130, + 601, + 482, + 666 + ], + "type": "text", + "content": " in the same way. With the updated queries, we refine the boundaries to be better aligned with those of ground truths using sigmoid-based refinement as similar in the anchor update, leading to the refined" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "text", + "content": "P. Lee and H. Byun" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 165 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 165 + ], + "type": "text", + "content": "predictions " + }, + { + "bbox": [ + 131, + 114, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{(l + 1)} = [\\mathbf{p}^{(l + 1)};\\mathbf{d}_s^{(l + 1)};\\mathbf{d}_e^{(l + 1)}]" + }, + { + "bbox": [ + 131, + 114, + 479, + 165 + ], + "type": "text", + "content": ". Note that we utilize the deformable attention for the purpose of local feature aggregation, which sharply differs from the original purpose of efficient multi-scale global operation [78]. We provide comparison experiments regarding their roles in Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "text", + "content": "Moment prediction. We repeatedly update the predictions through a total of " + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "inline_equation", + "content": "L_{D}" + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "text", + "content": " dual-pathway decoding layers. We denote the resulting predictions by " + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = \\left[\\mathbf{p};\\mathbf{d}_s;\\mathbf{d}_e\\right]" + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "text", + "content": ", the anchor queries by " + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_p" + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "text", + "content": ", and the boundary queries by " + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_s" + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_e" + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "text", + "content": ". Then we cast the predictions in the form of starting and ending timestamps, i.e., " + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\hat{\\varphi} = \\left[\\mathbf{p} - \\mathbf{d}_s;\\mathbf{p} + \\mathbf{d}_e\\right] \\in \\mathbb{R}^{M \\times 2}" + }, + { + "bbox": [ + 132, + 175, + 479, + 236 + ], + "type": "text", + "content": ", which serves as the final results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 251, + 272, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 272, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 272, + 262 + ], + "type": "text", + "content": "3.5 Quality-based Scoring" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 269, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 269, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 269, + 479, + 353 + ], + "type": "text", + "content": "After producing the moment predictions, we opt to rank them for evaluation. In the convention, query-based models utilize classification scores as the measure, which exhibits how well the proposals semantically match the sentence. However, it does not necessarily represent the localization qualities of proposals. Hence we propose to estimate the localization quality of each moment prediction. Formally, the quality score can be derived as " + }, + { + "bbox": [ + 131, + 269, + 479, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = \\sigma (\\mathrm{MLP}([\\mathbf{C}_p\\parallel \\mathbf{C}_s\\parallel \\mathbf{C}_e]))\\in \\mathbb{R}^M" + }, + { + "bbox": [ + 131, + 269, + 479, + 353 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 269, + 479, + 353 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 269, + 479, + 353 + ], + "type": "text", + "content": " is sigmoid activation. Then the quality loss is defined as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 223, + 360, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 360, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 223, + 360, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {q u a l}} = \\sum_ {m = 1} ^ {M} \\left| q _ {m} - \\max _ {\\forall n} \\left(\\frac {\\left| \\hat {\\varphi} _ {m} \\cap \\varphi_ {n} \\right|}{\\left| \\hat {\\varphi} _ {m} \\cup \\varphi_ {n} \\right|}\\right) \\right|, \\tag {7}", + "image_path": "cdbb69a6337401f99dcdb1522e7f6350e9d5fa3a568d7aac292b84c450f5c774.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 398, + 479, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 398, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 398, + 479, + 422 + ], + "type": "text", + "content": "where the objective of the quality head is to predict the maximum IoUs of the proposals with ground-truth moments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 438, + 209, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 209, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 209, + 450 + ], + "type": "text", + "content": "3.6 Matching" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 456, + 479, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 479, + 492 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 479, + 492 + ], + "type": "text", + "content": "As in the standard of query-based models [3,21], we perform Hungarian matching [19] between predictions and ground truths. The optimal matching results " + }, + { + "bbox": [ + 131, + 456, + 479, + 492 + ], + "type": "inline_equation", + "content": "\\psi^{*}" + }, + { + "bbox": [ + 131, + 456, + 479, + 492 + ], + "type": "text", + "content": " can be derived as follows." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 238, + 494, + 480, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 494, + 480, + 526 + ], + "spans": [ + { + "bbox": [ + 238, + 494, + 480, + 526 + ], + "type": "interline_equation", + "content": "\\psi^ {*} = \\arg \\min _ {\\psi \\in \\mathfrak {G} _ {N}} \\sum_ {n = 1} ^ {N} \\mathcal {C} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right), \\tag {8}", + "image_path": "8f07e6d85d6965bb2ce2b12d97d45b17b3d1763f8e61eaf8cc3f37644c28b377.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 184, + 529, + 427, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 529, + 427, + 543 + ], + "spans": [ + { + "bbox": [ + 184, + 529, + 427, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {C} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right) = \\lambda_ {l _ {1}} \\mathcal {L} _ {l _ {1}} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right) + \\lambda_ {\\mathrm {i o u}} \\mathcal {L} _ {\\mathrm {i o u}} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi (n)}\\right),", + "image_path": "16b07dc52fc295da4c3a02472f6684f0782b5dac3a2c58ab834b7bd19761fbce.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\mathfrak{G}_N" + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "content": " denotes the combination pool and " + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\psi (n)" + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "content": " is the index of the prediction matched by the " + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "content": "-th ground truth. " + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{l_1}" + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{iou}}" + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "content": " respectively represent the " + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "content": " distance and the generalized IoU [46] between the moments, while " + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{l_1}" + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{iou}}" + }, + { + "bbox": [ + 131, + 544, + 479, + 604 + ], + "type": "text", + "content": " are their weights. Note that in contrast to other works, the classification term is not involved in the matching process, leading to localization-oriented matching." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 605, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 479, + 628 + ], + "type": "text", + "content": "Once the matching is completed, we minimize the matching cost between each pair of the matching results " + }, + { + "bbox": [ + 131, + 605, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\psi^{*}" + }, + { + "bbox": [ + 131, + 605, + 479, + 628 + ], + "type": "text", + "content": ". The localization loss is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 250, + 635, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 635, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 250, + 635, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {l o c}} = \\sum_ {n = 1} ^ {N} \\mathcal {C} \\left(\\varphi_ {n}, \\hat {\\varphi} _ {\\psi^ {*} (n)}\\right). \\tag {9}", + "image_path": "563f2513fd47830a5f6fff078aa431ac22456522555a20539a05801a5a36611e.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAM-DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "Overall training objectives. Our model is trained in an end-to-end fashion and the overall training objective is defined as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 143, + 481, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 143, + 481, + 156 + ], + "spans": [ + { + "bbox": [ + 195, + 143, + 481, + 156 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {l o c}} + \\lambda_ {\\text {q u a l}} \\mathcal {L} _ {\\text {q u a l}} + \\lambda_ {\\text {s a l}} \\mathcal {L} _ {\\text {s a l}} + \\lambda_ {\\text {r e g u l}} \\mathcal {L} _ {\\text {r e g u l}}, \\tag {10}", + "image_path": "4c0a2aa93d9593153293fcf12a9acff662e6726e151d9ca78fe3d868d6388e25.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 158, + 307, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 307, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 307, + 171 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 158, + 307, + 171 + ], + "type": "inline_equation", + "content": "\\lambda_{*}" + }, + { + "bbox": [ + 132, + 158, + 307, + 171 + ], + "type": "text", + "content": " are the balancing parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 188, + 230, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 188, + 230, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 188, + 230, + 202 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 272, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 272, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 272, + 226 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 233, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 481, + 365 + ], + "type": "text", + "content": "Datasets. QVHighlights [21] is a recently built dataset, containing a total of 10,148 videos and 10,310 sentences from vlog and news domains. In addition to moments, it provides segment-level saliency score annotations within each moment. Importantly, it allows a single sentence corresponding to multiple disjoint moments (1.8 on average). Due to this practical setup, we utilize QVHighlights as the main benchmark. Charades-STA [11] includes 16,128 sentence-moment pairs with 9,848 indoor videos. The average duration of videos and moments is 30.6 and 8.1 seconds, respectively. TACoS [45] contains 127 cooking videos encompassing a total of 18,818 sentence-moment pairs. This dataset is known to be challenging since the moments occupy only a small portion (6.1 sec on average) within considerably long videos (4.8 min on average)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 460 + ], + "type": "text", + "content": "Evaluation metrics. Following the standard protocol, we measure the Recall@1 (R1) under the IoU thresholds of 0.3, 0.5, and 0.7 by default. Since QVHighlights contains multiple ground-truth moments per sentence, we report the mean average precision (mAP) with IoU thresholds of 0.5 and 0.75 as well as the average mAP over a set of IoU thresholds [0.5:0.05:0.95]. Meanwhile, we compute the mean IoU of top-1 predictions on Charades-STA and TACoS. Note that the performances at high IoU thresholds (e.g., 0.7) exhibit how well the predictions align with the ground truths." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 478, + 279, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 279, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 279, + 490 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 593 + ], + "type": "text", + "content": "For a fair comparison, we adopt the same feature extraction strategy with the competitors [21,26,41]. Specifically, we adopt the CLIP [44] text features for text and the concatenation of Slowfast [10] (ResNet-50) and CLIP [44] (ViT-B/32) features for videos unless otherwise specified. The video features are extracted every 1 second for Charades-STA and 2 seconds for QVHighlihtgs and TACoS. To compare with audio-augmented models [37], we optionally employ audio features extracted by PANNs [18] pre-trained on AudioSet [14]. Due to the long video duration, we uniformly sample 200 feature vectors from each video for TACoS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "We set the embedding dimension " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " to 256, the number of attention heads to 8, the number of queries " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " to 10, the number of boundary points " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " to 3, and the margin " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " to 0.2. We determine the numbers of encoding and decoding layers as same with the prior work [41], i.e., " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "L_{E} = L_{D} = 2" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": ". The balancing parameters are set as: " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{l1} = 10" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{iou}} = \\lambda_{\\mathrm{sal}} = \\lambda_{\\mathrm{regul}} = 1" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{qual}} = 2" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": ". As in previous works [26,41], we increase " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{sal}}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " to 4 when saliency labels are unavailable, i.e., for Charades-STA" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "text", + "content": "P. Lee and H. Byun" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 173, + 134, + 442, + 367 + ], + "blocks": [ + { + "bbox": [ + 206, + 114, + 406, + 126 + ], + "lines": [ + { + "bbox": [ + 206, + 114, + 406, + 126 + ], + "spans": [ + { + "bbox": [ + 206, + 114, + 406, + 126 + ], + "type": "text", + "content": "Table 2: Results on the QVHighlights test split." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 134, + 442, + 367 + ], + "lines": [ + { + "bbox": [ + 173, + 134, + 442, + 367 + ], + "spans": [ + { + "bbox": [ + 173, + 134, + 442, + 367 + ], + "type": "table", + "html": "
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
w/o pre-training
MCN [1]11.412.7224.948.2210.67
CAL [9]25.4911.5423.407.659.89
CLIP [44]16.885.1918.117.007.67
XML [22]46.6933.4647.8934.6734.90
Moment-DETR [21]52.8933.0254.8229.4030.73
UMT† [37]56.2341.1853.8337.0136.12
MH-DETR [60]60.0542.4860.7538.1338.38
QD-DETR [41]62.4044.9862.5239.8839.86
QD-DETR† [41]63.0645.1063.0440.1040.19
UniVTG [28]58.8640.8657.6035.5935.47
EaTR‡ [17]57.9842.4159.9539.2939.00
MomentDiff [26]57.4239.6654.0235.7335.95
BAM-DETR62.7148.6464.5746.3345.36
BAM-DETR†64.0748.1265.6147.5146.91
w/ pre-training on 4.2M data labeled by CLIP
UniVTG [28]65.4350.0664.0645.0243.63
w/ pre-training on 236K ASR captions
Moment-DETR [21]59.7840.3360.5135.3636.14
UMT† [37]60.8343.2657.3339.1238.08
QD-DETR [41]63.1845.1963.3740.3539.96
BAM-DETR63.8847.9266.3348.2246.67
", + "image_path": "93207d62e389d9eed764733d63b49577816eb983f6dcef3dc8e1784e891fc688.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 173, + 368, + 414, + 377 + ], + "lines": [ + { + "bbox": [ + 173, + 368, + 414, + 377 + ], + "spans": [ + { + "bbox": [ + 173, + 368, + 414, + 377 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 173, + 368, + 414, + 377 + ], + "type": "text", + "content": "additional use of audio modality " + }, + { + "bbox": [ + 173, + 368, + 414, + 377 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 173, + 368, + 414, + 377 + ], + "type": "text", + "content": "reproduced by official checkpoint" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 384, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 422 + ], + "type": "text", + "content": "and TACoS. Our model is trained from scratch for 200 epochs on QVHighlights and 100 epochs on the other datasets using the AdamW optimizer [38] with a learning rate of 1e-4 and a batch size of 32." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 441, + 334, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 334, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 334, + 453 + ], + "type": "text", + "content": "4.3 Comparison with State-of-the-arts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "text", + "content": "Results on QVHighlights. We compare our model with existing state-of-the-arts including recent query-based approaches [17, 21, 26, 37, 41, 60] on the test split. As shown in Table 2, our BAM-DETR consistently outperforms the comparative models under various settings. In detail, without pre-training, our model surpasses the previous state-of-the-art model [41] by large margins, e.g., " + }, + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "inline_equation", + "content": "3.66\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "text", + "content": " in R1@0.7 and " + }, + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "inline_equation", + "content": "6.45\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "text", + "content": " in mAP@0.75. These improvements under strict IoU thresholds verify the superior localization ability of our method. When leveraging auxiliary audio features, the performance further boosts especially in R1@0.5 and mAPs, enlarging the gap between the competitors including those with audio features [37, 41]. To compare with the methods with pretraining, we pretrain our model on middle-scale ASR caption data [21]. Again, our BAM-DETR achieves state-of-the-art results, while showing the least gap with the method [28] that leverages large-scale data for pre-training. Notably, our model even outperforms it in terms of mAPs with much fewer (about " + }, + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "inline_equation", + "content": "18\\times" + }, + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "text", + "content": ") pre-training data, manifesting the effectiveness of the proposed methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Results on Charades-STA. Experimental results on the test split are shown in Table 3. Not only does our BAM-DETR outperform the query-based com" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAM-DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 134, + 478, + 231 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 434, + 126 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 434, + 126 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 434, + 126 + ], + "type": "text", + "content": "Table 3: Results on the Charades-STA and TACoS test splits." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 134, + 478, + 231 + ], + "lines": [ + { + "bbox": [ + 136, + 134, + 478, + 231 + ], + "spans": [ + { + "bbox": [ + 136, + 134, + 478, + 231 + ], + "type": "table", + "html": "
MethodCharades-STATACoS
R1@0.3R1@0.5R1@0.7mIoUR1@0.3R1@0.5R1@0.7mIoU
2D-TAN [72]58.7646.0227.4041.2540.0127.9912.9227.22
VSLNet [70]60.3042.6924.1441.5835.5423.5413.1524.99
Moment-DETR [21]65.8352.0730.5945.5437.9724.6711.9725.49
QD-DETR [41]-57.3132.55-52.3936.7721.0735.76
UniVTG [28]70.8158.0135.6550.1051.4434.9717.3533.60
MomentDiff [26]-55.5732.42-46.6428.9212.3730.36
BAM-DETR72.9359.9539.3852.3356.6941.5426.7739.31
", + "image_path": "07fb522f5ce489018a86bff3bf088735c4d3fa27ffcbb9433a67cc8687f7975e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 266, + 478, + 354 + ], + "blocks": [ + { + "bbox": [ + 132, + 236, + 480, + 258 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 258 + ], + "type": "text", + "content": "Table 4: Results on the anti-biased Charades-STA test split against the moment location and length. VGG [52] and Glove [43] features are employed for all models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 266, + 478, + 354 + ], + "lines": [ + { + "bbox": [ + 136, + 266, + 478, + 354 + ], + "spans": [ + { + "bbox": [ + 136, + 266, + 478, + 354 + ], + "type": "table", + "html": "
Methodw.r.t. moment locationw.r.t. moment length
R1@0.3R1@0.5R1@0.7\\( \\mathrm{mAP}_{avg} \\)R1@0.3R1@0.5R1@0.7\\( \\mathrm{mAP}_{avg} \\)
2D-TAN [72]27.8120.4410.8417.2339.6828.6817.7222.79
MMN [56]33.5827.2014.1219.1843.5834.3119.9426.85
Moment-DETR [21]29.9422.1611.5618.6642.7334.3916.1224.02
\\( QD-DETR^‡ \\)[41]56.1746.8228.1330.7067.3954.4432.8736.99
MomentDiff [26]48.3933.5915.7121.3751.2538.3223.3828.19
BAM-DETR59.8350.0032.0831.6868.4055.4640.7443.21
", + "image_path": "a880065b250b21766432508c9956ba74da706f939c0db24a10566ba0f2ccbfbe.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 136, + 354, + 251, + 363 + ], + "lines": [ + { + "bbox": [ + 136, + 354, + 251, + 363 + ], + "spans": [ + { + "bbox": [ + 136, + 354, + 251, + 363 + ], + "type": "text", + "content": " reproduced by official codebase" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": "petitors [21, 26, 41], but it achieves a new state-of-the-art by surpassing the best performing anchor-free model [28] for all metrics. Notably, a large gap of " + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "inline_equation", + "content": "3.73\\%" + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": " is observed in R1@0.7, which confirms the strong localization ability of our model. Results on TACoS. We present the comparison results on the test set in Table 3. It can be observed that our BAM-DETR achieves a new state-of-the-art with pronounced performances under the strict IoU thresholds, which is consistent with the above results on other datasets. On this challenging benchmark, our model surpasses the previous best model [41] by " + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": " (relatively " + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": ") in R1@0.7. These results clearly exhibit the superiority of the proposed model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 495, + 275, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 275, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 275, + 506 + ], + "type": "text", + "content": "4.4 Robustness Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": "Query-based models potentially have a temporal bias [16, 63] against the locations and lengths of moments. To measure robustness, we evaluate our model on the anti-biased Charades-STA [26] with distribution shifts of the moment location and length between training and test sets. Table 4 summarizes the results, where our model outperforms all competitors under both anti-biased settings. Especially, it shows significant performance gaps under the moment length bias. This can be expected since our model directly localizes boundaries instead of predicting lengths, which lessens the effect of bias. This robustness test corroborates the advantage of our boundary-oriented moment modeling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 635, + 204, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 635, + 204, + 647 + ], + "spans": [ + { + "bbox": [ + 132, + 635, + 204, + 647 + ], + "type": "text", + "content": "4.5 Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 653, + 453, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 453, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 453, + 666 + ], + "type": "text", + "content": "We conduct analytical experiments on the QVHighlights validation split." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "P. Lee and H. Byun" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 169, + 134, + 445, + 201 + ], + "blocks": [ + { + "bbox": [ + 187, + 114, + 425, + 126 + ], + "lines": [ + { + "bbox": [ + 187, + 114, + 425, + 126 + ], + "spans": [ + { + "bbox": [ + 187, + 114, + 425, + 126 + ], + "type": "text", + "content": "Table 5: Ablation study of components on QVHighlights." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 169, + 134, + 445, + 201 + ], + "lines": [ + { + "bbox": [ + 169, + 134, + 445, + 201 + ], + "spans": [ + { + "bbox": [ + 169, + 134, + 445, + 201 + ], + "type": "table", + "html": "
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
Baseline62.3947.8762.6441.5441.75
+ boundary-oriented modeling63.4249.2362.8643.2442.42
+ dual-pathway decoder63.6150.2663.0144.9844.16
+ quality-based scoring65.1051.6165.4148.5647.61
", + "image_path": "dc8e96552aa8987ff809aa7a69ec5eab6f1dc392042cec87c8a756cd28c850f0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 240, + 234, + 306 + ], + "blocks": [ + { + "bbox": [ + 133, + 208, + 478, + 220 + ], + "lines": [ + { + "bbox": [ + 133, + 208, + 478, + 220 + ], + "spans": [ + { + "bbox": [ + 133, + 208, + 478, + 220 + ], + "type": "text", + "content": "Table 6: Ablative experiments on QVHighlights. Default settings are marked gray." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 229, + 232, + 239 + ], + "lines": [ + { + "bbox": [ + 137, + 229, + 232, + 239 + ], + "spans": [ + { + "bbox": [ + 137, + 229, + 232, + 239 + ], + "type": "text", + "content": "(a) Number of predictions" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 240, + 234, + 306 + ], + "lines": [ + { + "bbox": [ + 138, + 240, + 234, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 234, + 306 + ], + "type": "table", + "html": "
MR1mAP
@0.5@0.7Avg.
562.1949.0345.18
1065.1051.6147.61
1565.4850.3248.23
2065.2351.6148.01
", + "image_path": "234e5cb5642634cffd27e99598d79a0cdbc0f6c470d0f86cb13723faa9f2a8c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 243, + 240, + 357, + 297 + ], + "blocks": [ + { + "bbox": [ + 243, + 229, + 353, + 239 + ], + "lines": [ + { + "bbox": [ + 243, + 229, + 353, + 239 + ], + "spans": [ + { + "bbox": [ + 243, + 229, + 353, + 239 + ], + "type": "text", + "content": "(b) Choice of memory features" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 243, + 240, + 357, + 297 + ], + "lines": [ + { + "bbox": [ + 243, + 240, + 357, + 297 + ], + "spans": [ + { + "bbox": [ + 243, + 240, + 357, + 297 + ], + "type": "table", + "html": "
MemoryR1mAP Avg.
@0.5@0.7
plain63.7449.6846.21
sensitive63.8149.8746.34
merged65.1051.6147.61
", + "image_path": "10002ff49de3faa01d89f91c605ba683bf6a32a4ea96ea0a2d790248f1f6494b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 366, + 240, + 479, + 306 + ], + "blocks": [ + { + "bbox": [ + 363, + 229, + 476, + 239 + ], + "lines": [ + { + "bbox": [ + 363, + 229, + 476, + 239 + ], + "spans": [ + { + "bbox": [ + 363, + 229, + 476, + 239 + ], + "type": "text", + "content": "(c) Number of boundary points" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 366, + 240, + 479, + 306 + ], + "lines": [ + { + "bbox": [ + 366, + 240, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 366, + 240, + 479, + 306 + ], + "type": "table", + "html": "
KR1mAP
@0.5@0.7Avg.
1 (fixed)63.2348.3245.17
164.3249.8747.14
365.1051.6147.61
564.9050.4546.70
", + "image_path": "9dc75ae1b163f66d079f699c997a35da42920a65438b87d3eb10f30467e9ad52.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 140, + 321, + 272, + 386 + ], + "blocks": [ + { + "bbox": [ + 133, + 309, + 484, + 319 + ], + "lines": [ + { + "bbox": [ + 133, + 309, + 484, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 309, + 484, + 319 + ], + "type": "text", + "content": "(d) Query choices for quality prediction (e) Combinations of attention layers for updating pathways" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 321, + 272, + 386 + ], + "lines": [ + { + "bbox": [ + 140, + 321, + 272, + 386 + ], + "spans": [ + { + "bbox": [ + 140, + 321, + 272, + 386 + ], + "type": "table", + "html": "
CpCsCeR1mAP
@0.5@0.7Avg.
✓62.9749.0346.26
✓64.3950.7146.66
✓64.5849.2346.71
✓✓✓65.1051.6147.61
", + "image_path": "9f735c3269b61d11fb96d4e56382a4bab13832c8a4d5b91bc96412dbf6c13532.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 280, + 321, + 483, + 386 + ], + "blocks": [ + { + "bbox": [ + 280, + 321, + 483, + 386 + ], + "lines": [ + { + "bbox": [ + 280, + 321, + 483, + 386 + ], + "spans": [ + { + "bbox": [ + 280, + 321, + 483, + 386 + ], + "type": "table", + "html": "
Anchor queryBoundary queryR1 @0.5@0.7mAP Avg.FLOPsParams
Global (shared)62.2649.0344.180.71G8.2M
GlobalGlobal63.6149.8144.540.72G11.5M
GlobalFocused63.7449.6846.210.65G9.5M
FocusedFocused62.4849.1645.870.62G7.6M
", + "image_path": "bf269cb49dee9ba5148519c8fa1a15fd9183910a1a7bd35d0a9ed542f0aad8de.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 399, + 481, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 471 + ], + "type": "text", + "content": "Effect of each component. We analyze the effect of each component in Table 5. The direct adoption of our moment modeling solely improves the performance, particularly at strict thresholds, showcasing its advantage in boundary alignment. Employing the dual-pathway decoder leads to further gains, which suggests the essential role of separate pathways. Lastly, the quality-based scoring considerably elevates the scores, especially in terms of mAPs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 472, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 509 + ], + "type": "text", + "content": "Number of predictions. We experiment with varying numbers of predictions in Table 6a, where the model achieves robust results when " + }, + { + "bbox": [ + 130, + 472, + 481, + 509 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 472, + 481, + 509 + ], + "type": "text", + "content": " is sufficiently large. By default, we set " + }, + { + "bbox": [ + 130, + 472, + 481, + 509 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 472, + 481, + 509 + ], + "type": "text", + "content": " to 10 for a fair comparison with the previous works [21,41]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 509, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 556 + ], + "type": "text", + "content": "Locality-enhanced features. We analyze the effect of the choice of input features for boundary-focused attention in Table 6b. The results indicate that boundary-sensitive features are slightly more helpful than plain ones in precise localization. In addition, the merged features achieve the best performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "text", + "content": "Number of sampled boundary points. We analyze the effect of the number of sampled boundary points " + }, + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "text", + "content": " in Table 6c. For comparisons, we report the case where the fixed boundary features are sampled without using offsets (" + }, + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "text", + "content": " row). As shown in the table, the dynamic selection of neighborhoods rather than fixed boundaries is important for accurate moment localization. The performance improves when sampling multiple points, while it saturates at " + }, + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 130, + 557, + 481, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Query choices for quality prediction. Our model leverages both anchor and boundary queries for quality prediction. We investigate the effect of query choices in Table 6d. As a result, utilizing all the queries shows better localization" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAM-DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 115, + 239, + 188 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 239, + 188 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 239, + 188 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 239, + 188 + ], + "type": "image", + "image_path": "a3183881a6d67efc4689850fe951014885c2be812c4cc1e918f170c9e79e6fe6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 190, + 226, + 198 + ], + "lines": [ + { + "bbox": [ + 146, + 190, + 226, + 198 + ], + "spans": [ + { + "bbox": [ + 146, + 190, + 226, + 198 + ], + "type": "text", + "content": "(a) Boundary hit rate" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 206, + 208, + 406, + 219 + ], + "lines": [ + { + "bbox": [ + 206, + 208, + 406, + 219 + ], + "spans": [ + { + "bbox": [ + 206, + 208, + 406, + 219 + ], + "type": "text", + "content": "Fig. 4: Analytical experiments on QVHighlights." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 482, + 187 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 482, + 187 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 482, + 187 + ], + "type": "image", + "image_path": "f84c2f8966cf19bf98b821dd9769ed1120c79d299a12ce45c3faec02403e29e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 190, + 410, + 198 + ], + "lines": [ + { + "bbox": [ + 322, + 190, + 410, + 198 + ], + "spans": [ + { + "bbox": [ + 322, + 190, + 410, + 198 + ], + "type": "text", + "content": "(b) Visualization results" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 479, + 251 + ], + "type": "text", + "content": "performance than using one of them. This might be trivial as different queries contain complementary information for quality prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 252, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 480, + 346 + ], + "type": "text", + "content": "Attention layers. Table 6e compares different combinations of attention layers for updating pathways in terms of performances and costs. For an apple-to/apple comparison, we leverage the plain memory features for both attention layers. First of all, employing separate global cross-attention for different queries leads to a huge parameter increase yet limited score gains. Replacing the global one for boundary queries with our focused attention layer substantially reduces the cost, while achieving the best performance. Meanwhile, the use of focused attention for both types of queries leads to inferior results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 348, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 467 + ], + "type": "text", + "content": "Boundary alignment. Standard IoU-based metrics are indirect measures for boundary alignment. To precisely diagnose the ability, inspired by Deeplab [5], we compute the boundary hit rate of predictions with varying band widths. We expand ground-truth boundary points with a band width to form starting/ending zones and regard a prediction as correct if both of its boundaries fall within the corresponding zones. More details can be found in Appendix. To disentangle the effect of ranking, we mark a video as correct if at least one prediction is correct and measure the video-level hit rate. As shown in Fig. 4a, our model greatly outperforms the recent competitors. The sharp increase in low band widths clearly validates the superiority of our model in boundary alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 467, + 480, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 480, + 528 + ], + "type": "text", + "content": "Qualitative results. As shown in Fig. 4b, previous models fail to localize accurate moments with misleading center predictions. Especially, QD-DETR [41] accurately predicts the moment length but suffers from the misaligned center, leading to the limited IoU. In contrast, thanks to the novel moment modeling, our model predicts well-aligned boundaries without relying on center prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 544, + 220, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 220, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 220, + 558 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "In this paper, we identified the center misalignment issue of existing query-based models for sentence grounding. To address it, we presented boundary-oriented moment modeling where boundaries are directly predicted without relying on centers. Based on the modeling, we designed the boundary-aligned moment detection transformer characterized by dual-pathway decoding. Further, we proposed localization quality-based scoring of predictions. The efficacy of the proposed methods is validated by thorough examinations. We hope this work sheds light on the issue of center-based moment modeling in detection transformers." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "P. Lee and H. Byun" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 174 + ], + "type": "text", + "content": "This project was supported by the National Research Foundation of Korea grant funded by the Korea government (MSIT) (No. 2022R1A2B5B02001467; RS-2024-00346364)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 212, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 480, + 243 + ], + "type": "text", + "content": "1. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017) 1, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 267 + ], + "type": "text", + "content": "2. Cao, M., Chen, L., Shou, M.Z., Zhang, C., Zou, Y.: On pursuit of designing multimodal transformer for video grounding. In: EMNLP. pp. 9810-9823 (2021) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 298 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229. Springer (2020) 2, 4, 6, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 320 + ], + "type": "text", + "content": "4. Chen, J., Chen, X., Ma, L., Jie, Z., Chua, T.S.: Temporally grounding natural sentence in video. In: EMNLP. pp. 162-171 (2018) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 321, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 353 + ], + "type": "text", + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. PAMI 40(4), 834-848 (2017) 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "type": "text", + "content": "6. Chen, L., Lu, C., Tang, S., Xiao, J., Zhang, D., Tan, C., Li, X.: Rethinking the bottom-up framework for query-based video localization. In: AAAI. vol. 34, pp. 10551-10558 (2020) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 386, + 480, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 416 + ], + "type": "text", + "content": "7. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: ICCV. pp. 2988-2997 (2021) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 450 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 480, + 482 + ], + "type": "text", + "content": "9. Escorcia, V., Soldan, M., Sivic, J., Ghanem, B., Russell, B.: Temporal localization of moments in video collections with natural language. arXiv preprint arXiv:1907.12763 (2019) 1, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 504 + ], + "type": "text", + "content": "0. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6202-6211 (2019) 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 525 + ], + "type": "text", + "content": "1. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: CVPR. pp. 5267-5275 (2017) 1, 4, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 546 + ], + "type": "text", + "content": "2. Gao, P., Zheng, M., Wang, X., Dai, J., Li, H.: Fast convergence of detr with spatially modulated co-attention. In: ICCV. pp. 3621-3630 (2021) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "text", + "content": "3. Ge, R., Gao, J., Chen, K., Nevatia, R.: Mac: Mining activity concepts for language-based temporal localization. In: WACV. pp. 245-253. IEEE (2019) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "text", + "content": "4. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: ICASSP. pp. 776-780. IEEE (2017) 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 601, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 631 + ], + "type": "text", + "content": "5. Ghosh, S., Agarwal, A., Parekh, Z., Hauptmann, A.G.: Excl: Extractive clip localization using natural language descriptions. In: NAACL. pp. 1984-1990 (2019) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "6. Hao, J., Sun, H., Ren, P., Wang, J., Qi, Q., Liao, J.: Can shuffling video benefit temporal bias problem: A novel training framework for temporal grounding. In: ECCV. pp. 130-147. Springer (2022) 12" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAM-DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "text", + "content": "17. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: ICCV. pp. 13846-13856 (2023) 2, 4, 6, 7, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "18. Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Trans. on Audio, Speech, and Language Processing 28, 2880-2894 (2020) 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "text", + "content": "19. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "20. Lee, P., Kim, T., Shim, M., Wee, D., Byun, H.: Decomposed cross-modal distillation for rgb-based temporal action detection. In: CVPR. pp. 2373-2383 (2023) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "text", + "content": "21. Lei, J., Berg, T.L., Bansal, M.: Detecting moments and highlights in videos via natural language queries. In: Neurips. vol. 34, pp. 11846-11858 (2021) 2, 3, 4, 5, 6, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "text", + "content": "22. Lei, J., Yu, L., Berg, T.L., Bansal, M.: Tv: A large-scale dataset for video-subtitle moment retrieval. In: ECCV. pp. 447-463. Springer (2020) 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 308 + ], + "type": "text", + "content": "23. Li, F., Zeng, A., Liu, S., Zhang, H., Li, H., Zhang, L., Ni, L.M.: Lite detr: An interleaved multi-scale encoder for efficient detr. In: CVPR. pp. 18558-18567 (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 310, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 332 + ], + "type": "text", + "content": "24. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: CVPR. pp. 13619-13627 (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "25. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In: ICCV. pp. 12032-12042 (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "26. Li, P., Xie, C.W., Xie, H., Zhao, L., Zhang, L., Zheng, Y., Zhao, D., Zhang, Y.: Momentdiff: Generative video moment retrieval from random to real. In: Neurips (2023) 2, 4, 6, 10, 11, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "text", + "content": "27. Lin, C., Xu, C., Luo, D., Wang, Y., Tai, Y., Wang, C., Li, J., Huang, F., Fu, Y.: Learning salient boundary feature for anchor-free temporal action localization. In: CVPR. pp. 3320-3329 (2021) 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "28. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: ICCV. pp. 2794-2804 (2023) 4, 5, 6, 11, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "type": "text", + "content": "29. Lin, T., Zhao, X., Su, H., Wang, C., Yang, M.: Bsn: Boundary sensitive network for temporal action proposal generation. In: ECCV. pp. 3-19 (2018) 1, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "30. Lin, Y., Yuan, Y., Zhang, Z., Li, C., Zheng, N., Hu, H.: Detr does not need multiscale or locality design. In: ICCV. pp. 6545-6554 (2023) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "31. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: CVPR. pp. 11235-11244 (2021) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "32. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: CVPR. pp. 11235-11244 (2021) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "33. Liu, F., Wei, H., Zhao, W., Li, G., Peng, J., Li, Z.: Wb-detr: transformer-based detector without backbone. In: ICCV. pp. 2979-2987 (2021) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "34. Liu, M., Wang, X., Nie, L., Tian, Q., Chen, B., Chua, T.S.: Cross-modal moment localization in videos. In: ACM MM. pp. 843-851 (2018) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "35. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. In: ICLR (2022) 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "36. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., Zhang, L.: Detection transformer with stable matching. In: ICCV. pp. 6491-6500 (2023) 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "type": "text", + "content": "P. Lee and H. Byun" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: CVPR. pp. 3042-3051 (2022) 2, 4, 5, 10, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "38. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 204 + ], + "type": "text", + "content": "39. Lu, C., Chen, L., Tan, C., Li, X., Xiao, J.: Debug: A dense bottom-up grounding approach for natural language video localization. In: EMNLP-IJCNLP. pp. 5144-5153 (2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 236 + ], + "type": "text", + "content": "40. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: ICCV. pp. 3651-3660 (2021) 4, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "41. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: CVPR. pp. 23023-23033 (2023) 2, 3, 4, 6, 7, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "42. Mun, J., Cho, M., Han, B.: Local-global video-text interactions for temporal grounding. In: CVPR. pp. 10810-10819 (2020) 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "43. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP. pp. 1532-1543 (2014) 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021) 10, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 378 + ], + "type": "text", + "content": "45. Regneri, M., Rohrbach, M., Wetzel, D., Thater, S., Schiele, B., Pinkal, M.: Grounding action descriptions in videos. Trans. Assoc. Comput. Linguistics 1, 25-36 (2013) 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "46. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: CVPR. pp. 658-666 (2019) 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "text", + "content": "47. Roh, B., Shin, J., Shin, W., Kim, S.: Sparse detr: Efficient end-to-end object detection with learnable sparsity. In: ICLR (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "48. Rui, Y., Gupta, A., Acero, A.: Automatically extracting highlights for tv baseball programs. In: ACM MM. pp. 105-115 (2000) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "text", + "content": "49. Shao, D., Xiong, Y., Zhao, Y., Huang, Q., Qiao, Y., Lin, D.: Find and focus: Retrieve and localize video events with natural language queries. In: ECCV. pp. 200-216 (2018) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "50. Sharghi, A., Laurel, J.S., Gong, B.: Query-focused video summarization: Dataset, evaluation, and a memory network based approach. In: CVPR. pp. 4788-4797 (2017) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "51. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: CVPR. pp. 18857-18866 (2023) 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "52. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "text", + "content": "53. Sun, Z., Cao, S., Yang, Y., Kitani, K.M.: Rethinking transformer-based set prediction for object detection. In: ICCV. pp. 3611-3620 (2021) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Neurips. vol. 30 (2017) 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "55. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: AAAI. vol. 36, pp. 2567-2575 (2022) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "56. Wang, Z., Wang, L., Wu, T., Li, T., Wu, G.: Negative sample matters: A renaissance of metric learning for temporal grounding. In: AAAI. vol. 36, pp. 2613-2623 (2022) 4, 12" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAM-DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "57. Xiao, S., Chen, L., Zhang, S., Ji, W., Shao, J., Ye, L., Xiao, J.: Boundary proposal network for two-stage natural language video localization. In: AAAI. vol. 35, pp. 2986-2994 (2021) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "58. Xiong, B., Kalantidis, Y., Ghadiyaram, D., Grauman, K.: Less is more: Learning highlight detection from video duration. In: CVPR. pp. 1258-1267 (2019) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "text", + "content": "59. Xu, H., He, K., Plummer, B.A., Sigal, L., Sclaroff, S., Saenko, K.: Multilevel language and vision integration for text-to-clip retrieval. In: AAAI. vol. 33, pp. 9062–9069 (2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "60. Xu, Y., Sun, Y., Li, Y., Shi, Y., Zhu, X., Du, S.: Mh-detr: Video moment and highlight detection with cross-modal transformer. In: ACM MM (2023) 4, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "61. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: Unloc: A unified framework for video localization tasks. In: ICCV. pp. 13623-13633 (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "62. Ye, M., Ke, L., Li, S., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Cascade-detr: Delving into high-quality universal object detection. In: ICCV. pp. 6704-6714 (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "63. Yuan, Y., Lan, X., Wang, X., Chen, L., Wang, Z., Zhu, W.: A closer look at temporal sentence grounding in videos: Dataset and metric. In: Proc. 2nd Int. Workshop on Human-Centric Multimedia Analysis. pp. 13-21 (2021) 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "64. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In: Neurips. vol. 32 (2019) 1, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 357, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 480, + 390 + ], + "type": "text", + "content": "65. Yuan, Y., Mei, T., Zhu, W.: To find where you talk: Temporal sentence localization in video with attention based location regression. In: AAAI. vol. 33, pp. 9159-9166 (2019) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "text", + "content": "66. Zeng, R., Xu, H., Huang, W., Chen, P., Tan, M., Gan, C.: Dense regression network for video grounding. In: CVPR. pp. 10287-10296 (2020) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "67. Zhang, D., Dai, X., Wang, X., Wang, Y.F., Davis, L.S.: Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In: CVPR. pp. 1247-1257 (2019) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "68. Zhang, G., Luo, Z., Yu, Y., Cui, K., Lu, S.: Accelerating detr convergence via semantic-aligned matching. In: CVPR. pp. 949-958 (2022) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "69. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: ICLR (2023) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "70. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. In: ACL (2020) 4, 5, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "71. Zhang, K., Chao, W.L., Sha, F., Grauman, K.: Video summarization with long short-term memory. In: ECCV. pp. 766-782. Springer (2016) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "72. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. vol. 34, pp. 12870-12877 (2020) 1, 4, 5, 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "type": "text", + "content": "73. Zhang, S., Su, J., Luo, J.: Exploiting temporal relationships in video moment localization with natural language. In: ACM MM. pp. 1230-1238 (2019) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "74. Zhang, Z., Lin, Z., Zhao, Z., Xiao, Z.: Cross-modal interaction networks for query-based moment retrieval in videos. In: ACM SIGIR. pp. 655-664 (2019) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "75. Zhao, Y., Xiong, Y., Wang, L., Wu, Z., Tang, X., Lin, D.: Temporal action detection with structured segment networks. In: ICCV. pp. 2914-2923 (2017) 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "76. Zheng, D., Dong, W., Hu, H., Chen, X., Wang, Y.: Less is more: Focus attention for efficient detr. In: ICCV. pp. 6674-6683 (2023) 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "text", + "content": "P. Lee and H. Byun" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 162 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "77. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: CVPR. pp. 9308-9316 (2019) 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 162 + ], + "type": "text", + "content": "78. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021) 4, 7, 8, 9" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAM-DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BAMM_ Bidirectional Autoregressive Motion Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_content_list.json b/2024/BAMM_ Bidirectional Autoregressive Motion Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7279701cc5ed2573228dc987e9e832cc30ce1819 --- /dev/null +++ b/2024/BAMM_ Bidirectional Autoregressive Motion Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_content_list.json @@ -0,0 +1,1496 @@ +[ + { + "type": "text", + "text": "BAMM: Bidirectional Autoregressive Motion Model", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ekkasit Pinyoanuntapong1, Muhammad Usama Saleem1, Pu Wang1, Minwoo Lee1, Srijan Das1, and Chen Chen2", + "bbox": [ + 236, + 212, + 764, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 University of North Carolina at Charlotte {epinyoan,msaleem2, Pu.Wang, minwoo.lee, sdas24}@uncc.edu", + "2 University of Central Florida chen.chen@crcv.ucf.edu" + ], + "bbox": [ + 241, + 253, + 756, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ada26cc70a325b7862be6684fdc5a829b2bccc205a06f6c55132ff335e6673d.jpg", + "image_caption": [ + "a) Motion Length Prediction" + ], + "image_footnote": [], + "bbox": [ + 220, + 332, + 354, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4664df2cc75ea62b3c0b5fd7146c0bb9c31fdcbef7f2b93467c1cf2fe01760d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 321, + 511, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cbbc3c6b7b3e0e524c61171c8e742a57eb7bda0e64be05a2de188aed2007ca99.jpg", + "image_caption": [ + "Fig. 1: (a) Motion Length Prediction: Text-to-motion models often require specific input lengths, making them sensitive to motion generation. In contrast, BAMM automatically predicts the end of the motion, thus avoiding reliance on inaccurate motion length estimations. (b) High-quality Text-to-Motion: BAMM generates natural human movements precisely aligned with detailed textual descriptions. (c) Motion Editing: BAMM is capable of multiple editing tasks, such as inpainting (as demonstrated), outpainting, prefix prediction, suffix completion, and arbitrarily long motion sequence synthesis." + ], + "image_footnote": [], + "bbox": [ + 366, + 345, + 511, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff1e38a1bdc1f8d193e019b03893cea99147415a4afa7af04fae0ec07277b385.jpg", + "image_caption": [ + "c) Motion Editing" + ], + "image_footnote": [], + "bbox": [ + 542, + 333, + 781, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Generating human motion from text has been dominated by denoising motion models either through diffusion or generative masking process. However, these models face great limitations in usability by requiring prior knowledge of the motion length. Conversely, autoregressive motion models address this limitation by adaptively predicting motion endpoints, at the cost of degraded generation quality and editing capabilities. To address these challenges, we propose Bidirectional Autoregressive Motion Model (BAMM), a novel text-to-motion generation framework. BAMM consists of two key components: (1) a motion tokenizer that transforms 3D human motion into discrete tokens in latent space, and (2) a masked self-attention transformer that autoregressively predicts randomly masked tokens via a hybrid attention masking strategy. By unifying generative masked modeling and autoregressive modeling, BAMM captures rich and bidirectional dependencies among motion tokens, while learning the probabilistic mapping from textual inputs to motion outputs with dynamically-adjusted motion sequence length. This feature enables BAMM to simultaneously achieving high-quality motion generation with enhanced usability and built-in motion editability. Extensive experiments on HumanML3D and KIT-ML datasets demonstrate that BAMM surpasses current state-of-the-art methods in both", + "bbox": [ + 259, + 561, + 740, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qualitative and quantitative measures. Our project page is available at https://exitudio.github.io/BAMM-page", + "bbox": [ + 259, + 147, + 738, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Text to Motion $\\cdot$ Autoregressive Motion Model $\\cdot$ Generative Masked Motion Model", + "bbox": [ + 261, + 188, + 740, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 248, + 375, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-motion generation is a promising interdisciplinary field that uses natural language to generate 3D human movements. This emerging field holds vast potential to transform animation, gaming, filming, and VR/AR/MR domains by enabling easy and intuitive creation of 3D assets through user-friendly textual inputs. However, bridging the semantic gap between textual descriptions and intricate motion sequences presents a significant challenge. To address this challenge, recent efforts have been focusing on two methods: (1) conditional denoising motion model and (2) conditional autoregressive motion model. Both methods can greatly improve motion generation quality by learning the probabilistic distribution of motion sequences, conditioned on the textural descriptors. However, both methods face fundamental limitations.", + "bbox": [ + 212, + 280, + 787, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditional denoising motion models are trained to restore corrupted motion sequences to their original state, guided by textual prompts. These models operate through two primary mechanisms: diffusion and generative masking. Diffusion models apply structured Gaussian noise to the original motion data for corruption [22,37,41,46], whereas generative masked models [14,29] corrupt motion sequence by substituting selected motion tokens with [MASK] tokens. The denoising process, followed by motion corruption procedure, considers motion tokens from both directions, effectively capturing the intricate dependencies among tokens. This leads to enhanced motion generation quality. Furthermore, these models inherently ease the motion editing tasks. By selectively corrupting and recovering motion tokens in areas needing modifications, they ensure seamless transitions between edited and unedited segments.", + "bbox": [ + 212, + 446, + 787, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While denoising motion models excel in generation quality and editability, a significant limitation of these models is their usability because these models depend on prior knowledge of motion length for each text prompt, a requirement that proves impractical in real-world scenarios. Utilizing incorrect motion lengths can result in a significant decline in generation quality. To address this fundamental limitation, conditional motion autoregressive models emerge as a solution, capable of simultaneously predicting the sequence length and content of generated motions. Inspired by large language models like GPT [5], motion autoregressive models sequentially predict one motion token at a time from left to right until the [END] token is predicted, guided by the textual description [21, 45, 49]. As a result, the generated motions are not only well aligned with the text inputs but also appropriately scaled in duration. However, the sequential token decoding of autoregressive models cannot fully capture the dependencies between the motion tokens, potentially compromising generation quality, and complicating", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "E. Pinyoanuntapong et al.", + "bbox": [ + 271, + 114, + 446, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the editing process, because edited parts need to be conditioned on the unedited parts from both directions to ensure overall continuity and coherence.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As previously highlighted, existing text-driven motion generation models encounter a comprise among usability, quality of generation, and editability. To address this challenge, we propose Bidirectional Autoregressive Motion Model (BAMM), a novel text-to-motion generation framework. It consists of two pivotal components: a motion tokenizer and a conditional masked self-attention transformer. In a two-stage training paradigm, the motion tokenizer performs initial training based on Vector Quantized Variational Autoencoders (VQ-VAE) [26]. This tokenizer encodes the raw motion sequence into discrete motion tokens within the latent space leveraging learned codebooks. In the subsequent phase, motion tokens are randomly masked out. A conditional masked self-attention transformer is then trained to autoregressively predict these masked tokens, adopting the causal attention masking strategy. This strategy is largely departing from the traditional [MASK] token replacement approach in generative masked models. In particular, it does not substitute the input motion tokens with [MASK] tokens. Instead, it adjusts the attention score matrix according to both unidirectional and bidirectional causal masks. The unidirectional causal mask enables adaptive prediction of [END] token based on text prompts, while bidirectional causal mask forces the model to predict the next motion token not only based on past tokens but also conditioned on future unmasked tokens. This facilitates bidirectional autoregressive training for enhanced predictive capabilities.", + "bbox": [ + 212, + 178, + 787, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: Comparison of quality and capability of generation on text-to-motion to state-of-the-art models on the largest text-to-motion dataset [16]. $\\checkmark$ means capability while $X$ is not. \"Predict Length\" denotes the ability to generate motion without prior knowledge of motion length. \"Input Length\" refers to the ability to take input length as a constraint, while \"Edit\" indicates motion editability. Since MMM and MoMask require ground-truth motion length as input, we use predicted motion length from pretrained length estimator by [16]. The lowest FID score means the best overall quality of the generated motion, ensuring that its authenticity and naturalness is very close to the ground-truth human movements. High R-precision and low MM-dist means accurate alignment between the generated motion and the text prompts.", + "bbox": [ + 212, + 494, + 787, + 633 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/2a2aa0787a8b78d64dab04b95050582396014ca6c6c133ffd58506d27e7d6419.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTop-1 ↑FID ↓MM-Dist ↓Predict LengthInput LengthEdit
T2M-GPT0.4910.1163.118✓✗✗
AttT2M0.4990.1123.038✓✗✗
MMM0.5040.0802.998✗✓✓
MoMask0.5220.0902.945✗✓✓
BAMM (Ours)0.5250.0552.919✓✓✓
", + "bbox": [ + 313, + 645, + 687, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By unifying masked and autoregressive prediction during training, BAMM captures rich and bidirectional dependencies among motion tokens, while learning a direct probabilistic mapping from textual inputs to motion outputs with dynamically-adjusted motion sequence length. Leveraging such unique feature, we propose cascaded motion generation during inference, where BAMM first leverages unidirectional autoregressive decoding to implicitly predict motion se", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BAMM", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "quence length and generate coarse-grained motion sequence. Such motion sequence is then refined by masking and regenerating a portion of motion tokens in a bidirectional autoregressive fashion. This feature allows BAMM to achieve high-quality motion generation with high usability. Moreover, BAMM naturally supports zero-shot motion editing without specially being trained for such task. By treating the masked motion tokens as the contents that need editing, BAMM can predict the masked tokens based on the surrounding context and the text description. Our main contributions can be summarized as follows.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the bidirectional autoregressive motion model, which is a novel text-to-motion generation framework. It effectively harnesses the complementary benefits of denoising and autoregressive models, thus simultaneously achieving high-quality motion generation with enhanced usability and innate motion editability, as showcased in Fig. 1 and Table 1.", + "- We demonstrate that our model outperforms current state-of-the-art methods qualitatively and quantitatively on two standard text-to-motion generation datasets, HumanML3D [16] and KIT-ML [30].", + "- We showcase that our model supports a variety of motion editing tasks in the zero-shot manner without specially training for these tasks, including motion inpainting, outpainting, prefix prediction, suffix completion, and long sequence generation." + ], + "bbox": [ + 225, + 281, + 782, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 488, + 387, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Synthesis with Latent Space Alignment. Early text-to-motion approaches typically adopt a two-stage scheme: learning separate latent representations for text and motion sequences, followed by latent space alignment using distance losses like cosine similarity or KL divergence [2,16,27,28,36,42]. For example, Language2Pose [2] aimed to establish a shared latent space for both language descriptions and motion sequences. MotionCLIP [36] simplified this concept by incorporating stylization and diversity techniques, leveraging the pre-trained text-image latent space of the CLIP model [32]. However, this strategy inherently struggles with generating high-fidelity motions due to the difficulty of perfectly aligning these inherently disparate latent spaces.", + "bbox": [ + 212, + 522, + 787, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditional Denoising Motion Model. Inspired by the success of denoising diffusion models (DDMs) [19,35] in text-to-image and text-to-video generation [18, 25, 33, 34], diffusion models have been applied for text-to-motion generation. MDM [37], MotionDiffuse [46], MLD [8], and FLAME [22] are the representative examples. Meanwhile, BERT-type masked generative models have shown their success in both text generation tasks, e.g., Q&A and language translations [9,12,31], and text-to-image synthesis [6,7,10,39,43,47,48]. Following this trend, MMM [29] and MoMask [14] are recent attempts to propose conditional masked motion modeling to enable high-fidelity motion synthesis that is precisely aligned with text prompts. Both diffusion and generative masked modeling share the same principle: \"denoising\" corrupted data [3]. Meanwhile, they also face the", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "E. Pinyoanuntapong et al.", + "bbox": [ + 271, + 114, + 446, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "same usability limitation: they require prior knowledge of the motion sequence length for the denoising process. This length information is obtained either by intuitive guess from users or via a separately pre-trained predictor that predicts the motion length distribution based on the input text prompts [16]. Both approaches, however, lead to significant motion quality degradation as demonstrated in the experiments section. This is because the distributions of motion length and motion contents are inherently coupled, which need to be jointly aligned with the textual context.", + "bbox": [ + 212, + 146, + 787, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditional Autoregressive Motion Model. The autoregressive motion models, such as T2M-GPT [45], AttT2M [49], and MotionGPT [21], can effectively address the usability challenge faced by denoising models because they follow the GPT-type training and inference [5] to implicitly predict motion sequence length by generating the [END] token conditioned on both previously generated motion tokens and text inputs. However, the limitation of these models lies in their use of causal attention for unidirectional and sequential motion token prediction. This practice not only hinders model's motion editability but also jeopardizes motion generation quality. To address this limitation, we propose the first bidirectional autoregressive modeling approach for human motion generation, which draws inspiration from the self-attention masking employed by large language model pretraining [11,38].", + "bbox": [ + 212, + 268, + 787, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 474, + 330, + 489 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/84c38308060370eb216a17d42ed87844ef9168e44236fe87b0e6ecfd71087d4f.jpg", + "image_caption": [ + "Fig.2: Overall architecture of BAMM. (a) Motion Tokenizer encodes the raw motion sequence into discrete motion tokens according to a learned codebook. (b) Masked Self-attention Transformer learns to sequentially predict next tokens conditioned on text embedding from CLIP model and future unmasked tokens. Masked self-attention mechanism unifies autoregressive model and generative masked motion via bidirectional and unidirectional causal masks." + ], + "image_footnote": [], + "bbox": [ + 218, + 518, + 759, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our objective is to create a text-to-motion synthesis framework that simultaneously achieves high-quality motion generation with enhanced usability and innate motion editability. Towards this goal, our framework, as illustrated in Fig. 2, consists of two key components: motion tokenizer that compresses and converts raw 3D human motion into a sequence of discrete motion tokens in the", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BAMM", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "latent space (Section 3.1) and conditional masked self-attention transformer that leverages unidirectional and bidirectional casual masks to integrate autoregressive model and generative masked model into a unified framework (Section 3.2). The training procedure follows a hybrid attention masking strategy, where the two causal masks are applied randomly and the model is forced to reconstruct the motion sequence under both cases (Section 3.3). The cascaded motion decoding is introduced for motion generation during the inference phase. It uses unidirectional autoregressive decoding to jointly predict motion sequence length and its contents, which are refined via bidirectional autoregressive decoding (Section 3.4).", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Motion Tokenizer", + "text_level": 1, + "bbox": [ + 215, + 319, + 408, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The objective of this stage is to learn the discrete representation of motion by quantizing the embedding $z$ from the output of the encoder into codebook $\\mathcal{C}$ . We first pretrain a motion tokenizer based on VQ-VAE [26]. In particular, given a motion sequence $\\mathcal{M} = [m_1, m_2, m_3, \\dots, m_\\tau]$ where $m \\in \\mathbb{R}^D$ , $\\tau$ is the total frames of motion, and $D$ is the dimension of the 3D pose in each frame, encoder is used to encode motion $\\mathcal{M}$ to the latent embedding $z \\in \\mathbb{R}^{t \\times d}$ with a downsampling rate of $\\tau / t$ . The embedding $z$ is quantized into codes $c \\in \\mathcal{C}$ . Codebook $\\mathcal{C} = \\{\\gamma_k\\}_{k=1}^K$ contains $K$ number of codes. The nearest Euclidean distance between the embedding $\\mathbf{z}$ and the code of vector is computed by $\\hat{z}_i = \\mathrm{argmin}_j \\| \\mathbf{z} - \\mathcal{C}_j \\|_2^2$ . The loss function is defined as", + "bbox": [ + 212, + 344, + 787, + 494 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V Q} = \\left\\| \\operatorname {s g} (\\mathbf {z}) - \\mathbf {e} \\right\\| _ {2} ^ {2} + \\beta \\| \\mathbf {z} - \\operatorname {s g} (\\mathbf {e}) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 508, + 784, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}(\\cdot)$ is the stop-gradient operator, $\\beta$ refers the hyper-parameter for commitment loss. The loss function is optimized via a straight-through gradient estimator. We apply exponential moving average for codebooks update and codebook reset by following [45] [14] [29].", + "bbox": [ + 212, + 536, + 787, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Conditional Masked Self-attention Transformer", + "text_level": 1, + "bbox": [ + 215, + 619, + 651, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model employs a standard multi-layer transformer, whose inputs are the concatenation of the motion tokens $x_{1:t}$ from the tokenizer with $t$ as the sequence length, the text embedding $x_0$ from the pre-trained CLIP model [32], and the [END] token $x_{t + 1}$ that serves as the indicator of the motion's endpoint. The input tokens $x_{0:t + 1}$ are masked out strategically. Different from generative masked models, we do not replace the input tokens with [MASK] ones. Instead, we adopt causal attention mask $M$ as shown in Fig. 2 (b) to specify the attention relations among the input tokens. In particular, the token in the masked areas, indicated by $\\blacksquare$ , can be attended to itself, all the tokens on its left, and the unmasked tokens on its right. The unmasked token can be attended to by other unmasked tokens in both directions. In particular, we employ two causal masks: the unidirectional one, where only text token is unmasked and all other tokens are in the masked areas, and the bidirectional one, where text and [END]", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "E. Pinyoanuntapong et al.", + "bbox": [ + 271, + 114, + 446, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tokens are unmasked, while a random number of motion tokens are put into the masked areas. Consequently, Masked Self-attention Transformer retains the causal aspect of autoregressive motion generation while also being capable of conditioning on future tokens. The output of masked self-attention is as follows:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA t t e n t i o n = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} + \\mathrm {M}\\right) \\cdot V \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 218, + 785, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{Q}$ , $\\mathrm{K}$ , and $\\mathrm{V}$ , indicate queries, keys, and values respectively while $d_{k}$ represents the dimension of queries and keys. The self-attention mask $M \\in \\mathbb{R}^{(t + 1)\\times (t + 1)}$ is assigned to zero in the positions where attention is allowed, and to negative infinity otherwise. Adding negative infinity forces attention score to be zero after Softmax(·) operation. Therefore, bidirectional causal mask $M^{bc}$ can be written as", + "bbox": [ + 214, + 256, + 787, + 347 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i j} = \\left\\{ \\begin{array}{l l} 0, & \\text {w h e r e} (i \\geq j \\wedge i \\notin U) \\vee (j \\in U) \\\\ - \\infty , & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 356, + 785, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $i,j\\in [0,1,2,\\ldots ,t + 1]$ is the index of query $Q$ and key $K$ . $U = [u_{0},u_{1},\\dots ]$ contains the indices of unmasked tokens. The unidirectional causal mask $M^{uc}$ is a special case of the bidirectional one when $U = \\emptyset$ .", + "bbox": [ + 214, + 398, + 787, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Training: Hybrid Attention Masking", + "text_level": 1, + "bbox": [ + 214, + 467, + 563, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a discrete representation of the motion sequence $x_{1:t}$ , our model is trained to reconstruct the motion sequence, conditioned on the text token $x_0$ under both unidirectional and bidirectional causal masking strategies, i.e., $M^{uc}$ and $M^{bc}$ . The reconstruction probability of each motion token under each masking case is $p_{\\theta}(x_i \\mid M^{uc})$ and $p_{\\theta}(x_i \\mid M^{bc})$ , respectively. The training objective is to minimize the negative log-likelihood of the motion sequence prediction", + "bbox": [ + 214, + 492, + 787, + 583 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {h y b r i d}} = - \\underset {\\mathbf {X} \\in p (\\mathbf {X})} {\\mathbb {E}} \\left[ \\lambda \\sum_ {\\forall i \\in [ 1, t ]} \\log p _ {\\theta} \\left(x _ {i} \\mid M ^ {u c}\\right) + (1 - \\lambda) \\sum_ {\\forall i \\in [ 1, t ]} \\log p _ {\\theta} \\left(x _ {i} \\mid M ^ {b c}\\right) \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 604, + 785, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda$ is the probability of selecting unidirectional causal mask. Through experiments, we found $\\lambda = 0.5$ yields best performance. In addition, when bidirectional causal mask is selected, we randomly put $50\\% - 100\\%$ motion tokens in the masked areas.", + "bbox": [ + 214, + 656, + 787, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Inference: Cascaded Motion Decoding", + "text_level": 1, + "bbox": [ + 214, + 739, + 576, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To generate motion sequence during inference phase, dual-iteration cascaded decoding is introduced. In the first iteration, autoregressive decoding is applied, where the motion tokens are sequentially and stochastically sampled according to unidirectional prediction distribution $p_{\\theta}(x_i \\mid M^{uc})$ , concluding upon predicting the [END] token. Since autoregressive decoding could accumulate prediction", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BAMM", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/587bdad98e09ba99338a55c64f5f5647584f6ef15bde944017a33dcbc25559c6.jpg", + "image_caption": [ + "Fig. 3: Inference: Dual-iteration Cascaded Motion Decoding. In the first iteration, autoregressive decoding is applied by adopting unidirectional causal mask to generate coarse-grained motion and predict motion sequence length. In the second iteration, bidirectional autoregressive decoding is performed via bidirectional causal mask to removing and repredicting low-confidence motion tokens autoregressively." + ], + "image_footnote": [], + "bbox": [ + 271, + 146, + 725, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "errors from the previously generated tokens, the generated motion sequence is refined in the second iteration by masking out a subset of motion tokens and then resampling these masked tokens according to the bidirectional prediction distribution $p_{\\theta}(x_i \\mid M^{bc})$ . In this refinement iteration, [END] is positioned where it was predicted in the first iteration. The model can attend to unmasked tokens in all directions to re-predict low-confidence tokens based on rich surrounding context. The choice of masking strategies has an impact on the refinement gain, which is evaluated in Section 6.", + "bbox": [ + 212, + 349, + 787, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hybrid Classifier-free Guidance. At training time, we randomly drop textual tokens to teach the model to generate motion unconditionally. During inference, we apply classifier-free guidance (CFG) [20] for cascaded motion decoding. In particular, we generate the final motion sequence by a linear combination of the conditioned logits $\\ell_c$ and unconditional logits $\\ell_u$ with guidance scale $s$ as", + "bbox": [ + 212, + 470, + 787, + 559 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {g} = (1 + s) \\cdot \\ell_ {c} - s \\cdot \\ell_ {u}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 561, + 785, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We apply different CFG scale $s$ for each iteration during cascaded decoding. The effectiveness of CFG in each iteration is evaluated in Section 5.", + "bbox": [ + 212, + 583, + 785, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Residual Motion Refinement. To further enhance motion generation quality, the motion sequence yielded from cascaded decoding can be refined by another refinement transformer based on residual vector quantization (RVQ) [44]. By utilizing RVQ, the raw motion sequence is encoded into multiple token sequences in the latent space. Each token sequence is generated by a separate quantizer and each quantizer encodes the quantization error left by the previous quantizer. As a result, the token sequence from the first quantizer encodes the most of information of the original motion sequence. The rest of the token sequences from other quantizers only encode quantization errors. Through RVQ, information loss can be minimized during the embedding quantization process. As a result, our masked self-attention transformer is used to generate the most informative token sequence from the first quantizer. Using this sequence as input, another refinement transformer is trained to predict the remaining token sequences, which are merged into a single token sequence for final motion decoding. This RVQ-based refinement has been adopted by audio generation", + "bbox": [ + 212, + 613, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "E. Pinyoanuntapong et al.", + "bbox": [ + 271, + 114, + 447, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/942fb72941539d12adda89d4bf0d03b8e140b7a9d554dfb51921daabe696239f.jpg", + "image_caption": [ + "Fig. 4: Residual Motion Refinement. The residual vector quantization encodes the raw motion sequence into multiple token sequences in different colors (left). The base token sequence from the first vector quantizer is generated via cascaded decoding by masked self-attention transformer. The base token sequence is used as the input of the refinement transformer to predict the residual token sequences from other quantizers. The combined sequences are fed into tokenizer's decoder for motion generation. The refinement transformer shares the same architecture as the masked self-attention transformer with a full attention mask(right)." + ], + "image_footnote": [], + "bbox": [ + 235, + 143, + 789, + 247 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "models [4, 11, 40] and recently demonstrates its benefits in motion generation task [14].", + "bbox": [ + 212, + 386, + 785, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Motion Editability", + "text_level": 1, + "bbox": [ + 214, + 439, + 415, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The autoregressive approach naturally lacks the ability for temporal motion editing as it cannot leverage future motion tokens. In contrast, BAMM enables temporal motion editing through its bidirectional causal mask, which allows the model to access information from all directions of the conditioned tokens. Consequently, temporal motion editing becomes achievable by merely applying masked attention to the positions that require medications. We illustrate various editing tasks, i.e. motion inpainting (in-betweening), outpainting, prefix prediction, and suffix completion qualitatively and quantitively in Fig. 7 and Table 5. BAMM also allows us to generate the arbitrarily long motion sequence according to a sequence of text prompts, which is showcased in Supplementary Material.", + "bbox": [ + 212, + 465, + 787, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 641, + 375, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we provide a comprehensive evaluation along with empirical results for our proposed motion generation model, BAMM. In Section 4.1, we adhere to standard evaluation protocols on two datasets, demonstrating that our model outperforms current state-of-the-art methods both quantitatively and qualitatively. Furthermore, in Section 4.2, we showcase our predictive length and editing capabilities in comparison to state-of-the-art methods. The results reveal the robust performance of our model, displaying its effectiveness even in cases without prior information about motion length.", + "bbox": [ + 212, + 672, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We evaluated our model using the evaluation protocol proposed by [16] for text-driven motion generation with two datasets. KIT-ML dataset contains 3,911 motion sequences, each with one to four textual annotations,", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BAMM", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "amounting to a total of 6,278 annotations. These motion sequences are derived from the KIT and CMU motion [1] databases and have been adjusted to 12.5 FPS. The dataset is divided into training, validation, and testing sets, with proportions of $80\\%$ , $5\\%$ , and $15\\%$ respectively. HumanML3D dataset includes a wide variety of human activities, such as exercise and dancing, and consists of 14,616 motion sequences paired with 44,970 textual descriptions. These descriptions come from a vocabulary of 5,371 unique words. The motion sequences, sourced from AMASS [24] and HumanAct12 [17], have been standardized to 20 FPS and are limited to a maximum duration of 10 seconds, with the actual lengths varying between 2 to 10 seconds. Each sequence is accompanied by at least three descriptive annotations, averaging 12 words in length.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt the standard evaluation framework from T2M [16], employing pre-trained models that encode text and motion information to evaluate text and motion tokens in the embedding space. R-precision (Top-1, 2, 3 accuracy) measures how well the generated motions align with the text prompts, while Multimodal Distance (MM-Dist) quantifies the distance between generated and ground-truth motions in a shared feature space. Frechet Inception Distance (FID) assesses the statistical similarity between the feature distributions of generated and real motions. Additionally, we evaluate diversity (the average Euclidean distance between random motion pairs) and multimodality (the average variance across Euclidean distances between generated motion pairs for a single prompt) to capture the range of possible motion interpretations and their consistency with the text description.", + "bbox": [ + 212, + 313, + 787, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Comparison to State-of-the-art Approaches", + "text_level": 1, + "bbox": [ + 215, + 512, + 619, + 527 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/51e3fd425f56f2d73a7412299cc7e0febd0f7cbff430f3eabb270ce84eaf68b6.jpg", + "table_caption": [ + "Table 2: Comparison of text-conditional motion synthesis on HumanML3D [16] test set. We repeat the evaluation 20 times for each metric and report the average with $95\\%$ confidence interval. Red and Blue indicate the best and the second best result. Methods with gray highlight $\\S$ report motion generation results using the ground-truth motion length." + ], + "table_footnote": [], + "table_body": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Hier [13]0.301±.0020.425±.0020.552±.0046.523±.0245.012±.0188.332±.042-
TEMOS§ [27]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
TM2T [15]0.424±.0030.618±.0030.729±.0021.501±.0173.467±.0118.589±.0762.424±.093
T2M [16]0.455±.0030.636±.0030.736±.0021.087±.0213.347±.0089.175±.0832.219±.074
MDM§ [37]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse§ [46]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD§ [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
Fg-T2M§ [41]0.492±.0020.683±.0030.783±.0020.243±.0193.109±.0079.278±.0721.614±.049
M2DM§ [23]0.497±.0030.682±.0020.763±.0030.352±.0053.134±.0109.926±.0733.587±.072
T2M-GPT [45]0.491±.0030.680±.0030.775±.0020.116±.0043.118±.0119.761±.0811.856±.011
AttT2M [49]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
MMM§ [29]0.515±.0020.708±.0020.804±.0020.089±.0052.926±.0079.577±.0501.226±.035
MoMask§ [14]0.521±.0020.713±.0030.807±.0020.045±.0022.958±.008-1.241±.040
BAMM (ours)0.525±.0020.720±.0030.814±.0030.055±.0022.919±.0089.717±.0891.687±.051
BAMM§ (ours)0.522±.0030.715±.0030.808±.0030.055±.0022.936±.0779.636±.0091.732±.055
", + "bbox": [ + 253, + 632, + 743, + 810 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "E. Pinyoanuntapong et al.", + "bbox": [ + 271, + 114, + 446, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8de3d4f769cec2dda829f6583657935d4e440860b67637f0a3116dcf068b1e28.jpg", + "table_caption": [ + "Table 3: Comparison of text-conditional motion synthesis on KIT-ML [30] test set. We repeat the evaluation 20 times for each metric and report the average with $95\\%$ confidence interval. Red and Blue indicate the best and the second best result. Methods with gray highlight $\\S$ report ground-truth motion length for generation." + ], + "table_footnote": [], + "table_body": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Hier [13]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.072-
TEMOS$ [27]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
TM2T [15]0.280±.0050.463±.0060.587±.0053.599±.1534.591±.0269.473±.1173.292±.081
T2M [16]0.361±.0060.559±.0070.681±.0073.022±.1073.488±.02810.72±.1452.052±.107
MDM$ [37]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MotionDiffuse$ [46]0.417±.0040.621±.0040.739±.0041.954±.0642.958±.00511.10±.1430.730±.013
MLD$ [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
Fg-T2M$ [41]0.418±.0050.626±.0040.745±.0040.571±.0473.114±.01510.93±.0831.019±.029
M2DM$ [23]0.416±.0040.628±.0040.743±.0040.515±.0293.015±.01711.417±.973.325±.37
T2M-GPT [45]0.402±.0060.619±.0050.737±.0060.717±.0413.053±.02610.86±.0941.912±.036
AttT2M [49]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
MMM$ [29]0.404±.0050.621±.0050.744±.0040.316±.0282.977±.01910.910±.1011.232±.039
MoMask$ [14]0.433±.0070.656±.0050.781±.0050.204±.0112.779±.022-1.131±.043
BAMM (ours)0.438±.0090.661±.0090.788±.0050.183±.0132.723±.02611.008±.0941.609±.065
BAMM (ours)$0.436±.0070.660±.0060.791±.0050.200±.0112.714±.01610.914±.0971.517±.058
", + "bbox": [ + 251, + 204, + 741, + 383 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Results. We evaluate our model on the HumanML3D [16] and KIT-ML [30] datasets, reporting the results in Table 2 and 3, respectively, in comparison to state-of-the-art methods. Following the standard evaluation protocol from [16], we report the average of 20 generations with a $95\\%$ confidence interval. Our model consistently outperforms other methods in terms of R-precision, FID, and MM-Distance, while maintaining comparability in terms of Diversity and Multimodal Distance, which typically represent the trade-off between high quality and diversity. This suggests that our model generates very high-quality outputs while retaining a good degree of diversity. Moreover, whereas most models utilize the ground truth length for evaluation, our model supports both predicted length and takes length as an input, outperforming other methods in both scenarios. We further investigate the effect of a separate length estimator which can significantly impact performance in Section 4.2.", + "bbox": [ + 212, + 402, + 785, + 598 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 5 presents a qualitative comparison with T2M-GPT [45], MoMask [14], and MDM [37]. BAMM and T2M-GPT generate motion without requiring input length. Despite this, BAMM accurately generates motion from textual descriptions even without input length. We utilize a pre-trained length estimator from [16] for MoMask [14] and MDM [37]. Notably, BAMM generates motion accurately aligned with the provided text, whereas T2M-GPT and MoMask produce erroneous motion, and MDM generates entirely inaccurate motion. Additional visualizations for BAMM are shown in Fig. 6, indicating its ability to generate high detail such as complex trajectories and interactions with invisible objects. Moreover, Fig. 7 demonstrates BAMM's capability in various temporal editing tasks, including inpainting, outpainting, prefix, and suffix.", + "bbox": [ + 212, + 599, + 785, + 765 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Length Prediction and Editability", + "text_level": 1, + "bbox": [ + 214, + 785, + 535, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Predict Length vs ground truth length. In real-world scenarios, the ground truth length is often unknown. Many methods cannot predict length autonomously.", + "bbox": [ + 214, + 809, + 795, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BAMM", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fb9d9b4266dfb039b411d8c9130d9bea94029c973f00f07f6134f9aaa5e28663.jpg", + "image_caption": [ + "Fig. 5: Visualization comparison of textual to motion to state-of-the-art methods. BAMM and T2M-GPT do not require motion length as an input. We use a pre-trained length estimator from [16] for MoMask and MDM. BAMM generates higher quality and is more correlated with textual descriptions." + ], + "image_footnote": [], + "bbox": [ + 282, + 132, + 723, + 318 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f9a8183e46fdfe1d46c91fee66ac7289637c9c226c43fe5717c09e1ebdec8a1b.jpg", + "image_caption": [ + "Fig. 6: Visualization of text-to-motion generation by BAMM. BAMM can generate high-quality motion with complex descriptions, such as intricate trajectories and interactions with invisible objects." + ], + "image_footnote": [], + "bbox": [ + 277, + 397, + 728, + 494 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Consequently, they evaluate using ground truth length, indicated by gray highlight $\\S$ in Table 2 and Table 3. We illustrate the impact of inaccurate length prediction by comparing our model to state-of-the-art denoising models that require motion length as an input, such as MoMask and MMM. In particular, both models take the estimated length obtained from a standalone length predictor pretrained on the HumanML3D dataset [16]. Table 4 reveals that the predicted length significantly worsens MoMask's FID score, from 0.045 to 0.090, while MMM's R-precision Top1 drops from 0.515 to 0.504. In contrast, our BAMM not only maintains the best performance but also outperforms MMM and MoMask in both scenarios.", + "bbox": [ + 214, + 560, + 787, + 709 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/868598c7e15351d35a88dcfbe03a8144e8224e5ca365941b89ed96528f261ed2.jpg", + "table_caption": [ + "Table 4: Comparison of text-conditional motion synthesis using predicted and ground truth length on HumanML3D [16] dataset" + ], + "table_footnote": [], + "table_body": "
LengthMethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
GroundMMM0.5150.7080.8040.0892.9269.5771.226
TruthMoMask0.5210.7130.8070.0452.958-1.241
BAMM (Ours)0.5220.7150.8080.0552.9369.6361.732
PredictedMMM0.5040.6960.7940.0802.9989.4111.164
MoMask0.5220.7150.8110.0902.9459.6471.239
BAMM (Ours)0.5250.7200.8140.0552.9199.7171.687
", + "bbox": [ + 305, + 768, + 683, + 844 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "E. Pinyoanuntapong et al.", + "bbox": [ + 271, + 114, + 444, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9294446434a39066cbf67430dd738105598b9d0308a9f9dde20d04c5e6ab1cde.jpg", + "image_caption": [ + "Fig. 7: Visualization of temporal editing tasks, inpainting (in-betweening), outpainting, prefix, and suffix where blue indicates conditioned motion and red refers to generated parts." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 781, + 246 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Editability. We conduct experiments in four temporal editing tasks, namely inpainting (in-betweening), outpainting, prefix, and suffix, comparing them with MDM [37] and MoMask [14] on the HumanML3D dataset. Inpainting is evaluated by generating $50\\%$ of the motion sequence given the first and last $25\\%$ . Outpainting is its opposite counterpart. Prefix is conditioned by the first $50\\%$ of the ground truth motion and generates the remaining portion. Similarly, suffix operates in the reverse manner. Note that editing tasks require prior knowledge of conditioned motion and motion length, for which MDM and MoMask are specifically designed. Despite this, Table 5 shows that our model significantly outperforms MDM in all editing tasks and surpasses MoMask in outpainting, while performing equally well in the other tasks.", + "bbox": [ + 212, + 311, + 787, + 477 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7ac77609d2847caaaa2afabb12182bde39cd2b8e4482a1e29c996a84a4cab61c.jpg", + "table_caption": [ + "Table 5: Evaluation on temporal editing tasks on HumanML3D [16] dataset" + ], + "table_footnote": [], + "table_body": "
TasksMethodsR-Precision ↑FID ↓ MM-Dist ↓ Diversity ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Temporal Inpainting (In-betweenging)MDM0.3910.5780.6922.3623.8598.014
MoMask0.5340.7270.820.042.8789.64
BAMM0.5350.7290.8210.0562.8639.629
Temporal OutpaintingMDM0.4150.6130.7272.0573.6198.199
MoMask0.5310.7260.8180.0572.8899.619
BAMM0.5350.730.8220.0562.8569.659
PrefixMDM0.420.6130.7251.463.5638.312
MoMask0.5360.730.8220.062.8759.607
BAMM0.5320.7270.8210.0582.8689.612
SuffixMDM0.4030.5970.7112.5623.7318.088
MoMask0.5320.7260.8190.0522.8819.659
BAMM0.5270.720.8140.052.8919.721
", + "bbox": [ + 320, + 512, + 671, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 671, + 401, + 690 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the ablation study, we study the impacts of the adaptive classifier-free guidance scales in each iteration of the Masked Self-attention Transformer, the masking strategy, and the number of iterations, as shown in Table 6.", + "bbox": [ + 212, + 704, + 785, + 748 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Adaptive Classifier-Free Guidance Scales (CFG): We conducted experiments with different CFG values for both the first and second iterations and found that CFG=4 in the first iteration and CFG=3 in the second iteration works best. Although CFG=5 in the first iteration yields a better FID score, the other scores are worse. Mask Strategy: The first strategy, \"50% of low confidence\", applies mask attention on 50% of the lowest confident positions from the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BAMM", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/18d0fe4153160804ea94ed98cbf326533e2d26b26a5598c70b3e54322bb3aca4.jpg", + "table_caption": [ + "Table 6: Ablation Study" + ], + "table_footnote": [], + "table_body": "
AblationsTypesR-Precision ↑FID ↓ MM-Dist ↓ Diversity ↑ MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
CFG of 1st iteration1st iter2nd iter
CFG=2CFG=30.5170.7110.8050.1052.9569.841.907
CFG=3CFG=30.5210.7140.8080.072.9449.7771.766
CFG=4CFG=30.5250.720.8140.0552.9199.7171.687
CFG=5CFG=30.5220.7160.810.0522.9279.6471.691
CFG=6CFG=30.520.7130.810.062.949.6211.697
CFG of 2nd iterationCFG=4CFG=10.5210.7160.810.0582.9439.7431.744
CFG=4CFG=20.5240.7190.8140.0572.9249.7521.698
CFG=4CFG=30.5250.720.8140.0552.9199.7171.687
CFG=4CFG=40.5220.7190.8120.0562.9249.6911.698
CFG=4CFG=50.5210.7170.8120.0652.9319.6381.735
Mask Strategy50% of low confidence confidence < .50.5250.720.8130.0652.9219.7321.67
suffix0.5250.7180.810.0642.9239.7651.656
%2=00.5190.7150.810.0522.9439.6831.841
0.5250.720.8140.0552.9199.7171.687
# of iterations1 iteration0.5240.7180.8120.0642.9269.7201.644
2 iterations0.5250.7200.8140.0552.9199.7171.687
3 iterations0.5250.7190.8140.0552.9179.7271.69
", + "bbox": [ + 267, + 159, + 725, + 345 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "first iteration. Similarly, \"confidence $< .5$ \" uses low confidence but sets a threshold, masking out positions where confidence is below 0.5. Specifically, the former strictly masks out $50\\%$ of the total token sequence while the latter masks the token with a confidence lower than .5. \"Suffix\" strategy masks out the first $50\\%$ of the token sequence, utilizing the remaining $50\\%$ as condition tokens. Lastly, $\\% 2 = 0$ \"strategy masks every other token. BAMM produces decent performances across the tested masking strategies while the simple masking of every other token works the best. Number of Iterations: We conducted experiments with one to three iterations. \"1 iteration\" refers to the first iteration, as described in 3.4, which utilizes a unidirectional causal mask. Similarly, \"2 iterations\" involves applying bidirectional causal mask to the Masked Self-attention Transformer to re-predict the tokens from the first iteration, as illustrated in 3. In \"3 iterations\", we repeat the second iteration (bidirectional causal mask) but apply masks to $1/3$ of the sequence to re-predict the motion tokens instead. The experiments indicate that \"2 iterations\" clearly demonstrate improvement over one iteration, while \"3 iterations\" do not significantly improve upon \"2 iterations\". This suggests that \"2 iterations\" are sufficient.", + "bbox": [ + 212, + 367, + 787, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 652, + 359, + 669 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce the Bidirectional Autoregressive Motion Model (BAMM), a novel framework for text-to-motion generation. BAMM combines a motion tokenizer, which encodes 3D human motion into discrete latent tokens, with a masked self-attention transformer that autoregressively predicts the masked tokens through a masked casual attention approach. BAMM integrates of generative masked and autoregressive modeling into an unified framework, This features allows it to understand complex motion relationships and precisely map text inputs to high-quality motion outputs with adaptively adjusted sequence lengths. Our extensive testing on HumanML3D and KIT-ML datasets confirms BAMM's superiority in both qualitative and quantitative evaluations over existing methods.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "E. Pinyoanuntapong et al.", + "bbox": [ + 271, + 114, + 446, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Cmu graphics lab motion capture database, http://mocap.cs.cmu.edu/, accessed: 2022-11-11 10", + "2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728 (2019). https://doi.org/10.1109/3DV.2019.000844", + "3. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems 34, 17981-17993 (2021) 4", + "4. Borsos, Z., Sharifi, M., Vincent, D., Kharitonov, E., Zeghidour, N., Tagliasacchi, M.: Soundstorm: Efficient parallel audio generation. ArXiv abs/2305.09636 (2023), https://api-semanticscholar.org/CorpusID:258715176 9", + "5. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T.J., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. ArXiv abs/2005.14165 (2020), https://api_semanticscholar.org/CorpusID:218971783 2, 5", + "6. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M., Murphy, K.P., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D.: Muse: Text-to-image generation via masked generative transformers. ArXiv abs/2301.00704 (2023), https://api(semanticscholar.org/CorpusID:255372955 4", + "7. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11305-11315 (2022), https://apisemantic scholar.org/CorpusID:246680316 4", + "8. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, J., Yu, G.: Executing your commands via motion diffusion in latent space. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 18000-18010 (2022), https://api-semanticscholar.org/CorpusID:254408910 4, 10, 11", + "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019), https://apisemantic scholar.org/CorpusID:52967399 4", + "0. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. ArXiv abs/2204.14217 (2022), https://api(semanticscholar.org/CorpusID:248476190 4", + "1. Du, Z., Qian, Y., Liu, X., Ding, M., Qiu, J., Yang, Z., Tang, J.: Glm: General language model pretraining with autoregressive blank infilling. In: Annual Meeting of the Association for Computational Linguistics (2021), https://apisemantic scholar.org/CorpusID:247519241 5, 9", + "2. Ghazvininejad, M., Levy, O., Liu, Y., Zettlemoyer, L.: Mask-predict: Parallel decoding of conditional masked language models. In: Conference on Empirical Methods in Natural Language Processing (2019), https://api_semanticscholar.org/ CorpusID:202538740 4", + "3. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 1376-1386 (2021), https://apisemantic scholar.org/CorpusID:232404671 10, 11" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BAMM", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions (2023) 2, 4, 6, 9, 10, 11, 13", + "15. Guo, C., Xuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. ArXiv abs/2207.01696 (2022), https://api_semanticscholar.org/CorpusID: 250280248 10, 11", + "16. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5142-5151 (2022). https://doi.org/10.1109/CVPR52688.2022.005093, 4, 5, 9, 10, 11, 12, 13, 2", + "17. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. Proceedings of the 28th ACM International Conference on Multimedia (2020), https://apisemantic scholar.org/CorpusID:220870974 10", + "18. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A.A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models. ArXiv abs/2210.02303 (2022), https://api_semanticscholar.org/CorpusID:252715883 4", + "19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. ArXiv abs/2006.11239 (2020), https://apisemantic scholar.org/CorpusID: 219955663 4", + "20. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 8", + "21. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. ArXiv abs/2306.14795 (2023), https://api.sementicscholar.org/CorpusID:259262201 2, 5", + "22. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI Conference on Artificial Intelligence (2022), https://api-semanticscholar.org/CorpusID:251979380 2, 4", + "23. Kong, H., Gong, K., Lian, D., Mi, M.B., Wang, X.: Priority-centric human motion generation in discrete latent space. ArXiv abs/2308.14480 (2023), https://api.sementicscholar.org/CorpusID:261245369 10, 11", + "24. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 5441-5450 (2019), https://apisemanticscholar.org/CorpusID:102351100 10", + "25. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning (2021), https://api-semanticscholar.org/CorpusID:245335086 4", + "26. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning. ArXiv abs/1711.00937 (2017), https://api-semanticscholar.org/CorpusID:20282961 3, 6", + "27. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. ArXiv abs/2204.14109 (2022), https://api.sementicscholar.org/CorpusID:248476220 4, 10, 11", + "28. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. ArXiv abs/2305.00976 (2023), https://api_semanticscholar.org/CorpusID:258436810 4" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "E. Pinyoanuntapong et al.", + "bbox": [ + 271, + 114, + 444, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Pinyoanuntapong, E., Wang, P., Lee, M., Chen, C.: Mmm: Generative masked motion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 2, 4, 6, 10, 11", + "30. Plappert, M., Mandery, C., Asfour, T.: The KIT motion-language dataset. Big Data 4(4), 236-252 (dec 2016). https://doi.org/10.1089/big.2016.0028, http://dx.doi.org/10.1089/big.2016.0028 4, 11, 5", + "31. Qian, L., Zhou, H., Bao, Y., Wang, M., Qiu, L., Zhang, W., Yu, Y., Li, L.: Glancing transformer for non-autoregressive neural machine translation. In: Annual Meeting of the Association for Computational Linguistics (2020), https://api_semanticscholar.org/CorpusID:221150562 4", + "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021), https://api_semanticscholar.org/CorpusID: 231591445 4, 6", + "33. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. ArXiv abs/2205.11487 (2022), https://api_semanticscholar.org/CorpusID:248986576 4", + "34. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data. ArXiv abs/2209.14792 (2022), https://api_semanticscholar.org/CorpusID:252595919 4", + "35. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ArXiv abs/2010.02502 (2020), https://api.sementicscholar.org/CorpusID:222140788 4", + "36. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision (2022), https://api-semanticscholar.org/CorpusID:247450907 4", + "37. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. ArXiv abs/2209.14916 (2022), https://api.sementicscholar.org/CorpusID:252595883 2, 4, 10, 11, 13", + "38. Vaswani, A., Shazeer, N.M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:137564895", + "39. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. ArXiv abs/2210.02399 (2022), https://apisemantic scholar.org/CorpusID:252715594 4", + "40. Wang, C., Chen, S., Wu, Y., Zhang, Z.H., Zhou, L., Liu, S., Chen, Z., Liu, Y., Wang, H., Li, J., He, L., Zhao, S., Wei, F.: Neural codec language models are zero-shot text to speech synthesizers. ArXiv abs/2301.02111 (2023), https://api.sementicscholar.org/CorpusID:255440307 9", + "41. Wang, Y., Leng, Z., Li, F.W.B., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. ArXiv abs/2309.06284 (2023), https://api.sementicscholar.org/CorpusID:261697123 2, 10, 11", + "42. Yan, S., Liu, Y., Wang, H., Du, X., Liu, M., Liu, H.: Cross-modal retrieval for motion and text via doptriple loss (2023), https://api-semanticscholar.org/CorpusID:263610212 4" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BAMM", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., Hutchinson, B.C., Han, W., Parekh, Z., Li, X., Zhang, H., Baldridge, J., Wu, Y.: Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res. 2022 (2022), https://api_semanticscholar.org/CorpusID:249926846 4", + "44. Zeghidour, N., Luebs, A., Omran, A., Skoglund, J., Tagliasacchi, M.: Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30, 495-507 (2021), https://api-semanticscholar.org/CorpusID:236149944 8", + "45. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: Generating human motion from textual descriptions with discrete representations. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14730-14740 (2023), https://api_semanticscholar.org/CorpusID:255942203 2, 5, 6, 10, 11", + "46. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. ArXiv abs/2208.15001 (2022), https://api_semanticscholar.org/CorpusID: 251953565 2, 4, 10, 11", + "47. Zhang, Z., Ma, J., Zhou, C., Men, R., Li, Z., Ding, M., Tang, J., Zhou, J., Yang, H.: M6-ufc: Unifying multi-modal controls for conditional image synthesis via non-autoregressive generative transformers (2021), https://api_semanticscholar.org/CorpusID:237204528 4", + "48. Zhang, Z., Ma, J., Zhou, C., Men, R., Li, Z., Ding, M., Tang, J., Zhou, J., Yang, H.: Ufc-bert: Unifying multi-modal controls for conditional image synthesis. In: Neural Information Processing Systems (2021), https://apisemantic scholar.org/CorpusID:235253928 4", + "49. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. ArXiv abs/2309.00796 (2023), https://api.sementicscholar.org/CorpusID:261530775 2, 5, 10, 11" + ], + "bbox": [ + 215, + 146, + 785, + 550 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "E. Pinyoanuntapong et al.", + "bbox": [ + 271, + 114, + 444, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/BAMM_ Bidirectional Autoregressive Motion Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_model.json b/2024/BAMM_ Bidirectional Autoregressive Motion Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40d331f5aa922bcad8ae4cdf6b0aad57b989d6e2 --- /dev/null +++ b/2024/BAMM_ Bidirectional Autoregressive Motion Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_model.json @@ -0,0 +1,2128 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.244, + 0.141, + 0.761, + 0.184 + ], + "angle": 0, + "content": "BAMM: Bidirectional Autoregressive Motion Model" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.213, + 0.765, + 0.245 + ], + "angle": 0, + "content": "Ekkasit Pinyoanuntapong1, Muhammad Usama Saleem1, Pu Wang1, Minwoo Lee1, Srijan Das1, and Chen Chen2" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.255, + 0.757, + 0.282 + ], + "angle": 0, + "content": "1 University of North Carolina at Charlotte {epinyoan,msaleem2, Pu.Wang, minwoo.lee, sdas24}@uncc.edu" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.283, + 0.694, + 0.298 + ], + "angle": 0, + "content": "2 University of Central Florida chen.chen@crcv.ucf.edu" + }, + { + "type": "list", + "bbox": [ + 0.243, + 0.255, + 0.757, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.323, + 0.348, + 0.332 + ], + "angle": 0, + "content": "a) Motion Length Prediction" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.333, + 0.356, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.323, + 0.513, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.346, + 0.512, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.323, + 0.706, + 0.333 + ], + "angle": 0, + "content": "c) Motion Editing" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.334, + 0.782, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Fig. 1: (a) Motion Length Prediction: Text-to-motion models often require specific input lengths, making them sensitive to motion generation. In contrast, BAMM automatically predicts the end of the motion, thus avoiding reliance on inaccurate motion length estimations. (b) High-quality Text-to-Motion: BAMM generates natural human movements precisely aligned with detailed textual descriptions. (c) Motion Editing: BAMM is capable of multiple editing tasks, such as inpainting (as demonstrated), outpainting, prefix prediction, suffix completion, and arbitrarily long motion sequence synthesis." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.563, + 0.741, + 0.84 + ], + "angle": 0, + "content": "Abstract. Generating human motion from text has been dominated by denoising motion models either through diffusion or generative masking process. However, these models face great limitations in usability by requiring prior knowledge of the motion length. Conversely, autoregressive motion models address this limitation by adaptively predicting motion endpoints, at the cost of degraded generation quality and editing capabilities. To address these challenges, we propose Bidirectional Autoregressive Motion Model (BAMM), a novel text-to-motion generation framework. BAMM consists of two key components: (1) a motion tokenizer that transforms 3D human motion into discrete tokens in latent space, and (2) a masked self-attention transformer that autoregressively predicts randomly masked tokens via a hybrid attention masking strategy. By unifying generative masked modeling and autoregressive modeling, BAMM captures rich and bidirectional dependencies among motion tokens, while learning the probabilistic mapping from textual inputs to motion outputs with dynamically-adjusted motion sequence length. This feature enables BAMM to simultaneously achieving high-quality motion generation with enhanced usability and built-in motion editability. Extensive experiments on HumanML3D and KIT-ML datasets demonstrate that BAMM surpasses current state-of-the-art methods in both" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.447, + 0.129 + ], + "angle": 0, + "content": "E. Pinyoanuntapong et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.74, + 0.176 + ], + "angle": 0, + "content": "qualitative and quantitative measures. Our project page is available at https://exitudio.github.io/BAMM-page" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.189, + 0.741, + 0.217 + ], + "angle": 0, + "content": "Keywords: Text to Motion \\(\\cdot\\) Autoregressive Motion Model \\(\\cdot\\) Generative Masked Motion Model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.249, + 0.376, + 0.266 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Text-to-motion generation is a promising interdisciplinary field that uses natural language to generate 3D human movements. This emerging field holds vast potential to transform animation, gaming, filming, and VR/AR/MR domains by enabling easy and intuitive creation of 3D assets through user-friendly textual inputs. However, bridging the semantic gap between textual descriptions and intricate motion sequences presents a significant challenge. To address this challenge, recent efforts have been focusing on two methods: (1) conditional denoising motion model and (2) conditional autoregressive motion model. Both methods can greatly improve motion generation quality by learning the probabilistic distribution of motion sequences, conditioned on the textural descriptors. However, both methods face fundamental limitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Conditional denoising motion models are trained to restore corrupted motion sequences to their original state, guided by textual prompts. These models operate through two primary mechanisms: diffusion and generative masking. Diffusion models apply structured Gaussian noise to the original motion data for corruption [22,37,41,46], whereas generative masked models [14,29] corrupt motion sequence by substituting selected motion tokens with [MASK] tokens. The denoising process, followed by motion corruption procedure, considers motion tokens from both directions, effectively capturing the intricate dependencies among tokens. This leads to enhanced motion generation quality. Furthermore, these models inherently ease the motion editing tasks. By selectively corrupting and recovering motion tokens in areas needing modifications, they ensure seamless transitions between edited and unedited segments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.842 + ], + "angle": 0, + "content": "While denoising motion models excel in generation quality and editability, a significant limitation of these models is their usability because these models depend on prior knowledge of motion length for each text prompt, a requirement that proves impractical in real-world scenarios. Utilizing incorrect motion lengths can result in a significant decline in generation quality. To address this fundamental limitation, conditional motion autoregressive models emerge as a solution, capable of simultaneously predicting the sequence length and content of generated motions. Inspired by large language models like GPT [5], motion autoregressive models sequentially predict one motion token at a time from left to right until the [END] token is predicted, guided by the textual description [21, 45, 49]. As a result, the generated motions are not only well aligned with the text inputs but also appropriately scaled in duration. However, the sequential token decoding of autoregressive models cannot fully capture the dependencies between the motion tokens, potentially compromising generation quality, and complicating" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAMM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "the editing process, because edited parts need to be conditioned on the unedited parts from both directions to ensure overall continuity and coherence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.481 + ], + "angle": 0, + "content": "As previously highlighted, existing text-driven motion generation models encounter a comprise among usability, quality of generation, and editability. To address this challenge, we propose Bidirectional Autoregressive Motion Model (BAMM), a novel text-to-motion generation framework. It consists of two pivotal components: a motion tokenizer and a conditional masked self-attention transformer. In a two-stage training paradigm, the motion tokenizer performs initial training based on Vector Quantized Variational Autoencoders (VQ-VAE) [26]. This tokenizer encodes the raw motion sequence into discrete motion tokens within the latent space leveraging learned codebooks. In the subsequent phase, motion tokens are randomly masked out. A conditional masked self-attention transformer is then trained to autoregressively predict these masked tokens, adopting the causal attention masking strategy. This strategy is largely departing from the traditional [MASK] token replacement approach in generative masked models. In particular, it does not substitute the input motion tokens with [MASK] tokens. Instead, it adjusts the attention score matrix according to both unidirectional and bidirectional causal masks. The unidirectional causal mask enables adaptive prediction of [END] token based on text prompts, while bidirectional causal mask forces the model to predict the next motion token not only based on past tokens but also conditioned on future unmasked tokens. This facilitates bidirectional autoregressive training for enhanced predictive capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.634 + ], + "angle": 0, + "content": "Table 1: Comparison of quality and capability of generation on text-to-motion to state-of-the-art models on the largest text-to-motion dataset [16]. \\(\\checkmark\\) means capability while \\(X\\) is not. \"Predict Length\" denotes the ability to generate motion without prior knowledge of motion length. \"Input Length\" refers to the ability to take input length as a constraint, while \"Edit\" indicates motion editability. Since MMM and MoMask require ground-truth motion length as input, we use predicted motion length from pretrained length estimator by [16]. The lowest FID score means the best overall quality of the generated motion, ensuring that its authenticity and naturalness is very close to the ground-truth human movements. High R-precision and low MM-dist means accurate alignment between the generated motion and the text prompts." + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.646, + 0.688, + 0.719 + ], + "angle": 0, + "content": "
MethodsTop-1 ↑FID ↓MM-Dist ↓Predict LengthInput LengthEdit
T2M-GPT0.4910.1163.118✓✗✗
AttT2M0.4990.1123.038✓✗✗
MMM0.5040.0802.998✗✓✓
MoMask0.5220.0902.945✗✓✓
BAMM (Ours)0.5250.0552.919✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "By unifying masked and autoregressive prediction during training, BAMM captures rich and bidirectional dependencies among motion tokens, while learning a direct probabilistic mapping from textual inputs to motion outputs with dynamically-adjusted motion sequence length. Leveraging such unique feature, we propose cascaded motion generation during inference, where BAMM first leverages unidirectional autoregressive decoding to implicitly predict motion se" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.447, + 0.129 + ], + "angle": 0, + "content": "E. Pinyoanuntapong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "quence length and generate coarse-grained motion sequence. Such motion sequence is then refined by masking and regenerating a portion of motion tokens in a bidirectional autoregressive fashion. This feature allows BAMM to achieve high-quality motion generation with high usability. Moreover, BAMM naturally supports zero-shot motion editing without specially being trained for such task. By treating the masked motion tokens as the contents that need editing, BAMM can predict the masked tokens based on the surrounding context and the text description. Our main contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.282, + 0.784, + 0.355 + ], + "angle": 0, + "content": "- We introduce the bidirectional autoregressive motion model, which is a novel text-to-motion generation framework. It effectively harnesses the complementary benefits of denoising and autoregressive models, thus simultaneously achieving high-quality motion generation with enhanced usability and innate motion editability, as showcased in Fig. 1 and Table 1." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.358, + 0.784, + 0.402 + ], + "angle": 0, + "content": "- We demonstrate that our model outperforms current state-of-the-art methods qualitatively and quantitatively on two standard text-to-motion generation datasets, HumanML3D [16] and KIT-ML [30]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.404, + 0.784, + 0.462 + ], + "angle": 0, + "content": "- We showcase that our model supports a variety of motion editing tasks in the zero-shot manner without specially training for these tasks, including motion inpainting, outpainting, prefix prediction, suffix completion, and long sequence generation." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.282, + 0.784, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.388, + 0.505 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Motion Synthesis with Latent Space Alignment. Early text-to-motion approaches typically adopt a two-stage scheme: learning separate latent representations for text and motion sequences, followed by latent space alignment using distance losses like cosine similarity or KL divergence [2,16,27,28,36,42]. For example, Language2Pose [2] aimed to establish a shared latent space for both language descriptions and motion sequences. MotionCLIP [36] simplified this concept by incorporating stylization and diversity techniques, leveraging the pre-trained text-image latent space of the CLIP model [32]. However, this strategy inherently struggles with generating high-fidelity motions due to the difficulty of perfectly aligning these inherently disparate latent spaces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Conditional Denoising Motion Model. Inspired by the success of denoising diffusion models (DDMs) [19,35] in text-to-image and text-to-video generation [18, 25, 33, 34], diffusion models have been applied for text-to-motion generation. MDM [37], MotionDiffuse [46], MLD [8], and FLAME [22] are the representative examples. Meanwhile, BERT-type masked generative models have shown their success in both text generation tasks, e.g., Q&A and language translations [9,12,31], and text-to-image synthesis [6,7,10,39,43,47,48]. Following this trend, MMM [29] and MoMask [14] are recent attempts to propose conditional masked motion modeling to enable high-fidelity motion synthesis that is precisely aligned with text prompts. Both diffusion and generative masked modeling share the same principle: \"denoising\" corrupted data [3]. Meanwhile, they also face the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAMM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.266 + ], + "angle": 0, + "content": "same usability limitation: they require prior knowledge of the motion sequence length for the denoising process. This length information is obtained either by intuitive guess from users or via a separately pre-trained predictor that predicts the motion length distribution based on the input text prompts [16]. Both approaches, however, lead to significant motion quality degradation as demonstrated in the experiments section. This is because the distributions of motion length and motion contents are inherently coupled, which need to be jointly aligned with the textual context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.789, + 0.451 + ], + "angle": 0, + "content": "Conditional Autoregressive Motion Model. The autoregressive motion models, such as T2M-GPT [45], AttT2M [49], and MotionGPT [21], can effectively address the usability challenge faced by denoising models because they follow the GPT-type training and inference [5] to implicitly predict motion sequence length by generating the [END] token conditioned on both previously generated motion tokens and text inputs. However, the limitation of these models lies in their use of causal attention for unidirectional and sequential motion token prediction. This practice not only hinders model's motion editability but also jeopardizes motion generation quality. To address this limitation, we propose the first bidirectional autoregressive modeling approach for human motion generation, which draws inspiration from the self-attention masking employed by large language model pretraining [11,38]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.476, + 0.331, + 0.491 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.519, + 0.761, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Fig.2: Overall architecture of BAMM. (a) Motion Tokenizer encodes the raw motion sequence into discrete motion tokens according to a learned codebook. (b) Masked Self-attention Transformer learns to sequentially predict next tokens conditioned on text embedding from CLIP model and future unmasked tokens. Masked self-attention mechanism unifies autoregressive model and generative masked motion via bidirectional and unidirectional causal masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our objective is to create a text-to-motion synthesis framework that simultaneously achieves high-quality motion generation with enhanced usability and innate motion editability. Towards this goal, our framework, as illustrated in Fig. 2, consists of two key components: motion tokenizer that compresses and converts raw 3D human motion into a sequence of discrete motion tokens in the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.447, + 0.129 + ], + "angle": 0, + "content": "E. Pinyoanuntapong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "latent space (Section 3.1) and conditional masked self-attention transformer that leverages unidirectional and bidirectional casual masks to integrate autoregressive model and generative masked model into a unified framework (Section 3.2). The training procedure follows a hybrid attention masking strategy, where the two causal masks are applied randomly and the model is forced to reconstruct the motion sequence under both cases (Section 3.3). The cascaded motion decoding is introduced for motion generation during the inference phase. It uses unidirectional autoregressive decoding to jointly predict motion sequence length and its contents, which are refined via bidirectional autoregressive decoding (Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.32, + 0.409, + 0.334 + ], + "angle": 0, + "content": "3.1 Motion Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.789, + 0.496 + ], + "angle": 0, + "content": "The objective of this stage is to learn the discrete representation of motion by quantizing the embedding \\(z\\) from the output of the encoder into codebook \\(\\mathcal{C}\\). We first pretrain a motion tokenizer based on VQ-VAE [26]. In particular, given a motion sequence \\(\\mathcal{M} = [m_1, m_2, m_3, \\dots, m_\\tau]\\) where \\(m \\in \\mathbb{R}^D\\), \\(\\tau\\) is the total frames of motion, and \\(D\\) is the dimension of the 3D pose in each frame, encoder is used to encode motion \\(\\mathcal{M}\\) to the latent embedding \\(z \\in \\mathbb{R}^{t \\times d}\\) with a downsampling rate of \\(\\tau / t\\). The embedding \\(z\\) is quantized into codes \\(c \\in \\mathcal{C}\\). Codebook \\(\\mathcal{C} = \\{\\gamma_k\\}_{k=1}^K\\) contains \\(K\\) number of codes. The nearest Euclidean distance between the embedding \\(\\mathbf{z}\\) and the code of vector is computed by \\(\\hat{z}_i = \\mathrm{argmin}_j \\| \\mathbf{z} - \\mathcal{C}_j \\|_2^2\\). The loss function is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.509, + 0.785, + 0.526 + ], + "angle": 0, + "content": "\\[\nL _ {V Q} = \\left\\| \\operatorname {s g} (\\mathbf {z}) - \\mathbf {e} \\right\\| _ {2} ^ {2} + \\beta \\| \\mathbf {z} - \\operatorname {s g} (\\mathbf {e}) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.599 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}(\\cdot)\\) is the stop-gradient operator, \\(\\beta\\) refers the hyper-parameter for commitment loss. The loss function is optimized via a straight-through gradient estimator. We apply exponential moving average for codebooks update and codebook reset by following [45] [14] [29]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.652, + 0.634 + ], + "angle": 0, + "content": "3.2 Conditional Masked Self-attention Transformer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Our model employs a standard multi-layer transformer, whose inputs are the concatenation of the motion tokens \\( x_{1:t} \\) from the tokenizer with \\( t \\) as the sequence length, the text embedding \\( x_0 \\) from the pre-trained CLIP model [32], and the [END] token \\( x_{t + 1} \\) that serves as the indicator of the motion's endpoint. The input tokens \\( x_{0:t + 1} \\) are masked out strategically. Different from generative masked models, we do not replace the input tokens with [MASK] ones. Instead, we adopt causal attention mask \\( M \\) as shown in Fig. 2 (b) to specify the attention relations among the input tokens. In particular, the token in the masked areas, indicated by \\( \\blacksquare \\), can be attended to itself, all the tokens on its left, and the unmasked tokens on its right. The unmasked token can be attended to by other unmasked tokens in both directions. In particular, we employ two causal masks: the unidirectional one, where only text token is unmasked and all other tokens are in the masked areas, and the bidirectional one, where text and [END]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAMM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "tokens are unmasked, while a random number of motion tokens are put into the masked areas. Consequently, Masked Self-attention Transformer retains the causal aspect of autoregressive motion generation while also being capable of conditioning on future tokens. The output of masked self-attention is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.219, + 0.786, + 0.253 + ], + "angle": 0, + "content": "\\[\nA t t e n t i o n = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} + \\mathrm {M}\\right) \\cdot V \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.257, + 0.788, + 0.348 + ], + "angle": 0, + "content": "where \\( \\mathrm{Q} \\), \\( \\mathrm{K} \\), and \\( \\mathrm{V} \\), indicate queries, keys, and values respectively while \\( d_{k} \\) represents the dimension of queries and keys. The self-attention mask \\( M \\in \\mathbb{R}^{(t + 1)\\times (t + 1)} \\) is assigned to zero in the positions where attention is allowed, and to negative infinity otherwise. Adding negative infinity forces attention score to be zero after Softmax(·) operation. Therefore, bidirectional causal mask \\( M^{bc} \\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.357, + 0.786, + 0.389 + ], + "angle": 0, + "content": "\\[\nM _ {i j} = \\left\\{ \\begin{array}{l l} 0, & \\text {w h e r e} (i \\geq j \\wedge i \\notin U) \\vee (j \\in U) \\\\ - \\infty , & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.399, + 0.788, + 0.446 + ], + "angle": 0, + "content": "where \\(i,j\\in [0,1,2,\\ldots ,t + 1]\\) is the index of query \\(Q\\) and key \\(K\\). \\(U = [u_{0},u_{1},\\dots ]\\) contains the indices of unmasked tokens. The unidirectional causal mask \\(M^{uc}\\) is a special case of the bidirectional one when \\(U = \\emptyset\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.468, + 0.564, + 0.484 + ], + "angle": 0, + "content": "3.3 Training: Hybrid Attention Masking" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.493, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Given a discrete representation of the motion sequence \\( x_{1:t} \\), our model is trained to reconstruct the motion sequence, conditioned on the text token \\( x_0 \\) under both unidirectional and bidirectional causal masking strategies, i.e., \\( M^{uc} \\) and \\( M^{bc} \\). The reconstruction probability of each motion token under each masking case is \\( p_{\\theta}(x_i \\mid M^{uc}) \\) and \\( p_{\\theta}(x_i \\mid M^{bc}) \\), respectively. The training objective is to minimize the negative log-likelihood of the motion sequence prediction" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.606, + 0.787, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {h y b r i d}} = - \\underset {\\mathbf {X} \\in p (\\mathbf {X})} {\\mathbb {E}} \\left[ \\lambda \\sum_ {\\forall i \\in [ 1, t ]} \\log p _ {\\theta} \\left(x _ {i} \\mid M ^ {u c}\\right) + (1 - \\lambda) \\sum_ {\\forall i \\in [ 1, t ]} \\log p _ {\\theta} \\left(x _ {i} \\mid M ^ {b c}\\right) \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.657, + 0.788, + 0.716 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the probability of selecting unidirectional causal mask. Through experiments, we found \\(\\lambda = 0.5\\) yields best performance. In addition, when bidirectional causal mask is selected, we randomly put \\(50\\% - 100\\%\\) motion tokens in the masked areas." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.578, + 0.755 + ], + "angle": 0, + "content": "3.4 Inference: Cascaded Motion Decoding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To generate motion sequence during inference phase, dual-iteration cascaded decoding is introduced. In the first iteration, autoregressive decoding is applied, where the motion tokens are sequentially and stochastically sampled according to unidirectional prediction distribution \\( p_{\\theta}(x_i \\mid M^{uc}) \\), concluding upon predicting the [END] token. Since autoregressive decoding could accumulate prediction" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.449, + 0.13 + ], + "angle": 0, + "content": "E. Pinyoanuntapong et al." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.147, + 0.726, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Fig. 3: Inference: Dual-iteration Cascaded Motion Decoding. In the first iteration, autoregressive decoding is applied by adopting unidirectional causal mask to generate coarse-grained motion and predict motion sequence length. In the second iteration, bidirectional autoregressive decoding is performed via bidirectional causal mask to removing and repredicting low-confidence motion tokens autoregressively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.47 + ], + "angle": 0, + "content": "errors from the previously generated tokens, the generated motion sequence is refined in the second iteration by masking out a subset of motion tokens and then resampling these masked tokens according to the bidirectional prediction distribution \\( p_{\\theta}(x_i \\mid M^{bc}) \\). In this refinement iteration, [END] is positioned where it was predicted in the first iteration. The model can attend to unmasked tokens in all directions to re-predict low-confidence tokens based on rich surrounding context. The choice of masking strategies has an impact on the refinement gain, which is evaluated in Section 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Hybrid Classifier-free Guidance. At training time, we randomly drop textual tokens to teach the model to generate motion unconditionally. During inference, we apply classifier-free guidance (CFG) [20] for cascaded motion decoding. In particular, we generate the final motion sequence by a linear combination of the conditioned logits \\(\\ell_c\\) and unconditional logits \\(\\ell_u\\) with guidance scale \\(s\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.562, + 0.786, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\ell_ {g} = (1 + s) \\cdot \\ell_ {c} - s \\cdot \\ell_ {u}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.613 + ], + "angle": 0, + "content": "We apply different CFG scale \\( s \\) for each iteration during cascaded decoding. The effectiveness of CFG in each iteration is evaluated in Section 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Residual Motion Refinement. To further enhance motion generation quality, the motion sequence yielded from cascaded decoding can be refined by another refinement transformer based on residual vector quantization (RVQ) [44]. By utilizing RVQ, the raw motion sequence is encoded into multiple token sequences in the latent space. Each token sequence is generated by a separate quantizer and each quantizer encodes the quantization error left by the previous quantizer. As a result, the token sequence from the first quantizer encodes the most of information of the original motion sequence. The rest of the token sequences from other quantizers only encode quantization errors. Through RVQ, information loss can be minimized during the embedding quantization process. As a result, our masked self-attention transformer is used to generate the most informative token sequence from the first quantizer. Using this sequence as input, another refinement transformer is trained to predict the remaining token sequences, which are merged into a single token sequence for final motion decoding. This RVQ-based refinement has been adopted by audio generation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAMM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.145, + 0.79, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.37 + ], + "angle": 0, + "content": "Fig. 4: Residual Motion Refinement. The residual vector quantization encodes the raw motion sequence into multiple token sequences in different colors (left). The base token sequence from the first vector quantizer is generated via cascaded decoding by masked self-attention transformer. The base token sequence is used as the input of the refinement transformer to predict the residual token sequences from other quantizers. The combined sequences are fed into tokenizer's decoder for motion generation. The refinement transformer shares the same architecture as the masked self-attention transformer with a full attention mask(right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.786, + 0.418 + ], + "angle": 0, + "content": "models [4, 11, 40] and recently demonstrates its benefits in motion generation task [14]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.44, + 0.416, + 0.456 + ], + "angle": 0, + "content": "3.5 Motion Editability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.618 + ], + "angle": 0, + "content": "The autoregressive approach naturally lacks the ability for temporal motion editing as it cannot leverage future motion tokens. In contrast, BAMM enables temporal motion editing through its bidirectional causal mask, which allows the model to access information from all directions of the conditioned tokens. Consequently, temporal motion editing becomes achievable by merely applying masked attention to the positions that require medications. We illustrate various editing tasks, i.e. motion inpainting (in-betweening), outpainting, prefix prediction, and suffix completion qualitatively and quantitively in Fig. 7 and Table 5. BAMM also allows us to generate the arbitrarily long motion sequence according to a sequence of text prompts, which is showcased in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.642, + 0.377, + 0.659 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.795 + ], + "angle": 0, + "content": "In this section, we provide a comprehensive evaluation along with empirical results for our proposed motion generation model, BAMM. In Section 4.1, we adhere to standard evaluation protocols on two datasets, demonstrating that our model outperforms current state-of-the-art methods both quantitatively and qualitatively. Furthermore, in Section 4.2, we showcase our predictive length and editing capabilities in comparison to state-of-the-art methods. The results reveal the robust performance of our model, displaying its effectiveness even in cases without prior information about motion length." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Datasets. We evaluated our model using the evaluation protocol proposed by [16] for text-driven motion generation with two datasets. KIT-ML dataset contains 3,911 motion sequences, each with one to four textual annotations," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.447, + 0.129 + ], + "angle": 0, + "content": "E. Pinyoanuntapong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "amounting to a total of 6,278 annotations. These motion sequences are derived from the KIT and CMU motion [1] databases and have been adjusted to 12.5 FPS. The dataset is divided into training, validation, and testing sets, with proportions of \\(80\\%\\), \\(5\\%\\), and \\(15\\%\\) respectively. HumanML3D dataset includes a wide variety of human activities, such as exercise and dancing, and consists of 14,616 motion sequences paired with 44,970 textual descriptions. These descriptions come from a vocabulary of 5,371 unique words. The motion sequences, sourced from AMASS [24] and HumanAct12 [17], have been standardized to 20 FPS and are limited to a maximum duration of 10 seconds, with the actual lengths varying between 2 to 10 seconds. Each sequence is accompanied by at least three descriptive annotations, averaging 12 words in length." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt the standard evaluation framework from T2M [16], employing pre-trained models that encode text and motion information to evaluate text and motion tokens in the embedding space. R-precision (Top-1, 2, 3 accuracy) measures how well the generated motions align with the text prompts, while Multimodal Distance (MM-Dist) quantifies the distance between generated and ground-truth motions in a shared feature space. Frechet Inception Distance (FID) assesses the statistical similarity between the feature distributions of generated and real motions. Additionally, we evaluate diversity (the average Euclidean distance between random motion pairs) and multimodality (the average variance across Euclidean distances between generated motion pairs for a single prompt) to capture the range of possible motion interpretations and their consistency with the text description." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.62, + 0.529 + ], + "angle": 0, + "content": "4.1 Comparison to State-of-the-art Approaches" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Table 2: Comparison of text-conditional motion synthesis on HumanML3D [16] test set. We repeat the evaluation 20 times for each metric and report the average with \\(95\\%\\) confidence interval. Red and Blue indicate the best and the second best result. Methods with gray highlight \\(\\S\\) report motion generation results using the ground-truth motion length." + }, + { + "type": "table", + "bbox": [ + 0.254, + 0.633, + 0.744, + 0.811 + ], + "angle": 0, + "content": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Hier [13]0.301±.0020.425±.0020.552±.0046.523±.0245.012±.0188.332±.042-
TEMOS§ [27]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
TM2T [15]0.424±.0030.618±.0030.729±.0021.501±.0173.467±.0118.589±.0762.424±.093
T2M [16]0.455±.0030.636±.0030.736±.0021.087±.0213.347±.0089.175±.0832.219±.074
MDM§ [37]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse§ [46]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD§ [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
Fg-T2M§ [41]0.492±.0020.683±.0030.783±.0020.243±.0193.109±.0079.278±.0721.614±.049
M2DM§ [23]0.497±.0030.682±.0020.763±.0030.352±.0053.134±.0109.926±.0733.587±.072
T2M-GPT [45]0.491±.0030.680±.0030.775±.0020.116±.0043.118±.0119.761±.0811.856±.011
AttT2M [49]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
MMM§ [29]0.515±.0020.708±.0020.804±.0020.089±.0052.926±.0079.577±.0501.226±.035
MoMask§ [14]0.521±.0020.713±.0030.807±.0020.045±.0022.958±.008-1.241±.040
BAMM (ours)0.525±.0020.720±.0030.814±.0030.055±.0022.919±.0089.717±.0891.687±.051
BAMM§ (ours)0.522±.0030.715±.0030.808±.0030.055±.0022.936±.0779.636±.0091.732±.055
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAMM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Comparison of text-conditional motion synthesis on KIT-ML [30] test set. We repeat the evaluation 20 times for each metric and report the average with \\(95\\%\\) confidence interval. Red and Blue indicate the best and the second best result. Methods with gray highlight \\(\\S\\) report ground-truth motion length for generation." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.205, + 0.742, + 0.384 + ], + "angle": 0, + "content": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Hier [13]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.072-
TEMOS$ [27]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
TM2T [15]0.280±.0050.463±.0060.587±.0053.599±.1534.591±.0269.473±.1173.292±.081
T2M [16]0.361±.0060.559±.0070.681±.0073.022±.1073.488±.02810.72±.1452.052±.107
MDM$ [37]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MotionDiffuse$ [46]0.417±.0040.621±.0040.739±.0041.954±.0642.958±.00511.10±.1430.730±.013
MLD$ [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
Fg-T2M$ [41]0.418±.0050.626±.0040.745±.0040.571±.0473.114±.01510.93±.0831.019±.029
M2DM$ [23]0.416±.0040.628±.0040.743±.0040.515±.0293.015±.01711.417±.973.325±.37
T2M-GPT [45]0.402±.0060.619±.0050.737±.0060.717±.0413.053±.02610.86±.0941.912±.036
AttT2M [49]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
MMM$ [29]0.404±.0050.621±.0050.744±.0040.316±.0282.977±.01910.910±.1011.232±.039
MoMask$ [14]0.433±.0070.656±.0050.781±.0050.204±.0112.779±.022-1.131±.043
BAMM (ours)0.438±.0090.661±.0090.788±.0050.183±.0132.723±.02611.008±.0941.609±.065
BAMM (ours)$0.436±.0070.660±.0060.791±.0050.200±.0112.714±.01610.914±.0971.517±.058
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.599 + ], + "angle": 0, + "content": "Quantitative Results. We evaluate our model on the HumanML3D [16] and KIT-ML [30] datasets, reporting the results in Table 2 and 3, respectively, in comparison to state-of-the-art methods. Following the standard evaluation protocol from [16], we report the average of 20 generations with a \\(95\\%\\) confidence interval. Our model consistently outperforms other methods in terms of R-precision, FID, and MM-Distance, while maintaining comparability in terms of Diversity and Multimodal Distance, which typically represent the trade-off between high quality and diversity. This suggests that our model generates very high-quality outputs while retaining a good degree of diversity. Moreover, whereas most models utilize the ground truth length for evaluation, our model supports both predicted length and takes length as an input, outperforming other methods in both scenarios. We further investigate the effect of a separate length estimator which can significantly impact performance in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.787, + 0.766 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 5 presents a qualitative comparison with T2M-GPT [45], MoMask [14], and MDM [37]. BAMM and T2M-GPT generate motion without requiring input length. Despite this, BAMM accurately generates motion from textual descriptions even without input length. We utilize a pre-trained length estimator from [16] for MoMask [14] and MDM [37]. Notably, BAMM generates motion accurately aligned with the provided text, whereas T2M-GPT and MoMask produce erroneous motion, and MDM generates entirely inaccurate motion. Additional visualizations for BAMM are shown in Fig. 6, indicating its ability to generate high detail such as complex trajectories and interactions with invisible objects. Moreover, Fig. 7 demonstrates BAMM's capability in various temporal editing tasks, including inpainting, outpainting, prefix, and suffix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.536, + 0.802 + ], + "angle": 0, + "content": "4.2 Length Prediction and Editability" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.797, + 0.84 + ], + "angle": 0, + "content": "Predict Length vs ground truth length. In real-world scenarios, the ground truth length is often unknown. Many methods cannot predict length autonomously." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.446, + 0.129 + ], + "angle": 0, + "content": "E. Pinyoanuntapong et al." + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.133, + 0.725, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.326, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Fig. 5: Visualization comparison of textual to motion to state-of-the-art methods. BAMM and T2M-GPT do not require motion length as an input. We use a pre-trained length estimator from [16] for MoMask and MDM. BAMM generates higher quality and is more correlated with textual descriptions." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.398, + 0.73, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.507, + 0.785, + 0.548 + ], + "angle": 0, + "content": "Fig. 6: Visualization of text-to-motion generation by BAMM. BAMM can generate high-quality motion with complex descriptions, such as intricate trajectories and interactions with invisible objects." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.561, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Consequently, they evaluate using ground truth length, indicated by gray highlight \\(\\S\\) in Table 2 and Table 3. We illustrate the impact of inaccurate length prediction by comparing our model to state-of-the-art denoising models that require motion length as an input, such as MoMask and MMM. In particular, both models take the estimated length obtained from a standalone length predictor pretrained on the HumanML3D dataset [16]. Table 4 reveals that the predicted length significantly worsens MoMask's FID score, from 0.045 to 0.090, while MMM's R-precision Top1 drops from 0.515 to 0.504. In contrast, our BAMM not only maintains the best performance but also outperforms MMM and MoMask in both scenarios." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.736, + 0.785, + 0.765 + ], + "angle": 0, + "content": "Table 4: Comparison of text-conditional motion synthesis using predicted and ground truth length on HumanML3D [16] dataset" + }, + { + "type": "table", + "bbox": [ + 0.307, + 0.769, + 0.684, + 0.845 + ], + "angle": 0, + "content": "
LengthMethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
GroundMMM0.5150.7080.8040.0892.9269.5771.226
TruthMoMask0.5210.7130.8070.0452.958-1.241
BAMM (Ours)0.5220.7150.8080.0552.9369.6361.732
PredictedMMM0.5040.6960.7940.0802.9989.4111.164
MoMask0.5220.7150.8110.0902.9459.6471.239
BAMM (Ours)0.5250.7200.8140.0552.9199.7171.687
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAMM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.782, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "Fig. 7: Visualization of temporal editing tasks, inpainting (in-betweening), outpainting, prefix, and suffix where blue indicates conditioned motion and red refers to generated parts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.478 + ], + "angle": 0, + "content": "Editability. We conduct experiments in four temporal editing tasks, namely inpainting (in-betweening), outpainting, prefix, and suffix, comparing them with MDM [37] and MoMask [14] on the HumanML3D dataset. Inpainting is evaluated by generating \\(50\\%\\) of the motion sequence given the first and last \\(25\\%\\). Outpainting is its opposite counterpart. Prefix is conditioned by the first \\(50\\%\\) of the ground truth motion and generates the remaining portion. Similarly, suffix operates in the reverse manner. Note that editing tasks require prior knowledge of conditioned motion and motion length, for which MDM and MoMask are specifically designed. Despite this, Table 5 shows that our model significantly outperforms MDM in all editing tasks and surpasses MoMask in outpainting, while performing equally well in the other tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.492, + 0.787, + 0.507 + ], + "angle": 0, + "content": "Table 5: Evaluation on temporal editing tasks on HumanML3D [16] dataset" + }, + { + "type": "table", + "bbox": [ + 0.321, + 0.513, + 0.673, + 0.646 + ], + "angle": 0, + "content": "
TasksMethodsR-Precision ↑FID ↓ MM-Dist ↓ Diversity ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Temporal Inpainting (In-betweenging)MDM0.3910.5780.6922.3623.8598.014
MoMask0.5340.7270.820.042.8789.64
BAMM0.5350.7290.8210.0562.8639.629
Temporal OutpaintingMDM0.4150.6130.7272.0573.6198.199
MoMask0.5310.7260.8180.0572.8899.619
BAMM0.5350.730.8220.0562.8569.659
PrefixMDM0.420.6130.7251.463.5638.312
MoMask0.5360.730.8220.062.8759.607
BAMM0.5320.7270.8210.0582.8689.612
SuffixMDM0.4030.5970.7112.5623.7318.088
MoMask0.5320.7260.8190.0522.8819.659
BAMM0.5270.720.8140.052.8919.721
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.403, + 0.691 + ], + "angle": 0, + "content": "5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.786, + 0.749 + ], + "angle": 0, + "content": "In the ablation study, we study the impacts of the adaptive classifier-free guidance scales in each iteration of the Masked Self-attention Transformer, the masking strategy, and the number of iterations, as shown in Table 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Adaptive Classifier-Free Guidance Scales (CFG): We conducted experiments with different CFG values for both the first and second iterations and found that CFG=4 in the first iteration and CFG=3 in the second iteration works best. Although CFG=5 in the first iteration yields a better FID score, the other scores are worse. Mask Strategy: The first strategy, \"50% of low confidence\", applies mask attention on 50% of the lowest confident positions from the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.447, + 0.129 + ], + "angle": 0, + "content": "E. Pinyoanuntapong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.409, + 0.145, + 0.594, + 0.159 + ], + "angle": 0, + "content": "Table 6: Ablation Study" + }, + { + "type": "table", + "bbox": [ + 0.268, + 0.16, + 0.727, + 0.346 + ], + "angle": 0, + "content": "
AblationsTypesR-Precision ↑FID ↓ MM-Dist ↓ Diversity ↑ MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
CFG of 1st iteration1st iter2nd iter
CFG=2CFG=30.5170.7110.8050.1052.9569.841.907
CFG=3CFG=30.5210.7140.8080.072.9449.7771.766
CFG=4CFG=30.5250.720.8140.0552.9199.7171.687
CFG=5CFG=30.5220.7160.810.0522.9279.6471.691
CFG=6CFG=30.520.7130.810.062.949.6211.697
CFG of 2nd iterationCFG=4CFG=10.5210.7160.810.0582.9439.7431.744
CFG=4CFG=20.5240.7190.8140.0572.9249.7521.698
CFG=4CFG=30.5250.720.8140.0552.9199.7171.687
CFG=4CFG=40.5220.7190.8120.0562.9249.6911.698
CFG=4CFG=50.5210.7170.8120.0652.9319.6381.735
Mask Strategy50% of low confidence confidence < .50.5250.720.8130.0652.9219.7321.67
suffix0.5250.7180.810.0642.9239.7651.656
%2=00.5190.7150.810.0522.9439.6831.841
0.5250.720.8140.0552.9199.7171.687
# of iterations1 iteration0.5240.7180.8120.0642.9269.7201.644
2 iterations0.5250.7200.8140.0552.9199.7171.687
3 iterations0.5250.7190.8140.0552.9179.7271.69
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.626 + ], + "angle": 0, + "content": "first iteration. Similarly, \"confidence \\(< .5\\) \" uses low confidence but sets a threshold, masking out positions where confidence is below 0.5. Specifically, the former strictly masks out \\(50\\%\\) of the total token sequence while the latter masks the token with a confidence lower than .5. \"Suffix\" strategy masks out the first \\(50\\%\\) of the token sequence, utilizing the remaining \\(50\\%\\) as condition tokens. Lastly, \\(\\% 2 = 0\\) \"strategy masks every other token. BAMM produces decent performances across the tested masking strategies while the simple masking of every other token works the best. Number of Iterations: We conducted experiments with one to three iterations. \"1 iteration\" refers to the first iteration, as described in 3.4, which utilizes a unidirectional causal mask. Similarly, \"2 iterations\" involves applying bidirectional causal mask to the Masked Self-attention Transformer to re-predict the tokens from the first iteration, as illustrated in 3. In \"3 iterations\", we repeat the second iteration (bidirectional causal mask) but apply masks to \\(1/3\\) of the sequence to re-predict the motion tokens instead. The experiments indicate that \"2 iterations\" clearly demonstrate improvement over one iteration, while \"3 iterations\" do not significantly improve upon \"2 iterations\". This suggests that \"2 iterations\" are sufficient." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.653, + 0.36, + 0.67 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduce the Bidirectional Autoregressive Motion Model (BAMM), a novel framework for text-to-motion generation. BAMM combines a motion tokenizer, which encodes 3D human motion into discrete latent tokens, with a masked self-attention transformer that autoregressively predicts the masked tokens through a masked casual attention approach. BAMM integrates of generative masked and autoregressive modeling into an unified framework, This features allows it to understand complex motion relationships and precisely map text inputs to high-quality motion outputs with adaptively adjusted sequence lengths. Our extensive testing on HumanML3D and KIT-ML datasets confirms BAMM's superiority in both qualitative and quantitative evaluations over existing methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAMM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.198 + ], + "angle": 0, + "content": "1. Cmu graphics lab motion capture database, http://mocap.cs.cmu.edu/, accessed: 2022-11-11 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.2, + 0.785, + 0.24 + ], + "angle": 0, + "content": "2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728 (2019). https://doi.org/10.1109/3DV.2019.000844" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.241, + 0.785, + 0.281 + ], + "angle": 0, + "content": "3. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems 34, 17981-17993 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.322 + ], + "angle": 0, + "content": "4. Borsos, Z., Sharifi, M., Vincent, D., Kharitonov, E., Zeghidour, N., Tagliasacchi, M.: Soundstorm: Efficient parallel audio generation. ArXiv abs/2305.09636 (2023), https://api-semanticscholar.org/CorpusID:258715176 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.322, + 0.785, + 0.417 + ], + "angle": 0, + "content": "5. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T.J., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. ArXiv abs/2005.14165 (2020), https://api_semanticscholar.org/CorpusID:218971783 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.785, + 0.472 + ], + "angle": 0, + "content": "6. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M., Murphy, K.P., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D.: Muse: Text-to-image generation via masked generative transformers. ArXiv abs/2301.00704 (2023), https://api(semanticscholar.org/CorpusID:255372955 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.527 + ], + "angle": 0, + "content": "7. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11305-11315 (2022), https://apisemantic scholar.org/CorpusID:246680316 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.581 + ], + "angle": 0, + "content": "8. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, J., Yu, G.: Executing your commands via motion diffusion in latent space. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 18000-18010 (2022), https://api-semanticscholar.org/CorpusID:254408910 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.636 + ], + "angle": 0, + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019), https://apisemantic scholar.org/CorpusID:52967399 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.676 + ], + "angle": 0, + "content": "0. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. ArXiv abs/2204.14217 (2022), https://api(semanticscholar.org/CorpusID:248476190 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.731 + ], + "angle": 0, + "content": "1. Du, Z., Qian, Y., Liu, X., Ding, M., Qiu, J., Yang, Z., Tang, J.: Glm: General language model pretraining with autoregressive blank infilling. In: Annual Meeting of the Association for Computational Linguistics (2021), https://apisemantic scholar.org/CorpusID:247519241 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.785 + ], + "angle": 0, + "content": "2. Ghazvininejad, M., Levy, O., Liu, Y., Zettlemoyer, L.: Mask-predict: Parallel decoding of conditional masked language models. In: Conference on Empirical Methods in Natural Language Processing (2019), https://api_semanticscholar.org/ CorpusID:202538740 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 1376-1386 (2021), https://apisemantic scholar.org/CorpusID:232404671 10, 11" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.446, + 0.129 + ], + "angle": 0, + "content": "E. Pinyoanuntapong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions (2023) 2, 4, 6, 9, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.232 + ], + "angle": 0, + "content": "15. Guo, C., Xuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. ArXiv abs/2207.01696 (2022), https://api_semanticscholar.org/CorpusID: 250280248 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.288 + ], + "angle": 0, + "content": "16. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5142-5151 (2022). https://doi.org/10.1109/CVPR52688.2022.005093, 4, 5, 9, 10, 11, 12, 13, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.344 + ], + "angle": 0, + "content": "17. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. Proceedings of the 28th ACM International Conference on Multimedia (2020), https://apisemantic scholar.org/CorpusID:220870974 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.786, + 0.401 + ], + "angle": 0, + "content": "18. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A.A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models. ArXiv abs/2210.02303 (2022), https://api_semanticscholar.org/CorpusID:252715883 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.786, + 0.441 + ], + "angle": 0, + "content": "19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. ArXiv abs/2006.11239 (2020), https://apisemantic scholar.org/CorpusID: 219955663 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.444, + 0.786, + 0.472 + ], + "angle": 0, + "content": "20. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.786, + 0.514 + ], + "angle": 0, + "content": "21. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. ArXiv abs/2306.14795 (2023), https://api.sementicscholar.org/CorpusID:259262201 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.786, + 0.557 + ], + "angle": 0, + "content": "22. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI Conference on Artificial Intelligence (2022), https://api-semanticscholar.org/CorpusID:251979380 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.786, + 0.599 + ], + "angle": 0, + "content": "23. Kong, H., Gong, K., Lian, D., Mi, M.B., Wang, X.: Priority-centric human motion generation in discrete latent space. ArXiv abs/2308.14480 (2023), https://api.sementicscholar.org/CorpusID:261245369 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.786, + 0.657 + ], + "angle": 0, + "content": "24. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 5441-5450 (2019), https://apisemanticscholar.org/CorpusID:102351100 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.786, + 0.713 + ], + "angle": 0, + "content": "25. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning (2021), https://api-semanticscholar.org/CorpusID:245335086 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.786, + 0.755 + ], + "angle": 0, + "content": "26. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning. ArXiv abs/1711.00937 (2017), https://api-semanticscholar.org/CorpusID:20282961 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.786, + 0.798 + ], + "angle": 0, + "content": "27. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. ArXiv abs/2204.14109 (2022), https://api.sementicscholar.org/CorpusID:248476220 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "28. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. ArXiv abs/2305.00976 (2023), https://api_semanticscholar.org/CorpusID:258436810 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BAMM" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Pinyoanuntapong, E., Wang, P., Lee, M., Chen, C.: Mmm: Generative masked motion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 2, 4, 6, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "30. Plappert, M., Mandery, C., Asfour, T.: The KIT motion-language dataset. Big Data 4(4), 236-252 (dec 2016). https://doi.org/10.1089/big.2016.0028, http://dx.doi.org/10.1089/big.2016.0028 4, 11, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.284 + ], + "angle": 0, + "content": "31. Qian, L., Zhou, H., Bao, Y., Wang, M., Qiu, L., Zhang, W., Yu, Y., Li, L.: Glancing transformer for non-autoregressive neural machine translation. In: Annual Meeting of the Association for Computational Linguistics (2020), https://api_semanticscholar.org/CorpusID:221150562 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.352 + ], + "angle": 0, + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021), https://api_semanticscholar.org/CorpusID: 231591445 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.42 + ], + "angle": 0, + "content": "33. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. ArXiv abs/2205.11487 (2022), https://api_semanticscholar.org/CorpusID:248986576 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.475 + ], + "angle": 0, + "content": "34. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data. ArXiv abs/2209.14792 (2022), https://api_semanticscholar.org/CorpusID:252595919 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.514 + ], + "angle": 0, + "content": "35. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ArXiv abs/2010.02502 (2020), https://api.sementicscholar.org/CorpusID:222140788 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.556 + ], + "angle": 0, + "content": "36. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision (2022), https://api-semanticscholar.org/CorpusID:247450907 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.596 + ], + "angle": 0, + "content": "37. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. ArXiv abs/2209.14916 (2022), https://api.sementicscholar.org/CorpusID:252595883 2, 4, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.65 + ], + "angle": 0, + "content": "38. Vaswani, A., Shazeer, N.M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:137564895" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.704 + ], + "angle": 0, + "content": "39. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. ArXiv abs/2210.02399 (2022), https://apisemantic scholar.org/CorpusID:252715594 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.759 + ], + "angle": 0, + "content": "40. Wang, C., Chen, S., Wu, Y., Zhang, Z.H., Zhou, L., Liu, S., Chen, Z., Liu, Y., Wang, H., Li, J., He, L., Zhao, S., Wei, F.: Neural codec language models are zero-shot text to speech synthesizers. ArXiv abs/2301.02111 (2023), https://api.sementicscholar.org/CorpusID:255440307 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "41. Wang, Y., Leng, Z., Li, F.W.B., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. ArXiv abs/2309.06284 (2023), https://api.sementicscholar.org/CorpusID:261697123 2, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.84 + ], + "angle": 0, + "content": "42. Yan, S., Liu, Y., Wang, H., Du, X., Liu, M., Liu, H.: Cross-modal retrieval for motion and text via doptriple loss (2023), https://api-semanticscholar.org/CorpusID:263610212 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.446, + 0.129 + ], + "angle": 0, + "content": "E. Pinyoanuntapong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.217 + ], + "angle": 0, + "content": "43. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., Hutchinson, B.C., Han, W., Parekh, Z., Li, X., Zhang, H., Baldridge, J., Wu, Y.: Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res. 2022 (2022), https://api_semanticscholar.org/CorpusID:249926846 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.272 + ], + "angle": 0, + "content": "44. Zeghidour, N., Luebs, A., Omran, A., Skoglund, J., Tagliasacchi, M.: Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30, 495-507 (2021), https://api-semanticscholar.org/CorpusID:236149944 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.342 + ], + "angle": 0, + "content": "45. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: Generating human motion from textual descriptions with discrete representations. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14730-14740 (2023), https://api_semanticscholar.org/CorpusID:255942203 2, 5, 6, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.396 + ], + "angle": 0, + "content": "46. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. ArXiv abs/2208.15001 (2022), https://api_semanticscholar.org/CorpusID: 251953565 2, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.452 + ], + "angle": 0, + "content": "47. Zhang, Z., Ma, J., Zhou, C., Men, R., Li, Z., Ding, M., Tang, J., Zhou, J., Yang, H.: M6-ufc: Unifying multi-modal controls for conditional image synthesis via non-autoregressive generative transformers (2021), https://api_semanticscholar.org/CorpusID:237204528 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.507 + ], + "angle": 0, + "content": "48. Zhang, Z., Ma, J., Zhou, C., Men, R., Li, Z., Ding, M., Tang, J., Zhou, J., Yang, H.: Ufc-bert: Unifying multi-modal controls for conditional image synthesis. In: Neural Information Processing Systems (2021), https://apisemantic scholar.org/CorpusID:235253928 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.551 + ], + "angle": 0, + "content": "49. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. ArXiv abs/2309.00796 (2023), https://api.sementicscholar.org/CorpusID:261530775 2, 5, 10, 11" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.551 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BAMM_ Bidirectional Autoregressive Motion Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_origin.pdf b/2024/BAMM_ Bidirectional Autoregressive Motion Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dfd0ac7dfc9b589cb4b55630fe1cecde911a92d2 --- /dev/null +++ b/2024/BAMM_ Bidirectional Autoregressive Motion Model/2e83647b-3608-45ae-8b2a-4eda59a2fde7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bd6702bb75aa5707ad8b1428cc36d210003af7db1e1414a55163fcd71275ab6 +size 12390500 diff --git a/2024/BAMM_ Bidirectional Autoregressive Motion Model/full.md b/2024/BAMM_ Bidirectional Autoregressive Motion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9aec274954c6ac3863a0240705646fdd5d003231 --- /dev/null +++ b/2024/BAMM_ Bidirectional Autoregressive Motion Model/full.md @@ -0,0 +1,252 @@ +# BAMM: Bidirectional Autoregressive Motion Model + +Ekkasit Pinyoanuntapong1, Muhammad Usama Saleem1, Pu Wang1, Minwoo Lee1, Srijan Das1, and Chen Chen2 + +1 University of North Carolina at Charlotte {epinyoan,msaleem2, Pu.Wang, minwoo.lee, sdas24}@uncc.edu +2 University of Central Florida chen.chen@crcv.ucf.edu + +![](images/7ada26cc70a325b7862be6684fdc5a829b2bccc205a06f6c55132ff335e6673d.jpg) +a) Motion Length Prediction + +![](images/4664df2cc75ea62b3c0b5fd7146c0bb9c31fdcbef7f2b93467c1cf2fe01760d6.jpg) + +![](images/cbbc3c6b7b3e0e524c61171c8e742a57eb7bda0e64be05a2de188aed2007ca99.jpg) +Fig. 1: (a) Motion Length Prediction: Text-to-motion models often require specific input lengths, making them sensitive to motion generation. In contrast, BAMM automatically predicts the end of the motion, thus avoiding reliance on inaccurate motion length estimations. (b) High-quality Text-to-Motion: BAMM generates natural human movements precisely aligned with detailed textual descriptions. (c) Motion Editing: BAMM is capable of multiple editing tasks, such as inpainting (as demonstrated), outpainting, prefix prediction, suffix completion, and arbitrarily long motion sequence synthesis. + +![](images/ff1e38a1bdc1f8d193e019b03893cea99147415a4afa7af04fae0ec07277b385.jpg) +c) Motion Editing + +Abstract. Generating human motion from text has been dominated by denoising motion models either through diffusion or generative masking process. However, these models face great limitations in usability by requiring prior knowledge of the motion length. Conversely, autoregressive motion models address this limitation by adaptively predicting motion endpoints, at the cost of degraded generation quality and editing capabilities. To address these challenges, we propose Bidirectional Autoregressive Motion Model (BAMM), a novel text-to-motion generation framework. BAMM consists of two key components: (1) a motion tokenizer that transforms 3D human motion into discrete tokens in latent space, and (2) a masked self-attention transformer that autoregressively predicts randomly masked tokens via a hybrid attention masking strategy. By unifying generative masked modeling and autoregressive modeling, BAMM captures rich and bidirectional dependencies among motion tokens, while learning the probabilistic mapping from textual inputs to motion outputs with dynamically-adjusted motion sequence length. This feature enables BAMM to simultaneously achieving high-quality motion generation with enhanced usability and built-in motion editability. Extensive experiments on HumanML3D and KIT-ML datasets demonstrate that BAMM surpasses current state-of-the-art methods in both + +qualitative and quantitative measures. Our project page is available at https://exitudio.github.io/BAMM-page + +Keywords: Text to Motion $\cdot$ Autoregressive Motion Model $\cdot$ Generative Masked Motion Model + +# 1 Introduction + +Text-to-motion generation is a promising interdisciplinary field that uses natural language to generate 3D human movements. This emerging field holds vast potential to transform animation, gaming, filming, and VR/AR/MR domains by enabling easy and intuitive creation of 3D assets through user-friendly textual inputs. However, bridging the semantic gap between textual descriptions and intricate motion sequences presents a significant challenge. To address this challenge, recent efforts have been focusing on two methods: (1) conditional denoising motion model and (2) conditional autoregressive motion model. Both methods can greatly improve motion generation quality by learning the probabilistic distribution of motion sequences, conditioned on the textural descriptors. However, both methods face fundamental limitations. + +Conditional denoising motion models are trained to restore corrupted motion sequences to their original state, guided by textual prompts. These models operate through two primary mechanisms: diffusion and generative masking. Diffusion models apply structured Gaussian noise to the original motion data for corruption [22,37,41,46], whereas generative masked models [14,29] corrupt motion sequence by substituting selected motion tokens with [MASK] tokens. The denoising process, followed by motion corruption procedure, considers motion tokens from both directions, effectively capturing the intricate dependencies among tokens. This leads to enhanced motion generation quality. Furthermore, these models inherently ease the motion editing tasks. By selectively corrupting and recovering motion tokens in areas needing modifications, they ensure seamless transitions between edited and unedited segments. + +While denoising motion models excel in generation quality and editability, a significant limitation of these models is their usability because these models depend on prior knowledge of motion length for each text prompt, a requirement that proves impractical in real-world scenarios. Utilizing incorrect motion lengths can result in a significant decline in generation quality. To address this fundamental limitation, conditional motion autoregressive models emerge as a solution, capable of simultaneously predicting the sequence length and content of generated motions. Inspired by large language models like GPT [5], motion autoregressive models sequentially predict one motion token at a time from left to right until the [END] token is predicted, guided by the textual description [21, 45, 49]. As a result, the generated motions are not only well aligned with the text inputs but also appropriately scaled in duration. However, the sequential token decoding of autoregressive models cannot fully capture the dependencies between the motion tokens, potentially compromising generation quality, and complicating + +the editing process, because edited parts need to be conditioned on the unedited parts from both directions to ensure overall continuity and coherence. + +As previously highlighted, existing text-driven motion generation models encounter a comprise among usability, quality of generation, and editability. To address this challenge, we propose Bidirectional Autoregressive Motion Model (BAMM), a novel text-to-motion generation framework. It consists of two pivotal components: a motion tokenizer and a conditional masked self-attention transformer. In a two-stage training paradigm, the motion tokenizer performs initial training based on Vector Quantized Variational Autoencoders (VQ-VAE) [26]. This tokenizer encodes the raw motion sequence into discrete motion tokens within the latent space leveraging learned codebooks. In the subsequent phase, motion tokens are randomly masked out. A conditional masked self-attention transformer is then trained to autoregressively predict these masked tokens, adopting the causal attention masking strategy. This strategy is largely departing from the traditional [MASK] token replacement approach in generative masked models. In particular, it does not substitute the input motion tokens with [MASK] tokens. Instead, it adjusts the attention score matrix according to both unidirectional and bidirectional causal masks. The unidirectional causal mask enables adaptive prediction of [END] token based on text prompts, while bidirectional causal mask forces the model to predict the next motion token not only based on past tokens but also conditioned on future unmasked tokens. This facilitates bidirectional autoregressive training for enhanced predictive capabilities. + +Table 1: Comparison of quality and capability of generation on text-to-motion to state-of-the-art models on the largest text-to-motion dataset [16]. $\checkmark$ means capability while $X$ is not. "Predict Length" denotes the ability to generate motion without prior knowledge of motion length. "Input Length" refers to the ability to take input length as a constraint, while "Edit" indicates motion editability. Since MMM and MoMask require ground-truth motion length as input, we use predicted motion length from pretrained length estimator by [16]. The lowest FID score means the best overall quality of the generated motion, ensuring that its authenticity and naturalness is very close to the ground-truth human movements. High R-precision and low MM-dist means accurate alignment between the generated motion and the text prompts. + +
MethodsTop-1 ↑FID ↓MM-Dist ↓Predict LengthInput LengthEdit
T2M-GPT0.4910.1163.118✓✗✗
AttT2M0.4990.1123.038✓✗✗
MMM0.5040.0802.998✗✓✓
MoMask0.5220.0902.945✗✓✓
BAMM (Ours)0.5250.0552.919✓✓✓
+ +By unifying masked and autoregressive prediction during training, BAMM captures rich and bidirectional dependencies among motion tokens, while learning a direct probabilistic mapping from textual inputs to motion outputs with dynamically-adjusted motion sequence length. Leveraging such unique feature, we propose cascaded motion generation during inference, where BAMM first leverages unidirectional autoregressive decoding to implicitly predict motion se + +quence length and generate coarse-grained motion sequence. Such motion sequence is then refined by masking and regenerating a portion of motion tokens in a bidirectional autoregressive fashion. This feature allows BAMM to achieve high-quality motion generation with high usability. Moreover, BAMM naturally supports zero-shot motion editing without specially being trained for such task. By treating the masked motion tokens as the contents that need editing, BAMM can predict the masked tokens based on the surrounding context and the text description. Our main contributions can be summarized as follows. + +- We introduce the bidirectional autoregressive motion model, which is a novel text-to-motion generation framework. It effectively harnesses the complementary benefits of denoising and autoregressive models, thus simultaneously achieving high-quality motion generation with enhanced usability and innate motion editability, as showcased in Fig. 1 and Table 1. +- We demonstrate that our model outperforms current state-of-the-art methods qualitatively and quantitatively on two standard text-to-motion generation datasets, HumanML3D [16] and KIT-ML [30]. +- We showcase that our model supports a variety of motion editing tasks in the zero-shot manner without specially training for these tasks, including motion inpainting, outpainting, prefix prediction, suffix completion, and long sequence generation. + +# 2 Related Work + +Motion Synthesis with Latent Space Alignment. Early text-to-motion approaches typically adopt a two-stage scheme: learning separate latent representations for text and motion sequences, followed by latent space alignment using distance losses like cosine similarity or KL divergence [2,16,27,28,36,42]. For example, Language2Pose [2] aimed to establish a shared latent space for both language descriptions and motion sequences. MotionCLIP [36] simplified this concept by incorporating stylization and diversity techniques, leveraging the pre-trained text-image latent space of the CLIP model [32]. However, this strategy inherently struggles with generating high-fidelity motions due to the difficulty of perfectly aligning these inherently disparate latent spaces. + +Conditional Denoising Motion Model. Inspired by the success of denoising diffusion models (DDMs) [19,35] in text-to-image and text-to-video generation [18, 25, 33, 34], diffusion models have been applied for text-to-motion generation. MDM [37], MotionDiffuse [46], MLD [8], and FLAME [22] are the representative examples. Meanwhile, BERT-type masked generative models have shown their success in both text generation tasks, e.g., Q&A and language translations [9,12,31], and text-to-image synthesis [6,7,10,39,43,47,48]. Following this trend, MMM [29] and MoMask [14] are recent attempts to propose conditional masked motion modeling to enable high-fidelity motion synthesis that is precisely aligned with text prompts. Both diffusion and generative masked modeling share the same principle: "denoising" corrupted data [3]. Meanwhile, they also face the + +same usability limitation: they require prior knowledge of the motion sequence length for the denoising process. This length information is obtained either by intuitive guess from users or via a separately pre-trained predictor that predicts the motion length distribution based on the input text prompts [16]. Both approaches, however, lead to significant motion quality degradation as demonstrated in the experiments section. This is because the distributions of motion length and motion contents are inherently coupled, which need to be jointly aligned with the textual context. + +Conditional Autoregressive Motion Model. The autoregressive motion models, such as T2M-GPT [45], AttT2M [49], and MotionGPT [21], can effectively address the usability challenge faced by denoising models because they follow the GPT-type training and inference [5] to implicitly predict motion sequence length by generating the [END] token conditioned on both previously generated motion tokens and text inputs. However, the limitation of these models lies in their use of causal attention for unidirectional and sequential motion token prediction. This practice not only hinders model's motion editability but also jeopardizes motion generation quality. To address this limitation, we propose the first bidirectional autoregressive modeling approach for human motion generation, which draws inspiration from the self-attention masking employed by large language model pretraining [11,38]. + +# 3 Method + +![](images/84c38308060370eb216a17d42ed87844ef9168e44236fe87b0e6ecfd71087d4f.jpg) +Fig.2: Overall architecture of BAMM. (a) Motion Tokenizer encodes the raw motion sequence into discrete motion tokens according to a learned codebook. (b) Masked Self-attention Transformer learns to sequentially predict next tokens conditioned on text embedding from CLIP model and future unmasked tokens. Masked self-attention mechanism unifies autoregressive model and generative masked motion via bidirectional and unidirectional causal masks. + +Our objective is to create a text-to-motion synthesis framework that simultaneously achieves high-quality motion generation with enhanced usability and innate motion editability. Towards this goal, our framework, as illustrated in Fig. 2, consists of two key components: motion tokenizer that compresses and converts raw 3D human motion into a sequence of discrete motion tokens in the + +latent space (Section 3.1) and conditional masked self-attention transformer that leverages unidirectional and bidirectional casual masks to integrate autoregressive model and generative masked model into a unified framework (Section 3.2). The training procedure follows a hybrid attention masking strategy, where the two causal masks are applied randomly and the model is forced to reconstruct the motion sequence under both cases (Section 3.3). The cascaded motion decoding is introduced for motion generation during the inference phase. It uses unidirectional autoregressive decoding to jointly predict motion sequence length and its contents, which are refined via bidirectional autoregressive decoding (Section 3.4). + +# 3.1 Motion Tokenizer + +The objective of this stage is to learn the discrete representation of motion by quantizing the embedding $z$ from the output of the encoder into codebook $\mathcal{C}$ . We first pretrain a motion tokenizer based on VQ-VAE [26]. In particular, given a motion sequence $\mathcal{M} = [m_1, m_2, m_3, \dots, m_\tau]$ where $m \in \mathbb{R}^D$ , $\tau$ is the total frames of motion, and $D$ is the dimension of the 3D pose in each frame, encoder is used to encode motion $\mathcal{M}$ to the latent embedding $z \in \mathbb{R}^{t \times d}$ with a downsampling rate of $\tau / t$ . The embedding $z$ is quantized into codes $c \in \mathcal{C}$ . Codebook $\mathcal{C} = \{\gamma_k\}_{k=1}^K$ contains $K$ number of codes. The nearest Euclidean distance between the embedding $\mathbf{z}$ and the code of vector is computed by $\hat{z}_i = \mathrm{argmin}_j \| \mathbf{z} - \mathcal{C}_j \|_2^2$ . The loss function is defined as + +$$ +L _ {V Q} = \left\| \operatorname {s g} (\mathbf {z}) - \mathbf {e} \right\| _ {2} ^ {2} + \beta \| \mathbf {z} - \operatorname {s g} (\mathbf {e}) \| _ {2} ^ {2}, \tag {1} +$$ + +where $\mathrm{sg}(\cdot)$ is the stop-gradient operator, $\beta$ refers the hyper-parameter for commitment loss. The loss function is optimized via a straight-through gradient estimator. We apply exponential moving average for codebooks update and codebook reset by following [45] [14] [29]. + +# 3.2 Conditional Masked Self-attention Transformer + +Our model employs a standard multi-layer transformer, whose inputs are the concatenation of the motion tokens $x_{1:t}$ from the tokenizer with $t$ as the sequence length, the text embedding $x_0$ from the pre-trained CLIP model [32], and the [END] token $x_{t + 1}$ that serves as the indicator of the motion's endpoint. The input tokens $x_{0:t + 1}$ are masked out strategically. Different from generative masked models, we do not replace the input tokens with [MASK] ones. Instead, we adopt causal attention mask $M$ as shown in Fig. 2 (b) to specify the attention relations among the input tokens. In particular, the token in the masked areas, indicated by $\blacksquare$ , can be attended to itself, all the tokens on its left, and the unmasked tokens on its right. The unmasked token can be attended to by other unmasked tokens in both directions. In particular, we employ two causal masks: the unidirectional one, where only text token is unmasked and all other tokens are in the masked areas, and the bidirectional one, where text and [END] + +tokens are unmasked, while a random number of motion tokens are put into the masked areas. Consequently, Masked Self-attention Transformer retains the causal aspect of autoregressive motion generation while also being capable of conditioning on future tokens. The output of masked self-attention is as follows: + +$$ +A t t e n t i o n = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}} + \mathrm {M}\right) \cdot V \tag {2} +$$ + +where $\mathrm{Q}$ , $\mathrm{K}$ , and $\mathrm{V}$ , indicate queries, keys, and values respectively while $d_{k}$ represents the dimension of queries and keys. The self-attention mask $M \in \mathbb{R}^{(t + 1)\times (t + 1)}$ is assigned to zero in the positions where attention is allowed, and to negative infinity otherwise. Adding negative infinity forces attention score to be zero after Softmax(·) operation. Therefore, bidirectional causal mask $M^{bc}$ can be written as + +$$ +M _ {i j} = \left\{ \begin{array}{l l} 0, & \text {w h e r e} (i \geq j \wedge i \notin U) \vee (j \in U) \\ - \infty , & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +where $i,j\in [0,1,2,\ldots ,t + 1]$ is the index of query $Q$ and key $K$ . $U = [u_{0},u_{1},\dots ]$ contains the indices of unmasked tokens. The unidirectional causal mask $M^{uc}$ is a special case of the bidirectional one when $U = \emptyset$ . + +# 3.3 Training: Hybrid Attention Masking + +Given a discrete representation of the motion sequence $x_{1:t}$ , our model is trained to reconstruct the motion sequence, conditioned on the text token $x_0$ under both unidirectional and bidirectional causal masking strategies, i.e., $M^{uc}$ and $M^{bc}$ . The reconstruction probability of each motion token under each masking case is $p_{\theta}(x_i \mid M^{uc})$ and $p_{\theta}(x_i \mid M^{bc})$ , respectively. The training objective is to minimize the negative log-likelihood of the motion sequence prediction + +$$ +\mathcal {L} _ {\mathrm {h y b r i d}} = - \underset {\mathbf {X} \in p (\mathbf {X})} {\mathbb {E}} \left[ \lambda \sum_ {\forall i \in [ 1, t ]} \log p _ {\theta} \left(x _ {i} \mid M ^ {u c}\right) + (1 - \lambda) \sum_ {\forall i \in [ 1, t ]} \log p _ {\theta} \left(x _ {i} \mid M ^ {b c}\right) \right]. \tag {4} +$$ + +where $\lambda$ is the probability of selecting unidirectional causal mask. Through experiments, we found $\lambda = 0.5$ yields best performance. In addition, when bidirectional causal mask is selected, we randomly put $50\% - 100\%$ motion tokens in the masked areas. + +# 3.4 Inference: Cascaded Motion Decoding + +To generate motion sequence during inference phase, dual-iteration cascaded decoding is introduced. In the first iteration, autoregressive decoding is applied, where the motion tokens are sequentially and stochastically sampled according to unidirectional prediction distribution $p_{\theta}(x_i \mid M^{uc})$ , concluding upon predicting the [END] token. Since autoregressive decoding could accumulate prediction + +![](images/587bdad98e09ba99338a55c64f5f5647584f6ef15bde944017a33dcbc25559c6.jpg) +Fig. 3: Inference: Dual-iteration Cascaded Motion Decoding. In the first iteration, autoregressive decoding is applied by adopting unidirectional causal mask to generate coarse-grained motion and predict motion sequence length. In the second iteration, bidirectional autoregressive decoding is performed via bidirectional causal mask to removing and repredicting low-confidence motion tokens autoregressively. + +errors from the previously generated tokens, the generated motion sequence is refined in the second iteration by masking out a subset of motion tokens and then resampling these masked tokens according to the bidirectional prediction distribution $p_{\theta}(x_i \mid M^{bc})$ . In this refinement iteration, [END] is positioned where it was predicted in the first iteration. The model can attend to unmasked tokens in all directions to re-predict low-confidence tokens based on rich surrounding context. The choice of masking strategies has an impact on the refinement gain, which is evaluated in Section 6. + +Hybrid Classifier-free Guidance. At training time, we randomly drop textual tokens to teach the model to generate motion unconditionally. During inference, we apply classifier-free guidance (CFG) [20] for cascaded motion decoding. In particular, we generate the final motion sequence by a linear combination of the conditioned logits $\ell_c$ and unconditional logits $\ell_u$ with guidance scale $s$ as + +$$ +\ell_ {g} = (1 + s) \cdot \ell_ {c} - s \cdot \ell_ {u}. \tag {5} +$$ + +We apply different CFG scale $s$ for each iteration during cascaded decoding. The effectiveness of CFG in each iteration is evaluated in Section 5. + +Residual Motion Refinement. To further enhance motion generation quality, the motion sequence yielded from cascaded decoding can be refined by another refinement transformer based on residual vector quantization (RVQ) [44]. By utilizing RVQ, the raw motion sequence is encoded into multiple token sequences in the latent space. Each token sequence is generated by a separate quantizer and each quantizer encodes the quantization error left by the previous quantizer. As a result, the token sequence from the first quantizer encodes the most of information of the original motion sequence. The rest of the token sequences from other quantizers only encode quantization errors. Through RVQ, information loss can be minimized during the embedding quantization process. As a result, our masked self-attention transformer is used to generate the most informative token sequence from the first quantizer. Using this sequence as input, another refinement transformer is trained to predict the remaining token sequences, which are merged into a single token sequence for final motion decoding. This RVQ-based refinement has been adopted by audio generation + +![](images/942fb72941539d12adda89d4bf0d03b8e140b7a9d554dfb51921daabe696239f.jpg) +Fig. 4: Residual Motion Refinement. The residual vector quantization encodes the raw motion sequence into multiple token sequences in different colors (left). The base token sequence from the first vector quantizer is generated via cascaded decoding by masked self-attention transformer. The base token sequence is used as the input of the refinement transformer to predict the residual token sequences from other quantizers. The combined sequences are fed into tokenizer's decoder for motion generation. The refinement transformer shares the same architecture as the masked self-attention transformer with a full attention mask(right). + +models [4, 11, 40] and recently demonstrates its benefits in motion generation task [14]. + +# 3.5 Motion Editability + +The autoregressive approach naturally lacks the ability for temporal motion editing as it cannot leverage future motion tokens. In contrast, BAMM enables temporal motion editing through its bidirectional causal mask, which allows the model to access information from all directions of the conditioned tokens. Consequently, temporal motion editing becomes achievable by merely applying masked attention to the positions that require medications. We illustrate various editing tasks, i.e. motion inpainting (in-betweening), outpainting, prefix prediction, and suffix completion qualitatively and quantitively in Fig. 7 and Table 5. BAMM also allows us to generate the arbitrarily long motion sequence according to a sequence of text prompts, which is showcased in Supplementary Material. + +# 4 Experiments + +In this section, we provide a comprehensive evaluation along with empirical results for our proposed motion generation model, BAMM. In Section 4.1, we adhere to standard evaluation protocols on two datasets, demonstrating that our model outperforms current state-of-the-art methods both quantitatively and qualitatively. Furthermore, in Section 4.2, we showcase our predictive length and editing capabilities in comparison to state-of-the-art methods. The results reveal the robust performance of our model, displaying its effectiveness even in cases without prior information about motion length. + +Datasets. We evaluated our model using the evaluation protocol proposed by [16] for text-driven motion generation with two datasets. KIT-ML dataset contains 3,911 motion sequences, each with one to four textual annotations, + +amounting to a total of 6,278 annotations. These motion sequences are derived from the KIT and CMU motion [1] databases and have been adjusted to 12.5 FPS. The dataset is divided into training, validation, and testing sets, with proportions of $80\%$ , $5\%$ , and $15\%$ respectively. HumanML3D dataset includes a wide variety of human activities, such as exercise and dancing, and consists of 14,616 motion sequences paired with 44,970 textual descriptions. These descriptions come from a vocabulary of 5,371 unique words. The motion sequences, sourced from AMASS [24] and HumanAct12 [17], have been standardized to 20 FPS and are limited to a maximum duration of 10 seconds, with the actual lengths varying between 2 to 10 seconds. Each sequence is accompanied by at least three descriptive annotations, averaging 12 words in length. + +Evaluation Metrics. We adopt the standard evaluation framework from T2M [16], employing pre-trained models that encode text and motion information to evaluate text and motion tokens in the embedding space. R-precision (Top-1, 2, 3 accuracy) measures how well the generated motions align with the text prompts, while Multimodal Distance (MM-Dist) quantifies the distance between generated and ground-truth motions in a shared feature space. Frechet Inception Distance (FID) assesses the statistical similarity between the feature distributions of generated and real motions. Additionally, we evaluate diversity (the average Euclidean distance between random motion pairs) and multimodality (the average variance across Euclidean distances between generated motion pairs for a single prompt) to capture the range of possible motion interpretations and their consistency with the text description. + +# 4.1 Comparison to State-of-the-art Approaches + +Table 2: Comparison of text-conditional motion synthesis on HumanML3D [16] test set. We repeat the evaluation 20 times for each metric and report the average with $95\%$ confidence interval. Red and Blue indicate the best and the second best result. Methods with gray highlight $\S$ report motion generation results using the ground-truth motion length. + +
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Hier [13]0.301±.0020.425±.0020.552±.0046.523±.0245.012±.0188.332±.042-
TEMOS§ [27]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
TM2T [15]0.424±.0030.618±.0030.729±.0021.501±.0173.467±.0118.589±.0762.424±.093
T2M [16]0.455±.0030.636±.0030.736±.0021.087±.0213.347±.0089.175±.0832.219±.074
MDM§ [37]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse§ [46]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD§ [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
Fg-T2M§ [41]0.492±.0020.683±.0030.783±.0020.243±.0193.109±.0079.278±.0721.614±.049
M2DM§ [23]0.497±.0030.682±.0020.763±.0030.352±.0053.134±.0109.926±.0733.587±.072
T2M-GPT [45]0.491±.0030.680±.0030.775±.0020.116±.0043.118±.0119.761±.0811.856±.011
AttT2M [49]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
MMM§ [29]0.515±.0020.708±.0020.804±.0020.089±.0052.926±.0079.577±.0501.226±.035
MoMask§ [14]0.521±.0020.713±.0030.807±.0020.045±.0022.958±.008-1.241±.040
BAMM (ours)0.525±.0020.720±.0030.814±.0030.055±.0022.919±.0089.717±.0891.687±.051
BAMM§ (ours)0.522±.0030.715±.0030.808±.0030.055±.0022.936±.0779.636±.0091.732±.055
+ +Table 3: Comparison of text-conditional motion synthesis on KIT-ML [30] test set. We repeat the evaluation 20 times for each metric and report the average with $95\%$ confidence interval. Red and Blue indicate the best and the second best result. Methods with gray highlight $\S$ report ground-truth motion length for generation. + +
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Hier [13]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.072-
TEMOS$ [27]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
TM2T [15]0.280±.0050.463±.0060.587±.0053.599±.1534.591±.0269.473±.1173.292±.081
T2M [16]0.361±.0060.559±.0070.681±.0073.022±.1073.488±.02810.72±.1452.052±.107
MDM$ [37]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MotionDiffuse$ [46]0.417±.0040.621±.0040.739±.0041.954±.0642.958±.00511.10±.1430.730±.013
MLD$ [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
Fg-T2M$ [41]0.418±.0050.626±.0040.745±.0040.571±.0473.114±.01510.93±.0831.019±.029
M2DM$ [23]0.416±.0040.628±.0040.743±.0040.515±.0293.015±.01711.417±.973.325±.37
T2M-GPT [45]0.402±.0060.619±.0050.737±.0060.717±.0413.053±.02610.86±.0941.912±.036
AttT2M [49]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
MMM$ [29]0.404±.0050.621±.0050.744±.0040.316±.0282.977±.01910.910±.1011.232±.039
MoMask$ [14]0.433±.0070.656±.0050.781±.0050.204±.0112.779±.022-1.131±.043
BAMM (ours)0.438±.0090.661±.0090.788±.0050.183±.0132.723±.02611.008±.0941.609±.065
BAMM (ours)$0.436±.0070.660±.0060.791±.0050.200±.0112.714±.01610.914±.0971.517±.058
+ +Quantitative Results. We evaluate our model on the HumanML3D [16] and KIT-ML [30] datasets, reporting the results in Table 2 and 3, respectively, in comparison to state-of-the-art methods. Following the standard evaluation protocol from [16], we report the average of 20 generations with a $95\%$ confidence interval. Our model consistently outperforms other methods in terms of R-precision, FID, and MM-Distance, while maintaining comparability in terms of Diversity and Multimodal Distance, which typically represent the trade-off between high quality and diversity. This suggests that our model generates very high-quality outputs while retaining a good degree of diversity. Moreover, whereas most models utilize the ground truth length for evaluation, our model supports both predicted length and takes length as an input, outperforming other methods in both scenarios. We further investigate the effect of a separate length estimator which can significantly impact performance in Section 4.2. + +Qualitative Results. Fig. 5 presents a qualitative comparison with T2M-GPT [45], MoMask [14], and MDM [37]. BAMM and T2M-GPT generate motion without requiring input length. Despite this, BAMM accurately generates motion from textual descriptions even without input length. We utilize a pre-trained length estimator from [16] for MoMask [14] and MDM [37]. Notably, BAMM generates motion accurately aligned with the provided text, whereas T2M-GPT and MoMask produce erroneous motion, and MDM generates entirely inaccurate motion. Additional visualizations for BAMM are shown in Fig. 6, indicating its ability to generate high detail such as complex trajectories and interactions with invisible objects. Moreover, Fig. 7 demonstrates BAMM's capability in various temporal editing tasks, including inpainting, outpainting, prefix, and suffix. + +# 4.2 Length Prediction and Editability + +Predict Length vs ground truth length. In real-world scenarios, the ground truth length is often unknown. Many methods cannot predict length autonomously. + +![](images/fb9d9b4266dfb039b411d8c9130d9bea94029c973f00f07f6134f9aaa5e28663.jpg) +Fig. 5: Visualization comparison of textual to motion to state-of-the-art methods. BAMM and T2M-GPT do not require motion length as an input. We use a pre-trained length estimator from [16] for MoMask and MDM. BAMM generates higher quality and is more correlated with textual descriptions. + +![](images/f9a8183e46fdfe1d46c91fee66ac7289637c9c226c43fe5717c09e1ebdec8a1b.jpg) +Fig. 6: Visualization of text-to-motion generation by BAMM. BAMM can generate high-quality motion with complex descriptions, such as intricate trajectories and interactions with invisible objects. + +Consequently, they evaluate using ground truth length, indicated by gray highlight $\S$ in Table 2 and Table 3. We illustrate the impact of inaccurate length prediction by comparing our model to state-of-the-art denoising models that require motion length as an input, such as MoMask and MMM. In particular, both models take the estimated length obtained from a standalone length predictor pretrained on the HumanML3D dataset [16]. Table 4 reveals that the predicted length significantly worsens MoMask's FID score, from 0.045 to 0.090, while MMM's R-precision Top1 drops from 0.515 to 0.504. In contrast, our BAMM not only maintains the best performance but also outperforms MMM and MoMask in both scenarios. + +Table 4: Comparison of text-conditional motion synthesis using predicted and ground truth length on HumanML3D [16] dataset + +
LengthMethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
GroundMMM0.5150.7080.8040.0892.9269.5771.226
TruthMoMask0.5210.7130.8070.0452.958-1.241
BAMM (Ours)0.5220.7150.8080.0552.9369.6361.732
PredictedMMM0.5040.6960.7940.0802.9989.4111.164
MoMask0.5220.7150.8110.0902.9459.6471.239
BAMM (Ours)0.5250.7200.8140.0552.9199.7171.687
+ +![](images/9294446434a39066cbf67430dd738105598b9d0308a9f9dde20d04c5e6ab1cde.jpg) +Fig. 7: Visualization of temporal editing tasks, inpainting (in-betweening), outpainting, prefix, and suffix where blue indicates conditioned motion and red refers to generated parts. + +Editability. We conduct experiments in four temporal editing tasks, namely inpainting (in-betweening), outpainting, prefix, and suffix, comparing them with MDM [37] and MoMask [14] on the HumanML3D dataset. Inpainting is evaluated by generating $50\%$ of the motion sequence given the first and last $25\%$ . Outpainting is its opposite counterpart. Prefix is conditioned by the first $50\%$ of the ground truth motion and generates the remaining portion. Similarly, suffix operates in the reverse manner. Note that editing tasks require prior knowledge of conditioned motion and motion length, for which MDM and MoMask are specifically designed. Despite this, Table 5 shows that our model significantly outperforms MDM in all editing tasks and surpasses MoMask in outpainting, while performing equally well in the other tasks. + +Table 5: Evaluation on temporal editing tasks on HumanML3D [16] dataset + +
TasksMethodsR-Precision ↑FID ↓ MM-Dist ↓ Diversity ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Temporal Inpainting (In-betweenging)MDM0.3910.5780.6922.3623.8598.014
MoMask0.5340.7270.820.042.8789.64
BAMM0.5350.7290.8210.0562.8639.629
Temporal OutpaintingMDM0.4150.6130.7272.0573.6198.199
MoMask0.5310.7260.8180.0572.8899.619
BAMM0.5350.730.8220.0562.8569.659
PrefixMDM0.420.6130.7251.463.5638.312
MoMask0.5360.730.8220.062.8759.607
BAMM0.5320.7270.8210.0582.8689.612
SuffixMDM0.4030.5970.7112.5623.7318.088
MoMask0.5320.7260.8190.0522.8819.659
BAMM0.5270.720.8140.052.8919.721
+ +# 5 Ablation Study + +In the ablation study, we study the impacts of the adaptive classifier-free guidance scales in each iteration of the Masked Self-attention Transformer, the masking strategy, and the number of iterations, as shown in Table 6. + +Adaptive Classifier-Free Guidance Scales (CFG): We conducted experiments with different CFG values for both the first and second iterations and found that CFG=4 in the first iteration and CFG=3 in the second iteration works best. Although CFG=5 in the first iteration yields a better FID score, the other scores are worse. Mask Strategy: The first strategy, "50% of low confidence", applies mask attention on 50% of the lowest confident positions from the + +Table 6: Ablation Study + +
AblationsTypesR-Precision ↑FID ↓ MM-Dist ↓ Diversity ↑ MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
CFG of 1st iteration1st iter2nd iter
CFG=2CFG=30.5170.7110.8050.1052.9569.841.907
CFG=3CFG=30.5210.7140.8080.072.9449.7771.766
CFG=4CFG=30.5250.720.8140.0552.9199.7171.687
CFG=5CFG=30.5220.7160.810.0522.9279.6471.691
CFG=6CFG=30.520.7130.810.062.949.6211.697
CFG of 2nd iterationCFG=4CFG=10.5210.7160.810.0582.9439.7431.744
CFG=4CFG=20.5240.7190.8140.0572.9249.7521.698
CFG=4CFG=30.5250.720.8140.0552.9199.7171.687
CFG=4CFG=40.5220.7190.8120.0562.9249.6911.698
CFG=4CFG=50.5210.7170.8120.0652.9319.6381.735
Mask Strategy50% of low confidence confidence < .50.5250.720.8130.0652.9219.7321.67
suffix0.5250.7180.810.0642.9239.7651.656
%2=00.5190.7150.810.0522.9439.6831.841
0.5250.720.8140.0552.9199.7171.687
# of iterations1 iteration0.5240.7180.8120.0642.9269.7201.644
2 iterations0.5250.7200.8140.0552.9199.7171.687
3 iterations0.5250.7190.8140.0552.9179.7271.69
+ +first iteration. Similarly, "confidence $< .5$ " uses low confidence but sets a threshold, masking out positions where confidence is below 0.5. Specifically, the former strictly masks out $50\%$ of the total token sequence while the latter masks the token with a confidence lower than .5. "Suffix" strategy masks out the first $50\%$ of the token sequence, utilizing the remaining $50\%$ as condition tokens. Lastly, $\% 2 = 0$ "strategy masks every other token. BAMM produces decent performances across the tested masking strategies while the simple masking of every other token works the best. Number of Iterations: We conducted experiments with one to three iterations. "1 iteration" refers to the first iteration, as described in 3.4, which utilizes a unidirectional causal mask. Similarly, "2 iterations" involves applying bidirectional causal mask to the Masked Self-attention Transformer to re-predict the tokens from the first iteration, as illustrated in 3. In "3 iterations", we repeat the second iteration (bidirectional causal mask) but apply masks to $1/3$ of the sequence to re-predict the motion tokens instead. The experiments indicate that "2 iterations" clearly demonstrate improvement over one iteration, while "3 iterations" do not significantly improve upon "2 iterations". This suggests that "2 iterations" are sufficient. + +# 6 Conclusion + +We introduce the Bidirectional Autoregressive Motion Model (BAMM), a novel framework for text-to-motion generation. BAMM combines a motion tokenizer, which encodes 3D human motion into discrete latent tokens, with a masked self-attention transformer that autoregressively predicts the masked tokens through a masked casual attention approach. BAMM integrates of generative masked and autoregressive modeling into an unified framework, This features allows it to understand complex motion relationships and precisely map text inputs to high-quality motion outputs with adaptively adjusted sequence lengths. Our extensive testing on HumanML3D and KIT-ML datasets confirms BAMM's superiority in both qualitative and quantitative evaluations over existing methods. + +# References + +1. Cmu graphics lab motion capture database, http://mocap.cs.cmu.edu/, accessed: 2022-11-11 10 +2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728 (2019). https://doi.org/10.1109/3DV.2019.000844 +3. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems 34, 17981-17993 (2021) 4 +4. Borsos, Z., Sharifi, M., Vincent, D., Kharitonov, E., Zeghidour, N., Tagliasacchi, M.: Soundstorm: Efficient parallel audio generation. ArXiv abs/2305.09636 (2023), https://api-semanticscholar.org/CorpusID:258715176 9 +5. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T.J., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. ArXiv abs/2005.14165 (2020), https://api_semanticscholar.org/CorpusID:218971783 2, 5 +6. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M., Murphy, K.P., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D.: Muse: Text-to-image generation via masked generative transformers. ArXiv abs/2301.00704 (2023), https://api(semanticscholar.org/CorpusID:255372955 4 +7. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11305-11315 (2022), https://apisemantic scholar.org/CorpusID:246680316 4 +8. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, J., Yu, G.: Executing your commands via motion diffusion in latent space. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 18000-18010 (2022), https://api-semanticscholar.org/CorpusID:254408910 4, 10, 11 +9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019), https://apisemantic scholar.org/CorpusID:52967399 4 +0. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. ArXiv abs/2204.14217 (2022), https://api(semanticscholar.org/CorpusID:248476190 4 +1. Du, Z., Qian, Y., Liu, X., Ding, M., Qiu, J., Yang, Z., Tang, J.: Glm: General language model pretraining with autoregressive blank infilling. In: Annual Meeting of the Association for Computational Linguistics (2021), https://apisemantic scholar.org/CorpusID:247519241 5, 9 +2. Ghazvininejad, M., Levy, O., Liu, Y., Zettlemoyer, L.: Mask-predict: Parallel decoding of conditional masked language models. In: Conference on Empirical Methods in Natural Language Processing (2019), https://api_semanticscholar.org/ CorpusID:202538740 4 +3. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 1376-1386 (2021), https://apisemantic scholar.org/CorpusID:232404671 10, 11 + +14. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions (2023) 2, 4, 6, 9, 10, 11, 13 +15. Guo, C., Xuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. ArXiv abs/2207.01696 (2022), https://api_semanticscholar.org/CorpusID: 250280248 10, 11 +16. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5142-5151 (2022). https://doi.org/10.1109/CVPR52688.2022.005093, 4, 5, 9, 10, 11, 12, 13, 2 +17. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. Proceedings of the 28th ACM International Conference on Multimedia (2020), https://apisemantic scholar.org/CorpusID:220870974 10 +18. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A.A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models. ArXiv abs/2210.02303 (2022), https://api_semanticscholar.org/CorpusID:252715883 4 +19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. ArXiv abs/2006.11239 (2020), https://apisemantic scholar.org/CorpusID: 219955663 4 +20. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 8 +21. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. ArXiv abs/2306.14795 (2023), https://api.sementicscholar.org/CorpusID:259262201 2, 5 +22. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI Conference on Artificial Intelligence (2022), https://api-semanticscholar.org/CorpusID:251979380 2, 4 +23. Kong, H., Gong, K., Lian, D., Mi, M.B., Wang, X.: Priority-centric human motion generation in discrete latent space. ArXiv abs/2308.14480 (2023), https://api.sementicscholar.org/CorpusID:261245369 10, 11 +24. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 5441-5450 (2019), https://apisemanticscholar.org/CorpusID:102351100 10 +25. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning (2021), https://api-semanticscholar.org/CorpusID:245335086 4 +26. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning. ArXiv abs/1711.00937 (2017), https://api-semanticscholar.org/CorpusID:20282961 3, 6 +27. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. ArXiv abs/2204.14109 (2022), https://api.sementicscholar.org/CorpusID:248476220 4, 10, 11 +28. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. ArXiv abs/2305.00976 (2023), https://api_semanticscholar.org/CorpusID:258436810 4 + +29. Pinyoanuntapong, E., Wang, P., Lee, M., Chen, C.: Mmm: Generative masked motion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 2, 4, 6, 10, 11 +30. Plappert, M., Mandery, C., Asfour, T.: The KIT motion-language dataset. Big Data 4(4), 236-252 (dec 2016). https://doi.org/10.1089/big.2016.0028, http://dx.doi.org/10.1089/big.2016.0028 4, 11, 5 +31. Qian, L., Zhou, H., Bao, Y., Wang, M., Qiu, L., Zhang, W., Yu, Y., Li, L.: Glancing transformer for non-autoregressive neural machine translation. In: Annual Meeting of the Association for Computational Linguistics (2020), https://api_semanticscholar.org/CorpusID:221150562 4 +32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021), https://api_semanticscholar.org/CorpusID: 231591445 4, 6 +33. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. ArXiv abs/2205.11487 (2022), https://api_semanticscholar.org/CorpusID:248986576 4 +34. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data. ArXiv abs/2209.14792 (2022), https://api_semanticscholar.org/CorpusID:252595919 4 +35. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ArXiv abs/2010.02502 (2020), https://api.sementicscholar.org/CorpusID:222140788 4 +36. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision (2022), https://api-semanticscholar.org/CorpusID:247450907 4 +37. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. ArXiv abs/2209.14916 (2022), https://api.sementicscholar.org/CorpusID:252595883 2, 4, 10, 11, 13 +38. Vaswani, A., Shazeer, N.M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:137564895 +39. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. ArXiv abs/2210.02399 (2022), https://apisemantic scholar.org/CorpusID:252715594 4 +40. Wang, C., Chen, S., Wu, Y., Zhang, Z.H., Zhou, L., Liu, S., Chen, Z., Liu, Y., Wang, H., Li, J., He, L., Zhao, S., Wei, F.: Neural codec language models are zero-shot text to speech synthesizers. ArXiv abs/2301.02111 (2023), https://api.sementicscholar.org/CorpusID:255440307 9 +41. Wang, Y., Leng, Z., Li, F.W.B., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. ArXiv abs/2309.06284 (2023), https://api.sementicscholar.org/CorpusID:261697123 2, 10, 11 +42. Yan, S., Liu, Y., Wang, H., Du, X., Liu, M., Liu, H.: Cross-modal retrieval for motion and text via doptriple loss (2023), https://api-semanticscholar.org/CorpusID:263610212 4 + +43. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., Hutchinson, B.C., Han, W., Parekh, Z., Li, X., Zhang, H., Baldridge, J., Wu, Y.: Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res. 2022 (2022), https://api_semanticscholar.org/CorpusID:249926846 4 +44. Zeghidour, N., Luebs, A., Omran, A., Skoglund, J., Tagliasacchi, M.: Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30, 495-507 (2021), https://api-semanticscholar.org/CorpusID:236149944 8 +45. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: Generating human motion from textual descriptions with discrete representations. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14730-14740 (2023), https://api_semanticscholar.org/CorpusID:255942203 2, 5, 6, 10, 11 +46. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. ArXiv abs/2208.15001 (2022), https://api_semanticscholar.org/CorpusID: 251953565 2, 4, 10, 11 +47. Zhang, Z., Ma, J., Zhou, C., Men, R., Li, Z., Ding, M., Tang, J., Zhou, J., Yang, H.: M6-ufc: Unifying multi-modal controls for conditional image synthesis via non-autoregressive generative transformers (2021), https://api_semanticscholar.org/CorpusID:237204528 4 +48. Zhang, Z., Ma, J., Zhou, C., Men, R., Li, Z., Ding, M., Tang, J., Zhou, J., Yang, H.: Ufc-bert: Unifying multi-modal controls for conditional image synthesis. In: Neural Information Processing Systems (2021), https://apisemantic scholar.org/CorpusID:235253928 4 +49. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. ArXiv abs/2309.00796 (2023), https://api.sementicscholar.org/CorpusID:261530775 2, 5, 10, 11 \ No newline at end of file diff --git a/2024/BAMM_ Bidirectional Autoregressive Motion Model/images.zip b/2024/BAMM_ Bidirectional Autoregressive Motion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fd7a1341bf8c01ef1fca1dad33d254474956a0eb --- /dev/null +++ b/2024/BAMM_ Bidirectional Autoregressive Motion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6b815a3f902511fe41838e2f64d0a0403e1a44b88a37ad0ddaddf2eb3347d69 +size 630399 diff --git a/2024/BAMM_ Bidirectional Autoregressive Motion Model/layout.json b/2024/BAMM_ Bidirectional Autoregressive Motion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4bb1d3cf80901812a1e3012e2203136936154565 --- /dev/null +++ b/2024/BAMM_ Bidirectional Autoregressive Motion Model/layout.json @@ -0,0 +1,8022 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 145 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 145 + ], + "type": "text", + "content": "BAMM: Bidirectional Autoregressive Motion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 168, + 468, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 168, + 468, + 194 + ], + "spans": [ + { + "bbox": [ + 145, + 168, + 468, + 194 + ], + "type": "text", + "content": "Ekkasit Pinyoanuntapong1, Muhammad Usama Saleem1, Pu Wang1, Minwoo Lee1, Srijan Das1, and Chen Chen2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 201, + 463, + 236 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 148, + 201, + 463, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 201, + 463, + 223 + ], + "spans": [ + { + "bbox": [ + 148, + 201, + 463, + 223 + ], + "type": "text", + "content": "1 University of North Carolina at Charlotte {epinyoan,msaleem2, Pu.Wang, minwoo.lee, sdas24}@uncc.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 224, + 424, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 224, + 424, + 236 + ], + "spans": [ + { + "bbox": [ + 189, + 224, + 424, + 236 + ], + "type": "text", + "content": "2 University of Central Florida chen.chen@crcv.ucf.edu" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 135, + 263, + 217, + 338 + ], + "blocks": [ + { + "bbox": [ + 136, + 255, + 212, + 262 + ], + "lines": [ + { + "bbox": [ + 136, + 255, + 212, + 262 + ], + "spans": [ + { + "bbox": [ + 136, + 255, + 212, + 262 + ], + "type": "text", + "content": "a) Motion Length Prediction" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 263, + 217, + 338 + ], + "lines": [ + { + "bbox": [ + 135, + 263, + 217, + 338 + ], + "spans": [ + { + "bbox": [ + 135, + 263, + 217, + 338 + ], + "type": "image", + "image_path": "7ada26cc70a325b7862be6684fdc5a829b2bccc205a06f6c55132ff335e6673d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 221, + 255, + 313, + 273 + ], + "blocks": [ + { + "bbox": [ + 221, + 255, + 313, + 273 + ], + "lines": [ + { + "bbox": [ + 221, + 255, + 313, + 273 + ], + "spans": [ + { + "bbox": [ + 221, + 255, + 313, + 273 + ], + "type": "image", + "image_path": "4664df2cc75ea62b3c0b5fd7146c0bb9c31fdcbef7f2b93467c1cf2fe01760d6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 224, + 274, + 313, + 335 + ], + "blocks": [ + { + "bbox": [ + 224, + 274, + 313, + 335 + ], + "lines": [ + { + "bbox": [ + 224, + 274, + 313, + 335 + ], + "spans": [ + { + "bbox": [ + 224, + 274, + 313, + 335 + ], + "type": "image", + "image_path": "cbbc3c6b7b3e0e524c61171c8e742a57eb7bda0e64be05a2de188aed2007ca99.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 346, + 482, + 434 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 434 + ], + "type": "text", + "content": "Fig. 1: (a) Motion Length Prediction: Text-to-motion models often require specific input lengths, making them sensitive to motion generation. In contrast, BAMM automatically predicts the end of the motion, thus avoiding reliance on inaccurate motion length estimations. (b) High-quality Text-to-Motion: BAMM generates natural human movements precisely aligned with detailed textual descriptions. (c) Motion Editing: BAMM is capable of multiple editing tasks, such as inpainting (as demonstrated), outpainting, prefix prediction, suffix completion, and arbitrarily long motion sequence synthesis." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 332, + 264, + 478, + 329 + ], + "blocks": [ + { + "bbox": [ + 385, + 255, + 432, + 263 + ], + "lines": [ + { + "bbox": [ + 385, + 255, + 432, + 263 + ], + "spans": [ + { + "bbox": [ + 385, + 255, + 432, + 263 + ], + "type": "text", + "content": "c) Motion Editing" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 264, + 478, + 329 + ], + "lines": [ + { + "bbox": [ + 332, + 264, + 478, + 329 + ], + "spans": [ + { + "bbox": [ + 332, + 264, + 478, + 329 + ], + "type": "image", + "image_path": "ff1e38a1bdc1f8d193e019b03893cea99147415a4afa7af04fae0ec07277b385.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 159, + 445, + 453, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 445, + 453, + 665 + ], + "spans": [ + { + "bbox": [ + 159, + 445, + 453, + 665 + ], + "type": "text", + "content": "Abstract. Generating human motion from text has been dominated by denoising motion models either through diffusion or generative masking process. However, these models face great limitations in usability by requiring prior knowledge of the motion length. Conversely, autoregressive motion models address this limitation by adaptively predicting motion endpoints, at the cost of degraded generation quality and editing capabilities. To address these challenges, we propose Bidirectional Autoregressive Motion Model (BAMM), a novel text-to-motion generation framework. BAMM consists of two key components: (1) a motion tokenizer that transforms 3D human motion into discrete tokens in latent space, and (2) a masked self-attention transformer that autoregressively predicts randomly masked tokens via a hybrid attention masking strategy. By unifying generative masked modeling and autoregressive modeling, BAMM captures rich and bidirectional dependencies among motion tokens, while learning the probabilistic mapping from textual inputs to motion outputs with dynamically-adjusted motion sequence length. This feature enables BAMM to simultaneously achieving high-quality motion generation with enhanced usability and built-in motion editability. Extensive experiments on HumanML3D and KIT-ML datasets demonstrate that BAMM surpasses current state-of-the-art methods in both" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 452, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 452, + 139 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 452, + 139 + ], + "type": "text", + "content": "qualitative and quantitative measures. Our project page is available at https://exitudio.github.io/BAMM-page" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 149, + 453, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 149, + 453, + 171 + ], + "spans": [ + { + "bbox": [ + 160, + 149, + 453, + 171 + ], + "type": "text", + "content": "Keywords: Text to Motion " + }, + { + "bbox": [ + 160, + 149, + 453, + 171 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 149, + 453, + 171 + ], + "type": "text", + "content": " Autoregressive Motion Model " + }, + { + "bbox": [ + 160, + 149, + 453, + 171 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 149, + 453, + 171 + ], + "type": "text", + "content": " Generative Masked Motion Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 197, + 230, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 230, + 210 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 230, + 210 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 222, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 482, + 353 + ], + "type": "text", + "content": "Text-to-motion generation is a promising interdisciplinary field that uses natural language to generate 3D human movements. This emerging field holds vast potential to transform animation, gaming, filming, and VR/AR/MR domains by enabling easy and intuitive creation of 3D assets through user-friendly textual inputs. However, bridging the semantic gap between textual descriptions and intricate motion sequences presents a significant challenge. To address this challenge, recent efforts have been focusing on two methods: (1) conditional denoising motion model and (2) conditional autoregressive motion model. Both methods can greatly improve motion generation quality by learning the probabilistic distribution of motion sequences, conditioned on the textural descriptors. However, both methods face fundamental limitations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "content": "Conditional denoising motion models are trained to restore corrupted motion sequences to their original state, guided by textual prompts. These models operate through two primary mechanisms: diffusion and generative masking. Diffusion models apply structured Gaussian noise to the original motion data for corruption [22,37,41,46], whereas generative masked models [14,29] corrupt motion sequence by substituting selected motion tokens with [MASK] tokens. The denoising process, followed by motion corruption procedure, considers motion tokens from both directions, effectively capturing the intricate dependencies among tokens. This leads to enhanced motion generation quality. Furthermore, these models inherently ease the motion editing tasks. By selectively corrupting and recovering motion tokens in areas needing modifications, they ensure seamless transitions between edited and unedited segments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "While denoising motion models excel in generation quality and editability, a significant limitation of these models is their usability because these models depend on prior knowledge of motion length for each text prompt, a requirement that proves impractical in real-world scenarios. Utilizing incorrect motion lengths can result in a significant decline in generation quality. To address this fundamental limitation, conditional motion autoregressive models emerge as a solution, capable of simultaneously predicting the sequence length and content of generated motions. Inspired by large language models like GPT [5], motion autoregressive models sequentially predict one motion token at a time from left to right until the [END] token is predicted, guided by the textual description [21, 45, 49]. As a result, the generated motions are not only well aligned with the text inputs but also appropriately scaled in duration. However, the sequential token decoding of autoregressive models cannot fully capture the dependencies between the motion tokens, potentially compromising generation quality, and complicating" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "text", + "content": "E. Pinyoanuntapong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "the editing process, because edited parts need to be conditioned on the unedited parts from both directions to ensure overall continuity and coherence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 380 + ], + "type": "text", + "content": "As previously highlighted, existing text-driven motion generation models encounter a comprise among usability, quality of generation, and editability. To address this challenge, we propose Bidirectional Autoregressive Motion Model (BAMM), a novel text-to-motion generation framework. It consists of two pivotal components: a motion tokenizer and a conditional masked self-attention transformer. In a two-stage training paradigm, the motion tokenizer performs initial training based on Vector Quantized Variational Autoencoders (VQ-VAE) [26]. This tokenizer encodes the raw motion sequence into discrete motion tokens within the latent space leveraging learned codebooks. In the subsequent phase, motion tokens are randomly masked out. A conditional masked self-attention transformer is then trained to autoregressively predict these masked tokens, adopting the causal attention masking strategy. This strategy is largely departing from the traditional [MASK] token replacement approach in generative masked models. In particular, it does not substitute the input motion tokens with [MASK] tokens. Instead, it adjusts the attention score matrix according to both unidirectional and bidirectional causal masks. The unidirectional causal mask enables adaptive prediction of [END] token based on text prompts, while bidirectional causal mask forces the model to predict the next motion token not only based on past tokens but also conditioned on future unmasked tokens. This facilitates bidirectional autoregressive training for enhanced predictive capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 502 + ], + "type": "text", + "content": "Table 1: Comparison of quality and capability of generation on text-to-motion to state-of-the-art models on the largest text-to-motion dataset [16]. " + }, + { + "bbox": [ + 130, + 392, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 130, + 392, + 482, + 502 + ], + "type": "text", + "content": " means capability while " + }, + { + "bbox": [ + 130, + 392, + 482, + 502 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 392, + 482, + 502 + ], + "type": "text", + "content": " is not. \"Predict Length\" denotes the ability to generate motion without prior knowledge of motion length. \"Input Length\" refers to the ability to take input length as a constraint, while \"Edit\" indicates motion editability. Since MMM and MoMask require ground-truth motion length as input, we use predicted motion length from pretrained length estimator by [16]. The lowest FID score means the best overall quality of the generated motion, ensuring that its authenticity and naturalness is very close to the ground-truth human movements. High R-precision and low MM-dist means accurate alignment between the generated motion and the text prompts." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 192, + 511, + 421, + 569 + ], + "blocks": [ + { + "bbox": [ + 192, + 511, + 421, + 569 + ], + "lines": [ + { + "bbox": [ + 192, + 511, + 421, + 569 + ], + "spans": [ + { + "bbox": [ + 192, + 511, + 421, + 569 + ], + "type": "table", + "html": "
MethodsTop-1 ↑FID ↓MM-Dist ↓Predict LengthInput LengthEdit
T2M-GPT0.4910.1163.118✓✗✗
AttT2M0.4990.1123.038✓✗✗
MMM0.5040.0802.998✗✓✓
MoMask0.5220.0902.945✗✓✓
BAMM (Ours)0.5250.0552.919✓✓✓
", + "image_path": "2a2aa0787a8b78d64dab04b95050582396014ca6c6c133ffd58506d27e7d6419.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "By unifying masked and autoregressive prediction during training, BAMM captures rich and bidirectional dependencies among motion tokens, while learning a direct probabilistic mapping from textual inputs to motion outputs with dynamically-adjusted motion sequence length. Leveraging such unique feature, we propose cascaded motion generation during inference, where BAMM first leverages unidirectional autoregressive decoding to implicitly predict motion se" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "quence length and generate coarse-grained motion sequence. Such motion sequence is then refined by masking and regenerating a portion of motion tokens in a bidirectional autoregressive fashion. This feature allows BAMM to achieve high-quality motion generation with high usability. Moreover, BAMM naturally supports zero-shot motion editing without specially being trained for such task. By treating the masked motion tokens as the contents that need editing, BAMM can predict the masked tokens based on the surrounding context and the text description. Our main contributions can be summarized as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 223, + 479, + 365 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 223, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 479, + 281 + ], + "type": "text", + "content": "- We introduce the bidirectional autoregressive motion model, which is a novel text-to-motion generation framework. It effectively harnesses the complementary benefits of denoising and autoregressive models, thus simultaneously achieving high-quality motion generation with enhanced usability and innate motion editability, as showcased in Fig. 1 and Table 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 283, + 479, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 479, + 318 + ], + "type": "text", + "content": "- We demonstrate that our model outperforms current state-of-the-art methods qualitatively and quantitatively on two standard text-to-motion generation datasets, HumanML3D [16] and KIT-ML [30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 319, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 479, + 365 + ], + "type": "text", + "content": "- We showcase that our model supports a variety of motion editing tasks in the zero-shot manner without specially training for these tasks, including motion inpainting, outpainting, prefix prediction, suffix completion, and long sequence generation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 387, + 237, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 237, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 237, + 399 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": "Motion Synthesis with Latent Space Alignment. Early text-to-motion approaches typically adopt a two-stage scheme: learning separate latent representations for text and motion sequences, followed by latent space alignment using distance losses like cosine similarity or KL divergence [2,16,27,28,36,42]. For example, Language2Pose [2] aimed to establish a shared latent space for both language descriptions and motion sequences. MotionCLIP [36] simplified this concept by incorporating stylization and diversity techniques, leveraging the pre-trained text-image latent space of the CLIP model [32]. However, this strategy inherently struggles with generating high-fidelity motions due to the difficulty of perfectly aligning these inherently disparate latent spaces." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Conditional Denoising Motion Model. Inspired by the success of denoising diffusion models (DDMs) [19,35] in text-to-image and text-to-video generation [18, 25, 33, 34], diffusion models have been applied for text-to-motion generation. MDM [37], MotionDiffuse [46], MLD [8], and FLAME [22] are the representative examples. Meanwhile, BERT-type masked generative models have shown their success in both text generation tasks, e.g., Q&A and language translations [9,12,31], and text-to-image synthesis [6,7,10,39,43,47,48]. Following this trend, MMM [29] and MoMask [14] are recent attempts to propose conditional masked motion modeling to enable high-fidelity motion synthesis that is precisely aligned with text prompts. Both diffusion and generative masked modeling share the same principle: \"denoising\" corrupted data [3]. Meanwhile, they also face the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "text", + "content": "E. Pinyoanuntapong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "content": "same usability limitation: they require prior knowledge of the motion sequence length for the denoising process. This length information is obtained either by intuitive guess from users or via a separately pre-trained predictor that predicts the motion length distribution based on the input text prompts [16]. Both approaches, however, lead to significant motion quality degradation as demonstrated in the experiments section. This is because the distributions of motion length and motion contents are inherently coupled, which need to be jointly aligned with the textual context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 357 + ], + "type": "text", + "content": "Conditional Autoregressive Motion Model. The autoregressive motion models, such as T2M-GPT [45], AttT2M [49], and MotionGPT [21], can effectively address the usability challenge faced by denoising models because they follow the GPT-type training and inference [5] to implicitly predict motion sequence length by generating the [END] token conditioned on both previously generated motion tokens and text inputs. However, the limitation of these models lies in their use of causal attention for unidirectional and sequential motion token prediction. This practice not only hinders model's motion editability but also jeopardizes motion generation quality. To address this limitation, we propose the first bidirectional autoregressive modeling approach for human motion generation, which draws inspiration from the self-attention masking employed by large language model pretraining [11,38]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 376, + 202, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 202, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 202, + 388 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 411, + 465, + 517 + ], + "blocks": [ + { + "bbox": [ + 134, + 411, + 465, + 517 + ], + "lines": [ + { + "bbox": [ + 134, + 411, + 465, + 517 + ], + "spans": [ + { + "bbox": [ + 134, + 411, + 465, + 517 + ], + "type": "image", + "image_path": "84c38308060370eb216a17d42ed87844ef9168e44236fe87b0e6ecfd71087d4f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 522, + 482, + 588 + ], + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 588 + ], + "type": "text", + "content": "Fig.2: Overall architecture of BAMM. (a) Motion Tokenizer encodes the raw motion sequence into discrete motion tokens according to a learned codebook. (b) Masked Self-attention Transformer learns to sequentially predict next tokens conditioned on text embedding from CLIP model and future unmasked tokens. Masked self-attention mechanism unifies autoregressive model and generative masked motion via bidirectional and unidirectional causal masks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Our objective is to create a text-to-motion synthesis framework that simultaneously achieves high-quality motion generation with enhanced usability and innate motion editability. Towards this goal, our framework, as illustrated in Fig. 2, consists of two key components: motion tokenizer that compresses and converts raw 3D human motion into a sequence of discrete motion tokens in the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "latent space (Section 3.1) and conditional masked self-attention transformer that leverages unidirectional and bidirectional casual masks to integrate autoregressive model and generative masked model into a unified framework (Section 3.2). The training procedure follows a hybrid attention masking strategy, where the two causal masks are applied randomly and the model is forced to reconstruct the motion sequence under both cases (Section 3.3). The cascaded motion decoding is introduced for motion generation during the inference phase. It uses unidirectional autoregressive decoding to jointly predict motion sequence length and its contents, which are refined via bidirectional autoregressive decoding (Section 3.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 253, + 250, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 253, + 250, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 250, + 264 + ], + "type": "text", + "content": "3.1 Motion Tokenizer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": "The objective of this stage is to learn the discrete representation of motion by quantizing the embedding " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " from the output of the encoder into codebook " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": ". We first pretrain a motion tokenizer based on VQ-VAE [26]. In particular, given a motion sequence " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = [m_1, m_2, m_3, \\dots, m_\\tau]" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "m \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " is the total frames of motion, and " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " is the dimension of the 3D pose in each frame, encoder is used to encode motion " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " to the latent embedding " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{t \\times d}" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " with a downsampling rate of " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\tau / t" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": ". The embedding " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " is quantized into codes " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": ". Codebook " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\gamma_k\\}_{k=1}^K" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " number of codes. The nearest Euclidean distance between the embedding " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": " and the code of vector is computed by " + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\hat{z}_i = \\mathrm{argmin}_j \\| \\mathbf{z} - \\mathcal{C}_j \\|_2^2" + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": ". The loss function is defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 403, + 480, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 403, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 225, + 403, + 480, + 416 + ], + "type": "interline_equation", + "content": "L _ {V Q} = \\left\\| \\operatorname {s g} (\\mathbf {z}) - \\mathbf {e} \\right\\| _ {2} ^ {2} + \\beta \\| \\mathbf {z} - \\operatorname {s g} (\\mathbf {e}) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "78b6f9e6e852a48ce18b809c132dd8107f764459165d754fc4e286e3f276d246.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 425, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(\\cdot)" + }, + { + "bbox": [ + 130, + 425, + 482, + 474 + ], + "type": "text", + "content": " is the stop-gradient operator, " + }, + { + "bbox": [ + 130, + 425, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 425, + 482, + 474 + ], + "type": "text", + "content": " refers the hyper-parameter for commitment loss. The loss function is optimized via a straight-through gradient estimator. We apply exponential moving average for codebooks update and codebook reset by following [45] [14] [29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 491, + 399, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 399, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 399, + 502 + ], + "type": "text", + "content": "3.2 Conditional Masked Self-attention Transformer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Our model employs a standard multi-layer transformer, whose inputs are the concatenation of the motion tokens " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{1:t}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " from the tokenizer with " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " as the sequence length, the text embedding " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " from the pre-trained CLIP model [32], and the [END] token " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{t + 1}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " that serves as the indicator of the motion's endpoint. The input tokens " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{0:t + 1}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " are masked out strategically. Different from generative masked models, we do not replace the input tokens with [MASK] ones. Instead, we adopt causal attention mask " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " as shown in Fig. 2 (b) to specify the attention relations among the input tokens. In particular, the token in the masked areas, indicated by " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\blacksquare" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", can be attended to itself, all the tokens on its left, and the unmasked tokens on its right. The unmasked token can be attended to by other unmasked tokens in both directions. In particular, we employ two causal masks: the unidirectional one, where only text token is unmasked and all other tokens are in the masked areas, and the bidirectional one, where text and [END]" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "text", + "content": "E. Pinyoanuntapong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "tokens are unmasked, while a random number of motion tokens are put into the masked areas. Consequently, Masked Self-attention Transformer retains the causal aspect of autoregressive motion generation while also being capable of conditioning on future tokens. The output of masked self-attention is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 218, + 173, + 481, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 173, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 218, + 173, + 481, + 200 + ], + "type": "interline_equation", + "content": "A t t e n t i o n = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} + \\mathrm {M}\\right) \\cdot V \\tag {2}", + "image_path": "bebbf165e610fd93a655d244eacd7c7255a6a69b05efece3414749cd7aad2477.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}" + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{K}" + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{V}" + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "text", + "content": ", indicate queries, keys, and values respectively while " + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "text", + "content": " represents the dimension of queries and keys. The self-attention mask " + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{(t + 1)\\times (t + 1)}" + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "text", + "content": " is assigned to zero in the positions where attention is allowed, and to negative infinity otherwise. Adding negative infinity forces attention score to be zero after Softmax(·) operation. Therefore, bidirectional causal mask " + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "inline_equation", + "content": "M^{bc}" + }, + { + "bbox": [ + 131, + 203, + 482, + 275 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 282, + 481, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 282, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 206, + 282, + 481, + 308 + ], + "type": "interline_equation", + "content": "M _ {i j} = \\left\\{ \\begin{array}{l l} 0, & \\text {w h e r e} (i \\geq j \\wedge i \\notin U) \\vee (j \\in U) \\\\ - \\infty , & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}", + "image_path": "a7d540fbf4f60ee1bb192d4b2b34697f796814d67e4683f272ea50e741f25484.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "inline_equation", + "content": "i,j\\in [0,1,2,\\ldots ,t + 1]" + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "text", + "content": " is the index of query " + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "inline_equation", + "content": "U = [u_{0},u_{1},\\dots ]" + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "text", + "content": " contains the indices of unmasked tokens. The unidirectional causal mask " + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "inline_equation", + "content": "M^{uc}" + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "text", + "content": " is a special case of the bidirectional one when " + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "inline_equation", + "content": "U = \\emptyset" + }, + { + "bbox": [ + 131, + 316, + 482, + 353 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 370, + 345, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 370, + 345, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 345, + 383 + ], + "type": "text", + "content": "3.3 Training: Hybrid Attention Masking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "text", + "content": "Given a discrete representation of the motion sequence " + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "inline_equation", + "content": "x_{1:t}" + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "text", + "content": ", our model is trained to reconstruct the motion sequence, conditioned on the text token " + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "text", + "content": " under both unidirectional and bidirectional causal masking strategies, i.e., " + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "inline_equation", + "content": "M^{uc}" + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "inline_equation", + "content": "M^{bc}" + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "text", + "content": ". The reconstruction probability of each motion token under each masking case is " + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_i \\mid M^{uc})" + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_i \\mid M^{bc})" + }, + { + "bbox": [ + 131, + 390, + 482, + 462 + ], + "type": "text", + "content": ", respectively. The training objective is to minimize the negative log-likelihood of the motion sequence prediction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 479, + 481, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 479, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 141, + 479, + 481, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {h y b r i d}} = - \\underset {\\mathbf {X} \\in p (\\mathbf {X})} {\\mathbb {E}} \\left[ \\lambda \\sum_ {\\forall i \\in [ 1, t ]} \\log p _ {\\theta} \\left(x _ {i} \\mid M ^ {u c}\\right) + (1 - \\lambda) \\sum_ {\\forall i \\in [ 1, t ]} \\log p _ {\\theta} \\left(x _ {i} \\mid M ^ {b c}\\right) \\right]. \\tag {4}", + "image_path": "4ff53c5f63b2a846d79d0a2f41d41d2add8bdee83a4c1dfdb6a7ecd02adbfd22.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 520, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 520, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 520, + 482, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 520, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 520, + 482, + 567 + ], + "type": "text", + "content": " is the probability of selecting unidirectional causal mask. Through experiments, we found " + }, + { + "bbox": [ + 131, + 520, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 131, + 520, + 482, + 567 + ], + "type": "text", + "content": " yields best performance. In addition, when bidirectional causal mask is selected, we randomly put " + }, + { + "bbox": [ + 131, + 520, + 482, + 567 + ], + "type": "inline_equation", + "content": "50\\% - 100\\%" + }, + { + "bbox": [ + 131, + 520, + 482, + 567 + ], + "type": "text", + "content": " motion tokens in the masked areas." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 586, + 353, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 353, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 353, + 597 + ], + "type": "text", + "content": "3.4 Inference: Cascaded Motion Decoding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "To generate motion sequence during inference phase, dual-iteration cascaded decoding is introduced. In the first iteration, autoregressive decoding is applied, where the motion tokens are sequentially and stochastically sampled according to unidirectional prediction distribution " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_i \\mid M^{uc})" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", concluding upon predicting the [END] token. Since autoregressive decoding could accumulate prediction" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 116, + 444, + 205 + ], + "blocks": [ + { + "bbox": [ + 166, + 116, + 444, + 205 + ], + "lines": [ + { + "bbox": [ + 166, + 116, + 444, + 205 + ], + "spans": [ + { + "bbox": [ + 166, + 116, + 444, + 205 + ], + "type": "image", + "image_path": "587bdad98e09ba99338a55c64f5f5647584f6ef15bde944017a33dcbc25559c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 208, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 264 + ], + "type": "text", + "content": "Fig. 3: Inference: Dual-iteration Cascaded Motion Decoding. In the first iteration, autoregressive decoding is applied by adopting unidirectional causal mask to generate coarse-grained motion and predict motion sequence length. In the second iteration, bidirectional autoregressive decoding is performed via bidirectional causal mask to removing and repredicting low-confidence motion tokens autoregressively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 372 + ], + "type": "text", + "content": "errors from the previously generated tokens, the generated motion sequence is refined in the second iteration by masking out a subset of motion tokens and then resampling these masked tokens according to the bidirectional prediction distribution " + }, + { + "bbox": [ + 130, + 277, + 482, + 372 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_i \\mid M^{bc})" + }, + { + "bbox": [ + 130, + 277, + 482, + 372 + ], + "type": "text", + "content": ". In this refinement iteration, [END] is positioned where it was predicted in the first iteration. The model can attend to unmasked tokens in all directions to re-predict low-confidence tokens based on rich surrounding context. The choice of masking strategies has an impact on the refinement gain, which is evaluated in Section 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 373, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 443 + ], + "type": "text", + "content": "Hybrid Classifier-free Guidance. At training time, we randomly drop textual tokens to teach the model to generate motion unconditionally. During inference, we apply classifier-free guidance (CFG) [20] for cascaded motion decoding. In particular, we generate the final motion sequence by a linear combination of the conditioned logits " + }, + { + "bbox": [ + 130, + 373, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\ell_c" + }, + { + "bbox": [ + 130, + 373, + 482, + 443 + ], + "type": "text", + "content": " and unconditional logits " + }, + { + "bbox": [ + 130, + 373, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\ell_u" + }, + { + "bbox": [ + 130, + 373, + 482, + 443 + ], + "type": "text", + "content": " with guidance scale " + }, + { + "bbox": [ + 130, + 373, + 482, + 443 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 373, + 482, + 443 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 252, + 445, + 481, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 445, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 252, + 445, + 481, + 458 + ], + "type": "interline_equation", + "content": "\\ell_ {g} = (1 + s) \\cdot \\ell_ {c} - s \\cdot \\ell_ {u}. \\tag {5}", + "image_path": "22d881f7c395e6a02bbaeb0e840387a11cb61ebdfaa29b1c8f15df8ff8ea2e95.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 485 + ], + "type": "text", + "content": "We apply different CFG scale " + }, + { + "bbox": [ + 130, + 462, + 481, + 485 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 462, + 481, + 485 + ], + "type": "text", + "content": " for each iteration during cascaded decoding. The effectiveness of CFG in each iteration is evaluated in Section 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": "Residual Motion Refinement. To further enhance motion generation quality, the motion sequence yielded from cascaded decoding can be refined by another refinement transformer based on residual vector quantization (RVQ) [44]. By utilizing RVQ, the raw motion sequence is encoded into multiple token sequences in the latent space. Each token sequence is generated by a separate quantizer and each quantizer encodes the quantization error left by the previous quantizer. As a result, the token sequence from the first quantizer encodes the most of information of the original motion sequence. The rest of the token sequences from other quantizers only encode quantization errors. Through RVQ, information loss can be minimized during the embedding quantization process. As a result, our masked self-attention transformer is used to generate the most informative token sequence from the first quantizer. Using this sequence as input, another refinement transformer is trained to predict the remaining token sequences, which are merged into a single token sequence for final motion decoding. This RVQ-based refinement has been adopted by audio generation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 274, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 274, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 274, + 102 + ], + "type": "text", + "content": "E. Pinyoanuntapong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 144, + 114, + 483, + 196 + ], + "blocks": [ + { + "bbox": [ + 144, + 114, + 483, + 196 + ], + "lines": [ + { + "bbox": [ + 144, + 114, + 483, + 196 + ], + "spans": [ + { + "bbox": [ + 144, + 114, + 483, + 196 + ], + "type": "image", + "image_path": "942fb72941539d12adda89d4bf0d03b8e140b7a9d554dfb51921daabe696239f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 204, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 293 + ], + "type": "text", + "content": "Fig. 4: Residual Motion Refinement. The residual vector quantization encodes the raw motion sequence into multiple token sequences in different colors (left). The base token sequence from the first vector quantizer is generated via cascaded decoding by masked self-attention transformer. The base token sequence is used as the input of the refinement transformer to predict the residual token sequences from other quantizers. The combined sequences are fed into tokenizer's decoder for motion generation. The refinement transformer shares the same architecture as the masked self-attention transformer with a full attention mask(right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 306, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 481, + 331 + ], + "type": "text", + "content": "models [4, 11, 40] and recently demonstrates its benefits in motion generation task [14]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 348, + 254, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 254, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 254, + 361 + ], + "type": "text", + "content": "3.5 Motion Editability" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 489 + ], + "type": "text", + "content": "The autoregressive approach naturally lacks the ability for temporal motion editing as it cannot leverage future motion tokens. In contrast, BAMM enables temporal motion editing through its bidirectional causal mask, which allows the model to access information from all directions of the conditioned tokens. Consequently, temporal motion editing becomes achievable by merely applying masked attention to the positions that require medications. We illustrate various editing tasks, i.e. motion inpainting (in-betweening), outpainting, prefix prediction, and suffix completion qualitatively and quantitively in Fig. 7 and Table 5. BAMM also allows us to generate the arbitrarily long motion sequence according to a sequence of text prompts, which is showcased in Supplementary Material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 508, + 230, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 230, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 230, + 521 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": "In this section, we provide a comprehensive evaluation along with empirical results for our proposed motion generation model, BAMM. In Section 4.1, we adhere to standard evaluation protocols on two datasets, demonstrating that our model outperforms current state-of-the-art methods both quantitatively and qualitatively. Furthermore, in Section 4.2, we showcase our predictive length and editing capabilities in comparison to state-of-the-art methods. The results reveal the robust performance of our model, displaying its effectiveness even in cases without prior information about motion length." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Datasets. We evaluated our model using the evaluation protocol proposed by [16] for text-driven motion generation with two datasets. KIT-ML dataset contains 3,911 motion sequences, each with one to four textual annotations," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "amounting to a total of 6,278 annotations. These motion sequences are derived from the KIT and CMU motion [1] databases and have been adjusted to 12.5 FPS. The dataset is divided into training, validation, and testing sets, with proportions of " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " respectively. HumanML3D dataset includes a wide variety of human activities, such as exercise and dancing, and consists of 14,616 motion sequences paired with 44,970 textual descriptions. These descriptions come from a vocabulary of 5,371 unique words. The motion sequences, sourced from AMASS [24] and HumanAct12 [17], have been standardized to 20 FPS and are limited to a maximum duration of 10 seconds, with the actual lengths varying between 2 to 10 seconds. Each sequence is accompanied by at least three descriptive annotations, averaging 12 words in length." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt the standard evaluation framework from T2M [16], employing pre-trained models that encode text and motion information to evaluate text and motion tokens in the embedding space. R-precision (Top-1, 2, 3 accuracy) measures how well the generated motions align with the text prompts, while Multimodal Distance (MM-Dist) quantifies the distance between generated and ground-truth motions in a shared feature space. Frechet Inception Distance (FID) assesses the statistical similarity between the feature distributions of generated and real motions. Additionally, we evaluate diversity (the average Euclidean distance between random motion pairs) and multimodality (the average variance across Euclidean distances between generated motion pairs for a single prompt) to capture the range of possible motion interpretations and their consistency with the text description." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 406, + 379, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 379, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 379, + 418 + ], + "type": "text", + "content": "4.1 Comparison to State-of-the-art Approaches" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 155, + 501, + 455, + 642 + ], + "blocks": [ + { + "bbox": [ + 130, + 442, + 482, + 498 + ], + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 498 + ], + "type": "text", + "content": "Table 2: Comparison of text-conditional motion synthesis on HumanML3D [16] test set. We repeat the evaluation 20 times for each metric and report the average with " + }, + { + "bbox": [ + 130, + 442, + 482, + 498 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 442, + 482, + 498 + ], + "type": "text", + "content": " confidence interval. Red and Blue indicate the best and the second best result. Methods with gray highlight " + }, + { + "bbox": [ + 130, + 442, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 130, + 442, + 482, + 498 + ], + "type": "text", + "content": " report motion generation results using the ground-truth motion length." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 155, + 501, + 455, + 642 + ], + "lines": [ + { + "bbox": [ + 155, + 501, + 455, + 642 + ], + "spans": [ + { + "bbox": [ + 155, + 501, + 455, + 642 + ], + "type": "table", + "html": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Hier [13]0.301±.0020.425±.0020.552±.0046.523±.0245.012±.0188.332±.042-
TEMOS§ [27]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
TM2T [15]0.424±.0030.618±.0030.729±.0021.501±.0173.467±.0118.589±.0762.424±.093
T2M [16]0.455±.0030.636±.0030.736±.0021.087±.0213.347±.0089.175±.0832.219±.074
MDM§ [37]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse§ [46]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD§ [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
Fg-T2M§ [41]0.492±.0020.683±.0030.783±.0020.243±.0193.109±.0079.278±.0721.614±.049
M2DM§ [23]0.497±.0030.682±.0020.763±.0030.352±.0053.134±.0109.926±.0733.587±.072
T2M-GPT [45]0.491±.0030.680±.0030.775±.0020.116±.0043.118±.0119.761±.0811.856±.011
AttT2M [49]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
MMM§ [29]0.515±.0020.708±.0020.804±.0020.089±.0052.926±.0079.577±.0501.226±.035
MoMask§ [14]0.521±.0020.713±.0030.807±.0020.045±.0022.958±.008-1.241±.040
BAMM (ours)0.525±.0020.720±.0030.814±.0030.055±.0022.919±.0089.717±.0891.687±.051
BAMM§ (ours)0.522±.0030.715±.0030.808±.0030.055±.0022.936±.0779.636±.0091.732±.055
", + "image_path": "51e3fd425f56f2d73a7412299cc7e0febd0f7cbff430f3eabb270ce84eaf68b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "text", + "content": "E. Pinyoanuntapong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 162, + 454, + 304 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Comparison of text-conditional motion synthesis on KIT-ML [30] test set. We repeat the evaluation 20 times for each metric and report the average with " + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": " confidence interval. Red and Blue indicate the best and the second best result. Methods with gray highlight " + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": " report ground-truth motion length for generation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 162, + 454, + 304 + ], + "lines": [ + { + "bbox": [ + 154, + 162, + 454, + 304 + ], + "spans": [ + { + "bbox": [ + 154, + 162, + 454, + 304 + ], + "type": "table", + "html": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Hier [13]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.072-
TEMOS$ [27]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
TM2T [15]0.280±.0050.463±.0060.587±.0053.599±.1534.591±.0269.473±.1173.292±.081
T2M [16]0.361±.0060.559±.0070.681±.0073.022±.1073.488±.02810.72±.1452.052±.107
MDM$ [37]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MotionDiffuse$ [46]0.417±.0040.621±.0040.739±.0041.954±.0642.958±.00511.10±.1430.730±.013
MLD$ [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
Fg-T2M$ [41]0.418±.0050.626±.0040.745±.0040.571±.0473.114±.01510.93±.0831.019±.029
M2DM$ [23]0.416±.0040.628±.0040.743±.0040.515±.0293.015±.01711.417±.973.325±.37
T2M-GPT [45]0.402±.0060.619±.0050.737±.0060.717±.0413.053±.02610.86±.0941.912±.036
AttT2M [49]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
MMM$ [29]0.404±.0050.621±.0050.744±.0040.316±.0282.977±.01910.910±.1011.232±.039
MoMask$ [14]0.433±.0070.656±.0050.781±.0050.204±.0112.779±.022-1.131±.043
BAMM (ours)0.438±.0090.661±.0090.788±.0050.183±.0132.723±.02611.008±.0941.609±.065
BAMM (ours)$0.436±.0070.660±.0060.791±.0050.200±.0112.714±.01610.914±.0971.517±.058
", + "image_path": "8de3d4f769cec2dda829f6583657935d4e440860b67637f0a3116dcf068b1e28.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 474 + ], + "type": "text", + "content": "Quantitative Results. We evaluate our model on the HumanML3D [16] and KIT-ML [30] datasets, reporting the results in Table 2 and 3, respectively, in comparison to state-of-the-art methods. Following the standard evaluation protocol from [16], we report the average of 20 generations with a " + }, + { + "bbox": [ + 130, + 319, + 481, + 474 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 319, + 481, + 474 + ], + "type": "text", + "content": " confidence interval. Our model consistently outperforms other methods in terms of R-precision, FID, and MM-Distance, while maintaining comparability in terms of Diversity and Multimodal Distance, which typically represent the trade-off between high quality and diversity. This suggests that our model generates very high-quality outputs while retaining a good degree of diversity. Moreover, whereas most models utilize the ground truth length for evaluation, our model supports both predicted length and takes length as an input, outperforming other methods in both scenarios. We further investigate the effect of a separate length estimator which can significantly impact performance in Section 4.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 475, + 481, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 606 + ], + "type": "text", + "content": "Qualitative Results. Fig. 5 presents a qualitative comparison with T2M-GPT [45], MoMask [14], and MDM [37]. BAMM and T2M-GPT generate motion without requiring input length. Despite this, BAMM accurately generates motion from textual descriptions even without input length. We utilize a pre-trained length estimator from [16] for MoMask [14] and MDM [37]. Notably, BAMM generates motion accurately aligned with the provided text, whereas T2M-GPT and MoMask produce erroneous motion, and MDM generates entirely inaccurate motion. Additional visualizations for BAMM are shown in Fig. 6, indicating its ability to generate high detail such as complex trajectories and interactions with invisible objects. Moreover, Fig. 7 demonstrates BAMM's capability in various temporal editing tasks, including inpainting, outpainting, prefix, and suffix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 622, + 328, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 328, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 328, + 635 + ], + "type": "text", + "content": "4.2 Length Prediction and Editability" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 487, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 487, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 487, + 665 + ], + "type": "text", + "content": "Predict Length vs ground truth length. In real-world scenarios, the ground truth length is often unknown. Many methods cannot predict length autonomously." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 105, + 443, + 252 + ], + "blocks": [ + { + "bbox": [ + 173, + 105, + 443, + 252 + ], + "lines": [ + { + "bbox": [ + 173, + 105, + 443, + 252 + ], + "spans": [ + { + "bbox": [ + 173, + 105, + 443, + 252 + ], + "type": "image", + "image_path": "fb9d9b4266dfb039b411d8c9130d9bea94029c973f00f07f6134f9aaa5e28663.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 258, + 480, + 301 + ], + "lines": [ + { + "bbox": [ + 131, + 258, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 480, + 301 + ], + "type": "text", + "content": "Fig. 5: Visualization comparison of textual to motion to state-of-the-art methods. BAMM and T2M-GPT do not require motion length as an input. We use a pre-trained length estimator from [16] for MoMask and MDM. BAMM generates higher quality and is more correlated with textual descriptions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 315, + 446, + 392 + ], + "blocks": [ + { + "bbox": [ + 170, + 315, + 446, + 392 + ], + "lines": [ + { + "bbox": [ + 170, + 315, + 446, + 392 + ], + "spans": [ + { + "bbox": [ + 170, + 315, + 446, + 392 + ], + "type": "image", + "image_path": "f9a8183e46fdfe1d46c91fee66ac7289637c9c226c43fe5717c09e1ebdec8a1b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 401, + 480, + 434 + ], + "lines": [ + { + "bbox": [ + 131, + 401, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 480, + 434 + ], + "type": "text", + "content": "Fig. 6: Visualization of text-to-motion generation by BAMM. BAMM can generate high-quality motion with complex descriptions, such as intricate trajectories and interactions with invisible objects." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 444, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 482, + 562 + ], + "type": "text", + "content": "Consequently, they evaluate using ground truth length, indicated by gray highlight " + }, + { + "bbox": [ + 131, + 444, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 131, + 444, + 482, + 562 + ], + "type": "text", + "content": " in Table 2 and Table 3. We illustrate the impact of inaccurate length prediction by comparing our model to state-of-the-art denoising models that require motion length as an input, such as MoMask and MMM. In particular, both models take the estimated length obtained from a standalone length predictor pretrained on the HumanML3D dataset [16]. Table 4 reveals that the predicted length significantly worsens MoMask's FID score, from 0.045 to 0.090, while MMM's R-precision Top1 drops from 0.515 to 0.504. In contrast, our BAMM not only maintains the best performance but also outperforms MMM and MoMask in both scenarios." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 187, + 609, + 418, + 669 + ], + "blocks": [ + { + "bbox": [ + 131, + 582, + 480, + 605 + ], + "lines": [ + { + "bbox": [ + 131, + 582, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 480, + 605 + ], + "type": "text", + "content": "Table 4: Comparison of text-conditional motion synthesis using predicted and ground truth length on HumanML3D [16] dataset" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 609, + 418, + 669 + ], + "lines": [ + { + "bbox": [ + 187, + 609, + 418, + 669 + ], + "spans": [ + { + "bbox": [ + 187, + 609, + 418, + 669 + ], + "type": "table", + "html": "
LengthMethodsR-Precision ↑FID ↓MM-Dist ↓Diversity ↑MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
GroundMMM0.5150.7080.8040.0892.9269.5771.226
TruthMoMask0.5210.7130.8070.0452.958-1.241
BAMM (Ours)0.5220.7150.8080.0552.9369.6361.732
PredictedMMM0.5040.6960.7940.0802.9989.4111.164
MoMask0.5220.7150.8110.0902.9459.6471.239
BAMM (Ours)0.5250.7200.8140.0552.9199.7171.687
", + "image_path": "868598c7e15351d35a88dcfbe03a8144e8224e5ca365941b89ed96528f261ed2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "text", + "content": "E. Pinyoanuntapong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 478, + 195 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 478, + 195 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 478, + 195 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 478, + 195 + ], + "type": "image", + "image_path": "9294446434a39066cbf67430dd738105598b9d0308a9f9dde20d04c5e6ab1cde.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "Fig. 7: Visualization of temporal editing tasks, inpainting (in-betweening), outpainting, prefix, and suffix where blue indicates conditioned motion and red refers to generated parts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 378 + ], + "type": "text", + "content": "Editability. We conduct experiments in four temporal editing tasks, namely inpainting (in-betweening), outpainting, prefix, and suffix, comparing them with MDM [37] and MoMask [14] on the HumanML3D dataset. Inpainting is evaluated by generating " + }, + { + "bbox": [ + 130, + 247, + 482, + 378 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 247, + 482, + 378 + ], + "type": "text", + "content": " of the motion sequence given the first and last " + }, + { + "bbox": [ + 130, + 247, + 482, + 378 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 247, + 482, + 378 + ], + "type": "text", + "content": ". Outpainting is its opposite counterpart. Prefix is conditioned by the first " + }, + { + "bbox": [ + 130, + 247, + 482, + 378 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 247, + 482, + 378 + ], + "type": "text", + "content": " of the ground truth motion and generates the remaining portion. Similarly, suffix operates in the reverse manner. Note that editing tasks require prior knowledge of conditioned motion and motion length, for which MDM and MoMask are specifically designed. Despite this, Table 5 shows that our model significantly outperforms MDM in all editing tasks and surpasses MoMask in outpainting, while performing equally well in the other tasks." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 196, + 406, + 411, + 511 + ], + "blocks": [ + { + "bbox": [ + 132, + 389, + 481, + 401 + ], + "lines": [ + { + "bbox": [ + 132, + 389, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 481, + 401 + ], + "type": "text", + "content": "Table 5: Evaluation on temporal editing tasks on HumanML3D [16] dataset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 196, + 406, + 411, + 511 + ], + "lines": [ + { + "bbox": [ + 196, + 406, + 411, + 511 + ], + "spans": [ + { + "bbox": [ + 196, + 406, + 411, + 511 + ], + "type": "table", + "html": "
TasksMethodsR-Precision ↑FID ↓ MM-Dist ↓ Diversity ↑
Top-1 ↑Top-2 ↑Top-3 ↑
Temporal Inpainting (In-betweenging)MDM0.3910.5780.6922.3623.8598.014
MoMask0.5340.7270.820.042.8789.64
BAMM0.5350.7290.8210.0562.8639.629
Temporal OutpaintingMDM0.4150.6130.7272.0573.6198.199
MoMask0.5310.7260.8180.0572.8899.619
BAMM0.5350.730.8220.0562.8569.659
PrefixMDM0.420.6130.7251.463.5638.312
MoMask0.5360.730.8220.062.8759.607
BAMM0.5320.7270.8210.0582.8689.612
SuffixMDM0.4030.5970.7112.5623.7318.088
MoMask0.5320.7260.8190.0522.8819.659
BAMM0.5270.720.8140.052.8919.721
", + "image_path": "7ac77609d2847caaaa2afabb12182bde39cd2b8e4482a1e29c996a84a4cab61c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 532, + 246, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 246, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 246, + 547 + ], + "type": "text", + "content": "5 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 593 + ], + "type": "text", + "content": "In the ablation study, we study the impacts of the adaptive classifier-free guidance scales in each iteration of the Masked Self-attention Transformer, the masking strategy, and the number of iterations, as shown in Table 6." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Adaptive Classifier-Free Guidance Scales (CFG): We conducted experiments with different CFG values for both the first and second iterations and found that CFG=4 in the first iteration and CFG=3 in the second iteration works best. Although CFG=5 in the first iteration yields a better FID score, the other scores are worse. Mask Strategy: The first strategy, \"50% of low confidence\", applies mask attention on 50% of the lowest confident positions from the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 126, + 444, + 274 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 363, + 125 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 363, + 125 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 363, + 125 + ], + "type": "text", + "content": "Table 6: Ablation Study" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 126, + 444, + 274 + ], + "lines": [ + { + "bbox": [ + 164, + 126, + 444, + 274 + ], + "spans": [ + { + "bbox": [ + 164, + 126, + 444, + 274 + ], + "type": "table", + "html": "
AblationsTypesR-Precision ↑FID ↓ MM-Dist ↓ Diversity ↑ MModality ↑
Top-1 ↑Top-2 ↑Top-3 ↑
CFG of 1st iteration1st iter2nd iter
CFG=2CFG=30.5170.7110.8050.1052.9569.841.907
CFG=3CFG=30.5210.7140.8080.072.9449.7771.766
CFG=4CFG=30.5250.720.8140.0552.9199.7171.687
CFG=5CFG=30.5220.7160.810.0522.9279.6471.691
CFG=6CFG=30.520.7130.810.062.949.6211.697
CFG of 2nd iterationCFG=4CFG=10.5210.7160.810.0582.9439.7431.744
CFG=4CFG=20.5240.7190.8140.0572.9249.7521.698
CFG=4CFG=30.5250.720.8140.0552.9199.7171.687
CFG=4CFG=40.5220.7190.8120.0562.9249.6911.698
CFG=4CFG=50.5210.7170.8120.0652.9319.6381.735
Mask Strategy50% of low confidence confidence < .50.5250.720.8130.0652.9219.7321.67
suffix0.5250.7180.810.0642.9239.7651.656
%2=00.5190.7150.810.0522.9439.6831.841
0.5250.720.8140.0552.9199.7171.687
# of iterations1 iteration0.5240.7180.8120.0642.9269.7201.644
2 iterations0.5250.7200.8140.0552.9199.7171.687
3 iterations0.5250.7190.8140.0552.9179.7271.69
", + "image_path": "18d0fe4153160804ea94ed98cbf326533e2d26b26a5598c70b3e54322bb3aca4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "text", + "content": "first iteration. Similarly, \"confidence " + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "inline_equation", + "content": "< .5" + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "text", + "content": " \" uses low confidence but sets a threshold, masking out positions where confidence is below 0.5. Specifically, the former strictly masks out " + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "text", + "content": " of the total token sequence while the latter masks the token with a confidence lower than .5. \"Suffix\" strategy masks out the first " + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "text", + "content": " of the token sequence, utilizing the remaining " + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "text", + "content": " as condition tokens. Lastly, " + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\% 2 = 0" + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "text", + "content": " \"strategy masks every other token. BAMM produces decent performances across the tested masking strategies while the simple masking of every other token works the best. Number of Iterations: We conducted experiments with one to three iterations. \"1 iteration\" refers to the first iteration, as described in 3.4, which utilizes a unidirectional causal mask. Similarly, \"2 iterations\" involves applying bidirectional causal mask to the Masked Self-attention Transformer to re-predict the tokens from the first iteration, as illustrated in 3. In \"3 iterations\", we repeat the second iteration (bidirectional causal mask) but apply masks to " + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 130, + 291, + 482, + 495 + ], + "type": "text", + "content": " of the sequence to re-predict the motion tokens instead. The experiments indicate that \"2 iterations\" clearly demonstrate improvement over one iteration, while \"3 iterations\" do not significantly improve upon \"2 iterations\". This suggests that \"2 iterations\" are sufficient." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We introduce the Bidirectional Autoregressive Motion Model (BAMM), a novel framework for text-to-motion generation. BAMM combines a motion tokenizer, which encodes 3D human motion into discrete latent tokens, with a masked self-attention transformer that autoregressively predicts the masked tokens through a masked casual attention approach. BAMM integrates of generative masked and autoregressive modeling into an unified framework, This features allows it to understand complex motion relationships and precisely map text inputs to high-quality motion outputs with adaptively adjusted sequence lengths. Our extensive testing on HumanML3D and KIT-ML datasets confirms BAMM's superiority in both qualitative and quantitative evaluations over existing methods." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 273, + 102 + ], + "type": "text", + "content": "E. Pinyoanuntapong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 156 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 156 + ], + "type": "text", + "content": "1. Cmu graphics lab motion capture database, http://mocap.cs.cmu.edu/, accessed: 2022-11-11 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 158, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 158, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 158, + 480, + 190 + ], + "type": "text", + "content": "2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728 (2019). https://doi.org/10.1109/3DV.2019.000844" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 190, + 480, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 480, + 222 + ], + "type": "text", + "content": "3. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems 34, 17981-17993 (2021) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "type": "text", + "content": "4. Borsos, Z., Sharifi, M., Vincent, D., Kharitonov, E., Zeghidour, N., Tagliasacchi, M.: Soundstorm: Efficient parallel audio generation. ArXiv abs/2305.09636 (2023), https://api-semanticscholar.org/CorpusID:258715176 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 255, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 330 + ], + "type": "text", + "content": "5. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T.J., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. ArXiv abs/2005.14165 (2020), https://api_semanticscholar.org/CorpusID:218971783 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 331, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 373 + ], + "type": "text", + "content": "6. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M., Murphy, K.P., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D.: Muse: Text-to-image generation via masked generative transformers. ArXiv abs/2301.00704 (2023), https://api(semanticscholar.org/CorpusID:255372955 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 374, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 417 + ], + "type": "text", + "content": "7. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11305-11315 (2022), https://apisemantic scholar.org/CorpusID:246680316 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 418, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 460 + ], + "type": "text", + "content": "8. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, J., Yu, G.: Executing your commands via motion diffusion in latent space. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 18000-18010 (2022), https://api-semanticscholar.org/CorpusID:254408910 4, 10, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "type": "text", + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019), https://apisemantic scholar.org/CorpusID:52967399 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 503, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 535 + ], + "type": "text", + "content": "0. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. ArXiv abs/2204.14217 (2022), https://api(semanticscholar.org/CorpusID:248476190 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 536, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 578 + ], + "type": "text", + "content": "1. Du, Z., Qian, Y., Liu, X., Ding, M., Qiu, J., Yang, Z., Tang, J.: Glm: General language model pretraining with autoregressive blank infilling. In: Annual Meeting of the Association for Computational Linguistics (2021), https://apisemantic scholar.org/CorpusID:247519241 5, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 578, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 621 + ], + "type": "text", + "content": "2. Ghazvininejad, M., Levy, O., Liu, Y., Zettlemoyer, L.: Mask-predict: Parallel decoding of conditional masked language models. In: Conference on Empirical Methods in Natural Language Processing (2019), https://api_semanticscholar.org/ CorpusID:202538740 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "3. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 1376-1386 (2021), https://apisemantic scholar.org/CorpusID:232404671 10, 11" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "14. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions (2023) 2, 4, 6, 9, 10, 11, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "text", + "content": "15. Guo, C., Xuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. ArXiv abs/2207.01696 (2022), https://api_semanticscholar.org/CorpusID: 250280248 10, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 228 + ], + "type": "text", + "content": "16. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5142-5151 (2022). https://doi.org/10.1109/CVPR52688.2022.005093, 4, 5, 9, 10, 11, 12, 13, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 272 + ], + "type": "text", + "content": "17. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. Proceedings of the 28th ACM International Conference on Multimedia (2020), https://apisemantic scholar.org/CorpusID:220870974 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 317 + ], + "type": "text", + "content": "18. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A.A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models. ArXiv abs/2210.02303 (2022), https://api_semanticscholar.org/CorpusID:252715883 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 318, + 481, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 349 + ], + "type": "text", + "content": "19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. ArXiv abs/2006.11239 (2020), https://apisemantic scholar.org/CorpusID: 219955663 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 351, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 481, + 373 + ], + "type": "text", + "content": "20. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "21. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. ArXiv abs/2306.14795 (2023), https://api.sementicscholar.org/CorpusID:259262201 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "22. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI Conference on Artificial Intelligence (2022), https://api-semanticscholar.org/CorpusID:251979380 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "text", + "content": "23. Kong, H., Gong, K., Lian, D., Mi, M.B., Wang, X.: Priority-centric human motion generation in discrete latent space. ArXiv abs/2308.14480 (2023), https://api.sementicscholar.org/CorpusID:261245369 10, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "text", + "content": "24. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 5441-5450 (2019), https://apisemanticscholar.org/CorpusID:102351100 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 521, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 564 + ], + "type": "text", + "content": "25. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning (2021), https://api-semanticscholar.org/CorpusID:245335086 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "26. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning. ArXiv abs/1711.00937 (2017), https://api-semanticscholar.org/CorpusID:20282961 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "text", + "content": "27. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. ArXiv abs/2204.14109 (2022), https://api.sementicscholar.org/CorpusID:248476220 4, 10, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "28. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. ArXiv abs/2305.00976 (2023), https://api_semanticscholar.org/CorpusID:258436810 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "text", + "content": "E. Pinyoanuntapong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Pinyoanuntapong, E., Wang, P., Lee, M., Chen, C.: Mmm: Generative masked motion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 2, 4, 6, 10, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "text", + "content": "30. Plappert, M., Mandery, C., Asfour, T.: The KIT motion-language dataset. Big Data 4(4), 236-252 (dec 2016). https://doi.org/10.1089/big.2016.0028, http://dx.doi.org/10.1089/big.2016.0028 4, 11, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "type": "text", + "content": "31. Qian, L., Zhou, H., Bao, Y., Wang, M., Qiu, L., Zhang, W., Yu, Y., Li, L.: Glancing transformer for non-autoregressive neural machine translation. In: Annual Meeting of the Association for Computational Linguistics (2020), https://api_semanticscholar.org/CorpusID:221150562 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 278 + ], + "type": "text", + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021), https://api_semanticscholar.org/CorpusID: 231591445 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 278, + 482, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 332 + ], + "type": "text", + "content": "33. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. ArXiv abs/2205.11487 (2022), https://api_semanticscholar.org/CorpusID:248986576 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 332, + 482, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 376 + ], + "type": "text", + "content": "34. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data. ArXiv abs/2209.14792 (2022), https://api_semanticscholar.org/CorpusID:252595919 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 376, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 407 + ], + "type": "text", + "content": "35. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ArXiv abs/2010.02502 (2020), https://api.sementicscholar.org/CorpusID:222140788 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 407, + 482, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 440 + ], + "type": "text", + "content": "36. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision (2022), https://api-semanticscholar.org/CorpusID:247450907 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 440, + 482, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 472 + ], + "type": "text", + "content": "37. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. ArXiv abs/2209.14916 (2022), https://api.sementicscholar.org/CorpusID:252595883 2, 4, 10, 11, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 472, + 482, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 514 + ], + "type": "text", + "content": "38. Vaswani, A., Shazeer, N.M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:137564895" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "type": "text", + "content": "39. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. ArXiv abs/2210.02399 (2022), https://apisemantic scholar.org/CorpusID:252715594 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "text", + "content": "40. Wang, C., Chen, S., Wu, Y., Zhang, Z.H., Zhou, L., Liu, S., Chen, Z., Liu, Y., Wang, H., Li, J., He, L., Zhao, S., Wei, F.: Neural codec language models are zero-shot text to speech synthesizers. ArXiv abs/2301.02111 (2023), https://api.sementicscholar.org/CorpusID:255440307 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "41. Wang, Y., Leng, Z., Li, F.W.B., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. ArXiv abs/2309.06284 (2023), https://api.sementicscholar.org/CorpusID:261697123 2, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "text", + "content": "42. Yan, S., Liu, Y., Wang, H., Du, X., Liu, M., Liu, H.: Cross-modal retrieval for motion and text via doptriple loss (2023), https://api-semanticscholar.org/CorpusID:263610212 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "BAMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 436 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "type": "text", + "content": "43. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., Hutchinson, B.C., Han, W., Parekh, Z., Li, X., Zhang, H., Baldridge, J., Wu, Y.: Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res. 2022 (2022), https://api_semanticscholar.org/CorpusID:249926846 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "44. Zeghidour, N., Luebs, A., Omran, A., Skoglund, J., Tagliasacchi, M.: Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30, 495-507 (2021), https://api-semanticscholar.org/CorpusID:236149944 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 270 + ], + "type": "text", + "content": "45. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: Generating human motion from textual descriptions with discrete representations. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14730-14740 (2023), https://api_semanticscholar.org/CorpusID:255942203 2, 5, 6, 10, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 271, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 313 + ], + "type": "text", + "content": "46. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. ArXiv abs/2208.15001 (2022), https://api_semanticscholar.org/CorpusID: 251953565 2, 4, 10, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "text", + "content": "47. Zhang, Z., Ma, J., Zhou, C., Men, R., Li, Z., Ding, M., Tang, J., Zhou, J., Yang, H.: M6-ufc: Unifying multi-modal controls for conditional image synthesis via non-autoregressive generative transformers (2021), https://api_semanticscholar.org/CorpusID:237204528 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 358, + 481, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 401 + ], + "type": "text", + "content": "48. Zhang, Z., Ma, J., Zhou, C., Men, R., Li, Z., Ding, M., Tang, J., Zhou, J., Yang, H.: Ufc-bert: Unifying multi-modal controls for conditional image synthesis. In: Neural Information Processing Systems (2021), https://apisemantic scholar.org/CorpusID:235253928 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 402, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 436 + ], + "type": "text", + "content": "49. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. ArXiv abs/2309.00796 (2023), https://api.sementicscholar.org/CorpusID:261530775 2, 5, 10, 11" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "text", + "content": "E. Pinyoanuntapong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_content_list.json b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04bbbc9ccaec0c32d4eaebf2fbb9a843017d6193 --- /dev/null +++ b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_content_list.json @@ -0,0 +1,1608 @@ +[ + { + "type": "text", + "text": "BEAF: Observing BEFORE-AFTER Changes to Evaluate Hallucination in Vision-language Models", + "text_level": 1, + "bbox": [ + 217, + 141, + 787, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moon Ye-Bin $^{1*}$ Nam Hyeon-Woo $^{1*}$ Wonseok Choi $^{2}$ Tae-Hyun Oh $^{1,2,3}$", + "bbox": [ + 218, + 210, + 782, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. of EE, POSTECH, Korea $^{2}$ Grad. School of AI, POSTECH, Korea $^{3}$ Institute for Convergence Research and Education in Advanced Technology, Yonsei University, Korea.", + "bbox": [ + 220, + 238, + 781, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ybmoon, hyeonw.nam, wonseok.c, taehyun}@postech.ac.kr", + "bbox": [ + 290, + 282, + 709, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Vision language models (VLMs) perceive the world through a combination of a visual encoder and a large language model (LLM). The visual encoder, pre-trained on large-scale vision-text datasets, provides zero-shot generalization to visual data, and the LLM endows its high reasoning ability to VLMs. It leads VLMs to achieve high performance on wide benchmarks without fine-tuning, exhibiting zero or few-shot capability. However, recent studies show that VLMs are vulnerable to hallucination. This undesirable behavior degrades reliability and credibility, thereby making users unable to fully trust the output from VLMs. To enhance trustworthiness and better tackle the hallucination of VLMs, we curate a new evaluation dataset, called the BEfore-AFTER hallucination dataset (BEAF), and introduce new metrics: True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID). Unlike prior works that focus only on constructing questions and answers, the key idea of our benchmark is to manipulate visual scene information by image editing models and to design the metrics based on scene changes. This allows us to clearly assess whether VLMs correctly understand a given scene by observing the ability to perceive changes. We also visualize image-wise object relationship by virtue of our two-axis view: vision and text. Upon evaluating VLMs with our dataset, we observed that our metrics reveal different aspects of VLM hallucination that have not been reported before. Project page: https://beafbench.github.io/", + "bbox": [ + 261, + 333, + 743, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 664, + 375, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision language models (VLMs) [1, 5, 8, 23, 24] have recently emerged by its multi-modal reasoning capability. It is indeed promising research toward making intelligent agents perceive the world. VLMs comprise a vision encoder as a visual perception module and a large language model (LLM) as a reasoning module. Recent LLMs [6, 32, 33] have exhibited powerful reasoning capabilities across various tasks, including commonsense reasoning [3, 16] and math [7], in zero-/few-shot ways. This success motivates researchers to integrate LLMs with vision encoders to open LLMs' eyes to see the world. Leveraging the reasoning", + "bbox": [ + 212, + 695, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Authors contributed equally to this work.", + "bbox": [ + 217, + 824, + 511, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "capabilities of LLMs alongside visual perception enables VLMs to excel in various visual tasks, such as image captioning [2,38] and visual question answering [11, 15] without task-specific fine-tuning.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite this encouraging demonstration, VLMs are susceptible to the hallucination [12,14,19,34,39,42] wherein their outputs do not reflect facts presented in input images. For example, although a given input image does not present an object, VLMs often incorrectly answer for the presence of the object. This tendency is problematic, akin to spreading falsehoods in human communication. Hallucination prevents genuine communication, consequently diminishing the reliability of interactions between agents or humans. No matter how good VLM's answer is, if hallucination is included in the answer, it can undermine trust in the system's reliability and credibility. Therefore, it is crucial to address hallucinations to improve the system's responsibility and trustworthiness.", + "bbox": [ + 212, + 195, + 784, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Understanding the impact of hallucinations on model performance requires the development of an assessment framework, which requires a deep exploration of the underlying causes. The prior works [14, 41, 42] have been proposed to unveil the causation of hallucinations in terms of data and models by reporting their performance on existing hallucination benchmarks. As a widely used hallucination benchmark, POPE [19] follows the question-and-answer (QnA) format. For the convenience of measuring, POPE uses a discriminative question, e.g., \"Is there {object} in the image?\", and gets a YES or NO answer. The following hallucination benchmarks [14, 19, 34] commonly leverage LLMs or human annotators to generate question prompts with available scene information such as bounding boxes and captions. This text-axis evaluation may assess certain aspects of the hallucination performance well, but it is hard to disentangle sources of hallucination, particularly in VLMs that handle both visual and text modalities. Thus, to thoroughly analyze and understand the behavior of VLMs, it is both natural and necessary to consider both modalities.", + "bbox": [ + 212, + 349, + 785, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a BEFORE-AFTER (BEAF) hallucination evaluation benchmark. This benchmark comprises a dataset enhanced along the vision-axis and metrics that are aware of changes. We manipulate scenes by selectively removing objects on the vision-axis while augmenting QnA pairs along the text-axis. We can observe how the models' answers change as the image is manipulated for the same question. For example, as illustrated in Fig. 1, if we remove an apple from the given image and then ask \"Is there an apple?\", the answer from the model should change from \"Yes\" to \"No\", between the original and modified images, if it truly relies on what it sees. Therefore, for a more thorough evaluation and analysis, we need change-aware metrics beyond just using standard accuracy [14, 19, 34].", + "bbox": [ + 212, + 579, + 785, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose four new metrics: True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID), which are enabled by our vision and text augmentation. TU measures whether the model answers correctly before and after the manipulation. IG evaluates whether the models perceive the images consistently both before and after manipulation. SB is related to the phenomenon of giving the same answer instead of correctly recognizing the image. ID measures", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Ye-Bin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec68970f63d0e326873895ee4d8be40f9eed1d21bad8bd22363188c352a6c607.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 143, + 493, + 299 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/1dff7727fd68446bb59c22ae30d3e783e5fbe4d9a5c3deba59ef8c167d18814d.jpg", + "table_caption": [], + "table_footnote": [ + "MT: Manipulation Target" + ], + "table_body": "
MTQuestionMetricsCorrectnessAnswer
Text-Acc○-
Img. & TextRemoved ObjectTU○ → ○Y → N
IG× → ×N → Y
SBp○ → ×Y → Y
SBn× → ○N → N
Unremoved ObjectID○ → ×\n× → ○Y → N\nN → Y
", + "bbox": [ + 504, + 146, + 759, + 290 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/44cd7026e1a3dc75993230db3fa4f4fb633fa90398a910778da9f45b8e53c474.jpg", + "image_caption": [ + "BEAF Evaluation (Ours)", + "Fig. 1: Illustration of BEbefore-AFTER (BEAF) benchmark. We present a comparison between traditional evaluation benchmarks and our BEAF benchmark for assessing the hallucination behavior in VLMs. Traditional evaluation methods solely manipulate questions based on the existence of an object and measure accuracy or F1 score. In contrast, our BEAF benchmark not only constructs questions but also manipulates images and tracks changes in answers as the images undergo manipulation. The BEAF benchmark introduces novel metrics including True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID), which consider the changes in answers." + ], + "image_footnote": [], + "bbox": [ + 230, + 318, + 767, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cases where the answer changes even though the target object does not change or does not exist in the image. The first three metrics address questions regarding target objects that have been removed, while the last one belongs to questions about other objects not related to the targets.", + "bbox": [ + 212, + 657, + 787, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on our BEAF dataset and change-aware metrics, we can evaluate and understand the models, even the parts we did not know through the previous methods. The evaluation results imply that answers assumed as non-hallucination from the existing hallucination benchmarks could be a hallucination. We also analyze the influence of each object in the scene based on the variation along the vision and text axes. In addition, we include a discussion on open-generation answers and CLIPScore results. Our key contributions are as follows:", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BEAF Benchmark", + "bbox": [ + 607, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a benchmark, called BEFORE-AFTER (BEAF), to evaluate the hallucination of large vision-language models. Our benchmark comprises a dataset that considers both text- and vision-axes, and change-aware metrics, which allow us to perform granular evaluation.", + "- We demonstrate that previously reported outcomes deemed non-hallucinatory in the prior text-axis-only evaluation benchmarks may be hallucinations.", + "- We also visualize the impact of relationships between objects within hallucinatory images based on the BEAF results." + ], + "bbox": [ + 214, + 146, + 782, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 292, + 387, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision Language Models (VLMs). VLMs [1, 4, 5, 8, 10, 23-25, 40, 43] have demonstrated their impressive image describing and reasoning ability based on powerful large language models (LLMs) [6, 31-33]. Generally, VLMs consist of three parts: vision encoder, LLM, and vision-language alignment module. The vision information extracted from the vision encoder is translated to the LLM-understandable token by passing through the alignment module. LLM takes a prompt text and visual tokens as input and outputs response text tokens according to a given image and prompt. Their training process is divided into multiple steps depending on which parts of VLMs are trained.", + "bbox": [ + 212, + 325, + 784, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hallucination in VLMs. Hallucination [12, 19, 22, 34, 35, 39] in VLMs refers to the case where the generated text answer does not reflect the true contents of the provided images but rather relies on the internal knowledge of the models. This behavior might cause harmful consequences in high-risk applications such as the medical domain or auto-driving scenarios. This undesirable behavior could be caused by various factors, e.g., data imbalance, image resolution, and model capacity. The assessment of hallucination is also related to the true understanding of VLMs. If VLMs respond to the given input correctly rather than relying on innate knowledge, VLMs are considered to have a perfect understanding of the perceived world. Numerous methodologies [12, 21, 42] have been developed to tackle the issue of hallucination in VLMs by addressing data biases [14, 21], optimization techniques [12, 18, 29], and post-processing methodologies [9, 37, 42].", + "bbox": [ + 212, + 462, + 785, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluating Hallucination in VLMs. Evaluating hallucination performance is essential to guide the research of VLM. Previous studies have introduced hallucination benchmarks to understand and assess these undesirable phenomena of VLMs. Motivated by traditional visual and question benchmarks, existing hallucination benchmarks [14, 19, 34] follow this VQA style that provides a simple QnA evaluation pipeline and intuitive metrics. POPE [19] and CIEM [14] curate images and QnA pairs with metrics such as accuracy, precision, and recall, where they use a specific question type inducing multi-choice or yes/no answers, called discriminative type. AMBER [34] includes an additional question type that induces open-ended caption generation as answers, called generative type. For such open-generation assessment, they extract the objects from the output text answers and compare them with the ground-truth objects in the image. These prior studies focus on manipulating texts of questions and answers only,", + "bbox": [ + 212, + 643, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Ye-Bin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/cf69220a14cf4c8f672df7ef0b1b365ad4252825f0556c840607853a7ea98c4a.jpg", + "table_caption": [ + "Table 1: Data statistics. Our BEAF dataset contains 26K image-question pairs, consisting of the original and manipulated ones. On average, an original image is associated with 3.45 manipulated images and 11.72 questions." + ], + "table_footnote": [], + "table_body": "
ImageImg-Q PairYesNo
Original5005,6532,0263,627
Manipulated1,72720,4656,26714,198
Total2,22726,1188,29317,825
", + "bbox": [ + 277, + 198, + 727, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ratio of upper-class categories", + "text_level": 1, + "bbox": [ + 416, + 292, + 583, + 304 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e24d158a9e1388e6e8825eb08336756f3f01b27fd5cd513cf3da0bd940306a87.jpg", + "image_caption": [ + "Removed objects (total: 1,727)" + ], + "image_footnote": [], + "bbox": [ + 228, + 316, + 485, + 434 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1fdf2aa9cd32bd516a63a06377f0481df96a98192bd056694af6b05a9c3f628.jpg", + "image_caption": [ + "Objects in question (total: 26,118)", + "Fig. 2: Ratio of removed objects from image and object in question. For convenience, we report the ratio of upper-class categories instead of the object class itself. The total number of removed objects is the same as the number of manipulated images, and one object in question is the same as the number of image-question pairs." + ], + "image_footnote": [], + "bbox": [ + 522, + 316, + 767, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which we call text-axis evaluation. However, if some objects always co-occur with another object, we cannot judge whether VLMs truly understand scene information accurately. Also, VLMs take multi-modal inputs, both images and text prompts; thus, evaluation considering only the text-axis would be insufficient as a hallucination assessment of VLMs. In this work, we manipulate the scene information and questions together, i.e., vision-/text-axes; it enables the fine-grained analysis of hallucination. We call it vision-/text-axes evaluation.", + "bbox": [ + 212, + 569, + 784, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 BEAF Evaluation Benchmark for Hallucination", + "text_level": 1, + "bbox": [ + 214, + 715, + 707, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We explore another axis, i.e., vision-axis, to make up for the deficiencies in the text-axis evaluation. Our key idea is to manipulate a visual scene by removing and inpainting objects within an image. This allows us to focus on changes and to check whether the VLMs are correctly aware of it. We describe how to curate our dataset in Sec. 3.1, and then elaborate our new metrics in Sec. 3.2", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BEAF Benchmark", + "bbox": [ + 606, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/52c8686b18ccccc1418f502deb7ae6977b05d55c26ff0796a3a3c7847d506d72.jpg", + "image_caption": [ + "Fig. 3: Image manipulation pipeline. We illustrate the image manipulation pipeline comprised of three stages. In Stage 1, we automatically remove target objects sharing the same semantic class from the given images. Stage 2 is to filter the automatically manipulated results based on the predefined rules, such as mask errors, remained shadows, and low-quality outcomes. Undesirable manipulations are either corrected or discarded during this stage. Finally, in Stage 3, human annotators engage in human-guided manipulation of the filtered images to achieve high-quality results." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 787, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Evaluation Dataset", + "text_level": 1, + "bbox": [ + 215, + 515, + 419, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Overview. Table 1 and Fig. 2 summarize our data statistics. The dataset consists of images and their associated QnA. The image type is divided into original and manipulated. Five hundred original images are sampled from the validation set of the MS-COCO dataset [20], which shares the same split with POPE [19], and their associated 1,727 manipulated images are generated by removing objects from the original images. On average, 3.45 manipulated images are created from the original images. The main question format is \"Is there {object} in this image?\" and {object} can be a positive object within the image or a negative object not in the image; the positive object is one of the objects in the image, and the negative object is the one not in the image. We adopt the positive and negative objects from the POPE for the original image and adapt the question for the manipulated image. If we remove the positive object from the original image to make the manipulated one, the positive object becomes the negative object in the manipulated one. The total dataset comprised 2,227 images and 26,118 image-question pairs. The \"Yes\" and \"No\" ground truth answers are 8,293 (31.75%) and 17,825 (68.25%), respectively.", + "bbox": [ + 212, + 550, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image Dataset Construction Pipeline. We manipulate the original images by removing objects within the image. To collect the manipulated images, we pass through a three-stage pipeline (see Fig. 3): the first stage involves auto", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Ye-Bin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "matic manipulation, followed by a filtering stage, and finally, a human-guided manipulation stage. We offer a detailed explanation of each stage as follows.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1) First Stage: Automatic Manipulation. For automatic object removal, we use an off-the-shelf inpainting model, LaMa [30], with mask guidance. While the initial option is to use the ground truth masks provided within the original dataset, we opt not to use them because of the low quality ground truth masks in MS-COCO. For example, polygon masks often cannot cover all areas of the corresponding object and can be propagated to the inpainting quality. Therefore, we employ a pre-trained mask extractor called SAM [17], utilizing bounding box annotations. Subsequently, we dilate the extracted mask to have a margin for more effective object removal, and remove the object with the inpainting model given the dilated mask guidance. We empirically choose to use LaMa over diffusion models, e.g., [28, 44], as we find that this specialized model performs better in the context of object removal.", + "bbox": [ + 212, + 176, + 785, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2) Second Stage: Filtering. Although we remove objects from the original images, achieving consistent quality in the outcome can be influenced by a myriad of factors. These factors include, but are not limited to, the size of the objects being removed and the complexity of the backgrounds they inhabit. Additionally, the presence of shadows remains unaddressed, as they are not effectively masked using SAM; the ground truth masks also do not capture the shadow. As shown on the right side of Fig. 3, it displays undesirable images due to the presence of shadow or a faucet. The human annotators classified the results obtained from the previous stage into usable or unusable, where only the usable results are further manipulated. The detailed filtering rules can be found in the supplementary material. We directly use the usable results and hand over the results that need further manipulation to the next stage.", + "bbox": [ + 212, + 359, + 787, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3) Third Stage: Human-guided Manipulation. The third stage of our workflow relies on human annotators to intervene and refine the images obtained from the second stage. As mentioned above, these filtered images often exhibit undesirable artifacts such as ghost shadows, afterimages, and fragmented objects. Even such a small artifact may provide strong cues indicating the past presence of objects that both humans and machines can deduce, e.g., a cat's tail suggests the past presence of the cat before the object removal. Thus, human annotators are asked to meticulously manipulate the images, either correcting or enhancing them. Annotators effectively refine visual data, aligning it more closely with real natural images in human perception.", + "bbox": [ + 212, + 541, + 787, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 3, our annotation process results in images that are free of the aforementioned artifacts. This manual intervention is crucial, as it ensures the fidelity and reliability of our benchmark dataset.", + "bbox": [ + 212, + 693, + 785, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 765, + 419, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The prior studies [14, 19, 34] suggest using traditional metrics such as accuracy, precision, recall, and F1 score, which we can also use the metrics in our dataset. In addition, we propose four new metrics: True Understanding (TU), IGnorance", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BEAF Benchmark", + "bbox": [ + 607, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(IG), StuBbornness (SB), and InDecision (ID), for more detailed evaluation by exploiting the distinctive configuration of before-/after-changes in our dataset.", + "bbox": [ + 215, + 145, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let $\\mathcal{S} = \\{(I_o, I_m, Q, A_o, A_m, R)_i\\}_{i=1}^{|S|}$ denote our dataset, where each tuple consists of an original image $I_o$ , a version of its manipulated image $I_m$ , their associated question $Q$ , their respective corresponding answers $A_o$ and $A_m$ , and a True/False flag $R$ that indicates whether the question is associated with the object being removed in the manipulated image $I_m$ . For compactness of our metric definitions, we define $\\text{Filter}(\\cdot)$ to extract the subset of $S$ satisfying input conditions as:", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F i l t e r} \\left(b _ {o}, b _ {m}, b _ {r}\\right) = \\{t | \\text {I s C o r r e c t} \\left(A _ {o}\\right) = b _ {o}, \\text {I s C o r r e c t} \\left(A _ {m}\\right) = b _ {m}, R = b _ {r}, t \\in \\mathcal {S} \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 295, + 772, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the tuple $t = (I_o, I_m, Q, A_o, A_m, R)$ . For example, Filter(True, False, True) is the subset of $S$ of which the element satisfy the conditions that are 1) the answers corresponding to respective original and manipulated images are correct and wrong, and 2) the question is the one related to the object removed in the manipulated images.", + "bbox": [ + 212, + 321, + 782, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "True Understanding. True Understanding (TU) measures whether the models truly understand the scene. It evaluates if the model can accurately answer to the question about the removed object. As shown in Fig. 1, if the apple is removed, a model that truly understands the change should switch its answer from \"Yes\" to \"No\" to the question \"Is there an apple?\" We define the TU metric as follows:", + "bbox": [ + 212, + 398, + 787, + 488 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T U} = \\frac {\\left| \\text {F i l t e r} (\\text {T r u e} , \\text {T r u e} , \\text {T r u e}) \\right|}{\\left| \\text {F i l t e r} (\\text {T r u e} \\vee \\text {F a l s e} , \\text {T r u e} \\vee \\text {F a l s e} , \\text {T r u e}) \\right|} \\cdot 1 0 0. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 497, + 784, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared to the traditional accuracy, our metric specifically considers the correctness of answers before and after manipulation. Therefore, it provides a conditional measure of accuracy, taking into account the model's response to changes within the scene. A higher value of this metric indicates a more accurate understanding of the scene by the model, as it demonstrates the model's ability to adapt its answers appropriately to the scene changes.", + "bbox": [ + 212, + 527, + 782, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ignorance. The second metric is IGnorance (IG), which measures the extent to which models lack knowledge about specific scene information. This metric quantifies ignorance by identifying instances where models provide incorrect answers to the question about the removed objects. As shown in Fig. 1, ignorance is evident if models incorrectly answer \"No\" to the presence of an apple before its removal and incorrectly answer \"Yes\" after the apple is removed, indicating a failure to recognize it. IG is defined as follows:", + "bbox": [ + 212, + 619, + 787, + 724 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I G} = \\frac {\\left| \\text {F i l t e r (F a l s e , F a l s e , T r u e)} \\right|}{\\left| \\text {F i l t e r (T r u e V F a l s e , T r u e V F a l s e , T r u e)} \\right|} \\cdot 1 0 0. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 734, + 784, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The lower the metric, the less ignorant the model is.", + "bbox": [ + 215, + 763, + 593, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stubbornness. What we observe is that VLMs tend to output the same answer repeatedly. It means that VLMs are biased toward the specific answer rather than arbitrary one when they do not know. This repetitive behavior is undesirable; thus, we introduce the metric of StuBbornness (SB), which measures the extent", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Ye-Bin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to which models adhere to their initial answers. Furthermore, we categorize SB into $\\mathrm{SB}_p$ and $\\mathrm{SB}_n$ where the subscripts of $p$ and $n$ correspond to the consistent positive (\"Yes\") and negative (\"No\") answers, respectively. We define them as follows:", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S B} _ {p} = \\frac {1 0 0 \\cdot | \\text {F i l t e r (T r u e , F a l s e , T r u e)} |}{| \\text {F i l t e r (T r u e \\vee F a l s e , T r u e \\vee F a l s e , T r u e)} |}, \\quad \\mathrm {S B} _ {n} = \\frac {1 0 0 \\cdot | \\text {F i l t e r (F a l s e , T r u e , T r u e)} |}{| \\text {F i l t e r (T r u e \\vee F a l s e , T r u e \\vee F a l s e , T r u e)} |}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 215, + 785, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SB is the summation of $\\mathrm{SB}_p$ and $\\mathrm{SB}_n$ and can be computed from $100 - \\mathrm{TU} - \\mathrm{IG}$ . The lower the metric, the less ignorant the model is.", + "bbox": [ + 212, + 250, + 782, + 279 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Indecision. Unlike the previous three metrics that focus on the removed objects in the manipulated images, InDecision (ID) focuses on the answers to the questions that are not relevant to the removed objects. These answers should not be changed even after the manipulation. Nevertheless, models, often change their answers about unmanipulated objects, which would imply that the model predicts a random answer without grounding it on the image. We track these behaviors with the ID metric defined as follows:", + "bbox": [ + 212, + 280, + 784, + 383 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I D} = \\frac {\\left| \\text {F i l t e r} (\\text {T r u e} , \\text {F a l s e} , \\text {F a l s e}) \\right| + \\left| \\text {F i l t e r} (\\text {F a l s e} , \\text {T r u e} , \\text {F a l s e}) \\right|}{\\left| \\text {F i l t e r} (\\text {T r u e} \\vee \\text {F a l s e} , \\text {T r u e} \\vee \\text {F a l s e} , \\text {F a l s e}) \\right|} \\cdot 1 0 0. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 395, + 785, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The lower the metric, the less ignorant the model is. Additionally, we combine TU and ID using a harmonic mean to make it easier to compare on our benchmark. We refer to it as F1-Score, which is computed as follows: $F1 = \\frac{2}{TU^{-1} + (100 - ID)^{-1}}$ .", + "bbox": [ + 214, + 428, + 782, + 477 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 214, + 498, + 366, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Sec. 4.1, we evaluate the current VLMs using our BEAF dataset. We employ both proposed and traditional metrics and compare their performance with other benchmarks. Furthermore, we conduct an analysis of image-wise dynamic correctness by visualizing image-wise object relationship containing vision and text axes. We provide a report on the object-wise error rate. In Sec. 4.2, we assess the performance of VLMs on the open generation task using our dataset, while also analyzing both the dataset and VLMs based on the CLIP score.", + "bbox": [ + 212, + 532, + 784, + 637 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "All the reported performance is the zero-shot performance of VLMs, including LLaVA [23], InstructBLIP [8], Shikra [5], and mPLUG-Owl [36]. We use prompt templates provided in their released codebases or papers.", + "bbox": [ + 212, + 638, + 785, + 683 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Main Results", + "text_level": 1, + "bbox": [ + 214, + 707, + 372, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation on BEAF benchmark. We evaluate several VLMs on our BEAF dataset using the proposed change-aware metrics, TU, IG, SB, ID, and F1. The evaluation results are presented in Table 2. Among the 13B-size models, LLaVA outperforms InstructBLIP in terms of TU, IG, and SB because InstructBLIP consistently answers \"Yes\" regardless of scene manipulation, and this result is related to its high value of $\\mathrm{SB}_p$ . InstructBLIP achieves better ID compared to LLaVA. Among 7B-size models, Shikra exhibits the best in TU and $\\mathrm{SB}_p$ . Shikra", + "bbox": [ + 212, + 733, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BEAF Benchmark", + "bbox": [ + 606, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1dc58c44bc9fccc29173c45ba95e04206fb309b910058f5978e8f47fb8b20805.jpg", + "table_caption": [ + "Table 2: Evaluation results on BEAF benchmark. We evaluate various VLMs on our benchmark. F1 is a harmonic mean of TU and 1-ID. Yes ratio is the percentage of total answers that are answered \"Yes\" for the removed object-related question. The former is for the original image, and the latter stands for the manipulated image. Underline stands for the best." + ], + "table_footnote": [], + "table_body": "
SizeModelTU(↑)IG(↓)SBp(↓)SBn(↓)ID(↓)F1(↑)Yes (%)
13BLLaVA-v1.524.30.272.03.56.438.696.4 → 72.2
InstructBLIP11.10.384.44.26.119.995.5 → 84.7
7BLLaVA-v1.532.60.161.36.05.648.593.9 → 61.4
InstructBLIP33.50.651.314.65.249.584.8 → 51.9
Shikra52.70.431.215.75.767.683.9 → 31.6
mPLUG-Owl224.60.072.13.37.038.896.7 → 72.1
", + "bbox": [ + 220, + 226, + 782, + 318 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4d80e89772a6212511b50847784212d31a115207be49af89d3dbc834fe96ca15.jpg", + "table_caption": [ + "Table 3: Traditional evaluation results on BEAF dataset. We evaluate various VLMs on our dataset in terms of accuracy, precision, recall, and F1 score. Note that the BEAF dataset includes both original and manipulated data but does not consider the changes for this traditional evaluation. Underline stands for the best." + ], + "table_footnote": [], + "table_body": "
SizeModelAcc.(↑)Precision(↑)Recall(↑)F1(↑)Yes (%)
13BLLaVA-v1.574.956.393.470.352.6
InstructBLIP71.552.894.367.756.7
7BLLaVA-v1.579.762.689.573.745.4
InstructBLIP82.068.380.774.037.5
Shikra84.574.976.975.932.6
mPLUG-Owl269.150.793.865.858.7
", + "bbox": [ + 220, + 406, + 782, + 511 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "is trained using a location-aware strategy, which could be beneficial for hallucination evaluation because knowledge of location aids in judging the existence of an object. The overall results indicate that, although it is rare to get all the answers wrong, it is common for the answers to remain the same even though the image has changed. This indicates that recent VLMs are not adequately reflective of such changes.", + "bbox": [ + 215, + 546, + 785, + 637 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Traditional evaluation on BEAF dataset. In addition to the proposed change-aware metrics, we also evaluate using traditional metrics such as accuracy, precision, recall, and F1 score, as shown in Table 3. While the trends in accuracy or F1 score are similar to those shown in TU results of Table 2, the absolute values of TU are much lower than those of the traditional metrics. This discrepancy suggests that, although a model achieves high performance in the existing benchmarks, our benchmark tells us that we cannot conclude low hallucination and the models may still suffer from hallucination. For example, although LLaVA-v1.5-13B achieves high performance on accuracy, the Yes ratio in Table 2 remains high even after the manipulation. It may include instances where the correct answer was guessed without actual reference to the given image. The experimental findings underscore the importance of observing changes in answers along the vision-axis to accurately evaluate hallucination in VLMs.", + "bbox": [ + 215, + 641, + 787, + 837 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Ye-Bin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b51017f2148922e95dfd337608e7b4b084a057d71780364f1045ca279e89e131.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 142, + 395, + 233 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f12711c2b3f2e8474e8f0e351ac8d7378ffc8cc32c2422eb301276e374107087.jpg", + "image_caption": [ + "Fig. 4: Visualization of image-wise object relationship. We visualize image-wise object relationship along the text and vision axes from Shikra. [Top] Original image samples. [Bottom] The object relation table given the manipulated images and questions. We color the correct answer blue and the wrong one red. The text axis stands for the target object queried in the question, and the vision axis for the removed object in the image. The \"none\" in the vision axis means the original image (not manipulated). Thereby, we can analyze influence between objects within a scene at once." + ], + "image_footnote": [], + "bbox": [ + 217, + 233, + 388, + 316 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b73916b14e5e81779ff87ab4407c068f9a741ba32636110748280b05d0049b06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 142, + 589, + 233 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3f1f3b70a1cf372774eaecff39359a2f173e26b2917dba53ed6db7db28fd64b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 239, + 581, + 316 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e3bb7ece665e16e8f5678700b7d6cba2f09b196ee3b9dd7aab732df981b17a4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 142, + 782, + 233 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/86ff2b31abb6e666dba9e9d57516fd95e90dcd065577e66d302d253a7b0b419d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 234, + 782, + 316 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with other benchmarks. We compare our metric TU with the performance for other benchmarks of VQAv2 [11] and GQA [15], as shown in Table 4. Each VQA performance value is reported from the proposed papers. The results indicate that, while mPLUG-Owl2 outperforms others on VQAv2 and LLaVA excels on GQA, Shikra leads in per", + "bbox": [ + 212, + 460, + 486, + 611 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "formance on our metric. We observe that there is a negative correlation between performance and the incidence of hallucination, which aligns with the previous observation [35]. Although improving performance on various benchmarks is challenging, it is essential to check multiple aspects of the performance, including hallucination. We believe that our benchmark provides valuable insights in conjunction with existing benchmarks.", + "bbox": [ + 212, + 612, + 787, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Image-wise object relationship. We visualize changes in dynamic object relationship along the vision and text axes, as shown in Fig. 4. Compared to our metrics providing comprehensive statistics as summarized scores, this visualization enables to track the changes in VLMs' responses at object-wise and scene-wise levels. The x-axis is the text axis where each object is associated with the question, \"Is there {object}?\"; the y-axis is the vision axis where the object is removed from the original image. The blue color is correct, the red one is wrong.", + "bbox": [ + 212, + 703, + 787, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the first column in Fig. 4, altering the scene affects the textual correctness. Specifically, removing either the couch or the book shifts the book's textual cor-", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b15836e1aaea5bdd368ef1f665bd88ae2f0ec0e8cb30a870673882619e587768.jpg", + "table_caption": [ + "Table 4: Comparison of performance on VQA and TU. We compare the accuracy on VQA datasets, VQAv2 and GQA, as well as TU on BEAF dataset." + ], + "table_footnote": [], + "table_body": "
Model (7B)VQAv2GQATU
LLaVA-v1.578.562.032.6
InstructBLIP-49.233.5
Shikra77.4-52.7
mPLUG-Owl279.456.124.6
", + "bbox": [ + 524, + 532, + 767, + 604 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BEAF Benchmark", + "bbox": [ + 607, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c6b8a87ffc47806c59d71c613b7a4c02da54362f5a20f63f3880eceba953a9d6.jpg", + "image_caption": [ + "Fig. 5: Object-wise error rate on the manipulated image-question pairs. We plot the error rate of each object to investigate which object is inaccurately inferred when it is removed. We exclude the objects manipulated less than 20 times. The solid line represents the average error rate of VLMs (7B), while the purple area indicates the $95\\%$ confidence interval." + ], + "image_footnote": [], + "bbox": [ + 266, + 148, + 733, + 280 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "rectness from correct to incorrect. This suggests a correlation between the couch and the book, indicating that complex scenes can lead to model hallucinations even in the absence of the book. In the second scene, when the object vase or book is removed, most of the incorrectness becomes correct. This could imply that the presence of a vase or a book might lead the model to misinterpret the scene. Through the simultaneous analysis of both vision and text axes, we observe immediate changes in the answers and offer interpretations for the dynamic interactions observed in the second and third scenes illustrated in the figure.", + "bbox": [ + 212, + 388, + 787, + 511 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Object-wise error rate. We compute the mean and $95\\%$ confidence interval of the object-wise error after the manipulation in Fig. 5. It represents whether the VLMs can accurately detect the removal. The results show that VLMs frequently fail to recognize certain objects like traffic lights, vases, and remotes. We speculate that the difficulty in recognizing these objects may stem from their complexity or their infrequent appearance in the training data.", + "bbox": [ + 212, + 511, + 787, + 602 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Additional Study", + "text_level": 1, + "bbox": [ + 214, + 621, + 406, + 636 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluation of open generation. In Table 5, we compare the results of the CHAIR [27] and TU metrics to evaluate the hallucination in the open-ended generation answers on our dataset. Briefly, $\\mathrm{CHAIR}_I$ and $\\mathrm{CHAIR}_S$ count the hallucinated instances and sentences containing these objects in the generated output, respectively. The lower the score, the better. We observe that if the open generation contains the hallucinated object and sentences, TU is also lower, as seen with InstructBLIP-13B.", + "bbox": [ + 212, + 643, + 517, + 839 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1a1e42b1e7959d0e82e4f758036dc32d89aa4110ceaa67f0da089b46fba4c820.jpg", + "table_caption": [ + "Table 5: Comparison of CHAIR metric and TU. We measure the CHAIR score based on the answers derived from the images in BEAF using the prompt \"Describe this image.\"" + ], + "table_footnote": [], + "table_body": "
SizeModelCHAIR1CHAIRSTU
13BLLaVA-v1.521.672.324.3
InstructBLIP28.181.111.1
7BLLaVA-v1.522.373.932.6
InstructBLIP21.067.533.5
Shikra23.677.052.7
mPLUG-Owl225.178.424.6
", + "bbox": [ + 529, + 729, + 782, + 814 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Ye-Bin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d22c0ec6b2135da95057ec1495e3184f36228cc8eeee92da5e0b8be0e81f4d8b.jpg", + "image_caption": [ + "Fig. 6: Samples with CLIP Score. We visualize samples and display the CLIP Score below. Additionally, we report the results of Faster R-CNN [26] using the magnifying glass icon. After removing the object, the removed object is not detected, which demonstrates the manipulation quality." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 787, + 311 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "CILP Score. We compare the CLIP Score [13] between original and manipulated images to ensure the target object has been well-removed. A lot of VLMs utilize the visual encoder from CLIP, prompting us to analyze CLIP's performance with our datasets. We measure the CLIP score between an object and a prompt \"a photo", + "bbox": [ + 214, + 412, + 521, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of {object}. We use a cropped object image based on the GT bounding box. If the image has multiple objects of the same semantic class, we use the average CLIP score of all the same semantic objects. For the manipulated image, {object} is the removed object. As shown in Table 6, the manipulated images have lower CLIP scores compared to their original counterparts.", + "bbox": [ + 212, + 534, + 784, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also visualize the images with the CLIP Score and the multi-class classification results in Fig. 6. The multi-class classification correctly identifies the depicted objects. In addition, Fig. 7 visualizes an open generation sample result, where red-colored text indicates hallucinated words. Before manipulation, all the VLMs mention \"people\" in the background; however, they still mention \"people\" even though all the persons are removed from the image. This implies that the commonly used visual encoder might be a non-negligible source of hallucination.", + "bbox": [ + 212, + 609, + 784, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 215, + 743, + 508, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We present a comprehensive analysis of hallucination using our BEAF benchmark. Unlike conventional hallucination benchmarks that measure accuracy by changing only the text-axis question, we also manipulate images by removing objects and propose four sophisticated metrics to track the changes in answers", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e95344c6ed683babee8348d30542b6e5b59694babaa6e6c3649a54e3c052e472.jpg", + "table_caption": [ + "Table 6: CLIP Score. We measure the CLIP Score of original and manipulated images with the text prompt." + ], + "table_footnote": [], + "table_body": "
TypeCLIP Score
Original53.7
Manipulated47.9
", + "bbox": [ + 562, + 470, + 756, + 523 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BEAF Benchmark", + "bbox": [ + 607, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f6d9f70d17cefcae75666afbb809b71be2e291da3b7df22fffa144d108365209.jpg", + "image_caption": [ + "Fig. 7: Qualitative example on open-ended generation. We report the changes in answers for the open-ended generation task with the prompt \"Describe this image.\" as the image undergoes manipulation. The red color highlights the hallucination parts." + ], + "image_footnote": [], + "bbox": [ + 251, + 143, + 750, + 397 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "according to image changes. This approach enables us to a more discerning analysis of the hallucination in two-axis views, e.g., vision and text axes. We can find with our metric that traditional accuracy may include instances where the correct answer was guessed without actual reference to the given image. We found that our TU is a more calibrated metric, which allows us to observe the intriguing phenomenon that has not been revealed in the prior work: VLMs tend to be confused if the complexity of an image is high. (refer to Sec. A in the supplementary material). Additionally, we introduce the visualization method to track the correctness as the objects are removed. We believe that our benchmark presents a comprehensive analysis.", + "bbox": [ + 212, + 488, + 787, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation. While our benchmark offers a more fine-grained evaluation, it still faces the constraints of the MS-COCO dataset. These include a limited variety of object categories and an uneven distribution of object instances. While we utilized current state-of-the-art in-painting models and mask extractors, we manually adjusted the images to ensure image quality and dataset integrity. This step prevented us from achieving a fully automated creation process. Although such limitations exist, these aspects are expected to naturally diminish or disappear as in-painting models advance, making our dataset construction pipeline fully automatic and our method applicable to all images. Furthermore, we foresee future evaluations expanding in scale through manipulations like object removal, insertion, or resizing. While our current method may appear as a modest evaluation benchmark, we hope that our work becomes the first step toward controllable hallucination evaluation.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Ye-Bin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 217, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub; No.2022-0-00124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities; No.RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH))", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 292, + 323, + 308 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv (2023) 1, 4", + "2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: IEEE International Conference on Computer Vision (ICCV) (2019) 2", + "3. Bisk, Y., Zellers, R., Gao, J., Choi, Y., et al.: Piqa: Reasoning about physical commonsense in natural language. In: AAAI Conference on Artificial Intelligence (AAAI) (2020) 1", + "4. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv (2023) 4", + "5. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv (2023) 1, 4, 9", + "6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/1, 4", + "7. Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., Schulman, J.: Training verifiers to solve math word problems. arXiv (2021) 1", + "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 1, 4, 9", + "9. Deng, A., Chen, Z., Hooi, B.: Seeing is believing: Mitigating hallucination in large vision-language models via clip-guided decoding. arXiv (2024) 4", + "10. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv (2023) 4", + "11. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 2, 11", + "12. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. In: AAAI Conference on Artificial Intelligence (AAAI) (2024) 2, 4" + ], + "bbox": [ + 217, + 325, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BEAF Benchmark", + "bbox": [ + 606, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv (2021) 13", + "14. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv (2023) 2, 4, 7", + "15. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 2, 11", + "16. Hyun, L., Sung-Bin, K., Han, S., Yu, Y., Oh, T.H.: Smile: Multimodal dataset for understanding laughter in video with language models. arXiv (2023) 1", + "17. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv (2023) 7", + "18. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv (2023) 4", + "19. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. In: Conference on Empirical Methods in Natural Language Processing (EMNLP) (2023) 2, 4, 6, 7", + "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV) (2014) 6", + "21. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: International Conference on Learning Representations (ICLR) (2024) 4", + "22. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models. arXiv (2024) 4", + "23. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv (2023) 1, 4, 9", + "24. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 1, 4", + "25. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. In: International Conference on Learning Representations (ICLR) (2024) 4", + "26. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: Advances in Neural Information Processing Systems (NeurIPS) (2015) 13", + "27. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: Conference on Empirical Methods in Natural Language Processing (EMNLP) (2018) 12", + "28. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 7", + "29. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv (2023) 4", + "30. Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. In: IEEE Winter Conf. on Applications of Computer Vision (WACV) (2022) 7", + "31. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. 3(6), 7 (2023) 4" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Ye-Bin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv (2023) 1, 4", + "33. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv (2023) 1, 4", + "34. Wang, J., Wang, Y., Xu, G., Zhang, J., Gu, Y., Jia, H., Yan, M., Zhang, J., Sang, J.: An llm-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv (2023) 2, 4, 7", + "35. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023) 4, 11", + "36. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv (2023) 9", + "37. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv (2023) 4", + "38. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics (TACL) (2014) 2", + "39. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023) 2, 4", + "40. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. In: International Conference on Learning Representations (ICLR) (2024) 4", + "41. Zhao, Y., Li, Z., Jin, Z., Zhang, F., Zhao, H., Dou, C., Tao, Z., Xu, X., Liu, D.: Enhancing the spatial awareness capability of multi-modal large language model. arXiv (2023) 2", + "42. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models. In: International Conference on Learning Representations (ICLR) (2024) 2, 4", + "43. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv (2023) 4", + "44. Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv (2023) 7" + ], + "bbox": [ + 212, + 146, + 787, + 674 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BEAF Benchmark", + "bbox": [ + 606, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_model.json b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1cdbbbc0b545afc76cfcf8330f7ea2666f148804 --- /dev/null +++ b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_model.json @@ -0,0 +1,2192 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.142, + 0.788, + 0.187 + ], + "angle": 0, + "content": "BEAF: Observing BEFORE-AFTER Changes to Evaluate Hallucination in Vision-language Models" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.212, + 0.784, + 0.228 + ], + "angle": 0, + "content": "Moon Ye-Bin\\(^{1*}\\) Nam Hyeon-Woo\\(^{1*}\\) Wonseok Choi\\(^{2}\\) Tae-Hyun Oh\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.239, + 0.782, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\)Dept. of EE, POSTECH, Korea \\(^{2}\\)Grad. School of AI, POSTECH, Korea \\(^{3}\\)Institute for Convergence Research and Education in Advanced Technology, Yonsei University, Korea." + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.283, + 0.71, + 0.296 + ], + "angle": 0, + "content": "{ybmoon, hyeonw.nam, wonseok.c, taehyun}@postech.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.334, + 0.744, + 0.64 + ], + "angle": 0, + "content": "Abstract. Vision language models (VLMs) perceive the world through a combination of a visual encoder and a large language model (LLM). The visual encoder, pre-trained on large-scale vision-text datasets, provides zero-shot generalization to visual data, and the LLM endows its high reasoning ability to VLMs. It leads VLMs to achieve high performance on wide benchmarks without fine-tuning, exhibiting zero or few-shot capability. However, recent studies show that VLMs are vulnerable to hallucination. This undesirable behavior degrades reliability and credibility, thereby making users unable to fully trust the output from VLMs. To enhance trustworthiness and better tackle the hallucination of VLMs, we curate a new evaluation dataset, called the BEfore-AFTER hallucination dataset (BEAF), and introduce new metrics: True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID). Unlike prior works that focus only on constructing questions and answers, the key idea of our benchmark is to manipulate visual scene information by image editing models and to design the metrics based on scene changes. This allows us to clearly assess whether VLMs correctly understand a given scene by observing the ability to perceive changes. We also visualize image-wise object relationship by virtue of our two-axis view: vision and text. Upon evaluating VLMs with our dataset, we observed that our metrics reveal different aspects of VLM hallucination that have not been reported before. Project page: https://beafbench.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.665, + 0.376, + 0.68 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Vision language models (VLMs) [1, 5, 8, 23, 24] have recently emerged by its multi-modal reasoning capability. It is indeed promising research toward making intelligent agents perceive the world. VLMs comprise a vision encoder as a visual perception module and a large language model (LLM) as a reasoning module. Recent LLMs [6, 32, 33] have exhibited powerful reasoning capabilities across various tasks, including commonsense reasoning [3, 16] and math [7], in zero-/few-shot ways. This success motivates researchers to integrate LLMs with vision encoders to open LLMs' eyes to see the world. Leveraging the reasoning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.513, + 0.84 + ], + "angle": 0, + "content": "* Authors contributed equally to this work." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Ye-Bin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "capabilities of LLMs alongside visual perception enables VLMs to excel in various visual tasks, such as image captioning [2,38] and visual question answering [11, 15] without task-specific fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.196, + 0.785, + 0.346 + ], + "angle": 0, + "content": "Despite this encouraging demonstration, VLMs are susceptible to the hallucination [12,14,19,34,39,42] wherein their outputs do not reflect facts presented in input images. For example, although a given input image does not present an object, VLMs often incorrectly answer for the presence of the object. This tendency is problematic, akin to spreading falsehoods in human communication. Hallucination prevents genuine communication, consequently diminishing the reliability of interactions between agents or humans. No matter how good VLM's answer is, if hallucination is included in the answer, it can undermine trust in the system's reliability and credibility. Therefore, it is crucial to address hallucinations to improve the system's responsibility and trustworthiness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.787, + 0.576 + ], + "angle": 0, + "content": "Understanding the impact of hallucinations on model performance requires the development of an assessment framework, which requires a deep exploration of the underlying causes. The prior works [14, 41, 42] have been proposed to unveil the causation of hallucinations in terms of data and models by reporting their performance on existing hallucination benchmarks. As a widely used hallucination benchmark, POPE [19] follows the question-and-answer (QnA) format. For the convenience of measuring, POPE uses a discriminative question, e.g., \"Is there {object} in the image?\", and gets a YES or NO answer. The following hallucination benchmarks [14, 19, 34] commonly leverage LLMs or human annotators to generate question prompts with available scene information such as bounding boxes and captions. This text-axis evaluation may assess certain aspects of the hallucination performance well, but it is hard to disentangle sources of hallucination, particularly in VLMs that handle both visual and text modalities. Thus, to thoroughly analyze and understand the behavior of VLMs, it is both natural and necessary to consider both modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.787, + 0.746 + ], + "angle": 0, + "content": "In this work, we propose a BEFORE-AFTER (BEAF) hallucination evaluation benchmark. This benchmark comprises a dataset enhanced along the vision-axis and metrics that are aware of changes. We manipulate scenes by selectively removing objects on the vision-axis while augmenting QnA pairs along the text-axis. We can observe how the models' answers change as the image is manipulated for the same question. For example, as illustrated in Fig. 1, if we remove an apple from the given image and then ask \"Is there an apple?\", the answer from the model should change from \"Yes\" to \"No\", between the original and modified images, if it truly relies on what it sees. Therefore, for a more thorough evaluation and analysis, we need change-aware metrics beyond just using standard accuracy [14, 19, 34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We propose four new metrics: True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID), which are enabled by our vision and text augmentation. TU measures whether the model answers correctly before and after the manipulation. IG evaluates whether the models perceive the images consistently both before and after manipulation. SB is related to the phenomenon of giving the same answer instead of correctly recognizing the image. ID measures" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.608, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEAF Benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.145, + 0.495, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.147, + 0.761, + 0.291 + ], + "angle": 0, + "content": "
MTQuestionMetricsCorrectnessAnswer
Text-Acc○-
Img. & TextRemoved ObjectTU○ → ○Y → N
IG× → ×N → Y
SBp○ → ×Y → Y
SBn× → ○N → N
Unremoved ObjectID○ → ×\n× → ○Y → N\nN → Y
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.512, + 0.292, + 0.601, + 0.3 + ], + "angle": 0, + "content": "MT: Manipulation Target" + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.306, + 0.558, + 0.317 + ], + "angle": 0, + "content": "BEAF Evaluation (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.319, + 0.769, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Fig. 1: Illustration of BEbefore-AFTER (BEAF) benchmark. We present a comparison between traditional evaluation benchmarks and our BEAF benchmark for assessing the hallucination behavior in VLMs. Traditional evaluation methods solely manipulate questions based on the existence of an object and measure accuracy or F1 score. In contrast, our BEAF benchmark not only constructs questions but also manipulates images and tracks changes in answers as the images undergo manipulation. The BEAF benchmark introduces novel metrics including True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID), which consider the changes in answers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.719 + ], + "angle": 0, + "content": "cases where the answer changes even though the target object does not change or does not exist in the image. The first three metrics address questions regarding target objects that have been removed, while the last one belongs to questions about other objects not related to the targets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Based on our BEAF dataset and change-aware metrics, we can evaluate and understand the models, even the parts we did not know through the previous methods. The evaluation results imply that answers assumed as non-hallucination from the existing hallucination benchmarks could be a hallucination. We also analyze the influence of each object in the scene based on the variation along the vision and text axes. In addition, we include a discussion on open-generation answers and CLIPScore results. Our key contributions are as follows:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Ye-Bin et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.783, + 0.205 + ], + "angle": 0, + "content": "- We propose a benchmark, called BEFORE-AFTER (BEAF), to evaluate the hallucination of large vision-language models. Our benchmark comprises a dataset that considers both text- and vision-axes, and change-aware metrics, which allow us to perform granular evaluation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.208, + 0.784, + 0.235 + ], + "angle": 0, + "content": "- We demonstrate that previously reported outcomes deemed non-hallucinatory in the prior text-axis-only evaluation benchmarks may be hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.238, + 0.784, + 0.265 + ], + "angle": 0, + "content": "- We also visualize the impact of relationships between objects within hallucinatory images based on the BEAF results." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.388, + 0.309 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.785, + 0.461 + ], + "angle": 0, + "content": "Vision Language Models (VLMs). VLMs [1, 4, 5, 8, 10, 23-25, 40, 43] have demonstrated their impressive image describing and reasoning ability based on powerful large language models (LLMs) [6, 31-33]. Generally, VLMs consist of three parts: vision encoder, LLM, and vision-language alignment module. The vision information extracted from the vision encoder is translated to the LLM-understandable token by passing through the alignment module. LLM takes a prompt text and visual tokens as input and outputs response text tokens according to a given image and prompt. Their training process is divided into multiple steps depending on which parts of VLMs are trained." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.787, + 0.644 + ], + "angle": 0, + "content": "Hallucination in VLMs. Hallucination [12, 19, 22, 34, 35, 39] in VLMs refers to the case where the generated text answer does not reflect the true contents of the provided images but rather relies on the internal knowledge of the models. This behavior might cause harmful consequences in high-risk applications such as the medical domain or auto-driving scenarios. This undesirable behavior could be caused by various factors, e.g., data imbalance, image resolution, and model capacity. The assessment of hallucination is also related to the true understanding of VLMs. If VLMs respond to the given input correctly rather than relying on innate knowledge, VLMs are considered to have a perfect understanding of the perceived world. Numerous methodologies [12, 21, 42] have been developed to tackle the issue of hallucination in VLMs by addressing data biases [14, 21], optimization techniques [12, 18, 29], and post-processing methodologies [9, 37, 42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Evaluating Hallucination in VLMs. Evaluating hallucination performance is essential to guide the research of VLM. Previous studies have introduced hallucination benchmarks to understand and assess these undesirable phenomena of VLMs. Motivated by traditional visual and question benchmarks, existing hallucination benchmarks [14, 19, 34] follow this VQA style that provides a simple QnA evaluation pipeline and intuitive metrics. POPE [19] and CIEM [14] curate images and QnA pairs with metrics such as accuracy, precision, and recall, where they use a specific question type inducing multi-choice or yes/no answers, called discriminative type. AMBER [34] includes an additional question type that induces open-ended caption generation as answers, called generative type. For such open-generation assessment, they extract the objects from the output text answers and compare them with the ground-truth objects in the image. These prior studies focus on manipulating texts of questions and answers only," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEAF Benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Table 1: Data statistics. Our BEAF dataset contains 26K image-question pairs, consisting of the original and manipulated ones. On average, an original image is associated with 3.45 manipulated images and 11.72 questions." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.199, + 0.728, + 0.271 + ], + "angle": 0, + "content": "
ImageImg-Q PairYesNo
Original5005,6532,0263,627
Manipulated1,72720,4656,26714,198
Total2,22726,1188,29317,825
" + }, + { + "type": "title", + "bbox": [ + 0.418, + 0.293, + 0.584, + 0.305 + ], + "angle": 0, + "content": "Ratio of upper-class categories" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.317, + 0.486, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.449, + 0.445, + 0.46 + ], + "angle": 0, + "content": "Removed objects (total: 1,727)" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.318, + 0.768, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.449, + 0.736, + 0.46 + ], + "angle": 0, + "content": "Objects in question (total: 26,118)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.477, + 0.785, + 0.533 + ], + "angle": 0, + "content": "Fig. 2: Ratio of removed objects from image and object in question. For convenience, we report the ratio of upper-class categories instead of the object class itself. The total number of removed objects is the same as the number of manipulated images, and one object in question is the same as the number of image-question pairs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.785, + 0.675 + ], + "angle": 0, + "content": "which we call text-axis evaluation. However, if some objects always co-occur with another object, we cannot judge whether VLMs truly understand scene information accurately. Also, VLMs take multi-modal inputs, both images and text prompts; thus, evaluation considering only the text-axis would be insufficient as a hallucination assessment of VLMs. In this work, we manipulate the scene information and questions together, i.e., vision-/text-axes; it enables the fine-grained analysis of hallucination. We call it vision-/text-axes evaluation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.716, + 0.709, + 0.733 + ], + "angle": 0, + "content": "3 BEAF Evaluation Benchmark for Hallucination" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We explore another axis, i.e., vision-axis, to make up for the deficiencies in the text-axis evaluation. Our key idea is to manipulate a visual scene by removing and inpainting objects within an image. This allows us to focus on changes and to check whether the VLMs are correctly aware of it. We describe how to curate our dataset in Sec. 3.1, and then elaborate our new metrics in Sec. 3.2" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Ye-Bin et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.788, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.385, + 0.789, + 0.483 + ], + "angle": 0, + "content": "Fig. 3: Image manipulation pipeline. We illustrate the image manipulation pipeline comprised of three stages. In Stage 1, we automatically remove target objects sharing the same semantic class from the given images. Stage 2 is to filter the automatically manipulated results based on the predefined rules, such as mask errors, remained shadows, and low-quality outcomes. Undesirable manipulations are either corrected or discarded during this stage. Finally, in Stage 3, human annotators engage in human-guided manipulation of the filtered images to achieve high-quality results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.516, + 0.421, + 0.53 + ], + "angle": 0, + "content": "3.1 Evaluation Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Dataset Overview. Table 1 and Fig. 2 summarize our data statistics. The dataset consists of images and their associated QnA. The image type is divided into original and manipulated. Five hundred original images are sampled from the validation set of the MS-COCO dataset [20], which shares the same split with POPE [19], and their associated 1,727 manipulated images are generated by removing objects from the original images. On average, 3.45 manipulated images are created from the original images. The main question format is \"Is there {object} in this image?\" and {object} can be a positive object within the image or a negative object not in the image; the positive object is one of the objects in the image, and the negative object is the one not in the image. We adopt the positive and negative objects from the POPE for the original image and adapt the question for the manipulated image. If we remove the positive object from the original image to make the manipulated one, the positive object becomes the negative object in the manipulated one. The total dataset comprised 2,227 images and 26,118 image-question pairs. The \"Yes\" and \"No\" ground truth answers are 8,293 (31.75%) and 17,825 (68.25%), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Image Dataset Construction Pipeline. We manipulate the original images by removing objects within the image. To collect the manipulated images, we pass through a three-stage pipeline (see Fig. 3): the first stage involves auto" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.608, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "BEAF Benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "matic manipulation, followed by a filtering stage, and finally, a human-guided manipulation stage. We offer a detailed explanation of each stage as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.359 + ], + "angle": 0, + "content": "1) First Stage: Automatic Manipulation. For automatic object removal, we use an off-the-shelf inpainting model, LaMa [30], with mask guidance. While the initial option is to use the ground truth masks provided within the original dataset, we opt not to use them because of the low quality ground truth masks in MS-COCO. For example, polygon masks often cannot cover all areas of the corresponding object and can be propagated to the inpainting quality. Therefore, we employ a pre-trained mask extractor called SAM [17], utilizing bounding box annotations. Subsequently, we dilate the extracted mask to have a margin for more effective object removal, and remove the object with the inpainting model given the dilated mask guidance. We empirically choose to use LaMa over diffusion models, e.g., [28, 44], as we find that this specialized model performs better in the context of object removal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.542 + ], + "angle": 0, + "content": "2) Second Stage: Filtering. Although we remove objects from the original images, achieving consistent quality in the outcome can be influenced by a myriad of factors. These factors include, but are not limited to, the size of the objects being removed and the complexity of the backgrounds they inhabit. Additionally, the presence of shadows remains unaddressed, as they are not effectively masked using SAM; the ground truth masks also do not capture the shadow. As shown on the right side of Fig. 3, it displays undesirable images due to the presence of shadow or a faucet. The human annotators classified the results obtained from the previous stage into usable or unusable, where only the usable results are further manipulated. The detailed filtering rules can be found in the supplementary material. We directly use the usable results and hand over the results that need further manipulation to the next stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.693 + ], + "angle": 0, + "content": "3) Third Stage: Human-guided Manipulation. The third stage of our workflow relies on human annotators to intervene and refine the images obtained from the second stage. As mentioned above, these filtered images often exhibit undesirable artifacts such as ghost shadows, afterimages, and fragmented objects. Even such a small artifact may provide strong cues indicating the past presence of objects that both humans and machines can deduce, e.g., a cat's tail suggests the past presence of the cat before the object removal. Thus, human annotators are asked to meticulously manipulate the images, either correcting or enhancing them. Annotators effectively refine visual data, aligning it more closely with real natural images in human perception." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.787, + 0.739 + ], + "angle": 0, + "content": "As shown in Fig. 3, our annotation process results in images that are free of the aforementioned artifacts. This manual intervention is crucial, as it ensures the fidelity and reliability of our benchmark dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.766, + 0.421, + 0.78 + ], + "angle": 0, + "content": "3.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The prior studies [14, 19, 34] suggest using traditional metrics such as accuracy, precision, recall, and F1 score, which we can also use the metrics in our dataset. In addition, we propose four new metrics: True Understanding (TU), IGnorance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Ye-Bin et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "(IG), StuBbornness (SB), and InDecision (ID), for more detailed evaluation by exploiting the distinctive configuration of before-/after-changes in our dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Let \\( \\mathcal{S} = \\{(I_o, I_m, Q, A_o, A_m, R)_i\\}_{i=1}^{|S|} \\) denote our dataset, where each tuple consists of an original image \\( I_o \\), a version of its manipulated image \\( I_m \\), their associated question \\( Q \\), their respective corresponding answers \\( A_o \\) and \\( A_m \\), and a True/False flag \\( R \\) that indicates whether the question is associated with the object being removed in the manipulated image \\( I_m \\). For compactness of our metric definitions, we define \\( \\text{Filter}(\\cdot) \\) to extract the subset of \\( S \\) satisfying input conditions as:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.296, + 0.773, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F i l t e r} \\left(b _ {o}, b _ {m}, b _ {r}\\right) = \\{t | \\text {I s C o r r e c t} \\left(A _ {o}\\right) = b _ {o}, \\text {I s C o r r e c t} \\left(A _ {m}\\right) = b _ {m}, R = b _ {r}, t \\in \\mathcal {S} \\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.784, + 0.398 + ], + "angle": 0, + "content": "where the tuple \\( t = (I_o, I_m, Q, A_o, A_m, R) \\). For example, Filter(True, False, True) is the subset of \\( S \\) of which the element satisfy the conditions that are 1) the answers corresponding to respective original and manipulated images are correct and wrong, and 2) the question is the one related to the object removed in the manipulated images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.788, + 0.489 + ], + "angle": 0, + "content": "True Understanding. True Understanding (TU) measures whether the models truly understand the scene. It evaluates if the model can accurately answer to the question about the removed object. As shown in Fig. 1, if the apple is removed, a model that truly understands the change should switch its answer from \"Yes\" to \"No\" to the question \"Is there an apple?\" We define the TU metric as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.498, + 0.785, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T U} = \\frac {\\left| \\text {F i l t e r} (\\text {T r u e} , \\text {T r u e} , \\text {T r u e}) \\right|}{\\left| \\text {F i l t e r} (\\text {T r u e} \\vee \\text {F a l s e} , \\text {T r u e} \\vee \\text {F a l s e} , \\text {T r u e}) \\right|} \\cdot 1 0 0. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.784, + 0.619 + ], + "angle": 0, + "content": "Compared to the traditional accuracy, our metric specifically considers the correctness of answers before and after manipulation. Therefore, it provides a conditional measure of accuracy, taking into account the model's response to changes within the scene. A higher value of this metric indicates a more accurate understanding of the scene by the model, as it demonstrates the model's ability to adapt its answers appropriately to the scene changes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.725 + ], + "angle": 0, + "content": "Ignorance. The second metric is IGnorance (IG), which measures the extent to which models lack knowledge about specific scene information. This metric quantifies ignorance by identifying instances where models provide incorrect answers to the question about the removed objects. As shown in Fig. 1, ignorance is evident if models incorrectly answer \"No\" to the presence of an apple before its removal and incorrectly answer \"Yes\" after the apple is removed, indicating a failure to recognize it. IG is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.735, + 0.785, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I G} = \\frac {\\left| \\text {F i l t e r (F a l s e , F a l s e , T r u e)} \\right|}{\\left| \\text {F i l t e r (T r u e V F a l s e , T r u e V F a l s e , T r u e)} \\right|} \\cdot 1 0 0. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.765, + 0.594, + 0.779 + ], + "angle": 0, + "content": "The lower the metric, the less ignorant the model is." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Stubbornness. What we observe is that VLMs tend to output the same answer repeatedly. It means that VLMs are biased toward the specific answer rather than arbitrary one when they do not know. This repetitive behavior is undesirable; thus, we introduce the metric of StuBbornness (SB), which measures the extent" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "BEAF Benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "to which models adhere to their initial answers. Furthermore, we categorize SB into \\(\\mathrm{SB}_p\\) and \\(\\mathrm{SB}_n\\) where the subscripts of \\(p\\) and \\(n\\) correspond to the consistent positive (\"Yes\") and negative (\"No\") answers, respectively. We define them as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.217, + 0.786, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S B} _ {p} = \\frac {1 0 0 \\cdot | \\text {F i l t e r (T r u e , F a l s e , T r u e)} |}{| \\text {F i l t e r (T r u e \\vee F a l s e , T r u e \\vee F a l s e , T r u e)} |}, \\quad \\mathrm {S B} _ {n} = \\frac {1 0 0 \\cdot | \\text {F i l t e r (F a l s e , T r u e , T r u e)} |}{| \\text {F i l t e r (T r u e \\vee F a l s e , T r u e \\vee F a l s e , T r u e)} |}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.784, + 0.28 + ], + "angle": 0, + "content": "SB is the summation of \\(\\mathrm{SB}_p\\) and \\(\\mathrm{SB}_n\\) and can be computed from \\(100 - \\mathrm{TU} - \\mathrm{IG}\\). The lower the metric, the less ignorant the model is." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.785, + 0.385 + ], + "angle": 0, + "content": "Indecision. Unlike the previous three metrics that focus on the removed objects in the manipulated images, InDecision (ID) focuses on the answers to the questions that are not relevant to the removed objects. These answers should not be changed even after the manipulation. Nevertheless, models, often change their answers about unmanipulated objects, which would imply that the model predicts a random answer without grounding it on the image. We track these behaviors with the ID metric defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.396, + 0.786, + 0.419 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I D} = \\frac {\\left| \\text {F i l t e r} (\\text {T r u e} , \\text {F a l s e} , \\text {F a l s e}) \\right| + \\left| \\text {F i l t e r} (\\text {F a l s e} , \\text {T r u e} , \\text {F a l s e}) \\right|}{\\left| \\text {F i l t e r} (\\text {T r u e} \\vee \\text {F a l s e} , \\text {T r u e} \\vee \\text {F a l s e} , \\text {F a l s e}) \\right|} \\cdot 1 0 0. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.429, + 0.784, + 0.478 + ], + "angle": 0, + "content": "The lower the metric, the less ignorant the model is. Additionally, we combine TU and ID using a harmonic mean to make it easier to compare on our benchmark. We refer to it as F1-Score, which is computed as follows: \\( F1 = \\frac{2}{TU^{-1} + (100 - ID)^{-1}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.499, + 0.367, + 0.517 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.785, + 0.638 + ], + "angle": 0, + "content": "In Sec. 4.1, we evaluate the current VLMs using our BEAF dataset. We employ both proposed and traditional metrics and compare their performance with other benchmarks. Furthermore, we conduct an analysis of image-wise dynamic correctness by visualizing image-wise object relationship containing vision and text axes. We provide a report on the object-wise error rate. In Sec. 4.2, we assess the performance of VLMs on the open generation task using our dataset, while also analyzing both the dataset and VLMs based on the CLIP score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.787, + 0.684 + ], + "angle": 0, + "content": "All the reported performance is the zero-shot performance of VLMs, including LLaVA [23], InstructBLIP [8], Shikra [5], and mPLUG-Owl [36]. We use prompt templates provided in their released codebases or papers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.373, + 0.723 + ], + "angle": 0, + "content": "4.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Evaluation on BEAF benchmark. We evaluate several VLMs on our BEAF dataset using the proposed change-aware metrics, TU, IG, SB, ID, and F1. The evaluation results are presented in Table 2. Among the 13B-size models, LLaVA outperforms InstructBLIP in terms of TU, IG, and SB because InstructBLIP consistently answers \"Yes\" regardless of scene manipulation, and this result is related to its high value of \\(\\mathrm{SB}_p\\). InstructBLIP achieves better ID compared to LLaVA. Among 7B-size models, Shikra exhibits the best in TU and \\(\\mathrm{SB}_p\\). Shikra" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Ye-Bin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.214 + ], + "angle": 0, + "content": "Table 2: Evaluation results on BEAF benchmark. We evaluate various VLMs on our benchmark. F1 is a harmonic mean of TU and 1-ID. Yes ratio is the percentage of total answers that are answered \"Yes\" for the removed object-related question. The former is for the original image, and the latter stands for the manipulated image. Underline stands for the best." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.227, + 0.784, + 0.319 + ], + "angle": 0, + "content": "
SizeModelTU(↑)IG(↓)SBp(↓)SBn(↓)ID(↓)F1(↑)Yes (%)
13BLLaVA-v1.524.30.272.03.56.438.696.4 → 72.2
InstructBLIP11.10.384.44.26.119.995.5 → 84.7
7BLLaVA-v1.532.60.161.36.05.648.593.9 → 61.4
InstructBLIP33.50.651.314.65.249.584.8 → 51.9
Shikra52.70.431.215.75.767.683.9 → 31.6
mPLUG-Owl224.60.072.13.37.038.896.7 → 72.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.339, + 0.787, + 0.394 + ], + "angle": 0, + "content": "Table 3: Traditional evaluation results on BEAF dataset. We evaluate various VLMs on our dataset in terms of accuracy, precision, recall, and F1 score. Note that the BEAF dataset includes both original and manipulated data but does not consider the changes for this traditional evaluation. Underline stands for the best." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.407, + 0.784, + 0.512 + ], + "angle": 0, + "content": "
SizeModelAcc.(↑)Precision(↑)Recall(↑)F1(↑)Yes (%)
13BLLaVA-v1.574.956.393.470.352.6
InstructBLIP71.552.894.367.756.7
7BLLaVA-v1.579.762.689.573.745.4
InstructBLIP82.068.380.774.037.5
Shikra84.574.976.975.932.6
mPLUG-Owl269.150.793.865.858.7
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.547, + 0.787, + 0.638 + ], + "angle": 0, + "content": "is trained using a location-aware strategy, which could be beneficial for hallucination evaluation because knowledge of location aids in judging the existence of an object. The overall results indicate that, although it is rare to get all the answers wrong, it is common for the answers to remain the same even though the image has changed. This indicates that recent VLMs are not adequately reflective of such changes." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.642, + 0.788, + 0.838 + ], + "angle": 0, + "content": "Traditional evaluation on BEAF dataset. In addition to the proposed change-aware metrics, we also evaluate using traditional metrics such as accuracy, precision, recall, and F1 score, as shown in Table 3. While the trends in accuracy or F1 score are similar to those shown in TU results of Table 2, the absolute values of TU are much lower than those of the traditional metrics. This discrepancy suggests that, although a model achieves high performance in the existing benchmarks, our benchmark tells us that we cannot conclude low hallucination and the models may still suffer from hallucination. For example, although LLaVA-v1.5-13B achieves high performance on accuracy, the Yes ratio in Table 2 remains high even after the manipulation. It may include instances where the correct answer was guessed without actual reference to the given image. The experimental findings underscore the importance of observing changes in answers along the vision-axis to accurately evaluate hallucination in VLMs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.608, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEAF Benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.396, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.234, + 0.39, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.143, + 0.591, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.241, + 0.582, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.143, + 0.784, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.235, + 0.784, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 4: Visualization of image-wise object relationship. We visualize image-wise object relationship along the text and vision axes from Shikra. [Top] Original image samples. [Bottom] The object relation table given the manipulated images and questions. We color the correct answer blue and the wrong one red. The text axis stands for the target object queried in the question, and the vision axis for the removed object in the image. The \"none\" in the vision axis means the original image (not manipulated). Thereby, we can analyze influence between objects within a scene at once." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.487, + 0.612 + ], + "angle": 0, + "content": "Comparison with other benchmarks. We compare our metric TU with the performance for other benchmarks of VQAv2 [11] and GQA [15], as shown in Table 4. Each VQA performance value is reported from the proposed papers. The results indicate that, while mPLUG-Owl2 outperforms others on VQAv2 and LLaVA excels on GQA, Shikra leads in per" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.702 + ], + "angle": 0, + "content": "formance on our metric. We observe that there is a negative correlation between performance and the incidence of hallucination, which aligns with the previous observation [35]. Although improving performance on various benchmarks is challenging, it is essential to check multiple aspects of the performance, including hallucination. We believe that our benchmark provides valuable insights in conjunction with existing benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Image-wise object relationship. We visualize changes in dynamic object relationship along the vision and text axes, as shown in Fig. 4. Compared to our metrics providing comprehensive statistics as summarized scores, this visualization enables to track the changes in VLMs' responses at object-wise and scene-wise levels. The x-axis is the text axis where each object is associated with the question, \"Is there {object}?\"; the y-axis is the vision axis where the object is removed from the original image. The blue color is correct, the red one is wrong." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the first column in Fig. 4, altering the scene affects the textual correctness. Specifically, removing either the couch or the book shifts the book's textual cor-" + }, + { + "type": "table_caption", + "bbox": [ + 0.497, + 0.477, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Table 4: Comparison of performance on VQA and TU. We compare the accuracy on VQA datasets, VQAv2 and GQA, as well as TU on BEAF dataset." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.533, + 0.768, + 0.605 + ], + "angle": 0, + "content": "
Model (7B)VQAv2GQATU
LLaVA-v1.578.562.032.6
InstructBLIP-49.233.5
Shikra77.4-52.7
mPLUG-Owl279.456.124.6
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Ye-Bin et al." + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.149, + 0.734, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.293, + 0.789, + 0.364 + ], + "angle": 0, + "content": "Fig. 5: Object-wise error rate on the manipulated image-question pairs. We plot the error rate of each object to investigate which object is inaccurately inferred when it is removed. We exclude the objects manipulated less than 20 times. The solid line represents the average error rate of VLMs (7B), while the purple area indicates the \\(95\\%\\) confidence interval." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.789, + 0.512 + ], + "angle": 0, + "content": "rectness from correct to incorrect. This suggests a correlation between the couch and the book, indicating that complex scenes can lead to model hallucinations even in the absence of the book. In the second scene, when the object vase or book is removed, most of the incorrectness becomes correct. This could imply that the presence of a vase or a book might lead the model to misinterpret the scene. Through the simultaneous analysis of both vision and text axes, we observe immediate changes in the answers and offer interpretations for the dynamic interactions observed in the second and third scenes illustrated in the figure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.789, + 0.603 + ], + "angle": 0, + "content": "Object-wise error rate. We compute the mean and \\(95\\%\\) confidence interval of the object-wise error after the manipulation in Fig. 5. It represents whether the VLMs can accurately detect the removal. The results show that VLMs frequently fail to recognize certain objects like traffic lights, vases, and remotes. We speculate that the difficulty in recognizing these objects may stem from their complexity or their infrequent appearance in the training data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.622, + 0.407, + 0.637 + ], + "angle": 0, + "content": "4.2 Additional Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.518, + 0.84 + ], + "angle": 0, + "content": "Evaluation of open generation. In Table 5, we compare the results of the CHAIR [27] and TU metrics to evaluate the hallucination in the open-ended generation answers on our dataset. Briefly, \\(\\mathrm{CHAIR}_I\\) and \\(\\mathrm{CHAIR}_S\\) count the hallucinated instances and sentences containing these objects in the generated output, respectively. The lower the score, the better. We observe that if the open generation contains the hallucinated object and sentences, TU is also lower, as seen with InstructBLIP-13B." + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.659, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Table 5: Comparison of CHAIR metric and TU. We measure the CHAIR score based on the answers derived from the images in BEAF using the prompt \"Describe this image.\"" + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.731, + 0.784, + 0.815 + ], + "angle": 0, + "content": "
SizeModelCHAIR1CHAIRSTU
13BLLaVA-v1.521.672.324.3
InstructBLIP28.181.111.1
7BLLaVA-v1.522.373.932.6
InstructBLIP21.067.533.5
Shikra23.677.052.7
mPLUG-Owl225.178.424.6
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.608, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEAF Benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.788, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.325, + 0.785, + 0.382 + ], + "angle": 0, + "content": "Fig. 6: Samples with CLIP Score. We visualize samples and display the CLIP Score below. Additionally, we report the results of Faster R-CNN [26] using the magnifying glass icon. After removing the object, the removed object is not detected, which demonstrates the manipulation quality." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.413, + 0.522, + 0.534 + ], + "angle": 0, + "content": "CILP Score. We compare the CLIP Score [13] between original and manipulated images to ensure the target object has been well-removed. A lot of VLMs utilize the visual encoder from CLIP, prompting us to analyze CLIP's performance with our datasets. We measure the CLIP score between an object and a prompt \"a photo" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.785, + 0.61 + ], + "angle": 0, + "content": "of {object}. We use a cropped object image based on the GT bounding box. If the image has multiple objects of the same semantic class, we use the average CLIP score of all the same semantic objects. For the manipulated image, {object} is the removed object. As shown in Table 6, the manipulated images have lower CLIP scores compared to their original counterparts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.785, + 0.716 + ], + "angle": 0, + "content": "We also visualize the images with the CLIP Score and the multi-class classification results in Fig. 6. The multi-class classification correctly identifies the depicted objects. In addition, Fig. 7 visualizes an open generation sample result, where red-colored text indicates hallucinated words. Before manipulation, all the VLMs mention \"people\" in the background; however, they still mention \"people\" even though all the persons are removed from the image. This implies that the commonly used visual encoder might be a non-negligible source of hallucination." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.744, + 0.509, + 0.76 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We present a comprehensive analysis of hallucination using our BEAF benchmark. Unlike conventional hallucination benchmarks that measure accuracy by changing only the text-axis question, we also manipulate images by removing objects and propose four sophisticated metrics to track the changes in answers" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.429, + 0.785, + 0.47 + ], + "angle": 0, + "content": "Table 6: CLIP Score. We measure the CLIP Score of original and manipulated images with the text prompt." + }, + { + "type": "table", + "bbox": [ + 0.563, + 0.472, + 0.757, + 0.524 + ], + "angle": 0, + "content": "
TypeCLIP Score
Original53.7
Manipulated47.9
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Ye-Bin et al." + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.145, + 0.75, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.408, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Fig. 7: Qualitative example on open-ended generation. We report the changes in answers for the open-ended generation task with the prompt \"Describe this image.\" as the image undergoes manipulation. The red color highlights the hallucination parts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.789, + 0.64 + ], + "angle": 0, + "content": "according to image changes. This approach enables us to a more discerning analysis of the hallucination in two-axis views, e.g., vision and text axes. We can find with our metric that traditional accuracy may include instances where the correct answer was guessed without actual reference to the given image. We found that our TU is a more calibrated metric, which allows us to observe the intriguing phenomenon that has not been revealed in the prior work: VLMs tend to be confused if the complexity of an image is high. (refer to Sec. A in the supplementary material). Additionally, we introduce the visualization method to track the correctness as the objects are removed. We believe that our benchmark presents a comprehensive analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Limitation. While our benchmark offers a more fine-grained evaluation, it still faces the constraints of the MS-COCO dataset. These include a limited variety of object categories and an uneven distribution of object instances. While we utilized current state-of-the-art in-painting models and mask extractors, we manually adjusted the images to ensure image quality and dataset integrity. This step prevented us from achieving a fully automated creation process. Although such limitations exist, these aspects are expected to naturally diminish or disappear as in-painting models advance, making our dataset construction pipeline fully automatic and our method applicable to all images. Furthermore, we foresee future evaluations expanding in scale through manipulations like object removal, insertion, or resizing. While our current method may appear as a modest evaluation benchmark, we hope that our work becomes the first step toward controllable hallucination evaluation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "BEAF Benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.27 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub; No.2022-0-00124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities; No.RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH))" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.324, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.367 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv (2023) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.369, + 0.785, + 0.409 + ], + "angle": 0, + "content": "2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: IEEE International Conference on Computer Vision (ICCV) (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.785, + 0.451 + ], + "angle": 0, + "content": "3. Bisk, Y., Zellers, R., Gao, J., Choi, Y., et al.: Piqa: Reasoning about physical commonsense in natural language. In: AAAI Conference on Artificial Intelligence (AAAI) (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.785, + 0.492 + ], + "angle": 0, + "content": "4. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.521 + ], + "angle": 0, + "content": "5. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv (2023) 1, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.576 + ], + "angle": 0, + "content": "6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "7. Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., Schulman, J.: Training verifiers to solve math word problems. arXiv (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.673 + ], + "angle": 0, + "content": "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 1, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.785, + 0.702 + ], + "angle": 0, + "content": "9. Deng, A., Chen, Z., Hooi, B.: Seeing is believing: Mitigating hallucination in large vision-language models via clip-guided decoding. arXiv (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "10. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.743, + 0.785, + 0.799 + ], + "angle": 0, + "content": "11. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.839 + ], + "angle": 0, + "content": "12. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. In: AAAI Conference on Artificial Intelligence (AAAI) (2024) 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.326, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Ye-Bin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "13. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv (2021) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "14. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv (2023) 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "15. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "16. Hyun, L., Sung-Bin, K., Han, S., Yu, Y., Oh, T.H.: Smile: Multimodal dataset for understanding laughter in video with language models. arXiv (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "17. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.785, + 0.337 + ], + "angle": 0, + "content": "18. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.785, + 0.378 + ], + "angle": 0, + "content": "19. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. In: Conference on Empirical Methods in Natural Language Processing (EMNLP) (2023) 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV) (2014) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "21. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: International Conference on Learning Representations (ICLR) (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "22. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models. arXiv (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "23. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv (2023) 1, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.54 + ], + "angle": 0, + "content": "24. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.581 + ], + "angle": 0, + "content": "25. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. In: International Conference on Learning Representations (ICLR) (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.622 + ], + "angle": 0, + "content": "26. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: Advances in Neural Information Processing Systems (NeurIPS) (2015) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "27. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: Conference on Empirical Methods in Natural Language Processing (EMNLP) (2018) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.703 + ], + "angle": 0, + "content": "28. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "29. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "30. Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. In: IEEE Winter Conf. on Applications of Computer Vision (WACV) (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. 3(6), 7 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEAF Benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv (2023) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "33. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv (2023) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "34. Wang, J., Wang, Y., Xu, G., Zhang, J., Gu, Y., Jia, H., Yan, M., Zhang, J., Sang, J.: An llm-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv (2023) 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "35. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.355 + ], + "angle": 0, + "content": "36. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "37. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.438 + ], + "angle": 0, + "content": "38. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics (TACL) (2014) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.493 + ], + "angle": 0, + "content": "39. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.535 + ], + "angle": 0, + "content": "40. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. In: International Conference on Learning Representations (ICLR) (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.577 + ], + "angle": 0, + "content": "41. Zhao, Y., Li, Z., Jin, Z., Zhang, F., Zhao, H., Dou, C., Tao, Z., Xu, X., Liu, D.: Enhancing the spatial awareness capability of multi-modal large language model. arXiv (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.619 + ], + "angle": 0, + "content": "42. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models. In: International Conference on Learning Representations (ICLR) (2024) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.646 + ], + "angle": 0, + "content": "43. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.675 + ], + "angle": 0, + "content": "44. Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv (2023) 7" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_origin.pdf b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c92bc0f6604ab1c52a0be2630c6d8d7fd6eb6f4 --- /dev/null +++ b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/a88124a3-f7bb-4ef8-ac8f-7a55200b58dc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a69f14885b6d249ea997cee04ccf50d8acac5be2f7701f27f3d8a5c2837fcb45 +size 3782072 diff --git a/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/full.md b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..84511f4a6f23fa9cddaa79bc105a6206e46333e1 --- /dev/null +++ b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/full.md @@ -0,0 +1,276 @@ +# BEAF: Observing BEFORE-AFTER Changes to Evaluate Hallucination in Vision-language Models + +Moon Ye-Bin $^{1*}$ Nam Hyeon-Woo $^{1*}$ Wonseok Choi $^{2}$ Tae-Hyun Oh $^{1,2,3}$ + +$^{1}$ Dept. of EE, POSTECH, Korea $^{2}$ Grad. School of AI, POSTECH, Korea $^{3}$ Institute for Convergence Research and Education in Advanced Technology, Yonsei University, Korea. + +{ybmoon, hyeonw.nam, wonseok.c, taehyun}@postech.ac.kr + +Abstract. Vision language models (VLMs) perceive the world through a combination of a visual encoder and a large language model (LLM). The visual encoder, pre-trained on large-scale vision-text datasets, provides zero-shot generalization to visual data, and the LLM endows its high reasoning ability to VLMs. It leads VLMs to achieve high performance on wide benchmarks without fine-tuning, exhibiting zero or few-shot capability. However, recent studies show that VLMs are vulnerable to hallucination. This undesirable behavior degrades reliability and credibility, thereby making users unable to fully trust the output from VLMs. To enhance trustworthiness and better tackle the hallucination of VLMs, we curate a new evaluation dataset, called the BEfore-AFTER hallucination dataset (BEAF), and introduce new metrics: True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID). Unlike prior works that focus only on constructing questions and answers, the key idea of our benchmark is to manipulate visual scene information by image editing models and to design the metrics based on scene changes. This allows us to clearly assess whether VLMs correctly understand a given scene by observing the ability to perceive changes. We also visualize image-wise object relationship by virtue of our two-axis view: vision and text. Upon evaluating VLMs with our dataset, we observed that our metrics reveal different aspects of VLM hallucination that have not been reported before. Project page: https://beafbench.github.io/ + +# 1 Introduction + +Vision language models (VLMs) [1, 5, 8, 23, 24] have recently emerged by its multi-modal reasoning capability. It is indeed promising research toward making intelligent agents perceive the world. VLMs comprise a vision encoder as a visual perception module and a large language model (LLM) as a reasoning module. Recent LLMs [6, 32, 33] have exhibited powerful reasoning capabilities across various tasks, including commonsense reasoning [3, 16] and math [7], in zero-/few-shot ways. This success motivates researchers to integrate LLMs with vision encoders to open LLMs' eyes to see the world. Leveraging the reasoning + +capabilities of LLMs alongside visual perception enables VLMs to excel in various visual tasks, such as image captioning [2,38] and visual question answering [11, 15] without task-specific fine-tuning. + +Despite this encouraging demonstration, VLMs are susceptible to the hallucination [12,14,19,34,39,42] wherein their outputs do not reflect facts presented in input images. For example, although a given input image does not present an object, VLMs often incorrectly answer for the presence of the object. This tendency is problematic, akin to spreading falsehoods in human communication. Hallucination prevents genuine communication, consequently diminishing the reliability of interactions between agents or humans. No matter how good VLM's answer is, if hallucination is included in the answer, it can undermine trust in the system's reliability and credibility. Therefore, it is crucial to address hallucinations to improve the system's responsibility and trustworthiness. + +Understanding the impact of hallucinations on model performance requires the development of an assessment framework, which requires a deep exploration of the underlying causes. The prior works [14, 41, 42] have been proposed to unveil the causation of hallucinations in terms of data and models by reporting their performance on existing hallucination benchmarks. As a widely used hallucination benchmark, POPE [19] follows the question-and-answer (QnA) format. For the convenience of measuring, POPE uses a discriminative question, e.g., "Is there {object} in the image?", and gets a YES or NO answer. The following hallucination benchmarks [14, 19, 34] commonly leverage LLMs or human annotators to generate question prompts with available scene information such as bounding boxes and captions. This text-axis evaluation may assess certain aspects of the hallucination performance well, but it is hard to disentangle sources of hallucination, particularly in VLMs that handle both visual and text modalities. Thus, to thoroughly analyze and understand the behavior of VLMs, it is both natural and necessary to consider both modalities. + +In this work, we propose a BEFORE-AFTER (BEAF) hallucination evaluation benchmark. This benchmark comprises a dataset enhanced along the vision-axis and metrics that are aware of changes. We manipulate scenes by selectively removing objects on the vision-axis while augmenting QnA pairs along the text-axis. We can observe how the models' answers change as the image is manipulated for the same question. For example, as illustrated in Fig. 1, if we remove an apple from the given image and then ask "Is there an apple?", the answer from the model should change from "Yes" to "No", between the original and modified images, if it truly relies on what it sees. Therefore, for a more thorough evaluation and analysis, we need change-aware metrics beyond just using standard accuracy [14, 19, 34]. + +We propose four new metrics: True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID), which are enabled by our vision and text augmentation. TU measures whether the model answers correctly before and after the manipulation. IG evaluates whether the models perceive the images consistently both before and after manipulation. SB is related to the phenomenon of giving the same answer instead of correctly recognizing the image. ID measures + +![](images/ec68970f63d0e326873895ee4d8be40f9eed1d21bad8bd22363188c352a6c607.jpg) + +
MTQuestionMetricsCorrectnessAnswer
Text-Acc○-
Img. & TextRemoved ObjectTU○ → ○Y → N
IG× → ×N → Y
SBp○ → ×Y → Y
SBn× → ○N → N
Unremoved ObjectID○ → × +× → ○Y → N +N → Y
+ +MT: Manipulation Target + +![](images/44cd7026e1a3dc75993230db3fa4f4fb633fa90398a910778da9f45b8e53c474.jpg) +BEAF Evaluation (Ours) +Fig. 1: Illustration of BEbefore-AFTER (BEAF) benchmark. We present a comparison between traditional evaluation benchmarks and our BEAF benchmark for assessing the hallucination behavior in VLMs. Traditional evaluation methods solely manipulate questions based on the existence of an object and measure accuracy or F1 score. In contrast, our BEAF benchmark not only constructs questions but also manipulates images and tracks changes in answers as the images undergo manipulation. The BEAF benchmark introduces novel metrics including True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID), which consider the changes in answers. + +cases where the answer changes even though the target object does not change or does not exist in the image. The first three metrics address questions regarding target objects that have been removed, while the last one belongs to questions about other objects not related to the targets. + +Based on our BEAF dataset and change-aware metrics, we can evaluate and understand the models, even the parts we did not know through the previous methods. The evaluation results imply that answers assumed as non-hallucination from the existing hallucination benchmarks could be a hallucination. We also analyze the influence of each object in the scene based on the variation along the vision and text axes. In addition, we include a discussion on open-generation answers and CLIPScore results. Our key contributions are as follows: + +- We propose a benchmark, called BEFORE-AFTER (BEAF), to evaluate the hallucination of large vision-language models. Our benchmark comprises a dataset that considers both text- and vision-axes, and change-aware metrics, which allow us to perform granular evaluation. +- We demonstrate that previously reported outcomes deemed non-hallucinatory in the prior text-axis-only evaluation benchmarks may be hallucinations. +- We also visualize the impact of relationships between objects within hallucinatory images based on the BEAF results. + +# 2 Related Work + +Vision Language Models (VLMs). VLMs [1, 4, 5, 8, 10, 23-25, 40, 43] have demonstrated their impressive image describing and reasoning ability based on powerful large language models (LLMs) [6, 31-33]. Generally, VLMs consist of three parts: vision encoder, LLM, and vision-language alignment module. The vision information extracted from the vision encoder is translated to the LLM-understandable token by passing through the alignment module. LLM takes a prompt text and visual tokens as input and outputs response text tokens according to a given image and prompt. Their training process is divided into multiple steps depending on which parts of VLMs are trained. + +Hallucination in VLMs. Hallucination [12, 19, 22, 34, 35, 39] in VLMs refers to the case where the generated text answer does not reflect the true contents of the provided images but rather relies on the internal knowledge of the models. This behavior might cause harmful consequences in high-risk applications such as the medical domain or auto-driving scenarios. This undesirable behavior could be caused by various factors, e.g., data imbalance, image resolution, and model capacity. The assessment of hallucination is also related to the true understanding of VLMs. If VLMs respond to the given input correctly rather than relying on innate knowledge, VLMs are considered to have a perfect understanding of the perceived world. Numerous methodologies [12, 21, 42] have been developed to tackle the issue of hallucination in VLMs by addressing data biases [14, 21], optimization techniques [12, 18, 29], and post-processing methodologies [9, 37, 42]. + +Evaluating Hallucination in VLMs. Evaluating hallucination performance is essential to guide the research of VLM. Previous studies have introduced hallucination benchmarks to understand and assess these undesirable phenomena of VLMs. Motivated by traditional visual and question benchmarks, existing hallucination benchmarks [14, 19, 34] follow this VQA style that provides a simple QnA evaluation pipeline and intuitive metrics. POPE [19] and CIEM [14] curate images and QnA pairs with metrics such as accuracy, precision, and recall, where they use a specific question type inducing multi-choice or yes/no answers, called discriminative type. AMBER [34] includes an additional question type that induces open-ended caption generation as answers, called generative type. For such open-generation assessment, they extract the objects from the output text answers and compare them with the ground-truth objects in the image. These prior studies focus on manipulating texts of questions and answers only, + +Table 1: Data statistics. Our BEAF dataset contains 26K image-question pairs, consisting of the original and manipulated ones. On average, an original image is associated with 3.45 manipulated images and 11.72 questions. + +
ImageImg-Q PairYesNo
Original5005,6532,0263,627
Manipulated1,72720,4656,26714,198
Total2,22726,1188,29317,825
+ +# Ratio of upper-class categories + +![](images/e24d158a9e1388e6e8825eb08336756f3f01b27fd5cd513cf3da0bd940306a87.jpg) +Removed objects (total: 1,727) + +![](images/a1fdf2aa9cd32bd516a63a06377f0481df96a98192bd056694af6b05a9c3f628.jpg) +Objects in question (total: 26,118) +Fig. 2: Ratio of removed objects from image and object in question. For convenience, we report the ratio of upper-class categories instead of the object class itself. The total number of removed objects is the same as the number of manipulated images, and one object in question is the same as the number of image-question pairs. + +which we call text-axis evaluation. However, if some objects always co-occur with another object, we cannot judge whether VLMs truly understand scene information accurately. Also, VLMs take multi-modal inputs, both images and text prompts; thus, evaluation considering only the text-axis would be insufficient as a hallucination assessment of VLMs. In this work, we manipulate the scene information and questions together, i.e., vision-/text-axes; it enables the fine-grained analysis of hallucination. We call it vision-/text-axes evaluation. + +# 3 BEAF Evaluation Benchmark for Hallucination + +We explore another axis, i.e., vision-axis, to make up for the deficiencies in the text-axis evaluation. Our key idea is to manipulate a visual scene by removing and inpainting objects within an image. This allows us to focus on changes and to check whether the VLMs are correctly aware of it. We describe how to curate our dataset in Sec. 3.1, and then elaborate our new metrics in Sec. 3.2 + +![](images/52c8686b18ccccc1418f502deb7ae6977b05d55c26ff0796a3a3c7847d506d72.jpg) +Fig. 3: Image manipulation pipeline. We illustrate the image manipulation pipeline comprised of three stages. In Stage 1, we automatically remove target objects sharing the same semantic class from the given images. Stage 2 is to filter the automatically manipulated results based on the predefined rules, such as mask errors, remained shadows, and low-quality outcomes. Undesirable manipulations are either corrected or discarded during this stage. Finally, in Stage 3, human annotators engage in human-guided manipulation of the filtered images to achieve high-quality results. + +# 3.1 Evaluation Dataset + +Dataset Overview. Table 1 and Fig. 2 summarize our data statistics. The dataset consists of images and their associated QnA. The image type is divided into original and manipulated. Five hundred original images are sampled from the validation set of the MS-COCO dataset [20], which shares the same split with POPE [19], and their associated 1,727 manipulated images are generated by removing objects from the original images. On average, 3.45 manipulated images are created from the original images. The main question format is "Is there {object} in this image?" and {object} can be a positive object within the image or a negative object not in the image; the positive object is one of the objects in the image, and the negative object is the one not in the image. We adopt the positive and negative objects from the POPE for the original image and adapt the question for the manipulated image. If we remove the positive object from the original image to make the manipulated one, the positive object becomes the negative object in the manipulated one. The total dataset comprised 2,227 images and 26,118 image-question pairs. The "Yes" and "No" ground truth answers are 8,293 (31.75%) and 17,825 (68.25%), respectively. + +Image Dataset Construction Pipeline. We manipulate the original images by removing objects within the image. To collect the manipulated images, we pass through a three-stage pipeline (see Fig. 3): the first stage involves auto + +matic manipulation, followed by a filtering stage, and finally, a human-guided manipulation stage. We offer a detailed explanation of each stage as follows. + +1) First Stage: Automatic Manipulation. For automatic object removal, we use an off-the-shelf inpainting model, LaMa [30], with mask guidance. While the initial option is to use the ground truth masks provided within the original dataset, we opt not to use them because of the low quality ground truth masks in MS-COCO. For example, polygon masks often cannot cover all areas of the corresponding object and can be propagated to the inpainting quality. Therefore, we employ a pre-trained mask extractor called SAM [17], utilizing bounding box annotations. Subsequently, we dilate the extracted mask to have a margin for more effective object removal, and remove the object with the inpainting model given the dilated mask guidance. We empirically choose to use LaMa over diffusion models, e.g., [28, 44], as we find that this specialized model performs better in the context of object removal. + +2) Second Stage: Filtering. Although we remove objects from the original images, achieving consistent quality in the outcome can be influenced by a myriad of factors. These factors include, but are not limited to, the size of the objects being removed and the complexity of the backgrounds they inhabit. Additionally, the presence of shadows remains unaddressed, as they are not effectively masked using SAM; the ground truth masks also do not capture the shadow. As shown on the right side of Fig. 3, it displays undesirable images due to the presence of shadow or a faucet. The human annotators classified the results obtained from the previous stage into usable or unusable, where only the usable results are further manipulated. The detailed filtering rules can be found in the supplementary material. We directly use the usable results and hand over the results that need further manipulation to the next stage. + +3) Third Stage: Human-guided Manipulation. The third stage of our workflow relies on human annotators to intervene and refine the images obtained from the second stage. As mentioned above, these filtered images often exhibit undesirable artifacts such as ghost shadows, afterimages, and fragmented objects. Even such a small artifact may provide strong cues indicating the past presence of objects that both humans and machines can deduce, e.g., a cat's tail suggests the past presence of the cat before the object removal. Thus, human annotators are asked to meticulously manipulate the images, either correcting or enhancing them. Annotators effectively refine visual data, aligning it more closely with real natural images in human perception. + +As shown in Fig. 3, our annotation process results in images that are free of the aforementioned artifacts. This manual intervention is crucial, as it ensures the fidelity and reliability of our benchmark dataset. + +# 3.2 Evaluation Metrics + +The prior studies [14, 19, 34] suggest using traditional metrics such as accuracy, precision, recall, and F1 score, which we can also use the metrics in our dataset. In addition, we propose four new metrics: True Understanding (TU), IGnorance + +(IG), StuBbornness (SB), and InDecision (ID), for more detailed evaluation by exploiting the distinctive configuration of before-/after-changes in our dataset. + +Let $\mathcal{S} = \{(I_o, I_m, Q, A_o, A_m, R)_i\}_{i=1}^{|S|}$ denote our dataset, where each tuple consists of an original image $I_o$ , a version of its manipulated image $I_m$ , their associated question $Q$ , their respective corresponding answers $A_o$ and $A_m$ , and a True/False flag $R$ that indicates whether the question is associated with the object being removed in the manipulated image $I_m$ . For compactness of our metric definitions, we define $\text{Filter}(\cdot)$ to extract the subset of $S$ satisfying input conditions as: + +$$ +\operatorname {F i l t e r} \left(b _ {o}, b _ {m}, b _ {r}\right) = \{t | \text {I s C o r r e c t} \left(A _ {o}\right) = b _ {o}, \text {I s C o r r e c t} \left(A _ {m}\right) = b _ {m}, R = b _ {r}, t \in \mathcal {S} \} +$$ + +where the tuple $t = (I_o, I_m, Q, A_o, A_m, R)$ . For example, Filter(True, False, True) is the subset of $S$ of which the element satisfy the conditions that are 1) the answers corresponding to respective original and manipulated images are correct and wrong, and 2) the question is the one related to the object removed in the manipulated images. + +True Understanding. True Understanding (TU) measures whether the models truly understand the scene. It evaluates if the model can accurately answer to the question about the removed object. As shown in Fig. 1, if the apple is removed, a model that truly understands the change should switch its answer from "Yes" to "No" to the question "Is there an apple?" We define the TU metric as follows: + +$$ +\mathrm {T U} = \frac {\left| \text {F i l t e r} (\text {T r u e} , \text {T r u e} , \text {T r u e}) \right|}{\left| \text {F i l t e r} (\text {T r u e} \vee \text {F a l s e} , \text {T r u e} \vee \text {F a l s e} , \text {T r u e}) \right|} \cdot 1 0 0. \tag {1} +$$ + +Compared to the traditional accuracy, our metric specifically considers the correctness of answers before and after manipulation. Therefore, it provides a conditional measure of accuracy, taking into account the model's response to changes within the scene. A higher value of this metric indicates a more accurate understanding of the scene by the model, as it demonstrates the model's ability to adapt its answers appropriately to the scene changes. + +Ignorance. The second metric is IGnorance (IG), which measures the extent to which models lack knowledge about specific scene information. This metric quantifies ignorance by identifying instances where models provide incorrect answers to the question about the removed objects. As shown in Fig. 1, ignorance is evident if models incorrectly answer "No" to the presence of an apple before its removal and incorrectly answer "Yes" after the apple is removed, indicating a failure to recognize it. IG is defined as follows: + +$$ +\mathrm {I G} = \frac {\left| \text {F i l t e r (F a l s e , F a l s e , T r u e)} \right|}{\left| \text {F i l t e r (T r u e V F a l s e , T r u e V F a l s e , T r u e)} \right|} \cdot 1 0 0. \tag {2} +$$ + +The lower the metric, the less ignorant the model is. + +Stubbornness. What we observe is that VLMs tend to output the same answer repeatedly. It means that VLMs are biased toward the specific answer rather than arbitrary one when they do not know. This repetitive behavior is undesirable; thus, we introduce the metric of StuBbornness (SB), which measures the extent + +to which models adhere to their initial answers. Furthermore, we categorize SB into $\mathrm{SB}_p$ and $\mathrm{SB}_n$ where the subscripts of $p$ and $n$ correspond to the consistent positive ("Yes") and negative ("No") answers, respectively. We define them as follows: + +$$ +\mathrm {S B} _ {p} = \frac {1 0 0 \cdot | \text {F i l t e r (T r u e , F a l s e , T r u e)} |}{| \text {F i l t e r (T r u e \vee F a l s e , T r u e \vee F a l s e , T r u e)} |}, \quad \mathrm {S B} _ {n} = \frac {1 0 0 \cdot | \text {F i l t e r (F a l s e , T r u e , T r u e)} |}{| \text {F i l t e r (T r u e \vee F a l s e , T r u e \vee F a l s e , T r u e)} |}. \tag {3} +$$ + +SB is the summation of $\mathrm{SB}_p$ and $\mathrm{SB}_n$ and can be computed from $100 - \mathrm{TU} - \mathrm{IG}$ . The lower the metric, the less ignorant the model is. + +Indecision. Unlike the previous three metrics that focus on the removed objects in the manipulated images, InDecision (ID) focuses on the answers to the questions that are not relevant to the removed objects. These answers should not be changed even after the manipulation. Nevertheless, models, often change their answers about unmanipulated objects, which would imply that the model predicts a random answer without grounding it on the image. We track these behaviors with the ID metric defined as follows: + +$$ +\mathrm {I D} = \frac {\left| \text {F i l t e r} (\text {T r u e} , \text {F a l s e} , \text {F a l s e}) \right| + \left| \text {F i l t e r} (\text {F a l s e} , \text {T r u e} , \text {F a l s e}) \right|}{\left| \text {F i l t e r} (\text {T r u e} \vee \text {F a l s e} , \text {T r u e} \vee \text {F a l s e} , \text {F a l s e}) \right|} \cdot 1 0 0. \tag {4} +$$ + +The lower the metric, the less ignorant the model is. Additionally, we combine TU and ID using a harmonic mean to make it easier to compare on our benchmark. We refer to it as F1-Score, which is computed as follows: $F1 = \frac{2}{TU^{-1} + (100 - ID)^{-1}}$ . + +# 4 Experiment + +In Sec. 4.1, we evaluate the current VLMs using our BEAF dataset. We employ both proposed and traditional metrics and compare their performance with other benchmarks. Furthermore, we conduct an analysis of image-wise dynamic correctness by visualizing image-wise object relationship containing vision and text axes. We provide a report on the object-wise error rate. In Sec. 4.2, we assess the performance of VLMs on the open generation task using our dataset, while also analyzing both the dataset and VLMs based on the CLIP score. + +All the reported performance is the zero-shot performance of VLMs, including LLaVA [23], InstructBLIP [8], Shikra [5], and mPLUG-Owl [36]. We use prompt templates provided in their released codebases or papers. + +# 4.1 Main Results + +Evaluation on BEAF benchmark. We evaluate several VLMs on our BEAF dataset using the proposed change-aware metrics, TU, IG, SB, ID, and F1. The evaluation results are presented in Table 2. Among the 13B-size models, LLaVA outperforms InstructBLIP in terms of TU, IG, and SB because InstructBLIP consistently answers "Yes" regardless of scene manipulation, and this result is related to its high value of $\mathrm{SB}_p$ . InstructBLIP achieves better ID compared to LLaVA. Among 7B-size models, Shikra exhibits the best in TU and $\mathrm{SB}_p$ . Shikra + +Table 2: Evaluation results on BEAF benchmark. We evaluate various VLMs on our benchmark. F1 is a harmonic mean of TU and 1-ID. Yes ratio is the percentage of total answers that are answered "Yes" for the removed object-related question. The former is for the original image, and the latter stands for the manipulated image. Underline stands for the best. + +
SizeModelTU(↑)IG(↓)SBp(↓)SBn(↓)ID(↓)F1(↑)Yes (%)
13BLLaVA-v1.524.30.272.03.56.438.696.4 → 72.2
InstructBLIP11.10.384.44.26.119.995.5 → 84.7
7BLLaVA-v1.532.60.161.36.05.648.593.9 → 61.4
InstructBLIP33.50.651.314.65.249.584.8 → 51.9
Shikra52.70.431.215.75.767.683.9 → 31.6
mPLUG-Owl224.60.072.13.37.038.896.7 → 72.1
+ +Table 3: Traditional evaluation results on BEAF dataset. We evaluate various VLMs on our dataset in terms of accuracy, precision, recall, and F1 score. Note that the BEAF dataset includes both original and manipulated data but does not consider the changes for this traditional evaluation. Underline stands for the best. + +
SizeModelAcc.(↑)Precision(↑)Recall(↑)F1(↑)Yes (%)
13BLLaVA-v1.574.956.393.470.352.6
InstructBLIP71.552.894.367.756.7
7BLLaVA-v1.579.762.689.573.745.4
InstructBLIP82.068.380.774.037.5
Shikra84.574.976.975.932.6
mPLUG-Owl269.150.793.865.858.7
+ +is trained using a location-aware strategy, which could be beneficial for hallucination evaluation because knowledge of location aids in judging the existence of an object. The overall results indicate that, although it is rare to get all the answers wrong, it is common for the answers to remain the same even though the image has changed. This indicates that recent VLMs are not adequately reflective of such changes. + +Traditional evaluation on BEAF dataset. In addition to the proposed change-aware metrics, we also evaluate using traditional metrics such as accuracy, precision, recall, and F1 score, as shown in Table 3. While the trends in accuracy or F1 score are similar to those shown in TU results of Table 2, the absolute values of TU are much lower than those of the traditional metrics. This discrepancy suggests that, although a model achieves high performance in the existing benchmarks, our benchmark tells us that we cannot conclude low hallucination and the models may still suffer from hallucination. For example, although LLaVA-v1.5-13B achieves high performance on accuracy, the Yes ratio in Table 2 remains high even after the manipulation. It may include instances where the correct answer was guessed without actual reference to the given image. The experimental findings underscore the importance of observing changes in answers along the vision-axis to accurately evaluate hallucination in VLMs. + +![](images/b51017f2148922e95dfd337608e7b4b084a057d71780364f1045ca279e89e131.jpg) + +![](images/f12711c2b3f2e8474e8f0e351ac8d7378ffc8cc32c2422eb301276e374107087.jpg) +Fig. 4: Visualization of image-wise object relationship. We visualize image-wise object relationship along the text and vision axes from Shikra. [Top] Original image samples. [Bottom] The object relation table given the manipulated images and questions. We color the correct answer blue and the wrong one red. The text axis stands for the target object queried in the question, and the vision axis for the removed object in the image. The "none" in the vision axis means the original image (not manipulated). Thereby, we can analyze influence between objects within a scene at once. + +![](images/b73916b14e5e81779ff87ab4407c068f9a741ba32636110748280b05d0049b06.jpg) + +![](images/3f1f3b70a1cf372774eaecff39359a2f173e26b2917dba53ed6db7db28fd64b7.jpg) + +![](images/e3bb7ece665e16e8f5678700b7d6cba2f09b196ee3b9dd7aab732df981b17a4d.jpg) + +![](images/86ff2b31abb6e666dba9e9d57516fd95e90dcd065577e66d302d253a7b0b419d.jpg) + +Comparison with other benchmarks. We compare our metric TU with the performance for other benchmarks of VQAv2 [11] and GQA [15], as shown in Table 4. Each VQA performance value is reported from the proposed papers. The results indicate that, while mPLUG-Owl2 outperforms others on VQAv2 and LLaVA excels on GQA, Shikra leads in per + +formance on our metric. We observe that there is a negative correlation between performance and the incidence of hallucination, which aligns with the previous observation [35]. Although improving performance on various benchmarks is challenging, it is essential to check multiple aspects of the performance, including hallucination. We believe that our benchmark provides valuable insights in conjunction with existing benchmarks. + +Image-wise object relationship. We visualize changes in dynamic object relationship along the vision and text axes, as shown in Fig. 4. Compared to our metrics providing comprehensive statistics as summarized scores, this visualization enables to track the changes in VLMs' responses at object-wise and scene-wise levels. The x-axis is the text axis where each object is associated with the question, "Is there {object}?"; the y-axis is the vision axis where the object is removed from the original image. The blue color is correct, the red one is wrong. + +In the first column in Fig. 4, altering the scene affects the textual correctness. Specifically, removing either the couch or the book shifts the book's textual cor- + +Table 4: Comparison of performance on VQA and TU. We compare the accuracy on VQA datasets, VQAv2 and GQA, as well as TU on BEAF dataset. + +
Model (7B)VQAv2GQATU
LLaVA-v1.578.562.032.6
InstructBLIP-49.233.5
Shikra77.4-52.7
mPLUG-Owl279.456.124.6
+ +![](images/c6b8a87ffc47806c59d71c613b7a4c02da54362f5a20f63f3880eceba953a9d6.jpg) +Fig. 5: Object-wise error rate on the manipulated image-question pairs. We plot the error rate of each object to investigate which object is inaccurately inferred when it is removed. We exclude the objects manipulated less than 20 times. The solid line represents the average error rate of VLMs (7B), while the purple area indicates the $95\%$ confidence interval. + +rectness from correct to incorrect. This suggests a correlation between the couch and the book, indicating that complex scenes can lead to model hallucinations even in the absence of the book. In the second scene, when the object vase or book is removed, most of the incorrectness becomes correct. This could imply that the presence of a vase or a book might lead the model to misinterpret the scene. Through the simultaneous analysis of both vision and text axes, we observe immediate changes in the answers and offer interpretations for the dynamic interactions observed in the second and third scenes illustrated in the figure. + +Object-wise error rate. We compute the mean and $95\%$ confidence interval of the object-wise error after the manipulation in Fig. 5. It represents whether the VLMs can accurately detect the removal. The results show that VLMs frequently fail to recognize certain objects like traffic lights, vases, and remotes. We speculate that the difficulty in recognizing these objects may stem from their complexity or their infrequent appearance in the training data. + +# 4.2 Additional Study + +Evaluation of open generation. In Table 5, we compare the results of the CHAIR [27] and TU metrics to evaluate the hallucination in the open-ended generation answers on our dataset. Briefly, $\mathrm{CHAIR}_I$ and $\mathrm{CHAIR}_S$ count the hallucinated instances and sentences containing these objects in the generated output, respectively. The lower the score, the better. We observe that if the open generation contains the hallucinated object and sentences, TU is also lower, as seen with InstructBLIP-13B. + +Table 5: Comparison of CHAIR metric and TU. We measure the CHAIR score based on the answers derived from the images in BEAF using the prompt "Describe this image." + +
SizeModelCHAIR1CHAIRSTU
13BLLaVA-v1.521.672.324.3
InstructBLIP28.181.111.1
7BLLaVA-v1.522.373.932.6
InstructBLIP21.067.533.5
Shikra23.677.052.7
mPLUG-Owl225.178.424.6
+ +![](images/d22c0ec6b2135da95057ec1495e3184f36228cc8eeee92da5e0b8be0e81f4d8b.jpg) +Fig. 6: Samples with CLIP Score. We visualize samples and display the CLIP Score below. Additionally, we report the results of Faster R-CNN [26] using the magnifying glass icon. After removing the object, the removed object is not detected, which demonstrates the manipulation quality. + +CILP Score. We compare the CLIP Score [13] between original and manipulated images to ensure the target object has been well-removed. A lot of VLMs utilize the visual encoder from CLIP, prompting us to analyze CLIP's performance with our datasets. We measure the CLIP score between an object and a prompt "a photo + +of {object}. We use a cropped object image based on the GT bounding box. If the image has multiple objects of the same semantic class, we use the average CLIP score of all the same semantic objects. For the manipulated image, {object} is the removed object. As shown in Table 6, the manipulated images have lower CLIP scores compared to their original counterparts. + +We also visualize the images with the CLIP Score and the multi-class classification results in Fig. 6. The multi-class classification correctly identifies the depicted objects. In addition, Fig. 7 visualizes an open generation sample result, where red-colored text indicates hallucinated words. Before manipulation, all the VLMs mention "people" in the background; however, they still mention "people" even though all the persons are removed from the image. This implies that the commonly used visual encoder might be a non-negligible source of hallucination. + +# 5 Discussion and Conclusion + +We present a comprehensive analysis of hallucination using our BEAF benchmark. Unlike conventional hallucination benchmarks that measure accuracy by changing only the text-axis question, we also manipulate images by removing objects and propose four sophisticated metrics to track the changes in answers + +Table 6: CLIP Score. We measure the CLIP Score of original and manipulated images with the text prompt. + +
TypeCLIP Score
Original53.7
Manipulated47.9
+ +![](images/f6d9f70d17cefcae75666afbb809b71be2e291da3b7df22fffa144d108365209.jpg) +Fig. 7: Qualitative example on open-ended generation. We report the changes in answers for the open-ended generation task with the prompt "Describe this image." as the image undergoes manipulation. The red color highlights the hallucination parts. + +according to image changes. This approach enables us to a more discerning analysis of the hallucination in two-axis views, e.g., vision and text axes. We can find with our metric that traditional accuracy may include instances where the correct answer was guessed without actual reference to the given image. We found that our TU is a more calibrated metric, which allows us to observe the intriguing phenomenon that has not been revealed in the prior work: VLMs tend to be confused if the complexity of an image is high. (refer to Sec. A in the supplementary material). Additionally, we introduce the visualization method to track the correctness as the objects are removed. We believe that our benchmark presents a comprehensive analysis. + +Limitation. While our benchmark offers a more fine-grained evaluation, it still faces the constraints of the MS-COCO dataset. These include a limited variety of object categories and an uneven distribution of object instances. While we utilized current state-of-the-art in-painting models and mask extractors, we manually adjusted the images to ensure image quality and dataset integrity. This step prevented us from achieving a fully automated creation process. Although such limitations exist, these aspects are expected to naturally diminish or disappear as in-painting models advance, making our dataset construction pipeline fully automatic and our method applicable to all images. Furthermore, we foresee future evaluations expanding in scale through manipulations like object removal, insertion, or resizing. While our current method may appear as a modest evaluation benchmark, we hope that our work becomes the first step toward controllable hallucination evaluation. + +# Acknowledgment + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub; No.2022-0-00124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities; No.RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH)) + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv (2023) 1, 4 +2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: IEEE International Conference on Computer Vision (ICCV) (2019) 2 +3. Bisk, Y., Zellers, R., Gao, J., Choi, Y., et al.: Piqa: Reasoning about physical commonsense in natural language. In: AAAI Conference on Artificial Intelligence (AAAI) (2020) 1 +4. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv (2023) 4 +5. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv (2023) 1, 4, 9 +6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/1, 4 +7. Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., Schulman, J.: Training verifiers to solve math word problems. arXiv (2021) 1 +8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 1, 4, 9 +9. Deng, A., Chen, Z., Hooi, B.: Seeing is believing: Mitigating hallucination in large vision-language models via clip-guided decoding. arXiv (2024) 4 +10. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv (2023) 4 +11. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 2, 11 +12. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. In: AAAI Conference on Artificial Intelligence (AAAI) (2024) 2, 4 + +13. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv (2021) 13 +14. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv (2023) 2, 4, 7 +15. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 2, 11 +16. Hyun, L., Sung-Bin, K., Han, S., Yu, Y., Oh, T.H.: Smile: Multimodal dataset for understanding laughter in video with language models. arXiv (2023) 1 +17. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv (2023) 7 +18. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv (2023) 4 +19. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. In: Conference on Empirical Methods in Natural Language Processing (EMNLP) (2023) 2, 4, 6, 7 +20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV) (2014) 6 +21. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: International Conference on Learning Representations (ICLR) (2024) 4 +22. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models. arXiv (2024) 4 +23. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv (2023) 1, 4, 9 +24. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 1, 4 +25. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. In: International Conference on Learning Representations (ICLR) (2024) 4 +26. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: Advances in Neural Information Processing Systems (NeurIPS) (2015) 13 +27. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: Conference on Empirical Methods in Natural Language Processing (EMNLP) (2018) 12 +28. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 7 +29. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv (2023) 4 +30. Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. In: IEEE Winter Conf. on Applications of Computer Vision (WACV) (2022) 7 +31. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. 3(6), 7 (2023) 4 + +32. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv (2023) 1, 4 +33. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv (2023) 1, 4 +34. Wang, J., Wang, Y., Xu, G., Zhang, J., Gu, Y., Jia, H., Yan, M., Zhang, J., Sang, J.: An llm-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv (2023) 2, 4, 7 +35. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023) 4, 11 +36. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv (2023) 9 +37. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv (2023) 4 +38. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics (TACL) (2014) 2 +39. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023) 2, 4 +40. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. In: International Conference on Learning Representations (ICLR) (2024) 4 +41. Zhao, Y., Li, Z., Jin, Z., Zhang, F., Zhao, H., Dou, C., Tao, Z., Xu, X., Liu, D.: Enhancing the spatial awareness capability of multi-modal large language model. arXiv (2023) 2 +42. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models. In: International Conference on Learning Representations (ICLR) (2024) 2, 4 +43. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv (2023) 4 +44. Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv (2023) 7 \ No newline at end of file diff --git a/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/images.zip b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..36b9138a183841dc04083026c15c7eac5cdf3dd1 --- /dev/null +++ b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c70ff44809aa2c57b5240a5078653f0b1c9c6efc3c20b59d9bf2d71a3ad523c6 +size 603249 diff --git a/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/layout.json b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..90305850e20c49d33b0fc1cd395b1f8efd527175 --- /dev/null +++ b/2024/BEAF_ Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models/layout.json @@ -0,0 +1,7509 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 112, + 482, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 112, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 482, + 148 + ], + "type": "text", + "content": "BEAF: Observing BEFORE-AFTER Changes to Evaluate Hallucination in Vision-language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "spans": [ + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": "Moon Ye-Bin" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": " Nam Hyeon-Woo" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": " Wonseok Choi" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": " Tae-Hyun Oh" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 189, + 478, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 189, + 478, + 223 + ], + "spans": [ + { + "bbox": [ + 135, + 189, + 478, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 189, + 478, + 223 + ], + "type": "text", + "content": "Dept. of EE, POSTECH, Korea " + }, + { + "bbox": [ + 135, + 189, + 478, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 189, + 478, + 223 + ], + "type": "text", + "content": "Grad. School of AI, POSTECH, Korea " + }, + { + "bbox": [ + 135, + 189, + 478, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 189, + 478, + 223 + ], + "type": "text", + "content": "Institute for Convergence Research and Education in Advanced Technology, Yonsei University, Korea." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 178, + 224, + 434, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 224, + 434, + 234 + ], + "spans": [ + { + "bbox": [ + 178, + 224, + 434, + 234 + ], + "type": "text", + "content": "{ybmoon, hyeonw.nam, wonseok.c, taehyun}@postech.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 264, + 455, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 264, + 455, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 264, + 455, + 506 + ], + "type": "text", + "content": "Abstract. Vision language models (VLMs) perceive the world through a combination of a visual encoder and a large language model (LLM). The visual encoder, pre-trained on large-scale vision-text datasets, provides zero-shot generalization to visual data, and the LLM endows its high reasoning ability to VLMs. It leads VLMs to achieve high performance on wide benchmarks without fine-tuning, exhibiting zero or few-shot capability. However, recent studies show that VLMs are vulnerable to hallucination. This undesirable behavior degrades reliability and credibility, thereby making users unable to fully trust the output from VLMs. To enhance trustworthiness and better tackle the hallucination of VLMs, we curate a new evaluation dataset, called the BEfore-AFTER hallucination dataset (BEAF), and introduce new metrics: True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID). Unlike prior works that focus only on constructing questions and answers, the key idea of our benchmark is to manipulate visual scene information by image editing models and to design the metrics based on scene changes. This allows us to clearly assess whether VLMs correctly understand a given scene by observing the ability to perceive changes. We also visualize image-wise object relationship by virtue of our two-axis view: vision and text. Upon evaluating VLMs with our dataset, we observed that our metrics reveal different aspects of VLM hallucination that have not been reported before. Project page: https://beafbench.github.io/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 526, + 230, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 230, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 230, + 538 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 551, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 647 + ], + "type": "text", + "content": "Vision language models (VLMs) [1, 5, 8, 23, 24] have recently emerged by its multi-modal reasoning capability. It is indeed promising research toward making intelligent agents perceive the world. VLMs comprise a vision encoder as a visual perception module and a large language model (LLM) as a reasoning module. Recent LLMs [6, 32, 33] have exhibited powerful reasoning capabilities across various tasks, including commonsense reasoning [3, 16] and math [7], in zero-/few-shot ways. This success motivates researchers to integrate LLMs with vision encoders to open LLMs' eyes to see the world. Leveraging the reasoning" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 313, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 313, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 313, + 665 + ], + "type": "text", + "content": "* Authors contributed equally to this work." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "capabilities of LLMs alongside visual perception enables VLMs to excel in various visual tasks, such as image captioning [2,38] and visual question answering [11, 15] without task-specific fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 155, + 480, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 155, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 155, + 480, + 274 + ], + "type": "text", + "content": "Despite this encouraging demonstration, VLMs are susceptible to the hallucination [12,14,19,34,39,42] wherein their outputs do not reflect facts presented in input images. For example, although a given input image does not present an object, VLMs often incorrectly answer for the presence of the object. This tendency is problematic, akin to spreading falsehoods in human communication. Hallucination prevents genuine communication, consequently diminishing the reliability of interactions between agents or humans. No matter how good VLM's answer is, if hallucination is included in the answer, it can undermine trust in the system's reliability and credibility. Therefore, it is crucial to address hallucinations to improve the system's responsibility and trustworthiness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 481, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 481, + 456 + ], + "type": "text", + "content": "Understanding the impact of hallucinations on model performance requires the development of an assessment framework, which requires a deep exploration of the underlying causes. The prior works [14, 41, 42] have been proposed to unveil the causation of hallucinations in terms of data and models by reporting their performance on existing hallucination benchmarks. As a widely used hallucination benchmark, POPE [19] follows the question-and-answer (QnA) format. For the convenience of measuring, POPE uses a discriminative question, e.g., \"Is there {object} in the image?\", and gets a YES or NO answer. The following hallucination benchmarks [14, 19, 34] commonly leverage LLMs or human annotators to generate question prompts with available scene information such as bounding boxes and captions. This text-axis evaluation may assess certain aspects of the hallucination performance well, but it is hard to disentangle sources of hallucination, particularly in VLMs that handle both visual and text modalities. Thus, to thoroughly analyze and understand the behavior of VLMs, it is both natural and necessary to consider both modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 459, + 481, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 481, + 590 + ], + "type": "text", + "content": "In this work, we propose a BEFORE-AFTER (BEAF) hallucination evaluation benchmark. This benchmark comprises a dataset enhanced along the vision-axis and metrics that are aware of changes. We manipulate scenes by selectively removing objects on the vision-axis while augmenting QnA pairs along the text-axis. We can observe how the models' answers change as the image is manipulated for the same question. For example, as illustrated in Fig. 1, if we remove an apple from the given image and then ask \"Is there an apple?\", the answer from the model should change from \"Yes\" to \"No\", between the original and modified images, if it truly relies on what it sees. Therefore, for a more thorough evaluation and analysis, we need change-aware metrics beyond just using standard accuracy [14, 19, 34]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "We propose four new metrics: True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID), which are enabled by our vision and text augmentation. TU measures whether the model answers correctly before and after the manipulation. IG evaluates whether the models perceive the images consistently both before and after manipulation. SB is related to the phenomenon of giving the same answer instead of correctly recognizing the image. ID measures" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Ye-Bin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 302, + 237 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 302, + 237 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 302, + 237 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 302, + 237 + ], + "type": "image", + "image_path": "ec68970f63d0e326873895ee4d8be40f9eed1d21bad8bd22363188c352a6c607.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 309, + 116, + 465, + 230 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 465, + 230 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 465, + 230 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 465, + 230 + ], + "type": "table", + "html": "
MTQuestionMetricsCorrectnessAnswer
Text-Acc○-
Img. & TextRemoved ObjectTU○ → ○Y → N
IG× → ×N → Y
SBp○ → ×Y → Y
SBn× → ○N → N
Unremoved ObjectID○ → ×\n× → ○Y → N\nN → Y
", + "image_path": "1dff7727fd68446bb59c22ae30d3e783e5fbe4d9a5c3deba59ef8c167d18814d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 231, + 367, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 231, + 367, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 367, + 237 + ], + "type": "text", + "content": "MT: Manipulation Target" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 141, + 252, + 470, + 399 + ], + "blocks": [ + { + "bbox": [ + 267, + 242, + 341, + 251 + ], + "lines": [ + { + "bbox": [ + 267, + 242, + 341, + 251 + ], + "spans": [ + { + "bbox": [ + 267, + 242, + 341, + 251 + ], + "type": "text", + "content": "BEAF Evaluation (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 252, + 470, + 399 + ], + "lines": [ + { + "bbox": [ + 141, + 252, + 470, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 252, + 470, + 399 + ], + "type": "image", + "image_path": "44cd7026e1a3dc75993230db3fa4f4fb633fa90398a910778da9f45b8e53c474.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 408, + 482, + 497 + ], + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 497 + ], + "type": "text", + "content": "Fig. 1: Illustration of BEbefore-AFTER (BEAF) benchmark. We present a comparison between traditional evaluation benchmarks and our BEAF benchmark for assessing the hallucination behavior in VLMs. Traditional evaluation methods solely manipulate questions based on the existence of an object and measure accuracy or F1 score. In contrast, our BEAF benchmark not only constructs questions but also manipulates images and tracks changes in answers as the images undergo manipulation. The BEAF benchmark introduces novel metrics including True Understanding (TU), Ignorance (IG), StuBbornness (SB), and InDecision (ID), which consider the changes in answers." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 569 + ], + "type": "text", + "content": "cases where the answer changes even though the target object does not change or does not exist in the image. The first three metrics address questions regarding target objects that have been removed, while the last one belongs to questions about other objects not related to the targets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Based on our BEAF dataset and change-aware metrics, we can evaluate and understand the models, even the parts we did not know through the previous methods. The evaluation results imply that answers assumed as non-hallucination from the existing hallucination benchmarks could be a hallucination. We also analyze the influence of each object in the scene based on the variation along the vision and text axes. In addition, we include a discussion on open-generation answers and CLIPScore results. Our key contributions are as follows:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 372, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 372, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEAF Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 209 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 162 + ], + "type": "text", + "content": "- We propose a benchmark, called BEFORE-AFTER (BEAF), to evaluate the hallucination of large vision-language models. Our benchmark comprises a dataset that considers both text- and vision-axes, and change-aware metrics, which allow us to perform granular evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 164, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 164, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 131, + 164, + 479, + 186 + ], + "type": "text", + "content": "- We demonstrate that previously reported outcomes deemed non-hallucinatory in the prior text-axis-only evaluation benchmarks may be hallucinations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 188, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 188, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 188, + 479, + 209 + ], + "type": "text", + "content": "- We also visualize the impact of relationships between objects within hallucinatory images based on the BEAF results." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 232, + 237, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 237, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 237, + 244 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 258, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 365 + ], + "type": "text", + "content": "Vision Language Models (VLMs). VLMs [1, 4, 5, 8, 10, 23-25, 40, 43] have demonstrated their impressive image describing and reasoning ability based on powerful large language models (LLMs) [6, 31-33]. Generally, VLMs consist of three parts: vision encoder, LLM, and vision-language alignment module. The vision information extracted from the vision encoder is translated to the LLM-understandable token by passing through the alignment module. LLM takes a prompt text and visual tokens as input and outputs response text tokens according to a given image and prompt. Their training process is divided into multiple steps depending on which parts of VLMs are trained." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 366, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 481, + 510 + ], + "type": "text", + "content": "Hallucination in VLMs. Hallucination [12, 19, 22, 34, 35, 39] in VLMs refers to the case where the generated text answer does not reflect the true contents of the provided images but rather relies on the internal knowledge of the models. This behavior might cause harmful consequences in high-risk applications such as the medical domain or auto-driving scenarios. This undesirable behavior could be caused by various factors, e.g., data imbalance, image resolution, and model capacity. The assessment of hallucination is also related to the true understanding of VLMs. If VLMs respond to the given input correctly rather than relying on innate knowledge, VLMs are considered to have a perfect understanding of the perceived world. Numerous methodologies [12, 21, 42] have been developed to tackle the issue of hallucination in VLMs by addressing data biases [14, 21], optimization techniques [12, 18, 29], and post-processing methodologies [9, 37, 42]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "content": "Evaluating Hallucination in VLMs. Evaluating hallucination performance is essential to guide the research of VLM. Previous studies have introduced hallucination benchmarks to understand and assess these undesirable phenomena of VLMs. Motivated by traditional visual and question benchmarks, existing hallucination benchmarks [14, 19, 34] follow this VQA style that provides a simple QnA evaluation pipeline and intuitive metrics. POPE [19] and CIEM [14] curate images and QnA pairs with metrics such as accuracy, precision, and recall, where they use a specific question type inducing multi-choice or yes/no answers, called discriminative type. AMBER [34] includes an additional question type that induces open-ended caption generation as answers, called generative type. For such open-generation assessment, they extract the objects from the output text answers and compare them with the ground-truth objects in the image. These prior studies focus on manipulating texts of questions and answers only," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Ye-Bin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 157, + 445, + 214 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 148 + ], + "type": "text", + "content": "Table 1: Data statistics. Our BEAF dataset contains 26K image-question pairs, consisting of the original and manipulated ones. On average, an original image is associated with 3.45 manipulated images and 11.72 questions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 157, + 445, + 214 + ], + "lines": [ + { + "bbox": [ + 170, + 157, + 445, + 214 + ], + "spans": [ + { + "bbox": [ + 170, + 157, + 445, + 214 + ], + "type": "table", + "html": "
ImageImg-Q PairYesNo
Original5005,6532,0263,627
Manipulated1,72720,4656,26714,198
Total2,22726,1188,29317,825
", + "image_path": "cf69220a14cf4c8f672df7ef0b1b365ad4252825f0556c840607853a7ea98c4a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 232, + 357, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 232, + 357, + 241 + ], + "spans": [ + { + "bbox": [ + 255, + 232, + 357, + 241 + ], + "type": "text", + "content": "Ratio of upper-class categories" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 140, + 251, + 297, + 344 + ], + "blocks": [ + { + "bbox": [ + 140, + 251, + 297, + 344 + ], + "lines": [ + { + "bbox": [ + 140, + 251, + 297, + 344 + ], + "spans": [ + { + "bbox": [ + 140, + 251, + 297, + 344 + ], + "type": "image", + "image_path": "e24d158a9e1388e6e8825eb08336756f3f01b27fd5cd513cf3da0bd940306a87.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 355, + 272, + 364 + ], + "lines": [ + { + "bbox": [ + 168, + 355, + 272, + 364 + ], + "spans": [ + { + "bbox": [ + 168, + 355, + 272, + 364 + ], + "type": "text", + "content": "Removed objects (total: 1,727)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 320, + 251, + 470, + 344 + ], + "blocks": [ + { + "bbox": [ + 320, + 251, + 470, + 344 + ], + "lines": [ + { + "bbox": [ + 320, + 251, + 470, + 344 + ], + "spans": [ + { + "bbox": [ + 320, + 251, + 470, + 344 + ], + "type": "image", + "image_path": "a1fdf2aa9cd32bd516a63a06377f0481df96a98192bd056694af6b05a9c3f628.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 355, + 450, + 364 + ], + "lines": [ + { + "bbox": [ + 335, + 355, + 450, + 364 + ], + "spans": [ + { + "bbox": [ + 335, + 355, + 450, + 364 + ], + "type": "text", + "content": "Objects in question (total: 26,118)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 377, + 480, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 377, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 480, + 422 + ], + "type": "text", + "content": "Fig. 2: Ratio of removed objects from image and object in question. For convenience, we report the ratio of upper-class categories instead of the object class itself. The total number of removed objects is the same as the number of manipulated images, and one object in question is the same as the number of image-question pairs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 451, + 480, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 480, + 534 + ], + "type": "text", + "content": "which we call text-axis evaluation. However, if some objects always co-occur with another object, we cannot judge whether VLMs truly understand scene information accurately. Also, VLMs take multi-modal inputs, both images and text prompts; thus, evaluation considering only the text-axis would be insufficient as a hallucination assessment of VLMs. In this work, we manipulate the scene information and questions together, i.e., vision-/text-axes; it enables the fine-grained analysis of hallucination. We call it vision-/text-axes evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 567, + 433, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 567, + 433, + 580 + ], + "spans": [ + { + "bbox": [ + 131, + 567, + 433, + 580 + ], + "type": "text", + "content": "3 BEAF Evaluation Benchmark for Hallucination" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "We explore another axis, i.e., vision-axis, to make up for the deficiencies in the text-axis evaluation. Our key idea is to manipulate a visual scene by removing and inpainting objects within an image. This allows us to focus on changes and to check whether the VLMs are correctly aware of it. We describe how to curate our dataset in Sec. 3.1, and then elaborate our new metrics in Sec. 3.2" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEAF Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 482, + 293 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 482, + 293 + ], + "type": "image", + "image_path": "52c8686b18ccccc1418f502deb7ae6977b05d55c26ff0796a3a3c7847d506d72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 304, + 482, + 382 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 382 + ], + "type": "text", + "content": "Fig. 3: Image manipulation pipeline. We illustrate the image manipulation pipeline comprised of three stages. In Stage 1, we automatically remove target objects sharing the same semantic class from the given images. Stage 2 is to filter the automatically manipulated results based on the predefined rules, such as mask errors, remained shadows, and low-quality outcomes. Undesirable manipulations are either corrected or discarded during this stage. Finally, in Stage 3, human annotators engage in human-guided manipulation of the filtered images to achieve high-quality results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 408, + 257, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 257, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 257, + 419 + ], + "type": "text", + "content": "3.1 Evaluation Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 436, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 628 + ], + "type": "text", + "content": "Dataset Overview. Table 1 and Fig. 2 summarize our data statistics. The dataset consists of images and their associated QnA. The image type is divided into original and manipulated. Five hundred original images are sampled from the validation set of the MS-COCO dataset [20], which shares the same split with POPE [19], and their associated 1,727 manipulated images are generated by removing objects from the original images. On average, 3.45 manipulated images are created from the original images. The main question format is \"Is there {object} in this image?\" and {object} can be a positive object within the image or a negative object not in the image; the positive object is one of the objects in the image, and the negative object is the one not in the image. We adopt the positive and negative objects from the POPE for the original image and adapt the question for the manipulated image. If we remove the positive object from the original image to make the manipulated one, the positive object becomes the negative object in the manipulated one. The total dataset comprised 2,227 images and 26,118 image-question pairs. The \"Yes\" and \"No\" ground truth answers are 8,293 (31.75%) and 17,825 (68.25%), respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Image Dataset Construction Pipeline. We manipulate the original images by removing objects within the image. To collect the manipulated images, we pass through a three-stage pipeline (see Fig. 3): the first stage involves auto" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Ye-Bin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "matic manipulation, followed by a filtering stage, and finally, a human-guided manipulation stage. We offer a detailed explanation of each stage as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 284 + ], + "type": "text", + "content": "1) First Stage: Automatic Manipulation. For automatic object removal, we use an off-the-shelf inpainting model, LaMa [30], with mask guidance. While the initial option is to use the ground truth masks provided within the original dataset, we opt not to use them because of the low quality ground truth masks in MS-COCO. For example, polygon masks often cannot cover all areas of the corresponding object and can be propagated to the inpainting quality. Therefore, we employ a pre-trained mask extractor called SAM [17], utilizing bounding box annotations. Subsequently, we dilate the extracted mask to have a margin for more effective object removal, and remove the object with the inpainting model given the dilated mask guidance. We empirically choose to use LaMa over diffusion models, e.g., [28, 44], as we find that this specialized model performs better in the context of object removal." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 429 + ], + "type": "text", + "content": "2) Second Stage: Filtering. Although we remove objects from the original images, achieving consistent quality in the outcome can be influenced by a myriad of factors. These factors include, but are not limited to, the size of the objects being removed and the complexity of the backgrounds they inhabit. Additionally, the presence of shadows remains unaddressed, as they are not effectively masked using SAM; the ground truth masks also do not capture the shadow. As shown on the right side of Fig. 3, it displays undesirable images due to the presence of shadow or a faucet. The human annotators classified the results obtained from the previous stage into usable or unusable, where only the usable results are further manipulated. The detailed filtering rules can be found in the supplementary material. We directly use the usable results and hand over the results that need further manipulation to the next stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 429, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 548 + ], + "type": "text", + "content": "3) Third Stage: Human-guided Manipulation. The third stage of our workflow relies on human annotators to intervene and refine the images obtained from the second stage. As mentioned above, these filtered images often exhibit undesirable artifacts such as ghost shadows, afterimages, and fragmented objects. Even such a small artifact may provide strong cues indicating the past presence of objects that both humans and machines can deduce, e.g., a cat's tail suggests the past presence of the cat before the object removal. Thus, human annotators are asked to meticulously manipulate the images, either correcting or enhancing them. Annotators effectively refine visual data, aligning it more closely with real natural images in human perception." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 549, + 481, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 481, + 585 + ], + "type": "text", + "content": "As shown in Fig. 3, our annotation process results in images that are free of the aforementioned artifacts. This manual intervention is crucial, as it ensures the fidelity and reliability of our benchmark dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 606, + 257, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 606, + 257, + 617 + ], + "spans": [ + { + "bbox": [ + 131, + 606, + 257, + 617 + ], + "type": "text", + "content": "3.2 Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "The prior studies [14, 19, 34] suggest using traditional metrics such as accuracy, precision, recall, and F1 score, which we can also use the metrics in our dataset. In addition, we propose four new metrics: True Understanding (TU), IGnorance" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 372, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 372, + 91, + 447, + 101 + ], + "type": "text", + "content": "BEAF Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "type": "text", + "content": "(IG), StuBbornness (SB), and InDecision (ID), for more detailed evaluation by exploiting the distinctive configuration of before-/after-changes in our dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{(I_o, I_m, Q, A_o, A_m, R)_i\\}_{i=1}^{|S|}" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " denote our dataset, where each tuple consists of an original image " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "I_o" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ", a version of its manipulated image " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "I_m" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ", their associated question " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ", their respective corresponding answers " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "A_o" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "A_m" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ", and a True/False flag " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " that indicates whether the question is associated with the object being removed in the manipulated image " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "I_m" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ". For compactness of our metric definitions, we define " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\text{Filter}(\\cdot)" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " to extract the subset of " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " satisfying input conditions as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 234, + 473, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 234, + 473, + 248 + ], + "spans": [ + { + "bbox": [ + 140, + 234, + 473, + 248 + ], + "type": "interline_equation", + "content": "\\operatorname {F i l t e r} \\left(b _ {o}, b _ {m}, b _ {r}\\right) = \\{t | \\text {I s C o r r e c t} \\left(A _ {o}\\right) = b _ {o}, \\text {I s C o r r e c t} \\left(A _ {m}\\right) = b _ {m}, R = b _ {r}, t \\in \\mathcal {S} \\}", + "image_path": "8e6d3644ed57cf9d7a86559573dc26c935df98448f8a5762a69187ce2ced33d4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "text", + "content": "where the tuple " + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "inline_equation", + "content": "t = (I_o, I_m, Q, A_o, A_m, R)" + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "text", + "content": ". For example, Filter(True, False, True) is the subset of " + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "text", + "content": " of which the element satisfy the conditions that are 1) the answers corresponding to respective original and manipulated images are correct and wrong, and 2) the question is the one related to the object removed in the manipulated images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 316, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 387 + ], + "type": "text", + "content": "True Understanding. True Understanding (TU) measures whether the models truly understand the scene. It evaluates if the model can accurately answer to the question about the removed object. As shown in Fig. 1, if the apple is removed, a model that truly understands the change should switch its answer from \"Yes\" to \"No\" to the question \"Is there an apple?\" We define the TU metric as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 216, + 394, + 480, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 394, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 216, + 394, + 480, + 411 + ], + "type": "interline_equation", + "content": "\\mathrm {T U} = \\frac {\\left| \\text {F i l t e r} (\\text {T r u e} , \\text {T r u e} , \\text {T r u e}) \\right|}{\\left| \\text {F i l t e r} (\\text {T r u e} \\vee \\text {F a l s e} , \\text {T r u e} \\vee \\text {F a l s e} , \\text {T r u e}) \\right|} \\cdot 1 0 0. \\tag {1}", + "image_path": "0d00a0e3948c9f13721a5f689306ddafb0b88e7e80b16e2112e897b4f568b9e7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 418, + 479, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 479, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 479, + 490 + ], + "type": "text", + "content": "Compared to the traditional accuracy, our metric specifically considers the correctness of answers before and after manipulation. Therefore, it provides a conditional measure of accuracy, taking into account the model's response to changes within the scene. A higher value of this metric indicates a more accurate understanding of the scene by the model, as it demonstrates the model's ability to adapt its answers appropriately to the scene changes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 491, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 574 + ], + "type": "text", + "content": "Ignorance. The second metric is IGnorance (IG), which measures the extent to which models lack knowledge about specific scene information. This metric quantifies ignorance by identifying instances where models provide incorrect answers to the question about the removed objects. As shown in Fig. 1, ignorance is evident if models incorrectly answer \"No\" to the presence of an apple before its removal and incorrectly answer \"Yes\" after the apple is removed, indicating a failure to recognize it. IG is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 219, + 582, + 480, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 582, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 219, + 582, + 480, + 598 + ], + "type": "interline_equation", + "content": "\\mathrm {I G} = \\frac {\\left| \\text {F i l t e r (F a l s e , F a l s e , T r u e)} \\right|}{\\left| \\text {F i l t e r (T r u e V F a l s e , T r u e V F a l s e , T r u e)} \\right|} \\cdot 1 0 0. \\tag {2}", + "image_path": "abf1f63330f635dc5de4d47a5eff2e41f01b2c5779333ca552468e9294daa7c1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 605, + 363, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 363, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 363, + 616 + ], + "type": "text", + "content": "The lower the metric, the less ignorant the model is." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Stubbornness. What we observe is that VLMs tend to output the same answer repeatedly. It means that VLMs are biased toward the specific answer rather than arbitrary one when they do not know. This repetitive behavior is undesirable; thus, we introduce the metric of StuBbornness (SB), which measures the extent" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Ye-Bin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "to which models adhere to their initial answers. Furthermore, we categorize SB into " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathrm{SB}_p" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathrm{SB}_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " where the subscripts of " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " correspond to the consistent positive (\"Yes\") and negative (\"No\") answers, respectively. We define them as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 481, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 190 + ], + "type": "interline_equation", + "content": "\\mathrm {S B} _ {p} = \\frac {1 0 0 \\cdot | \\text {F i l t e r (T r u e , F a l s e , T r u e)} |}{| \\text {F i l t e r (T r u e \\vee F a l s e , T r u e \\vee F a l s e , T r u e)} |}, \\quad \\mathrm {S B} _ {n} = \\frac {1 0 0 \\cdot | \\text {F i l t e r (F a l s e , T r u e , T r u e)} |}{| \\text {F i l t e r (T r u e \\vee F a l s e , T r u e \\vee F a l s e , T r u e)} |}. \\tag {3}", + "image_path": "7d82aa4259e7e20dc4af00defbc651f85b03cc67d66a115bc8ba1c1c1a00b84e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 479, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 479, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 479, + 221 + ], + "type": "text", + "content": "SB is the summation of " + }, + { + "bbox": [ + 130, + 198, + 479, + 221 + ], + "type": "inline_equation", + "content": "\\mathrm{SB}_p" + }, + { + "bbox": [ + 130, + 198, + 479, + 221 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 198, + 479, + 221 + ], + "type": "inline_equation", + "content": "\\mathrm{SB}_n" + }, + { + "bbox": [ + 130, + 198, + 479, + 221 + ], + "type": "text", + "content": " and can be computed from " + }, + { + "bbox": [ + 130, + 198, + 479, + 221 + ], + "type": "inline_equation", + "content": "100 - \\mathrm{TU} - \\mathrm{IG}" + }, + { + "bbox": [ + 130, + 198, + 479, + 221 + ], + "type": "text", + "content": ". The lower the metric, the less ignorant the model is." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 222, + 480, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 480, + 304 + ], + "type": "text", + "content": "Indecision. Unlike the previous three metrics that focus on the removed objects in the manipulated images, InDecision (ID) focuses on the answers to the questions that are not relevant to the removed objects. These answers should not be changed even after the manipulation. Nevertheless, models, often change their answers about unmanipulated objects, which would imply that the model predicts a random answer without grounding it on the image. We track these behaviors with the ID metric defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 313, + 481, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 313, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 194, + 313, + 481, + 331 + ], + "type": "interline_equation", + "content": "\\mathrm {I D} = \\frac {\\left| \\text {F i l t e r} (\\text {T r u e} , \\text {F a l s e} , \\text {F a l s e}) \\right| + \\left| \\text {F i l t e r} (\\text {F a l s e} , \\text {T r u e} , \\text {F a l s e}) \\right|}{\\left| \\text {F i l t e r} (\\text {T r u e} \\vee \\text {F a l s e} , \\text {T r u e} \\vee \\text {F a l s e} , \\text {F a l s e}) \\right|} \\cdot 1 0 0. \\tag {4}", + "image_path": "63c1b9f78c86c615471b887a14c4c302440c974273cc5a5b8bb4de4501c1d0a9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 339, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 479, + 378 + ], + "type": "text", + "content": "The lower the metric, the less ignorant the model is. Additionally, we combine TU and ID using a harmonic mean to make it easier to compare on our benchmark. We refer to it as F1-Score, which is computed as follows: " + }, + { + "bbox": [ + 131, + 339, + 479, + 378 + ], + "type": "inline_equation", + "content": "F1 = \\frac{2}{TU^{-1} + (100 - ID)^{-1}}" + }, + { + "bbox": [ + 131, + 339, + 479, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 395, + 224, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 395, + 224, + 409 + ], + "spans": [ + { + "bbox": [ + 131, + 395, + 224, + 409 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 422, + 480, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 505 + ], + "type": "text", + "content": "In Sec. 4.1, we evaluate the current VLMs using our BEAF dataset. We employ both proposed and traditional metrics and compare their performance with other benchmarks. Furthermore, we conduct an analysis of image-wise dynamic correctness by visualizing image-wise object relationship containing vision and text axes. We provide a report on the object-wise error rate. In Sec. 4.2, we assess the performance of VLMs on the open generation task using our dataset, while also analyzing both the dataset and VLMs based on the CLIP score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 506, + 481, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 481, + 541 + ], + "type": "text", + "content": "All the reported performance is the zero-shot performance of VLMs, including LLaVA [23], InstructBLIP [8], Shikra [5], and mPLUG-Owl [36]. We use prompt templates provided in their released codebases or papers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 560, + 228, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 228, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 228, + 572 + ], + "type": "text", + "content": "4.1 Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": "Evaluation on BEAF benchmark. We evaluate several VLMs on our BEAF dataset using the proposed change-aware metrics, TU, IG, SB, ID, and F1. The evaluation results are presented in Table 2. Among the 13B-size models, LLaVA outperforms InstructBLIP in terms of TU, IG, and SB because InstructBLIP consistently answers \"Yes\" regardless of scene manipulation, and this result is related to its high value of " + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{SB}_p" + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": ". InstructBLIP achieves better ID compared to LLaVA. Among 7B-size models, Shikra exhibits the best in TU and " + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{SB}_p" + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": ". Shikra" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 101 + ], + "type": "text", + "content": "BEAF Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 179, + 479, + 252 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 169 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 169 + ], + "type": "text", + "content": "Table 2: Evaluation results on BEAF benchmark. We evaluate various VLMs on our benchmark. F1 is a harmonic mean of TU and 1-ID. Yes ratio is the percentage of total answers that are answered \"Yes\" for the removed object-related question. The former is for the original image, and the latter stands for the manipulated image. Underline stands for the best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 179, + 479, + 252 + ], + "lines": [ + { + "bbox": [ + 135, + 179, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 135, + 179, + 479, + 252 + ], + "type": "table", + "html": "
SizeModelTU(↑)IG(↓)SBp(↓)SBn(↓)ID(↓)F1(↑)Yes (%)
13BLLaVA-v1.524.30.272.03.56.438.696.4 → 72.2
InstructBLIP11.10.384.44.26.119.995.5 → 84.7
7BLLaVA-v1.532.60.161.36.05.648.593.9 → 61.4
InstructBLIP33.50.651.314.65.249.584.8 → 51.9
Shikra52.70.431.215.75.767.683.9 → 31.6
mPLUG-Owl224.60.072.13.37.038.896.7 → 72.1
", + "image_path": "1dc58c44bc9fccc29173c45ba95e04206fb309b910058f5978e8f47fb8b20805.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 322, + 479, + 405 + ], + "blocks": [ + { + "bbox": [ + 132, + 268, + 481, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 312 + ], + "type": "text", + "content": "Table 3: Traditional evaluation results on BEAF dataset. We evaluate various VLMs on our dataset in terms of accuracy, precision, recall, and F1 score. Note that the BEAF dataset includes both original and manipulated data but does not consider the changes for this traditional evaluation. Underline stands for the best." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 322, + 479, + 405 + ], + "lines": [ + { + "bbox": [ + 135, + 322, + 479, + 405 + ], + "spans": [ + { + "bbox": [ + 135, + 322, + 479, + 405 + ], + "type": "table", + "html": "
SizeModelAcc.(↑)Precision(↑)Recall(↑)F1(↑)Yes (%)
13BLLaVA-v1.574.956.393.470.352.6
InstructBLIP71.552.894.367.756.7
7BLLaVA-v1.579.762.689.573.745.4
InstructBLIP82.068.380.774.037.5
Shikra84.574.976.975.932.6
mPLUG-Owl269.150.793.865.858.7
", + "image_path": "4d80e89772a6212511b50847784212d31a115207be49af89d3dbc834fe96ca15.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 433, + 481, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 481, + 505 + ], + "type": "text", + "content": "is trained using a location-aware strategy, which could be beneficial for hallucination evaluation because knowledge of location aids in judging the existence of an object. The overall results indicate that, although it is rare to get all the answers wrong, it is common for the answers to remain the same even though the image has changed. This indicates that recent VLMs are not adequately reflective of such changes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 508, + 482, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 482, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 482, + 663 + ], + "type": "text", + "content": "Traditional evaluation on BEAF dataset. In addition to the proposed change-aware metrics, we also evaluate using traditional metrics such as accuracy, precision, recall, and F1 score, as shown in Table 3. While the trends in accuracy or F1 score are similar to those shown in TU results of Table 2, the absolute values of TU are much lower than those of the traditional metrics. This discrepancy suggests that, although a model achieves high performance in the existing benchmarks, our benchmark tells us that we cannot conclude low hallucination and the models may still suffer from hallucination. For example, although LLaVA-v1.5-13B achieves high performance on accuracy, the Yes ratio in Table 2 remains high even after the manipulation. It may include instances where the correct answer was guessed without actual reference to the given image. The experimental findings underscore the importance of observing changes in answers along the vision-axis to accurately evaluate hallucination in VLMs." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Ye-Bin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 242, + 185 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 242, + 185 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 242, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 242, + 185 + ], + "type": "image", + "image_path": "b51017f2148922e95dfd337608e7b4b084a057d71780364f1045ca279e89e131.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 185, + 238, + 251 + ], + "blocks": [ + { + "bbox": [ + 133, + 185, + 238, + 251 + ], + "lines": [ + { + "bbox": [ + 133, + 185, + 238, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 238, + 251 + ], + "type": "image", + "image_path": "f12711c2b3f2e8474e8f0e351ac8d7378ffc8cc32c2422eb301276e374107087.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 340 + ], + "type": "text", + "content": "Fig. 4: Visualization of image-wise object relationship. We visualize image-wise object relationship along the text and vision axes from Shikra. [Top] Original image samples. [Bottom] The object relation table given the manipulated images and questions. We color the correct answer blue and the wrong one red. The text axis stands for the target object queried in the question, and the vision axis for the removed object in the image. The \"none\" in the vision axis means the original image (not manipulated). Thereby, we can analyze influence between objects within a scene at once." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 255, + 113, + 361, + 185 + ], + "blocks": [ + { + "bbox": [ + 255, + 113, + 361, + 185 + ], + "lines": [ + { + "bbox": [ + 255, + 113, + 361, + 185 + ], + "spans": [ + { + "bbox": [ + 255, + 113, + 361, + 185 + ], + "type": "image", + "image_path": "b73916b14e5e81779ff87ab4407c068f9a741ba32636110748280b05d0049b06.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 260, + 190, + 356, + 251 + ], + "blocks": [ + { + "bbox": [ + 260, + 190, + 356, + 251 + ], + "lines": [ + { + "bbox": [ + 260, + 190, + 356, + 251 + ], + "spans": [ + { + "bbox": [ + 260, + 190, + 356, + 251 + ], + "type": "image", + "image_path": "3f1f3b70a1cf372774eaecff39359a2f173e26b2917dba53ed6db7db28fd64b7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 376, + 113, + 479, + 185 + ], + "blocks": [ + { + "bbox": [ + 376, + 113, + 479, + 185 + ], + "lines": [ + { + "bbox": [ + 376, + 113, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 376, + 113, + 479, + 185 + ], + "type": "image", + "image_path": "e3bb7ece665e16e8f5678700b7d6cba2f09b196ee3b9dd7aab732df981b17a4d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 373, + 186, + 479, + 251 + ], + "blocks": [ + { + "bbox": [ + 373, + 186, + 479, + 251 + ], + "lines": [ + { + "bbox": [ + 373, + 186, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 373, + 186, + 479, + 251 + ], + "type": "image", + "image_path": "86ff2b31abb6e666dba9e9d57516fd95e90dcd065577e66d302d253a7b0b419d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 298, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 298, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 298, + 484 + ], + "type": "text", + "content": "Comparison with other benchmarks. We compare our metric TU with the performance for other benchmarks of VQAv2 [11] and GQA [15], as shown in Table 4. Each VQA performance value is reported from the proposed papers. The results indicate that, while mPLUG-Owl2 outperforms others on VQAv2 and LLaVA excels on GQA, Shikra leads in per" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 485, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 555 + ], + "type": "text", + "content": "formance on our metric. We observe that there is a negative correlation between performance and the incidence of hallucination, which aligns with the previous observation [35]. Although improving performance on various benchmarks is challenging, it is essential to check multiple aspects of the performance, including hallucination. We believe that our benchmark provides valuable insights in conjunction with existing benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": "Image-wise object relationship. We visualize changes in dynamic object relationship along the vision and text axes, as shown in Fig. 4. Compared to our metrics providing comprehensive statistics as summarized scores, this visualization enables to track the changes in VLMs' responses at object-wise and scene-wise levels. The x-axis is the text axis where each object is associated with the question, \"Is there {object}?\"; the y-axis is the vision axis where the object is removed from the original image. The blue color is correct, the red one is wrong." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In the first column in Fig. 4, altering the scene affects the textual correctness. Specifically, removing either the couch or the book shifts the book's textual cor-" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 321, + 422, + 470, + 479 + ], + "blocks": [ + { + "bbox": [ + 304, + 377, + 482, + 421 + ], + "lines": [ + { + "bbox": [ + 304, + 377, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 482, + 421 + ], + "type": "text", + "content": "Table 4: Comparison of performance on VQA and TU. We compare the accuracy on VQA datasets, VQAv2 and GQA, as well as TU on BEAF dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 422, + 470, + 479 + ], + "lines": [ + { + "bbox": [ + 321, + 422, + 470, + 479 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 470, + 479 + ], + "type": "table", + "html": "
Model (7B)VQAv2GQATU
LLaVA-v1.578.562.032.6
InstructBLIP-49.233.5
Shikra77.4-52.7
mPLUG-Owl279.456.124.6
", + "image_path": "b15836e1aaea5bdd368ef1f665bd88ae2f0ec0e8cb30a870673882619e587768.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 372, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 372, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEAF Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 163, + 118, + 449, + 222 + ], + "blocks": [ + { + "bbox": [ + 163, + 118, + 449, + 222 + ], + "lines": [ + { + "bbox": [ + 163, + 118, + 449, + 222 + ], + "spans": [ + { + "bbox": [ + 163, + 118, + 449, + 222 + ], + "type": "image", + "image_path": "c6b8a87ffc47806c59d71c613b7a4c02da54362f5a20f63f3880eceba953a9d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 288 + ], + "type": "text", + "content": "Fig. 5: Object-wise error rate on the manipulated image-question pairs. We plot the error rate of each object to investigate which object is inaccurately inferred when it is removed. We exclude the objects manipulated less than 20 times. The solid line represents the average error rate of VLMs (7B), while the purple area indicates the " + }, + { + "bbox": [ + 130, + 232, + 482, + 288 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 232, + 482, + 288 + ], + "type": "text", + "content": " confidence interval." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 405 + ], + "type": "text", + "content": "rectness from correct to incorrect. This suggests a correlation between the couch and the book, indicating that complex scenes can lead to model hallucinations even in the absence of the book. In the second scene, when the object vase or book is removed, most of the incorrectness becomes correct. This could imply that the presence of a vase or a book might lead the model to misinterpret the scene. Through the simultaneous analysis of both vision and text axes, we observe immediate changes in the answers and offer interpretations for the dynamic interactions observed in the second and third scenes illustrated in the figure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 405, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 477 + ], + "type": "text", + "content": "Object-wise error rate. We compute the mean and " + }, + { + "bbox": [ + 130, + 405, + 482, + 477 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 405, + 482, + 477 + ], + "type": "text", + "content": " confidence interval of the object-wise error after the manipulation in Fig. 5. It represents whether the VLMs can accurately detect the removal. The results show that VLMs frequently fail to recognize certain objects like traffic lights, vases, and remotes. We speculate that the difficulty in recognizing these objects may stem from their complexity or their infrequent appearance in the training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 492, + 249, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 492, + 249, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 249, + 504 + ], + "type": "text", + "content": "4.2 Additional Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 317, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 317, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 317, + 665 + ], + "type": "text", + "content": "Evaluation of open generation. In Table 5, we compare the results of the CHAIR [27] and TU metrics to evaluate the hallucination in the open-ended generation answers on our dataset. Briefly, " + }, + { + "bbox": [ + 130, + 510, + 317, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{CHAIR}_I" + }, + { + "bbox": [ + 130, + 510, + 317, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 317, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{CHAIR}_S" + }, + { + "bbox": [ + 130, + 510, + 317, + 665 + ], + "type": "text", + "content": " count the hallucinated instances and sentences containing these objects in the generated output, respectively. The lower the score, the better. We observe that if the open generation contains the hallucinated object and sentences, TU is also lower, as seen with InstructBLIP-13B." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 324, + 578, + 479, + 645 + ], + "blocks": [ + { + "bbox": [ + 321, + 521, + 482, + 578 + ], + "lines": [ + { + "bbox": [ + 321, + 521, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 521, + 482, + 578 + ], + "type": "text", + "content": "Table 5: Comparison of CHAIR metric and TU. We measure the CHAIR score based on the answers derived from the images in BEAF using the prompt \"Describe this image.\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 578, + 479, + 645 + ], + "lines": [ + { + "bbox": [ + 324, + 578, + 479, + 645 + ], + "spans": [ + { + "bbox": [ + 324, + 578, + 479, + 645 + ], + "type": "table", + "html": "
SizeModelCHAIR1CHAIRSTU
13BLLaVA-v1.521.672.324.3
InstructBLIP28.181.111.1
7BLLaVA-v1.522.373.932.6
InstructBLIP21.067.533.5
Shikra23.677.052.7
mPLUG-Owl225.178.424.6
", + "image_path": "1a1e42b1e7959d0e82e4f758036dc32d89aa4110ceaa67f0da089b46fba4c820.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Ye-Bin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 482, + 247 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 482, + 247 + ], + "type": "image", + "image_path": "d22c0ec6b2135da95057ec1495e3184f36228cc8eeee92da5e0b8be0e81f4d8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 257, + 480, + 302 + ], + "lines": [ + { + "bbox": [ + 131, + 257, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 257, + 480, + 302 + ], + "type": "text", + "content": "Fig. 6: Samples with CLIP Score. We visualize samples and display the CLIP Score below. Additionally, we report the results of Faster R-CNN [26] using the magnifying glass icon. After removing the object, the removed object is not detected, which demonstrates the manipulation quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 327, + 319, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 327, + 319, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 327, + 319, + 422 + ], + "type": "text", + "content": "CILP Score. We compare the CLIP Score [13] between original and manipulated images to ensure the target object has been well-removed. A lot of VLMs utilize the visual encoder from CLIP, prompting us to analyze CLIP's performance with our datasets. We measure the CLIP score between an object and a prompt \"a photo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 480, + 483 + ], + "type": "text", + "content": "of {object}. We use a cropped object image based on the GT bounding box. If the image has multiple objects of the same semantic class, we use the average CLIP score of all the same semantic objects. For the manipulated image, {object} is the removed object. As shown in Table 6, the manipulated images have lower CLIP scores compared to their original counterparts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "text", + "content": "We also visualize the images with the CLIP Score and the multi-class classification results in Fig. 6. The multi-class classification correctly identifies the depicted objects. In addition, Fig. 7 visualizes an open generation sample result, where red-colored text indicates hallucinated words. Before manipulation, all the VLMs mention \"people\" in the background; however, they still mention \"people\" even though all the persons are removed from the image. This implies that the commonly used visual encoder might be a non-negligible source of hallucination." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 589, + 311, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 311, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 311, + 601 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "We present a comprehensive analysis of hallucination using our BEAF benchmark. Unlike conventional hallucination benchmarks that measure accuracy by changing only the text-axis question, we also manipulate images by removing objects and propose four sophisticated metrics to track the changes in answers" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 344, + 373, + 463, + 415 + ], + "blocks": [ + { + "bbox": [ + 325, + 339, + 480, + 372 + ], + "lines": [ + { + "bbox": [ + 325, + 339, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 325, + 339, + 480, + 372 + ], + "type": "text", + "content": "Table 6: CLIP Score. We measure the CLIP Score of original and manipulated images with the text prompt." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 344, + 373, + 463, + 415 + ], + "lines": [ + { + "bbox": [ + 344, + 373, + 463, + 415 + ], + "spans": [ + { + "bbox": [ + 344, + 373, + 463, + 415 + ], + "type": "table", + "html": "
TypeCLIP Score
Original53.7
Manipulated47.9
", + "image_path": "e95344c6ed683babee8348d30542b6e5b59694babaa6e6c3649a54e3c052e472.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 372, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 372, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEAF Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 114, + 459, + 315 + ], + "blocks": [ + { + "bbox": [ + 154, + 114, + 459, + 315 + ], + "lines": [ + { + "bbox": [ + 154, + 114, + 459, + 315 + ], + "spans": [ + { + "bbox": [ + 154, + 114, + 459, + 315 + ], + "type": "image", + "image_path": "f6d9f70d17cefcae75666afbb809b71be2e291da3b7df22fffa144d108365209.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 323, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 132, + 323, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 482, + 357 + ], + "type": "text", + "content": "Fig. 7: Qualitative example on open-ended generation. We report the changes in answers for the open-ended generation task with the prompt \"Describe this image.\" as the image undergoes manipulation. The red color highlights the hallucination parts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 387, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 506 + ], + "type": "text", + "content": "according to image changes. This approach enables us to a more discerning analysis of the hallucination in two-axis views, e.g., vision and text axes. We can find with our metric that traditional accuracy may include instances where the correct answer was guessed without actual reference to the given image. We found that our TU is a more calibrated metric, which allows us to observe the intriguing phenomenon that has not been revealed in the prior work: VLMs tend to be confused if the complexity of an image is high. (refer to Sec. A in the supplementary material). Additionally, we introduce the visualization method to track the correctness as the objects are removed. We believe that our benchmark presents a comprehensive analysis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "Limitation. While our benchmark offers a more fine-grained evaluation, it still faces the constraints of the MS-COCO dataset. These include a limited variety of object categories and an uneven distribution of object instances. While we utilized current state-of-the-art in-painting models and mask extractors, we manually adjusted the images to ensure image quality and dataset integrity. This step prevented us from achieving a fully automated creation process. Although such limitations exist, these aspects are expected to naturally diminish or disappear as in-painting models advance, making our dataset construction pipeline fully automatic and our method applicable to all images. Furthermore, we foresee future evaluations expanding in scale through manipulations like object removal, insertion, or resizing. While our current method may appear as a modest evaluation benchmark, we hope that our work becomes the first step toward controllable hallucination evaluation." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Ye-Bin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub; No.2022-0-00124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities; No.RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH))" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 198, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 198, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 198, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 258, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv (2023) 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 292, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 292, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 480, + 323 + ], + "type": "text", + "content": "2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: IEEE International Conference on Computer Vision (ICCV) (2019) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 324, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 480, + 357 + ], + "type": "text", + "content": "3. Bisk, Y., Zellers, R., Gao, J., Choi, Y., et al.: Piqa: Reasoning about physical commonsense in natural language. In: AAAI Conference on Artificial Intelligence (AAAI) (2020) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "type": "text", + "content": "4. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 390, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 412 + ], + "type": "text", + "content": "5. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv (2023) 1, 4, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 413, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 456 + ], + "type": "text", + "content": "6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 413, + 480, + 456 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 413, + 480, + 456 + ], + "type": "text", + "content": " * chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/1, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "type": "text", + "content": "7. Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., Schulman, J.: Training verifiers to solve math word problems. arXiv (2021) 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 490, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 533 + ], + "type": "text", + "content": "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 1, 4, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "type": "text", + "content": "9. Deng, A., Chen, Z., Hooi, B.: Seeing is believing: Mitigating hallucination in large vision-language models via clip-guided decoding. arXiv (2024) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 587 + ], + "type": "text", + "content": "10. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv (2023) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 588, + 480, + 632 + ], + "type": "text", + "content": "11. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 2, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 632, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 664 + ], + "type": "text", + "content": "12. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. In: AAAI Conference on Artificial Intelligence (AAAI) (2024) 2, 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 101 + ], + "type": "text", + "content": "BEAF Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "13. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv (2021) 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "14. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv (2023) 2, 4, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "text", + "content": "15. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 2, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "text", + "content": "16. Hyun, L., Sung-Bin, K., Han, S., Yu, Y., Oh, T.H.: Smile: Multimodal dataset for understanding laughter in video with language models. arXiv (2023) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 235 + ], + "type": "text", + "content": "17. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv (2023) 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 236, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 480, + 266 + ], + "type": "text", + "content": "18. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 267, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 480, + 299 + ], + "type": "text", + "content": "19. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. In: Conference on Empirical Methods in Natural Language Processing (EMNLP) (2023) 2, 4, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "text", + "content": "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV) (2014) 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "21. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: International Conference on Learning Representations (ICLR) (2024) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "22. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models. arXiv (2024) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "23. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv (2023) 1, 4, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 427 + ], + "type": "text", + "content": "24. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 1, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 460 + ], + "type": "text", + "content": "25. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. In: International Conference on Learning Representations (ICLR) (2024) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 460, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 492 + ], + "type": "text", + "content": "26. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: Advances in Neural Information Processing Systems (NeurIPS) (2015) 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "text", + "content": "27. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: Conference on Empirical Methods in Natural Language Processing (EMNLP) (2018) 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "type": "text", + "content": "28. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "text", + "content": "29. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv (2023) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 632 + ], + "type": "text", + "content": "30. Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. In: IEEE Winter Conf. on Applications of Computer Vision (WACV) (2022) 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "31. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. 3(6), 7 (2023) 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Ye-Bin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 534 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv (2023) 1, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "33. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv (2023) 1, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "34. Wang, J., Wang, Y., Xu, G., Zhang, J., Gu, Y., Jia, H., Yan, M., Zhang, J., Sang, J.: An llm-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv (2023) 2, 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "text", + "content": "35. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023) 4, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "text", + "content": "36. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv (2023) 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "37. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 346 + ], + "type": "text", + "content": "38. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics (TACL) (2014) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 347, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 390 + ], + "type": "text", + "content": "39. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023) 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "text", + "content": "40. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. In: International Conference on Learning Representations (ICLR) (2024) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "type": "text", + "content": "41. Zhao, Y., Li, Z., Jin, Z., Zhang, F., Zhao, H., Dou, C., Tao, Z., Xu, X., Liu, D.: Enhancing the spatial awareness capability of multi-modal large language model. arXiv (2023) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 457, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 490 + ], + "type": "text", + "content": "42. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models. In: International Conference on Learning Representations (ICLR) (2024) 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 491, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 511 + ], + "type": "text", + "content": "43. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv (2023) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 512, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 534 + ], + "type": "text", + "content": "44. Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv (2023) 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEAF Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_content_list.json b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..495c5135716c2b4e66d637a23e67c7bee234d088 --- /dev/null +++ b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_content_list.json @@ -0,0 +1,1832 @@ +[ + { + "type": "text", + "text": "BI-MDRG: Bridging Image History in Multimodal Dialogue Response Generation", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hee Suk Yoon $^{1\\star}$ , Eunseop Yoon $^{1\\star}$ , Joshua Tian Jin Tee $^{1\\star}$ , Kang Zhang $^{1}$ , Yu-Jung Heo $^{2}$ , Du-Seong Chang $^{2}$ , and Chang D. Yoo $^{1\\star\\star}$", + "bbox": [ + 215, + 210, + 785, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Korea Advanced Institute of Science and Technology (KAIST)", + "bbox": [ + 282, + 253, + 715, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ KT Corporation", + "bbox": [ + 437, + 268, + 563, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hskyoon, esyoon97, joshuateetj, zhangkang, cd_yoo}@kaist.ac.kr", + "bbox": [ + 282, + 282, + 718, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yj.heo,dschang}@kt.com", + "bbox": [ + 416, + 297, + 588, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Multimodal Dialogue Response Generation (MDRG) is a recently proposed task where the model needs to generate responses in texts, images, or a blend of both based on the dialogue context. Due to the lack of a large-scale dataset specifically for this task and the benefits of leveraging powerful pre-trained models, previous work relies on the text modality as an intermediary step for both the image input and output of the model rather than adopting an end-to-end approach. However, this approach can overlook crucial information about the image, hindering 1) image-grounded text response and 2) consistency of objects in the image response. In this paper, we propose BI-MDRG that bridges the response generation path such that the image history information is utilized for enhanced relevance of text responses to the image content and the consistency of objects in sequential image responses. Through extensive experiments on the multimodal dialogue benchmark dataset, we show that BI-MDRG can effectively increase the quality of multimodal dialogue. Additionally, recognizing the gap in benchmark datasets for evaluating the image consistency in multimodal dialogue, we have created a curated set of 300 dialogues annotated to track object consistency across conversations. The code and the dataset is publicly available at https://github.com/hee-suk-yoon/BI-MDRG.", + "bbox": [ + 259, + 340, + 743, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multimodal Dialogue $\\cdot$ Image Grounding $\\cdot$ Image Consistency", + "bbox": [ + 259, + 631, + 740, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 683, + 375, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of instant messaging technology, visual modalities are increasingly used alongside text in online communication. To enhance user interaction with intelligent agents, a new task, Multimodal Dialogue Response Generation (MDRG) [43], has been proposed. This task requires models to generate both text and image responses based on dialogue history containing texts and images. Since learning an effective multimodal generation model with a sin", + "bbox": [ + 212, + 712, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution", + "bbox": [ + 217, + 811, + 364, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Corresponding Author", + "bbox": [ + 215, + 825, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75f9b85ffddd660a717389a9fcac18ecb8a61899dbad8137abf85be598972259.jpg", + "image_caption": [ + "(a) Overall framework of previous MDRG system" + ], + "image_footnote": [], + "bbox": [ + 222, + 147, + 367, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e8d04309ebe22b6b530bc8e25fc1dcc85d57332f800708fbe0baebf1eaa9a2b2.jpg", + "image_caption": [ + "(b) Loss of Image History Information when generating text and image response." + ], + "image_footnote": [], + "bbox": [ + 377, + 145, + 535, + 267 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1cd0ceb2f7bea0c731501004a2fdcd1fbcf2235fa4a0b5fdff99546ff8594b83.jpg", + "image_caption": [ + "Fig. 1: (a) Outlines the framework of previous Multimodal Dialogue Response Generation (MDRG) systems, which uses the textual descriptions of images $(u_{t})$ as an intermediary step toward generating image responses $(r_{t}^{\\mathrm{Image}})$ . (b) Highlights the limitations of these systems, particularly their failure to fully leverage image history $(r_{1:t-1}^{\\mathrm{Image}})$ in crafting both the textual response $(r_{t}^{\\mathrm{Text}})$ and the image response $(r_{t}^{\\mathrm{Image}})$ . (c) illustrates the consequences of this oversight, including responses that lack grounding in image context and consistency in image-based replies." + ], + "image_footnote": [], + "bbox": [ + 547, + 147, + 666, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d460dab270f987011460d6970a27df4d199b9a97f88de4980c66ffcd8c390ce9.jpg", + "image_caption": [ + "(c) Problems of Previous MDRG system" + ], + "image_footnote": [], + "bbox": [ + 666, + 148, + 784, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gle sequence-to-sequence model requires a large number of training instances, the lack of a large-scale multimodal dialogue dataset poses a significant challenge for the task.", + "bbox": [ + 212, + 443, + 787, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve this issue, the previous MDRG model [43] disentangles the textual response and image response generation, as illustrated in Figure 1-(a). Specifically, given the dialogue history $D$ at $t$ -th conversation turn, the MDRG model first generates the textual response $r_t^{\\mathrm{Text}}$ and the intermediary textual image description $u_t$ . This image description $u_t$ is subsequently fed into a text-to-image model to generate the corresponding image $r_t^{\\mathrm{Image}}$ . Such an approach allows the utilization of powerful models pre-trained on vast amounts of available data for text-to-text and text-to-image pairs, bypassing the need for a large-scale multimodal dialogue dataset for direct end-to-end training.", + "bbox": [ + 212, + 494, + 787, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nonetheless, due to its reliance on text as an intermediary representation of images, the previous MDRG model overlooks crucial information about the image, as shown in Figure 1-(b). By converting images from the dialogue history into textual descriptions $(u_{1:t-1})$ , these models fail to fully utilize the rich visual content in the actual images $(r_{1:t-1}^{\\mathrm{Image}})$ . This leads to two major issues: a lack of image-grounded context in textual responses (Figure 1-(c) (left)) and inconsistencies in image generation across dialogues (Figure 1-(c) (right)).", + "bbox": [ + 212, + 636, + 787, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For instance, Figure 1-(c) (left) shows the existing MDRG model failing to provide an image-grounded response to 'What breed is your dog?' because it only perceives the dog through the text 'a dog eating a watermelon'. Figure 1-(c) (right) illustrates the model's inconsistency, where the 'dog' in the image history is not maintained in the generated response. These issues highlight the need for improved MDRG models that effectively utilize image history.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "HS. Yoon et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contribution This paper introduces BI-MDRG to enhance text and image responses by bridging the image history.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Bridging Image History in Text Responses: In Section 3.1, we propose an architectural modification in which the visual features, extracted using a visual encoder, are integrated into the cross-attention layers of the core language model. This is complemented by a novel multimodal causal attention mask modulation tailored for the MDRG task to improve image-grounded textual responses.", + "bbox": [ + 225, + 181, + 787, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Bridging Image History in Image Responses: In Section 3.2, we propose a citation framework where we use the Citation Module designed to augment textual image descriptions by tagging objects with citation tags that identify recurring objects throughout a dialogue. In Section 3.3, we show that training with these augmented data enables the model to recognize and maintain the consistency of objects in subsequent image responses during inference using Customized Text-to-Image Models. Due to the absence of benchmark datasets for evaluating image consistency in multimodal dialogues, we have created the Multimodal Dialogue Image Consistency (MDIC) dataset containing dialogues annotated to track object consistency across conversations.", + "bbox": [ + 227, + 272, + 787, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 441, + 387, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Multimodal Dialogue Datasets", + "text_level": 1, + "bbox": [ + 215, + 470, + 511, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal dialogue datasets generally fall into three categories: question and answering [1,8] (where the task involves asking and answering questions about a specific image), in-scene [12,31,35,44,49] (where each dialogue turn corresponds to a scene from a movie or video), and conversation-based [9,20,21,27,42,47] (which engage in natural dialogue about a given image or involves image sharing within natural conversations). Further details are in Appendix C.1.", + "bbox": [ + 215, + 492, + 785, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper primarily explores the model's capability for natural dialogue within the conversation-based category. Notable datasets in this segment include ImageChat [42], PhotoChat [47], MMDD [20], DialogCC [21], MMChat [51], TikTalk [27], and MMDialog [9]. Given our focus on English-language scenarios, we exclude MMChat and Tiktalk from our evaluation, as it is a dataset primarily in Chinese. Additionally, DialogCC (not publicly available) and MMDD, which are synthesized by algorithmically pairing images with text-only dialogues for random turns, are also excluded from our analysis. Therefore, our evaluation is centered on ImageChat, PhotoChat, and MMDialog.", + "bbox": [ + 215, + 583, + 787, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ImageChat [42] consists of image-centered dialogues, where each dialogue is centered around a single given image. PhotoChat [47] features dialogues collected from social media, where a single image is shared in one of the conversation turns, which better mirrors everyday human interaction. Still, their limited scale and domain diversity restrict their applicability. Overcoming these limitations, MMDialog [9] features over a million diverse dialogues from social media, where multiple images are shared across numerous conversation turns, providing a more realistic representation of open-domain multimodal conversations.", + "bbox": [ + 215, + 719, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Briding Image History in Multimodal Dialogue Response Generation", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Multimodal Dialogue Modeling", + "text_level": 1, + "bbox": [ + 215, + 146, + 519, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pioneering studies [11, 32, 36] have delved into improving the performance of image-grounded conversational agents. While [14, 45, 46, 51] introduced a Seq2Seq based model focusing on multimodal dialogues, it primarily generated textual responses, not fully embracing the multimodal response scenario. In a notable advancement, [43] presented Divter, which not only produces informative text but also generates high-resolution images, marking a significant leap forward in multimodal dialogue response generation (MDRG). It is important to note that our focus in this paper is generation-based models, contrasting with retrieval-based [3, 16, 22], which output image responses by retrieving existing images from a corpus instead of generating new ones.", + "bbox": [ + 212, + 169, + 787, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Customized Text-to-Image", + "text_level": 1, + "bbox": [ + 215, + 340, + 480, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent studies on text-to-image diffusion models [5, 17, 38, 40] focus on customization [7, 10, 18, 39], learning specific concepts from a few images. Following this, users can flexibly generate the learned concepts into new scenes. Text inversion [10] generates varied contexts for a concept by updating the text embedding without altering the model. Dreambooth [39] and Custom Diffusion [18] fine-tune the U-Net architecture using an identifier, class label, and images. In a notable enhancement, BLIP-Diffusion [23] enables zero-shot subject-driven generation, allowing fast customized text-to-image generation.", + "bbox": [ + 212, + 364, + 787, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 BI-MDRG: Bridging the Image History in Multimodal Dialogue Response Generation", + "text_level": 1, + "bbox": [ + 215, + 506, + 782, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce BI-MDRG, a conversational agent designed to produce both textual and visual responses with enhanced awareness of image history. Sections 3.1 and 3.2 detail the training procedure, effectively integrating image history information into text responses and textual image descriptions. Section 3.3 outlines the inference process, wherein the image history informs the image responses by leveraging the captured details from enhanced textual image descriptions.", + "bbox": [ + 212, + 554, + 787, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Bridging the Image History for Image-Grounded Text Response", + "text_level": 1, + "bbox": [ + 215, + 665, + 785, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The example shown in Figure 1-(c) (left) highlighted one of the crucial limitations in the previous MDRG system: their reliance on textual descriptions for understanding image history, which hinders the image-grounded textual responses. To overcome this, we have adopted an architectural change along with a multimodal causal attention mask modulation to effectively bridge the image history information $r_{1:t-1}^{\\mathrm{Image}}$ to the text response $r_t^{\\mathrm{Text}}$ (as depicted in Figure 2).", + "bbox": [ + 212, + 688, + 573, + 840 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d67d544ddc906675d546ef7170d529ff8bc6fe5f82de2b5bb2592bc52056686b.jpg", + "image_caption": [ + "Fig.2: Bridging Image History to the Text Response." + ], + "image_footnote": [], + "bbox": [ + 611, + 688, + 763, + 806 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "HS. Yoon et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96ac44f7b984c60534151a2f0a80dfdaea3a98ce27b5b186fa9bc9761a69d60f.jpg", + "image_caption": [ + "Fig. 3: Training of BI-MDRG. (a) Textual Dialogue Response Generator $\\mathcal{G}$ cross-attends to the image features from the Visual Encoder $\\mathcal{V}$ . (b) Attention Mask Modulation alters the causal attention to prioritize image features over textual image descriptions. (c) Citation Module $\\mathcal{C}$ generates citation-augmented textual image descriptions, enabling the tracking of objects within image history for consistency maintenance." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 756, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture Our Textual Dialogue Response Generator $\\mathcal{G}$ (Figure 3-(a)), consists of a decoder-only language model with added visual cross-attention layers. These layers directly engage with image features provided by the Visual Encoder $\\mathcal{V}$ , drawing inspiration from Flamingo [2], to reduce dependence on textual image descriptions for perceiving images. In our framework, a dialogue context $D = \\{(r_i^{\\mathrm{Text}}, r_i^{\\mathrm{Image}})\\}_{i=1}^n$ comprises multiple turns, each with an associated text response $r_i^{\\mathrm{Text}}$ and an image response $r_i^{\\mathrm{Image}}$ . An image captioning model produces textual descriptions $u_i$ for each image $r_i^{\\mathrm{Image}}$ , which are then transformed into citation-augmented descriptions $u_i'$ as shown in Figure 3-(c) and further detailed in Section 3.2 (if there is no image response for a turn, $r_i^{\\mathrm{Image}} = \\emptyset$ and $u_i' = \\emptyset$ ). These descriptions and the text responses $\\{r_1^{\\mathrm{Text}}, u_1', \\ldots, r_n^{\\mathrm{Text}}, u_n'\\}$ are fed into $\\mathcal{G}$ , while the images $\\{r_1^{\\mathrm{Image}}, \\ldots, r_n^{\\mathrm{Image}}\\}$ are processed by $\\mathcal{V}$ to extract image features which are fed to the cross-attention layers in $\\mathcal{G}$ . Although our model directly cross-attends to the inputted image features, we retain the textual descriptions $u_{1:n}'$ as essential inputs since we require the generation of textual description by our $\\mathcal{G}$ model, which subsequently gets used by a text-to-image model for constructing the image response. Keeping the textual description inputs allows efficient teacher-forced next token prediction training.", + "bbox": [ + 212, + 426, + 787, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Causal Attention Mask Modulation For the input sequence $\\{r_1^{\\mathrm{Text}}, u_1', \\ldots, r_n^{\\mathrm{Text}}, u_n'\\}$ , we use a specialized mask along with the standard causal mask (Figure 3-(b)). The traditional causal mask allows each text response $r_i^{\\mathrm{Text}}$ to access previous textual image descriptions $u_{1:i-1}'$ , leading to reliance on textual information over visual context. Our masking strategy prevents $r_i^{\\mathrm{Text}}$ from accessing $u_{1:i-1}'$ , redirecting focus to the actual image features of $r_{1:i-1}^{\\mathrm{Image}}$ , ensuring text responses are grounded on raw image features.", + "bbox": [ + 212, + 732, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Briding Image History in Multimodal Dialogue Response Generation", + "bbox": [ + 274, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c950ce3eb160bc479415c7845d7b2d0efd442bf2637b77a13e9117f7781fe267.jpg", + "image_caption": [ + "Fig. 4: Illustration of the Citation Module. Citation Module recognizes identical objects within image history and injects this information into the textual image description with citation tags (e.g., [cite]0[/cite])." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 781, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Citation Module: Bridging the Image History to the Textual Image Description", + "text_level": 1, + "bbox": [ + 214, + 297, + 764, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The example shown in Figure 1-(c) (right) highlighted another critical limitation of the previous MDRG system: its inability to ensure consistency in image responses. To address this, we propose the Citation Module (Figure 3-(c)) to bridge the image history $r_{1:t-1}^{\\mathrm{Image}}$ and the textual image description $u_t$ (as depicted in Figure 5) by ensuring that the textual image description accurately relays which objects should persist in subsequent images.", + "bbox": [ + 212, + 335, + 550, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd64c838406fde07f3b821a888ccf254647639771228009cc870e00cae4adecb.jpg", + "image_caption": [ + "Fig. 5: Bridging Image History to the Textual Image Description." + ], + "image_footnote": [], + "bbox": [ + 598, + 316, + 751, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Citation Module Citation Module plays a pivotal role in tracking recurring objects in the dialogue using textual image descriptions. For instance, descriptions like \"a dog is in front of a fireplace\" and \"a dog running in the snow\" are augmented to \"a dog[cite]0[/cite] is in front of a fireplace\" and \"a dog[cite]0[/cite] running in the snow,\" respectively if they reference the same dog.", + "bbox": [ + 212, + 507, + 784, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motivated by [30], Figure 4 details the citation process for textual image descriptions $\\{u_1, \\dots, u_n\\}$ . For each textual image description $u_i$ , a Part of Speech (POS) tagging processor $\\mathcal{P}$ is employed to tag words and pinpoint the key object word $o_i$ in the description. The word $o_i$ , along with its corresponding image $r_i^{\\mathrm{Image}}$ , is processed through an open-set object detector $\\mathcal{M}$ to obtain the bounding box of the detected object, which is then input to a segmentation model $\\mathcal{S}$ for generating object segmentation mask $s_i$ . These masks are applied to isolate the objects from their backgrounds in $r_i^{\\mathrm{Image}}$ since background removal has been proven helpful for better extraction of object features [6]. These isolated objects are then analyzed by a visual feature extractor $\\mathcal{E}$ to extract features $f_i$ . The resulting feature set $\\{f_1, \\dots, f_n\\}$ undergoes clustering based on cosine similarity to identify identical objects across images, as outlined in Algorithm 1; this involves assigning each element a cluster id $c_i$ based on the similarity of their features. For each $o_i$ in $u_i$ , we augment the word so that it is followed by its corresponding cluster id (i.e., citation tag) $c_i$ of $f_i$ , resulting in the citation augmented textual description $u_i'$ . This Citation Module operates with off-the-shelf components for citation tags, requiring no training on the target dataset.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "HS. Yoon et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Citation Module" + ], + "code_body": "Require: image response $(r_1^{\\mathrm{Image}},\\dots r_n^{\\mathrm{Image}})$ , textual image description $(u_{1},\\ldots u_{n})$ , similarity threshold $\\tau$ , POS tagging processor $\\mathcal{P}$ , open-set object detector $\\mathcal{M}$ , segmentation model $S$ , visual feature extractor $\\mathcal{E}$ , object cluster dictionary $\\kappa = \\{\\}$ \nfor $i\\gets 1$ to $n$ do $\\triangleright$ Identify principal object word in $u_{i}$ $o_i\\gets \\mathcal{P}(u_i)$ $\\triangleright$ Obtain bounding box for $o_i$ $box_{i}\\gets \\mathcal{M}(o_{i},r_{i}^{\\mathrm{Image}})$ $\\triangleright$ Obtain segmentation mask for $o_i$ $s_i\\gets S(box_i)$ $\\triangleright$ Extract visual feature for $o_i$ $f_{i}\\gets \\mathcal{V}(s_{i}\\odot r_{i}^{\\mathrm{Image}})$ \nend for \ncluster_id $\\leftarrow 0$ \nfor $i\\gets 1$ to $n$ do Perform clustering of $o_i$ . if not $\\kappa$ has key(o) then $\\mathcal{K}[o_i]\\gets \\overline{\\mathsf{clust}} e r\\_ i d$ Get's own cluster. cluster_id $\\leftarrow$ cluster_id+1 \nend if \nfor $j\\gets i + 1$ to $n$ do \nif $\\operatorname {SIM}(f_i,f_j)\\geq \\tau$ and not $\\kappa$ has key(oj) then $\\kappa [o_j]\\gets \\kappa [o_i]$ Puts into existing cluster. end if \nend for \nend for \nreturn K", + "bbox": [ + 215, + 164, + 785, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generative Training Objective We use the next token prediction training via teacher forcing, which is used in standard auto-regressive language models [37]. Specifically, given the token sequence $w = \\{w_{j}\\}_{j=1}^{N}$ of the input sequence $\\{r_{1}^{\\mathrm{Text}}, u_{1}', \\dots, r_{n}^{\\mathrm{Text}}, u_{n}'\\}$ and the images $\\{r_{1}^{\\mathrm{Image}}, \\dots, r_{n}^{\\mathrm{Image}}\\}$ , we minimize the negative log-likelihood:", + "bbox": [ + 212, + 434, + 782, + 512 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (w) = \\sum_ {j = 1} ^ {N} \\log P \\left(w _ {j} \\mid w _ {< j}, \\left\\{r _ {i} ^ {\\text {I m a g e}} \\right\\} _ {i = 1} ^ {n}; \\mathcal {G}, \\mathcal {V}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 522, + 784, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With such training, our model can generate textual image descriptions during inference with citation tags that reflect the objects needing consistency.", + "bbox": [ + 215, + 574, + 784, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Inference Procedure: Bridging the Image History for Consistent Image Response", + "text_level": 1, + "bbox": [ + 215, + 625, + 782, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section outlines our inference procedure, which employs Customized Text-to-Image Model [23] in conjunction with citation-augmented textual image descriptions (Figure 7). This bridges the image history $r_{1:t-1}^{\\mathrm{Image}}$ and image response $r_t^{\\mathrm{Image}}$ (as depicted in Figure 6) allowing for consistent generation of the image response.", + "bbox": [ + 212, + 662, + 549, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference For an incoming dialogue context $D = \\{(r_i^{\\mathrm{Text}}, r_i^{\\mathrm{Image}})\\}_{i=1}^{t-1}$ , we initially construct", + "bbox": [ + 215, + 809, + 549, + 840 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43ac84ee766bcd4ed4b8efdde64500e5f29bbd01860ac7023f516fde318cb8f8.jpg", + "image_caption": [ + "Fig.6: Bridging the Image History to the Image Response." + ], + "image_footnote": [], + "bbox": [ + 586, + 652, + 766, + 790 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Briding Image History in Multimodal Dialogue Response Generation", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0104f449cfe2e2b9824cbfd435d51e66e118a4ff90939f235dabff6c85ef0488.jpg", + "image_caption": [ + "Fig. 7: Overall Inference Procedure. During inference, the Textual Response Generator $\\mathcal{G}$ generates the citation-augmented textual image description $u_{t}^{\\prime}$ . By utilizing its citation tag $c_{t}$ , object consistency can be maintained by feeding all preceding image responses with the identical citation tag into a Customized Text-to-Image Model $\\mathcal{F}$ ." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 784, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the sequence of textual image descriptions and corresponding text responses as $\\{r_1^{\\mathrm{Text}}, u_1, \\ldots, r_{t-1}^{\\mathrm{Text}}, u_{t-1}\\}$ . Our Citation Module $\\mathcal{C}$ then augments the textual image descriptions with citation tags to create $\\{r_1^{\\mathrm{Text}}, u_1', \\ldots, r_{t-1}^{\\mathrm{Text}}, u_{t-1}'\\}$ , which serves as the input for the Textual Response Generator $\\mathcal{G}$ . During inference, if $\\mathcal{G}$ predicts an [IMG] token, it initiates the generation of a citation-augmented textual image description $u_t'$ . We extract the citation $c_t$ and the core image description $u_t$ from $u_t'$ . Subsequently, the image description $u_t$ is input to our Customized Text-to-Image Model $\\mathcal{F}$ , along with all preceding image responses sharing the same citation $c_t$ as follows:", + "bbox": [ + 212, + 385, + 787, + 521 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} ^ {\\mathrm {I m a g e}} = \\mathcal {F} \\left(u _ {t} \\mid \\{r _ {i} ^ {\\mathrm {I m a g e}} \\mid c _ {i} = c _ {t} \\} _ {i = 1} ^ {t - 1}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 531, + 785, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This approach ensures the consistent generation of specific objects across the conversation.", + "bbox": [ + 212, + 566, + 787, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 622, + 375, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section presents the implementation details, evaluation benchmarks, and the experimental results of our approach. Section 4.1 outlines the implementation specifics. In Section 4.2, we assess the overall quality of our model against the standard benchmarks established in previous works [9, 47]. Section 4.3 is dedicated to evaluating the image grounding effectiveness of our model. Finally, Section 4.4 examines the consistency of the image responses generated by our system.", + "bbox": [ + 212, + 654, + 787, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 782, + 428, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset As mentioned in Section 2.1, our benchmark datasets include ImageChat [42], PhotoChat [47], and MMDialog [9]. Specifically, for the overall", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "HS. Yoon et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "multimodal dialogue evaluation in Section 4.2, we train and evaluate using the PhotoChat and MMDialog. For the image grounding evaluation in Section 4.3, we use ImageChat. For the image consistency evaluation in Section 4.4, we create the MDIC dataset by hand-labeling a subset of MMDialog test set. For Section 4.3 and Section 4.4, we use the model trained on MMDialog from Section 4.2.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details For the Textual Dialogue Response Generator $\\mathcal{G}$ and the Visual Encoder $\\mathcal{V}$ , we initialize with the pre-trained OpenFlamingo 4B model $^3$ [4]. During the fine-tuning phase, we employ special tokens to structure our inputs and outputs: [IMG] and [/IMG] encapsulate textual image descriptions, while [EOT] signifies the end of conversation turns. The BLIP2-flan-t5-xl model [24] is used for converting image responses to corresponding textual image descriptions. Additionally, our Citation Module $\\mathcal{C}$ uses [CITE] and [/CITE] tokens to mark the beginning and end of citations linked to key objects within the textual image descriptions.", + "bbox": [ + 212, + 244, + 787, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As noted in Section 3.2, the Citation Module $\\mathcal{C}$ is composed of four key components: the POS tagging processor $(\\mathcal{P})$ , the open-set object detector $(\\mathcal{M})$ , the segmentation model $(\\mathcal{S})$ , and the visual feature extractor $(\\mathcal{E})$ . Specifically, we employ spaCy [13] for $\\mathcal{P}$ , GroundingDino [28] for $\\mathcal{M}$ , the Segment Anything Model (SAM) [15] for $\\mathcal{S}$ , and DINOv2 [33] for $\\mathcal{E}$ . The similarity threshold $\\tau$ is set to 0.6. No further learning is done for these modules; they are utilized as pre-trained components within our system for citation tagging. For our customized text-to-image generation model, $\\mathcal{F}$ , we used BLIP-Diffusion [23] when conditioning on the input image and the standard Stable Diffusion 2.1 [38] when there is no input image conditioning.", + "bbox": [ + 212, + 381, + 787, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Learning Details Let us denote $\\{\\theta_{\\mathcal{V}},\\theta_{\\mathcal{G}_v},\\theta_{\\mathcal{G}_l}\\}$ as the parameters of the perceiver resampler of the Visual Encoder $\\nu$ , the visual cross-attention layers of the Textual Dialogue Response Generator $\\mathcal{G}$ , and the language model layers of $\\mathcal{G}$ , respectively.", + "bbox": [ + 212, + 555, + 787, + 614 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the first stage of training, we train $\\theta_{\\mathcal{G}_l}$ . The batch size is set to 256 with a maximum token length set to 256. In the second stage of training, we jointly train $\\theta_{\\mathcal{V}}$ and $\\theta_{\\mathcal{G}_v}$ . The batch size is set to 128 with a maximum token length set to 512. Both the first and second stage is trained by minimizing the next token prediction loss (Eq. 1) using the AdamW optimizer [29] with a learning rate set to 1e-4. The trainings were conducted using $16 \\times$ NVIDIA A100 80GB PCIe.", + "bbox": [ + 212, + 616, + 787, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Multimodal Dialogue Evaluation", + "text_level": 1, + "bbox": [ + 214, + 729, + 532, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Dataset We evaluate the overall performance of our BI-MDRG system on the test set of PhotoChat [47] and MMDialog [9] dataset. PhotoChat contains a single image per dialogue, while the MMDialog includes dialogues with multiple images across turns, offering a more complex context for assessing", + "bbox": [ + 212, + 755, + 787, + 816 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Briding Image History in Multimodal Dialogue Response Generation", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "3 https://huggingface.co/openflamingo/OpenFlamingo-4B-vitl-rpj3b-langinstruct", + "bbox": [ + 217, + 824, + 761, + 840 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/da7d86fd2271b296f0ffcd36700de4eeb85d9a6677a98d112aa26958b3807e8e.jpg", + "table_caption": [ + "Table 1: Automatic evaluation results of BI-MDRG on the PhotoChat and the MM-Dialog test set. Numbers in bold represent the best scores." + ], + "table_footnote": [], + "table_body": "
ModelIntent F1Image Response ISTextual B1Image Description B2R-1R-LText Response
B1B2R1R-L
PhotoChat Dataset [47]
BERT-base [47]53.2---------
T5-3B [47]58.9---------
Divter [43]56.215.815.111.4-15.86.521.66-5.69
DivterLLM54.116.141.327.143.341.611.44.7511.210.8
BI-MDRG55.716.742.128.244.642.512.45.1212.111.2
MMDialog Dataset [9]
PaCE [25]77.6---------
Divter [9]71.820.5----9.447.45-11.2
MiniGPT5 [50]-20.2----29.119.5-12.1
DivterLLM67.321.044.235.745.543.621.316.220.419.4
BI-MDRG70.522.452.244.753.251.627.623.525.724.8
", + "bbox": [ + 235, + 184, + 767, + 369 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "our system's capability. We perform the tests for all turns except the first turn of each dialogue and consider all previous turns as context.", + "bbox": [ + 212, + 398, + 782, + 429 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metric The performance evaluation is carried out using automatic metrics across four key dimensions: (1) Image Intent Prediction - assessing the need for an image response in the current turn; (2) Image Response Quality; (3) Textual Image Description Quality; and (4) Text Response Generation. We employ the F1 metric for Image Intent Prediction, following the binary classification approach detailed in [9, 43]. BLEU [34] and ROUGE [26] metrics are used for evaluating both the Textual Image Descriptions and Text Responses. The Image Response Quality is measured using the Inception Score (IS) [41], in line with [9].", + "bbox": [ + 212, + 454, + 787, + 592 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines The BI-MDRG system is compared against Divter [43], a prior Multimodal Dialogue Response Generator that utilizes DialoGPT (762M) [48] as its language model backbone. For a fair comparison, we also train Divter on the same backbone Language Model as our BI-MDRG, denoted as $\\mathrm{Divter}_{\\mathrm{LLM}}$ (3B). Moreover, for the MMDialog dataset, we compare with MiniGPT5 [50] which uses a 9B backbone VLM. For Intent-prediction baselines, we also include BERT-base, T5-3B reported in [47] for the PhotoChat dataset and PaCE reported in [25] for the MMDialog dataset.", + "bbox": [ + 212, + 616, + 787, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results Table 1 presents the evaluation results, where our BI-MDRG system demonstrates outstanding performance on both datasets. In Textual Image Description and Text Response Generation, BI-MDRG achieves the highest BLEU and ROUGE metrics scores in most cases, indicating the system's proficiency in generating relevant and coherent image responses and text responses. Notably,", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "HS. Yoon et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "BI-MDRG, which uses a 4B backbone VLM, outperforms MiniGPT5, which utilizes a 9B backbone VLM.", + "bbox": [ + 214, + 145, + 782, + 176 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Image Grounding Evaluation", + "text_level": 1, + "bbox": [ + 215, + 196, + 501, + 213 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Dataset To evaluate the ability to generate image-grounded textual responses, we utilize the ImageChat dataset [42], consisting of dialogues centered around a single image (an example is shown in Appendix C.4). Since the conversations are grounded in an image, the dataset is suitable for the evaluation of the image grounding capability of the model. We use the same model trained on the MMDialog dataset in Section 4.2 without further tuning on the ImageChat train set.", + "bbox": [ + 212, + 222, + 787, + 328 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metric In order to evaluate the image grounding capability, we use BLEU and ROUGE as the metrics. We perform the evaluation only on the last turn of each dialogue and consider all previous turns as the input context.", + "bbox": [ + 214, + 349, + 787, + 396 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results The performance of our BI-MDRG system on the ImageChat benchmark is summarized in Table 2. These results demonstrate the system's superior image grounding capabilities in text responses. BI-MDRG achieves the highest scores in BLEU-1, ROUGE-1, and ROUGE-L metrics, indicating its effectiveness in generating contextually relevant and coherent text responses grounded in the visual elements of the conversation. In comparison to DivterLLM", + "bbox": [ + 212, + 431, + 516, + 611 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "and BI-MDRG $_{\\mathrm{w/o mask}}$ , our BI-MDRG model shows an improvement, highlighting the significance of our proposed architectural enhancements and the multimodal causal attention mask in understanding and integrating visual context into text responses effectively.", + "bbox": [ + 212, + 612, + 784, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moreover, we show an example of the model prediction between the Divter $_{\\text{LLM}}$ and our BI-MDRG in Figure 8-(a). Figure 8-(a) (right-top) shows that Divter $_{\\text{LLM}}$ produces contextually irrelevant text response due to its lack of image understanding. However, Figure 8-(a) (right-bottom) shows that since our BI-MDRG has access to the image, it can produce a text response grounded on the image.", + "bbox": [ + 212, + 672, + 787, + 750 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1326e66b9cf5f1c8706394f003ed7c2554829cb22c4fb0a64e50f2ecf2724bf4.jpg", + "table_caption": [ + "Table 2: Automatic evaluation results on the ImageChat test set. Numbers in bold represent the best scores." + ], + "table_footnote": [], + "table_body": "
ModelB1R-1R-L
ImageChat Dataset [42]
DivterLLM8.610.39.6
BI-MDRGw/o mask10.011.110.2
BI-MDRG10.911.710.9
", + "bbox": [ + 531, + 494, + 772, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Image Response Consistency Evaluation", + "text_level": 1, + "bbox": [ + 214, + 770, + 591, + 786 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We integrated the Citation Module into our BI-MDRG system to improve consistency in image responses within the dialogue. This module assigns citation tags to primary objects in textual image descriptions, tracking their presence across", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Briding Image History in Multimodal Dialogue Response Generation", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1d8cf49b30fa7b47b5c0510d2738d84c156c7af343ac58c23cc6935f7f301a0d.jpg", + "image_caption": [ + "(a) Enhanced Image Grounded Text Response" + ], + "image_footnote": [], + "bbox": [ + 246, + 146, + 496, + 248 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/166b2d24bdda312aa2d9a8fec5129b8ce8e4b320934159d8740b6dfbb7102fbf.jpg", + "image_caption": [ + "(b) Enhanced Consistency of Image Response" + ], + "image_footnote": [], + "bbox": [ + 506, + 146, + 754, + 248 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/77093220b173eb862eb9eedb7e970696f8f95cc63fbe9570bf8637e52aea3de8.jpg", + "image_caption": [ + "Fig. 8: Example of predictions from Divter $_{\\text{LLM}}$ [43] vs. BI-MDRG (ours). Additional examples can be found in Appendix E.", + "(a) Citation Module Evaluation" + ], + "image_footnote": [], + "bbox": [ + 223, + 318, + 405, + 412 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5fe84f13a5cf58fb0d7d07f879f86630cc4838439cf799a0b7db173e12debde9.jpg", + "image_caption": [ + "(b) Citation Token Prediction Evaluation" + ], + "image_footnote": [], + "bbox": [ + 418, + 318, + 576, + 415 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6468590d985393f283360dc290a6ec3c00fb2a28cdb2211de3668c6b494ce6a6.jpg", + "image_caption": [ + "(c) Image Consistency Evaluation", + "Fig. 9: Three evaluation aspects for image response consistency. (a) We assess the accuracy of the pseudo-labels of citation tags created by the Citation Module. (b) We evaluate how well the model learns from these pseudo-labels and predict the citation tag $\\hat{c}_t$ , given the dialogue history, current text response $r_t^{\\mathrm{Text}}$ , and $\\tilde{u}_t$ which denotes the textual image description leading up to the primary object word $o_t$ . (c) We evaluate the consistency of the resulting image response $\\hat{r}_t^{\\mathrm{Image}}$ ." + ], + "image_footnote": [], + "bbox": [ + 588, + 319, + 785, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the dialogue. As shown in Figure 9, we focus on three aspects for evaluating the image response consistency: (a) Citation Module Evaluation, (b) Citation Token Prediction Evaluation, and (c) Image Consistency Evaluation.", + "bbox": [ + 214, + 555, + 787, + 602 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Multimodal Dialogue Image Consistency (MDIC) Dataset Creation Due to the absence of benchmark datasets for evaluating the image consistency in multimodal dialogue scenarios, we created a dataset of 300 dialogues, each manually annotated to assign correct citation tags to objects based on their appearances in image responses. We use this created dataset for the subsequent evaluations. Details of the annotation process can be found in Appendix C.5.", + "bbox": [ + 214, + 621, + 787, + 713 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Citation Module Evaluation Figure 9-(a) shows the evaluation method for the Citation Module. As described in Section 3.2, the module $\\mathcal{C}$ processes visual features $\\{f_1,\\ldots ,f_n\\}$ corresponding to primary objects $\\{o_1,\\dots,o_n\\}$ in image descriptions. The module generates citations based on the cosine similarity between these features. We constructed a similarity matrix $\\mathcal{X}$ , where each element $\\mathcal{X}_{ij}$ represents the cosine similarity between the features $f_{i}$ and $f_{j}$ . A feature pair $(f_i,f_j)$ is deemed 'similar' if $\\mathcal{X}_{ij} \\geq \\tau$ and 'not similar' otherwise,", + "bbox": [ + 214, + 733, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "HS. Yoon et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where $\\tau$ is a predetermined threshold. The module's performance was quantified using the F1 score, and we achieved an F1 score of 0.72, demonstrating the module's high accuracy in assigning pseudo-labels of citation tags for training.", + "bbox": [ + 212, + 146, + 787, + 194 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Citation Token Prediction Evaluation Our model, trained with citation-augmented textual image descriptions, can generate these descriptions with accurate citation tags for the primary objects in the descriptions. Figure 9-(b) shows the evaluation method where we assess our model's ability to predict the citation tags correctly. We input the ground truth sequence $\\{r_1^{\\mathrm{Text}},u_1',\\ldots ,r_t^{\\mathrm{Text}},\\tilde{u}_t\\}$ from the MDIC dataset along with the image re", + "bbox": [ + 212, + 228, + 517, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "sponse history $\\{r_1^{\\mathrm{Image}},\\ldots ,r_{t - 1}^{\\mathrm{Image}}\\}$ . Here, $\\tilde{u}_t$ denotes the textual image description leading up to the principal object word $o_t$ . The goal is to verify whether the model's predicted citation $\\hat{c}_t$ aligns with the actual citation $c_t$ . The accuracy (Acc.) of our citation token prediction is presented in Table 3. We compare with a baseline denoted as Divter $_{\\mathrm{LLM}} + \\mathrm{LLMCite}$ where we assign citation tags by instructing an external LLM (Details in Appendix D). The results show that BI-MDRG achieves $84.0\\%$ accuracy compared to $33.5\\%$ of the LLMCite baseline, showing the effectiveness of our model in identifying objects needing consistency.", + "bbox": [ + 212, + 393, + 787, + 517 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0ab76e6c1b01e3e9d3ac53a9e3eacde4574a6339dd0fe146d5cb42c235c6e559.jpg", + "table_caption": [ + "Table 3: Citation Token Prediction (Acc.) and Image Consistency Evaluation (DINOv2). Numbers in bold represent the best scores." + ], + "table_footnote": [], + "table_body": "
ModelAcc.DINOv2(↑)
MDIC Dataset
DivterLLM + LLMCite33.50.32
BI-MDRG (ours)84.00.53
", + "bbox": [ + 531, + 305, + 774, + 373 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Image Consistency Evaluation Figure 9-(c) shows the evaluation method for the Image Consistency Evaluation. To assess the consistency of the image response, we employ metrics commonly found in the literature on custom object text-to-image generation [7, 10, 18, 19, 39]. DINO features outperform CLIP in capturing image alignment, specifically in distinguishing different objects of the same class [39]. Therefore, for our evaluation, we measure image alignment scores using the DINOv2 [33] features.", + "bbox": [ + 212, + 537, + 787, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In our proposed framework, correctly using the citation tag is essential to ensure image consistency in the conversation. This involves correctly predicting a new citation tag or referencing an existing tag in history. Therefore, our evaluation includes not only scenarios where the model's predicted citation tag aligns with the ground truth $(c_{t} = \\hat{c}_{t})$ and $c_{t}$ exists in previous dialogue history, but also instances where the model's predicted citation tag do not match the ground truth $(c_{t} \\neq \\hat{c}_{t})$ and either $\\hat{c}_{t}$ or $c_{t}$ is in previous dialogue history.", + "bbox": [ + 212, + 643, + 787, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We report the DINOv2 score as an average of both the successful and failed citation cases mentioned above, where successful prediction positively influences the score, and failed prediction reduces it, offering a comprehensive measure of the model's performance in image consistency. Table 3 shows that our method successfully maintains the consistency of the image response in a dialogue setting, with the DINOv2 score achieving 0.53 compared to 0.32 of baseline. Furthermore,", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Briding Image History in Multimodal Dialogue Response Generation", + "bbox": [ + 272, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/fe856db88f4fb1a43d18c5ede305f6ec65977e4bbc512a20817513628c91b156.jpg", + "table_caption": [ + "Table 4: Evaluation of the citation framework on various settings. The image consistency is measured using DINOv2 (↑) on MDIC dataset, and TID (Textual Image Description) and TR (Text Response) are measured on MMDialog [9] test set. Custom uses customized text-to-image generation with image conditioning, while Diffusion employs a standard text-to-image diffusion model without conditioning. The citation framework allows to selectively use between these for consistent image response. The first row represents the performance of BI-MDRG (ours)." + ], + "table_footnote": [], + "table_body": "
CitationVLM SizeDiffusionDINOv2TIDTR
B1B2R-1R-LB1B2R-1
Citation Module4BCustom+Diffusion0.5352.244.753.251.627.623.525.7
LLMCite x4BCustom+Diffusion0.3452.044.953.351.327.423.425.8
4BDiffusion0.2552.044.953.351.327.423.425.8
LLMCite x9BCustom+Diffusion0.3356.048.754.552.130.325.728.8
9BDiffusion0.2656.048.754.552.130.325.728.8
", + "bbox": [ + 217, + 247, + 784, + 351 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "in Figure 8-(b), we present a comparative example of enhanced image consistency of prediction from BI-MDRG compared to Divter $_{\\text{LLM}}$ .", + "bbox": [ + 212, + 369, + 782, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Importance of Citation Tags for Image Consistency", + "text_level": 1, + "bbox": [ + 214, + 425, + 750, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Due to powerful pre-trained models for text-to-text and text-to-image, adopting text as an intermediary for image responses is a practical solution for the MDRG [43] task. However, due to the inherent information loss of images during this process, achieving image consistency is infeasible without a targeted framework for consistency maintenance.", + "bbox": [ + 212, + 460, + 787, + 536 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 4 shows image consistency and dialogue response performance across various settings. Not using citations shows similar dialogue response performance (Intent, TID, TR) compared to LLMCite and our Citation Module. However, LLMCite improves image consistency (DINOv2) from 0.25 to 0.34 (4B) and 0.26 to 0.33 (9B). Our Citation Module further boosts this from 0.34 to 0.53, indicating its importance for image consistency without affecting dialogue response performance. Notably, scaling the model size (4B to 9B) improves textual response but fails to maintain image consistency without our citation framework, as also evident by ChatGPT's shortcomings in Appendix E.", + "bbox": [ + 212, + 537, + 787, + 674 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 699, + 359, + 715 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents BI-MDRG, a novel framework for Multimodal Dialogue Response Generation (MDRG) aimed at bridging the image history for enhanced text and image response. Our model's innovative use of image history to inform both text and image responses addresses fundamental limitations in previous methodologies, particularly in maintaining consistency in multimodal interactions. The effectiveness of BI-MDRG has been demonstrated through rigorous evaluations using multiple benchmark datasets and a custom-annotated dataset.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "HS. Yoon et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by a grant of the KAIST-KT joint research project through AI2X Lab., Tech Innovation Group, funded by KT (No. D23000019, Developing Visual and Language Capabilities for AI-Based Dialogue Systems), and by Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2021-0-01381, Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments).", + "bbox": [ + 212, + 174, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 301, + 321, + 316 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alamri, H., Cartillier, V., Das, A., Wang, J., Cherian, A., Essa, I., Batra, D., Marks, T.K., Hori, C., Anderson, P., Lee, S., Parikh, D.: Audio-visual scene-aware dialog. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2019)", + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=EbMuimAbPbs", + "3. Anonymous: Chatsearch: a dataset and a generative retrieval model for general conversational image retrieval (2023), https://openreview.net/forum?id=0unbjYPmbC", + "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)", + "5. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Zhang, Q., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., Karras, T., Liu, M.Y.: ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)", + "6. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023)", + "7. Cohen, N., Gal, R., Meirom, E.A., Chechik, G., Atzmon, Y.: \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In: European Conference on Computer Vision. pp. 558-577. Springer (2022)", + "8. Das, A., Kottur, S., Gupta, K., Singh, A., Yadav, D., Moura, J.M., Parikh, D., Batra, D.: Visual dialog. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 326-335 (2017)", + "9. Feng, J., Sun, Q., Xu, C., Zhao, P., Yang, Y., Tao, C., Zhao, D., Lin, Q.: MMDialog: A large-scale multi-turn dialogue dataset towards multi-modal open-domain conversation. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 7348-7363. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-long.405, https://aclanthology.org/2023.acl-long.405" + ], + "bbox": [ + 225, + 330, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Briding Image History in Multimodal Dialogue Response Generation", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "10. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618", + "11. Gan, Z., Cheng, Y., Kholy, A., Li, L., Liu, J., Gao, J.: Multi-step reasoning via recurrent dual attention for visual dialog. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 6463-6474. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1648, https://aclanthology.org/P19-1648", + "12. Han, S., Hessel, J., Dziri, N., Choi, Y., Yu, Y.: Champagne: Learning real-world conversation from large-scale web videos. arXiv preprint arXiv:2303.09713 (2023)", + "13. Honnibal, M., Montani, I.: spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing (2017), to appear", + "14. Kim, J., Yoon, S., Kim, D., Yoo, C.D.: Structured co-reference graph attention for video-grounded dialogue (2021), https://arxiv.org/abs/2103.13361", + "15. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)", + "16. Koh, J.Y., Salakhutdinov, R., Fried, D.: Grounding language models to images for multimodal inputs and outputs. ICML (2023)", + "17. Koo, G., Yoon, S., Yoo, C.D.: Wavelet-guided acceleration of text inversion in diffusion-based image editing. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4380-4384. IEEE (2024)", + "18. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion (2023)", + "19. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion (2023)", + "20. Lee, N., Shin, S., Choo, J., Choi, H.J., Myaeng, S.H.: Constructing multi-modal dialogue dataset by replacing text with semantically relevant images. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). pp. 897-906. Association for Computational Linguistics, Online (Aug 2021), https://aclanthology.org/2021.acl-short.113", + "21. Lee, Y.J., Ko, B., Kim, H.G., Choi, H.J.: Dialogcc: Large-scale multi-modal dialogue dataset. arXiv preprint arXiv:2212.04119 (2022)", + "22. Levy, M., Ben-Ari, R., Darshan, N., Lischinski, D.: Chatting makes perfect: Chat-based image retrieval. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 61437-61449. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/c1b3d1e2cf53bb28cabd801bd58b3521-Paper-Conference.pdf", + "23. Li, D., Li, J., Hoi, S.C.H.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing (2023)", + "24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "HS. Yoon et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "25. Li, Y., Hui, B., Yin, Z., Yang, M., Huang, F., Li, Y.: PaCE: Unified multi-modal dialogue pre-training with progressive and compositional experts. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 13402-13416. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-long.749, https://aclanthology.org/2023.acl-long.749", + "26. Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out. pp. 74-81. Association for Computational Linguistics, Barcelona, Spain (Jul 2004), https://aclanthology.org/W04-1013", + "27. Lin, H., Ruan, L., Xia, W., Liu, P., Wen, J., Xu, Y., Hu, D., Song, R., Zhao, W.X., Jin, Q., et al.: Tiktalk: A video-based dialogue dataset for multi-modal chitchat in real world. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1303-1313 (2023)", + "28. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations", + "30. Ma, J., Liang, J., Chen, C., Lu, H.: Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410 (2023)", + "31. Meng, Y., Wang, S., Han, Q., Sun, X., Wu, F., Yan, R., Li, J.: Openvidual: A large-scale, open-domain dialogue dataset with visual contexts. arXiv preprint arXiv:2012.15015 (2020)", + "32. Niu, Y., Zhang, H., Zhang, M., Zhang, J., Lu, Z., Wen, J.R.: Recursive visual attention in visual dialog. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6679-6688 (2019)", + "33. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)", + "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: A method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting on Association for Computational Linguistics. p. 311-318. ACL '02, Association for Computational Linguistics, USA (2002). https://doi.org/10.3115/1073083.1073135, https://doi.org/10.3115/1073083.1073135", + "35. Poria, S., Hazarika, D., Majumder, N., Naik, G., Cambria, E., Mihalcea, R.: MELD: A multimodal multi-party dataset for emotion recognition in conversations. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 527-536. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1050, https://aclanthology.org/P19-1050", + "36. Qi, J., Niu, Y., Huang, J., Zhang, H.: Two causal principles for improving visual dialog. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10860-10869 (2020)", + "37. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners (2019)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Briding Image History in Multimodal Dialogue Response Generation", + "bbox": [ + 274, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)", + "39. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)", + "40. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=08Yk-n512A1", + "41. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X., Chen, X.: Improved techniques for training gans. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 29. Curran Associates, Inc. (2016), https://proceedings.neurips.cc/paper_files/paper/2016/file/8a3363abe792db2d8761d6403605aab7-Paper.pdf", + "42. Shuster, K., Humeau, S., Bordes, A., Weston, J.: Image-chat: Engaging grounded conversations. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 2414-2429. Association for Computational Linguistics, Online (Jul 2020). https://doi.org/10.18653/v1/2020.acl-main.219, https://aclanthology.org/2020.acl-main.219", + "43. Sun, Q., Wang, Y., Xu, C., Zheng, K., Yang, Y., Hu, H., Xu, F., Zhang, J., Geng, X., Jiang, D.: Multimodal dialogue response generation. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2854-2866. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.204, https://aclanthology.org/2022.acl-long.204", + "44. Wang, S., Meng, Y., Li, X., Sun, X., Ouyang, R., Li, J.: Openvidual 2.0: A larger-scale, open-domain dialogue generation dataset with visual contexts. arXiv preprint arXiv:2109.12761 (2021)", + "45. Yoon, S., Kim, D., Yoon, E., Yoon, H.S., Kim, J., Yoo, C.D.: Hear: Hearing enhanced audio response for video-grounded dialogue. arXiv preprint arXiv:2312.09736 (2023)", + "46. Yoon, S., Yoon, E., Yoon, H.S., Kim, J., Yoo, C.: Information-theoretic text hallucination reduction for video-grounded dialogue. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 4182-4193. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.280, https://aclanthology.org/2022.emnlp-main.280", + "47. Zang, X., Liu, L., Wang, M., Song, Y., Zhang, H., Chen, J.: PhotoChat: A human-human dialogue dataset with photo sharing behavior for joint image-text modeling. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 6142-6152. Association for Computational Linguis" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "HS. Yoon et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.479, https://aclanthology.org/2021.acl-long.479", + "48. Zhang, Y., Sun, S., Galley, M., Chen, Y.C., Brockett, C., Gao, X., Gao, J., Liu, J., Dolan, B.: Dialogpt: Large-scale generative pre-training for conversational response generation. In: ACL, system demonstration (2020)", + "49. Zhao, J., Zhang, T., Hu, J., Liu, Y., Jin, Q., Wang, X., Li, H.: M3ED: Multi-modal multi-scene multi-label emotional dialogue database. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 5699-5710. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.391, https://aclanthology.org/2022.acl-long.391", + "50. Zheng, K., He, X., Wang, X.E.: Minigpt-5: Interleaved vision-and-language generation via generative vocens. arXiv preprint arXiv:2310.02239 (2023)", + "51. Zheng, Y., Chen, G., Liu, X., Sun, J.: MMChat: Multi-modal chat dataset on social media. In: Calzolari, N., Bechet, F., Blache, P., Choukri, K., Cieri, C., Declerck, T., Goggi, S., Isahara, H., Maegaard, B., Mariani, J., Mazo, H., Odijk, J., Piperidis, S. (eds.) Proceedings of the Thirteenth Language Resources and Evaluation Conference. pp. 5778-5786. European Language Resources Association, Marseille, France (Jun 2022), https://aclanthology.org/2022.lrec-1.621" + ], + "bbox": [ + 217, + 147, + 785, + 410 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Briding Image History in Multimodal Dialogue Response Generation", + "bbox": [ + 274, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_model.json b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..652e21755c8162e22f07d1919d0ccd233b576974 --- /dev/null +++ b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_model.json @@ -0,0 +1,2526 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.256, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "BI-MDRG: Bridging Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.786, + 0.245 + ], + "angle": 0, + "content": "Hee Suk Yoon\\(^{1\\star}\\), Eunseop Yoon\\(^{1\\star}\\), Joshua Tian Jin Tee\\(^{1\\star}\\), Kang Zhang\\(^{1}\\), Yu-Jung Heo\\(^{2}\\), Du-Seong Chang\\(^{2}\\), and Chang D. Yoo\\(^{1\\star\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.255, + 0.717, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Korea Advanced Institute of Science and Technology (KAIST)" + }, + { + "type": "text", + "bbox": [ + 0.438, + 0.269, + 0.564, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\) KT Corporation" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.284, + 0.72, + 0.297 + ], + "angle": 0, + "content": "{hskyoon, esyoon97, joshuateetj, zhangkang, cd_yoo}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.298, + 0.589, + 0.311 + ], + "angle": 0, + "content": "{yj.heo,dschang}@kt.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.342, + 0.744, + 0.621 + ], + "angle": 0, + "content": "Abstract. Multimodal Dialogue Response Generation (MDRG) is a recently proposed task where the model needs to generate responses in texts, images, or a blend of both based on the dialogue context. Due to the lack of a large-scale dataset specifically for this task and the benefits of leveraging powerful pre-trained models, previous work relies on the text modality as an intermediary step for both the image input and output of the model rather than adopting an end-to-end approach. However, this approach can overlook crucial information about the image, hindering 1) image-grounded text response and 2) consistency of objects in the image response. In this paper, we propose BI-MDRG that bridges the response generation path such that the image history information is utilized for enhanced relevance of text responses to the image content and the consistency of objects in sequential image responses. Through extensive experiments on the multimodal dialogue benchmark dataset, we show that BI-MDRG can effectively increase the quality of multimodal dialogue. Additionally, recognizing the gap in benchmark datasets for evaluating the image consistency in multimodal dialogue, we have created a curated set of 300 dialogues annotated to track object consistency across conversations. The code and the dataset is publicly available at https://github.com/hee-suk-yoon/BI-MDRG." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.632, + 0.741, + 0.661 + ], + "angle": 0, + "content": "Keywords: Multimodal Dialogue \\(\\cdot\\) Image Grounding \\(\\cdot\\) Image Consistency" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.684, + 0.377, + 0.7 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.806 + ], + "angle": 0, + "content": "With the development of instant messaging technology, visual modalities are increasingly used alongside text in online communication. To enhance user interaction with intelligent agents, a new task, Multimodal Dialogue Response Generation (MDRG) [43], has been proposed. This task requires models to generate both text and image responses based on dialogue history containing texts and images. Since learning an effective multimodal generation model with a sin" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.812, + 0.365, + 0.825 + ], + "angle": 0, + "content": "* Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.217, + 0.827, + 0.385, + 0.84 + ], + "angle": 0, + "content": "\\*\\* Corresponding Author" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.812, + 0.385, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "HS. Yoon et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.148, + 0.368, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.268, + 0.362, + 0.288 + ], + "angle": 0, + "content": "(a) Overall framework of previous MDRG system" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.146, + 0.536, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.269, + 0.555, + 0.289 + ], + "angle": 0, + "content": "(b) Loss of Image History Information when generating text and image response." + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.148, + 0.667, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.271, + 0.747, + 0.283 + ], + "angle": 0, + "content": "(c) Problems of Previous MDRG system" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.149, + 0.785, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.789, + 0.404 + ], + "angle": 0, + "content": "Fig. 1: (a) Outlines the framework of previous Multimodal Dialogue Response Generation (MDRG) systems, which uses the textual descriptions of images \\((u_{t})\\) as an intermediary step toward generating image responses \\((r_{t}^{\\mathrm{Image}})\\). (b) Highlights the limitations of these systems, particularly their failure to fully leverage image history \\((r_{1:t-1}^{\\mathrm{Image}})\\) in crafting both the textual response \\((r_{t}^{\\mathrm{Text}})\\) and the image response \\((r_{t}^{\\mathrm{Image}})\\). (c) illustrates the consequences of this oversight, including responses that lack grounding in image context and consistency in image-based replies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.489 + ], + "angle": 0, + "content": "gle sequence-to-sequence model requires a large number of training instances, the lack of a large-scale multimodal dialogue dataset poses a significant challenge for the task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.789, + 0.632 + ], + "angle": 0, + "content": "To solve this issue, the previous MDRG model [43] disentangles the textual response and image response generation, as illustrated in Figure 1-(a). Specifically, given the dialogue history \\( D \\) at \\( t \\)-th conversation turn, the MDRG model first generates the textual response \\( r_t^{\\mathrm{Text}} \\) and the intermediary textual image description \\( u_t \\). This image description \\( u_t \\) is subsequently fed into a text-to-image model to generate the corresponding image \\( r_t^{\\mathrm{Image}} \\). Such an approach allows the utilization of powerful models pre-trained on vast amounts of available data for text-to-text and text-to-image pairs, bypassing the need for a large-scale multimodal dialogue dataset for direct end-to-end training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.789, + 0.746 + ], + "angle": 0, + "content": "Nonetheless, due to its reliance on text as an intermediary representation of images, the previous MDRG model overlooks crucial information about the image, as shown in Figure 1-(b). By converting images from the dialogue history into textual descriptions \\((u_{1:t-1})\\), these models fail to fully utilize the rich visual content in the actual images \\((r_{1:t-1}^{\\mathrm{Image}})\\). This leads to two major issues: a lack of image-grounded context in textual responses (Figure 1-(c) (left)) and inconsistencies in image generation across dialogues (Figure 1-(c) (right))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.842 + ], + "angle": 0, + "content": "For instance, Figure 1-(c) (left) shows the existing MDRG model failing to provide an image-grounded response to 'What breed is your dog?' because it only perceives the dog through the text 'a dog eating a watermelon'. Figure 1-(c) (right) illustrates the model's inconsistency, where the 'dog' in the image history is not maintained in the generated response. These issues highlight the need for improved MDRG models that effectively utilize image history." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.274, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Briding Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "Contribution This paper introduces BI-MDRG to enhance text and image responses by bridging the image history." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.183, + 0.788, + 0.272 + ], + "angle": 0, + "content": "- Bridging Image History in Text Responses: In Section 3.1, we propose an architectural modification in which the visual features, extracted using a visual encoder, are integrated into the cross-attention layers of the core language model. This is complemented by a novel multimodal causal attention mask modulation tailored for the MDRG task to improve image-grounded textual responses." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.273, + 0.788, + 0.424 + ], + "angle": 0, + "content": "- Bridging Image History in Image Responses: In Section 3.2, we propose a citation framework where we use the Citation Module designed to augment textual image descriptions by tagging objects with citation tags that identify recurring objects throughout a dialogue. In Section 3.3, we show that training with these augmented data enables the model to recognize and maintain the consistency of objects in subsequent image responses during inference using Customized Text-to-Image Models. Due to the absence of benchmark datasets for evaluating image consistency in multimodal dialogues, we have created the Multimodal Dialogue Image Consistency (MDIC) dataset containing dialogues annotated to track object consistency across conversations." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.443, + 0.388, + 0.459 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.471, + 0.513, + 0.485 + ], + "angle": 0, + "content": "2.1 Multimodal Dialogue Datasets" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.493, + 0.787, + 0.583 + ], + "angle": 0, + "content": "Multimodal dialogue datasets generally fall into three categories: question and answering [1,8] (where the task involves asking and answering questions about a specific image), in-scene [12,31,35,44,49] (where each dialogue turn corresponds to a scene from a movie or video), and conversation-based [9,20,21,27,42,47] (which engage in natural dialogue about a given image or involves image sharing within natural conversations). Further details are in Appendix C.1." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.584, + 0.788, + 0.719 + ], + "angle": 0, + "content": "This paper primarily explores the model's capability for natural dialogue within the conversation-based category. Notable datasets in this segment include ImageChat [42], PhotoChat [47], MMDD [20], DialogCC [21], MMChat [51], TikTalk [27], and MMDialog [9]. Given our focus on English-language scenarios, we exclude MMChat and Tiktalk from our evaluation, as it is a dataset primarily in Chinese. Additionally, DialogCC (not publicly available) and MMDD, which are synthesized by algorithmically pairing images with text-only dialogues for random turns, are also excluded from our analysis. Therefore, our evaluation is centered on ImageChat, PhotoChat, and MMDialog." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "ImageChat [42] consists of image-centered dialogues, where each dialogue is centered around a single given image. PhotoChat [47] features dialogues collected from social media, where a single image is shared in one of the conversation turns, which better mirrors everyday human interaction. Still, their limited scale and domain diversity restrict their applicability. Overcoming these limitations, MMDialog [9] features over a million diverse dialogues from social media, where multiple images are shared across numerous conversation turns, providing a more realistic representation of open-domain multimodal conversations." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "HS. Yoon et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.521, + 0.163 + ], + "angle": 0, + "content": "2.2 Multimodal Dialogue Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.322 + ], + "angle": 0, + "content": "Pioneering studies [11, 32, 36] have delved into improving the performance of image-grounded conversational agents. While [14, 45, 46, 51] introduced a Seq2Seq based model focusing on multimodal dialogues, it primarily generated textual responses, not fully embracing the multimodal response scenario. In a notable advancement, [43] presented Divter, which not only produces informative text but also generates high-resolution images, marking a significant leap forward in multimodal dialogue response generation (MDRG). It is important to note that our focus in this paper is generation-based models, contrasting with retrieval-based [3, 16, 22], which output image responses by retrieving existing images from a corpus instead of generating new ones." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.342, + 0.482, + 0.358 + ], + "angle": 0, + "content": "2.3 Customized Text-to-Image" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.487 + ], + "angle": 0, + "content": "Recent studies on text-to-image diffusion models [5, 17, 38, 40] focus on customization [7, 10, 18, 39], learning specific concepts from a few images. Following this, users can flexibly generate the learned concepts into new scenes. Text inversion [10] generates varied contexts for a concept by updating the text embedding without altering the model. Dreambooth [39] and Custom Diffusion [18] fine-tune the U-Net architecture using an identifier, class label, and images. In a notable enhancement, BLIP-Diffusion [23] enables zero-shot subject-driven generation, allowing fast customized text-to-image generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.507, + 0.784, + 0.543 + ], + "angle": 0, + "content": "3 BI-MDRG: Bridging the Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.646 + ], + "angle": 0, + "content": "We introduce BI-MDRG, a conversational agent designed to produce both textual and visual responses with enhanced awareness of image history. Sections 3.1 and 3.2 detail the training procedure, effectively integrating image history information into text responses and textual image descriptions. Section 3.3 outlines the inference process, wherein the image history informs the image responses by leveraging the captured details from enhanced textual image descriptions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.666, + 0.787, + 0.683 + ], + "angle": 0, + "content": "3.1 Bridging the Image History for Image-Grounded Text Response" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.574, + 0.842 + ], + "angle": 0, + "content": "The example shown in Figure 1-(c) (left) highlighted one of the crucial limitations in the previous MDRG system: their reliance on textual descriptions for understanding image history, which hinders the image-grounded textual responses. To overcome this, we have adopted an architectural change along with a multimodal causal attention mask modulation to effectively bridge the image history information \\( r_{1:t-1}^{\\mathrm{Image}} \\) to the text response \\( r_t^{\\mathrm{Text}} \\) (as depicted in Figure 2)." + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.689, + 0.764, + 0.808 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.817, + 0.787, + 0.846 + ], + "angle": 0, + "content": "Fig.2: Bridging Image History to the Text Response." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.275, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Briding Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.145, + 0.758, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 3: Training of BI-MDRG. (a) Textual Dialogue Response Generator \\(\\mathcal{G}\\) cross-attends to the image features from the Visual Encoder \\(\\mathcal{V}\\). (b) Attention Mask Modulation alters the causal attention to prioritize image features over textual image descriptions. (c) Citation Module \\(\\mathcal{C}\\) generates citation-augmented textual image descriptions, enabling the tracking of objects within image history for consistency maintenance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.789, + 0.707 + ], + "angle": 0, + "content": "Architecture Our Textual Dialogue Response Generator \\(\\mathcal{G}\\) (Figure 3-(a)), consists of a decoder-only language model with added visual cross-attention layers. These layers directly engage with image features provided by the Visual Encoder \\(\\mathcal{V}\\), drawing inspiration from Flamingo [2], to reduce dependence on textual image descriptions for perceiving images. In our framework, a dialogue context \\(D = \\{(r_i^{\\mathrm{Text}}, r_i^{\\mathrm{Image}})\\}_{i=1}^n\\) comprises multiple turns, each with an associated text response \\(r_i^{\\mathrm{Text}}\\) and an image response \\(r_i^{\\mathrm{Image}}\\). An image captioning model produces textual descriptions \\(u_i\\) for each image \\(r_i^{\\mathrm{Image}}\\), which are then transformed into citation-augmented descriptions \\(u_i'\\) as shown in Figure 3-(c) and further detailed in Section 3.2 (if there is no image response for a turn, \\(r_i^{\\mathrm{Image}} = \\emptyset\\) and \\(u_i' = \\emptyset\\)). These descriptions and the text responses \\(\\{r_1^{\\mathrm{Text}}, u_1', \\ldots, r_n^{\\mathrm{Text}}, u_n'\\}\\) are fed into \\(\\mathcal{G}\\), while the images \\(\\{r_1^{\\mathrm{Image}}, \\ldots, r_n^{\\mathrm{Image}}\\}\\) are processed by \\(\\mathcal{V}\\) to extract image features which are fed to the cross-attention layers in \\(\\mathcal{G}\\). Although our model directly cross-attends to the inputted image features, we retain the textual descriptions \\(u_{1:n}'\\) as essential inputs since we require the generation of textual description by our \\(\\mathcal{G}\\) model, which subsequently gets used by a text-to-image model for constructing the image response. Keeping the textual description inputs allows efficient teacher-forced next token prediction training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Multimodal Causal Attention Mask Modulation For the input sequence \\(\\{r_1^{\\mathrm{Text}}, u_1', \\ldots, r_n^{\\mathrm{Text}}, u_n'\\}\\), we use a specialized mask along with the standard causal mask (Figure 3-(b)). The traditional causal mask allows each text response \\(r_i^{\\mathrm{Text}}\\) to access previous textual image descriptions \\(u_{1:i-1}'\\), leading to reliance on textual information over visual context. Our masking strategy prevents \\(r_i^{\\mathrm{Text}}\\) from accessing \\(u_{1:i-1}'\\), redirecting focus to the actual image features of \\(r_{1:i-1}^{\\mathrm{Image}}\\), ensuring text responses are grounded on raw image features." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "HS. Yoon et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.782, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.273 + ], + "angle": 0, + "content": "Fig. 4: Illustration of the Citation Module. Citation Module recognizes identical objects within image history and injects this information into the textual image description with citation tags (e.g., [cite]0[/cite])." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.299, + 0.766, + 0.33 + ], + "angle": 0, + "content": "3.2 Citation Module: Bridging the Image History to the Textual Image Description" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.551, + 0.489 + ], + "angle": 0, + "content": "The example shown in Figure 1-(c) (right) highlighted another critical limitation of the previous MDRG system: its inability to ensure consistency in image responses. To address this, we propose the Citation Module (Figure 3-(c)) to bridge the image history \\( r_{1:t-1}^{\\mathrm{Image}} \\) and the textual image description \\( u_t \\) (as depicted in Figure 5) by ensuring that the textual image description accurately relays which objects should persist in subsequent images." + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.318, + 0.752, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.444, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Fig. 5: Bridging Image History to the Textual Image Description." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.584 + ], + "angle": 0, + "content": "Citation Module Citation Module plays a pivotal role in tracking recurring objects in the dialogue using textual image descriptions. For instance, descriptions like \"a dog is in front of a fireplace\" and \"a dog running in the snow\" are augmented to \"a dog[cite]0[/cite] is in front of a fireplace\" and \"a dog[cite]0[/cite] running in the snow,\" respectively if they reference the same dog." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Motivated by [30], Figure 4 details the citation process for textual image descriptions \\(\\{u_1, \\dots, u_n\\}\\). For each textual image description \\(u_i\\), a Part of Speech (POS) tagging processor \\(\\mathcal{P}\\) is employed to tag words and pinpoint the key object word \\(o_i\\) in the description. The word \\(o_i\\), along with its corresponding image \\(r_i^{\\mathrm{Image}}\\), is processed through an open-set object detector \\(\\mathcal{M}\\) to obtain the bounding box of the detected object, which is then input to a segmentation model \\(\\mathcal{S}\\) for generating object segmentation mask \\(s_i\\). These masks are applied to isolate the objects from their backgrounds in \\(r_i^{\\mathrm{Image}}\\) since background removal has been proven helpful for better extraction of object features [6]. These isolated objects are then analyzed by a visual feature extractor \\(\\mathcal{E}\\) to extract features \\(f_i\\). The resulting feature set \\(\\{f_1, \\dots, f_n\\}\\) undergoes clustering based on cosine similarity to identify identical objects across images, as outlined in Algorithm 1; this involves assigning each element a cluster id \\(c_i\\) based on the similarity of their features. For each \\(o_i\\) in \\(u_i\\), we augment the word so that it is followed by its corresponding cluster id (i.e., citation tag) \\(c_i\\) of \\(f_i\\), resulting in the citation augmented textual description \\(u_i'\\). This Citation Module operates with off-the-shelf components for citation tags, requiring no training on the target dataset." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.274, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Briding Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.445, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Citation Module" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.787, + 0.407 + ], + "angle": 0, + "content": "Require: image response \\((r_1^{\\mathrm{Image}},\\dots r_n^{\\mathrm{Image}})\\) , textual image description \\((u_{1},\\ldots u_{n})\\) , similarity threshold \\(\\tau\\) , POS tagging processor \\(\\mathcal{P}\\) , open-set object detector \\(\\mathcal{M}\\) , segmentation model \\(S\\) , visual feature extractor \\(\\mathcal{E}\\) , object cluster dictionary \\(\\kappa = \\{\\}\\) \nfor \\(i\\gets 1\\) to \\(n\\) do \\(\\triangleright\\) Identify principal object word in \\(u_{i}\\) \n\\(o_i\\gets \\mathcal{P}(u_i)\\) \\(\\triangleright\\) Obtain bounding box for \\(o_i\\) \n\\(box_{i}\\gets \\mathcal{M}(o_{i},r_{i}^{\\mathrm{Image}})\\) \\(\\triangleright\\) Obtain segmentation mask for \\(o_i\\) \n\\(s_i\\gets S(box_i)\\) \\(\\triangleright\\) Extract visual feature for \\(o_i\\) \n\\(f_{i}\\gets \\mathcal{V}(s_{i}\\odot r_{i}^{\\mathrm{Image}})\\) \nend for \ncluster_id \\(\\leftarrow 0\\) \nfor \\(i\\gets 1\\) to \\(n\\) do Perform clustering of \\(o_i\\) . if not \\(\\kappa\\) has key(o) then \\(\\mathcal{K}[o_i]\\gets \\overline{\\mathsf{clust}} e r\\_ i d\\) Get's own cluster. cluster_id \\(\\leftarrow\\) cluster_id+1 \nend if \nfor \\(j\\gets i + 1\\) to \\(n\\) do \nif \\(\\operatorname {SIM}(f_i,f_j)\\geq \\tau\\) and not \\(\\kappa\\) has key(oj) then \\(\\kappa [o_j]\\gets \\kappa [o_i]\\) Puts into existing cluster. end if \nend for \nend for \nreturn K" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.784, + 0.513 + ], + "angle": 0, + "content": "Generative Training Objective We use the next token prediction training via teacher forcing, which is used in standard auto-regressive language models [37]. Specifically, given the token sequence \\( w = \\{w_{j}\\}_{j=1}^{N} \\) of the input sequence \\( \\{r_{1}^{\\mathrm{Text}}, u_{1}', \\dots, r_{n}^{\\mathrm{Text}}, u_{n}'\\} \\) and the images \\( \\{r_{1}^{\\mathrm{Image}}, \\dots, r_{n}^{\\mathrm{Image}}\\} \\), we minimize the negative log-likelihood:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.523, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (w) = \\sum_ {j = 1} ^ {N} \\log P \\left(w _ {j} \\mid w _ {< j}, \\left\\{r _ {i} ^ {\\text {I m a g e}} \\right\\} _ {i = 1} ^ {n}; \\mathcal {G}, \\mathcal {V}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.575, + 0.785, + 0.606 + ], + "angle": 0, + "content": "With such training, our model can generate textual image descriptions during inference with citation tags that reflect the objects needing consistency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.784, + 0.657 + ], + "angle": 0, + "content": "3.3 Inference Procedure: Bridging the Image History for Consistent Image Response" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.664, + 0.55, + 0.789 + ], + "angle": 0, + "content": "This section outlines our inference procedure, which employs Customized Text-to-Image Model [23] in conjunction with citation-augmented textual image descriptions (Figure 7). This bridges the image history \\( r_{1:t-1}^{\\mathrm{Image}} \\) and image response \\( r_t^{\\mathrm{Image}} \\) (as depicted in Figure 6) allowing for consistent generation of the image response." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.55, + 0.842 + ], + "angle": 0, + "content": "Inference For an incoming dialogue context \\(D = \\{(r_i^{\\mathrm{Text}}, r_i^{\\mathrm{Image}})\\}_{i=1}^{t-1}\\), we initially construct" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.653, + 0.767, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.802, + 0.785, + 0.831 + ], + "angle": 0, + "content": "Fig.6: Bridging the Image History to the Image Response." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "HS. Yoon et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.785, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Fig. 7: Overall Inference Procedure. During inference, the Textual Response Generator \\(\\mathcal{G}\\) generates the citation-augmented textual image description \\(u_{t}^{\\prime}\\). By utilizing its citation tag \\(c_{t}\\), object consistency can be maintained by feeding all preceding image responses with the identical citation tag into a Customized Text-to-Image Model \\(\\mathcal{F}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.789, + 0.522 + ], + "angle": 0, + "content": "the sequence of textual image descriptions and corresponding text responses as \\(\\{r_1^{\\mathrm{Text}}, u_1, \\ldots, r_{t-1}^{\\mathrm{Text}}, u_{t-1}\\}\\). Our Citation Module \\(\\mathcal{C}\\) then augments the textual image descriptions with citation tags to create \\(\\{r_1^{\\mathrm{Text}}, u_1', \\ldots, r_{t-1}^{\\mathrm{Text}}, u_{t-1}'\\}\\), which serves as the input for the Textual Response Generator \\(\\mathcal{G}\\). During inference, if \\(\\mathcal{G}\\) predicts an [IMG] token, it initiates the generation of a citation-augmented textual image description \\(u_t'\\). We extract the citation \\(c_t\\) and the core image description \\(u_t\\) from \\(u_t'\\). Subsequently, the image description \\(u_t\\) is input to our Customized Text-to-Image Model \\(\\mathcal{F}\\), along with all preceding image responses sharing the same citation \\(c_t\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.532, + 0.787, + 0.559 + ], + "angle": 0, + "content": "\\[\nr _ {t} ^ {\\mathrm {I m a g e}} = \\mathcal {F} \\left(u _ {t} \\mid \\{r _ {i} ^ {\\mathrm {I m a g e}} \\mid c _ {i} = c _ {t} \\} _ {i = 1} ^ {t - 1}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.598 + ], + "angle": 0, + "content": "This approach ensures the consistent generation of specific objects across the conversation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.623, + 0.377, + 0.64 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.761 + ], + "angle": 0, + "content": "This section presents the implementation details, evaluation benchmarks, and the experimental results of our approach. Section 4.1 outlines the implementation specifics. In Section 4.2, we assess the overall quality of our model against the standard benchmarks established in previous works [9, 47]. Section 4.3 is dedicated to evaluating the image grounding effectiveness of our model. Finally, Section 4.4 examines the consistency of the image responses generated by our system." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.429, + 0.801 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Dataset As mentioned in Section 2.1, our benchmark datasets include ImageChat [42], PhotoChat [47], and MMDialog [9]. Specifically, for the overall" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.274, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Briding Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "multimodal dialogue evaluation in Section 4.2, we train and evaluate using the PhotoChat and MMDialog. For the image grounding evaluation in Section 4.3, we use ImageChat. For the image consistency evaluation in Section 4.4, we create the MDIC dataset by hand-labeling a subset of MMDialog test set. For Section 4.3 and Section 4.4, we use the model trained on MMDialog from Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Implementation Details For the Textual Dialogue Response Generator \\(\\mathcal{G}\\) and the Visual Encoder \\(\\mathcal{V}\\), we initialize with the pre-trained OpenFlamingo 4B model\\(^3\\) [4]. During the fine-tuning phase, we employ special tokens to structure our inputs and outputs: [IMG] and [/IMG] encapsulate textual image descriptions, while [EOT] signifies the end of conversation turns. The BLIP2-flan-t5-xl model [24] is used for converting image responses to corresponding textual image descriptions. Additionally, our Citation Module \\(\\mathcal{C}\\) uses [CITE] and [/CITE] tokens to mark the beginning and end of citations linked to key objects within the textual image descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.534 + ], + "angle": 0, + "content": "As noted in Section 3.2, the Citation Module \\(\\mathcal{C}\\) is composed of four key components: the POS tagging processor \\((\\mathcal{P})\\), the open-set object detector \\((\\mathcal{M})\\), the segmentation model \\((\\mathcal{S})\\), and the visual feature extractor \\((\\mathcal{E})\\). Specifically, we employ spaCy [13] for \\(\\mathcal{P}\\), GroundingDino [28] for \\(\\mathcal{M}\\), the Segment Anything Model (SAM) [15] for \\(\\mathcal{S}\\), and DINOv2 [33] for \\(\\mathcal{E}\\). The similarity threshold \\(\\tau\\) is set to 0.6. No further learning is done for these modules; they are utilized as pre-trained components within our system for citation tagging. For our customized text-to-image generation model, \\(\\mathcal{F}\\), we used BLIP-Diffusion [23] when conditioning on the input image and the standard Stable Diffusion 2.1 [38] when there is no input image conditioning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.616 + ], + "angle": 0, + "content": "Learning Details Let us denote \\(\\{\\theta_{\\mathcal{V}},\\theta_{\\mathcal{G}_v},\\theta_{\\mathcal{G}_l}\\}\\) as the parameters of the perceiver resampler of the Visual Encoder \\(\\nu\\), the visual cross-attention layers of the Textual Dialogue Response Generator \\(\\mathcal{G}\\), and the language model layers of \\(\\mathcal{G}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.707 + ], + "angle": 0, + "content": "In the first stage of training, we train \\(\\theta_{\\mathcal{G}_l}\\). The batch size is set to 256 with a maximum token length set to 256. In the second stage of training, we jointly train \\(\\theta_{\\mathcal{V}}\\) and \\(\\theta_{\\mathcal{G}_v}\\). The batch size is set to 128 with a maximum token length set to 512. Both the first and second stage is trained by minimizing the next token prediction loss (Eq. 1) using the AdamW optimizer [29] with a learning rate set to 1e-4. The trainings were conducted using \\(16 \\times\\) NVIDIA A100 80GB PCIe." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.73, + 0.533, + 0.745 + ], + "angle": 0, + "content": "4.2 Multimodal Dialogue Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Evaluation Dataset We evaluate the overall performance of our BI-MDRG system on the test set of PhotoChat [47] and MMDialog [9] dataset. PhotoChat contains a single image per dialogue, while the MMDialog includes dialogues with multiple images across turns, offering a more complex context for assessing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.763, + 0.841 + ], + "angle": 0, + "content": "3 https://huggingface.co/openflamingo/OpenFlamingo-4B-vitl-rpj3b-langinstruct" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "HS. Yoon et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: Automatic evaluation results of BI-MDRG on the PhotoChat and the MM-Dialog test set. Numbers in bold represent the best scores." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.185, + 0.769, + 0.37 + ], + "angle": 0, + "content": "
ModelIntent F1Image Response ISTextual B1Image Description B2R-1R-LText Response
B1B2R1R-L
PhotoChat Dataset [47]
BERT-base [47]53.2---------
T5-3B [47]58.9---------
Divter [43]56.215.815.111.4-15.86.521.66-5.69
DivterLLM54.116.141.327.143.341.611.44.7511.210.8
BI-MDRG55.716.742.128.244.642.512.45.1212.111.2
MMDialog Dataset [9]
PaCE [25]77.6---------
Divter [9]71.820.5----9.447.45-11.2
MiniGPT5 [50]-20.2----29.119.5-12.1
DivterLLM67.321.044.235.745.543.621.316.220.419.4
BI-MDRG70.522.452.244.753.251.627.623.525.724.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.43 + ], + "angle": 0, + "content": "our system's capability. We perform the tests for all turns except the first turn of each dialogue and consider all previous turns as context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Evaluation Metric The performance evaluation is carried out using automatic metrics across four key dimensions: (1) Image Intent Prediction - assessing the need for an image response in the current turn; (2) Image Response Quality; (3) Textual Image Description Quality; and (4) Text Response Generation. We employ the F1 metric for Image Intent Prediction, following the binary classification approach detailed in [9, 43]. BLEU [34] and ROUGE [26] metrics are used for evaluating both the Textual Image Descriptions and Text Responses. The Image Response Quality is measured using the Inception Score (IS) [41], in line with [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Baselines The BI-MDRG system is compared against Divter [43], a prior Multimodal Dialogue Response Generator that utilizes DialoGPT (762M) [48] as its language model backbone. For a fair comparison, we also train Divter on the same backbone Language Model as our BI-MDRG, denoted as \\(\\mathrm{Divter}_{\\mathrm{LLM}}\\) (3B). Moreover, for the MMDialog dataset, we compare with MiniGPT5 [50] which uses a 9B backbone VLM. For Intent-prediction baselines, we also include BERT-base, T5-3B reported in [47] for the PhotoChat dataset and PaCE reported in [25] for the MMDialog dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Results Table 1 presents the evaluation results, where our BI-MDRG system demonstrates outstanding performance on both datasets. In Textual Image Description and Text Response Generation, BI-MDRG achieves the highest BLEU and ROUGE metrics scores in most cases, indicating the system's proficiency in generating relevant and coherent image responses and text responses. Notably," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.274, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Briding Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "BI-MDRG, which uses a 4B backbone VLM, outperforms MiniGPT5, which utilizes a 9B backbone VLM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.198, + 0.502, + 0.214 + ], + "angle": 0, + "content": "4.3 Image Grounding Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Evaluation Dataset To evaluate the ability to generate image-grounded textual responses, we utilize the ImageChat dataset [42], consisting of dialogues centered around a single image (an example is shown in Appendix C.4). Since the conversations are grounded in an image, the dataset is suitable for the evaluation of the image grounding capability of the model. We use the same model trained on the MMDialog dataset in Section 4.2 without further tuning on the ImageChat train set." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.35, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Evaluation Metric In order to evaluate the image grounding capability, we use BLEU and ROUGE as the metrics. We perform the evaluation only on the last turn of each dialogue and consider all previous turns as the input context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.517, + 0.612 + ], + "angle": 0, + "content": "Results The performance of our BI-MDRG system on the ImageChat benchmark is summarized in Table 2. These results demonstrate the system's superior image grounding capabilities in text responses. BI-MDRG achieves the highest scores in BLEU-1, ROUGE-1, and ROUGE-L metrics, indicating its effectiveness in generating contextually relevant and coherent text responses grounded in the visual elements of the conversation. In comparison to DivterLLM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.785, + 0.673 + ], + "angle": 0, + "content": "and BI-MDRG\\(_{\\mathrm{w/o mask}}\\), our BI-MDRG model shows an improvement, highlighting the significance of our proposed architectural enhancements and the multimodal causal attention mask in understanding and integrating visual context into text responses effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Moreover, we show an example of the model prediction between the Divter\\(_{\\text{LLM}}\\) and our BI-MDRG in Figure 8-(a). Figure 8-(a) (right-top) shows that Divter\\(_{\\text{LLM}}\\) produces contextually irrelevant text response due to its lack of image understanding. However, Figure 8-(a) (right-bottom) shows that since our BI-MDRG has access to the image, it can produce a text response grounded on the image." + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.439, + 0.788, + 0.482 + ], + "angle": 0, + "content": "Table 2: Automatic evaluation results on the ImageChat test set. Numbers in bold represent the best scores." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.495, + 0.773, + 0.591 + ], + "angle": 0, + "content": "
ModelB1R-1R-L
ImageChat Dataset [42]
DivterLLM8.610.39.6
BI-MDRGw/o mask10.011.110.2
BI-MDRG10.911.710.9
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.593, + 0.787 + ], + "angle": 0, + "content": "4.4 Image Response Consistency Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We integrated the Citation Module into our BI-MDRG system to improve consistency in image responses within the dialogue. This module assigns citation tags to primary objects in textual image descriptions, tracking their presence across" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "HS. Yoon et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.147, + 0.497, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.253, + 0.464, + 0.263 + ], + "angle": 0, + "content": "(a) Enhanced Image Grounded Text Response" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.147, + 0.756, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.253, + 0.726, + 0.264 + ], + "angle": 0, + "content": "(b) Enhanced Consistency of Image Response" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.278, + 0.785, + 0.307 + ], + "angle": 0, + "content": "Fig. 8: Example of predictions from Divter\\(_{\\text{LLM}}\\) [43] vs. BI-MDRG (ours). Additional examples can be found in Appendix E." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.319, + 0.406, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.419, + 0.383, + 0.429 + ], + "angle": 0, + "content": "(a) Citation Module Evaluation" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.319, + 0.578, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.419, + 0.582, + 0.429 + ], + "angle": 0, + "content": "(b) Citation Token Prediction Evaluation" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.32, + 0.787, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.42, + 0.762, + 0.43 + ], + "angle": 0, + "content": "(c) Image Consistency Evaluation" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.444, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Fig. 9: Three evaluation aspects for image response consistency. (a) We assess the accuracy of the pseudo-labels of citation tags created by the Citation Module. (b) We evaluate how well the model learns from these pseudo-labels and predict the citation tag \\(\\hat{c}_t\\), given the dialogue history, current text response \\(r_t^{\\mathrm{Text}}\\), and \\(\\tilde{u}_t\\) which denotes the textual image description leading up to the primary object word \\(o_t\\). (c) We evaluate the consistency of the resulting image response \\(\\hat{r}_t^{\\mathrm{Image}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.556, + 0.788, + 0.603 + ], + "angle": 0, + "content": "the dialogue. As shown in Figure 9, we focus on three aspects for evaluating the image response consistency: (a) Citation Module Evaluation, (b) Citation Token Prediction Evaluation, and (c) Image Consistency Evaluation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.622, + 0.788, + 0.714 + ], + "angle": 0, + "content": "Multimodal Dialogue Image Consistency (MDIC) Dataset Creation Due to the absence of benchmark datasets for evaluating the image consistency in multimodal dialogue scenarios, we created a dataset of 300 dialogues, each manually annotated to assign correct citation tags to objects based on their appearances in image responses. We use this created dataset for the subsequent evaluations. Details of the annotation process can be found in Appendix C.5." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Citation Module Evaluation Figure 9-(a) shows the evaluation method for the Citation Module. As described in Section 3.2, the module \\(\\mathcal{C}\\) processes visual features \\(\\{f_1,\\ldots ,f_n\\}\\) corresponding to primary objects \\(\\{o_1,\\dots,o_n\\}\\) in image descriptions. The module generates citations based on the cosine similarity between these features. We constructed a similarity matrix \\(\\mathcal{X}\\), where each element \\(\\mathcal{X}_{ij}\\) represents the cosine similarity between the features \\(f_{i}\\) and \\(f_{j}\\). A feature pair \\((f_i,f_j)\\) is deemed 'similar' if \\(\\mathcal{X}_{ij} \\geq \\tau\\) and 'not similar' otherwise," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.274, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Briding Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.195 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a predetermined threshold. The module's performance was quantified using the F1 score, and we achieved an F1 score of 0.72, demonstrating the module's high accuracy in assigning pseudo-labels of citation tags for training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.229, + 0.518, + 0.395 + ], + "angle": 0, + "content": "Citation Token Prediction Evaluation Our model, trained with citation-augmented textual image descriptions, can generate these descriptions with accurate citation tags for the primary objects in the descriptions. Figure 9-(b) shows the evaluation method where we assess our model's ability to predict the citation tags correctly. We input the ground truth sequence \\(\\{r_1^{\\mathrm{Text}},u_1',\\ldots ,r_t^{\\mathrm{Text}},\\tilde{u}_t\\}\\) from the MDIC dataset along with the image re" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.518 + ], + "angle": 0, + "content": "sponse history \\(\\{r_1^{\\mathrm{Image}},\\ldots ,r_{t - 1}^{\\mathrm{Image}}\\}\\). Here, \\(\\tilde{u}_t\\) denotes the textual image description leading up to the principal object word \\(o_t\\). The goal is to verify whether the model's predicted citation \\(\\hat{c}_t\\) aligns with the actual citation \\(c_t\\). The accuracy (Acc.) of our citation token prediction is presented in Table 3. We compare with a baseline denoted as Divter\\(_{\\mathrm{LLM}} + \\mathrm{LLMCite}\\) where we assign citation tags by instructing an external LLM (Details in Appendix D). The results show that BI-MDRG achieves \\(84.0\\%\\) accuracy compared to \\(33.5\\%\\) of the LLMCite baseline, showing the effectiveness of our model in identifying objects needing consistency." + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.236, + 0.789, + 0.293 + ], + "angle": 0, + "content": "Table 3: Citation Token Prediction (Acc.) and Image Consistency Evaluation (DINOv2). Numbers in bold represent the best scores." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.306, + 0.776, + 0.374 + ], + "angle": 0, + "content": "
ModelAcc.DINOv2(↑)
MDIC Dataset
DivterLLM + LLMCite33.50.32
BI-MDRG (ours)84.00.53
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Image Consistency Evaluation Figure 9-(c) shows the evaluation method for the Image Consistency Evaluation. To assess the consistency of the image response, we employ metrics commonly found in the literature on custom object text-to-image generation [7, 10, 18, 19, 39]. DINO features outperform CLIP in capturing image alignment, specifically in distinguishing different objects of the same class [39]. Therefore, for our evaluation, we measure image alignment scores using the DINOv2 [33] features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.75 + ], + "angle": 0, + "content": "In our proposed framework, correctly using the citation tag is essential to ensure image consistency in the conversation. This involves correctly predicting a new citation tag or referencing an existing tag in history. Therefore, our evaluation includes not only scenarios where the model's predicted citation tag aligns with the ground truth \\( (c_{t} = \\hat{c}_{t}) \\) and \\( c_{t} \\) exists in previous dialogue history, but also instances where the model's predicted citation tag do not match the ground truth \\( (c_{t} \\neq \\hat{c}_{t}) \\) and either \\( \\hat{c}_{t} \\) or \\( c_{t} \\) is in previous dialogue history." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We report the DINOv2 score as an average of both the successful and failed citation cases mentioned above, where successful prediction positively influences the score, and failed prediction reduces it, offering a comprehensive measure of the model's performance in image consistency. Table 3 shows that our method successfully maintains the consistency of the image response in a dialogue setting, with the DINOv2 score achieving 0.53 compared to 0.32 of baseline. Furthermore," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "HS. Yoon et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 4: Evaluation of the citation framework on various settings. The image consistency is measured using DINOv2 (↑) on MDIC dataset, and TID (Textual Image Description) and TR (Text Response) are measured on MMDialog [9] test set. Custom uses customized text-to-image generation with image conditioning, while Diffusion employs a standard text-to-image diffusion model without conditioning. The citation framework allows to selectively use between these for consistent image response. The first row represents the performance of BI-MDRG (ours)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.248, + 0.785, + 0.352 + ], + "angle": 0, + "content": "
CitationVLM SizeDiffusionDINOv2TIDTR
B1B2R-1R-LB1B2R-1
Citation Module4BCustom+Diffusion0.5352.244.753.251.627.623.525.7
LLMCite x4BCustom+Diffusion0.3452.044.953.351.327.423.425.8
4BDiffusion0.2552.044.953.351.327.423.425.8
LLMCite x9BCustom+Diffusion0.3356.048.754.552.130.325.728.8
9BDiffusion0.2656.048.754.552.130.325.728.8
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.37, + 0.784, + 0.403 + ], + "angle": 0, + "content": "in Figure 8-(b), we present a comparative example of enhanced image consistency of prediction from BI-MDRG compared to Divter\\(_{\\text{LLM}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.426, + 0.75, + 0.446 + ], + "angle": 0, + "content": "5 Importance of Citation Tags for Image Consistency" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.537 + ], + "angle": 0, + "content": "Due to powerful pre-trained models for text-to-text and text-to-image, adopting text as an intermediary for image responses is a practical solution for the MDRG [43] task. However, due to the inherent information loss of images during this process, achieving image consistency is infeasible without a targeted framework for consistency maintenance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.675 + ], + "angle": 0, + "content": "Table 4 shows image consistency and dialogue response performance across various settings. Not using citations shows similar dialogue response performance (Intent, TID, TR) compared to LLMCite and our Citation Module. However, LLMCite improves image consistency (DINOv2) from 0.25 to 0.34 (4B) and 0.26 to 0.33 (9B). Our Citation Module further boosts this from 0.34 to 0.53, indicating its importance for image consistency without affecting dialogue response performance. Notably, scaling the model size (4B to 9B) improves textual response but fails to maintain image consistency without our citation framework, as also evident by ChatGPT's shortcomings in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.7, + 0.36, + 0.716 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "This paper presents BI-MDRG, a novel framework for Multimodal Dialogue Response Generation (MDRG) aimed at bridging the image history for enhanced text and image response. Our model's innovative use of image history to inform both text and image responses addresses fundamental limitations in previous methodologies, particularly in maintaining consistency in multimodal interactions. The effectiveness of BI-MDRG has been demonstrated through rigorous evaluations using multiple benchmark datasets and a custom-annotated dataset." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.274, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Briding Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.283 + ], + "angle": 0, + "content": "This work was supported by a grant of the KAIST-KT joint research project through AI2X Lab., Tech Innovation Group, funded by KT (No. D23000019, Developing Visual and Language Capabilities for AI-Based Dialogue Systems), and by Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2021-0-01381, Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.302, + 0.323, + 0.318 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.387 + ], + "angle": 0, + "content": "1. Alamri, H., Cartillier, V., Das, A., Wang, J., Cherian, A., Essa, I., Batra, D., Marks, T.K., Hori, C., Anderson, P., Lee, S., Parikh, D.: Audio-visual scene-aware dialog. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.785, + 0.483 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=EbMuimAbPbs" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.51 + ], + "angle": 0, + "content": "3. Anonymous: Chatsearch: a dataset and a generative retrieval model for general conversational image retrieval (2023), https://openreview.net/forum?id=0unbjYPmbC" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.579 + ], + "angle": 0, + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.634 + ], + "angle": 0, + "content": "5. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Zhang, Q., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., Karras, T., Liu, M.Y.: ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.661 + ], + "angle": 0, + "content": "6. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.662, + 0.785, + 0.702 + ], + "angle": 0, + "content": "7. Cohen, N., Gal, R., Meirom, E.A., Chechik, G., Atzmon, Y.: \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In: European Conference on Computer Vision. pp. 558-577. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.743 + ], + "angle": 0, + "content": "8. Das, A., Kottur, S., Gupta, K., Singh, A., Yadav, D., Moura, J.M., Parikh, D., Batra, D.: Visual dialog. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 326-335 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.84 + ], + "angle": 0, + "content": "9. Feng, J., Sun, Q., Xu, C., Zhao, P., Yang, Y., Tao, C., Zhao, D., Lin, Q.: MMDialog: A large-scale multi-turn dialogue dataset towards multi-modal open-domain conversation. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 7348-7363. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-long.405, https://aclanthology.org/2023.acl-long.405" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "HS. Yoon et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "10. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.288 + ], + "angle": 0, + "content": "11. Gan, Z., Cheng, Y., Kholy, A., Li, L., Liu, J., Gao, J.: Multi-step reasoning via recurrent dual attention for visual dialog. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 6463-6474. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1648, https://aclanthology.org/P19-1648" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.317 + ], + "angle": 0, + "content": "12. Han, S., Hessel, J., Dziri, N., Choi, Y., Yu, Y.: Champagne: Learning real-world conversation from large-scale web videos. arXiv preprint arXiv:2303.09713 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.358 + ], + "angle": 0, + "content": "13. Honnibal, M., Montani, I.: spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing (2017), to appear" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.387 + ], + "angle": 0, + "content": "14. Kim, J., Yoon, S., Kim, D., Yoo, C.D.: Structured co-reference graph attention for video-grounded dialogue (2021), https://arxiv.org/abs/2103.13361" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.787, + 0.43 + ], + "angle": 0, + "content": "15. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.459 + ], + "angle": 0, + "content": "16. Koh, J.Y., Salakhutdinov, R., Fried, D.: Grounding language models to images for multimodal inputs and outputs. ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.515 + ], + "angle": 0, + "content": "17. Koo, G., Yoon, S., Yoo, C.D.: Wavelet-guided acceleration of text inversion in diffusion-based image editing. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4380-4384. IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.544 + ], + "angle": 0, + "content": "18. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.787, + 0.573 + ], + "angle": 0, + "content": "19. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.574, + 0.787, + 0.657 + ], + "angle": 0, + "content": "20. Lee, N., Shin, S., Choo, J., Choi, H.J., Myaeng, S.H.: Constructing multi-modal dialogue dataset by replacing text with semantically relevant images. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). pp. 897-906. Association for Computational Linguistics, Online (Aug 2021), https://aclanthology.org/2021.acl-short.113" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.685 + ], + "angle": 0, + "content": "21. Lee, Y.J., Ko, B., Kim, H.G., Choi, H.J.: Dialogcc: Large-scale multi-modal dialogue dataset. arXiv preprint arXiv:2212.04119 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.787, + 0.769 + ], + "angle": 0, + "content": "22. Levy, M., Ben-Ari, R., Darshan, N., Lischinski, D.: Chatting makes perfect: Chat-based image retrieval. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 61437-61449. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/c1b3d1e2cf53bb28cabd801bd58b3521-Paper-Conference.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "23. Li, D., Li, J., Hoi, S.C.H.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.275, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Briding Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.244 + ], + "angle": 0, + "content": "25. Li, Y., Hui, B., Yin, Z., Yang, M., Huang, F., Li, Y.: PaCE: Unified multi-modal dialogue pre-training with progressive and compositional experts. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 13402-13416. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-long.749, https://aclanthology.org/2023.acl-long.749" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.288 + ], + "angle": 0, + "content": "26. Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out. pp. 74-81. Association for Computational Linguistics, Barcelona, Spain (Jul 2004), https://aclanthology.org/W04-1013" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.343 + ], + "angle": 0, + "content": "27. Lin, H., Ruan, L., Xia, W., Liu, P., Wen, J., Xu, Y., Hu, D., Song, R., Zhao, W.X., Jin, Q., et al.: Tiktalk: A video-based dialogue dataset for multi-modal chitchat in real world. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1303-1313 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.345, + 0.787, + 0.387 + ], + "angle": 0, + "content": "28. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.787, + 0.416 + ], + "angle": 0, + "content": "29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.459 + ], + "angle": 0, + "content": "30. Ma, J., Liang, J., Chen, C., Lu, H.: Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.5 + ], + "angle": 0, + "content": "31. Meng, Y., Wang, S., Han, Q., Sun, X., Wu, F., Yan, R., Li, J.: Openvidual: A large-scale, open-domain dialogue dataset with visual contexts. arXiv preprint arXiv:2012.15015 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.543 + ], + "angle": 0, + "content": "32. Niu, Y., Zhang, H., Zhang, M., Zhang, J., Lu, Z., Wen, J.R.: Recursive visual attention in visual dialog. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6679-6688 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.614 + ], + "angle": 0, + "content": "33. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.685 + ], + "angle": 0, + "content": "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: A method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting on Association for Computational Linguistics. p. 311-318. ACL '02, Association for Computational Linguistics, USA (2002). https://doi.org/10.3115/1073083.1073135, https://doi.org/10.3115/1073083.1073135" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.768 + ], + "angle": 0, + "content": "35. Poria, S., Hazarika, D., Majumder, N., Naik, G., Cambria, E., Mihalcea, R.: MELD: A multimodal multi-party dataset for emotion recognition in conversations. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 527-536. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1050, https://aclanthology.org/P19-1050" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.769, + 0.787, + 0.811 + ], + "angle": 0, + "content": "36. Qi, J., Niu, Y., Huang, J., Zhang, H.: Two causal principles for improving visual dialog. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10860-10869 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "37. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "HS. Yoon et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.203 + ], + "angle": 0, + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.259 + ], + "angle": 0, + "content": "39. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.34 + ], + "angle": 0, + "content": "40. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=08Yk-n512A1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.424 + ], + "angle": 0, + "content": "41. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X., Chen, X.: Improved techniques for training gans. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 29. Curran Associates, Inc. (2016), https://proceedings.neurips.cc/paper_files/paper/2016/file/8a3363abe792db2d8761d6403605aab7-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.507 + ], + "angle": 0, + "content": "42. Shuster, K., Humeau, S., Bordes, A., Weston, J.: Image-chat: Engaging grounded conversations. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 2414-2429. Association for Computational Linguistics, Online (Jul 2020). https://doi.org/10.18653/v1/2020.acl-main.219, https://aclanthology.org/2020.acl-main.219" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.591 + ], + "angle": 0, + "content": "43. Sun, Q., Wang, Y., Xu, C., Zheng, K., Yang, Y., Hu, H., Xu, F., Zhang, J., Geng, X., Jiang, D.: Multimodal dialogue response generation. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2854-2866. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.204, https://aclanthology.org/2022.acl-long.204" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.633 + ], + "angle": 0, + "content": "44. Wang, S., Meng, Y., Li, X., Sun, X., Ouyang, R., Li, J.: Openvidual 2.0: A larger-scale, open-domain dialogue generation dataset with visual contexts. arXiv preprint arXiv:2109.12761 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "45. Yoon, S., Kim, D., Yoon, E., Yoon, H.S., Kim, J., Yoo, C.D.: Hear: Hearing enhanced audio response for video-grounded dialogue. arXiv preprint arXiv:2312.09736 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.787, + 0.757 + ], + "angle": 0, + "content": "46. Yoon, S., Yoon, E., Yoon, H.S., Kim, J., Yoo, C.: Information-theoretic text hallucination reduction for video-grounded dialogue. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 4182-4193. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.280, https://aclanthology.org/2022.emnlp-main.280" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Zang, X., Liu, L., Wang, M., Song, Y., Zhang, H., Chen, J.: PhotoChat: A human-human dialogue dataset with photo sharing behavior for joint image-text modeling. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 6142-6152. Association for Computational Linguis" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.275, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Briding Image History in Multimodal Dialogue Response Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "tics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.479, https://aclanthology.org/2021.acl-long.479" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "48. Zhang, Y., Sun, S., Galley, M., Chen, Y.C., Brockett, C., Gao, X., Gao, J., Liu, J., Dolan, B.: Dialogpt: Large-scale generative pre-training for conversational response generation. In: ACL, system demonstration (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.3 + ], + "angle": 0, + "content": "49. Zhao, J., Zhang, T., Hu, J., Liu, Y., Jin, Q., Wang, X., Li, H.: M3ED: Multi-modal multi-scene multi-label emotional dialogue database. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 5699-5710. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.391, https://aclanthology.org/2022.acl-long.391" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.787, + 0.328 + ], + "angle": 0, + "content": "50. Zheng, K., He, X., Wang, X.E.: Minigpt-5: Interleaved vision-and-language generation via generative vocens. arXiv preprint arXiv:2310.02239 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.329, + 0.787, + 0.411 + ], + "angle": 0, + "content": "51. Zheng, Y., Chen, G., Liu, X., Sun, J.: MMChat: Multi-modal chat dataset on social media. In: Calzolari, N., Bechet, F., Blache, P., Choukri, K., Cieri, C., Declerck, T., Goggi, S., Isahara, H., Maegaard, B., Mariani, J., Mazo, H., Odijk, J., Piperidis, S. (eds.) Proceedings of the Thirteenth Language Resources and Evaluation Conference. pp. 5778-5786. European Language Resources Association, Marseille, France (Jun 2022), https://aclanthology.org/2022.lrec-1.621" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.411 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_origin.pdf b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d24af9090883a4ab89942e3860da6302950448a7 --- /dev/null +++ b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/636c15ae-30f9-4ded-b5a0-a79ba850c29c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:070ae57d968a20cc761667fee6dab2e442878807f45e794e352b0ae70481642e +size 9190917 diff --git a/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/full.md b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8d9d95719de39ae1604473833e881c5b182f1df --- /dev/null +++ b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/full.md @@ -0,0 +1,310 @@ +# BI-MDRG: Bridging Image History in Multimodal Dialogue Response Generation + +Hee Suk Yoon $^{1\star}$ , Eunseop Yoon $^{1\star}$ , Joshua Tian Jin Tee $^{1\star}$ , Kang Zhang $^{1}$ , Yu-Jung Heo $^{2}$ , Du-Seong Chang $^{2}$ , and Chang D. Yoo $^{1\star\star}$ + +$^{1}$ Korea Advanced Institute of Science and Technology (KAIST) + +$^{2}$ KT Corporation + +{hskyoon, esyoon97, joshuateetj, zhangkang, cd_yoo}@kaist.ac.kr + +{yj.heo,dschang}@kt.com + +Abstract. Multimodal Dialogue Response Generation (MDRG) is a recently proposed task where the model needs to generate responses in texts, images, or a blend of both based on the dialogue context. Due to the lack of a large-scale dataset specifically for this task and the benefits of leveraging powerful pre-trained models, previous work relies on the text modality as an intermediary step for both the image input and output of the model rather than adopting an end-to-end approach. However, this approach can overlook crucial information about the image, hindering 1) image-grounded text response and 2) consistency of objects in the image response. In this paper, we propose BI-MDRG that bridges the response generation path such that the image history information is utilized for enhanced relevance of text responses to the image content and the consistency of objects in sequential image responses. Through extensive experiments on the multimodal dialogue benchmark dataset, we show that BI-MDRG can effectively increase the quality of multimodal dialogue. Additionally, recognizing the gap in benchmark datasets for evaluating the image consistency in multimodal dialogue, we have created a curated set of 300 dialogues annotated to track object consistency across conversations. The code and the dataset is publicly available at https://github.com/hee-suk-yoon/BI-MDRG. + +Keywords: Multimodal Dialogue $\cdot$ Image Grounding $\cdot$ Image Consistency + +# 1 Introduction + +With the development of instant messaging technology, visual modalities are increasingly used alongside text in online communication. To enhance user interaction with intelligent agents, a new task, Multimodal Dialogue Response Generation (MDRG) [43], has been proposed. This task requires models to generate both text and image responses based on dialogue history containing texts and images. Since learning an effective multimodal generation model with a sin + +![](images/75f9b85ffddd660a717389a9fcac18ecb8a61899dbad8137abf85be598972259.jpg) +(a) Overall framework of previous MDRG system + +![](images/e8d04309ebe22b6b530bc8e25fc1dcc85d57332f800708fbe0baebf1eaa9a2b2.jpg) +(b) Loss of Image History Information when generating text and image response. + +![](images/1cd0ceb2f7bea0c731501004a2fdcd1fbcf2235fa4a0b5fdff99546ff8594b83.jpg) +Fig. 1: (a) Outlines the framework of previous Multimodal Dialogue Response Generation (MDRG) systems, which uses the textual descriptions of images $(u_{t})$ as an intermediary step toward generating image responses $(r_{t}^{\mathrm{Image}})$ . (b) Highlights the limitations of these systems, particularly their failure to fully leverage image history $(r_{1:t-1}^{\mathrm{Image}})$ in crafting both the textual response $(r_{t}^{\mathrm{Text}})$ and the image response $(r_{t}^{\mathrm{Image}})$ . (c) illustrates the consequences of this oversight, including responses that lack grounding in image context and consistency in image-based replies. + +![](images/d460dab270f987011460d6970a27df4d199b9a97f88de4980c66ffcd8c390ce9.jpg) +(c) Problems of Previous MDRG system + +gle sequence-to-sequence model requires a large number of training instances, the lack of a large-scale multimodal dialogue dataset poses a significant challenge for the task. + +To solve this issue, the previous MDRG model [43] disentangles the textual response and image response generation, as illustrated in Figure 1-(a). Specifically, given the dialogue history $D$ at $t$ -th conversation turn, the MDRG model first generates the textual response $r_t^{\mathrm{Text}}$ and the intermediary textual image description $u_t$ . This image description $u_t$ is subsequently fed into a text-to-image model to generate the corresponding image $r_t^{\mathrm{Image}}$ . Such an approach allows the utilization of powerful models pre-trained on vast amounts of available data for text-to-text and text-to-image pairs, bypassing the need for a large-scale multimodal dialogue dataset for direct end-to-end training. + +Nonetheless, due to its reliance on text as an intermediary representation of images, the previous MDRG model overlooks crucial information about the image, as shown in Figure 1-(b). By converting images from the dialogue history into textual descriptions $(u_{1:t-1})$ , these models fail to fully utilize the rich visual content in the actual images $(r_{1:t-1}^{\mathrm{Image}})$ . This leads to two major issues: a lack of image-grounded context in textual responses (Figure 1-(c) (left)) and inconsistencies in image generation across dialogues (Figure 1-(c) (right)). + +For instance, Figure 1-(c) (left) shows the existing MDRG model failing to provide an image-grounded response to 'What breed is your dog?' because it only perceives the dog through the text 'a dog eating a watermelon'. Figure 1-(c) (right) illustrates the model's inconsistency, where the 'dog' in the image history is not maintained in the generated response. These issues highlight the need for improved MDRG models that effectively utilize image history. + +Contribution This paper introduces BI-MDRG to enhance text and image responses by bridging the image history. + +- Bridging Image History in Text Responses: In Section 3.1, we propose an architectural modification in which the visual features, extracted using a visual encoder, are integrated into the cross-attention layers of the core language model. This is complemented by a novel multimodal causal attention mask modulation tailored for the MDRG task to improve image-grounded textual responses. + +- Bridging Image History in Image Responses: In Section 3.2, we propose a citation framework where we use the Citation Module designed to augment textual image descriptions by tagging objects with citation tags that identify recurring objects throughout a dialogue. In Section 3.3, we show that training with these augmented data enables the model to recognize and maintain the consistency of objects in subsequent image responses during inference using Customized Text-to-Image Models. Due to the absence of benchmark datasets for evaluating image consistency in multimodal dialogues, we have created the Multimodal Dialogue Image Consistency (MDIC) dataset containing dialogues annotated to track object consistency across conversations. + +# 2 Related Work + +# 2.1 Multimodal Dialogue Datasets + +Multimodal dialogue datasets generally fall into three categories: question and answering [1,8] (where the task involves asking and answering questions about a specific image), in-scene [12,31,35,44,49] (where each dialogue turn corresponds to a scene from a movie or video), and conversation-based [9,20,21,27,42,47] (which engage in natural dialogue about a given image or involves image sharing within natural conversations). Further details are in Appendix C.1. + +This paper primarily explores the model's capability for natural dialogue within the conversation-based category. Notable datasets in this segment include ImageChat [42], PhotoChat [47], MMDD [20], DialogCC [21], MMChat [51], TikTalk [27], and MMDialog [9]. Given our focus on English-language scenarios, we exclude MMChat and Tiktalk from our evaluation, as it is a dataset primarily in Chinese. Additionally, DialogCC (not publicly available) and MMDD, which are synthesized by algorithmically pairing images with text-only dialogues for random turns, are also excluded from our analysis. Therefore, our evaluation is centered on ImageChat, PhotoChat, and MMDialog. + +ImageChat [42] consists of image-centered dialogues, where each dialogue is centered around a single given image. PhotoChat [47] features dialogues collected from social media, where a single image is shared in one of the conversation turns, which better mirrors everyday human interaction. Still, their limited scale and domain diversity restrict their applicability. Overcoming these limitations, MMDialog [9] features over a million diverse dialogues from social media, where multiple images are shared across numerous conversation turns, providing a more realistic representation of open-domain multimodal conversations. + +# 2.2 Multimodal Dialogue Modeling + +Pioneering studies [11, 32, 36] have delved into improving the performance of image-grounded conversational agents. While [14, 45, 46, 51] introduced a Seq2Seq based model focusing on multimodal dialogues, it primarily generated textual responses, not fully embracing the multimodal response scenario. In a notable advancement, [43] presented Divter, which not only produces informative text but also generates high-resolution images, marking a significant leap forward in multimodal dialogue response generation (MDRG). It is important to note that our focus in this paper is generation-based models, contrasting with retrieval-based [3, 16, 22], which output image responses by retrieving existing images from a corpus instead of generating new ones. + +# 2.3 Customized Text-to-Image + +Recent studies on text-to-image diffusion models [5, 17, 38, 40] focus on customization [7, 10, 18, 39], learning specific concepts from a few images. Following this, users can flexibly generate the learned concepts into new scenes. Text inversion [10] generates varied contexts for a concept by updating the text embedding without altering the model. Dreambooth [39] and Custom Diffusion [18] fine-tune the U-Net architecture using an identifier, class label, and images. In a notable enhancement, BLIP-Diffusion [23] enables zero-shot subject-driven generation, allowing fast customized text-to-image generation. + +# 3 BI-MDRG: Bridging the Image History in Multimodal Dialogue Response Generation + +We introduce BI-MDRG, a conversational agent designed to produce both textual and visual responses with enhanced awareness of image history. Sections 3.1 and 3.2 detail the training procedure, effectively integrating image history information into text responses and textual image descriptions. Section 3.3 outlines the inference process, wherein the image history informs the image responses by leveraging the captured details from enhanced textual image descriptions. + +# 3.1 Bridging the Image History for Image-Grounded Text Response + +The example shown in Figure 1-(c) (left) highlighted one of the crucial limitations in the previous MDRG system: their reliance on textual descriptions for understanding image history, which hinders the image-grounded textual responses. To overcome this, we have adopted an architectural change along with a multimodal causal attention mask modulation to effectively bridge the image history information $r_{1:t-1}^{\mathrm{Image}}$ to the text response $r_t^{\mathrm{Text}}$ (as depicted in Figure 2). + +![](images/d67d544ddc906675d546ef7170d529ff8bc6fe5f82de2b5bb2592bc52056686b.jpg) +Fig.2: Bridging Image History to the Text Response. + +![](images/96ac44f7b984c60534151a2f0a80dfdaea3a98ce27b5b186fa9bc9761a69d60f.jpg) +Fig. 3: Training of BI-MDRG. (a) Textual Dialogue Response Generator $\mathcal{G}$ cross-attends to the image features from the Visual Encoder $\mathcal{V}$ . (b) Attention Mask Modulation alters the causal attention to prioritize image features over textual image descriptions. (c) Citation Module $\mathcal{C}$ generates citation-augmented textual image descriptions, enabling the tracking of objects within image history for consistency maintenance. + +Architecture Our Textual Dialogue Response Generator $\mathcal{G}$ (Figure 3-(a)), consists of a decoder-only language model with added visual cross-attention layers. These layers directly engage with image features provided by the Visual Encoder $\mathcal{V}$ , drawing inspiration from Flamingo [2], to reduce dependence on textual image descriptions for perceiving images. In our framework, a dialogue context $D = \{(r_i^{\mathrm{Text}}, r_i^{\mathrm{Image}})\}_{i=1}^n$ comprises multiple turns, each with an associated text response $r_i^{\mathrm{Text}}$ and an image response $r_i^{\mathrm{Image}}$ . An image captioning model produces textual descriptions $u_i$ for each image $r_i^{\mathrm{Image}}$ , which are then transformed into citation-augmented descriptions $u_i'$ as shown in Figure 3-(c) and further detailed in Section 3.2 (if there is no image response for a turn, $r_i^{\mathrm{Image}} = \emptyset$ and $u_i' = \emptyset$ ). These descriptions and the text responses $\{r_1^{\mathrm{Text}}, u_1', \ldots, r_n^{\mathrm{Text}}, u_n'\}$ are fed into $\mathcal{G}$ , while the images $\{r_1^{\mathrm{Image}}, \ldots, r_n^{\mathrm{Image}}\}$ are processed by $\mathcal{V}$ to extract image features which are fed to the cross-attention layers in $\mathcal{G}$ . Although our model directly cross-attends to the inputted image features, we retain the textual descriptions $u_{1:n}'$ as essential inputs since we require the generation of textual description by our $\mathcal{G}$ model, which subsequently gets used by a text-to-image model for constructing the image response. Keeping the textual description inputs allows efficient teacher-forced next token prediction training. + +Multimodal Causal Attention Mask Modulation For the input sequence $\{r_1^{\mathrm{Text}}, u_1', \ldots, r_n^{\mathrm{Text}}, u_n'\}$ , we use a specialized mask along with the standard causal mask (Figure 3-(b)). The traditional causal mask allows each text response $r_i^{\mathrm{Text}}$ to access previous textual image descriptions $u_{1:i-1}'$ , leading to reliance on textual information over visual context. Our masking strategy prevents $r_i^{\mathrm{Text}}$ from accessing $u_{1:i-1}'$ , redirecting focus to the actual image features of $r_{1:i-1}^{\mathrm{Image}}$ , ensuring text responses are grounded on raw image features. + +![](images/c950ce3eb160bc479415c7845d7b2d0efd442bf2637b77a13e9117f7781fe267.jpg) +Fig. 4: Illustration of the Citation Module. Citation Module recognizes identical objects within image history and injects this information into the textual image description with citation tags (e.g., [cite]0[/cite]). + +# 3.2 Citation Module: Bridging the Image History to the Textual Image Description + +The example shown in Figure 1-(c) (right) highlighted another critical limitation of the previous MDRG system: its inability to ensure consistency in image responses. To address this, we propose the Citation Module (Figure 3-(c)) to bridge the image history $r_{1:t-1}^{\mathrm{Image}}$ and the textual image description $u_t$ (as depicted in Figure 5) by ensuring that the textual image description accurately relays which objects should persist in subsequent images. + +![](images/dd64c838406fde07f3b821a888ccf254647639771228009cc870e00cae4adecb.jpg) +Fig. 5: Bridging Image History to the Textual Image Description. + +Citation Module Citation Module plays a pivotal role in tracking recurring objects in the dialogue using textual image descriptions. For instance, descriptions like "a dog is in front of a fireplace" and "a dog running in the snow" are augmented to "a dog[cite]0[/cite] is in front of a fireplace" and "a dog[cite]0[/cite] running in the snow," respectively if they reference the same dog. + +Motivated by [30], Figure 4 details the citation process for textual image descriptions $\{u_1, \dots, u_n\}$ . For each textual image description $u_i$ , a Part of Speech (POS) tagging processor $\mathcal{P}$ is employed to tag words and pinpoint the key object word $o_i$ in the description. The word $o_i$ , along with its corresponding image $r_i^{\mathrm{Image}}$ , is processed through an open-set object detector $\mathcal{M}$ to obtain the bounding box of the detected object, which is then input to a segmentation model $\mathcal{S}$ for generating object segmentation mask $s_i$ . These masks are applied to isolate the objects from their backgrounds in $r_i^{\mathrm{Image}}$ since background removal has been proven helpful for better extraction of object features [6]. These isolated objects are then analyzed by a visual feature extractor $\mathcal{E}$ to extract features $f_i$ . The resulting feature set $\{f_1, \dots, f_n\}$ undergoes clustering based on cosine similarity to identify identical objects across images, as outlined in Algorithm 1; this involves assigning each element a cluster id $c_i$ based on the similarity of their features. For each $o_i$ in $u_i$ , we augment the word so that it is followed by its corresponding cluster id (i.e., citation tag) $c_i$ of $f_i$ , resulting in the citation augmented textual description $u_i'$ . This Citation Module operates with off-the-shelf components for citation tags, requiring no training on the target dataset. + +Algorithm 1 Citation Module +Require: image response $(r_1^{\mathrm{Image}},\dots r_n^{\mathrm{Image}})$ , textual image description $(u_{1},\ldots u_{n})$ , similarity threshold $\tau$ , POS tagging processor $\mathcal{P}$ , open-set object detector $\mathcal{M}$ , segmentation model $S$ , visual feature extractor $\mathcal{E}$ , object cluster dictionary $\kappa = \{\}$ +for $i\gets 1$ to $n$ do $\triangleright$ Identify principal object word in $u_{i}$ $o_i\gets \mathcal{P}(u_i)$ $\triangleright$ Obtain bounding box for $o_i$ $box_{i}\gets \mathcal{M}(o_{i},r_{i}^{\mathrm{Image}})$ $\triangleright$ Obtain segmentation mask for $o_i$ $s_i\gets S(box_i)$ $\triangleright$ Extract visual feature for $o_i$ $f_{i}\gets \mathcal{V}(s_{i}\odot r_{i}^{\mathrm{Image}})$ +end for +cluster_id $\leftarrow 0$ +for $i\gets 1$ to $n$ do Perform clustering of $o_i$ . if not $\kappa$ has key(o) then $\mathcal{K}[o_i]\gets \overline{\mathsf{clust}} e r\_ i d$ Get's own cluster. cluster_id $\leftarrow$ cluster_id+1 +end if +for $j\gets i + 1$ to $n$ do +if $\operatorname {SIM}(f_i,f_j)\geq \tau$ and not $\kappa$ has key(oj) then $\kappa [o_j]\gets \kappa [o_i]$ Puts into existing cluster. end if +end for +end for +return K + +Generative Training Objective We use the next token prediction training via teacher forcing, which is used in standard auto-regressive language models [37]. Specifically, given the token sequence $w = \{w_{j}\}_{j=1}^{N}$ of the input sequence $\{r_{1}^{\mathrm{Text}}, u_{1}', \dots, r_{n}^{\mathrm{Text}}, u_{n}'\}$ and the images $\{r_{1}^{\mathrm{Image}}, \dots, r_{n}^{\mathrm{Image}}\}$ , we minimize the negative log-likelihood: + +$$ +\mathcal {L} (w) = \sum_ {j = 1} ^ {N} \log P \left(w _ {j} \mid w _ {< j}, \left\{r _ {i} ^ {\text {I m a g e}} \right\} _ {i = 1} ^ {n}; \mathcal {G}, \mathcal {V}\right). \tag {1} +$$ + +With such training, our model can generate textual image descriptions during inference with citation tags that reflect the objects needing consistency. + +# 3.3 Inference Procedure: Bridging the Image History for Consistent Image Response + +This section outlines our inference procedure, which employs Customized Text-to-Image Model [23] in conjunction with citation-augmented textual image descriptions (Figure 7). This bridges the image history $r_{1:t-1}^{\mathrm{Image}}$ and image response $r_t^{\mathrm{Image}}$ (as depicted in Figure 6) allowing for consistent generation of the image response. + +Inference For an incoming dialogue context $D = \{(r_i^{\mathrm{Text}}, r_i^{\mathrm{Image}})\}_{i=1}^{t-1}$ , we initially construct + +![](images/43ac84ee766bcd4ed4b8efdde64500e5f29bbd01860ac7023f516fde318cb8f8.jpg) +Fig.6: Bridging the Image History to the Image Response. + +![](images/0104f449cfe2e2b9824cbfd435d51e66e118a4ff90939f235dabff6c85ef0488.jpg) +Fig. 7: Overall Inference Procedure. During inference, the Textual Response Generator $\mathcal{G}$ generates the citation-augmented textual image description $u_{t}^{\prime}$ . By utilizing its citation tag $c_{t}$ , object consistency can be maintained by feeding all preceding image responses with the identical citation tag into a Customized Text-to-Image Model $\mathcal{F}$ . + +the sequence of textual image descriptions and corresponding text responses as $\{r_1^{\mathrm{Text}}, u_1, \ldots, r_{t-1}^{\mathrm{Text}}, u_{t-1}\}$ . Our Citation Module $\mathcal{C}$ then augments the textual image descriptions with citation tags to create $\{r_1^{\mathrm{Text}}, u_1', \ldots, r_{t-1}^{\mathrm{Text}}, u_{t-1}'\}$ , which serves as the input for the Textual Response Generator $\mathcal{G}$ . During inference, if $\mathcal{G}$ predicts an [IMG] token, it initiates the generation of a citation-augmented textual image description $u_t'$ . We extract the citation $c_t$ and the core image description $u_t$ from $u_t'$ . Subsequently, the image description $u_t$ is input to our Customized Text-to-Image Model $\mathcal{F}$ , along with all preceding image responses sharing the same citation $c_t$ as follows: + +$$ +r _ {t} ^ {\mathrm {I m a g e}} = \mathcal {F} \left(u _ {t} \mid \{r _ {i} ^ {\mathrm {I m a g e}} \mid c _ {i} = c _ {t} \} _ {i = 1} ^ {t - 1}\right). \tag {2} +$$ + +This approach ensures the consistent generation of specific objects across the conversation. + +# 4 Experiments + +This section presents the implementation details, evaluation benchmarks, and the experimental results of our approach. Section 4.1 outlines the implementation specifics. In Section 4.2, we assess the overall quality of our model against the standard benchmarks established in previous works [9, 47]. Section 4.3 is dedicated to evaluating the image grounding effectiveness of our model. Finally, Section 4.4 examines the consistency of the image responses generated by our system. + +# 4.1 Experimental Setup + +Dataset As mentioned in Section 2.1, our benchmark datasets include ImageChat [42], PhotoChat [47], and MMDialog [9]. Specifically, for the overall + +multimodal dialogue evaluation in Section 4.2, we train and evaluate using the PhotoChat and MMDialog. For the image grounding evaluation in Section 4.3, we use ImageChat. For the image consistency evaluation in Section 4.4, we create the MDIC dataset by hand-labeling a subset of MMDialog test set. For Section 4.3 and Section 4.4, we use the model trained on MMDialog from Section 4.2. + +Implementation Details For the Textual Dialogue Response Generator $\mathcal{G}$ and the Visual Encoder $\mathcal{V}$ , we initialize with the pre-trained OpenFlamingo 4B model $^3$ [4]. During the fine-tuning phase, we employ special tokens to structure our inputs and outputs: [IMG] and [/IMG] encapsulate textual image descriptions, while [EOT] signifies the end of conversation turns. The BLIP2-flan-t5-xl model [24] is used for converting image responses to corresponding textual image descriptions. Additionally, our Citation Module $\mathcal{C}$ uses [CITE] and [/CITE] tokens to mark the beginning and end of citations linked to key objects within the textual image descriptions. + +As noted in Section 3.2, the Citation Module $\mathcal{C}$ is composed of four key components: the POS tagging processor $(\mathcal{P})$ , the open-set object detector $(\mathcal{M})$ , the segmentation model $(\mathcal{S})$ , and the visual feature extractor $(\mathcal{E})$ . Specifically, we employ spaCy [13] for $\mathcal{P}$ , GroundingDino [28] for $\mathcal{M}$ , the Segment Anything Model (SAM) [15] for $\mathcal{S}$ , and DINOv2 [33] for $\mathcal{E}$ . The similarity threshold $\tau$ is set to 0.6. No further learning is done for these modules; they are utilized as pre-trained components within our system for citation tagging. For our customized text-to-image generation model, $\mathcal{F}$ , we used BLIP-Diffusion [23] when conditioning on the input image and the standard Stable Diffusion 2.1 [38] when there is no input image conditioning. + +Learning Details Let us denote $\{\theta_{\mathcal{V}},\theta_{\mathcal{G}_v},\theta_{\mathcal{G}_l}\}$ as the parameters of the perceiver resampler of the Visual Encoder $\nu$ , the visual cross-attention layers of the Textual Dialogue Response Generator $\mathcal{G}$ , and the language model layers of $\mathcal{G}$ , respectively. + +In the first stage of training, we train $\theta_{\mathcal{G}_l}$ . The batch size is set to 256 with a maximum token length set to 256. In the second stage of training, we jointly train $\theta_{\mathcal{V}}$ and $\theta_{\mathcal{G}_v}$ . The batch size is set to 128 with a maximum token length set to 512. Both the first and second stage is trained by minimizing the next token prediction loss (Eq. 1) using the AdamW optimizer [29] with a learning rate set to 1e-4. The trainings were conducted using $16 \times$ NVIDIA A100 80GB PCIe. + +# 4.2 Multimodal Dialogue Evaluation + +Evaluation Dataset We evaluate the overall performance of our BI-MDRG system on the test set of PhotoChat [47] and MMDialog [9] dataset. PhotoChat contains a single image per dialogue, while the MMDialog includes dialogues with multiple images across turns, offering a more complex context for assessing + +Table 1: Automatic evaluation results of BI-MDRG on the PhotoChat and the MM-Dialog test set. Numbers in bold represent the best scores. + +
ModelIntent F1Image Response ISTextual B1Image Description B2R-1R-LText Response
B1B2R1R-L
PhotoChat Dataset [47]
BERT-base [47]53.2---------
T5-3B [47]58.9---------
Divter [43]56.215.815.111.4-15.86.521.66-5.69
DivterLLM54.116.141.327.143.341.611.44.7511.210.8
BI-MDRG55.716.742.128.244.642.512.45.1212.111.2
MMDialog Dataset [9]
PaCE [25]77.6---------
Divter [9]71.820.5----9.447.45-11.2
MiniGPT5 [50]-20.2----29.119.5-12.1
DivterLLM67.321.044.235.745.543.621.316.220.419.4
BI-MDRG70.522.452.244.753.251.627.623.525.724.8
+ +our system's capability. We perform the tests for all turns except the first turn of each dialogue and consider all previous turns as context. + +Evaluation Metric The performance evaluation is carried out using automatic metrics across four key dimensions: (1) Image Intent Prediction - assessing the need for an image response in the current turn; (2) Image Response Quality; (3) Textual Image Description Quality; and (4) Text Response Generation. We employ the F1 metric for Image Intent Prediction, following the binary classification approach detailed in [9, 43]. BLEU [34] and ROUGE [26] metrics are used for evaluating both the Textual Image Descriptions and Text Responses. The Image Response Quality is measured using the Inception Score (IS) [41], in line with [9]. + +Baselines The BI-MDRG system is compared against Divter [43], a prior Multimodal Dialogue Response Generator that utilizes DialoGPT (762M) [48] as its language model backbone. For a fair comparison, we also train Divter on the same backbone Language Model as our BI-MDRG, denoted as $\mathrm{Divter}_{\mathrm{LLM}}$ (3B). Moreover, for the MMDialog dataset, we compare with MiniGPT5 [50] which uses a 9B backbone VLM. For Intent-prediction baselines, we also include BERT-base, T5-3B reported in [47] for the PhotoChat dataset and PaCE reported in [25] for the MMDialog dataset. + +Results Table 1 presents the evaluation results, where our BI-MDRG system demonstrates outstanding performance on both datasets. In Textual Image Description and Text Response Generation, BI-MDRG achieves the highest BLEU and ROUGE metrics scores in most cases, indicating the system's proficiency in generating relevant and coherent image responses and text responses. Notably, + +BI-MDRG, which uses a 4B backbone VLM, outperforms MiniGPT5, which utilizes a 9B backbone VLM. + +# 4.3 Image Grounding Evaluation + +Evaluation Dataset To evaluate the ability to generate image-grounded textual responses, we utilize the ImageChat dataset [42], consisting of dialogues centered around a single image (an example is shown in Appendix C.4). Since the conversations are grounded in an image, the dataset is suitable for the evaluation of the image grounding capability of the model. We use the same model trained on the MMDialog dataset in Section 4.2 without further tuning on the ImageChat train set. + +Evaluation Metric In order to evaluate the image grounding capability, we use BLEU and ROUGE as the metrics. We perform the evaluation only on the last turn of each dialogue and consider all previous turns as the input context. + +Results The performance of our BI-MDRG system on the ImageChat benchmark is summarized in Table 2. These results demonstrate the system's superior image grounding capabilities in text responses. BI-MDRG achieves the highest scores in BLEU-1, ROUGE-1, and ROUGE-L metrics, indicating its effectiveness in generating contextually relevant and coherent text responses grounded in the visual elements of the conversation. In comparison to DivterLLM + +and BI-MDRG $_{\mathrm{w/o mask}}$ , our BI-MDRG model shows an improvement, highlighting the significance of our proposed architectural enhancements and the multimodal causal attention mask in understanding and integrating visual context into text responses effectively. + +Moreover, we show an example of the model prediction between the Divter $_{\text{LLM}}$ and our BI-MDRG in Figure 8-(a). Figure 8-(a) (right-top) shows that Divter $_{\text{LLM}}$ produces contextually irrelevant text response due to its lack of image understanding. However, Figure 8-(a) (right-bottom) shows that since our BI-MDRG has access to the image, it can produce a text response grounded on the image. + +Table 2: Automatic evaluation results on the ImageChat test set. Numbers in bold represent the best scores. + +
ModelB1R-1R-L
ImageChat Dataset [42]
DivterLLM8.610.39.6
BI-MDRGw/o mask10.011.110.2
BI-MDRG10.911.710.9
+ +# 4.4 Image Response Consistency Evaluation + +We integrated the Citation Module into our BI-MDRG system to improve consistency in image responses within the dialogue. This module assigns citation tags to primary objects in textual image descriptions, tracking their presence across + +![](images/1d8cf49b30fa7b47b5c0510d2738d84c156c7af343ac58c23cc6935f7f301a0d.jpg) +(a) Enhanced Image Grounded Text Response + +![](images/166b2d24bdda312aa2d9a8fec5129b8ce8e4b320934159d8740b6dfbb7102fbf.jpg) +(b) Enhanced Consistency of Image Response + +![](images/77093220b173eb862eb9eedb7e970696f8f95cc63fbe9570bf8637e52aea3de8.jpg) +Fig. 8: Example of predictions from Divter $_{\text{LLM}}$ [43] vs. BI-MDRG (ours). Additional examples can be found in Appendix E. +(a) Citation Module Evaluation + +![](images/5fe84f13a5cf58fb0d7d07f879f86630cc4838439cf799a0b7db173e12debde9.jpg) +(b) Citation Token Prediction Evaluation + +![](images/6468590d985393f283360dc290a6ec3c00fb2a28cdb2211de3668c6b494ce6a6.jpg) +(c) Image Consistency Evaluation +Fig. 9: Three evaluation aspects for image response consistency. (a) We assess the accuracy of the pseudo-labels of citation tags created by the Citation Module. (b) We evaluate how well the model learns from these pseudo-labels and predict the citation tag $\hat{c}_t$ , given the dialogue history, current text response $r_t^{\mathrm{Text}}$ , and $\tilde{u}_t$ which denotes the textual image description leading up to the primary object word $o_t$ . (c) We evaluate the consistency of the resulting image response $\hat{r}_t^{\mathrm{Image}}$ . + +the dialogue. As shown in Figure 9, we focus on three aspects for evaluating the image response consistency: (a) Citation Module Evaluation, (b) Citation Token Prediction Evaluation, and (c) Image Consistency Evaluation. + +Multimodal Dialogue Image Consistency (MDIC) Dataset Creation Due to the absence of benchmark datasets for evaluating the image consistency in multimodal dialogue scenarios, we created a dataset of 300 dialogues, each manually annotated to assign correct citation tags to objects based on their appearances in image responses. We use this created dataset for the subsequent evaluations. Details of the annotation process can be found in Appendix C.5. + +Citation Module Evaluation Figure 9-(a) shows the evaluation method for the Citation Module. As described in Section 3.2, the module $\mathcal{C}$ processes visual features $\{f_1,\ldots ,f_n\}$ corresponding to primary objects $\{o_1,\dots,o_n\}$ in image descriptions. The module generates citations based on the cosine similarity between these features. We constructed a similarity matrix $\mathcal{X}$ , where each element $\mathcal{X}_{ij}$ represents the cosine similarity between the features $f_{i}$ and $f_{j}$ . A feature pair $(f_i,f_j)$ is deemed 'similar' if $\mathcal{X}_{ij} \geq \tau$ and 'not similar' otherwise, + +where $\tau$ is a predetermined threshold. The module's performance was quantified using the F1 score, and we achieved an F1 score of 0.72, demonstrating the module's high accuracy in assigning pseudo-labels of citation tags for training. + +Citation Token Prediction Evaluation Our model, trained with citation-augmented textual image descriptions, can generate these descriptions with accurate citation tags for the primary objects in the descriptions. Figure 9-(b) shows the evaluation method where we assess our model's ability to predict the citation tags correctly. We input the ground truth sequence $\{r_1^{\mathrm{Text}},u_1',\ldots ,r_t^{\mathrm{Text}},\tilde{u}_t\}$ from the MDIC dataset along with the image re + +sponse history $\{r_1^{\mathrm{Image}},\ldots ,r_{t - 1}^{\mathrm{Image}}\}$ . Here, $\tilde{u}_t$ denotes the textual image description leading up to the principal object word $o_t$ . The goal is to verify whether the model's predicted citation $\hat{c}_t$ aligns with the actual citation $c_t$ . The accuracy (Acc.) of our citation token prediction is presented in Table 3. We compare with a baseline denoted as Divter $_{\mathrm{LLM}} + \mathrm{LLMCite}$ where we assign citation tags by instructing an external LLM (Details in Appendix D). The results show that BI-MDRG achieves $84.0\%$ accuracy compared to $33.5\%$ of the LLMCite baseline, showing the effectiveness of our model in identifying objects needing consistency. + +Table 3: Citation Token Prediction (Acc.) and Image Consistency Evaluation (DINOv2). Numbers in bold represent the best scores. + +
ModelAcc.DINOv2(↑)
MDIC Dataset
DivterLLM + LLMCite33.50.32
BI-MDRG (ours)84.00.53
+ +Image Consistency Evaluation Figure 9-(c) shows the evaluation method for the Image Consistency Evaluation. To assess the consistency of the image response, we employ metrics commonly found in the literature on custom object text-to-image generation [7, 10, 18, 19, 39]. DINO features outperform CLIP in capturing image alignment, specifically in distinguishing different objects of the same class [39]. Therefore, for our evaluation, we measure image alignment scores using the DINOv2 [33] features. + +In our proposed framework, correctly using the citation tag is essential to ensure image consistency in the conversation. This involves correctly predicting a new citation tag or referencing an existing tag in history. Therefore, our evaluation includes not only scenarios where the model's predicted citation tag aligns with the ground truth $(c_{t} = \hat{c}_{t})$ and $c_{t}$ exists in previous dialogue history, but also instances where the model's predicted citation tag do not match the ground truth $(c_{t} \neq \hat{c}_{t})$ and either $\hat{c}_{t}$ or $c_{t}$ is in previous dialogue history. + +We report the DINOv2 score as an average of both the successful and failed citation cases mentioned above, where successful prediction positively influences the score, and failed prediction reduces it, offering a comprehensive measure of the model's performance in image consistency. Table 3 shows that our method successfully maintains the consistency of the image response in a dialogue setting, with the DINOv2 score achieving 0.53 compared to 0.32 of baseline. Furthermore, + +Table 4: Evaluation of the citation framework on various settings. The image consistency is measured using DINOv2 (↑) on MDIC dataset, and TID (Textual Image Description) and TR (Text Response) are measured on MMDialog [9] test set. Custom uses customized text-to-image generation with image conditioning, while Diffusion employs a standard text-to-image diffusion model without conditioning. The citation framework allows to selectively use between these for consistent image response. The first row represents the performance of BI-MDRG (ours). + +
CitationVLM SizeDiffusionDINOv2TIDTR
B1B2R-1R-LB1B2R-1
Citation Module4BCustom+Diffusion0.5352.244.753.251.627.623.525.7
LLMCite x4BCustom+Diffusion0.3452.044.953.351.327.423.425.8
4BDiffusion0.2552.044.953.351.327.423.425.8
LLMCite x9BCustom+Diffusion0.3356.048.754.552.130.325.728.8
9BDiffusion0.2656.048.754.552.130.325.728.8
+ +in Figure 8-(b), we present a comparative example of enhanced image consistency of prediction from BI-MDRG compared to Divter $_{\text{LLM}}$ . + +# 5 Importance of Citation Tags for Image Consistency + +Due to powerful pre-trained models for text-to-text and text-to-image, adopting text as an intermediary for image responses is a practical solution for the MDRG [43] task. However, due to the inherent information loss of images during this process, achieving image consistency is infeasible without a targeted framework for consistency maintenance. + +Table 4 shows image consistency and dialogue response performance across various settings. Not using citations shows similar dialogue response performance (Intent, TID, TR) compared to LLMCite and our Citation Module. However, LLMCite improves image consistency (DINOv2) from 0.25 to 0.34 (4B) and 0.26 to 0.33 (9B). Our Citation Module further boosts this from 0.34 to 0.53, indicating its importance for image consistency without affecting dialogue response performance. Notably, scaling the model size (4B to 9B) improves textual response but fails to maintain image consistency without our citation framework, as also evident by ChatGPT's shortcomings in Appendix E. + +# 6 Conclusion + +This paper presents BI-MDRG, a novel framework for Multimodal Dialogue Response Generation (MDRG) aimed at bridging the image history for enhanced text and image response. Our model's innovative use of image history to inform both text and image responses addresses fundamental limitations in previous methodologies, particularly in maintaining consistency in multimodal interactions. The effectiveness of BI-MDRG has been demonstrated through rigorous evaluations using multiple benchmark datasets and a custom-annotated dataset. + +# Acknowledgement + +This work was supported by a grant of the KAIST-KT joint research project through AI2X Lab., Tech Innovation Group, funded by KT (No. D23000019, Developing Visual and Language Capabilities for AI-Based Dialogue Systems), and by Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2021-0-01381, Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments). + +# References + +1. Alamri, H., Cartillier, V., Das, A., Wang, J., Cherian, A., Essa, I., Batra, D., Marks, T.K., Hori, C., Anderson, P., Lee, S., Parikh, D.: Audio-visual scene-aware dialog. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2019) +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=EbMuimAbPbs +3. Anonymous: Chatsearch: a dataset and a generative retrieval model for general conversational image retrieval (2023), https://openreview.net/forum?id=0unbjYPmbC +4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023) +5. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Zhang, Q., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., Karras, T., Liu, M.Y.: ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) +6. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023) +7. Cohen, N., Gal, R., Meirom, E.A., Chechik, G., Atzmon, Y.: "this is my unicorn, fluffy": Personalizing frozen vision-language representations. In: European Conference on Computer Vision. pp. 558-577. Springer (2022) +8. Das, A., Kottur, S., Gupta, K., Singh, A., Yadav, D., Moura, J.M., Parikh, D., Batra, D.: Visual dialog. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 326-335 (2017) +9. Feng, J., Sun, Q., Xu, C., Zhao, P., Yang, Y., Tao, C., Zhao, D., Lin, Q.: MMDialog: A large-scale multi-turn dialogue dataset towards multi-modal open-domain conversation. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 7348-7363. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-long.405, https://aclanthology.org/2023.acl-long.405 + +10. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618 +11. Gan, Z., Cheng, Y., Kholy, A., Li, L., Liu, J., Gao, J.: Multi-step reasoning via recurrent dual attention for visual dialog. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 6463-6474. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1648, https://aclanthology.org/P19-1648 +12. Han, S., Hessel, J., Dziri, N., Choi, Y., Yu, Y.: Champagne: Learning real-world conversation from large-scale web videos. arXiv preprint arXiv:2303.09713 (2023) +13. Honnibal, M., Montani, I.: spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing (2017), to appear +14. Kim, J., Yoon, S., Kim, D., Yoo, C.D.: Structured co-reference graph attention for video-grounded dialogue (2021), https://arxiv.org/abs/2103.13361 +15. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) +16. Koh, J.Y., Salakhutdinov, R., Fried, D.: Grounding language models to images for multimodal inputs and outputs. ICML (2023) +17. Koo, G., Yoon, S., Yoo, C.D.: Wavelet-guided acceleration of text inversion in diffusion-based image editing. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4380-4384. IEEE (2024) +18. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion (2023) +19. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion (2023) +20. Lee, N., Shin, S., Choo, J., Choi, H.J., Myaeng, S.H.: Constructing multi-modal dialogue dataset by replacing text with semantically relevant images. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). pp. 897-906. Association for Computational Linguistics, Online (Aug 2021), https://aclanthology.org/2021.acl-short.113 +21. Lee, Y.J., Ko, B., Kim, H.G., Choi, H.J.: Dialogcc: Large-scale multi-modal dialogue dataset. arXiv preprint arXiv:2212.04119 (2022) +22. Levy, M., Ben-Ari, R., Darshan, N., Lischinski, D.: Chatting makes perfect: Chat-based image retrieval. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 61437-61449. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/c1b3d1e2cf53bb28cabd801bd58b3521-Paper-Conference.pdf +23. Li, D., Li, J., Hoi, S.C.H.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing (2023) +24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) + +25. Li, Y., Hui, B., Yin, Z., Yang, M., Huang, F., Li, Y.: PaCE: Unified multi-modal dialogue pre-training with progressive and compositional experts. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 13402-13416. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-long.749, https://aclanthology.org/2023.acl-long.749 +26. Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out. pp. 74-81. Association for Computational Linguistics, Barcelona, Spain (Jul 2004), https://aclanthology.org/W04-1013 +27. Lin, H., Ruan, L., Xia, W., Liu, P., Wen, J., Xu, Y., Hu, D., Song, R., Zhao, W.X., Jin, Q., et al.: Tiktalk: A video-based dialogue dataset for multi-modal chitchat in real world. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1303-1313 (2023) +28. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations +30. Ma, J., Liang, J., Chen, C., Lu, H.: Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410 (2023) +31. Meng, Y., Wang, S., Han, Q., Sun, X., Wu, F., Yan, R., Li, J.: Openvidual: A large-scale, open-domain dialogue dataset with visual contexts. arXiv preprint arXiv:2012.15015 (2020) +32. Niu, Y., Zhang, H., Zhang, M., Zhang, J., Lu, Z., Wen, J.R.: Recursive visual attention in visual dialog. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6679-6688 (2019) +33. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) +34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: A method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting on Association for Computational Linguistics. p. 311-318. ACL '02, Association for Computational Linguistics, USA (2002). https://doi.org/10.3115/1073083.1073135, https://doi.org/10.3115/1073083.1073135 +35. Poria, S., Hazarika, D., Majumder, N., Naik, G., Cambria, E., Mihalcea, R.: MELD: A multimodal multi-party dataset for emotion recognition in conversations. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 527-536. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1050, https://aclanthology.org/P19-1050 +36. Qi, J., Niu, Y., Huang, J., Zhang, H.: Two causal principles for improving visual dialog. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10860-10869 (2020) +37. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners (2019) + +38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) +39. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) +40. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=08Yk-n512A1 +41. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X., Chen, X.: Improved techniques for training gans. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 29. Curran Associates, Inc. (2016), https://proceedings.neurips.cc/paper_files/paper/2016/file/8a3363abe792db2d8761d6403605aab7-Paper.pdf +42. Shuster, K., Humeau, S., Bordes, A., Weston, J.: Image-chat: Engaging grounded conversations. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 2414-2429. Association for Computational Linguistics, Online (Jul 2020). https://doi.org/10.18653/v1/2020.acl-main.219, https://aclanthology.org/2020.acl-main.219 +43. Sun, Q., Wang, Y., Xu, C., Zheng, K., Yang, Y., Hu, H., Xu, F., Zhang, J., Geng, X., Jiang, D.: Multimodal dialogue response generation. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2854-2866. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.204, https://aclanthology.org/2022.acl-long.204 +44. Wang, S., Meng, Y., Li, X., Sun, X., Ouyang, R., Li, J.: Openvidual 2.0: A larger-scale, open-domain dialogue generation dataset with visual contexts. arXiv preprint arXiv:2109.12761 (2021) +45. Yoon, S., Kim, D., Yoon, E., Yoon, H.S., Kim, J., Yoo, C.D.: Hear: Hearing enhanced audio response for video-grounded dialogue. arXiv preprint arXiv:2312.09736 (2023) +46. Yoon, S., Yoon, E., Yoon, H.S., Kim, J., Yoo, C.: Information-theoretic text hallucination reduction for video-grounded dialogue. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 4182-4193. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.280, https://aclanthology.org/2022.emnlp-main.280 +47. Zang, X., Liu, L., Wang, M., Song, Y., Zhang, H., Chen, J.: PhotoChat: A human-human dialogue dataset with photo sharing behavior for joint image-text modeling. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 6142-6152. Association for Computational Linguis + +tics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.479, https://aclanthology.org/2021.acl-long.479 +48. Zhang, Y., Sun, S., Galley, M., Chen, Y.C., Brockett, C., Gao, X., Gao, J., Liu, J., Dolan, B.: Dialogpt: Large-scale generative pre-training for conversational response generation. In: ACL, system demonstration (2020) +49. Zhao, J., Zhang, T., Hu, J., Liu, Y., Jin, Q., Wang, X., Li, H.: M3ED: Multi-modal multi-scene multi-label emotional dialogue database. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 5699-5710. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.391, https://aclanthology.org/2022.acl-long.391 +50. Zheng, K., He, X., Wang, X.E.: Minigpt-5: Interleaved vision-and-language generation via generative vocens. arXiv preprint arXiv:2310.02239 (2023) +51. Zheng, Y., Chen, G., Liu, X., Sun, J.: MMChat: Multi-modal chat dataset on social media. In: Calzolari, N., Bechet, F., Blache, P., Choukri, K., Cieri, C., Declerck, T., Goggi, S., Isahara, H., Maegaard, B., Mariani, J., Mazo, H., Odijk, J., Piperidis, S. (eds.) Proceedings of the Thirteenth Language Resources and Evaluation Conference. pp. 5778-5786. European Language Resources Association, Marseille, France (Jun 2022), https://aclanthology.org/2022.lrec-1.621 \ No newline at end of file diff --git a/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/images.zip b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d2b92063320ba3491aa944a88025294d35f8ca1c --- /dev/null +++ b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f7a74597121c593d914f29b100b2d87e98780897c13cd494c787c7f94c2aede +size 412857 diff --git a/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/layout.json b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..701dd31233f9fd1f432f9f7cf9eb83e5e35cbbf5 --- /dev/null +++ b/2024/BI-MDRG_ Bridging Image History in Multimodal Dialogue Response Generation/layout.json @@ -0,0 +1,11432 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "BI-MDRG: Bridging Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": "Hee Suk Yoon" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Eunseop Yoon" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Joshua Tian Jin Tee" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Kang Zhang" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Yu-Jung Heo" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Du-Seong Chang" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", and Chang D. Yoo" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 201, + 438, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 201, + 438, + 213 + ], + "spans": [ + { + "bbox": [ + 173, + 201, + 438, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 201, + 438, + 213 + ], + "type": "text", + "content": " Korea Advanced Institute of Science and Technology (KAIST)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "spans": [ + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "type": "text", + "content": " KT Corporation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 224, + 440, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 224, + 440, + 235 + ], + "spans": [ + { + "bbox": [ + 173, + 224, + 440, + 235 + ], + "type": "text", + "content": "{hskyoon, esyoon97, joshuateetj, zhangkang, cd_yoo}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 236, + 360, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 236, + 360, + 246 + ], + "spans": [ + { + "bbox": [ + 255, + 236, + 360, + 246 + ], + "type": "text", + "content": "{yj.heo,dschang}@kt.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 270, + 455, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 270, + 455, + 491 + ], + "spans": [ + { + "bbox": [ + 159, + 270, + 455, + 491 + ], + "type": "text", + "content": "Abstract. Multimodal Dialogue Response Generation (MDRG) is a recently proposed task where the model needs to generate responses in texts, images, or a blend of both based on the dialogue context. Due to the lack of a large-scale dataset specifically for this task and the benefits of leveraging powerful pre-trained models, previous work relies on the text modality as an intermediary step for both the image input and output of the model rather than adopting an end-to-end approach. However, this approach can overlook crucial information about the image, hindering 1) image-grounded text response and 2) consistency of objects in the image response. In this paper, we propose BI-MDRG that bridges the response generation path such that the image history information is utilized for enhanced relevance of text responses to the image content and the consistency of objects in sequential image responses. Through extensive experiments on the multimodal dialogue benchmark dataset, we show that BI-MDRG can effectively increase the quality of multimodal dialogue. Additionally, recognizing the gap in benchmark datasets for evaluating the image consistency in multimodal dialogue, we have created a curated set of 300 dialogues annotated to track object consistency across conversations. The code and the dataset is publicly available at https://github.com/hee-suk-yoon/BI-MDRG." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 500, + 453, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 500, + 453, + 523 + ], + "spans": [ + { + "bbox": [ + 159, + 500, + 453, + 523 + ], + "type": "text", + "content": "Keywords: Multimodal Dialogue " + }, + { + "bbox": [ + 159, + 500, + 453, + 523 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 500, + 453, + 523 + ], + "type": "text", + "content": " Image Grounding " + }, + { + "bbox": [ + 159, + 500, + 453, + 523 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 500, + 453, + 523 + ], + "type": "text", + "content": " Image Consistency" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 541, + 230, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 230, + 554 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 230, + 554 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 564, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 638 + ], + "type": "text", + "content": "With the development of instant messaging technology, visual modalities are increasingly used alongside text in online communication. To enhance user interaction with intelligent agents, a new task, Multimodal Dialogue Response Generation (MDRG) [43], has been proposed. This task requires models to generate both text and image responses based on dialogue history containing texts and images. Since learning an effective multimodal generation model with a sin" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 643, + 223, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 223, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 223, + 653 + ], + "type": "text", + "content": "* Equal Contribution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "type": "text", + "content": "\\*\\* Corresponding Author" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 117, + 225, + 210 + ], + "blocks": [ + { + "bbox": [ + 136, + 117, + 225, + 210 + ], + "lines": [ + { + "bbox": [ + 136, + 117, + 225, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 117, + 225, + 210 + ], + "type": "image", + "image_path": "75f9b85ffddd660a717389a9fcac18ecb8a61899dbad8137abf85be598972259.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 212, + 221, + 228 + ], + "lines": [ + { + "bbox": [ + 136, + 212, + 221, + 228 + ], + "spans": [ + { + "bbox": [ + 136, + 212, + 221, + 228 + ], + "type": "text", + "content": "(a) Overall framework of previous MDRG system" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 231, + 115, + 328, + 212 + ], + "blocks": [ + { + "bbox": [ + 231, + 115, + 328, + 212 + ], + "lines": [ + { + "bbox": [ + 231, + 115, + 328, + 212 + ], + "spans": [ + { + "bbox": [ + 231, + 115, + 328, + 212 + ], + "type": "image", + "image_path": "e8d04309ebe22b6b530bc8e25fc1dcc85d57332f800708fbe0baebf1eaa9a2b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 213, + 339, + 228 + ], + "lines": [ + { + "bbox": [ + 230, + 213, + 339, + 228 + ], + "spans": [ + { + "bbox": [ + 230, + 213, + 339, + 228 + ], + "type": "text", + "content": "(b) Loss of Image History Information when generating text and image response." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 335, + 117, + 408, + 210 + ], + "blocks": [ + { + "bbox": [ + 335, + 117, + 408, + 210 + ], + "lines": [ + { + "bbox": [ + 335, + 117, + 408, + 210 + ], + "spans": [ + { + "bbox": [ + 335, + 117, + 408, + 210 + ], + "type": "image", + "image_path": "1cd0ceb2f7bea0c731501004a2fdcd1fbcf2235fa4a0b5fdff99546ff8594b83.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "text", + "content": "Fig. 1: (a) Outlines the framework of previous Multimodal Dialogue Response Generation (MDRG) systems, which uses the textual descriptions of images " + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "inline_equation", + "content": "(u_{t})" + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "text", + "content": " as an intermediary step toward generating image responses " + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "inline_equation", + "content": "(r_{t}^{\\mathrm{Image}})" + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "text", + "content": ". (b) Highlights the limitations of these systems, particularly their failure to fully leverage image history " + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "inline_equation", + "content": "(r_{1:t-1}^{\\mathrm{Image}})" + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "text", + "content": " in crafting both the textual response " + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "inline_equation", + "content": "(r_{t}^{\\mathrm{Text}})" + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "text", + "content": " and the image response " + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "inline_equation", + "content": "(r_{t}^{\\mathrm{Image}})" + }, + { + "bbox": [ + 130, + 239, + 482, + 319 + ], + "type": "text", + "content": ". (c) illustrates the consequences of this oversight, including responses that lack grounding in image context and consistency in image-based replies." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 408, + 118, + 480, + 210 + ], + "blocks": [ + { + "bbox": [ + 358, + 214, + 457, + 224 + ], + "lines": [ + { + "bbox": [ + 358, + 214, + 457, + 224 + ], + "spans": [ + { + "bbox": [ + 358, + 214, + 457, + 224 + ], + "type": "text", + "content": "(c) Problems of Previous MDRG system" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 408, + 118, + 480, + 210 + ], + "lines": [ + { + "bbox": [ + 408, + 118, + 480, + 210 + ], + "spans": [ + { + "bbox": [ + 408, + 118, + 480, + 210 + ], + "type": "image", + "image_path": "d460dab270f987011460d6970a27df4d199b9a97f88de4980c66ffcd8c390ce9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "text", + "content": "gle sequence-to-sequence model requires a large number of training instances, the lack of a large-scale multimodal dialogue dataset poses a significant challenge for the task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "text", + "content": "To solve this issue, the previous MDRG model [43] disentangles the textual response and image response generation, as illustrated in Figure 1-(a). Specifically, given the dialogue history " + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "text", + "content": "-th conversation turn, the MDRG model first generates the textual response " + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "inline_equation", + "content": "r_t^{\\mathrm{Text}}" + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "text", + "content": " and the intermediary textual image description " + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "text", + "content": ". This image description " + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "text", + "content": " is subsequently fed into a text-to-image model to generate the corresponding image " + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "inline_equation", + "content": "r_t^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 392, + 482, + 500 + ], + "type": "text", + "content": ". Such an approach allows the utilization of powerful models pre-trained on vast amounts of available data for text-to-text and text-to-image pairs, bypassing the need for a large-scale multimodal dialogue dataset for direct end-to-end training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 504, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 590 + ], + "type": "text", + "content": "Nonetheless, due to its reliance on text as an intermediary representation of images, the previous MDRG model overlooks crucial information about the image, as shown in Figure 1-(b). By converting images from the dialogue history into textual descriptions " + }, + { + "bbox": [ + 130, + 504, + 482, + 590 + ], + "type": "inline_equation", + "content": "(u_{1:t-1})" + }, + { + "bbox": [ + 130, + 504, + 482, + 590 + ], + "type": "text", + "content": ", these models fail to fully utilize the rich visual content in the actual images " + }, + { + "bbox": [ + 130, + 504, + 482, + 590 + ], + "type": "inline_equation", + "content": "(r_{1:t-1}^{\\mathrm{Image}})" + }, + { + "bbox": [ + 130, + 504, + 482, + 590 + ], + "type": "text", + "content": ". This leads to two major issues: a lack of image-grounded context in textual responses (Figure 1-(c) (left)) and inconsistencies in image generation across dialogues (Figure 1-(c) (right))." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "For instance, Figure 1-(c) (left) shows the existing MDRG model failing to provide an image-grounded response to 'What breed is your dog?' because it only perceives the dog through the text 'a dog eating a watermelon'. Figure 1-(c) (right) illustrates the model's inconsistency, where the 'dog' in the image history is not maintained in the generated response. These issues highlight the need for improved MDRG models that effectively utilize image history." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "HS. Yoon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "Contribution This paper introduces BI-MDRG to enhance text and image responses by bridging the image history." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 144, + 482, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 144, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 144, + 482, + 215 + ], + "type": "text", + "content": "- Bridging Image History in Text Responses: In Section 3.1, we propose an architectural modification in which the visual features, extracted using a visual encoder, are integrated into the cross-attention layers of the core language model. This is complemented by a novel multimodal causal attention mask modulation tailored for the MDRG task to improve image-grounded textual responses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 216, + 482, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 216, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 139, + 216, + 482, + 335 + ], + "type": "text", + "content": "- Bridging Image History in Image Responses: In Section 3.2, we propose a citation framework where we use the Citation Module designed to augment textual image descriptions by tagging objects with citation tags that identify recurring objects throughout a dialogue. In Section 3.3, we show that training with these augmented data enables the model to recognize and maintain the consistency of objects in subsequent image responses during inference using Customized Text-to-Image Models. Due to the absence of benchmark datasets for evaluating image consistency in multimodal dialogues, we have created the Multimodal Dialogue Image Consistency (MDIC) dataset containing dialogues annotated to track object consistency across conversations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 350, + 237, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 237, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 237, + 363 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 373, + 313, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 313, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 313, + 384 + ], + "type": "text", + "content": "2.1 Multimodal Dialogue Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 390, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 481, + 461 + ], + "type": "text", + "content": "Multimodal dialogue datasets generally fall into three categories: question and answering [1,8] (where the task involves asking and answering questions about a specific image), in-scene [12,31,35,44,49] (where each dialogue turn corresponds to a scene from a movie or video), and conversation-based [9,20,21,27,42,47] (which engage in natural dialogue about a given image or involves image sharing within natural conversations). Further details are in Appendix C.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 462, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 482, + 569 + ], + "type": "text", + "content": "This paper primarily explores the model's capability for natural dialogue within the conversation-based category. Notable datasets in this segment include ImageChat [42], PhotoChat [47], MMDD [20], DialogCC [21], MMChat [51], TikTalk [27], and MMDialog [9]. Given our focus on English-language scenarios, we exclude MMChat and Tiktalk from our evaluation, as it is a dataset primarily in Chinese. Additionally, DialogCC (not publicly available) and MMDD, which are synthesized by algorithmically pairing images with text-only dialogues for random turns, are also excluded from our analysis. Therefore, our evaluation is centered on ImageChat, PhotoChat, and MMDialog." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 482, + 665 + ], + "type": "text", + "content": "ImageChat [42] consists of image-centered dialogues, where each dialogue is centered around a single given image. PhotoChat [47] features dialogues collected from social media, where a single image is shared in one of the conversation turns, which better mirrors everyday human interaction. Still, their limited scale and domain diversity restrict their applicability. Overcoming these limitations, MMDialog [9] features over a million diverse dialogues from social media, where multiple images are shared across numerous conversation turns, providing a more realistic representation of open-domain multimodal conversations." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Briding Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 318, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 318, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 318, + 129 + ], + "type": "text", + "content": "2.2 Multimodal Dialogue Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 255 + ], + "type": "text", + "content": "Pioneering studies [11, 32, 36] have delved into improving the performance of image-grounded conversational agents. While [14, 45, 46, 51] introduced a Seq2Seq based model focusing on multimodal dialogues, it primarily generated textual responses, not fully embracing the multimodal response scenario. In a notable advancement, [43] presented Divter, which not only produces informative text but also generates high-resolution images, marking a significant leap forward in multimodal dialogue response generation (MDRG). It is important to note that our focus in this paper is generation-based models, contrasting with retrieval-based [3, 16, 22], which output image responses by retrieving existing images from a corpus instead of generating new ones." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 270, + 294, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 294, + 283 + ], + "type": "text", + "content": "2.3 Customized Text-to-Image" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 385 + ], + "type": "text", + "content": "Recent studies on text-to-image diffusion models [5, 17, 38, 40] focus on customization [7, 10, 18, 39], learning specific concepts from a few images. Following this, users can flexibly generate the learned concepts into new scenes. Text inversion [10] generates varied contexts for a concept by updating the text embedding without altering the model. Dreambooth [39] and Custom Diffusion [18] fine-tune the U-Net architecture using an identifier, class label, and images. In a notable enhancement, BLIP-Diffusion [23] enables zero-shot subject-driven generation, allowing fast customized text-to-image generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 401, + 479, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 401, + 479, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 479, + 430 + ], + "type": "text", + "content": "3 BI-MDRG: Bridging the Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "type": "text", + "content": "We introduce BI-MDRG, a conversational agent designed to produce both textual and visual responses with enhanced awareness of image history. Sections 3.1 and 3.2 detail the training procedure, effectively integrating image history information into text responses and textual image descriptions. Section 3.3 outlines the inference process, wherein the image history informs the image responses by leveraging the captured details from enhanced textual image descriptions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 527, + 481, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 481, + 540 + ], + "type": "text", + "content": "3.1 Bridging the Image History for Image-Grounded Text Response" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 351, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 351, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 351, + 666 + ], + "type": "text", + "content": "The example shown in Figure 1-(c) (left) highlighted one of the crucial limitations in the previous MDRG system: their reliance on textual descriptions for understanding image history, which hinders the image-grounded textual responses. To overcome this, we have adopted an architectural change along with a multimodal causal attention mask modulation to effectively bridge the image history information " + }, + { + "bbox": [ + 130, + 545, + 351, + 666 + ], + "type": "inline_equation", + "content": "r_{1:t-1}^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 545, + 351, + 666 + ], + "type": "text", + "content": " to the text response " + }, + { + "bbox": [ + 130, + 545, + 351, + 666 + ], + "type": "inline_equation", + "content": "r_t^{\\mathrm{Text}}" + }, + { + "bbox": [ + 130, + 545, + 351, + 666 + ], + "type": "text", + "content": " (as depicted in Figure 2)." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 374, + 545, + 467, + 639 + ], + "blocks": [ + { + "bbox": [ + 374, + 545, + 467, + 639 + ], + "lines": [ + { + "bbox": [ + 374, + 545, + 467, + 639 + ], + "spans": [ + { + "bbox": [ + 374, + 545, + 467, + 639 + ], + "type": "image", + "image_path": "d67d544ddc906675d546ef7170d529ff8bc6fe5f82de2b5bb2592bc52056686b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 647, + 481, + 670 + ], + "lines": [ + { + "bbox": [ + 356, + 647, + 481, + 670 + ], + "spans": [ + { + "bbox": [ + 356, + 647, + 481, + 670 + ], + "type": "text", + "content": "Fig.2: Bridging Image History to the Text Response." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "HS. Yoon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 463, + 248 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 463, + 248 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 463, + 248 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 463, + 248 + ], + "type": "image", + "image_path": "96ac44f7b984c60534151a2f0a80dfdaea3a98ce27b5b186fa9bc9761a69d60f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "text", + "content": "Fig. 3: Training of BI-MDRG. (a) Textual Dialogue Response Generator " + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "text", + "content": " cross-attends to the image features from the Visual Encoder " + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "text", + "content": ". (b) Attention Mask Modulation alters the causal attention to prioritize image features over textual image descriptions. (c) Citation Module " + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "text", + "content": " generates citation-augmented textual image descriptions, enabling the tracking of objects within image history for consistency maintenance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": "Architecture Our Textual Dialogue Response Generator " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " (Figure 3-(a)), consists of a decoder-only language model with added visual cross-attention layers. These layers directly engage with image features provided by the Visual Encoder " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": ", drawing inspiration from Flamingo [2], to reduce dependence on textual image descriptions for perceiving images. In our framework, a dialogue context " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "D = \\{(r_i^{\\mathrm{Text}}, r_i^{\\mathrm{Image}})\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " comprises multiple turns, each with an associated text response " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "r_i^{\\mathrm{Text}}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " and an image response " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "r_i^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": ". An image captioning model produces textual descriptions " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " for each image " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "r_i^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": ", which are then transformed into citation-augmented descriptions " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "u_i'" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " as shown in Figure 3-(c) and further detailed in Section 3.2 (if there is no image response for a turn, " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "r_i^{\\mathrm{Image}} = \\emptyset" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "u_i' = \\emptyset" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": "). These descriptions and the text responses " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\{r_1^{\\mathrm{Text}}, u_1', \\ldots, r_n^{\\mathrm{Text}}, u_n'\\}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " are fed into " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": ", while the images " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\{r_1^{\\mathrm{Image}}, \\ldots, r_n^{\\mathrm{Image}}\\}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " are processed by " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " to extract image features which are fed to the cross-attention layers in " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": ". Although our model directly cross-attends to the inputted image features, we retain the textual descriptions " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "u_{1:n}'" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " as essential inputs since we require the generation of textual description by our " + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 338, + 482, + 559 + ], + "type": "text", + "content": " model, which subsequently gets used by a text-to-image model for constructing the image response. Keeping the textual description inputs allows efficient teacher-forced next token prediction training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "text", + "content": "Multimodal Causal Attention Mask Modulation For the input sequence " + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{r_1^{\\mathrm{Text}}, u_1', \\ldots, r_n^{\\mathrm{Text}}, u_n'\\}" + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "text", + "content": ", we use a specialized mask along with the standard causal mask (Figure 3-(b)). The traditional causal mask allows each text response " + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "inline_equation", + "content": "r_i^{\\mathrm{Text}}" + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "text", + "content": " to access previous textual image descriptions " + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "inline_equation", + "content": "u_{1:i-1}'" + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "text", + "content": ", leading to reliance on textual information over visual context. Our masking strategy prevents " + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "inline_equation", + "content": "r_i^{\\mathrm{Text}}" + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "text", + "content": " from accessing " + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "inline_equation", + "content": "u_{1:i-1}'" + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "text", + "content": ", redirecting focus to the actual image features of " + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "inline_equation", + "content": "r_{1:i-1}^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 580, + 482, + 666 + ], + "type": "text", + "content": ", ensuring text responses are grounded on raw image features." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "type": "text", + "content": "Briding Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 478, + 174 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 478, + 174 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 478, + 174 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 478, + 174 + ], + "type": "image", + "image_path": "c950ce3eb160bc479415c7845d7b2d0efd442bf2637b77a13e9117f7781fe267.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 182, + 482, + 216 + ], + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 216 + ], + "type": "text", + "content": "Fig. 4: Illustration of the Citation Module. Citation Module recognizes identical objects within image history and injects this information into the textual image description with citation tags (e.g., [cite]0[/cite])." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 236, + 468, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 468, + 261 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 468, + 261 + ], + "type": "text", + "content": "3.2 Citation Module: Bridging the Image History to the Textual Image Description" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 337, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 337, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 337, + 387 + ], + "type": "text", + "content": "The example shown in Figure 1-(c) (right) highlighted another critical limitation of the previous MDRG system: its inability to ensure consistency in image responses. To address this, we propose the Citation Module (Figure 3-(c)) to bridge the image history " + }, + { + "bbox": [ + 130, + 266, + 337, + 387 + ], + "type": "inline_equation", + "content": "r_{1:t-1}^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 266, + 337, + 387 + ], + "type": "text", + "content": " and the textual image description " + }, + { + "bbox": [ + 130, + 266, + 337, + 387 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 130, + 266, + 337, + 387 + ], + "type": "text", + "content": " (as depicted in Figure 5) by ensuring that the textual image description accurately relays which objects should persist in subsequent images." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 366, + 251, + 460, + 342 + ], + "blocks": [ + { + "bbox": [ + 366, + 251, + 460, + 342 + ], + "lines": [ + { + "bbox": [ + 366, + 251, + 460, + 342 + ], + "spans": [ + { + "bbox": [ + 366, + 251, + 460, + 342 + ], + "type": "image", + "image_path": "dd64c838406fde07f3b821a888ccf254647639771228009cc870e00cae4adecb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 351, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 342, + 351, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 342, + 351, + 482, + 374 + ], + "type": "text", + "content": "Fig. 5: Bridging Image History to the Textual Image Description." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 480, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 462 + ], + "type": "text", + "content": "Citation Module Citation Module plays a pivotal role in tracking recurring objects in the dialogue using textual image descriptions. For instance, descriptions like \"a dog is in front of a fireplace\" and \"a dog running in the snow\" are augmented to \"a dog[cite]0[/cite] is in front of a fireplace\" and \"a dog[cite]0[/cite] running in the snow,\" respectively if they reference the same dog." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Motivated by [30], Figure 4 details the citation process for textual image descriptions " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{u_1, \\dots, u_n\\}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ". For each textual image description " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ", a Part of Speech (POS) tagging processor " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " is employed to tag words and pinpoint the key object word " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " in the description. The word " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ", along with its corresponding image " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "r_i^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ", is processed through an open-set object detector " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " to obtain the bounding box of the detected object, which is then input to a segmentation model " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " for generating object segmentation mask " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ". These masks are applied to isolate the objects from their backgrounds in " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "r_i^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " since background removal has been proven helpful for better extraction of object features [6]. These isolated objects are then analyzed by a visual feature extractor " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " to extract features " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ". The resulting feature set " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{f_1, \\dots, f_n\\}" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " undergoes clustering based on cosine similarity to identify identical objects across images, as outlined in Algorithm 1; this involves assigning each element a cluster id " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " based on the similarity of their features. For each " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ", we augment the word so that it is followed by its corresponding cluster id (i.e., citation tag) " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ", resulting in the citation augmented textual description " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "u_i'" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ". This Citation Module operates with off-the-shelf components for citation tags, requiring no training on the target dataset." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "HS. Yoon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 481, + 322 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 272, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 272, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 272, + 128 + ], + "type": "text", + "content": "Algorithm 1 Citation Module" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": "Require: image response " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "(r_1^{\\mathrm{Image}},\\dots r_n^{\\mathrm{Image}})" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " , textual image description " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "(u_{1},\\ldots u_{n})" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " , similarity threshold " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " , POS tagging processor " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " , open-set object detector " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " , segmentation model " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " , visual feature extractor " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " , object cluster dictionary " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\kappa = \\{\\}" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "i\\gets 1" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " Identify principal object word in " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "o_i\\gets \\mathcal{P}(u_i)" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " Obtain bounding box for " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "box_{i}\\gets \\mathcal{M}(o_{i},r_{i}^{\\mathrm{Image}})" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " Obtain segmentation mask for " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "s_i\\gets S(box_i)" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " Extract visual feature for " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "f_{i}\\gets \\mathcal{V}(s_{i}\\odot r_{i}^{\\mathrm{Image}})" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " \nend for \ncluster_id " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\leftarrow 0" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "i\\gets 1" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " do Perform clustering of " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " . if not " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " has key(o) then " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{K}[o_i]\\gets \\overline{\\mathsf{clust}} e r\\_ i d" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " Get's own cluster. cluster_id " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " cluster_id+1 \nend if \nfor " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "j\\gets i + 1" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " do \nif " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\operatorname {SIM}(f_i,f_j)\\geq \\tau" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " and not " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " has key(oj) then " + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\kappa [o_j]\\gets \\kappa [o_i]" + }, + { + "bbox": [ + 132, + 130, + 481, + 322 + ], + "type": "text", + "content": " Puts into existing cluster. end if \nend for \nend for \nreturn K" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 344, + 479, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 479, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 479, + 406 + ], + "type": "text", + "content": "Generative Training Objective We use the next token prediction training via teacher forcing, which is used in standard auto-regressive language models [37]. Specifically, given the token sequence " + }, + { + "bbox": [ + 130, + 344, + 479, + 406 + ], + "type": "inline_equation", + "content": "w = \\{w_{j}\\}_{j=1}^{N}" + }, + { + "bbox": [ + 130, + 344, + 479, + 406 + ], + "type": "text", + "content": " of the input sequence " + }, + { + "bbox": [ + 130, + 344, + 479, + 406 + ], + "type": "inline_equation", + "content": "\\{r_{1}^{\\mathrm{Text}}, u_{1}', \\dots, r_{n}^{\\mathrm{Text}}, u_{n}'\\}" + }, + { + "bbox": [ + 130, + 344, + 479, + 406 + ], + "type": "text", + "content": " and the images " + }, + { + "bbox": [ + 130, + 344, + 479, + 406 + ], + "type": "inline_equation", + "content": "\\{r_{1}^{\\mathrm{Image}}, \\dots, r_{n}^{\\mathrm{Image}}\\}" + }, + { + "bbox": [ + 130, + 344, + 479, + 406 + ], + "type": "text", + "content": ", we minimize the negative log-likelihood:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 414, + 480, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 414, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 208, + 414, + 480, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (w) = \\sum_ {j = 1} ^ {N} \\log P \\left(w _ {j} \\mid w _ {< j}, \\left\\{r _ {i} ^ {\\text {I m a g e}} \\right\\} _ {i = 1} ^ {n}; \\mathcal {G}, \\mathcal {V}\\right). \\tag {1}", + "image_path": "7d517e6dc580c0b8ac2d80d9a2598c6e8fa7b4bee21f265aa3339c83f50bc6c9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 455, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 480, + 479 + ], + "type": "text", + "content": "With such training, our model can generate textual image descriptions during inference with citation tags that reflect the objects needing consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 495, + 479, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 479, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 479, + 520 + ], + "type": "text", + "content": "3.3 Inference Procedure: Bridging the Image History for Consistent Image Response" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 525, + 336, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 336, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 336, + 624 + ], + "type": "text", + "content": "This section outlines our inference procedure, which employs Customized Text-to-Image Model [23] in conjunction with citation-augmented textual image descriptions (Figure 7). This bridges the image history " + }, + { + "bbox": [ + 130, + 525, + 336, + 624 + ], + "type": "inline_equation", + "content": "r_{1:t-1}^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 525, + 336, + 624 + ], + "type": "text", + "content": " and image response " + }, + { + "bbox": [ + 130, + 525, + 336, + 624 + ], + "type": "inline_equation", + "content": "r_t^{\\mathrm{Image}}" + }, + { + "bbox": [ + 130, + 525, + 336, + 624 + ], + "type": "text", + "content": " (as depicted in Figure 6) allowing for consistent generation of the image response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 336, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 336, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 336, + 666 + ], + "type": "text", + "content": "Inference For an incoming dialogue context " + }, + { + "bbox": [ + 132, + 641, + 336, + 666 + ], + "type": "inline_equation", + "content": "D = \\{(r_i^{\\mathrm{Text}}, r_i^{\\mathrm{Image}})\\}_{i=1}^{t-1}" + }, + { + "bbox": [ + 132, + 641, + 336, + 666 + ], + "type": "text", + "content": ", we initially construct" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 359, + 517, + 469, + 626 + ], + "blocks": [ + { + "bbox": [ + 359, + 517, + 469, + 626 + ], + "lines": [ + { + "bbox": [ + 359, + 517, + 469, + 626 + ], + "spans": [ + { + "bbox": [ + 359, + 517, + 469, + 626 + ], + "type": "image", + "image_path": "43ac84ee766bcd4ed4b8efdde64500e5f29bbd01860ac7023f516fde318cb8f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 635, + 480, + 658 + ], + "lines": [ + { + "bbox": [ + 342, + 635, + 480, + 658 + ], + "spans": [ + { + "bbox": [ + 342, + 635, + 480, + 658 + ], + "type": "text", + "content": "Fig.6: Bridging the Image History to the Image Response." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Briding Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 480, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 228 + ], + "type": "image", + "image_path": "0104f449cfe2e2b9824cbfd435d51e66e118a4ff90939f235dabff6c85ef0488.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "text", + "content": "Fig. 7: Overall Inference Procedure. During inference, the Textual Response Generator " + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "text", + "content": " generates the citation-augmented textual image description " + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "inline_equation", + "content": "u_{t}^{\\prime}" + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "text", + "content": ". By utilizing its citation tag " + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "text", + "content": ", object consistency can be maintained by feeding all preceding image responses with the identical citation tag into a Customized Text-to-Image Model " + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": "the sequence of textual image descriptions and corresponding text responses as " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\{r_1^{\\mathrm{Text}}, u_1, \\ldots, r_{t-1}^{\\mathrm{Text}}, u_{t-1}\\}" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": ". Our Citation Module " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": " then augments the textual image descriptions with citation tags to create " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\{r_1^{\\mathrm{Text}}, u_1', \\ldots, r_{t-1}^{\\mathrm{Text}}, u_{t-1}'\\}" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": ", which serves as the input for the Textual Response Generator " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": ". During inference, if " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": " predicts an [IMG] token, it initiates the generation of a citation-augmented textual image description " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "u_t'" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": ". We extract the citation " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": " and the core image description " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "u_t'" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": ". Subsequently, the image description " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": " is input to our Customized Text-to-Image Model " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": ", along with all preceding image responses sharing the same citation " + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 219, + 421, + 481, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 421, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 219, + 421, + 481, + 442 + ], + "type": "interline_equation", + "content": "r _ {t} ^ {\\mathrm {I m a g e}} = \\mathcal {F} \\left(u _ {t} \\mid \\{r _ {i} ^ {\\mathrm {I m a g e}} \\mid c _ {i} = c _ {t} \\} _ {i = 1} ^ {t - 1}\\right). \\tag {2}", + "image_path": "84dba54cd3845a95f29b23873410e6fc60d31074ccc1d76fce63faaadc12e4e7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "text", + "content": "This approach ensures the consistent generation of specific objects across the conversation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 493, + 230, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 230, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 230, + 506 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 518, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 602 + ], + "type": "text", + "content": "This section presents the implementation details, evaluation benchmarks, and the experimental results of our approach. Section 4.1 outlines the implementation specifics. In Section 4.2, we assess the overall quality of our model against the standard benchmarks established in previous works [9, 47]. Section 4.3 is dedicated to evaluating the image grounding effectiveness of our model. Finally, Section 4.4 examines the consistency of the image responses generated by our system." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 620, + 262, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 262, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 262, + 634 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Dataset As mentioned in Section 2.1, our benchmark datasets include ImageChat [42], PhotoChat [47], and MMDialog [9]. Specifically, for the overall" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "HS. Yoon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "multimodal dialogue evaluation in Section 4.2, we train and evaluate using the PhotoChat and MMDialog. For the image grounding evaluation in Section 4.3, we use ImageChat. For the image consistency evaluation in Section 4.4, we create the MDIC dataset by hand-labeling a subset of MMDialog test set. For Section 4.3 and Section 4.4, we use the model trained on MMDialog from Section 4.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "text", + "content": "Implementation Details For the Textual Dialogue Response Generator " + }, + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "text", + "content": " and the Visual Encoder " + }, + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "text", + "content": ", we initialize with the pre-trained OpenFlamingo 4B model" + }, + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "text", + "content": " [4]. During the fine-tuning phase, we employ special tokens to structure our inputs and outputs: [IMG] and [/IMG] encapsulate textual image descriptions, while [EOT] signifies the end of conversation turns. The BLIP2-flan-t5-xl model [24] is used for converting image responses to corresponding textual image descriptions. Additionally, our Citation Module " + }, + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 194, + 482, + 302 + ], + "type": "text", + "content": " uses [CITE] and [/CITE] tokens to mark the beginning and end of citations linked to key objects within the textual image descriptions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": "As noted in Section 3.2, the Citation Module " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": " is composed of four key components: the POS tagging processor " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "(\\mathcal{P})" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", the open-set object detector " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "(\\mathcal{M})" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", the segmentation model " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "(\\mathcal{S})" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", and the visual feature extractor " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "(\\mathcal{E})" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ". Specifically, we employ spaCy [13] for " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", GroundingDino [28] for " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", the Segment Anything Model (SAM) [15] for " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", and DINOv2 [33] for " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ". The similarity threshold " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": " is set to 0.6. No further learning is done for these modules; they are utilized as pre-trained components within our system for citation tagging. For our customized text-to-image generation model, " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", we used BLIP-Diffusion [23] when conditioning on the input image and the standard Stable Diffusion 2.1 [38] when there is no input image conditioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": "Learning Details Let us denote " + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\{\\theta_{\\mathcal{V}},\\theta_{\\mathcal{G}_v},\\theta_{\\mathcal{G}_l}\\}" + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": " as the parameters of the perceiver resampler of the Visual Encoder " + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": ", the visual cross-attention layers of the Textual Dialogue Response Generator " + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": ", and the language model layers of " + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 440, + 482, + 487 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "text", + "content": "In the first stage of training, we train " + }, + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}_l}" + }, + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "text", + "content": ". The batch size is set to 256 with a maximum token length set to 256. In the second stage of training, we jointly train " + }, + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}_v}" + }, + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "text", + "content": ". The batch size is set to 128 with a maximum token length set to 512. Both the first and second stage is trained by minimizing the next token prediction loss (Eq. 1) using the AdamW optimizer [29] with a learning rate set to 1e-4. The trainings were conducted using " + }, + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 130, + 488, + 482, + 559 + ], + "type": "text", + "content": " NVIDIA A100 80GB PCIe." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 578, + 326, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 578, + 326, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 578, + 326, + 590 + ], + "type": "text", + "content": "4.2 Multimodal Dialogue Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": "Evaluation Dataset We evaluate the overall performance of our BI-MDRG system on the test set of PhotoChat [47] and MMDialog [9] dataset. PhotoChat contains a single image per dialogue, while the MMDialog includes dialogues with multiple images across turns, offering a more complex context for assessing" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Briding Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 466, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 466, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 466, + 666 + ], + "type": "text", + "content": "3 https://huggingface.co/openflamingo/OpenFlamingo-4B-vitl-rpj3b-langinstruct" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 146, + 470, + 293 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Automatic evaluation results of BI-MDRG on the PhotoChat and the MM-Dialog test set. Numbers in bold represent the best scores." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 146, + 470, + 293 + ], + "lines": [ + { + "bbox": [ + 144, + 146, + 470, + 293 + ], + "spans": [ + { + "bbox": [ + 144, + 146, + 470, + 293 + ], + "type": "table", + "html": "
ModelIntent F1Image Response ISTextual B1Image Description B2R-1R-LText Response
B1B2R1R-L
PhotoChat Dataset [47]
BERT-base [47]53.2---------
T5-3B [47]58.9---------
Divter [43]56.215.815.111.4-15.86.521.66-5.69
DivterLLM54.116.141.327.143.341.611.44.7511.210.8
BI-MDRG55.716.742.128.244.642.512.45.1212.111.2
MMDialog Dataset [9]
PaCE [25]77.6---------
Divter [9]71.820.5----9.447.45-11.2
MiniGPT5 [50]-20.2----29.119.5-12.1
DivterLLM67.321.044.235.745.543.621.316.220.419.4
BI-MDRG70.522.452.244.753.251.627.623.525.724.8
", + "image_path": "da7d86fd2271b296f0ffcd36700de4eeb85d9a6677a98d112aa26958b3807e8e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 316, + 479, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 340 + ], + "type": "text", + "content": "our system's capability. We perform the tests for all turns except the first turn of each dialogue and consider all previous turns as context." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 360, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 469 + ], + "type": "text", + "content": "Evaluation Metric The performance evaluation is carried out using automatic metrics across four key dimensions: (1) Image Intent Prediction - assessing the need for an image response in the current turn; (2) Image Response Quality; (3) Textual Image Description Quality; and (4) Text Response Generation. We employ the F1 metric for Image Intent Prediction, following the binary classification approach detailed in [9, 43]. BLEU [34] and ROUGE [26] metrics are used for evaluating both the Textual Image Descriptions and Text Responses. The Image Response Quality is measured using the Inception Score (IS) [41], in line with [9]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 488, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 586 + ], + "type": "text", + "content": "Baselines The BI-MDRG system is compared against Divter [43], a prior Multimodal Dialogue Response Generator that utilizes DialoGPT (762M) [48] as its language model backbone. For a fair comparison, we also train Divter on the same backbone Language Model as our BI-MDRG, denoted as " + }, + { + "bbox": [ + 130, + 488, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{Divter}_{\\mathrm{LLM}}" + }, + { + "bbox": [ + 130, + 488, + 482, + 586 + ], + "type": "text", + "content": " (3B). Moreover, for the MMDialog dataset, we compare with MiniGPT5 [50] which uses a 9B backbone VLM. For Intent-prediction baselines, we also include BERT-base, T5-3B reported in [47] for the PhotoChat dataset and PaCE reported in [25] for the MMDialog dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Results Table 1 presents the evaluation results, where our BI-MDRG system demonstrates outstanding performance on both datasets. In Textual Image Description and Text Response Generation, BI-MDRG achieves the highest BLEU and ROUGE metrics scores in most cases, indicating the system's proficiency in generating relevant and coherent image responses and text responses. Notably," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "HS. Yoon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "text", + "content": "BI-MDRG, which uses a 4B backbone VLM, outperforms MiniGPT5, which utilizes a 9B backbone VLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 307, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 307, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 307, + 169 + ], + "type": "text", + "content": "4.3 Image Grounding Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 176, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 260 + ], + "type": "text", + "content": "Evaluation Dataset To evaluate the ability to generate image-grounded textual responses, we utilize the ImageChat dataset [42], consisting of dialogues centered around a single image (an example is shown in Appendix C.4). Since the conversations are grounded in an image, the dataset is suitable for the evaluation of the image grounding capability of the model. We use the same model trained on the MMDialog dataset in Section 4.2 without further tuning on the ImageChat train set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 277, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 482, + 314 + ], + "type": "text", + "content": "Evaluation Metric In order to evaluate the image grounding capability, we use BLEU and ROUGE as the metrics. We perform the evaluation only on the last turn of each dialogue and consider all previous turns as the input context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 316, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 316, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 316, + 484 + ], + "type": "text", + "content": "Results The performance of our BI-MDRG system on the ImageChat benchmark is summarized in Table 2. These results demonstrate the system's superior image grounding capabilities in text responses. BI-MDRG achieves the highest scores in BLEU-1, ROUGE-1, and ROUGE-L metrics, indicating its effectiveness in generating contextually relevant and coherent text responses grounded in the visual elements of the conversation. In comparison to DivterLLM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 480, + 533 + ], + "type": "text", + "content": "and BI-MDRG" + }, + { + "bbox": [ + 130, + 485, + 480, + 533 + ], + "type": "inline_equation", + "content": "_{\\mathrm{w/o mask}}" + }, + { + "bbox": [ + 130, + 485, + 480, + 533 + ], + "type": "text", + "content": ", our BI-MDRG model shows an improvement, highlighting the significance of our proposed architectural enhancements and the multimodal causal attention mask in understanding and integrating visual context into text responses effectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": "Moreover, we show an example of the model prediction between the Divter" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "_{\\text{LLM}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": " and our BI-MDRG in Figure 8-(a). Figure 8-(a) (right-top) shows that Divter" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "_{\\text{LLM}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": " produces contextually irrelevant text response due to its lack of image understanding. However, Figure 8-(a) (right-bottom) shows that since our BI-MDRG has access to the image, it can produce a text response grounded on the image." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 325, + 392, + 473, + 468 + ], + "blocks": [ + { + "bbox": [ + 321, + 347, + 482, + 381 + ], + "lines": [ + { + "bbox": [ + 321, + 347, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 321, + 347, + 482, + 381 + ], + "type": "text", + "content": "Table 2: Automatic evaluation results on the ImageChat test set. Numbers in bold represent the best scores." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 392, + 473, + 468 + ], + "lines": [ + { + "bbox": [ + 325, + 392, + 473, + 468 + ], + "spans": [ + { + "bbox": [ + 325, + 392, + 473, + 468 + ], + "type": "table", + "html": "
ModelB1R-1R-L
ImageChat Dataset [42]
DivterLLM8.610.39.6
BI-MDRGw/o mask10.011.110.2
BI-MDRG10.911.710.9
", + "image_path": "1326e66b9cf5f1c8706394f003ed7c2554829cb22c4fb0a64e50f2ecf2724bf4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 610, + 362, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 362, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 362, + 623 + ], + "type": "text", + "content": "4.4 Image Response Consistency Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We integrated the Citation Module into our BI-MDRG system to improve consistency in image responses within the dialogue. This module assigns citation tags to primary objects in textual image descriptions, tracking their presence across" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Briding Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 116, + 304, + 197 + ], + "blocks": [ + { + "bbox": [ + 151, + 116, + 304, + 197 + ], + "lines": [ + { + "bbox": [ + 151, + 116, + 304, + 197 + ], + "spans": [ + { + "bbox": [ + 151, + 116, + 304, + 197 + ], + "type": "image", + "image_path": "1d8cf49b30fa7b47b5c0510d2738d84c156c7af343ac58c23cc6935f7f301a0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 200, + 283, + 208 + ], + "lines": [ + { + "bbox": [ + 180, + 200, + 283, + 208 + ], + "spans": [ + { + "bbox": [ + 180, + 200, + 283, + 208 + ], + "type": "text", + "content": "(a) Enhanced Image Grounded Text Response" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 116, + 462, + 197 + ], + "blocks": [ + { + "bbox": [ + 310, + 116, + 462, + 197 + ], + "lines": [ + { + "bbox": [ + 310, + 116, + 462, + 197 + ], + "spans": [ + { + "bbox": [ + 310, + 116, + 462, + 197 + ], + "type": "image", + "image_path": "166b2d24bdda312aa2d9a8fec5129b8ce8e4b320934159d8740b6dfbb7102fbf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 200, + 444, + 209 + ], + "lines": [ + { + "bbox": [ + 340, + 200, + 444, + 209 + ], + "spans": [ + { + "bbox": [ + 340, + 200, + 444, + 209 + ], + "type": "text", + "content": "(b) Enhanced Consistency of Image Response" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 137, + 252, + 248, + 327 + ], + "blocks": [ + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "lines": [ + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "type": "text", + "content": "Fig. 8: Example of predictions from Divter" + }, + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "type": "inline_equation", + "content": "_{\\text{LLM}}" + }, + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "type": "text", + "content": " [43] vs. BI-MDRG (ours). Additional examples can be found in Appendix E." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 252, + 248, + 327 + ], + "lines": [ + { + "bbox": [ + 137, + 252, + 248, + 327 + ], + "spans": [ + { + "bbox": [ + 137, + 252, + 248, + 327 + ], + "type": "image", + "image_path": "77093220b173eb862eb9eedb7e970696f8f95cc63fbe9570bf8637e52aea3de8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 331, + 234, + 339 + ], + "lines": [ + { + "bbox": [ + 156, + 331, + 234, + 339 + ], + "spans": [ + { + "bbox": [ + 156, + 331, + 234, + 339 + ], + "type": "text", + "content": "(a) Citation Module Evaluation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 256, + 252, + 353, + 329 + ], + "blocks": [ + { + "bbox": [ + 256, + 252, + 353, + 329 + ], + "lines": [ + { + "bbox": [ + 256, + 252, + 353, + 329 + ], + "spans": [ + { + "bbox": [ + 256, + 252, + 353, + 329 + ], + "type": "image", + "image_path": "5fe84f13a5cf58fb0d7d07f879f86630cc4838439cf799a0b7db173e12debde9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 331, + 356, + 339 + ], + "lines": [ + { + "bbox": [ + 253, + 331, + 356, + 339 + ], + "spans": [ + { + "bbox": [ + 253, + 331, + 356, + 339 + ], + "type": "text", + "content": "(b) Citation Token Prediction Evaluation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 360, + 253, + 481, + 331 + ], + "blocks": [ + { + "bbox": [ + 360, + 253, + 481, + 331 + ], + "lines": [ + { + "bbox": [ + 360, + 253, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 360, + 253, + 481, + 331 + ], + "type": "image", + "image_path": "6468590d985393f283360dc290a6ec3c00fb2a28cdb2211de3668c6b494ce6a6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 332, + 466, + 340 + ], + "lines": [ + { + "bbox": [ + 381, + 332, + 466, + 340 + ], + "spans": [ + { + "bbox": [ + 381, + 332, + 466, + 340 + ], + "type": "text", + "content": "(c) Image Consistency Evaluation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "lines": [ + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "text", + "content": "Fig. 9: Three evaluation aspects for image response consistency. (a) We assess the accuracy of the pseudo-labels of citation tags created by the Citation Module. (b) We evaluate how well the model learns from these pseudo-labels and predict the citation tag " + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\hat{c}_t" + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "text", + "content": ", given the dialogue history, current text response " + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "inline_equation", + "content": "r_t^{\\mathrm{Text}}" + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\tilde{u}_t" + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "text", + "content": " which denotes the textual image description leading up to the primary object word " + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "text", + "content": ". (c) We evaluate the consistency of the resulting image response " + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\hat{r}_t^{\\mathrm{Image}}" + }, + { + "bbox": [ + 131, + 351, + 482, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 440, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 440, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 440, + 482, + 477 + ], + "type": "text", + "content": "the dialogue. As shown in Figure 9, we focus on three aspects for evaluating the image response consistency: (a) Citation Module Evaluation, (b) Citation Token Prediction Evaluation, and (c) Image Consistency Evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 492, + 482, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 492, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 482, + 565 + ], + "type": "text", + "content": "Multimodal Dialogue Image Consistency (MDIC) Dataset Creation Due to the absence of benchmark datasets for evaluating the image consistency in multimodal dialogue scenarios, we created a dataset of 300 dialogues, each manually annotated to assign correct citation tags to objects based on their appearances in image responses. We use this created dataset for the subsequent evaluations. Details of the annotation process can be found in Appendix C.5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": "Citation Module Evaluation Figure 9-(a) shows the evaluation method for the Citation Module. As described in Section 3.2, the module " + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": " processes visual features " + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{f_1,\\ldots ,f_n\\}" + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": " corresponding to primary objects " + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{o_1,\\dots,o_n\\}" + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": " in image descriptions. The module generates citations based on the cosine similarity between these features. We constructed a similarity matrix " + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": ", where each element " + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ij}" + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": " represents the cosine similarity between the features " + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": ". A feature pair " + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "(f_i,f_j)" + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": " is deemed 'similar' if " + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ij} \\geq \\tau" + }, + { + "bbox": [ + 131, + 581, + 482, + 667 + ], + "type": "text", + "content": " and 'not similar' otherwise," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "HS. Yoon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": " is a predetermined threshold. The module's performance was quantified using the F1 score, and we achieved an F1 score of 0.72, demonstrating the module's high accuracy in assigning pseudo-labels of citation tags for training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 181, + 317, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 317, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 317, + 312 + ], + "type": "text", + "content": "Citation Token Prediction Evaluation Our model, trained with citation-augmented textual image descriptions, can generate these descriptions with accurate citation tags for the primary objects in the descriptions. Figure 9-(b) shows the evaluation method where we assess our model's ability to predict the citation tags correctly. We input the ground truth sequence " + }, + { + "bbox": [ + 130, + 181, + 317, + 312 + ], + "type": "inline_equation", + "content": "\\{r_1^{\\mathrm{Text}},u_1',\\ldots ,r_t^{\\mathrm{Text}},\\tilde{u}_t\\}" + }, + { + "bbox": [ + 130, + 181, + 317, + 312 + ], + "type": "text", + "content": " from the MDIC dataset along with the image re" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "content": "sponse history " + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\{r_1^{\\mathrm{Image}},\\ldots ,r_{t - 1}^{\\mathrm{Image}}\\}" + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\tilde{u}_t" + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "content": " denotes the textual image description leading up to the principal object word " + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "content": ". The goal is to verify whether the model's predicted citation " + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\hat{c}_t" + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "content": " aligns with the actual citation " + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "content": ". The accuracy (Acc.) of our citation token prediction is presented in Table 3. We compare with a baseline denoted as Divter" + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "inline_equation", + "content": "_{\\mathrm{LLM}} + \\mathrm{LLMCite}" + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "content": " where we assign citation tags by instructing an external LLM (Details in Appendix D). The results show that BI-MDRG achieves " + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "content": " accuracy compared to " + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "inline_equation", + "content": "33.5\\%" + }, + { + "bbox": [ + 130, + 312, + 482, + 410 + ], + "type": "text", + "content": " of the LLMCite baseline, showing the effectiveness of our model in identifying objects needing consistency." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 325, + 242, + 474, + 296 + ], + "blocks": [ + { + "bbox": [ + 321, + 186, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 321, + 186, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 482, + 232 + ], + "type": "text", + "content": "Table 3: Citation Token Prediction (Acc.) and Image Consistency Evaluation (DINOv2). Numbers in bold represent the best scores." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 242, + 474, + 296 + ], + "lines": [ + { + "bbox": [ + 325, + 242, + 474, + 296 + ], + "spans": [ + { + "bbox": [ + 325, + 242, + 474, + 296 + ], + "type": "table", + "html": "
ModelAcc.DINOv2(↑)
MDIC Dataset
DivterLLM + LLMCite33.50.32
BI-MDRG (ours)84.00.53
", + "image_path": "0ab76e6c1b01e3e9d3ac53a9e3eacde4574a6339dd0fe146d5cb42c235c6e559.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": "Image Consistency Evaluation Figure 9-(c) shows the evaluation method for the Image Consistency Evaluation. To assess the consistency of the image response, we employ metrics commonly found in the literature on custom object text-to-image generation [7, 10, 18, 19, 39]. DINO features outperform CLIP in capturing image alignment, specifically in distinguishing different objects of the same class [39]. Therefore, for our evaluation, we measure image alignment scores using the DINOv2 [33] features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": "In our proposed framework, correctly using the citation tag is essential to ensure image consistency in the conversation. This involves correctly predicting a new citation tag or referencing an existing tag in history. Therefore, our evaluation includes not only scenarios where the model's predicted citation tag aligns with the ground truth " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "(c_{t} = \\hat{c}_{t})" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": " exists in previous dialogue history, but also instances where the model's predicted citation tag do not match the ground truth " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "(c_{t} \\neq \\hat{c}_{t})" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": " and either " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\hat{c}_{t}" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": " is in previous dialogue history." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We report the DINOv2 score as an average of both the successful and failed citation cases mentioned above, where successful prediction positively influences the score, and failed prediction reduces it, offering a comprehensive measure of the model's performance in image consistency. Table 3 shows that our method successfully maintains the consistency of the image response in a dialogue setting, with the DINOv2 score achieving 0.53 compared to 0.32 of baseline. Furthermore," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 90, + 447, + 102 + ], + "type": "text", + "content": "Briding Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 196, + 480, + 278 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 4: Evaluation of the citation framework on various settings. The image consistency is measured using DINOv2 (↑) on MDIC dataset, and TID (Textual Image Description) and TR (Text Response) are measured on MMDialog [9] test set. Custom uses customized text-to-image generation with image conditioning, while Diffusion employs a standard text-to-image diffusion model without conditioning. The citation framework allows to selectively use between these for consistent image response. The first row represents the performance of BI-MDRG (ours)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 196, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 133, + 196, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 196, + 480, + 278 + ], + "type": "table", + "html": "
CitationVLM SizeDiffusionDINOv2TIDTR
B1B2R-1R-LB1B2R-1
Citation Module4BCustom+Diffusion0.5352.244.753.251.627.623.525.7
LLMCite x4BCustom+Diffusion0.3452.044.953.351.327.423.425.8
4BDiffusion0.2552.044.953.351.327.423.425.8
LLMCite x9BCustom+Diffusion0.3356.048.754.552.130.325.728.8
9BDiffusion0.2656.048.754.552.130.325.728.8
", + "image_path": "fe856db88f4fb1a43d18c5ede305f6ec65977e4bbc512a20817513628c91b156.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 293, + 479, + 319 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 479, + 319 + ], + "type": "text", + "content": "in Figure 8-(b), we present a comparative example of enhanced image consistency of prediction from BI-MDRG compared to Divter" + }, + { + "bbox": [ + 130, + 293, + 479, + 319 + ], + "type": "inline_equation", + "content": "_{\\text{LLM}}" + }, + { + "bbox": [ + 130, + 293, + 479, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "type": "text" + }, + { + "bbox": [ + 131, + 337, + 459, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 459, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 459, + 353 + ], + "type": "text", + "content": "5 Importance of Citation Tags for Image Consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 425 + ], + "type": "text", + "content": "Due to powerful pre-trained models for text-to-text and text-to-image, adopting text as an intermediary for image responses is a practical solution for the MDRG [43] task. However, due to the inherent information loss of images during this process, achieving image consistency is infeasible without a targeted framework for consistency maintenance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 534 + ], + "type": "text", + "content": "Table 4 shows image consistency and dialogue response performance across various settings. Not using citations shows similar dialogue response performance (Intent, TID, TR) compared to LLMCite and our Citation Module. However, LLMCite improves image consistency (DINOv2) from 0.25 to 0.34 (4B) and 0.26 to 0.33 (9B). Our Citation Module further boosts this from 0.34 to 0.53, indicating its importance for image consistency without affecting dialogue response performance. Notably, scaling the model size (4B to 9B) improves textual response but fails to maintain image consistency without our citation framework, as also evident by ChatGPT's shortcomings in Appendix E." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 554, + 220, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 554, + 220, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 554, + 220, + 567 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "This paper presents BI-MDRG, a novel framework for Multimodal Dialogue Response Generation (MDRG) aimed at bridging the image history for enhanced text and image response. Our model's innovative use of image history to inform both text and image responses addresses fundamental limitations in previous methodologies, particularly in maintaining consistency in multimodal interactions. The effectiveness of BI-MDRG has been demonstrated through rigorous evaluations using multiple benchmark datasets and a custom-annotated dataset." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "HS. Yoon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "content": "This work was supported by a grant of the KAIST-KT joint research project through AI2X Lab., Tech Innovation Group, funded by KT (No. D23000019, Developing Visual and Language Capabilities for AI-Based Dialogue Systems), and by Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2021-0-01381, Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 262, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 262, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 306 + ], + "type": "text", + "content": "1. Alamri, H., Cartillier, V., Das, A., Wang, J., Cherian, A., Essa, I., Batra, D., Marks, T.K., Hori, C., Anderson, P., Lee, S., Parikh, D.: Audio-visual scene-aware dialog. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 307, + 480, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 480, + 382 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=EbMuimAbPbs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "text", + "content": "3. Anonymous: Chatsearch: a dataset and a generative retrieval model for general conversational image retrieval (2023), https://openreview.net/forum?id=0unbjYPmbC" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 404, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 458 + ], + "type": "text", + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "type": "text", + "content": "5. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Zhang, Q., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., Karras, T., Liu, M.Y.: ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "text", + "content": "6. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 524, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 480, + 555 + ], + "type": "text", + "content": "7. Cohen, N., Gal, R., Meirom, E.A., Chechik, G., Atzmon, Y.: \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In: European Conference on Computer Vision. pp. 558-577. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 556, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 588 + ], + "type": "text", + "content": "8. Das, A., Kottur, S., Gupta, K., Singh, A., Yadav, D., Moura, J.M., Parikh, D., Batra, D.: Visual dialog. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 326-335 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 589, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 665 + ], + "type": "text", + "content": "9. Feng, J., Sun, Q., Xu, C., Zhao, P., Yang, Y., Tao, C., Zhao, D., Lin, Q.: MMDialog: A large-scale multi-turn dialogue dataset towards multi-modal open-domain conversation. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 7348-7363. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-long.405, https://aclanthology.org/2023.acl-long.405" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Briding Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "10. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 228 + ], + "type": "text", + "content": "11. Gan, Z., Cheng, Y., Kholy, A., Li, L., Liu, J., Gao, J.: Multi-step reasoning via recurrent dual attention for visual dialog. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 6463-6474. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1648, https://aclanthology.org/P19-1648" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "type": "text", + "content": "12. Han, S., Hessel, J., Dziri, N., Choi, Y., Yu, Y.: Champagne: Learning real-world conversation from large-scale web videos. arXiv preprint arXiv:2303.09713 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "type": "text", + "content": "13. Honnibal, M., Montani, I.: spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing (2017), to appear" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 306 + ], + "type": "text", + "content": "14. Kim, J., Yoon, S., Kim, D., Yoo, C.D.: Structured co-reference graph attention for video-grounded dialogue (2021), https://arxiv.org/abs/2103.13361" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "type": "text", + "content": "15. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "text", + "content": "16. Koh, J.Y., Salakhutdinov, R., Fried, D.: Grounding language models to images for multimodal inputs and outputs. ICML (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 407 + ], + "type": "text", + "content": "17. Koo, G., Yoon, S., Yoo, C.D.: Wavelet-guided acceleration of text inversion in diffusion-based image editing. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4380-4384. IEEE (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 409, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 430 + ], + "type": "text", + "content": "18. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 453 + ], + "type": "text", + "content": "19. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 454, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 481, + 520 + ], + "type": "text", + "content": "20. Lee, N., Shin, S., Choo, J., Choi, H.J., Myaeng, S.H.: Constructing multi-modal dialogue dataset by replacing text with semantically relevant images. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). pp. 897-906. Association for Computational Linguistics, Online (Aug 2021), https://aclanthology.org/2021.acl-short.113" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 521, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 542 + ], + "type": "text", + "content": "21. Lee, Y.J., Ko, B., Kim, H.G., Choi, H.J.: Dialogcc: Large-scale multi-modal dialogue dataset. arXiv preprint arXiv:2212.04119 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 544, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 609 + ], + "type": "text", + "content": "22. Levy, M., Ben-Ari, R., Darshan, N., Lischinski, D.: Chatting makes perfect: Chat-based image retrieval. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 61437-61449. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/c1b3d1e2cf53bb28cabd801bd58b3521-Paper-Conference.pdf" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "text", + "content": "23. Li, D., Li, J., Hoi, S.C.H.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "HS. Yoon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 193 + ], + "type": "text", + "content": "25. Li, Y., Hui, B., Yin, Z., Yang, M., Huang, F., Li, Y.: PaCE: Unified multi-modal dialogue pre-training with progressive and compositional experts. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 13402-13416. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-long.749, https://aclanthology.org/2023.acl-long.749" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 194, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 228 + ], + "type": "text", + "content": "26. Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out. pp. 74-81. Association for Computational Linguistics, Barcelona, Spain (Jul 2004), https://aclanthology.org/W04-1013" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 228, + 481, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 271 + ], + "type": "text", + "content": "27. Lin, H., Ruan, L., Xia, W., Liu, P., Wen, J., Xu, Y., Hu, D., Song, R., Zhao, W.X., Jin, Q., et al.: Tiktalk: A video-based dialogue dataset for multi-modal chitchat in real world. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1303-1313 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 273, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 481, + 306 + ], + "type": "text", + "content": "28. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "text", + "content": "29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "text", + "content": "30. Ma, J., Liang, J., Chen, C., Lu, H.: Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "text", + "content": "31. Meng, Y., Wang, S., Han, Q., Sun, X., Wu, F., Yan, R., Li, J.: Openvidual: A large-scale, open-domain dialogue dataset with visual contexts. arXiv preprint arXiv:2012.15015 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "text", + "content": "32. Niu, Y., Zhang, H., Zhang, M., Zhang, J., Lu, Z., Wen, J.R.: Recursive visual attention in visual dialog. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6679-6688 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 430, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 486 + ], + "type": "text", + "content": "33. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 487, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 542 + ], + "type": "text", + "content": "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: A method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting on Association for Computational Linguistics. p. 311-318. ACL '02, Association for Computational Linguistics, USA (2002). https://doi.org/10.3115/1073083.1073135, https://doi.org/10.3115/1073083.1073135" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 543, + 481, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 608 + ], + "type": "text", + "content": "35. Poria, S., Hazarika, D., Majumder, N., Naik, G., Cambria, E., Mihalcea, R.: MELD: A multimodal multi-party dataset for emotion recognition in conversations. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 527-536. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1050, https://aclanthology.org/P19-1050" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 609, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 642 + ], + "type": "text", + "content": "36. Qi, J., Niu, Y., Huang, J., Zhang, H.: Two causal principles for improving visual dialog. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10860-10869 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "37. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners (2019)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "type": "text", + "content": "Briding Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "text", + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "text", + "content": "39. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 269 + ], + "type": "text", + "content": "40. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=08Yk-n512A1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 270, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 335 + ], + "type": "text", + "content": "41. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X., Chen, X.: Improved techniques for training gans. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 29. Curran Associates, Inc. (2016), https://proceedings.neurips.cc/paper_files/paper/2016/file/8a3363abe792db2d8761d6403605aab7-Paper.pdf" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 336, + 481, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 401 + ], + "type": "text", + "content": "42. Shuster, K., Humeau, S., Bordes, A., Weston, J.: Image-chat: Engaging grounded conversations. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 2414-2429. Association for Computational Linguistics, Online (Jul 2020). https://doi.org/10.18653/v1/2020.acl-main.219, https://aclanthology.org/2020.acl-main.219" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 402, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 468 + ], + "type": "text", + "content": "43. Sun, Q., Wang, Y., Xu, C., Zheng, K., Yang, Y., Hu, H., Xu, F., Zhang, J., Geng, X., Jiang, D.: Multimodal dialogue response generation. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2854-2866. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.204, https://aclanthology.org/2022.acl-long.204" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "44. Wang, S., Meng, Y., Li, X., Sun, X., Ouyang, R., Li, J.: Openvidual 2.0: A larger-scale, open-domain dialogue generation dataset with visual contexts. arXiv preprint arXiv:2109.12761 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "45. Yoon, S., Kim, D., Yoon, E., Yoon, H.S., Kim, J., Yoo, C.D.: Hear: Hearing enhanced audio response for video-grounded dialogue. arXiv preprint arXiv:2312.09736 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 533, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 599 + ], + "type": "text", + "content": "46. Yoon, S., Yoon, E., Yoon, H.S., Kim, J., Yoo, C.: Information-theoretic text hallucination reduction for video-grounded dialogue. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 4182-4193. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.280, https://aclanthology.org/2022.emnlp-main.280" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 600, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 665 + ], + "type": "text", + "content": "47. Zang, X., Liu, L., Wang, M., Song, Y., Zhang, H., Chen, J.: PhotoChat: A human-human dialogue dataset with photo sharing behavior for joint image-text modeling. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 6142-6152. Association for Computational Linguis" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "HS. Yoon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 325 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "text", + "content": "tics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.479, https://aclanthology.org/2021.acl-long.479" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "48. Zhang, Y., Sun, S., Galley, M., Chen, Y.C., Brockett, C., Gao, X., Gao, J., Liu, J., Dolan, B.: Dialogpt: Large-scale generative pre-training for conversational response generation. In: ACL, system demonstration (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 237 + ], + "type": "text", + "content": "49. Zhao, J., Zhang, T., Hu, J., Liu, Y., Jin, Q., Wang, X., Li, H.: M3ED: Multi-modal multi-scene multi-label emotional dialogue database. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 5699-5710. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.391, https://aclanthology.org/2022.acl-long.391" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 238, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 259 + ], + "type": "text", + "content": "50. Zheng, K., He, X., Wang, X.E.: Minigpt-5: Interleaved vision-and-language generation via generative vocens. arXiv preprint arXiv:2310.02239 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 260, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 260, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 260, + 481, + 325 + ], + "type": "text", + "content": "51. Zheng, Y., Chen, G., Liu, X., Sun, J.: MMChat: Multi-modal chat dataset on social media. In: Calzolari, N., Bechet, F., Blache, P., Choukri, K., Cieri, C., Declerck, T., Goggi, S., Isahara, H., Maegaard, B., Mariani, J., Mazo, H., Odijk, J., Piperidis, S. (eds.) Proceedings of the Thirteenth Language Resources and Evaluation Conference. pp. 5778-5786. European Language Resources Association, Marseille, France (Jun 2022), https://aclanthology.org/2022.lrec-1.621" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 91, + 447, + 102 + ], + "type": "text", + "content": "Briding Image History in Multimodal Dialogue Response Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_content_list.json b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..27f7fb2e8d6258fcfbaca64f641070af3aa20c28 --- /dev/null +++ b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_content_list.json @@ -0,0 +1,2071 @@ +[ + { + "type": "text", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bo-Kyeong Kim1, Hyoung-Kyu Song2, Thibault Castells1, and Shinkook Choi1", + "bbox": [ + 258, + 210, + 743, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nota Inc. $^{2}$ Captions Research", + "bbox": [ + 392, + 253, + 609, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bokyeong.kim, thibault, shinkook.choi}@nota.ai, kyu@captions.ai", + "bbox": [ + 251, + 268, + 746, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6b9bcc8e0f2fcb2509332954f1cf0ba0cdc8d89478f5152869bb15fec8b3bb1.jpg", + "image_caption": [ + "(a) Efficient General-purpose T2I" + ], + "image_footnote": [], + "bbox": [ + 217, + 316, + 442, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a880b37aa421bf7c7039a28540b468c220868d370fd1272d98fdae10ecec360.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 316, + 663, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/faeb0988ef932fc94156e39fcdd65b7a628762c3e12c300ec4344a0908708d1e.jpg", + "image_caption": [ + "(b) Efficient Personalized T2I", + "(c) Efficient I2I (Different Styles)", + "Fig. 1: Our compressed model enables efficient (a) zero-shot text-to-image generation, (b) personalized synthesis, (c) image-to-image translation, and (d) mobile deployment. Samples from BK-SDM-Small with $36\\%$ reduced parameters and latency are shown." + ], + "image_footnote": [], + "bbox": [ + 442, + 393, + 663, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b925a38ebba87d72e561973bbc2e826e809aca9bcfe9a5ed2ed5188e1d327b1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 316, + 784, + 409 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8674b31599d7a1041f81bceee447e162397c6b1e78c4ce1cfd30ca333a9c0277.jpg", + "image_caption": [ + "(d) Mobile Deployment" + ], + "image_footnote": [], + "bbox": [ + 665, + 410, + 784, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-to-image (T2I) generation with Stable Diffusion models (SDMs) involves high computing demands due to billion-scale parameters. To enhance efficiency, recent studies have reduced sampling steps and applied network quantization while retaining the original architectures. The lack of architectural reduction attempts may stem from worries over expensive retraining for such massive models. In this work, we uncover the surprising potential of block pruning and feature distillation for low-cost general-purpose T2I. By removing several residual and attention blocks from the U-Net of SDMs, we achieve $30\\% \\sim 50\\%$ reduction in model size, MACs, and latency. We show that distillation retraining is effective even under limited resources: using only 13 A100 days and a tiny dataset, our compact models can imitate the original SDMs (v1.4 and v2.1-base with over 6,000 A100 days). Benefiting from the transferred knowledge, our BK-SDMs deliver competitive results on zero-shot MS-COCO against larger multi-billion parameter models. We further demonstrate the applicability of our lightweight backbones in personalized generation and image-to-image translation. Deployment of our models on edge devices attains 4-second inference. Code and models can be found at: https://github.com/Nota-NetsPresso/BK-SDM.", + "bbox": [ + 259, + 575, + 743, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33d091d3cee1eeb8313e995c1c2da5619ece4832718611e4730fb952929b77d0.jpg", + "image_caption": [ + "Fig. 2: Computation of Stable Diffusion. The denoising U-Net is the main processing bottleneck. THOP [91] is used to measure MACs in generating a $512 \\times 512$ image." + ], + "image_footnote": [], + "bbox": [ + 303, + 145, + 467, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dcff685fa1d1cfe7d2f65477e552672de268f7f4144efe386bcce74762c4bd0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 143, + 699, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 324, + 374, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stable Diffusion models (SDMs) [58,60-62] are one of the most renowned open-source models for text-to-image (T2I) synthesis, and their exceptional capability has begun to be leveraged as a backbone in several text-guided vision applications [2,3,83,87]. SDMs are T2I-specialized latent diffusion models (LDMs) [58], which employ diffusion operations [23,37,76] in a semantically compressed space for compute efficiency. Within a SDM, a U-Net [9,64] performs iterative sampling to progressively denoise a random latent code and is aided by a text encoder [53] and an image decoder [12,81] to produce text-aligned images. This inference process still involves excessive computational requirements (see Fig. 2), which often hinder the utilization of SDMs despite their rapidly growing usage.", + "bbox": [ + 212, + 357, + 787, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate this issue, numerous approaches toward efficient SDMs have been introduced. A pretrained diffusion model is distilled to reduce the number of denoising steps, enabling an identically architectured model with fewer sampling steps [40, 43]. Post-training quantization [24, 34, 74] and implementation optimization [6, 7] methods are also leveraged. However, the removal of architectural elements in large diffusion models remains less explored.", + "bbox": [ + 212, + 508, + 787, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This study unlocks the immense potential of classical architectural compression in attaining smaller and faster diffusion models. We eliminate multiple residual and attention blocks from the U-Net of a SDM and retrain it with feature-level knowledge distillation (KD) [18, 63] for general-purpose T2I. Under restricted training resources, our compact models can mimic the original SDM by leveraging transferred knowledge. Our work effectively reduces the computation of SDM-v1.4 [60] and SDM-v2.1-base [62] while achieving compelling zero-shot results on par with multi-billion parameter models [10, 11, 56]. Our contributions are summarized as follows:", + "bbox": [ + 212, + 599, + 787, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We compress SDMs by removing architectural blocks from the U-Net, achieving up to $51\\%$ reduction in model size and $43\\%$ improvement in latency on CPU and GPU. Previous pruning studies [13,47,84] focused on small models ( $< 100\\mathrm{M}$ parameters) like ResNet50 and DeiT-B, not on foundation models like SDMs ( $>1,000\\mathrm{M} = 1\\mathrm{B}$ ), possibly due to the lack of economic retraining for such large models. Moreover, U-Net architectures are arguably more com", + "bbox": [ + 225, + 748, + 789, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "B.-K. Kim et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "plex due to the necessity of considering skip connections across the network, making the structural block removal inside them not straightforward.", + "bbox": [ + 240, + 146, + 785, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, we first demonstrate the notable benefit of feature distillation for training diffusion models, which enables competitive T2I even with significantly fewer resources (using only 13 A100 days and 0.22M LAION pairs [70]). Considering the vast expense of training SDMs from scratch (surpassing 6,000 A100 days and 2,000M pairs), our study indicates that network compression is a remarkably cost-effective strategy in building compact general-purpose diffusion models.", + "bbox": [ + 227, + 176, + 785, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We show the practicality of our work across various aspects. Our lightweight backbones are readily applicable to customized generation [65] and image-to-image translation [44], effectively lowering finetuning and inference costs. T2I synthesis on Jetson AGX Orin and iPhone 14 using our models takes less than 4 seconds.", + "bbox": [ + 227, + 282, + 785, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We have publicly released our approach, model weights, and source code, motivating subsequent works by other researchers (e.g., block pruning and KD for the SDM-v1 variant [5,72] and SDXL [31, 73]).", + "bbox": [ + 227, + 359, + 785, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 428, + 387, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large T2I diffusion models. By gradually removing noise from corrupted data, diffusion-based generative models [9, 22, 76] enable high-fidelity synthesis with broad mode coverage. Integrating these merits with the advancement of pretrained language models [8, 53, 54] has significantly improved the quality of T2I synthesis. In GLIDE [48] and Imagen [66], a text-conditional diffusion model generates a small image, which is upsampled via super-resolution modules. In DALL-E-2 [55], a text-conditional prior network produces an image embedding, which is transformed into an image via a diffusion decoder and further upscaled into higher resolutions. SDMs [58, 60–62] perform the diffusion modeling in a low-dimensional latent space constructed through a pixel-space autoencoder. We use SDMs as our baseline because of its open-access and gaining popularity over numerous downstream tasks [2, 3, 65, 83].", + "bbox": [ + 212, + 460, + 787, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient diffusion models. Several studies have addressed the slow sampling process. Diffusion-tailored distillation [42, 43, 68] progressively transfers knowledge from a pretrained diffusion model to a fewer-step model with the same architecture. Fast high-order solvers [38, 39, 88] for diffusion ordinary differential equations boost the sampling speed. Complementarily, our network compression approach reduces per-step computation and can be easily integrated with less sampling steps. Leveraging quantization [24, 34, 74] and implementation optimizations [6, 7] for SDMs can also be combined with our compact models for further efficiency.", + "bbox": [ + 212, + 643, + 787, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distillation-based compression. KD enhances the performance of small-size models by exploiting output-level [21, 49] and feature-level [18, 63, 85] information of large source models. Although this classical KD has been actively used for efficient GANs [33, 57, 86], its power has not been explored for structurally compressed diffusion models. Distillation pretraining enables small yet capable", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "bbox": [ + 264, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "general-purpose language models [26, 69, 77] and vision transformers [16, 80]. Beyond such models, we show that its success can be extended to diffusion models with iterative sampling.", + "bbox": [ + 212, + 146, + 785, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concurrent studies. SnapFusion [35] and MobileDiffusion [89] achieve an efficient U-Net for SDMs through architecture optimization and step reduction. Würstchen [50] introduces two diffusion processes on low- and high-resolution latent spaces for economic training. These works are valuable but require much larger training resources than our work (see Tab. 1). In contrast, we utilize considerably fewer resources, leveraging the surprising benefits of classical KD for foundational diffusion models. While not demonstrated on SDMs, Diff-Pruning [14] proposes structured pruning based on Taylor expansion tailored for diffusion models. We emphasize the use of depth (block) pruning in our work, which often leads to greater enhancements in inference speeds compared to the width (channel) pruning approach of Diff-Pruning. Moreover, Tab. 5 employs a comparable Taylor pruning criterion, which results in an insufficient decrease in MACs.", + "bbox": [ + 212, + 191, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Compression Method", + "text_level": 1, + "bbox": [ + 214, + 398, + 460, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compress the U-Net [64] in SDMs, which is the most compute-heavy component (see Fig. 2). Conditioned on the text and time-step embeddings, the U-Net performs multiple denoising steps on latent representations. At each step, the U-Net produces the noise residual to compute the latent for the next step. We reduce this per-step computation, leading to Block-removed Knowledge-distilled SDMs (BK-SDMs).", + "bbox": [ + 212, + 431, + 787, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Compact U-Net Architecture", + "text_level": 1, + "bbox": [ + 214, + 547, + 504, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The following architectures are obtained by compressing SDM-v1 (1.04B parameters), as shown in Fig. 3:", + "bbox": [ + 212, + 574, + 785, + 604 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\circ$ BK-SDM-Base (0.76B) obtained with Sec. 3.1.(1).", + "$\\circ$ BK-SDM-Small (0.66B) with Secs. 3.1.(1) and 3.1.(2).", + "$\\circ$ BK-SDM-Tiny (0.50B) with Secs. 3.1.(1), 3.1.(2), and 3.1.(3)." + ], + "bbox": [ + 225, + 619, + 686, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach can be identically applied to SDM-v2 (1.26B parameters), leading to BK-SDM-v2-{Base (0.98B), Small (0.88B), Tiny (0.72B)}.", + "bbox": [ + 212, + 679, + 785, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) Fewer Blocks in the Down and Up Stages. This approach is closely aligned with DistilBERT [69] which halves the number of layers and initializes the compact model with the original weights by benefiting from the shared dimensionality. In the original U-Net, each stage with a common spatial size consists of multiple blocks, and most stages contain pairs of residual (R) [17] and cross-attention (A) [25,82] blocks. We hypothesize the existence of some unnecessary pairs and use the following removal strategies.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "B.-K. Kim et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/803b8479453aadc16b24f01220daa9eca29fa7f525a80a1d4c9280434523e209.jpg", + "image_caption": [ + "Fig. 3: Block removal from the denoising U-Net. Our approach is applicable to all the SDM versions in v1 and v2, which share the same U-Net block configuration. For experiments, we used v1.4 [60] and v2.1-base [62]. See Sec. A for the details." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/045c11e1473319a1d92139396107748fe87d6c61f39c893953e1ad42ed9066bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 301, + 383, + 383 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c8b74f10abcd00fdd81656ce0301e279fb1681c1b47af206d773c8eb77b02c7e.jpg", + "image_caption": [ + "Fig. 4: Minor impact of removing the mid-stage from the U-Net. Results without retraining. See Sec. B for additional results." + ], + "image_footnote": [], + "bbox": [ + 217, + 387, + 383, + 458 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/55b1a0e44a56537e20c067b3453203869a67a4c7124e7cc205fb5d42039cc9aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 301, + 784, + 386 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/64d1c5df0da70129ad90c294ff86c741ea16ab828b27d5496039d5bcc608c790.jpg", + "image_caption": [ + "Fig.5: Importance of (a) each block and (b) each group of paired/triplet blocks. Higher score implies removable blocks. The results are aligned with our architectures (e.g., removal of innermost stages and the second R-A pairs in down stages). See Sec. C for further analysis." + ], + "image_footnote": [], + "bbox": [ + 398, + 388, + 784, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Down Stages. We maintain the first R-A pairs while eliminating the second pairs, because the first pairs process the changed spatial information and would be more important than the second pairs. This design is consistent with the sensitivity analysis that measures the block-level significance (see Fig. 5). Our approach also does not harm the dimensionality of the original U-Net, enabling the use of the corresponding pretrained weights for initialization [69]. \n- Up Stages. While adhering to the aforementioned scheme, we retain the third R-A pairs. This allows us to utilize the output feature maps at the end of each down stage and the corresponding skip connections between the down and up stages. The same process is applied to the innermost down and up stages that contain only R blocks.", + "bbox": [ + 227, + 585, + 787, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) Removal of the Entire Mid-Stage. Surprisingly, removing the entire mid-stage from the original U-Net does not noticeably degrade the generation quality while effectively reducing the parameters by $11\\%$ (see Fig. 4). This observation is consistent with the minor role of inner layers in the U-Net generator of GANs [28].", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "bbox": [ + 264, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/be833c8fd0127a7f45a4ec7885c92d3cfd2540f798b19ff4799b9fc3003f4948.jpg", + "image_caption": [ + "Fig. 6: Distillation-based retraining. The block-removed U-Net is trained effectively through the guidance of the original U-Net." + ], + "image_footnote": [], + "bbox": [ + 330, + 143, + 674, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Integrating the mid-stage removal with fewer blocks in Sec. 3.1 further decreases compute burdens (Tab. 2) at the cost of a slight decline in performance (Tab. 1). Therefore, we offer this mid-stage elimination as an option, depending on the priority between compute efficiency (using BK-SDM-Small) and generation quality (BK-SDM-Base).", + "bbox": [ + 212, + 297, + 787, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(3) Further Removal of the Innermost Stages. For additional compression, the innermost down and up stages can also be pruned, leading to our lightest model BK-SDM-Tiny. This implies that outer stages with larger spatial dimensions and their skip connections play a crucial role in the U-Net for T2I synthesis.", + "bbox": [ + 214, + 395, + 789, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alignment with Pruning Sensitivity Analysis. To support the properness of our architectures, we measure the importance of each block (see Fig. 5) and show that unimportant blocks match with our design choices. The importance is measured by how generation scores vary when removing each residual or attention block from the U-Net. A significant drop in performance highlights the essential role of that block. Note that some blocks are not directly removable due to different channel dimensions between input and output; we replace such blocks with channel interpolation modules (denoted by “*” in Fig. 5) to mimic the removal while retaining the information.", + "bbox": [ + 212, + 477, + 787, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The sensitivity analysis implies that the innermost down-mid-up stages and the second R-A pairs in the down stages play relatively minor roles. Pruning these blocks aligns with our architectures, designed based on human knowledge (e.g., prioritizing blocks with altered channel dimensions) and previous studies [28, 69].", + "bbox": [ + 212, + 613, + 787, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Though some results aligned, the importance criterion from the CLIP Score pruning sensitivity (in Fig. 5) leads to excessive pruning of attention blocks, eventually yielding inferior performance compared to our approach (see Table 5).", + "bbox": [ + 212, + 674, + 787, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Distillation-based Retraining", + "text_level": 1, + "bbox": [ + 214, + 739, + 501, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For general-purpose T2I, we train our block-removed U-Net to mimic the behavior of the original U-Net (see Fig. 6). To obtain the input of U-Net, we use pretrained-and-frozen encoders [58] for images and text prompts.", + "bbox": [ + 212, + 763, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the latent representation $\\mathbf{z}$ of an image and its paired text embedding $\\mathbf{y}$ , the task loss for the reverse denoising process [22, 58] is computed as:", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "B.-K. Kim et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T a s k}} = \\mathbb {E} _ {\\mathbf {z}, \\epsilon , \\mathbf {y}, t} \\left[ | | \\epsilon - \\epsilon_ {\\mathrm {S}} (\\mathbf {z} _ {t}, \\mathbf {y}, t) | | _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 157, + 787, + 184 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t$ is a noisy latent code from the diffusion process [22] with the sampled noise $\\epsilon \\sim N(\\mathbf{0},\\mathbf{I})$ and time step $t \\sim \\mathrm{Uniform}(1,T)$ , and $\\epsilon_{\\mathrm{S}}(\\circ)$ indicates the estimated noise from our compact U-Net student. For brevity, we omit the subscripts of $\\mathbb{E}_{\\mathbf{z},\\epsilon,\\mathbf{y},t}[\\circ]$ in the following notations.", + "bbox": [ + 214, + 188, + 787, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The compact student is also trained to imitate the outputs of the original U-Net teacher, $\\epsilon_{\\mathrm{T}}(\\circ)$ , with the following output-level KD objective [21]:", + "bbox": [ + 214, + 250, + 785, + 280 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {O u t K D}} = \\mathbb {E} \\left[ \\left| \\left| \\epsilon_ {\\mathrm {T}} (\\mathbf {z} _ {t}, \\mathbf {y}, t) - \\epsilon_ {\\mathrm {S}} (\\mathbf {z} _ {t}, \\mathbf {y}, t) \\right| \\right| _ {2} ^ {2} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 289, + 787, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A key to our approach is feature-level KD [18, 63] that provides abundant guidance for the student's training:", + "bbox": [ + 214, + 321, + 785, + 354 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {F e a t K D}} = \\mathbb {E} \\left[ \\sum_ {l} | | f _ {\\mathrm {T}} ^ {l} (\\mathbf {z} _ {t}, \\mathbf {y}, t) - f _ {\\mathrm {S}} ^ {l} (\\mathbf {z} _ {t}, \\mathbf {y}, t) | | _ {2} ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 362, + 787, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{T}}^{l}(\\circ)$ and $f_{\\mathrm{S}}^{l}(\\circ)$ represent the feature maps of the $l$ -th layer in a predefined set of distilled layers from the teacher and the student, respectively. While learnable regressors (e.g., $1 \\times 1$ convolutions to match the number of channels) have been commonly used [57,63,75], our approach circumvents this requirement. By applying distillation at the end of each stage in both models, we ensure that the dimensionality of the feature maps already matches, thus eliminating the need for additional regressors.", + "bbox": [ + 214, + 402, + 787, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The final objective is shown below, and we simply set $\\lambda_{\\mathrm{OutKD}}$ and $\\lambda_{\\mathrm{FeatKD}}$ as 1. Without loss-weight tuning, our approach is effective in empirical validation.", + "bbox": [ + 214, + 510, + 785, + 541 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {T a s k}} + \\lambda_ {\\text {O u t K D}} \\mathcal {L} _ {\\text {O u t K D}} + \\lambda_ {\\text {F e a t K D}} \\mathcal {L} _ {\\text {F e a t K D}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 556, + 787, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 595, + 447, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distillation Retraining. We primarily use 0.22M image-text pairs from LAION-Aesthetics V2 (L-Aes) $6.5 + [70,71]$ , which are significantly fewer than the original training data used for SDMs [60,62] ( $>2,000\\mathrm{M}$ pairs). In Fig. 11, dataset sizes smaller than 0.22M are randomly sampled from L-Aes $6.5+$ , while those larger than 0.22M are from L-Aes $6.25+$ .", + "bbox": [ + 214, + 627, + 785, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot T2I Evaluation. Following the popular protocol [56, 58, 66], we use 30K prompts from the MS-COCO validation split [36], downsample the $512 \\times 512$ generated images to $256 \\times 256$ , and compare them with the entire validation set. We compute Fréchet Inception Distance (FID) [20] and Inception Score (IS) [67] to assess visual quality. We measure CLIP score [19, 53] with CLIP-ViT-g/14 model to assess text-image correspondence.", + "bbox": [ + 214, + 703, + 785, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Downstream Tasks. For personalized generation, we use the DreamBooth dataset [65] (30 subjects $\\times$ 25 prompts $\\times$ 4~6 images) and perform per-subject finetuning. Following the evaluation protocol [65], we use ViT-S/16 model [4]", + "bbox": [ + 214, + 794, + 785, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "bbox": [ + 264, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e13e3ffb39ed17b06a09be93a8dabcc72e0939a78a4b57d1f8ac4bde3a50d74.jpg", + "table_caption": [ + "Table 1: Results on zero-shot MS-COCO $256 \\times 256$ 30K. Training resources include image-text pairs, batch size, iterations, and A100 days. Despite far smaller resources, our compact models outperform prior studies [10, 11, 56, 79, 90], showing the benefit of compressing existing powerful models. Note that FID fluctuates more than the other metrics over training progress in our experiments (see Figs. 9 and 11)." + ], + "table_footnote": [ + "$\\dagger$ : Evaluated with the released checkpoints. $\\ddagger$ : Total parameters for T2I synthesis. $*$ : Estimated based on public information. DF and AR: diffusion and autoregressive models. $\\downarrow$ and $\\uparrow$ : lower and higher values are better." + ], + "table_body": "
Model\nNameType #Param‡Generation ScoreTraining Resource
FID↓IS↑CLIP↑Data Size (Batch, # Iter)A100 Days
SDM-v1.4 [58,60]†DF1.04B13.0536.760.2958>2000M*(2048, 1171K)6250
Small Stable Diffusion [51]†DF0.76B12.7632.330.2851229M(128, 1100K)-
BK-SDM-Base [Ours]†DF0.76B15.7633.790.28780.22M(256, 50K)13
BK-SDM-Small [Ours]†DF0.66B16.9831.680.26770.22M(256, 50K)13
BK-SDM-Tiny [Ours]†DF0.50B17.1230.090.26530.22M(256, 50K)13
SDM-v2.1-base [58,62]†DF1.26B13.9335.930.3075>2000M*(2048, 1620K)8334
BK-SDM-v2-Base [Ours]†DF0.98B15.8531.700.28680.22M(128, 50K)4
BK-SDM-v2-Small [Ours]†DF0.88B16.6131.730.29010.22M(128, 50K)4
BK-SDM-v2-Tiny [Ours]†DF0.72B15.6831.640.28970.22M(128, 50K)4
DALL-E [56]AR12B27.517.9-250M(1024, 430K)-
CogView [10]AR4B27.118.2-30M(6144, 144K)-
CogView2 [11]AR6B2422.4-30M(4096, 300K)-
Make-A-Scene [15]AR4B11.84--35M(1024, 170K)-
LAFITE [90]GAN0.23B26.9426.02-3M--
GALIP (CC12M) [79]†GAN0.32B13.8625.160.281712M--
GigaGAN [27]GAN1.1B9.09-->100M*(512, 1350K)4783
GLIDE [48]DF3.5B12.24--250M(2048, 2500K)-
LDM-KL-8-G [58]DF1.45B12.6330.29-400M(680, 390K)-
DALL-E-2 [55]DF5.2B10.39--250M(4096, 3400K)-
SnapFusion [35]DF0.99B~13.6-~0.295>100M*(2048, -)>128*
Würstchen-v2 [50]†DF3.1B22.4032.870.26761700M(1536, 1725K)1484
MobileDiffusion [89]†DF0.52B9.01--150M(2048, 1000K)-
", + "bbox": [ + 222, + 227, + 781, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for DINO score and CLIP-ViT-g/14 model for CLIP-I and CLIP-T scores. For image-to-image translation, input images are sourced from Meng et al. [43].", + "bbox": [ + 212, + 564, + 787, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. We adjust the codes in Diffusers [52] and PEFT [41]. We use a single NVIDIA A100 80G GPU for main retraining and a single NVIDIA GeForce RTX 3090 GPU for per-subject finetuning. For compute efficiency, we always opt for 25 denoising steps of the U-Net at the inference phase, unless specified. The classifier-free guidance scale [23,66] is set to the default value of 7.5. The latent resolution is set to the default ( $H = W = 64$ in Fig. 3), yielding $512 \\times 512$ images. See Sec. J for the details.", + "bbox": [ + 212, + 597, + 787, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Results", + "text_level": 1, + "bbox": [ + 215, + 744, + 323, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All the results in Secs. 5.1-5.6 were obtained with the full benchmark protocol (MS-COCO $256 \\times 256$ 30K samples), except for Fig. 10 ( $512 \\times 512$ 5K samples). Unless specified, the training setup in Tab. 1 was used.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "B.-K. Kim et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfb355cb7ac4f643ced756f39c9166f92ab7017d45c90a21a398387b8461f6f9.jpg", + "image_caption": [ + "Fig. 7: Visual comparison with open-sourced models. The results [11,48,50,79,90] were obtained with their official codes.", + "Table 2: Impact of compute reduction in U-Net on the entire SDM. The number of sampling steps is indicated with the parentheses, e.g., U-Net (1) for one step. The full computation (denoted by \"Whole\") covers the text encoder, U-Net, and image decoder. All corresponding values are obtained on the generation of a single $512 \\times 512$ image with 25 denoising steps. The latency was measured on Xeon Silver 4210R CPU 2.40GHz and NVIDIA GeForce RTX 3090 GPU." + ], + "image_footnote": [ + "Prompt: A bowl that has vegetables inside of it; A brown and white cat staring off with pretty green eyes.; A toy raccoon standing on a pile of broccoli." + ], + "bbox": [ + 220, + 143, + 782, + 305 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0c03c70a3823e356667282666c42e6ec0553c7eba452dbd491aa2fe9ba88a43a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model# ParamMACsCPU LatencyGPU Latency
U-NetWholeU-Net (1)U-Net (25)WholeU-Net (1)U-Net (25)WholeU-Net (1)U-Net (25)Whole
SDM-v1.4 [60]860M1033M339G8469G9716G5.63s146.28s153.00s0.049s1.28s1.41s
BK-SDM-Base [Ours]580M752M224G5594G6841G3.84s99.95s106.67s0.032s0.83s0.96s
(-32.6%)(-27.1%)(-33.9%)(-33.9%)(-29.5%)(-31.8%)(-31.7%)(-30.3%)(-34.6%)(-35.2%)(-31.9%)
BK-SDM-Small [Ours]483M655M218G5444G6690G3.45s89.78s96.50s0.030s0.77s0.90s
(-43.9%)(-36.5%)(-35.7%)(-35.7%)(-31.1%)(-38.7%)(-38.6%)(-36.9%)(-38.7%)(-39.8%)(-36.1%)
BK-SDM-Tiny [Ours]324M496M206G5126G6373G3.03s78.77s85.49s0.026s0.67s0.80s
(-62.4%)(-51.9%)(-39.5%)(-39.5%)(-34.4%)(-46.2%)(-46.1%)(-44.1%)(-46.9%)(-47.7%)(-43.2%)
", + "bbox": [ + 217, + 465, + 787, + 560 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Comparison with Existing Works", + "text_level": 1, + "bbox": [ + 214, + 588, + 535, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative Comparison. Tab. 1 shows the zero-shot results for general-purpose T2I. Despite being trained with only 0.22M samples and having fewer than 1B parameters, our compressed models demonstrate competitive performance on par with existing large models.", + "bbox": [ + 212, + 612, + 785, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visual Comparison. Fig. 7 depicts synthesized images with some MS-COCO captions. Our compact models inherit the superiority of SDM and produce more photorealistic images compared to the AR-based [11] and GAN-based [79, 90] baselines. Noticeably, the same latent code results in a shared visual style between the original and our models (6th–9th columns in Fig. 7), similar to the observation in transfer learning for GANs [45]. See Sec. D for additional results.", + "bbox": [ + 212, + 672, + 785, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Computational Gain", + "text_level": 1, + "bbox": [ + 214, + 785, + 429, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tab. 2 shows how the compute reduction for each sampling step of the U-Net affects the overall process. The per-step reduction effectively decreases MACs", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4a3176249c534bb2256fc922e395cb48c2c6a8fafac118c6a56351595d8a3a1c.jpg", + "table_caption": [ + "Table 3: Quality gains from distillation retraining. MS-COCO 30K." + ], + "table_footnote": [], + "table_body": "
BK-SDM\nTypeKDGenerationScore# Param\nU-Net
FID↓IS↑CLIP↑
BaseX23.5723.020.24080.58B
✓15.7633.790.2878
v2-BaseX16.7625.880.26610.59B
✓15.8531.700.2868
v2-SmallX16.7125.770.26550.49B
✓16.6131.730.2901
v2-TinyX16.8726.060.26780.33B
✓15.6831.640.2897
", + "bbox": [ + 220, + 186, + 470, + 301 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ebaaa7a0297bb970b20e7e9242421c9d17e52116cde7be5b002cea6374312756.jpg", + "table_caption": [ + "Table 4: Significance of each element in transferred knowledge. Results of BK-SDMv2-Small on MS-COCO 30K." + ], + "table_footnote": [], + "table_body": "
Weight InitializationOutput FeatureGeneration Score
KDKDFID↓IS↑CLIP↑
RandomXX41.7515.420.1733
Random✓X29.0118.290.1967
Random✓✓24.4722.370.2323
TeacherXX16.7125.770.2655
Teacher✓X14.2729.470.2777
Teacher✓✓16.6131.730.2901
", + "bbox": [ + 486, + 202, + 781, + 300 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a0558c4dd4c4092b6bb06418cb3166a283d39ead661c10f564a4f82e1dce13ae.jpg", + "image_caption": [ + "Fig. 8: Image areas affected by each word. KD enables our models to mimic the SDM, yielding similar per-word attribution maps. The model without KD behaves differently, causing dissimilar maps and inaccurate generation (e.g., two sheep and unusual bird shapes)." + ], + "image_footnote": [], + "bbox": [ + 251, + 313, + 751, + 473 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and inference time by more than $30\\%$ . Notably, BK-SDM-Tiny has $50\\%$ fewer parameters than the original SDM.", + "bbox": [ + 212, + 569, + 787, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Benefit of Distillation Retraining", + "text_level": 1, + "bbox": [ + 214, + 619, + 535, + 636 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "T2I Performance. Tab. 3 summarizes the results from ablating the total KD objective (Eq. 2+Eq. 3). Across various model types, distillation brings a clear improvement in generation quality. Tab. 4 analyzes the effect of each element in transferred knowledge. Exploiting output-level KD (Eq. 2) boosts the performance compared to using only the denoising task loss. Leveraging feature-level KD (Eq. 3) yields further score enhancement. Additionally, using the teacher weights for initialization is highly beneficial.", + "bbox": [ + 212, + 643, + 787, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Cross-attention Resemblance. Fig. 8 displays the per-word attribution maps [78] created by aggregating cross-attention scores over spatiotemporal dimensions. The attribution maps of our models are semantically and spatially similar to those of the original model, indicating the merit of supervisory signals at multiple stages via KD. In contrast, the baseline model without KD activates incorrect areas, leading to text-mismatched generation results.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "B.-K. Kim et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f6d779e2f565da131a56c95ae295d76bf605647dc0c3e59797a2ba1158d0d112.jpg", + "image_caption": [ + "Fig.9: Zero-shot results over training progress. The architecture size of BK-SDM, usage of KD, and batch size are denoted. Results on MS-COCO 30K." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 338, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6bd5100732c37d9bde039be5b4df979ab4b09b2e1904ac0adf1e7ac7829ba24b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 145, + 459, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8da766e3190a210889185f8bf4d6cc18f628b90837a238fef02cfdd3da1ee28b.jpg", + "image_caption": [ + "Fig.10:Trade-off curves Left:FID vs.CLIP score Right: quality vs.efficiency Ours-Base on MS-COCO 5K." + ], + "image_footnote": [], + "bbox": [ + 459, + 145, + 581, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/54b695e4dd62078adc4fd40badb1b3a3ff07671d80e92840633383e7c4ad1566.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 143, + 785, + 242 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training Progress. Fig. 9 shows the results over training iterations. Without KD, training solely with the denoising task loss causes fluctuations or sudden drops in performance (indicated with green and cyan). In contrast, distillation (purple and pink) stabilizes and accelerates the training process, demonstrating the benefit of providing sufficient hints for training guidance. Notably, our small-and tiny-size models trained with KD (yellow and red) outperform the bigger base-size model without KD (cyan). Additionally, while the best FID score is observed early on, IS and CLIP score exhibit ongoing improvement, implying that judging models merely with FID may be suboptimal.", + "bbox": [ + 212, + 349, + 787, + 487 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Trade-off Results. Fig. 10 shows the results of BK-SDM-Base with and without KD on MS-COCO $512 \\times 512$ 5K. Higher classifier-free guidance scales [23,66] lead to better text-aligned images at the cost of less diversity. More denoising steps improve generation quality at the cost of slower inference. Distillation retraining achieves much better trade-off curves than the baseline without KD.", + "bbox": [ + 212, + 487, + 787, + 561 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.4 Comparison with Different Pruning Criteria", + "text_level": 1, + "bbox": [ + 214, + 584, + 625, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We extend our baseline comparisons by calculating several block-level importance scores [29].1 Furthermore, the impact of applying KD at the end of each stage is also examined. Tab. 5 shows the results. Magnitude (Mag) pruning [32] removes critical outer blocks, causing irrecoverable loss. Taylor pruning [14,30,46] removes only the inner part, causing ineffective reduction in MACs and speed; additionally, this method demands the intensive calculation of backward gradients. The CLIP Score method (based on Fig. 5) results in severe pruning of", + "bbox": [ + 212, + 607, + 787, + 715 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "bbox": [ + 264, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "1 Let the $(h_{\\mathrm{out}}, h_{\\mathrm{in}})$ -size weight matrix in the $b$ -th block be $\\mathbf{W}^{l,b} = \\left[W_{i,j}^{l,b}\\right]$ , where $l$ denotes the layer type (e.g., convolution (flattened) or attention's key projection). The scores at the output neuron level [1] are aggregated for the block-level importance criteria, $S_{\\mathrm{Magnitude}}^b = \\mathbb{E}_{l,i}\\left[\\sum_j\\left|W_{i,j}^{l,b}\\right|\\right]$ and $S_{\\mathrm{Taylor}}^b = \\mathbb{E}_{l,i}\\left[\\sum_j\\left|\\frac{\\partial\\mathcal{L}(D)}{\\partial W_{i,j}^{l,b}}W_{i,j}^{l,b}\\right|\\right]$ . Here, $\\mathcal{L}$ and $D$ denote the denoising task loss and a calibration set of 1K samples. The final scores are then ranked to remove unimportant blocks, or to replace them with interpolation for unremovable blocks.", + "bbox": [ + 217, + 723, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 5: Different block-level criteria at similar parameters. Taylor pruning removes solely the inner blocks, leading to insufficient reduction in MACs. Our method attains a favorable compromise between performance and inference speed. Results on MS-COCO 30K.", + "bbox": [ + 212, + 143, + 527, + 227 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e86b102d0e614bc381f723b7a952e9983c66c4064285f9163699156b87681190.jpg", + "table_caption": [], + "table_footnote": [ + "Retraining with batch 128, 0.22M data, and 50K iters. Original SD-v1.4's U-Net: (MACs, Params) = (338.7G, 859.5M)." + ], + "table_body": "
Pruning \nCriterionFID↓CLIP Score↑U-Net
KD×KD√KD×KD√MACs# Param
Mag [32]>150>150<0.1<0.1122.2G495.5M
Taylor [14,30]18.8516.730.24290.2749305.4G493.4M
CLIP [Fig. 5]32.9321.860.19360.2283204.8G496.8M
Ours-Small22.9716.830.22870.2668217.7G482.3M
", + "bbox": [ + 220, + 241, + 522, + 305 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c316aaee828a6ce6bd3459f5f0f84700674fb9554353b7d2a10d6f90c13de787.jpg", + "table_caption": [ + "Table 6: A/B-testing preference study. Total 2,501 comparisons." + ], + "table_footnote": [ + "Participants selected preferred images from pairs with the same prompts. One image in each pair was always from BK-SDM-Small, while the other was randomly chosen from the four methods." + ], + "table_body": "
Ours-SmallGLIDE-filt [48]GALIP [79]Würstchen-v2 SDM-v1.4 [50][60]
# Win†453 (71.5%)379 (61.5%)319 (51.4%)202 (32.1%)
# Defeat†181 (28.5%)237 (38.5%)302 (48.6%)428 (67.9%)
", + "bbox": [ + 535, + 207, + 782, + 267 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e9ac6b9fba2754ca57a4f8f6a1f5be639a7fef48ece41e149d2523ba21edb759.jpg", + "table_caption": [ + "Table 7: Impact of training batch size and iterations. Results on MS-COCO 30K." + ], + "table_footnote": [], + "table_body": "
BK-SDMBaseSmallTiny
Batch Size642566425664256
FID↓14.6115.7616.8716.9817.2817.12
IS↑31.4433.7929.5131.6828.3330.09
CLIP↑0.28260.28780.26440.26770.26070.2653
BK-SDMBase (Data 2.3M)Small (Data 2.3M)Tiny (Data 2.3M)
#Iter50K100K50K100K50K100K
FID↓14.8115.3917.0517.0117.5317.63
IS↑34.1734.7633.1033.1431.3232.26
CLIP↑0.28830.28890.27340.27540.26900.2713
", + "bbox": [ + 220, + 396, + 527, + 503 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b57e17a2fb09f53b6f478788d756b0a1b43a0ec56f8217cedd37595525806530.jpg", + "image_caption": [ + "Fig. 11: Impact of training data volume. Results of BK-SDM-Small on MS-COCO 30K." + ], + "image_footnote": [], + "bbox": [ + 537, + 339, + 658, + 452 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9cb8d13181c912f273bc88dddfc2ffe89b622298ff8d4564dc05a8142a604557.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 339, + 784, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "attention blocks, adversely affecting performance. Our approach achieves superior or comparable results against the benchmark methods and can be directly applied to all the SDM versions in v1 and v2. This advantage comes without the extra phase of calculating pruning criteria while ensuring a favorable balance between performance and inference speed. Notably, KD always remains effective.", + "bbox": [ + 212, + 549, + 787, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.5 Human Preference Assessment", + "text_level": 1, + "bbox": [ + 214, + 648, + 516, + 662 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Despite their extensive use, automated metrics for assessing image quality (e.g., FID) do not consistently match with human evaluation [55, 66]. To further address this, we have conducted an A/B testing-based study: participants were shown pairs of images generated using the same MS-COCO prompt and asked to choose their preferred image in relation to the prompt. One image in each pair was always from our model, while the other was randomly selected from one of four comparative methods. A total of 25 participants conducted 2,501 comparisons online. The participants had no bias towards the examined models and were not compensated. The prompts were randomly drawn from the MS-COCO set and varied for 3 to 5 participants. As shown in Tab. 6, our model (0.66B parameters) excels over GLIDE-filtered, GALIP, and Wüstchen-v2 (3.1B).", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "B.-K. Kim et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 8: Personalized generation with finetuning over different backbones. Our compact models can preserve subject fidelity (DINO and CLIP-I) and prompt fidelity (CLIP-T) of the original SDM with reduced finetuning (FT) costs and fewer parameters.", + "bbox": [ + 214, + 143, + 602, + 214 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/25960ae8bbd48c0a2738589b6f455242df720db080625ce42b3a8196cda2836f.jpg", + "table_caption": [], + "table_footnote": [ + "†: Per-subject FT time and GPU memory for 800 iterations on RTX 3090." + ], + "table_body": "
Backbone {Param}DINO↑CLIP-I↑CLIP-T↑FT (Time, Mem)†
SDM-v1.4 [60] {1.04B}0.7280.7250.263(882s, 23.0GB)
BK-SDM-Base {0.76B}0.7230.7170.260(623s, 18.7GB)
BK-SDM-Small {0.66B}0.7200.7050.259(604s, 17.2GB)
BK-SDM-Tiny {0.50B}0.7150.6930.261(560s, 13.1GB)
Base (Batch 64) {0.76B}0.7180.7080.262(623s, 18.7GB)
- No KD & Random Init.0.5940.4650.191(623s, 18.7GB)
- No KD & Teacher Init.0.7160.6690.258(623s, 18.7GB)
", + "bbox": [ + 220, + 227, + 596, + 323 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/dc2f6913ab6103edb4a9b261e1e3d5bcf4067f47fa0948e5cb6e2ca02b08e5ea.jpg", + "table_caption": [ + "Table 9: Speedups on edge devices." + ], + "table_footnote": [], + "table_body": "
ModelAGX Orin 32GBiPhone 14
SDM-v14.9s [61]5.6s [60]
BK-SDM-Base3.4s (-31%)4.0s (-29%)
BK-SDM-Small3.2s (-35%)3.9s (-29%)
BK-SDM-Tiny2.8s (-43%)3.9s (-29%)
", + "bbox": [ + 611, + 210, + 782, + 313 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/64329bbe172f3cba1d82952c1f0148ea45368e99bc02947d14dbbdec5319fe22.jpg", + "image_caption": [ + "Fig. 12: Visual results of personalized generation. Each subject is marked as \"a [identifier] [class noun]\" (e.g., \"a [V] dog\")." + ], + "image_footnote": [], + "bbox": [ + 217, + 349, + 589, + 470 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/37a3ea26f54b4cc2ad8d07c1029c0ce3d40a053089701804f24f11a73c238dd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 351, + 784, + 469 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a9a460561ad1244a1c71559b8eb11c8333a389e403460129ff3bb113e3f27d36.jpg", + "image_caption": [ + "(left-to-right, top-to-bottom) Target style: Art station; Oil painting; Photograph; Watercolor", + "Fig. 13: Text-guided image-to-image translation." + ], + "image_footnote": [], + "bbox": [ + 217, + 525, + 406, + 638 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/46e70f60fbb2ae51979cf68452df557d01bb97b385c525b482320cc4d2c315df.jpg", + "image_caption": [ + "Fig. 14: Small LDM for face synthesis." + ], + "image_footnote": [], + "bbox": [ + 406, + 525, + 596, + 638 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ec1f16501c99b509c4228d59a5421bd5e4f0df0fe8459e0e144a5c81ec676e4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 525, + 785, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.6 Impact of Training Resources on Performance", + "text_level": 1, + "bbox": [ + 214, + 719, + 640, + 734 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Consistent with existing evidence, the use of larger batch sizes, more extensive data, and longer iterations for training enhances performance in our work (see Tab. 7, Fig. 11, and Sec. H). However, this benefit requires increased resource demands (e.g., extended training days without multiple high-spec GPUs and greater data storage capacity). As such, despite the better performing models in Tab. 7, we primarily report the models with fewer resources. We believe that", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "accessible training costs by many researchers can help drive advancement in massive models.", + "bbox": [ + 215, + 146, + 784, + 175 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.7 Application", + "text_level": 1, + "bbox": [ + 215, + 202, + 357, + 217 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Personalized T2I Synthesis. Tab. 8 compares the fine-tuning results with DreamBooth [65] over different backbones to create images about a given subject. BK-SDMs can preserve $95\\% \\sim 99\\%$ scores of the original SDM while cutting fine-tuning costs. Fig. 12 depicts that our models can accurately capture the subject details and generate various scenes. Over the models retrained with a batch size of 64, the baselines without KD fail to generate the subjects or cannot maintain the identity details. See Sec. E for further results.", + "bbox": [ + 212, + 231, + 787, + 335 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Image-to-Image Translation. Fig. 13 presents the text-guided stylization results with SDEdit [44]. Our model, resembling the ability of the original SDM, faithfully produces images given style-specified prompts and content pictures. See Sec. F for additional results.", + "bbox": [ + 212, + 337, + 787, + 396 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Deployment on Edge Devices. We deploy our models trained with 2.3M pairs and compare them against the original SDM under the same setup on edge devices (20 denoising steps on NVIDIA Jetson AGX Orin 32GB and 10 steps on iPhone 14). Our models produce a $512 \\times 512$ image within 4 seconds (see Tab. 9), while maintaining acceptable image quality (Fig. 1(d) and Sec. G).", + "bbox": [ + 212, + 398, + 787, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Another LDM. SDMs are derived from LDMs [58], which share a similar U-Net design across many tasks. To validate the generality of our work, we compress an LDM (with 308M parameters and 410K training iterations) for unconditional generation on CelebA-HQ [59] by using the same approach of BK-SDM-Small (187M parameters and 30K iterations). Fig. 14 shows the efficacy of our architecture and distillation retraining.", + "bbox": [ + 212, + 474, + 787, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 590, + 508, + 607 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We uncover the potential of architectural compression for general-purpose text-to-image synthesis with a renowned model, Stable Diffusion. Our block-removed lightweight models are effective for zero-shot generation, achieving competitive results against large-scale baselines. Distillation is a key of our method, leading to powerful retraining even under very constrained resources. Our work is orthogonal to previous works for efficient diffusion models, e.g., enabling fewer sampling steps, and can be readily combined with them. We hope our study can facilitate future research on structural compression of large diffusion models.", + "bbox": [ + 212, + 627, + 787, + 747 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Our compact models inherit the capability of the source model for high-fidelity image generation, but they have shortcomings such as inaccurate generation of full-body human appearance.", + "bbox": [ + 212, + 750, + 787, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Negative Social Impacts. Recent large generative models are capable of creating high-quality plausible content. To avoid causing unintended social usage, researchers should take steps to ensure the appropriateness of the training data.", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "B.-K. Kim et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank the Microsoft Startups Founders Hub program and the AI Industrial Convergence Cluster Development project funded by the Ministry of Science and ICT (MSIT, Korea) and Gwangju Metropolitan City for their generous support of GPU resources.", + "bbox": [ + 212, + 176, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. A Simple and Effective Pruning Approach for LLMs. ICLR (2024) 11", + "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) 2, 3", + "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023) 2, 3", + "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) 7", + "5. Castells, T., Song, H.K., Piao, T., Choi, S., Kim, B.K., Yim, H., Lee, C., Kim, J.G., Kim, T.H.: Edgefusion: On-device text-to-image generation. In: CVPR Workshop (2024) 3", + "6. Chen, Y.H., Sarokin, R., Lee, J., Tang, J., Chang, C.L., Kulik, A., Grundmann, M.: Speed is all you need: On-device acceleration of large diffusion models viagpu-aware optimizations. In: CVPR Workshop (2023) 2, 3", + "7. Choi, J., Kim, M., Ahn, D., Kim, T., Kim, Y., Jo, D., Jeon, H., Kim, J.J., Kim, H.: Squeezing large-scale diffusion models for mobile. In: ICML Workshop (2023) 2, 3", + "8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 3", + "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021) 2, 3", + "10. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. In: NeurIPS (2021) 2, 8", + "11. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. In: NeurIPS (2022) 2, 8, 9", + "12. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) 2", + "13. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: CVPR (2023) 2", + "14. Fang, G., Ma, X., Wang, X.: Structural pruning for diffusion models. In: NeurIPS (2023) 4, 11, 12", + "15. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV (2022) 8", + "16. Hao, Z., Guo, J., Jia, D., Han, K., Tang, Y., Zhang, C., Hu, H., Wang, Y.: Learning efficient vision transformers via fine-grained manifold distillation. In: NeurIPS (2022) 4", + "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 4" + ], + "bbox": [ + 218, + 295, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "bbox": [ + 264, + 113, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Heo, B., Kim, J., Yun, S., Park, H., Kwak, N., Choi, J.Y.: A comprehensive overhaul of feature distillation. In: ICCV (2019) 2, 3, 7", + "19. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: EMNLP (2021) 7", + "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 7", + "21. Hinton, G., Vinyls, O., Dean, J.: Distilling the knowledge in a neural network. In: NeurIPS Workshop (2014) 3, 7", + "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) 3, 6, 7", + "23. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop (2021) 2, 8, 11", + "24. Hou, J., Asghar, Z.: World's first on-device demonstration of stable diffusion on an android phone. https://wwwQUALCOMM.com/news (2023) 2, 3", + "25. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: ICML (2021) 4", + "26. Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., Wang, F., Liu, Q.: Tinybert: Distilling bert for natural language understanding. In: Findings of EMNLP (2020) 4", + "27. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) 8", + "28. Kim, B.K., Choi, S., Park, H.: Cut inner layers: A structured pruning strategy for efficient u-net gans. In: ICML Workshop (2022) 5, 6", + "29. Kim, B.K., Kim, G., Kim, T.H., Castells, T., Choi, S., Shin, J., Song, H.K.: Shortened llama: A simple depth pruning for large language models. arXiv preprint arXiv:2402.02834 (2024) 11", + "30. LeCun, Y., Denker, J., Solla, S.: Optimal brain damage. In: NeurIPS (1989) 11, 12", + "31. Lee, Y., Park, K., Cho, Y., Lee, Y.J., Hwang, S.J.: Koala: Self-attention matters in knowledge distillation of latent diffusion models for memory-efficient and fast image synthesis. arXiv preprint arXiv:2312.04005v1 (2023) 3", + "32. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. In: ICLR (2017) 11, 12", + "33. Li, M., Lin, J., Ding, Y., Liu, Z., Zhu, J.Y., Han, S.: Gan compression: Efficient architectures for interactive conditional gans. In: CVPR (2020) 3", + "34. Li, X., Liu, Y., Lian, L., Yang, H., Dong, Z., Kang, D., Zhang, S., Keutzer, K.: Q-diffusion: Quantizing diffusion models. In: ICCV (2023) 2, 3", + "35. Li, Y., Wang, H., Jin, Q., Hu, J., Chemerys, P., Fu, Y., Wang, Y., Tulyakov, S., Ren, J.: Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. In: NeurIPS (2023) 4, 8", + "36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dolkar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) 7", + "37. Liu, L., Ren, Y., Lin, Z., Zhao, Z.: Pseudo numerical methods for diffusion models on manifolds. In: ICLR (2022) 2", + "38. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In: NeurIPS (2022) 3", + "39. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095 (2022) 3" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "B.-K. Kim et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023) 2", + "41. Mangrulkar, S., Gugger, S., Debut, L., Belkada, Y., Paul, S.: Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft (2022) 8", + "42. Meng, C., Gao, R., Kingma, D.P., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: NeurIPS Workshop (2022) 3", + "43. Meng, C., Gao, R., Kingma, D.P., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR (2023) 2, 3, 8", + "44. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022) 3, 14", + "45. Mo, S., Cho, M., Shin, J.: Freeze the discriminator: a simple baseline for fine-tuning gans. In: CVPR Workshop (2020) 9", + "46. Molchanov, P., Mallya, A., Tyree, S., Frosio, I., Kautz, J.: Importance estimation for neural network pruning. In: CVPR (2019) 11", + "47. Murti, C., Narshana, T., Bhattacharyya, C.: TVSPrune - pruning non-discriminative filters via total variation separability of intermediate representations without fine tuning. In: ICLR (2023) 2", + "48. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: ICML (2022) 3, 8, 9, 12", + "49. Park, W., Kim, D., Lu, Y., Cho, M.: Relational knowledge distillation. In: CVPR (2019) 3", + "50. Pernias, P., Rampas, D., Richter, M.L., Pal, C.J., Aubreville, M.: Wierstchen: An efficient architecture for large-scale text-to-image diffusion models. In: ICLR (2024) 4, 8, 9, 12", + "51. Pinkney, J.: Small stable diffusion. https://huggingface.co/OFA-Sys/small-stable-diffusion-v0 (2023) 8", + "52. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022) 8", + "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 3, 7", + "54. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019) 3", + "55. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) 3, 8, 12", + "56. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) 2, 7, 8", + "57. Ren, Y., Wu, J., Xiao, X., Yang, J.: Online multi-granularity distillation for gan compression. In: ICCV (2021) 3, 7", + "58. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 3, 6, 7, 8, 14", + "59. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: Ldm on celeba-hq. https://huggingface.co/CompVis/ldm-celebahq-256 (2022) 14", + "60. Rombach, R., Esser, P.: Stable diffusion v1-4. https://huggingface.co/CompVis/stable-diffusion-v1-4 (2022) 2, 3, 5, 7, 8, 9, 12, 13" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Rombach, R., Esser, P.: Stable diffusion v1-5. https://huggingface.co/runwayml/stable-diffusion-v1-5 (2022) 2, 3, 13", + "62. Rombach, R., Esser, P., Ha, D.: Stable diffusion v2-1-base. https://huggingface.co/stabilityai/stable-diffusion-2-1-base (2022) 2, 3, 5, 7, 8", + "63. Romero, A., Ballas, N., Kahou, S.E., Chassang, A., Gatta, C., Bengio, Y.: Fitnets: Hints for thin deep nets. In: ICLR (2015) 2, 3, 7", + "64. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) 2, 4", + "65. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023) 3, 7, 14", + "66. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 3, 7, 8, 11, 12", + "67. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: NeurIPS (2016) 7", + "68. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2022) 3", + "69. Sanh, V., Debut, L., Chaumont, J., Wolf, T.: Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In: NeurIPS Workshop (2019) 4, 5, 6", + "70. Schuhmann, C., Beaumont, R.: Laion-aesthetics. https://laion.ai/blog/laion-aesthetics (2022) 3, 7", + "71. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS Workshop (2022) 7", + "72. Segmind: Segmind-distill-sd. https://github.com/segmind/distill-sd/tree/c1e97a70d141df09e6fe5cc7dbd66e0cbae3eeb (2023) 3", + "73. Segmind: Ssd-1b. https://github.com/segmind/SSD-1B/tree/d2ff723ea8ecf5ddb86f3aac0af1db30e88a2e2d (2023) 3", + "74. Shen, H., Cheng, P., Ye, X., Cheng, W., Abidi, H.: Accelerate stable diffusion with intel neural compressor. https://medium.com/intel-analytics-software (2022) 2, 3", + "75. Shu, C., Liu, Y., Gao, J., Yan, Z., Shen, C.: Channel-wise knowledge distillation for dense prediction. In: ICCV (2021) 7", + "76. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) 2, 3", + "77. Sun, Z., Yu, H., Song, X., Liu, R., Yang, Y., Zhou, D.: Mobilebert: a compact task-agnostic bert for resource-limited devices. In: ACL (2020) 4", + "78. Tang, R., Liu, L., Pandey, A., Jiang, Z., Yang, G., Kumar, K., Stenetorp, P., Lin, J., Ture, F.: What the DAAM: Interpreting stable diffusion using cross attention. In: ACL (2023) 10", + "79. Tao, M., Bao, B.K., Tang, H., Xu, C.: Galip: Generative adversarial clips for text-to-image synthesis. In: CVPR (2023) 8, 9, 12", + "80. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers and distillation through attention. In: ICML (2021) 4", + "81. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: NeurIPS (2017) 2" + ], + "bbox": [ + 215, + 147, + 787, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "B.-K. Kim et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "82. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 4", + "83. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: CVPR (2023) 2, 3", + "84. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: CVPR (2023) 2", + "85. Zagoruyko, S., Komodakis, N.: Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In: ICLR (2017) 3", + "86. Zhang, L., Chen, X., Tu, X., Wan, P., Xu, N., Ma, K.: Wavelet knowledge distillation: Towards efficient image-to-image translation. In: CVPR (2022) 3", + "87. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) 2", + "88. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: ICLR (2023) 3", + "89. Zhao, Y., Xu, Y., Xiao, Z., Hou, T.: Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567 (2023) 4, 8", + "90. Zhou, Y., Zhang, R., Chen, C., Li, C., Tensmeyer, C., Yu, T., Gu, J., Xu, J., Sun, T.: Towards language-free training for text-to-image generation. In: CVPR (2022) 8, 9", + "91. Zhu, L.: Thop: Pytorch-ocounter. https://github.com/Lyken17/pytorchOpCounter (2018) 2" + ], + "bbox": [ + 212, + 146, + 787, + 452 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_model.json b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_model.json new file mode 100644 index 0000000000000000000000000000000000000000..59ffdc36574cbba3616ba0069c624ad47bc641f5 --- /dev/null +++ b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_model.json @@ -0,0 +1,3219 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.185 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.212, + 0.744, + 0.243 + ], + "angle": 0, + "content": "Bo-Kyeong Kim1, Hyoung-Kyu Song2, Thibault Castells1, and Shinkook Choi1" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.254, + 0.61, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\)Nota Inc. \\(^{2}\\)Captions Research" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.27, + 0.747, + 0.284 + ], + "angle": 0, + "content": "{bokyeong.kim, thibault, shinkook.choi}@nota.ai, kyu@captions.ai" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.318, + 0.443, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.463, + 0.398, + 0.472 + ], + "angle": 0, + "content": "(a) Efficient General-purpose T2I" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.318, + 0.665, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.387, + 0.612, + 0.394 + ], + "angle": 0, + "content": "(b) Efficient Personalized T2I" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.395, + 0.664, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.488, + 0.463, + 0.619, + 0.472 + ], + "angle": 0, + "content": "(c) Efficient I2I (Different Styles)" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.318, + 0.785, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.411, + 0.785, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.463, + 0.771, + 0.472 + ], + "angle": 0, + "content": "(d) Mobile Deployment" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Fig. 1: Our compressed model enables efficient (a) zero-shot text-to-image generation, (b) personalized synthesis, (c) image-to-image translation, and (d) mobile deployment. Samples from BK-SDM-Small with \\(36\\%\\) reduced parameters and latency are shown." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.577, + 0.744, + 0.84 + ], + "angle": 0, + "content": "Abstract. Text-to-image (T2I) generation with Stable Diffusion models (SDMs) involves high computing demands due to billion-scale parameters. To enhance efficiency, recent studies have reduced sampling steps and applied network quantization while retaining the original architectures. The lack of architectural reduction attempts may stem from worries over expensive retraining for such massive models. In this work, we uncover the surprising potential of block pruning and feature distillation for low-cost general-purpose T2I. By removing several residual and attention blocks from the U-Net of SDMs, we achieve \\(30\\% \\sim 50\\%\\) reduction in model size, MACs, and latency. We show that distillation retraining is effective even under limited resources: using only 13 A100 days and a tiny dataset, our compact models can imitate the original SDMs (v1.4 and v2.1-base with over 6,000 A100 days). Benefiting from the transferred knowledge, our BK-SDMs deliver competitive results on zero-shot MS-COCO against larger multi-billion parameter models. We further demonstrate the applicability of our lightweight backbones in personalized generation and image-to-image translation. Deployment of our models on edge devices attains 4-second inference. Code and models can be found at: https://github.com/Nota-NetsPresso/BK-SDM." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.128 + ], + "angle": 0, + "content": "B.-K. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.146, + 0.468, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.145, + 0.7, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.268, + 0.789, + 0.297 + ], + "angle": 0, + "content": "Fig. 2: Computation of Stable Diffusion. The denoising U-Net is the main processing bottleneck. THOP [91] is used to measure MACs in generating a \\(512 \\times 512\\) image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.325, + 0.375, + 0.34 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.509 + ], + "angle": 0, + "content": "Stable Diffusion models (SDMs) [58,60-62] are one of the most renowned open-source models for text-to-image (T2I) synthesis, and their exceptional capability has begun to be leveraged as a backbone in several text-guided vision applications [2,3,83,87]. SDMs are T2I-specialized latent diffusion models (LDMs) [58], which employ diffusion operations [23,37,76] in a semantically compressed space for compute efficiency. Within a SDM, a U-Net [9,64] performs iterative sampling to progressively denoise a random latent code and is aided by a text encoder [53] and an image decoder [12,81] to produce text-aligned images. This inference process still involves excessive computational requirements (see Fig. 2), which often hinder the utilization of SDMs despite their rapidly growing usage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.6 + ], + "angle": 0, + "content": "To alleviate this issue, numerous approaches toward efficient SDMs have been introduced. A pretrained diffusion model is distilled to reduce the number of denoising steps, enabling an identically architectured model with fewer sampling steps [40, 43]. Post-training quantization [24, 34, 74] and implementation optimization [6, 7] methods are also leveraged. However, the removal of architectural elements in large diffusion models remains less explored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.735 + ], + "angle": 0, + "content": "This study unlocks the immense potential of classical architectural compression in attaining smaller and faster diffusion models. We eliminate multiple residual and attention blocks from the U-Net of a SDM and retrain it with feature-level knowledge distillation (KD) [18, 63] for general-purpose T2I. Under restricted training resources, our compact models can mimic the original SDM by leveraging transferred knowledge. Our work effectively reduces the computation of SDM-v1.4 [60] and SDM-v2.1-base [62] while achieving compelling zero-shot results on par with multi-billion parameter models [10, 11, 56]. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.749, + 0.79, + 0.84 + ], + "angle": 0, + "content": "We compress SDMs by removing architectural blocks from the U-Net, achieving up to \\(51\\%\\) reduction in model size and \\(43\\%\\) improvement in latency on CPU and GPU. Previous pruning studies [13,47,84] focused on small models (\\(< 100\\mathrm{M}\\) parameters) like ResNet50 and DeiT-B, not on foundation models like SDMs (\\(>1,000\\mathrm{M} = 1\\mathrm{B}\\)), possibly due to the lack of economic retraining for such large models. Moreover, U-Net architectures are arguably more com" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.177 + ], + "angle": 0, + "content": "plex due to the necessity of considering skip connections across the network, making the structural block removal inside them not straightforward." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.178, + 0.786, + 0.282 + ], + "angle": 0, + "content": "To the best of our knowledge, we first demonstrate the notable benefit of feature distillation for training diffusion models, which enables competitive T2I even with significantly fewer resources (using only 13 A100 days and 0.22M LAION pairs [70]). Considering the vast expense of training SDMs from scratch (surpassing 6,000 A100 days and 2,000M pairs), our study indicates that network compression is a remarkably cost-effective strategy in building compact general-purpose diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.284, + 0.786, + 0.357 + ], + "angle": 0, + "content": "We show the practicality of our work across various aspects. Our lightweight backbones are readily applicable to customized generation [65] and image-to-image translation [44], effectively lowering finetuning and inference costs. T2I synthesis on Jetson AGX Orin and iPhone 14 using our models takes less than 4 seconds." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.36, + 0.786, + 0.405 + ], + "angle": 0, + "content": "We have publicly released our approach, model weights, and source code, motivating subsequent works by other researchers (e.g., block pruning and KD for the SDM-v1 variant [5,72] and SDXL [31, 73])." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.429, + 0.388, + 0.445 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Large T2I diffusion models. By gradually removing noise from corrupted data, diffusion-based generative models [9, 22, 76] enable high-fidelity synthesis with broad mode coverage. Integrating these merits with the advancement of pretrained language models [8, 53, 54] has significantly improved the quality of T2I synthesis. In GLIDE [48] and Imagen [66], a text-conditional diffusion model generates a small image, which is upsampled via super-resolution modules. In DALL-E-2 [55], a text-conditional prior network produces an image embedding, which is transformed into an image via a diffusion decoder and further upscaled into higher resolutions. SDMs [58, 60–62] perform the diffusion modeling in a low-dimensional latent space constructed through a pixel-space autoencoder. We use SDMs as our baseline because of its open-access and gaining popularity over numerous downstream tasks [2, 3, 65, 83]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Efficient diffusion models. Several studies have addressed the slow sampling process. Diffusion-tailored distillation [42, 43, 68] progressively transfers knowledge from a pretrained diffusion model to a fewer-step model with the same architecture. Fast high-order solvers [38, 39, 88] for diffusion ordinary differential equations boost the sampling speed. Complementarily, our network compression approach reduces per-step computation and can be easily integrated with less sampling steps. Leveraging quantization [24, 34, 74] and implementation optimizations [6, 7] for SDMs can also be combined with our compact models for further efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Distillation-based compression. KD enhances the performance of small-size models by exploiting output-level [21, 49] and feature-level [18, 63, 85] information of large source models. Although this classical KD has been actively used for efficient GANs [33, 57, 86], its power has not been explored for structurally compressed diffusion models. Distillation pretraining enables small yet capable" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "B.-K. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.193 + ], + "angle": 0, + "content": "general-purpose language models [26, 69, 77] and vision transformers [16, 80]. Beyond such models, we show that its success can be extended to diffusion models with iterative sampling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.789, + 0.375 + ], + "angle": 0, + "content": "Concurrent studies. SnapFusion [35] and MobileDiffusion [89] achieve an efficient U-Net for SDMs through architecture optimization and step reduction. Würstchen [50] introduces two diffusion processes on low- and high-resolution latent spaces for economic training. These works are valuable but require much larger training resources than our work (see Tab. 1). In contrast, we utilize considerably fewer resources, leveraging the surprising benefits of classical KD for foundational diffusion models. While not demonstrated on SDMs, Diff-Pruning [14] proposes structured pruning based on Taylor expansion tailored for diffusion models. We emphasize the use of depth (block) pruning in our work, which often leads to greater enhancements in inference speeds compared to the width (channel) pruning approach of Diff-Pruning. Moreover, Tab. 5 employs a comparable Taylor pruning criterion, which results in an insufficient decrease in MACs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.399, + 0.461, + 0.417 + ], + "angle": 0, + "content": "3 Compression Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.525 + ], + "angle": 0, + "content": "We compress the U-Net [64] in SDMs, which is the most compute-heavy component (see Fig. 2). Conditioned on the text and time-step embeddings, the U-Net performs multiple denoising steps on latent representations. At each step, the U-Net produces the noise residual to compute the latent for the next step. We reduce this per-step computation, leading to Block-removed Knowledge-distilled SDMs (BK-SDMs)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.548, + 0.506, + 0.563 + ], + "angle": 0, + "content": "3.1 Compact U-Net Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.787, + 0.606 + ], + "angle": 0, + "content": "The following architectures are obtained by compressing SDM-v1 (1.04B parameters), as shown in Fig. 3:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.62, + 0.604, + 0.634 + ], + "angle": 0, + "content": "\\(\\circ\\) BK-SDM-Base (0.76B) obtained with Sec. 3.1.(1)." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.636, + 0.633, + 0.65 + ], + "angle": 0, + "content": "\\(\\circ\\) BK-SDM-Small (0.66B) with Secs. 3.1.(1) and 3.1.(2)." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.651, + 0.687, + 0.666 + ], + "angle": 0, + "content": "\\(\\circ\\) BK-SDM-Tiny (0.50B) with Secs. 3.1.(1), 3.1.(2), and 3.1.(3)." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.62, + 0.687, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.787, + 0.711 + ], + "angle": 0, + "content": "Our approach can be identically applied to SDM-v2 (1.26B parameters), leading to BK-SDM-v2-{Base (0.98B), Small (0.88B), Tiny (0.72B)}." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "(1) Fewer Blocks in the Down and Up Stages. This approach is closely aligned with DistilBERT [69] which halves the number of layers and initializes the compact model with the original weights by benefiting from the shared dimensionality. In the original U-Net, each stage with a common spatial size consists of multiple blocks, and most stages contain pairs of residual (R) [17] and cross-attention (A) [25,82] blocks. We hypothesize the existence of some unnecessary pairs and use the following removal strategies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.29 + ], + "angle": 0, + "content": "Fig. 3: Block removal from the denoising U-Net. Our approach is applicable to all the SDM versions in v1 and v2, which share the same U-Net block configuration. For experiments, we used v1.4 [60] and v2.1-base [62]. See Sec. A for the details." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.303, + 0.385, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.388, + 0.385, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.463, + 0.387, + 0.545 + ], + "angle": 0, + "content": "Fig. 4: Minor impact of removing the mid-stage from the U-Net. Results without retraining. See Sec. B for additional results." + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.303, + 0.785, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.39, + 0.785, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.475, + 0.788, + 0.545 + ], + "angle": 0, + "content": "Fig.5: Importance of (a) each block and (b) each group of paired/triplet blocks. Higher score implies removable blocks. The results are aligned with our architectures (e.g., removal of innermost stages and the second R-A pairs in down stages). See Sec. C for further analysis." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.586, + 0.789, + 0.751 + ], + "angle": 0, + "content": "- Down Stages. We maintain the first R-A pairs while eliminating the second pairs, because the first pairs process the changed spatial information and would be more important than the second pairs. This design is consistent with the sensitivity analysis that measures the block-level significance (see Fig. 5). Our approach also does not harm the dimensionality of the original U-Net, enabling the use of the corresponding pretrained weights for initialization [69]. \n- Up Stages. While adhering to the aforementioned scheme, we retain the third R-A pairs. This allows us to utilize the output feature maps at the end of each down stage and the corresponding skip connections between the down and up stages. The same process is applied to the innermost down and up stages that contain only R blocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "(2) Removal of the Entire Mid-Stage. Surprisingly, removing the entire mid-stage from the original U-Net does not noticeably degrade the generation quality while effectively reducing the parameters by \\(11\\%\\) (see Fig. 4). This observation is consistent with the minor role of inner layers in the U-Net generator of GANs [28]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.128 + ], + "angle": 0, + "content": "B.-K. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.145, + 0.675, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.242, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Fig. 6: Distillation-based retraining. The block-removed U-Net is trained effectively through the guidance of the original U-Net." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.789, + 0.378 + ], + "angle": 0, + "content": "Integrating the mid-stage removal with fewer blocks in Sec. 3.1 further decreases compute burdens (Tab. 2) at the cost of a slight decline in performance (Tab. 1). Therefore, we offer this mid-stage elimination as an option, depending on the priority between compute efficiency (using BK-SDM-Small) and generation quality (BK-SDM-Base)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.396, + 0.79, + 0.458 + ], + "angle": 0, + "content": "(3) Further Removal of the Innermost Stages. For additional compression, the innermost down and up stages can also be pruned, leading to our lightest model BK-SDM-Tiny. This implies that outer stages with larger spatial dimensions and their skip connections play a crucial role in the U-Net for T2I synthesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.613 + ], + "angle": 0, + "content": "Alignment with Pruning Sensitivity Analysis. To support the properness of our architectures, we measure the importance of each block (see Fig. 5) and show that unimportant blocks match with our design choices. The importance is measured by how generation scores vary when removing each residual or attention block from the U-Net. A significant drop in performance highlights the essential role of that block. Note that some blocks are not directly removable due to different channel dimensions between input and output; we replace such blocks with channel interpolation modules (denoted by “*” in Fig. 5) to mimic the removal while retaining the information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.675 + ], + "angle": 0, + "content": "The sensitivity analysis implies that the innermost down-mid-up stages and the second R-A pairs in the down stages play relatively minor roles. Pruning these blocks aligns with our architectures, designed based on human knowledge (e.g., prioritizing blocks with altered channel dimensions) and previous studies [28, 69]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.721 + ], + "angle": 0, + "content": "Though some results aligned, the importance criterion from the CLIP Score pruning sensitivity (in Fig. 5) leads to excessive pruning of attention blocks, eventually yielding inferior performance compared to our approach (see Table 5)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.502, + 0.758 + ], + "angle": 0, + "content": "3.2 Distillation-based Retraining" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.81 + ], + "angle": 0, + "content": "For general-purpose T2I, we train our block-removed U-Net to mimic the behavior of the original U-Net (see Fig. 6). To obtain the input of U-Net, we use pretrained-and-frozen encoders [58] for images and text prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Given the latent representation \\(\\mathbf{z}\\) of an image and its paired text embedding \\(\\mathbf{y}\\), the task loss for the reverse denoising process [22, 58] is computed as:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.158, + 0.788, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T a s k}} = \\mathbb {E} _ {\\mathbf {z}, \\epsilon , \\mathbf {y}, t} \\left[ | | \\epsilon - \\epsilon_ {\\mathrm {S}} (\\mathbf {z} _ {t}, \\mathbf {y}, t) | | _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.189, + 0.789, + 0.25 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t\\) is a noisy latent code from the diffusion process [22] with the sampled noise \\(\\epsilon \\sim N(\\mathbf{0},\\mathbf{I})\\) and time step \\(t \\sim \\mathrm{Uniform}(1,T)\\), and \\(\\epsilon_{\\mathrm{S}}(\\circ)\\) indicates the estimated noise from our compact U-Net student. For brevity, we omit the subscripts of \\(\\mathbb{E}_{\\mathbf{z},\\epsilon,\\mathbf{y},t}[\\circ]\\) in the following notations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.25, + 0.787, + 0.281 + ], + "angle": 0, + "content": "The compact student is also trained to imitate the outputs of the original U-Net teacher, \\(\\epsilon_{\\mathrm{T}}(\\circ)\\), with the following output-level KD objective [21]:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.29, + 0.788, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {O u t K D}} = \\mathbb {E} \\left[ \\left| \\left| \\epsilon_ {\\mathrm {T}} (\\mathbf {z} _ {t}, \\mathbf {y}, t) - \\epsilon_ {\\mathrm {S}} (\\mathbf {z} _ {t}, \\mathbf {y}, t) \\right| \\right| _ {2} ^ {2} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.323, + 0.787, + 0.355 + ], + "angle": 0, + "content": "A key to our approach is feature-level KD [18, 63] that provides abundant guidance for the student's training:" + }, + { + "type": "equation", + "bbox": [ + 0.332, + 0.363, + 0.788, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {F e a t K D}} = \\mathbb {E} \\left[ \\sum_ {l} | | f _ {\\mathrm {T}} ^ {l} (\\mathbf {z} _ {t}, \\mathbf {y}, t) - f _ {\\mathrm {S}} ^ {l} (\\mathbf {z} _ {t}, \\mathbf {y}, t) | | _ {2} ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.404, + 0.789, + 0.51 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{T}}^{l}(\\circ) \\) and \\( f_{\\mathrm{S}}^{l}(\\circ) \\) represent the feature maps of the \\( l \\)-th layer in a predefined set of distilled layers from the teacher and the student, respectively. While learnable regressors (e.g., \\( 1 \\times 1 \\) convolutions to match the number of channels) have been commonly used [57,63,75], our approach circumvents this requirement. By applying distillation at the end of each stage in both models, we ensure that the dimensionality of the feature maps already matches, thus eliminating the need for additional regressors." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.511, + 0.787, + 0.542 + ], + "angle": 0, + "content": "The final objective is shown below, and we simply set \\(\\lambda_{\\mathrm{OutKD}}\\) and \\(\\lambda_{\\mathrm{FeatKD}}\\) as 1. Without loss-weight tuning, our approach is effective in empirical validation." + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.557, + 0.788, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {T a s k}} + \\lambda_ {\\text {O u t K D}} \\mathcal {L} _ {\\text {O u t K D}} + \\lambda_ {\\text {F e a t K D}} \\mathcal {L} _ {\\text {F e a t K D}}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.448, + 0.615 + ], + "angle": 0, + "content": "4 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.628, + 0.787, + 0.703 + ], + "angle": 0, + "content": "Distillation Retraining. We primarily use 0.22M image-text pairs from LAION-Aesthetics V2 (L-Aes) \\(6.5 + [70,71]\\), which are significantly fewer than the original training data used for SDMs [60,62] (\\(>2,000\\mathrm{M}\\) pairs). In Fig. 11, dataset sizes smaller than 0.22M are randomly sampled from L-Aes \\(6.5+\\), while those larger than 0.22M are from L-Aes \\(6.25+\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.704, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Zero-shot T2I Evaluation. Following the popular protocol [56, 58, 66], we use 30K prompts from the MS-COCO validation split [36], downsample the \\(512 \\times 512\\) generated images to \\(256 \\times 256\\), and compare them with the entire validation set. We compute Fréchet Inception Distance (FID) [20] and Inception Score (IS) [67] to assess visual quality. We measure CLIP score [19, 53] with CLIP-ViT-g/14 model to assess text-image correspondence." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Downstream Tasks. For personalized generation, we use the DreamBooth dataset [65] (30 subjects \\(\\times\\) 25 prompts \\(\\times\\) 4~6 images) and perform per-subject finetuning. Following the evaluation protocol [65], we use ViT-S/16 model [4]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "B.-K. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Results on zero-shot MS-COCO \\(256 \\times 256\\) 30K. Training resources include image-text pairs, batch size, iterations, and A100 days. Despite far smaller resources, our compact models outperform prior studies [10, 11, 56, 79, 90], showing the benefit of compressing existing powerful models. Note that FID fluctuates more than the other metrics over training progress in our experiments (see Figs. 9 and 11)." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.228, + 0.782, + 0.505 + ], + "angle": 0, + "content": "
Model\nNameType #Param‡Generation ScoreTraining Resource
FID↓IS↑CLIP↑Data Size (Batch, # Iter)A100 Days
SDM-v1.4 [58,60]†DF1.04B13.0536.760.2958>2000M*(2048, 1171K)6250
Small Stable Diffusion [51]†DF0.76B12.7632.330.2851229M(128, 1100K)-
BK-SDM-Base [Ours]†DF0.76B15.7633.790.28780.22M(256, 50K)13
BK-SDM-Small [Ours]†DF0.66B16.9831.680.26770.22M(256, 50K)13
BK-SDM-Tiny [Ours]†DF0.50B17.1230.090.26530.22M(256, 50K)13
SDM-v2.1-base [58,62]†DF1.26B13.9335.930.3075>2000M*(2048, 1620K)8334
BK-SDM-v2-Base [Ours]†DF0.98B15.8531.700.28680.22M(128, 50K)4
BK-SDM-v2-Small [Ours]†DF0.88B16.6131.730.29010.22M(128, 50K)4
BK-SDM-v2-Tiny [Ours]†DF0.72B15.6831.640.28970.22M(128, 50K)4
DALL-E [56]AR12B27.517.9-250M(1024, 430K)-
CogView [10]AR4B27.118.2-30M(6144, 144K)-
CogView2 [11]AR6B2422.4-30M(4096, 300K)-
Make-A-Scene [15]AR4B11.84--35M(1024, 170K)-
LAFITE [90]GAN0.23B26.9426.02-3M--
GALIP (CC12M) [79]†GAN0.32B13.8625.160.281712M--
GigaGAN [27]GAN1.1B9.09-->100M*(512, 1350K)4783
GLIDE [48]DF3.5B12.24--250M(2048, 2500K)-
LDM-KL-8-G [58]DF1.45B12.6330.29-400M(680, 390K)-
DALL-E-2 [55]DF5.2B10.39--250M(4096, 3400K)-
SnapFusion [35]DF0.99B~13.6-~0.295>100M*(2048, -)>128*
Würstchen-v2 [50]†DF3.1B22.4032.870.26761700M(1536, 1725K)1484
MobileDiffusion [89]†DF0.52B9.01--150M(2048, 1000K)-
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.221, + 0.507, + 0.782, + 0.53 + ], + "angle": 0, + "content": "\\(\\dagger\\): Evaluated with the released checkpoints. \\(\\ddagger\\): Total parameters for T2I synthesis. \\(*\\): Estimated based on public information. DF and AR: diffusion and autoregressive models. \\(\\downarrow\\) and \\(\\uparrow\\): lower and higher values are better." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.597 + ], + "angle": 0, + "content": "for DINO score and CLIP-ViT-g/14 model for CLIP-I and CLIP-T scores. For image-to-image translation, input images are sourced from Meng et al. [43]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Implementation Details. We adjust the codes in Diffusers [52] and PEFT [41]. We use a single NVIDIA A100 80G GPU for main retraining and a single NVIDIA GeForce RTX 3090 GPU for per-subject finetuning. For compute efficiency, we always opt for 25 denoising steps of the U-Net at the inference phase, unless specified. The classifier-free guidance scale [23,66] is set to the default value of 7.5. The latent resolution is set to the default (\\(H = W = 64\\) in Fig. 3), yielding \\(512 \\times 512\\) images. See Sec. J for the details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.746, + 0.325, + 0.761 + ], + "angle": 0, + "content": "5 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "All the results in Secs. 5.1-5.6 were obtained with the full benchmark protocol (MS-COCO \\(256 \\times 256\\) 30K samples), except for Fig. 10 (\\(512 \\times 512\\) 5K samples). Unless specified, the training setup in Tab. 1 was used." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.218, + 0.306, + 0.778, + 0.315 + ], + "angle": 0, + "content": "Prompt: A bowl that has vegetables inside of it; A brown and white cat staring off with pretty green eyes.; A toy raccoon standing on a pile of broccoli." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.354 + ], + "angle": 0, + "content": "Fig. 7: Visual comparison with open-sourced models. The results [11,48,50,79,90] were obtained with their official codes." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.453 + ], + "angle": 0, + "content": "Table 2: Impact of compute reduction in U-Net on the entire SDM. The number of sampling steps is indicated with the parentheses, e.g., U-Net (1) for one step. The full computation (denoted by \"Whole\") covers the text encoder, U-Net, and image decoder. All corresponding values are obtained on the generation of a single \\(512 \\times 512\\) image with 25 denoising steps. The latency was measured on Xeon Silver 4210R CPU 2.40GHz and NVIDIA GeForce RTX 3090 GPU." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.467, + 0.789, + 0.561 + ], + "angle": 0, + "content": "
Model# ParamMACsCPU LatencyGPU Latency
U-NetWholeU-Net (1)U-Net (25)WholeU-Net (1)U-Net (25)WholeU-Net (1)U-Net (25)Whole
SDM-v1.4 [60]860M1033M339G8469G9716G5.63s146.28s153.00s0.049s1.28s1.41s
BK-SDM-Base [Ours]580M752M224G5594G6841G3.84s99.95s106.67s0.032s0.83s0.96s
(-32.6%)(-27.1%)(-33.9%)(-33.9%)(-29.5%)(-31.8%)(-31.7%)(-30.3%)(-34.6%)(-35.2%)(-31.9%)
BK-SDM-Small [Ours]483M655M218G5444G6690G3.45s89.78s96.50s0.030s0.77s0.90s
(-43.9%)(-36.5%)(-35.7%)(-35.7%)(-31.1%)(-38.7%)(-38.6%)(-36.9%)(-38.7%)(-39.8%)(-36.1%)
BK-SDM-Tiny [Ours]324M496M206G5126G6373G3.03s78.77s85.49s0.026s0.67s0.80s
(-62.4%)(-51.9%)(-39.5%)(-39.5%)(-34.4%)(-46.2%)(-46.1%)(-44.1%)(-46.9%)(-47.7%)(-43.2%)
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.589, + 0.536, + 0.604 + ], + "angle": 0, + "content": "5.1 Comparison with Existing Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.673 + ], + "angle": 0, + "content": "Quantitative Comparison. Tab. 1 shows the zero-shot results for general-purpose T2I. Despite being trained with only 0.22M samples and having fewer than 1B parameters, our compressed models demonstrate competitive performance on par with existing large models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.787, + 0.765 + ], + "angle": 0, + "content": "Visual Comparison. Fig. 7 depicts synthesized images with some MS-COCO captions. Our compact models inherit the superiority of SDM and produce more photorealistic images compared to the AR-based [11] and GAN-based [79, 90] baselines. Noticeably, the same latent code results in a shared visual style between the original and our models (6th–9th columns in Fig. 7), similar to the observation in transfer learning for GANs [45]. See Sec. D for additional results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.431, + 0.801 + ], + "angle": 0, + "content": "5.2 Computational Gain" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Tab. 2 shows how the compute reduction for each sampling step of the U-Net affects the overall process. The per-step reduction effectively decreases MACs" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "B.-K. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.477, + 0.173 + ], + "angle": 0, + "content": "Table 3: Quality gains from distillation retraining. MS-COCO 30K." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.187, + 0.472, + 0.303 + ], + "angle": 0, + "content": "
BK-SDM\nTypeKDGenerationScore# Param\nU-Net
FID↓IS↑CLIP↑
BaseX23.5723.020.24080.58B
✓15.7633.790.2878
v2-BaseX16.7625.880.26610.59B
✓15.8531.700.2868
v2-SmallX16.7125.770.26550.49B
✓16.6131.730.2901
v2-TinyX16.8726.060.26780.33B
✓15.6831.640.2897
" + }, + { + "type": "table_caption", + "bbox": [ + 0.481, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 4: Significance of each element in transferred knowledge. Results of BK-SDMv2-Small on MS-COCO 30K." + }, + { + "type": "table", + "bbox": [ + 0.487, + 0.203, + 0.782, + 0.301 + ], + "angle": 0, + "content": "
Weight InitializationOutput FeatureGeneration Score
KDKDFID↓IS↑CLIP↑
RandomXX41.7515.420.1733
Random✓X29.0118.290.1967
Random✓✓24.4722.370.2323
TeacherXX16.7125.770.2655
Teacher✓X14.2729.470.2777
Teacher✓✓16.6131.730.2901
" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.314, + 0.753, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Fig. 8: Image areas affected by each word. KD enables our models to mimic the SDM, yielding similar per-word attribution maps. The model without KD behaves differently, causing dissimilar maps and inaccurate generation (e.g., two sheep and unusual bird shapes)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.601 + ], + "angle": 0, + "content": "and inference time by more than \\(30\\%\\). Notably, BK-SDM-Tiny has \\(50\\%\\) fewer parameters than the original SDM." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.536, + 0.637 + ], + "angle": 0, + "content": "5.3 Benefit of Distillation Retraining" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.749 + ], + "angle": 0, + "content": "T2I Performance. Tab. 3 summarizes the results from ablating the total KD objective (Eq. 2+Eq. 3). Across various model types, distillation brings a clear improvement in generation quality. Tab. 4 analyzes the effect of each element in transferred knowledge. Exploiting output-level KD (Eq. 2) boosts the performance compared to using only the denoising task loss. Leveraging feature-level KD (Eq. 3) yields further score enhancement. Additionally, using the teacher weights for initialization is highly beneficial." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Cross-attention Resemblance. Fig. 8 displays the per-word attribution maps [78] created by aggregating cross-attention scores over spatiotemporal dimensions. The attribution maps of our models are semantically and spatially similar to those of the original model, indicating the merit of supervisory signals at multiple stages via KD. In contrast, the baseline model without KD activates incorrect areas, leading to text-mismatched generation results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.339, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.146, + 0.46, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.146, + 0.583, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.267, + 0.587, + 0.308 + ], + "angle": 0, + "content": "Fig.9: Zero-shot results over training progress. The architecture size of BK-SDM, usage of KD, and batch size are denoted. Results on MS-COCO 30K." + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.144, + 0.787, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.254, + 0.788, + 0.308 + ], + "angle": 0, + "content": "Fig.10:Trade-off curves Left:FID vs.CLIP score Right: quality vs.efficiency Ours-Base on MS-COCO 5K." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.488 + ], + "angle": 0, + "content": "Training Progress. Fig. 9 shows the results over training iterations. Without KD, training solely with the denoising task loss causes fluctuations or sudden drops in performance (indicated with green and cyan). In contrast, distillation (purple and pink) stabilizes and accelerates the training process, demonstrating the benefit of providing sufficient hints for training guidance. Notably, our small-and tiny-size models trained with KD (yellow and red) outperform the bigger base-size model without KD (cyan). Additionally, while the best FID score is observed early on, IS and CLIP score exhibit ongoing improvement, implying that judging models merely with FID may be suboptimal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Trade-off Results. Fig. 10 shows the results of BK-SDM-Base with and without KD on MS-COCO \\(512 \\times 512\\) 5K. Higher classifier-free guidance scales [23,66] lead to better text-aligned images at the cost of less diversity. More denoising steps improve generation quality at the cost of slower inference. Distillation retraining achieves much better trade-off curves than the baseline without KD." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.585, + 0.627, + 0.6 + ], + "angle": 0, + "content": "5.4 Comparison with Different Pruning Criteria" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.789, + 0.716 + ], + "angle": 0, + "content": "We extend our baseline comparisons by calculating several block-level importance scores [29].1 Furthermore, the impact of applying KD at the end of each stage is also examined. Tab. 5 shows the results. Magnitude (Mag) pruning [32] removes critical outer blocks, causing irrecoverable loss. Taylor pruning [14,30,46] removes only the inner part, causing ineffective reduction in MACs and speed; additionally, this method demands the intensive calculation of backward gradients. The CLIP Score method (based on Fig. 5) results in severe pruning of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.724, + 0.789, + 0.84 + ], + "angle": 0, + "content": "1 Let the \\((h_{\\mathrm{out}}, h_{\\mathrm{in}})\\)-size weight matrix in the \\(b\\)-th block be \\(\\mathbf{W}^{l,b} = \\left[W_{i,j}^{l,b}\\right]\\), where \\(l\\) denotes the layer type (e.g., convolution (flattened) or attention's key projection). The scores at the output neuron level [1] are aggregated for the block-level importance criteria, \\(S_{\\mathrm{Magnitude}}^b = \\mathbb{E}_{l,i}\\left[\\sum_j\\left|W_{i,j}^{l,b}\\right|\\right]\\) and \\(S_{\\mathrm{Taylor}}^b = \\mathbb{E}_{l,i}\\left[\\sum_j\\left|\\frac{\\partial\\mathcal{L}(D)}{\\partial W_{i,j}^{l,b}}W_{i,j}^{l,b}\\right|\\right]\\). Here, \\(\\mathcal{L}\\) and \\(D\\) denote the denoising task loss and a calibration set of 1K samples. The final scores are then ranked to remove unimportant blocks, or to replace them with interpolation for unremovable blocks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "B.-K. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.528, + 0.228 + ], + "angle": 0, + "content": "Table 5: Different block-level criteria at similar parameters. Taylor pruning removes solely the inner blocks, leading to insufficient reduction in MACs. Our method attains a favorable compromise between performance and inference speed. Results on MS-COCO 30K." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.242, + 0.524, + 0.306 + ], + "angle": 0, + "content": "
Pruning \nCriterionFID↓CLIP Score↑U-Net
KD×KD√KD×KD√MACs# Param
Mag [32]>150>150<0.1<0.1122.2G495.5M
Taylor [14,30]18.8516.730.24290.2749305.4G493.4M
CLIP [Fig. 5]32.9321.860.19360.2283204.8G496.8M
Ours-Small22.9716.830.22870.2668217.7G482.3M
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.22, + 0.307, + 0.51, + 0.327 + ], + "angle": 0, + "content": "Retraining with batch 128, 0.22M data, and 50K iters. Original SD-v1.4's U-Net: (MACs, Params) = (338.7G, 859.5M)." + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.165, + 0.786, + 0.194 + ], + "angle": 0, + "content": "Table 6: A/B-testing preference study. Total 2,501 comparisons." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.208, + 0.784, + 0.268 + ], + "angle": 0, + "content": "
Ours-SmallGLIDE-filt [48]GALIP [79]Würstchen-v2 SDM-v1.4 [50][60]
# Win†453 (71.5%)379 (61.5%)319 (51.4%)202 (32.1%)
# Defeat†181 (28.5%)237 (38.5%)302 (48.6%)428 (67.9%)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.535, + 0.27, + 0.784, + 0.307 + ], + "angle": 0, + "content": "Participants selected preferred images from pairs with the same prompts. One image in each pair was always from BK-SDM-Small, while the other was randomly chosen from the four methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.355, + 0.531, + 0.383 + ], + "angle": 0, + "content": "Table 7: Impact of training batch size and iterations. Results on MS-COCO 30K." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.397, + 0.528, + 0.505 + ], + "angle": 0, + "content": "
BK-SDMBaseSmallTiny
Batch Size642566425664256
FID↓14.6115.7616.8716.9817.2817.12
IS↑31.4433.7929.5131.6828.3330.09
CLIP↑0.28260.28780.26440.26770.26070.2653
BK-SDMBase (Data 2.3M)Small (Data 2.3M)Tiny (Data 2.3M)
#Iter50K100K50K100K50K100K
FID↓14.8115.3917.0517.0117.5317.63
IS↑34.1734.7633.1033.1431.3232.26
CLIP↑0.28830.28890.27340.27540.26900.2713
" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.34, + 0.66, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.34, + 0.785, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.466, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 11: Impact of training data volume. Results of BK-SDM-Small on MS-COCO 30K." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.626 + ], + "angle": 0, + "content": "attention blocks, adversely affecting performance. Our approach achieves superior or comparable results against the benchmark methods and can be directly applied to all the SDM versions in v1 and v2. This advantage comes without the extra phase of calculating pruning criteria while ensuring a favorable balance between performance and inference speed. Notably, KD always remains effective." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.649, + 0.517, + 0.663 + ], + "angle": 0, + "content": "5.5 Human Preference Assessment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Despite their extensive use, automated metrics for assessing image quality (e.g., FID) do not consistently match with human evaluation [55, 66]. To further address this, we have conducted an A/B testing-based study: participants were shown pairs of images generated using the same MS-COCO prompt and asked to choose their preferred image in relation to the prompt. One image in each pair was always from our model, while the other was randomly selected from one of four comparative methods. A total of 25 participants conducted 2,501 comparisons online. The participants had no bias towards the examined models and were not compensated. The prompts were randomly drawn from the MS-COCO set and varied for 3 to 5 participants. As shown in Tab. 6, our model (0.66B parameters) excels over GLIDE-filtered, GALIP, and Wüstchen-v2 (3.1B)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.603, + 0.215 + ], + "angle": 0, + "content": "Table 8: Personalized generation with finetuning over different backbones. Our compact models can preserve subject fidelity (DINO and CLIP-I) and prompt fidelity (CLIP-T) of the original SDM with reduced finetuning (FT) costs and fewer parameters." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.228, + 0.597, + 0.324 + ], + "angle": 0, + "content": "
Backbone {Param}DINO↑CLIP-I↑CLIP-T↑FT (Time, Mem)†
SDM-v1.4 [60] {1.04B}0.7280.7250.263(882s, 23.0GB)
BK-SDM-Base {0.76B}0.7230.7170.260(623s, 18.7GB)
BK-SDM-Small {0.66B}0.7200.7050.259(604s, 17.2GB)
BK-SDM-Tiny {0.50B}0.7150.6930.261(560s, 13.1GB)
Base (Batch 64) {0.76B}0.7180.7080.262(623s, 18.7GB)
- No KD & Random Init.0.5940.4650.191(623s, 18.7GB)
- No KD & Teacher Init.0.7160.6690.258(623s, 18.7GB)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.221, + 0.326, + 0.597, + 0.338 + ], + "angle": 0, + "content": "†: Per-subject FT time and GPU memory for 800 iterations on RTX 3090." + }, + { + "type": "table_caption", + "bbox": [ + 0.607, + 0.169, + 0.787, + 0.196 + ], + "angle": 0, + "content": "Table 9: Speedups on edge devices." + }, + { + "type": "table", + "bbox": [ + 0.612, + 0.211, + 0.784, + 0.314 + ], + "angle": 0, + "content": "
ModelAGX Orin 32GBiPhone 14
SDM-v14.9s [61]5.6s [60]
BK-SDM-Base3.4s (-31%)4.0s (-29%)
BK-SDM-Small3.2s (-35%)3.9s (-29%)
BK-SDM-Tiny2.8s (-43%)3.9s (-29%)
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.351, + 0.59, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.352, + 0.785, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.482, + 0.788, + 0.511 + ], + "angle": 0, + "content": "Fig. 12: Visual results of personalized generation. Each subject is marked as \"a [identifier] [class noun]\" (e.g., \"a [V] dog\")." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.526, + 0.407, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.526, + 0.597, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.639, + 0.544, + 0.648 + ], + "angle": 0, + "content": "(left-to-right, top-to-bottom) Target style: Art station; Oil painting; Photograph; Watercolor" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.526, + 0.787, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.646, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Fig. 14: Small LDM for face synthesis." + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.661, + 0.591, + 0.675 + ], + "angle": 0, + "content": "Fig. 13: Text-guided image-to-image translation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.72, + 0.642, + 0.735 + ], + "angle": 0, + "content": "5.6 Impact of Training Resources on Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Consistent with existing evidence, the use of larger batch sizes, more extensive data, and longer iterations for training enhances performance in our work (see Tab. 7, Fig. 11, and Sec. H). However, this benefit requires increased resource demands (e.g., extended training days without multiple high-spec GPUs and greater data storage capacity). As such, despite the better performing models in Tab. 7, we primarily report the models with fewer resources. We believe that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.128 + ], + "angle": 0, + "content": "B.-K. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "accessible training costs by many researchers can help drive advancement in massive models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.203, + 0.359, + 0.218 + ], + "angle": 0, + "content": "5.7 Application" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Personalized T2I Synthesis. Tab. 8 compares the fine-tuning results with DreamBooth [65] over different backbones to create images about a given subject. BK-SDMs can preserve \\(95\\% \\sim 99\\%\\) scores of the original SDM while cutting fine-tuning costs. Fig. 12 depicts that our models can accurately capture the subject details and generate various scenes. Over the models retrained with a batch size of 64, the baselines without KD fail to generate the subjects or cannot maintain the identity details. See Sec. E for further results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Image-to-Image Translation. Fig. 13 presents the text-guided stylization results with SDEdit [44]. Our model, resembling the ability of the original SDM, faithfully produces images given style-specified prompts and content pictures. See Sec. F for additional results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Deployment on Edge Devices. We deploy our models trained with 2.3M pairs and compare them against the original SDM under the same setup on edge devices (20 denoising steps on NVIDIA Jetson AGX Orin 32GB and 10 steps on iPhone 14). Our models produce a \\(512 \\times 512\\) image within 4 seconds (see Tab. 9), while maintaining acceptable image quality (Fig. 1(d) and Sec. G)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Another LDM. SDMs are derived from LDMs [58], which share a similar U-Net design across many tasks. To validate the generality of our work, we compress an LDM (with 308M parameters and 410K training iterations) for unconditional generation on CelebA-HQ [59] by using the same approach of BK-SDM-Small (187M parameters and 30K iterations). Fig. 14 shows the efficacy of our architecture and distillation retraining." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.592, + 0.509, + 0.608 + ], + "angle": 0, + "content": "6 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.748 + ], + "angle": 0, + "content": "We uncover the potential of architectural compression for general-purpose text-to-image synthesis with a renowned model, Stable Diffusion. Our block-removed lightweight models are effective for zero-shot generation, achieving competitive results against large-scale baselines. Distillation is a key of our method, leading to powerful retraining even under very constrained resources. Our work is orthogonal to previous works for efficient diffusion models, e.g., enabling fewer sampling steps, and can be readily combined with them. We hope our study can facilitate future research on structural compression of large diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Limitations. Our compact models inherit the capability of the source model for high-fidelity image generation, but they have shortcomings such as inaccurate generation of full-body human appearance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Negative Social Impacts. Recent large generative models are capable of creating high-quality plausible content. To avoid causing unintended social usage, researchers should take steps to ensure the appropriateness of the training data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.239 + ], + "angle": 0, + "content": "We thank the Microsoft Startups Founders Hub program and the AI Industrial Convergence Cluster Development project funded by the Ministry of Science and ICT (MSIT, Korea) and Gwangju Metropolitan City for their generous support of GPU resources." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.707, + 0.31 + ], + "angle": 0, + "content": "1. A Simple and Effective Pruning Approach for LLMs. ICLR (2024) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.351 + ], + "angle": 0, + "content": "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.786, + 0.379 + ], + "angle": 0, + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.787, + 0.407 + ], + "angle": 0, + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.787, + 0.449 + ], + "angle": 0, + "content": "5. Castells, T., Song, H.K., Piao, T., Choi, S., Kim, B.K., Yim, H., Lee, C., Kim, J.G., Kim, T.H.: Edgefusion: On-device text-to-image generation. In: CVPR Workshop (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.787, + 0.491 + ], + "angle": 0, + "content": "6. Chen, Y.H., Sarokin, R., Lee, J., Tang, J., Chang, C.L., Kulik, A., Grundmann, M.: Speed is all you need: On-device acceleration of large diffusion models viagpu-aware optimizations. In: CVPR Workshop (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.787, + 0.519 + ], + "angle": 0, + "content": "7. Choi, J., Kim, M., Ahn, D., Kim, T., Kim, Y., Jo, D., Jeon, H., Kim, J.J., Kim, H.: Squeezing large-scale diffusion models for mobile. In: ICML Workshop (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.787, + 0.547 + ], + "angle": 0, + "content": "8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.787, + 0.574 + ], + "angle": 0, + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.576, + 0.787, + 0.617 + ], + "angle": 0, + "content": "10. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. In: NeurIPS (2021) 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.618, + 0.787, + 0.644 + ], + "angle": 0, + "content": "11. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. In: NeurIPS (2022) 2, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.646, + 0.787, + 0.672 + ], + "angle": 0, + "content": "12. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.674, + 0.787, + 0.701 + ], + "angle": 0, + "content": "13. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: CVPR (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.787, + 0.728 + ], + "angle": 0, + "content": "14. Fang, G., Ma, X., Wang, X.: Structural pruning for diffusion models. In: NeurIPS (2023) 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.729, + 0.787, + 0.768 + ], + "angle": 0, + "content": "15. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "16. Hao, Z., Guo, J., Jia, D., Han, K., Tang, Y., Zhang, C., Hu, H., Wang, Y.: Learning efficient vision transformers via fine-grained manifold distillation. In: NeurIPS (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 4" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.296, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "B.-K. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Heo, B., Kim, J., Yun, S., Park, H., Kwak, N., Choi, J.Y.: A comprehensive overhaul of feature distillation. In: ICCV (2019) 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.204 + ], + "angle": 0, + "content": "19. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: EMNLP (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.274 + ], + "angle": 0, + "content": "21. Hinton, G., Vinyls, O., Dean, J.: Distilling the knowledge in a neural network. In: NeurIPS Workshop (2014) 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.275, + 0.785, + 0.302 + ], + "angle": 0, + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.33 + ], + "angle": 0, + "content": "23. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop (2021) 2, 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.36 + ], + "angle": 0, + "content": "24. Hou, J., Asghar, Z.: World's first on-device demonstration of stable diffusion on an android phone. https://wwwQUALCOMM.com/news (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.388 + ], + "angle": 0, + "content": "25. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: ICML (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.429 + ], + "angle": 0, + "content": "26. Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., Wang, F., Liu, Q.: Tinybert: Distilling bert for natural language understanding. In: Findings of EMNLP (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.785, + 0.459 + ], + "angle": 0, + "content": "27. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "28. Kim, B.K., Choi, S., Park, H.: Cut inner layers: A structured pruning strategy for efficient u-net gans. In: ICML Workshop (2022) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "29. Kim, B.K., Kim, G., Kim, T.H., Castells, T., Choi, S., Shin, J., Song, H.K.: Shortened llama: A simple depth pruning for large language models. arXiv preprint arXiv:2402.02834 (2024) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.544 + ], + "angle": 0, + "content": "30. LeCun, Y., Denker, J., Solla, S.: Optimal brain damage. In: NeurIPS (1989) 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.586 + ], + "angle": 0, + "content": "31. Lee, Y., Park, K., Cho, Y., Lee, Y.J., Hwang, S.J.: Koala: Self-attention matters in knowledge distillation of latent diffusion models for memory-efficient and fast image synthesis. arXiv preprint arXiv:2312.04005v1 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.614 + ], + "angle": 0, + "content": "32. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. In: ICLR (2017) 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.642 + ], + "angle": 0, + "content": "33. Li, M., Lin, J., Ding, Y., Liu, Z., Zhu, J.Y., Han, S.: Gan compression: Efficient architectures for interactive conditional gans. In: CVPR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.671 + ], + "angle": 0, + "content": "34. Li, X., Liu, Y., Lian, L., Yang, H., Dong, Z., Kang, D., Zhang, S., Keutzer, K.: Q-diffusion: Quantizing diffusion models. In: ICCV (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.712 + ], + "angle": 0, + "content": "35. Li, Y., Wang, H., Jin, Q., Hu, J., Chemerys, P., Fu, Y., Wang, Y., Tulyakov, S., Ren, J.: Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. In: NeurIPS (2023) 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.742 + ], + "angle": 0, + "content": "36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dolkar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.769 + ], + "angle": 0, + "content": "37. Liu, L., Ren, Y., Lin, Z., Zhao, Z.: Pseudo numerical methods for diffusion models on manifolds. In: ICLR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.798 + ], + "angle": 0, + "content": "38. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In: NeurIPS (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095 (2022) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "40. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "41. Mangrulkar, S., Gugger, S., Debut, L., Belkada, Y., Paul, S.: Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.229, + 0.787, + 0.257 + ], + "angle": 0, + "content": "42. Meng, C., Gao, R., Kingma, D.P., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: NeurIPS Workshop (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.284 + ], + "angle": 0, + "content": "43. Meng, C., Gao, R., Kingma, D.P., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR (2023) 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.324 + ], + "angle": 0, + "content": "44. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022) 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.352 + ], + "angle": 0, + "content": "45. Mo, S., Cho, M., Shin, J.: Freeze the discriminator: a simple baseline for fine-tuning gans. In: CVPR Workshop (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.379 + ], + "angle": 0, + "content": "46. Molchanov, P., Mallya, A., Tyree, S., Frosio, I., Kautz, J.: Importance estimation for neural network pruning. In: CVPR (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.419 + ], + "angle": 0, + "content": "47. Murti, C., Narshana, T., Bhattacharyya, C.: TVSPrune - pruning non-discriminative filters via total variation separability of intermediate representations without fine tuning. In: ICLR (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "48. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: ICML (2022) 3, 8, 9, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.487 + ], + "angle": 0, + "content": "49. Park, W., Kim, D., Lu, Y., Cho, M.: Relational knowledge distillation. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.787, + 0.528 + ], + "angle": 0, + "content": "50. Pernias, P., Rampas, D., Richter, M.L., Pal, C.J., Aubreville, M.: Wierstchen: An efficient architecture for large-scale text-to-image diffusion models. In: ICLR (2024) 4, 8, 9, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.555 + ], + "angle": 0, + "content": "51. Pinkney, J.: Small stable diffusion. https://huggingface.co/OFA-Sys/small-stable-diffusion-v0 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.597 + ], + "angle": 0, + "content": "52. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.637 + ], + "angle": 0, + "content": "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.664 + ], + "angle": 0, + "content": "54. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "55. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) 3, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.732 + ], + "angle": 0, + "content": "56. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "57. Ren, Y., Wu, J., Xiao, X., Yang, J.: Online multi-granularity distillation for gan compression. In: ICCV (2021) 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.786 + ], + "angle": 0, + "content": "58. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 3, 6, 7, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "59. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: Ldm on celeba-hq. https://huggingface.co/CompVis/ldm-celebahq-256 (2022) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "60. Rombach, R., Esser, P.: Stable diffusion v1-4. https://huggingface.co/CompVis/stable-diffusion-v1-4 (2022) 2, 3, 5, 7, 8, 9, 12, 13" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "B.-K. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.788, + 0.175 + ], + "angle": 0, + "content": "61. Rombach, R., Esser, P.: Stable diffusion v1-5. https://huggingface.co/runwayml/stable-diffusion-v1-5 (2022) 2, 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.788, + 0.203 + ], + "angle": 0, + "content": "62. Rombach, R., Esser, P., Ha, D.: Stable diffusion v2-1-base. https://huggingface.co/stabilityai/stable-diffusion-2-1-base (2022) 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "63. Romero, A., Ballas, N., Kahou, S.E., Chassang, A., Gatta, C., Bengio, Y.: Fitnets: Hints for thin deep nets. In: ICLR (2015) 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "64. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.299 + ], + "angle": 0, + "content": "65. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023) 3, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.355 + ], + "angle": 0, + "content": "66. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 3, 7, 8, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "67. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: NeurIPS (2016) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.787, + 0.411 + ], + "angle": 0, + "content": "68. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.439 + ], + "angle": 0, + "content": "69. Sanh, V., Debut, L., Chaumont, J., Wolf, T.: Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In: NeurIPS Workshop (2019) 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.466 + ], + "angle": 0, + "content": "70. Schuhmann, C., Beaumont, R.: Laion-aesthetics. https://laion.ai/blog/laion-aesthetics (2022) 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "71. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS Workshop (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.55 + ], + "angle": 0, + "content": "72. Segmind: Segmind-distill-sd. https://github.com/segmind/distill-sd/tree/c1e97a70d141df09e6fe5cc7dbd66e0cbae3eeb (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "73. Segmind: Ssd-1b. https://github.com/segmind/SSD-1B/tree/d2ff723ea8ecf5ddb86f3aac0af1db30e88a2e2d (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.617 + ], + "angle": 0, + "content": "74. Shen, H., Cheng, P., Ye, X., Cheng, W., Abidi, H.: Accelerate stable diffusion with intel neural compressor. https://medium.com/intel-analytics-software (2022) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.646 + ], + "angle": 0, + "content": "75. Shu, C., Liu, Y., Gao, J., Yan, Z., Shen, C.: Channel-wise knowledge distillation for dense prediction. In: ICCV (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.673 + ], + "angle": 0, + "content": "76. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "77. Sun, Z., Yu, H., Song, X., Liu, R., Yang, Y., Zhou, D.: Mobilebert: a compact task-agnostic bert for resource-limited devices. In: ACL (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "78. Tang, R., Liu, L., Pandey, A., Jiang, Z., Yang, G., Kumar, K., Stenetorp, P., Lin, J., Ture, F.: What the DAAM: Interpreting stable diffusion using cross attention. In: ACL (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.771 + ], + "angle": 0, + "content": "79. Tao, M., Bao, B.K., Tang, H., Xu, C.: Galip: Generative adversarial clips for text-to-image synthesis. In: CVPR (2023) 8, 9, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "80. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers and distillation through attention. In: ICML (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "81. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: NeurIPS (2017) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "82. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "83. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: CVPR (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.231 + ], + "angle": 0, + "content": "84. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: CVPR (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "85. Zagoruyko, S., Komodakis, N.: Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In: ICLR (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.3 + ], + "angle": 0, + "content": "86. Zhang, L., Chen, X., Tu, X., Wan, P., Xu, N., Ma, K.: Wavelet knowledge distillation: Towards efficient image-to-image translation. In: CVPR (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "87. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.356 + ], + "angle": 0, + "content": "88. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: ICLR (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.384 + ], + "angle": 0, + "content": "89. Zhao, Y., Xu, Y., Xiao, Z., Hou, T.: Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567 (2023) 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.424 + ], + "angle": 0, + "content": "90. Zhou, Y., Zhang, R., Chen, C., Li, C., Tensmeyer, C., Yu, T., Gu, J., Xu, J., Sun, T.: Towards language-free training for text-to-image generation. In: CVPR (2022) 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.453 + ], + "angle": 0, + "content": "91. Zhu, L.: Thop: Pytorch-ocounter. https://github.com/Lyken17/pytorchOpCounter (2018) 2" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.453 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_origin.pdf b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..08d6fec06c654e9fa03dd4c151da51d46cb6ee45 --- /dev/null +++ b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/945ce569-fef9-4655-9b9a-e82e1c16cd44_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:438e57f9f956f6ea7374aa5a6414728edd9f321f78e20d0633f2cc2e1a45eb17 +size 4898709 diff --git a/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/full.md b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cb3573a8c0352fc3fcd7912e668867e8b034fefb --- /dev/null +++ b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/full.md @@ -0,0 +1,395 @@ +# BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion + +Bo-Kyeong Kim1, Hyoung-Kyu Song2, Thibault Castells1, and Shinkook Choi1 + +$^{1}$ Nota Inc. $^{2}$ Captions Research + +{bokyeong.kim, thibault, shinkook.choi}@nota.ai, kyu@captions.ai + +![](images/b6b9bcc8e0f2fcb2509332954f1cf0ba0cdc8d89478f5152869bb15fec8b3bb1.jpg) +(a) Efficient General-purpose T2I + +![](images/8a880b37aa421bf7c7039a28540b468c220868d370fd1272d98fdae10ecec360.jpg) + +![](images/faeb0988ef932fc94156e39fcdd65b7a628762c3e12c300ec4344a0908708d1e.jpg) +(b) Efficient Personalized T2I +(c) Efficient I2I (Different Styles) +Fig. 1: Our compressed model enables efficient (a) zero-shot text-to-image generation, (b) personalized synthesis, (c) image-to-image translation, and (d) mobile deployment. Samples from BK-SDM-Small with $36\%$ reduced parameters and latency are shown. + +![](images/b925a38ebba87d72e561973bbc2e826e809aca9bcfe9a5ed2ed5188e1d327b1f.jpg) + +![](images/8674b31599d7a1041f81bceee447e162397c6b1e78c4ce1cfd30ca333a9c0277.jpg) +(d) Mobile Deployment + +Abstract. Text-to-image (T2I) generation with Stable Diffusion models (SDMs) involves high computing demands due to billion-scale parameters. To enhance efficiency, recent studies have reduced sampling steps and applied network quantization while retaining the original architectures. The lack of architectural reduction attempts may stem from worries over expensive retraining for such massive models. In this work, we uncover the surprising potential of block pruning and feature distillation for low-cost general-purpose T2I. By removing several residual and attention blocks from the U-Net of SDMs, we achieve $30\% \sim 50\%$ reduction in model size, MACs, and latency. We show that distillation retraining is effective even under limited resources: using only 13 A100 days and a tiny dataset, our compact models can imitate the original SDMs (v1.4 and v2.1-base with over 6,000 A100 days). Benefiting from the transferred knowledge, our BK-SDMs deliver competitive results on zero-shot MS-COCO against larger multi-billion parameter models. We further demonstrate the applicability of our lightweight backbones in personalized generation and image-to-image translation. Deployment of our models on edge devices attains 4-second inference. Code and models can be found at: https://github.com/Nota-NetsPresso/BK-SDM. + +![](images/33d091d3cee1eeb8313e995c1c2da5619ece4832718611e4730fb952929b77d0.jpg) +Fig. 2: Computation of Stable Diffusion. The denoising U-Net is the main processing bottleneck. THOP [91] is used to measure MACs in generating a $512 \times 512$ image. + +![](images/dcff685fa1d1cfe7d2f65477e552672de268f7f4144efe386bcce74762c4bd0b.jpg) + +# 1 Introduction + +Stable Diffusion models (SDMs) [58,60-62] are one of the most renowned open-source models for text-to-image (T2I) synthesis, and their exceptional capability has begun to be leveraged as a backbone in several text-guided vision applications [2,3,83,87]. SDMs are T2I-specialized latent diffusion models (LDMs) [58], which employ diffusion operations [23,37,76] in a semantically compressed space for compute efficiency. Within a SDM, a U-Net [9,64] performs iterative sampling to progressively denoise a random latent code and is aided by a text encoder [53] and an image decoder [12,81] to produce text-aligned images. This inference process still involves excessive computational requirements (see Fig. 2), which often hinder the utilization of SDMs despite their rapidly growing usage. + +To alleviate this issue, numerous approaches toward efficient SDMs have been introduced. A pretrained diffusion model is distilled to reduce the number of denoising steps, enabling an identically architectured model with fewer sampling steps [40, 43]. Post-training quantization [24, 34, 74] and implementation optimization [6, 7] methods are also leveraged. However, the removal of architectural elements in large diffusion models remains less explored. + +This study unlocks the immense potential of classical architectural compression in attaining smaller and faster diffusion models. We eliminate multiple residual and attention blocks from the U-Net of a SDM and retrain it with feature-level knowledge distillation (KD) [18, 63] for general-purpose T2I. Under restricted training resources, our compact models can mimic the original SDM by leveraging transferred knowledge. Our work effectively reduces the computation of SDM-v1.4 [60] and SDM-v2.1-base [62] while achieving compelling zero-shot results on par with multi-billion parameter models [10, 11, 56]. Our contributions are summarized as follows: + +We compress SDMs by removing architectural blocks from the U-Net, achieving up to $51\%$ reduction in model size and $43\%$ improvement in latency on CPU and GPU. Previous pruning studies [13,47,84] focused on small models ( $< 100\mathrm{M}$ parameters) like ResNet50 and DeiT-B, not on foundation models like SDMs ( $>1,000\mathrm{M} = 1\mathrm{B}$ ), possibly due to the lack of economic retraining for such large models. Moreover, U-Net architectures are arguably more com + +plex due to the necessity of considering skip connections across the network, making the structural block removal inside them not straightforward. + +To the best of our knowledge, we first demonstrate the notable benefit of feature distillation for training diffusion models, which enables competitive T2I even with significantly fewer resources (using only 13 A100 days and 0.22M LAION pairs [70]). Considering the vast expense of training SDMs from scratch (surpassing 6,000 A100 days and 2,000M pairs), our study indicates that network compression is a remarkably cost-effective strategy in building compact general-purpose diffusion models. + +We show the practicality of our work across various aspects. Our lightweight backbones are readily applicable to customized generation [65] and image-to-image translation [44], effectively lowering finetuning and inference costs. T2I synthesis on Jetson AGX Orin and iPhone 14 using our models takes less than 4 seconds. + +We have publicly released our approach, model weights, and source code, motivating subsequent works by other researchers (e.g., block pruning and KD for the SDM-v1 variant [5,72] and SDXL [31, 73]). + +# 2 Related Work + +Large T2I diffusion models. By gradually removing noise from corrupted data, diffusion-based generative models [9, 22, 76] enable high-fidelity synthesis with broad mode coverage. Integrating these merits with the advancement of pretrained language models [8, 53, 54] has significantly improved the quality of T2I synthesis. In GLIDE [48] and Imagen [66], a text-conditional diffusion model generates a small image, which is upsampled via super-resolution modules. In DALL-E-2 [55], a text-conditional prior network produces an image embedding, which is transformed into an image via a diffusion decoder and further upscaled into higher resolutions. SDMs [58, 60–62] perform the diffusion modeling in a low-dimensional latent space constructed through a pixel-space autoencoder. We use SDMs as our baseline because of its open-access and gaining popularity over numerous downstream tasks [2, 3, 65, 83]. + +Efficient diffusion models. Several studies have addressed the slow sampling process. Diffusion-tailored distillation [42, 43, 68] progressively transfers knowledge from a pretrained diffusion model to a fewer-step model with the same architecture. Fast high-order solvers [38, 39, 88] for diffusion ordinary differential equations boost the sampling speed. Complementarily, our network compression approach reduces per-step computation and can be easily integrated with less sampling steps. Leveraging quantization [24, 34, 74] and implementation optimizations [6, 7] for SDMs can also be combined with our compact models for further efficiency. + +Distillation-based compression. KD enhances the performance of small-size models by exploiting output-level [21, 49] and feature-level [18, 63, 85] information of large source models. Although this classical KD has been actively used for efficient GANs [33, 57, 86], its power has not been explored for structurally compressed diffusion models. Distillation pretraining enables small yet capable + +general-purpose language models [26, 69, 77] and vision transformers [16, 80]. Beyond such models, we show that its success can be extended to diffusion models with iterative sampling. + +Concurrent studies. SnapFusion [35] and MobileDiffusion [89] achieve an efficient U-Net for SDMs through architecture optimization and step reduction. Würstchen [50] introduces two diffusion processes on low- and high-resolution latent spaces for economic training. These works are valuable but require much larger training resources than our work (see Tab. 1). In contrast, we utilize considerably fewer resources, leveraging the surprising benefits of classical KD for foundational diffusion models. While not demonstrated on SDMs, Diff-Pruning [14] proposes structured pruning based on Taylor expansion tailored for diffusion models. We emphasize the use of depth (block) pruning in our work, which often leads to greater enhancements in inference speeds compared to the width (channel) pruning approach of Diff-Pruning. Moreover, Tab. 5 employs a comparable Taylor pruning criterion, which results in an insufficient decrease in MACs. + +# 3 Compression Method + +We compress the U-Net [64] in SDMs, which is the most compute-heavy component (see Fig. 2). Conditioned on the text and time-step embeddings, the U-Net performs multiple denoising steps on latent representations. At each step, the U-Net produces the noise residual to compute the latent for the next step. We reduce this per-step computation, leading to Block-removed Knowledge-distilled SDMs (BK-SDMs). + +# 3.1 Compact U-Net Architecture + +The following architectures are obtained by compressing SDM-v1 (1.04B parameters), as shown in Fig. 3: + +$\circ$ BK-SDM-Base (0.76B) obtained with Sec. 3.1.(1). +$\circ$ BK-SDM-Small (0.66B) with Secs. 3.1.(1) and 3.1.(2). +$\circ$ BK-SDM-Tiny (0.50B) with Secs. 3.1.(1), 3.1.(2), and 3.1.(3). + +Our approach can be identically applied to SDM-v2 (1.26B parameters), leading to BK-SDM-v2-{Base (0.98B), Small (0.88B), Tiny (0.72B)}. + +(1) Fewer Blocks in the Down and Up Stages. This approach is closely aligned with DistilBERT [69] which halves the number of layers and initializes the compact model with the original weights by benefiting from the shared dimensionality. In the original U-Net, each stage with a common spatial size consists of multiple blocks, and most stages contain pairs of residual (R) [17] and cross-attention (A) [25,82] blocks. We hypothesize the existence of some unnecessary pairs and use the following removal strategies. + +![](images/803b8479453aadc16b24f01220daa9eca29fa7f525a80a1d4c9280434523e209.jpg) +Fig. 3: Block removal from the denoising U-Net. Our approach is applicable to all the SDM versions in v1 and v2, which share the same U-Net block configuration. For experiments, we used v1.4 [60] and v2.1-base [62]. See Sec. A for the details. + +![](images/045c11e1473319a1d92139396107748fe87d6c61f39c893953e1ad42ed9066bf.jpg) + +![](images/c8b74f10abcd00fdd81656ce0301e279fb1681c1b47af206d773c8eb77b02c7e.jpg) +Fig. 4: Minor impact of removing the mid-stage from the U-Net. Results without retraining. See Sec. B for additional results. + +![](images/55b1a0e44a56537e20c067b3453203869a67a4c7124e7cc205fb5d42039cc9aa.jpg) + +![](images/64d1c5df0da70129ad90c294ff86c741ea16ab828b27d5496039d5bcc608c790.jpg) +Fig.5: Importance of (a) each block and (b) each group of paired/triplet blocks. Higher score implies removable blocks. The results are aligned with our architectures (e.g., removal of innermost stages and the second R-A pairs in down stages). See Sec. C for further analysis. + +- Down Stages. We maintain the first R-A pairs while eliminating the second pairs, because the first pairs process the changed spatial information and would be more important than the second pairs. This design is consistent with the sensitivity analysis that measures the block-level significance (see Fig. 5). Our approach also does not harm the dimensionality of the original U-Net, enabling the use of the corresponding pretrained weights for initialization [69]. +- Up Stages. While adhering to the aforementioned scheme, we retain the third R-A pairs. This allows us to utilize the output feature maps at the end of each down stage and the corresponding skip connections between the down and up stages. The same process is applied to the innermost down and up stages that contain only R blocks. + +(2) Removal of the Entire Mid-Stage. Surprisingly, removing the entire mid-stage from the original U-Net does not noticeably degrade the generation quality while effectively reducing the parameters by $11\%$ (see Fig. 4). This observation is consistent with the minor role of inner layers in the U-Net generator of GANs [28]. + +![](images/be833c8fd0127a7f45a4ec7885c92d3cfd2540f798b19ff4799b9fc3003f4948.jpg) +Fig. 6: Distillation-based retraining. The block-removed U-Net is trained effectively through the guidance of the original U-Net. + +Integrating the mid-stage removal with fewer blocks in Sec. 3.1 further decreases compute burdens (Tab. 2) at the cost of a slight decline in performance (Tab. 1). Therefore, we offer this mid-stage elimination as an option, depending on the priority between compute efficiency (using BK-SDM-Small) and generation quality (BK-SDM-Base). + +(3) Further Removal of the Innermost Stages. For additional compression, the innermost down and up stages can also be pruned, leading to our lightest model BK-SDM-Tiny. This implies that outer stages with larger spatial dimensions and their skip connections play a crucial role in the U-Net for T2I synthesis. + +Alignment with Pruning Sensitivity Analysis. To support the properness of our architectures, we measure the importance of each block (see Fig. 5) and show that unimportant blocks match with our design choices. The importance is measured by how generation scores vary when removing each residual or attention block from the U-Net. A significant drop in performance highlights the essential role of that block. Note that some blocks are not directly removable due to different channel dimensions between input and output; we replace such blocks with channel interpolation modules (denoted by “*” in Fig. 5) to mimic the removal while retaining the information. + +The sensitivity analysis implies that the innermost down-mid-up stages and the second R-A pairs in the down stages play relatively minor roles. Pruning these blocks aligns with our architectures, designed based on human knowledge (e.g., prioritizing blocks with altered channel dimensions) and previous studies [28, 69]. + +Though some results aligned, the importance criterion from the CLIP Score pruning sensitivity (in Fig. 5) leads to excessive pruning of attention blocks, eventually yielding inferior performance compared to our approach (see Table 5). + +# 3.2 Distillation-based Retraining + +For general-purpose T2I, we train our block-removed U-Net to mimic the behavior of the original U-Net (see Fig. 6). To obtain the input of U-Net, we use pretrained-and-frozen encoders [58] for images and text prompts. + +Given the latent representation $\mathbf{z}$ of an image and its paired text embedding $\mathbf{y}$ , the task loss for the reverse denoising process [22, 58] is computed as: + +$$ +\mathcal {L} _ {\text {T a s k}} = \mathbb {E} _ {\mathbf {z}, \epsilon , \mathbf {y}, t} \left[ | | \epsilon - \epsilon_ {\mathrm {S}} (\mathbf {z} _ {t}, \mathbf {y}, t) | | _ {2} ^ {2} \right], \tag {1} +$$ + +where $\mathbf{z}_t$ is a noisy latent code from the diffusion process [22] with the sampled noise $\epsilon \sim N(\mathbf{0},\mathbf{I})$ and time step $t \sim \mathrm{Uniform}(1,T)$ , and $\epsilon_{\mathrm{S}}(\circ)$ indicates the estimated noise from our compact U-Net student. For brevity, we omit the subscripts of $\mathbb{E}_{\mathbf{z},\epsilon,\mathbf{y},t}[\circ]$ in the following notations. + +The compact student is also trained to imitate the outputs of the original U-Net teacher, $\epsilon_{\mathrm{T}}(\circ)$ , with the following output-level KD objective [21]: + +$$ +\mathcal {L} _ {\text {O u t K D}} = \mathbb {E} \left[ \left| \left| \epsilon_ {\mathrm {T}} (\mathbf {z} _ {t}, \mathbf {y}, t) - \epsilon_ {\mathrm {S}} (\mathbf {z} _ {t}, \mathbf {y}, t) \right| \right| _ {2} ^ {2} \right]. \tag {2} +$$ + +A key to our approach is feature-level KD [18, 63] that provides abundant guidance for the student's training: + +$$ +\mathcal {L} _ {\text {F e a t K D}} = \mathbb {E} \left[ \sum_ {l} | | f _ {\mathrm {T}} ^ {l} (\mathbf {z} _ {t}, \mathbf {y}, t) - f _ {\mathrm {S}} ^ {l} (\mathbf {z} _ {t}, \mathbf {y}, t) | | _ {2} ^ {2} \right], \tag {3} +$$ + +where $f_{\mathrm{T}}^{l}(\circ)$ and $f_{\mathrm{S}}^{l}(\circ)$ represent the feature maps of the $l$ -th layer in a predefined set of distilled layers from the teacher and the student, respectively. While learnable regressors (e.g., $1 \times 1$ convolutions to match the number of channels) have been commonly used [57,63,75], our approach circumvents this requirement. By applying distillation at the end of each stage in both models, we ensure that the dimensionality of the feature maps already matches, thus eliminating the need for additional regressors. + +The final objective is shown below, and we simply set $\lambda_{\mathrm{OutKD}}$ and $\lambda_{\mathrm{FeatKD}}$ as 1. Without loss-weight tuning, our approach is effective in empirical validation. + +$$ +\mathcal {L} = \mathcal {L} _ {\text {T a s k}} + \lambda_ {\text {O u t K D}} \mathcal {L} _ {\text {O u t K D}} + \lambda_ {\text {F e a t K D}} \mathcal {L} _ {\text {F e a t K D}}. \tag {4} +$$ + +# 4 Experimental Setup + +Distillation Retraining. We primarily use 0.22M image-text pairs from LAION-Aesthetics V2 (L-Aes) $6.5 + [70,71]$ , which are significantly fewer than the original training data used for SDMs [60,62] ( $>2,000\mathrm{M}$ pairs). In Fig. 11, dataset sizes smaller than 0.22M are randomly sampled from L-Aes $6.5+$ , while those larger than 0.22M are from L-Aes $6.25+$ . + +Zero-shot T2I Evaluation. Following the popular protocol [56, 58, 66], we use 30K prompts from the MS-COCO validation split [36], downsample the $512 \times 512$ generated images to $256 \times 256$ , and compare them with the entire validation set. We compute Fréchet Inception Distance (FID) [20] and Inception Score (IS) [67] to assess visual quality. We measure CLIP score [19, 53] with CLIP-ViT-g/14 model to assess text-image correspondence. + +Downstream Tasks. For personalized generation, we use the DreamBooth dataset [65] (30 subjects $\times$ 25 prompts $\times$ 4~6 images) and perform per-subject finetuning. Following the evaluation protocol [65], we use ViT-S/16 model [4] + +Table 1: Results on zero-shot MS-COCO $256 \times 256$ 30K. Training resources include image-text pairs, batch size, iterations, and A100 days. Despite far smaller resources, our compact models outperform prior studies [10, 11, 56, 79, 90], showing the benefit of compressing existing powerful models. Note that FID fluctuates more than the other metrics over training progress in our experiments (see Figs. 9 and 11). + +
Model +NameType #Param‡Generation ScoreTraining Resource
FID↓IS↑CLIP↑Data Size (Batch, # Iter)A100 Days
SDM-v1.4 [58,60]†DF1.04B13.0536.760.2958>2000M*(2048, 1171K)6250
Small Stable Diffusion [51]†DF0.76B12.7632.330.2851229M(128, 1100K)-
BK-SDM-Base [Ours]†DF0.76B15.7633.790.28780.22M(256, 50K)13
BK-SDM-Small [Ours]†DF0.66B16.9831.680.26770.22M(256, 50K)13
BK-SDM-Tiny [Ours]†DF0.50B17.1230.090.26530.22M(256, 50K)13
SDM-v2.1-base [58,62]†DF1.26B13.9335.930.3075>2000M*(2048, 1620K)8334
BK-SDM-v2-Base [Ours]†DF0.98B15.8531.700.28680.22M(128, 50K)4
BK-SDM-v2-Small [Ours]†DF0.88B16.6131.730.29010.22M(128, 50K)4
BK-SDM-v2-Tiny [Ours]†DF0.72B15.6831.640.28970.22M(128, 50K)4
DALL-E [56]AR12B27.517.9-250M(1024, 430K)-
CogView [10]AR4B27.118.2-30M(6144, 144K)-
CogView2 [11]AR6B2422.4-30M(4096, 300K)-
Make-A-Scene [15]AR4B11.84--35M(1024, 170K)-
LAFITE [90]GAN0.23B26.9426.02-3M--
GALIP (CC12M) [79]†GAN0.32B13.8625.160.281712M--
GigaGAN [27]GAN1.1B9.09-->100M*(512, 1350K)4783
GLIDE [48]DF3.5B12.24--250M(2048, 2500K)-
LDM-KL-8-G [58]DF1.45B12.6330.29-400M(680, 390K)-
DALL-E-2 [55]DF5.2B10.39--250M(4096, 3400K)-
SnapFusion [35]DF0.99B~13.6-~0.295>100M*(2048, -)>128*
Würstchen-v2 [50]†DF3.1B22.4032.870.26761700M(1536, 1725K)1484
MobileDiffusion [89]†DF0.52B9.01--150M(2048, 1000K)-
+ +$\dagger$ : Evaluated with the released checkpoints. $\ddagger$ : Total parameters for T2I synthesis. $*$ : Estimated based on public information. DF and AR: diffusion and autoregressive models. $\downarrow$ and $\uparrow$ : lower and higher values are better. + +for DINO score and CLIP-ViT-g/14 model for CLIP-I and CLIP-T scores. For image-to-image translation, input images are sourced from Meng et al. [43]. + +Implementation Details. We adjust the codes in Diffusers [52] and PEFT [41]. We use a single NVIDIA A100 80G GPU for main retraining and a single NVIDIA GeForce RTX 3090 GPU for per-subject finetuning. For compute efficiency, we always opt for 25 denoising steps of the U-Net at the inference phase, unless specified. The classifier-free guidance scale [23,66] is set to the default value of 7.5. The latent resolution is set to the default ( $H = W = 64$ in Fig. 3), yielding $512 \times 512$ images. See Sec. J for the details. + +# 5 Results + +All the results in Secs. 5.1-5.6 were obtained with the full benchmark protocol (MS-COCO $256 \times 256$ 30K samples), except for Fig. 10 ( $512 \times 512$ 5K samples). Unless specified, the training setup in Tab. 1 was used. + +Fig. 7: Visual comparison with open-sourced models. The results [11,48,50,79,90] were obtained with their official codes. +Table 2: Impact of compute reduction in U-Net on the entire SDM. The number of sampling steps is indicated with the parentheses, e.g., U-Net (1) for one step. The full computation (denoted by "Whole") covers the text encoder, U-Net, and image decoder. All corresponding values are obtained on the generation of a single $512 \times 512$ image with 25 denoising steps. The latency was measured on Xeon Silver 4210R CPU 2.40GHz and NVIDIA GeForce RTX 3090 GPU. +![](images/cfb355cb7ac4f643ced756f39c9166f92ab7017d45c90a21a398387b8461f6f9.jpg) +Prompt: A bowl that has vegetables inside of it; A brown and white cat staring off with pretty green eyes.; A toy raccoon standing on a pile of broccoli. + +
Model# ParamMACsCPU LatencyGPU Latency
U-NetWholeU-Net (1)U-Net (25)WholeU-Net (1)U-Net (25)WholeU-Net (1)U-Net (25)Whole
SDM-v1.4 [60]860M1033M339G8469G9716G5.63s146.28s153.00s0.049s1.28s1.41s
BK-SDM-Base [Ours]580M752M224G5594G6841G3.84s99.95s106.67s0.032s0.83s0.96s
(-32.6%)(-27.1%)(-33.9%)(-33.9%)(-29.5%)(-31.8%)(-31.7%)(-30.3%)(-34.6%)(-35.2%)(-31.9%)
BK-SDM-Small [Ours]483M655M218G5444G6690G3.45s89.78s96.50s0.030s0.77s0.90s
(-43.9%)(-36.5%)(-35.7%)(-35.7%)(-31.1%)(-38.7%)(-38.6%)(-36.9%)(-38.7%)(-39.8%)(-36.1%)
BK-SDM-Tiny [Ours]324M496M206G5126G6373G3.03s78.77s85.49s0.026s0.67s0.80s
(-62.4%)(-51.9%)(-39.5%)(-39.5%)(-34.4%)(-46.2%)(-46.1%)(-44.1%)(-46.9%)(-47.7%)(-43.2%)
+ +# 5.1 Comparison with Existing Works + +Quantitative Comparison. Tab. 1 shows the zero-shot results for general-purpose T2I. Despite being trained with only 0.22M samples and having fewer than 1B parameters, our compressed models demonstrate competitive performance on par with existing large models. + +Visual Comparison. Fig. 7 depicts synthesized images with some MS-COCO captions. Our compact models inherit the superiority of SDM and produce more photorealistic images compared to the AR-based [11] and GAN-based [79, 90] baselines. Noticeably, the same latent code results in a shared visual style between the original and our models (6th–9th columns in Fig. 7), similar to the observation in transfer learning for GANs [45]. See Sec. D for additional results. + +# 5.2 Computational Gain + +Tab. 2 shows how the compute reduction for each sampling step of the U-Net affects the overall process. The per-step reduction effectively decreases MACs + +Table 3: Quality gains from distillation retraining. MS-COCO 30K. + +
BK-SDM +TypeKDGenerationScore# Param +U-Net
FID↓IS↑CLIP↑
BaseX23.5723.020.24080.58B
✓15.7633.790.2878
v2-BaseX16.7625.880.26610.59B
✓15.8531.700.2868
v2-SmallX16.7125.770.26550.49B
✓16.6131.730.2901
v2-TinyX16.8726.060.26780.33B
✓15.6831.640.2897
+ +Table 4: Significance of each element in transferred knowledge. Results of BK-SDMv2-Small on MS-COCO 30K. + +
Weight InitializationOutput FeatureGeneration Score
KDKDFID↓IS↑CLIP↑
RandomXX41.7515.420.1733
Random✓X29.0118.290.1967
Random✓✓24.4722.370.2323
TeacherXX16.7125.770.2655
Teacher✓X14.2729.470.2777
Teacher✓✓16.6131.730.2901
+ +![](images/a0558c4dd4c4092b6bb06418cb3166a283d39ead661c10f564a4f82e1dce13ae.jpg) +Fig. 8: Image areas affected by each word. KD enables our models to mimic the SDM, yielding similar per-word attribution maps. The model without KD behaves differently, causing dissimilar maps and inaccurate generation (e.g., two sheep and unusual bird shapes). + +and inference time by more than $30\%$ . Notably, BK-SDM-Tiny has $50\%$ fewer parameters than the original SDM. + +# 5.3 Benefit of Distillation Retraining + +T2I Performance. Tab. 3 summarizes the results from ablating the total KD objective (Eq. 2+Eq. 3). Across various model types, distillation brings a clear improvement in generation quality. Tab. 4 analyzes the effect of each element in transferred knowledge. Exploiting output-level KD (Eq. 2) boosts the performance compared to using only the denoising task loss. Leveraging feature-level KD (Eq. 3) yields further score enhancement. Additionally, using the teacher weights for initialization is highly beneficial. + +Cross-attention Resemblance. Fig. 8 displays the per-word attribution maps [78] created by aggregating cross-attention scores over spatiotemporal dimensions. The attribution maps of our models are semantically and spatially similar to those of the original model, indicating the merit of supervisory signals at multiple stages via KD. In contrast, the baseline model without KD activates incorrect areas, leading to text-mismatched generation results. + +![](images/f6d779e2f565da131a56c95ae295d76bf605647dc0c3e59797a2ba1158d0d112.jpg) +Fig.9: Zero-shot results over training progress. The architecture size of BK-SDM, usage of KD, and batch size are denoted. Results on MS-COCO 30K. + +![](images/6bd5100732c37d9bde039be5b4df979ab4b09b2e1904ac0adf1e7ac7829ba24b.jpg) + +![](images/8da766e3190a210889185f8bf4d6cc18f628b90837a238fef02cfdd3da1ee28b.jpg) +Fig.10:Trade-off curves Left:FID vs.CLIP score Right: quality vs.efficiency Ours-Base on MS-COCO 5K. + +![](images/54b695e4dd62078adc4fd40badb1b3a3ff07671d80e92840633383e7c4ad1566.jpg) + +Training Progress. Fig. 9 shows the results over training iterations. Without KD, training solely with the denoising task loss causes fluctuations or sudden drops in performance (indicated with green and cyan). In contrast, distillation (purple and pink) stabilizes and accelerates the training process, demonstrating the benefit of providing sufficient hints for training guidance. Notably, our small-and tiny-size models trained with KD (yellow and red) outperform the bigger base-size model without KD (cyan). Additionally, while the best FID score is observed early on, IS and CLIP score exhibit ongoing improvement, implying that judging models merely with FID may be suboptimal. + +Trade-off Results. Fig. 10 shows the results of BK-SDM-Base with and without KD on MS-COCO $512 \times 512$ 5K. Higher classifier-free guidance scales [23,66] lead to better text-aligned images at the cost of less diversity. More denoising steps improve generation quality at the cost of slower inference. Distillation retraining achieves much better trade-off curves than the baseline without KD. + +# 5.4 Comparison with Different Pruning Criteria + +We extend our baseline comparisons by calculating several block-level importance scores [29].1 Furthermore, the impact of applying KD at the end of each stage is also examined. Tab. 5 shows the results. Magnitude (Mag) pruning [32] removes critical outer blocks, causing irrecoverable loss. Taylor pruning [14,30,46] removes only the inner part, causing ineffective reduction in MACs and speed; additionally, this method demands the intensive calculation of backward gradients. The CLIP Score method (based on Fig. 5) results in severe pruning of + +Table 5: Different block-level criteria at similar parameters. Taylor pruning removes solely the inner blocks, leading to insufficient reduction in MACs. Our method attains a favorable compromise between performance and inference speed. Results on MS-COCO 30K. + +
Pruning +CriterionFID↓CLIP Score↑U-Net
KD×KD√KD×KD√MACs# Param
Mag [32]>150>150<0.1<0.1122.2G495.5M
Taylor [14,30]18.8516.730.24290.2749305.4G493.4M
CLIP [Fig. 5]32.9321.860.19360.2283204.8G496.8M
Ours-Small22.9716.830.22870.2668217.7G482.3M
+ +Retraining with batch 128, 0.22M data, and 50K iters. Original SD-v1.4's U-Net: (MACs, Params) = (338.7G, 859.5M). + +Table 6: A/B-testing preference study. Total 2,501 comparisons. + +
Ours-SmallGLIDE-filt [48]GALIP [79]Würstchen-v2 SDM-v1.4 [50][60]
# Win†453 (71.5%)379 (61.5%)319 (51.4%)202 (32.1%)
# Defeat†181 (28.5%)237 (38.5%)302 (48.6%)428 (67.9%)
+ +Participants selected preferred images from pairs with the same prompts. One image in each pair was always from BK-SDM-Small, while the other was randomly chosen from the four methods. + +Table 7: Impact of training batch size and iterations. Results on MS-COCO 30K. + +
BK-SDMBaseSmallTiny
Batch Size642566425664256
FID↓14.6115.7616.8716.9817.2817.12
IS↑31.4433.7929.5131.6828.3330.09
CLIP↑0.28260.28780.26440.26770.26070.2653
BK-SDMBase (Data 2.3M)Small (Data 2.3M)Tiny (Data 2.3M)
#Iter50K100K50K100K50K100K
FID↓14.8115.3917.0517.0117.5317.63
IS↑34.1734.7633.1033.1431.3232.26
CLIP↑0.28830.28890.27340.27540.26900.2713
+ +![](images/b57e17a2fb09f53b6f478788d756b0a1b43a0ec56f8217cedd37595525806530.jpg) +Fig. 11: Impact of training data volume. Results of BK-SDM-Small on MS-COCO 30K. + +![](images/9cb8d13181c912f273bc88dddfc2ffe89b622298ff8d4564dc05a8142a604557.jpg) + +attention blocks, adversely affecting performance. Our approach achieves superior or comparable results against the benchmark methods and can be directly applied to all the SDM versions in v1 and v2. This advantage comes without the extra phase of calculating pruning criteria while ensuring a favorable balance between performance and inference speed. Notably, KD always remains effective. + +# 5.5 Human Preference Assessment + +Despite their extensive use, automated metrics for assessing image quality (e.g., FID) do not consistently match with human evaluation [55, 66]. To further address this, we have conducted an A/B testing-based study: participants were shown pairs of images generated using the same MS-COCO prompt and asked to choose their preferred image in relation to the prompt. One image in each pair was always from our model, while the other was randomly selected from one of four comparative methods. A total of 25 participants conducted 2,501 comparisons online. The participants had no bias towards the examined models and were not compensated. The prompts were randomly drawn from the MS-COCO set and varied for 3 to 5 participants. As shown in Tab. 6, our model (0.66B parameters) excels over GLIDE-filtered, GALIP, and Wüstchen-v2 (3.1B). + +Table 8: Personalized generation with finetuning over different backbones. Our compact models can preserve subject fidelity (DINO and CLIP-I) and prompt fidelity (CLIP-T) of the original SDM with reduced finetuning (FT) costs and fewer parameters. + +
Backbone {Param}DINO↑CLIP-I↑CLIP-T↑FT (Time, Mem)†
SDM-v1.4 [60] {1.04B}0.7280.7250.263(882s, 23.0GB)
BK-SDM-Base {0.76B}0.7230.7170.260(623s, 18.7GB)
BK-SDM-Small {0.66B}0.7200.7050.259(604s, 17.2GB)
BK-SDM-Tiny {0.50B}0.7150.6930.261(560s, 13.1GB)
Base (Batch 64) {0.76B}0.7180.7080.262(623s, 18.7GB)
- No KD & Random Init.0.5940.4650.191(623s, 18.7GB)
- No KD & Teacher Init.0.7160.6690.258(623s, 18.7GB)
+ +†: Per-subject FT time and GPU memory for 800 iterations on RTX 3090. + +Table 9: Speedups on edge devices. + +
ModelAGX Orin 32GBiPhone 14
SDM-v14.9s [61]5.6s [60]
BK-SDM-Base3.4s (-31%)4.0s (-29%)
BK-SDM-Small3.2s (-35%)3.9s (-29%)
BK-SDM-Tiny2.8s (-43%)3.9s (-29%)
+ +![](images/64329bbe172f3cba1d82952c1f0148ea45368e99bc02947d14dbbdec5319fe22.jpg) +Fig. 12: Visual results of personalized generation. Each subject is marked as "a [identifier] [class noun]" (e.g., "a [V] dog"). + +![](images/37a3ea26f54b4cc2ad8d07c1029c0ce3d40a053089701804f24f11a73c238dd1.jpg) + +![](images/a9a460561ad1244a1c71559b8eb11c8333a389e403460129ff3bb113e3f27d36.jpg) +(left-to-right, top-to-bottom) Target style: Art station; Oil painting; Photograph; Watercolor +Fig. 13: Text-guided image-to-image translation. + +![](images/46e70f60fbb2ae51979cf68452df557d01bb97b385c525b482320cc4d2c315df.jpg) +Fig. 14: Small LDM for face synthesis. + +![](images/ec1f16501c99b509c4228d59a5421bd5e4f0df0fe8459e0e144a5c81ec676e4f.jpg) + +# 5.6 Impact of Training Resources on Performance + +Consistent with existing evidence, the use of larger batch sizes, more extensive data, and longer iterations for training enhances performance in our work (see Tab. 7, Fig. 11, and Sec. H). However, this benefit requires increased resource demands (e.g., extended training days without multiple high-spec GPUs and greater data storage capacity). As such, despite the better performing models in Tab. 7, we primarily report the models with fewer resources. We believe that + +accessible training costs by many researchers can help drive advancement in massive models. + +# 5.7 Application + +Personalized T2I Synthesis. Tab. 8 compares the fine-tuning results with DreamBooth [65] over different backbones to create images about a given subject. BK-SDMs can preserve $95\% \sim 99\%$ scores of the original SDM while cutting fine-tuning costs. Fig. 12 depicts that our models can accurately capture the subject details and generate various scenes. Over the models retrained with a batch size of 64, the baselines without KD fail to generate the subjects or cannot maintain the identity details. See Sec. E for further results. + +Image-to-Image Translation. Fig. 13 presents the text-guided stylization results with SDEdit [44]. Our model, resembling the ability of the original SDM, faithfully produces images given style-specified prompts and content pictures. See Sec. F for additional results. + +Deployment on Edge Devices. We deploy our models trained with 2.3M pairs and compare them against the original SDM under the same setup on edge devices (20 denoising steps on NVIDIA Jetson AGX Orin 32GB and 10 steps on iPhone 14). Our models produce a $512 \times 512$ image within 4 seconds (see Tab. 9), while maintaining acceptable image quality (Fig. 1(d) and Sec. G). + +Another LDM. SDMs are derived from LDMs [58], which share a similar U-Net design across many tasks. To validate the generality of our work, we compress an LDM (with 308M parameters and 410K training iterations) for unconditional generation on CelebA-HQ [59] by using the same approach of BK-SDM-Small (187M parameters and 30K iterations). Fig. 14 shows the efficacy of our architecture and distillation retraining. + +# 6 Conclusion and Discussion + +We uncover the potential of architectural compression for general-purpose text-to-image synthesis with a renowned model, Stable Diffusion. Our block-removed lightweight models are effective for zero-shot generation, achieving competitive results against large-scale baselines. Distillation is a key of our method, leading to powerful retraining even under very constrained resources. Our work is orthogonal to previous works for efficient diffusion models, e.g., enabling fewer sampling steps, and can be readily combined with them. We hope our study can facilitate future research on structural compression of large diffusion models. + +Limitations. Our compact models inherit the capability of the source model for high-fidelity image generation, but they have shortcomings such as inaccurate generation of full-body human appearance. + +Negative Social Impacts. Recent large generative models are capable of creating high-quality plausible content. To avoid causing unintended social usage, researchers should take steps to ensure the appropriateness of the training data. + +# Acknowledgments + +We thank the Microsoft Startups Founders Hub program and the AI Industrial Convergence Cluster Development project funded by the Ministry of Science and ICT (MSIT, Korea) and Gwangju Metropolitan City for their generous support of GPU resources. + +# References + +1. A Simple and Effective Pruning Approach for LLMs. ICLR (2024) 11 +2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) 2, 3 +3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023) 2, 3 +4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) 7 +5. Castells, T., Song, H.K., Piao, T., Choi, S., Kim, B.K., Yim, H., Lee, C., Kim, J.G., Kim, T.H.: Edgefusion: On-device text-to-image generation. In: CVPR Workshop (2024) 3 +6. Chen, Y.H., Sarokin, R., Lee, J., Tang, J., Chang, C.L., Kulik, A., Grundmann, M.: Speed is all you need: On-device acceleration of large diffusion models viagpu-aware optimizations. In: CVPR Workshop (2023) 2, 3 +7. Choi, J., Kim, M., Ahn, D., Kim, T., Kim, Y., Jo, D., Jeon, H., Kim, J.J., Kim, H.: Squeezing large-scale diffusion models for mobile. In: ICML Workshop (2023) 2, 3 +8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 3 +9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021) 2, 3 +10. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. In: NeurIPS (2021) 2, 8 +11. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. In: NeurIPS (2022) 2, 8, 9 +12. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) 2 +13. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: CVPR (2023) 2 +14. Fang, G., Ma, X., Wang, X.: Structural pruning for diffusion models. In: NeurIPS (2023) 4, 11, 12 +15. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV (2022) 8 +16. Hao, Z., Guo, J., Jia, D., Han, K., Tang, Y., Zhang, C., Hu, H., Wang, Y.: Learning efficient vision transformers via fine-grained manifold distillation. In: NeurIPS (2022) 4 +17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 4 + +18. Heo, B., Kim, J., Yun, S., Park, H., Kwak, N., Choi, J.Y.: A comprehensive overhaul of feature distillation. In: ICCV (2019) 2, 3, 7 +19. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: EMNLP (2021) 7 +20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 7 +21. Hinton, G., Vinyls, O., Dean, J.: Distilling the knowledge in a neural network. In: NeurIPS Workshop (2014) 3, 7 +22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) 3, 6, 7 +23. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop (2021) 2, 8, 11 +24. Hou, J., Asghar, Z.: World's first on-device demonstration of stable diffusion on an android phone. https://wwwQUALCOMM.com/news (2023) 2, 3 +25. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: ICML (2021) 4 +26. Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., Wang, F., Liu, Q.: Tinybert: Distilling bert for natural language understanding. In: Findings of EMNLP (2020) 4 +27. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) 8 +28. Kim, B.K., Choi, S., Park, H.: Cut inner layers: A structured pruning strategy for efficient u-net gans. In: ICML Workshop (2022) 5, 6 +29. Kim, B.K., Kim, G., Kim, T.H., Castells, T., Choi, S., Shin, J., Song, H.K.: Shortened llama: A simple depth pruning for large language models. arXiv preprint arXiv:2402.02834 (2024) 11 +30. LeCun, Y., Denker, J., Solla, S.: Optimal brain damage. In: NeurIPS (1989) 11, 12 +31. Lee, Y., Park, K., Cho, Y., Lee, Y.J., Hwang, S.J.: Koala: Self-attention matters in knowledge distillation of latent diffusion models for memory-efficient and fast image synthesis. arXiv preprint arXiv:2312.04005v1 (2023) 3 +32. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. In: ICLR (2017) 11, 12 +33. Li, M., Lin, J., Ding, Y., Liu, Z., Zhu, J.Y., Han, S.: Gan compression: Efficient architectures for interactive conditional gans. In: CVPR (2020) 3 +34. Li, X., Liu, Y., Lian, L., Yang, H., Dong, Z., Kang, D., Zhang, S., Keutzer, K.: Q-diffusion: Quantizing diffusion models. In: ICCV (2023) 2, 3 +35. Li, Y., Wang, H., Jin, Q., Hu, J., Chemerys, P., Fu, Y., Wang, Y., Tulyakov, S., Ren, J.: Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. In: NeurIPS (2023) 4, 8 +36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dolkar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) 7 +37. Liu, L., Ren, Y., Lin, Z., Zhao, Z.: Pseudo numerical methods for diffusion models on manifolds. In: ICLR (2022) 2 +38. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In: NeurIPS (2022) 3 +39. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095 (2022) 3 + +40. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023) 2 +41. Mangrulkar, S., Gugger, S., Debut, L., Belkada, Y., Paul, S.: Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft (2022) 8 +42. Meng, C., Gao, R., Kingma, D.P., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: NeurIPS Workshop (2022) 3 +43. Meng, C., Gao, R., Kingma, D.P., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR (2023) 2, 3, 8 +44. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022) 3, 14 +45. Mo, S., Cho, M., Shin, J.: Freeze the discriminator: a simple baseline for fine-tuning gans. In: CVPR Workshop (2020) 9 +46. Molchanov, P., Mallya, A., Tyree, S., Frosio, I., Kautz, J.: Importance estimation for neural network pruning. In: CVPR (2019) 11 +47. Murti, C., Narshana, T., Bhattacharyya, C.: TVSPrune - pruning non-discriminative filters via total variation separability of intermediate representations without fine tuning. In: ICLR (2023) 2 +48. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: ICML (2022) 3, 8, 9, 12 +49. Park, W., Kim, D., Lu, Y., Cho, M.: Relational knowledge distillation. In: CVPR (2019) 3 +50. Pernias, P., Rampas, D., Richter, M.L., Pal, C.J., Aubreville, M.: Wierstchen: An efficient architecture for large-scale text-to-image diffusion models. In: ICLR (2024) 4, 8, 9, 12 +51. Pinkney, J.: Small stable diffusion. https://huggingface.co/OFA-Sys/small-stable-diffusion-v0 (2023) 8 +52. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022) 8 +53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 3, 7 +54. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019) 3 +55. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) 3, 8, 12 +56. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) 2, 7, 8 +57. Ren, Y., Wu, J., Xiao, X., Yang, J.: Online multi-granularity distillation for gan compression. In: ICCV (2021) 3, 7 +58. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 3, 6, 7, 8, 14 +59. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: Ldm on celeba-hq. https://huggingface.co/CompVis/ldm-celebahq-256 (2022) 14 +60. Rombach, R., Esser, P.: Stable diffusion v1-4. https://huggingface.co/CompVis/stable-diffusion-v1-4 (2022) 2, 3, 5, 7, 8, 9, 12, 13 + +61. Rombach, R., Esser, P.: Stable diffusion v1-5. https://huggingface.co/runwayml/stable-diffusion-v1-5 (2022) 2, 3, 13 +62. Rombach, R., Esser, P., Ha, D.: Stable diffusion v2-1-base. https://huggingface.co/stabilityai/stable-diffusion-2-1-base (2022) 2, 3, 5, 7, 8 +63. Romero, A., Ballas, N., Kahou, S.E., Chassang, A., Gatta, C., Bengio, Y.: Fitnets: Hints for thin deep nets. In: ICLR (2015) 2, 3, 7 +64. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) 2, 4 +65. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023) 3, 7, 14 +66. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 3, 7, 8, 11, 12 +67. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: NeurIPS (2016) 7 +68. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2022) 3 +69. Sanh, V., Debut, L., Chaumont, J., Wolf, T.: Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In: NeurIPS Workshop (2019) 4, 5, 6 +70. Schuhmann, C., Beaumont, R.: Laion-aesthetics. https://laion.ai/blog/laion-aesthetics (2022) 3, 7 +71. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS Workshop (2022) 7 +72. Segmind: Segmind-distill-sd. https://github.com/segmind/distill-sd/tree/c1e97a70d141df09e6fe5cc7dbd66e0cbae3eeb (2023) 3 +73. Segmind: Ssd-1b. https://github.com/segmind/SSD-1B/tree/d2ff723ea8ecf5ddb86f3aac0af1db30e88a2e2d (2023) 3 +74. Shen, H., Cheng, P., Ye, X., Cheng, W., Abidi, H.: Accelerate stable diffusion with intel neural compressor. https://medium.com/intel-analytics-software (2022) 2, 3 +75. Shu, C., Liu, Y., Gao, J., Yan, Z., Shen, C.: Channel-wise knowledge distillation for dense prediction. In: ICCV (2021) 7 +76. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) 2, 3 +77. Sun, Z., Yu, H., Song, X., Liu, R., Yang, Y., Zhou, D.: Mobilebert: a compact task-agnostic bert for resource-limited devices. In: ACL (2020) 4 +78. Tang, R., Liu, L., Pandey, A., Jiang, Z., Yang, G., Kumar, K., Stenetorp, P., Lin, J., Ture, F.: What the DAAM: Interpreting stable diffusion using cross attention. In: ACL (2023) 10 +79. Tao, M., Bao, B.K., Tang, H., Xu, C.: Galip: Generative adversarial clips for text-to-image synthesis. In: CVPR (2023) 8, 9, 12 +80. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers and distillation through attention. In: ICML (2021) 4 +81. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: NeurIPS (2017) 2 + +82. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 4 +83. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: CVPR (2023) 2, 3 +84. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: CVPR (2023) 2 +85. Zagoruyko, S., Komodakis, N.: Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In: ICLR (2017) 3 +86. Zhang, L., Chen, X., Tu, X., Wan, P., Xu, N., Ma, K.: Wavelet knowledge distillation: Towards efficient image-to-image translation. In: CVPR (2022) 3 +87. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) 2 +88. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: ICLR (2023) 3 +89. Zhao, Y., Xu, Y., Xiao, Z., Hou, T.: Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567 (2023) 4, 8 +90. Zhou, Y., Zhang, R., Chen, C., Li, C., Tensmeyer, C., Yu, T., Gu, J., Xu, J., Sun, T.: Towards language-free training for text-to-image generation. In: CVPR (2022) 8, 9 +91. Zhu, L.: Thop: Pytorch-ocounter. https://github.com/Lyken17/pytorchOpCounter (2018) 2 \ No newline at end of file diff --git a/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/images.zip b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..abec66d6ff94fc16242ccd7af944fa2ed1e13b1c --- /dev/null +++ b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37406b09dcd2dba4dc6f5d82d0c6e2be22042348aa6865865fd7591fb89a1a76 +size 973812 diff --git a/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/layout.json b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac7ed09bdc545c46c809d71545ab576bfa69a1b2 --- /dev/null +++ b/2024/BK-SDM_ A Lightweight, Fast, and Cheap Version of Stable Diffusion/layout.json @@ -0,0 +1,11200 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 146 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 146 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 158, + 167, + 455, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 167, + 455, + 192 + ], + "spans": [ + { + "bbox": [ + 158, + 167, + 455, + 192 + ], + "type": "text", + "content": "Bo-Kyeong Kim1, Hyoung-Kyu Song2, Thibault Castells1, and Shinkook Choi1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 240, + 201, + 373, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 201, + 373, + 213 + ], + "spans": [ + { + "bbox": [ + 240, + 201, + 373, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 240, + 201, + 373, + 213 + ], + "type": "text", + "content": "Nota Inc. " + }, + { + "bbox": [ + 240, + 201, + 373, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 240, + 201, + 373, + 213 + ], + "type": "text", + "content": "Captions Research" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 213, + 457, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 213, + 457, + 224 + ], + "spans": [ + { + "bbox": [ + 154, + 213, + 457, + 224 + ], + "type": "text", + "content": "{bokyeong.kim, thibault, shinkook.choi}@nota.ai, kyu@captions.ai" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 251, + 271, + 365 + ], + "blocks": [ + { + "bbox": [ + 133, + 251, + 271, + 365 + ], + "lines": [ + { + "bbox": [ + 133, + 251, + 271, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 251, + 271, + 365 + ], + "type": "image", + "image_path": "b6b9bcc8e0f2fcb2509332954f1cf0ba0cdc8d89478f5152869bb15fec8b3bb1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 366, + 243, + 373 + ], + "lines": [ + { + "bbox": [ + 161, + 366, + 243, + 373 + ], + "spans": [ + { + "bbox": [ + 161, + 366, + 243, + 373 + ], + "type": "text", + "content": "(a) Efficient General-purpose T2I" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 271, + 251, + 406, + 305 + ], + "blocks": [ + { + "bbox": [ + 271, + 251, + 406, + 305 + ], + "lines": [ + { + "bbox": [ + 271, + 251, + 406, + 305 + ], + "spans": [ + { + "bbox": [ + 271, + 251, + 406, + 305 + ], + "type": "image", + "image_path": "8a880b37aa421bf7c7039a28540b468c220868d370fd1272d98fdae10ecec360.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 271, + 312, + 406, + 365 + ], + "blocks": [ + { + "bbox": [ + 302, + 306, + 374, + 312 + ], + "lines": [ + { + "bbox": [ + 302, + 306, + 374, + 312 + ], + "spans": [ + { + "bbox": [ + 302, + 306, + 374, + 312 + ], + "type": "text", + "content": "(b) Efficient Personalized T2I" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 271, + 312, + 406, + 365 + ], + "lines": [ + { + "bbox": [ + 271, + 312, + 406, + 365 + ], + "spans": [ + { + "bbox": [ + 271, + 312, + 406, + 365 + ], + "type": "image", + "image_path": "faeb0988ef932fc94156e39fcdd65b7a628762c3e12c300ec4344a0908708d1e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 366, + 378, + 373 + ], + "lines": [ + { + "bbox": [ + 298, + 366, + 378, + 373 + ], + "spans": [ + { + "bbox": [ + 298, + 366, + 378, + 373 + ], + "type": "text", + "content": "(c) Efficient I2I (Different Styles)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "type": "text", + "content": "Fig. 1: Our compressed model enables efficient (a) zero-shot text-to-image generation, (b) personalized synthesis, (c) image-to-image translation, and (d) mobile deployment. Samples from BK-SDM-Small with " + }, + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "type": "text", + "content": " reduced parameters and latency are shown." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 406, + 251, + 480, + 324 + ], + "blocks": [ + { + "bbox": [ + 406, + 251, + 480, + 324 + ], + "lines": [ + { + "bbox": [ + 406, + 251, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 406, + 251, + 480, + 324 + ], + "type": "image", + "image_path": "b925a38ebba87d72e561973bbc2e826e809aca9bcfe9a5ed2ed5188e1d327b1f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 407, + 325, + 480, + 365 + ], + "blocks": [ + { + "bbox": [ + 407, + 325, + 480, + 365 + ], + "lines": [ + { + "bbox": [ + 407, + 325, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 407, + 325, + 480, + 365 + ], + "type": "image", + "image_path": "8674b31599d7a1041f81bceee447e162397c6b1e78c4ce1cfd30ca333a9c0277.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 366, + 471, + 373 + ], + "lines": [ + { + "bbox": [ + 414, + 366, + 471, + 373 + ], + "spans": [ + { + "bbox": [ + 414, + 366, + 471, + 373 + ], + "type": "text", + "content": "(d) Mobile Deployment" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 159, + 456, + 455, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 456, + 455, + 665 + ], + "spans": [ + { + "bbox": [ + 159, + 456, + 455, + 665 + ], + "type": "text", + "content": "Abstract. Text-to-image (T2I) generation with Stable Diffusion models (SDMs) involves high computing demands due to billion-scale parameters. To enhance efficiency, recent studies have reduced sampling steps and applied network quantization while retaining the original architectures. The lack of architectural reduction attempts may stem from worries over expensive retraining for such massive models. In this work, we uncover the surprising potential of block pruning and feature distillation for low-cost general-purpose T2I. By removing several residual and attention blocks from the U-Net of SDMs, we achieve " + }, + { + "bbox": [ + 159, + 456, + 455, + 665 + ], + "type": "inline_equation", + "content": "30\\% \\sim 50\\%" + }, + { + "bbox": [ + 159, + 456, + 455, + 665 + ], + "type": "text", + "content": " reduction in model size, MACs, and latency. We show that distillation retraining is effective even under limited resources: using only 13 A100 days and a tiny dataset, our compact models can imitate the original SDMs (v1.4 and v2.1-base with over 6,000 A100 days). Benefiting from the transferred knowledge, our BK-SDMs deliver competitive results on zero-shot MS-COCO against larger multi-billion parameter models. We further demonstrate the applicability of our lightweight backbones in personalized generation and image-to-image translation. Deployment of our models on edge devices attains 4-second inference. Code and models can be found at: https://github.com/Nota-NetsPresso/BK-SDM." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 115, + 286, + 203 + ], + "blocks": [ + { + "bbox": [ + 186, + 115, + 286, + 203 + ], + "lines": [ + { + "bbox": [ + 186, + 115, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 186, + 115, + 286, + 203 + ], + "type": "image", + "image_path": "33d091d3cee1eeb8313e995c1c2da5619ece4832718611e4730fb952929b77d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 212, + 482, + 235 + ], + "lines": [ + { + "bbox": [ + 131, + 212, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 482, + 235 + ], + "type": "text", + "content": "Fig. 2: Computation of Stable Diffusion. The denoising U-Net is the main processing bottleneck. THOP [91] is used to measure MACs in generating a " + }, + { + "bbox": [ + 131, + 212, + 482, + 235 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 131, + 212, + 482, + 235 + ], + "type": "text", + "content": " image." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 288, + 114, + 428, + 204 + ], + "blocks": [ + { + "bbox": [ + 288, + 114, + 428, + 204 + ], + "lines": [ + { + "bbox": [ + 288, + 114, + 428, + 204 + ], + "spans": [ + { + "bbox": [ + 288, + 114, + 428, + 204 + ], + "type": "image", + "image_path": "dcff685fa1d1cfe7d2f65477e552672de268f7f4144efe386bcce74762c4bd0b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 257, + 229, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 229, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 229, + 269 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 403 + ], + "type": "text", + "content": "Stable Diffusion models (SDMs) [58,60-62] are one of the most renowned open-source models for text-to-image (T2I) synthesis, and their exceptional capability has begun to be leveraged as a backbone in several text-guided vision applications [2,3,83,87]. SDMs are T2I-specialized latent diffusion models (LDMs) [58], which employ diffusion operations [23,37,76] in a semantically compressed space for compute efficiency. Within a SDM, a U-Net [9,64] performs iterative sampling to progressively denoise a random latent code and is aided by a text encoder [53] and an image decoder [12,81] to produce text-aligned images. This inference process still involves excessive computational requirements (see Fig. 2), which often hinder the utilization of SDMs despite their rapidly growing usage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": "To alleviate this issue, numerous approaches toward efficient SDMs have been introduced. A pretrained diffusion model is distilled to reduce the number of denoising steps, enabling an identically architectured model with fewer sampling steps [40, 43]. Post-training quantization [24, 34, 74] and implementation optimization [6, 7] methods are also leveraged. However, the removal of architectural elements in large diffusion models remains less explored." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 475, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 582 + ], + "type": "text", + "content": "This study unlocks the immense potential of classical architectural compression in attaining smaller and faster diffusion models. We eliminate multiple residual and attention blocks from the U-Net of a SDM and retrain it with feature-level knowledge distillation (KD) [18, 63] for general-purpose T2I. Under restricted training resources, our compact models can mimic the original SDM by leveraging transferred knowledge. Our work effectively reduces the computation of SDM-v1.4 [60] and SDM-v2.1-base [62] while achieving compelling zero-shot results on par with multi-billion parameter models [10, 11, 56]. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "text", + "content": "We compress SDMs by removing architectural blocks from the U-Net, achieving up to " + }, + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "text", + "content": " reduction in model size and " + }, + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "text", + "content": " improvement in latency on CPU and GPU. Previous pruning studies [13,47,84] focused on small models (" + }, + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "inline_equation", + "content": "< 100\\mathrm{M}" + }, + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "text", + "content": " parameters) like ResNet50 and DeiT-B, not on foundation models like SDMs (" + }, + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "inline_equation", + "content": ">1,000\\mathrm{M} = 1\\mathrm{B}" + }, + { + "bbox": [ + 138, + 593, + 483, + 665 + ], + "type": "text", + "content": "), possibly due to the lack of economic retraining for such large models. Moreover, U-Net architectures are arguably more com" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "B.-K. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 140 + ], + "type": "text", + "content": "plex due to the necessity of considering skip connections across the network, making the structural block removal inside them not straightforward." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 140, + 481, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 140, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 139, + 140, + 481, + 223 + ], + "type": "text", + "content": "To the best of our knowledge, we first demonstrate the notable benefit of feature distillation for training diffusion models, which enables competitive T2I even with significantly fewer resources (using only 13 A100 days and 0.22M LAION pairs [70]). Considering the vast expense of training SDMs from scratch (surpassing 6,000 A100 days and 2,000M pairs), our study indicates that network compression is a remarkably cost-effective strategy in building compact general-purpose diffusion models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 224, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 224, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 139, + 224, + 481, + 282 + ], + "type": "text", + "content": "We show the practicality of our work across various aspects. Our lightweight backbones are readily applicable to customized generation [65] and image-to-image translation [44], effectively lowering finetuning and inference costs. T2I synthesis on Jetson AGX Orin and iPhone 14 using our models takes less than 4 seconds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 285, + 481, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 285, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 139, + 285, + 481, + 320 + ], + "type": "text", + "content": "We have publicly released our approach, model weights, and source code, motivating subsequent works by other researchers (e.g., block pruning and KD for the SDM-v1 variant [5,72] and SDXL [31, 73])." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 339, + 237, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 339, + 237, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 237, + 352 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 509 + ], + "type": "text", + "content": "Large T2I diffusion models. By gradually removing noise from corrupted data, diffusion-based generative models [9, 22, 76] enable high-fidelity synthesis with broad mode coverage. Integrating these merits with the advancement of pretrained language models [8, 53, 54] has significantly improved the quality of T2I synthesis. In GLIDE [48] and Imagen [66], a text-conditional diffusion model generates a small image, which is upsampled via super-resolution modules. In DALL-E-2 [55], a text-conditional prior network produces an image embedding, which is transformed into an image via a diffusion decoder and further upscaled into higher resolutions. SDMs [58, 60–62] perform the diffusion modeling in a low-dimensional latent space constructed through a pixel-space autoencoder. We use SDMs as our baseline because of its open-access and gaining popularity over numerous downstream tasks [2, 3, 65, 83]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 604 + ], + "type": "text", + "content": "Efficient diffusion models. Several studies have addressed the slow sampling process. Diffusion-tailored distillation [42, 43, 68] progressively transfers knowledge from a pretrained diffusion model to a fewer-step model with the same architecture. Fast high-order solvers [38, 39, 88] for diffusion ordinary differential equations boost the sampling speed. Complementarily, our network compression approach reduces per-step computation and can be easily integrated with less sampling steps. Leveraging quantization [24, 34, 74] and implementation optimizations [6, 7] for SDMs can also be combined with our compact models for further efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Distillation-based compression. KD enhances the performance of small-size models by exploiting output-level [21, 49] and feature-level [18, 63, 85] information of large source models. Although this classical KD has been actively used for efficient GANs [33, 57, 86], its power has not been explored for structurally compressed diffusion models. Distillation pretraining enables small yet capable" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": "general-purpose language models [26, 69, 77] and vision transformers [16, 80]. Beyond such models, we show that its success can be extended to diffusion models with iterative sampling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 297 + ], + "type": "text", + "content": "Concurrent studies. SnapFusion [35] and MobileDiffusion [89] achieve an efficient U-Net for SDMs through architecture optimization and step reduction. Würstchen [50] introduces two diffusion processes on low- and high-resolution latent spaces for economic training. These works are valuable but require much larger training resources than our work (see Tab. 1). In contrast, we utilize considerably fewer resources, leveraging the surprising benefits of classical KD for foundational diffusion models. While not demonstrated on SDMs, Diff-Pruning [14] proposes structured pruning based on Taylor expansion tailored for diffusion models. We emphasize the use of depth (block) pruning in our work, which often leads to greater enhancements in inference speeds compared to the width (channel) pruning approach of Diff-Pruning. Moreover, Tab. 5 employs a comparable Taylor pruning criterion, which results in an insufficient decrease in MACs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 316, + 282, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 282, + 330 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 282, + 330 + ], + "type": "text", + "content": "3 Compression Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "type": "text", + "content": "We compress the U-Net [64] in SDMs, which is the most compute-heavy component (see Fig. 2). Conditioned on the text and time-step embeddings, the U-Net performs multiple denoising steps on latent representations. At each step, the U-Net produces the noise residual to compute the latent for the next step. We reduce this per-step computation, leading to Block-removed Knowledge-distilled SDMs (BK-SDMs)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 434, + 309, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 309, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 309, + 445 + ], + "type": "text", + "content": "3.1 Compact U-Net Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "type": "text", + "content": "The following architectures are obtained by compressing SDM-v1 (1.04B parameters), as shown in Fig. 3:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 491, + 420, + 527 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 138, + 491, + 369, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 369, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 369, + 502 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 138, + 491, + 369, + 502 + ], + "type": "text", + "content": " BK-SDM-Base (0.76B) obtained with Sec. 3.1.(1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 503, + 387, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 387, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 387, + 514 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 138, + 503, + 387, + 514 + ], + "type": "text", + "content": " BK-SDM-Small (0.66B) with Secs. 3.1.(1) and 3.1.(2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 515, + 420, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 515, + 420, + 527 + ], + "spans": [ + { + "bbox": [ + 138, + 515, + 420, + 527 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 138, + 515, + 420, + 527 + ], + "type": "text", + "content": " BK-SDM-Tiny (0.50B) with Secs. 3.1.(1), 3.1.(2), and 3.1.(3)." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 538, + 481, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 481, + 563 + ], + "type": "text", + "content": "Our approach can be identically applied to SDM-v2 (1.26B parameters), leading to BK-SDM-v2-{Base (0.98B), Small (0.88B), Tiny (0.72B)}." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "(1) Fewer Blocks in the Down and Up Stages. This approach is closely aligned with DistilBERT [69] which halves the number of layers and initializes the compact model with the original weights by benefiting from the shared dimensionality. In the original U-Net, each stage with a common spatial size consists of multiple blocks, and most stages contain pairs of residual (R) [17] and cross-attention (A) [25,82] blocks. We hypothesize the existence of some unnecessary pairs and use the following removal strategies." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "B.-K. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 187 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 187 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 187 + ], + "type": "image", + "image_path": "803b8479453aadc16b24f01220daa9eca29fa7f525a80a1d4c9280434523e209.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 194, + 482, + 229 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 229 + ], + "type": "text", + "content": "Fig. 3: Block removal from the denoising U-Net. Our approach is applicable to all the SDM versions in v1 and v2, which share the same U-Net block configuration. For experiments, we used v1.4 [60] and v2.1-base [62]. See Sec. A for the details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 239, + 235, + 304 + ], + "blocks": [ + { + "bbox": [ + 133, + 239, + 235, + 304 + ], + "lines": [ + { + "bbox": [ + 133, + 239, + 235, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 235, + 304 + ], + "type": "image", + "image_path": "045c11e1473319a1d92139396107748fe87d6c61f39c893953e1ad42ed9066bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 307, + 235, + 363 + ], + "blocks": [ + { + "bbox": [ + 133, + 307, + 235, + 363 + ], + "lines": [ + { + "bbox": [ + 133, + 307, + 235, + 363 + ], + "spans": [ + { + "bbox": [ + 133, + 307, + 235, + 363 + ], + "type": "image", + "image_path": "c8b74f10abcd00fdd81656ce0301e279fb1681c1b47af206d773c8eb77b02c7e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 366, + 236, + 431 + ], + "lines": [ + { + "bbox": [ + 132, + 366, + 236, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 236, + 431 + ], + "type": "text", + "content": "Fig. 4: Minor impact of removing the mid-stage from the U-Net. Results without retraining. See Sec. B for additional results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 244, + 239, + 480, + 306 + ], + "blocks": [ + { + "bbox": [ + 244, + 239, + 480, + 306 + ], + "lines": [ + { + "bbox": [ + 244, + 239, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 244, + 239, + 480, + 306 + ], + "type": "image", + "image_path": "55b1a0e44a56537e20c067b3453203869a67a4c7124e7cc205fb5d42039cc9aa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 244, + 308, + 480, + 376 + ], + "blocks": [ + { + "bbox": [ + 244, + 308, + 480, + 376 + ], + "lines": [ + { + "bbox": [ + 244, + 308, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 244, + 308, + 480, + 376 + ], + "type": "image", + "image_path": "64d1c5df0da70129ad90c294ff86c741ea16ab828b27d5496039d5bcc608c790.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 376, + 482, + 431 + ], + "lines": [ + { + "bbox": [ + 242, + 376, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 242, + 376, + 482, + 431 + ], + "type": "text", + "content": "Fig.5: Importance of (a) each block and (b) each group of paired/triplet blocks. Higher score implies removable blocks. The results are aligned with our architectures (e.g., removal of innermost stages and the second R-A pairs in down stages). See Sec. C for further analysis." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 464, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 464, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 139, + 464, + 482, + 594 + ], + "type": "text", + "content": "- Down Stages. We maintain the first R-A pairs while eliminating the second pairs, because the first pairs process the changed spatial information and would be more important than the second pairs. This design is consistent with the sensitivity analysis that measures the block-level significance (see Fig. 5). Our approach also does not harm the dimensionality of the original U-Net, enabling the use of the corresponding pretrained weights for initialization [69]. \n- Up Stages. While adhering to the aforementioned scheme, we retain the third R-A pairs. This allows us to utilize the output feature maps at the end of each down stage and the corresponding skip connections between the down and up stages. The same process is applied to the innermost down and up stages that contain only R blocks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "(2) Removal of the Entire Mid-Stage. Surprisingly, removing the entire mid-stage from the original U-Net does not noticeably degrade the generation quality while effectively reducing the parameters by " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " (see Fig. 4). This observation is consistent with the minor role of inner layers in the U-Net generator of GANs [28]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 202, + 114, + 413, + 184 + ], + "blocks": [ + { + "bbox": [ + 202, + 114, + 413, + 184 + ], + "lines": [ + { + "bbox": [ + 202, + 114, + 413, + 184 + ], + "spans": [ + { + "bbox": [ + 202, + 114, + 413, + 184 + ], + "type": "image", + "image_path": "be833c8fd0127a7f45a4ec7885c92d3cfd2540f798b19ff4799b9fc3003f4948.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 191, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 131, + 191, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 131, + 191, + 482, + 214 + ], + "type": "text", + "content": "Fig. 6: Distillation-based retraining. The block-removed U-Net is trained effectively through the guidance of the original U-Net." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 299 + ], + "type": "text", + "content": "Integrating the mid-stage removal with fewer blocks in Sec. 3.1 further decreases compute burdens (Tab. 2) at the cost of a slight decline in performance (Tab. 1). Therefore, we offer this mid-stage elimination as an option, depending on the priority between compute efficiency (using BK-SDM-Small) and generation quality (BK-SDM-Base)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 313, + 483, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 313, + 483, + 362 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 483, + 362 + ], + "type": "text", + "content": "(3) Further Removal of the Innermost Stages. For additional compression, the innermost down and up stages can also be pruned, leading to our lightest model BK-SDM-Tiny. This implies that outer stages with larger spatial dimensions and their skip connections play a crucial role in the U-Net for T2I synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 485 + ], + "type": "text", + "content": "Alignment with Pruning Sensitivity Analysis. To support the properness of our architectures, we measure the importance of each block (see Fig. 5) and show that unimportant blocks match with our design choices. The importance is measured by how generation scores vary when removing each residual or attention block from the U-Net. A significant drop in performance highlights the essential role of that block. Note that some blocks are not directly removable due to different channel dimensions between input and output; we replace such blocks with channel interpolation modules (denoted by “*” in Fig. 5) to mimic the removal while retaining the information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 534 + ], + "type": "text", + "content": "The sensitivity analysis implies that the innermost down-mid-up stages and the second R-A pairs in the down stages play relatively minor roles. Pruning these blocks aligns with our architectures, designed based on human knowledge (e.g., prioritizing blocks with altered channel dimensions) and previous studies [28, 69]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 571 + ], + "type": "text", + "content": "Though some results aligned, the importance criterion from the CLIP Score pruning sensitivity (in Fig. 5) leads to excessive pruning of attention blocks, eventually yielding inferior performance compared to our approach (see Table 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 586, + 307, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 307, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 307, + 600 + ], + "type": "text", + "content": "3.2 Distillation-based Retraining" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 641 + ], + "type": "text", + "content": "For general-purpose T2I, we train our block-removed U-Net to mimic the behavior of the original U-Net (see Fig. 6). To obtain the input of U-Net, we use pretrained-and-frozen encoders [58] for images and text prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Given the latent representation " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " of an image and its paired text embedding " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", the task loss for the reverse denoising process [22, 58] is computed as:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "B.-K. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 227, + 125, + 482, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 125, + 482, + 146 + ], + "spans": [ + { + "bbox": [ + 227, + 125, + 482, + 146 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T a s k}} = \\mathbb {E} _ {\\mathbf {z}, \\epsilon , \\mathbf {y}, t} \\left[ | | \\epsilon - \\epsilon_ {\\mathrm {S}} (\\mathbf {z} _ {t}, \\mathbf {y}, t) | | _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "4cc2b6be1f9fbf0a72cf1e061827d610db715bd0301daa117cea205faf2d9ae2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "spans": [ + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "text", + "content": " is a noisy latent code from the diffusion process [22] with the sampled noise " + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim N(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "text", + "content": " and time step " + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "inline_equation", + "content": "t \\sim \\mathrm{Uniform}(1,T)" + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{S}}(\\circ)" + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "text", + "content": " indicates the estimated noise from our compact U-Net student. For brevity, we omit the subscripts of " + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathbf{z},\\epsilon,\\mathbf{y},t}[\\circ]" + }, + { + "bbox": [ + 131, + 149, + 482, + 198 + ], + "type": "text", + "content": " in the following notations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "text", + "content": "The compact student is also trained to imitate the outputs of the original U-Net teacher, " + }, + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{T}}(\\circ)" + }, + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "text", + "content": ", with the following output-level KD objective [21]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 214, + 229, + 482, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 229, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 214, + 229, + 482, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {O u t K D}} = \\mathbb {E} \\left[ \\left| \\left| \\epsilon_ {\\mathrm {T}} (\\mathbf {z} _ {t}, \\mathbf {y}, t) - \\epsilon_ {\\mathrm {S}} (\\mathbf {z} _ {t}, \\mathbf {y}, t) \\right| \\right| _ {2} ^ {2} \\right]. \\tag {2}", + "image_path": "8a723021ebcf6937e668e1830227a91c32db96b342ae1936235db101c8b50375.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 255, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 481, + 281 + ], + "type": "text", + "content": "A key to our approach is feature-level KD [18, 63] that provides abundant guidance for the student's training:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 203, + 287, + 482, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 287, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 203, + 287, + 482, + 314 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {F e a t K D}} = \\mathbb {E} \\left[ \\sum_ {l} | | f _ {\\mathrm {T}} ^ {l} (\\mathbf {z} _ {t}, \\mathbf {y}, t) - f _ {\\mathrm {S}} ^ {l} (\\mathbf {z} _ {t}, \\mathbf {y}, t) | | _ {2} ^ {2} \\right], \\tag {3}", + "image_path": "9130872305c2c1f38e648638d2fec05cb02bf4978e4500c9bd185e23cd0a9a33.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{T}}^{l}(\\circ)" + }, + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{S}}^{l}(\\circ)" + }, + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "text", + "content": " represent the feature maps of the " + }, + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "text", + "content": "-th layer in a predefined set of distilled layers from the teacher and the student, respectively. While learnable regressors (e.g., " + }, + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 131, + 319, + 482, + 403 + ], + "type": "text", + "content": " convolutions to match the number of channels) have been commonly used [57,63,75], our approach circumvents this requirement. By applying distillation at the end of each stage in both models, we ensure that the dimensionality of the feature maps already matches, thus eliminating the need for additional regressors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 404, + 481, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 404, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 481, + 429 + ], + "type": "text", + "content": "The final objective is shown below, and we simply set " + }, + { + "bbox": [ + 131, + 404, + 481, + 429 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{OutKD}}" + }, + { + "bbox": [ + 131, + 404, + 481, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 404, + 481, + 429 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{FeatKD}}" + }, + { + "bbox": [ + 131, + 404, + 481, + 429 + ], + "type": "text", + "content": " as 1. Without loss-weight tuning, our approach is effective in empirical validation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 202, + 441, + 482, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 441, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 202, + 441, + 482, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {T a s k}} + \\lambda_ {\\text {O u t K D}} \\mathcal {L} _ {\\text {O u t K D}} + \\lambda_ {\\text {F e a t K D}} \\mathcal {L} _ {\\text {F e a t K D}}. \\tag {4}", + "image_path": "b455d094377f67cff0d3a7e514237b878001e2353bfb834130820c753a887667.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 472, + 274, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 274, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 274, + 487 + ], + "type": "text", + "content": "4 Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "text", + "content": "Distillation Retraining. We primarily use 0.22M image-text pairs from LAION-Aesthetics V2 (L-Aes) " + }, + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "inline_equation", + "content": "6.5 + [70,71]" + }, + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "text", + "content": ", which are significantly fewer than the original training data used for SDMs [60,62] (" + }, + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "inline_equation", + "content": ">2,000\\mathrm{M}" + }, + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "text", + "content": " pairs). In Fig. 11, dataset sizes smaller than 0.22M are randomly sampled from L-Aes " + }, + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "inline_equation", + "content": "6.5+" + }, + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "text", + "content": ", while those larger than 0.22M are from L-Aes " + }, + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "inline_equation", + "content": "6.25+" + }, + { + "bbox": [ + 131, + 497, + 481, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 557, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 481, + 629 + ], + "type": "text", + "content": "Zero-shot T2I Evaluation. Following the popular protocol [56, 58, 66], we use 30K prompts from the MS-COCO validation split [36], downsample the " + }, + { + "bbox": [ + 131, + 557, + 481, + 629 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 131, + 557, + 481, + 629 + ], + "type": "text", + "content": " generated images to " + }, + { + "bbox": [ + 131, + 557, + 481, + 629 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 131, + 557, + 481, + 629 + ], + "type": "text", + "content": ", and compare them with the entire validation set. We compute Fréchet Inception Distance (FID) [20] and Inception Score (IS) [67] to assess visual quality. We measure CLIP score [19, 53] with CLIP-ViT-g/14 model to assess text-image correspondence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "content": "Downstream Tasks. For personalized generation, we use the DreamBooth dataset [65] (30 subjects " + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "content": " 25 prompts " + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 131, + 629, + 481, + 667 + ], + "type": "text", + "content": " 4~6 images) and perform per-subject finetuning. Following the evaluation protocol [65], we use ViT-S/16 model [4]" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 90, + 448, + 102 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 180, + 478, + 399 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Results on zero-shot MS-COCO " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " 30K. Training resources include image-text pairs, batch size, iterations, and A100 days. Despite far smaller resources, our compact models outperform prior studies [10, 11, 56, 79, 90], showing the benefit of compressing existing powerful models. Note that FID fluctuates more than the other metrics over training progress in our experiments (see Figs. 9 and 11)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 180, + 478, + 399 + ], + "lines": [ + { + "bbox": [ + 136, + 180, + 478, + 399 + ], + "spans": [ + { + "bbox": [ + 136, + 180, + 478, + 399 + ], + "type": "table", + "html": "
Model\nNameType #Param‡Generation ScoreTraining Resource
FID↓IS↑CLIP↑Data Size (Batch, # Iter)A100 Days
SDM-v1.4 [58,60]†DF1.04B13.0536.760.2958>2000M*(2048, 1171K)6250
Small Stable Diffusion [51]†DF0.76B12.7632.330.2851229M(128, 1100K)-
BK-SDM-Base [Ours]†DF0.76B15.7633.790.28780.22M(256, 50K)13
BK-SDM-Small [Ours]†DF0.66B16.9831.680.26770.22M(256, 50K)13
BK-SDM-Tiny [Ours]†DF0.50B17.1230.090.26530.22M(256, 50K)13
SDM-v2.1-base [58,62]†DF1.26B13.9335.930.3075>2000M*(2048, 1620K)8334
BK-SDM-v2-Base [Ours]†DF0.98B15.8531.700.28680.22M(128, 50K)4
BK-SDM-v2-Small [Ours]†DF0.88B16.6131.730.29010.22M(128, 50K)4
BK-SDM-v2-Tiny [Ours]†DF0.72B15.6831.640.28970.22M(128, 50K)4
DALL-E [56]AR12B27.517.9-250M(1024, 430K)-
CogView [10]AR4B27.118.2-30M(6144, 144K)-
CogView2 [11]AR6B2422.4-30M(4096, 300K)-
Make-A-Scene [15]AR4B11.84--35M(1024, 170K)-
LAFITE [90]GAN0.23B26.9426.02-3M--
GALIP (CC12M) [79]†GAN0.32B13.8625.160.281712M--
GigaGAN [27]GAN1.1B9.09-->100M*(512, 1350K)4783
GLIDE [48]DF3.5B12.24--250M(2048, 2500K)-
LDM-KL-8-G [58]DF1.45B12.6330.29-400M(680, 390K)-
DALL-E-2 [55]DF5.2B10.39--250M(4096, 3400K)-
SnapFusion [35]DF0.99B~13.6-~0.295>100M*(2048, -)>128*
Würstchen-v2 [50]†DF3.1B22.4032.870.26761700M(1536, 1725K)1484
MobileDiffusion [89]†DF0.52B9.01--150M(2048, 1000K)-
", + "image_path": "4e13e3ffb39ed17b06a09be93a8dabcc72e0939a78a4b57d1f8ac4bde3a50d74.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "lines": [ + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "spans": [ + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "text", + "content": ": Evaluated with the released checkpoints. " + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "text", + "content": ": Total parameters for T2I synthesis. " + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "text", + "content": ": Estimated based on public information. DF and AR: diffusion and autoregressive models. " + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 135, + 401, + 478, + 419 + ], + "type": "text", + "content": ": lower and higher values are better." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 447, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 472 + ], + "type": "text", + "content": "for DINO score and CLIP-ViT-g/14 model for CLIP-I and CLIP-T scores. For image-to-image translation, input images are sourced from Meng et al. [43]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 473, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 558 + ], + "type": "text", + "content": "Implementation Details. We adjust the codes in Diffusers [52] and PEFT [41]. We use a single NVIDIA A100 80G GPU for main retraining and a single NVIDIA GeForce RTX 3090 GPU for per-subject finetuning. For compute efficiency, we always opt for 25 denoising steps of the U-Net at the inference phase, unless specified. The classifier-free guidance scale [23,66] is set to the default value of 7.5. The latent resolution is set to the default (" + }, + { + "bbox": [ + 130, + 473, + 482, + 558 + ], + "type": "inline_equation", + "content": "H = W = 64" + }, + { + "bbox": [ + 130, + 473, + 482, + 558 + ], + "type": "text", + "content": " in Fig. 3), yielding " + }, + { + "bbox": [ + 130, + 473, + 482, + 558 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 473, + 482, + 558 + ], + "type": "text", + "content": " images. See Sec. J for the details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 590, + 198, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 198, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 198, + 602 + ], + "type": "text", + "content": "5 Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "All the results in Secs. 5.1-5.6 were obtained with the full benchmark protocol (MS-COCO " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " 30K samples), except for Fig. 10 (" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " 5K samples). Unless specified, the training setup in Tab. 1 was used." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "B.-K. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 242 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 242 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 242 + ], + "type": "image", + "image_path": "cfb355cb7ac4f643ced756f39c9166f92ab7017d45c90a21a398387b8461f6f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 242, + 476, + 249 + ], + "lines": [ + { + "bbox": [ + 133, + 242, + 476, + 249 + ], + "spans": [ + { + "bbox": [ + 133, + 242, + 476, + 249 + ], + "type": "text", + "content": "Prompt: A bowl that has vegetables inside of it; A brown and white cat staring off with pretty green eyes.; A toy raccoon standing on a pile of broccoli." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 259, + 479, + 280 + ], + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 280 + ], + "type": "text", + "content": "Fig. 7: Visual comparison with open-sourced models. The results [11,48,50,79,90] were obtained with their official codes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "text", + "content": "Table 2: Impact of compute reduction in U-Net on the entire SDM. The number of sampling steps is indicated with the parentheses, e.g., U-Net (1) for one step. The full computation (denoted by \"Whole\") covers the text encoder, U-Net, and image decoder. All corresponding values are obtained on the generation of a single " + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "text", + "content": " image with 25 denoising steps. The latency was measured on Xeon Silver 4210R CPU 2.40GHz and NVIDIA GeForce RTX 3090 GPU." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 369, + 482, + 444 + ], + "blocks": [ + { + "bbox": [ + 133, + 369, + 482, + 444 + ], + "lines": [ + { + "bbox": [ + 133, + 369, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 482, + 444 + ], + "type": "table", + "html": "
Model# ParamMACsCPU LatencyGPU Latency
U-NetWholeU-Net (1)U-Net (25)WholeU-Net (1)U-Net (25)WholeU-Net (1)U-Net (25)Whole
SDM-v1.4 [60]860M1033M339G8469G9716G5.63s146.28s153.00s0.049s1.28s1.41s
BK-SDM-Base [Ours]580M752M224G5594G6841G3.84s99.95s106.67s0.032s0.83s0.96s
(-32.6%)(-27.1%)(-33.9%)(-33.9%)(-29.5%)(-31.8%)(-31.7%)(-30.3%)(-34.6%)(-35.2%)(-31.9%)
BK-SDM-Small [Ours]483M655M218G5444G6690G3.45s89.78s96.50s0.030s0.77s0.90s
(-43.9%)(-36.5%)(-35.7%)(-35.7%)(-31.1%)(-38.7%)(-38.6%)(-36.9%)(-38.7%)(-39.8%)(-36.1%)
BK-SDM-Tiny [Ours]324M496M206G5126G6373G3.03s78.77s85.49s0.026s0.67s0.80s
(-62.4%)(-51.9%)(-39.5%)(-39.5%)(-34.4%)(-46.2%)(-46.1%)(-44.1%)(-46.9%)(-47.7%)(-43.2%)
", + "image_path": "0c03c70a3823e356667282666c42e6ec0553c7eba452dbd491aa2fe9ba88a43a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 466, + 328, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 328, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 328, + 478 + ], + "type": "text", + "content": "5.1 Comparison with Existing Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 533 + ], + "type": "text", + "content": "Quantitative Comparison. Tab. 1 shows the zero-shot results for general-purpose T2I. Despite being trained with only 0.22M samples and having fewer than 1B parameters, our compressed models demonstrate competitive performance on par with existing large models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 605 + ], + "type": "text", + "content": "Visual Comparison. Fig. 7 depicts synthesized images with some MS-COCO captions. Our compact models inherit the superiority of SDM and produce more photorealistic images compared to the AR-based [11] and GAN-based [79, 90] baselines. Noticeably, the same latent code results in a shared visual style between the original and our models (6th–9th columns in Fig. 7), similar to the observation in transfer learning for GANs [45]. See Sec. D for additional results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 622, + 263, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 263, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 263, + 634 + ], + "type": "text", + "content": "5.2 Computational Gain" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Tab. 2 shows how the compute reduction for each sampling step of the U-Net affects the overall process. The per-step reduction effectively decreases MACs" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 148, + 288, + 239 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 291, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 291, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 291, + 137 + ], + "type": "text", + "content": "Table 3: Quality gains from distillation retraining. MS-COCO 30K." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 148, + 288, + 239 + ], + "lines": [ + { + "bbox": [ + 135, + 148, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 135, + 148, + 288, + 239 + ], + "type": "table", + "html": "
BK-SDM\nTypeKDGenerationScore# Param\nU-Net
FID↓IS↑CLIP↑
BaseX23.5723.020.24080.58B
✓15.7633.790.2878
v2-BaseX16.7625.880.26610.59B
✓15.8531.700.2868
v2-SmallX16.7125.770.26550.49B
✓16.6131.730.2901
v2-TinyX16.8726.060.26780.33B
✓15.6831.640.2897
", + "image_path": "4a3176249c534bb2256fc922e395cb48c2c6a8fafac118c6a56351595d8a3a1c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 298, + 160, + 478, + 238 + ], + "blocks": [ + { + "bbox": [ + 294, + 116, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 294, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 294, + 116, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Significance of each element in transferred knowledge. Results of BK-SDMv2-Small on MS-COCO 30K." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 160, + 478, + 238 + ], + "lines": [ + { + "bbox": [ + 298, + 160, + 478, + 238 + ], + "spans": [ + { + "bbox": [ + 298, + 160, + 478, + 238 + ], + "type": "table", + "html": "
Weight InitializationOutput FeatureGeneration Score
KDKDFID↓IS↑CLIP↑
RandomXX41.7515.420.1733
Random✓X29.0118.290.1967
Random✓✓24.4722.370.2323
TeacherXX16.7125.770.2655
Teacher✓X14.2729.470.2777
Teacher✓✓16.6131.730.2901
", + "image_path": "ebaaa7a0297bb970b20e7e9242421c9d17e52116cde7be5b002cea6374312756.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 154, + 248, + 460, + 375 + ], + "blocks": [ + { + "bbox": [ + 154, + 248, + 460, + 375 + ], + "lines": [ + { + "bbox": [ + 154, + 248, + 460, + 375 + ], + "spans": [ + { + "bbox": [ + 154, + 248, + 460, + 375 + ], + "type": "image", + "image_path": "a0558c4dd4c4092b6bb06418cb3166a283d39ead661c10f564a4f82e1dce13ae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 384, + 482, + 430 + ], + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 430 + ], + "type": "text", + "content": "Fig. 8: Image areas affected by each word. KD enables our models to mimic the SDM, yielding similar per-word attribution maps. The model without KD behaves differently, causing dissimilar maps and inaccurate generation (e.g., two sheep and unusual bird shapes)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "content": "and inference time by more than " + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "content": ". Notably, BK-SDM-Tiny has " + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "content": " fewer parameters than the original SDM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 491, + 328, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 328, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 328, + 504 + ], + "type": "text", + "content": "5.3 Benefit of Distillation Retraining" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": "T2I Performance. Tab. 3 summarizes the results from ablating the total KD objective (Eq. 2+Eq. 3). Across various model types, distillation brings a clear improvement in generation quality. Tab. 4 analyzes the effect of each element in transferred knowledge. Exploiting output-level KD (Eq. 2) boosts the performance compared to using only the denoising task loss. Leveraging feature-level KD (Eq. 3) yields further score enhancement. Additionally, using the teacher weights for initialization is highly beneficial." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Cross-attention Resemblance. Fig. 8 displays the per-word attribution maps [78] created by aggregating cross-attention scores over spatiotemporal dimensions. The attribution maps of our models are semantically and spatially similar to those of the original model, indicating the merit of supervisory signals at multiple stages via KD. In contrast, the baseline model without KD activates incorrect areas, leading to text-mismatched generation results." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "B.-K. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 207, + 201 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 207, + 201 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 207, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 207, + 201 + ], + "type": "image", + "image_path": "f6d779e2f565da131a56c95ae295d76bf605647dc0c3e59797a2ba1158d0d112.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 211, + 359, + 243 + ], + "lines": [ + { + "bbox": [ + 131, + 211, + 359, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 359, + 243 + ], + "type": "text", + "content": "Fig.9: Zero-shot results over training progress. The architecture size of BK-SDM, usage of KD, and batch size are denoted. Results on MS-COCO 30K." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 208, + 115, + 281, + 201 + ], + "blocks": [ + { + "bbox": [ + 208, + 115, + 281, + 201 + ], + "lines": [ + { + "bbox": [ + 208, + 115, + 281, + 201 + ], + "spans": [ + { + "bbox": [ + 208, + 115, + 281, + 201 + ], + "type": "image", + "image_path": "6bd5100732c37d9bde039be5b4df979ab4b09b2e1904ac0adf1e7ac7829ba24b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 281, + 115, + 356, + 201 + ], + "blocks": [ + { + "bbox": [ + 281, + 115, + 356, + 201 + ], + "lines": [ + { + "bbox": [ + 281, + 115, + 356, + 201 + ], + "spans": [ + { + "bbox": [ + 281, + 115, + 356, + 201 + ], + "type": "image", + "image_path": "8da766e3190a210889185f8bf4d6cc18f628b90837a238fef02cfdd3da1ee28b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 201, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 364, + 201, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 364, + 201, + 482, + 243 + ], + "type": "text", + "content": "Fig.10:Trade-off curves Left:FID vs.CLIP score Right: quality vs.efficiency Ours-Base on MS-COCO 5K." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 481, + 192 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 481, + 192 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 481, + 192 + ], + "type": "image", + "image_path": "54b695e4dd62078adc4fd40badb1b3a3ff07671d80e92840633383e7c4ad1566.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 277, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 386 + ], + "type": "text", + "content": "Training Progress. Fig. 9 shows the results over training iterations. Without KD, training solely with the denoising task loss causes fluctuations or sudden drops in performance (indicated with green and cyan). In contrast, distillation (purple and pink) stabilizes and accelerates the training process, demonstrating the benefit of providing sufficient hints for training guidance. Notably, our small-and tiny-size models trained with KD (yellow and red) outperform the bigger base-size model without KD (cyan). Additionally, while the best FID score is observed early on, IS and CLIP score exhibit ongoing improvement, implying that judging models merely with FID may be suboptimal." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 386, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 445 + ], + "type": "text", + "content": "Trade-off Results. Fig. 10 shows the results of BK-SDM-Base with and without KD on MS-COCO " + }, + { + "bbox": [ + 130, + 386, + 482, + 445 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 386, + 482, + 445 + ], + "type": "text", + "content": " 5K. Higher classifier-free guidance scales [23,66] lead to better text-aligned images at the cost of less diversity. More denoising steps improve generation quality at the cost of slower inference. Distillation retraining achieves much better trade-off curves than the baseline without KD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 463, + 383, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 383, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 383, + 475 + ], + "type": "text", + "content": "5.4 Comparison with Different Pruning Criteria" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 481, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 567 + ], + "type": "text", + "content": "We extend our baseline comparisons by calculating several block-level importance scores [29].1 Furthermore, the impact of applying KD at the end of each stage is also examined. Tab. 5 shows the results. Magnitude (Mag) pruning [32] removes critical outer blocks, causing irrecoverable loss. Taylor pruning [14,30,46] removes only the inner part, causing ineffective reduction in MACs and speed; additionally, this method demands the intensive calculation of backward gradients. The CLIP Score method (based on Fig. 5) results in severe pruning of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "text", + "content": "1 Let the " + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "inline_equation", + "content": "(h_{\\mathrm{out}}, h_{\\mathrm{in}})" + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "text", + "content": "-size weight matrix in the " + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "text", + "content": "-th block be " + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{l,b} = \\left[W_{i,j}^{l,b}\\right]" + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "text", + "content": " denotes the layer type (e.g., convolution (flattened) or attention's key projection). The scores at the output neuron level [1] are aggregated for the block-level importance criteria, " + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{Magnitude}}^b = \\mathbb{E}_{l,i}\\left[\\sum_j\\left|W_{i,j}^{l,b}\\right|\\right]" + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{Taylor}}^b = \\mathbb{E}_{l,i}\\left[\\sum_j\\left|\\frac{\\partial\\mathcal{L}(D)}{\\partial W_{i,j}^{l,b}}W_{i,j}^{l,b}\\right|\\right]" + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 133, + 573, + 482, + 665 + ], + "type": "text", + "content": " denote the denoising task loss and a calibration set of 1K samples. The final scores are then ranked to remove unimportant blocks, or to replace them with interpolation for unremovable blocks." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 323, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 323, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 323, + 180 + ], + "type": "text", + "content": "Table 5: Different block-level criteria at similar parameters. Taylor pruning removes solely the inner blocks, leading to insufficient reduction in MACs. Our method attains a favorable compromise between performance and inference speed. Results on MS-COCO 30K." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 135, + 191, + 320, + 242 + ], + "blocks": [ + { + "bbox": [ + 135, + 191, + 320, + 242 + ], + "lines": [ + { + "bbox": [ + 135, + 191, + 320, + 242 + ], + "spans": [ + { + "bbox": [ + 135, + 191, + 320, + 242 + ], + "type": "table", + "html": "
Pruning \nCriterionFID↓CLIP Score↑U-Net
KD×KD√KD×KD√MACs# Param
Mag [32]>150>150<0.1<0.1122.2G495.5M
Taylor [14,30]18.8516.730.24290.2749305.4G493.4M
CLIP [Fig. 5]32.9321.860.19360.2283204.8G496.8M
Ours-Small22.9716.830.22870.2668217.7G482.3M
", + "image_path": "e86b102d0e614bc381f723b7a952e9983c66c4064285f9163699156b87681190.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 134, + 243, + 312, + 258 + ], + "lines": [ + { + "bbox": [ + 134, + 243, + 312, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 243, + 312, + 258 + ], + "type": "text", + "content": "Retraining with batch 128, 0.22M data, and 50K iters. Original SD-v1.4's U-Net: (MACs, Params) = (338.7G, 859.5M)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 328, + 164, + 479, + 212 + ], + "blocks": [ + { + "bbox": [ + 325, + 130, + 481, + 153 + ], + "lines": [ + { + "bbox": [ + 325, + 130, + 481, + 153 + ], + "spans": [ + { + "bbox": [ + 325, + 130, + 481, + 153 + ], + "type": "text", + "content": "Table 6: A/B-testing preference study. Total 2,501 comparisons." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 164, + 479, + 212 + ], + "lines": [ + { + "bbox": [ + 328, + 164, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 328, + 164, + 479, + 212 + ], + "type": "table", + "html": "
Ours-SmallGLIDE-filt [48]GALIP [79]Würstchen-v2 SDM-v1.4 [50][60]
# Win†453 (71.5%)379 (61.5%)319 (51.4%)202 (32.1%)
# Defeat†181 (28.5%)237 (38.5%)302 (48.6%)428 (67.9%)
", + "image_path": "c316aaee828a6ce6bd3459f5f0f84700674fb9554353b7d2a10d6f90c13de787.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 327, + 213, + 479, + 243 + ], + "lines": [ + { + "bbox": [ + 327, + 213, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 327, + 213, + 479, + 243 + ], + "type": "text", + "content": "Participants selected preferred images from pairs with the same prompts. One image in each pair was always from BK-SDM-Small, while the other was randomly chosen from the four methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 135, + 314, + 323, + 399 + ], + "blocks": [ + { + "bbox": [ + 131, + 281, + 324, + 303 + ], + "lines": [ + { + "bbox": [ + 131, + 281, + 324, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 281, + 324, + 303 + ], + "type": "text", + "content": "Table 7: Impact of training batch size and iterations. Results on MS-COCO 30K." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 314, + 323, + 399 + ], + "lines": [ + { + "bbox": [ + 135, + 314, + 323, + 399 + ], + "spans": [ + { + "bbox": [ + 135, + 314, + 323, + 399 + ], + "type": "table", + "html": "
BK-SDMBaseSmallTiny
Batch Size642566425664256
FID↓14.6115.7616.8716.9817.2817.12
IS↑31.4433.7929.5131.6828.3330.09
CLIP↑0.28260.28780.26440.26770.26070.2653
BK-SDMBase (Data 2.3M)Small (Data 2.3M)Tiny (Data 2.3M)
#Iter50K100K50K100K50K100K
FID↓14.8115.3917.0517.0117.5317.63
IS↑34.1734.7633.1033.1431.3232.26
CLIP↑0.28830.28890.27340.27540.26900.2713
", + "image_path": "e9ac6b9fba2754ca57a4f8f6a1f5be639a7fef48ece41e149d2523ba21edb759.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 269, + 403, + 358 + ], + "blocks": [ + { + "bbox": [ + 329, + 269, + 403, + 358 + ], + "lines": [ + { + "bbox": [ + 329, + 269, + 403, + 358 + ], + "spans": [ + { + "bbox": [ + 329, + 269, + 403, + 358 + ], + "type": "image", + "image_path": "b57e17a2fb09f53b6f478788d756b0a1b43a0ec56f8217cedd37595525806530.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 369, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 328, + 369, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 328, + 369, + 482, + 401 + ], + "type": "text", + "content": "Fig. 11: Impact of training data volume. Results of BK-SDM-Small on MS-COCO 30K." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 404, + 269, + 480, + 357 + ], + "blocks": [ + { + "bbox": [ + 404, + 269, + 480, + 357 + ], + "lines": [ + { + "bbox": [ + 404, + 269, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 404, + 269, + 480, + 357 + ], + "type": "image", + "image_path": "9cb8d13181c912f273bc88dddfc2ffe89b622298ff8d4564dc05a8142a604557.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 495 + ], + "type": "text", + "content": "attention blocks, adversely affecting performance. Our approach achieves superior or comparable results against the benchmark methods and can be directly applied to all the SDM versions in v1 and v2. This advantage comes without the extra phase of calculating pruning criteria while ensuring a favorable balance between performance and inference speed. Notably, KD always remains effective." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 514, + 316, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 316, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 316, + 525 + ], + "type": "text", + "content": "5.5 Human Preference Assessment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Despite their extensive use, automated metrics for assessing image quality (e.g., FID) do not consistently match with human evaluation [55, 66]. To further address this, we have conducted an A/B testing-based study: participants were shown pairs of images generated using the same MS-COCO prompt and asked to choose their preferred image in relation to the prompt. One image in each pair was always from our model, while the other was randomly selected from one of four comparative methods. A total of 25 participants conducted 2,501 comparisons online. The participants had no bias towards the examined models and were not compensated. The prompts were randomly drawn from the MS-COCO set and varied for 3 to 5 participants. As shown in Tab. 6, our model (0.66B parameters) excels over GLIDE-filtered, GALIP, and Wüstchen-v2 (3.1B)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "B.-K. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 114, + 369, + 170 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 369, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 369, + 170 + ], + "type": "text", + "content": "Table 8: Personalized generation with finetuning over different backbones. Our compact models can preserve subject fidelity (DINO and CLIP-I) and prompt fidelity (CLIP-T) of the original SDM with reduced finetuning (FT) costs and fewer parameters." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 135, + 180, + 365, + 256 + ], + "blocks": [ + { + "bbox": [ + 135, + 180, + 365, + 256 + ], + "lines": [ + { + "bbox": [ + 135, + 180, + 365, + 256 + ], + "spans": [ + { + "bbox": [ + 135, + 180, + 365, + 256 + ], + "type": "table", + "html": "
Backbone {Param}DINO↑CLIP-I↑CLIP-T↑FT (Time, Mem)†
SDM-v1.4 [60] {1.04B}0.7280.7250.263(882s, 23.0GB)
BK-SDM-Base {0.76B}0.7230.7170.260(623s, 18.7GB)
BK-SDM-Small {0.66B}0.7200.7050.259(604s, 17.2GB)
BK-SDM-Tiny {0.50B}0.7150.6930.261(560s, 13.1GB)
Base (Batch 64) {0.76B}0.7180.7080.262(623s, 18.7GB)
- No KD & Random Init.0.5940.4650.191(623s, 18.7GB)
- No KD & Teacher Init.0.7160.6690.258(623s, 18.7GB)
", + "image_path": "25960ae8bbd48c0a2738589b6f455242df720db080625ce42b3a8196cda2836f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 135, + 258, + 365, + 267 + ], + "lines": [ + { + "bbox": [ + 135, + 258, + 365, + 267 + ], + "spans": [ + { + "bbox": [ + 135, + 258, + 365, + 267 + ], + "type": "text", + "content": "†: Per-subject FT time and GPU memory for 800 iterations on RTX 3090." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 374, + 167, + 479, + 248 + ], + "blocks": [ + { + "bbox": [ + 371, + 133, + 481, + 155 + ], + "lines": [ + { + "bbox": [ + 371, + 133, + 481, + 155 + ], + "spans": [ + { + "bbox": [ + 371, + 133, + 481, + 155 + ], + "type": "text", + "content": "Table 9: Speedups on edge devices." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 374, + 167, + 479, + 248 + ], + "lines": [ + { + "bbox": [ + 374, + 167, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 374, + 167, + 479, + 248 + ], + "type": "table", + "html": "
ModelAGX Orin 32GBiPhone 14
SDM-v14.9s [61]5.6s [60]
BK-SDM-Base3.4s (-31%)4.0s (-29%)
BK-SDM-Small3.2s (-35%)3.9s (-29%)
BK-SDM-Tiny2.8s (-43%)3.9s (-29%)
", + "image_path": "dc2f6913ab6103edb4a9b261e1e3d5bcf4067f47fa0948e5cb6e2ca02b08e5ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 277, + 361, + 373 + ], + "blocks": [ + { + "bbox": [ + 133, + 277, + 361, + 373 + ], + "lines": [ + { + "bbox": [ + 133, + 277, + 361, + 373 + ], + "spans": [ + { + "bbox": [ + 133, + 277, + 361, + 373 + ], + "type": "image", + "image_path": "64329bbe172f3cba1d82952c1f0148ea45368e99bc02947d14dbbdec5319fe22.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 381, + 482, + 404 + ], + "lines": [ + { + "bbox": [ + 131, + 381, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 482, + 404 + ], + "type": "text", + "content": "Fig. 12: Visual results of personalized generation. Each subject is marked as \"a [identifier] [class noun]\" (e.g., \"a [V] dog\")." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 362, + 278, + 480, + 372 + ], + "blocks": [ + { + "bbox": [ + 362, + 278, + 480, + 372 + ], + "lines": [ + { + "bbox": [ + 362, + 278, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 362, + 278, + 480, + 372 + ], + "type": "image", + "image_path": "37a3ea26f54b4cc2ad8d07c1029c0ce3d40a053089701804f24f11a73c238dd1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 416, + 249, + 506 + ], + "blocks": [ + { + "bbox": [ + 133, + 416, + 249, + 506 + ], + "lines": [ + { + "bbox": [ + 133, + 416, + 249, + 506 + ], + "spans": [ + { + "bbox": [ + 133, + 416, + 249, + 506 + ], + "type": "image", + "image_path": "a9a460561ad1244a1c71559b8eb11c8333a389e403460129ff3bb113e3f27d36.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 506, + 332, + 513 + ], + "lines": [ + { + "bbox": [ + 133, + 506, + 332, + 513 + ], + "spans": [ + { + "bbox": [ + 133, + 506, + 332, + 513 + ], + "type": "text", + "content": "(left-to-right, top-to-bottom) Target style: Art station; Oil painting; Photograph; Watercolor" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 523, + 361, + 534 + ], + "lines": [ + { + "bbox": [ + 135, + 523, + 361, + 534 + ], + "spans": [ + { + "bbox": [ + 135, + 523, + 361, + 534 + ], + "type": "text", + "content": "Fig. 13: Text-guided image-to-image translation." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 249, + 416, + 365, + 506 + ], + "blocks": [ + { + "bbox": [ + 249, + 416, + 365, + 506 + ], + "lines": [ + { + "bbox": [ + 249, + 416, + 365, + 506 + ], + "spans": [ + { + "bbox": [ + 249, + 416, + 365, + 506 + ], + "type": "image", + "image_path": "46e70f60fbb2ae51979cf68452df557d01bb97b385c525b482320cc4d2c315df.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 511, + 481, + 533 + ], + "lines": [ + { + "bbox": [ + 370, + 511, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 370, + 511, + 481, + 533 + ], + "type": "text", + "content": "Fig. 14: Small LDM for face synthesis." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 372, + 416, + 481, + 502 + ], + "blocks": [ + { + "bbox": [ + 372, + 416, + 481, + 502 + ], + "lines": [ + { + "bbox": [ + 372, + 416, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 372, + 416, + 481, + 502 + ], + "type": "image", + "image_path": "ec1f16501c99b509c4228d59a5421bd5e4f0df0fe8459e0e144a5c81ec676e4f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 570, + 392, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 392, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 392, + 582 + ], + "type": "text", + "content": "5.6 Impact of Training Resources on Performance" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Consistent with existing evidence, the use of larger batch sizes, more extensive data, and longer iterations for training enhances performance in our work (see Tab. 7, Fig. 11, and Sec. H). However, this benefit requires increased resource demands (e.g., extended training days without multiple high-spec GPUs and greater data storage capacity). As such, despite the better performing models in Tab. 7, we primarily report the models with fewer resources. We believe that" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "type": "text", + "content": "accessible training costs by many researchers can help drive advancement in massive models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 219, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 219, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 219, + 172 + ], + "type": "text", + "content": "5.7 Application" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 266 + ], + "type": "text", + "content": "Personalized T2I Synthesis. Tab. 8 compares the fine-tuning results with DreamBooth [65] over different backbones to create images about a given subject. BK-SDMs can preserve " + }, + { + "bbox": [ + 130, + 183, + 482, + 266 + ], + "type": "inline_equation", + "content": "95\\% \\sim 99\\%" + }, + { + "bbox": [ + 130, + 183, + 482, + 266 + ], + "type": "text", + "content": " scores of the original SDM while cutting fine-tuning costs. Fig. 12 depicts that our models can accurately capture the subject details and generate various scenes. Over the models retrained with a batch size of 64, the baselines without KD fail to generate the subjects or cannot maintain the identity details. See Sec. E for further results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 267, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 314 + ], + "type": "text", + "content": "Image-to-Image Translation. Fig. 13 presents the text-guided stylization results with SDEdit [44]. Our model, resembling the ability of the original SDM, faithfully produces images given style-specified prompts and content pictures. See Sec. F for additional results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "text", + "content": "Deployment on Edge Devices. We deploy our models trained with 2.3M pairs and compare them against the original SDM under the same setup on edge devices (20 denoising steps on NVIDIA Jetson AGX Orin 32GB and 10 steps on iPhone 14). Our models produce a " + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "text", + "content": " image within 4 seconds (see Tab. 9), while maintaining acceptable image quality (Fig. 1(d) and Sec. G)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 376, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 449 + ], + "type": "text", + "content": "Another LDM. SDMs are derived from LDMs [58], which share a similar U-Net design across many tasks. To validate the generality of our work, we compress an LDM (with 308M parameters and 410K training iterations) for unconditional generation on CelebA-HQ [59] by using the same approach of BK-SDM-Small (187M parameters and 30K iterations). Fig. 14 shows the efficacy of our architecture and distillation retraining." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 468, + 311, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 311, + 481 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 311, + 481 + ], + "type": "text", + "content": "6 Conclusion and Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "content": "We uncover the potential of architectural compression for general-purpose text-to-image synthesis with a renowned model, Stable Diffusion. Our block-removed lightweight models are effective for zero-shot generation, achieving competitive results against large-scale baselines. Distillation is a key of our method, leading to powerful retraining even under very constrained resources. Our work is orthogonal to previous works for efficient diffusion models, e.g., enabling fewer sampling steps, and can be readily combined with them. We hope our study can facilitate future research on structural compression of large diffusion models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 628 + ], + "type": "text", + "content": "Limitations. Our compact models inherit the capability of the source model for high-fidelity image generation, but they have shortcomings such as inaccurate generation of full-body human appearance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Negative Social Impacts. Recent large generative models are capable of creating high-quality plausible content. To avoid causing unintended social usage, researchers should take steps to ensure the appropriateness of the training data." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "B.-K. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "content": "We thank the Microsoft Startups Founders Hub program and the AI Industrial Convergence Cluster Development project funded by the Ministry of Science and ICT (MSIT, Korea) and Gwangju Metropolitan City for their generous support of GPU resources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 234, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 234, + 432, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 432, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 432, + 245 + ], + "type": "text", + "content": "1. A Simple and Effective Pruning Approach for LLMs. ICLR (2024) 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "text", + "content": "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 279, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 481, + 300 + ], + "type": "text", + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023) 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "type": "text", + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "text", + "content": "5. Castells, T., Song, H.K., Piao, T., Choi, S., Kim, B.K., Yim, H., Lee, C., Kim, J.G., Kim, T.H.: Edgefusion: On-device text-to-image generation. In: CVPR Workshop (2024) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "type": "text", + "content": "6. Chen, Y.H., Sarokin, R., Lee, J., Tang, J., Chang, C.L., Kulik, A., Grundmann, M.: Speed is all you need: On-device acceleration of large diffusion models viagpu-aware optimizations. In: CVPR Workshop (2023) 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "type": "text", + "content": "7. Choi, J., Kim, M., Ahn, D., Kim, T., Kim, Y., Jo, D., Jeon, H., Kim, J.J., Kim, H.: Squeezing large-scale diffusion models for mobile. In: ICML Workshop (2023) 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 412, + 481, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 481, + 433 + ], + "type": "text", + "content": "8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 434, + 481, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 454 + ], + "type": "text", + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021) 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 456, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 456, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 134, + 456, + 481, + 488 + ], + "type": "text", + "content": "10. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. In: NeurIPS (2021) 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 489, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 489, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 134, + 489, + 481, + 510 + ], + "type": "text", + "content": "11. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. In: NeurIPS (2022) 2, 8, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 511, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 511, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 134, + 511, + 481, + 532 + ], + "type": "text", + "content": "12. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 533, + 481, + 555 + ], + "type": "text", + "content": "13. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: CVPR (2023) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 555, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 481, + 576 + ], + "type": "text", + "content": "14. Fang, G., Ma, X., Wang, X.: Structural pruning for diffusion models. In: NeurIPS (2023) 4, 11, 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 577, + 481, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 481, + 608 + ], + "type": "text", + "content": "15. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV (2022) 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "type": "text", + "content": "16. Hao, Z., Guo, J., Jia, D., Han, K., Tang, Y., Zhang, C., Hu, H., Wang, Y.: Learning efficient vision transformers via fine-grained manifold distillation. In: NeurIPS (2022) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 90, + 447, + 102 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Heo, B., Kim, J., Yun, S., Park, H., Kwak, N., Choi, J.Y.: A comprehensive overhaul of feature distillation. In: ICCV (2019) 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 161 + ], + "type": "text", + "content": "19. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: EMNLP (2021) 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 480, + 194 + ], + "type": "text", + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "type": "text", + "content": "21. Hinton, G., Vinyls, O., Dean, J.: Distilling the knowledge in a neural network. In: NeurIPS Workshop (2014) 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 217, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 480, + 239 + ], + "type": "text", + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 240, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 261 + ], + "type": "text", + "content": "23. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop (2021) 2, 8, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "text", + "content": "24. Hou, J., Asghar, Z.: World's first on-device demonstration of stable diffusion on an android phone. https://wwwQUALCOMM.com/news (2023) 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "text", + "content": "25. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: ICML (2021) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 308, + 480, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 339 + ], + "type": "text", + "content": "26. Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., Wang, F., Liu, Q.: Tinybert: Distilling bert for natural language understanding. In: Findings of EMNLP (2020) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 341, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 480, + 363 + ], + "type": "text", + "content": "27. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "28. Kim, B.K., Choi, S., Park, H.: Cut inner layers: A structured pruning strategy for efficient u-net gans. In: ICML Workshop (2022) 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "29. Kim, B.K., Kim, G., Kim, T.H., Castells, T., Choi, S., Shin, J., Song, H.K.: Shortened llama: A simple depth pruning for large language models. arXiv preprint arXiv:2402.02834 (2024) 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 419, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 430 + ], + "type": "text", + "content": "30. LeCun, Y., Denker, J., Solla, S.: Optimal brain damage. In: NeurIPS (1989) 11, 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 431, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 464 + ], + "type": "text", + "content": "31. Lee, Y., Park, K., Cho, Y., Lee, Y.J., Hwang, S.J.: Koala: Self-attention matters in knowledge distillation of latent diffusion models for memory-efficient and fast image synthesis. arXiv preprint arXiv:2312.04005v1 (2023) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "text", + "content": "32. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. In: ICLR (2017) 11, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "text", + "content": "33. Li, M., Lin, J., Ding, Y., Liu, Z., Zhu, J.Y., Han, S.: Gan compression: Efficient architectures for interactive conditional gans. In: CVPR (2020) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "text", + "content": "34. Li, X., Liu, Y., Lian, L., Yang, H., Dong, Z., Kang, D., Zhang, S., Keutzer, K.: Q-diffusion: Quantizing diffusion models. In: ICCV (2023) 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "text", + "content": "35. Li, Y., Wang, H., Jin, Q., Hu, J., Chemerys, P., Fu, Y., Wang, Y., Tulyakov, S., Ren, J.: Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. In: NeurIPS (2023) 4, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 565, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 587 + ], + "type": "text", + "content": "36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dolkar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 588, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 609 + ], + "type": "text", + "content": "37. Liu, L., Ren, Y., Lin, Z., Zhao, Z.: Pseudo numerical methods for diffusion models on manifolds. In: ICLR (2022) 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "text", + "content": "38. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In: NeurIPS (2022) 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "39. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095 (2022) 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "B.-K. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "40. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 482, + 181 + ], + "type": "text", + "content": "41. Mangrulkar, S., Gugger, S., Debut, L., Belkada, Y., Paul, S.: Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft (2022) 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 181, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 481, + 203 + ], + "type": "text", + "content": "42. Meng, C., Gao, R., Kingma, D.P., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: NeurIPS Workshop (2022) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "type": "text", + "content": "43. Meng, C., Gao, R., Kingma, D.P., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR (2023) 2, 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 256 + ], + "type": "text", + "content": "44. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022) 3, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "text", + "content": "45. Mo, S., Cho, M., Shin, J.: Freeze the discriminator: a simple baseline for fine-tuning gans. In: CVPR Workshop (2020) 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "text", + "content": "46. Molchanov, P., Mallya, A., Tyree, S., Frosio, I., Kautz, J.: Importance estimation for neural network pruning. In: CVPR (2019) 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 331 + ], + "type": "text", + "content": "47. Murti, C., Narshana, T., Bhattacharyya, C.: TVSPrune - pruning non-discriminative filters via total variation separability of intermediate representations without fine tuning. In: ICLR (2023) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "48. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: ICML (2022) 3, 8, 9, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "text", + "content": "49. Park, W., Kim, D., Lu, Y., Cho, M.: Relational knowledge distillation. In: CVPR (2019) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 418 + ], + "type": "text", + "content": "50. Pernias, P., Rampas, D., Richter, M.L., Pal, C.J., Aubreville, M.: Wierstchen: An efficient architecture for large-scale text-to-image diffusion models. In: ICLR (2024) 4, 8, 9, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "text", + "content": "51. Pinkney, J.: Small stable diffusion. https://huggingface.co/OFA-Sys/small-stable-diffusion-v0 (2023) 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "52. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022) 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 525 + ], + "type": "text", + "content": "54. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "55. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) 3, 8, 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "text", + "content": "56. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) 2, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "text", + "content": "57. Ren, Y., Wu, J., Xiao, X., Yang, J.: Online multi-granularity distillation for gan compression. In: ICCV (2021) 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "text", + "content": "58. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 3, 6, 7, 8, 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "59. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: Ldm on celeba-hq. https://huggingface.co/CompVis/ldm-celebahq-256 (2022) 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "60. Rombach, R., Esser, P.: Stable diffusion v1-4. https://huggingface.co/CompVis/stable-diffusion-v1-4 (2022) 2, 3, 5, 7, 8, 9, 12, 13" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 117, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 482, + 138 + ], + "type": "text", + "content": "61. Rombach, R., Esser, P.: Stable diffusion v1-5. https://huggingface.co/runwayml/stable-diffusion-v1-5 (2022) 2, 3, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 482, + 160 + ], + "type": "text", + "content": "62. Rombach, R., Esser, P., Ha, D.: Stable diffusion v2-1-base. https://huggingface.co/stabilityai/stable-diffusion-2-1-base (2022) 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "text", + "content": "63. Romero, A., Ballas, N., Kahou, S.E., Chassang, A., Gatta, C., Bengio, Y.: Fitnets: Hints for thin deep nets. In: ICLR (2015) 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "64. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 236 + ], + "type": "text", + "content": "65. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023) 3, 7, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "type": "text", + "content": "66. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 3, 7, 8, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "67. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: NeurIPS (2016) 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "68. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2022) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "69. Sanh, V., Debut, L., Chaumont, J., Wolf, T.: Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In: NeurIPS Workshop (2019) 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "text", + "content": "70. Schuhmann, C., Beaumont, R.: Laion-aesthetics. https://laion.ai/blog/laion-aesthetics (2022) 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 481, + 413 + ], + "type": "text", + "content": "71. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS Workshop (2022) 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "text", + "content": "72. Segmind: Segmind-distill-sd. https://github.com/segmind/distill-sd/tree/c1e97a70d141df09e6fe5cc7dbd66e0cbae3eeb (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "73. Segmind: Ssd-1b. https://github.com/segmind/SSD-1B/tree/d2ff723ea8ecf5ddb86f3aac0af1db30e88a2e2d (2023) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 488 + ], + "type": "text", + "content": "74. Shen, H., Cheng, P., Ye, X., Cheng, W., Abidi, H.: Accelerate stable diffusion with intel neural compressor. https://medium.com/intel-analytics-software (2022) 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "text", + "content": "75. Shu, C., Liu, Y., Gao, J., Yan, Z., Shen, C.: Channel-wise knowledge distillation for dense prediction. In: ICCV (2021) 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "76. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "77. Sun, Z., Yu, H., Song, X., Liu, R., Yang, Y., Zhou, D.: Mobilebert: a compact task-agnostic bert for resource-limited devices. In: ACL (2020) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "78. Tang, R., Liu, L., Pandey, A., Jiang, Z., Yang, G., Kumar, K., Stenetorp, P., Lin, J., Ture, F.: What the DAAM: Interpreting stable diffusion using cross attention. In: ACL (2023) 10" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "79. Tao, M., Bao, B.K., Tang, H., Xu, C.: Galip: Generative adversarial clips for text-to-image synthesis. In: CVPR (2023) 8, 9, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "80. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers and distillation through attention. In: ICML (2021) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "81. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: NeurIPS (2017) 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "B.-K. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 358 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "82. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "83. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: CVPR (2023) 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "text", + "content": "84. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: CVPR (2023) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "85. Zagoruyko, S., Komodakis, N.: Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In: ICLR (2017) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "text", + "content": "86. Zhang, L., Chen, X., Tu, X., Wan, P., Xu, N., Ma, K.: Wavelet knowledge distillation: Towards efficient image-to-image translation. In: CVPR (2022) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "87. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "text", + "content": "88. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: ICLR (2023) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "type": "text", + "content": "89. Zhao, Y., Xu, Y., Xiao, Z., Hou, T.: Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567 (2023) 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "text", + "content": "90. Zhou, Y., Zhang, R., Chen, C., Li, C., Tensmeyer, C., Yu, T., Gu, J., Xu, J., Sun, T.: Towards language-free training for text-to-image generation. In: CVPR (2022) 8, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 335, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 358 + ], + "type": "text", + "content": "91. Zhu, L.: Thop: Pytorch-ocounter. https://github.com/Lyken17/pytorchOpCounter (2018) 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_content_list.json b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3ccd075a7f31ef4e044559f7b64873bc64ef2aed --- /dev/null +++ b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_content_list.json @@ -0,0 +1,1842 @@ +[ + { + "type": "text", + "text": "BKDSNN: Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zekai $\\mathrm{Xu}^{1}$ , Kang You $^{1}$ , Qinghai Guo $^{2}$ , Xiang Wang $^{2}$ , and Zhezhi He $^{1\\dagger}$", + "bbox": [ + 240, + 233, + 759, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China", + "bbox": [ + 225, + 261, + 776, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Huawei Technologies, Shenzhen, China", + "bbox": [ + 364, + 290, + 638, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sherlock.holmes.xu,kang_you, $^\\dagger$ zhezhi.he}@sjtu.edu.cn", + "bbox": [ + 300, + 305, + 702, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/Intelligent-Computing-Research-Group/BKDSNN", + "bbox": [ + 259, + 319, + 741, + 332 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Spiking neural networks (SNNs), which mimic biological neural systems to convey information via discrete spikes, are well-known as brain-inspired models with excellent computing efficiency. By utilizing the surrogate gradient estimation for discrete spikes, learning-based SNN training methods that can achieve ultra-low inference latency (i.e., number of time-step) have emerged recently. Nevertheless, due to the difficulty of deriving precise gradient for discrete spikes in learning-based methods, a distinct accuracy gap persists between SNNs and their artificial neural networks (ANNs) counterparts. To address the aforementioned issue, we propose a blurred knowledge distillation (BKD) technique, which leverages randomly blurred SNN features to restore and imitate the ANN features. Note that, our BKD is applied upon the feature map right before the last layer of SNNs, which can also mix with prior logits-based knowledge distillation for maximal accuracy boost. In the category of learning-based methods, our work achieves state-of-the-art performance for training SNNs on both static and neuromorphic datasets. On the ImageNet dataset, BKDSNN outperforms prior best results by $4.51\\%$ and $0.93\\%$ with the network topology of CNN and Transformer, respectively.", + "bbox": [ + 259, + 364, + 743, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Learning-based Spiking Neural Networks $\\cdot$ Knowledge Distillation", + "bbox": [ + 259, + 642, + 740, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 694, + 375, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the neural firing mechanism in the biological nervous system, spiking neural network (SNN) [37] emerges as an alternative framework for the traditional artificial neural network (ANN). ANNs rely on continuous activation values to convey information between neurons [31], while SNNs utilize discrete spiking events to process and propagate information [6,38]. Given the remarkable energy efficiency achieved by such event-driven information propagation [18], SNNs", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "show promising prospects on computational intelligence tasks [44] with strong autonomous learning capabilities and ultra-low power consumption [2,11,41,58]. Apart from these advantages, SNNs are also natively compatible with event-based neuromorphic datasets (e.g., CIFAR10-DVS) [14,61,62,65].", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, conversion-based and learning-based are two common methods to train SNNs. Conversion-based methods attempt to leverage information from ANNs by transferring the parameters of a pre-trained ANN to its SNN counterpart, which have achieved promising results in several tasks [2, 8, 33, 34]. However, event-triggered neurons convey limited information with ultra-low inference time-steps (e.g., 4 time-steps), leading to degraded performance of conversion-based methods [14, 65]. Consequently, conversion-based methods usually take a long inference time (e.g., 32 time-steps) to get comparable accuracy as the original ANN [3, 47] based on the equivalence of the ReLU activation and the firing rate of integrate-and-fire (IF) neuron [16]. In order to further improve the performance of SNNs with ultra-low inference time-steps, learning-based methods unfold SNN training over the inference time-steps and leverage back-propagation through time (known as BPTT) [52, 59] to train SNNs. Inspired by [27], several works [14, 35, 62] adopt surrogate gradient estimation as an implicit differentiation for direct training of SNNs. With the combination of BPTT and surrogate gradient estimation, learning-based methods exhibit advantages in processing event-based neuromorphic datasets (e.g., CIFAR10-DVS [22]) [14, 61, 62, 65]. Furthermore, direct training of CNN- [14, 23] and Transformer-based [61, 62, 65] SNNs also achieves promising performance with ultra-low time-steps on static datasets (e.g., CIFAR10 [28], CIFAR100 [28] and ImageNet [7]).", + "bbox": [ + 212, + 210, + 787, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite that numerous efforts have been made in learning-based methods, SNNs still need improvement compared with ANNs. [53, 62] claim that imprecise surrogate gradient estimation of spiking neurons leads to performance degradation. On the other hand, KDSNN [53] and LaSNN [21] introduce layer-wise knowledge distillation (KD) to train a student SNN with rich information from the ANN teacher, achieving promising performance on several classical vision datasets (MNIST, CIFAR10/100) with CNN-based models. However, according to our preliminary experimental results, both KDSNN and LaSNN struggle with more complex datasets (ImageNet) and Transformer-based models. We hypothesize that the direct distillation between discrete SNN features and continuous ANN features [21, 53] might be the source of the limitations.", + "bbox": [ + 212, + 518, + 787, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, to further alleviate performance degradation of learning-based SNNs, we introduce a blurred knowledge distillation (BKD) technique to leverage randomly blurred SNN features to restore and mimic ANN features. We apply BKD to the intermediate feature before the last layer of the student SNN and prove its superiority over layer-wise knowledge distillation in learning-based SNNs. Furthermore, we combine prior logits-based distillation with BKD, which provides mutual benefits for improving performance. We evaluate BKDSNN on both static datasets (i.e., CIFAR10, CIFAR100, and ImageNet) and a neuromorphic dataset (i.e., CIFAR10-DVS), where we achieve state-of-the-art (SOTA) performance with both CNN- and Transformer-based SNN models.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zekai Xu et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are threefold:", + "bbox": [ + 238, + 146, + 576, + 160 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a blurred knowledge distillation (BKD) technique to guide the training of student SNNs through feature imitation and demonstrate the advantages of BKD on learning-based SNNs.", + "- We apply BKD to the intermediate feature before the last layer and prove its superiority over previous layer-wise distillation. Combined with prior logits-based knowledge distillation, our BKDSNN maximizes accuracy boost on learning-based SNNs.", + "- We benchmark our method on both static datasets (i.e., CIFAR10, CIFAR100 and ImageNet) and a neuromorphic dataset (i.e., CIFAR10-DVS). Our method outperforms prior SOTA learning-based methods for both CNN-and Transformer-based models." + ], + "bbox": [ + 225, + 172, + 782, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Background and Related Work", + "text_level": 1, + "bbox": [ + 215, + 361, + 552, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spiking Neuron Model. Previous works [2,3,8,14,24,61,62,65] have widely used the classic integrate-and-fire (IF) neuron as the fundamental unit in SNNs. IF neuron integrates input currents through changes in its membrane potential, which is compared against the threshold for spike generation. The membrane potential is reset after emitting a spike. The dynamics of the IF neuron is depicted as follows:", + "bbox": [ + 214, + 393, + 782, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH [ t ] = V [ t - 1 ] + \\frac {1}{\\tau} \\left(I [ t ] - (V [ t - 1 ] - V _ {\\mathrm {r e s e t}})\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 494, + 666, + 523 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS [ t ] = \\Theta (H [ t ] - V _ {\\mathrm {t h}}) = \\left\\{ \\begin{array}{l l} 1, & H [ t ] \\geq V _ {\\mathrm {t h}} \\\\ 0, & H [ t ] < V _ {\\mathrm {t h}} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 526, + 784, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV [ t ] = H [ t ] (1 - S [ t ]) + V _ {\\text {r e s e t}} S [ t ]\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 568, + 620, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is the membrane time constant, $I[t]$ is the receiving current at time step $t$ and $\\Theta (\\cdot)$ is the heaviside function. The neuron accumulates $I[t]$ to change the membrane potential $H[t]$ at time step $t$ . Once $H[t]$ exceeds the firing threshold $V_{\\mathrm{th}}$ , the neuron generates a spike and resets the membrane potential to $V_{\\mathrm{reset}}$ .", + "bbox": [ + 214, + 597, + 782, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conversion-based Methods. To fully make use of the knowledge from pre-trained ANN, conversion-based methods are designed to transfer the learned parameters of ANNs into corresponding SNNs, without further training of SNNs. Previous works [17, 18] replace ReLU activations in ANNs with spiking neurons and add scaling operations like weight normalization and threshold balancing [10] to generate corresponding SNNs. Although several works [8, 17, 18, 34] have achieved nearly loss-less accuracy with VGG-16 and ResNet, the converted SNNs still require long inference time to match the original ANNs in precision. [45] claims that low firing rates caused by weight normalization result in increasing latency for the information reaching higher layers, which limits the practical application of the converted SNNs.", + "bbox": [ + 214, + 672, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Enhancing SNN with Blurred Knowledge Distillation", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9bd6ac2317b1c9e4bbfbdd0bd2e2823cffac267c74d3507665209de20b5e90ac.jpg", + "image_caption": [ + "Fig.1: Overview of training SNN with blurred knowledge distillation (BKD). The SNN is directly trained with back-propagation through time (BPTT) [52], where we utilize a blurred variant of KD (i.e., BKD) to achieve higher accuracy. BKD highlighted in yellow shaded region differs from prior SNN-oriented KD in three perspectives: 1) A blurred matrix $\\pmb{B}$ is randomly sampled on the fly (per input image) to mask out the feature of student SNN; 2) A restoration block $\\mathcal{G}$ consisting of two convolutional layers connected by a ReLU layer is applied on blurred SNN features to restore and mimic ANN features; 3) Such blurred knowledge distillation is applied only to the intermediate features before the last layer." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning-based Methods. With the development of surrogate gradient estimation [39], learning-based methods leverage continuous-value-based loss functions to handle nondifferentiable difficulties in SNN training. Ref. [39] introduces surrogate gradient estimation to conduct SNN training on the event-based DVS gesture dataset, which achieves better accuracy after a few training iterations. In an effort to improve computational resources in SNNs, joint CNN-SNN [54] puts forward a novel augmented spiking-based framework and achieves promising results in several image classification tasks. DCT-SNN [15] reduces the number of inference time-steps by combining frequency-domain learning with surrogate gradient descent. Spiking ResNet [60] uses the directly-trained deep SNN to achieve high performance on ImageNet. SEW ResNet [14] proposes the spike-element-wise block to counter gradient vanishing and gradient exploding, which successfully extends the depth of directly-trained SNN to more than 100 layers. When it comes to Transformer-based SNN, Spikformer [65] with a novel spiking self-attention block achieves significant improvement in image classification on both neuromorphic and static datasets. Transformer-based SNN is further improved by successive Spikingformer [61], ConvBN-MaxPooling-LIF (CML) [62] and Meta-SpikeFormer [57], achieving SOTA performance.", + "bbox": [ + 212, + 417, + 787, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowledge Distillation for SNN Training. Knowledge distillation (KD) is a classical transfer learning method that has been proven to be effective on various tasks, such as model compression [20, 36, 49, 56], incremental learning [4, 5, 42, 55], etc. Recently, several works [30, 32, 50] extend KD for SNN training, such as distilling spikes from a large SNN to a small SNN or logits-based knowledge from a well-trained ANN to a simple SNN. KDSNN [53] proposes a novel framework embracing both logits-based and feature-based knowledge distillation, which enhances performance on several image classification datasets (MNIST, CIFAR10, and CIFAR100). LaSNN [21] puts forward a layer-wise feature-based", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zekai Xu et al.", + "bbox": [ + 271, + 114, + 372, + 126 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7cda1c924515d3ca30a0ac07251e511b5d957ff6ae4f1f34fd98c3efd8bd4e39.jpg", + "table_caption": [ + "Table 1: Summary of notations in this paper" + ], + "table_footnote": [], + "table_body": "
NotationDescription
Ttime-step
λblurred ratio
R∈Rb×ctearandom matrix R~U(0,1)
B∈Rb×cteablurred matrix
ˆy∈Rb×nclsground truth label
ystu,ytea∈Rb×nclsoutput logits of student SNN and teacher ANN
qstu,qtea∈Rb×nclsystuand ytea after flattened processing
Ftea∈Rb×ctea×h×wfeature before the last layer of teacher ANN
Fstu∈RT×b×cstu×h×wfeature before the last layer of student SNN
Fstu∈Rb×cstu×h×waverage of Fstu at time-step dimension (T)
fadpadaptive layer to align Fstu with Ftea
Grestoration block for Fstu
ˆFstu∈Rb×ctea×h×wblurred restoration feature from Fstu
", + "bbox": [ + 307, + 170, + 696, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ANN-to-SNN distillation framework, achieving comparable top-1 accuracy to ANNs on the challenging Tiny ImageNet dataset. Furthermore, KDSNN [53] and LaSNN [21] are applicable to both homogeneous and heterogeneous networks. Despite the superiority of KDSNN [53] and LaSNN [21], they struggle with more complex datasets (e.g., ImageNet) with CNN-based and Transformer-based models according to our preliminary experimental results in Sec. 4.2.", + "bbox": [ + 212, + 375, + 787, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 493, + 330, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we propose a novel blurred knowledge distillation SNN (overall architecture of BKDSNN is in Fig. 1a) training method that leverages knowledge from the teacher ANN to guide the training of the student SNN. As depicted in Fig. 1b, our BKD first introduces a randomly blurred matrix $\\pmb{B}$ to mask the student SNN features. Unlike KDSNN [53] and LaSNN [21], which only apply an adaptive layer to align the SNN feature with the ANN feature before distillation, BKD introduces a restoration block $\\mathcal{G}$ on blurred SNN feature to restore and mimic ANN feature. The proposed BKD is applied only to the intermediate feature before the last layer, different from the layer-wise distillation [21, 53]. As shown in Fig. 1a, we combine the logits-based distillation with BKD to leverage both logits and feature knowledge from the teacher ANN.", + "bbox": [ + 212, + 527, + 787, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Blurred Knowledge Distillation. To facilitate understanding, we present important notations in Tab. 1. The proposed blurred knowledge distillation is illustrated in Fig. 1b. We first introduce a random matrix $\\pmb{R}$ which obeys uniform distribution between 0 and 1. Then we apply a blurred ratio $\\lambda$ to $\\pmb{R}$ to generate a blurred matrix $\\pmb{B}$ through Eq. (2).", + "bbox": [ + 212, + 714, + 787, + 789 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {B} _ {i, j} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\boldsymbol {R} _ {i, j} < \\lambda \\\\ 1, & \\text {O t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 801, + 785, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Enhancing SNN with Blurred Knowledge Distillation", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4925b5de32efb33a919fb07dc670420e409e93a90cb6338ae193e59a3c2937be.jpg", + "image_caption": [ + "(a) Feature visualization of ResNet-18." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 509, + 289 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6fb5a2b82e913d8ee5e3659413dd8637e02440c423fff0da061be68b47328ff3.jpg", + "image_caption": [ + "(b) Feature distribution of SNN and ANN.", + "Fig.2: (a) Comparison of feature visualization in ResNet-18 under various methods; (b) Histogram of SNN and ANN feature." + ], + "image_footnote": [], + "bbox": [ + 511, + 147, + 784, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For feature $\\mathbf{F}^{\\mathrm{stu}}$ at the last layer of student SNN, we first average $\\mathbf{F}^{\\mathrm{stu}}$ at timestep dimension $(T)$ as:", + "bbox": [ + 212, + 354, + 782, + 386 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {F}} ^ {\\mathrm {s t u}} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\mathbf {F} _ {i} ^ {\\mathrm {s t u}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 386, + 784, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After the averaging operation above, we apply a convolutional adaptive layer $f_{\\mathrm{adp}}$ to align $\\tilde{\\mathbf{F}}^{\\mathrm{stu}}$ with $\\mathsf{F}_i^{\\mathrm{tea}}$ when student feature dimension $c^{\\mathrm{stu}}$ mismatches with teacher feature dimension $c^{\\mathrm{tea}}$ . After the alignment, we construct a restoration block $\\mathcal{G}$ with two convolutional layers connected by a ReLU layer and apply it to $\\tilde{\\mathbf{F}}^{\\mathrm{stu}}$ with the blurred matrix $B$ to achieve blurred restoration feature $\\hat{\\mathbf{F}}^{\\mathrm{stu}}$ as illustrated in Eq. (4).", + "bbox": [ + 214, + 431, + 787, + 529 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} ^ {\\mathrm {s t u}} = \\left\\{ \\begin{array}{l l} \\mathcal {G} \\left(f _ {\\mathrm {a d p}} \\left(\\tilde {\\mathbf {F}} ^ {\\mathrm {s t u}}\\right) * \\boldsymbol {B}\\right), & c ^ {\\mathrm {s t u}} \\neq c ^ {\\mathrm {t e a}} \\\\ \\mathcal {G} \\left(\\tilde {\\mathbf {F}} ^ {\\mathrm {s t u}} * \\boldsymbol {B}\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 539, + 784, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We calculate blurred knowledge distillation loss $\\mathcal{L}_{\\mathrm{BKD}}$ using quadratic norm $(L^2)$ as follows.", + "bbox": [ + 214, + 599, + 782, + 627 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B K D}} = \\sum_ {i} ^ {b} \\sum_ {j} ^ {c ^ {\\mathrm {s t u}}} \\sum_ {k} ^ {h} \\sum_ {l} ^ {w} \\left(\\hat {\\mathbf {F}} _ {i, j, k, l} ^ {\\mathrm {s t u}} - \\mathbf {F} _ {i, j, k, l} ^ {\\mathrm {t e a}}\\right) ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 627, + 784, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then combine blurred knowledge distillation loss $\\mathcal{L}_{\\mathrm{BKD}}$ with the original cross entropy loss $\\mathcal{L}_{\\mathrm{CE}}$ to get the total loss $\\mathcal{L}_{\\mathrm{total}}$ as follows.", + "bbox": [ + 214, + 675, + 782, + 708 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {C E}} (\\boldsymbol {y} ^ {\\mathrm {s t u}}, \\hat {\\boldsymbol {y}}) + \\mathcal {L} _ {\\mathrm {B K D}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 718, + 784, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of $BKD$ . For both CNN- and Transformer-based models, features of deeper layers have better representation for the original input image [26,43,63]. In other words, features of deeper layers contain sufficient information of adjacent pixels (for CNN-based models) [26] or patch pairs (for Transformer-based models) [43, 63]. Therefore, we can use partial student SNN feature before the last layer to restore and mimic teacher ANN feature through a restoration block", + "bbox": [ + 214, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zekai Xu et al.", + "bbox": [ + 271, + 114, + 372, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{G}$ , which can help the student SNN achieves better representation and effectively mimics teacher ANN. We then apply Grad-CAM which can intuitively demonstrate the feature similarity [46] on each block in ResNet-18 under various methods. As depicted in Fig. 2a, compared with previous SNN (SEW ResNet-18 [14]) and KDSNN [53], our BKDSNN transfers sufficient knowledge from ANN and effectively mimics ANN feature. Meanwhile, as illustrated in Fig. 2b, student SNN feature $\\tilde{\\mathsf{F}}^{\\mathrm{stu}}$ is discrete while teacher ANN feature $\\mathsf{F}^{\\mathrm{tea}}$ is continuous. Therefore, directly applying distillation to encourage discrete SNN feature $\\tilde{\\mathsf{F}}^{\\mathrm{stu}}$ to imitate continuous ANN feature $\\mathsf{F}^{\\mathrm{tea}}$ [21, 53] is not plausible. However, the proposed BKD leverages the restoration block $\\mathcal{G}$ to map discrete feature $\\tilde{\\mathsf{F}}^{\\mathrm{stu}}$ to continuous feature $\\hat{\\mathsf{F}}^{\\mathrm{stu}}$ for imitating continuous ANN feature $\\mathsf{F}^{\\mathrm{tea}}$ , which is a more proper spiking-friendly feature-based distillation method.", + "bbox": [ + 214, + 146, + 787, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mixed Distillation. As depicted in Fig. 1a, BKDSNN further applies logits-based distillation which transfers knowledge from teacher ANN logits to student SNN logits. As soft logits from the teacher model embrace sufficient hidden knowledge, which serves as a good regularization to student model [64], we introduce the temperature $\\tau$ to make the logits distribution flatter. The output logits of both teacher and student are processed as follows:", + "bbox": [ + 214, + 356, + 787, + 446 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} _ {i} = \\frac {\\exp \\left(\\boldsymbol {y} _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {n ^ {\\mathrm {c l s}}} \\exp \\left(\\boldsymbol {y} _ {j} / \\tau\\right)}, \\quad i \\in \\{1, 2, \\dots , n ^ {\\mathrm {c l s}} \\} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 458, + 785, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then logits-based distillation loss $\\mathcal{L}_{\\mathrm{LD}}$ is defined as:", + "bbox": [ + 215, + 508, + 591, + 523 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L D}} = \\tau^ {2} \\cdot \\mathcal {L} _ {\\text {C r o s s E n t r o p y}} \\left(\\boldsymbol {q} ^ {\\text {s t u}}, \\boldsymbol {q} ^ {\\text {t e a}}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 535, + 785, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We combine $\\mathcal{L}_{\\mathrm{LD}}$ with $\\mathcal{L}_{\\mathrm{BKD}}$ to achieve mixed distillation loss $\\mathcal{L}_{\\mathrm{MD}}$ (Eq. (9)), which contains both logits and feature information from the teacher ANN.", + "bbox": [ + 214, + 565, + 782, + 594 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M D}} = w _ {\\mathrm {L D}} \\cdot \\mathcal {L} _ {\\mathrm {L D}} + w _ {\\mathrm {B K D}} \\cdot \\mathcal {L} _ {\\mathrm {B K D}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 608, + 785, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $w_{\\mathrm{LD}}$ and $w_{\\mathrm{BKD}}$ are weights for $\\mathcal{L}_{\\mathrm{LD}}$ and $\\mathcal{L}_{\\mathrm{BKD}}$ , respectively. Thus, the overall loss $\\mathcal{L}$ is defined as follows:", + "bbox": [ + 214, + 637, + 782, + 666 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {C r o s s E n t r o p y}} \\left(\\boldsymbol {y} ^ {\\text {s t u}}, \\hat {\\boldsymbol {y}}\\right) + \\mathcal {L} _ {\\mathrm {M D}} = \\mathcal {L} _ {\\text {t a s k}} + \\mathcal {L} _ {\\mathrm {M D}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 315, + 679, + 785, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 734, + 375, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental Configuration", + "text_level": 1, + "bbox": [ + 215, + 767, + 493, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Platform. All experiments are evaluated on a server equipped with 32 cores Intel Xeon Platinum 8352V CPU with $2.10\\mathrm{GHz}$ and 8-way NVIDIA GPUs. We use SpikingJelly [13] to simulate IF neurons.", + "bbox": [ + 214, + 794, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Enhancing SNN with Blurred Knowledge Distillation", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Static Datasets and Settings. We first evaluate BKDSNN on three static datasets, i.e., CIFAR10/100 and ImageNet.", + "bbox": [ + 215, + 145, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CIFAR10/CIFAR100 [28] contains 50k training and 10k testing images with $32 \\times 32$ resolution, and is labeled with 10- or 100-categories respectively for classification. For CNN-based models, we use ResNet-19 from TET [9] and lightweight ResNet-20 from SEW ResNet [14] with corresponding ANNs as teacher models. For Transformer-based models, we train ViT-S/16 on CIFAR10 and CIFAR100 to make it serve as the teacher ANN model for Sformer-4-256, Sformer-2-384, and Sformer-4-384 (We abbreviate Spikingformer to Sformer for convenience). Sformer-L-D represents a Sformer model with L Sformer encoder blocks and D feature embedding dimensions.", + "bbox": [ + 212, + 179, + 787, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ImageNet-1K [7] contains 1.3 millions of training images and 50k validation images with $224 \\times 224$ resolution, labeled with 1k classes. We choose ResNet-18, ResNet-34, and ResNet-50 pre-trained on ImageNet-1K as teacher ANN models for the corresponding SEW ResNet models [14]. For Transformer-based models, we introduce ViT-B/16 [12] pre-trained on ImageNet1K as the teacher ANN model for Sformer-8-384, Sformer-8-512 and Sformer-8-768. We also apply our BKDSNN on the latest Meta-SpikeFormer [57] and conduct experiments with Mformer-8-256, Mformer-8-384 and Mformer-8-512 on ImageNet-1K (We abbreviate Meta-SpikeFormer to Mformer for convenience.).", + "bbox": [ + 212, + 318, + 787, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For both CNN- and Transformer-based models, we set temperature $\\tau$ to 2.0, blurred ratio $\\lambda$ to 0.15 and logits-based distillation weight $w_{\\mathrm{LD}}$ to 1. Specifically for CNN-based models, we set blurred knowledge distillation weight $w_{\\mathrm{BKD}}$ to $7e^{-4}$ for CIFAR10/100 and $7e^{-5}$ for ImageNet. We follow all the training configurations (batch size, learning rate, optimizer, etc.) from SEW ResNet [14] except ResNet-19, where we use training configuration from TET [9]. For Transformer-based models, we set $w_{\\mathrm{BKD}}$ to $2e^{-3}$ for both CIFAR10/100 and ImageNet, and we follow all the training configurations (batch size, learning rate, optimizer, etc.) from Sformer+CML [62].", + "bbox": [ + 212, + 458, + 787, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Neuromorphic Dataset and Networks. We also evaluate BKDSNN on CIFAR10-DVS [22], which is a neuromorphic dataset that leverages dynamic vision sensor (DVS) to convert 10k frame-based images from the CIFAR10 dataset into 10k event streams, providing an event-stream dataset with 10 different classes. It includes 9k training samples and 1k testing samples. Note that, CIFAR10-DVS is an event-stream dataset tailored to SNN, so we use SNN models with 16 timesteps as teacher models to guide the training of SNN models with lower timesteps (i.e., 4 and 8). We utilize Wide-7B-Net [14] and Sformer-2-256 [62] with 4 and 8 inference time-steps as student models for CNN-based and Transformer-based models, respectively.", + "bbox": [ + 212, + 625, + 787, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For both Wide-7B-Net [14] and former-2-256 [62], we set $\\tau$ to 2.0, $\\lambda$ to 0.40 and $w_{\\mathrm{LD}}$ to 1. Then we set $w_{\\mathrm{BKD}}$ to $2e^{-3}$ for Wide-7B-Net [14] and $7e^{-5}$ for Sformer-2-256 [62]. We follow all the training configurations from SEW ResNet [14] for Wide-7B-Net [14], and Sformer+CML [62] for Sformer-2-256 [62].", + "bbox": [ + 212, + 777, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zekai Xu et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f50c13423d5fd3f2ec88ee40973f76e7bbb410f75078a27b0688e7d954803583.jpg", + "table_caption": [ + "Table 2: Comparison with previous works, using various networks and multiple datasets. $P$ denotes #parameters. $KD$ denotes knowledge distillation. $\\star$ : results taken from [14]. †: prior state-of-the-art conversion-based method QCFS [2]. The prior SOTA results are marked in gray, which is used to compare with our BKDSNN and report the accuracy difference. Best results are in bold. - indicates results not available (not reported in original article or cannot reproduce)." + ], + "table_footnote": [], + "table_body": "
w/o KD MethodArchitectureR-19 (P: 12.63M)R-20 (P: 0.27M)Time StepArchitecture MethodSformer-4-256 (P: 4.15M)Sformer-2-384 (P: 5.76M)Sformer-4-384 (P: 9.32M)Time Step
CF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 Acc
QCS [2]†--83.7555.234w/o KD Spíkarfer [65]93.9475.9694.8076.9595.1977.864
TET [9]94.4474.4787.7559.144+CML [62]94.8277.6495.6378.7595.9379.654
STBP [60]92.9270.86--4+CML [62]94.7777.4395.2278.3495.6179.094
Spíkarfer-R [23]92.1370.1387.3259.934
SEW-R [14]93.2470.8489.0760.164
Teacher ANN95.6079.7891.7768.401Teacher ANN96.7582.2296.7582.2296.7582.221
KDSNN [53]94.3674.0889.0360.184KDSNN [53]95.0078.3895.5979.2595.8880.334
LaSNN [21]94.2173.9288.4159.174LaSNN [21]94.9778.0295.5578.9195.7979.994
BKDSNN (ours)94.6474.9589.2960.924BKDSNN (ours)95.29 (+0.29)79.41 (+1.03)95.90 (+0.31)80.63 (+1.38)96.06 (+0.18)81.26 (+0.93)4
", + "bbox": [ + 222, + 232, + 787, + 351 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/04976da06a9a63d7477de263bede0aa8fa98d836e264c6ef675a219aa3080384.jpg", + "table_caption": [ + "(a) CNN-based SNN on $CF10 / 100$ dataset.", + "(b) Transformer-based SNN on $CF10/100$ dataset." + ], + "table_footnote": [], + "table_body": "
w/o KD MethodArchitectureR-18(P: 11.69M)Top1 AccR-34(P: 21.79M)Top1 AccR-50(P: 25.56M)Top1 AccTimeStep
QCFS [2]†-69.37-32
TET [9]-68.00-4
STBP [60]-67.04-4
Spiking-R [23]★62.3261.8657.664
SEW-R [14]63.1867.0467.784
with KD (ours)Teacher ANN69.7673.3176.311
KDSNN [53]63.4267.1867.724
LaSNN [21]63.3166.9467.814
BKDSNN(ours)65.60(+2.18)71.24(+4.06)72.32(+4.51)4
", + "bbox": [ + 225, + 377, + 482, + 500 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/de814e0dc9ec1d4aefe249892bf4d849b39f39b3ef7a860245d36c1182eefc0c.jpg", + "table_caption": [ + "(c) CNN-based SNN on ImageNet dataset." + ], + "table_footnote": [], + "table_body": "
Architecture MethodSformer-8-384(P: 16.81M)Top1 Acc.Sformer-8-512(P: 29.68M)Top1 Acc.Sformer-8-768(P: 66.34M)Top1 Acc.Time Step
w/o KDSpikformer [65]70.2473.3874.814
+CML [62]72.7375.6177.344
Sformer [61]72.4574.7975.854
+CML [62]74.3576.5477.644
with KDTeacher ANN81.7881.7881.781
KDSNN [53]74.6276.4477.834
LaSNN [21]73.8576.3877.664
BKDSNN(ours)75.48(+0.86)77.24(+0.70)79.93(+2.10)4
", + "bbox": [ + 491, + 377, + 787, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(d) Transformer-based SNN on ImageNet dataset.", + "bbox": [ + 495, + 501, + 784, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Settings for Prior KD Methods on SNNs. Furthermore, to compare BKDSNN with previous KD methods on SNNs (i.e., KDSNN [53] and LaSNN [21]), we follow the settings from [53] and [21] to reproduce the experimental results on all datasets with CNN-based and Transformer-based models.", + "bbox": [ + 212, + 527, + 787, + 587 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparison on Static Datasets", + "text_level": 1, + "bbox": [ + 214, + 614, + 517, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison on CIFAR10/100 between BKDSNN and previous methods are listed in Tab. 2a and Tab. 2b, where BKDSNN surpasses prior SOTA methods with both CNN- and Transformer-based models. Specifically, BKD applied upon Sformer-4-384 outperforms previous SOTA by $0.18\\%$ on CIFAR10 and $0.93\\%$ on CIFAR100, which achieves $96.06\\%$ and $81.26\\%$ accuracy respectively.", + "bbox": [ + 212, + 642, + 787, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Intuitively, it is difficult for both KDSNN [53] and LaSNN [21] to improve the performance of CNN-based models on the CIFAR dataset. For ResNet-19, the knowledge of the ANN teacher is limited for the SNN student, leading to the above difficulties. When it comes to ResNet-20, the lightweight model conveys limited information which results in the ineffectiveness of previous knowledge distillation methods. In contrast, our BKDSNN is effective on both limited teacher knowledge as well as lightweight student models, which verifies the superiority of our method.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Enhancing SNN with Blurred Knowledge Distillation", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7b35e568c8f534962dd5e0d72abd877d7d76dc672d21410cab022d6e11f896bc.jpg", + "table_caption": [ + "Table 3: (a) Comparison with Meta-SpikeFormer [57] on ImageNet. (b) Experimental results on CIFAR10-DVS. Mformer represents Meta-SpikeFormer [57]. $\\star$ : SNN with high inference time-step (i.e., 16) serves as the teacher model for corresponding SNN with low inference time-step (i.e., 4 and 8). The prior state-of-the-art results are marked in gray. Best results are in bold." + ], + "table_footnote": [], + "table_body": "
Architecture MethodMformer-8-256(P: 15.11M)Top1 Acc.Mformer-8-384(P: 31.33M)Top1 Acc.Mformer-8-512(P: 55.35M)Top1 Acc.Time StepTransformer-basedCNN-based
MethodAccTime StepAccTime StepMethodAccTime StepAccTime Step
wDMformer [57]71.8275.4077.961Sfomer-2-256+CML [62]★81.41681.416Wide-7B-Net[14]★74.41674.416
74.1177.2078.814
Teacher ANN81.7881.7881.781Sfomer-2-256+CML [62]76.4479.28Wide-7B-Net[14]64.8470.28
with KDMformer [57]--79.111KDSNN [53]78.1479.98KDSNN [53]66.1470.98
--80.004LaSNN [21]78.3480.18LaSNN [21]66.6470.58
BKDSNN(ours)72.9376.1079.821BKDSNN(ours)79.3(+1.0)480.8(+0.7)8BKDSNN(ours)68.3(+1.7)472.2(+1.3)8
", + "bbox": [ + 217, + 217, + 785, + 314 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7768b11dc02747ab46c270075a11e43623bda4ab48ca2722b155caf2f5db2d24.jpg", + "image_caption": [ + "(a) Meta-SpikeFormer on ImageNet dataset. (b) Experimental results on CIFAR10-DVS.", + "(a) Feature map visualization.", + "Fig.3: (a) Feature map visualization of different methods on SEW ResNet-18 (CNN) and Sformer-8-384+CML (Transformer). Input images are sampled from the ImageNet validation dataset. Shaded colors from blue to red indicate the impacts of the regions on the classification scores from low to high. (b) Top1 accuracy versus GPU hours on ImageNet with SEW ResNet-50 and Spikingformer-8-768." + ], + "image_footnote": [], + "bbox": [ + 215, + 327, + 785, + 501 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison on ImageNet between BKDSNN and previous methods are tabulated in Tab. 2c, Tab. 2d and Tab. 3a. As anticipated, BKDSNN outperforms previous SOTA methods with both CNN- and Transformer-based models on ImageNet. BKDSNN achieves $72.32\\%$ top1 accuracy with SEW ResNet-50 and $80.93\\%$ with Mformer-8-512, with $4.51\\%$ and $0.93\\%$ improvement, respectively.", + "bbox": [ + 212, + 593, + 787, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As reported in Tab. 2c, Tab. 2d and Tab. 3a, ANN outperforms previous learning-based SNN by a large margin, which will serve as the teacher model to provide adequate knowledge for the student SNN. Based on our experiments, the accuracy improvement of SNN trained by KDSNN [53] and LaSNN [21] is limited, due to the inappropriate distillation between student and teacher models. Nevertheless, thanks to the blurred student features that effectively mimic and restore the teacher features, BKDSNN achieves promising results on the large-scale dataset (i.e. ImageNet).", + "bbox": [ + 212, + 669, + 787, + 790 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Feature Visualization. To further interpret BKDSNN, we apply Grad-CAM [46] to visualize the intermediate feature before the last layer using different methods.", + "bbox": [ + 212, + 808, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zekai Xu et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In $2^{\\mathrm{nd}}$ (ANN) to $3^{\\mathrm{rd}}$ (SNN) column of Fig. 3a, ANN properly focuses on the regions with vital information (highlighted in red), while learning-based SNN is distracted by the background area. Such erroneous attention of SNN ( $3^{\\mathrm{rd}}$ column in Fig. 3a) does not get calibrated by the previous KD methods, i.e., KDSNN [53] ( $4^{\\mathrm{th}}$ column) and LaSNN [21] ( $5^{\\mathrm{th}}$ column). However, the attention (shaded) region of our BKDSNN ( $6^{\\mathrm{th}}$ column) is quite similar to the pre-trained ANN, which demonstrates the effectiveness of BKD in learning the knowledge of ANN by mimicking and restoring teacher ANN features with blurred student SNN features.", + "bbox": [ + 212, + 145, + 787, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison of Training Cost. We plot the evolution curves of validation top1 accuracy during GPU hours as depicted in Fig. 3b. Without applying knowledge distillation (denoted as w/o KD in blue lines), the training of SEW ResNet-50 and Spikingformer-8-768 consume the least GPU hours. With the introduction of layer-wise distillation (LaSNN [21] in green lines and KDSNN [53] in yellow lines), the training of SEW ResNet-50 and Spikingformer-8-768 consume more hours while the improvement is limited. Compared with the layer-wise distillation methods above, our single-layer BKDSNN (in red lines) saves GPU hours and boosts the performance significantly.", + "bbox": [ + 212, + 296, + 787, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparison on Neuromorphic Dataset", + "text_level": 1, + "bbox": [ + 214, + 453, + 581, + 469 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also evaluate BKDSNN on CIFAR10-DVS with results reported in Tab. 3b. Specifically, BKDSNN outperforms previous SOTA results by $1.0\\%$ on models with 4 time-steps and $0.7\\%$ on models with 8 time-steps. Note that, the improvement of BKDSNN on Wide-7B-Net [14] is more pronounced than that on Sformer-2-256 [62], which is probably due to the fact that Wide-7B-Net [14] with 16 inference time-steps provides more knowledge for the corresponding student model with lower time-steps, compared to Sformer-2-256 [62].", + "bbox": [ + 212, + 477, + 787, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 604, + 388, + 619 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation on multiple losses $(\\mathcal{L}_{LD},\\mathcal{L}_{BKD}$ and $\\mathcal{L}_{MD})$ . We test different settings of BKDSNN, including versions without any knowledge distillation (w/o KD), with logit-based KD $(+{\\mathcal{L}}_{\\mathrm{LD}})$ , with BKD $(+{\\mathcal{L}}_{\\mathrm{BKD}})$ , with mixed distillation (logitbased $^+$ BKD, $+{\\mathcal{L}}_{\\mathrm{MD}}$ ). Results are tabulated in Tab. 4b. Compared to logitbased KD, applying BKD to the input feature of the last layer shows on-par accuracy improvement on both CNN- and Transformer-based SNNs. Furthermore, thanks to the compatibility between the blurred feature-based KD and logit-based KD, their mixed version shows the maximal accuracy improvement.", + "bbox": [ + 212, + 628, + 787, + 750 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition to the accuracy obtained for SEW ResNet-18 and Sformer-8-384 discussed above, we also plot their evolution curves of validation accuracy during training as in Fig. 4. We observe that with the introduction of logits-based distillation $\\mathcal{L}_{\\mathrm{LD}}$ (yellow curve) and feature-based blurred knowledge distillation $\\mathcal{L}_{\\mathrm{BKD}}$ (green curve), the performance is significantly improved compared to the vanilla counterpart (blue curve). Moreover, as anticipated, the mixed distillation", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Enhancing SNN with Blurred Knowledge Distillation", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/84461836aba2f02b603d0a2fd319dd16f2bb2eb08c7ae1d81093affe16ba43be.jpg", + "table_caption": [ + "Table 4: Ablation study with multiple settings. $R$ , $S$ , and $KD$ denotes ResNet, Sformer and Knowledge Distillation respectively. $S-4-384-400\\mathrm{E}$ refers to Sformer-4-384 after 400 epochs training. $\\star$ : results from [14], $\\diamond$ : results from [62]. Best results are in bold, runner-up results are marked in gray." + ], + "table_footnote": [], + "table_body": "
MethodConfigS-4-384-400ER-19
τλCF10 AccCF100 AccCF10 AccCF100 Acc
+Ld1.5-96.1381.4094.5074.66
2.0-96.1781.2594.5374.57
4.0-96.3180.8894.5674.52
+BKD-0.1596.1381.3894.5274.76
-0.4096.2681.4394.5674.83
-0.6096.3181.5394.6174.89
+MD2.00.1596.1881.6394.6474.95
2.00.4096.2081.6794.6874.98
2.00.6096.37 (+0.11)81.86 (+0.19)94.76 (+0.08)75.07 (+0.09)
", + "bbox": [ + 220, + 214, + 477, + 359 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(a) Ablation of $\\tau$ and $\\lambda$ on CIFAR.", + "bbox": [ + 243, + 362, + 449, + 375 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9ac73f7c110b69099a63b6e1f134593689997fda3ccb77d3475b260c56daf201.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCNN-basedTransformer-based
Student SNNTeacher AccSNN AccStudent SNNTeacher AccSNN Acc
w/o KDR-18*-63.18S-8-384-74.35
R-34*-67.04S-8-512-76.54
R-50*-67.78S-8-768-77.64
+LDR-1869.7665.36(+2.18)S-8-38481.7874.69(+0.24)
R-3473.3169.90(+2.86)S-8-51276.99(+0.45)
R-5076.1371.78(+4.00)S-8-76879.04(+1.40)
+BKDR-1869.7665.16(+1.98)S-8-38481.7875.08(+0.73)
R-3473.3170.23(+3.19)S-8-51277.29(+0.75)
R-5076.1371.24(+3.46)S-8-76879.64(+2.00)
+MDR-1869.7665.60(+2.42)S-8-38481.7875.48(+1.13)
R-3473.3171.24(+4.20)S-8-51277.24(+0.70)
R-5076.1372.32(+4.54)S-8-76879.93(+2.29)
", + "bbox": [ + 486, + 214, + 782, + 359 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f9d13f25f3ad7473ca473049f5ad61621df36c0fe4203d97ba4cd32dfdc924b7.jpg", + "table_caption": [ + "(b) Ablation of multiple losses on ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodSEW R-18 Top1 AccS-8-384 Top1 AccKD location
LaSNN [21]63.3173.85layer-wise
+LBKD63.8274.47layer-wise
KDSNN [53]63.4274.62layer-wise
+LBKD63.9474.93layer-wise
KDSNN [53]63.8774.96single-layer
BKDSNN (ours)65.60 (+1.66)75.48 (+0.52)single-layer
", + "bbox": [ + 220, + 378, + 426, + 477 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0f567ce202c7afe9801035d214a948550c7335b866e0087304b661fde50bcb0f.jpg", + "table_caption": [ + "(c) Ablation of BKD on ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodλSpikingformer-4-384-400EResNet-19
CIFAR10CIFAR100CIFAR10CIFAR100
AccAccAccAcc
w Gw/o Gw Gw/o Gw Gw/o Gw Gw/o G
+LBKD0.0095.3894.3880.4476.6494.4793.4274.6870.97
0.1596.1394.6081.3876.8194.5293.5874.7671.22
0.4096.2694.8281.4377.1494.5693.6474.8371.29
0.6096.3194.8981.5377.1694.6193.7274.8971.37
", + "bbox": [ + 436, + 378, + 782, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(d) Ablation of $\\lambda$ and $\\mathcal{G}$ on CIFAR.", + "bbox": [ + 503, + 479, + 710, + 489 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(red curve) in KDSNN leads to faster convergence with higher validation accuracy during the training. These experimental results highlight the important role of logits-based and feature-based knowledge from the teacher ANN in improving the performance of student SNNs.", + "bbox": [ + 212, + 500, + 787, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation of BKD and its location We have reported the KDSNN accuracy under different settings in Tab. 4b where KD was applied to logits, the last layer feature, or both. Two more questions still remain unanswered: 1) whether BKD outperforms the vanilla implementation? 2) Is applying KD/BKD to the last layer better than the layer-wise implementation?", + "bbox": [ + 212, + 580, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To answer these questions, we perform experiments whose results are reported in Tab. 4c. LaSNN [21] and KDSNN [53] are two recent works of layer-wise implementation of KD for SNN training. We replace the original feature-based distillation loss in LaSNN and KDSNN with BKD loss $\\mathcal{L}_{\\mathrm{BKD}}$ . In LaSNN, the replacement improves accuracy by $0.51\\%$ on SEW ResNet-18 and by $0.62\\%$ on Sformer-8-384. In KDSNN, the replacement improves accuracy by $0.52\\%$ on SEW ResNet-18 and by $0.31\\%$ on Sformer-8-384. These experiments confirm the advantages of the blurred implementation of KD. Moreover, we further test layer-wise (every layer feature) or single-layer (intermediate feature before the last layer) implementations of KD. We observe similar accuracy boosts for the two implementations. Thus, we conclude that applying BKD only to the intermediate feature map before the last layer is the adequate setup.", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zekai Xu et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4a27c357dba9210983d0e08554470aab891a49ac80cbd24454339def593cb341.jpg", + "image_caption": [ + "(a) SEW ResNet-18" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 405, + 238 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d05f73ad5df0d98ed2cc07fa4eda4f0ae980aa61b0a49ac65a315a0ffd04c0ca.jpg", + "image_caption": [ + "(b) SEW ResNet-34" + ], + "image_footnote": [], + "bbox": [ + 406, + 145, + 594, + 238 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f6726c3ec05329b99cb19edb09c94685fa9941cc5b72c1e0d82657e1456fd5d0.jpg", + "image_caption": [ + "(c) SEW ResNet-50" + ], + "image_footnote": [], + "bbox": [ + 596, + 145, + 787, + 238 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/51a6c32e82dbf0d63276c7af2ed063f77bdc120843ef6174cce873dbff01af8c.jpg", + "image_caption": [ + "(d) Sformer-8-384" + ], + "image_footnote": [], + "bbox": [ + 217, + 253, + 406, + 347 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e77baedc6659d2c1109995edc3480a43c647928c12e951e99e40fe4c46214845.jpg", + "image_caption": [ + "(e) Sformer-8-512" + ], + "image_footnote": [], + "bbox": [ + 406, + 253, + 594, + 347 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/634e91bd9d0ad6119c9d37866bc6c1e6bb9c7012739d37177316595dcd73114e.jpg", + "image_caption": [ + "(f) Sformer-8-768", + "Fig. 4: Evolution curve of validation accuracy during training, with CNN- (top 3 sub-figures) and Transformer-based (bottom 3 sub-figures) SNNs on ImageNet. The labels in the legends are the same as in Tab. 4b." + ], + "image_footnote": [], + "bbox": [ + 596, + 253, + 785, + 347 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of temperature coefficient, blurred ratio and restoration block. To examine the optimal setting of hyper-parameters, i.e., temperature $\\tau$ in Eqs. (7) and (8), blurred ration $\\lambda$ in Eq. (2) and restoration block $\\mathcal{G}$ in Eq. (4), we sweep $\\tau$ , $\\lambda$ and $\\mathcal{G}$ with ResNet-19 and Sformer-4-384, under varying BKDSNN settings. The results of the CIFAR-10 and CIFAR-100 datasets are reported in Tab. 4a and Tab. 4d. Note that we increase the training epochs of Sformer-4-384 to 400 for more distinct results. For logits-based distillation $\\mathcal{L}_{\\mathrm{LD}}$ , we find that flatter logits distribution (with higher $\\tau$ ) is better for classification with smaller categories (CIFAR10), while sharper logits distribution (with lower $\\tau$ ) is more effective for classification with larger categories (CIFAR100). For BKD $\\mathcal{L}_{\\mathrm{BKD}}$ and mixed distillation $\\mathcal{L}_{\\mathrm{MD}}$ , training with a higher blurred ratio $\\lambda$ (i.e., 0.60) achieves greater accuracy across various datasets and models. As reported in Tab. 4d, with the introduction of blurred ration $\\lambda$ and restoration block $\\mathcal{G}$ , student SNN better mimics teacher ANN and conveys more sufficient information, which enhances performance on CIFAR-10 and CIFAR-100 with both CNN- and Transformer-based models.", + "bbox": [ + 212, + 407, + 787, + 650 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Firing Rate Analysis", + "text_level": 1, + "bbox": [ + 215, + 671, + 433, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Firing Rate. We compare the firing rate $r$ of the intermediate feature before last layer on CIFAR with Sformer-4-384 and ResNet-19 under multiple settings. The firing rate $r$ is defined as:", + "bbox": [ + 212, + 695, + 787, + 742 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\nr = \\frac {1}{b c ^ {\\mathrm {s t u}} h w} \\sum_ {i} ^ {b} \\sum_ {j} ^ {c ^ {\\mathrm {s t u}}} \\sum_ {k} ^ {h} \\sum_ {l} ^ {w} \\tilde {\\mathbf {F}} _ {i, j, k, l} ^ {\\mathrm {s t u}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 752, + 785, + 795 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As reported in Tab. 5a, it reveals that on both CNN- and Transformer-based models, setting a higher firing rate $r$ achieves higher accuracy. Note that, with the introduction of $\\mathcal{L}_{\\mathrm{BKD}}$ and $\\mathcal{L}_{\\mathrm{MD}}$ , high blurred ratio $\\lambda$ encourages model to", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Enhancing SNN with Blurred Knowledge Distillation", + "bbox": [ + 377, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/46ff6d7ad4aaf974511f88c026d019b0211f34857ff523dcdfe0e31e701f4563.jpg", + "table_caption": [ + "Table 5: Firing rate $r$ under multiple settings on CIFAR and Energy consumption of ANN and SNN on ImageNet. $T$ denotes Transformer." + ], + "table_footnote": [ + "(a) Firing rate $r$ on CIFAR.", + "(b) Energy consumption on ImageNet." + ], + "table_body": "
MethodConfigτλS-4-384-400ER-19TimeStepSOPS(G)↓Energy Consumption (mJ)(↓)Top1Acc (↑)
CF10r (e-2)CF10AccCF10e(10)e-2CF10AccCF10e(10)e-2CF10AccCF10e(10)e-2CF10Acc
+LLD1.5-5.5596.131.5281.403.9494.500.9374.66R-50 [19]14.12 (FLOPS)18.9576.13
2.0-5.7296.171.3281.254.0394.531.0174.57Spiking R-50 [14]45.31 (28.88%↑)5.44 (71.29%↓)57.66 (-13.47)
4.0-6.0796.311.1880.884.1194.561.0874.52SEW R-50 [14]44.83 (17.23%↑)4.89 (74.20%↓)67.78 (-8.35)
+LBDK-0.154.6796.130.9781.383.8294.520.8274.76BKDSNN (ours)44.81 (16.75%↑)4.85 (74.40%↓)72.32 (-3.81)
-0.405.3596.261.0381.433.9494.560.8974.83
-0.605.5596.311.1481.534.0894.611.0274.89
+LMD2.0 0.155.3996.181.1481.633.9294.641.0374.95T-8-768 [12]118.60 (FLOPS)85.5681.16
2.0 0.406.1796.201.3481.673.9994.681.0774.98S-8-768 [61]412.54 (32.58%↓)13.68 (84.01%↓)75.85 (-5.31)
2.0 0.606.6696.371.5581.864.1894.761.1375.07S-8-768 + CML [62]412.30 (33.87%↓)13.42 (84.32%↓)77.64 (-3.52)
BKDSNN (ours)412.60 (32.26%↓)13.72 (83.96%↓)79.93 (-1.23)
", + "bbox": [ + 217, + 174, + 785, + 273 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "fire more spikes. Furthermore, compared with $\\mathcal{L}_{\\mathrm{LD}}$ , our $\\mathcal{L}_{\\mathrm{BKD}}$ reduces the firing rate while improving the performance, which assists the energy consumption reduction. Meanwhile, in virtue of $\\mathcal{L}_{\\mathrm{MD}}$ , BKDSNN significantly improves the performance with a slight boost on firing rate, which leads to state-of-the-art.", + "bbox": [ + 212, + 291, + 785, + 353 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.6 Energy Consumption Analysis", + "text_level": 1, + "bbox": [ + 215, + 371, + 511, + 387 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As tabulated in Tab. 5b, we calculate the energy consumption of BKDSNN according to the energy model in [25,29,65] and compare it with its ANN counterparts and SNNs trained by previous learning-based methods. In terms of the OPS, SNN variants are higher than ANNs, while ANNs use FLOPS as the operation counting metric. Nevertheless, since the energy per synaptic operation is more efficient than the operation in ANN. For ResNet-50, BKDSNN consumes the lowest energy while optimally suppressing accuracy loss. For Sformer-8-768, our BKDSNN consumes more energy to increase firing rates for maintaining similar performance as SOTA methods for SNN training.", + "bbox": [ + 212, + 393, + 787, + 530 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 550, + 359, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "BKDSNN constructs a knowledge distillation based framework for learning-based SNN training, achieving promising performance on both static and neuromorphic datasets supported by comprehensive experiments. We expect that BKDSNN will contribute to future investigations on brain-inspired SNN, which can further narrow the performance gap between SNN and ANN. In this work, we mainly focus on learning-based SNN training methods due to their remarkable performance under ultra-low inference time-step. We anticipate to extend BKDSNN on conversion-based methods, which is expected to overcome the existing bottlenecks of conversion-based methods under ultra-low inference time-step.", + "bbox": [ + 212, + 578, + 787, + 715 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 734, + 403, + 753 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is partially supported by National Key R&D Program of China (2022YFB4500200), National Natural Science Foundation of China (No. 62102257), Biren Technology-Shanghai Jiao Tong University Joint Laboratory Open Research Fund, Microsoft Research Asia Gift Fund, Lingang Laboratory Open Research Fund (No.LGQS-202202-11).", + "bbox": [ + 212, + 763, + 802, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zekai Xu et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Amir, A., Taba, B., Berg, D., Melano, T., McKinstry, J., Di Nolfo, C., Nayak, T., Andreopoulos, A., Garreau, G., Mendoza, M., et al.: A low power, fully event-based gesture recognition system. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7243-7252 (2017) 6", + "2. Bu, T., Fang, W., Ding, J., Dai, P., Yu, Z., Huang, T.: Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347 (2023) 2, 3, 9, 1", + "3. Cao, Y., Chen, Y., Khosla, D.: Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision 113, 54-66 (2015) 2, 3", + "4. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9233-9242 (2020) 4", + "5. Chen, L., Yu, C., Chen, L.: A new knowledge distillation for incremental object detection. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2019) 4", + "6. Davies, M., Srinivasa, N., Lin, T.H., Chinya, G., Cao, Y., Choday, S.H., Dimou, G., Joshi, P., Imam, N., Jain, S., et al.: Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro 38(1), 82-99 (2018) 1", + "7. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 2, 8", + "8. Deng, S., Gu, S.: Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476 (2021) 2, 3", + "9. Deng, S., Li, Y., Zhang, S., Gu, S.: Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946 (2022) 8, 9, 4", + "0. Diehl, P., Neil, D., Binas, J., Cook, M., Liu, S.C., Pfeiffer, M.: Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing (07 2015). https://doi.org/10.1109/IJCNN.2015.7280696 3, 1", + "1. Ding, J., Bu, T., Yu, Z., Huang, T., Liu, J.: Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training. Advances in Neural Information Processing Systems 35, 24780-24793 (2022) 2", + "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8, 14", + "3. Fang, W., Chen, Y., Ding, J., Yu, Z., Masquelier, T., Chen, D., Huang, L., Zhou, H., Li, G., Tian, Y.: Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances 9(40), eadi1480 (2023). https://doi.org/10.1126/sciadv.adi1480, https://www.science.org/doi/abs/10.1126/sciadv.adi1480 7", + "4. Fang, W., Yu, Z., Chen, Y., Huang, T., Masquelier, T., Tian, Y.: Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems 34, 21056-21069 (2021) 2, 3, 4, 7, 8, 9, 10, 11, 12, 14", + "5. Garg, I., Chowdhury, S.S., Roy, K.: Dct-snn: Using dct to distribute spatial information over time for low-latency spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4671-4680 (2021) 4" + ], + "bbox": [ + 225, + 172, + 785, + 837 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Enhancing SNN with Blurred Knowledge Distillation", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Gerstner, W., Kistler, W.M.: Spiking neuron models: Single neurons, populations, plasticity. Cambridge university press (2002) 2", + "17. Han, B., Roy, K.: Deep spiking neural network: Energy efficiency through time based coding. In: European Conference on Computer Vision. pp. 388-404. Springer (2020) 3", + "18. Han, B., Srinivasan, G., Roy, K.: Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13558-13567 (2020) 1, 3", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 14", + "20. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 4", + "21. Hong, D., Shen, J., Qi, Y., Wang, Y.: Lasnn: Layer-wise ann-to-snn distillation for effective and efficient training in deep spiking neural networks. arXiv preprint arXiv:2304.09101 (2023) 2, 4, 5, 7, 9, 10, 11, 12", + "22. Hongmin, L., Hanchao, L., Xiangyang, J., Guoqi, L., Luping, S.: Cifar10-dvs: An event-stream dataset for object classification. Frontiers in Neuroscience 11 (2017) 2, 8", + "23. Hu, Y., Tang, H., Pan, G.: Spiking deep residual network (2020) 2, 9", + "24. Hu, Y., Zheng, Q., Jiang, X., Pan, G.: Fast-snn: Fast spiking neural network by converting quantized ann. arXiv preprint arXiv:2305.19868 (2023) 3", + "25. Hu, Y., Deng, L., Wu, Y., Yao, M., Li, G.: Advancing spiking neural networks towards deep residual learning (2023) 14, 5", + "26. Khan, A., Sohail, A., Zahoora, U., Qureshi, A.S.: A survey of the recent architectures of deep convolutional neural networks. Artificial intelligence review 53, 5455-5516 (2020) 6", + "27. Kheradpisheh, S.R., Ganjtabesh, M., Thorpe, S.J., Masquelier, T.: Stdp-based spiking deep convolutional neural networks for object recognition. Neural Networks 99, 56-67 (2018) 2", + "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 2, 8", + "29. Kundu, S., Pedram, M., Beerel, P.A.: Hire-snn: Harnessing the inherent robustness of energy-efficient deep spiking neural networks by training with crafted input noise (2021) 14, 5", + "30. Kushawaha, R.K., Kumar, S., Banerjee, B., Velmurugan, R.: Distilling spikes: Knowledge distillation in spiking neural networks. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 4536-4543. IEEE (2021) 4", + "31. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015) 1", + "32. Lee, D., Park, S., Kim, J., Doh, W., Yoon, S.: Energy-efficient knowledge distillation for spiking neural networks. arXiv preprint arXiv:2106.07172 (2021) 4", + "33. Li, C., Ma, L., Furber, S.: Quantization framework for fast spiking neural networks. Frontiers in Neuroscience 16, 918793 (2022) 2, 1", + "34. Li, Y., Deng, S., Dong, X., Gong, R., Gu, S.: A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In: International conference on machine learning. pp. 6316-6325. PMLR (2021) 2, 3", + "35. Liu, F., Zhao, W., Chen, Y., Wang, Z., Yang, T., Li, J.: Sstdp: Supervised spike timing dependent plasticity for efficient spiking neural network training. Frontiers in neuroscience p. 1413 (2021) 2" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zekai Xu et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Liu, Y., Chen, K., Liu, C., Qin, Z., Luo, Z., Wang, J.: Structured knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2604-2613 (2019) 4", + "37. Maass, W.: Networks of spiking neurons: the third generation of neural network models. Neural networks 10(9), 1659-1671 (1997) 1", + "38. Merolla, P.A., Arthur, J.V., Alvarez-Icaza, R., Cassidy, A.S., Sawada, J., Akopyan, F., Jackson, B.L., Imam, N., Guo, C., Nakamura, Y., et al.: A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345(6197), 668-673 (2014) 1", + "39. Neftci, E.O., Mostafa, H., Zenke, F.: Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36(6), 51-63 (2019) 4, 2", + "40. Orchard, G., Jayawant, A., Cohen, G.K., Thakor, N.: Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience 9, 437 (2015) 6", + "41. Ostojic, S.: Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons. Nature neuroscience 17(4), 594-600 (2014) 2", + "42. Peng, C., Zhao, K., Maksoud, S., Li, M., Lovell, B.C.: Sid: Incremental learning for anchor-free object detection via selective and inter-related distillation. Computer vision and image understanding 210, 103229 (2021) 4", + "43. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021) 6", + "44. Roy, K., Jaiswal, A., Panda, P.: Towards spike-based machine intelligence with neuromorphic computing. Nature 575(7784), 607-617 (2019) 2", + "45. Rueckauer, B., Lungu, I.A., Hu, Y., Pfeiffer, M., Liu, S.C.: Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in neuroscience 11, 682 (2017) 3, 1", + "46. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision 128(2), 336-359 (Oct 2019). https://doi.org/10.1007/s11263-019-01228-7, http://dx.doi.org/10.1007/s11263-019-01228-7 7, 10", + "47. Sengupta, A., Ye, Y., Wang, R., Liu, C., Roy, K.: Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience 13, 95 (2019) 2", + "48. Sironi, A., Brambilla, M., Bourdis, N., Lagorce, X., Benosman, R.: Hats: Histograms of averaged time surfaces for robust event-based object classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1731-1740 (2018) 6", + "49. Sun, S., Cheng, Y., Gan, Z., Liu, J.: Patient knowledge distillation for bert model compression. arXiv preprint arXiv:1908.09355 (2019) 4", + "50. Takuya, S., Zhang, R., Nakashima, Y.: Training low-latency spiking neural network through knowledge distillation. In: 2021 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPS). pp. 1-3. IEEE (2021) 4", + "51. Tan, G., Wang, Y., Han, H., Cao, Y., Wu, F., Zha, Z.J.: Multi-grained spatiotemporal features perceived network for event-based lip-reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20094-20103 (2022) 6" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Enhancing SNN with Blurred Knowledge Distillation", + "bbox": [ + 377, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Wu, Y., Deng, L., Li, G., Zhu, J., Xie, Y., Shi, L.: Direct training for spiking neural networks: Faster, larger, better. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1311-1318 (2019) 2, 4, 1", + "53. Xu, Q., Li, Y., Shen, J., Liu, J.K., Tang, H., Pan, G.: Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7886-7895 (2023) 2, 4, 5, 7, 9, 10, 11, 12, 3", + "54. Xu, Q., Li, Y., Shen, J., Zhang, P., Liu, J.K., Tang, H., Pan, G.: Hierarchical spiking-based model for efficient image classification with enhanced feature extraction and encoding. IEEE Transactions on Neural Networks and Learning Systems (2022) 4", + "55. Xu, Z., Zhang, M., Hou, J., Gong, X., Wen, C., Wang, C., Zhang, J.: Delving into transformer for incremental semantic segmentation (2022) 4", + "56. Yang, Z., Li, Z., Jiang, X., Gong, Y., Yuan, Z., Zhao, D., Yuan, C.: Focal and global knowledge distillation for detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4643-4652 (2022) 4", + "57. Yao, M., Hu, J., Hu, T., Xu, Y., Zhou, Z., Tian, Y., Bo, X., Li, G.: Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In: The Twelfth International Conference on Learning Representations (2023) 4, 8, 10, 6", + "58. Zenke, F., Agnes, E.J., Gerstner, W.: Diverse synaptic plasticity mechanisms orchestrated to form and retrieve memories in spiking neural networks. Nature communications 6(1), 6922 (2015) 2", + "59. Zenke, F., Vogels, T.P.: The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation 33(4), 899-925 (2021) 2, 1", + "60. Zheng, H., Wu, Y., Deng, L., Hu, Y., Li, G.: Going deeper with directly-trained larger spiking neural networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 11062-11070 (2021) 4, 9", + "61. Zhou, C., Yu, L., Zhou, Z., Zhang, H., Ma, Z., Zhou, H., Tian, Y.: Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954 (2023) 2, 3, 4, 9, 14", + "62. Zhou, C., Zhang, H., Zhou, Z., Yu, L., Ma, Z., Zhou, H., Fan, X., Tian, Y.: Enhancing the performance of transformer-based spiking neural networks by improved downsampling with precise gradient backpropagation. arXiv preprint arXiv:2305.05954 (2023) 2, 3, 4, 8, 9, 10, 11, 12, 14", + "63. Zhou, D., Kang, B., Jin, X., Yang, L., Lian, X., Jiang, Z., Hou, Q., Feng, J.: Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886 (2021) 6", + "64. Zhou, H., Song, L., Chen, J., Zhou, Y., Wang, G., Yuan, J., Zhang, Q.: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. arXiv preprint arXiv:2102.00650 (2021) 7", + "65. Zhou, Z., Zhu, Y., He, C., Wang, Y., Yan, S., Tian, Y., Yuan, L.: Spikformer: When spiking neural network meets transformer (2022) 2, 3, 4, 9, 14, 5" + ], + "bbox": [ + 215, + 146, + 785, + 743 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Zekai Xu et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_model.json b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..965959796a8ae84faeffb2ea40baa59de45c0ad8 --- /dev/null +++ b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_model.json @@ -0,0 +1,2711 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.242, + 0.141, + 0.761, + 0.21 + ], + "angle": 0, + "content": "BKDSNN: Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.234, + 0.761, + 0.252 + ], + "angle": 0, + "content": "Zekai \\(\\mathrm{Xu}^{1}\\), Kang You\\(^{1}\\), Qinghai Guo\\(^{2}\\), Xiang Wang\\(^{2}\\), and Zhezhi He\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.262, + 0.777, + 0.29 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.291, + 0.639, + 0.304 + ], + "angle": 0, + "content": "\\(^{2}\\) Huawei Technologies, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.306, + 0.703, + 0.319 + ], + "angle": 0, + "content": "{sherlock.holmes.xu,kang_you,\\(^\\dagger\\)zhezhi.he}@sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.32, + 0.742, + 0.333 + ], + "angle": 0, + "content": "https://github.com/Intelligent-Computing-Research-Group/BKDSNN" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.366, + 0.744, + 0.63 + ], + "angle": 0, + "content": "Abstract. Spiking neural networks (SNNs), which mimic biological neural systems to convey information via discrete spikes, are well-known as brain-inspired models with excellent computing efficiency. By utilizing the surrogate gradient estimation for discrete spikes, learning-based SNN training methods that can achieve ultra-low inference latency (i.e., number of time-step) have emerged recently. Nevertheless, due to the difficulty of deriving precise gradient for discrete spikes in learning-based methods, a distinct accuracy gap persists between SNNs and their artificial neural networks (ANNs) counterparts. To address the aforementioned issue, we propose a blurred knowledge distillation (BKD) technique, which leverages randomly blurred SNN features to restore and imitate the ANN features. Note that, our BKD is applied upon the feature map right before the last layer of SNNs, which can also mix with prior logits-based knowledge distillation for maximal accuracy boost. In the category of learning-based methods, our work achieves state-of-the-art performance for training SNNs on both static and neuromorphic datasets. On the ImageNet dataset, BKDSNN outperforms prior best results by \\(4.51\\%\\) and \\(0.93\\%\\) with the network topology of CNN and Transformer, respectively." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.643, + 0.741, + 0.671 + ], + "angle": 0, + "content": "Keywords: Learning-based Spiking Neural Networks \\(\\cdot\\) Knowledge Distillation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.377, + 0.712 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Inspired by the neural firing mechanism in the biological nervous system, spiking neural network (SNN) [37] emerges as an alternative framework for the traditional artificial neural network (ANN). ANNs rely on continuous activation values to convey information between neurons [31], while SNNs utilize discrete spiking events to process and propagate information [6,38]. Given the remarkable energy efficiency achieved by such event-driven information propagation [18], SNNs" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "† Corresponding Author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "Zekai Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "show promising prospects on computational intelligence tasks [44] with strong autonomous learning capabilities and ultra-low power consumption [2,11,41,58]. Apart from these advantages, SNNs are also natively compatible with event-based neuromorphic datasets (e.g., CIFAR10-DVS) [14,61,62,65]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.212, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Currently, conversion-based and learning-based are two common methods to train SNNs. Conversion-based methods attempt to leverage information from ANNs by transferring the parameters of a pre-trained ANN to its SNN counterpart, which have achieved promising results in several tasks [2, 8, 33, 34]. However, event-triggered neurons convey limited information with ultra-low inference time-steps (e.g., 4 time-steps), leading to degraded performance of conversion-based methods [14, 65]. Consequently, conversion-based methods usually take a long inference time (e.g., 32 time-steps) to get comparable accuracy as the original ANN [3, 47] based on the equivalence of the ReLU activation and the firing rate of integrate-and-fire (IF) neuron [16]. In order to further improve the performance of SNNs with ultra-low inference time-steps, learning-based methods unfold SNN training over the inference time-steps and leverage back-propagation through time (known as BPTT) [52, 59] to train SNNs. Inspired by [27], several works [14, 35, 62] adopt surrogate gradient estimation as an implicit differentiation for direct training of SNNs. With the combination of BPTT and surrogate gradient estimation, learning-based methods exhibit advantages in processing event-based neuromorphic datasets (e.g., CIFAR10-DVS [22]) [14, 61, 62, 65]. Furthermore, direct training of CNN- [14, 23] and Transformer-based [61, 62, 65] SNNs also achieves promising performance with ultra-low time-steps on static datasets (e.g., CIFAR10 [28], CIFAR100 [28] and ImageNet [7])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Despite that numerous efforts have been made in learning-based methods, SNNs still need improvement compared with ANNs. [53, 62] claim that imprecise surrogate gradient estimation of spiking neurons leads to performance degradation. On the other hand, KDSNN [53] and LaSNN [21] introduce layer-wise knowledge distillation (KD) to train a student SNN with rich information from the ANN teacher, achieving promising performance on several classical vision datasets (MNIST, CIFAR10/100) with CNN-based models. However, according to our preliminary experimental results, both KDSNN and LaSNN struggle with more complex datasets (ImageNet) and Transformer-based models. We hypothesize that the direct distillation between discrete SNN features and continuous ANN features [21, 53] might be the source of the limitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, to further alleviate performance degradation of learning-based SNNs, we introduce a blurred knowledge distillation (BKD) technique to leverage randomly blurred SNN features to restore and mimic ANN features. We apply BKD to the intermediate feature before the last layer of the student SNN and prove its superiority over layer-wise knowledge distillation in learning-based SNNs. Furthermore, we combine prior logits-based distillation with BKD, which provides mutual benefits for improving performance. We evaluate BKDSNN on both static datasets (i.e., CIFAR10, CIFAR100, and ImageNet) and a neuromorphic dataset (i.e., CIFAR10-DVS), where we achieve state-of-the-art (SOTA) performance with both CNN- and Transformer-based SNN models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Enhancing SNN with Blurred Knowledge Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.147, + 0.578, + 0.161 + ], + "angle": 0, + "content": "To summarize, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.173, + 0.784, + 0.215 + ], + "angle": 0, + "content": "- We introduce a blurred knowledge distillation (BKD) technique to guide the training of student SNNs through feature imitation and demonstrate the advantages of BKD on learning-based SNNs." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.218, + 0.784, + 0.275 + ], + "angle": 0, + "content": "- We apply BKD to the intermediate feature before the last layer and prove its superiority over previous layer-wise distillation. Combined with prior logits-based knowledge distillation, our BKDSNN maximizes accuracy boost on learning-based SNNs." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.278, + 0.784, + 0.335 + ], + "angle": 0, + "content": "- We benchmark our method on both static datasets (i.e., CIFAR10, CIFAR100 and ImageNet) and a neuromorphic dataset (i.e., CIFAR10-DVS). Our method outperforms prior SOTA learning-based methods for both CNN-and Transformer-based models." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.173, + 0.784, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.362, + 0.553, + 0.38 + ], + "angle": 0, + "content": "2 Background and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.395, + 0.784, + 0.485 + ], + "angle": 0, + "content": "Spiking Neuron Model. Previous works [2,3,8,14,24,61,62,65] have widely used the classic integrate-and-fire (IF) neuron as the fundamental unit in SNNs. IF neuron integrates input currents through changes in its membrane potential, which is compared against the threshold for spike generation. The membrane potential is reset after emitting a spike. The dynamics of the IF neuron is depicted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.496, + 0.667, + 0.525 + ], + "angle": 0, + "content": "\\[\nH [ t ] = V [ t - 1 ] + \\frac {1}{\\tau} \\left(I [ t ] - (V [ t - 1 ] - V _ {\\mathrm {r e s e t}})\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.527, + 0.785, + 0.567 + ], + "angle": 0, + "content": "\\[\nS [ t ] = \\Theta (H [ t ] - V _ {\\mathrm {t h}}) = \\left\\{ \\begin{array}{l l} 1, & H [ t ] \\geq V _ {\\mathrm {t h}} \\\\ 0, & H [ t ] < V _ {\\mathrm {t h}} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.569, + 0.621, + 0.587 + ], + "angle": 0, + "content": "\\[\nV [ t ] = H [ t ] (1 - S [ t ]) + V _ {\\text {r e s e t}} S [ t ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.598, + 0.784, + 0.66 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the membrane time constant, \\(I[t]\\) is the receiving current at time step \\(t\\) and \\(\\Theta (\\cdot)\\) is the heaviside function. The neuron accumulates \\(I[t]\\) to change the membrane potential \\(H[t]\\) at time step \\(t\\). Once \\(H[t]\\) exceeds the firing threshold \\(V_{\\mathrm{th}}\\), the neuron generates a spike and resets the membrane potential to \\(V_{\\mathrm{reset}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Conversion-based Methods. To fully make use of the knowledge from pre-trained ANN, conversion-based methods are designed to transfer the learned parameters of ANNs into corresponding SNNs, without further training of SNNs. Previous works [17, 18] replace ReLU activations in ANNs with spiking neurons and add scaling operations like weight normalization and threshold balancing [10] to generate corresponding SNNs. Although several works [8, 17, 18, 34] have achieved nearly loss-less accuracy with VGG-16 and ResNet, the converted SNNs still require long inference time to match the original ANNs in precision. [45] claims that low firing rates caused by weight normalization result in increasing latency for the information reaching higher layers, which limits the practical application of the converted SNNs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.373, + 0.127 + ], + "angle": 0, + "content": "Zekai Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.789, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.284, + 0.789, + 0.41 + ], + "angle": 0, + "content": "Fig.1: Overview of training SNN with blurred knowledge distillation (BKD). The SNN is directly trained with back-propagation through time (BPTT) [52], where we utilize a blurred variant of KD (i.e., BKD) to achieve higher accuracy. BKD highlighted in yellow shaded region differs from prior SNN-oriented KD in three perspectives: 1) A blurred matrix \\(\\pmb{B}\\) is randomly sampled on the fly (per input image) to mask out the feature of student SNN; 2) A restoration block \\(\\mathcal{G}\\) consisting of two convolutional layers connected by a ReLU layer is applied on blurred SNN features to restore and mimic ANN features; 3) Such blurred knowledge distillation is applied only to the intermediate features before the last layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.789, + 0.693 + ], + "angle": 0, + "content": "Learning-based Methods. With the development of surrogate gradient estimation [39], learning-based methods leverage continuous-value-based loss functions to handle nondifferentiable difficulties in SNN training. Ref. [39] introduces surrogate gradient estimation to conduct SNN training on the event-based DVS gesture dataset, which achieves better accuracy after a few training iterations. In an effort to improve computational resources in SNNs, joint CNN-SNN [54] puts forward a novel augmented spiking-based framework and achieves promising results in several image classification tasks. DCT-SNN [15] reduces the number of inference time-steps by combining frequency-domain learning with surrogate gradient descent. Spiking ResNet [60] uses the directly-trained deep SNN to achieve high performance on ImageNet. SEW ResNet [14] proposes the spike-element-wise block to counter gradient vanishing and gradient exploding, which successfully extends the depth of directly-trained SNN to more than 100 layers. When it comes to Transformer-based SNN, Spikformer [65] with a novel spiking self-attention block achieves significant improvement in image classification on both neuromorphic and static datasets. Transformer-based SNN is further improved by successive Spikingformer [61], ConvBN-MaxPooling-LIF (CML) [62] and Meta-SpikeFormer [57], achieving SOTA performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Knowledge Distillation for SNN Training. Knowledge distillation (KD) is a classical transfer learning method that has been proven to be effective on various tasks, such as model compression [20, 36, 49, 56], incremental learning [4, 5, 42, 55], etc. Recently, several works [30, 32, 50] extend KD for SNN training, such as distilling spikes from a large SNN to a small SNN or logits-based knowledge from a well-trained ANN to a simple SNN. KDSNN [53] proposes a novel framework embracing both logits-based and feature-based knowledge distillation, which enhances performance on several image classification datasets (MNIST, CIFAR10, and CIFAR100). LaSNN [21] puts forward a layer-wise feature-based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Enhancing SNN with Blurred Knowledge Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.348, + 0.145, + 0.655, + 0.16 + ], + "angle": 0, + "content": "Table 1: Summary of notations in this paper" + }, + { + "type": "table", + "bbox": [ + 0.308, + 0.171, + 0.697, + 0.364 + ], + "angle": 0, + "content": "
NotationDescription
Ttime-step
λblurred ratio
R∈Rb×ctearandom matrix R~U(0,1)
B∈Rb×cteablurred matrix
ˆy∈Rb×nclsground truth label
ystu,ytea∈Rb×nclsoutput logits of student SNN and teacher ANN
qstu,qtea∈Rb×nclsystuand ytea after flattened processing
Ftea∈Rb×ctea×h×wfeature before the last layer of teacher ANN
Fstu∈RT×b×cstu×h×wfeature before the last layer of student SNN
Fstu∈Rb×cstu×h×waverage of Fstu at time-step dimension (T)
fadpadaptive layer to align Fstu with Ftea
Grestoration block for Fstu
ˆFstu∈Rb×ctea×h×wblurred restoration feature from Fstu
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.466 + ], + "angle": 0, + "content": "ANN-to-SNN distillation framework, achieving comparable top-1 accuracy to ANNs on the challenging Tiny ImageNet dataset. Furthermore, KDSNN [53] and LaSNN [21] are applicable to both homogeneous and heterogeneous networks. Despite the superiority of KDSNN [53] and LaSNN [21], they struggle with more complex datasets (e.g., ImageNet) with CNN-based and Transformer-based models according to our preliminary experimental results in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.494, + 0.331, + 0.509 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.697 + ], + "angle": 0, + "content": "In this paper, we propose a novel blurred knowledge distillation SNN (overall architecture of BKDSNN is in Fig. 1a) training method that leverages knowledge from the teacher ANN to guide the training of the student SNN. As depicted in Fig. 1b, our BKD first introduces a randomly blurred matrix \\(\\pmb{B}\\) to mask the student SNN features. Unlike KDSNN [53] and LaSNN [21], which only apply an adaptive layer to align the SNN feature with the ANN feature before distillation, BKD introduces a restoration block \\(\\mathcal{G}\\) on blurred SNN feature to restore and mimic ANN feature. The proposed BKD is applied only to the intermediate feature before the last layer, different from the layer-wise distillation [21, 53]. As shown in Fig. 1a, we combine the logits-based distillation with BKD to leverage both logits and feature knowledge from the teacher ANN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.79 + ], + "angle": 0, + "content": "Blurred Knowledge Distillation. To facilitate understanding, we present important notations in Tab. 1. The proposed blurred knowledge distillation is illustrated in Fig. 1b. We first introduce a random matrix \\(\\pmb{R}\\) which obeys uniform distribution between 0 and 1. Then we apply a blurred ratio \\(\\lambda\\) to \\(\\pmb{R}\\) to generate a blurred matrix \\(\\pmb{B}\\) through Eq. (2)." + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.803, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {B} _ {i, j} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\boldsymbol {R} _ {i, j} < \\lambda \\\\ 1, & \\text {O t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.127 + ], + "angle": 0, + "content": "Zekai Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.511, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.292, + 0.485, + 0.304 + ], + "angle": 0, + "content": "(a) Feature visualization of ResNet-18." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.148, + 0.785, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.292, + 0.781, + 0.304 + ], + "angle": 0, + "content": "(b) Feature distribution of SNN and ANN." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.315, + 0.784, + 0.345 + ], + "angle": 0, + "content": "Fig.2: (a) Comparison of feature visualization in ResNet-18 under various methods; (b) Histogram of SNN and ANN feature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.784, + 0.387 + ], + "angle": 0, + "content": "For feature \\(\\mathbf{F}^{\\mathrm{stu}}\\) at the last layer of student SNN, we first average \\(\\mathbf{F}^{\\mathrm{stu}}\\) at timestep dimension \\((T)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.387, + 0.785, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {F}} ^ {\\mathrm {s t u}} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\mathbf {F} _ {i} ^ {\\mathrm {s t u}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.432, + 0.788, + 0.53 + ], + "angle": 0, + "content": "After the averaging operation above, we apply a convolutional adaptive layer \\( f_{\\mathrm{adp}} \\) to align \\( \\tilde{\\mathbf{F}}^{\\mathrm{stu}} \\) with \\( \\mathsf{F}_i^{\\mathrm{tea}} \\) when student feature dimension \\( c^{\\mathrm{stu}} \\) mismatches with teacher feature dimension \\( c^{\\mathrm{tea}} \\). After the alignment, we construct a restoration block \\( \\mathcal{G} \\) with two convolutional layers connected by a ReLU layer and apply it to \\( \\tilde{\\mathbf{F}}^{\\mathrm{stu}} \\) with the blurred matrix \\( B \\) to achieve blurred restoration feature \\( \\hat{\\mathbf{F}}^{\\mathrm{stu}} \\) as illustrated in Eq. (4)." + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.54, + 0.785, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} ^ {\\mathrm {s t u}} = \\left\\{ \\begin{array}{l l} \\mathcal {G} \\left(f _ {\\mathrm {a d p}} \\left(\\tilde {\\mathbf {F}} ^ {\\mathrm {s t u}}\\right) * \\boldsymbol {B}\\right), & c ^ {\\mathrm {s t u}} \\neq c ^ {\\mathrm {t e a}} \\\\ \\mathcal {G} \\left(\\tilde {\\mathbf {F}} ^ {\\mathrm {s t u}} * \\boldsymbol {B}\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.6, + 0.784, + 0.628 + ], + "angle": 0, + "content": "We calculate blurred knowledge distillation loss \\(\\mathcal{L}_{\\mathrm{BKD}}\\) using quadratic norm \\((L^2)\\) as follows." + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.628, + 0.785, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B K D}} = \\sum_ {i} ^ {b} \\sum_ {j} ^ {c ^ {\\mathrm {s t u}}} \\sum_ {k} ^ {h} \\sum_ {l} ^ {w} \\left(\\hat {\\mathbf {F}} _ {i, j, k, l} ^ {\\mathrm {s t u}} - \\mathbf {F} _ {i, j, k, l} ^ {\\mathrm {t e a}}\\right) ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.676, + 0.784, + 0.709 + ], + "angle": 0, + "content": "We then combine blurred knowledge distillation loss \\(\\mathcal{L}_{\\mathrm{BKD}}\\) with the original cross entropy loss \\(\\mathcal{L}_{\\mathrm{CE}}\\) to get the total loss \\(\\mathcal{L}_{\\mathrm{total}}\\) as follows." + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.719, + 0.785, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {C E}} (\\boldsymbol {y} ^ {\\mathrm {s t u}}, \\hat {\\boldsymbol {y}}) + \\mathcal {L} _ {\\mathrm {B K D}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of \\( BKD \\). For both CNN- and Transformer-based models, features of deeper layers have better representation for the original input image [26,43,63]. In other words, features of deeper layers contain sufficient information of adjacent pixels (for CNN-based models) [26] or patch pairs (for Transformer-based models) [43, 63]. Therefore, we can use partial student SNN feature before the last layer to restore and mimic teacher ANN feature through a restoration block" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Enhancing SNN with Blurred Knowledge Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.34 + ], + "angle": 0, + "content": "\\(\\mathcal{G}\\), which can help the student SNN achieves better representation and effectively mimics teacher ANN. We then apply Grad-CAM which can intuitively demonstrate the feature similarity [46] on each block in ResNet-18 under various methods. As depicted in Fig. 2a, compared with previous SNN (SEW ResNet-18 [14]) and KDSNN [53], our BKDSNN transfers sufficient knowledge from ANN and effectively mimics ANN feature. Meanwhile, as illustrated in Fig. 2b, student SNN feature \\(\\tilde{\\mathsf{F}}^{\\mathrm{stu}}\\) is discrete while teacher ANN feature \\(\\mathsf{F}^{\\mathrm{tea}}\\) is continuous. Therefore, directly applying distillation to encourage discrete SNN feature \\(\\tilde{\\mathsf{F}}^{\\mathrm{stu}}\\) to imitate continuous ANN feature \\(\\mathsf{F}^{\\mathrm{tea}}\\) [21, 53] is not plausible. However, the proposed BKD leverages the restoration block \\(\\mathcal{G}\\) to map discrete feature \\(\\tilde{\\mathsf{F}}^{\\mathrm{stu}}\\) to continuous feature \\(\\hat{\\mathsf{F}}^{\\mathrm{stu}}\\) for imitating continuous ANN feature \\(\\mathsf{F}^{\\mathrm{tea}}\\), which is a more proper spiking-friendly feature-based distillation method." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.357, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Mixed Distillation. As depicted in Fig. 1a, BKDSNN further applies logits-based distillation which transfers knowledge from teacher ANN logits to student SNN logits. As soft logits from the teacher model embrace sufficient hidden knowledge, which serves as a good regularization to student model [64], we introduce the temperature \\(\\tau\\) to make the logits distribution flatter. The output logits of both teacher and student are processed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.459, + 0.786, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} _ {i} = \\frac {\\exp \\left(\\boldsymbol {y} _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {n ^ {\\mathrm {c l s}}} \\exp \\left(\\boldsymbol {y} _ {j} / \\tau\\right)}, \\quad i \\in \\{1, 2, \\dots , n ^ {\\mathrm {c l s}} \\} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.509, + 0.592, + 0.525 + ], + "angle": 0, + "content": "Then logits-based distillation loss \\(\\mathcal{L}_{\\mathrm{LD}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.536, + 0.786, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L D}} = \\tau^ {2} \\cdot \\mathcal {L} _ {\\text {C r o s s E n t r o p y}} \\left(\\boldsymbol {q} ^ {\\text {s t u}}, \\boldsymbol {q} ^ {\\text {t e a}}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.566, + 0.784, + 0.595 + ], + "angle": 0, + "content": "We combine \\(\\mathcal{L}_{\\mathrm{LD}}\\) with \\(\\mathcal{L}_{\\mathrm{BKD}}\\) to achieve mixed distillation loss \\(\\mathcal{L}_{\\mathrm{MD}}\\) (Eq. (9)), which contains both logits and feature information from the teacher ANN." + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.609, + 0.786, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M D}} = w _ {\\mathrm {L D}} \\cdot \\mathcal {L} _ {\\mathrm {L D}} + w _ {\\mathrm {B K D}} \\cdot \\mathcal {L} _ {\\mathrm {B K D}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.638, + 0.784, + 0.667 + ], + "angle": 0, + "content": "where \\(w_{\\mathrm{LD}}\\) and \\(w_{\\mathrm{BKD}}\\) are weights for \\(\\mathcal{L}_{\\mathrm{LD}}\\) and \\(\\mathcal{L}_{\\mathrm{BKD}}\\), respectively. Thus, the overall loss \\(\\mathcal{L}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.316, + 0.68, + 0.786, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {C r o s s E n t r o p y}} \\left(\\boldsymbol {y} ^ {\\text {s t u}}, \\hat {\\boldsymbol {y}}\\right) + \\mathcal {L} _ {\\mathrm {M D}} = \\mathcal {L} _ {\\text {t a s k}} + \\mathcal {L} _ {\\mathrm {M D}} \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.735, + 0.376, + 0.751 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.768, + 0.494, + 0.784 + ], + "angle": 0, + "content": "4.1 Experimental Configuration" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Platform. All experiments are evaluated on a server equipped with 32 cores Intel Xeon Platinum 8352V CPU with \\(2.10\\mathrm{GHz}\\) and 8-way NVIDIA GPUs. We use SpikingJelly [13] to simulate IF neurons." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "Zekai Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Static Datasets and Settings. We first evaluate BKDSNN on three static datasets, i.e., CIFAR10/100 and ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.317 + ], + "angle": 0, + "content": "CIFAR10/CIFAR100 [28] contains 50k training and 10k testing images with \\(32 \\times 32\\) resolution, and is labeled with 10- or 100-categories respectively for classification. For CNN-based models, we use ResNet-19 from TET [9] and lightweight ResNet-20 from SEW ResNet [14] with corresponding ANNs as teacher models. For Transformer-based models, we train ViT-S/16 on CIFAR10 and CIFAR100 to make it serve as the teacher ANN model for Sformer-4-256, Sformer-2-384, and Sformer-4-384 (We abbreviate Spikingformer to Sformer for convenience). Sformer-L-D represents a Sformer model with L Sformer encoder blocks and D feature embedding dimensions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.788, + 0.456 + ], + "angle": 0, + "content": "ImageNet-1K [7] contains 1.3 millions of training images and 50k validation images with \\(224 \\times 224\\) resolution, labeled with 1k classes. We choose ResNet-18, ResNet-34, and ResNet-50 pre-trained on ImageNet-1K as teacher ANN models for the corresponding SEW ResNet models [14]. For Transformer-based models, we introduce ViT-B/16 [12] pre-trained on ImageNet1K as the teacher ANN model for Sformer-8-384, Sformer-8-512 and Sformer-8-768. We also apply our BKDSNN on the latest Meta-SpikeFormer [57] and conduct experiments with Mformer-8-256, Mformer-8-384 and Mformer-8-512 on ImageNet-1K (We abbreviate Meta-SpikeFormer to Mformer for convenience.)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.596 + ], + "angle": 0, + "content": "For both CNN- and Transformer-based models, we set temperature \\(\\tau\\) to 2.0, blurred ratio \\(\\lambda\\) to 0.15 and logits-based distillation weight \\(w_{\\mathrm{LD}}\\) to 1. Specifically for CNN-based models, we set blurred knowledge distillation weight \\(w_{\\mathrm{BKD}}\\) to \\(7e^{-4}\\) for CIFAR10/100 and \\(7e^{-5}\\) for ImageNet. We follow all the training configurations (batch size, learning rate, optimizer, etc.) from SEW ResNet [14] except ResNet-19, where we use training configuration from TET [9]. For Transformer-based models, we set \\(w_{\\mathrm{BKD}}\\) to \\(2e^{-3}\\) for both CIFAR10/100 and ImageNet, and we follow all the training configurations (batch size, learning rate, optimizer, etc.) from Sformer+CML [62]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Neuromorphic Dataset and Networks. We also evaluate BKDSNN on CIFAR10-DVS [22], which is a neuromorphic dataset that leverages dynamic vision sensor (DVS) to convert 10k frame-based images from the CIFAR10 dataset into 10k event streams, providing an event-stream dataset with 10 different classes. It includes 9k training samples and 1k testing samples. Note that, CIFAR10-DVS is an event-stream dataset tailored to SNN, so we use SNN models with 16 timesteps as teacher models to guide the training of SNN models with lower timesteps (i.e., 4 and 8). We utilize Wide-7B-Net [14] and Sformer-2-256 [62] with 4 and 8 inference time-steps as student models for CNN-based and Transformer-based models, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.843 + ], + "angle": 0, + "content": "For both Wide-7B-Net [14] and former-2-256 [62], we set \\(\\tau\\) to 2.0, \\(\\lambda\\) to 0.40 and \\(w_{\\mathrm{LD}}\\) to 1. Then we set \\(w_{\\mathrm{BKD}}\\) to \\(2e^{-3}\\) for Wide-7B-Net [14] and \\(7e^{-5}\\) for Sformer-2-256 [62]. We follow all the training configurations from SEW ResNet [14] for Wide-7B-Net [14], and Sformer+CML [62] for Sformer-2-256 [62]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Enhancing SNN with Blurred Knowledge Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.23 + ], + "angle": 0, + "content": "Table 2: Comparison with previous works, using various networks and multiple datasets. \\(P\\) denotes #parameters. \\(KD\\) denotes knowledge distillation. \\(\\star\\): results taken from [14]. †: prior state-of-the-art conversion-based method QCFS [2]. The prior SOTA results are marked in gray, which is used to compare with our BKDSNN and report the accuracy difference. Best results are in bold. - indicates results not available (not reported in original article or cannot reproduce)." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.233, + 0.788, + 0.352 + ], + "angle": 0, + "content": "
w/o KD MethodArchitectureR-19 (P: 12.63M)R-20 (P: 0.27M)Time StepArchitecture MethodSformer-4-256 (P: 4.15M)Sformer-2-384 (P: 5.76M)Sformer-4-384 (P: 9.32M)Time Step
CF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 Acc
QCS [2]†--83.7555.234w/o KD Spíkarfer [65]93.9475.9694.8076.9595.1977.864
TET [9]94.4474.4787.7559.144+CML [62]94.8277.6495.6378.7595.9379.654
STBP [60]92.9270.86--4+CML [62]94.7777.4395.2278.3495.6179.094
Spíkarfer-R [23]92.1370.1387.3259.934
SEW-R [14]93.2470.8489.0760.164
Teacher ANN95.6079.7891.7768.401Teacher ANN96.7582.2296.7582.2296.7582.221
KDSNN [53]94.3674.0889.0360.184KDSNN [53]95.0078.3895.5979.2595.8880.334
LaSNN [21]94.2173.9288.4159.174LaSNN [21]94.9778.0295.5578.9195.7979.994
BKDSNN (ours)94.6474.9589.2960.924BKDSNN (ours)95.29 (+0.29)79.41 (+1.03)95.90 (+0.31)80.63 (+1.38)96.06 (+0.18)81.26 (+0.93)4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.354, + 0.463, + 0.365 + ], + "angle": 0, + "content": "(a) CNN-based SNN on \\( CF10 / 100 \\) dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.48, + 0.353, + 0.774, + 0.365 + ], + "angle": 0, + "content": "(b) Transformer-based SNN on \\( CF10/100 \\) dataset." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.378, + 0.483, + 0.501 + ], + "angle": 0, + "content": "
w/o KD MethodArchitectureR-18(P: 11.69M)Top1 AccR-34(P: 21.79M)Top1 AccR-50(P: 25.56M)Top1 AccTimeStep
QCFS [2]†-69.37-32
TET [9]-68.00-4
STBP [60]-67.04-4
Spiking-R [23]★62.3261.8657.664
SEW-R [14]63.1867.0467.784
with KD (ours)Teacher ANN69.7673.3176.311
KDSNN [53]63.4267.1867.724
LaSNN [21]63.3166.9467.814
BKDSNN(ours)65.60(+2.18)71.24(+4.06)72.32(+4.51)4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.231, + 0.503, + 0.479, + 0.514 + ], + "angle": 0, + "content": "(c) CNN-based SNN on ImageNet dataset." + }, + { + "type": "table", + "bbox": [ + 0.493, + 0.378, + 0.789, + 0.501 + ], + "angle": 0, + "content": "
Architecture MethodSformer-8-384(P: 16.81M)Top1 Acc.Sformer-8-512(P: 29.68M)Top1 Acc.Sformer-8-768(P: 66.34M)Top1 Acc.Time Step
w/o KDSpikformer [65]70.2473.3874.814
+CML [62]72.7375.6177.344
Sformer [61]72.4574.7975.854
+CML [62]74.3576.5477.644
with KDTeacher ANN81.7881.7881.781
KDSNN [53]74.6276.4477.834
LaSNN [21]73.8576.3877.664
BKDSNN(ours)75.48(+0.86)77.24(+0.70)79.93(+2.10)4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.496, + 0.502, + 0.785, + 0.514 + ], + "angle": 0, + "content": "(d) Transformer-based SNN on ImageNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.588 + ], + "angle": 0, + "content": "Settings for Prior KD Methods on SNNs. Furthermore, to compare BKDSNN with previous KD methods on SNNs (i.e., KDSNN [53] and LaSNN [21]), we follow the settings from [53] and [21] to reproduce the experimental results on all datasets with CNN-based and Transformer-based models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.615, + 0.519, + 0.63 + ], + "angle": 0, + "content": "4.2 Comparison on Static Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Comparison on CIFAR10/100 between BKDSNN and previous methods are listed in Tab. 2a and Tab. 2b, where BKDSNN surpasses prior SOTA methods with both CNN- and Transformer-based models. Specifically, BKD applied upon Sformer-4-384 outperforms previous SOTA by \\(0.18\\%\\) on CIFAR10 and \\(0.93\\%\\) on CIFAR100, which achieves \\(96.06\\%\\) and \\(81.26\\%\\) accuracy respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Intuitively, it is difficult for both KDSNN [53] and LaSNN [21] to improve the performance of CNN-based models on the CIFAR dataset. For ResNet-19, the knowledge of the ANN teacher is limited for the SNN student, leading to the above difficulties. When it comes to ResNet-20, the lightweight model conveys limited information which results in the ineffectiveness of previous knowledge distillation methods. In contrast, our BKDSNN is effective on both limited teacher knowledge as well as lightweight student models, which verifies the superiority of our method." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "Zekai Xu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 3: (a) Comparison with Meta-SpikeFormer [57] on ImageNet. (b) Experimental results on CIFAR10-DVS. Mformer represents Meta-SpikeFormer [57]. \\(\\star\\): SNN with high inference time-step (i.e., 16) serves as the teacher model for corresponding SNN with low inference time-step (i.e., 4 and 8). The prior state-of-the-art results are marked in gray. Best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.315 + ], + "angle": 0, + "content": "
Architecture MethodMformer-8-256(P: 15.11M)Top1 Acc.Mformer-8-384(P: 31.33M)Top1 Acc.Mformer-8-512(P: 55.35M)Top1 Acc.Time StepTransformer-basedCNN-based
MethodAccTime StepAccTime StepMethodAccTime StepAccTime Step
wDMformer [57]71.8275.4077.961Sfomer-2-256+CML [62]★81.41681.416Wide-7B-Net[14]★74.41674.416
74.1177.2078.814
Teacher ANN81.7881.7881.781Sfomer-2-256+CML [62]76.4479.28Wide-7B-Net[14]64.8470.28
with KDMformer [57]--79.111KDSNN [53]78.1479.98KDSNN [53]66.1470.98
--80.004LaSNN [21]78.3480.18LaSNN [21]66.6470.58
BKDSNN(ours)72.9376.1079.821BKDSNN(ours)79.3(+1.0)480.8(+0.7)8BKDSNN(ours)68.3(+1.7)472.2(+1.3)8
" + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.316, + 0.761, + 0.327 + ], + "angle": 0, + "content": "(a) Meta-SpikeFormer on ImageNet dataset. (b) Experimental results on CIFAR10-DVS." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.505, + 0.488, + 0.516 + ], + "angle": 0, + "content": "(a) Feature map visualization." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Fig.3: (a) Feature map visualization of different methods on SEW ResNet-18 (CNN) and Sformer-8-384+CML (Transformer). Input images are sampled from the ImageNet validation dataset. Shaded colors from blue to red indicate the impacts of the regions on the classification scores from low to high. (b) Top1 accuracy versus GPU hours on ImageNet with SEW ResNet-50 and Spikingformer-8-768." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Comparison on ImageNet between BKDSNN and previous methods are tabulated in Tab. 2c, Tab. 2d and Tab. 3a. As anticipated, BKDSNN outperforms previous SOTA methods with both CNN- and Transformer-based models on ImageNet. BKDSNN achieves \\(72.32\\%\\) top1 accuracy with SEW ResNet-50 and \\(80.93\\%\\) with Mformer-8-512, with \\(4.51\\%\\) and \\(0.93\\%\\) improvement, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.788, + 0.791 + ], + "angle": 0, + "content": "As reported in Tab. 2c, Tab. 2d and Tab. 3a, ANN outperforms previous learning-based SNN by a large margin, which will serve as the teacher model to provide adequate knowledge for the student SNN. Based on our experiments, the accuracy improvement of SNN trained by KDSNN [53] and LaSNN [21] is limited, due to the inappropriate distillation between student and teacher models. Nevertheless, thanks to the blurred student features that effectively mimic and restore the teacher features, BKDSNN achieves promising results on the large-scale dataset (i.e. ImageNet)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Feature Visualization. To further interpret BKDSNN, we apply Grad-CAM [46] to visualize the intermediate feature before the last layer using different methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Enhancing SNN with Blurred Knowledge Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.283 + ], + "angle": 0, + "content": "In \\(2^{\\mathrm{nd}}\\) (ANN) to \\(3^{\\mathrm{rd}}\\) (SNN) column of Fig. 3a, ANN properly focuses on the regions with vital information (highlighted in red), while learning-based SNN is distracted by the background area. Such erroneous attention of SNN (\\(3^{\\mathrm{rd}}\\) column in Fig. 3a) does not get calibrated by the previous KD methods, i.e., KDSNN [53] (\\(4^{\\mathrm{th}}\\) column) and LaSNN [21] (\\(5^{\\mathrm{th}}\\) column). However, the attention (shaded) region of our BKDSNN (\\(6^{\\mathrm{th}}\\) column) is quite similar to the pre-trained ANN, which demonstrates the effectiveness of BKD in learning the knowledge of ANN by mimicking and restoring teacher ANN features with blurred student SNN features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Comparison of Training Cost. We plot the evolution curves of validation top1 accuracy during GPU hours as depicted in Fig. 3b. Without applying knowledge distillation (denoted as w/o KD in blue lines), the training of SEW ResNet-50 and Spikingformer-8-768 consume the least GPU hours. With the introduction of layer-wise distillation (LaSNN [21] in green lines and KDSNN [53] in yellow lines), the training of SEW ResNet-50 and Spikingformer-8-768 consume more hours while the improvement is limited. Compared with the layer-wise distillation methods above, our single-layer BKDSNN (in red lines) saves GPU hours and boosts the performance significantly." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.454, + 0.583, + 0.47 + ], + "angle": 0, + "content": "4.3 Comparison on Neuromorphic Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.586 + ], + "angle": 0, + "content": "We also evaluate BKDSNN on CIFAR10-DVS with results reported in Tab. 3b. Specifically, BKDSNN outperforms previous SOTA results by \\(1.0\\%\\) on models with 4 time-steps and \\(0.7\\%\\) on models with 8 time-steps. Note that, the improvement of BKDSNN on Wide-7B-Net [14] is more pronounced than that on Sformer-2-256 [62], which is probably due to the fact that Wide-7B-Net [14] with 16 inference time-steps provides more knowledge for the corresponding student model with lower time-steps, compared to Sformer-2-256 [62]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.605, + 0.39, + 0.621 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Ablation on multiple losses \\((\\mathcal{L}_{LD},\\mathcal{L}_{BKD}\\) and \\(\\mathcal{L}_{MD})\\) . We test different settings of BKDSNN, including versions without any knowledge distillation (w/o KD), with logit-based KD \\((+{\\mathcal{L}}_{\\mathrm{LD}})\\) , with BKD \\((+{\\mathcal{L}}_{\\mathrm{BKD}})\\) , with mixed distillation (logitbased \\(^+\\) BKD, \\(+{\\mathcal{L}}_{\\mathrm{MD}}\\) ). Results are tabulated in Tab. 4b. Compared to logitbased KD, applying BKD to the input feature of the last layer shows on-par accuracy improvement on both CNN- and Transformer-based SNNs. Furthermore, thanks to the compatibility between the blurred feature-based KD and logit-based KD, their mixed version shows the maximal accuracy improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In addition to the accuracy obtained for SEW ResNet-18 and Sformer-8-384 discussed above, we also plot their evolution curves of validation accuracy during training as in Fig. 4. We observe that with the introduction of logits-based distillation \\(\\mathcal{L}_{\\mathrm{LD}}\\) (yellow curve) and feature-based blurred knowledge distillation \\(\\mathcal{L}_{\\mathrm{BKD}}\\) (green curve), the performance is significantly improved compared to the vanilla counterpart (blue curve). Moreover, as anticipated, the mixed distillation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "Zekai Xu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 4: Ablation study with multiple settings. \\( R \\), \\( S \\), and \\( KD \\) denotes ResNet, Sformer and Knowledge Distillation respectively. \\( S-4-384-400\\mathrm{E} \\) refers to Sformer-4-384 after 400 epochs training. \\( \\star \\): results from [14], \\( \\diamond \\): results from [62]. Best results are in bold, runner-up results are marked in gray." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.215, + 0.478, + 0.361 + ], + "angle": 0, + "content": "
MethodConfigS-4-384-400ER-19
τλCF10 AccCF100 AccCF10 AccCF100 Acc
+Ld1.5-96.1381.4094.5074.66
2.0-96.1781.2594.5374.57
4.0-96.3180.8894.5674.52
+BKD-0.1596.1381.3894.5274.76
-0.4096.2681.4394.5674.83
-0.6096.3181.5394.6174.89
+MD2.00.1596.1881.6394.6474.95
2.00.4096.2081.6794.6874.98
2.00.6096.37 (+0.11)81.86 (+0.19)94.76 (+0.08)75.07 (+0.09)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.245, + 0.363, + 0.45, + 0.375 + ], + "angle": 0, + "content": "(a) Ablation of \\(\\tau\\) and \\(\\lambda\\) on CIFAR." + }, + { + "type": "table", + "bbox": [ + 0.487, + 0.215, + 0.784, + 0.361 + ], + "angle": 0, + "content": "
MethodCNN-basedTransformer-based
Student SNNTeacher AccSNN AccStudent SNNTeacher AccSNN Acc
w/o KDR-18*-63.18S-8-384-74.35
R-34*-67.04S-8-512-76.54
R-50*-67.78S-8-768-77.64
+LDR-1869.7665.36(+2.18)S-8-38481.7874.69(+0.24)
R-3473.3169.90(+2.86)S-8-51276.99(+0.45)
R-5076.1371.78(+4.00)S-8-76879.04(+1.40)
+BKDR-1869.7665.16(+1.98)S-8-38481.7875.08(+0.73)
R-3473.3170.23(+3.19)S-8-51277.29(+0.75)
R-5076.1371.24(+3.46)S-8-76879.64(+2.00)
+MDR-1869.7665.60(+2.42)S-8-38481.7875.48(+1.13)
R-3473.3171.24(+4.20)S-8-51277.24(+0.70)
R-5076.1372.32(+4.54)S-8-76879.93(+2.29)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.363, + 0.765, + 0.375 + ], + "angle": 0, + "content": "(b) Ablation of multiple losses on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.38, + 0.428, + 0.478 + ], + "angle": 0, + "content": "
MethodSEW R-18 Top1 AccS-8-384 Top1 AccKD location
LaSNN [21]63.3173.85layer-wise
+LBKD63.8274.47layer-wise
KDSNN [53]63.4274.62layer-wise
+LBKD63.9474.93layer-wise
KDSNN [53]63.8774.96single-layer
BKDSNN (ours)65.60 (+1.66)75.48 (+0.52)single-layer
" + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.48, + 0.426, + 0.492 + ], + "angle": 0, + "content": "(c) Ablation of BKD on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.437, + 0.38, + 0.783, + 0.478 + ], + "angle": 0, + "content": "
MethodλSpikingformer-4-384-400EResNet-19
CIFAR10CIFAR100CIFAR10CIFAR100
AccAccAccAcc
w Gw/o Gw Gw/o Gw Gw/o Gw Gw/o G
+LBKD0.0095.3894.3880.4476.6494.4793.4274.6870.97
0.1596.1394.6081.3876.8194.5293.5874.7671.22
0.4096.2694.8281.4377.1494.5693.6474.8371.29
0.6096.3194.8981.5377.1694.6193.7274.8971.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.48, + 0.711, + 0.491 + ], + "angle": 0, + "content": "(d) Ablation of \\(\\lambda\\) and \\(\\mathcal{G}\\) on CIFAR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.789, + 0.562 + ], + "angle": 0, + "content": "(red curve) in KDSNN leads to faster convergence with higher validation accuracy during the training. These experimental results highlight the important role of logits-based and feature-based knowledge from the teacher ANN in improving the performance of student SNNs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.658 + ], + "angle": 0, + "content": "Ablation of BKD and its location We have reported the KDSNN accuracy under different settings in Tab. 4b where KD was applied to logits, the last layer feature, or both. Two more questions still remain unanswered: 1) whether BKD outperforms the vanilla implementation? 2) Is applying KD/BKD to the last layer better than the layer-wise implementation?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To answer these questions, we perform experiments whose results are reported in Tab. 4c. LaSNN [21] and KDSNN [53] are two recent works of layer-wise implementation of KD for SNN training. We replace the original feature-based distillation loss in LaSNN and KDSNN with BKD loss \\(\\mathcal{L}_{\\mathrm{BKD}}\\). In LaSNN, the replacement improves accuracy by \\(0.51\\%\\) on SEW ResNet-18 and by \\(0.62\\%\\) on Sformer-8-384. In KDSNN, the replacement improves accuracy by \\(0.52\\%\\) on SEW ResNet-18 and by \\(0.31\\%\\) on Sformer-8-384. These experiments confirm the advantages of the blurred implementation of KD. Moreover, we further test layer-wise (every layer feature) or single-layer (intermediate feature before the last layer) implementations of KD. We observe similar accuracy boosts for the two implementations. Thus, we conclude that applying BKD only to the intermediate feature map before the last layer is the adequate setup." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Enhancing SNN with Blurred Knowledge Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.406, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.241, + 0.371, + 0.252 + ], + "angle": 0, + "content": "(a) SEW ResNet-18" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.146, + 0.596, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.241, + 0.562, + 0.252 + ], + "angle": 0, + "content": "(b) SEW ResNet-34" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.146, + 0.788, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.241, + 0.75, + 0.252 + ], + "angle": 0, + "content": "(c) SEW ResNet-50" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.255, + 0.407, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.35, + 0.365, + 0.361 + ], + "angle": 0, + "content": "(d) Sformer-8-384" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.255, + 0.596, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.35, + 0.555, + 0.361 + ], + "angle": 0, + "content": "(e) Sformer-8-512" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.255, + 0.787, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.35, + 0.744, + 0.361 + ], + "angle": 0, + "content": "(f) Sformer-8-768" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Fig. 4: Evolution curve of validation accuracy during training, with CNN- (top 3 sub-figures) and Transformer-based (bottom 3 sub-figures) SNNs on ImageNet. The labels in the legends are the same as in Tab. 4b." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.651 + ], + "angle": 0, + "content": "Ablation of temperature coefficient, blurred ratio and restoration block. To examine the optimal setting of hyper-parameters, i.e., temperature \\(\\tau\\) in Eqs. (7) and (8), blurred ration \\(\\lambda\\) in Eq. (2) and restoration block \\(\\mathcal{G}\\) in Eq. (4), we sweep \\(\\tau\\), \\(\\lambda\\) and \\(\\mathcal{G}\\) with ResNet-19 and Sformer-4-384, under varying BKDSNN settings. The results of the CIFAR-10 and CIFAR-100 datasets are reported in Tab. 4a and Tab. 4d. Note that we increase the training epochs of Sformer-4-384 to 400 for more distinct results. For logits-based distillation \\(\\mathcal{L}_{\\mathrm{LD}}\\), we find that flatter logits distribution (with higher \\(\\tau\\)) is better for classification with smaller categories (CIFAR10), while sharper logits distribution (with lower \\(\\tau\\)) is more effective for classification with larger categories (CIFAR100). For BKD \\(\\mathcal{L}_{\\mathrm{BKD}}\\) and mixed distillation \\(\\mathcal{L}_{\\mathrm{MD}}\\), training with a higher blurred ratio \\(\\lambda\\) (i.e., 0.60) achieves greater accuracy across various datasets and models. As reported in Tab. 4d, with the introduction of blurred ration \\(\\lambda\\) and restoration block \\(\\mathcal{G}\\), student SNN better mimics teacher ANN and conveys more sufficient information, which enhances performance on CIFAR-10 and CIFAR-100 with both CNN- and Transformer-based models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.434, + 0.687 + ], + "angle": 0, + "content": "4.5 Firing Rate Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Firing Rate. We compare the firing rate \\( r \\) of the intermediate feature before last layer on CIFAR with Sformer-4-384 and ResNet-19 under multiple settings. The firing rate \\( r \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.753, + 0.786, + 0.796 + ], + "angle": 0, + "content": "\\[\nr = \\frac {1}{b c ^ {\\mathrm {s t u}} h w} \\sum_ {i} ^ {b} \\sum_ {j} ^ {c ^ {\\mathrm {s t u}}} \\sum_ {k} ^ {h} \\sum_ {l} ^ {w} \\tilde {\\mathbf {F}} _ {i, j, k, l} ^ {\\mathrm {s t u}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "As reported in Tab. 5a, it reveals that on both CNN- and Transformer-based models, setting a higher firing rate \\( r \\) achieves higher accuracy. Note that, with the introduction of \\( \\mathcal{L}_{\\mathrm{BKD}} \\) and \\( \\mathcal{L}_{\\mathrm{MD}} \\), high blurred ratio \\( \\lambda \\) encourages model to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "Zekai Xu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 5: Firing rate \\( r \\) under multiple settings on CIFAR and Energy consumption of ANN and SNN on ImageNet. \\( T \\) denotes Transformer." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.274 + ], + "angle": 0, + "content": "
MethodConfigτλS-4-384-400ER-19TimeStepSOPS(G)↓Energy Consumption (mJ)(↓)Top1Acc (↑)
CF10r (e-2)CF10AccCF10e(10)e-2CF10AccCF10e(10)e-2CF10AccCF10e(10)e-2CF10Acc
+LLD1.5-5.5596.131.5281.403.9494.500.9374.66R-50 [19]14.12 (FLOPS)18.9576.13
2.0-5.7296.171.3281.254.0394.531.0174.57Spiking R-50 [14]45.31 (28.88%↑)5.44 (71.29%↓)57.66 (-13.47)
4.0-6.0796.311.1880.884.1194.561.0874.52SEW R-50 [14]44.83 (17.23%↑)4.89 (74.20%↓)67.78 (-8.35)
+LBDK-0.154.6796.130.9781.383.8294.520.8274.76BKDSNN (ours)44.81 (16.75%↑)4.85 (74.40%↓)72.32 (-3.81)
-0.405.3596.261.0381.433.9494.560.8974.83
-0.605.5596.311.1481.534.0894.611.0274.89
+LMD2.0 0.155.3996.181.1481.633.9294.641.0374.95T-8-768 [12]118.60 (FLOPS)85.5681.16
2.0 0.406.1796.201.3481.673.9994.681.0774.98S-8-768 [61]412.54 (32.58%↓)13.68 (84.01%↓)75.85 (-5.31)
2.0 0.606.6696.371.5581.864.1894.761.1375.07S-8-768 + CML [62]412.30 (33.87%↓)13.42 (84.32%↓)77.64 (-3.52)
BKDSNN (ours)412.60 (32.26%↓)13.72 (83.96%↓)79.93 (-1.23)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.264, + 0.276, + 0.428, + 0.287 + ], + "angle": 0, + "content": "(a) Firing rate \\( r \\) on CIFAR." + }, + { + "type": "table_footnote", + "bbox": [ + 0.521, + 0.275, + 0.744, + 0.287 + ], + "angle": 0, + "content": "(b) Energy consumption on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.787, + 0.354 + ], + "angle": 0, + "content": "fire more spikes. Furthermore, compared with \\(\\mathcal{L}_{\\mathrm{LD}}\\), our \\(\\mathcal{L}_{\\mathrm{BKD}}\\) reduces the firing rate while improving the performance, which assists the energy consumption reduction. Meanwhile, in virtue of \\(\\mathcal{L}_{\\mathrm{MD}}\\), BKDSNN significantly improves the performance with a slight boost on firing rate, which leads to state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.372, + 0.513, + 0.388 + ], + "angle": 0, + "content": "4.6 Energy Consumption Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.531 + ], + "angle": 0, + "content": "As tabulated in Tab. 5b, we calculate the energy consumption of BKDSNN according to the energy model in [25,29,65] and compare it with its ANN counterparts and SNNs trained by previous learning-based methods. In terms of the OPS, SNN variants are higher than ANNs, while ANNs use FLOPS as the operation counting metric. Nevertheless, since the energy per synaptic operation is more efficient than the operation in ANN. For ResNet-50, BKDSNN consumes the lowest energy while optimally suppressing accuracy loss. For Sformer-8-768, our BKDSNN consumes more energy to increase firing rates for maintaining similar performance as SOTA methods for SNN training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.551, + 0.36, + 0.567 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.789, + 0.716 + ], + "angle": 0, + "content": "BKDSNN constructs a knowledge distillation based framework for learning-based SNN training, achieving promising performance on both static and neuromorphic datasets supported by comprehensive experiments. We expect that BKDSNN will contribute to future investigations on brain-inspired SNN, which can further narrow the performance gap between SNN and ANN. In this work, we mainly focus on learning-based SNN training methods due to their remarkable performance under ultra-low inference time-step. We anticipate to extend BKDSNN on conversion-based methods, which is expected to overcome the existing bottlenecks of conversion-based methods under ultra-low inference time-step." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.736, + 0.405, + 0.754 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.803, + 0.842 + ], + "angle": 0, + "content": "This work is partially supported by National Key R&D Program of China (2022YFB4500200), National Natural Science Foundation of China (No. 62102257), Biren Technology-Shanghai Jiao Tong University Joint Laboratory Open Research Fund, Microsoft Research Asia Gift Fund, Lingang Laboratory Open Research Fund (No.LGQS-202202-11)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Enhancing SNN with Blurred Knowledge Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.228 + ], + "angle": 0, + "content": "1. Amir, A., Taba, B., Berg, D., Melano, T., McKinstry, J., Di Nolfo, C., Nayak, T., Andreopoulos, A., Garreau, G., Mendoza, M., et al.: A low power, fully event-based gesture recognition system. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7243-7252 (2017) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.229, + 0.786, + 0.269 + ], + "angle": 0, + "content": "2. Bu, T., Fang, W., Ding, J., Dai, P., Yu, Z., Huang, T.: Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347 (2023) 2, 3, 9, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.27, + 0.786, + 0.31 + ], + "angle": 0, + "content": "3. Cao, Y., Chen, Y., Khosla, D.: Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision 113, 54-66 (2015) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.786, + 0.364 + ], + "angle": 0, + "content": "4. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9233-9242 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.786, + 0.405 + ], + "angle": 0, + "content": "5. Chen, L., Yu, C., Chen, L.: A new knowledge distillation for incremental object detection. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.786, + 0.446 + ], + "angle": 0, + "content": "6. Davies, M., Srinivasa, N., Lin, T.H., Chinya, G., Cao, Y., Choday, S.H., Dimou, G., Joshi, P., Imam, N., Jain, S., et al.: Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro 38(1), 82-99 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.786, + 0.487 + ], + "angle": 0, + "content": "7. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.786, + 0.514 + ], + "angle": 0, + "content": "8. Deng, S., Gu, S.: Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476 (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.786, + 0.54 + ], + "angle": 0, + "content": "9. Deng, S., Li, Y., Zhang, S., Gu, S.: Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946 (2022) 8, 9, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.786, + 0.581 + ], + "angle": 0, + "content": "0. Diehl, P., Neil, D., Binas, J., Cook, M., Liu, S.C., Pfeiffer, M.: Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing (07 2015). https://doi.org/10.1109/IJCNN.2015.7280696 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.786, + 0.622 + ], + "angle": 0, + "content": "1. Ding, J., Bu, T., Yu, Z., Huang, T., Liu, J.: Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training. Advances in Neural Information Processing Systems 35, 24780-24793 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.676 + ], + "angle": 0, + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.786, + 0.744 + ], + "angle": 0, + "content": "3. Fang, W., Chen, Y., Ding, J., Yu, Z., Masquelier, T., Chen, D., Huang, L., Zhou, H., Li, G., Tian, Y.: Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances 9(40), eadi1480 (2023). https://doi.org/10.1126/sciadv.adi1480, https://www.science.org/doi/abs/10.1126/sciadv.adi1480 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.786, + 0.785 + ], + "angle": 0, + "content": "4. Fang, W., Yu, Z., Chen, Y., Huang, T., Masquelier, T., Tian, Y.: Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems 34, 21056-21069 (2021) 2, 3, 4, 7, 8, 9, 10, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.786, + 0.838 + ], + "angle": 0, + "content": "5. Garg, I., Chowdhury, S.S., Roy, K.: Dct-snn: Using dct to distribute spatial information over time for low-latency spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4671-4680 (2021) 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "Zekai Xu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Gerstner, W., Kistler, W.M.: Spiking neuron models: Single neurons, populations, plasticity. Cambridge university press (2002) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.215 + ], + "angle": 0, + "content": "17. Han, B., Roy, K.: Deep spiking neural network: Energy efficiency through time based coding. In: European Conference on Computer Vision. pp. 388-404. Springer (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.271 + ], + "angle": 0, + "content": "18. Han, B., Srinivasan, G., Roy, K.: Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13558-13567 (2020) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.272, + 0.785, + 0.312 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.313, + 0.785, + 0.339 + ], + "angle": 0, + "content": "20. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.34, + 0.785, + 0.379 + ], + "angle": 0, + "content": "21. Hong, D., Shen, J., Qi, Y., Wang, Y.: Lasnn: Layer-wise ann-to-snn distillation for effective and efficient training in deep spiking neural networks. arXiv preprint arXiv:2304.09101 (2023) 2, 4, 5, 7, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.38, + 0.785, + 0.419 + ], + "angle": 0, + "content": "22. Hongmin, L., Hanchao, L., Xiangyang, J., Guoqi, L., Luping, S.: Cifar10-dvs: An event-stream dataset for object classification. Frontiers in Neuroscience 11 (2017) 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.704, + 0.433 + ], + "angle": 0, + "content": "23. Hu, Y., Tang, H., Pan, G.: Spiking deep residual network (2020) 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.46 + ], + "angle": 0, + "content": "24. Hu, Y., Zheng, Q., Jiang, X., Pan, G.: Fast-snn: Fast spiking neural network by converting quantized ann. arXiv preprint arXiv:2305.19868 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "25. Hu, Y., Deng, L., Wu, Y., Yao, M., Li, G.: Advancing spiking neural networks towards deep residual learning (2023) 14, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "26. Khan, A., Sohail, A., Zahoora, U., Qureshi, A.S.: A survey of the recent architectures of deep convolutional neural networks. Artificial intelligence review 53, 5455-5516 (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.569 + ], + "angle": 0, + "content": "27. Kheradpisheh, S.R., Ganjtabesh, M., Thorpe, S.J., Masquelier, T.: Stdp-based spiking deep convolutional neural networks for object recognition. Neural Networks 99, 56-67 (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.596 + ], + "angle": 0, + "content": "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "29. Kundu, S., Pedram, M., Beerel, P.A.: Hire-snn: Harnessing the inherent robustness of energy-efficient deep spiking neural networks by training with crafted input noise (2021) 14, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "30. Kushawaha, R.K., Kumar, S., Banerjee, B., Velmurugan, R.: Distilling spikes: Knowledge distillation in spiking neural networks. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 4536-4543. IEEE (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.679, + 0.785, + 0.704 + ], + "angle": 0, + "content": "31. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.732 + ], + "angle": 0, + "content": "32. Lee, D., Park, S., Kim, J., Doh, W., Yoon, S.: Energy-efficient knowledge distillation for spiking neural networks. arXiv preprint arXiv:2106.07172 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.785, + 0.758 + ], + "angle": 0, + "content": "33. Li, C., Ma, L., Furber, S.: Quantization framework for fast spiking neural networks. Frontiers in Neuroscience 16, 918793 (2022) 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "34. Li, Y., Deng, S., Dong, X., Gong, R., Gu, S.: A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In: International conference on machine learning. pp. 6316-6325. PMLR (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "35. Liu, F., Zhao, W., Chen, Y., Wang, Z., Yang, T., Li, J.: Sstdp: Supervised spike timing dependent plasticity for efficient spiking neural network training. Frontiers in neuroscience p. 1413 (2021) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.378, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Enhancing SNN with Blurred Knowledge Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "36. Liu, Y., Chen, K., Liu, C., Qin, Z., Luo, Z., Wang, J.: Structured knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2604-2613 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.219 + ], + "angle": 0, + "content": "37. Maass, W.: Networks of spiking neurons: the third generation of neural network models. Neural networks 10(9), 1659-1671 (1997) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.219, + 0.787, + 0.275 + ], + "angle": 0, + "content": "38. Merolla, P.A., Arthur, J.V., Alvarez-Icaza, R., Cassidy, A.S., Sawada, J., Akopyan, F., Jackson, B.L., Imam, N., Guo, C., Nakamura, Y., et al.: A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345(6197), 668-673 (2014) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "39. Neftci, E.O., Mostafa, H., Zenke, F.: Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36(6), 51-63 (2019) 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.317, + 0.787, + 0.36 + ], + "angle": 0, + "content": "40. Orchard, G., Jayawant, A., Cohen, G.K., Thakor, N.: Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience 9, 437 (2015) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.36, + 0.787, + 0.389 + ], + "angle": 0, + "content": "41. Ostojic, S.: Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons. Nature neuroscience 17(4), 594-600 (2014) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.389, + 0.787, + 0.431 + ], + "angle": 0, + "content": "42. Peng, C., Zhao, K., Maksoud, S., Li, M., Lovell, B.C.: Sid: Incremental learning for anchor-free object detection via selective and inter-related distillation. Computer vision and image understanding 210, 103229 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.431, + 0.787, + 0.473 + ], + "angle": 0, + "content": "43. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.787, + 0.502 + ], + "angle": 0, + "content": "44. Roy, K., Jaiswal, A., Panda, P.: Towards spike-based machine intelligence with neuromorphic computing. Nature 575(7784), 607-617 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.502, + 0.787, + 0.544 + ], + "angle": 0, + "content": "45. Rueckauer, B., Lungu, I.A., Hu, Y., Pfeiffer, M., Liu, S.C.: Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in neuroscience 11, 682 (2017) 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.544, + 0.787, + 0.613 + ], + "angle": 0, + "content": "46. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision 128(2), 336-359 (Oct 2019). https://doi.org/10.1007/s11263-019-01228-7, http://dx.doi.org/10.1007/s11263-019-01228-7 7, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.655 + ], + "angle": 0, + "content": "47. Sengupta, A., Ye, Y., Wang, R., Liu, C., Roy, K.: Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience 13, 95 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.656, + 0.787, + 0.713 + ], + "angle": 0, + "content": "48. Sironi, A., Brambilla, M., Bourdis, N., Lagorce, X., Benosman, R.: Hats: Histograms of averaged time surfaces for robust event-based object classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1731-1740 (2018) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.713, + 0.787, + 0.742 + ], + "angle": 0, + "content": "49. Sun, S., Cheng, Y., Gan, Z., Liu, J.: Patient knowledge distillation for bert model compression. arXiv preprint arXiv:1908.09355 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.742, + 0.787, + 0.785 + ], + "angle": 0, + "content": "50. Takuya, S., Zhang, R., Nakashima, Y.: Training low-latency spiking neural network through knowledge distillation. In: 2021 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPS). pp. 1-3. IEEE (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "51. Tan, G., Wang, Y., Han, H., Cao, Y., Wu, F., Zha, Z.J.: Multi-grained spatiotemporal features perceived network for event-based lip-reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20094-20103 (2022) 6" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "Zekai Xu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "52. Wu, Y., Deng, L., Li, G., Zhu, J., Xie, Y., Shi, L.: Direct training for spiking neural networks: Faster, larger, better. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1311-1318 (2019) 2, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "53. Xu, Q., Li, Y., Shen, J., Liu, J.K., Tang, H., Pan, G.: Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7886-7895 (2023) 2, 4, 5, 7, 9, 10, 11, 12, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.3 + ], + "angle": 0, + "content": "54. Xu, Q., Li, Y., Shen, J., Zhang, P., Liu, J.K., Tang, H., Pan, G.: Hierarchical spiking-based model for efficient image classification with enhanced feature extraction and encoding. IEEE Transactions on Neural Networks and Learning Systems (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.328 + ], + "angle": 0, + "content": "55. Xu, Z., Zhang, M., Hou, J., Gong, X., Wen, C., Wang, C., Zhang, J.: Delving into transformer for incremental semantic segmentation (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "56. Yang, Z., Li, Z., Jiang, X., Gong, Y., Yuan, Z., Zhao, D., Yuan, C.: Focal and global knowledge distillation for detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4643-4652 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.424 + ], + "angle": 0, + "content": "57. Yao, M., Hu, J., Hu, T., Xu, Y., Zhou, Z., Tian, Y., Bo, X., Li, G.: Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In: The Twelfth International Conference on Learning Representations (2023) 4, 8, 10, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.467 + ], + "angle": 0, + "content": "58. Zenke, F., Agnes, E.J., Gerstner, W.: Diverse synaptic plasticity mechanisms orchestrated to form and retrieve memories in spiking neural networks. Nature communications 6(1), 6922 (2015) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "59. Zenke, F., Vogels, T.P.: The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation 33(4), 899-925 (2021) 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.55 + ], + "angle": 0, + "content": "60. Zheng, H., Wu, Y., Deng, L., Hu, Y., Li, G.: Going deeper with directly-trained larger spiking neural networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 11062-11070 (2021) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.592 + ], + "angle": 0, + "content": "61. Zhou, C., Yu, L., Zhou, Z., Zhang, H., Ma, Z., Zhou, H., Tian, Y.: Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954 (2023) 2, 3, 4, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.647 + ], + "angle": 0, + "content": "62. Zhou, C., Zhang, H., Zhou, Z., Yu, L., Ma, Z., Zhou, H., Fan, X., Tian, Y.: Enhancing the performance of transformer-based spiking neural networks by improved downsampling with precise gradient backpropagation. arXiv preprint arXiv:2305.05954 (2023) 2, 3, 4, 8, 9, 10, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.675 + ], + "angle": 0, + "content": "63. Zhou, D., Kang, B., Jin, X., Yang, L., Lian, X., Jiang, Z., Hou, Q., Feng, J.: Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886 (2021) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.716 + ], + "angle": 0, + "content": "64. Zhou, H., Song, L., Chen, J., Zhou, Y., Wang, G., Yuan, J., Zhang, Q.: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. arXiv preprint arXiv:2102.00650 (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.786, + 0.744 + ], + "angle": 0, + "content": "65. Zhou, Z., Zhu, Y., He, C., Wang, Y., Yan, S., Tian, Y., Yuan, L.: Spikformer: When spiking neural network meets transformer (2022) 2, 3, 4, 9, 14, 5" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.744 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_origin.pdf b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e49fb575a59a889bf5f16541a3f62079f140b8e --- /dev/null +++ b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/1b8780bf-0cd6-4729-8941-c8d2548b5afe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07b0b4c5d63d22224d87a7d17471c05abd46471ce8453c27e88bb225c88208ca +size 2348822 diff --git a/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/full.md b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f304c185ece780afb946ffc1b160c1e1df52bc08 --- /dev/null +++ b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/full.md @@ -0,0 +1,352 @@ +# BKDSNN: Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation + +Zekai $\mathrm{Xu}^{1}$ , Kang You $^{1}$ , Qinghai Guo $^{2}$ , Xiang Wang $^{2}$ , and Zhezhi He $^{1\dagger}$ + +$^{1}$ School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China + +$^{2}$ Huawei Technologies, Shenzhen, China + +{sherlock.holmes.xu,kang_you, $^\dagger$ zhezhi.he}@sjtu.edu.cn + +https://github.com/Intelligent-Computing-Research-Group/BKDSNN + +Abstract. Spiking neural networks (SNNs), which mimic biological neural systems to convey information via discrete spikes, are well-known as brain-inspired models with excellent computing efficiency. By utilizing the surrogate gradient estimation for discrete spikes, learning-based SNN training methods that can achieve ultra-low inference latency (i.e., number of time-step) have emerged recently. Nevertheless, due to the difficulty of deriving precise gradient for discrete spikes in learning-based methods, a distinct accuracy gap persists between SNNs and their artificial neural networks (ANNs) counterparts. To address the aforementioned issue, we propose a blurred knowledge distillation (BKD) technique, which leverages randomly blurred SNN features to restore and imitate the ANN features. Note that, our BKD is applied upon the feature map right before the last layer of SNNs, which can also mix with prior logits-based knowledge distillation for maximal accuracy boost. In the category of learning-based methods, our work achieves state-of-the-art performance for training SNNs on both static and neuromorphic datasets. On the ImageNet dataset, BKDSNN outperforms prior best results by $4.51\%$ and $0.93\%$ with the network topology of CNN and Transformer, respectively. + +Keywords: Learning-based Spiking Neural Networks $\cdot$ Knowledge Distillation + +# 1 Introduction + +Inspired by the neural firing mechanism in the biological nervous system, spiking neural network (SNN) [37] emerges as an alternative framework for the traditional artificial neural network (ANN). ANNs rely on continuous activation values to convey information between neurons [31], while SNNs utilize discrete spiking events to process and propagate information [6,38]. Given the remarkable energy efficiency achieved by such event-driven information propagation [18], SNNs + +show promising prospects on computational intelligence tasks [44] with strong autonomous learning capabilities and ultra-low power consumption [2,11,41,58]. Apart from these advantages, SNNs are also natively compatible with event-based neuromorphic datasets (e.g., CIFAR10-DVS) [14,61,62,65]. + +Currently, conversion-based and learning-based are two common methods to train SNNs. Conversion-based methods attempt to leverage information from ANNs by transferring the parameters of a pre-trained ANN to its SNN counterpart, which have achieved promising results in several tasks [2, 8, 33, 34]. However, event-triggered neurons convey limited information with ultra-low inference time-steps (e.g., 4 time-steps), leading to degraded performance of conversion-based methods [14, 65]. Consequently, conversion-based methods usually take a long inference time (e.g., 32 time-steps) to get comparable accuracy as the original ANN [3, 47] based on the equivalence of the ReLU activation and the firing rate of integrate-and-fire (IF) neuron [16]. In order to further improve the performance of SNNs with ultra-low inference time-steps, learning-based methods unfold SNN training over the inference time-steps and leverage back-propagation through time (known as BPTT) [52, 59] to train SNNs. Inspired by [27], several works [14, 35, 62] adopt surrogate gradient estimation as an implicit differentiation for direct training of SNNs. With the combination of BPTT and surrogate gradient estimation, learning-based methods exhibit advantages in processing event-based neuromorphic datasets (e.g., CIFAR10-DVS [22]) [14, 61, 62, 65]. Furthermore, direct training of CNN- [14, 23] and Transformer-based [61, 62, 65] SNNs also achieves promising performance with ultra-low time-steps on static datasets (e.g., CIFAR10 [28], CIFAR100 [28] and ImageNet [7]). + +Despite that numerous efforts have been made in learning-based methods, SNNs still need improvement compared with ANNs. [53, 62] claim that imprecise surrogate gradient estimation of spiking neurons leads to performance degradation. On the other hand, KDSNN [53] and LaSNN [21] introduce layer-wise knowledge distillation (KD) to train a student SNN with rich information from the ANN teacher, achieving promising performance on several classical vision datasets (MNIST, CIFAR10/100) with CNN-based models. However, according to our preliminary experimental results, both KDSNN and LaSNN struggle with more complex datasets (ImageNet) and Transformer-based models. We hypothesize that the direct distillation between discrete SNN features and continuous ANN features [21, 53] might be the source of the limitations. + +In this paper, to further alleviate performance degradation of learning-based SNNs, we introduce a blurred knowledge distillation (BKD) technique to leverage randomly blurred SNN features to restore and mimic ANN features. We apply BKD to the intermediate feature before the last layer of the student SNN and prove its superiority over layer-wise knowledge distillation in learning-based SNNs. Furthermore, we combine prior logits-based distillation with BKD, which provides mutual benefits for improving performance. We evaluate BKDSNN on both static datasets (i.e., CIFAR10, CIFAR100, and ImageNet) and a neuromorphic dataset (i.e., CIFAR10-DVS), where we achieve state-of-the-art (SOTA) performance with both CNN- and Transformer-based SNN models. + +To summarize, our contributions are threefold: + +- We introduce a blurred knowledge distillation (BKD) technique to guide the training of student SNNs through feature imitation and demonstrate the advantages of BKD on learning-based SNNs. +- We apply BKD to the intermediate feature before the last layer and prove its superiority over previous layer-wise distillation. Combined with prior logits-based knowledge distillation, our BKDSNN maximizes accuracy boost on learning-based SNNs. +- We benchmark our method on both static datasets (i.e., CIFAR10, CIFAR100 and ImageNet) and a neuromorphic dataset (i.e., CIFAR10-DVS). Our method outperforms prior SOTA learning-based methods for both CNN-and Transformer-based models. + +# 2 Background and Related Work + +Spiking Neuron Model. Previous works [2,3,8,14,24,61,62,65] have widely used the classic integrate-and-fire (IF) neuron as the fundamental unit in SNNs. IF neuron integrates input currents through changes in its membrane potential, which is compared against the threshold for spike generation. The membrane potential is reset after emitting a spike. The dynamics of the IF neuron is depicted as follows: + +$$ +H [ t ] = V [ t - 1 ] + \frac {1}{\tau} \left(I [ t ] - (V [ t - 1 ] - V _ {\mathrm {r e s e t}})\right) +$$ + +$$ +S [ t ] = \Theta (H [ t ] - V _ {\mathrm {t h}}) = \left\{ \begin{array}{l l} 1, & H [ t ] \geq V _ {\mathrm {t h}} \\ 0, & H [ t ] < V _ {\mathrm {t h}} \end{array} \right. \tag {1} +$$ + +$$ +V [ t ] = H [ t ] (1 - S [ t ]) + V _ {\text {r e s e t}} S [ t ] +$$ + +where $\tau$ is the membrane time constant, $I[t]$ is the receiving current at time step $t$ and $\Theta (\cdot)$ is the heaviside function. The neuron accumulates $I[t]$ to change the membrane potential $H[t]$ at time step $t$ . Once $H[t]$ exceeds the firing threshold $V_{\mathrm{th}}$ , the neuron generates a spike and resets the membrane potential to $V_{\mathrm{reset}}$ . + +Conversion-based Methods. To fully make use of the knowledge from pre-trained ANN, conversion-based methods are designed to transfer the learned parameters of ANNs into corresponding SNNs, without further training of SNNs. Previous works [17, 18] replace ReLU activations in ANNs with spiking neurons and add scaling operations like weight normalization and threshold balancing [10] to generate corresponding SNNs. Although several works [8, 17, 18, 34] have achieved nearly loss-less accuracy with VGG-16 and ResNet, the converted SNNs still require long inference time to match the original ANNs in precision. [45] claims that low firing rates caused by weight normalization result in increasing latency for the information reaching higher layers, which limits the practical application of the converted SNNs. + +![](images/9bd6ac2317b1c9e4bbfbdd0bd2e2823cffac267c74d3507665209de20b5e90ac.jpg) +Fig.1: Overview of training SNN with blurred knowledge distillation (BKD). The SNN is directly trained with back-propagation through time (BPTT) [52], where we utilize a blurred variant of KD (i.e., BKD) to achieve higher accuracy. BKD highlighted in yellow shaded region differs from prior SNN-oriented KD in three perspectives: 1) A blurred matrix $\pmb{B}$ is randomly sampled on the fly (per input image) to mask out the feature of student SNN; 2) A restoration block $\mathcal{G}$ consisting of two convolutional layers connected by a ReLU layer is applied on blurred SNN features to restore and mimic ANN features; 3) Such blurred knowledge distillation is applied only to the intermediate features before the last layer. + +Learning-based Methods. With the development of surrogate gradient estimation [39], learning-based methods leverage continuous-value-based loss functions to handle nondifferentiable difficulties in SNN training. Ref. [39] introduces surrogate gradient estimation to conduct SNN training on the event-based DVS gesture dataset, which achieves better accuracy after a few training iterations. In an effort to improve computational resources in SNNs, joint CNN-SNN [54] puts forward a novel augmented spiking-based framework and achieves promising results in several image classification tasks. DCT-SNN [15] reduces the number of inference time-steps by combining frequency-domain learning with surrogate gradient descent. Spiking ResNet [60] uses the directly-trained deep SNN to achieve high performance on ImageNet. SEW ResNet [14] proposes the spike-element-wise block to counter gradient vanishing and gradient exploding, which successfully extends the depth of directly-trained SNN to more than 100 layers. When it comes to Transformer-based SNN, Spikformer [65] with a novel spiking self-attention block achieves significant improvement in image classification on both neuromorphic and static datasets. Transformer-based SNN is further improved by successive Spikingformer [61], ConvBN-MaxPooling-LIF (CML) [62] and Meta-SpikeFormer [57], achieving SOTA performance. + +Knowledge Distillation for SNN Training. Knowledge distillation (KD) is a classical transfer learning method that has been proven to be effective on various tasks, such as model compression [20, 36, 49, 56], incremental learning [4, 5, 42, 55], etc. Recently, several works [30, 32, 50] extend KD for SNN training, such as distilling spikes from a large SNN to a small SNN or logits-based knowledge from a well-trained ANN to a simple SNN. KDSNN [53] proposes a novel framework embracing both logits-based and feature-based knowledge distillation, which enhances performance on several image classification datasets (MNIST, CIFAR10, and CIFAR100). LaSNN [21] puts forward a layer-wise feature-based + +Table 1: Summary of notations in this paper + +
NotationDescription
Ttime-step
λblurred ratio
R∈Rb×ctearandom matrix R~U(0,1)
B∈Rb×cteablurred matrix
ˆy∈Rb×nclsground truth label
ystu,ytea∈Rb×nclsoutput logits of student SNN and teacher ANN
qstu,qtea∈Rb×nclsystuand ytea after flattened processing
Ftea∈Rb×ctea×h×wfeature before the last layer of teacher ANN
Fstu∈RT×b×cstu×h×wfeature before the last layer of student SNN
Fstu∈Rb×cstu×h×waverage of Fstu at time-step dimension (T)
fadpadaptive layer to align Fstu with Ftea
Grestoration block for Fstu
ˆFstu∈Rb×ctea×h×wblurred restoration feature from Fstu
+ +ANN-to-SNN distillation framework, achieving comparable top-1 accuracy to ANNs on the challenging Tiny ImageNet dataset. Furthermore, KDSNN [53] and LaSNN [21] are applicable to both homogeneous and heterogeneous networks. Despite the superiority of KDSNN [53] and LaSNN [21], they struggle with more complex datasets (e.g., ImageNet) with CNN-based and Transformer-based models according to our preliminary experimental results in Sec. 4.2. + +# 3 Method + +In this paper, we propose a novel blurred knowledge distillation SNN (overall architecture of BKDSNN is in Fig. 1a) training method that leverages knowledge from the teacher ANN to guide the training of the student SNN. As depicted in Fig. 1b, our BKD first introduces a randomly blurred matrix $\pmb{B}$ to mask the student SNN features. Unlike KDSNN [53] and LaSNN [21], which only apply an adaptive layer to align the SNN feature with the ANN feature before distillation, BKD introduces a restoration block $\mathcal{G}$ on blurred SNN feature to restore and mimic ANN feature. The proposed BKD is applied only to the intermediate feature before the last layer, different from the layer-wise distillation [21, 53]. As shown in Fig. 1a, we combine the logits-based distillation with BKD to leverage both logits and feature knowledge from the teacher ANN. + +Blurred Knowledge Distillation. To facilitate understanding, we present important notations in Tab. 1. The proposed blurred knowledge distillation is illustrated in Fig. 1b. We first introduce a random matrix $\pmb{R}$ which obeys uniform distribution between 0 and 1. Then we apply a blurred ratio $\lambda$ to $\pmb{R}$ to generate a blurred matrix $\pmb{B}$ through Eq. (2). + +$$ +\boldsymbol {B} _ {i, j} = \left\{ \begin{array}{l l} 0, & \text {i f} \boldsymbol {R} _ {i, j} < \lambda \\ 1, & \text {O t h e r w i s e} \end{array} \right. \tag {2} +$$ + +![](images/4925b5de32efb33a919fb07dc670420e409e93a90cb6338ae193e59a3c2937be.jpg) +(a) Feature visualization of ResNet-18. + +![](images/6fb5a2b82e913d8ee5e3659413dd8637e02440c423fff0da061be68b47328ff3.jpg) +(b) Feature distribution of SNN and ANN. +Fig.2: (a) Comparison of feature visualization in ResNet-18 under various methods; (b) Histogram of SNN and ANN feature. + +For feature $\mathbf{F}^{\mathrm{stu}}$ at the last layer of student SNN, we first average $\mathbf{F}^{\mathrm{stu}}$ at timestep dimension $(T)$ as: + +$$ +\tilde {\mathbf {F}} ^ {\mathrm {s t u}} = \frac {1}{T} \sum_ {i = 1} ^ {T} \mathbf {F} _ {i} ^ {\mathrm {s t u}} \tag {3} +$$ + +After the averaging operation above, we apply a convolutional adaptive layer $f_{\mathrm{adp}}$ to align $\tilde{\mathbf{F}}^{\mathrm{stu}}$ with $\mathsf{F}_i^{\mathrm{tea}}$ when student feature dimension $c^{\mathrm{stu}}$ mismatches with teacher feature dimension $c^{\mathrm{tea}}$ . After the alignment, we construct a restoration block $\mathcal{G}$ with two convolutional layers connected by a ReLU layer and apply it to $\tilde{\mathbf{F}}^{\mathrm{stu}}$ with the blurred matrix $B$ to achieve blurred restoration feature $\hat{\mathbf{F}}^{\mathrm{stu}}$ as illustrated in Eq. (4). + +$$ +\hat {\mathbf {F}} ^ {\mathrm {s t u}} = \left\{ \begin{array}{l l} \mathcal {G} \left(f _ {\mathrm {a d p}} \left(\tilde {\mathbf {F}} ^ {\mathrm {s t u}}\right) * \boldsymbol {B}\right), & c ^ {\mathrm {s t u}} \neq c ^ {\mathrm {t e a}} \\ \mathcal {G} \left(\tilde {\mathbf {F}} ^ {\mathrm {s t u}} * \boldsymbol {B}\right), & \text {o t h e r w i s e} \end{array} \right. \tag {4} +$$ + +We calculate blurred knowledge distillation loss $\mathcal{L}_{\mathrm{BKD}}$ using quadratic norm $(L^2)$ as follows. + +$$ +\mathcal {L} _ {\mathrm {B K D}} = \sum_ {i} ^ {b} \sum_ {j} ^ {c ^ {\mathrm {s t u}}} \sum_ {k} ^ {h} \sum_ {l} ^ {w} \left(\hat {\mathbf {F}} _ {i, j, k, l} ^ {\mathrm {s t u}} - \mathbf {F} _ {i, j, k, l} ^ {\mathrm {t e a}}\right) ^ {2} \tag {5} +$$ + +We then combine blurred knowledge distillation loss $\mathcal{L}_{\mathrm{BKD}}$ with the original cross entropy loss $\mathcal{L}_{\mathrm{CE}}$ to get the total loss $\mathcal{L}_{\mathrm{total}}$ as follows. + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {C E}} (\boldsymbol {y} ^ {\mathrm {s t u}}, \hat {\boldsymbol {y}}) + \mathcal {L} _ {\mathrm {B K D}} \tag {6} +$$ + +Effectiveness of $BKD$ . For both CNN- and Transformer-based models, features of deeper layers have better representation for the original input image [26,43,63]. In other words, features of deeper layers contain sufficient information of adjacent pixels (for CNN-based models) [26] or patch pairs (for Transformer-based models) [43, 63]. Therefore, we can use partial student SNN feature before the last layer to restore and mimic teacher ANN feature through a restoration block + +$\mathcal{G}$ , which can help the student SNN achieves better representation and effectively mimics teacher ANN. We then apply Grad-CAM which can intuitively demonstrate the feature similarity [46] on each block in ResNet-18 under various methods. As depicted in Fig. 2a, compared with previous SNN (SEW ResNet-18 [14]) and KDSNN [53], our BKDSNN transfers sufficient knowledge from ANN and effectively mimics ANN feature. Meanwhile, as illustrated in Fig. 2b, student SNN feature $\tilde{\mathsf{F}}^{\mathrm{stu}}$ is discrete while teacher ANN feature $\mathsf{F}^{\mathrm{tea}}$ is continuous. Therefore, directly applying distillation to encourage discrete SNN feature $\tilde{\mathsf{F}}^{\mathrm{stu}}$ to imitate continuous ANN feature $\mathsf{F}^{\mathrm{tea}}$ [21, 53] is not plausible. However, the proposed BKD leverages the restoration block $\mathcal{G}$ to map discrete feature $\tilde{\mathsf{F}}^{\mathrm{stu}}$ to continuous feature $\hat{\mathsf{F}}^{\mathrm{stu}}$ for imitating continuous ANN feature $\mathsf{F}^{\mathrm{tea}}$ , which is a more proper spiking-friendly feature-based distillation method. + +Mixed Distillation. As depicted in Fig. 1a, BKDSNN further applies logits-based distillation which transfers knowledge from teacher ANN logits to student SNN logits. As soft logits from the teacher model embrace sufficient hidden knowledge, which serves as a good regularization to student model [64], we introduce the temperature $\tau$ to make the logits distribution flatter. The output logits of both teacher and student are processed as follows: + +$$ +\boldsymbol {q} _ {i} = \frac {\exp \left(\boldsymbol {y} _ {i} / \tau\right)}{\sum_ {j = 1} ^ {n ^ {\mathrm {c l s}}} \exp \left(\boldsymbol {y} _ {j} / \tau\right)}, \quad i \in \{1, 2, \dots , n ^ {\mathrm {c l s}} \} \tag {7} +$$ + +Then logits-based distillation loss $\mathcal{L}_{\mathrm{LD}}$ is defined as: + +$$ +\mathcal {L} _ {\mathrm {L D}} = \tau^ {2} \cdot \mathcal {L} _ {\text {C r o s s E n t r o p y}} \left(\boldsymbol {q} ^ {\text {s t u}}, \boldsymbol {q} ^ {\text {t e a}}\right) \tag {8} +$$ + +We combine $\mathcal{L}_{\mathrm{LD}}$ with $\mathcal{L}_{\mathrm{BKD}}$ to achieve mixed distillation loss $\mathcal{L}_{\mathrm{MD}}$ (Eq. (9)), which contains both logits and feature information from the teacher ANN. + +$$ +\mathcal {L} _ {\mathrm {M D}} = w _ {\mathrm {L D}} \cdot \mathcal {L} _ {\mathrm {L D}} + w _ {\mathrm {B K D}} \cdot \mathcal {L} _ {\mathrm {B K D}} \tag {9} +$$ + +where $w_{\mathrm{LD}}$ and $w_{\mathrm{BKD}}$ are weights for $\mathcal{L}_{\mathrm{LD}}$ and $\mathcal{L}_{\mathrm{BKD}}$ , respectively. Thus, the overall loss $\mathcal{L}$ is defined as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {C r o s s E n t r o p y}} \left(\boldsymbol {y} ^ {\text {s t u}}, \hat {\boldsymbol {y}}\right) + \mathcal {L} _ {\mathrm {M D}} = \mathcal {L} _ {\text {t a s k}} + \mathcal {L} _ {\mathrm {M D}} \tag {10} +$$ + +# 4 Experiments + +# 4.1 Experimental Configuration + +Platform. All experiments are evaluated on a server equipped with 32 cores Intel Xeon Platinum 8352V CPU with $2.10\mathrm{GHz}$ and 8-way NVIDIA GPUs. We use SpikingJelly [13] to simulate IF neurons. + +Static Datasets and Settings. We first evaluate BKDSNN on three static datasets, i.e., CIFAR10/100 and ImageNet. + +CIFAR10/CIFAR100 [28] contains 50k training and 10k testing images with $32 \times 32$ resolution, and is labeled with 10- or 100-categories respectively for classification. For CNN-based models, we use ResNet-19 from TET [9] and lightweight ResNet-20 from SEW ResNet [14] with corresponding ANNs as teacher models. For Transformer-based models, we train ViT-S/16 on CIFAR10 and CIFAR100 to make it serve as the teacher ANN model for Sformer-4-256, Sformer-2-384, and Sformer-4-384 (We abbreviate Spikingformer to Sformer for convenience). Sformer-L-D represents a Sformer model with L Sformer encoder blocks and D feature embedding dimensions. + +ImageNet-1K [7] contains 1.3 millions of training images and 50k validation images with $224 \times 224$ resolution, labeled with 1k classes. We choose ResNet-18, ResNet-34, and ResNet-50 pre-trained on ImageNet-1K as teacher ANN models for the corresponding SEW ResNet models [14]. For Transformer-based models, we introduce ViT-B/16 [12] pre-trained on ImageNet1K as the teacher ANN model for Sformer-8-384, Sformer-8-512 and Sformer-8-768. We also apply our BKDSNN on the latest Meta-SpikeFormer [57] and conduct experiments with Mformer-8-256, Mformer-8-384 and Mformer-8-512 on ImageNet-1K (We abbreviate Meta-SpikeFormer to Mformer for convenience.). + +For both CNN- and Transformer-based models, we set temperature $\tau$ to 2.0, blurred ratio $\lambda$ to 0.15 and logits-based distillation weight $w_{\mathrm{LD}}$ to 1. Specifically for CNN-based models, we set blurred knowledge distillation weight $w_{\mathrm{BKD}}$ to $7e^{-4}$ for CIFAR10/100 and $7e^{-5}$ for ImageNet. We follow all the training configurations (batch size, learning rate, optimizer, etc.) from SEW ResNet [14] except ResNet-19, where we use training configuration from TET [9]. For Transformer-based models, we set $w_{\mathrm{BKD}}$ to $2e^{-3}$ for both CIFAR10/100 and ImageNet, and we follow all the training configurations (batch size, learning rate, optimizer, etc.) from Sformer+CML [62]. + +Neuromorphic Dataset and Networks. We also evaluate BKDSNN on CIFAR10-DVS [22], which is a neuromorphic dataset that leverages dynamic vision sensor (DVS) to convert 10k frame-based images from the CIFAR10 dataset into 10k event streams, providing an event-stream dataset with 10 different classes. It includes 9k training samples and 1k testing samples. Note that, CIFAR10-DVS is an event-stream dataset tailored to SNN, so we use SNN models with 16 timesteps as teacher models to guide the training of SNN models with lower timesteps (i.e., 4 and 8). We utilize Wide-7B-Net [14] and Sformer-2-256 [62] with 4 and 8 inference time-steps as student models for CNN-based and Transformer-based models, respectively. + +For both Wide-7B-Net [14] and former-2-256 [62], we set $\tau$ to 2.0, $\lambda$ to 0.40 and $w_{\mathrm{LD}}$ to 1. Then we set $w_{\mathrm{BKD}}$ to $2e^{-3}$ for Wide-7B-Net [14] and $7e^{-5}$ for Sformer-2-256 [62]. We follow all the training configurations from SEW ResNet [14] for Wide-7B-Net [14], and Sformer+CML [62] for Sformer-2-256 [62]. + +Table 2: Comparison with previous works, using various networks and multiple datasets. $P$ denotes #parameters. $KD$ denotes knowledge distillation. $\star$ : results taken from [14]. †: prior state-of-the-art conversion-based method QCFS [2]. The prior SOTA results are marked in gray, which is used to compare with our BKDSNN and report the accuracy difference. Best results are in bold. - indicates results not available (not reported in original article or cannot reproduce). + +
w/o KD MethodArchitectureR-19 (P: 12.63M)R-20 (P: 0.27M)Time StepArchitecture MethodSformer-4-256 (P: 4.15M)Sformer-2-384 (P: 5.76M)Sformer-4-384 (P: 9.32M)Time Step
CF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 Acc
QCS [2]†--83.7555.234w/o KD Spíkarfer [65]93.9475.9694.8076.9595.1977.864
TET [9]94.4474.4787.7559.144+CML [62]94.8277.6495.6378.7595.9379.654
STBP [60]92.9270.86--4+CML [62]94.7777.4395.2278.3495.6179.094
Spíkarfer-R [23]92.1370.1387.3259.934
SEW-R [14]93.2470.8489.0760.164
Teacher ANN95.6079.7891.7768.401Teacher ANN96.7582.2296.7582.2296.7582.221
KDSNN [53]94.3674.0889.0360.184KDSNN [53]95.0078.3895.5979.2595.8880.334
LaSNN [21]94.2173.9288.4159.174LaSNN [21]94.9778.0295.5578.9195.7979.994
BKDSNN (ours)94.6474.9589.2960.924BKDSNN (ours)95.29 (+0.29)79.41 (+1.03)95.90 (+0.31)80.63 (+1.38)96.06 (+0.18)81.26 (+0.93)4
+ +(a) CNN-based SNN on $CF10 / 100$ dataset. +(b) Transformer-based SNN on $CF10/100$ dataset. + +
w/o KD MethodArchitectureR-18(P: 11.69M)Top1 AccR-34(P: 21.79M)Top1 AccR-50(P: 25.56M)Top1 AccTimeStep
QCFS [2]†-69.37-32
TET [9]-68.00-4
STBP [60]-67.04-4
Spiking-R [23]★62.3261.8657.664
SEW-R [14]63.1867.0467.784
with KD (ours)Teacher ANN69.7673.3176.311
KDSNN [53]63.4267.1867.724
LaSNN [21]63.3166.9467.814
BKDSNN(ours)65.60(+2.18)71.24(+4.06)72.32(+4.51)4
+ +(c) CNN-based SNN on ImageNet dataset. + +
Architecture MethodSformer-8-384(P: 16.81M)Top1 Acc.Sformer-8-512(P: 29.68M)Top1 Acc.Sformer-8-768(P: 66.34M)Top1 Acc.Time Step
w/o KDSpikformer [65]70.2473.3874.814
+CML [62]72.7375.6177.344
Sformer [61]72.4574.7975.854
+CML [62]74.3576.5477.644
with KDTeacher ANN81.7881.7881.781
KDSNN [53]74.6276.4477.834
LaSNN [21]73.8576.3877.664
BKDSNN(ours)75.48(+0.86)77.24(+0.70)79.93(+2.10)4
+ +(d) Transformer-based SNN on ImageNet dataset. + +Settings for Prior KD Methods on SNNs. Furthermore, to compare BKDSNN with previous KD methods on SNNs (i.e., KDSNN [53] and LaSNN [21]), we follow the settings from [53] and [21] to reproduce the experimental results on all datasets with CNN-based and Transformer-based models. + +# 4.2 Comparison on Static Datasets + +Comparison on CIFAR10/100 between BKDSNN and previous methods are listed in Tab. 2a and Tab. 2b, where BKDSNN surpasses prior SOTA methods with both CNN- and Transformer-based models. Specifically, BKD applied upon Sformer-4-384 outperforms previous SOTA by $0.18\%$ on CIFAR10 and $0.93\%$ on CIFAR100, which achieves $96.06\%$ and $81.26\%$ accuracy respectively. + +Intuitively, it is difficult for both KDSNN [53] and LaSNN [21] to improve the performance of CNN-based models on the CIFAR dataset. For ResNet-19, the knowledge of the ANN teacher is limited for the SNN student, leading to the above difficulties. When it comes to ResNet-20, the lightweight model conveys limited information which results in the ineffectiveness of previous knowledge distillation methods. In contrast, our BKDSNN is effective on both limited teacher knowledge as well as lightweight student models, which verifies the superiority of our method. + +Table 3: (a) Comparison with Meta-SpikeFormer [57] on ImageNet. (b) Experimental results on CIFAR10-DVS. Mformer represents Meta-SpikeFormer [57]. $\star$ : SNN with high inference time-step (i.e., 16) serves as the teacher model for corresponding SNN with low inference time-step (i.e., 4 and 8). The prior state-of-the-art results are marked in gray. Best results are in bold. + +
Architecture MethodMformer-8-256(P: 15.11M)Top1 Acc.Mformer-8-384(P: 31.33M)Top1 Acc.Mformer-8-512(P: 55.35M)Top1 Acc.Time StepTransformer-basedCNN-based
MethodAccTime StepAccTime StepMethodAccTime StepAccTime Step
wDMformer [57]71.8275.4077.961Sfomer-2-256+CML [62]★81.41681.416Wide-7B-Net[14]★74.41674.416
74.1177.2078.814
Teacher ANN81.7881.7881.781Sfomer-2-256+CML [62]76.4479.28Wide-7B-Net[14]64.8470.28
with KDMformer [57]--79.111KDSNN [53]78.1479.98KDSNN [53]66.1470.98
--80.004LaSNN [21]78.3480.18LaSNN [21]66.6470.58
BKDSNN(ours)72.9376.1079.821BKDSNN(ours)79.3(+1.0)480.8(+0.7)8BKDSNN(ours)68.3(+1.7)472.2(+1.3)8
+ +![](images/7768b11dc02747ab46c270075a11e43623bda4ab48ca2722b155caf2f5db2d24.jpg) +(a) Meta-SpikeFormer on ImageNet dataset. (b) Experimental results on CIFAR10-DVS. +(a) Feature map visualization. +Fig.3: (a) Feature map visualization of different methods on SEW ResNet-18 (CNN) and Sformer-8-384+CML (Transformer). Input images are sampled from the ImageNet validation dataset. Shaded colors from blue to red indicate the impacts of the regions on the classification scores from low to high. (b) Top1 accuracy versus GPU hours on ImageNet with SEW ResNet-50 and Spikingformer-8-768. + +Comparison on ImageNet between BKDSNN and previous methods are tabulated in Tab. 2c, Tab. 2d and Tab. 3a. As anticipated, BKDSNN outperforms previous SOTA methods with both CNN- and Transformer-based models on ImageNet. BKDSNN achieves $72.32\%$ top1 accuracy with SEW ResNet-50 and $80.93\%$ with Mformer-8-512, with $4.51\%$ and $0.93\%$ improvement, respectively. + +As reported in Tab. 2c, Tab. 2d and Tab. 3a, ANN outperforms previous learning-based SNN by a large margin, which will serve as the teacher model to provide adequate knowledge for the student SNN. Based on our experiments, the accuracy improvement of SNN trained by KDSNN [53] and LaSNN [21] is limited, due to the inappropriate distillation between student and teacher models. Nevertheless, thanks to the blurred student features that effectively mimic and restore the teacher features, BKDSNN achieves promising results on the large-scale dataset (i.e. ImageNet). + +Feature Visualization. To further interpret BKDSNN, we apply Grad-CAM [46] to visualize the intermediate feature before the last layer using different methods. + +In $2^{\mathrm{nd}}$ (ANN) to $3^{\mathrm{rd}}$ (SNN) column of Fig. 3a, ANN properly focuses on the regions with vital information (highlighted in red), while learning-based SNN is distracted by the background area. Such erroneous attention of SNN ( $3^{\mathrm{rd}}$ column in Fig. 3a) does not get calibrated by the previous KD methods, i.e., KDSNN [53] ( $4^{\mathrm{th}}$ column) and LaSNN [21] ( $5^{\mathrm{th}}$ column). However, the attention (shaded) region of our BKDSNN ( $6^{\mathrm{th}}$ column) is quite similar to the pre-trained ANN, which demonstrates the effectiveness of BKD in learning the knowledge of ANN by mimicking and restoring teacher ANN features with blurred student SNN features. + +Comparison of Training Cost. We plot the evolution curves of validation top1 accuracy during GPU hours as depicted in Fig. 3b. Without applying knowledge distillation (denoted as w/o KD in blue lines), the training of SEW ResNet-50 and Spikingformer-8-768 consume the least GPU hours. With the introduction of layer-wise distillation (LaSNN [21] in green lines and KDSNN [53] in yellow lines), the training of SEW ResNet-50 and Spikingformer-8-768 consume more hours while the improvement is limited. Compared with the layer-wise distillation methods above, our single-layer BKDSNN (in red lines) saves GPU hours and boosts the performance significantly. + +# 4.3 Comparison on Neuromorphic Dataset + +We also evaluate BKDSNN on CIFAR10-DVS with results reported in Tab. 3b. Specifically, BKDSNN outperforms previous SOTA results by $1.0\%$ on models with 4 time-steps and $0.7\%$ on models with 8 time-steps. Note that, the improvement of BKDSNN on Wide-7B-Net [14] is more pronounced than that on Sformer-2-256 [62], which is probably due to the fact that Wide-7B-Net [14] with 16 inference time-steps provides more knowledge for the corresponding student model with lower time-steps, compared to Sformer-2-256 [62]. + +# 4.4 Ablation Study + +Ablation on multiple losses $(\mathcal{L}_{LD},\mathcal{L}_{BKD}$ and $\mathcal{L}_{MD})$ . We test different settings of BKDSNN, including versions without any knowledge distillation (w/o KD), with logit-based KD $(+{\mathcal{L}}_{\mathrm{LD}})$ , with BKD $(+{\mathcal{L}}_{\mathrm{BKD}})$ , with mixed distillation (logitbased $^+$ BKD, $+{\mathcal{L}}_{\mathrm{MD}}$ ). Results are tabulated in Tab. 4b. Compared to logitbased KD, applying BKD to the input feature of the last layer shows on-par accuracy improvement on both CNN- and Transformer-based SNNs. Furthermore, thanks to the compatibility between the blurred feature-based KD and logit-based KD, their mixed version shows the maximal accuracy improvement. + +In addition to the accuracy obtained for SEW ResNet-18 and Sformer-8-384 discussed above, we also plot their evolution curves of validation accuracy during training as in Fig. 4. We observe that with the introduction of logits-based distillation $\mathcal{L}_{\mathrm{LD}}$ (yellow curve) and feature-based blurred knowledge distillation $\mathcal{L}_{\mathrm{BKD}}$ (green curve), the performance is significantly improved compared to the vanilla counterpart (blue curve). Moreover, as anticipated, the mixed distillation + +Table 4: Ablation study with multiple settings. $R$ , $S$ , and $KD$ denotes ResNet, Sformer and Knowledge Distillation respectively. $S-4-384-400\mathrm{E}$ refers to Sformer-4-384 after 400 epochs training. $\star$ : results from [14], $\diamond$ : results from [62]. Best results are in bold, runner-up results are marked in gray. + +
MethodConfigS-4-384-400ER-19
τλCF10 AccCF100 AccCF10 AccCF100 Acc
+Ld1.5-96.1381.4094.5074.66
2.0-96.1781.2594.5374.57
4.0-96.3180.8894.5674.52
+BKD-0.1596.1381.3894.5274.76
-0.4096.2681.4394.5674.83
-0.6096.3181.5394.6174.89
+MD2.00.1596.1881.6394.6474.95
2.00.4096.2081.6794.6874.98
2.00.6096.37 (+0.11)81.86 (+0.19)94.76 (+0.08)75.07 (+0.09)
+ +(a) Ablation of $\tau$ and $\lambda$ on CIFAR. + +
MethodCNN-basedTransformer-based
Student SNNTeacher AccSNN AccStudent SNNTeacher AccSNN Acc
w/o KDR-18*-63.18S-8-384-74.35
R-34*-67.04S-8-512-76.54
R-50*-67.78S-8-768-77.64
+LDR-1869.7665.36(+2.18)S-8-38481.7874.69(+0.24)
R-3473.3169.90(+2.86)S-8-51276.99(+0.45)
R-5076.1371.78(+4.00)S-8-76879.04(+1.40)
+BKDR-1869.7665.16(+1.98)S-8-38481.7875.08(+0.73)
R-3473.3170.23(+3.19)S-8-51277.29(+0.75)
R-5076.1371.24(+3.46)S-8-76879.64(+2.00)
+MDR-1869.7665.60(+2.42)S-8-38481.7875.48(+1.13)
R-3473.3171.24(+4.20)S-8-51277.24(+0.70)
R-5076.1372.32(+4.54)S-8-76879.93(+2.29)
+ +(b) Ablation of multiple losses on ImageNet. + +
MethodSEW R-18 Top1 AccS-8-384 Top1 AccKD location
LaSNN [21]63.3173.85layer-wise
+LBKD63.8274.47layer-wise
KDSNN [53]63.4274.62layer-wise
+LBKD63.9474.93layer-wise
KDSNN [53]63.8774.96single-layer
BKDSNN (ours)65.60 (+1.66)75.48 (+0.52)single-layer
+ +(c) Ablation of BKD on ImageNet. + +
MethodλSpikingformer-4-384-400EResNet-19
CIFAR10CIFAR100CIFAR10CIFAR100
AccAccAccAcc
w Gw/o Gw Gw/o Gw Gw/o Gw Gw/o G
+LBKD0.0095.3894.3880.4476.6494.4793.4274.6870.97
0.1596.1394.6081.3876.8194.5293.5874.7671.22
0.4096.2694.8281.4377.1494.5693.6474.8371.29
0.6096.3194.8981.5377.1694.6193.7274.8971.37
+ +(d) Ablation of $\lambda$ and $\mathcal{G}$ on CIFAR. + +(red curve) in KDSNN leads to faster convergence with higher validation accuracy during the training. These experimental results highlight the important role of logits-based and feature-based knowledge from the teacher ANN in improving the performance of student SNNs. + +Ablation of BKD and its location We have reported the KDSNN accuracy under different settings in Tab. 4b where KD was applied to logits, the last layer feature, or both. Two more questions still remain unanswered: 1) whether BKD outperforms the vanilla implementation? 2) Is applying KD/BKD to the last layer better than the layer-wise implementation? + +To answer these questions, we perform experiments whose results are reported in Tab. 4c. LaSNN [21] and KDSNN [53] are two recent works of layer-wise implementation of KD for SNN training. We replace the original feature-based distillation loss in LaSNN and KDSNN with BKD loss $\mathcal{L}_{\mathrm{BKD}}$ . In LaSNN, the replacement improves accuracy by $0.51\%$ on SEW ResNet-18 and by $0.62\%$ on Sformer-8-384. In KDSNN, the replacement improves accuracy by $0.52\%$ on SEW ResNet-18 and by $0.31\%$ on Sformer-8-384. These experiments confirm the advantages of the blurred implementation of KD. Moreover, we further test layer-wise (every layer feature) or single-layer (intermediate feature before the last layer) implementations of KD. We observe similar accuracy boosts for the two implementations. Thus, we conclude that applying BKD only to the intermediate feature map before the last layer is the adequate setup. + +![](images/4a27c357dba9210983d0e08554470aab891a49ac80cbd24454339def593cb341.jpg) +(a) SEW ResNet-18 + +![](images/d05f73ad5df0d98ed2cc07fa4eda4f0ae980aa61b0a49ac65a315a0ffd04c0ca.jpg) +(b) SEW ResNet-34 + +![](images/f6726c3ec05329b99cb19edb09c94685fa9941cc5b72c1e0d82657e1456fd5d0.jpg) +(c) SEW ResNet-50 + +![](images/51a6c32e82dbf0d63276c7af2ed063f77bdc120843ef6174cce873dbff01af8c.jpg) +(d) Sformer-8-384 + +![](images/e77baedc6659d2c1109995edc3480a43c647928c12e951e99e40fe4c46214845.jpg) +(e) Sformer-8-512 + +![](images/634e91bd9d0ad6119c9d37866bc6c1e6bb9c7012739d37177316595dcd73114e.jpg) +(f) Sformer-8-768 +Fig. 4: Evolution curve of validation accuracy during training, with CNN- (top 3 sub-figures) and Transformer-based (bottom 3 sub-figures) SNNs on ImageNet. The labels in the legends are the same as in Tab. 4b. + +Ablation of temperature coefficient, blurred ratio and restoration block. To examine the optimal setting of hyper-parameters, i.e., temperature $\tau$ in Eqs. (7) and (8), blurred ration $\lambda$ in Eq. (2) and restoration block $\mathcal{G}$ in Eq. (4), we sweep $\tau$ , $\lambda$ and $\mathcal{G}$ with ResNet-19 and Sformer-4-384, under varying BKDSNN settings. The results of the CIFAR-10 and CIFAR-100 datasets are reported in Tab. 4a and Tab. 4d. Note that we increase the training epochs of Sformer-4-384 to 400 for more distinct results. For logits-based distillation $\mathcal{L}_{\mathrm{LD}}$ , we find that flatter logits distribution (with higher $\tau$ ) is better for classification with smaller categories (CIFAR10), while sharper logits distribution (with lower $\tau$ ) is more effective for classification with larger categories (CIFAR100). For BKD $\mathcal{L}_{\mathrm{BKD}}$ and mixed distillation $\mathcal{L}_{\mathrm{MD}}$ , training with a higher blurred ratio $\lambda$ (i.e., 0.60) achieves greater accuracy across various datasets and models. As reported in Tab. 4d, with the introduction of blurred ration $\lambda$ and restoration block $\mathcal{G}$ , student SNN better mimics teacher ANN and conveys more sufficient information, which enhances performance on CIFAR-10 and CIFAR-100 with both CNN- and Transformer-based models. + +# 4.5 Firing Rate Analysis + +Firing Rate. We compare the firing rate $r$ of the intermediate feature before last layer on CIFAR with Sformer-4-384 and ResNet-19 under multiple settings. The firing rate $r$ is defined as: + +$$ +r = \frac {1}{b c ^ {\mathrm {s t u}} h w} \sum_ {i} ^ {b} \sum_ {j} ^ {c ^ {\mathrm {s t u}}} \sum_ {k} ^ {h} \sum_ {l} ^ {w} \tilde {\mathbf {F}} _ {i, j, k, l} ^ {\mathrm {s t u}} \tag {11} +$$ + +As reported in Tab. 5a, it reveals that on both CNN- and Transformer-based models, setting a higher firing rate $r$ achieves higher accuracy. Note that, with the introduction of $\mathcal{L}_{\mathrm{BKD}}$ and $\mathcal{L}_{\mathrm{MD}}$ , high blurred ratio $\lambda$ encourages model to + +Table 5: Firing rate $r$ under multiple settings on CIFAR and Energy consumption of ANN and SNN on ImageNet. $T$ denotes Transformer. + +
MethodConfigτλS-4-384-400ER-19TimeStepSOPS(G)↓Energy Consumption (mJ)(↓)Top1Acc (↑)
CF10r (e-2)CF10AccCF10e(10)e-2CF10AccCF10e(10)e-2CF10AccCF10e(10)e-2CF10Acc
+LLD1.5-5.5596.131.5281.403.9494.500.9374.66R-50 [19]14.12 (FLOPS)18.9576.13
2.0-5.7296.171.3281.254.0394.531.0174.57Spiking R-50 [14]45.31 (28.88%↑)5.44 (71.29%↓)57.66 (-13.47)
4.0-6.0796.311.1880.884.1194.561.0874.52SEW R-50 [14]44.83 (17.23%↑)4.89 (74.20%↓)67.78 (-8.35)
+LBDK-0.154.6796.130.9781.383.8294.520.8274.76BKDSNN (ours)44.81 (16.75%↑)4.85 (74.40%↓)72.32 (-3.81)
-0.405.3596.261.0381.433.9494.560.8974.83
-0.605.5596.311.1481.534.0894.611.0274.89
+LMD2.0 0.155.3996.181.1481.633.9294.641.0374.95T-8-768 [12]118.60 (FLOPS)85.5681.16
2.0 0.406.1796.201.3481.673.9994.681.0774.98S-8-768 [61]412.54 (32.58%↓)13.68 (84.01%↓)75.85 (-5.31)
2.0 0.606.6696.371.5581.864.1894.761.1375.07S-8-768 + CML [62]412.30 (33.87%↓)13.42 (84.32%↓)77.64 (-3.52)
BKDSNN (ours)412.60 (32.26%↓)13.72 (83.96%↓)79.93 (-1.23)
+ +(a) Firing rate $r$ on CIFAR. +(b) Energy consumption on ImageNet. + +fire more spikes. Furthermore, compared with $\mathcal{L}_{\mathrm{LD}}$ , our $\mathcal{L}_{\mathrm{BKD}}$ reduces the firing rate while improving the performance, which assists the energy consumption reduction. Meanwhile, in virtue of $\mathcal{L}_{\mathrm{MD}}$ , BKDSNN significantly improves the performance with a slight boost on firing rate, which leads to state-of-the-art. + +# 4.6 Energy Consumption Analysis + +As tabulated in Tab. 5b, we calculate the energy consumption of BKDSNN according to the energy model in [25,29,65] and compare it with its ANN counterparts and SNNs trained by previous learning-based methods. In terms of the OPS, SNN variants are higher than ANNs, while ANNs use FLOPS as the operation counting metric. Nevertheless, since the energy per synaptic operation is more efficient than the operation in ANN. For ResNet-50, BKDSNN consumes the lowest energy while optimally suppressing accuracy loss. For Sformer-8-768, our BKDSNN consumes more energy to increase firing rates for maintaining similar performance as SOTA methods for SNN training. + +# 5 Conclusion + +BKDSNN constructs a knowledge distillation based framework for learning-based SNN training, achieving promising performance on both static and neuromorphic datasets supported by comprehensive experiments. We expect that BKDSNN will contribute to future investigations on brain-inspired SNN, which can further narrow the performance gap between SNN and ANN. In this work, we mainly focus on learning-based SNN training methods due to their remarkable performance under ultra-low inference time-step. We anticipate to extend BKDSNN on conversion-based methods, which is expected to overcome the existing bottlenecks of conversion-based methods under ultra-low inference time-step. + +# Acknowledgements + +This work is partially supported by National Key R&D Program of China (2022YFB4500200), National Natural Science Foundation of China (No. 62102257), Biren Technology-Shanghai Jiao Tong University Joint Laboratory Open Research Fund, Microsoft Research Asia Gift Fund, Lingang Laboratory Open Research Fund (No.LGQS-202202-11). + +# References + +1. Amir, A., Taba, B., Berg, D., Melano, T., McKinstry, J., Di Nolfo, C., Nayak, T., Andreopoulos, A., Garreau, G., Mendoza, M., et al.: A low power, fully event-based gesture recognition system. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7243-7252 (2017) 6 +2. Bu, T., Fang, W., Ding, J., Dai, P., Yu, Z., Huang, T.: Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347 (2023) 2, 3, 9, 1 +3. Cao, Y., Chen, Y., Khosla, D.: Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision 113, 54-66 (2015) 2, 3 +4. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9233-9242 (2020) 4 +5. Chen, L., Yu, C., Chen, L.: A new knowledge distillation for incremental object detection. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2019) 4 +6. Davies, M., Srinivasa, N., Lin, T.H., Chinya, G., Cao, Y., Choday, S.H., Dimou, G., Joshi, P., Imam, N., Jain, S., et al.: Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro 38(1), 82-99 (2018) 1 +7. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 2, 8 +8. Deng, S., Gu, S.: Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476 (2021) 2, 3 +9. Deng, S., Li, Y., Zhang, S., Gu, S.: Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946 (2022) 8, 9, 4 +0. Diehl, P., Neil, D., Binas, J., Cook, M., Liu, S.C., Pfeiffer, M.: Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing (07 2015). https://doi.org/10.1109/IJCNN.2015.7280696 3, 1 +1. Ding, J., Bu, T., Yu, Z., Huang, T., Liu, J.: Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training. Advances in Neural Information Processing Systems 35, 24780-24793 (2022) 2 +2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8, 14 +3. Fang, W., Chen, Y., Ding, J., Yu, Z., Masquelier, T., Chen, D., Huang, L., Zhou, H., Li, G., Tian, Y.: Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances 9(40), eadi1480 (2023). https://doi.org/10.1126/sciadv.adi1480, https://www.science.org/doi/abs/10.1126/sciadv.adi1480 7 +4. Fang, W., Yu, Z., Chen, Y., Huang, T., Masquelier, T., Tian, Y.: Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems 34, 21056-21069 (2021) 2, 3, 4, 7, 8, 9, 10, 11, 12, 14 +5. Garg, I., Chowdhury, S.S., Roy, K.: Dct-snn: Using dct to distribute spatial information over time for low-latency spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4671-4680 (2021) 4 + +16. Gerstner, W., Kistler, W.M.: Spiking neuron models: Single neurons, populations, plasticity. Cambridge university press (2002) 2 +17. Han, B., Roy, K.: Deep spiking neural network: Energy efficiency through time based coding. In: European Conference on Computer Vision. pp. 388-404. Springer (2020) 3 +18. Han, B., Srinivasan, G., Roy, K.: Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13558-13567 (2020) 1, 3 +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 14 +20. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 4 +21. Hong, D., Shen, J., Qi, Y., Wang, Y.: Lasnn: Layer-wise ann-to-snn distillation for effective and efficient training in deep spiking neural networks. arXiv preprint arXiv:2304.09101 (2023) 2, 4, 5, 7, 9, 10, 11, 12 +22. Hongmin, L., Hanchao, L., Xiangyang, J., Guoqi, L., Luping, S.: Cifar10-dvs: An event-stream dataset for object classification. Frontiers in Neuroscience 11 (2017) 2, 8 +23. Hu, Y., Tang, H., Pan, G.: Spiking deep residual network (2020) 2, 9 +24. Hu, Y., Zheng, Q., Jiang, X., Pan, G.: Fast-snn: Fast spiking neural network by converting quantized ann. arXiv preprint arXiv:2305.19868 (2023) 3 +25. Hu, Y., Deng, L., Wu, Y., Yao, M., Li, G.: Advancing spiking neural networks towards deep residual learning (2023) 14, 5 +26. Khan, A., Sohail, A., Zahoora, U., Qureshi, A.S.: A survey of the recent architectures of deep convolutional neural networks. Artificial intelligence review 53, 5455-5516 (2020) 6 +27. Kheradpisheh, S.R., Ganjtabesh, M., Thorpe, S.J., Masquelier, T.: Stdp-based spiking deep convolutional neural networks for object recognition. Neural Networks 99, 56-67 (2018) 2 +28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 2, 8 +29. Kundu, S., Pedram, M., Beerel, P.A.: Hire-snn: Harnessing the inherent robustness of energy-efficient deep spiking neural networks by training with crafted input noise (2021) 14, 5 +30. Kushawaha, R.K., Kumar, S., Banerjee, B., Velmurugan, R.: Distilling spikes: Knowledge distillation in spiking neural networks. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 4536-4543. IEEE (2021) 4 +31. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015) 1 +32. Lee, D., Park, S., Kim, J., Doh, W., Yoon, S.: Energy-efficient knowledge distillation for spiking neural networks. arXiv preprint arXiv:2106.07172 (2021) 4 +33. Li, C., Ma, L., Furber, S.: Quantization framework for fast spiking neural networks. Frontiers in Neuroscience 16, 918793 (2022) 2, 1 +34. Li, Y., Deng, S., Dong, X., Gong, R., Gu, S.: A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In: International conference on machine learning. pp. 6316-6325. PMLR (2021) 2, 3 +35. Liu, F., Zhao, W., Chen, Y., Wang, Z., Yang, T., Li, J.: Sstdp: Supervised spike timing dependent plasticity for efficient spiking neural network training. Frontiers in neuroscience p. 1413 (2021) 2 + +36. Liu, Y., Chen, K., Liu, C., Qin, Z., Luo, Z., Wang, J.: Structured knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2604-2613 (2019) 4 +37. Maass, W.: Networks of spiking neurons: the third generation of neural network models. Neural networks 10(9), 1659-1671 (1997) 1 +38. Merolla, P.A., Arthur, J.V., Alvarez-Icaza, R., Cassidy, A.S., Sawada, J., Akopyan, F., Jackson, B.L., Imam, N., Guo, C., Nakamura, Y., et al.: A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345(6197), 668-673 (2014) 1 +39. Neftci, E.O., Mostafa, H., Zenke, F.: Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36(6), 51-63 (2019) 4, 2 +40. Orchard, G., Jayawant, A., Cohen, G.K., Thakor, N.: Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience 9, 437 (2015) 6 +41. Ostojic, S.: Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons. Nature neuroscience 17(4), 594-600 (2014) 2 +42. Peng, C., Zhao, K., Maksoud, S., Li, M., Lovell, B.C.: Sid: Incremental learning for anchor-free object detection via selective and inter-related distillation. Computer vision and image understanding 210, 103229 (2021) 4 +43. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021) 6 +44. Roy, K., Jaiswal, A., Panda, P.: Towards spike-based machine intelligence with neuromorphic computing. Nature 575(7784), 607-617 (2019) 2 +45. Rueckauer, B., Lungu, I.A., Hu, Y., Pfeiffer, M., Liu, S.C.: Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in neuroscience 11, 682 (2017) 3, 1 +46. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision 128(2), 336-359 (Oct 2019). https://doi.org/10.1007/s11263-019-01228-7, http://dx.doi.org/10.1007/s11263-019-01228-7 7, 10 +47. Sengupta, A., Ye, Y., Wang, R., Liu, C., Roy, K.: Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience 13, 95 (2019) 2 +48. Sironi, A., Brambilla, M., Bourdis, N., Lagorce, X., Benosman, R.: Hats: Histograms of averaged time surfaces for robust event-based object classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1731-1740 (2018) 6 +49. Sun, S., Cheng, Y., Gan, Z., Liu, J.: Patient knowledge distillation for bert model compression. arXiv preprint arXiv:1908.09355 (2019) 4 +50. Takuya, S., Zhang, R., Nakashima, Y.: Training low-latency spiking neural network through knowledge distillation. In: 2021 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPS). pp. 1-3. IEEE (2021) 4 +51. Tan, G., Wang, Y., Han, H., Cao, Y., Wu, F., Zha, Z.J.: Multi-grained spatiotemporal features perceived network for event-based lip-reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20094-20103 (2022) 6 + +52. Wu, Y., Deng, L., Li, G., Zhu, J., Xie, Y., Shi, L.: Direct training for spiking neural networks: Faster, larger, better. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1311-1318 (2019) 2, 4, 1 +53. Xu, Q., Li, Y., Shen, J., Liu, J.K., Tang, H., Pan, G.: Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7886-7895 (2023) 2, 4, 5, 7, 9, 10, 11, 12, 3 +54. Xu, Q., Li, Y., Shen, J., Zhang, P., Liu, J.K., Tang, H., Pan, G.: Hierarchical spiking-based model for efficient image classification with enhanced feature extraction and encoding. IEEE Transactions on Neural Networks and Learning Systems (2022) 4 +55. Xu, Z., Zhang, M., Hou, J., Gong, X., Wen, C., Wang, C., Zhang, J.: Delving into transformer for incremental semantic segmentation (2022) 4 +56. Yang, Z., Li, Z., Jiang, X., Gong, Y., Yuan, Z., Zhao, D., Yuan, C.: Focal and global knowledge distillation for detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4643-4652 (2022) 4 +57. Yao, M., Hu, J., Hu, T., Xu, Y., Zhou, Z., Tian, Y., Bo, X., Li, G.: Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In: The Twelfth International Conference on Learning Representations (2023) 4, 8, 10, 6 +58. Zenke, F., Agnes, E.J., Gerstner, W.: Diverse synaptic plasticity mechanisms orchestrated to form and retrieve memories in spiking neural networks. Nature communications 6(1), 6922 (2015) 2 +59. Zenke, F., Vogels, T.P.: The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation 33(4), 899-925 (2021) 2, 1 +60. Zheng, H., Wu, Y., Deng, L., Hu, Y., Li, G.: Going deeper with directly-trained larger spiking neural networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 11062-11070 (2021) 4, 9 +61. Zhou, C., Yu, L., Zhou, Z., Zhang, H., Ma, Z., Zhou, H., Tian, Y.: Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954 (2023) 2, 3, 4, 9, 14 +62. Zhou, C., Zhang, H., Zhou, Z., Yu, L., Ma, Z., Zhou, H., Fan, X., Tian, Y.: Enhancing the performance of transformer-based spiking neural networks by improved downsampling with precise gradient backpropagation. arXiv preprint arXiv:2305.05954 (2023) 2, 3, 4, 8, 9, 10, 11, 12, 14 +63. Zhou, D., Kang, B., Jin, X., Yang, L., Lian, X., Jiang, Z., Hou, Q., Feng, J.: Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886 (2021) 6 +64. Zhou, H., Song, L., Chen, J., Zhou, Y., Wang, G., Yuan, J., Zhang, Q.: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. arXiv preprint arXiv:2102.00650 (2021) 7 +65. Zhou, Z., Zhu, Y., He, C., Wang, Y., Yan, S., Tian, Y., Yuan, L.: Spikformer: When spiking neural network meets transformer (2022) 2, 3, 4, 9, 14, 5 \ No newline at end of file diff --git a/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/images.zip b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a9959a88df29e98103a4f6240f4aaecf38e42f6b --- /dev/null +++ b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a44e5d086350bbce6665cef7ba3d46b1c504a9b8235486d6ae626863dcf67ee +size 733689 diff --git a/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/layout.json b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..247eab18e3d40d6322104c0ebbd8f392253f4cdc --- /dev/null +++ b/2024/BKDSNN_ Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation/layout.json @@ -0,0 +1,11431 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 166 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 166 + ], + "type": "text", + "content": "BKDSNN: Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "spans": [ + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "text", + "content": "Zekai " + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{1}" + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "text", + "content": ", Kang You" + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "text", + "content": ", Qinghai Guo" + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "text", + "content": ", Xiang Wang" + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "text", + "content": ", and Zhezhi He" + }, + { + "bbox": [ + 147, + 185, + 465, + 199 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 207, + 475, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 207, + 475, + 229 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 475, + 229 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 207, + 475, + 229 + ], + "type": "text", + "content": " School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 230, + 391, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 230, + 391, + 240 + ], + "spans": [ + { + "bbox": [ + 223, + 230, + 391, + 240 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 223, + 230, + 391, + 240 + ], + "type": "text", + "content": " Huawei Technologies, Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 242, + 430, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 242, + 430, + 252 + ], + "spans": [ + { + "bbox": [ + 184, + 242, + 430, + 252 + ], + "type": "text", + "content": "{sherlock.holmes.xu,kang_you," + }, + { + "bbox": [ + 184, + 242, + 430, + 252 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 184, + 242, + 430, + 252 + ], + "type": "text", + "content": "zhezhi.he}@sjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 253, + 454, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 253, + 454, + 263 + ], + "spans": [ + { + "bbox": [ + 159, + 253, + 454, + 263 + ], + "type": "text", + "content": "https://github.com/Intelligent-Computing-Research-Group/BKDSNN" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 289, + 455, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 289, + 455, + 498 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 455, + 498 + ], + "type": "text", + "content": "Abstract. Spiking neural networks (SNNs), which mimic biological neural systems to convey information via discrete spikes, are well-known as brain-inspired models with excellent computing efficiency. By utilizing the surrogate gradient estimation for discrete spikes, learning-based SNN training methods that can achieve ultra-low inference latency (i.e., number of time-step) have emerged recently. Nevertheless, due to the difficulty of deriving precise gradient for discrete spikes in learning-based methods, a distinct accuracy gap persists between SNNs and their artificial neural networks (ANNs) counterparts. To address the aforementioned issue, we propose a blurred knowledge distillation (BKD) technique, which leverages randomly blurred SNN features to restore and imitate the ANN features. Note that, our BKD is applied upon the feature map right before the last layer of SNNs, which can also mix with prior logits-based knowledge distillation for maximal accuracy boost. In the category of learning-based methods, our work achieves state-of-the-art performance for training SNNs on both static and neuromorphic datasets. On the ImageNet dataset, BKDSNN outperforms prior best results by " + }, + { + "bbox": [ + 159, + 289, + 455, + 498 + ], + "type": "inline_equation", + "content": "4.51\\%" + }, + { + "bbox": [ + 159, + 289, + 455, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 159, + 289, + 455, + 498 + ], + "type": "inline_equation", + "content": "0.93\\%" + }, + { + "bbox": [ + 159, + 289, + 455, + 498 + ], + "type": "text", + "content": " with the network topology of CNN and Transformer, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 509, + 453, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 509, + 453, + 531 + ], + "spans": [ + { + "bbox": [ + 159, + 509, + 453, + 531 + ], + "type": "text", + "content": "Keywords: Learning-based Spiking Neural Networks " + }, + { + "bbox": [ + 159, + 509, + 453, + 531 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 509, + 453, + 531 + ], + "type": "text", + "content": " Knowledge Distillation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 550, + 230, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 230, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 230, + 563 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Inspired by the neural firing mechanism in the biological nervous system, spiking neural network (SNN) [37] emerges as an alternative framework for the traditional artificial neural network (ANN). ANNs rely on continuous activation values to convey information between neurons [31], while SNNs utilize discrete spiking events to process and propagate information [6,38]. Given the remarkable energy efficiency achieved by such event-driven information propagation [18], SNNs" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "† Corresponding Author" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "show promising prospects on computational intelligence tasks [44] with strong autonomous learning capabilities and ultra-low power consumption [2,11,41,58]. Apart from these advantages, SNNs are also natively compatible with event-based neuromorphic datasets (e.g., CIFAR10-DVS) [14,61,62,65]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "text", + "content": "Currently, conversion-based and learning-based are two common methods to train SNNs. Conversion-based methods attempt to leverage information from ANNs by transferring the parameters of a pre-trained ANN to its SNN counterpart, which have achieved promising results in several tasks [2, 8, 33, 34]. However, event-triggered neurons convey limited information with ultra-low inference time-steps (e.g., 4 time-steps), leading to degraded performance of conversion-based methods [14, 65]. Consequently, conversion-based methods usually take a long inference time (e.g., 32 time-steps) to get comparable accuracy as the original ANN [3, 47] based on the equivalence of the ReLU activation and the firing rate of integrate-and-fire (IF) neuron [16]. In order to further improve the performance of SNNs with ultra-low inference time-steps, learning-based methods unfold SNN training over the inference time-steps and leverage back-propagation through time (known as BPTT) [52, 59] to train SNNs. Inspired by [27], several works [14, 35, 62] adopt surrogate gradient estimation as an implicit differentiation for direct training of SNNs. With the combination of BPTT and surrogate gradient estimation, learning-based methods exhibit advantages in processing event-based neuromorphic datasets (e.g., CIFAR10-DVS [22]) [14, 61, 62, 65]. Furthermore, direct training of CNN- [14, 23] and Transformer-based [61, 62, 65] SNNs also achieves promising performance with ultra-low time-steps on static datasets (e.g., CIFAR10 [28], CIFAR100 [28] and ImageNet [7])." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "text", + "content": "Despite that numerous efforts have been made in learning-based methods, SNNs still need improvement compared with ANNs. [53, 62] claim that imprecise surrogate gradient estimation of spiking neurons leads to performance degradation. On the other hand, KDSNN [53] and LaSNN [21] introduce layer-wise knowledge distillation (KD) to train a student SNN with rich information from the ANN teacher, achieving promising performance on several classical vision datasets (MNIST, CIFAR10/100) with CNN-based models. However, according to our preliminary experimental results, both KDSNN and LaSNN struggle with more complex datasets (ImageNet) and Transformer-based models. We hypothesize that the direct distillation between discrete SNN features and continuous ANN features [21, 53] might be the source of the limitations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "In this paper, to further alleviate performance degradation of learning-based SNNs, we introduce a blurred knowledge distillation (BKD) technique to leverage randomly blurred SNN features to restore and mimic ANN features. We apply BKD to the intermediate feature before the last layer of the student SNN and prove its superiority over layer-wise knowledge distillation in learning-based SNNs. Furthermore, we combine prior logits-based distillation with BKD, which provides mutual benefits for improving performance. We evaluate BKDSNN on both static datasets (i.e., CIFAR10, CIFAR100, and ImageNet) and a neuromorphic dataset (i.e., CIFAR10-DVS), where we achieve state-of-the-art (SOTA) performance with both CNN- and Transformer-based SNN models." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Zekai Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 146, + 116, + 353, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 116, + 353, + 127 + ], + "spans": [ + { + "bbox": [ + 146, + 116, + 353, + 127 + ], + "type": "text", + "content": "To summarize, our contributions are threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 479, + 265 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 137, + 479, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 479, + 170 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 479, + 170 + ], + "type": "text", + "content": "- We introduce a blurred knowledge distillation (BKD) technique to guide the training of student SNNs through feature imitation and demonstrate the advantages of BKD on learning-based SNNs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 172, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 172, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 139, + 172, + 479, + 217 + ], + "type": "text", + "content": "- We apply BKD to the intermediate feature before the last layer and prove its superiority over previous layer-wise distillation. Combined with prior logits-based knowledge distillation, our BKDSNN maximizes accuracy boost on learning-based SNNs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 220, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 479, + 265 + ], + "type": "text", + "content": "- We benchmark our method on both static datasets (i.e., CIFAR10, CIFAR100 and ImageNet) and a neuromorphic dataset (i.e., CIFAR10-DVS). Our method outperforms prior SOTA learning-based methods for both CNN-and Transformer-based models." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 286, + 338, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 338, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 338, + 300 + ], + "type": "text", + "content": "2 Background and Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 312, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 479, + 384 + ], + "type": "text", + "content": "Spiking Neuron Model. Previous works [2,3,8,14,24,61,62,65] have widely used the classic integrate-and-fire (IF) neuron as the fundamental unit in SNNs. IF neuron integrates input currents through changes in its membrane potential, which is compared against the threshold for spike generation. The membrane potential is reset after emitting a spike. The dynamics of the IF neuron is depicted as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 205, + 392, + 408, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 392, + 408, + 415 + ], + "spans": [ + { + "bbox": [ + 205, + 392, + 408, + 415 + ], + "type": "interline_equation", + "content": "H [ t ] = V [ t - 1 ] + \\frac {1}{\\tau} \\left(I [ t ] - (V [ t - 1 ] - V _ {\\mathrm {r e s e t}})\\right)", + "image_path": "c2c136f603750f825666ddccdeb79ec1f4bda38faa7843ff06d579bf58c39102.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 216, + 417, + 480, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 417, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 216, + 417, + 480, + 449 + ], + "type": "interline_equation", + "content": "S [ t ] = \\Theta (H [ t ] - V _ {\\mathrm {t h}}) = \\left\\{ \\begin{array}{l l} 1, & H [ t ] \\geq V _ {\\mathrm {t h}} \\\\ 0, & H [ t ] < V _ {\\mathrm {t h}} \\end{array} \\right. \\tag {1}", + "image_path": "67e359a1ba7699ad4f3425fe1f359ec759dd4408bca33eb448a0cc0ad4a0174b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 235, + 450, + 380, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 450, + 380, + 464 + ], + "spans": [ + { + "bbox": [ + 235, + 450, + 380, + 464 + ], + "type": "interline_equation", + "content": "V [ t ] = H [ t ] (1 - S [ t ]) + V _ {\\text {r e s e t}} S [ t ]", + "image_path": "1d4c1fb37b816fecbb12e0b9e879a96e7bc2e891ca3164fcb8afcfd5b2883cfb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": " is the membrane time constant, " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "I[t]" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": " is the receiving current at time step " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "\\Theta (\\cdot)" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": " is the heaviside function. The neuron accumulates " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "I[t]" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": " to change the membrane potential " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "H[t]" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": ". Once " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "H[t]" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": " exceeds the firing threshold " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{th}}" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": ", the neuron generates a spike and resets the membrane potential to " + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{reset}}" + }, + { + "bbox": [ + 131, + 473, + 479, + 522 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 481, + 665 + ], + "type": "text", + "content": "Conversion-based Methods. To fully make use of the knowledge from pre-trained ANN, conversion-based methods are designed to transfer the learned parameters of ANNs into corresponding SNNs, without further training of SNNs. Previous works [17, 18] replace ReLU activations in ANNs with spiking neurons and add scaling operations like weight normalization and threshold balancing [10] to generate corresponding SNNs. Although several works [8, 17, 18, 34] have achieved nearly loss-less accuracy with VGG-16 and ResNet, the converted SNNs still require long inference time to match the original ANNs in precision. [45] claims that low firing rates caused by weight normalization result in increasing latency for the information reaching higher layers, which limits the practical application of the converted SNNs." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Enhancing SNN with Blurred Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 217 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 217 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 217 + ], + "type": "image", + "image_path": "9bd6ac2317b1c9e4bbfbdd0bd2e2823cffac267c74d3507665209de20b5e90ac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 324 + ], + "type": "text", + "content": "Fig.1: Overview of training SNN with blurred knowledge distillation (BKD). The SNN is directly trained with back-propagation through time (BPTT) [52], where we utilize a blurred variant of KD (i.e., BKD) to achieve higher accuracy. BKD highlighted in yellow shaded region differs from prior SNN-oriented KD in three perspectives: 1) A blurred matrix " + }, + { + "bbox": [ + 130, + 224, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\pmb{B}" + }, + { + "bbox": [ + 130, + 224, + 482, + 324 + ], + "type": "text", + "content": " is randomly sampled on the fly (per input image) to mask out the feature of student SNN; 2) A restoration block " + }, + { + "bbox": [ + 130, + 224, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 224, + 482, + 324 + ], + "type": "text", + "content": " consisting of two convolutional layers connected by a ReLU layer is applied on blurred SNN features to restore and mimic ANN features; 3) Such blurred knowledge distillation is applied only to the intermediate features before the last layer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 331, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 548 + ], + "type": "text", + "content": "Learning-based Methods. With the development of surrogate gradient estimation [39], learning-based methods leverage continuous-value-based loss functions to handle nondifferentiable difficulties in SNN training. Ref. [39] introduces surrogate gradient estimation to conduct SNN training on the event-based DVS gesture dataset, which achieves better accuracy after a few training iterations. In an effort to improve computational resources in SNNs, joint CNN-SNN [54] puts forward a novel augmented spiking-based framework and achieves promising results in several image classification tasks. DCT-SNN [15] reduces the number of inference time-steps by combining frequency-domain learning with surrogate gradient descent. Spiking ResNet [60] uses the directly-trained deep SNN to achieve high performance on ImageNet. SEW ResNet [14] proposes the spike-element-wise block to counter gradient vanishing and gradient exploding, which successfully extends the depth of directly-trained SNN to more than 100 layers. When it comes to Transformer-based SNN, Spikformer [65] with a novel spiking self-attention block achieves significant improvement in image classification on both neuromorphic and static datasets. Transformer-based SNN is further improved by successive Spikingformer [61], ConvBN-MaxPooling-LIF (CML) [62] and Meta-SpikeFormer [57], achieving SOTA performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Knowledge Distillation for SNN Training. Knowledge distillation (KD) is a classical transfer learning method that has been proven to be effective on various tasks, such as model compression [20, 36, 49, 56], incremental learning [4, 5, 42, 55], etc. Recently, several works [30, 32, 50] extend KD for SNN training, such as distilling spikes from a large SNN to a small SNN or logits-based knowledge from a well-trained ANN to a simple SNN. KDSNN [53] proposes a novel framework embracing both logits-based and feature-based knowledge distillation, which enhances performance on several image classification datasets (MNIST, CIFAR10, and CIFAR100). LaSNN [21] puts forward a layer-wise feature-based" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 100 + ], + "type": "text", + "content": "Zekai Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 188, + 135, + 426, + 288 + ], + "blocks": [ + { + "bbox": [ + 212, + 114, + 400, + 126 + ], + "lines": [ + { + "bbox": [ + 212, + 114, + 400, + 126 + ], + "spans": [ + { + "bbox": [ + 212, + 114, + 400, + 126 + ], + "type": "text", + "content": "Table 1: Summary of notations in this paper" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 188, + 135, + 426, + 288 + ], + "lines": [ + { + "bbox": [ + 188, + 135, + 426, + 288 + ], + "spans": [ + { + "bbox": [ + 188, + 135, + 426, + 288 + ], + "type": "table", + "html": "
NotationDescription
Ttime-step
λblurred ratio
R∈Rb×ctearandom matrix R~U(0,1)
B∈Rb×cteablurred matrix
ˆy∈Rb×nclsground truth label
ystu,ytea∈Rb×nclsoutput logits of student SNN and teacher ANN
qstu,qtea∈Rb×nclsystuand ytea after flattened processing
Ftea∈Rb×ctea×h×wfeature before the last layer of teacher ANN
Fstu∈RT×b×cstu×h×wfeature before the last layer of student SNN
Fstu∈Rb×cstu×h×waverage of Fstu at time-step dimension (T)
fadpadaptive layer to align Fstu with Ftea
Grestoration block for Fstu
ˆFstu∈Rb×ctea×h×wblurred restoration feature from Fstu
", + "image_path": "7cda1c924515d3ca30a0ac07251e511b5d957ff6ae4f1f34fd98c3efd8bd4e39.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 369 + ], + "type": "text", + "content": "ANN-to-SNN distillation framework, achieving comparable top-1 accuracy to ANNs on the challenging Tiny ImageNet dataset. Furthermore, KDSNN [53] and LaSNN [21] are applicable to both homogeneous and heterogeneous networks. Despite the superiority of KDSNN [53] and LaSNN [21], they struggle with more complex datasets (e.g., ImageNet) with CNN-based and Transformer-based models according to our preliminary experimental results in Sec. 4.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 391, + 202, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 202, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 202, + 403 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": "In this paper, we propose a novel blurred knowledge distillation SNN (overall architecture of BKDSNN is in Fig. 1a) training method that leverages knowledge from the teacher ANN to guide the training of the student SNN. As depicted in Fig. 1b, our BKD first introduces a randomly blurred matrix " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{B}" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": " to mask the student SNN features. Unlike KDSNN [53] and LaSNN [21], which only apply an adaptive layer to align the SNN feature with the ANN feature before distillation, BKD introduces a restoration block " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": " on blurred SNN feature to restore and mimic ANN feature. The proposed BKD is applied only to the intermediate feature before the last layer, different from the layer-wise distillation [21, 53]. As shown in Fig. 1a, we combine the logits-based distillation with BKD to leverage both logits and feature knowledge from the teacher ANN." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "text", + "content": "Blurred Knowledge Distillation. To facilitate understanding, we present important notations in Tab. 1. The proposed blurred knowledge distillation is illustrated in Fig. 1b. We first introduce a random matrix " + }, + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "text", + "content": " which obeys uniform distribution between 0 and 1. Then we apply a blurred ratio " + }, + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "text", + "content": " to generate a blurred matrix " + }, + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\pmb{B}" + }, + { + "bbox": [ + 130, + 566, + 482, + 625 + ], + "type": "text", + "content": " through Eq. (2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 250, + 635, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 635, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 250, + 635, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\boldsymbol {B} _ {i, j} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\boldsymbol {R} _ {i, j} < \\lambda \\\\ 1, & \\text {O t h e r w i s e} \\end{array} \\right. \\tag {2}", + "image_path": "abd76b7263d1ba8e684ded1626610748a8fd1339753dffbf2a634e12798d5108.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Enhancing SNN with Blurred Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 312, + 229 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 312, + 229 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 312, + 229 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 312, + 229 + ], + "type": "image", + "image_path": "4925b5de32efb33a919fb07dc670420e409e93a90cb6338ae193e59a3c2937be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 231, + 296, + 240 + ], + "lines": [ + { + "bbox": [ + 159, + 231, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 159, + 231, + 296, + 240 + ], + "type": "text", + "content": "(a) Feature visualization of ResNet-18." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 117, + 480, + 229 + ], + "blocks": [ + { + "bbox": [ + 313, + 117, + 480, + 229 + ], + "lines": [ + { + "bbox": [ + 313, + 117, + 480, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 480, + 229 + ], + "type": "image", + "image_path": "6fb5a2b82e913d8ee5e3659413dd8637e02440c423fff0da061be68b47328ff3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 231, + 477, + 240 + ], + "lines": [ + { + "bbox": [ + 326, + 231, + 477, + 240 + ], + "spans": [ + { + "bbox": [ + 326, + 231, + 477, + 240 + ], + "type": "text", + "content": "(b) Feature distribution of SNN and ANN." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 249, + 479, + 273 + ], + "lines": [ + { + "bbox": [ + 131, + 249, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 479, + 273 + ], + "type": "text", + "content": "Fig.2: (a) Comparison of feature visualization in ResNet-18 under various methods; (b) Histogram of SNN and ANN feature." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": "For feature " + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{stu}}" + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": " at the last layer of student SNN, we first average " + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{stu}}" + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": " at timestep dimension " + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 266, + 306, + 480, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 306, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 266, + 306, + 480, + 338 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {F}} ^ {\\mathrm {s t u}} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\mathbf {F} _ {i} ^ {\\mathrm {s t u}} \\tag {3}", + "image_path": "319b2385ac4648f0f8d5a79ca851556e9993ef5065a36d3b8ce54a76671bbbf5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": "After the averaging operation above, we apply a convolutional adaptive layer " + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{adp}}" + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": " to align " + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}^{\\mathrm{stu}}" + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathsf{F}_i^{\\mathrm{tea}}" + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": " when student feature dimension " + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{stu}}" + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": " mismatches with teacher feature dimension " + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{tea}}" + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": ". After the alignment, we construct a restoration block " + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": " with two convolutional layers connected by a ReLU layer and apply it to " + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}^{\\mathrm{stu}}" + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": " with the blurred matrix " + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": " to achieve blurred restoration feature " + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^{\\mathrm{stu}}" + }, + { + "bbox": [ + 131, + 342, + 482, + 419 + ], + "type": "text", + "content": " as illustrated in Eq. (4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 213, + 427, + 480, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 427, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 213, + 427, + 480, + 466 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} ^ {\\mathrm {s t u}} = \\left\\{ \\begin{array}{l l} \\mathcal {G} \\left(f _ {\\mathrm {a d p}} \\left(\\tilde {\\mathbf {F}} ^ {\\mathrm {s t u}}\\right) * \\boldsymbol {B}\\right), & c ^ {\\mathrm {s t u}} \\neq c ^ {\\mathrm {t e a}} \\\\ \\mathcal {G} \\left(\\tilde {\\mathbf {F}} ^ {\\mathrm {s t u}} * \\boldsymbol {B}\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}", + "image_path": "e5dbf4b7fdad5bf6f585fbe4ad7b66e786da9032970d45ba1a1a3741a2e969f6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 475, + 479, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 479, + 497 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 479, + 497 + ], + "type": "text", + "content": "We calculate blurred knowledge distillation loss " + }, + { + "bbox": [ + 131, + 475, + 479, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 131, + 475, + 479, + 497 + ], + "type": "text", + "content": " using quadratic norm " + }, + { + "bbox": [ + 131, + 475, + 479, + 497 + ], + "type": "inline_equation", + "content": "(L^2)" + }, + { + "bbox": [ + 131, + 475, + 479, + 497 + ], + "type": "text", + "content": " as follows." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 212, + 497, + 480, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 497, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 212, + 497, + 480, + 530 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B K D}} = \\sum_ {i} ^ {b} \\sum_ {j} ^ {c ^ {\\mathrm {s t u}}} \\sum_ {k} ^ {h} \\sum_ {l} ^ {w} \\left(\\hat {\\mathbf {F}} _ {i, j, k, l} ^ {\\mathrm {s t u}} - \\mathbf {F} _ {i, j, k, l} ^ {\\mathrm {t e a}}\\right) ^ {2} \\tag {5}", + "image_path": "d243215fcb050a40a3e27a1307c59b1816f64020eaf7f5390556f1758704d36b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 535, + 479, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 535, + 479, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 535, + 479, + 561 + ], + "type": "text", + "content": "We then combine blurred knowledge distillation loss " + }, + { + "bbox": [ + 131, + 535, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 131, + 535, + 479, + 561 + ], + "type": "text", + "content": " with the original cross entropy loss " + }, + { + "bbox": [ + 131, + 535, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 131, + 535, + 479, + 561 + ], + "type": "text", + "content": " to get the total loss " + }, + { + "bbox": [ + 131, + 535, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 131, + 535, + 479, + 561 + ], + "type": "text", + "content": " as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 240, + 569, + 480, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 569, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 240, + 569, + 480, + 583 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {C E}} (\\boldsymbol {y} ^ {\\mathrm {s t u}}, \\hat {\\boldsymbol {y}}) + \\mathcal {L} _ {\\mathrm {B K D}} \\tag {6}", + "image_path": "66e8f295e9a9a8031af2899778adcab1b7d235d025a2d2e25dc4eb201272ad08.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of " + }, + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "BKD" + }, + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "content": ". For both CNN- and Transformer-based models, features of deeper layers have better representation for the original input image [26,43,63]. In other words, features of deeper layers contain sufficient information of adjacent pixels (for CNN-based models) [26] or patch pairs (for Transformer-based models) [43, 63]. Therefore, we can use partial student SNN feature before the last layer to restore and mimic teacher ANN feature through a restoration block" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 100 + ], + "type": "text", + "content": "Zekai Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "content": ", which can help the student SNN achieves better representation and effectively mimics teacher ANN. We then apply Grad-CAM which can intuitively demonstrate the feature similarity [46] on each block in ResNet-18 under various methods. As depicted in Fig. 2a, compared with previous SNN (SEW ResNet-18 [14]) and KDSNN [53], our BKDSNN transfers sufficient knowledge from ANN and effectively mimics ANN feature. Meanwhile, as illustrated in Fig. 2b, student SNN feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathsf{F}}^{\\mathrm{stu}}" + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "content": " is discrete while teacher ANN feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathsf{F}^{\\mathrm{tea}}" + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "content": " is continuous. Therefore, directly applying distillation to encourage discrete SNN feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathsf{F}}^{\\mathrm{stu}}" + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "content": " to imitate continuous ANN feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathsf{F}^{\\mathrm{tea}}" + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "content": " [21, 53] is not plausible. However, the proposed BKD leverages the restoration block " + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "content": " to map discrete feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathsf{F}}^{\\mathrm{stu}}" + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "content": " to continuous feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\hat{\\mathsf{F}}^{\\mathrm{stu}}" + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "content": " for imitating continuous ANN feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathsf{F}^{\\mathrm{tea}}" + }, + { + "bbox": [ + 131, + 116, + 482, + 269 + ], + "type": "text", + "content": ", which is a more proper spiking-friendly feature-based distillation method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 282, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 282, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 282, + 482, + 354 + ], + "type": "text", + "content": "Mixed Distillation. As depicted in Fig. 1a, BKDSNN further applies logits-based distillation which transfers knowledge from teacher ANN logits to student SNN logits. As soft logits from the teacher model embrace sufficient hidden knowledge, which serves as a good regularization to student model [64], we introduce the temperature " + }, + { + "bbox": [ + 131, + 282, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 282, + 482, + 354 + ], + "type": "text", + "content": " to make the logits distribution flatter. The output logits of both teacher and student are processed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 210, + 363, + 481, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 363, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 210, + 363, + 481, + 395 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} _ {i} = \\frac {\\exp \\left(\\boldsymbol {y} _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {n ^ {\\mathrm {c l s}}} \\exp \\left(\\boldsymbol {y} _ {j} / \\tau\\right)}, \\quad i \\in \\{1, 2, \\dots , n ^ {\\mathrm {c l s}} \\} \\tag {7}", + "image_path": "ff8d127b89a3a56bc716f762f447db90076bbc8353ca4e3d8cdd1cb5907a8d09.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 403, + 362, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 362, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 362, + 415 + ], + "type": "text", + "content": "Then logits-based distillation loss " + }, + { + "bbox": [ + 132, + 403, + 362, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LD}}" + }, + { + "bbox": [ + 132, + 403, + 362, + 415 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 424, + 481, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 424, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 225, + 424, + 481, + 438 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L D}} = \\tau^ {2} \\cdot \\mathcal {L} _ {\\text {C r o s s E n t r o p y}} \\left(\\boldsymbol {q} ^ {\\text {s t u}}, \\boldsymbol {q} ^ {\\text {t e a}}\\right) \\tag {8}", + "image_path": "b2ff0d8dd78ffe36bc0f9b2fcae30e5d9fe0ffa181ea7965096c5298c74cce21.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 448, + 479, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 448, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 131, + 448, + 479, + 471 + ], + "type": "text", + "content": "We combine " + }, + { + "bbox": [ + 131, + 448, + 479, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LD}}" + }, + { + "bbox": [ + 131, + 448, + 479, + 471 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 448, + 479, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 131, + 448, + 479, + 471 + ], + "type": "text", + "content": " to achieve mixed distillation loss " + }, + { + "bbox": [ + 131, + 448, + 479, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MD}}" + }, + { + "bbox": [ + 131, + 448, + 479, + 471 + ], + "type": "text", + "content": " (Eq. (9)), which contains both logits and feature information from the teacher ANN." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 231, + 482, + 481, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 482, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 231, + 482, + 481, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M D}} = w _ {\\mathrm {L D}} \\cdot \\mathcal {L} _ {\\mathrm {L D}} + w _ {\\mathrm {B K D}} \\cdot \\mathcal {L} _ {\\mathrm {B K D}} \\tag {9}", + "image_path": "ad48dda1c6278d4942fa8060745e51bbf5c98d88c3b0e79926fad8d68a140cb3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{LD}}" + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "text", + "content": " are weights for " + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LD}}" + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "text", + "content": ", respectively. Thus, the overall loss " + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 505, + 479, + 528 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 193, + 538, + 481, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 538, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 193, + 538, + 481, + 553 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {C r o s s E n t r o p y}} \\left(\\boldsymbol {y} ^ {\\text {s t u}}, \\hat {\\boldsymbol {y}}\\right) + \\mathcal {L} _ {\\mathrm {M D}} = \\mathcal {L} _ {\\text {t a s k}} + \\mathcal {L} _ {\\mathrm {M D}} \\tag {10}", + "image_path": "51581af0f3c870f203afd1ec7e1e39f06b6b0eed02fab4c34dd9dafbf2b8d087.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 582, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 230, + 594 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 608, + 302, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 608, + 302, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 608, + 302, + 620 + ], + "type": "text", + "content": "4.1 Experimental Configuration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": "Platform. All experiments are evaluated on a server equipped with 32 cores Intel Xeon Platinum 8352V CPU with " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "2.10\\mathrm{GHz}" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": " and 8-way NVIDIA GPUs. We use SpikingJelly [13] to simulate IF neurons." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Enhancing SNN with Blurred Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 479, + 140 + ], + "type": "text", + "content": "Static Datasets and Settings. We first evaluate BKDSNN on three static datasets, i.e., CIFAR10/100 and ImageNet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 251 + ], + "type": "text", + "content": "CIFAR10/CIFAR100 [28] contains 50k training and 10k testing images with " + }, + { + "bbox": [ + 130, + 142, + 482, + 251 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 142, + 482, + 251 + ], + "type": "text", + "content": " resolution, and is labeled with 10- or 100-categories respectively for classification. For CNN-based models, we use ResNet-19 from TET [9] and lightweight ResNet-20 from SEW ResNet [14] with corresponding ANNs as teacher models. For Transformer-based models, we train ViT-S/16 on CIFAR10 and CIFAR100 to make it serve as the teacher ANN model for Sformer-4-256, Sformer-2-384, and Sformer-4-384 (We abbreviate Spikingformer to Sformer for convenience). Sformer-L-D represents a Sformer model with L Sformer encoder blocks and D feature embedding dimensions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 252, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 361 + ], + "type": "text", + "content": "ImageNet-1K [7] contains 1.3 millions of training images and 50k validation images with " + }, + { + "bbox": [ + 130, + 252, + 482, + 361 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 252, + 482, + 361 + ], + "type": "text", + "content": " resolution, labeled with 1k classes. We choose ResNet-18, ResNet-34, and ResNet-50 pre-trained on ImageNet-1K as teacher ANN models for the corresponding SEW ResNet models [14]. For Transformer-based models, we introduce ViT-B/16 [12] pre-trained on ImageNet1K as the teacher ANN model for Sformer-8-384, Sformer-8-512 and Sformer-8-768. We also apply our BKDSNN on the latest Meta-SpikeFormer [57] and conduct experiments with Mformer-8-256, Mformer-8-384 and Mformer-8-512 on ImageNet-1K (We abbreviate Meta-SpikeFormer to Mformer for convenience.)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": "For both CNN- and Transformer-based models, we set temperature " + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": " to 2.0, blurred ratio " + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": " to 0.15 and logits-based distillation weight " + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{LD}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": " to 1. Specifically for CNN-based models, we set blurred knowledge distillation weight " + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "inline_equation", + "content": "7e^{-4}" + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": " for CIFAR10/100 and " + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "inline_equation", + "content": "7e^{-5}" + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": " for ImageNet. We follow all the training configurations (batch size, learning rate, optimizer, etc.) from SEW ResNet [14] except ResNet-19, where we use training configuration from TET [9]. For Transformer-based models, we set " + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "inline_equation", + "content": "2e^{-3}" + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": " for both CIFAR10/100 and ImageNet, and we follow all the training configurations (batch size, learning rate, optimizer, etc.) from Sformer+CML [62]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "type": "text", + "content": "Neuromorphic Dataset and Networks. We also evaluate BKDSNN on CIFAR10-DVS [22], which is a neuromorphic dataset that leverages dynamic vision sensor (DVS) to convert 10k frame-based images from the CIFAR10 dataset into 10k event streams, providing an event-stream dataset with 10 different classes. It includes 9k training samples and 1k testing samples. Note that, CIFAR10-DVS is an event-stream dataset tailored to SNN, so we use SNN models with 16 timesteps as teacher models to guide the training of SNN models with lower timesteps (i.e., 4 and 8). We utilize Wide-7B-Net [14] and Sformer-2-256 [62] with 4 and 8 inference time-steps as student models for CNN-based and Transformer-based models, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": "For both Wide-7B-Net [14] and former-2-256 [62], we set " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " to 2.0, " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " to 0.40 and " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{LD}}" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " to 1. Then we set " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "2e^{-3}" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " for Wide-7B-Net [14] and " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "7e^{-5}" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " for Sformer-2-256 [62]. We follow all the training configurations from SEW ResNet [14] for Wide-7B-Net [14], and Sformer+CML [62] for Sformer-2-256 [62]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Zekai Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 184, + 482, + 278 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "text", + "content": "Table 2: Comparison with previous works, using various networks and multiple datasets. " + }, + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "text", + "content": " denotes #parameters. " + }, + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "inline_equation", + "content": "KD" + }, + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "text", + "content": " denotes knowledge distillation. " + }, + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "text", + "content": ": results taken from [14]. †: prior state-of-the-art conversion-based method QCFS [2]. The prior SOTA results are marked in gray, which is used to compare with our BKDSNN and report the accuracy difference. Best results are in bold. - indicates results not available (not reported in original article or cannot reproduce)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 184, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 136, + 184, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 136, + 184, + 482, + 278 + ], + "type": "table", + "html": "
w/o KD MethodArchitectureR-19 (P: 12.63M)R-20 (P: 0.27M)Time StepArchitecture MethodSformer-4-256 (P: 4.15M)Sformer-2-384 (P: 5.76M)Sformer-4-384 (P: 9.32M)Time Step
CF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 AccCF10 Acc
QCS [2]†--83.7555.234w/o KD Spíkarfer [65]93.9475.9694.8076.9595.1977.864
TET [9]94.4474.4787.7559.144+CML [62]94.8277.6495.6378.7595.9379.654
STBP [60]92.9270.86--4+CML [62]94.7777.4395.2278.3495.6179.094
Spíkarfer-R [23]92.1370.1387.3259.934
SEW-R [14]93.2470.8489.0760.164
Teacher ANN95.6079.7891.7768.401Teacher ANN96.7582.2296.7582.2296.7582.221
KDSNN [53]94.3674.0889.0360.184KDSNN [53]95.0078.3895.5979.2595.8880.334
LaSNN [21]94.2173.9288.4159.174LaSNN [21]94.9778.0295.5578.9195.7979.994
BKDSNN (ours)94.6474.9589.2960.924BKDSNN (ours)95.29 (+0.29)79.41 (+1.03)95.90 (+0.31)80.63 (+1.38)96.06 (+0.18)81.26 (+0.93)4
", + "image_path": "f50c13423d5fd3f2ec88ee40973f76e7bbb410f75078a27b0688e7d954803583.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 299, + 295, + 396 + ], + "blocks": [ + { + "bbox": [ + 133, + 280, + 283, + 289 + ], + "lines": [ + { + "bbox": [ + 133, + 280, + 283, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 280, + 283, + 289 + ], + "type": "text", + "content": "(a) CNN-based SNN on " + }, + { + "bbox": [ + 133, + 280, + 283, + 289 + ], + "type": "inline_equation", + "content": "CF10 / 100" + }, + { + "bbox": [ + 133, + 280, + 283, + 289 + ], + "type": "text", + "content": " dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 293, + 279, + 473, + 289 + ], + "lines": [ + { + "bbox": [ + 293, + 279, + 473, + 289 + ], + "spans": [ + { + "bbox": [ + 293, + 279, + 473, + 289 + ], + "type": "text", + "content": "(b) Transformer-based SNN on " + }, + { + "bbox": [ + 293, + 279, + 473, + 289 + ], + "type": "inline_equation", + "content": "CF10/100" + }, + { + "bbox": [ + 293, + 279, + 473, + 289 + ], + "type": "text", + "content": " dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 299, + 295, + 396 + ], + "lines": [ + { + "bbox": [ + 138, + 299, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 295, + 396 + ], + "type": "table", + "html": "
w/o KD MethodArchitectureR-18(P: 11.69M)Top1 AccR-34(P: 21.79M)Top1 AccR-50(P: 25.56M)Top1 AccTimeStep
QCFS [2]†-69.37-32
TET [9]-68.00-4
STBP [60]-67.04-4
Spiking-R [23]★62.3261.8657.664
SEW-R [14]63.1867.0467.784
with KD (ours)Teacher ANN69.7673.3176.311
KDSNN [53]63.4267.1867.724
LaSNN [21]63.3166.9467.814
BKDSNN(ours)65.60(+2.18)71.24(+4.06)72.32(+4.51)4
", + "image_path": "04976da06a9a63d7477de263bede0aa8fa98d836e264c6ef675a219aa3080384.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 301, + 299, + 482, + 396 + ], + "blocks": [ + { + "bbox": [ + 141, + 398, + 293, + 407 + ], + "lines": [ + { + "bbox": [ + 141, + 398, + 293, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 398, + 293, + 407 + ], + "type": "text", + "content": "(c) CNN-based SNN on ImageNet dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 301, + 299, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 301, + 299, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 301, + 299, + 482, + 396 + ], + "type": "table", + "html": "
Architecture MethodSformer-8-384(P: 16.81M)Top1 Acc.Sformer-8-512(P: 29.68M)Top1 Acc.Sformer-8-768(P: 66.34M)Top1 Acc.Time Step
w/o KDSpikformer [65]70.2473.3874.814
+CML [62]72.7375.6177.344
Sformer [61]72.4574.7975.854
+CML [62]74.3576.5477.644
with KDTeacher ANN81.7881.7881.781
KDSNN [53]74.6276.4477.834
LaSNN [21]73.8576.3877.664
BKDSNN(ours)75.48(+0.86)77.24(+0.70)79.93(+2.10)4
", + "image_path": "de814e0dc9ec1d4aefe249892bf4d849b39f39b3ef7a860245d36c1182eefc0c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 397, + 480, + 407 + ], + "lines": [ + { + "bbox": [ + 303, + 397, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 303, + 397, + 480, + 407 + ], + "type": "text", + "content": "(d) Transformer-based SNN on ImageNet dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 418, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 465 + ], + "type": "text", + "content": "Settings for Prior KD Methods on SNNs. Furthermore, to compare BKDSNN with previous KD methods on SNNs (i.e., KDSNN [53] and LaSNN [21]), we follow the settings from [53] and [21] to reproduce the experimental results on all datasets with CNN-based and Transformer-based models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 487, + 317, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 317, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 317, + 498 + ], + "type": "text", + "content": "4.2 Comparison on Static Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": "Comparison on CIFAR10/100 between BKDSNN and previous methods are listed in Tab. 2a and Tab. 2b, where BKDSNN surpasses prior SOTA methods with both CNN- and Transformer-based models. Specifically, BKD applied upon Sformer-4-384 outperforms previous SOTA by " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "0.18\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": " on CIFAR10 and " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "0.93\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": " on CIFAR100, which achieves " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "96.06\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "81.26\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": " accuracy respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Intuitively, it is difficult for both KDSNN [53] and LaSNN [21] to improve the performance of CNN-based models on the CIFAR dataset. For ResNet-19, the knowledge of the ANN teacher is limited for the SNN student, leading to the above difficulties. When it comes to ResNet-20, the lightweight model conveys limited information which results in the ineffectiveness of previous knowledge distillation methods. In contrast, our BKDSNN is effective on both limited teacher knowledge as well as lightweight student models, which verifies the superiority of our method." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Enhancing SNN with Blurred Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 172, + 481, + 249 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 3: (a) Comparison with Meta-SpikeFormer [57] on ImageNet. (b) Experimental results on CIFAR10-DVS. Mformer represents Meta-SpikeFormer [57]. " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": ": SNN with high inference time-step (i.e., 16) serves as the teacher model for corresponding SNN with low inference time-step (i.e., 4 and 8). The prior state-of-the-art results are marked in gray. Best results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 172, + 481, + 249 + ], + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 249 + ], + "type": "table", + "html": "
Architecture MethodMformer-8-256(P: 15.11M)Top1 Acc.Mformer-8-384(P: 31.33M)Top1 Acc.Mformer-8-512(P: 55.35M)Top1 Acc.Time StepTransformer-basedCNN-based
MethodAccTime StepAccTime StepMethodAccTime StepAccTime Step
wDMformer [57]71.8275.4077.961Sfomer-2-256+CML [62]★81.41681.416Wide-7B-Net[14]★74.41674.416
74.1177.2078.814
Teacher ANN81.7881.7881.781Sfomer-2-256+CML [62]76.4479.28Wide-7B-Net[14]64.8470.28
with KDMformer [57]--79.111KDSNN [53]78.1479.98KDSNN [53]66.1470.98
--80.004LaSNN [21]78.3480.18LaSNN [21]66.6470.58
BKDSNN(ours)72.9376.1079.821BKDSNN(ours)79.3(+1.0)480.8(+0.7)8BKDSNN(ours)68.3(+1.7)472.2(+1.3)8
", + "image_path": "7b35e568c8f534962dd5e0d72abd877d7d76dc672d21410cab022d6e11f896bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 132, + 259, + 481, + 397 + ], + "blocks": [ + { + "bbox": [ + 133, + 250, + 465, + 258 + ], + "lines": [ + { + "bbox": [ + 133, + 250, + 465, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 250, + 465, + 258 + ], + "type": "text", + "content": "(a) Meta-SpikeFormer on ImageNet dataset. (b) Experimental results on CIFAR10-DVS." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 259, + 481, + 397 + ], + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 397 + ], + "type": "image", + "image_path": "7768b11dc02747ab46c270075a11e43623bda4ab48ca2722b155caf2f5db2d24.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 399, + 298, + 408 + ], + "lines": [ + { + "bbox": [ + 190, + 399, + 298, + 408 + ], + "spans": [ + { + "bbox": [ + 190, + 399, + 298, + 408 + ], + "type": "text", + "content": "(a) Feature map visualization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 409, + 482, + 464 + ], + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 464 + ], + "type": "text", + "content": "Fig.3: (a) Feature map visualization of different methods on SEW ResNet-18 (CNN) and Sformer-8-384+CML (Transformer). Input images are sampled from the ImageNet validation dataset. Shaded colors from blue to red indicate the impacts of the regions on the classification scores from low to high. (b) Top1 accuracy versus GPU hours on ImageNet with SEW ResNet-50 and Spikingformer-8-768." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "text", + "content": "Comparison on ImageNet between BKDSNN and previous methods are tabulated in Tab. 2c, Tab. 2d and Tab. 3a. As anticipated, BKDSNN outperforms previous SOTA methods with both CNN- and Transformer-based models on ImageNet. BKDSNN achieves " + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "inline_equation", + "content": "72.32\\%" + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "text", + "content": " top1 accuracy with SEW ResNet-50 and " + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "inline_equation", + "content": "80.93\\%" + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "text", + "content": " with Mformer-8-512, with " + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "inline_equation", + "content": "4.51\\%" + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "inline_equation", + "content": "0.93\\%" + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "text", + "content": " improvement, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": "As reported in Tab. 2c, Tab. 2d and Tab. 3a, ANN outperforms previous learning-based SNN by a large margin, which will serve as the teacher model to provide adequate knowledge for the student SNN. Based on our experiments, the accuracy improvement of SNN trained by KDSNN [53] and LaSNN [21] is limited, due to the inappropriate distillation between student and teacher models. Nevertheless, thanks to the blurred student features that effectively mimic and restore the teacher features, BKDSNN achieves promising results on the large-scale dataset (i.e. ImageNet)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": "Feature Visualization. To further interpret BKDSNN, we apply Grad-CAM [46] to visualize the intermediate feature before the last layer using different methods." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "Zekai Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": "In " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " (ANN) to " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " (SNN) column of Fig. 3a, ANN properly focuses on the regions with vital information (highlighted in red), while learning-based SNN is distracted by the background area. Such erroneous attention of SNN (" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " column in Fig. 3a) does not get calibrated by the previous KD methods, i.e., KDSNN [53] (" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " column) and LaSNN [21] (" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "5^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " column). However, the attention (shaded) region of our BKDSNN (" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "6^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " column) is quite similar to the pre-trained ANN, which demonstrates the effectiveness of BKD in learning the knowledge of ANN by mimicking and restoring teacher ANN features with blurred student SNN features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 235, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 344 + ], + "type": "text", + "content": "Comparison of Training Cost. We plot the evolution curves of validation top1 accuracy during GPU hours as depicted in Fig. 3b. Without applying knowledge distillation (denoted as w/o KD in blue lines), the training of SEW ResNet-50 and Spikingformer-8-768 consume the least GPU hours. With the introduction of layer-wise distillation (LaSNN [21] in green lines and KDSNN [53] in yellow lines), the training of SEW ResNet-50 and Spikingformer-8-768 consume more hours while the improvement is limited. Compared with the layer-wise distillation methods above, our single-layer BKDSNN (in red lines) saves GPU hours and boosts the performance significantly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 359, + 356, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 359, + 356, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 359, + 356, + 372 + ], + "type": "text", + "content": "4.3 Comparison on Neuromorphic Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 464 + ], + "type": "text", + "content": "We also evaluate BKDSNN on CIFAR10-DVS with results reported in Tab. 3b. Specifically, BKDSNN outperforms previous SOTA results by " + }, + { + "bbox": [ + 130, + 378, + 482, + 464 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 464 + ], + "type": "text", + "content": " on models with 4 time-steps and " + }, + { + "bbox": [ + 130, + 378, + 482, + 464 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 464 + ], + "type": "text", + "content": " on models with 8 time-steps. Note that, the improvement of BKDSNN on Wide-7B-Net [14] is more pronounced than that on Sformer-2-256 [62], which is probably due to the fact that Wide-7B-Net [14] with 16 inference time-steps provides more knowledge for the corresponding student model with lower time-steps, compared to Sformer-2-256 [62]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 479, + 238, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 238, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 238, + 491 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": "Ablation on multiple losses " + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{LD},\\mathcal{L}_{BKD}" + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MD})" + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": " . We test different settings of BKDSNN, including versions without any knowledge distillation (w/o KD), with logit-based KD " + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "inline_equation", + "content": "(+{\\mathcal{L}}_{\\mathrm{LD}})" + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": " , with BKD " + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "inline_equation", + "content": "(+{\\mathcal{L}}_{\\mathrm{BKD}})" + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": " , with mixed distillation (logitbased " + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": " BKD, " + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "inline_equation", + "content": "+{\\mathcal{L}}_{\\mathrm{MD}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": " ). Results are tabulated in Tab. 4b. Compared to logitbased KD, applying BKD to the input feature of the last layer shows on-par accuracy improvement on both CNN- and Transformer-based SNNs. Furthermore, thanks to the compatibility between the blurred feature-based KD and logit-based KD, their mixed version shows the maximal accuracy improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In addition to the accuracy obtained for SEW ResNet-18 and Sformer-8-384 discussed above, we also plot their evolution curves of validation accuracy during training as in Fig. 4. We observe that with the introduction of logits-based distillation " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LD}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " (yellow curve) and feature-based blurred knowledge distillation " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " (green curve), the performance is significantly improved compared to the vanilla counterpart (blue curve). Moreover, as anticipated, the mixed distillation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Enhancing SNN with Blurred Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 170, + 292, + 285 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 4: Ablation study with multiple settings. " + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "KD" + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": " denotes ResNet, Sformer and Knowledge Distillation respectively. " + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "S-4-384-400\\mathrm{E}" + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": " refers to Sformer-4-384 after 400 epochs training. " + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": ": results from [14], " + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": ": results from [62]. Best results are in bold, runner-up results are marked in gray." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 170, + 292, + 285 + ], + "lines": [ + { + "bbox": [ + 135, + 170, + 292, + 285 + ], + "spans": [ + { + "bbox": [ + 135, + 170, + 292, + 285 + ], + "type": "table", + "html": "
MethodConfigS-4-384-400ER-19
τλCF10 AccCF100 AccCF10 AccCF100 Acc
+Ld1.5-96.1381.4094.5074.66
2.0-96.1781.2594.5374.57
4.0-96.3180.8894.5674.52
+BKD-0.1596.1381.3894.5274.76
-0.4096.2681.4394.5674.83
-0.6096.3181.5394.6174.89
+MD2.00.1596.1881.6394.6474.95
2.00.4096.2081.6794.6874.98
2.00.6096.37 (+0.11)81.86 (+0.19)94.76 (+0.08)75.07 (+0.09)
", + "image_path": "84461836aba2f02b603d0a2fd319dd16f2bb2eb08c7ae1d81093affe16ba43be.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 287, + 275, + 297 + ], + "lines": [ + { + "bbox": [ + 149, + 287, + 275, + 297 + ], + "spans": [ + { + "bbox": [ + 149, + 287, + 275, + 297 + ], + "type": "text", + "content": "(a) Ablation of " + }, + { + "bbox": [ + 149, + 287, + 275, + 297 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 149, + 287, + 275, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 149, + 287, + 275, + 297 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 149, + 287, + 275, + 297 + ], + "type": "text", + "content": " on CIFAR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 298, + 170, + 479, + 285 + ], + "blocks": [ + { + "bbox": [ + 298, + 170, + 479, + 285 + ], + "lines": [ + { + "bbox": [ + 298, + 170, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 298, + 170, + 479, + 285 + ], + "type": "table", + "html": "
MethodCNN-basedTransformer-based
Student SNNTeacher AccSNN AccStudent SNNTeacher AccSNN Acc
w/o KDR-18*-63.18S-8-384-74.35
R-34*-67.04S-8-512-76.54
R-50*-67.78S-8-768-77.64
+LDR-1869.7665.36(+2.18)S-8-38481.7874.69(+0.24)
R-3473.3169.90(+2.86)S-8-51276.99(+0.45)
R-5076.1371.78(+4.00)S-8-76879.04(+1.40)
+BKDR-1869.7665.16(+1.98)S-8-38481.7875.08(+0.73)
R-3473.3170.23(+3.19)S-8-51277.29(+0.75)
R-5076.1371.24(+3.46)S-8-76879.64(+2.00)
+MDR-1869.7665.60(+2.42)S-8-38481.7875.48(+1.13)
R-3473.3171.24(+4.20)S-8-51277.24(+0.70)
R-5076.1372.32(+4.54)S-8-76879.93(+2.29)
", + "image_path": "9ac73f7c110b69099a63b6e1f134593689997fda3ccb77d3475b260c56daf201.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 135, + 300, + 261, + 378 + ], + "blocks": [ + { + "bbox": [ + 309, + 287, + 468, + 297 + ], + "lines": [ + { + "bbox": [ + 309, + 287, + 468, + 297 + ], + "spans": [ + { + "bbox": [ + 309, + 287, + 468, + 297 + ], + "type": "text", + "content": "(b) Ablation of multiple losses on ImageNet." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 300, + 261, + 378 + ], + "lines": [ + { + "bbox": [ + 135, + 300, + 261, + 378 + ], + "spans": [ + { + "bbox": [ + 135, + 300, + 261, + 378 + ], + "type": "table", + "html": "
MethodSEW R-18 Top1 AccS-8-384 Top1 AccKD location
LaSNN [21]63.3173.85layer-wise
+LBKD63.8274.47layer-wise
KDSNN [53]63.4274.62layer-wise
+LBKD63.9474.93layer-wise
KDSNN [53]63.8774.96single-layer
BKDSNN (ours)65.60 (+1.66)75.48 (+0.52)single-layer
", + "image_path": "f9d13f25f3ad7473ca473049f5ad61621df36c0fe4203d97ba4cd32dfdc924b7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 267, + 300, + 479, + 378 + ], + "blocks": [ + { + "bbox": [ + 136, + 380, + 260, + 389 + ], + "lines": [ + { + "bbox": [ + 136, + 380, + 260, + 389 + ], + "spans": [ + { + "bbox": [ + 136, + 380, + 260, + 389 + ], + "type": "text", + "content": "(c) Ablation of BKD on ImageNet." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 267, + 300, + 479, + 378 + ], + "lines": [ + { + "bbox": [ + 267, + 300, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 267, + 300, + 479, + 378 + ], + "type": "table", + "html": "
MethodλSpikingformer-4-384-400EResNet-19
CIFAR10CIFAR100CIFAR10CIFAR100
AccAccAccAcc
w Gw/o Gw Gw/o Gw Gw/o Gw Gw/o G
+LBKD0.0095.3894.3880.4476.6494.4793.4274.6870.97
0.1596.1394.6081.3876.8194.5293.5874.7671.22
0.4096.2694.8281.4377.1494.5693.6474.8371.29
0.6096.3194.8981.5377.1694.6193.7274.8971.37
", + "image_path": "0f567ce202c7afe9801035d214a948550c7335b866e0087304b661fde50bcb0f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 380, + 435, + 388 + ], + "lines": [ + { + "bbox": [ + 308, + 380, + 435, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 380, + 435, + 388 + ], + "type": "text", + "content": "(d) Ablation of " + }, + { + "bbox": [ + 308, + 380, + 435, + 388 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 308, + 380, + 435, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 308, + 380, + 435, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 308, + 380, + 435, + 388 + ], + "type": "text", + "content": " on CIFAR." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": "(red curve) in KDSNN leads to faster convergence with higher validation accuracy during the training. These experimental results highlight the important role of logits-based and feature-based knowledge from the teacher ANN in improving the performance of student SNNs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": "Ablation of BKD and its location We have reported the KDSNN accuracy under different settings in Tab. 4b where KD was applied to logits, the last layer feature, or both. Two more questions still remain unanswered: 1) whether BKD outperforms the vanilla implementation? 2) Is applying KD/BKD to the last layer better than the layer-wise implementation?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "To answer these questions, we perform experiments whose results are reported in Tab. 4c. LaSNN [21] and KDSNN [53] are two recent works of layer-wise implementation of KD for SNN training. We replace the original feature-based distillation loss in LaSNN and KDSNN with BKD loss " + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": ". In LaSNN, the replacement improves accuracy by " + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.51\\%" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " on SEW ResNet-18 and by " + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.62\\%" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " on Sformer-8-384. In KDSNN, the replacement improves accuracy by " + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.52\\%" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " on SEW ResNet-18 and by " + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.31\\%" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " on Sformer-8-384. These experiments confirm the advantages of the blurred implementation of KD. Moreover, we further test layer-wise (every layer feature) or single-layer (intermediate feature before the last layer) implementations of KD. We observe similar accuracy boosts for the two implementations. Thus, we conclude that applying BKD only to the intermediate feature map before the last layer is the adequate setup." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Zekai Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 248, + 189 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 248, + 189 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 248, + 189 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 248, + 189 + ], + "type": "image", + "image_path": "4a27c357dba9210983d0e08554470aab891a49ac80cbd24454339def593cb341.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 190, + 227, + 199 + ], + "lines": [ + { + "bbox": [ + 155, + 190, + 227, + 199 + ], + "spans": [ + { + "bbox": [ + 155, + 190, + 227, + 199 + ], + "type": "text", + "content": "(a) SEW ResNet-18" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 115, + 364, + 189 + ], + "blocks": [ + { + "bbox": [ + 249, + 115, + 364, + 189 + ], + "lines": [ + { + "bbox": [ + 249, + 115, + 364, + 189 + ], + "spans": [ + { + "bbox": [ + 249, + 115, + 364, + 189 + ], + "type": "image", + "image_path": "d05f73ad5df0d98ed2cc07fa4eda4f0ae980aa61b0a49ac65a315a0ffd04c0ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 190, + 343, + 199 + ], + "lines": [ + { + "bbox": [ + 271, + 190, + 343, + 199 + ], + "spans": [ + { + "bbox": [ + 271, + 190, + 343, + 199 + ], + "type": "text", + "content": "(b) SEW ResNet-34" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 115, + 482, + 189 + ], + "blocks": [ + { + "bbox": [ + 365, + 115, + 482, + 189 + ], + "lines": [ + { + "bbox": [ + 365, + 115, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 365, + 115, + 482, + 189 + ], + "type": "image", + "image_path": "f6726c3ec05329b99cb19edb09c94685fa9941cc5b72c1e0d82657e1456fd5d0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 190, + 459, + 199 + ], + "lines": [ + { + "bbox": [ + 388, + 190, + 459, + 199 + ], + "spans": [ + { + "bbox": [ + 388, + 190, + 459, + 199 + ], + "type": "text", + "content": "(c) SEW ResNet-50" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 201, + 249, + 275 + ], + "blocks": [ + { + "bbox": [ + 133, + 201, + 249, + 275 + ], + "lines": [ + { + "bbox": [ + 133, + 201, + 249, + 275 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 249, + 275 + ], + "type": "image", + "image_path": "51a6c32e82dbf0d63276c7af2ed063f77bdc120843ef6174cce873dbff01af8c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 277, + 223, + 285 + ], + "lines": [ + { + "bbox": [ + 159, + 277, + 223, + 285 + ], + "spans": [ + { + "bbox": [ + 159, + 277, + 223, + 285 + ], + "type": "text", + "content": "(d) Sformer-8-384" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 249, + 201, + 364, + 275 + ], + "blocks": [ + { + "bbox": [ + 249, + 201, + 364, + 275 + ], + "lines": [ + { + "bbox": [ + 249, + 201, + 364, + 275 + ], + "spans": [ + { + "bbox": [ + 249, + 201, + 364, + 275 + ], + "type": "image", + "image_path": "e77baedc6659d2c1109995edc3480a43c647928c12e951e99e40fe4c46214845.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 277, + 339, + 285 + ], + "lines": [ + { + "bbox": [ + 276, + 277, + 339, + 285 + ], + "spans": [ + { + "bbox": [ + 276, + 277, + 339, + 285 + ], + "type": "text", + "content": "(e) Sformer-8-512" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 201, + 481, + 275 + ], + "blocks": [ + { + "bbox": [ + 365, + 201, + 481, + 275 + ], + "lines": [ + { + "bbox": [ + 365, + 201, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 365, + 201, + 481, + 275 + ], + "type": "image", + "image_path": "634e91bd9d0ad6119c9d37866bc6c1e6bb9c7012739d37177316595dcd73114e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 277, + 455, + 285 + ], + "lines": [ + { + "bbox": [ + 392, + 277, + 455, + 285 + ], + "spans": [ + { + "bbox": [ + 392, + 277, + 455, + 285 + ], + "type": "text", + "content": "(f) Sformer-8-768" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "type": "text", + "content": "Fig. 4: Evolution curve of validation accuracy during training, with CNN- (top 3 sub-figures) and Transformer-based (bottom 3 sub-figures) SNNs on ImageNet. The labels in the legends are the same as in Tab. 4b." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": "Ablation of temperature coefficient, blurred ratio and restoration block. To examine the optimal setting of hyper-parameters, i.e., temperature " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": " in Eqs. (7) and (8), blurred ration " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": " in Eq. (2) and restoration block " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": " in Eq. (4), we sweep " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": " with ResNet-19 and Sformer-4-384, under varying BKDSNN settings. The results of the CIFAR-10 and CIFAR-100 datasets are reported in Tab. 4a and Tab. 4d. Note that we increase the training epochs of Sformer-4-384 to 400 for more distinct results. For logits-based distillation " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LD}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": ", we find that flatter logits distribution (with higher " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": ") is better for classification with smaller categories (CIFAR10), while sharper logits distribution (with lower " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": ") is more effective for classification with larger categories (CIFAR100). For BKD " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": " and mixed distillation " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MD}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": ", training with a higher blurred ratio " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": " (i.e., 0.60) achieves greater accuracy across various datasets and models. As reported in Tab. 4d, with the introduction of blurred ration " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": " and restoration block " + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 323, + 482, + 515 + ], + "type": "text", + "content": ", student SNN better mimics teacher ANN and conveys more sufficient information, which enhances performance on CIFAR-10 and CIFAR-100 with both CNN- and Transformer-based models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 532, + 265, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 265, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 265, + 544 + ], + "type": "text", + "content": "4.5 Firing Rate Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "text", + "content": "Firing Rate. We compare the firing rate " + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "text", + "content": " of the intermediate feature before last layer on CIFAR with Sformer-4-384 and ResNet-19 under multiple settings. The firing rate " + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 231, + 596, + 481, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 596, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 231, + 596, + 481, + 630 + ], + "type": "interline_equation", + "content": "r = \\frac {1}{b c ^ {\\mathrm {s t u}} h w} \\sum_ {i} ^ {b} \\sum_ {j} ^ {c ^ {\\mathrm {s t u}}} \\sum_ {k} ^ {h} \\sum_ {l} ^ {w} \\tilde {\\mathbf {F}} _ {i, j, k, l} ^ {\\mathrm {s t u}} \\tag {11}", + "image_path": "eb3fb5982557f6af23916b3003225c44438c0d8d0ec3cba2a5e9452337015cd1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "As reported in Tab. 5a, it reveals that on both CNN- and Transformer-based models, setting a higher firing rate " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " achieves higher accuracy. Note that, with the introduction of " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MD}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", high blurred ratio " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " encourages model to" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 448, + 102 + ], + "type": "text", + "content": "Enhancing SNN with Blurred Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 138, + 481, + 217 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 5: Firing rate " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " under multiple settings on CIFAR and Energy consumption of ANN and SNN on ImageNet. " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " denotes Transformer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 138, + 481, + 217 + ], + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 217 + ], + "type": "table", + "html": "
MethodConfigτλS-4-384-400ER-19TimeStepSOPS(G)↓Energy Consumption (mJ)(↓)Top1Acc (↑)
CF10r (e-2)CF10AccCF10e(10)e-2CF10AccCF10e(10)e-2CF10AccCF10e(10)e-2CF10Acc
+LLD1.5-5.5596.131.5281.403.9494.500.9374.66R-50 [19]14.12 (FLOPS)18.9576.13
2.0-5.7296.171.3281.254.0394.531.0174.57Spiking R-50 [14]45.31 (28.88%↑)5.44 (71.29%↓)57.66 (-13.47)
4.0-6.0796.311.1880.884.1194.561.0874.52SEW R-50 [14]44.83 (17.23%↑)4.89 (74.20%↓)67.78 (-8.35)
+LBDK-0.154.6796.130.9781.383.8294.520.8274.76BKDSNN (ours)44.81 (16.75%↑)4.85 (74.40%↓)72.32 (-3.81)
-0.405.3596.261.0381.433.9494.560.8974.83
-0.605.5596.311.1481.534.0894.611.0274.89
+LMD2.0 0.155.3996.181.1481.633.9294.641.0374.95T-8-768 [12]118.60 (FLOPS)85.5681.16
2.0 0.406.1796.201.3481.673.9994.681.0774.98S-8-768 [61]412.54 (32.58%↓)13.68 (84.01%↓)75.85 (-5.31)
2.0 0.606.6696.371.5581.864.1894.761.1375.07S-8-768 + CML [62]412.30 (33.87%↓)13.42 (84.32%↓)77.64 (-3.52)
BKDSNN (ours)412.60 (32.26%↓)13.72 (83.96%↓)79.93 (-1.23)
", + "image_path": "46ff6d7ad4aaf974511f88c026d019b0211f34857ff523dcdfe0e31e701f4563.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 161, + 218, + 261, + 227 + ], + "lines": [ + { + "bbox": [ + 161, + 218, + 261, + 227 + ], + "spans": [ + { + "bbox": [ + 161, + 218, + 261, + 227 + ], + "type": "text", + "content": "(a) Firing rate " + }, + { + "bbox": [ + 161, + 218, + 261, + 227 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 161, + 218, + 261, + 227 + ], + "type": "text", + "content": " on CIFAR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 318, + 217, + 455, + 227 + ], + "lines": [ + { + "bbox": [ + 318, + 217, + 455, + 227 + ], + "spans": [ + { + "bbox": [ + 318, + 217, + 455, + 227 + ], + "type": "text", + "content": "(b) Energy consumption on ImageNet." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 481, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 481, + 280 + ], + "type": "text", + "content": "fire more spikes. Furthermore, compared with " + }, + { + "bbox": [ + 130, + 231, + 481, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LD}}" + }, + { + "bbox": [ + 130, + 231, + 481, + 280 + ], + "type": "text", + "content": ", our " + }, + { + "bbox": [ + 130, + 231, + 481, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BKD}}" + }, + { + "bbox": [ + 130, + 231, + 481, + 280 + ], + "type": "text", + "content": " reduces the firing rate while improving the performance, which assists the energy consumption reduction. Meanwhile, in virtue of " + }, + { + "bbox": [ + 130, + 231, + 481, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MD}}" + }, + { + "bbox": [ + 130, + 231, + 481, + 280 + ], + "type": "text", + "content": ", BKDSNN significantly improves the performance with a slight boost on firing rate, which leads to state-of-the-art." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 294, + 313, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 294, + 313, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 294, + 313, + 307 + ], + "type": "text", + "content": "4.6 Energy Consumption Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 312, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 420 + ], + "type": "text", + "content": "As tabulated in Tab. 5b, we calculate the energy consumption of BKDSNN according to the energy model in [25,29,65] and compare it with its ANN counterparts and SNNs trained by previous learning-based methods. In terms of the OPS, SNN variants are higher than ANNs, while ANNs use FLOPS as the operation counting metric. Nevertheless, since the energy per synaptic operation is more efficient than the operation in ANN. For ResNet-50, BKDSNN consumes the lowest energy while optimally suppressing accuracy loss. For Sformer-8-768, our BKDSNN consumes more energy to increase firing rates for maintaining similar performance as SOTA methods for SNN training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 436, + 220, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 220, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 220, + 449 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 458, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 567 + ], + "type": "text", + "content": "BKDSNN constructs a knowledge distillation based framework for learning-based SNN training, achieving promising performance on both static and neuromorphic datasets supported by comprehensive experiments. We expect that BKDSNN will contribute to future investigations on brain-inspired SNN, which can further narrow the performance gap between SNN and ANN. In this work, we mainly focus on learning-based SNN training methods due to their remarkable performance under ultra-low inference time-step. We anticipate to extend BKDSNN on conversion-based methods, which is expected to overcome the existing bottlenecks of conversion-based methods under ultra-low inference time-step." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 582, + 247, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 247, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 247, + 597 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 491, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 491, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 491, + 666 + ], + "type": "text", + "content": "This work is partially supported by National Key R&D Program of China (2022YFB4500200), National Natural Science Foundation of China (No. 62102257), Biren Technology-Shanghai Jiao Tong University Joint Laboratory Open Research Fund, Microsoft Research Asia Gift Fund, Lingang Laboratory Open Research Fund (No.LGQS-202202-11)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Zekai Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 663 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "text", + "content": "1. Amir, A., Taba, B., Berg, D., Melano, T., McKinstry, J., Di Nolfo, C., Nayak, T., Andreopoulos, A., Garreau, G., Mendoza, M., et al.: A low power, fully event-based gesture recognition system. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7243-7252 (2017) 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 181, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 181, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 181, + 481, + 213 + ], + "type": "text", + "content": "2. Bu, T., Fang, W., Ding, J., Dai, P., Yu, Z., Huang, T.: Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347 (2023) 2, 3, 9, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "type": "text", + "content": "3. Cao, Y., Chen, Y., Khosla, D.: Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision 113, 54-66 (2015) 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 288 + ], + "type": "text", + "content": "4. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9233-9242 (2020) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "type": "text", + "content": "5. Chen, L., Yu, C., Chen, L.: A new knowledge distillation for incremental object detection. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2019) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "type": "text", + "content": "6. Davies, M., Srinivasa, N., Lin, T.H., Chinya, G., Cao, Y., Choday, S.H., Dimou, G., Joshi, P., Imam, N., Jain, S., et al.: Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro 38(1), 82-99 (2018) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 385 + ], + "type": "text", + "content": "7. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 481, + 407 + ], + "type": "text", + "content": "8. Deng, S., Gu, S.: Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476 (2021) 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 481, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 427 + ], + "type": "text", + "content": "9. Deng, S., Li, Y., Zhang, S., Gu, S.: Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946 (2022) 8, 9, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 428, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 481, + 460 + ], + "type": "text", + "content": "0. Diehl, P., Neil, D., Binas, J., Cook, M., Liu, S.C., Pfeiffer, M.: Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing (07 2015). https://doi.org/10.1109/IJCNN.2015.7280696 3, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 481, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 492 + ], + "type": "text", + "content": "1. Ding, J., Bu, T., Yu, Z., Huang, T., Liu, J.: Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training. Advances in Neural Information Processing Systems 35, 24780-24793 (2022) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 493, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 535 + ], + "type": "text", + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8, 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 589 + ], + "type": "text", + "content": "3. Fang, W., Chen, Y., Ding, J., Yu, Z., Masquelier, T., Chen, D., Huang, L., Zhou, H., Li, G., Tian, Y.: Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances 9(40), eadi1480 (2023). https://doi.org/10.1126/sciadv.adi1480, https://www.science.org/doi/abs/10.1126/sciadv.adi1480 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "type": "text", + "content": "4. Fang, W., Yu, Z., Chen, Y., Huang, T., Masquelier, T., Tian, Y.: Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems 34, 21056-21069 (2021) 2, 3, 4, 7, 8, 9, 10, 11, 12, 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 622, + 481, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 663 + ], + "type": "text", + "content": "5. Garg, I., Chowdhury, S.S., Roy, K.: Dct-snn: Using dct to distribute spatial information over time for low-latency spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4671-4680 (2021) 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Enhancing SNN with Blurred Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Gerstner, W., Kistler, W.M.: Spiking neuron models: Single neurons, populations, plasticity. Cambridge university press (2002) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 170 + ], + "type": "text", + "content": "17. Han, B., Roy, K.: Deep spiking neural network: Energy efficiency through time based coding. In: European Conference on Computer Vision. pp. 388-404. Springer (2020) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "text", + "content": "18. Han, B., Srinivasan, G., Roy, K.: Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13558-13567 (2020) 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 215, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 215, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 215, + 480, + 247 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "type": "text", + "content": "20. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 269, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 480, + 300 + ], + "type": "text", + "content": "21. Hong, D., Shen, J., Qi, Y., Wang, Y.: Lasnn: Layer-wise ann-to-snn distillation for effective and efficient training in deep spiking neural networks. arXiv preprint arXiv:2304.09101 (2023) 2, 4, 5, 7, 9, 10, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "text", + "content": "22. Hongmin, L., Hanchao, L., Xiangyang, J., Guoqi, L., Luping, S.: Cifar10-dvs: An event-stream dataset for object classification. Frontiers in Neuroscience 11 (2017) 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 430, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 430, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 430, + 342 + ], + "type": "text", + "content": "23. Hu, Y., Tang, H., Pan, G.: Spiking deep residual network (2020) 2, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "type": "text", + "content": "24. Hu, Y., Zheng, Q., Jiang, X., Pan, G.: Fast-snn: Fast spiking neural network by converting quantized ann. arXiv preprint arXiv:2305.19868 (2023) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "25. Hu, Y., Deng, L., Wu, Y., Yao, M., Li, G.: Advancing spiking neural networks towards deep residual learning (2023) 14, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "26. Khan, A., Sohail, A., Zahoora, U., Qureshi, A.S.: A survey of the recent architectures of deep convolutional neural networks. Artificial intelligence review 53, 5455-5516 (2020) 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "text", + "content": "27. Kheradpisheh, S.R., Ganjtabesh, M., Thorpe, S.J., Masquelier, T.: Stdp-based spiking deep convolutional neural networks for object recognition. Neural Networks 99, 56-67 (2018) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "text", + "content": "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "29. Kundu, S., Pedram, M., Beerel, P.A.: Hire-snn: Harnessing the inherent robustness of energy-efficient deep spiking neural networks by training with crafted input noise (2021) 14, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "30. Kushawaha, R.K., Kumar, S., Banerjee, B., Velmurugan, R.: Distilling spikes: Knowledge distillation in spiking neural networks. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 4536-4543. IEEE (2021) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 537, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 480, + 557 + ], + "type": "text", + "content": "31. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "type": "text", + "content": "32. Lee, D., Park, S., Kim, J., Doh, W., Yoon, S.: Energy-efficient knowledge distillation for spiking neural networks. arXiv preprint arXiv:2106.07172 (2021) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 580, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 480, + 600 + ], + "type": "text", + "content": "33. Li, C., Ma, L., Furber, S.: Quantization framework for fast spiking neural networks. Frontiers in Neuroscience 16, 918793 (2022) 2, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "34. Li, Y., Deng, S., Dong, X., Gong, R., Gu, S.: A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In: International conference on machine learning. pp. 6316-6325. PMLR (2021) 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "35. Liu, F., Zhao, W., Chen, Y., Wang, Z., Yang, T., Li, J.: Sstdp: Supervised spike timing dependent plasticity for efficient spiking neural network training. Frontiers in neuroscience p. 1413 (2021) 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "Zekai Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "36. Liu, Y., Chen, K., Liu, C., Qin, Z., Luo, Z., Wang, J.: Structured knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2604-2613 (2019) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 173 + ], + "type": "text", + "content": "37. Maass, W.: Networks of spiking neurons: the third generation of neural network models. Neural networks 10(9), 1659-1671 (1997) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 173, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 173, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 481, + 217 + ], + "type": "text", + "content": "38. Merolla, P.A., Arthur, J.V., Alvarez-Icaza, R., Cassidy, A.S., Sawada, J., Akopyan, F., Jackson, B.L., Imam, N., Guo, C., Nakamura, Y., et al.: A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345(6197), 668-673 (2014) 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 481, + 251 + ], + "type": "text", + "content": "39. Neftci, E.O., Mostafa, H., Zenke, F.: Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36(6), 51-63 (2019) 4, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 481, + 285 + ], + "type": "text", + "content": "40. Orchard, G., Jayawant, A., Cohen, G.K., Thakor, N.: Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience 9, 437 (2015) 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 285, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 481, + 308 + ], + "type": "text", + "content": "41. Ostojic, S.: Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons. Nature neuroscience 17(4), 594-600 (2014) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 481, + 341 + ], + "type": "text", + "content": "42. Peng, C., Zhao, K., Maksoud, S., Li, M., Lovell, B.C.: Sid: Incremental learning for anchor-free object detection via selective and inter-related distillation. Computer vision and image understanding 210, 103229 (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 341, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 481, + 374 + ], + "type": "text", + "content": "43. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021) 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 374, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 481, + 397 + ], + "type": "text", + "content": "44. Roy, K., Jaiswal, A., Panda, P.: Towards spike-based machine intelligence with neuromorphic computing. Nature 575(7784), 607-617 (2019) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 481, + 430 + ], + "type": "text", + "content": "45. Rueckauer, B., Lungu, I.A., Hu, Y., Pfeiffer, M., Liu, S.C.: Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in neuroscience 11, 682 (2017) 3, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 430, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 485 + ], + "type": "text", + "content": "46. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision 128(2), 336-359 (Oct 2019). https://doi.org/10.1007/s11263-019-01228-7, http://dx.doi.org/10.1007/s11263-019-01228-7 7, 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 485, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 518 + ], + "type": "text", + "content": "47. Sengupta, A., Ye, Y., Wang, R., Liu, C., Roy, K.: Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience 13, 95 (2019) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 519, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 481, + 564 + ], + "type": "text", + "content": "48. Sironi, A., Brambilla, M., Bourdis, N., Lagorce, X., Benosman, R.: Hats: Histograms of averaged time surfaces for robust event-based object classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1731-1740 (2018) 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 564, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 481, + 587 + ], + "type": "text", + "content": "49. Sun, S., Cheng, Y., Gan, Z., Liu, J.: Patient knowledge distillation for bert model compression. arXiv preprint arXiv:1908.09355 (2019) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 587, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 481, + 621 + ], + "type": "text", + "content": "50. Takuya, S., Zhang, R., Nakashima, Y.: Training low-latency spiking neural network through knowledge distillation. In: 2021 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPS). pp. 1-3. IEEE (2021) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "type": "text", + "content": "51. Tan, G., Wang, Y., Han, H., Cao, Y., Wu, F., Zha, Z.J.: Multi-grained spatiotemporal features perceived network for event-based lip-reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20094-20103 (2022) 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 231, + 91, + 447, + 102 + ], + "type": "text", + "content": "Enhancing SNN with Blurred Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 589 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "52. Wu, Y., Deng, L., Li, G., Zhu, J., Xie, Y., Shi, L.: Direct training for spiking neural networks: Faster, larger, better. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1311-1318 (2019) 2, 4, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "53. Xu, Q., Li, Y., Shen, J., Liu, J.K., Tang, H., Pan, G.: Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7886-7895 (2023) 2, 4, 5, 7, 9, 10, 11, 12, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "text", + "content": "54. Xu, Q., Li, Y., Shen, J., Zhang, P., Liu, J.K., Tang, H., Pan, G.: Hierarchical spiking-based model for efficient image classification with enhanced feature extraction and encoding. IEEE Transactions on Neural Networks and Learning Systems (2022) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "55. Xu, Z., Zhang, M., Hou, J., Gong, X., Wen, C., Wang, C., Zhang, J.: Delving into transformer for incremental semantic segmentation (2022) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "56. Yang, Z., Li, Z., Jiang, X., Gong, Y., Yuan, Z., Zhao, D., Yuan, C.: Focal and global knowledge distillation for detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4643-4652 (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "57. Yao, M., Hu, J., Hu, T., Xu, Y., Zhou, Z., Tian, Y., Bo, X., Li, G.: Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In: The Twelfth International Conference on Learning Representations (2023) 4, 8, 10, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "58. Zenke, F., Agnes, E.J., Gerstner, W.: Diverse synaptic plasticity mechanisms orchestrated to form and retrieve memories in spiking neural networks. Nature communications 6(1), 6922 (2015) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "59. Zenke, F., Vogels, T.P.: The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation 33(4), 899-925 (2021) 2, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "60. Zheng, H., Wu, Y., Deng, L., Hu, Y., Li, G.: Going deeper with directly-trained larger spiking neural networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 11062-11070 (2021) 4, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "61. Zhou, C., Yu, L., Zhou, Z., Zhang, H., Ma, Z., Zhou, H., Tian, Y.: Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954 (2023) 2, 3, 4, 9, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "text", + "content": "62. Zhou, C., Zhang, H., Zhou, Z., Yu, L., Ma, Z., Zhou, H., Fan, X., Tian, Y.: Enhancing the performance of transformer-based spiking neural networks by improved downsampling with precise gradient backpropagation. arXiv preprint arXiv:2305.05954 (2023) 2, 3, 4, 8, 9, 10, 11, 12, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 534 + ], + "type": "text", + "content": "63. Zhou, D., Kang, B., Jin, X., Yang, L., Lian, X., Jiang, Z., Hou, Q., Feng, J.: Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886 (2021) 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "64. Zhou, H., Song, L., Chen, J., Zhou, Y., Wang, G., Yuan, J., Zhang, Q.: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. arXiv preprint arXiv:2102.00650 (2021) 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "text", + "content": "65. Zhou, Z., Zhu, Y., He, C., Wang, Y., Yan, S., Tian, Y., Yuan, L.: Spikformer: When spiking neural network meets transformer (2022) 2, 3, 4, 9, 14, 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "Zekai Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_content_list.json b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44ae3abf74fb8fae8f75fc6043fee1cb04467ad7 --- /dev/null +++ b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_content_list.json @@ -0,0 +1,2810 @@ +[ + { + "type": "text", + "text": "BLINK: Multimodal Large Language Models Can See but Not Perceive", + "text_level": 1, + "bbox": [ + 227, + 145, + 776, + 189 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingyu Fu $^{1*}$ , Yushi Hu $^{2,3*}$ , Bangzheng Li $^{4}$ , Yu Feng $^{1}$ , Haoyu Wang $^{1}$ , Xudong Lin $^{5}$ , Dan Roth $^{1}$ , Noah A. Smith $^{2,3}$ , Wei-Chiu Ma $^{3\\dagger}$ , and Ranjay Krishna $^{2,3\\dagger}$", + "bbox": [ + 225, + 215, + 776, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Pennsylvania 2 University of Washington 3 Allen Institute for AI 4 University of California, Davis 5 Columbia University https://zeyofu.github.io/blink/", + "bbox": [ + 228, + 256, + 771, + 301 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0fc6fa44ce957b1569e768062818938bf03ae186fbe98712022e11cf3b04054a.jpg", + "image_caption": [ + "Figure 1: The BLINK Benchmark. BLINK contains 14 visual perception tasks that can be solved by humans \"within a blink\", but pose significant challenges for current multimodal LLMs. These tasks are inspired by classical computer vision problems and recast into multiple-choice questions for multimodal LLMs to answer. Notice that the visual prompts and questions in this figure are different from the actual ones used in the benchmark for illustrative purposes, and answers of the samples are provided in fn 6" + ], + "image_footnote": [], + "bbox": [ + 254, + 327, + 751, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce BLINK, a new benchmark for multimodal language models (LLMs) that focuses on core visual perception abilities not found in other evaluations. Most of the BLINK tasks can be solved by humans \"within a blink\" (e.g., relative depth estimation, visual correspondence, forensics detection, and multi-view reasoning). However, we find these perception-demanding tasks cast significant challenges for current multimodal LLMs because they resist mediation through natural language. BLINK reforms 14 classic computer vision tasks into 3,807 multiple-choice questions, paired with single or multiple images and visual prompting. While humans get $95.70\\%$ accuracy on average, BLINK is surprisingly challenging for existing multimodal LLMs: even the", + "bbox": [ + 259, + 686, + 743, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "best-performing GPT-4V and Gemini achieve accuracies of $51.26\\%$ and $45.72\\%$ , only $13.17\\%$ and $7.63\\%$ higher than random guessing, indicating that such perception abilities have not \"emerged\" yet in recent multimodal LLMs. Our analysis also highlights that specialist CV models could solve these problems much better, suggesting potential pathways for future improvements. We believe BLINK will stimulate the community to help multimodal LLMs catch up with human-level visual perception.", + "bbox": [ + 259, + 147, + 743, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Multi-modal Large Language Models $\\cdot$ Vision-Language Benchmark $\\cdot$ Visual Perception Evaluation", + "bbox": [ + 261, + 257, + 743, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 310, + 375, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to today, computer vision was originally attempting to interpret images as projections of 3D scenes, not just processing 2D arrays of flat \"patterns\" [25, 59, 62]. In this pursuit, early research developed a series of intermediate tasks: they focused on understanding optical properties like reflectance [12, 79], 3D primitives through multi-view reasoning [38, 60], geometric reasoning through depth estimation [76], instance recognition through visual correspondence [56], affordance through keypoint grounding [37], and forensics through intrinsic images [9]. Yet in the modern era of large language models (LLMs), we, as a community, have focused less on such perceptual tasks, and instead have developed new tasks, mostly expressed in natural language, emphasizing the vision-language connection learned by multimodal LLMs [3, 6, 18, 20, 24, 26, 51, 52, 57, 63, 66, 73, 80]. This might be because many traditional computer vision tasks resist mediation through natural language, due to the inherent imprecision of language (e.g., it is challenging to precisely pinpoint a spatial keypoint through language).", + "bbox": [ + 212, + 340, + 789, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper aims to highlight crucial aspects of visual perception that have been overlooked when evaluating multimodal LLMs. To appropriately position our paper, let us revisit how we currently evaluate perception through using multimodal LLMs [44,45,48,53,54,58,90]. While many of these benchmarks have been popularized as the de facto evaluation measures for influential models like GPT-4V and Gemini-Pro, they conflate perception with language knowledge and reasoning. At the risk of singling out one benchmark, let us consider two questions highlighted in the popular MMBench [53]: “ Why is this hummingbird called ruby-throated?” and “ What will happen next? A: the person is gonna laugh B: the person is gonna cry.” For the first question, the vision subpart is to recognize the hummingbird. For the second, it only needs a coarse description of the image. Everything else is left to the language model to solve. Such a conflation has also been reported for other benchmarks by previous work [11,39,87]. Our experiments show that this conflation reductively evaluates perception as a dense captioning task. In other words, by replacing the image with a task-agnostic dense caption, our experiments show that a “blind” GPT-4 performs well on these “multimodal tasks”.", + "bbox": [ + 212, + 553, + 789, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response, we propose BLINK. BLINK reimagines traditional computer vision problems through a format that allows us to evaluate multimodal LLMs.", + "bbox": [ + 212, + 810, + 789, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8e11d99406a2568a4fec377d15b45b5eae34e4f22ea46b266503916f332bbefc.jpg", + "image_caption": [ + "Figure 2: Comparison between BLINK and previous benchmarks. BLINK has several novel features: (1) BLINK incorporates diverse visual prompts, like circles, boxes, and image masks, while previous benchmarks only have text questions and answers. (2) BLINK evaluates a more comprehensive range of visual perception abilities, like multi-view reasoning, depth estimation, and reflectance estimation. Prior benchmarks are generally more focused on recognition-based VQA. (3) BLINK contains \"visual\" commonsense problems that humans can answer within seconds, while prior benchmarks like [90] require domain knowledge. The samples of previous benchmarks are from [45, 53, 90]. Part of our samples are curated from [10, 19, 29, 32, 36, 43, 91]." + ], + "image_footnote": [], + "bbox": [ + 243, + 157, + 751, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As partially demonstrated in Figure 15, BLINK consists of 14 classic computer vision tasks, ranging from low-level pattern matching (e.g., visual correspondences estimation) to mid-level reasoning (e.g., relative depth estimation), and extending to high-level visual understanding (e.g., visual similarity). The image tasks are meticulously selected such that they are difficult to solve by reducing the evaluation using dense captioning; instead, the models must perceive the contents of the image(s) to answer. We recast each traditional task into a modern question-answering format, where answer choices are either images or text. BLINK contains 3.8K questions across 7.3K images, where questions may contain multiple images that are curated from a wide range of datasets [8, 10, 19, 29, 32, 36, 42, 47], encompassing indoor household scenes as well as outdoor urban or natural environments. The questions and choices are either derived from the datasets, or", + "bbox": [ + 212, + 585, + 787, + 768 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BLINK", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "6 The answers of the examples in Figure 1 are as follows. Relative depth: B; jigsaw: A; multi-view reasoning: right; visual correspondence: A; semantic correspondence: C; forensics detection: final image; IQ test: D; visual similarity: upper one; functional correspondence: A; relative reflectance: they are about the same.", + "bbox": [ + 217, + 782, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "manually written by humans. On average, each question can be solved by a human subject within a BLINK of an eye, except the IQ test.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We carefully evaluate 17 multimodal LLMs with various sizes (i.e., 7B, 13B, 34B) on BLINK. We observe the paradox that while these problems are easy for humans (95.70% average accuracy), they are extremely hard for existing machinery - even GPT-4V model can only achieve 51.26% accuracy on average, which is 44.44% worse than humans, and 13.17% better than random guessing. We also experiment with specialist vision models and find that they perform much better than multimodal LLMs. For example, the specialist outperforms GPT-4V by 62.8% on visual correspondence estimation, 38.7% on relative depth estimation, and 34.6% on multi-view reasoning, in terms of absolute accuracy. Our findings indicate that the perceptual abilities of multimodal LLMs have been previously overestimated. Furthermore, these models may benefit from integrating insights from specialized models that excel in these areas. We believe BLINK can serve as an effective testbed for bridging the gap between traditional notions of perception and the modern generative capabilities of multimodal LLMs.", + "bbox": [ + 212, + 176, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 412, + 387, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal models: Inspired by the impressive success in recent large language models (LLMs) [13,21,63,77,93], a sequence of studies explore multimodal LMMs that can jointly understand vision and language information and generate textual answers through adding a modality adaption structure between a frozen visual encoder [27,65,71] and a frozen LLM [77,93]. Flamingo [3] and BLIP-2 [46] are two of the earliest works to explore these transformer-based multimodality conjunction structures. They first pre-train on image-text matching datasets [15,42,47,67] and then fine-tune on task-specific datasets such as visual question answer (VQA) [4,34]. Starting from LLaVA [49,51], people use LLM synthesized instruction-following chat data (which are in VQA format) for instruction tuning and achieve much better results [7,16,24,73]. There have been extended studies that explore further capabilities of multimodal LLMs, especially on VQA reasoning [30,33,39,40,68,78,84,92]. However, they mainly focus on the textual reasoning abilities within the multimodal LLMs and do not emphasize visual perceptions.", + "bbox": [ + 212, + 446, + 787, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal benchmarks: Traditional vision-language datasets are designed to assess single-task capabilities, such as optical character recognition (OCR) [55], image captioning [47], and visual question answering [4,34]. However, these datasets are often not comprehensive enough to holistically assess multimodal LMMs on general perception and reasoning abilities. Many recent papers have built more comprehensive benchmarks. MME [28] is one of the earliest holistic benchmarks containing multi-modal Yes/No questions on the defined visual perception and language reasoning tasks. MM-Vet [89] includes six sub-features from the previous datasets including recognition-focused questions, OCR, and math, providing a diverse while discrete evaluation set. MMBench [53] covers more subjects and provides a more robust circular evaluation setting. Seed", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bench [44,45] benchmark has a more diverse source of inputs, including multiple-image inputs and video, and includes more tasks. However, the visual perception questions in MME, MMBench, MM-Vet, and Seed-Bench are mainly extracted from existing VQA datasets or generated by GPT [63] from image descriptions such as COCO-Caption [47], and are recognition focused, covering topics such as object (attribute)recognition, and OCR. In contrast, we focus on multiple distinct nuanced perception abilities and recognition-level perception is only one of our focus. Some other multimodal benchmarks have distinct focuses. MMMU [90] aims at achieving expert-level artificial general intelligence by collecting domain-knowledge-required questions. HallusionBench [35] mainly tests the language hallucination and visual illusion phenomena. MathVista [58] presents exclusively mathematical domain visual questions based on images such as charts, tables, and diagrams. These benchmarks do not require human-level perception abilities as in BLINK and therefore cannot measure model visual perceptions holistically.", + "bbox": [ + 212, + 145, + 787, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The BLINK Benchmark", + "text_level": 1, + "bbox": [ + 215, + 385, + 477, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to faithfully evaluate the visual perception capabilities of existing Multimodal LLMs. We seek to study the visual perception gap between humans and machineries, and offer deeper insights into potential pathways towards achieving more generalized machine perception. Based on the observation that existing benchmarks predominantly focus on evaluating visual recognition abilities, we introduce a novel benchmark, BLINK, designed to enable both quantitative and qualitative evaluation of the nuanced perception capabilities of multimodal LLM across various dimensions. We unfold this section by illustrating the overall design of BLINK (§3.1) and discussing its unique features comparing with previous benchmarks. Then we describe each task in detail, providing an in-depth explanation of the data curation process (§3.2).", + "bbox": [ + 212, + 421, + 789, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview of BLINK", + "text_level": 1, + "bbox": [ + 215, + 614, + 423, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure that one can effectively measure what Multimodal LLMs can or cannot perceive, we carefully select 14 tasks (see §3.2 for the full list) that are difficult to solve by reducing the evaluation into text-only questions using dense captioning. The tasks are drawn from either classic computer vision problems or recent applications of Multimodal LLMs, each of which requires a nuanced understanding of the visual data. They range from low-level pattern matching (e.g., visual correspondence) to mid-level spatial reasoning (e.g., relative depth), and up to high-level visual understanding (e.g., visual similarity). This variety allows for a systematic exploration of Multimodal LLMs' capabilities across different perceptual complexity layers. Furthermore, these visual tasks vary in granularity, ranging from pixels (e.g., relative reflectance) to patches (e.g., jigsaw) and extending to the full image (e.g., forensic detection), enabling us to evaluate models' proficiency in observing at various scales.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BLINK", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ef107ef8238b4231e2ce547624ec425dbe05a0809cf83f9b2cc0efff16dc2a14.jpg", + "image_caption": [ + "Figure 3: Statistics of BLINK. 14 tasks range from pixel-level to image-level; from low-level pattern matching (e.g., visual correspondences) to mid-level reasoning (e.g., relative depth), further to high-level visual understanding (e.g., visual similarity)." + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 485, + 330 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8e5fdbf538a17f6d2bd882f0e8a08ae04b366d052f4a47df03065d39ebd23db.jpg", + "image_caption": [ + "Figure 4: Accuracies of multimodal LLMs on BLINK test set. Please refer to Table [1] and §4.2 for more results and discussions." + ], + "image_footnote": [], + "bbox": [ + 511, + 167, + 784, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To facilitate the evaluation of multimodal LLMs, we recast all tasks as multiple-choice question-answering problems. The options for answers may include images or texts, while the questions themselves can feature either single or multiple images. Prompts are designed to be both textual and visual in nature. We re-purposed several existing vision datasets as well as collected new data. In total, we contribute 3.9K multiple-choice questions and 7.3K images, with an even distribution between the validation and test sets. Numbers of each task are reported in Figure 3, and more detailed statistics can be found in Appendix A.5. Key features of BLINK: Comparing with previous benchmarks, BLINK has the following novel features:", + "bbox": [ + 212, + 476, + 787, + 630 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Visual prompting: Unlike existing benchmarks that support only text prompting, BLINK features a variety of visual prompts. This enables one to highlight specific areas within images, facilitating the evaluation of Multi-modal LLMs' detailed understanding of these regions. It also offers an interface for researchers to investigate the impact of visual prompting techniques.", + "- Perception beyond recognition: Besides visual recognition, BLINK considers a diverse set of visual perception abilities, such as 3D reasoning, geometric understanding, affordance reasoning, etc. The breadth allows one to evaluate Multimodal LLMs from an unique array of perspectives.", + "- \"Visual commonsense\" that does not require domain knowledge: The questions in BLINK are intentionally designed to be straightforward, requiring neither domain-specific knowledge nor expertise to answer. They are crafted in such a way that humans can solve them almost instantaneously," + ], + "bbox": [ + 223, + 641, + 789, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 385, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "typically within a few seconds. This allows us to explore the fundamental gap in visual perception gap between humans and Multimodal LLMs, highlighting the paradox that problems easily solved by humans often pose significant challenges for machines.", + "bbox": [ + 240, + 146, + 785, + 205 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Interleaved image-text formats: BLINK features a heterogeneous question-answering format, wherein both questions and choices can be presented as text or images. This diversity compels Multimodal LLMs to genuinely understand the questions, pushing the boundaries of their interpretative capabilities.", + "- Diverse image sources: BLINK comprises a wide range of in-the-wild images sourced from various origins, covering everything from indoor and outdoor scenes to object-centric views and landscapes. This collection spans abstract diagrams, synthesized images, and authentic photographs, ensuring a comprehensive examination of visual perception" + ], + "bbox": [ + 225, + 207, + 785, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The design principles of BLINK are also illustrated in Figure 2. We will now describe each task in detail.", + "bbox": [ + 212, + 352, + 782, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Dataset Collection Process", + "text_level": 1, + "bbox": [ + 215, + 406, + 482, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BLINK comprises 14 tasks, all of which have been repurposed into a multiple-choice question-answering format. These tasks utilize a diverse collection of images from various sources and we ensure that each image is unique.", + "bbox": [ + 212, + 431, + 784, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual correspondence: This task aims to evaluate the ability of Multimodal LLMs to understand and identify the same scene point across various viewpoints, lighting conditions, or time. We exploit HPatches [8] for this task. HPatches contains a number of image sequences, each of which are composed of images taken under different illuminations and/or viewpoints of a scene. For each question, we randomly sample two images and an interest point within them. Then we exploit the ground-truth homography to compute its correspondence. Finally, we randomly select three more interest points to serve as other choices.", + "bbox": [ + 212, + 477, + 785, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relative reflectance: This task aims to compare the reflectance (albedo) of two pixels. It allows us to evaluate Multimodal LLMs' understanding of material properties and their interaction with light, which is crucial for applications requiring high-fidelity visual interpretations. We curate our samples using human annotations from the Intrinsic Images in the Wild (IIW) dataset [10]. Each question is based on an image and two specified points. The goal is to identify which point is darker, or whether the two points have similar reflectance.", + "bbox": [ + 212, + 598, + 785, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relative depth: Humans are good at judging relative depth [19]. This task can thus serve as a proxy to validate if the geometric understanding capabilities of existing multimodal LLMs are close to human. We curate our samples using human annotations from the Depth in the Wild [19] dataset. Given an image and two specified points, the task is to determine which point is closer.", + "bbox": [ + 212, + 703, + 785, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatial relation: Understanding spatial relationships between objects in a scene is essential for interpreting complex visual environments. However, modern Multimodal LLMs often struggles with spatial concepts such as \"left\" and \"right\" [88]. This task helps us evaluate whether the models finally possess this", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BLINK", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2b6ee94633975a0e058fc3150fdd291369c324fa70bf054a07440df8075483c.jpg", + "image_caption": [ + "Visual correspondence" + ], + "image_footnote": [ + "A: Which point corresponds to the reference point 1?" + ], + "bbox": [ + 253, + 162, + 359, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5689ea0115dd42d13288f048d8b9b2dfee37e3cd72546038249cfa6c6305b164.jpg", + "image_caption": [], + "image_footnote": [ + "(a) A" + ], + "bbox": [ + 372, + 191, + 393, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42351252cbe567794610834a137b92fcb8209d07233d8dccdcfda37ed4e3640f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 191, + 418, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe48018c0a23e379c08d9faff86652e87a3b6a1b1225022555518de4ce257a06.jpg", + "image_caption": [], + "image_footnote": [ + "(b) B" + ], + "bbox": [ + 429, + 191, + 447, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ea7490b1452a576c92d4774fd6e1272b5e2a6fc52b6b6507b7787d0d53a6e9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 198, + 393, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80889ee5fd0d100618eaef38dd2324c727c0dcbc55e11014bec1d6bb57b7a45a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 198, + 418, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3478181c78b7cf1ecb16e18d01233255f363f91303a3a9f62f402839382a1c30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 198, + 447, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e4d21331d2bae3ba8a9deb5b1548a1edb38db2187ec4cbd25f84931b6df2adf.jpg", + "image_caption": [ + "Relative depth" + ], + "image_footnote": [ + "Q: Which point is closer?" + ], + "bbox": [ + 256, + 229, + 344, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/651d8987645fbe4fd4f03467e483671b0d06f9e388ca4e8c3315bb7883b40f80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 251, + 379, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f30e061ade7dc0195bf14bcbb63fc65828a43a30d7fe27f7d812f4b8f399bd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 251, + 393, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e76796139a94c9a6aa99d5eff68d41bdcf3204680d0473a4472ef424b4e8d612.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 251, + 418, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e1a45817760d86c08b5914f146029c5bab69308ef7c4322cdb3f5bd4f42ede4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 251, + 446, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6d450547c40a69d3bad0349117f19cf4bb4144c6f4edbabd3475573bc115c16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 258, + 380, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/070b3cea6cc4151c43caa320a7e8ea0556ce8438961535030fb57ae8ac6aa891.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 258, + 442, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b5dd7ad0226225c452ac6c3d90ec69d3bfca4ed92b749f113530608c58960d8.jpg", + "image_caption": [ + "Multi-view reasoning" + ], + "image_footnote": [ + "Q. The first image is from the beginning of the video and the second image is from the end. Is the camera moving towards left or right when shooting the video?" + ], + "bbox": [ + 253, + 297, + 351, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c07bb3cb678fcdbd3f0427849f29d0b99357966bd06659d47016af5ab92caaf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 325, + 390, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/134c023a930b90a42c0aa5f4b9e710a809fc725c52f7c18b02e8f0acd3e124bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 332, + 403, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5aca460fd98be45c3282bd208fdfb6c882271a97cf133fc43d9514d890f67037.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 325, + 442, + 332 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4619dea0e7f5eff96a0cfaaa33f2bfa79f888bcf39f49aa0229c6ee76bde353.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 332, + 447, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6712115a8d9c2df0620071d2e014b4f7288c427aa46643ac03fc4de0b949ba45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 339, + 447, + 344 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab505912deabe487dc23c629b6b127b48a41d38e476edc200f091dfceb5f5b8f.jpg", + "image_caption": [ + "Art style" + ], + "image_footnote": [], + "bbox": [ + 253, + 367, + 300, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/feb01dd766f52d22f28bdec52168544036307b8907d6dad44b55ecafd0993de2.jpg", + "image_caption": [], + "image_footnote": [ + "Q: Which image fits the missing part in the first image?" + ], + "bbox": [ + 313, + 378, + 362, + 409 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/154e14039f83a667fe3bff9ade0e7844eed9da8138a41633888cf4c21578a2d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 382, + 393, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e4e3d0b464a53fd1734bfe604369ac137643e80e69c1b45acda61f865379cd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 380, + 416, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2c958e71850115845c9835a4f4d1e8d0c716fc17a7068573aa3db64eb4267d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 380, + 467, + 409 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9de8f24018cadd0c26138372fcb3e43dbead6766466b1f956eae7dfa62cbd0d5.jpg", + "image_caption": [ + "Counting" + ], + "image_footnote": [], + "bbox": [ + 251, + 431, + 336, + 481 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b3f5d32d8c88b9fe2cf844fca2748e74ef13af4cdaa0139ac5dd2842f6eeb15.jpg", + "image_caption": [], + "image_footnote": [ + "Q: How many fingers are in front of the bathtub?" + ], + "bbox": [ + 352, + 452, + 370, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/016972865dae4d5cac90fcd7fc1c56814ffa00408a963f38caf78d4735eed579.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 460, + 369, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9c77077853e56a314a0b106f94c9de75869a2935d8f6b141d39a19a6761f69e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 467, + 369, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61901e0975f80a2ce43de3c3d06f9fdc4520ce8dc33dae75197da8c52f6c0c6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 453, + 442, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe66627b4d8ec6d0b645f6b539e64c83dd2ecc2057cc9bbd07fd0a98ddb7b58a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 467, + 431, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 434, + 453, + 442, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/985c6dd4a3161130177d1cb56a73df4a568eeec89a64b8861bda530bc9bbe0cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 467, + 457, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a025d187995d994c23a13f4b25f5ebce5d55927b498ade3796f87e1dbdddac8.jpg", + "image_caption": [ + "IQ test" + ], + "image_footnote": [ + "Q: Which image comes at the end?" + ], + "bbox": [ + 254, + 506, + 388, + 550 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cddbf4b51915617b679eb5a0f0a9f57710bececf4db69b62d1e267ce7842d654.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 527, + 421, + 534 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d47c3c59997047da3f010529362000f9b62977a788ceeaee514cd0c4eb22f190.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 527, + 470, + 535 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c12a496d68093d9ff5b9c06b21a55d99662c09003dcdacd8dd57fc863bfce5aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 536, + 433, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bbc26ee5e64d405602bb624c425e2117ad46e75b77aed97d4e36857121215c3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 536, + 470, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77c6b5c4b66ae600a2954ca2707715063f2a81aaa78cf46e014618a2c056c41e.jpg", + "image_caption": [ + "1 Semantic" + ], + "image_footnote": [ + "Q: Which point is semantically similar to the reference point 1?" + ], + "bbox": [ + 253, + 571, + 361, + 619 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/766ac0850e000c625cad38a6a112b6ddb0193aa3d8a493bd0a6adb1e58ec9673.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 588, + 410, + 594 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b159e2ae552fe6069d25ca9125b310ba3e584c3d3d68ba062a01bec1e50ca515.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 588, + 467, + 595 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0127801748bc39d4a68bdeaa24b9171a0512c8c55b7a533f547d9da87fe06f10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 601, + 418, + 607 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b058eb3109671b70f69ca0cc10b29832c017f26d616aaf21132916232097d408.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 601, + 480, + 617 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/977586505e10e895db3137716321027709a98775c83bea83af6a3cd51c1fd8f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 626, + 348, + 636 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5258312ff30e6325337d3434304f118059e78e2275466fcca4615ac864c652aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 628, + 382, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2450015f28bc8359038ab5baf011ff1918fb929b8b830774f95f499c75d2f956.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 627, + 431, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4479acecd70be14838dac929070a3013b4eced580ddc069b3e2167bf06e46ea5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 628, + 478, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fabb53e70b0df933cfcf2ff058aded627f18874381583252c0218147899ea8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 628, + 532, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 545, + 627, + 570, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e2877f232288738e9d83959b67ee06a6ee1415b0637bdfa1a58075086cec9b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 628, + 655, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b9ed0e4b8e4697279d0dc0192a5e54f4b7024770bb8ddf9dd79de1465aa97d2.jpg", + "image_caption": [ + "Relative reflectance" + ], + "image_footnote": [ + "Q: Consider the surface color (color without shading) of the two points in the image. Which one is darker, or the color is about the same?" + ], + "bbox": [ + 509, + 162, + 609, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61094acecc0cfd4a3d870f961d1a833799c2c9ee349c9fa8fb80ffecc1df1768.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 186, + 627, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d07bf20a09e40b882ae823b58833fd10757047409b80aa37ab46da4ad70ffb7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 195, + 630, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37f550064cadd3bc93ca905ee7f88bc63fe3e59ac06e768b27826ecbdeae7f0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 195, + 648, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba3421d9168789866e2a712b775e62f18274de1b216c7fc8acffc845ca438723.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 195, + 689, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53a6af5ff11348ee5f46e3fce202019be61cfb32b164026ced9f061b9af3f65e.jpg", + "image_caption": [], + "image_footnote": [ + "(b) B is darker", + "(b) B is darker" + ], + "bbox": [ + 691, + 186, + 710, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1afaa30997f6e8156423f465444dde3c653d50e3524a6df6f4daa7f0ebde1d07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 195, + 710, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d70b02c9f50fcc837aa4a8af57130ce5f3e41edf09f97b2e5b61584b85e38ad5.jpg", + "image_caption": [ + "Spatial reasoning" + ], + "image_footnote": [], + "bbox": [ + 511, + 229, + 560, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1ae464d1233777e17c372c003736fb132d33cd218ca889dfe7b701bad78f747f.jpg", + "image_caption": [], + "image_footnote": [ + "Q: Is the bed at the right side of the dining table?" + ], + "bbox": [ + 578, + 250, + 599, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9ef6393e6e18068156a0ac93d7488e5612c66ac0055792d3aa09df9b1ddb629.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 250, + 625, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fec253e3c0343e070193a68c69e4cbf2d407e8066b8b8b25525ab29732f205aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 250, + 651, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa86a63d228c6745c047284f68d6e85f4d3213c8afb5aaac2397caac7d845733.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 250, + 679, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dde2ecb0cfb6bffb2124c796d412b0f9ee6bef3260a2c8390e3d27d10d4f58d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 257, + 691, + 266 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/534f20444fc29b577843427ed47bff55b404994ed4050e585dbe7b96549931c4.jpg", + "image_caption": [ + "Jigsaw" + ], + "image_footnote": [ + "Q: Which image fits the missing part in the first image?" + ], + "bbox": [ + 509, + 299, + 573, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e36f4e07db8660f686e01ac09f3299c4812718d7106ccf2ac2d87230118f7d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 313, + 650, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fce774468fa3f1902ecaa222e8647770907bf29306f418dba5d08bb278458d42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 313, + 661, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b76e2ef980b6667ae40db8d8d66970f42d07944af034495d92f17fcc4251cd93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 313, + 725, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cec570b58da17bf2015e26eed1f874264fa4083f6c980e1f1a82f7a489ee45fd.jpg", + "image_caption": [ + "Object localization", + "Forensics detection" + ], + "image_footnote": [ + "Q: Which bounding box more accurately encloses the bun?" + ], + "bbox": [ + 509, + 368, + 604, + 416 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34fb30e756a205b5554d37df4d0d7feb712b48471f673b7080410d072e0395ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 392, + 689, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 671, + 400, + 689, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98eeba09cb4216392f2875bf2691bf06c6add1f2b0634634aa06160206c84710.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 392, + 710, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 671, + 400, + 689, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/999c312274771ddff17a6a6482752b6694d6a8a2b161d6e3fed808a77eabe7d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 407, + 705, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b4c98f281143e0e58fe60448be2482885d4ffe636add08314aa1c9d262dd0c2.jpg", + "image_caption": [], + "image_footnote": [ + "Q: Which image is most likely to be a real photograph?" + ], + "bbox": [ + 511, + 444, + 560, + 482 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/393bef96cbf2b2051f12af281ca418e5a2b802d8ba319d1770338acf32cfe6fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 445, + 612, + 473 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23dd786dbd85f9401a43a47adae456a8a1e6cda94a89954b14b006d7143cde32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 445, + 665, + 473 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18b7e8e079a07bfb9190fa99c9aa5eb35eb09742fa0b4ee4f5ddbfa78c224308.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 445, + 718, + 473 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da075ed77930d542e54affdbc601981218f67d91191fb6db799dde6aa0a37c07.jpg", + "image_caption": [ + "Visual similarity" + ], + "image_footnote": [ + "Q: Which image is most similar to the reference image?" + ], + "bbox": [ + 509, + 502, + 573, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94dc6ee69e3b8b2824b26ccad8a65d799f519415bf62c14f5cdbf649b2f71ed4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 517, + 635, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/362cdf0b32cc3029051dd9242451a363e060c252e4db786d2952879d6f41a03e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 517, + 661, + 539 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f00cff1e151a939a03b146f4ff23b6e8a694109dae9ca33b50ba59a33aa55dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 517, + 689, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02055b2953abfab149c98e35ba75da7478fa9cae5729ea08de56951123680174.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 517, + 728, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40e6e0ee580a2ec686de7bf6f98978ae662f3af20794de0da89b93f937560ff1.jpg", + "image_caption": [ + "Functional correspondence" + ], + "image_footnote": [ + "Q Which point is functionally different from reference point 1 during mixing?" + ], + "bbox": [ + 509, + 571, + 653, + 621 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62cf777146c3edb5dbeeb788cf96cda371fcbf5fa010ecad725546adc555109f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 595, + 679, + 602 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4467d0ee1e86e3cc91aa13bcc4ee39ab78b66913806341aea73dfa072d6c3454.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 595, + 707, + 602 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbba412a63df6715cc5cddb77f69099e640442411b1b8afc32ee77ad4f1a1b3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 602, + 679, + 608 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e4b3eea5e9df564b65ef4e09cbdd1a0cb75aff3623dbb660d06a30c39079003.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 602, + 707, + 608 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e1576154b3887c795ba3271aff1dc324f3050072bb87dfa87d50f9216715db1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 609, + 679, + 614 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/083be8e78cb1fbec184a77b304474a13e9151a088f6ec3266305c6e2fc450524.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 609, + 725, + 614 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07dffa79710d93202c148ad77796d3ac999e063be525ca284d5a86ebff76811e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 628, + 535, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 545, + 627, + 570, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 573, + 628, + 591, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 591, + 628, + 655, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5eaae702c43824d0cfc94c3fb0d0264c19e004968e40cf4660eba92685aec38a.jpg", + "image_caption": [ + "Figure 5: Qualitative results on BLINK. For each task, we show the choice of LLaVAv1.6-34B [51], Qwen-VL-Max [7], Gemini Pro [73], GPT-4V [63], and humans. Red choice indicates the ground truth. Notice that the markers are intentionally enlarged for visualization purposes, and we make some images inset images to save space. For IQ test, the third image is constructed by overlaying the first and second images." + ], + "image_footnote": [], + "bbox": [ + 609, + 635, + 627, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34535316fb4354c16131856371a6f902772a22d071b46d16c79882790df0789e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 630, + 655, + 636 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/811ebca9e7c09a042df12987a66b1f0f8a8804714ce607c17c10323becd89d9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 630, + 681, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vital skill. We curate our samples from the Visual Spatial Reasoning [48] dataset. Each sample contains an image and a claim. The task is to determine if the claim is true or false. We reformat the claims into binary questions via GPT-3.5 [13]. Multi-view reasoning: This task is centered on evaluating the multi-view reasoning capabilities of Multimodal LLMs. The objective is to deduce the", + "bbox": [ + 212, + 760, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "relative camera motion based on two images of an object captured from different viewpoints. Our data is sourced from the Wild6D dataset [91], which features videos of various objects recorded in diverse settings. We select two random frames from each video to calculate the relative camera motion. Recognizing that even humans might struggle to precisely articulate 3D motion details, we simplify the task by classifying motions into two broad categories: moving towards the left or moving towards the right. Despite the simplicity of these questions, as we will later demonstrate, they pose significant challenges for current models.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jigsaw: This task assesses the ability of Multimodal LLMs to recognize and group patterns, aligning patches based on continuity in shape, color, and texture. We utilize images from the TARA dataset [32] and segment each of them into a 3x3 grid. We retain the three segments from the upper left corner as the reference image, and treat the central segment along with a randomly chosen segment as options. The objective is to identify the correct patch (i.e., the central patch).", + "bbox": [ + 212, + 270, + 789, + 361 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Art style: This task evaluates Multimodal LLMs capability to analyze and discern both local and global similarities in art styles among multiple images. Although there have been prior efforts to incorporate art-related questions into evaluation [90], such attempts primarily focused on questions requiring expert-level knowledge, including deducing an artist's name and understanding historical contexts, rather than on direct image comparison. For this task, we collect paintings and their stylistic information from WikiArt. Given one reference painting image and two other paintings as options, the model is tasked with identifying the one that most closely shares the art style of the reference painting.", + "bbox": [ + 212, + 363, + 789, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Object localization: The ability to accurately detect and localize objects is critical for scene understanding. While previous benchmarks [53] have explored this task, their focus was primarily on coarse localization. For instance, they might only ask the model if an object is located at the \"top\" or \"right\" side of an image. BLINK, in contrast, aims for a more fine-grained evaluation. We exploit images from LVIS [36], randomly sampling one object per image along with its ground-truth bounding box. Then we add Gaussian noise to the ground-truth box to create a confounding box. The goal is to select the correct one.", + "bbox": [ + 212, + 502, + 787, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Counting: This task evaluate Multimodal LLMs' abilities in detection, recognition, and compositional reasoning, particularly in complex scenes where objects may overlap, be occluded, or vary in size and appearance. We select our questions from the TallyQA dataset [2], known for its challenging human-written counting questions. Each sample comprises an image, a question, and a numerical answer. In addition, we randomly select three numbers to serve as confounding options.", + "bbox": [ + 212, + 626, + 787, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Forensic detection: Recent advances in generative AI have raised concerns about malicious uses and have prompted calls for the automatic detection of fake content. To evaluate whether Multimodal LLMs can fulfill such a role, we construct sets of real and synthesized images that describe similar scenes and ask the models to identify the real ones. Specifically, we first generate synthetic images using Stable Diffusion XL [64], employing COCO captions [47] as prompts. Then, we manually search online using these captions as descriptions and select high-quality photographs as the real images.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BLINK", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "IQ test: This task evaluates the ability of Multimodal LLMs to engage in graphical reasoning, without requiring any domain-specific knowledge. We manually collect test samples, along with human explanations, from various public, license-friendly online sources. Given visual examples and a selection of images, the objective is to identify the image that either continues the pattern established by the examples or is spatially consistent with them.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Visual similarity: This task aims to verify whether Multimodal LLMs possess a nuanced understanding of visual features, patterns, and aesthetics at a level comparable to humans. We select our samples from the DreamSim dataset [29]. Given a reference image alongside two alternative images, the objective is to identify the image that more closely resembles the reference image perceptually.", + "bbox": [ + 212, + 237, + 787, + 311 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Semantic correspondence: This task focuses on identifying and matching semantically similar yet visually distinct elements across images, thereby evaluating the ability of Multimodal LLMs to understand the underlying semantics of object parts. Our samples are sourced from the SPair-71k dataset [61], which features pairs of images with multiple corresponding semantic points. For each task, we randomly select one semantic point in an image as a reference, and provide the matching point alongside three random semantic points in the paired image as options. The objective is to accurately identify the correct matches.", + "bbox": [ + 212, + 313, + 787, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Functional correspondence: The task aims to identify points that are functionally similar across objects. It challenges Multimodal LLMs to extend their understanding beyond mere semantics, enabling them to infer the diverse functions an object can perform in various contexts. Such capability is crucial for applications in robotics. We derive our samples from the FunKPoint dataset [43], which features paired images annotated for functional correspondences. Following a method analogous to semantic correspondence, we present an action alongside two object images. One image includes a reference point, while the other offers four potential points. The objective is to select the point that best matches the reference in terms of functional affordances.", + "bbox": [ + 212, + 433, + 787, + 583 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data quality control: To guarantee the quality of BLINK, we manually go through all collected data and filter out data that are ambiguous.", + "bbox": [ + 212, + 584, + 785, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 638, + 375, + 655 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we first describe the experimental setup and the baselines (\\$4.1). Then we present a comprehensive evaluation of 16 recent multimodal LLMs (\\$4.2). We demonstrate that while humans can answer the questions with high accuracy, BLINK is challenging for existing models. Finally, we provide detailed analyses on multiple experimental settings, including the effect of reducing images to captions, sensitivity to different visual prompts, and error analysis (\\$4.3).", + "bbox": [ + 212, + 670, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 784, + 428, + 799 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Multimodal LLMs: We evaluate BLINK on 16 recent models: MiniGPT-4-v2 [16], OpenFlamingo-v2 [5], InstructBLIP (7B and 13B) [24], CogVLM [80],", + "bbox": [ + 212, + 809, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "LLaVA(v1, v1.5, v1.6, internLM, and xtuner versions, model size 7B, 13B, and 34B) [23,26,50-52], Yi-VL (6B and 34B), Qwen-VL-MAX [7], Gemini Pro [73], Claude 3 Opus [1] and GPT-4V(vision) [63]. See Appendix B for more details.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation setup: We follow standard setups as in the VLMEvalKit [22], where the temperature is set to 0 and retry is set to 10. However, we do not resize the images during any experiment. For the models that do not support multiple images as input, we concatenate the images as input. We extract the choice from the models' output with a set of pre-defined rules and GPT-3.5-turbo [13]. We refer the readers to Appendix A for more details on visual prompting, how we generate the answers in BLINK, and the human evaluation protocol.", + "bbox": [ + 212, + 193, + 787, + 299 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 321, + 372, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Overall performance: As shown in Table 1, the mean accuracy of 7B and 13B open-source Multimodal LLMs hover around $35 - 42\\%$ , which is similar to random guess $(38.09\\%)$ . The most proficient open-source model, LLaVA-v1.6-34B, achieves an accuracy of $45.05\\%$ . Even the most advanced models, GPT-4V and Gemini Pro and Claude 3 OPUS, achieve accuracies of only $51.26\\%$ , $45.72\\%$ , and $44.11\\%$ respectively. Their performance are merely $13.17\\%$ , $7.63\\%$ and $6.02\\%$ better than random guessing and lag behind human performance by $44.44\\%$ , $49.98\\%$ and $51.59\\%$ . Notably, for certain tasks such as jigsaw, semantic correspondence, multi-view reasoning, object localization, and relative reflectance, some multimodal LLMs even underperform compared to random guessing. Some qualitative results are shown in Figure 5. Detailed scores on the validation set are in Appendix C. In which tasks do multimodal LLMs show relative strengths and weaknesses? Figure 4 shows the accuracies of the best-performing models on BLINK: LLaVA-v1.6-34B [51], Gemini Pro [73], and GPT-4V [63]. We observe that multimodal LLMs perform relatively better on spatial reasoning, art style, and counting tasks, in which they are much better than random guessing. The models also demonstrate some capability in relative depth and forensics detection. Overall, they are doing relatively well on mid-level perception tasks. In terms of granularity, the models in general perform better on image-level tasks and struggle on pixel-level and crop-level tasks.", + "bbox": [ + 212, + 348, + 787, + 651 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "GPT-4V behaves differently: Figure 4 and Table 1 show an interesting phenomenon: GPT-4V's performance pattern is different from other models. Compared with its counterparts, GPT-4V is much better in visual similarity, art style, jigsaw, and multi-view reasoning. Specifically, its performance on visual similarity is $29\\%$ better than Gemini Pro, demonstrating that GPT-4V possesses a nuanced understanding of visual patterns and aesthetics that is similar to humans. In contrast, Gemini Pro and LLaVA have similar performance patterns. Human performance: Human evaluators achieve over $95\\%$ accuracy across most tasks, with an average accuracy of $95.70\\%$ . This performance disparity between", + "bbox": [ + 212, + 651, + 787, + 787 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BLINK", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "7 More details are at the official website at https://www.01.ai/", + "bbox": [ + 217, + 796, + 653, + 811 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "Note that the human score for IQ test is annotated by authors. It may not reflect typical human performance, which is also expected to vary.", + "bbox": [ + 218, + 811, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3aed257f90f930164ba0c81c7909ac4f0b808f12c171311ece365c442b5d7f63.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Validation (1,901)Test (1,906)Similarity (136)Counting (120)Depth (124)Jigsaw (150)Art (117)Fun.Corr. (130)
Random Choice38.0938.09502550505025
Human95.6795.7096.7093.7599.1999.0095.3080.77
Open-source multimodal LLMs
MiniGPT-4-v2 1634.2334.5752.9410.8349.1926.0047.8618.46
OpenFlamingo-v2 539.1838.3255.1521.6754.0346.0052.1436.15
InstructBLIP-7B 2439.7238.6546.3229.1750.8154.0047.8623.85
InstructBLIP-13B 2442.2439.5846.3230.8350.0054.0050.4322.31
LLaVAInternLM2-7B 7437.7136.0652.9452.5052.4234.6730.7723.08
Yi-VL-6B 738.7241.2446.3246.6756.4550.0053.8523.85
Yi-VL-34B 741.6842.7850.0058.3353.2354.0046.1539.23
LLaVA-v1.5-7B-xtuner 2339.3640.8146.3253.3350.8154.0047.8623.85
LLaVA-v1.5-13B-xtuner 2342.0041.3146.3245.0054.0353.3347.8626.15
CogVLM 8041.5439.3846.3238.3350.8152.6749.5723.85
LLaVA-v1.5-7B 4937.1338.0146.3243.3351.6111.3347.8621.54
LLaVA-v1.5-13B 4942.6640.5546.3250.0047.5854.0047.8620.77
LLaVA-v1.6-34B 5146.8045.0546.3268.3364.5256.6747.0130.77
API-based models
Qwen-VL-Max 740.2841.9451.4755.8358.873.3337.6128.46
Gemini Pro 7345.1645.7255.8865.0050.0054.0049.5732.31
Claude 3 OPUS 144.0544.1170.5949.1757.2632.6760.6822.31
GPT-4V(ison) 6351.1451.2683.0960.8358.8762.6778.6331.54
GPT-4 Turbo 6354.6153.8983.0960.8366.9466.0081.2031.54
GPT-4o 6360.0459.0365.4451.6764.5258.0082.9139.23
Sem.Corr. (140)Spatial (143)Local. (125)Vis.Corr. (172)Multi-view (133)Reflect. (134)Forensic (132)IQ (150)
Random Choice255050255033.332525
Human96.0798.2598.0099.4292.4895.14100.0080.00
Open-source multimodal LLMs
MiniGPT-4-v2 1626.4351.7556.0023.8452.6331.3417.4219.33
OpenFlamingo-v2 523.5746.8552.0025.0041.3543.2815.9123.33
InstructBLIP-7B 2425.0055.2444.8022.6758.6529.8529.5523.33
InstructBLIP-13B 2422.8664.3452.0020.9354.1446.2713.6426.00
LLaVAInternLM2-7B 7422.1474.1348.0021.5141.3532.843.7914.67
Yi-VL-6B 726.4372.7349.6029.6548.1229.8520.4523.33
Yi-VL-34B 721.4370.6354.4023.8441.3546.2717.4222.67
LLaVA-v1.5-7B-xtuner 2324.2974.8345.6023.8442.1126.8736.3621.33
LLaVA-v1.5-13B-xtuner 2322.1477.6248.0022.0941.3546.2729.5518.67
CogVLM 8023.5767.1343.2020.9357.1426.8724.2426.67
LLaVA-v1.5-7B 4932.1470.6348.8020.3549.6236.5728.0324.00
LLaVA-v1.5-13B 4923.5767.8347.2020.3541.3545.5227.2728.00
LLaVA-v1.6-34B 5127.8676.2241.6027.3346.6229.8541.6726.00
API-based models
Qwen-VL-Max 729.2977.6249.6022.6753.3849.2547.7322.00
Gemini Pro 7322.1467.1346.4037.2141.3546.2745.4527.33
Claude 3 OPUS 120.7157.3446.4031.4057.8927.6162.1221.33
GPT-4V(ison) 6330.0072.0350.4037.2158.6538.8130.3024.67
GPT-4 Turbo 6332.8667.1348.8042.4457.1434.3351.5230.67
GPT-4o 6345.7176.9256.0071.5160.1538.8185.6130.00
", + "bbox": [ + 228, + 143, + 777, + 710 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 1: Results of different models on the BLINK test set. The first row shows task names and number of test data. Average scores on validation set are also included.", + "bbox": [ + 215, + 710, + 784, + 738 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "humans and multimodal LLMs highlights the significant visual perception gap that exists between current machine learning models and humans in perceiving, processing, and understanding complex visual and textual context.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3bb53de13f40437cd78c2a10177ee3ef19aad7e7ca0e50653921245457a2c3e7.jpg", + "image_caption": [ + "Figure 6: Performance of using image caption + text-only GPT-4 vs. GPT-4 Vision on MMBench [53], MMMU [90], and BLINK (\\$4.3)." + ], + "image_footnote": [], + "bbox": [ + 222, + 147, + 442, + 237 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/97243bb6d556d0e1285a0c95cd16652f285abb6f77c05cc2543cc425f2a8db59.jpg", + "image_caption": [ + "Figure 7: Accuracy of GPT-4V with different visual prompts (e.g., different circle sizes, colors) on relative depth, relative reflectance, and visual correspondence tasks. More discussions in §4.3." + ], + "image_footnote": [], + "bbox": [ + 467, + 152, + 784, + 233 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Analysis", + "text_level": 1, + "bbox": [ + 215, + 340, + 331, + 356 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Is dense captioning all you need for a multimodal LLM benchmark? To answer the question, we reduce multimodal benchmarks to a text-only problem. Specifically, we convert images into task-agnostic dense image captions with GPT-4V. The dense caption describes detailed information about the image and the visual prompts (e.g., where each circle is), using language. For each multimodal question, we prompt the text-only GPT-4-0125-preview model with image captions and the textual question and evaluate if the \"blind\" GPT-4 can answer the question. We call this Caption + LLM. This experiment is predicated on the hypothesis that captioning involves predominantly recognition-centric perception. If using captions along with text-only LLMs yields performance comparable to or surpassing that achieved through the integration of images with multimodal LLMs, then the perception demands of that benchmark are primarily confined to recognition only.", + "bbox": [ + 212, + 369, + 787, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We experiment with BLINK, MMBench [53] and MMMU [90], as illustrated in Figure 6. Surprisingly, we find that the Caption + LLM setting achieves better results on MMBench than GPT-4V (with $5.7\\%$ increase in accuracy). On MMMU, Caption + LLM achieves $47.2\\%$ accuracy, which is $9.6\\%$ lower than GPT-4V performance, but is still much better than random guessing. On BLINK, Caption + LLM fails, achieving random guessing performance. These results indicate that dense captions cover the visual information needed for MMBench. For MMMU, image captions carry a large portion of visual information needed to answer the domain-knowledge-specific questions. Meanwhile, the performance decrease observed in BLINK suggests the necessity for advanced perceptual abilities beyond what is currently attainable with general captions. This variance highlights the limitations of existing multimodal LLM benchmarks in addressing the full spectrum of visual perception.", + "bbox": [ + 212, + 566, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of visual prompting on BLINK: Several BLINK tasks involve visual prompting. Prior work [69] shows that factors like shape, size, and color may affect task performance, and circles give the best overall performance. Following [69], we adopt circles in BLINK and analyze the effect of circle sizes and colors on multiple tasks in Figure 7. We experiment with relative depth, relative reflectance,", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BLINK", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/528c48b84e106d58d4cdeb7c98380ef8230ed55c98fbcd55a1ec39ed268d0a23.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskVis.Corr.DepthMulti-viewSem.Corr.ForensicReflect.
Random25.0050.0050.0025.0025.0033.33
Human99.5699.5992.1094.60100.0099.63
Gemini Pro42.4440.3244.3626.6250.7645.52
GPT-4V33.7259.6855.6428.7834.0938.81
SpecialistDIFT 72DepthAnything86LoFTR 70DIFT 72ORDINAL SHADING 14
96.5197.5890.2271.2268.9477.61
", + "bbox": [ + 241, + 143, + 756, + 231 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 2: Comparison between multimodal LLMs, specialists, and human performance on the BLINK dev set. The specialists perform much better than multimodal LLMs.", + "bbox": [ + 212, + 233, + 785, + 260 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "and visual correspondence, with 100 validation set samples per task. The images are all reshaped to 1024px height. We experiment with circles with 5px, 10px, and 15px radius, and with red or gray color. We find that red is better than gray for all tasks. Also, the optimal circle size is task-dependent. On average 10px circles work the best, and we use it for all evaluations in this paper. The experiments suggest that visual prompting can have a big impact on multimodal LLM performance, and improving visual prompts or improving model robustness to prompt variation is a promising direction for future research [85].", + "bbox": [ + 212, + 303, + 787, + 424 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Can specialist models solve BLINK tasks? Specialists can serve as a proxy upper bound of how good multimodal LLMs could be. We download the trained checkpoints for six specialist models and evaluate them on BLINK. As shown in Table 2, the specialists perform much better than GPT-4V and Gemini Pro, outperforming the best multimodal LLM by $18\\%$ to $57\\%$ on these tasks. Specifically, DepthAnything [86] and DIFT [72] achieve human-level performance on depth estimation and visual correspondence, whereas multimodal LLMs fail miserably. This sheds light on the possibility that multimodal LLMs may progress on these tasks given the correct data and training strategy. For instance, one possible way is to distill existing specialist models into multimodal LLMs [41].", + "bbox": [ + 212, + 424, + 789, + 575 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Error analysis of GPT-4V: We randomly sampled 140 error instances made by GPT-4V on BLINK, 10 per task, and meticulously examined them. The most common types of errors are: Hallucinate fine-grained patterns and attributes (24.2%): the model hallucinates the nuanced details of objects. This error is most common for relative reflectance, forensics detection, and jigsaw tasks. Hallucinate visual prompt locations (20.0%): the circle location described by the model is wrong. This is common for visual correspondence and relative depth tasks. Other errors include Failures on capturing overall setting or style (8.6%), and Failures on grounding an object (5.7%). More details are in Appendix C.3", + "bbox": [ + 212, + 575, + 787, + 712 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 733, + 359, + 750 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced BLINK, a new multimodal LLM benchmark that evaluates core visual perception abilities not found in existing evaluations. While these tasks seem trivial for humans to solve \"within a blink\", we find they pose significant challenges for current multimodal LLMs. Ultimately, Blink provides an effective testbed for multimodal LLMs to catch up with human-level visual perception.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was funded in part by ONR Contract N00014-23-1-2417, and supported by NSF grant IIS-2212433.", + "bbox": [ + 212, + 178, + 787, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 233, + 323, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family (March 2024)", + "2. Acharya, M., Kafle, K., Kanan, C.: Tallyqa: Answering complex counting questions. In: AAAI (2019)", + "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "4. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015)", + "5. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)", + "6. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023)", + "7. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023)", + "8. Balntas, V., Lenc, K., Vedaldi, A., Mikolajczyk, K.: Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In: CVPR (2017)", + "9. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978)", + "0. Bell, S., Bala, K., Snavely, N.: Intrinsic images in the wild. ACM Trans. on Graphics (SIGGRAPH) 33(4) (2014)", + "1. Berrios, W., Mittal, G., Thrush, T., Kiela, D., Singh, A.: Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410 (2023)", + "2. Black, M.J., Anandan, P.: A framework for the robust estimation of optical flow. In: 1993 (4th) International Conference on Computer Vision. pp. 231-236. IEEE (1993)", + "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "4. Careaga, C., Aksoy, Y.: Intrinsic image decomposition via ordinal shading. ACM Trans. Graph. (2023)", + "5. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: CVPR (2021)" + ], + "bbox": [ + 225, + 266, + 787, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BLINK", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning (2023)", + "17. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)", + "18. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023)", + "19. Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. Advances in neural information processing systems 29 (2016)", + "20. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Changpinyo, S., Wu, J., Ruiz, C.R., Goodman, S., Wang, X., Tay, Y., Shakeri, S., Dehghani, M., Salz, D., Lucic, M., Tschannen, M., Nagrani, A., Hu, H., Joshi, M., Pang, B., Montgomery, C., Pietrzyk, P., Ritter, M., Piergiovanni, A., Minderer, M., Pavetic, F., Waters, A., Li, G., Alabdulmohsin, I., Beyer, L., Amelot, J., Lee, K., Steiner, A.P., Li, Y., Keysers, D., Arnab, A., Xu, Y., Rong, K., Kolesnikov, A., Seyedhosseini, M., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali-x: On scaling up a multilingual vision and language model (2023)", + "21. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levskaya, A., Ghemawat, S., Dev, S., Michalewski, H., Garcia, X., Misra, V., Robinson, K., Fedus, L., Zhou, D., Ippolito, D., Luan, D., Lim, H., Zoph, B., Spiridonov, A., Sepassi, R., Dohan, D., Agrawal, S., Omernick, M., Dai, A.M., Pillai, T.S., Pellat, M., Lewkowycz, A., Moreira, E., Child, R., Polozov, O., Lee, K., Zhou, Z., Wang, X., Saeta, B., Diaz, M., Firat, O., Catasta, M., Wei, J., Meier-Hellstern, K., Eck, D., Dean, J., Petrov, S., Fiedel, N.: Palm: Scaling language modeling with pathways (2022)", + "22. Contributors, O.: Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass (2023)", + "23. Contributors, X.: Xtuner: A toolkit for efficiently fine-tuning llm. https://github.com/InternLM/xtuner (2023)", + "24. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)", + "25. DO CT OR OF, P.E.: MACHINE PERCEPTION OF THREE-DIMENSIONAL, SO LIDS. Ph.D. thesis, MASSACHUSETTS INSTITUTE OF TECHNOLOGY (1961)", + "26. Dong, X., Zhang, P., Zang, Y., Cao, Y., Wang, B., Ouyang, L., Wei, X., Zhang, S., Duan, H., Cao, M., Zhang, W., Li, Y., Yan, H., Gao, Y., Zhang, X., Li, W., Li, J., Chen, K., He, C., Zhang, X., Qiao, Y., Lin, D., Wang, J.: Internl m-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420 (2024)", + "27. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)", + "29. Fu, S., Tamir, N., Sundaram, S., Chai, L., Zhang, R., Dekel, T., Isola, P.: Dreamsim: Learning new dimensions of human visual similarity using synthetic data (2023)", + "30. Fu, X., He, M., Lu, Y., Wang, W.Y., Roth, D.: Commonsense-t2i challenge: Can text-to-image generation models understand commonsense? arXiv preprint arXiv:2406.07546 (2024)", + "31. Fu, X., Zhang, S., Kwon, G., Perera, P., Zhu, H., Zhang, Y., Li, A.H., Wang, W.Y., Wang, Z., Castelli, V., Ng, P., Roth, D., Xiang, B.: Generate then select: Open-ended visual question answering guided by world knowledge. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Findings of the Association for Computational Linguistics: ACL 2023. pp. 2333–2346. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023-findings-acl.147 https://aclanthology.org/2023-findings-acl.147", + "32. Fu, X., Zhou, B., Chandratreya, I., Vondrick, C., Roth, D.: There's a time and place for reasoning beyond the image. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 1138-1149. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.81, https://aclanthology.org/2022.acl-long.81", + "33. Fu, X., Zhou, B., Chen, S., Yatskar, M., Roth, D.: Interpretable by design visual question answering. arXiv preprint arXiv:2305.14882 (2023)", + "34. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)", + "35. Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., Manocha, D., Zhou, T.: Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models (2023)", + "36. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)", + "37. Harris, C., Stephens, M., et al.: A combined corner and edge detector. In: Alvey vision conference. vol. 15, pp. 10-5244. Citeseer (1988)", + "38. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)", + "39. Hu, Y., Hua, H., Yang, Z., Shi, W., Smith, N.A., Luo, J.: Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699 (2022)", + "40. Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897 (2023)", + "41. Hu, Y., Stretcu, O., Lu, C.T., Viswanathan, K., Hata, K., Luo, E., Krishna, R., Fuxman, A.: Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. arXiv preprint arXiv:2312.03052 (2023)", + "42. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BLINK", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Lai, Z., Purushwalkam, S., Gupta, A.: The functional correspondence problem. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15772-15781 (2021)", + "44. Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092 (2023)", + "45. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)", + "46. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "48. Liu, F., Emerson, G., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics 11, 635-651 (2023)", + "49. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)", + "50. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)", + "51. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/", + "52. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)", + "53. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., Chen, K., Lin, D.: Mmbench: Is your multi-modal model an all-around player? (2023)", + "54. Liu, Y., Li, Z., Li, H., Yu, W., Huang, M., Peng, D., Liu, M., Chen, M., Li, C., Jin, L., et al.: On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895 (2023)", + "55. Liu, Y., Li, Z., Yang, B., Li, C., Yin, X., lin Liu, C., Jin, L., Bai, X.: On the hidden mystery ofOCR in large multimodal models (2024)", + "56. Lowe, D.G.: Object recognition from local scale-invariant features. In: Proceedings of the seventh IEEE international conference on computer vision. vol. 2, pp. 1150-1157. IEEE (1999)", + "57. Lu, J., Clark, C., Lee, S., Zhang, Z., Khosla, S., Marten, R., Hoiem, D., Kembhavi, A.: Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172 (2023)", + "58. Lu, P., Bansal, H., Xia, T., Liu, J., Li, C., Hajishirzi, H., Cheng, H., Chang, K.W., Galley, M., Gao, J.: Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255 (2023)", + "59. Marr, D.: Vision: A computational investigation into the human representation and processing of visual information. MIT press (2010)", + "60. Marr, D., Poggio, T.: Cooperative computation of stereo disparity: A cooperative algorithm is derived for extracting disparity information from stereo image pairs. Science 194(4262), 283-287 (1976)", + "61. Min, J., Lee, J., Ponce, J., Cho, M.: Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543 (2019)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Minsky, M., Papert, S.: An introduction to computational geometry. Cambridge, tiass., HIT 479(480), 104 (1969)", + "63. OpenAI: Gpt-4 technical report (2023)", + "64. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "66. Sarkar, A., Mai, H., Mahapatra, A., Lazebnik, S., Bhattachad, A.: Shadows don't lie and lines can't bend! generative models don't know projective geometry... for now. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28140-28149 (2024)", + "67. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "68. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022)", + "69. Shtedritski, A., Rupprecht, C., Vedaldi, A.: What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712 (2023)", + "70. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. CVPR (2021)", + "71. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)", + "72. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881 (2023)", + "73. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)", + "74. Team, I.: InternlM: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM (2023)", + "75. Team, M.N.: Introducing mpt-7b: A new standard for open-source, commercially usable lms (2023), www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-05", + "76. Torralba, A., Oliva, A.: Depth estimation from image structure. IEEE Transactions on pattern analysis and machine intelligence 24(9), 1226-1238 (2002)", + "77. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "78. Wang, F., Fu, X., Huang, J.Y., Li, Z., Liu, Q., Liu, X., Ma, M.D., Xu, N., Zhou, W., Zhang, K., et al.: Muirbench: A comprehensive benchmark for robust multi-image understanding. arXiv preprint arXiv:2406.09411 (2024)", + "79. Wang, J.Y., Adelson, E.H.: Layered representation for motion analysis. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. pp. 361-366. IEEE (1993)", + "80. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., Xu, J., Xu, B., Li, J., Dong, Y., Ding, M., Tang, J.: Cogvlm: Visual expert for pretrained language models (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BLINK", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "81. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022)", + "82. Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., Li, H.: Dire for diffusion-generated image detection. arXiv preprint arXiv:2303.09295 (2023)", + "83. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., Zhou, D.: Chain-of-thought prompting elicits reasoning in large language models (2022)", + "84. Yan, A., Yang, Z., Wu, J., Zhu, W., Yang, J., Li, L., Lin, K., Wang, J., McAuley, J., Gao, J., et al.: List items one by one: A new data source and learning paradigm for multimodal llms. arXiv preprint arXiv:2404.16375 (2024)", + "85. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023)", + "86. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)", + "87. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3081-3089 (2022)", + "88. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023)", + "89. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)", + "90. Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W., Sun, Y., Wei, C., Yu, B., Yuan, R., Sun, R., Yin, M., Zheng, B., Yang, Z., Liu, Y., Huang, W., Sun, H., Su, Y., Chen, W.: Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502 (2023)", + "91. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems 35, 27469-27483 (2022)", + "92. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "93. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 215, + 146, + 787, + 674 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Fu and Hu et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_model.json b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8c288e5bf1b19a122eb2b2dcbd73b1be018db170 --- /dev/null +++ b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_model.json @@ -0,0 +1,3969 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.228, + 0.146, + 0.777, + 0.19 + ], + "angle": 0, + "content": "BLINK: Multimodal Large Language Models Can See but Not Perceive" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.217, + 0.777, + 0.248 + ], + "angle": 0, + "content": "Xingyu Fu\\(^{1*}\\), Yushi Hu\\(^{2,3*}\\), Bangzheng Li\\(^{4}\\), Yu Feng\\(^{1}\\), Haoyu Wang\\(^{1}\\), Xudong Lin\\(^{5}\\), Dan Roth\\(^{1}\\), Noah A. Smith\\(^{2,3}\\), Wei-Chiu Ma\\(^{3\\dagger}\\), and Ranjay Krishna\\(^{2,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.257, + 0.772, + 0.302 + ], + "angle": 0, + "content": "1 University of Pennsylvania 2 University of Washington 3 Allen Institute for AI 4 University of California, Davis 5 Columbia University https://zeyofu.github.io/blink/" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.328, + 0.752, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.663 + ], + "angle": 0, + "content": "Figure 1: The BLINK Benchmark. BLINK contains 14 visual perception tasks that can be solved by humans \"within a blink\", but pose significant challenges for current multimodal LLMs. These tasks are inspired by classical computer vision problems and recast into multiple-choice questions for multimodal LLMs to answer. Notice that the visual prompts and questions in this figure are different from the actual ones used in the benchmark for illustrative purposes, and answers of the samples are provided in fn 6" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.688, + 0.744, + 0.84 + ], + "angle": 0, + "content": "Abstract. We introduce BLINK, a new benchmark for multimodal language models (LLMs) that focuses on core visual perception abilities not found in other evaluations. Most of the BLINK tasks can be solved by humans \"within a blink\" (e.g., relative depth estimation, visual correspondence, forensics detection, and multi-view reasoning). However, we find these perception-demanding tasks cast significant challenges for current multimodal LLMs because they resist mediation through natural language. BLINK reforms 14 classic computer vision tasks into 3,807 multiple-choice questions, paired with single or multiple images and visual prompting. While humans get \\(95.70\\%\\) accuracy on average, BLINK is surprisingly challenging for existing multimodal LLMs: even the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.246 + ], + "angle": 0, + "content": "best-performing GPT-4V and Gemini achieve accuracies of \\(51.26\\%\\) and \\(45.72\\%\\), only \\(13.17\\%\\) and \\(7.63\\%\\) higher than random guessing, indicating that such perception abilities have not \"emerged\" yet in recent multimodal LLMs. Our analysis also highlights that specialist CV models could solve these problems much better, suggesting potential pathways for future improvements. We believe BLINK will stimulate the community to help multimodal LLMs catch up with human-level visual perception." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.258, + 0.744, + 0.288 + ], + "angle": 0, + "content": "Keywords: Multi-modal Large Language Models \\(\\cdot\\) Vision-Language Benchmark \\(\\cdot\\) Visual Perception Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.311, + 0.376, + 0.327 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.79, + 0.553 + ], + "angle": 0, + "content": "Compared to today, computer vision was originally attempting to interpret images as projections of 3D scenes, not just processing 2D arrays of flat \"patterns\" [25, 59, 62]. In this pursuit, early research developed a series of intermediate tasks: they focused on understanding optical properties like reflectance [12, 79], 3D primitives through multi-view reasoning [38, 60], geometric reasoning through depth estimation [76], instance recognition through visual correspondence [56], affordance through keypoint grounding [37], and forensics through intrinsic images [9]. Yet in the modern era of large language models (LLMs), we, as a community, have focused less on such perceptual tasks, and instead have developed new tasks, mostly expressed in natural language, emphasizing the vision-language connection learned by multimodal LLMs [3, 6, 18, 20, 24, 26, 51, 52, 57, 63, 66, 73, 80]. This might be because many traditional computer vision tasks resist mediation through natural language, due to the inherent imprecision of language (e.g., it is challenging to precisely pinpoint a spatial keypoint through language)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.79, + 0.809 + ], + "angle": 0, + "content": "This paper aims to highlight crucial aspects of visual perception that have been overlooked when evaluating multimodal LLMs. To appropriately position our paper, let us revisit how we currently evaluate perception through using multimodal LLMs [44,45,48,53,54,58,90]. While many of these benchmarks have been popularized as the de facto evaluation measures for influential models like GPT-4V and Gemini-Pro, they conflate perception with language knowledge and reasoning. At the risk of singling out one benchmark, let us consider two questions highlighted in the popular MMBench [53]: “ Why is this hummingbird called ruby-throated?” and “ What will happen next? A: the person is gonna laugh B: the person is gonna cry.” For the first question, the vision subpart is to recognize the hummingbird. For the second, it only needs a coarse description of the image. Everything else is left to the language model to solve. Such a conflation has also been reported for other benchmarks by previous work [11,39,87]. Our experiments show that this conflation reductively evaluates perception as a dense captioning task. In other words, by replacing the image with a task-agnostic dense caption, our experiments show that a “blind” GPT-4 performs well on these “multimodal tasks”." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.79, + 0.84 + ], + "angle": 0, + "content": "In response, we propose BLINK. BLINK reimagines traditional computer vision problems through a format that allows us to evaluate multimodal LLMs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.688, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BLINK" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.159, + 0.752, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.429, + 0.79, + 0.556 + ], + "angle": 0, + "content": "Figure 2: Comparison between BLINK and previous benchmarks. BLINK has several novel features: (1) BLINK incorporates diverse visual prompts, like circles, boxes, and image masks, while previous benchmarks only have text questions and answers. (2) BLINK evaluates a more comprehensive range of visual perception abilities, like multi-view reasoning, depth estimation, and reflectance estimation. Prior benchmarks are generally more focused on recognition-based VQA. (3) BLINK contains \"visual\" commonsense problems that humans can answer within seconds, while prior benchmarks like [90] require domain knowledge. The samples of previous benchmarks are from [45, 53, 90]. Part of our samples are curated from [10, 19, 29, 32, 36, 43, 91]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.789, + 0.769 + ], + "angle": 0, + "content": "As partially demonstrated in Figure 15, BLINK consists of 14 classic computer vision tasks, ranging from low-level pattern matching (e.g., visual correspondences estimation) to mid-level reasoning (e.g., relative depth estimation), and extending to high-level visual understanding (e.g., visual similarity). The image tasks are meticulously selected such that they are difficult to solve by reducing the evaluation using dense captioning; instead, the models must perceive the contents of the image(s) to answer. We recast each traditional task into a modern question-answering format, where answer choices are either images or text. BLINK contains 3.8K questions across 7.3K images, where questions may contain multiple images that are curated from a wide range of datasets [8, 10, 19, 29, 32, 36, 42, 47], encompassing indoor household scenes as well as outdoor urban or natural environments. The questions and choices are either derived from the datasets, or" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.783, + 0.789, + 0.84 + ], + "angle": 0, + "content": "6 The answers of the examples in Figure 1 are as follows. Relative depth: B; jigsaw: A; multi-view reasoning: right; visual correspondence: A; semantic correspondence: C; forensics detection: final image; IQ test: D; visual similarity: upper one; functional correspondence: A; relative reflectance: they are about the same." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "manually written by humans. On average, each question can be solved by a human subject within a BLINK of an eye, except the IQ test." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.39 + ], + "angle": 0, + "content": "We carefully evaluate 17 multimodal LLMs with various sizes (i.e., 7B, 13B, 34B) on BLINK. We observe the paradox that while these problems are easy for humans (95.70% average accuracy), they are extremely hard for existing machinery - even GPT-4V model can only achieve 51.26% accuracy on average, which is 44.44% worse than humans, and 13.17% better than random guessing. We also experiment with specialist vision models and find that they perform much better than multimodal LLMs. For example, the specialist outperforms GPT-4V by 62.8% on visual correspondence estimation, 38.7% on relative depth estimation, and 34.6% on multi-view reasoning, in terms of absolute accuracy. Our findings indicate that the perceptual abilities of multimodal LLMs have been previously overestimated. Furthermore, these models may benefit from integrating insights from specialized models that excel in these areas. We believe BLINK can serve as an effective testbed for bridging the gap between traditional notions of perception and the modern generative capabilities of multimodal LLMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.388, + 0.43 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.789, + 0.674 + ], + "angle": 0, + "content": "Multimodal models: Inspired by the impressive success in recent large language models (LLMs) [13,21,63,77,93], a sequence of studies explore multimodal LMMs that can jointly understand vision and language information and generate textual answers through adding a modality adaption structure between a frozen visual encoder [27,65,71] and a frozen LLM [77,93]. Flamingo [3] and BLIP-2 [46] are two of the earliest works to explore these transformer-based multimodality conjunction structures. They first pre-train on image-text matching datasets [15,42,47,67] and then fine-tune on task-specific datasets such as visual question answer (VQA) [4,34]. Starting from LLaVA [49,51], people use LLM synthesized instruction-following chat data (which are in VQA format) for instruction tuning and achieve much better results [7,16,24,73]. There have been extended studies that explore further capabilities of multimodal LLMs, especially on VQA reasoning [30,33,39,40,68,78,84,92]. However, they mainly focus on the textual reasoning abilities within the multimodal LLMs and do not emphasize visual perceptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Multimodal benchmarks: Traditional vision-language datasets are designed to assess single-task capabilities, such as optical character recognition (OCR) [55], image captioning [47], and visual question answering [4,34]. However, these datasets are often not comprehensive enough to holistically assess multimodal LMMs on general perception and reasoning abilities. Many recent papers have built more comprehensive benchmarks. MME [28] is one of the earliest holistic benchmarks containing multi-modal Yes/No questions on the defined visual perception and language reasoning tasks. MM-Vet [89] includes six sub-features from the previous datasets including recognition-focused questions, OCR, and math, providing a diverse while discrete evaluation set. MMBench [53] covers more subjects and provides a more robust circular evaluation setting. Seed" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BLINK" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Bench [44,45] benchmark has a more diverse source of inputs, including multiple-image inputs and video, and includes more tasks. However, the visual perception questions in MME, MMBench, MM-Vet, and Seed-Bench are mainly extracted from existing VQA datasets or generated by GPT [63] from image descriptions such as COCO-Caption [47], and are recognition focused, covering topics such as object (attribute)recognition, and OCR. In contrast, we focus on multiple distinct nuanced perception abilities and recognition-level perception is only one of our focus. Some other multimodal benchmarks have distinct focuses. MMMU [90] aims at achieving expert-level artificial general intelligence by collecting domain-knowledge-required questions. HallusionBench [35] mainly tests the language hallucination and visual illusion phenomena. MathVista [58] presents exclusively mathematical domain visual questions based on images such as charts, tables, and diagrams. These benchmarks do not require human-level perception abilities as in BLINK and therefore cannot measure model visual perceptions holistically." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.386, + 0.478, + 0.403 + ], + "angle": 0, + "content": "3 The BLINK Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.79, + 0.59 + ], + "angle": 0, + "content": "Our goal is to faithfully evaluate the visual perception capabilities of existing Multimodal LLMs. We seek to study the visual perception gap between humans and machineries, and offer deeper insights into potential pathways towards achieving more generalized machine perception. Based on the observation that existing benchmarks predominantly focus on evaluating visual recognition abilities, we introduce a novel benchmark, BLINK, designed to enable both quantitative and qualitative evaluation of the nuanced perception capabilities of multimodal LLM across various dimensions. We unfold this section by illustrating the overall design of BLINK (§3.1) and discussing its unique features comparing with previous benchmarks. Then we describe each task in detail, providing an in-depth explanation of the data curation process (§3.2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.424, + 0.63 + ], + "angle": 0, + "content": "3.1 Overview of BLINK" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To ensure that one can effectively measure what Multimodal LLMs can or cannot perceive, we carefully select 14 tasks (see §3.2 for the full list) that are difficult to solve by reducing the evaluation into text-only questions using dense captioning. The tasks are drawn from either classic computer vision problems or recent applications of Multimodal LLMs, each of which requires a nuanced understanding of the visual data. They range from low-level pattern matching (e.g., visual correspondence) to mid-level spatial reasoning (e.g., relative depth), and up to high-level visual understanding (e.g., visual similarity). This variety allows for a systematic exploration of Multimodal LLMs' capabilities across different perceptual complexity layers. Furthermore, these visual tasks vary in granularity, ranging from pixels (e.g., relative reflectance) to patches (e.g., jigsaw) and extending to the full image (e.g., forensic detection), enabling us to evaluate models' proficiency in observing at various scales." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.128 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.148, + 0.486, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.336, + 0.496, + 0.435 + ], + "angle": 0, + "content": "Figure 3: Statistics of BLINK. 14 tasks range from pixel-level to image-level; from low-level pattern matching (e.g., visual correspondences) to mid-level reasoning (e.g., relative depth), further to high-level visual understanding (e.g., visual similarity)." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.169, + 0.785, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.37, + 0.791, + 0.414 + ], + "angle": 0, + "content": "Figure 4: Accuracies of multimodal LLMs on BLINK test set. Please refer to Table [1] and §4.2 for more results and discussions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.631 + ], + "angle": 0, + "content": "To facilitate the evaluation of multimodal LLMs, we recast all tasks as multiple-choice question-answering problems. The options for answers may include images or texts, while the questions themselves can feature either single or multiple images. Prompts are designed to be both textual and visual in nature. We re-purposed several existing vision datasets as well as collected new data. In total, we contribute 3.9K multiple-choice questions and 7.3K images, with an even distribution between the validation and test sets. Numbers of each task are reported in Figure 3, and more detailed statistics can be found in Appendix A.5. Key features of BLINK: Comparing with previous benchmarks, BLINK has the following novel features:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.642, + 0.789, + 0.718 + ], + "angle": 0, + "content": "- Visual prompting: Unlike existing benchmarks that support only text prompting, BLINK features a variety of visual prompts. This enables one to highlight specific areas within images, facilitating the evaluation of Multi-modal LLMs' detailed understanding of these regions. It also offers an interface for researchers to investigate the impact of visual prompting techniques." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.719, + 0.789, + 0.779 + ], + "angle": 0, + "content": "- Perception beyond recognition: Besides visual recognition, BLINK considers a diverse set of visual perception abilities, such as 3D reasoning, geometric understanding, affordance reasoning, etc. The breadth allows one to evaluate Multimodal LLMs from an unique array of perspectives." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.78, + 0.79, + 0.841 + ], + "angle": 0, + "content": "- \"Visual commonsense\" that does not require domain knowledge: The questions in BLINK are intentionally designed to be straightforward, requiring neither domain-specific knowledge nor expertise to answer. They are crafted in such a way that humans can solve them almost instantaneously," + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.642, + 0.79, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BLINK" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.207 + ], + "angle": 0, + "content": "typically within a few seconds. This allows us to explore the fundamental gap in visual perception gap between humans and Multimodal LLMs, highlighting the paradox that problems easily solved by humans often pose significant challenges for machines." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.208, + 0.787, + 0.265 + ], + "angle": 0, + "content": "- Interleaved image-text formats: BLINK features a heterogeneous question-answering format, wherein both questions and choices can be presented as text or images. This diversity compels Multimodal LLMs to genuinely understand the questions, pushing the boundaries of their interpretative capabilities." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.268, + 0.785, + 0.342 + ], + "angle": 0, + "content": "- Diverse image sources: BLINK comprises a wide range of in-the-wild images sourced from various origins, covering everything from indoor and outdoor scenes to object-centric views and landscapes. This collection spans abstract diagrams, synthesized images, and authentic photographs, ensuring a comprehensive examination of visual perception" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.208, + 0.787, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.384 + ], + "angle": 0, + "content": "The design principles of BLINK are also illustrated in Figure 2. We will now describe each task in detail." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.407, + 0.483, + 0.421 + ], + "angle": 0, + "content": "3.2 Dataset Collection Process" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.478 + ], + "angle": 0, + "content": "BLINK comprises 14 tasks, all of which have been repurposed into a multiple-choice question-answering format. These tasks utilize a diverse collection of images from various sources and we ensure that each image is unique." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.787, + 0.597 + ], + "angle": 0, + "content": "Visual correspondence: This task aims to evaluate the ability of Multimodal LLMs to understand and identify the same scene point across various viewpoints, lighting conditions, or time. We exploit HPatches [8] for this task. HPatches contains a number of image sequences, each of which are composed of images taken under different illuminations and/or viewpoints of a scene. For each question, we randomly sample two images and an interest point within them. Then we exploit the ground-truth homography to compute its correspondence. Finally, we randomly select three more interest points to serve as other choices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.703 + ], + "angle": 0, + "content": "Relative reflectance: This task aims to compare the reflectance (albedo) of two pixels. It allows us to evaluate Multimodal LLMs' understanding of material properties and their interaction with light, which is crucial for applications requiring high-fidelity visual interpretations. We curate our samples using human annotations from the Intrinsic Images in the Wild (IIW) dataset [10]. Each question is based on an image and two specified points. The goal is to identify which point is darker, or whether the two points have similar reflectance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Relative depth: Humans are good at judging relative depth [19]. This task can thus serve as a proxy to validate if the geometric understanding capabilities of existing multimodal LLMs are close to human. We curate our samples using human annotations from the Depth in the Wild [19] dataset. Given an image and two specified points, the task is to determine which point is closer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Spatial relation: Understanding spatial relationships between objects in a scene is essential for interpreting complex visual environments. However, modern Multimodal LLMs often struggles with spatial concepts such as \"left\" and \"right\" [88]. This task helps us evaluate whether the models finally possess this" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.153, + 0.364, + 0.162 + ], + "angle": 0, + "content": "Visual correspondence" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.163, + 0.36, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.352, + 0.165, + 0.457, + 0.178 + ], + "angle": 0, + "content": "A: Which point corresponds to the reference point 1?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.375, + 0.18, + 0.394, + 0.187 + ], + "angle": 0, + "content": "(a) A" + }, + { + "type": "image_footnote", + "bbox": [ + 0.43, + 0.18, + 0.447, + 0.187 + ], + "angle": 0, + "content": "(b) B" + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.192, + 0.394, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.192, + 0.419, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.192, + 0.448, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.199, + 0.394, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.199, + 0.419, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.199, + 0.448, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.22, + 0.329, + 0.23 + ], + "angle": 0, + "content": "Relative depth" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.23, + 0.346, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.356, + 0.237, + 0.432, + 0.244 + ], + "angle": 0, + "content": "Q: Which point is closer?" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.252, + 0.38, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.252, + 0.394, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.252, + 0.419, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.252, + 0.447, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.26, + 0.381, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.26, + 0.443, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.29, + 0.358, + 0.298 + ], + "angle": 0, + "content": "Multi-view reasoning" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.299, + 0.352, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.354, + 0.299, + 0.481, + 0.325 + ], + "angle": 0, + "content": "Q. The first image is from the beginning of the video and the second image is from the end. Is the camera moving towards left or right when shooting the video?" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.327, + 0.391, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.333, + 0.404, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.327, + 0.444, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.333, + 0.448, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.34, + 0.448, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.36, + 0.297, + 0.367 + ], + "angle": 0, + "content": "Art style" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.368, + 0.302, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.308, + 0.369, + 0.472, + 0.377 + ], + "angle": 0, + "content": "Q: Which image fits the missing part in the first image?" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.38, + 0.364, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.383, + 0.394, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.381, + 0.418, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.381, + 0.468, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.424, + 0.304, + 0.432 + ], + "angle": 0, + "content": "Counting" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.433, + 0.337, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.343, + 0.439, + 0.46, + 0.451 + ], + "angle": 0, + "content": "Q: How many fingers are in front of the bathtub?" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.453, + 0.371, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.461, + 0.37, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.468, + 0.37, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.454, + 0.444, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.468, + 0.433, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.454, + 0.444, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.468, + 0.458, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.496, + 0.29, + 0.503 + ], + "angle": 0, + "content": "IQ test" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.507, + 0.389, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.395, + 0.509, + 0.478, + 0.522 + ], + "angle": 0, + "content": "Q: Which image comes at the end?" + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.529, + 0.423, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.529, + 0.471, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.537, + 0.434, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.537, + 0.471, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.564, + 0.299, + 0.571 + ], + "angle": 0, + "content": "1 Semantic" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.573, + 0.362, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.366, + 0.574, + 0.483, + 0.586 + ], + "angle": 0, + "content": "Q: Which point is semantically similar to the reference point 1?" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.589, + 0.411, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.589, + 0.468, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.602, + 0.419, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.602, + 0.481, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.627, + 0.349, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.629, + 0.383, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.628, + 0.432, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.629, + 0.48, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.629, + 0.534, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.628, + 0.571, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.629, + 0.656, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.154, + 0.606, + 0.162 + ], + "angle": 0, + "content": "Relative reflectance" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.163, + 0.61, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.562, + 0.165, + 0.718, + 0.184 + ], + "angle": 0, + "content": "Q: Consider the surface color (color without shading) of the two points in the image. Which one is darker, or the color is about the same?" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.187, + 0.629, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.632, + 0.187, + 0.712, + 0.195 + ], + "angle": 0, + "content": "(b) B is darker" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.196, + 0.631, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.196, + 0.65, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.652, + 0.187, + 0.713, + 0.195 + ], + "angle": 0, + "content": "(b) B is darker" + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.196, + 0.691, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.187, + 0.712, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.196, + 0.712, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.221, + 0.595, + 0.23 + ], + "angle": 0, + "content": "Spatial reasoning" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.231, + 0.561, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.57, + 0.233, + 0.71, + 0.245 + ], + "angle": 0, + "content": "Q: Is the bed at the right side of the dining table?" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.25, + 0.6, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.251, + 0.627, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.251, + 0.653, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.251, + 0.68, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.258, + 0.692, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.291, + 0.546, + 0.299 + ], + "angle": 0, + "content": "Jigsaw" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.3, + 0.575, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.577, + 0.304, + 0.739, + 0.311 + ], + "angle": 0, + "content": "Q: Which image fits the missing part in the first image?" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.314, + 0.651, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.314, + 0.662, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.314, + 0.726, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.36, + 0.603, + 0.368 + ], + "angle": 0, + "content": "Object localization" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.369, + 0.606, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.612, + 0.373, + 0.721, + 0.385 + ], + "angle": 0, + "content": "Q: Which bounding box more accurately encloses the bun?" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.393, + 0.69, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.401, + 0.69, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.393, + 0.712, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.401, + 0.69, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.408, + 0.706, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.425, + 0.606, + 0.433 + ], + "angle": 0, + "content": "Forensics detection" + }, + { + "type": "image_footnote", + "bbox": [ + 0.512, + 0.437, + 0.677, + 0.445 + ], + "angle": 0, + "content": "Q: Which image is most likely to be a real photograph?" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.445, + 0.561, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.446, + 0.613, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.446, + 0.666, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.446, + 0.72, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.495, + 0.584, + 0.503 + ], + "angle": 0, + "content": "Visual similarity" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.503, + 0.574, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.577, + 0.507, + 0.741, + 0.515 + ], + "angle": 0, + "content": "Q: Which image is most similar to the reference image?" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.518, + 0.637, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.518, + 0.663, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.518, + 0.69, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.518, + 0.73, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.564, + 0.637, + 0.572 + ], + "angle": 0, + "content": "Functional correspondence" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.572, + 0.655, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.655, + 0.576, + 0.743, + 0.594 + ], + "angle": 0, + "content": "Q Which point is functionally different from reference point 1 during mixing?" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.596, + 0.68, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.596, + 0.709, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.603, + 0.68, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.603, + 0.709, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.61, + 0.68, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.61, + 0.726, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.629, + 0.536, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.628, + 0.571, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.629, + 0.592, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.629, + 0.656, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.611, + 0.636, + 0.629, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.631, + 0.656, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.631, + 0.682, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.657, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Figure 5: Qualitative results on BLINK. For each task, we show the choice of LLaVAv1.6-34B [51], Qwen-VL-Max [7], Gemini Pro [73], GPT-4V [63], and humans. Red choice indicates the ground truth. Notice that the markers are intentionally enlarged for visualization purposes, and we make some images inset images to save space. For IQ test, the third image is constructed by overlaying the first and second images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.761, + 0.789, + 0.841 + ], + "angle": 0, + "content": "vital skill. We curate our samples from the Visual Spatial Reasoning [48] dataset. Each sample contains an image and a claim. The task is to determine if the claim is true or false. We reformat the claims into binary questions via GPT-3.5 [13]. Multi-view reasoning: This task is centered on evaluating the multi-view reasoning capabilities of Multimodal LLMs. The objective is to deduce the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BLINK" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "relative camera motion based on two images of an object captured from different viewpoints. Our data is sourced from the Wild6D dataset [91], which features videos of various objects recorded in diverse settings. We select two random frames from each video to calculate the relative camera motion. Recognizing that even humans might struggle to precisely articulate 3D motion details, we simplify the task by classifying motions into two broad categories: moving towards the left or moving towards the right. Despite the simplicity of these questions, as we will later demonstrate, they pose significant challenges for current models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.79, + 0.362 + ], + "angle": 0, + "content": "Jigsaw: This task assesses the ability of Multimodal LLMs to recognize and group patterns, aligning patches based on continuity in shape, color, and texture. We utilize images from the TARA dataset [32] and segment each of them into a 3x3 grid. We retain the three segments from the upper left corner as the reference image, and treat the central segment along with a randomly chosen segment as options. The objective is to identify the correct patch (i.e., the central patch)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.79, + 0.501 + ], + "angle": 0, + "content": "Art style: This task evaluates Multimodal LLMs capability to analyze and discern both local and global similarities in art styles among multiple images. Although there have been prior efforts to incorporate art-related questions into evaluation [90], such attempts primarily focused on questions requiring expert-level knowledge, including deducing an artist's name and understanding historical contexts, rather than on direct image comparison. For this task, we collect paintings and their stylistic information from WikiArt. Given one reference painting image and two other paintings as options, the model is tasked with identifying the one that most closely shares the art style of the reference painting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Object localization: The ability to accurately detect and localize objects is critical for scene understanding. While previous benchmarks [53] have explored this task, their focus was primarily on coarse localization. For instance, they might only ask the model if an object is located at the \"top\" or \"right\" side of an image. BLINK, in contrast, aims for a more fine-grained evaluation. We exploit images from LVIS [36], randomly sampling one object per image along with its ground-truth bounding box. Then we add Gaussian noise to the ground-truth box to create a confounding box. The goal is to select the correct one." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Counting: This task evaluate Multimodal LLMs' abilities in detection, recognition, and compositional reasoning, particularly in complex scenes where objects may overlap, be occluded, or vary in size and appearance. We select our questions from the TallyQA dataset [2], known for its challenging human-written counting questions. Each sample comprises an image, a question, and a numerical answer. In addition, we randomly select three numbers to serve as confounding options." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Forensic detection: Recent advances in generative AI have raised concerns about malicious uses and have prompted calls for the automatic detection of fake content. To evaluate whether Multimodal LLMs can fulfill such a role, we construct sets of real and synthesized images that describe similar scenes and ask the models to identify the real ones. Specifically, we first generate synthetic images using Stable Diffusion XL [64], employing COCO captions [47] as prompts. Then, we manually search online using these captions as descriptions and select high-quality photographs as the real images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "IQ test: This task evaluates the ability of Multimodal LLMs to engage in graphical reasoning, without requiring any domain-specific knowledge. We manually collect test samples, along with human explanations, from various public, license-friendly online sources. Given visual examples and a selection of images, the objective is to identify the image that either continues the pattern established by the examples or is spatially consistent with them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Visual similarity: This task aims to verify whether Multimodal LLMs possess a nuanced understanding of visual features, patterns, and aesthetics at a level comparable to humans. We select our samples from the DreamSim dataset [29]. Given a reference image alongside two alternative images, the objective is to identify the image that more closely resembles the reference image perceptually." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Semantic correspondence: This task focuses on identifying and matching semantically similar yet visually distinct elements across images, thereby evaluating the ability of Multimodal LLMs to understand the underlying semantics of object parts. Our samples are sourced from the SPair-71k dataset [61], which features pairs of images with multiple corresponding semantic points. For each task, we randomly select one semantic point in an image as a reference, and provide the matching point alongside three random semantic points in the paired image as options. The objective is to accurately identify the correct matches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Functional correspondence: The task aims to identify points that are functionally similar across objects. It challenges Multimodal LLMs to extend their understanding beyond mere semantics, enabling them to infer the diverse functions an object can perform in various contexts. Such capability is crucial for applications in robotics. We derive our samples from the FunKPoint dataset [43], which features paired images annotated for functional correspondences. Following a method analogous to semantic correspondence, we present an action alongside two object images. One image includes a reference point, while the other offers four potential points. The objective is to select the point that best matches the reference in terms of functional affordances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.787, + 0.615 + ], + "angle": 0, + "content": "Data quality control: To guarantee the quality of BLINK, we manually go through all collected data and filter out data that are ambiguous." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.639, + 0.377, + 0.656 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.764 + ], + "angle": 0, + "content": "In this section, we first describe the experimental setup and the baselines (\\$4.1). Then we present a comprehensive evaluation of 16 recent multimodal LLMs (\\$4.2). We demonstrate that while humans can answer the questions with high accuracy, BLINK is challenging for existing models. Finally, we provide detailed analyses on multiple experimental settings, including the effect of reducing images to captions, sensitivity to different visual prompts, and error analysis (\\$4.3)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.785, + 0.429, + 0.8 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Multimodal LLMs: We evaluate BLINK on 16 recent models: MiniGPT-4-v2 [16], OpenFlamingo-v2 [5], InstructBLIP (7B and 13B) [24], CogVLM [80]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BLINK" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "LLaVA(v1, v1.5, v1.6, internLM, and xtuner versions, model size 7B, 13B, and 34B) [23,26,50-52], Yi-VL (6B and 34B), Qwen-VL-MAX [7], Gemini Pro [73], Claude 3 Opus [1] and GPT-4V(vision) [63]. See Appendix B for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.194, + 0.788, + 0.3 + ], + "angle": 0, + "content": "Evaluation setup: We follow standard setups as in the VLMEvalKit [22], where the temperature is set to 0 and retry is set to 10. However, we do not resize the images during any experiment. For the models that do not support multiple images as input, we concatenate the images as input. We extract the choice from the models' output with a set of pre-defined rules and GPT-3.5-turbo [13]. We refer the readers to Appendix A for more details on visual prompting, how we generate the answers in BLINK, and the human evaluation protocol." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.322, + 0.373, + 0.336 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.789, + 0.652 + ], + "angle": 0, + "content": "Overall performance: As shown in Table 1, the mean accuracy of 7B and 13B open-source Multimodal LLMs hover around \\(35 - 42\\%\\), which is similar to random guess \\((38.09\\%)\\). The most proficient open-source model, LLaVA-v1.6-34B, achieves an accuracy of \\(45.05\\%\\). Even the most advanced models, GPT-4V and Gemini Pro and Claude 3 OPUS, achieve accuracies of only \\(51.26\\%\\), \\(45.72\\%\\), and \\(44.11\\%\\) respectively. Their performance are merely \\(13.17\\%\\), \\(7.63\\%\\) and \\(6.02\\%\\) better than random guessing and lag behind human performance by \\(44.44\\%\\), \\(49.98\\%\\) and \\(51.59\\%\\). Notably, for certain tasks such as jigsaw, semantic correspondence, multi-view reasoning, object localization, and relative reflectance, some multimodal LLMs even underperform compared to random guessing. Some qualitative results are shown in Figure 5. Detailed scores on the validation set are in Appendix C. In which tasks do multimodal LLMs show relative strengths and weaknesses? Figure 4 shows the accuracies of the best-performing models on BLINK: LLaVA-v1.6-34B [51], Gemini Pro [73], and GPT-4V [63]. We observe that multimodal LLMs perform relatively better on spatial reasoning, art style, and counting tasks, in which they are much better than random guessing. The models also demonstrate some capability in relative depth and forensics detection. Overall, they are doing relatively well on mid-level perception tasks. In terms of granularity, the models in general perform better on image-level tasks and struggle on pixel-level and crop-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.789, + 0.788 + ], + "angle": 0, + "content": "GPT-4V behaves differently: Figure 4 and Table 1 show an interesting phenomenon: GPT-4V's performance pattern is different from other models. Compared with its counterparts, GPT-4V is much better in visual similarity, art style, jigsaw, and multi-view reasoning. Specifically, its performance on visual similarity is \\(29\\%\\) better than Gemini Pro, demonstrating that GPT-4V possesses a nuanced understanding of visual patterns and aesthetics that is similar to humans. In contrast, Gemini Pro and LLaVA have similar performance patterns. Human performance: Human evaluators achieve over \\(95\\%\\) accuracy across most tasks, with an average accuracy of \\(95.70\\%\\). This performance disparity between" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.654, + 0.812 + ], + "angle": 0, + "content": "7 More details are at the official website at https://www.01.ai/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Note that the human score for IQ test is annotated by authors. It may not reflect typical human performance, which is also expected to vary." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.145, + 0.778, + 0.712 + ], + "angle": 0, + "content": "
Validation (1,901)Test (1,906)Similarity (136)Counting (120)Depth (124)Jigsaw (150)Art (117)Fun.Corr. (130)
Random Choice38.0938.09502550505025
Human95.6795.7096.7093.7599.1999.0095.3080.77
Open-source multimodal LLMs
MiniGPT-4-v2 1634.2334.5752.9410.8349.1926.0047.8618.46
OpenFlamingo-v2 539.1838.3255.1521.6754.0346.0052.1436.15
InstructBLIP-7B 2439.7238.6546.3229.1750.8154.0047.8623.85
InstructBLIP-13B 2442.2439.5846.3230.8350.0054.0050.4322.31
LLaVAInternLM2-7B 7437.7136.0652.9452.5052.4234.6730.7723.08
Yi-VL-6B 738.7241.2446.3246.6756.4550.0053.8523.85
Yi-VL-34B 741.6842.7850.0058.3353.2354.0046.1539.23
LLaVA-v1.5-7B-xtuner 2339.3640.8146.3253.3350.8154.0047.8623.85
LLaVA-v1.5-13B-xtuner 2342.0041.3146.3245.0054.0353.3347.8626.15
CogVLM 8041.5439.3846.3238.3350.8152.6749.5723.85
LLaVA-v1.5-7B 4937.1338.0146.3243.3351.6111.3347.8621.54
LLaVA-v1.5-13B 4942.6640.5546.3250.0047.5854.0047.8620.77
LLaVA-v1.6-34B 5146.8045.0546.3268.3364.5256.6747.0130.77
API-based models
Qwen-VL-Max 740.2841.9451.4755.8358.873.3337.6128.46
Gemini Pro 7345.1645.7255.8865.0050.0054.0049.5732.31
Claude 3 OPUS 144.0544.1170.5949.1757.2632.6760.6822.31
GPT-4V(ison) 6351.1451.2683.0960.8358.8762.6778.6331.54
GPT-4 Turbo 6354.6153.8983.0960.8366.9466.0081.2031.54
GPT-4o 6360.0459.0365.4451.6764.5258.0082.9139.23
Sem.Corr. (140)Spatial (143)Local. (125)Vis.Corr. (172)Multi-view (133)Reflect. (134)Forensic (132)IQ (150)
Random Choice255050255033.332525
Human96.0798.2598.0099.4292.4895.14100.0080.00
Open-source multimodal LLMs
MiniGPT-4-v2 1626.4351.7556.0023.8452.6331.3417.4219.33
OpenFlamingo-v2 523.5746.8552.0025.0041.3543.2815.9123.33
InstructBLIP-7B 2425.0055.2444.8022.6758.6529.8529.5523.33
InstructBLIP-13B 2422.8664.3452.0020.9354.1446.2713.6426.00
LLaVAInternLM2-7B 7422.1474.1348.0021.5141.3532.843.7914.67
Yi-VL-6B 726.4372.7349.6029.6548.1229.8520.4523.33
Yi-VL-34B 721.4370.6354.4023.8441.3546.2717.4222.67
LLaVA-v1.5-7B-xtuner 2324.2974.8345.6023.8442.1126.8736.3621.33
LLaVA-v1.5-13B-xtuner 2322.1477.6248.0022.0941.3546.2729.5518.67
CogVLM 8023.5767.1343.2020.9357.1426.8724.2426.67
LLaVA-v1.5-7B 4932.1470.6348.8020.3549.6236.5728.0324.00
LLaVA-v1.5-13B 4923.5767.8347.2020.3541.3545.5227.2728.00
LLaVA-v1.6-34B 5127.8676.2241.6027.3346.6229.8541.6726.00
API-based models
Qwen-VL-Max 729.2977.6249.6022.6753.3849.2547.7322.00
Gemini Pro 7322.1467.1346.4037.2141.3546.2745.4527.33
Claude 3 OPUS 120.7157.3446.4031.4057.8927.6162.1221.33
GPT-4V(ison) 6330.0072.0350.4037.2158.6538.8130.3024.67
GPT-4 Turbo 6332.8667.1348.8042.4457.1434.3351.5230.67
GPT-4o 6345.7176.9256.0071.5160.1538.8185.6130.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.712, + 0.785, + 0.739 + ], + "angle": 0, + "content": "Table 1: Results of different models on the BLINK test set. The first row shows task names and number of test data. Average scores on validation set are also included." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "humans and multimodal LLMs highlights the significant visual perception gap that exists between current machine learning models and humans in perceiving, processing, and understanding complex visual and textual context." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BLINK" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.148, + 0.444, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.242, + 0.45, + 0.3 + ], + "angle": 0, + "content": "Figure 6: Performance of using image caption + text-only GPT-4 vs. GPT-4 Vision on MMBench [53], MMMU [90], and BLINK (\\$4.3)." + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.153, + 0.785, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.464, + 0.234, + 0.788, + 0.294 + ], + "angle": 0, + "content": "Figure 7: Accuracy of GPT-4V with different visual prompts (e.g., different circle sizes, colors) on relative depth, relative reflectance, and visual correspondence tasks. More discussions in §4.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.341, + 0.332, + 0.357 + ], + "angle": 0, + "content": "4.3 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.789, + 0.567 + ], + "angle": 0, + "content": "Is dense captioning all you need for a multimodal LLM benchmark? To answer the question, we reduce multimodal benchmarks to a text-only problem. Specifically, we convert images into task-agnostic dense image captions with GPT-4V. The dense caption describes detailed information about the image and the visual prompts (e.g., where each circle is), using language. For each multimodal question, we prompt the text-only GPT-4-0125-preview model with image captions and the textual question and evaluate if the \"blind\" GPT-4 can answer the question. We call this Caption + LLM. This experiment is predicated on the hypothesis that captioning involves predominantly recognition-centric perception. If using captions along with text-only LLMs yields performance comparable to or surpassing that achieved through the integration of images with multimodal LLMs, then the perception demands of that benchmark are primarily confined to recognition only." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.765 + ], + "angle": 0, + "content": "We experiment with BLINK, MMBench [53] and MMMU [90], as illustrated in Figure 6. Surprisingly, we find that the Caption + LLM setting achieves better results on MMBench than GPT-4V (with \\(5.7\\%\\) increase in accuracy). On MMMU, Caption + LLM achieves \\(47.2\\%\\) accuracy, which is \\(9.6\\%\\) lower than GPT-4V performance, but is still much better than random guessing. On BLINK, Caption + LLM fails, achieving random guessing performance. These results indicate that dense captions cover the visual information needed for MMBench. For MMMU, image captions carry a large portion of visual information needed to answer the domain-knowledge-specific questions. Meanwhile, the performance decrease observed in BLINK suggests the necessity for advanced perceptual abilities beyond what is currently attainable with general captions. This variance highlights the limitations of existing multimodal LLM benchmarks in addressing the full spectrum of visual perception." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Effect of visual prompting on BLINK: Several BLINK tasks involve visual prompting. Prior work [69] shows that factors like shape, size, and color may affect task performance, and circles give the best overall performance. Following [69], we adopt circles in BLINK and analyze the effect of circle sizes and colors on multiple tasks in Figure 7. We experiment with relative depth, relative reflectance," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.145, + 0.758, + 0.232 + ], + "angle": 0, + "content": "
TaskVis.Corr.DepthMulti-viewSem.Corr.ForensicReflect.
Random25.0050.0050.0025.0025.0033.33
Human99.5699.5992.1094.60100.0099.63
Gemini Pro42.4440.3244.3626.6250.7645.52
GPT-4V33.7259.6855.6428.7834.0938.81
SpecialistDIFT 72DepthAnything86LoFTR 70DIFT 72ORDINAL SHADING 14
96.5197.5890.2271.2268.9477.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.234, + 0.787, + 0.261 + ], + "angle": 0, + "content": "Table 2: Comparison between multimodal LLMs, specialists, and human performance on the BLINK dev set. The specialists perform much better than multimodal LLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.425 + ], + "angle": 0, + "content": "and visual correspondence, with 100 validation set samples per task. The images are all reshaped to 1024px height. We experiment with circles with 5px, 10px, and 15px radius, and with red or gray color. We find that red is better than gray for all tasks. Also, the optimal circle size is task-dependent. On average 10px circles work the best, and we use it for all evaluations in this paper. The experiments suggest that visual prompting can have a big impact on multimodal LLM performance, and improving visual prompts or improving model robustness to prompt variation is a promising direction for future research [85]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.79, + 0.577 + ], + "angle": 0, + "content": "Can specialist models solve BLINK tasks? Specialists can serve as a proxy upper bound of how good multimodal LLMs could be. We download the trained checkpoints for six specialist models and evaluate them on BLINK. As shown in Table 2, the specialists perform much better than GPT-4V and Gemini Pro, outperforming the best multimodal LLM by \\(18\\%\\) to \\(57\\%\\) on these tasks. Specifically, DepthAnything [86] and DIFT [72] achieve human-level performance on depth estimation and visual correspondence, whereas multimodal LLMs fail miserably. This sheds light on the possibility that multimodal LLMs may progress on these tasks given the correct data and training strategy. For instance, one possible way is to distill existing specialist models into multimodal LLMs [41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.789, + 0.713 + ], + "angle": 0, + "content": "Error analysis of GPT-4V: We randomly sampled 140 error instances made by GPT-4V on BLINK, 10 per task, and meticulously examined them. The most common types of errors are: Hallucinate fine-grained patterns and attributes (24.2%): the model hallucinates the nuanced details of objects. This error is most common for relative reflectance, forensics detection, and jigsaw tasks. Hallucinate visual prompt locations (20.0%): the circle location described by the model is wrong. This is common for visual correspondence and relative depth tasks. Other errors include Failures on capturing overall setting or style (8.6%), and Failures on grounding an object (5.7%). More details are in Appendix C.3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.36, + 0.75 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduced BLINK, a new multimodal LLM benchmark that evaluates core visual perception abilities not found in existing evaluations. While these tasks seem trivial for humans to solve \"within a blink\", we find they pose significant challenges for current multimodal LLMs. Ultimately, Blink provides an effective testbed for multimodal LLMs to catch up with human-level visual perception." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BLINK" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.21 + ], + "angle": 0, + "content": "This work was funded in part by ONR Contract N00014-23-1-2417, and supported by NSF grant IIS-2212433." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.234, + 0.324, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.788, + 0.296 + ], + "angle": 0, + "content": "1. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family (March 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.788, + 0.323 + ], + "angle": 0, + "content": "2. Acharya, M., Kafle, K., Kanan, C.: Tallyqa: Answering complex counting questions. In: AAAI (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.788, + 0.379 + ], + "angle": 0, + "content": "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.38, + 0.788, + 0.421 + ], + "angle": 0, + "content": "4. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.788, + 0.477 + ], + "angle": 0, + "content": "5. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.478, + 0.788, + 0.518 + ], + "angle": 0, + "content": "6. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.788, + 0.56 + ], + "angle": 0, + "content": "7. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.788, + 0.589 + ], + "angle": 0, + "content": "8. Balntas, V., Lenc, K., Vedaldi, A., Mikolajczyk, K.: Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.788, + 0.617 + ], + "angle": 0, + "content": "9. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.788, + 0.645 + ], + "angle": 0, + "content": "0. Bell, S., Bala, K., Snavely, N.: Intrinsic images in the wild. ACM Trans. on Graphics (SIGGRAPH) 33(4) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.788, + 0.686 + ], + "angle": 0, + "content": "1. Berrios, W., Mittal, G., Thrush, T., Kiela, D., Singh, A.: Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.788, + 0.728 + ], + "angle": 0, + "content": "2. Black, M.J., Anandan, P.: A framework for the robust estimation of optical flow. In: 1993 (4th) International Conference on Computer Vision. pp. 231-236. IEEE (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.788, + 0.771 + ], + "angle": 0, + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.788, + 0.799 + ], + "angle": 0, + "content": "4. Careaga, C., Aksoy, Y.: Intrinsic image decomposition via ordinal shading. ACM Trans. Graph. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "5. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: CVPR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "17. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.275 + ], + "angle": 0, + "content": "18. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.277, + 0.787, + 0.304 + ], + "angle": 0, + "content": "19. Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.305, + 0.787, + 0.416 + ], + "angle": 0, + "content": "20. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Changpinyo, S., Wu, J., Ruiz, C.R., Goodman, S., Wang, X., Tay, Y., Shakeri, S., Dehghani, M., Salz, D., Lucic, M., Tschannen, M., Nagrani, A., Hu, H., Joshi, M., Pang, B., Montgomery, C., Pietrzyk, P., Ritter, M., Piergiovanni, A., Minderer, M., Pavetic, F., Waters, A., Li, G., Alabdulmohsin, I., Beyer, L., Amelot, J., Lee, K., Steiner, A.P., Li, Y., Keysers, D., Arnab, A., Xu, Y., Rong, K., Kolesnikov, A., Seyedhosseini, M., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali-x: On scaling up a multilingual vision and language model (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.787, + 0.569 + ], + "angle": 0, + "content": "21. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levskaya, A., Ghemawat, S., Dev, S., Michalewski, H., Garcia, X., Misra, V., Robinson, K., Fedus, L., Zhou, D., Ippolito, D., Luan, D., Lim, H., Zoph, B., Spiridonov, A., Sepassi, R., Dohan, D., Agrawal, S., Omernick, M., Dai, A.M., Pillai, T.S., Pellat, M., Lewkowycz, A., Moreira, E., Child, R., Polozov, O., Lee, K., Zhou, Z., Wang, X., Saeta, B., Diaz, M., Firat, O., Catasta, M., Wei, J., Meier-Hellstern, K., Eck, D., Dean, J., Petrov, S., Fiedel, N.: Palm: Scaling language modeling with pathways (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.571, + 0.787, + 0.599 + ], + "angle": 0, + "content": "22. Contributors, O.: Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.6, + 0.787, + 0.628 + ], + "angle": 0, + "content": "23. Contributors, X.: Xtuner: A toolkit for efficiently fine-tuning llm. https://github.com/InternLM/xtuner (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.67 + ], + "angle": 0, + "content": "24. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.713 + ], + "angle": 0, + "content": "25. DO CT OR OF, P.E.: MACHINE PERCEPTION OF THREE-DIMENSIONAL, SO LIDS. Ph.D. thesis, MASSACHUSETTS INSTITUTE OF TECHNOLOGY (1961)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.783 + ], + "angle": 0, + "content": "26. Dong, X., Zhang, P., Zang, Y., Cao, Y., Wang, B., Ouyang, L., Wei, X., Zhang, S., Duan, H., Cao, M., Zhang, W., Li, Y., Yan, H., Gao, Y., Zhang, X., Li, W., Li, J., Chen, K., He, C., Zhang, X., Qiao, Y., Lin, D., Wang, J.: Internl m-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "27. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BLINK" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "28. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.191, + 0.788, + 0.219 + ], + "angle": 0, + "content": "29. Fu, S., Tamir, N., Sundaram, S., Chai, L., Zhang, R., Dekel, T., Isola, P.: Dreamsim: Learning new dimensions of human visual similarity using synthetic data (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.22, + 0.788, + 0.26 + ], + "angle": 0, + "content": "30. Fu, X., He, M., Lu, Y., Wang, W.Y., Roth, D.: Commonsense-t2i challenge: Can text-to-image generation models understand commonsense? arXiv preprint arXiv:2406.07546 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.789, + 0.359 + ], + "angle": 0, + "content": "31. Fu, X., Zhang, S., Kwon, G., Perera, P., Zhu, H., Zhang, Y., Li, A.H., Wang, W.Y., Wang, Z., Castelli, V., Ng, P., Roth, D., Xiang, B.: Generate then select: Open-ended visual question answering guided by world knowledge. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Findings of the Association for Computational Linguistics: ACL 2023. pp. 2333–2346. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023-findings-acl.147 https://aclanthology.org/2023-findings-acl.147" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.789, + 0.444 + ], + "angle": 0, + "content": "32. Fu, X., Zhou, B., Chandratreya, I., Vondrick, C., Roth, D.: There's a time and place for reasoning beyond the image. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 1138-1149. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.81, https://aclanthology.org/2022.acl-long.81" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.788, + 0.472 + ], + "angle": 0, + "content": "33. Fu, X., Zhou, B., Chen, S., Yatskar, M., Roth, D.: Interpretable by design visual question answering. arXiv preprint arXiv:2305.14882 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.788, + 0.514 + ], + "angle": 0, + "content": "34. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.788, + 0.57 + ], + "angle": 0, + "content": "35. Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., Manocha, D., Zhou, T.: Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.571, + 0.788, + 0.613 + ], + "angle": 0, + "content": "36. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.788, + 0.642 + ], + "angle": 0, + "content": "37. Harris, C., Stephens, M., et al.: A combined corner and edge detector. In: Alvey vision conference. vol. 15, pp. 10-5244. Citeseer (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.788, + 0.671 + ], + "angle": 0, + "content": "38. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.788, + 0.699 + ], + "angle": 0, + "content": "39. Hu, Y., Hua, H., Yang, Z., Shi, W., Smith, N.A., Luo, J.: Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.788, + 0.741 + ], + "angle": 0, + "content": "40. Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.788, + 0.784 + ], + "angle": 0, + "content": "41. Hu, Y., Stretcu, O., Lu, C.T., Viswanathan, K., Hata, K., Luo, E., Krishna, R., Fuxman, A.: Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. arXiv preprint arXiv:2312.03052 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.788, + 0.84 + ], + "angle": 0, + "content": "42. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.789, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "43. Lai, Z., Purushwalkam, S., Gupta, A.: The functional correspondence problem. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15772-15781 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "44. Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.788, + 0.271 + ], + "angle": 0, + "content": "45. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.788, + 0.311 + ], + "angle": 0, + "content": "46. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.788, + 0.367 + ], + "angle": 0, + "content": "47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.788, + 0.394 + ], + "angle": 0, + "content": "48. Liu, F., Emerson, G., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics 11, 635-651 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.788, + 0.421 + ], + "angle": 0, + "content": "49. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.788, + 0.447 + ], + "angle": 0, + "content": "50. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.788, + 0.488 + ], + "angle": 0, + "content": "51. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.788, + 0.515 + ], + "angle": 0, + "content": "52. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.788, + 0.555 + ], + "angle": 0, + "content": "53. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., Chen, K., Lin, D.: Mmbench: Is your multi-modal model an all-around player? (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.788, + 0.596 + ], + "angle": 0, + "content": "54. Liu, Y., Li, Z., Li, H., Yu, W., Huang, M., Peng, D., Liu, M., Chen, M., Li, C., Jin, L., et al.: On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.788, + 0.623 + ], + "angle": 0, + "content": "55. Liu, Y., Li, Z., Yang, B., Li, C., Yin, X., lin Liu, C., Jin, L., Bai, X.: On the hidden mystery ofOCR in large multimodal models (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.788, + 0.663 + ], + "angle": 0, + "content": "56. Lowe, D.G.: Object recognition from local scale-invariant features. In: Proceedings of the seventh IEEE international conference on computer vision. vol. 2, pp. 1150-1157. IEEE (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.788, + 0.704 + ], + "angle": 0, + "content": "57. Lu, J., Clark, C., Lee, S., Zhang, Z., Khosla, S., Marten, R., Hoiem, D., Kembhavi, A.: Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.788, + 0.746 + ], + "angle": 0, + "content": "58. Lu, P., Bansal, H., Xia, T., Liu, J., Li, C., Hajishirzi, H., Cheng, H., Chang, K.W., Galley, M., Gao, J.: Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.788, + 0.772 + ], + "angle": 0, + "content": "59. Marr, D.: Vision: A computational investigation into the human representation and processing of visual information. MIT press (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.788, + 0.813 + ], + "angle": 0, + "content": "60. Marr, D., Poggio, T.: Cooperative computation of stereo disparity: A cooperative algorithm is derived for extracting disparity information from stereo image pairs. Science 194(4262), 283-287 (1976)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "61. Min, J., Lee, J., Ponce, J., Cho, M.: Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BLINK" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "62. Minsky, M., Papert, S.: An introduction to computational geometry. Cambridge, tiass., HIT 479(480), 104 (1969)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.503, + 0.19 + ], + "angle": 0, + "content": "63. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.192, + 0.787, + 0.233 + ], + "angle": 0, + "content": "64. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.288 + ], + "angle": 0, + "content": "65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.345 + ], + "angle": 0, + "content": "66. Sarkar, A., Mai, H., Mahapatra, A., Lazebnik, S., Bhattachad, A.: Shadows don't lie and lines can't bend! generative models don't know projective geometry... for now. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28140-28149 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.787, + 0.387 + ], + "angle": 0, + "content": "67. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.43 + ], + "angle": 0, + "content": "68. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.459 + ], + "angle": 0, + "content": "69. Shtedritski, A., Rupprecht, C., Vedaldi, A.: What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.486 + ], + "angle": 0, + "content": "70. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.515 + ], + "angle": 0, + "content": "71. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.543 + ], + "angle": 0, + "content": "72. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.787, + 0.586 + ], + "angle": 0, + "content": "73. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.787, + 0.614 + ], + "angle": 0, + "content": "74. Team, I.: InternlM: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.787, + 0.643 + ], + "angle": 0, + "content": "75. Team, M.N.: Introducing mpt-7b: A new standard for open-source, commercially usable lms (2023), www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-05" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.645, + 0.787, + 0.671 + ], + "angle": 0, + "content": "76. Torralba, A., Oliva, A.: Depth estimation from image structure. IEEE Transactions on pattern analysis and machine intelligence 24(9), 1226-1238 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.673, + 0.787, + 0.713 + ], + "angle": 0, + "content": "77. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.787, + 0.755 + ], + "angle": 0, + "content": "78. Wang, F., Fu, X., Huang, J.Y., Li, Z., Liu, Q., Liu, X., Ma, M.D., Xu, N., Zhou, W., Zhang, K., et al.: Muirbench: A comprehensive benchmark for robust multi-image understanding. arXiv preprint arXiv:2406.09411 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.797 + ], + "angle": 0, + "content": "79. Wang, J.Y., Adelson, E.H.: Layered representation for motion analysis. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. pp. 361-366. IEEE (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "80. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., Xu, J., Xu, B., Li, J., Dong, Y., Ding, M., Tang, J.: Cogvlm: Visual expert for pretrained language models (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "Fu and Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "81. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.218 + ], + "angle": 0, + "content": "82. Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., Li, H.: Dire for diffusion-generated image detection. arXiv preprint arXiv:2303.09295 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.788, + 0.259 + ], + "angle": 0, + "content": "83. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., Zhou, D.: Chain-of-thought prompting elicits reasoning in large language models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "84. Yan, A., Yang, Z., Wu, J., Zhu, W., Yang, J., Li, L., Lin, K., Wang, J., McAuley, J., Gao, J., et al.: List items one by one: A new data source and learning paradigm for multimodal llms. arXiv preprint arXiv:2404.16375 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "85. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.788, + 0.357 + ], + "angle": 0, + "content": "86. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.788, + 0.398 + ], + "angle": 0, + "content": "87. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3081-3089 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.788, + 0.438 + ], + "angle": 0, + "content": "88. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.438, + 0.788, + 0.481 + ], + "angle": 0, + "content": "89. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.788, + 0.55 + ], + "angle": 0, + "content": "90. Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W., Sun, Y., Wei, C., Yu, B., Yuan, R., Sun, R., Yin, M., Zheng, B., Yang, Z., Liu, Y., Huang, W., Sun, H., Su, Y., Chen, W.: Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "91. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems 35, 27469-27483 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.788, + 0.632 + ], + "angle": 0, + "content": "92. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.632, + 0.788, + 0.675 + ], + "angle": 0, + "content": "93. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_origin.pdf b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..367f8a5b6ef53dba9cdc2a64137afe8955046e12 --- /dev/null +++ b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/30b449a7-1670-4dcd-9e2d-3bee255a2246_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:308a1a92b952612f6138667de1556732aa1342d343b2179cc77c8be00d2a8119 +size 9333507 diff --git a/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/full.md b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d23c8ee64d2ccab36865c6c2d3bfa6f0a11b3e8 --- /dev/null +++ b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/full.md @@ -0,0 +1,470 @@ +# BLINK: Multimodal Large Language Models Can See but Not Perceive + +Xingyu Fu $^{1*}$ , Yushi Hu $^{2,3*}$ , Bangzheng Li $^{4}$ , Yu Feng $^{1}$ , Haoyu Wang $^{1}$ , Xudong Lin $^{5}$ , Dan Roth $^{1}$ , Noah A. Smith $^{2,3}$ , Wei-Chiu Ma $^{3\dagger}$ , and Ranjay Krishna $^{2,3\dagger}$ + +1 University of Pennsylvania 2 University of Washington 3 Allen Institute for AI 4 University of California, Davis 5 Columbia University https://zeyofu.github.io/blink/ + +![](images/0fc6fa44ce957b1569e768062818938bf03ae186fbe98712022e11cf3b04054a.jpg) +Figure 1: The BLINK Benchmark. BLINK contains 14 visual perception tasks that can be solved by humans "within a blink", but pose significant challenges for current multimodal LLMs. These tasks are inspired by classical computer vision problems and recast into multiple-choice questions for multimodal LLMs to answer. Notice that the visual prompts and questions in this figure are different from the actual ones used in the benchmark for illustrative purposes, and answers of the samples are provided in fn 6 + +Abstract. We introduce BLINK, a new benchmark for multimodal language models (LLMs) that focuses on core visual perception abilities not found in other evaluations. Most of the BLINK tasks can be solved by humans "within a blink" (e.g., relative depth estimation, visual correspondence, forensics detection, and multi-view reasoning). However, we find these perception-demanding tasks cast significant challenges for current multimodal LLMs because they resist mediation through natural language. BLINK reforms 14 classic computer vision tasks into 3,807 multiple-choice questions, paired with single or multiple images and visual prompting. While humans get $95.70\%$ accuracy on average, BLINK is surprisingly challenging for existing multimodal LLMs: even the + +best-performing GPT-4V and Gemini achieve accuracies of $51.26\%$ and $45.72\%$ , only $13.17\%$ and $7.63\%$ higher than random guessing, indicating that such perception abilities have not "emerged" yet in recent multimodal LLMs. Our analysis also highlights that specialist CV models could solve these problems much better, suggesting potential pathways for future improvements. We believe BLINK will stimulate the community to help multimodal LLMs catch up with human-level visual perception. + +Keywords: Multi-modal Large Language Models $\cdot$ Vision-Language Benchmark $\cdot$ Visual Perception Evaluation + +# 1 Introduction + +Compared to today, computer vision was originally attempting to interpret images as projections of 3D scenes, not just processing 2D arrays of flat "patterns" [25, 59, 62]. In this pursuit, early research developed a series of intermediate tasks: they focused on understanding optical properties like reflectance [12, 79], 3D primitives through multi-view reasoning [38, 60], geometric reasoning through depth estimation [76], instance recognition through visual correspondence [56], affordance through keypoint grounding [37], and forensics through intrinsic images [9]. Yet in the modern era of large language models (LLMs), we, as a community, have focused less on such perceptual tasks, and instead have developed new tasks, mostly expressed in natural language, emphasizing the vision-language connection learned by multimodal LLMs [3, 6, 18, 20, 24, 26, 51, 52, 57, 63, 66, 73, 80]. This might be because many traditional computer vision tasks resist mediation through natural language, due to the inherent imprecision of language (e.g., it is challenging to precisely pinpoint a spatial keypoint through language). + +This paper aims to highlight crucial aspects of visual perception that have been overlooked when evaluating multimodal LLMs. To appropriately position our paper, let us revisit how we currently evaluate perception through using multimodal LLMs [44,45,48,53,54,58,90]. While many of these benchmarks have been popularized as the de facto evaluation measures for influential models like GPT-4V and Gemini-Pro, they conflate perception with language knowledge and reasoning. At the risk of singling out one benchmark, let us consider two questions highlighted in the popular MMBench [53]: “ Why is this hummingbird called ruby-throated?” and “ What will happen next? A: the person is gonna laugh B: the person is gonna cry.” For the first question, the vision subpart is to recognize the hummingbird. For the second, it only needs a coarse description of the image. Everything else is left to the language model to solve. Such a conflation has also been reported for other benchmarks by previous work [11,39,87]. Our experiments show that this conflation reductively evaluates perception as a dense captioning task. In other words, by replacing the image with a task-agnostic dense caption, our experiments show that a “blind” GPT-4 performs well on these “multimodal tasks”. + +In response, we propose BLINK. BLINK reimagines traditional computer vision problems through a format that allows us to evaluate multimodal LLMs. + +![](images/8e11d99406a2568a4fec377d15b45b5eae34e4f22ea46b266503916f332bbefc.jpg) +Figure 2: Comparison between BLINK and previous benchmarks. BLINK has several novel features: (1) BLINK incorporates diverse visual prompts, like circles, boxes, and image masks, while previous benchmarks only have text questions and answers. (2) BLINK evaluates a more comprehensive range of visual perception abilities, like multi-view reasoning, depth estimation, and reflectance estimation. Prior benchmarks are generally more focused on recognition-based VQA. (3) BLINK contains "visual" commonsense problems that humans can answer within seconds, while prior benchmarks like [90] require domain knowledge. The samples of previous benchmarks are from [45, 53, 90]. Part of our samples are curated from [10, 19, 29, 32, 36, 43, 91]. + +As partially demonstrated in Figure 15, BLINK consists of 14 classic computer vision tasks, ranging from low-level pattern matching (e.g., visual correspondences estimation) to mid-level reasoning (e.g., relative depth estimation), and extending to high-level visual understanding (e.g., visual similarity). The image tasks are meticulously selected such that they are difficult to solve by reducing the evaluation using dense captioning; instead, the models must perceive the contents of the image(s) to answer. We recast each traditional task into a modern question-answering format, where answer choices are either images or text. BLINK contains 3.8K questions across 7.3K images, where questions may contain multiple images that are curated from a wide range of datasets [8, 10, 19, 29, 32, 36, 42, 47], encompassing indoor household scenes as well as outdoor urban or natural environments. The questions and choices are either derived from the datasets, or + +manually written by humans. On average, each question can be solved by a human subject within a BLINK of an eye, except the IQ test. + +We carefully evaluate 17 multimodal LLMs with various sizes (i.e., 7B, 13B, 34B) on BLINK. We observe the paradox that while these problems are easy for humans (95.70% average accuracy), they are extremely hard for existing machinery - even GPT-4V model can only achieve 51.26% accuracy on average, which is 44.44% worse than humans, and 13.17% better than random guessing. We also experiment with specialist vision models and find that they perform much better than multimodal LLMs. For example, the specialist outperforms GPT-4V by 62.8% on visual correspondence estimation, 38.7% on relative depth estimation, and 34.6% on multi-view reasoning, in terms of absolute accuracy. Our findings indicate that the perceptual abilities of multimodal LLMs have been previously overestimated. Furthermore, these models may benefit from integrating insights from specialized models that excel in these areas. We believe BLINK can serve as an effective testbed for bridging the gap between traditional notions of perception and the modern generative capabilities of multimodal LLMs. + +# 2 Related Work + +Multimodal models: Inspired by the impressive success in recent large language models (LLMs) [13,21,63,77,93], a sequence of studies explore multimodal LMMs that can jointly understand vision and language information and generate textual answers through adding a modality adaption structure between a frozen visual encoder [27,65,71] and a frozen LLM [77,93]. Flamingo [3] and BLIP-2 [46] are two of the earliest works to explore these transformer-based multimodality conjunction structures. They first pre-train on image-text matching datasets [15,42,47,67] and then fine-tune on task-specific datasets such as visual question answer (VQA) [4,34]. Starting from LLaVA [49,51], people use LLM synthesized instruction-following chat data (which are in VQA format) for instruction tuning and achieve much better results [7,16,24,73]. There have been extended studies that explore further capabilities of multimodal LLMs, especially on VQA reasoning [30,33,39,40,68,78,84,92]. However, they mainly focus on the textual reasoning abilities within the multimodal LLMs and do not emphasize visual perceptions. + +Multimodal benchmarks: Traditional vision-language datasets are designed to assess single-task capabilities, such as optical character recognition (OCR) [55], image captioning [47], and visual question answering [4,34]. However, these datasets are often not comprehensive enough to holistically assess multimodal LMMs on general perception and reasoning abilities. Many recent papers have built more comprehensive benchmarks. MME [28] is one of the earliest holistic benchmarks containing multi-modal Yes/No questions on the defined visual perception and language reasoning tasks. MM-Vet [89] includes six sub-features from the previous datasets including recognition-focused questions, OCR, and math, providing a diverse while discrete evaluation set. MMBench [53] covers more subjects and provides a more robust circular evaluation setting. Seed + +Bench [44,45] benchmark has a more diverse source of inputs, including multiple-image inputs and video, and includes more tasks. However, the visual perception questions in MME, MMBench, MM-Vet, and Seed-Bench are mainly extracted from existing VQA datasets or generated by GPT [63] from image descriptions such as COCO-Caption [47], and are recognition focused, covering topics such as object (attribute)recognition, and OCR. In contrast, we focus on multiple distinct nuanced perception abilities and recognition-level perception is only one of our focus. Some other multimodal benchmarks have distinct focuses. MMMU [90] aims at achieving expert-level artificial general intelligence by collecting domain-knowledge-required questions. HallusionBench [35] mainly tests the language hallucination and visual illusion phenomena. MathVista [58] presents exclusively mathematical domain visual questions based on images such as charts, tables, and diagrams. These benchmarks do not require human-level perception abilities as in BLINK and therefore cannot measure model visual perceptions holistically. + +# 3 The BLINK Benchmark + +Our goal is to faithfully evaluate the visual perception capabilities of existing Multimodal LLMs. We seek to study the visual perception gap between humans and machineries, and offer deeper insights into potential pathways towards achieving more generalized machine perception. Based on the observation that existing benchmarks predominantly focus on evaluating visual recognition abilities, we introduce a novel benchmark, BLINK, designed to enable both quantitative and qualitative evaluation of the nuanced perception capabilities of multimodal LLM across various dimensions. We unfold this section by illustrating the overall design of BLINK (§3.1) and discussing its unique features comparing with previous benchmarks. Then we describe each task in detail, providing an in-depth explanation of the data curation process (§3.2). + +# 3.1 Overview of BLINK + +To ensure that one can effectively measure what Multimodal LLMs can or cannot perceive, we carefully select 14 tasks (see §3.2 for the full list) that are difficult to solve by reducing the evaluation into text-only questions using dense captioning. The tasks are drawn from either classic computer vision problems or recent applications of Multimodal LLMs, each of which requires a nuanced understanding of the visual data. They range from low-level pattern matching (e.g., visual correspondence) to mid-level spatial reasoning (e.g., relative depth), and up to high-level visual understanding (e.g., visual similarity). This variety allows for a systematic exploration of Multimodal LLMs' capabilities across different perceptual complexity layers. Furthermore, these visual tasks vary in granularity, ranging from pixels (e.g., relative reflectance) to patches (e.g., jigsaw) and extending to the full image (e.g., forensic detection), enabling us to evaluate models' proficiency in observing at various scales. + +![](images/ef107ef8238b4231e2ce547624ec425dbe05a0809cf83f9b2cc0efff16dc2a14.jpg) +Figure 3: Statistics of BLINK. 14 tasks range from pixel-level to image-level; from low-level pattern matching (e.g., visual correspondences) to mid-level reasoning (e.g., relative depth), further to high-level visual understanding (e.g., visual similarity). + +![](images/c8e5fdbf538a17f6d2bd882f0e8a08ae04b366d052f4a47df03065d39ebd23db.jpg) +Figure 4: Accuracies of multimodal LLMs on BLINK test set. Please refer to Table [1] and §4.2 for more results and discussions. + +To facilitate the evaluation of multimodal LLMs, we recast all tasks as multiple-choice question-answering problems. The options for answers may include images or texts, while the questions themselves can feature either single or multiple images. Prompts are designed to be both textual and visual in nature. We re-purposed several existing vision datasets as well as collected new data. In total, we contribute 3.9K multiple-choice questions and 7.3K images, with an even distribution between the validation and test sets. Numbers of each task are reported in Figure 3, and more detailed statistics can be found in Appendix A.5. Key features of BLINK: Comparing with previous benchmarks, BLINK has the following novel features: + +- Visual prompting: Unlike existing benchmarks that support only text prompting, BLINK features a variety of visual prompts. This enables one to highlight specific areas within images, facilitating the evaluation of Multi-modal LLMs' detailed understanding of these regions. It also offers an interface for researchers to investigate the impact of visual prompting techniques. +- Perception beyond recognition: Besides visual recognition, BLINK considers a diverse set of visual perception abilities, such as 3D reasoning, geometric understanding, affordance reasoning, etc. The breadth allows one to evaluate Multimodal LLMs from an unique array of perspectives. +- "Visual commonsense" that does not require domain knowledge: The questions in BLINK are intentionally designed to be straightforward, requiring neither domain-specific knowledge nor expertise to answer. They are crafted in such a way that humans can solve them almost instantaneously, + +typically within a few seconds. This allows us to explore the fundamental gap in visual perception gap between humans and Multimodal LLMs, highlighting the paradox that problems easily solved by humans often pose significant challenges for machines. + +- Interleaved image-text formats: BLINK features a heterogeneous question-answering format, wherein both questions and choices can be presented as text or images. This diversity compels Multimodal LLMs to genuinely understand the questions, pushing the boundaries of their interpretative capabilities. +- Diverse image sources: BLINK comprises a wide range of in-the-wild images sourced from various origins, covering everything from indoor and outdoor scenes to object-centric views and landscapes. This collection spans abstract diagrams, synthesized images, and authentic photographs, ensuring a comprehensive examination of visual perception + +The design principles of BLINK are also illustrated in Figure 2. We will now describe each task in detail. + +# 3.2 Dataset Collection Process + +BLINK comprises 14 tasks, all of which have been repurposed into a multiple-choice question-answering format. These tasks utilize a diverse collection of images from various sources and we ensure that each image is unique. + +Visual correspondence: This task aims to evaluate the ability of Multimodal LLMs to understand and identify the same scene point across various viewpoints, lighting conditions, or time. We exploit HPatches [8] for this task. HPatches contains a number of image sequences, each of which are composed of images taken under different illuminations and/or viewpoints of a scene. For each question, we randomly sample two images and an interest point within them. Then we exploit the ground-truth homography to compute its correspondence. Finally, we randomly select three more interest points to serve as other choices. + +Relative reflectance: This task aims to compare the reflectance (albedo) of two pixels. It allows us to evaluate Multimodal LLMs' understanding of material properties and their interaction with light, which is crucial for applications requiring high-fidelity visual interpretations. We curate our samples using human annotations from the Intrinsic Images in the Wild (IIW) dataset [10]. Each question is based on an image and two specified points. The goal is to identify which point is darker, or whether the two points have similar reflectance. + +Relative depth: Humans are good at judging relative depth [19]. This task can thus serve as a proxy to validate if the geometric understanding capabilities of existing multimodal LLMs are close to human. We curate our samples using human annotations from the Depth in the Wild [19] dataset. Given an image and two specified points, the task is to determine which point is closer. + +Spatial relation: Understanding spatial relationships between objects in a scene is essential for interpreting complex visual environments. However, modern Multimodal LLMs often struggles with spatial concepts such as "left" and "right" [88]. This task helps us evaluate whether the models finally possess this + +Visual correspondence +![](images/e2b6ee94633975a0e058fc3150fdd291369c324fa70bf054a07440df8075483c.jpg) +A: Which point corresponds to the reference point 1? + +![](images/5689ea0115dd42d13288f048d8b9b2dfee37e3cd72546038249cfa6c6305b164.jpg) +(a) A + +![](images/42351252cbe567794610834a137b92fcb8209d07233d8dccdcfda37ed4e3640f.jpg) + +![](images/fe48018c0a23e379c08d9faff86652e87a3b6a1b1225022555518de4ce257a06.jpg) +(b) B + +![](images/7ea7490b1452a576c92d4774fd6e1272b5e2a6fc52b6b6507b7787d0d53a6e9a.jpg) + +![](images/80889ee5fd0d100618eaef38dd2324c727c0dcbc55e11014bec1d6bb57b7a45a.jpg) + +![](images/3478181c78b7cf1ecb16e18d01233255f363f91303a3a9f62f402839382a1c30.jpg) + +Relative depth +![](images/6e4d21331d2bae3ba8a9deb5b1548a1edb38db2187ec4cbd25f84931b6df2adf.jpg) +Q: Which point is closer? + +![](images/651d8987645fbe4fd4f03467e483671b0d06f9e388ca4e8c3315bb7883b40f80.jpg) + +![](images/9f30e061ade7dc0195bf14bcbb63fc65828a43a30d7fe27f7d812f4b8f399bd2.jpg) + +![](images/e76796139a94c9a6aa99d5eff68d41bdcf3204680d0473a4472ef424b4e8d612.jpg) + +![](images/1e1a45817760d86c08b5914f146029c5bab69308ef7c4322cdb3f5bd4f42ede4.jpg) + +![](images/d6d450547c40a69d3bad0349117f19cf4bb4144c6f4edbabd3475573bc115c16.jpg) + +![](images/070b3cea6cc4151c43caa320a7e8ea0556ce8438961535030fb57ae8ac6aa891.jpg) + +Multi-view reasoning +![](images/8b5dd7ad0226225c452ac6c3d90ec69d3bfca4ed92b749f113530608c58960d8.jpg) +Q. The first image is from the beginning of the video and the second image is from the end. Is the camera moving towards left or right when shooting the video? + +![](images/c07bb3cb678fcdbd3f0427849f29d0b99357966bd06659d47016af5ab92caaf2.jpg) + +![](images/134c023a930b90a42c0aa5f4b9e710a809fc725c52f7c18b02e8f0acd3e124bc.jpg) + +![](images/5aca460fd98be45c3282bd208fdfb6c882271a97cf133fc43d9514d890f67037.jpg) + +![](images/a4619dea0e7f5eff96a0cfaaa33f2bfa79f888bcf39f49aa0229c6ee76bde353.jpg) + +![](images/6712115a8d9c2df0620071d2e014b4f7288c427aa46643ac03fc4de0b949ba45.jpg) + +![](images/ab505912deabe487dc23c629b6b127b48a41d38e476edc200f091dfceb5f5b8f.jpg) +Art style + +![](images/feb01dd766f52d22f28bdec52168544036307b8907d6dad44b55ecafd0993de2.jpg) +Q: Which image fits the missing part in the first image? + +![](images/154e14039f83a667fe3bff9ade0e7844eed9da8138a41633888cf4c21578a2d6.jpg) + +![](images/6e4e3d0b464a53fd1734bfe604369ac137643e80e69c1b45acda61f865379cd6.jpg) + +![](images/a2c958e71850115845c9835a4f4d1e8d0c716fc17a7068573aa3db64eb4267d3.jpg) + +![](images/9de8f24018cadd0c26138372fcb3e43dbead6766466b1f956eae7dfa62cbd0d5.jpg) +Counting + +![](images/5b3f5d32d8c88b9fe2cf844fca2748e74ef13af4cdaa0139ac5dd2842f6eeb15.jpg) +Q: How many fingers are in front of the bathtub? + +![](images/016972865dae4d5cac90fcd7fc1c56814ffa00408a963f38caf78d4735eed579.jpg) + +![](images/e9c77077853e56a314a0b106f94c9de75869a2935d8f6b141d39a19a6761f69e.jpg) + +![](images/61901e0975f80a2ce43de3c3d06f9fdc4520ce8dc33dae75197da8c52f6c0c6d.jpg) + +![](images/fe66627b4d8ec6d0b645f6b539e64c83dd2ecc2057cc9bbd07fd0a98ddb7b58a.jpg) + +![](images/985c6dd4a3161130177d1cb56a73df4a568eeec89a64b8861bda530bc9bbe0cd.jpg) + +IQ test +![](images/2a025d187995d994c23a13f4b25f5ebce5d55927b498ade3796f87e1dbdddac8.jpg) +Q: Which image comes at the end? + +![](images/cddbf4b51915617b679eb5a0f0a9f57710bececf4db69b62d1e267ce7842d654.jpg) + +![](images/d47c3c59997047da3f010529362000f9b62977a788ceeaee514cd0c4eb22f190.jpg) + +![](images/c12a496d68093d9ff5b9c06b21a55d99662c09003dcdacd8dd57fc863bfce5aa.jpg) + +![](images/bbc26ee5e64d405602bb624c425e2117ad46e75b77aed97d4e36857121215c3b.jpg) + +1 Semantic +![](images/77c6b5c4b66ae600a2954ca2707715063f2a81aaa78cf46e014618a2c056c41e.jpg) +Q: Which point is semantically similar to the reference point 1? + +![](images/766ac0850e000c625cad38a6a112b6ddb0193aa3d8a493bd0a6adb1e58ec9673.jpg) + +![](images/b159e2ae552fe6069d25ca9125b310ba3e584c3d3d68ba062a01bec1e50ca515.jpg) + +![](images/0127801748bc39d4a68bdeaa24b9171a0512c8c55b7a533f547d9da87fe06f10.jpg) + +![](images/b058eb3109671b70f69ca0cc10b29832c017f26d616aaf21132916232097d408.jpg) + +![](images/977586505e10e895db3137716321027709a98775c83bea83af6a3cd51c1fd8f3.jpg) + +![](images/5258312ff30e6325337d3434304f118059e78e2275466fcca4615ac864c652aa.jpg) + +![](images/2450015f28bc8359038ab5baf011ff1918fb929b8b830774f95f499c75d2f956.jpg) + +![](images/4479acecd70be14838dac929070a3013b4eced580ddc069b3e2167bf06e46ea5.jpg) + +![](images/3fabb53e70b0df933cfcf2ff058aded627f18874381583252c0218147899ea8b.jpg) + +![](images/3e2877f232288738e9d83959b67ee06a6ee1415b0637bdfa1a58075086cec9b6.jpg) + +Relative reflectance +![](images/0b9ed0e4b8e4697279d0dc0192a5e54f4b7024770bb8ddf9dd79de1465aa97d2.jpg) +Q: Consider the surface color (color without shading) of the two points in the image. Which one is darker, or the color is about the same? + +![](images/61094acecc0cfd4a3d870f961d1a833799c2c9ee349c9fa8fb80ffecc1df1768.jpg) + +![](images/d07bf20a09e40b882ae823b58833fd10757047409b80aa37ab46da4ad70ffb7f.jpg) + +![](images/37f550064cadd3bc93ca905ee7f88bc63fe3e59ac06e768b27826ecbdeae7f0e.jpg) + +![](images/ba3421d9168789866e2a712b775e62f18274de1b216c7fc8acffc845ca438723.jpg) + +![](images/53a6af5ff11348ee5f46e3fce202019be61cfb32b164026ced9f061b9af3f65e.jpg) +(b) B is darker +(b) B is darker + +![](images/1afaa30997f6e8156423f465444dde3c653d50e3524a6df6f4daa7f0ebde1d07.jpg) + +![](images/d70b02c9f50fcc837aa4a8af57130ce5f3e41edf09f97b2e5b61584b85e38ad5.jpg) +Spatial reasoning + +![](images/1ae464d1233777e17c372c003736fb132d33cd218ca889dfe7b701bad78f747f.jpg) +Q: Is the bed at the right side of the dining table? + +![](images/d9ef6393e6e18068156a0ac93d7488e5612c66ac0055792d3aa09df9b1ddb629.jpg) + +![](images/fec253e3c0343e070193a68c69e4cbf2d407e8066b8b8b25525ab29732f205aa.jpg) + +![](images/aa86a63d228c6745c047284f68d6e85f4d3213c8afb5aaac2397caac7d845733.jpg) + +![](images/dde2ecb0cfb6bffb2124c796d412b0f9ee6bef3260a2c8390e3d27d10d4f58d2.jpg) + +Jigsaw +![](images/534f20444fc29b577843427ed47bff55b404994ed4050e585dbe7b96549931c4.jpg) +Q: Which image fits the missing part in the first image? + +![](images/6e36f4e07db8660f686e01ac09f3299c4812718d7106ccf2ac2d87230118f7d2.jpg) + +![](images/fce774468fa3f1902ecaa222e8647770907bf29306f418dba5d08bb278458d42.jpg) + +![](images/b76e2ef980b6667ae40db8d8d66970f42d07944af034495d92f17fcc4251cd93.jpg) + +Object localization +Forensics detection +![](images/cec570b58da17bf2015e26eed1f874264fa4083f6c980e1f1a82f7a489ee45fd.jpg) +Q: Which bounding box more accurately encloses the bun? + +![](images/34fb30e756a205b5554d37df4d0d7feb712b48471f673b7080410d072e0395ec.jpg) + +![](images/98eeba09cb4216392f2875bf2691bf06c6add1f2b0634634aa06160206c84710.jpg) + +![](images/999c312274771ddff17a6a6482752b6694d6a8a2b161d6e3fed808a77eabe7d1.jpg) + +![](images/2b4c98f281143e0e58fe60448be2482885d4ffe636add08314aa1c9d262dd0c2.jpg) +Q: Which image is most likely to be a real photograph? + +![](images/393bef96cbf2b2051f12af281ca418e5a2b802d8ba319d1770338acf32cfe6fe.jpg) + +![](images/23dd786dbd85f9401a43a47adae456a8a1e6cda94a89954b14b006d7143cde32.jpg) + +![](images/18b7e8e079a07bfb9190fa99c9aa5eb35eb09742fa0b4ee4f5ddbfa78c224308.jpg) + +Visual similarity +![](images/da075ed77930d542e54affdbc601981218f67d91191fb6db799dde6aa0a37c07.jpg) +Q: Which image is most similar to the reference image? + +![](images/94dc6ee69e3b8b2824b26ccad8a65d799f519415bf62c14f5cdbf649b2f71ed4.jpg) + +![](images/362cdf0b32cc3029051dd9242451a363e060c252e4db786d2952879d6f41a03e.jpg) + +![](images/4f00cff1e151a939a03b146f4ff23b6e8a694109dae9ca33b50ba59a33aa55dd.jpg) + +![](images/02055b2953abfab149c98e35ba75da7478fa9cae5729ea08de56951123680174.jpg) + +Functional correspondence +![](images/40e6e0ee580a2ec686de7bf6f98978ae662f3af20794de0da89b93f937560ff1.jpg) +Q Which point is functionally different from reference point 1 during mixing? + +![](images/62cf777146c3edb5dbeeb788cf96cda371fcbf5fa010ecad725546adc555109f.jpg) + +![](images/4467d0ee1e86e3cc91aa13bcc4ee39ab78b66913806341aea73dfa072d6c3454.jpg) + +![](images/dbba412a63df6715cc5cddb77f69099e640442411b1b8afc32ee77ad4f1a1b3e.jpg) + +![](images/9e4b3eea5e9df564b65ef4e09cbdd1a0cb75aff3623dbb660d06a30c39079003.jpg) + +![](images/2e1576154b3887c795ba3271aff1dc324f3050072bb87dfa87d50f9216715db1.jpg) + +![](images/083be8e78cb1fbec184a77b304474a13e9151a088f6ec3266305c6e2fc450524.jpg) + +![](images/07dffa79710d93202c148ad77796d3ac999e063be525ca284d5a86ebff76811e.jpg) + +![](images/5eaae702c43824d0cfc94c3fb0d0264c19e004968e40cf4660eba92685aec38a.jpg) +Figure 5: Qualitative results on BLINK. For each task, we show the choice of LLaVAv1.6-34B [51], Qwen-VL-Max [7], Gemini Pro [73], GPT-4V [63], and humans. Red choice indicates the ground truth. Notice that the markers are intentionally enlarged for visualization purposes, and we make some images inset images to save space. For IQ test, the third image is constructed by overlaying the first and second images. + +![](images/34535316fb4354c16131856371a6f902772a22d071b46d16c79882790df0789e.jpg) + +![](images/811ebca9e7c09a042df12987a66b1f0f8a8804714ce607c17c10323becd89d9b.jpg) + +vital skill. We curate our samples from the Visual Spatial Reasoning [48] dataset. Each sample contains an image and a claim. The task is to determine if the claim is true or false. We reformat the claims into binary questions via GPT-3.5 [13]. Multi-view reasoning: This task is centered on evaluating the multi-view reasoning capabilities of Multimodal LLMs. The objective is to deduce the + +relative camera motion based on two images of an object captured from different viewpoints. Our data is sourced from the Wild6D dataset [91], which features videos of various objects recorded in diverse settings. We select two random frames from each video to calculate the relative camera motion. Recognizing that even humans might struggle to precisely articulate 3D motion details, we simplify the task by classifying motions into two broad categories: moving towards the left or moving towards the right. Despite the simplicity of these questions, as we will later demonstrate, they pose significant challenges for current models. + +Jigsaw: This task assesses the ability of Multimodal LLMs to recognize and group patterns, aligning patches based on continuity in shape, color, and texture. We utilize images from the TARA dataset [32] and segment each of them into a 3x3 grid. We retain the three segments from the upper left corner as the reference image, and treat the central segment along with a randomly chosen segment as options. The objective is to identify the correct patch (i.e., the central patch). + +Art style: This task evaluates Multimodal LLMs capability to analyze and discern both local and global similarities in art styles among multiple images. Although there have been prior efforts to incorporate art-related questions into evaluation [90], such attempts primarily focused on questions requiring expert-level knowledge, including deducing an artist's name and understanding historical contexts, rather than on direct image comparison. For this task, we collect paintings and their stylistic information from WikiArt. Given one reference painting image and two other paintings as options, the model is tasked with identifying the one that most closely shares the art style of the reference painting. + +Object localization: The ability to accurately detect and localize objects is critical for scene understanding. While previous benchmarks [53] have explored this task, their focus was primarily on coarse localization. For instance, they might only ask the model if an object is located at the "top" or "right" side of an image. BLINK, in contrast, aims for a more fine-grained evaluation. We exploit images from LVIS [36], randomly sampling one object per image along with its ground-truth bounding box. Then we add Gaussian noise to the ground-truth box to create a confounding box. The goal is to select the correct one. + +Counting: This task evaluate Multimodal LLMs' abilities in detection, recognition, and compositional reasoning, particularly in complex scenes where objects may overlap, be occluded, or vary in size and appearance. We select our questions from the TallyQA dataset [2], known for its challenging human-written counting questions. Each sample comprises an image, a question, and a numerical answer. In addition, we randomly select three numbers to serve as confounding options. + +Forensic detection: Recent advances in generative AI have raised concerns about malicious uses and have prompted calls for the automatic detection of fake content. To evaluate whether Multimodal LLMs can fulfill such a role, we construct sets of real and synthesized images that describe similar scenes and ask the models to identify the real ones. Specifically, we first generate synthetic images using Stable Diffusion XL [64], employing COCO captions [47] as prompts. Then, we manually search online using these captions as descriptions and select high-quality photographs as the real images. + +IQ test: This task evaluates the ability of Multimodal LLMs to engage in graphical reasoning, without requiring any domain-specific knowledge. We manually collect test samples, along with human explanations, from various public, license-friendly online sources. Given visual examples and a selection of images, the objective is to identify the image that either continues the pattern established by the examples or is spatially consistent with them. + +Visual similarity: This task aims to verify whether Multimodal LLMs possess a nuanced understanding of visual features, patterns, and aesthetics at a level comparable to humans. We select our samples from the DreamSim dataset [29]. Given a reference image alongside two alternative images, the objective is to identify the image that more closely resembles the reference image perceptually. + +Semantic correspondence: This task focuses on identifying and matching semantically similar yet visually distinct elements across images, thereby evaluating the ability of Multimodal LLMs to understand the underlying semantics of object parts. Our samples are sourced from the SPair-71k dataset [61], which features pairs of images with multiple corresponding semantic points. For each task, we randomly select one semantic point in an image as a reference, and provide the matching point alongside three random semantic points in the paired image as options. The objective is to accurately identify the correct matches. + +Functional correspondence: The task aims to identify points that are functionally similar across objects. It challenges Multimodal LLMs to extend their understanding beyond mere semantics, enabling them to infer the diverse functions an object can perform in various contexts. Such capability is crucial for applications in robotics. We derive our samples from the FunKPoint dataset [43], which features paired images annotated for functional correspondences. Following a method analogous to semantic correspondence, we present an action alongside two object images. One image includes a reference point, while the other offers four potential points. The objective is to select the point that best matches the reference in terms of functional affordances. + +Data quality control: To guarantee the quality of BLINK, we manually go through all collected data and filter out data that are ambiguous. + +# 4 Experiments + +In this section, we first describe the experimental setup and the baselines (\$4.1). Then we present a comprehensive evaluation of 16 recent multimodal LLMs (\$4.2). We demonstrate that while humans can answer the questions with high accuracy, BLINK is challenging for existing models. Finally, we provide detailed analyses on multiple experimental settings, including the effect of reducing images to captions, sensitivity to different visual prompts, and error analysis (\$4.3). + +# 4.1 Experimental Setup + +Multimodal LLMs: We evaluate BLINK on 16 recent models: MiniGPT-4-v2 [16], OpenFlamingo-v2 [5], InstructBLIP (7B and 13B) [24], CogVLM [80], + +LLaVA(v1, v1.5, v1.6, internLM, and xtuner versions, model size 7B, 13B, and 34B) [23,26,50-52], Yi-VL (6B and 34B), Qwen-VL-MAX [7], Gemini Pro [73], Claude 3 Opus [1] and GPT-4V(vision) [63]. See Appendix B for more details. + +Evaluation setup: We follow standard setups as in the VLMEvalKit [22], where the temperature is set to 0 and retry is set to 10. However, we do not resize the images during any experiment. For the models that do not support multiple images as input, we concatenate the images as input. We extract the choice from the models' output with a set of pre-defined rules and GPT-3.5-turbo [13]. We refer the readers to Appendix A for more details on visual prompting, how we generate the answers in BLINK, and the human evaluation protocol. + +# 4.2 Main Results + +Overall performance: As shown in Table 1, the mean accuracy of 7B and 13B open-source Multimodal LLMs hover around $35 - 42\%$ , which is similar to random guess $(38.09\%)$ . The most proficient open-source model, LLaVA-v1.6-34B, achieves an accuracy of $45.05\%$ . Even the most advanced models, GPT-4V and Gemini Pro and Claude 3 OPUS, achieve accuracies of only $51.26\%$ , $45.72\%$ , and $44.11\%$ respectively. Their performance are merely $13.17\%$ , $7.63\%$ and $6.02\%$ better than random guessing and lag behind human performance by $44.44\%$ , $49.98\%$ and $51.59\%$ . Notably, for certain tasks such as jigsaw, semantic correspondence, multi-view reasoning, object localization, and relative reflectance, some multimodal LLMs even underperform compared to random guessing. Some qualitative results are shown in Figure 5. Detailed scores on the validation set are in Appendix C. In which tasks do multimodal LLMs show relative strengths and weaknesses? Figure 4 shows the accuracies of the best-performing models on BLINK: LLaVA-v1.6-34B [51], Gemini Pro [73], and GPT-4V [63]. We observe that multimodal LLMs perform relatively better on spatial reasoning, art style, and counting tasks, in which they are much better than random guessing. The models also demonstrate some capability in relative depth and forensics detection. Overall, they are doing relatively well on mid-level perception tasks. In terms of granularity, the models in general perform better on image-level tasks and struggle on pixel-level and crop-level tasks. + +GPT-4V behaves differently: Figure 4 and Table 1 show an interesting phenomenon: GPT-4V's performance pattern is different from other models. Compared with its counterparts, GPT-4V is much better in visual similarity, art style, jigsaw, and multi-view reasoning. Specifically, its performance on visual similarity is $29\%$ better than Gemini Pro, demonstrating that GPT-4V possesses a nuanced understanding of visual patterns and aesthetics that is similar to humans. In contrast, Gemini Pro and LLaVA have similar performance patterns. Human performance: Human evaluators achieve over $95\%$ accuracy across most tasks, with an average accuracy of $95.70\%$ . This performance disparity between + +
Validation (1,901)Test (1,906)Similarity (136)Counting (120)Depth (124)Jigsaw (150)Art (117)Fun.Corr. (130)
Random Choice38.0938.09502550505025
Human95.6795.7096.7093.7599.1999.0095.3080.77
Open-source multimodal LLMs
MiniGPT-4-v2 1634.2334.5752.9410.8349.1926.0047.8618.46
OpenFlamingo-v2 539.1838.3255.1521.6754.0346.0052.1436.15
InstructBLIP-7B 2439.7238.6546.3229.1750.8154.0047.8623.85
InstructBLIP-13B 2442.2439.5846.3230.8350.0054.0050.4322.31
LLaVAInternLM2-7B 7437.7136.0652.9452.5052.4234.6730.7723.08
Yi-VL-6B 738.7241.2446.3246.6756.4550.0053.8523.85
Yi-VL-34B 741.6842.7850.0058.3353.2354.0046.1539.23
LLaVA-v1.5-7B-xtuner 2339.3640.8146.3253.3350.8154.0047.8623.85
LLaVA-v1.5-13B-xtuner 2342.0041.3146.3245.0054.0353.3347.8626.15
CogVLM 8041.5439.3846.3238.3350.8152.6749.5723.85
LLaVA-v1.5-7B 4937.1338.0146.3243.3351.6111.3347.8621.54
LLaVA-v1.5-13B 4942.6640.5546.3250.0047.5854.0047.8620.77
LLaVA-v1.6-34B 5146.8045.0546.3268.3364.5256.6747.0130.77
API-based models
Qwen-VL-Max 740.2841.9451.4755.8358.873.3337.6128.46
Gemini Pro 7345.1645.7255.8865.0050.0054.0049.5732.31
Claude 3 OPUS 144.0544.1170.5949.1757.2632.6760.6822.31
GPT-4V(ison) 6351.1451.2683.0960.8358.8762.6778.6331.54
GPT-4 Turbo 6354.6153.8983.0960.8366.9466.0081.2031.54
GPT-4o 6360.0459.0365.4451.6764.5258.0082.9139.23
Sem.Corr. (140)Spatial (143)Local. (125)Vis.Corr. (172)Multi-view (133)Reflect. (134)Forensic (132)IQ (150)
Random Choice255050255033.332525
Human96.0798.2598.0099.4292.4895.14100.0080.00
Open-source multimodal LLMs
MiniGPT-4-v2 1626.4351.7556.0023.8452.6331.3417.4219.33
OpenFlamingo-v2 523.5746.8552.0025.0041.3543.2815.9123.33
InstructBLIP-7B 2425.0055.2444.8022.6758.6529.8529.5523.33
InstructBLIP-13B 2422.8664.3452.0020.9354.1446.2713.6426.00
LLaVAInternLM2-7B 7422.1474.1348.0021.5141.3532.843.7914.67
Yi-VL-6B 726.4372.7349.6029.6548.1229.8520.4523.33
Yi-VL-34B 721.4370.6354.4023.8441.3546.2717.4222.67
LLaVA-v1.5-7B-xtuner 2324.2974.8345.6023.8442.1126.8736.3621.33
LLaVA-v1.5-13B-xtuner 2322.1477.6248.0022.0941.3546.2729.5518.67
CogVLM 8023.5767.1343.2020.9357.1426.8724.2426.67
LLaVA-v1.5-7B 4932.1470.6348.8020.3549.6236.5728.0324.00
LLaVA-v1.5-13B 4923.5767.8347.2020.3541.3545.5227.2728.00
LLaVA-v1.6-34B 5127.8676.2241.6027.3346.6229.8541.6726.00
API-based models
Qwen-VL-Max 729.2977.6249.6022.6753.3849.2547.7322.00
Gemini Pro 7322.1467.1346.4037.2141.3546.2745.4527.33
Claude 3 OPUS 120.7157.3446.4031.4057.8927.6162.1221.33
GPT-4V(ison) 6330.0072.0350.4037.2158.6538.8130.3024.67
GPT-4 Turbo 6332.8667.1348.8042.4457.1434.3351.5230.67
GPT-4o 6345.7176.9256.0071.5160.1538.8185.6130.00
+ +Table 1: Results of different models on the BLINK test set. The first row shows task names and number of test data. Average scores on validation set are also included. + +humans and multimodal LLMs highlights the significant visual perception gap that exists between current machine learning models and humans in perceiving, processing, and understanding complex visual and textual context. + +![](images/3bb53de13f40437cd78c2a10177ee3ef19aad7e7ca0e50653921245457a2c3e7.jpg) +Figure 6: Performance of using image caption + text-only GPT-4 vs. GPT-4 Vision on MMBench [53], MMMU [90], and BLINK (\$4.3). + +![](images/97243bb6d556d0e1285a0c95cd16652f285abb6f77c05cc2543cc425f2a8db59.jpg) +Figure 7: Accuracy of GPT-4V with different visual prompts (e.g., different circle sizes, colors) on relative depth, relative reflectance, and visual correspondence tasks. More discussions in §4.3. + +# 4.3 Analysis + +Is dense captioning all you need for a multimodal LLM benchmark? To answer the question, we reduce multimodal benchmarks to a text-only problem. Specifically, we convert images into task-agnostic dense image captions with GPT-4V. The dense caption describes detailed information about the image and the visual prompts (e.g., where each circle is), using language. For each multimodal question, we prompt the text-only GPT-4-0125-preview model with image captions and the textual question and evaluate if the "blind" GPT-4 can answer the question. We call this Caption + LLM. This experiment is predicated on the hypothesis that captioning involves predominantly recognition-centric perception. If using captions along with text-only LLMs yields performance comparable to or surpassing that achieved through the integration of images with multimodal LLMs, then the perception demands of that benchmark are primarily confined to recognition only. + +We experiment with BLINK, MMBench [53] and MMMU [90], as illustrated in Figure 6. Surprisingly, we find that the Caption + LLM setting achieves better results on MMBench than GPT-4V (with $5.7\%$ increase in accuracy). On MMMU, Caption + LLM achieves $47.2\%$ accuracy, which is $9.6\%$ lower than GPT-4V performance, but is still much better than random guessing. On BLINK, Caption + LLM fails, achieving random guessing performance. These results indicate that dense captions cover the visual information needed for MMBench. For MMMU, image captions carry a large portion of visual information needed to answer the domain-knowledge-specific questions. Meanwhile, the performance decrease observed in BLINK suggests the necessity for advanced perceptual abilities beyond what is currently attainable with general captions. This variance highlights the limitations of existing multimodal LLM benchmarks in addressing the full spectrum of visual perception. + +Effect of visual prompting on BLINK: Several BLINK tasks involve visual prompting. Prior work [69] shows that factors like shape, size, and color may affect task performance, and circles give the best overall performance. Following [69], we adopt circles in BLINK and analyze the effect of circle sizes and colors on multiple tasks in Figure 7. We experiment with relative depth, relative reflectance, + +
TaskVis.Corr.DepthMulti-viewSem.Corr.ForensicReflect.
Random25.0050.0050.0025.0025.0033.33
Human99.5699.5992.1094.60100.0099.63
Gemini Pro42.4440.3244.3626.6250.7645.52
GPT-4V33.7259.6855.6428.7834.0938.81
SpecialistDIFT 72DepthAnything86LoFTR 70DIFT 72ORDINAL SHADING 14
96.5197.5890.2271.2268.9477.61
+ +Table 2: Comparison between multimodal LLMs, specialists, and human performance on the BLINK dev set. The specialists perform much better than multimodal LLMs. + +and visual correspondence, with 100 validation set samples per task. The images are all reshaped to 1024px height. We experiment with circles with 5px, 10px, and 15px radius, and with red or gray color. We find that red is better than gray for all tasks. Also, the optimal circle size is task-dependent. On average 10px circles work the best, and we use it for all evaluations in this paper. The experiments suggest that visual prompting can have a big impact on multimodal LLM performance, and improving visual prompts or improving model robustness to prompt variation is a promising direction for future research [85]. + +Can specialist models solve BLINK tasks? Specialists can serve as a proxy upper bound of how good multimodal LLMs could be. We download the trained checkpoints for six specialist models and evaluate them on BLINK. As shown in Table 2, the specialists perform much better than GPT-4V and Gemini Pro, outperforming the best multimodal LLM by $18\%$ to $57\%$ on these tasks. Specifically, DepthAnything [86] and DIFT [72] achieve human-level performance on depth estimation and visual correspondence, whereas multimodal LLMs fail miserably. This sheds light on the possibility that multimodal LLMs may progress on these tasks given the correct data and training strategy. For instance, one possible way is to distill existing specialist models into multimodal LLMs [41]. + +Error analysis of GPT-4V: We randomly sampled 140 error instances made by GPT-4V on BLINK, 10 per task, and meticulously examined them. The most common types of errors are: Hallucinate fine-grained patterns and attributes (24.2%): the model hallucinates the nuanced details of objects. This error is most common for relative reflectance, forensics detection, and jigsaw tasks. Hallucinate visual prompt locations (20.0%): the circle location described by the model is wrong. This is common for visual correspondence and relative depth tasks. Other errors include Failures on capturing overall setting or style (8.6%), and Failures on grounding an object (5.7%). More details are in Appendix C.3 + +# 5 Conclusion + +We introduced BLINK, a new multimodal LLM benchmark that evaluates core visual perception abilities not found in existing evaluations. While these tasks seem trivial for humans to solve "within a blink", we find they pose significant challenges for current multimodal LLMs. Ultimately, Blink provides an effective testbed for multimodal LLMs to catch up with human-level visual perception. + +# Acknowledgements + +This work was funded in part by ONR Contract N00014-23-1-2417, and supported by NSF grant IIS-2212433. + +# References + +1. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family (March 2024) +2. Acharya, M., Kafle, K., Kanan, C.: Tallyqa: Answering complex counting questions. In: AAAI (2019) +3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +4. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015) +5. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023) +6. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023) +7. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023) +8. Balntas, V., Lenc, K., Vedaldi, A., Mikolajczyk, K.: Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In: CVPR (2017) +9. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978) +0. Bell, S., Bala, K., Snavely, N.: Intrinsic images in the wild. ACM Trans. on Graphics (SIGGRAPH) 33(4) (2014) +1. Berrios, W., Mittal, G., Thrush, T., Kiela, D., Singh, A.: Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410 (2023) +2. Black, M.J., Anandan, P.: A framework for the robust estimation of optical flow. In: 1993 (4th) International Conference on Computer Vision. pp. 231-236. IEEE (1993) +3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +4. Careaga, C., Aksoy, Y.: Intrinsic image decomposition via ordinal shading. ACM Trans. Graph. (2023) +5. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: CVPR (2021) + +16. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning (2023) +17. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023) +18. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023) +19. Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. Advances in neural information processing systems 29 (2016) +20. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Changpinyo, S., Wu, J., Ruiz, C.R., Goodman, S., Wang, X., Tay, Y., Shakeri, S., Dehghani, M., Salz, D., Lucic, M., Tschannen, M., Nagrani, A., Hu, H., Joshi, M., Pang, B., Montgomery, C., Pietrzyk, P., Ritter, M., Piergiovanni, A., Minderer, M., Pavetic, F., Waters, A., Li, G., Alabdulmohsin, I., Beyer, L., Amelot, J., Lee, K., Steiner, A.P., Li, Y., Keysers, D., Arnab, A., Xu, Y., Rong, K., Kolesnikov, A., Seyedhosseini, M., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali-x: On scaling up a multilingual vision and language model (2023) +21. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levskaya, A., Ghemawat, S., Dev, S., Michalewski, H., Garcia, X., Misra, V., Robinson, K., Fedus, L., Zhou, D., Ippolito, D., Luan, D., Lim, H., Zoph, B., Spiridonov, A., Sepassi, R., Dohan, D., Agrawal, S., Omernick, M., Dai, A.M., Pillai, T.S., Pellat, M., Lewkowycz, A., Moreira, E., Child, R., Polozov, O., Lee, K., Zhou, Z., Wang, X., Saeta, B., Diaz, M., Firat, O., Catasta, M., Wei, J., Meier-Hellstern, K., Eck, D., Dean, J., Petrov, S., Fiedel, N.: Palm: Scaling language modeling with pathways (2022) +22. Contributors, O.: Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass (2023) +23. Contributors, X.: Xtuner: A toolkit for efficiently fine-tuning llm. https://github.com/InternLM/xtuner (2023) +24. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023) +25. DO CT OR OF, P.E.: MACHINE PERCEPTION OF THREE-DIMENSIONAL, SO LIDS. Ph.D. thesis, MASSACHUSETTS INSTITUTE OF TECHNOLOGY (1961) +26. Dong, X., Zhang, P., Zang, Y., Cao, Y., Wang, B., Ouyang, L., Wei, X., Zhang, S., Duan, H., Cao, M., Zhang, W., Li, Y., Yan, H., Gao, Y., Zhang, X., Li, W., Li, J., Chen, K., He, C., Zhang, X., Qiao, Y., Lin, D., Wang, J.: Internl m-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420 (2024) +27. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023) + +28. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) +29. Fu, S., Tamir, N., Sundaram, S., Chai, L., Zhang, R., Dekel, T., Isola, P.: Dreamsim: Learning new dimensions of human visual similarity using synthetic data (2023) +30. Fu, X., He, M., Lu, Y., Wang, W.Y., Roth, D.: Commonsense-t2i challenge: Can text-to-image generation models understand commonsense? arXiv preprint arXiv:2406.07546 (2024) +31. Fu, X., Zhang, S., Kwon, G., Perera, P., Zhu, H., Zhang, Y., Li, A.H., Wang, W.Y., Wang, Z., Castelli, V., Ng, P., Roth, D., Xiang, B.: Generate then select: Open-ended visual question answering guided by world knowledge. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Findings of the Association for Computational Linguistics: ACL 2023. pp. 2333–2346. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023-findings-acl.147 https://aclanthology.org/2023-findings-acl.147 +32. Fu, X., Zhou, B., Chandratreya, I., Vondrick, C., Roth, D.: There's a time and place for reasoning beyond the image. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 1138-1149. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.81, https://aclanthology.org/2022.acl-long.81 +33. Fu, X., Zhou, B., Chen, S., Yatskar, M., Roth, D.: Interpretable by design visual question answering. arXiv preprint arXiv:2305.14882 (2023) +34. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017) +35. Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., Manocha, D., Zhou, T.: Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models (2023) +36. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) +37. Harris, C., Stephens, M., et al.: A combined corner and edge detector. In: Alvey vision conference. vol. 15, pp. 10-5244. Citeseer (1988) +38. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003) +39. Hu, Y., Hua, H., Yang, Z., Shi, W., Smith, N.A., Luo, J.: Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699 (2022) +40. Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897 (2023) +41. Hu, Y., Stretcu, O., Lu, C.T., Viswanathan, K., Hata, K., Luo, E., Krishna, R., Fuxman, A.: Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. arXiv preprint arXiv:2312.03052 (2023) +42. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) + +43. Lai, Z., Purushwalkam, S., Gupta, A.: The functional correspondence problem. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15772-15781 (2021) +44. Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092 (2023) +45. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023) +46. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +48. Liu, F., Emerson, G., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics 11, 635-651 (2023) +49. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023) +50. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023) +51. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/ +52. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024) +53. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., Chen, K., Lin, D.: Mmbench: Is your multi-modal model an all-around player? (2023) +54. Liu, Y., Li, Z., Li, H., Yu, W., Huang, M., Peng, D., Liu, M., Chen, M., Li, C., Jin, L., et al.: On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895 (2023) +55. Liu, Y., Li, Z., Yang, B., Li, C., Yin, X., lin Liu, C., Jin, L., Bai, X.: On the hidden mystery ofOCR in large multimodal models (2024) +56. Lowe, D.G.: Object recognition from local scale-invariant features. In: Proceedings of the seventh IEEE international conference on computer vision. vol. 2, pp. 1150-1157. IEEE (1999) +57. Lu, J., Clark, C., Lee, S., Zhang, Z., Khosla, S., Marten, R., Hoiem, D., Kembhavi, A.: Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172 (2023) +58. Lu, P., Bansal, H., Xia, T., Liu, J., Li, C., Hajishirzi, H., Cheng, H., Chang, K.W., Galley, M., Gao, J.: Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255 (2023) +59. Marr, D.: Vision: A computational investigation into the human representation and processing of visual information. MIT press (2010) +60. Marr, D., Poggio, T.: Cooperative computation of stereo disparity: A cooperative algorithm is derived for extracting disparity information from stereo image pairs. Science 194(4262), 283-287 (1976) +61. Min, J., Lee, J., Ponce, J., Cho, M.: Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543 (2019) + +62. Minsky, M., Papert, S.: An introduction to computational geometry. Cambridge, tiass., HIT 479(480), 104 (1969) +63. OpenAI: Gpt-4 technical report (2023) +64. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +66. Sarkar, A., Mai, H., Mahapatra, A., Lazebnik, S., Bhattachad, A.: Shadows don't lie and lines can't bend! generative models don't know projective geometry... for now. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28140-28149 (2024) +67. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +68. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022) +69. Shtedritski, A., Rupprecht, C., Vedaldi, A.: What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712 (2023) +70. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. CVPR (2021) +71. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023) +72. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881 (2023) +73. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) +74. Team, I.: InternlM: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM (2023) +75. Team, M.N.: Introducing mpt-7b: A new standard for open-source, commercially usable lms (2023), www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-05 +76. Torralba, A., Oliva, A.: Depth estimation from image structure. IEEE Transactions on pattern analysis and machine intelligence 24(9), 1226-1238 (2002) +77. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +78. Wang, F., Fu, X., Huang, J.Y., Li, Z., Liu, Q., Liu, X., Ma, M.D., Xu, N., Zhou, W., Zhang, K., et al.: Muirbench: A comprehensive benchmark for robust multi-image understanding. arXiv preprint arXiv:2406.09411 (2024) +79. Wang, J.Y., Adelson, E.H.: Layered representation for motion analysis. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. pp. 361-366. IEEE (1993) +80. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., Xu, J., Xu, B., Li, J., Dong, Y., Ding, M., Tang, J.: Cogvlm: Visual expert for pretrained language models (2023) + +81. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022) +82. Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., Li, H.: Dire for diffusion-generated image detection. arXiv preprint arXiv:2303.09295 (2023) +83. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., Zhou, D.: Chain-of-thought prompting elicits reasoning in large language models (2022) +84. Yan, A., Yang, Z., Wu, J., Zhu, W., Yang, J., Li, L., Lin, K., Wang, J., McAuley, J., Gao, J., et al.: List items one by one: A new data source and learning paradigm for multimodal llms. arXiv preprint arXiv:2404.16375 (2024) +85. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023) +86. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024) +87. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3081-3089 (2022) +88. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023) +89. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023) +90. Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W., Sun, Y., Wei, C., Yu, B., Yuan, R., Sun, R., Yin, M., Zheng, B., Yang, Z., Liu, Y., Huang, W., Sun, H., Su, Y., Chen, W.: Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502 (2023) +91. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems 35, 27469-27483 (2022) +92. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +93. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems 36 (2024) \ No newline at end of file diff --git a/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/images.zip b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cf84163305b57ffb0263179007d4be878ab8eddc --- /dev/null +++ b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40b92710ed9bd1a68c4a65674ee46e1594e597590ecc8dcf294f20aa89365007 +size 708465 diff --git a/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/layout.json b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a418b723466760951eed974d85830e885cf86cde --- /dev/null +++ b/2024/BLINK_ Multimodal Large Language Models Can See but Not Perceive/layout.json @@ -0,0 +1,13717 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 115, + 475, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 115, + 475, + 150 + ], + "spans": [ + { + "bbox": [ + 139, + 115, + 475, + 150 + ], + "type": "text", + "content": "BLINK: Multimodal Large Language Models Can See but Not Perceive" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": "Xingyu Fu" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": ", Yushi Hu" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": ", Bangzheng Li" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": ", Yu Feng" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": ", Haoyu Wang" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": ", Xudong Lin" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": ", Dan Roth" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": ", Noah A. Smith" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": ", Wei-Chiu Ma" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "text", + "content": ", and Ranjay Krishna" + }, + { + "bbox": [ + 138, + 171, + 475, + 196 + ], + "type": "inline_equation", + "content": "^{2,3\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 203, + 472, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 203, + 472, + 239 + ], + "spans": [ + { + "bbox": [ + 140, + 203, + 472, + 239 + ], + "type": "text", + "content": "1 University of Pennsylvania 2 University of Washington 3 Allen Institute for AI 4 University of California, Davis 5 Columbia University https://zeyofu.github.io/blink/" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 156, + 259, + 460, + 445 + ], + "blocks": [ + { + "bbox": [ + 156, + 259, + 460, + 445 + ], + "lines": [ + { + "bbox": [ + 156, + 259, + 460, + 445 + ], + "spans": [ + { + "bbox": [ + 156, + 259, + 460, + 445 + ], + "type": "image", + "image_path": "0fc6fa44ce957b1569e768062818938bf03ae186fbe98712022e11cf3b04054a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 457, + 482, + 525 + ], + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 525 + ], + "type": "text", + "content": "Figure 1: The BLINK Benchmark. BLINK contains 14 visual perception tasks that can be solved by humans \"within a blink\", but pose significant challenges for current multimodal LLMs. These tasks are inspired by classical computer vision problems and recast into multiple-choice questions for multimodal LLMs to answer. Notice that the visual prompts and questions in this figure are different from the actual ones used in the benchmark for illustrative purposes, and answers of the samples are provided in fn 6" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 544, + 455, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 544, + 455, + 665 + ], + "spans": [ + { + "bbox": [ + 159, + 544, + 455, + 665 + ], + "type": "text", + "content": "Abstract. We introduce BLINK, a new benchmark for multimodal language models (LLMs) that focuses on core visual perception abilities not found in other evaluations. Most of the BLINK tasks can be solved by humans \"within a blink\" (e.g., relative depth estimation, visual correspondence, forensics detection, and multi-view reasoning). However, we find these perception-demanding tasks cast significant challenges for current multimodal LLMs because they resist mediation through natural language. BLINK reforms 14 classic computer vision tasks into 3,807 multiple-choice questions, paired with single or multiple images and visual prompting. While humans get " + }, + { + "bbox": [ + 159, + 544, + 455, + 665 + ], + "type": "inline_equation", + "content": "95.70\\%" + }, + { + "bbox": [ + 159, + 544, + 455, + 665 + ], + "type": "text", + "content": " accuracy on average, BLINK is surprisingly challenging for existing multimodal LLMs: even the" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "text", + "content": "best-performing GPT-4V and Gemini achieve accuracies of " + }, + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "inline_equation", + "content": "51.26\\%" + }, + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "inline_equation", + "content": "45.72\\%" + }, + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "text", + "content": ", only " + }, + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "inline_equation", + "content": "13.17\\%" + }, + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "inline_equation", + "content": "7.63\\%" + }, + { + "bbox": [ + 159, + 117, + 455, + 194 + ], + "type": "text", + "content": " higher than random guessing, indicating that such perception abilities have not \"emerged\" yet in recent multimodal LLMs. Our analysis also highlights that specialist CV models could solve these problems much better, suggesting potential pathways for future improvements. We believe BLINK will stimulate the community to help multimodal LLMs catch up with human-level visual perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 204, + 455, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 204, + 455, + 228 + ], + "spans": [ + { + "bbox": [ + 160, + 204, + 455, + 228 + ], + "type": "text", + "content": "Keywords: Multi-modal Large Language Models " + }, + { + "bbox": [ + 160, + 204, + 455, + 228 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 204, + 455, + 228 + ], + "type": "text", + "content": " Vision-Language Benchmark " + }, + { + "bbox": [ + 160, + 204, + 455, + 228 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 204, + 455, + 228 + ], + "type": "text", + "content": " Visual Perception Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 246, + 230, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 230, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 230, + 258 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 483, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 483, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 483, + 437 + ], + "type": "text", + "content": "Compared to today, computer vision was originally attempting to interpret images as projections of 3D scenes, not just processing 2D arrays of flat \"patterns\" [25, 59, 62]. In this pursuit, early research developed a series of intermediate tasks: they focused on understanding optical properties like reflectance [12, 79], 3D primitives through multi-view reasoning [38, 60], geometric reasoning through depth estimation [76], instance recognition through visual correspondence [56], affordance through keypoint grounding [37], and forensics through intrinsic images [9]. Yet in the modern era of large language models (LLMs), we, as a community, have focused less on such perceptual tasks, and instead have developed new tasks, mostly expressed in natural language, emphasizing the vision-language connection learned by multimodal LLMs [3, 6, 18, 20, 24, 26, 51, 52, 57, 63, 66, 73, 80]. This might be because many traditional computer vision tasks resist mediation through natural language, due to the inherent imprecision of language (e.g., it is challenging to precisely pinpoint a spatial keypoint through language)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 483, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 483, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 483, + 640 + ], + "type": "text", + "content": "This paper aims to highlight crucial aspects of visual perception that have been overlooked when evaluating multimodal LLMs. To appropriately position our paper, let us revisit how we currently evaluate perception through using multimodal LLMs [44,45,48,53,54,58,90]. While many of these benchmarks have been popularized as the de facto evaluation measures for influential models like GPT-4V and Gemini-Pro, they conflate perception with language knowledge and reasoning. At the risk of singling out one benchmark, let us consider two questions highlighted in the popular MMBench [53]: “ Why is this hummingbird called ruby-throated?” and “ What will happen next? A: the person is gonna laugh B: the person is gonna cry.” For the first question, the vision subpart is to recognize the hummingbird. For the second, it only needs a coarse description of the image. Everything else is left to the language model to solve. Such a conflation has also been reported for other benchmarks by previous work [11,39,87]. Our experiments show that this conflation reductively evaluates perception as a dense captioning task. In other words, by replacing the image with a task-agnostic dense caption, our experiments show that a “blind” GPT-4 performs well on these “multimodal tasks”." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 642, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 483, + 665 + ], + "type": "text", + "content": "In response, we propose BLINK. BLINK reimagines traditional computer vision problems through a format that allows us to evaluate multimodal LLMs." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 125, + 460, + 328 + ], + "blocks": [ + { + "bbox": [ + 149, + 125, + 460, + 328 + ], + "lines": [ + { + "bbox": [ + 149, + 125, + 460, + 328 + ], + "spans": [ + { + "bbox": [ + 149, + 125, + 460, + 328 + ], + "type": "image", + "image_path": "8e11d99406a2568a4fec377d15b45b5eae34e4f22ea46b266503916f332bbefc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 339, + 483, + 440 + ], + "lines": [ + { + "bbox": [ + 130, + 339, + 483, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 483, + 440 + ], + "type": "text", + "content": "Figure 2: Comparison between BLINK and previous benchmarks. BLINK has several novel features: (1) BLINK incorporates diverse visual prompts, like circles, boxes, and image masks, while previous benchmarks only have text questions and answers. (2) BLINK evaluates a more comprehensive range of visual perception abilities, like multi-view reasoning, depth estimation, and reflectance estimation. Prior benchmarks are generally more focused on recognition-based VQA. (3) BLINK contains \"visual\" commonsense problems that humans can answer within seconds, while prior benchmarks like [90] require domain knowledge. The samples of previous benchmarks are from [45, 53, 90]. Part of our samples are curated from [10, 19, 29, 32, 36, 43, 91]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "type": "text", + "content": "As partially demonstrated in Figure 15, BLINK consists of 14 classic computer vision tasks, ranging from low-level pattern matching (e.g., visual correspondences estimation) to mid-level reasoning (e.g., relative depth estimation), and extending to high-level visual understanding (e.g., visual similarity). The image tasks are meticulously selected such that they are difficult to solve by reducing the evaluation using dense captioning; instead, the models must perceive the contents of the image(s) to answer. We recast each traditional task into a modern question-answering format, where answer choices are either images or text. BLINK contains 3.8K questions across 7.3K images, where questions may contain multiple images that are curated from a wide range of datasets [8, 10, 19, 29, 32, 36, 42, 47], encompassing indoor household scenes as well as outdoor urban or natural environments. The questions and choices are either derived from the datasets, or" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "BLINK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 620, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 620, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 620, + 482, + 665 + ], + "type": "text", + "content": "6 The answers of the examples in Figure 1 are as follows. Relative depth: B; jigsaw: A; multi-view reasoning: right; visual correspondence: A; semantic correspondence: C; forensics detection: final image; IQ test: D; visual similarity: upper one; functional correspondence: A; relative reflectance: they are about the same." + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "manually written by humans. On average, each question can be solved by a human subject within a BLINK of an eye, except the IQ test." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": "We carefully evaluate 17 multimodal LLMs with various sizes (i.e., 7B, 13B, 34B) on BLINK. We observe the paradox that while these problems are easy for humans (95.70% average accuracy), they are extremely hard for existing machinery - even GPT-4V model can only achieve 51.26% accuracy on average, which is 44.44% worse than humans, and 13.17% better than random guessing. We also experiment with specialist vision models and find that they perform much better than multimodal LLMs. For example, the specialist outperforms GPT-4V by 62.8% on visual correspondence estimation, 38.7% on relative depth estimation, and 34.6% on multi-view reasoning, in terms of absolute accuracy. Our findings indicate that the perceptual abilities of multimodal LLMs have been previously overestimated. Furthermore, these models may benefit from integrating insights from specialized models that excel in these areas. We believe BLINK can serve as an effective testbed for bridging the gap between traditional notions of perception and the modern generative capabilities of multimodal LLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 327, + 237, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 237, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 237, + 340 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "type": "text", + "content": "Multimodal models: Inspired by the impressive success in recent large language models (LLMs) [13,21,63,77,93], a sequence of studies explore multimodal LMMs that can jointly understand vision and language information and generate textual answers through adding a modality adaption structure between a frozen visual encoder [27,65,71] and a frozen LLM [77,93]. Flamingo [3] and BLIP-2 [46] are two of the earliest works to explore these transformer-based multimodality conjunction structures. They first pre-train on image-text matching datasets [15,42,47,67] and then fine-tune on task-specific datasets such as visual question answer (VQA) [4,34]. Starting from LLaVA [49,51], people use LLM synthesized instruction-following chat data (which are in VQA format) for instruction tuning and achieve much better results [7,16,24,73]. There have been extended studies that explore further capabilities of multimodal LLMs, especially on VQA reasoning [30,33,39,40,68,78,84,92]. However, they mainly focus on the textual reasoning abilities within the multimodal LLMs and do not emphasize visual perceptions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Multimodal benchmarks: Traditional vision-language datasets are designed to assess single-task capabilities, such as optical character recognition (OCR) [55], image captioning [47], and visual question answering [4,34]. However, these datasets are often not comprehensive enough to holistically assess multimodal LMMs on general perception and reasoning abilities. Many recent papers have built more comprehensive benchmarks. MME [28] is one of the earliest holistic benchmarks containing multi-modal Yes/No questions on the defined visual perception and language reasoning tasks. MM-Vet [89] includes six sub-features from the previous datasets including recognition-focused questions, OCR, and math, providing a diverse while discrete evaluation set. MMBench [53] covers more subjects and provides a more robust circular evaluation setting. Seed" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": "Bench [44,45] benchmark has a more diverse source of inputs, including multiple-image inputs and video, and includes more tasks. However, the visual perception questions in MME, MMBench, MM-Vet, and Seed-Bench are mainly extracted from existing VQA datasets or generated by GPT [63] from image descriptions such as COCO-Caption [47], and are recognition focused, covering topics such as object (attribute)recognition, and OCR. In contrast, we focus on multiple distinct nuanced perception abilities and recognition-level perception is only one of our focus. Some other multimodal benchmarks have distinct focuses. MMMU [90] aims at achieving expert-level artificial general intelligence by collecting domain-knowledge-required questions. HallusionBench [35] mainly tests the language hallucination and visual illusion phenomena. MathVista [58] presents exclusively mathematical domain visual questions based on images such as charts, tables, and diagrams. These benchmarks do not require human-level perception abilities as in BLINK and therefore cannot measure model visual perceptions holistically." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 305, + 292, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 305, + 292, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 305, + 292, + 319 + ], + "type": "text", + "content": "3 The BLINK Benchmark" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 334, + 483, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 483, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 483, + 467 + ], + "type": "text", + "content": "Our goal is to faithfully evaluate the visual perception capabilities of existing Multimodal LLMs. We seek to study the visual perception gap between humans and machineries, and offer deeper insights into potential pathways towards achieving more generalized machine perception. Based on the observation that existing benchmarks predominantly focus on evaluating visual recognition abilities, we introduce a novel benchmark, BLINK, designed to enable both quantitative and qualitative evaluation of the nuanced perception capabilities of multimodal LLM across various dimensions. We unfold this section by illustrating the overall design of BLINK (§3.1) and discussing its unique features comparing with previous benchmarks. Then we describe each task in detail, providing an in-depth explanation of the data curation process (§3.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 487, + 259, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 259, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 259, + 498 + ], + "type": "text", + "content": "3.1 Overview of BLINK" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "To ensure that one can effectively measure what Multimodal LLMs can or cannot perceive, we carefully select 14 tasks (see §3.2 for the full list) that are difficult to solve by reducing the evaluation into text-only questions using dense captioning. The tasks are drawn from either classic computer vision problems or recent applications of Multimodal LLMs, each of which requires a nuanced understanding of the visual data. They range from low-level pattern matching (e.g., visual correspondence) to mid-level spatial reasoning (e.g., relative depth), and up to high-level visual understanding (e.g., visual similarity). This variety allows for a systematic exploration of Multimodal LLMs' capabilities across different perceptual complexity layers. Furthermore, these visual tasks vary in granularity, ranging from pixels (e.g., relative reflectance) to patches (e.g., jigsaw) and extending to the full image (e.g., forensic detection), enabling us to evaluate models' proficiency in observing at various scales." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BLINK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 297, + 262 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 297, + 262 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 297, + 262 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 297, + 262 + ], + "type": "image", + "image_path": "ef107ef8238b4231e2ce547624ec425dbe05a0809cf83f9b2cc0efff16dc2a14.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 266, + 303, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 303, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 303, + 344 + ], + "type": "text", + "content": "Figure 3: Statistics of BLINK. 14 tasks range from pixel-level to image-level; from low-level pattern matching (e.g., visual correspondences) to mid-level reasoning (e.g., relative depth), further to high-level visual understanding (e.g., visual similarity)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 133, + 480, + 291 + ], + "blocks": [ + { + "bbox": [ + 313, + 133, + 480, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 133, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 480, + 291 + ], + "type": "image", + "image_path": "c8e5fdbf538a17f6d2bd882f0e8a08ae04b366d052f4a47df03065d39ebd23db.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 293, + 484, + 327 + ], + "lines": [ + { + "bbox": [ + 311, + 293, + 484, + 327 + ], + "spans": [ + { + "bbox": [ + 311, + 293, + 484, + 327 + ], + "type": "text", + "content": "Figure 4: Accuracies of multimodal LLMs on BLINK test set. Please refer to Table [1] and §4.2 for more results and discussions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 499 + ], + "type": "text", + "content": "To facilitate the evaluation of multimodal LLMs, we recast all tasks as multiple-choice question-answering problems. The options for answers may include images or texts, while the questions themselves can feature either single or multiple images. Prompts are designed to be both textual and visual in nature. We re-purposed several existing vision datasets as well as collected new data. In total, we contribute 3.9K multiple-choice questions and 7.3K images, with an even distribution between the validation and test sets. Numbers of each task are reported in Figure 3, and more detailed statistics can be found in Appendix A.5. Key features of BLINK: Comparing with previous benchmarks, BLINK has the following novel features:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 508, + 483, + 666 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 508, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 508, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 508, + 482, + 568 + ], + "type": "text", + "content": "- Visual prompting: Unlike existing benchmarks that support only text prompting, BLINK features a variety of visual prompts. This enables one to highlight specific areas within images, facilitating the evaluation of Multi-modal LLMs' detailed understanding of these regions. It also offers an interface for researchers to investigate the impact of visual prompting techniques." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 569, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 569, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 137, + 569, + 482, + 616 + ], + "type": "text", + "content": "- Perception beyond recognition: Besides visual recognition, BLINK considers a diverse set of visual perception abilities, such as 3D reasoning, geometric understanding, affordance reasoning, etc. The breadth allows one to evaluate Multimodal LLMs from an unique array of perspectives." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 483, + 666 + ], + "type": "text", + "content": "- \"Visual commonsense\" that does not require domain knowledge: The questions in BLINK are intentionally designed to be straightforward, requiring neither domain-specific knowledge nor expertise to answer. They are crafted in such a way that humans can solve them almost instantaneously," + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 163 + ], + "type": "text", + "content": "typically within a few seconds. This allows us to explore the fundamental gap in visual perception gap between humans and Multimodal LLMs, highlighting the paradox that problems easily solved by humans often pose significant challenges for machines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 164, + 481, + 270 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 138, + 164, + 481, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 164, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 164, + 481, + 209 + ], + "type": "text", + "content": "- Interleaved image-text formats: BLINK features a heterogeneous question-answering format, wherein both questions and choices can be presented as text or images. This diversity compels Multimodal LLMs to genuinely understand the questions, pushing the boundaries of their interpretative capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 212, + 480, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 480, + 270 + ], + "type": "text", + "content": "- Diverse image sources: BLINK comprises a wide range of in-the-wild images sourced from various origins, covering everything from indoor and outdoor scenes to object-centric views and landscapes. This collection spans abstract diagrams, synthesized images, and authentic photographs, ensuring a comprehensive examination of visual perception" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "content": "The design principles of BLINK are also illustrated in Figure 2. We will now describe each task in detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 322, + 295, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 295, + 333 + ], + "type": "text", + "content": "3.2 Dataset Collection Process" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 378 + ], + "type": "text", + "content": "BLINK comprises 14 tasks, all of which have been repurposed into a multiple-choice question-answering format. These tasks utilize a diverse collection of images from various sources and we ensure that each image is unique." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 378, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 481, + 472 + ], + "type": "text", + "content": "Visual correspondence: This task aims to evaluate the ability of Multimodal LLMs to understand and identify the same scene point across various viewpoints, lighting conditions, or time. We exploit HPatches [8] for this task. HPatches contains a number of image sequences, each of which are composed of images taken under different illuminations and/or viewpoints of a scene. For each question, we randomly sample two images and an interest point within them. Then we exploit the ground-truth homography to compute its correspondence. Finally, we randomly select three more interest points to serve as other choices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 556 + ], + "type": "text", + "content": "Relative reflectance: This task aims to compare the reflectance (albedo) of two pixels. It allows us to evaluate Multimodal LLMs' understanding of material properties and their interaction with light, which is crucial for applications requiring high-fidelity visual interpretations. We curate our samples using human annotations from the Intrinsic Images in the Wild (IIW) dataset [10]. Each question is based on an image and two specified points. The goal is to identify which point is darker, or whether the two points have similar reflectance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 617 + ], + "type": "text", + "content": "Relative depth: Humans are good at judging relative depth [19]. This task can thus serve as a proxy to validate if the geometric understanding capabilities of existing multimodal LLMs are close to human. We curate our samples using human annotations from the Depth in the Wild [19] dataset. Given an image and two specified points, the task is to determine which point is closer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Spatial relation: Understanding spatial relationships between objects in a scene is essential for interpreting complex visual environments. However, modern Multimodal LLMs often struggles with spatial concepts such as \"left\" and \"right\" [88]. This task helps us evaluate whether the models finally possess this" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BLINK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 129, + 220, + 167 + ], + "blocks": [ + { + "bbox": [ + 155, + 121, + 222, + 128 + ], + "lines": [ + { + "bbox": [ + 155, + 121, + 222, + 128 + ], + "spans": [ + { + "bbox": [ + 155, + 121, + 222, + 128 + ], + "type": "text", + "content": "Visual correspondence" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 129, + 220, + 167 + ], + "lines": [ + { + "bbox": [ + 155, + 129, + 220, + 167 + ], + "spans": [ + { + "bbox": [ + 155, + 129, + 220, + 167 + ], + "type": "image", + "image_path": "e2b6ee94633975a0e058fc3150fdd291369c324fa70bf054a07440df8075483c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 130, + 279, + 140 + ], + "lines": [ + { + "bbox": [ + 215, + 130, + 279, + 140 + ], + "spans": [ + { + "bbox": [ + 215, + 130, + 279, + 140 + ], + "type": "text", + "content": "A: Which point corresponds to the reference point 1?" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 228, + 152, + 241, + 156 + ], + "blocks": [ + { + "bbox": [ + 229, + 142, + 241, + 148 + ], + "lines": [ + { + "bbox": [ + 229, + 142, + 241, + 148 + ], + "spans": [ + { + "bbox": [ + 229, + 142, + 241, + 148 + ], + "type": "text", + "content": "(a) A" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 228, + 152, + 241, + 156 + ], + "lines": [ + { + "bbox": [ + 228, + 152, + 241, + 156 + ], + "spans": [ + { + "bbox": [ + 228, + 152, + 241, + 156 + ], + "type": "image", + "image_path": "5689ea0115dd42d13288f048d8b9b2dfee37e3cd72546038249cfa6c6305b164.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 245, + 152, + 256, + 156 + ], + "blocks": [ + { + "bbox": [ + 245, + 152, + 256, + 156 + ], + "lines": [ + { + "bbox": [ + 245, + 152, + 256, + 156 + ], + "spans": [ + { + "bbox": [ + 245, + 152, + 256, + 156 + ], + "type": "image", + "image_path": "42351252cbe567794610834a137b92fcb8209d07233d8dccdcfda37ed4e3640f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 263, + 152, + 274, + 156 + ], + "blocks": [ + { + "bbox": [ + 263, + 142, + 273, + 148 + ], + "lines": [ + { + "bbox": [ + 263, + 142, + 273, + 148 + ], + "spans": [ + { + "bbox": [ + 263, + 142, + 273, + 148 + ], + "type": "text", + "content": "(b) B" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 263, + 152, + 274, + 156 + ], + "lines": [ + { + "bbox": [ + 263, + 152, + 274, + 156 + ], + "spans": [ + { + "bbox": [ + 263, + 152, + 274, + 156 + ], + "type": "image", + "image_path": "fe48018c0a23e379c08d9faff86652e87a3b6a1b1225022555518de4ce257a06.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 228, + 157, + 241, + 162 + ], + "blocks": [ + { + "bbox": [ + 228, + 157, + 241, + 162 + ], + "lines": [ + { + "bbox": [ + 228, + 157, + 241, + 162 + ], + "spans": [ + { + "bbox": [ + 228, + 157, + 241, + 162 + ], + "type": "image", + "image_path": "7ea7490b1452a576c92d4774fd6e1272b5e2a6fc52b6b6507b7787d0d53a6e9a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 245, + 157, + 256, + 162 + ], + "blocks": [ + { + "bbox": [ + 245, + 157, + 256, + 162 + ], + "lines": [ + { + "bbox": [ + 245, + 157, + 256, + 162 + ], + "spans": [ + { + "bbox": [ + 245, + 157, + 256, + 162 + ], + "type": "image", + "image_path": "80889ee5fd0d100618eaef38dd2324c727c0dcbc55e11014bec1d6bb57b7a45a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 263, + 157, + 274, + 162 + ], + "blocks": [ + { + "bbox": [ + 263, + 157, + 274, + 162 + ], + "lines": [ + { + "bbox": [ + 263, + 157, + 274, + 162 + ], + "spans": [ + { + "bbox": [ + 263, + 157, + 274, + 162 + ], + "type": "image", + "image_path": "3478181c78b7cf1ecb16e18d01233255f363f91303a3a9f62f402839382a1c30.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 157, + 182, + 211, + 220 + ], + "blocks": [ + { + "bbox": [ + 159, + 174, + 201, + 182 + ], + "lines": [ + { + "bbox": [ + 159, + 174, + 201, + 182 + ], + "spans": [ + { + "bbox": [ + 159, + 174, + 201, + 182 + ], + "type": "text", + "content": "Relative depth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 182, + 211, + 220 + ], + "lines": [ + { + "bbox": [ + 157, + 182, + 211, + 220 + ], + "spans": [ + { + "bbox": [ + 157, + 182, + 211, + 220 + ], + "type": "image", + "image_path": "6e4d21331d2bae3ba8a9deb5b1548a1edb38db2187ec4cbd25f84931b6df2adf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 187, + 264, + 193 + ], + "lines": [ + { + "bbox": [ + 217, + 187, + 264, + 193 + ], + "spans": [ + { + "bbox": [ + 217, + 187, + 264, + 193 + ], + "type": "text", + "content": "Q: Which point is closer?" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 220, + 199, + 232, + 204 + ], + "blocks": [ + { + "bbox": [ + 220, + 199, + 232, + 204 + ], + "lines": [ + { + "bbox": [ + 220, + 199, + 232, + 204 + ], + "spans": [ + { + "bbox": [ + 220, + 199, + 232, + 204 + ], + "type": "image", + "image_path": "651d8987645fbe4fd4f03467e483671b0d06f9e388ca4e8c3315bb7883b40f80.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 235, + 199, + 241, + 204 + ], + "blocks": [ + { + "bbox": [ + 235, + 199, + 241, + 204 + ], + "lines": [ + { + "bbox": [ + 235, + 199, + 241, + 204 + ], + "spans": [ + { + "bbox": [ + 235, + 199, + 241, + 204 + ], + "type": "image", + "image_path": "9f30e061ade7dc0195bf14bcbb63fc65828a43a30d7fe27f7d812f4b8f399bd2.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 246, + 199, + 256, + 204 + ], + "blocks": [ + { + "bbox": [ + 246, + 199, + 256, + 204 + ], + "lines": [ + { + "bbox": [ + 246, + 199, + 256, + 204 + ], + "spans": [ + { + "bbox": [ + 246, + 199, + 256, + 204 + ], + "type": "image", + "image_path": "e76796139a94c9a6aa99d5eff68d41bdcf3204680d0473a4472ef424b4e8d612.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 263, + 199, + 273, + 204 + ], + "blocks": [ + { + "bbox": [ + 263, + 199, + 273, + 204 + ], + "lines": [ + { + "bbox": [ + 263, + 199, + 273, + 204 + ], + "spans": [ + { + "bbox": [ + 263, + 199, + 273, + 204 + ], + "type": "image", + "image_path": "1e1a45817760d86c08b5914f146029c5bab69308ef7c4322cdb3f5bd4f42ede4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 220, + 205, + 233, + 210 + ], + "blocks": [ + { + "bbox": [ + 220, + 205, + 233, + 210 + ], + "lines": [ + { + "bbox": [ + 220, + 205, + 233, + 210 + ], + "spans": [ + { + "bbox": [ + 220, + 205, + 233, + 210 + ], + "type": "image", + "image_path": "d6d450547c40a69d3bad0349117f19cf4bb4144c6f4edbabd3475573bc115c16.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 248, + 205, + 271, + 210 + ], + "blocks": [ + { + "bbox": [ + 248, + 205, + 271, + 210 + ], + "lines": [ + { + "bbox": [ + 248, + 205, + 271, + 210 + ], + "spans": [ + { + "bbox": [ + 248, + 205, + 271, + 210 + ], + "type": "image", + "image_path": "070b3cea6cc4151c43caa320a7e8ea0556ce8438961535030fb57ae8ac6aa891.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 155, + 236, + 215, + 275 + ], + "blocks": [ + { + "bbox": [ + 159, + 229, + 219, + 236 + ], + "lines": [ + { + "bbox": [ + 159, + 229, + 219, + 236 + ], + "spans": [ + { + "bbox": [ + 159, + 229, + 219, + 236 + ], + "type": "text", + "content": "Multi-view reasoning" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 236, + 215, + 275 + ], + "lines": [ + { + "bbox": [ + 155, + 236, + 215, + 275 + ], + "spans": [ + { + "bbox": [ + 155, + 236, + 215, + 275 + ], + "type": "image", + "image_path": "8b5dd7ad0226225c452ac6c3d90ec69d3bfca4ed92b749f113530608c58960d8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 236, + 294, + 257 + ], + "lines": [ + { + "bbox": [ + 216, + 236, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 216, + 236, + 294, + 257 + ], + "type": "text", + "content": "Q. The first image is from the beginning of the video and the second image is from the end. Is the camera moving towards left or right when shooting the video?" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 225, + 258, + 239, + 262 + ], + "blocks": [ + { + "bbox": [ + 225, + 258, + 239, + 262 + ], + "lines": [ + { + "bbox": [ + 225, + 258, + 239, + 262 + ], + "spans": [ + { + "bbox": [ + 225, + 258, + 239, + 262 + ], + "type": "image", + "image_path": "c07bb3cb678fcdbd3f0427849f29d0b99357966bd06659d47016af5ab92caaf2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 225, + 263, + 247, + 270 + ], + "blocks": [ + { + "bbox": [ + 225, + 263, + 247, + 270 + ], + "lines": [ + { + "bbox": [ + 225, + 263, + 247, + 270 + ], + "spans": [ + { + "bbox": [ + 225, + 263, + 247, + 270 + ], + "type": "image", + "image_path": "134c023a930b90a42c0aa5f4b9e710a809fc725c52f7c18b02e8f0acd3e124bc.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 255, + 258, + 271, + 263 + ], + "blocks": [ + { + "bbox": [ + 255, + 258, + 271, + 263 + ], + "lines": [ + { + "bbox": [ + 255, + 258, + 271, + 263 + ], + "spans": [ + { + "bbox": [ + 255, + 258, + 271, + 263 + ], + "type": "image", + "image_path": "5aca460fd98be45c3282bd208fdfb6c882271a97cf133fc43d9514d890f67037.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 255, + 263, + 274, + 269 + ], + "blocks": [ + { + "bbox": [ + 255, + 263, + 274, + 269 + ], + "lines": [ + { + "bbox": [ + 255, + 263, + 274, + 269 + ], + "spans": [ + { + "bbox": [ + 255, + 263, + 274, + 269 + ], + "type": "image", + "image_path": "a4619dea0e7f5eff96a0cfaaa33f2bfa79f888bcf39f49aa0229c6ee76bde353.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 255, + 269, + 274, + 273 + ], + "blocks": [ + { + "bbox": [ + 255, + 269, + 274, + 273 + ], + "lines": [ + { + "bbox": [ + 255, + 269, + 274, + 273 + ], + "spans": [ + { + "bbox": [ + 255, + 269, + 274, + 273 + ], + "type": "image", + "image_path": "6712115a8d9c2df0620071d2e014b4f7288c427aa46643ac03fc4de0b949ba45.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 155, + 291, + 184, + 328 + ], + "blocks": [ + { + "bbox": [ + 157, + 285, + 181, + 290 + ], + "lines": [ + { + "bbox": [ + 157, + 285, + 181, + 290 + ], + "spans": [ + { + "bbox": [ + 157, + 285, + 181, + 290 + ], + "type": "text", + "content": "Art style" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 291, + 184, + 328 + ], + "lines": [ + { + "bbox": [ + 155, + 291, + 184, + 328 + ], + "spans": [ + { + "bbox": [ + 155, + 291, + 184, + 328 + ], + "type": "image", + "image_path": "ab505912deabe487dc23c629b6b127b48a41d38e476edc200f091dfceb5f5b8f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 192, + 300, + 222, + 324 + ], + "blocks": [ + { + "bbox": [ + 188, + 292, + 288, + 298 + ], + "lines": [ + { + "bbox": [ + 188, + 292, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 188, + 292, + 288, + 298 + ], + "type": "text", + "content": "Q: Which image fits the missing part in the first image?" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 192, + 300, + 222, + 324 + ], + "lines": [ + { + "bbox": [ + 192, + 300, + 222, + 324 + ], + "spans": [ + { + "bbox": [ + 192, + 300, + 222, + 324 + ], + "type": "image", + "image_path": "feb01dd766f52d22f28bdec52168544036307b8907d6dad44b55ecafd0993de2.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 223, + 303, + 241, + 323 + ], + "blocks": [ + { + "bbox": [ + 223, + 303, + 241, + 323 + ], + "lines": [ + { + "bbox": [ + 223, + 303, + 241, + 323 + ], + "spans": [ + { + "bbox": [ + 223, + 303, + 241, + 323 + ], + "type": "image", + "image_path": "154e14039f83a667fe3bff9ade0e7844eed9da8138a41633888cf4c21578a2d6.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 244, + 301, + 255, + 306 + ], + "blocks": [ + { + "bbox": [ + 244, + 301, + 255, + 306 + ], + "lines": [ + { + "bbox": [ + 244, + 301, + 255, + 306 + ], + "spans": [ + { + "bbox": [ + 244, + 301, + 255, + 306 + ], + "type": "image", + "image_path": "6e4e3d0b464a53fd1734bfe604369ac137643e80e69c1b45acda61f865379cd6.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 256, + 301, + 286, + 324 + ], + "blocks": [ + { + "bbox": [ + 256, + 301, + 286, + 324 + ], + "lines": [ + { + "bbox": [ + 256, + 301, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 256, + 301, + 286, + 324 + ], + "type": "image", + "image_path": "a2c958e71850115845c9835a4f4d1e8d0c716fc17a7068573aa3db64eb4267d3.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 154, + 342, + 206, + 381 + ], + "blocks": [ + { + "bbox": [ + 157, + 335, + 186, + 342 + ], + "lines": [ + { + "bbox": [ + 157, + 335, + 186, + 342 + ], + "spans": [ + { + "bbox": [ + 157, + 335, + 186, + 342 + ], + "type": "text", + "content": "Counting" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 342, + 206, + 381 + ], + "lines": [ + { + "bbox": [ + 154, + 342, + 206, + 381 + ], + "spans": [ + { + "bbox": [ + 154, + 342, + 206, + 381 + ], + "type": "image", + "image_path": "9de8f24018cadd0c26138372fcb3e43dbead6766466b1f956eae7dfa62cbd0d5.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 216, + 358, + 227, + 363 + ], + "blocks": [ + { + "bbox": [ + 209, + 347, + 281, + 357 + ], + "lines": [ + { + "bbox": [ + 209, + 347, + 281, + 357 + ], + "spans": [ + { + "bbox": [ + 209, + 347, + 281, + 357 + ], + "type": "text", + "content": "Q: How many fingers are in front of the bathtub?" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 216, + 358, + 227, + 363 + ], + "lines": [ + { + "bbox": [ + 216, + 358, + 227, + 363 + ], + "spans": [ + { + "bbox": [ + 216, + 358, + 227, + 363 + ], + "type": "image", + "image_path": "5b3f5d32d8c88b9fe2cf844fca2748e74ef13af4cdaa0139ac5dd2842f6eeb15.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 216, + 365, + 226, + 369 + ], + "blocks": [ + { + "bbox": [ + 216, + 365, + 226, + 369 + ], + "lines": [ + { + "bbox": [ + 216, + 365, + 226, + 369 + ], + "spans": [ + { + "bbox": [ + 216, + 365, + 226, + 369 + ], + "type": "image", + "image_path": "016972865dae4d5cac90fcd7fc1c56814ffa00408a963f38caf78d4735eed579.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 216, + 370, + 226, + 376 + ], + "blocks": [ + { + "bbox": [ + 216, + 370, + 226, + 376 + ], + "lines": [ + { + "bbox": [ + 216, + 370, + 226, + 376 + ], + "spans": [ + { + "bbox": [ + 216, + 370, + 226, + 376 + ], + "type": "image", + "image_path": "e9c77077853e56a314a0b106f94c9de75869a2935d8f6b141d39a19a6761f69e.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 253, + 359, + 271, + 365 + ], + "blocks": [ + { + "bbox": [ + 253, + 359, + 271, + 365 + ], + "lines": [ + { + "bbox": [ + 253, + 359, + 271, + 365 + ], + "spans": [ + { + "bbox": [ + 253, + 359, + 271, + 365 + ], + "type": "image", + "image_path": "61901e0975f80a2ce43de3c3d06f9fdc4520ce8dc33dae75197da8c52f6c0c6d.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 253, + 370, + 264, + 376 + ], + "blocks": [ + { + "bbox": [ + 253, + 370, + 264, + 376 + ], + "lines": [ + { + "bbox": [ + 253, + 370, + 264, + 376 + ], + "spans": [ + { + "bbox": [ + 253, + 370, + 264, + 376 + ], + "type": "image", + "image_path": "fe66627b4d8ec6d0b645f6b539e64c83dd2ecc2057cc9bbd07fd0a98ddb7b58a.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "bbox": [ + 266, + 359, + 271, + 365 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 359, + 271, + 365 + ], + "spans": [ + { + "bbox": [ + 266, + 359, + 271, + 365 + ], + "type": "image", + "image_path": "a995b9a520f0fabe40c6bd8adc8ff896424ac9c8d24ebe81c112c2dea89a7e15.jpg" + } + ] + } + ], + "index": 45, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 266, + 370, + 280, + 376 + ], + "blocks": [ + { + "bbox": [ + 266, + 370, + 280, + 376 + ], + "lines": [ + { + "bbox": [ + 266, + 370, + 280, + 376 + ], + "spans": [ + { + "bbox": [ + 266, + 370, + 280, + 376 + ], + "type": "image", + "image_path": "985c6dd4a3161130177d1cb56a73df4a568eeec89a64b8861bda530bc9bbe0cd.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 156, + 401, + 238, + 436 + ], + "blocks": [ + { + "bbox": [ + 157, + 392, + 177, + 398 + ], + "lines": [ + { + "bbox": [ + 157, + 392, + 177, + 398 + ], + "spans": [ + { + "bbox": [ + 157, + 392, + 177, + 398 + ], + "type": "text", + "content": "IQ test" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 401, + 238, + 436 + ], + "lines": [ + { + "bbox": [ + 156, + 401, + 238, + 436 + ], + "spans": [ + { + "bbox": [ + 156, + 401, + 238, + 436 + ], + "type": "image", + "image_path": "2a025d187995d994c23a13f4b25f5ebce5d55927b498ade3796f87e1dbdddac8.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 403, + 292, + 413 + ], + "lines": [ + { + "bbox": [ + 241, + 403, + 292, + 413 + ], + "spans": [ + { + "bbox": [ + 241, + 403, + 292, + 413 + ], + "type": "text", + "content": "Q: Which image comes at the end?" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 241, + 418, + 258, + 423 + ], + "blocks": [ + { + "bbox": [ + 241, + 418, + 258, + 423 + ], + "lines": [ + { + "bbox": [ + 241, + 418, + 258, + 423 + ], + "spans": [ + { + "bbox": [ + 241, + 418, + 258, + 423 + ], + "type": "image", + "image_path": "cddbf4b51915617b679eb5a0f0a9f57710bececf4db69b62d1e267ce7842d654.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 269, + 418, + 288, + 424 + ], + "blocks": [ + { + "bbox": [ + 269, + 418, + 288, + 424 + ], + "lines": [ + { + "bbox": [ + 269, + 418, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 269, + 418, + 288, + 424 + ], + "type": "image", + "image_path": "d47c3c59997047da3f010529362000f9b62977a788ceeaee514cd0c4eb22f190.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 241, + 425, + 265, + 430 + ], + "blocks": [ + { + "bbox": [ + 241, + 425, + 265, + 430 + ], + "lines": [ + { + "bbox": [ + 241, + 425, + 265, + 430 + ], + "spans": [ + { + "bbox": [ + 241, + 425, + 265, + 430 + ], + "type": "image", + "image_path": "c12a496d68093d9ff5b9c06b21a55d99662c09003dcdacd8dd57fc863bfce5aa.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 269, + 425, + 288, + 430 + ], + "blocks": [ + { + "bbox": [ + 269, + 425, + 288, + 430 + ], + "lines": [ + { + "bbox": [ + 269, + 425, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 269, + 425, + 288, + 430 + ], + "type": "image", + "image_path": "bbc26ee5e64d405602bb624c425e2117ad46e75b77aed97d4e36857121215c3b.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 155, + 453, + 221, + 491 + ], + "blocks": [ + { + "bbox": [ + 157, + 446, + 182, + 452 + ], + "lines": [ + { + "bbox": [ + 157, + 446, + 182, + 452 + ], + "spans": [ + { + "bbox": [ + 157, + 446, + 182, + 452 + ], + "type": "text", + "content": "1 Semantic" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 453, + 221, + 491 + ], + "lines": [ + { + "bbox": [ + 155, + 453, + 221, + 491 + ], + "spans": [ + { + "bbox": [ + 155, + 453, + 221, + 491 + ], + "type": "image", + "image_path": "77c6b5c4b66ae600a2954ca2707715063f2a81aaa78cf46e014618a2c056c41e.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 454, + 295, + 464 + ], + "lines": [ + { + "bbox": [ + 223, + 454, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 223, + 454, + 295, + 464 + ], + "type": "text", + "content": "Q: Which point is semantically similar to the reference point 1?" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 225, + 466, + 251, + 471 + ], + "blocks": [ + { + "bbox": [ + 225, + 466, + 251, + 471 + ], + "lines": [ + { + "bbox": [ + 225, + 466, + 251, + 471 + ], + "spans": [ + { + "bbox": [ + 225, + 466, + 251, + 471 + ], + "type": "image", + "image_path": "766ac0850e000c625cad38a6a112b6ddb0193aa3d8a493bd0a6adb1e58ec9673.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 269, + 466, + 286, + 472 + ], + "blocks": [ + { + "bbox": [ + 269, + 466, + 286, + 472 + ], + "lines": [ + { + "bbox": [ + 269, + 466, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 269, + 466, + 286, + 472 + ], + "type": "image", + "image_path": "b159e2ae552fe6069d25ca9125b310ba3e584c3d3d68ba062a01bec1e50ca515.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 241, + 476, + 256, + 481 + ], + "blocks": [ + { + "bbox": [ + 241, + 476, + 256, + 481 + ], + "lines": [ + { + "bbox": [ + 241, + 476, + 256, + 481 + ], + "spans": [ + { + "bbox": [ + 241, + 476, + 256, + 481 + ], + "type": "image", + "image_path": "0127801748bc39d4a68bdeaa24b9171a0512c8c55b7a533f547d9da87fe06f10.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 269, + 476, + 294, + 489 + ], + "blocks": [ + { + "bbox": [ + 269, + 476, + 294, + 489 + ], + "lines": [ + { + "bbox": [ + 269, + 476, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 269, + 476, + 294, + 489 + ], + "type": "image", + "image_path": "b058eb3109671b70f69ca0cc10b29832c017f26d616aaf21132916232097d408.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 201, + 496, + 213, + 504 + ], + "blocks": [ + { + "bbox": [ + 201, + 496, + 213, + 504 + ], + "lines": [ + { + "bbox": [ + 201, + 496, + 213, + 504 + ], + "spans": [ + { + "bbox": [ + 201, + 496, + 213, + 504 + ], + "type": "image", + "image_path": "977586505e10e895db3137716321027709a98775c83bea83af6a3cd51c1fd8f3.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 211, + 498, + 234, + 503 + ], + "blocks": [ + { + "bbox": [ + 211, + 498, + 234, + 503 + ], + "lines": [ + { + "bbox": [ + 211, + 498, + 234, + 503 + ], + "spans": [ + { + "bbox": [ + 211, + 498, + 234, + 503 + ], + "type": "image", + "image_path": "5258312ff30e6325337d3434304f118059e78e2275466fcca4615ac864c652aa.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 242, + 497, + 264, + 503 + ], + "blocks": [ + { + "bbox": [ + 242, + 497, + 264, + 503 + ], + "lines": [ + { + "bbox": [ + 242, + 497, + 264, + 503 + ], + "spans": [ + { + "bbox": [ + 242, + 497, + 264, + 503 + ], + "type": "image", + "image_path": "2450015f28bc8359038ab5baf011ff1918fb929b8b830774f95f499c75d2f956.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 272, + 498, + 293, + 503 + ], + "blocks": [ + { + "bbox": [ + 272, + 498, + 293, + 503 + ], + "lines": [ + { + "bbox": [ + 272, + 498, + 293, + 503 + ], + "spans": [ + { + "bbox": [ + 272, + 498, + 293, + 503 + ], + "type": "image", + "image_path": "4479acecd70be14838dac929070a3013b4eced580ddc069b3e2167bf06e46ea5.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 295, + 498, + 326, + 503 + ], + "blocks": [ + { + "bbox": [ + 295, + 498, + 326, + 503 + ], + "lines": [ + { + "bbox": [ + 295, + 498, + 326, + 503 + ], + "spans": [ + { + "bbox": [ + 295, + 498, + 326, + 503 + ], + "type": "image", + "image_path": "3fabb53e70b0df933cfcf2ff058aded627f18874381583252c0218147899ea8b.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "bbox": [ + 334, + 497, + 349, + 503 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 497, + 349, + 503 + ], + "spans": [ + { + "bbox": [ + 334, + 497, + 349, + 503 + ], + "type": "image", + "image_path": "84901e43ad62eff7aea41c2fcac0b0597bd6aa9a1a530115bd53fb6ad53ad5f0.jpg" + } + ] + } + ], + "index": 66, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 351, + 498, + 401, + 503 + ], + "blocks": [ + { + "bbox": [ + 351, + 498, + 401, + 503 + ], + "lines": [ + { + "bbox": [ + 351, + 498, + 401, + 503 + ], + "spans": [ + { + "bbox": [ + 351, + 498, + 401, + 503 + ], + "type": "image", + "image_path": "3e2877f232288738e9d83959b67ee06a6ee1415b0637bdfa1a58075086cec9b6.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 312, + 129, + 373, + 167 + ], + "blocks": [ + { + "bbox": [ + 315, + 121, + 370, + 128 + ], + "lines": [ + { + "bbox": [ + 315, + 121, + 370, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 121, + 370, + 128 + ], + "type": "text", + "content": "Relative reflectance" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 129, + 373, + 167 + ], + "lines": [ + { + "bbox": [ + 312, + 129, + 373, + 167 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 373, + 167 + ], + "type": "image", + "image_path": "0b9ed0e4b8e4697279d0dc0192a5e54f4b7024770bb8ddf9dd79de1465aa97d2.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 130, + 439, + 145 + ], + "lines": [ + { + "bbox": [ + 343, + 130, + 439, + 145 + ], + "spans": [ + { + "bbox": [ + 343, + 130, + 439, + 145 + ], + "type": "text", + "content": "Q: Consider the surface color (color without shading) of the two points in the image. Which one is darker, or the color is about the same?" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 350, + 148, + 384, + 154 + ], + "blocks": [ + { + "bbox": [ + 350, + 148, + 384, + 154 + ], + "lines": [ + { + "bbox": [ + 350, + 148, + 384, + 154 + ], + "spans": [ + { + "bbox": [ + 350, + 148, + 384, + 154 + ], + "type": "image", + "image_path": "61094acecc0cfd4a3d870f961d1a833799c2c9ee349c9fa8fb80ffecc1df1768.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 350, + 155, + 386, + 160 + ], + "blocks": [ + { + "bbox": [ + 350, + 155, + 386, + 160 + ], + "lines": [ + { + "bbox": [ + 350, + 155, + 386, + 160 + ], + "spans": [ + { + "bbox": [ + 350, + 155, + 386, + 160 + ], + "type": "image", + "image_path": "d07bf20a09e40b882ae823b58833fd10757047409b80aa37ab46da4ad70ffb7f.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 386, + 155, + 397, + 160 + ], + "blocks": [ + { + "bbox": [ + 386, + 155, + 397, + 160 + ], + "lines": [ + { + "bbox": [ + 386, + 155, + 397, + 160 + ], + "spans": [ + { + "bbox": [ + 386, + 155, + 397, + 160 + ], + "type": "image", + "image_path": "37f550064cadd3bc93ca905ee7f88bc63fe3e59ac06e768b27826ecbdeae7f0e.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 404, + 155, + 422, + 160 + ], + "blocks": [ + { + "bbox": [ + 404, + 155, + 422, + 160 + ], + "lines": [ + { + "bbox": [ + 404, + 155, + 422, + 160 + ], + "spans": [ + { + "bbox": [ + 404, + 155, + 422, + 160 + ], + "type": "image", + "image_path": "ba3421d9168789866e2a712b775e62f18274de1b216c7fc8acffc845ca438723.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + } + ], + "index": 76 + }, + { + "type": "image", + "bbox": [ + 423, + 148, + 435, + 154 + ], + "blocks": [ + { + "bbox": [ + 386, + 148, + 435, + 154 + ], + "lines": [ + { + "bbox": [ + 386, + 148, + 435, + 154 + ], + "spans": [ + { + "bbox": [ + 386, + 148, + 435, + 154 + ], + "type": "text", + "content": "(b) B is darker" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 399, + 148, + 436, + 154 + ], + "lines": [ + { + "bbox": [ + 399, + 148, + 436, + 154 + ], + "spans": [ + { + "bbox": [ + 399, + 148, + 436, + 154 + ], + "type": "text", + "content": "(b) B is darker" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 423, + 148, + 435, + 154 + ], + "lines": [ + { + "bbox": [ + 423, + 148, + 435, + 154 + ], + "spans": [ + { + "bbox": [ + 423, + 148, + 435, + 154 + ], + "type": "image", + "image_path": "53a6af5ff11348ee5f46e3fce202019be61cfb32b164026ced9f061b9af3f65e.jpg" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_body" + } + ], + "index": 77 + }, + { + "type": "image", + "bbox": [ + 423, + 155, + 435, + 160 + ], + "blocks": [ + { + "bbox": [ + 423, + 155, + 435, + 160 + ], + "lines": [ + { + "bbox": [ + 423, + 155, + 435, + 160 + ], + "spans": [ + { + "bbox": [ + 423, + 155, + 435, + 160 + ], + "type": "image", + "image_path": "1afaa30997f6e8156423f465444dde3c653d50e3524a6df6f4daa7f0ebde1d07.jpg" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_body" + } + ], + "index": 78 + }, + { + "type": "image", + "bbox": [ + 313, + 182, + 343, + 218 + ], + "blocks": [ + { + "bbox": [ + 314, + 175, + 364, + 182 + ], + "lines": [ + { + "bbox": [ + 314, + 175, + 364, + 182 + ], + "spans": [ + { + "bbox": [ + 314, + 175, + 364, + 182 + ], + "type": "text", + "content": "Spatial reasoning" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 182, + 343, + 218 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 343, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 343, + 218 + ], + "type": "image", + "image_path": "d70b02c9f50fcc837aa4a8af57130ce5f3e41edf09f97b2e5b61584b85e38ad5.jpg" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_body" + } + ], + "index": 80 + }, + { + "type": "image", + "bbox": [ + 354, + 198, + 367, + 203 + ], + "blocks": [ + { + "bbox": [ + 348, + 184, + 434, + 194 + ], + "lines": [ + { + "bbox": [ + 348, + 184, + 434, + 194 + ], + "spans": [ + { + "bbox": [ + 348, + 184, + 434, + 194 + ], + "type": "text", + "content": "Q: Is the bed at the right side of the dining table?" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 354, + 198, + 367, + 203 + ], + "lines": [ + { + "bbox": [ + 354, + 198, + 367, + 203 + ], + "spans": [ + { + "bbox": [ + 354, + 198, + 367, + 203 + ], + "type": "image", + "image_path": "1ae464d1233777e17c372c003736fb132d33cd218ca889dfe7b701bad78f747f.jpg" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_body" + } + ], + "index": 82 + }, + { + "type": "image", + "bbox": [ + 371, + 198, + 383, + 204 + ], + "blocks": [ + { + "bbox": [ + 371, + 198, + 383, + 204 + ], + "lines": [ + { + "bbox": [ + 371, + 198, + 383, + 204 + ], + "spans": [ + { + "bbox": [ + 371, + 198, + 383, + 204 + ], + "type": "image", + "image_path": "d9ef6393e6e18068156a0ac93d7488e5612c66ac0055792d3aa09df9b1ddb629.jpg" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_body" + } + ], + "index": 83 + }, + { + "type": "image", + "bbox": [ + 388, + 198, + 399, + 204 + ], + "blocks": [ + { + "bbox": [ + 388, + 198, + 399, + 204 + ], + "lines": [ + { + "bbox": [ + 388, + 198, + 399, + 204 + ], + "spans": [ + { + "bbox": [ + 388, + 198, + 399, + 204 + ], + "type": "image", + "image_path": "fec253e3c0343e070193a68c69e4cbf2d407e8066b8b8b25525ab29732f205aa.jpg" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_body" + } + ], + "index": 84 + }, + { + "type": "image", + "bbox": [ + 405, + 198, + 416, + 204 + ], + "blocks": [ + { + "bbox": [ + 405, + 198, + 416, + 204 + ], + "lines": [ + { + "bbox": [ + 405, + 198, + 416, + 204 + ], + "spans": [ + { + "bbox": [ + 405, + 198, + 416, + 204 + ], + "type": "image", + "image_path": "aa86a63d228c6745c047284f68d6e85f4d3213c8afb5aaac2397caac7d845733.jpg" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_body" + } + ], + "index": 85 + }, + { + "type": "image", + "bbox": [ + 405, + 204, + 423, + 211 + ], + "blocks": [ + { + "bbox": [ + 405, + 204, + 423, + 211 + ], + "lines": [ + { + "bbox": [ + 405, + 204, + 423, + 211 + ], + "spans": [ + { + "bbox": [ + 405, + 204, + 423, + 211 + ], + "type": "image", + "image_path": "dde2ecb0cfb6bffb2124c796d412b0f9ee6bef3260a2c8390e3d27d10d4f58d2.jpg" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "image_body" + } + ], + "index": 86 + }, + { + "type": "image", + "bbox": [ + 312, + 237, + 351, + 275 + ], + "blocks": [ + { + "bbox": [ + 314, + 230, + 334, + 236 + ], + "lines": [ + { + "bbox": [ + 314, + 230, + 334, + 236 + ], + "spans": [ + { + "bbox": [ + 314, + 230, + 334, + 236 + ], + "type": "text", + "content": "Jigsaw" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 237, + 351, + 275 + ], + "lines": [ + { + "bbox": [ + 312, + 237, + 351, + 275 + ], + "spans": [ + { + "bbox": [ + 312, + 237, + 351, + 275 + ], + "type": "image", + "image_path": "534f20444fc29b577843427ed47bff55b404994ed4050e585dbe7b96549931c4.jpg" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 240, + 452, + 246 + ], + "lines": [ + { + "bbox": [ + 353, + 240, + 452, + 246 + ], + "spans": [ + { + "bbox": [ + 353, + 240, + 452, + 246 + ], + "type": "text", + "content": "Q: Which image fits the missing part in the first image?" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 88 + }, + { + "type": "image", + "bbox": [ + 361, + 248, + 398, + 272 + ], + "blocks": [ + { + "bbox": [ + 361, + 248, + 398, + 272 + ], + "lines": [ + { + "bbox": [ + 361, + 248, + 398, + 272 + ], + "spans": [ + { + "bbox": [ + 361, + 248, + 398, + 272 + ], + "type": "image", + "image_path": "6e36f4e07db8660f686e01ac09f3299c4812718d7106ccf2ac2d87230118f7d2.jpg" + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "image_body" + } + ], + "index": 90 + }, + { + "type": "image", + "bbox": [ + 399, + 248, + 405, + 254 + ], + "blocks": [ + { + "bbox": [ + 399, + 248, + 405, + 254 + ], + "lines": [ + { + "bbox": [ + 399, + 248, + 405, + 254 + ], + "spans": [ + { + "bbox": [ + 399, + 248, + 405, + 254 + ], + "type": "image", + "image_path": "fce774468fa3f1902ecaa222e8647770907bf29306f418dba5d08bb278458d42.jpg" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_body" + } + ], + "index": 91 + }, + { + "type": "image", + "bbox": [ + 405, + 248, + 444, + 272 + ], + "blocks": [ + { + "bbox": [ + 405, + 248, + 444, + 272 + ], + "lines": [ + { + "bbox": [ + 405, + 248, + 444, + 272 + ], + "spans": [ + { + "bbox": [ + 405, + 248, + 444, + 272 + ], + "type": "image", + "image_path": "b76e2ef980b6667ae40db8d8d66970f42d07944af034495d92f17fcc4251cd93.jpg" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "image_body" + } + ], + "index": 92 + }, + { + "type": "image", + "bbox": [ + 312, + 292, + 370, + 330 + ], + "blocks": [ + { + "bbox": [ + 314, + 285, + 369, + 291 + ], + "lines": [ + { + "bbox": [ + 314, + 285, + 369, + 291 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 369, + 291 + ], + "type": "text", + "content": "Object localization" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 292, + 370, + 330 + ], + "lines": [ + { + "bbox": [ + 312, + 292, + 370, + 330 + ], + "spans": [ + { + "bbox": [ + 312, + 292, + 370, + 330 + ], + "type": "image", + "image_path": "cec570b58da17bf2015e26eed1f874264fa4083f6c980e1f1a82f7a489ee45fd.jpg" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 295, + 441, + 304 + ], + "lines": [ + { + "bbox": [ + 374, + 295, + 441, + 304 + ], + "spans": [ + { + "bbox": [ + 374, + 295, + 441, + 304 + ], + "type": "text", + "content": "Q: Which bounding box more accurately encloses the bun?" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 315, + 336, + 370, + 342 + ], + "lines": [ + { + "bbox": [ + 315, + 336, + 370, + 342 + ], + "spans": [ + { + "bbox": [ + 315, + 336, + 370, + 342 + ], + "type": "text", + "content": "Forensics detection" + } + ] + } + ], + "index": 101, + "angle": 0, + "type": "image_caption" + } + ], + "index": 94 + }, + { + "type": "image", + "bbox": [ + 412, + 311, + 422, + 316 + ], + "blocks": [ + { + "bbox": [ + 412, + 311, + 422, + 316 + ], + "lines": [ + { + "bbox": [ + 412, + 311, + 422, + 316 + ], + "spans": [ + { + "bbox": [ + 412, + 311, + 422, + 316 + ], + "type": "image", + "image_path": "34fb30e756a205b5554d37df4d0d7feb712b48471f673b7080410d072e0395ec.jpg" + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "image_body" + } + ], + "index": 96 + }, + { + "bbox": [ + 411, + 317, + 422, + 323 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 317, + 422, + 323 + ], + "spans": [ + { + "bbox": [ + 411, + 317, + 422, + 323 + ], + "type": "image", + "image_path": "c1001e5222e5b1e7bc8bc45e86ac3dd276734c57de5550e7f680ecf3bed48ca4.jpg" + } + ] + } + ], + "index": 97, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 424, + 311, + 435, + 316 + ], + "blocks": [ + { + "bbox": [ + 424, + 311, + 435, + 316 + ], + "lines": [ + { + "bbox": [ + 424, + 311, + 435, + 316 + ], + "spans": [ + { + "bbox": [ + 424, + 311, + 435, + 316 + ], + "type": "image", + "image_path": "98eeba09cb4216392f2875bf2691bf06c6add1f2b0634634aa06160206c84710.jpg" + } + ] + } + ], + "index": 98, + "angle": 0, + "type": "image_body" + } + ], + "index": 98 + }, + { + "bbox": [ + 411, + 317, + 422, + 323 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 317, + 422, + 323 + ], + "spans": [ + { + "bbox": [ + 411, + 317, + 422, + 323 + ], + "type": "image", + "image_path": "c1001e5222e5b1e7bc8bc45e86ac3dd276734c57de5550e7f680ecf3bed48ca4.jpg" + } + ] + } + ], + "index": 99, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 424, + 323, + 432, + 327 + ], + "blocks": [ + { + "bbox": [ + 424, + 323, + 432, + 327 + ], + "lines": [ + { + "bbox": [ + 424, + 323, + 432, + 327 + ], + "spans": [ + { + "bbox": [ + 424, + 323, + 432, + 327 + ], + "type": "image", + "image_path": "999c312274771ddff17a6a6482752b6694d6a8a2b161d6e3fed808a77eabe7d1.jpg" + } + ] + } + ], + "index": 100, + "angle": 0, + "type": "image_body" + } + ], + "index": 100 + }, + { + "type": "image", + "bbox": [ + 313, + 352, + 343, + 382 + ], + "blocks": [ + { + "bbox": [ + 313, + 346, + 414, + 352 + ], + "lines": [ + { + "bbox": [ + 313, + 346, + 414, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 414, + 352 + ], + "type": "text", + "content": "Q: Which image is most likely to be a real photograph?" + } + ] + } + ], + "index": 102, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 313, + 352, + 343, + 382 + ], + "lines": [ + { + "bbox": [ + 313, + 352, + 343, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 343, + 382 + ], + "type": "image", + "image_path": "2b4c98f281143e0e58fe60448be2482885d4ffe636add08314aa1c9d262dd0c2.jpg" + } + ] + } + ], + "index": 103, + "angle": 0, + "type": "image_body" + } + ], + "index": 103 + }, + { + "type": "image", + "bbox": [ + 347, + 353, + 375, + 375 + ], + "blocks": [ + { + "bbox": [ + 347, + 353, + 375, + 375 + ], + "lines": [ + { + "bbox": [ + 347, + 353, + 375, + 375 + ], + "spans": [ + { + "bbox": [ + 347, + 353, + 375, + 375 + ], + "type": "image", + "image_path": "393bef96cbf2b2051f12af281ca418e5a2b802d8ba319d1770338acf32cfe6fe.jpg" + } + ] + } + ], + "index": 104, + "angle": 0, + "type": "image_body" + } + ], + "index": 104 + }, + { + "type": "image", + "bbox": [ + 376, + 353, + 407, + 375 + ], + "blocks": [ + { + "bbox": [ + 376, + 353, + 407, + 375 + ], + "lines": [ + { + "bbox": [ + 376, + 353, + 407, + 375 + ], + "spans": [ + { + "bbox": [ + 376, + 353, + 407, + 375 + ], + "type": "image", + "image_path": "23dd786dbd85f9401a43a47adae456a8a1e6cda94a89954b14b006d7143cde32.jpg" + } + ] + } + ], + "index": 105, + "angle": 0, + "type": "image_body" + } + ], + "index": 105 + }, + { + "type": "image", + "bbox": [ + 413, + 353, + 440, + 375 + ], + "blocks": [ + { + "bbox": [ + 413, + 353, + 440, + 375 + ], + "lines": [ + { + "bbox": [ + 413, + 353, + 440, + 375 + ], + "spans": [ + { + "bbox": [ + 413, + 353, + 440, + 375 + ], + "type": "image", + "image_path": "18b7e8e079a07bfb9190fa99c9aa5eb35eb09742fa0b4ee4f5ddbfa78c224308.jpg" + } + ] + } + ], + "index": 106, + "angle": 0, + "type": "image_body" + } + ], + "index": 106 + }, + { + "type": "image", + "bbox": [ + 312, + 398, + 351, + 437 + ], + "blocks": [ + { + "bbox": [ + 314, + 392, + 357, + 398 + ], + "lines": [ + { + "bbox": [ + 314, + 392, + 357, + 398 + ], + "spans": [ + { + "bbox": [ + 314, + 392, + 357, + 398 + ], + "type": "text", + "content": "Visual similarity" + } + ] + } + ], + "index": 107, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 398, + 351, + 437 + ], + "lines": [ + { + "bbox": [ + 312, + 398, + 351, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 398, + 351, + 437 + ], + "type": "image", + "image_path": "da075ed77930d542e54affdbc601981218f67d91191fb6db799dde6aa0a37c07.jpg" + } + ] + } + ], + "index": 108, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 401, + 453, + 407 + ], + "lines": [ + { + "bbox": [ + 353, + 401, + 453, + 407 + ], + "spans": [ + { + "bbox": [ + 353, + 401, + 453, + 407 + ], + "type": "text", + "content": "Q: Which image is most similar to the reference image?" + } + ] + } + ], + "index": 109, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 108 + }, + { + "type": "image", + "bbox": [ + 355, + 410, + 389, + 434 + ], + "blocks": [ + { + "bbox": [ + 355, + 410, + 389, + 434 + ], + "lines": [ + { + "bbox": [ + 355, + 410, + 389, + 434 + ], + "spans": [ + { + "bbox": [ + 355, + 410, + 389, + 434 + ], + "type": "image", + "image_path": "94dc6ee69e3b8b2824b26ccad8a65d799f519415bf62c14f5cdbf649b2f71ed4.jpg" + } + ] + } + ], + "index": 110, + "angle": 0, + "type": "image_body" + } + ], + "index": 110 + }, + { + "type": "image", + "bbox": [ + 390, + 410, + 405, + 427 + ], + "blocks": [ + { + "bbox": [ + 390, + 410, + 405, + 427 + ], + "lines": [ + { + "bbox": [ + 390, + 410, + 405, + 427 + ], + "spans": [ + { + "bbox": [ + 390, + 410, + 405, + 427 + ], + "type": "image", + "image_path": "362cdf0b32cc3029051dd9242451a363e060c252e4db786d2952879d6f41a03e.jpg" + } + ] + } + ], + "index": 111, + "angle": 0, + "type": "image_body" + } + ], + "index": 111 + }, + { + "type": "image", + "bbox": [ + 406, + 410, + 422, + 434 + ], + "blocks": [ + { + "bbox": [ + 406, + 410, + 422, + 434 + ], + "lines": [ + { + "bbox": [ + 406, + 410, + 422, + 434 + ], + "spans": [ + { + "bbox": [ + 406, + 410, + 422, + 434 + ], + "type": "image", + "image_path": "4f00cff1e151a939a03b146f4ff23b6e8a694109dae9ca33b50ba59a33aa55dd.jpg" + } + ] + } + ], + "index": 112, + "angle": 0, + "type": "image_body" + } + ], + "index": 112 + }, + { + "type": "image", + "bbox": [ + 423, + 410, + 446, + 434 + ], + "blocks": [ + { + "bbox": [ + 423, + 410, + 446, + 434 + ], + "lines": [ + { + "bbox": [ + 423, + 410, + 446, + 434 + ], + "spans": [ + { + "bbox": [ + 423, + 410, + 446, + 434 + ], + "type": "image", + "image_path": "02055b2953abfab149c98e35ba75da7478fa9cae5729ea08de56951123680174.jpg" + } + ] + } + ], + "index": 113, + "angle": 0, + "type": "image_body" + } + ], + "index": 113 + }, + { + "type": "image", + "bbox": [ + 312, + 453, + 400, + 492 + ], + "blocks": [ + { + "bbox": [ + 314, + 446, + 389, + 453 + ], + "lines": [ + { + "bbox": [ + 314, + 446, + 389, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 446, + 389, + 453 + ], + "type": "text", + "content": "Functional correspondence" + } + ] + } + ], + "index": 114, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 453, + 400, + 492 + ], + "lines": [ + { + "bbox": [ + 312, + 453, + 400, + 492 + ], + "spans": [ + { + "bbox": [ + 312, + 453, + 400, + 492 + ], + "type": "image", + "image_path": "40e6e0ee580a2ec686de7bf6f98978ae662f3af20794de0da89b93f937560ff1.jpg" + } + ] + } + ], + "index": 115, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 456, + 454, + 470 + ], + "lines": [ + { + "bbox": [ + 400, + 456, + 454, + 470 + ], + "spans": [ + { + "bbox": [ + 400, + 456, + 454, + 470 + ], + "type": "text", + "content": "Q Which point is functionally different from reference point 1 during mixing?" + } + ] + } + ], + "index": 116, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 115 + }, + { + "type": "image", + "bbox": [ + 405, + 472, + 416, + 477 + ], + "blocks": [ + { + "bbox": [ + 405, + 472, + 416, + 477 + ], + "lines": [ + { + "bbox": [ + 405, + 472, + 416, + 477 + ], + "spans": [ + { + "bbox": [ + 405, + 472, + 416, + 477 + ], + "type": "image", + "image_path": "62cf777146c3edb5dbeeb788cf96cda371fcbf5fa010ecad725546adc555109f.jpg" + } + ] + } + ], + "index": 117, + "angle": 0, + "type": "image_body" + } + ], + "index": 117 + }, + { + "type": "image", + "bbox": [ + 417, + 472, + 433, + 477 + ], + "blocks": [ + { + "bbox": [ + 417, + 472, + 433, + 477 + ], + "lines": [ + { + "bbox": [ + 417, + 472, + 433, + 477 + ], + "spans": [ + { + "bbox": [ + 417, + 472, + 433, + 477 + ], + "type": "image", + "image_path": "4467d0ee1e86e3cc91aa13bcc4ee39ab78b66913806341aea73dfa072d6c3454.jpg" + } + ] + } + ], + "index": 118, + "angle": 0, + "type": "image_body" + } + ], + "index": 118 + }, + { + "type": "image", + "bbox": [ + 405, + 477, + 416, + 482 + ], + "blocks": [ + { + "bbox": [ + 405, + 477, + 416, + 482 + ], + "lines": [ + { + "bbox": [ + 405, + 477, + 416, + 482 + ], + "spans": [ + { + "bbox": [ + 405, + 477, + 416, + 482 + ], + "type": "image", + "image_path": "dbba412a63df6715cc5cddb77f69099e640442411b1b8afc32ee77ad4f1a1b3e.jpg" + } + ] + } + ], + "index": 119, + "angle": 0, + "type": "image_body" + } + ], + "index": 119 + }, + { + "type": "image", + "bbox": [ + 417, + 477, + 433, + 482 + ], + "blocks": [ + { + "bbox": [ + 417, + 477, + 433, + 482 + ], + "lines": [ + { + "bbox": [ + 417, + 477, + 433, + 482 + ], + "spans": [ + { + "bbox": [ + 417, + 477, + 433, + 482 + ], + "type": "image", + "image_path": "9e4b3eea5e9df564b65ef4e09cbdd1a0cb75aff3623dbb660d06a30c39079003.jpg" + } + ] + } + ], + "index": 120, + "angle": 0, + "type": "image_body" + } + ], + "index": 120 + }, + { + "type": "image", + "bbox": [ + 405, + 483, + 416, + 487 + ], + "blocks": [ + { + "bbox": [ + 405, + 483, + 416, + 487 + ], + "lines": [ + { + "bbox": [ + 405, + 483, + 416, + 487 + ], + "spans": [ + { + "bbox": [ + 405, + 483, + 416, + 487 + ], + "type": "image", + "image_path": "2e1576154b3887c795ba3271aff1dc324f3050072bb87dfa87d50f9216715db1.jpg" + } + ] + } + ], + "index": 121, + "angle": 0, + "type": "image_body" + } + ], + "index": 121 + }, + { + "type": "image", + "bbox": [ + 417, + 483, + 444, + 487 + ], + "blocks": [ + { + "bbox": [ + 417, + 483, + 444, + 487 + ], + "lines": [ + { + "bbox": [ + 417, + 483, + 444, + 487 + ], + "spans": [ + { + "bbox": [ + 417, + 483, + 444, + 487 + ], + "type": "image", + "image_path": "083be8e78cb1fbec184a77b304474a13e9151a088f6ec3266305c6e2fc450524.jpg" + } + ] + } + ], + "index": 122, + "angle": 0, + "type": "image_body" + } + ], + "index": 122 + }, + { + "type": "image", + "bbox": [ + 315, + 498, + 328, + 505 + ], + "blocks": [ + { + "bbox": [ + 315, + 498, + 328, + 505 + ], + "lines": [ + { + "bbox": [ + 315, + 498, + 328, + 505 + ], + "spans": [ + { + "bbox": [ + 315, + 498, + 328, + 505 + ], + "type": "image", + "image_path": "07dffa79710d93202c148ad77796d3ac999e063be525ca284d5a86ebff76811e.jpg" + } + ] + } + ], + "index": 123, + "angle": 0, + "type": "image_body" + } + ], + "index": 123 + }, + { + "bbox": [ + 334, + 497, + 349, + 503 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 497, + 349, + 503 + ], + "spans": [ + { + "bbox": [ + 334, + 497, + 349, + 503 + ], + "type": "image", + "image_path": "84901e43ad62eff7aea41c2fcac0b0597bd6aa9a1a530115bd53fb6ad53ad5f0.jpg" + } + ] + } + ], + "index": 124, + "type": "text" + }, + { + "bbox": [ + 351, + 498, + 362, + 503 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 498, + 362, + 503 + ], + "spans": [ + { + "bbox": [ + 351, + 498, + 362, + 503 + ], + "type": "image", + "image_path": "cc1a59a78bbb3f6252df4b92eae0b77e5c744212c4ad6f64a4d2a64e25bb6c86.jpg" + } + ] + } + ], + "index": 125, + "type": "text" + }, + { + "bbox": [ + 362, + 498, + 401, + 503 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 498, + 401, + 503 + ], + "spans": [ + { + "bbox": [ + 362, + 498, + 401, + 503 + ], + "type": "image", + "image_path": "3ed93850585e0edaf843a0444d5d5e1a1a28b29c06d3844da044a8ef94a1d617.jpg" + } + ] + } + ], + "index": 126, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 373, + 503, + 384, + 506 + ], + "blocks": [ + { + "bbox": [ + 373, + 503, + 384, + 506 + ], + "lines": [ + { + "bbox": [ + 373, + 503, + 384, + 506 + ], + "spans": [ + { + "bbox": [ + 373, + 503, + 384, + 506 + ], + "type": "image", + "image_path": "5eaae702c43824d0cfc94c3fb0d0264c19e004968e40cf4660eba92685aec38a.jpg" + } + ] + } + ], + "index": 127, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 520, + 482, + 575 + ], + "lines": [ + { + "bbox": [ + 132, + 520, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 482, + 575 + ], + "type": "text", + "content": "Figure 5: Qualitative results on BLINK. For each task, we show the choice of LLaVAv1.6-34B [51], Qwen-VL-Max [7], Gemini Pro [73], GPT-4V [63], and humans. Red choice indicates the ground truth. Notice that the markers are intentionally enlarged for visualization purposes, and we make some images inset images to save space. For IQ test, the third image is constructed by overlaying the first and second images." + } + ] + } + ], + "index": 130, + "angle": 0, + "type": "image_caption" + } + ], + "index": 127 + }, + { + "type": "image", + "bbox": [ + 386, + 499, + 401, + 504 + ], + "blocks": [ + { + "bbox": [ + 386, + 499, + 401, + 504 + ], + "lines": [ + { + "bbox": [ + 386, + 499, + 401, + 504 + ], + "spans": [ + { + "bbox": [ + 386, + 499, + 401, + 504 + ], + "type": "image", + "image_path": "34535316fb4354c16131856371a6f902772a22d071b46d16c79882790df0789e.jpg" + } + ] + } + ], + "index": 128, + "angle": 0, + "type": "image_body" + } + ], + "index": 128 + }, + { + "type": "image", + "bbox": [ + 402, + 499, + 417, + 504 + ], + "blocks": [ + { + "bbox": [ + 402, + 499, + 417, + 504 + ], + "lines": [ + { + "bbox": [ + 402, + 499, + 417, + 504 + ], + "spans": [ + { + "bbox": [ + 402, + 499, + 417, + 504 + ], + "type": "image", + "image_path": "811ebca9e7c09a042df12987a66b1f0f8a8804714ce607c17c10323becd89d9b.jpg" + } + ] + } + ], + "index": 129, + "angle": 0, + "type": "image_body" + } + ], + "index": 129 + }, + { + "bbox": [ + 130, + 602, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 602, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 602, + 482, + 666 + ], + "type": "text", + "content": "vital skill. We curate our samples from the Visual Spatial Reasoning [48] dataset. Each sample contains an image and a claim. The task is to determine if the claim is true or false. We reformat the claims into binary questions via GPT-3.5 [13]. Multi-view reasoning: This task is centered on evaluating the multi-view reasoning capabilities of Multimodal LLMs. The objective is to deduce the" + } + ] + } + ], + "index": 131 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "relative camera motion based on two images of an object captured from different viewpoints. Our data is sourced from the Wild6D dataset [91], which features videos of various objects recorded in diverse settings. We select two random frames from each video to calculate the relative camera motion. Recognizing that even humans might struggle to precisely articulate 3D motion details, we simplify the task by classifying motions into two broad categories: moving towards the left or moving towards the right. Despite the simplicity of these questions, as we will later demonstrate, they pose significant challenges for current models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 214, + 483, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 483, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 483, + 286 + ], + "type": "text", + "content": "Jigsaw: This task assesses the ability of Multimodal LLMs to recognize and group patterns, aligning patches based on continuity in shape, color, and texture. We utilize images from the TARA dataset [32] and segment each of them into a 3x3 grid. We retain the three segments from the upper left corner as the reference image, and treat the central segment along with a randomly chosen segment as options. The objective is to identify the correct patch (i.e., the central patch)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 288, + 483, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 483, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 483, + 396 + ], + "type": "text", + "content": "Art style: This task evaluates Multimodal LLMs capability to analyze and discern both local and global similarities in art styles among multiple images. Although there have been prior efforts to incorporate art-related questions into evaluation [90], such attempts primarily focused on questions requiring expert-level knowledge, including deducing an artist's name and understanding historical contexts, rather than on direct image comparison. For this task, we collect paintings and their stylistic information from WikiArt. Given one reference painting image and two other paintings as options, the model is tasked with identifying the one that most closely shares the art style of the reference painting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 398, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 495 + ], + "type": "text", + "content": "Object localization: The ability to accurately detect and localize objects is critical for scene understanding. While previous benchmarks [53] have explored this task, their focus was primarily on coarse localization. For instance, they might only ask the model if an object is located at the \"top\" or \"right\" side of an image. BLINK, in contrast, aims for a more fine-grained evaluation. We exploit images from LVIS [36], randomly sampling one object per image along with its ground-truth bounding box. Then we add Gaussian noise to the ground-truth box to create a confounding box. The goal is to select the correct one." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "text", + "content": "Counting: This task evaluate Multimodal LLMs' abilities in detection, recognition, and compositional reasoning, particularly in complex scenes where objects may overlap, be occluded, or vary in size and appearance. We select our questions from the TallyQA dataset [2], known for its challenging human-written counting questions. Each sample comprises an image, a question, and a numerical answer. In addition, we randomly select three numbers to serve as confounding options." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Forensic detection: Recent advances in generative AI have raised concerns about malicious uses and have prompted calls for the automatic detection of fake content. To evaluate whether Multimodal LLMs can fulfill such a role, we construct sets of real and synthesized images that describe similar scenes and ask the models to identify the real ones. Specifically, we first generate synthetic images using Stable Diffusion XL [64], employing COCO captions [47] as prompts. Then, we manually search online using these captions as descriptions and select high-quality photographs as the real images." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BLINK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "IQ test: This task evaluates the ability of Multimodal LLMs to engage in graphical reasoning, without requiring any domain-specific knowledge. We manually collect test samples, along with human explanations, from various public, license-friendly online sources. Given visual examples and a selection of images, the objective is to identify the image that either continues the pattern established by the examples or is spatially consistent with them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 247 + ], + "type": "text", + "content": "Visual similarity: This task aims to verify whether Multimodal LLMs possess a nuanced understanding of visual features, patterns, and aesthetics at a level comparable to humans. We select our samples from the DreamSim dataset [29]. Given a reference image alongside two alternative images, the objective is to identify the image that more closely resembles the reference image perceptually." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 342 + ], + "type": "text", + "content": "Semantic correspondence: This task focuses on identifying and matching semantically similar yet visually distinct elements across images, thereby evaluating the ability of Multimodal LLMs to understand the underlying semantics of object parts. Our samples are sourced from the SPair-71k dataset [61], which features pairs of images with multiple corresponding semantic points. For each task, we randomly select one semantic point in an image as a reference, and provide the matching point alongside three random semantic points in the paired image as options. The objective is to accurately identify the correct matches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 462 + ], + "type": "text", + "content": "Functional correspondence: The task aims to identify points that are functionally similar across objects. It challenges Multimodal LLMs to extend their understanding beyond mere semantics, enabling them to infer the diverse functions an object can perform in various contexts. Such capability is crucial for applications in robotics. We derive our samples from the FunKPoint dataset [43], which features paired images annotated for functional correspondences. Following a method analogous to semantic correspondence, we present an action alongside two object images. One image includes a reference point, while the other offers four potential points. The objective is to select the point that best matches the reference in terms of functional affordances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 463, + 481, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 487 + ], + "type": "text", + "content": "Data quality control: To guarantee the quality of BLINK, we manually go through all collected data and filter out data that are ambiguous." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 506, + 230, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 230, + 519 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 230, + 519 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 531, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 605 + ], + "type": "text", + "content": "In this section, we first describe the experimental setup and the baselines (\\$4.1). Then we present a comprehensive evaluation of 16 recent multimodal LLMs (\\$4.2). We demonstrate that while humans can answer the questions with high accuracy, BLINK is challenging for existing models. Finally, we provide detailed analyses on multiple experimental settings, including the effect of reducing images to captions, sensitivity to different visual prompts, and error analysis (\\$4.3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 621, + 262, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 262, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 262, + 633 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "text", + "content": "Multimodal LLMs: We evaluate BLINK on 16 recent models: MiniGPT-4-v2 [16], OpenFlamingo-v2 [5], InstructBLIP (7B and 13B) [24], CogVLM [80]," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "LLaVA(v1, v1.5, v1.6, internLM, and xtuner versions, model size 7B, 13B, and 34B) [23,26,50-52], Yi-VL (6B and 34B), Qwen-VL-MAX [7], Gemini Pro [73], Claude 3 Opus [1] and GPT-4V(vision) [63]. See Appendix B for more details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 153, + 482, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 153, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 153, + 482, + 237 + ], + "type": "text", + "content": "Evaluation setup: We follow standard setups as in the VLMEvalKit [22], where the temperature is set to 0 and retry is set to 10. However, we do not resize the images during any experiment. For the models that do not support multiple images as input, we concatenate the images as input. We extract the choice from the models' output with a set of pre-defined rules and GPT-3.5-turbo [13]. We refer the readers to Appendix A for more details on visual prompting, how we generate the answers in BLINK, and the human evaluation protocol." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 255, + 228, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 228, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 228, + 266 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": "Overall performance: As shown in Table 1, the mean accuracy of 7B and 13B open-source Multimodal LLMs hover around " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "35 - 42\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": ", which is similar to random guess " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "(38.09\\%)" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": ". The most proficient open-source model, LLaVA-v1.6-34B, achieves an accuracy of " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "45.05\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": ". Even the most advanced models, GPT-4V and Gemini Pro and Claude 3 OPUS, achieve accuracies of only " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "51.26\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "45.72\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "44.11\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": " respectively. Their performance are merely " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "13.17\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "7.63\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "6.02\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": " better than random guessing and lag behind human performance by " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "44.44\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "49.98\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "inline_equation", + "content": "51.59\\%" + }, + { + "bbox": [ + 130, + 276, + 482, + 516 + ], + "type": "text", + "content": ". Notably, for certain tasks such as jigsaw, semantic correspondence, multi-view reasoning, object localization, and relative reflectance, some multimodal LLMs even underperform compared to random guessing. Some qualitative results are shown in Figure 5. Detailed scores on the validation set are in Appendix C. In which tasks do multimodal LLMs show relative strengths and weaknesses? Figure 4 shows the accuracies of the best-performing models on BLINK: LLaVA-v1.6-34B [51], Gemini Pro [73], and GPT-4V [63]. We observe that multimodal LLMs perform relatively better on spatial reasoning, art style, and counting tasks, in which they are much better than random guessing. The models also demonstrate some capability in relative depth and forensics detection. Overall, they are doing relatively well on mid-level perception tasks. In terms of granularity, the models in general perform better on image-level tasks and struggle on pixel-level and crop-level tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": "GPT-4V behaves differently: Figure 4 and Table 1 show an interesting phenomenon: GPT-4V's performance pattern is different from other models. Compared with its counterparts, GPT-4V is much better in visual similarity, art style, jigsaw, and multi-view reasoning. Specifically, its performance on visual similarity is " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": " better than Gemini Pro, demonstrating that GPT-4V possesses a nuanced understanding of visual patterns and aesthetics that is similar to humans. In contrast, Gemini Pro and LLaVA have similar performance patterns. Human performance: Human evaluators achieve over " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": " accuracy across most tasks, with an average accuracy of " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "95.70\\%" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": ". This performance disparity between" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BLINK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 400, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 400, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 400, + 643 + ], + "type": "text", + "content": "7 More details are at the official website at https://www.01.ai/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "Note that the human score for IQ test is annotated by authors. It may not reflect typical human performance, which is also expected to vary." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 114, + 476, + 563 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 476, + 563 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 476, + 563 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 476, + 563 + ], + "type": "table", + "html": "
Validation (1,901)Test (1,906)Similarity (136)Counting (120)Depth (124)Jigsaw (150)Art (117)Fun.Corr. (130)
Random Choice38.0938.09502550505025
Human95.6795.7096.7093.7599.1999.0095.3080.77
Open-source multimodal LLMs
MiniGPT-4-v2 1634.2334.5752.9410.8349.1926.0047.8618.46
OpenFlamingo-v2 539.1838.3255.1521.6754.0346.0052.1436.15
InstructBLIP-7B 2439.7238.6546.3229.1750.8154.0047.8623.85
InstructBLIP-13B 2442.2439.5846.3230.8350.0054.0050.4322.31
LLaVAInternLM2-7B 7437.7136.0652.9452.5052.4234.6730.7723.08
Yi-VL-6B 738.7241.2446.3246.6756.4550.0053.8523.85
Yi-VL-34B 741.6842.7850.0058.3353.2354.0046.1539.23
LLaVA-v1.5-7B-xtuner 2339.3640.8146.3253.3350.8154.0047.8623.85
LLaVA-v1.5-13B-xtuner 2342.0041.3146.3245.0054.0353.3347.8626.15
CogVLM 8041.5439.3846.3238.3350.8152.6749.5723.85
LLaVA-v1.5-7B 4937.1338.0146.3243.3351.6111.3347.8621.54
LLaVA-v1.5-13B 4942.6640.5546.3250.0047.5854.0047.8620.77
LLaVA-v1.6-34B 5146.8045.0546.3268.3364.5256.6747.0130.77
API-based models
Qwen-VL-Max 740.2841.9451.4755.8358.873.3337.6128.46
Gemini Pro 7345.1645.7255.8865.0050.0054.0049.5732.31
Claude 3 OPUS 144.0544.1170.5949.1757.2632.6760.6822.31
GPT-4V(ison) 6351.1451.2683.0960.8358.8762.6778.6331.54
GPT-4 Turbo 6354.6153.8983.0960.8366.9466.0081.2031.54
GPT-4o 6360.0459.0365.4451.6764.5258.0082.9139.23
Sem.Corr. (140)Spatial (143)Local. (125)Vis.Corr. (172)Multi-view (133)Reflect. (134)Forensic (132)IQ (150)
Random Choice255050255033.332525
Human96.0798.2598.0099.4292.4895.14100.0080.00
Open-source multimodal LLMs
MiniGPT-4-v2 1626.4351.7556.0023.8452.6331.3417.4219.33
OpenFlamingo-v2 523.5746.8552.0025.0041.3543.2815.9123.33
InstructBLIP-7B 2425.0055.2444.8022.6758.6529.8529.5523.33
InstructBLIP-13B 2422.8664.3452.0020.9354.1446.2713.6426.00
LLaVAInternLM2-7B 7422.1474.1348.0021.5141.3532.843.7914.67
Yi-VL-6B 726.4372.7349.6029.6548.1229.8520.4523.33
Yi-VL-34B 721.4370.6354.4023.8441.3546.2717.4222.67
LLaVA-v1.5-7B-xtuner 2324.2974.8345.6023.8442.1126.8736.3621.33
LLaVA-v1.5-13B-xtuner 2322.1477.6248.0022.0941.3546.2729.5518.67
CogVLM 8023.5767.1343.2020.9357.1426.8724.2426.67
LLaVA-v1.5-7B 4932.1470.6348.8020.3549.6236.5728.0324.00
LLaVA-v1.5-13B 4923.5767.8347.2020.3541.3545.5227.2728.00
LLaVA-v1.6-34B 5127.8676.2241.6027.3346.6229.8541.6726.00
API-based models
Qwen-VL-Max 729.2977.6249.6022.6753.3849.2547.7322.00
Gemini Pro 7322.1467.1346.4037.2141.3546.2745.4527.33
Claude 3 OPUS 120.7157.3446.4031.4057.8927.6162.1221.33
GPT-4V(ison) 6330.0072.0350.4037.2158.6538.8130.3024.67
GPT-4 Turbo 6332.8667.1348.8042.4457.1434.3351.5230.67
GPT-4o 6345.7176.9256.0071.5160.1538.8185.6130.00
", + "image_path": "3aed257f90f930164ba0c81c7909ac4f0b808f12c171311ece365c442b5d7f63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 563, + 480, + 585 + ], + "lines": [ + { + "bbox": [ + 132, + 563, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 480, + 585 + ], + "type": "text", + "content": "Table 1: Results of different models on the BLINK test set. The first row shows task names and number of test data. Average scores on validation set are also included." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "humans and multimodal LLMs highlights the significant visual perception gap that exists between current machine learning models and humans in perceiving, processing, and understanding complex visual and textual context." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 117, + 271, + 188 + ], + "blocks": [ + { + "bbox": [ + 136, + 117, + 271, + 188 + ], + "lines": [ + { + "bbox": [ + 136, + 117, + 271, + 188 + ], + "spans": [ + { + "bbox": [ + 136, + 117, + 271, + 188 + ], + "type": "image", + "image_path": "3bb53de13f40437cd78c2a10177ee3ef19aad7e7ca0e50653921245457a2c3e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 191, + 275, + 237 + ], + "lines": [ + { + "bbox": [ + 132, + 191, + 275, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 275, + 237 + ], + "type": "text", + "content": "Figure 6: Performance of using image caption + text-only GPT-4 vs. GPT-4 Vision on MMBench [53], MMMU [90], and BLINK (\\$4.3)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 286, + 121, + 480, + 185 + ], + "blocks": [ + { + "bbox": [ + 286, + 121, + 480, + 185 + ], + "lines": [ + { + "bbox": [ + 286, + 121, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 286, + 121, + 480, + 185 + ], + "type": "image", + "image_path": "97243bb6d556d0e1285a0c95cd16652f285abb6f77c05cc2543cc425f2a8db59.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 185, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 283, + 185, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 283, + 185, + 482, + 232 + ], + "type": "text", + "content": "Figure 7: Accuracy of GPT-4V with different visual prompts (e.g., different circle sizes, colors) on relative depth, relative reflectance, and visual correspondence tasks. More discussions in §4.3." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 270, + 203, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 203, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 203, + 282 + ], + "type": "text", + "content": "4.3 Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 449 + ], + "type": "text", + "content": "Is dense captioning all you need for a multimodal LLM benchmark? To answer the question, we reduce multimodal benchmarks to a text-only problem. Specifically, we convert images into task-agnostic dense image captions with GPT-4V. The dense caption describes detailed information about the image and the visual prompts (e.g., where each circle is), using language. For each multimodal question, we prompt the text-only GPT-4-0125-preview model with image captions and the textual question and evaluate if the \"blind\" GPT-4 can answer the question. We call this Caption + LLM. This experiment is predicated on the hypothesis that captioning involves predominantly recognition-centric perception. If using captions along with text-only LLMs yields performance comparable to or surpassing that achieved through the integration of images with multimodal LLMs, then the perception demands of that benchmark are primarily confined to recognition only." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "content": "We experiment with BLINK, MMBench [53] and MMMU [90], as illustrated in Figure 6. Surprisingly, we find that the Caption + LLM setting achieves better results on MMBench than GPT-4V (with " + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "content": " increase in accuracy). On MMMU, Caption + LLM achieves " + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "inline_equation", + "content": "47.2\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "content": " accuracy, which is " + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "content": " lower than GPT-4V performance, but is still much better than random guessing. On BLINK, Caption + LLM fails, achieving random guessing performance. These results indicate that dense captions cover the visual information needed for MMBench. For MMMU, image captions carry a large portion of visual information needed to answer the domain-knowledge-specific questions. Meanwhile, the performance decrease observed in BLINK suggests the necessity for advanced perceptual abilities beyond what is currently attainable with general captions. This variance highlights the limitations of existing multimodal LLM benchmarks in addressing the full spectrum of visual perception." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Effect of visual prompting on BLINK: Several BLINK tasks involve visual prompting. Prior work [69] shows that factors like shape, size, and color may affect task performance, and circles give the best overall performance. Following [69], we adopt circles in BLINK and analyze the effect of circle sizes and colors on multiple tasks in Figure 7. We experiment with relative depth, relative reflectance," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BLINK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 114, + 463, + 183 + ], + "blocks": [ + { + "bbox": [ + 148, + 114, + 463, + 183 + ], + "lines": [ + { + "bbox": [ + 148, + 114, + 463, + 183 + ], + "spans": [ + { + "bbox": [ + 148, + 114, + 463, + 183 + ], + "type": "table", + "html": "
TaskVis.Corr.DepthMulti-viewSem.Corr.ForensicReflect.
Random25.0050.0050.0025.0025.0033.33
Human99.5699.5992.1094.60100.0099.63
Gemini Pro42.4440.3244.3626.6250.7645.52
GPT-4V33.7259.6855.6428.7834.0938.81
SpecialistDIFT 72DepthAnything86LoFTR 70DIFT 72ORDINAL SHADING 14
96.5197.5890.2271.2268.9477.61
", + "image_path": "528c48b84e106d58d4cdeb7c98380ef8230ed55c98fbcd55a1ec39ed268d0a23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 185, + 481, + 206 + ], + "lines": [ + { + "bbox": [ + 130, + 185, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 481, + 206 + ], + "type": "text", + "content": "Table 2: Comparison between multimodal LLMs, specialists, and human performance on the BLINK dev set. The specialists perform much better than multimodal LLMs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 240, + 482, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 336 + ], + "type": "text", + "content": "and visual correspondence, with 100 validation set samples per task. The images are all reshaped to 1024px height. We experiment with circles with 5px, 10px, and 15px radius, and with red or gray color. We find that red is better than gray for all tasks. Also, the optimal circle size is task-dependent. On average 10px circles work the best, and we use it for all evaluations in this paper. The experiments suggest that visual prompting can have a big impact on multimodal LLM performance, and improving visual prompts or improving model robustness to prompt variation is a promising direction for future research [85]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 336, + 483, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 483, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 483, + 456 + ], + "type": "text", + "content": "Can specialist models solve BLINK tasks? Specialists can serve as a proxy upper bound of how good multimodal LLMs could be. We download the trained checkpoints for six specialist models and evaluate them on BLINK. As shown in Table 2, the specialists perform much better than GPT-4V and Gemini Pro, outperforming the best multimodal LLM by " + }, + { + "bbox": [ + 130, + 336, + 483, + 456 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 130, + 336, + 483, + 456 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 336, + 483, + 456 + ], + "type": "inline_equation", + "content": "57\\%" + }, + { + "bbox": [ + 130, + 336, + 483, + 456 + ], + "type": "text", + "content": " on these tasks. Specifically, DepthAnything [86] and DIFT [72] achieve human-level performance on depth estimation and visual correspondence, whereas multimodal LLMs fail miserably. This sheds light on the possibility that multimodal LLMs may progress on these tasks given the correct data and training strategy. For instance, one possible way is to distill existing specialist models into multimodal LLMs [41]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 456, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 564 + ], + "type": "text", + "content": "Error analysis of GPT-4V: We randomly sampled 140 error instances made by GPT-4V on BLINK, 10 per task, and meticulously examined them. The most common types of errors are: Hallucinate fine-grained patterns and attributes (24.2%): the model hallucinates the nuanced details of objects. This error is most common for relative reflectance, forensics detection, and jigsaw tasks. Hallucinate visual prompt locations (20.0%): the circle location described by the model is wrong. This is common for visual correspondence and relative depth tasks. Other errors include Failures on capturing overall setting or style (8.6%), and Failures on grounding an object (5.7%). More details are in Appendix C.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 581, + 220, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 220, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 220, + 594 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We introduced BLINK, a new multimodal LLM benchmark that evaluates core visual perception abilities not found in existing evaluations. While these tasks seem trivial for humans to solve \"within a blink\", we find they pose significant challenges for current multimodal LLMs. Ultimately, Blink provides an effective testbed for multimodal LLMs to catch up with human-level visual perception." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 166 + ], + "type": "text", + "content": "This work was funded in part by ONR Contract N00014-23-1-2417, and supported by NSF grant IIS-2212433." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 198, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 198, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 198, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 211, + 482, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 482, + 234 + ], + "type": "text", + "content": "1. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family (March 2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 482, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 482, + 255 + ], + "type": "text", + "content": "2. Acharya, M., Kafle, K., Kanan, C.: Tallyqa: Answering complex counting questions. In: AAAI (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 482, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 482, + 300 + ], + "type": "text", + "content": "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 482, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 482, + 333 + ], + "type": "text", + "content": "4. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 334, + 482, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 482, + 377 + ], + "type": "text", + "content": "5. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 378, + 482, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 482, + 410 + ], + "type": "text", + "content": "6. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 411, + 482, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 482, + 443 + ], + "type": "text", + "content": "7. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 444, + 482, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 482, + 466 + ], + "type": "text", + "content": "8. Balntas, V., Lenc, K., Vedaldi, A., Mikolajczyk, K.: Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In: CVPR (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 467, + 482, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 482, + 488 + ], + "type": "text", + "content": "9. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 489, + 482, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 482, + 510 + ], + "type": "text", + "content": "0. Bell, S., Bala, K., Snavely, N.: Intrinsic images in the wild. ACM Trans. on Graphics (SIGGRAPH) 33(4) (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 511, + 482, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 482, + 543 + ], + "type": "text", + "content": "1. Berrios, W., Mittal, G., Thrush, T., Kiela, D., Singh, A.: Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 544, + 482, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 482, + 576 + ], + "type": "text", + "content": "2. Black, M.J., Anandan, P.: A framework for the robust estimation of optical flow. In: 1993 (4th) International Conference on Computer Vision. pp. 231-236. IEEE (1993)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 482, + 610 + ], + "type": "text", + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 482, + 632 + ], + "type": "text", + "content": "4. Careaga, C., Aksoy, Y.: Intrinsic image decomposition via ordinal shading. ACM Trans. Graph. (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 482, + 665 + ], + "type": "text", + "content": "5. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: CVPR (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BLINK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "text", + "content": "17. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "text", + "content": "18. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 219, + 481, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 481, + 240 + ], + "type": "text", + "content": "19. Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 241, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 481, + 329 + ], + "type": "text", + "content": "20. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Changpinyo, S., Wu, J., Ruiz, C.R., Goodman, S., Wang, X., Tay, Y., Shakeri, S., Dehghani, M., Salz, D., Lucic, M., Tschannen, M., Nagrani, A., Hu, H., Joshi, M., Pang, B., Montgomery, C., Pietrzyk, P., Ritter, M., Piergiovanni, A., Minderer, M., Pavetic, F., Waters, A., Li, G., Alabdulmohsin, I., Beyer, L., Amelot, J., Lee, K., Steiner, A.P., Li, Y., Keysers, D., Arnab, A., Xu, Y., Rong, K., Kolesnikov, A., Seyedhosseini, M., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali-x: On scaling up a multilingual vision and language model (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 331, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 450 + ], + "type": "text", + "content": "21. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levskaya, A., Ghemawat, S., Dev, S., Michalewski, H., Garcia, X., Misra, V., Robinson, K., Fedus, L., Zhou, D., Ippolito, D., Luan, D., Lim, H., Zoph, B., Spiridonov, A., Sepassi, R., Dohan, D., Agrawal, S., Omernick, M., Dai, A.M., Pillai, T.S., Pellat, M., Lewkowycz, A., Moreira, E., Child, R., Polozov, O., Lee, K., Zhou, Z., Wang, X., Saeta, B., Diaz, M., Firat, O., Catasta, M., Wei, J., Meier-Hellstern, K., Eck, D., Dean, J., Petrov, S., Fiedel, N.: Palm: Scaling language modeling with pathways (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 452, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 481, + 474 + ], + "type": "text", + "content": "22. Contributors, O.: Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "type": "text", + "content": "23. Contributors, X.: Xtuner: A toolkit for efficiently fine-tuning llm. https://github.com/InternLM/xtuner (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "text", + "content": "24. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "text", + "content": "25. DO CT OR OF, P.E.: MACHINE PERCEPTION OF THREE-DIMENSIONAL, SO LIDS. Ph.D. thesis, MASSACHUSETTS INSTITUTE OF TECHNOLOGY (1961)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 565, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 620 + ], + "type": "text", + "content": "26. Dong, X., Zhang, P., Zang, Y., Cao, Y., Wang, B., Ouyang, L., Wei, X., Zhang, S., Duan, H., Cao, M., Zhang, W., Li, Y., Yan, H., Gao, Y., Zhang, X., Li, W., Li, J., Chen, K., He, C., Zhang, X., Qiao, Y., Lin, D., Wang, J.: Internl m-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "27. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "28. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 151, + 482, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 151, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 130, + 151, + 482, + 173 + ], + "type": "text", + "content": "29. Fu, S., Tamir, N., Sundaram, S., Chai, L., Zhang, R., Dekel, T., Isola, P.: Dreamsim: Learning new dimensions of human visual similarity using synthetic data (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 174, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 482, + 205 + ], + "type": "text", + "content": "30. Fu, X., He, M., Lu, Y., Wang, W.Y., Roth, D.: Commonsense-t2i challenge: Can text-to-image generation models understand commonsense? arXiv preprint arXiv:2406.07546 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 482, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 482, + 284 + ], + "type": "text", + "content": "31. Fu, X., Zhang, S., Kwon, G., Perera, P., Zhu, H., Zhang, Y., Li, A.H., Wang, W.Y., Wang, Z., Castelli, V., Ng, P., Roth, D., Xiang, B.: Generate then select: Open-ended visual question answering guided by world knowledge. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Findings of the Association for Computational Linguistics: ACL 2023. pp. 2333–2346. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023-findings-acl.147 https://aclanthology.org/2023-findings-acl.147" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 482, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 482, + 351 + ], + "type": "text", + "content": "32. Fu, X., Zhou, B., Chandratreya, I., Vondrick, C., Roth, D.: There's a time and place for reasoning beyond the image. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 1138-1149. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.81, https://aclanthology.org/2022.acl-long.81" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 352, + 482, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 482, + 373 + ], + "type": "text", + "content": "33. Fu, X., Zhou, B., Chen, S., Yatskar, M., Roth, D.: Interpretable by design visual question answering. arXiv preprint arXiv:2305.14882 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 374, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 482, + 407 + ], + "type": "text", + "content": "34. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 407, + 482, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 482, + 451 + ], + "type": "text", + "content": "35. Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., Manocha, D., Zhou, T.: Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 452, + 482, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 482, + 485 + ], + "type": "text", + "content": "36. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 486, + 482, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 482, + 508 + ], + "type": "text", + "content": "37. Harris, C., Stephens, M., et al.: A combined corner and edge detector. In: Alvey vision conference. vol. 15, pp. 10-5244. Citeseer (1988)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 509, + 482, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 482, + 531 + ], + "type": "text", + "content": "38. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 532, + 482, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 482, + 553 + ], + "type": "text", + "content": "39. Hu, Y., Hua, H., Yang, Z., Shi, W., Smith, N.A., Luo, J.: Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 482, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 482, + 586 + ], + "type": "text", + "content": "40. Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 587, + 482, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 482, + 620 + ], + "type": "text", + "content": "41. Hu, Y., Stretcu, O., Lu, C.T., Viswanathan, K., Hata, K., Luo, E., Krishna, R., Fuxman, A.: Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. arXiv preprint arXiv:2312.03052 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 482, + 665 + ], + "type": "text", + "content": "42. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BLINK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 149 + ], + "type": "text", + "content": "43. Lai, Z., Purushwalkam, S., Gupta, A.: The functional correspondence problem. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15772-15781 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 482, + 182 + ], + "type": "text", + "content": "44. Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 482, + 214 + ], + "type": "text", + "content": "45. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 482, + 246 + ], + "type": "text", + "content": "46. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 482, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 482, + 290 + ], + "type": "text", + "content": "47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 482, + 312 + ], + "type": "text", + "content": "48. Liu, F., Emerson, G., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics 11, 635-651 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 312, + 482, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 482, + 333 + ], + "type": "text", + "content": "49. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 333, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 482, + 354 + ], + "type": "text", + "content": "50. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 482, + 386 + ], + "type": "text", + "content": "51. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 482, + 407 + ], + "type": "text", + "content": "52. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 482, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 482, + 439 + ], + "type": "text", + "content": "53. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., Chen, K., Lin, D.: Mmbench: Is your multi-modal model an all-around player? (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 439, + 482, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 482, + 472 + ], + "type": "text", + "content": "54. Liu, Y., Li, Z., Li, H., Yu, W., Huang, M., Peng, D., Liu, M., Chen, M., Li, C., Jin, L., et al.: On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 472, + 482, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 482, + 493 + ], + "type": "text", + "content": "55. Liu, Y., Li, Z., Yang, B., Li, C., Yin, X., lin Liu, C., Jin, L., Bai, X.: On the hidden mystery ofOCR in large multimodal models (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 493, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 482, + 525 + ], + "type": "text", + "content": "56. Lowe, D.G.: Object recognition from local scale-invariant features. In: Proceedings of the seventh IEEE international conference on computer vision. vol. 2, pp. 1150-1157. IEEE (1999)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 482, + 557 + ], + "type": "text", + "content": "57. Lu, J., Clark, C., Lee, S., Zhang, Z., Khosla, S., Marten, R., Hoiem, D., Kembhavi, A.: Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 557, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 482, + 590 + ], + "type": "text", + "content": "58. Lu, P., Bansal, H., Xia, T., Liu, J., Li, C., Hajishirzi, H., Cheng, H., Chang, K.W., Galley, M., Gao, J.: Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 590, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 482, + 611 + ], + "type": "text", + "content": "59. Marr, D.: Vision: A computational investigation into the human representation and processing of visual information. MIT press (2010)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "type": "text", + "content": "60. Marr, D., Poggio, T.: Cooperative computation of stereo disparity: A cooperative algorithm is derived for extracting disparity information from stereo image pairs. Science 194(4262), 283-287 (1976)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 482, + 665 + ], + "type": "text", + "content": "61. Min, J., Lee, J., Ponce, J., Cho, M.: Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543 (2019)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "62. Minsky, M., Papert, S.: An introduction to computational geometry. Cambridge, tiass., HIT 479(480), 104 (1969)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 307, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 307, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 307, + 150 + ], + "type": "text", + "content": "63. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 152, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 152, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 152, + 481, + 184 + ], + "type": "text", + "content": "64. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 185, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 228 + ], + "type": "text", + "content": "65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 481, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 273 + ], + "type": "text", + "content": "66. Sarkar, A., Mai, H., Mahapatra, A., Lazebnik, S., Bhattachad, A.: Shadows don't lie and lines can't bend! generative models don't know projective geometry... for now. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28140-28149 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "text", + "content": "67. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "type": "text", + "content": "68. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "text", + "content": "69. Shtedritski, A., Rupprecht, C., Vedaldi, A.: What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 384 + ], + "type": "text", + "content": "70. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. CVPR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "71. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 409, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 430 + ], + "type": "text", + "content": "72. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 431, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 464 + ], + "type": "text", + "content": "73. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 465, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 481, + 486 + ], + "type": "text", + "content": "74. Team, I.: InternlM: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 487, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 509 + ], + "type": "text", + "content": "75. Team, M.N.: Introducing mpt-7b: A new standard for open-source, commercially usable lms (2023), www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-05" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 510, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 531 + ], + "type": "text", + "content": "76. Torralba, A., Oliva, A.: Depth estimation from image structure. IEEE Transactions on pattern analysis and machine intelligence 24(9), 1226-1238 (2002)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 533, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 564 + ], + "type": "text", + "content": "77. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 566, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 481, + 597 + ], + "type": "text", + "content": "78. Wang, F., Fu, X., Huang, J.Y., Li, Z., Liu, Q., Liu, X., Ma, M.D., Xu, N., Zhou, W., Zhang, K., et al.: Muirbench: A comprehensive benchmark for robust multi-image understanding. arXiv preprint arXiv:2406.09411 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "text", + "content": "79. Wang, J.Y., Adelson, E.H.: Layered representation for motion analysis. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. pp. 361-366. IEEE (1993)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "80. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., Xu, J., Xu, B., Li, J., Dong, Y., Ding, M., Tang, J.: Cogvlm: Visual expert for pretrained language models (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BLINK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 534 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "81. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "text", + "content": "82. Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., Li, H.: Dire for diffusion-generated image detection. arXiv preprint arXiv:2303.09295 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "type": "text", + "content": "83. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., Zhou, D.: Chain-of-thought prompting elicits reasoning in large language models (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "type": "text", + "content": "84. Yan, A., Yang, Z., Wu, J., Zhu, W., Yang, J., Li, L., Lin, K., Wang, J., McAuley, J., Gao, J., et al.: List items one by one: A new data source and learning paradigm for multimodal llms. arXiv preprint arXiv:2404.16375 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 482, + 259 + ], + "type": "text", + "content": "85. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 482, + 282 + ], + "type": "text", + "content": "86. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 482, + 315 + ], + "type": "text", + "content": "87. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3081-3089 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 482, + 346 + ], + "type": "text", + "content": "88. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 346, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 482, + 380 + ], + "type": "text", + "content": "89. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 482, + 435 + ], + "type": "text", + "content": "90. Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W., Sun, Y., Wei, C., Yu, B., Yuan, R., Sun, R., Yin, M., Zheng, B., Yang, Z., Liu, Y., Huang, W., Sun, H., Su, Y., Chen, W.: Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 482, + 468 + ], + "type": "text", + "content": "91. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems 35, 27469-27483 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 482, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 482, + 500 + ], + "type": "text", + "content": "92. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 500, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 482, + 534 + ], + "type": "text", + "content": "93. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "Fu and Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BRAVE_ Broadening the visual encoding of vision-language models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_content_list.json b/2024/BRAVE_ Broadening the visual encoding of vision-language models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..441876f27ac66407046fe3fa074df0283d65160a --- /dev/null +++ b/2024/BRAVE_ Broadening the visual encoding of vision-language models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_content_list.json @@ -0,0 +1,1712 @@ +[ + { + "type": "text", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "text_level": 1, + "bbox": [ + 235, + 138, + 769, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Oğuzhan Fatih Kar1,2, Alessio Tonioni1, Petra Poklukar1, Achin Kulshrestha1, Amir Zamir2, and Federico Tombari1,3", + "bbox": [ + 251, + 212, + 750, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Google $^{2}$ Swiss Federal Institute of Technology Lausanne (EPFL) $^{3}$ TU Munich https://brave-vlms.epfl.ch", + "bbox": [ + 218, + 253, + 782, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Improving visual capabilities of VLMs with BRAVE", + "text_level": 1, + "bbox": [ + 218, + 338, + 486, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/275a45e841d9c5e8a9bd0345cc017349e824ce91657b6e6f4df78ba4013d30ac.jpg", + "image_caption": [ + "Is there a hand using the mouse in this image?", + "(a) Yes" + ], + "image_footnote": [], + "bbox": [ + 225, + 368, + 323, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cc9d2159ed7352bcc7782dfe32972a6beed12bfbc71df98d0e85c5d0fd1cdfb.jpg", + "image_caption": [ + "Are the butterfly's feet visible?", + "(a) Yes" + ], + "image_footnote": [], + "bbox": [ + 328, + 368, + 426, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8140a21dcf553b1758543b938e1149a60334dd3dcc11687e31f71664bd2d248e.jpg", + "image_caption": [ + "Is the door of the truck open?", + "(a) Yes", + "(b) No" + ], + "image_footnote": [], + "bbox": [ + 431, + 368, + 478, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf4e78dc2116e5408bee825b78eba03968f4cb69291d468c78e18adc9b7c42ac.jpg", + "image_caption": [ + "InstructBLIP (a) (b) (EVA Encoder)", + "BRAVE (a) (b) (Multiple Encoders)" + ], + "image_footnote": [], + "bbox": [ + 218, + 446, + 321, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b117e0eed6c3be4ce2226d4bd53a70bea0295ca6f45c8ca4652199a5c775d819.jpg", + "image_caption": [ + "(a) (b)" + ], + "image_footnote": [], + "bbox": [ + 379, + 422, + 413, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f989f58577877184e44fa199a8243811051743e678ae65762159580486d2004b.jpg", + "image_caption": [ + "(a) (a)" + ], + "image_footnote": [], + "bbox": [ + 470, + 422, + 508, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ff5d117d7eeb0caabe9a75c7786d2668c66042e25e0c051be06ca598ae1e526.jpg", + "image_caption": [ + "Fig. 1: We propose BRAVE to broaden the visual capabilities of vision-language models (VLMs). Left: In contrast to existing methods, e.g. InstructBLIP [25] or LLaVA-1.5 [64], that use a single vision encoder [59,84], BRAVE combines diverse features from multiple vision encoders into a more versatile and compact representation. The examples are taken from [84] and assess the VLM's ability to differentiate images with visual differences. Right: BRAVE leads to state-of-the-art performance on a wide range of captioning and visual question answering tasks. Furthermore, it significantly improves the performance on benchmarks [84] where commonly employed vision encoders fail.", + "SoTA Performance in Captioning & VQA" + ], + "image_footnote": [], + "bbox": [ + 535, + 351, + 785, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. \"blindness\" to certain image features, visual hallucination, etc. To address these issues, we study broadening the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features", + "bbox": [ + 259, + 672, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs.", + "bbox": [ + 259, + 147, + 743, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Vision-Language $\\cdot$ Visual Encoding $\\cdot$ Multi-modal Learning", + "bbox": [ + 261, + 271, + 741, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 310, + 375, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) have recently seen significant improvements on solving tasks requiring both visual and text understanding capabilities such as captioning, visual question answering (VQA), and instruction following. These advancements are fueled by the progress in uni-modal vision encoders [30, 75] and language models (LMs) [22, 24], which are then combined with different \"bridging\" techniques [3,56,66] to create VLMs. Improved training recipes [25,64] and scalability efforts in terms of training data and model size [6,19,77,87,99] further boost the performance.", + "bbox": [ + 212, + 340, + 785, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the progress, VLMs are subject to several shortcomings: on the language side, the LMs are known to be susceptible to hallucinations and logical faults [8,37,79,81], while on the vision side, they are limited by the capabilities of the vision encoder. For example, Tong et al. [84] observed that commonly employed vision encoders such as CLIP [75] can exhibit \"blindness\", i.e. fail to differentiate images with clear visual differences (See Fig. 1, left). Similarly, Li et al. [59] showed that VLMs are prone to visual hallucinations where the models imagine incorrect details about a given image. These limitations create a bottleneck for developing performant and visually grounded VLMs.", + "bbox": [ + 212, + 462, + 785, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these, we study broadening the visual encoding capabilities of VLMs. It is known in machine learning that different representations of an input can yield different generalization properties, and each can play a role when ensembled to create a more complete representation [33]. Inspired by this, we first perform a comprehensive evaluation of VLMs which differ only in their vision encoders, i.e. visual biases. For this, we consider encoders with different biases, e.g. different objectives, training data, and model sizes. Our results in Table 2 show that different encoders lead to varying performance across vision-language tasks and there is no single encoder achieving a consistent top performance across tasks. Furthermore, we find that encoders with different biases can perform surprisingly similarly, suggesting that there could be multiple cues [35,45] that can be leveraged to solve the vision-language tasks.", + "bbox": [ + 212, + 598, + 785, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these findings, we propose to employ various vision encoders for VLMs and introduce a method to learn how to combine them efficiently. We denote the method as BRAVE, which stands for broadening the visual encoding of VLMs. As shown in Fig. 2, BRAVE combines features from an arbitrary number", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of vision encoders into a compressed fixed-length visual representation which is then fed as a soft visual prompt to a frozen LM. We achieve the bridging of encoders and the LM by introducing a lightweight multi-encoder querying transformer (MEQ-Former), which is a multi-encoder generalization of the Q-Former proposed in the single encoder BLIP-2 framework [56]. In particular, MEQ-Former accepts a textual prompt and a set of learnable queries as inputs and jointly cross-attends to the features from different vision encoders (Fig. 2). We provide a simple yet effective recipe in Sec. 3 for training the MEQ-Former in a single stage that bypasses the two-stage pre-training paradigm of [56]. Notably, this is achieved with a smaller number of trainable parameters during pre-training than the existing methods [3,6,19,25,61,64,87] (Sec. 4).", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our evaluation on a wide range of captioning and VQA tasks shows that BRAVE effectively consolidates diverse visual signals into a broad and contextual representation, leading to consistently better performance over the state-of-the-art and improved robustness against out-of-distribution inputs, as shown in Fig. 1 and Sec. 4. Furthermore, in contrast to previous works that benefit from scaling the language axis [3, 19, 25, 64] by using larger LMs, our work demonstrates that scaling along the vision axis has a significant potential for VLMs.", + "bbox": [ + 212, + 313, + 787, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 238, + 417, + 593, + 431 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct a systematic analysis of several vision encoders with different inductive biases, e.g. training data, objective, model size, on solving vision-language tasks under a unified training and evaluation framework.", + "- We introduce BRAVE, a method that efficiently consolidates features from any number of vision encoders into a compressed and contextual representation, leading to 1) state-of-the-art performance on several captioning and VQA tasks and 2) significantly improved robustness against visual hallucinations and out-of-distribution inputs.", + "- We perform a comprehensive ablation study of BRAVE, highlighting the impact of the design choices, which we hope provide useful insights to the researchers for further advancements in VLMs." + ], + "bbox": [ + 225, + 445, + 784, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Impact of vision encoders for vision-language models", + "text_level": 1, + "bbox": [ + 214, + 633, + 763, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To quantify the impact of visual biases on the performance of VLMs, we compare VLMs with different vision encoders on commonly evaluated VQA and captioning tasks. For this, we develop a pre-training, fine-tuning and evaluation setup, as explained next. To the best of our knowledge, this is the first unified and comprehensive evaluation of different vision encoders for vision-language understanding and generation tasks.", + "bbox": [ + 212, + 666, + 787, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLM architecture. We use a frozen vision encoder which is connected to a frozen language model by a bridge network with trainable parameters. For the bridge, we adopt the Q-Former module from [56]. This choice is based on the following reasons: 1) the Q-Former resamples visual features to a fixed-length", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/f0e7e9e66090fed507e442d144606fee7f7c6943c0045c3f52d31295f93e20ce.jpg", + "table_caption": [ + "Table 1: Overview of vision encoders we benchmarked. All encoders use a ViT [27] backbone, yet they differ in terms of objective, training data, and model size. ITC: Image-text contrastive learning [75], MIM: Masked image modelling [41], LGC: Local-to-global correspondence learning [11]. See Sec. 2 for details." + ], + "table_footnote": [], + "table_body": "
Vision encoderParametersTraining dataObjective
CLIP-L/14 [75]0.3BOpenAI WITITC
OpenCLIP-G/14 [21]1.8BLAION-2BITC
EVA-CLIP-g [30]1BLAION-400MMIM (CLIP features)
SIGLIP-G/14 [100]1.8BWebLIITC (Sigmoid loss)
SILC-G/16 [73]1.8BWebLIITC + LGC
ViT-e [19]3.8BJFT-3BClassification
ViT-G [99]1.8BJFT-3BClassification
DINOv2-L/14 [74]0.3BLVD-142MLGC + iBOT [105]
", + "bbox": [ + 271, + 212, + 727, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "output before feeding them to the LM which enables a fair comparison of vision encoders with different output dimensionalities, 2) the Q-Former is efficient to train and evaluate as it is compatible with frozen pre-trained vision encoders and LMs. For the vision encoder, we consider a diverse set of options as described in the next paragraph. For the LM, we follow [56] and use FlanT5-XL [24]. The Q-Former has 110M parameters consisting of a series of cross- and self-attention layers and a fixed number of learnable queries. Unless specified otherwise, we use 32 queries of dimension 768 which is also the hidden dimension of the Q-Former. The queries interact with visual features through cross-attention layers, enabling a significant reduction in their dimensionality. For example, CLIP-L/14 [75] visual features are reduced from $257 \\times 1024$ to $32 \\times 768$ . The output of the Q-Former is then linearly projected to the textual embedding space of the LM to create a soft visual prompt, and then prepended to the textual prompt that specifies the task. Both visual and textual prompts are fed as inputs to the frozen LM. For captioning, we use the textual prompt \"A photo of\" and for VQA, we directly use the question as the prompt. Similar to InstructBLIP [25], we give the textual prompt as additional input to the Q-Former together with the learnable queries to extract more task-aligned visual prompts.", + "bbox": [ + 212, + 376, + 787, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision encoders. As summarized in Table 1, we consider eight recently introduced vision encoders: CLIP [75], OpenCLIP [21], EVA-CLIP [30], SIGLIP [100], SILC [73], ViT-e [19], ViT-G [99], and DINOv2 [74]. While they all use ViT-based backbones [27], they differ in terms of 1) training data, e.g. LAION [77], WebLI [19], LVD-142M [74], etc., 2) training objective, e.g. image-text contrastive learning, masked image modelling, classification, etc., and 3) model size, e.g. 300M to 4B parameters. Due to this diversity, they incorporate different biases and thus potentially capture different aspects of the underlying depicted scene.", + "bbox": [ + 212, + 667, + 787, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-training data and objectives. We pre-train the Q-Former using WebLI [19] dataset at $224 \\times 224$ resolution. We use the filtered and de-duplicated", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/43ed85e5c7189e1753bdeacda3934b93642cb4317c8ccdccf001cf052a6eb211.jpg", + "table_caption": [ + "Table 2: Benchmarking vision encoders on different vision-language tasks. For COCO captioning, we report CIDEr [86] score. For VQA tasks, we report top-1 accuracy. For MMVP, we report average pair accuracy [84]. Top-3 best results are shown with Blue, Green, and Orange, respectively. All VLMs use FlanT5-XL [24] as the language model and a Q-Former [56] to bridge vision and language modalities. See Sec. 2 for details." + ], + "table_footnote": [], + "table_body": "
Vision encoderCOCO cap. ↑ Karpathy valVQAv2 ↑ Karpathy valOKVQA ↑ valGQA ↑ test-devMMVP ↑ test
CLIP-L/14133.074.461.048.715.3
OpenCLIP-G128.373.360.648.022.0
EVA-CLIP-g140.977.063.050.127.3
SIGLIP-G/14133.074.762.548.624.0
SILC-G/16141.177.063.449.724.0
ViT-e137.875.661.949.125.3
ViT-G133.874.261.248.320.7
DINOv2-L/14127.671.359.048.022.0
", + "bbox": [ + 272, + 243, + 733, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "English-only subset with 100 million image-text pairs [19, 73]. We keep both the vision encoder and the LM frozen and only train the parameters of the Q-Former. The model is trained with captioning objective using the alt-text in WebLI as target, as it has been shown to be effective [92]. Similar to [56], we remove the last Transformer layer of the encoder. We refer the reader to the supplementary material for additional training details and hyperparameters.", + "bbox": [ + 212, + 424, + 787, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation tasks. We evaluate the obtained VLMs on standard captioning and VQA tasks. For captioning, we use the COCO captioning benchmark [20] and fine-tune the pre-trained VLM on the commonly employed Karpathy training split [53]. Similar to the pre-training, only the Q-Former parameters are updated. For VQA, we again follow the standard practices [19, 25, 56, 64] and fine-tune on a mixture of VQA data that enforces the Q-Former to extract more task-aligned visual features. Our mixture includes the training sets of VQAv2 [36] and OKVQA [70] as well as the synthetically generated VQA data from $\\mathrm{VQ^2A}$ [12], resulting in total 17M training examples. We then evaluate the fine-tuned models on VQAv2 and OKVQA validation sets. Similar to [25, 56], we also evaluate the VLMs' zero-shot VQA capabilities on GQA [48]. Finally, we evaluate the zero-shot performance on MMVP [84] that includes images with semantic differences that are challenging for the current VLMs to handle. Unlike captioning, we fine-tune both Q-Former and the LM parameters while keeping the vision encoder frozen, as we find this to improve the performance, similar to [64]. For both captioning and VQA tasks, we use $224\\times 224$ image resolution. See the supplementary for more details and results.", + "bbox": [ + 212, + 541, + 787, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. We report the findings in Table 2 and make the following observations:", + "bbox": [ + 214, + 824, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Encoders with different biases can obtain similar performance. While we observe some variation in VLMs' performance across different encoders, the gap remains small, especially among the top-performing ones. For example, EVA-CLIP-g, SILC-G/16, and ViT-e achieve the best results for VQAv2, GQA and COCO captioning despite their differences in training objective and data, as shown in Table 1. Similar observations can be made for OpenCLIP and DINOv2 encoders that have a small performance gap.", + "- MMVP stays challenging for all encoders. While it is originally curated with \"CLIP-blind\" pairs in [84], i.e. images that CLIP perceives as similar despite their visual differences (See Fig. 1 for examples), most of the encoders stay below the random guess accuracy of $25\\%$ . Exceptions are EVA-CLIP and ViT-e which slightly surpass the random guess level ( $27.3\\%$ and $25.3\\%$ ).", + "- For tasks requiring compositional reasoning and open-world knowledge, the performance decreases and the gap among VLMs diminishes, as shown by the results on GQA and OKVQA. More precisely, the standard deviation across all the VLMs' performance on GQA and OKVQA is 0.8 and 1.36, respectively, but 1.74 on VQAv2 and 4.91 on COCO.", + "- Scaling the size of vision encoder can improve the performance as seen by the consistent improvements obtained by ViT-e over ViT-G when scaling from 1.8B to 3.8B parameters.", + "- Pre-training data distribution can play a critical role. OpenCLIP-G/14 model is larger than CLIP-L/14 and uses more training data, yet it noticeably underperforms on most of the evaluated VQA and captioning tasks except the MMVP benchmark, where OpenCLIP significantly outperforms the CLIP model. Since both models are trained using the same training objective, this indicates that the impact of the training objective and dataset to the performance of the VLM can be disentangled, and both are important.", + "- Performance correlates between VQA and captioning tasks. Encoders with better performance on VQA tasks generally also perform better on captioning, and vice versa. Spearman's rank correlation between VQA tasks and COCO captioning is 0.90, 0.80, 0.79, 0.56 for VQAv2, OK-VQA, GQA, and MMVP, respectively. Thus, improving the visual component in VLMs is expected to improve performance for a broad range of tasks." + ], + "bbox": [ + 225, + 146, + 785, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above results highlight the importance of using the right vision encoder for VLMs, motivating us to focus on the following question: Can we broaden the vision capabilities of VLMs through combining vision encoders with different visual biases? We answer positively in the next section.", + "bbox": [ + 215, + 651, + 787, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Combining vision encoders with BRAVE", + "text_level": 1, + "bbox": [ + 215, + 733, + 620, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motivated by the findings in Sec. 2, we propose using multiple vision encoders with different, and potentially complementary, visual biases to create more capable VLMs. To do this, we introduce BRAVE, a method that combines the strengths of different vision encoders while staying efficient in terms of number of trainable parameters. Below, we provide details of BRAVE.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db020df293b626c30f01d0e6866cc2becb612f2e138bec09372ce27aecc1fc3f.jpg", + "image_caption": [ + "Fig. 2: Overview of BRAVE. Left: We keep all the vision encoders (VEs) and the language model (LM) frozen. The linear projection layers are used to concatenate features from $K$ different VEs, e.g. $K = 5$ , sequence-wise. These are then resampled by the MEQ-Former which accepts a set of learnable queries and a text prompt describing the task as inputs. The output of MEQ-Former is projected to the input space of the LM using fully-connected (FC) layers. The total number of trainable parameters is 116M ( $\\approx 1\\%$ of the total parameters). Right: Architecture of the MEQ-Former with $N = 12$ transformer layers. It interacts with the concatenated features through crossattention layers and produces a fixed-length output to be fed as input to the LM." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 553, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f32267e7f2c26754790dc029424ed0747f84db25d18d015b9c239c6ff803703.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 148, + 777, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overview. As depicted in Fig. 2, we introduce the multi-encoder querying transformer, or MEQ-Former, that combines visual features from an arbitrary set of encoders and outputs a fixed-length visual representation that is given as a soft prompt to a frozen LM. It takes as an input a sequence comprised of a fixed number of learnable queries and embedded text tokens which describe the task - for captioning, they are the textual prompt \"A photo of\", while for VQA, they are the question prompt for a given image (See supplementary for examples). MEQ-Former interacts with visual features through cross-attention layers. The visual features from different encoders are first linearly projected to the same dimension and then concatenated sequence-wise. The obtained concatenated feature is given as a key and value pair to the cross-attention layers in MEQ-Former, and is cross-attended to by the MEQ-Former's query sequence (Fig. 2). This resampling enables to efficiently processing a large number of visual features since it bypasses the quadratic complexity of self-attention. It also acts as a \"bottleneck\" that keeps the total number of VLM's parameters low compared to a naive ensembling of VLMs that is prohibitively expensive (See a comparison in Sec. 4.3). Moreover, the visual features are not \"marked\" with any encoder-specific embedding, thus MEQ-Former is not biased to differentiate between encoders, allowing for a simpler design.", + "bbox": [ + 212, + 487, + 787, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Differences with previous work. Our work differs from other popular approaches like LLaVa [64] and PaLI [19] that require significantly more trainable parameters ( $\\approx 10\\mathrm{B}$ ) and do not have a resampling mechanism. Resampling visual", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "features from a single encoder has been successfully demonstrated in previous work for single- [25,56] and multiple-frames [3,55,97]. To the best of our knowledge, our work is the first one demonstrating an effective resampling mechanism to consolidate visual features from several arbitrarily different encoders while keeping the number of trainable parameters small (our MEQ-Former having 116M parameters vs BLIP-2's Q-Former is 188M). See Sec. 5 for more comparisons.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pre-training details. We use the same data and objective as in Sec. 2 and only train the MEQ-Former while keeping the vision encoders and the LM frozen. During pre-training, we randomly mask features from each encoder with $20\\%$ probability as we found that this can act as a regularizer and enforces MEQ-Former to avoid local minima by attending only to the features of a single encoder (Ablated in Sec. 4.3). After pre-training, we fine-tune the MEQ-Former and (optionally) the LM on the downstream tasks, as explained in Sec. 4.", + "bbox": [ + 212, + 258, + 789, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architecture. For the main results, we combine five vision encoders, namely EVA-CLIP- $g$ [30], CLIP- $L/14$ [75], SILC- $G/16$ [73], ViT- $e$ [19], and DINOv2- $L/14$ [74] to cover all training datasets and objectives from Table 1. Please see Sec. 4.3 and the supplementary for analysis on the contributions of encoders. For the LM, we employ FlanT5-XL [24] as before. We set the visual feature dimension after linear projection as 1408 for concatenation. For MEQ-Former, we use $32 \\times 5 = 160$ learnable queries to proportionally scale the model capacity to the number of vision encoders and set the hidden dimension as 768. The MEQ-Former resampling results in $14 \\times$ reduction in total feature size (from $1223 \\times 1408$ to $160 \\times 768$ ). The total number of trainable parameters during pre-training is 116M which is about $1\\%$ of the total number of parameters in the VLM (10.3B).", + "bbox": [ + 212, + 387, + 789, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation. We use FLAX [42] with the Scenic [26] library to implement our models and training pipeline. For evaluations, we use both Scenic and EvalAI API [93]. We pre-train with batch size of 1024 on 64 TPUv5 chips.", + "bbox": [ + 212, + 574, + 787, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 642, + 377, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform evaluations on a broad range of captioning (Sec. 4.1) and VQA tasks (Sec. 4.2) to show the effectiveness of BRAVE. Please see Fig. 1 right for a summary of the results and supplementary for an overview of the tasks. We then provide a comprehensive analysis of BRAVE in Sec. 4.3.", + "bbox": [ + 212, + 672, + 787, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For captioning, we evaluate on COCO [20] and NoCaps [2] benchmarks. The latter includes out-of-distribution samples with novel classes that do not exist in the COCO benchmark, hence it serves as a challenging testbed to assess robustness. The results (Table 3) show that BRAVE pushes the state-of-the-art for both benchmarks while staying efficient in terms of trainable parameters.", + "bbox": [ + 212, + 734, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For VQA, we evaluate on VQAv2 [36], OKVQA [70], GQA [48], and VizWiz-QA [39] benchmarks. To assess the robustness of VLMs, we additionally include", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6172d64b7d8a4ded4e57e27e6fe71d7b1bc8775babe58dbafeb23849044c9057.jpg", + "table_caption": [ + "Table 3: Captioning. We compare BRAVE with state-of-the-art captioning methods on COCO and NoCaps (zero-shot). CIDEr [86] score is reported (Best in bold, second best is underlined). Numbers of other methods are taken from the corresponding publications. Dashed line means no result was reported. BRAVE achieves the best results in NoCaps evaluation sets (both out-domain and overall in validation & test) and second best for COCO, while having much fewer trainable parameters than other methods. For example, compared to PaLI-17B [19], BRAVE uses significantly fewer trainable parameters, training data, and image pixels. Yet, it outperforms the baselines on NoCaps and stays competitive on COCO, demonstrating the effectiveness of having different visual biases for generalization. See Sec. 4.1 for more details." + ], + "table_footnote": [], + "table_body": "
Method# paramsCOCO (fine-tuned)NoCaps (zero-shot, val)NoCaps (zero-shot, test)
TrainableTotalKarpathy testout-domainoverallout-domainoverall
Flamingo [3]10.6B80B138.1----
SimVLM [91]632M632M143.3113.7112.2-110.3
Qwen-VL [6]9.6B9.6B--121.4--
BLIP-2 [56]1.1B4.1B144.5124.8121.6--
InstructBLIP [25]188M14.2B--121.9--
CoCa [96]2.1B2.1B143.6-122.4-120.6
GiT2 [87]5.1B5.1B145.0130.6126.9122.3124.8
PaLI-17B [19]16.9B16.9B149.1-127.0126.7124.4
BRAVE116M10.3B148.0133.3127.6127.1125.6
", + "bbox": [ + 248, + 295, + 754, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "POPE [59] that measures object hallucinations, and MMVP [84] which exposes the shortcomings of CLIP-based vision encoders that are widely used in VLMs. Our results (Table 4) show that BRAVE achieves consistent improvements over recent methods for all the tasks by leveraging diverse visual features.", + "bbox": [ + 212, + 448, + 787, + 510 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Image captioning", + "text_level": 1, + "bbox": [ + 214, + 532, + 405, + 549 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fine-tuning details. Similar to Sec. 2, we only fine-tune the MEQ-Former on COCO captioning and keep the vision encoders and the LM frozen. During fine-tuning, we set a higher input resolution as $336 \\times 336$ to further boost the performance, similar to previous works [19,56,64] (See Sec. 4.3 for an ablation). Following [19,25,56,87,96], we also zero-shot evaluated the COCO fine-tuned model on NoCaps [2]. Please see the supplementary for more training details.", + "bbox": [ + 212, + 559, + 787, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results. We summarize the captioning evaluations in Table 3. BRAVE uses the least number of trainable parameters (116M) yet achieves strong results for both COCO and NoCaps benchmarks. For NoCaps, BRAVE is the best performing method with significant gains over recent methods. This is especially the case for out-domain samples with novel classes, demonstrating the usefulness of diversity in visual features for robustness. For COCO, BRAVE stays competitive with the best performing model, PaLI-17B [19]. Notably, BRAVE achieves this while using $150 \\times$ fewer trainable parameters (116M vs 16.9B), $16 \\times$ less pre-training data (100M vs 1.6B) and $3 \\times$ fewer image pixels ( $336 \\times 336$ vs $588 \\times 588$ ) than PaLI-17B, suggesting that having different visual biases is effective for generalization, while keeping the sample complexity low.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Visual question answering. We compare BRAVE with state-of-the-art VQA methods on VQAv2, OKVQA, GQA, VizWiz-QA, MMVP, and POPE. Top-1 accuracies are reported (Best in bold, second best is underlined). For MMVP, we report average pair accuracy [84]. Numbers of other methods are taken from the corresponding publications. Dashed line means no result was reported. BRAVE achieves the best results for six out of seven benchmarks, while staying efficient in terms of trainable parameters. Similar to the COCO captioning results in Table 3, BRAVE stays competitive with PaLI-17B [19] for VQAv2, while using significantly less pre-training data, image resolution, and model parameters (both trainable and total). See Sec. 4.2 for details.", + "bbox": [ + 212, + 143, + 787, + 270 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a2474afb0011e5310d7878e0f96accbdcbc2fc9f1fc0696d0c6755a8eecb0584.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# paramsFine-tunedZero-shot
TrainableTotalVQAv2 test-devOKVQA valGQA test-devVizWiz-QA test-devGQA test-devMMVP testPOPE test
SimVLM [91]632M632M80.0------
Flamingo [3]10.2B80B82.057.8-31.6---
MiniGPT-v2 [14]7B8B-57.860.153.6---
GiT2 [87]5.1B5.1B81.7------
Qwen-VL [6]9.6B9.6B79.558.659.335.2---
SPHINX-2k [61]13B16.5B80.762.663.144.9--87.2
PaLI-17B [19]16.9B16.9B84.364.5-----
BLIP-2 [56]1.2B12.1B81.654.7-29.444.7-85.3
InstructBLIP [25]188M14.2B-55.5-33.449.516.778.9
ShareGPT4V [16]13.4B13.4B81.0-64.8----
LLaVa1.5 [64]13B13.4B80.0-63.353.6-24.785.9
LLaVa1.6 [65]13B13.4B-46.365.4---86.3
LLaVa1.5 (I-MoF) [84]13B13.6B79.3----31.386.7
BRAVE3B10.3B82.566.066.354.252.742.087.6
", + "bbox": [ + 246, + 282, + 754, + 469 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Visual question answering", + "text_level": 1, + "bbox": [ + 215, + 496, + 478, + 512 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fine-tuning details. We use the same VQA mixture as in Sec. 2 and fine-tune both the MEQ-Former and the LM at $224 \\times 224$ resolution. Vision encoders are kept frozen. This is followed by a high-resolution fine-tuning stage on VQAv2 and OKVQA training sets at $336 \\times 336$ resolution, similar to previous work [19,56,64]. The resulting model is evaluated on VQAv2 and OKVQA benchmarks as well as zero-shot evaluated on GQA [48], VizWiz-QA [39], MMVP [84] and POPE [59], following [25,56,64,84]. To be able to compare to the methods that fine-tune on GQA training set, e.g. [6,61,64], we also perform an additional training stage on top of the mixture-finetuned model, and evaluate the model on the GQA evaluation set (test-dev). Please see the supplementary for the details.", + "bbox": [ + 212, + 520, + 787, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. We summarize the VQA evaluations in Table 4. BRAVE achieves strong results for all the benchmarks in both fine-tuned and zero-shot evaluation scenarios. For OKVQA, GQA, VizWiz-QA, MMVP, and POPE benchmarks, BRAVE achieves the best results, and for VQAv2, it stays competitive with PaLI-17B:", + "bbox": [ + 212, + 691, + 787, + 753 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For MMVP, BRAVE significantly improves the performance over previous approaches, making it less susceptible to failure modes of CLIP-based encoders. This can be further seen from the qualitatives in Figures 1 and 3.", + "- For VizWiz-QA, BRAVE achieves the best results without using any scene text understanding or OCR data during pre-training, as opposed to [6,14,64]." + ], + "bbox": [ + 223, + 763, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bfda9a6e63e1900bbaf032351484f0a6836e6058b510ba6dcf807e09244f6d5e.jpg", + "image_caption": [ + "Fig. 3: Qualitative results. We compare predictions of BRAVE and the VLMs with different vision encoders, e.g. CLIP, on samples from the MMVP benchmark. Following [84], a model is considered correct only if it answers to both images in a pair correctly, i.e. if it can successfully differentiate between images with semantic differences. Note that the images in a pair are seen independently, i.e. neither of the images is provided as context for the other one. All encoders output some correct predictions, yet none of them performs consistently well on a broad range of inputs. BRAVE alleviates this by combining diverse visual features, leading to a more consistent performance. The quantitative difference is indeed stark: $42\\%$ for BRAVE vs $27.3\\%$ for best single encoder (Tables 2 and 4). See supplementary for more qualitative results." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 787, + 402 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For GQA and OKVQA that requires different capabilities, i.e. spatial reasoning and open-world knowledge, BRAVE noticeably improves the performance.", + "- For POPE, BRAVE further reduces the visual hallucinations over dual encoder methods [61,84], confirming that using multiple encoders can be effective for visual grounding (Also see Sec. 5 for a discussion).", + "- For VQAv2, BRAVE is the second best after PaLI-17B, while staying more efficient in terms of training data, model size, and input resolution, similar to the COCO captioning result in Sec. 4.1. We expect BRAVE to benefit from such scaling efforts as well, which is a future direction to investigate." + ], + "bbox": [ + 223, + 580, + 784, + 714 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Analysis of BRAVE", + "text_level": 1, + "bbox": [ + 215, + 742, + 403, + 757 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We study different aspects of BRAVE including an ablation study, contribution of vision encoders to the final performance, comparing MEQ-Former to an ensembling strategy, and the role of pre-training data. Due to the space limit, please see the supplementary for more qualitative and quantitative results.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/26b0088519087d6b637a8ddf2533a74d5c6bd308c92d976018e5fc1fab6e545a.jpg", + "table_caption": [ + "Table 5: Ablation study of BRAVE. We start from the base pre-training setup explained in Sec. 3, denoted as A0. Each row corresponds to a particular deviation from A0, while keeping the rest fixed. For example, A5 evaluates the performance when high-resolution fine-tuning is disabled. N/A means the ablation is not applicable for that evaluation. See Sec. 4.3 for the discussions and supplementary for more results." + ], + "table_footnote": [], + "table_body": "
Ablation IDStudy SubjectOriginalChangeCOCO Captioning (Karpathy val)VQAv2 (Karpathy val)OKVQA (val)
A0Base--147.081.865.7
A1LM fine-tuning for VQAFull fine-tuningNo fine-tuningN/A78.657.5
LoRA (r=64)N/A80.762.8
LoRA (r=128)N/A81.062.9
A2Synthetic VQA data [12]✓✗N/A81.164.0
A3Encoder dropout✓✗145.381.366.0
A4MEQ-Former text input✓✗145.981.464.9
A5High-res fine-tuning✓✗145.279.665.0
A6A3 + A5✓✗144.079.064.5
A7A4 + A5✓✗145.178.363.4
A8Language modelFlanT5-XLFlanT5-L142.579.965.5
", + "bbox": [ + 272, + 226, + 727, + 381 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation study. We perform a comprehensive study in Table 5 where we ablate different design choices, e.g. training data and choice of language model. The ablations are labeled from A1 to A8, each showing a particular deviation from the original pre-training setup A0 (Sec. 3). We summarize the key findings: (A1): LM fine-tuning for VQA significantly boosts the performance compared to frozen LM, similar to the observation in [64]. To minimize the performance loss, we perform a LoRA [46] fine-tuning by inserting LoRA layers to the LM. With a rank of 128, we compensated for $70\\%$ of the performance gap while using $10\\times$ fewer trainable parameters than full fine-tuning (324M vs 3B). This underscores the development of parameter-efficient fine-tuning methods as an important direction for more efficient VLMs.", + "bbox": [ + 212, + 412, + 787, + 579 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(A2): Using synthetic VQA data [12] noticeably enhances our performance, while being much cheaper to collect than human annotated data. This can potentially be further improved by using more diverse templates [38, 83].", + "bbox": [ + 212, + 580, + 785, + 626 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(A3,A4,A5): Encoder dropout during training improves the performance for COCO captioning and VQAv2 but slightly degrades for OKVQA. Using text prompt as an additional input to MEQ-Former helps with extracting better task-aligned features. Similarly, high-resolution fine-tuning stage gives a boost to the performance at the expense of processing more visual tokens.", + "bbox": [ + 212, + 627, + 785, + 702 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(A6,A7): These ablations combine the impact of previously ablated design choices, further demonstrating their contribution to the final performance.", + "bbox": [ + 212, + 703, + 785, + 734 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(A8): A stronger LM significantly helps for all the tasks, underscoring that scaling the VLM along the vision and language axes simultaneously have complementary benefits, and both are important.", + "bbox": [ + 212, + 734, + 785, + 781 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Contribution of vision encoders. To understand the impact of vision encoders to the final performance of BRAVE, we perform additional evaluations:", + "bbox": [ + 212, + 808, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7112a9140dd9ea5ba4c641a7c29b3ca11f169f479fde8f9f890463b9b1294523.jpg", + "image_caption": [ + "Fig. 4: Contribution of vision encoders to BRAVE. Left: We analyze the robustness of BRAVE when a subset of encoders are removed during evaluation. We report the average drop in CIDEr for COCO and accuracy for VQAv2. Right: We compute average attention scores for different vision encoders cross-attended by the MEQ-Former for COCO and VQAv2. See Sec. 4.3 for the discussions." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 730, + 246 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "- In Fig. 4-left, we perform a robustness analysis by removing a subset of vision encoders during inference. This is performed for all the combinations and we report the average drop in performance for COCO captioning and VQAv2 tasks (Full results in supplementary). The results show that the performance of BRAVE degrades gracefully up to two encoder removals for both tasks, indicating some redundancy among the encoders. Because of this, even if some encoders are missing, BRAVE can make use of the remaining set of encoders to compensate. Beyond the two encoder removal, the drop in the performance is more severe, suggesting that some features are more unique and consequently harder to replace using the remaining set.", + "bbox": [ + 225, + 353, + 787, + 503 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "- In Fig. 4-right, we compute attention scores of features from different encoders that are cross-attended by the MEQ-Former queries. These are averaged for randomly sampled 1k samples from COCO and VQAv2 validation sets. The results show that the MEQ-Former can cross-attend, or \"pay attention to\", some encoders more than the others depending on the downstream task adaptively, further confirming the usefulness of having different biases.", + "bbox": [ + 225, + 505, + 787, + 594 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Related work", + "text_level": 1, + "bbox": [ + 215, + 614, + 382, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Several works have focused on training VLMs to solve tasks requiring both vision and text understanding capabilities. Most of these works make use of pre-trained vision encoders [9,27,31,49,75,99] and LMs [6,22,24,44,76,85,103]. Architecture-wise, different approaches were developed to project visual features to the textual embedding space, e.g. Q-Former [25, 56, 106], P-Former [51], cross-attention [3, 6, 19, 95], linear projector [14, 15, 66] and MLP [64, 89]. Different objectives were also considered depending on the downstream tasks, e.g. image-text contrastive learning [50,73,75,94,101], auto-regressive text generation [17,19,23,87,88,91] or both [56-58,96] (see [102] for a survey). In contrast to ours, these works consider a single vision encoder which has inherent limitations [47,59,82,84] (Sec. 2).", + "bbox": [ + 212, + 643, + 787, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Concurrent to our work, LLaVa-MoF [84] proposes combining CLIP [75] and DINOv2 [74] encoders using separate adapters. SPHINX [61] mixes CLIP, DINOv2, and Q-Former outputs with linear projection layers. Both of the methods", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "input a concatenation of the extracted visual embeddings to the LM, which is not scalable to multiple encoders. We differ from these works by proposing a flexible and unified resampling mechanism that combines features from several vision encoders while staying efficient and improving the performance (Tables 3, 4).", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our work focuses on scaling the VLMs along the vision axis, unlike other works that explored scaling homogenously model size across the visual and/or language component [1,3,4,18,19,28,34,80,90], training data [6,10,13,49,50,60, 71,78,87,89] and number of tasks and modalities [5,7,40,43,67-69,72,98]. BRAVE can be combined with these efforts to further improve performance. Moreover, several studies investigated the visual aspect of VLMs by evaluating them against hallucinations and other robustness issues [32,47,59,62,63,82,104]. BRAVE improves the robustness to those issues by broadening visual capabilities of VLMs.", + "bbox": [ + 212, + 207, + 787, + 328 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions and Limitations", + "text_level": 1, + "bbox": [ + 215, + 352, + 526, + 369 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We focus on broadening the visual encoding aspect of VLMs. Our evaluation of vision encoders revealed that there is no encoder achieving a consistent top performance across tasks and encoders with different biases can perform similarly. Motivated by this, we introduce BRAVE that combines diverse features from several vision encoders using a small number of trainable parameters. Extensive evaluations show that BRAVE achieves state-of-the-art performance and improves robustness of VLMs against out-of-distribution inputs and visual hallucinations. We discuss limitations of our work and future directions below:", + "bbox": [ + 212, + 385, + 787, + 505 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Adaptive mechanisms. While MEQ-Former can resample from the useful features and discard the irrelevant ones, it still needs forward passes from all the encoders. A future direction could be to explore adaptive mechanisms to pre-select which encoders to resample from, reducing the inference cost.", + "bbox": [ + 212, + 507, + 787, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Improving sample efficiency. BRAVE uses around 100M image-text pairs during pre-training which is an order of magnitude less than recent methods [6, 19, 87], yet can benefit from further lowering the sample complexity similar to [51]. More broad vision encoders. The set of encoders we investigated do not completely cover all types of visual biases, e.g. including the ones with strong 3D [29, 52] or semantic [54] priors could further improve the performance. Our results show that the current set is a good starting point to see the benefits of consolidation. Another interesting direction is to distill from the MEQ-Former-combined features back into a single encoder for a more unified solution.", + "bbox": [ + 212, + 566, + 787, + 702 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Different modalities and multiple frames. While we focus on image and text modalities, BRAVE is a general method that could be potentially used to fuse more modalities, e.g. audio or 3D, or extended to incorporate multiple frames, e.g. for video understanding or in-context learning.", + "bbox": [ + 212, + 703, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Biases of text-generative models. As we employ pre-trained LMs in our setup, the resulting VLMs are susceptible to the longstanding bias and fairness issues of these models, thus they should be used with caution.", + "bbox": [ + 212, + 763, + 787, + 809 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Diego Martin Arroyo, Ferjad Naeem, Xingyi Zhou, Yannick Strümpler and Yongqin Xian for their help with the project.", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8948-8957 (2019)", + "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)", + "5. Bachmann, R., Kar, O.F., Mizrahi, D., Garjani, A., Gao, M., Griffiths, D., Hu, J., Dehghan, A., Zamir, A.: 4m-21: An any-to-any vision model for tens of tasks and modalities. arXiv preprint arXiv:2406.09406 (2024)", + "6. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)", + "7. Bai, Y., Geng, X., Mangalam, K., Bar, A., Yuille, A., Darrell, T., Malik, J., Efros, A.A.: Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785 (2023)", + "8. Bang, Y., Cahyawijaya, S., Lee, N., Dai, W., Su, D., Wilie, B., Lovenia, H., Ji, Z., Yu, T., Chung, W., et al.: A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023 (2023)", + "9. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: International Conference on Machine Learning. pp. 1059-1071. PMLR (2021)", + "10. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022)", + "11. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "12. Changpinyo, S., Kukliansky, D., Szpektor, I., Chen, X., Ding, N., Soricut, R.: All you may need for vqa are image captions. arXiv preprint arXiv:2205.01883 (2022)", + "13. Changpinyo, S., Sharma, P.K., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3557-3567 (2021)", + "14. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)", + "15. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023)", + "17. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. In: International Conference on Learning Representations (2022)", + "18. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Changpinyo, S., Wu, J., Ruiz, C.R., Goodman, S., Wang, X., Tay, Y., et al.: Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565 (2023)", + "19. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)", + "20. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)", + "21. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023)", + "22. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)", + "23. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021)", + "24. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)", + "25. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=vvoWPYqZJA", + "26. Dehghani, M., Gritsenko, A., Arnab, A., Minderer, M., Tay, Y.: Scenic: A jax library for computer vision research and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21393-21398 (2022)", + "27. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)", + "28. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)", + "29. Eftekhar, A., Sax, A., Bachmann, R., Malik, J., Zamir, A.R.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 10766-10776 (2021)", + "30. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Fang, Y., Wang, W., Xie, B., Sun, Q.S., Wu, L.Y., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA: Exploring the limits of masked visual representation learning at scale. ArXiv abs/2211.07636 (2022)", + "32. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)", + "33. Geman, S., Bienenstock, E., Doursat, R.: Neural networks and the bias/variance dilemma. Neural computation 4(1), 1-58 (1992)", + "34. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790 (2023)", + "35. Goodwin, C.: Seeing in depth. Social studies of science 25(2), 237-274 (1995)", + "36. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6904-6913 (2017)", + "37. Guo, B., Zhang, X., Wang, Z., Jiang, M., Nie, J., Ding, Y., Yue, J., Wu, Y.: How close is chatgpt to human experts? comparison corpus, evaluation, and detection. arXiv preprint arXiv:2301.07597 (2023)", + "38. Guo, J., Li, J., Li, D., Tiong, A.M.H., Li, B., Tao, D., Hoi, S.: From images to textual prompts: Zero-shot visual question answering with frozen large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10867-10877 (2023)", + "39. Gurari, D., Li, Q., Stangl, A.J., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3608-3617 (2018)", + "40. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)", + "41. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021)", + "42. Heek, J., Levskaya, A., Oliver, A., Ritter, M., Rondepierre, B., Steiner, A., van Zee, M.: Flax: A neural network library and ecosystem for jax, 2020. URL http://github.com/google/flax1 (2020)", + "43. Herzig, R., Abramovich, O., Ben Avraham, E., Arbelle, A., Karlinsky, L., Shamir, A., Darrell, T., Globerson, A.: Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6803-6815 (2024)", + "44. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022)", + "45. Howard, I.P., Rogers, B.J.: Binocular vision and stereopsis. Oxford University Press, USA (1995)", + "46. Hu, J.E., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Chen, W.: LoRA: Low-rank adaptation of large language models. ArXiv abs/2106.09685 (2021), https://api-semanticscholar.org/CorpusID:235458009", + "47. Huang, W., Liu, H., Guo, M., Gong, N.Z.: Visual hallucinations of multi-modal large language models. arXiv preprint arXiv:2402.14683 (2024)" + ], + "bbox": [ + 222, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019)", + "49. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below.", + "50. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)", + "51. Jian, Y., Gao, C., Vosoughi, S.: Bootstrapping vision-language learning with decoupled language pre-training. Advances in Neural Information Processing Systems 36 (2024)", + "52. Kar, O.F., Yeo, T., Atanov, A., Zamir, A.: 3d common corruptions and data augmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18963-18974 (2022)", + "53. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3128-3137 (2015)", + "54. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "55. Korbar, B., Xian, Y., Tonioni, A., Zisserman, A., Tombari, F.: Text-conditioned resampler for long form video understanding. arXiv preprint arXiv:2312.11897 (2023)", + "56. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "57. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: BLIP: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: International Conference on Machine Learning (2022)", + "58. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)", + "59. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)", + "60. Lin, J., Yin, H., Ping, W., Lu, Y., Molchanov, P., Tao, A., Mao, H., Kautz, J., Shoeybi, M., Han, S.: Vila: On pre-training for visual language models. arXiv preprint arXiv:2312.07533 (2023)", + "61. Lin, Z., Liu, C., Zhang, R., Gao, P., Qiu, L., Xiao, H., Qiu, H., Lin, C., Shao, W., Chen, K., et al.: Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575 (2023)", + "62. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023)", + "63. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253 (2024)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)", + "65. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (2024)", + "66. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "67. Liu, S., Fan, L.J., Johns, E., Yu, Z., Xiao, C., Anandkumar, A.: Prismer: A vision-language model with an ensemble of experts. ArXiv abs/2303.02506 (2023)", + "68. Lu, J., Clark, C., Lee, S., Zhang, Z., Khosla, S., Marten, R., Hoiem, D., Kembhavi, A.: Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172 (2023)", + "69. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-IO: A unified model for vision, language, and multi-modal tasks. In: The Eleventh International Conference on Learning Representations (2023)", + "70. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 3195-3204 (2019)", + "71. McKinzie, B., Gan, Z., Fauconnier, J.P., Dodge, S., Zhang, B., Dufter, P., Shah, D., Du, X., Peng, F., Weers, F., et al.: Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611 (2024)", + "72. Mizrahi, D., Bachmann, R., Kar, O.F., Yeo, T., Gao, M., Dehghan, A., Zamir, A.: 4M: Massively multimodal masked modeling. In: Advances in Neural Information Processing Systems (2023)", + "73. Naeem, M.F., Xian, Y., Zhai, X., Hoyer, L., Van Gool, L., Tombari, F.: Silc: Improving vision language pretraining with self-distillation. arXiv preprint arXiv:2310.13355 (2023)", + "74. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.Q., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y.B., Li, S.W., Misra, I., Rabbat, M.G., Sharma, V., Synnaeve, G., Xu, H., Jégou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINoV2: Learning robust visual features without supervision. ArXiv abs/2304.07193 (2023)", + "75. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)", + "76. Raffel, C., Shazeer, N.M., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "77. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "78. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + ], + "bbox": [ + 223, + 147, + 784, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "79. Shen, Y., Heacock, L., Elias, J., Hentel, K.D., Reig, B., Shih, G., Moy, L.: Chatgpt and other large language models are double-edged swords (2023)", + "80. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)", + "81. Thorp, H.H.: Chatgpt is fun, but not an author (2023)", + "82. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022)", + "83. Tiong, A.M.H., Li, J., Li, B., Savarese, S., Hoi, S.C.: Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. arXiv preprint arXiv:2210.08773 (2022)", + "84. Tong, S., Liu, Z., Zhai, Y., Ma, Y., LeCun, Y., Xie, S.: Eyes wide shut? exploring the visual shortcomings of multimodal llms. arXiv preprint arXiv:2401.06209 (2024)", + "85. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "86. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)", + "87. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022)", + "88. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022)", + "89. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)", + "90. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., Wei, F.: Image as a foreign language: BEiT pretraining for all vision and vision-language tasks. ArXiv abs/2208.10442 (2022)", + "91. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021)", + "92. Xu, J., Zhou, X., Yan, S., Gu, X., Arnab, A., Sun, C., Wang, X., Schmid, C.: Pixel Aligned Language Models. arXiv preprint arXiv: 2312.09237 (2023)", + "93. Yadav, D., Jain, R., Agrawal, H., Chattopadhyay, P., Singh, T., Jain, A., Singh, S.B., Lee, S., Batra, D.: Towards better evaluation systems for ai agents (2019)", + "94. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021)", + "95. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + ], + "bbox": [ + 223, + 146, + 784, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "O.F. Kar et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "96. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (2022)", + "97. Yu, S., Cho, J., Yadav, P., Bansal, M.: Self-chained image-language model for video localization and question answering. arXiv preprint arXiv:2305.06988 (2023)", + "98. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)", + "99. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)", + "100. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023)", + "101. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022)", + "102. Zhang, D., Yu, Y., Li, C., Dong, J., Su, D., Chu, C., Yu, D.: Mm-llms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601 (2024)", + "103. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)", + "104. Zhao, Y., Pang, T., Du, C., Yang, X., Li, C., Cheung, N.M.M., Lin, M.: On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems 36 (2024)", + "105. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: iBoT: Image BERT pre-training with online tokenizer. In: International Conference on Learning Representations (2022)", + "106. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + ], + "bbox": [ + 217, + 146, + 787, + 590 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "BRAVE: Broadening the visual encoding of vision-language models", + "bbox": [ + 284, + 114, + 730, + 128 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 20 + } +] \ No newline at end of file diff --git a/2024/BRAVE_ Broadening the visual encoding of vision-language models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_model.json b/2024/BRAVE_ Broadening the visual encoding of vision-language models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_model.json new file mode 100644 index 0000000000000000000000000000000000000000..12c3dc2918f9297443e245fa9971f7a8aa80473e --- /dev/null +++ b/2024/BRAVE_ Broadening the visual encoding of vision-language models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_model.json @@ -0,0 +1,3069 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.236, + 0.139, + 0.77, + 0.187 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.213, + 0.751, + 0.244 + ], + "angle": 0, + "content": "Oğuzhan Fatih Kar1,2, Alessio Tonioni1, Petra Poklukar1, Achin Kulshrestha1, Amir Zamir2, and Federico Tombari1,3" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.254, + 0.784, + 0.285 + ], + "angle": 0, + "content": "\\(^{1}\\)Google \\(^{2}\\)Swiss Federal Institute of Technology Lausanne (EPFL) \\(^{3}\\)TU Munich https://brave-vlms.epfl.ch" + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.339, + 0.488, + 0.351 + ], + "angle": 0, + "content": "Improving visual capabilities of VLMs with BRAVE" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.352, + 0.327, + 0.369 + ], + "angle": 0, + "content": "Is there a hand using the mouse in this image?" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.369, + 0.324, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.407, + 0.262, + 0.415 + ], + "angle": 0, + "content": "(a) Yes" + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.352, + 0.425, + 0.369 + ], + "angle": 0, + "content": "Are the butterfly's feet visible?" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.369, + 0.427, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.407, + 0.366, + 0.415 + ], + "angle": 0, + "content": "(a) Yes" + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.352, + 0.529, + 0.369 + ], + "angle": 0, + "content": "Is the door of the truck open?" + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.369, + 0.48, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.407, + 0.468, + 0.415 + ], + "angle": 0, + "content": "(a) Yes" + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.407, + 0.515, + 0.415 + ], + "angle": 0, + "content": "(b) No" + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.423, + 0.323, + 0.442 + ], + "angle": 0, + "content": "InstructBLIP (a) (b) (EVA Encoder)" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.447, + 0.322, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.473, + 0.322, + 0.492 + ], + "angle": 0, + "content": "BRAVE (a) (b) (Multiple Encoders)" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.423, + 0.415, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.449, + 0.416, + 0.457 + ], + "angle": 0, + "content": "(a) (b)" + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.423, + 0.509, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.449, + 0.509, + 0.457 + ], + "angle": 0, + "content": "(a) (a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.499, + 0.789, + 0.617 + ], + "angle": 0, + "content": "Fig. 1: We propose BRAVE to broaden the visual capabilities of vision-language models (VLMs). Left: In contrast to existing methods, e.g. InstructBLIP [25] or LLaVA-1.5 [64], that use a single vision encoder [59,84], BRAVE combines diverse features from multiple vision encoders into a more versatile and compact representation. The examples are taken from [84] and assess the VLM's ability to differentiate images with visual differences. Right: BRAVE leads to state-of-the-art performance on a wide range of captioning and visual question answering tasks. Furthermore, it significantly improves the performance on benchmarks [84] where commonly employed vision encoders fail." + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.339, + 0.76, + 0.352 + ], + "angle": 0, + "content": "SoTA Performance in Captioning & VQA" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.352, + 0.787, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.674, + 0.744, + 0.841 + ], + "angle": 0, + "content": "Abstract. Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. \"blindness\" to certain image features, visual hallucination, etc. To address these issues, we study broadening the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.26 + ], + "angle": 0, + "content": "from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.272, + 0.743, + 0.288 + ], + "angle": 0, + "content": "Keywords: Vision-Language \\(\\cdot\\) Visual Encoding \\(\\cdot\\) Multi-modal Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.311, + 0.377, + 0.327 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.787, + 0.462 + ], + "angle": 0, + "content": "Vision-language models (VLMs) have recently seen significant improvements on solving tasks requiring both visual and text understanding capabilities such as captioning, visual question answering (VQA), and instruction following. These advancements are fueled by the progress in uni-modal vision encoders [30, 75] and language models (LMs) [22, 24], which are then combined with different \"bridging\" techniques [3,56,66] to create VLMs. Improved training recipes [25,64] and scalability efforts in terms of training data and model size [6,19,77,87,99] further boost the performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.787, + 0.599 + ], + "angle": 0, + "content": "Despite the progress, VLMs are subject to several shortcomings: on the language side, the LMs are known to be susceptible to hallucinations and logical faults [8,37,79,81], while on the vision side, they are limited by the capabilities of the vision encoder. For example, Tong et al. [84] observed that commonly employed vision encoders such as CLIP [75] can exhibit \"blindness\", i.e. fail to differentiate images with clear visual differences (See Fig. 1, left). Similarly, Li et al. [59] showed that VLMs are prone to visual hallucinations where the models imagine incorrect details about a given image. These limitations create a bottleneck for developing performant and visually grounded VLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.78 + ], + "angle": 0, + "content": "To address these, we study broadening the visual encoding capabilities of VLMs. It is known in machine learning that different representations of an input can yield different generalization properties, and each can play a role when ensembled to create a more complete representation [33]. Inspired by this, we first perform a comprehensive evaluation of VLMs which differ only in their vision encoders, i.e. visual biases. For this, we consider encoders with different biases, e.g. different objectives, training data, and model sizes. Our results in Table 2 show that different encoders lead to varying performance across vision-language tasks and there is no single encoder achieving a consistent top performance across tasks. Furthermore, we find that encoders with different biases can perform surprisingly similarly, suggesting that there could be multiple cues [35,45] that can be leveraged to solve the vision-language tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Motivated by these findings, we propose to employ various vision encoders for VLMs and introduce a method to learn how to combine them efficiently. We denote the method as BRAVE, which stands for broadening the visual encoding of VLMs. As shown in Fig. 2, BRAVE combines features from an arbitrary number" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "of vision encoders into a compressed fixed-length visual representation which is then fed as a soft visual prompt to a frozen LM. We achieve the bridging of encoders and the LM by introducing a lightweight multi-encoder querying transformer (MEQ-Former), which is a multi-encoder generalization of the Q-Former proposed in the single encoder BLIP-2 framework [56]. In particular, MEQ-Former accepts a textual prompt and a set of learnable queries as inputs and jointly cross-attends to the features from different vision encoders (Fig. 2). We provide a simple yet effective recipe in Sec. 3 for training the MEQ-Former in a single stage that bypasses the two-stage pre-training paradigm of [56]. Notably, this is achieved with a smaller number of trainable parameters during pre-training than the existing methods [3,6,19,25,61,64,87] (Sec. 4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Our evaluation on a wide range of captioning and VQA tasks shows that BRAVE effectively consolidates diverse visual signals into a broad and contextual representation, leading to consistently better performance over the state-of-the-art and improved robustness against out-of-distribution inputs, as shown in Fig. 1 and Sec. 4. Furthermore, in contrast to previous works that benefit from scaling the language axis [3, 19, 25, 64] by using larger LMs, our work demonstrates that scaling along the vision axis has a significant potential for VLMs." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.419, + 0.594, + 0.433 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.446, + 0.784, + 0.49 + ], + "angle": 0, + "content": "- We conduct a systematic analysis of several vision encoders with different inductive biases, e.g. training data, objective, model size, on solving vision-language tasks under a unified training and evaluation framework." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.492, + 0.785, + 0.566 + ], + "angle": 0, + "content": "- We introduce BRAVE, a method that efficiently consolidates features from any number of vision encoders into a compressed and contextual representation, leading to 1) state-of-the-art performance on several captioning and VQA tasks and 2) significantly improved robustness against visual hallucinations and out-of-distribution inputs." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.567, + 0.785, + 0.61 + ], + "angle": 0, + "content": "- We perform a comprehensive ablation study of BRAVE, highlighting the impact of the design choices, which we hope provide useful insights to the researchers for further advancements in VLMs." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.635, + 0.764, + 0.654 + ], + "angle": 0, + "content": "2 Impact of vision encoders for vision-language models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.757 + ], + "angle": 0, + "content": "To quantify the impact of visual biases on the performance of VLMs, we compare VLMs with different vision encoders on commonly evaluated VQA and captioning tasks. For this, we develop a pre-training, fine-tuning and evaluation setup, as explained next. To the best of our knowledge, this is the first unified and comprehensive evaluation of different vision encoders for vision-language understanding and generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "VLM architecture. We use a frozen vision encoder which is connected to a frozen language model by a bridge network with trainable parameters. For the bridge, we adopt the Q-Former module from [56]. This choice is based on the following reasons: 1) the Q-Former resamples visual features to a fixed-length" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Overview of vision encoders we benchmarked. All encoders use a ViT [27] backbone, yet they differ in terms of objective, training data, and model size. ITC: Image-text contrastive learning [75], MIM: Masked image modelling [41], LGC: Local-to-global correspondence learning [11]. See Sec. 2 for details." + }, + { + "type": "table", + "bbox": [ + 0.272, + 0.213, + 0.728, + 0.351 + ], + "angle": 0, + "content": "
Vision encoderParametersTraining dataObjective
CLIP-L/14 [75]0.3BOpenAI WITITC
OpenCLIP-G/14 [21]1.8BLAION-2BITC
EVA-CLIP-g [30]1BLAION-400MMIM (CLIP features)
SIGLIP-G/14 [100]1.8BWebLIITC (Sigmoid loss)
SILC-G/16 [73]1.8BWebLIITC + LGC
ViT-e [19]3.8BJFT-3BClassification
ViT-G [99]1.8BJFT-3BClassification
DINOv2-L/14 [74]0.3BLVD-142MLGC + iBOT [105]
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.789, + 0.651 + ], + "angle": 0, + "content": "output before feeding them to the LM which enables a fair comparison of vision encoders with different output dimensionalities, 2) the Q-Former is efficient to train and evaluate as it is compatible with frozen pre-trained vision encoders and LMs. For the vision encoder, we consider a diverse set of options as described in the next paragraph. For the LM, we follow [56] and use FlanT5-XL [24]. The Q-Former has 110M parameters consisting of a series of cross- and self-attention layers and a fixed number of learnable queries. Unless specified otherwise, we use 32 queries of dimension 768 which is also the hidden dimension of the Q-Former. The queries interact with visual features through cross-attention layers, enabling a significant reduction in their dimensionality. For example, CLIP-L/14 [75] visual features are reduced from \\(257 \\times 1024\\) to \\(32 \\times 768\\). The output of the Q-Former is then linearly projected to the textual embedding space of the LM to create a soft visual prompt, and then prepended to the textual prompt that specifies the task. Both visual and textual prompts are fed as inputs to the frozen LM. For captioning, we use the textual prompt \"A photo of\" and for VQA, we directly use the question as the prompt. Similar to InstructBLIP [25], we give the textual prompt as additional input to the Q-Former together with the learnable queries to extract more task-aligned visual prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.788, + 0.792 + ], + "angle": 0, + "content": "Vision encoders. As summarized in Table 1, we consider eight recently introduced vision encoders: CLIP [75], OpenCLIP [21], EVA-CLIP [30], SIGLIP [100], SILC [73], ViT-e [19], ViT-G [99], and DINOv2 [74]. While they all use ViT-based backbones [27], they differ in terms of 1) training data, e.g. LAION [77], WebLI [19], LVD-142M [74], etc., 2) training objective, e.g. image-text contrastive learning, masked image modelling, classification, etc., and 3) model size, e.g. 300M to 4B parameters. Due to this diversity, they incorporate different biases and thus potentially capture different aspects of the underlying depicted scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Pre-training data and objectives. We pre-train the Q-Former using WebLI [19] dataset at \\(224 \\times 224\\) resolution. We use the filtered and de-duplicated" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.231 + ], + "angle": 0, + "content": "Table 2: Benchmarking vision encoders on different vision-language tasks. For COCO captioning, we report CIDEr [86] score. For VQA tasks, we report top-1 accuracy. For MMVP, we report average pair accuracy [84]. Top-3 best results are shown with Blue, Green, and Orange, respectively. All VLMs use FlanT5-XL [24] as the language model and a Q-Former [56] to bridge vision and language modalities. See Sec. 2 for details." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.244, + 0.734, + 0.396 + ], + "angle": 0, + "content": "
Vision encoderCOCO cap. ↑ Karpathy valVQAv2 ↑ Karpathy valOKVQA ↑ valGQA ↑ test-devMMVP ↑ test
CLIP-L/14133.074.461.048.715.3
OpenCLIP-G128.373.360.648.022.0
EVA-CLIP-g140.977.063.050.127.3
SIGLIP-G/14133.074.762.548.624.0
SILC-G/16141.177.063.449.724.0
ViT-e137.875.661.949.125.3
ViT-G133.874.261.248.320.7
DINOv2-L/14127.671.359.048.022.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.519 + ], + "angle": 0, + "content": "English-only subset with 100 million image-text pairs [19, 73]. We keep both the vision encoder and the LM frozen and only train the parameters of the Q-Former. The model is trained with captioning objective using the alt-text in WebLI as target, as it has been shown to be effective [92]. Similar to [56], we remove the last Transformer layer of the encoder. We refer the reader to the supplementary material for additional training details and hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.789, + 0.801 + ], + "angle": 0, + "content": "Evaluation tasks. We evaluate the obtained VLMs on standard captioning and VQA tasks. For captioning, we use the COCO captioning benchmark [20] and fine-tune the pre-trained VLM on the commonly employed Karpathy training split [53]. Similar to the pre-training, only the Q-Former parameters are updated. For VQA, we again follow the standard practices [19, 25, 56, 64] and fine-tune on a mixture of VQA data that enforces the Q-Former to extract more task-aligned visual features. Our mixture includes the training sets of VQAv2 [36] and OKVQA [70] as well as the synthetically generated VQA data from \\(\\mathrm{VQ^2A}\\) [12], resulting in total 17M training examples. We then evaluate the fine-tuned models on VQAv2 and OKVQA validation sets. Similar to [25, 56], we also evaluate the VLMs' zero-shot VQA capabilities on GQA [48]. Finally, we evaluate the zero-shot performance on MMVP [84] that includes images with semantic differences that are challenging for the current VLMs to handle. Unlike captioning, we fine-tune both Q-Former and the LM parameters while keeping the vision encoder frozen, as we find this to improve the performance, similar to [64]. For both captioning and VQA tasks, we use \\(224\\times 224\\) image resolution. See the supplementary for more details and results." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Results. We report the findings in Table 2 and make the following observations:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.253 + ], + "angle": 0, + "content": "- Encoders with different biases can obtain similar performance. While we observe some variation in VLMs' performance across different encoders, the gap remains small, especially among the top-performing ones. For example, EVA-CLIP-g, SILC-G/16, and ViT-e achieve the best results for VQAv2, GQA and COCO captioning despite their differences in training objective and data, as shown in Table 1. Similar observations can be made for OpenCLIP and DINOv2 encoders that have a small performance gap." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.253, + 0.787, + 0.328 + ], + "angle": 0, + "content": "- MMVP stays challenging for all encoders. While it is originally curated with \"CLIP-blind\" pairs in [84], i.e. images that CLIP perceives as similar despite their visual differences (See Fig. 1 for examples), most of the encoders stay below the random guess accuracy of \\(25\\%\\). Exceptions are EVA-CLIP and ViT-e which slightly surpass the random guess level (\\(27.3\\%\\) and \\(25.3\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.328, + 0.787, + 0.403 + ], + "angle": 0, + "content": "- For tasks requiring compositional reasoning and open-world knowledge, the performance decreases and the gap among VLMs diminishes, as shown by the results on GQA and OKVQA. More precisely, the standard deviation across all the VLMs' performance on GQA and OKVQA is 0.8 and 1.36, respectively, but 1.74 on VQAv2 and 4.91 on COCO." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.403, + 0.787, + 0.448 + ], + "angle": 0, + "content": "- Scaling the size of vision encoder can improve the performance as seen by the consistent improvements obtained by ViT-e over ViT-G when scaling from 1.8B to 3.8B parameters." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.448, + 0.787, + 0.553 + ], + "angle": 0, + "content": "- Pre-training data distribution can play a critical role. OpenCLIP-G/14 model is larger than CLIP-L/14 and uses more training data, yet it noticeably underperforms on most of the evaluated VQA and captioning tasks except the MMVP benchmark, where OpenCLIP significantly outperforms the CLIP model. Since both models are trained using the same training objective, this indicates that the impact of the training objective and dataset to the performance of the VLM can be disentangled, and both are important." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.553, + 0.787, + 0.644 + ], + "angle": 0, + "content": "- Performance correlates between VQA and captioning tasks. Encoders with better performance on VQA tasks generally also perform better on captioning, and vice versa. Spearman's rank correlation between VQA tasks and COCO captioning is 0.90, 0.80, 0.79, 0.56 for VQAv2, OK-VQA, GQA, and MMVP, respectively. Thus, improving the visual component in VLMs is expected to improve performance for a broad range of tasks." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.652, + 0.788, + 0.713 + ], + "angle": 0, + "content": "The above results highlight the importance of using the right vision encoder for VLMs, motivating us to focus on the following question: Can we broaden the vision capabilities of VLMs through combining vision encoders with different visual biases? We answer positively in the next section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.622, + 0.751 + ], + "angle": 0, + "content": "3 Combining vision encoders with BRAVE" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Motivated by the findings in Sec. 2, we propose using multiple vision encoders with different, and potentially complementary, visual biases to create more capable VLMs. To do this, we introduce BRAVE, a method that combines the strengths of different vision encoders while staying efficient in terms of number of trainable parameters. Below, we provide details of BRAVE." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.554, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.149, + 0.778, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.46 + ], + "angle": 0, + "content": "Fig. 2: Overview of BRAVE. Left: We keep all the vision encoders (VEs) and the language model (LM) frozen. The linear projection layers are used to concatenate features from \\( K \\) different VEs, e.g. \\( K = 5 \\), sequence-wise. These are then resampled by the MEQ-Former which accepts a set of learnable queries and a text prompt describing the task as inputs. The output of MEQ-Former is projected to the input space of the LM using fully-connected (FC) layers. The total number of trainable parameters is 116M (\\( \\approx 1\\% \\) of the total parameters). Right: Architecture of the MEQ-Former with \\( N = 12 \\) transformer layers. It interacts with the concatenated features through crossattention layers and produces a fixed-length output to be fed as input to the LM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.789, + 0.777 + ], + "angle": 0, + "content": "Overview. As depicted in Fig. 2, we introduce the multi-encoder querying transformer, or MEQ-Former, that combines visual features from an arbitrary set of encoders and outputs a fixed-length visual representation that is given as a soft prompt to a frozen LM. It takes as an input a sequence comprised of a fixed number of learnable queries and embedded text tokens which describe the task - for captioning, they are the textual prompt \"A photo of\", while for VQA, they are the question prompt for a given image (See supplementary for examples). MEQ-Former interacts with visual features through cross-attention layers. The visual features from different encoders are first linearly projected to the same dimension and then concatenated sequence-wise. The obtained concatenated feature is given as a key and value pair to the cross-attention layers in MEQ-Former, and is cross-attended to by the MEQ-Former's query sequence (Fig. 2). This resampling enables to efficiently processing a large number of visual features since it bypasses the quadratic complexity of self-attention. It also acts as a \"bottleneck\" that keeps the total number of VLM's parameters low compared to a naive ensembling of VLMs that is prohibitively expensive (See a comparison in Sec. 4.3). Moreover, the visual features are not \"marked\" with any encoder-specific embedding, thus MEQ-Former is not biased to differentiate between encoders, allowing for a simpler design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Differences with previous work. Our work differs from other popular approaches like LLaVa [64] and PaLI [19] that require significantly more trainable parameters (\\(\\approx 10\\mathrm{B}\\)) and do not have a resampling mechanism. Resampling visual" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "features from a single encoder has been successfully demonstrated in previous work for single- [25,56] and multiple-frames [3,55,97]. To the best of our knowledge, our work is the first one demonstrating an effective resampling mechanism to consolidate visual features from several arbitrarily different encoders while keeping the number of trainable parameters small (our MEQ-Former having 116M parameters vs BLIP-2's Q-Former is 188M). See Sec. 5 for more comparisons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.79, + 0.368 + ], + "angle": 0, + "content": "Pre-training details. We use the same data and objective as in Sec. 2 and only train the MEQ-Former while keeping the vision encoders and the LM frozen. During pre-training, we randomly mask features from each encoder with \\(20\\%\\) probability as we found that this can act as a regularizer and enforces MEQ-Former to avoid local minima by attending only to the features of a single encoder (Ablated in Sec. 4.3). After pre-training, we fine-tune the MEQ-Former and (optionally) the LM on the downstream tasks, as explained in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.79, + 0.555 + ], + "angle": 0, + "content": "Architecture. For the main results, we combine five vision encoders, namely EVA-CLIP-\\(g\\) [30], CLIP-\\(L/14\\) [75], SILC-\\(G/16\\) [73], ViT-\\(e\\) [19], and DINOv2-\\(L/14\\) [74] to cover all training datasets and objectives from Table 1. Please see Sec. 4.3 and the supplementary for analysis on the contributions of encoders. For the LM, we employ FlanT5-XL [24] as before. We set the visual feature dimension after linear projection as 1408 for concatenation. For MEQ-Former, we use \\(32 \\times 5 = 160\\) learnable queries to proportionally scale the model capacity to the number of vision encoders and set the hidden dimension as 768. The MEQ-Former resampling results in \\(14 \\times\\) reduction in total feature size (from \\(1223 \\times 1408\\) to \\(160 \\times 768\\)). The total number of trainable parameters during pre-training is 116M which is about \\(1\\%\\) of the total number of parameters in the VLM (10.3B)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.622 + ], + "angle": 0, + "content": "Implementation. We use FLAX [42] with the Scenic [26] library to implement our models and training pipeline. For evaluations, we use both Scenic and EvalAI API [93]. We pre-train with batch size of 1024 on 64 TPUv5 chips." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.643, + 0.378, + 0.661 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.735 + ], + "angle": 0, + "content": "We perform evaluations on a broad range of captioning (Sec. 4.1) and VQA tasks (Sec. 4.2) to show the effectiveness of BRAVE. Please see Fig. 1 right for a summary of the results and supplementary for an overview of the tasks. We then provide a comprehensive analysis of BRAVE in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.81 + ], + "angle": 0, + "content": "For captioning, we evaluate on COCO [20] and NoCaps [2] benchmarks. The latter includes out-of-distribution samples with novel classes that do not exist in the COCO benchmark, hence it serves as a challenging testbed to assess robustness. The results (Table 3) show that BRAVE pushes the state-of-the-art for both benchmarks while staying efficient in terms of trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "For VQA, we evaluate on VQAv2 [36], OKVQA [70], GQA [48], and VizWiz-QA [39] benchmarks. To assess the robustness of VLMs, we additionally include" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Table 3: Captioning. We compare BRAVE with state-of-the-art captioning methods on COCO and NoCaps (zero-shot). CIDEr [86] score is reported (Best in bold, second best is underlined). Numbers of other methods are taken from the corresponding publications. Dashed line means no result was reported. BRAVE achieves the best results in NoCaps evaluation sets (both out-domain and overall in validation & test) and second best for COCO, while having much fewer trainable parameters than other methods. For example, compared to PaLI-17B [19], BRAVE uses significantly fewer trainable parameters, training data, and image pixels. Yet, it outperforms the baselines on NoCaps and stays competitive on COCO, demonstrating the effectiveness of having different visual biases for generalization. See Sec. 4.1 for more details." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.296, + 0.756, + 0.421 + ], + "angle": 0, + "content": "
Method# paramsCOCO (fine-tuned)NoCaps (zero-shot, val)NoCaps (zero-shot, test)
TrainableTotalKarpathy testout-domainoverallout-domainoverall
Flamingo [3]10.6B80B138.1----
SimVLM [91]632M632M143.3113.7112.2-110.3
Qwen-VL [6]9.6B9.6B--121.4--
BLIP-2 [56]1.1B4.1B144.5124.8121.6--
InstructBLIP [25]188M14.2B--121.9--
CoCa [96]2.1B2.1B143.6-122.4-120.6
GiT2 [87]5.1B5.1B145.0130.6126.9122.3124.8
PaLI-17B [19]16.9B16.9B149.1-127.0126.7124.4
BRAVE116M10.3B148.0133.3127.6127.1125.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.511 + ], + "angle": 0, + "content": "POPE [59] that measures object hallucinations, and MMVP [84] which exposes the shortcomings of CLIP-based vision encoders that are widely used in VLMs. Our results (Table 4) show that BRAVE achieves consistent improvements over recent methods for all the tasks by leveraging diverse visual features." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.534, + 0.406, + 0.55 + ], + "angle": 0, + "content": "4.1 Image captioning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Fine-tuning details. Similar to Sec. 2, we only fine-tune the MEQ-Former on COCO captioning and keep the vision encoders and the LM frozen. During fine-tuning, we set a higher input resolution as \\(336 \\times 336\\) to further boost the performance, similar to previous works [19,56,64] (See Sec. 4.3 for an ablation). Following [19,25,56,87,96], we also zero-shot evaluated the COCO fine-tuned model on NoCaps [2]. Please see the supplementary for more training details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Results. We summarize the captioning evaluations in Table 3. BRAVE uses the least number of trainable parameters (116M) yet achieves strong results for both COCO and NoCaps benchmarks. For NoCaps, BRAVE is the best performing method with significant gains over recent methods. This is especially the case for out-domain samples with novel classes, demonstrating the usefulness of diversity in visual features for robustness. For COCO, BRAVE stays competitive with the best performing model, PaLI-17B [19]. Notably, BRAVE achieves this while using \\(150 \\times\\) fewer trainable parameters (116M vs 16.9B), \\(16 \\times\\) less pre-training data (100M vs 1.6B) and \\(3 \\times\\) fewer image pixels (\\(336 \\times 336\\) vs \\(588 \\times 588\\)) than PaLI-17B, suggesting that having different visual biases is effective for generalization, while keeping the sample complexity low." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Table 4: Visual question answering. We compare BRAVE with state-of-the-art VQA methods on VQAv2, OKVQA, GQA, VizWiz-QA, MMVP, and POPE. Top-1 accuracies are reported (Best in bold, second best is underlined). For MMVP, we report average pair accuracy [84]. Numbers of other methods are taken from the corresponding publications. Dashed line means no result was reported. BRAVE achieves the best results for six out of seven benchmarks, while staying efficient in terms of trainable parameters. Similar to the COCO captioning results in Table 3, BRAVE stays competitive with PaLI-17B [19] for VQAv2, while using significantly less pre-training data, image resolution, and model parameters (both trainable and total). See Sec. 4.2 for details." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.283, + 0.756, + 0.47 + ], + "angle": 0, + "content": "
Method# paramsFine-tunedZero-shot
TrainableTotalVQAv2 test-devOKVQA valGQA test-devVizWiz-QA test-devGQA test-devMMVP testPOPE test
SimVLM [91]632M632M80.0------
Flamingo [3]10.2B80B82.057.8-31.6---
MiniGPT-v2 [14]7B8B-57.860.153.6---
GiT2 [87]5.1B5.1B81.7------
Qwen-VL [6]9.6B9.6B79.558.659.335.2---
SPHINX-2k [61]13B16.5B80.762.663.144.9--87.2
PaLI-17B [19]16.9B16.9B84.364.5-----
BLIP-2 [56]1.2B12.1B81.654.7-29.444.7-85.3
InstructBLIP [25]188M14.2B-55.5-33.449.516.778.9
ShareGPT4V [16]13.4B13.4B81.0-64.8----
LLaVa1.5 [64]13B13.4B80.0-63.353.6-24.785.9
LLaVa1.6 [65]13B13.4B-46.365.4---86.3
LLaVa1.5 (I-MoF) [84]13B13.6B79.3----31.386.7
BRAVE3B10.3B82.566.066.354.252.742.087.6
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.497, + 0.479, + 0.513 + ], + "angle": 0, + "content": "4.2 Visual question answering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Fine-tuning details. We use the same VQA mixture as in Sec. 2 and fine-tune both the MEQ-Former and the LM at \\(224 \\times 224\\) resolution. Vision encoders are kept frozen. This is followed by a high-resolution fine-tuning stage on VQAv2 and OKVQA training sets at \\(336 \\times 336\\) resolution, similar to previous work [19,56,64]. The resulting model is evaluated on VQAv2 and OKVQA benchmarks as well as zero-shot evaluated on GQA [48], VizWiz-QA [39], MMVP [84] and POPE [59], following [25,56,64,84]. To be able to compare to the methods that fine-tune on GQA training set, e.g. [6,61,64], we also perform an additional training stage on top of the mixture-finetuned model, and evaluate the model on the GQA evaluation set (test-dev). Please see the supplementary for the details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.755 + ], + "angle": 0, + "content": "Results. We summarize the VQA evaluations in Table 4. BRAVE achieves strong results for all the benchmarks in both fine-tuned and zero-shot evaluation scenarios. For OKVQA, GQA, VizWiz-QA, MMVP, and POPE benchmarks, BRAVE achieves the best results, and for VQAv2, it stays competitive with PaLI-17B:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.765, + 0.784, + 0.809 + ], + "angle": 0, + "content": "- For MMVP, BRAVE significantly improves the performance over previous approaches, making it less susceptible to failure modes of CLIP-based encoders. This can be further seen from the qualitatives in Figures 1 and 3." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.811, + 0.784, + 0.84 + ], + "angle": 0, + "content": "- For VizWiz-QA, BRAVE achieves the best results without using any scene text understanding or OCR data during pre-training, as opposed to [6,14,64]." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.788, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Fig. 3: Qualitative results. We compare predictions of BRAVE and the VLMs with different vision encoders, e.g. CLIP, on samples from the MMVP benchmark. Following [84], a model is considered correct only if it answers to both images in a pair correctly, i.e. if it can successfully differentiate between images with semantic differences. Note that the images in a pair are seen independently, i.e. neither of the images is provided as context for the other one. All encoders output some correct predictions, yet none of them performs consistently well on a broad range of inputs. BRAVE alleviates this by combining diverse visual features, leading to a more consistent performance. The quantitative difference is indeed stark: \\(42\\%\\) for BRAVE vs \\(27.3\\%\\) for best single encoder (Tables 2 and 4). See supplementary for more qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.581, + 0.785, + 0.611 + ], + "angle": 0, + "content": "- For GQA and OKVQA that requires different capabilities, i.e. spatial reasoning and open-world knowledge, BRAVE noticeably improves the performance." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.612, + 0.785, + 0.655 + ], + "angle": 0, + "content": "- For POPE, BRAVE further reduces the visual hallucinations over dual encoder methods [61,84], confirming that using multiple encoders can be effective for visual grounding (Also see Sec. 5 for a discussion)." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.655, + 0.785, + 0.715 + ], + "angle": 0, + "content": "- For VQAv2, BRAVE is the second best after PaLI-17B, while staying more efficient in terms of training data, model size, and input resolution, similar to the COCO captioning result in Sec. 4.1. We expect BRAVE to benefit from such scaling efforts as well, which is a future direction to investigate." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.581, + 0.785, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.743, + 0.405, + 0.758 + ], + "angle": 0, + "content": "4.3 Analysis of BRAVE" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We study different aspects of BRAVE including an ablation study, contribution of vision encoders to the final performance, comparing MEQ-Former to an ensembling strategy, and the role of pre-training data. Due to the space limit, please see the supplementary for more qualitative and quantitative results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 5: Ablation study of BRAVE. We start from the base pre-training setup explained in Sec. 3, denoted as A0. Each row corresponds to a particular deviation from A0, while keeping the rest fixed. For example, A5 evaluates the performance when high-resolution fine-tuning is disabled. N/A means the ablation is not applicable for that evaluation. See Sec. 4.3 for the discussions and supplementary for more results." + }, + { + "type": "table", + "bbox": [ + 0.273, + 0.227, + 0.728, + 0.382 + ], + "angle": 0, + "content": "
Ablation IDStudy SubjectOriginalChangeCOCO Captioning (Karpathy val)VQAv2 (Karpathy val)OKVQA (val)
A0Base--147.081.865.7
A1LM fine-tuning for VQAFull fine-tuningNo fine-tuningN/A78.657.5
LoRA (r=64)N/A80.762.8
LoRA (r=128)N/A81.062.9
A2Synthetic VQA data [12]✓✗N/A81.164.0
A3Encoder dropout✓✗145.381.366.0
A4MEQ-Former text input✓✗145.981.464.9
A5High-res fine-tuning✓✗145.279.665.0
A6A3 + A5✓✗144.079.064.5
A7A4 + A5✓✗145.178.363.4
A8Language modelFlanT5-XLFlanT5-L142.579.965.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.58 + ], + "angle": 0, + "content": "Ablation study. We perform a comprehensive study in Table 5 where we ablate different design choices, e.g. training data and choice of language model. The ablations are labeled from A1 to A8, each showing a particular deviation from the original pre-training setup A0 (Sec. 3). We summarize the key findings: (A1): LM fine-tuning for VQA significantly boosts the performance compared to frozen LM, similar to the observation in [64]. To minimize the performance loss, we perform a LoRA [46] fine-tuning by inserting LoRA layers to the LM. With a rank of 128, we compensated for \\(70\\%\\) of the performance gap while using \\(10\\times\\) fewer trainable parameters than full fine-tuning (324M vs 3B). This underscores the development of parameter-efficient fine-tuning methods as an important direction for more efficient VLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.787, + 0.627 + ], + "angle": 0, + "content": "(A2): Using synthetic VQA data [12] noticeably enhances our performance, while being much cheaper to collect than human annotated data. This can potentially be further improved by using more diverse templates [38, 83]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.787, + 0.703 + ], + "angle": 0, + "content": "(A3,A4,A5): Encoder dropout during training improves the performance for COCO captioning and VQAv2 but slightly degrades for OKVQA. Using text prompt as an additional input to MEQ-Former helps with extracting better task-aligned features. Similarly, high-resolution fine-tuning stage gives a boost to the performance at the expense of processing more visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.735 + ], + "angle": 0, + "content": "(A6,A7): These ablations combine the impact of previously ablated design choices, further demonstrating their contribution to the final performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.736, + 0.787, + 0.782 + ], + "angle": 0, + "content": "(A8): A stronger LM significantly helps for all the tasks, underscoring that scaling the VLM along the vision and language axes simultaneously have complementary benefits, and both are important." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Contribution of vision encoders. To understand the impact of vision encoders to the final performance of BRAVE, we perform additional evaluations:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.144, + 0.731, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 4: Contribution of vision encoders to BRAVE. Left: We analyze the robustness of BRAVE when a subset of encoders are removed during evaluation. We report the average drop in CIDEr for COCO and accuracy for VQAv2. Right: We compute average attention scores for different vision encoders cross-attended by the MEQ-Former for COCO and VQAv2. See Sec. 4.3 for the discussions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.354, + 0.788, + 0.505 + ], + "angle": 0, + "content": "- In Fig. 4-left, we perform a robustness analysis by removing a subset of vision encoders during inference. This is performed for all the combinations and we report the average drop in performance for COCO captioning and VQAv2 tasks (Full results in supplementary). The results show that the performance of BRAVE degrades gracefully up to two encoder removals for both tasks, indicating some redundancy among the encoders. Because of this, even if some encoders are missing, BRAVE can make use of the remaining set of encoders to compensate. Beyond the two encoder removal, the drop in the performance is more severe, suggesting that some features are more unique and consequently harder to replace using the remaining set." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.506, + 0.788, + 0.595 + ], + "angle": 0, + "content": "- In Fig. 4-right, we compute attention scores of features from different encoders that are cross-attended by the MEQ-Former queries. These are averaged for randomly sampled 1k samples from COCO and VQAv2 validation sets. The results show that the MEQ-Former can cross-attend, or \"pay attention to\", some encoders more than the others depending on the downstream task adaptively, further confirming the usefulness of having different biases." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.383, + 0.631 + ], + "angle": 0, + "content": "5 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Several works have focused on training VLMs to solve tasks requiring both vision and text understanding capabilities. Most of these works make use of pre-trained vision encoders [9,27,31,49,75,99] and LMs [6,22,24,44,76,85,103]. Architecture-wise, different approaches were developed to project visual features to the textual embedding space, e.g. Q-Former [25, 56, 106], P-Former [51], cross-attention [3, 6, 19, 95], linear projector [14, 15, 66] and MLP [64, 89]. Different objectives were also considered depending on the downstream tasks, e.g. image-text contrastive learning [50,73,75,94,101], auto-regressive text generation [17,19,23,87,88,91] or both [56-58,96] (see [102] for a survey). In contrast to ours, these works consider a single vision encoder which has inherent limitations [47,59,82,84] (Sec. 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Concurrent to our work, LLaVa-MoF [84] proposes combining CLIP [75] and DINOv2 [74] encoders using separate adapters. SPHINX [61] mixes CLIP, DINOv2, and Q-Former outputs with linear projection layers. Both of the methods" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "input a concatenation of the extracted visual embeddings to the LM, which is not scalable to multiple encoders. We differ from these works by proposing a flexible and unified resampling mechanism that combines features from several vision encoders while staying efficient and improving the performance (Tables 3, 4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Our work focuses on scaling the VLMs along the vision axis, unlike other works that explored scaling homogenously model size across the visual and/or language component [1,3,4,18,19,28,34,80,90], training data [6,10,13,49,50,60, 71,78,87,89] and number of tasks and modalities [5,7,40,43,67-69,72,98]. BRAVE can be combined with these efforts to further improve performance. Moreover, several studies investigated the visual aspect of VLMs by evaluating them against hallucinations and other robustness issues [32,47,59,62,63,82,104]. BRAVE improves the robustness to those issues by broadening visual capabilities of VLMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.353, + 0.527, + 0.37 + ], + "angle": 0, + "content": "6 Conclusions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.506 + ], + "angle": 0, + "content": "We focus on broadening the visual encoding aspect of VLMs. Our evaluation of vision encoders revealed that there is no encoder achieving a consistent top performance across tasks and encoders with different biases can perform similarly. Motivated by this, we introduce BRAVE that combines diverse features from several vision encoders using a small number of trainable parameters. Extensive evaluations show that BRAVE achieves state-of-the-art performance and improves robustness of VLMs against out-of-distribution inputs and visual hallucinations. We discuss limitations of our work and future directions below:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Adaptive mechanisms. While MEQ-Former can resample from the useful features and discard the irrelevant ones, it still needs forward passes from all the encoders. A future direction could be to explore adaptive mechanisms to pre-select which encoders to resample from, reducing the inference cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Improving sample efficiency. BRAVE uses around 100M image-text pairs during pre-training which is an order of magnitude less than recent methods [6, 19, 87], yet can benefit from further lowering the sample complexity similar to [51]. More broad vision encoders. The set of encoders we investigated do not completely cover all types of visual biases, e.g. including the ones with strong 3D [29, 52] or semantic [54] priors could further improve the performance. Our results show that the current set is a good starting point to see the benefits of consolidation. Another interesting direction is to distill from the MEQ-Former-combined features back into a single encoder for a more unified solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Different modalities and multiple frames. While we focus on image and text modalities, BRAVE is a general method that could be potentially used to fuse more modalities, e.g. audio or 3D, or extended to incorporate multiple frames, e.g. for video understanding or in-context learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Biases of text-generative models. As we employ pre-trained LMs in our setup, the resulting VLMs are susceptible to the longstanding bias and fairness issues of these models, thus they should be used with caution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Acknowledgements. We thank Diego Martin Arroyo, Ferjad Naeem, Xingyi Zhou, Yannick Strümpler and Yongqin Xian for their help with the project." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.215, + 0.785, + 0.268 + ], + "angle": 0, + "content": "2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8948-8957 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.269, + 0.785, + 0.323 + ], + "angle": 0, + "content": "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.324, + 0.785, + 0.377 + ], + "angle": 0, + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.379, + 0.785, + 0.418 + ], + "angle": 0, + "content": "5. Bachmann, R., Kar, O.F., Mizrahi, D., Garjani, A., Gao, M., Griffiths, D., Hu, J., Dehghan, A., Zamir, A.: 4m-21: An any-to-any vision model for tens of tasks and modalities. arXiv preprint arXiv:2406.09406 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "6. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.447, + 0.785, + 0.486 + ], + "angle": 0, + "content": "7. Bai, Y., Geng, X., Mangalam, K., Bar, A., Yuille, A., Darrell, T., Malik, J., Efros, A.A.: Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.487, + 0.785, + 0.54 + ], + "angle": 0, + "content": "8. Bang, Y., Cahyawijaya, S., Lee, N., Dai, W., Su, D., Wilie, B., Lovenia, H., Ji, Z., Yu, T., Chung, W., et al.: A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.542, + 0.785, + 0.581 + ], + "angle": 0, + "content": "9. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: International Conference on Machine Learning. pp. 1059-1071. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.582, + 0.785, + 0.609 + ], + "angle": 0, + "content": "10. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.61, + 0.785, + 0.663 + ], + "angle": 0, + "content": "11. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "12. Changpinyo, S., Kukliansky, D., Szpektor, I., Chen, X., Ding, N., Soricut, R.: All you may need for vqa are image captions. arXiv preprint arXiv:2205.01883 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.691, + 0.785, + 0.745 + ], + "angle": 0, + "content": "13. Changpinyo, S., Sharma, P.K., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3557-3567 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.746, + 0.785, + 0.799 + ], + "angle": 0, + "content": "14. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "17. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.229, + 0.785, + 0.271 + ], + "angle": 0, + "content": "18. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Changpinyo, S., Wu, J., Ruiz, C.R., Goodman, S., Wang, X., Tay, Y., et al.: Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "19. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "20. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.406 + ], + "angle": 0, + "content": "21. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.461 + ], + "angle": 0, + "content": "22. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "23. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "24. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.595 + ], + "angle": 0, + "content": "25. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=vvoWPYqZJA" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.637 + ], + "angle": 0, + "content": "26. Dehghani, M., Gritsenko, A., Arnab, A., Minderer, M., Tay, Y.: Scenic: A jax library for computer vision research and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21393-21398 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.691 + ], + "angle": 0, + "content": "27. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "28. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.785 + ], + "angle": 0, + "content": "29. Eftekhar, A., Sax, A., Bachmann, R., Malik, J., Zamir, A.R.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 10766-10776 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Fang, Y., Wang, W., Xie, B., Sun, Q.S., Wu, L.Y., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA: Exploring the limits of masked visual representation learning at scale. ArXiv abs/2211.07636 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "32. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.23, + 0.787, + 0.256 + ], + "angle": 0, + "content": "33. Geman, S., Bienenstock, E., Doursat, R.: Neural networks and the bias/variance dilemma. Neural computation 4(1), 1-58 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.257, + 0.787, + 0.297 + ], + "angle": 0, + "content": "34. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.297, + 0.787, + 0.311 + ], + "angle": 0, + "content": "35. Goodwin, C.: Seeing in depth. Social studies of science 25(2), 237-274 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.311, + 0.787, + 0.365 + ], + "angle": 0, + "content": "36. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6904-6913 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.365, + 0.787, + 0.405 + ], + "angle": 0, + "content": "37. Guo, B., Zhang, X., Wang, Z., Jiang, M., Nie, J., Ding, Y., Yue, J., Wu, Y.: How close is chatgpt to human experts? comparison corpus, evaluation, and detection. arXiv preprint arXiv:2301.07597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.405, + 0.787, + 0.461 + ], + "angle": 0, + "content": "38. Guo, J., Li, J., Li, D., Tiong, A.M.H., Li, B., Tao, D., Hoi, S.: From images to textual prompts: Zero-shot visual question answering with frozen large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10867-10877 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.461, + 0.787, + 0.514 + ], + "angle": 0, + "content": "39. Gurari, D., Li, Q., Stangl, A.J., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3608-3617 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "40. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.555, + 0.787, + 0.596 + ], + "angle": 0, + "content": "41. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.596, + 0.787, + 0.637 + ], + "angle": 0, + "content": "42. Heek, J., Levskaya, A., Oliver, A., Ritter, M., Rondepierre, B., Steiner, A., van Zee, M.: Flax: A neural network library and ecosystem for jax, 2020. URL http://github.com/google/flax1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.637, + 0.787, + 0.704 + ], + "angle": 0, + "content": "43. Herzig, R., Abramovich, O., Ben Avraham, E., Arbelle, A., Karlinsky, L., Shamir, A., Darrell, T., Globerson, A.: Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6803-6815 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.705, + 0.787, + 0.746 + ], + "angle": 0, + "content": "44. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.746, + 0.787, + 0.772 + ], + "angle": 0, + "content": "45. Howard, I.P., Rogers, B.J.: Binocular vision and stereopsis. Oxford University Press, USA (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "46. Hu, J.E., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Chen, W.: LoRA: Low-rank adaptation of large language models. ArXiv abs/2106.09685 (2021), https://api-semanticscholar.org/CorpusID:235458009" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Huang, W., Liu, H., Guo, M., Gong, N.Z.: Visual hallucinations of multi-modal large language models. arXiv preprint arXiv:2402.14683 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.787, + 0.257 + ], + "angle": 0, + "content": "49. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.785, + 0.311 + ], + "angle": 0, + "content": "50. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.353 + ], + "angle": 0, + "content": "51. Jian, Y., Gao, C., Vosoughi, S.: Bootstrapping vision-language learning with decoupled language pre-training. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.785, + 0.393 + ], + "angle": 0, + "content": "52. Kar, O.F., Yeo, T., Atanov, A., Zamir, A.: 3d common corruptions and data augmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18963-18974 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "53. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3128-3137 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "54. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "55. Korbar, B., Xian, Y., Tonioni, A., Zisserman, A., Tombari, F.: Text-conditioned resampler for long form video understanding. arXiv preprint arXiv:2312.11897 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.556 + ], + "angle": 0, + "content": "56. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.556, + 0.785, + 0.597 + ], + "angle": 0, + "content": "57. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: BLIP: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: International Conference on Machine Learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "58. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "59. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "60. Lin, J., Yin, H., Ping, W., Lu, Y., Molchanov, P., Tao, A., Mao, H., Kautz, J., Shoeybi, M., Han, S.: Vila: On pre-training for visual language models. arXiv preprint arXiv:2312.07533 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.759 + ], + "angle": 0, + "content": "61. Lin, Z., Liu, C., Zhang, R., Gao, P., Qiu, L., Xiao, H., Qiu, H., Lin, C., Shao, W., Chen, K., et al.: Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "62. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.148, + 0.784, + 0.175 + ], + "angle": 0, + "content": "64. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "65. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "66. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "67. Liu, S., Fan, L.J., Johns, E., Yu, Z., Xiao, C., Anandkumar, A.: Prismer: A vision-language model with an ensemble of experts. ArXiv abs/2303.02506 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "68. Lu, J., Clark, C., Lee, S., Zhang, Z., Khosla, S., Marten, R., Hoiem, D., Kembhavi, A.: Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "69. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-IO: A unified model for vision, language, and multi-modal tasks. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.396 + ], + "angle": 0, + "content": "70. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 3195-3204 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "71. McKinzie, B., Gan, Z., Fauconnier, J.P., Dodge, S., Zhang, B., Dufter, P., Shah, D., Du, X., Peng, F., Weers, F., et al.: Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "72. Mizrahi, D., Bachmann, R., Kar, O.F., Yeo, T., Gao, M., Dehghan, A., Zamir, A.: 4M: Massively multimodal masked modeling. In: Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "73. Naeem, M.F., Xian, Y., Zhai, X., Hoyer, L., Van Gool, L., Tombari, F.: Silc: Improving vision language pretraining with self-distillation. arXiv preprint arXiv:2310.13355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.605 + ], + "angle": 0, + "content": "74. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.Q., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y.B., Li, S.W., Misra, I., Rabbat, M.G., Sharma, V., Synnaeve, G., Xu, H., Jégou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINoV2: Learning robust visual features without supervision. ArXiv abs/2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.66 + ], + "angle": 0, + "content": "75. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "76. Raffel, C., Shazeer, N.M., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.785 + ], + "angle": 0, + "content": "77. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "78. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "O.F. Kar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "79. Shen, Y., Heacock, L., Elias, J., Hentel, K.D., Reig, B., Shih, G., Moy, L.: Chatgpt and other large language models are double-edged swords (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "80. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.218, + 0.615, + 0.231 + ], + "angle": 0, + "content": "81. Thorp, H.H.: Chatgpt is fun, but not an author (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "82. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "83. Tiong, A.M.H., Li, J., Li, B., Savarese, S., Hoi, S.C.: Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. arXiv preprint arXiv:2210.08773 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.785, + 0.369 + ], + "angle": 0, + "content": "84. Tong, S., Liu, Z., Zhai, Y., Ma, Y., LeCun, Y., Xie, S.: Eyes wide shut? exploring the visual shortcomings of multimodal llms. arXiv preprint arXiv:2401.06209 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "85. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "86. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "87. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.549 + ], + "angle": 0, + "content": "88. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "89. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.646 + ], + "angle": 0, + "content": "90. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., Wei, F.: Image as a foreign language: BEiT pretraining for all vision and vision-language tasks. ArXiv abs/2208.10442 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "91. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "92. Xu, J., Zhou, X., Yan, S., Gu, X., Arnab, A., Sun, C., Wang, X., Schmid, C.: Pixel Aligned Language Models. arXiv preprint arXiv: 2312.09237 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "93. Yadav, D., Jain, R., Agrawal, H., Chattopadhyay, P., Singh, T., Jain, A., Singh, S.B., Lee, S., Batra, D.: Towards better evaluation systems for ai agents (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "94. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "95. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.285, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BRAVE: Broadening the visual encoding of vision-language models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "96. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "97. Yu, S., Cho, J., Yadav, P., Bansal, M.: Self-chained image-language model for video localization and question answering. arXiv preprint arXiv:2305.06988 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "98. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "99. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "100. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.787, + 0.383 + ], + "angle": 0, + "content": "101. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.383, + 0.787, + 0.424 + ], + "angle": 0, + "content": "102. Zhang, D., Yu, Y., Li, C., Dong, J., Su, D., Chu, C., Yu, D.: Mm-llms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.424, + 0.787, + 0.467 + ], + "angle": 0, + "content": "103. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "104. Zhao, Y., Pang, T., Du, C., Yang, X., Li, C., Cheung, N.M.M., Lin, M.: On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.508, + 0.787, + 0.55 + ], + "angle": 0, + "content": "105. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: iBoT: Image BERT pre-training with online tokenizer. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "106. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BRAVE_ Broadening the visual encoding of vision-language models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_origin.pdf b/2024/BRAVE_ Broadening the visual encoding of vision-language models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..25f38fd21f918cfd32bfceee6c41667ae73a1041 --- /dev/null +++ b/2024/BRAVE_ Broadening the visual encoding of vision-language models/bd8562df-9a3d-4ea9-9664-44ca2f41f576_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d1650b1d1a1fd8a5541429577944e0ade07363f66ae5b03ce8c2a7e1263028a +size 1458884 diff --git a/2024/BRAVE_ Broadening the visual encoding of vision-language models/full.md b/2024/BRAVE_ Broadening the visual encoding of vision-language models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..38b9d1c7a39b9e9f4a28142f0fefaf30e45b2383 --- /dev/null +++ b/2024/BRAVE_ Broadening the visual encoding of vision-language models/full.md @@ -0,0 +1,322 @@ +# BRAVE: Broadening the visual encoding of vision-language models + +Oğuzhan Fatih Kar1,2, Alessio Tonioni1, Petra Poklukar1, Achin Kulshrestha1, Amir Zamir2, and Federico Tombari1,3 + +$^{1}$ Google $^{2}$ Swiss Federal Institute of Technology Lausanne (EPFL) $^{3}$ TU Munich https://brave-vlms.epfl.ch + +# Improving visual capabilities of VLMs with BRAVE + +![](images/275a45e841d9c5e8a9bd0345cc017349e824ce91657b6e6f4df78ba4013d30ac.jpg) +Is there a hand using the mouse in this image? +(a) Yes + +![](images/3cc9d2159ed7352bcc7782dfe32972a6beed12bfbc71df98d0e85c5d0fd1cdfb.jpg) +Are the butterfly's feet visible? +(a) Yes + +![](images/8140a21dcf553b1758543b938e1149a60334dd3dcc11687e31f71664bd2d248e.jpg) +Is the door of the truck open? +(a) Yes +(b) No + +![](images/bf4e78dc2116e5408bee825b78eba03968f4cb69291d468c78e18adc9b7c42ac.jpg) +InstructBLIP (a) (b) (EVA Encoder) +BRAVE (a) (b) (Multiple Encoders) + +![](images/b117e0eed6c3be4ce2226d4bd53a70bea0295ca6f45c8ca4652199a5c775d819.jpg) +(a) (b) + +![](images/f989f58577877184e44fa199a8243811051743e678ae65762159580486d2004b.jpg) +(a) (a) + +![](images/1ff5d117d7eeb0caabe9a75c7786d2668c66042e25e0c051be06ca598ae1e526.jpg) +Fig. 1: We propose BRAVE to broaden the visual capabilities of vision-language models (VLMs). Left: In contrast to existing methods, e.g. InstructBLIP [25] or LLaVA-1.5 [64], that use a single vision encoder [59,84], BRAVE combines diverse features from multiple vision encoders into a more versatile and compact representation. The examples are taken from [84] and assess the VLM's ability to differentiate images with visual differences. Right: BRAVE leads to state-of-the-art performance on a wide range of captioning and visual question answering tasks. Furthermore, it significantly improves the performance on benchmarks [84] where commonly employed vision encoders fail. +SoTA Performance in Captioning & VQA + +Abstract. Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. "blindness" to certain image features, visual hallucination, etc. To address these issues, we study broadening the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features + +from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs. + +Keywords: Vision-Language $\cdot$ Visual Encoding $\cdot$ Multi-modal Learning + +# 1 Introduction + +Vision-language models (VLMs) have recently seen significant improvements on solving tasks requiring both visual and text understanding capabilities such as captioning, visual question answering (VQA), and instruction following. These advancements are fueled by the progress in uni-modal vision encoders [30, 75] and language models (LMs) [22, 24], which are then combined with different "bridging" techniques [3,56,66] to create VLMs. Improved training recipes [25,64] and scalability efforts in terms of training data and model size [6,19,77,87,99] further boost the performance. + +Despite the progress, VLMs are subject to several shortcomings: on the language side, the LMs are known to be susceptible to hallucinations and logical faults [8,37,79,81], while on the vision side, they are limited by the capabilities of the vision encoder. For example, Tong et al. [84] observed that commonly employed vision encoders such as CLIP [75] can exhibit "blindness", i.e. fail to differentiate images with clear visual differences (See Fig. 1, left). Similarly, Li et al. [59] showed that VLMs are prone to visual hallucinations where the models imagine incorrect details about a given image. These limitations create a bottleneck for developing performant and visually grounded VLMs. + +To address these, we study broadening the visual encoding capabilities of VLMs. It is known in machine learning that different representations of an input can yield different generalization properties, and each can play a role when ensembled to create a more complete representation [33]. Inspired by this, we first perform a comprehensive evaluation of VLMs which differ only in their vision encoders, i.e. visual biases. For this, we consider encoders with different biases, e.g. different objectives, training data, and model sizes. Our results in Table 2 show that different encoders lead to varying performance across vision-language tasks and there is no single encoder achieving a consistent top performance across tasks. Furthermore, we find that encoders with different biases can perform surprisingly similarly, suggesting that there could be multiple cues [35,45] that can be leveraged to solve the vision-language tasks. + +Motivated by these findings, we propose to employ various vision encoders for VLMs and introduce a method to learn how to combine them efficiently. We denote the method as BRAVE, which stands for broadening the visual encoding of VLMs. As shown in Fig. 2, BRAVE combines features from an arbitrary number + +of vision encoders into a compressed fixed-length visual representation which is then fed as a soft visual prompt to a frozen LM. We achieve the bridging of encoders and the LM by introducing a lightweight multi-encoder querying transformer (MEQ-Former), which is a multi-encoder generalization of the Q-Former proposed in the single encoder BLIP-2 framework [56]. In particular, MEQ-Former accepts a textual prompt and a set of learnable queries as inputs and jointly cross-attends to the features from different vision encoders (Fig. 2). We provide a simple yet effective recipe in Sec. 3 for training the MEQ-Former in a single stage that bypasses the two-stage pre-training paradigm of [56]. Notably, this is achieved with a smaller number of trainable parameters during pre-training than the existing methods [3,6,19,25,61,64,87] (Sec. 4). + +Our evaluation on a wide range of captioning and VQA tasks shows that BRAVE effectively consolidates diverse visual signals into a broad and contextual representation, leading to consistently better performance over the state-of-the-art and improved robustness against out-of-distribution inputs, as shown in Fig. 1 and Sec. 4. Furthermore, in contrast to previous works that benefit from scaling the language axis [3, 19, 25, 64] by using larger LMs, our work demonstrates that scaling along the vision axis has a significant potential for VLMs. + +Our contributions can be summarized as follows: + +- We conduct a systematic analysis of several vision encoders with different inductive biases, e.g. training data, objective, model size, on solving vision-language tasks under a unified training and evaluation framework. +- We introduce BRAVE, a method that efficiently consolidates features from any number of vision encoders into a compressed and contextual representation, leading to 1) state-of-the-art performance on several captioning and VQA tasks and 2) significantly improved robustness against visual hallucinations and out-of-distribution inputs. +- We perform a comprehensive ablation study of BRAVE, highlighting the impact of the design choices, which we hope provide useful insights to the researchers for further advancements in VLMs. + +# 2 Impact of vision encoders for vision-language models + +To quantify the impact of visual biases on the performance of VLMs, we compare VLMs with different vision encoders on commonly evaluated VQA and captioning tasks. For this, we develop a pre-training, fine-tuning and evaluation setup, as explained next. To the best of our knowledge, this is the first unified and comprehensive evaluation of different vision encoders for vision-language understanding and generation tasks. + +VLM architecture. We use a frozen vision encoder which is connected to a frozen language model by a bridge network with trainable parameters. For the bridge, we adopt the Q-Former module from [56]. This choice is based on the following reasons: 1) the Q-Former resamples visual features to a fixed-length + +Table 1: Overview of vision encoders we benchmarked. All encoders use a ViT [27] backbone, yet they differ in terms of objective, training data, and model size. ITC: Image-text contrastive learning [75], MIM: Masked image modelling [41], LGC: Local-to-global correspondence learning [11]. See Sec. 2 for details. + +
Vision encoderParametersTraining dataObjective
CLIP-L/14 [75]0.3BOpenAI WITITC
OpenCLIP-G/14 [21]1.8BLAION-2BITC
EVA-CLIP-g [30]1BLAION-400MMIM (CLIP features)
SIGLIP-G/14 [100]1.8BWebLIITC (Sigmoid loss)
SILC-G/16 [73]1.8BWebLIITC + LGC
ViT-e [19]3.8BJFT-3BClassification
ViT-G [99]1.8BJFT-3BClassification
DINOv2-L/14 [74]0.3BLVD-142MLGC + iBOT [105]
+ +output before feeding them to the LM which enables a fair comparison of vision encoders with different output dimensionalities, 2) the Q-Former is efficient to train and evaluate as it is compatible with frozen pre-trained vision encoders and LMs. For the vision encoder, we consider a diverse set of options as described in the next paragraph. For the LM, we follow [56] and use FlanT5-XL [24]. The Q-Former has 110M parameters consisting of a series of cross- and self-attention layers and a fixed number of learnable queries. Unless specified otherwise, we use 32 queries of dimension 768 which is also the hidden dimension of the Q-Former. The queries interact with visual features through cross-attention layers, enabling a significant reduction in their dimensionality. For example, CLIP-L/14 [75] visual features are reduced from $257 \times 1024$ to $32 \times 768$ . The output of the Q-Former is then linearly projected to the textual embedding space of the LM to create a soft visual prompt, and then prepended to the textual prompt that specifies the task. Both visual and textual prompts are fed as inputs to the frozen LM. For captioning, we use the textual prompt "A photo of" and for VQA, we directly use the question as the prompt. Similar to InstructBLIP [25], we give the textual prompt as additional input to the Q-Former together with the learnable queries to extract more task-aligned visual prompts. + +Vision encoders. As summarized in Table 1, we consider eight recently introduced vision encoders: CLIP [75], OpenCLIP [21], EVA-CLIP [30], SIGLIP [100], SILC [73], ViT-e [19], ViT-G [99], and DINOv2 [74]. While they all use ViT-based backbones [27], they differ in terms of 1) training data, e.g. LAION [77], WebLI [19], LVD-142M [74], etc., 2) training objective, e.g. image-text contrastive learning, masked image modelling, classification, etc., and 3) model size, e.g. 300M to 4B parameters. Due to this diversity, they incorporate different biases and thus potentially capture different aspects of the underlying depicted scene. + +Pre-training data and objectives. We pre-train the Q-Former using WebLI [19] dataset at $224 \times 224$ resolution. We use the filtered and de-duplicated + +Table 2: Benchmarking vision encoders on different vision-language tasks. For COCO captioning, we report CIDEr [86] score. For VQA tasks, we report top-1 accuracy. For MMVP, we report average pair accuracy [84]. Top-3 best results are shown with Blue, Green, and Orange, respectively. All VLMs use FlanT5-XL [24] as the language model and a Q-Former [56] to bridge vision and language modalities. See Sec. 2 for details. + +
Vision encoderCOCO cap. ↑ Karpathy valVQAv2 ↑ Karpathy valOKVQA ↑ valGQA ↑ test-devMMVP ↑ test
CLIP-L/14133.074.461.048.715.3
OpenCLIP-G128.373.360.648.022.0
EVA-CLIP-g140.977.063.050.127.3
SIGLIP-G/14133.074.762.548.624.0
SILC-G/16141.177.063.449.724.0
ViT-e137.875.661.949.125.3
ViT-G133.874.261.248.320.7
DINOv2-L/14127.671.359.048.022.0
+ +English-only subset with 100 million image-text pairs [19, 73]. We keep both the vision encoder and the LM frozen and only train the parameters of the Q-Former. The model is trained with captioning objective using the alt-text in WebLI as target, as it has been shown to be effective [92]. Similar to [56], we remove the last Transformer layer of the encoder. We refer the reader to the supplementary material for additional training details and hyperparameters. + +Evaluation tasks. We evaluate the obtained VLMs on standard captioning and VQA tasks. For captioning, we use the COCO captioning benchmark [20] and fine-tune the pre-trained VLM on the commonly employed Karpathy training split [53]. Similar to the pre-training, only the Q-Former parameters are updated. For VQA, we again follow the standard practices [19, 25, 56, 64] and fine-tune on a mixture of VQA data that enforces the Q-Former to extract more task-aligned visual features. Our mixture includes the training sets of VQAv2 [36] and OKVQA [70] as well as the synthetically generated VQA data from $\mathrm{VQ^2A}$ [12], resulting in total 17M training examples. We then evaluate the fine-tuned models on VQAv2 and OKVQA validation sets. Similar to [25, 56], we also evaluate the VLMs' zero-shot VQA capabilities on GQA [48]. Finally, we evaluate the zero-shot performance on MMVP [84] that includes images with semantic differences that are challenging for the current VLMs to handle. Unlike captioning, we fine-tune both Q-Former and the LM parameters while keeping the vision encoder frozen, as we find this to improve the performance, similar to [64]. For both captioning and VQA tasks, we use $224\times 224$ image resolution. See the supplementary for more details and results. + +Results. We report the findings in Table 2 and make the following observations: + +- Encoders with different biases can obtain similar performance. While we observe some variation in VLMs' performance across different encoders, the gap remains small, especially among the top-performing ones. For example, EVA-CLIP-g, SILC-G/16, and ViT-e achieve the best results for VQAv2, GQA and COCO captioning despite their differences in training objective and data, as shown in Table 1. Similar observations can be made for OpenCLIP and DINOv2 encoders that have a small performance gap. +- MMVP stays challenging for all encoders. While it is originally curated with "CLIP-blind" pairs in [84], i.e. images that CLIP perceives as similar despite their visual differences (See Fig. 1 for examples), most of the encoders stay below the random guess accuracy of $25\%$ . Exceptions are EVA-CLIP and ViT-e which slightly surpass the random guess level ( $27.3\%$ and $25.3\%$ ). +- For tasks requiring compositional reasoning and open-world knowledge, the performance decreases and the gap among VLMs diminishes, as shown by the results on GQA and OKVQA. More precisely, the standard deviation across all the VLMs' performance on GQA and OKVQA is 0.8 and 1.36, respectively, but 1.74 on VQAv2 and 4.91 on COCO. +- Scaling the size of vision encoder can improve the performance as seen by the consistent improvements obtained by ViT-e over ViT-G when scaling from 1.8B to 3.8B parameters. +- Pre-training data distribution can play a critical role. OpenCLIP-G/14 model is larger than CLIP-L/14 and uses more training data, yet it noticeably underperforms on most of the evaluated VQA and captioning tasks except the MMVP benchmark, where OpenCLIP significantly outperforms the CLIP model. Since both models are trained using the same training objective, this indicates that the impact of the training objective and dataset to the performance of the VLM can be disentangled, and both are important. +- Performance correlates between VQA and captioning tasks. Encoders with better performance on VQA tasks generally also perform better on captioning, and vice versa. Spearman's rank correlation between VQA tasks and COCO captioning is 0.90, 0.80, 0.79, 0.56 for VQAv2, OK-VQA, GQA, and MMVP, respectively. Thus, improving the visual component in VLMs is expected to improve performance for a broad range of tasks. + +The above results highlight the importance of using the right vision encoder for VLMs, motivating us to focus on the following question: Can we broaden the vision capabilities of VLMs through combining vision encoders with different visual biases? We answer positively in the next section. + +# 3 Combining vision encoders with BRAVE + +Motivated by the findings in Sec. 2, we propose using multiple vision encoders with different, and potentially complementary, visual biases to create more capable VLMs. To do this, we introduce BRAVE, a method that combines the strengths of different vision encoders while staying efficient in terms of number of trainable parameters. Below, we provide details of BRAVE. + +![](images/db020df293b626c30f01d0e6866cc2becb612f2e138bec09372ce27aecc1fc3f.jpg) +Fig. 2: Overview of BRAVE. Left: We keep all the vision encoders (VEs) and the language model (LM) frozen. The linear projection layers are used to concatenate features from $K$ different VEs, e.g. $K = 5$ , sequence-wise. These are then resampled by the MEQ-Former which accepts a set of learnable queries and a text prompt describing the task as inputs. The output of MEQ-Former is projected to the input space of the LM using fully-connected (FC) layers. The total number of trainable parameters is 116M ( $\approx 1\%$ of the total parameters). Right: Architecture of the MEQ-Former with $N = 12$ transformer layers. It interacts with the concatenated features through crossattention layers and produces a fixed-length output to be fed as input to the LM. + +![](images/1f32267e7f2c26754790dc029424ed0747f84db25d18d015b9c239c6ff803703.jpg) + +Overview. As depicted in Fig. 2, we introduce the multi-encoder querying transformer, or MEQ-Former, that combines visual features from an arbitrary set of encoders and outputs a fixed-length visual representation that is given as a soft prompt to a frozen LM. It takes as an input a sequence comprised of a fixed number of learnable queries and embedded text tokens which describe the task - for captioning, they are the textual prompt "A photo of", while for VQA, they are the question prompt for a given image (See supplementary for examples). MEQ-Former interacts with visual features through cross-attention layers. The visual features from different encoders are first linearly projected to the same dimension and then concatenated sequence-wise. The obtained concatenated feature is given as a key and value pair to the cross-attention layers in MEQ-Former, and is cross-attended to by the MEQ-Former's query sequence (Fig. 2). This resampling enables to efficiently processing a large number of visual features since it bypasses the quadratic complexity of self-attention. It also acts as a "bottleneck" that keeps the total number of VLM's parameters low compared to a naive ensembling of VLMs that is prohibitively expensive (See a comparison in Sec. 4.3). Moreover, the visual features are not "marked" with any encoder-specific embedding, thus MEQ-Former is not biased to differentiate between encoders, allowing for a simpler design. + +Differences with previous work. Our work differs from other popular approaches like LLaVa [64] and PaLI [19] that require significantly more trainable parameters ( $\approx 10\mathrm{B}$ ) and do not have a resampling mechanism. Resampling visual + +features from a single encoder has been successfully demonstrated in previous work for single- [25,56] and multiple-frames [3,55,97]. To the best of our knowledge, our work is the first one demonstrating an effective resampling mechanism to consolidate visual features from several arbitrarily different encoders while keeping the number of trainable parameters small (our MEQ-Former having 116M parameters vs BLIP-2's Q-Former is 188M). See Sec. 5 for more comparisons. + +Pre-training details. We use the same data and objective as in Sec. 2 and only train the MEQ-Former while keeping the vision encoders and the LM frozen. During pre-training, we randomly mask features from each encoder with $20\%$ probability as we found that this can act as a regularizer and enforces MEQ-Former to avoid local minima by attending only to the features of a single encoder (Ablated in Sec. 4.3). After pre-training, we fine-tune the MEQ-Former and (optionally) the LM on the downstream tasks, as explained in Sec. 4. + +Architecture. For the main results, we combine five vision encoders, namely EVA-CLIP- $g$ [30], CLIP- $L/14$ [75], SILC- $G/16$ [73], ViT- $e$ [19], and DINOv2- $L/14$ [74] to cover all training datasets and objectives from Table 1. Please see Sec. 4.3 and the supplementary for analysis on the contributions of encoders. For the LM, we employ FlanT5-XL [24] as before. We set the visual feature dimension after linear projection as 1408 for concatenation. For MEQ-Former, we use $32 \times 5 = 160$ learnable queries to proportionally scale the model capacity to the number of vision encoders and set the hidden dimension as 768. The MEQ-Former resampling results in $14 \times$ reduction in total feature size (from $1223 \times 1408$ to $160 \times 768$ ). The total number of trainable parameters during pre-training is 116M which is about $1\%$ of the total number of parameters in the VLM (10.3B). + +Implementation. We use FLAX [42] with the Scenic [26] library to implement our models and training pipeline. For evaluations, we use both Scenic and EvalAI API [93]. We pre-train with batch size of 1024 on 64 TPUv5 chips. + +# 4 Experiments + +We perform evaluations on a broad range of captioning (Sec. 4.1) and VQA tasks (Sec. 4.2) to show the effectiveness of BRAVE. Please see Fig. 1 right for a summary of the results and supplementary for an overview of the tasks. We then provide a comprehensive analysis of BRAVE in Sec. 4.3. + +For captioning, we evaluate on COCO [20] and NoCaps [2] benchmarks. The latter includes out-of-distribution samples with novel classes that do not exist in the COCO benchmark, hence it serves as a challenging testbed to assess robustness. The results (Table 3) show that BRAVE pushes the state-of-the-art for both benchmarks while staying efficient in terms of trainable parameters. + +For VQA, we evaluate on VQAv2 [36], OKVQA [70], GQA [48], and VizWiz-QA [39] benchmarks. To assess the robustness of VLMs, we additionally include + +Table 3: Captioning. We compare BRAVE with state-of-the-art captioning methods on COCO and NoCaps (zero-shot). CIDEr [86] score is reported (Best in bold, second best is underlined). Numbers of other methods are taken from the corresponding publications. Dashed line means no result was reported. BRAVE achieves the best results in NoCaps evaluation sets (both out-domain and overall in validation & test) and second best for COCO, while having much fewer trainable parameters than other methods. For example, compared to PaLI-17B [19], BRAVE uses significantly fewer trainable parameters, training data, and image pixels. Yet, it outperforms the baselines on NoCaps and stays competitive on COCO, demonstrating the effectiveness of having different visual biases for generalization. See Sec. 4.1 for more details. + +
Method# paramsCOCO (fine-tuned)NoCaps (zero-shot, val)NoCaps (zero-shot, test)
TrainableTotalKarpathy testout-domainoverallout-domainoverall
Flamingo [3]10.6B80B138.1----
SimVLM [91]632M632M143.3113.7112.2-110.3
Qwen-VL [6]9.6B9.6B--121.4--
BLIP-2 [56]1.1B4.1B144.5124.8121.6--
InstructBLIP [25]188M14.2B--121.9--
CoCa [96]2.1B2.1B143.6-122.4-120.6
GiT2 [87]5.1B5.1B145.0130.6126.9122.3124.8
PaLI-17B [19]16.9B16.9B149.1-127.0126.7124.4
BRAVE116M10.3B148.0133.3127.6127.1125.6
+ +POPE [59] that measures object hallucinations, and MMVP [84] which exposes the shortcomings of CLIP-based vision encoders that are widely used in VLMs. Our results (Table 4) show that BRAVE achieves consistent improvements over recent methods for all the tasks by leveraging diverse visual features. + +# 4.1 Image captioning + +Fine-tuning details. Similar to Sec. 2, we only fine-tune the MEQ-Former on COCO captioning and keep the vision encoders and the LM frozen. During fine-tuning, we set a higher input resolution as $336 \times 336$ to further boost the performance, similar to previous works [19,56,64] (See Sec. 4.3 for an ablation). Following [19,25,56,87,96], we also zero-shot evaluated the COCO fine-tuned model on NoCaps [2]. Please see the supplementary for more training details. + +Results. We summarize the captioning evaluations in Table 3. BRAVE uses the least number of trainable parameters (116M) yet achieves strong results for both COCO and NoCaps benchmarks. For NoCaps, BRAVE is the best performing method with significant gains over recent methods. This is especially the case for out-domain samples with novel classes, demonstrating the usefulness of diversity in visual features for robustness. For COCO, BRAVE stays competitive with the best performing model, PaLI-17B [19]. Notably, BRAVE achieves this while using $150 \times$ fewer trainable parameters (116M vs 16.9B), $16 \times$ less pre-training data (100M vs 1.6B) and $3 \times$ fewer image pixels ( $336 \times 336$ vs $588 \times 588$ ) than PaLI-17B, suggesting that having different visual biases is effective for generalization, while keeping the sample complexity low. + +Table 4: Visual question answering. We compare BRAVE with state-of-the-art VQA methods on VQAv2, OKVQA, GQA, VizWiz-QA, MMVP, and POPE. Top-1 accuracies are reported (Best in bold, second best is underlined). For MMVP, we report average pair accuracy [84]. Numbers of other methods are taken from the corresponding publications. Dashed line means no result was reported. BRAVE achieves the best results for six out of seven benchmarks, while staying efficient in terms of trainable parameters. Similar to the COCO captioning results in Table 3, BRAVE stays competitive with PaLI-17B [19] for VQAv2, while using significantly less pre-training data, image resolution, and model parameters (both trainable and total). See Sec. 4.2 for details. + +
Method# paramsFine-tunedZero-shot
TrainableTotalVQAv2 test-devOKVQA valGQA test-devVizWiz-QA test-devGQA test-devMMVP testPOPE test
SimVLM [91]632M632M80.0------
Flamingo [3]10.2B80B82.057.8-31.6---
MiniGPT-v2 [14]7B8B-57.860.153.6---
GiT2 [87]5.1B5.1B81.7------
Qwen-VL [6]9.6B9.6B79.558.659.335.2---
SPHINX-2k [61]13B16.5B80.762.663.144.9--87.2
PaLI-17B [19]16.9B16.9B84.364.5-----
BLIP-2 [56]1.2B12.1B81.654.7-29.444.7-85.3
InstructBLIP [25]188M14.2B-55.5-33.449.516.778.9
ShareGPT4V [16]13.4B13.4B81.0-64.8----
LLaVa1.5 [64]13B13.4B80.0-63.353.6-24.785.9
LLaVa1.6 [65]13B13.4B-46.365.4---86.3
LLaVa1.5 (I-MoF) [84]13B13.6B79.3----31.386.7
BRAVE3B10.3B82.566.066.354.252.742.087.6
+ +# 4.2 Visual question answering + +Fine-tuning details. We use the same VQA mixture as in Sec. 2 and fine-tune both the MEQ-Former and the LM at $224 \times 224$ resolution. Vision encoders are kept frozen. This is followed by a high-resolution fine-tuning stage on VQAv2 and OKVQA training sets at $336 \times 336$ resolution, similar to previous work [19,56,64]. The resulting model is evaluated on VQAv2 and OKVQA benchmarks as well as zero-shot evaluated on GQA [48], VizWiz-QA [39], MMVP [84] and POPE [59], following [25,56,64,84]. To be able to compare to the methods that fine-tune on GQA training set, e.g. [6,61,64], we also perform an additional training stage on top of the mixture-finetuned model, and evaluate the model on the GQA evaluation set (test-dev). Please see the supplementary for the details. + +Results. We summarize the VQA evaluations in Table 4. BRAVE achieves strong results for all the benchmarks in both fine-tuned and zero-shot evaluation scenarios. For OKVQA, GQA, VizWiz-QA, MMVP, and POPE benchmarks, BRAVE achieves the best results, and for VQAv2, it stays competitive with PaLI-17B: + +- For MMVP, BRAVE significantly improves the performance over previous approaches, making it less susceptible to failure modes of CLIP-based encoders. This can be further seen from the qualitatives in Figures 1 and 3. +- For VizWiz-QA, BRAVE achieves the best results without using any scene text understanding or OCR data during pre-training, as opposed to [6,14,64]. + +![](images/bfda9a6e63e1900bbaf032351484f0a6836e6058b510ba6dcf807e09244f6d5e.jpg) +Fig. 3: Qualitative results. We compare predictions of BRAVE and the VLMs with different vision encoders, e.g. CLIP, on samples from the MMVP benchmark. Following [84], a model is considered correct only if it answers to both images in a pair correctly, i.e. if it can successfully differentiate between images with semantic differences. Note that the images in a pair are seen independently, i.e. neither of the images is provided as context for the other one. All encoders output some correct predictions, yet none of them performs consistently well on a broad range of inputs. BRAVE alleviates this by combining diverse visual features, leading to a more consistent performance. The quantitative difference is indeed stark: $42\%$ for BRAVE vs $27.3\%$ for best single encoder (Tables 2 and 4). See supplementary for more qualitative results. + +- For GQA and OKVQA that requires different capabilities, i.e. spatial reasoning and open-world knowledge, BRAVE noticeably improves the performance. +- For POPE, BRAVE further reduces the visual hallucinations over dual encoder methods [61,84], confirming that using multiple encoders can be effective for visual grounding (Also see Sec. 5 for a discussion). +- For VQAv2, BRAVE is the second best after PaLI-17B, while staying more efficient in terms of training data, model size, and input resolution, similar to the COCO captioning result in Sec. 4.1. We expect BRAVE to benefit from such scaling efforts as well, which is a future direction to investigate. + +# 4.3 Analysis of BRAVE + +We study different aspects of BRAVE including an ablation study, contribution of vision encoders to the final performance, comparing MEQ-Former to an ensembling strategy, and the role of pre-training data. Due to the space limit, please see the supplementary for more qualitative and quantitative results. + +Table 5: Ablation study of BRAVE. We start from the base pre-training setup explained in Sec. 3, denoted as A0. Each row corresponds to a particular deviation from A0, while keeping the rest fixed. For example, A5 evaluates the performance when high-resolution fine-tuning is disabled. N/A means the ablation is not applicable for that evaluation. See Sec. 4.3 for the discussions and supplementary for more results. + +
Ablation IDStudy SubjectOriginalChangeCOCO Captioning (Karpathy val)VQAv2 (Karpathy val)OKVQA (val)
A0Base--147.081.865.7
A1LM fine-tuning for VQAFull fine-tuningNo fine-tuningN/A78.657.5
LoRA (r=64)N/A80.762.8
LoRA (r=128)N/A81.062.9
A2Synthetic VQA data [12]✓✗N/A81.164.0
A3Encoder dropout✓✗145.381.366.0
A4MEQ-Former text input✓✗145.981.464.9
A5High-res fine-tuning✓✗145.279.665.0
A6A3 + A5✓✗144.079.064.5
A7A4 + A5✓✗145.178.363.4
A8Language modelFlanT5-XLFlanT5-L142.579.965.5
+ +Ablation study. We perform a comprehensive study in Table 5 where we ablate different design choices, e.g. training data and choice of language model. The ablations are labeled from A1 to A8, each showing a particular deviation from the original pre-training setup A0 (Sec. 3). We summarize the key findings: (A1): LM fine-tuning for VQA significantly boosts the performance compared to frozen LM, similar to the observation in [64]. To minimize the performance loss, we perform a LoRA [46] fine-tuning by inserting LoRA layers to the LM. With a rank of 128, we compensated for $70\%$ of the performance gap while using $10\times$ fewer trainable parameters than full fine-tuning (324M vs 3B). This underscores the development of parameter-efficient fine-tuning methods as an important direction for more efficient VLMs. + +(A2): Using synthetic VQA data [12] noticeably enhances our performance, while being much cheaper to collect than human annotated data. This can potentially be further improved by using more diverse templates [38, 83]. + +(A3,A4,A5): Encoder dropout during training improves the performance for COCO captioning and VQAv2 but slightly degrades for OKVQA. Using text prompt as an additional input to MEQ-Former helps with extracting better task-aligned features. Similarly, high-resolution fine-tuning stage gives a boost to the performance at the expense of processing more visual tokens. + +(A6,A7): These ablations combine the impact of previously ablated design choices, further demonstrating their contribution to the final performance. + +(A8): A stronger LM significantly helps for all the tasks, underscoring that scaling the VLM along the vision and language axes simultaneously have complementary benefits, and both are important. + +Contribution of vision encoders. To understand the impact of vision encoders to the final performance of BRAVE, we perform additional evaluations: + +![](images/7112a9140dd9ea5ba4c641a7c29b3ca11f169f479fde8f9f890463b9b1294523.jpg) +Fig. 4: Contribution of vision encoders to BRAVE. Left: We analyze the robustness of BRAVE when a subset of encoders are removed during evaluation. We report the average drop in CIDEr for COCO and accuracy for VQAv2. Right: We compute average attention scores for different vision encoders cross-attended by the MEQ-Former for COCO and VQAv2. See Sec. 4.3 for the discussions. + +- In Fig. 4-left, we perform a robustness analysis by removing a subset of vision encoders during inference. This is performed for all the combinations and we report the average drop in performance for COCO captioning and VQAv2 tasks (Full results in supplementary). The results show that the performance of BRAVE degrades gracefully up to two encoder removals for both tasks, indicating some redundancy among the encoders. Because of this, even if some encoders are missing, BRAVE can make use of the remaining set of encoders to compensate. Beyond the two encoder removal, the drop in the performance is more severe, suggesting that some features are more unique and consequently harder to replace using the remaining set. + +- In Fig. 4-right, we compute attention scores of features from different encoders that are cross-attended by the MEQ-Former queries. These are averaged for randomly sampled 1k samples from COCO and VQAv2 validation sets. The results show that the MEQ-Former can cross-attend, or "pay attention to", some encoders more than the others depending on the downstream task adaptively, further confirming the usefulness of having different biases. + +# 5 Related work + +Several works have focused on training VLMs to solve tasks requiring both vision and text understanding capabilities. Most of these works make use of pre-trained vision encoders [9,27,31,49,75,99] and LMs [6,22,24,44,76,85,103]. Architecture-wise, different approaches were developed to project visual features to the textual embedding space, e.g. Q-Former [25, 56, 106], P-Former [51], cross-attention [3, 6, 19, 95], linear projector [14, 15, 66] and MLP [64, 89]. Different objectives were also considered depending on the downstream tasks, e.g. image-text contrastive learning [50,73,75,94,101], auto-regressive text generation [17,19,23,87,88,91] or both [56-58,96] (see [102] for a survey). In contrast to ours, these works consider a single vision encoder which has inherent limitations [47,59,82,84] (Sec. 2). + +Concurrent to our work, LLaVa-MoF [84] proposes combining CLIP [75] and DINOv2 [74] encoders using separate adapters. SPHINX [61] mixes CLIP, DINOv2, and Q-Former outputs with linear projection layers. Both of the methods + +input a concatenation of the extracted visual embeddings to the LM, which is not scalable to multiple encoders. We differ from these works by proposing a flexible and unified resampling mechanism that combines features from several vision encoders while staying efficient and improving the performance (Tables 3, 4). + +Our work focuses on scaling the VLMs along the vision axis, unlike other works that explored scaling homogenously model size across the visual and/or language component [1,3,4,18,19,28,34,80,90], training data [6,10,13,49,50,60, 71,78,87,89] and number of tasks and modalities [5,7,40,43,67-69,72,98]. BRAVE can be combined with these efforts to further improve performance. Moreover, several studies investigated the visual aspect of VLMs by evaluating them against hallucinations and other robustness issues [32,47,59,62,63,82,104]. BRAVE improves the robustness to those issues by broadening visual capabilities of VLMs. + +# 6 Conclusions and Limitations + +We focus on broadening the visual encoding aspect of VLMs. Our evaluation of vision encoders revealed that there is no encoder achieving a consistent top performance across tasks and encoders with different biases can perform similarly. Motivated by this, we introduce BRAVE that combines diverse features from several vision encoders using a small number of trainable parameters. Extensive evaluations show that BRAVE achieves state-of-the-art performance and improves robustness of VLMs against out-of-distribution inputs and visual hallucinations. We discuss limitations of our work and future directions below: + +Adaptive mechanisms. While MEQ-Former can resample from the useful features and discard the irrelevant ones, it still needs forward passes from all the encoders. A future direction could be to explore adaptive mechanisms to pre-select which encoders to resample from, reducing the inference cost. + +Improving sample efficiency. BRAVE uses around 100M image-text pairs during pre-training which is an order of magnitude less than recent methods [6, 19, 87], yet can benefit from further lowering the sample complexity similar to [51]. More broad vision encoders. The set of encoders we investigated do not completely cover all types of visual biases, e.g. including the ones with strong 3D [29, 52] or semantic [54] priors could further improve the performance. Our results show that the current set is a good starting point to see the benefits of consolidation. Another interesting direction is to distill from the MEQ-Former-combined features back into a single encoder for a more unified solution. + +Different modalities and multiple frames. While we focus on image and text modalities, BRAVE is a general method that could be potentially used to fuse more modalities, e.g. audio or 3D, or extended to incorporate multiple frames, e.g. for video understanding or in-context learning. + +Biases of text-generative models. As we employ pre-trained LMs in our setup, the resulting VLMs are susceptible to the longstanding bias and fairness issues of these models, thus they should be used with caution. + +Acknowledgements. We thank Diego Martin Arroyo, Ferjad Naeem, Xingyi Zhou, Yannick Strümpler and Yongqin Xian for their help with the project. + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8948-8957 (2019) +3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023) +5. Bachmann, R., Kar, O.F., Mizrahi, D., Garjani, A., Gao, M., Griffiths, D., Hu, J., Dehghan, A., Zamir, A.: 4m-21: An any-to-any vision model for tens of tasks and modalities. arXiv preprint arXiv:2406.09406 (2024) +6. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023) +7. Bai, Y., Geng, X., Mangalam, K., Bar, A., Yuille, A., Darrell, T., Malik, J., Efros, A.A.: Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785 (2023) +8. Bang, Y., Cahyawijaya, S., Lee, N., Dai, W., Su, D., Wilie, B., Lovenia, H., Ji, Z., Yu, T., Chung, W., et al.: A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023 (2023) +9. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: International Conference on Machine Learning. pp. 1059-1071. PMLR (2021) +10. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022) +11. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +12. Changpinyo, S., Kukliansky, D., Szpektor, I., Chen, X., Ding, N., Soricut, R.: All you may need for vqa are image captions. arXiv preprint arXiv:2205.01883 (2022) +13. Changpinyo, S., Sharma, P.K., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3557-3567 (2021) +14. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023) +15. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023) + +16. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023) +17. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. In: International Conference on Learning Representations (2022) +18. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Changpinyo, S., Wu, J., Ruiz, C.R., Goodman, S., Wang, X., Tay, Y., et al.: Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565 (2023) +19. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022) +20. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) +21. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023) +22. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023) +23. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021) +24. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) +25. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=vvoWPYqZJA +26. Dehghani, M., Gritsenko, A., Arnab, A., Minderer, M., Tay, Y.: Scenic: A jax library for computer vision research and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21393-21398 (2022) +27. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021) +28. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023) +29. Eftekhar, A., Sax, A., Bachmann, R., Malik, J., Zamir, A.R.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 10766-10776 (2021) +30. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023) + +31. Fang, Y., Wang, W., Xie, B., Sun, Q.S., Wu, L.Y., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA: Exploring the limits of masked visual representation learning at scale. ArXiv abs/2211.07636 (2022) +32. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) +33. Geman, S., Bienenstock, E., Doursat, R.: Neural networks and the bias/variance dilemma. Neural computation 4(1), 1-58 (1992) +34. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790 (2023) +35. Goodwin, C.: Seeing in depth. Social studies of science 25(2), 237-274 (1995) +36. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6904-6913 (2017) +37. Guo, B., Zhang, X., Wang, Z., Jiang, M., Nie, J., Ding, Y., Yue, J., Wu, Y.: How close is chatgpt to human experts? comparison corpus, evaluation, and detection. arXiv preprint arXiv:2301.07597 (2023) +38. Guo, J., Li, J., Li, D., Tiong, A.M.H., Li, B., Tao, D., Hoi, S.: From images to textual prompts: Zero-shot visual question answering with frozen large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10867-10877 (2023) +39. Gurari, D., Li, Q., Stangl, A.J., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3608-3617 (2018) +40. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023) +41. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021) +42. Heek, J., Levskaya, A., Oliver, A., Ritter, M., Rondepierre, B., Steiner, A., van Zee, M.: Flax: A neural network library and ecosystem for jax, 2020. URL http://github.com/google/flax1 (2020) +43. Herzig, R., Abramovich, O., Ben Avraham, E., Arbelle, A., Karlinsky, L., Shamir, A., Darrell, T., Globerson, A.: Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6803-6815 (2024) +44. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022) +45. Howard, I.P., Rogers, B.J.: Binocular vision and stereopsis. Oxford University Press, USA (1995) +46. Hu, J.E., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Chen, W.: LoRA: Low-rank adaptation of large language models. ArXiv abs/2106.09685 (2021), https://api-semanticscholar.org/CorpusID:235458009 +47. Huang, W., Liu, H., Guo, M., Gong, N.Z.: Visual hallucinations of multi-modal large language models. arXiv preprint arXiv:2402.14683 (2024) + +48. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019) +49. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below. +50. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021) +51. Jian, Y., Gao, C., Vosoughi, S.: Bootstrapping vision-language learning with decoupled language pre-training. Advances in Neural Information Processing Systems 36 (2024) +52. Kar, O.F., Yeo, T., Atanov, A., Zamir, A.: 3d common corruptions and data augmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18963-18974 (2022) +53. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3128-3137 (2015) +54. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +55. Korbar, B., Xian, Y., Tonioni, A., Zisserman, A., Tombari, F.: Text-conditioned resampler for long form video understanding. arXiv preprint arXiv:2312.11897 (2023) +56. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +57. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: BLIP: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: International Conference on Machine Learning (2022) +58. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021) +59. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023) +60. Lin, J., Yin, H., Ping, W., Lu, Y., Molchanov, P., Tao, A., Mao, H., Kautz, J., Shoeybi, M., Han, S.: Vila: On pre-training for visual language models. arXiv preprint arXiv:2312.07533 (2023) +61. Lin, Z., Liu, C., Zhang, R., Gao, P., Qiu, L., Xiao, H., Qiu, H., Lin, C., Shao, W., Chen, K., et al.: Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575 (2023) +62. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023) +63. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253 (2024) + +64. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) +65. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (2024) +66. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +67. Liu, S., Fan, L.J., Johns, E., Yu, Z., Xiao, C., Anandkumar, A.: Prismer: A vision-language model with an ensemble of experts. ArXiv abs/2303.02506 (2023) +68. Lu, J., Clark, C., Lee, S., Zhang, Z., Khosla, S., Marten, R., Hoiem, D., Kembhavi, A.: Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172 (2023) +69. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-IO: A unified model for vision, language, and multi-modal tasks. In: The Eleventh International Conference on Learning Representations (2023) +70. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 3195-3204 (2019) +71. McKinzie, B., Gan, Z., Fauconnier, J.P., Dodge, S., Zhang, B., Dufter, P., Shah, D., Du, X., Peng, F., Weers, F., et al.: Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611 (2024) +72. Mizrahi, D., Bachmann, R., Kar, O.F., Yeo, T., Gao, M., Dehghan, A., Zamir, A.: 4M: Massively multimodal masked modeling. In: Advances in Neural Information Processing Systems (2023) +73. Naeem, M.F., Xian, Y., Zhai, X., Hoyer, L., Van Gool, L., Tombari, F.: Silc: Improving vision language pretraining with self-distillation. arXiv preprint arXiv:2310.13355 (2023) +74. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.Q., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y.B., Li, S.W., Misra, I., Rabbat, M.G., Sharma, V., Synnaeve, G., Xu, H., Jégou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINoV2: Learning robust visual features without supervision. ArXiv abs/2304.07193 (2023) +75. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021) +76. Raffel, C., Shazeer, N.M., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +77. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +78. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) + +79. Shen, Y., Heacock, L., Elias, J., Hentel, K.D., Reig, B., Shih, G., Moy, L.: Chatgpt and other large language models are double-edged swords (2023) +80. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) +81. Thorp, H.H.: Chatgpt is fun, but not an author (2023) +82. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022) +83. Tiong, A.M.H., Li, J., Li, B., Savarese, S., Hoi, S.C.: Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. arXiv preprint arXiv:2210.08773 (2022) +84. Tong, S., Liu, Z., Zhai, Y., Ma, Y., LeCun, Y., Xie, S.: Eyes wide shut? exploring the visual shortcomings of multimodal llms. arXiv preprint arXiv:2401.06209 (2024) +85. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +86. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015) +87. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022) +88. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022) +89. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023) +90. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., Wei, F.: Image as a foreign language: BEiT pretraining for all vision and vision-language tasks. ArXiv abs/2208.10442 (2022) +91. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021) +92. Xu, J., Zhou, X., Yan, S., Gu, X., Arnab, A., Sun, C., Wang, X., Schmid, C.: Pixel Aligned Language Models. arXiv preprint arXiv: 2312.09237 (2023) +93. Yadav, D., Jain, R., Agrawal, H., Chattopadhyay, P., Singh, T., Jain, A., Singh, S.B., Lee, S., Batra, D.: Towards better evaluation systems for ai agents (2019) +94. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021) +95. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) + +96. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (2022) +97. Yu, S., Cho, J., Yadav, P., Bansal, M.: Self-chained image-language model for video localization and question answering. arXiv preprint arXiv:2305.06988 (2023) +98. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) +99. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022) +100. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023) +101. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022) +102. Zhang, D., Yu, Y., Li, C., Dong, J., Su, D., Chu, C., Yu, D.: Mm-llms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601 (2024) +103. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) +104. Zhao, Y., Pang, T., Du, C., Yang, X., Li, C., Cheung, N.M.M., Lin, M.: On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems 36 (2024) +105. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: iBoT: Image BERT pre-training with online tokenizer. In: International Conference on Learning Representations (2022) +106. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) \ No newline at end of file diff --git a/2024/BRAVE_ Broadening the visual encoding of vision-language models/images.zip b/2024/BRAVE_ Broadening the visual encoding of vision-language models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a972896ba8e24660e4b1c01c7ad38ff95355dcd4 --- /dev/null +++ b/2024/BRAVE_ Broadening the visual encoding of vision-language models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc48b62847781cc39cc14dcb4a26207e577bc73bbd3a61ff4ef4339a6b657a4d +size 442799 diff --git a/2024/BRAVE_ Broadening the visual encoding of vision-language models/layout.json b/2024/BRAVE_ Broadening the visual encoding of vision-language models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d4caefb10d061c924d70015b91286a78b6d586d2 --- /dev/null +++ b/2024/BRAVE_ Broadening the visual encoding of vision-language models/layout.json @@ -0,0 +1,10129 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 110, + 471, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 110, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 144, + 110, + 471, + 148 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 168, + 459, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 168, + 459, + 193 + ], + "spans": [ + { + "bbox": [ + 154, + 168, + 459, + 193 + ], + "type": "text", + "content": "Oğuzhan Fatih Kar1,2, Alessio Tonioni1, Petra Poklukar1, Achin Kulshrestha1, Amir Zamir2, and Federico Tombari1,3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 201, + 479, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 201, + 479, + 225 + ], + "spans": [ + { + "bbox": [ + 134, + 201, + 479, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 201, + 479, + 225 + ], + "type": "text", + "content": "Google " + }, + { + "bbox": [ + 134, + 201, + 479, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 201, + 479, + 225 + ], + "type": "text", + "content": "Swiss Federal Institute of Technology Lausanne (EPFL) " + }, + { + "bbox": [ + 134, + 201, + 479, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 201, + 479, + 225 + ], + "type": "text", + "content": "TU Munich https://brave-vlms.epfl.ch" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 268, + 298, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 268, + 298, + 277 + ], + "spans": [ + { + "bbox": [ + 134, + 268, + 298, + 277 + ], + "type": "text", + "content": "Improving visual capabilities of VLMs with BRAVE" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 292, + 198, + 321 + ], + "blocks": [ + { + "bbox": [ + 137, + 278, + 200, + 292 + ], + "lines": [ + { + "bbox": [ + 137, + 278, + 200, + 292 + ], + "spans": [ + { + "bbox": [ + 137, + 278, + 200, + 292 + ], + "type": "text", + "content": "Is there a hand using the mouse in this image?" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 292, + 198, + 321 + ], + "lines": [ + { + "bbox": [ + 138, + 292, + 198, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 198, + 321 + ], + "type": "image", + "image_path": "275a45e841d9c5e8a9bd0345cc017349e824ce91657b6e6f4df78ba4013d30ac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 322, + 160, + 328 + ], + "lines": [ + { + "bbox": [ + 144, + 322, + 160, + 328 + ], + "spans": [ + { + "bbox": [ + 144, + 322, + 160, + 328 + ], + "type": "text", + "content": "(a) Yes" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 201, + 292, + 261, + 321 + ], + "blocks": [ + { + "bbox": [ + 202, + 278, + 260, + 292 + ], + "lines": [ + { + "bbox": [ + 202, + 278, + 260, + 292 + ], + "spans": [ + { + "bbox": [ + 202, + 278, + 260, + 292 + ], + "type": "text", + "content": "Are the butterfly's feet visible?" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 292, + 261, + 321 + ], + "lines": [ + { + "bbox": [ + 201, + 292, + 261, + 321 + ], + "spans": [ + { + "bbox": [ + 201, + 292, + 261, + 321 + ], + "type": "image", + "image_path": "3cc9d2159ed7352bcc7782dfe32972a6beed12bfbc71df98d0e85c5d0fd1cdfb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 322, + 223, + 328 + ], + "lines": [ + { + "bbox": [ + 208, + 322, + 223, + 328 + ], + "spans": [ + { + "bbox": [ + 208, + 322, + 223, + 328 + ], + "type": "text", + "content": "(a) Yes" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 264, + 292, + 293, + 321 + ], + "blocks": [ + { + "bbox": [ + 265, + 278, + 323, + 292 + ], + "lines": [ + { + "bbox": [ + 265, + 278, + 323, + 292 + ], + "spans": [ + { + "bbox": [ + 265, + 278, + 323, + 292 + ], + "type": "text", + "content": "Is the door of the truck open?" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 264, + 292, + 293, + 321 + ], + "lines": [ + { + "bbox": [ + 264, + 292, + 293, + 321 + ], + "spans": [ + { + "bbox": [ + 264, + 292, + 293, + 321 + ], + "type": "image", + "image_path": "8140a21dcf553b1758543b938e1149a60334dd3dcc11687e31f71664bd2d248e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 322, + 286, + 328 + ], + "lines": [ + { + "bbox": [ + 269, + 322, + 286, + 328 + ], + "spans": [ + { + "bbox": [ + 269, + 322, + 286, + 328 + ], + "type": "text", + "content": "(a) Yes" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 322, + 315, + 328 + ], + "lines": [ + { + "bbox": [ + 299, + 322, + 315, + 328 + ], + "spans": [ + { + "bbox": [ + 299, + 322, + 315, + 328 + ], + "type": "text", + "content": "(b) No" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 134, + 354, + 197, + 370 + ], + "blocks": [ + { + "bbox": [ + 134, + 335, + 197, + 350 + ], + "lines": [ + { + "bbox": [ + 134, + 335, + 197, + 350 + ], + "spans": [ + { + "bbox": [ + 134, + 335, + 197, + 350 + ], + "type": "text", + "content": "InstructBLIP (a) (b) (EVA Encoder)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 354, + 197, + 370 + ], + "lines": [ + { + "bbox": [ + 134, + 354, + 197, + 370 + ], + "spans": [ + { + "bbox": [ + 134, + 354, + 197, + 370 + ], + "type": "image", + "image_path": "bf4e78dc2116e5408bee825b78eba03968f4cb69291d468c78e18adc9b7c42ac.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 374, + 197, + 389 + ], + "lines": [ + { + "bbox": [ + 134, + 374, + 197, + 389 + ], + "spans": [ + { + "bbox": [ + 134, + 374, + 197, + 389 + ], + "type": "text", + "content": "BRAVE (a) (b) (Multiple Encoders)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 232, + 335, + 253, + 342 + ], + "blocks": [ + { + "bbox": [ + 232, + 335, + 253, + 342 + ], + "lines": [ + { + "bbox": [ + 232, + 335, + 253, + 342 + ], + "spans": [ + { + "bbox": [ + 232, + 335, + 253, + 342 + ], + "type": "image", + "image_path": "b117e0eed6c3be4ce2226d4bd53a70bea0295ca6f45c8ca4652199a5c775d819.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 232, + 355, + 254, + 361 + ], + "lines": [ + { + "bbox": [ + 232, + 355, + 254, + 361 + ], + "spans": [ + { + "bbox": [ + 232, + 355, + 254, + 361 + ], + "type": "text", + "content": "(a) (b)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 288, + 335, + 311, + 342 + ], + "blocks": [ + { + "bbox": [ + 288, + 335, + 311, + 342 + ], + "lines": [ + { + "bbox": [ + 288, + 335, + 311, + 342 + ], + "spans": [ + { + "bbox": [ + 288, + 335, + 311, + 342 + ], + "type": "image", + "image_path": "f989f58577877184e44fa199a8243811051743e678ae65762159580486d2004b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 355, + 311, + 361 + ], + "lines": [ + { + "bbox": [ + 288, + 355, + 311, + 361 + ], + "spans": [ + { + "bbox": [ + 288, + 355, + 311, + 361 + ], + "type": "text", + "content": "(a) (a)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 328, + 278, + 481, + 390 + ], + "blocks": [ + { + "bbox": [ + 134, + 395, + 482, + 488 + ], + "lines": [ + { + "bbox": [ + 134, + 395, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 134, + 395, + 482, + 488 + ], + "type": "text", + "content": "Fig. 1: We propose BRAVE to broaden the visual capabilities of vision-language models (VLMs). Left: In contrast to existing methods, e.g. InstructBLIP [25] or LLaVA-1.5 [64], that use a single vision encoder [59,84], BRAVE combines diverse features from multiple vision encoders into a more versatile and compact representation. The examples are taken from [84] and assess the VLM's ability to differentiate images with visual differences. Right: BRAVE leads to state-of-the-art performance on a wide range of captioning and visual question answering tasks. Furthermore, it significantly improves the performance on benchmarks [84] where commonly employed vision encoders fail." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 268, + 465, + 278 + ], + "lines": [ + { + "bbox": [ + 332, + 268, + 465, + 278 + ], + "spans": [ + { + "bbox": [ + 332, + 268, + 465, + 278 + ], + "type": "text", + "content": "SoTA Performance in Captioning & VQA" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 278, + 481, + 390 + ], + "lines": [ + { + "bbox": [ + 328, + 278, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 328, + 278, + 481, + 390 + ], + "type": "image", + "image_path": "1ff5d117d7eeb0caabe9a75c7786d2668c66042e25e0c051be06ca598ae1e526.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 159, + 533, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 533, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 533, + 455, + 666 + ], + "type": "text", + "content": "Abstract. Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. \"blindness\" to certain image features, visual hallucination, etc. To address these issues, we study broadening the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 205 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 205 + ], + "type": "text", + "content": "from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 215, + 454, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 215, + 454, + 228 + ], + "spans": [ + { + "bbox": [ + 160, + 215, + 454, + 228 + ], + "type": "text", + "content": "Keywords: Vision-Language " + }, + { + "bbox": [ + 160, + 215, + 454, + 228 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 215, + 454, + 228 + ], + "type": "text", + "content": " Visual Encoding " + }, + { + "bbox": [ + 160, + 215, + 454, + 228 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 215, + 454, + 228 + ], + "type": "text", + "content": " Multi-modal Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 246, + 230, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 230, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 230, + 258 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 365 + ], + "type": "text", + "content": "Vision-language models (VLMs) have recently seen significant improvements on solving tasks requiring both visual and text understanding capabilities such as captioning, visual question answering (VQA), and instruction following. These advancements are fueled by the progress in uni-modal vision encoders [30, 75] and language models (LMs) [22, 24], which are then combined with different \"bridging\" techniques [3,56,66] to create VLMs. Improved training recipes [25,64] and scalability efforts in terms of training data and model size [6,19,77,87,99] further boost the performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 481, + 474 + ], + "type": "text", + "content": "Despite the progress, VLMs are subject to several shortcomings: on the language side, the LMs are known to be susceptible to hallucinations and logical faults [8,37,79,81], while on the vision side, they are limited by the capabilities of the vision encoder. For example, Tong et al. [84] observed that commonly employed vision encoders such as CLIP [75] can exhibit \"blindness\", i.e. fail to differentiate images with clear visual differences (See Fig. 1, left). Similarly, Li et al. [59] showed that VLMs are prone to visual hallucinations where the models imagine incorrect details about a given image. These limitations create a bottleneck for developing performant and visually grounded VLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 617 + ], + "type": "text", + "content": "To address these, we study broadening the visual encoding capabilities of VLMs. It is known in machine learning that different representations of an input can yield different generalization properties, and each can play a role when ensembled to create a more complete representation [33]. Inspired by this, we first perform a comprehensive evaluation of VLMs which differ only in their vision encoders, i.e. visual biases. For this, we consider encoders with different biases, e.g. different objectives, training data, and model sizes. Our results in Table 2 show that different encoders lead to varying performance across vision-language tasks and there is no single encoder achieving a consistent top performance across tasks. Furthermore, we find that encoders with different biases can perform surprisingly similarly, suggesting that there could be multiple cues [35,45] that can be leveraged to solve the vision-language tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "Motivated by these findings, we propose to employ various vision encoders for VLMs and introduce a method to learn how to combine them efficiently. We denote the method as BRAVE, which stands for broadening the visual encoding of VLMs. As shown in Fig. 2, BRAVE combines features from an arbitrary number" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "of vision encoders into a compressed fixed-length visual representation which is then fed as a soft visual prompt to a frozen LM. We achieve the bridging of encoders and the LM by introducing a lightweight multi-encoder querying transformer (MEQ-Former), which is a multi-encoder generalization of the Q-Former proposed in the single encoder BLIP-2 framework [56]. In particular, MEQ-Former accepts a textual prompt and a set of learnable queries as inputs and jointly cross-attends to the features from different vision encoders (Fig. 2). We provide a simple yet effective recipe in Sec. 3 for training the MEQ-Former in a single stage that bypasses the two-stage pre-training paradigm of [56]. Notably, this is achieved with a smaller number of trainable parameters during pre-training than the existing methods [3,6,19,25,61,64,87] (Sec. 4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 331 + ], + "type": "text", + "content": "Our evaluation on a wide range of captioning and VQA tasks shows that BRAVE effectively consolidates diverse visual signals into a broad and contextual representation, leading to consistently better performance over the state-of-the-art and improved robustness against out-of-distribution inputs, as shown in Fig. 1 and Sec. 4. Furthermore, in contrast to previous works that benefit from scaling the language axis [3, 19, 25, 64] by using larger LMs, our work demonstrates that scaling along the vision axis has a significant potential for VLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 331, + 363, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 331, + 363, + 342 + ], + "spans": [ + { + "bbox": [ + 146, + 331, + 363, + 342 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 353, + 480, + 483 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 353, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 479, + 388 + ], + "type": "text", + "content": "- We conduct a systematic analysis of several vision encoders with different inductive biases, e.g. training data, objective, model size, on solving vision-language tasks under a unified training and evaluation framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 389, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 480, + 448 + ], + "type": "text", + "content": "- We introduce BRAVE, a method that efficiently consolidates features from any number of vision encoders into a compressed and contextual representation, leading to 1) state-of-the-art performance on several captioning and VQA tasks and 2) significantly improved robustness against visual hallucinations and out-of-distribution inputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 449, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 483 + ], + "type": "text", + "content": "- We perform a comprehensive ablation study of BRAVE, highlighting the impact of the design choices, which we hope provide useful insights to the researchers for further advancements in VLMs." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 502, + 467, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 467, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 467, + 517 + ], + "type": "text", + "content": "2 Impact of vision encoders for vision-language models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 528, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 599 + ], + "type": "text", + "content": "To quantify the impact of visual biases on the performance of VLMs, we compare VLMs with different vision encoders on commonly evaluated VQA and captioning tasks. For this, we develop a pre-training, fine-tuning and evaluation setup, as explained next. To the best of our knowledge, this is the first unified and comprehensive evaluation of different vision encoders for vision-language understanding and generation tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "VLM architecture. We use a frozen vision encoder which is connected to a frozen language model by a bridge network with trainable parameters. For the bridge, we adopt the Q-Former module from [56]. This choice is based on the following reasons: 1) the Q-Former resamples visual features to a fixed-length" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 448, + 102 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 166, + 168, + 445, + 277 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Overview of vision encoders we benchmarked. All encoders use a ViT [27] backbone, yet they differ in terms of objective, training data, and model size. ITC: Image-text contrastive learning [75], MIM: Masked image modelling [41], LGC: Local-to-global correspondence learning [11]. See Sec. 2 for details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 166, + 168, + 445, + 277 + ], + "lines": [ + { + "bbox": [ + 166, + 168, + 445, + 277 + ], + "spans": [ + { + "bbox": [ + 166, + 168, + 445, + 277 + ], + "type": "table", + "html": "
Vision encoderParametersTraining dataObjective
CLIP-L/14 [75]0.3BOpenAI WITITC
OpenCLIP-G/14 [21]1.8BLAION-2BITC
EVA-CLIP-g [30]1BLAION-400MMIM (CLIP features)
SIGLIP-G/14 [100]1.8BWebLIITC (Sigmoid loss)
SILC-G/16 [73]1.8BWebLIITC + LGC
ViT-e [19]3.8BJFT-3BClassification
ViT-G [99]1.8BJFT-3BClassification
DINOv2-L/14 [74]0.3BLVD-142MLGC + iBOT [105]
", + "image_path": "f0e7e9e66090fed507e442d144606fee7f7c6943c0045c3f52d31295f93e20ce.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 298, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 515 + ], + "type": "text", + "content": "output before feeding them to the LM which enables a fair comparison of vision encoders with different output dimensionalities, 2) the Q-Former is efficient to train and evaluate as it is compatible with frozen pre-trained vision encoders and LMs. For the vision encoder, we consider a diverse set of options as described in the next paragraph. For the LM, we follow [56] and use FlanT5-XL [24]. The Q-Former has 110M parameters consisting of a series of cross- and self-attention layers and a fixed number of learnable queries. Unless specified otherwise, we use 32 queries of dimension 768 which is also the hidden dimension of the Q-Former. The queries interact with visual features through cross-attention layers, enabling a significant reduction in their dimensionality. For example, CLIP-L/14 [75] visual features are reduced from " + }, + { + "bbox": [ + 130, + 298, + 482, + 515 + ], + "type": "inline_equation", + "content": "257 \\times 1024" + }, + { + "bbox": [ + 130, + 298, + 482, + 515 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 298, + 482, + 515 + ], + "type": "inline_equation", + "content": "32 \\times 768" + }, + { + "bbox": [ + 130, + 298, + 482, + 515 + ], + "type": "text", + "content": ". The output of the Q-Former is then linearly projected to the textual embedding space of the LM to create a soft visual prompt, and then prepended to the textual prompt that specifies the task. Both visual and textual prompts are fed as inputs to the frozen LM. For captioning, we use the textual prompt \"A photo of\" and for VQA, we directly use the question as the prompt. Similar to InstructBLIP [25], we give the textual prompt as additional input to the Q-Former together with the learnable queries to extract more task-aligned visual prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 529, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 627 + ], + "type": "text", + "content": "Vision encoders. As summarized in Table 1, we consider eight recently introduced vision encoders: CLIP [75], OpenCLIP [21], EVA-CLIP [30], SIGLIP [100], SILC [73], ViT-e [19], ViT-G [99], and DINOv2 [74]. While they all use ViT-based backbones [27], they differ in terms of 1) training data, e.g. LAION [77], WebLI [19], LVD-142M [74], etc., 2) training objective, e.g. image-text contrastive learning, masked image modelling, classification, etc., and 3) model size, e.g. 300M to 4B parameters. Due to this diversity, they incorporate different biases and thus potentially capture different aspects of the underlying depicted scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Pre-training data and objectives. We pre-train the Q-Former using WebLI [19] dataset at " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " resolution. We use the filtered and de-duplicated" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 193, + 449, + 313 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "text", + "content": "Table 2: Benchmarking vision encoders on different vision-language tasks. For COCO captioning, we report CIDEr [86] score. For VQA tasks, we report top-1 accuracy. For MMVP, we report average pair accuracy [84]. Top-3 best results are shown with Blue, Green, and Orange, respectively. All VLMs use FlanT5-XL [24] as the language model and a Q-Former [56] to bridge vision and language modalities. See Sec. 2 for details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 193, + 449, + 313 + ], + "lines": [ + { + "bbox": [ + 167, + 193, + 449, + 313 + ], + "spans": [ + { + "bbox": [ + 167, + 193, + 449, + 313 + ], + "type": "table", + "html": "
Vision encoderCOCO cap. ↑ Karpathy valVQAv2 ↑ Karpathy valOKVQA ↑ valGQA ↑ test-devMMVP ↑ test
CLIP-L/14133.074.461.048.715.3
OpenCLIP-G128.373.360.648.022.0
EVA-CLIP-g140.977.063.050.127.3
SIGLIP-G/14133.074.762.548.624.0
SILC-G/16141.177.063.449.724.0
ViT-e137.875.661.949.125.3
ViT-G133.874.261.248.320.7
DINOv2-L/14127.671.359.048.022.0
", + "image_path": "43ed85e5c7189e1753bdeacda3934b93642cb4317c8ccdccf001cf052a6eb211.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 336, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 411 + ], + "type": "text", + "content": "English-only subset with 100 million image-text pairs [19, 73]. We keep both the vision encoder and the LM frozen and only train the parameters of the Q-Former. The model is trained with captioning objective using the alt-text in WebLI as target, as it has been shown to be effective [92]. Similar to [56], we remove the last Transformer layer of the encoder. We refer the reader to the supplementary material for additional training details and hyperparameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 429, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 634 + ], + "type": "text", + "content": "Evaluation tasks. We evaluate the obtained VLMs on standard captioning and VQA tasks. For captioning, we use the COCO captioning benchmark [20] and fine-tune the pre-trained VLM on the commonly employed Karpathy training split [53]. Similar to the pre-training, only the Q-Former parameters are updated. For VQA, we again follow the standard practices [19, 25, 56, 64] and fine-tune on a mixture of VQA data that enforces the Q-Former to extract more task-aligned visual features. Our mixture includes the training sets of VQAv2 [36] and OKVQA [70] as well as the synthetically generated VQA data from " + }, + { + "bbox": [ + 130, + 429, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ^2A}" + }, + { + "bbox": [ + 130, + 429, + 482, + 634 + ], + "type": "text", + "content": " [12], resulting in total 17M training examples. We then evaluate the fine-tuned models on VQAv2 and OKVQA validation sets. Similar to [25, 56], we also evaluate the VLMs' zero-shot VQA capabilities on GQA [48]. Finally, we evaluate the zero-shot performance on MMVP [84] that includes images with semantic differences that are challenging for the current VLMs to handle. Unlike captioning, we fine-tune both Q-Former and the LM parameters while keeping the vision encoder frozen, as we find this to improve the performance, similar to [64]. For both captioning and VQA tasks, we use " + }, + { + "bbox": [ + 130, + 429, + 482, + 634 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 130, + 429, + 482, + 634 + ], + "type": "text", + "content": " image resolution. See the supplementary for more details and results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 653, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 481, + 666 + ], + "type": "text", + "content": "Results. We report the findings in Table 2 and make the following observations:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 448, + 103 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 510 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 200 + ], + "type": "text", + "content": "- Encoders with different biases can obtain similar performance. While we observe some variation in VLMs' performance across different encoders, the gap remains small, especially among the top-performing ones. For example, EVA-CLIP-g, SILC-G/16, and ViT-e achieve the best results for VQAv2, GQA and COCO captioning despite their differences in training objective and data, as shown in Table 1. Similar observations can be made for OpenCLIP and DINOv2 encoders that have a small performance gap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 200, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 200, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 481, + 259 + ], + "type": "text", + "content": "- MMVP stays challenging for all encoders. While it is originally curated with \"CLIP-blind\" pairs in [84], i.e. images that CLIP perceives as similar despite their visual differences (See Fig. 1 for examples), most of the encoders stay below the random guess accuracy of " + }, + { + "bbox": [ + 138, + 200, + 481, + 259 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 138, + 200, + 481, + 259 + ], + "type": "text", + "content": ". Exceptions are EVA-CLIP and ViT-e which slightly surpass the random guess level (" + }, + { + "bbox": [ + 138, + 200, + 481, + 259 + ], + "type": "inline_equation", + "content": "27.3\\%" + }, + { + "bbox": [ + 138, + 200, + 481, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 200, + 481, + 259 + ], + "type": "inline_equation", + "content": "25.3\\%" + }, + { + "bbox": [ + 138, + 200, + 481, + 259 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 259, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 481, + 319 + ], + "type": "text", + "content": "- For tasks requiring compositional reasoning and open-world knowledge, the performance decreases and the gap among VLMs diminishes, as shown by the results on GQA and OKVQA. More precisely, the standard deviation across all the VLMs' performance on GQA and OKVQA is 0.8 and 1.36, respectively, but 1.74 on VQAv2 and 4.91 on COCO." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 319, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 481, + 354 + ], + "type": "text", + "content": "- Scaling the size of vision encoder can improve the performance as seen by the consistent improvements obtained by ViT-e over ViT-G when scaling from 1.8B to 3.8B parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 354, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 437 + ], + "type": "text", + "content": "- Pre-training data distribution can play a critical role. OpenCLIP-G/14 model is larger than CLIP-L/14 and uses more training data, yet it noticeably underperforms on most of the evaluated VQA and captioning tasks except the MMVP benchmark, where OpenCLIP significantly outperforms the CLIP model. Since both models are trained using the same training objective, this indicates that the impact of the training objective and dataset to the performance of the VLM can be disentangled, and both are important." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 437, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 510 + ], + "type": "text", + "content": "- Performance correlates between VQA and captioning tasks. Encoders with better performance on VQA tasks generally also perform better on captioning, and vice versa. Spearman's rank correlation between VQA tasks and COCO captioning is 0.90, 0.80, 0.79, 0.56 for VQAv2, OK-VQA, GQA, and MMVP, respectively. Thus, improving the visual component in VLMs is expected to improve performance for a broad range of tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 516, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 482, + 564 + ], + "type": "text", + "content": "The above results highlight the importance of using the right vision encoder for VLMs, motivating us to focus on the following question: Can we broaden the vision capabilities of VLMs through combining vision encoders with different visual biases? We answer positively in the next section." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 581, + 380, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 380, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 380, + 594 + ], + "type": "text", + "content": "3 Combining vision encoders with BRAVE" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Motivated by the findings in Sec. 2, we propose using multiple vision encoders with different, and potentially complementary, visual biases to create more capable VLMs. To do this, we introduce BRAVE, a method that combines the strengths of different vision encoders while staying efficient in terms of number of trainable parameters. Below, we provide details of BRAVE." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 339, + 255 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 339, + 255 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 339, + 255 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 339, + 255 + ], + "type": "image", + "image_path": "db020df293b626c30f01d0e6866cc2becb612f2e138bec09372ce27aecc1fc3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "type": "text", + "content": "Fig. 2: Overview of BRAVE. Left: We keep all the vision encoders (VEs) and the language model (LM) frozen. The linear projection layers are used to concatenate features from " + }, + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "type": "text", + "content": " different VEs, e.g. " + }, + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "type": "text", + "content": ", sequence-wise. These are then resampled by the MEQ-Former which accepts a set of learnable queries and a text prompt describing the task as inputs. The output of MEQ-Former is projected to the input space of the LM using fully-connected (FC) layers. The total number of trainable parameters is 116M (" + }, + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\approx 1\\%" + }, + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "type": "text", + "content": " of the total parameters). Right: Architecture of the MEQ-Former with " + }, + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "type": "inline_equation", + "content": "N = 12" + }, + { + "bbox": [ + 130, + 264, + 482, + 364 + ], + "type": "text", + "content": " transformer layers. It interacts with the concatenated features through crossattention layers and produces a fixed-length output to be fed as input to the LM." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 345, + 118, + 476, + 255 + ], + "blocks": [ + { + "bbox": [ + 345, + 118, + 476, + 255 + ], + "lines": [ + { + "bbox": [ + 345, + 118, + 476, + 255 + ], + "spans": [ + { + "bbox": [ + 345, + 118, + 476, + 255 + ], + "type": "image", + "image_path": "1f32267e7f2c26754790dc029424ed0747f84db25d18d015b9c239c6ff803703.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 386, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 615 + ], + "type": "text", + "content": "Overview. As depicted in Fig. 2, we introduce the multi-encoder querying transformer, or MEQ-Former, that combines visual features from an arbitrary set of encoders and outputs a fixed-length visual representation that is given as a soft prompt to a frozen LM. It takes as an input a sequence comprised of a fixed number of learnable queries and embedded text tokens which describe the task - for captioning, they are the textual prompt \"A photo of\", while for VQA, they are the question prompt for a given image (See supplementary for examples). MEQ-Former interacts with visual features through cross-attention layers. The visual features from different encoders are first linearly projected to the same dimension and then concatenated sequence-wise. The obtained concatenated feature is given as a key and value pair to the cross-attention layers in MEQ-Former, and is cross-attended to by the MEQ-Former's query sequence (Fig. 2). This resampling enables to efficiently processing a large number of visual features since it bypasses the quadratic complexity of self-attention. It also acts as a \"bottleneck\" that keeps the total number of VLM's parameters low compared to a naive ensembling of VLMs that is prohibitively expensive (See a comparison in Sec. 4.3). Moreover, the visual features are not \"marked\" with any encoder-specific embedding, thus MEQ-Former is not biased to differentiate between encoders, allowing for a simpler design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Differences with previous work. Our work differs from other popular approaches like LLaVa [64] and PaLI [19] that require significantly more trainable parameters (" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\approx 10\\mathrm{B}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ") and do not have a resampling mechanism. Resampling visual" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 448, + 102 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "features from a single encoder has been successfully demonstrated in previous work for single- [25,56] and multiple-frames [3,55,97]. To the best of our knowledge, our work is the first one demonstrating an effective resampling mechanism to consolidate visual features from several arbitrarily different encoders while keeping the number of trainable parameters small (our MEQ-Former having 116M parameters vs BLIP-2's Q-Former is 188M). See Sec. 5 for more comparisons." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 205, + 483, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 483, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 483, + 291 + ], + "type": "text", + "content": "Pre-training details. We use the same data and objective as in Sec. 2 and only train the MEQ-Former while keeping the vision encoders and the LM frozen. During pre-training, we randomly mask features from each encoder with " + }, + { + "bbox": [ + 130, + 205, + 483, + 291 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 205, + 483, + 291 + ], + "type": "text", + "content": " probability as we found that this can act as a regularizer and enforces MEQ-Former to avoid local minima by attending only to the features of a single encoder (Ablated in Sec. 4.3). After pre-training, we fine-tune the MEQ-Former and (optionally) the LM on the downstream tasks, as explained in Sec. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": "Architecture. For the main results, we combine five vision encoders, namely EVA-CLIP-" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": " [30], CLIP-" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "L/14" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": " [75], SILC-" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "G/16" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": " [73], ViT-" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": " [19], and DINOv2-" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "L/14" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": " [74] to cover all training datasets and objectives from Table 1. Please see Sec. 4.3 and the supplementary for analysis on the contributions of encoders. For the LM, we employ FlanT5-XL [24] as before. We set the visual feature dimension after linear projection as 1408 for concatenation. For MEQ-Former, we use " + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "32 \\times 5 = 160" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": " learnable queries to proportionally scale the model capacity to the number of vision encoders and set the hidden dimension as 768. The MEQ-Former resampling results in " + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "14 \\times" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": " reduction in total feature size (from " + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "1223 \\times 1408" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "160 \\times 768" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": "). The total number of trainable parameters during pre-training is 116M which is about " + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 307, + 483, + 439 + ], + "type": "text", + "content": " of the total number of parameters in the VLM (10.3B)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 455, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 492 + ], + "type": "text", + "content": "Implementation. We use FLAX [42] with the Scenic [26] library to implement our models and training pipeline. For evaluations, we use both Scenic and EvalAI API [93]. We pre-train with batch size of 1024 on 64 TPUv5 chips." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 509, + 231, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 231, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 231, + 523 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": "We perform evaluations on a broad range of captioning (Sec. 4.1) and VQA tasks (Sec. 4.2) to show the effectiveness of BRAVE. Please see Fig. 1 right for a summary of the results and supplementary for an overview of the tasks. We then provide a comprehensive analysis of BRAVE in Sec. 4.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "text", + "content": "For captioning, we evaluate on COCO [20] and NoCaps [2] benchmarks. The latter includes out-of-distribution samples with novel classes that do not exist in the COCO benchmark, hence it serves as a challenging testbed to assess robustness. The results (Table 3) show that BRAVE pushes the state-of-the-art for both benchmarks while staying efficient in terms of trainable parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "For VQA, we evaluate on VQAv2 [36], OKVQA [70], GQA [48], and VizWiz-QA [39] benchmarks. To assess the robustness of VLMs, we additionally include" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 234, + 462, + 333 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 224 + ], + "type": "text", + "content": "Table 3: Captioning. We compare BRAVE with state-of-the-art captioning methods on COCO and NoCaps (zero-shot). CIDEr [86] score is reported (Best in bold, second best is underlined). Numbers of other methods are taken from the corresponding publications. Dashed line means no result was reported. BRAVE achieves the best results in NoCaps evaluation sets (both out-domain and overall in validation & test) and second best for COCO, while having much fewer trainable parameters than other methods. For example, compared to PaLI-17B [19], BRAVE uses significantly fewer trainable parameters, training data, and image pixels. Yet, it outperforms the baselines on NoCaps and stays competitive on COCO, demonstrating the effectiveness of having different visual biases for generalization. See Sec. 4.1 for more details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 234, + 462, + 333 + ], + "lines": [ + { + "bbox": [ + 152, + 234, + 462, + 333 + ], + "spans": [ + { + "bbox": [ + 152, + 234, + 462, + 333 + ], + "type": "table", + "html": "
Method# paramsCOCO (fine-tuned)NoCaps (zero-shot, val)NoCaps (zero-shot, test)
TrainableTotalKarpathy testout-domainoverallout-domainoverall
Flamingo [3]10.6B80B138.1----
SimVLM [91]632M632M143.3113.7112.2-110.3
Qwen-VL [6]9.6B9.6B--121.4--
BLIP-2 [56]1.1B4.1B144.5124.8121.6--
InstructBLIP [25]188M14.2B--121.9--
CoCa [96]2.1B2.1B143.6-122.4-120.6
GiT2 [87]5.1B5.1B145.0130.6126.9122.3124.8
PaLI-17B [19]16.9B16.9B149.1-127.0126.7124.4
BRAVE116M10.3B148.0133.3127.6127.1125.6
", + "image_path": "6172d64b7d8a4ded4e57e27e6fe71d7b1bc8775babe58dbafeb23849044c9057.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 355, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 404 + ], + "type": "text", + "content": "POPE [59] that measures object hallucinations, and MMVP [84] which exposes the shortcomings of CLIP-based vision encoders that are widely used in VLMs. Our results (Table 4) show that BRAVE achieves consistent improvements over recent methods for all the tasks by leveraging diverse visual features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 422, + 248, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 248, + 435 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 248, + 435 + ], + "type": "text", + "content": "4.1 Image captioning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 443, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 515 + ], + "type": "text", + "content": "Fine-tuning details. Similar to Sec. 2, we only fine-tune the MEQ-Former on COCO captioning and keep the vision encoders and the LM frozen. During fine-tuning, we set a higher input resolution as " + }, + { + "bbox": [ + 130, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 130, + 443, + 482, + 515 + ], + "type": "text", + "content": " to further boost the performance, similar to previous works [19,56,64] (See Sec. 4.3 for an ablation). Following [19,25,56,87,96], we also zero-shot evaluated the COCO fine-tuned model on NoCaps [2]. Please see the supplementary for more training details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Results. We summarize the captioning evaluations in Table 3. BRAVE uses the least number of trainable parameters (116M) yet achieves strong results for both COCO and NoCaps benchmarks. For NoCaps, BRAVE is the best performing method with significant gains over recent methods. This is especially the case for out-domain samples with novel classes, demonstrating the usefulness of diversity in visual features for robustness. For COCO, BRAVE stays competitive with the best performing model, PaLI-17B [19]. Notably, BRAVE achieves this while using " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "150 \\times" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " fewer trainable parameters (116M vs 16.9B), " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " less pre-training data (100M vs 1.6B) and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " fewer image pixels (" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "588 \\times 588" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ") than PaLI-17B, suggesting that having different visual biases is effective for generalization, while keeping the sample complexity low." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 448, + 103 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": "Table 4: Visual question answering. We compare BRAVE with state-of-the-art VQA methods on VQAv2, OKVQA, GQA, VizWiz-QA, MMVP, and POPE. Top-1 accuracies are reported (Best in bold, second best is underlined). For MMVP, we report average pair accuracy [84]. Numbers of other methods are taken from the corresponding publications. Dashed line means no result was reported. BRAVE achieves the best results for six out of seven benchmarks, while staying efficient in terms of trainable parameters. Similar to the COCO captioning results in Table 3, BRAVE stays competitive with PaLI-17B [19] for VQAv2, while using significantly less pre-training data, image resolution, and model parameters (both trainable and total). See Sec. 4.2 for details." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 151, + 224, + 462, + 372 + ], + "blocks": [ + { + "bbox": [ + 151, + 224, + 462, + 372 + ], + "lines": [ + { + "bbox": [ + 151, + 224, + 462, + 372 + ], + "spans": [ + { + "bbox": [ + 151, + 224, + 462, + 372 + ], + "type": "table", + "html": "
Method# paramsFine-tunedZero-shot
TrainableTotalVQAv2 test-devOKVQA valGQA test-devVizWiz-QA test-devGQA test-devMMVP testPOPE test
SimVLM [91]632M632M80.0------
Flamingo [3]10.2B80B82.057.8-31.6---
MiniGPT-v2 [14]7B8B-57.860.153.6---
GiT2 [87]5.1B5.1B81.7------
Qwen-VL [6]9.6B9.6B79.558.659.335.2---
SPHINX-2k [61]13B16.5B80.762.663.144.9--87.2
PaLI-17B [19]16.9B16.9B84.364.5-----
BLIP-2 [56]1.2B12.1B81.654.7-29.444.7-85.3
InstructBLIP [25]188M14.2B-55.5-33.449.516.778.9
ShareGPT4V [16]13.4B13.4B81.0-64.8----
LLaVa1.5 [64]13B13.4B80.0-63.353.6-24.785.9
LLaVa1.6 [65]13B13.4B-46.365.4---86.3
LLaVa1.5 (I-MoF) [84]13B13.6B79.3----31.386.7
BRAVE3B10.3B82.566.066.354.252.742.087.6
", + "image_path": "a2474afb0011e5310d7878e0f96accbdcbc2fc9f1fc0696d0c6755a8eecb0584.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 393, + 293, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 393, + 293, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 293, + 406 + ], + "type": "text", + "content": "4.2 Visual question answering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "content": "Fine-tuning details. We use the same VQA mixture as in Sec. 2 and fine-tune both the MEQ-Former and the LM at " + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "content": " resolution. Vision encoders are kept frozen. This is followed by a high-resolution fine-tuning stage on VQAv2 and OKVQA training sets at " + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "content": " resolution, similar to previous work [19,56,64]. The resulting model is evaluated on VQAv2 and OKVQA benchmarks as well as zero-shot evaluated on GQA [48], VizWiz-QA [39], MMVP [84] and POPE [59], following [25,56,64,84]. To be able to compare to the methods that fine-tune on GQA training set, e.g. [6,61,64], we also perform an additional training stage on top of the mixture-finetuned model, and evaluate the model on the GQA evaluation set (test-dev). Please see the supplementary for the details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "text", + "content": "Results. We summarize the VQA evaluations in Table 4. BRAVE achieves strong results for all the benchmarks in both fine-tuned and zero-shot evaluation scenarios. For OKVQA, GQA, VizWiz-QA, MMVP, and POPE benchmarks, BRAVE achieves the best results, and for VQAv2, it stays competitive with PaLI-17B:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 605, + 479, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 137, + 605, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 605, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 137, + 605, + 479, + 640 + ], + "type": "text", + "content": "- For MMVP, BRAVE significantly improves the performance over previous approaches, making it less susceptible to failure modes of CLIP-based encoders. This can be further seen from the qualitatives in Figures 1 and 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "text", + "content": "- For VizWiz-QA, BRAVE achieves the best results without using any scene text understanding or OCR data during pre-training, as opposed to [6,14,64]." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 482, + 319 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 319 + ], + "type": "image", + "image_path": "bfda9a6e63e1900bbaf032351484f0a6836e6058b510ba6dcf807e09244f6d5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 329, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 439 + ], + "type": "text", + "content": "Fig. 3: Qualitative results. We compare predictions of BRAVE and the VLMs with different vision encoders, e.g. CLIP, on samples from the MMVP benchmark. Following [84], a model is considered correct only if it answers to both images in a pair correctly, i.e. if it can successfully differentiate between images with semantic differences. Note that the images in a pair are seen independently, i.e. neither of the images is provided as context for the other one. All encoders output some correct predictions, yet none of them performs consistently well on a broad range of inputs. BRAVE alleviates this by combining diverse visual features, leading to a more consistent performance. The quantitative difference is indeed stark: " + }, + { + "bbox": [ + 130, + 329, + 482, + 439 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 130, + 329, + 482, + 439 + ], + "type": "text", + "content": " for BRAVE vs " + }, + { + "bbox": [ + 130, + 329, + 482, + 439 + ], + "type": "inline_equation", + "content": "27.3\\%" + }, + { + "bbox": [ + 130, + 329, + 482, + 439 + ], + "type": "text", + "content": " for best single encoder (Tables 2 and 4). See supplementary for more qualitative results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 460, + 480, + 566 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 137, + 460, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 460, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 137, + 460, + 480, + 483 + ], + "type": "text", + "content": "- For GQA and OKVQA that requires different capabilities, i.e. spatial reasoning and open-world knowledge, BRAVE noticeably improves the performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 484, + 480, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 484, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 138, + 484, + 480, + 518 + ], + "type": "text", + "content": "- For POPE, BRAVE further reduces the visual hallucinations over dual encoder methods [61,84], confirming that using multiple encoders can be effective for visual grounding (Also see Sec. 5 for a discussion)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 518, + 480, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 518, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 137, + 518, + 480, + 566 + ], + "type": "text", + "content": "- For VQAv2, BRAVE is the second best after PaLI-17B, while staying more efficient in terms of training data, model size, and input resolution, similar to the COCO captioning result in Sec. 4.1. We expect BRAVE to benefit from such scaling efforts as well, which is a future direction to investigate." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 588, + 247, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 247, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 247, + 600 + ], + "type": "text", + "content": "4.3 Analysis of BRAVE" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "We study different aspects of BRAVE including an ablation study, contribution of vision encoders to the final performance, comparing MEQ-Former to an ensembling strategy, and the role of pre-training data. Due to the space limit, please see the supplementary for more qualitative and quantitative results." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 448, + 102 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 179, + 445, + 302 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 5: Ablation study of BRAVE. We start from the base pre-training setup explained in Sec. 3, denoted as A0. Each row corresponds to a particular deviation from A0, while keeping the rest fixed. For example, A5 evaluates the performance when high-resolution fine-tuning is disabled. N/A means the ablation is not applicable for that evaluation. See Sec. 4.3 for the discussions and supplementary for more results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 179, + 445, + 302 + ], + "lines": [ + { + "bbox": [ + 167, + 179, + 445, + 302 + ], + "spans": [ + { + "bbox": [ + 167, + 179, + 445, + 302 + ], + "type": "table", + "html": "
Ablation IDStudy SubjectOriginalChangeCOCO Captioning (Karpathy val)VQAv2 (Karpathy val)OKVQA (val)
A0Base--147.081.865.7
A1LM fine-tuning for VQAFull fine-tuningNo fine-tuningN/A78.657.5
LoRA (r=64)N/A80.762.8
LoRA (r=128)N/A81.062.9
A2Synthetic VQA data [12]✓✗N/A81.164.0
A3Encoder dropout✓✗145.381.366.0
A4MEQ-Former text input✓✗145.981.464.9
A5High-res fine-tuning✓✗145.279.665.0
A6A3 + A5✓✗144.079.064.5
A7A4 + A5✓✗145.178.363.4
A8Language modelFlanT5-XLFlanT5-L142.579.965.5
", + "image_path": "26b0088519087d6b637a8ddf2533a74d5c6bd308c92d976018e5fc1fab6e545a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 459 + ], + "type": "text", + "content": "Ablation study. We perform a comprehensive study in Table 5 where we ablate different design choices, e.g. training data and choice of language model. The ablations are labeled from A1 to A8, each showing a particular deviation from the original pre-training setup A0 (Sec. 3). We summarize the key findings: (A1): LM fine-tuning for VQA significantly boosts the performance compared to frozen LM, similar to the observation in [64]. To minimize the performance loss, we perform a LoRA [46] fine-tuning by inserting LoRA layers to the LM. With a rank of 128, we compensated for " + }, + { + "bbox": [ + 130, + 327, + 482, + 459 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 327, + 482, + 459 + ], + "type": "text", + "content": " of the performance gap while using " + }, + { + "bbox": [ + 130, + 327, + 482, + 459 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 130, + 327, + 482, + 459 + ], + "type": "text", + "content": " fewer trainable parameters than full fine-tuning (324M vs 3B). This underscores the development of parameter-efficient fine-tuning methods as an important direction for more efficient VLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 460, + 481, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 496 + ], + "type": "text", + "content": "(A2): Using synthetic VQA data [12] noticeably enhances our performance, while being much cheaper to collect than human annotated data. This can potentially be further improved by using more diverse templates [38, 83]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 556 + ], + "type": "text", + "content": "(A3,A4,A5): Encoder dropout during training improves the performance for COCO captioning and VQAv2 but slightly degrades for OKVQA. Using text prompt as an additional input to MEQ-Former helps with extracting better task-aligned features. Similarly, high-resolution fine-tuning stage gives a boost to the performance at the expense of processing more visual tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 582 + ], + "type": "text", + "content": "(A6,A7): These ablations combine the impact of previously ablated design choices, further demonstrating their contribution to the final performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 481, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 619 + ], + "type": "text", + "content": "(A8): A stronger LM significantly helps for all the tasks, underscoring that scaling the VLM along the vision and language axes simultaneously have complementary benefits, and both are important." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": "Contribution of vision encoders. To understand the impact of vision encoders to the final performance of BRAVE, we perform additional evaluations:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 447, + 195 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 447, + 195 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 447, + 195 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 447, + 195 + ], + "type": "image", + "image_path": "7112a9140dd9ea5ba4c641a7c29b3ca11f169f479fde8f9f890463b9b1294523.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 203, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 258 + ], + "type": "text", + "content": "Fig. 4: Contribution of vision encoders to BRAVE. Left: We analyze the robustness of BRAVE when a subset of encoders are removed during evaluation. We report the average drop in CIDEr for COCO and accuracy for VQAv2. Right: We compute average attention scores for different vision encoders cross-attended by the MEQ-Former for COCO and VQAv2. See Sec. 4.3 for the discussions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 280, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 482, + 399 + ], + "type": "text", + "content": "- In Fig. 4-left, we perform a robustness analysis by removing a subset of vision encoders during inference. This is performed for all the combinations and we report the average drop in performance for COCO captioning and VQAv2 tasks (Full results in supplementary). The results show that the performance of BRAVE degrades gracefully up to two encoder removals for both tasks, indicating some redundancy among the encoders. Because of this, even if some encoders are missing, BRAVE can make use of the remaining set of encoders to compensate. Beyond the two encoder removal, the drop in the performance is more severe, suggesting that some features are more unique and consequently harder to replace using the remaining set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 400, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 400, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 400, + 482, + 471 + ], + "type": "text", + "content": "- In Fig. 4-right, we compute attention scores of features from different encoders that are cross-attended by the MEQ-Former queries. These are averaged for randomly sampled 1k samples from COCO and VQAv2 validation sets. The results show that the MEQ-Former can cross-attend, or \"pay attention to\", some encoders more than the others depending on the downstream task adaptively, further confirming the usefulness of having different biases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 487, + 234, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 234, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 234, + 499 + ], + "type": "text", + "content": "5 Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "Several works have focused on training VLMs to solve tasks requiring both vision and text understanding capabilities. Most of these works make use of pre-trained vision encoders [9,27,31,49,75,99] and LMs [6,22,24,44,76,85,103]. Architecture-wise, different approaches were developed to project visual features to the textual embedding space, e.g. Q-Former [25, 56, 106], P-Former [51], cross-attention [3, 6, 19, 95], linear projector [14, 15, 66] and MLP [64, 89]. Different objectives were also considered depending on the downstream tasks, e.g. image-text contrastive learning [50,73,75,94,101], auto-regressive text generation [17,19,23,87,88,91] or both [56-58,96] (see [102] for a survey). In contrast to ours, these works consider a single vision encoder which has inherent limitations [47,59,82,84] (Sec. 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Concurrent to our work, LLaVa-MoF [84] proposes combining CLIP [75] and DINOv2 [74] encoders using separate adapters. SPHINX [61] mixes CLIP, DINOv2, and Q-Former outputs with linear projection layers. Both of the methods" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 448, + 103 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "input a concatenation of the extracted visual embeddings to the LM, which is not scalable to multiple encoders. We differ from these works by proposing a flexible and unified resampling mechanism that combines features from several vision encoders while staying efficient and improving the performance (Tables 3, 4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 260 + ], + "type": "text", + "content": "Our work focuses on scaling the VLMs along the vision axis, unlike other works that explored scaling homogenously model size across the visual and/or language component [1,3,4,18,19,28,34,80,90], training data [6,10,13,49,50,60, 71,78,87,89] and number of tasks and modalities [5,7,40,43,67-69,72,98]. BRAVE can be combined with these efforts to further improve performance. Moreover, several studies investigated the visual aspect of VLMs by evaluating them against hallucinations and other robustness issues [32,47,59,62,63,82,104]. BRAVE improves the robustness to those issues by broadening visual capabilities of VLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 279, + 322, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 322, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 322, + 293 + ], + "type": "text", + "content": "6 Conclusions and Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 305, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 400 + ], + "type": "text", + "content": "We focus on broadening the visual encoding aspect of VLMs. Our evaluation of vision encoders revealed that there is no encoder achieving a consistent top performance across tasks and encoders with different biases can perform similarly. Motivated by this, we introduce BRAVE that combines diverse features from several vision encoders using a small number of trainable parameters. Extensive evaluations show that BRAVE achieves state-of-the-art performance and improves robustness of VLMs against out-of-distribution inputs and visual hallucinations. We discuss limitations of our work and future directions below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 449 + ], + "type": "text", + "content": "Adaptive mechanisms. While MEQ-Former can resample from the useful features and discard the irrelevant ones, it still needs forward passes from all the encoders. A future direction could be to explore adaptive mechanisms to pre-select which encoders to resample from, reducing the inference cost." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 556 + ], + "type": "text", + "content": "Improving sample efficiency. BRAVE uses around 100M image-text pairs during pre-training which is an order of magnitude less than recent methods [6, 19, 87], yet can benefit from further lowering the sample complexity similar to [51]. More broad vision encoders. The set of encoders we investigated do not completely cover all types of visual biases, e.g. including the ones with strong 3D [29, 52] or semantic [54] priors could further improve the performance. Our results show that the current set is a good starting point to see the benefits of consolidation. Another interesting direction is to distill from the MEQ-Former-combined features back into a single encoder for a more unified solution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "text", + "content": "Different modalities and multiple frames. While we focus on image and text modalities, BRAVE is a general method that could be potentially used to fuse more modalities, e.g. audio or 3D, or extended to incorporate multiple frames, e.g. for video understanding or in-context learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 641 + ], + "type": "text", + "content": "Biases of text-generative models. As we employ pre-trained LMs in our setup, the resulting VLMs are susceptible to the longstanding bias and fairness issues of these models, thus they should be used with caution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgements. We thank Diego Martin Arroyo, Ferjad Naeem, Xingyi Zhou, Yannick Strümpler and Yongqin Xian for their help with the project." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 142, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 142, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 170, + 480, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 170, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 142, + 170, + 480, + 212 + ], + "type": "text", + "content": "2. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8948-8957 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 213, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 213, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 480, + 255 + ], + "type": "text", + "content": "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 256, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 256, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 142, + 256, + 480, + 298 + ], + "type": "text", + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 142, + 300, + 480, + 331 + ], + "type": "text", + "content": "5. Bachmann, R., Kar, O.F., Mizrahi, D., Garjani, A., Gao, M., Griffiths, D., Hu, J., Dehghan, A., Zamir, A.: 4m-21: An any-to-any vision model for tens of tasks and modalities. arXiv preprint arXiv:2406.09406 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 332, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 142, + 332, + 480, + 353 + ], + "type": "text", + "content": "6. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 354, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 354, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 142, + 354, + 480, + 384 + ], + "type": "text", + "content": "7. Bai, Y., Geng, X., Mangalam, K., Bar, A., Yuille, A., Darrell, T., Malik, J., Efros, A.A.: Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 385, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 385, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 142, + 385, + 480, + 427 + ], + "type": "text", + "content": "8. Bang, Y., Cahyawijaya, S., Lee, N., Dai, W., Su, D., Wilie, B., Lovenia, H., Ji, Z., Yu, T., Chung, W., et al.: A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 142, + 429, + 480, + 460 + ], + "type": "text", + "content": "9. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: International Conference on Machine Learning. pp. 1059-1071. PMLR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "text", + "content": "10. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 483, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 525 + ], + "type": "text", + "content": "11. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 546 + ], + "type": "text", + "content": "12. Changpinyo, S., Kukliansky, D., Szpektor, I., Chen, X., Ding, N., Soricut, R.: All you may need for vqa are image captions. arXiv preprint arXiv:2205.01883 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 547, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 590 + ], + "type": "text", + "content": "13. Changpinyo, S., Sharma, P.K., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3557-3567 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "14. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "15. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 480, + 181 + ], + "type": "text", + "content": "17. Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.: Pix2seq: A language modeling framework for object detection. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 181, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 181, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 138, + 181, + 480, + 214 + ], + "type": "text", + "content": "18. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Changpinyo, S., Wu, J., Ruiz, C.R., Goodman, S., Wang, X., Tay, Y., et al.: Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 480, + 246 + ], + "type": "text", + "content": "19. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "type": "text", + "content": "20. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 321 + ], + "type": "text", + "content": "21. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 365 + ], + "type": "text", + "content": "22. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 321, + 480, + 365 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 321, + 480, + 365 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "text", + "content": "23. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "text", + "content": "24. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 471 + ], + "type": "text", + "content": "25. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=vvoWPYqZJA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 471, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 504 + ], + "type": "text", + "content": "26. Dehghani, M., Gritsenko, A., Arnab, A., Minderer, M., Tay, Y.: Scenic: A jax library for computer vision research and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21393-21398 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 503, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 547 + ], + "type": "text", + "content": "27. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "text", + "content": "28. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 579, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 621 + ], + "type": "text", + "content": "29. Eftekhar, A., Sax, A., Bachmann, R., Malik, J., Zamir, A.R.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 10766-10776 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "30. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Fang, Y., Wang, W., Xie, B., Sun, Q.S., Wu, L.Y., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA: Exploring the limits of masked visual representation learning at scale. ArXiv abs/2211.07636 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 149, + 481, + 182 + ], + "type": "text", + "content": "32. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 182, + 481, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 182, + 481, + 202 + ], + "spans": [ + { + "bbox": [ + 137, + 182, + 481, + 202 + ], + "type": "text", + "content": "33. Geman, S., Bienenstock, E., Doursat, R.: Neural networks and the bias/variance dilemma. Neural computation 4(1), 1-58 (1992)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 203, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 203, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 137, + 203, + 481, + 235 + ], + "type": "text", + "content": "34. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 235, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 235, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 137, + 235, + 481, + 246 + ], + "type": "text", + "content": "35. Goodwin, C.: Seeing in depth. Social studies of science 25(2), 237-274 (1995)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 246, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 246, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 137, + 246, + 481, + 289 + ], + "type": "text", + "content": "36. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6904-6913 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 289, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 289, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 137, + 289, + 481, + 320 + ], + "type": "text", + "content": "37. Guo, B., Zhang, X., Wang, Z., Jiang, M., Nie, J., Ding, Y., Yue, J., Wu, Y.: How close is chatgpt to human experts? comparison corpus, evaluation, and detection. arXiv preprint arXiv:2301.07597 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 320, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 320, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 137, + 320, + 481, + 365 + ], + "type": "text", + "content": "38. Guo, J., Li, J., Li, D., Tiong, A.M.H., Li, B., Tao, D., Hoi, S.: From images to textual prompts: Zero-shot visual question answering with frozen large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10867-10877 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 365, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 365, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 137, + 365, + 481, + 407 + ], + "type": "text", + "content": "39. Gurari, D., Li, Q., Stangl, A.J., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3608-3617 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 137, + 407, + 481, + 439 + ], + "type": "text", + "content": "40. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 137, + 439, + 481, + 472 + ], + "type": "text", + "content": "41. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 137, + 472, + 481, + 504 + ], + "type": "text", + "content": "42. Heek, J., Levskaya, A., Oliver, A., Ritter, M., Rondepierre, B., Steiner, A., van Zee, M.: Flax: A neural network library and ecosystem for jax, 2020. URL http://github.com/google/flax1 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 504, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 504, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 137, + 504, + 481, + 557 + ], + "type": "text", + "content": "43. Herzig, R., Abramovich, O., Ben Avraham, E., Arbelle, A., Karlinsky, L., Shamir, A., Darrell, T., Globerson, A.: Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6803-6815 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 137, + 558, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 558, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 137, + 558, + 481, + 590 + ], + "type": "text", + "content": "44. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 137, + 590, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 590, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 137, + 590, + 481, + 611 + ], + "type": "text", + "content": "45. Howard, I.P., Rogers, B.J.: Binocular vision and stereopsis. Oxford University Press, USA (1995)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 137, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 137, + 611, + 481, + 643 + ], + "type": "text", + "content": "46. Hu, J.E., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Chen, W.: LoRA: Low-rank adaptation of large language models. ArXiv abs/2106.09685 (2021), https://api-semanticscholar.org/CorpusID:235458009" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 137, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 643, + 481, + 665 + ], + "type": "text", + "content": "47. Huang, W., Liu, H., Guo, M., Gong, N.Z.: Visual hallucinations of multi-modal large language models. arXiv preprint arXiv:2402.14683 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "48. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 481, + 203 + ], + "type": "text", + "content": "49. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 203, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 480, + 246 + ], + "type": "text", + "content": "50. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 246, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 279 + ], + "type": "text", + "content": "51. Jian, Y., Gao, C., Vosoughi, S.: Bootstrapping vision-language learning with decoupled language pre-training. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 279, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 480, + 311 + ], + "type": "text", + "content": "52. Kar, O.F., Yeo, T., Atanov, A., Zamir, A.: 3d common corruptions and data augmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18963-18974 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 343 + ], + "type": "text", + "content": "53. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3128-3137 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "text", + "content": "54. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "text", + "content": "55. Korbar, B., Xian, Y., Tonioni, A., Zisserman, A., Tombari, F.: Text-conditioned resampler for long form video understanding. arXiv preprint arXiv:2312.11897 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 440 + ], + "type": "text", + "content": "56. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 440, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 480, + 472 + ], + "type": "text", + "content": "57. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: BLIP: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: International Conference on Machine Learning (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "type": "text", + "content": "58. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "text", + "content": "59. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "60. Lin, J., Yin, H., Ping, W., Lu, Y., Molchanov, P., Tao, A., Mao, H., Kautz, J., Shoeybi, M., Han, S.: Vila: On pre-training for visual language models. arXiv preprint arXiv:2312.07533 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 601 + ], + "type": "text", + "content": "61. Lin, Z., Liu, C., Zhang, R., Gao, P., Qiu, L., Xiao, H., Qiu, H., Lin, C., Shao, W., Chen, K., et al.: Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "62. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "63. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253 (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 137, + 117, + 479, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 117, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 479, + 138 + ], + "type": "text", + "content": "64. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 137, + 139, + 480, + 160 + ], + "type": "text", + "content": "65. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 161, + 480, + 182 + ], + "type": "text", + "content": "66. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "text", + "content": "67. Liu, S., Fan, L.J., Johns, E., Yu, Z., Xiao, C., Anandkumar, A.: Prismer: A vision-language model with an ensemble of experts. ArXiv abs/2303.02506 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 237 + ], + "type": "text", + "content": "68. Lu, J., Clark, C., Lee, S., Zhang, Z., Khosla, S., Marten, R., Hoiem, D., Kembhavi, A.: Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "text", + "content": "69. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-IO: A unified model for vision, language, and multi-modal tasks. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 313 + ], + "type": "text", + "content": "70. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 3195-3204 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 346 + ], + "type": "text", + "content": "71. McKinzie, B., Gan, Z., Fauconnier, J.P., Dodge, S., Zhang, B., Dufter, P., Shah, D., Du, X., Peng, F., Weers, F., et al.: Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 480, + 380 + ], + "type": "text", + "content": "72. Mizrahi, D., Bachmann, R., Kar, O.F., Yeo, T., Gao, M., Dehghan, A., Zamir, A.: 4M: Massively multimodal masked modeling. In: Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "text", + "content": "73. Naeem, M.F., Xian, Y., Zhai, X., Hoyer, L., Van Gool, L., Tombari, F.: Silc: Improving vision language pretraining with self-distillation. arXiv preprint arXiv:2310.13355 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 414, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 479 + ], + "type": "text", + "content": "74. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.Q., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y.B., Li, S.W., Misra, I., Rabbat, M.G., Sharma, V., Synnaeve, G., Xu, H., Jégou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINoV2: Learning robust visual features without supervision. ArXiv abs/2304.07193 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 522 + ], + "type": "text", + "content": "75. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 566 + ], + "type": "text", + "content": "76. Raffel, C., Shazeer, N.M., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 621 + ], + "type": "text", + "content": "77. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "78. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 480, + 138 + ], + "type": "text", + "content": "79. Shen, Y., Heacock, L., Elias, J., Hentel, K.D., Reig, B., Shih, G., Moy, L.: Chatgpt and other large language models are double-edged swords (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 140, + 480, + 171 + ], + "type": "text", + "content": "80. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 172, + 376, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 172, + 376, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 172, + 376, + 182 + ], + "type": "text", + "content": "81. Thorp, H.H.: Chatgpt is fun, but not an author (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 226 + ], + "type": "text", + "content": "82. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "text", + "content": "83. Tiong, A.M.H., Li, J., Li, B., Savarese, S., Hoi, S.C.: Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. arXiv preprint arXiv:2210.08773 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 260, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 480, + 292 + ], + "type": "text", + "content": "84. Tong, S., Liu, Z., Zhai, Y., Ma, Y., LeCun, Y., Xie, S.: Eyes wide shut? exploring the visual shortcomings of multimodal llms. arXiv preprint arXiv:2401.06209 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "text", + "content": "85. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "type": "text", + "content": "86. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 390 + ], + "type": "text", + "content": "87. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 434 + ], + "type": "text", + "content": "88. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "text", + "content": "89. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 511 + ], + "type": "text", + "content": "90. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., Wei, F.: Image as a foreign language: BEiT pretraining for all vision and vision-language tasks. ArXiv abs/2208.10442 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "91. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "92. Xu, J., Zhou, X., Yan, S., Gu, X., Arnab, A., Sun, C., Wang, X., Schmid, C.: Pixel Aligned Language Models. arXiv preprint arXiv: 2312.09237 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "93. Yadav, D., Jain, R., Agrawal, H., Chattopadhyay, P., Singh, T., Jain, A., Singh, S.B., Lee, S., Batra, D.: Towards better evaluation systems for ai agents (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "94. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "95. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "O.F. Kar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 468 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "text", + "content": "96. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 137, + 150, + 481, + 172 + ], + "type": "text", + "content": "97. Yu, S., Cho, J., Yadav, P., Bansal, M.: Self-chained image-language model for video localization and question answering. arXiv preprint arXiv:2305.06988 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 172, + 481, + 205 + ], + "type": "text", + "content": "98. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 137, + 205, + 481, + 237 + ], + "type": "text", + "content": "99. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "type": "text", + "content": "100. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 481, + 303 + ], + "type": "text", + "content": "101. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 303, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 303, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 303, + 481, + 335 + ], + "type": "text", + "content": "102. Zhang, D., Yu, Y., Li, C., Dong, J., Su, D., Chu, C., Yu, D.: Mm-llms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 335, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 335, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 335, + 481, + 369 + ], + "type": "text", + "content": "103. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 481, + 402 + ], + "type": "text", + "content": "104. Zhao, Y., Pang, T., Du, C., Yang, X., Li, C., Cheung, N.M.M., Lin, M.: On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 133, + 402, + 481, + 435 + ], + "type": "text", + "content": "105. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: iBoT: Image BERT pre-training with online tokenizer. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 481, + 468 + ], + "type": "text", + "content": "106. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 174, + 91, + 447, + 102 + ], + "type": "text", + "content": "BRAVE: Broadening the visual encoding of vision-language models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_content_list.json b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1d26acc90cd23a2fa2c5bcf38b4ec05cb6fafa9a --- /dev/null +++ b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_content_list.json @@ -0,0 +1,1871 @@ +[ + { + "type": "text", + "text": "BRIDGE: Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues", + "text_level": 1, + "bbox": [ + 238, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sara Sarto1, Marcella Cornia1, Lorenzo Baraldi1, and Rita Cucchiara1,2", + "bbox": [ + 218, + 210, + 781, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Modena and Reggio Emilia, Italy", + "bbox": [ + 336, + 238, + 668, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 IIT-CNR, Italy", + "bbox": [ + 441, + 253, + 558, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{name.surname}@unimore.it", + "bbox": [ + 403, + 268, + 598, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Effectively aligning with human judgment when evaluating machine-generated image captions represents a complex yet intriguing challenge. Existing evaluation metrics like CIDEr or CLIP-Score fall short in this regard as they do not take into account the corresponding image or lack the capability of encoding fine-grained details and penalizing hallucinations. To overcome these issues, in this paper, we propose BRIDGE, a new learnable and reference-free image captioning metric that employs a novel module to map visual features into dense vectors and integrates them into multi-modal pseudo-captions which are built during the evaluation process. This approach results in a multimodal metric that properly incorporates information from the input image without relying on reference captions, bridging the gap between human judgment and machine-generated image captions. Experiments spanning several datasets demonstrate that our proposal achieves state-of-the-art results compared to existing reference-free evaluation scores. Our source code and trained models are publicly available at: https://github.com/aimagelab/bridge-score", + "bbox": [ + 259, + 314, + 743, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Captioning Evaluation $\\cdot$ Vision-and-Language", + "bbox": [ + 261, + 563, + 656, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 601, + 375, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The objective of image captioning is to produce natural language descriptions conditioned on input images, that closely resemble human language and align to human intentions. As such, the captioning task involves the recognition and understanding of the visual content of the image, including fine-grained elements such as objects, attributes, and their relationships. Advances in training methodologies and architectures have contributed to the progress in the field, significantly improving the generation quality. Recent innovations include fully-attentive models [13, 14, 17], improved connections between visual and textual modalities [14, 40], and the incorporation of objects and tags at an architectural level [3, 31, 62]. Additionally, there has been a notable focus on increasing the robustness of cross-modal features [5, 32, 48], which consequently can increase description accuracy.", + "bbox": [ + 212, + 628, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As constant improvements are made on the generation side, it becomes crucial to enhance the evaluation process as well. In this regard, image captioning", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f147fd6145523e55daa46720fb0e57567e8ca3669b2da5b0801f9c62bd99591c.jpg", + "image_caption": [ + "Fig. 1: Comparison between different captioning evaluation approaches: (a) CIDEr [55] scores candidate and reference captions without considering the input image; (b) CLIP-Score [18] compares text and images using global vectors in a shared embedding space; (c) our BRIDGE, internally builds multimodal pseudo-captions by translating fine-grained image features into pseudo-tokens thanks to a mapping module." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 782, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "evaluation aims to assess the quality of a generated caption given an image and, potentially, human-written reference captions. However, it is important to note that obtaining these reference captions can often be challenging and expensive, adding complexity to the evaluation process. Despite the recent advancements in captioning capabilities, standard automatic evaluation metrics have mainly relied on translation metrics [4,33,41] or text-only ones [2,55,63] which often fall short in capturing aspects such as grammatical correctness, semantic relevance, and specificity. These limitations are worsened by the limited coverage of image content in available references, resulting in inaccurate penalties when generated captions accurately describe novel elements not mentioned in the references.", + "bbox": [ + 212, + 311, + 784, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these limitations, advanced metrics aligning visual and textual data have emerged [18,24,26,27]. Notably, recent metrics leverage the CLIP embedding space [42], which shows a strong correlation with human judgment. Despite the effectiveness of contrastive-based embedding spaces, metrics based on dual-encoder architectures tend to focus on global alignment between an image and its caption, often ignoring fine-grained details or penalizing hallucinations.", + "bbox": [ + 212, + 462, + 784, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following this insight, in this paper, we introduce BRIDGE, a novel learnable and reference-free image captioning metric that enhances the alignment of more fine-grained visual features. Specifically, our model provides a pre-trained dual-encoder architecture with a mapping module designed to effectively exploit visual cues (Fig. 1). This is done by internally creating multimodal pseudo-captions, containing both textual and dense visual features. The process for building these pseudo-captions involves the creation of a template caption, which focuses on the syntactical structure of the scene, and a mapping module. The latter refines the template caption by enriching it with more fine-grained visual features about the subjects depicted in the image. Subsequently, the overall model is trained with a combination of contrastive losses which promote multimodal alignment.", + "bbox": [ + 212, + 553, + 785, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments are conducted on a variety of datasets for image captioning annotated with human rankings, including Flickr8k-Expert, Flickr8k-CF [19], Composite [1], and Pascal-50S [55]. Through comprehensive analysis, we demonstrate the efficacy of the proposed metric and show its ability to overcome the limitations of existing state-of-the-art reference-free alternatives. Additionally, we evaluate its sensitivity to object hallucination by conducting experiments on the FOIL dataset [49]. Overall, BRIDGE outperforms previous metrics and showcases superior performance compared to CLIP-Score [18] and PAC-Score [47].", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Sarto et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. In summary, our contributions are as follows:", + "bbox": [ + 215, + 146, + 666, + 160 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We tackle the limitations of existing image captioning metrics by proposing the first learnable reference-free metric, termed as BRIDGE, that focuses on more fine-grained visual features. Our proposal integrates a dual-encoder architecture with a mapping module which is in charge of producing multimodal pseudo-captions that combine text and richer dense visual features.", + "- Multimodal pseudo-captions are built by creating template captions that describe the scene from a syntactical point of view. These templates are subsequently enriched with fine-grained features through a mapping network.", + "- Experiments, carried out on different datasets with human preferences, demonstrate a higher degree of correlation with respect to existing metrics. We also evaluate the sensitivity of the proposal to objects hallucinations." + ], + "bbox": [ + 223, + 162, + 782, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 352, + 387, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classical Reference-based Captioning Metrics. Several widely used captioning evaluation metrics were originally developed in the context of NLP tasks and rely on n-gram matching techniques. Among these classical metrics, BLEU [41] is designed to focus on precision and incorporates a penalty for sentence brevity. METEOR [4], instead, combines precision and recall to evaluate the quality of captions, while others, such as ROUGE [33], were initially born for summarization tasks and later adapted to image captioning. More recently, two metrics tailored for visual captioning have emerged: CIDEr [55], which measures n-gram similarity and is based on TF-IDF, and SPICE [2], which quantifies graph-based similarity through scene graphs constructed from candidate and reference captions. Overall, focusing on textual-level comparisons, these metrics assume that human-written references accurately represent the image content.", + "bbox": [ + 212, + 381, + 784, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learnable Captioning Metrics. With the rise of large pre-trained models, image captioning evaluation now frequently exploits these models to compare textual-only [60, 63] or visual-textual [18, 21, 22, 26, 27, 57] contents. Notably, the BERT score and its improved version use pre-trained BERT embeddings to compare word tokens in generated and ground-truth sentences.", + "bbox": [ + 212, + 566, + 784, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some metrics, like BLEU and CIDEr, rely solely on text matching between reference captions and machine-generated captions, potentially introducing bias in evaluations due to non-accurate reference captions. To mitigate these issues, alternative solutions leverage the multimodal nature of vision-and-language models. As an example, TIGEr [22] considers the similarities between words and image regions and assesses how well machine-generated captions represent image content and their alignment with human-generated captions.", + "bbox": [ + 212, + 643, + 784, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, other approaches [18, 24, 27] leverage web-scale vision-and-language models such as VilBERT [38] and CLIP [42] for more robust metrics. For example, in [24], CLIP visual-textual features are used to compute negative Gaussian cross-mutual information. Other works, instead, have employed diffusion models in text-only tasks [64] or exploited the zero-shot language modeling capabilities of large language models [8] to evaluate candidate captions.", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reference-free Captioning Metrics. While all aforementioned metrics rely on a set of ground-truth captions to compute the final score, a few attempts have been made to introduce reference-free evaluations, only taking into account the correlation of the candidate caption with the image. In this regard, Lee et al. [26] proposed to fine-tune the UNITER model [10] via contrastive learning to let the model to discriminate between positive and synthetically-generated negative captions. On a different line, Hessel et al. [18] introduced the CLIP-Score, which only relies on a modified cosine similarity between image and candidate caption representations coming from the CLIP model. The recently proposed PAC-Score [47], instead, builds upon the usage of CLIP but incorporates a fine-tuning phase with positive augmentation, further enhancing the accuracy of evaluation. In this paper, we follow this research path and propose a novel learnable, and reference-free evaluation metric that can effectively incorporate fine-grained visual features for evaluating the correlation of image-text pairs.", + "bbox": [ + 212, + 146, + 787, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 BRIDGE for Captioning Evaluation", + "text_level": 1, + "bbox": [ + 215, + 383, + 599, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we present our reference-free captioning evaluation approach, termed BRIDGE. Our approach leverages a dual-encoder architecture, which comprises both a language encoder and a vision encoder. Given a frozen pretrained model, we train a mapping module responsible for filling the holes of a masked template caption with pseudo language tokens that are enriched with visual information. An overview of our model is depicted in Fig. 2.", + "bbox": [ + 212, + 414, + 787, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 530, + 372, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach relies on CLIP (Contrastive Language-Image Pre-training) [42], a powerful vision and language model designed to align images and corresponding text captions within a shared embedding space. For a given input image $I$ , the image encoder $E_V$ extracts the visual information $v = E_V(I) \\in \\mathbb{R}^d$ . On the textual side, an input caption $T$ is tokenized and a textual representation is obtained by passing it through the textual encoder $E_T$ , obtaining $t = E_T(T) \\in \\mathbb{R}^d$ . Once the textual and visual features, $t$ and $v$ respectively, are projected in a common space, visual and textual inputs can be compared via cosine similarity.", + "bbox": [ + 212, + 556, + 787, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The relationships learned by CLIP can be exploited to build an image captioning evaluator. In CLIP-Score [18] the authors directly compare candidate captions and images in the embedding space and show that this achieves a good correlation with human judgments. In detail, to assess the quality of a candidate generation, they feed both the image and the candidate caption through their respective feature extractors, and compute the cosine similarity of the resultant embeddings:", + "bbox": [ + 212, + 679, + 787, + 784 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C L I P} \\text {- S c o r e} (I, T) = w \\cdot \\max (\\cos (v, t), 0), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 785, + 785, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w$ is a rescaling factor employed to stretch the score distribution while ensuring the ranking results remain unchanged.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Sarto et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c2d21a265ec7cb0e78f844dea2608c71b92de6044e3a19960a8c5df6c8fbaee.jpg", + "image_caption": [ + "Fig. 2: Overview of the BRIDGE evaluation approach. Starting from a template caption, a mapping network augments it with dense visual features, obtaining a pseudocaption which is then used for computing image-text similarities." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 782, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Injecting Fine-Grained Visual Features", + "text_level": 1, + "bbox": [ + 214, + 409, + 584, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike CLIP-Score, our approach does not rely exclusively on global image descriptors for evaluating image-text alignments. Instead, we focus on employing stronger visual information. To do so, we draw inspiration from the Pic2Word approach [46] and represent the input image through a multimodal pseudo-caption, an embedding representation that contains stronger visual elements.", + "bbox": [ + 212, + 431, + 784, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building Template Captions. In order to create multimodal pseudo-captions, we first build template captions for a given input image. These are skeletal textual representations of the image, obtained by masking out all the relevant textual concepts from the descriptions generated by a captioner. Through these template captions, we aim to provide the model only with a templated textual structure which can then be filled with more fine-grained visual features. In particular, given an automatically generated caption describing the input image, such as 'A man running with a white dog', we remove the main subjects within the caption (e.g. 'man' and 'white dog') and mask them with a [MASK] token. This will allow the model to fill in these gaps by incorporating more fine-grained features from the image encoder. Since a primary subject might be described by words other than just its corresponding nouns (e.g. adjectives), we utilize noun chunks. Fig. 3 reports template captions and corresponding noun chunks.", + "bbox": [ + 212, + 511, + 785, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a sentence containing $N$ noun chunks, we independently encode them through the mapping network. To this aim, we replicate the template caption as many times as the number of noun chunks and mask a different noun chunk in each of the replicas. We thus obtain $N$ different versions of the template caption, each one masking a single noun chunk, for instance", + "bbox": [ + 212, + 707, + 785, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[A [MASK] running with a white dog',", + "bbox": [ + 343, + 792, + 658, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "'A man running with a [MASK]'.", + "bbox": [ + 367, + 811, + 632, + 825 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues", + "bbox": [ + 236, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e9b1d339c07d6c2584bb07f66cbc26842c5bb1254eaf9bbe4b187ed5e3d5e3dd.jpg", + "image_caption": [ + "COCO", + "TEMPLATE", + "NOUN CHUNKS", + "yellow bus", + "Fig. 3: COCO captions with template captions and associated noun chunks." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 316, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a46b7b22878b3171115f7ebea4c2888bb32cfcce876bfbac649c85e857d428d0.jpg", + "image_caption": [ + "People on a motorcycle in front of a road", + "A yellow bus driving down a street next to a park.", + "street", + "park", + "COCO", + "TEMPLATE", + "NOUN CHUNKS", + "Little dog laying on a pillow next to the laptop.", + "A dog laying on a couch next to a PC.", + "PC", + "dog", + "couch" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 599, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NOUN CHUNKS", + "bbox": [ + 607, + 190, + 640, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mapping with Fine-Grained Visual Features. The above-described masked replicas are then fed to the mapping module $\\psi$ . Specifically, our approach exploits the visual information extracted from the visual encoder $E_V$ to enrich the replicas with the informative content of the image $I$ . To get a more fine-grained representation of the image, we directly take the grid of features from the last layer, $\\widehat{v}$ . For instance, in the case of a ViT-B/32 backbone, this will have a shape of $50 \\times d$ , where $d$ is the dimensionality of the last embedding of the network.", + "bbox": [ + 212, + 244, + 785, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The mapping network is implemented as a stack of Transformer [54] encoder layers interleaved with cross-attention layers. Its role is to refine each template captions with visual information. Since each template caption is processed independently, the mapping module returns a set of sequences, each with the same length as the corresponding input template caption. From the output of the mapping module, we keep only the predictions for the masked tokens in each template caption and copy them back into the original templates.", + "bbox": [ + 212, + 351, + 785, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, by providing a masked input template in the form $\\hat{T_i} = [w_1,\\dots,w_{j - 1},\\mathsf{MASK},w_{j + 1},\\dots,w_T]$ , where $\\{w_j\\}_j$ represent original tokens from the input caption, we obtain $\\hat{T_i}^* = \\left[w_1,\\dots,w_{j - 1},\\psi (\\hat{T_i})_j,w_{j + 1},\\dots,w_T\\right]$ , where $\\psi (\\hat{T_i},\\widehat{v})_j$ represents the output of the mapping network at the position corresponding to the masked input token position. In the case of noun chunks consisting of more than one token, multiple consecutive tokens are replaced with the corresponding outputs from the mapping network. By injecting the outputs of the mapping token into the initial template caption, we effectively complete the original templates with visually enriched vectors. Notably, these newly generated pseudo-captions combine word sequences from the template captions with dense vectors obtained by the mapping module. Consequently, they cannot be decoded as standard captions. As a last step, the obtained pseudo-captions are fed into the pre-trained CLIP language encoder.", + "bbox": [ + 212, + 457, + 787, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Training Protocol", + "text_level": 1, + "bbox": [ + 214, + 690, + 408, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train our mapping network, the loss is defined as a weighted version of the symmetric InfoNCE loss [39], where positive and negative items are weighted according to the number of noun chunks in each caption.", + "bbox": [ + 212, + 718, + 785, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, given a batch in the form $\\mathcal{B} = \\{(I_i,T_i)\\}_{i = 1}^N$ , where $I_{i}$ and $T_{i}$ represent image-caption pairs, each image $I_{i}$ is expanded in $N_{i}$ multimodal pseudocaptions as outlined above, where $N_{i}$ is the number of noun chunks in caption $T_{i}$ . Further, let $t_{ij}^{*}$ represent the embedding vector of the $j$ -th pseudo-caption derived from the $i$ -th image, $v_{i}$ the embedding vector of the $i$ -th image and $t_{i}$", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Sarto et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the embedding vector of the $i$ -th ground-truth caption. Finally, let $M$ be the total number of noun chunks in the mini-batch, i.e. $M = \\sum_{i=1}^{N} N_i$ .", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The first loss, denoted as $\\mathcal{L}_1$ , tries to align the pseudo-captions $t_{ij}^*$ with the global visual features of the corresponding images $v_i$ . In addition to this loss term, we define a second loss component $\\mathcal{L}_2$ that promotes the alignment between pseudo-captions $t_{ij}^*$ and the textual feature vector of the ground-truth caption $t_i$ corresponding to the input image. This ensures that pseduo-captions are aligned also on a textual space, in addition to being aligned in the image space. Additional details can be found in the supplementary materials.", + "bbox": [ + 212, + 179, + 784, + 285 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regularization Branch. With the aforementioned loss terms, our objective is encouraging an association between each pseudo-caption and its corresponding image and caption. However, it is also important to differentiate each pseudocaption from the others of the same image. To achieve this, we define a regularization loss which promotes a precise alignment between each pseudo-caption and the corresponding noun chunk.", + "bbox": [ + 212, + 291, + 782, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we create prompts like \"a photo of a \" and encode them with the text encoder $E_{T}$ . In parallel, each pseudo-caption is fed into a dedicated multilayer perceptron (MLP) projection, which consists of two linear layers with a ReLU activation in between. Formally, the branch is defined as", + "bbox": [ + 212, + 383, + 782, + 444 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} (x) = \\operatorname {L i n e a r} (\\operatorname {R e L U} (\\operatorname {L i n e a r} (x))). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 460, + 782, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our goal is to emphasize each pseudo-caption's alignment with its corresponding noun chunk, we employ a regular contrastive loss $\\mathcal{L}_3$ between the prompts mentioned earlier and the outputs of the projection branch.", + "bbox": [ + 212, + 489, + 782, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the overall loss function we train BRIDGE is defined as a weighted summation of two aforementioned losses, plus the regularization loss, as", + "bbox": [ + 212, + 537, + 782, + 566 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {2} \\mathcal {L} _ {2} + \\lambda_ {3} \\mathcal {L} _ {r}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 583, + 782, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Inference and Score Computation", + "text_level": 1, + "bbox": [ + 214, + 626, + 539, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At inference time, given an image-candidate caption pair $(I,T)$ , we extract all pseudo-captions from $I$ using our mapping network. Subsequently, we compute the mean pseudo-caption embedding as $t^{*} = \\frac{1}{N}\\hat{t}_{i}^{*}$ , where $\\hat{t}_i^*$ indicates the $i$ -th pseudo-caption extracted from $I$ and $N$ here indicates the overall number of pseudo-captions associated with $I$ .", + "bbox": [ + 212, + 657, + 784, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At that point, given the visual embedding $v$ of the image and the embedding of the candidate caption $t$ , the matching score between $I$ and $T$ is defined as", + "bbox": [ + 212, + 734, + 782, + 763 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {B R I D G E} (I, T) = 0. 5 \\cdot [ \\operatorname {C L I P - S c o r e} (I, T) + w \\cdot \\max (\\cos \\left(t ^ {*}, t\\right), 0) ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 795, + 782, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\cos$ indicates the cosine similarity and $w$ is a constant scaling factor.", + "bbox": [ + 214, + 824, + 751, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues", + "bbox": [ + 236, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Evaluation", + "text_level": 1, + "bbox": [ + 215, + 143, + 491, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 172, + 455, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architecture and Training Details. Building upon prior research [18,24,50], we use either CLIP [42] ViT-B/32 or ViT-L/14 as backbone for the visual and textual encoder. The mapping module is composed of two Transformer layers and is trained on the COCO dataset [34], which contains more than 120k images annotated with five captions. In particular, we employ the splits introduced by Karpathy et al. [23], where 5,000 images are used for both validation and testing and the rest for training. To map the grid visual features to an embedding space of dimension 512, we employ a simple linear projection. For the regularization branch, we utilize a two-layer multi-layer perceptron.", + "bbox": [ + 212, + 196, + 785, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During training, we use AdamW [37] as optimizer with a learning rate equal to 0.0001 and a batch size of 256. The $\\lambda_1$ , $\\lambda_2$ , and $\\lambda_3$ values are selected with a grid search, choosing the combination that provides the best validation loss. Specifically, we set both $\\lambda_1$ and $\\lambda_3$ to 0.01, while $\\lambda_2$ is set to 1.0. The training stage lasts around one day on a single A100 GPU.", + "bbox": [ + 212, + 333, + 785, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Template Caption Generation. The template captions used as input for the mapping module are generated using the BLIP model [29]. In particular, we use the ViT-L/14 version pre-trained on 129M image-text pairs and finetuned on the COCO dataset. After this generation phase, the primary subjects of the template sentences are extracted by using the NLTK library [6]. During training, two noun chunks are randomly chosen from the set identified during the extraction step. In the evaluation phase, otherwise, all identified noun chunks are included.", + "bbox": [ + 212, + 412, + 785, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Datasets", + "text_level": 1, + "bbox": [ + 215, + 540, + 333, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the correlation of the proposed metric with human ratings, we conduct experiments on the Flickr8k-Expert, Flickr8k-CF, Composite, and Pascal50-S datasets [1,19,55]. In addition, for detecting hallucinations in textual sentences, we extend our analysis to the FOIL dataset [49]. Except for Pascal50S and FOIL in where accuracy scores are used, evaluation on all other datasets relies on Kendall $\\tau_{b}$ , Kendall $\\tau_{c}$ , and Spearman $\\rho$ correlation scores.", + "bbox": [ + 212, + 564, + 785, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Flickr8k-Expert and Flickr8k-CF [19]. These datasets consist of image-caption pairs with corresponding human ratings. Specifically, Flickr8k-Expert comprises 17k expert annotations for visual-textual pairs, with a total of 5,664 different images. Each pair receives a score ranging from 1 (lack of correlation) to 4 (accurate depiction), where 1 indicates a lack of correlation between the caption and the image, and 4 indicates an accurate depiction of the image without errors. On the other hand, Flickr8k-CF is composed of 145k binary quality judgments, collected from CrowdFlower, for 48k image-caption pairs containing 1,000 unique images. Each pair is annotated with at least three binary scores, where “yes” denotes that the caption correlates with the image. To measure the alignment with human judgment, we compute the mean proportion of “yes” annotations as the score for each pair.", + "bbox": [ + 212, + 657, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Sarto et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ddb339a0e6cc0d78a427ea55c7cd61961f6c1ac1f9bdd7e34776ded7888346e2.jpg", + "table_caption": [ + "Table 1: Ablation study results. US indicates the number of pseudo tokens." + ], + "table_footnote": [], + "table_body": "
Flickr8k-ExpertFlickr8k-CFPascal-50S
Kend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρAccuracy
Architectural Components
w/o mapping module (i.e. MLP)53.153.565.335.518.343.581.9
w/o template captions53.754.166.035.518.443.682.5
w/o regularization branch54.154.566.535.718.543.682.7
Score Formulation
w/o textual similarity51.151.263.034.417.730.580.9
w/o visual similarity53.854.266.035.418.343.781.9
Pseudo-token Size
w/ US = 154.154.566.435.117.130.382.6
w/ US = 254.154.566.436.118.744.482.8
w/ US = 454.354.766.635.418.343.882.4
w/ US = 854.054.466.335.918.644.581.7
BRIDGE (US = 3)54.454.867.736.118.744.582.6
", + "bbox": [ + 218, + 167, + 782, + 358 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Composite [1]. It contains 12k human ratings for image-caption pairs including a combination of images taken from COCO [34] (2,007 images), Flickr8k [19] (997 images), and Flickr30k [61] (991 images). In this dataset, human evaluators assess each image-caption pair, assigning a score within the range of 1 to 5 to estimate the correspondence of the caption with the associated image.", + "bbox": [ + 212, + 373, + 784, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pascal50-S [55]. It presents pairwise preference judgments between two captions. Overall, the dataset consists of 4,000 sentence pairs, each of them associated with an image from the UIUC Pascal sentence dataset [45]. Each pair is associated with 48 human judgments, where each evaluation indicates which sentence better describes the given image. The sentence pairs are categorized into four groups: (i) both human-written and correct captions (HC), (ii) both human-written captions where one is correct and the other is wrong (HI), (iii) both correct captions but one written by humans and the other machine-generated (HM), (iv) both machine-generated and correct captions (MM).", + "bbox": [ + 212, + 452, + 784, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "FOIL [49]. The dataset comprises image-caption pairs from the COCO dataset [34]. In this scenario, captions undergo perturbation by generating modified versions that closely resemble the originals but introduce a single error, referred to as \"foil word\". For a fair comparison, we select the subset of the validation set that does not overlap with the portion of COCO used during training, resulting in 8,000 images, each paired with a foil-correct textual counterpart.", + "bbox": [ + 212, + 590, + 784, + 683 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Ablation Studies and Analysis", + "text_level": 1, + "bbox": [ + 214, + 705, + 511, + 720 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our metric, we start by analyzing variations of our main architectural components. Then, we assess the impact of caption templates in our score formulation. All these experiments are performed using CLIP ViT-B/32 as backbone and reported in Table 1.", + "bbox": [ + 212, + 729, + 784, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Contribution of Architectural Components. We first investigate the performance of the most straightforward implementation of a mapping module, structured as a two-layer MLP following [46]. We also validate the importance", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b343b42c025e3a411eed6cd5f9aabb952228e74361d48fe93557f21f576fed8c.jpg", + "table_caption": [ + "Table 2: Impact of different template captions." + ], + "table_footnote": [], + "table_body": "
ExpertCFPascal-50S
Kend. τbKend. τbAcc.
Transformer Templates
w/o mapping module46.131.680.4
BRIDGE54.035.982.7
(+7.9)(+4.3)(+2.3)
BLIP Templates
w/o mapping module48.633.881.0
BRIDGE54.436.182.6
(+5.8)(+2.3)(+1.6)
BLIP-2 Templates
w/o mapping module49.434.282.4
BRIDGE54.436.282.9
(+5.0)(+2.0)(+0.5)
", + "bbox": [ + 222, + 171, + 529, + 329 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/612a9f2728812130403915490f8143dc29b7208409534a8d155ef174c35bc96e.jpg", + "image_caption": [ + "C C C C C C", + "#", + "B" + ], + "image_footnote": [], + "bbox": [ + 576, + 145, + 669, + 198 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4bddc7b4c62a1eac6bbaba29d702bab35e35a37a90d2c72e28642d5fba601265.jpg", + "image_caption": [ + "TRANSE" + ], + "image_footnote": [], + "bbox": [ + 678, + 146, + 769, + 198 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A young girl is preparing to blow out her candle", + "bbox": [ + 584, + 203, + 665, + 232 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A painting of a table with fruit on top of it", + "bbox": [ + 684, + 203, + 764, + 231 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A young girl is eating a piece of cake", + "bbox": [ + 584, + 234, + 660, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A painting of a painting of fruits and a painting", + "bbox": [ + 686, + 234, + 759, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A woman blowing out a candle on a cake", + "bbox": [ + 583, + 268, + 663, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A painting of a pitcher and oranges on a table", + "bbox": [ + 683, + 268, + 766, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A girl blowing out a candle on a cupcake", + "bbox": [ + 581, + 301, + 663, + 330 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A painting of a still life with oranges and a candle", + "bbox": [ + 684, + 301, + 766, + 330 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of the template captions through a model variant in which a set of learnable tokens $\\mathsf{S}^*$ serves as input for the mapping module, without relying on template captions. In both variants, given the absence of template captions, we construct a template such as 'a photo of $\\mathsf{S}^*$ ' and extract its features using the CLIP text encoder. In the Table, it can be seen that, regardless of any architectural changes, it is important to provide a simple sentence structure to the mapping module to achieve competitive performance.", + "bbox": [ + 212, + 349, + 787, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In addition to these baselines, we devise a variant to analyze the contribution of the regularization branch. In this setting, we employ template captions as input for the mapping module, resulting in a substantial improvement of $+1.0$ Kendall $\\tau_{b}$ and $+0.8$ accuracy points compared to the MLP variant, respectively on the Flickr8k-Expert and on the Pascal-50S dataset. When introducing the regularization branch (i.e. the complete BRIDGE architecture), further enhancements can be observed especially on the Flickr8k-CF with an improvement of $+0.4$ points in terms of the Kendall $\\tau_{b}$ correlation score.", + "bbox": [ + 212, + 455, + 787, + 577 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We also emphasize the importance of each component in our score formulation. Specifically, we present correlation results when employing only the visual similarity within our architecture, which is the original CLIP-Score formulation. As observed, performance drops drastically when relying only on visual information. A less significant drop is observed when employing only textual similarity.", + "bbox": [ + 212, + 578, + 787, + 654 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As an additional analysis, we report the effect of changing the number of the pseudo tokens for each noun chunk, denoting it as unit size (US). Specifically, we compute the scores employing $\\mathrm{US} = 1,2,3,4,8$ . From the results, it can be seen that $\\mathrm{US} = 3$ generally leads to the best performance across nearly all evaluation metrics. This configuration is used in all experiments reported in the paper.", + "bbox": [ + 212, + 654, + 787, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Analysis on Caption Templates. We analyze the effect of changing the initial template captions for our model. Specifically, we employ template captions generated by a conventional Transformer-based captioner that uses CLIP features as input and is trained only on the COCO dataset, as well as those generated by BLIP [29], and BLIP-2 [28]. Notably, we select template captions of different quality based on both standard metric evaluations and correlations with human judgment. To assess the quality of the raw generated templates, we observe that", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Sarto et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the CIDEr score of these models on the COCO test set is equal to 114.2, 131.4, and 145.8 respectively for the standard Transformer model, BLIP, and BLIP-2. Note that all models were trained with cross-entropy loss only. Results on the Flickr8k-Expert, Flickr8k-CF, and Pascal-50S datasets are reported in Table 2. To qualitatively validate the generated templates, we include sample captions generated by the three models compared to a ground-truth caption from the COCO test set. Specifically, captions generated by BLIP-2 are generally more detailed and effectively describe the visual content of the input image compared to those generated by BLIP and, notably, the standard Transformer model.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For each caption template source, we compute the correlation scores when the mapping module is disabled and the captions are directly fed to the text encoder. We compare it with our standard BRIDGE score, considering the different caption templates. Across all datasets, it is evident that directly using the caption templates as input to the text encoder leads to poor performance. This highlights the intended flexibility of template captions as skeletal representations, allowing the model to enhance them with fine-grained visual features.", + "bbox": [ + 212, + 282, + 787, + 388 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In fact, starting from these simple template captions and following our approach, we achieve improvements of $+5.8$ and $+2.3$ Kendall $\\tau_{b}$ points and $+1.6$ accuracy points, respectively, on the Flickr8k-Expert, Flickr8k-CF, and Pascal50-S datasets when using BLIP caption templates. The overall best results are with captions from the BLIP-2 model, confirming that better templates can indeed lead to improved results. However, even when using lower-quality captions, the final correlation results are very close to those obtained with higher-quality captions. This highlights the robustness of our metric to caption templates of varying quality and that it is not necessary to rely on captions generated by large-scale captioners to achieve strong correlation scores.", + "bbox": [ + 212, + 388, + 787, + 540 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Comparison with State-of-the-Art Captioning Metrics", + "text_level": 1, + "bbox": [ + 214, + 559, + 705, + 575 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluating Sample-Level Human Correlation. We evaluate the sample-level human correlation on the Flickr8k [19] and Composite [1] datasets. Following previous works [18,63], we compute Kendall correlation scores in both $\\tau_{b}$ and $\\tau_{c}$ versions and also include the Spearman $\\rho$ score. Results are reported in Table 3, where we compare our proposed BRIDGE metric against other reference-free evaluation scores like UMIC [26], CLIP-S [18], and PAC-S [47]. Moreover, we also compare with standard captioning evaluation metrics (i.e. BLEU [41], ME-TEOR [4], CIDEr [55], and SPICE [2]) and more recent reference-based solutions that exploit text-only or cross-modal learned embeddings, such as BERT-S [63], BERT-S++ [60], TIGEr [22], VilBERTScore [27], and MID [24]. For completeness, we also include the reference-based versions of CLIP-S and PAC-S, termed RefCLIP-S and RefPAC-S, which however are not directly comparable with our solution as both rely on a set of five reference captions.", + "bbox": [ + 212, + 583, + 787, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As it can be seen, BRIDGE outperforms other reference-free metrics in terms of correlation with human judgment, achieving the highest scores on almost all datasets. Specifically, compared to CLIP-S, BRIDGE shows improvements in terms of Kendall $\\tau_{c}$ of $+3.6$ and $+2.8$ points when using ViT-B/32 and ViT-L/14", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues", + "bbox": [ + 236, + 113, + 732, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f005db42a60070404eee56074b7b2bbc930d6acdd0e2a1312a840639c14ff025.jpg", + "table_caption": [ + "Table 3: Correlation scores on Flickr8k-Expert, Flickr8k-CF, and Composite [1, 19]." + ], + "table_footnote": [], + "table_body": "
Flickr8k-ExpertFlickr8k-CFComposite
Kend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρ
Reference-based metrics
BLEU-4 [41]30.630.838.716.98.721.028.330.638.1
METEOR [4]41.541.851.922.211.527.636.038.948.1
CIDEr [55]43.643.954.324.612.730.534.937.747.0
SPICE [2]51.744.955.124.412.031.338.840.349.1
BERT-S [63]-39.250.322.8--39.930.148.6
BERT-S++ [60]48.146.756.9---42.344.952.1
TIGER [22]51.449.348.6---47.545.455.3
ViLBERTScore [27]54.250.161.3---51.452.458.7
MID [24]-54.9-37.3-----
RefCLIP-S [18]52.653.064.636.418.844.751.255.466.2
RefPAC-S [47]55.555.967.637.619.546.253.057.368.0
Reference-free metrics
UMIC [26]-46.8-------
CLIP-S (ViT-B/32) [18]51.151.263.034.417.730.549.853.864.3
PAC-S (ViT-B/32) [47]53.954.366.136.018.644.451.555.766.3
BRIDGE (ViT-B/32)54.454.866.736.118.744.550.955.065.4
CLIP-S (ViT-L/14) [18]52.653.064.735.218.243.351.355.465.9
PAC-S (ViT-L/14) [47]55.155.567.336.819.045.352.356.567.1
BRIDGE (ViT-L/14)55.455.867.736.319.044.752.957.267.8
", + "bbox": [ + 218, + 166, + 784, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "backbone on the Flickr8k-Expert dataset. These improvements extend consistently across all datasets and correlation scores. Compared to PAC-S, BRIDGE still achieves superior results on both Flickr8k-Expert and Flickr8k-CF, while performing on par on the Composite dataset (i.e. PAC-S achieves the best results when using ViT-B/32, while BRIDGE outperforms PAC-S using ViT-L/14).", + "bbox": [ + 212, + 453, + 787, + 530 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Discriminating Correct Captions. We also evaluate the effectiveness of our metric on the PASCAL-50S dataset [55]. In this context, instead of", + "bbox": [ + 212, + 537, + 787, + 568 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "calculating correlation scores, we compute accuracy by determining, for each pair, the caption favored by the majority of human ratings as correct (with ties being resolved randomly). We then measure how frequently the evaluation metric assigns a higher score to the chosen caption. Following previous works [18], we randomly select five reference captions from the set of 48 provided by the dataset and average the results over five distinct draws. Accuracy values are reported in Table 4. The results show that when using", + "bbox": [ + 212, + 569, + 444, + 840 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/17092aa887071e5d2f2852c70f9bfbf5b37655bd9f053d097ebcf3629bb65d97.jpg", + "table_caption": [ + "Table 4: Accuracy results on Pascal-50S [55] averaged over five random draws of reference captions. The $\\dagger$ marker indicates scores from previous works." + ], + "table_footnote": [], + "table_body": "
HCHIHMMMMean
Reference-based metrics
BLEU-4 [41]60.393.185.757.074.0
METEOR [4]66.097.794.066.681.1
CIDEr [55]66.597.990.765.280.1
BERT-S++†[60]65.498.196.460.380.1
TIGER†[22]56.099.892.874.280.7
MID†[24]67.099.797.476.885.2
RefCLIP-S [18]64.999.595.573.383.3
RefPAC-S [48]67.799.696.075.684.7
Reference-free metrics
CLIP-S (ViT-B/32) [18]55.999.396.572.080.9
PAC-S (ViT-B/32) [47]60.699.396.972.982.4
BRIDGE (ViT-B/32)59.499.497.574.082.6
CLIP-S (ViT-L/14) [18]57.099.696.773.581.7
PAC-S (ViT-L/14) [47]59.599.495.874.782.2
BRIDGE (ViT-L/14)61.299.696.674.182.9
", + "bbox": [ + 465, + 626, + 769, + 830 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Sarto et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b8227c3c7cdf1b96f90487fce67a5fa807108b59008efc1a9fd3a6303ba53ff3.jpg", + "image_caption": [ + "Fig. 4: Sample images from the FOIL dataset [49] and corresponding scores generated by our proposed metric compared with CLIP-S and PAC-S." + ], + "image_footnote": [], + "bbox": [ + 218, + 152, + 383, + 315 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b984061a8407b42408d3544b43b77986c02ab8fcce872b39e45f71c29da2d9eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 152, + 557, + 315 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2c0964f9a6d38e4c216aecf52f65f8ef967b3eafdee40e1f2dc30b8cf7e7ed15.jpg", + "table_caption": [ + "Table 5: Accuracy results on the FOIL [49] dataset." + ], + "table_footnote": [], + "table_body": "
Acc.
Reference-based metrics
BLEU-4 [41]66.2
METEOR [4]70.1
CIDEr [55]85.7
MID [24]90.5
RefCLIP-S [18]91.0
RefPAC-S [18]93.7
Reference-free metrics
CLIP-S (ViT-B/32) [18]87.2
PAC-S (ViT-B/32) [47]89.9
BRIDGE (ViT-B/32)91.5
CLIP-S (ViT-L/14) [18]90.9
PAC-S (ViT-L/14) [47]91.9
BRIDGE (ViT-L/14)93.0
", + "bbox": [ + 575, + 189, + 751, + 373 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "both ViT-B/32 and ViT-L/14, BRIDGE consistently outperform CLIP-S across all categories, showcasing an average accuracy increase of $+1.7$ and $+1.2$ points, respectively. When comparing with PAC-S, our solution can better discriminate the correct captions on both backbones with an average accuracy increase of $+0.2$ and $+0.7$ using ViT-B/32 and ViT-L/14 respectively.", + "bbox": [ + 212, + 386, + 787, + 463 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Object Hallucination Analysis. We then extend our analysis to the FOIL dataset [49] for correctly identifying captions that may contain object hallucinations. Table 5 shows the accuracy results. As it can be seen, BRIDGE outperforms CLIP-S and PAC-S, exhibiting an increase respectively of $+4.3$ and $+1.6$ points when using ViT-B/32 as backbone. Similar improvements can also be observed when employing more robust visual features. These results demonstrate the capabilities of our metric also to identify hallucinated objects correctly. In Fig 4, we present sample results comparing our metric with CLIP-S and PAC-S.", + "bbox": [ + 212, + 467, + 787, + 590 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 System-level Correlation", + "text_level": 1, + "bbox": [ + 215, + 614, + 465, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Finally, we delve into the efficacy of our proposed metric when evaluating popular existing captioning models. To this aim, we generate predictions of several state-of-the-art captioning models on the COCO test set, including Show and Tell and Show, Attend and Tell which are among the first image captioning models based on deep learning, Up-Down [3], SGAE [59], AoANet [20], $\\mathcal{M}^2$ Transformer [14], X-Transformer [40] which all include region-based image features with either LSTM-based or Transformer-based language models, and the recently proposed COS-Net model [32] that incorporates CLIP features. In addition to reporting evaluations on traditional captioning models, we also include recent LLM-based captioning models, including ZeroCap [51] and SmallCap [44], which are based on GPT-2 [43], and MiniGPT-v2 [9], BLIP-2 [28], IDEFICS [25], LLaVA-1.5 [35,36], and InstructBLIP [15] which instead are based on larger-scale LLMs like FlanT5 [12], Vicuna [11], or LLaMA [52,53].", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e28418cfee9f72c1a8927c81fd0535480991b5e00c82ca2fc3af6f962b755e3f.jpg", + "table_caption": [ + "Table 6: Evaluation scores of traditional and LLM-based captioners on COCO test set ( $\\spadesuit$ : not trained/fine-tuned on COCO)." + ], + "table_footnote": [], + "table_body": "
BLEU-4METEORCIDErCLIP-SPAC-SBRIDGE
TraditionalShow and Tell [56]31.425.097.20.5720.7720.788
Show, Attend and Tell [58]33.426.2104.60.5820.7850.804
Up-Down [3]36.727.9122.70.7230.8030.821
SGAE [59]39.028.4129.10.7340.8120.833
AoANet [20]38.929.2129.80.7370.8150.836
M2Transformer [14]39.129.2131.20.7340.8130.841
X-Transformer [40]39.729.5132.80.6100.8120.845
COS-Net [32]42.030.6141.10.7580.8320.859
LLM-basedZeroCap♦[51]2.310.115.10.8100.8160.862
SmallCap [44]37.027.9119.70.7480.8260.847
MiniGPT-v2 [9]18.824.680.40.7520.8180.845
BLIP-2 [28]43.732.0145.80.7670.8370.868
IDEFICS-9B♦[25]4.319.150.00.7400.7860.838
LLaVA-1.5-7B♦[35]8.128.069.60.7840.8090.867
InstructBLIP-Flan-T5-XL♦[15]6.128.138.10.8170.8370.902
Humans-24.187.60.7740.8230.856
", + "bbox": [ + 243, + 178, + 750, + 377 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The results are presented in Table 6, where we evaluate our BRIDGE scores against both standard metrics, such as BLEU-4, METEOR, and CIDEr, and more recent ones like CLIP-S and PAC-S. Captioning models are compared to a human baseline, in which, for each sample, one human-annotated sentence (selected randomly from the five provided by the COCO dataset) serves as a candidate caption. As shown in the table, BRIDGE can effectively evaluate human-annotated sentences which obtain a score similar to recent state-of-the-art captioning models such as COS-Net. This capability lacks in standard metrics such as METEOR and CIDEr which rank human captions lower than those generated by less-performing captioning models like Show, Attend, and Tell or Up-Down.", + "bbox": [ + 212, + 393, + 784, + 545 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "When considering the evaluation of captions generated by existing models, our metric shows a strong correlation with standard evaluation metrics when evaluating traditional image captioners or large-scale models that are fine-tuned on the COCO dataset like BLIP-2. When instead considering more recent approaches that are based on large language models and are not fine-tuned on COCO, BRIDGE still recognizes the goodness of generated captions, raking InstructBLIP as the best-performing approach. This demonstrates the capabilities of our metric to correctly evaluate longer and more detailed captions which are typically generated by LLM-based multimodal models [7, 16, 30] and that might be very different from captions contained in the COCO dataset.", + "bbox": [ + 212, + 546, + 785, + 696 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 720, + 359, + 736 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have presented a novel learnable, and reference-free image captioning metric that combines text and dense visual features. Our proposal, BRIDGE, employs templated captions that are enriched with fine-grained visual cues thanks to a mapping network. Through experimental evaluation, we demonstrate that BRIDGE outperforms existing reference-free metrics in terms of correlation with human judgment and sensitivity to hallucinated objects.", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Sarto et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been conducted under a research grant co-funded by Leonardo S.p.A. and supported by the PNRR-M4C2 (PE0000013) project \"FAIR - Future Artificial Intelligence Research\" and by the PRIN project \"MUSMA\" (CUP G53D23002930006 - M4C2 I1.1), both funded by EU - Next-Generation EU.", + "bbox": [ + 212, + 175, + 787, + 266 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 287, + 321, + 304 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aditya, S., Yang, Y., Baral, C., Fermuller, C., Aloimonos, Y.: From Images to Sentences through Scene Description Graphs using Commonsense Reasoning and Knowledge. arXiv preprint arXiv:1511.03292 (2015)", + "2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: SPICE: Semantic Propositional Image Caption Evaluation. In: ECCV (2016)", + "3. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR (2018)", + "4. Banerjee, S., Lavie, A.: METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In: ACL Workshops (2005)", + "5. Barraco, M., Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: With a Little Help from your own Past: Prototypical Memory Networks for Image Captioning. In: ICCV (2023)", + "6. Bird, S., Klein, E., Loper, E.: Natural Language Processing with Python: Analyzing Text with the Natural Language Toolkit. O'Reilly Media, Inc. (2009)", + "7. Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Baraldi, L., Cornia, M., Cucchiara, R.: The Revolution of Multimodal Large Language Models: A Survey. In: ACL Findings (2024)", + "8. Chan, D., Petryk, S., Gonzalez, J.E., Darrell, T., Canny, J.: CLAIR: Evaluating Image Captions with Large Language Models. In: EMNLP (2023)", + "9. Chen, J., Li, D.Z.X.S.X., Zhang, Z.L.P., Xiong, R.K.V.C.Y., Elhoseiny, M.: MiniGPT-v2: Large Language Model as a Unified Interface for Vision-Language Multi-task Learning. arXiv preprint arXiv:2310.09478 (2023)", + "0. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: UNITER: Universal Image-TExt Representation Learning. In: ECCV (2020)", + "1. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An Open-Source Chatbot Improressing GPT-4 with $90\\%$ * ChatGPT Quality (2023)", + "12. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S.S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Castro-Ros, A., Pellat, M., Robinson, K., Valter, D., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E.H., Dean, J., Devlin, J., Roberts, A., Zhou, D., Le, Q.V., Wei, J.: Scaling Instruction-Finetuned Language Models. JMLR 25(70), 1-53 (2024)", + "13. Cornia, M., Baraldi, L., Cucchiara, R.: SMArT: Training Shallow Memory-aware Transformers for Robotic Explainability. In: ICRA (2020)", + "14. Cornia, M., Stefanini, M., Baraldi, L., Cucchiara, R.: Meshed-Memory Transformer for Image Captioning. In: CVPR (2020)" + ], + "bbox": [ + 218, + 316, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning. arXiv preprint arXiv:2305.06500 (2023)", + "16. Dong, H., Li, J., Wu, B., Wang, J., Zhang, Y., Guo, H.: Benchmarking and Improving Detail Image Caption. arXiv preprint arXiv:2405.19092 (2024)", + "17. Herdade, S., Kappeler, A., Boakye, K., Soares, J.: Image captioning: Transforming objects into words. In: NeurIPS (2019)", + "18. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In: EMNLP (2021)", + "19. Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking task: Data, models and evaluation metrics. JAIR 47, 853-899 (2013)", + "20. Huang, L., Wang, W., Chen, J., Wei, X.Y.: Attention on Attention for Image Captioning. In: ICCV (2019)", + "21. Jiang, M., Hu, J., Huang, Q., Zhang, L., Diesner, J., Gao, J.: REO-Relevance, Extraness, Omission: A Fine-grained Evaluation for Image Captioning. In: EMNLP (2019)", + "22. Jiang, M., Huang, Q., Zhang, L., Wang, X., Zhang, P., Gan, Z., Diesner, J., Gao, J.: TIGER: Text-to-Image Grounding for Image Caption Evaluation. In: EMNLP (2019)", + "23. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015)", + "24. Kim, J.H., Kim, Y., Lee, J., Yoo, K.M., Lee, S.W.: Mutual Information Divergence: A Unified Metric for Multimodal Generative Models. In: NeurIPS (2022)", + "25. Laurencon, H., Saulnier, L., Tronchon, L., Bekman, S., Singh, A., Lozhkov, A., Wang, T., Karamcheti, S., Rush, A.M., Kiela, D., Cord, M., Sanh, V.: OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents. In: NeurIPS (2023)", + "26. Lee, H., Yoon, S., Dernoncourt, F., Bui, T., Jung, K.: UMIC: An Unreferenced Metric for Image Captioning via Contrastive Learning. In: ACL (2021)", + "27. Lee, H., Yoon, S., Dernoncourt, F., Kim, D.S., Bui, T., Jung, K.: ViLBERTScore: Evaluating Image Caption Using Vision-and-Language BERT. In: EMNLP Workshops (2020)", + "28. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping Language-Image Pretraining with Frozen Image Encoders and Large Language Models. In: ICML (2023)", + "29. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In: ICML (2022)", + "30. Li, X., Tu, H., Hui, M., Wang, Z., Zhao, B., Xiao, J., Ren, S., Mei, J., Liu, Q., Zheng, H., et al.: What If We Recap tion Billions of Web Images with LLaMA-3? arXiv preprint arXiv:2406.08478 (2024)", + "31. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: ECCV (2020)", + "32. Li, Y., Pan, Y., Yao, T., Mei, T.: Comprehending and ordering semantics for image captioning. In: CVPR (2022)", + "33. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL Workshops (2004)", + "34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: ECCV (2014)", + "35. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved Baselines with Visual Instruction Tuning. In: CVPR (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Sarto et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual Instruction Tuning. In: NeurIPS (2023)", + "37. Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization. In: ICLR (2019)", + "38. Lu, J., Batra, D., Parikh, D., Lee, S.: ViLBERT: Pretraining Task-Agnostic Visi-olinguistic Representations for Vision-and-Language Tasks. In: NeurIPS (2019)", + "39. Oord, A.v.d., Li, Y., Vinyals, O.: Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748 (2018)", + "40. Pan, Y., Yao, T., Li, Y., Mei, T.: X-Linear Attention Networks for Image Captioning. In: CVPR (2020)", + "41. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: BLEU: a method for automatic evaluation of machine translation. In: ACL (2002)", + "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transferable Visual Models From Natural Language Supervision. In: ICML (2021)", + "43. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners. OpenAI Blog 1(8), 9 (2019)", + "44. Ramos, R., Martins, B., Elliott, D., Kementchedjhieva, Y.: SmallCap: lightweight image captioning prompted with retrieval augmentation. In: CVPR (2023)", + "45. Rashtchian, C., Young, P., Hodosh, M., Hockenmaier, J.: Collecting Image Annotations Using Amazon's Mechanical Turk. In: NAACL Workshops (2010)", + "46. Saito, K., Sohn, K., Zhang, X., Li, C.L., Lee, C.Y., Saenko, K., Pfister, T.: Pic2Word: Mapping Pictures to Words for Zero-Shot Composed Image Retrieval. In: CVPR (2023)", + "47. Sarto, S., Barraco, M., Cornia, M., Baraldi, L., Cucchiara, R.: Positive-Augmented Contrastive Learning for Image and Video Captioning Evaluation. In: CVPR (2023)", + "48. Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: Retrieval-Augmented Transformer for Image Captioning. In: CBMI (2022)", + "49. Shekhar, R., Pezzelle, S., Klimovich, Y., Herbelot, A., Nabi, M., Sangineto, E., Bernardi, R.: FOIL it! Find One mismatch between Image and Language caption. In: ACL (2017)", + "50. Shi, Y., Yang, X., Xu, H., Yuan, C., Li, B., Hu, W., Zha, Z.J.: EMScore: Evaluating Video Captioning via Coarse-Grained and Fine-Grained Embedding Matching. In: CVPR (2022)", + "51. Tewel, Y., Shalev, Y., Schwartz, I., Wolf, L.: Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In: CVPR (2022)", + "52. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023)", + "53. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "55. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: CIDEr: Consensus-based Image Description Evaluation. In: CVPR (2015)", + "56. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: CVPR (2015)", + "57. Wang, S., Yao, Z., Wang, R., Wu, Z., Chen, X.: FAIER: Fidelity and Adequacy Ensured Image Caption Evaluation. In: CVPR (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "58. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhutdinov, R., Zemel, R.S., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015)", + "59. Yang, X., Tang, K., Zhang, H., Cai, J.: Auto-Encoding Scene Graphs for Image Captioning. In: CVPR (2019)", + "60. Yi, Y., Deng, H., Hu, J.: Improving Image Captioning Evaluation by Considered Inter Refinences Variance. In: ACL (2020)", + "61. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL 2, 67-78 (2014)", + "62. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: VinVL: Revisiting visual representations in vision-language models. In: CVPR (2021)", + "63. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: BERTScore: Evaluating Text Generation with BERT. In: ICLR (2020)", + "64. Zhu, W., Wang, X.E., Yan, A., Eckstein, M., Wang, W.Y.: ImaginE: An Imagination-Based Automatic Evaluation Metric for Natural Language Generation. In: EACL (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 383 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Sarto et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_model.json b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3ad1c989eec7c5e9610b6fbb9d185bd341c37e23 --- /dev/null +++ b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_model.json @@ -0,0 +1,2799 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "BRIDGE: Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.212, + 0.782, + 0.227 + ], + "angle": 0, + "content": "Sara Sarto1, Marcella Cornia1, Lorenzo Baraldi1, and Rita Cucchiara1,2" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.239, + 0.669, + 0.254 + ], + "angle": 0, + "content": "1 University of Modena and Reggio Emilia, Italy" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.254, + 0.56, + 0.268 + ], + "angle": 0, + "content": "2 IIT-CNR, Italy" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.27, + 0.599, + 0.281 + ], + "angle": 0, + "content": "{name.surname}@unimore.it" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.315, + 0.744, + 0.552 + ], + "angle": 0, + "content": "Abstract. Effectively aligning with human judgment when evaluating machine-generated image captions represents a complex yet intriguing challenge. Existing evaluation metrics like CIDEr or CLIP-Score fall short in this regard as they do not take into account the corresponding image or lack the capability of encoding fine-grained details and penalizing hallucinations. To overcome these issues, in this paper, we propose BRIDGE, a new learnable and reference-free image captioning metric that employs a novel module to map visual features into dense vectors and integrates them into multi-modal pseudo-captions which are built during the evaluation process. This approach results in a multimodal metric that properly incorporates information from the input image without relying on reference captions, bridging the gap between human judgment and machine-generated image captions. Experiments spanning several datasets demonstrate that our proposal achieves state-of-the-art results compared to existing reference-free evaluation scores. Our source code and trained models are publicly available at: https://github.com/aimagelab/bridge-score" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.564, + 0.658, + 0.579 + ], + "angle": 0, + "content": "Keywords: Captioning Evaluation \\(\\cdot\\) Vision-and-Language" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.377, + 0.618 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.81 + ], + "angle": 0, + "content": "The objective of image captioning is to produce natural language descriptions conditioned on input images, that closely resemble human language and align to human intentions. As such, the captioning task involves the recognition and understanding of the visual content of the image, including fine-grained elements such as objects, attributes, and their relationships. Advances in training methodologies and architectures have contributed to the progress in the field, significantly improving the generation quality. Recent innovations include fully-attentive models [13, 14, 17], improved connections between visual and textual modalities [14, 40], and the incorporation of objects and tags at an architectural level [3, 31, 62]. Additionally, there has been a notable focus on increasing the robustness of cross-modal features [5, 32, 48], which consequently can increase description accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "As constant improvements are made on the generation side, it becomes crucial to enhance the evaluation process as well. In this regard, image captioning" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Sarto et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.147, + 0.784, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.226, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Fig. 1: Comparison between different captioning evaluation approaches: (a) CIDEr [55] scores candidate and reference captions without considering the input image; (b) CLIP-Score [18] compares text and images using global vectors in a shared embedding space; (c) our BRIDGE, internally builds multimodal pseudo-captions by translating fine-grained image features into pseudo-tokens thanks to a mapping module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.785, + 0.462 + ], + "angle": 0, + "content": "evaluation aims to assess the quality of a generated caption given an image and, potentially, human-written reference captions. However, it is important to note that obtaining these reference captions can often be challenging and expensive, adding complexity to the evaluation process. Despite the recent advancements in captioning capabilities, standard automatic evaluation metrics have mainly relied on translation metrics [4,33,41] or text-only ones [2,55,63] which often fall short in capturing aspects such as grammatical correctness, semantic relevance, and specificity. These limitations are worsened by the limited coverage of image content in available references, resulting in inaccurate penalties when generated captions accurately describe novel elements not mentioned in the references." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.553 + ], + "angle": 0, + "content": "In response to these limitations, advanced metrics aligning visual and textual data have emerged [18,24,26,27]. Notably, recent metrics leverage the CLIP embedding space [42], which shows a strong correlation with human judgment. Despite the effectiveness of contrastive-based embedding spaces, metrics based on dual-encoder architectures tend to focus on global alignment between an image and its caption, often ignoring fine-grained details or penalizing hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.786, + 0.719 + ], + "angle": 0, + "content": "Following this insight, in this paper, we introduce BRIDGE, a novel learnable and reference-free image captioning metric that enhances the alignment of more fine-grained visual features. Specifically, our model provides a pre-trained dual-encoder architecture with a mapping module designed to effectively exploit visual cues (Fig. 1). This is done by internally creating multimodal pseudo-captions, containing both textual and dense visual features. The process for building these pseudo-captions involves the creation of a template caption, which focuses on the syntactical structure of the scene, and a mapping module. The latter refines the template caption by enriching it with more fine-grained visual features about the subjects depicted in the image. Subsequently, the overall model is trained with a combination of contrastive losses which promote multimodal alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Experiments are conducted on a variety of datasets for image captioning annotated with human rankings, including Flickr8k-Expert, Flickr8k-CF [19], Composite [1], and Pascal-50S [55]. Through comprehensive analysis, we demonstrate the efficacy of the proposed metric and show its ability to overcome the limitations of existing state-of-the-art reference-free alternatives. Additionally, we evaluate its sensitivity to object hallucination by conducting experiments on the FOIL dataset [49]. Overall, BRIDGE outperforms previous metrics and showcases superior performance compared to CLIP-Score [18] and PAC-Score [47]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.667, + 0.161 + ], + "angle": 0, + "content": "Contributions. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.163, + 0.784, + 0.235 + ], + "angle": 0, + "content": "- We tackle the limitations of existing image captioning metrics by proposing the first learnable reference-free metric, termed as BRIDGE, that focuses on more fine-grained visual features. Our proposal integrates a dual-encoder architecture with a mapping module which is in charge of producing multimodal pseudo-captions that combine text and richer dense visual features." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.238, + 0.784, + 0.281 + ], + "angle": 0, + "content": "- Multimodal pseudo-captions are built by creating template captions that describe the scene from a syntactical point of view. These templates are subsequently enriched with fine-grained features through a mapping network." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.284, + 0.784, + 0.325 + ], + "angle": 0, + "content": "- Experiments, carried out on different datasets with human preferences, demonstrate a higher degree of correlation with respect to existing metrics. We also evaluate the sensitivity of the proposal to objects hallucinations." + }, + { + "type": "list", + "bbox": [ + 0.224, + 0.163, + 0.784, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.353, + 0.388, + 0.369 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.563 + ], + "angle": 0, + "content": "Classical Reference-based Captioning Metrics. Several widely used captioning evaluation metrics were originally developed in the context of NLP tasks and rely on n-gram matching techniques. Among these classical metrics, BLEU [41] is designed to focus on precision and incorporates a penalty for sentence brevity. METEOR [4], instead, combines precision and recall to evaluate the quality of captions, while others, such as ROUGE [33], were initially born for summarization tasks and later adapted to image captioning. More recently, two metrics tailored for visual captioning have emerged: CIDEr [55], which measures n-gram similarity and is based on TF-IDF, and SPICE [2], which quantifies graph-based similarity through scene graphs constructed from candidate and reference captions. Overall, focusing on textual-level comparisons, these metrics assume that human-written references accurately represent the image content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.643 + ], + "angle": 0, + "content": "Learnable Captioning Metrics. With the rise of large pre-trained models, image captioning evaluation now frequently exploits these models to compare textual-only [60, 63] or visual-textual [18, 21, 22, 26, 27, 57] contents. Notably, the BERT score and its improved version use pre-trained BERT embeddings to compare word tokens in generated and ground-truth sentences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.748 + ], + "angle": 0, + "content": "Some metrics, like BLEU and CIDEr, rely solely on text matching between reference captions and machine-generated captions, potentially introducing bias in evaluations due to non-accurate reference captions. To mitigate these issues, alternative solutions leverage the multimodal nature of vision-and-language models. As an example, TIGEr [22] considers the similarities between words and image regions and assesses how well machine-generated captions represent image content and their alignment with human-generated captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In contrast, other approaches [18, 24, 27] leverage web-scale vision-and-language models such as VilBERT [38] and CLIP [42] for more robust metrics. For example, in [24], CLIP visual-textual features are used to compute negative Gaussian cross-mutual information. Other works, instead, have employed diffusion models in text-only tasks [64] or exploited the zero-shot language modeling capabilities of large language models [8] to evaluate candidate captions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Sarto et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Reference-free Captioning Metrics. While all aforementioned metrics rely on a set of ground-truth captions to compute the final score, a few attempts have been made to introduce reference-free evaluations, only taking into account the correlation of the candidate caption with the image. In this regard, Lee et al. [26] proposed to fine-tune the UNITER model [10] via contrastive learning to let the model to discriminate between positive and synthetically-generated negative captions. On a different line, Hessel et al. [18] introduced the CLIP-Score, which only relies on a modified cosine similarity between image and candidate caption representations coming from the CLIP model. The recently proposed PAC-Score [47], instead, builds upon the usage of CLIP but incorporates a fine-tuning phase with positive augmentation, further enhancing the accuracy of evaluation. In this paper, we follow this research path and propose a novel learnable, and reference-free evaluation metric that can effectively incorporate fine-grained visual features for evaluating the correlation of image-text pairs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.384, + 0.6, + 0.403 + ], + "angle": 0, + "content": "3 BRIDGE for Captioning Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.507 + ], + "angle": 0, + "content": "In the following, we present our reference-free captioning evaluation approach, termed BRIDGE. Our approach leverages a dual-encoder architecture, which comprises both a language encoder and a vision encoder. Given a frozen pretrained model, we train a mapping module responsible for filling the holes of a masked template caption with pseudo language tokens that are enriched with visual information. An overview of our model is depicted in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.373, + 0.545 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Our approach relies on CLIP (Contrastive Language-Image Pre-training) [42], a powerful vision and language model designed to align images and corresponding text captions within a shared embedding space. For a given input image \\( I \\), the image encoder \\( E_V \\) extracts the visual information \\( v = E_V(I) \\in \\mathbb{R}^d \\). On the textual side, an input caption \\( T \\) is tokenized and a textual representation is obtained by passing it through the textual encoder \\( E_T \\), obtaining \\( t = E_T(T) \\in \\mathbb{R}^d \\). Once the textual and visual features, \\( t \\) and \\( v \\) respectively, are projected in a common space, visual and textual inputs can be compared via cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.785 + ], + "angle": 0, + "content": "The relationships learned by CLIP can be exploited to build an image captioning evaluator. In CLIP-Score [18] the authors directly compare candidate captions and images in the embedding space and show that this achieves a good correlation with human judgments. In detail, to assess the quality of a candidate generation, they feed both the image and the candidate caption through their respective feature extractors, and compute the cosine similarity of the resultant embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.786, + 0.786, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C L I P} \\text {- S c o r e} (I, T) = w \\cdot \\max (\\cos (v, t), 0), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\( w \\) is a rescaling factor employed to stretch the score distribution while ensuring the ranking results remain unchanged." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.784, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.353, + 0.785, + 0.396 + ], + "angle": 0, + "content": "Fig. 2: Overview of the BRIDGE evaluation approach. Starting from a template caption, a mapping network augments it with dense visual features, obtaining a pseudocaption which is then used for computing image-text similarities." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.41, + 0.585, + 0.426 + ], + "angle": 0, + "content": "3.2 Injecting Fine-Grained Visual Features" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.508 + ], + "angle": 0, + "content": "Unlike CLIP-Score, our approach does not rely exclusively on global image descriptors for evaluating image-text alignments. Instead, we focus on employing stronger visual information. To do so, we draw inspiration from the Pic2Word approach [46] and represent the input image through a multimodal pseudo-caption, an embedding representation that contains stronger visual elements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.787, + 0.708 + ], + "angle": 0, + "content": "Building Template Captions. In order to create multimodal pseudo-captions, we first build template captions for a given input image. These are skeletal textual representations of the image, obtained by masking out all the relevant textual concepts from the descriptions generated by a captioner. Through these template captions, we aim to provide the model only with a templated textual structure which can then be filled with more fine-grained visual features. In particular, given an automatically generated caption describing the input image, such as 'A man running with a white dog', we remove the main subjects within the caption (e.g. 'man' and 'white dog') and mask them with a [MASK] token. This will allow the model to fill in these gaps by incorporating more fine-grained features from the image encoder. Since a primary subject might be described by words other than just its corresponding nouns (e.g. adjectives), we utilize noun chunks. Fig. 3 reports template captions and corresponding noun chunks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.708, + 0.787, + 0.784 + ], + "angle": 0, + "content": "Given a sentence containing \\( N \\) noun chunks, we independently encode them through the mapping network. To this aim, we replicate the template caption as many times as the number of noun chunks and mask a different noun chunk in each of the replicas. We thus obtain \\( N \\) different versions of the template caption, each one masking a single noun chunk, for instance" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.793, + 0.66, + 0.808 + ], + "angle": 0, + "content": "[A [MASK] running with a white dog'," + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.812, + 0.633, + 0.827 + ], + "angle": 0, + "content": "'A man running with a [MASK]'." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Sarto et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.318, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.149, + 0.357, + 0.161 + ], + "angle": 0, + "content": "COCO" + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.173, + 0.363, + 0.183 + ], + "angle": 0, + "content": "TEMPLATE" + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.191, + 0.361, + 0.206 + ], + "angle": 0, + "content": "NOUN CHUNKS" + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.149, + 0.49, + 0.161 + ], + "angle": 0, + "content": "People on a motorcycle in front of a road" + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.169, + 0.495, + 0.187 + ], + "angle": 0, + "content": "A yellow bus driving down a street next to a park." + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.191, + 0.421, + 0.206 + ], + "angle": 0, + "content": "yellow bus" + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.191, + 0.448, + 0.205 + ], + "angle": 0, + "content": "street" + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.192, + 0.488, + 0.205 + ], + "angle": 0, + "content": "park" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.145, + 0.601, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.151, + 0.638, + 0.162 + ], + "angle": 0, + "content": "COCO" + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.173, + 0.645, + 0.184 + ], + "angle": 0, + "content": "TEMPLATE" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.19, + 0.641, + 0.206 + ], + "angle": 0, + "content": "NOUN CHUNKS" + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.149, + 0.776, + 0.167 + ], + "angle": 0, + "content": "Little dog laying on a pillow next to the laptop." + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.17, + 0.778, + 0.187 + ], + "angle": 0, + "content": "A dog laying on a couch next to a PC." + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.191, + 0.642, + 0.206 + ], + "angle": 0, + "content": "NOUN CHUNKS" + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.193, + 0.68, + 0.204 + ], + "angle": 0, + "content": "PC" + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.193, + 0.725, + 0.204 + ], + "angle": 0, + "content": "dog" + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.193, + 0.775, + 0.204 + ], + "angle": 0, + "content": "couch" + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.215, + 0.756, + 0.23 + ], + "angle": 0, + "content": "Fig. 3: COCO captions with template captions and associated noun chunks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.245, + 0.787, + 0.352 + ], + "angle": 0, + "content": "Mapping with Fine-Grained Visual Features. The above-described masked replicas are then fed to the mapping module \\(\\psi\\). Specifically, our approach exploits the visual information extracted from the visual encoder \\(E_V\\) to enrich the replicas with the informative content of the image \\(I\\). To get a more fine-grained representation of the image, we directly take the grid of features from the last layer, \\(\\widehat{v}\\). For instance, in the case of a ViT-B/32 backbone, this will have a shape of \\(50 \\times d\\), where \\(d\\) is the dimensionality of the last embedding of the network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.787, + 0.457 + ], + "angle": 0, + "content": "The mapping network is implemented as a stack of Transformer [54] encoder layers interleaved with cross-attention layers. Its role is to refine each template captions with visual information. Since each template caption is processed independently, the mapping module returns a set of sequences, each with the same length as the corresponding input template caption. From the output of the mapping module, we keep only the predictions for the masked tokens in each template caption and copy them back into the original templates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.789, + 0.665 + ], + "angle": 0, + "content": "Therefore, by providing a masked input template in the form \\(\\hat{T_i} = [w_1,\\dots,w_{j - 1},\\mathsf{MASK},w_{j + 1},\\dots,w_T]\\), where \\(\\{w_j\\}_j\\) represent original tokens from the input caption, we obtain \\(\\hat{T_i}^* = \\left[w_1,\\dots,w_{j - 1},\\psi (\\hat{T_i})_j,w_{j + 1},\\dots,w_T\\right]\\), where \\(\\psi (\\hat{T_i},\\widehat{v})_j\\) represents the output of the mapping network at the position corresponding to the masked input token position. In the case of noun chunks consisting of more than one token, multiple consecutive tokens are replaced with the corresponding outputs from the mapping network. By injecting the outputs of the mapping token into the initial template caption, we effectively complete the original templates with visually enriched vectors. Notably, these newly generated pseudo-captions combine word sequences from the template captions with dense vectors obtained by the mapping module. Consequently, they cannot be decoded as standard captions. As a last step, the obtained pseudo-captions are fed into the pre-trained CLIP language encoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.691, + 0.41, + 0.706 + ], + "angle": 0, + "content": "3.3 Training Protocol" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.763 + ], + "angle": 0, + "content": "To train our mapping network, the loss is defined as a weighted version of the symmetric InfoNCE loss [39], where positive and negative items are weighted according to the number of noun chunks in each caption." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Specifically, given a batch in the form \\(\\mathcal{B} = \\{(I_i,T_i)\\}_{i = 1}^N\\), where \\(I_{i}\\) and \\(T_{i}\\) represent image-caption pairs, each image \\(I_{i}\\) is expanded in \\(N_{i}\\) multimodal pseudocaptions as outlined above, where \\(N_{i}\\) is the number of noun chunks in caption \\(T_{i}\\). Further, let \\(t_{ij}^{*}\\) represent the embedding vector of the \\(j\\)-th pseudo-caption derived from the \\(i\\)-th image, \\(v_{i}\\) the embedding vector of the \\(i\\)-th image and \\(t_{i}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "the embedding vector of the \\(i\\)-th ground-truth caption. Finally, let \\(M\\) be the total number of noun chunks in the mini-batch, i.e. \\(M = \\sum_{i=1}^{N} N_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.785, + 0.286 + ], + "angle": 0, + "content": "The first loss, denoted as \\(\\mathcal{L}_1\\), tries to align the pseudo-captions \\(t_{ij}^*\\) with the global visual features of the corresponding images \\(v_i\\). In addition to this loss term, we define a second loss component \\(\\mathcal{L}_2\\) that promotes the alignment between pseudo-captions \\(t_{ij}^*\\) and the textual feature vector of the ground-truth caption \\(t_i\\) corresponding to the input image. This ensures that pseduo-captions are aligned also on a textual space, in addition to being aligned in the image space. Additional details can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.784, + 0.382 + ], + "angle": 0, + "content": "Regularization Branch. With the aforementioned loss terms, our objective is encouraging an association between each pseudo-caption and its corresponding image and caption. However, it is also important to differentiate each pseudocaption from the others of the same image. To achieve this, we define a regularization loss which promotes a precise alignment between each pseudo-caption and the corresponding noun chunk." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.784, + 0.445 + ], + "angle": 0, + "content": "First, we create prompts like \"a photo of a \" and encode them with the text encoder \\( E_{T} \\). In parallel, each pseudo-caption is fed into a dedicated multilayer perceptron (MLP) projection, which consists of two linear layers with a ReLU activation in between. Formally, the branch is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.461, + 0.784, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} (x) = \\operatorname {L i n e a r} (\\operatorname {R e L U} (\\operatorname {L i n e a r} (x))). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.784, + 0.536 + ], + "angle": 0, + "content": "Since our goal is to emphasize each pseudo-caption's alignment with its corresponding noun chunk, we employ a regular contrastive loss \\(\\mathcal{L}_3\\) between the prompts mentioned earlier and the outputs of the projection branch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.784, + 0.568 + ], + "angle": 0, + "content": "Finally, the overall loss function we train BRIDGE is defined as a weighted summation of two aforementioned losses, plus the regularization loss, as" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.584, + 0.784, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {2} \\mathcal {L} _ {2} + \\lambda_ {3} \\mathcal {L} _ {r}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.627, + 0.54, + 0.642 + ], + "angle": 0, + "content": "3.4 Inference and Score Computation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.785, + 0.733 + ], + "angle": 0, + "content": "At inference time, given an image-candidate caption pair \\((I,T)\\), we extract all pseudo-captions from \\(I\\) using our mapping network. Subsequently, we compute the mean pseudo-caption embedding as \\(t^{*} = \\frac{1}{N}\\hat{t}_{i}^{*}\\), where \\(\\hat{t}_i^*\\) indicates the \\(i\\)-th pseudo-caption extracted from \\(I\\) and \\(N\\) here indicates the overall number of pseudo-captions associated with \\(I\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.765 + ], + "angle": 0, + "content": "At that point, given the visual embedding \\( v \\) of the image and the embedding of the candidate caption \\( t \\), the matching score between \\( I \\) and \\( T \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.796, + 0.784, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\operatorname {B R I D G E} (I, T) = 0. 5 \\cdot [ \\operatorname {C L I P - S c o r e} (I, T) + w \\cdot \\max (\\cos \\left(t ^ {*}, t\\right), 0) ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.753, + 0.84 + ], + "angle": 0, + "content": "where \\( \\cos \\) indicates the cosine similarity and \\( w \\) is a constant scaling factor." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Sarto et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.492, + 0.163 + ], + "angle": 0, + "content": "4 Experimental Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.173, + 0.457, + 0.188 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.787, + 0.333 + ], + "angle": 0, + "content": "Architecture and Training Details. Building upon prior research [18,24,50], we use either CLIP [42] ViT-B/32 or ViT-L/14 as backbone for the visual and textual encoder. The mapping module is composed of two Transformer layers and is trained on the COCO dataset [34], which contains more than 120k images annotated with five captions. In particular, we employ the splits introduced by Karpathy et al. [23], where 5,000 images are used for both validation and testing and the rest for training. To map the grid visual features to an embedding space of dimension 512, we employ a simple linear projection. For the regularization branch, we utilize a two-layer multi-layer perceptron." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.787, + 0.409 + ], + "angle": 0, + "content": "During training, we use AdamW [37] as optimizer with a learning rate equal to 0.0001 and a batch size of 256. The \\(\\lambda_1\\), \\(\\lambda_2\\), and \\(\\lambda_3\\) values are selected with a grid search, choosing the combination that provides the best validation loss. Specifically, we set both \\(\\lambda_1\\) and \\(\\lambda_3\\) to 0.01, while \\(\\lambda_2\\) is set to 1.0. The training stage lasts around one day on a single A100 GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.787, + 0.518 + ], + "angle": 0, + "content": "Template Caption Generation. The template captions used as input for the mapping module are generated using the BLIP model [29]. In particular, we use the ViT-L/14 version pre-trained on 129M image-text pairs and finetuned on the COCO dataset. After this generation phase, the primary subjects of the template sentences are extracted by using the NLTK library [6]. During training, two noun chunks are randomly chosen from the set identified during the extraction step. In the evaluation phase, otherwise, all identified noun chunks are included." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.541, + 0.334, + 0.554 + ], + "angle": 0, + "content": "4.2 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.787, + 0.656 + ], + "angle": 0, + "content": "To evaluate the correlation of the proposed metric with human ratings, we conduct experiments on the Flickr8k-Expert, Flickr8k-CF, Composite, and Pascal50-S datasets [1,19,55]. In addition, for detecting hallucinations in textual sentences, we extend our analysis to the FOIL dataset [49]. Except for Pascal50S and FOIL in where accuracy scores are used, evaluation on all other datasets relies on Kendall \\(\\tau_{b}\\), Kendall \\(\\tau_{c}\\), and Spearman \\(\\rho\\) correlation scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Flickr8k-Expert and Flickr8k-CF [19]. These datasets consist of image-caption pairs with corresponding human ratings. Specifically, Flickr8k-Expert comprises 17k expert annotations for visual-textual pairs, with a total of 5,664 different images. Each pair receives a score ranging from 1 (lack of correlation) to 4 (accurate depiction), where 1 indicates a lack of correlation between the caption and the image, and 4 indicates an accurate depiction of the image without errors. On the other hand, Flickr8k-CF is composed of 145k binary quality judgments, collected from CrowdFlower, for 48k image-caption pairs containing 1,000 unique images. Each pair is annotated with at least three binary scores, where “yes” denotes that the caption correlates with the image. To measure the alignment with human judgment, we compute the mean proportion of “yes” annotations as the score for each pair." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.246, + 0.145, + 0.755, + 0.159 + ], + "angle": 0, + "content": "Table 1: Ablation study results. US indicates the number of pseudo tokens." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.168, + 0.783, + 0.359 + ], + "angle": 0, + "content": "
Flickr8k-ExpertFlickr8k-CFPascal-50S
Kend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρAccuracy
Architectural Components
w/o mapping module (i.e. MLP)53.153.565.335.518.343.581.9
w/o template captions53.754.166.035.518.443.682.5
w/o regularization branch54.154.566.535.718.543.682.7
Score Formulation
w/o textual similarity51.151.263.034.417.730.580.9
w/o visual similarity53.854.266.035.418.343.781.9
Pseudo-token Size
w/ US = 154.154.566.435.117.130.382.6
w/ US = 254.154.566.436.118.744.482.8
w/ US = 454.354.766.635.418.343.882.4
w/ US = 854.054.466.335.918.644.581.7
BRIDGE (US = 3)54.454.867.736.118.744.582.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.785, + 0.45 + ], + "angle": 0, + "content": "Composite [1]. It contains 12k human ratings for image-caption pairs including a combination of images taken from COCO [34] (2,007 images), Flickr8k [19] (997 images), and Flickr30k [61] (991 images). In this dataset, human evaluators assess each image-caption pair, assigning a score within the range of 1 to 5 to estimate the correspondence of the caption with the associated image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.785, + 0.59 + ], + "angle": 0, + "content": "Pascal50-S [55]. It presents pairwise preference judgments between two captions. Overall, the dataset consists of 4,000 sentence pairs, each of them associated with an image from the UIUC Pascal sentence dataset [45]. Each pair is associated with 48 human judgments, where each evaluation indicates which sentence better describes the given image. The sentence pairs are categorized into four groups: (i) both human-written and correct captions (HC), (ii) both human-written captions where one is correct and the other is wrong (HI), (iii) both correct captions but one written by humans and the other machine-generated (HM), (iv) both machine-generated and correct captions (MM)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.785, + 0.684 + ], + "angle": 0, + "content": "FOIL [49]. The dataset comprises image-caption pairs from the COCO dataset [34]. In this scenario, captions undergo perturbation by generating modified versions that closely resemble the originals but introduce a single error, referred to as \"foil word\". For a fair comparison, we select the subset of the validation set that does not overlap with the portion of COCO used during training, resulting in 8,000 images, each paired with a foil-correct textual counterpart." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.706, + 0.512, + 0.722 + ], + "angle": 0, + "content": "4.3 Ablation Studies and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.785, + 0.793 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our metric, we start by analyzing variations of our main architectural components. Then, we assess the impact of caption templates in our score formulation. All these experiments are performed using CLIP ViT-B/32 as backbone and reported in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Contribution of Architectural Components. We first investigate the performance of the most straightforward implementation of a mapping module, structured as a two-layer MLP following [46]. We also validate the importance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Sarto et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.148, + 0.533, + 0.163 + ], + "angle": 0, + "content": "Table 2: Impact of different template captions." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.172, + 0.53, + 0.33 + ], + "angle": 0, + "content": "
ExpertCFPascal-50S
Kend. τbKend. τbAcc.
Transformer Templates
w/o mapping module46.131.680.4
BRIDGE54.035.982.7
(+7.9)(+4.3)(+2.3)
BLIP Templates
w/o mapping module48.633.881.0
BRIDGE54.436.182.6
(+5.8)(+2.3)(+1.6)
BLIP-2 Templates
w/o mapping module49.434.282.4
BRIDGE54.436.282.9
(+5.0)(+2.0)(+0.5)
" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.146, + 0.671, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.147, + 0.771, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.203, + 0.565, + 0.232 + ], + "angle": 0, + "content": "C C C C C C" + }, + { + "type": "text", + "bbox": [ + 0.585, + 0.204, + 0.666, + 0.233 + ], + "angle": 0, + "content": "A young girl is preparing to blow out her candle" + }, + { + "type": "text", + "bbox": [ + 0.685, + 0.204, + 0.765, + 0.232 + ], + "angle": 0, + "content": "A painting of a table with fruit on top of it" + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.236, + 0.566, + 0.265 + ], + "angle": 0, + "content": "TRANSE" + }, + { + "type": "text", + "bbox": [ + 0.586, + 0.236, + 0.661, + 0.265 + ], + "angle": 0, + "content": "A young girl is eating a piece of cake" + }, + { + "type": "text", + "bbox": [ + 0.687, + 0.236, + 0.761, + 0.265 + ], + "angle": 0, + "content": "A painting of a painting of fruits and a painting" + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.273, + 0.566, + 0.292 + ], + "angle": 0, + "content": "#" + }, + { + "type": "text", + "bbox": [ + 0.584, + 0.27, + 0.665, + 0.298 + ], + "angle": 0, + "content": "A woman blowing out a candle on a cake" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.27, + 0.767, + 0.298 + ], + "angle": 0, + "content": "A painting of a pitcher and oranges on a table" + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.303, + 0.566, + 0.329 + ], + "angle": 0, + "content": "B" + }, + { + "type": "text", + "bbox": [ + 0.583, + 0.303, + 0.664, + 0.332 + ], + "angle": 0, + "content": "A girl blowing out a candle on a cupcake" + }, + { + "type": "text", + "bbox": [ + 0.685, + 0.303, + 0.767, + 0.332 + ], + "angle": 0, + "content": "A painting of a still life with oranges and a candle" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.457 + ], + "angle": 0, + "content": "of the template captions through a model variant in which a set of learnable tokens \\(\\mathsf{S}^*\\) serves as input for the mapping module, without relying on template captions. In both variants, given the absence of template captions, we construct a template such as 'a photo of \\(\\mathsf{S}^*\\)' and extract its features using the CLIP text encoder. In the Table, it can be seen that, regardless of any architectural changes, it is important to provide a simple sentence structure to the mapping module to achieve competitive performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.578 + ], + "angle": 0, + "content": "In addition to these baselines, we devise a variant to analyze the contribution of the regularization branch. In this setting, we employ template captions as input for the mapping module, resulting in a substantial improvement of \\(+1.0\\) Kendall \\(\\tau_{b}\\) and \\(+0.8\\) accuracy points compared to the MLP variant, respectively on the Flickr8k-Expert and on the Pascal-50S dataset. When introducing the regularization branch (i.e. the complete BRIDGE architecture), further enhancements can be observed especially on the Flickr8k-CF with an improvement of \\(+0.4\\) points in terms of the Kendall \\(\\tau_{b}\\) correlation score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.655 + ], + "angle": 0, + "content": "We also emphasize the importance of each component in our score formulation. Specifically, we present correlation results when employing only the visual similarity within our architecture, which is the original CLIP-Score formulation. As observed, performance drops drastically when relying only on visual information. A less significant drop is observed when employing only textual similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.732 + ], + "angle": 0, + "content": "As an additional analysis, we report the effect of changing the number of the pseudo tokens for each noun chunk, denoting it as unit size (US). Specifically, we compute the scores employing \\(\\mathrm{US} = 1,2,3,4,8\\). From the results, it can be seen that \\(\\mathrm{US} = 3\\) generally leads to the best performance across nearly all evaluation metrics. This configuration is used in all experiments reported in the paper." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Analysis on Caption Templates. We analyze the effect of changing the initial template captions for our model. Specifically, we employ template captions generated by a conventional Transformer-based captioner that uses CLIP features as input and is trained only on the COCO dataset, as well as those generated by BLIP [29], and BLIP-2 [28]. Notably, we select template captions of different quality based on both standard metric evaluations and correlations with human judgment. To assess the quality of the raw generated templates, we observe that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "the CIDEr score of these models on the COCO test set is equal to 114.2, 131.4, and 145.8 respectively for the standard Transformer model, BLIP, and BLIP-2. Note that all models were trained with cross-entropy loss only. Results on the Flickr8k-Expert, Flickr8k-CF, and Pascal-50S datasets are reported in Table 2. To qualitatively validate the generated templates, we include sample captions generated by the three models compared to a ground-truth caption from the COCO test set. Specifically, captions generated by BLIP-2 are generally more detailed and effectively describe the visual content of the input image compared to those generated by BLIP and, notably, the standard Transformer model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.389 + ], + "angle": 0, + "content": "For each caption template source, we compute the correlation scores when the mapping module is disabled and the captions are directly fed to the text encoder. We compare it with our standard BRIDGE score, considering the different caption templates. Across all datasets, it is evident that directly using the caption templates as input to the text encoder leads to poor performance. This highlights the intended flexibility of template captions as skeletal representations, allowing the model to enhance them with fine-grained visual features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.541 + ], + "angle": 0, + "content": "In fact, starting from these simple template captions and following our approach, we achieve improvements of \\(+5.8\\) and \\(+2.3\\) Kendall \\(\\tau_{b}\\) points and \\(+1.6\\) accuracy points, respectively, on the Flickr8k-Expert, Flickr8k-CF, and Pascal50-S datasets when using BLIP caption templates. The overall best results are with captions from the BLIP-2 model, confirming that better templates can indeed lead to improved results. However, even when using lower-quality captions, the final correlation results are very close to those obtained with higher-quality captions. This highlights the robustness of our metric to caption templates of varying quality and that it is not necessary to rely on captions generated by large-scale captioners to achieve strong correlation scores." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.707, + 0.577 + ], + "angle": 0, + "content": "4.4 Comparison with State-of-the-Art Captioning Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Evaluating Sample-Level Human Correlation. We evaluate the sample-level human correlation on the Flickr8k [19] and Composite [1] datasets. Following previous works [18,63], we compute Kendall correlation scores in both \\(\\tau_{b}\\) and \\(\\tau_{c}\\) versions and also include the Spearman \\(\\rho\\) score. Results are reported in Table 3, where we compare our proposed BRIDGE metric against other reference-free evaluation scores like UMIC [26], CLIP-S [18], and PAC-S [47]. Moreover, we also compare with standard captioning evaluation metrics (i.e. BLEU [41], ME-TEOR [4], CIDEr [55], and SPICE [2]) and more recent reference-based solutions that exploit text-only or cross-modal learned embeddings, such as BERT-S [63], BERT-S++ [60], TIGEr [22], VilBERTScore [27], and MID [24]. For completeness, we also include the reference-based versions of CLIP-S and PAC-S, termed RefCLIP-S and RefPAC-S, which however are not directly comparable with our solution as both rely on a set of five reference captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As it can be seen, BRIDGE outperforms other reference-free metrics in terms of correlation with human judgment, achieving the highest scores on almost all datasets. Specifically, compared to CLIP-S, BRIDGE shows improvements in terms of Kendall \\(\\tau_{c}\\) of \\(+3.6\\) and \\(+2.8\\) points when using ViT-B/32 and ViT-L/14" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Sarto et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.144, + 0.782, + 0.16 + ], + "angle": 0, + "content": "Table 3: Correlation scores on Flickr8k-Expert, Flickr8k-CF, and Composite [1, 19]." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.167, + 0.785, + 0.435 + ], + "angle": 0, + "content": "
Flickr8k-ExpertFlickr8k-CFComposite
Kend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρ
Reference-based metrics
BLEU-4 [41]30.630.838.716.98.721.028.330.638.1
METEOR [4]41.541.851.922.211.527.636.038.948.1
CIDEr [55]43.643.954.324.612.730.534.937.747.0
SPICE [2]51.744.955.124.412.031.338.840.349.1
BERT-S [63]-39.250.322.8--39.930.148.6
BERT-S++ [60]48.146.756.9---42.344.952.1
TIGER [22]51.449.348.6---47.545.455.3
ViLBERTScore [27]54.250.161.3---51.452.458.7
MID [24]-54.9-37.3-----
RefCLIP-S [18]52.653.064.636.418.844.751.255.466.2
RefPAC-S [47]55.555.967.637.619.546.253.057.368.0
Reference-free metrics
UMIC [26]-46.8-------
CLIP-S (ViT-B/32) [18]51.151.263.034.417.730.549.853.864.3
PAC-S (ViT-B/32) [47]53.954.366.136.018.644.451.555.766.3
BRIDGE (ViT-B/32)54.454.866.736.118.744.550.955.065.4
CLIP-S (ViT-L/14) [18]52.653.064.735.218.243.351.355.465.9
PAC-S (ViT-L/14) [47]55.155.567.336.819.045.352.356.567.1
BRIDGE (ViT-L/14)55.455.867.736.319.044.752.957.267.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.531 + ], + "angle": 0, + "content": "backbone on the Flickr8k-Expert dataset. These improvements extend consistently across all datasets and correlation scores. Compared to PAC-S, BRIDGE still achieves superior results on both Flickr8k-Expert and Flickr8k-CF, while performing on par on the Composite dataset (i.e. PAC-S achieves the best results when using ViT-B/32, while BRIDGE outperforms PAC-S using ViT-L/14)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Discriminating Correct Captions. We also evaluate the effectiveness of our metric on the PASCAL-50S dataset [55]. In this context, instead of" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.445, + 0.842 + ], + "angle": 0, + "content": "calculating correlation scores, we compute accuracy by determining, for each pair, the caption favored by the majority of human ratings as correct (with ties being resolved randomly). We then measure how frequently the evaluation metric assigns a higher score to the chosen caption. Following previous works [18], we randomly select five reference captions from the set of 48 provided by the dataset and average the results over five distinct draws. Accuracy values are reported in Table 4. The results show that when using" + }, + { + "type": "table_caption", + "bbox": [ + 0.452, + 0.58, + 0.788, + 0.623 + ], + "angle": 0, + "content": "Table 4: Accuracy results on Pascal-50S [55] averaged over five random draws of reference captions. The \\(\\dagger\\) marker indicates scores from previous works." + }, + { + "type": "table", + "bbox": [ + 0.467, + 0.627, + 0.77, + 0.832 + ], + "angle": 0, + "content": "
HCHIHMMMMean
Reference-based metrics
BLEU-4 [41]60.393.185.757.074.0
METEOR [4]66.097.794.066.681.1
CIDEr [55]66.597.990.765.280.1
BERT-S++†[60]65.498.196.460.380.1
TIGER†[22]56.099.892.874.280.7
MID†[24]67.099.797.476.885.2
RefCLIP-S [18]64.999.595.573.383.3
RefPAC-S [48]67.799.696.075.684.7
Reference-free metrics
CLIP-S (ViT-B/32) [18]55.999.396.572.080.9
PAC-S (ViT-B/32) [47]60.699.396.972.982.4
BRIDGE (ViT-B/32)59.499.497.574.082.6
CLIP-S (ViT-L/14) [18]57.099.696.773.581.7
PAC-S (ViT-L/14) [47]59.599.495.874.782.2
BRIDGE (ViT-L/14)61.299.696.674.182.9
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.154, + 0.385, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.328, + 0.561, + 0.371 + ], + "angle": 0, + "content": "Fig. 4: Sample images from the FOIL dataset [49] and corresponding scores generated by our proposed metric compared with CLIP-S and PAC-S." + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.154, + 0.558, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.566, + 0.158, + 0.763, + 0.186 + ], + "angle": 0, + "content": "Table 5: Accuracy results on the FOIL [49] dataset." + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.19, + 0.753, + 0.374 + ], + "angle": 0, + "content": "
Acc.
Reference-based metrics
BLEU-4 [41]66.2
METEOR [4]70.1
CIDEr [55]85.7
MID [24]90.5
RefCLIP-S [18]91.0
RefPAC-S [18]93.7
Reference-free metrics
CLIP-S (ViT-B/32) [18]87.2
PAC-S (ViT-B/32) [47]89.9
BRIDGE (ViT-B/32)91.5
CLIP-S (ViT-L/14) [18]90.9
PAC-S (ViT-L/14) [47]91.9
BRIDGE (ViT-L/14)93.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.464 + ], + "angle": 0, + "content": "both ViT-B/32 and ViT-L/14, BRIDGE consistently outperform CLIP-S across all categories, showcasing an average accuracy increase of \\(+1.7\\) and \\(+1.2\\) points, respectively. When comparing with PAC-S, our solution can better discriminate the correct captions on both backbones with an average accuracy increase of \\(+0.2\\) and \\(+0.7\\) using ViT-B/32 and ViT-L/14 respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.591 + ], + "angle": 0, + "content": "Object Hallucination Analysis. We then extend our analysis to the FOIL dataset [49] for correctly identifying captions that may contain object hallucinations. Table 5 shows the accuracy results. As it can be seen, BRIDGE outperforms CLIP-S and PAC-S, exhibiting an increase respectively of \\(+4.3\\) and \\(+1.6\\) points when using ViT-B/32 as backbone. Similar improvements can also be observed when employing more robust visual features. These results demonstrate the capabilities of our metric also to identify hallucinated objects correctly. In Fig 4, we present sample results comparing our metric with CLIP-S and PAC-S." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.466, + 0.631 + ], + "angle": 0, + "content": "4.5 System-level Correlation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Finally, we delve into the efficacy of our proposed metric when evaluating popular existing captioning models. To this aim, we generate predictions of several state-of-the-art captioning models on the COCO test set, including Show and Tell and Show, Attend and Tell which are among the first image captioning models based on deep learning, Up-Down [3], SGAE [59], AoANet [20], \\(\\mathcal{M}^2\\) Transformer [14], X-Transformer [40] which all include region-based image features with either LSTM-based or Transformer-based language models, and the recently proposed COS-Net model [32] that incorporates CLIP features. In addition to reporting evaluations on traditional captioning models, we also include recent LLM-based captioning models, including ZeroCap [51] and SmallCap [44], which are based on GPT-2 [43], and MiniGPT-v2 [9], BLIP-2 [28], IDEFICS [25], LLaVA-1.5 [35,36], and InstructBLIP [15] which instead are based on larger-scale LLMs like FlanT5 [12], Vicuna [11], or LLaMA [52,53]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Sarto et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 6: Evaluation scores of traditional and LLM-based captioners on COCO test set (\\(\\spadesuit\\): not trained/fine-tuned on COCO)." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.179, + 0.751, + 0.378 + ], + "angle": 0, + "content": "
BLEU-4METEORCIDErCLIP-SPAC-SBRIDGE
TraditionalShow and Tell [56]31.425.097.20.5720.7720.788
Show, Attend and Tell [58]33.426.2104.60.5820.7850.804
Up-Down [3]36.727.9122.70.7230.8030.821
SGAE [59]39.028.4129.10.7340.8120.833
AoANet [20]38.929.2129.80.7370.8150.836
M2Transformer [14]39.129.2131.20.7340.8130.841
X-Transformer [40]39.729.5132.80.6100.8120.845
COS-Net [32]42.030.6141.10.7580.8320.859
LLM-basedZeroCap♦[51]2.310.115.10.8100.8160.862
SmallCap [44]37.027.9119.70.7480.8260.847
MiniGPT-v2 [9]18.824.680.40.7520.8180.845
BLIP-2 [28]43.732.0145.80.7670.8370.868
IDEFICS-9B♦[25]4.319.150.00.7400.7860.838
LLaVA-1.5-7B♦[35]8.128.069.60.7840.8090.867
InstructBLIP-Flan-T5-XL♦[15]6.128.138.10.8170.8370.902
Humans-24.187.60.7740.8230.856
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.785, + 0.546 + ], + "angle": 0, + "content": "The results are presented in Table 6, where we evaluate our BRIDGE scores against both standard metrics, such as BLEU-4, METEOR, and CIDEr, and more recent ones like CLIP-S and PAC-S. Captioning models are compared to a human baseline, in which, for each sample, one human-annotated sentence (selected randomly from the five provided by the COCO dataset) serves as a candidate caption. As shown in the table, BRIDGE can effectively evaluate human-annotated sentences which obtain a score similar to recent state-of-the-art captioning models such as COS-Net. This capability lacks in standard metrics such as METEOR and CIDEr which rank human captions lower than those generated by less-performing captioning models like Show, Attend, and Tell or Up-Down." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.787, + 0.697 + ], + "angle": 0, + "content": "When considering the evaluation of captions generated by existing models, our metric shows a strong correlation with standard evaluation metrics when evaluating traditional image captioners or large-scale models that are fine-tuned on the COCO dataset like BLIP-2. When instead considering more recent approaches that are based on large language models and are not fine-tuned on COCO, BRIDGE still recognizes the goodness of generated captions, raking InstructBLIP as the best-performing approach. This demonstrates the capabilities of our metric to correctly evaluate longer and more detailed captions which are typically generated by LLM-based multimodal models [7, 16, 30] and that might be very different from captions contained in the COCO dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.721, + 0.36, + 0.737 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this paper, we have presented a novel learnable, and reference-free image captioning metric that combines text and dense visual features. Our proposal, BRIDGE, employs templated captions that are enriched with fine-grained visual cues thanks to a mapping network. Through experimental evaluation, we demonstrate that BRIDGE outperforms existing reference-free metrics in terms of correlation with human judgment and sensitivity to hallucinated objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.267 + ], + "angle": 0, + "content": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been conducted under a research grant co-funded by Leonardo S.p.A. and supported by the PNRR-M4C2 (PE0000013) project \"FAIR - Future Artificial Intelligence Research\" and by the PRIN project \"MUSMA\" (CUP G53D23002930006 - M4C2 I1.1), both funded by EU - Next-Generation EU." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.289, + 0.323, + 0.305 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.359 + ], + "angle": 0, + "content": "1. Aditya, S., Yang, Y., Baral, C., Fermuller, C., Aloimonos, Y.: From Images to Sentences through Scene Description Graphs using Commonsense Reasoning and Knowledge. arXiv preprint arXiv:1511.03292 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.785, + 0.387 + ], + "angle": 0, + "content": "2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: SPICE: Semantic Propositional Image Caption Evaluation. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.785, + 0.428 + ], + "angle": 0, + "content": "3. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.455 + ], + "angle": 0, + "content": "4. Banerjee, S., Lavie, A.: METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In: ACL Workshops (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.496 + ], + "angle": 0, + "content": "5. Barraco, M., Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: With a Little Help from your own Past: Prototypical Memory Networks for Image Captioning. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.785, + 0.524 + ], + "angle": 0, + "content": "6. Bird, S., Klein, E., Loper, E.: Natural Language Processing with Python: Analyzing Text with the Natural Language Toolkit. O'Reilly Media, Inc. (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.526, + 0.785, + 0.565 + ], + "angle": 0, + "content": "7. Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Baraldi, L., Cornia, M., Cucchiara, R.: The Revolution of Multimodal Large Language Models: A Survey. In: ACL Findings (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.785, + 0.592 + ], + "angle": 0, + "content": "8. Chan, D., Petryk, S., Gonzalez, J.E., Darrell, T., Canny, J.: CLAIR: Evaluating Image Captions with Large Language Models. In: EMNLP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.785, + 0.633 + ], + "angle": 0, + "content": "9. Chen, J., Li, D.Z.X.S.X., Zhang, Z.L.P., Xiong, R.K.V.C.Y., Elhoseiny, M.: MiniGPT-v2: Large Language Model as a Unified Interface for Vision-Language Multi-task Learning. arXiv preprint arXiv:2310.09478 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.661 + ], + "angle": 0, + "content": "0. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: UNITER: Universal Image-TExt Representation Learning. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.702 + ], + "angle": 0, + "content": "1. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An Open-Source Chatbot Improressing GPT-4 with \\(90\\%\\) * ChatGPT Quality (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.704, + 0.785, + 0.785 + ], + "angle": 0, + "content": "12. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S.S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Castro-Ros, A., Pellat, M., Robinson, K., Valter, D., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E.H., Dean, J., Devlin, J., Roberts, A., Zhou, D., Le, Q.V., Wei, J.: Scaling Instruction-Finetuned Language Models. JMLR 25(70), 1-53 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.787, + 0.785, + 0.812 + ], + "angle": 0, + "content": "13. Cornia, M., Baraldi, L., Cucchiara, R.: SMArT: Training Shallow Memory-aware Transformers for Robotic Explainability. In: ICRA (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Cornia, M., Stefanini, M., Baraldi, L., Cucchiara, R.: Meshed-Memory Transformer for Image Captioning. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.318, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Sarto et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning. arXiv preprint arXiv:2305.06500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "16. Dong, H., Li, J., Wu, B., Wang, J., Zhang, Y., Guo, H.: Benchmarking and Improving Detail Image Caption. arXiv preprint arXiv:2405.19092 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "17. Herdade, S., Kappeler, A., Boakye, K., Soares, J.: Image captioning: Transforming objects into words. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.785, + 0.273 + ], + "angle": 0, + "content": "18. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In: EMNLP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "19. Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking task: Data, models and evaluation metrics. JAIR 47, 853-899 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "20. Huang, L., Wang, W., Chen, J., Wei, X.Y.: Attention on Attention for Image Captioning. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "21. Jiang, M., Hu, J., Huang, Q., Zhang, L., Diesner, J., Gao, J.: REO-Relevance, Extraness, Omission: A Fine-grained Evaluation for Image Captioning. In: EMNLP (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "22. Jiang, M., Huang, Q., Zhang, L., Wang, X., Zhang, P., Gan, Z., Diesner, J., Gao, J.: TIGER: Text-to-Image Grounding for Image Caption Evaluation. In: EMNLP (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "23. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.467 + ], + "angle": 0, + "content": "24. Kim, J.H., Kim, Y., Lee, J., Yoo, K.M., Lee, S.W.: Mutual Information Divergence: A Unified Metric for Multimodal Generative Models. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "25. Laurencon, H., Saulnier, L., Tronchon, L., Bekman, S., Singh, A., Lozhkov, A., Wang, T., Karamcheti, S., Rush, A.M., Kiela, D., Cord, M., Sanh, V.: OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "26. Lee, H., Yoon, S., Dernoncourt, F., Bui, T., Jung, K.: UMIC: An Unreferenced Metric for Image Captioning via Contrastive Learning. In: ACL (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "27. Lee, H., Yoon, S., Dernoncourt, F., Kim, D.S., Bui, T., Jung, K.: ViLBERTScore: Evaluating Image Caption Using Vision-and-Language BERT. In: EMNLP Workshops (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.619 + ], + "angle": 0, + "content": "28. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping Language-Image Pretraining with Frozen Image Encoders and Large Language Models. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.647 + ], + "angle": 0, + "content": "29. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "30. Li, X., Tu, H., Hui, M., Wang, Z., Zhao, B., Xiao, J., Ren, S., Mei, J., Liu, Q., Zheng, H., et al.: What If We Recap tion Billions of Web Images with LLaMA-3? arXiv preprint arXiv:2406.08478 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.73 + ], + "angle": 0, + "content": "31. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.758 + ], + "angle": 0, + "content": "32. Li, Y., Pan, Y., Yao, T., Mei, T.: Comprehending and ordering semantics for image captioning. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "33. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL Workshops (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "35. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved Baselines with Visual Instruction Tuning. In: CVPR (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.782, + 0.161 + ], + "angle": 0, + "content": "36. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual Instruction Tuning. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.162, + 0.785, + 0.188 + ], + "angle": 0, + "content": "37. Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.216 + ], + "angle": 0, + "content": "38. Lu, J., Batra, D., Parikh, D., Lee, S.: ViLBERT: Pretraining Task-Agnostic Visi-olinguistic Representations for Vision-and-Language Tasks. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.216, + 0.785, + 0.242 + ], + "angle": 0, + "content": "39. Oord, A.v.d., Li, Y., Vinyals, O.: Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.269 + ], + "angle": 0, + "content": "40. Pan, Y., Yao, T., Li, Y., Mei, T.: X-Linear Attention Networks for Image Captioning. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.785, + 0.296 + ], + "angle": 0, + "content": "41. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: BLEU: a method for automatic evaluation of machine translation. In: ACL (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.785, + 0.338 + ], + "angle": 0, + "content": "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transferable Visual Models From Natural Language Supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "43. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners. OpenAI Blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.392 + ], + "angle": 0, + "content": "44. Ramos, R., Martins, B., Elliott, D., Kementchedjhieva, Y.: SmallCap: lightweight image captioning prompted with retrieval augmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.418 + ], + "angle": 0, + "content": "45. Rashtchian, C., Young, P., Hodosh, M., Hockenmaier, J.: Collecting Image Annotations Using Amazon's Mechanical Turk. In: NAACL Workshops (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.785, + 0.459 + ], + "angle": 0, + "content": "46. Saito, K., Sohn, K., Zhang, X., Li, C.L., Lee, C.Y., Saenko, K., Pfister, T.: Pic2Word: Mapping Pictures to Words for Zero-Shot Composed Image Retrieval. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "47. Sarto, S., Barraco, M., Cornia, M., Baraldi, L., Cucchiara, R.: Positive-Augmented Contrastive Learning for Image and Video Captioning Evaluation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.527 + ], + "angle": 0, + "content": "48. Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: Retrieval-Augmented Transformer for Image Captioning. In: CBMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.568 + ], + "angle": 0, + "content": "49. Shekhar, R., Pezzelle, S., Klimovich, Y., Herbelot, A., Nabi, M., Sangineto, E., Bernardi, R.: FOIL it! Find One mismatch between Image and Language caption. In: ACL (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.609 + ], + "angle": 0, + "content": "50. Shi, Y., Yang, X., Xu, H., Yuan, C., Li, B., Hu, W., Zha, Z.J.: EMScore: Evaluating Video Captioning via Coarse-Grained and Fine-Grained Embedding Matching. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.636 + ], + "angle": 0, + "content": "51. Tewel, Y., Shalev, Y., Schwartz, I., Wolf, L.: Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.785, + 0.691 + ], + "angle": 0, + "content": "52. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "53. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.785, + 0.759 + ], + "angle": 0, + "content": "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "55. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: CIDEr: Consensus-based Image Description Evaluation. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "56. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "57. Wang, S., Yao, Z., Wang, R., Wu, Z., Chen, X.: FAIER: Fidelity and Adequacy Ensured Image Caption Evaluation. In: CVPR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Sarto et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "58. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhutdinov, R., Zemel, R.S., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "59. Yang, X., Tang, K., Zhang, H., Cai, J.: Auto-Encoding Scene Graphs for Image Captioning. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "60. Yi, Y., Deng, H., Hu, J.: Improving Image Captioning Evaluation by Considered Inter Refinences Variance. In: ACL (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "61. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL 2, 67-78 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "62. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: VinVL: Revisiting visual representations in vision-language models. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "63. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: BERTScore: Evaluating Text Generation with BERT. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "64. Zhu, W., Wang, X.E., Yan, A., Eckstein, M., Wang, W.Y.: ImaginE: An Imagination-Based Automatic Evaluation Metric for Natural Language Generation. In: EACL (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_origin.pdf b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..73dbf6a9c2c2c29702c38be372104a67ba1ddef8 --- /dev/null +++ b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/f7f05300-4ad0-468f-b4b8-6f20ff2ba5a0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:134d474cd562a66a7d337164fcb2df9d02b1152a7badbd9f15f9ea86195528a4 +size 1830126 diff --git a/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/full.md b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/full.md new file mode 100644 index 0000000000000000000000000000000000000000..706295ed92fc59ccfa25991d897cc5204d32b07b --- /dev/null +++ b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/full.md @@ -0,0 +1,345 @@ +# BRIDGE: Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues + +Sara Sarto1, Marcella Cornia1, Lorenzo Baraldi1, and Rita Cucchiara1,2 + +1 University of Modena and Reggio Emilia, Italy + +2 IIT-CNR, Italy + +{name.surname}@unimore.it + +Abstract. Effectively aligning with human judgment when evaluating machine-generated image captions represents a complex yet intriguing challenge. Existing evaluation metrics like CIDEr or CLIP-Score fall short in this regard as they do not take into account the corresponding image or lack the capability of encoding fine-grained details and penalizing hallucinations. To overcome these issues, in this paper, we propose BRIDGE, a new learnable and reference-free image captioning metric that employs a novel module to map visual features into dense vectors and integrates them into multi-modal pseudo-captions which are built during the evaluation process. This approach results in a multimodal metric that properly incorporates information from the input image without relying on reference captions, bridging the gap between human judgment and machine-generated image captions. Experiments spanning several datasets demonstrate that our proposal achieves state-of-the-art results compared to existing reference-free evaluation scores. Our source code and trained models are publicly available at: https://github.com/aimagelab/bridge-score + +Keywords: Captioning Evaluation $\cdot$ Vision-and-Language + +# 1 Introduction + +The objective of image captioning is to produce natural language descriptions conditioned on input images, that closely resemble human language and align to human intentions. As such, the captioning task involves the recognition and understanding of the visual content of the image, including fine-grained elements such as objects, attributes, and their relationships. Advances in training methodologies and architectures have contributed to the progress in the field, significantly improving the generation quality. Recent innovations include fully-attentive models [13, 14, 17], improved connections between visual and textual modalities [14, 40], and the incorporation of objects and tags at an architectural level [3, 31, 62]. Additionally, there has been a notable focus on increasing the robustness of cross-modal features [5, 32, 48], which consequently can increase description accuracy. + +As constant improvements are made on the generation side, it becomes crucial to enhance the evaluation process as well. In this regard, image captioning + +![](images/f147fd6145523e55daa46720fb0e57567e8ca3669b2da5b0801f9c62bd99591c.jpg) +Fig. 1: Comparison between different captioning evaluation approaches: (a) CIDEr [55] scores candidate and reference captions without considering the input image; (b) CLIP-Score [18] compares text and images using global vectors in a shared embedding space; (c) our BRIDGE, internally builds multimodal pseudo-captions by translating fine-grained image features into pseudo-tokens thanks to a mapping module. + +evaluation aims to assess the quality of a generated caption given an image and, potentially, human-written reference captions. However, it is important to note that obtaining these reference captions can often be challenging and expensive, adding complexity to the evaluation process. Despite the recent advancements in captioning capabilities, standard automatic evaluation metrics have mainly relied on translation metrics [4,33,41] or text-only ones [2,55,63] which often fall short in capturing aspects such as grammatical correctness, semantic relevance, and specificity. These limitations are worsened by the limited coverage of image content in available references, resulting in inaccurate penalties when generated captions accurately describe novel elements not mentioned in the references. + +In response to these limitations, advanced metrics aligning visual and textual data have emerged [18,24,26,27]. Notably, recent metrics leverage the CLIP embedding space [42], which shows a strong correlation with human judgment. Despite the effectiveness of contrastive-based embedding spaces, metrics based on dual-encoder architectures tend to focus on global alignment between an image and its caption, often ignoring fine-grained details or penalizing hallucinations. + +Following this insight, in this paper, we introduce BRIDGE, a novel learnable and reference-free image captioning metric that enhances the alignment of more fine-grained visual features. Specifically, our model provides a pre-trained dual-encoder architecture with a mapping module designed to effectively exploit visual cues (Fig. 1). This is done by internally creating multimodal pseudo-captions, containing both textual and dense visual features. The process for building these pseudo-captions involves the creation of a template caption, which focuses on the syntactical structure of the scene, and a mapping module. The latter refines the template caption by enriching it with more fine-grained visual features about the subjects depicted in the image. Subsequently, the overall model is trained with a combination of contrastive losses which promote multimodal alignment. + +Experiments are conducted on a variety of datasets for image captioning annotated with human rankings, including Flickr8k-Expert, Flickr8k-CF [19], Composite [1], and Pascal-50S [55]. Through comprehensive analysis, we demonstrate the efficacy of the proposed metric and show its ability to overcome the limitations of existing state-of-the-art reference-free alternatives. Additionally, we evaluate its sensitivity to object hallucination by conducting experiments on the FOIL dataset [49]. Overall, BRIDGE outperforms previous metrics and showcases superior performance compared to CLIP-Score [18] and PAC-Score [47]. + +Contributions. In summary, our contributions are as follows: + +- We tackle the limitations of existing image captioning metrics by proposing the first learnable reference-free metric, termed as BRIDGE, that focuses on more fine-grained visual features. Our proposal integrates a dual-encoder architecture with a mapping module which is in charge of producing multimodal pseudo-captions that combine text and richer dense visual features. +- Multimodal pseudo-captions are built by creating template captions that describe the scene from a syntactical point of view. These templates are subsequently enriched with fine-grained features through a mapping network. +- Experiments, carried out on different datasets with human preferences, demonstrate a higher degree of correlation with respect to existing metrics. We also evaluate the sensitivity of the proposal to objects hallucinations. + +# 2 Related Work + +Classical Reference-based Captioning Metrics. Several widely used captioning evaluation metrics were originally developed in the context of NLP tasks and rely on n-gram matching techniques. Among these classical metrics, BLEU [41] is designed to focus on precision and incorporates a penalty for sentence brevity. METEOR [4], instead, combines precision and recall to evaluate the quality of captions, while others, such as ROUGE [33], were initially born for summarization tasks and later adapted to image captioning. More recently, two metrics tailored for visual captioning have emerged: CIDEr [55], which measures n-gram similarity and is based on TF-IDF, and SPICE [2], which quantifies graph-based similarity through scene graphs constructed from candidate and reference captions. Overall, focusing on textual-level comparisons, these metrics assume that human-written references accurately represent the image content. + +Learnable Captioning Metrics. With the rise of large pre-trained models, image captioning evaluation now frequently exploits these models to compare textual-only [60, 63] or visual-textual [18, 21, 22, 26, 27, 57] contents. Notably, the BERT score and its improved version use pre-trained BERT embeddings to compare word tokens in generated and ground-truth sentences. + +Some metrics, like BLEU and CIDEr, rely solely on text matching between reference captions and machine-generated captions, potentially introducing bias in evaluations due to non-accurate reference captions. To mitigate these issues, alternative solutions leverage the multimodal nature of vision-and-language models. As an example, TIGEr [22] considers the similarities between words and image regions and assesses how well machine-generated captions represent image content and their alignment with human-generated captions. + +In contrast, other approaches [18, 24, 27] leverage web-scale vision-and-language models such as VilBERT [38] and CLIP [42] for more robust metrics. For example, in [24], CLIP visual-textual features are used to compute negative Gaussian cross-mutual information. Other works, instead, have employed diffusion models in text-only tasks [64] or exploited the zero-shot language modeling capabilities of large language models [8] to evaluate candidate captions. + +Reference-free Captioning Metrics. While all aforementioned metrics rely on a set of ground-truth captions to compute the final score, a few attempts have been made to introduce reference-free evaluations, only taking into account the correlation of the candidate caption with the image. In this regard, Lee et al. [26] proposed to fine-tune the UNITER model [10] via contrastive learning to let the model to discriminate between positive and synthetically-generated negative captions. On a different line, Hessel et al. [18] introduced the CLIP-Score, which only relies on a modified cosine similarity between image and candidate caption representations coming from the CLIP model. The recently proposed PAC-Score [47], instead, builds upon the usage of CLIP but incorporates a fine-tuning phase with positive augmentation, further enhancing the accuracy of evaluation. In this paper, we follow this research path and propose a novel learnable, and reference-free evaluation metric that can effectively incorporate fine-grained visual features for evaluating the correlation of image-text pairs. + +# 3 BRIDGE for Captioning Evaluation + +In the following, we present our reference-free captioning evaluation approach, termed BRIDGE. Our approach leverages a dual-encoder architecture, which comprises both a language encoder and a vision encoder. Given a frozen pretrained model, we train a mapping module responsible for filling the holes of a masked template caption with pseudo language tokens that are enriched with visual information. An overview of our model is depicted in Fig. 2. + +# 3.1 Preliminaries + +Our approach relies on CLIP (Contrastive Language-Image Pre-training) [42], a powerful vision and language model designed to align images and corresponding text captions within a shared embedding space. For a given input image $I$ , the image encoder $E_V$ extracts the visual information $v = E_V(I) \in \mathbb{R}^d$ . On the textual side, an input caption $T$ is tokenized and a textual representation is obtained by passing it through the textual encoder $E_T$ , obtaining $t = E_T(T) \in \mathbb{R}^d$ . Once the textual and visual features, $t$ and $v$ respectively, are projected in a common space, visual and textual inputs can be compared via cosine similarity. + +The relationships learned by CLIP can be exploited to build an image captioning evaluator. In CLIP-Score [18] the authors directly compare candidate captions and images in the embedding space and show that this achieves a good correlation with human judgments. In detail, to assess the quality of a candidate generation, they feed both the image and the candidate caption through their respective feature extractors, and compute the cosine similarity of the resultant embeddings: + +$$ +\operatorname {C L I P} \text {- S c o r e} (I, T) = w \cdot \max (\cos (v, t), 0), \tag {1} +$$ + +where $w$ is a rescaling factor employed to stretch the score distribution while ensuring the ranking results remain unchanged. + +![](images/3c2d21a265ec7cb0e78f844dea2608c71b92de6044e3a19960a8c5df6c8fbaee.jpg) +Fig. 2: Overview of the BRIDGE evaluation approach. Starting from a template caption, a mapping network augments it with dense visual features, obtaining a pseudocaption which is then used for computing image-text similarities. + +# 3.2 Injecting Fine-Grained Visual Features + +Unlike CLIP-Score, our approach does not rely exclusively on global image descriptors for evaluating image-text alignments. Instead, we focus on employing stronger visual information. To do so, we draw inspiration from the Pic2Word approach [46] and represent the input image through a multimodal pseudo-caption, an embedding representation that contains stronger visual elements. + +Building Template Captions. In order to create multimodal pseudo-captions, we first build template captions for a given input image. These are skeletal textual representations of the image, obtained by masking out all the relevant textual concepts from the descriptions generated by a captioner. Through these template captions, we aim to provide the model only with a templated textual structure which can then be filled with more fine-grained visual features. In particular, given an automatically generated caption describing the input image, such as 'A man running with a white dog', we remove the main subjects within the caption (e.g. 'man' and 'white dog') and mask them with a [MASK] token. This will allow the model to fill in these gaps by incorporating more fine-grained features from the image encoder. Since a primary subject might be described by words other than just its corresponding nouns (e.g. adjectives), we utilize noun chunks. Fig. 3 reports template captions and corresponding noun chunks. + +Given a sentence containing $N$ noun chunks, we independently encode them through the mapping network. To this aim, we replicate the template caption as many times as the number of noun chunks and mask a different noun chunk in each of the replicas. We thus obtain $N$ different versions of the template caption, each one masking a single noun chunk, for instance + +[A [MASK] running with a white dog', + +'A man running with a [MASK]'. + +![](images/e9b1d339c07d6c2584bb07f66cbc26842c5bb1254eaf9bbe4b187ed5e3d5e3dd.jpg) +COCO +TEMPLATE +NOUN CHUNKS +yellow bus +Fig. 3: COCO captions with template captions and associated noun chunks. + +![](images/a46b7b22878b3171115f7ebea4c2888bb32cfcce876bfbac649c85e857d428d0.jpg) +People on a motorcycle in front of a road +A yellow bus driving down a street next to a park. +street +park +COCO +TEMPLATE +NOUN CHUNKS +Little dog laying on a pillow next to the laptop. +A dog laying on a couch next to a PC. +PC +dog +couch + +NOUN CHUNKS + +Mapping with Fine-Grained Visual Features. The above-described masked replicas are then fed to the mapping module $\psi$ . Specifically, our approach exploits the visual information extracted from the visual encoder $E_V$ to enrich the replicas with the informative content of the image $I$ . To get a more fine-grained representation of the image, we directly take the grid of features from the last layer, $\widehat{v}$ . For instance, in the case of a ViT-B/32 backbone, this will have a shape of $50 \times d$ , where $d$ is the dimensionality of the last embedding of the network. + +The mapping network is implemented as a stack of Transformer [54] encoder layers interleaved with cross-attention layers. Its role is to refine each template captions with visual information. Since each template caption is processed independently, the mapping module returns a set of sequences, each with the same length as the corresponding input template caption. From the output of the mapping module, we keep only the predictions for the masked tokens in each template caption and copy them back into the original templates. + +Therefore, by providing a masked input template in the form $\hat{T_i} = [w_1,\dots,w_{j - 1},\mathsf{MASK},w_{j + 1},\dots,w_T]$ , where $\{w_j\}_j$ represent original tokens from the input caption, we obtain $\hat{T_i}^* = \left[w_1,\dots,w_{j - 1},\psi (\hat{T_i})_j,w_{j + 1},\dots,w_T\right]$ , where $\psi (\hat{T_i},\widehat{v})_j$ represents the output of the mapping network at the position corresponding to the masked input token position. In the case of noun chunks consisting of more than one token, multiple consecutive tokens are replaced with the corresponding outputs from the mapping network. By injecting the outputs of the mapping token into the initial template caption, we effectively complete the original templates with visually enriched vectors. Notably, these newly generated pseudo-captions combine word sequences from the template captions with dense vectors obtained by the mapping module. Consequently, they cannot be decoded as standard captions. As a last step, the obtained pseudo-captions are fed into the pre-trained CLIP language encoder. + +# 3.3 Training Protocol + +To train our mapping network, the loss is defined as a weighted version of the symmetric InfoNCE loss [39], where positive and negative items are weighted according to the number of noun chunks in each caption. + +Specifically, given a batch in the form $\mathcal{B} = \{(I_i,T_i)\}_{i = 1}^N$ , where $I_{i}$ and $T_{i}$ represent image-caption pairs, each image $I_{i}$ is expanded in $N_{i}$ multimodal pseudocaptions as outlined above, where $N_{i}$ is the number of noun chunks in caption $T_{i}$ . Further, let $t_{ij}^{*}$ represent the embedding vector of the $j$ -th pseudo-caption derived from the $i$ -th image, $v_{i}$ the embedding vector of the $i$ -th image and $t_{i}$ + +the embedding vector of the $i$ -th ground-truth caption. Finally, let $M$ be the total number of noun chunks in the mini-batch, i.e. $M = \sum_{i=1}^{N} N_i$ . + +The first loss, denoted as $\mathcal{L}_1$ , tries to align the pseudo-captions $t_{ij}^*$ with the global visual features of the corresponding images $v_i$ . In addition to this loss term, we define a second loss component $\mathcal{L}_2$ that promotes the alignment between pseudo-captions $t_{ij}^*$ and the textual feature vector of the ground-truth caption $t_i$ corresponding to the input image. This ensures that pseduo-captions are aligned also on a textual space, in addition to being aligned in the image space. Additional details can be found in the supplementary materials. + +Regularization Branch. With the aforementioned loss terms, our objective is encouraging an association between each pseudo-caption and its corresponding image and caption. However, it is also important to differentiate each pseudocaption from the others of the same image. To achieve this, we define a regularization loss which promotes a precise alignment between each pseudo-caption and the corresponding noun chunk. + +First, we create prompts like "a photo of a " and encode them with the text encoder $E_{T}$ . In parallel, each pseudo-caption is fed into a dedicated multilayer perceptron (MLP) projection, which consists of two linear layers with a ReLU activation in between. Formally, the branch is defined as + +$$ +\mathcal {C} (x) = \operatorname {L i n e a r} (\operatorname {R e L U} (\operatorname {L i n e a r} (x))). \tag {2} +$$ + +Since our goal is to emphasize each pseudo-caption's alignment with its corresponding noun chunk, we employ a regular contrastive loss $\mathcal{L}_3$ between the prompts mentioned earlier and the outputs of the projection branch. + +Finally, the overall loss function we train BRIDGE is defined as a weighted summation of two aforementioned losses, plus the regularization loss, as + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {1} + \lambda_ {2} \mathcal {L} _ {2} + \lambda_ {3} \mathcal {L} _ {r}. \tag {3} +$$ + +# 3.4 Inference and Score Computation + +At inference time, given an image-candidate caption pair $(I,T)$ , we extract all pseudo-captions from $I$ using our mapping network. Subsequently, we compute the mean pseudo-caption embedding as $t^{*} = \frac{1}{N}\hat{t}_{i}^{*}$ , where $\hat{t}_i^*$ indicates the $i$ -th pseudo-caption extracted from $I$ and $N$ here indicates the overall number of pseudo-captions associated with $I$ . + +At that point, given the visual embedding $v$ of the image and the embedding of the candidate caption $t$ , the matching score between $I$ and $T$ is defined as + +$$ +\operatorname {B R I D G E} (I, T) = 0. 5 \cdot [ \operatorname {C L I P - S c o r e} (I, T) + w \cdot \max (\cos \left(t ^ {*}, t\right), 0) ], \tag {4} +$$ + +where $\cos$ indicates the cosine similarity and $w$ is a constant scaling factor. + +# 4 Experimental Evaluation + +# 4.1 Implementation Details + +Architecture and Training Details. Building upon prior research [18,24,50], we use either CLIP [42] ViT-B/32 or ViT-L/14 as backbone for the visual and textual encoder. The mapping module is composed of two Transformer layers and is trained on the COCO dataset [34], which contains more than 120k images annotated with five captions. In particular, we employ the splits introduced by Karpathy et al. [23], where 5,000 images are used for both validation and testing and the rest for training. To map the grid visual features to an embedding space of dimension 512, we employ a simple linear projection. For the regularization branch, we utilize a two-layer multi-layer perceptron. + +During training, we use AdamW [37] as optimizer with a learning rate equal to 0.0001 and a batch size of 256. The $\lambda_1$ , $\lambda_2$ , and $\lambda_3$ values are selected with a grid search, choosing the combination that provides the best validation loss. Specifically, we set both $\lambda_1$ and $\lambda_3$ to 0.01, while $\lambda_2$ is set to 1.0. The training stage lasts around one day on a single A100 GPU. + +Template Caption Generation. The template captions used as input for the mapping module are generated using the BLIP model [29]. In particular, we use the ViT-L/14 version pre-trained on 129M image-text pairs and finetuned on the COCO dataset. After this generation phase, the primary subjects of the template sentences are extracted by using the NLTK library [6]. During training, two noun chunks are randomly chosen from the set identified during the extraction step. In the evaluation phase, otherwise, all identified noun chunks are included. + +# 4.2 Datasets + +To evaluate the correlation of the proposed metric with human ratings, we conduct experiments on the Flickr8k-Expert, Flickr8k-CF, Composite, and Pascal50-S datasets [1,19,55]. In addition, for detecting hallucinations in textual sentences, we extend our analysis to the FOIL dataset [49]. Except for Pascal50S and FOIL in where accuracy scores are used, evaluation on all other datasets relies on Kendall $\tau_{b}$ , Kendall $\tau_{c}$ , and Spearman $\rho$ correlation scores. + +Flickr8k-Expert and Flickr8k-CF [19]. These datasets consist of image-caption pairs with corresponding human ratings. Specifically, Flickr8k-Expert comprises 17k expert annotations for visual-textual pairs, with a total of 5,664 different images. Each pair receives a score ranging from 1 (lack of correlation) to 4 (accurate depiction), where 1 indicates a lack of correlation between the caption and the image, and 4 indicates an accurate depiction of the image without errors. On the other hand, Flickr8k-CF is composed of 145k binary quality judgments, collected from CrowdFlower, for 48k image-caption pairs containing 1,000 unique images. Each pair is annotated with at least three binary scores, where “yes” denotes that the caption correlates with the image. To measure the alignment with human judgment, we compute the mean proportion of “yes” annotations as the score for each pair. + +Table 1: Ablation study results. US indicates the number of pseudo tokens. + +
Flickr8k-ExpertFlickr8k-CFPascal-50S
Kend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρAccuracy
Architectural Components
w/o mapping module (i.e. MLP)53.153.565.335.518.343.581.9
w/o template captions53.754.166.035.518.443.682.5
w/o regularization branch54.154.566.535.718.543.682.7
Score Formulation
w/o textual similarity51.151.263.034.417.730.580.9
w/o visual similarity53.854.266.035.418.343.781.9
Pseudo-token Size
w/ US = 154.154.566.435.117.130.382.6
w/ US = 254.154.566.436.118.744.482.8
w/ US = 454.354.766.635.418.343.882.4
w/ US = 854.054.466.335.918.644.581.7
BRIDGE (US = 3)54.454.867.736.118.744.582.6
+ +Composite [1]. It contains 12k human ratings for image-caption pairs including a combination of images taken from COCO [34] (2,007 images), Flickr8k [19] (997 images), and Flickr30k [61] (991 images). In this dataset, human evaluators assess each image-caption pair, assigning a score within the range of 1 to 5 to estimate the correspondence of the caption with the associated image. + +Pascal50-S [55]. It presents pairwise preference judgments between two captions. Overall, the dataset consists of 4,000 sentence pairs, each of them associated with an image from the UIUC Pascal sentence dataset [45]. Each pair is associated with 48 human judgments, where each evaluation indicates which sentence better describes the given image. The sentence pairs are categorized into four groups: (i) both human-written and correct captions (HC), (ii) both human-written captions where one is correct and the other is wrong (HI), (iii) both correct captions but one written by humans and the other machine-generated (HM), (iv) both machine-generated and correct captions (MM). + +FOIL [49]. The dataset comprises image-caption pairs from the COCO dataset [34]. In this scenario, captions undergo perturbation by generating modified versions that closely resemble the originals but introduce a single error, referred to as "foil word". For a fair comparison, we select the subset of the validation set that does not overlap with the portion of COCO used during training, resulting in 8,000 images, each paired with a foil-correct textual counterpart. + +# 4.3 Ablation Studies and Analysis + +To evaluate the effectiveness of our metric, we start by analyzing variations of our main architectural components. Then, we assess the impact of caption templates in our score formulation. All these experiments are performed using CLIP ViT-B/32 as backbone and reported in Table 1. + +Contribution of Architectural Components. We first investigate the performance of the most straightforward implementation of a mapping module, structured as a two-layer MLP following [46]. We also validate the importance + +Table 2: Impact of different template captions. + +
ExpertCFPascal-50S
Kend. τbKend. τbAcc.
Transformer Templates
w/o mapping module46.131.680.4
BRIDGE54.035.982.7
(+7.9)(+4.3)(+2.3)
BLIP Templates
w/o mapping module48.633.881.0
BRIDGE54.436.182.6
(+5.8)(+2.3)(+1.6)
BLIP-2 Templates
w/o mapping module49.434.282.4
BRIDGE54.436.282.9
(+5.0)(+2.0)(+0.5)
+ +![](images/612a9f2728812130403915490f8143dc29b7208409534a8d155ef174c35bc96e.jpg) +C C C C C C +# +B + +![](images/4bddc7b4c62a1eac6bbaba29d702bab35e35a37a90d2c72e28642d5fba601265.jpg) +TRANSE + +A young girl is preparing to blow out her candle + +A painting of a table with fruit on top of it + +A young girl is eating a piece of cake + +A painting of a painting of fruits and a painting + +A woman blowing out a candle on a cake + +A painting of a pitcher and oranges on a table + +A girl blowing out a candle on a cupcake + +A painting of a still life with oranges and a candle + +of the template captions through a model variant in which a set of learnable tokens $\mathsf{S}^*$ serves as input for the mapping module, without relying on template captions. In both variants, given the absence of template captions, we construct a template such as 'a photo of $\mathsf{S}^*$ ' and extract its features using the CLIP text encoder. In the Table, it can be seen that, regardless of any architectural changes, it is important to provide a simple sentence structure to the mapping module to achieve competitive performance. + +In addition to these baselines, we devise a variant to analyze the contribution of the regularization branch. In this setting, we employ template captions as input for the mapping module, resulting in a substantial improvement of $+1.0$ Kendall $\tau_{b}$ and $+0.8$ accuracy points compared to the MLP variant, respectively on the Flickr8k-Expert and on the Pascal-50S dataset. When introducing the regularization branch (i.e. the complete BRIDGE architecture), further enhancements can be observed especially on the Flickr8k-CF with an improvement of $+0.4$ points in terms of the Kendall $\tau_{b}$ correlation score. + +We also emphasize the importance of each component in our score formulation. Specifically, we present correlation results when employing only the visual similarity within our architecture, which is the original CLIP-Score formulation. As observed, performance drops drastically when relying only on visual information. A less significant drop is observed when employing only textual similarity. + +As an additional analysis, we report the effect of changing the number of the pseudo tokens for each noun chunk, denoting it as unit size (US). Specifically, we compute the scores employing $\mathrm{US} = 1,2,3,4,8$ . From the results, it can be seen that $\mathrm{US} = 3$ generally leads to the best performance across nearly all evaluation metrics. This configuration is used in all experiments reported in the paper. + +Analysis on Caption Templates. We analyze the effect of changing the initial template captions for our model. Specifically, we employ template captions generated by a conventional Transformer-based captioner that uses CLIP features as input and is trained only on the COCO dataset, as well as those generated by BLIP [29], and BLIP-2 [28]. Notably, we select template captions of different quality based on both standard metric evaluations and correlations with human judgment. To assess the quality of the raw generated templates, we observe that + +the CIDEr score of these models on the COCO test set is equal to 114.2, 131.4, and 145.8 respectively for the standard Transformer model, BLIP, and BLIP-2. Note that all models were trained with cross-entropy loss only. Results on the Flickr8k-Expert, Flickr8k-CF, and Pascal-50S datasets are reported in Table 2. To qualitatively validate the generated templates, we include sample captions generated by the three models compared to a ground-truth caption from the COCO test set. Specifically, captions generated by BLIP-2 are generally more detailed and effectively describe the visual content of the input image compared to those generated by BLIP and, notably, the standard Transformer model. + +For each caption template source, we compute the correlation scores when the mapping module is disabled and the captions are directly fed to the text encoder. We compare it with our standard BRIDGE score, considering the different caption templates. Across all datasets, it is evident that directly using the caption templates as input to the text encoder leads to poor performance. This highlights the intended flexibility of template captions as skeletal representations, allowing the model to enhance them with fine-grained visual features. + +In fact, starting from these simple template captions and following our approach, we achieve improvements of $+5.8$ and $+2.3$ Kendall $\tau_{b}$ points and $+1.6$ accuracy points, respectively, on the Flickr8k-Expert, Flickr8k-CF, and Pascal50-S datasets when using BLIP caption templates. The overall best results are with captions from the BLIP-2 model, confirming that better templates can indeed lead to improved results. However, even when using lower-quality captions, the final correlation results are very close to those obtained with higher-quality captions. This highlights the robustness of our metric to caption templates of varying quality and that it is not necessary to rely on captions generated by large-scale captioners to achieve strong correlation scores. + +# 4.4 Comparison with State-of-the-Art Captioning Metrics + +Evaluating Sample-Level Human Correlation. We evaluate the sample-level human correlation on the Flickr8k [19] and Composite [1] datasets. Following previous works [18,63], we compute Kendall correlation scores in both $\tau_{b}$ and $\tau_{c}$ versions and also include the Spearman $\rho$ score. Results are reported in Table 3, where we compare our proposed BRIDGE metric against other reference-free evaluation scores like UMIC [26], CLIP-S [18], and PAC-S [47]. Moreover, we also compare with standard captioning evaluation metrics (i.e. BLEU [41], ME-TEOR [4], CIDEr [55], and SPICE [2]) and more recent reference-based solutions that exploit text-only or cross-modal learned embeddings, such as BERT-S [63], BERT-S++ [60], TIGEr [22], VilBERTScore [27], and MID [24]. For completeness, we also include the reference-based versions of CLIP-S and PAC-S, termed RefCLIP-S and RefPAC-S, which however are not directly comparable with our solution as both rely on a set of five reference captions. + +As it can be seen, BRIDGE outperforms other reference-free metrics in terms of correlation with human judgment, achieving the highest scores on almost all datasets. Specifically, compared to CLIP-S, BRIDGE shows improvements in terms of Kendall $\tau_{c}$ of $+3.6$ and $+2.8$ points when using ViT-B/32 and ViT-L/14 + +Table 3: Correlation scores on Flickr8k-Expert, Flickr8k-CF, and Composite [1, 19]. + +
Flickr8k-ExpertFlickr8k-CFComposite
Kend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρ
Reference-based metrics
BLEU-4 [41]30.630.838.716.98.721.028.330.638.1
METEOR [4]41.541.851.922.211.527.636.038.948.1
CIDEr [55]43.643.954.324.612.730.534.937.747.0
SPICE [2]51.744.955.124.412.031.338.840.349.1
BERT-S [63]-39.250.322.8--39.930.148.6
BERT-S++ [60]48.146.756.9---42.344.952.1
TIGER [22]51.449.348.6---47.545.455.3
ViLBERTScore [27]54.250.161.3---51.452.458.7
MID [24]-54.9-37.3-----
RefCLIP-S [18]52.653.064.636.418.844.751.255.466.2
RefPAC-S [47]55.555.967.637.619.546.253.057.368.0
Reference-free metrics
UMIC [26]-46.8-------
CLIP-S (ViT-B/32) [18]51.151.263.034.417.730.549.853.864.3
PAC-S (ViT-B/32) [47]53.954.366.136.018.644.451.555.766.3
BRIDGE (ViT-B/32)54.454.866.736.118.744.550.955.065.4
CLIP-S (ViT-L/14) [18]52.653.064.735.218.243.351.355.465.9
PAC-S (ViT-L/14) [47]55.155.567.336.819.045.352.356.567.1
BRIDGE (ViT-L/14)55.455.867.736.319.044.752.957.267.8
+ +backbone on the Flickr8k-Expert dataset. These improvements extend consistently across all datasets and correlation scores. Compared to PAC-S, BRIDGE still achieves superior results on both Flickr8k-Expert and Flickr8k-CF, while performing on par on the Composite dataset (i.e. PAC-S achieves the best results when using ViT-B/32, while BRIDGE outperforms PAC-S using ViT-L/14). + +Discriminating Correct Captions. We also evaluate the effectiveness of our metric on the PASCAL-50S dataset [55]. In this context, instead of + +calculating correlation scores, we compute accuracy by determining, for each pair, the caption favored by the majority of human ratings as correct (with ties being resolved randomly). We then measure how frequently the evaluation metric assigns a higher score to the chosen caption. Following previous works [18], we randomly select five reference captions from the set of 48 provided by the dataset and average the results over five distinct draws. Accuracy values are reported in Table 4. The results show that when using + +Table 4: Accuracy results on Pascal-50S [55] averaged over five random draws of reference captions. The $\dagger$ marker indicates scores from previous works. + +
HCHIHMMMMean
Reference-based metrics
BLEU-4 [41]60.393.185.757.074.0
METEOR [4]66.097.794.066.681.1
CIDEr [55]66.597.990.765.280.1
BERT-S++†[60]65.498.196.460.380.1
TIGER†[22]56.099.892.874.280.7
MID†[24]67.099.797.476.885.2
RefCLIP-S [18]64.999.595.573.383.3
RefPAC-S [48]67.799.696.075.684.7
Reference-free metrics
CLIP-S (ViT-B/32) [18]55.999.396.572.080.9
PAC-S (ViT-B/32) [47]60.699.396.972.982.4
BRIDGE (ViT-B/32)59.499.497.574.082.6
CLIP-S (ViT-L/14) [18]57.099.696.773.581.7
PAC-S (ViT-L/14) [47]59.599.495.874.782.2
BRIDGE (ViT-L/14)61.299.696.674.182.9
+ +![](images/b8227c3c7cdf1b96f90487fce67a5fa807108b59008efc1a9fd3a6303ba53ff3.jpg) +Fig. 4: Sample images from the FOIL dataset [49] and corresponding scores generated by our proposed metric compared with CLIP-S and PAC-S. + +![](images/b984061a8407b42408d3544b43b77986c02ab8fcce872b39e45f71c29da2d9eb.jpg) + +Table 5: Accuracy results on the FOIL [49] dataset. + +
Acc.
Reference-based metrics
BLEU-4 [41]66.2
METEOR [4]70.1
CIDEr [55]85.7
MID [24]90.5
RefCLIP-S [18]91.0
RefPAC-S [18]93.7
Reference-free metrics
CLIP-S (ViT-B/32) [18]87.2
PAC-S (ViT-B/32) [47]89.9
BRIDGE (ViT-B/32)91.5
CLIP-S (ViT-L/14) [18]90.9
PAC-S (ViT-L/14) [47]91.9
BRIDGE (ViT-L/14)93.0
+ +both ViT-B/32 and ViT-L/14, BRIDGE consistently outperform CLIP-S across all categories, showcasing an average accuracy increase of $+1.7$ and $+1.2$ points, respectively. When comparing with PAC-S, our solution can better discriminate the correct captions on both backbones with an average accuracy increase of $+0.2$ and $+0.7$ using ViT-B/32 and ViT-L/14 respectively. + +Object Hallucination Analysis. We then extend our analysis to the FOIL dataset [49] for correctly identifying captions that may contain object hallucinations. Table 5 shows the accuracy results. As it can be seen, BRIDGE outperforms CLIP-S and PAC-S, exhibiting an increase respectively of $+4.3$ and $+1.6$ points when using ViT-B/32 as backbone. Similar improvements can also be observed when employing more robust visual features. These results demonstrate the capabilities of our metric also to identify hallucinated objects correctly. In Fig 4, we present sample results comparing our metric with CLIP-S and PAC-S. + +# 4.5 System-level Correlation + +Finally, we delve into the efficacy of our proposed metric when evaluating popular existing captioning models. To this aim, we generate predictions of several state-of-the-art captioning models on the COCO test set, including Show and Tell and Show, Attend and Tell which are among the first image captioning models based on deep learning, Up-Down [3], SGAE [59], AoANet [20], $\mathcal{M}^2$ Transformer [14], X-Transformer [40] which all include region-based image features with either LSTM-based or Transformer-based language models, and the recently proposed COS-Net model [32] that incorporates CLIP features. In addition to reporting evaluations on traditional captioning models, we also include recent LLM-based captioning models, including ZeroCap [51] and SmallCap [44], which are based on GPT-2 [43], and MiniGPT-v2 [9], BLIP-2 [28], IDEFICS [25], LLaVA-1.5 [35,36], and InstructBLIP [15] which instead are based on larger-scale LLMs like FlanT5 [12], Vicuna [11], or LLaMA [52,53]. + +Table 6: Evaluation scores of traditional and LLM-based captioners on COCO test set ( $\spadesuit$ : not trained/fine-tuned on COCO). + +
BLEU-4METEORCIDErCLIP-SPAC-SBRIDGE
TraditionalShow and Tell [56]31.425.097.20.5720.7720.788
Show, Attend and Tell [58]33.426.2104.60.5820.7850.804
Up-Down [3]36.727.9122.70.7230.8030.821
SGAE [59]39.028.4129.10.7340.8120.833
AoANet [20]38.929.2129.80.7370.8150.836
M2Transformer [14]39.129.2131.20.7340.8130.841
X-Transformer [40]39.729.5132.80.6100.8120.845
COS-Net [32]42.030.6141.10.7580.8320.859
LLM-basedZeroCap♦[51]2.310.115.10.8100.8160.862
SmallCap [44]37.027.9119.70.7480.8260.847
MiniGPT-v2 [9]18.824.680.40.7520.8180.845
BLIP-2 [28]43.732.0145.80.7670.8370.868
IDEFICS-9B♦[25]4.319.150.00.7400.7860.838
LLaVA-1.5-7B♦[35]8.128.069.60.7840.8090.867
InstructBLIP-Flan-T5-XL♦[15]6.128.138.10.8170.8370.902
Humans-24.187.60.7740.8230.856
+ +The results are presented in Table 6, where we evaluate our BRIDGE scores against both standard metrics, such as BLEU-4, METEOR, and CIDEr, and more recent ones like CLIP-S and PAC-S. Captioning models are compared to a human baseline, in which, for each sample, one human-annotated sentence (selected randomly from the five provided by the COCO dataset) serves as a candidate caption. As shown in the table, BRIDGE can effectively evaluate human-annotated sentences which obtain a score similar to recent state-of-the-art captioning models such as COS-Net. This capability lacks in standard metrics such as METEOR and CIDEr which rank human captions lower than those generated by less-performing captioning models like Show, Attend, and Tell or Up-Down. + +When considering the evaluation of captions generated by existing models, our metric shows a strong correlation with standard evaluation metrics when evaluating traditional image captioners or large-scale models that are fine-tuned on the COCO dataset like BLIP-2. When instead considering more recent approaches that are based on large language models and are not fine-tuned on COCO, BRIDGE still recognizes the goodness of generated captions, raking InstructBLIP as the best-performing approach. This demonstrates the capabilities of our metric to correctly evaluate longer and more detailed captions which are typically generated by LLM-based multimodal models [7, 16, 30] and that might be very different from captions contained in the COCO dataset. + +# 5 Conclusion + +In this paper, we have presented a novel learnable, and reference-free image captioning metric that combines text and dense visual features. Our proposal, BRIDGE, employs templated captions that are enriched with fine-grained visual cues thanks to a mapping network. Through experimental evaluation, we demonstrate that BRIDGE outperforms existing reference-free metrics in terms of correlation with human judgment and sensitivity to hallucinated objects. + +# Acknowledgments + +We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been conducted under a research grant co-funded by Leonardo S.p.A. and supported by the PNRR-M4C2 (PE0000013) project "FAIR - Future Artificial Intelligence Research" and by the PRIN project "MUSMA" (CUP G53D23002930006 - M4C2 I1.1), both funded by EU - Next-Generation EU. + +# References + +1. Aditya, S., Yang, Y., Baral, C., Fermuller, C., Aloimonos, Y.: From Images to Sentences through Scene Description Graphs using Commonsense Reasoning and Knowledge. arXiv preprint arXiv:1511.03292 (2015) +2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: SPICE: Semantic Propositional Image Caption Evaluation. In: ECCV (2016) +3. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR (2018) +4. Banerjee, S., Lavie, A.: METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In: ACL Workshops (2005) +5. Barraco, M., Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: With a Little Help from your own Past: Prototypical Memory Networks for Image Captioning. In: ICCV (2023) +6. Bird, S., Klein, E., Loper, E.: Natural Language Processing with Python: Analyzing Text with the Natural Language Toolkit. O'Reilly Media, Inc. (2009) +7. Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Baraldi, L., Cornia, M., Cucchiara, R.: The Revolution of Multimodal Large Language Models: A Survey. In: ACL Findings (2024) +8. Chan, D., Petryk, S., Gonzalez, J.E., Darrell, T., Canny, J.: CLAIR: Evaluating Image Captions with Large Language Models. In: EMNLP (2023) +9. Chen, J., Li, D.Z.X.S.X., Zhang, Z.L.P., Xiong, R.K.V.C.Y., Elhoseiny, M.: MiniGPT-v2: Large Language Model as a Unified Interface for Vision-Language Multi-task Learning. arXiv preprint arXiv:2310.09478 (2023) +0. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: UNITER: Universal Image-TExt Representation Learning. In: ECCV (2020) +1. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An Open-Source Chatbot Improressing GPT-4 with $90\%$ * ChatGPT Quality (2023) +12. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S.S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Castro-Ros, A., Pellat, M., Robinson, K., Valter, D., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E.H., Dean, J., Devlin, J., Roberts, A., Zhou, D., Le, Q.V., Wei, J.: Scaling Instruction-Finetuned Language Models. JMLR 25(70), 1-53 (2024) +13. Cornia, M., Baraldi, L., Cucchiara, R.: SMArT: Training Shallow Memory-aware Transformers for Robotic Explainability. In: ICRA (2020) +14. Cornia, M., Stefanini, M., Baraldi, L., Cucchiara, R.: Meshed-Memory Transformer for Image Captioning. In: CVPR (2020) + +15. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning. arXiv preprint arXiv:2305.06500 (2023) +16. Dong, H., Li, J., Wu, B., Wang, J., Zhang, Y., Guo, H.: Benchmarking and Improving Detail Image Caption. arXiv preprint arXiv:2405.19092 (2024) +17. Herdade, S., Kappeler, A., Boakye, K., Soares, J.: Image captioning: Transforming objects into words. In: NeurIPS (2019) +18. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In: EMNLP (2021) +19. Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking task: Data, models and evaluation metrics. JAIR 47, 853-899 (2013) +20. Huang, L., Wang, W., Chen, J., Wei, X.Y.: Attention on Attention for Image Captioning. In: ICCV (2019) +21. Jiang, M., Hu, J., Huang, Q., Zhang, L., Diesner, J., Gao, J.: REO-Relevance, Extraness, Omission: A Fine-grained Evaluation for Image Captioning. In: EMNLP (2019) +22. Jiang, M., Huang, Q., Zhang, L., Wang, X., Zhang, P., Gan, Z., Diesner, J., Gao, J.: TIGER: Text-to-Image Grounding for Image Caption Evaluation. In: EMNLP (2019) +23. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015) +24. Kim, J.H., Kim, Y., Lee, J., Yoo, K.M., Lee, S.W.: Mutual Information Divergence: A Unified Metric for Multimodal Generative Models. In: NeurIPS (2022) +25. Laurencon, H., Saulnier, L., Tronchon, L., Bekman, S., Singh, A., Lozhkov, A., Wang, T., Karamcheti, S., Rush, A.M., Kiela, D., Cord, M., Sanh, V.: OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents. In: NeurIPS (2023) +26. Lee, H., Yoon, S., Dernoncourt, F., Bui, T., Jung, K.: UMIC: An Unreferenced Metric for Image Captioning via Contrastive Learning. In: ACL (2021) +27. Lee, H., Yoon, S., Dernoncourt, F., Kim, D.S., Bui, T., Jung, K.: ViLBERTScore: Evaluating Image Caption Using Vision-and-Language BERT. In: EMNLP Workshops (2020) +28. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping Language-Image Pretraining with Frozen Image Encoders and Large Language Models. In: ICML (2023) +29. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In: ICML (2022) +30. Li, X., Tu, H., Hui, M., Wang, Z., Zhao, B., Xiao, J., Ren, S., Mei, J., Liu, Q., Zheng, H., et al.: What If We Recap tion Billions of Web Images with LLaMA-3? arXiv preprint arXiv:2406.08478 (2024) +31. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: ECCV (2020) +32. Li, Y., Pan, Y., Yao, T., Mei, T.: Comprehending and ordering semantics for image captioning. In: CVPR (2022) +33. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL Workshops (2004) +34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: ECCV (2014) +35. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved Baselines with Visual Instruction Tuning. In: CVPR (2024) + +36. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual Instruction Tuning. In: NeurIPS (2023) +37. Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization. In: ICLR (2019) +38. Lu, J., Batra, D., Parikh, D., Lee, S.: ViLBERT: Pretraining Task-Agnostic Visi-olinguistic Representations for Vision-and-Language Tasks. In: NeurIPS (2019) +39. Oord, A.v.d., Li, Y., Vinyals, O.: Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748 (2018) +40. Pan, Y., Yao, T., Li, Y., Mei, T.: X-Linear Attention Networks for Image Captioning. In: CVPR (2020) +41. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: BLEU: a method for automatic evaluation of machine translation. In: ACL (2002) +42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transferable Visual Models From Natural Language Supervision. In: ICML (2021) +43. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners. OpenAI Blog 1(8), 9 (2019) +44. Ramos, R., Martins, B., Elliott, D., Kementchedjhieva, Y.: SmallCap: lightweight image captioning prompted with retrieval augmentation. In: CVPR (2023) +45. Rashtchian, C., Young, P., Hodosh, M., Hockenmaier, J.: Collecting Image Annotations Using Amazon's Mechanical Turk. In: NAACL Workshops (2010) +46. Saito, K., Sohn, K., Zhang, X., Li, C.L., Lee, C.Y., Saenko, K., Pfister, T.: Pic2Word: Mapping Pictures to Words for Zero-Shot Composed Image Retrieval. In: CVPR (2023) +47. Sarto, S., Barraco, M., Cornia, M., Baraldi, L., Cucchiara, R.: Positive-Augmented Contrastive Learning for Image and Video Captioning Evaluation. In: CVPR (2023) +48. Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: Retrieval-Augmented Transformer for Image Captioning. In: CBMI (2022) +49. Shekhar, R., Pezzelle, S., Klimovich, Y., Herbelot, A., Nabi, M., Sangineto, E., Bernardi, R.: FOIL it! Find One mismatch between Image and Language caption. In: ACL (2017) +50. Shi, Y., Yang, X., Xu, H., Yuan, C., Li, B., Hu, W., Zha, Z.J.: EMScore: Evaluating Video Captioning via Coarse-Grained and Fine-Grained Embedding Matching. In: CVPR (2022) +51. Tewel, Y., Shalev, Y., Schwartz, I., Wolf, L.: Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In: CVPR (2022) +52. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023) +53. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +55. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: CIDEr: Consensus-based Image Description Evaluation. In: CVPR (2015) +56. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: CVPR (2015) +57. Wang, S., Yao, Z., Wang, R., Wu, Z., Chen, X.: FAIER: Fidelity and Adequacy Ensured Image Caption Evaluation. In: CVPR (2021) + +58. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhutdinov, R., Zemel, R.S., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015) +59. Yang, X., Tang, K., Zhang, H., Cai, J.: Auto-Encoding Scene Graphs for Image Captioning. In: CVPR (2019) +60. Yi, Y., Deng, H., Hu, J.: Improving Image Captioning Evaluation by Considered Inter Refinences Variance. In: ACL (2020) +61. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL 2, 67-78 (2014) +62. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: VinVL: Revisiting visual representations in vision-language models. In: CVPR (2021) +63. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: BERTScore: Evaluating Text Generation with BERT. In: ICLR (2020) +64. Zhu, W., Wang, X.E., Yan, A., Eckstein, M., Wang, W.Y.: ImaginE: An Imagination-Based Automatic Evaluation Metric for Natural Language Generation. In: EACL (2023) \ No newline at end of file diff --git a/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/images.zip b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f9677afd7d44b0d68eee8ee25067ad62248b1079 --- /dev/null +++ b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3671d8dda9025bd317df50d2491cc18f4d33ab6aeefbc8a7f84178553a4bc62 +size 486702 diff --git a/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/layout.json b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..42a29f035cf6f87d33c041bde57f85720f0df6d3 --- /dev/null +++ b/2024/BRIDGE_ Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues/layout.json @@ -0,0 +1,10519 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "text", + "content": "BRIDGE: Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 167, + 478, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 167, + 478, + 179 + ], + "spans": [ + { + "bbox": [ + 134, + 167, + 478, + 179 + ], + "type": "text", + "content": "Sara Sarto1, Marcella Cornia1, Lorenzo Baraldi1, and Rita Cucchiara1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 206, + 189, + 409, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 189, + 409, + 201 + ], + "spans": [ + { + "bbox": [ + 206, + 189, + 409, + 201 + ], + "type": "text", + "content": "1 University of Modena and Reggio Emilia, Italy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 270, + 201, + 342, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 201, + 342, + 212 + ], + "spans": [ + { + "bbox": [ + 270, + 201, + 342, + 212 + ], + "type": "text", + "content": "2 IIT-CNR, Italy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 247, + 213, + 366, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 213, + 366, + 222 + ], + "spans": [ + { + "bbox": [ + 247, + 213, + 366, + 222 + ], + "type": "text", + "content": "{name.surname}@unimore.it" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 249, + 455, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 249, + 455, + 437 + ], + "spans": [ + { + "bbox": [ + 159, + 249, + 455, + 437 + ], + "type": "text", + "content": "Abstract. Effectively aligning with human judgment when evaluating machine-generated image captions represents a complex yet intriguing challenge. Existing evaluation metrics like CIDEr or CLIP-Score fall short in this regard as they do not take into account the corresponding image or lack the capability of encoding fine-grained details and penalizing hallucinations. To overcome these issues, in this paper, we propose BRIDGE, a new learnable and reference-free image captioning metric that employs a novel module to map visual features into dense vectors and integrates them into multi-modal pseudo-captions which are built during the evaluation process. This approach results in a multimodal metric that properly incorporates information from the input image without relying on reference captions, bridging the gap between human judgment and machine-generated image captions. Experiments spanning several datasets demonstrate that our proposal achieves state-of-the-art results compared to existing reference-free evaluation scores. Our source code and trained models are publicly available at: https://github.com/aimagelab/bridge-score" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 446, + 402, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 446, + 402, + 458 + ], + "spans": [ + { + "bbox": [ + 160, + 446, + 402, + 458 + ], + "type": "text", + "content": "Keywords: Captioning Evaluation " + }, + { + "bbox": [ + 160, + 446, + 402, + 458 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 446, + 402, + 458 + ], + "type": "text", + "content": " Vision-and-Language" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 476, + 230, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 230, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 230, + 489 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": "The objective of image captioning is to produce natural language descriptions conditioned on input images, that closely resemble human language and align to human intentions. As such, the captioning task involves the recognition and understanding of the visual content of the image, including fine-grained elements such as objects, attributes, and their relationships. Advances in training methodologies and architectures have contributed to the progress in the field, significantly improving the generation quality. Recent innovations include fully-attentive models [13, 14, 17], improved connections between visual and textual modalities [14, 40], and the incorporation of objects and tags at an architectural level [3, 31, 62]. Additionally, there has been a notable focus on increasing the robustness of cross-modal features [5, 32, 48], which consequently can increase description accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "As constant improvements are made on the generation side, it becomes crucial to enhance the evaluation process as well. In this regard, image captioning" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 479, + 173 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 479, + 173 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 479, + 173 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 479, + 173 + ], + "type": "image", + "image_path": "f147fd6145523e55daa46720fb0e57567e8ca3669b2da5b0801f9c62bd99591c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 178, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 178, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 480, + 236 + ], + "type": "text", + "content": "Fig. 1: Comparison between different captioning evaluation approaches: (a) CIDEr [55] scores candidate and reference captions without considering the input image; (b) CLIP-Score [18] compares text and images using global vectors in a shared embedding space; (c) our BRIDGE, internally builds multimodal pseudo-captions by translating fine-grained image features into pseudo-tokens thanks to a mapping module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 365 + ], + "type": "text", + "content": "evaluation aims to assess the quality of a generated caption given an image and, potentially, human-written reference captions. However, it is important to note that obtaining these reference captions can often be challenging and expensive, adding complexity to the evaluation process. Despite the recent advancements in captioning capabilities, standard automatic evaluation metrics have mainly relied on translation metrics [4,33,41] or text-only ones [2,55,63] which often fall short in capturing aspects such as grammatical correctness, semantic relevance, and specificity. These limitations are worsened by the limited coverage of image content in available references, resulting in inaccurate penalties when generated captions accurately describe novel elements not mentioned in the references." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 437 + ], + "type": "text", + "content": "In response to these limitations, advanced metrics aligning visual and textual data have emerged [18,24,26,27]. Notably, recent metrics leverage the CLIP embedding space [42], which shows a strong correlation with human judgment. Despite the effectiveness of contrastive-based embedding spaces, metrics based on dual-encoder architectures tend to focus on global alignment between an image and its caption, often ignoring fine-grained details or penalizing hallucinations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 481, + 569 + ], + "type": "text", + "content": "Following this insight, in this paper, we introduce BRIDGE, a novel learnable and reference-free image captioning metric that enhances the alignment of more fine-grained visual features. Specifically, our model provides a pre-trained dual-encoder architecture with a mapping module designed to effectively exploit visual cues (Fig. 1). This is done by internally creating multimodal pseudo-captions, containing both textual and dense visual features. The process for building these pseudo-captions involves the creation of a template caption, which focuses on the syntactical structure of the scene, and a mapping module. The latter refines the template caption by enriching it with more fine-grained visual features about the subjects depicted in the image. Subsequently, the overall model is trained with a combination of contrastive losses which promote multimodal alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "Experiments are conducted on a variety of datasets for image captioning annotated with human rankings, including Flickr8k-Expert, Flickr8k-CF [19], Composite [1], and Pascal-50S [55]. Through comprehensive analysis, we demonstrate the efficacy of the proposed metric and show its ability to overcome the limitations of existing state-of-the-art reference-free alternatives. Additionally, we evaluate its sensitivity to object hallucination by conducting experiments on the FOIL dataset [49]. Overall, BRIDGE outperforms previous metrics and showcases superior performance compared to CLIP-Score [18] and PAC-Score [47]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Sarto et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 408, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 408, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 408, + 127 + ], + "type": "text", + "content": "Contributions. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 129, + 479, + 257 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 137, + 129, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 129, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 137, + 129, + 479, + 186 + ], + "type": "text", + "content": "- We tackle the limitations of existing image captioning metrics by proposing the first learnable reference-free metric, termed as BRIDGE, that focuses on more fine-grained visual features. Our proposal integrates a dual-encoder architecture with a mapping module which is in charge of producing multimodal pseudo-captions that combine text and richer dense visual features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 188, + 479, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 479, + 222 + ], + "type": "text", + "content": "- Multimodal pseudo-captions are built by creating template captions that describe the scene from a syntactical point of view. These templates are subsequently enriched with fine-grained features through a mapping network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 224, + 479, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 224, + 479, + 257 + ], + "spans": [ + { + "bbox": [ + 137, + 224, + 479, + 257 + ], + "type": "text", + "content": "- Experiments, carried out on different datasets with human preferences, demonstrate a higher degree of correlation with respect to existing metrics. We also evaluate the sensitivity of the proposal to objects hallucinations." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 279, + 237, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 237, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 237, + 292 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 302, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 445 + ], + "type": "text", + "content": "Classical Reference-based Captioning Metrics. Several widely used captioning evaluation metrics were originally developed in the context of NLP tasks and rely on n-gram matching techniques. Among these classical metrics, BLEU [41] is designed to focus on precision and incorporates a penalty for sentence brevity. METEOR [4], instead, combines precision and recall to evaluate the quality of captions, while others, such as ROUGE [33], were initially born for summarization tasks and later adapted to image captioning. More recently, two metrics tailored for visual captioning have emerged: CIDEr [55], which measures n-gram similarity and is based on TF-IDF, and SPICE [2], which quantifies graph-based similarity through scene graphs constructed from candidate and reference captions. Overall, focusing on textual-level comparisons, these metrics assume that human-written references accurately represent the image content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 480, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 509 + ], + "type": "text", + "content": "Learnable Captioning Metrics. With the rise of large pre-trained models, image captioning evaluation now frequently exploits these models to compare textual-only [60, 63] or visual-textual [18, 21, 22, 26, 27, 57] contents. Notably, the BERT score and its improved version use pre-trained BERT embeddings to compare word tokens in generated and ground-truth sentences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 480, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 592 + ], + "type": "text", + "content": "Some metrics, like BLEU and CIDEr, rely solely on text matching between reference captions and machine-generated captions, potentially introducing bias in evaluations due to non-accurate reference captions. To mitigate these issues, alternative solutions leverage the multimodal nature of vision-and-language models. As an example, TIGEr [22] considers the similarities between words and image regions and assesses how well machine-generated captions represent image content and their alignment with human-generated captions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "In contrast, other approaches [18, 24, 27] leverage web-scale vision-and-language models such as VilBERT [38] and CLIP [42] for more robust metrics. For example, in [24], CLIP visual-textual features are used to compute negative Gaussian cross-mutual information. Other works, instead, have employed diffusion models in text-only tasks [64] or exploited the zero-shot language modeling capabilities of large language models [8] to evaluate candidate captions." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": "Reference-free Captioning Metrics. While all aforementioned metrics rely on a set of ground-truth captions to compute the final score, a few attempts have been made to introduce reference-free evaluations, only taking into account the correlation of the candidate caption with the image. In this regard, Lee et al. [26] proposed to fine-tune the UNITER model [10] via contrastive learning to let the model to discriminate between positive and synthetically-generated negative captions. On a different line, Hessel et al. [18] introduced the CLIP-Score, which only relies on a modified cosine similarity between image and candidate caption representations coming from the CLIP model. The recently proposed PAC-Score [47], instead, builds upon the usage of CLIP but incorporates a fine-tuning phase with positive augmentation, further enhancing the accuracy of evaluation. In this paper, we follow this research path and propose a novel learnable, and reference-free evaluation metric that can effectively incorporate fine-grained visual features for evaluating the correlation of image-text pairs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 304, + 367, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 367, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 367, + 319 + ], + "type": "text", + "content": "3 BRIDGE for Captioning Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 328, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 401 + ], + "type": "text", + "content": "In the following, we present our reference-free captioning evaluation approach, termed BRIDGE. Our approach leverages a dual-encoder architecture, which comprises both a language encoder and a vision encoder. Given a frozen pretrained model, we train a mapping module responsible for filling the holes of a masked template caption with pseudo language tokens that are enriched with visual information. An overview of our model is depicted in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 420, + 228, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 228, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 228, + 431 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "content": "Our approach relies on CLIP (Contrastive Language-Image Pre-training) [42], a powerful vision and language model designed to align images and corresponding text captions within a shared embedding space. For a given input image " + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "content": ", the image encoder " + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "inline_equation", + "content": "E_V" + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "content": " extracts the visual information " + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "inline_equation", + "content": "v = E_V(I) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "content": ". On the textual side, an input caption " + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "content": " is tokenized and a textual representation is obtained by passing it through the textual encoder " + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "inline_equation", + "content": "E_T" + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "content": ", obtaining " + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "inline_equation", + "content": "t = E_T(T) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "content": ". Once the textual and visual features, " + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 441, + 482, + 537 + ], + "type": "text", + "content": " respectively, are projected in a common space, visual and textual inputs can be compared via cosine similarity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 538, + 482, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 621 + ], + "type": "text", + "content": "The relationships learned by CLIP can be exploited to build an image captioning evaluator. In CLIP-Score [18] the authors directly compare candidate captions and images in the embedding space and show that this achieves a good correlation with human judgments. In detail, to assess the quality of a candidate generation, they feed both the image and the candidate caption through their respective feature extractors, and compute the cosine similarity of the resultant embeddings:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 622, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 622, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 219, + 622, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\operatorname {C L I P} \\text {- S c o r e} (I, T) = w \\cdot \\max (\\cos (v, t), 0), \\tag {1}", + "image_path": "3ae0b4977db8b25b39a5e40048e34af84a0e276aaaeffa9b68cd983bbe2d27f9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " is a rescaling factor employed to stretch the score distribution while ensuring the ranking results remain unchanged." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Sarto et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 479, + 271 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 271 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 271 + ], + "type": "image", + "image_path": "3c2d21a265ec7cb0e78f844dea2608c71b92de6044e3a19960a8c5df6c8fbaee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 279, + 480, + 313 + ], + "lines": [ + { + "bbox": [ + 130, + 279, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 480, + 313 + ], + "type": "text", + "content": "Fig. 2: Overview of the BRIDGE evaluation approach. Starting from a template caption, a mapping network augments it with dense visual features, obtaining a pseudocaption which is then used for computing image-text similarities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 324, + 358, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 324, + 358, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 324, + 358, + 337 + ], + "type": "text", + "content": "3.2 Injecting Fine-Grained Visual Features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 402 + ], + "type": "text", + "content": "Unlike CLIP-Score, our approach does not rely exclusively on global image descriptors for evaluating image-text alignments. Instead, we focus on employing stronger visual information. To do so, we draw inspiration from the Pic2Word approach [46] and represent the input image through a multimodal pseudo-caption, an embedding representation that contains stronger visual elements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 405, + 481, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 481, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 481, + 560 + ], + "type": "text", + "content": "Building Template Captions. In order to create multimodal pseudo-captions, we first build template captions for a given input image. These are skeletal textual representations of the image, obtained by masking out all the relevant textual concepts from the descriptions generated by a captioner. Through these template captions, we aim to provide the model only with a templated textual structure which can then be filled with more fine-grained visual features. In particular, given an automatically generated caption describing the input image, such as 'A man running with a white dog', we remove the main subjects within the caption (e.g. 'man' and 'white dog') and mask them with a [MASK] token. This will allow the model to fill in these gaps by incorporating more fine-grained features from the image encoder. Since a primary subject might be described by words other than just its corresponding nouns (e.g. adjectives), we utilize noun chunks. Fig. 3 reports template captions and corresponding noun chunks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 560, + 481, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 481, + 620 + ], + "type": "text", + "content": "Given a sentence containing " + }, + { + "bbox": [ + 130, + 560, + 481, + 620 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 560, + 481, + 620 + ], + "type": "text", + "content": " noun chunks, we independently encode them through the mapping network. To this aim, we replicate the template caption as many times as the number of noun chunks and mask a different noun chunk in each of the replicas. We thus obtain " + }, + { + "bbox": [ + 130, + 560, + 481, + 620 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 560, + 481, + 620 + ], + "type": "text", + "content": " different versions of the template caption, each one masking a single noun chunk, for instance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 210, + 628, + 403, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 628, + 403, + 639 + ], + "spans": [ + { + "bbox": [ + 210, + 628, + 403, + 639 + ], + "type": "text", + "content": "[A [MASK] running with a white dog'," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 225, + 643, + 387, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 643, + 387, + 654 + ], + "spans": [ + { + "bbox": [ + 225, + 643, + 387, + 654 + ], + "type": "text", + "content": "'A man running with a [MASK]'." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 448, + 102 + ], + "type": "text", + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 194, + 163 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 194, + 163 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 194, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 194, + 163 + ], + "type": "image", + "image_path": "e9b1d339c07d6c2584bb07f66cbc26842c5bb1254eaf9bbe4b187ed5e3d5e3dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 118, + 218, + 127 + ], + "lines": [ + { + "bbox": [ + 196, + 118, + 218, + 127 + ], + "spans": [ + { + "bbox": [ + 196, + 118, + 218, + 127 + ], + "type": "text", + "content": "COCO" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 137, + 222, + 144 + ], + "lines": [ + { + "bbox": [ + 196, + 137, + 222, + 144 + ], + "spans": [ + { + "bbox": [ + 196, + 137, + 222, + 144 + ], + "type": "text", + "content": "TEMPLATE" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 198, + 151, + 220, + 163 + ], + "lines": [ + { + "bbox": [ + 198, + 151, + 220, + 163 + ], + "spans": [ + { + "bbox": [ + 198, + 151, + 220, + 163 + ], + "type": "text", + "content": "NOUN CHUNKS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 151, + 257, + 163 + ], + "lines": [ + { + "bbox": [ + 230, + 151, + 257, + 163 + ], + "spans": [ + { + "bbox": [ + 230, + 151, + 257, + 163 + ], + "type": "text", + "content": "yellow bus" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 170, + 462, + 182 + ], + "lines": [ + { + "bbox": [ + 149, + 170, + 462, + 182 + ], + "spans": [ + { + "bbox": [ + 149, + 170, + 462, + 182 + ], + "type": "text", + "content": "Fig. 3: COCO captions with template captions and associated noun chunks." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 367, + 163 + ], + "blocks": [ + { + "bbox": [ + 232, + 118, + 299, + 127 + ], + "lines": [ + { + "bbox": [ + 232, + 118, + 299, + 127 + ], + "spans": [ + { + "bbox": [ + 232, + 118, + 299, + 127 + ], + "type": "text", + "content": "People on a motorcycle in front of a road" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 133, + 302, + 148 + ], + "lines": [ + { + "bbox": [ + 230, + 133, + 302, + 148 + ], + "spans": [ + { + "bbox": [ + 230, + 133, + 302, + 148 + ], + "type": "text", + "content": "A yellow bus driving down a street next to a park." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 257, + 151, + 274, + 162 + ], + "lines": [ + { + "bbox": [ + 257, + 151, + 274, + 162 + ], + "spans": [ + { + "bbox": [ + 257, + 151, + 274, + 162 + ], + "type": "text", + "content": "street" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 287, + 152, + 298, + 162 + ], + "lines": [ + { + "bbox": [ + 287, + 152, + 298, + 162 + ], + "spans": [ + { + "bbox": [ + 287, + 152, + 298, + 162 + ], + "type": "text", + "content": "park" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 114, + 367, + 163 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 367, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 367, + 163 + ], + "type": "image", + "image_path": "a46b7b22878b3171115f7ebea4c2888bb32cfcce876bfbac649c85e857d428d0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 119, + 390, + 128 + ], + "lines": [ + { + "bbox": [ + 372, + 119, + 390, + 128 + ], + "spans": [ + { + "bbox": [ + 372, + 119, + 390, + 128 + ], + "type": "text", + "content": "COCO" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 137, + 394, + 145 + ], + "lines": [ + { + "bbox": [ + 370, + 137, + 394, + 145 + ], + "spans": [ + { + "bbox": [ + 370, + 137, + 394, + 145 + ], + "type": "text", + "content": "TEMPLATE" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 150, + 392, + 163 + ], + "lines": [ + { + "bbox": [ + 371, + 150, + 392, + 163 + ], + "spans": [ + { + "bbox": [ + 371, + 150, + 392, + 163 + ], + "type": "text", + "content": "NOUN CHUNKS" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 118, + 474, + 132 + ], + "lines": [ + { + "bbox": [ + 403, + 118, + 474, + 132 + ], + "spans": [ + { + "bbox": [ + 403, + 118, + 474, + 132 + ], + "type": "text", + "content": "Little dog laying on a pillow next to the laptop." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 134, + 476, + 148 + ], + "lines": [ + { + "bbox": [ + 400, + 134, + 476, + 148 + ], + "spans": [ + { + "bbox": [ + 400, + 134, + 476, + 148 + ], + "type": "text", + "content": "A dog laying on a couch next to a PC." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 152, + 416, + 161 + ], + "lines": [ + { + "bbox": [ + 406, + 152, + 416, + 161 + ], + "spans": [ + { + "bbox": [ + 406, + 152, + 416, + 161 + ], + "type": "text", + "content": "PC" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 152, + 443, + 161 + ], + "lines": [ + { + "bbox": [ + 432, + 152, + 443, + 161 + ], + "spans": [ + { + "bbox": [ + 432, + 152, + 443, + 161 + ], + "type": "text", + "content": "dog" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 152, + 474, + 161 + ], + "lines": [ + { + "bbox": [ + 457, + 152, + 474, + 161 + ], + "spans": [ + { + "bbox": [ + 457, + 152, + 474, + 161 + ], + "type": "text", + "content": "couch" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 151, + 392, + 163 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 151, + 392, + 163 + ], + "spans": [ + { + "bbox": [ + 372, + 151, + 392, + 163 + ], + "type": "text", + "content": "NOUN CHUNKS" + } + ] + } + ], + "index": 17, + "type": "text" + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "text", + "content": "Mapping with Fine-Grained Visual Features. The above-described masked replicas are then fed to the mapping module " + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "text", + "content": ". Specifically, our approach exploits the visual information extracted from the visual encoder " + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "inline_equation", + "content": "E_V" + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "text", + "content": " to enrich the replicas with the informative content of the image " + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "text", + "content": ". To get a more fine-grained representation of the image, we directly take the grid of features from the last layer, " + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "inline_equation", + "content": "\\widehat{v}" + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "text", + "content": ". For instance, in the case of a ViT-B/32 backbone, this will have a shape of " + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "inline_equation", + "content": "50 \\times d" + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 194, + 481, + 278 + ], + "type": "text", + "content": " is the dimensionality of the last embedding of the network." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 278, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 481, + 361 + ], + "type": "text", + "content": "The mapping network is implemented as a stack of Transformer [54] encoder layers interleaved with cross-attention layers. Its role is to refine each template captions with visual information. Since each template caption is processed independently, the mapping module returns a set of sequences, each with the same length as the corresponding input template caption. From the output of the mapping module, we keep only the predictions for the masked tokens in each template caption and copy them back into the original templates." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "text", + "content": "Therefore, by providing a masked input template in the form " + }, + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "inline_equation", + "content": "\\hat{T_i} = [w_1,\\dots,w_{j - 1},\\mathsf{MASK},w_{j + 1},\\dots,w_T]" + }, + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "inline_equation", + "content": "\\{w_j\\}_j" + }, + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "text", + "content": " represent original tokens from the input caption, we obtain " + }, + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "inline_equation", + "content": "\\hat{T_i}^* = \\left[w_1,\\dots,w_{j - 1},\\psi (\\hat{T_i})_j,w_{j + 1},\\dots,w_T\\right]" + }, + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "inline_equation", + "content": "\\psi (\\hat{T_i},\\widehat{v})_j" + }, + { + "bbox": [ + 130, + 362, + 482, + 526 + ], + "type": "text", + "content": " represents the output of the mapping network at the position corresponding to the masked input token position. In the case of noun chunks consisting of more than one token, multiple consecutive tokens are replaced with the corresponding outputs from the mapping network. By injecting the outputs of the mapping token into the initial template caption, we effectively complete the original templates with visually enriched vectors. Notably, these newly generated pseudo-captions combine word sequences from the template captions with dense vectors obtained by the mapping module. Consequently, they cannot be decoded as standard captions. As a last step, the obtained pseudo-captions are fed into the pre-trained CLIP language encoder." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 547, + 250, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 250, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 250, + 559 + ], + "type": "text", + "content": "3.3 Training Protocol" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 569, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 604 + ], + "type": "text", + "content": "To train our mapping network, the loss is defined as a weighted version of the symmetric InfoNCE loss [39], where positive and negative items are weighted according to the number of noun chunks in each caption." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Specifically, given a batch in the form " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{(I_i,T_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " represent image-caption pairs, each image " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " is expanded in " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " multimodal pseudocaptions as outlined above, where " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " is the number of noun chunks in caption " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": ". Further, let " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "t_{ij}^{*}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " represent the embedding vector of the " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "-th pseudo-caption derived from the " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "-th image, " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " the embedding vector of the " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "-th image and " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "t_{i}" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Sarto et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "the embedding vector of the " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "-th ground-truth caption. Finally, let " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " be the total number of noun chunks in the mini-batch, i.e. " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "M = \\sum_{i=1}^{N} N_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "text", + "content": "The first loss, denoted as " + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "text", + "content": ", tries to align the pseudo-captions " + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "inline_equation", + "content": "t_{ij}^*" + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "text", + "content": " with the global visual features of the corresponding images " + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "text", + "content": ". In addition to this loss term, we define a second loss component " + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "text", + "content": " that promotes the alignment between pseudo-captions " + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "inline_equation", + "content": "t_{ij}^*" + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "text", + "content": " and the textual feature vector of the ground-truth caption " + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 142, + 480, + 226 + ], + "type": "text", + "content": " corresponding to the input image. This ensures that pseduo-captions are aligned also on a textual space, in addition to being aligned in the image space. Additional details can be found in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 479, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 479, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 479, + 302 + ], + "type": "text", + "content": "Regularization Branch. With the aforementioned loss terms, our objective is encouraging an association between each pseudo-caption and its corresponding image and caption. However, it is also important to differentiate each pseudocaption from the others of the same image. To achieve this, we define a regularization loss which promotes a precise alignment between each pseudo-caption and the corresponding noun chunk." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "type": "text", + "content": "First, we create prompts like \"a photo of a \" and encode them with the text encoder " + }, + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "type": "text", + "content": ". In parallel, each pseudo-caption is fed into a dedicated multilayer perceptron (MLP) projection, which consists of two linear layers with a ReLU activation in between. Formally, the branch is defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 365, + 479, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 365, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 233, + 365, + 479, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {C} (x) = \\operatorname {L i n e a r} (\\operatorname {R e L U} (\\operatorname {L i n e a r} (x))). \\tag {2}", + "image_path": "1bec8a52c5dc39a6e6f9f6268f6baf11e460c566842bc972b29699caae32bd7d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "text", + "content": "Since our goal is to emphasize each pseudo-caption's alignment with its corresponding noun chunk, we employ a regular contrastive loss " + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_3" + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "text", + "content": " between the prompts mentioned earlier and the outputs of the projection branch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 449 + ], + "type": "text", + "content": "Finally, the overall loss function we train BRIDGE is defined as a weighted summation of two aforementioned losses, plus the regularization loss, as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 251, + 462, + 479, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 462, + 479, + 473 + ], + "spans": [ + { + "bbox": [ + 251, + 462, + 479, + 473 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {2} \\mathcal {L} _ {2} + \\lambda_ {3} \\mathcal {L} _ {r}. \\tag {3}", + "image_path": "088796290f6b323500049879044946f3084570bc495ea2bbb8c116edf7dea8cb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 496, + 330, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 330, + 508 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 330, + 508 + ], + "type": "text", + "content": "3.4 Inference and Score Computation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "content": "At inference time, given an image-candidate caption pair " + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "inline_equation", + "content": "(I,T)" + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "content": ", we extract all pseudo-captions from " + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "content": " using our mapping network. Subsequently, we compute the mean pseudo-caption embedding as " + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "inline_equation", + "content": "t^{*} = \\frac{1}{N}\\hat{t}_{i}^{*}" + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "inline_equation", + "content": "\\hat{t}_i^*" + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "content": "-th pseudo-caption extracted from " + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "content": " here indicates the overall number of pseudo-captions associated with " + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 521, + 480, + 580 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": "At that point, given the visual embedding " + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": " of the image and the embedding of the candidate caption " + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": ", the matching score between " + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 167, + 630, + 479, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 630, + 479, + 642 + ], + "spans": [ + { + "bbox": [ + 167, + 630, + 479, + 642 + ], + "type": "interline_equation", + "content": "\\operatorname {B R I D G E} (I, T) = 0. 5 \\cdot [ \\operatorname {C L I P - S c o r e} (I, T) + w \\cdot \\max (\\cos \\left(t ^ {*}, t\\right), 0) ], \\tag {4}", + "image_path": "59f5e5c2bba8adb66dda69b1c74715e135e2c695d1b96c469aa54c6d08ea5114.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 653, + 460, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 460, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 460, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 653, + 460, + 665 + ], + "type": "inline_equation", + "content": "\\cos" + }, + { + "bbox": [ + 131, + 653, + 460, + 665 + ], + "type": "text", + "content": " indicates the cosine similarity and " + }, + { + "bbox": [ + 131, + 653, + 460, + 665 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 131, + 653, + 460, + 665 + ], + "type": "text", + "content": " is a constant scaling factor." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 448, + 102 + ], + "type": "text", + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 129 + ], + "type": "text", + "content": "4 Experimental Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 279, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 279, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 279, + 148 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 156, + 481, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 481, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 481, + 263 + ], + "type": "text", + "content": "Architecture and Training Details. Building upon prior research [18,24,50], we use either CLIP [42] ViT-B/32 or ViT-L/14 as backbone for the visual and textual encoder. The mapping module is composed of two Transformer layers and is trained on the COCO dataset [34], which contains more than 120k images annotated with five captions. In particular, we employ the splits introduced by Karpathy et al. [23], where 5,000 images are used for both validation and testing and the rest for training. To map the grid visual features to an embedding space of dimension 512, we employ a simple linear projection. For the regularization branch, we utilize a two-layer multi-layer perceptron." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "text", + "content": "During training, we use AdamW [37] as optimizer with a learning rate equal to 0.0001 and a batch size of 256. The " + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "text", + "content": " values are selected with a grid search, choosing the combination that provides the best validation loss. Specifically, we set both " + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "text", + "content": " to 0.01, while " + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 130, + 264, + 481, + 323 + ], + "type": "text", + "content": " is set to 1.0. The training stage lasts around one day on a single A100 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 327, + 481, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 481, + 410 + ], + "type": "text", + "content": "Template Caption Generation. The template captions used as input for the mapping module are generated using the BLIP model [29]. In particular, we use the ViT-L/14 version pre-trained on 129M image-text pairs and finetuned on the COCO dataset. After this generation phase, the primary subjects of the template sentences are extracted by using the NLTK library [6]. During training, two noun chunks are randomly chosen from the set identified during the extraction step. In the evaluation phase, otherwise, all identified noun chunks are included." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 428, + 204, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 204, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 204, + 438 + ], + "type": "text", + "content": "4.2 Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 447, + 481, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 481, + 519 + ], + "type": "text", + "content": "To evaluate the correlation of the proposed metric with human ratings, we conduct experiments on the Flickr8k-Expert, Flickr8k-CF, Composite, and Pascal50-S datasets [1,19,55]. In addition, for detecting hallucinations in textual sentences, we extend our analysis to the FOIL dataset [49]. Except for Pascal50S and FOIL in where accuracy scores are used, evaluation on all other datasets relies on Kendall " + }, + { + "bbox": [ + 130, + 447, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\tau_{b}" + }, + { + "bbox": [ + 130, + 447, + 481, + 519 + ], + "type": "text", + "content": ", Kendall " + }, + { + "bbox": [ + 130, + 447, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 130, + 447, + 481, + 519 + ], + "type": "text", + "content": ", and Spearman " + }, + { + "bbox": [ + 130, + 447, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 447, + 481, + 519 + ], + "type": "text", + "content": " correlation scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "content": "Flickr8k-Expert and Flickr8k-CF [19]. These datasets consist of image-caption pairs with corresponding human ratings. Specifically, Flickr8k-Expert comprises 17k expert annotations for visual-textual pairs, with a total of 5,664 different images. Each pair receives a score ranging from 1 (lack of correlation) to 4 (accurate depiction), where 1 indicates a lack of correlation between the caption and the image, and 4 indicates an accurate depiction of the image without errors. On the other hand, Flickr8k-CF is composed of 145k binary quality judgments, collected from CrowdFlower, for 48k image-caption pairs containing 1,000 unique images. Each pair is annotated with at least three binary scores, where “yes” denotes that the caption correlates with the image. To measure the alignment with human judgment, we compute the mean proportion of “yes” annotations as the score for each pair." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Sarto et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 133, + 479, + 284 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 462, + 125 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 462, + 125 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 462, + 125 + ], + "type": "text", + "content": "Table 1: Ablation study results. US indicates the number of pseudo tokens." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 133, + 479, + 284 + ], + "lines": [ + { + "bbox": [ + 134, + 133, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 134, + 133, + 479, + 284 + ], + "type": "table", + "html": "
Flickr8k-ExpertFlickr8k-CFPascal-50S
Kend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρAccuracy
Architectural Components
w/o mapping module (i.e. MLP)53.153.565.335.518.343.581.9
w/o template captions53.754.166.035.518.443.682.5
w/o regularization branch54.154.566.535.718.543.682.7
Score Formulation
w/o textual similarity51.151.263.034.417.730.580.9
w/o visual similarity53.854.266.035.418.343.781.9
Pseudo-token Size
w/ US = 154.154.566.435.117.130.382.6
w/ US = 254.154.566.436.118.744.482.8
w/ US = 454.354.766.635.418.343.882.4
w/ US = 854.054.466.335.918.644.581.7
BRIDGE (US = 3)54.454.867.736.118.744.582.6
", + "image_path": "ddb339a0e6cc0d78a427ea55c7cd61961f6c1ac1f9bdd7e34776ded7888346e2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": "Composite [1]. It contains 12k human ratings for image-caption pairs including a combination of images taken from COCO [34] (2,007 images), Flickr8k [19] (997 images), and Flickr30k [61] (991 images). In this dataset, human evaluators assess each image-caption pair, assigning a score within the range of 1 to 5 to estimate the correspondence of the caption with the associated image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 358, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 480, + 467 + ], + "type": "text", + "content": "Pascal50-S [55]. It presents pairwise preference judgments between two captions. Overall, the dataset consists of 4,000 sentence pairs, each of them associated with an image from the UIUC Pascal sentence dataset [45]. Each pair is associated with 48 human judgments, where each evaluation indicates which sentence better describes the given image. The sentence pairs are categorized into four groups: (i) both human-written and correct captions (HC), (ii) both human-written captions where one is correct and the other is wrong (HI), (iii) both correct captions but one written by humans and the other machine-generated (HM), (iv) both machine-generated and correct captions (MM)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 480, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 480, + 541 + ], + "type": "text", + "content": "FOIL [49]. The dataset comprises image-caption pairs from the COCO dataset [34]. In this scenario, captions undergo perturbation by generating modified versions that closely resemble the originals but introduce a single error, referred to as \"foil word\". For a fair comparison, we select the subset of the validation set that does not overlap with the portion of COCO used during training, resulting in 8,000 images, each paired with a foil-correct textual counterpart." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 559, + 313, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 313, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 313, + 571 + ], + "type": "text", + "content": "4.3 Ablation Studies and Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 578, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 480, + 628 + ], + "type": "text", + "content": "To evaluate the effectiveness of our metric, we start by analyzing variations of our main architectural components. Then, we assess the impact of caption templates in our score formulation. All these experiments are performed using CLIP ViT-B/32 as backbone and reported in Table 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Contribution of Architectural Components. We first investigate the performance of the most straightforward implementation of a mapping module, structured as a two-layer MLP following [46]. We also validate the importance" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 136, + 324, + 261 + ], + "blocks": [ + { + "bbox": [ + 133, + 117, + 326, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 117, + 326, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 326, + 129 + ], + "type": "text", + "content": "Table 2: Impact of different template captions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 136, + 324, + 261 + ], + "lines": [ + { + "bbox": [ + 136, + 136, + 324, + 261 + ], + "spans": [ + { + "bbox": [ + 136, + 136, + 324, + 261 + ], + "type": "table", + "html": "
ExpertCFPascal-50S
Kend. τbKend. τbAcc.
Transformer Templates
w/o mapping module46.131.680.4
BRIDGE54.035.982.7
(+7.9)(+4.3)(+2.3)
BLIP Templates
w/o mapping module48.633.881.0
BRIDGE54.436.182.6
(+5.8)(+2.3)(+1.6)
BLIP-2 Templates
w/o mapping module49.434.282.4
BRIDGE54.436.282.9
(+5.0)(+2.0)(+0.5)
", + "image_path": "b343b42c025e3a411eed6cd5f9aabb952228e74361d48fe93557f21f576fed8c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 353, + 115, + 410, + 157 + ], + "blocks": [ + { + "bbox": [ + 353, + 115, + 410, + 157 + ], + "lines": [ + { + "bbox": [ + 353, + 115, + 410, + 157 + ], + "spans": [ + { + "bbox": [ + 353, + 115, + 410, + 157 + ], + "type": "image", + "image_path": "612a9f2728812130403915490f8143dc29b7208409534a8d155ef174c35bc96e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 160, + 345, + 183 + ], + "lines": [ + { + "bbox": [ + 339, + 160, + 345, + 183 + ], + "spans": [ + { + "bbox": [ + 339, + 160, + 345, + 183 + ], + "type": "text", + "content": "C C C C C C" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 339, + 216, + 346, + 231 + ], + "lines": [ + { + "bbox": [ + 339, + 216, + 346, + 231 + ], + "spans": [ + { + "bbox": [ + 339, + 216, + 346, + 231 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 339, + 239, + 346, + 260 + ], + "lines": [ + { + "bbox": [ + 339, + 239, + 346, + 260 + ], + "spans": [ + { + "bbox": [ + 339, + 239, + 346, + 260 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 415, + 116, + 471, + 157 + ], + "blocks": [ + { + "bbox": [ + 415, + 116, + 471, + 157 + ], + "lines": [ + { + "bbox": [ + 415, + 116, + 471, + 157 + ], + "spans": [ + { + "bbox": [ + 415, + 116, + 471, + 157 + ], + "type": "image", + "image_path": "4bddc7b4c62a1eac6bbaba29d702bab35e35a37a90d2c72e28642d5fba601265.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 186, + 346, + 209 + ], + "lines": [ + { + "bbox": [ + 339, + 186, + 346, + 209 + ], + "spans": [ + { + "bbox": [ + 339, + 186, + 346, + 209 + ], + "type": "text", + "content": "TRANSE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 358, + 161, + 407, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 161, + 407, + 184 + ], + "spans": [ + { + "bbox": [ + 358, + 161, + 407, + 184 + ], + "type": "text", + "content": "A young girl is preparing to blow out her candle" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 419, + 161, + 468, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 161, + 468, + 183 + ], + "spans": [ + { + "bbox": [ + 419, + 161, + 468, + 183 + ], + "type": "text", + "content": "A painting of a table with fruit on top of it" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 358, + 186, + 404, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 186, + 404, + 209 + ], + "spans": [ + { + "bbox": [ + 358, + 186, + 404, + 209 + ], + "type": "text", + "content": "A young girl is eating a piece of cake" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 420, + 186, + 465, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 186, + 465, + 209 + ], + "spans": [ + { + "bbox": [ + 420, + 186, + 465, + 209 + ], + "type": "text", + "content": "A painting of a painting of fruits and a painting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 357, + 213, + 406, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 213, + 406, + 236 + ], + "spans": [ + { + "bbox": [ + 357, + 213, + 406, + 236 + ], + "type": "text", + "content": "A woman blowing out a candle on a cake" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 418, + 213, + 469, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 213, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 418, + 213, + 469, + 236 + ], + "type": "text", + "content": "A painting of a pitcher and oranges on a table" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 239, + 406, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 239, + 406, + 262 + ], + "spans": [ + { + "bbox": [ + 356, + 239, + 406, + 262 + ], + "type": "text", + "content": "A girl blowing out a candle on a cupcake" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 419, + 239, + 469, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 239, + 469, + 262 + ], + "spans": [ + { + "bbox": [ + 419, + 239, + 469, + 262 + ], + "type": "text", + "content": "A painting of a still life with oranges and a candle" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": "of the template captions through a model variant in which a set of learnable tokens " + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathsf{S}^*" + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": " serves as input for the mapping module, without relying on template captions. In both variants, given the absence of template captions, we construct a template such as 'a photo of " + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathsf{S}^*" + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": "' and extract its features using the CLIP text encoder. In the Table, it can be seen that, regardless of any architectural changes, it is important to provide a simple sentence structure to the mapping module to achieve competitive performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": "In addition to these baselines, we devise a variant to analyze the contribution of the regularization branch. In this setting, we employ template captions as input for the mapping module, resulting in a substantial improvement of " + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "inline_equation", + "content": "+1.0" + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": " Kendall " + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "inline_equation", + "content": "\\tau_{b}" + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "inline_equation", + "content": "+0.8" + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": " accuracy points compared to the MLP variant, respectively on the Flickr8k-Expert and on the Pascal-50S dataset. When introducing the regularization branch (i.e. the complete BRIDGE architecture), further enhancements can be observed especially on the Flickr8k-CF with an improvement of " + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "inline_equation", + "content": "+0.4" + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": " points in terms of the Kendall " + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "inline_equation", + "content": "\\tau_{b}" + }, + { + "bbox": [ + 130, + 361, + 482, + 457 + ], + "type": "text", + "content": " correlation score." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 458, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 518 + ], + "type": "text", + "content": "We also emphasize the importance of each component in our score formulation. Specifically, we present correlation results when employing only the visual similarity within our architecture, which is the original CLIP-Score formulation. As observed, performance drops drastically when relying only on visual information. A less significant drop is observed when employing only textual similarity." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "type": "text", + "content": "As an additional analysis, we report the effect of changing the number of the pseudo tokens for each noun chunk, denoting it as unit size (US). Specifically, we compute the scores employing " + }, + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{US} = 1,2,3,4,8" + }, + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "type": "text", + "content": ". From the results, it can be seen that " + }, + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{US} = 3" + }, + { + "bbox": [ + 130, + 518, + 482, + 579 + ], + "type": "text", + "content": " generally leads to the best performance across nearly all evaluation metrics. This configuration is used in all experiments reported in the paper." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Analysis on Caption Templates. We analyze the effect of changing the initial template captions for our model. Specifically, we employ template captions generated by a conventional Transformer-based captioner that uses CLIP features as input and is trained only on the COCO dataset, as well as those generated by BLIP [29], and BLIP-2 [28]. Notably, we select template captions of different quality based on both standard metric evaluations and correlations with human judgment. To assess the quality of the raw generated templates, we observe that" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Sarto et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "the CIDEr score of these models on the COCO test set is equal to 114.2, 131.4, and 145.8 respectively for the standard Transformer model, BLIP, and BLIP-2. Note that all models were trained with cross-entropy loss only. Results on the Flickr8k-Expert, Flickr8k-CF, and Pascal-50S datasets are reported in Table 2. To qualitatively validate the generated templates, we include sample captions generated by the three models compared to a ground-truth caption from the COCO test set. Specifically, captions generated by BLIP-2 are generally more detailed and effectively describe the visual content of the input image compared to those generated by BLIP and, notably, the standard Transformer model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": "For each caption template source, we compute the correlation scores when the mapping module is disabled and the captions are directly fed to the text encoder. We compare it with our standard BRIDGE score, considering the different caption templates. Across all datasets, it is evident that directly using the caption templates as input to the text encoder leads to poor performance. This highlights the intended flexibility of template captions as skeletal representations, allowing the model to enhance them with fine-grained visual features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "text", + "content": "In fact, starting from these simple template captions and following our approach, we achieve improvements of " + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "inline_equation", + "content": "+5.8" + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "inline_equation", + "content": "+2.3" + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "text", + "content": " Kendall " + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "inline_equation", + "content": "\\tau_{b}" + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "text", + "content": " points and " + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "inline_equation", + "content": "+1.6" + }, + { + "bbox": [ + 130, + 308, + 482, + 428 + ], + "type": "text", + "content": " accuracy points, respectively, on the Flickr8k-Expert, Flickr8k-CF, and Pascal50-S datasets when using BLIP caption templates. The overall best results are with captions from the BLIP-2 model, confirming that better templates can indeed lead to improved results. However, even when using lower-quality captions, the final correlation results are very close to those obtained with higher-quality captions. This highlights the robustness of our metric to caption templates of varying quality and that it is not necessary to rely on captions generated by large-scale captioners to achieve strong correlation scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 443, + 432, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 432, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 432, + 456 + ], + "type": "text", + "content": "4.4 Comparison with State-of-the-Art Captioning Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "text", + "content": "Evaluating Sample-Level Human Correlation. We evaluate the sample-level human correlation on the Flickr8k [19] and Composite [1] datasets. Following previous works [18,63], we compute Kendall correlation scores in both " + }, + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\tau_{b}" + }, + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "text", + "content": " versions and also include the Spearman " + }, + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "text", + "content": " score. Results are reported in Table 3, where we compare our proposed BRIDGE metric against other reference-free evaluation scores like UMIC [26], CLIP-S [18], and PAC-S [47]. Moreover, we also compare with standard captioning evaluation metrics (i.e. BLEU [41], ME-TEOR [4], CIDEr [55], and SPICE [2]) and more recent reference-based solutions that exploit text-only or cross-modal learned embeddings, such as BERT-S [63], BERT-S++ [60], TIGEr [22], VilBERTScore [27], and MID [24]. For completeness, we also include the reference-based versions of CLIP-S and PAC-S, termed RefCLIP-S and RefPAC-S, which however are not directly comparable with our solution as both rely on a set of five reference captions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "As it can be seen, BRIDGE outperforms other reference-free metrics in terms of correlation with human judgment, achieving the highest scores on almost all datasets. Specifically, compared to CLIP-S, BRIDGE shows improvements in terms of Kendall " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "+3.6" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "+2.8" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " points when using ViT-B/32 and ViT-L/14" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 448, + 103 + ], + "type": "text", + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 132, + 480, + 344 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "type": "text", + "content": "Table 3: Correlation scores on Flickr8k-Expert, Flickr8k-CF, and Composite [1, 19]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 132, + 480, + 344 + ], + "lines": [ + { + "bbox": [ + 134, + 132, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 134, + 132, + 480, + 344 + ], + "type": "table", + "html": "
Flickr8k-ExpertFlickr8k-CFComposite
Kend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρKend. τbKend. τcSpear. ρ
Reference-based metrics
BLEU-4 [41]30.630.838.716.98.721.028.330.638.1
METEOR [4]41.541.851.922.211.527.636.038.948.1
CIDEr [55]43.643.954.324.612.730.534.937.747.0
SPICE [2]51.744.955.124.412.031.338.840.349.1
BERT-S [63]-39.250.322.8--39.930.148.6
BERT-S++ [60]48.146.756.9---42.344.952.1
TIGER [22]51.449.348.6---47.545.455.3
ViLBERTScore [27]54.250.161.3---51.452.458.7
MID [24]-54.9-37.3-----
RefCLIP-S [18]52.653.064.636.418.844.751.255.466.2
RefPAC-S [47]55.555.967.637.619.546.253.057.368.0
Reference-free metrics
UMIC [26]-46.8-------
CLIP-S (ViT-B/32) [18]51.151.263.034.417.730.549.853.864.3
PAC-S (ViT-B/32) [47]53.954.366.136.018.644.451.555.766.3
BRIDGE (ViT-B/32)54.454.866.736.118.744.550.955.065.4
CLIP-S (ViT-L/14) [18]52.653.064.735.218.243.351.355.465.9
PAC-S (ViT-L/14) [47]55.155.567.336.819.045.352.356.567.1
BRIDGE (ViT-L/14)55.455.867.736.319.044.752.957.267.8
", + "image_path": "f005db42a60070404eee56074b7b2bbc930d6acdd0e2a1312a840639c14ff025.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 359, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 420 + ], + "type": "text", + "content": "backbone on the Flickr8k-Expert dataset. These improvements extend consistently across all datasets and correlation scores. Compared to PAC-S, BRIDGE still achieves superior results on both Flickr8k-Expert and Flickr8k-CF, while performing on par on the Composite dataset (i.e. PAC-S achieves the best results when using ViT-B/32, while BRIDGE outperforms PAC-S using ViT-L/14)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 450 + ], + "type": "text", + "content": "Discriminating Correct Captions. We also evaluate the effectiveness of our metric on the PASCAL-50S dataset [55]. In this context, instead of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 451, + 272, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 272, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 272, + 666 + ], + "type": "text", + "content": "calculating correlation scores, we compute accuracy by determining, for each pair, the caption favored by the majority of human ratings as correct (with ties being resolved randomly). We then measure how frequently the evaluation metric assigns a higher score to the chosen caption. Following previous works [18], we randomly select five reference captions from the set of 48 provided by the dataset and average the results over five distinct draws. Accuracy values are reported in Table 4. The results show that when using" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 285, + 496, + 471, + 658 + ], + "blocks": [ + { + "bbox": [ + 276, + 459, + 482, + 493 + ], + "lines": [ + { + "bbox": [ + 276, + 459, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 276, + 459, + 482, + 493 + ], + "type": "text", + "content": "Table 4: Accuracy results on Pascal-50S [55] averaged over five random draws of reference captions. The " + }, + { + "bbox": [ + 276, + 459, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 276, + 459, + 482, + 493 + ], + "type": "text", + "content": " marker indicates scores from previous works." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 285, + 496, + 471, + 658 + ], + "lines": [ + { + "bbox": [ + 285, + 496, + 471, + 658 + ], + "spans": [ + { + "bbox": [ + 285, + 496, + 471, + 658 + ], + "type": "table", + "html": "
HCHIHMMMMean
Reference-based metrics
BLEU-4 [41]60.393.185.757.074.0
METEOR [4]66.097.794.066.681.1
CIDEr [55]66.597.990.765.280.1
BERT-S++†[60]65.498.196.460.380.1
TIGER†[22]56.099.892.874.280.7
MID†[24]67.099.797.476.885.2
RefCLIP-S [18]64.999.595.573.383.3
RefPAC-S [48]67.799.696.075.684.7
Reference-free metrics
CLIP-S (ViT-B/32) [18]55.999.396.572.080.9
PAC-S (ViT-B/32) [47]60.699.396.972.982.4
BRIDGE (ViT-B/32)59.499.497.574.082.6
CLIP-S (ViT-L/14) [18]57.099.696.773.581.7
PAC-S (ViT-L/14) [47]59.599.495.874.782.2
BRIDGE (ViT-L/14)61.299.696.674.182.9
", + "image_path": "17092aa887071e5d2f2852c70f9bfbf5b37655bd9f053d097ebcf3629bb65d97.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Sarto et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 121, + 235, + 250 + ], + "blocks": [ + { + "bbox": [ + 134, + 121, + 235, + 250 + ], + "lines": [ + { + "bbox": [ + 134, + 121, + 235, + 250 + ], + "spans": [ + { + "bbox": [ + 134, + 121, + 235, + 250 + ], + "type": "image", + "image_path": "b8227c3c7cdf1b96f90487fce67a5fa807108b59008efc1a9fd3a6303ba53ff3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 259, + 343, + 293 + ], + "lines": [ + { + "bbox": [ + 132, + 259, + 343, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 343, + 293 + ], + "type": "text", + "content": "Fig. 4: Sample images from the FOIL dataset [49] and corresponding scores generated by our proposed metric compared with CLIP-S and PAC-S." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 121, + 341, + 250 + ], + "blocks": [ + { + "bbox": [ + 238, + 121, + 341, + 250 + ], + "lines": [ + { + "bbox": [ + 238, + 121, + 341, + 250 + ], + "spans": [ + { + "bbox": [ + 238, + 121, + 341, + 250 + ], + "type": "image", + "image_path": "b984061a8407b42408d3544b43b77986c02ab8fcce872b39e45f71c29da2d9eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 352, + 150, + 460, + 296 + ], + "blocks": [ + { + "bbox": [ + 346, + 125, + 466, + 147 + ], + "lines": [ + { + "bbox": [ + 346, + 125, + 466, + 147 + ], + "spans": [ + { + "bbox": [ + 346, + 125, + 466, + 147 + ], + "type": "text", + "content": "Table 5: Accuracy results on the FOIL [49] dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 150, + 460, + 296 + ], + "lines": [ + { + "bbox": [ + 352, + 150, + 460, + 296 + ], + "spans": [ + { + "bbox": [ + 352, + 150, + 460, + 296 + ], + "type": "table", + "html": "
Acc.
Reference-based metrics
BLEU-4 [41]66.2
METEOR [4]70.1
CIDEr [55]85.7
MID [24]90.5
RefCLIP-S [18]91.0
RefPAC-S [18]93.7
Reference-free metrics
CLIP-S (ViT-B/32) [18]87.2
PAC-S (ViT-B/32) [47]89.9
BRIDGE (ViT-B/32)91.5
CLIP-S (ViT-L/14) [18]90.9
PAC-S (ViT-L/14) [47]91.9
BRIDGE (ViT-L/14)93.0
", + "image_path": "2c0964f9a6d38e4c216aecf52f65f8ef967b3eafdee40e1f2dc30b8cf7e7ed15.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": "both ViT-B/32 and ViT-L/14, BRIDGE consistently outperform CLIP-S across all categories, showcasing an average accuracy increase of " + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "inline_equation", + "content": "+1.7" + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": " points, respectively. When comparing with PAC-S, our solution can better discriminate the correct captions on both backbones with an average accuracy increase of " + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "inline_equation", + "content": "+0.2" + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "inline_equation", + "content": "+0.7" + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": " using ViT-B/32 and ViT-L/14 respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 370, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 468 + ], + "type": "text", + "content": "Object Hallucination Analysis. We then extend our analysis to the FOIL dataset [49] for correctly identifying captions that may contain object hallucinations. Table 5 shows the accuracy results. As it can be seen, BRIDGE outperforms CLIP-S and PAC-S, exhibiting an increase respectively of " + }, + { + "bbox": [ + 130, + 370, + 482, + 468 + ], + "type": "inline_equation", + "content": "+4.3" + }, + { + "bbox": [ + 130, + 370, + 482, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 370, + 482, + 468 + ], + "type": "inline_equation", + "content": "+1.6" + }, + { + "bbox": [ + 130, + 370, + 482, + 468 + ], + "type": "text", + "content": " points when using ViT-B/32 as backbone. Similar improvements can also be observed when employing more robust visual features. These results demonstrate the capabilities of our metric also to identify hallucinated objects correctly. In Fig 4, we present sample results comparing our metric with CLIP-S and PAC-S." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 487, + 285, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 285, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 285, + 499 + ], + "type": "text", + "content": "4.5 System-level Correlation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Finally, we delve into the efficacy of our proposed metric when evaluating popular existing captioning models. To this aim, we generate predictions of several state-of-the-art captioning models on the COCO test set, including Show and Tell and Show, Attend and Tell which are among the first image captioning models based on deep learning, Up-Down [3], SGAE [59], AoANet [20], " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^2" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " Transformer [14], X-Transformer [40] which all include region-based image features with either LSTM-based or Transformer-based language models, and the recently proposed COS-Net model [32] that incorporates CLIP features. In addition to reporting evaluations on traditional captioning models, we also include recent LLM-based captioning models, including ZeroCap [51] and SmallCap [44], which are based on GPT-2 [43], and MiniGPT-v2 [9], BLIP-2 [28], IDEFICS [25], LLaVA-1.5 [35,36], and InstructBLIP [15] which instead are based on larger-scale LLMs like FlanT5 [12], Vicuna [11], or LLaMA [52,53]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 141, + 459, + 299 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 6: Evaluation scores of traditional and LLM-based captioners on COCO test set (" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": ": not trained/fine-tuned on COCO)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 141, + 459, + 299 + ], + "lines": [ + { + "bbox": [ + 149, + 141, + 459, + 299 + ], + "spans": [ + { + "bbox": [ + 149, + 141, + 459, + 299 + ], + "type": "table", + "html": "
BLEU-4METEORCIDErCLIP-SPAC-SBRIDGE
TraditionalShow and Tell [56]31.425.097.20.5720.7720.788
Show, Attend and Tell [58]33.426.2104.60.5820.7850.804
Up-Down [3]36.727.9122.70.7230.8030.821
SGAE [59]39.028.4129.10.7340.8120.833
AoANet [20]38.929.2129.80.7370.8150.836
M2Transformer [14]39.129.2131.20.7340.8130.841
X-Transformer [40]39.729.5132.80.6100.8120.845
COS-Net [32]42.030.6141.10.7580.8320.859
LLM-basedZeroCap♦[51]2.310.115.10.8100.8160.862
SmallCap [44]37.027.9119.70.7480.8260.847
MiniGPT-v2 [9]18.824.680.40.7520.8180.845
BLIP-2 [28]43.732.0145.80.7670.8370.868
IDEFICS-9B♦[25]4.319.150.00.7400.7860.838
LLaVA-1.5-7B♦[35]8.128.069.60.7840.8090.867
InstructBLIP-Flan-T5-XL♦[15]6.128.138.10.8170.8370.902
Humans-24.187.60.7740.8230.856
", + "image_path": "e28418cfee9f72c1a8927c81fd0535480991b5e00c82ca2fc3af6f962b755e3f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "type": "text", + "content": "The results are presented in Table 6, where we evaluate our BRIDGE scores against both standard metrics, such as BLEU-4, METEOR, and CIDEr, and more recent ones like CLIP-S and PAC-S. Captioning models are compared to a human baseline, in which, for each sample, one human-annotated sentence (selected randomly from the five provided by the COCO dataset) serves as a candidate caption. As shown in the table, BRIDGE can effectively evaluate human-annotated sentences which obtain a score similar to recent state-of-the-art captioning models such as COS-Net. This capability lacks in standard metrics such as METEOR and CIDEr which rank human captions lower than those generated by less-performing captioning models like Show, Attend, and Tell or Up-Down." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 433, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 481, + 552 + ], + "type": "text", + "content": "When considering the evaluation of captions generated by existing models, our metric shows a strong correlation with standard evaluation metrics when evaluating traditional image captioners or large-scale models that are fine-tuned on the COCO dataset like BLIP-2. When instead considering more recent approaches that are based on large language models and are not fine-tuned on COCO, BRIDGE still recognizes the goodness of generated captions, raking InstructBLIP as the best-performing approach. This demonstrates the capabilities of our metric to correctly evaluate longer and more detailed captions which are typically generated by LLM-based multimodal models [7, 16, 30] and that might be very different from captions contained in the COCO dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 571, + 220, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 220, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 220, + 583 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "In this paper, we have presented a novel learnable, and reference-free image captioning metric that combines text and dense visual features. Our proposal, BRIDGE, employs templated captions that are enriched with fine-grained visual cues thanks to a mapping network. Through experimental evaluation, we demonstrate that BRIDGE outperforms existing reference-free metrics in terms of correlation with human judgment and sensitivity to hallucinated objects." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Sarto et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 211 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 211 + ], + "type": "text", + "content": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been conducted under a research grant co-funded by Leonardo S.p.A. and supported by the PNRR-M4C2 (PE0000013) project \"FAIR - Future Artificial Intelligence Research\" and by the PRIN project \"MUSMA\" (CUP G53D23002930006 - M4C2 I1.1), both funded by EU - Next-Generation EU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 228, + 197, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 197, + 241 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 197, + 241 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 251, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 284 + ], + "type": "text", + "content": "1. Aditya, S., Yang, Y., Baral, C., Fermuller, C., Aloimonos, Y.: From Images to Sentences through Scene Description Graphs using Commonsense Reasoning and Knowledge. arXiv preprint arXiv:1511.03292 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "type": "text", + "content": "2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: SPICE: Semantic Propositional Image Caption Evaluation. In: ECCV (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 307, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 480, + 338 + ], + "type": "text", + "content": "3. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "type": "text", + "content": "4. Banerjee, S., Lavie, A.: METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In: ACL Workshops (2005)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "text", + "content": "5. Barraco, M., Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: With a Little Help from your own Past: Prototypical Memory Networks for Image Captioning. In: ICCV (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 394, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 480, + 415 + ], + "type": "text", + "content": "6. Bird, S., Klein, E., Loper, E.: Natural Language Processing with Python: Analyzing Text with the Natural Language Toolkit. O'Reilly Media, Inc. (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 416, + 480, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 416, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 416, + 480, + 447 + ], + "type": "text", + "content": "7. Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Baraldi, L., Cornia, M., Cucchiara, R.: The Revolution of Multimodal Large Language Models: A Survey. In: ACL Findings (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 449, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 468 + ], + "type": "text", + "content": "8. Chan, D., Petryk, S., Gonzalez, J.E., Darrell, T., Canny, J.: CLAIR: Evaluating Image Captions with Large Language Models. In: EMNLP (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 470, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 480, + 501 + ], + "type": "text", + "content": "9. Chen, J., Li, D.Z.X.S.X., Zhang, Z.L.P., Xiong, R.K.V.C.Y., Elhoseiny, M.: MiniGPT-v2: Large Language Model as a Unified Interface for Vision-Language Multi-task Learning. arXiv preprint arXiv:2310.09478 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "text", + "content": "0. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: UNITER: Universal Image-TExt Representation Learning. In: ECCV (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 555 + ], + "type": "text", + "content": "1. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An Open-Source Chatbot Improressing GPT-4 with " + }, + { + "bbox": [ + 138, + 525, + 480, + 555 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 525, + 480, + 555 + ], + "type": "text", + "content": " * ChatGPT Quality (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 557, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 557, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 557, + 480, + 621 + ], + "type": "text", + "content": "12. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S.S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Castro-Ros, A., Pellat, M., Robinson, K., Valter, D., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E.H., Dean, J., Devlin, J., Roberts, A., Zhou, D., Le, Q.V., Wei, J.: Scaling Instruction-Finetuned Language Models. JMLR 25(70), 1-53 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 623, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 623, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 623, + 480, + 643 + ], + "type": "text", + "content": "13. Cornia, M., Baraldi, L., Cucchiara, R.: SMArT: Training Shallow Memory-aware Transformers for Robotic Explainability. In: ICRA (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "14. Cornia, M., Stefanini, M., Baraldi, L., Cucchiara, R.: Meshed-Memory Transformer for Image Captioning. In: CVPR (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning. arXiv preprint arXiv:2305.06500 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "16. Dong, H., Li, J., Wu, B., Wang, J., Zhang, Y., Guo, H.: Benchmarking and Improving Detail Image Caption. arXiv preprint arXiv:2405.19092 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "text", + "content": "17. Herdade, S., Kappeler, A., Boakye, K., Soares, J.: Image captioning: Transforming objects into words. In: NeurIPS (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "text", + "content": "18. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In: EMNLP (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "text", + "content": "19. Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking task: Data, models and evaluation metrics. JAIR 47, 853-899 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 259 + ], + "type": "text", + "content": "20. Huang, L., Wang, W., Chen, J., Wei, X.Y.: Attention on Attention for Image Captioning. In: ICCV (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 292 + ], + "type": "text", + "content": "21. Jiang, M., Hu, J., Huang, Q., Zhang, L., Diesner, J., Gao, J.: REO-Relevance, Extraness, Omission: A Fine-grained Evaluation for Image Captioning. In: EMNLP (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "22. Jiang, M., Huang, Q., Zhang, L., Wang, X., Zhang, P., Gan, Z., Diesner, J., Gao, J.: TIGER: Text-to-Image Grounding for Image Caption Evaluation. In: EMNLP (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "23. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "24. Kim, J.H., Kim, Y., Lee, J., Yoo, K.M., Lee, S.W.: Mutual Information Divergence: A Unified Metric for Multimodal Generative Models. In: NeurIPS (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "type": "text", + "content": "25. Laurencon, H., Saulnier, L., Tronchon, L., Bekman, S., Singh, A., Lozhkov, A., Wang, T., Karamcheti, S., Rush, A.M., Kiela, D., Cord, M., Sanh, V.: OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents. In: NeurIPS (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "26. Lee, H., Yoon, S., Dernoncourt, F., Bui, T., Jung, K.: UMIC: An Unreferenced Metric for Image Captioning via Contrastive Learning. In: ACL (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "27. Lee, H., Yoon, S., Dernoncourt, F., Kim, D.S., Bui, T., Jung, K.: ViLBERTScore: Evaluating Image Caption Using Vision-and-Language BERT. In: EMNLP Workshops (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "28. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping Language-Image Pretraining with Frozen Image Encoders and Large Language Models. In: ICML (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "text", + "content": "29. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In: ICML (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "30. Li, X., Tu, H., Hui, M., Wang, Z., Zhao, B., Xiao, J., Ren, S., Mei, J., Liu, Q., Zheng, H., et al.: What If We Recap tion Billions of Web Images with LLaMA-3? arXiv preprint arXiv:2406.08478 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "text", + "content": "31. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: ECCV (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 600 + ], + "type": "text", + "content": "32. Li, Y., Pan, Y., Yao, T., Mei, T.: Comprehending and ordering semantics for image captioning. In: CVPR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "33. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL Workshops (2004)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: ECCV (2014)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "35. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved Baselines with Visual Instruction Tuning. In: CVPR (2024)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Sarto et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 478, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 478, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 478, + 127 + ], + "type": "text", + "content": "36. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual Instruction Tuning. In: NeurIPS (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 128, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 128, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 480, + 148 + ], + "type": "text", + "content": "37. Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization. In: ICLR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "38. Lu, J., Batra, D., Parikh, D., Lee, S.: ViLBERT: Pretraining Task-Agnostic Visi-olinguistic Representations for Vision-and-Language Tasks. In: NeurIPS (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "text", + "content": "39. Oord, A.v.d., Li, Y., Vinyals, O.: Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "text", + "content": "40. Pan, Y., Yao, T., Li, Y., Mei, T.: X-Linear Attention Networks for Image Captioning. In: CVPR (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 213, + 480, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 234 + ], + "type": "text", + "content": "41. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: BLEU: a method for automatic evaluation of machine translation. In: ACL (2002)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 235, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 267 + ], + "type": "text", + "content": "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transferable Visual Models From Natural Language Supervision. In: ICML (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "text", + "content": "43. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners. OpenAI Blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "type": "text", + "content": "44. Ramos, R., Martins, B., Elliott, D., Kementchedjhieva, Y.: SmallCap: lightweight image captioning prompted with retrieval augmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 310, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 331 + ], + "type": "text", + "content": "45. Rashtchian, C., Young, P., Hodosh, M., Hockenmaier, J.: Collecting Image Annotations Using Amazon's Mechanical Turk. In: NAACL Workshops (2010)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "text", + "content": "46. Saito, K., Sohn, K., Zhang, X., Li, C.L., Lee, C.Y., Saenko, K., Pfister, T.: Pic2Word: Mapping Pictures to Words for Zero-Shot Composed Image Retrieval. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "47. Sarto, S., Barraco, M., Cornia, M., Baraldi, L., Cucchiara, R.: Positive-Augmented Contrastive Learning for Image and Video Captioning Evaluation. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "type": "text", + "content": "48. Sarto, S., Cornia, M., Baraldi, L., Cucchiara, R.: Retrieval-Augmented Transformer for Image Captioning. In: CBMI (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "text", + "content": "49. Shekhar, R., Pezzelle, S., Klimovich, Y., Herbelot, A., Nabi, M., Sangineto, E., Bernardi, R.: FOIL it! Find One mismatch between Image and Language caption. In: ACL (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "type": "text", + "content": "50. Shi, Y., Yang, X., Xu, H., Yuan, C., Li, B., Hu, W., Zha, Z.J.: EMScore: Evaluating Video Captioning via Coarse-Grained and Fine-Grained Embedding Matching. In: CVPR (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 482, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 503 + ], + "type": "text", + "content": "51. Tewel, Y., Shalev, Y., Schwartz, I., Wolf, L.: Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In: CVPR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 503, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 480, + 547 + ], + "type": "text", + "content": "52. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "text", + "content": "53. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "text", + "content": "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "55. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: CIDEr: Consensus-based Image Description Evaluation. In: CVPR (2015)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "56. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: CVPR (2015)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "57. Wang, S., Yao, Z., Wang, R., Wu, Z., Chen, X.: FAIER: Fidelity and Adequacy Ensured Image Caption Evaluation. In: CVPR (2021)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 304 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "58. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhutdinov, R., Zemel, R.S., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "59. Yang, X., Tang, K., Zhang, H., Cai, J.: Auto-Encoding Scene Graphs for Image Captioning. In: CVPR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "60. Yi, Y., Deng, H., Hu, J.: Improving Image Captioning Evaluation by Considered Inter Refinences Variance. In: ACL (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "61. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL 2, 67-78 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "62. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: VinVL: Revisiting visual representations in vision-language models. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "63. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: BERTScore: Evaluating Text Generation with BERT. In: ICLR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "text", + "content": "64. Zhu, W., Wang, X.E., Yan, A., Eckstein, M., Wang, W.Y.: ImaginE: An Imagination-Based Automatic Evaluation Metric for Natural Language Generation. In: EACL (2023)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Sarto et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/360699b9-a034-4ed5-b537-07332b36832e_content_list.json b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/360699b9-a034-4ed5-b537-07332b36832e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9b7defb3cebc179fbc1c3013666a119865cca378 --- /dev/null +++ b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/360699b9-a034-4ed5-b537-07332b36832e_content_list.json @@ -0,0 +1,1849 @@ +[ + { + "type": "text", + "text": "BaSIC: BayesNet Structure Learning for Computational Scalable Neural Image Compression", + "text_level": 1, + "bbox": [ + 269, + 140, + 733, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yufeng Zhang $^{1}$ , Hang Yu $^{2}$ , Shizhan Liu $^{1}$ , Wenrui Dai $^{1}$ , and Weiyao Lin $^{\\star 1}$", + "bbox": [ + 326, + 234, + 673, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University, China. {worldlife,shanluzuode,daiwenrui,wylin}@sjtu.edu.cn $^{2}$ Ant Group, China. hyu.hugo@antgroup.com", + "bbox": [ + 303, + 276, + 697, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Despite superior rate-distortion performance over traditional codecs, Neural Image Compression (NIC) is limited by its computational scalability in practical deployment. Prevailing research focuses on accelerating specific NIC modules but is restricted in controlling overall computational complexity. To this end, this work introduces BaSIC (BayesNet structure learning for computational Scalable neural Image Compression), a comprehensive, computationally scalable framework that affords full control over NIC processes. We learn the Bayesian network (BayesNet) structure of NIC for controlling both neural network backbones and autoregressive units. The learning of BayesNet is achieved by solving two sub-problems, i.e., learning a heterogeneous bipartite BayesNet for the inter-node structure to regulate backbone complexity, and a multipartite BayesNet for the intra-node structure to optimize parallel computation in autoregressive units. Experiments demonstrate that our method not only facilitates full computational scalability with more accurate complexity control but also maintains competitive compression performance compared to other computation scalable frameworks under equivalent computational constraints. Code will be available in https://github.com/worldlife123/cbench_BaSIC.", + "bbox": [ + 261, + 359, + 738, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image Compression $\\cdot$ Computational Scalable Compression $\\cdot$ BayesNet Structure Learning $\\cdot$ Autoregressive Model", + "bbox": [ + 261, + 636, + 738, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 691, + 375, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Image Compression (NIC) technologies [7, 11] have significantly outstripped the performance of conventional image CODECs, ushering in an era of intelligent image compression. These advanced CODEcs are intended for deployment across a spectrum of devices, from low-end devices like mobile phones to high-end devices like GPU computation clusters. Each type of device necessitates that CODEcs tailor their performance to specific computational or memory", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author.", + "bbox": [ + 218, + 824, + 388, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7e40b9eb529fe36748b0cfe1367f5f601611e68962e546044fcb5a92bd5724d.jpg", + "image_caption": [ + "Fig. 1: Bayesian network diagrams for various NIC frameworks. Our proposed framework BaSIC (d) generalizes the concepts in (a), (b), and (c)." + ], + "image_footnote": [], + "bbox": [ + 236, + 142, + 467, + 345 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/99a995a3c841f19090b791a708db2180f96b1e6ef47cc7830340e0c89c41ad5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 152, + 767, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "constraints, a requirement known as Computational Scalability. High-end devices can endure more intricate computations for superior encoding quality, whereas low-end devices may prioritize reduced computational load over encoding efficiency to conserve energy.", + "bbox": [ + 212, + 416, + 782, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While conventional image CODECs typically accommodate this functionality (e.g. WebP [28] offers seven levels of computational complexity to regulate lossy image compression), incorporating this attribute into NIC remains a significant hurdle for its broader adoption. Recent studies have focused predominantly on accelerating NIC computations rather than addressing computational scalability. Works such as [11, 12, 25] target the acceleration of autoregressive modules on GPUs, which are frequently used in NIC. Others, including [15, 31, 38], explore neural network acceleration techniques—given that NIC fundamentally relies on neural network architectures. On the other hand, some recent efforts have attempted to integrate dynamic neural networks like slimmable networks [32] within NIC frameworks to adjust the computational complexity [27, 29, 36]. However, these approaches have only partially addressed computational scalability and often neglect crucial components in NIC, such as autoregressive modules, which are essential for compression performance.", + "bbox": [ + 212, + 477, + 784, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we endeavor to create a NIC framework with enhanced computational scalability, incorporating computational control across all NIC components. We call this framework BaSIC (BayesNet structure learning for computational Scalable neural Image Compression). To this end, we re-examine NIC through the broader lens of the Bayesian network (BayesNet) paradigm. Figure 1 illustrates the BayesNet depiction of several exemplary NIC architectures. Specifically, our approach is rooted in the hyperprior BayesNet framework (Figure 1a), which postulates that the generative model can be factorized as $p(\\mathbf{x}|\\mathbf{y})p(\\mathbf{y}|\\mathbf{z})p(\\mathbf{z})$ , where $\\mathbf{x}$ represents observed data, $\\mathbf{y}$ the prior, and $\\mathbf{z}$ the hyperprior. In this context, SlimCAE (Figure 1b) can be interpreted as allowing", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "selective edge types between $\\mathbf{y}$ and $\\mathbf{x}$ , and between $\\hat{\\mathbf{x}}$ and $\\hat{\\mathbf{y}}$ , with each type manifesting different conditional dependencies, modeled via slimmable neural networks with varied channel widths. On the other hand, the joint autoregressive framework (Figure 1c) introduces autoregressive intra-node dependencies within $\\hat{y}$ to enhance compression, at the expense of increased computational demands due to the requisite sequential processing [12]. These examples clearly demonstrate the intimate connection between the structure of the BayesNet and the complexity inherent in NIC frameworks.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our advancement is a NIC framework that exploits end-to-end BayesNet structure learning [23, 33] to meticulously optimize computational scalability. This innovative framework is not bound to specific neural network backbones, making it versatile enough to enhance computational scalability in most existing NIC frameworks with a hyperprior BayesNet representation. Our proposed framework is visualized in Figure 1d. In particular, we represent the BayesNet structure as $\\mathbf{G}$ and impose a prior $p(\\mathbf{G})$ on it. Given that the structure we seek to learn must conform to a directed acyclic graph (DAG), we tailor the prior $p(\\mathbf{G})$ for compression frameworks within this DAG constraint, considering two distinct instances: the inter-node BayesNet, representing conditional dependencies between nodes, and the intra-node BayesNet, which pertains to dependencies within nodes. We depict the former as a heterogeneous bipartite graph and design a structure learning method to seamlessly incorporate dynamic neural networks like slimmable networks [32]. For the latter, we design a network architecture parameterized by multipartite or $k$ -partite graphs to enable a parallel autoregressive process. By manipulating the partite number $k$ , we can adjust the temporal demands of the autoregressive process, thereby directly impacting decompression speed.", + "bbox": [ + 212, + 270, + 787, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When deploying the codec, the configurations of both the inter-node and intra-node BayesNets can be tailored to match the computational budget of target devices. For instance, on less capable devices, we might employ simpler neural network edges within the inter-node BayesNet and a smaller partite number within the intra-node BayesNet, whereas more capable devices can leverage a more complex network and a larger partite number, allowing for scalable computational resource allocation based on the specifics of the device at hand.", + "bbox": [ + 212, + 546, + 789, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are mainly three-fold:", + "bbox": [ + 238, + 656, + 635, + 671 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an end-to-end computationally scalable NIC framework founded on BayesNet structure learning, facilitating fine-tuned adjustments to computational complexity across all facets of the NIC process.", + "- We devise a generative model based on a heterogeneous bipartite structure for the inter-node BayesNet, harnessing the capabilities of existing dynamic neural networks to realize computational scalability effectively.", + "- We design a generative model structured as a multipartite graph for the intra-node BayesNet. This model is tailored to accommodate and streamline parallel computation in practical applications." + ], + "bbox": [ + 223, + 696, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we briefly review recent works regarding the computational complexity of NIC as well as BayesNet structure learning.", + "bbox": [ + 215, + 175, + 784, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 End-to-end Neural Image Compression", + "text_level": 1, + "bbox": [ + 215, + 227, + 583, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The hyperprior framework [4] serves as a foundation for most end-to-end NIC models, favored for its effective compression performance and straightforward implementation. Subsequent studies [7, 19, 24] have enhanced the hyperprior by integrating a masked convolution-based autoregressive model in the latent space to reduce bit-rate. While these improvements offer better compression, they often overlook the increased computational demands. To apply NIC in practice, some recent works are now addressing the need to expedite these models.", + "bbox": [ + 215, + 250, + 785, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Autoregressive Model Acceleration To circumvent the slow serial iteration steps of masked convolutions on GPUs, newer studies introduce parallelized autoregressive models for GPU acceleration. For instance, [25] presents a channelwise parallel autoregressive model, while [12] proposes a spatially parallel model using a checkerboard mask, requiring just two parallel stages. This concept is extended to a multistage model by [21]. Further, [11] amalgamates channel-wise and spatially parallel autoregressive models. Additional acceleration methods include content-adaptive adjustments to parallel autoregressive models [35] and designing simplified autoregressive models [16]. Yet, such models are typically hand-crafted and may not represent the most efficient trade-off between compression complexity and performance.", + "bbox": [ + 215, + 357, + 785, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Network Acceleration Considering that neural networks (i.e., edges in the BayesNets) represent a significant portion of NIC's computational complexity, some works focus on accelerating these directly. For example, there are designs for more efficient neural network operators like a simplified GDN [15] and CPU-friendly down/up-samplers [38]. Others apply general neural network acceleration techniques within compression models, such as MorphNet [15] and learnable channel pruning [30, 31].", + "bbox": [ + 215, + 523, + 785, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Computation Scalable Framework A handful of recent works have also noticed the importance of computational scalability for NIC. SlimCAE [29] employs slimmable networks [32] as the framework's backbone, enabling dynamic adjustments of bit-rate and computation through channel width modulation. Further developments include universal slimmable networks for more precise scalability [36] and the integration of input-adaptive complexity allocation methods [27,36]. Nonetheless, these approaches focus solely on the scalability of neural networks, largely disregarding the autoregressive models' scalability.", + "bbox": [ + 215, + 628, + 785, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 BayesNet Structure Learning", + "text_level": 1, + "bbox": [ + 215, + 770, + 503, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BayesNet structure learning aims to discern the conditional relationships among a set of variables with a graph representation, which ideally should be a directed acyclic graph (DAG). Traditional approaches often involve score-based [8, 34]", + "bbox": [ + 215, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and constraint-based methods [6, 18]. However, they typically operate within a constrained search space, which can lead to less-than-optimal graphs. More recent initiatives introduce neural network techniques for more robust graph generation within the realm of BayesNet structure learning. Some align this learning task with variational inference to facilitate graph structure development within an end-to-end framework [23, 33], utilizing graph acyclicity regularizers [37] that may not guarantee acyclicity. Alternatively, other research employs recurrent-style graph generators that assure acyclicity through iterations [9], but these can become inefficient with an increase in the number of nodes. In our work, we diverge from these established methods due to their inefficiency when dealing with the dense pixelation of images. Instead, we analyze the BayesNet representations inherent in existing NIC frameworks and directly parameterize the graph structure as DAGs, ensuring both acyclicity and efficiency.", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Overview of BaSIC", + "text_level": 1, + "bbox": [ + 215, + 367, + 439, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assume we have a set of observations $\\mathbf{X} = \\{\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_N\\}^3$ , the objective of BayesNet structure learning is to infer a distribution over the structure of the BayesNet $\\mathbf{G} = (\\mathbf{V}, \\mathbf{E}, \\boldsymbol{\\theta})$ that best describes the observations. According to the Bayesian formalism, such optimization problem could be solved by minimizing expectation over $p(\\mathbf{G}|\\mathbf{X})$ of any function $f$ related to $\\mathbf{G}$ , under the constraint that $\\mathbf{G}$ is acyclic:", + "bbox": [ + 212, + 400, + 787, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\mathbf {G}} {\\min } \\mathbb {E} _ {p (\\mathbf {G} | \\mathbf {X})} [ f (\\mathbf {G}) ], \\quad \\mathrm {s . t .} \\mathbf {G} \\text {i s a c y c l i c .} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 503, + 785, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a computationally scalable compression framework, we tackle BayesNet structure learning, linking BayesNet configurations to computational complexity and shaping the function $f(\\mathbf{G})$ accordingly.", + "bbox": [ + 212, + 536, + 784, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Apart from modeling of $\\mathbf{X}$ , the compression framework also involves latent variables $\\mathbf{L} = \\{\\mathbf{y}_1, \\mathbf{y}_2, \\ldots, \\mathbf{y}_N, \\mathbf{z}_1, \\mathbf{z}_2, \\ldots, \\mathbf{z}_N, \\ldots\\}$ to generate $\\mathbf{X}$ . The resulting BayesNet should capture conditional dependencies within and between $\\mathbf{X}$ and $\\mathbf{L}$ . This generative relationship is described by:", + "bbox": [ + 212, + 582, + 787, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {G}, \\mathbf {L}, \\mathbf {X}) = p (\\mathbf {G}) p (\\mathbf {L} | \\mathbf {G}) p (\\mathbf {X} | \\mathbf {G}, \\mathbf {L}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 656, + 785, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The expectation from Equation 1 is then:", + "bbox": [ + 215, + 683, + 517, + 698 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {p (\\mathbf {G}, \\mathbf {L} | \\mathbf {X})} [ f (\\mathbf {G}, \\mathbf {L}) ] = \\mathbb {E} _ {p (\\mathbf {G} | \\mathbf {X}) p (\\mathbf {L} | \\mathbf {X}, \\mathbf {G})} [ f (\\mathbf {G}, \\mathbf {L}) ]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 712, + 785, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing compression frameworks without $\\mathbf{G}$ approximate $p(\\mathbf{L}|\\mathbf{X})$ with an inference network $q(\\mathbf{L}|\\mathbf{X})$ , as in VAEs [17]. Incorporating $\\mathbf{G}$ into our design, we refine our inference network to $q(\\mathbf{G},\\mathbf{L}|\\mathbf{X})$ for enhanced approximation:", + "bbox": [ + 212, + 739, + 787, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {p (\\mathbf {G}, \\mathbf {L} | \\mathbf {X})} [ f (\\mathbf {G}, \\mathbf {L}) ] \\approx \\mathbb {E} _ {q (\\mathbf {G} | \\mathbf {X}) q (\\mathbf {L} | \\mathbf {X}, \\mathbf {G})} [ f (\\mathbf {G}, \\mathbf {L}) ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 797, + 785, + 815 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3 All symbols used in this paper are listed in Appendix.", + "bbox": [ + 217, + 824, + 594, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This approach allows $\\mathbf{G}$ to direct the computational complexity of both the inference and the generative processes. Furthermore, we adopt a differentiable approach to learning $\\mathbf{G}$ , in line with recent advances [23, 33], and incorporate this aspect into our model specification.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition of $f(\\mathbf{G}, \\mathbf{L})$ In line with lossy image compression standards [15,31], our optimization combines bit-rate $\\mathcal{L}_R$ , distortion $\\mathcal{L}_D$ , and complexity $\\mathcal{L}_C$ :", + "bbox": [ + 212, + 207, + 782, + 238 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {q (\\mathbf {G} | \\mathbf {X}) q (\\mathbf {L} | \\mathbf {X}, \\mathbf {G})} [ f (\\mathbf {G}, \\mathbf {L}) ] = \\mathcal {L} _ {R} (\\mathbf {G}, \\mathbf {L}) + \\lambda_ {D} \\mathcal {L} _ {D} (\\mathbf {G}, \\mathbf {L}) + \\lambda_ {C} \\mathcal {L} _ {C} (\\mathbf {G}, \\mathbf {L}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 250, + 785, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Concretely, $\\mathcal{L}_R$ is computed by enacting the BayesNet's generative process with latent node samples $\\hat{\\mathbf{L}}$ :", + "bbox": [ + 214, + 277, + 782, + 308 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} (\\mathbf {G}, \\mathbf {L}) = \\log_ {2} p _ {\\mathbf {L} | \\mathbf {G}} (\\hat {\\mathbf {L}} | \\mathbf {G}), \\quad \\hat {\\mathbf {L}} \\sim q (\\mathbf {L} | \\mathbf {X}, \\mathbf {G}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 320, + 785, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, $\\mathcal{L}_D$ and $\\mathcal{L}_C$ can be tailored to specific optimization objectives. $\\mathcal{L}_D$ may encompass a range of complexity measures or their amalgamations, such as conventional metrics like Mean-Squared Error (MSE) or more advanced machine learning metrics like Frechet Inception Distance (FID) [13]. Conversely, $\\mathcal{L}_C$ can include various complexity metrics or their combinations, such as the number of Floating-point Operations (FLOPs), network parameters, time, or energy consumption. $\\lambda_{D}$ and $\\lambda_{C}$ are the corresponding weights. By setting $f(\\mathbf{G},\\mathbf{L})$ in line with the established rate-distortion (R-D) trade-off, Equation 4 falls in step with loss functions prevalent in standard compression frameworks.", + "bbox": [ + 212, + 349, + 784, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition of $q(\\mathbf{G}|\\mathbf{X})$ As a determiner of the BayesNet structure, $\\mathbf{G}$ affects computational complexity. In practical compression scenarios, including $\\mathbf{G}$ in the bitstream could impact the bit-rate. Typically, it may be more advantageous to treat $\\mathbf{G}$ as independent of $\\mathbf{X}$ , suggesting $q(\\mathbf{G}|\\mathbf{X}) = p(\\mathbf{G})$ . However, in scenarios where complexity is tailored to input data, the model $q(\\mathbf{G}|\\mathbf{X})$ can be employed. DAG Constraint of $p(\\mathbf{G})$ As previously noted in Equation 1, $p(\\mathbf{G})$ should reflect a distribution over DAG structures. For edges between different node groups in compression models, such as $p(\\mathbf{x}|\\mathbf{y})$ or $p(\\mathbf{y}|\\mathbf{z})$ , ensuring a DAG structure is straightforward since bipartite graphs are inherently acyclic; we call this component the Inter-node BayesNet ( $\\mathbf{G}_{inter}$ ). For conditional dependencies within node groups, like $p(\\mathrm{y}_i|\\mathrm{y}_{i-1})$ which also require DAG adherence, we refer to it as the Intra-node BayesNet ( $\\mathbf{G}_{intra}$ ). The forthcoming sections will delve into the methodology for learning both $p(\\mathbf{G}_{inter})$ and $p(\\mathbf{G}_{intra})$ as DAGs.", + "bbox": [ + 212, + 484, + 785, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Structure Learning for Inter and Intra-Node BayesNets", + "text_level": 1, + "bbox": [ + 214, + 705, + 782, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Heterogeneous Bipartite Graphs for Inter-node BayesNet", + "text_level": 1, + "bbox": [ + 214, + 738, + 733, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we specify $p(\\mathbf{G}_{inter})$ . Note that $\\mathbf{G}_{inter}$ is a bipartite graph, and so it inherently satisfies the acyclic constraint.", + "bbox": [ + 212, + 763, + 782, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Naive Bipartite Parameterization of $p(\\mathbf{G}_{inter})$ Consider the task of learning a bipartite graph for $p(\\mathbf{y}|\\mathbf{z})$ . We split the nodes into $\\mathbf{z}_1, \\mathbf{z}_2, \\ldots, \\mathbf{z}_I$ and $\\mathbf{y}_1, \\mathbf{y}_2, \\ldots, \\mathbf{y}_J$ . Bipartite graphs can be represented by an $I \\times J$ adjacency matrix", + "bbox": [ + 212, + 795, + 784, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with each element indicating an edge between $\\mathbf{y}$ and $\\mathbf{z}$ . Consequently, $\\mathbf{G}_{inter}$ can be modeled as $I \\times J$ Bernoulli distributions:", + "bbox": [ + 214, + 146, + 785, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}\\right) = \\operatorname {B e r n o u l l i} \\left(\\pi_ {\\mathbf {y}, \\mathbf {z}}\\right), \\quad \\pi_ {\\mathbf {y}, \\mathbf {z}} \\in \\mathbb {R} ^ {I \\times J}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 184, + 785, + 202 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The edge $p(\\mathbf{y}|\\mathbf{z},\\mathbf{G}_{inter}^{\\mathbf{y};\\mathbf{z}})$ is then defined as:", + "bbox": [ + 215, + 207, + 524, + 224 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {y} | \\mathbf {z}, \\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}) = \\prod_ {i, j} ^ {I, J} p (\\mathbf {y} _ {j} | \\mathbf {z} _ {i}) \\hat {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}} [ i, j ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 232, + 785, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hat{G}_{inter}^{\\mathbf{y},\\mathbf{z}}$ are samples from $p(\\mathbf{G}_{inter}^{\\mathbf{y},\\mathbf{z}})$ . Nonetheless, implementing such a BayesNet can present challenges. For instance, normalization layers such as GDN [3], often integral to compression models, are difficult to integrate because the nodes are presumed to be mutually independent.", + "bbox": [ + 214, + 284, + 782, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Heterogeneous Parameterization of $p(\\mathbf{G}_{inter})$ We offer an innovative parameterization for the bipartite graph, tailored for heterogeneous DAG structure learning within image compression frameworks. Rather than dividing nodes, we maintain individual nodes in adjacent layers and diversify only the connecting edges. Each node pair is linked through multiple neural network edges of varying sizes and complexities, from which only one is selected. In this context, $\\mathbf{G}_{inter}$ is characterized by $N$ categorical distributions:", + "bbox": [ + 214, + 345, + 784, + 450 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}\\right) = \\operatorname {C a t e g o r i c a l} \\left(\\pi_ {\\mathbf {y}, \\mathbf {z}}\\right), \\quad \\pi_ {\\mathbf {y}, \\mathbf {z}} \\in \\mathbb {R} ^ {N}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 458, + 785, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thus, the connection $p(\\mathbf{y}|\\mathbf{z},\\mathbf{G}_{inter}^{\\mathbf{y},\\mathbf{z}})$ can be expressed as:", + "bbox": [ + 215, + 482, + 624, + 500 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {y} | \\mathbf {z}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {z}}) = \\sum_ {n} ^ {N} \\hat {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {z}} [ n ] p _ {n} (\\mathbf {y} | \\mathbf {z}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 507, + 785, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This heterogeneous DAG learning approach does not merely determine whether nodes are interdependent; it specifies how they are interconnected. Given the inherent dependencies among latent nodes in compression frameworks, this nuanced parameterization is particularly apt for representing inter-node BayesNets.", + "bbox": [ + 214, + 553, + 785, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation of the edges Once $\\mathbf{G}_{inter}$ is established, we proceed to construct the heterogeneous edges, represented as $p_n(\\mathbf{y}|\\mathbf{z})$ , utilizing neural networks. Following [29], we employ slimmable networks and select the $n$ -th channel width configuration for each $p_n(\\mathbf{y}|\\mathbf{z})$ . This approach is memory-efficient as it does not necessitate additional memory overhead compared to a singular network implementation for $p(\\mathbf{y}|\\mathbf{z})$ . A comprehensive example detailing this implementation is provided in Section 5 for further reference.", + "bbox": [ + 214, + 614, + 785, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimization of $p(\\mathbf{G}_{inter})$ We can leverage samples from $p(\\mathbf{G}_{inter})$ as weights to blend the various edges, enabling end-to-end optimization via the Gumbel-softmax reparameterization trick [14], eliminating the need for an additional loss function. The computational complexity associated with this optimization can be quantified using the same mixing weights:", + "bbox": [ + 214, + 719, + 785, + 795 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} \\left(\\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}\\right) = \\sum_ {n} ^ {N} \\hat {\\mathbf {G}} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}} [ n ] C \\left(p _ {n} (\\mathbf {y} | \\mathbf {z})\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 803, + 785, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1beb671ff72c3787c11bdd6d2a87208eee8e890c410e5fcb35cb57d7158a1ef5.jpg", + "table_caption": [ + "Table 1: Statistics of different intra-node BayesNet configurations for a $3 \\times 32 \\times 32$ tensor. Numbers inside the nodes represent the topological indices of the partites. For Maskconv each partite only include one node, so we omit the topological index." + ], + "table_footnote": [], + "table_body": "
MaskConv [24]Zigzag [20]Checkerboard [12]Multistage [21]Channel [25]
Stages102463243
FLOPs1171888352976117182048
Time (ms)11783.0727.346.582.671.2
", + "bbox": [ + 222, + 198, + 782, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $C(p_{m}(\\mathbf{y}|\\mathbf{z}))$ represents the complexity measure for a single edge such as FLOPs.", + "bbox": [ + 212, + 349, + 784, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Relation to Neural Architecture Search (NAS) Our approach to parameterizing $p(\\mathbf{G}_{inter})$ bears resemblance to Differentiable Architecture Search [22] in both concept and optimization techniques. Both methods utilize continuous relaxation to combine outputs from diverse neural network modules. However, the underlying objectives differ. NAS focuses on developing neural networks with enhanced performance by refining the computational graph. In contrast, our objective is to construct optimized probabilistic graphs tailored for compression, with neural networks serving as approximators of the conditional probabilities.", + "bbox": [ + 212, + 381, + 784, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Multipartite Graph for Intra-node BayesNet", + "text_level": 1, + "bbox": [ + 215, + 523, + 629, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As noted in Section 2.2, traditional BayesNet structure learning has been constrained by the need to maintain a DAG, which, when applied to image data with its vast number of pixels, can greatly reduce efficiency. Table 1 illuminates an alternative approach that can satisfy both acyclic and parallel constraints effectively. It reveals that the number of parallel stages corresponds to the multipartite quantity within each BayesNet. Since edges can only direct from topologically earlier nodes to later ones, nodes within the same partite can be processed simultaneously. Bayesian inference can thus be executed by sequentially processing each partite from the first to the last. This process not only maintains acyclicity but also confines the number of parallel stages to the number of partites. Hence, we propose to parameterize $p(\\mathbf{G}_{intra})$ with categorical distributions at each node, each category corresponding to the node's topological index within the partites. Consequently, $p(\\mathbf{G}_{intra})$ takes the following form:", + "bbox": [ + 212, + 547, + 787, + 746 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right) = \\prod_ {c, h, w} ^ {C, H, W} p \\left(\\mathrm {T} _ {c, h, w}\\right), \\quad \\mathrm {T} _ {c, h, w} = \\text {C a t e g o r i c a l} \\left(\\pi_ {c, h, w}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 755, + 785, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{T}_{c,h,w}$ denotes the topological indices of partites per categorical distribution, with $C,H,W$ being the node count along the channel, height, and width", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dimensions, respectively. Adjusting the number of partites (or parallel stages) $\\dim \\pi_{c,h,w} = S_{intra}$ can modulate the computation time.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Optimization of $p(\\mathbf{G}_{\\text {intra }})$ Having constrained the number of parties to manage computational time, we now focus on learning the intra-node BayesNet structure to minimize bit-rate. The optimization challenge is formulated as:", + "bbox": [ + 212, + 176, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {p \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right)} \\mathbb {E} _ {p \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right)} p \\left(\\mathbf {X}, \\mathbf {L} \\mid \\mathbf {G} _ {\\text {i n t r a}}\\right), \\quad \\text {s . t .} \\hat {\\mathbf {G}} _ {\\text {i n t r a}} \\text {i s a c y c l i c ,} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 231, + 782, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{G}}_{\\text {intra }}$ is sampled from $p(\\mathbf{G}_{\\text {intra }})$ . However, unlike the optimization for $\\mathbf{G}_{\\text {inter }}$ , partite indices are ordinal and don't lend themselves to being used as mixing weights, making the Gumbel-softmax reparameterization trick inapplicable. We therefore approach $\\mathbf{G}_{\\text {intra }}$ as discrete variables and employ the Monte-Carlo method with the VIMCO objective [26] for optimization:", + "bbox": [ + 214, + 265, + 784, + 340 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathbf {G} _ {i n t r a}} = \\mathbb {E} _ {q \\left(\\mathbf {G} _ {i n t r a} ^ {1: M} \\mid \\mathbf {X}\\right)} \\log \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\log p (\\mathbf {X}, \\mathbf {L} | \\mathbf {G} _ {i n t r a} ^ {i}), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 351, + 782, + 390 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{G}_{intra}^{i}$ is the i-th Monte-Carlo sample and $M$ denotes number of samples.", + "bbox": [ + 214, + 400, + 782, + 416 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Moreover, note that Equation 12 presumes all partites are independent—an unlikely scenario as proximal nodes often exhibit stronger correlations. In practice, we can approximate $p(\\mathbf{G}_{intra})$ with an unconditional generative model to consider such correlations and optimize this proxy instead, described as:", + "bbox": [ + 212, + 417, + 782, + 476 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right) \\propto g (\\mathbf {N}), \\quad \\mathbf {N} \\sim \\mathcal {N} (0, 1), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 486, + 782, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "with $g(\\mathbf{N})$ symbolizing a generative network derived from standard normal noise. Implementation of the Intra-node BayesNet With the model $p(\\mathbf{G}_{\\text{intra}})$ established, we can draw samples of topological indices for parties, denoted $\\hat{T}_{c,h,w}$ , from each respective $\\mathrm{T}_{c,h,w}$ . The subsequent step involves establishing conditional dependencies among these groups. Since dependencies are restricted to nodes from different parties, a dynamic masked convolution operation is utilized. This operation ensures each node's dependencies are limited to preceding nodes within the convolutional kernel's range. For instance, the intra-node BayesNet for $\\mathbf{y}$ can be characterized by:", + "bbox": [ + 214, + 510, + 784, + 646 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\prod_ {c, h, w} ^ {C, H, W} p \\left(\\mathrm {y} _ {c, h, w} \\mid \\mathbf {G} _ {\\text {i n t r a}}\\right) \\propto \\operatorname {d y n c o n v} \\left(\\hat {\\mathrm {y}} _ {c, h, w}, W _ {c, h, w} \\left(\\hat {T} _ {c, h, w}\\right)\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 655, + 782, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{y}}_{c,h,w}$ are samples from $\\mathrm{y}_{c,h,w}$ , and $W_{c,h,w}$ represents a dynamically adjusted convolutional kernel influenced by adjacent nodes' topological indices. As an illustration, consider a 2D $5\\times 5$ convolution: the dynamic kernel $W_{c,h,w}$ used for node $\\mathrm{y}_{c,h,w}$ is defined by:", + "bbox": [ + 212, + 705, + 782, + 767 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nW _ {c, h, w} [ i, j + 2, k + 2 ] = \\left\\{ \\begin{array}{l l} W _ {i, j + 2, k + 2} & , \\hat {T} _ {c, h, w} < \\hat {T} _ {i, h + j, w + k} \\\\ 0 & , \\hat {T} _ {c, h, w} \\geq \\hat {T} _ {i, h + j, w + k} \\end{array} , \\left\\{ \\begin{array}{l l} i & \\in \\{0, 1, \\dots , C - 1 \\} \\\\ j, k & \\in \\{- 2, - 1, 0, 1, 2 \\} \\end{array} . \\right. \\right. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 775, + 782, + 824 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For in-depth guidance of its implementation, please refer to Appendix.", + "bbox": [ + 214, + 825, + 705, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/18a6da2e6d528ffb3a32f95858e869156b4741c3c3f1867ceaf43d40964a58d8.jpg", + "image_caption": [ + "Fig. 2: The architecture of the implemented example based on the joint autoregressive framework [24]. Comp/Decomp means Compression / Decompression. AE/AD symbolize the arithmetic encoder and decoder, respectively. AR denotes the autoregressive model that executes the learned intra-node BayesNet. Q stands for quantization. $\\pmb{x}$ correspond to input image and $\\hat{\\pmb{x}}$ the reconstructed image." + ], + "image_footnote": [], + "bbox": [ + 250, + 142, + 754, + 333 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 From BayesNet to BaSIC", + "text_level": 1, + "bbox": [ + 215, + 441, + 501, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Upon determining the BayesNet structure, we proceed to specify the nodes and edges in this BayesNet, crafting a novel NIC framework, BaSIC. This section details an implementation instance of the envisioned computationally scalable NIC, leveraging the prevalent joint autoregressive and hyperprior compression model [24]. It should be noted that this represents just one of the viable strategies for implementing BaSIC, with the potential for alternatives to offer improved results. The architecture of our example is depicted in Figure 2, and we will explore how it seamlessly integrates with the aforementioned BayesNets. $^{4}$", + "bbox": [ + 212, + 477, + 787, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Inter-node BayesNet The inter-node BayesNet is constructed over each edge between nodes within the original BayesNet, specifically:", + "bbox": [ + 212, + 599, + 787, + 630 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q (\\mathbf {y} | \\mathbf {x}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {x}}) \\propto g _ {a} (\\mathbf {x}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {x}}), \\quad q (\\mathbf {z} | \\mathbf {y}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {z}, \\mathbf {y}}) \\propto h _ {a} (\\mathbf {y}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {z}, \\mathbf {y}}), \\\\ (1 1 1, \\hat {\\mathbf {x}} ^ {\\hat {\\mathbf {y}}} \\hat {\\mathbf {z}}), (1 1 1, \\hat {\\mathbf {z}} ^ {\\hat {\\mathbf {y}}} \\hat {\\mathbf {z}}). \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 640, + 712, + 667 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\np (\\hat {\\mathbf {y}} | \\hat {\\mathbf {z}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {y}}, \\hat {\\mathbf {z}}}) \\propto h _ {s} (\\hat {\\mathbf {z}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {y}}, \\hat {\\mathbf {z}}}), \\quad p (\\hat {\\mathbf {x}} | \\hat {\\mathbf {y}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {x}}, \\hat {\\mathbf {y}}}) \\propto g _ {s} (\\hat {\\mathbf {y}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {x}}, \\hat {\\mathbf {y}}})\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 662, + 705, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To achieve computational scalability, heterogeneous bipartite graphs are employed as outlined in Equation 10 for $\\mathbf{G}_{inter}^{\\mathbf{y},\\mathbf{x}}$ , $\\mathbf{G}_{inter}^{\\mathbf{z},\\mathbf{y}}$ , $\\mathbf{G}_{inter}^{\\hat{\\mathbf{y}},\\hat{\\mathbf{z}}}$ , and $\\mathbf{G}_{inter}^{\\hat{\\mathbf{x}},\\hat{\\mathbf{y}}}$ . The implementation of $g_{a}, g_{s}, h_{a}$ , and $h_{s}$ follow [29]'s slimmable networks, with dynamic channel widths including 48, 72, 96, 144, and 192 for computational scalability. Note that we use a fixed 192 channels for all latent variables $\\mathbf{y}, \\mathbf{z}$ , and only channels in the middle layers of each slimmable network are variable. This modification decouples bitrate scalability from computational scalability, enhancing the convenience in real-world applications.", + "bbox": [ + 212, + 693, + 787, + 814 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "Further details are provided in Appendix.", + "bbox": [ + 217, + 824, + 511, + 839 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Intra-node BayesNet In the established framework, the intra-node BayesNet employs masked convolution applied to $\\hat{\\mathbf{y}}$ . We substitute this with our suggested multipartite graph-based intra-node BayesNet $p(\\hat{\\mathbf{y}}|\\mathbf{G}_{intra}^{\\hat{\\mathbf{y}}})$ , which enables parallel computing on GPUs. The intra-node BayesNet uses a $5\\times 5$ dynamic masked convolution as detailed in Equation 17, with $C$ equals to number of partites and $H,W$ equals to image height and width, respectively. Note that to facilitate the optimization of the intra-node BayesNet, the additional VIMCO loss expressed in Equation 14 should be incorporated into the overall loss function.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Optimization The optimization process involves a loss function that simultaneously accounts for both the BayesNet structure (Equation 14) and the neural network parameters (Equation 4):", + "bbox": [ + 212, + 268, + 782, + 311 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {R} + \\lambda_ {D} \\mathcal {L} _ {D} + \\lambda_ {C} \\mathcal {L} _ {C} + \\mathcal {L} _ {\\mathbf {G} _ {\\text {i n t r a}}} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 328, + 784, + 342 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Computational Scalability The described implementation enables computational scalability across two dimensions. Firstly, by modulating all $\\mathbf{G}_{inter}$ , which govern the channel width in slimmable networks, we can fine-tune the computational load of the inter-node neural networks. $\\mathbf{G}_{inter}$ can be modulated during training through the $\\lambda_{C}$ parameter, or during inference by selecting from $\\mathbf{G}_{inter}$ to alter channel widths. Notably, the latter can be streamlined by training with a uniform $\\mathbf{G}_{inter}$ distribution and setting $\\lambda_{C} = 0$ . Secondly, by varying the number of partites in $\\mathbf{G}_{intra}^{\\hat{y}}$ , we can adjust the stages of parallel computations in autoregressive models, thereby influencing the decompression time. In practice, we could train multiple autoregressive models with varying stages and choose among them during inference to modify decompression times.", + "bbox": [ + 212, + 356, + 784, + 521 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 214, + 549, + 375, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 214, + 584, + 413, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experiment Tasks Following previous works [29], our experiments focus on the task of lossy image compression with computational scalability. We benchmark our BaSIC framework against other computational scalable frameworks, as well as conduct ablation studies to validate the efficacy of the proposed intranode BayesNet learning methods. Further experimental results and details can be found in Appendix, including compression performance evaluation and comparison under fixed computational constraints and more ablation studies regarding the inter-node BayesNet. For evaluation, we employ Pixel Signal-Noise Ratio (PSNR) for distortion and Bits Per Pixel (BPP) for bit-rate, unless stated otherwise.", + "bbox": [ + 212, + 612, + 782, + 760 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Datasets We use a merged training dataset from ImageNet [10] and CLIC [2], preprocessed following the method in [11]. The Kodak [1] dataset serves for validation and testing.", + "bbox": [ + 212, + 763, + 782, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details For the proposed method, we use the joint autoregressive framework [24] as the basis, as described in Section 5. Unless otherwise", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/37f1dc03aa2b1ae77866c639d05da9f06da593941c7ed1f3296224cdd07abfea.jpg", + "image_caption": [ + "Fig. 3: Comparative results for different computational scalable frameworks." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 488, + 292 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/08457f6224055bd3195c2cf2d1d289282646b78868c5913c862c42702b6c0fb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 145, + 763, + 289 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "specified, we follow CompressAI [5] to set the hyperparameter $\\lambda_{D}$ , while employing $\\lambda_{C} = 0$ during training and a greedy search among 16 intra-node BayesNets with varied FLOPs during inference. Training spans approximately 500k iterations, employing the Adam optimizer with a learning rate of 0.0001. To compare running times equitably, all experiments are performed on identical hardware (an Intel i7-6800K CPU, NVIDIA Titan Xp GPU, and 32GB of memory).", + "bbox": [ + 212, + 345, + 787, + 436 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.2 Comparison to Other Computation Scalable Frameworks", + "text_level": 1, + "bbox": [ + 214, + 462, + 730, + 478 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This section assesses our computational scalable BaSIC method against other scalable frameworks. We configure our framework without autoregressive modules, similar to SlimCAE [29], to compare performance at various computational levels. We adopt Hyperprior [4] as a reference, reporting BD-rate for diverse MACs for both methods. Also, the rate-distortion plots for SlimCAE's full range of complexity and three levels (L0, L7, L15) for our method are included for reference. Note that $\\lambda$ -scheduling is deactivated in SlimCAE to facilitate uniform BD-rate analysis across levels. The results are depicted in Figure 3.", + "bbox": [ + 212, + 491, + 787, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our method, employing the same slimmable network backbone as SlimCAE, demonstrates a marginal performance edge at most computation levels, likely attributable to the fixed channels of latent variables that enhance higher bit-rate performance. Specifically, our method saves around $2\\%$ more bits than Slim-CAE under the maximum complexity, and around $10\\%$ more bits under the minimum complexity. While SlimCAE [29] employs $\\lambda$ -scheduling to tackle this, it inadvertently ties bit-rate and distortion to computational complexity, which is less desirable for cases like low-powered devices needing high-quality images. In contrast, our inter-node BayesNet independently modulates the computational complexity of individual neural network modules, as evidenced by the rate-distortion plots that show minimal impact on performance with reduced computation. Moreover, the varied complexity levels indicate our method's superior control over computational complexity, allowing us to assign different inter-node BayesNets (channel widths) to distinct layers. SlimCAE, conversely, applies the same channel width across all layers, limiting flexibility. For instance,", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0d03adae42eec569e08e84cc415f652f6f63c3ecd1fc27570c64c391341ea509.jpg", + "table_caption": [ + "Table 2: Bit-rate comparison for different autoregressive modules." + ], + "table_footnote": [], + "table_body": "
Stages-2410
MethodsMaskConvBaselineCBCWOursMSOursELICOurs
BPP0.31500.33330.32160.32220.32070.32100.31830.31740.3156
Speed (MB/s)0.32345.55191.23173.46172.57107.9374.6717.8517.32
", + "bbox": [ + 246, + 171, + 754, + 233 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "to meet a 200kMAC/pixel threshold, our closest model uses 158kMAC/pixel, while SlimCAE's nearest option drops to 111kMAC/pixel, as shown in Figure 3.", + "bbox": [ + 212, + 262, + 782, + 292 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.3 Comparison to Other Autoregressive Models for Intra-Node BayesNet", + "text_level": 1, + "bbox": [ + 212, + 316, + 761, + 347 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we evaluate the effectiveness of the proposed intra-node BayesNet structure learning method, by comparing it to other hand-crafted autoregressive models. For a simple and fair comparison, we first establish a baseline using a joint autoregressive framework [24] with a null context model, trained for about 500k iterations. Subsequently, we retain all other components while training various autoregressive models to replace the baseline context model for an additional 500k iterations. This approach ensures the reconstructed image remains unchanged, allowing us to isolate and evaluate the bit-rate to reflect the effectiveness of different autoregressive models. Moreover, the decompression speed is also reported to evaluate the efficiency of It's important to note that the neural network architecture for the context model is kept constant across all methods, with the sole variability stemming from the structure of the intra-node BayesNet.", + "bbox": [ + 212, + 359, + 785, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The benchmark autoregressive models include Channel-wise (CW, 2-stage) [25], Checkerboard [12] (CB, 2-stage), Multi-stage spatial $2 \\times 2$ [21] (MS, 4-stage) and the combined unevenly grouped channel-wise and checkerboard [11] (ELIC, 10-stage in total). We also train the MaskConv context model from [24] as a baseline. Our model is trained under constraints of 2, 4, and 10 parallel stages and evaluated against counterparts with equivalent stage configurations. The findings are presented in Table 2.", + "bbox": [ + 212, + 541, + 787, + 647 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Overall, our learning-based BayesNet structures consistently demonstrate slightly better performance compared to the manually-crafted models under the same parallel stages, with similar decompression speed. Interestingly, in each group our learned method reaches close BPP to the hand-crafted ones in the next level of parallel stage, and in the 10-stage case, our method even approaches the BPP of non-parallel MaskConv model. This proves that our proposed method excels in learning effective intra-node BayesNets for autoregressive models.", + "bbox": [ + 212, + 648, + 787, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Results We also illustrate the learned intra-node BayesNet by our method as well as the manually-designed one of ELIC in Figure 4 to analyze their difference with hand-crafted models. We could see that the BayesNet of ELIC uses a checkerboard context model over the spatial dimension, including only two", + "bbox": [ + 212, + 753, + 787, + 814 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "5 More visualization results are shown in Appendix.", + "bbox": [ + 217, + 824, + 566, + 839 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c1faa6ec35b794b24ae4d04e1afa693041b07d6f2bde0e22603d802af43ef0fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 143, + 759, + 237 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a8b4921d5297f5a418bfcf70c08d0fc9cef573a0a4f74633ad210f16d3b6b3b5.jpg", + "image_caption": [ + "(a) ELIC", + "(b) Ours-10stage", + "Fig. 4: Partite topological indices visualization for intra-node BayesNet in the 10-stage case. Each image patch indicates topological indices of nodes in a group of channels. Blacker pixels indicate lower topological indices and whiter indicate higher topological indices. E.g., in the 10-stage case, pure-black indicates index 0 which is first iterated, and pure-white indicates group 9 which is last iterated. Better viewed with zoom-in." + ], + "image_footnote": [], + "bbox": [ + 243, + 250, + 759, + 343 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "partites within each channel. In comparison, our learned BayesNet incorporates more different partites (mostly three or four) in the spatial dimension, which captures more context information than the checkerboard case. This is reasonable as images are usually heavily correlated in spatial dimensions. Moreover, the learned topological indices also give some insight for designing parallel autoregressive models, that channel and spatial context models could be combined more compactly to improve their performance.", + "bbox": [ + 212, + 468, + 787, + 575 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusions and Discussions", + "text_level": 1, + "bbox": [ + 215, + 597, + 526, + 614 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents a novel approach, BaSIC, to addressing computational scalability in Neural Image Compression (NIC) by focusing on BayesNet structure learning. We decompose the challenge into two subproblems: heterogeneous bipartite graph learning for inter-node BayesNet, and multipartite graph learning for intra-node BayesNet, proposing solutions for each. Experimental results, using a joint autoregressive framework as a test case, show that our work offers improved control over computational demands relative to other scalable computation models.", + "bbox": [ + 212, + 628, + 787, + 750 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The paper is supported in part by the National Natural Science Foundation of China (No. 62325109, U21B2013).", + "bbox": [ + 215, + 176, + 785, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Kodak lossless true color image suite. https://r0k.us/graphics/kodak/ (2023), accessed: 2023-11-12", + "2. Workshop and challenge on learned image compression. https://click.compression.cc/ (2023), accessed: 2023-01-09", + "3. Balle, J., Laparra, V., Simoncelli, E.P.: Density modeling of images using a generalized normalization transformation. CoRR abs/1511.06281 (2015), https://apisemantic scholar.org/CorpusID:2684987", + "4. Balle, J., Minnen, D.C., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. ArXiv abs/1802.01436 (2018)", + "5. Bégaint, J., Racapé, F., Feltman, S., Pushparaja, A.: Compressai: a pytorch library and evaluation platform for end-to-end compression research. arXiv preprint arXiv:2011.03029 (2020)", + "6. Cheng, J., Greiner, R., Kelly, J., Bell, D.A., Liu, W.: Learning bayesian networks from data: An information-theory based approach. Artif. Intell. 137, 43-90 (2002), https://api-semanticscholar.org/CorpusID:7821347", + "7. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 7936-7945 (2020)", + "8. Chickering, D.M.: Optimal structure identification with greedy search. J. Mach. Learn. Res. 3, 507-554 (2003), https://api-semanticscholar.org/CorpusID:1191614", + "9. Deleu, T., G'ois, A., Emezue, C.C., Rankawat, M., Lacoste-Julien, S., Bauer, S., Bengio, Y.: Bayesian structure learning with generative flow networks. ArXiv abs/2202.13903 (2022), https://api(semanticscholar.org/CorpusID: 247158659", + "10. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition pp. 248-255 (2009)", + "11. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5708-5717 (2022), https://apisemantic scholar.org/CorpusID:247594672", + "12. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14766-14775 (2021), https://apisemantic scholar.org/CorpusID:232404386", + "13. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:326772" + ], + "bbox": [ + 218, + 258, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Jang, E., Gu, S.S., Poole, B.: Categorical reparameterization with gumbel-softmax. ArXiv abs/1611.01144 (2016)", + "15. Johnston, N., Eban, E., Gordon, A., Ball'e, J.: Computationally efficient neural image compression. ArXiv abs/1912.08771 (2019), https://apisemantic scholar.org/CorpusID:209404937", + "16. Kang, N., Qiu, S., Zhang, S., Li, Z., Xia, S.: Pilc: Practical image lossless compression with an end-to-endgpu oriented neural framework. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3729-3738 (2022), https://apisemantic scholar.org/CorpusID:249625569", + "17. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. CoRR abs/1312.6114 (2013)", + "18. Koivisto, M., Sood, K.: Exact bayesian structure discovery in bayesian networks. J. Mach. Learn. Res. 5, 549-573 (2004), https://api(semanticscholar.org/ CorpusID:12412451", + "19. Lee, J., Cho, S., Beack, S.: Context-adaptive entropy model for end-to-end optimized image compression. In: International Conference on Learning Representations (2018)", + "20. Li, M., Ma, K., You, J.J., Zhang, D., Zuo, W.: Efficient and effective context-based convolutional entropy modeling for image compression. IEEE Transactions on Image Processing 29, 5900-5911 (2019), https://api-semanticscholar.org/CorpusID:195345342", + "21. Lin, F., Sun, H., Liu, J., Katto, J.: Multistage spatial context models for learned image compression. ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 1-5 (2023), https://apisemantic scholar.org/CorpusID:257038187", + "22. Liu, H., Simonyan, K., Yang, Y.: Darts: Differentiable architecture search. ArXiv abs/1806.09055 (2018), https://api-semanticscholar.org/CorpusID: 49411844", + "23. Lorch, L., Rothfuss, J., Scholkopf, B., Krause, A.: Dibs: Differentiable bayesian structure learning. ArXiv abs/2105.11839 (2021), https://apisemantic scholar.org/CorpusID:235187432", + "24. Minnen, D.C., Balle, J., Toderici, G.: Joint autoregressive and hierarchical priors for learned image compression. ArXiv abs/1809.02736 (2018)", + "25. Minnen, D.C., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. 2020 IEEE International Conference on Image Processing (ICIP) pp. 3339-3343 (2020), https://api-semanticscholar.org/CorpusID:220633468", + "26. Mnih, A., Rezende, D.J.: Variational inference for monte carlo objectives. In: International Conference on Machine Learning (2016), https://api-semanticscholar.org/CorpusID:5859948", + "27. Tao, L., Gao, W., Li, G., Zhang, C.: Adanic: Towards practical neural image compression via dynamic transform routing. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 16833-16842 (2023), https://api.sementicscholar.org/CorpusID:267026318", + "28. webmproject: libwebp. https://github.com/webmproject/libwebp (2023), accessed: 2023-01-09", + "29. Yang, F., Herranz, L., Cheng, Y., Mozerov, M.G.: Slimmable compressive autoencoders for practical neural image compression. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 4996-5005 (2021), https://api-semanticscholar.org/CorpusID:232417114" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Yin, S., Li, C., Bao, Y., Liang, Y.: Universal efficient variable-rate neural image compression. ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 2025-2029 (2021), https://api.sementicscholar.org/CorpusID:244478380", + "31. Yin, S., Meng, F., Tan, W.T., Li, C., Bao, Y., Liang, Y., Liu, W.: Exploring structural sparsity in neural image compression. 2022 IEEE International Conference on Image Processing (ICIP) pp. 471-475 (2022), https://api-semanticscholar.org/CorpusID:246680031", + "32. Yu, J., Yang, L., Xu, N., Yang, J., Huang, T.S.: Slimmable neural networks. ArXiv abs/1812.08928 (2018), https://api(semanticscholar.org/CorpusID:56657799", + "33. Yu, Y., Chen, J., Gao, T., Yu, M.: Dag-gnn: Dag structure learning with graph neural networks. In: International Conference on Machine Learning (2019), https://apisemantic scholar.org/CorpusID:128358697", + "34. Yuan, C., Malone, B.M.: Learning optimal bayesian networks: A shortest path perspective. J. Artif. Intell. Res. 48, 23-65 (2013), https://api-semanticscholar.org/CorpusID:5951724", + "35. Zhang, Y., Lu, G., Feng, D., Zhu, C., Song, L.: Content adaptive checkerboard context model for learned image compression. 2023 IEEE International Symposium on Circuits and Systems (ISCAS) pp. 1-5 (2023), https://api-semanticscholar.org/CorpusID:260002843", + "36. Zhang, Z., Chen, B., Lin, H., Lin, J., Wang, X., Zhao, T.: Elfic: A learning-based flexible image codec with rate-distortion-complexity optimization. Proceedings of the 31st ACM International Conference on Multimedia (2023), https://apisemantic scholar.org/CorpusID:264492165", + "37. Zheng, X., Aragam, B., Ravikumar, P., Xing, E.P.: Dags with no tears: Continuous optimization for structure learning. In: Neural Information Processing Systems (2018), https://api_semanticscholar.org/CorpusID:53217974", + "38. Zheng, Z., Wang, X., Lin, X., Lv, S.: Get the best of the three worlds: Real-time neural image compression in a non-gpu environment. Proceedings of the 29th ACM International Conference on Multimedia (2021), https://api-semanticscholar.org/CorpusID:239011955" + ], + "bbox": [ + 212, + 146, + 787, + 590 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/360699b9-a034-4ed5-b537-07332b36832e_model.json b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/360699b9-a034-4ed5-b537-07332b36832e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a9dc2d0edf9a12190dfbaa682d86696ebb92942 --- /dev/null +++ b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/360699b9-a034-4ed5-b537-07332b36832e_model.json @@ -0,0 +1,2302 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.271, + 0.141, + 0.734, + 0.21 + ], + "angle": 0, + "content": "BaSIC: BayesNet Structure Learning for Computational Scalable Neural Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.235, + 0.674, + 0.266 + ], + "angle": 0, + "content": "Yufeng Zhang\\(^{1}\\), Hang Yu\\(^{2}\\), Shizhan Liu\\(^{1}\\), Wenrui Dai\\(^{1}\\), and Weiyao Lin\\(^{\\star 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.277, + 0.699, + 0.321 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Jiao Tong University, China. {worldlife,shanluzuode,daiwenrui,wylin}@sjtu.edu.cn \\(^{2}\\) Ant Group, China. hyu.hugo@antgroup.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.36, + 0.74, + 0.623 + ], + "angle": 0, + "content": "Abstract. Despite superior rate-distortion performance over traditional codecs, Neural Image Compression (NIC) is limited by its computational scalability in practical deployment. Prevailing research focuses on accelerating specific NIC modules but is restricted in controlling overall computational complexity. To this end, this work introduces BaSIC (BayesNet structure learning for computational Scalable neural Image Compression), a comprehensive, computationally scalable framework that affords full control over NIC processes. We learn the Bayesian network (BayesNet) structure of NIC for controlling both neural network backbones and autoregressive units. The learning of BayesNet is achieved by solving two sub-problems, i.e., learning a heterogeneous bipartite BayesNet for the inter-node structure to regulate backbone complexity, and a multipartite BayesNet for the intra-node structure to optimize parallel computation in autoregressive units. Experiments demonstrate that our method not only facilitates full computational scalability with more accurate complexity control but also maintains competitive compression performance compared to other computation scalable frameworks under equivalent computational constraints. Code will be available in https://github.com/worldlife123/cbench_BaSIC." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.637, + 0.74, + 0.665 + ], + "angle": 0, + "content": "Keywords: Image Compression \\(\\cdot\\) Computational Scalable Compression \\(\\cdot\\) BayesNet Structure Learning \\(\\cdot\\) Autoregressive Model" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.692, + 0.377, + 0.708 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Neural Image Compression (NIC) technologies [7, 11] have significantly outstripped the performance of conventional image CODECs, ushering in an era of intelligent image compression. These advanced CODEcs are intended for deployment across a spectrum of devices, from low-end devices like mobile phones to high-end devices like GPU computation clusters. Each type of device necessitates that CODEcs tailor their performance to specific computational or memory" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.39, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.143, + 0.468, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.153, + 0.768, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.358, + 0.785, + 0.387 + ], + "angle": 0, + "content": "Fig. 1: Bayesian network diagrams for various NIC frameworks. Our proposed framework BaSIC (d) generalizes the concepts in (a), (b), and (c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.784, + 0.476 + ], + "angle": 0, + "content": "constraints, a requirement known as Computational Scalability. High-end devices can endure more intricate computations for superior encoding quality, whereas low-end devices may prioritize reduced computational load over encoding efficiency to conserve energy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.785, + 0.688 + ], + "angle": 0, + "content": "While conventional image CODECs typically accommodate this functionality (e.g. WebP [28] offers seven levels of computational complexity to regulate lossy image compression), incorporating this attribute into NIC remains a significant hurdle for its broader adoption. Recent studies have focused predominantly on accelerating NIC computations rather than addressing computational scalability. Works such as [11, 12, 25] target the acceleration of autoregressive modules on GPUs, which are frequently used in NIC. Others, including [15, 31, 38], explore neural network acceleration techniques—given that NIC fundamentally relies on neural network architectures. On the other hand, some recent efforts have attempted to integrate dynamic neural networks like slimmable networks [32] within NIC frameworks to adjust the computational complexity [27, 29, 36]. However, these approaches have only partially addressed computational scalability and often neglect crucial components in NIC, such as autoregressive modules, which are essential for compression performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this work, we endeavor to create a NIC framework with enhanced computational scalability, incorporating computational control across all NIC components. We call this framework BaSIC (BayesNet structure learning for computational Scalable neural Image Compression). To this end, we re-examine NIC through the broader lens of the Bayesian network (BayesNet) paradigm. Figure 1 illustrates the BayesNet depiction of several exemplary NIC architectures. Specifically, our approach is rooted in the hyperprior BayesNet framework (Figure 1a), which postulates that the generative model can be factorized as \\( p(\\mathbf{x}|\\mathbf{y})p(\\mathbf{y}|\\mathbf{z})p(\\mathbf{z}) \\), where \\( \\mathbf{x} \\) represents observed data, \\( \\mathbf{y} \\) the prior, and \\( \\mathbf{z} \\) the hyperprior. In this context, SlimCAE (Figure 1b) can be interpreted as allowing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "selective edge types between \\(\\mathbf{y}\\) and \\(\\mathbf{x}\\), and between \\(\\hat{\\mathbf{x}}\\) and \\(\\hat{\\mathbf{y}}\\), with each type manifesting different conditional dependencies, modeled via slimmable neural networks with varied channel widths. On the other hand, the joint autoregressive framework (Figure 1c) introduces autoregressive intra-node dependencies within \\(\\hat{y}\\) to enhance compression, at the expense of increased computational demands due to the requisite sequential processing [12]. These examples clearly demonstrate the intimate connection between the structure of the BayesNet and the complexity inherent in NIC frameworks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.789, + 0.543 + ], + "angle": 0, + "content": "Our advancement is a NIC framework that exploits end-to-end BayesNet structure learning [23, 33] to meticulously optimize computational scalability. This innovative framework is not bound to specific neural network backbones, making it versatile enough to enhance computational scalability in most existing NIC frameworks with a hyperprior BayesNet representation. Our proposed framework is visualized in Figure 1d. In particular, we represent the BayesNet structure as \\(\\mathbf{G}\\) and impose a prior \\(p(\\mathbf{G})\\) on it. Given that the structure we seek to learn must conform to a directed acyclic graph (DAG), we tailor the prior \\(p(\\mathbf{G})\\) for compression frameworks within this DAG constraint, considering two distinct instances: the inter-node BayesNet, representing conditional dependencies between nodes, and the intra-node BayesNet, which pertains to dependencies within nodes. We depict the former as a heterogeneous bipartite graph and design a structure learning method to seamlessly incorporate dynamic neural networks like slimmable networks [32]. For the latter, we design a network architecture parameterized by multipartite or \\(k\\)-partite graphs to enable a parallel autoregressive process. By manipulating the partite number \\(k\\), we can adjust the temporal demands of the autoregressive process, thereby directly impacting decompression speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.79, + 0.653 + ], + "angle": 0, + "content": "When deploying the codec, the configurations of both the inter-node and intra-node BayesNets can be tailored to match the computational budget of target devices. For instance, on less capable devices, we might employ simpler neural network edges within the inter-node BayesNet and a smaller partite number within the intra-node BayesNet, whereas more capable devices can leverage a more complex network and a larger partite number, allowing for scalable computational resource allocation based on the specifics of the device at hand." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.657, + 0.637, + 0.672 + ], + "angle": 0, + "content": "To summarize, our contributions are mainly three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.697, + 0.787, + 0.742 + ], + "angle": 0, + "content": "- We propose an end-to-end computationally scalable NIC framework founded on BayesNet structure learning, facilitating fine-tuned adjustments to computational complexity across all facets of the NIC process." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.746, + 0.787, + 0.79 + ], + "angle": 0, + "content": "- We devise a generative model based on a heterogeneous bipartite structure for the inter-node BayesNet, harnessing the capabilities of existing dynamic neural networks to realize computational scalability effectively." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We design a generative model structured as a multipartite graph for the intra-node BayesNet. This model is tailored to accommodate and streamline parallel computation in practical applications." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.697, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.231, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.176, + 0.785, + 0.207 + ], + "angle": 0, + "content": "In this section, we briefly review recent works regarding the computational complexity of NIC as well as BayesNet structure learning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.228, + 0.584, + 0.243 + ], + "angle": 0, + "content": "2.1 End-to-end Neural Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.251, + 0.787, + 0.357 + ], + "angle": 0, + "content": "The hyperprior framework [4] serves as a foundation for most end-to-end NIC models, favored for its effective compression performance and straightforward implementation. Subsequent studies [7, 19, 24] have enhanced the hyperprior by integrating a masked convolution-based autoregressive model in the latent space to reduce bit-rate. While these improvements offer better compression, they often overlook the increased computational demands. To apply NIC in practice, some recent works are now addressing the need to expedite these models." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.358, + 0.787, + 0.522 + ], + "angle": 0, + "content": "Autoregressive Model Acceleration To circumvent the slow serial iteration steps of masked convolutions on GPUs, newer studies introduce parallelized autoregressive models for GPU acceleration. For instance, [25] presents a channelwise parallel autoregressive model, while [12] proposes a spatially parallel model using a checkerboard mask, requiring just two parallel stages. This concept is extended to a multistage model by [21]. Further, [11] amalgamates channel-wise and spatially parallel autoregressive models. Additional acceleration methods include content-adaptive adjustments to parallel autoregressive models [35] and designing simplified autoregressive models [16]. Yet, such models are typically hand-crafted and may not represent the most efficient trade-off between compression complexity and performance." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.524, + 0.787, + 0.629 + ], + "angle": 0, + "content": "Neural Network Acceleration Considering that neural networks (i.e., edges in the BayesNets) represent a significant portion of NIC's computational complexity, some works focus on accelerating these directly. For example, there are designs for more efficient neural network operators like a simplified GDN [15] and CPU-friendly down/up-samplers [38]. Others apply general neural network acceleration techniques within compression models, such as MorphNet [15] and learnable channel pruning [30, 31]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.63, + 0.787, + 0.75 + ], + "angle": 0, + "content": "Computation Scalable Framework A handful of recent works have also noticed the importance of computational scalability for NIC. SlimCAE [29] employs slimmable networks [32] as the framework's backbone, enabling dynamic adjustments of bit-rate and computation through channel width modulation. Further developments include universal slimmable networks for more precise scalability [36] and the integration of input-adaptive complexity allocation methods [27,36]. Nonetheless, these approaches focus solely on the scalability of neural networks, largely disregarding the autoregressive models' scalability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.504, + 0.787 + ], + "angle": 0, + "content": "2.2 BayesNet Structure Learning" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "BayesNet structure learning aims to discern the conditional relationships among a set of variables with a graph representation, which ideally should be a directed acyclic graph (DAG). Traditional approaches often involve score-based [8, 34]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "and constraint-based methods [6, 18]. However, they typically operate within a constrained search space, which can lead to less-than-optimal graphs. More recent initiatives introduce neural network techniques for more robust graph generation within the realm of BayesNet structure learning. Some align this learning task with variational inference to facilitate graph structure development within an end-to-end framework [23, 33], utilizing graph acyclicity regularizers [37] that may not guarantee acyclicity. Alternatively, other research employs recurrent-style graph generators that assure acyclicity through iterations [9], but these can become inefficient with an increase in the number of nodes. In our work, we diverge from these established methods due to their inefficiency when dealing with the dense pixelation of images. Instead, we analyze the BayesNet representations inherent in existing NIC frameworks and directly parameterize the graph structure as DAGs, ensuring both acyclicity and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.368, + 0.441, + 0.384 + ], + "angle": 0, + "content": "3 Overview of BaSIC" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.789, + 0.492 + ], + "angle": 0, + "content": "Assume we have a set of observations \\(\\mathbf{X} = \\{\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_N\\}^3\\), the objective of BayesNet structure learning is to infer a distribution over the structure of the BayesNet \\(\\mathbf{G} = (\\mathbf{V}, \\mathbf{E}, \\boldsymbol{\\theta})\\) that best describes the observations. According to the Bayesian formalism, such optimization problem could be solved by minimizing expectation over \\(p(\\mathbf{G}|\\mathbf{X})\\) of any function \\(f\\) related to \\(\\mathbf{G}\\), under the constraint that \\(\\mathbf{G}\\) is acyclic:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.505, + 0.787, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\underset {\\mathbf {G}} {\\min } \\mathbb {E} _ {p (\\mathbf {G} | \\mathbf {X})} [ f (\\mathbf {G}) ], \\quad \\mathrm {s . t .} \\mathbf {G} \\text {i s a c y c l i c .} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.785, + 0.583 + ], + "angle": 0, + "content": "For a computationally scalable compression framework, we tackle BayesNet structure learning, linking BayesNet configurations to computational complexity and shaping the function \\( f(\\mathbf{G}) \\) accordingly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Apart from modeling of \\(\\mathbf{X}\\), the compression framework also involves latent variables \\(\\mathbf{L} = \\{\\mathbf{y}_1, \\mathbf{y}_2, \\ldots, \\mathbf{y}_N, \\mathbf{z}_1, \\mathbf{z}_2, \\ldots, \\mathbf{z}_N, \\ldots\\}\\) to generate \\(\\mathbf{X}\\). The resulting BayesNet should capture conditional dependencies within and between \\(\\mathbf{X}\\) and \\(\\mathbf{L}\\). This generative relationship is described by:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.657, + 0.787, + 0.673 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {G}, \\mathbf {L}, \\mathbf {X}) = p (\\mathbf {G}) p (\\mathbf {L} | \\mathbf {G}) p (\\mathbf {X} | \\mathbf {G}, \\mathbf {L}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.684, + 0.518, + 0.699 + ], + "angle": 0, + "content": "The expectation from Equation 1 is then:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.713, + 0.787, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {p (\\mathbf {G}, \\mathbf {L} | \\mathbf {X})} [ f (\\mathbf {G}, \\mathbf {L}) ] = \\mathbb {E} _ {p (\\mathbf {G} | \\mathbf {X}) p (\\mathbf {L} | \\mathbf {X}, \\mathbf {G})} [ f (\\mathbf {G}, \\mathbf {L}) ]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.788, + 0.788 + ], + "angle": 0, + "content": "Existing compression frameworks without \\(\\mathbf{G}\\) approximate \\(p(\\mathbf{L}|\\mathbf{X})\\) with an inference network \\(q(\\mathbf{L}|\\mathbf{X})\\), as in VAEs [17]. Incorporating \\(\\mathbf{G}\\) into our design, we refine our inference network to \\(q(\\mathbf{G},\\mathbf{L}|\\mathbf{X})\\) for enhanced approximation:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.799, + 0.787, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {p (\\mathbf {G}, \\mathbf {L} | \\mathbf {X})} [ f (\\mathbf {G}, \\mathbf {L}) ] \\approx \\mathbb {E} _ {q (\\mathbf {G} | \\mathbf {X}) q (\\mathbf {L} | \\mathbf {X}, \\mathbf {G})} [ f (\\mathbf {G}, \\mathbf {L}) ]. \\tag {4}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.595, + 0.841 + ], + "angle": 0, + "content": "3 All symbols used in this paper are listed in Appendix." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "This approach allows \\(\\mathbf{G}\\) to direct the computational complexity of both the inference and the generative processes. Furthermore, we adopt a differentiable approach to learning \\(\\mathbf{G}\\), in line with recent advances [23, 33], and incorporate this aspect into our model specification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.784, + 0.239 + ], + "angle": 0, + "content": "Definition of \\( f(\\mathbf{G}, \\mathbf{L}) \\) In line with lossy image compression standards [15,31], our optimization combines bit-rate \\( \\mathcal{L}_R \\), distortion \\( \\mathcal{L}_D \\), and complexity \\( \\mathcal{L}_C \\):" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.251, + 0.786, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {q (\\mathbf {G} | \\mathbf {X}) q (\\mathbf {L} | \\mathbf {X}, \\mathbf {G})} [ f (\\mathbf {G}, \\mathbf {L}) ] = \\mathcal {L} _ {R} (\\mathbf {G}, \\mathbf {L}) + \\lambda_ {D} \\mathcal {L} _ {D} (\\mathbf {G}, \\mathbf {L}) + \\lambda_ {C} \\mathcal {L} _ {C} (\\mathbf {G}, \\mathbf {L}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.279, + 0.784, + 0.309 + ], + "angle": 0, + "content": "Concretely, \\(\\mathcal{L}_R\\) is computed by enacting the BayesNet's generative process with latent node samples \\(\\hat{\\mathbf{L}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.321, + 0.786, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} (\\mathbf {G}, \\mathbf {L}) = \\log_ {2} p _ {\\mathbf {L} | \\mathbf {G}} (\\hat {\\mathbf {L}} | \\mathbf {G}), \\quad \\hat {\\mathbf {L}} \\sim q (\\mathbf {L} | \\mathbf {X}, \\mathbf {G}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.785, + 0.485 + ], + "angle": 0, + "content": "Moreover, \\(\\mathcal{L}_D\\) and \\(\\mathcal{L}_C\\) can be tailored to specific optimization objectives. \\(\\mathcal{L}_D\\) may encompass a range of complexity measures or their amalgamations, such as conventional metrics like Mean-Squared Error (MSE) or more advanced machine learning metrics like Frechet Inception Distance (FID) [13]. Conversely, \\(\\mathcal{L}_C\\) can include various complexity metrics or their combinations, such as the number of Floating-point Operations (FLOPs), network parameters, time, or energy consumption. \\(\\lambda_{D}\\) and \\(\\lambda_{C}\\) are the corresponding weights. By setting \\(f(\\mathbf{G},\\mathbf{L})\\) in line with the established rate-distortion (R-D) trade-off, Equation 4 falls in step with loss functions prevalent in standard compression frameworks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.787, + 0.683 + ], + "angle": 0, + "content": "Definition of \\( q(\\mathbf{G}|\\mathbf{X}) \\) As a determiner of the BayesNet structure, \\( \\mathbf{G} \\) affects computational complexity. In practical compression scenarios, including \\( \\mathbf{G} \\) in the bitstream could impact the bit-rate. Typically, it may be more advantageous to treat \\( \\mathbf{G} \\) as independent of \\( \\mathbf{X} \\), suggesting \\( q(\\mathbf{G}|\\mathbf{X}) = p(\\mathbf{G}) \\). However, in scenarios where complexity is tailored to input data, the model \\( q(\\mathbf{G}|\\mathbf{X}) \\) can be employed. DAG Constraint of \\( p(\\mathbf{G}) \\) As previously noted in Equation 1, \\( p(\\mathbf{G}) \\) should reflect a distribution over DAG structures. For edges between different node groups in compression models, such as \\( p(\\mathbf{x}|\\mathbf{y}) \\) or \\( p(\\mathbf{y}|\\mathbf{z}) \\), ensuring a DAG structure is straightforward since bipartite graphs are inherently acyclic; we call this component the Inter-node BayesNet (\\( \\mathbf{G}_{inter} \\)). For conditional dependencies within node groups, like \\( p(\\mathrm{y}_i|\\mathrm{y}_{i-1}) \\) which also require DAG adherence, we refer to it as the Intra-node BayesNet (\\( \\mathbf{G}_{intra} \\)). The forthcoming sections will delve into the methodology for learning both \\( p(\\mathbf{G}_{inter}) \\) and \\( p(\\mathbf{G}_{intra}) \\) as DAGs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.706, + 0.784, + 0.723 + ], + "angle": 0, + "content": "4 Structure Learning for Inter and Intra-Node BayesNets" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.739, + 0.734, + 0.755 + ], + "angle": 0, + "content": "4.1 Heterogeneous Bipartite Graphs for Inter-node BayesNet" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.794 + ], + "angle": 0, + "content": "In this section, we specify \\( p(\\mathbf{G}_{inter}) \\). Note that \\( \\mathbf{G}_{inter} \\) is a bipartite graph, and so it inherently satisfies the acyclic constraint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Naive Bipartite Parameterization of \\( p(\\mathbf{G}_{inter}) \\) Consider the task of learning a bipartite graph for \\( p(\\mathbf{y}|\\mathbf{z}) \\). We split the nodes into \\( \\mathbf{z}_1, \\mathbf{z}_2, \\ldots, \\mathbf{z}_I \\) and \\( \\mathbf{y}_1, \\mathbf{y}_2, \\ldots, \\mathbf{y}_J \\). Bipartite graphs can be represented by an \\( I \\times J \\) adjacency matrix" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.786, + 0.177 + ], + "angle": 0, + "content": "with each element indicating an edge between \\( \\mathbf{y} \\) and \\( \\mathbf{z} \\). Consequently, \\( \\mathbf{G}_{inter} \\) can be modeled as \\( I \\times J \\) Bernoulli distributions:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.185, + 0.786, + 0.203 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}\\right) = \\operatorname {B e r n o u l l i} \\left(\\pi_ {\\mathbf {y}, \\mathbf {z}}\\right), \\quad \\pi_ {\\mathbf {y}, \\mathbf {z}} \\in \\mathbb {R} ^ {I \\times J}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.208, + 0.525, + 0.225 + ], + "angle": 0, + "content": "The edge \\(p(\\mathbf{y}|\\mathbf{z},\\mathbf{G}_{inter}^{\\mathbf{y};\\mathbf{z}})\\) is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.233, + 0.786, + 0.276 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {y} | \\mathbf {z}, \\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}) = \\prod_ {i, j} ^ {I, J} p (\\mathbf {y} _ {j} | \\mathbf {z} _ {i}) \\hat {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}} [ i, j ], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.285, + 0.784, + 0.346 + ], + "angle": 0, + "content": "where \\(\\hat{G}_{inter}^{\\mathbf{y},\\mathbf{z}}\\) are samples from \\(p(\\mathbf{G}_{inter}^{\\mathbf{y},\\mathbf{z}})\\). Nonetheless, implementing such a BayesNet can present challenges. For instance, normalization layers such as GDN [3], often integral to compression models, are difficult to integrate because the nodes are presumed to be mutually independent." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.347, + 0.785, + 0.451 + ], + "angle": 0, + "content": "Heterogeneous Parameterization of \\( p(\\mathbf{G}_{inter}) \\) We offer an innovative parameterization for the bipartite graph, tailored for heterogeneous DAG structure learning within image compression frameworks. Rather than dividing nodes, we maintain individual nodes in adjacent layers and diversify only the connecting edges. Each node pair is linked through multiple neural network edges of varying sizes and complexities, from which only one is selected. In this context, \\( \\mathbf{G}_{inter} \\) is characterized by \\( N \\) categorical distributions:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.459, + 0.786, + 0.477 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}\\right) = \\operatorname {C a t e g o r i c a l} \\left(\\pi_ {\\mathbf {y}, \\mathbf {z}}\\right), \\quad \\pi_ {\\mathbf {y}, \\mathbf {z}} \\in \\mathbb {R} ^ {N}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.483, + 0.625, + 0.501 + ], + "angle": 0, + "content": "Thus, the connection \\(p(\\mathbf{y}|\\mathbf{z},\\mathbf{G}_{inter}^{\\mathbf{y},\\mathbf{z}})\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.508, + 0.786, + 0.548 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {y} | \\mathbf {z}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {z}}) = \\sum_ {n} ^ {N} \\hat {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {z}} [ n ] p _ {n} (\\mathbf {y} | \\mathbf {z}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.554, + 0.786, + 0.614 + ], + "angle": 0, + "content": "This heterogeneous DAG learning approach does not merely determine whether nodes are interdependent; it specifies how they are interconnected. Given the inherent dependencies among latent nodes in compression frameworks, this nuanced parameterization is particularly apt for representing inter-node BayesNets." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.615, + 0.786, + 0.719 + ], + "angle": 0, + "content": "Implementation of the edges Once \\(\\mathbf{G}_{inter}\\) is established, we proceed to construct the heterogeneous edges, represented as \\(p_n(\\mathbf{y}|\\mathbf{z})\\), utilizing neural networks. Following [29], we employ slimmable networks and select the \\(n\\)-th channel width configuration for each \\(p_n(\\mathbf{y}|\\mathbf{z})\\). This approach is memory-efficient as it does not necessitate additional memory overhead compared to a singular network implementation for \\(p(\\mathbf{y}|\\mathbf{z})\\). A comprehensive example detailing this implementation is provided in Section 5 for further reference." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.72, + 0.786, + 0.796 + ], + "angle": 0, + "content": "Optimization of \\( p(\\mathbf{G}_{inter}) \\) We can leverage samples from \\( p(\\mathbf{G}_{inter}) \\) as weights to blend the various edges, enabling end-to-end optimization via the Gumbel-softmax reparameterization trick [14], eliminating the need for an additional loss function. The computational complexity associated with this optimization can be quantified using the same mixing weights:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.804, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} \\left(\\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}\\right) = \\sum_ {n} ^ {N} \\hat {\\mathbf {G}} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}} [ n ] C \\left(p _ {n} (\\mathbf {y} | \\mathbf {z})\\right), \\tag {11}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Statistics of different intra-node BayesNet configurations for a \\(3 \\times 32 \\times 32\\) tensor. Numbers inside the nodes represent the topological indices of the partites. For Maskconv each partite only include one node, so we omit the topological index." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.199, + 0.784, + 0.322 + ], + "angle": 0, + "content": "
MaskConv [24]Zigzag [20]Checkerboard [12]Multistage [21]Channel [25]
Stages102463243
FLOPs1171888352976117182048
Time (ms)11783.0727.346.582.671.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.785, + 0.38 + ], + "angle": 0, + "content": "where \\( C(p_{m}(\\mathbf{y}|\\mathbf{z})) \\) represents the complexity measure for a single edge such as FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.503 + ], + "angle": 0, + "content": "Relation to Neural Architecture Search (NAS) Our approach to parameterizing \\( p(\\mathbf{G}_{inter}) \\) bears resemblance to Differentiable Architecture Search [22] in both concept and optimization techniques. Both methods utilize continuous relaxation to combine outputs from diverse neural network modules. However, the underlying objectives differ. NAS focuses on developing neural networks with enhanced performance by refining the computational graph. In contrast, our objective is to construct optimized probabilistic graphs tailored for compression, with neural networks serving as approximators of the conditional probabilities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.524, + 0.63, + 0.54 + ], + "angle": 0, + "content": "4.2 Multipartite Graph for Intra-node BayesNet" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.747 + ], + "angle": 0, + "content": "As noted in Section 2.2, traditional BayesNet structure learning has been constrained by the need to maintain a DAG, which, when applied to image data with its vast number of pixels, can greatly reduce efficiency. Table 1 illuminates an alternative approach that can satisfy both acyclic and parallel constraints effectively. It reveals that the number of parallel stages corresponds to the multipartite quantity within each BayesNet. Since edges can only direct from topologically earlier nodes to later ones, nodes within the same partite can be processed simultaneously. Bayesian inference can thus be executed by sequentially processing each partite from the first to the last. This process not only maintains acyclicity but also confines the number of parallel stages to the number of partites. Hence, we propose to parameterize \\( p(\\mathbf{G}_{intra}) \\) with categorical distributions at each node, each category corresponding to the node's topological index within the partites. Consequently, \\( p(\\mathbf{G}_{intra}) \\) takes the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.756, + 0.786, + 0.801 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right) = \\prod_ {c, h, w} ^ {C, H, W} p \\left(\\mathrm {T} _ {c, h, w}\\right), \\quad \\mathrm {T} _ {c, h, w} = \\text {C a t e g o r i c a l} \\left(\\pi_ {c, h, w}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathrm{T}_{c,h,w}\\) denotes the topological indices of partites per categorical distribution, with \\(C,H,W\\) being the node count along the channel, height, and width" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "dimensions, respectively. Adjusting the number of partites (or parallel stages) \\(\\dim \\pi_{c,h,w} = S_{intra}\\) can modulate the computation time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.223 + ], + "angle": 0, + "content": "Optimization of \\( p(\\mathbf{G}_{\\text {intra }}) \\) Having constrained the number of parties to manage computational time, we now focus on learning the intra-node BayesNet structure to minimize bit-rate. The optimization challenge is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.232, + 0.784, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\min _ {p \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right)} \\mathbb {E} _ {p \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right)} p \\left(\\mathbf {X}, \\mathbf {L} \\mid \\mathbf {G} _ {\\text {i n t r a}}\\right), \\quad \\text {s . t .} \\hat {\\mathbf {G}} _ {\\text {i n t r a}} \\text {i s a c y c l i c ,} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.266, + 0.785, + 0.342 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{G}}_{\\text {intra }}\\) is sampled from \\(p(\\mathbf{G}_{\\text {intra }})\\). However, unlike the optimization for \\(\\mathbf{G}_{\\text {inter }}\\), partite indices are ordinal and don't lend themselves to being used as mixing weights, making the Gumbel-softmax reparameterization trick inapplicable. We therefore approach \\(\\mathbf{G}_{\\text {intra }}\\) as discrete variables and employ the Monte-Carlo method with the VIMCO objective [26] for optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.352, + 0.784, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathbf {G} _ {i n t r a}} = \\mathbb {E} _ {q \\left(\\mathbf {G} _ {i n t r a} ^ {1: M} \\mid \\mathbf {X}\\right)} \\log \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\log p (\\mathbf {X}, \\mathbf {L} | \\mathbf {G} _ {i n t r a} ^ {i}), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.401, + 0.784, + 0.417 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}_{intra}^{i}\\) is the i-th Monte-Carlo sample and \\(M\\) denotes number of samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.784, + 0.477 + ], + "angle": 0, + "content": "Moreover, note that Equation 12 presumes all partites are independent—an unlikely scenario as proximal nodes often exhibit stronger correlations. In practice, we can approximate \\( p(\\mathbf{G}_{intra}) \\) with an unconditional generative model to consider such correlations and optimize this proxy instead, described as:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.487, + 0.784, + 0.502 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right) \\propto g (\\mathbf {N}), \\quad \\mathbf {N} \\sim \\mathcal {N} (0, 1), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.511, + 0.785, + 0.647 + ], + "angle": 0, + "content": "with \\( g(\\mathbf{N}) \\) symbolizing a generative network derived from standard normal noise. Implementation of the Intra-node BayesNet With the model \\( p(\\mathbf{G}_{\\text{intra}}) \\) established, we can draw samples of topological indices for parties, denoted \\( \\hat{T}_{c,h,w} \\), from each respective \\( \\mathrm{T}_{c,h,w} \\). The subsequent step involves establishing conditional dependencies among these groups. Since dependencies are restricted to nodes from different parties, a dynamic masked convolution operation is utilized. This operation ensures each node's dependencies are limited to preceding nodes within the convolutional kernel's range. For instance, the intra-node BayesNet for \\( \\mathbf{y} \\) can be characterized by:" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.656, + 0.784, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\prod_ {c, h, w} ^ {C, H, W} p \\left(\\mathrm {y} _ {c, h, w} \\mid \\mathbf {G} _ {\\text {i n t r a}}\\right) \\propto \\operatorname {d y n c o n v} \\left(\\hat {\\mathrm {y}} _ {c, h, w}, W _ {c, h, w} \\left(\\hat {T} _ {c, h, w}\\right)\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.784, + 0.768 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{y}}_{c,h,w}\\) are samples from \\(\\mathrm{y}_{c,h,w}\\), and \\(W_{c,h,w}\\) represents a dynamically adjusted convolutional kernel influenced by adjacent nodes' topological indices. As an illustration, consider a 2D \\(5\\times 5\\) convolution: the dynamic kernel \\(W_{c,h,w}\\) used for node \\(\\mathrm{y}_{c,h,w}\\) is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.776, + 0.784, + 0.825 + ], + "angle": 0, + "content": "\\[\nW _ {c, h, w} [ i, j + 2, k + 2 ] = \\left\\{ \\begin{array}{l l} W _ {i, j + 2, k + 2} & , \\hat {T} _ {c, h, w} < \\hat {T} _ {i, h + j, w + k} \\\\ 0 & , \\hat {T} _ {c, h, w} \\geq \\hat {T} _ {i, h + j, w + k} \\end{array} , \\left\\{ \\begin{array}{l l} i & \\in \\{0, 1, \\dots , C - 1 \\} \\\\ j, k & \\in \\{- 2, - 1, 0, 1, 2 \\} \\end{array} . \\right. \\right. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.826, + 0.707, + 0.84 + ], + "angle": 0, + "content": "For in-depth guidance of its implementation, please refer to Appendix." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.143, + 0.756, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.343, + 0.789, + 0.414 + ], + "angle": 0, + "content": "Fig. 2: The architecture of the implemented example based on the joint autoregressive framework [24]. Comp/Decomp means Compression / Decompression. AE/AD symbolize the arithmetic encoder and decoder, respectively. AR denotes the autoregressive model that executes the learned intra-node BayesNet. Q stands for quantization. \\(\\pmb{x}\\) correspond to input image and \\(\\hat{\\pmb{x}}\\) the reconstructed image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.443, + 0.502, + 0.461 + ], + "angle": 0, + "content": "5 From BayesNet to BaSIC" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.599 + ], + "angle": 0, + "content": "Upon determining the BayesNet structure, we proceed to specify the nodes and edges in this BayesNet, crafting a novel NIC framework, BaSIC. This section details an implementation instance of the envisioned computationally scalable NIC, leveraging the prevalent joint autoregressive and hyperprior compression model [24]. It should be noted that this represents just one of the viable strategies for implementing BaSIC, with the potential for alternatives to offer improved results. The architecture of our example is depicted in Figure 2, and we will explore how it seamlessly integrates with the aforementioned BayesNets.\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.789, + 0.631 + ], + "angle": 0, + "content": "Inter-node BayesNet The inter-node BayesNet is constructed over each edge between nodes within the original BayesNet, specifically:" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.641, + 0.713, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q (\\mathbf {y} | \\mathbf {x}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {x}}) \\propto g _ {a} (\\mathbf {x}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {x}}), \\quad q (\\mathbf {z} | \\mathbf {y}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {z}, \\mathbf {y}}) \\propto h _ {a} (\\mathbf {y}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {z}, \\mathbf {y}}), \\\\ (1 1 1, \\hat {\\mathbf {x}} ^ {\\hat {\\mathbf {y}}} \\hat {\\mathbf {z}}), (1 1 1, \\hat {\\mathbf {z}} ^ {\\hat {\\mathbf {y}}} \\hat {\\mathbf {z}}). \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.663, + 0.706, + 0.682 + ], + "angle": 0, + "content": "\\[\np (\\hat {\\mathbf {y}} | \\hat {\\mathbf {z}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {y}}, \\hat {\\mathbf {z}}}) \\propto h _ {s} (\\hat {\\mathbf {z}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {y}}, \\hat {\\mathbf {z}}}), \\quad p (\\hat {\\mathbf {x}} | \\hat {\\mathbf {y}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {x}}, \\hat {\\mathbf {y}}}) \\propto g _ {s} (\\hat {\\mathbf {y}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {x}}, \\hat {\\mathbf {y}}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.789, + 0.815 + ], + "angle": 0, + "content": "To achieve computational scalability, heterogeneous bipartite graphs are employed as outlined in Equation 10 for \\(\\mathbf{G}_{inter}^{\\mathbf{y},\\mathbf{x}}\\), \\(\\mathbf{G}_{inter}^{\\mathbf{z},\\mathbf{y}}\\), \\(\\mathbf{G}_{inter}^{\\hat{\\mathbf{y}},\\hat{\\mathbf{z}}}\\), and \\(\\mathbf{G}_{inter}^{\\hat{\\mathbf{x}},\\hat{\\mathbf{y}}}\\). The implementation of \\(g_{a}, g_{s}, h_{a}\\), and \\(h_{s}\\) follow [29]'s slimmable networks, with dynamic channel widths including 48, 72, 96, 144, and 192 for computational scalability. Note that we use a fixed 192 channels for all latent variables \\(\\mathbf{y}, \\mathbf{z}\\), and only channels in the middle layers of each slimmable network are variable. This modification decouples bitrate scalability from computational scalability, enhancing the convenience in real-world applications." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.513, + 0.84 + ], + "angle": 0, + "content": "Further details are provided in Appendix." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "Intra-node BayesNet In the established framework, the intra-node BayesNet employs masked convolution applied to \\(\\hat{\\mathbf{y}}\\). We substitute this with our suggested multipartite graph-based intra-node BayesNet \\(p(\\hat{\\mathbf{y}}|\\mathbf{G}_{intra}^{\\hat{\\mathbf{y}}})\\), which enables parallel computing on GPUs. The intra-node BayesNet uses a \\(5\\times 5\\) dynamic masked convolution as detailed in Equation 17, with \\(C\\) equals to number of partites and \\(H,W\\) equals to image height and width, respectively. Note that to facilitate the optimization of the intra-node BayesNet, the additional VIMCO loss expressed in Equation 14 should be incorporated into the overall loss function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.784, + 0.313 + ], + "angle": 0, + "content": "Optimization The optimization process involves a loss function that simultaneously accounts for both the BayesNet structure (Equation 14) and the neural network parameters (Equation 4):" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.329, + 0.785, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {R} + \\lambda_ {D} \\mathcal {L} _ {D} + \\lambda_ {C} \\mathcal {L} _ {C} + \\mathcal {L} _ {\\mathbf {G} _ {\\text {i n t r a}}} \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.785, + 0.522 + ], + "angle": 0, + "content": "Computational Scalability The described implementation enables computational scalability across two dimensions. Firstly, by modulating all \\(\\mathbf{G}_{inter}\\), which govern the channel width in slimmable networks, we can fine-tune the computational load of the inter-node neural networks. \\(\\mathbf{G}_{inter}\\) can be modulated during training through the \\(\\lambda_{C}\\) parameter, or during inference by selecting from \\(\\mathbf{G}_{inter}\\) to alter channel widths. Notably, the latter can be streamlined by training with a uniform \\(\\mathbf{G}_{inter}\\) distribution and setting \\(\\lambda_{C} = 0\\). Secondly, by varying the number of partites in \\(\\mathbf{G}_{intra}^{\\hat{y}}\\), we can adjust the stages of parallel computations in autoregressive models, thereby influencing the decompression time. In practice, we could train multiple autoregressive models with varying stages and choose among them during inference to modify decompression times." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.55, + 0.376, + 0.567 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.585, + 0.414, + 0.6 + ], + "angle": 0, + "content": "6.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.784, + 0.761 + ], + "angle": 0, + "content": "Experiment Tasks Following previous works [29], our experiments focus on the task of lossy image compression with computational scalability. We benchmark our BaSIC framework against other computational scalable frameworks, as well as conduct ablation studies to validate the efficacy of the proposed intranode BayesNet learning methods. Further experimental results and details can be found in Appendix, including compression performance evaluation and comparison under fixed computational constraints and more ablation studies regarding the inter-node BayesNet. For evaluation, we employ Pixel Signal-Noise Ratio (PSNR) for distortion and Bits Per Pixel (BPP) for bit-rate, unless stated otherwise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.764, + 0.784, + 0.809 + ], + "angle": 0, + "content": "Datasets We use a merged training dataset from ImageNet [10] and CLIC [2], preprocessed following the method in [11]. The Kodak [1] dataset serves for validation and testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Implementation Details For the proposed method, we use the joint autoregressive framework [24] as the basis, as described in Section 5. Unless otherwise" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.49, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.146, + 0.764, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.3, + 0.756, + 0.314 + ], + "angle": 0, + "content": "Fig. 3: Comparative results for different computational scalable frameworks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.438 + ], + "angle": 0, + "content": "specified, we follow CompressAI [5] to set the hyperparameter \\(\\lambda_{D}\\), while employing \\(\\lambda_{C} = 0\\) during training and a greedy search among 16 intra-node BayesNets with varied FLOPs during inference. Training spans approximately 500k iterations, employing the Adam optimizer with a learning rate of 0.0001. To compare running times equitably, all experiments are performed on identical hardware (an Intel i7-6800K CPU, NVIDIA Titan Xp GPU, and 32GB of memory)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.463, + 0.731, + 0.479 + ], + "angle": 0, + "content": "6.2 Comparison to Other Computation Scalable Frameworks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.613 + ], + "angle": 0, + "content": "This section assesses our computational scalable BaSIC method against other scalable frameworks. We configure our framework without autoregressive modules, similar to SlimCAE [29], to compare performance at various computational levels. We adopt Hyperprior [4] as a reference, reporting BD-rate for diverse MACs for both methods. Also, the rate-distortion plots for SlimCAE's full range of complexity and three levels (L0, L7, L15) for our method are included for reference. Note that \\(\\lambda\\)-scheduling is deactivated in SlimCAE to facilitate uniform BD-rate analysis across levels. The results are depicted in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Our method, employing the same slimmable network backbone as SlimCAE, demonstrates a marginal performance edge at most computation levels, likely attributable to the fixed channels of latent variables that enhance higher bit-rate performance. Specifically, our method saves around \\(2\\%\\) more bits than Slim-CAE under the maximum complexity, and around \\(10\\%\\) more bits under the minimum complexity. While SlimCAE [29] employs \\(\\lambda\\)-scheduling to tackle this, it inadvertently ties bit-rate and distortion to computational complexity, which is less desirable for cases like low-powered devices needing high-quality images. In contrast, our inter-node BayesNet independently modulates the computational complexity of individual neural network modules, as evidenced by the rate-distortion plots that show minimal impact on performance with reduced computation. Moreover, the varied complexity levels indicate our method's superior control over computational complexity, allowing us to assign different inter-node BayesNets (channel widths) to distinct layers. SlimCAE, conversely, applies the same channel width across all layers, limiting flexibility. For instance," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.278, + 0.145, + 0.724, + 0.16 + ], + "angle": 0, + "content": "Table 2: Bit-rate comparison for different autoregressive modules." + }, + { + "type": "table", + "bbox": [ + 0.247, + 0.172, + 0.756, + 0.234 + ], + "angle": 0, + "content": "
Stages-2410
MethodsMaskConvBaselineCBCWOursMSOursELICOurs
BPP0.31500.33330.32160.32220.32070.32100.31830.31740.3156
Speed (MB/s)0.32345.55191.23173.46172.57107.9374.6717.8517.32
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.263, + 0.784, + 0.294 + ], + "angle": 0, + "content": "to meet a 200kMAC/pixel threshold, our closest model uses 158kMAC/pixel, while SlimCAE's nearest option drops to 111kMAC/pixel, as shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.318, + 0.762, + 0.348 + ], + "angle": 0, + "content": "6.3 Comparison to Other Autoregressive Models for Intra-Node BayesNet" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.787, + 0.542 + ], + "angle": 0, + "content": "In this section, we evaluate the effectiveness of the proposed intra-node BayesNet structure learning method, by comparing it to other hand-crafted autoregressive models. For a simple and fair comparison, we first establish a baseline using a joint autoregressive framework [24] with a null context model, trained for about 500k iterations. Subsequently, we retain all other components while training various autoregressive models to replace the baseline context model for an additional 500k iterations. This approach ensures the reconstructed image remains unchanged, allowing us to isolate and evaluate the bit-rate to reflect the effectiveness of different autoregressive models. Moreover, the decompression speed is also reported to evaluate the efficiency of It's important to note that the neural network architecture for the context model is kept constant across all methods, with the sole variability stemming from the structure of the intra-node BayesNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.648 + ], + "angle": 0, + "content": "The benchmark autoregressive models include Channel-wise (CW, 2-stage) [25], Checkerboard [12] (CB, 2-stage), Multi-stage spatial \\(2 \\times 2\\) [21] (MS, 4-stage) and the combined unevenly grouped channel-wise and checkerboard [11] (ELIC, 10-stage in total). We also train the MaskConv context model from [24] as a baseline. Our model is trained under constraints of 2, 4, and 10 parallel stages and evaluated against counterparts with equivalent stage configurations. The findings are presented in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.649, + 0.788, + 0.754 + ], + "angle": 0, + "content": "Overall, our learning-based BayesNet structures consistently demonstrate slightly better performance compared to the manually-crafted models under the same parallel stages, with similar decompression speed. Interestingly, in each group our learned method reaches close BPP to the hand-crafted ones in the next level of parallel stage, and in the 10-stage case, our method even approaches the BPP of non-parallel MaskConv model. This proves that our proposed method excels in learning effective intra-node BayesNets for autoregressive models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.788, + 0.815 + ], + "angle": 0, + "content": "Qualitative Results We also illustrate the learned intra-node BayesNet by our method as well as the manually-designed one of ELIC in Figure 4 to analyze their difference with hand-crafted models. We could see that the BayesNet of ELIC uses a checkerboard context model over the spatial dimension, including only two" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.568, + 0.84 + ], + "angle": 0, + "content": "5 More visualization results are shown in Appendix." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.761, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.239, + 0.531, + 0.251 + ], + "angle": 0, + "content": "(a) ELIC" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.251, + 0.76, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.346, + 0.553, + 0.358 + ], + "angle": 0, + "content": "(b) Ours-10stage" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.44 + ], + "angle": 0, + "content": "Fig. 4: Partite topological indices visualization for intra-node BayesNet in the 10-stage case. Each image patch indicates topological indices of nodes in a group of channels. Blacker pixels indicate lower topological indices and whiter indicate higher topological indices. E.g., in the 10-stage case, pure-black indicates index 0 which is first iterated, and pure-white indicates group 9 which is last iterated. Better viewed with zoom-in." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.577 + ], + "angle": 0, + "content": "partites within each channel. In comparison, our learned BayesNet incorporates more different partites (mostly three or four) in the spatial dimension, which captures more context information than the checkerboard case. This is reasonable as images are usually heavily correlated in spatial dimensions. Moreover, the learned topological indices also give some insight for designing parallel autoregressive models, that channel and spatial context models could be combined more compactly to improve their performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.598, + 0.527, + 0.615 + ], + "angle": 0, + "content": "7 Conclusions and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.751 + ], + "angle": 0, + "content": "This paper presents a novel approach, BaSIC, to addressing computational scalability in Neural Image Compression (NIC) by focusing on BayesNet structure learning. We decompose the challenge into two subproblems: heterogeneous bipartite graph learning for inter-node BayesNet, and multipartite graph learning for intra-node BayesNet, proposing solutions for each. Experimental results, using a joint autoregressive framework as a test case, show that our work offers improved control over computational demands relative to other scalable computation models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.207 + ], + "angle": 0, + "content": "The paper is supported in part by the National Natural Science Foundation of China (No. 62325109, U21B2013)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.245 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.288 + ], + "angle": 0, + "content": "1. Kodak lossless true color image suite. https://r0k.us/graphics/kodak/ (2023), accessed: 2023-11-12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.315 + ], + "angle": 0, + "content": "2. Workshop and challenge on learned image compression. https://click.compression.cc/ (2023), accessed: 2023-01-09" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.357 + ], + "angle": 0, + "content": "3. Balle, J., Laparra, V., Simoncelli, E.P.: Density modeling of images using a generalized normalization transformation. CoRR abs/1511.06281 (2015), https://apisemantic scholar.org/CorpusID:2684987" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.384 + ], + "angle": 0, + "content": "4. Balle, J., Minnen, D.C., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. ArXiv abs/1802.01436 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.425 + ], + "angle": 0, + "content": "5. Bégaint, J., Racapé, F., Feltman, S., Pushparaja, A.: Compressai: a pytorch library and evaluation platform for end-to-end compression research. arXiv preprint arXiv:2011.03029 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.468 + ], + "angle": 0, + "content": "6. Cheng, J., Greiner, R., Kelly, J., Bell, D.A., Liu, W.: Learning bayesian networks from data: An information-theory based approach. Artif. Intell. 137, 43-90 (2002), https://api-semanticscholar.org/CorpusID:7821347" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.785, + 0.522 + ], + "angle": 0, + "content": "7. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 7936-7945 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.562 + ], + "angle": 0, + "content": "8. Chickering, D.M.: Optimal structure identification with greedy search. J. Mach. Learn. Res. 3, 507-554 (2003), https://api-semanticscholar.org/CorpusID:1191614" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.617 + ], + "angle": 0, + "content": "9. Deleu, T., G'ois, A., Emezue, C.C., Rankawat, M., Lacoste-Julien, S., Bauer, S., Bengio, Y.: Bayesian structure learning with generative flow networks. ArXiv abs/2202.13903 (2022), https://api(semanticscholar.org/CorpusID: 247158659" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "10. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition pp. 248-255 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.661, + 0.785, + 0.728 + ], + "angle": 0, + "content": "11. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5708-5717 (2022), https://apisemantic scholar.org/CorpusID:247594672" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.729, + 0.785, + 0.784 + ], + "angle": 0, + "content": "12. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14766-14775 (2021), https://apisemantic scholar.org/CorpusID:232404386" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:326772" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.26, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Jang, E., Gu, S.S., Poole, B.: Categorical reparameterization with gumbel-softmax. ArXiv abs/1611.01144 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "15. Johnston, N., Eban, E., Gordon, A., Ball'e, J.: Computationally efficient neural image compression. ArXiv abs/1912.08771 (2019), https://apisemantic scholar.org/CorpusID:209404937" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.272 + ], + "angle": 0, + "content": "16. Kang, N., Qiu, S., Zhang, S., Li, Z., Xia, S.: Pilc: Practical image lossless compression with an end-to-endgpu oriented neural framework. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3729-3738 (2022), https://apisemantic scholar.org/CorpusID:249625569" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "17. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. CoRR abs/1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "18. Koivisto, M., Sood, K.: Exact bayesian structure discovery in bayesian networks. J. Mach. Learn. Res. 5, 549-573 (2004), https://api(semanticscholar.org/ CorpusID:12412451" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "19. Lee, J., Cho, S., Beack, S.: Context-adaptive entropy model for end-to-end optimized image compression. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.438 + ], + "angle": 0, + "content": "20. Li, M., Ma, K., You, J.J., Zhang, D., Zuo, W.: Efficient and effective context-based convolutional entropy modeling for image compression. IEEE Transactions on Image Processing 29, 5900-5911 (2019), https://api-semanticscholar.org/CorpusID:195345342" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.494 + ], + "angle": 0, + "content": "21. Lin, F., Sun, H., Liu, J., Katto, J.: Multistage spatial context models for learned image compression. ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 1-5 (2023), https://apisemantic scholar.org/CorpusID:257038187" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "22. Liu, H., Simonyan, K., Yang, Y.: Darts: Differentiable architecture search. ArXiv abs/1806.09055 (2018), https://api-semanticscholar.org/CorpusID: 49411844" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.535, + 0.785, + 0.577 + ], + "angle": 0, + "content": "23. Lorch, L., Rothfuss, J., Scholkopf, B., Krause, A.: Dibs: Differentiable bayesian structure learning. ArXiv abs/2105.11839 (2021), https://apisemantic scholar.org/CorpusID:235187432" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.605 + ], + "angle": 0, + "content": "24. Minnen, D.C., Balle, J., Toderici, G.: Joint autoregressive and hierarchical priors for learned image compression. ArXiv abs/1809.02736 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.659 + ], + "angle": 0, + "content": "25. Minnen, D.C., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. 2020 IEEE International Conference on Image Processing (ICIP) pp. 3339-3343 (2020), https://api-semanticscholar.org/CorpusID:220633468" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "26. Mnih, A., Rezende, D.J.: Variational inference for monte carlo objectives. In: International Conference on Machine Learning (2016), https://api-semanticscholar.org/CorpusID:5859948" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "27. Tao, L., Gao, W., Li, G., Zhang, C.: Adanic: Towards practical neural image compression via dynamic transform routing. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 16833-16842 (2023), https://api.sementicscholar.org/CorpusID:267026318" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "28. webmproject: libwebp. https://github.com/webmproject/libwebp (2023), accessed: 2023-01-09" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Yang, F., Herranz, L., Cheng, Y., Mozerov, M.G.: Slimmable compressive autoencoders for practical neural image compression. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 4996-5005 (2021), https://api-semanticscholar.org/CorpusID:232417114" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "30. Yin, S., Li, C., Bao, Y., Liang, Y.: Universal efficient variable-rate neural image compression. ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 2025-2029 (2021), https://api.sementicscholar.org/CorpusID:244478380" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.258 + ], + "angle": 0, + "content": "31. Yin, S., Meng, F., Tan, W.T., Li, C., Bao, Y., Liang, Y., Liu, W.: Exploring structural sparsity in neural image compression. 2022 IEEE International Conference on Image Processing (ICIP) pp. 471-475 (2022), https://api-semanticscholar.org/CorpusID:246680031" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.299 + ], + "angle": 0, + "content": "32. Yu, J., Yang, L., Xu, N., Yang, J., Huang, T.S.: Slimmable neural networks. ArXiv abs/1812.08928 (2018), https://api(semanticscholar.org/CorpusID:56657799" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "33. Yu, Y., Chen, J., Gao, T., Yu, M.: Dag-gnn: Dag structure learning with graph neural networks. In: International Conference on Machine Learning (2019), https://apisemantic scholar.org/CorpusID:128358697" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.383 + ], + "angle": 0, + "content": "34. Yuan, C., Malone, B.M.: Learning optimal bayesian networks: A shortest path perspective. J. Artif. Intell. Res. 48, 23-65 (2013), https://api-semanticscholar.org/CorpusID:5951724" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.438 + ], + "angle": 0, + "content": "35. Zhang, Y., Lu, G., Feng, D., Zhu, C., Song, L.: Content adaptive checkerboard context model for learned image compression. 2023 IEEE International Symposium on Circuits and Systems (ISCAS) pp. 1-5 (2023), https://api-semanticscholar.org/CorpusID:260002843" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.493 + ], + "angle": 0, + "content": "36. Zhang, Z., Chen, B., Lin, H., Lin, J., Wang, X., Zhao, T.: Elfic: A learning-based flexible image codec with rate-distortion-complexity optimization. Proceedings of the 31st ACM International Conference on Multimedia (2023), https://apisemantic scholar.org/CorpusID:264492165" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.535 + ], + "angle": 0, + "content": "37. Zheng, X., Aragam, B., Ravikumar, P., Xing, E.P.: Dags with no tears: Continuous optimization for structure learning. In: Neural Information Processing Systems (2018), https://api_semanticscholar.org/CorpusID:53217974" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.591 + ], + "angle": 0, + "content": "38. Zheng, Z., Wang, X., Lin, X., Lv, S.: Get the best of the three worlds: Real-time neural image compression in a non-gpu environment. Proceedings of the 29th ACM International Conference on Multimedia (2021), https://api-semanticscholar.org/CorpusID:239011955" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/360699b9-a034-4ed5-b537-07332b36832e_origin.pdf b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/360699b9-a034-4ed5-b537-07332b36832e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2d7d14a1951d0dd36c14f61fe81dbd28c53f6062 --- /dev/null +++ b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/360699b9-a034-4ed5-b537-07332b36832e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:132638e86a75a3633969075c3fa74c5a0a5d1fd4b44579d188a1e826fdee34d1 +size 1094245 diff --git a/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/full.md b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b8337f35d5788fa193a5d6f923e5030c1037d2a0 --- /dev/null +++ b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/full.md @@ -0,0 +1,328 @@ +# BaSIC: BayesNet Structure Learning for Computational Scalable Neural Image Compression + +Yufeng Zhang $^{1}$ , Hang Yu $^{2}$ , Shizhan Liu $^{1}$ , Wenrui Dai $^{1}$ , and Weiyao Lin $^{\star 1}$ + +$^{1}$ Shanghai Jiao Tong University, China. {worldlife,shanluzuode,daiwenrui,wylin}@sjtu.edu.cn $^{2}$ Ant Group, China. hyu.hugo@antgroup.com + +Abstract. Despite superior rate-distortion performance over traditional codecs, Neural Image Compression (NIC) is limited by its computational scalability in practical deployment. Prevailing research focuses on accelerating specific NIC modules but is restricted in controlling overall computational complexity. To this end, this work introduces BaSIC (BayesNet structure learning for computational Scalable neural Image Compression), a comprehensive, computationally scalable framework that affords full control over NIC processes. We learn the Bayesian network (BayesNet) structure of NIC for controlling both neural network backbones and autoregressive units. The learning of BayesNet is achieved by solving two sub-problems, i.e., learning a heterogeneous bipartite BayesNet for the inter-node structure to regulate backbone complexity, and a multipartite BayesNet for the intra-node structure to optimize parallel computation in autoregressive units. Experiments demonstrate that our method not only facilitates full computational scalability with more accurate complexity control but also maintains competitive compression performance compared to other computation scalable frameworks under equivalent computational constraints. Code will be available in https://github.com/worldlife123/cbench_BaSIC. + +Keywords: Image Compression $\cdot$ Computational Scalable Compression $\cdot$ BayesNet Structure Learning $\cdot$ Autoregressive Model + +# 1 Introduction + +Neural Image Compression (NIC) technologies [7, 11] have significantly outstripped the performance of conventional image CODECs, ushering in an era of intelligent image compression. These advanced CODEcs are intended for deployment across a spectrum of devices, from low-end devices like mobile phones to high-end devices like GPU computation clusters. Each type of device necessitates that CODEcs tailor their performance to specific computational or memory + +![](images/b7e40b9eb529fe36748b0cfe1367f5f601611e68962e546044fcb5a92bd5724d.jpg) +Fig. 1: Bayesian network diagrams for various NIC frameworks. Our proposed framework BaSIC (d) generalizes the concepts in (a), (b), and (c). + +![](images/99a995a3c841f19090b791a708db2180f96b1e6ef47cc7830340e0c89c41ad5a.jpg) + +constraints, a requirement known as Computational Scalability. High-end devices can endure more intricate computations for superior encoding quality, whereas low-end devices may prioritize reduced computational load over encoding efficiency to conserve energy. + +While conventional image CODECs typically accommodate this functionality (e.g. WebP [28] offers seven levels of computational complexity to regulate lossy image compression), incorporating this attribute into NIC remains a significant hurdle for its broader adoption. Recent studies have focused predominantly on accelerating NIC computations rather than addressing computational scalability. Works such as [11, 12, 25] target the acceleration of autoregressive modules on GPUs, which are frequently used in NIC. Others, including [15, 31, 38], explore neural network acceleration techniques—given that NIC fundamentally relies on neural network architectures. On the other hand, some recent efforts have attempted to integrate dynamic neural networks like slimmable networks [32] within NIC frameworks to adjust the computational complexity [27, 29, 36]. However, these approaches have only partially addressed computational scalability and often neglect crucial components in NIC, such as autoregressive modules, which are essential for compression performance. + +In this work, we endeavor to create a NIC framework with enhanced computational scalability, incorporating computational control across all NIC components. We call this framework BaSIC (BayesNet structure learning for computational Scalable neural Image Compression). To this end, we re-examine NIC through the broader lens of the Bayesian network (BayesNet) paradigm. Figure 1 illustrates the BayesNet depiction of several exemplary NIC architectures. Specifically, our approach is rooted in the hyperprior BayesNet framework (Figure 1a), which postulates that the generative model can be factorized as $p(\mathbf{x}|\mathbf{y})p(\mathbf{y}|\mathbf{z})p(\mathbf{z})$ , where $\mathbf{x}$ represents observed data, $\mathbf{y}$ the prior, and $\mathbf{z}$ the hyperprior. In this context, SlimCAE (Figure 1b) can be interpreted as allowing + +selective edge types between $\mathbf{y}$ and $\mathbf{x}$ , and between $\hat{\mathbf{x}}$ and $\hat{\mathbf{y}}$ , with each type manifesting different conditional dependencies, modeled via slimmable neural networks with varied channel widths. On the other hand, the joint autoregressive framework (Figure 1c) introduces autoregressive intra-node dependencies within $\hat{y}$ to enhance compression, at the expense of increased computational demands due to the requisite sequential processing [12]. These examples clearly demonstrate the intimate connection between the structure of the BayesNet and the complexity inherent in NIC frameworks. + +Our advancement is a NIC framework that exploits end-to-end BayesNet structure learning [23, 33] to meticulously optimize computational scalability. This innovative framework is not bound to specific neural network backbones, making it versatile enough to enhance computational scalability in most existing NIC frameworks with a hyperprior BayesNet representation. Our proposed framework is visualized in Figure 1d. In particular, we represent the BayesNet structure as $\mathbf{G}$ and impose a prior $p(\mathbf{G})$ on it. Given that the structure we seek to learn must conform to a directed acyclic graph (DAG), we tailor the prior $p(\mathbf{G})$ for compression frameworks within this DAG constraint, considering two distinct instances: the inter-node BayesNet, representing conditional dependencies between nodes, and the intra-node BayesNet, which pertains to dependencies within nodes. We depict the former as a heterogeneous bipartite graph and design a structure learning method to seamlessly incorporate dynamic neural networks like slimmable networks [32]. For the latter, we design a network architecture parameterized by multipartite or $k$ -partite graphs to enable a parallel autoregressive process. By manipulating the partite number $k$ , we can adjust the temporal demands of the autoregressive process, thereby directly impacting decompression speed. + +When deploying the codec, the configurations of both the inter-node and intra-node BayesNets can be tailored to match the computational budget of target devices. For instance, on less capable devices, we might employ simpler neural network edges within the inter-node BayesNet and a smaller partite number within the intra-node BayesNet, whereas more capable devices can leverage a more complex network and a larger partite number, allowing for scalable computational resource allocation based on the specifics of the device at hand. + +To summarize, our contributions are mainly three-fold: + +- We propose an end-to-end computationally scalable NIC framework founded on BayesNet structure learning, facilitating fine-tuned adjustments to computational complexity across all facets of the NIC process. +- We devise a generative model based on a heterogeneous bipartite structure for the inter-node BayesNet, harnessing the capabilities of existing dynamic neural networks to realize computational scalability effectively. +- We design a generative model structured as a multipartite graph for the intra-node BayesNet. This model is tailored to accommodate and streamline parallel computation in practical applications. + +# 2 Related Work + +In this section, we briefly review recent works regarding the computational complexity of NIC as well as BayesNet structure learning. + +# 2.1 End-to-end Neural Image Compression + +The hyperprior framework [4] serves as a foundation for most end-to-end NIC models, favored for its effective compression performance and straightforward implementation. Subsequent studies [7, 19, 24] have enhanced the hyperprior by integrating a masked convolution-based autoregressive model in the latent space to reduce bit-rate. While these improvements offer better compression, they often overlook the increased computational demands. To apply NIC in practice, some recent works are now addressing the need to expedite these models. + +Autoregressive Model Acceleration To circumvent the slow serial iteration steps of masked convolutions on GPUs, newer studies introduce parallelized autoregressive models for GPU acceleration. For instance, [25] presents a channelwise parallel autoregressive model, while [12] proposes a spatially parallel model using a checkerboard mask, requiring just two parallel stages. This concept is extended to a multistage model by [21]. Further, [11] amalgamates channel-wise and spatially parallel autoregressive models. Additional acceleration methods include content-adaptive adjustments to parallel autoregressive models [35] and designing simplified autoregressive models [16]. Yet, such models are typically hand-crafted and may not represent the most efficient trade-off between compression complexity and performance. + +Neural Network Acceleration Considering that neural networks (i.e., edges in the BayesNets) represent a significant portion of NIC's computational complexity, some works focus on accelerating these directly. For example, there are designs for more efficient neural network operators like a simplified GDN [15] and CPU-friendly down/up-samplers [38]. Others apply general neural network acceleration techniques within compression models, such as MorphNet [15] and learnable channel pruning [30, 31]. + +Computation Scalable Framework A handful of recent works have also noticed the importance of computational scalability for NIC. SlimCAE [29] employs slimmable networks [32] as the framework's backbone, enabling dynamic adjustments of bit-rate and computation through channel width modulation. Further developments include universal slimmable networks for more precise scalability [36] and the integration of input-adaptive complexity allocation methods [27,36]. Nonetheless, these approaches focus solely on the scalability of neural networks, largely disregarding the autoregressive models' scalability. + +# 2.2 BayesNet Structure Learning + +BayesNet structure learning aims to discern the conditional relationships among a set of variables with a graph representation, which ideally should be a directed acyclic graph (DAG). Traditional approaches often involve score-based [8, 34] + +and constraint-based methods [6, 18]. However, they typically operate within a constrained search space, which can lead to less-than-optimal graphs. More recent initiatives introduce neural network techniques for more robust graph generation within the realm of BayesNet structure learning. Some align this learning task with variational inference to facilitate graph structure development within an end-to-end framework [23, 33], utilizing graph acyclicity regularizers [37] that may not guarantee acyclicity. Alternatively, other research employs recurrent-style graph generators that assure acyclicity through iterations [9], but these can become inefficient with an increase in the number of nodes. In our work, we diverge from these established methods due to their inefficiency when dealing with the dense pixelation of images. Instead, we analyze the BayesNet representations inherent in existing NIC frameworks and directly parameterize the graph structure as DAGs, ensuring both acyclicity and efficiency. + +# 3 Overview of BaSIC + +Assume we have a set of observations $\mathbf{X} = \{\mathbf{x}_1, \mathbf{x}_2, \dots, \mathbf{x}_N\}^3$ , the objective of BayesNet structure learning is to infer a distribution over the structure of the BayesNet $\mathbf{G} = (\mathbf{V}, \mathbf{E}, \boldsymbol{\theta})$ that best describes the observations. According to the Bayesian formalism, such optimization problem could be solved by minimizing expectation over $p(\mathbf{G}|\mathbf{X})$ of any function $f$ related to $\mathbf{G}$ , under the constraint that $\mathbf{G}$ is acyclic: + +$$ +\underset {\mathbf {G}} {\min } \mathbb {E} _ {p (\mathbf {G} | \mathbf {X})} [ f (\mathbf {G}) ], \quad \mathrm {s . t .} \mathbf {G} \text {i s a c y c l i c .} \tag {1} +$$ + +For a computationally scalable compression framework, we tackle BayesNet structure learning, linking BayesNet configurations to computational complexity and shaping the function $f(\mathbf{G})$ accordingly. + +Apart from modeling of $\mathbf{X}$ , the compression framework also involves latent variables $\mathbf{L} = \{\mathbf{y}_1, \mathbf{y}_2, \ldots, \mathbf{y}_N, \mathbf{z}_1, \mathbf{z}_2, \ldots, \mathbf{z}_N, \ldots\}$ to generate $\mathbf{X}$ . The resulting BayesNet should capture conditional dependencies within and between $\mathbf{X}$ and $\mathbf{L}$ . This generative relationship is described by: + +$$ +p (\mathbf {G}, \mathbf {L}, \mathbf {X}) = p (\mathbf {G}) p (\mathbf {L} | \mathbf {G}) p (\mathbf {X} | \mathbf {G}, \mathbf {L}). \tag {2} +$$ + +The expectation from Equation 1 is then: + +$$ +\mathbb {E} _ {p (\mathbf {G}, \mathbf {L} | \mathbf {X})} [ f (\mathbf {G}, \mathbf {L}) ] = \mathbb {E} _ {p (\mathbf {G} | \mathbf {X}) p (\mathbf {L} | \mathbf {X}, \mathbf {G})} [ f (\mathbf {G}, \mathbf {L}) ]. \tag {3} +$$ + +Existing compression frameworks without $\mathbf{G}$ approximate $p(\mathbf{L}|\mathbf{X})$ with an inference network $q(\mathbf{L}|\mathbf{X})$ , as in VAEs [17]. Incorporating $\mathbf{G}$ into our design, we refine our inference network to $q(\mathbf{G},\mathbf{L}|\mathbf{X})$ for enhanced approximation: + +$$ +\mathbb {E} _ {p (\mathbf {G}, \mathbf {L} | \mathbf {X})} [ f (\mathbf {G}, \mathbf {L}) ] \approx \mathbb {E} _ {q (\mathbf {G} | \mathbf {X}) q (\mathbf {L} | \mathbf {X}, \mathbf {G})} [ f (\mathbf {G}, \mathbf {L}) ]. \tag {4} +$$ + +This approach allows $\mathbf{G}$ to direct the computational complexity of both the inference and the generative processes. Furthermore, we adopt a differentiable approach to learning $\mathbf{G}$ , in line with recent advances [23, 33], and incorporate this aspect into our model specification. + +Definition of $f(\mathbf{G}, \mathbf{L})$ In line with lossy image compression standards [15,31], our optimization combines bit-rate $\mathcal{L}_R$ , distortion $\mathcal{L}_D$ , and complexity $\mathcal{L}_C$ : + +$$ +\mathbb {E} _ {q (\mathbf {G} | \mathbf {X}) q (\mathbf {L} | \mathbf {X}, \mathbf {G})} [ f (\mathbf {G}, \mathbf {L}) ] = \mathcal {L} _ {R} (\mathbf {G}, \mathbf {L}) + \lambda_ {D} \mathcal {L} _ {D} (\mathbf {G}, \mathbf {L}) + \lambda_ {C} \mathcal {L} _ {C} (\mathbf {G}, \mathbf {L}). \tag {5} +$$ + +Concretely, $\mathcal{L}_R$ is computed by enacting the BayesNet's generative process with latent node samples $\hat{\mathbf{L}}$ : + +$$ +\mathcal {L} _ {R} (\mathbf {G}, \mathbf {L}) = \log_ {2} p _ {\mathbf {L} | \mathbf {G}} (\hat {\mathbf {L}} | \mathbf {G}), \quad \hat {\mathbf {L}} \sim q (\mathbf {L} | \mathbf {X}, \mathbf {G}) \tag {6} +$$ + +Moreover, $\mathcal{L}_D$ and $\mathcal{L}_C$ can be tailored to specific optimization objectives. $\mathcal{L}_D$ may encompass a range of complexity measures or their amalgamations, such as conventional metrics like Mean-Squared Error (MSE) or more advanced machine learning metrics like Frechet Inception Distance (FID) [13]. Conversely, $\mathcal{L}_C$ can include various complexity metrics or their combinations, such as the number of Floating-point Operations (FLOPs), network parameters, time, or energy consumption. $\lambda_{D}$ and $\lambda_{C}$ are the corresponding weights. By setting $f(\mathbf{G},\mathbf{L})$ in line with the established rate-distortion (R-D) trade-off, Equation 4 falls in step with loss functions prevalent in standard compression frameworks. + +Definition of $q(\mathbf{G}|\mathbf{X})$ As a determiner of the BayesNet structure, $\mathbf{G}$ affects computational complexity. In practical compression scenarios, including $\mathbf{G}$ in the bitstream could impact the bit-rate. Typically, it may be more advantageous to treat $\mathbf{G}$ as independent of $\mathbf{X}$ , suggesting $q(\mathbf{G}|\mathbf{X}) = p(\mathbf{G})$ . However, in scenarios where complexity is tailored to input data, the model $q(\mathbf{G}|\mathbf{X})$ can be employed. DAG Constraint of $p(\mathbf{G})$ As previously noted in Equation 1, $p(\mathbf{G})$ should reflect a distribution over DAG structures. For edges between different node groups in compression models, such as $p(\mathbf{x}|\mathbf{y})$ or $p(\mathbf{y}|\mathbf{z})$ , ensuring a DAG structure is straightforward since bipartite graphs are inherently acyclic; we call this component the Inter-node BayesNet ( $\mathbf{G}_{inter}$ ). For conditional dependencies within node groups, like $p(\mathrm{y}_i|\mathrm{y}_{i-1})$ which also require DAG adherence, we refer to it as the Intra-node BayesNet ( $\mathbf{G}_{intra}$ ). The forthcoming sections will delve into the methodology for learning both $p(\mathbf{G}_{inter})$ and $p(\mathbf{G}_{intra})$ as DAGs. + +# 4 Structure Learning for Inter and Intra-Node BayesNets + +# 4.1 Heterogeneous Bipartite Graphs for Inter-node BayesNet + +In this section, we specify $p(\mathbf{G}_{inter})$ . Note that $\mathbf{G}_{inter}$ is a bipartite graph, and so it inherently satisfies the acyclic constraint. + +Naive Bipartite Parameterization of $p(\mathbf{G}_{inter})$ Consider the task of learning a bipartite graph for $p(\mathbf{y}|\mathbf{z})$ . We split the nodes into $\mathbf{z}_1, \mathbf{z}_2, \ldots, \mathbf{z}_I$ and $\mathbf{y}_1, \mathbf{y}_2, \ldots, \mathbf{y}_J$ . Bipartite graphs can be represented by an $I \times J$ adjacency matrix + +with each element indicating an edge between $\mathbf{y}$ and $\mathbf{z}$ . Consequently, $\mathbf{G}_{inter}$ can be modeled as $I \times J$ Bernoulli distributions: + +$$ +p \left(\mathbf {G} _ {\text {i n t e r}} ^ {\mathbf {y}, \mathbf {z}}\right) = \operatorname {B e r n o u l l i} \left(\pi_ {\mathbf {y}, \mathbf {z}}\right), \quad \pi_ {\mathbf {y}, \mathbf {z}} \in \mathbb {R} ^ {I \times J}. \tag {7} +$$ + +The edge $p(\mathbf{y}|\mathbf{z},\mathbf{G}_{inter}^{\mathbf{y};\mathbf{z}})$ is then defined as: + +$$ +p (\mathbf {y} | \mathbf {z}, \mathbf {G} _ {\text {i n t e r}} ^ {\mathbf {y}, \mathbf {z}}) = \prod_ {i, j} ^ {I, J} p (\mathbf {y} _ {j} | \mathbf {z} _ {i}) \hat {G} _ {\text {i n t e r}} ^ {\mathbf {y}, \mathbf {z}} [ i, j ], \tag {8} +$$ + +where $\hat{G}_{inter}^{\mathbf{y},\mathbf{z}}$ are samples from $p(\mathbf{G}_{inter}^{\mathbf{y},\mathbf{z}})$ . Nonetheless, implementing such a BayesNet can present challenges. For instance, normalization layers such as GDN [3], often integral to compression models, are difficult to integrate because the nodes are presumed to be mutually independent. + +Heterogeneous Parameterization of $p(\mathbf{G}_{inter})$ We offer an innovative parameterization for the bipartite graph, tailored for heterogeneous DAG structure learning within image compression frameworks. Rather than dividing nodes, we maintain individual nodes in adjacent layers and diversify only the connecting edges. Each node pair is linked through multiple neural network edges of varying sizes and complexities, from which only one is selected. In this context, $\mathbf{G}_{inter}$ is characterized by $N$ categorical distributions: + +$$ +p \left(\mathbf {G} _ {\text {i n t e r}} ^ {\mathbf {y}, \mathbf {z}}\right) = \operatorname {C a t e g o r i c a l} \left(\pi_ {\mathbf {y}, \mathbf {z}}\right), \quad \pi_ {\mathbf {y}, \mathbf {z}} \in \mathbb {R} ^ {N}. \tag {9} +$$ + +Thus, the connection $p(\mathbf{y}|\mathbf{z},\mathbf{G}_{inter}^{\mathbf{y},\mathbf{z}})$ can be expressed as: + +$$ +p (\mathbf {y} | \mathbf {z}, \mathbf {G} _ {i n t e r} ^ {\mathbf {y}, \mathbf {z}}) = \sum_ {n} ^ {N} \hat {G} _ {i n t e r} ^ {\mathbf {y}, \mathbf {z}} [ n ] p _ {n} (\mathbf {y} | \mathbf {z}). \tag {10} +$$ + +This heterogeneous DAG learning approach does not merely determine whether nodes are interdependent; it specifies how they are interconnected. Given the inherent dependencies among latent nodes in compression frameworks, this nuanced parameterization is particularly apt for representing inter-node BayesNets. + +Implementation of the edges Once $\mathbf{G}_{inter}$ is established, we proceed to construct the heterogeneous edges, represented as $p_n(\mathbf{y}|\mathbf{z})$ , utilizing neural networks. Following [29], we employ slimmable networks and select the $n$ -th channel width configuration for each $p_n(\mathbf{y}|\mathbf{z})$ . This approach is memory-efficient as it does not necessitate additional memory overhead compared to a singular network implementation for $p(\mathbf{y}|\mathbf{z})$ . A comprehensive example detailing this implementation is provided in Section 5 for further reference. + +Optimization of $p(\mathbf{G}_{inter})$ We can leverage samples from $p(\mathbf{G}_{inter})$ as weights to blend the various edges, enabling end-to-end optimization via the Gumbel-softmax reparameterization trick [14], eliminating the need for an additional loss function. The computational complexity associated with this optimization can be quantified using the same mixing weights: + +$$ +\mathcal {L} _ {C} \left(\mathbf {G} _ {\text {i n t e r}} ^ {\mathbf {y}, \mathbf {z}}\right) = \sum_ {n} ^ {N} \hat {\mathbf {G}} _ {\text {i n t e r}} ^ {\mathbf {y}, \mathbf {z}} [ n ] C \left(p _ {n} (\mathbf {y} | \mathbf {z})\right), \tag {11} +$$ + +Table 1: Statistics of different intra-node BayesNet configurations for a $3 \times 32 \times 32$ tensor. Numbers inside the nodes represent the topological indices of the partites. For Maskconv each partite only include one node, so we omit the topological index. + +
MaskConv [24]Zigzag [20]Checkerboard [12]Multistage [21]Channel [25]
Stages102463243
FLOPs1171888352976117182048
Time (ms)11783.0727.346.582.671.2
+ +where $C(p_{m}(\mathbf{y}|\mathbf{z}))$ represents the complexity measure for a single edge such as FLOPs. + +Relation to Neural Architecture Search (NAS) Our approach to parameterizing $p(\mathbf{G}_{inter})$ bears resemblance to Differentiable Architecture Search [22] in both concept and optimization techniques. Both methods utilize continuous relaxation to combine outputs from diverse neural network modules. However, the underlying objectives differ. NAS focuses on developing neural networks with enhanced performance by refining the computational graph. In contrast, our objective is to construct optimized probabilistic graphs tailored for compression, with neural networks serving as approximators of the conditional probabilities. + +# 4.2 Multipartite Graph for Intra-node BayesNet + +As noted in Section 2.2, traditional BayesNet structure learning has been constrained by the need to maintain a DAG, which, when applied to image data with its vast number of pixels, can greatly reduce efficiency. Table 1 illuminates an alternative approach that can satisfy both acyclic and parallel constraints effectively. It reveals that the number of parallel stages corresponds to the multipartite quantity within each BayesNet. Since edges can only direct from topologically earlier nodes to later ones, nodes within the same partite can be processed simultaneously. Bayesian inference can thus be executed by sequentially processing each partite from the first to the last. This process not only maintains acyclicity but also confines the number of parallel stages to the number of partites. Hence, we propose to parameterize $p(\mathbf{G}_{intra})$ with categorical distributions at each node, each category corresponding to the node's topological index within the partites. Consequently, $p(\mathbf{G}_{intra})$ takes the following form: + +$$ +p \left(\mathbf {G} _ {\text {i n t r a}}\right) = \prod_ {c, h, w} ^ {C, H, W} p \left(\mathrm {T} _ {c, h, w}\right), \quad \mathrm {T} _ {c, h, w} = \text {C a t e g o r i c a l} \left(\pi_ {c, h, w}\right), \tag {12} +$$ + +where $\mathrm{T}_{c,h,w}$ denotes the topological indices of partites per categorical distribution, with $C,H,W$ being the node count along the channel, height, and width + +dimensions, respectively. Adjusting the number of partites (or parallel stages) $\dim \pi_{c,h,w} = S_{intra}$ can modulate the computation time. + +Optimization of $p(\mathbf{G}_{\text {intra }})$ Having constrained the number of parties to manage computational time, we now focus on learning the intra-node BayesNet structure to minimize bit-rate. The optimization challenge is formulated as: + +$$ +\min _ {p \left(\mathbf {G} _ {\text {i n t r a}}\right)} \mathbb {E} _ {p \left(\mathbf {G} _ {\text {i n t r a}}\right)} p \left(\mathbf {X}, \mathbf {L} \mid \mathbf {G} _ {\text {i n t r a}}\right), \quad \text {s . t .} \hat {\mathbf {G}} _ {\text {i n t r a}} \text {i s a c y c l i c ,} \tag {13} +$$ + +where $\hat{\mathbf{G}}_{\text {intra }}$ is sampled from $p(\mathbf{G}_{\text {intra }})$ . However, unlike the optimization for $\mathbf{G}_{\text {inter }}$ , partite indices are ordinal and don't lend themselves to being used as mixing weights, making the Gumbel-softmax reparameterization trick inapplicable. We therefore approach $\mathbf{G}_{\text {intra }}$ as discrete variables and employ the Monte-Carlo method with the VIMCO objective [26] for optimization: + +$$ +\mathcal {L} _ {\mathbf {G} _ {i n t r a}} = \mathbb {E} _ {q \left(\mathbf {G} _ {i n t r a} ^ {1: M} \mid \mathbf {X}\right)} \log \frac {1}{M} \sum_ {i = 1} ^ {M} \log p (\mathbf {X}, \mathbf {L} | \mathbf {G} _ {i n t r a} ^ {i}), \tag {14} +$$ + +where $\mathbf{G}_{intra}^{i}$ is the i-th Monte-Carlo sample and $M$ denotes number of samples. + +Moreover, note that Equation 12 presumes all partites are independent—an unlikely scenario as proximal nodes often exhibit stronger correlations. In practice, we can approximate $p(\mathbf{G}_{intra})$ with an unconditional generative model to consider such correlations and optimize this proxy instead, described as: + +$$ +p \left(\mathbf {G} _ {\text {i n t r a}}\right) \propto g (\mathbf {N}), \quad \mathbf {N} \sim \mathcal {N} (0, 1), \tag {15} +$$ + +with $g(\mathbf{N})$ symbolizing a generative network derived from standard normal noise. Implementation of the Intra-node BayesNet With the model $p(\mathbf{G}_{\text{intra}})$ established, we can draw samples of topological indices for parties, denoted $\hat{T}_{c,h,w}$ , from each respective $\mathrm{T}_{c,h,w}$ . The subsequent step involves establishing conditional dependencies among these groups. Since dependencies are restricted to nodes from different parties, a dynamic masked convolution operation is utilized. This operation ensures each node's dependencies are limited to preceding nodes within the convolutional kernel's range. For instance, the intra-node BayesNet for $\mathbf{y}$ can be characterized by: + +$$ +\prod_ {c, h, w} ^ {C, H, W} p \left(\mathrm {y} _ {c, h, w} \mid \mathbf {G} _ {\text {i n t r a}}\right) \propto \operatorname {d y n c o n v} \left(\hat {\mathrm {y}} _ {c, h, w}, W _ {c, h, w} \left(\hat {T} _ {c, h, w}\right)\right), \tag {16} +$$ + +where $\hat{\mathbf{y}}_{c,h,w}$ are samples from $\mathrm{y}_{c,h,w}$ , and $W_{c,h,w}$ represents a dynamically adjusted convolutional kernel influenced by adjacent nodes' topological indices. As an illustration, consider a 2D $5\times 5$ convolution: the dynamic kernel $W_{c,h,w}$ used for node $\mathrm{y}_{c,h,w}$ is defined by: + +$$ +W _ {c, h, w} [ i, j + 2, k + 2 ] = \left\{ \begin{array}{l l} W _ {i, j + 2, k + 2} & , \hat {T} _ {c, h, w} < \hat {T} _ {i, h + j, w + k} \\ 0 & , \hat {T} _ {c, h, w} \geq \hat {T} _ {i, h + j, w + k} \end{array} , \left\{ \begin{array}{l l} i & \in \{0, 1, \dots , C - 1 \} \\ j, k & \in \{- 2, - 1, 0, 1, 2 \} \end{array} . \right. \right. \tag {17} +$$ + +For in-depth guidance of its implementation, please refer to Appendix. + +![](images/18a6da2e6d528ffb3a32f95858e869156b4741c3c3f1867ceaf43d40964a58d8.jpg) +Fig. 2: The architecture of the implemented example based on the joint autoregressive framework [24]. Comp/Decomp means Compression / Decompression. AE/AD symbolize the arithmetic encoder and decoder, respectively. AR denotes the autoregressive model that executes the learned intra-node BayesNet. Q stands for quantization. $\pmb{x}$ correspond to input image and $\hat{\pmb{x}}$ the reconstructed image. + +# 5 From BayesNet to BaSIC + +Upon determining the BayesNet structure, we proceed to specify the nodes and edges in this BayesNet, crafting a novel NIC framework, BaSIC. This section details an implementation instance of the envisioned computationally scalable NIC, leveraging the prevalent joint autoregressive and hyperprior compression model [24]. It should be noted that this represents just one of the viable strategies for implementing BaSIC, with the potential for alternatives to offer improved results. The architecture of our example is depicted in Figure 2, and we will explore how it seamlessly integrates with the aforementioned BayesNets. $^{4}$ + +Inter-node BayesNet The inter-node BayesNet is constructed over each edge between nodes within the original BayesNet, specifically: + +$$ +\begin{array}{l} q (\mathbf {y} | \mathbf {x}, \mathbf {G} _ {i n t e r} ^ {\mathbf {y}, \mathbf {x}}) \propto g _ {a} (\mathbf {x}, \mathbf {G} _ {i n t e r} ^ {\mathbf {y}, \mathbf {x}}), \quad q (\mathbf {z} | \mathbf {y}, \mathbf {G} _ {i n t e r} ^ {\mathbf {z}, \mathbf {y}}) \propto h _ {a} (\mathbf {y}, \mathbf {G} _ {i n t e r} ^ {\mathbf {z}, \mathbf {y}}), \\ (1 1 1, \hat {\mathbf {x}} ^ {\hat {\mathbf {y}}} \hat {\mathbf {z}}), (1 1 1, \hat {\mathbf {z}} ^ {\hat {\mathbf {y}}} \hat {\mathbf {z}}). \end{array} +$$ + +$$ +p (\hat {\mathbf {y}} | \hat {\mathbf {z}}, \mathbf {G} _ {i n t e r} ^ {\hat {\mathbf {y}}, \hat {\mathbf {z}}}) \propto h _ {s} (\hat {\mathbf {z}}, \mathbf {G} _ {i n t e r} ^ {\hat {\mathbf {y}}, \hat {\mathbf {z}}}), \quad p (\hat {\mathbf {x}} | \hat {\mathbf {y}}, \mathbf {G} _ {i n t e r} ^ {\hat {\mathbf {x}}, \hat {\mathbf {y}}}) \propto g _ {s} (\hat {\mathbf {y}}, \mathbf {G} _ {i n t e r} ^ {\hat {\mathbf {x}}, \hat {\mathbf {y}}}) +$$ + +To achieve computational scalability, heterogeneous bipartite graphs are employed as outlined in Equation 10 for $\mathbf{G}_{inter}^{\mathbf{y},\mathbf{x}}$ , $\mathbf{G}_{inter}^{\mathbf{z},\mathbf{y}}$ , $\mathbf{G}_{inter}^{\hat{\mathbf{y}},\hat{\mathbf{z}}}$ , and $\mathbf{G}_{inter}^{\hat{\mathbf{x}},\hat{\mathbf{y}}}$ . The implementation of $g_{a}, g_{s}, h_{a}$ , and $h_{s}$ follow [29]'s slimmable networks, with dynamic channel widths including 48, 72, 96, 144, and 192 for computational scalability. Note that we use a fixed 192 channels for all latent variables $\mathbf{y}, \mathbf{z}$ , and only channels in the middle layers of each slimmable network are variable. This modification decouples bitrate scalability from computational scalability, enhancing the convenience in real-world applications. + +Intra-node BayesNet In the established framework, the intra-node BayesNet employs masked convolution applied to $\hat{\mathbf{y}}$ . We substitute this with our suggested multipartite graph-based intra-node BayesNet $p(\hat{\mathbf{y}}|\mathbf{G}_{intra}^{\hat{\mathbf{y}}})$ , which enables parallel computing on GPUs. The intra-node BayesNet uses a $5\times 5$ dynamic masked convolution as detailed in Equation 17, with $C$ equals to number of partites and $H,W$ equals to image height and width, respectively. Note that to facilitate the optimization of the intra-node BayesNet, the additional VIMCO loss expressed in Equation 14 should be incorporated into the overall loss function. + +Optimization The optimization process involves a loss function that simultaneously accounts for both the BayesNet structure (Equation 14) and the neural network parameters (Equation 4): + +$$ +\mathcal {L} = \mathcal {L} _ {R} + \lambda_ {D} \mathcal {L} _ {D} + \lambda_ {C} \mathcal {L} _ {C} + \mathcal {L} _ {\mathbf {G} _ {\text {i n t r a}}} \tag {19} +$$ + +Computational Scalability The described implementation enables computational scalability across two dimensions. Firstly, by modulating all $\mathbf{G}_{inter}$ , which govern the channel width in slimmable networks, we can fine-tune the computational load of the inter-node neural networks. $\mathbf{G}_{inter}$ can be modulated during training through the $\lambda_{C}$ parameter, or during inference by selecting from $\mathbf{G}_{inter}$ to alter channel widths. Notably, the latter can be streamlined by training with a uniform $\mathbf{G}_{inter}$ distribution and setting $\lambda_{C} = 0$ . Secondly, by varying the number of partites in $\mathbf{G}_{intra}^{\hat{y}}$ , we can adjust the stages of parallel computations in autoregressive models, thereby influencing the decompression time. In practice, we could train multiple autoregressive models with varying stages and choose among them during inference to modify decompression times. + +# 6 Experiments + +# 6.1 Experiment Setup + +Experiment Tasks Following previous works [29], our experiments focus on the task of lossy image compression with computational scalability. We benchmark our BaSIC framework against other computational scalable frameworks, as well as conduct ablation studies to validate the efficacy of the proposed intranode BayesNet learning methods. Further experimental results and details can be found in Appendix, including compression performance evaluation and comparison under fixed computational constraints and more ablation studies regarding the inter-node BayesNet. For evaluation, we employ Pixel Signal-Noise Ratio (PSNR) for distortion and Bits Per Pixel (BPP) for bit-rate, unless stated otherwise. + +Datasets We use a merged training dataset from ImageNet [10] and CLIC [2], preprocessed following the method in [11]. The Kodak [1] dataset serves for validation and testing. + +Implementation Details For the proposed method, we use the joint autoregressive framework [24] as the basis, as described in Section 5. Unless otherwise + +![](images/37f1dc03aa2b1ae77866c639d05da9f06da593941c7ed1f3296224cdd07abfea.jpg) +Fig. 3: Comparative results for different computational scalable frameworks. + +![](images/08457f6224055bd3195c2cf2d1d289282646b78868c5913c862c42702b6c0fb9.jpg) + +specified, we follow CompressAI [5] to set the hyperparameter $\lambda_{D}$ , while employing $\lambda_{C} = 0$ during training and a greedy search among 16 intra-node BayesNets with varied FLOPs during inference. Training spans approximately 500k iterations, employing the Adam optimizer with a learning rate of 0.0001. To compare running times equitably, all experiments are performed on identical hardware (an Intel i7-6800K CPU, NVIDIA Titan Xp GPU, and 32GB of memory). + +# 6.2 Comparison to Other Computation Scalable Frameworks + +This section assesses our computational scalable BaSIC method against other scalable frameworks. We configure our framework without autoregressive modules, similar to SlimCAE [29], to compare performance at various computational levels. We adopt Hyperprior [4] as a reference, reporting BD-rate for diverse MACs for both methods. Also, the rate-distortion plots for SlimCAE's full range of complexity and three levels (L0, L7, L15) for our method are included for reference. Note that $\lambda$ -scheduling is deactivated in SlimCAE to facilitate uniform BD-rate analysis across levels. The results are depicted in Figure 3. + +Our method, employing the same slimmable network backbone as SlimCAE, demonstrates a marginal performance edge at most computation levels, likely attributable to the fixed channels of latent variables that enhance higher bit-rate performance. Specifically, our method saves around $2\%$ more bits than Slim-CAE under the maximum complexity, and around $10\%$ more bits under the minimum complexity. While SlimCAE [29] employs $\lambda$ -scheduling to tackle this, it inadvertently ties bit-rate and distortion to computational complexity, which is less desirable for cases like low-powered devices needing high-quality images. In contrast, our inter-node BayesNet independently modulates the computational complexity of individual neural network modules, as evidenced by the rate-distortion plots that show minimal impact on performance with reduced computation. Moreover, the varied complexity levels indicate our method's superior control over computational complexity, allowing us to assign different inter-node BayesNets (channel widths) to distinct layers. SlimCAE, conversely, applies the same channel width across all layers, limiting flexibility. For instance, + +Table 2: Bit-rate comparison for different autoregressive modules. + +
Stages-2410
MethodsMaskConvBaselineCBCWOursMSOursELICOurs
BPP0.31500.33330.32160.32220.32070.32100.31830.31740.3156
Speed (MB/s)0.32345.55191.23173.46172.57107.9374.6717.8517.32
+ +to meet a 200kMAC/pixel threshold, our closest model uses 158kMAC/pixel, while SlimCAE's nearest option drops to 111kMAC/pixel, as shown in Figure 3. + +# 6.3 Comparison to Other Autoregressive Models for Intra-Node BayesNet + +In this section, we evaluate the effectiveness of the proposed intra-node BayesNet structure learning method, by comparing it to other hand-crafted autoregressive models. For a simple and fair comparison, we first establish a baseline using a joint autoregressive framework [24] with a null context model, trained for about 500k iterations. Subsequently, we retain all other components while training various autoregressive models to replace the baseline context model for an additional 500k iterations. This approach ensures the reconstructed image remains unchanged, allowing us to isolate and evaluate the bit-rate to reflect the effectiveness of different autoregressive models. Moreover, the decompression speed is also reported to evaluate the efficiency of It's important to note that the neural network architecture for the context model is kept constant across all methods, with the sole variability stemming from the structure of the intra-node BayesNet. + +The benchmark autoregressive models include Channel-wise (CW, 2-stage) [25], Checkerboard [12] (CB, 2-stage), Multi-stage spatial $2 \times 2$ [21] (MS, 4-stage) and the combined unevenly grouped channel-wise and checkerboard [11] (ELIC, 10-stage in total). We also train the MaskConv context model from [24] as a baseline. Our model is trained under constraints of 2, 4, and 10 parallel stages and evaluated against counterparts with equivalent stage configurations. The findings are presented in Table 2. + +Overall, our learning-based BayesNet structures consistently demonstrate slightly better performance compared to the manually-crafted models under the same parallel stages, with similar decompression speed. Interestingly, in each group our learned method reaches close BPP to the hand-crafted ones in the next level of parallel stage, and in the 10-stage case, our method even approaches the BPP of non-parallel MaskConv model. This proves that our proposed method excels in learning effective intra-node BayesNets for autoregressive models. + +Qualitative Results We also illustrate the learned intra-node BayesNet by our method as well as the manually-designed one of ELIC in Figure 4 to analyze their difference with hand-crafted models. We could see that the BayesNet of ELIC uses a checkerboard context model over the spatial dimension, including only two + +![](images/c1faa6ec35b794b24ae4d04e1afa693041b07d6f2bde0e22603d802af43ef0fe.jpg) + +![](images/a8b4921d5297f5a418bfcf70c08d0fc9cef573a0a4f74633ad210f16d3b6b3b5.jpg) +(a) ELIC +(b) Ours-10stage +Fig. 4: Partite topological indices visualization for intra-node BayesNet in the 10-stage case. Each image patch indicates topological indices of nodes in a group of channels. Blacker pixels indicate lower topological indices and whiter indicate higher topological indices. E.g., in the 10-stage case, pure-black indicates index 0 which is first iterated, and pure-white indicates group 9 which is last iterated. Better viewed with zoom-in. + +partites within each channel. In comparison, our learned BayesNet incorporates more different partites (mostly three or four) in the spatial dimension, which captures more context information than the checkerboard case. This is reasonable as images are usually heavily correlated in spatial dimensions. Moreover, the learned topological indices also give some insight for designing parallel autoregressive models, that channel and spatial context models could be combined more compactly to improve their performance. + +# 7 Conclusions and Discussions + +This paper presents a novel approach, BaSIC, to addressing computational scalability in Neural Image Compression (NIC) by focusing on BayesNet structure learning. We decompose the challenge into two subproblems: heterogeneous bipartite graph learning for inter-node BayesNet, and multipartite graph learning for intra-node BayesNet, proposing solutions for each. Experimental results, using a joint autoregressive framework as a test case, show that our work offers improved control over computational demands relative to other scalable computation models. + +# Acknowledgement + +The paper is supported in part by the National Natural Science Foundation of China (No. 62325109, U21B2013). + +# References + +1. Kodak lossless true color image suite. https://r0k.us/graphics/kodak/ (2023), accessed: 2023-11-12 +2. Workshop and challenge on learned image compression. https://click.compression.cc/ (2023), accessed: 2023-01-09 +3. Balle, J., Laparra, V., Simoncelli, E.P.: Density modeling of images using a generalized normalization transformation. CoRR abs/1511.06281 (2015), https://apisemantic scholar.org/CorpusID:2684987 +4. Balle, J., Minnen, D.C., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. ArXiv abs/1802.01436 (2018) +5. Bégaint, J., Racapé, F., Feltman, S., Pushparaja, A.: Compressai: a pytorch library and evaluation platform for end-to-end compression research. arXiv preprint arXiv:2011.03029 (2020) +6. Cheng, J., Greiner, R., Kelly, J., Bell, D.A., Liu, W.: Learning bayesian networks from data: An information-theory based approach. Artif. Intell. 137, 43-90 (2002), https://api-semanticscholar.org/CorpusID:7821347 +7. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 7936-7945 (2020) +8. Chickering, D.M.: Optimal structure identification with greedy search. J. Mach. Learn. Res. 3, 507-554 (2003), https://api-semanticscholar.org/CorpusID:1191614 +9. Deleu, T., G'ois, A., Emezue, C.C., Rankawat, M., Lacoste-Julien, S., Bauer, S., Bengio, Y.: Bayesian structure learning with generative flow networks. ArXiv abs/2202.13903 (2022), https://api(semanticscholar.org/CorpusID: 247158659 +10. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition pp. 248-255 (2009) +11. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5708-5717 (2022), https://apisemantic scholar.org/CorpusID:247594672 +12. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14766-14775 (2021), https://apisemantic scholar.org/CorpusID:232404386 +13. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:326772 + +14. Jang, E., Gu, S.S., Poole, B.: Categorical reparameterization with gumbel-softmax. ArXiv abs/1611.01144 (2016) +15. Johnston, N., Eban, E., Gordon, A., Ball'e, J.: Computationally efficient neural image compression. ArXiv abs/1912.08771 (2019), https://apisemantic scholar.org/CorpusID:209404937 +16. Kang, N., Qiu, S., Zhang, S., Li, Z., Xia, S.: Pilc: Practical image lossless compression with an end-to-endgpu oriented neural framework. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3729-3738 (2022), https://apisemantic scholar.org/CorpusID:249625569 +17. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. CoRR abs/1312.6114 (2013) +18. Koivisto, M., Sood, K.: Exact bayesian structure discovery in bayesian networks. J. Mach. Learn. Res. 5, 549-573 (2004), https://api(semanticscholar.org/ CorpusID:12412451 +19. Lee, J., Cho, S., Beack, S.: Context-adaptive entropy model for end-to-end optimized image compression. In: International Conference on Learning Representations (2018) +20. Li, M., Ma, K., You, J.J., Zhang, D., Zuo, W.: Efficient and effective context-based convolutional entropy modeling for image compression. IEEE Transactions on Image Processing 29, 5900-5911 (2019), https://api-semanticscholar.org/CorpusID:195345342 +21. Lin, F., Sun, H., Liu, J., Katto, J.: Multistage spatial context models for learned image compression. ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 1-5 (2023), https://apisemantic scholar.org/CorpusID:257038187 +22. Liu, H., Simonyan, K., Yang, Y.: Darts: Differentiable architecture search. ArXiv abs/1806.09055 (2018), https://api-semanticscholar.org/CorpusID: 49411844 +23. Lorch, L., Rothfuss, J., Scholkopf, B., Krause, A.: Dibs: Differentiable bayesian structure learning. ArXiv abs/2105.11839 (2021), https://apisemantic scholar.org/CorpusID:235187432 +24. Minnen, D.C., Balle, J., Toderici, G.: Joint autoregressive and hierarchical priors for learned image compression. ArXiv abs/1809.02736 (2018) +25. Minnen, D.C., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. 2020 IEEE International Conference on Image Processing (ICIP) pp. 3339-3343 (2020), https://api-semanticscholar.org/CorpusID:220633468 +26. Mnih, A., Rezende, D.J.: Variational inference for monte carlo objectives. In: International Conference on Machine Learning (2016), https://api-semanticscholar.org/CorpusID:5859948 +27. Tao, L., Gao, W., Li, G., Zhang, C.: Adanic: Towards practical neural image compression via dynamic transform routing. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 16833-16842 (2023), https://api.sementicscholar.org/CorpusID:267026318 +28. webmproject: libwebp. https://github.com/webmproject/libwebp (2023), accessed: 2023-01-09 +29. Yang, F., Herranz, L., Cheng, Y., Mozerov, M.G.: Slimmable compressive autoencoders for practical neural image compression. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 4996-5005 (2021), https://api-semanticscholar.org/CorpusID:232417114 + +30. Yin, S., Li, C., Bao, Y., Liang, Y.: Universal efficient variable-rate neural image compression. ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 2025-2029 (2021), https://api.sementicscholar.org/CorpusID:244478380 +31. Yin, S., Meng, F., Tan, W.T., Li, C., Bao, Y., Liang, Y., Liu, W.: Exploring structural sparsity in neural image compression. 2022 IEEE International Conference on Image Processing (ICIP) pp. 471-475 (2022), https://api-semanticscholar.org/CorpusID:246680031 +32. Yu, J., Yang, L., Xu, N., Yang, J., Huang, T.S.: Slimmable neural networks. ArXiv abs/1812.08928 (2018), https://api(semanticscholar.org/CorpusID:56657799 +33. Yu, Y., Chen, J., Gao, T., Yu, M.: Dag-gnn: Dag structure learning with graph neural networks. In: International Conference on Machine Learning (2019), https://apisemantic scholar.org/CorpusID:128358697 +34. Yuan, C., Malone, B.M.: Learning optimal bayesian networks: A shortest path perspective. J. Artif. Intell. Res. 48, 23-65 (2013), https://api-semanticscholar.org/CorpusID:5951724 +35. Zhang, Y., Lu, G., Feng, D., Zhu, C., Song, L.: Content adaptive checkerboard context model for learned image compression. 2023 IEEE International Symposium on Circuits and Systems (ISCAS) pp. 1-5 (2023), https://api-semanticscholar.org/CorpusID:260002843 +36. Zhang, Z., Chen, B., Lin, H., Lin, J., Wang, X., Zhao, T.: Elfic: A learning-based flexible image codec with rate-distortion-complexity optimization. Proceedings of the 31st ACM International Conference on Multimedia (2023), https://apisemantic scholar.org/CorpusID:264492165 +37. Zheng, X., Aragam, B., Ravikumar, P., Xing, E.P.: Dags with no tears: Continuous optimization for structure learning. In: Neural Information Processing Systems (2018), https://api_semanticscholar.org/CorpusID:53217974 +38. Zheng, Z., Wang, X., Lin, X., Lv, S.: Get the best of the three worlds: Real-time neural image compression in a non-gpu environment. Proceedings of the 29th ACM International Conference on Multimedia (2021), https://api-semanticscholar.org/CorpusID:239011955 \ No newline at end of file diff --git a/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/images.zip b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c2292df15433c9304087bf572cae9edc66a7447e --- /dev/null +++ b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d8c34e24a0143444f4b1585fbde2c0551e653485b33a03cd57249c8ec0b0664 +size 494134 diff --git a/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/layout.json b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3d208d520f947f60da46449feffdb7e8f1941b44 --- /dev/null +++ b/2024/BaSIC_ BayesNet Structure Learning for Computational Scalable Neural Image Compression/layout.json @@ -0,0 +1,10075 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 111, + 449, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 111, + 449, + 166 + ], + "spans": [ + { + "bbox": [ + 165, + 111, + 449, + 166 + ], + "type": "text", + "content": "BaSIC: BayesNet Structure Learning for Computational Scalable Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "spans": [ + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "text", + "content": "Yufeng Zhang" + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "text", + "content": ", Hang Yu" + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "text", + "content": ", Shizhan Liu" + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "text", + "content": ", Wenrui Dai" + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "text", + "content": ", and Weiyao Lin" + }, + { + "bbox": [ + 200, + 186, + 412, + 210 + ], + "type": "inline_equation", + "content": "^{\\star 1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 186, + 219, + 427, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 219, + 427, + 254 + ], + "spans": [ + { + "bbox": [ + 186, + 219, + 427, + 254 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 219, + 427, + 254 + ], + "type": "text", + "content": " Shanghai Jiao Tong University, China. {worldlife,shanluzuode,daiwenrui,wylin}@sjtu.edu.cn " + }, + { + "bbox": [ + 186, + 219, + 427, + 254 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 186, + 219, + 427, + 254 + ], + "type": "text", + "content": " Ant Group, China. hyu.hugo@antgroup.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 285, + 452, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 285, + 452, + 493 + ], + "spans": [ + { + "bbox": [ + 160, + 285, + 452, + 493 + ], + "type": "text", + "content": "Abstract. Despite superior rate-distortion performance over traditional codecs, Neural Image Compression (NIC) is limited by its computational scalability in practical deployment. Prevailing research focuses on accelerating specific NIC modules but is restricted in controlling overall computational complexity. To this end, this work introduces BaSIC (BayesNet structure learning for computational Scalable neural Image Compression), a comprehensive, computationally scalable framework that affords full control over NIC processes. We learn the Bayesian network (BayesNet) structure of NIC for controlling both neural network backbones and autoregressive units. The learning of BayesNet is achieved by solving two sub-problems, i.e., learning a heterogeneous bipartite BayesNet for the inter-node structure to regulate backbone complexity, and a multipartite BayesNet for the intra-node structure to optimize parallel computation in autoregressive units. Experiments demonstrate that our method not only facilitates full computational scalability with more accurate complexity control but also maintains competitive compression performance compared to other computation scalable frameworks under equivalent computational constraints. Code will be available in https://github.com/worldlife123/cbench_BaSIC." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 504, + 452, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 504, + 452, + 526 + ], + "spans": [ + { + "bbox": [ + 160, + 504, + 452, + 526 + ], + "type": "text", + "content": "Keywords: Image Compression " + }, + { + "bbox": [ + 160, + 504, + 452, + 526 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 504, + 452, + 526 + ], + "type": "text", + "content": " Computational Scalable Compression " + }, + { + "bbox": [ + 160, + 504, + 452, + 526 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 504, + 452, + 526 + ], + "type": "text", + "content": " BayesNet Structure Learning " + }, + { + "bbox": [ + 160, + 504, + 452, + 526 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 504, + 452, + 526 + ], + "type": "text", + "content": " Autoregressive Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 548, + 230, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 548, + 230, + 560 + ], + "spans": [ + { + "bbox": [ + 133, + 548, + 230, + 560 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Neural Image Compression (NIC) technologies [7, 11] have significantly outstripped the performance of conventional image CODECs, ushering in an era of intelligent image compression. These advanced CODEcs are intended for deployment across a spectrum of devices, from low-end devices like mobile phones to high-end devices like GPU computation clusters. Each type of device necessitates that CODEcs tailor their performance to specific computational or memory" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 653, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 238, + 665 + ], + "type": "text", + "content": "* Corresponding Author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 113, + 286, + 274 + ], + "blocks": [ + { + "bbox": [ + 145, + 113, + 286, + 274 + ], + "lines": [ + { + "bbox": [ + 145, + 113, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 145, + 113, + 286, + 274 + ], + "type": "image", + "image_path": "b7e40b9eb529fe36748b0cfe1367f5f601611e68962e546044fcb5a92bd5724d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 283, + 480, + 306 + ], + "lines": [ + { + "bbox": [ + 131, + 283, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 283, + 480, + 306 + ], + "type": "text", + "content": "Fig. 1: Bayesian network diagrams for various NIC frameworks. Our proposed framework BaSIC (d) generalizes the concepts in (a), (b), and (c)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 288, + 121, + 470, + 274 + ], + "blocks": [ + { + "bbox": [ + 288, + 121, + 470, + 274 + ], + "lines": [ + { + "bbox": [ + 288, + 121, + 470, + 274 + ], + "spans": [ + { + "bbox": [ + 288, + 121, + 470, + 274 + ], + "type": "image", + "image_path": "99a995a3c841f19090b791a708db2180f96b1e6ef47cc7830340e0c89c41ad5a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 330, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 479, + 376 + ], + "type": "text", + "content": "constraints, a requirement known as Computational Scalability. High-end devices can endure more intricate computations for superior encoding quality, whereas low-end devices may prioritize reduced computational load over encoding efficiency to conserve energy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 480, + 544 + ], + "type": "text", + "content": "While conventional image CODECs typically accommodate this functionality (e.g. WebP [28] offers seven levels of computational complexity to regulate lossy image compression), incorporating this attribute into NIC remains a significant hurdle for its broader adoption. Recent studies have focused predominantly on accelerating NIC computations rather than addressing computational scalability. Works such as [11, 12, 25] target the acceleration of autoregressive modules on GPUs, which are frequently used in NIC. Others, including [15, 31, 38], explore neural network acceleration techniques—given that NIC fundamentally relies on neural network architectures. On the other hand, some recent efforts have attempted to integrate dynamic neural networks like slimmable networks [32] within NIC frameworks to adjust the computational complexity [27, 29, 36]. However, these approaches have only partially addressed computational scalability and often neglect crucial components in NIC, such as autoregressive modules, which are essential for compression performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "In this work, we endeavor to create a NIC framework with enhanced computational scalability, incorporating computational control across all NIC components. We call this framework BaSIC (BayesNet structure learning for computational Scalable neural Image Compression). To this end, we re-examine NIC through the broader lens of the Bayesian network (BayesNet) paradigm. Figure 1 illustrates the BayesNet depiction of several exemplary NIC architectures. Specifically, our approach is rooted in the hyperprior BayesNet framework (Figure 1a), which postulates that the generative model can be factorized as " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{y})p(\\mathbf{y}|\\mathbf{z})p(\\mathbf{z})" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " represents observed data, " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " the prior, and " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " the hyperprior. In this context, SlimCAE (Figure 1b) can be interpreted as allowing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "selective edge types between " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ", and between " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ", with each type manifesting different conditional dependencies, modeled via slimmable neural networks with varied channel widths. On the other hand, the joint autoregressive framework (Figure 1c) introduces autoregressive intra-node dependencies within " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " to enhance compression, at the expense of increased computational demands due to the requisite sequential processing [12]. These examples clearly demonstrate the intimate connection between the structure of the BayesNet and the complexity inherent in NIC frameworks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "text", + "content": "Our advancement is a NIC framework that exploits end-to-end BayesNet structure learning [23, 33] to meticulously optimize computational scalability. This innovative framework is not bound to specific neural network backbones, making it versatile enough to enhance computational scalability in most existing NIC frameworks with a hyperprior BayesNet representation. Our proposed framework is visualized in Figure 1d. In particular, we represent the BayesNet structure as " + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "text", + "content": " and impose a prior " + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G})" + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "text", + "content": " on it. Given that the structure we seek to learn must conform to a directed acyclic graph (DAG), we tailor the prior " + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G})" + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "text", + "content": " for compression frameworks within this DAG constraint, considering two distinct instances: the inter-node BayesNet, representing conditional dependencies between nodes, and the intra-node BayesNet, which pertains to dependencies within nodes. We depict the former as a heterogeneous bipartite graph and design a structure learning method to seamlessly incorporate dynamic neural networks like slimmable networks [32]. For the latter, we design a network architecture parameterized by multipartite or " + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "text", + "content": "-partite graphs to enable a parallel autoregressive process. By manipulating the partite number " + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 214, + 482, + 430 + ], + "type": "text", + "content": ", we can adjust the temporal demands of the autoregressive process, thereby directly impacting decompression speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 433, + 483, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 483, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 483, + 517 + ], + "type": "text", + "content": "When deploying the codec, the configurations of both the inter-node and intra-node BayesNets can be tailored to match the computational budget of target devices. For instance, on less capable devices, we might employ simpler neural network edges within the inter-node BayesNet and a smaller partite number within the intra-node BayesNet, whereas more capable devices can leverage a more complex network and a larger partite number, allowing for scalable computational resource allocation based on the specifics of the device at hand." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 520, + 389, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 520, + 389, + 532 + ], + "spans": [ + { + "bbox": [ + 146, + 520, + 389, + 532 + ], + "type": "text", + "content": "To summarize, our contributions are mainly three-fold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 552, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 552, + 481, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 552, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 552, + 481, + 587 + ], + "type": "text", + "content": "- We propose an end-to-end computationally scalable NIC framework founded on BayesNet structure learning, facilitating fine-tuned adjustments to computational complexity across all facets of the NIC process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 590, + 481, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 625 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 625 + ], + "type": "text", + "content": "- We devise a generative model based on a heterogeneous bipartite structure for the inter-node BayesNet, harnessing the capabilities of existing dynamic neural networks to realize computational scalability effectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "type": "text", + "content": "- We design a generative model structured as a multipartite graph for the intra-node BayesNet. This model is tailored to accommodate and streamline parallel computation in practical applications." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 163 + ], + "type": "text", + "content": "In this section, we briefly review recent works regarding the computational complexity of NIC as well as BayesNet structure learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 180, + 357, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 180, + 357, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 357, + 192 + ], + "type": "text", + "content": "2.1 End-to-end Neural Image Compression" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 198, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 198, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 481, + 282 + ], + "type": "text", + "content": "The hyperprior framework [4] serves as a foundation for most end-to-end NIC models, favored for its effective compression performance and straightforward implementation. Subsequent studies [7, 19, 24] have enhanced the hyperprior by integrating a masked convolution-based autoregressive model in the latent space to reduce bit-rate. While these improvements offer better compression, they often overlook the increased computational demands. To apply NIC in practice, some recent works are now addressing the need to expedite these models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 283, + 481, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 283, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 481, + 413 + ], + "type": "text", + "content": "Autoregressive Model Acceleration To circumvent the slow serial iteration steps of masked convolutions on GPUs, newer studies introduce parallelized autoregressive models for GPU acceleration. For instance, [25] presents a channelwise parallel autoregressive model, while [12] proposes a spatially parallel model using a checkerboard mask, requiring just two parallel stages. This concept is extended to a multistage model by [21]. Further, [11] amalgamates channel-wise and spatially parallel autoregressive models. Additional acceleration methods include content-adaptive adjustments to parallel autoregressive models [35] and designing simplified autoregressive models [16]. Yet, such models are typically hand-crafted and may not represent the most efficient trade-off between compression complexity and performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 415, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 481, + 498 + ], + "type": "text", + "content": "Neural Network Acceleration Considering that neural networks (i.e., edges in the BayesNets) represent a significant portion of NIC's computational complexity, some works focus on accelerating these directly. For example, there are designs for more efficient neural network operators like a simplified GDN [15] and CPU-friendly down/up-samplers [38]. Others apply general neural network acceleration techniques within compression models, such as MorphNet [15] and learnable channel pruning [30, 31]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 498, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 594 + ], + "type": "text", + "content": "Computation Scalable Framework A handful of recent works have also noticed the importance of computational scalability for NIC. SlimCAE [29] employs slimmable networks [32] as the framework's backbone, enabling dynamic adjustments of bit-rate and computation through channel width modulation. Further developments include universal slimmable networks for more precise scalability [36] and the integration of input-adaptive complexity allocation methods [27,36]. Nonetheless, these approaches focus solely on the scalability of neural networks, largely disregarding the autoregressive models' scalability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 610, + 308, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 308, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 308, + 623 + ], + "type": "text", + "content": "2.2 BayesNet Structure Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "content": "BayesNet structure learning aims to discern the conditional relationships among a set of variables with a graph representation, which ideally should be a directed acyclic graph (DAG). Traditional approaches often involve score-based [8, 34]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": "and constraint-based methods [6, 18]. However, they typically operate within a constrained search space, which can lead to less-than-optimal graphs. More recent initiatives introduce neural network techniques for more robust graph generation within the realm of BayesNet structure learning. Some align this learning task with variational inference to facilitate graph structure development within an end-to-end framework [23, 33], utilizing graph acyclicity regularizers [37] that may not guarantee acyclicity. Alternatively, other research employs recurrent-style graph generators that assure acyclicity through iterations [9], but these can become inefficient with an increase in the number of nodes. In our work, we diverge from these established methods due to their inefficiency when dealing with the dense pixelation of images. Instead, we analyze the BayesNet representations inherent in existing NIC frameworks and directly parameterize the graph structure as DAGs, ensuring both acyclicity and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 291, + 269, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 269, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 269, + 304 + ], + "type": "text", + "content": "3 Overview of BaSIC" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "text", + "content": "Assume we have a set of observations " + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_N\\}^3" + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "text", + "content": ", the objective of BayesNet structure learning is to infer a distribution over the structure of the BayesNet " + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{G} = (\\mathbf{V}, \\mathbf{E}, \\boldsymbol{\\theta})" + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "text", + "content": " that best describes the observations. According to the Bayesian formalism, such optimization problem could be solved by minimizing expectation over " + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}|\\mathbf{X})" + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "text", + "content": " of any function " + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "text", + "content": " related to " + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "text", + "content": ", under the constraint that " + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 317, + 482, + 389 + ], + "type": "text", + "content": " is acyclic:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 224, + 399, + 481, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 399, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 224, + 399, + 481, + 417 + ], + "type": "interline_equation", + "content": "\\underset {\\mathbf {G}} {\\min } \\mathbb {E} _ {p (\\mathbf {G} | \\mathbf {X})} [ f (\\mathbf {G}) ], \\quad \\mathrm {s . t .} \\mathbf {G} \\text {i s a c y c l i c .} \\tag {1}", + "image_path": "656a5d90b05c95ce6f63c176b865a0251a0b63d4f6491410d8e43081ce416c9b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 480, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 480, + 461 + ], + "type": "text", + "content": "For a computationally scalable compression framework, we tackle BayesNet structure learning, linking BayesNet configurations to computational complexity and shaping the function " + }, + { + "bbox": [ + 130, + 425, + 480, + 461 + ], + "type": "inline_equation", + "content": "f(\\mathbf{G})" + }, + { + "bbox": [ + 130, + 425, + 480, + 461 + ], + "type": "text", + "content": " accordingly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": "Apart from modeling of " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": ", the compression framework also involves latent variables " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = \\{\\mathbf{y}_1, \\mathbf{y}_2, \\ldots, \\mathbf{y}_N, \\mathbf{z}_1, \\mathbf{z}_2, \\ldots, \\mathbf{z}_N, \\ldots\\}" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": ". The resulting BayesNet should capture conditional dependencies within and between " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": ". This generative relationship is described by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 520, + 481, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 520, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 227, + 520, + 481, + 533 + ], + "type": "interline_equation", + "content": "p (\\mathbf {G}, \\mathbf {L}, \\mathbf {X}) = p (\\mathbf {G}) p (\\mathbf {L} | \\mathbf {G}) p (\\mathbf {X} | \\mathbf {G}, \\mathbf {L}). \\tag {2}", + "image_path": "0f6c46a18089aeec02a328d97739560dc506a7b656e4289150d7934d8113d033.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 541, + 317, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 317, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 317, + 553 + ], + "type": "text", + "content": "The expectation from Equation 1 is then:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 205, + 564, + 481, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 564, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 205, + 564, + 481, + 578 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {p (\\mathbf {G}, \\mathbf {L} | \\mathbf {X})} [ f (\\mathbf {G}, \\mathbf {L}) ] = \\mathbb {E} _ {p (\\mathbf {G} | \\mathbf {X}) p (\\mathbf {L} | \\mathbf {X}, \\mathbf {G})} [ f (\\mathbf {G}, \\mathbf {L}) ]. \\tag {3}", + "image_path": "69e45abb1b34d95fa420568715855e286dbb19e6944e3b12de3b230895e7644f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "text", + "content": "Existing compression frameworks without " + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "text", + "content": " approximate " + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "inline_equation", + "content": "p(\\mathbf{L}|\\mathbf{X})" + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "text", + "content": " with an inference network " + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "inline_equation", + "content": "q(\\mathbf{L}|\\mathbf{X})" + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "text", + "content": ", as in VAEs [17]. Incorporating " + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "text", + "content": " into our design, we refine our inference network to " + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "inline_equation", + "content": "q(\\mathbf{G},\\mathbf{L}|\\mathbf{X})" + }, + { + "bbox": [ + 130, + 586, + 482, + 624 + ], + "type": "text", + "content": " for enhanced approximation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 205, + 632, + 481, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 632, + 481, + 646 + ], + "spans": [ + { + "bbox": [ + 205, + 632, + 481, + 646 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {p (\\mathbf {G}, \\mathbf {L} | \\mathbf {X})} [ f (\\mathbf {G}, \\mathbf {L}) ] \\approx \\mathbb {E} _ {q (\\mathbf {G} | \\mathbf {X}) q (\\mathbf {L} | \\mathbf {X}, \\mathbf {G})} [ f (\\mathbf {G}, \\mathbf {L}) ]. \\tag {4}", + "image_path": "7b780bcbe6121987cbfde7d55f471074bd61ef10c5a0736395bc3f26a809b53d.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 364, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 364, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 364, + 666 + ], + "type": "text", + "content": "3 All symbols used in this paper are listed in Appendix." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "This approach allows " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " to direct the computational complexity of both the inference and the generative processes. Furthermore, we adopt a differentiable approach to learning " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ", in line with recent advances [23, 33], and incorporate this aspect into our model specification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "content": "Definition of " + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "inline_equation", + "content": "f(\\mathbf{G}, \\mathbf{L})" + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "content": " In line with lossy image compression standards [15,31], our optimization combines bit-rate " + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "content": ", distortion " + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "content": ", and complexity " + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 150, + 198, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 198, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 150, + 198, + 481, + 213 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {q (\\mathbf {G} | \\mathbf {X}) q (\\mathbf {L} | \\mathbf {X}, \\mathbf {G})} [ f (\\mathbf {G}, \\mathbf {L}) ] = \\mathcal {L} _ {R} (\\mathbf {G}, \\mathbf {L}) + \\lambda_ {D} \\mathcal {L} _ {D} (\\mathbf {G}, \\mathbf {L}) + \\lambda_ {C} \\mathcal {L} _ {C} (\\mathbf {G}, \\mathbf {L}). \\tag {5}", + "image_path": "ab1e06449e8e9909949a2907e7bffdd472335de39790f130060c7f49a715b4f6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 220, + 479, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 479, + 244 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 479, + 244 + ], + "type": "text", + "content": "Concretely, " + }, + { + "bbox": [ + 131, + 220, + 479, + 244 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 131, + 220, + 479, + 244 + ], + "type": "text", + "content": " is computed by enacting the BayesNet's generative process with latent node samples " + }, + { + "bbox": [ + 131, + 220, + 479, + 244 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}" + }, + { + "bbox": [ + 131, + 220, + 479, + 244 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 207, + 254, + 481, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 254, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 207, + 254, + 481, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} (\\mathbf {G}, \\mathbf {L}) = \\log_ {2} p _ {\\mathbf {L} | \\mathbf {G}} (\\hat {\\mathbf {L}} | \\mathbf {G}), \\quad \\hat {\\mathbf {L}} \\sim q (\\mathbf {L} | \\mathbf {X}, \\mathbf {G}) \\tag {6}", + "image_path": "dbef8eafd200d5cc586b968737f329e5f26df515edc60edb09ef2b565e8dfa5b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "text", + "content": "Moreover, " + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "text", + "content": " can be tailored to specific optimization objectives. " + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "text", + "content": " may encompass a range of complexity measures or their amalgamations, such as conventional metrics like Mean-Squared Error (MSE) or more advanced machine learning metrics like Frechet Inception Distance (FID) [13]. Conversely, " + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "text", + "content": " can include various complexity metrics or their combinations, such as the number of Floating-point Operations (FLOPs), network parameters, time, or energy consumption. " + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\lambda_{D}" + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\lambda_{C}" + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "text", + "content": " are the corresponding weights. By setting " + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "inline_equation", + "content": "f(\\mathbf{G},\\mathbf{L})" + }, + { + "bbox": [ + 130, + 277, + 480, + 384 + ], + "type": "text", + "content": " in line with the established rate-distortion (R-D) trade-off, Equation 4 falls in step with loss functions prevalent in standard compression frameworks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": "Definition of " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "q(\\mathbf{G}|\\mathbf{X})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " As a determiner of the BayesNet structure, " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " affects computational complexity. In practical compression scenarios, including " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " in the bitstream could impact the bit-rate. Typically, it may be more advantageous to treat " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " as independent of " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": ", suggesting " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "q(\\mathbf{G}|\\mathbf{X}) = p(\\mathbf{G})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": ". However, in scenarios where complexity is tailored to input data, the model " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "q(\\mathbf{G}|\\mathbf{X})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " can be employed. DAG Constraint of " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " As previously noted in Equation 1, " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " should reflect a distribution over DAG structures. For edges between different node groups in compression models, such as " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y}|\\mathbf{z})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": ", ensuring a DAG structure is straightforward since bipartite graphs are inherently acyclic; we call this component the Inter-node BayesNet (" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": "). For conditional dependencies within node groups, like " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "p(\\mathrm{y}_i|\\mathrm{y}_{i-1})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " which also require DAG adherence, we refer to it as the Intra-node BayesNet (" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{intra}" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": "). The forthcoming sections will delve into the methodology for learning both " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{inter})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{intra})" + }, + { + "bbox": [ + 130, + 384, + 481, + 540 + ], + "type": "text", + "content": " as DAGs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 559, + 479, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 479, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 479, + 572 + ], + "type": "text", + "content": "4 Structure Learning for Inter and Intra-Node BayesNets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 585, + 449, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 449, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 449, + 597 + ], + "type": "text", + "content": "4.1 Heterogeneous Bipartite Graphs for Inter-node BayesNet" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "text", + "content": "In this section, we specify " + }, + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{inter})" + }, + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}" + }, + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "text", + "content": " is a bipartite graph, and so it inherently satisfies the acyclic constraint." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "text", + "content": "Naive Bipartite Parameterization of " + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{inter})" + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "text", + "content": " Consider the task of learning a bipartite graph for " + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y}|\\mathbf{z})" + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "text", + "content": ". We split the nodes into " + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1, \\mathbf{z}_2, \\ldots, \\mathbf{z}_I" + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_1, \\mathbf{y}_2, \\ldots, \\mathbf{y}_J" + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "text", + "content": ". Bipartite graphs can be represented by an " + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "inline_equation", + "content": "I \\times J" + }, + { + "bbox": [ + 130, + 630, + 480, + 666 + ], + "type": "text", + "content": " adjacency matrix" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": "with each element indicating an edge between " + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": ". Consequently, " + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}" + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": " can be modeled as " + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "inline_equation", + "content": "I \\times J" + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": " Bernoulli distributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 146, + 481, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 146, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 212, + 146, + 481, + 160 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}\\right) = \\operatorname {B e r n o u l l i} \\left(\\pi_ {\\mathbf {y}, \\mathbf {z}}\\right), \\quad \\pi_ {\\mathbf {y}, \\mathbf {z}} \\in \\mathbb {R} ^ {I \\times J}. \\tag {7}", + "image_path": "8a17a77118dea69b04cf10e59a8b940e18776515aaa2894d4024269f5de21efd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 164, + 321, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 164, + 321, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 164, + 321, + 178 + ], + "type": "text", + "content": "The edge " + }, + { + "bbox": [ + 132, + 164, + 321, + 178 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y}|\\mathbf{z},\\mathbf{G}_{inter}^{\\mathbf{y};\\mathbf{z}})" + }, + { + "bbox": [ + 132, + 164, + 321, + 178 + ], + "type": "text", + "content": " is then defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 220, + 184, + 481, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 184, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 220, + 184, + 481, + 218 + ], + "type": "interline_equation", + "content": "p (\\mathbf {y} | \\mathbf {z}, \\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}) = \\prod_ {i, j} ^ {I, J} p (\\mathbf {y} _ {j} | \\mathbf {z} _ {i}) \\hat {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}} [ i, j ], \\tag {8}", + "image_path": "ad8eba65577f658866912143b89c8588d34cd557facafdce1265b4b486aab772.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 225, + 479, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 225, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 479, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 225, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\hat{G}_{inter}^{\\mathbf{y},\\mathbf{z}}" + }, + { + "bbox": [ + 131, + 225, + 479, + 274 + ], + "type": "text", + "content": " are samples from " + }, + { + "bbox": [ + 131, + 225, + 479, + 274 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{inter}^{\\mathbf{y},\\mathbf{z}})" + }, + { + "bbox": [ + 131, + 225, + 479, + 274 + ], + "type": "text", + "content": ". Nonetheless, implementing such a BayesNet can present challenges. For instance, normalization layers such as GDN [3], often integral to compression models, are difficult to integrate because the nodes are presumed to be mutually independent." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 274, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 274, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 480, + 357 + ], + "type": "text", + "content": "Heterogeneous Parameterization of " + }, + { + "bbox": [ + 131, + 274, + 480, + 357 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{inter})" + }, + { + "bbox": [ + 131, + 274, + 480, + 357 + ], + "type": "text", + "content": " We offer an innovative parameterization for the bipartite graph, tailored for heterogeneous DAG structure learning within image compression frameworks. Rather than dividing nodes, we maintain individual nodes in adjacent layers and diversify only the connecting edges. Each node pair is linked through multiple neural network edges of varying sizes and complexities, from which only one is selected. In this context, " + }, + { + "bbox": [ + 131, + 274, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}" + }, + { + "bbox": [ + 131, + 274, + 480, + 357 + ], + "type": "text", + "content": " is characterized by " + }, + { + "bbox": [ + 131, + 274, + 480, + 357 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 274, + 480, + 357 + ], + "type": "text", + "content": " categorical distributions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 363, + 481, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 363, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 212, + 363, + 481, + 377 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}\\right) = \\operatorname {C a t e g o r i c a l} \\left(\\pi_ {\\mathbf {y}, \\mathbf {z}}\\right), \\quad \\pi_ {\\mathbf {y}, \\mathbf {z}} \\in \\mathbb {R} ^ {N}. \\tag {9}", + "image_path": "72d192bb8817075503a393044290f3a6a927775a2f5f26aa73d2eb92aa50fb4b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 382, + 382, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 382, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 382, + 396 + ], + "type": "text", + "content": "Thus, the connection " + }, + { + "bbox": [ + 132, + 382, + 382, + 396 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y}|\\mathbf{z},\\mathbf{G}_{inter}^{\\mathbf{y},\\mathbf{z}})" + }, + { + "bbox": [ + 132, + 382, + 382, + 396 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 224, + 402, + 481, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 402, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 224, + 402, + 481, + 434 + ], + "type": "interline_equation", + "content": "p (\\mathbf {y} | \\mathbf {z}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {z}}) = \\sum_ {n} ^ {N} \\hat {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {z}} [ n ] p _ {n} (\\mathbf {y} | \\mathbf {z}). \\tag {10}", + "image_path": "abf61b6dc69ecaa0d097b4bc9bc2504f9550885fea5b2c11a40b5884077a6c77.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 438, + 481, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 438, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 131, + 438, + 481, + 486 + ], + "type": "text", + "content": "This heterogeneous DAG learning approach does not merely determine whether nodes are interdependent; it specifies how they are interconnected. Given the inherent dependencies among latent nodes in compression frameworks, this nuanced parameterization is particularly apt for representing inter-node BayesNets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "text", + "content": "Implementation of the edges Once " + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}" + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "text", + "content": " is established, we proceed to construct the heterogeneous edges, represented as " + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "inline_equation", + "content": "p_n(\\mathbf{y}|\\mathbf{z})" + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "text", + "content": ", utilizing neural networks. Following [29], we employ slimmable networks and select the " + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "text", + "content": "-th channel width configuration for each " + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "inline_equation", + "content": "p_n(\\mathbf{y}|\\mathbf{z})" + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "text", + "content": ". This approach is memory-efficient as it does not necessitate additional memory overhead compared to a singular network implementation for " + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y}|\\mathbf{z})" + }, + { + "bbox": [ + 131, + 487, + 481, + 569 + ], + "type": "text", + "content": ". A comprehensive example detailing this implementation is provided in Section 5 for further reference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 570, + 481, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 481, + 630 + ], + "type": "text", + "content": "Optimization of " + }, + { + "bbox": [ + 131, + 570, + 481, + 630 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{inter})" + }, + { + "bbox": [ + 131, + 570, + 481, + 630 + ], + "type": "text", + "content": " We can leverage samples from " + }, + { + "bbox": [ + 131, + 570, + 481, + 630 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{inter})" + }, + { + "bbox": [ + 131, + 570, + 481, + 630 + ], + "type": "text", + "content": " as weights to blend the various edges, enabling end-to-end optimization via the Gumbel-softmax reparameterization trick [14], eliminating the need for an additional loss function. The computational complexity associated with this optimization can be quantified using the same mixing weights:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 222, + 636, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 636, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 222, + 636, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} \\left(\\mathbf {G} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}}\\right) = \\sum_ {n} ^ {N} \\hat {\\mathbf {G}} _ {\\text {i n t e r}} ^ {\\mathbf {y}, \\mathbf {z}} [ n ] C \\left(p _ {n} (\\mathbf {y} | \\mathbf {z})\\right), \\tag {11}", + "image_path": "91ea955b9af20ca1cc35872564e9fdba68cc908ec2de2b047966e7b21297eeaf.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 157, + 479, + 255 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Statistics of different intra-node BayesNet configurations for a " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "3 \\times 32 \\times 32" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " tensor. Numbers inside the nodes represent the topological indices of the partites. For Maskconv each partite only include one node, so we omit the topological index." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 157, + 479, + 255 + ], + "lines": [ + { + "bbox": [ + 136, + 157, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 479, + 255 + ], + "type": "table", + "html": "
MaskConv [24]Zigzag [20]Checkerboard [12]Multistage [21]Channel [25]
Stages102463243
FLOPs1171888352976117182048
Time (ms)11783.0727.346.582.671.2
", + "image_path": "1beb671ff72c3787c11bdd6d2a87208eee8e890c410e5fcb35cb57d7158a1ef5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 480, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 277, + 480, + 300 + ], + "type": "inline_equation", + "content": "C(p_{m}(\\mathbf{y}|\\mathbf{z}))" + }, + { + "bbox": [ + 130, + 277, + 480, + 300 + ], + "type": "text", + "content": " represents the complexity measure for a single edge such as FLOPs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 480, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 398 + ], + "type": "text", + "content": "Relation to Neural Architecture Search (NAS) Our approach to parameterizing " + }, + { + "bbox": [ + 130, + 302, + 480, + 398 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{inter})" + }, + { + "bbox": [ + 130, + 302, + 480, + 398 + ], + "type": "text", + "content": " bears resemblance to Differentiable Architecture Search [22] in both concept and optimization techniques. Both methods utilize continuous relaxation to combine outputs from diverse neural network modules. However, the underlying objectives differ. NAS focuses on developing neural networks with enhanced performance by refining the computational graph. In contrast, our objective is to construct optimized probabilistic graphs tailored for compression, with neural networks serving as approximators of the conditional probabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 415, + 385, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 385, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 385, + 427 + ], + "type": "text", + "content": "4.2 Multipartite Graph for Intra-node BayesNet" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 591 + ], + "type": "text", + "content": "As noted in Section 2.2, traditional BayesNet structure learning has been constrained by the need to maintain a DAG, which, when applied to image data with its vast number of pixels, can greatly reduce efficiency. Table 1 illuminates an alternative approach that can satisfy both acyclic and parallel constraints effectively. It reveals that the number of parallel stages corresponds to the multipartite quantity within each BayesNet. Since edges can only direct from topologically earlier nodes to later ones, nodes within the same partite can be processed simultaneously. Bayesian inference can thus be executed by sequentially processing each partite from the first to the last. This process not only maintains acyclicity but also confines the number of parallel stages to the number of partites. Hence, we propose to parameterize " + }, + { + "bbox": [ + 130, + 434, + 482, + 591 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{intra})" + }, + { + "bbox": [ + 130, + 434, + 482, + 591 + ], + "type": "text", + "content": " with categorical distributions at each node, each category corresponding to the node's topological index within the partites. Consequently, " + }, + { + "bbox": [ + 130, + 434, + 482, + 591 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{intra})" + }, + { + "bbox": [ + 130, + 434, + 482, + 591 + ], + "type": "text", + "content": " takes the following form:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 173, + 598, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 598, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 173, + 598, + 481, + 634 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right) = \\prod_ {c, h, w} ^ {C, H, W} p \\left(\\mathrm {T} _ {c, h, w}\\right), \\quad \\mathrm {T} _ {c, h, w} = \\text {C a t e g o r i c a l} \\left(\\pi_ {c, h, w}\\right), \\tag {12}", + "image_path": "3eb65100983fa101d920982bd4392a3bef4760349632a5d15ac47c448ab9080e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{c,h,w}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " denotes the topological indices of partites per categorical distribution, with " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "C,H,W" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " being the node count along the channel, height, and width" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "dimensions, respectively. Adjusting the number of partites (or parallel stages) " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\dim \\pi_{c,h,w} = S_{intra}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " can modulate the computation time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "content": "Optimization of " + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{\\text {intra }})" + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "content": " Having constrained the number of parties to manage computational time, we now focus on learning the intra-node BayesNet structure to minimize bit-rate. The optimization challenge is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 183, + 479, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 183, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 182, + 183, + 479, + 202 + ], + "type": "interline_equation", + "content": "\\min _ {p \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right)} \\mathbb {E} _ {p \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right)} p \\left(\\mathbf {X}, \\mathbf {L} \\mid \\mathbf {G} _ {\\text {i n t r a}}\\right), \\quad \\text {s . t .} \\hat {\\mathbf {G}} _ {\\text {i n t r a}} \\text {i s a c y c l i c ,} \\tag {13}", + "image_path": "598f6108f73da2b5f6f1a898f57a6debe4a5b4d833c2672efde6d7f89f61dcd6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}_{\\text {intra }}" + }, + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "text", + "content": " is sampled from " + }, + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{\\text {intra }})" + }, + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "text", + "content": ". However, unlike the optimization for " + }, + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\text {inter }}" + }, + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "text", + "content": ", partite indices are ordinal and don't lend themselves to being used as mixing weights, making the Gumbel-softmax reparameterization trick inapplicable. We therefore approach " + }, + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\text {intra }}" + }, + { + "bbox": [ + 131, + 210, + 480, + 270 + ], + "type": "text", + "content": " as discrete variables and employ the Monte-Carlo method with the VIMCO objective [26] for optimization:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 278, + 479, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 278, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 194, + 278, + 479, + 309 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathbf {G} _ {i n t r a}} = \\mathbb {E} _ {q \\left(\\mathbf {G} _ {i n t r a} ^ {1: M} \\mid \\mathbf {X}\\right)} \\log \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\log p (\\mathbf {X}, \\mathbf {L} | \\mathbf {G} _ {i n t r a} ^ {i}), \\tag {14}", + "image_path": "dae3cd9a3cb3a74a99707990257c4c69c05d33e530cc8dabca1ea657553d10f0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 317, + 479, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 317, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 131, + 317, + 479, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 317, + 479, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{intra}^{i}" + }, + { + "bbox": [ + 131, + 317, + 479, + 330 + ], + "type": "text", + "content": " is the i-th Monte-Carlo sample and " + }, + { + "bbox": [ + 131, + 317, + 479, + 330 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 317, + 479, + 330 + ], + "type": "text", + "content": " denotes number of samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 331, + 479, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 377 + ], + "type": "text", + "content": "Moreover, note that Equation 12 presumes all partites are independent—an unlikely scenario as proximal nodes often exhibit stronger correlations. In practice, we can approximate " + }, + { + "bbox": [ + 130, + 331, + 479, + 377 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{intra})" + }, + { + "bbox": [ + 130, + 331, + 479, + 377 + ], + "type": "text", + "content": " with an unconditional generative model to consider such correlations and optimize this proxy instead, described as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 233, + 385, + 479, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 385, + 479, + 397 + ], + "spans": [ + { + "bbox": [ + 233, + 385, + 479, + 397 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {G} _ {\\text {i n t r a}}\\right) \\propto g (\\mathbf {N}), \\quad \\mathbf {N} \\sim \\mathcal {N} (0, 1), \\tag {15}", + "image_path": "bbe5519ad11ba24fb46bd817b0863831d1e8187a0293c627f43579e55649152d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "g(\\mathbf{N})" + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "text", + "content": " symbolizing a generative network derived from standard normal noise. Implementation of the Intra-node BayesNet With the model " + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "p(\\mathbf{G}_{\\text{intra}})" + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "text", + "content": " established, we can draw samples of topological indices for parties, denoted " + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "\\hat{T}_{c,h,w}" + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "text", + "content": ", from each respective " + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{c,h,w}" + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "text", + "content": ". The subsequent step involves establishing conditional dependencies among these groups. Since dependencies are restricted to nodes from different parties, a dynamic masked convolution operation is utilized. This operation ensures each node's dependencies are limited to preceding nodes within the convolutional kernel's range. For instance, the intra-node BayesNet for " + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 131, + 404, + 480, + 512 + ], + "type": "text", + "content": " can be characterized by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 184, + 519, + 479, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 519, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 184, + 519, + 479, + 552 + ], + "type": "interline_equation", + "content": "\\prod_ {c, h, w} ^ {C, H, W} p \\left(\\mathrm {y} _ {c, h, w} \\mid \\mathbf {G} _ {\\text {i n t r a}}\\right) \\propto \\operatorname {d y n c o n v} \\left(\\hat {\\mathrm {y}} _ {c, h, w}, W _ {c, h, w} \\left(\\hat {T} _ {c, h, w}\\right)\\right), \\tag {16}", + "image_path": "3b7828d90ab2e6aef8c8ab6b15aa0b8ba7788f2e76305b9d197e7a1a464e06ab.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{c,h,w}" + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "content": " are samples from " + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "inline_equation", + "content": "\\mathrm{y}_{c,h,w}" + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "inline_equation", + "content": "W_{c,h,w}" + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "content": " represents a dynamically adjusted convolutional kernel influenced by adjacent nodes' topological indices. As an illustration, consider a 2D " + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "inline_equation", + "content": "5\\times 5" + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "content": " convolution: the dynamic kernel " + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "inline_equation", + "content": "W_{c,h,w}" + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "content": " used for node " + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "inline_equation", + "content": "\\mathrm{y}_{c,h,w}" + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "content": " is defined by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 614, + 479, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 614, + 479, + 653 + ], + "spans": [ + { + "bbox": [ + 134, + 614, + 479, + 653 + ], + "type": "interline_equation", + "content": "W _ {c, h, w} [ i, j + 2, k + 2 ] = \\left\\{ \\begin{array}{l l} W _ {i, j + 2, k + 2} & , \\hat {T} _ {c, h, w} < \\hat {T} _ {i, h + j, w + k} \\\\ 0 & , \\hat {T} _ {c, h, w} \\geq \\hat {T} _ {i, h + j, w + k} \\end{array} , \\left\\{ \\begin{array}{l l} i & \\in \\{0, 1, \\dots , C - 1 \\} \\\\ j, k & \\in \\{- 2, - 1, 0, 1, 2 \\} \\end{array} . \\right. \\right. \\tag {17}", + "image_path": "5c16cbb52991ce5fe015c259e4aeefc125e4291f3fe3d7e1e01369968bec8c92.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 654, + 432, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 654, + 432, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 654, + 432, + 665 + ], + "type": "text", + "content": "For in-depth guidance of its implementation, please refer to Appendix." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 113, + 462, + 264 + ], + "blocks": [ + { + "bbox": [ + 153, + 113, + 462, + 264 + ], + "lines": [ + { + "bbox": [ + 153, + 113, + 462, + 264 + ], + "spans": [ + { + "bbox": [ + 153, + 113, + 462, + 264 + ], + "type": "image", + "image_path": "18a6da2e6d528ffb3a32f95858e869156b4741c3c3f1867ceaf43d40964a58d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "text", + "content": "Fig. 2: The architecture of the implemented example based on the joint autoregressive framework [24]. Comp/Decomp means Compression / Decompression. AE/AD symbolize the arithmetic encoder and decoder, respectively. AR denotes the autoregressive model that executes the learned intra-node BayesNet. Q stands for quantization. " + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "text", + "content": " correspond to input image and " + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "text", + "content": " the reconstructed image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 350, + 307, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 307, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 307, + 365 + ], + "type": "text", + "content": "5 From BayesNet to BaSIC" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "type": "text", + "content": "Upon determining the BayesNet structure, we proceed to specify the nodes and edges in this BayesNet, crafting a novel NIC framework, BaSIC. This section details an implementation instance of the envisioned computationally scalable NIC, leveraging the prevalent joint autoregressive and hyperprior compression model [24]. It should be noted that this represents just one of the viable strategies for implementing BaSIC, with the potential for alternatives to offer improved results. The architecture of our example is depicted in Figure 2, and we will explore how it seamlessly integrates with the aforementioned BayesNets." + }, + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 475, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 499 + ], + "type": "text", + "content": "Inter-node BayesNet The inter-node BayesNet is constructed over each edge between nodes within the original BayesNet, specifically:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 157, + 507, + 436, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 507, + 436, + 529 + ], + "spans": [ + { + "bbox": [ + 157, + 507, + 436, + 529 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q (\\mathbf {y} | \\mathbf {x}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {x}}) \\propto g _ {a} (\\mathbf {x}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {y}, \\mathbf {x}}), \\quad q (\\mathbf {z} | \\mathbf {y}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {z}, \\mathbf {y}}) \\propto h _ {a} (\\mathbf {y}, \\mathbf {G} _ {i n t e r} ^ {\\mathbf {z}, \\mathbf {y}}), \\\\ (1 1 1, \\hat {\\mathbf {x}} ^ {\\hat {\\mathbf {y}}} \\hat {\\mathbf {z}}), (1 1 1, \\hat {\\mathbf {z}} ^ {\\hat {\\mathbf {y}}} \\hat {\\mathbf {z}}). \\end{array}", + "image_path": "797e559457ecd63a97bd038b2ad46d55635fb11f97987fe4f5dbc3adba91c619.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 525, + 432, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 525, + 432, + 540 + ], + "spans": [ + { + "bbox": [ + 159, + 525, + 432, + 540 + ], + "type": "interline_equation", + "content": "p (\\hat {\\mathbf {y}} | \\hat {\\mathbf {z}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {y}}, \\hat {\\mathbf {z}}}) \\propto h _ {s} (\\hat {\\mathbf {z}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {y}}, \\hat {\\mathbf {z}}}), \\quad p (\\hat {\\mathbf {x}} | \\hat {\\mathbf {y}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {x}}, \\hat {\\mathbf {y}}}) \\propto g _ {s} (\\hat {\\mathbf {y}}, \\mathbf {G} _ {i n t e r} ^ {\\hat {\\mathbf {x}}, \\hat {\\mathbf {y}}})", + "image_path": "86251011544862cbb039b07010c304d208efea11e5a8d8ca4b09ee1dc782d16c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "text", + "content": "To achieve computational scalability, heterogeneous bipartite graphs are employed as outlined in Equation 10 for " + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}^{\\mathbf{y},\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}^{\\mathbf{z},\\mathbf{y}}" + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}^{\\hat{\\mathbf{y}},\\hat{\\mathbf{z}}}" + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}^{\\hat{\\mathbf{x}},\\hat{\\mathbf{y}}}" + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "text", + "content": ". The implementation of " + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "inline_equation", + "content": "g_{a}, g_{s}, h_{a}" + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "inline_equation", + "content": "h_{s}" + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "text", + "content": " follow [29]'s slimmable networks, with dynamic channel widths including 48, 72, 96, 144, and 192 for computational scalability. Note that we use a fixed 192 channels for all latent variables " + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{y}, \\mathbf{z}" + }, + { + "bbox": [ + 130, + 549, + 482, + 645 + ], + "type": "text", + "content": ", and only channels in the middle layers of each slimmable network are variable. This modification decouples bitrate scalability from computational scalability, enhancing the convenience in real-world applications." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 313, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 313, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 313, + 665 + ], + "type": "text", + "content": "Further details are provided in Appendix." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "Intra-node BayesNet In the established framework, the intra-node BayesNet employs masked convolution applied to " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": ". We substitute this with our suggested multipartite graph-based intra-node BayesNet " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "p(\\hat{\\mathbf{y}}|\\mathbf{G}_{intra}^{\\hat{\\mathbf{y}}})" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": ", which enables parallel computing on GPUs. The intra-node BayesNet uses a " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "5\\times 5" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " dynamic masked convolution as detailed in Equation 17, with " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " equals to number of partites and " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "H,W" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " equals to image height and width, respectively. Note that to facilitate the optimization of the intra-node BayesNet, the additional VIMCO loss expressed in Equation 14 should be incorporated into the overall loss function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 479, + 247 + ], + "type": "text", + "content": "Optimization The optimization process involves a loss function that simultaneously accounts for both the BayesNet structure (Equation 14) and the neural network parameters (Equation 4):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 230, + 260, + 480, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 260, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 230, + 260, + 480, + 271 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {R} + \\lambda_ {D} \\mathcal {L} _ {D} + \\lambda_ {C} \\mathcal {L} _ {C} + \\mathcal {L} _ {\\mathbf {G} _ {\\text {i n t r a}}} \\tag {19}", + "image_path": "0b44c6760db9814821a79bafecc5c66f4a567b737cc5148a1bdc04ea183e110e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "text", + "content": "Computational Scalability The described implementation enables computational scalability across two dimensions. Firstly, by modulating all " + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}" + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "text", + "content": ", which govern the channel width in slimmable networks, we can fine-tune the computational load of the inter-node neural networks. " + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}" + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "text", + "content": " can be modulated during training through the " + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{C}" + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "text", + "content": " parameter, or during inference by selecting from " + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}" + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "text", + "content": " to alter channel widths. Notably, the latter can be streamlined by training with a uniform " + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{inter}" + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "text", + "content": " distribution and setting " + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{C} = 0" + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "text", + "content": ". Secondly, by varying the number of partites in " + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{intra}^{\\hat{y}}" + }, + { + "bbox": [ + 130, + 282, + 480, + 413 + ], + "type": "text", + "content": ", we can adjust the stages of parallel computations in autoregressive models, thereby influencing the decompression time. In practice, we could train multiple autoregressive models with varying stages and choose among them during inference to modify decompression times." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 435, + 230, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 435, + 230, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 435, + 230, + 449 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 463, + 253, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 253, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 253, + 475 + ], + "type": "text", + "content": "6.1 Experiment Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 479, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 479, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 479, + 602 + ], + "type": "text", + "content": "Experiment Tasks Following previous works [29], our experiments focus on the task of lossy image compression with computational scalability. We benchmark our BaSIC framework against other computational scalable frameworks, as well as conduct ablation studies to validate the efficacy of the proposed intranode BayesNet learning methods. Further experimental results and details can be found in Appendix, including compression performance evaluation and comparison under fixed computational constraints and more ablation studies regarding the inter-node BayesNet. For evaluation, we employ Pixel Signal-Noise Ratio (PSNR) for distortion and Bits Per Pixel (BPP) for bit-rate, unless stated otherwise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "text", + "content": "Datasets We use a merged training dataset from ImageNet [10] and CLIC [2], preprocessed following the method in [11]. The Kodak [1] dataset serves for validation and testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Implementation Details For the proposed method, we use the joint autoregressive framework [24] as the basis, as described in Section 5. Unless otherwise" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 299, + 232 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 299, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 299, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 299, + 232 + ], + "type": "image", + "image_path": "37f1dc03aa2b1ae77866c639d05da9f06da593941c7ed1f3296224cdd07abfea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 237, + 462, + 248 + ], + "lines": [ + { + "bbox": [ + 149, + 237, + 462, + 248 + ], + "spans": [ + { + "bbox": [ + 149, + 237, + 462, + 248 + ], + "type": "text", + "content": "Fig. 3: Comparative results for different computational scalable frameworks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 115, + 467, + 229 + ], + "blocks": [ + { + "bbox": [ + 304, + 115, + 467, + 229 + ], + "lines": [ + { + "bbox": [ + 304, + 115, + 467, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 467, + 229 + ], + "type": "image", + "image_path": "08457f6224055bd3195c2cf2d1d289282646b78868c5913c862c42702b6c0fb9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "content": "specified, we follow CompressAI [5] to set the hyperparameter " + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\lambda_{D}" + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "content": ", while employing " + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\lambda_{C} = 0" + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "content": " during training and a greedy search among 16 intra-node BayesNets with varied FLOPs during inference. Training spans approximately 500k iterations, employing the Adam optimizer with a learning rate of 0.0001. To compare running times equitably, all experiments are performed on identical hardware (an Intel i7-6800K CPU, NVIDIA Titan Xp GPU, and 32GB of memory)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 366, + 447, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 366, + 447, + 379 + ], + "spans": [ + { + "bbox": [ + 131, + 366, + 447, + 379 + ], + "type": "text", + "content": "6.2 Comparison to Other Computation Scalable Frameworks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "content": "This section assesses our computational scalable BaSIC method against other scalable frameworks. We configure our framework without autoregressive modules, similar to SlimCAE [29], to compare performance at various computational levels. We adopt Hyperprior [4] as a reference, reporting BD-rate for diverse MACs for both methods. Also, the rate-distortion plots for SlimCAE's full range of complexity and three levels (L0, L7, L15) for our method are included for reference. Note that " + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "content": "-scheduling is deactivated in SlimCAE to facilitate uniform BD-rate analysis across levels. The results are depicted in Figure 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Our method, employing the same slimmable network backbone as SlimCAE, demonstrates a marginal performance edge at most computation levels, likely attributable to the fixed channels of latent variables that enhance higher bit-rate performance. Specifically, our method saves around " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " more bits than Slim-CAE under the maximum complexity, and around " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " more bits under the minimum complexity. While SlimCAE [29] employs " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "-scheduling to tackle this, it inadvertently ties bit-rate and distortion to computational complexity, which is less desirable for cases like low-powered devices needing high-quality images. In contrast, our inter-node BayesNet independently modulates the computational complexity of individual neural network modules, as evidenced by the rate-distortion plots that show minimal impact on performance with reduced computation. Moreover, the varied complexity levels indicate our method's superior control over computational complexity, allowing us to assign different inter-node BayesNets (channel widths) to distinct layers. SlimCAE, conversely, applies the same channel width across all layers, limiting flexibility. For instance," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 136, + 462, + 185 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 443, + 126 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 443, + 126 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 443, + 126 + ], + "type": "text", + "content": "Table 2: Bit-rate comparison for different autoregressive modules." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 136, + 462, + 185 + ], + "lines": [ + { + "bbox": [ + 151, + 136, + 462, + 185 + ], + "spans": [ + { + "bbox": [ + 151, + 136, + 462, + 185 + ], + "type": "table", + "html": "
Stages-2410
MethodsMaskConvBaselineCBCWOursMSOursELICOurs
BPP0.31500.33330.32160.32220.32070.32100.31830.31740.3156
Speed (MB/s)0.32345.55191.23173.46172.57107.9374.6717.8517.32
", + "image_path": "0d03adae42eec569e08e84cc415f652f6f63c3ecd1fc27570c64c391341ea509.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "type": "text", + "content": "to meet a 200kMAC/pixel threshold, our closest model uses 158kMAC/pixel, while SlimCAE's nearest option drops to 111kMAC/pixel, as shown in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 251, + 466, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 466, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 466, + 275 + ], + "type": "text", + "content": "6.3 Comparison to Other Autoregressive Models for Intra-Node BayesNet" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 481, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 481, + 429 + ], + "type": "text", + "content": "In this section, we evaluate the effectiveness of the proposed intra-node BayesNet structure learning method, by comparing it to other hand-crafted autoregressive models. For a simple and fair comparison, we first establish a baseline using a joint autoregressive framework [24] with a null context model, trained for about 500k iterations. Subsequently, we retain all other components while training various autoregressive models to replace the baseline context model for an additional 500k iterations. This approach ensures the reconstructed image remains unchanged, allowing us to isolate and evaluate the bit-rate to reflect the effectiveness of different autoregressive models. Moreover, the decompression speed is also reported to evaluate the efficiency of It's important to note that the neural network architecture for the context model is kept constant across all methods, with the sole variability stemming from the structure of the intra-node BayesNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": "The benchmark autoregressive models include Channel-wise (CW, 2-stage) [25], Checkerboard [12] (CB, 2-stage), Multi-stage spatial " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": " [21] (MS, 4-stage) and the combined unevenly grouped channel-wise and checkerboard [11] (ELIC, 10-stage in total). We also train the MaskConv context model from [24] as a baseline. Our model is trained under constraints of 2, 4, and 10 parallel stages and evaluated against counterparts with equivalent stage configurations. The findings are presented in Table 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 514, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 597 + ], + "type": "text", + "content": "Overall, our learning-based BayesNet structures consistently demonstrate slightly better performance compared to the manually-crafted models under the same parallel stages, with similar decompression speed. Interestingly, in each group our learned method reaches close BPP to the hand-crafted ones in the next level of parallel stage, and in the 10-stage case, our method even approaches the BPP of non-parallel MaskConv model. This proves that our proposed method excels in learning effective intra-node BayesNets for autoregressive models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 597, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 645 + ], + "type": "text", + "content": "Qualitative Results We also illustrate the learned intra-node BayesNet by our method as well as the manually-designed one of ELIC in Figure 4 to analyze their difference with hand-crafted models. We could see that the BayesNet of ELIC uses a checkerboard context model over the spatial dimension, including only two" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 347, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 347, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 347, + 665 + ], + "type": "text", + "content": "5 More visualization results are shown in Appendix." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 465, + 188 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 465, + 188 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 465, + 188 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 465, + 188 + ], + "type": "image", + "image_path": "c1faa6ec35b794b24ae4d04e1afa693041b07d6f2bde0e22603d802af43ef0fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 198, + 465, + 272 + ], + "blocks": [ + { + "bbox": [ + 290, + 189, + 324, + 198 + ], + "lines": [ + { + "bbox": [ + 290, + 189, + 324, + 198 + ], + "spans": [ + { + "bbox": [ + 290, + 189, + 324, + 198 + ], + "type": "text", + "content": "(a) ELIC" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 198, + 465, + 272 + ], + "lines": [ + { + "bbox": [ + 149, + 198, + 465, + 272 + ], + "spans": [ + { + "bbox": [ + 149, + 198, + 465, + 272 + ], + "type": "image", + "image_path": "a8b4921d5297f5a418bfcf70c08d0fc9cef573a0a4f74633ad210f16d3b6b3b5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 274, + 338, + 283 + ], + "lines": [ + { + "bbox": [ + 277, + 274, + 338, + 283 + ], + "spans": [ + { + "bbox": [ + 277, + 274, + 338, + 283 + ], + "type": "text", + "content": "(b) Ours-10stage" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 291, + 482, + 348 + ], + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 348 + ], + "type": "text", + "content": "Fig. 4: Partite topological indices visualization for intra-node BayesNet in the 10-stage case. Each image patch indicates topological indices of nodes in a group of channels. Blacker pixels indicate lower topological indices and whiter indicate higher topological indices. E.g., in the 10-stage case, pure-black indicates index 0 which is first iterated, and pure-white indicates group 9 which is last iterated. Better viewed with zoom-in." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 371, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 456 + ], + "type": "text", + "content": "partites within each channel. In comparison, our learned BayesNet incorporates more different partites (mostly three or four) in the spatial dimension, which captures more context information than the checkerboard case. This is reasonable as images are usually heavily correlated in spatial dimensions. Moreover, the learned topological indices also give some insight for designing parallel autoregressive models, that channel and spatial context models could be combined more compactly to improve their performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 473, + 322, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 322, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 322, + 487 + ], + "type": "text", + "content": "7 Conclusions and Discussions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": "This paper presents a novel approach, BaSIC, to addressing computational scalability in Neural Image Compression (NIC) by focusing on BayesNet structure learning. We decompose the challenge into two subproblems: heterogeneous bipartite graph learning for inter-node BayesNet, and multipartite graph learning for intra-node BayesNet, proposing solutions for each. Experimental results, using a joint autoregressive framework as a test case, show that our work offers improved control over computational demands relative to other scalable computation models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 163 + ], + "type": "text", + "content": "The paper is supported in part by the National Natural Science Foundation of China (No. 62325109, U21B2013)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 205, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 228 + ], + "type": "text", + "content": "1. Kodak lossless true color image suite. https://r0k.us/graphics/kodak/ (2023), accessed: 2023-11-12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "text", + "content": "2. Workshop and challenge on learned image compression. https://click.compression.cc/ (2023), accessed: 2023-01-09" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 250, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 282 + ], + "type": "text", + "content": "3. Balle, J., Laparra, V., Simoncelli, E.P.: Density modeling of images using a generalized normalization transformation. CoRR abs/1511.06281 (2015), https://apisemantic scholar.org/CorpusID:2684987" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "text", + "content": "4. Balle, J., Minnen, D.C., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. ArXiv abs/1802.01436 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "text", + "content": "5. Bégaint, J., Racapé, F., Feltman, S., Pushparaja, A.: Compressai: a pytorch library and evaluation platform for end-to-end compression research. arXiv preprint arXiv:2011.03029 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 370 + ], + "type": "text", + "content": "6. Cheng, J., Greiner, R., Kelly, J., Bell, D.A., Liu, W.: Learning bayesian networks from data: An information-theory based approach. Artif. Intell. 137, 43-90 (2002), https://api-semanticscholar.org/CorpusID:7821347" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 371, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 480, + 413 + ], + "type": "text", + "content": "7. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 7936-7945 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "text", + "content": "8. Chickering, D.M.: Optimal structure identification with greedy search. J. Mach. Learn. Res. 3, 507-554 (2003), https://api-semanticscholar.org/CorpusID:1191614" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 446, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 488 + ], + "type": "text", + "content": "9. Deleu, T., G'ois, A., Emezue, C.C., Rankawat, M., Lacoste-Julien, S., Bauer, S., Bengio, Y.: Bayesian structure learning with generative flow networks. ArXiv abs/2202.13903 (2022), https://api(semanticscholar.org/CorpusID: 247158659" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 135, + 490, + 480, + 522 + ], + "type": "text", + "content": "10. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition pp. 248-255 (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 523, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 523, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 523, + 480, + 576 + ], + "type": "text", + "content": "11. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5708-5717 (2022), https://apisemantic scholar.org/CorpusID:247594672" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 577, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 480, + 620 + ], + "type": "text", + "content": "12. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14766-14775 (2021), https://apisemantic scholar.org/CorpusID:232404386" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "text", + "content": "13. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:326772" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "14. Jang, E., Gu, S.S., Poole, B.: Categorical reparameterization with gumbel-softmax. ArXiv abs/1611.01144 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "15. Johnston, N., Eban, E., Gordon, A., Ball'e, J.: Computationally efficient neural image compression. ArXiv abs/1912.08771 (2019), https://apisemantic scholar.org/CorpusID:209404937" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 215 + ], + "type": "text", + "content": "16. Kang, N., Qiu, S., Zhang, S., Li, Z., Xia, S.: Pilc: Practical image lossless compression with an end-to-endgpu oriented neural framework. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3729-3738 (2022), https://apisemantic scholar.org/CorpusID:249625569" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "text", + "content": "17. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. CoRR abs/1312.6114 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "text", + "content": "18. Koivisto, M., Sood, K.: Exact bayesian structure discovery in bayesian networks. J. Mach. Learn. Res. 5, 549-573 (2004), https://api(semanticscholar.org/ CorpusID:12412451" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "19. Lee, J., Cho, S., Beack, S.: Context-adaptive entropy model for end-to-end optimized image compression. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 346 + ], + "type": "text", + "content": "20. Li, M., Ma, K., You, J.J., Zhang, D., Zuo, W.: Efficient and effective context-based convolutional entropy modeling for image compression. IEEE Transactions on Image Processing 29, 5900-5911 (2019), https://api-semanticscholar.org/CorpusID:195345342" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 391 + ], + "type": "text", + "content": "21. Lin, F., Sun, H., Liu, J., Katto, J.: Multistage spatial context models for learned image compression. ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 1-5 (2023), https://apisemantic scholar.org/CorpusID:257038187" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "22. Liu, H., Simonyan, K., Yang, Y.: Darts: Differentiable architecture search. ArXiv abs/1806.09055 (2018), https://api-semanticscholar.org/CorpusID: 49411844" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "type": "text", + "content": "23. Lorch, L., Rothfuss, J., Scholkopf, B., Krause, A.: Dibs: Differentiable bayesian structure learning. ArXiv abs/2105.11839 (2021), https://apisemantic scholar.org/CorpusID:235187432" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "type": "text", + "content": "24. Minnen, D.C., Balle, J., Toderici, G.: Joint autoregressive and hierarchical priors for learned image compression. ArXiv abs/1809.02736 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 521 + ], + "type": "text", + "content": "25. Minnen, D.C., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. 2020 IEEE International Conference on Image Processing (ICIP) pp. 3339-3343 (2020), https://api-semanticscholar.org/CorpusID:220633468" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "26. Mnih, A., Rezende, D.J.: Variational inference for monte carlo objectives. In: International Conference on Machine Learning (2016), https://api-semanticscholar.org/CorpusID:5859948" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "text", + "content": "27. Tao, L., Gao, W., Li, G., Zhang, C.: Adanic: Towards practical neural image compression via dynamic transform routing. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 16833-16842 (2023), https://api.sementicscholar.org/CorpusID:267026318" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "text", + "content": "28. webmproject: libwebp. https://github.com/webmproject/libwebp (2023), accessed: 2023-01-09" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "text", + "content": "29. Yang, F., Herranz, L., Cheng, Y., Mozerov, M.G.: Slimmable compressive autoencoders for practical neural image compression. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 4996-5005 (2021), https://api-semanticscholar.org/CorpusID:232417114" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 468 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "30. Yin, S., Li, C., Bao, Y., Liang, Y.: Universal efficient variable-rate neural image compression. ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 2025-2029 (2021), https://api.sementicscholar.org/CorpusID:244478380" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 204 + ], + "type": "text", + "content": "31. Yin, S., Meng, F., Tan, W.T., Li, C., Bao, Y., Liang, Y., Liu, W.: Exploring structural sparsity in neural image compression. 2022 IEEE International Conference on Image Processing (ICIP) pp. 471-475 (2022), https://api-semanticscholar.org/CorpusID:246680031" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 236 + ], + "type": "text", + "content": "32. Yu, J., Yang, L., Xu, N., Yang, J., Huang, T.S.: Slimmable neural networks. ArXiv abs/1812.08928 (2018), https://api(semanticscholar.org/CorpusID:56657799" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "33. Yu, Y., Chen, J., Gao, T., Yu, M.: Dag-gnn: Dag structure learning with graph neural networks. In: International Conference on Machine Learning (2019), https://apisemantic scholar.org/CorpusID:128358697" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 271, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 303 + ], + "type": "text", + "content": "34. Yuan, C., Malone, B.M.: Learning optimal bayesian networks: A shortest path perspective. J. Artif. Intell. Res. 48, 23-65 (2013), https://api-semanticscholar.org/CorpusID:5951724" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 346 + ], + "type": "text", + "content": "35. Zhang, Y., Lu, G., Feng, D., Zhu, C., Song, L.: Content adaptive checkerboard context model for learned image compression. 2023 IEEE International Symposium on Circuits and Systems (ISCAS) pp. 1-5 (2023), https://api-semanticscholar.org/CorpusID:260002843" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 347, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 390 + ], + "type": "text", + "content": "36. Zhang, Z., Chen, B., Lin, H., Lin, J., Wang, X., Zhao, T.: Elfic: A learning-based flexible image codec with rate-distortion-complexity optimization. Proceedings of the 31st ACM International Conference on Multimedia (2023), https://apisemantic scholar.org/CorpusID:264492165" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "text", + "content": "37. Zheng, X., Aragam, B., Ravikumar, P., Xing, E.P.: Dags with no tears: Continuous optimization for structure learning. In: Neural Information Processing Systems (2018), https://api_semanticscholar.org/CorpusID:53217974" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 424, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 468 + ], + "type": "text", + "content": "38. Zheng, Z., Wang, X., Lin, X., Lv, S.: Get the best of the three worlds: Real-time neural image compression in a non-gpu environment. Proceedings of the 29th ACM International Conference on Multimedia (2021), https://api-semanticscholar.org/CorpusID:239011955" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_content_list.json b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3fe719342aa4ea928057586495fa689405b8155a --- /dev/null +++ b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_content_list.json @@ -0,0 +1,2015 @@ +[ + { + "type": "text", + "text": "Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anqi Zhang and Guangyu Gao\\*", + "bbox": [ + 374, + 234, + 630, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science and Technology,", + "bbox": [ + 344, + 262, + 656, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beijing Institute of Technology, Beijing, China", + "bbox": [ + 344, + 277, + 656, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "andy_zaq@outlook.com", + "bbox": [ + 421, + 292, + 581, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guangyugao@bit.edu.cn", + "bbox": [ + 418, + 305, + 583, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Class Incremental Semantic Segmentation (CISS), within Incremental Learning for semantic segmentation, targets segmenting new categories while reducing the catastrophic forgetting on the old categories. Besides, background shifting, where the background category changes constantly in each step, is a special challenge for CISS. Current methods with a shared background classifier struggle to keep up with these changes, leading to decreased stability in background predictions and reduced accuracy of segmentation. For this special challenge, we designed a novel background adaptation mechanism, which explicitly models the background residual rather than the background itself in each step, and aggregates these residuals to represent the evolving background. Therefore, the background adaptation mechanism ensures the stability of previous background classifiers, while enabling the model to concentrate on the easy-learned residuals from the additional channel, which enhances background discernment for better prediction of novel categories. To precisely optimize the background adaptation mechanism, we propose Pseudo Background Binary Cross-Entropy loss and Background Adaptation losses, which amplify the adaptation effect. Group Knowledge Distillation and Background Feature Distillation strategies are designed to prevent forgetting old categories. Our approach, evaluated across various incremental scenarios on Pascal VOC 2012 and ADE20K datasets, outperforms prior exemplar-free state-of-the-art methods with mIoU of $3.0\\%$ in VOC 10-1 and $2.0\\%$ in ADE 100-5, notably enhancing the accuracy of new classes while mitigating catastrophic forgetting. Code is available in https://andyzaq.github.io/barmsite/.", + "bbox": [ + 259, + 354, + 754, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 724, + 375, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of deep learning and the support of convolutional neural networks (CNNs) [27,30,31,51] and Transformers [20,33,53,54], current semantic segmentation networks [10-13,15,16,24,34,46] have achieved remarkable results on standard supervised learning tasks. However, real-world data, arriving as an", + "bbox": [ + 212, + 756, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author.", + "bbox": [ + 217, + 824, + 388, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "unstable continuous stream, poses the challenge of retaining prior knowledge while accommodating new concepts, termed as catastrophic forgetting [38, 39].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this problem, incremental learning is proposed to adapt to changing data streams for new concepts, but also avoid forgetting old knowledge, especially for the classification task, i.e., Class-Incremental Learning (CIL) [14,28,32,59,65]. Cermelli et al. [6] introduced CIL into Semantic Segmentation as Class Incremental Semantic Segmentation (CISS). Since the semantic segmentation tasks involve dense predictions, the issue of catastrophic forgetting typically becomes more challenging. Most recent works [6,8,21] have struggled to mitigate catastrophic forgetting through knowledge distillation [3,21,40] or parameter freezing [8,61]. Although these strategies effectively alleviate catastrophic forgetting, it comes at the cost of the model's plasticity, making it challenging to learn novel classes. In current state-of-the-art approaches [3,61], the performance of established old classes approaches the upper limit of joint training, yet there still remains a significant gap in mIoU for novel classes. However, in the context of lifelong learning scenarios featuring an infinite influx of novel classes, the incremental learner's plasticity becomes increasingly crucial. Therefore, the optimal solution should be fine-tuning more parameters for novel classes, while catastrophic forgetting needs to be handled properly.", + "bbox": [ + 212, + 176, + 787, + 434 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f1a97bce118baed3dfb84f220a1bb8d7b7c02e7f289ad416a81c6f26cfa2cd09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 460, + 315, + 513 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/52291a177dbe141e77873c4a765f55fb624edf986307970b19b2913f3ce63fe4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 516, + 315, + 569 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/780f1a4c66c6324deaca6da641e912d66cb9013e421fd1f5712ab0723899e31b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 570, + 313, + 623 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1bfa3e4687e4bb266c521ac63d31e170a4597d701dfdfff6729bbe3e78f01c46.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 246, + 625, + 313, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/54abe5470d97ff2fcb271e85f9335835839a86f12f9d2710bd17e12f9f9c6d63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 463, + 387, + 515 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9c923c8c6b31da1df51e481aa2dec6f2c7838bd6f635ddb8b215848e43d99bf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 516, + 387, + 569 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ba8ea4ef1691fa9293e1aed119de98eb4790862f0c633259f4e9c1966235bc92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 570, + 387, + 625 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50a424ac6880efea796179f10a94248efcea1528be35847f48da90f067f77de9.jpg", + "image_caption": [ + "Prediction of step $t - 1$" + ], + "image_footnote": [], + "bbox": [ + 320, + 625, + 387, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/925175c29e97205e711d865ef9b4745345856c901a63aa2aa40dd768635df746.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 463, + 457, + 515 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0cadcbab6f0839911eafc8a5f265cb9a1c521be89736643d9befd226e4742d09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 516, + 457, + 569 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d449847987a191e2b729857e4a7096aa660651c52b115b1e250d18b4eda0f75b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 570, + 457, + 625 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9a5e34868501b1c68574ae3f3fec5ab60266504d55d154b3e6d02cf9dc9a7f21.jpg", + "image_caption": [ + "Bg logits of step $t - 1$" + ], + "image_footnote": [], + "bbox": [ + 395, + 625, + 457, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/49d8fe702ce2bf0ac94e5d316896ee6ac13d5fd8b749305a88d4e131f64eaa53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 463, + 529, + 515 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8848eb1b153c2e2cc16ff4a40576b1edb9e00940ab58a7c6023ba537cb07f995.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 516, + 529, + 569 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/908428f9671c0ded36545d3fc3bb4a8d533438f5fa1cb0cac83a6f5123ab7db3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 570, + 529, + 625 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/73776b77fed59101a8a1357dceb0e4e4803f080cb003a4e6d493409d3bb6171d.jpg", + "image_caption": [ + "Bg Adaptation of step $t$", + "Fig. 1: 3D Visualization of Background Adaptation results. The background logits for step $t$ combine those from step $t - 1$ with the adaptation logits learned in step $t$ , preventing disorderly adjustments and focusing on residuals. The current step background logits are processed using the Sigmoid function. Note that colors range from red (small values) to blue (large values)." + ], + "image_footnote": [], + "bbox": [ + 465, + 625, + 529, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7ba554256d6202cf91461801b388df15e9e13a712691e63778f88a8dbd910a43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 463, + 599, + 515 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dfad77d6c94cd1ea015d31579029aee42b99cc1299eeef0f0fec7d2bd9f6a384.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 516, + 599, + 569 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/99c7d18dca28bc05b9f75bfd01b480310c277fe64fd9b6e7e5cf09aee87e9bbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 570, + 601, + 625 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5d25eb747cc238b802967016b7a200872de0fd744b5e29de0b37f7f21225f55d.jpg", + "image_caption": [ + "Bg logits of step $t$ (Sigmoid)" + ], + "image_footnote": [], + "bbox": [ + 537, + 625, + 601, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0213258c00ff83766b4feb3341ab120259e859d57b62324e427bf191ef1dc28c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 460, + 676, + 515 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67829c7c655502e26655348e53d001d0aabe2efe0dbeb63e59a0e7fe5b63afec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 516, + 676, + 569 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2b02acc1185ac69de183ed6526314247802a9a173b1313e8698099bb7a60dcb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 570, + 676, + 625 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/464261f58dfc47ccb59e8eddb657715410f9c07ee1e878d0d037d52691a34286.jpg", + "image_caption": [ + "Prediction of step t" + ], + "image_footnote": [], + "bbox": [ + 607, + 625, + 676, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3abc89263b7272830f9a538f6e653f4a0cad8274b0f00d22a09748ac9cf6ee65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 460, + 753, + 515 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d74ed47d3da7515ea88233b0488d9379d4cd5bbe9e6d76a73f25863e81913f67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 516, + 753, + 569 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/819b40936f42cf055bc5906169a1c5ab18bc3098957d3ce7f2a9e62974fea72b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 570, + 753, + 625 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a1ed7d4703a540e65796bf292c82d4de1fc4a292d7cac88f8baf23334b8be20d.jpg", + "image_caption": [ + "Categories in step t" + ], + "image_footnote": [], + "bbox": [ + 681, + 625, + 753, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, background shifting, where the background category changes constantly in each step, is another challenge for CISS. Upon further investiga", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion, it was observed that previous methods within CISS often utilize a shared background classifier across each step. However, novel classes are typically considered as the background in previous learning steps, indicating that the background class undergoes continuous changes throughout incremental learning. This way results in an incessantly shifting background in each step, leading to disruption in background prediction and inaccurate learning of novel classes. Some approaches [8,9,41] store old exemplars in extra memory space to prevent background shifting, while long-term storage of such data may not be feasible in real-world scenarios. To enhance the incremental nature of the CISS method and reduce background instability, we revise the representation of the background class and propose an exemplar-free background adaptation mechanism. In particular, the background adaptation mechanism explicitly models the background residual rather than the background itself in each incremental step. The learned residuals of the additional channel are aggregated with the previously learned background logits to represent the evolving background, as illustrated in Fig. 1. By employing the background adaptation mechanism, we only need to optimize the additional channel of residuals for constructing the background during the incremental steps. This approach eliminates the need for repeatedly learning background classifiers associated with old classes, thereby avoiding the instability of optimizing a shared background classifier.", + "bbox": [ + 217, + 146, + 785, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, we customize our training approaches in order to fit the training process of the background adaptation structure and promote its performance. We first introduce the widely used pseudo-labeling [8,21] to generate a pseudobackground, and then we design a PB-BCE loss to focus on optimizing the classifier of novel classes and related background adaptation residual. Besides, for the Background Adaptation channel, the objective is to minimize the values in the regions of novel categories, while ensuring that the values in other regions remain above 0. Thus, two auxiliary Background Adaptation losses are specifically applied to the background adaptation channel for precise adjustment. Given our pursuit of plasticity on novel categories without external memory, we have devised more comprehensive exemplar-free knowledge distillation strategies to combat catastrophic forgetting. We further adopt knowledge distillation strategies on the logit maps of old categories and the intermediate features, namely the Group Knowledge Distillation and Background Feature Distillation, respectively. Finally, our approach outperforms previous methods in multiple benchmarks, especially in realistic and challenging long-term scenarios such as 10-1 of VOC2012 and 100-5 of ADE20K. Our approach achieves significant performance gains of $4.5\\%$ and $5.2\\%$ on novel classes, respectively, compared to the previous state-of-the-art.", + "bbox": [ + 217, + 452, + 785, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, our contributions are summarized as follows:", + "bbox": [ + 240, + 742, + 624, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose a Background Adaptation strategy in incremental steps by optimizing an additional channel for residual modeling instead of the shared background classifier, which avoids unstable optimization and prevents chaos in background logits.", + "bbox": [ + 225, + 779, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Background Adaptation with Residual Modeling", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We design PB-BCE loss and Background Adaptation losses to enhance the effectiveness of Background Adaptation, while two distillation strategies ensure maintaining knowledge of old categories.", + "bbox": [ + 223, + 146, + 782, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Experiments validate the effectiveness of our proposed method, showcasing a new state-of-the-art performance on CISS benchmarks.", + "bbox": [ + 225, + 191, + 784, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 239, + 397, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Class Incremental Learning", + "text_level": 1, + "bbox": [ + 215, + 267, + 488, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Class Incremental Learning (CIL) breaks the constraints of the standard training process in an incremental form on new classes. CIL task aims to continually learn the new classes without retraining on the data of old classes, which significantly reduces the training cost. As a compromise, the CIL task has to effectively solve the problem of catastrophic forgetting [38, 39]. Current CIL methods are divided into several main categories. Replay-based methods recall the data of old classes during incremental training. Some of the approaches [2, 4, 35] reserve memory buffers and temporarily store representative past data. Others [42, 49, 55] introduce powerful generative models to generate pseudo-data for rehearsal. Regularization-based methods introduce regularization terms to prevent forgetting the old classes. These methods utilize loss functions as penalties [1, 29, 45, 47] or knowledge distillation from the previous model [19, 32, 44]. Representation-based methods attempt to transfer representations to certain categories, which apply adaptation on a fixed pre-trained backbone [5, 22, 50, 60] or self-supervised learning [7, 36, 43].", + "bbox": [ + 212, + 287, + 787, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Class Incremental Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 532, + 607, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The achievements in CIL tasks encourage the Incremental Learning pipeline to develop in other dense prediction tasks, including Semantic Segmentation. Class Incremental Semantic Segmentation (CISS) task is first introduced in MiB [6], which focused on reconstructing the background region under the restriction from knowledge distillation. Since then, various methods [57, 63] have been introduced in the CISS task. PLOP [21] defined the pseudo-labeling and multi-scale local distillation to remember the old categories. SDR [40] introduced category-wise prototype matching and contrastive learning methods to enlarge the difference between categories. RCIL [58] and EWF [56] explored methods for integrating parameters of new and old models. DKD [3] decomposed the class logits into positive and negative logits for better knowledge distillation and applied threshold to generate background regions. Incrementer [48] borrows prompt learning from ViT [20] and applies the structure of Segmenter [52] instead of DeeplabV3 [12], which is the basic structure of other methods. Besides, several approaches utilize external memory [66] or models [37, 62] for better performance. SSUL [8] introduced a Saliency-map detector to identify the saliency region and store some of the previous instances in memory. MicroSeg [61] utilized proposals from pretrained Mask2Former [15] as a guidance. ALIFE [41] memorized category features and adapted for remembering the old classes.", + "bbox": [ + 212, + 551, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/779f2a42ff61ffb57b6df982cc03cef055be9ba6e9995a83b67d19385e26245b.jpg", + "image_caption": [ + "(a) The overall pipeline of our approach." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 656, + 316 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c2e6f0a7d7c85d9a9fcce5af4d17978e190e2324486eaa9048eff006192d01bc.jpg", + "image_caption": [ + "(b) $\\mathrm{BgA}$ losses.", + "Fig. 2: Overview of our framework that consists of a Background Adaptation mechanism and proposed losses. (a) The Background Adaptation mechanism uses the negative part of the adaptation channel to rectify the background logits. (b) Two Background Adaptation losses separately optimize the regions $w/$ and $w/o$ novel categories to enhance the performance of the Background Adaptation." + ], + "image_footnote": [], + "bbox": [ + 676, + 195, + 779, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 441, + 338, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 474, + 372, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Class-Incremental Semantic Segmentation (CISS) task performs training in $T$ steps. During step $t$ , the training data $D^{t}$ containing the labels of the novel classes $C^{t}$ are applied for training, while the labels of old classes $C^{1:t-1}$ are expunged. In particular, each sample of the training data $D^{t}$ consists of an RGB image $x^{t} \\in \\mathbb{R}^{3 \\times H \\times W}$ and a corresponding ground truth label $y^{t}$ on novel classes $C^{t}$ . Besides, the ground truth label $y^{t}$ replaces the labels of the $C^{1:t-1}$ with the background label $c_{b}$ as the old classes will not appear in $C^{t}$ to ensure $C^{1:t-1} \\cap C^{t} = \\emptyset$ . Therefore, the model $f^{t}$ being trained at step $t$ not only needs to learn to segment novel categories but also has to retain the capability to segment old categories. However, while learning the novel categories, the model $f^{t}$ tends to overfit the novel categories $C^{t}$ as $f^{t}$ cannot contact with samples of $C^{1:t-1}$ , causing severe catastrophic forgetting. Moreover, the novel categories can only appear in the background region during the previous steps. We are more likely to classify them from the previous background prediction region, which motivates us to focus on background adaptation.", + "bbox": [ + 212, + 501, + 787, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Background Adaptation", + "text_level": 1, + "bbox": [ + 215, + 752, + 459, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous studies often apply only one background classifier. They have to readjust the parameters of the background classifier whenever the model learns novel categories, which causes catastrophic forgotten even if the classifier of the old categories remains the same probability output. The unstable background classifier", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Background Adaptation with Residual Modeling", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "could be a thorny constraint factor. To tackle the problems above, we design a Background Adaptation mechanism. Different from the previous distinct background classifier, we introduce an additional background adaptation channel for each classifier of the incremental process.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the training process of incremental step $t$ , the current model $f^t$ inherits the parameters from the previous model $f^{t-1}$ and adds a new classifier $f_{C^t}^t$ for parallel classification of categories in $C^t$ alongside the previous classifiers. The encoder $f_e^t$ and the decoder $f_d^t$ transform the input image $x^t$ into features map $F_d^t \\in \\mathbb{R}^{h \\times w \\times c}$ :", + "bbox": [ + 212, + 207, + 784, + 284 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {d} ^ {t} = f _ {d} ^ {t} \\left(f _ {e} ^ {t} \\left(x ^ {t}\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 286, + 784, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to the standard classifier of the previous methods [8, 21, 41], our classifier of each incremental step has an additional channel for background adaptation. Therefore, for each group of categories $C^i$ from step $i$ , the classifier $f_{C^i}^t$ of the model $f^t$ generates the corresponding classification logit maps $\\mu_{C^i \\cup \\{b_i\\}}^t \\in \\mathbb{R}^{h \\times w \\times (|C^i| + 1)}$ :", + "bbox": [ + 212, + 311, + 782, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {C ^ {i} \\cup \\{b _ {i} \\}} ^ {t} = f _ {C ^ {i}} ^ {t} \\left(F _ {d} ^ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 407, + 784, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $i \\in \\{1, 2, \\ldots, t\\}$ . We define $b^1$ to represent the background class of the initial step, and $b^i$ denotes the background adaptation for subsequent incremental steps when $i >= 2$ . The symbol $b$ refers to the adapted background class. Thus, $\\mu_{C^i \\cup \\{b_i\\}}^t$ contains the probability maps $\\mu_{C^i}$ of $|C^i|$ categories from step $i$ and an additional channel $\\mu_{b^i}^t$ for constructing the background logits $\\mu_b^t$ . Considering that the region of the novel categories $C^t$ is more likely inherited from the background region of the previous step, the classifier $f_{C^t}^t$ adjusts the background logits through $\\mu_{b^t}^t$ , which decreases the probability values in the region $\\mathcal{R}_{C^t}$ of $C^t$ . The parameters of previous classifiers are not required to continuously optimize for new background regions, thereby avoiding instability in the classification of old categories.", + "bbox": [ + 212, + 439, + 784, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The background adaptation channel $\\mu_{b^t}^t$ can correct the background logits $\\mu_b^t\\in \\mathbb{R}^{h\\times w\\times 1}$ in $\\mathcal{R}_{C^t}$ , yet the noise in other regions of $\\mu_{b^t}^t$ may interfere with the background prediction. Particularly, when old categories are present in the image, there is a risk that these regions could be incorrectly covered by the adapted background. Hence, we design a Filter operation that replaces positive values of $\\mu_{b^i}^t$ with 0. As shown in Fig. 2a, with the Filter operation applied, the residual modeling of background logits during the inference process is as follows:", + "bbox": [ + 212, + 609, + 784, + 715 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {b} ^ {t} = \\mu_ {b ^ {1}} ^ {t} + \\sum_ {i = 2} ^ {t} F i l t e r \\left(\\mu_ {b ^ {i}} ^ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 729, + 784, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the background logits from the initial step $\\mu_{b^1}^t$ stay the same and the background adaptation logits from the incremental steps $\\mu_{b^i}^t$ only remain the negative values via the Filter. In particular, to make $\\mu_{b}^{t}$ more friendly for gradient backpropagation, we abandoned the Filter on the background adaptation logits", + "bbox": [ + 212, + 779, + 782, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of novel classes $\\mu_{b^t}^t$ during the training process:", + "bbox": [ + 214, + 146, + 555, + 162 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {b} ^ {t} = \\mu_ {b ^ {1}} ^ {t} + \\sum_ {i = 2} ^ {t - 1} \\operatorname {F i l t e r} \\left(\\mu_ {b ^ {i}} ^ {t}\\right) + \\mu_ {b ^ {t}} ^ {t}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 175, + 785, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Incremental Training Strategy", + "text_level": 1, + "bbox": [ + 215, + 239, + 511, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pseudo Labeling. During the incremental steps, even if objects of old categories $C^{1:t-1}$ exist in the image, we cannot obtain corresponding labels from the ground truth. In order to simulate a ground truth label that contains the categories we have learned, we follow [8,21,61] to introduce the prediction probability of the previous step $\\varphi^{t-1} \\in \\mathbb{R}^{h \\times w \\times |C^{1:t-1}|}$ from $f^{t-1}(x^t)$ with Sigmoid function and combine it with the current ground truth label $y^t$ :", + "bbox": [ + 214, + 267, + 782, + 361 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} ^ {t} = \\left\\{ \\begin{array}{l l} y ^ {t}, & \\left(y ^ {t} \\in C ^ {t}\\right) \\vee \\left\\{\\left(y ^ {t} \\in c _ {b}\\right) \\wedge \\left(\\max \\left(\\varphi^ {t - 1}\\right) < \\tau\\right) \\right\\} \\\\ \\tilde {y} ^ {t - 1}, & \\left(y ^ {t} \\in c _ {b}\\right) \\wedge \\left(\\max \\left(\\varphi^ {t - 1}\\right) \\geq \\tau\\right) \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 372, + 785, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hat{y}^t\\in \\mathbb{R}^{h\\times w}$ denotes the generated pseudo label and $\\tilde{y}^{t - 1} = \\arg \\max (\\varphi^{t - 1})$ represents the final prediction from $f^{t - 1}(x^t)$ .", + "bbox": [ + 214, + 425, + 782, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PB-BCE Loss. Binary Cross-Entropy (BCE) loss is extensively employed for learning to segment objects of novel categories. We follow the usage of the Sigmoid function and activate $\\mu^t$ as $\\varphi^t$ . Some of the CISS methods [8,21,61] utilize the pseudo mask as the ground truth for supervision, which effectively replaces the background label in the regions of the old categories. However, the ambiguous supervision of the old categories could improperly optimize the classifiers of $C^{1:t - 1}$ and destroy their identification capability. Therefore, we design a Pseudo Background BCE (PB-BCE) loss which cuts off the backpropagation of old categories, meaning that we only take the labels of current categories $\\hat{y}_{C^t}^t$ and pseudo-background $\\hat{y}_{c_b}^t$ for optimization:", + "bbox": [ + 214, + 481, + 787, + 635 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P B - B C E} = - \\frac {1}{h w} \\sum_ {k = 1} ^ {h w} \\sum_ {c \\in \\{c _ {b} \\} \\cap C ^ {t}} C E \\left(\\varphi_ {c, k} ^ {t}, \\hat {\\psi} _ {c, k} ^ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 647, + 785, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $k$ represents the index of pixel and $\\hat{\\psi}_c^t$ denotes the binary map of class c in pseudo map $\\hat{y}^t$ . $CE$ is the cross-entropy calculation:", + "bbox": [ + 214, + 705, + 782, + 737 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nC E \\left(\\varphi_ {*} ^ {t}, \\hat {\\psi} _ {*} ^ {t}\\right) = \\hat {\\psi} _ {*} ^ {t} \\cdot \\log \\varphi_ {*} ^ {t} + \\left(1 - \\hat {\\psi} _ {*} ^ {t}\\right) \\cdot \\log \\left(1 - \\varphi_ {*} ^ {t}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 747, + 785, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although the whole background prediction logits $\\mu_b^t$ are involved in the loss calculation, only the background adaptation channel for current categories $\\mu_{b^t}^t$ serves for optimization. Previous background logits are detached to prevent disorderly evolving on parameters of the old classifiers.", + "bbox": [ + 214, + 779, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Background Adaptation with Residual Modeling", + "bbox": [ + 408, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background Adaptation Losses. The Background Adaptation (BgA) losses, as shown in Fig. 2b, are designed for further refined adjustments on the background adaptation channel $\\mu_{b^t}^t$ . We propose two losses for two different regions of $\\mu_{b^t}^t$ according to the target effect. For regions $\\mathcal{R}_{C^t}$ belonging to the current classes $C^t$ , the ideal target is to have negative values as small as possible. To achieve this, we introduce cross-entropy loss $\\mathcal{L}_{BgA^{+}}$ for supervision:", + "bbox": [ + 212, + 146, + 787, + 239 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B g A ^ {+}} = - \\frac {1}{| \\mathcal {R} _ {C ^ {t}} |} \\sum_ {k \\in \\mathcal {R} _ {C ^ {t}}} \\log \\frac {\\exp \\left(- \\mu_ {b ^ {t} , k} ^ {t}\\right)}{1 + \\exp \\left(- \\mu_ {b ^ {t} , k} ^ {t}\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 247, + 785, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $|\\mathcal{R}_{C^t}|$ represents the number of pixels in $\\mathcal{R}_{C^t}$ . Besides, as for the regions of the background that might contain old classes, we hope the value of $\\mu_{bt}^{t}$ could be 0 for maintaining the historical background prediction result. However, achieving precise control over the whole region to be a precise value is challenging and may be impractical. Therefore, we set the positive logits in the adaptation channel of the old categories to 0 in Eq. 4 to avoid the noise and apply a Triplet loss $\\mathcal{L}_{BgA^{-}}$ to supervise the background region:", + "bbox": [ + 214, + 296, + 787, + 404 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B g A ^ {-}} = \\frac {1}{h w - \\left| \\mathcal {R} _ {C ^ {t}} \\right|} \\sum_ {k \\notin \\mathcal {R} _ {C ^ {t}}} \\max \\left[ 0, \\left(1 - \\varphi_ {b ^ {t}, k} ^ {t}\\right) ^ {2} - \\left(\\varphi_ {b ^ {t}, k} ^ {t} - 0\\right) ^ {2} \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 412, + 785, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first activate the background adaptation channel as $\\varphi_{b^t}^t$ via Sigmoid function, then the logits are mapped to the range from 0 to 1. We aim for the logits $\\mu_{b^t}^t$ of this region to be greater than 0 while not overfitting to the objects in the background, meaning that values of $\\varphi_{b^t}^t$ lie within the range of 0.5 to 1 and not close to 1. In particular, The Triplet loss $\\mathcal{L}_{BgA^{-}}$ respectively set 0 and 1 as negative and positive anchors. We simultaneously constrain the loss not to go below 0 by setting a lower bound. Therefore, once the value of the pixel exceeds 0.5, i.e., the middle of the positive and negative anchors, there will be no more gradient on the pixel for backpropagation, which mitigates overfitting to the objects in the background.", + "bbox": [ + 212, + 459, + 787, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Knowledge Distillation Strategies. In the training process of incremental steps for CISS tasks, substantial alterations to the existing parameters of $f^t$ frequently occur, giving rise to situations of catastrophic forgetting. Guided by the previous methods, we have devised a series of knowledge distillation strategies tailored to our background adaptation mechanism.", + "bbox": [ + 212, + 630, + 787, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Group Knowledge Distillation. In Sec. 3.2, we have introduced the classification logit map $\\mu_{C^i\\cup \\{b_i\\}}^t$ for each group of categories $C^i$ , which contains an additional channel for background adaptation. We introduce a Group Knowledge Distillation loss $\\mathcal{L}_{GKD}$ to ensure the stability of the old category logits, as well as the previous background adaptation channel:", + "bbox": [ + 223, + 713, + 787, + 789 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G K D} = - \\frac {1}{h w} \\sum_ {k = 1} ^ {h w} \\sum_ {i = 1} ^ {t - 1} \\sum_ {c \\in \\{c _ {b} ^ {i} \\} \\cup C ^ {i}} C E \\left(\\varphi_ {c, k} ^ {t}, \\varphi_ {c, k} ^ {t - 1}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 799, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\varphi_{*}^{t - 1}$ and $\\varphi_{*}^{t}$ denote the activated logits of $\\mu_{*}^{t - 1}$ and $\\mu_{*}^{t}$ , respectively. \n- Background Feature Distillation. When the network learning novel categories, not only the parameters of the novel classifier $f_{C^t}^t$ are optimized, but the encoder $f_{e}^{t}$ and the decoder $f_{d}^{t}$ are affected as well. Merely utilizing Group Knowledge Distillation loss $\\mathcal{L}_{GKD}$ is insufficient in effectively mitigating catastrophic forgetting as it primarily focuses on constraining classifiers of the old categories. However, excessively constraining parameter optimization can impede learning for novel categories. Taking all these factors into account, we designed a Background Feature Distillation (BFD) method that only acts on the features out of the regions of current categories $\\mathcal{R}_{C^t}$ :", + "bbox": [ + 225, + 145, + 782, + 296 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B F D} = \\sum_ {i = 1} ^ {t - 1} \\| \\psi_ {b} ^ {t} \\cdot F _ {m ^ {i}} ^ {t} - \\psi_ {b} ^ {t} \\cdot F _ {m ^ {i}} ^ {t - 1} \\| _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 306, + 785, + 347 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $F_{m^i}^t$ denotes the intermediate features in $f_{C^i}^t$ , $\\psi_b^t$ represents the binary mask out of $\\mathcal{R}_{C^t}$ , and $\\| \\cdot \\|_2$ calculates the Euclidean distance between the two matrix. The Background Feature Distillation allows other losses to supervise the region of novel categories $\\mathcal{R}_{C^t}$ while restricting the negative effect outside $\\mathcal{R}_{C^t}$ , where old categories $C^{1:t-1}$ could appear.", + "bbox": [ + 241, + 358, + 782, + 435 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Objective Function", + "text_level": 1, + "bbox": [ + 215, + 455, + 419, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With our proposed losses above, we construct our objective function as follows:", + "bbox": [ + 215, + 481, + 779, + 496 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {P B - B C E} + \\lambda_ {1} \\cdot \\mathcal {L} _ {B g A ^ {+}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {B g A ^ {-}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {G K D} + \\lambda_ {4} \\cdot \\mathcal {L} _ {B F D}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 507, + 784, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{PB - BCE}$ supervises the learning of novel classes, $\\mathcal{L}_{PB - BCE}$ and $\\mathcal{L}_{BgA^{-}}$ enhance the effect of Background Adaptation, $\\mathcal{L}_{GKD}$ and $\\mathcal{L}_{BFD}$ mitigate forgetting old classes. Hyperparameters $\\lambda_1,\\lambda_2,\\lambda_3,\\lambda_4$ balance these constraints.", + "bbox": [ + 215, + 535, + 782, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 603, + 374, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 633, + 426, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We follow the settings of most previous work and evaluate our proposed methods on Pascal VOC 2012 [23,26] and ADE20K [64] datasets. Pascal VOC 2012 is composed of 20 different classes with 10,582 images for training and 1,449 images for validation. ADE20K has 150 classes that contain 100 thing classes and 50 stuff classes. There are 20,210 images for training and 2,000 images for validation in ADE20K.", + "bbox": [ + 212, + 657, + 782, + 747 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Protocols. Earlier work [6, 21] has defined two types of training settings, i.e. overlapping and disjoint. The overlapping setting contains future classes in the previous steps, whereas the disjoint setting forbids future classes until they belong to the training classes. We follow recent work [3, 41, 56, 63] and adopt the overlapping setting for evaluating as it is more realistic. Besides, following the previous training protocols [6, 21, 56], we set each sub-task from the dataset as", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Background Adaptation with Residual Modeling", + "bbox": [ + 408, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$N_{ini} - N_{inc}$ , where $N_{ini}$ and $N_{inc}$ denotes the quantity of the categories in the initial step and a single incremental step, respectively. For example, the 15-1 sub-task of Pascal VOC 2012 has 15 classes for the initial training step. During each incremental step, one novel class is applied for incremental training without previous classes. Specifically, we define 4 sub-tasks on Pascal VOC 2012, including 19-1 (2 steps), 15-1 (6 steps), 10-1 (11 steps), and 5-3 (6 steps). On ADE20K, we define 4 sub-tasks containing 100-50 (2 steps), 100-10 (6 steps), 50-50 (3 steps), and 100-5 (11 steps).", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate our proposed method with widely-used mean Intersection-over-Union (mIoU). The Intersection-over-Union (IoU) is defined as $IoU = \\frac{TP}{TP + FP + FN}$ , where $TP$ , $FP$ , and $FN$ represent the quantity of true-positive, false-positive, and false-negative prediction pixels. The mIoU calculates the mean value of IoU corresponding to each category. To effectively measure the performance of CISS tasks, we compute the mIoU values of the initial classes $C^0$ , the incremental classes $C^{1:t}$ , and all classes $C^{0:t}$ , respectively.", + "bbox": [ + 212, + 267, + 787, + 372 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. Following the previous studies [6, 8, 40], we apply Deeplab-v3 [12] as the segmentation network with a ResNet-101 [27] for feature extraction. The ResNet-101 is initially pretrained on ImageNet-1K [18]. We use SGD optimizer to optimize the network. The initial learning rate of the initial step is set to $10^{-2}$ , while in the incremental steps, the initial learning rate is set to $10^{-3}$ for Pascal VOC 2012 and $10^{-2}$ for ADE20K. The momentum value is 0.9 and the weight decay is $10^{-4}$ in all steps. We apply the poly learning rate schedule following [3, 8, 12, 21]. During the initial step, the training epoch is set to 50 for Pascal VOC 2012 and 60 for ADE20K, while for each incremental step, we train the network for 20 epochs on Pascal VOC and 100 epochs on ADE20K. The batch size is 16 for Pascal VOC 2012 and 8 for ADE20K in all steps. The data augmentation follows the standard of previous work, which is composed of random scaling, random flipping, and random crop. The coefficients $\\lambda_{1}, \\lambda_{2}, \\lambda_{3}, \\lambda_{4}$ , of the objective function are set to 1, 5, 1, 4. The threshold of pseudo labeling $\\tau = 0.7$ is set to all settings. The implementation is built on Pytorch. The experiments are conducted on NVIDIA RTX2080Ti GPUs.", + "bbox": [ + 212, + 373, + 787, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 635, + 433, + 650 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison on Pascal VOC 2012. We evaluate our proposed method with current exemplar-free DeeplabV3-based methods on the Pascal VOC 2012 dataset. The sub-tasks of Pascal VOC 2012 have distinct characteristics. For example, the sub-task 19-1 has more categories in the initial step, while 5-3 has fewer classes for the initial training and more incremental steps, which is much more challenging. The results in Tab. 1 show that our proposed method outperforms other methods on novel classes and overall mIoU results of all sub-tasks. Particularly, our approach has $1.7\\%$ and $3.0\\%$ promotion of mIoU in more challenging sub-tasks of 10-1 and 5-3, respectively. Except for DKD [3] which refine the background class, our approach achieves at least $3.0\\%$ and $7.0\\%$ of promotion on mIoU Compared to the methods with a standard shared background classifier, demonstrating the effectiveness of the Background Adaptation mechanism.", + "bbox": [ + 212, + 657, + 789, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/43660eabb450ab643263f2647755ce6e30492e27b2c649a8390d5f763322e74f.jpg", + "table_caption": [ + "Table 1: Quantitative comparison with state-of-the-art exemplar-free methods on Pascal VOC 2012 in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best." + ], + "table_footnote": [], + "table_body": "
Method15-1 (6 steps)10-1 (11 steps)5-3 (6 steps)19-1 (1 step)
0-1516-20all0-1011-20all0-56-20all0-1920all
MiB [CVPR20] [6]38.013.532.212.213.112.657.142.546.771.222.168.9
SDR [CVPR21] [40]47.314.739.532.417.125.1---69.132.667.4
PLOP [CVPR21] [21]65.121.154.644.015.530.525.730.028.775.437.373.5
RCIL [CVPR22] [58]70.623.759.455.415.134.363.134.642.877.031.574.7
EWF [CVPR23] [56]77.732.767.071.530.351.961.742.247.777.96.774.5
AWT [WACV23] [25]59.117.249.133.218.026.061.845.950.4---
IDEC [TPAMI23] [63]77.036.567.370.746.359.167.149.054.1---
DKD [NeurIPS22] [3]78.142.769.773.146.560.469.653.558.177.841.576.0
Ours77.645.970.072.251.062.171.357.061.178.242.276.4
Joint79.872.477.478.476.477.476.977.677.477.577.077.4
", + "bbox": [ + 217, + 198, + 784, + 383 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ca7fb1225a0e710d7c023d178794f6b71262ccd0421ce0eb2cd34b11e94413a9.jpg", + "image_caption": [ + "(a) Task 15-1 of VOC2012." + ], + "image_footnote": [], + "bbox": [ + 256, + 402, + 480, + 535 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d6a7066a6e6348b8d583e1d7eb9a596af3caabffe512fba29c05505e365ca214.jpg", + "image_caption": [ + "(b) Task 10-1 of VOC2012.", + "Fig. 3: The step-wise mIoU comparison of our approach with previous methods under the sub-tasks 10-1 and 15-1 of Pascal VOC 2012." + ], + "image_footnote": [], + "bbox": [ + 519, + 402, + 754, + 534 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Besides, our method achieve an impressive advantage on the novel classes, with significant improvements of $4.5\\%$ and $3.5\\%$ compared to the previous state-of-the-art method. Many previous methods emphasize restricting catastrophic forgetting by boosting the mIoU of larger-proportion old classes in specific tasks (e.g. 15-1) to make the overall mIoU appear to be higher. However, our approach not only retains the ability to segment old classes but also significantly improves the accuracy of novel classes, aligning more closely with the incremental objective emphasized in CISS.", + "bbox": [ + 212, + 627, + 787, + 747 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison on ADE20K. We evaluate our method on the sub-tasks of more challenging ADE20K, including 100-50, 100-10, 100-5, and 50-50, as shown in Tab. 2. Although our method is slightly lower than the state-of-the-art method in short-term sub-tasks, we are remarkably ahead of other methods in more challenging long-term 100-10 and 100-5 tasks. The performance in the novel classes is still outstanding, which demonstrates the universality of our method.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Background Adaptation with Residual Modeling", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/82e5faaac809b7c7908c01cb9f183ad5261857483af98a6d0f98874583e0a5bc.jpg", + "table_caption": [ + "Table 2: Quantitative comparison with state-of-the-art exemplar-free methods on ADE20K in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best." + ], + "table_footnote": [], + "table_body": "
Method100-50 (2 steps)100-10 (6 steps)50-50 (3 steps)100-5 (11 steps)
0-100101-150all0-100101-150all0-5051-150all0-100101-150all
MiB [6]40.517.232.838.211.129.245.621.029.336.05.726.0
PLOP [21]41.914.932.940.513.631.648.821.030.439.17.828.8
RCIL [58]42.318.834.539.317.632.148.325.032.538.511.529.6
EWF [56]41.221.334.641.516.633.2---41.413.432.1
AWT [25]40.924.735.639.121.433.246.627.033.538.616.031.1
GSC [17]42.419.234.840.816.232.646.226.433.0---
IDEC [63]42.018.234.140.317.632.747.426.033.139.214.631.0
DKD [3]42.422.936.041.619.534.348.826.333.9---
Ours42.023.035.741.123.135.247.926.533.740.521.234.1
Joint44.328.238.944.328.238.951.133.338.944.328.238.9
", + "bbox": [ + 215, + 199, + 802, + 383 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 412, + 401, + 426 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Component Ablations. In this section, we analyze our proposed components, including the Background Adaptation mechanism, Background Feature Distillation loss $\\mathcal{L}_{BFD}$ , Group Knowledge Distillation loss $\\mathcal{L}_{GKD}$ , Background Adaptation losses $\\mathcal{L}_{BgA^{-}}$ and $\\mathcal{L}_{BgA^{+}}$ . The experiments are carried out on the representative sub-task 15-1 of Pascal VOC 2012. We separately evaluate different combinations of components and present the results in Tab. 3. The baseline method applies pseudo-labeling, $\\mathcal{L}_{PB-BCE}$ , and freezes the whole network except the classifiers of background and novel classes. Under the same experimental settings, the Background Adaptation mechanism significantly promotes $5.2\\%$ of overall mIoU, demonstrating its effectiveness in retaining the knowledge of old classes, and more notably, improving the performance of novel classes. The usage of distillation strategies with $\\mathcal{L}_{BFD}$ and $\\mathcal{L}_{GKD}$ preserves the performance of old classes while further boosting mIoU by $6.0\\%$ on novel classes. Moreover, the precise supervision of the Background Adaptation using $\\mathcal{L}_{BgA^{-}}$ and $\\mathcal{L}_{BgA^{+}}$ once again enhances the segmentation of novel classes with an additional $9.4\\%$ of mIoU. Subsequently, there is a growth of $2.4\\%$ on the overall mIoU, which proves that additional precise supervision is necessary. The results above validate the effectiveness of the Background Adaptation mechanism and other related losses. There is a notable improvement, particularly in the performance of novel classes, exceeding twice that of the baseline.", + "bbox": [ + 212, + 438, + 787, + 741 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Background Adaptation Schemes. Sec. 4.3 has presented the importance of the Background Adaptation losses. We have devised multiple schemes to achieve more optimal results of Background Adaptation, as shown in Tab. 4. Initially, we aim to have the Background Adaptation value for the current ground truth background region equal to 0, using Mean Square Error (MSE) loss to minimize", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d22940dadb8761178176141fc53c45f8d7cda9f502c1ee321f928bc0e7a6e67e.jpg", + "table_caption": [ + "Table 3: Ablation Study on the components of our proposed method on 15-1. BgA denotes Background Adaptation. In particular, the experiments $w / o$ distillation losses (i.e. $\\mathcal{L}_{BFD}$ and $\\mathcal{L}_{GKD}$ ) freeze the whole network except the classifier $f^t$ following the settings in [8]. All experiments contain $\\mathcal{L}_{PB-BCE}$ ." + ], + "table_footnote": [], + "table_body": "
BgA\\( \\mathcal{L}_{BFD} \\)\\( \\mathcal{L}_{GKD} \\)\\( \\mathcal{L}_{BgA^{-}} \\)\\( \\mathcal{L}_{BgA^{+}} \\)mIoUnimIoUincmIoU
73.421.861.1
✓77.430.566.3
✓✓76.737.167.3
✓✓✓77.336.567.6
✓✓✓✓77.143.469.0
✓✓✓✓76.634.166.5
✓✓✓✓✓77.645.970.0
", + "bbox": [ + 330, + 212, + 668, + 325 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/925fba54260f49aca97d4ce616742aa823bdd61cfdaa1c7e7a423639b75150dc.jpg", + "table_caption": [ + "Table 4: Ablation study on schemes of Table 5: Ablation Study on the feature Background Adaptation. distillation (FD) strategies." + ], + "table_footnote": [], + "table_body": "
FiltermIoUinimIoUincmIoU
MSE-0X74.241.866.5
✓74.642.266.9
BCE-1✓77.444.769.6
OursX65.745.460.8
✓77.645.970.0
", + "bbox": [ + 259, + 383, + 454, + 463 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9e692f94a2114a567603de853113a42a6767d1ddaa3e63428085f25b4a7c9619.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
mIoUinimIoUincmIoU
w/o FD56.933.251.2
LKD70.138.362.6
LMSE77.544.369.6
LBFD77.645.970.0
", + "bbox": [ + 547, + 385, + 743, + 463 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the deviation from 0 (MSE-0). However, MSE-0 couldn't restrict the background value to absolute 0, leading to noise accumulation and a lower mIoU compared to Background Adaptation $w/$ knowledge distillation. Later, we incorporate considerations for the current class region and introduce a Binary Cross-Entropy (BCE) loss (BCE-1). The expectation is to have the value in the current class region as small as possible, while the value in the background class region is as large as possible. Besides, we apply the Filter to remove the noise above 0. The BCE-1 scheme with Filter improved mIoU by $2\\%$ , but supervision in the background region tends to overfit to background objects. Therefore, we retain the BCE loss in the novel class region and redesign a triplet loss as in Sec. 3.3. The final Background Adaptation losses further enhance the segmentation capability for novel classes, resulting in a $1.2\\%$ improvement.", + "bbox": [ + 212, + 498, + 787, + 681 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Background Feature Distillation. To validate the effectiveness of Background Feature Distillation loss $\\mathcal{L}_{BFD}$ in Sec. 3.3, we compare $\\mathcal{L}_{BFD}$ with the method $w / o$ feature distillation ( $w / o$ FD), $w/$ standard knowledge distillation ( $\\mathcal{L}_{KD}$ ), and $w/$ mean square error ( $\\mathcal{L}_{MSE}$ ). As shown in Tab. 5, our $\\mathcal{L}_{BFD}$ outperforms other representative distillation methods. Particularly, the Background distillation with MSE, i.e., $\\mathcal{L}_{BFD}$ , shows a $1.6\\%$ promotion of mIoU on novel classes with no decline in old classes compared to $\\mathcal{L}_{MSE}$ . This demonstrates the necessity of preserving a tunnel for learning novel classes.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Background Adaptation with Residual Modeling", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8ce6c0bc2e05004791d7289fed033d4fbefd6fdae47a6344232430c3d2348247.jpg", + "image_caption": [ + "Fig. 4: Qualitative analysis on 15-1 of Pascal VOC 2012. The Baseline and Ours $w/o$ BgA losses are in the $1^{st}$ row and the $4^{th}$ row of Tab. 3. Ours applies additional Background Adaptation losses, compared to the Ours $w/o$ BgA losses." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 736, + 431 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 512, + 429, + 527 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We've conducted a qualitative analysis of our approach with Baseline and Ours $w / o$ BgA losses in the $1^{st}$ row and the $4^{th}$ row of Tab. 3. The results shown in Fig. 4 are visualized from the 15-1 sub-task of VOC 2012. Our approach in the last three columns reduces the misclassification (e.g., train in the first two columns of each method) and enhances the precision (e.g., sofa in the last column of each method) of the novel classes, while maintaining segmentation ability on the old classes. The results showcase the stability and plasticity of our approach.", + "bbox": [ + 212, + 534, + 787, + 642 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 660, + 369, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have introduced a novel CISS method featuring a Background Adaptation mechanism, designed to prevent the continual shifting of a shared background class and promote the learning of novel classes. The mechanism explicitly models the background residual rather than the background at each step and aggregates these residuals to represent the evolving background. To further enhance the performance of the Background Adaptation, we customized corresponding losses on the adaptation channel and improved the distillation scheme. Experiments on our proposed components verify their effectiveness under the Background Adaptation mechanism. Moreover, our method surpasses previous state-of-the-art methods in most sub-tasks, notably excelling in performance of novel classes.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No. 61972036) and the Industry-University-Institute Cooperation Foundation of the Eighth Research Institute of China Aerospace Science and Technology Corporation (No. SAST2022-049).", + "bbox": [ + 215, + 175, + 784, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, H., Cha, S., Lee, D., Moon, T.: Uncertainty-based continual learning with adaptive regularization. NeurIPS 32 (2019) 4", + "2. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. NeurIPS 32 (2019) 4", + "3. Baek, D., Oh, Y., Lee, S., Lee, J., Ham, B.: Decomposed knowledge distillation for class-incremental semantic segmentation. NeurIPS 35, 10380-10392 (2022) 2, 4, 9, 10, 11, 12", + "4. Borsos, Z., Mutny, M., Krause, A.: Coresets via bilevel optimization for continual learning and streaming. NeurIPS 33, 14879-14890 (2020) 4", + "5. Boschini, M., Bonicelli, L., Porrello, A., Bellitto, G., Pennisi, M., Palazzo, S., Spampinato, C., Calderara, S.: Transfer without forgetting. In: ECCV. pp. 692-709 (2022) 4", + "6. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: CVPR. pp. 9233-9242 (2020) 2, 4, 9, 10, 11, 12", + "7. Cha, H., Lee, J., Shin, J.: Co2l: Contrastive continual learning. In: ICCV. pp. 9516-9525 (2021) 4", + "8. Cha, S., Yoo, Y., Moon, T., et al.: Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. NeurIPS 34, 10919-10930 (2021) 2, 3, 4, 6, 7, 10, 13", + "9. Chen, J., Cong, R., Luo, Y., Ip, H., Kwong, S.: Saving 100x storage: Prototype replay for reconstructing training sample distribution in class-incremental semantic segmentation. Advances in Neural Information Processing Systems 36 (2024) 3", + "10. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062 (2014) 1", + "11. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI 40(4), 834-848 (2017) 1", + "12. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017) 1, 4, 10", + "13. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: ECCV. pp. 801-818 (2018) 1", + "14. Chen, Z., Liu, B.: Lifelong machine learning, vol. 1. Springer (2018) 2", + "15. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022) 1, 4", + "16. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. NeurIPS 34, 17864-17875 (2021) 1" + ], + "bbox": [ + 217, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Background Adaptation with Residual Modeling", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Cong, W., Cong, Y., Dong, J., Sun, G., Ding, H.: Gradient-semantic compensation for incremental semantic segmentation. arXiv preprint arXiv:2307.10822 (2023) 12", + "18. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009) 10", + "19. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: CVPR. pp. 5138-5146 (2019) 4", + "20. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 1, 4", + "21. Douillard, A., Chen, Y., Dapogny, A., Cord, M.: Plop: Learning without forgetting for continual semantic segmentation. In: CVPR. pp. 4040-4050 (2021) 2, 3, 4, 6, 7, 9, 10, 11, 12", + "22. Ermis, B., Zappella, G., Wistuba, M., Rawal, A., Archambeau, C.: Memory efficient continual learning with transformers. NeurIPS 35, 10629-10642 (2022) 4", + "23. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. IJCV 88, 303-338 (2010) 9", + "24. Fang, Y., Zhu, F., Cheng, B., Liu, L., Zhao, Y., Wei, Y.: Locating noise is halfway denoising for semi-supervised segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16612-16622 (2023) 1", + "25. Goswami, D., Schuster, R., van de Weijer, J., Stricker, D.: Attribution-aware weight transfer: A warm-start initialization for class-incremental semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3195-3204 (2023) 11, 12", + "26. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: ICCV. pp. 991-998 (2011) 9", + "27. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) 1, 10", + "28. Jung, H., Ju, J., Jung, M., Kim, J.: Less-forgetting learning in deep neural networks. arXiv preprint arXiv:1607.00122 (2016) 2", + "29. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13), 3521-3526 (2017) 4", + "30. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. NeurIPS 25 (2012) 1", + "31. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) 1", + "32. Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI 40(12), 2935-2947 (2017) 2, 4", + "33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012-10022 (2021) 1", + "34. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015) 1", + "35. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. NeurIPS 30 (2017) 4", + "36. Madaan, D., Yoon, J., Li, Y., Liu, Y., Hwang, S.J.: Representational continuity for unsupervised continual learning. In: ICLR (2022) 4", + "37. Maracani, A., Michieli, U., Toldo, M., Zanuttigh, P.: Recall: Replay-based continual learning in semantic segmentation. In: ICCV. pp. 7026-7035 (2021) 4" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. McClelland, J.L., McNaughton, B.L., O'Reilly, R.C.: Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review 102(3), 419 (1995) 2, 4", + "39. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989) 2, 4", + "40. Michieli, U., Zanuttigh, P.: Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In: CVPR. pp. 1114-1124 (2021) 2, 4, 10, 11", + "41. Oh, Y., Baek, D., Ham, B.: Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. NeurIPS 35, 14516-14528 (2022) 3, 4, 6, 9", + "42. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: CVPR. pp. 11321-11329 (2019) 4", + "43. Pham, Q., Liu, C., Hoi, S.: Dualnet: Continual learning, fast and slow. NeurIPS 34, 16131-16144 (2021) 4", + "44. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 4", + "45. Ritter, H., Botev, A., Barber, D.: Online structured laplace approximations for overcoming catastrophic forgetting. NeurIPS 31 (2018) 4", + "46. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 234-241 (2015) 1", + "47. Schwarz, J., Czarnecki, W., Luketina, J., Grabska-Barwinska, A., Teh, Y.W., Pascanu, R., Hadsell, R.: Progress & compress: A scalable framework for continual learning. In: Proc. International Conference on Machine Learning. pp. 4528-4537. PMLR (2018) 4", + "48. Shang, C., Li, H., Meng, F., Wu, Q., Qiu, H., Wang, L.: Incrementer: Transformer for class-incremental semantic segmentation with knowledge distillation focusing on old class. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7214-7224 (2023) 4", + "49. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. NeurIPS 30 (2017) 4", + "50. Shon, H., Lee, J., Kim, S.H., Kim, J.: Dlcft: Deep linear continual fine-tuning for general incremental learning. In: ECCV. pp. 513-529 (2022) 4", + "51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) 1", + "52. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021) 4", + "53. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: Proc. International Conference on Machine Learning. pp. 10347-10357 (2021) 1", + "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) 1", + "55. Wu, C., Herranz, L., Liu, X., Van De Weijer, J., Raducanu, B., et al.: Memory replay gans: Learning to generate new categories without forgetting. NeurIPS 31 (2018) 4" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Background Adaptation with Residual Modeling", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Xiao, J.W., Zhang, C.B., Feng, J., Liu, X., van de Weijer, J., Cheng, M.M.: Endpoints weight fusion for class incremental semantic segmentation. In: CVPR. pp. 7204-7213 (2023) 4, 9, 11, 12", + "57. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Nabi, M., Alameda-Pineda, X., Ricci, E.: Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE TPAMI 45(2), 2567-2581 (2022) 4", + "58. Zhang, C.B., Xiao, J.W., Liu, X., Chen, Y.C., Cheng, M.M.: Representation compensation networks for continual semantic segmentation. In: CVPR. pp. 7053-7064 (2022) 4, 11, 12", + "59. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19148-19158 (2023) 2", + "60. Zhang, J.O., Sax, A., Zamir, A., Guibas, L., Malik, J.: Side-tuning: a baseline for network adaptation via additive side networks. In: ECCV. pp. 698-714 (2020) 4", + "61. Zhang, Z., Gao, G., Fang, Z., Jiao, J., Wei, Y.: Mining unseen classes via regional objectness: A simple baseline for incremental segmentation. NeurIPS 35, 24340-24353 (2022) 2, 4, 7", + "62. Zhang, Z., Gao, G., Jiao, J., Liu, C.H., Wei, Y.: Coinseg: Contrast inter-and intra-class representations for incremental segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 843-853 (2023) 4", + "63. Zhao, D., Yuan, B., Shi, Z.: Inherit with distillation and evolve with contrast: Exploring class incremental semantic segmentation without exemplar memory. IEEE TPAMI (2023) 4, 9, 11, 12", + "64. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: CVPR. pp. 633-641 (2017) 9", + "65. Zhu, H., Wei, Y., Liang, X., Zhang, C., Zhao, Y.: Ctp: Towards vision-language continual pretraining via compatible momentum contrast and topology preservation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22257-22267 (2023) 2", + "66. Zhu, L., Chen, T., Yin, J., See, S., Liu, J.: Continual semantic segmentation with automatic memory sample selection. In: CVPR. pp. 3082-3092 (2023) 4" + ], + "bbox": [ + 215, + 146, + 784, + 590 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_model.json b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c4c54214a5867153b7bef7754f1ee76c813c4c42 --- /dev/null +++ b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_model.json @@ -0,0 +1,2766 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.226, + 0.141, + 0.78, + 0.21 + ], + "angle": 0, + "content": "Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.235, + 0.631, + 0.251 + ], + "angle": 0, + "content": "Anqi Zhang and Guangyu Gao\\*" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.263, + 0.657, + 0.278 + ], + "angle": 0, + "content": "School of Computer Science and Technology," + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.278, + 0.657, + 0.291 + ], + "angle": 0, + "content": "Beijing Institute of Technology, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.293, + 0.582, + 0.305 + ], + "angle": 0, + "content": "andy_zaq@outlook.com" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.306, + 0.584, + 0.319 + ], + "angle": 0, + "content": "guangyugao@bit.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.355, + 0.756, + 0.701 + ], + "angle": 0, + "content": "Abstract. Class Incremental Semantic Segmentation (CISS), within Incremental Learning for semantic segmentation, targets segmenting new categories while reducing the catastrophic forgetting on the old categories. Besides, background shifting, where the background category changes constantly in each step, is a special challenge for CISS. Current methods with a shared background classifier struggle to keep up with these changes, leading to decreased stability in background predictions and reduced accuracy of segmentation. For this special challenge, we designed a novel background adaptation mechanism, which explicitly models the background residual rather than the background itself in each step, and aggregates these residuals to represent the evolving background. Therefore, the background adaptation mechanism ensures the stability of previous background classifiers, while enabling the model to concentrate on the easy-learned residuals from the additional channel, which enhances background discernment for better prediction of novel categories. To precisely optimize the background adaptation mechanism, we propose Pseudo Background Binary Cross-Entropy loss and Background Adaptation losses, which amplify the adaptation effect. Group Knowledge Distillation and Background Feature Distillation strategies are designed to prevent forgetting old categories. Our approach, evaluated across various incremental scenarios on Pascal VOC 2012 and ADE20K datasets, outperforms prior exemplar-free state-of-the-art methods with mIoU of \\(3.0\\%\\) in VOC 10-1 and \\(2.0\\%\\) in ADE 100-5, notably enhancing the accuracy of new classes while mitigating catastrophic forgetting. Code is available in https://andyzaq.github.io/barmsite/." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.726, + 0.377, + 0.741 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.819 + ], + "angle": 0, + "content": "With the development of deep learning and the support of convolutional neural networks (CNNs) [27,30,31,51] and Transformers [20,33,53,54], current semantic segmentation networks [10-13,15,16,24,34,46] have achieved remarkable results on standard supervised learning tasks. However, real-world data, arriving as an" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.39, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "unstable continuous stream, poses the challenge of retaining prior knowledge while accommodating new concepts, termed as catastrophic forgetting [38, 39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.435 + ], + "angle": 0, + "content": "To tackle this problem, incremental learning is proposed to adapt to changing data streams for new concepts, but also avoid forgetting old knowledge, especially for the classification task, i.e., Class-Incremental Learning (CIL) [14,28,32,59,65]. Cermelli et al. [6] introduced CIL into Semantic Segmentation as Class Incremental Semantic Segmentation (CISS). Since the semantic segmentation tasks involve dense predictions, the issue of catastrophic forgetting typically becomes more challenging. Most recent works [6,8,21] have struggled to mitigate catastrophic forgetting through knowledge distillation [3,21,40] or parameter freezing [8,61]. Although these strategies effectively alleviate catastrophic forgetting, it comes at the cost of the model's plasticity, making it challenging to learn novel classes. In current state-of-the-art approaches [3,61], the performance of established old classes approaches the upper limit of joint training, yet there still remains a significant gap in mIoU for novel classes. However, in the context of lifelong learning scenarios featuring an infinite influx of novel classes, the incremental learner's plasticity becomes increasingly crucial. Therefore, the optimal solution should be fine-tuning more parameters for novel classes, while catastrophic forgetting needs to be handled properly." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.462, + 0.316, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.517, + 0.316, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.571, + 0.315, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.626, + 0.315, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.68, + 0.297, + 0.689 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.464, + 0.388, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.517, + 0.388, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.571, + 0.388, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.626, + 0.388, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.68, + 0.381, + 0.698 + ], + "angle": 0, + "content": "Prediction of step \\(t - 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.464, + 0.458, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.517, + 0.458, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.571, + 0.458, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.626, + 0.458, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.68, + 0.45, + 0.698 + ], + "angle": 0, + "content": "Bg logits of step \\(t - 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.464, + 0.53, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.517, + 0.53, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.571, + 0.53, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.626, + 0.53, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.68, + 0.533, + 0.698 + ], + "angle": 0, + "content": "Bg Adaptation of step \\(t\\)" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.464, + 0.601, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.517, + 0.601, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.571, + 0.602, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.626, + 0.602, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.68, + 0.604, + 0.698 + ], + "angle": 0, + "content": "Bg logits of step \\(t\\) (Sigmoid)" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.461, + 0.678, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.517, + 0.678, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.571, + 0.678, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.626, + 0.678, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.68, + 0.67, + 0.698 + ], + "angle": 0, + "content": "Prediction of step t" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.461, + 0.754, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.517, + 0.754, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.571, + 0.754, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.626, + 0.754, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.68, + 0.744, + 0.698 + ], + "angle": 0, + "content": "Categories in step t" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.71, + 0.788, + 0.781 + ], + "angle": 0, + "content": "Fig. 1: 3D Visualization of Background Adaptation results. The background logits for step \\( t \\) combine those from step \\( t - 1 \\) with the adaptation logits learned in step \\( t \\), preventing disorderly adjustments and focusing on residuals. The current step background logits are processed using the Sigmoid function. Note that colors range from red (small values) to blue (large values)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Nevertheless, background shifting, where the background category changes constantly in each step, is another challenge for CISS. Upon further investiga" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Background Adaptation with Residual Modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.448 + ], + "angle": 0, + "content": "tion, it was observed that previous methods within CISS often utilize a shared background classifier across each step. However, novel classes are typically considered as the background in previous learning steps, indicating that the background class undergoes continuous changes throughout incremental learning. This way results in an incessantly shifting background in each step, leading to disruption in background prediction and inaccurate learning of novel classes. Some approaches [8,9,41] store old exemplars in extra memory space to prevent background shifting, while long-term storage of such data may not be feasible in real-world scenarios. To enhance the incremental nature of the CISS method and reduce background instability, we revise the representation of the background class and propose an exemplar-free background adaptation mechanism. In particular, the background adaptation mechanism explicitly models the background residual rather than the background itself in each incremental step. The learned residuals of the additional channel are aggregated with the previously learned background logits to represent the evolving background, as illustrated in Fig. 1. By employing the background adaptation mechanism, we only need to optimize the additional channel of residuals for constructing the background during the incremental steps. This approach eliminates the need for repeatedly learning background classifiers associated with old classes, thereby avoiding the instability of optimizing a shared background classifier." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.453, + 0.786, + 0.738 + ], + "angle": 0, + "content": "Moreover, we customize our training approaches in order to fit the training process of the background adaptation structure and promote its performance. We first introduce the widely used pseudo-labeling [8,21] to generate a pseudobackground, and then we design a PB-BCE loss to focus on optimizing the classifier of novel classes and related background adaptation residual. Besides, for the Background Adaptation channel, the objective is to minimize the values in the regions of novel categories, while ensuring that the values in other regions remain above 0. Thus, two auxiliary Background Adaptation losses are specifically applied to the background adaptation channel for precise adjustment. Given our pursuit of plasticity on novel categories without external memory, we have devised more comprehensive exemplar-free knowledge distillation strategies to combat catastrophic forgetting. We further adopt knowledge distillation strategies on the logit maps of old categories and the intermediate features, namely the Group Knowledge Distillation and Background Feature Distillation, respectively. Finally, our approach outperforms previous methods in multiple benchmarks, especially in realistic and challenging long-term scenarios such as 10-1 of VOC2012 and 100-5 of ADE20K. Our approach achieves significant performance gains of \\(4.5\\%\\) and \\(5.2\\%\\) on novel classes, respectively, compared to the previous state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.743, + 0.625, + 0.756 + ], + "angle": 0, + "content": "Overall, our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "- We propose a Background Adaptation strategy in incremental steps by optimizing an additional channel for residual modeling instead of the shared background classifier, which avoids unstable optimization and prevents chaos in background logits." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "- We design PB-BCE loss and Background Adaptation losses to enhance the effectiveness of Background Adaptation, while two distillation strategies ensure maintaining knowledge of old categories." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.192, + 0.785, + 0.22 + ], + "angle": 0, + "content": "- Experiments validate the effectiveness of our proposed method, showcasing a new state-of-the-art performance on CISS benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.24, + 0.398, + 0.256 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.268, + 0.489, + 0.284 + ], + "angle": 0, + "content": "2.1 Class Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.788, + 0.516 + ], + "angle": 0, + "content": "Class Incremental Learning (CIL) breaks the constraints of the standard training process in an incremental form on new classes. CIL task aims to continually learn the new classes without retraining on the data of old classes, which significantly reduces the training cost. As a compromise, the CIL task has to effectively solve the problem of catastrophic forgetting [38, 39]. Current CIL methods are divided into several main categories. Replay-based methods recall the data of old classes during incremental training. Some of the approaches [2, 4, 35] reserve memory buffers and temporarily store representative past data. Others [42, 49, 55] introduce powerful generative models to generate pseudo-data for rehearsal. Regularization-based methods introduce regularization terms to prevent forgetting the old classes. These methods utilize loss functions as penalties [1, 29, 45, 47] or knowledge distillation from the previous model [19, 32, 44]. Representation-based methods attempt to transfer representations to certain categories, which apply adaptation on a fixed pre-trained backbone [5, 22, 50, 60] or self-supervised learning [7, 36, 43]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.533, + 0.609, + 0.548 + ], + "angle": 0, + "content": "2.2 Class Incremental Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The achievements in CIL tasks encourage the Incremental Learning pipeline to develop in other dense prediction tasks, including Semantic Segmentation. Class Incremental Semantic Segmentation (CISS) task is first introduced in MiB [6], which focused on reconstructing the background region under the restriction from knowledge distillation. Since then, various methods [57, 63] have been introduced in the CISS task. PLOP [21] defined the pseudo-labeling and multi-scale local distillation to remember the old categories. SDR [40] introduced category-wise prototype matching and contrastive learning methods to enlarge the difference between categories. RCIL [58] and EWF [56] explored methods for integrating parameters of new and old models. DKD [3] decomposed the class logits into positive and negative logits for better knowledge distillation and applied threshold to generate background regions. Incrementer [48] borrows prompt learning from ViT [20] and applies the structure of Segmenter [52] instead of DeeplabV3 [12], which is the basic structure of other methods. Besides, several approaches utilize external memory [66] or models [37, 62] for better performance. SSUL [8] introduced a Saliency-map detector to identify the saliency region and store some of the previous instances in memory. MicroSeg [61] utilized proposals from pretrained Mask2Former [15] as a guidance. ALIFE [41] memorized category features and adapted for remembering the old classes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Background Adaptation with Residual Modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.145, + 0.657, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.32, + 0.555, + 0.332 + ], + "angle": 0, + "content": "(a) The overall pipeline of our approach." + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.196, + 0.781, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.319, + 0.769, + 0.332 + ], + "angle": 0, + "content": "(b) \\(\\mathrm{BgA}\\) losses." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig. 2: Overview of our framework that consists of a Background Adaptation mechanism and proposed losses. (a) The Background Adaptation mechanism uses the negative part of the adaptation channel to rectify the background logits. (b) Two Background Adaptation losses separately optimize the regions \\( w/ \\) and \\( w/o \\) novel categories to enhance the performance of the Background Adaptation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.442, + 0.339, + 0.457 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.475, + 0.373, + 0.489 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.789, + 0.73 + ], + "angle": 0, + "content": "The Class-Incremental Semantic Segmentation (CISS) task performs training in \\( T \\) steps. During step \\( t \\), the training data \\( D^{t} \\) containing the labels of the novel classes \\( C^{t} \\) are applied for training, while the labels of old classes \\( C^{1:t-1} \\) are expunged. In particular, each sample of the training data \\( D^{t} \\) consists of an RGB image \\( x^{t} \\in \\mathbb{R}^{3 \\times H \\times W} \\) and a corresponding ground truth label \\( y^{t} \\) on novel classes \\( C^{t} \\). Besides, the ground truth label \\( y^{t} \\) replaces the labels of the \\( C^{1:t-1} \\) with the background label \\( c_{b} \\) as the old classes will not appear in \\( C^{t} \\) to ensure \\( C^{1:t-1} \\cap C^{t} = \\emptyset \\). Therefore, the model \\( f^{t} \\) being trained at step \\( t \\) not only needs to learn to segment novel categories but also has to retain the capability to segment old categories. However, while learning the novel categories, the model \\( f^{t} \\) tends to overfit the novel categories \\( C^{t} \\) as \\( f^{t} \\) cannot contact with samples of \\( C^{1:t-1} \\), causing severe catastrophic forgetting. Moreover, the novel categories can only appear in the background region during the previous steps. We are more likely to classify them from the previous background prediction region, which motivates us to focus on background adaptation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.46, + 0.769 + ], + "angle": 0, + "content": "3.2 Background Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Previous studies often apply only one background classifier. They have to readjust the parameters of the background classifier whenever the model learns novel categories, which causes catastrophic forgotten even if the classifier of the old categories remains the same probability output. The unstable background classifier" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "could be a thorny constraint factor. To tackle the problems above, we design a Background Adaptation mechanism. Different from the previous distinct background classifier, we introduce an additional background adaptation channel for each classifier of the incremental process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.285 + ], + "angle": 0, + "content": "During the training process of incremental step \\( t \\), the current model \\( f^t \\) inherits the parameters from the previous model \\( f^{t-1} \\) and adds a new classifier \\( f_{C^t}^t \\) for parallel classification of categories in \\( C^t \\) alongside the previous classifiers. The encoder \\( f_e^t \\) and the decoder \\( f_d^t \\) transform the input image \\( x^t \\) into features map \\( F_d^t \\in \\mathbb{R}^{h \\times w \\times c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.287, + 0.785, + 0.303 + ], + "angle": 0, + "content": "\\[\nF _ {d} ^ {t} = f _ {d} ^ {t} \\left(f _ {e} ^ {t} \\left(x ^ {t}\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.784, + 0.395 + ], + "angle": 0, + "content": "Compared to the standard classifier of the previous methods [8, 21, 41], our classifier of each incremental step has an additional channel for background adaptation. Therefore, for each group of categories \\( C^i \\) from step \\( i \\), the classifier \\( f_{C^i}^t \\) of the model \\( f^t \\) generates the corresponding classification logit maps \\( \\mu_{C^i \\cup \\{b_i\\}}^t \\in \\mathbb{R}^{h \\times w \\times (|C^i| + 1)} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.408, + 0.785, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mu_ {C ^ {i} \\cup \\{b _ {i} \\}} ^ {t} = f _ {C ^ {i}} ^ {t} \\left(F _ {d} ^ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.608 + ], + "angle": 0, + "content": "where \\( i \\in \\{1, 2, \\ldots, t\\} \\). We define \\( b^1 \\) to represent the background class of the initial step, and \\( b^i \\) denotes the background adaptation for subsequent incremental steps when \\( i >= 2 \\). The symbol \\( b \\) refers to the adapted background class. Thus, \\( \\mu_{C^i \\cup \\{b_i\\}}^t \\) contains the probability maps \\( \\mu_{C^i} \\) of \\( |C^i| \\) categories from step \\( i \\) and an additional channel \\( \\mu_{b^i}^t \\) for constructing the background logits \\( \\mu_b^t \\). Considering that the region of the novel categories \\( C^t \\) is more likely inherited from the background region of the previous step, the classifier \\( f_{C^t}^t \\) adjusts the background logits through \\( \\mu_{b^t}^t \\), which decreases the probability values in the region \\( \\mathcal{R}_{C^t} \\) of \\( C^t \\). The parameters of previous classifiers are not required to continuously optimize for new background regions, thereby avoiding instability in the classification of old categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.785, + 0.716 + ], + "angle": 0, + "content": "The background adaptation channel \\(\\mu_{b^t}^t\\) can correct the background logits \\(\\mu_b^t\\in \\mathbb{R}^{h\\times w\\times 1}\\) in \\(\\mathcal{R}_{C^t}\\), yet the noise in other regions of \\(\\mu_{b^t}^t\\) may interfere with the background prediction. Particularly, when old categories are present in the image, there is a risk that these regions could be incorrectly covered by the adapted background. Hence, we design a Filter operation that replaces positive values of \\(\\mu_{b^i}^t\\) with 0. As shown in Fig. 2a, with the Filter operation applied, the residual modeling of background logits during the inference process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.731, + 0.785, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mu_ {b} ^ {t} = \\mu_ {b ^ {1}} ^ {t} + \\sum_ {i = 2} ^ {t} F i l t e r \\left(\\mu_ {b ^ {i}} ^ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.841 + ], + "angle": 0, + "content": "where the background logits from the initial step \\(\\mu_{b^1}^t\\) stay the same and the background adaptation logits from the incremental steps \\(\\mu_{b^i}^t\\) only remain the negative values via the Filter. In particular, to make \\(\\mu_{b}^{t}\\) more friendly for gradient backpropagation, we abandoned the Filter on the background adaptation logits" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Background Adaptation with Residual Modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.557, + 0.164 + ], + "angle": 0, + "content": "of novel classes \\(\\mu_{b^t}^t\\) during the training process:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.176, + 0.786, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mu_ {b} ^ {t} = \\mu_ {b ^ {1}} ^ {t} + \\sum_ {i = 2} ^ {t - 1} \\operatorname {F i l t e r} \\left(\\mu_ {b ^ {i}} ^ {t}\\right) + \\mu_ {b ^ {t}} ^ {t}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.24, + 0.513, + 0.257 + ], + "angle": 0, + "content": "3.3 Incremental Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.268, + 0.784, + 0.362 + ], + "angle": 0, + "content": "Pseudo Labeling. During the incremental steps, even if objects of old categories \\( C^{1:t-1} \\) exist in the image, we cannot obtain corresponding labels from the ground truth. In order to simulate a ground truth label that contains the categories we have learned, we follow [8,21,61] to introduce the prediction probability of the previous step \\( \\varphi^{t-1} \\in \\mathbb{R}^{h \\times w \\times |C^{1:t-1}|} \\) from \\( f^{t-1}(x^t) \\) with Sigmoid function and combine it with the current ground truth label \\( y^t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.373, + 0.786, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\hat {y} ^ {t} = \\left\\{ \\begin{array}{l l} y ^ {t}, & \\left(y ^ {t} \\in C ^ {t}\\right) \\vee \\left\\{\\left(y ^ {t} \\in c _ {b}\\right) \\wedge \\left(\\max \\left(\\varphi^ {t - 1}\\right) < \\tau\\right) \\right\\} \\\\ \\tilde {y} ^ {t - 1}, & \\left(y ^ {t} \\in c _ {b}\\right) \\wedge \\left(\\max \\left(\\varphi^ {t - 1}\\right) \\geq \\tau\\right) \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.426, + 0.784, + 0.459 + ], + "angle": 0, + "content": "where \\(\\hat{y}^t\\in \\mathbb{R}^{h\\times w}\\) denotes the generated pseudo label and \\(\\tilde{y}^{t - 1} = \\arg \\max (\\varphi^{t - 1})\\) represents the final prediction from \\(f^{t - 1}(x^t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.482, + 0.788, + 0.636 + ], + "angle": 0, + "content": "PB-BCE Loss. Binary Cross-Entropy (BCE) loss is extensively employed for learning to segment objects of novel categories. We follow the usage of the Sigmoid function and activate \\(\\mu^t\\) as \\(\\varphi^t\\). Some of the CISS methods [8,21,61] utilize the pseudo mask as the ground truth for supervision, which effectively replaces the background label in the regions of the old categories. However, the ambiguous supervision of the old categories could improperly optimize the classifiers of \\(C^{1:t - 1}\\) and destroy their identification capability. Therefore, we design a Pseudo Background BCE (PB-BCE) loss which cuts off the backpropagation of old categories, meaning that we only take the labels of current categories \\(\\hat{y}_{C^t}^t\\) and pseudo-background \\(\\hat{y}_{c_b}^t\\) for optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.648, + 0.786, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P B - B C E} = - \\frac {1}{h w} \\sum_ {k = 1} ^ {h w} \\sum_ {c \\in \\{c _ {b} \\} \\cap C ^ {t}} C E \\left(\\varphi_ {c, k} ^ {t}, \\hat {\\psi} _ {c, k} ^ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.707, + 0.784, + 0.738 + ], + "angle": 0, + "content": "where \\( k \\) represents the index of pixel and \\( \\hat{\\psi}_c^t \\) denotes the binary map of class c in pseudo map \\( \\hat{y}^t \\). \\( CE \\) is the cross-entropy calculation:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.748, + 0.786, + 0.768 + ], + "angle": 0, + "content": "\\[\nC E \\left(\\varphi_ {*} ^ {t}, \\hat {\\psi} _ {*} ^ {t}\\right) = \\hat {\\psi} _ {*} ^ {t} \\cdot \\log \\varphi_ {*} ^ {t} + \\left(1 - \\hat {\\psi} _ {*} ^ {t}\\right) \\cdot \\log \\left(1 - \\varphi_ {*} ^ {t}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Although the whole background prediction logits \\(\\mu_b^t\\) are involved in the loss calculation, only the background adaptation channel for current categories \\(\\mu_{b^t}^t\\) serves for optimization. Previous background logits are detached to prevent disorderly evolving on parameters of the old classifiers." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.24 + ], + "angle": 0, + "content": "Background Adaptation Losses. The Background Adaptation (BgA) losses, as shown in Fig. 2b, are designed for further refined adjustments on the background adaptation channel \\(\\mu_{b^t}^t\\). We propose two losses for two different regions of \\(\\mu_{b^t}^t\\) according to the target effect. For regions \\(\\mathcal{R}_{C^t}\\) belonging to the current classes \\(C^t\\), the ideal target is to have negative values as small as possible. To achieve this, we introduce cross-entropy loss \\(\\mathcal{L}_{BgA^{+}}\\) for supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.248, + 0.786, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B g A ^ {+}} = - \\frac {1}{| \\mathcal {R} _ {C ^ {t}} |} \\sum_ {k \\in \\mathcal {R} _ {C ^ {t}}} \\log \\frac {\\exp \\left(- \\mu_ {b ^ {t} , k} ^ {t}\\right)}{1 + \\exp \\left(- \\mu_ {b ^ {t} , k} ^ {t}\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.297, + 0.789, + 0.405 + ], + "angle": 0, + "content": "where \\( |\\mathcal{R}_{C^t}| \\) represents the number of pixels in \\( \\mathcal{R}_{C^t} \\). Besides, as for the regions of the background that might contain old classes, we hope the value of \\( \\mu_{bt}^{t} \\) could be 0 for maintaining the historical background prediction result. However, achieving precise control over the whole region to be a precise value is challenging and may be impractical. Therefore, we set the positive logits in the adaptation channel of the old categories to 0 in Eq. 4 to avoid the noise and apply a Triplet loss \\( \\mathcal{L}_{BgA^{-}} \\) to supervise the background region:" + }, + { + "type": "equation", + "bbox": [ + 0.278, + 0.414, + 0.786, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B g A ^ {-}} = \\frac {1}{h w - \\left| \\mathcal {R} _ {C ^ {t}} \\right|} \\sum_ {k \\notin \\mathcal {R} _ {C ^ {t}}} \\max \\left[ 0, \\left(1 - \\varphi_ {b ^ {t}, k} ^ {t}\\right) ^ {2} - \\left(\\varphi_ {b ^ {t}, k} ^ {t} - 0\\right) ^ {2} \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.611 + ], + "angle": 0, + "content": "We first activate the background adaptation channel as \\(\\varphi_{b^t}^t\\) via Sigmoid function, then the logits are mapped to the range from 0 to 1. We aim for the logits \\(\\mu_{b^t}^t\\) of this region to be greater than 0 while not overfitting to the objects in the background, meaning that values of \\(\\varphi_{b^t}^t\\) lie within the range of 0.5 to 1 and not close to 1. In particular, The Triplet loss \\(\\mathcal{L}_{BgA^{-}}\\) respectively set 0 and 1 as negative and positive anchors. We simultaneously constrain the loss not to go below 0 by setting a lower bound. Therefore, once the value of the pixel exceeds 0.5, i.e., the middle of the positive and negative anchors, there will be no more gradient on the pixel for backpropagation, which mitigates overfitting to the objects in the background." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Knowledge Distillation Strategies. In the training process of incremental steps for CISS tasks, substantial alterations to the existing parameters of \\( f^t \\) frequently occur, giving rise to situations of catastrophic forgetting. Guided by the previous methods, we have devised a series of knowledge distillation strategies tailored to our background adaptation mechanism." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.714, + 0.788, + 0.79 + ], + "angle": 0, + "content": "- Group Knowledge Distillation. In Sec. 3.2, we have introduced the classification logit map \\(\\mu_{C^i\\cup \\{b_i\\}}^t\\) for each group of categories \\(C^i\\), which contains an additional channel for background adaptation. We introduce a Group Knowledge Distillation loss \\(\\mathcal{L}_{GKD}\\) to ensure the stability of the old category logits, as well as the previous background adaptation channel:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.8, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G K D} = - \\frac {1}{h w} \\sum_ {k = 1} ^ {h w} \\sum_ {i = 1} ^ {t - 1} \\sum_ {c \\in \\{c _ {b} ^ {i} \\} \\cup C ^ {i}} C E \\left(\\varphi_ {c, k} ^ {t}, \\varphi_ {c, k} ^ {t - 1}\\right) \\tag {10}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Background Adaptation with Residual Modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.146, + 0.784, + 0.297 + ], + "angle": 0, + "content": "where \\(\\varphi_{*}^{t - 1}\\) and \\(\\varphi_{*}^{t}\\) denote the activated logits of \\(\\mu_{*}^{t - 1}\\) and \\(\\mu_{*}^{t}\\), respectively. \n- Background Feature Distillation. When the network learning novel categories, not only the parameters of the novel classifier \\(f_{C^t}^t\\) are optimized, but the encoder \\(f_{e}^{t}\\) and the decoder \\(f_{d}^{t}\\) are affected as well. Merely utilizing Group Knowledge Distillation loss \\(\\mathcal{L}_{GKD}\\) is insufficient in effectively mitigating catastrophic forgetting as it primarily focuses on constraining classifiers of the old categories. However, excessively constraining parameter optimization can impede learning for novel categories. Taking all these factors into account, we designed a Background Feature Distillation (BFD) method that only acts on the features out of the regions of current categories \\(\\mathcal{R}_{C^t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.307, + 0.786, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B F D} = \\sum_ {i = 1} ^ {t - 1} \\| \\psi_ {b} ^ {t} \\cdot F _ {m ^ {i}} ^ {t} - \\psi_ {b} ^ {t} \\cdot F _ {m ^ {i}} ^ {t - 1} \\| _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.359, + 0.784, + 0.436 + ], + "angle": 0, + "content": "where \\( F_{m^i}^t \\) denotes the intermediate features in \\( f_{C^i}^t \\), \\( \\psi_b^t \\) represents the binary mask out of \\( \\mathcal{R}_{C^t} \\), and \\( \\| \\cdot \\|_2 \\) calculates the Euclidean distance between the two matrix. The Background Feature Distillation allows other losses to supervise the region of novel categories \\( \\mathcal{R}_{C^t} \\) while restricting the negative effect outside \\( \\mathcal{R}_{C^t} \\), where old categories \\( C^{1:t-1} \\) could appear." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.42, + 0.472 + ], + "angle": 0, + "content": "3.4 Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.482, + 0.78, + 0.497 + ], + "angle": 0, + "content": "With our proposed losses above, we construct our objective function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.508, + 0.785, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {P B - B C E} + \\lambda_ {1} \\cdot \\mathcal {L} _ {B g A ^ {+}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {B g A ^ {-}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {G K D} + \\lambda_ {4} \\cdot \\mathcal {L} _ {B F D}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.536, + 0.784, + 0.581 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{PB - BCE}\\) supervises the learning of novel classes, \\(\\mathcal{L}_{PB - BCE}\\) and \\(\\mathcal{L}_{BgA^{-}}\\) enhance the effect of Background Adaptation, \\(\\mathcal{L}_{GKD}\\) and \\(\\mathcal{L}_{BFD}\\) mitigate forgetting old classes. Hyperparameters \\(\\lambda_1,\\lambda_2,\\lambda_3,\\lambda_4\\) balance these constraints." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.375, + 0.621 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.427, + 0.65 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.748 + ], + "angle": 0, + "content": "Datasets. We follow the settings of most previous work and evaluate our proposed methods on Pascal VOC 2012 [23,26] and ADE20K [64] datasets. Pascal VOC 2012 is composed of 20 different classes with 10,582 images for training and 1,449 images for validation. ADE20K has 150 classes that contain 100 thing classes and 50 stuff classes. There are 20,210 images for training and 2,000 images for validation in ADE20K." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Protocols. Earlier work [6, 21] has defined two types of training settings, i.e. overlapping and disjoint. The overlapping setting contains future classes in the previous steps, whereas the disjoint setting forbids future classes until they belong to the training classes. We follow recent work [3, 41, 56, 63] and adopt the overlapping setting for evaluating as it is more realistic. Besides, following the previous training protocols [6, 21, 56], we set each sub-task from the dataset as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "A. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "\\(N_{ini} - N_{inc}\\), where \\(N_{ini}\\) and \\(N_{inc}\\) denotes the quantity of the categories in the initial step and a single incremental step, respectively. For example, the 15-1 sub-task of Pascal VOC 2012 has 15 classes for the initial training step. During each incremental step, one novel class is applied for incremental training without previous classes. Specifically, we define 4 sub-tasks on Pascal VOC 2012, including 19-1 (2 steps), 15-1 (6 steps), 10-1 (11 steps), and 5-3 (6 steps). On ADE20K, we define 4 sub-tasks containing 100-50 (2 steps), 100-10 (6 steps), 50-50 (3 steps), and 100-5 (11 steps)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate our proposed method with widely-used mean Intersection-over-Union (mIoU). The Intersection-over-Union (IoU) is defined as \\( IoU = \\frac{TP}{TP + FP + FN} \\), where \\( TP \\), \\( FP \\), and \\( FN \\) represent the quantity of true-positive, false-positive, and false-negative prediction pixels. The mIoU calculates the mean value of IoU corresponding to each category. To effectively measure the performance of CISS tasks, we compute the mIoU values of the initial classes \\( C^0 \\), the incremental classes \\( C^{1:t} \\), and all classes \\( C^{0:t} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.616 + ], + "angle": 0, + "content": "Implementation details. Following the previous studies [6, 8, 40], we apply Deeplab-v3 [12] as the segmentation network with a ResNet-101 [27] for feature extraction. The ResNet-101 is initially pretrained on ImageNet-1K [18]. We use SGD optimizer to optimize the network. The initial learning rate of the initial step is set to \\(10^{-2}\\), while in the incremental steps, the initial learning rate is set to \\(10^{-3}\\) for Pascal VOC 2012 and \\(10^{-2}\\) for ADE20K. The momentum value is 0.9 and the weight decay is \\(10^{-4}\\) in all steps. We apply the poly learning rate schedule following [3, 8, 12, 21]. During the initial step, the training epoch is set to 50 for Pascal VOC 2012 and 60 for ADE20K, while for each incremental step, we train the network for 20 epochs on Pascal VOC and 100 epochs on ADE20K. The batch size is 16 for Pascal VOC 2012 and 8 for ADE20K in all steps. The data augmentation follows the standard of previous work, which is composed of random scaling, random flipping, and random crop. The coefficients \\(\\lambda_{1}, \\lambda_{2}, \\lambda_{3}, \\lambda_{4}\\), of the objective function are set to 1, 5, 1, 4. The threshold of pseudo labeling \\(\\tau = 0.7\\) is set to all settings. The implementation is built on Pytorch. The experiments are conducted on NVIDIA RTX2080Ti GPUs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.434, + 0.651 + ], + "angle": 0, + "content": "4.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Comparison on Pascal VOC 2012. We evaluate our proposed method with current exemplar-free DeeplabV3-based methods on the Pascal VOC 2012 dataset. The sub-tasks of Pascal VOC 2012 have distinct characteristics. For example, the sub-task 19-1 has more categories in the initial step, while 5-3 has fewer classes for the initial training and more incremental steps, which is much more challenging. The results in Tab. 1 show that our proposed method outperforms other methods on novel classes and overall mIoU results of all sub-tasks. Particularly, our approach has \\(1.7\\%\\) and \\(3.0\\%\\) promotion of mIoU in more challenging sub-tasks of 10-1 and 5-3, respectively. Except for DKD [3] which refine the background class, our approach achieves at least \\(3.0\\%\\) and \\(7.0\\%\\) of promotion on mIoU Compared to the methods with a standard shared background classifier, demonstrating the effectiveness of the Background Adaptation mechanism." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Background Adaptation with Residual Modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison with state-of-the-art exemplar-free methods on Pascal VOC 2012 in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.785, + 0.384 + ], + "angle": 0, + "content": "
Method15-1 (6 steps)10-1 (11 steps)5-3 (6 steps)19-1 (1 step)
0-1516-20all0-1011-20all0-56-20all0-1920all
MiB [CVPR20] [6]38.013.532.212.213.112.657.142.546.771.222.168.9
SDR [CVPR21] [40]47.314.739.532.417.125.1---69.132.667.4
PLOP [CVPR21] [21]65.121.154.644.015.530.525.730.028.775.437.373.5
RCIL [CVPR22] [58]70.623.759.455.415.134.363.134.642.877.031.574.7
EWF [CVPR23] [56]77.732.767.071.530.351.961.742.247.777.96.774.5
AWT [WACV23] [25]59.117.249.133.218.026.061.845.950.4---
IDEC [TPAMI23] [63]77.036.567.370.746.359.167.149.054.1---
DKD [NeurIPS22] [3]78.142.769.773.146.560.469.653.558.177.841.576.0
Ours77.645.970.072.251.062.171.357.061.178.242.276.4
Joint79.872.477.478.476.477.476.977.677.477.577.077.4
" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.404, + 0.482, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.544, + 0.442, + 0.556 + ], + "angle": 0, + "content": "(a) Task 15-1 of VOC2012." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.404, + 0.756, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.544, + 0.71, + 0.556 + ], + "angle": 0, + "content": "(b) Task 10-1 of VOC2012." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.568, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Fig. 3: The step-wise mIoU comparison of our approach with previous methods under the sub-tasks 10-1 and 15-1 of Pascal VOC 2012." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Besides, our method achieve an impressive advantage on the novel classes, with significant improvements of \\(4.5\\%\\) and \\(3.5\\%\\) compared to the previous state-of-the-art method. Many previous methods emphasize restricting catastrophic forgetting by boosting the mIoU of larger-proportion old classes in specific tasks (e.g. 15-1) to make the overall mIoU appear to be higher. However, our approach not only retains the ability to segment old classes but also significantly improves the accuracy of novel classes, aligning more closely with the incremental objective emphasized in CISS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Comparison on ADE20K. We evaluate our method on the sub-tasks of more challenging ADE20K, including 100-50, 100-10, 100-5, and 50-50, as shown in Tab. 2. Although our method is slightly lower than the state-of-the-art method in short-term sub-tasks, we are remarkably ahead of other methods in more challenging long-term 100-10 and 100-5 tasks. The performance in the novel classes is still outstanding, which demonstrates the universality of our method." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison with state-of-the-art exemplar-free methods on ADE20K in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.2, + 0.803, + 0.384 + ], + "angle": 0, + "content": "
Method100-50 (2 steps)100-10 (6 steps)50-50 (3 steps)100-5 (11 steps)
0-100101-150all0-100101-150all0-5051-150all0-100101-150all
MiB [6]40.517.232.838.211.129.245.621.029.336.05.726.0
PLOP [21]41.914.932.940.513.631.648.821.030.439.17.828.8
RCIL [58]42.318.834.539.317.632.148.325.032.538.511.529.6
EWF [56]41.221.334.641.516.633.2---41.413.432.1
AWT [25]40.924.735.639.121.433.246.627.033.538.616.031.1
GSC [17]42.419.234.840.816.232.646.226.433.0---
IDEC [63]42.018.234.140.317.632.747.426.033.139.214.631.0
DKD [3]42.422.936.041.619.534.348.826.333.9---
Ours42.023.035.741.123.135.247.926.533.740.521.234.1
Joint44.328.238.944.328.238.951.133.338.944.328.238.9
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.403, + 0.428 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.789, + 0.742 + ], + "angle": 0, + "content": "Component Ablations. In this section, we analyze our proposed components, including the Background Adaptation mechanism, Background Feature Distillation loss \\(\\mathcal{L}_{BFD}\\), Group Knowledge Distillation loss \\(\\mathcal{L}_{GKD}\\), Background Adaptation losses \\(\\mathcal{L}_{BgA^{-}}\\) and \\(\\mathcal{L}_{BgA^{+}}\\). The experiments are carried out on the representative sub-task 15-1 of Pascal VOC 2012. We separately evaluate different combinations of components and present the results in Tab. 3. The baseline method applies pseudo-labeling, \\(\\mathcal{L}_{PB-BCE}\\), and freezes the whole network except the classifiers of background and novel classes. Under the same experimental settings, the Background Adaptation mechanism significantly promotes \\(5.2\\%\\) of overall mIoU, demonstrating its effectiveness in retaining the knowledge of old classes, and more notably, improving the performance of novel classes. The usage of distillation strategies with \\(\\mathcal{L}_{BFD}\\) and \\(\\mathcal{L}_{GKD}\\) preserves the performance of old classes while further boosting mIoU by \\(6.0\\%\\) on novel classes. Moreover, the precise supervision of the Background Adaptation using \\(\\mathcal{L}_{BgA^{-}}\\) and \\(\\mathcal{L}_{BgA^{+}}\\) once again enhances the segmentation of novel classes with an additional \\(9.4\\%\\) of mIoU. Subsequently, there is a growth of \\(2.4\\%\\) on the overall mIoU, which proves that additional precise supervision is necessary. The results above validate the effectiveness of the Background Adaptation mechanism and other related losses. There is a notable improvement, particularly in the performance of novel classes, exceeding twice that of the baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Background Adaptation Schemes. Sec. 4.3 has presented the importance of the Background Adaptation losses. We have devised multiple schemes to achieve more optimal results of Background Adaptation, as shown in Tab. 4. Initially, we aim to have the Background Adaptation value for the current ground truth background region equal to 0, using Mean Square Error (MSE) loss to minimize" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Background Adaptation with Residual Modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 3: Ablation Study on the components of our proposed method on 15-1. BgA denotes Background Adaptation. In particular, the experiments \\( w / o \\) distillation losses (i.e. \\( \\mathcal{L}_{BFD} \\) and \\( \\mathcal{L}_{GKD} \\)) freeze the whole network except the classifier \\( f^t \\) following the settings in [8]. All experiments contain \\( \\mathcal{L}_{PB-BCE} \\)." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.213, + 0.669, + 0.326 + ], + "angle": 0, + "content": "
BgA\\( \\mathcal{L}_{BFD} \\)\\( \\mathcal{L}_{GKD} \\)\\( \\mathcal{L}_{BgA^{-}} \\)\\( \\mathcal{L}_{BgA^{+}} \\)mIoUnimIoUincmIoU
73.421.861.1
✓77.430.566.3
✓✓76.737.167.3
✓✓✓77.336.567.6
✓✓✓✓77.143.469.0
✓✓✓✓76.634.166.5
✓✓✓✓✓77.645.970.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.345, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Table 4: Ablation study on schemes of Table 5: Ablation Study on the feature Background Adaptation. distillation (FD) strategies." + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.385, + 0.455, + 0.464 + ], + "angle": 0, + "content": "
FiltermIoUinimIoUincmIoU
MSE-0X74.241.866.5
✓74.642.266.9
BCE-1✓77.444.769.6
OursX65.745.460.8
✓77.645.970.0
" + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.386, + 0.744, + 0.464 + ], + "angle": 0, + "content": "
mIoUinimIoUincmIoU
w/o FD56.933.251.2
LKD70.138.362.6
LMSE77.544.369.6
LBFD77.645.970.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.683 + ], + "angle": 0, + "content": "the deviation from 0 (MSE-0). However, MSE-0 couldn't restrict the background value to absolute 0, leading to noise accumulation and a lower mIoU compared to Background Adaptation \\( w/ \\) knowledge distillation. Later, we incorporate considerations for the current class region and introduce a Binary Cross-Entropy (BCE) loss (BCE-1). The expectation is to have the value in the current class region as small as possible, while the value in the background class region is as large as possible. Besides, we apply the Filter to remove the noise above 0. The BCE-1 scheme with Filter improved mIoU by \\( 2\\% \\), but supervision in the background region tends to overfit to background objects. Therefore, we retain the BCE loss in the novel class region and redesign a triplet loss as in Sec. 3.3. The final Background Adaptation losses further enhance the segmentation capability for novel classes, resulting in a \\( 1.2\\% \\) improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Background Feature Distillation. To validate the effectiveness of Background Feature Distillation loss \\(\\mathcal{L}_{BFD}\\) in Sec. 3.3, we compare \\(\\mathcal{L}_{BFD}\\) with the method \\(w / o\\) feature distillation (\\(w / o\\) FD), \\(w/\\) standard knowledge distillation (\\(\\mathcal{L}_{KD}\\)), and \\(w/\\) mean square error (\\(\\mathcal{L}_{MSE}\\)). As shown in Tab. 5, our \\(\\mathcal{L}_{BFD}\\) outperforms other representative distillation methods. Particularly, the Background distillation with MSE, i.e., \\(\\mathcal{L}_{BFD}\\), shows a \\(1.6\\%\\) promotion of mIoU on novel classes with no decline in old classes compared to \\(\\mathcal{L}_{MSE}\\). This demonstrates the necessity of preserving a tunnel for learning novel classes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.144, + 0.737, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.445, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Fig. 4: Qualitative analysis on 15-1 of Pascal VOC 2012. The Baseline and Ours \\( w/o \\) BgA losses are in the \\( 1^{st} \\) row and the \\( 4^{th} \\) row of Tab. 3. Ours applies additional Background Adaptation losses, compared to the Ours \\( w/o \\) BgA losses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.431, + 0.529 + ], + "angle": 0, + "content": "4.4 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.643 + ], + "angle": 0, + "content": "We've conducted a qualitative analysis of our approach with Baseline and Ours \\( w / o \\) BgA losses in the \\( 1^{st} \\) row and the \\( 4^{th} \\) row of Tab. 3. The results shown in Fig. 4 are visualized from the 15-1 sub-task of VOC 2012. Our approach in the last three columns reduces the misclassification (e.g., train in the first two columns of each method) and enhances the precision (e.g., sofa in the last column of each method) of the novel classes, while maintaining segmentation ability on the old classes. The results showcase the stability and plasticity of our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.37, + 0.677 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We have introduced a novel CISS method featuring a Background Adaptation mechanism, designed to prevent the continual shifting of a shared background class and promote the learning of novel classes. The mechanism explicitly models the background residual rather than the background at each step and aggregates these residuals to represent the evolving background. To further enhance the performance of the Background Adaptation, we customized corresponding losses on the adaptation channel and improved the distillation scheme. Experiments on our proposed components verify their effectiveness under the Background Adaptation mechanism. Moreover, our method surpasses previous state-of-the-art methods in most sub-tasks, notably excelling in performance of novel classes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Background Adaptation with Residual Modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.237 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. 61972036) and the Industry-University-Institute Cooperation Foundation of the Eighth Research Institute of China Aerospace Science and Technology Corporation (No. SAST2022-049)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.317 + ], + "angle": 0, + "content": "1. Ahn, H., Cha, S., Lee, D., Moon, T.: Uncertainty-based continual learning with adaptive regularization. NeurIPS 32 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.345 + ], + "angle": 0, + "content": "2. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. NeurIPS 32 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.385 + ], + "angle": 0, + "content": "3. Baek, D., Oh, Y., Lee, S., Lee, J., Ham, B.: Decomposed knowledge distillation for class-incremental semantic segmentation. NeurIPS 35, 10380-10392 (2022) 2, 4, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.785, + 0.414 + ], + "angle": 0, + "content": "4. Borsos, Z., Mutny, M., Krause, A.: Coresets via bilevel optimization for continual learning and streaming. NeurIPS 33, 14879-14890 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.455 + ], + "angle": 0, + "content": "5. Boschini, M., Bonicelli, L., Porrello, A., Bellitto, G., Pennisi, M., Palazzo, S., Spampinato, C., Calderara, S.: Transfer without forgetting. In: ECCV. pp. 692-709 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.785, + 0.495 + ], + "angle": 0, + "content": "6. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: CVPR. pp. 9233-9242 (2020) 2, 4, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.523 + ], + "angle": 0, + "content": "7. Cha, H., Lee, J., Shin, J.: Co2l: Contrastive continual learning. In: ICCV. pp. 9516-9525 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.564 + ], + "angle": 0, + "content": "8. Cha, S., Yoo, Y., Moon, T., et al.: Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. NeurIPS 34, 10919-10930 (2021) 2, 3, 4, 6, 7, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.606 + ], + "angle": 0, + "content": "9. Chen, J., Cong, R., Luo, Y., Ip, H., Kwong, S.: Saving 100x storage: Prototype replay for reconstructing training sample distribution in class-incremental semantic segmentation. Advances in Neural Information Processing Systems 36 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.607, + 0.785, + 0.647 + ], + "angle": 0, + "content": "10. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062 (2014) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.648, + 0.785, + 0.689 + ], + "angle": 0, + "content": "11. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI 40(4), 834-848 (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.69, + 0.785, + 0.716 + ], + "angle": 0, + "content": "12. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017) 1, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "13. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: ECCV. pp. 801-818 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.71, + 0.771 + ], + "angle": 0, + "content": "14. Chen, Z., Liu, B.: Lifelong machine learning, vol. 1. Springer (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "15. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "16. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. NeurIPS 34, 17864-17875 (2021) 1" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Cong, W., Cong, Y., Dong, J., Sun, G., Ding, H.: Gradient-semantic compensation for incremental semantic segmentation. arXiv preprint arXiv:2307.10822 (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "18. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.233 + ], + "angle": 0, + "content": "19. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: CVPR. pp. 5138-5146 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.273 + ], + "angle": 0, + "content": "20. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.316 + ], + "angle": 0, + "content": "21. Douillard, A., Chen, Y., Dapogny, A., Cord, M.: Plop: Learning without forgetting for continual semantic segmentation. In: CVPR. pp. 4040-4050 (2021) 2, 3, 4, 6, 7, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.345 + ], + "angle": 0, + "content": "22. Ermis, B., Zappella, G., Wistuba, M., Rawal, A., Archambeau, C.: Memory efficient continual learning with transformers. NeurIPS 35, 10629-10642 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.373 + ], + "angle": 0, + "content": "23. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. IJCV 88, 303-338 (2010) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.415 + ], + "angle": 0, + "content": "24. Fang, Y., Zhu, F., Cheng, B., Liu, L., Zhao, Y., Wei, Y.: Locating noise is halfway denoising for semi-supervised segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16612-16622 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.471 + ], + "angle": 0, + "content": "25. Goswami, D., Schuster, R., van de Weijer, J., Stricker, D.: Attribution-aware weight transfer: A warm-start initialization for class-incremental semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3195-3204 (2023) 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.5 + ], + "angle": 0, + "content": "26. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: ICCV. pp. 991-998 (2011) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.528 + ], + "angle": 0, + "content": "27. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) 1, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.556 + ], + "angle": 0, + "content": "28. Jung, H., Ju, J., Jung, M., Kim, J.: Less-forgetting learning in deep neural networks. arXiv preprint arXiv:1607.00122 (2016) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.785, + 0.599 + ], + "angle": 0, + "content": "29. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13), 3521-3526 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.627 + ], + "angle": 0, + "content": "30. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. NeurIPS 25 (2012) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.655 + ], + "angle": 0, + "content": "31. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.785, + 0.684 + ], + "angle": 0, + "content": "32. Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI 40(12), 2935-2947 (2017) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.726 + ], + "angle": 0, + "content": "33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012-10022 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.754 + ], + "angle": 0, + "content": "34. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.785, + 0.783 + ], + "angle": 0, + "content": "35. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. NeurIPS 30 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.811 + ], + "angle": 0, + "content": "36. Madaan, D., Yoon, J., Li, Y., Liu, Y., Hwang, S.J.: Representational continuity for unsupervised continual learning. In: ICLR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "37. Maracani, A., Michieli, U., Toldo, M., Zanuttigh, P.: Recall: Replay-based continual learning in semantic segmentation. In: ICCV. pp. 7026-7035 (2021) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Background Adaptation with Residual Modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "38. McClelland, J.L., McNaughton, B.L., O'Reilly, R.C.: Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review 102(3), 419 (1995) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "39. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.248, + 0.787, + 0.288 + ], + "angle": 0, + "content": "40. Michieli, U., Zanuttigh, P.: Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In: CVPR. pp. 1114-1124 (2021) 2, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.289, + 0.787, + 0.317 + ], + "angle": 0, + "content": "41. Oh, Y., Baek, D., Ham, B.: Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. NeurIPS 35, 14516-14528 (2022) 3, 4, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.318, + 0.787, + 0.359 + ], + "angle": 0, + "content": "42. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: CVPR. pp. 11321-11329 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.36, + 0.787, + 0.388 + ], + "angle": 0, + "content": "43. Pham, Q., Liu, C., Hoi, S.: Dualnet: Continual learning, fast and slow. NeurIPS 34, 16131-16144 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.389, + 0.787, + 0.417 + ], + "angle": 0, + "content": "44. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.418, + 0.787, + 0.445 + ], + "angle": 0, + "content": "45. Ritter, H., Botev, A., Barber, D.: Online structured laplace approximations for overcoming catastrophic forgetting. NeurIPS 31 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.446, + 0.787, + 0.488 + ], + "angle": 0, + "content": "46. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 234-241 (2015) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.488, + 0.787, + 0.543 + ], + "angle": 0, + "content": "47. Schwarz, J., Czarnecki, W., Luketina, J., Grabska-Barwinska, A., Teh, Y.W., Pascanu, R., Hadsell, R.: Progress & compress: A scalable framework for continual learning. In: Proc. International Conference on Machine Learning. pp. 4528-4537. PMLR (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.545, + 0.787, + 0.6 + ], + "angle": 0, + "content": "48. Shang, C., Li, H., Meng, F., Wu, Q., Qiu, H., Wang, L.: Incrementer: Transformer for class-incremental semantic segmentation with knowledge distillation focusing on old class. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7214-7224 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.601, + 0.787, + 0.628 + ], + "angle": 0, + "content": "49. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. NeurIPS 30 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.629, + 0.787, + 0.657 + ], + "angle": 0, + "content": "50. Shon, H., Lee, J., Kim, S.H., Kim, J.: Dlcft: Deep linear continual fine-tuning for general incremental learning. In: ECCV. pp. 513-529 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.658, + 0.787, + 0.685 + ], + "angle": 0, + "content": "51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.686, + 0.787, + 0.728 + ], + "angle": 0, + "content": "52. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.729, + 0.787, + 0.77 + ], + "angle": 0, + "content": "53. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: Proc. International Conference on Machine Learning. pp. 10347-10357 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.771, + 0.787, + 0.798 + ], + "angle": 0, + "content": "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "55. Wu, C., Herranz, L., Liu, X., Van De Weijer, J., Raducanu, B., et al.: Memory replay gans: Learning to generate new categories without forgetting. NeurIPS 31 (2018) 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "56. Xiao, J.W., Zhang, C.B., Feng, J., Liu, X., van de Weijer, J., Cheng, M.M.: Endpoints weight fusion for class incremental semantic segmentation. In: CVPR. pp. 7204-7213 (2023) 4, 9, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "57. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Nabi, M., Alameda-Pineda, X., Ricci, E.: Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE TPAMI 45(2), 2567-2581 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "58. Zhang, C.B., Xiao, J.W., Liu, X., Chen, Y.C., Cheng, M.M.: Representation compensation networks for continual semantic segmentation. In: CVPR. pp. 7053-7064 (2022) 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.325 + ], + "angle": 0, + "content": "59. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19148-19158 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "60. Zhang, J.O., Sax, A., Zamir, A., Guibas, L., Malik, J.: Side-tuning: a baseline for network adaptation via additive side networks. In: ECCV. pp. 698-714 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.396 + ], + "angle": 0, + "content": "61. Zhang, Z., Gao, G., Fang, Z., Jiao, J., Wei, Y.: Mining unseen classes via regional objectness: A simple baseline for incremental segmentation. NeurIPS 35, 24340-24353 (2022) 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "62. Zhang, Z., Gao, G., Jiao, J., Liu, C.H., Wei, Y.: Coinseg: Contrast inter-and intra-class representations for incremental segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 843-853 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.48 + ], + "angle": 0, + "content": "63. Zhao, D., Yuan, B., Shi, Z.: Inherit with distillation and evolve with contrast: Exploring class incremental semantic segmentation without exemplar memory. IEEE TPAMI (2023) 4, 9, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "64. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: CVPR. pp. 633-641 (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.563 + ], + "angle": 0, + "content": "65. Zhu, H., Wei, Y., Liang, X., Zhang, C., Zhao, Y.: Ctp: Towards vision-language continual pretraining via compatible momentum contrast and topology preservation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22257-22267 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "66. Zhu, L., Chen, T., Yin, J., See, S., Liu, J.: Continual semantic segmentation with automatic memory sample selection. In: CVPR. pp. 3082-3092 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_origin.pdf b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4456da7b704e77225ae3d2d877064a3db2e71cde --- /dev/null +++ b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/0e537afc-06b2-4d04-8f8e-b30eb0f0595c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31664bf17e1b3025eca296ebee28bf8e0c0a9aedde89b2b642e1fbb743a3a8a6 +size 2154217 diff --git a/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/full.md b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6fcf53734dfc5b928e64627138aef993e7e24df6 --- /dev/null +++ b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/full.md @@ -0,0 +1,365 @@ +# Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation + +Anqi Zhang and Guangyu Gao\* + +School of Computer Science and Technology, + +Beijing Institute of Technology, Beijing, China + +andy_zaq@outlook.com + +guangyugao@bit.edu.cn + +Abstract. Class Incremental Semantic Segmentation (CISS), within Incremental Learning for semantic segmentation, targets segmenting new categories while reducing the catastrophic forgetting on the old categories. Besides, background shifting, where the background category changes constantly in each step, is a special challenge for CISS. Current methods with a shared background classifier struggle to keep up with these changes, leading to decreased stability in background predictions and reduced accuracy of segmentation. For this special challenge, we designed a novel background adaptation mechanism, which explicitly models the background residual rather than the background itself in each step, and aggregates these residuals to represent the evolving background. Therefore, the background adaptation mechanism ensures the stability of previous background classifiers, while enabling the model to concentrate on the easy-learned residuals from the additional channel, which enhances background discernment for better prediction of novel categories. To precisely optimize the background adaptation mechanism, we propose Pseudo Background Binary Cross-Entropy loss and Background Adaptation losses, which amplify the adaptation effect. Group Knowledge Distillation and Background Feature Distillation strategies are designed to prevent forgetting old categories. Our approach, evaluated across various incremental scenarios on Pascal VOC 2012 and ADE20K datasets, outperforms prior exemplar-free state-of-the-art methods with mIoU of $3.0\%$ in VOC 10-1 and $2.0\%$ in ADE 100-5, notably enhancing the accuracy of new classes while mitigating catastrophic forgetting. Code is available in https://andyzaq.github.io/barmsite/. + +# 1 Introduction + +With the development of deep learning and the support of convolutional neural networks (CNNs) [27,30,31,51] and Transformers [20,33,53,54], current semantic segmentation networks [10-13,15,16,24,34,46] have achieved remarkable results on standard supervised learning tasks. However, real-world data, arriving as an + +unstable continuous stream, poses the challenge of retaining prior knowledge while accommodating new concepts, termed as catastrophic forgetting [38, 39]. + +To tackle this problem, incremental learning is proposed to adapt to changing data streams for new concepts, but also avoid forgetting old knowledge, especially for the classification task, i.e., Class-Incremental Learning (CIL) [14,28,32,59,65]. Cermelli et al. [6] introduced CIL into Semantic Segmentation as Class Incremental Semantic Segmentation (CISS). Since the semantic segmentation tasks involve dense predictions, the issue of catastrophic forgetting typically becomes more challenging. Most recent works [6,8,21] have struggled to mitigate catastrophic forgetting through knowledge distillation [3,21,40] or parameter freezing [8,61]. Although these strategies effectively alleviate catastrophic forgetting, it comes at the cost of the model's plasticity, making it challenging to learn novel classes. In current state-of-the-art approaches [3,61], the performance of established old classes approaches the upper limit of joint training, yet there still remains a significant gap in mIoU for novel classes. However, in the context of lifelong learning scenarios featuring an infinite influx of novel classes, the incremental learner's plasticity becomes increasingly crucial. Therefore, the optimal solution should be fine-tuning more parameters for novel classes, while catastrophic forgetting needs to be handled properly. + +![](images/f1a97bce118baed3dfb84f220a1bb8d7b7c02e7f289ad416a81c6f26cfa2cd09.jpg) + +![](images/52291a177dbe141e77873c4a765f55fb624edf986307970b19b2913f3ce63fe4.jpg) + +![](images/780f1a4c66c6324deaca6da641e912d66cb9013e421fd1f5712ab0723899e31b.jpg) + +![](images/1bfa3e4687e4bb266c521ac63d31e170a4597d701dfdfff6729bbe3e78f01c46.jpg) +Image + +![](images/54abe5470d97ff2fcb271e85f9335835839a86f12f9d2710bd17e12f9f9c6d63.jpg) + +![](images/9c923c8c6b31da1df51e481aa2dec6f2c7838bd6f635ddb8b215848e43d99bf4.jpg) + +![](images/ba8ea4ef1691fa9293e1aed119de98eb4790862f0c633259f4e9c1966235bc92.jpg) + +![](images/50a424ac6880efea796179f10a94248efcea1528be35847f48da90f067f77de9.jpg) +Prediction of step $t - 1$ + +![](images/925175c29e97205e711d865ef9b4745345856c901a63aa2aa40dd768635df746.jpg) + +![](images/0cadcbab6f0839911eafc8a5f265cb9a1c521be89736643d9befd226e4742d09.jpg) + +![](images/d449847987a191e2b729857e4a7096aa660651c52b115b1e250d18b4eda0f75b.jpg) + +![](images/9a5e34868501b1c68574ae3f3fec5ab60266504d55d154b3e6d02cf9dc9a7f21.jpg) +Bg logits of step $t - 1$ + +![](images/49d8fe702ce2bf0ac94e5d316896ee6ac13d5fd8b749305a88d4e131f64eaa53.jpg) + +![](images/8848eb1b153c2e2cc16ff4a40576b1edb9e00940ab58a7c6023ba537cb07f995.jpg) + +![](images/908428f9671c0ded36545d3fc3bb4a8d533438f5fa1cb0cac83a6f5123ab7db3.jpg) + +![](images/73776b77fed59101a8a1357dceb0e4e4803f080cb003a4e6d493409d3bb6171d.jpg) +Bg Adaptation of step $t$ +Fig. 1: 3D Visualization of Background Adaptation results. The background logits for step $t$ combine those from step $t - 1$ with the adaptation logits learned in step $t$ , preventing disorderly adjustments and focusing on residuals. The current step background logits are processed using the Sigmoid function. Note that colors range from red (small values) to blue (large values). + +![](images/7ba554256d6202cf91461801b388df15e9e13a712691e63778f88a8dbd910a43.jpg) + +![](images/dfad77d6c94cd1ea015d31579029aee42b99cc1299eeef0f0fec7d2bd9f6a384.jpg) + +![](images/99c7d18dca28bc05b9f75bfd01b480310c277fe64fd9b6e7e5cf09aee87e9bbb.jpg) + +![](images/5d25eb747cc238b802967016b7a200872de0fd744b5e29de0b37f7f21225f55d.jpg) +Bg logits of step $t$ (Sigmoid) + +![](images/0213258c00ff83766b4feb3341ab120259e859d57b62324e427bf191ef1dc28c.jpg) + +![](images/67829c7c655502e26655348e53d001d0aabe2efe0dbeb63e59a0e7fe5b63afec.jpg) + +![](images/2b02acc1185ac69de183ed6526314247802a9a173b1313e8698099bb7a60dcb6.jpg) + +![](images/464261f58dfc47ccb59e8eddb657715410f9c07ee1e878d0d037d52691a34286.jpg) +Prediction of step t + +![](images/3abc89263b7272830f9a538f6e653f4a0cad8274b0f00d22a09748ac9cf6ee65.jpg) + +![](images/d74ed47d3da7515ea88233b0488d9379d4cd5bbe9e6d76a73f25863e81913f67.jpg) + +![](images/819b40936f42cf055bc5906169a1c5ab18bc3098957d3ce7f2a9e62974fea72b.jpg) + +![](images/a1ed7d4703a540e65796bf292c82d4de1fc4a292d7cac88f8baf23334b8be20d.jpg) +Categories in step t + +Nevertheless, background shifting, where the background category changes constantly in each step, is another challenge for CISS. Upon further investiga + +tion, it was observed that previous methods within CISS often utilize a shared background classifier across each step. However, novel classes are typically considered as the background in previous learning steps, indicating that the background class undergoes continuous changes throughout incremental learning. This way results in an incessantly shifting background in each step, leading to disruption in background prediction and inaccurate learning of novel classes. Some approaches [8,9,41] store old exemplars in extra memory space to prevent background shifting, while long-term storage of such data may not be feasible in real-world scenarios. To enhance the incremental nature of the CISS method and reduce background instability, we revise the representation of the background class and propose an exemplar-free background adaptation mechanism. In particular, the background adaptation mechanism explicitly models the background residual rather than the background itself in each incremental step. The learned residuals of the additional channel are aggregated with the previously learned background logits to represent the evolving background, as illustrated in Fig. 1. By employing the background adaptation mechanism, we only need to optimize the additional channel of residuals for constructing the background during the incremental steps. This approach eliminates the need for repeatedly learning background classifiers associated with old classes, thereby avoiding the instability of optimizing a shared background classifier. + +Moreover, we customize our training approaches in order to fit the training process of the background adaptation structure and promote its performance. We first introduce the widely used pseudo-labeling [8,21] to generate a pseudobackground, and then we design a PB-BCE loss to focus on optimizing the classifier of novel classes and related background adaptation residual. Besides, for the Background Adaptation channel, the objective is to minimize the values in the regions of novel categories, while ensuring that the values in other regions remain above 0. Thus, two auxiliary Background Adaptation losses are specifically applied to the background adaptation channel for precise adjustment. Given our pursuit of plasticity on novel categories without external memory, we have devised more comprehensive exemplar-free knowledge distillation strategies to combat catastrophic forgetting. We further adopt knowledge distillation strategies on the logit maps of old categories and the intermediate features, namely the Group Knowledge Distillation and Background Feature Distillation, respectively. Finally, our approach outperforms previous methods in multiple benchmarks, especially in realistic and challenging long-term scenarios such as 10-1 of VOC2012 and 100-5 of ADE20K. Our approach achieves significant performance gains of $4.5\%$ and $5.2\%$ on novel classes, respectively, compared to the previous state-of-the-art. + +Overall, our contributions are summarized as follows: + +- We propose a Background Adaptation strategy in incremental steps by optimizing an additional channel for residual modeling instead of the shared background classifier, which avoids unstable optimization and prevents chaos in background logits. + +- We design PB-BCE loss and Background Adaptation losses to enhance the effectiveness of Background Adaptation, while two distillation strategies ensure maintaining knowledge of old categories. + +- Experiments validate the effectiveness of our proposed method, showcasing a new state-of-the-art performance on CISS benchmarks. + +# 2 Related Works + +# 2.1 Class Incremental Learning + +Class Incremental Learning (CIL) breaks the constraints of the standard training process in an incremental form on new classes. CIL task aims to continually learn the new classes without retraining on the data of old classes, which significantly reduces the training cost. As a compromise, the CIL task has to effectively solve the problem of catastrophic forgetting [38, 39]. Current CIL methods are divided into several main categories. Replay-based methods recall the data of old classes during incremental training. Some of the approaches [2, 4, 35] reserve memory buffers and temporarily store representative past data. Others [42, 49, 55] introduce powerful generative models to generate pseudo-data for rehearsal. Regularization-based methods introduce regularization terms to prevent forgetting the old classes. These methods utilize loss functions as penalties [1, 29, 45, 47] or knowledge distillation from the previous model [19, 32, 44]. Representation-based methods attempt to transfer representations to certain categories, which apply adaptation on a fixed pre-trained backbone [5, 22, 50, 60] or self-supervised learning [7, 36, 43]. + +# 2.2 Class Incremental Semantic Segmentation + +The achievements in CIL tasks encourage the Incremental Learning pipeline to develop in other dense prediction tasks, including Semantic Segmentation. Class Incremental Semantic Segmentation (CISS) task is first introduced in MiB [6], which focused on reconstructing the background region under the restriction from knowledge distillation. Since then, various methods [57, 63] have been introduced in the CISS task. PLOP [21] defined the pseudo-labeling and multi-scale local distillation to remember the old categories. SDR [40] introduced category-wise prototype matching and contrastive learning methods to enlarge the difference between categories. RCIL [58] and EWF [56] explored methods for integrating parameters of new and old models. DKD [3] decomposed the class logits into positive and negative logits for better knowledge distillation and applied threshold to generate background regions. Incrementer [48] borrows prompt learning from ViT [20] and applies the structure of Segmenter [52] instead of DeeplabV3 [12], which is the basic structure of other methods. Besides, several approaches utilize external memory [66] or models [37, 62] for better performance. SSUL [8] introduced a Saliency-map detector to identify the saliency region and store some of the previous instances in memory. MicroSeg [61] utilized proposals from pretrained Mask2Former [15] as a guidance. ALIFE [41] memorized category features and adapted for remembering the old classes. + +![](images/779f2a42ff61ffb57b6df982cc03cef055be9ba6e9995a83b67d19385e26245b.jpg) +(a) The overall pipeline of our approach. + +![](images/c2e6f0a7d7c85d9a9fcce5af4d17978e190e2324486eaa9048eff006192d01bc.jpg) +(b) $\mathrm{BgA}$ losses. +Fig. 2: Overview of our framework that consists of a Background Adaptation mechanism and proposed losses. (a) The Background Adaptation mechanism uses the negative part of the adaptation channel to rectify the background logits. (b) Two Background Adaptation losses separately optimize the regions $w/$ and $w/o$ novel categories to enhance the performance of the Background Adaptation. + +# 3 Methods + +# 3.1 Preliminaries + +The Class-Incremental Semantic Segmentation (CISS) task performs training in $T$ steps. During step $t$ , the training data $D^{t}$ containing the labels of the novel classes $C^{t}$ are applied for training, while the labels of old classes $C^{1:t-1}$ are expunged. In particular, each sample of the training data $D^{t}$ consists of an RGB image $x^{t} \in \mathbb{R}^{3 \times H \times W}$ and a corresponding ground truth label $y^{t}$ on novel classes $C^{t}$ . Besides, the ground truth label $y^{t}$ replaces the labels of the $C^{1:t-1}$ with the background label $c_{b}$ as the old classes will not appear in $C^{t}$ to ensure $C^{1:t-1} \cap C^{t} = \emptyset$ . Therefore, the model $f^{t}$ being trained at step $t$ not only needs to learn to segment novel categories but also has to retain the capability to segment old categories. However, while learning the novel categories, the model $f^{t}$ tends to overfit the novel categories $C^{t}$ as $f^{t}$ cannot contact with samples of $C^{1:t-1}$ , causing severe catastrophic forgetting. Moreover, the novel categories can only appear in the background region during the previous steps. We are more likely to classify them from the previous background prediction region, which motivates us to focus on background adaptation. + +# 3.2 Background Adaptation + +Previous studies often apply only one background classifier. They have to readjust the parameters of the background classifier whenever the model learns novel categories, which causes catastrophic forgotten even if the classifier of the old categories remains the same probability output. The unstable background classifier + +could be a thorny constraint factor. To tackle the problems above, we design a Background Adaptation mechanism. Different from the previous distinct background classifier, we introduce an additional background adaptation channel for each classifier of the incremental process. + +During the training process of incremental step $t$ , the current model $f^t$ inherits the parameters from the previous model $f^{t-1}$ and adds a new classifier $f_{C^t}^t$ for parallel classification of categories in $C^t$ alongside the previous classifiers. The encoder $f_e^t$ and the decoder $f_d^t$ transform the input image $x^t$ into features map $F_d^t \in \mathbb{R}^{h \times w \times c}$ : + +$$ +F _ {d} ^ {t} = f _ {d} ^ {t} \left(f _ {e} ^ {t} \left(x ^ {t}\right)\right). \tag {1} +$$ + +Compared to the standard classifier of the previous methods [8, 21, 41], our classifier of each incremental step has an additional channel for background adaptation. Therefore, for each group of categories $C^i$ from step $i$ , the classifier $f_{C^i}^t$ of the model $f^t$ generates the corresponding classification logit maps $\mu_{C^i \cup \{b_i\}}^t \in \mathbb{R}^{h \times w \times (|C^i| + 1)}$ : + +$$ +\mu_ {C ^ {i} \cup \{b _ {i} \}} ^ {t} = f _ {C ^ {i}} ^ {t} \left(F _ {d} ^ {t}\right), \tag {2} +$$ + +where $i \in \{1, 2, \ldots, t\}$ . We define $b^1$ to represent the background class of the initial step, and $b^i$ denotes the background adaptation for subsequent incremental steps when $i >= 2$ . The symbol $b$ refers to the adapted background class. Thus, $\mu_{C^i \cup \{b_i\}}^t$ contains the probability maps $\mu_{C^i}$ of $|C^i|$ categories from step $i$ and an additional channel $\mu_{b^i}^t$ for constructing the background logits $\mu_b^t$ . Considering that the region of the novel categories $C^t$ is more likely inherited from the background region of the previous step, the classifier $f_{C^t}^t$ adjusts the background logits through $\mu_{b^t}^t$ , which decreases the probability values in the region $\mathcal{R}_{C^t}$ of $C^t$ . The parameters of previous classifiers are not required to continuously optimize for new background regions, thereby avoiding instability in the classification of old categories. + +The background adaptation channel $\mu_{b^t}^t$ can correct the background logits $\mu_b^t\in \mathbb{R}^{h\times w\times 1}$ in $\mathcal{R}_{C^t}$ , yet the noise in other regions of $\mu_{b^t}^t$ may interfere with the background prediction. Particularly, when old categories are present in the image, there is a risk that these regions could be incorrectly covered by the adapted background. Hence, we design a Filter operation that replaces positive values of $\mu_{b^i}^t$ with 0. As shown in Fig. 2a, with the Filter operation applied, the residual modeling of background logits during the inference process is as follows: + +$$ +\mu_ {b} ^ {t} = \mu_ {b ^ {1}} ^ {t} + \sum_ {i = 2} ^ {t} F i l t e r \left(\mu_ {b ^ {i}} ^ {t}\right), \tag {3} +$$ + +where the background logits from the initial step $\mu_{b^1}^t$ stay the same and the background adaptation logits from the incremental steps $\mu_{b^i}^t$ only remain the negative values via the Filter. In particular, to make $\mu_{b}^{t}$ more friendly for gradient backpropagation, we abandoned the Filter on the background adaptation logits + +of novel classes $\mu_{b^t}^t$ during the training process: + +$$ +\mu_ {b} ^ {t} = \mu_ {b ^ {1}} ^ {t} + \sum_ {i = 2} ^ {t - 1} \operatorname {F i l t e r} \left(\mu_ {b ^ {i}} ^ {t}\right) + \mu_ {b ^ {t}} ^ {t}. \tag {4} +$$ + +# 3.3 Incremental Training Strategy + +Pseudo Labeling. During the incremental steps, even if objects of old categories $C^{1:t-1}$ exist in the image, we cannot obtain corresponding labels from the ground truth. In order to simulate a ground truth label that contains the categories we have learned, we follow [8,21,61] to introduce the prediction probability of the previous step $\varphi^{t-1} \in \mathbb{R}^{h \times w \times |C^{1:t-1}|}$ from $f^{t-1}(x^t)$ with Sigmoid function and combine it with the current ground truth label $y^t$ : + +$$ +\hat {y} ^ {t} = \left\{ \begin{array}{l l} y ^ {t}, & \left(y ^ {t} \in C ^ {t}\right) \vee \left\{\left(y ^ {t} \in c _ {b}\right) \wedge \left(\max \left(\varphi^ {t - 1}\right) < \tau\right) \right\} \\ \tilde {y} ^ {t - 1}, & \left(y ^ {t} \in c _ {b}\right) \wedge \left(\max \left(\varphi^ {t - 1}\right) \geq \tau\right) \end{array} \right. \tag {5} +$$ + +where $\hat{y}^t\in \mathbb{R}^{h\times w}$ denotes the generated pseudo label and $\tilde{y}^{t - 1} = \arg \max (\varphi^{t - 1})$ represents the final prediction from $f^{t - 1}(x^t)$ . + +PB-BCE Loss. Binary Cross-Entropy (BCE) loss is extensively employed for learning to segment objects of novel categories. We follow the usage of the Sigmoid function and activate $\mu^t$ as $\varphi^t$ . Some of the CISS methods [8,21,61] utilize the pseudo mask as the ground truth for supervision, which effectively replaces the background label in the regions of the old categories. However, the ambiguous supervision of the old categories could improperly optimize the classifiers of $C^{1:t - 1}$ and destroy their identification capability. Therefore, we design a Pseudo Background BCE (PB-BCE) loss which cuts off the backpropagation of old categories, meaning that we only take the labels of current categories $\hat{y}_{C^t}^t$ and pseudo-background $\hat{y}_{c_b}^t$ for optimization: + +$$ +\mathcal {L} _ {P B - B C E} = - \frac {1}{h w} \sum_ {k = 1} ^ {h w} \sum_ {c \in \{c _ {b} \} \cap C ^ {t}} C E \left(\varphi_ {c, k} ^ {t}, \hat {\psi} _ {c, k} ^ {t}\right), \tag {6} +$$ + +where $k$ represents the index of pixel and $\hat{\psi}_c^t$ denotes the binary map of class c in pseudo map $\hat{y}^t$ . $CE$ is the cross-entropy calculation: + +$$ +C E \left(\varphi_ {*} ^ {t}, \hat {\psi} _ {*} ^ {t}\right) = \hat {\psi} _ {*} ^ {t} \cdot \log \varphi_ {*} ^ {t} + \left(1 - \hat {\psi} _ {*} ^ {t}\right) \cdot \log \left(1 - \varphi_ {*} ^ {t}\right). \tag {7} +$$ + +Although the whole background prediction logits $\mu_b^t$ are involved in the loss calculation, only the background adaptation channel for current categories $\mu_{b^t}^t$ serves for optimization. Previous background logits are detached to prevent disorderly evolving on parameters of the old classifiers. + +Background Adaptation Losses. The Background Adaptation (BgA) losses, as shown in Fig. 2b, are designed for further refined adjustments on the background adaptation channel $\mu_{b^t}^t$ . We propose two losses for two different regions of $\mu_{b^t}^t$ according to the target effect. For regions $\mathcal{R}_{C^t}$ belonging to the current classes $C^t$ , the ideal target is to have negative values as small as possible. To achieve this, we introduce cross-entropy loss $\mathcal{L}_{BgA^{+}}$ for supervision: + +$$ +\mathcal {L} _ {B g A ^ {+}} = - \frac {1}{| \mathcal {R} _ {C ^ {t}} |} \sum_ {k \in \mathcal {R} _ {C ^ {t}}} \log \frac {\exp \left(- \mu_ {b ^ {t} , k} ^ {t}\right)}{1 + \exp \left(- \mu_ {b ^ {t} , k} ^ {t}\right)}, \tag {8} +$$ + +where $|\mathcal{R}_{C^t}|$ represents the number of pixels in $\mathcal{R}_{C^t}$ . Besides, as for the regions of the background that might contain old classes, we hope the value of $\mu_{bt}^{t}$ could be 0 for maintaining the historical background prediction result. However, achieving precise control over the whole region to be a precise value is challenging and may be impractical. Therefore, we set the positive logits in the adaptation channel of the old categories to 0 in Eq. 4 to avoid the noise and apply a Triplet loss $\mathcal{L}_{BgA^{-}}$ to supervise the background region: + +$$ +\mathcal {L} _ {B g A ^ {-}} = \frac {1}{h w - \left| \mathcal {R} _ {C ^ {t}} \right|} \sum_ {k \notin \mathcal {R} _ {C ^ {t}}} \max \left[ 0, \left(1 - \varphi_ {b ^ {t}, k} ^ {t}\right) ^ {2} - \left(\varphi_ {b ^ {t}, k} ^ {t} - 0\right) ^ {2} \right]. \tag {9} +$$ + +We first activate the background adaptation channel as $\varphi_{b^t}^t$ via Sigmoid function, then the logits are mapped to the range from 0 to 1. We aim for the logits $\mu_{b^t}^t$ of this region to be greater than 0 while not overfitting to the objects in the background, meaning that values of $\varphi_{b^t}^t$ lie within the range of 0.5 to 1 and not close to 1. In particular, The Triplet loss $\mathcal{L}_{BgA^{-}}$ respectively set 0 and 1 as negative and positive anchors. We simultaneously constrain the loss not to go below 0 by setting a lower bound. Therefore, once the value of the pixel exceeds 0.5, i.e., the middle of the positive and negative anchors, there will be no more gradient on the pixel for backpropagation, which mitigates overfitting to the objects in the background. + +Knowledge Distillation Strategies. In the training process of incremental steps for CISS tasks, substantial alterations to the existing parameters of $f^t$ frequently occur, giving rise to situations of catastrophic forgetting. Guided by the previous methods, we have devised a series of knowledge distillation strategies tailored to our background adaptation mechanism. + +- Group Knowledge Distillation. In Sec. 3.2, we have introduced the classification logit map $\mu_{C^i\cup \{b_i\}}^t$ for each group of categories $C^i$ , which contains an additional channel for background adaptation. We introduce a Group Knowledge Distillation loss $\mathcal{L}_{GKD}$ to ensure the stability of the old category logits, as well as the previous background adaptation channel: + +$$ +\mathcal {L} _ {G K D} = - \frac {1}{h w} \sum_ {k = 1} ^ {h w} \sum_ {i = 1} ^ {t - 1} \sum_ {c \in \{c _ {b} ^ {i} \} \cup C ^ {i}} C E \left(\varphi_ {c, k} ^ {t}, \varphi_ {c, k} ^ {t - 1}\right) \tag {10} +$$ + +where $\varphi_{*}^{t - 1}$ and $\varphi_{*}^{t}$ denote the activated logits of $\mu_{*}^{t - 1}$ and $\mu_{*}^{t}$ , respectively. +- Background Feature Distillation. When the network learning novel categories, not only the parameters of the novel classifier $f_{C^t}^t$ are optimized, but the encoder $f_{e}^{t}$ and the decoder $f_{d}^{t}$ are affected as well. Merely utilizing Group Knowledge Distillation loss $\mathcal{L}_{GKD}$ is insufficient in effectively mitigating catastrophic forgetting as it primarily focuses on constraining classifiers of the old categories. However, excessively constraining parameter optimization can impede learning for novel categories. Taking all these factors into account, we designed a Background Feature Distillation (BFD) method that only acts on the features out of the regions of current categories $\mathcal{R}_{C^t}$ : + +$$ +\mathcal {L} _ {B F D} = \sum_ {i = 1} ^ {t - 1} \| \psi_ {b} ^ {t} \cdot F _ {m ^ {i}} ^ {t} - \psi_ {b} ^ {t} \cdot F _ {m ^ {i}} ^ {t - 1} \| _ {2} ^ {2}, \tag {11} +$$ + +where $F_{m^i}^t$ denotes the intermediate features in $f_{C^i}^t$ , $\psi_b^t$ represents the binary mask out of $\mathcal{R}_{C^t}$ , and $\| \cdot \|_2$ calculates the Euclidean distance between the two matrix. The Background Feature Distillation allows other losses to supervise the region of novel categories $\mathcal{R}_{C^t}$ while restricting the negative effect outside $\mathcal{R}_{C^t}$ , where old categories $C^{1:t-1}$ could appear. + +# 3.4 Objective Function + +With our proposed losses above, we construct our objective function as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {P B - B C E} + \lambda_ {1} \cdot \mathcal {L} _ {B g A ^ {+}} + \lambda_ {2} \cdot \mathcal {L} _ {B g A ^ {-}} + \lambda_ {3} \cdot \mathcal {L} _ {G K D} + \lambda_ {4} \cdot \mathcal {L} _ {B F D}, \tag {12} +$$ + +where $\mathcal{L}_{PB - BCE}$ supervises the learning of novel classes, $\mathcal{L}_{PB - BCE}$ and $\mathcal{L}_{BgA^{-}}$ enhance the effect of Background Adaptation, $\mathcal{L}_{GKD}$ and $\mathcal{L}_{BFD}$ mitigate forgetting old classes. Hyperparameters $\lambda_1,\lambda_2,\lambda_3,\lambda_4$ balance these constraints. + +# 4 Experiments + +# 4.1 Experimental Setup + +Datasets. We follow the settings of most previous work and evaluate our proposed methods on Pascal VOC 2012 [23,26] and ADE20K [64] datasets. Pascal VOC 2012 is composed of 20 different classes with 10,582 images for training and 1,449 images for validation. ADE20K has 150 classes that contain 100 thing classes and 50 stuff classes. There are 20,210 images for training and 2,000 images for validation in ADE20K. + +Protocols. Earlier work [6, 21] has defined two types of training settings, i.e. overlapping and disjoint. The overlapping setting contains future classes in the previous steps, whereas the disjoint setting forbids future classes until they belong to the training classes. We follow recent work [3, 41, 56, 63] and adopt the overlapping setting for evaluating as it is more realistic. Besides, following the previous training protocols [6, 21, 56], we set each sub-task from the dataset as + +$N_{ini} - N_{inc}$ , where $N_{ini}$ and $N_{inc}$ denotes the quantity of the categories in the initial step and a single incremental step, respectively. For example, the 15-1 sub-task of Pascal VOC 2012 has 15 classes for the initial training step. During each incremental step, one novel class is applied for incremental training without previous classes. Specifically, we define 4 sub-tasks on Pascal VOC 2012, including 19-1 (2 steps), 15-1 (6 steps), 10-1 (11 steps), and 5-3 (6 steps). On ADE20K, we define 4 sub-tasks containing 100-50 (2 steps), 100-10 (6 steps), 50-50 (3 steps), and 100-5 (11 steps). + +Evaluation metrics. We evaluate our proposed method with widely-used mean Intersection-over-Union (mIoU). The Intersection-over-Union (IoU) is defined as $IoU = \frac{TP}{TP + FP + FN}$ , where $TP$ , $FP$ , and $FN$ represent the quantity of true-positive, false-positive, and false-negative prediction pixels. The mIoU calculates the mean value of IoU corresponding to each category. To effectively measure the performance of CISS tasks, we compute the mIoU values of the initial classes $C^0$ , the incremental classes $C^{1:t}$ , and all classes $C^{0:t}$ , respectively. + +Implementation details. Following the previous studies [6, 8, 40], we apply Deeplab-v3 [12] as the segmentation network with a ResNet-101 [27] for feature extraction. The ResNet-101 is initially pretrained on ImageNet-1K [18]. We use SGD optimizer to optimize the network. The initial learning rate of the initial step is set to $10^{-2}$ , while in the incremental steps, the initial learning rate is set to $10^{-3}$ for Pascal VOC 2012 and $10^{-2}$ for ADE20K. The momentum value is 0.9 and the weight decay is $10^{-4}$ in all steps. We apply the poly learning rate schedule following [3, 8, 12, 21]. During the initial step, the training epoch is set to 50 for Pascal VOC 2012 and 60 for ADE20K, while for each incremental step, we train the network for 20 epochs on Pascal VOC and 100 epochs on ADE20K. The batch size is 16 for Pascal VOC 2012 and 8 for ADE20K in all steps. The data augmentation follows the standard of previous work, which is composed of random scaling, random flipping, and random crop. The coefficients $\lambda_{1}, \lambda_{2}, \lambda_{3}, \lambda_{4}$ , of the objective function are set to 1, 5, 1, 4. The threshold of pseudo labeling $\tau = 0.7$ is set to all settings. The implementation is built on Pytorch. The experiments are conducted on NVIDIA RTX2080Ti GPUs. + +# 4.2 Quantitative Results + +Comparison on Pascal VOC 2012. We evaluate our proposed method with current exemplar-free DeeplabV3-based methods on the Pascal VOC 2012 dataset. The sub-tasks of Pascal VOC 2012 have distinct characteristics. For example, the sub-task 19-1 has more categories in the initial step, while 5-3 has fewer classes for the initial training and more incremental steps, which is much more challenging. The results in Tab. 1 show that our proposed method outperforms other methods on novel classes and overall mIoU results of all sub-tasks. Particularly, our approach has $1.7\%$ and $3.0\%$ promotion of mIoU in more challenging sub-tasks of 10-1 and 5-3, respectively. Except for DKD [3] which refine the background class, our approach achieves at least $3.0\%$ and $7.0\%$ of promotion on mIoU Compared to the methods with a standard shared background classifier, demonstrating the effectiveness of the Background Adaptation mechanism. + +Table 1: Quantitative comparison with state-of-the-art exemplar-free methods on Pascal VOC 2012 in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best. + +
Method15-1 (6 steps)10-1 (11 steps)5-3 (6 steps)19-1 (1 step)
0-1516-20all0-1011-20all0-56-20all0-1920all
MiB [CVPR20] [6]38.013.532.212.213.112.657.142.546.771.222.168.9
SDR [CVPR21] [40]47.314.739.532.417.125.1---69.132.667.4
PLOP [CVPR21] [21]65.121.154.644.015.530.525.730.028.775.437.373.5
RCIL [CVPR22] [58]70.623.759.455.415.134.363.134.642.877.031.574.7
EWF [CVPR23] [56]77.732.767.071.530.351.961.742.247.777.96.774.5
AWT [WACV23] [25]59.117.249.133.218.026.061.845.950.4---
IDEC [TPAMI23] [63]77.036.567.370.746.359.167.149.054.1---
DKD [NeurIPS22] [3]78.142.769.773.146.560.469.653.558.177.841.576.0
Ours77.645.970.072.251.062.171.357.061.178.242.276.4
Joint79.872.477.478.476.477.476.977.677.477.577.077.4
+ +![](images/ca7fb1225a0e710d7c023d178794f6b71262ccd0421ce0eb2cd34b11e94413a9.jpg) +(a) Task 15-1 of VOC2012. + +![](images/d6a7066a6e6348b8d583e1d7eb9a596af3caabffe512fba29c05505e365ca214.jpg) +(b) Task 10-1 of VOC2012. +Fig. 3: The step-wise mIoU comparison of our approach with previous methods under the sub-tasks 10-1 and 15-1 of Pascal VOC 2012. + +Besides, our method achieve an impressive advantage on the novel classes, with significant improvements of $4.5\%$ and $3.5\%$ compared to the previous state-of-the-art method. Many previous methods emphasize restricting catastrophic forgetting by boosting the mIoU of larger-proportion old classes in specific tasks (e.g. 15-1) to make the overall mIoU appear to be higher. However, our approach not only retains the ability to segment old classes but also significantly improves the accuracy of novel classes, aligning more closely with the incremental objective emphasized in CISS. + +Comparison on ADE20K. We evaluate our method on the sub-tasks of more challenging ADE20K, including 100-50, 100-10, 100-5, and 50-50, as shown in Tab. 2. Although our method is slightly lower than the state-of-the-art method in short-term sub-tasks, we are remarkably ahead of other methods in more challenging long-term 100-10 and 100-5 tasks. The performance in the novel classes is still outstanding, which demonstrates the universality of our method. + +Table 2: Quantitative comparison with state-of-the-art exemplar-free methods on ADE20K in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best. + +
Method100-50 (2 steps)100-10 (6 steps)50-50 (3 steps)100-5 (11 steps)
0-100101-150all0-100101-150all0-5051-150all0-100101-150all
MiB [6]40.517.232.838.211.129.245.621.029.336.05.726.0
PLOP [21]41.914.932.940.513.631.648.821.030.439.17.828.8
RCIL [58]42.318.834.539.317.632.148.325.032.538.511.529.6
EWF [56]41.221.334.641.516.633.2---41.413.432.1
AWT [25]40.924.735.639.121.433.246.627.033.538.616.031.1
GSC [17]42.419.234.840.816.232.646.226.433.0---
IDEC [63]42.018.234.140.317.632.747.426.033.139.214.631.0
DKD [3]42.422.936.041.619.534.348.826.333.9---
Ours42.023.035.741.123.135.247.926.533.740.521.234.1
Joint44.328.238.944.328.238.951.133.338.944.328.238.9
+ +# 4.3 Ablation Studies + +Component Ablations. In this section, we analyze our proposed components, including the Background Adaptation mechanism, Background Feature Distillation loss $\mathcal{L}_{BFD}$ , Group Knowledge Distillation loss $\mathcal{L}_{GKD}$ , Background Adaptation losses $\mathcal{L}_{BgA^{-}}$ and $\mathcal{L}_{BgA^{+}}$ . The experiments are carried out on the representative sub-task 15-1 of Pascal VOC 2012. We separately evaluate different combinations of components and present the results in Tab. 3. The baseline method applies pseudo-labeling, $\mathcal{L}_{PB-BCE}$ , and freezes the whole network except the classifiers of background and novel classes. Under the same experimental settings, the Background Adaptation mechanism significantly promotes $5.2\%$ of overall mIoU, demonstrating its effectiveness in retaining the knowledge of old classes, and more notably, improving the performance of novel classes. The usage of distillation strategies with $\mathcal{L}_{BFD}$ and $\mathcal{L}_{GKD}$ preserves the performance of old classes while further boosting mIoU by $6.0\%$ on novel classes. Moreover, the precise supervision of the Background Adaptation using $\mathcal{L}_{BgA^{-}}$ and $\mathcal{L}_{BgA^{+}}$ once again enhances the segmentation of novel classes with an additional $9.4\%$ of mIoU. Subsequently, there is a growth of $2.4\%$ on the overall mIoU, which proves that additional precise supervision is necessary. The results above validate the effectiveness of the Background Adaptation mechanism and other related losses. There is a notable improvement, particularly in the performance of novel classes, exceeding twice that of the baseline. + +Background Adaptation Schemes. Sec. 4.3 has presented the importance of the Background Adaptation losses. We have devised multiple schemes to achieve more optimal results of Background Adaptation, as shown in Tab. 4. Initially, we aim to have the Background Adaptation value for the current ground truth background region equal to 0, using Mean Square Error (MSE) loss to minimize + +Table 3: Ablation Study on the components of our proposed method on 15-1. BgA denotes Background Adaptation. In particular, the experiments $w / o$ distillation losses (i.e. $\mathcal{L}_{BFD}$ and $\mathcal{L}_{GKD}$ ) freeze the whole network except the classifier $f^t$ following the settings in [8]. All experiments contain $\mathcal{L}_{PB-BCE}$ . + +
BgA\( \mathcal{L}_{BFD} \)\( \mathcal{L}_{GKD} \)\( \mathcal{L}_{BgA^{-}} \)\( \mathcal{L}_{BgA^{+}} \)mIoUnimIoUincmIoU
73.421.861.1
✓77.430.566.3
✓✓76.737.167.3
✓✓✓77.336.567.6
✓✓✓✓77.143.469.0
✓✓✓✓76.634.166.5
✓✓✓✓✓77.645.970.0
+ +Table 4: Ablation study on schemes of Table 5: Ablation Study on the feature Background Adaptation. distillation (FD) strategies. + +
FiltermIoUinimIoUincmIoU
MSE-0X74.241.866.5
✓74.642.266.9
BCE-1✓77.444.769.6
OursX65.745.460.8
✓77.645.970.0
+ +
mIoUinimIoUincmIoU
w/o FD56.933.251.2
LKD70.138.362.6
LMSE77.544.369.6
LBFD77.645.970.0
+ +the deviation from 0 (MSE-0). However, MSE-0 couldn't restrict the background value to absolute 0, leading to noise accumulation and a lower mIoU compared to Background Adaptation $w/$ knowledge distillation. Later, we incorporate considerations for the current class region and introduce a Binary Cross-Entropy (BCE) loss (BCE-1). The expectation is to have the value in the current class region as small as possible, while the value in the background class region is as large as possible. Besides, we apply the Filter to remove the noise above 0. The BCE-1 scheme with Filter improved mIoU by $2\%$ , but supervision in the background region tends to overfit to background objects. Therefore, we retain the BCE loss in the novel class region and redesign a triplet loss as in Sec. 3.3. The final Background Adaptation losses further enhance the segmentation capability for novel classes, resulting in a $1.2\%$ improvement. + +Background Feature Distillation. To validate the effectiveness of Background Feature Distillation loss $\mathcal{L}_{BFD}$ in Sec. 3.3, we compare $\mathcal{L}_{BFD}$ with the method $w / o$ feature distillation ( $w / o$ FD), $w/$ standard knowledge distillation ( $\mathcal{L}_{KD}$ ), and $w/$ mean square error ( $\mathcal{L}_{MSE}$ ). As shown in Tab. 5, our $\mathcal{L}_{BFD}$ outperforms other representative distillation methods. Particularly, the Background distillation with MSE, i.e., $\mathcal{L}_{BFD}$ , shows a $1.6\%$ promotion of mIoU on novel classes with no decline in old classes compared to $\mathcal{L}_{MSE}$ . This demonstrates the necessity of preserving a tunnel for learning novel classes. + +![](images/8ce6c0bc2e05004791d7289fed033d4fbefd6fdae47a6344232430c3d2348247.jpg) +Fig. 4: Qualitative analysis on 15-1 of Pascal VOC 2012. The Baseline and Ours $w/o$ BgA losses are in the $1^{st}$ row and the $4^{th}$ row of Tab. 3. Ours applies additional Background Adaptation losses, compared to the Ours $w/o$ BgA losses. + +# 4.4 Qualitative Analysis + +We've conducted a qualitative analysis of our approach with Baseline and Ours $w / o$ BgA losses in the $1^{st}$ row and the $4^{th}$ row of Tab. 3. The results shown in Fig. 4 are visualized from the 15-1 sub-task of VOC 2012. Our approach in the last three columns reduces the misclassification (e.g., train in the first two columns of each method) and enhances the precision (e.g., sofa in the last column of each method) of the novel classes, while maintaining segmentation ability on the old classes. The results showcase the stability and plasticity of our approach. + +# 5 Conclusions + +We have introduced a novel CISS method featuring a Background Adaptation mechanism, designed to prevent the continual shifting of a shared background class and promote the learning of novel classes. The mechanism explicitly models the background residual rather than the background at each step and aggregates these residuals to represent the evolving background. To further enhance the performance of the Background Adaptation, we customized corresponding losses on the adaptation channel and improved the distillation scheme. Experiments on our proposed components verify their effectiveness under the Background Adaptation mechanism. Moreover, our method surpasses previous state-of-the-art methods in most sub-tasks, notably excelling in performance of novel classes. + +# Acknowledgement + +This work was supported by the National Natural Science Foundation of China (No. 61972036) and the Industry-University-Institute Cooperation Foundation of the Eighth Research Institute of China Aerospace Science and Technology Corporation (No. SAST2022-049). + +# References + +1. Ahn, H., Cha, S., Lee, D., Moon, T.: Uncertainty-based continual learning with adaptive regularization. NeurIPS 32 (2019) 4 +2. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. NeurIPS 32 (2019) 4 +3. Baek, D., Oh, Y., Lee, S., Lee, J., Ham, B.: Decomposed knowledge distillation for class-incremental semantic segmentation. NeurIPS 35, 10380-10392 (2022) 2, 4, 9, 10, 11, 12 +4. Borsos, Z., Mutny, M., Krause, A.: Coresets via bilevel optimization for continual learning and streaming. NeurIPS 33, 14879-14890 (2020) 4 +5. Boschini, M., Bonicelli, L., Porrello, A., Bellitto, G., Pennisi, M., Palazzo, S., Spampinato, C., Calderara, S.: Transfer without forgetting. In: ECCV. pp. 692-709 (2022) 4 +6. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: CVPR. pp. 9233-9242 (2020) 2, 4, 9, 10, 11, 12 +7. Cha, H., Lee, J., Shin, J.: Co2l: Contrastive continual learning. In: ICCV. pp. 9516-9525 (2021) 4 +8. Cha, S., Yoo, Y., Moon, T., et al.: Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. NeurIPS 34, 10919-10930 (2021) 2, 3, 4, 6, 7, 10, 13 +9. Chen, J., Cong, R., Luo, Y., Ip, H., Kwong, S.: Saving 100x storage: Prototype replay for reconstructing training sample distribution in class-incremental semantic segmentation. Advances in Neural Information Processing Systems 36 (2024) 3 +10. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062 (2014) 1 +11. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI 40(4), 834-848 (2017) 1 +12. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017) 1, 4, 10 +13. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: ECCV. pp. 801-818 (2018) 1 +14. Chen, Z., Liu, B.: Lifelong machine learning, vol. 1. Springer (2018) 2 +15. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022) 1, 4 +16. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. NeurIPS 34, 17864-17875 (2021) 1 + +17. Cong, W., Cong, Y., Dong, J., Sun, G., Ding, H.: Gradient-semantic compensation for incremental semantic segmentation. arXiv preprint arXiv:2307.10822 (2023) 12 +18. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009) 10 +19. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: CVPR. pp. 5138-5146 (2019) 4 +20. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 1, 4 +21. Douillard, A., Chen, Y., Dapogny, A., Cord, M.: Plop: Learning without forgetting for continual semantic segmentation. In: CVPR. pp. 4040-4050 (2021) 2, 3, 4, 6, 7, 9, 10, 11, 12 +22. Ermis, B., Zappella, G., Wistuba, M., Rawal, A., Archambeau, C.: Memory efficient continual learning with transformers. NeurIPS 35, 10629-10642 (2022) 4 +23. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. IJCV 88, 303-338 (2010) 9 +24. Fang, Y., Zhu, F., Cheng, B., Liu, L., Zhao, Y., Wei, Y.: Locating noise is halfway denoising for semi-supervised segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16612-16622 (2023) 1 +25. Goswami, D., Schuster, R., van de Weijer, J., Stricker, D.: Attribution-aware weight transfer: A warm-start initialization for class-incremental semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3195-3204 (2023) 11, 12 +26. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: ICCV. pp. 991-998 (2011) 9 +27. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) 1, 10 +28. Jung, H., Ju, J., Jung, M., Kim, J.: Less-forgetting learning in deep neural networks. arXiv preprint arXiv:1607.00122 (2016) 2 +29. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13), 3521-3526 (2017) 4 +30. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. NeurIPS 25 (2012) 1 +31. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) 1 +32. Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI 40(12), 2935-2947 (2017) 2, 4 +33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012-10022 (2021) 1 +34. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015) 1 +35. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. NeurIPS 30 (2017) 4 +36. Madaan, D., Yoon, J., Li, Y., Liu, Y., Hwang, S.J.: Representational continuity for unsupervised continual learning. In: ICLR (2022) 4 +37. Maracani, A., Michieli, U., Toldo, M., Zanuttigh, P.: Recall: Replay-based continual learning in semantic segmentation. In: ICCV. pp. 7026-7035 (2021) 4 + +38. McClelland, J.L., McNaughton, B.L., O'Reilly, R.C.: Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review 102(3), 419 (1995) 2, 4 +39. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989) 2, 4 +40. Michieli, U., Zanuttigh, P.: Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In: CVPR. pp. 1114-1124 (2021) 2, 4, 10, 11 +41. Oh, Y., Baek, D., Ham, B.: Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. NeurIPS 35, 14516-14528 (2022) 3, 4, 6, 9 +42. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: CVPR. pp. 11321-11329 (2019) 4 +43. Pham, Q., Liu, C., Hoi, S.: Dualnet: Continual learning, fast and slow. NeurIPS 34, 16131-16144 (2021) 4 +44. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 4 +45. Ritter, H., Botev, A., Barber, D.: Online structured laplace approximations for overcoming catastrophic forgetting. NeurIPS 31 (2018) 4 +46. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 234-241 (2015) 1 +47. Schwarz, J., Czarnecki, W., Luketina, J., Grabska-Barwinska, A., Teh, Y.W., Pascanu, R., Hadsell, R.: Progress & compress: A scalable framework for continual learning. In: Proc. International Conference on Machine Learning. pp. 4528-4537. PMLR (2018) 4 +48. Shang, C., Li, H., Meng, F., Wu, Q., Qiu, H., Wang, L.: Incrementer: Transformer for class-incremental semantic segmentation with knowledge distillation focusing on old class. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7214-7224 (2023) 4 +49. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. NeurIPS 30 (2017) 4 +50. Shon, H., Lee, J., Kim, S.H., Kim, J.: Dlcft: Deep linear continual fine-tuning for general incremental learning. In: ECCV. pp. 513-529 (2022) 4 +51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) 1 +52. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021) 4 +53. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: Proc. International Conference on Machine Learning. pp. 10347-10357 (2021) 1 +54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) 1 +55. Wu, C., Herranz, L., Liu, X., Van De Weijer, J., Raducanu, B., et al.: Memory replay gans: Learning to generate new categories without forgetting. NeurIPS 31 (2018) 4 + +56. Xiao, J.W., Zhang, C.B., Feng, J., Liu, X., van de Weijer, J., Cheng, M.M.: Endpoints weight fusion for class incremental semantic segmentation. In: CVPR. pp. 7204-7213 (2023) 4, 9, 11, 12 +57. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Nabi, M., Alameda-Pineda, X., Ricci, E.: Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE TPAMI 45(2), 2567-2581 (2022) 4 +58. Zhang, C.B., Xiao, J.W., Liu, X., Chen, Y.C., Cheng, M.M.: Representation compensation networks for continual semantic segmentation. In: CVPR. pp. 7053-7064 (2022) 4, 11, 12 +59. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19148-19158 (2023) 2 +60. Zhang, J.O., Sax, A., Zamir, A., Guibas, L., Malik, J.: Side-tuning: a baseline for network adaptation via additive side networks. In: ECCV. pp. 698-714 (2020) 4 +61. Zhang, Z., Gao, G., Fang, Z., Jiao, J., Wei, Y.: Mining unseen classes via regional objectness: A simple baseline for incremental segmentation. NeurIPS 35, 24340-24353 (2022) 2, 4, 7 +62. Zhang, Z., Gao, G., Jiao, J., Liu, C.H., Wei, Y.: Coinseg: Contrast inter-and intra-class representations for incremental segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 843-853 (2023) 4 +63. Zhao, D., Yuan, B., Shi, Z.: Inherit with distillation and evolve with contrast: Exploring class incremental semantic segmentation without exemplar memory. IEEE TPAMI (2023) 4, 9, 11, 12 +64. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: CVPR. pp. 633-641 (2017) 9 +65. Zhu, H., Wei, Y., Liang, X., Zhang, C., Zhao, Y.: Ctp: Towards vision-language continual pretraining via compatible momentum contrast and topology preservation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22257-22267 (2023) 2 +66. Zhu, L., Chen, T., Yin, J., See, S., Liu, J.: Continual semantic segmentation with automatic memory sample selection. In: CVPR. pp. 3082-3092 (2023) 4 \ No newline at end of file diff --git a/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/images.zip b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..71028fb0102eea268a450f37749651a3ee56ae23 --- /dev/null +++ b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a896143f38ec22c47c0ada6f730bbd4b18c5f97af5b1c38b6e3d687024b6470 +size 530479 diff --git a/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/layout.json b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6dd76ee9ee7fe286b469148dec10ca5962cb7c8a --- /dev/null +++ b/2024/Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation/layout.json @@ -0,0 +1,12605 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "type": "text", + "content": "Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 229, + 186, + 386, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 186, + 386, + 198 + ], + "spans": [ + { + "bbox": [ + 229, + 186, + 386, + 198 + ], + "type": "text", + "content": "Anqi Zhang and Guangyu Gao\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 211, + 208, + 402, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 208, + 402, + 220 + ], + "spans": [ + { + "bbox": [ + 211, + 208, + 402, + 220 + ], + "type": "text", + "content": "School of Computer Science and Technology," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 220, + 402, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 220, + 402, + 230 + ], + "spans": [ + { + "bbox": [ + 211, + 220, + 402, + 230 + ], + "type": "text", + "content": "Beijing Institute of Technology, Beijing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 232, + 356, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 232, + 356, + 241 + ], + "spans": [ + { + "bbox": [ + 258, + 232, + 356, + 241 + ], + "type": "text", + "content": "andy_zaq@outlook.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 256, + 242, + 357, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 242, + 357, + 252 + ], + "spans": [ + { + "bbox": [ + 256, + 242, + 357, + 252 + ], + "type": "text", + "content": "guangyugao@bit.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 281, + 462, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 281, + 462, + 555 + ], + "spans": [ + { + "bbox": [ + 159, + 281, + 462, + 555 + ], + "type": "text", + "content": "Abstract. Class Incremental Semantic Segmentation (CISS), within Incremental Learning for semantic segmentation, targets segmenting new categories while reducing the catastrophic forgetting on the old categories. Besides, background shifting, where the background category changes constantly in each step, is a special challenge for CISS. Current methods with a shared background classifier struggle to keep up with these changes, leading to decreased stability in background predictions and reduced accuracy of segmentation. For this special challenge, we designed a novel background adaptation mechanism, which explicitly models the background residual rather than the background itself in each step, and aggregates these residuals to represent the evolving background. Therefore, the background adaptation mechanism ensures the stability of previous background classifiers, while enabling the model to concentrate on the easy-learned residuals from the additional channel, which enhances background discernment for better prediction of novel categories. To precisely optimize the background adaptation mechanism, we propose Pseudo Background Binary Cross-Entropy loss and Background Adaptation losses, which amplify the adaptation effect. Group Knowledge Distillation and Background Feature Distillation strategies are designed to prevent forgetting old categories. Our approach, evaluated across various incremental scenarios on Pascal VOC 2012 and ADE20K datasets, outperforms prior exemplar-free state-of-the-art methods with mIoU of " + }, + { + "bbox": [ + 159, + 281, + 462, + 555 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 159, + 281, + 462, + 555 + ], + "type": "text", + "content": " in VOC 10-1 and " + }, + { + "bbox": [ + 159, + 281, + 462, + 555 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 159, + 281, + 462, + 555 + ], + "type": "text", + "content": " in ADE 100-5, notably enhancing the accuracy of new classes while mitigating catastrophic forgetting. Code is available in https://andyzaq.github.io/barmsite/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 574, + 230, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 230, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 230, + 586 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "type": "text", + "content": "With the development of deep learning and the support of convolutional neural networks (CNNs) [27,30,31,51] and Transformers [20,33,53,54], current semantic segmentation networks [10-13,15,16,24,34,46] have achieved remarkable results on standard supervised learning tasks. However, real-world data, arriving as an" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "text", + "content": "* Corresponding Author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "unstable continuous stream, poses the challenge of retaining prior knowledge while accommodating new concepts, termed as catastrophic forgetting [38, 39]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 344 + ], + "type": "text", + "content": "To tackle this problem, incremental learning is proposed to adapt to changing data streams for new concepts, but also avoid forgetting old knowledge, especially for the classification task, i.e., Class-Incremental Learning (CIL) [14,28,32,59,65]. Cermelli et al. [6] introduced CIL into Semantic Segmentation as Class Incremental Semantic Segmentation (CISS). Since the semantic segmentation tasks involve dense predictions, the issue of catastrophic forgetting typically becomes more challenging. Most recent works [6,8,21] have struggled to mitigate catastrophic forgetting through knowledge distillation [3,21,40] or parameter freezing [8,61]. Although these strategies effectively alleviate catastrophic forgetting, it comes at the cost of the model's plasticity, making it challenging to learn novel classes. In current state-of-the-art approaches [3,61], the performance of established old classes approaches the upper limit of joint training, yet there still remains a significant gap in mIoU for novel classes. However, in the context of lifelong learning scenarios featuring an infinite influx of novel classes, the incremental learner's plasticity becomes increasingly crucial. Therefore, the optimal solution should be fine-tuning more parameters for novel classes, while catastrophic forgetting needs to be handled properly." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 365, + 193, + 407 + ], + "blocks": [ + { + "bbox": [ + 151, + 365, + 193, + 407 + ], + "lines": [ + { + "bbox": [ + 151, + 365, + 193, + 407 + ], + "spans": [ + { + "bbox": [ + 151, + 365, + 193, + 407 + ], + "type": "image", + "image_path": "f1a97bce118baed3dfb84f220a1bb8d7b7c02e7f289ad416a81c6f26cfa2cd09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 151, + 409, + 193, + 451 + ], + "blocks": [ + { + "bbox": [ + 151, + 409, + 193, + 451 + ], + "lines": [ + { + "bbox": [ + 151, + 409, + 193, + 451 + ], + "spans": [ + { + "bbox": [ + 151, + 409, + 193, + 451 + ], + "type": "image", + "image_path": "52291a177dbe141e77873c4a765f55fb624edf986307970b19b2913f3ce63fe4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 151, + 452, + 192, + 494 + ], + "blocks": [ + { + "bbox": [ + 151, + 452, + 192, + 494 + ], + "lines": [ + { + "bbox": [ + 151, + 452, + 192, + 494 + ], + "spans": [ + { + "bbox": [ + 151, + 452, + 192, + 494 + ], + "type": "image", + "image_path": "780f1a4c66c6324deaca6da641e912d66cb9013e421fd1f5712ab0723899e31b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 151, + 495, + 192, + 536 + ], + "blocks": [ + { + "bbox": [ + 151, + 495, + 192, + 536 + ], + "lines": [ + { + "bbox": [ + 151, + 495, + 192, + 536 + ], + "spans": [ + { + "bbox": [ + 151, + 495, + 192, + 536 + ], + "type": "image", + "image_path": "1bfa3e4687e4bb266c521ac63d31e170a4597d701dfdfff6729bbe3e78f01c46.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 538, + 181, + 545 + ], + "lines": [ + { + "bbox": [ + 164, + 538, + 181, + 545 + ], + "spans": [ + { + "bbox": [ + 164, + 538, + 181, + 545 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 196, + 367, + 237, + 408 + ], + "blocks": [ + { + "bbox": [ + 196, + 367, + 237, + 408 + ], + "lines": [ + { + "bbox": [ + 196, + 367, + 237, + 408 + ], + "spans": [ + { + "bbox": [ + 196, + 367, + 237, + 408 + ], + "type": "image", + "image_path": "54abe5470d97ff2fcb271e85f9335835839a86f12f9d2710bd17e12f9f9c6d63.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 196, + 409, + 237, + 451 + ], + "blocks": [ + { + "bbox": [ + 196, + 409, + 237, + 451 + ], + "lines": [ + { + "bbox": [ + 196, + 409, + 237, + 451 + ], + "spans": [ + { + "bbox": [ + 196, + 409, + 237, + 451 + ], + "type": "image", + "image_path": "9c923c8c6b31da1df51e481aa2dec6f2c7838bd6f635ddb8b215848e43d99bf4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 196, + 452, + 237, + 495 + ], + "blocks": [ + { + "bbox": [ + 196, + 452, + 237, + 495 + ], + "lines": [ + { + "bbox": [ + 196, + 452, + 237, + 495 + ], + "spans": [ + { + "bbox": [ + 196, + 452, + 237, + 495 + ], + "type": "image", + "image_path": "ba8ea4ef1691fa9293e1aed119de98eb4790862f0c633259f4e9c1966235bc92.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 196, + 495, + 237, + 536 + ], + "blocks": [ + { + "bbox": [ + 196, + 495, + 237, + 536 + ], + "lines": [ + { + "bbox": [ + 196, + 495, + 237, + 536 + ], + "spans": [ + { + "bbox": [ + 196, + 495, + 237, + 536 + ], + "type": "image", + "image_path": "50a424ac6880efea796179f10a94248efcea1528be35847f48da90f067f77de9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 538, + 233, + 552 + ], + "lines": [ + { + "bbox": [ + 198, + 538, + 233, + 552 + ], + "spans": [ + { + "bbox": [ + 198, + 538, + 233, + 552 + ], + "type": "text", + "content": "Prediction of step " + }, + { + "bbox": [ + 198, + 538, + 233, + 552 + ], + "type": "inline_equation", + "content": "t - 1" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 242, + 367, + 280, + 408 + ], + "blocks": [ + { + "bbox": [ + 242, + 367, + 280, + 408 + ], + "lines": [ + { + "bbox": [ + 242, + 367, + 280, + 408 + ], + "spans": [ + { + "bbox": [ + 242, + 367, + 280, + 408 + ], + "type": "image", + "image_path": "925175c29e97205e711d865ef9b4745345856c901a63aa2aa40dd768635df746.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 242, + 409, + 280, + 451 + ], + "blocks": [ + { + "bbox": [ + 242, + 409, + 280, + 451 + ], + "lines": [ + { + "bbox": [ + 242, + 409, + 280, + 451 + ], + "spans": [ + { + "bbox": [ + 242, + 409, + 280, + 451 + ], + "type": "image", + "image_path": "0cadcbab6f0839911eafc8a5f265cb9a1c521be89736643d9befd226e4742d09.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 242, + 452, + 280, + 495 + ], + "blocks": [ + { + "bbox": [ + 242, + 452, + 280, + 495 + ], + "lines": [ + { + "bbox": [ + 242, + 452, + 280, + 495 + ], + "spans": [ + { + "bbox": [ + 242, + 452, + 280, + 495 + ], + "type": "image", + "image_path": "d449847987a191e2b729857e4a7096aa660651c52b115b1e250d18b4eda0f75b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 242, + 495, + 280, + 536 + ], + "blocks": [ + { + "bbox": [ + 242, + 495, + 280, + 536 + ], + "lines": [ + { + "bbox": [ + 242, + 495, + 280, + 536 + ], + "spans": [ + { + "bbox": [ + 242, + 495, + 280, + 536 + ], + "type": "image", + "image_path": "9a5e34868501b1c68574ae3f3fec5ab60266504d55d154b3e6d02cf9dc9a7f21.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 538, + 275, + 552 + ], + "lines": [ + { + "bbox": [ + 247, + 538, + 275, + 552 + ], + "spans": [ + { + "bbox": [ + 247, + 538, + 275, + 552 + ], + "type": "text", + "content": "Bg logits of step " + }, + { + "bbox": [ + 247, + 538, + 275, + 552 + ], + "type": "inline_equation", + "content": "t - 1" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 285, + 367, + 324, + 408 + ], + "blocks": [ + { + "bbox": [ + 285, + 367, + 324, + 408 + ], + "lines": [ + { + "bbox": [ + 285, + 367, + 324, + 408 + ], + "spans": [ + { + "bbox": [ + 285, + 367, + 324, + 408 + ], + "type": "image", + "image_path": "49d8fe702ce2bf0ac94e5d316896ee6ac13d5fd8b749305a88d4e131f64eaa53.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 285, + 409, + 324, + 451 + ], + "blocks": [ + { + "bbox": [ + 285, + 409, + 324, + 451 + ], + "lines": [ + { + "bbox": [ + 285, + 409, + 324, + 451 + ], + "spans": [ + { + "bbox": [ + 285, + 409, + 324, + 451 + ], + "type": "image", + "image_path": "8848eb1b153c2e2cc16ff4a40576b1edb9e00940ab58a7c6023ba537cb07f995.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 285, + 452, + 324, + 495 + ], + "blocks": [ + { + "bbox": [ + 285, + 452, + 324, + 495 + ], + "lines": [ + { + "bbox": [ + 285, + 452, + 324, + 495 + ], + "spans": [ + { + "bbox": [ + 285, + 452, + 324, + 495 + ], + "type": "image", + "image_path": "908428f9671c0ded36545d3fc3bb4a8d533438f5fa1cb0cac83a6f5123ab7db3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 285, + 495, + 324, + 536 + ], + "blocks": [ + { + "bbox": [ + 285, + 495, + 324, + 536 + ], + "lines": [ + { + "bbox": [ + 285, + 495, + 324, + 536 + ], + "spans": [ + { + "bbox": [ + 285, + 495, + 324, + 536 + ], + "type": "image", + "image_path": "73776b77fed59101a8a1357dceb0e4e4803f080cb003a4e6d493409d3bb6171d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 538, + 326, + 552 + ], + "lines": [ + { + "bbox": [ + 284, + 538, + 326, + 552 + ], + "spans": [ + { + "bbox": [ + 284, + 538, + 326, + 552 + ], + "type": "text", + "content": "Bg Adaptation of step " + }, + { + "bbox": [ + 284, + 538, + 326, + 552 + ], + "type": "inline_equation", + "content": "t" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 562, + 482, + 618 + ], + "lines": [ + { + "bbox": [ + 130, + 562, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 482, + 618 + ], + "type": "text", + "content": "Fig. 1: 3D Visualization of Background Adaptation results. The background logits for step " + }, + { + "bbox": [ + 130, + 562, + 482, + 618 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 562, + 482, + 618 + ], + "type": "text", + "content": " combine those from step " + }, + { + "bbox": [ + 130, + 562, + 482, + 618 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 562, + 482, + 618 + ], + "type": "text", + "content": " with the adaptation logits learned in step " + }, + { + "bbox": [ + 130, + 562, + 482, + 618 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 562, + 482, + 618 + ], + "type": "text", + "content": ", preventing disorderly adjustments and focusing on residuals. The current step background logits are processed using the Sigmoid function. Note that colors range from red (small values) to blue (large values)." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 329, + 367, + 367, + 408 + ], + "blocks": [ + { + "bbox": [ + 329, + 367, + 367, + 408 + ], + "lines": [ + { + "bbox": [ + 329, + 367, + 367, + 408 + ], + "spans": [ + { + "bbox": [ + 329, + 367, + 367, + 408 + ], + "type": "image", + "image_path": "7ba554256d6202cf91461801b388df15e9e13a712691e63778f88a8dbd910a43.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 329, + 409, + 367, + 451 + ], + "blocks": [ + { + "bbox": [ + 329, + 409, + 367, + 451 + ], + "lines": [ + { + "bbox": [ + 329, + 409, + 367, + 451 + ], + "spans": [ + { + "bbox": [ + 329, + 409, + 367, + 451 + ], + "type": "image", + "image_path": "dfad77d6c94cd1ea015d31579029aee42b99cc1299eeef0f0fec7d2bd9f6a384.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 329, + 452, + 368, + 495 + ], + "blocks": [ + { + "bbox": [ + 329, + 452, + 368, + 495 + ], + "lines": [ + { + "bbox": [ + 329, + 452, + 368, + 495 + ], + "spans": [ + { + "bbox": [ + 329, + 452, + 368, + 495 + ], + "type": "image", + "image_path": "99c7d18dca28bc05b9f75bfd01b480310c277fe64fd9b6e7e5cf09aee87e9bbb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 329, + 495, + 368, + 536 + ], + "blocks": [ + { + "bbox": [ + 329, + 495, + 368, + 536 + ], + "lines": [ + { + "bbox": [ + 329, + 495, + 368, + 536 + ], + "spans": [ + { + "bbox": [ + 329, + 495, + 368, + 536 + ], + "type": "image", + "image_path": "5d25eb747cc238b802967016b7a200872de0fd744b5e29de0b37f7f21225f55d.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 538, + 369, + 552 + ], + "lines": [ + { + "bbox": [ + 329, + 538, + 369, + 552 + ], + "spans": [ + { + "bbox": [ + 329, + 538, + 369, + 552 + ], + "type": "text", + "content": "Bg logits of step " + }, + { + "bbox": [ + 329, + 538, + 369, + 552 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 329, + 538, + 369, + 552 + ], + "type": "text", + "content": " (Sigmoid)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 372, + 365, + 414, + 408 + ], + "blocks": [ + { + "bbox": [ + 372, + 365, + 414, + 408 + ], + "lines": [ + { + "bbox": [ + 372, + 365, + 414, + 408 + ], + "spans": [ + { + "bbox": [ + 372, + 365, + 414, + 408 + ], + "type": "image", + "image_path": "0213258c00ff83766b4feb3341ab120259e859d57b62324e427bf191ef1dc28c.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 372, + 409, + 414, + 451 + ], + "blocks": [ + { + "bbox": [ + 372, + 409, + 414, + 451 + ], + "lines": [ + { + "bbox": [ + 372, + 409, + 414, + 451 + ], + "spans": [ + { + "bbox": [ + 372, + 409, + 414, + 451 + ], + "type": "image", + "image_path": "67829c7c655502e26655348e53d001d0aabe2efe0dbeb63e59a0e7fe5b63afec.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 372, + 452, + 414, + 495 + ], + "blocks": [ + { + "bbox": [ + 372, + 452, + 414, + 495 + ], + "lines": [ + { + "bbox": [ + 372, + 452, + 414, + 495 + ], + "spans": [ + { + "bbox": [ + 372, + 452, + 414, + 495 + ], + "type": "image", + "image_path": "2b02acc1185ac69de183ed6526314247802a9a173b1313e8698099bb7a60dcb6.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 372, + 495, + 414, + 536 + ], + "blocks": [ + { + "bbox": [ + 372, + 495, + 414, + 536 + ], + "lines": [ + { + "bbox": [ + 372, + 495, + 414, + 536 + ], + "spans": [ + { + "bbox": [ + 372, + 495, + 414, + 536 + ], + "type": "image", + "image_path": "464261f58dfc47ccb59e8eddb657715410f9c07ee1e878d0d037d52691a34286.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 538, + 410, + 552 + ], + "lines": [ + { + "bbox": [ + 373, + 538, + 410, + 552 + ], + "spans": [ + { + "bbox": [ + 373, + 538, + 410, + 552 + ], + "type": "text", + "content": "Prediction of step t" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 417, + 365, + 461, + 408 + ], + "blocks": [ + { + "bbox": [ + 417, + 365, + 461, + 408 + ], + "lines": [ + { + "bbox": [ + 417, + 365, + 461, + 408 + ], + "spans": [ + { + "bbox": [ + 417, + 365, + 461, + 408 + ], + "type": "image", + "image_path": "3abc89263b7272830f9a538f6e653f4a0cad8274b0f00d22a09748ac9cf6ee65.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 417, + 409, + 461, + 451 + ], + "blocks": [ + { + "bbox": [ + 417, + 409, + 461, + 451 + ], + "lines": [ + { + "bbox": [ + 417, + 409, + 461, + 451 + ], + "spans": [ + { + "bbox": [ + 417, + 409, + 461, + 451 + ], + "type": "image", + "image_path": "d74ed47d3da7515ea88233b0488d9379d4cd5bbe9e6d76a73f25863e81913f67.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 417, + 452, + 461, + 495 + ], + "blocks": [ + { + "bbox": [ + 417, + 452, + 461, + 495 + ], + "lines": [ + { + "bbox": [ + 417, + 452, + 461, + 495 + ], + "spans": [ + { + "bbox": [ + 417, + 452, + 461, + 495 + ], + "type": "image", + "image_path": "819b40936f42cf055bc5906169a1c5ab18bc3098957d3ce7f2a9e62974fea72b.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 417, + 495, + 461, + 536 + ], + "blocks": [ + { + "bbox": [ + 417, + 495, + 461, + 536 + ], + "lines": [ + { + "bbox": [ + 417, + 495, + 461, + 536 + ], + "spans": [ + { + "bbox": [ + 417, + 495, + 461, + 536 + ], + "type": "image", + "image_path": "a1ed7d4703a540e65796bf292c82d4de1fc4a292d7cac88f8baf23334b8be20d.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 538, + 455, + 552 + ], + "lines": [ + { + "bbox": [ + 417, + 538, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 417, + 538, + 455, + 552 + ], + "type": "text", + "content": "Categories in step t" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Nevertheless, background shifting, where the background category changes constantly in each step, is another challenge for CISS. Upon further investiga" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 354 + ], + "type": "text", + "content": "tion, it was observed that previous methods within CISS often utilize a shared background classifier across each step. However, novel classes are typically considered as the background in previous learning steps, indicating that the background class undergoes continuous changes throughout incremental learning. This way results in an incessantly shifting background in each step, leading to disruption in background prediction and inaccurate learning of novel classes. Some approaches [8,9,41] store old exemplars in extra memory space to prevent background shifting, while long-term storage of such data may not be feasible in real-world scenarios. To enhance the incremental nature of the CISS method and reduce background instability, we revise the representation of the background class and propose an exemplar-free background adaptation mechanism. In particular, the background adaptation mechanism explicitly models the background residual rather than the background itself in each incremental step. The learned residuals of the additional channel are aggregated with the previously learned background logits to represent the evolving background, as illustrated in Fig. 1. By employing the background adaptation mechanism, we only need to optimize the additional channel of residuals for constructing the background during the incremental steps. This approach eliminates the need for repeatedly learning background classifiers associated with old classes, thereby avoiding the instability of optimizing a shared background classifier." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 358, + 481, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 481, + 584 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 481, + 584 + ], + "type": "text", + "content": "Moreover, we customize our training approaches in order to fit the training process of the background adaptation structure and promote its performance. We first introduce the widely used pseudo-labeling [8,21] to generate a pseudobackground, and then we design a PB-BCE loss to focus on optimizing the classifier of novel classes and related background adaptation residual. Besides, for the Background Adaptation channel, the objective is to minimize the values in the regions of novel categories, while ensuring that the values in other regions remain above 0. Thus, two auxiliary Background Adaptation losses are specifically applied to the background adaptation channel for precise adjustment. Given our pursuit of plasticity on novel categories without external memory, we have devised more comprehensive exemplar-free knowledge distillation strategies to combat catastrophic forgetting. We further adopt knowledge distillation strategies on the logit maps of old categories and the intermediate features, namely the Group Knowledge Distillation and Background Feature Distillation, respectively. Finally, our approach outperforms previous methods in multiple benchmarks, especially in realistic and challenging long-term scenarios such as 10-1 of VOC2012 and 100-5 of ADE20K. Our approach achieves significant performance gains of " + }, + { + "bbox": [ + 133, + 358, + 481, + 584 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 133, + 358, + 481, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 358, + 481, + 584 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 133, + 358, + 481, + 584 + ], + "type": "text", + "content": " on novel classes, respectively, compared to the previous state-of-the-art." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 588, + 382, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 588, + 382, + 598 + ], + "spans": [ + { + "bbox": [ + 147, + 588, + 382, + 598 + ], + "type": "text", + "content": "Overall, our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 481, + 665 + ], + "type": "text", + "content": "- We propose a Background Adaptation strategy in incremental steps by optimizing an additional channel for residual modeling instead of the shared background classifier, which avoids unstable optimization and prevents chaos in background logits." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Background Adaptation with Residual Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "content": "- We design PB-BCE loss and Background Adaptation losses to enhance the effectiveness of Background Adaptation, while two distillation strategies ensure maintaining knowledge of old categories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 480, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 480, + 174 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 480, + 174 + ], + "type": "text", + "content": "- Experiments validate the effectiveness of our proposed method, showcasing a new state-of-the-art performance on CISS benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 190, + 243, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 190, + 243, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 243, + 202 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 212, + 299, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 212, + 299, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 299, + 224 + ], + "type": "text", + "content": "2.1 Class Incremental Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 228, + 482, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 408 + ], + "type": "text", + "content": "Class Incremental Learning (CIL) breaks the constraints of the standard training process in an incremental form on new classes. CIL task aims to continually learn the new classes without retraining on the data of old classes, which significantly reduces the training cost. As a compromise, the CIL task has to effectively solve the problem of catastrophic forgetting [38, 39]. Current CIL methods are divided into several main categories. Replay-based methods recall the data of old classes during incremental training. Some of the approaches [2, 4, 35] reserve memory buffers and temporarily store representative past data. Others [42, 49, 55] introduce powerful generative models to generate pseudo-data for rehearsal. Regularization-based methods introduce regularization terms to prevent forgetting the old classes. These methods utilize loss functions as penalties [1, 29, 45, 47] or knowledge distillation from the previous model [19, 32, 44]. Representation-based methods attempt to transfer representations to certain categories, which apply adaptation on a fixed pre-trained backbone [5, 22, 50, 60] or self-supervised learning [7, 36, 43]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 422, + 372, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 372, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 372, + 434 + ], + "type": "text", + "content": "2.2 Class Incremental Semantic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": "The achievements in CIL tasks encourage the Incremental Learning pipeline to develop in other dense prediction tasks, including Semantic Segmentation. Class Incremental Semantic Segmentation (CISS) task is first introduced in MiB [6], which focused on reconstructing the background region under the restriction from knowledge distillation. Since then, various methods [57, 63] have been introduced in the CISS task. PLOP [21] defined the pseudo-labeling and multi-scale local distillation to remember the old categories. SDR [40] introduced category-wise prototype matching and contrastive learning methods to enlarge the difference between categories. RCIL [58] and EWF [56] explored methods for integrating parameters of new and old models. DKD [3] decomposed the class logits into positive and negative logits for better knowledge distillation and applied threshold to generate background regions. Incrementer [48] borrows prompt learning from ViT [20] and applies the structure of Segmenter [52] instead of DeeplabV3 [12], which is the basic structure of other methods. Besides, several approaches utilize external memory [66] or models [37, 62] for better performance. SSUL [8] introduced a Saliency-map detector to identify the saliency region and store some of the previous instances in memory. MicroSeg [61] utilized proposals from pretrained Mask2Former [15] as a guidance. ALIFE [41] memorized category features and adapted for remembering the old classes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 402, + 251 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 402, + 251 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 402, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 402, + 251 + ], + "type": "image", + "image_path": "779f2a42ff61ffb57b6df982cc03cef055be9ba6e9995a83b67d19385e26245b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 253, + 339, + 262 + ], + "lines": [ + { + "bbox": [ + 195, + 253, + 339, + 262 + ], + "spans": [ + { + "bbox": [ + 195, + 253, + 339, + 262 + ], + "type": "text", + "content": "(a) The overall pipeline of our approach." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 414, + 155, + 477, + 246 + ], + "blocks": [ + { + "bbox": [ + 414, + 155, + 477, + 246 + ], + "lines": [ + { + "bbox": [ + 414, + 155, + 477, + 246 + ], + "spans": [ + { + "bbox": [ + 414, + 155, + 477, + 246 + ], + "type": "image", + "image_path": "c2e6f0a7d7c85d9a9fcce5af4d17978e190e2324486eaa9048eff006192d01bc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 252, + 470, + 262 + ], + "lines": [ + { + "bbox": [ + 414, + 252, + 470, + 262 + ], + "spans": [ + { + "bbox": [ + 414, + 252, + 470, + 262 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 414, + 252, + 470, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{BgA}" + }, + { + "bbox": [ + 414, + 252, + 470, + 262 + ], + "type": "text", + "content": " losses." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "text", + "content": "Fig. 2: Overview of our framework that consists of a Background Adaptation mechanism and proposed losses. (a) The Background Adaptation mechanism uses the negative part of the adaptation channel to rectify the background logits. (b) Two Background Adaptation losses separately optimize the regions " + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 130, + 271, + 482, + 327 + ], + "type": "text", + "content": " novel categories to enhance the performance of the Background Adaptation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 350, + 207, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 207, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 207, + 361 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 376, + 228, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 228, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 228, + 387 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": "The Class-Incremental Semantic Segmentation (CISS) task performs training in " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " steps. During step " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": ", the training data " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "D^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " containing the labels of the novel classes " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "C^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " are applied for training, while the labels of old classes " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " are expunged. In particular, each sample of the training data " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "D^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " consists of an RGB image " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "x^{t} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " and a corresponding ground truth label " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " on novel classes " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "C^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": ". Besides, the ground truth label " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " replaces the labels of the " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " with the background label " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "c_{b}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " as the old classes will not appear in " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "C^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " to ensure " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "C^{1:t-1} \\cap C^{t} = \\emptyset" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": ". Therefore, the model " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "f^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " being trained at step " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " not only needs to learn to segment novel categories but also has to retain the capability to segment old categories. However, while learning the novel categories, the model " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "f^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " tends to overfit the novel categories " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "C^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "f^{t}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": " cannot contact with samples of " + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 130, + 397, + 482, + 578 + ], + "type": "text", + "content": ", causing severe catastrophic forgetting. Moreover, the novel categories can only appear in the background region during the previous steps. We are more likely to classify them from the previous background prediction region, which motivates us to focus on background adaptation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 596, + 281, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 281, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 281, + 609 + ], + "type": "text", + "content": "3.2 Background Adaptation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Previous studies often apply only one background classifier. They have to readjust the parameters of the background classifier whenever the model learns novel categories, which causes catastrophic forgotten even if the classifier of the old categories remains the same probability output. The unstable background classifier" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Background Adaptation with Residual Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "could be a thorny constraint factor. To tackle the problems above, we design a Background Adaptation mechanism. Different from the previous distinct background classifier, we introduce an additional background adaptation channel for each classifier of the incremental process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": "During the training process of incremental step " + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": ", the current model " + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "inline_equation", + "content": "f^t" + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": " inherits the parameters from the previous model " + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "inline_equation", + "content": "f^{t-1}" + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": " and adds a new classifier " + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "inline_equation", + "content": "f_{C^t}^t" + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": " for parallel classification of categories in " + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": " alongside the previous classifiers. The encoder " + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "inline_equation", + "content": "f_e^t" + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "inline_equation", + "content": "f_d^t" + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": " transform the input image " + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "inline_equation", + "content": "x^t" + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": " into features map " + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "inline_equation", + "content": "F_d^t \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 164, + 480, + 225 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 272, + 227, + 480, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 227, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 272, + 227, + 480, + 239 + ], + "type": "interline_equation", + "content": "F _ {d} ^ {t} = f _ {d} ^ {t} \\left(f _ {e} ^ {t} \\left(x ^ {t}\\right)\\right). \\tag {1}", + "image_path": "e55bfa95d94acdb330dd50399c2d9423c29bde97a16cefd42d4277db87e2e035.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "text", + "content": "Compared to the standard classifier of the previous methods [8, 21, 41], our classifier of each incremental step has an additional channel for background adaptation. Therefore, for each group of categories " + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "text", + "content": " from step " + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "text", + "content": ", the classifier " + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "inline_equation", + "content": "f_{C^i}^t" + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "text", + "content": " of the model " + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "inline_equation", + "content": "f^t" + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "text", + "content": " generates the corresponding classification logit maps " + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "inline_equation", + "content": "\\mu_{C^i \\cup \\{b_i\\}}^t \\in \\mathbb{R}^{h \\times w \\times (|C^i| + 1)}" + }, + { + "bbox": [ + 130, + 247, + 479, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 263, + 323, + 480, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 323, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 263, + 323, + 480, + 338 + ], + "type": "interline_equation", + "content": "\\mu_ {C ^ {i} \\cup \\{b _ {i} \\}} ^ {t} = f _ {C ^ {i}} ^ {t} \\left(F _ {d} ^ {t}\\right), \\tag {2}", + "image_path": "652778abaeb8cb3fdd549251b9a1d0f20641522569b9b9cc08c91716f3b9cdb7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "i \\in \\{1, 2, \\ldots, t\\}" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "b^1" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " to represent the background class of the initial step, and " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "b^i" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " denotes the background adaptation for subsequent incremental steps when " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "i >= 2" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": ". The symbol " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " refers to the adapted background class. Thus, " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "\\mu_{C^i \\cup \\{b_i\\}}^t" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " contains the probability maps " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "\\mu_{C^i}" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "|C^i|" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " categories from step " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " and an additional channel " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "\\mu_{b^i}^t" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " for constructing the background logits " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "\\mu_b^t" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": ". Considering that the region of the novel categories " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " is more likely inherited from the background region of the previous step, the classifier " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "f_{C^t}^t" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " adjusts the background logits through " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "\\mu_{b^t}^t" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": ", which decreases the probability values in the region " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C^t}" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 130, + 348, + 480, + 481 + ], + "type": "text", + "content": ". The parameters of previous classifiers are not required to continuously optimize for new background regions, thereby avoiding instability in the classification of old categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "text", + "content": "The background adaptation channel " + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mu_{b^t}^t" + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "text", + "content": " can correct the background logits " + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mu_b^t\\in \\mathbb{R}^{h\\times w\\times 1}" + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C^t}" + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "text", + "content": ", yet the noise in other regions of " + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mu_{b^t}^t" + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "text", + "content": " may interfere with the background prediction. Particularly, when old categories are present in the image, there is a risk that these regions could be incorrectly covered by the adapted background. Hence, we design a Filter operation that replaces positive values of " + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mu_{b^i}^t" + }, + { + "bbox": [ + 130, + 483, + 480, + 567 + ], + "type": "text", + "content": " with 0. As shown in Fig. 2a, with the Filter operation applied, the residual modeling of background logits during the inference process is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 578, + 480, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 249, + 578, + 480, + 609 + ], + "type": "interline_equation", + "content": "\\mu_ {b} ^ {t} = \\mu_ {b ^ {1}} ^ {t} + \\sum_ {i = 2} ^ {t} F i l t e r \\left(\\mu_ {b ^ {i}} ^ {t}\\right), \\tag {3}", + "image_path": "fe6fa0d642f75b588a172a6136961125939add29446bdc93e5ab7bcc1e22f51b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "content": "where the background logits from the initial step " + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mu_{b^1}^t" + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "content": " stay the same and the background adaptation logits from the incremental steps " + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mu_{b^i}^t" + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "content": " only remain the negative values via the Filter. In particular, to make " + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mu_{b}^{t}" + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "content": " more friendly for gradient backpropagation, we abandoned the Filter on the background adaptation logits" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 340, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 340, + 129 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 340, + 129 + ], + "type": "text", + "content": "of novel classes " + }, + { + "bbox": [ + 131, + 116, + 340, + 129 + ], + "type": "inline_equation", + "content": "\\mu_{b^t}^t" + }, + { + "bbox": [ + 131, + 116, + 340, + 129 + ], + "type": "text", + "content": " during the training process:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 139, + 481, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 233, + 139, + 481, + 171 + ], + "type": "interline_equation", + "content": "\\mu_ {b} ^ {t} = \\mu_ {b ^ {1}} ^ {t} + \\sum_ {i = 2} ^ {t - 1} \\operatorname {F i l t e r} \\left(\\mu_ {b ^ {i}} ^ {t}\\right) + \\mu_ {b ^ {t}} ^ {t}. \\tag {4}", + "image_path": "9769fdaf97988ed138d110712d179a303b9fa67d84d2c4acf6c047a37aeedd3b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 190, + 313, + 203 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 190, + 313, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 313, + 203 + ], + "type": "text", + "content": "3.3 Incremental Training Strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "text", + "content": "Pseudo Labeling. During the incremental steps, even if objects of old categories " + }, + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "text", + "content": " exist in the image, we cannot obtain corresponding labels from the ground truth. In order to simulate a ground truth label that contains the categories we have learned, we follow [8,21,61] to introduce the prediction probability of the previous step " + }, + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "inline_equation", + "content": "\\varphi^{t-1} \\in \\mathbb{R}^{h \\times w \\times |C^{1:t-1}|}" + }, + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "inline_equation", + "content": "f^{t-1}(x^t)" + }, + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "text", + "content": " with Sigmoid function and combine it with the current ground truth label " + }, + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 131, + 212, + 479, + 286 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 181, + 295, + 481, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 295, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 181, + 295, + 481, + 327 + ], + "type": "interline_equation", + "content": "\\hat {y} ^ {t} = \\left\\{ \\begin{array}{l l} y ^ {t}, & \\left(y ^ {t} \\in C ^ {t}\\right) \\vee \\left\\{\\left(y ^ {t} \\in c _ {b}\\right) \\wedge \\left(\\max \\left(\\varphi^ {t - 1}\\right) < \\tau\\right) \\right\\} \\\\ \\tilde {y} ^ {t - 1}, & \\left(y ^ {t} \\in c _ {b}\\right) \\wedge \\left(\\max \\left(\\varphi^ {t - 1}\\right) \\geq \\tau\\right) \\end{array} \\right. \\tag {5}", + "image_path": "4d6c00653a5f31e885d29bd234b613fbc095300e80491891dc15d205398d7b67.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 337, + 479, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 479, + 363 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 479, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 337, + 479, + 363 + ], + "type": "inline_equation", + "content": "\\hat{y}^t\\in \\mathbb{R}^{h\\times w}" + }, + { + "bbox": [ + 131, + 337, + 479, + 363 + ], + "type": "text", + "content": " denotes the generated pseudo label and " + }, + { + "bbox": [ + 131, + 337, + 479, + 363 + ], + "type": "inline_equation", + "content": "\\tilde{y}^{t - 1} = \\arg \\max (\\varphi^{t - 1})" + }, + { + "bbox": [ + 131, + 337, + 479, + 363 + ], + "type": "text", + "content": " represents the final prediction from " + }, + { + "bbox": [ + 131, + 337, + 479, + 363 + ], + "type": "inline_equation", + "content": "f^{t - 1}(x^t)" + }, + { + "bbox": [ + 131, + 337, + 479, + 363 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "text", + "content": "PB-BCE Loss. Binary Cross-Entropy (BCE) loss is extensively employed for learning to segment objects of novel categories. We follow the usage of the Sigmoid function and activate " + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mu^t" + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\varphi^t" + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "text", + "content": ". Some of the CISS methods [8,21,61] utilize the pseudo mask as the ground truth for supervision, which effectively replaces the background label in the regions of the old categories. However, the ambiguous supervision of the old categories could improperly optimize the classifiers of " + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "inline_equation", + "content": "C^{1:t - 1}" + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "text", + "content": " and destroy their identification capability. Therefore, we design a Pseudo Background BCE (PB-BCE) loss which cuts off the backpropagation of old categories, meaning that we only take the labels of current categories " + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\hat{y}_{C^t}^t" + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "text", + "content": " and pseudo-background " + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\hat{y}_{c_b}^t" + }, + { + "bbox": [ + 131, + 381, + 482, + 503 + ], + "type": "text", + "content": " for optimization:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 201, + 513, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 513, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 201, + 513, + 481, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P B - B C E} = - \\frac {1}{h w} \\sum_ {k = 1} ^ {h w} \\sum_ {c \\in \\{c _ {b} \\} \\cap C ^ {t}} C E \\left(\\varphi_ {c, k} ^ {t}, \\hat {\\psi} _ {c, k} ^ {t}\\right), \\tag {6}", + "image_path": "6cd8cf44fb013898890cccd50852bc74abc9a333e4625dc1203ee156c9c77009.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "text", + "content": " represents the index of pixel and " + }, + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "inline_equation", + "content": "\\hat{\\psi}_c^t" + }, + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "text", + "content": " denotes the binary map of class c in pseudo map " + }, + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "inline_equation", + "content": "\\hat{y}^t" + }, + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "inline_equation", + "content": "CE" + }, + { + "bbox": [ + 131, + 559, + 479, + 584 + ], + "type": "text", + "content": " is the cross-entropy calculation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 198, + 592, + 481, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 592, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 198, + 592, + 481, + 608 + ], + "type": "interline_equation", + "content": "C E \\left(\\varphi_ {*} ^ {t}, \\hat {\\psi} _ {*} ^ {t}\\right) = \\hat {\\psi} _ {*} ^ {t} \\cdot \\log \\varphi_ {*} ^ {t} + \\left(1 - \\hat {\\psi} _ {*} ^ {t}\\right) \\cdot \\log \\left(1 - \\varphi_ {*} ^ {t}\\right). \\tag {7}", + "image_path": "a1a7b75b072170999f6b6f4a7af0a907ddfbf06c4982bef15c937f1f6d80eeed.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "content": "Although the whole background prediction logits " + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mu_b^t" + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "content": " are involved in the loss calculation, only the background adaptation channel for current categories " + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mu_{b^t}^t" + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "content": " serves for optimization. Previous background logits are detached to prevent disorderly evolving on parameters of the old classifiers." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "type": "text", + "content": "Background Adaptation with Residual Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "text", + "content": "Background Adaptation Losses. The Background Adaptation (BgA) losses, as shown in Fig. 2b, are designed for further refined adjustments on the background adaptation channel " + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "\\mu_{b^t}^t" + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "text", + "content": ". We propose two losses for two different regions of " + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "\\mu_{b^t}^t" + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "text", + "content": " according to the target effect. For regions " + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C^t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "text", + "content": " belonging to the current classes " + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "text", + "content": ", the ideal target is to have negative values as small as possible. To achieve this, we introduce cross-entropy loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BgA^{+}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "text", + "content": " for supervision:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 196, + 481, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 211, + 196, + 481, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B g A ^ {+}} = - \\frac {1}{| \\mathcal {R} _ {C ^ {t}} |} \\sum_ {k \\in \\mathcal {R} _ {C ^ {t}}} \\log \\frac {\\exp \\left(- \\mu_ {b ^ {t} , k} ^ {t}\\right)}{1 + \\exp \\left(- \\mu_ {b ^ {t} , k} ^ {t}\\right)}, \\tag {8}", + "image_path": "03e6f8cc2886422df41805aef940f6ce9aca37ebc1108cfa8b1af80b5be62f49.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "inline_equation", + "content": "|\\mathcal{R}_{C^t}|" + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "content": " represents the number of pixels in " + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C^t}" + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "content": ". Besides, as for the regions of the background that might contain old classes, we hope the value of " + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mu_{bt}^{t}" + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "content": " could be 0 for maintaining the historical background prediction result. However, achieving precise control over the whole region to be a precise value is challenging and may be impractical. Therefore, we set the positive logits in the adaptation channel of the old categories to 0 in Eq. 4 to avoid the noise and apply a Triplet loss " + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BgA^{-}}" + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "content": " to supervise the background region:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 170, + 327, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 327, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 170, + 327, + 481, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B g A ^ {-}} = \\frac {1}{h w - \\left| \\mathcal {R} _ {C ^ {t}} \\right|} \\sum_ {k \\notin \\mathcal {R} _ {C ^ {t}}} \\max \\left[ 0, \\left(1 - \\varphi_ {b ^ {t}, k} ^ {t}\\right) ^ {2} - \\left(\\varphi_ {b ^ {t}, k} ^ {t} - 0\\right) ^ {2} \\right]. \\tag {9}", + "image_path": "07b4c49462ec699ec5c369dcf04376ff44a1da4a5e9843a3b8bd3d80a4ed3994.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "text", + "content": "We first activate the background adaptation channel as " + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\varphi_{b^t}^t" + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "text", + "content": " via Sigmoid function, then the logits are mapped to the range from 0 to 1. We aim for the logits " + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mu_{b^t}^t" + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "text", + "content": " of this region to be greater than 0 while not overfitting to the objects in the background, meaning that values of " + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\varphi_{b^t}^t" + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "text", + "content": " lie within the range of 0.5 to 1 and not close to 1. In particular, The Triplet loss " + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BgA^{-}}" + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "text", + "content": " respectively set 0 and 1 as negative and positive anchors. We simultaneously constrain the loss not to go below 0 by setting a lower bound. Therefore, once the value of the pixel exceeds 0.5, i.e., the middle of the positive and negative anchors, there will be no more gradient on the pixel for backpropagation, which mitigates overfitting to the objects in the background." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 499, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 559 + ], + "type": "text", + "content": "Knowledge Distillation Strategies. In the training process of incremental steps for CISS tasks, substantial alterations to the existing parameters of " + }, + { + "bbox": [ + 130, + 499, + 482, + 559 + ], + "type": "inline_equation", + "content": "f^t" + }, + { + "bbox": [ + 130, + 499, + 482, + 559 + ], + "type": "text", + "content": " frequently occur, giving rise to situations of catastrophic forgetting. Guided by the previous methods, we have devised a series of knowledge distillation strategies tailored to our background adaptation mechanism." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 565, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 565, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 137, + 565, + 482, + 625 + ], + "type": "text", + "content": "- Group Knowledge Distillation. In Sec. 3.2, we have introduced the classification logit map " + }, + { + "bbox": [ + 137, + 565, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\mu_{C^i\\cup \\{b_i\\}}^t" + }, + { + "bbox": [ + 137, + 565, + 482, + 625 + ], + "type": "text", + "content": " for each group of categories " + }, + { + "bbox": [ + 137, + 565, + 482, + 625 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 137, + 565, + 482, + 625 + ], + "type": "text", + "content": ", which contains an additional channel for background adaptation. We introduce a Group Knowledge Distillation loss " + }, + { + "bbox": [ + 137, + 565, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GKD}" + }, + { + "bbox": [ + 137, + 565, + 482, + 625 + ], + "type": "text", + "content": " to ensure the stability of the old category logits, as well as the previous background adaptation channel:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 213, + 633, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 633, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 213, + 633, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G K D} = - \\frac {1}{h w} \\sum_ {k = 1} ^ {h w} \\sum_ {i = 1} ^ {t - 1} \\sum_ {c \\in \\{c _ {b} ^ {i} \\} \\cup C ^ {i}} C E \\left(\\varphi_ {c, k} ^ {t}, \\varphi_ {c, k} ^ {t - 1}\\right) \\tag {10}", + "image_path": "ba33e01bcd3002e92bf47e6ec538fb4c65ddca3d177b241fb113ecde61554d39.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "inline_equation", + "content": "\\varphi_{*}^{t - 1}" + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "inline_equation", + "content": "\\varphi_{*}^{t}" + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": " denote the activated logits of " + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "inline_equation", + "content": "\\mu_{*}^{t - 1}" + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "inline_equation", + "content": "\\mu_{*}^{t}" + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": ", respectively. \n- Background Feature Distillation. When the network learning novel categories, not only the parameters of the novel classifier " + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "inline_equation", + "content": "f_{C^t}^t" + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": " are optimized, but the encoder " + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "inline_equation", + "content": "f_{e}^{t}" + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "inline_equation", + "content": "f_{d}^{t}" + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": " are affected as well. Merely utilizing Group Knowledge Distillation loss " + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GKD}" + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": " is insufficient in effectively mitigating catastrophic forgetting as it primarily focuses on constraining classifiers of the old categories. However, excessively constraining parameter optimization can impede learning for novel categories. Taking all these factors into account, we designed a Background Feature Distillation (BFD) method that only acts on the features out of the regions of current categories " + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C^t}" + }, + { + "bbox": [ + 138, + 115, + 479, + 235 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 243, + 481, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 243, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 233, + 243, + 481, + 275 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B F D} = \\sum_ {i = 1} ^ {t - 1} \\| \\psi_ {b} ^ {t} \\cdot F _ {m ^ {i}} ^ {t} - \\psi_ {b} ^ {t} \\cdot F _ {m ^ {i}} ^ {t - 1} \\| _ {2} ^ {2}, \\tag {11}", + "image_path": "aa02d039cd069ef77c3c928d2f75a39a46f35c0ffb7c89d02b386f696f1538d9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "spans": [ + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "inline_equation", + "content": "F_{m^i}^t" + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "content": " denotes the intermediate features in " + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "inline_equation", + "content": "f_{C^i}^t" + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "inline_equation", + "content": "\\psi_b^t" + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "content": " represents the binary mask out of " + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C^t}" + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "content": " calculates the Euclidean distance between the two matrix. The Background Feature Distillation allows other losses to supervise the region of novel categories " + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C^t}" + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "content": " while restricting the negative effect outside " + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C^t}" + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "content": ", where old categories " + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 148, + 284, + 479, + 345 + ], + "type": "text", + "content": " could appear." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 257, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 257, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 257, + 373 + ], + "type": "text", + "content": "3.4 Objective Function" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 381, + 477, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 477, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 477, + 393 + ], + "type": "text", + "content": "With our proposed losses above, we construct our objective function as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 145, + 402, + 480, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 402, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 145, + 402, + 480, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {P B - B C E} + \\lambda_ {1} \\cdot \\mathcal {L} _ {B g A ^ {+}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {B g A ^ {-}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {G K D} + \\lambda_ {4} \\cdot \\mathcal {L} _ {B F D}, \\tag {12}", + "image_path": "c0e5e7f9c1aa39008a59729f62fe2ccd1b6e8b7f7cb2eeac7852597bdeb29299.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PB - BCE}" + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "text", + "content": " supervises the learning of novel classes, " + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PB - BCE}" + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BgA^{-}}" + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "text", + "content": " enhance the effect of Background Adaptation, " + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GKD}" + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BFD}" + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "text", + "content": " mitigate forgetting old classes. Hyperparameters " + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_1,\\lambda_2,\\lambda_3,\\lambda_4" + }, + { + "bbox": [ + 132, + 424, + 479, + 460 + ], + "type": "text", + "content": " balance these constraints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 478, + 229, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 229, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 229, + 491 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 502, + 261, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 261, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 261, + 514 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 479, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 592 + ], + "type": "text", + "content": "Datasets. We follow the settings of most previous work and evaluate our proposed methods on Pascal VOC 2012 [23,26] and ADE20K [64] datasets. Pascal VOC 2012 is composed of 20 different classes with 10,582 images for training and 1,449 images for validation. ADE20K has 150 classes that contain 100 thing classes and 50 stuff classes. There are 20,210 images for training and 2,000 images for validation in ADE20K." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "Protocols. Earlier work [6, 21] has defined two types of training settings, i.e. overlapping and disjoint. The overlapping setting contains future classes in the previous steps, whereas the disjoint setting forbids future classes until they belong to the training classes. We follow recent work [3, 41, 56, 63] and adopt the overlapping setting for evaluating as it is more realistic. Besides, following the previous training protocols [6, 21, 56], we set each sub-task from the dataset as" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "type": "text", + "content": "Background Adaptation with Residual Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "N_{ini} - N_{inc}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "N_{ini}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "N_{inc}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " denotes the quantity of the categories in the initial step and a single incremental step, respectively. For example, the 15-1 sub-task of Pascal VOC 2012 has 15 classes for the initial training step. During each incremental step, one novel class is applied for incremental training without previous classes. Specifically, we define 4 sub-tasks on Pascal VOC 2012, including 19-1 (2 steps), 15-1 (6 steps), 10-1 (11 steps), and 5-3 (6 steps). On ADE20K, we define 4 sub-tasks containing 100-50 (2 steps), 100-10 (6 steps), 50-50 (3 steps), and 100-5 (11 steps)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate our proposed method with widely-used mean Intersection-over-Union (mIoU). The Intersection-over-Union (IoU) is defined as " + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "inline_equation", + "content": "IoU = \\frac{TP}{TP + FP + FN}" + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "inline_equation", + "content": "TP" + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "content": " represent the quantity of true-positive, false-positive, and false-negative prediction pixels. The mIoU calculates the mean value of IoU corresponding to each category. To effectively measure the performance of CISS tasks, we compute the mIoU values of the initial classes " + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "inline_equation", + "content": "C^0" + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "content": ", the incremental classes " + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "inline_equation", + "content": "C^{1:t}" + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "content": ", and all classes " + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "inline_equation", + "content": "C^{0:t}" + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "content": "Implementation details. Following the previous studies [6, 8, 40], we apply Deeplab-v3 [12] as the segmentation network with a ResNet-101 [27] for feature extraction. The ResNet-101 is initially pretrained on ImageNet-1K [18]. We use SGD optimizer to optimize the network. The initial learning rate of the initial step is set to " + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "content": ", while in the incremental steps, the initial learning rate is set to " + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "content": " for Pascal VOC 2012 and " + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "content": " for ADE20K. The momentum value is 0.9 and the weight decay is " + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "content": " in all steps. We apply the poly learning rate schedule following [3, 8, 12, 21]. During the initial step, the training epoch is set to 50 for Pascal VOC 2012 and 60 for ADE20K, while for each incremental step, we train the network for 20 epochs on Pascal VOC and 100 epochs on ADE20K. The batch size is 16 for Pascal VOC 2012 and 8 for ADE20K in all steps. The data augmentation follows the standard of previous work, which is composed of random scaling, random flipping, and random crop. The coefficients " + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\lambda_{1}, \\lambda_{2}, \\lambda_{3}, \\lambda_{4}" + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "content": ", of the objective function are set to 1, 5, 1, 4. The threshold of pseudo labeling " + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\tau = 0.7" + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "content": " is set to all settings. The implementation is built on Pytorch. The experiments are conducted on NVIDIA RTX2080Ti GPUs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 503, + 265, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 265, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 265, + 515 + ], + "type": "text", + "content": "4.2 Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": "Comparison on Pascal VOC 2012. We evaluate our proposed method with current exemplar-free DeeplabV3-based methods on the Pascal VOC 2012 dataset. The sub-tasks of Pascal VOC 2012 have distinct characteristics. For example, the sub-task 19-1 has more categories in the initial step, while 5-3 has fewer classes for the initial training and more incremental steps, which is much more challenging. The results in Tab. 1 show that our proposed method outperforms other methods on novel classes and overall mIoU results of all sub-tasks. Particularly, our approach has " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " promotion of mIoU in more challenging sub-tasks of 10-1 and 5-3, respectively. Except for DKD [3] which refine the background class, our approach achieves at least " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " of promotion on mIoU Compared to the methods with a standard shared background classifier, demonstrating the effectiveness of the Background Adaptation mechanism." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "A. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 480, + 304 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparison with state-of-the-art exemplar-free methods on Pascal VOC 2012 in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 480, + 304 + ], + "type": "table", + "html": "
Method15-1 (6 steps)10-1 (11 steps)5-3 (6 steps)19-1 (1 step)
0-1516-20all0-1011-20all0-56-20all0-1920all
MiB [CVPR20] [6]38.013.532.212.213.112.657.142.546.771.222.168.9
SDR [CVPR21] [40]47.314.739.532.417.125.1---69.132.667.4
PLOP [CVPR21] [21]65.121.154.644.015.530.525.730.028.775.437.373.5
RCIL [CVPR22] [58]70.623.759.455.415.134.363.134.642.877.031.574.7
EWF [CVPR23] [56]77.732.767.071.530.351.961.742.247.777.96.774.5
AWT [WACV23] [25]59.117.249.133.218.026.061.845.950.4---
IDEC [TPAMI23] [63]77.036.567.370.746.359.167.149.054.1---
DKD [NeurIPS22] [3]78.142.769.773.146.560.469.653.558.177.841.576.0
Ours77.645.970.072.251.062.171.357.061.178.242.276.4
Joint79.872.477.478.476.477.476.977.677.477.577.077.4
", + "image_path": "43660eabb450ab643263f2647755ce6e30492e27b2c649a8390d5f763322e74f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 157, + 319, + 294, + 424 + ], + "blocks": [ + { + "bbox": [ + 157, + 319, + 294, + 424 + ], + "lines": [ + { + "bbox": [ + 157, + 319, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 157, + 319, + 294, + 424 + ], + "type": "image", + "image_path": "ca7fb1225a0e710d7c023d178794f6b71262ccd0421ce0eb2cd34b11e94413a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 430, + 270, + 440 + ], + "lines": [ + { + "bbox": [ + 173, + 430, + 270, + 440 + ], + "spans": [ + { + "bbox": [ + 173, + 430, + 270, + 440 + ], + "type": "text", + "content": "(a) Task 15-1 of VOC2012." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 318, + 319, + 462, + 423 + ], + "blocks": [ + { + "bbox": [ + 318, + 319, + 462, + 423 + ], + "lines": [ + { + "bbox": [ + 318, + 319, + 462, + 423 + ], + "spans": [ + { + "bbox": [ + 318, + 319, + 462, + 423 + ], + "type": "image", + "image_path": "d6a7066a6e6348b8d583e1d7eb9a596af3caabffe512fba29c05505e365ca214.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 430, + 434, + 440 + ], + "lines": [ + { + "bbox": [ + 337, + 430, + 434, + 440 + ], + "spans": [ + { + "bbox": [ + 337, + 430, + 434, + 440 + ], + "type": "text", + "content": "(b) Task 10-1 of VOC2012." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 449, + 482, + 471 + ], + "lines": [ + { + "bbox": [ + 131, + 449, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 482, + 471 + ], + "type": "text", + "content": "Fig. 3: The step-wise mIoU comparison of our approach with previous methods under the sub-tasks 10-1 and 15-1 of Pascal VOC 2012." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "content": "Besides, our method achieve an impressive advantage on the novel classes, with significant improvements of " + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "content": " compared to the previous state-of-the-art method. Many previous methods emphasize restricting catastrophic forgetting by boosting the mIoU of larger-proportion old classes in specific tasks (e.g. 15-1) to make the overall mIoU appear to be higher. However, our approach not only retains the ability to segment old classes but also significantly improves the accuracy of novel classes, aligning more closely with the incremental objective emphasized in CISS." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Comparison on ADE20K. We evaluate our method on the sub-tasks of more challenging ADE20K, including 100-50, 100-10, 100-5, and 50-50, as shown in Tab. 2. Although our method is slightly lower than the state-of-the-art method in short-term sub-tasks, we are remarkably ahead of other methods in more challenging long-term 100-10 and 100-5 tasks. The performance in the novel classes is still outstanding, which demonstrates the universality of our method." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Background Adaptation with Residual Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 158, + 491, + 304 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Quantitative comparison with state-of-the-art exemplar-free methods on ADE20K in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 158, + 491, + 304 + ], + "lines": [ + { + "bbox": [ + 132, + 158, + 491, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 491, + 304 + ], + "type": "table", + "html": "
Method100-50 (2 steps)100-10 (6 steps)50-50 (3 steps)100-5 (11 steps)
0-100101-150all0-100101-150all0-5051-150all0-100101-150all
MiB [6]40.517.232.838.211.129.245.621.029.336.05.726.0
PLOP [21]41.914.932.940.513.631.648.821.030.439.17.828.8
RCIL [58]42.318.834.539.317.632.148.325.032.538.511.529.6
EWF [56]41.221.334.641.516.633.2---41.413.432.1
AWT [25]40.924.735.639.121.433.246.627.033.538.616.031.1
GSC [17]42.419.234.840.816.232.646.226.433.0---
IDEC [63]42.018.234.140.317.632.747.426.033.139.214.631.0
DKD [3]42.422.936.041.619.534.348.826.333.9---
Ours42.023.035.741.123.135.247.926.533.740.521.234.1
Joint44.328.238.944.328.238.951.133.338.944.328.238.9
", + "image_path": "82e5faaac809b7c7908c01cb9f183ad5261857483af98a6d0f98874583e0a5bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 327, + 246, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 246, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 246, + 338 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": "Component Ablations. In this section, we analyze our proposed components, including the Background Adaptation mechanism, Background Feature Distillation loss " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BFD}" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": ", Group Knowledge Distillation loss " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GKD}" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": ", Background Adaptation losses " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BgA^{-}}" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BgA^{+}}" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": ". The experiments are carried out on the representative sub-task 15-1 of Pascal VOC 2012. We separately evaluate different combinations of components and present the results in Tab. 3. The baseline method applies pseudo-labeling, " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PB-BCE}" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": ", and freezes the whole network except the classifiers of background and novel classes. Under the same experimental settings, the Background Adaptation mechanism significantly promotes " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": " of overall mIoU, demonstrating its effectiveness in retaining the knowledge of old classes, and more notably, improving the performance of novel classes. The usage of distillation strategies with " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BFD}" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GKD}" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": " preserves the performance of old classes while further boosting mIoU by " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "6.0\\%" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": " on novel classes. Moreover, the precise supervision of the Background Adaptation using " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BgA^{-}}" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BgA^{+}}" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": " once again enhances the segmentation of novel classes with an additional " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "9.4\\%" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": " of mIoU. Subsequently, there is a growth of " + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 130, + 347, + 482, + 587 + ], + "type": "text", + "content": " on the overall mIoU, which proves that additional precise supervision is necessary. The results above validate the effectiveness of the Background Adaptation mechanism and other related losses. There is a notable improvement, particularly in the performance of novel classes, exceeding twice that of the baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Background Adaptation Schemes. Sec. 4.3 has presented the importance of the Background Adaptation losses. We have devised multiple schemes to achieve more optimal results of Background Adaptation, as shown in Tab. 4. Initially, we aim to have the Background Adaptation value for the current ground truth background region equal to 0, using Mean Square Error (MSE) loss to minimize" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 168, + 409, + 258 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 3: Ablation Study on the components of our proposed method on 15-1. BgA denotes Background Adaptation. In particular, the experiments " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " distillation losses (i.e. " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BFD}" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GKD}" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": ") freeze the whole network except the classifier " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "f^t" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " following the settings in [8]. All experiments contain " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PB-BCE}" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 202, + 168, + 409, + 258 + ], + "lines": [ + { + "bbox": [ + 202, + 168, + 409, + 258 + ], + "spans": [ + { + "bbox": [ + 202, + 168, + 409, + 258 + ], + "type": "table", + "html": "
BgA\\( \\mathcal{L}_{BFD} \\)\\( \\mathcal{L}_{GKD} \\)\\( \\mathcal{L}_{BgA^{-}} \\)\\( \\mathcal{L}_{BgA^{+}} \\)mIoUnimIoUincmIoU
73.421.861.1
✓77.430.566.3
✓✓76.737.167.3
✓✓✓77.336.567.6
✓✓✓✓77.143.469.0
✓✓✓✓76.634.166.5
✓✓✓✓✓77.645.970.0
", + "image_path": "d22940dadb8761178176141fc53c45f8d7cda9f502c1ee321f928bc0e7a6e67e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 159, + 304, + 278, + 367 + ], + "blocks": [ + { + "bbox": [ + 131, + 273, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 131, + 273, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 131, + 273, + 482, + 296 + ], + "type": "text", + "content": "Table 4: Ablation study on schemes of Table 5: Ablation Study on the feature Background Adaptation. distillation (FD) strategies." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 304, + 278, + 367 + ], + "lines": [ + { + "bbox": [ + 159, + 304, + 278, + 367 + ], + "spans": [ + { + "bbox": [ + 159, + 304, + 278, + 367 + ], + "type": "table", + "html": "
FiltermIoUinimIoUincmIoU
MSE-0X74.241.866.5
✓74.642.266.9
BCE-1✓77.444.769.6
OursX65.745.460.8
✓77.645.970.0
", + "image_path": "925fba54260f49aca97d4ce616742aa823bdd61cfdaa1c7e7a423639b75150dc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 335, + 305, + 455, + 367 + ], + "blocks": [ + { + "bbox": [ + 335, + 305, + 455, + 367 + ], + "lines": [ + { + "bbox": [ + 335, + 305, + 455, + 367 + ], + "spans": [ + { + "bbox": [ + 335, + 305, + 455, + 367 + ], + "type": "table", + "html": "
mIoUinimIoUincmIoU
w/o FD56.933.251.2
LKD70.138.362.6
LMSE77.544.369.6
LBFD77.645.970.0
", + "image_path": "9e692f94a2114a567603de853113a42a6767d1ddaa3e63428085f25b4a7c9619.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 395, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 540 + ], + "type": "text", + "content": "the deviation from 0 (MSE-0). However, MSE-0 couldn't restrict the background value to absolute 0, leading to noise accumulation and a lower mIoU compared to Background Adaptation " + }, + { + "bbox": [ + 130, + 395, + 482, + 540 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 130, + 395, + 482, + 540 + ], + "type": "text", + "content": " knowledge distillation. Later, we incorporate considerations for the current class region and introduce a Binary Cross-Entropy (BCE) loss (BCE-1). The expectation is to have the value in the current class region as small as possible, while the value in the background class region is as large as possible. Besides, we apply the Filter to remove the noise above 0. The BCE-1 scheme with Filter improved mIoU by " + }, + { + "bbox": [ + 130, + 395, + 482, + 540 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 395, + 482, + 540 + ], + "type": "text", + "content": ", but supervision in the background region tends to overfit to background objects. Therefore, we retain the BCE loss in the novel class region and redesign a triplet loss as in Sec. 3.3. The final Background Adaptation losses further enhance the segmentation capability for novel classes, resulting in a " + }, + { + "bbox": [ + 130, + 395, + 482, + 540 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 130, + 395, + 482, + 540 + ], + "type": "text", + "content": " improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Background Feature Distillation. To validate the effectiveness of Background Feature Distillation loss " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BFD}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " in Sec. 3.3, we compare " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BFD}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " with the method " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " feature distillation (" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " FD), " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " standard knowledge distillation (" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KD}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " mean square error (" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MSE}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "). As shown in Tab. 5, our " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BFD}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " outperforms other representative distillation methods. Particularly, the Background distillation with MSE, i.e., " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BFD}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", shows a " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " promotion of mIoU on novel classes with no decline in old classes compared to " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MSE}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ". This demonstrates the necessity of preserving a tunnel for learning novel classes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Background Adaptation with Residual Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 451, + 342 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 451, + 342 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 451, + 342 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 451, + 342 + ], + "type": "image", + "image_path": "8ce6c0bc2e05004791d7289fed033d4fbefd6fdae47a6344232430c3d2348247.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "type": "text", + "content": "Fig. 4: Qualitative analysis on 15-1 of Pascal VOC 2012. The Baseline and Ours " + }, + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "type": "text", + "content": " BgA losses are in the " + }, + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "type": "text", + "content": " row and the " + }, + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "type": "text", + "content": " row of Tab. 3. Ours applies additional Background Adaptation losses, compared to the Ours " + }, + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 132, + 352, + 482, + 384 + ], + "type": "text", + "content": " BgA losses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 406, + 263, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 263, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 263, + 418 + ], + "type": "text", + "content": "4.4 Qualitative Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "text", + "content": "We've conducted a qualitative analysis of our approach with Baseline and Ours " + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "text", + "content": " BgA losses in the " + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "text", + "content": " row and the " + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "text", + "content": " row of Tab. 3. The results shown in Fig. 4 are visualized from the 15-1 sub-task of VOC 2012. Our approach in the last three columns reduces the misclassification (e.g., train in the first two columns of each method) and enhances the precision (e.g., sofa in the last column of each method) of the novel classes, while maintaining segmentation ability on the old classes. The results showcase the stability and plasticity of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 523, + 226, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 226, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 226, + 536 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We have introduced a novel CISS method featuring a Background Adaptation mechanism, designed to prevent the continual shifting of a shared background class and promote the learning of novel classes. The mechanism explicitly models the background residual rather than the background at each step and aggregates these residuals to represent the evolving background. To further enhance the performance of the Background Adaptation, we customized corresponding losses on the adaptation channel and improved the distillation scheme. Experiments on our proposed components verify their effectiveness under the Background Adaptation mechanism. Moreover, our method surpasses previous state-of-the-art methods in most sub-tasks, notably excelling in performance of novel classes." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 187 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. 61972036) and the Industry-University-Institute Cooperation Foundation of the Eighth Research Institute of China Aerospace Science and Technology Corporation (No. SAST2022-049)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "text", + "content": "1. Ahn, H., Cha, S., Lee, D., Moon, T.: Uncertainty-based continual learning with adaptive regularization. NeurIPS 32 (2019) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "text", + "content": "2. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. NeurIPS 32 (2019) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 304 + ], + "type": "text", + "content": "3. Baek, D., Oh, Y., Lee, S., Lee, J., Ham, B.: Decomposed knowledge distillation for class-incremental semantic segmentation. NeurIPS 35, 10380-10392 (2022) 2, 4, 9, 10, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 306, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 480, + 327 + ], + "type": "text", + "content": "4. Borsos, Z., Mutny, M., Krause, A.: Coresets via bilevel optimization for continual learning and streaming. NeurIPS 33, 14879-14890 (2020) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "type": "text", + "content": "5. Boschini, M., Bonicelli, L., Porrello, A., Bellitto, G., Pennisi, M., Palazzo, S., Spampinato, C., Calderara, S.: Transfer without forgetting. In: ECCV. pp. 692-709 (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "text", + "content": "6. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: CVPR. pp. 9233-9242 (2020) 2, 4, 9, 10, 11, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 393, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 414 + ], + "type": "text", + "content": "7. Cha, H., Lee, J., Shin, J.: Co2l: Contrastive continual learning. In: ICCV. pp. 9516-9525 (2021) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "text", + "content": "8. Cha, S., Yoo, Y., Moon, T., et al.: Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. NeurIPS 34, 10919-10930 (2021) 2, 3, 4, 6, 7, 10, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "9. Chen, J., Cong, R., Luo, Y., Ip, H., Kwong, S.: Saving 100x storage: Prototype replay for reconstructing training sample distribution in class-incremental semantic segmentation. Advances in Neural Information Processing Systems 36 (2024) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 480, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 480, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 133, + 480, + 480, + 512 + ], + "type": "text", + "content": "10. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062 (2014) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 513, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 513, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 133, + 513, + 480, + 545 + ], + "type": "text", + "content": "11. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI 40(4), 834-848 (2017) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 546, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 546, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 133, + 546, + 480, + 567 + ], + "type": "text", + "content": "12. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017) 1, 4, 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 600 + ], + "type": "text", + "content": "13. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: ECCV. pp. 801-818 (2018) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 601, + 434, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 434, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 434, + 610 + ], + "type": "text", + "content": "14. Chen, Z., Liu, B.: Lifelong machine learning, vol. 1. Springer (2018) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "text", + "content": "15. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022) 1, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "16. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. NeurIPS 34, 17864-17875 (2021) 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Background Adaptation with Residual Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Cong, W., Cong, Y., Dong, J., Sun, G., Ding, H.: Gradient-semantic compensation for incremental semantic segmentation. arXiv preprint arXiv:2307.10822 (2023) 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "text", + "content": "18. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009) 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 162, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 184 + ], + "type": "text", + "content": "19. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: CVPR. pp. 5138-5146 (2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 185, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 216 + ], + "type": "text", + "content": "20. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 218, + 480, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 250 + ], + "type": "text", + "content": "21. Douillard, A., Chen, Y., Dapogny, A., Cord, M.: Plop: Learning without forgetting for continual semantic segmentation. In: CVPR. pp. 4040-4050 (2021) 2, 3, 4, 6, 7, 9, 10, 11, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 273 + ], + "type": "text", + "content": "22. Ermis, B., Zappella, G., Wistuba, M., Rawal, A., Archambeau, C.: Memory efficient continual learning with transformers. NeurIPS 35, 10629-10642 (2022) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "type": "text", + "content": "23. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. IJCV 88, 303-338 (2010) 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "text", + "content": "24. Fang, Y., Zhu, F., Cheng, B., Liu, L., Zhao, Y., Wei, Y.: Locating noise is halfway denoising for semi-supervised segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16612-16622 (2023) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 330, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 373 + ], + "type": "text", + "content": "25. Goswami, D., Schuster, R., van de Weijer, J., Stricker, D.: Attribution-aware weight transfer: A warm-start initialization for class-incremental semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3195-3204 (2023) 11, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "26. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: ICCV. pp. 991-998 (2011) 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "27. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) 1, 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 419, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 440 + ], + "type": "text", + "content": "28. Jung, H., Ju, J., Jung, M., Kim, J.: Less-forgetting learning in deep neural networks. arXiv preprint arXiv:1607.00122 (2016) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 441, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 474 + ], + "type": "text", + "content": "29. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13), 3521-3526 (2017) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 475, + 480, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 496 + ], + "type": "text", + "content": "30. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. NeurIPS 25 (2012) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 498, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 518 + ], + "type": "text", + "content": "31. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 520, + 480, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 480, + 541 + ], + "type": "text", + "content": "32. Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI 40(12), 2935-2947 (2017) 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 543, + 480, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 574 + ], + "type": "text", + "content": "33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012-10022 (2021) 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "text", + "content": "34. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015) 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "text", + "content": "35. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. NeurIPS 30 (2017) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 621, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 642 + ], + "type": "text", + "content": "36. Madaan, D., Yoon, J., Li, Y., Liu, Y., Hwang, S.J.: Representational continuity for unsupervised continual learning. In: ICLR (2022) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "37. Maracani, A., Michieli, U., Toldo, M., Zanuttigh, P.: Recall: Replay-based continual learning in semantic segmentation. In: ICCV. pp. 7026-7035 (2021) 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "38. McClelland, J.L., McNaughton, B.L., O'Reilly, R.C.: Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review 102(3), 419 (1995) 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 195 + ], + "type": "text", + "content": "39. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989) 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "text", + "content": "40. Michieli, U., Zanuttigh, P.: Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In: CVPR. pp. 1114-1124 (2021) 2, 4, 10, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "type": "text", + "content": "41. Oh, Y., Baek, D., Ham, B.: Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. NeurIPS 35, 14516-14528 (2022) 3, 4, 6, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "text", + "content": "42. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: CVPR. pp. 11321-11329 (2019) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "text", + "content": "43. Pham, Q., Liu, C., Hoi, S.: Dualnet: Continual learning, fast and slow. NeurIPS 34, 16131-16144 (2021) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 330 + ], + "type": "text", + "content": "44. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 352 + ], + "type": "text", + "content": "45. Ritter, H., Botev, A., Barber, D.: Online structured laplace approximations for overcoming catastrophic forgetting. NeurIPS 31 (2018) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "text", + "content": "46. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 234-241 (2015) 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "text", + "content": "47. Schwarz, J., Czarnecki, W., Luketina, J., Grabska-Barwinska, A., Teh, Y.W., Pascanu, R., Hadsell, R.: Progress & compress: A scalable framework for continual learning. In: Proc. International Conference on Machine Learning. pp. 4528-4537. PMLR (2018) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 431, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 475 + ], + "type": "text", + "content": "48. Shang, C., Li, H., Meng, F., Wu, Q., Qiu, H., Wang, L.: Incrementer: Transformer for class-incremental semantic segmentation with knowledge distillation focusing on old class. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7214-7224 (2023) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "type": "text", + "content": "49. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. NeurIPS 30 (2017) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "type": "text", + "content": "50. Shon, H., Lee, J., Kim, S.H., Kim, J.: Dlcft: Deep linear continual fine-tuning for general incremental learning. In: ECCV. pp. 513-529 (2022) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 521, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 542 + ], + "type": "text", + "content": "51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 576 + ], + "type": "text", + "content": "52. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "text", + "content": "53. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: Proc. International Conference on Machine Learning. pp. 10347-10357 (2021) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "55. Wu, C., Herranz, L., Liu, X., Van De Weijer, J., Raducanu, B., et al.: Memory replay gans: Learning to generate new categories without forgetting. NeurIPS 31 (2018) 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Background Adaptation with Residual Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 468 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "56. Xiao, J.W., Zhang, C.B., Feng, J., Liu, X., van de Weijer, J., Cheng, M.M.: Endpoints weight fusion for class incremental semantic segmentation. In: CVPR. pp. 7204-7213 (2023) 4, 9, 11, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "57. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Nabi, M., Alameda-Pineda, X., Ricci, E.: Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE TPAMI 45(2), 2567-2581 (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "text", + "content": "58. Zhang, C.B., Xiao, J.W., Liu, X., Chen, Y.C., Cheng, M.M.: Representation compensation networks for continual semantic segmentation. In: CVPR. pp. 7053-7064 (2022) 4, 11, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 257 + ], + "type": "text", + "content": "59. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19148-19158 (2023) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "60. Zhang, J.O., Sax, A., Zamir, A., Guibas, L., Malik, J.: Side-tuning: a baseline for network adaptation via additive side networks. In: ECCV. pp. 698-714 (2020) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "text", + "content": "61. Zhang, Z., Gao, G., Fang, Z., Jiao, J., Wei, Y.: Mining unseen classes via regional objectness: A simple baseline for incremental segmentation. NeurIPS 35, 24340-24353 (2022) 2, 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "62. Zhang, Z., Gao, G., Jiao, J., Liu, C.H., Wei, Y.: Coinseg: Contrast inter-and intra-class representations for incremental segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 843-853 (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "63. Zhao, D., Yuan, B., Shi, Z.: Inherit with distillation and evolve with contrast: Exploring class incremental semantic segmentation without exemplar memory. IEEE TPAMI (2023) 4, 9, 11, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "64. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: CVPR. pp. 633-641 (2017) 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 445 + ], + "type": "text", + "content": "65. Zhu, H., Wei, Y., Liang, X., Zhang, C., Zhao, Y.: Ctp: Towards vision-language continual pretraining via compatible momentum contrast and topology preservation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22257-22267 (2023) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "66. Zhu, L., Chen, T., Yin, J., See, S., Liu, J.: Continual semantic segmentation with automatic memory sample selection. In: CVPR. pp. 3082-3092 (2023) 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_content_list.json b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc72fa8913e662cd01807a0ec5e5135c7138e3a5 --- /dev/null +++ b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "Bad Students Make Great Teachers: Active Learning Accelerates Large-Scale Visual Understanding", + "text_level": 1, + "bbox": [ + 248, + 191, + 754, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Talfan Evans1,* Shreya Pathak1,* Hamza Merzic1,2,* Jonathan Schwarz1,† Ryutaro Tanno1 Olivier J. Henaff1,*", + "bbox": [ + 300, + 276, + 695, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Google DeepMind $^{2}$ University College London", + "bbox": [ + 344, + 319, + 657, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Power-law scaling indicates that large-scale training with uniform sampling is prohibitively slow. Active learning methods aim to increase data efficiency by prioritizing learning on the most relevant examples. Despite their appeal, these methods have yet to be widely adopted since no one algorithm has been shown to a) generalize across models and tasks b) scale to large datasets and c) yield overall FLOP savings when accounting for the overhead of data selection. In this work we propose a method which satisfies these three properties, leveraging small, cheap proxy models to estimate \"learnability\" scores for datapoints, which are used to prioritize data for training much larger models. As a result, models trained using our methods - ClassAct and ActiveCLIP - require $46\\%$ and $51\\%$ fewer training updates and up to $25\\%$ less total computation to reach the same performance as uniformly-trained visual classifiers on JFT and multimodal models on ALIGN, respectively. Finally, we find our data-prioritization scheme to be complementary with recent data-curation and learning objectives, yielding a new state-of-the-art in several multimodal transfer tasks.", + "bbox": [ + 273, + 358, + 727, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 233, + 580, + 383, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Power-law scaling for vision and language models [19,43] indicates that incremental improvements in model performance require orders of magnitude increases in computation. One of the key features of these empirical power-laws is that training data is sampled uniformly. In contrast, active data selection prioritizes computation on the data that maximally contributes to task performance [21,24,34], with the ultimate goals of improving data efficiency and reducing the cost of training. However, active data selection has yet to become a mainstay of large model training, since no existing algorithm satisfies the three properties of being a) robust to the choice of model and training task, b) scalable to large datasets and architectures, and c) more compute efficient end-to-end than training with uniform samples.", + "bbox": [ + 231, + 605, + 771, + 748 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal technical contribution. Email correspondence to and . †Current affiliation: Harvard University, work done while at Google DeepMind.", + "bbox": [ + 233, + 754, + 771, + 792 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b276811fb9999805e31e07d2fd1e3f9e55823a6a65f03efcfa09c4a679132726.jpg", + "image_caption": [ + "Fig. 1: Active learning accelerates large-scale visual understanding. For large-scale classification and multimodal learning tasks, prioritised training on data selected using our active selection methods ClassAct (left) and ActiveCLIP (right) requires significantly fewer updates to reach the final performance of uniform training." + ], + "image_footnote": [], + "bbox": [ + 236, + 195, + 500, + 280 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/976b696fe455fa1da0a8f10fbcf0eadb744dc221dd135b88e425ea61e7dd4035.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 502, + 195, + 764, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first instance, data selection based on hand-engineered filters (e.g. removing incorrectly shaped images or that only contain a single colour [2]) can trivially improve training efficiency at minimal computational overhead. However, such heuristics are limited in their effectiveness by the expertise of the human designer, and are not guaranteed to transfer to training of different models, data modalities, or tasks and incur significant effort to develop and tune.", + "bbox": [ + 231, + 346, + 769, + 423 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, model-based curation methods, which use the loss of the model itself to score examples, have shown promise by focusing training on 'hard' and omitting 'easy' data [35], but also the opposite (to exclude noise or other low-quality data) in both language modeling [13] and multimodal learning [15]. However, these methods often spend as much computation on the curation of datasets as is gained from subsequent pretraining, making them less compute-efficient than training on uniformly-sampled data. Finally, while several compute-efficient methods have been successfully deployed at small scale [8], they generally do not scale to even medium-sized datasets such as ImageNet.", + "bbox": [ + 231, + 423, + 769, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose an algorithm that satisfies the three properties of generality, scalability, and compute-positivity. The proposed framework uses small proxy models to compute learnability scores for candidate training data, resulting in significant training efficiency gains at an almost negligible overhead over standard uniform-training. We test two instantiations of this framework, ClassAct and ActiveCLIP / ActiveSigLIP, for large scale classification and multimodal pretraining, respectively. Our findings are summarized below:", + "bbox": [ + 231, + 541, + 769, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benchmarking heuristics for large-scale pretraining: We investigate loss- and learnability-based prioritization [15, 27, 35] for large-scale classification and find that pretrained reference models are an essential component for accelerating learning, producing efficiency gains of up to $46\\%$ .", + "bbox": [ + 231, + 637, + 769, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalizing selection policies across model scale: Secondly, we show that smaller models act as effective proxies for much larger ( $\\sim 1000 \\times$ ) models in the context of learnability but not loss-based scoring, resulting in compute-positive gains of up to $25\\%$ over uniform training, a first in the context of large-scale pretraining.", + "bbox": [ + 231, + 695, + 769, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accelerating multimodal pretraining: Using reference models pretrained on small, clean datasets, we substantially accelerate pretraining on much larger,", + "bbox": [ + 231, + 764, + 769, + 793 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 233, + 169, + 243, + 179 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Evans et al.", + "bbox": [ + 284, + 169, + 379, + 179 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "noisier datasets. Moreover, we find ActiveCLIP to be complementary to recent data-curation techniques [12] and learning objectives [44], yielding a new state-of-the-art in several multimodal understanding tasks.", + "bbox": [ + 231, + 195, + 769, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Amortizing data selection policies: Data-selection policies trained on one task can also accelerate the training of subsequent models on different but related tasks, suggesting that such policies can be easily derived from pre-trained models.", + "bbox": [ + 231, + 241, + 771, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Simplifying active-learning with online reference models: Lastly, we demonstrate that pre-trained reference models may not be necessary at all, where these models are small and can be trained in parallel on larger batches than the learner model, while remaining compute-positive.", + "bbox": [ + 231, + 286, + 771, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 233, + 359, + 394, + 374 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data pruning. One approach to data-selection is to identify and sub-select data ahead of training. For example, [28] and [35] show that the training loss and gradients can be used to discard large portions of small-to-medium sized datasets (e.g. CIFAR10 and ImageNet) with little loss in performance. These methods have since been deployed for the curation of web-scale datasets in both language modeling [26] and multimodal learning [1, 12, 25], demonstrating large reductions in the amount of data required together with performance improvements. However, in the single-epoch regime that is becoming typical of large model training [16, 19], pre-filtering can be as expensive as learning from it, a shortcoming which we address in this work. Nevertheless, we show that our method for dynamic data selection is complementary to and benefits from such data curation techniques.", + "bbox": [ + 231, + 390, + 771, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online active learning. Unlike data-pruning, online active learning continuously filters data throughout training and applies naturally to the semi-infinite, single-epoch regime. Online Batch Selection [22] scores and filters using the learner model, which has the theoretical advantage that the importance of data can be determined relative to the current state of the learner. In terms of metrics, the Reducible Holdout Loss (RHO) [27] also uses the concept of a reference model to identify learnable data points not yet well represented. Other proposed heuristics include memorization for long-tailed data [11] and assigning \"complexity\" scores based on the number of times the example is forgotten during training [38]. None of these approaches however have demonstrated that the cost of scoring can be reduced to the point of justifying learner efficiency gains.", + "bbox": [ + 231, + 552, + 771, + 697 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compute-efficient data selection. Several works have demonstrated the benefits of selecting data based on simple heuristics, such as low-level image properties [2] or proximity to high-quality text corpora [3, 7, 29, 40]. While cheap to compute, these statistics often require domain-specific knowledge which limits their applicability across tasks. Domain-agnostic methods such as core-sets alleviate this by selecting data based on the geometry of their embeddings [4, 14] which can be efficiently computed, however these algorithms generally do not", + "bbox": [ + 231, + 701, + 771, + 794 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Active Learning Accelerates Large-Scale Visual Understanding", + "bbox": [ + 326, + 168, + 719, + 181 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 757, + 169, + 769, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scale to large-scale datasets [8]. Most related to our work is DoReMi [39] which uses domain-general, scalable, and compute-efficient proxy models for the simpler problem of determining optimal data-mixtures for the subsequent training of a larger language model.", + "bbox": [ + 231, + 195, + 771, + 248 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 231, + 266, + 347, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Data selection as prioritized replay", + "text_level": 1, + "bbox": [ + 231, + 290, + 547, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use online batch selection [22] to apply our scoring heuristics to standard visual learning tasks: firstly, we sample uniformly from the training set $\\boldsymbol{x}_i \\stackrel{\\mathcal{U}}{\\sim} \\mathcal{D}$ and compute a score $s_i = s(\\boldsymbol{x}_i | \\theta) \\in \\mathbb{R}$ to each data point $\\boldsymbol{x}_i$ using model parameters $\\theta$ . Given a large enough collection of scored examples stored in a memory bank $\\mathcal{M} = \\{\\boldsymbol{x}_i\\}_{i \\in (0, \\dots, M-1)}$ , we then sample non-uniformly according to their scores $\\boldsymbol{x}_i \\stackrel{\\pi}{\\sim} \\mathcal{M}$ [31], where $\\pi(\\boldsymbol{x}_i) = \\text{Softmax}(\\{s_i\\}_{i \\in (0, \\dots, M-1)})$ . A batch of such examples is used to update the learner model. Following convention in reinforcement learning, we refer to the scoring and target models as actors and learners respectively.", + "bbox": [ + 231, + 311, + 771, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Statistics for data selection", + "text_level": 1, + "bbox": [ + 231, + 451, + 485, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We explore a few statistics for model-based prioritization, grouped into two categories.", + "bbox": [ + 231, + 471, + 769, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Example difficulty: given the current state of the learner, an intuitive prioritization scheme might favour 'difficult' examples (as measured by their training loss), while removing 'easy' examples that are trivially classified and which yield small gradients. This loss-based prioritization:", + "bbox": [ + 231, + 502, + 769, + 554 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {\\text {h a r d}} \\left(\\boldsymbol {x} _ {i} \\mid \\theta\\right) = \\ell \\left(\\boldsymbol {x} _ {i} \\mid \\theta\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 562, + 768, + 576 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "can use the current parameters of the learner $\\theta^t$ or those of a fixed model $\\theta^{*}$ . The opposite argument can be made for favoring examples that are easily solved by a well-trained model, as such a prioritization removes the noisy examples present in large-scale datasets:", + "bbox": [ + 231, + 584, + 769, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {\\text {e a s y}} \\left(\\boldsymbol {x} _ {i} | \\theta\\right) = - \\ell \\left(\\boldsymbol {x} _ {i} | \\theta\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 644, + 768, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This scheme is commonly used in multimodal learning for identifying high-quality examples with pre-trained models [15,32,33].", + "bbox": [ + 231, + 665, + 769, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Example learnability: Given that favoring easy and hard examples target different and potentially orthogonal properties of the data, a natural question is whether these policies can be combined. Learnability criteria straightforwardly combine the two as", + "bbox": [ + 231, + 697, + 769, + 748 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s ^ {\\text {l e a r n}} \\left(\\boldsymbol {x} _ {i} | \\theta^ {t}, \\theta^ {*}\\right) = s ^ {\\text {h a r d}} \\left(\\boldsymbol {x} _ {i} | \\theta^ {t}\\right) + s ^ {\\text {e a s y}} \\left(\\boldsymbol {x} _ {i} | \\theta^ {*}\\right) (3) \\\\ = \\ell \\left(\\boldsymbol {x} _ {i} \\mid \\theta^ {t}\\right) - \\ell \\left(\\boldsymbol {x} _ {i} \\mid \\theta^ {*}\\right), (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 757, + 768, + 788 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 233, + 169, + 245, + 179 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Evans et al.", + "bbox": [ + 284, + 169, + 379, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "favoring examples that are easily solved by a well-trained model $\\theta^{*}$ but challenging to the learner in its current state $\\theta^t$ , such that more computation dedicated to this example could lower its loss. Conversely, examples that are trivially classified by the learner (or mislabeled) will yield low (or high) losses for both the current learner and the well-trained one, leading to low learnability scores.", + "bbox": [ + 231, + 195, + 771, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A special case of learnability scores (the RHO loss, [27]) uses a model $\\theta^{\\mathrm{ho}}$ specifically trained on a held-out dataset to ensure the independence of its predictions from those of the current learner $s^{\\mathrm{learn}}(\\pmb{x}_i|\\theta^t,\\theta^{\\mathrm{ho}})$ . We assess in Section 4.2 whether this is necessary when training on large-scale image datasets.", + "bbox": [ + 231, + 260, + 771, + 313 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6dc4c9e120688085749e7bef4da43f4d298326e970f431a67df88065169b146e.jpg", + "image_caption": [ + "Fig. 2: Amortizing the cost of data selection. Drawn to scale: length of bars indicates number of FLOPs required to reach the accuracy of a ViT-L trained with uniform sampling (\"ViT-L Uniform Sampling\", see Figure 4). Expensive model policies (e.g. a ViT-B scores data for the ViT-L learner, or $\\mathrm{B} \\rightarrow \\mathrm{L}$ ) produce large learner speedups, at the expense of the additional FLOPs associate with data selection. This overhead can be reduced by deriving the data-selection policies from smaller models (e.g. ViT-S, ViT-Ti or ViT-Mu score data for the ViT-L learner), at the expense of marginal decreases in the learner speedup. Costs could be additionally amortized by using off-the-shelf reference models, removing the need to train from scratch (yellow). Since the reference model is fixed throughout training, scores can be assigned once to a 'foundation dataset' and amortized across many training runs (lime green; [27,35]). Since the online model is independent of the learner model and generalizes across scale, data selection policies can also be distilled as a fixed ordering of a given dataset (a 'foundation curriculum')." + ], + "image_footnote": [], + "bbox": [ + 238, + 339, + 766, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Unlocking compute-positive training", + "text_level": 1, + "bbox": [ + 231, + 623, + 559, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scoring requires inference passes over the actor and learner models. We assume that gradient updates to cost $3\\mathrm{x}$ inference passes. The cost of scoring data $F$ thus scales with the proportion of data which is being rejected (e.g. retaining only $20\\%$ of the data requires 5 inference passes per trained batch). The requirements for compute-positivity can therefore be expressed as:", + "bbox": [ + 231, + 644, + 771, + 709 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\left(3 F _ {\\text {l e a r n}} + \\rho F _ {\\text {a c t}}\\right) \\beta + 3 F _ {\\text {r e f}}} _ {\\text {A c t i v e L e a r n i n g}} < \\underbrace {3 F _ {\\text {l e a r n}}} _ {\\text {U n i f o r m S a m p l i n g}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 718, + 769, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{\\mathrm{act}}$ is the cost of scoring an example, $\\rho$ is the number of examples scored per training example, and $\\beta$ is the efficiency gain (in terms of the learner update)", + "bbox": [ + 231, + 764, + 769, + 792 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Active Learning Accelerates Large-Scale Visual Understanding", + "bbox": [ + 326, + 168, + 719, + 181 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 757, + 169, + 769, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "relative to training with uniform sampling (see Appendix Section A.1 for more details). The RHS term is the cost per update of uniform sampling. The first LHS term inside the brackets is the cost of the learner training during AL, the second term is the scoring cost and the last term outside the brackets the cost of training the reference model. We illustrate in Figure 2 the different contexts in which parts of this computation may be effectively amortized.", + "bbox": [ + 231, + 195, + 768, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the large-scale training regime where data is neither repeated nor seen before, compute-positivity requires that either or all of the reference model training, actor scoring and learner efficiency $\\beta$ terms must be made smaller to produce net savings relative to uniform sampling. Typical prioritization schemes can produce saving on the order of $50\\%$ (ie $\\beta = 0.5$ ), suggesting that savings must also be made by down-scaling the other terms $F_{\\mathrm{act}}$ and $F_{\\mathrm{ref}}$ .", + "bbox": [ + 231, + 275, + 769, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cost of easy-reference scoring. While both the cost of the reference model and example scoring can be scaled down in the case of easy-reference scoring with small models ( $F_{\\mathrm{act}} = F_{\\mathrm{ref}}$ , see Equation 2), it is unclear whether the efficiency gains $\\beta$ are robust to this down-scaling (see section 4.2).", + "bbox": [ + 231, + 359, + 769, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cost of RHO learnability scoring. The original definition of learnability scores [27] requires inference passes through both the learner and a reference model ( $F_{\\mathrm{act}} = F_{\\mathrm{ref}} + F_{\\mathrm{learn}}$ , see Equation 3), meaning that although the cost of the reference model can be reduced by using a smaller model, the cost of example scoring cannot.", + "bbox": [ + 231, + 416, + 769, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cost of ClassAct / ActiveCLIP. For this reason, we explore whether replacing the learner model in term 1 of Equation 3 with a much smaller model can still produce comparable learner efficiency gains to those already observed (see Appendix Algorithm 1). Specifically, we introduce a third \"online\" model, which has the same architecture and size as the reference model, but is trained in parallel with the learner. In this case, the cost of scoring examples reduces to:", + "bbox": [ + 231, + 489, + 769, + 568 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {a c t}} = F _ {\\text {r e f}} + F _ {\\text {o n l i n e}} = 2 F _ {\\text {r e f}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 585, + 768, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and can be scaled down along with the reference model. We instantiate our method for two canonical pre-training tasks: visual classification and multimodal learning, which we call ClassAct, ActiveCLIP and ActiveSigLIP respectively. In Appendix Section A.6, we describe an asynchronous active learning framework where scoring and learning is performed in parallel on separate devices. This setup does not affect the FLOP calculations but can mitigate the overhead in time even when using non-approximate scoring.", + "bbox": [ + 231, + 608, + 769, + 701 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Losses for canonical visual pre-training tasks", + "text_level": 1, + "bbox": [ + 231, + 725, + 623, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For visual classification with ClassACT, we use the standard cross-entropy loss for both actors and learners. For multimodal learning with ActiveCLIP, learners optimize the contrastive loss $\\ell_{\\mathrm{learn}} = \\ell_{\\mathrm{learn}}^{\\mathrm{im,txt}} + \\ell_{\\mathrm{learn}}^{\\mathrm{txt,im}}$ , whereas the actor loss $\\ell_{\\mathrm{act}}$", + "bbox": [ + 231, + 752, + 769, + 795 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 233, + 169, + 243, + 178 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Evans et al.", + "bbox": [ + 284, + 169, + 379, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 ClassAct / ActiveCLIP", + "text_level": 1, + "bbox": [ + 233, + 195, + 492, + 209 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: Randomly initialized learner model $\\theta_{l}$ and small online model $\\theta_{o}$ , small pre-trained reference model $\\theta_{r}$ . Models use loss $\\ell_{\\mathrm{act}}$ for scoring data and $\\ell_{\\mathrm{learn}}$ for computing updates. Dataset $\\mathcal{D}$ , batch size $B$ , sub-batch size $b < B$ .", + "2: while training do", + "3: $X\\sim \\mathcal{D}$ , where $|X| = B$ Sample uniformly", + "4: $S = \\ell_{\\mathrm{act}}(X|\\theta_o) - \\ell_{\\mathrm{act}}(X|\\theta_r)$", + "5: $I\\sim \\mathrm{SoftMax}(S)$ , where $|I| = b$ Sample indices", + "6: $Y = X[I]$ $\\triangleright$ Collect sub-batch", + "7: $\\theta_{l}\\gets \\mathrm{Adam}[\\nabla_{\\theta_{l}}\\ell_{\\mathrm{learn}}(Y|\\theta_{l})]$", + "8: $\\theta_o\\gets \\mathrm{Adam}[\\nabla_{\\theta_o}\\ell_{\\mathrm{learn}}(Y|\\theta_o)]$ Update online model", + "9: end while" + ], + "bbox": [ + 236, + 212, + 771, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is simply the dot-product similarity between image and text embeddings:", + "bbox": [ + 231, + 370, + 727, + 384 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {a c t}} \\left(\\boldsymbol {x} _ {i} \\mid \\theta\\right) = - \\boldsymbol {z} _ {i} ^ {\\mathrm {i m}} \\cdot \\boldsymbol {z} _ {i} ^ {\\mathrm {t x t}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 386, + 390, + 768, + 406 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {l e a r n}} ^ {\\mathrm {i m}, \\mathrm {t x t}} (\\boldsymbol {x} _ {i} | \\theta) = - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {i m}} \\cdot \\boldsymbol {z} _ {i} ^ {\\mathrm {t x t}}\\right)}{\\sum_ {j} \\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {i m}} \\cdot \\boldsymbol {z} _ {j} ^ {\\mathrm {t x t}}\\right)} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 408, + 768, + 440 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\boldsymbol{z}_i^{\\mathrm{im}} = f^{\\mathrm{im}}(\\boldsymbol{x}_i; \\theta)$ and $\\boldsymbol{z}_i^{\\mathrm{txt}} = f^{\\mathrm{txt}}(\\boldsymbol{x}_i; \\theta)$ are image and text embeddings respectively, and $\\ell_{\\mathrm{learn}}^{\\mathrm{txt}, \\mathrm{im}}(\\boldsymbol{x}_i; \\theta)$ is defined analogously to $\\ell_{\\mathrm{learn}}^{\\mathrm{im}, \\mathrm{txt}}(\\boldsymbol{x}_i; \\theta)$ . Similarly, ActiveSigLIP instead uses the sigmoid loss [44] for the learner's objective and $\\ell_{\\mathrm{act}}$ for scoring.", + "bbox": [ + 231, + 447, + 769, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 231, + 520, + 381, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All our experiments were conducted with Vision Transformers [9] for which strong baselines are available across model sizes [43]. Unless specified, we adopt models with patch-size 16 throughout (ViT-S refers to ViT-S/16 and similar). We consider two canonical tasks for large-scale pretraining: classification on JFT-300M [36] and multimodal contrastive learning [30] on large image-text datasets. When pre-training with JFT classification we use held-out classification performance as the evaluation metric. When pre-training on image-text data we evaluate with standard multimodal transfer tasks: ImageNet zero-shot classification and image-to-text / text-to-image retrieval on COCO.", + "bbox": [ + 231, + 545, + 769, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Throughout, we will refer to the large batch of size $B$ sampled uniformly from the training data as the 'super-batch', and the prioritised smaller batch of size $b < B$ as the 'sub-batch'. In all our experiment, we filter 50% of uniformly sampled data such that $\\rho = B / b = 2$ , although more aggressive filtering regimes warrant investigation [35].", + "bbox": [ + 231, + 662, + 769, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Evaluating loss-based scoring heuristics in the large-data regime", + "text_level": 1, + "bbox": [ + 231, + 745, + 769, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We begin by evaluating loss- and learnability-based heuristics on their ability to accelerate supervised classification on JFT (Fig. 3). Arguably the most intuitive", + "bbox": [ + 231, + 764, + 769, + 792 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Active Learning Accelerates Large-Scale Visual Understanding", + "bbox": [ + 326, + 168, + 719, + 181 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 757, + 169, + 769, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd270feee29f95d773fd59a75a31aca907bb97690a0fb075dfbd83db1bb0884e.jpg", + "image_caption": [ + "Fig. 3: Evaluation of loss-based data-selection criteria for large-scale classification. We train a ViT-B on JFT-300M with different data-selection policies. Prioritising hard data under the learner (green curve) produced marginal gains over the uniform sampling baseline. Prioritizing data using both learnability (blue curve, [27]) and easy reference prioritization (red curve, [15]) produced significant speedups and performance gains." + ], + "image_footnote": [], + "bbox": [ + 290, + 201, + 709, + 332 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fda2293d0d429e1180327358be41fc151e8dabef54c06ba30731ca0947e9ce1.jpg", + "image_caption": [ + "Fig. 4: Generalization of data-selection policies across models scales. Left: We train a ViT-L for 3 epochs on JFT using uniform sampling (grey) or prioritized data sampling using example learnability (blue) or low-loss under the reference model (red). Example scores are computed using ViT-B actors (dark), or cheaper ViT-S or ViT-Tiny models (light). While both example learnability and \"easy reference\" yield good speedups with expensive actors, learnability criteria are much more robust to approximate scoring. Top right: Learner (ViT-L) speedup is computed as the fraction of learner iterations saved in order to attain the baseline's top performance. Actor overhead is computed as the additional computation in FLOPs required to score examples with a particular actor architecture (varying from ViT-Mu to ViT-L, see Appendix Table 4). Example learnability yields robust learner speedups across actor scales, \"easy reference\" scoring does not. Lower right: total compute efficiency is calculated as a product of learner efficiency and actor overhead, indicating the amount of computation required to reach baseline performance. Approximate actors (i.e. ViT-S or smaller) computing example learnability enable total compute speedups, other schemes do not." + ], + "image_footnote": [], + "bbox": [ + 236, + 427, + 500, + 589 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/434dd7f91d42456df41f0e7afdf1d4742ab31b1a6f6cef39ed93055f372a7bf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 428, + 766, + 589 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 233, + 169, + 243, + 178 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Evans et al.", + "bbox": [ + 284, + 169, + 379, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method to score data is to prioritise training on data with high loss under the learner (hard learner). In our experiments, this strategy (Eq. 1) only marginally improved performance over the uniform sampling baseline, despite requiring an additional inference pass over the super-batch. This is perhaps not surprising - data points with high loss may also be unlearnable due to e.g. label noise, such that training on those data points does not result in the model performing any better on the held-out test set. Large scale datasets are more likely to be noisy.", + "bbox": [ + 231, + 195, + 771, + 286 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Scoring methods based on pre-trained reference models performed much better—both easy reference (equation 2) and learnability (equation 3) -based prioritization produced significant gains over uniform sampling. Here, we pretrained an identical ViT-B for the same 3 epochs to use as a reference model for a second training run displayed above. producing speed-ups of $33\\%$ (Fig. 3).", + "bbox": [ + 231, + 286, + 771, + 353 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/af975683167a5ea5d947d933a3f525fa4937cf3e5330025768dc300311e5ed53.jpg", + "image_caption": [ + "Fig. 5: Scaling laws for active learning. We trained a baseline ViT-L over a range of compute budgets (for which ViT-L is compute optimal, see Zhai et al., 2021). We also trained the same ViT-L with both ViT-Ti and ViT-S reference policies, pre-trained for the same number of epochs. Left: Small model policies produce robust savings in learner compute. Right: When accounting for total compute (learner + actor training and data scoring), small model policies in all compute budgets produce FLOP savings over training with uniform samples. These scaling laws generalize those measured empirically in the uniform sampling setting [43] to the case of non-uniform data selection." + ], + "image_footnote": [], + "bbox": [ + 236, + 376, + 500, + 488 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4918a10c2f728b7e4d882f333f6af743f3035ac6b058739f4c3c669bca288946.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 376, + 766, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Generalising data-selection policies across scale", + "text_level": 1, + "bbox": [ + 231, + 638, + 643, + 653 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The speed-ups afforded by learnability prioritization (Fig. 3) come at the cost of the additional inference passes required to score the data during learner training, plus the cost of training the reference model. This makes the overall gains strongly compute-negative relative to training with uniform samples. Even if the size of the reference model is scaled down [27], these methods still incur the cost of additional learner inference passes to score the data during training.", + "bbox": [ + 231, + 661, + 769, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Unlocking compute-positive active learning. To address this issue, we introduce a set of down-scaled models with the same ViT architecture that we use to score data for training a larger ViT-L model (see Methods Section 3.3. We use ViT-B, ViT-S or ViT-Ti variants (which are $4 \\times$ , $13 \\times$ and $47 \\times$ cheaper than", + "bbox": [ + 231, + 739, + 769, + 792 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Active Learning Accelerates Large-Scale Visual Understanding", + "bbox": [ + 326, + 168, + 719, + 181 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 757, + 169, + 769, + 179 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/33b8f873928de77211adbd3807ef21c545899669f25a39dd3056894af65855cc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodViT model capacityReference TypeSpeed-up
Reference Online LearnerLearner speedupCompute speedup
Uniform Sampling (ViT-B)B0%0%
RHOTinyBBHeld-out, fixed0%-79%
ClassAct-HOTinyTinyBHeld-out, fixed18%3%
ClassActTinyTinyBIn-domain, fixed18%3%
ClassAct-OnlineTinyTinyBTrained online17%2%
", + "bbox": [ + 233, + 193, + 793, + 295 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Simplifying and accelerating the computation of learnability scores.", + "bbox": [ + 231, + 296, + 769, + 307 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Relative to RHO [27], ClassAct makes two changes: replacing the reference model with one trained in-domain (removing the need for bespoke held-out sets), and dramatically reducing capacity of the online actor models used for scoring examples. All experiments were conducted on 3 epochs of one-half of JFT to enable the held-out ablations. RHO with a small reference model did not produce a learner speedup in our experiments.", + "bbox": [ + 231, + 307, + 769, + 366 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the learner) for both the online and reference model (see Appendix Algorithm 1). In Figure 4 (left) we assess the impact of these cheaper scoring models on learner efficiency. First, we find that easy reference prioritization to be very sensitive to the capacity of the scoring model: while ViT-B scoring models yield reasonable gains over uniform sampling, prioritizing with ViT-S and ViT-Ti scoring models underperforms significantly (Fig. 4, red curves).", + "bbox": [ + 231, + 396, + 769, + 475 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In contrast, we find that learnability based prioritization yields robust gains, even when the scoring models are significantly scaled down (ClassAct; Fig. 4, blue curves). For example, while ViT-B scoring models yield a $31\\%$ learner speedup, the $50 \\times$ smaller ViT-Ti scoring models still provide a $26\\%$ speedup. We pushed this logic by using even smaller scoring models (the ViT-Mu family which we introduce, see Appendix) which are up to $1000 \\times$ smaller than the learner. Despite this, prioritizing data based on their scores yields non-negligible speedups (e.g. $16\\%$ for the smallest actors we consider; Figure 4, top right).", + "bbox": [ + 231, + 477, + 769, + 581 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "These experiments demonstrate that, with the appropriate scoring criterion, online and reference models can be significantly downscaled and still produce comparable gains to larger models, with learner efficiency degrading gracefully with the actor overhead (i.e. the cost of the reference model and data scoring). As a result, our method ClassAct quickly becomes FLOP positive as the online + reference models are downscaled (Figure 4, bottom right), while at the same time producing speed-ups in wall-clock time for a given learner batch size.", + "bbox": [ + 231, + 583, + 769, + 673 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Together, our results expose a pareto front across which to determine an optimal context-specific data selection strategy (Figure 2). Where pre-trained models are available, some of the cost of larger data selection policies can be discounted. If savings in wall-clock time supersede the associated cost of scoring, large models can be tolerated for data selection. Reference model costs can also be amortized across many training runs by appending scores to 'foundation datasets'. However, in the case where no component of the framework can amortized (as in the case of large-scale pretraining), prioritizing data with small ClassAct models can deliver large savings in total computation.", + "bbox": [ + 231, + 675, + 769, + 792 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 233, + 169, + 252, + 179 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Evans et al.", + "bbox": [ + 284, + 169, + 379, + 179 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Generalising neural scaling laws to the active-learning setting", + "text_level": 1, + "bbox": [ + 231, + 195, + 754, + 210 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We next investigated the scaling behaviour of ClassAct by experimenting over large learner compute budgets (Fig. 5), using both ViT-Ti and ViT-S models as actors for training a ViT-L. Predictably, the ViT-S produced larger, although marginal gains over the ViT-Ti actors when not accounting for scoring FLOPs (Figure 5, left). However, when accounting for total FLOPs, the difference was less pronounced (Figure 5, right). Our results generalize large scale uniform sampling scaling laws such as uncovered by [19] for LLMs and reproduced for large vision transformers by [43] to the case of non-uniform sampling. For the first time, we demonstrate that these scaling laws can be shifted in our favour by selecting data using general model-based scoring heuristics.", + "bbox": [ + 231, + 219, + 771, + 350 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Training the reference model in parallel", + "text_level": 1, + "bbox": [ + 231, + 371, + 584, + 384 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The reference model needs to be trained if none is already available. This two-step process adds complexity to active model training, especially if using large scale infrastructure. However, an interesting consequence of down-scaling the reference model is that both inference passes and gradients can be computed over a much larger batch than can be computed on the learner. In theory, this would mean that the small reference model could instead be trained online, in parallel with the large learner and small online model.", + "bbox": [ + 231, + 395, + 769, + 485 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We confirmed our hypothesis by running an experiment in which we trained our reference model on a super-batch of size $B = 10b$ and trained the online and learner model in sequence with the sub-batch of size $b$ . To make sure the reference model quickly converges, we additionally set the learning rate to double that of the online and learner models (this would cause instability for the learner and online models because of the additional variance from the smaller batch). We also verified that training the reference model on a held-out set of data performed equally in our experiments to reference models trained on the same data as the learner model [27]. Our 'one-pass' setup, Online-ClassAct, produces the same performance as the pre-trained ClassAct pipeline in our Ti-trains-B experiments (Table 1). Pseudocode is shown in appendix Algorithm 1.", + "bbox": [ + 231, + 486, + 769, + 629 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We have shown that by decoupling the scoring models from the learner model entirely, it is possible to significantly downscale the scoring models with minor degradation to performance (see Table 1). Unlike RHO, which can train a large 'learner' model with a small 'reference' model, we introduce a third 'online' model, with the same architecture and parameter count as the reference model, enabling the reduction of actor computation (see Table 1).", + "bbox": [ + 231, + 629, + 771, + 707 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 ActiveCLIP: active multimodal learning", + "text_level": 1, + "bbox": [ + 231, + 728, + 586, + 742 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We have so far demonstrated that large scale image classifiers can be trained with lower total compute by actively selecting the data used for training. However, classification has largely been superseded as a large scale pre-training method", + "bbox": [ + 231, + 752, + 769, + 792 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Active Learning Accelerates Large-Scale Visual Understanding", + "bbox": [ + 326, + 168, + 719, + 181 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 752, + 169, + 766, + 179 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/414efc23932b836c592dc6c59ee3c6fb7b68df0bf13d7ac35da3c865c38fe505.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ImageNet0-ShotCOCO (im2txtR1)COCO (txt2imR1)(Learner)
Speed-up % (vs. Uniform)23.048.128.648.026.827.0ALIGN
0.016.318.822.128.217.5LTIP
Performance (vs. Uniform)47.8 (2.2)53.2 (7.4)40.9 (2.4)44.8 (6.3)27.3 (1.8)30.5 (5.0)ALIGN
46.5 (-0.1)47.2 (0.6)45.6 (1.2)45.4 (1.0)31.8 (1.7)31.0 (0.9)LTIP
(Reference)ALIGNLTIPALIGNLTIPALIGNLTIP
", + "bbox": [ + 257, + 193, + 742, + 262 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Generalizing data-selection policies across datasets and tasks. We pretrain reference models on the large but noisy ALIGN dataset, or the smaller and more curated LTIP dataset [2]. Consistent with [2], we find training with uniform sampling on LTIP to yield stronger transfer learning results than pre-training on ALIGN. These reference models can be used very effectively for data-selection on both LTIP and ALIGN, whereas ALIGN-pretrained reference models yield more modest speedups. All models are provided with 800M training images at resolution $128 \\times 128$ , speedups are shown relative to the time at which the uniform sampling baseline was reached for that evaluation metric. Colour indicates performance relative to uniform (brackets).", + "bbox": [ + 231, + 268, + 771, + 377 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b840da8d161994ef6e8d8f5a76ac04441e1a2f3be9b0ce6906093daaa4b6e49b.jpg", + "image_caption": [ + "Fig. 6: Reference policies generalize across tasks. We train a ViT-B reference model on either ALIGN or LTIP, then use it to train a second ViT-B learner on ALIGN with ActiveCLIP. The biggest gains were found with an LTIP reference model, despite it needing to perform out-of-domain generalization. In 50k iterations, ActiveCLIP selects the 800M \"cleanest\" examples from the ALIGN dataset, whose size is 1.6B in total." + ], + "image_footnote": [], + "bbox": [ + 236, + 399, + 500, + 504 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5fa7cd19b63e436f2c84975c09634e2d56b67ded39bd586dfea33e1978a14457.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 502, + 399, + 764, + 504 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "by CLIP-style multimodal training [30]. Figure 6 demonstrates that our CLIP-adapted active learning method ActiveCLIP (see Methods) produces similar speedups in terms of learner computation as observed in JFT classification. Specifically, we find that prioritized sampling with learnability scores accelerates multimodal pre-training by $18 - 48\\%$ , depending on the evaluation metric (ImageNet zero-shot accuracy or COCO retrieval) and reference model configuration, which we explore below.", + "bbox": [ + 231, + 591, + 769, + 682 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Policy generalization across tasks", + "text_level": 1, + "bbox": [ + 231, + 703, + 534, + 716 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To fully leverage off-the-shelf pre-trained image models for training (Figure 2), our results up to now suggest that the reference model should be trained on the same task that the learner model is being trained for. In Figure 6 (see Table 2 for full results), we show that we can in fact pre-train our reference models on related but distinct datasets. Moreover results suggest that there may even be a benefit", + "bbox": [ + 231, + 726, + 769, + 792 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 233, + 169, + 252, + 179 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Evans et al.", + "bbox": [ + 284, + 169, + 379, + 179 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1f4b4319b4f6f88c3ecccf5b128d2d33b845ad0f6872e07e8728ec8f31e9c94f.jpg", + "image_caption": [ + "Fig. 7: Reference models trained on curated datasets are powerful data selectors. Active data selection using ActiveSigLIP with reference models trained on increasingly curated subsets of Webli. Webli-350M reference models are consistently more effective than those trained on raw or 1B subsets." + ], + "image_footnote": [], + "bbox": [ + 235, + 194, + 766, + 300 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "to cross-training in some-cases; an ALIGN $\\rightarrow$ ALIGN reference $\\rightarrow$ learner model combination ('in-domain' ActiveCLIP) produced similar speedups to ClassAct on JFT. However, these gains were greatly surpassed by an LTIP $\\rightarrow$ ALIGN combination. One possibility is that LTIP is less noisily labeled such that the scoring policies it produces are 'clean' - i.e. more able to filter for 'clean' data. The corollary may also be true; active data selection appears to greatly improve the utility of ALIGN, suggesting that it contains a large proportion of data that is of low quality for training.", + "bbox": [ + 231, + 380, + 769, + 484 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We observed the same effect when training with ActiveSigLIP on the large scale Webli dataset [6] (results summarised in Table 3). We trained reference models on both the raw 4B set as well as two extensively curated 1B / 350M subsets, which were then used to train subsequent learner models (Fig. 7). In all-but-one cases, filtering data with the 350M-trained referenced model produced the best results when transferring to 350M, 1B and 4B learner datasets. Notably, the significant gains observed over training with uniform sampling on the 350M subset suggest that our method can still improve over methods that pre-filter datasets once but then train with uniform samples [12].", + "bbox": [ + 231, + 488, + 771, + 605 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.7 Comparison to prior multimodal art", + "text_level": 1, + "bbox": [ + 233, + 638, + 556, + 650 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We leverage the insight that pre-training a reference model on clean data can facilitate learning on larger, noisy data for training our final model. Here, we train a reference model on LTIP, then use it to train a new model on the much larger mixture of LTIP and ALIGN, following [2], for a total of 3 and 8 billion training examples. Table 3 shows that in this training regime, ActiveCLIP surpasses models trained with significantly more data on ImageNet 0-Shot classification and COCO retrieval. Finally, we find that our active learning method is complementary to recent advances in multimodal learning: ActiveSigLIP significantly improves uniformly-trained SigLIP [44] in both COCO retrieval metrics.", + "bbox": [ + 231, + 674, + 769, + 792 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Active Learning Accelerates Large-Scale Visual Understanding", + "bbox": [ + 326, + 168, + 717, + 181 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 752, + 169, + 768, + 179 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6916f601796f74c8ef924d536a053e2c05522d10f5599990a0b811dba914e8f9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrain ex.IN-1K ZS Top-1COCO
im2txttxt2im
CLIP13B68.352.433.1
EVA-CLIP3B+2B69.7†
ActiveCLIP3B71.357.743.0
OpenCLIP34B70.259.442.3
EVA-CLIP8B+2B74.7†58.742.2
ActiveCLIP8B72.260.744.9
SigLIP3B72.160.742.7
ActiveSigLIP3B72.063.545.3
", + "bbox": [ + 322, + 193, + 673, + 334 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 3: Comparison of ActiveCLIP to public multimodal pre-taining methods. ActiveCLIP outperforms models trained with the same or more data (CLIP, [30]; EVA-CLIP, [37]; and OpenCLIP, [17]). ActiveSigLIP produced significant gains over the baseline SOTA performance of SigLIP [44]. $^\\dagger$ benefits from additional ImageNet21K pretraining (+2B training examples). All ActiveCLIP/SigLIP models use a reference model trained on LTIP to guide learning on a mixture of ALIGN and LTIP.", + "bbox": [ + 231, + 342, + 771, + 415 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 233, + 442, + 363, + 456 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we have presented a new method for active data selection that builds upon and simplifies the concept of 'learnability'. Our experiments demonstrate that this approach can significantly reduce the computation required for large-scale pretraining, compared to training with uniform samples. To our knowledge, this is the first active learning method that is more efficient than training with uniform samples when accounting for total FLOPs, and that does not rely on hand-designed features, allowing broad application across training setups. We have validated this by showing results on classification and contrastive pretraining, and found that our data selection policies continue to produce efficiency gains in the large-scale regime and can generalize effectively across task modalities. Collectively, our experiments also illustrate a Pareto frontier that allows trading off actor/data-selection computation against savings in training iterations, suggesting an alternative path to improved performance beyond scaling training batch sizes.", + "bbox": [ + 231, + 500, + 769, + 682 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work focused on supervised pre-training for images, but further work could involve extending our method to other modalities and training schemes such as language, video, and generative modeling. An important note is that all our experiments present results from filtering only $50\\%$ of the data; further gains may be possible by filtering more aggressively, at the cost of further overheads. In particular, aggressive data-selection coupled with efficient scoring schemes such as the ones proposed here could test the hypothesis that large-scale pretraining can benefit from exponential, rather than power-law, scaling behavior.", + "bbox": [ + 231, + 687, + 769, + 792 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 233, + 169, + 252, + 179 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Evans et al.", + "bbox": [ + 284, + 169, + 379, + 179 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 233, + 194, + 334, + 207 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abbas, A., Tirumala, K., Simig, D., Ganguli, S., Morcos, A.S.: Semdedup: Data-efficient learning at web-scale through semantic dedduplication. arXiv preprint arXiv:2303.09540 (2023)", + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: A visual language model for few-shot learning. Advances in Neural Information Processing Systems (2022)", + "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems (2020)", + "4. Campbell, T., Broderick, T.: Bayesian coreset construction via greedy iterative geodesic ascent. In: International Conference on Machine Learning. pp. 698-706. PMLR (2018)", + "5. Cassirer, A., Barth-Maron, G., Brevdo, E., Ramos, S., Boyd, T., Sottiaux, T., Kroiss, M.: Reverb: A framework for experience replay (2021)", + "6. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)", + "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)", + "8. Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., Zaharia, M.: Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829 (2019)", + "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)", + "10. Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih, V., Ward, T., Doron, Y., Firoiu, V., Harley, T., Dunning, I., et al.: Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures. In: International conference on machine learning. pp. 1407-1416. PMLR (2018)", + "11. Feldman, V.: Does learning require memorization? a short tale about a long tail. In: Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing. pp. 954-959 (2020)", + "12. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)", + "13. Gunasekar, S., Zhang, Y., Aneja, J., Mendes, C.C.T., Del Giorno, A., Gopi, S., Javaheripi, M., Kauffmann, P., de Rosa, G., Saarikivi, O., et al.: Textbooks are all you need. arXiv preprint arXiv:2306.11644 (2023)", + "14. Har-Peled, S., Kushal, A.: Smaller coresets for k-median and k-means clustering. In: Proceedings of the twenty-first annual symposium on Computational geometry. pp. 126-134 (2005)", + "15. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + ], + "bbox": [ + 233, + 226, + 768, + 792 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Active Learning Accelerates Large-Scale Visual Understanding", + "bbox": [ + 327, + 168, + 719, + 181 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 752, + 169, + 768, + 179 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. In: Advances in Neural Information Processing Systems (2022)", + "17. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below.", + "18. Jouppi, N.P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., Bates, S., Bhatia, S., Boden, N., Borchers, A., et al.: In-datacenter performance analysis of a tensor processing unit. In: Proceedings of the 44th annual international symposium on computer architecture. pp. 1-12 (2017)", + "19. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)", + "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Springer (2014)", + "21. Lindley, D.V.: On a measure of the information provided by an experiment. The Annals of Mathematical Statistics 27(4), 986-1005 (1956)", + "22. Loshchilov, I., Hutter, F.: Online batch selection for faster training of neural networks. arXiv preprint arXiv:1511.06343 (2015)", + "23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)", + "24. MacKay, D.J.: Information-based objective functions for active data selection. Neural computation 4(4), 590-604 (1992)", + "25. Mahmoud, A., Elhoushi, M., Abbas, A., Yang, Y., Ardalani, N., Leather, H., Morcos, A.: Sieve: Multimodal dataset pruning using image captioning models. arXiv preprint arXiv:2310.02110 (2023)", + "26. Marion, M., Üstün, A., Pozzobon, L., Wang, A., Fadaee, M., Hooker, S.: When less is more: Investigating data pruning for pretraining llms at scale. arXiv preprint arXiv:2309.04564 (2023)", + "27. Mindermann, S., Brauner, J.M., Razzak, M.T., Sharma, M., Kirsch, A., Xu, W., Höltgen, B., Gomez, A.N., Morisot, A., Farquhar, S., et al.: Prioritized training on points that are learnable, worth learning, and not yet learnt. In: International Conference on Machine Learning. pp. 15630-15649. PMLR (2022)", + "28. Paul, M., Ganguli, S., Dziugaite, G.K.: Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems 34, 20596-20607 (2021)", + "29. Prabhu, A., Dognin, C., Singh, M.: Sampling bias in deep active classification: An empirical study. arXiv preprint arXiv:1909.09389 (2019)", + "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)", + "31. Schaul, T., Quan, J., Antonoglou, I., Silver, D.: Prioritized experience replay. arXiv preprint arXiv:1511.05952 (2015)", + "32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + ], + "bbox": [ + 233, + 195, + 769, + 790 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 233, + 169, + 252, + 179 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Evans et al.", + "bbox": [ + 284, + 169, + 379, + 179 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "34. Settles, B.: Active learning literature survey (2009)", + "35. Sorscher, B., Geirhos, R., Shekhar, S., Ganguli, S., Morcos, A.: Beyond neural scaling laws: beating power law scaling via data pruning. Advances in Neural Information Processing Systems 35, 19523-19536 (2022)", + "36. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the IEEE international conference on computer vision. pp. 843-852 (2017)", + "37. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)", + "38. Toneva, M., Sordoni, A., Combes, R.T.d., Trischler, A., Bengio, Y., Gordon, G.J.: An empirical study of example forgetting during deep neural network learning. arXiv preprint arXiv:1812.05159 (2018)", + "39. Xie, S.M., Pham, H., Dong, X., Du, N., Liu, H., Lu, Y., Liang, P., Le, Q.V., Ma, T., Yu, A.W.: Doremi: Optimizing data mixtures speeds up language model pretraining (2023)", + "40. Xie, S.M., Santurkar, S., Ma, T., Liang, P.: Data selection for language models via importance resampling. arXiv preprint arXiv:2302.03169 (2023)", + "41. Yang, F., Barth-Maron, G., Stanczyk, P., Hoffman, M., Liu, S., Kroiss, M., Pope, A., Rrustemi, A.: Launchpad: A programming model for distributed machine learning research. arXiv preprint arXiv:2106.04516 (2021), https://arxiv.org/abs/2106.04516", + "42. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: CoCa: Contrastive captioners are image-text foundation models. In: Transactions on Machine Learning Research (2022)", + "43. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)", + "44. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023)" + ], + "bbox": [ + 233, + 195, + 768, + 578 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Active Learning Accelerates Large-Scale Visual Understanding", + "bbox": [ + 326, + 168, + 719, + 181 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 752, + 169, + 768, + 179 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_model.json b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_model.json new file mode 100644 index 0000000000000000000000000000000000000000..074acfbdec66ade6f670f460d82ed2d1f29ac73d --- /dev/null +++ b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_model.json @@ -0,0 +1,2236 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.192, + 0.756, + 0.251 + ], + "angle": 0, + "content": "Bad Students Make Great Teachers: Active Learning Accelerates Large-Scale Visual Understanding" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.277, + 0.696, + 0.305 + ], + "angle": 0, + "content": "Talfan Evans1,* Shreya Pathak1,* Hamza Merzic1,2,* Jonathan Schwarz1,† Ryutaro Tanno1 Olivier J. Henaff1,*" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.32, + 0.658, + 0.334 + ], + "angle": 0, + "content": "\\(^{1}\\)Google DeepMind \\(^{2}\\)University College London" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.359, + 0.728, + 0.562 + ], + "angle": 0, + "content": "Abstract. Power-law scaling indicates that large-scale training with uniform sampling is prohibitively slow. Active learning methods aim to increase data efficiency by prioritizing learning on the most relevant examples. Despite their appeal, these methods have yet to be widely adopted since no one algorithm has been shown to a) generalize across models and tasks b) scale to large datasets and c) yield overall FLOP savings when accounting for the overhead of data selection. In this work we propose a method which satisfies these three properties, leveraging small, cheap proxy models to estimate \"learnability\" scores for datapoints, which are used to prioritize data for training much larger models. As a result, models trained using our methods - ClassAct and ActiveCLIP - require \\(46\\%\\) and \\(51\\%\\) fewer training updates and up to \\(25\\%\\) less total computation to reach the same performance as uniformly-trained visual classifiers on JFT and multimodal models on ALIGN, respectively. Finally, we find our data-prioritization scheme to be complementary with recent data-curation and learning objectives, yielding a new state-of-the-art in several multimodal transfer tasks." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.581, + 0.384, + 0.594 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.606, + 0.772, + 0.749 + ], + "angle": 0, + "content": "Power-law scaling for vision and language models [19,43] indicates that incremental improvements in model performance require orders of magnitude increases in computation. One of the key features of these empirical power-laws is that training data is sampled uniformly. In contrast, active data selection prioritizes computation on the data that maximally contributes to task performance [21,24,34], with the ultimate goals of improving data efficiency and reducing the cost of training. However, active data selection has yet to become a mainstay of large model training, since no existing algorithm satisfies the three properties of being a) robust to the choice of model and training task, b) scalable to large datasets and architectures, and c) more compute efficient end-to-end than training with uniform samples." + }, + { + "type": "page_footnote", + "bbox": [ + 0.234, + 0.755, + 0.772, + 0.793 + ], + "angle": 0, + "content": "* Equal technical contribution. Email correspondence to and . †Current affiliation: Harvard University, work done while at Google DeepMind." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.234, + 0.17, + 0.245, + 0.18 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.17, + 0.381, + 0.18 + ], + "angle": 0, + "content": "T. Evans et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.196, + 0.502, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.196, + 0.766, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.292, + 0.772, + 0.342 + ], + "angle": 0, + "content": "Fig. 1: Active learning accelerates large-scale visual understanding. For large-scale classification and multimodal learning tasks, prioritised training on data selected using our active selection methods ClassAct (left) and ActiveCLIP (right) requires significantly fewer updates to reach the final performance of uniform training." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.347, + 0.77, + 0.424 + ], + "angle": 0, + "content": "In the first instance, data selection based on hand-engineered filters (e.g. removing incorrectly shaped images or that only contain a single colour [2]) can trivially improve training efficiency at minimal computational overhead. However, such heuristics are limited in their effectiveness by the expertise of the human designer, and are not guaranteed to transfer to training of different models, data modalities, or tasks and incur significant effort to develop and tune." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.425, + 0.771, + 0.542 + ], + "angle": 0, + "content": "In contrast, model-based curation methods, which use the loss of the model itself to score examples, have shown promise by focusing training on 'hard' and omitting 'easy' data [35], but also the opposite (to exclude noise or other low-quality data) in both language modeling [13] and multimodal learning [15]. However, these methods often spend as much computation on the curation of datasets as is gained from subsequent pretraining, making them less compute-efficient than training on uniformly-sampled data. Finally, while several compute-efficient methods have been successfully deployed at small scale [8], they generally do not scale to even medium-sized datasets such as ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.542, + 0.77, + 0.633 + ], + "angle": 0, + "content": "In this work, we propose an algorithm that satisfies the three properties of generality, scalability, and compute-positivity. The proposed framework uses small proxy models to compute learnability scores for candidate training data, resulting in significant training efficiency gains at an almost negligible overhead over standard uniform-training. We test two instantiations of this framework, ClassAct and ActiveCLIP / ActiveSigLIP, for large scale classification and multimodal pretraining, respectively. Our findings are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.638, + 0.771, + 0.692 + ], + "angle": 0, + "content": "Benchmarking heuristics for large-scale pretraining: We investigate loss- and learnability-based prioritization [15, 27, 35] for large-scale classification and find that pretrained reference models are an essential component for accelerating learning, producing efficiency gains of up to \\(46\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.696, + 0.771, + 0.763 + ], + "angle": 0, + "content": "Generalizing selection policies across model scale: Secondly, we show that smaller models act as effective proxies for much larger (\\(\\sim 1000 \\times\\)) models in the context of learnability but not loss-based scoring, resulting in compute-positive gains of up to \\(25\\%\\) over uniform training, a first in the context of large-scale pretraining." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.766, + 0.771, + 0.794 + ], + "angle": 0, + "content": "Accelerating multimodal pretraining: Using reference models pretrained on small, clean datasets, we substantially accelerate pretraining on much larger," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.169, + 0.721, + 0.182 + ], + "angle": 0, + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.759, + 0.17, + 0.771, + 0.18 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.197, + 0.771, + 0.237 + ], + "angle": 0, + "content": "noisier datasets. Moreover, we find ActiveCLIP to be complementary to recent data-curation techniques [12] and learning objectives [44], yielding a new state-of-the-art in several multimodal understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.242, + 0.772, + 0.283 + ], + "angle": 0, + "content": "Amortizing data selection policies: Data-selection policies trained on one task can also accelerate the training of subsequent models on different but related tasks, suggesting that such policies can be easily derived from pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.287, + 0.772, + 0.34 + ], + "angle": 0, + "content": "Simplifying active-learning with online reference models: Lastly, we demonstrate that pre-trained reference models may not be necessary at all, where these models are small and can be trained in parallel on larger batches than the learner model, while remaining compute-positive." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.361, + 0.396, + 0.375 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.391, + 0.772, + 0.548 + ], + "angle": 0, + "content": "Data pruning. One approach to data-selection is to identify and sub-select data ahead of training. For example, [28] and [35] show that the training loss and gradients can be used to discard large portions of small-to-medium sized datasets (e.g. CIFAR10 and ImageNet) with little loss in performance. These methods have since been deployed for the curation of web-scale datasets in both language modeling [26] and multimodal learning [1, 12, 25], demonstrating large reductions in the amount of data required together with performance improvements. However, in the single-epoch regime that is becoming typical of large model training [16, 19], pre-filtering can be as expensive as learning from it, a shortcoming which we address in this work. Nevertheless, we show that our method for dynamic data selection is complementary to and benefits from such data curation techniques." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.553, + 0.772, + 0.698 + ], + "angle": 0, + "content": "Online active learning. Unlike data-pruning, online active learning continuously filters data throughout training and applies naturally to the semi-infinite, single-epoch regime. Online Batch Selection [22] scores and filters using the learner model, which has the theoretical advantage that the importance of data can be determined relative to the current state of the learner. In terms of metrics, the Reducible Holdout Loss (RHO) [27] also uses the concept of a reference model to identify learnable data points not yet well represented. Other proposed heuristics include memorization for long-tailed data [11] and assigning \"complexity\" scores based on the number of times the example is forgotten during training [38]. None of these approaches however have demonstrated that the cost of scoring can be reduced to the point of justifying learner efficiency gains." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.702, + 0.772, + 0.795 + ], + "angle": 0, + "content": "Compute-efficient data selection. Several works have demonstrated the benefits of selecting data based on simple heuristics, such as low-level image properties [2] or proximity to high-quality text corpora [3, 7, 29, 40]. While cheap to compute, these statistics often require domain-specific knowledge which limits their applicability across tasks. Domain-agnostic methods such as core-sets alleviate this by selecting data based on the geometry of their embeddings [4, 14] which can be efficiently computed, however these algorithms generally do not" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.234, + 0.17, + 0.246, + 0.18 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.17, + 0.381, + 0.18 + ], + "angle": 0, + "content": "T. Evans et al." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.196, + 0.772, + 0.249 + ], + "angle": 0, + "content": "scale to large-scale datasets [8]. Most related to our work is DoReMi [39] which uses domain-general, scalable, and compute-efficient proxy models for the simpler problem of determining optimal data-mixtures for the subsequent training of a larger language model." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.267, + 0.348, + 0.28 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.292, + 0.548, + 0.306 + ], + "angle": 0, + "content": "3.1 Data selection as prioritized replay" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.312, + 0.772, + 0.435 + ], + "angle": 0, + "content": "We use online batch selection [22] to apply our scoring heuristics to standard visual learning tasks: firstly, we sample uniformly from the training set \\( \\boldsymbol{x}_i \\stackrel{\\mathcal{U}}{\\sim} \\mathcal{D} \\) and compute a score \\( s_i = s(\\boldsymbol{x}_i | \\theta) \\in \\mathbb{R} \\) to each data point \\( \\boldsymbol{x}_i \\) using model parameters \\( \\theta \\). Given a large enough collection of scored examples stored in a memory bank \\( \\mathcal{M} = \\{\\boldsymbol{x}_i\\}_{i \\in (0, \\dots, M-1)} \\), we then sample non-uniformly according to their scores \\( \\boldsymbol{x}_i \\stackrel{\\pi}{\\sim} \\mathcal{M} \\) [31], where \\( \\pi(\\boldsymbol{x}_i) = \\text{Softmax}(\\{s_i\\}_{i \\in (0, \\dots, M-1)}) \\). A batch of such examples is used to update the learner model. Following convention in reinforcement learning, we refer to the scoring and target models as actors and learners respectively." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.452, + 0.487, + 0.464 + ], + "angle": 0, + "content": "3.2 Statistics for data selection" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.472, + 0.771, + 0.497 + ], + "angle": 0, + "content": "We explore a few statistics for model-based prioritization, grouped into two categories." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.503, + 0.771, + 0.555 + ], + "angle": 0, + "content": "Example difficulty: given the current state of the learner, an intuitive prioritization scheme might favour 'difficult' examples (as measured by their training loss), while removing 'easy' examples that are trivially classified and which yield small gradients. This loss-based prioritization:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.563, + 0.769, + 0.577 + ], + "angle": 0, + "content": "\\[\ns ^ {\\text {h a r d}} \\left(\\boldsymbol {x} _ {i} \\mid \\theta\\right) = \\ell \\left(\\boldsymbol {x} _ {i} \\mid \\theta\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.585, + 0.77, + 0.638 + ], + "angle": 0, + "content": "can use the current parameters of the learner \\(\\theta^t\\) or those of a fixed model \\(\\theta^{*}\\). The opposite argument can be made for favoring examples that are easily solved by a well-trained model, as such a prioritization removes the noisy examples present in large-scale datasets:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.646, + 0.769, + 0.659 + ], + "angle": 0, + "content": "\\[\ns ^ {\\text {e a s y}} \\left(\\boldsymbol {x} _ {i} | \\theta\\right) = - \\ell \\left(\\boldsymbol {x} _ {i} | \\theta\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.666, + 0.77, + 0.694 + ], + "angle": 0, + "content": "This scheme is commonly used in multimodal learning for identifying high-quality examples with pre-trained models [15,32,33]." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.698, + 0.771, + 0.749 + ], + "angle": 0, + "content": "Example learnability: Given that favoring easy and hard examples target different and potentially orthogonal properties of the data, a natural question is whether these policies can be combined. Learnability criteria straightforwardly combine the two as" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.758, + 0.769, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s ^ {\\text {l e a r n}} \\left(\\boldsymbol {x} _ {i} | \\theta^ {t}, \\theta^ {*}\\right) = s ^ {\\text {h a r d}} \\left(\\boldsymbol {x} _ {i} | \\theta^ {t}\\right) + s ^ {\\text {e a s y}} \\left(\\boldsymbol {x} _ {i} | \\theta^ {*}\\right) (3) \\\\ = \\ell \\left(\\boldsymbol {x} _ {i} \\mid \\theta^ {t}\\right) - \\ell \\left(\\boldsymbol {x} _ {i} \\mid \\theta^ {*}\\right), (4) \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.169, + 0.721, + 0.182 + ], + "angle": 0, + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.759, + 0.17, + 0.771, + 0.18 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.196, + 0.772, + 0.261 + ], + "angle": 0, + "content": "favoring examples that are easily solved by a well-trained model \\(\\theta^{*}\\) but challenging to the learner in its current state \\(\\theta^t\\), such that more computation dedicated to this example could lower its loss. Conversely, examples that are trivially classified by the learner (or mislabeled) will yield low (or high) losses for both the current learner and the well-trained one, leading to low learnability scores." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.261, + 0.772, + 0.314 + ], + "angle": 0, + "content": "A special case of learnability scores (the RHO loss, [27]) uses a model \\(\\theta^{\\mathrm{ho}}\\) specifically trained on a held-out dataset to ensure the independence of its predictions from those of the current learner \\(s^{\\mathrm{learn}}(\\pmb{x}_i|\\theta^t,\\theta^{\\mathrm{ho}})\\). We assess in Section 4.2 whether this is necessary when training on large-scale image datasets." + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.34, + 0.768, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.424, + 0.774, + 0.591 + ], + "angle": 0, + "content": "Fig. 2: Amortizing the cost of data selection. Drawn to scale: length of bars indicates number of FLOPs required to reach the accuracy of a ViT-L trained with uniform sampling (\"ViT-L Uniform Sampling\", see Figure 4). Expensive model policies (e.g. a ViT-B scores data for the ViT-L learner, or \\(\\mathrm{B} \\rightarrow \\mathrm{L}\\)) produce large learner speedups, at the expense of the additional FLOPs associate with data selection. This overhead can be reduced by deriving the data-selection policies from smaller models (e.g. ViT-S, ViT-Ti or ViT-Mu score data for the ViT-L learner), at the expense of marginal decreases in the learner speedup. Costs could be additionally amortized by using off-the-shelf reference models, removing the need to train from scratch (yellow). Since the reference model is fixed throughout training, scores can be assigned once to a 'foundation dataset' and amortized across many training runs (lime green; [27,35]). Since the online model is independent of the learner model and generalizes across scale, data selection policies can also be distilled as a fixed ordering of a given dataset (a 'foundation curriculum')." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.624, + 0.56, + 0.639 + ], + "angle": 0, + "content": "3.3 Unlocking compute-positive training" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.645, + 0.772, + 0.71 + ], + "angle": 0, + "content": "Scoring requires inference passes over the actor and learner models. We assume that gradient updates to cost \\(3\\mathrm{x}\\) inference passes. The cost of scoring data \\(F\\) thus scales with the proportion of data which is being rejected (e.g. retaining only \\(20\\%\\) of the data requires 5 inference passes per trained batch). The requirements for compute-positivity can therefore be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.719, + 0.77, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\left(3 F _ {\\text {l e a r n}} + \\rho F _ {\\text {a c t}}\\right) \\beta + 3 F _ {\\text {r e f}}} _ {\\text {A c t i v e L e a r n i n g}} < \\underbrace {3 F _ {\\text {l e a r n}}} _ {\\text {U n i f o r m S a m p l i n g}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.766, + 0.771, + 0.793 + ], + "angle": 0, + "content": "where \\( F_{\\mathrm{act}} \\) is the cost of scoring an example, \\( \\rho \\) is the number of examples scored per training example, and \\( \\beta \\) is the efficiency gain (in terms of the learner update)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.234, + 0.17, + 0.245, + 0.179 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.17, + 0.381, + 0.18 + ], + "angle": 0, + "content": "T. Evans et al." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.197, + 0.769, + 0.274 + ], + "angle": 0, + "content": "relative to training with uniform sampling (see Appendix Section A.1 for more details). The RHS term is the cost per update of uniform sampling. The first LHS term inside the brackets is the cost of the learner training during AL, the second term is the scoring cost and the last term outside the brackets the cost of training the reference model. We illustrate in Figure 2 the different contexts in which parts of this computation may be effectively amortized." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.276, + 0.77, + 0.355 + ], + "angle": 0, + "content": "In the large-scale training regime where data is neither repeated nor seen before, compute-positivity requires that either or all of the reference model training, actor scoring and learner efficiency \\(\\beta\\) terms must be made smaller to produce net savings relative to uniform sampling. Typical prioritization schemes can produce saving on the order of \\(50\\%\\) (ie \\(\\beta = 0.5\\)), suggesting that savings must also be made by down-scaling the other terms \\(F_{\\mathrm{act}}\\) and \\(F_{\\mathrm{ref}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.36, + 0.771, + 0.412 + ], + "angle": 0, + "content": "Cost of easy-reference scoring. While both the cost of the reference model and example scoring can be scaled down in the case of easy-reference scoring with small models (\\( F_{\\mathrm{act}} = F_{\\mathrm{ref}} \\), see Equation 2), it is unclear whether the efficiency gains \\( \\beta \\) are robust to this down-scaling (see section 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.418, + 0.771, + 0.485 + ], + "angle": 0, + "content": "Cost of RHO learnability scoring. The original definition of learnability scores [27] requires inference passes through both the learner and a reference model (\\( F_{\\mathrm{act}} = F_{\\mathrm{ref}} + F_{\\mathrm{learn}} \\), see Equation 3), meaning that although the cost of the reference model can be reduced by using a smaller model, the cost of example scoring cannot." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.49, + 0.77, + 0.569 + ], + "angle": 0, + "content": "Cost of ClassAct / ActiveCLIP. For this reason, we explore whether replacing the learner model in term 1 of Equation 3 with a much smaller model can still produce comparable learner efficiency gains to those already observed (see Appendix Algorithm 1). Specifically, we introduce a third \"online\" model, which has the same architecture and size as the reference model, but is trained in parallel with the learner. In this case, the cost of scoring examples reduces to:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.586, + 0.769, + 0.6 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {a c t}} = F _ {\\text {r e f}} + F _ {\\text {o n l i n e}} = 2 F _ {\\text {r e f}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.609, + 0.77, + 0.702 + ], + "angle": 0, + "content": "and can be scaled down along with the reference model. We instantiate our method for two canonical pre-training tasks: visual classification and multimodal learning, which we call ClassAct, ActiveCLIP and ActiveSigLIP respectively. In Appendix Section A.6, we describe an asynchronous active learning framework where scoring and learning is performed in parallel on separate devices. This setup does not affect the FLOP calculations but can mitigate the overhead in time even when using non-approximate scoring." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.726, + 0.625, + 0.74 + ], + "angle": 0, + "content": "3.4 Losses for canonical visual pre-training tasks" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.753, + 0.771, + 0.796 + ], + "angle": 0, + "content": "For visual classification with ClassACT, we use the standard cross-entropy loss for both actors and learners. For multimodal learning with ActiveCLIP, learners optimize the contrastive loss \\(\\ell_{\\mathrm{learn}} = \\ell_{\\mathrm{learn}}^{\\mathrm{im,txt}} + \\ell_{\\mathrm{learn}}^{\\mathrm{txt,im}}\\), whereas the actor loss \\(\\ell_{\\mathrm{act}}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.169, + 0.72, + 0.182 + ], + "angle": 0, + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.759, + 0.17, + 0.77, + 0.18 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.196, + 0.494, + 0.21 + ], + "angle": 0, + "content": "Algorithm 1 ClassAct / ActiveCLIP" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.213, + 0.772, + 0.25 + ], + "angle": 0, + "content": "1: Input: Randomly initialized learner model \\(\\theta_{l}\\) and small online model \\(\\theta_{o}\\), small pre-trained reference model \\(\\theta_{r}\\). Models use loss \\(\\ell_{\\mathrm{act}}\\) for scoring data and \\(\\ell_{\\mathrm{learn}}\\) for computing updates. Dataset \\(\\mathcal{D}\\), batch size \\(B\\), sub-batch size \\(b < B\\)." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.25, + 0.377, + 0.261 + ], + "angle": 0, + "content": "2: while training do" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.261, + 0.77, + 0.274 + ], + "angle": 0, + "content": "3: \\(X\\sim \\mathcal{D}\\) , where \\(|X| = B\\) Sample uniformly" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.274, + 0.769, + 0.285 + ], + "angle": 0, + "content": "4: \\(S = \\ell_{\\mathrm{act}}(X|\\theta_o) - \\ell_{\\mathrm{act}}(X|\\theta_r)\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.285, + 0.769, + 0.297 + ], + "angle": 0, + "content": "5: \\(I\\sim \\mathrm{SoftMax}(S)\\) , where \\(|I| = b\\) Sample indices" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.297, + 0.769, + 0.308 + ], + "angle": 0, + "content": "6: \\(Y = X[I]\\) \\(\\triangleright\\) Collect sub-batch" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.308, + 0.769, + 0.321 + ], + "angle": 0, + "content": "7: \\(\\theta_{l}\\gets \\mathrm{Adam}[\\nabla_{\\theta_{l}}\\ell_{\\mathrm{learn}}(Y|\\theta_{l})]\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.321, + 0.769, + 0.333 + ], + "angle": 0, + "content": "8: \\(\\theta_o\\gets \\mathrm{Adam}[\\nabla_{\\theta_o}\\ell_{\\mathrm{learn}}(Y|\\theta_o)]\\) Update online model" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.333, + 0.33, + 0.344 + ], + "angle": 0, + "content": "9: end while" + }, + { + "type": "list", + "bbox": [ + 0.237, + 0.213, + 0.772, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.371, + 0.729, + 0.385 + ], + "angle": 0, + "content": "is simply the dot-product similarity between image and text embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.391, + 0.769, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {a c t}} \\left(\\boldsymbol {x} _ {i} \\mid \\theta\\right) = - \\boldsymbol {z} _ {i} ^ {\\mathrm {i m}} \\cdot \\boldsymbol {z} _ {i} ^ {\\mathrm {t x t}} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.409, + 0.769, + 0.441 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {l e a r n}} ^ {\\mathrm {i m}, \\mathrm {t x t}} (\\boldsymbol {x} _ {i} | \\theta) = - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {i m}} \\cdot \\boldsymbol {z} _ {i} ^ {\\mathrm {t x t}}\\right)}{\\sum_ {j} \\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {i m}} \\cdot \\boldsymbol {z} _ {j} ^ {\\mathrm {t x t}}\\right)} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.448, + 0.771, + 0.503 + ], + "angle": 0, + "content": "where \\( \\boldsymbol{z}_i^{\\mathrm{im}} = f^{\\mathrm{im}}(\\boldsymbol{x}_i; \\theta) \\) and \\( \\boldsymbol{z}_i^{\\mathrm{txt}} = f^{\\mathrm{txt}}(\\boldsymbol{x}_i; \\theta) \\) are image and text embeddings respectively, and \\( \\ell_{\\mathrm{learn}}^{\\mathrm{txt}, \\mathrm{im}}(\\boldsymbol{x}_i; \\theta) \\) is defined analogously to \\( \\ell_{\\mathrm{learn}}^{\\mathrm{im}, \\mathrm{txt}}(\\boldsymbol{x}_i; \\theta) \\). Similarly, ActiveSigLIP instead uses the sigmoid loss [44] for the learner's objective and \\( \\ell_{\\mathrm{act}} \\) for scoring." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.521, + 0.383, + 0.536 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.546, + 0.77, + 0.663 + ], + "angle": 0, + "content": "All our experiments were conducted with Vision Transformers [9] for which strong baselines are available across model sizes [43]. Unless specified, we adopt models with patch-size 16 throughout (ViT-S refers to ViT-S/16 and similar). We consider two canonical tasks for large-scale pretraining: classification on JFT-300M [36] and multimodal contrastive learning [30] on large image-text datasets. When pre-training with JFT classification we use held-out classification performance as the evaluation metric. When pre-training on image-text data we evaluate with standard multimodal transfer tasks: ImageNet zero-shot classification and image-to-text / text-to-image retrieval on COCO." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.663, + 0.771, + 0.73 + ], + "angle": 0, + "content": "Throughout, we will refer to the large batch of size \\( B \\) sampled uniformly from the training data as the 'super-batch', and the prioritised smaller batch of size \\( b < B \\) as the 'sub-batch'. In all our experiment, we filter 50% of uniformly sampled data such that \\( \\rho = B / b = 2 \\), although more aggressive filtering regimes warrant investigation [35]." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.746, + 0.77, + 0.76 + ], + "angle": 0, + "content": "4.1 Evaluating loss-based scoring heuristics in the large-data regime" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.766, + 0.771, + 0.793 + ], + "angle": 0, + "content": "We begin by evaluating loss- and learnability-based heuristics on their ability to accelerate supervised classification on JFT (Fig. 3). Arguably the most intuitive" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.234, + 0.17, + 0.245, + 0.179 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.17, + 0.38, + 0.18 + ], + "angle": 0, + "content": "T. Evans et al." + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.203, + 0.71, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.345, + 0.772, + 0.417 + ], + "angle": 0, + "content": "Fig. 3: Evaluation of loss-based data-selection criteria for large-scale classification. We train a ViT-B on JFT-300M with different data-selection policies. Prioritising hard data under the learner (green curve) produced marginal gains over the uniform sampling baseline. Prioritizing data using both learnability (blue curve, [27]) and easy reference prioritization (red curve, [15]) produced significant speedups and performance gains." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.428, + 0.501, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.429, + 0.768, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.602, + 0.772, + 0.782 + ], + "angle": 0, + "content": "Fig. 4: Generalization of data-selection policies across models scales. Left: We train a ViT-L for 3 epochs on JFT using uniform sampling (grey) or prioritized data sampling using example learnability (blue) or low-loss under the reference model (red). Example scores are computed using ViT-B actors (dark), or cheaper ViT-S or ViT-Tiny models (light). While both example learnability and \"easy reference\" yield good speedups with expensive actors, learnability criteria are much more robust to approximate scoring. Top right: Learner (ViT-L) speedup is computed as the fraction of learner iterations saved in order to attain the baseline's top performance. Actor overhead is computed as the additional computation in FLOPs required to score examples with a particular actor architecture (varying from ViT-Mu to ViT-L, see Appendix Table 4). Example learnability yields robust learner speedups across actor scales, \"easy reference\" scoring does not. Lower right: total compute efficiency is calculated as a product of learner efficiency and actor overhead, indicating the amount of computation required to reach baseline performance. Approximate actors (i.e. ViT-S or smaller) computing example learnability enable total compute speedups, other schemes do not." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.169, + 0.721, + 0.182 + ], + "angle": 0, + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.758, + 0.17, + 0.77, + 0.18 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.196, + 0.772, + 0.287 + ], + "angle": 0, + "content": "method to score data is to prioritise training on data with high loss under the learner (hard learner). In our experiments, this strategy (Eq. 1) only marginally improved performance over the uniform sampling baseline, despite requiring an additional inference pass over the super-batch. This is perhaps not surprising - data points with high loss may also be unlearnable due to e.g. label noise, such that training on those data points does not result in the model performing any better on the held-out test set. Large scale datasets are more likely to be noisy." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.287, + 0.772, + 0.354 + ], + "angle": 0, + "content": "Scoring methods based on pre-trained reference models performed much better—both easy reference (equation 2) and learnability (equation 3) -based prioritization produced significant gains over uniform sampling. Here, we pretrained an identical ViT-B for the same 3 epochs to use as a reference model for a second training run displayed above. producing speed-ups of \\(33\\%\\) (Fig. 3)." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.377, + 0.501, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.377, + 0.768, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.5, + 0.772, + 0.596 + ], + "angle": 0, + "content": "Fig. 5: Scaling laws for active learning. We trained a baseline ViT-L over a range of compute budgets (for which ViT-L is compute optimal, see Zhai et al., 2021). We also trained the same ViT-L with both ViT-Ti and ViT-S reference policies, pre-trained for the same number of epochs. Left: Small model policies produce robust savings in learner compute. Right: When accounting for total compute (learner + actor training and data scoring), small model policies in all compute budgets produce FLOP savings over training with uniform samples. These scaling laws generalize those measured empirically in the uniform sampling setting [43] to the case of non-uniform data selection." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.64, + 0.644, + 0.654 + ], + "angle": 0, + "content": "4.2 Generalising data-selection policies across scale" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.662, + 0.771, + 0.74 + ], + "angle": 0, + "content": "The speed-ups afforded by learnability prioritization (Fig. 3) come at the cost of the additional inference passes required to score the data during learner training, plus the cost of training the reference model. This makes the overall gains strongly compute-negative relative to training with uniform samples. Even if the size of the reference model is scaled down [27], these methods still incur the cost of additional learner inference passes to score the data during training." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.741, + 0.771, + 0.793 + ], + "angle": 0, + "content": "Unlocking compute-positive active learning. To address this issue, we introduce a set of down-scaled models with the same ViT architecture that we use to score data for training a larger ViT-L model (see Methods Section 3.3. We use ViT-B, ViT-S or ViT-Ti variants (which are \\(4 \\times\\), \\(13 \\times\\) and \\(47 \\times\\) cheaper than" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.235, + 0.17, + 0.253, + 0.18 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.17, + 0.381, + 0.18 + ], + "angle": 0, + "content": "T. Evans et al." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.194, + 0.794, + 0.296 + ], + "angle": 0, + "content": "
MethodViT model capacityReference TypeSpeed-up
Reference Online LearnerLearner speedupCompute speedup
Uniform Sampling (ViT-B)B0%0%
RHOTinyBBHeld-out, fixed0%-79%
ClassAct-HOTinyTinyBHeld-out, fixed18%3%
ClassActTinyTinyBIn-domain, fixed18%3%
ClassAct-OnlineTinyTinyBTrained online17%2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.233, + 0.297, + 0.77, + 0.308 + ], + "angle": 0, + "content": "Table 1: Simplifying and accelerating the computation of learnability scores." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.308, + 0.77, + 0.368 + ], + "angle": 0, + "content": "Relative to RHO [27], ClassAct makes two changes: replacing the reference model with one trained in-domain (removing the need for bespoke held-out sets), and dramatically reducing capacity of the online actor models used for scoring examples. All experiments were conducted on 3 epochs of one-half of JFT to enable the held-out ablations. RHO with a small reference model did not produce a learner speedup in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.397, + 0.771, + 0.476 + ], + "angle": 0, + "content": "the learner) for both the online and reference model (see Appendix Algorithm 1). In Figure 4 (left) we assess the impact of these cheaper scoring models on learner efficiency. First, we find that easy reference prioritization to be very sensitive to the capacity of the scoring model: while ViT-B scoring models yield reasonable gains over uniform sampling, prioritizing with ViT-S and ViT-Ti scoring models underperforms significantly (Fig. 4, red curves)." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.478, + 0.771, + 0.582 + ], + "angle": 0, + "content": "In contrast, we find that learnability based prioritization yields robust gains, even when the scoring models are significantly scaled down (ClassAct; Fig. 4, blue curves). For example, while ViT-B scoring models yield a \\(31\\%\\) learner speedup, the \\(50 \\times\\) smaller ViT-Ti scoring models still provide a \\(26\\%\\) speedup. We pushed this logic by using even smaller scoring models (the ViT-Mu family which we introduce, see Appendix) which are up to \\(1000 \\times\\) smaller than the learner. Despite this, prioritizing data based on their scores yields non-negligible speedups (e.g. \\(16\\%\\) for the smallest actors we consider; Figure 4, top right)." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.584, + 0.771, + 0.674 + ], + "angle": 0, + "content": "These experiments demonstrate that, with the appropriate scoring criterion, online and reference models can be significantly downscaled and still produce comparable gains to larger models, with learner efficiency degrading gracefully with the actor overhead (i.e. the cost of the reference model and data scoring). As a result, our method ClassAct quickly becomes FLOP positive as the online + reference models are downscaled (Figure 4, bottom right), while at the same time producing speed-ups in wall-clock time for a given learner batch size." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.676, + 0.771, + 0.793 + ], + "angle": 0, + "content": "Together, our results expose a pareto front across which to determine an optimal context-specific data selection strategy (Figure 2). Where pre-trained models are available, some of the cost of larger data selection policies can be discounted. If savings in wall-clock time supersede the associated cost of scoring, large models can be tolerated for data selection. Reference model costs can also be amortized across many training runs by appending scores to 'foundation datasets'. However, in the case where no component of the framework can amortized (as in the case of large-scale pretraining), prioritizing data with small ClassAct models can deliver large savings in total computation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.169, + 0.721, + 0.182 + ], + "angle": 0, + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.753, + 0.17, + 0.768, + 0.18 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.232, + 0.196, + 0.755, + 0.211 + ], + "angle": 0, + "content": "4.3 Generalising neural scaling laws to the active-learning setting" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.22, + 0.772, + 0.351 + ], + "angle": 0, + "content": "We next investigated the scaling behaviour of ClassAct by experimenting over large learner compute budgets (Fig. 5), using both ViT-Ti and ViT-S models as actors for training a ViT-L. Predictably, the ViT-S produced larger, although marginal gains over the ViT-Ti actors when not accounting for scoring FLOPs (Figure 5, left). However, when accounting for total FLOPs, the difference was less pronounced (Figure 5, right). Our results generalize large scale uniform sampling scaling laws such as uncovered by [19] for LLMs and reproduced for large vision transformers by [43] to the case of non-uniform sampling. For the first time, we demonstrate that these scaling laws can be shifted in our favour by selecting data using general model-based scoring heuristics." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.372, + 0.586, + 0.385 + ], + "angle": 0, + "content": "4.4 Training the reference model in parallel" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.396, + 0.77, + 0.486 + ], + "angle": 0, + "content": "The reference model needs to be trained if none is already available. This two-step process adds complexity to active model training, especially if using large scale infrastructure. However, an interesting consequence of down-scaling the reference model is that both inference passes and gradients can be computed over a much larger batch than can be computed on the learner. In theory, this would mean that the small reference model could instead be trained online, in parallel with the large learner and small online model." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.487, + 0.771, + 0.63 + ], + "angle": 0, + "content": "We confirmed our hypothesis by running an experiment in which we trained our reference model on a super-batch of size \\( B = 10b \\) and trained the online and learner model in sequence with the sub-batch of size \\( b \\). To make sure the reference model quickly converges, we additionally set the learning rate to double that of the online and learner models (this would cause instability for the learner and online models because of the additional variance from the smaller batch). We also verified that training the reference model on a held-out set of data performed equally in our experiments to reference models trained on the same data as the learner model [27]. Our 'one-pass' setup, Online-ClassAct, produces the same performance as the pre-trained ClassAct pipeline in our Ti-trains-B experiments (Table 1). Pseudocode is shown in appendix Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.63, + 0.772, + 0.709 + ], + "angle": 0, + "content": "We have shown that by decoupling the scoring models from the learner model entirely, it is possible to significantly downscale the scoring models with minor degradation to performance (see Table 1). Unlike RHO, which can train a large 'learner' model with a small 'reference' model, we introduce a third 'online' model, with the same architecture and parameter count as the reference model, enabling the reduction of actor computation (see Table 1)." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.729, + 0.588, + 0.743 + ], + "angle": 0, + "content": "4.5 ActiveCLIP: active multimodal learning" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.753, + 0.771, + 0.793 + ], + "angle": 0, + "content": "We have so far demonstrated that large scale image classifiers can be trained with lower total compute by actively selecting the data used for training. However, classification has largely been superseded as a large scale pre-training method" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.235, + 0.17, + 0.253, + 0.18 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.17, + 0.38, + 0.18 + ], + "angle": 0, + "content": "T. Evans et al." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.194, + 0.744, + 0.263 + ], + "angle": 0, + "content": "
ImageNet0-ShotCOCO (im2txtR1)COCO (txt2imR1)(Learner)
Speed-up % (vs. Uniform)23.048.128.648.026.827.0ALIGN
0.016.318.822.128.217.5LTIP
Performance (vs. Uniform)47.8 (2.2)53.2 (7.4)40.9 (2.4)44.8 (6.3)27.3 (1.8)30.5 (5.0)ALIGN
46.5 (-0.1)47.2 (0.6)45.6 (1.2)45.4 (1.0)31.8 (1.7)31.0 (0.9)LTIP
(Reference)ALIGNLTIPALIGNLTIPALIGNLTIP
" + }, + { + "type": "table_caption", + "bbox": [ + 0.232, + 0.269, + 0.772, + 0.378 + ], + "angle": 0, + "content": "Table 2: Generalizing data-selection policies across datasets and tasks. We pretrain reference models on the large but noisy ALIGN dataset, or the smaller and more curated LTIP dataset [2]. Consistent with [2], we find training with uniform sampling on LTIP to yield stronger transfer learning results than pre-training on ALIGN. These reference models can be used very effectively for data-selection on both LTIP and ALIGN, whereas ALIGN-pretrained reference models yield more modest speedups. All models are provided with 800M training images at resolution \\(128 \\times 128\\), speedups are shown relative to the time at which the uniform sampling baseline was reached for that evaluation metric. Colour indicates performance relative to uniform (brackets)." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.4, + 0.501, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.4, + 0.766, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.516, + 0.772, + 0.577 + ], + "angle": 0, + "content": "Fig. 6: Reference policies generalize across tasks. We train a ViT-B reference model on either ALIGN or LTIP, then use it to train a second ViT-B learner on ALIGN with ActiveCLIP. The biggest gains were found with an LTIP reference model, despite it needing to perform out-of-domain generalization. In 50k iterations, ActiveCLIP selects the 800M \"cleanest\" examples from the ALIGN dataset, whose size is 1.6B in total." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.592, + 0.77, + 0.683 + ], + "angle": 0, + "content": "by CLIP-style multimodal training [30]. Figure 6 demonstrates that our CLIP-adapted active learning method ActiveCLIP (see Methods) produces similar speedups in terms of learner computation as observed in JFT classification. Specifically, we find that prioritized sampling with learnability scores accelerates multimodal pre-training by \\(18 - 48\\%\\), depending on the evaluation metric (ImageNet zero-shot accuracy or COCO retrieval) and reference model configuration, which we explore below." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.704, + 0.536, + 0.717 + ], + "angle": 0, + "content": "4.6 Policy generalization across tasks" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.727, + 0.771, + 0.793 + ], + "angle": 0, + "content": "To fully leverage off-the-shelf pre-trained image models for training (Figure 2), our results up to now suggest that the reference model should be trained on the same task that the learner model is being trained for. In Figure 6 (see Table 2 for full results), we show that we can in fact pre-train our reference models on related but distinct datasets. Moreover results suggest that there may even be a benefit" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.169, + 0.718, + 0.182 + ], + "angle": 0, + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.753, + 0.17, + 0.769, + 0.18 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.195, + 0.768, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.312, + 0.772, + 0.36 + ], + "angle": 0, + "content": "Fig. 7: Reference models trained on curated datasets are powerful data selectors. Active data selection using ActiveSigLIP with reference models trained on increasingly curated subsets of Webli. Webli-350M reference models are consistently more effective than those trained on raw or 1B subsets." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.381, + 0.77, + 0.485 + ], + "angle": 0, + "content": "to cross-training in some-cases; an ALIGN \\(\\rightarrow\\) ALIGN reference \\(\\rightarrow\\) learner model combination ('in-domain' ActiveCLIP) produced similar speedups to ClassAct on JFT. However, these gains were greatly surpassed by an LTIP \\(\\rightarrow\\) ALIGN combination. One possibility is that LTIP is less noisily labeled such that the scoring policies it produces are 'clean' - i.e. more able to filter for 'clean' data. The corollary may also be true; active data selection appears to greatly improve the utility of ALIGN, suggesting that it contains a large proportion of data that is of low quality for training." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.489, + 0.772, + 0.606 + ], + "angle": 0, + "content": "We observed the same effect when training with ActiveSigLIP on the large scale Webli dataset [6] (results summarised in Table 3). We trained reference models on both the raw 4B set as well as two extensively curated 1B / 350M subsets, which were then used to train subsequent learner models (Fig. 7). In all-but-one cases, filtering data with the 350M-trained referenced model produced the best results when transferring to 350M, 1B and 4B learner datasets. Notably, the significant gains observed over training with uniform sampling on the 350M subset suggest that our method can still improve over methods that pre-filter datasets once but then train with uniform samples [12]." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.639, + 0.557, + 0.652 + ], + "angle": 0, + "content": "4.7 Comparison to prior multimodal art" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.675, + 0.77, + 0.793 + ], + "angle": 0, + "content": "We leverage the insight that pre-training a reference model on clean data can facilitate learning on larger, noisy data for training our final model. Here, we train a reference model on LTIP, then use it to train a new model on the much larger mixture of LTIP and ALIGN, following [2], for a total of 3 and 8 billion training examples. Table 3 shows that in this training regime, ActiveCLIP surpasses models trained with significantly more data on ImageNet 0-Shot classification and COCO retrieval. Finally, we find that our active learning method is complementary to recent advances in multimodal learning: ActiveSigLIP significantly improves uniformly-trained SigLIP [44] in both COCO retrieval metrics." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.235, + 0.17, + 0.253, + 0.18 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.17, + 0.381, + 0.18 + ], + "angle": 0, + "content": "T. Evans et al." + }, + { + "type": "table", + "bbox": [ + 0.324, + 0.194, + 0.675, + 0.335 + ], + "angle": 0, + "content": "
MethodTrain ex.IN-1K ZS Top-1COCO
im2txttxt2im
CLIP13B68.352.433.1
EVA-CLIP3B+2B69.7†
ActiveCLIP3B71.357.743.0
OpenCLIP34B70.259.442.3
EVA-CLIP8B+2B74.7†58.742.2
ActiveCLIP8B72.260.744.9
SigLIP3B72.160.742.7
ActiveSigLIP3B72.063.545.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.232, + 0.343, + 0.772, + 0.416 + ], + "angle": 0, + "content": "Table 3: Comparison of ActiveCLIP to public multimodal pre-taining methods. ActiveCLIP outperforms models trained with the same or more data (CLIP, [30]; EVA-CLIP, [37]; and OpenCLIP, [17]). ActiveSigLIP produced significant gains over the baseline SOTA performance of SigLIP [44]. \\(^\\dagger\\)benefits from additional ImageNet21K pretraining (+2B training examples). All ActiveCLIP/SigLIP models use a reference model trained on LTIP to guide learning on a mixture of ALIGN and LTIP." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.443, + 0.364, + 0.457 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.501, + 0.77, + 0.683 + ], + "angle": 0, + "content": "In this work, we have presented a new method for active data selection that builds upon and simplifies the concept of 'learnability'. Our experiments demonstrate that this approach can significantly reduce the computation required for large-scale pretraining, compared to training with uniform samples. To our knowledge, this is the first active learning method that is more efficient than training with uniform samples when accounting for total FLOPs, and that does not rely on hand-designed features, allowing broad application across training setups. We have validated this by showing results on classification and contrastive pretraining, and found that our data selection policies continue to produce efficiency gains in the large-scale regime and can generalize effectively across task modalities. Collectively, our experiments also illustrate a Pareto frontier that allows trading off actor/data-selection computation against savings in training iterations, suggesting an alternative path to improved performance beyond scaling training batch sizes." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.688, + 0.771, + 0.793 + ], + "angle": 0, + "content": "This work focused on supervised pre-training for images, but further work could involve extending our method to other modalities and training schemes such as language, video, and generative modeling. An important note is that all our experiments present results from filtering only \\(50\\%\\) of the data; further gains may be possible by filtering more aggressively, at the cost of further overheads. In particular, aggressive data-selection coupled with efficient scoring schemes such as the ones proposed here could test the hypothesis that large-scale pretraining can benefit from exponential, rather than power-law, scaling behavior." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.169, + 0.721, + 0.182 + ], + "angle": 0, + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.753, + 0.17, + 0.769, + 0.18 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.195, + 0.335, + 0.209 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.228, + 0.767, + 0.263 + ], + "angle": 0, + "content": "1. Abbas, A., Tirumala, K., Simig, D., Ganguli, S., Morcos, A.S.: Semdedup: Data-efficient learning at web-scale through semantic dedduplication. arXiv preprint arXiv:2303.09540 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.266, + 0.769, + 0.3 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: A visual language model for few-shot learning. Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.302, + 0.767, + 0.338 + ], + "angle": 0, + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.34, + 0.767, + 0.374 + ], + "angle": 0, + "content": "4. Campbell, T., Broderick, T.: Bayesian coreset construction via greedy iterative geodesic ascent. In: International Conference on Machine Learning. pp. 698-706. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.376, + 0.767, + 0.4 + ], + "angle": 0, + "content": "5. Cassirer, A., Barth-Maron, G., Brevdo, E., Ramos, S., Boyd, T., Sottiaux, T., Kroiss, M.: Reverb: A framework for experience replay (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.401, + 0.767, + 0.437 + ], + "angle": 0, + "content": "6. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.438, + 0.767, + 0.485 + ], + "angle": 0, + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.487, + 0.767, + 0.522 + ], + "angle": 0, + "content": "8. Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., Zaharia, M.: Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.524, + 0.767, + 0.571 + ], + "angle": 0, + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.572, + 0.767, + 0.62 + ], + "angle": 0, + "content": "10. Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih, V., Ward, T., Doron, Y., Firoiu, V., Harley, T., Dunning, I., et al.: Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures. In: International conference on machine learning. pp. 1407-1416. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.621, + 0.767, + 0.657 + ], + "angle": 0, + "content": "11. Feldman, V.: Does learning require memorization? a short tale about a long tail. In: Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing. pp. 954-959 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.658, + 0.767, + 0.694 + ], + "angle": 0, + "content": "12. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.695, + 0.767, + 0.731 + ], + "angle": 0, + "content": "13. Gunasekar, S., Zhang, Y., Aneja, J., Mendes, C.C.T., Del Giorno, A., Gopi, S., Javaheripi, M., Kauffmann, P., de Rosa, G., Saarikivi, O., et al.: Textbooks are all you need. arXiv preprint arXiv:2306.11644 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.732, + 0.767, + 0.768 + ], + "angle": 0, + "content": "14. Har-Peled, S., Kushal, A.: Smaller coresets for k-median and k-means clustering. In: Proceedings of the twenty-first annual symposium on Computational geometry. pp. 126-134 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.769, + 0.767, + 0.793 + ], + "angle": 0, + "content": "15. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.235, + 0.228, + 0.769, + 0.793 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.235, + 0.17, + 0.253, + 0.18 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.285, + 0.17, + 0.38, + 0.18 + ], + "angle": 0, + "content": "T. Evans et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.197, + 0.77, + 0.245 + ], + "angle": 0, + "content": "16. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.245, + 0.769, + 0.292 + ], + "angle": 0, + "content": "17. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below." + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.292, + 0.77, + 0.338 + ], + "angle": 0, + "content": "18. Jouppi, N.P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., Bates, S., Bhatia, S., Boden, N., Borchers, A., et al.: In-datacenter performance analysis of a tensor processing unit. In: Proceedings of the 44th annual international symposium on computer architecture. pp. 1-12 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.338, + 0.769, + 0.374 + ], + "angle": 0, + "content": "19. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.374, + 0.769, + 0.409 + ], + "angle": 0, + "content": "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.409, + 0.769, + 0.432 + ], + "angle": 0, + "content": "21. Lindley, D.V.: On a measure of the information provided by an experiment. The Annals of Mathematical Statistics 27(4), 986-1005 (1956)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.432, + 0.769, + 0.454 + ], + "angle": 0, + "content": "22. Loshchilov, I., Hutter, F.: Online batch selection for faster training of neural networks. arXiv preprint arXiv:1511.06343 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.455, + 0.769, + 0.478 + ], + "angle": 0, + "content": "23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.478, + 0.769, + 0.5 + ], + "angle": 0, + "content": "24. MacKay, D.J.: Information-based objective functions for active data selection. Neural computation 4(4), 590-604 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.501, + 0.769, + 0.535 + ], + "angle": 0, + "content": "25. Mahmoud, A., Elhoushi, M., Abbas, A., Yang, Y., Ardalani, N., Leather, H., Morcos, A.: Sieve: Multimodal dataset pruning using image captioning models. arXiv preprint arXiv:2310.02110 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.535, + 0.769, + 0.571 + ], + "angle": 0, + "content": "26. Marion, M., Üstün, A., Pozzobon, L., Wang, A., Fadaee, M., Hooker, S.: When less is more: Investigating data pruning for pretraining llms at scale. arXiv preprint arXiv:2309.04564 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.571, + 0.769, + 0.617 + ], + "angle": 0, + "content": "27. Mindermann, S., Brauner, J.M., Razzak, M.T., Sharma, M., Kirsch, A., Xu, W., Höltgen, B., Gomez, A.N., Morisot, A., Farquhar, S., et al.: Prioritized training on points that are learnable, worth learning, and not yet learnt. In: International Conference on Machine Learning. pp. 15630-15649. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.617, + 0.769, + 0.652 + ], + "angle": 0, + "content": "28. Paul, M., Ganguli, S., Dziugaite, G.K.: Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems 34, 20596-20607 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.653, + 0.769, + 0.675 + ], + "angle": 0, + "content": "29. Prabhu, A., Dognin, C., Singh, M.: Sampling bias in deep active classification: An empirical study. arXiv preprint arXiv:1909.09389 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.675, + 0.769, + 0.722 + ], + "angle": 0, + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.722, + 0.769, + 0.744 + ], + "angle": 0, + "content": "31. Schaul, T., Quan, J., Antonoglou, I., Silver, D.: Prioritized experience replay. arXiv preprint arXiv:1511.05952 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.744, + 0.769, + 0.792 + ], + "angle": 0, + "content": "32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.235, + 0.197, + 0.77, + 0.792 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.169, + 0.72, + 0.182 + ], + "angle": 0, + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.753, + 0.17, + 0.769, + 0.18 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.234, + 0.197, + 0.769, + 0.234 + ], + "angle": 0, + "content": "33. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.234, + 0.581, + 0.245 + ], + "angle": 0, + "content": "34. Settles, B.: Active learning literature survey (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.246, + 0.769, + 0.281 + ], + "angle": 0, + "content": "35. Sorscher, B., Geirhos, R., Shekhar, S., Ganguli, S., Morcos, A.: Beyond neural scaling laws: beating power law scaling via data pruning. Advances in Neural Information Processing Systems 35, 19523-19536 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.282, + 0.769, + 0.317 + ], + "angle": 0, + "content": "36. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the IEEE international conference on computer vision. pp. 843-852 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.318, + 0.768, + 0.34 + ], + "angle": 0, + "content": "37. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.342, + 0.768, + 0.376 + ], + "angle": 0, + "content": "38. Toneva, M., Sordoni, A., Combes, R.T.d., Trischler, A., Bengio, Y., Gordon, G.J.: An empirical study of example forgetting during deep neural network learning. arXiv preprint arXiv:1812.05159 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.377, + 0.768, + 0.412 + ], + "angle": 0, + "content": "39. Xie, S.M., Pham, H., Dong, X., Du, N., Liu, H., Lu, Y., Liang, P., Le, Q.V., Ma, T., Yu, A.W.: Doremi: Optimizing data mixtures speeds up language model pretraining (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.413, + 0.768, + 0.435 + ], + "angle": 0, + "content": "40. Xie, S.M., Santurkar, S., Ma, T., Liang, P.: Data selection for language models via importance resampling. arXiv preprint arXiv:2302.03169 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.436, + 0.769, + 0.482 + ], + "angle": 0, + "content": "41. Yang, F., Barth-Maron, G., Stanczyk, P., Hoffman, M., Liu, S., Kroiss, M., Pope, A., Rrustemi, A.: Launchpad: A programming model for distributed machine learning research. arXiv preprint arXiv:2106.04516 (2021), https://arxiv.org/abs/2106.04516" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.484, + 0.769, + 0.519 + ], + "angle": 0, + "content": "42. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: CoCa: Contrastive captioners are image-text foundation models. In: Transactions on Machine Learning Research (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.52, + 0.769, + 0.554 + ], + "angle": 0, + "content": "43. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.235, + 0.555, + 0.769, + 0.579 + ], + "angle": 0, + "content": "44. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.234, + 0.197, + 0.769, + 0.579 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_origin.pdf b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fd7376bdacc8120e5eb3005f618bf78baff84e9b --- /dev/null +++ b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/8ddfb6b7-27c3-448b-9e04-68cc0c428437_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88957330faaf82fb9f523cc67e2762bafbff9b873c30e04a630456e323da4542 +size 1605391 diff --git a/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/full.md b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c6d858c7296bad64911d50a3700a38e77673deb --- /dev/null +++ b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/full.md @@ -0,0 +1,283 @@ +# Bad Students Make Great Teachers: Active Learning Accelerates Large-Scale Visual Understanding + +Talfan Evans1,* Shreya Pathak1,* Hamza Merzic1,2,* Jonathan Schwarz1,† Ryutaro Tanno1 Olivier J. Henaff1,* + +$^{1}$ Google DeepMind $^{2}$ University College London + +Abstract. Power-law scaling indicates that large-scale training with uniform sampling is prohibitively slow. Active learning methods aim to increase data efficiency by prioritizing learning on the most relevant examples. Despite their appeal, these methods have yet to be widely adopted since no one algorithm has been shown to a) generalize across models and tasks b) scale to large datasets and c) yield overall FLOP savings when accounting for the overhead of data selection. In this work we propose a method which satisfies these three properties, leveraging small, cheap proxy models to estimate "learnability" scores for datapoints, which are used to prioritize data for training much larger models. As a result, models trained using our methods - ClassAct and ActiveCLIP - require $46\%$ and $51\%$ fewer training updates and up to $25\%$ less total computation to reach the same performance as uniformly-trained visual classifiers on JFT and multimodal models on ALIGN, respectively. Finally, we find our data-prioritization scheme to be complementary with recent data-curation and learning objectives, yielding a new state-of-the-art in several multimodal transfer tasks. + +# 1 Introduction + +Power-law scaling for vision and language models [19,43] indicates that incremental improvements in model performance require orders of magnitude increases in computation. One of the key features of these empirical power-laws is that training data is sampled uniformly. In contrast, active data selection prioritizes computation on the data that maximally contributes to task performance [21,24,34], with the ultimate goals of improving data efficiency and reducing the cost of training. However, active data selection has yet to become a mainstay of large model training, since no existing algorithm satisfies the three properties of being a) robust to the choice of model and training task, b) scalable to large datasets and architectures, and c) more compute efficient end-to-end than training with uniform samples. + +![](images/b276811fb9999805e31e07d2fd1e3f9e55823a6a65f03efcfa09c4a679132726.jpg) +Fig. 1: Active learning accelerates large-scale visual understanding. For large-scale classification and multimodal learning tasks, prioritised training on data selected using our active selection methods ClassAct (left) and ActiveCLIP (right) requires significantly fewer updates to reach the final performance of uniform training. + +![](images/976b696fe455fa1da0a8f10fbcf0eadb744dc221dd135b88e425ea61e7dd4035.jpg) + +In the first instance, data selection based on hand-engineered filters (e.g. removing incorrectly shaped images or that only contain a single colour [2]) can trivially improve training efficiency at minimal computational overhead. However, such heuristics are limited in their effectiveness by the expertise of the human designer, and are not guaranteed to transfer to training of different models, data modalities, or tasks and incur significant effort to develop and tune. + +In contrast, model-based curation methods, which use the loss of the model itself to score examples, have shown promise by focusing training on 'hard' and omitting 'easy' data [35], but also the opposite (to exclude noise or other low-quality data) in both language modeling [13] and multimodal learning [15]. However, these methods often spend as much computation on the curation of datasets as is gained from subsequent pretraining, making them less compute-efficient than training on uniformly-sampled data. Finally, while several compute-efficient methods have been successfully deployed at small scale [8], they generally do not scale to even medium-sized datasets such as ImageNet. + +In this work, we propose an algorithm that satisfies the three properties of generality, scalability, and compute-positivity. The proposed framework uses small proxy models to compute learnability scores for candidate training data, resulting in significant training efficiency gains at an almost negligible overhead over standard uniform-training. We test two instantiations of this framework, ClassAct and ActiveCLIP / ActiveSigLIP, for large scale classification and multimodal pretraining, respectively. Our findings are summarized below: + +Benchmarking heuristics for large-scale pretraining: We investigate loss- and learnability-based prioritization [15, 27, 35] for large-scale classification and find that pretrained reference models are an essential component for accelerating learning, producing efficiency gains of up to $46\%$ . + +Generalizing selection policies across model scale: Secondly, we show that smaller models act as effective proxies for much larger ( $\sim 1000 \times$ ) models in the context of learnability but not loss-based scoring, resulting in compute-positive gains of up to $25\%$ over uniform training, a first in the context of large-scale pretraining. + +Accelerating multimodal pretraining: Using reference models pretrained on small, clean datasets, we substantially accelerate pretraining on much larger, + +noisier datasets. Moreover, we find ActiveCLIP to be complementary to recent data-curation techniques [12] and learning objectives [44], yielding a new state-of-the-art in several multimodal understanding tasks. + +Amortizing data selection policies: Data-selection policies trained on one task can also accelerate the training of subsequent models on different but related tasks, suggesting that such policies can be easily derived from pre-trained models. + +Simplifying active-learning with online reference models: Lastly, we demonstrate that pre-trained reference models may not be necessary at all, where these models are small and can be trained in parallel on larger batches than the learner model, while remaining compute-positive. + +# 2 Related Work + +Data pruning. One approach to data-selection is to identify and sub-select data ahead of training. For example, [28] and [35] show that the training loss and gradients can be used to discard large portions of small-to-medium sized datasets (e.g. CIFAR10 and ImageNet) with little loss in performance. These methods have since been deployed for the curation of web-scale datasets in both language modeling [26] and multimodal learning [1, 12, 25], demonstrating large reductions in the amount of data required together with performance improvements. However, in the single-epoch regime that is becoming typical of large model training [16, 19], pre-filtering can be as expensive as learning from it, a shortcoming which we address in this work. Nevertheless, we show that our method for dynamic data selection is complementary to and benefits from such data curation techniques. + +Online active learning. Unlike data-pruning, online active learning continuously filters data throughout training and applies naturally to the semi-infinite, single-epoch regime. Online Batch Selection [22] scores and filters using the learner model, which has the theoretical advantage that the importance of data can be determined relative to the current state of the learner. In terms of metrics, the Reducible Holdout Loss (RHO) [27] also uses the concept of a reference model to identify learnable data points not yet well represented. Other proposed heuristics include memorization for long-tailed data [11] and assigning "complexity" scores based on the number of times the example is forgotten during training [38]. None of these approaches however have demonstrated that the cost of scoring can be reduced to the point of justifying learner efficiency gains. + +Compute-efficient data selection. Several works have demonstrated the benefits of selecting data based on simple heuristics, such as low-level image properties [2] or proximity to high-quality text corpora [3, 7, 29, 40]. While cheap to compute, these statistics often require domain-specific knowledge which limits their applicability across tasks. Domain-agnostic methods such as core-sets alleviate this by selecting data based on the geometry of their embeddings [4, 14] which can be efficiently computed, however these algorithms generally do not + +scale to large-scale datasets [8]. Most related to our work is DoReMi [39] which uses domain-general, scalable, and compute-efficient proxy models for the simpler problem of determining optimal data-mixtures for the subsequent training of a larger language model. + +# 3 Methods + +# 3.1 Data selection as prioritized replay + +We use online batch selection [22] to apply our scoring heuristics to standard visual learning tasks: firstly, we sample uniformly from the training set $\boldsymbol{x}_i \stackrel{\mathcal{U}}{\sim} \mathcal{D}$ and compute a score $s_i = s(\boldsymbol{x}_i | \theta) \in \mathbb{R}$ to each data point $\boldsymbol{x}_i$ using model parameters $\theta$ . Given a large enough collection of scored examples stored in a memory bank $\mathcal{M} = \{\boldsymbol{x}_i\}_{i \in (0, \dots, M-1)}$ , we then sample non-uniformly according to their scores $\boldsymbol{x}_i \stackrel{\pi}{\sim} \mathcal{M}$ [31], where $\pi(\boldsymbol{x}_i) = \text{Softmax}(\{s_i\}_{i \in (0, \dots, M-1)})$ . A batch of such examples is used to update the learner model. Following convention in reinforcement learning, we refer to the scoring and target models as actors and learners respectively. + +# 3.2 Statistics for data selection + +We explore a few statistics for model-based prioritization, grouped into two categories. + +Example difficulty: given the current state of the learner, an intuitive prioritization scheme might favour 'difficult' examples (as measured by their training loss), while removing 'easy' examples that are trivially classified and which yield small gradients. This loss-based prioritization: + +$$ +s ^ {\text {h a r d}} \left(\boldsymbol {x} _ {i} \mid \theta\right) = \ell \left(\boldsymbol {x} _ {i} \mid \theta\right) \tag {1} +$$ + +can use the current parameters of the learner $\theta^t$ or those of a fixed model $\theta^{*}$ . The opposite argument can be made for favoring examples that are easily solved by a well-trained model, as such a prioritization removes the noisy examples present in large-scale datasets: + +$$ +s ^ {\text {e a s y}} \left(\boldsymbol {x} _ {i} | \theta\right) = - \ell \left(\boldsymbol {x} _ {i} | \theta\right) \tag {2} +$$ + +This scheme is commonly used in multimodal learning for identifying high-quality examples with pre-trained models [15,32,33]. + +Example learnability: Given that favoring easy and hard examples target different and potentially orthogonal properties of the data, a natural question is whether these policies can be combined. Learnability criteria straightforwardly combine the two as + +$$ +\begin{array}{l} s ^ {\text {l e a r n}} \left(\boldsymbol {x} _ {i} | \theta^ {t}, \theta^ {*}\right) = s ^ {\text {h a r d}} \left(\boldsymbol {x} _ {i} | \theta^ {t}\right) + s ^ {\text {e a s y}} \left(\boldsymbol {x} _ {i} | \theta^ {*}\right) (3) \\ = \ell \left(\boldsymbol {x} _ {i} \mid \theta^ {t}\right) - \ell \left(\boldsymbol {x} _ {i} \mid \theta^ {*}\right), (4) \\ \end{array} +$$ + +favoring examples that are easily solved by a well-trained model $\theta^{*}$ but challenging to the learner in its current state $\theta^t$ , such that more computation dedicated to this example could lower its loss. Conversely, examples that are trivially classified by the learner (or mislabeled) will yield low (or high) losses for both the current learner and the well-trained one, leading to low learnability scores. + +A special case of learnability scores (the RHO loss, [27]) uses a model $\theta^{\mathrm{ho}}$ specifically trained on a held-out dataset to ensure the independence of its predictions from those of the current learner $s^{\mathrm{learn}}(\pmb{x}_i|\theta^t,\theta^{\mathrm{ho}})$ . We assess in Section 4.2 whether this is necessary when training on large-scale image datasets. + +![](images/6dc4c9e120688085749e7bef4da43f4d298326e970f431a67df88065169b146e.jpg) +Fig. 2: Amortizing the cost of data selection. Drawn to scale: length of bars indicates number of FLOPs required to reach the accuracy of a ViT-L trained with uniform sampling ("ViT-L Uniform Sampling", see Figure 4). Expensive model policies (e.g. a ViT-B scores data for the ViT-L learner, or $\mathrm{B} \rightarrow \mathrm{L}$ ) produce large learner speedups, at the expense of the additional FLOPs associate with data selection. This overhead can be reduced by deriving the data-selection policies from smaller models (e.g. ViT-S, ViT-Ti or ViT-Mu score data for the ViT-L learner), at the expense of marginal decreases in the learner speedup. Costs could be additionally amortized by using off-the-shelf reference models, removing the need to train from scratch (yellow). Since the reference model is fixed throughout training, scores can be assigned once to a 'foundation dataset' and amortized across many training runs (lime green; [27,35]). Since the online model is independent of the learner model and generalizes across scale, data selection policies can also be distilled as a fixed ordering of a given dataset (a 'foundation curriculum'). + +# 3.3 Unlocking compute-positive training + +Scoring requires inference passes over the actor and learner models. We assume that gradient updates to cost $3\mathrm{x}$ inference passes. The cost of scoring data $F$ thus scales with the proportion of data which is being rejected (e.g. retaining only $20\%$ of the data requires 5 inference passes per trained batch). The requirements for compute-positivity can therefore be expressed as: + +$$ +\underbrace {\left(3 F _ {\text {l e a r n}} + \rho F _ {\text {a c t}}\right) \beta + 3 F _ {\text {r e f}}} _ {\text {A c t i v e L e a r n i n g}} < \underbrace {3 F _ {\text {l e a r n}}} _ {\text {U n i f o r m S a m p l i n g}} \tag {5} +$$ + +where $F_{\mathrm{act}}$ is the cost of scoring an example, $\rho$ is the number of examples scored per training example, and $\beta$ is the efficiency gain (in terms of the learner update) + +relative to training with uniform sampling (see Appendix Section A.1 for more details). The RHS term is the cost per update of uniform sampling. The first LHS term inside the brackets is the cost of the learner training during AL, the second term is the scoring cost and the last term outside the brackets the cost of training the reference model. We illustrate in Figure 2 the different contexts in which parts of this computation may be effectively amortized. + +In the large-scale training regime where data is neither repeated nor seen before, compute-positivity requires that either or all of the reference model training, actor scoring and learner efficiency $\beta$ terms must be made smaller to produce net savings relative to uniform sampling. Typical prioritization schemes can produce saving on the order of $50\%$ (ie $\beta = 0.5$ ), suggesting that savings must also be made by down-scaling the other terms $F_{\mathrm{act}}$ and $F_{\mathrm{ref}}$ . + +Cost of easy-reference scoring. While both the cost of the reference model and example scoring can be scaled down in the case of easy-reference scoring with small models ( $F_{\mathrm{act}} = F_{\mathrm{ref}}$ , see Equation 2), it is unclear whether the efficiency gains $\beta$ are robust to this down-scaling (see section 4.2). + +Cost of RHO learnability scoring. The original definition of learnability scores [27] requires inference passes through both the learner and a reference model ( $F_{\mathrm{act}} = F_{\mathrm{ref}} + F_{\mathrm{learn}}$ , see Equation 3), meaning that although the cost of the reference model can be reduced by using a smaller model, the cost of example scoring cannot. + +Cost of ClassAct / ActiveCLIP. For this reason, we explore whether replacing the learner model in term 1 of Equation 3 with a much smaller model can still produce comparable learner efficiency gains to those already observed (see Appendix Algorithm 1). Specifically, we introduce a third "online" model, which has the same architecture and size as the reference model, but is trained in parallel with the learner. In this case, the cost of scoring examples reduces to: + +$$ +F _ {\text {a c t}} = F _ {\text {r e f}} + F _ {\text {o n l i n e}} = 2 F _ {\text {r e f}} \tag {6} +$$ + +and can be scaled down along with the reference model. We instantiate our method for two canonical pre-training tasks: visual classification and multimodal learning, which we call ClassAct, ActiveCLIP and ActiveSigLIP respectively. In Appendix Section A.6, we describe an asynchronous active learning framework where scoring and learning is performed in parallel on separate devices. This setup does not affect the FLOP calculations but can mitigate the overhead in time even when using non-approximate scoring. + +# 3.4 Losses for canonical visual pre-training tasks + +For visual classification with ClassACT, we use the standard cross-entropy loss for both actors and learners. For multimodal learning with ActiveCLIP, learners optimize the contrastive loss $\ell_{\mathrm{learn}} = \ell_{\mathrm{learn}}^{\mathrm{im,txt}} + \ell_{\mathrm{learn}}^{\mathrm{txt,im}}$ , whereas the actor loss $\ell_{\mathrm{act}}$ + +# Algorithm 1 ClassAct / ActiveCLIP + +1: Input: Randomly initialized learner model $\theta_{l}$ and small online model $\theta_{o}$ , small pre-trained reference model $\theta_{r}$ . Models use loss $\ell_{\mathrm{act}}$ for scoring data and $\ell_{\mathrm{learn}}$ for computing updates. Dataset $\mathcal{D}$ , batch size $B$ , sub-batch size $b < B$ . +2: while training do +3: $X\sim \mathcal{D}$ , where $|X| = B$ Sample uniformly +4: $S = \ell_{\mathrm{act}}(X|\theta_o) - \ell_{\mathrm{act}}(X|\theta_r)$ +5: $I\sim \mathrm{SoftMax}(S)$ , where $|I| = b$ Sample indices +6: $Y = X[I]$ $\triangleright$ Collect sub-batch +7: $\theta_{l}\gets \mathrm{Adam}[\nabla_{\theta_{l}}\ell_{\mathrm{learn}}(Y|\theta_{l})]$ +8: $\theta_o\gets \mathrm{Adam}[\nabla_{\theta_o}\ell_{\mathrm{learn}}(Y|\theta_o)]$ Update online model +9: end while + +is simply the dot-product similarity between image and text embeddings: + +$$ +\ell_ {\text {a c t}} \left(\boldsymbol {x} _ {i} \mid \theta\right) = - \boldsymbol {z} _ {i} ^ {\mathrm {i m}} \cdot \boldsymbol {z} _ {i} ^ {\mathrm {t x t}} \tag {7} +$$ + +$$ +\ell_ {\text {l e a r n}} ^ {\mathrm {i m}, \mathrm {t x t}} (\boldsymbol {x} _ {i} | \theta) = - \log \frac {\exp \left(\boldsymbol {z} _ {i} ^ {\mathrm {i m}} \cdot \boldsymbol {z} _ {i} ^ {\mathrm {t x t}}\right)}{\sum_ {j} \exp \left(\boldsymbol {z} _ {i} ^ {\mathrm {i m}} \cdot \boldsymbol {z} _ {j} ^ {\mathrm {t x t}}\right)} \tag {8} +$$ + +where $\boldsymbol{z}_i^{\mathrm{im}} = f^{\mathrm{im}}(\boldsymbol{x}_i; \theta)$ and $\boldsymbol{z}_i^{\mathrm{txt}} = f^{\mathrm{txt}}(\boldsymbol{x}_i; \theta)$ are image and text embeddings respectively, and $\ell_{\mathrm{learn}}^{\mathrm{txt}, \mathrm{im}}(\boldsymbol{x}_i; \theta)$ is defined analogously to $\ell_{\mathrm{learn}}^{\mathrm{im}, \mathrm{txt}}(\boldsymbol{x}_i; \theta)$ . Similarly, ActiveSigLIP instead uses the sigmoid loss [44] for the learner's objective and $\ell_{\mathrm{act}}$ for scoring. + +# 4 Experiments + +All our experiments were conducted with Vision Transformers [9] for which strong baselines are available across model sizes [43]. Unless specified, we adopt models with patch-size 16 throughout (ViT-S refers to ViT-S/16 and similar). We consider two canonical tasks for large-scale pretraining: classification on JFT-300M [36] and multimodal contrastive learning [30] on large image-text datasets. When pre-training with JFT classification we use held-out classification performance as the evaluation metric. When pre-training on image-text data we evaluate with standard multimodal transfer tasks: ImageNet zero-shot classification and image-to-text / text-to-image retrieval on COCO. + +Throughout, we will refer to the large batch of size $B$ sampled uniformly from the training data as the 'super-batch', and the prioritised smaller batch of size $b < B$ as the 'sub-batch'. In all our experiment, we filter 50% of uniformly sampled data such that $\rho = B / b = 2$ , although more aggressive filtering regimes warrant investigation [35]. + +# 4.1 Evaluating loss-based scoring heuristics in the large-data regime + +We begin by evaluating loss- and learnability-based heuristics on their ability to accelerate supervised classification on JFT (Fig. 3). Arguably the most intuitive + +![](images/dd270feee29f95d773fd59a75a31aca907bb97690a0fb075dfbd83db1bb0884e.jpg) +Fig. 3: Evaluation of loss-based data-selection criteria for large-scale classification. We train a ViT-B on JFT-300M with different data-selection policies. Prioritising hard data under the learner (green curve) produced marginal gains over the uniform sampling baseline. Prioritizing data using both learnability (blue curve, [27]) and easy reference prioritization (red curve, [15]) produced significant speedups and performance gains. + +![](images/3fda2293d0d429e1180327358be41fc151e8dabef54c06ba30731ca0947e9ce1.jpg) +Fig. 4: Generalization of data-selection policies across models scales. Left: We train a ViT-L for 3 epochs on JFT using uniform sampling (grey) or prioritized data sampling using example learnability (blue) or low-loss under the reference model (red). Example scores are computed using ViT-B actors (dark), or cheaper ViT-S or ViT-Tiny models (light). While both example learnability and "easy reference" yield good speedups with expensive actors, learnability criteria are much more robust to approximate scoring. Top right: Learner (ViT-L) speedup is computed as the fraction of learner iterations saved in order to attain the baseline's top performance. Actor overhead is computed as the additional computation in FLOPs required to score examples with a particular actor architecture (varying from ViT-Mu to ViT-L, see Appendix Table 4). Example learnability yields robust learner speedups across actor scales, "easy reference" scoring does not. Lower right: total compute efficiency is calculated as a product of learner efficiency and actor overhead, indicating the amount of computation required to reach baseline performance. Approximate actors (i.e. ViT-S or smaller) computing example learnability enable total compute speedups, other schemes do not. + +![](images/434dd7f91d42456df41f0e7afdf1d4742ab31b1a6f6cef39ed93055f372a7bf1.jpg) + +method to score data is to prioritise training on data with high loss under the learner (hard learner). In our experiments, this strategy (Eq. 1) only marginally improved performance over the uniform sampling baseline, despite requiring an additional inference pass over the super-batch. This is perhaps not surprising - data points with high loss may also be unlearnable due to e.g. label noise, such that training on those data points does not result in the model performing any better on the held-out test set. Large scale datasets are more likely to be noisy. + +Scoring methods based on pre-trained reference models performed much better—both easy reference (equation 2) and learnability (equation 3) -based prioritization produced significant gains over uniform sampling. Here, we pretrained an identical ViT-B for the same 3 epochs to use as a reference model for a second training run displayed above. producing speed-ups of $33\%$ (Fig. 3). + +![](images/af975683167a5ea5d947d933a3f525fa4937cf3e5330025768dc300311e5ed53.jpg) +Fig. 5: Scaling laws for active learning. We trained a baseline ViT-L over a range of compute budgets (for which ViT-L is compute optimal, see Zhai et al., 2021). We also trained the same ViT-L with both ViT-Ti and ViT-S reference policies, pre-trained for the same number of epochs. Left: Small model policies produce robust savings in learner compute. Right: When accounting for total compute (learner + actor training and data scoring), small model policies in all compute budgets produce FLOP savings over training with uniform samples. These scaling laws generalize those measured empirically in the uniform sampling setting [43] to the case of non-uniform data selection. + +![](images/4918a10c2f728b7e4d882f333f6af743f3035ac6b058739f4c3c669bca288946.jpg) + +# 4.2 Generalising data-selection policies across scale + +The speed-ups afforded by learnability prioritization (Fig. 3) come at the cost of the additional inference passes required to score the data during learner training, plus the cost of training the reference model. This makes the overall gains strongly compute-negative relative to training with uniform samples. Even if the size of the reference model is scaled down [27], these methods still incur the cost of additional learner inference passes to score the data during training. + +Unlocking compute-positive active learning. To address this issue, we introduce a set of down-scaled models with the same ViT architecture that we use to score data for training a larger ViT-L model (see Methods Section 3.3. We use ViT-B, ViT-S or ViT-Ti variants (which are $4 \times$ , $13 \times$ and $47 \times$ cheaper than + +
MethodViT model capacityReference TypeSpeed-up
Reference Online LearnerLearner speedupCompute speedup
Uniform Sampling (ViT-B)B0%0%
RHOTinyBBHeld-out, fixed0%-79%
ClassAct-HOTinyTinyBHeld-out, fixed18%3%
ClassActTinyTinyBIn-domain, fixed18%3%
ClassAct-OnlineTinyTinyBTrained online17%2%
+ +Table 1: Simplifying and accelerating the computation of learnability scores. + +Relative to RHO [27], ClassAct makes two changes: replacing the reference model with one trained in-domain (removing the need for bespoke held-out sets), and dramatically reducing capacity of the online actor models used for scoring examples. All experiments were conducted on 3 epochs of one-half of JFT to enable the held-out ablations. RHO with a small reference model did not produce a learner speedup in our experiments. + +the learner) for both the online and reference model (see Appendix Algorithm 1). In Figure 4 (left) we assess the impact of these cheaper scoring models on learner efficiency. First, we find that easy reference prioritization to be very sensitive to the capacity of the scoring model: while ViT-B scoring models yield reasonable gains over uniform sampling, prioritizing with ViT-S and ViT-Ti scoring models underperforms significantly (Fig. 4, red curves). + +In contrast, we find that learnability based prioritization yields robust gains, even when the scoring models are significantly scaled down (ClassAct; Fig. 4, blue curves). For example, while ViT-B scoring models yield a $31\%$ learner speedup, the $50 \times$ smaller ViT-Ti scoring models still provide a $26\%$ speedup. We pushed this logic by using even smaller scoring models (the ViT-Mu family which we introduce, see Appendix) which are up to $1000 \times$ smaller than the learner. Despite this, prioritizing data based on their scores yields non-negligible speedups (e.g. $16\%$ for the smallest actors we consider; Figure 4, top right). + +These experiments demonstrate that, with the appropriate scoring criterion, online and reference models can be significantly downscaled and still produce comparable gains to larger models, with learner efficiency degrading gracefully with the actor overhead (i.e. the cost of the reference model and data scoring). As a result, our method ClassAct quickly becomes FLOP positive as the online + reference models are downscaled (Figure 4, bottom right), while at the same time producing speed-ups in wall-clock time for a given learner batch size. + +Together, our results expose a pareto front across which to determine an optimal context-specific data selection strategy (Figure 2). Where pre-trained models are available, some of the cost of larger data selection policies can be discounted. If savings in wall-clock time supersede the associated cost of scoring, large models can be tolerated for data selection. Reference model costs can also be amortized across many training runs by appending scores to 'foundation datasets'. However, in the case where no component of the framework can amortized (as in the case of large-scale pretraining), prioritizing data with small ClassAct models can deliver large savings in total computation. + +# 4.3 Generalising neural scaling laws to the active-learning setting + +We next investigated the scaling behaviour of ClassAct by experimenting over large learner compute budgets (Fig. 5), using both ViT-Ti and ViT-S models as actors for training a ViT-L. Predictably, the ViT-S produced larger, although marginal gains over the ViT-Ti actors when not accounting for scoring FLOPs (Figure 5, left). However, when accounting for total FLOPs, the difference was less pronounced (Figure 5, right). Our results generalize large scale uniform sampling scaling laws such as uncovered by [19] for LLMs and reproduced for large vision transformers by [43] to the case of non-uniform sampling. For the first time, we demonstrate that these scaling laws can be shifted in our favour by selecting data using general model-based scoring heuristics. + +# 4.4 Training the reference model in parallel + +The reference model needs to be trained if none is already available. This two-step process adds complexity to active model training, especially if using large scale infrastructure. However, an interesting consequence of down-scaling the reference model is that both inference passes and gradients can be computed over a much larger batch than can be computed on the learner. In theory, this would mean that the small reference model could instead be trained online, in parallel with the large learner and small online model. + +We confirmed our hypothesis by running an experiment in which we trained our reference model on a super-batch of size $B = 10b$ and trained the online and learner model in sequence with the sub-batch of size $b$ . To make sure the reference model quickly converges, we additionally set the learning rate to double that of the online and learner models (this would cause instability for the learner and online models because of the additional variance from the smaller batch). We also verified that training the reference model on a held-out set of data performed equally in our experiments to reference models trained on the same data as the learner model [27]. Our 'one-pass' setup, Online-ClassAct, produces the same performance as the pre-trained ClassAct pipeline in our Ti-trains-B experiments (Table 1). Pseudocode is shown in appendix Algorithm 1. + +We have shown that by decoupling the scoring models from the learner model entirely, it is possible to significantly downscale the scoring models with minor degradation to performance (see Table 1). Unlike RHO, which can train a large 'learner' model with a small 'reference' model, we introduce a third 'online' model, with the same architecture and parameter count as the reference model, enabling the reduction of actor computation (see Table 1). + +# 4.5 ActiveCLIP: active multimodal learning + +We have so far demonstrated that large scale image classifiers can be trained with lower total compute by actively selecting the data used for training. However, classification has largely been superseded as a large scale pre-training method + +
ImageNet0-ShotCOCO (im2txtR1)COCO (txt2imR1)(Learner)
Speed-up % (vs. Uniform)23.048.128.648.026.827.0ALIGN
0.016.318.822.128.217.5LTIP
Performance (vs. Uniform)47.8 (2.2)53.2 (7.4)40.9 (2.4)44.8 (6.3)27.3 (1.8)30.5 (5.0)ALIGN
46.5 (-0.1)47.2 (0.6)45.6 (1.2)45.4 (1.0)31.8 (1.7)31.0 (0.9)LTIP
(Reference)ALIGNLTIPALIGNLTIPALIGNLTIP
+ +Table 2: Generalizing data-selection policies across datasets and tasks. We pretrain reference models on the large but noisy ALIGN dataset, or the smaller and more curated LTIP dataset [2]. Consistent with [2], we find training with uniform sampling on LTIP to yield stronger transfer learning results than pre-training on ALIGN. These reference models can be used very effectively for data-selection on both LTIP and ALIGN, whereas ALIGN-pretrained reference models yield more modest speedups. All models are provided with 800M training images at resolution $128 \times 128$ , speedups are shown relative to the time at which the uniform sampling baseline was reached for that evaluation metric. Colour indicates performance relative to uniform (brackets). + +![](images/b840da8d161994ef6e8d8f5a76ac04441e1a2f3be9b0ce6906093daaa4b6e49b.jpg) +Fig. 6: Reference policies generalize across tasks. We train a ViT-B reference model on either ALIGN or LTIP, then use it to train a second ViT-B learner on ALIGN with ActiveCLIP. The biggest gains were found with an LTIP reference model, despite it needing to perform out-of-domain generalization. In 50k iterations, ActiveCLIP selects the 800M "cleanest" examples from the ALIGN dataset, whose size is 1.6B in total. + +![](images/5fa7cd19b63e436f2c84975c09634e2d56b67ded39bd586dfea33e1978a14457.jpg) + +by CLIP-style multimodal training [30]. Figure 6 demonstrates that our CLIP-adapted active learning method ActiveCLIP (see Methods) produces similar speedups in terms of learner computation as observed in JFT classification. Specifically, we find that prioritized sampling with learnability scores accelerates multimodal pre-training by $18 - 48\%$ , depending on the evaluation metric (ImageNet zero-shot accuracy or COCO retrieval) and reference model configuration, which we explore below. + +# 4.6 Policy generalization across tasks + +To fully leverage off-the-shelf pre-trained image models for training (Figure 2), our results up to now suggest that the reference model should be trained on the same task that the learner model is being trained for. In Figure 6 (see Table 2 for full results), we show that we can in fact pre-train our reference models on related but distinct datasets. Moreover results suggest that there may even be a benefit + +![](images/1f4b4319b4f6f88c3ecccf5b128d2d33b845ad0f6872e07e8728ec8f31e9c94f.jpg) +Fig. 7: Reference models trained on curated datasets are powerful data selectors. Active data selection using ActiveSigLIP with reference models trained on increasingly curated subsets of Webli. Webli-350M reference models are consistently more effective than those trained on raw or 1B subsets. + +to cross-training in some-cases; an ALIGN $\rightarrow$ ALIGN reference $\rightarrow$ learner model combination ('in-domain' ActiveCLIP) produced similar speedups to ClassAct on JFT. However, these gains were greatly surpassed by an LTIP $\rightarrow$ ALIGN combination. One possibility is that LTIP is less noisily labeled such that the scoring policies it produces are 'clean' - i.e. more able to filter for 'clean' data. The corollary may also be true; active data selection appears to greatly improve the utility of ALIGN, suggesting that it contains a large proportion of data that is of low quality for training. + +We observed the same effect when training with ActiveSigLIP on the large scale Webli dataset [6] (results summarised in Table 3). We trained reference models on both the raw 4B set as well as two extensively curated 1B / 350M subsets, which were then used to train subsequent learner models (Fig. 7). In all-but-one cases, filtering data with the 350M-trained referenced model produced the best results when transferring to 350M, 1B and 4B learner datasets. Notably, the significant gains observed over training with uniform sampling on the 350M subset suggest that our method can still improve over methods that pre-filter datasets once but then train with uniform samples [12]. + +# 4.7 Comparison to prior multimodal art + +We leverage the insight that pre-training a reference model on clean data can facilitate learning on larger, noisy data for training our final model. Here, we train a reference model on LTIP, then use it to train a new model on the much larger mixture of LTIP and ALIGN, following [2], for a total of 3 and 8 billion training examples. Table 3 shows that in this training regime, ActiveCLIP surpasses models trained with significantly more data on ImageNet 0-Shot classification and COCO retrieval. Finally, we find that our active learning method is complementary to recent advances in multimodal learning: ActiveSigLIP significantly improves uniformly-trained SigLIP [44] in both COCO retrieval metrics. + +
MethodTrain ex.IN-1K ZS Top-1COCO
im2txttxt2im
CLIP13B68.352.433.1
EVA-CLIP3B+2B69.7†
ActiveCLIP3B71.357.743.0
OpenCLIP34B70.259.442.3
EVA-CLIP8B+2B74.7†58.742.2
ActiveCLIP8B72.260.744.9
SigLIP3B72.160.742.7
ActiveSigLIP3B72.063.545.3
+ +Table 3: Comparison of ActiveCLIP to public multimodal pre-taining methods. ActiveCLIP outperforms models trained with the same or more data (CLIP, [30]; EVA-CLIP, [37]; and OpenCLIP, [17]). ActiveSigLIP produced significant gains over the baseline SOTA performance of SigLIP [44]. $^\dagger$ benefits from additional ImageNet21K pretraining (+2B training examples). All ActiveCLIP/SigLIP models use a reference model trained on LTIP to guide learning on a mixture of ALIGN and LTIP. + +# 5 Discussion + +In this work, we have presented a new method for active data selection that builds upon and simplifies the concept of 'learnability'. Our experiments demonstrate that this approach can significantly reduce the computation required for large-scale pretraining, compared to training with uniform samples. To our knowledge, this is the first active learning method that is more efficient than training with uniform samples when accounting for total FLOPs, and that does not rely on hand-designed features, allowing broad application across training setups. We have validated this by showing results on classification and contrastive pretraining, and found that our data selection policies continue to produce efficiency gains in the large-scale regime and can generalize effectively across task modalities. Collectively, our experiments also illustrate a Pareto frontier that allows trading off actor/data-selection computation against savings in training iterations, suggesting an alternative path to improved performance beyond scaling training batch sizes. + +This work focused on supervised pre-training for images, but further work could involve extending our method to other modalities and training schemes such as language, video, and generative modeling. An important note is that all our experiments present results from filtering only $50\%$ of the data; further gains may be possible by filtering more aggressively, at the cost of further overheads. In particular, aggressive data-selection coupled with efficient scoring schemes such as the ones proposed here could test the hypothesis that large-scale pretraining can benefit from exponential, rather than power-law, scaling behavior. + +# References + +1. Abbas, A., Tirumala, K., Simig, D., Ganguli, S., Morcos, A.S.: Semdedup: Data-efficient learning at web-scale through semantic dedduplication. arXiv preprint arXiv:2303.09540 (2023) +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: A visual language model for few-shot learning. Advances in Neural Information Processing Systems (2022) +3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems (2020) +4. Campbell, T., Broderick, T.: Bayesian coreset construction via greedy iterative geodesic ascent. In: International Conference on Machine Learning. pp. 698-706. PMLR (2018) +5. Cassirer, A., Barth-Maron, G., Brevdo, E., Ramos, S., Boyd, T., Sottiaux, T., Kroiss, M.: Reverb: A framework for experience replay (2021) +6. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022) +7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023) +8. Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., Zaharia, M.: Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829 (2019) +9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021) +10. Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih, V., Ward, T., Doron, Y., Firoiu, V., Harley, T., Dunning, I., et al.: Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures. In: International conference on machine learning. pp. 1407-1416. PMLR (2018) +11. Feldman, V.: Does learning require memorization? a short tale about a long tail. In: Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing. pp. 954-959 (2020) +12. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023) +13. Gunasekar, S., Zhang, Y., Aneja, J., Mendes, C.C.T., Del Giorno, A., Gopi, S., Javaheripi, M., Kauffmann, P., de Rosa, G., Saarikivi, O., et al.: Textbooks are all you need. arXiv preprint arXiv:2306.11644 (2023) +14. Har-Peled, S., Kushal, A.: Smaller coresets for k-median and k-means clustering. In: Proceedings of the twenty-first annual symposium on Computational geometry. pp. 126-134 (2005) +15. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) + +16. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. In: Advances in Neural Information Processing Systems (2022) +17. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below. +18. Jouppi, N.P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., Bates, S., Bhatia, S., Boden, N., Borchers, A., et al.: In-datacenter performance analysis of a tensor processing unit. In: Proceedings of the 44th annual international symposium on computer architecture. pp. 1-12 (2017) +19. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020) +20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Springer (2014) +21. Lindley, D.V.: On a measure of the information provided by an experiment. The Annals of Mathematical Statistics 27(4), 986-1005 (1956) +22. Loshchilov, I., Hutter, F.: Online batch selection for faster training of neural networks. arXiv preprint arXiv:1511.06343 (2015) +23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019) +24. MacKay, D.J.: Information-based objective functions for active data selection. Neural computation 4(4), 590-604 (1992) +25. Mahmoud, A., Elhoushi, M., Abbas, A., Yang, Y., Ardalani, N., Leather, H., Morcos, A.: Sieve: Multimodal dataset pruning using image captioning models. arXiv preprint arXiv:2310.02110 (2023) +26. Marion, M., Üstün, A., Pozzobon, L., Wang, A., Fadaee, M., Hooker, S.: When less is more: Investigating data pruning for pretraining llms at scale. arXiv preprint arXiv:2309.04564 (2023) +27. Mindermann, S., Brauner, J.M., Razzak, M.T., Sharma, M., Kirsch, A., Xu, W., Höltgen, B., Gomez, A.N., Morisot, A., Farquhar, S., et al.: Prioritized training on points that are learnable, worth learning, and not yet learnt. In: International Conference on Machine Learning. pp. 15630-15649. PMLR (2022) +28. Paul, M., Ganguli, S., Dziugaite, G.K.: Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems 34, 20596-20607 (2021) +29. Prabhu, A., Dognin, C., Singh, M.: Sampling bias in deep active classification: An empirical study. arXiv preprint arXiv:1909.09389 (2019) +30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021) +31. Schaul, T., Quan, J., Antonoglou, I., Silver, D.: Prioritized experience replay. arXiv preprint arXiv:1511.05952 (2015) +32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) + +33. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +34. Settles, B.: Active learning literature survey (2009) +35. Sorscher, B., Geirhos, R., Shekhar, S., Ganguli, S., Morcos, A.: Beyond neural scaling laws: beating power law scaling via data pruning. Advances in Neural Information Processing Systems 35, 19523-19536 (2022) +36. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the IEEE international conference on computer vision. pp. 843-852 (2017) +37. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023) +38. Toneva, M., Sordoni, A., Combes, R.T.d., Trischler, A., Bengio, Y., Gordon, G.J.: An empirical study of example forgetting during deep neural network learning. arXiv preprint arXiv:1812.05159 (2018) +39. Xie, S.M., Pham, H., Dong, X., Du, N., Liu, H., Lu, Y., Liang, P., Le, Q.V., Ma, T., Yu, A.W.: Doremi: Optimizing data mixtures speeds up language model pretraining (2023) +40. Xie, S.M., Santurkar, S., Ma, T., Liang, P.: Data selection for language models via importance resampling. arXiv preprint arXiv:2302.03169 (2023) +41. Yang, F., Barth-Maron, G., Stanczyk, P., Hoffman, M., Liu, S., Kroiss, M., Pope, A., Rrustemi, A.: Launchpad: A programming model for distributed machine learning research. arXiv preprint arXiv:2106.04516 (2021), https://arxiv.org/abs/2106.04516 +42. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: CoCa: Contrastive captioners are image-text foundation models. In: Transactions on Machine Learning Research (2022) +43. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022) +44. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023) \ No newline at end of file diff --git a/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/images.zip b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9387c224e411da0ecb4293aa4f04416b3afb1253 --- /dev/null +++ b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7af951b9bfea01a645c4773cfa43398fd43f1f58e18880204de956b627b93b3c +size 358667 diff --git a/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/layout.json b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25434ffad8f181d66eaac5338cf0fee516d7d851 --- /dev/null +++ b/2024/Bad Students Make Great Teachers_ Active Learning Accelerates Large-Scale Visual Understanding/layout.json @@ -0,0 +1,8465 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 161, + 449, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 161, + 449, + 211 + ], + "spans": [ + { + "bbox": [ + 148, + 161, + 449, + 211 + ], + "type": "text", + "content": "Bad Students Make Great Teachers: Active Learning Accelerates Large-Scale Visual Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 233, + 414, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 233, + 414, + 256 + ], + "spans": [ + { + "bbox": [ + 179, + 233, + 414, + 256 + ], + "type": "text", + "content": "Talfan Evans1,* Shreya Pathak1,* Hamza Merzic1,2,* Jonathan Schwarz1,† Ryutaro Tanno1 Olivier J. Henaff1,*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 205, + 269, + 391, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 269, + 391, + 281 + ], + "spans": [ + { + "bbox": [ + 205, + 269, + 391, + 281 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 205, + 269, + 391, + 281 + ], + "type": "text", + "content": "Google DeepMind " + }, + { + "bbox": [ + 205, + 269, + 391, + 281 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 205, + 269, + 391, + 281 + ], + "type": "text", + "content": "University College London" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 163, + 302, + 433, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 302, + 433, + 473 + ], + "spans": [ + { + "bbox": [ + 163, + 302, + 433, + 473 + ], + "type": "text", + "content": "Abstract. Power-law scaling indicates that large-scale training with uniform sampling is prohibitively slow. Active learning methods aim to increase data efficiency by prioritizing learning on the most relevant examples. Despite their appeal, these methods have yet to be widely adopted since no one algorithm has been shown to a) generalize across models and tasks b) scale to large datasets and c) yield overall FLOP savings when accounting for the overhead of data selection. In this work we propose a method which satisfies these three properties, leveraging small, cheap proxy models to estimate \"learnability\" scores for datapoints, which are used to prioritize data for training much larger models. As a result, models trained using our methods - ClassAct and ActiveCLIP - require " + }, + { + "bbox": [ + 163, + 302, + 433, + 473 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 163, + 302, + 433, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 163, + 302, + 433, + 473 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 163, + 302, + 433, + 473 + ], + "type": "text", + "content": " fewer training updates and up to " + }, + { + "bbox": [ + 163, + 302, + 433, + 473 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 163, + 302, + 433, + 473 + ], + "type": "text", + "content": " less total computation to reach the same performance as uniformly-trained visual classifiers on JFT and multimodal models on ALIGN, respectively. Finally, we find our data-prioritization scheme to be complementary with recent data-curation and learning objectives, yielding a new state-of-the-art in several multimodal transfer tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 489, + 228, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 489, + 228, + 500 + ], + "spans": [ + { + "bbox": [ + 139, + 489, + 228, + 500 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 510, + 459, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 459, + 630 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 459, + 630 + ], + "type": "text", + "content": "Power-law scaling for vision and language models [19,43] indicates that incremental improvements in model performance require orders of magnitude increases in computation. One of the key features of these empirical power-laws is that training data is sampled uniformly. In contrast, active data selection prioritizes computation on the data that maximally contributes to task performance [21,24,34], with the ultimate goals of improving data efficiency and reducing the cost of training. However, active data selection has yet to become a mainstay of large model training, since no existing algorithm satisfies the three properties of being a) robust to the choice of model and training task, b) scalable to large datasets and architectures, and c) more compute efficient end-to-end than training with uniform samples." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 635, + 459, + 667 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 635, + 459, + 667 + ], + "spans": [ + { + "bbox": [ + 139, + 635, + 459, + 667 + ], + "type": "text", + "content": "* Equal technical contribution. Email correspondence to and . †Current affiliation: Harvard University, work done while at Google DeepMind." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 165, + 298, + 236 + ], + "blocks": [ + { + "bbox": [ + 141, + 165, + 298, + 236 + ], + "lines": [ + { + "bbox": [ + 141, + 165, + 298, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 165, + 298, + 236 + ], + "type": "image", + "image_path": "b276811fb9999805e31e07d2fd1e3f9e55823a6a65f03efcfa09c4a679132726.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 245, + 459, + 287 + ], + "lines": [ + { + "bbox": [ + 138, + 245, + 459, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 459, + 287 + ], + "type": "text", + "content": "Fig. 1: Active learning accelerates large-scale visual understanding. For large-scale classification and multimodal learning tasks, prioritised training on data selected using our active selection methods ClassAct (left) and ActiveCLIP (right) requires significantly fewer updates to reach the final performance of uniform training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 165, + 455, + 235 + ], + "blocks": [ + { + "bbox": [ + 299, + 165, + 455, + 235 + ], + "lines": [ + { + "bbox": [ + 299, + 165, + 455, + 235 + ], + "spans": [ + { + "bbox": [ + 299, + 165, + 455, + 235 + ], + "type": "image", + "image_path": "976b696fe455fa1da0a8f10fbcf0eadb744dc221dd135b88e425ea61e7dd4035.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 292, + 458, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 292, + 458, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 458, + 357 + ], + "type": "text", + "content": "In the first instance, data selection based on hand-engineered filters (e.g. removing incorrectly shaped images or that only contain a single colour [2]) can trivially improve training efficiency at minimal computational overhead. However, such heuristics are limited in their effectiveness by the expertise of the human designer, and are not guaranteed to transfer to training of different models, data modalities, or tasks and incur significant effort to develop and tune." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 357, + 458, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 458, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 458, + 456 + ], + "type": "text", + "content": "In contrast, model-based curation methods, which use the loss of the model itself to score examples, have shown promise by focusing training on 'hard' and omitting 'easy' data [35], but also the opposite (to exclude noise or other low-quality data) in both language modeling [13] and multimodal learning [15]. However, these methods often spend as much computation on the curation of datasets as is gained from subsequent pretraining, making them less compute-efficient than training on uniformly-sampled data. Finally, while several compute-efficient methods have been successfully deployed at small scale [8], they generally do not scale to even medium-sized datasets such as ImageNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 456, + 458, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 458, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 458, + 532 + ], + "type": "text", + "content": "In this work, we propose an algorithm that satisfies the three properties of generality, scalability, and compute-positivity. The proposed framework uses small proxy models to compute learnability scores for candidate training data, resulting in significant training efficiency gains at an almost negligible overhead over standard uniform-training. We test two instantiations of this framework, ClassAct and ActiveCLIP / ActiveSigLIP, for large scale classification and multimodal pretraining, respectively. Our findings are summarized below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 537, + 458, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 537, + 458, + 582 + ], + "spans": [ + { + "bbox": [ + 138, + 537, + 458, + 582 + ], + "type": "text", + "content": "Benchmarking heuristics for large-scale pretraining: We investigate loss- and learnability-based prioritization [15, 27, 35] for large-scale classification and find that pretrained reference models are an essential component for accelerating learning, producing efficiency gains of up to " + }, + { + "bbox": [ + 138, + 537, + 458, + 582 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 138, + 537, + 458, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 586, + 458, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 586, + 458, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 586, + 458, + 642 + ], + "type": "text", + "content": "Generalizing selection policies across model scale: Secondly, we show that smaller models act as effective proxies for much larger (" + }, + { + "bbox": [ + 138, + 586, + 458, + 642 + ], + "type": "inline_equation", + "content": "\\sim 1000 \\times" + }, + { + "bbox": [ + 138, + 586, + 458, + 642 + ], + "type": "text", + "content": ") models in the context of learnability but not loss-based scoring, resulting in compute-positive gains of up to " + }, + { + "bbox": [ + 138, + 586, + 458, + 642 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 138, + 586, + 458, + 642 + ], + "type": "text", + "content": " over uniform training, a first in the context of large-scale pretraining." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 644, + 458, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 458, + 668 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 458, + 668 + ], + "type": "text", + "content": "Accelerating multimodal pretraining: Using reference models pretrained on small, clean datasets, we substantially accelerate pretraining on much larger," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 143, + 145, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 143, + 145, + 151 + ], + "spans": [ + { + "bbox": [ + 139, + 143, + 145, + 151 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "spans": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "text", + "content": "T. Evans et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 165, + 458, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 458, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 458, + 199 + ], + "type": "text", + "content": "noisier datasets. Moreover, we find ActiveCLIP to be complementary to recent data-curation techniques [12] and learning objectives [44], yielding a new state-of-the-art in several multimodal understanding tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 203, + 459, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 459, + 238 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 459, + 238 + ], + "type": "text", + "content": "Amortizing data selection policies: Data-selection policies trained on one task can also accelerate the training of subsequent models on different but related tasks, suggesting that such policies can be easily derived from pre-trained models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 241, + 459, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 241, + 459, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 241, + 459, + 286 + ], + "type": "text", + "content": "Simplifying active-learning with online reference models: Lastly, we demonstrate that pre-trained reference models may not be necessary at all, where these models are small and can be trained in parallel on larger batches than the learner model, while remaining compute-positive." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 303, + 235, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 303, + 235, + 315 + ], + "spans": [ + { + "bbox": [ + 139, + 303, + 235, + 315 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 329, + 459, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 329, + 459, + 461 + ], + "spans": [ + { + "bbox": [ + 138, + 329, + 459, + 461 + ], + "type": "text", + "content": "Data pruning. One approach to data-selection is to identify and sub-select data ahead of training. For example, [28] and [35] show that the training loss and gradients can be used to discard large portions of small-to-medium sized datasets (e.g. CIFAR10 and ImageNet) with little loss in performance. These methods have since been deployed for the curation of web-scale datasets in both language modeling [26] and multimodal learning [1, 12, 25], demonstrating large reductions in the amount of data required together with performance improvements. However, in the single-epoch regime that is becoming typical of large model training [16, 19], pre-filtering can be as expensive as learning from it, a shortcoming which we address in this work. Nevertheless, we show that our method for dynamic data selection is complementary to and benefits from such data curation techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 465, + 459, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 459, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 459, + 587 + ], + "type": "text", + "content": "Online active learning. Unlike data-pruning, online active learning continuously filters data throughout training and applies naturally to the semi-infinite, single-epoch regime. Online Batch Selection [22] scores and filters using the learner model, which has the theoretical advantage that the importance of data can be determined relative to the current state of the learner. In terms of metrics, the Reducible Holdout Loss (RHO) [27] also uses the concept of a reference model to identify learnable data points not yet well represented. Other proposed heuristics include memorization for long-tailed data [11] and assigning \"complexity\" scores based on the number of times the example is forgotten during training [38]. None of these approaches however have demonstrated that the cost of scoring can be reduced to the point of justifying learner efficiency gains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 591, + 459, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 591, + 459, + 669 + ], + "spans": [ + { + "bbox": [ + 138, + 591, + 459, + 669 + ], + "type": "text", + "content": "Compute-efficient data selection. Several works have demonstrated the benefits of selecting data based on simple heuristics, such as low-level image properties [2] or proximity to high-quality text corpora [3, 7, 29, 40]. While cheap to compute, these statistics often require domain-specific knowledge which limits their applicability across tasks. Domain-agnostic methods such as core-sets alleviate this by selecting data based on the geometry of their embeddings [4, 14] which can be efficiently computed, however these algorithms generally do not" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "spans": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "text", + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "spans": [ + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 165, + 459, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 459, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 459, + 209 + ], + "type": "text", + "content": "scale to large-scale datasets [8]. Most related to our work is DoReMi [39] which uses domain-general, scalable, and compute-efficient proxy models for the simpler problem of determining optimal data-mixtures for the subsequent training of a larger language model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 224, + 207, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 207, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 207, + 235 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 245, + 326, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 326, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 326, + 257 + ], + "type": "text", + "content": "3.1 Data selection as prioritized replay" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "text", + "content": "We use online batch selection [22] to apply our scoring heuristics to standard visual learning tasks: firstly, we sample uniformly from the training set " + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i \\stackrel{\\mathcal{U}}{\\sim} \\mathcal{D}" + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "text", + "content": " and compute a score " + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "inline_equation", + "content": "s_i = s(\\boldsymbol{x}_i | \\theta) \\in \\mathbb{R}" + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "text", + "content": " to each data point " + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i" + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "text", + "content": " using model parameters " + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "text", + "content": ". Given a large enough collection of scored examples stored in a memory bank " + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\boldsymbol{x}_i\\}_{i \\in (0, \\dots, M-1)}" + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "text", + "content": ", we then sample non-uniformly according to their scores " + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i \\stackrel{\\pi}{\\sim} \\mathcal{M}" + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "text", + "content": " [31], where " + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "inline_equation", + "content": "\\pi(\\boldsymbol{x}_i) = \\text{Softmax}(\\{s_i\\}_{i \\in (0, \\dots, M-1)})" + }, + { + "bbox": [ + 138, + 262, + 459, + 366 + ], + "type": "text", + "content": ". A batch of such examples is used to update the learner model. Following convention in reinforcement learning, we refer to the scoring and target models as actors and learners respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 380, + 289, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 289, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 289, + 390 + ], + "type": "text", + "content": "3.2 Statistics for data selection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 397, + 458, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 458, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 458, + 418 + ], + "type": "text", + "content": "We explore a few statistics for model-based prioritization, grouped into two categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 423, + 458, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 458, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 458, + 467 + ], + "type": "text", + "content": "Example difficulty: given the current state of the learner, an intuitive prioritization scheme might favour 'difficult' examples (as measured by their training loss), while removing 'easy' examples that are trivially classified and which yield small gradients. This loss-based prioritization:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 257, + 474, + 457, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 474, + 457, + 485 + ], + "spans": [ + { + "bbox": [ + 257, + 474, + 457, + 485 + ], + "type": "interline_equation", + "content": "s ^ {\\text {h a r d}} \\left(\\boldsymbol {x} _ {i} \\mid \\theta\\right) = \\ell \\left(\\boldsymbol {x} _ {i} \\mid \\theta\\right) \\tag {1}", + "image_path": "28bdaec823d848e34736bda208d67d421b7b75b28d6513c15806001a1e32ad49.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 492, + 458, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 458, + 537 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 458, + 537 + ], + "type": "text", + "content": "can use the current parameters of the learner " + }, + { + "bbox": [ + 138, + 492, + 458, + 537 + ], + "type": "inline_equation", + "content": "\\theta^t" + }, + { + "bbox": [ + 138, + 492, + 458, + 537 + ], + "type": "text", + "content": " or those of a fixed model " + }, + { + "bbox": [ + 138, + 492, + 458, + 537 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 138, + 492, + 458, + 537 + ], + "type": "text", + "content": ". The opposite argument can be made for favoring examples that are easily solved by a well-trained model, as such a prioritization removes the noisy examples present in large-scale datasets:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 254, + 543, + 457, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 543, + 457, + 554 + ], + "spans": [ + { + "bbox": [ + 254, + 543, + 457, + 554 + ], + "type": "interline_equation", + "content": "s ^ {\\text {e a s y}} \\left(\\boldsymbol {x} _ {i} | \\theta\\right) = - \\ell \\left(\\boldsymbol {x} _ {i} | \\theta\\right) \\tag {2}", + "image_path": "0f2540ed710e6bf87508bdfb61a13c7cdbdf6df5f922876267b73e9a0667d488.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 560, + 458, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 560, + 458, + 584 + ], + "spans": [ + { + "bbox": [ + 138, + 560, + 458, + 584 + ], + "type": "text", + "content": "This scheme is commonly used in multimodal learning for identifying high-quality examples with pre-trained models [15,32,33]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 587, + 458, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 458, + 630 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 458, + 630 + ], + "type": "text", + "content": "Example learnability: Given that favoring easy and hard examples target different and potentially orthogonal properties of the data, a natural question is whether these policies can be combined. Learnability criteria straightforwardly combine the two as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 211, + 638, + 457, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 638, + 457, + 664 + ], + "spans": [ + { + "bbox": [ + 211, + 638, + 457, + 664 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s ^ {\\text {l e a r n}} \\left(\\boldsymbol {x} _ {i} | \\theta^ {t}, \\theta^ {*}\\right) = s ^ {\\text {h a r d}} \\left(\\boldsymbol {x} _ {i} | \\theta^ {t}\\right) + s ^ {\\text {e a s y}} \\left(\\boldsymbol {x} _ {i} | \\theta^ {*}\\right) (3) \\\\ = \\ell \\left(\\boldsymbol {x} _ {i} \\mid \\theta^ {t}\\right) - \\ell \\left(\\boldsymbol {x} _ {i} \\mid \\theta^ {*}\\right), (4) \\\\ \\end{array}", + "image_path": "a429981e52dae8bb1853c30d8e38fe32af2256d72d1207db6a20f005fef3be9c.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 143, + 146, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 143, + 146, + 151 + ], + "spans": [ + { + "bbox": [ + 139, + 143, + 146, + 151 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "spans": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "text", + "content": "T. Evans et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 165, + 459, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 459, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 459, + 219 + ], + "type": "text", + "content": "favoring examples that are easily solved by a well-trained model " + }, + { + "bbox": [ + 138, + 165, + 459, + 219 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 138, + 165, + 459, + 219 + ], + "type": "text", + "content": " but challenging to the learner in its current state " + }, + { + "bbox": [ + 138, + 165, + 459, + 219 + ], + "type": "inline_equation", + "content": "\\theta^t" + }, + { + "bbox": [ + 138, + 165, + 459, + 219 + ], + "type": "text", + "content": ", such that more computation dedicated to this example could lower its loss. Conversely, examples that are trivially classified by the learner (or mislabeled) will yield low (or high) losses for both the current learner and the well-trained one, leading to low learnability scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 219, + 459, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 219, + 459, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 219, + 459, + 264 + ], + "type": "text", + "content": "A special case of learnability scores (the RHO loss, [27]) uses a model " + }, + { + "bbox": [ + 138, + 219, + 459, + 264 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{ho}}" + }, + { + "bbox": [ + 138, + 219, + 459, + 264 + ], + "type": "text", + "content": " specifically trained on a held-out dataset to ensure the independence of its predictions from those of the current learner " + }, + { + "bbox": [ + 138, + 219, + 459, + 264 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{learn}}(\\pmb{x}_i|\\theta^t,\\theta^{\\mathrm{ho}})" + }, + { + "bbox": [ + 138, + 219, + 459, + 264 + ], + "type": "text", + "content": ". We assess in Section 4.2 whether this is necessary when training on large-scale image datasets." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 142, + 286, + 456, + 347 + ], + "blocks": [ + { + "bbox": [ + 142, + 286, + 456, + 347 + ], + "lines": [ + { + "bbox": [ + 142, + 286, + 456, + 347 + ], + "spans": [ + { + "bbox": [ + 142, + 286, + 456, + 347 + ], + "type": "image", + "image_path": "6dc4c9e120688085749e7bef4da43f4d298326e970f431a67df88065169b146e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 357, + 460, + 497 + ], + "lines": [ + { + "bbox": [ + 138, + 357, + 460, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 460, + 497 + ], + "type": "text", + "content": "Fig. 2: Amortizing the cost of data selection. Drawn to scale: length of bars indicates number of FLOPs required to reach the accuracy of a ViT-L trained with uniform sampling (\"ViT-L Uniform Sampling\", see Figure 4). Expensive model policies (e.g. a ViT-B scores data for the ViT-L learner, or " + }, + { + "bbox": [ + 138, + 357, + 460, + 497 + ], + "type": "inline_equation", + "content": "\\mathrm{B} \\rightarrow \\mathrm{L}" + }, + { + "bbox": [ + 138, + 357, + 460, + 497 + ], + "type": "text", + "content": ") produce large learner speedups, at the expense of the additional FLOPs associate with data selection. This overhead can be reduced by deriving the data-selection policies from smaller models (e.g. ViT-S, ViT-Ti or ViT-Mu score data for the ViT-L learner), at the expense of marginal decreases in the learner speedup. Costs could be additionally amortized by using off-the-shelf reference models, removing the need to train from scratch (yellow). Since the reference model is fixed throughout training, scores can be assigned once to a 'foundation dataset' and amortized across many training runs (lime green; [27,35]). Since the online model is independent of the learner model and generalizes across scale, data selection policies can also be distilled as a fixed ordering of a given dataset (a 'foundation curriculum')." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 525, + 333, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 333, + 538 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 333, + 538 + ], + "type": "text", + "content": "3.3 Unlocking compute-positive training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 543, + 459, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 459, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 459, + 597 + ], + "type": "text", + "content": "Scoring requires inference passes over the actor and learner models. We assume that gradient updates to cost " + }, + { + "bbox": [ + 138, + 543, + 459, + 597 + ], + "type": "inline_equation", + "content": "3\\mathrm{x}" + }, + { + "bbox": [ + 138, + 543, + 459, + 597 + ], + "type": "text", + "content": " inference passes. The cost of scoring data " + }, + { + "bbox": [ + 138, + 543, + 459, + 597 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 138, + 543, + 459, + 597 + ], + "type": "text", + "content": " thus scales with the proportion of data which is being rejected (e.g. retaining only " + }, + { + "bbox": [ + 138, + 543, + 459, + 597 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 138, + 543, + 459, + 597 + ], + "type": "text", + "content": " of the data requires 5 inference passes per trained batch). The requirements for compute-positivity can therefore be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 605, + 458, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 605, + 458, + 639 + ], + "spans": [ + { + "bbox": [ + 208, + 605, + 458, + 639 + ], + "type": "interline_equation", + "content": "\\underbrace {\\left(3 F _ {\\text {l e a r n}} + \\rho F _ {\\text {a c t}}\\right) \\beta + 3 F _ {\\text {r e f}}} _ {\\text {A c t i v e L e a r n i n g}} < \\underbrace {3 F _ {\\text {l e a r n}}} _ {\\text {U n i f o r m S a m p l i n g}} \\tag {5}", + "image_path": "4e2df8002b9c84ac6ba45a0947370947afb4f06305af82dd87136f43ec195f18.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{act}}" + }, + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "text", + "content": " is the cost of scoring an example, " + }, + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "text", + "content": " is the number of examples scored per training example, and " + }, + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "text", + "content": " is the efficiency gain (in terms of the learner update)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "spans": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "text", + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "spans": [ + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 165, + 457, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 457, + 230 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 457, + 230 + ], + "type": "text", + "content": "relative to training with uniform sampling (see Appendix Section A.1 for more details). The RHS term is the cost per update of uniform sampling. The first LHS term inside the brackets is the cost of the learner training during AL, the second term is the scoring cost and the last term outside the brackets the cost of training the reference model. We illustrate in Figure 2 the different contexts in which parts of this computation may be effectively amortized." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "text", + "content": "In the large-scale training regime where data is neither repeated nor seen before, compute-positivity requires that either or all of the reference model training, actor scoring and learner efficiency " + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "text", + "content": " terms must be made smaller to produce net savings relative to uniform sampling. Typical prioritization schemes can produce saving on the order of " + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "text", + "content": " (ie " + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "text", + "content": "), suggesting that savings must also be made by down-scaling the other terms " + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{act}}" + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ref}}" + }, + { + "bbox": [ + 138, + 232, + 458, + 298 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 303, + 458, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 458, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 458, + 346 + ], + "type": "text", + "content": "Cost of easy-reference scoring. While both the cost of the reference model and example scoring can be scaled down in the case of easy-reference scoring with small models (" + }, + { + "bbox": [ + 138, + 303, + 458, + 346 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{act}} = F_{\\mathrm{ref}}" + }, + { + "bbox": [ + 138, + 303, + 458, + 346 + ], + "type": "text", + "content": ", see Equation 2), it is unclear whether the efficiency gains " + }, + { + "bbox": [ + 138, + 303, + 458, + 346 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 138, + 303, + 458, + 346 + ], + "type": "text", + "content": " are robust to this down-scaling (see section 4.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 351, + 458, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 458, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 458, + 408 + ], + "type": "text", + "content": "Cost of RHO learnability scoring. The original definition of learnability scores [27] requires inference passes through both the learner and a reference model (" + }, + { + "bbox": [ + 138, + 351, + 458, + 408 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{act}} = F_{\\mathrm{ref}} + F_{\\mathrm{learn}}" + }, + { + "bbox": [ + 138, + 351, + 458, + 408 + ], + "type": "text", + "content": ", see Equation 3), meaning that although the cost of the reference model can be reduced by using a smaller model, the cost of example scoring cannot." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 412, + 458, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 458, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 458, + 479 + ], + "type": "text", + "content": "Cost of ClassAct / ActiveCLIP. For this reason, we explore whether replacing the learner model in term 1 of Equation 3 with a much smaller model can still produce comparable learner efficiency gains to those already observed (see Appendix Algorithm 1). Specifically, we introduce a third \"online\" model, which has the same architecture and size as the reference model, but is trained in parallel with the learner. In this case, the cost of scoring examples reduces to:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 493, + 457, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 493, + 457, + 505 + ], + "spans": [ + { + "bbox": [ + 242, + 493, + 457, + 505 + ], + "type": "interline_equation", + "content": "F _ {\\text {a c t}} = F _ {\\text {r e f}} + F _ {\\text {o n l i n e}} = 2 F _ {\\text {r e f}} \\tag {6}", + "image_path": "3d6d3f2899577a160f226db6b30376558e0d403680a7fcb1697c6bebda357844.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 512, + 458, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 458, + 591 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 458, + 591 + ], + "type": "text", + "content": "and can be scaled down along with the reference model. We instantiate our method for two canonical pre-training tasks: visual classification and multimodal learning, which we call ClassAct, ActiveCLIP and ActiveSigLIP respectively. In Appendix Section A.6, we describe an asynchronous active learning framework where scoring and learning is performed in parallel on separate devices. This setup does not affect the FLOP calculations but can mitigate the overhead in time even when using non-approximate scoring." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 611, + 371, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 371, + 623 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 371, + 623 + ], + "type": "text", + "content": "3.4 Losses for canonical visual pre-training tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 634, + 458, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 634, + 458, + 670 + ], + "spans": [ + { + "bbox": [ + 138, + 634, + 458, + 670 + ], + "type": "text", + "content": "For visual classification with ClassACT, we use the standard cross-entropy loss for both actors and learners. For multimodal learning with ActiveCLIP, learners optimize the contrastive loss " + }, + { + "bbox": [ + 138, + 634, + 458, + 670 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{learn}} = \\ell_{\\mathrm{learn}}^{\\mathrm{im,txt}} + \\ell_{\\mathrm{learn}}^{\\mathrm{txt,im}}" + }, + { + "bbox": [ + 138, + 634, + 458, + 670 + ], + "type": "text", + "content": ", whereas the actor loss " + }, + { + "bbox": [ + 138, + 634, + 458, + 670 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{act}}" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 143, + 145, + 150 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 143, + 145, + 150 + ], + "spans": [ + { + "bbox": [ + 139, + 143, + 145, + 150 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "spans": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "text", + "content": "T. Evans et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 139, + 165, + 293, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 165, + 293, + 176 + ], + "spans": [ + { + "bbox": [ + 139, + 165, + 293, + 176 + ], + "type": "text", + "content": "Algorithm 1 ClassAct / ActiveCLIP" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 179, + 459, + 289 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "spans": [ + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "content": "1: Input: Randomly initialized learner model " + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "content": " and small online model " + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "inline_equation", + "content": "\\theta_{o}" + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "content": ", small pre-trained reference model " + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "inline_equation", + "content": "\\theta_{r}" + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "content": ". Models use loss " + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{act}}" + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "content": " for scoring data and " + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{learn}}" + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "content": " for computing updates. Dataset " + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "content": ", batch size " + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "content": ", sub-batch size " + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "inline_equation", + "content": "b < B" + }, + { + "bbox": [ + 141, + 179, + 459, + 210 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 210, + 224, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 210, + 224, + 219 + ], + "spans": [ + { + "bbox": [ + 142, + 210, + 224, + 219 + ], + "type": "text", + "content": "2: while training do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 219, + 458, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 219, + 458, + 230 + ], + "spans": [ + { + "bbox": [ + 142, + 219, + 458, + 230 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 142, + 219, + 458, + 230 + ], + "type": "inline_equation", + "content": "X\\sim \\mathcal{D}" + }, + { + "bbox": [ + 142, + 219, + 458, + 230 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 142, + 219, + 458, + 230 + ], + "type": "inline_equation", + "content": "|X| = B" + }, + { + "bbox": [ + 142, + 219, + 458, + 230 + ], + "type": "text", + "content": " Sample uniformly" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 230, + 457, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 230, + 457, + 239 + ], + "spans": [ + { + "bbox": [ + 142, + 230, + 457, + 239 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 142, + 230, + 457, + 239 + ], + "type": "inline_equation", + "content": "S = \\ell_{\\mathrm{act}}(X|\\theta_o) - \\ell_{\\mathrm{act}}(X|\\theta_r)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 239, + 457, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 239, + 457, + 250 + ], + "spans": [ + { + "bbox": [ + 142, + 239, + 457, + 250 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 142, + 239, + 457, + 250 + ], + "type": "inline_equation", + "content": "I\\sim \\mathrm{SoftMax}(S)" + }, + { + "bbox": [ + 142, + 239, + 457, + 250 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 142, + 239, + 457, + 250 + ], + "type": "inline_equation", + "content": "|I| = b" + }, + { + "bbox": [ + 142, + 239, + 457, + 250 + ], + "type": "text", + "content": " Sample indices" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 250, + 457, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 250, + 457, + 259 + ], + "spans": [ + { + "bbox": [ + 142, + 250, + 457, + 259 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 142, + 250, + 457, + 259 + ], + "type": "inline_equation", + "content": "Y = X[I]" + }, + { + "bbox": [ + 142, + 250, + 457, + 259 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 142, + 250, + 457, + 259 + ], + "type": "text", + "content": " Collect sub-batch" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 259, + 457, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 259, + 457, + 270 + ], + "spans": [ + { + "bbox": [ + 142, + 259, + 457, + 270 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 142, + 259, + 457, + 270 + ], + "type": "inline_equation", + "content": "\\theta_{l}\\gets \\mathrm{Adam}[\\nabla_{\\theta_{l}}\\ell_{\\mathrm{learn}}(Y|\\theta_{l})]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 270, + 457, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 270, + 457, + 280 + ], + "spans": [ + { + "bbox": [ + 142, + 270, + 457, + 280 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 142, + 270, + 457, + 280 + ], + "type": "inline_equation", + "content": "\\theta_o\\gets \\mathrm{Adam}[\\nabla_{\\theta_o}\\ell_{\\mathrm{learn}}(Y|\\theta_o)]" + }, + { + "bbox": [ + 142, + 270, + 457, + 280 + ], + "type": "text", + "content": " Update online model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 280, + 196, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 280, + 196, + 289 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 196, + 289 + ], + "type": "text", + "content": "9: end while" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 138, + 312, + 433, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 433, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 433, + 324 + ], + "type": "text", + "content": "is simply the dot-product similarity between image and text embeddings:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 230, + 329, + 457, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 329, + 457, + 342 + ], + "spans": [ + { + "bbox": [ + 230, + 329, + 457, + 342 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {a c t}} \\left(\\boldsymbol {x} _ {i} \\mid \\theta\\right) = - \\boldsymbol {z} _ {i} ^ {\\mathrm {i m}} \\cdot \\boldsymbol {z} _ {i} ^ {\\mathrm {t x t}} \\tag {7}", + "image_path": "c1a72474d3f166aff9f8c08063184b3edfbba18283b9751e87352bfafef7ee93.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 220, + 344, + 457, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 344, + 457, + 371 + ], + "spans": [ + { + "bbox": [ + 220, + 344, + 457, + 371 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {l e a r n}} ^ {\\mathrm {i m}, \\mathrm {t x t}} (\\boldsymbol {x} _ {i} | \\theta) = - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {i m}} \\cdot \\boldsymbol {z} _ {i} ^ {\\mathrm {t x t}}\\right)}{\\sum_ {j} \\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {i m}} \\cdot \\boldsymbol {z} _ {j} ^ {\\mathrm {t x t}}\\right)} \\tag {8}", + "image_path": "e7dc24c97e892e70eae497fd420e42ab7b6f46e0625bb22aae72fec0fbaefa74.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_i^{\\mathrm{im}} = f^{\\mathrm{im}}(\\boldsymbol{x}_i; \\theta)" + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_i^{\\mathrm{txt}} = f^{\\mathrm{txt}}(\\boldsymbol{x}_i; \\theta)" + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "text", + "content": " are image and text embeddings respectively, and " + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{learn}}^{\\mathrm{txt}, \\mathrm{im}}(\\boldsymbol{x}_i; \\theta)" + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "text", + "content": " is defined analogously to " + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{learn}}^{\\mathrm{im}, \\mathrm{txt}}(\\boldsymbol{x}_i; \\theta)" + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "text", + "content": ". Similarly, ActiveSigLIP instead uses the sigmoid loss [44] for the learner's objective and " + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{act}}" + }, + { + "bbox": [ + 138, + 377, + 458, + 423 + ], + "type": "text", + "content": " for scoring." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 438, + 227, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 227, + 451 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 227, + 451 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 459, + 458, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 458, + 558 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 458, + 558 + ], + "type": "text", + "content": "All our experiments were conducted with Vision Transformers [9] for which strong baselines are available across model sizes [43]. Unless specified, we adopt models with patch-size 16 throughout (ViT-S refers to ViT-S/16 and similar). We consider two canonical tasks for large-scale pretraining: classification on JFT-300M [36] and multimodal contrastive learning [30] on large image-text datasets. When pre-training with JFT classification we use held-out classification performance as the evaluation metric. When pre-training on image-text data we evaluate with standard multimodal transfer tasks: ImageNet zero-shot classification and image-to-text / text-to-image retrieval on COCO." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 558, + 458, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 458, + 614 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 458, + 614 + ], + "type": "text", + "content": "Throughout, we will refer to the large batch of size " + }, + { + "bbox": [ + 138, + 558, + 458, + 614 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 138, + 558, + 458, + 614 + ], + "type": "text", + "content": " sampled uniformly from the training data as the 'super-batch', and the prioritised smaller batch of size " + }, + { + "bbox": [ + 138, + 558, + 458, + 614 + ], + "type": "inline_equation", + "content": "b < B" + }, + { + "bbox": [ + 138, + 558, + 458, + 614 + ], + "type": "text", + "content": " as the 'sub-batch'. In all our experiment, we filter 50% of uniformly sampled data such that " + }, + { + "bbox": [ + 138, + 558, + 458, + 614 + ], + "type": "inline_equation", + "content": "\\rho = B / b = 2" + }, + { + "bbox": [ + 138, + 558, + 458, + 614 + ], + "type": "text", + "content": ", although more aggressive filtering regimes warrant investigation [35]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 628, + 458, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 628, + 458, + 639 + ], + "spans": [ + { + "bbox": [ + 138, + 628, + 458, + 639 + ], + "type": "text", + "content": "4.1 Evaluating loss-based scoring heuristics in the large-data regime" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 458, + 667 + ], + "type": "text", + "content": "We begin by evaluating loss- and learnability-based heuristics on their ability to accelerate supervised classification on JFT (Fig. 3). Arguably the most intuitive" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "spans": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "text", + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "spans": [ + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 170, + 422, + 280 + ], + "blocks": [ + { + "bbox": [ + 173, + 170, + 422, + 280 + ], + "lines": [ + { + "bbox": [ + 173, + 170, + 422, + 280 + ], + "spans": [ + { + "bbox": [ + 173, + 170, + 422, + 280 + ], + "type": "image", + "image_path": "dd270feee29f95d773fd59a75a31aca907bb97690a0fb075dfbd83db1bb0884e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 290, + 459, + 351 + ], + "lines": [ + { + "bbox": [ + 138, + 290, + 459, + 351 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 459, + 351 + ], + "type": "text", + "content": "Fig. 3: Evaluation of loss-based data-selection criteria for large-scale classification. We train a ViT-B on JFT-300M with different data-selection policies. Prioritising hard data under the learner (green curve) produced marginal gains over the uniform sampling baseline. Prioritizing data using both learnability (blue curve, [27]) and easy reference prioritization (red curve, [15]) produced significant speedups and performance gains." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 360, + 298, + 496 + ], + "blocks": [ + { + "bbox": [ + 141, + 360, + 298, + 496 + ], + "lines": [ + { + "bbox": [ + 141, + 360, + 298, + 496 + ], + "spans": [ + { + "bbox": [ + 141, + 360, + 298, + 496 + ], + "type": "image", + "image_path": "3fda2293d0d429e1180327358be41fc151e8dabef54c06ba30731ca0947e9ce1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 506, + 459, + 658 + ], + "lines": [ + { + "bbox": [ + 138, + 506, + 459, + 658 + ], + "spans": [ + { + "bbox": [ + 138, + 506, + 459, + 658 + ], + "type": "text", + "content": "Fig. 4: Generalization of data-selection policies across models scales. Left: We train a ViT-L for 3 epochs on JFT using uniform sampling (grey) or prioritized data sampling using example learnability (blue) or low-loss under the reference model (red). Example scores are computed using ViT-B actors (dark), or cheaper ViT-S or ViT-Tiny models (light). While both example learnability and \"easy reference\" yield good speedups with expensive actors, learnability criteria are much more robust to approximate scoring. Top right: Learner (ViT-L) speedup is computed as the fraction of learner iterations saved in order to attain the baseline's top performance. Actor overhead is computed as the additional computation in FLOPs required to score examples with a particular actor architecture (varying from ViT-Mu to ViT-L, see Appendix Table 4). Example learnability yields robust learner speedups across actor scales, \"easy reference\" scoring does not. Lower right: total compute efficiency is calculated as a product of learner efficiency and actor overhead, indicating the amount of computation required to reach baseline performance. Approximate actors (i.e. ViT-S or smaller) computing example learnability enable total compute speedups, other schemes do not." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 303, + 361, + 456, + 496 + ], + "blocks": [ + { + "bbox": [ + 303, + 361, + 456, + 496 + ], + "lines": [ + { + "bbox": [ + 303, + 361, + 456, + 496 + ], + "spans": [ + { + "bbox": [ + 303, + 361, + 456, + 496 + ], + "type": "image", + "image_path": "434dd7f91d42456df41f0e7afdf1d4742ab31b1a6f6cef39ed93055f372a7bf1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 143, + 145, + 150 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 143, + 145, + 150 + ], + "spans": [ + { + "bbox": [ + 139, + 143, + 145, + 150 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "spans": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "text", + "content": "T. Evans et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 165, + 459, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 459, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 459, + 241 + ], + "type": "text", + "content": "method to score data is to prioritise training on data with high loss under the learner (hard learner). In our experiments, this strategy (Eq. 1) only marginally improved performance over the uniform sampling baseline, despite requiring an additional inference pass over the super-batch. This is perhaps not surprising - data points with high loss may also be unlearnable due to e.g. label noise, such that training on those data points does not result in the model performing any better on the held-out test set. Large scale datasets are more likely to be noisy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 241, + 459, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 241, + 459, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 241, + 459, + 298 + ], + "type": "text", + "content": "Scoring methods based on pre-trained reference models performed much better—both easy reference (equation 2) and learnability (equation 3) -based prioritization produced significant gains over uniform sampling. Here, we pretrained an identical ViT-B for the same 3 epochs to use as a reference model for a second training run displayed above. producing speed-ups of " + }, + { + "bbox": [ + 138, + 241, + 459, + 298 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 138, + 241, + 459, + 298 + ], + "type": "text", + "content": " (Fig. 3)." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 141, + 317, + 298, + 411 + ], + "blocks": [ + { + "bbox": [ + 141, + 317, + 298, + 411 + ], + "lines": [ + { + "bbox": [ + 141, + 317, + 298, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 317, + 298, + 411 + ], + "type": "image", + "image_path": "af975683167a5ea5d947d933a3f525fa4937cf3e5330025768dc300311e5ed53.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 421, + 459, + 501 + ], + "lines": [ + { + "bbox": [ + 138, + 421, + 459, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 459, + 501 + ], + "type": "text", + "content": "Fig. 5: Scaling laws for active learning. We trained a baseline ViT-L over a range of compute budgets (for which ViT-L is compute optimal, see Zhai et al., 2021). We also trained the same ViT-L with both ViT-Ti and ViT-S reference policies, pre-trained for the same number of epochs. Left: Small model policies produce robust savings in learner compute. Right: When accounting for total compute (learner + actor training and data scoring), small model policies in all compute budgets produce FLOP savings over training with uniform samples. These scaling laws generalize those measured empirically in the uniform sampling setting [43] to the case of non-uniform data selection." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 317, + 456, + 411 + ], + "blocks": [ + { + "bbox": [ + 306, + 317, + 456, + 411 + ], + "lines": [ + { + "bbox": [ + 306, + 317, + 456, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 456, + 411 + ], + "type": "image", + "image_path": "4918a10c2f728b7e4d882f333f6af743f3035ac6b058739f4c3c669bca288946.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 538, + 383, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 538, + 383, + 550 + ], + "spans": [ + { + "bbox": [ + 138, + 538, + 383, + 550 + ], + "type": "text", + "content": "4.2 Generalising data-selection policies across scale" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 557, + 458, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 458, + 623 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 458, + 623 + ], + "type": "text", + "content": "The speed-ups afforded by learnability prioritization (Fig. 3) come at the cost of the additional inference passes required to score the data during learner training, plus the cost of training the reference model. This makes the overall gains strongly compute-negative relative to training with uniform samples. Even if the size of the reference model is scaled down [27], these methods still incur the cost of additional learner inference passes to score the data during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 623, + 458, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 623, + 458, + 667 + ], + "spans": [ + { + "bbox": [ + 138, + 623, + 458, + 667 + ], + "type": "text", + "content": "Unlocking compute-positive active learning. To address this issue, we introduce a set of down-scaled models with the same ViT architecture that we use to score data for training a larger ViT-L model (see Methods Section 3.3. We use ViT-B, ViT-S or ViT-Ti variants (which are " + }, + { + "bbox": [ + 138, + 623, + 458, + 667 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 138, + 623, + 458, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 138, + 623, + 458, + 667 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 138, + 623, + 458, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 623, + 458, + 667 + ], + "type": "inline_equation", + "content": "47 \\times" + }, + { + "bbox": [ + 138, + 623, + 458, + 667 + ], + "type": "text", + "content": " cheaper than" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "spans": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "text", + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "spans": [ + { + "bbox": [ + 451, + 143, + 458, + 151 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 139, + 163, + 472, + 249 + ], + "blocks": [ + { + "bbox": [ + 139, + 163, + 472, + 249 + ], + "lines": [ + { + "bbox": [ + 139, + 163, + 472, + 249 + ], + "spans": [ + { + "bbox": [ + 139, + 163, + 472, + 249 + ], + "type": "table", + "html": "
MethodViT model capacityReference TypeSpeed-up
Reference Online LearnerLearner speedupCompute speedup
Uniform Sampling (ViT-B)B0%0%
RHOTinyBBHeld-out, fixed0%-79%
ClassAct-HOTinyTinyBHeld-out, fixed18%3%
ClassActTinyTinyBIn-domain, fixed18%3%
ClassAct-OnlineTinyTinyBTrained online17%2%
", + "image_path": "33b8f873928de77211adbd3807ef21c545899669f25a39dd3056894af65855cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 250, + 458, + 259 + ], + "lines": [ + { + "bbox": [ + 138, + 250, + 458, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 458, + 259 + ], + "type": "text", + "content": "Table 1: Simplifying and accelerating the computation of learnability scores." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 138, + 259, + 458, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 458, + 309 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 458, + 309 + ], + "type": "text", + "content": "Relative to RHO [27], ClassAct makes two changes: replacing the reference model with one trained in-domain (removing the need for bespoke held-out sets), and dramatically reducing capacity of the online actor models used for scoring examples. All experiments were conducted on 3 epochs of one-half of JFT to enable the held-out ablations. RHO with a small reference model did not produce a learner speedup in our experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 334, + 458, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 458, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 458, + 400 + ], + "type": "text", + "content": "the learner) for both the online and reference model (see Appendix Algorithm 1). In Figure 4 (left) we assess the impact of these cheaper scoring models on learner efficiency. First, we find that easy reference prioritization to be very sensitive to the capacity of the scoring model: while ViT-B scoring models yield reasonable gains over uniform sampling, prioritizing with ViT-S and ViT-Ti scoring models underperforms significantly (Fig. 4, red curves)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "text", + "content": "In contrast, we find that learnability based prioritization yields robust gains, even when the scoring models are significantly scaled down (ClassAct; Fig. 4, blue curves). For example, while ViT-B scoring models yield a " + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "text", + "content": " learner speedup, the " + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "text", + "content": " smaller ViT-Ti scoring models still provide a " + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "text", + "content": " speedup. We pushed this logic by using even smaller scoring models (the ViT-Mu family which we introduce, see Appendix) which are up to " + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "inline_equation", + "content": "1000 \\times" + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "text", + "content": " smaller than the learner. Despite this, prioritizing data based on their scores yields non-negligible speedups (e.g. " + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 138, + 402, + 458, + 490 + ], + "type": "text", + "content": " for the smallest actors we consider; Figure 4, top right)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 491, + 458, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 458, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 458, + 567 + ], + "type": "text", + "content": "These experiments demonstrate that, with the appropriate scoring criterion, online and reference models can be significantly downscaled and still produce comparable gains to larger models, with learner efficiency degrading gracefully with the actor overhead (i.e. the cost of the reference model and data scoring). As a result, our method ClassAct quickly becomes FLOP positive as the online + reference models are downscaled (Figure 4, bottom right), while at the same time producing speed-ups in wall-clock time for a given learner batch size." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 569, + 458, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 569, + 458, + 667 + ], + "spans": [ + { + "bbox": [ + 138, + 569, + 458, + 667 + ], + "type": "text", + "content": "Together, our results expose a pareto front across which to determine an optimal context-specific data selection strategy (Figure 2). Where pre-trained models are available, some of the cost of larger data selection policies can be discounted. If savings in wall-clock time supersede the associated cost of scoring, large models can be tolerated for data selection. Reference model costs can also be amortized across many training runs by appending scores to 'foundation datasets'. However, in the case where no component of the framework can amortized (as in the case of large-scale pretraining), prioritizing data with small ClassAct models can deliver large savings in total computation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "spans": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "spans": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "text", + "content": "T. Evans et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 165, + 449, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 449, + 177 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 449, + 177 + ], + "type": "text", + "content": "4.3 Generalising neural scaling laws to the active-learning setting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 185, + 459, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 185, + 459, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 185, + 459, + 295 + ], + "type": "text", + "content": "We next investigated the scaling behaviour of ClassAct by experimenting over large learner compute budgets (Fig. 5), using both ViT-Ti and ViT-S models as actors for training a ViT-L. Predictably, the ViT-S produced larger, although marginal gains over the ViT-Ti actors when not accounting for scoring FLOPs (Figure 5, left). However, when accounting for total FLOPs, the difference was less pronounced (Figure 5, right). Our results generalize large scale uniform sampling scaling laws such as uncovered by [19] for LLMs and reproduced for large vision transformers by [43] to the case of non-uniform sampling. For the first time, we demonstrate that these scaling laws can be shifted in our favour by selecting data using general model-based scoring heuristics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 313, + 348, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 348, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 348, + 324 + ], + "type": "text", + "content": "4.4 Training the reference model in parallel" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 333, + 458, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 333, + 458, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 333, + 458, + 409 + ], + "type": "text", + "content": "The reference model needs to be trained if none is already available. This two-step process adds complexity to active model training, especially if using large scale infrastructure. However, an interesting consequence of down-scaling the reference model is that both inference passes and gradients can be computed over a much larger batch than can be computed on the learner. In theory, this would mean that the small reference model could instead be trained online, in parallel with the large learner and small online model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 410, + 458, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 458, + 530 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 458, + 530 + ], + "type": "text", + "content": "We confirmed our hypothesis by running an experiment in which we trained our reference model on a super-batch of size " + }, + { + "bbox": [ + 138, + 410, + 458, + 530 + ], + "type": "inline_equation", + "content": "B = 10b" + }, + { + "bbox": [ + 138, + 410, + 458, + 530 + ], + "type": "text", + "content": " and trained the online and learner model in sequence with the sub-batch of size " + }, + { + "bbox": [ + 138, + 410, + 458, + 530 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 138, + 410, + 458, + 530 + ], + "type": "text", + "content": ". To make sure the reference model quickly converges, we additionally set the learning rate to double that of the online and learner models (this would cause instability for the learner and online models because of the additional variance from the smaller batch). We also verified that training the reference model on a held-out set of data performed equally in our experiments to reference models trained on the same data as the learner model [27]. Our 'one-pass' setup, Online-ClassAct, produces the same performance as the pre-trained ClassAct pipeline in our Ti-trains-B experiments (Table 1). Pseudocode is shown in appendix Algorithm 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 530, + 459, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 530, + 459, + 596 + ], + "spans": [ + { + "bbox": [ + 138, + 530, + 459, + 596 + ], + "type": "text", + "content": "We have shown that by decoupling the scoring models from the learner model entirely, it is possible to significantly downscale the scoring models with minor degradation to performance (see Table 1). Unlike RHO, which can train a large 'learner' model with a small 'reference' model, we introduce a third 'online' model, with the same architecture and parameter count as the reference model, enabling the reduction of actor computation (see Table 1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 613, + 349, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 613, + 349, + 625 + ], + "spans": [ + { + "bbox": [ + 138, + 613, + 349, + 625 + ], + "type": "text", + "content": "4.5 ActiveCLIP: active multimodal learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 634, + 458, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 634, + 458, + 667 + ], + "spans": [ + { + "bbox": [ + 138, + 634, + 458, + 667 + ], + "type": "text", + "content": "We have so far demonstrated that large scale image classifiers can be trained with lower total compute by actively selecting the data used for training. However, classification has largely been superseded as a large scale pre-training method" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "spans": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "text", + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 448, + 143, + 456, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 448, + 143, + 456, + 151 + ], + "spans": [ + { + "bbox": [ + 448, + 143, + 456, + 151 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 163, + 442, + 221 + ], + "blocks": [ + { + "bbox": [ + 153, + 163, + 442, + 221 + ], + "lines": [ + { + "bbox": [ + 153, + 163, + 442, + 221 + ], + "spans": [ + { + "bbox": [ + 153, + 163, + 442, + 221 + ], + "type": "table", + "html": "
ImageNet0-ShotCOCO (im2txtR1)COCO (txt2imR1)(Learner)
Speed-up % (vs. Uniform)23.048.128.648.026.827.0ALIGN
0.016.318.822.128.217.5LTIP
Performance (vs. Uniform)47.8 (2.2)53.2 (7.4)40.9 (2.4)44.8 (6.3)27.3 (1.8)30.5 (5.0)ALIGN
46.5 (-0.1)47.2 (0.6)45.6 (1.2)45.4 (1.0)31.8 (1.7)31.0 (0.9)LTIP
(Reference)ALIGNLTIPALIGNLTIPALIGNLTIP
", + "image_path": "414efc23932b836c592dc6c59ee3c6fb7b68df0bf13d7ac35da3c865c38fe505.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 226, + 459, + 318 + ], + "lines": [ + { + "bbox": [ + 138, + 226, + 459, + 318 + ], + "spans": [ + { + "bbox": [ + 138, + 226, + 459, + 318 + ], + "type": "text", + "content": "Table 2: Generalizing data-selection policies across datasets and tasks. We pretrain reference models on the large but noisy ALIGN dataset, or the smaller and more curated LTIP dataset [2]. Consistent with [2], we find training with uniform sampling on LTIP to yield stronger transfer learning results than pre-training on ALIGN. These reference models can be used very effectively for data-selection on both LTIP and ALIGN, whereas ALIGN-pretrained reference models yield more modest speedups. All models are provided with 800M training images at resolution " + }, + { + "bbox": [ + 138, + 226, + 459, + 318 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 138, + 226, + 459, + 318 + ], + "type": "text", + "content": ", speedups are shown relative to the time at which the uniform sampling baseline was reached for that evaluation metric. Colour indicates performance relative to uniform (brackets)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 141, + 336, + 298, + 425 + ], + "blocks": [ + { + "bbox": [ + 141, + 336, + 298, + 425 + ], + "lines": [ + { + "bbox": [ + 141, + 336, + 298, + 425 + ], + "spans": [ + { + "bbox": [ + 141, + 336, + 298, + 425 + ], + "type": "image", + "image_path": "b840da8d161994ef6e8d8f5a76ac04441e1a2f3be9b0ce6906093daaa4b6e49b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 434, + 459, + 485 + ], + "lines": [ + { + "bbox": [ + 138, + 434, + 459, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 459, + 485 + ], + "type": "text", + "content": "Fig. 6: Reference policies generalize across tasks. We train a ViT-B reference model on either ALIGN or LTIP, then use it to train a second ViT-B learner on ALIGN with ActiveCLIP. The biggest gains were found with an LTIP reference model, despite it needing to perform out-of-domain generalization. In 50k iterations, ActiveCLIP selects the 800M \"cleanest\" examples from the ALIGN dataset, whose size is 1.6B in total." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 336, + 455, + 425 + ], + "blocks": [ + { + "bbox": [ + 299, + 336, + 455, + 425 + ], + "lines": [ + { + "bbox": [ + 299, + 336, + 455, + 425 + ], + "spans": [ + { + "bbox": [ + 299, + 336, + 455, + 425 + ], + "type": "image", + "image_path": "5fa7cd19b63e436f2c84975c09634e2d56b67ded39bd586dfea33e1978a14457.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 498, + 458, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 458, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 458, + 575 + ], + "type": "text", + "content": "by CLIP-style multimodal training [30]. Figure 6 demonstrates that our CLIP-adapted active learning method ActiveCLIP (see Methods) produces similar speedups in terms of learner computation as observed in JFT classification. Specifically, we find that prioritized sampling with learnability scores accelerates multimodal pre-training by " + }, + { + "bbox": [ + 138, + 498, + 458, + 575 + ], + "type": "inline_equation", + "content": "18 - 48\\%" + }, + { + "bbox": [ + 138, + 498, + 458, + 575 + ], + "type": "text", + "content": ", depending on the evaluation metric (ImageNet zero-shot accuracy or COCO retrieval) and reference model configuration, which we explore below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 592, + 318, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 592, + 318, + 603 + ], + "spans": [ + { + "bbox": [ + 138, + 592, + 318, + 603 + ], + "type": "text", + "content": "4.6 Policy generalization across tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 612, + 458, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 612, + 458, + 667 + ], + "spans": [ + { + "bbox": [ + 138, + 612, + 458, + 667 + ], + "type": "text", + "content": "To fully leverage off-the-shelf pre-trained image models for training (Figure 2), our results up to now suggest that the reference model should be trained on the same task that the learner model is being trained for. In Figure 6 (see Table 2 for full results), we show that we can in fact pre-train our reference models on related but distinct datasets. Moreover results suggest that there may even be a benefit" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "spans": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "spans": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "text", + "content": "T. Evans et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 164, + 456, + 253 + ], + "blocks": [ + { + "bbox": [ + 140, + 164, + 456, + 253 + ], + "lines": [ + { + "bbox": [ + 140, + 164, + 456, + 253 + ], + "spans": [ + { + "bbox": [ + 140, + 164, + 456, + 253 + ], + "type": "image", + "image_path": "1f4b4319b4f6f88c3ecccf5b128d2d33b845ad0f6872e07e8728ec8f31e9c94f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 262, + 459, + 303 + ], + "lines": [ + { + "bbox": [ + 138, + 262, + 459, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 459, + 303 + ], + "type": "text", + "content": "Fig. 7: Reference models trained on curated datasets are powerful data selectors. Active data selection using ActiveSigLIP with reference models trained on increasingly curated subsets of Webli. Webli-350M reference models are consistently more effective than those trained on raw or 1B subsets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 320, + 458, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 458, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 458, + 408 + ], + "type": "text", + "content": "to cross-training in some-cases; an ALIGN " + }, + { + "bbox": [ + 138, + 320, + 458, + 408 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 138, + 320, + 458, + 408 + ], + "type": "text", + "content": " ALIGN reference " + }, + { + "bbox": [ + 138, + 320, + 458, + 408 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 138, + 320, + 458, + 408 + ], + "type": "text", + "content": " learner model combination ('in-domain' ActiveCLIP) produced similar speedups to ClassAct on JFT. However, these gains were greatly surpassed by an LTIP " + }, + { + "bbox": [ + 138, + 320, + 458, + 408 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 138, + 320, + 458, + 408 + ], + "type": "text", + "content": " ALIGN combination. One possibility is that LTIP is less noisily labeled such that the scoring policies it produces are 'clean' - i.e. more able to filter for 'clean' data. The corollary may also be true; active data selection appears to greatly improve the utility of ALIGN, suggesting that it contains a large proportion of data that is of low quality for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 411, + 459, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 459, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 459, + 510 + ], + "type": "text", + "content": "We observed the same effect when training with ActiveSigLIP on the large scale Webli dataset [6] (results summarised in Table 3). We trained reference models on both the raw 4B set as well as two extensively curated 1B / 350M subsets, which were then used to train subsequent learner models (Fig. 7). In all-but-one cases, filtering data with the 350M-trained referenced model produced the best results when transferring to 350M, 1B and 4B learner datasets. Notably, the significant gains observed over training with uniform sampling on the 350M subset suggest that our method can still improve over methods that pre-filter datasets once but then train with uniform samples [12]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 538, + 331, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 538, + 331, + 548 + ], + "spans": [ + { + "bbox": [ + 139, + 538, + 331, + 548 + ], + "type": "text", + "content": "4.7 Comparison to prior multimodal art" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 568, + 458, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 458, + 667 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 458, + 667 + ], + "type": "text", + "content": "We leverage the insight that pre-training a reference model on clean data can facilitate learning on larger, noisy data for training our final model. Here, we train a reference model on LTIP, then use it to train a new model on the much larger mixture of LTIP and ALIGN, following [2], for a total of 3 and 8 billion training examples. Table 3 shows that in this training regime, ActiveCLIP surpasses models trained with significantly more data on ImageNet 0-Shot classification and COCO retrieval. Finally, we find that our active learning method is complementary to recent advances in multimodal learning: ActiveSigLIP significantly improves uniformly-trained SigLIP [44] in both COCO retrieval metrics." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 142, + 427, + 153 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 142, + 427, + 153 + ], + "spans": [ + { + "bbox": [ + 194, + 142, + 427, + 153 + ], + "type": "text", + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 448, + 143, + 457, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 448, + 143, + 457, + 151 + ], + "spans": [ + { + "bbox": [ + 448, + 143, + 457, + 151 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 163, + 401, + 282 + ], + "blocks": [ + { + "bbox": [ + 192, + 163, + 401, + 282 + ], + "lines": [ + { + "bbox": [ + 192, + 163, + 401, + 282 + ], + "spans": [ + { + "bbox": [ + 192, + 163, + 401, + 282 + ], + "type": "table", + "html": "
MethodTrain ex.IN-1K ZS Top-1COCO
im2txttxt2im
CLIP13B68.352.433.1
EVA-CLIP3B+2B69.7†
ActiveCLIP3B71.357.743.0
OpenCLIP34B70.259.442.3
EVA-CLIP8B+2B74.7†58.742.2
ActiveCLIP8B72.260.744.9
SigLIP3B72.160.742.7
ActiveSigLIP3B72.063.545.3
", + "image_path": "6916f601796f74c8ef924d536a053e2c05522d10f5599990a0b811dba914e8f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 288, + 459, + 350 + ], + "lines": [ + { + "bbox": [ + 138, + 288, + 459, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 459, + 350 + ], + "type": "text", + "content": "Table 3: Comparison of ActiveCLIP to public multimodal pre-taining methods. ActiveCLIP outperforms models trained with the same or more data (CLIP, [30]; EVA-CLIP, [37]; and OpenCLIP, [17]). ActiveSigLIP produced significant gains over the baseline SOTA performance of SigLIP [44]. " + }, + { + "bbox": [ + 138, + 288, + 459, + 350 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 138, + 288, + 459, + 350 + ], + "type": "text", + "content": "benefits from additional ImageNet21K pretraining (+2B training examples). All ActiveCLIP/SigLIP models use a reference model trained on LTIP to guide learning on a mixture of ALIGN and LTIP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 139, + 373, + 216, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 373, + 216, + 384 + ], + "spans": [ + { + "bbox": [ + 139, + 373, + 216, + 384 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 421, + 458, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 458, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 458, + 575 + ], + "type": "text", + "content": "In this work, we have presented a new method for active data selection that builds upon and simplifies the concept of 'learnability'. Our experiments demonstrate that this approach can significantly reduce the computation required for large-scale pretraining, compared to training with uniform samples. To our knowledge, this is the first active learning method that is more efficient than training with uniform samples when accounting for total FLOPs, and that does not rely on hand-designed features, allowing broad application across training setups. We have validated this by showing results on classification and contrastive pretraining, and found that our data selection policies continue to produce efficiency gains in the large-scale regime and can generalize effectively across task modalities. Collectively, our experiments also illustrate a Pareto frontier that allows trading off actor/data-selection computation against savings in training iterations, suggesting an alternative path to improved performance beyond scaling training batch sizes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 579, + 458, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 458, + 667 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 458, + 667 + ], + "type": "text", + "content": "This work focused on supervised pre-training for images, but further work could involve extending our method to other modalities and training schemes such as language, video, and generative modeling. An important note is that all our experiments present results from filtering only " + }, + { + "bbox": [ + 138, + 579, + 458, + 667 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 138, + 579, + 458, + 667 + ], + "type": "text", + "content": " of the data; further gains may be possible by filtering more aggressively, at the cost of further overheads. In particular, aggressive data-selection coupled with efficient scoring schemes such as the ones proposed here could test the hypothesis that large-scale pretraining can benefit from exponential, rather than power-law, scaling behavior." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "spans": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "spans": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "text", + "content": "T. Evans et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 139, + 164, + 199, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 164, + 199, + 175 + ], + "spans": [ + { + "bbox": [ + 139, + 164, + 199, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 191, + 457, + 667 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 143, + 191, + 456, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 191, + 456, + 221 + ], + "spans": [ + { + "bbox": [ + 143, + 191, + 456, + 221 + ], + "type": "text", + "content": "1. Abbas, A., Tirumala, K., Simig, D., Ganguli, S., Morcos, A.S.: Semdedup: Data-efficient learning at web-scale through semantic dedduplication. arXiv preprint arXiv:2303.09540 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 223, + 457, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 223, + 457, + 252 + ], + "spans": [ + { + "bbox": [ + 143, + 223, + 457, + 252 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: A visual language model for few-shot learning. Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 254, + 456, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 254, + 456, + 284 + ], + "spans": [ + { + "bbox": [ + 143, + 254, + 456, + 284 + ], + "type": "text", + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 286, + 456, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 456, + 314 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 456, + 314 + ], + "type": "text", + "content": "4. Campbell, T., Broderick, T.: Bayesian coreset construction via greedy iterative geodesic ascent. In: International Conference on Machine Learning. pp. 698-706. PMLR (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 316, + 456, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 316, + 456, + 336 + ], + "spans": [ + { + "bbox": [ + 143, + 316, + 456, + 336 + ], + "type": "text", + "content": "5. Cassirer, A., Barth-Maron, G., Brevdo, E., Ramos, S., Boyd, T., Sottiaux, T., Kroiss, M.: Reverb: A framework for experience replay (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 337, + 456, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 337, + 456, + 367 + ], + "spans": [ + { + "bbox": [ + 143, + 337, + 456, + 367 + ], + "type": "text", + "content": "6. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 368, + 456, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 368, + 456, + 408 + ], + "spans": [ + { + "bbox": [ + 143, + 368, + 456, + 408 + ], + "type": "text", + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 410, + 456, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 410, + 456, + 439 + ], + "spans": [ + { + "bbox": [ + 143, + 410, + 456, + 439 + ], + "type": "text", + "content": "8. Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., Zaharia, M.: Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 441, + 456, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 441, + 456, + 480 + ], + "spans": [ + { + "bbox": [ + 143, + 441, + 456, + 480 + ], + "type": "text", + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 481, + 456, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 481, + 456, + 522 + ], + "spans": [ + { + "bbox": [ + 139, + 481, + 456, + 522 + ], + "type": "text", + "content": "10. Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih, V., Ward, T., Doron, Y., Firoiu, V., Harley, T., Dunning, I., et al.: Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures. In: International conference on machine learning. pp. 1407-1416. PMLR (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 522, + 456, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 522, + 456, + 553 + ], + "spans": [ + { + "bbox": [ + 139, + 522, + 456, + 553 + ], + "type": "text", + "content": "11. Feldman, V.: Does learning require memorization? a short tale about a long tail. In: Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing. pp. 954-959 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 554, + 456, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 554, + 456, + 584 + ], + "spans": [ + { + "bbox": [ + 139, + 554, + 456, + 584 + ], + "type": "text", + "content": "12. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 139, + 585, + 456, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 585, + 456, + 615 + ], + "spans": [ + { + "bbox": [ + 139, + 585, + 456, + 615 + ], + "type": "text", + "content": "13. Gunasekar, S., Zhang, Y., Aneja, J., Mendes, C.C.T., Del Giorno, A., Gopi, S., Javaheripi, M., Kauffmann, P., de Rosa, G., Saarikivi, O., et al.: Textbooks are all you need. arXiv preprint arXiv:2306.11644 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 139, + 616, + 456, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 616, + 456, + 646 + ], + "spans": [ + { + "bbox": [ + 139, + 616, + 456, + 646 + ], + "type": "text", + "content": "14. Har-Peled, S., Kushal, A.: Smaller coresets for k-median and k-means clustering. In: Proceedings of the twenty-first annual symposium on Computational geometry. pp. 126-134 (2005)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 139, + 647, + 456, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 647, + 456, + 667 + ], + "spans": [ + { + "bbox": [ + 139, + 647, + 456, + 667 + ], + "type": "text", + "content": "15. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 195, + 142, + 428, + 153 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 142, + 428, + 153 + ], + "spans": [ + { + "bbox": [ + 195, + 142, + 428, + 153 + ], + "type": "text", + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 448, + 143, + 457, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 448, + 143, + 457, + 151 + ], + "spans": [ + { + "bbox": [ + 448, + 143, + 457, + 151 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 139, + 165, + 458, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 139, + 165, + 458, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 165, + 458, + 206 + ], + "spans": [ + { + "bbox": [ + 139, + 165, + 458, + 206 + ], + "type": "text", + "content": "16. Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D.d.L., Hendricks, L.A., Welbl, J., Clark, A., et al.: Training compute-optimal large language models. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 206, + 457, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 206, + 457, + 245 + ], + "spans": [ + { + "bbox": [ + 139, + 206, + 457, + 245 + ], + "type": "text", + "content": "17. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 245, + 458, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 245, + 458, + 284 + ], + "spans": [ + { + "bbox": [ + 139, + 245, + 458, + 284 + ], + "type": "text", + "content": "18. Jouppi, N.P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., Bates, S., Bhatia, S., Boden, N., Borchers, A., et al.: In-datacenter performance analysis of a tensor processing unit. In: Proceedings of the 44th annual international symposium on computer architecture. pp. 1-12 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 284, + 457, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 284, + 457, + 314 + ], + "spans": [ + { + "bbox": [ + 139, + 284, + 457, + 314 + ], + "type": "text", + "content": "19. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 314, + 457, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 314, + 457, + 344 + ], + "spans": [ + { + "bbox": [ + 139, + 314, + 457, + 344 + ], + "type": "text", + "content": "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 344, + 457, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 344, + 457, + 363 + ], + "spans": [ + { + "bbox": [ + 139, + 344, + 457, + 363 + ], + "type": "text", + "content": "21. Lindley, D.V.: On a measure of the information provided by an experiment. The Annals of Mathematical Statistics 27(4), 986-1005 (1956)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 363, + 457, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 363, + 457, + 382 + ], + "spans": [ + { + "bbox": [ + 139, + 363, + 457, + 382 + ], + "type": "text", + "content": "22. Loshchilov, I., Hutter, F.: Online batch selection for faster training of neural networks. arXiv preprint arXiv:1511.06343 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 383, + 457, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 383, + 457, + 402 + ], + "spans": [ + { + "bbox": [ + 139, + 383, + 457, + 402 + ], + "type": "text", + "content": "23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 402, + 457, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 402, + 457, + 421 + ], + "spans": [ + { + "bbox": [ + 139, + 402, + 457, + 421 + ], + "type": "text", + "content": "24. MacKay, D.J.: Information-based objective functions for active data selection. Neural computation 4(4), 590-604 (1992)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 421, + 457, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 421, + 457, + 450 + ], + "spans": [ + { + "bbox": [ + 139, + 421, + 457, + 450 + ], + "type": "text", + "content": "25. Mahmoud, A., Elhoushi, M., Abbas, A., Yang, Y., Ardalani, N., Leather, H., Morcos, A.: Sieve: Multimodal dataset pruning using image captioning models. arXiv preprint arXiv:2310.02110 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 450, + 457, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 450, + 457, + 480 + ], + "spans": [ + { + "bbox": [ + 139, + 450, + 457, + 480 + ], + "type": "text", + "content": "26. Marion, M., Üstün, A., Pozzobon, L., Wang, A., Fadaee, M., Hooker, S.: When less is more: Investigating data pruning for pretraining llms at scale. arXiv preprint arXiv:2309.04564 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 480, + 457, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 480, + 457, + 519 + ], + "spans": [ + { + "bbox": [ + 139, + 480, + 457, + 519 + ], + "type": "text", + "content": "27. Mindermann, S., Brauner, J.M., Razzak, M.T., Sharma, M., Kirsch, A., Xu, W., Höltgen, B., Gomez, A.N., Morisot, A., Farquhar, S., et al.: Prioritized training on points that are learnable, worth learning, and not yet learnt. In: International Conference on Machine Learning. pp. 15630-15649. PMLR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 519, + 457, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 519, + 457, + 548 + ], + "spans": [ + { + "bbox": [ + 139, + 519, + 457, + 548 + ], + "type": "text", + "content": "28. Paul, M., Ganguli, S., Dziugaite, G.K.: Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems 34, 20596-20607 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 139, + 549, + 457, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 549, + 457, + 568 + ], + "spans": [ + { + "bbox": [ + 139, + 549, + 457, + 568 + ], + "type": "text", + "content": "29. Prabhu, A., Dognin, C., Singh, M.: Sampling bias in deep active classification: An empirical study. arXiv preprint arXiv:1909.09389 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 139, + 568, + 457, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 568, + 457, + 607 + ], + "spans": [ + { + "bbox": [ + 139, + 568, + 457, + 607 + ], + "type": "text", + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 139, + 607, + 457, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 607, + 457, + 626 + ], + "spans": [ + { + "bbox": [ + 139, + 607, + 457, + 626 + ], + "type": "text", + "content": "31. Schaul, T., Quan, J., Antonoglou, I., Silver, D.: Prioritized experience replay. arXiv preprint arXiv:1511.05952 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 139, + 626, + 457, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 626, + 457, + 666 + ], + "spans": [ + { + "bbox": [ + 139, + 626, + 457, + 666 + ], + "type": "text", + "content": "32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "spans": [ + { + "bbox": [ + 139, + 143, + 150, + 151 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "spans": [ + { + "bbox": [ + 169, + 143, + 226, + 151 + ], + "type": "text", + "content": "T. Evans et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 139, + 165, + 457, + 487 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 139, + 165, + 457, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 165, + 457, + 197 + ], + "spans": [ + { + "bbox": [ + 139, + 165, + 457, + 197 + ], + "type": "text", + "content": "33. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 197, + 345, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 197, + 345, + 206 + ], + "spans": [ + { + "bbox": [ + 139, + 197, + 345, + 206 + ], + "type": "text", + "content": "34. Settles, B.: Active learning literature survey (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 207, + 457, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 207, + 457, + 236 + ], + "spans": [ + { + "bbox": [ + 139, + 207, + 457, + 236 + ], + "type": "text", + "content": "35. Sorscher, B., Geirhos, R., Shekhar, S., Ganguli, S., Morcos, A.: Beyond neural scaling laws: beating power law scaling via data pruning. Advances in Neural Information Processing Systems 35, 19523-19536 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 237, + 457, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 237, + 457, + 266 + ], + "spans": [ + { + "bbox": [ + 139, + 237, + 457, + 266 + ], + "type": "text", + "content": "36. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the IEEE international conference on computer vision. pp. 843-852 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 267, + 456, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 267, + 456, + 286 + ], + "spans": [ + { + "bbox": [ + 139, + 267, + 456, + 286 + ], + "type": "text", + "content": "37. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 287, + 456, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 287, + 456, + 316 + ], + "spans": [ + { + "bbox": [ + 139, + 287, + 456, + 316 + ], + "type": "text", + "content": "38. Toneva, M., Sordoni, A., Combes, R.T.d., Trischler, A., Bengio, Y., Gordon, G.J.: An empirical study of example forgetting during deep neural network learning. arXiv preprint arXiv:1812.05159 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 317, + 456, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 317, + 456, + 346 + ], + "spans": [ + { + "bbox": [ + 139, + 317, + 456, + 346 + ], + "type": "text", + "content": "39. Xie, S.M., Pham, H., Dong, X., Du, N., Liu, H., Lu, Y., Liang, P., Le, Q.V., Ma, T., Yu, A.W.: Doremi: Optimizing data mixtures speeds up language model pretraining (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 347, + 456, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 347, + 456, + 366 + ], + "spans": [ + { + "bbox": [ + 139, + 347, + 456, + 366 + ], + "type": "text", + "content": "40. Xie, S.M., Santurkar, S., Ma, T., Liang, P.: Data selection for language models via importance resampling. arXiv preprint arXiv:2302.03169 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 367, + 457, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 367, + 457, + 405 + ], + "spans": [ + { + "bbox": [ + 139, + 367, + 457, + 405 + ], + "type": "text", + "content": "41. Yang, F., Barth-Maron, G., Stanczyk, P., Hoffman, M., Liu, S., Kroiss, M., Pope, A., Rrustemi, A.: Launchpad: A programming model for distributed machine learning research. arXiv preprint arXiv:2106.04516 (2021), https://arxiv.org/abs/2106.04516" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 407, + 457, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 407, + 457, + 436 + ], + "spans": [ + { + "bbox": [ + 139, + 407, + 457, + 436 + ], + "type": "text", + "content": "42. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: CoCa: Contrastive captioners are image-text foundation models. In: Transactions on Machine Learning Research (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 437, + 457, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 437, + 457, + 466 + ], + "spans": [ + { + "bbox": [ + 139, + 437, + 457, + 466 + ], + "type": "text", + "content": "43. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 467, + 457, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 467, + 457, + 487 + ], + "spans": [ + { + "bbox": [ + 139, + 467, + 457, + 487 + ], + "type": "text", + "content": "44. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "spans": [ + { + "bbox": [ + 194, + 142, + 428, + 153 + ], + "type": "text", + "content": "Active Learning Accelerates Large-Scale Visual Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 448, + 143, + 457, + 151 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 448, + 143, + 457, + 151 + ], + "spans": [ + { + "bbox": [ + 448, + 143, + 457, + 151 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_content_list.json b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a787b944ec9a97234534be4198cef9714947f3ac --- /dev/null +++ b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_content_list.json @@ -0,0 +1,1872 @@ +[ + { + "type": "text", + "text": "Bayesian Detector Combination for Object Detection with Crowdsourced Annotations", + "text_level": 1, + "bbox": [ + 256, + 140, + 746, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhi Qin $\\mathrm{Tan}^{1}$ , Olga Isupova $^{2}$ , Gustavo Carneiro $^{1}$ , Xiatian Zhu $^{1}$ , and Yunpeng Li $^{1}$", + "bbox": [ + 232, + 212, + 771, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 University of Surrey, Surrey, UK {z.tan,g.carneiro, xiatian.zhu, yunpeng.li}@surrey.ac.uk", + "$^{2}$ University of Oxford, Oxford, UK olga.isupova@ouce.ox.ac.uk" + ], + "bbox": [ + 290, + 253, + 710, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Acquiring fine-grained object detection annotations in unconstrained images is time-consuming, expensive, and prone to noise, especially in crowdsourcing scenarios. Most prior object detection methods assume accurate annotations; A few recent works have studied object detection with noisy crowdsourced annotations, with evaluation on distinct synthetic crowdsourced datasets of varying setups under artificial assumptions. To address these algorithmic limitations and evaluation inconsistency, we first propose a novel Bayesian Detector Combination (BDC) framework to more effectively train object detectors with noisy crowdsourced annotations, with the unique ability of automatically inferring the annotators' label qualities. Unlike previous approaches, BDC is model-agnostic, requires no prior knowledge of the annotators' skill level, and seamlessly integrates with existing object detection models. Due to the scarcity of real-world crowdsourced datasets, we introduce large synthetic datasets by simulating varying crowdsourcing scenarios. This allows consistent evaluation of different models at scale. Extensive experiments on both real and synthetic crowdsourced datasets show that BDC outperforms existing state-of-the-art methods, demonstrating its superiority in leveraging crowdsourced data for object detection. Our code and data are available at: https://github.com/zhiqin1998/bdc.", + "bbox": [ + 259, + 352, + 743, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Bayesian Detector Combination $\\cdot$ Object detection $\\cdot$ Crowdsourced annotations", + "bbox": [ + 259, + 643, + 740, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 699, + 375, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant advancements in the development of object detection with natural images [9, 14, 41, 53], existing methods typically assume the availability of large, accurately annotated datasets [37, 61]. Nonetheless, this assumption does not often hold true in practice. For example, accurate labels are often expensive to acquire in many applications. A large number of experts with varying professional skills can be involved to crowdsourced annotations of images sampled with replacement from a large dataset [8, 33, 50]. Usually, the annotations", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5581d5e72fc9c81d38c5028897e9a5518e186141340e1acc021947b725279667.jpg", + "image_caption": [ + "(a) ID: 70282" + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 330, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/05b271ffd556ac007a75bd771bc7ee4d9dad08270afa4c676a571740ba5cb60d.jpg", + "image_caption": [ + "(b) ID: 310072" + ], + "image_footnote": [], + "bbox": [ + 333, + 146, + 442, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7c49c0cf0db62ebd904ad3997472f95222e089f38f578d517c36193528ec48b6.jpg", + "image_caption": [ + "(c) ID: 516471" + ], + "image_footnote": [], + "bbox": [ + 446, + 143, + 555, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3820121acb52d132162c76ac90c25a9dc2bfbf62a610eba4b4327850149e7352.jpg", + "image_caption": [ + "(d) ID: 326247" + ], + "image_footnote": [], + "bbox": [ + 558, + 143, + 668, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/949e6bef41e345a45c73d9d09ad61e28e959f5bb3c172e14ff7b58426efc7c80.jpg", + "image_caption": [ + "(e) ID: 416149" + ], + "image_footnote": [], + "bbox": [ + 669, + 143, + 782, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8e042cdbfeb2aa75c73c2d5e938c17f218e83c28d8e6618f1b7e8573feddbc1f.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 220, + 219, + 330, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fff13b2842815147cfd3d4b49ef0b5c8d462983aec8ade574b71941c70c3c989.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 334, + 219, + 442, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/19b4649dc5193e2e3e0ee6906aaac5aa9fddf2a9cca64f8b72fde147e41ae2b3.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 446, + 218, + 555, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e92d201e483d90a1e0d17f8c202ec7ae4118fab72849de30b9418a0a491151d0.jpg", + "image_caption": [ + "(i)" + ], + "image_footnote": [], + "bbox": [ + 558, + 219, + 666, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/677293938fd74d7bc8b9f3c3ea8cc1df18e08e3ebec5c5329a09f7c8a9ab1d1d.jpg", + "image_caption": [ + "(j)" + ], + "image_footnote": [], + "bbox": [ + 669, + 219, + 782, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f8d60cf4d80ca7e1766cba646ba9a50bcb5940ed891190f255bbe9a17fd568bf.jpg", + "image_caption": [ + "(k)", + "Fig. 1: Examples of ambiguous cases with noisy or incorrect annotations on (a - e) MS COCO [29], (f - j) VinDr-CXR [36] and (k - o) disaster response dataset [19] to identify damaged and undamaged buildings with class names 'd' and 'u', respectively. (a) Mislabelling one object with two bounding boxes. (b - d) Duplicate annotations of visually similar objects (e.g., car/truck, donut/cake and fork/spoon). (d)(e) Incorrect annotation of object classes (e.g., cake labelled as a bowl, only annotating one of the cake decorations as 'person'). (f - o) Annotations from different annotators (represented by the different colours) of the VinDr-CXR and disaster response datasets show significant disagreement. (Details of the datasets are given in Section 4.1.)" + ], + "image_footnote": [], + "bbox": [ + 220, + 297, + 330, + 354 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d8cb7ef14e5c5572a1cd37f29efe4569feeb092a6453cfdbc8bf6a552cb71c2b.jpg", + "image_caption": [ + "(1)" + ], + "image_footnote": [], + "bbox": [ + 334, + 297, + 442, + 354 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9fea8b8b2113587af9b95723f320e5cf9a498386174aba9030159c584bcc2971.jpg", + "image_caption": [ + "(m)" + ], + "image_footnote": [], + "bbox": [ + 446, + 296, + 555, + 354 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/700a5490c136f968efa06c99faec1fd17bef5e9fbe67a58f51a1dc930e29ebf3.jpg", + "image_caption": [ + "(n)" + ], + "image_footnote": [], + "bbox": [ + 558, + 296, + 666, + 354 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4e93179fff9ab2410b3a0e9d3b0a501c7408214b322729916452c338258a2221.jpg", + "image_caption": [ + "(o)" + ], + "image_footnote": [], + "bbox": [ + 669, + 296, + 782, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from multiple annotators are unified and cleaned as the last step in annotation collection. However, disagreements arise when annotating complex images, making it challenging even for experienced expert annotators to achieve unanimity due to high interobserver variability [26,54] and thus, rendering the role of an annotation arbitrator nearly impossible (see Figure 1). As a result, multiple noisy annotations per image are obtained. Such a setting for annotation crowdsourcing is often called the multi-rater problem. The majority of methods on multi-rater learning have been developed for image classification [20,50,59] and segmentation [4,21,22,24,58,60], while we consider object detection here.", + "bbox": [ + 212, + 546, + 787, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several previous methods that exploit crowdsourced annotations have been proposed to improve object detection performance [18,27]. A common approach estimates the ground truth from noisy annotations through majority voting [46] which implicitly assumes that all annotators are equally accurate in their annotations. Alternatively, in [27], the weight of each annotation is predicted based on the annotators' proficiency levels which are assumed to be known a priori; however, this information is difficult to obtain in practice. On the other hand, in [18], their proposed method is designed specifically for Faster R-CNN (FRCNN) [41] and cannot be generalised to other object detectors. Moreover, these methods have been developed and evaluated on distinct, private synthetic", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z.Q. Tan et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/09a052399c6674ee465a327714efd9bba0fb41e3aabf7c7bd0d7375e20746c0e.jpg", + "image_caption": [ + "Fig. 2: Overall architecture of the proposed BDC. The updating of the aggregator's parameter and the object detector's parameters is repeated iteratively until convergence. Best viewed in colour." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 769, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "datasets constructed using different methods. Thus, we cannot directly compare their reported experiment results.", + "bbox": [ + 212, + 371, + 784, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these algorithmic and evaluation limitations, in this work, we first propose a model-agnostic annotation aggregation and object detector training pipeline, named Bayesian Detector Combination (BDC). BDC aggregates noisy bounding boxes and class labels by modelling the annotation reliability of each crowd member for both elements (Figure 2), without requiring any additional knowledge of the ground truth annotations or their quality. Furthermore, due to the limited access to the ground truth of real crowdsourced datasets, we synthesise multiple datasets simulating various crowdsourcing scenarios for systematic evaluation of the previous works and the proposed method. The main contributions of our work include:", + "bbox": [ + 212, + 401, + 785, + 551 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a model-agnostic Bayesian Detector Combination (BDC) framework that simultaneously infers the annotation quality of each crowd member and infers consensus bounding boxes and soft class labels from noisy, crowdsourced annotations to improve model performance.", + "- We implement and systematically evaluate the previous methods on real-world and synthetic datasets with crowdsourced annotations which simulate varying crowdsourcing scenarios.", + "- Extensive experiments conducted demonstrate the superior performance of BDC compared with prior alternatives in key metrics such as the average precision, the scalability with the increasing number of annotators, and the robustness against noisy annotators." + ], + "bbox": [ + 225, + 561, + 784, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 750, + 387, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classification with Crowdsourced Annotations. Learning with crowdsourced annotations for classification tasks has been studied in various works. Sheng et al. [46] adopted a simple approach to aggregate multiple labels via majority voting (MV) where the mode class is selected as the ground truth. Several", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bayesian Detector Combination", + "bbox": [ + 517, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "works [11,39,57] modelled the sensitivity and specificity of different annotators via confusion matrices obtained through the Expectation-Maximisation algorithm or a Bayesian approach. Kim and Ghahramani [23] introduced a general framework for Bayesian classifier combination (BCC) to model the relationship between the unknown true label and the annotated class label which is extended by Venanzi et al. [51] with the concept of annotator communities. Albarqouni et al. [3] and Isupova et al. [19] extended the idea of previous works [23,39] to convolutional neural networks (CNN) with softmax outputs. Several recent works use a trained classifier to rank consensus labels and incorporate the ranking scores into model learning [7,15]. Another research direction is to take all noisy annotations separately as the input during the training process [10,16,20,43,50,55,59], but Wei et al. [56] concluded that label aggregation is preferred over label separation when there are sufficient number of annotators or the noise rate is moderate. These methods are not directly applicable to object detection tasks since they do not involve regressing the locations of bounding boxes.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Detection with Crowdsourced Annotations. Hu and Meina [18] proposed Crowd R-CNN where one first pre-processes the annotations by combining the bounding boxes using a clustering algorithm, before aggregating the labels and learning the sensitivity of each annotator [39] to different object classes during model training. Crowd R-CNN is tailored for Faster R-CNN (FRCNN) [41] and is incompatible with one-stage object detectors such as YOLOv7 [53]. Le et al. [27] proposed to pre-process annotations using Weighted Boxes Fusion (WBF) [48] and introduced expert agreement re-weighted loss (EARL) to emphasise boxes with higher annotator agreement for model training. It requires access to each annotator's proficiency level to perform WBF and compute EARL, which is often challenging to obtain for real-world crowdsourced data. Compared with the aforementioned methods, the proposed BDC framework can incorporate any object detector in a plug-and-play fashion without requiring any additional inputs such as annotators' proficiency levels. Furthermore, previous works [18, 27] evaluated their methods on different synthetic datasets, with different numbers of annotators and proficiency levels. For example, Hu and Meina [18] performed evaluation on synthetic datasets derived from PASCAL VOC [13] and MS COCO [29]; Le et al. [27] synthesised a MED-MNIST dataset based on MNIST [28] and adopted the VinDr-CXR dataset [36]. Thus, their experiment results cannot be compared directly.", + "bbox": [ + 212, + 388, + 787, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 714, + 330, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Statement. We assume a training set $D = \\{x^{i},y^{i}\\}_{i = 1}^{N}$ with pairs of input images $x^{i}$ and crowdsourced annotations $y^{i} = \\{y_{m}^{i}\\}_{m = 1}^{M^{i}}$ , where $N$ is the total number of training samples and $M^{i}$ is the number of annotations of the $i$ -th image from all annotators. Each annotation is a tuple of a bounding box, class label and the annotator ID, $y_{m}^{i} = \\{(b_{m}^{i},c_{m}^{i},k_{m}^{i})\\}$ in which $b_{m}^{i}\\in \\mathbb{R}^{4}$ contains the x-y coordinates of the box centre as well as the width and height", + "bbox": [ + 212, + 746, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z.Q. Tan et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the bounding box, $c_{m}^{i} \\in \\{1, \\dots, J\\}$ represents the annotated class label of the object and $k_{m}^{i} \\in \\{1, \\dots, K\\}$ is the annotator ID. $J$ denotes the number of object classes while $K$ is the total number of annotators. Note that each annotator may annotate a varying number of object instances. For brevity, notations in the rest of this section are simplified to a single training sample $i$ . Our goal is to learn an object detector with noisy crowdsourced annotations while inferring and leveraging annotators' varying annotation abilities.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Bayesian Detector Combination", + "text_level": 1, + "bbox": [ + 214, + 273, + 524, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An overview of the proposed crowdsourced annotation aggregation framework is shown in Figure 2. At a high level, the proposed framework consists of four components, namely: the object detector (OD) module, the annotations-predictions matcher, the bounding box aggregator (BBA) and the class label aggregator (CLA). Firstly, we perform a one-to-many matching of the predictions generated by the OD to the crowdsourced annotations. The bounding box error is computed and used to update a Gaussian-Gamma prior distribution of the BBA for each annotator. Next, the class probability predictions by the OD and the class label annotations are used to update a Dirichlet prior in the CLA. The posterior distributions obtained from both aggregators are used to aggregate the bounding boxes and class labels. Finally, we train and optimise the OD's parameters and aggregators' parameters iteratively until convergence. We present a detailed introduction of each component below.", + "bbox": [ + 212, + 297, + 787, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object Detector Module. The object detector (OD) module with a parameter set, $\\theta_{O}$ , outputs a varying number of $P$ predictions, $\\hat{y} = \\{(\\hat{b}_n,\\hat{p}_n)\\}_{n = 1}^P$ , given an input image, $x$ , where $\\hat{b}_n\\in \\mathbb{R}^4$ and $\\hat{p}_n\\in [0,1]^J$ are the predicted bounding box and class probability, respectively. We train the model by minimising an overall loss function:", + "bbox": [ + 212, + 513, + 787, + 593 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {D}} = \\frac {1}{H} \\sum_ {n = 1} ^ {H} \\gamma_ {n} \\mathcal {L} \\left(\\mathfrak {b} _ {n}, \\rho_ {n}, \\hat {y}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 599, + 785, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma_{n}$ , $\\mathfrak{b}_{n}$ and $\\rho_{n}$ are the loss scale factor, the aggregated bounding box, and class probabilities obtained from the two aggregators defined in Section 3.1, respectively. $H$ is the number of aggregated annotations. $\\mathcal{L}$ is the loss function that the base OD model adopts. Since the OD operates as a black-box module in the proposed framework, any object detector can be used here as long as it outputs the predicted bounding boxes and the associated class label probabilities.", + "bbox": [ + 212, + 648, + 787, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotations-Predictions Matcher. To update the priors of the aggregators, we first match each annotation to one of the predictions. To find the optimal prediction, $\\hat{y}_m^*$ , for each annotation, $y_m$ , we minimise a matching cost function:", + "bbox": [ + 212, + 761, + 787, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {m} ^ {*} = \\underset {\\hat {y} _ {n} \\in \\hat {y}} {\\arg \\min } \\mathcal {L} _ {\\text {m a t c h}} (\\hat {y} _ {n}, y _ {m}) , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 816, + 785, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bayesian Detector Combination", + "bbox": [ + 517, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 146, + 264, + 159 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a t c h}} (\\hat {y} _ {n}, y _ {m}) = - \\hat {p} _ {n (c _ {m})} + \\lambda_ {1} \\mathcal {L} _ {I o U} (\\hat {b} _ {n}, b _ {m}) + \\lambda_ {2} | | \\hat {b} _ {n} - b _ {m} | | _ {1}. \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 167, + 785, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\lambda_{1}, \\lambda_{2} \\in \\mathbb{R}$ are hyperparameters and $\\mathcal{L}_{IoU}$ is the IoU loss [42]. In short, for each annotation, we aim to find the prediction with a higher predicted probability of the annotated class label, a higher IoU and a smaller L1 distance between the predicted and the annotated bounding box. This is inspired by the set prediction loss in DETR [9].", + "bbox": [ + 214, + 191, + 785, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instead of finding a one-to-one matching with a global minimum matching cost, we are interested in a one-to-many matching that one prediction can be matched to many crowdsourced annotations, with a local minimum matching cost, i.e. finding the prediction with minimum cost for each annotation individually. Moreover, a simple heuristic matching rule is sufficient instead of the Hungarian algorithm in DETR which is more computationally demanding.", + "bbox": [ + 214, + 268, + 787, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bounding Box Aggregator. In the Bounding Box Aggregator (BBA), the translation and scaling errors in the x and y-axis of each annotator are modelled with a Gaussian distribution:", + "bbox": [ + 214, + 378, + 787, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\epsilon_ {m} \\mid k _ {m} = k, \\boldsymbol {\\mu}, \\boldsymbol {\\sigma}\\right) = \\mathcal {N} \\left(\\mu^ {k}, \\sigma^ {k}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 436, + 785, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{m} \\in \\mathbb{R}^{4}$ contains the translation and scaling errors in the x and y-axis. $\\pmb{\\mu}$ and $\\pmb{\\sigma}$ are a collection of Gaussian parameters with $\\mu^k$ and $\\sigma^k$ representing the mean and covariance of $\\epsilon_{m}$ for the annotator $k$ , respectively. The Gaussian distributions have a Gaussian-Gamma prior with hyperparameters $\\mu_0^k$ , $v_0^k$ and $\\beta_0^k$ [6, 35]. To update the prior distribution, we compute the four error terms between all matched prediction and annotation boxes for each annotator with:", + "bbox": [ + 214, + 458, + 787, + 551 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {m} = \\left[ \\hat {b} _ {m (1)} ^ {*} - b _ {m (1)}, \\hat {b} _ {m (2)} ^ {*} - b _ {m (2)}, \\hat {b} _ {m (3)} ^ {*} \\div b _ {m (3)}, \\hat {b} _ {m (4)} ^ {*} \\div b _ {m (4)} \\right], \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 556, + 785, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{b}_m^*$ and $b_{m}$ are the matched prediction and annotation boxes, respectively. The subtraction and division operations are designed to make the error terms translation-invariant and scale-invariant. The Gaussian-Gamma prior is updated as follows:", + "bbox": [ + 214, + 589, + 787, + 648 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mu} ^ {k} = \\frac {\\mu_ {0} ^ {k} + M ^ {k} \\bar {\\epsilon} ^ {k}}{M ^ {k} + 1}, \\quad (6) \\quad \\tilde {v} ^ {k} = v _ {0} ^ {k} + \\frac {M ^ {k}}{2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 648, + 785, + 680 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\beta} ^ {k} = \\beta_ {0} ^ {k} + \\frac {M ^ {k} (\\bar {\\epsilon} ^ {k} - \\mu_ {0} ^ {k}) (\\bar {\\epsilon} ^ {k} - \\mu_ {0} ^ {k}) ^ {\\top}}{2 (M ^ {k} + 1)} + \\frac {1}{2} \\sum_ {m = 1} ^ {M} \\mathbb {1} [ k _ {m} = k ] (\\epsilon_ {m} - \\bar {\\epsilon} ^ {k}) (\\epsilon_ {m} - \\bar {\\epsilon} ^ {k}) ^ {\\top}, (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 691, + 785, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M^k$ is the number of annotation-prediction matches associated with the $k$ -th annotator, $\\overline{\\epsilon}^k$ is the mean of all $\\epsilon_{m}$ for the $k$ -th annotator and $\\mathbb{1}[\\cdot]$ is the indicator function.", + "bbox": [ + 214, + 738, + 785, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The obtained posterior mean (i.e. $\\mu^k = \\tilde{\\mu}^k$ ) is used to correct the annotated boxes for each annotator:", + "bbox": [ + 214, + 784, + 785, + 813 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nb _ {m} := \\left(b _ {m} + \\left[ \\mu_ {(1)} ^ {k}, \\mu_ {(2)} ^ {k}, 0, 0 \\right]\\right) \\odot \\left[ 1, 1, \\mu_ {(3)} ^ {k}, \\mu_ {(4)} ^ {k} \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 818, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z.Q. Tan et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All annotations matched to the same prediction are aggregated via a weighted average using the posterior precision, $P^k = 1 \\oslash \\text{diag}(\\sigma^k)$ , where $\\oslash$ denotes the element-wise division and $\\sigma^k = \\tilde{\\beta}^k / \\tilde{v}^k$ . The posterior precision, $P^k$ , of each annotator is used as weights to obtain the final boxes, $\\{\\mathfrak{b}_n\\}_{n=1}^H$ , with", + "bbox": [ + 214, + 146, + 787, + 209 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathfrak {b} _ {n} = \\frac {\\sum_ {(b , k) \\in \\kappa_ {n}} P ^ {k} \\odot b}{\\sum_ {(b , k) \\in \\kappa_ {n}} P ^ {k}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 219, + 785, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $H$ indicates the number of predictions matched to at least one annotation which is also equal to the number of aggregated annotations defined in Section 3.1. $\\kappa_{n}$ is the set of tuples consisting of the annotated bounding boxes and their annotator ID that are matched to the $n$ -th prediction. Inspired by EARL [27], a loss scale factor, $\\gamma_{n}$ , is computed for each aggregated box as the ratio of the number of matched annotations to $K$ and is used to re-weight the loss function of the OD in Equation (1). We refer the readers to Appendix A for further details on the updating procedure of the Bounding Box Aggregator.", + "bbox": [ + 214, + 265, + 787, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Class Label Aggregator. To aggregate crowdsourced class labels, we extend the Bayesian classifier combination neural network (BCCNet) [19] to the proposed BDC framework. In the class label aggregator (CLA), the annotated class label from each annotator, $c_{m}$ , is modelled as a multinomial distribution conditioning on the true label of the object:", + "bbox": [ + 214, + 407, + 787, + 484 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np \\left(c _ {m} \\mid k _ {m} = k, t _ {m} = j, \\boldsymbol {\\pi}\\right) = \\pi_ {j, c _ {m}} ^ {k}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 494, + 785, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pi$ is a collection of confusion matrices, $\\pi^k\\in \\mathbb{R}^{J\\times J}$ is the confusion matrix of the $k$ -th annotator and $t_m$ is the ground truth class label. Each row, $\\pi_j^k$ , provides the event probabilities of a multinomial distribution for the $k$ -th annotator when $t_m = j$ and is modelled with a Dirichlet prior with hyperparameters $\\alpha_{0j}^k = \\{\\alpha_{0j,1}^k,\\dots,\\alpha_{0j,J}^k\\}$ such that $\\pi_j^k |\\alpha_{0j}^k\\sim \\text{Dir}(\\alpha_{0j}^k)$ [23].", + "bbox": [ + 214, + 522, + 787, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The multinomial posterior distribution of the true label (i.e. the aggregated class label probability) can be obtained as", + "bbox": [ + 214, + 599, + 787, + 630 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {n, j} = \\exp \\left(\\ln \\hat {p} _ {n, j} + \\sum_ {(c, k) \\in \\tilde {\\kappa} _ {n}} \\mathbb {E} _ {\\pi_ {j} ^ {k}} \\ln \\pi_ {j, c} ^ {k}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 641, + 785, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tilde{\\kappa}_n$ is the set of tuples consisting of the annotated class labels and their annotator ID that are matched to the $n$ -th prediction. $\\hat{p}_{n,j}$ is the OD's class- $j$ probability of the $n$ -th prediction. Note that the CLA outputs class probabilities (soft labels) unlike other methods (e.g. MV) that output a single class (hard label). Appendix B provides the derivation for Equation 12 and a detailed discussion of the updating steps of the Class Label Aggregator.", + "bbox": [ + 214, + 696, + 787, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training. The optimal OD's parameters, $\\theta_{O}$ , and aggregators' parameters, $\\pmb{\\mu}$ , $\\pmb{v}$ , $\\pmb{\\beta}$ and $\\pmb{\\alpha}$ , are learnt iteratively in the BDC framework. Firstly, we initialise", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bayesian Detector Combination", + "bbox": [ + 517, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the parameters either randomly or with prior knowledge of the domain. Next, we train the OD by optimising the loss function, $\\mathcal{L}_{\\mathcal{D}}$ , with the aggregated annotations obtained using the current values of the aggregators' parameters. After one epoch of training, we obtain the OD's predictions using the current value of $\\theta_{O}$ to update the prior distributions of aggregators. These two steps are repeated iteratively until the parameters converge. See Appendix C for a further discussion of the training steps.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 273, + 375, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training details. To demonstrate the generalisability and robustness of BDC in incorporating different object detection algorithms, three common object detection algorithms are used in our experiments, YOLOv7 [53], FRCNN [41] with the ResNet50 backbone and EVA [14]. During training, we randomly apply data augmentation such as rotation, translation, zooming and horizontal flipping to the training images and annotations. $\\lambda_{1},\\lambda_{2}$ in Equation (3) are set to 2 and 5, respectively, following Carion et al. [9]. Lastly, the Gaussian-Gamma prior of the BBA, $\\pmb{\\mu}_{0}$ , $\\pmb{v}_{0}$ and $\\beta_0$ , are set to 0, 10 and 0.5, respectively, while the Dirichlet prior of the CLA, $\\alpha_0$ , is initialised to 1 and each diagonal entry is set to 10. The complete training configuration can be found in Appendix D.", + "bbox": [ + 212, + 301, + 787, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison against other aggregation methods. We compare the performance of BDC against multiple baselines as shown in Tables 1 to 4. 'NA' denotes the process of training the model with all the crowdsourced annotations as the ground truth annotations [56]. 'MV' stands for majority voting where we apply clustering to the crowdsourced annotations, then aggregate the class labels via MV and aggregate the bounding boxes as the biggest box where the majority had annotated pixel-wise, following the practice in [46]. We also implement Crowd R-CNN [18] and WBF-EARL [27] designed for crowdsourced object detection. For WBF-EARL [27] which requires the knowledge of annotators' proficiency levels, we followed the settings in [27] for the VinDR-CXR dataset while for the synthetic datasets, they are set to the overall accuracy of the annotators.", + "bbox": [ + 212, + 465, + 787, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Real Crowdsourced Datasets", + "text_level": 1, + "bbox": [ + 215, + 651, + 501, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VinDr-CXR. We evaluate the annotation aggregation methods on VinDrcxR [36], a thoracic abnormalities detection dataset. VinDr-CXR consists of 14 abnormalities classes and contains 15,000 training and 3,000 testing chest radiograph images. The training images are annotated by 17 experienced radiologists where at least three annotators are randomly assigned to annotate each image. Figures 1f to 1j show examples of the annotations with high bounding box and class label disagreement by the annotators. Furthermore, the testing images are annotated with the consensus of five expert radiologists through a two-stage labelling process and the test set annotations are not publicly available. The predictions are submitted to the VinBigData Kaggle [12] to obtain mean average precision (AP) at 0.4 IoU threshold on the private test set.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z.Q. Tan et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8c6eea403ff85663083e5499a0d405c493c5c56501ad5c2f240e7501f8dcd259.jpg", + "table_caption": [ + "Table 1: AP-4 results on VinDr-CXR [36] dataset obtained from the private score of VinBigData Kaggle competition [12]. Crowd R-CNN can only integrate FRCNN, hence there are no results reported for the two other entries corresponding to Crowd R-CNN." + ], + "table_footnote": [], + "table_body": "
MethodTest AP-4
YOLOv7FRCNNEVA
NA17.417.27.8
MV13.916.38.2
Crowd R-CNN [18]-16.7-
WBF-EARL [27]16.417.08.4
BDC (ours)19.217.98.9
", + "bbox": [ + 354, + 198, + 648, + 287 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/27d6d5e0ce6ea3eb674e5b640ea66d0fb302820d3a10fd5aa9d1791cd686ee87.jpg", + "image_caption": [ + "(a) NA" + ], + "image_footnote": [], + "bbox": [ + 217, + 301, + 328, + 417 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/83cedb638eeeab9d7d64990fdfdc8764761bfecc381e1887ff342da22235edd2.jpg", + "image_caption": [ + "(b) MV" + ], + "image_footnote": [], + "bbox": [ + 328, + 301, + 439, + 417 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/291058eff251e95d3af1d875e90d621628439a2b50e9c3bf8ae29686673dc5d7.jpg", + "image_caption": [ + "(c) WBF-EARL", + "Fig. 3: Comparison of aggregated labels from different methods on the VinDr-CXR [36]. For WBF-EARL [27], the number beside the class label is the annotators' level of agreement while for Crowd R-CNN [18] and BDC, the number indicates the class probability. For NA, the colours represent the different annotators." + ], + "image_footnote": [], + "bbox": [ + 442, + 301, + 553, + 417 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f90afadb7091e5e1cd925559bdc367793f7da1b9ae43343c7b16bda3ddec5519.jpg", + "image_caption": [ + "(d) Crowd R-CNN" + ], + "image_footnote": [], + "bbox": [ + 553, + 301, + 666, + 417 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d03061ac0e82e76322ef74896bf685d31fe6288b1f63cb222fcefe2a401f31a3.jpg", + "image_caption": [ + "(e) BDC (ours)" + ], + "image_footnote": [], + "bbox": [ + 666, + 301, + 777, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Disaster Response Dataset. We further use the disaster response dataset [19] where the task is to detect undamaged and damaged buildings on satellite images taken after hurricane Maria in 2017. The dataset consisting of 464 satellite images of Dominica is annotated by 13 semi-professional paid crowd members to ensure high labelling quality. Despite this effort, labels from different crowd members vary significantly (Figures 1k to 1o). Consequently, the dataset does not contain ground truth labels. Thus, its qualitative results are presented in Appendix E.", + "bbox": [ + 212, + 532, + 787, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative Results. Table 1 presents the AP.4 metric obtained from the VinBigData Kaggle competition and shows that the proposed BDC leads to higher AP.4 than other methods across different OD models on the VinDr-CXR test set. As Crowd R-CNN can only integrate FRCNN, there are no results reported for the two other entries corresponding to Crowd R-CNN.", + "bbox": [ + 212, + 662, + 787, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative Results. We visualise the results of the aggregated annotations from each method on the VinDr-CXR dataset in Figure 3. Since the true ground truth of the datasets is not provided, we use 'MV' as the best approximation of the ground truth for this qualitative assessment and observe that BDC produces more accurate aggregation compared to the other methods.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bayesian Detector Combination", + "bbox": [ + 517, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Synthetic Datasets", + "text_level": 1, + "bbox": [ + 215, + 146, + 416, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To alleviate the limitations in the available real-world crowdsourced datasets including the lack of ground truth labels, we synthesise several synthetic datasets with crowdsourced annotations to enable the systematic evaluation of the proposed method. This makes available the ground truth annotations to allow us to obtain detailed test evaluation metrics (e.g., AP $^{5}$ , AP $^{75}$ , AP $^{5::95}$ ) and verify the performance of the proposed method by evaluating the aggregated crowdsourced annotations with the true ground truth annotations of the training dataset (indicated as 'Train Aggregation AP' in Tables 2 to 4). Two common object detection datasets are used to construct the synthetic datasets: (1) Pascal Visual Object Classes (VOC) 2007 [13], a 20-class object detection dataset which consists of 5,011 and 4,952 images for training and testing, respectively; (2) Microsoft Common Objects in Context (COCO) 2017 [29], a large-scale 80-class object detection dataset consisting of 118,287 training and 5,000 validation images.", + "bbox": [ + 212, + 172, + 787, + 369 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Synthetic settings. Four synthetic settings are investigated in this work: (1) VOC-FULL and (2) COCO-FULL where 10 annotators with average skill levels annotate the whole datasets; (3) VOC-MIX where the first 5 annotators are experts while the remaining 20 annotators have an average skill level. All annotators annotate the whole dataset since Pascal VOC is a small dataset; and (4) COCO-MIX where 20 experts, 550 annotators with average skill level and 330 annotators with poor skill level work together to annotate the MS COCO dataset. Additionally, the last 100 are random annotators, i.e., the annotated class labels and bounding boxes are drawn from a uniform distribution. Since MS COCO is a large-scale dataset, we set that each annotator annotates a small subset of images, i.e., the coverage for the expert, average, poor and random annotators are $5\\%$ , $1\\%$ , $0.5\\%$ and $1\\%$ , respectively. Note that all annotators in these settings are simulated using deep learning models (See Appendix F for further discussions of the synthesis method and pseudocode). Additional experiments where we directly used the object detectors' predictions as synthetic crowdsourced annotations can be found in Appendix H.", + "bbox": [ + 212, + 385, + 787, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Results. Tables 2 to 4 report the experiment results of YOLOv7, FRCNN and EVA trained with different aggregation methods on all synthetic settings, respectively. Additionally, we trained each model with ground truth annotations of each dataset (denoted by 'GT') to obtain reference optimal model performance. The aggregation APs of the train dataset for NA and WBF-EARL are low because NA directly uses all the crowdsourced annotations while WBF outputs multiple bounding boxes per class per object instance, leading to a much lower precision (see Figure 4 for examples). We do not include performance from Crowd R-CNN for the COCO-MIX setting as its training time increases linearly with $K$ which means that the training time is prohibitively high for this dataset where $K = 1,000$ . In comparison, the extra computational cost introduced by the proposed BDC is small when compared to the much heavier training cost of the OD backbone (see Section 4.4 for an analysis of the computational cost).", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z.Q. Tan et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9c60d4a54f0850ae74defb8c521557560e8220123b9e4167bb380c0603f6fcb7.jpg", + "table_caption": [ + "Table 2: Test and train aggregation AP results for YOLOv7 model trained with different annotation aggregation methods on synthesised and ground truth datasets." + ], + "table_footnote": [], + "table_body": "
DatasetMethodTest APTrain Aggregation AP
AP.5AP.75AP.5:95AP.5AP.75AP.5:95
VOC-FULLNA59.745.340.710.29.99.9
MV72.252.347.362.048.548.5
WBF-EARL [27]69.352.345.722.316.414.5
BDC (ours)84.963.554.895.080.666.3
VOC-MIXNA46.127.627.44.44.34.3
MV77.161.053.775.770.668.6
WBF-EARL [27]61.047.240.927.924.920.0
BDC (ours)86.269.058.897.694.977.2
VOCGT87.774.065.3100100100
COCO-FULLNA53.441.437.48.68.28.2
MV61.946.442.752.544.443.3
WBF-EARL [27]55.642.538.914.68.98.6
BDC (ours)65.049.644.583.666.156.9
COCO-MIXNA53.441.237.58.57.98.0
MV60.946.442.445.842.341.1
WBF-EARL [27]56.242.639.112.17.07.0
BDC (ours)64.849.344.480.562.954.3
COCOGT67.552.848.6100100100
", + "bbox": [ + 267, + 184, + 736, + 429 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tables 2 to 4 demonstrate the superior performance of BDC compared to other methods from simpler synthetic settings to more complex synthetic crowdsourced annotations, i.e. from VOC-FULL to COCO-MIX. In addition, BDC achieves the highest train aggregation AP compared to other methods across all settings.", + "bbox": [ + 215, + 458, + 784, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 4 visualises the results of the aggregated annotations from each method on the VOC-MIX synthetic dataset. WBF performs box fusion on a per-class basis and results in significantly more boxes compared to other methods (excluding NA) which contributed to the low train aggregation AP and noisy labels during model training. On the other hand, although both Crowd R-CNN and BDC output class probabilities, the probability distribution of Crowd R-CNN is heavily skewed which is similar to hard labels, while BDC outputs the learnt class probabilities and can benefit from these soft labels during model training [34]. Besides, BDC can cope with failure cases of its aggregator as shown by the lower class probability of incorrect aggregation. Additional qualitative results can be found in Appendix I.", + "bbox": [ + 215, + 534, + 785, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Scalability and Robustness of BDC", + "text_level": 1, + "bbox": [ + 215, + 723, + 550, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We analyse the scalability and robustness of the aggregation methods using the Pascal VOC dataset. YOLOv7 is selected as the OD because it is lightweight and fast to train. Figure 5a demonstrates the effect of varying numbers of annotators, $K$ , where all annotations are synthesised using annotators of average skill level. Note that training YOLOv7 with the NA method results in a GPU out-of-memory error when $K \\geq 100$ . When the number of annotators increases, the", + "bbox": [ + 215, + 750, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bayesian Detector Combination", + "bbox": [ + 517, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b354d5653d9b37de644661a35a3f573e9bc31707fd8100abc6f2120d71b04116.jpg", + "table_caption": [ + "Table 3: Test and train aggregation AP results for FRCNN model trained with different annotation aggregation methods on synthesised and ground truth datasets. Crowd R-CNN has a training time that increases linearly with $K$ , hence there is no result reported for COCO-MIX dataset with $K = 1,000$ ." + ], + "table_footnote": [], + "table_body": "
DatasetMethodTest APTrain Aggregation AP
AP·5AP·75AP·5:95AP·5AP·75AP·5:95
VOC-FULLNA47.411.319.410.29.99.9
MV79.753.548.862.048.548.5
Crowd R-CNN [18]71.633.930.158.343.031.5
WBF-EARL [27]71.837.738.722.316.414.5
BDC (ours)83.756.351.095.080.466.3
VOC-MIXNA34.14.412.24.44.34.3
MV82.057.151.975.770.668.6
Crowd R-CNN [18]70.231.229.360.644.231.0
WBF-EARL [27]69.836.537.427.924.920.0
BDC (ours)84.759.352.997.694.676.9
VOCGT84.662.155.2100100100
COCO-FULLNA39.712.017.78.68.28.2
MV55.635.733.452.544.443.3
Crowd R-CNN [18]48.527.124.249.536.233.7
WBF-EARL [27]51.928.028.414.68.98.6
BDC (ours)56.635.333.583.366.056.7
COCO-MIXNA43.917.521.48.57.98.0
MV54.834.532.945.842.341.1
Crowd R-CNN [18]------
WBF-EARL [27]51.928.028.312.17.07.0
BDC (ours)56.134.633.080.362.754.1
COCOGT58.038.535.7100100100
", + "bbox": [ + 256, + 212, + 745, + 510 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "performance of YOLOv7 increases for both BDC and MV while the performance of both NA and WBF-EARL decreases when $K > 50$ . In addition, BDC has an average AP.5:95 improvement of 4.35 over MV.", + "bbox": [ + 212, + 537, + 785, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also investigate the effect of changing the percentage of noisy annotators while setting $K$ to 25. Here, noisy annotators are defined as annotators with a poor skill level while the other annotators are annotators with expert skill level. As illustrated in Figure 5b, we found that AP $^{5:95}$ of all aggregation methods reduce when we increase the percentage of noisy annotators. However, MV drops at a faster rate compared to the proposed method when the noisy annotator percentage is over $50\\%$ . We observe that BDC remains robust even when $100\\%$ of the annotators have low skill levels and the AP $^{5:95}$ is 18.5 higher than that obtained with MV. The consistent performance of BDC (up to $80\\%$ noise) can be attributed to the ability of BDC to correctly capture the ability of each annotator and assign proper weights when aggregating the annotations (See Appendix J for an analysis of the posterior distributions learnt in the aggregators).", + "bbox": [ + 212, + 583, + 787, + 765 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Computational Cost", + "text_level": 1, + "bbox": [ + 215, + 785, + 429, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 5 presents the wall time of each annotation aggregation method spent to pre-process the annotations and train the Faster R-CNN (FRCNN) [41] model for", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z.Q. Tan et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/dc0a73f645815cbf3a0953ba5622487b28ea06889d75fdac9062e1d065727273.jpg", + "table_caption": [ + "Table 4: Test and train aggregation AP results for EVA model trained with different annotation aggregation methods on synthesised and ground truth datasets." + ], + "table_footnote": [], + "table_body": "
DatasetMethodTest APTrain Aggregation AP
AP.5AP.75AP.5:95AP.5AP.75AP.5:95
VOC-FULLNA90.980.969.710.29.99.9
MV91.082.271.262.048.548.5
WBF-EARL [27]93.083.571.122.316.414.5
BDC (ours)94.285.671.795.381.266.7
VOC-MIXNA86.665.757.54.44.34.3
MV91.783.072.975.770.668.6
WBF-EARL [27]91.382.268.727.924.920.0
BDC (ours)94.986.573.898.895.077.3
VOCGT94.988.178.0100100100
COCO-FULLNA71.855.750.68.68.28.2
MV74.861.054.452.544.443.3
WBF-EARL [27]74.758.451.814.68.98.6
BDC (ours)78.062.755.183.365.155.3
COCO-MIXNA73.357.251.78.57.98.0
MV75.060.654.245.842.341.1
WBF-EARL [27]75.259.352.912.17.07.0
BDC (ours)77.162.554.879.962.253.2
COCOGT78.864.958.8100100100
", + "bbox": [ + 267, + 184, + 736, + 429 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "one epoch on the VOC-MIX synthetic dataset. The wall time is computed with a desktop workstation with an i9-10900X CPU and an RTX 3090 GPU. As shown in Table 5, MV, Crowd R-CNN and WBF-EARL pre-process the annotations before model training while NA and the proposed BDC require no pre-processing step. Crowd R-CNN has the longest training time which increases linearly with the number of annotators, $K$ . On the other hand, the proposed BDC performs the annotations-predictions matching step and the updating steps of the aggregators' parameters every training epoch and therefore has a slightly longer training time compared to NA, MV and WBF-EARL. However, this increase in training time is small when compared to the much higher training cost of the OD", + "bbox": [ + 212, + 462, + 787, + 613 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/093aae14ab363fa4e3b3131dd22cce74bb1972ca87d19eb8824a7afecef186e8.jpg", + "image_caption": [ + "(a) NA" + ], + "image_footnote": [], + "bbox": [ + 217, + 643, + 312, + 744 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/def6a5bc2d50d1d72a35dfb1849b5f556020f7051bc8e9d8a28769154e8f2b06.jpg", + "image_caption": [ + "(b) MV", + "(c) WBF-EARL" + ], + "image_footnote": [], + "bbox": [ + 313, + 643, + 406, + 744 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e6cc64f3f1cbea29b9cfa6a1a4ae3daa9a1cd47df2c99922548f9cc98e7c4836.jpg", + "image_caption": [ + "Fig. 4: Comparison of aggregated labels from different methods on the VOC-MIX synthetic dataset. For WBF-EARL [27], the number beside the class label is the annotators' level of agreement while for Crowd R-CNN [18] and BDC, the number indicates the class probability. For NA, the colours represent the different annotators." + ], + "image_footnote": [], + "bbox": [ + 408, + 643, + 500, + 744 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1c3e1b0d2274b0cd22b2237aff5ca34a32310e431309087b4cf609ac93475bc9.jpg", + "image_caption": [ + "(d) Crowd R-CNN" + ], + "image_footnote": [], + "bbox": [ + 501, + 643, + 596, + 744 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2f71fa5cd76551269c73ecefd75a9b3fb46fef874ecfd05e22d63ed71856b78a.jpg", + "image_caption": [ + "(e) BDC (ours)" + ], + "image_footnote": [], + "bbox": [ + 598, + 643, + 691, + 744 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/225f8100c52b5c363bd24eda2f5c284101bf1679f99ed6103d25bf8ecd789c77.jpg", + "image_caption": [ + "(f) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 694, + 643, + 785, + 744 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bayesian Detector Combination", + "bbox": [ + 517, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0364a45c8b39f31bd95e9f88021654048f93d6fc040889b4a71f36328d5aa8ba.jpg", + "image_caption": [ + "(a) Scalability of BDC" + ], + "image_footnote": [], + "bbox": [ + 238, + 143, + 500, + 309 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e5673ad679f6f627753d37a472c6491c794e5b0a36179bf78a139e8a5537a0f5.jpg", + "image_caption": [ + "(b) Robustness of BDC", + "Fig. 5: (a) Effect of the number of annotators, $K$ , and (b) effect of varying percentage of noisy annotators (with $K = 25$ ) on the test AP $^{5:95}$ of YOLOv7. Results for the 'NA' method are unavailable when $K \\geq 100$ due to GPU out-of-memory error." + ], + "image_footnote": [], + "bbox": [ + 511, + 143, + 769, + 310 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/f2912f150e357c736233d54852b23998c137b8dd7602bb7aa94aaefeb0fb615c.jpg", + "table_caption": [ + "Table 5: Annotations pre-processing time and FRCNN model training time per epoch of different annotation aggregation methods on the VOC-MIX dataset. The wall time is computed on a desktop workstation with an i9-10900X CPU and an RTX 3090 GPU." + ], + "table_footnote": [], + "table_body": "
MethodPre-processing time (s)Training time (s)
NA0.0537.0
MV19.4505.4
Crowd R-CNN [18]15.61175.5
WBF-EARL [27]4.9518.5
BDC (ours)0.0662.9
", + "bbox": [ + 297, + 446, + 705, + 527 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "module (618.6s OD training time out of 662.9s). Note that the model inference time is unaffected by the type of annotation aggregation method.", + "bbox": [ + 212, + 556, + 787, + 588 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 609, + 359, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We first presented a general annotation aggregation and object detection framework, named Bayesian Detector Combination (BDC), for learning a robust object detector from noisy crowdsourced annotations. BDC infers the consensus bounding boxes and corresponding soft labels without the need for ground truth annotations. BDC is model-agnostic and we demonstrated its superior performance by integrating it with three popular object detector backbones, namely YOLOv7, FRCNN, and EVA. Furthermore, we constructed four synthetic datasets to simulate different crowdsourcing scenarios involving annotators of various skill levels, enabling the systematic evaluation of the proposed BDC. We demonstrated the improved performance from BDC on real-world datasets and all the synthetic datasets when compared to other object detectors designed for crowdsourced annotations. We showed that BDC can scale against varying numbers of annotators and is robust against annotators with low annotation reliability.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z.Q. Tan et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. Z.Q. T. acknowledges the support of an Industrial Cooperative Awards in Science and Engineering (ICASE) studentship from the Engineering and Physical Sciences Research Council (EPSRC) for this work. G. C. acknowledges the support of the Engineering and Physical Sciences Research Council (EPSRC) through grant EP/Y018036/1 and the Australian Research Council (ARC) through grant FT190100525. The authors would like to thank the Satellite Applications Catapult for the provision of the disaster response dataset.", + "bbox": [ + 212, + 146, + 787, + 265 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 295, + 321, + 311 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Albarqouni, S., Baur, C., Achilles, F., Belagiannis, V., Demirci, S., Navab, N.: AggNet: Deep learning from crowds for mitosis detection in breast cancer histology images. IEEE Trans. Med. Imag. 35(5), 1313-1321 (May 2016)", + "2. Baumgartner, C.F., Tezcan, K.C., Chaitanya, K., Hötter, A.M., Muehlematter, U.J., Schawkat, K., Becker, A.S., Donati, O., Konukoglu, E.: PHiSeg: Capturing uncertainty in medical image segmentation. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 119-127. Shenzhen, China (2019)", + "3. Bernard, J.M., Smith, A.F.M.: Bayesian Theory, chap. 3, pp. 136-138. Wiley, West Sussex, UK (1993)", + "4. Bernhardt, M., Castro, D.C., Tanno, R., Schwaighofer, A., Tezcan, K.C., Monteiro, M., Bannur, S., Lungren, M.P., Nori, A., Glocker, B., Alvarez-Valle, J., Oktay, O.: Active label cleaning for improved dataset quality under resource constraints. Nature Commun. 13(1), 1161 (Mar 2022)", + "5. Budd, S., Day, T., Simpson, J., Lloyd, K., Matthew, J., Skelton, E., Razavi, R., Kainz, B.: Can non-specialists provide high quality gold standard labels in challenging modalities? In: Med. Imag. Comput. Comput. Assist. Interv. pp. 251-262. Strasbourg, France (2021)", + "6. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229 (2020)", + "7. Chen, Z., Wang, H., Sun, H., Chen, P., Han, T., Liu, X., Yang, J.: Structured probabilistic end-to-end learning from crowds. In: Int. Joint Conf. Artif. Intell. pp. 1512-1518 (2020)", + "8. Dawid, A.P., Skene, A.M.: Maximum likelihood estimation of observer error-rates using the EM algorithm. J. Roy. Statist. Soc. Ser. C (Appl. Statist.) 28(1), 20-28 (1979)", + "9. DungNB, Nguyen, H.Q., Elliott, J., KeepLearning, Nhan, N.T., Culliton, P.: VinBigData chest x-ray abnormalities detection (2020), https://kaggle.com/competitions/ vinbigdata-chest-xray-abnormalities-detection", + "10. Everingham, M., Luc Van Gool, C.K.I.W., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2007 (VOC2007). Int. J. Comput. Vis. 88, 303-338 (Jun 2010)", + "11. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA: Exploring the limits of masked visual representation learning at scale. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19358-19369. Vancouver, Canada (2023)" + ], + "bbox": [ + 217, + 330, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bayesian Detector Combination", + "bbox": [ + 517, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Goh, H.W., Tkachenko, U., Mueller, J.: CROWDLAB: Supervised learning to infer consensus labels and quality scores for data with multiple annotators (2022), arXiv:2210.06812", + "13. Guan, M.Y., Gulshan, V., Dai, A.M., Hinton, G.E.: Who said what: Modeling individual labelers improves classification. In: AAAI Conf. Artif. Intell. pp. 3109-3118. San Francisco, CA, USA (2017)", + "14. Hu, Y., Meina, S.: Crowd R-CNN: An object detection model utilizing crowdsourced labels. In: Int. Conf. Vis. Image Sig. Process. pp. 1-7. Bangkok, Thailand (2020)", + "15. Isupova, O., Li, Y., Kuzin, D., Roberts, S.J., Willis, K.J., Reece, S.: BCCNet: Bayesian classifier combination neural network. In: Neural Inf. Process. Syst. Workshop Mach. Learn. Develop. Montréal, Canada (2018)", + "16. Jensen, M.H., Jorgensen, D.R., Jalaboi, R., Hansen, M.E., Olsen, M.A.: Improving uncertainty estimation in convolutional neural networks using inter-rater agreement. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 540-548. Shenzhen, China (2019)", + "17. Ji, W., Yu, S., Wu, J., Ma, K., Bian, C., Bi, Q., Li, J., Liu, H., Cheng, L., Zheng, Y.: Learning calibrated medical image segmentation via multi-rater agreement modeling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12336-12346 (2021)", + "18. Jungo, A., Meier, R., Ermis, E., Blatti-Moreno, M., Herrmann, E., Wiest, R., Reyes, M.: On the effect of inter-observer variability for a reliable estimation of uncertainty of medical image segmentation. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 682-690. Granada, Spain (2018)", + "19. Kim, H., Ghahramani, Z.: Bayesian classifier combination. In: Int. Conf. Artif. Intell. Statist. pp. 619-627. La Palma, Canary Islands (2012)", + "20. Kohl, S.A.A., Romera-Paredes, B., Meyer, C., Fauw, J.D., Ledsam, J.R., Maier-Hein, K.H., Eslami, S.M.A., Rezende, D.J., Ronneberger, O.: A probabilistic u-net for segmentation of ambiguous images. In: Adv. Neural Inform. Process. Syst. pp. 6965–6975. Montréal, Canada (2018)", + "21. Lazarus, E., Mainiero, M.B., Schepps, B., Koelliker, S.L., Livingston, L.S.: BI-RADS lexicon for US and mammography: interobserver variability and positive predictive value. Radiology 239(2), 385-391 (Mar 2006)", + "22. Le, K., Tran, T., Pham, H., Nguyen Trung, H., Le, T., Nguyen, H.Q.: Learning from multiple expert annotators for enhancing anomaly detection in medical image analysis. IEEE Access 11, 14105-14114 (2023)", + "23. LeCun, Y., Cortes, C., Burges, C.: MNIST handwritten digit database (2010), http://yann.lecun.com/exdb/mnist", + "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Zurich, Switzerland (2014)", + "25. Maier-Hein, L., Mersmann, S., Kondermann, D., Stock, C., Kenngott, H.G., Sanchez, A., Wagner, M., Preukschas, A., Wekerle, A.L., Helfert, S., Bodenstedt, S., Speidel, S.: Crowdsourcing for reference correspondence generation in endoscopic images. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 349-356. Boston, MA, USA (2014)", + "26. Müller, R., Kornblith, S., Hinton, G.: When does label smoothing help? In: Adv. Neural Inform. Process. Syst. pp. 4694-4703. Vancouver, Canada (2019)", + "27. Murphy, K.: Conjugate bayesian analysis of the gaussian distribution (2007), https://www.cs.ubc.ca/~murphyk/Papers/bayesGauss.pdf" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z.Q. Tan et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Nguyen, H.Q., Lam, K., Le, L.T., Pham, H.H., Tran, D.Q., Nguyen, D.B., Le, D.D., Pham, C.M., Tong, H.T.T., Dinh, D.H., Do, C.D., Doan, L.T., Nguyen, C.N., Nguyen, B.T., Nguyen, Q.V., Hoang, A.D., Phan, H.N., Nguyen, A.T., Ho, P.H., Ngo, D.T., Nguyen, N.T., Nguyen, N.T., Dao, M., Vu, V.: VinDr-CXR: An open dataset of chest x-rays with radiologist's annotations. Sci. Data 9(1), 429 (Jul 2022)", + "29. Nowruzi, F.E., Kapoor, P., Kolhatkar, D., Hassanat, F., Laganiere, R., Rebut, J.: How much real data do we actually need: Analyzing object detection performance using synthetic and real data. In: Int. Conf. Mach. Learn. Workshop AI Auton. Driving. Long Beach, CA, USA (2019)", + "30. Raykar, V.C., Yu, S., Zhao, L.H., Valadez, G.H., Florin, C., Bogoni, L., Moy, L., Raykar, C., Valadez, G.H.: Learning from crowds. J. Mach. Learn. Res. 11(43), 1297-1322 (Aug 2010)", + "31. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: Adv. Neural Inform. Process. Syst. pp. 91-99. Montréal, Canada (2015)", + "32. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 658-666. Long Beach, CA, USA (2019)", + "33. Rodrigues, F., Pereira, F.C.: Deep learning from crowds. In: AAAI Conf. Artif. Intell. pp. 1611-1618. New Orleans, LA, USA (2018)", + "34. Sheng, V.S., Provost, F., Ipeirotis, P.G.: Get another label? improving data quality and data mining using multiple, noisy labelers. In: ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. pp. 614-622. Las Vegas, NV, USA (2008)", + "35. Solovyev, R., Wang, W., Gabruseva, T.: Weighted boxes fusion: Ensembling boxes from different object detection models. Image Vis. Comput. 107, 104-117 (Mar 2021)", + "36. Sudre, C.H., Anson, B.G., Ingala, S., Lane, C.D., Jimenez, D., Haider, L., Varsavsky, T., Tanno, R., Smith, L., Ourselin, S., Jäger, R.H., Cardoso, M.J.: Let's agree to disagree: Learning highly debatable multirater labelling. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 665-673. Shenzhen, China (2019)", + "37. Venanzi, M., Guiver, J., Kazai, G., Kohli, P., Shokouhi, M.: Community-based bayesian aggregation models for crowdsourcing. In: Int. Conf. World Wide Web. pp. 155-164. Seoul, Korea (2014)", + "38. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7464-7475. Vancouver, Canada (2023)", + "39. Watadani, T., Sakai, F., Johkoh, T., Noma, S., Akira, M., Fujimoto, K., Bankier, A.A., Lee, K.S., Müller, N.L., Song, J.W., Park, J.S., Lynch, D.A., Hansell, D.M., Remy-Jardin, M., Franquet, T., Sugiyama, Y.: Interobserver variability in the CT assessment of honeycombing in the lungs. Radiology 266(3), 936-944 (Dec 2012)", + "40. Wei, H., Xie, R., Feng, L., Han, B., An, B.: Deep learning from multiple noisy annotators as a union. IEEE Trans. Neural Netw. Learn. Syst. pp. 1-11 (Apr 2022)", + "41. Wei, J., Zhu, Z., Luo, T., Amid, E., Kumar, A., Liu, Y.: To aggregate or not? learning with separate noisy labels. In: ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. pp. 2523-2535. Long Beach, CA, USA (2023)", + "42. Whitehill, J., Wu, T.f., Bergsma, J., Movellan, J., Ruvolo, P.: Whose vote should count more: Optimal integration of labels from labelers of unknown expertise. In: Adv. Neural Inform. Process. Syst. pp. 2035-2043. Vancouver, Canada (2009)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bayesian Detector Combination", + "bbox": [ + 517, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Wu, J., Fang, H., Yang, Y., Liu, Y., Gao, J., Duan, L., Yang, W., Xu, Y.: Multi-rater prism: Learning self-calibrated medical image segmentation from multiple raters (2022), arXiv:2212.00601", + "44. Yu, S., Zhou, H.Y., Ma, K., Bian, C., Chu, C., Liu, H., Zheng, Y.: Difficulty-aware glaucoma classification with multi-rater consensus modeling. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 741-750 (2020)", + "45. Zhang, L., Tanno, R., Xu, M.C., Jin, C., Jacob, J., Ciccarelli, O., Barkhof, F., Alexander, D.C.: Disentangling human error from the ground truth in segmentation of medical images. In: Adv. Neural Inform. Process. Syst. pp. 15750-15762 (2020)", + "46. Zhu, X., Vondrick, C., Ramanan, D., Fowlkes, C.: Do we need more training data or better models for object detection? In: Brit. Mach. Vis. Conf. Surrey, UK (2012)" + ], + "bbox": [ + 215, + 146, + 785, + 300 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z.Q. Tan et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_model.json b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8010f84d36bac47dd292a156174fdc33a33708b9 --- /dev/null +++ b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_model.json @@ -0,0 +1,2656 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.258, + 0.141, + 0.747, + 0.185 + ], + "angle": 0, + "content": "Bayesian Detector Combination for Object Detection with Crowdsourced Annotations" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.213, + 0.772, + 0.245 + ], + "angle": 0, + "content": "Zhi Qin \\(\\mathrm{Tan}^{1}\\), Olga Isupova\\(^{2}\\), Gustavo Carneiro\\(^{1}\\), Xiatian Zhu\\(^{1}\\), and Yunpeng Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.255, + 0.712, + 0.284 + ], + "angle": 0, + "content": "1 University of Surrey, Surrey, UK {z.tan,g.carneiro, xiatian.zhu, yunpeng.li}@surrey.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.284, + 0.621, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) University of Oxford, Oxford, UK olga.isupova@ouce.ox.ac.uk" + }, + { + "type": "list", + "bbox": [ + 0.292, + 0.255, + 0.712, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.353, + 0.744, + 0.631 + ], + "angle": 0, + "content": "Abstract. Acquiring fine-grained object detection annotations in unconstrained images is time-consuming, expensive, and prone to noise, especially in crowdsourcing scenarios. Most prior object detection methods assume accurate annotations; A few recent works have studied object detection with noisy crowdsourced annotations, with evaluation on distinct synthetic crowdsourced datasets of varying setups under artificial assumptions. To address these algorithmic limitations and evaluation inconsistency, we first propose a novel Bayesian Detector Combination (BDC) framework to more effectively train object detectors with noisy crowdsourced annotations, with the unique ability of automatically inferring the annotators' label qualities. Unlike previous approaches, BDC is model-agnostic, requires no prior knowledge of the annotators' skill level, and seamlessly integrates with existing object detection models. Due to the scarcity of real-world crowdsourced datasets, we introduce large synthetic datasets by simulating varying crowdsourcing scenarios. This allows consistent evaluation of different models at scale. Extensive experiments on both real and synthetic crowdsourced datasets show that BDC outperforms existing state-of-the-art methods, demonstrating its superiority in leveraging crowdsourced data for object detection. Our code and data are available at: https://github.com/zhiqin1998/bdc." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.644, + 0.741, + 0.672 + ], + "angle": 0, + "content": "Keywords: Bayesian Detector Combination \\(\\cdot\\) Object detection \\(\\cdot\\) Crowdsourced annotations" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.7, + 0.377, + 0.716 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Despite significant advancements in the development of object detection with natural images [9, 14, 41, 53], existing methods typically assume the availability of large, accurately annotated datasets [37, 61]. Nonetheless, this assumption does not often hold true in practice. For example, accurate labels are often expensive to acquire in many applications. A large number of experts with varying professional skills can be involved to crowdsourced annotations of images sampled with replacement from a large dataset [8, 33, 50]. Usually, the annotations" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z.Q. Tan et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.146, + 0.331, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.207, + 0.318, + 0.218 + ], + "angle": 0, + "content": "(a) ID: 70282" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.147, + 0.444, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.207, + 0.434, + 0.218 + ], + "angle": 0, + "content": "(b) ID: 310072" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.145, + 0.557, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.207, + 0.545, + 0.218 + ], + "angle": 0, + "content": "(c) ID: 516471" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.145, + 0.669, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.207, + 0.659, + 0.218 + ], + "angle": 0, + "content": "(d) ID: 326247" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.145, + 0.783, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.207, + 0.771, + 0.218 + ], + "angle": 0, + "content": "(e) ID: 416149" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.22, + 0.331, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.284, + 0.286, + 0.296 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.22, + 0.444, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.285, + 0.399, + 0.296 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.219, + 0.557, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.285, + 0.512, + 0.296 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.22, + 0.668, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.285, + 0.622, + 0.296 + ], + "angle": 0, + "content": "(i)" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.22, + 0.783, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.285, + 0.735, + 0.296 + ], + "angle": 0, + "content": "(j)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.298, + 0.331, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.357, + 0.287, + 0.369 + ], + "angle": 0, + "content": "(k)" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.298, + 0.444, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.357, + 0.398, + 0.369 + ], + "angle": 0, + "content": "(1)" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.297, + 0.556, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.357, + 0.514, + 0.369 + ], + "angle": 0, + "content": "(m)" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.297, + 0.668, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.357, + 0.624, + 0.369 + ], + "angle": 0, + "content": "(n)" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.297, + 0.783, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.357, + 0.736, + 0.369 + ], + "angle": 0, + "content": "(o)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.38, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 1: Examples of ambiguous cases with noisy or incorrect annotations on (a - e) MS COCO [29], (f - j) VinDr-CXR [36] and (k - o) disaster response dataset [19] to identify damaged and undamaged buildings with class names 'd' and 'u', respectively. (a) Mislabelling one object with two bounding boxes. (b - d) Duplicate annotations of visually similar objects (e.g., car/truck, donut/cake and fork/spoon). (d)(e) Incorrect annotation of object classes (e.g., cake labelled as a bowl, only annotating one of the cake decorations as 'person'). (f - o) Annotations from different annotators (represented by the different colours) of the VinDr-CXR and disaster response datasets show significant disagreement. (Details of the datasets are given in Section 4.1.)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.684 + ], + "angle": 0, + "content": "from multiple annotators are unified and cleaned as the last step in annotation collection. However, disagreements arise when annotating complex images, making it challenging even for experienced expert annotators to achieve unanimity due to high interobserver variability [26,54] and thus, rendering the role of an annotation arbitrator nearly impossible (see Figure 1). As a result, multiple noisy annotations per image are obtained. Such a setting for annotation crowdsourcing is often called the multi-rater problem. The majority of methods on multi-rater learning have been developed for image classification [20,50,59] and segmentation [4,21,22,24,58,60], while we consider object detection here." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Several previous methods that exploit crowdsourced annotations have been proposed to improve object detection performance [18,27]. A common approach estimates the ground truth from noisy annotations through majority voting [46] which implicitly assumes that all annotators are equally accurate in their annotations. Alternatively, in [27], the weight of each annotation is predicted based on the annotators' proficiency levels which are assumed to be known a priori; however, this information is difficult to obtain in practice. On the other hand, in [18], their proposed method is designed specifically for Faster R-CNN (FRCNN) [41] and cannot be generalised to other object detectors. Moreover, these methods have been developed and evaluated on distinct, private synthetic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Bayesian Detector Combination" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.77, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Fig. 2: Overall architecture of the proposed BDC. The updating of the aggregator's parameter and the object detector's parameters is repeated iteratively until convergence. Best viewed in colour." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.401 + ], + "angle": 0, + "content": "datasets constructed using different methods. Thus, we cannot directly compare their reported experiment results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.787, + 0.552 + ], + "angle": 0, + "content": "To address these algorithmic and evaluation limitations, in this work, we first propose a model-agnostic annotation aggregation and object detector training pipeline, named Bayesian Detector Combination (BDC). BDC aggregates noisy bounding boxes and class labels by modelling the annotation reliability of each crowd member for both elements (Figure 2), without requiring any additional knowledge of the ground truth annotations or their quality. Furthermore, due to the limited access to the ground truth of real crowdsourced datasets, we synthesise multiple datasets simulating various crowdsourcing scenarios for systematic evaluation of the previous works and the proposed method. The main contributions of our work include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.623 + ], + "angle": 0, + "content": "- We propose a model-agnostic Bayesian Detector Combination (BDC) framework that simultaneously infers the annotation quality of each crowd member and infers consensus bounding boxes and soft class labels from noisy, crowdsourced annotations to improve model performance." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.623, + 0.785, + 0.667 + ], + "angle": 0, + "content": "- We implement and systematically evaluate the previous methods on real-world and synthetic datasets with crowdsourced annotations which simulate varying crowdsourcing scenarios." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.668, + 0.785, + 0.727 + ], + "angle": 0, + "content": "- Extensive experiments conducted demonstrate the superior performance of BDC compared with prior alternatives in key metrics such as the average precision, the scalability with the increasing number of annotators, and the robustness against noisy annotators." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.75, + 0.388, + 0.765 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Classification with Crowdsourced Annotations. Learning with crowdsourced annotations for classification tasks has been studied in various works. Sheng et al. [46] adopted a simple approach to aggregate multiple labels via majority voting (MV) where the mode class is selected as the ground truth. Several" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z.Q. Tan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "works [11,39,57] modelled the sensitivity and specificity of different annotators via confusion matrices obtained through the Expectation-Maximisation algorithm or a Bayesian approach. Kim and Ghahramani [23] introduced a general framework for Bayesian classifier combination (BCC) to model the relationship between the unknown true label and the annotated class label which is extended by Venanzi et al. [51] with the concept of annotator communities. Albarqouni et al. [3] and Isupova et al. [19] extended the idea of previous works [23,39] to convolutional neural networks (CNN) with softmax outputs. Several recent works use a trained classifier to rank consensus labels and incorporate the ranking scores into model learning [7,15]. Another research direction is to take all noisy annotations separately as the input during the training process [10,16,20,43,50,55,59], but Wei et al. [56] concluded that label aggregation is preferred over label separation when there are sufficient number of annotators or the noise rate is moderate. These methods are not directly applicable to object detection tasks since they do not involve regressing the locations of bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.693 + ], + "angle": 0, + "content": "Object Detection with Crowdsourced Annotations. Hu and Meina [18] proposed Crowd R-CNN where one first pre-processes the annotations by combining the bounding boxes using a clustering algorithm, before aggregating the labels and learning the sensitivity of each annotator [39] to different object classes during model training. Crowd R-CNN is tailored for Faster R-CNN (FRCNN) [41] and is incompatible with one-stage object detectors such as YOLOv7 [53]. Le et al. [27] proposed to pre-process annotations using Weighted Boxes Fusion (WBF) [48] and introduced expert agreement re-weighted loss (EARL) to emphasise boxes with higher annotator agreement for model training. It requires access to each annotator's proficiency level to perform WBF and compute EARL, which is often challenging to obtain for real-world crowdsourced data. Compared with the aforementioned methods, the proposed BDC framework can incorporate any object detector in a plug-and-play fashion without requiring any additional inputs such as annotators' proficiency levels. Furthermore, previous works [18, 27] evaluated their methods on different synthetic datasets, with different numbers of annotators and proficiency levels. For example, Hu and Meina [18] performed evaluation on synthetic datasets derived from PASCAL VOC [13] and MS COCO [29]; Le et al. [27] synthesised a MED-MNIST dataset based on MNIST [28] and adopted the VinDr-CXR dataset [36]. Thus, their experiment results cannot be compared directly." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.715, + 0.331, + 0.731 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Problem Statement. We assume a training set \\( D = \\{x^{i},y^{i}\\}_{i = 1}^{N} \\) with pairs of input images \\( x^{i} \\) and crowdsourced annotations \\( y^{i} = \\{y_{m}^{i}\\}_{m = 1}^{M^{i}} \\), where \\( N \\) is the total number of training samples and \\( M^{i} \\) is the number of annotations of the \\( i \\)-th image from all annotators. Each annotation is a tuple of a bounding box, class label and the annotator ID, \\( y_{m}^{i} = \\{(b_{m}^{i},c_{m}^{i},k_{m}^{i})\\} \\) in which \\( b_{m}^{i}\\in \\mathbb{R}^{4} \\) contains the x-y coordinates of the box centre as well as the width and height" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.518, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bayesian Detector Combination" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "of the bounding box, \\( c_{m}^{i} \\in \\{1, \\dots, J\\} \\) represents the annotated class label of the object and \\( k_{m}^{i} \\in \\{1, \\dots, K\\} \\) is the annotator ID. \\( J \\) denotes the number of object classes while \\( K \\) is the total number of annotators. Note that each annotator may annotate a varying number of object instances. For brevity, notations in the rest of this section are simplified to a single training sample \\( i \\). Our goal is to learn an object detector with noisy crowdsourced annotations while inferring and leveraging annotators' varying annotation abilities." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.274, + 0.526, + 0.289 + ], + "angle": 0, + "content": "3.1 Bayesian Detector Combination" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.789, + 0.495 + ], + "angle": 0, + "content": "An overview of the proposed crowdsourced annotation aggregation framework is shown in Figure 2. At a high level, the proposed framework consists of four components, namely: the object detector (OD) module, the annotations-predictions matcher, the bounding box aggregator (BBA) and the class label aggregator (CLA). Firstly, we perform a one-to-many matching of the predictions generated by the OD to the crowdsourced annotations. The bounding box error is computed and used to update a Gaussian-Gamma prior distribution of the BBA for each annotator. Next, the class probability predictions by the OD and the class label annotations are used to update a Dirichlet prior in the CLA. The posterior distributions obtained from both aggregators are used to aggregate the bounding boxes and class labels. Finally, we train and optimise the OD's parameters and aggregators' parameters iteratively until convergence. We present a detailed introduction of each component below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.789, + 0.594 + ], + "angle": 0, + "content": "Object Detector Module. The object detector (OD) module with a parameter set, \\(\\theta_{O}\\), outputs a varying number of \\(P\\) predictions, \\(\\hat{y} = \\{(\\hat{b}_n,\\hat{p}_n)\\}_{n = 1}^P\\), given an input image, \\(x\\), where \\(\\hat{b}_n\\in \\mathbb{R}^4\\) and \\(\\hat{p}_n\\in [0,1]^J\\) are the predicted bounding box and class probability, respectively. We train the model by minimising an overall loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.601, + 0.787, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {D}} = \\frac {1}{H} \\sum_ {n = 1} ^ {H} \\gamma_ {n} \\mathcal {L} \\left(\\mathfrak {b} _ {n}, \\rho_ {n}, \\hat {y}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.743 + ], + "angle": 0, + "content": "where \\(\\gamma_{n}\\), \\(\\mathfrak{b}_{n}\\) and \\(\\rho_{n}\\) are the loss scale factor, the aggregated bounding box, and class probabilities obtained from the two aggregators defined in Section 3.1, respectively. \\(H\\) is the number of aggregated annotations. \\(\\mathcal{L}\\) is the loss function that the base OD model adopts. Since the OD operates as a black-box module in the proposed framework, any object detector can be used here as long as it outputs the predicted bounding boxes and the associated class label probabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.762, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Annotations-Predictions Matcher. To update the priors of the aggregators, we first match each annotation to one of the predictions. To find the optimal prediction, \\(\\hat{y}_m^*\\), for each annotation, \\(y_m\\), we minimise a matching cost function:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.818, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {m} ^ {*} = \\underset {\\hat {y} _ {n} \\in \\hat {y}} {\\arg \\min } \\mathcal {L} _ {\\text {m a t c h}} (\\hat {y} _ {n}, y _ {m}) , \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z.Q. Tan et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.265, + 0.16 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.169, + 0.786, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a t c h}} (\\hat {y} _ {n}, y _ {m}) = - \\hat {p} _ {n (c _ {m})} + \\lambda_ {1} \\mathcal {L} _ {I o U} (\\hat {b} _ {n}, b _ {m}) + \\lambda_ {2} | | \\hat {b} _ {n} - b _ {m} | | _ {1}. \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.193, + 0.787, + 0.269 + ], + "angle": 0, + "content": "\\(\\lambda_{1}, \\lambda_{2} \\in \\mathbb{R}\\) are hyperparameters and \\(\\mathcal{L}_{IoU}\\) is the IoU loss [42]. In short, for each annotation, we aim to find the prediction with a higher predicted probability of the annotated class label, a higher IoU and a smaller L1 distance between the predicted and the annotated bounding box. This is inspired by the set prediction loss in DETR [9]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.27, + 0.788, + 0.362 + ], + "angle": 0, + "content": "Instead of finding a one-to-one matching with a global minimum matching cost, we are interested in a one-to-many matching that one prediction can be matched to many crowdsourced annotations, with a local minimum matching cost, i.e. finding the prediction with minimum cost for each annotation individually. Moreover, a simple heuristic matching rule is sufficient instead of the Hungarian algorithm in DETR which is more computationally demanding." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.379, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Bounding Box Aggregator. In the Bounding Box Aggregator (BBA), the translation and scaling errors in the x and y-axis of each annotator are modelled with a Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.438, + 0.786, + 0.457 + ], + "angle": 0, + "content": "\\[\np \\left(\\epsilon_ {m} \\mid k _ {m} = k, \\boldsymbol {\\mu}, \\boldsymbol {\\sigma}\\right) = \\mathcal {N} \\left(\\mu^ {k}, \\sigma^ {k}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.459, + 0.788, + 0.552 + ], + "angle": 0, + "content": "where \\(\\epsilon_{m} \\in \\mathbb{R}^{4}\\) contains the translation and scaling errors in the x and y-axis. \\(\\pmb{\\mu}\\) and \\(\\pmb{\\sigma}\\) are a collection of Gaussian parameters with \\(\\mu^k\\) and \\(\\sigma^k\\) representing the mean and covariance of \\(\\epsilon_{m}\\) for the annotator \\(k\\), respectively. The Gaussian distributions have a Gaussian-Gamma prior with hyperparameters \\(\\mu_0^k\\), \\(v_0^k\\) and \\(\\beta_0^k\\) [6, 35]. To update the prior distribution, we compute the four error terms between all matched prediction and annotation boxes for each annotator with:" + }, + { + "type": "equation", + "bbox": [ + 0.264, + 0.557, + 0.787, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {m} = \\left[ \\hat {b} _ {m (1)} ^ {*} - b _ {m (1)}, \\hat {b} _ {m (2)} ^ {*} - b _ {m (2)}, \\hat {b} _ {m (3)} ^ {*} \\div b _ {m (3)}, \\hat {b} _ {m (4)} ^ {*} \\div b _ {m (4)} \\right], \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.59, + 0.788, + 0.649 + ], + "angle": 0, + "content": "where \\(\\hat{b}_m^*\\) and \\(b_{m}\\) are the matched prediction and annotation boxes, respectively. The subtraction and division operations are designed to make the error terms translation-invariant and scale-invariant. The Gaussian-Gamma prior is updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.291, + 0.65, + 0.786, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mu} ^ {k} = \\frac {\\mu_ {0} ^ {k} + M ^ {k} \\bar {\\epsilon} ^ {k}}{M ^ {k} + 1}, \\quad (6) \\quad \\tilde {v} ^ {k} = v _ {0} ^ {k} + \\frac {M ^ {k}}{2}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.692, + 0.786, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\beta} ^ {k} = \\beta_ {0} ^ {k} + \\frac {M ^ {k} (\\bar {\\epsilon} ^ {k} - \\mu_ {0} ^ {k}) (\\bar {\\epsilon} ^ {k} - \\mu_ {0} ^ {k}) ^ {\\top}}{2 (M ^ {k} + 1)} + \\frac {1}{2} \\sum_ {m = 1} ^ {M} \\mathbb {1} [ k _ {m} = k ] (\\epsilon_ {m} - \\bar {\\epsilon} ^ {k}) (\\epsilon_ {m} - \\bar {\\epsilon} ^ {k}) ^ {\\top}, (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.739, + 0.787, + 0.783 + ], + "angle": 0, + "content": "where \\(M^k\\) is the number of annotation-prediction matches associated with the \\(k\\)-th annotator, \\(\\overline{\\epsilon}^k\\) is the mean of all \\(\\epsilon_{m}\\) for the \\(k\\)-th annotator and \\(\\mathbb{1}[\\cdot]\\) is the indicator function." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.785, + 0.787, + 0.814 + ], + "angle": 0, + "content": "The obtained posterior mean (i.e. \\(\\mu^k = \\tilde{\\mu}^k\\)) is used to correct the annotated boxes for each annotator:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.819, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\nb _ {m} := \\left(b _ {m} + \\left[ \\mu_ {(1)} ^ {k}, \\mu_ {(2)} ^ {k}, 0, 0 \\right]\\right) \\odot \\left[ 1, 1, \\mu_ {(3)} ^ {k}, \\mu_ {(4)} ^ {k} \\right]. \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.518, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Bayesian Detector Combination" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.21 + ], + "angle": 0, + "content": "All annotations matched to the same prediction are aggregated via a weighted average using the posterior precision, \\( P^k = 1 \\oslash \\text{diag}(\\sigma^k) \\), where \\( \\oslash \\) denotes the element-wise division and \\( \\sigma^k = \\tilde{\\beta}^k / \\tilde{v}^k \\). The posterior precision, \\( P^k \\), of each annotator is used as weights to obtain the final boxes, \\( \\{\\mathfrak{b}_n\\}_{n=1}^H \\), with" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.22, + 0.786, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\mathfrak {b} _ {n} = \\frac {\\sum_ {(b , k) \\in \\kappa_ {n}} P ^ {k} \\odot b}{\\sum_ {(b , k) \\in \\kappa_ {n}} P ^ {k}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.266, + 0.789, + 0.389 + ], + "angle": 0, + "content": "Here, \\( H \\) indicates the number of predictions matched to at least one annotation which is also equal to the number of aggregated annotations defined in Section 3.1. \\( \\kappa_{n} \\) is the set of tuples consisting of the annotated bounding boxes and their annotator ID that are matched to the \\( n \\)-th prediction. Inspired by EARL [27], a loss scale factor, \\( \\gamma_{n} \\), is computed for each aggregated box as the ratio of the number of matched annotations to \\( K \\) and is used to re-weight the loss function of the OD in Equation (1). We refer the readers to Appendix A for further details on the updating procedure of the Bounding Box Aggregator." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.408, + 0.789, + 0.485 + ], + "angle": 0, + "content": "Class Label Aggregator. To aggregate crowdsourced class labels, we extend the Bayesian classifier combination neural network (BCCNet) [19] to the proposed BDC framework. In the class label aggregator (CLA), the annotated class label from each annotator, \\( c_{m} \\), is modelled as a multinomial distribution conditioning on the true label of the object:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.495, + 0.786, + 0.515 + ], + "angle": 0, + "content": "\\[\np \\left(c _ {m} \\mid k _ {m} = k, t _ {m} = j, \\boldsymbol {\\pi}\\right) = \\pi_ {j, c _ {m}} ^ {k}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.523, + 0.789, + 0.601 + ], + "angle": 0, + "content": "where \\(\\pi\\) is a collection of confusion matrices, \\(\\pi^k\\in \\mathbb{R}^{J\\times J}\\) is the confusion matrix of the \\(k\\)-th annotator and \\(t_m\\) is the ground truth class label. Each row, \\(\\pi_j^k\\), provides the event probabilities of a multinomial distribution for the \\(k\\)-th annotator when \\(t_m = j\\) and is modelled with a Dirichlet prior with hyperparameters \\(\\alpha_{0j}^k = \\{\\alpha_{0j,1}^k,\\dots,\\alpha_{0j,J}^k\\}\\) such that \\(\\pi_j^k |\\alpha_{0j}^k\\sim \\text{Dir}(\\alpha_{0j}^k)\\) [23]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.601, + 0.789, + 0.631 + ], + "angle": 0, + "content": "The multinomial posterior distribution of the true label (i.e. the aggregated class label probability) can be obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.642, + 0.786, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\rho_ {n, j} = \\exp \\left(\\ln \\hat {p} _ {n, j} + \\sum_ {(c, k) \\in \\tilde {\\kappa} _ {n}} \\mathbb {E} _ {\\pi_ {j} ^ {k}} \\ln \\pi_ {j, c} ^ {k}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.698, + 0.789, + 0.79 + ], + "angle": 0, + "content": "where \\(\\tilde{\\kappa}_n\\) is the set of tuples consisting of the annotated class labels and their annotator ID that are matched to the \\(n\\)-th prediction. \\(\\hat{p}_{n,j}\\) is the OD's class-\\(j\\) probability of the \\(n\\)-th prediction. Note that the CLA outputs class probabilities (soft labels) unlike other methods (e.g. MV) that output a single class (hard label). Appendix B provides the derivation for Equation 12 and a detailed discussion of the updating steps of the Class Label Aggregator." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Training. The optimal OD's parameters, \\(\\theta_{O}\\), and aggregators' parameters, \\(\\pmb{\\mu}\\), \\(\\pmb{v}\\), \\(\\pmb{\\beta}\\) and \\(\\pmb{\\alpha}\\), are learnt iteratively in the BDC framework. Firstly, we initialise" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z.Q. Tan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "the parameters either randomly or with prior knowledge of the domain. Next, we train the OD by optimising the loss function, \\(\\mathcal{L}_{\\mathcal{D}}\\), with the aggregated annotations obtained using the current values of the aggregators' parameters. After one epoch of training, we obtain the OD's predictions using the current value of \\(\\theta_{O}\\) to update the prior distributions of aggregators. These two steps are repeated iteratively until the parameters converge. See Appendix C for a further discussion of the training steps." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.377, + 0.291 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.789, + 0.455 + ], + "angle": 0, + "content": "Training details. To demonstrate the generalisability and robustness of BDC in incorporating different object detection algorithms, three common object detection algorithms are used in our experiments, YOLOv7 [53], FRCNN [41] with the ResNet50 backbone and EVA [14]. During training, we randomly apply data augmentation such as rotation, translation, zooming and horizontal flipping to the training images and annotations. \\(\\lambda_{1},\\lambda_{2}\\) in Equation (3) are set to 2 and 5, respectively, following Carion et al. [9]. Lastly, the Gaussian-Gamma prior of the BBA, \\(\\pmb{\\mu}_{0}\\), \\(\\pmb{v}_{0}\\) and \\(\\beta_0\\), are set to 0, 10 and 0.5, respectively, while the Dirichlet prior of the CLA, \\(\\alpha_0\\), is initialised to 1 and each diagonal entry is set to 10. The complete training configuration can be found in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.789, + 0.633 + ], + "angle": 0, + "content": "Comparison against other aggregation methods. We compare the performance of BDC against multiple baselines as shown in Tables 1 to 4. 'NA' denotes the process of training the model with all the crowdsourced annotations as the ground truth annotations [56]. 'MV' stands for majority voting where we apply clustering to the crowdsourced annotations, then aggregate the class labels via MV and aggregate the bounding boxes as the biggest box where the majority had annotated pixel-wise, following the practice in [46]. We also implement Crowd R-CNN [18] and WBF-EARL [27] designed for crowdsourced object detection. For WBF-EARL [27] which requires the knowledge of annotators' proficiency levels, we followed the settings in [27] for the VinDR-CXR dataset while for the synthetic datasets, they are set to the overall accuracy of the annotators." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.652, + 0.502, + 0.667 + ], + "angle": 0, + "content": "4.1 Real Crowdsourced Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "VinDr-CXR. We evaluate the annotation aggregation methods on VinDrcxR [36], a thoracic abnormalities detection dataset. VinDr-CXR consists of 14 abnormalities classes and contains 15,000 training and 3,000 testing chest radiograph images. The training images are annotated by 17 experienced radiologists where at least three annotators are randomly assigned to annotate each image. Figures 1f to 1j show examples of the annotations with high bounding box and class label disagreement by the annotators. Furthermore, the testing images are annotated with the consensus of five expert radiologists through a two-stage labelling process and the test set annotations are not publicly available. The predictions are submitted to the VinBigData Kaggle [12] to obtain mean average precision (AP) at 0.4 IoU threshold on the private test set." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Bayesian Detector Combination" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: AP-4 results on VinDr-CXR [36] dataset obtained from the private score of VinBigData Kaggle competition [12]. Crowd R-CNN can only integrate FRCNN, hence there are no results reported for the two other entries corresponding to Crowd R-CNN." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.199, + 0.65, + 0.288 + ], + "angle": 0, + "content": "
MethodTest AP-4
YOLOv7FRCNNEVA
NA17.417.27.8
MV13.916.38.2
Crowd R-CNN [18]-16.7-
WBF-EARL [27]16.417.08.4
BDC (ours)19.217.98.9
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.303, + 0.329, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.42, + 0.295, + 0.431 + ], + "angle": 0, + "content": "(a) NA" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.303, + 0.441, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.42, + 0.409, + 0.431 + ], + "angle": 0, + "content": "(b) MV" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.303, + 0.554, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.42, + 0.546, + 0.431 + ], + "angle": 0, + "content": "(c) WBF-EARL" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.303, + 0.667, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.42, + 0.665, + 0.431 + ], + "angle": 0, + "content": "(d) Crowd R-CNN" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.303, + 0.779, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.42, + 0.767, + 0.431 + ], + "angle": 0, + "content": "(e) BDC (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.443, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Fig. 3: Comparison of aggregated labels from different methods on the VinDr-CXR [36]. For WBF-EARL [27], the number beside the class label is the annotators' level of agreement while for Crowd R-CNN [18] and BDC, the number indicates the class probability. For NA, the colours represent the different annotators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.64 + ], + "angle": 0, + "content": "Disaster Response Dataset. We further use the disaster response dataset [19] where the task is to detect undamaged and damaged buildings on satellite images taken after hurricane Maria in 2017. The dataset consisting of 464 satellite images of Dominica is annotated by 13 semi-professional paid crowd members to ensure high labelling quality. Despite this effort, labels from different crowd members vary significantly (Figures 1k to 1o). Consequently, the dataset does not contain ground truth labels. Thus, its qualitative results are presented in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Quantitative Results. Table 1 presents the AP.4 metric obtained from the VinBigData Kaggle competition and shows that the proposed BDC leads to higher AP.4 than other methods across different OD models on the VinDr-CXR test set. As Crowd R-CNN can only integrate FRCNN, there are no results reported for the two other entries corresponding to Crowd R-CNN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Qualitative Results. We visualise the results of the aggregated annotations from each method on the VinDr-CXR dataset in Figure 3. Since the true ground truth of the datasets is not provided, we use 'MV' as the best approximation of the ground truth for this qualitative assessment and observe that BDC produces more accurate aggregation compared to the other methods." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z.Q. Tan et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.418, + 0.162 + ], + "angle": 0, + "content": "4.2 Synthetic Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.371 + ], + "angle": 0, + "content": "To alleviate the limitations in the available real-world crowdsourced datasets including the lack of ground truth labels, we synthesise several synthetic datasets with crowdsourced annotations to enable the systematic evaluation of the proposed method. This makes available the ground truth annotations to allow us to obtain detailed test evaluation metrics (e.g., AP\\(^{5}\\), AP\\(^{75}\\), AP\\(^{5::95}\\)) and verify the performance of the proposed method by evaluating the aggregated crowdsourced annotations with the true ground truth annotations of the training dataset (indicated as 'Train Aggregation AP' in Tables 2 to 4). Two common object detection datasets are used to construct the synthetic datasets: (1) Pascal Visual Object Classes (VOC) 2007 [13], a 20-class object detection dataset which consists of 5,011 and 4,952 images for training and testing, respectively; (2) Microsoft Common Objects in Context (COCO) 2017 [29], a large-scale 80-class object detection dataset consisting of 118,287 training and 5,000 validation images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Synthetic settings. Four synthetic settings are investigated in this work: (1) VOC-FULL and (2) COCO-FULL where 10 annotators with average skill levels annotate the whole datasets; (3) VOC-MIX where the first 5 annotators are experts while the remaining 20 annotators have an average skill level. All annotators annotate the whole dataset since Pascal VOC is a small dataset; and (4) COCO-MIX where 20 experts, 550 annotators with average skill level and 330 annotators with poor skill level work together to annotate the MS COCO dataset. Additionally, the last 100 are random annotators, i.e., the annotated class labels and bounding boxes are drawn from a uniform distribution. Since MS COCO is a large-scale dataset, we set that each annotator annotates a small subset of images, i.e., the coverage for the expert, average, poor and random annotators are \\(5\\%\\), \\(1\\%\\), \\(0.5\\%\\) and \\(1\\%\\), respectively. Note that all annotators in these settings are simulated using deep learning models (See Appendix F for further discussions of the synthesis method and pseudocode). Additional experiments where we directly used the object detectors' predictions as synthetic crowdsourced annotations can be found in Appendix H." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Quantitative Results. Tables 2 to 4 report the experiment results of YOLOv7, FRCNN and EVA trained with different aggregation methods on all synthetic settings, respectively. Additionally, we trained each model with ground truth annotations of each dataset (denoted by 'GT') to obtain reference optimal model performance. The aggregation APs of the train dataset for NA and WBF-EARL are low because NA directly uses all the crowdsourced annotations while WBF outputs multiple bounding boxes per class per object instance, leading to a much lower precision (see Figure 4 for examples). We do not include performance from Crowd R-CNN for the COCO-MIX setting as its training time increases linearly with \\( K \\) which means that the training time is prohibitively high for this dataset where \\( K = 1,000 \\). In comparison, the extra computational cost introduced by the proposed BDC is small when compared to the much heavier training cost of the OD backbone (see Section 4.4 for an analysis of the computational cost)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bayesian Detector Combination" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 2: Test and train aggregation AP results for YOLOv7 model trained with different annotation aggregation methods on synthesised and ground truth datasets." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.185, + 0.737, + 0.43 + ], + "angle": 0, + "content": "
DatasetMethodTest APTrain Aggregation AP
AP.5AP.75AP.5:95AP.5AP.75AP.5:95
VOC-FULLNA59.745.340.710.29.99.9
MV72.252.347.362.048.548.5
WBF-EARL [27]69.352.345.722.316.414.5
BDC (ours)84.963.554.895.080.666.3
VOC-MIXNA46.127.627.44.44.34.3
MV77.161.053.775.770.668.6
WBF-EARL [27]61.047.240.927.924.920.0
BDC (ours)86.269.058.897.694.977.2
VOCGT87.774.065.3100100100
COCO-FULLNA53.441.437.48.68.28.2
MV61.946.442.752.544.443.3
WBF-EARL [27]55.642.538.914.68.98.6
BDC (ours)65.049.644.583.666.156.9
COCO-MIXNA53.441.237.58.57.98.0
MV60.946.442.445.842.341.1
WBF-EARL [27]56.242.639.112.17.07.0
BDC (ours)64.849.344.480.562.954.3
COCOGT67.552.848.6100100100
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.52 + ], + "angle": 0, + "content": "Tables 2 to 4 demonstrate the superior performance of BDC compared to other methods from simpler synthetic settings to more complex synthetic crowdsourced annotations, i.e. from VOC-FULL to COCO-MIX. In addition, BDC achieves the highest train aggregation AP compared to other methods across all settings." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.535, + 0.787, + 0.701 + ], + "angle": 0, + "content": "Qualitative Results. Figure 4 visualises the results of the aggregated annotations from each method on the VOC-MIX synthetic dataset. WBF performs box fusion on a per-class basis and results in significantly more boxes compared to other methods (excluding NA) which contributed to the low train aggregation AP and noisy labels during model training. On the other hand, although both Crowd R-CNN and BDC output class probabilities, the probability distribution of Crowd R-CNN is heavily skewed which is similar to hard labels, while BDC outputs the learnt class probabilities and can benefit from these soft labels during model training [34]. Besides, BDC can cope with failure cases of its aggregator as shown by the lower class probability of incorrect aggregation. Additional qualitative results can be found in Appendix I." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.724, + 0.551, + 0.74 + ], + "angle": 0, + "content": "4.3 Scalability and Robustness of BDC" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We analyse the scalability and robustness of the aggregation methods using the Pascal VOC dataset. YOLOv7 is selected as the OD because it is lightweight and fast to train. Figure 5a demonstrates the effect of varying numbers of annotators, \\( K \\), where all annotations are synthesised using annotators of average skill level. Note that training YOLOv7 with the NA method results in a GPU out-of-memory error when \\( K \\geq 100 \\). When the number of annotators increases, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z.Q. Tan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Test and train aggregation AP results for FRCNN model trained with different annotation aggregation methods on synthesised and ground truth datasets. Crowd R-CNN has a training time that increases linearly with \\( K \\), hence there is no result reported for COCO-MIX dataset with \\( K = 1,000 \\)." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.213, + 0.746, + 0.511 + ], + "angle": 0, + "content": "
DatasetMethodTest APTrain Aggregation AP
AP·5AP·75AP·5:95AP·5AP·75AP·5:95
VOC-FULLNA47.411.319.410.29.99.9
MV79.753.548.862.048.548.5
Crowd R-CNN [18]71.633.930.158.343.031.5
WBF-EARL [27]71.837.738.722.316.414.5
BDC (ours)83.756.351.095.080.466.3
VOC-MIXNA34.14.412.24.44.34.3
MV82.057.151.975.770.668.6
Crowd R-CNN [18]70.231.229.360.644.231.0
WBF-EARL [27]69.836.537.427.924.920.0
BDC (ours)84.759.352.997.694.676.9
VOCGT84.662.155.2100100100
COCO-FULLNA39.712.017.78.68.28.2
MV55.635.733.452.544.443.3
Crowd R-CNN [18]48.527.124.249.536.233.7
WBF-EARL [27]51.928.028.414.68.98.6
BDC (ours)56.635.333.583.366.056.7
COCO-MIXNA43.917.521.48.57.98.0
MV54.834.532.945.842.341.1
Crowd R-CNN [18]------
WBF-EARL [27]51.928.028.312.17.07.0
BDC (ours)56.134.633.080.362.754.1
COCOGT58.038.535.7100100100
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.787, + 0.584 + ], + "angle": 0, + "content": "performance of YOLOv7 increases for both BDC and MV while the performance of both NA and WBF-EARL decreases when \\( K > 50 \\). In addition, BDC has an average AP.5:95 improvement of 4.35 over MV." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.766 + ], + "angle": 0, + "content": "We also investigate the effect of changing the percentage of noisy annotators while setting \\( K \\) to 25. Here, noisy annotators are defined as annotators with a poor skill level while the other annotators are annotators with expert skill level. As illustrated in Figure 5b, we found that AP\\(^{5:95}\\) of all aggregation methods reduce when we increase the percentage of noisy annotators. However, MV drops at a faster rate compared to the proposed method when the noisy annotator percentage is over \\( 50\\% \\). We observe that BDC remains robust even when \\( 100\\% \\) of the annotators have low skill levels and the AP\\(^{5:95}\\) is 18.5 higher than that obtained with MV. The consistent performance of BDC (up to \\( 80\\% \\) noise) can be attributed to the ability of BDC to correctly capture the ability of each annotator and assign proper weights when aggregating the annotations (See Appendix J for an analysis of the posterior distributions learnt in the aggregators)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.431, + 0.802 + ], + "angle": 0, + "content": "4.4 Computational Cost" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Table 5 presents the wall time of each annotation aggregation method spent to pre-process the annotations and train the Faster R-CNN (FRCNN) [41] model for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Bayesian Detector Combination" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 4: Test and train aggregation AP results for EVA model trained with different annotation aggregation methods on synthesised and ground truth datasets." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.185, + 0.737, + 0.43 + ], + "angle": 0, + "content": "
DatasetMethodTest APTrain Aggregation AP
AP.5AP.75AP.5:95AP.5AP.75AP.5:95
VOC-FULLNA90.980.969.710.29.99.9
MV91.082.271.262.048.548.5
WBF-EARL [27]93.083.571.122.316.414.5
BDC (ours)94.285.671.795.381.266.7
VOC-MIXNA86.665.757.54.44.34.3
MV91.783.072.975.770.668.6
WBF-EARL [27]91.382.268.727.924.920.0
BDC (ours)94.986.573.898.895.077.3
VOCGT94.988.178.0100100100
COCO-FULLNA71.855.750.68.68.28.2
MV74.861.054.452.544.443.3
WBF-EARL [27]74.758.451.814.68.98.6
BDC (ours)78.062.755.183.365.155.3
COCO-MIXNA73.357.251.78.57.98.0
MV75.060.654.245.842.341.1
WBF-EARL [27]75.259.352.912.17.07.0
BDC (ours)77.162.554.879.962.253.2
COCOGT78.864.958.8100100100
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.614 + ], + "angle": 0, + "content": "one epoch on the VOC-MIX synthetic dataset. The wall time is computed with a desktop workstation with an i9-10900X CPU and an RTX 3090 GPU. As shown in Table 5, MV, Crowd R-CNN and WBF-EARL pre-process the annotations before model training while NA and the proposed BDC require no pre-processing step. Crowd R-CNN has the longest training time which increases linearly with the number of annotators, \\( K \\). On the other hand, the proposed BDC performs the annotations-predictions matching step and the updating steps of the aggregators' parameters every training epoch and therefore has a slightly longer training time compared to NA, MV and WBF-EARL. However, this increase in training time is small when compared to the much higher training cost of the OD" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.645, + 0.313, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.747, + 0.287, + 0.759 + ], + "angle": 0, + "content": "(a) NA" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.645, + 0.407, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.747, + 0.383, + 0.758 + ], + "angle": 0, + "content": "(b) MV" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.645, + 0.501, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.747, + 0.47, + 0.767 + ], + "angle": 0, + "content": "(c) WBF-EARL" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.645, + 0.597, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.747, + 0.572, + 0.767 + ], + "angle": 0, + "content": "(d) Crowd R-CNN" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.645, + 0.692, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.747, + 0.689, + 0.767 + ], + "angle": 0, + "content": "(e) BDC (ours)" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.645, + 0.787, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.747, + 0.761, + 0.767 + ], + "angle": 0, + "content": "(f) Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 4: Comparison of aggregated labels from different methods on the VOC-MIX synthetic dataset. For WBF-EARL [27], the number beside the class label is the annotators' level of agreement while for Crowd R-CNN [18] and BDC, the number indicates the class probability. For NA, the colours represent the different annotators." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z.Q. Tan et al." + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.144, + 0.501, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.313, + 0.434, + 0.325 + ], + "angle": 0, + "content": "(a) Scalability of BDC" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.144, + 0.77, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.313, + 0.708, + 0.325 + ], + "angle": 0, + "content": "(b) Robustness of BDC" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Fig. 5: (a) Effect of the number of annotators, \\( K \\), and (b) effect of varying percentage of noisy annotators (with \\( K = 25 \\)) on the test AP\\(^{5:95}\\) of YOLOv7. Results for the 'NA' method are unavailable when \\( K \\geq 100 \\) due to GPU out-of-memory error." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.437 + ], + "angle": 0, + "content": "Table 5: Annotations pre-processing time and FRCNN model training time per epoch of different annotation aggregation methods on the VOC-MIX dataset. The wall time is computed on a desktop workstation with an i9-10900X CPU and an RTX 3090 GPU." + }, + { + "type": "table", + "bbox": [ + 0.298, + 0.447, + 0.706, + 0.528 + ], + "angle": 0, + "content": "
MethodPre-processing time (s)Training time (s)
NA0.0537.0
MV19.4505.4
Crowd R-CNN [18]15.61175.5
WBF-EARL [27]4.9518.5
BDC (ours)0.0662.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.589 + ], + "angle": 0, + "content": "module (618.6s OD training time out of 662.9s). Note that the model inference time is unaffected by the type of annotation aggregation method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.611, + 0.36, + 0.627 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We first presented a general annotation aggregation and object detection framework, named Bayesian Detector Combination (BDC), for learning a robust object detector from noisy crowdsourced annotations. BDC infers the consensus bounding boxes and corresponding soft labels without the need for ground truth annotations. BDC is model-agnostic and we demonstrated its superior performance by integrating it with three popular object detector backbones, namely YOLOv7, FRCNN, and EVA. Furthermore, we constructed four synthetic datasets to simulate different crowdsourcing scenarios involving annotators of various skill levels, enabling the systematic evaluation of the proposed BDC. We demonstrated the improved performance from BDC on real-world datasets and all the synthetic datasets when compared to other object detectors designed for crowdsourced annotations. We showed that BDC can scale against varying numbers of annotators and is robust against annotators with low annotation reliability." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Bayesian Detector Combination" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.266 + ], + "angle": 0, + "content": "Acknowledgements. Z.Q. T. acknowledges the support of an Industrial Cooperative Awards in Science and Engineering (ICASE) studentship from the Engineering and Physical Sciences Research Council (EPSRC) for this work. G. C. acknowledges the support of the Engineering and Physical Sciences Research Council (EPSRC) through grant EP/Y018036/1 and the Australian Research Council (ARC) through grant FT190100525. The authors would like to thank the Satellite Applications Catapult for the provision of the disaster response dataset." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.296, + 0.323, + 0.312 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.375 + ], + "angle": 0, + "content": "1. Albarqouni, S., Baur, C., Achilles, F., Belagiannis, V., Demirci, S., Navab, N.: AggNet: Deep learning from crowds for mitosis detection in breast cancer histology images. IEEE Trans. Med. Imag. 35(5), 1313-1321 (May 2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.376, + 0.785, + 0.431 + ], + "angle": 0, + "content": "2. Baumgartner, C.F., Tezcan, K.C., Chaitanya, K., Hötter, A.M., Muehlematter, U.J., Schawkat, K., Becker, A.S., Donati, O., Konukoglu, E.: PHiSeg: Capturing uncertainty in medical image segmentation. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 119-127. Shenzhen, China (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.785, + 0.459 + ], + "angle": 0, + "content": "3. Bernard, J.M., Smith, A.F.M.: Bayesian Theory, chap. 3, pp. 136-138. Wiley, West Sussex, UK (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.516 + ], + "angle": 0, + "content": "4. Bernhardt, M., Castro, D.C., Tanno, R., Schwaighofer, A., Tezcan, K.C., Monteiro, M., Bannur, S., Lungren, M.P., Nori, A., Glocker, B., Alvarez-Valle, J., Oktay, O.: Active label cleaning for improved dataset quality under resource constraints. Nature Commun. 13(1), 1161 (Mar 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.785, + 0.572 + ], + "angle": 0, + "content": "5. Budd, S., Day, T., Simpson, J., Lloyd, K., Matthew, J., Skelton, E., Razavi, R., Kainz, B.: Can non-specialists provide high quality gold standard labels in challenging modalities? In: Med. Imag. Comput. Comput. Assist. Interv. pp. 251-262. Strasbourg, France (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.573, + 0.785, + 0.614 + ], + "angle": 0, + "content": "6. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.615, + 0.785, + 0.657 + ], + "angle": 0, + "content": "7. Chen, Z., Wang, H., Sun, H., Chen, P., Han, T., Liu, X., Yang, J.: Structured probabilistic end-to-end learning from crowds. In: Int. Joint Conf. Artif. Intell. pp. 1512-1518 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.658, + 0.785, + 0.699 + ], + "angle": 0, + "content": "8. Dawid, A.P., Skene, A.M.: Maximum likelihood estimation of observer error-rates using the EM algorithm. J. Roy. Statist. Soc. Ser. C (Appl. Statist.) 28(1), 20-28 (1979)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "9. DungNB, Nguyen, H.Q., Elliott, J., KeepLearning, Nhan, N.T., Culliton, P.: VinBigData chest x-ray abnormalities detection (2020), https://kaggle.com/competitions/ vinbigdata-chest-xray-abnormalities-detection" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.742, + 0.785, + 0.784 + ], + "angle": 0, + "content": "10. Everingham, M., Luc Van Gool, C.K.I.W., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2007 (VOC2007). Int. J. Comput. Vis. 88, 303-338 (Jun 2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "11. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA: Exploring the limits of masked visual representation learning at scale. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19358-19369. Vancouver, Canada (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.332, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z.Q. Tan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "12. Goh, H.W., Tkachenko, U., Mueller, J.: CROWDLAB: Supervised learning to infer consensus labels and quality scores for data with multiple annotators (2022), arXiv:2210.06812" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "13. Guan, M.Y., Gulshan, V., Dai, A.M., Hinton, G.E.: Who said what: Modeling individual labelers improves classification. In: AAAI Conf. Artif. Intell. pp. 3109-3118. San Francisco, CA, USA (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.274 + ], + "angle": 0, + "content": "14. Hu, Y., Meina, S.: Crowd R-CNN: An object detection model utilizing crowdsourced labels. In: Int. Conf. Vis. Image Sig. Process. pp. 1-7. Bangkok, Thailand (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "15. Isupova, O., Li, Y., Kuzin, D., Roberts, S.J., Willis, K.J., Reece, S.: BCCNet: Bayesian classifier combination neural network. In: Neural Inf. Process. Syst. Workshop Mach. Learn. Develop. Montréal, Canada (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.374 + ], + "angle": 0, + "content": "16. Jensen, M.H., Jorgensen, D.R., Jalaboi, R., Hansen, M.E., Olsen, M.A.: Improving uncertainty estimation in convolutional neural networks using inter-rater agreement. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 540-548. Shenzhen, China (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.416 + ], + "angle": 0, + "content": "17. Ji, W., Yu, S., Wu, J., Ma, K., Bian, C., Bi, Q., Li, J., Liu, H., Cheng, L., Zheng, Y.: Learning calibrated medical image segmentation via multi-rater agreement modeling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12336-12346 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.473 + ], + "angle": 0, + "content": "18. Jungo, A., Meier, R., Ermis, E., Blatti-Moreno, M., Herrmann, E., Wiest, R., Reyes, M.: On the effect of inter-observer variability for a reliable estimation of uncertainty of medical image segmentation. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 682-690. Granada, Spain (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.502 + ], + "angle": 0, + "content": "19. Kim, H., Ghahramani, Z.: Bayesian classifier combination. In: Int. Conf. Artif. Intell. Statist. pp. 619-627. La Palma, Canary Islands (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.557 + ], + "angle": 0, + "content": "20. Kohl, S.A.A., Romera-Paredes, B., Meyer, C., Fauw, J.D., Ledsam, J.R., Maier-Hein, K.H., Eslami, S.M.A., Rezende, D.J., Ronneberger, O.: A probabilistic u-net for segmentation of ambiguous images. In: Adv. Neural Inform. Process. Syst. pp. 6965–6975. Montréal, Canada (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.785, + 0.6 + ], + "angle": 0, + "content": "21. Lazarus, E., Mainiero, M.B., Schepps, B., Koelliker, S.L., Livingston, L.S.: BI-RADS lexicon for US and mammography: interobserver variability and positive predictive value. Radiology 239(2), 385-391 (Mar 2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "22. Le, K., Tran, T., Pham, H., Nguyen Trung, H., Le, T., Nguyen, H.Q.: Learning from multiple expert annotators for enhancing anomaly detection in medical image analysis. IEEE Access 11, 14105-14114 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.671 + ], + "angle": 0, + "content": "23. LeCun, Y., Cortes, C., Burges, C.: MNIST handwritten digit database (2010), http://yann.lecun.com/exdb/mnist" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.713 + ], + "angle": 0, + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Zurich, Switzerland (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.784 + ], + "angle": 0, + "content": "25. Maier-Hein, L., Mersmann, S., Kondermann, D., Stock, C., Kenngott, H.G., Sanchez, A., Wagner, M., Preukschas, A., Wekerle, A.L., Helfert, S., Bodenstedt, S., Speidel, S.: Crowdsourcing for reference correspondence generation in endoscopic images. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 349-356. Boston, MA, USA (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "26. Müller, R., Kornblith, S., Hinton, G.: When does label smoothing help? In: Adv. Neural Inform. Process. Syst. pp. 4694-4703. Vancouver, Canada (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.841 + ], + "angle": 0, + "content": "27. Murphy, K.: Conjugate bayesian analysis of the gaussian distribution (2007), https://www.cs.ubc.ca/~murphyk/Papers/bayesGauss.pdf" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Bayesian Detector Combination" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.23 + ], + "angle": 0, + "content": "28. Nguyen, H.Q., Lam, K., Le, L.T., Pham, H.H., Tran, D.Q., Nguyen, D.B., Le, D.D., Pham, C.M., Tong, H.T.T., Dinh, D.H., Do, C.D., Doan, L.T., Nguyen, C.N., Nguyen, B.T., Nguyen, Q.V., Hoang, A.D., Phan, H.N., Nguyen, A.T., Ho, P.H., Ngo, D.T., Nguyen, N.T., Nguyen, N.T., Dao, M., Vu, V.: VinDr-CXR: An open dataset of chest x-rays with radiologist's annotations. Sci. Data 9(1), 429 (Jul 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.787, + 0.285 + ], + "angle": 0, + "content": "29. Nowruzi, F.E., Kapoor, P., Kolhatkar, D., Hassanat, F., Laganiere, R., Rebut, J.: How much real data do we actually need: Analyzing object detection performance using synthetic and real data. In: Int. Conf. Mach. Learn. Workshop AI Auton. Driving. Long Beach, CA, USA (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.285, + 0.787, + 0.325 + ], + "angle": 0, + "content": "30. Raykar, V.C., Yu, S., Zhao, L.H., Valadez, G.H., Florin, C., Bogoni, L., Moy, L., Raykar, C., Valadez, G.H.: Learning from crowds. J. Mach. Learn. Res. 11(43), 1297-1322 (Aug 2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.787, + 0.367 + ], + "angle": 0, + "content": "31. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: Adv. Neural Inform. Process. Syst. pp. 91-99. Montréal, Canada (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.367, + 0.787, + 0.42 + ], + "angle": 0, + "content": "32. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 658-666. Long Beach, CA, USA (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.421, + 0.787, + 0.447 + ], + "angle": 0, + "content": "33. Rodrigues, F., Pereira, F.C.: Deep learning from crowds. In: AAAI Conf. Artif. Intell. pp. 1611-1618. New Orleans, LA, USA (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "34. Sheng, V.S., Provost, F., Ipeirotis, P.G.: Get another label? improving data quality and data mining using multiple, noisy labelers. In: ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. pp. 614-622. Las Vegas, NV, USA (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "35. Solovyev, R., Wang, W., Gabruseva, T.: Weighted boxes fusion: Ensembling boxes from different object detection models. Image Vis. Comput. 107, 104-117 (Mar 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.787, + 0.582 + ], + "angle": 0, + "content": "36. Sudre, C.H., Anson, B.G., Ingala, S., Lane, C.D., Jimenez, D., Haider, L., Varsavsky, T., Tanno, R., Smith, L., Ourselin, S., Jäger, R.H., Cardoso, M.J.: Let's agree to disagree: Learning highly debatable multirater labelling. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 665-673. Shenzhen, China (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.624 + ], + "angle": 0, + "content": "37. Venanzi, M., Guiver, J., Kazai, G., Kohli, P., Shokouhi, M.: Community-based bayesian aggregation models for crowdsourcing. In: Int. Conf. World Wide Web. pp. 155-164. Seoul, Korea (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.624, + 0.787, + 0.664 + ], + "angle": 0, + "content": "38. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7464-7475. Vancouver, Canada (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.664, + 0.787, + 0.718 + ], + "angle": 0, + "content": "39. Watadani, T., Sakai, F., Johkoh, T., Noma, S., Akira, M., Fujimoto, K., Bankier, A.A., Lee, K.S., Müller, N.L., Song, J.W., Park, J.S., Lynch, D.A., Hansell, D.M., Remy-Jardin, M., Franquet, T., Sugiyama, Y.: Interobserver variability in the CT assessment of honeycombing in the lungs. Radiology 266(3), 936-944 (Dec 2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.787, + 0.758 + ], + "angle": 0, + "content": "40. Wei, H., Xie, R., Feng, L., Han, B., An, B.: Deep learning from multiple noisy annotators as a union. IEEE Trans. Neural Netw. Learn. Syst. pp. 1-11 (Apr 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "41. Wei, J., Zhu, Z., Luo, T., Amid, E., Kumar, A., Liu, Y.: To aggregate or not? learning with separate noisy labels. In: ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. pp. 2523-2535. Long Beach, CA, USA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "42. Whitehill, J., Wu, T.f., Bergsma, J., Movellan, J., Ruvolo, P.: Whose vote should count more: Optimal integration of labels from labelers of unknown expertise. In: Adv. Neural Inform. Process. Syst. pp. 2035-2043. Vancouver, Canada (2009)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z.Q. Tan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "43. Wu, J., Fang, H., Yang, Y., Liu, Y., Gao, J., Duan, L., Yang, W., Xu, Y.: Multi-rater prism: Learning self-calibrated medical image segmentation from multiple raters (2022), arXiv:2212.00601" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "44. Yu, S., Zhou, H.Y., Ma, K., Bian, C., Chu, C., Liu, H., Zheng, Y.: Difficulty-aware glaucoma classification with multi-rater consensus modeling. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 741-750 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "45. Zhang, L., Tanno, R., Xu, M.C., Jin, C., Jacob, J., Ciccarelli, O., Barkhof, F., Alexander, D.C.: Disentangling human error from the ground truth in segmentation of medical images. In: Adv. Neural Inform. Process. Syst. pp. 15750-15762 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.301 + ], + "angle": 0, + "content": "46. Zhu, X., Vondrick, C., Ramanan, D., Fowlkes, C.: Do we need more training data or better models for object detection? In: Brit. Mach. Vis. Conf. Surrey, UK (2012)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.301 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_origin.pdf b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45d445d4f1b0c0359b9f838da798221e1b465a7b --- /dev/null +++ b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/1de40a86-4b30-47f0-9bdb-50adf2a69183_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2262ab7aa41b55c39359166da423652782d6ad3e35a54d783ac117fa02c51796 +size 3308196 diff --git a/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/full.md b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3fe3469d19b9d22f7801175b4c5f72262349e459 --- /dev/null +++ b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/full.md @@ -0,0 +1,339 @@ +# Bayesian Detector Combination for Object Detection with Crowdsourced Annotations + +Zhi Qin $\mathrm{Tan}^{1}$ , Olga Isupova $^{2}$ , Gustavo Carneiro $^{1}$ , Xiatian Zhu $^{1}$ , and Yunpeng Li $^{1}$ + +1 University of Surrey, Surrey, UK {z.tan,g.carneiro, xiatian.zhu, yunpeng.li}@surrey.ac.uk +$^{2}$ University of Oxford, Oxford, UK olga.isupova@ouce.ox.ac.uk + +Abstract. Acquiring fine-grained object detection annotations in unconstrained images is time-consuming, expensive, and prone to noise, especially in crowdsourcing scenarios. Most prior object detection methods assume accurate annotations; A few recent works have studied object detection with noisy crowdsourced annotations, with evaluation on distinct synthetic crowdsourced datasets of varying setups under artificial assumptions. To address these algorithmic limitations and evaluation inconsistency, we first propose a novel Bayesian Detector Combination (BDC) framework to more effectively train object detectors with noisy crowdsourced annotations, with the unique ability of automatically inferring the annotators' label qualities. Unlike previous approaches, BDC is model-agnostic, requires no prior knowledge of the annotators' skill level, and seamlessly integrates with existing object detection models. Due to the scarcity of real-world crowdsourced datasets, we introduce large synthetic datasets by simulating varying crowdsourcing scenarios. This allows consistent evaluation of different models at scale. Extensive experiments on both real and synthetic crowdsourced datasets show that BDC outperforms existing state-of-the-art methods, demonstrating its superiority in leveraging crowdsourced data for object detection. Our code and data are available at: https://github.com/zhiqin1998/bdc. + +Keywords: Bayesian Detector Combination $\cdot$ Object detection $\cdot$ Crowdsourced annotations + +# 1 Introduction + +Despite significant advancements in the development of object detection with natural images [9, 14, 41, 53], existing methods typically assume the availability of large, accurately annotated datasets [37, 61]. Nonetheless, this assumption does not often hold true in practice. For example, accurate labels are often expensive to acquire in many applications. A large number of experts with varying professional skills can be involved to crowdsourced annotations of images sampled with replacement from a large dataset [8, 33, 50]. Usually, the annotations + +![](images/5581d5e72fc9c81d38c5028897e9a5518e186141340e1acc021947b725279667.jpg) +(a) ID: 70282 + +![](images/05b271ffd556ac007a75bd771bc7ee4d9dad08270afa4c676a571740ba5cb60d.jpg) +(b) ID: 310072 + +![](images/7c49c0cf0db62ebd904ad3997472f95222e089f38f578d517c36193528ec48b6.jpg) +(c) ID: 516471 + +![](images/3820121acb52d132162c76ac90c25a9dc2bfbf62a610eba4b4327850149e7352.jpg) +(d) ID: 326247 + +![](images/949e6bef41e345a45c73d9d09ad61e28e959f5bb3c172e14ff7b58426efc7c80.jpg) +(e) ID: 416149 + +![](images/8e042cdbfeb2aa75c73c2d5e938c17f218e83c28d8e6618f1b7e8573feddbc1f.jpg) +(f) + +![](images/fff13b2842815147cfd3d4b49ef0b5c8d462983aec8ade574b71941c70c3c989.jpg) +(g) + +![](images/19b4649dc5193e2e3e0ee6906aaac5aa9fddf2a9cca64f8b72fde147e41ae2b3.jpg) +(h) + +![](images/e92d201e483d90a1e0d17f8c202ec7ae4118fab72849de30b9418a0a491151d0.jpg) +(i) + +![](images/677293938fd74d7bc8b9f3c3ea8cc1df18e08e3ebec5c5329a09f7c8a9ab1d1d.jpg) +(j) + +![](images/f8d60cf4d80ca7e1766cba646ba9a50bcb5940ed891190f255bbe9a17fd568bf.jpg) +(k) +Fig. 1: Examples of ambiguous cases with noisy or incorrect annotations on (a - e) MS COCO [29], (f - j) VinDr-CXR [36] and (k - o) disaster response dataset [19] to identify damaged and undamaged buildings with class names 'd' and 'u', respectively. (a) Mislabelling one object with two bounding boxes. (b - d) Duplicate annotations of visually similar objects (e.g., car/truck, donut/cake and fork/spoon). (d)(e) Incorrect annotation of object classes (e.g., cake labelled as a bowl, only annotating one of the cake decorations as 'person'). (f - o) Annotations from different annotators (represented by the different colours) of the VinDr-CXR and disaster response datasets show significant disagreement. (Details of the datasets are given in Section 4.1.) + +![](images/d8cb7ef14e5c5572a1cd37f29efe4569feeb092a6453cfdbc8bf6a552cb71c2b.jpg) +(1) + +![](images/9fea8b8b2113587af9b95723f320e5cf9a498386174aba9030159c584bcc2971.jpg) +(m) + +![](images/700a5490c136f968efa06c99faec1fd17bef5e9fbe67a58f51a1dc930e29ebf3.jpg) +(n) + +![](images/4e93179fff9ab2410b3a0e9d3b0a501c7408214b322729916452c338258a2221.jpg) +(o) + +from multiple annotators are unified and cleaned as the last step in annotation collection. However, disagreements arise when annotating complex images, making it challenging even for experienced expert annotators to achieve unanimity due to high interobserver variability [26,54] and thus, rendering the role of an annotation arbitrator nearly impossible (see Figure 1). As a result, multiple noisy annotations per image are obtained. Such a setting for annotation crowdsourcing is often called the multi-rater problem. The majority of methods on multi-rater learning have been developed for image classification [20,50,59] and segmentation [4,21,22,24,58,60], while we consider object detection here. + +Several previous methods that exploit crowdsourced annotations have been proposed to improve object detection performance [18,27]. A common approach estimates the ground truth from noisy annotations through majority voting [46] which implicitly assumes that all annotators are equally accurate in their annotations. Alternatively, in [27], the weight of each annotation is predicted based on the annotators' proficiency levels which are assumed to be known a priori; however, this information is difficult to obtain in practice. On the other hand, in [18], their proposed method is designed specifically for Faster R-CNN (FRCNN) [41] and cannot be generalised to other object detectors. Moreover, these methods have been developed and evaluated on distinct, private synthetic + +![](images/09a052399c6674ee465a327714efd9bba0fb41e3aabf7c7bd0d7375e20746c0e.jpg) +Fig. 2: Overall architecture of the proposed BDC. The updating of the aggregator's parameter and the object detector's parameters is repeated iteratively until convergence. Best viewed in colour. + +datasets constructed using different methods. Thus, we cannot directly compare their reported experiment results. + +To address these algorithmic and evaluation limitations, in this work, we first propose a model-agnostic annotation aggregation and object detector training pipeline, named Bayesian Detector Combination (BDC). BDC aggregates noisy bounding boxes and class labels by modelling the annotation reliability of each crowd member for both elements (Figure 2), without requiring any additional knowledge of the ground truth annotations or their quality. Furthermore, due to the limited access to the ground truth of real crowdsourced datasets, we synthesise multiple datasets simulating various crowdsourcing scenarios for systematic evaluation of the previous works and the proposed method. The main contributions of our work include: + +- We propose a model-agnostic Bayesian Detector Combination (BDC) framework that simultaneously infers the annotation quality of each crowd member and infers consensus bounding boxes and soft class labels from noisy, crowdsourced annotations to improve model performance. +- We implement and systematically evaluate the previous methods on real-world and synthetic datasets with crowdsourced annotations which simulate varying crowdsourcing scenarios. +- Extensive experiments conducted demonstrate the superior performance of BDC compared with prior alternatives in key metrics such as the average precision, the scalability with the increasing number of annotators, and the robustness against noisy annotators. + +# 2 Related Work + +Classification with Crowdsourced Annotations. Learning with crowdsourced annotations for classification tasks has been studied in various works. Sheng et al. [46] adopted a simple approach to aggregate multiple labels via majority voting (MV) where the mode class is selected as the ground truth. Several + +works [11,39,57] modelled the sensitivity and specificity of different annotators via confusion matrices obtained through the Expectation-Maximisation algorithm or a Bayesian approach. Kim and Ghahramani [23] introduced a general framework for Bayesian classifier combination (BCC) to model the relationship between the unknown true label and the annotated class label which is extended by Venanzi et al. [51] with the concept of annotator communities. Albarqouni et al. [3] and Isupova et al. [19] extended the idea of previous works [23,39] to convolutional neural networks (CNN) with softmax outputs. Several recent works use a trained classifier to rank consensus labels and incorporate the ranking scores into model learning [7,15]. Another research direction is to take all noisy annotations separately as the input during the training process [10,16,20,43,50,55,59], but Wei et al. [56] concluded that label aggregation is preferred over label separation when there are sufficient number of annotators or the noise rate is moderate. These methods are not directly applicable to object detection tasks since they do not involve regressing the locations of bounding boxes. + +Object Detection with Crowdsourced Annotations. Hu and Meina [18] proposed Crowd R-CNN where one first pre-processes the annotations by combining the bounding boxes using a clustering algorithm, before aggregating the labels and learning the sensitivity of each annotator [39] to different object classes during model training. Crowd R-CNN is tailored for Faster R-CNN (FRCNN) [41] and is incompatible with one-stage object detectors such as YOLOv7 [53]. Le et al. [27] proposed to pre-process annotations using Weighted Boxes Fusion (WBF) [48] and introduced expert agreement re-weighted loss (EARL) to emphasise boxes with higher annotator agreement for model training. It requires access to each annotator's proficiency level to perform WBF and compute EARL, which is often challenging to obtain for real-world crowdsourced data. Compared with the aforementioned methods, the proposed BDC framework can incorporate any object detector in a plug-and-play fashion without requiring any additional inputs such as annotators' proficiency levels. Furthermore, previous works [18, 27] evaluated their methods on different synthetic datasets, with different numbers of annotators and proficiency levels. For example, Hu and Meina [18] performed evaluation on synthetic datasets derived from PASCAL VOC [13] and MS COCO [29]; Le et al. [27] synthesised a MED-MNIST dataset based on MNIST [28] and adopted the VinDr-CXR dataset [36]. Thus, their experiment results cannot be compared directly. + +# 3 Method + +Problem Statement. We assume a training set $D = \{x^{i},y^{i}\}_{i = 1}^{N}$ with pairs of input images $x^{i}$ and crowdsourced annotations $y^{i} = \{y_{m}^{i}\}_{m = 1}^{M^{i}}$ , where $N$ is the total number of training samples and $M^{i}$ is the number of annotations of the $i$ -th image from all annotators. Each annotation is a tuple of a bounding box, class label and the annotator ID, $y_{m}^{i} = \{(b_{m}^{i},c_{m}^{i},k_{m}^{i})\}$ in which $b_{m}^{i}\in \mathbb{R}^{4}$ contains the x-y coordinates of the box centre as well as the width and height + +of the bounding box, $c_{m}^{i} \in \{1, \dots, J\}$ represents the annotated class label of the object and $k_{m}^{i} \in \{1, \dots, K\}$ is the annotator ID. $J$ denotes the number of object classes while $K$ is the total number of annotators. Note that each annotator may annotate a varying number of object instances. For brevity, notations in the rest of this section are simplified to a single training sample $i$ . Our goal is to learn an object detector with noisy crowdsourced annotations while inferring and leveraging annotators' varying annotation abilities. + +# 3.1 Bayesian Detector Combination + +An overview of the proposed crowdsourced annotation aggregation framework is shown in Figure 2. At a high level, the proposed framework consists of four components, namely: the object detector (OD) module, the annotations-predictions matcher, the bounding box aggregator (BBA) and the class label aggregator (CLA). Firstly, we perform a one-to-many matching of the predictions generated by the OD to the crowdsourced annotations. The bounding box error is computed and used to update a Gaussian-Gamma prior distribution of the BBA for each annotator. Next, the class probability predictions by the OD and the class label annotations are used to update a Dirichlet prior in the CLA. The posterior distributions obtained from both aggregators are used to aggregate the bounding boxes and class labels. Finally, we train and optimise the OD's parameters and aggregators' parameters iteratively until convergence. We present a detailed introduction of each component below. + +Object Detector Module. The object detector (OD) module with a parameter set, $\theta_{O}$ , outputs a varying number of $P$ predictions, $\hat{y} = \{(\hat{b}_n,\hat{p}_n)\}_{n = 1}^P$ , given an input image, $x$ , where $\hat{b}_n\in \mathbb{R}^4$ and $\hat{p}_n\in [0,1]^J$ are the predicted bounding box and class probability, respectively. We train the model by minimising an overall loss function: + +$$ +\mathcal {L} _ {\mathcal {D}} = \frac {1}{H} \sum_ {n = 1} ^ {H} \gamma_ {n} \mathcal {L} \left(\mathfrak {b} _ {n}, \rho_ {n}, \hat {y}\right), \tag {1} +$$ + +where $\gamma_{n}$ , $\mathfrak{b}_{n}$ and $\rho_{n}$ are the loss scale factor, the aggregated bounding box, and class probabilities obtained from the two aggregators defined in Section 3.1, respectively. $H$ is the number of aggregated annotations. $\mathcal{L}$ is the loss function that the base OD model adopts. Since the OD operates as a black-box module in the proposed framework, any object detector can be used here as long as it outputs the predicted bounding boxes and the associated class label probabilities. + +Annotations-Predictions Matcher. To update the priors of the aggregators, we first match each annotation to one of the predictions. To find the optimal prediction, $\hat{y}_m^*$ , for each annotation, $y_m$ , we minimise a matching cost function: + +$$ +\hat {y} _ {m} ^ {*} = \underset {\hat {y} _ {n} \in \hat {y}} {\arg \min } \mathcal {L} _ {\text {m a t c h}} (\hat {y} _ {n}, y _ {m}) , \tag {2} +$$ + +where + +$$ +\mathcal {L} _ {\text {m a t c h}} (\hat {y} _ {n}, y _ {m}) = - \hat {p} _ {n (c _ {m})} + \lambda_ {1} \mathcal {L} _ {I o U} (\hat {b} _ {n}, b _ {m}) + \lambda_ {2} | | \hat {b} _ {n} - b _ {m} | | _ {1}. \qquad (3) +$$ + +$\lambda_{1}, \lambda_{2} \in \mathbb{R}$ are hyperparameters and $\mathcal{L}_{IoU}$ is the IoU loss [42]. In short, for each annotation, we aim to find the prediction with a higher predicted probability of the annotated class label, a higher IoU and a smaller L1 distance between the predicted and the annotated bounding box. This is inspired by the set prediction loss in DETR [9]. + +Instead of finding a one-to-one matching with a global minimum matching cost, we are interested in a one-to-many matching that one prediction can be matched to many crowdsourced annotations, with a local minimum matching cost, i.e. finding the prediction with minimum cost for each annotation individually. Moreover, a simple heuristic matching rule is sufficient instead of the Hungarian algorithm in DETR which is more computationally demanding. + +Bounding Box Aggregator. In the Bounding Box Aggregator (BBA), the translation and scaling errors in the x and y-axis of each annotator are modelled with a Gaussian distribution: + +$$ +p \left(\epsilon_ {m} \mid k _ {m} = k, \boldsymbol {\mu}, \boldsymbol {\sigma}\right) = \mathcal {N} \left(\mu^ {k}, \sigma^ {k}\right), \tag {4} +$$ + +where $\epsilon_{m} \in \mathbb{R}^{4}$ contains the translation and scaling errors in the x and y-axis. $\pmb{\mu}$ and $\pmb{\sigma}$ are a collection of Gaussian parameters with $\mu^k$ and $\sigma^k$ representing the mean and covariance of $\epsilon_{m}$ for the annotator $k$ , respectively. The Gaussian distributions have a Gaussian-Gamma prior with hyperparameters $\mu_0^k$ , $v_0^k$ and $\beta_0^k$ [6, 35]. To update the prior distribution, we compute the four error terms between all matched prediction and annotation boxes for each annotator with: + +$$ +\epsilon_ {m} = \left[ \hat {b} _ {m (1)} ^ {*} - b _ {m (1)}, \hat {b} _ {m (2)} ^ {*} - b _ {m (2)}, \hat {b} _ {m (3)} ^ {*} \div b _ {m (3)}, \hat {b} _ {m (4)} ^ {*} \div b _ {m (4)} \right], \quad (5) +$$ + +where $\hat{b}_m^*$ and $b_{m}$ are the matched prediction and annotation boxes, respectively. The subtraction and division operations are designed to make the error terms translation-invariant and scale-invariant. The Gaussian-Gamma prior is updated as follows: + +$$ +\tilde {\mu} ^ {k} = \frac {\mu_ {0} ^ {k} + M ^ {k} \bar {\epsilon} ^ {k}}{M ^ {k} + 1}, \quad (6) \quad \tilde {v} ^ {k} = v _ {0} ^ {k} + \frac {M ^ {k}}{2}, \tag {7} +$$ + +$$ +\tilde {\beta} ^ {k} = \beta_ {0} ^ {k} + \frac {M ^ {k} (\bar {\epsilon} ^ {k} - \mu_ {0} ^ {k}) (\bar {\epsilon} ^ {k} - \mu_ {0} ^ {k}) ^ {\top}}{2 (M ^ {k} + 1)} + \frac {1}{2} \sum_ {m = 1} ^ {M} \mathbb {1} [ k _ {m} = k ] (\epsilon_ {m} - \bar {\epsilon} ^ {k}) (\epsilon_ {m} - \bar {\epsilon} ^ {k}) ^ {\top}, (8) +$$ + +where $M^k$ is the number of annotation-prediction matches associated with the $k$ -th annotator, $\overline{\epsilon}^k$ is the mean of all $\epsilon_{m}$ for the $k$ -th annotator and $\mathbb{1}[\cdot]$ is the indicator function. + +The obtained posterior mean (i.e. $\mu^k = \tilde{\mu}^k$ ) is used to correct the annotated boxes for each annotator: + +$$ +b _ {m} := \left(b _ {m} + \left[ \mu_ {(1)} ^ {k}, \mu_ {(2)} ^ {k}, 0, 0 \right]\right) \odot \left[ 1, 1, \mu_ {(3)} ^ {k}, \mu_ {(4)} ^ {k} \right]. \tag {9} +$$ + +All annotations matched to the same prediction are aggregated via a weighted average using the posterior precision, $P^k = 1 \oslash \text{diag}(\sigma^k)$ , where $\oslash$ denotes the element-wise division and $\sigma^k = \tilde{\beta}^k / \tilde{v}^k$ . The posterior precision, $P^k$ , of each annotator is used as weights to obtain the final boxes, $\{\mathfrak{b}_n\}_{n=1}^H$ , with + +$$ +\mathfrak {b} _ {n} = \frac {\sum_ {(b , k) \in \kappa_ {n}} P ^ {k} \odot b}{\sum_ {(b , k) \in \kappa_ {n}} P ^ {k}}. \tag {10} +$$ + +Here, $H$ indicates the number of predictions matched to at least one annotation which is also equal to the number of aggregated annotations defined in Section 3.1. $\kappa_{n}$ is the set of tuples consisting of the annotated bounding boxes and their annotator ID that are matched to the $n$ -th prediction. Inspired by EARL [27], a loss scale factor, $\gamma_{n}$ , is computed for each aggregated box as the ratio of the number of matched annotations to $K$ and is used to re-weight the loss function of the OD in Equation (1). We refer the readers to Appendix A for further details on the updating procedure of the Bounding Box Aggregator. + +Class Label Aggregator. To aggregate crowdsourced class labels, we extend the Bayesian classifier combination neural network (BCCNet) [19] to the proposed BDC framework. In the class label aggregator (CLA), the annotated class label from each annotator, $c_{m}$ , is modelled as a multinomial distribution conditioning on the true label of the object: + +$$ +p \left(c _ {m} \mid k _ {m} = k, t _ {m} = j, \boldsymbol {\pi}\right) = \pi_ {j, c _ {m}} ^ {k}, \tag {11} +$$ + +where $\pi$ is a collection of confusion matrices, $\pi^k\in \mathbb{R}^{J\times J}$ is the confusion matrix of the $k$ -th annotator and $t_m$ is the ground truth class label. Each row, $\pi_j^k$ , provides the event probabilities of a multinomial distribution for the $k$ -th annotator when $t_m = j$ and is modelled with a Dirichlet prior with hyperparameters $\alpha_{0j}^k = \{\alpha_{0j,1}^k,\dots,\alpha_{0j,J}^k\}$ such that $\pi_j^k |\alpha_{0j}^k\sim \text{Dir}(\alpha_{0j}^k)$ [23]. + +The multinomial posterior distribution of the true label (i.e. the aggregated class label probability) can be obtained as + +$$ +\rho_ {n, j} = \exp \left(\ln \hat {p} _ {n, j} + \sum_ {(c, k) \in \tilde {\kappa} _ {n}} \mathbb {E} _ {\pi_ {j} ^ {k}} \ln \pi_ {j, c} ^ {k}\right), \tag {12} +$$ + +where $\tilde{\kappa}_n$ is the set of tuples consisting of the annotated class labels and their annotator ID that are matched to the $n$ -th prediction. $\hat{p}_{n,j}$ is the OD's class- $j$ probability of the $n$ -th prediction. Note that the CLA outputs class probabilities (soft labels) unlike other methods (e.g. MV) that output a single class (hard label). Appendix B provides the derivation for Equation 12 and a detailed discussion of the updating steps of the Class Label Aggregator. + +Training. The optimal OD's parameters, $\theta_{O}$ , and aggregators' parameters, $\pmb{\mu}$ , $\pmb{v}$ , $\pmb{\beta}$ and $\pmb{\alpha}$ , are learnt iteratively in the BDC framework. Firstly, we initialise + +the parameters either randomly or with prior knowledge of the domain. Next, we train the OD by optimising the loss function, $\mathcal{L}_{\mathcal{D}}$ , with the aggregated annotations obtained using the current values of the aggregators' parameters. After one epoch of training, we obtain the OD's predictions using the current value of $\theta_{O}$ to update the prior distributions of aggregators. These two steps are repeated iteratively until the parameters converge. See Appendix C for a further discussion of the training steps. + +# 4 Experiments + +Training details. To demonstrate the generalisability and robustness of BDC in incorporating different object detection algorithms, three common object detection algorithms are used in our experiments, YOLOv7 [53], FRCNN [41] with the ResNet50 backbone and EVA [14]. During training, we randomly apply data augmentation such as rotation, translation, zooming and horizontal flipping to the training images and annotations. $\lambda_{1},\lambda_{2}$ in Equation (3) are set to 2 and 5, respectively, following Carion et al. [9]. Lastly, the Gaussian-Gamma prior of the BBA, $\pmb{\mu}_{0}$ , $\pmb{v}_{0}$ and $\beta_0$ , are set to 0, 10 and 0.5, respectively, while the Dirichlet prior of the CLA, $\alpha_0$ , is initialised to 1 and each diagonal entry is set to 10. The complete training configuration can be found in Appendix D. + +Comparison against other aggregation methods. We compare the performance of BDC against multiple baselines as shown in Tables 1 to 4. 'NA' denotes the process of training the model with all the crowdsourced annotations as the ground truth annotations [56]. 'MV' stands for majority voting where we apply clustering to the crowdsourced annotations, then aggregate the class labels via MV and aggregate the bounding boxes as the biggest box where the majority had annotated pixel-wise, following the practice in [46]. We also implement Crowd R-CNN [18] and WBF-EARL [27] designed for crowdsourced object detection. For WBF-EARL [27] which requires the knowledge of annotators' proficiency levels, we followed the settings in [27] for the VinDR-CXR dataset while for the synthetic datasets, they are set to the overall accuracy of the annotators. + +# 4.1 Real Crowdsourced Datasets + +VinDr-CXR. We evaluate the annotation aggregation methods on VinDrcxR [36], a thoracic abnormalities detection dataset. VinDr-CXR consists of 14 abnormalities classes and contains 15,000 training and 3,000 testing chest radiograph images. The training images are annotated by 17 experienced radiologists where at least three annotators are randomly assigned to annotate each image. Figures 1f to 1j show examples of the annotations with high bounding box and class label disagreement by the annotators. Furthermore, the testing images are annotated with the consensus of five expert radiologists through a two-stage labelling process and the test set annotations are not publicly available. The predictions are submitted to the VinBigData Kaggle [12] to obtain mean average precision (AP) at 0.4 IoU threshold on the private test set. + +Table 1: AP-4 results on VinDr-CXR [36] dataset obtained from the private score of VinBigData Kaggle competition [12]. Crowd R-CNN can only integrate FRCNN, hence there are no results reported for the two other entries corresponding to Crowd R-CNN. + +
MethodTest AP-4
YOLOv7FRCNNEVA
NA17.417.27.8
MV13.916.38.2
Crowd R-CNN [18]-16.7-
WBF-EARL [27]16.417.08.4
BDC (ours)19.217.98.9
+ +![](images/27d6d5e0ce6ea3eb674e5b640ea66d0fb302820d3a10fd5aa9d1791cd686ee87.jpg) +(a) NA + +![](images/83cedb638eeeab9d7d64990fdfdc8764761bfecc381e1887ff342da22235edd2.jpg) +(b) MV + +![](images/291058eff251e95d3af1d875e90d621628439a2b50e9c3bf8ae29686673dc5d7.jpg) +(c) WBF-EARL +Fig. 3: Comparison of aggregated labels from different methods on the VinDr-CXR [36]. For WBF-EARL [27], the number beside the class label is the annotators' level of agreement while for Crowd R-CNN [18] and BDC, the number indicates the class probability. For NA, the colours represent the different annotators. + +![](images/f90afadb7091e5e1cd925559bdc367793f7da1b9ae43343c7b16bda3ddec5519.jpg) +(d) Crowd R-CNN + +![](images/d03061ac0e82e76322ef74896bf685d31fe6288b1f63cb222fcefe2a401f31a3.jpg) +(e) BDC (ours) + +Disaster Response Dataset. We further use the disaster response dataset [19] where the task is to detect undamaged and damaged buildings on satellite images taken after hurricane Maria in 2017. The dataset consisting of 464 satellite images of Dominica is annotated by 13 semi-professional paid crowd members to ensure high labelling quality. Despite this effort, labels from different crowd members vary significantly (Figures 1k to 1o). Consequently, the dataset does not contain ground truth labels. Thus, its qualitative results are presented in Appendix E. + +Quantitative Results. Table 1 presents the AP.4 metric obtained from the VinBigData Kaggle competition and shows that the proposed BDC leads to higher AP.4 than other methods across different OD models on the VinDr-CXR test set. As Crowd R-CNN can only integrate FRCNN, there are no results reported for the two other entries corresponding to Crowd R-CNN. + +Qualitative Results. We visualise the results of the aggregated annotations from each method on the VinDr-CXR dataset in Figure 3. Since the true ground truth of the datasets is not provided, we use 'MV' as the best approximation of the ground truth for this qualitative assessment and observe that BDC produces more accurate aggregation compared to the other methods. + +# 4.2 Synthetic Datasets + +To alleviate the limitations in the available real-world crowdsourced datasets including the lack of ground truth labels, we synthesise several synthetic datasets with crowdsourced annotations to enable the systematic evaluation of the proposed method. This makes available the ground truth annotations to allow us to obtain detailed test evaluation metrics (e.g., AP $^{5}$ , AP $^{75}$ , AP $^{5::95}$ ) and verify the performance of the proposed method by evaluating the aggregated crowdsourced annotations with the true ground truth annotations of the training dataset (indicated as 'Train Aggregation AP' in Tables 2 to 4). Two common object detection datasets are used to construct the synthetic datasets: (1) Pascal Visual Object Classes (VOC) 2007 [13], a 20-class object detection dataset which consists of 5,011 and 4,952 images for training and testing, respectively; (2) Microsoft Common Objects in Context (COCO) 2017 [29], a large-scale 80-class object detection dataset consisting of 118,287 training and 5,000 validation images. + +Synthetic settings. Four synthetic settings are investigated in this work: (1) VOC-FULL and (2) COCO-FULL where 10 annotators with average skill levels annotate the whole datasets; (3) VOC-MIX where the first 5 annotators are experts while the remaining 20 annotators have an average skill level. All annotators annotate the whole dataset since Pascal VOC is a small dataset; and (4) COCO-MIX where 20 experts, 550 annotators with average skill level and 330 annotators with poor skill level work together to annotate the MS COCO dataset. Additionally, the last 100 are random annotators, i.e., the annotated class labels and bounding boxes are drawn from a uniform distribution. Since MS COCO is a large-scale dataset, we set that each annotator annotates a small subset of images, i.e., the coverage for the expert, average, poor and random annotators are $5\%$ , $1\%$ , $0.5\%$ and $1\%$ , respectively. Note that all annotators in these settings are simulated using deep learning models (See Appendix F for further discussions of the synthesis method and pseudocode). Additional experiments where we directly used the object detectors' predictions as synthetic crowdsourced annotations can be found in Appendix H. + +Quantitative Results. Tables 2 to 4 report the experiment results of YOLOv7, FRCNN and EVA trained with different aggregation methods on all synthetic settings, respectively. Additionally, we trained each model with ground truth annotations of each dataset (denoted by 'GT') to obtain reference optimal model performance. The aggregation APs of the train dataset for NA and WBF-EARL are low because NA directly uses all the crowdsourced annotations while WBF outputs multiple bounding boxes per class per object instance, leading to a much lower precision (see Figure 4 for examples). We do not include performance from Crowd R-CNN for the COCO-MIX setting as its training time increases linearly with $K$ which means that the training time is prohibitively high for this dataset where $K = 1,000$ . In comparison, the extra computational cost introduced by the proposed BDC is small when compared to the much heavier training cost of the OD backbone (see Section 4.4 for an analysis of the computational cost). + +Table 2: Test and train aggregation AP results for YOLOv7 model trained with different annotation aggregation methods on synthesised and ground truth datasets. + +
DatasetMethodTest APTrain Aggregation AP
AP.5AP.75AP.5:95AP.5AP.75AP.5:95
VOC-FULLNA59.745.340.710.29.99.9
MV72.252.347.362.048.548.5
WBF-EARL [27]69.352.345.722.316.414.5
BDC (ours)84.963.554.895.080.666.3
VOC-MIXNA46.127.627.44.44.34.3
MV77.161.053.775.770.668.6
WBF-EARL [27]61.047.240.927.924.920.0
BDC (ours)86.269.058.897.694.977.2
VOCGT87.774.065.3100100100
COCO-FULLNA53.441.437.48.68.28.2
MV61.946.442.752.544.443.3
WBF-EARL [27]55.642.538.914.68.98.6
BDC (ours)65.049.644.583.666.156.9
COCO-MIXNA53.441.237.58.57.98.0
MV60.946.442.445.842.341.1
WBF-EARL [27]56.242.639.112.17.07.0
BDC (ours)64.849.344.480.562.954.3
COCOGT67.552.848.6100100100
+ +Tables 2 to 4 demonstrate the superior performance of BDC compared to other methods from simpler synthetic settings to more complex synthetic crowdsourced annotations, i.e. from VOC-FULL to COCO-MIX. In addition, BDC achieves the highest train aggregation AP compared to other methods across all settings. + +Qualitative Results. Figure 4 visualises the results of the aggregated annotations from each method on the VOC-MIX synthetic dataset. WBF performs box fusion on a per-class basis and results in significantly more boxes compared to other methods (excluding NA) which contributed to the low train aggregation AP and noisy labels during model training. On the other hand, although both Crowd R-CNN and BDC output class probabilities, the probability distribution of Crowd R-CNN is heavily skewed which is similar to hard labels, while BDC outputs the learnt class probabilities and can benefit from these soft labels during model training [34]. Besides, BDC can cope with failure cases of its aggregator as shown by the lower class probability of incorrect aggregation. Additional qualitative results can be found in Appendix I. + +# 4.3 Scalability and Robustness of BDC + +We analyse the scalability and robustness of the aggregation methods using the Pascal VOC dataset. YOLOv7 is selected as the OD because it is lightweight and fast to train. Figure 5a demonstrates the effect of varying numbers of annotators, $K$ , where all annotations are synthesised using annotators of average skill level. Note that training YOLOv7 with the NA method results in a GPU out-of-memory error when $K \geq 100$ . When the number of annotators increases, the + +Table 3: Test and train aggregation AP results for FRCNN model trained with different annotation aggregation methods on synthesised and ground truth datasets. Crowd R-CNN has a training time that increases linearly with $K$ , hence there is no result reported for COCO-MIX dataset with $K = 1,000$ . + +
DatasetMethodTest APTrain Aggregation AP
AP·5AP·75AP·5:95AP·5AP·75AP·5:95
VOC-FULLNA47.411.319.410.29.99.9
MV79.753.548.862.048.548.5
Crowd R-CNN [18]71.633.930.158.343.031.5
WBF-EARL [27]71.837.738.722.316.414.5
BDC (ours)83.756.351.095.080.466.3
VOC-MIXNA34.14.412.24.44.34.3
MV82.057.151.975.770.668.6
Crowd R-CNN [18]70.231.229.360.644.231.0
WBF-EARL [27]69.836.537.427.924.920.0
BDC (ours)84.759.352.997.694.676.9
VOCGT84.662.155.2100100100
COCO-FULLNA39.712.017.78.68.28.2
MV55.635.733.452.544.443.3
Crowd R-CNN [18]48.527.124.249.536.233.7
WBF-EARL [27]51.928.028.414.68.98.6
BDC (ours)56.635.333.583.366.056.7
COCO-MIXNA43.917.521.48.57.98.0
MV54.834.532.945.842.341.1
Crowd R-CNN [18]------
WBF-EARL [27]51.928.028.312.17.07.0
BDC (ours)56.134.633.080.362.754.1
COCOGT58.038.535.7100100100
+ +performance of YOLOv7 increases for both BDC and MV while the performance of both NA and WBF-EARL decreases when $K > 50$ . In addition, BDC has an average AP.5:95 improvement of 4.35 over MV. + +We also investigate the effect of changing the percentage of noisy annotators while setting $K$ to 25. Here, noisy annotators are defined as annotators with a poor skill level while the other annotators are annotators with expert skill level. As illustrated in Figure 5b, we found that AP $^{5:95}$ of all aggregation methods reduce when we increase the percentage of noisy annotators. However, MV drops at a faster rate compared to the proposed method when the noisy annotator percentage is over $50\%$ . We observe that BDC remains robust even when $100\%$ of the annotators have low skill levels and the AP $^{5:95}$ is 18.5 higher than that obtained with MV. The consistent performance of BDC (up to $80\%$ noise) can be attributed to the ability of BDC to correctly capture the ability of each annotator and assign proper weights when aggregating the annotations (See Appendix J for an analysis of the posterior distributions learnt in the aggregators). + +# 4.4 Computational Cost + +Table 5 presents the wall time of each annotation aggregation method spent to pre-process the annotations and train the Faster R-CNN (FRCNN) [41] model for + +Table 4: Test and train aggregation AP results for EVA model trained with different annotation aggregation methods on synthesised and ground truth datasets. + +
DatasetMethodTest APTrain Aggregation AP
AP.5AP.75AP.5:95AP.5AP.75AP.5:95
VOC-FULLNA90.980.969.710.29.99.9
MV91.082.271.262.048.548.5
WBF-EARL [27]93.083.571.122.316.414.5
BDC (ours)94.285.671.795.381.266.7
VOC-MIXNA86.665.757.54.44.34.3
MV91.783.072.975.770.668.6
WBF-EARL [27]91.382.268.727.924.920.0
BDC (ours)94.986.573.898.895.077.3
VOCGT94.988.178.0100100100
COCO-FULLNA71.855.750.68.68.28.2
MV74.861.054.452.544.443.3
WBF-EARL [27]74.758.451.814.68.98.6
BDC (ours)78.062.755.183.365.155.3
COCO-MIXNA73.357.251.78.57.98.0
MV75.060.654.245.842.341.1
WBF-EARL [27]75.259.352.912.17.07.0
BDC (ours)77.162.554.879.962.253.2
COCOGT78.864.958.8100100100
+ +one epoch on the VOC-MIX synthetic dataset. The wall time is computed with a desktop workstation with an i9-10900X CPU and an RTX 3090 GPU. As shown in Table 5, MV, Crowd R-CNN and WBF-EARL pre-process the annotations before model training while NA and the proposed BDC require no pre-processing step. Crowd R-CNN has the longest training time which increases linearly with the number of annotators, $K$ . On the other hand, the proposed BDC performs the annotations-predictions matching step and the updating steps of the aggregators' parameters every training epoch and therefore has a slightly longer training time compared to NA, MV and WBF-EARL. However, this increase in training time is small when compared to the much higher training cost of the OD + +![](images/093aae14ab363fa4e3b3131dd22cce74bb1972ca87d19eb8824a7afecef186e8.jpg) +(a) NA + +![](images/def6a5bc2d50d1d72a35dfb1849b5f556020f7051bc8e9d8a28769154e8f2b06.jpg) +(b) MV +(c) WBF-EARL + +![](images/e6cc64f3f1cbea29b9cfa6a1a4ae3daa9a1cd47df2c99922548f9cc98e7c4836.jpg) +Fig. 4: Comparison of aggregated labels from different methods on the VOC-MIX synthetic dataset. For WBF-EARL [27], the number beside the class label is the annotators' level of agreement while for Crowd R-CNN [18] and BDC, the number indicates the class probability. For NA, the colours represent the different annotators. + +![](images/1c3e1b0d2274b0cd22b2237aff5ca34a32310e431309087b4cf609ac93475bc9.jpg) +(d) Crowd R-CNN + +![](images/2f71fa5cd76551269c73ecefd75a9b3fb46fef874ecfd05e22d63ed71856b78a.jpg) +(e) BDC (ours) + +![](images/225f8100c52b5c363bd24eda2f5c284101bf1679f99ed6103d25bf8ecd789c77.jpg) +(f) Ground truth + +![](images/0364a45c8b39f31bd95e9f88021654048f93d6fc040889b4a71f36328d5aa8ba.jpg) +(a) Scalability of BDC + +![](images/e5673ad679f6f627753d37a472c6491c794e5b0a36179bf78a139e8a5537a0f5.jpg) +(b) Robustness of BDC +Fig. 5: (a) Effect of the number of annotators, $K$ , and (b) effect of varying percentage of noisy annotators (with $K = 25$ ) on the test AP $^{5:95}$ of YOLOv7. Results for the 'NA' method are unavailable when $K \geq 100$ due to GPU out-of-memory error. + +Table 5: Annotations pre-processing time and FRCNN model training time per epoch of different annotation aggregation methods on the VOC-MIX dataset. The wall time is computed on a desktop workstation with an i9-10900X CPU and an RTX 3090 GPU. + +
MethodPre-processing time (s)Training time (s)
NA0.0537.0
MV19.4505.4
Crowd R-CNN [18]15.61175.5
WBF-EARL [27]4.9518.5
BDC (ours)0.0662.9
+ +module (618.6s OD training time out of 662.9s). Note that the model inference time is unaffected by the type of annotation aggregation method. + +# 5 Conclusion + +We first presented a general annotation aggregation and object detection framework, named Bayesian Detector Combination (BDC), for learning a robust object detector from noisy crowdsourced annotations. BDC infers the consensus bounding boxes and corresponding soft labels without the need for ground truth annotations. BDC is model-agnostic and we demonstrated its superior performance by integrating it with three popular object detector backbones, namely YOLOv7, FRCNN, and EVA. Furthermore, we constructed four synthetic datasets to simulate different crowdsourcing scenarios involving annotators of various skill levels, enabling the systematic evaluation of the proposed BDC. We demonstrated the improved performance from BDC on real-world datasets and all the synthetic datasets when compared to other object detectors designed for crowdsourced annotations. We showed that BDC can scale against varying numbers of annotators and is robust against annotators with low annotation reliability. + +Acknowledgements. Z.Q. T. acknowledges the support of an Industrial Cooperative Awards in Science and Engineering (ICASE) studentship from the Engineering and Physical Sciences Research Council (EPSRC) for this work. G. C. acknowledges the support of the Engineering and Physical Sciences Research Council (EPSRC) through grant EP/Y018036/1 and the Australian Research Council (ARC) through grant FT190100525. The authors would like to thank the Satellite Applications Catapult for the provision of the disaster response dataset. + +# References + +1. Albarqouni, S., Baur, C., Achilles, F., Belagiannis, V., Demirci, S., Navab, N.: AggNet: Deep learning from crowds for mitosis detection in breast cancer histology images. IEEE Trans. Med. Imag. 35(5), 1313-1321 (May 2016) +2. Baumgartner, C.F., Tezcan, K.C., Chaitanya, K., Hötter, A.M., Muehlematter, U.J., Schawkat, K., Becker, A.S., Donati, O., Konukoglu, E.: PHiSeg: Capturing uncertainty in medical image segmentation. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 119-127. Shenzhen, China (2019) +3. Bernard, J.M., Smith, A.F.M.: Bayesian Theory, chap. 3, pp. 136-138. Wiley, West Sussex, UK (1993) +4. Bernhardt, M., Castro, D.C., Tanno, R., Schwaighofer, A., Tezcan, K.C., Monteiro, M., Bannur, S., Lungren, M.P., Nori, A., Glocker, B., Alvarez-Valle, J., Oktay, O.: Active label cleaning for improved dataset quality under resource constraints. Nature Commun. 13(1), 1161 (Mar 2022) +5. Budd, S., Day, T., Simpson, J., Lloyd, K., Matthew, J., Skelton, E., Razavi, R., Kainz, B.: Can non-specialists provide high quality gold standard labels in challenging modalities? In: Med. Imag. Comput. Comput. Assist. Interv. pp. 251-262. Strasbourg, France (2021) +6. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229 (2020) +7. Chen, Z., Wang, H., Sun, H., Chen, P., Han, T., Liu, X., Yang, J.: Structured probabilistic end-to-end learning from crowds. In: Int. Joint Conf. Artif. Intell. pp. 1512-1518 (2020) +8. Dawid, A.P., Skene, A.M.: Maximum likelihood estimation of observer error-rates using the EM algorithm. J. Roy. Statist. Soc. Ser. C (Appl. Statist.) 28(1), 20-28 (1979) +9. DungNB, Nguyen, H.Q., Elliott, J., KeepLearning, Nhan, N.T., Culliton, P.: VinBigData chest x-ray abnormalities detection (2020), https://kaggle.com/competitions/ vinbigdata-chest-xray-abnormalities-detection +10. Everingham, M., Luc Van Gool, C.K.I.W., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2007 (VOC2007). Int. J. Comput. Vis. 88, 303-338 (Jun 2010) +11. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA: Exploring the limits of masked visual representation learning at scale. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19358-19369. Vancouver, Canada (2023) + +12. Goh, H.W., Tkachenko, U., Mueller, J.: CROWDLAB: Supervised learning to infer consensus labels and quality scores for data with multiple annotators (2022), arXiv:2210.06812 +13. Guan, M.Y., Gulshan, V., Dai, A.M., Hinton, G.E.: Who said what: Modeling individual labelers improves classification. In: AAAI Conf. Artif. Intell. pp. 3109-3118. San Francisco, CA, USA (2017) +14. Hu, Y., Meina, S.: Crowd R-CNN: An object detection model utilizing crowdsourced labels. In: Int. Conf. Vis. Image Sig. Process. pp. 1-7. Bangkok, Thailand (2020) +15. Isupova, O., Li, Y., Kuzin, D., Roberts, S.J., Willis, K.J., Reece, S.: BCCNet: Bayesian classifier combination neural network. In: Neural Inf. Process. Syst. Workshop Mach. Learn. Develop. Montréal, Canada (2018) +16. Jensen, M.H., Jorgensen, D.R., Jalaboi, R., Hansen, M.E., Olsen, M.A.: Improving uncertainty estimation in convolutional neural networks using inter-rater agreement. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 540-548. Shenzhen, China (2019) +17. Ji, W., Yu, S., Wu, J., Ma, K., Bian, C., Bi, Q., Li, J., Liu, H., Cheng, L., Zheng, Y.: Learning calibrated medical image segmentation via multi-rater agreement modeling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12336-12346 (2021) +18. Jungo, A., Meier, R., Ermis, E., Blatti-Moreno, M., Herrmann, E., Wiest, R., Reyes, M.: On the effect of inter-observer variability for a reliable estimation of uncertainty of medical image segmentation. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 682-690. Granada, Spain (2018) +19. Kim, H., Ghahramani, Z.: Bayesian classifier combination. In: Int. Conf. Artif. Intell. Statist. pp. 619-627. La Palma, Canary Islands (2012) +20. Kohl, S.A.A., Romera-Paredes, B., Meyer, C., Fauw, J.D., Ledsam, J.R., Maier-Hein, K.H., Eslami, S.M.A., Rezende, D.J., Ronneberger, O.: A probabilistic u-net for segmentation of ambiguous images. In: Adv. Neural Inform. Process. Syst. pp. 6965–6975. Montréal, Canada (2018) +21. Lazarus, E., Mainiero, M.B., Schepps, B., Koelliker, S.L., Livingston, L.S.: BI-RADS lexicon for US and mammography: interobserver variability and positive predictive value. Radiology 239(2), 385-391 (Mar 2006) +22. Le, K., Tran, T., Pham, H., Nguyen Trung, H., Le, T., Nguyen, H.Q.: Learning from multiple expert annotators for enhancing anomaly detection in medical image analysis. IEEE Access 11, 14105-14114 (2023) +23. LeCun, Y., Cortes, C., Burges, C.: MNIST handwritten digit database (2010), http://yann.lecun.com/exdb/mnist +24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Zurich, Switzerland (2014) +25. Maier-Hein, L., Mersmann, S., Kondermann, D., Stock, C., Kenngott, H.G., Sanchez, A., Wagner, M., Preukschas, A., Wekerle, A.L., Helfert, S., Bodenstedt, S., Speidel, S.: Crowdsourcing for reference correspondence generation in endoscopic images. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 349-356. Boston, MA, USA (2014) +26. Müller, R., Kornblith, S., Hinton, G.: When does label smoothing help? In: Adv. Neural Inform. Process. Syst. pp. 4694-4703. Vancouver, Canada (2019) +27. Murphy, K.: Conjugate bayesian analysis of the gaussian distribution (2007), https://www.cs.ubc.ca/~murphyk/Papers/bayesGauss.pdf + +28. Nguyen, H.Q., Lam, K., Le, L.T., Pham, H.H., Tran, D.Q., Nguyen, D.B., Le, D.D., Pham, C.M., Tong, H.T.T., Dinh, D.H., Do, C.D., Doan, L.T., Nguyen, C.N., Nguyen, B.T., Nguyen, Q.V., Hoang, A.D., Phan, H.N., Nguyen, A.T., Ho, P.H., Ngo, D.T., Nguyen, N.T., Nguyen, N.T., Dao, M., Vu, V.: VinDr-CXR: An open dataset of chest x-rays with radiologist's annotations. Sci. Data 9(1), 429 (Jul 2022) +29. Nowruzi, F.E., Kapoor, P., Kolhatkar, D., Hassanat, F., Laganiere, R., Rebut, J.: How much real data do we actually need: Analyzing object detection performance using synthetic and real data. In: Int. Conf. Mach. Learn. Workshop AI Auton. Driving. Long Beach, CA, USA (2019) +30. Raykar, V.C., Yu, S., Zhao, L.H., Valadez, G.H., Florin, C., Bogoni, L., Moy, L., Raykar, C., Valadez, G.H.: Learning from crowds. J. Mach. Learn. Res. 11(43), 1297-1322 (Aug 2010) +31. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: Adv. Neural Inform. Process. Syst. pp. 91-99. Montréal, Canada (2015) +32. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 658-666. Long Beach, CA, USA (2019) +33. Rodrigues, F., Pereira, F.C.: Deep learning from crowds. In: AAAI Conf. Artif. Intell. pp. 1611-1618. New Orleans, LA, USA (2018) +34. Sheng, V.S., Provost, F., Ipeirotis, P.G.: Get another label? improving data quality and data mining using multiple, noisy labelers. In: ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. pp. 614-622. Las Vegas, NV, USA (2008) +35. Solovyev, R., Wang, W., Gabruseva, T.: Weighted boxes fusion: Ensembling boxes from different object detection models. Image Vis. Comput. 107, 104-117 (Mar 2021) +36. Sudre, C.H., Anson, B.G., Ingala, S., Lane, C.D., Jimenez, D., Haider, L., Varsavsky, T., Tanno, R., Smith, L., Ourselin, S., Jäger, R.H., Cardoso, M.J.: Let's agree to disagree: Learning highly debatable multirater labelling. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 665-673. Shenzhen, China (2019) +37. Venanzi, M., Guiver, J., Kazai, G., Kohli, P., Shokouhi, M.: Community-based bayesian aggregation models for crowdsourcing. In: Int. Conf. World Wide Web. pp. 155-164. Seoul, Korea (2014) +38. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7464-7475. Vancouver, Canada (2023) +39. Watadani, T., Sakai, F., Johkoh, T., Noma, S., Akira, M., Fujimoto, K., Bankier, A.A., Lee, K.S., Müller, N.L., Song, J.W., Park, J.S., Lynch, D.A., Hansell, D.M., Remy-Jardin, M., Franquet, T., Sugiyama, Y.: Interobserver variability in the CT assessment of honeycombing in the lungs. Radiology 266(3), 936-944 (Dec 2012) +40. Wei, H., Xie, R., Feng, L., Han, B., An, B.: Deep learning from multiple noisy annotators as a union. IEEE Trans. Neural Netw. Learn. Syst. pp. 1-11 (Apr 2022) +41. Wei, J., Zhu, Z., Luo, T., Amid, E., Kumar, A., Liu, Y.: To aggregate or not? learning with separate noisy labels. In: ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. pp. 2523-2535. Long Beach, CA, USA (2023) +42. Whitehill, J., Wu, T.f., Bergsma, J., Movellan, J., Ruvolo, P.: Whose vote should count more: Optimal integration of labels from labelers of unknown expertise. In: Adv. Neural Inform. Process. Syst. pp. 2035-2043. Vancouver, Canada (2009) + +43. Wu, J., Fang, H., Yang, Y., Liu, Y., Gao, J., Duan, L., Yang, W., Xu, Y.: Multi-rater prism: Learning self-calibrated medical image segmentation from multiple raters (2022), arXiv:2212.00601 +44. Yu, S., Zhou, H.Y., Ma, K., Bian, C., Chu, C., Liu, H., Zheng, Y.: Difficulty-aware glaucoma classification with multi-rater consensus modeling. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 741-750 (2020) +45. Zhang, L., Tanno, R., Xu, M.C., Jin, C., Jacob, J., Ciccarelli, O., Barkhof, F., Alexander, D.C.: Disentangling human error from the ground truth in segmentation of medical images. In: Adv. Neural Inform. Process. Syst. pp. 15750-15762 (2020) +46. Zhu, X., Vondrick, C., Ramanan, D., Fowlkes, C.: Do we need more training data or better models for object detection? In: Brit. Mach. Vis. Conf. Surrey, UK (2012) \ No newline at end of file diff --git a/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/images.zip b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f40288523f3d99b16bd64bb2453423705c1e63b0 --- /dev/null +++ b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a1da423d26d0f582631025527d6bc468db7dcaf34c882f2b4a998d58680e0be +size 691997 diff --git a/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/layout.json b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de7d4d6ca3dc9a82573ec5e3acd90fc4a3cd17f0 --- /dev/null +++ b/2024/Bayesian Detector Combination for Object Detection with Crowdsourced Annotations/layout.json @@ -0,0 +1,10677 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 457, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 457, + 146 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 457, + 146 + ], + "type": "text", + "content": "Bayesian Detector Combination for Object Detection with Crowdsourced Annotations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "spans": [ + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "text", + "content": "Zhi Qin " + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{Tan}^{1}" + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "text", + "content": ", Olga Isupova" + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "text", + "content": ", Gustavo Carneiro" + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "text", + "content": ", Xiatian Zhu" + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "text", + "content": ", and Yunpeng Li" + }, + { + "bbox": [ + 142, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 178, + 201, + 435, + 246 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 178, + 201, + 435, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 201, + 435, + 224 + ], + "spans": [ + { + "bbox": [ + 178, + 201, + 435, + 224 + ], + "type": "text", + "content": "1 University of Surrey, Surrey, UK {z.tan,g.carneiro, xiatian.zhu, yunpeng.li}@surrey.ac.uk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 224, + 380, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 224, + 380, + 246 + ], + "spans": [ + { + "bbox": [ + 233, + 224, + 380, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 233, + 224, + 380, + 246 + ], + "type": "text", + "content": " University of Oxford, Oxford, UK olga.isupova@ouce.ox.ac.uk" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 279, + 455, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 279, + 455, + 499 + ], + "spans": [ + { + "bbox": [ + 159, + 279, + 455, + 499 + ], + "type": "text", + "content": "Abstract. Acquiring fine-grained object detection annotations in unconstrained images is time-consuming, expensive, and prone to noise, especially in crowdsourcing scenarios. Most prior object detection methods assume accurate annotations; A few recent works have studied object detection with noisy crowdsourced annotations, with evaluation on distinct synthetic crowdsourced datasets of varying setups under artificial assumptions. To address these algorithmic limitations and evaluation inconsistency, we first propose a novel Bayesian Detector Combination (BDC) framework to more effectively train object detectors with noisy crowdsourced annotations, with the unique ability of automatically inferring the annotators' label qualities. Unlike previous approaches, BDC is model-agnostic, requires no prior knowledge of the annotators' skill level, and seamlessly integrates with existing object detection models. Due to the scarcity of real-world crowdsourced datasets, we introduce large synthetic datasets by simulating varying crowdsourcing scenarios. This allows consistent evaluation of different models at scale. Extensive experiments on both real and synthetic crowdsourced datasets show that BDC outperforms existing state-of-the-art methods, demonstrating its superiority in leveraging crowdsourced data for object detection. Our code and data are available at: https://github.com/zhiqin1998/bdc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "spans": [ + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "text", + "content": "Keywords: Bayesian Detector Combination " + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "text", + "content": " Object detection " + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "text", + "content": " Crowdsourced annotations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Despite significant advancements in the development of object detection with natural images [9, 14, 41, 53], existing methods typically assume the availability of large, accurately annotated datasets [37, 61]. Nonetheless, this assumption does not often hold true in practice. For example, accurate labels are often expensive to acquire in many applications. A large number of experts with varying professional skills can be involved to crowdsourced annotations of images sampled with replacement from a large dataset [8, 33, 50]. Usually, the annotations" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 202, + 163 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 202, + 163 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 202, + 163 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 202, + 163 + ], + "type": "image", + "image_path": "5581d5e72fc9c81d38c5028897e9a5518e186141340e1acc021947b725279667.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 163, + 194, + 172 + ], + "lines": [ + { + "bbox": [ + 145, + 163, + 194, + 172 + ], + "spans": [ + { + "bbox": [ + 145, + 163, + 194, + 172 + ], + "type": "text", + "content": "(a) ID: 70282" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 204, + 116, + 271, + 163 + ], + "blocks": [ + { + "bbox": [ + 204, + 116, + 271, + 163 + ], + "lines": [ + { + "bbox": [ + 204, + 116, + 271, + 163 + ], + "spans": [ + { + "bbox": [ + 204, + 116, + 271, + 163 + ], + "type": "image", + "image_path": "05b271ffd556ac007a75bd771bc7ee4d9dad08270afa4c676a571740ba5cb60d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 163, + 265, + 172 + ], + "lines": [ + { + "bbox": [ + 212, + 163, + 265, + 172 + ], + "spans": [ + { + "bbox": [ + 212, + 163, + 265, + 172 + ], + "type": "text", + "content": "(b) ID: 310072" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 273, + 114, + 340, + 163 + ], + "blocks": [ + { + "bbox": [ + 273, + 114, + 340, + 163 + ], + "lines": [ + { + "bbox": [ + 273, + 114, + 340, + 163 + ], + "spans": [ + { + "bbox": [ + 273, + 114, + 340, + 163 + ], + "type": "image", + "image_path": "7c49c0cf0db62ebd904ad3997472f95222e089f38f578d517c36193528ec48b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 163, + 333, + 172 + ], + "lines": [ + { + "bbox": [ + 281, + 163, + 333, + 172 + ], + "spans": [ + { + "bbox": [ + 281, + 163, + 333, + 172 + ], + "type": "text", + "content": "(c) ID: 516471" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 409, + 163 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 409, + 163 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 409, + 163 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 409, + 163 + ], + "type": "image", + "image_path": "3820121acb52d132162c76ac90c25a9dc2bfbf62a610eba4b4327850149e7352.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 163, + 403, + 172 + ], + "lines": [ + { + "bbox": [ + 349, + 163, + 403, + 172 + ], + "spans": [ + { + "bbox": [ + 349, + 163, + 403, + 172 + ], + "type": "text", + "content": "(d) ID: 326247" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 479, + 163 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 479, + 163 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 479, + 163 + ], + "type": "image", + "image_path": "949e6bef41e345a45c73d9d09ad61e28e959f5bb3c172e14ff7b58426efc7c80.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 163, + 471, + 172 + ], + "lines": [ + { + "bbox": [ + 419, + 163, + 471, + 172 + ], + "spans": [ + { + "bbox": [ + 419, + 163, + 471, + 172 + ], + "type": "text", + "content": "(e) ID: 416149" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 135, + 174, + 202, + 224 + ], + "blocks": [ + { + "bbox": [ + 135, + 174, + 202, + 224 + ], + "lines": [ + { + "bbox": [ + 135, + 174, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 135, + 174, + 202, + 224 + ], + "type": "image", + "image_path": "8e042cdbfeb2aa75c73c2d5e938c17f218e83c28d8e6618f1b7e8573feddbc1f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 224, + 175, + 234 + ], + "lines": [ + { + "bbox": [ + 164, + 224, + 175, + 234 + ], + "spans": [ + { + "bbox": [ + 164, + 224, + 175, + 234 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 205, + 174, + 271, + 224 + ], + "blocks": [ + { + "bbox": [ + 205, + 174, + 271, + 224 + ], + "lines": [ + { + "bbox": [ + 205, + 174, + 271, + 224 + ], + "spans": [ + { + "bbox": [ + 205, + 174, + 271, + 224 + ], + "type": "image", + "image_path": "fff13b2842815147cfd3d4b49ef0b5c8d462983aec8ade574b71941c70c3c989.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 225, + 244, + 234 + ], + "lines": [ + { + "bbox": [ + 233, + 225, + 244, + 234 + ], + "spans": [ + { + "bbox": [ + 233, + 225, + 244, + 234 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 273, + 173, + 340, + 224 + ], + "blocks": [ + { + "bbox": [ + 273, + 173, + 340, + 224 + ], + "lines": [ + { + "bbox": [ + 273, + 173, + 340, + 224 + ], + "spans": [ + { + "bbox": [ + 273, + 173, + 340, + 224 + ], + "type": "image", + "image_path": "19b4649dc5193e2e3e0ee6906aaac5aa9fddf2a9cca64f8b72fde147e41ae2b3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 225, + 313, + 234 + ], + "lines": [ + { + "bbox": [ + 301, + 225, + 313, + 234 + ], + "spans": [ + { + "bbox": [ + 301, + 225, + 313, + 234 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 342, + 174, + 408, + 224 + ], + "blocks": [ + { + "bbox": [ + 342, + 174, + 408, + 224 + ], + "lines": [ + { + "bbox": [ + 342, + 174, + 408, + 224 + ], + "spans": [ + { + "bbox": [ + 342, + 174, + 408, + 224 + ], + "type": "image", + "image_path": "e92d201e483d90a1e0d17f8c202ec7ae4118fab72849de30b9418a0a491151d0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 225, + 380, + 234 + ], + "lines": [ + { + "bbox": [ + 371, + 225, + 380, + 234 + ], + "spans": [ + { + "bbox": [ + 371, + 225, + 380, + 234 + ], + "type": "text", + "content": "(i)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 410, + 174, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 410, + 174, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 410, + 174, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 410, + 174, + 479, + 224 + ], + "type": "image", + "image_path": "677293938fd74d7bc8b9f3c3ea8cc1df18e08e3ebec5c5329a09f7c8a9ab1d1d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 225, + 449, + 234 + ], + "lines": [ + { + "bbox": [ + 440, + 225, + 449, + 234 + ], + "spans": [ + { + "bbox": [ + 440, + 225, + 449, + 234 + ], + "type": "text", + "content": "(j)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 135, + 236, + 202, + 281 + ], + "blocks": [ + { + "bbox": [ + 135, + 236, + 202, + 281 + ], + "lines": [ + { + "bbox": [ + 135, + 236, + 202, + 281 + ], + "spans": [ + { + "bbox": [ + 135, + 236, + 202, + 281 + ], + "type": "image", + "image_path": "f8d60cf4d80ca7e1766cba646ba9a50bcb5940ed891190f255bbe9a17fd568bf.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 282, + 175, + 292 + ], + "lines": [ + { + "bbox": [ + 163, + 282, + 175, + 292 + ], + "spans": [ + { + "bbox": [ + 163, + 282, + 175, + 292 + ], + "type": "text", + "content": "(k)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 300, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 401 + ], + "type": "text", + "content": "Fig. 1: Examples of ambiguous cases with noisy or incorrect annotations on (a - e) MS COCO [29], (f - j) VinDr-CXR [36] and (k - o) disaster response dataset [19] to identify damaged and undamaged buildings with class names 'd' and 'u', respectively. (a) Mislabelling one object with two bounding boxes. (b - d) Duplicate annotations of visually similar objects (e.g., car/truck, donut/cake and fork/spoon). (d)(e) Incorrect annotation of object classes (e.g., cake labelled as a bowl, only annotating one of the cake decorations as 'person'). (f - o) Annotations from different annotators (represented by the different colours) of the VinDr-CXR and disaster response datasets show significant disagreement. (Details of the datasets are given in Section 4.1.)" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 205, + 236, + 271, + 281 + ], + "blocks": [ + { + "bbox": [ + 205, + 236, + 271, + 281 + ], + "lines": [ + { + "bbox": [ + 205, + 236, + 271, + 281 + ], + "spans": [ + { + "bbox": [ + 205, + 236, + 271, + 281 + ], + "type": "image", + "image_path": "d8cb7ef14e5c5572a1cd37f29efe4569feeb092a6453cfdbc8bf6a552cb71c2b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 282, + 243, + 292 + ], + "lines": [ + { + "bbox": [ + 233, + 282, + 243, + 292 + ], + "spans": [ + { + "bbox": [ + 233, + 282, + 243, + 292 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 273, + 235, + 340, + 281 + ], + "blocks": [ + { + "bbox": [ + 273, + 235, + 340, + 281 + ], + "lines": [ + { + "bbox": [ + 273, + 235, + 340, + 281 + ], + "spans": [ + { + "bbox": [ + 273, + 235, + 340, + 281 + ], + "type": "image", + "image_path": "9fea8b8b2113587af9b95723f320e5cf9a498386174aba9030159c584bcc2971.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 282, + 314, + 292 + ], + "lines": [ + { + "bbox": [ + 299, + 282, + 314, + 292 + ], + "spans": [ + { + "bbox": [ + 299, + 282, + 314, + 292 + ], + "type": "text", + "content": "(m)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 342, + 235, + 408, + 281 + ], + "blocks": [ + { + "bbox": [ + 342, + 235, + 408, + 281 + ], + "lines": [ + { + "bbox": [ + 342, + 235, + 408, + 281 + ], + "spans": [ + { + "bbox": [ + 342, + 235, + 408, + 281 + ], + "type": "image", + "image_path": "700a5490c136f968efa06c99faec1fd17bef5e9fbe67a58f51a1dc930e29ebf3.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 282, + 381, + 292 + ], + "lines": [ + { + "bbox": [ + 370, + 282, + 381, + 292 + ], + "spans": [ + { + "bbox": [ + 370, + 282, + 381, + 292 + ], + "type": "text", + "content": "(n)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 410, + 235, + 479, + 281 + ], + "blocks": [ + { + "bbox": [ + 410, + 235, + 479, + 281 + ], + "lines": [ + { + "bbox": [ + 410, + 235, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 410, + 235, + 479, + 281 + ], + "type": "image", + "image_path": "4e93179fff9ab2410b3a0e9d3b0a501c7408214b322729916452c338258a2221.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 282, + 450, + 292 + ], + "lines": [ + { + "bbox": [ + 439, + 282, + 450, + 292 + ], + "spans": [ + { + "bbox": [ + 439, + 282, + 450, + 292 + ], + "type": "text", + "content": "(o)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 433, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 541 + ], + "type": "text", + "content": "from multiple annotators are unified and cleaned as the last step in annotation collection. However, disagreements arise when annotating complex images, making it challenging even for experienced expert annotators to achieve unanimity due to high interobserver variability [26,54] and thus, rendering the role of an annotation arbitrator nearly impossible (see Figure 1). As a result, multiple noisy annotations per image are obtained. Such a setting for annotation crowdsourcing is often called the multi-rater problem. The majority of methods on multi-rater learning have been developed for image classification [20,50,59] and segmentation [4,21,22,24,58,60], while we consider object detection here." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Several previous methods that exploit crowdsourced annotations have been proposed to improve object detection performance [18,27]. A common approach estimates the ground truth from noisy annotations through majority voting [46] which implicitly assumes that all annotators are equally accurate in their annotations. Alternatively, in [27], the weight of each annotation is predicted based on the annotators' proficiency levels which are assumed to be known a priori; however, this information is difficult to obtain in practice. On the other hand, in [18], their proposed method is designed specifically for Faster R-CNN (FRCNN) [41] and cannot be generalised to other object detectors. Moreover, these methods have been developed and evaluated on distinct, private synthetic" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z.Q. Tan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 471, + 231 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 471, + 231 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 471, + 231 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 471, + 231 + ], + "type": "image", + "image_path": "09a052399c6674ee465a327714efd9bba0fb41e3aabf7c7bd0d7375e20746c0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 480, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 480, + 271 + ], + "type": "text", + "content": "Fig. 2: Overall architecture of the proposed BDC. The updating of the aggregator's parameter and the object detector's parameters is repeated iteratively until convergence. Best viewed in colour." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 294, + 480, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 317 + ], + "type": "text", + "content": "datasets constructed using different methods. Thus, we cannot directly compare their reported experiment results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 318, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 481, + 437 + ], + "type": "text", + "content": "To address these algorithmic and evaluation limitations, in this work, we first propose a model-agnostic annotation aggregation and object detector training pipeline, named Bayesian Detector Combination (BDC). BDC aggregates noisy bounding boxes and class labels by modelling the annotation reliability of each crowd member for both elements (Figure 2), without requiring any additional knowledge of the ground truth annotations or their quality. Furthermore, due to the limited access to the ground truth of real crowdsourced datasets, we synthesise multiple datasets simulating various crowdsourcing scenarios for systematic evaluation of the previous works and the proposed method. The main contributions of our work include:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 445, + 480, + 575 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 445, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 493 + ], + "type": "text", + "content": "- We propose a model-agnostic Bayesian Detector Combination (BDC) framework that simultaneously infers the annotation quality of each crowd member and infers consensus bounding boxes and soft class labels from noisy, crowdsourced annotations to improve model performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 493, + 480, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 480, + 528 + ], + "type": "text", + "content": "- We implement and systematically evaluate the previous methods on real-world and synthetic datasets with crowdsourced annotations which simulate varying crowdsourcing scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 529, + 480, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 529, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 529, + 480, + 575 + ], + "type": "text", + "content": "- Extensive experiments conducted demonstrate the superior performance of BDC compared with prior alternatives in key metrics such as the average precision, the scalability with the increasing number of annotators, and the robustness against noisy annotators." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 594, + 237, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 237, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 237, + 605 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Classification with Crowdsourced Annotations. Learning with crowdsourced annotations for classification tasks has been studied in various works. Sheng et al. [46] adopted a simple approach to aggregate multiple labels via majority voting (MV) where the mode class is selected as the ground truth. Several" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "text", + "content": "Bayesian Detector Combination" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "works [11,39,57] modelled the sensitivity and specificity of different annotators via confusion matrices obtained through the Expectation-Maximisation algorithm or a Bayesian approach. Kim and Ghahramani [23] introduced a general framework for Bayesian classifier combination (BCC) to model the relationship between the unknown true label and the annotated class label which is extended by Venanzi et al. [51] with the concept of annotator communities. Albarqouni et al. [3] and Isupova et al. [19] extended the idea of previous works [23,39] to convolutional neural networks (CNN) with softmax outputs. Several recent works use a trained classifier to rank consensus labels and incorporate the ranking scores into model learning [7,15]. Another research direction is to take all noisy annotations separately as the input during the training process [10,16,20,43,50,55,59], but Wei et al. [56] concluded that label aggregation is preferred over label separation when there are sufficient number of annotators or the noise rate is moderate. These methods are not directly applicable to object detection tasks since they do not involve regressing the locations of bounding boxes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 548 + ], + "type": "text", + "content": "Object Detection with Crowdsourced Annotations. Hu and Meina [18] proposed Crowd R-CNN where one first pre-processes the annotations by combining the bounding boxes using a clustering algorithm, before aggregating the labels and learning the sensitivity of each annotator [39] to different object classes during model training. Crowd R-CNN is tailored for Faster R-CNN (FRCNN) [41] and is incompatible with one-stage object detectors such as YOLOv7 [53]. Le et al. [27] proposed to pre-process annotations using Weighted Boxes Fusion (WBF) [48] and introduced expert agreement re-weighted loss (EARL) to emphasise boxes with higher annotator agreement for model training. It requires access to each annotator's proficiency level to perform WBF and compute EARL, which is often challenging to obtain for real-world crowdsourced data. Compared with the aforementioned methods, the proposed BDC framework can incorporate any object detector in a plug-and-play fashion without requiring any additional inputs such as annotators' proficiency levels. Furthermore, previous works [18, 27] evaluated their methods on different synthetic datasets, with different numbers of annotators and proficiency levels. For example, Hu and Meina [18] performed evaluation on synthetic datasets derived from PASCAL VOC [13] and MS COCO [29]; Le et al. [27] synthesised a MED-MNIST dataset based on MNIST [28] and adopted the VinDr-CXR dataset [36]. Thus, their experiment results cannot be compared directly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 566, + 202, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 202, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 202, + 578 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": "Problem Statement. We assume a training set " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "D = \\{x^{i},y^{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " with pairs of input images " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " and crowdsourced annotations " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "y^{i} = \\{y_{m}^{i}\\}_{m = 1}^{M^{i}}" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " is the total number of training samples and " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "M^{i}" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " is the number of annotations of the " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": "-th image from all annotators. Each annotation is a tuple of a bounding box, class label and the annotator ID, " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_{m}^{i} = \\{(b_{m}^{i},c_{m}^{i},k_{m}^{i})\\}" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "b_{m}^{i}\\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " contains the x-y coordinates of the box centre as well as the width and height" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z.Q. Tan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "of the bounding box, " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "c_{m}^{i} \\in \\{1, \\dots, J\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " represents the annotated class label of the object and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "k_{m}^{i} \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " is the annotator ID. " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " denotes the number of object classes while " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " is the total number of annotators. Note that each annotator may annotate a varying number of object instances. For brevity, notations in the rest of this section are simplified to a single training sample " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ". Our goal is to learn an object detector with noisy crowdsourced annotations while inferring and leveraging annotators' varying annotation abilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 217, + 321, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 321, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 321, + 228 + ], + "type": "text", + "content": "3.1 Bayesian Detector Combination" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 392 + ], + "type": "text", + "content": "An overview of the proposed crowdsourced annotation aggregation framework is shown in Figure 2. At a high level, the proposed framework consists of four components, namely: the object detector (OD) module, the annotations-predictions matcher, the bounding box aggregator (BBA) and the class label aggregator (CLA). Firstly, we perform a one-to-many matching of the predictions generated by the OD to the crowdsourced annotations. The bounding box error is computed and used to update a Gaussian-Gamma prior distribution of the BBA for each annotator. Next, the class probability predictions by the OD and the class label annotations are used to update a Dirichlet prior in the CLA. The posterior distributions obtained from both aggregators are used to aggregate the bounding boxes and class labels. Finally, we train and optimise the OD's parameters and aggregators' parameters iteratively until convergence. We present a detailed introduction of each component below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "text", + "content": "Object Detector Module. The object detector (OD) module with a parameter set, " + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\theta_{O}" + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "text", + "content": ", outputs a varying number of " + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "text", + "content": " predictions, " + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\{(\\hat{b}_n,\\hat{p}_n)\\}_{n = 1}^P" + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "text", + "content": ", given an input image, " + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\hat{b}_n\\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\hat{p}_n\\in [0,1]^J" + }, + { + "bbox": [ + 130, + 407, + 482, + 470 + ], + "type": "text", + "content": " are the predicted bounding box and class probability, respectively. We train the model by minimising an overall loss function:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 245, + 475, + 481, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 475, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 245, + 475, + 481, + 509 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {D}} = \\frac {1}{H} \\sum_ {n = 1} ^ {H} \\gamma_ {n} \\mathcal {L} \\left(\\mathfrak {b} _ {n}, \\rho_ {n}, \\hat {y}\\right), \\tag {1}", + "image_path": "39a685d2e04bfff5b43c29961c9e346afbf44028f0665127aeabddfc0ddbb267.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\gamma_{n}" + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\mathfrak{b}_{n}" + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\rho_{n}" + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "text", + "content": " are the loss scale factor, the aggregated bounding box, and class probabilities obtained from the two aggregators defined in Section 3.1, respectively. " + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "text", + "content": " is the number of aggregated annotations. " + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 514, + 482, + 588 + ], + "type": "text", + "content": " is the loss function that the base OD model adopts. Since the OD operates as a black-box module in the proposed framework, any object detector can be used here as long as it outputs the predicted bounding boxes and the associated class label probabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "text", + "content": "Annotations-Predictions Matcher. To update the priors of the aggregators, we first match each annotation to one of the predictions. To find the optimal prediction, " + }, + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\hat{y}_m^*" + }, + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "text", + "content": ", for each annotation, " + }, + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "inline_equation", + "content": "y_m" + }, + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "text", + "content": ", we minimise a matching cost function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 241, + 647, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 647, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 241, + 647, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {m} ^ {*} = \\underset {\\hat {y} _ {n} \\in \\hat {y}} {\\arg \\min } \\mathcal {L} _ {\\text {m a t c h}} (\\hat {y} _ {n}, y _ {m}) , \\tag {2}", + "image_path": "bf6eca1280a5a96d155dd3646214e543b876b4c8ca72187fb4a71910c774c1b2.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bayesian Detector Combination" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 162, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 162, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 162, + 126 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 172, + 133, + 481, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 133, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 172, + 133, + 481, + 148 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a t c h}} (\\hat {y} _ {n}, y _ {m}) = - \\hat {p} _ {n (c _ {m})} + \\lambda_ {1} \\mathcal {L} _ {I o U} (\\hat {b} _ {n}, b _ {m}) + \\lambda_ {2} | | \\hat {b} _ {n} - b _ {m} | | _ {1}. \\qquad (3)", + "image_path": "547ed815e2337e05244f9893870fd29bcaa19ac364b32ed10ac6e5b19fe15666.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 152, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 152, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 152, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\lambda_{1}, \\lambda_{2} \\in \\mathbb{R}" + }, + { + "bbox": [ + 131, + 152, + 481, + 213 + ], + "type": "text", + "content": " are hyperparameters and " + }, + { + "bbox": [ + 131, + 152, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IoU}" + }, + { + "bbox": [ + 131, + 152, + 481, + 213 + ], + "type": "text", + "content": " is the IoU loss [42]. In short, for each annotation, we aim to find the prediction with a higher predicted probability of the annotated class label, a higher IoU and a smaller L1 distance between the predicted and the annotated bounding box. This is inspired by the set prediction loss in DETR [9]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 213, + 482, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 286 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 286 + ], + "type": "text", + "content": "Instead of finding a one-to-one matching with a global minimum matching cost, we are interested in a one-to-many matching that one prediction can be matched to many crowdsourced annotations, with a local minimum matching cost, i.e. finding the prediction with minimum cost for each annotation individually. Moreover, a simple heuristic matching rule is sufficient instead of the Hungarian algorithm in DETR which is more computationally demanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 300, + 482, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 335 + ], + "type": "text", + "content": "Bounding Box Aggregator. In the Bounding Box Aggregator (BBA), the translation and scaling errors in the x and y-axis of each annotator are modelled with a Gaussian distribution:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 346, + 481, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 346, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 233, + 346, + 481, + 361 + ], + "type": "interline_equation", + "content": "p \\left(\\epsilon_ {m} \\mid k _ {m} = k, \\boldsymbol {\\mu}, \\boldsymbol {\\sigma}\\right) = \\mathcal {N} \\left(\\mu^ {k}, \\sigma^ {k}\\right), \\tag {4}", + "image_path": "1666ce06af731209e0f4d5c2b1d48f73bcaa518201ee991fdd10d84ee94ce07a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon_{m} \\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": " contains the translation and scaling errors in the x and y-axis. " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{\\sigma}" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": " are a collection of Gaussian parameters with " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mu^k" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\sigma^k" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": " representing the mean and covariance of " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon_{m}" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": " for the annotator " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": ", respectively. The Gaussian distributions have a Gaussian-Gamma prior with hyperparameters " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mu_0^k" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "v_0^k" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\beta_0^k" + }, + { + "bbox": [ + 131, + 363, + 482, + 437 + ], + "type": "text", + "content": " [6, 35]. To update the prior distribution, we compute the four error terms between all matched prediction and annotation boxes for each annotator with:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 161, + 441, + 481, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 441, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 161, + 441, + 481, + 461 + ], + "type": "interline_equation", + "content": "\\epsilon_ {m} = \\left[ \\hat {b} _ {m (1)} ^ {*} - b _ {m (1)}, \\hat {b} _ {m (2)} ^ {*} - b _ {m (2)}, \\hat {b} _ {m (3)} ^ {*} \\div b _ {m (3)}, \\hat {b} _ {m (4)} ^ {*} \\div b _ {m (4)} \\right], \\quad (5)", + "image_path": "b51a2c8eb996ccc0d0c56d17956bac881bb62f2f083fb6f6ce7119c99974c927.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 467, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 482, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 467, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\hat{b}_m^*" + }, + { + "bbox": [ + 131, + 467, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 467, + 482, + 514 + ], + "type": "inline_equation", + "content": "b_{m}" + }, + { + "bbox": [ + 131, + 467, + 482, + 514 + ], + "type": "text", + "content": " are the matched prediction and annotation boxes, respectively. The subtraction and division operations are designed to make the error terms translation-invariant and scale-invariant. The Gaussian-Gamma prior is updated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 178, + 514, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 514, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 178, + 514, + 481, + 539 + ], + "type": "interline_equation", + "content": "\\tilde {\\mu} ^ {k} = \\frac {\\mu_ {0} ^ {k} + M ^ {k} \\bar {\\epsilon} ^ {k}}{M ^ {k} + 1}, \\quad (6) \\quad \\tilde {v} ^ {k} = v _ {0} ^ {k} + \\frac {M ^ {k}}{2}, \\tag {7}", + "image_path": "0397261fcf06eca02b6e1b7bb70e4ed8b4b84617018e57028c6f7543a96d2067.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 548, + 481, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 548, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 548, + 481, + 579 + ], + "type": "interline_equation", + "content": "\\tilde {\\beta} ^ {k} = \\beta_ {0} ^ {k} + \\frac {M ^ {k} (\\bar {\\epsilon} ^ {k} - \\mu_ {0} ^ {k}) (\\bar {\\epsilon} ^ {k} - \\mu_ {0} ^ {k}) ^ {\\top}}{2 (M ^ {k} + 1)} + \\frac {1}{2} \\sum_ {m = 1} ^ {M} \\mathbb {1} [ k _ {m} = k ] (\\epsilon_ {m} - \\bar {\\epsilon} ^ {k}) (\\epsilon_ {m} - \\bar {\\epsilon} ^ {k}) ^ {\\top}, (8)", + "image_path": "8cddf42bb8348ff9dfedb1cc643d9e8017b4dca01f75c5cedc13df6fcd03a601.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "inline_equation", + "content": "M^k" + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "text", + "content": " is the number of annotation-prediction matches associated with the " + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "text", + "content": "-th annotator, " + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "inline_equation", + "content": "\\overline{\\epsilon}^k" + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "text", + "content": " is the mean of all " + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "inline_equation", + "content": "\\epsilon_{m}" + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "text", + "content": "-th annotator and " + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "inline_equation", + "content": "\\mathbb{1}[\\cdot]" + }, + { + "bbox": [ + 131, + 585, + 481, + 620 + ], + "type": "text", + "content": " is the indicator function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 621, + 481, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 481, + 644 + ], + "type": "text", + "content": "The obtained posterior mean (i.e. " + }, + { + "bbox": [ + 131, + 621, + 481, + 644 + ], + "type": "inline_equation", + "content": "\\mu^k = \\tilde{\\mu}^k" + }, + { + "bbox": [ + 131, + 621, + 481, + 644 + ], + "type": "text", + "content": ") is used to correct the annotated boxes for each annotator:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 198, + 648, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 648, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 198, + 648, + 481, + 669 + ], + "type": "interline_equation", + "content": "b _ {m} := \\left(b _ {m} + \\left[ \\mu_ {(1)} ^ {k}, \\mu_ {(2)} ^ {k}, 0, 0 \\right]\\right) \\odot \\left[ 1, 1, \\mu_ {(3)} ^ {k}, \\mu_ {(4)} ^ {k} \\right]. \\tag {9}", + "image_path": "37f69fb5e156b4107e3a7e5aba15f65d8282f2a409a31471a1bd79210f93295a.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z.Q. Tan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": "All annotations matched to the same prediction are aggregated via a weighted average using the posterior precision, " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "P^k = 1 \\oslash \\text{diag}(\\sigma^k)" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\oslash" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": " denotes the element-wise division and " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\sigma^k = \\tilde{\\beta}^k / \\tilde{v}^k" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": ". The posterior precision, " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "P^k" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": ", of each annotator is used as weights to obtain the final boxes, " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\{\\mathfrak{b}_n\\}_{n=1}^H" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": ", with" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 174, + 481, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 174, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 255, + 174, + 481, + 204 + ], + "type": "interline_equation", + "content": "\\mathfrak {b} _ {n} = \\frac {\\sum_ {(b , k) \\in \\kappa_ {n}} P ^ {k} \\odot b}{\\sum_ {(b , k) \\in \\kappa_ {n}} P ^ {k}}. \\tag {10}", + "image_path": "c887daddbde75ccc56b75a2322aeee369565a716ac0ece155534ebb7819b1e75.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "text", + "content": " indicates the number of predictions matched to at least one annotation which is also equal to the number of aggregated annotations defined in Section 3.1. " + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\kappa_{n}" + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "text", + "content": " is the set of tuples consisting of the annotated bounding boxes and their annotator ID that are matched to the " + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "text", + "content": "-th prediction. Inspired by EARL [27], a loss scale factor, " + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\gamma_{n}" + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "text", + "content": ", is computed for each aggregated box as the ratio of the number of matched annotations to " + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 210, + 482, + 308 + ], + "type": "text", + "content": " and is used to re-weight the loss function of the OD in Equation (1). We refer the readers to Appendix A for further details on the updating procedure of the Bounding Box Aggregator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 323, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 482, + 384 + ], + "type": "text", + "content": "Class Label Aggregator. To aggregate crowdsourced class labels, we extend the Bayesian classifier combination neural network (BCCNet) [19] to the proposed BDC framework. In the class label aggregator (CLA), the annotated class label from each annotator, " + }, + { + "bbox": [ + 131, + 323, + 482, + 384 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 131, + 323, + 482, + 384 + ], + "type": "text", + "content": ", is modelled as a multinomial distribution conditioning on the true label of the object:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 235, + 392, + 481, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 392, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 235, + 392, + 481, + 407 + ], + "type": "interline_equation", + "content": "p \\left(c _ {m} \\mid k _ {m} = k, t _ {m} = j, \\boldsymbol {\\pi}\\right) = \\pi_ {j, c _ {m}} ^ {k}, \\tag {11}", + "image_path": "4c0ab9cbb8b0dbd4c410e81d73e622cb1b3d1f682a89008033937413f2f5692c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": " is a collection of confusion matrices, " + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\pi^k\\in \\mathbb{R}^{J\\times J}" + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": " is the confusion matrix of the " + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": "-th annotator and " + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": " is the ground truth class label. Each row, " + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\pi_j^k" + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": ", provides the event probabilities of a multinomial distribution for the " + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": "-th annotator when " + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "inline_equation", + "content": "t_m = j" + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": " and is modelled with a Dirichlet prior with hyperparameters " + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\alpha_{0j}^k = \\{\\alpha_{0j,1}^k,\\dots,\\alpha_{0j,J}^k\\}" + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\pi_j^k |\\alpha_{0j}^k\\sim \\text{Dir}(\\alpha_{0j}^k)" + }, + { + "bbox": [ + 131, + 414, + 482, + 475 + ], + "type": "text", + "content": " [23]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 475, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 482, + 499 + ], + "type": "text", + "content": "The multinomial posterior distribution of the true label (i.e. the aggregated class label probability) can be obtained as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 213, + 508, + 481, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 508, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 213, + 508, + 481, + 546 + ], + "type": "interline_equation", + "content": "\\rho_ {n, j} = \\exp \\left(\\ln \\hat {p} _ {n, j} + \\sum_ {(c, k) \\in \\tilde {\\kappa} _ {n}} \\mathbb {E} _ {\\pi_ {j} ^ {k}} \\ln \\pi_ {j, c} ^ {k}\\right), \\tag {12}", + "image_path": "9725b9e53f83a35d8a6b82a01beb9794075772f2f8c812c9b2311ce2be2fbf3b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{\\kappa}_n" + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "text", + "content": " is the set of tuples consisting of the annotated class labels and their annotator ID that are matched to the " + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "text", + "content": "-th prediction. " + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\hat{p}_{n,j}" + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "text", + "content": " is the OD's class-" + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "text", + "content": " probability of the " + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 552, + 482, + 625 + ], + "type": "text", + "content": "-th prediction. Note that the CLA outputs class probabilities (soft labels) unlike other methods (e.g. MV) that output a single class (hard label). Appendix B provides the derivation for Equation 12 and a detailed discussion of the updating steps of the Class Label Aggregator." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Training. The optimal OD's parameters, " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta_{O}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": ", and aggregators' parameters, " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\beta}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\alpha}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": ", are learnt iteratively in the BDC framework. Firstly, we initialise" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "text", + "content": "Bayesian Detector Combination" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "the parameters either randomly or with prior knowledge of the domain. Next, we train the OD by optimising the loss function, " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{D}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", with the aggregated annotations obtained using the current values of the aggregators' parameters. After one epoch of training, we obtain the OD's predictions using the current value of " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\theta_{O}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " to update the prior distributions of aggregators. These two steps are repeated iteratively until the parameters converge. See Appendix C for a further discussion of the training steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 217, + 230, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 230, + 230 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 230, + 230 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "text", + "content": "Training details. To demonstrate the generalisability and robustness of BDC in incorporating different object detection algorithms, three common object detection algorithms are used in our experiments, YOLOv7 [53], FRCNN [41] with the ResNet50 backbone and EVA [14]. During training, we randomly apply data augmentation such as rotation, translation, zooming and horizontal flipping to the training images and annotations. " + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2}" + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "text", + "content": " in Equation (3) are set to 2 and 5, respectively, following Carion et al. [9]. Lastly, the Gaussian-Gamma prior of the BBA, " + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{0}" + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{0}" + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\beta_0" + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "text", + "content": ", are set to 0, 10 and 0.5, respectively, while the Dirichlet prior of the CLA, " + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 130, + 239, + 482, + 360 + ], + "type": "text", + "content": ", is initialised to 1 and each diagonal entry is set to 10. The complete training configuration can be found in Appendix D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 501 + ], + "type": "text", + "content": "Comparison against other aggregation methods. We compare the performance of BDC against multiple baselines as shown in Tables 1 to 4. 'NA' denotes the process of training the model with all the crowdsourced annotations as the ground truth annotations [56]. 'MV' stands for majority voting where we apply clustering to the crowdsourced annotations, then aggregate the class labels via MV and aggregate the bounding boxes as the biggest box where the majority had annotated pixel-wise, following the practice in [46]. We also implement Crowd R-CNN [18] and WBF-EARL [27] designed for crowdsourced object detection. For WBF-EARL [27] which requires the knowledge of annotators' proficiency levels, we followed the settings in [27] for the VinDR-CXR dataset while for the synthetic datasets, they are set to the overall accuracy of the annotators." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 516, + 307, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 307, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 307, + 528 + ], + "type": "text", + "content": "4.1 Real Crowdsourced Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "VinDr-CXR. We evaluate the annotation aggregation methods on VinDrcxR [36], a thoracic abnormalities detection dataset. VinDr-CXR consists of 14 abnormalities classes and contains 15,000 training and 3,000 testing chest radiograph images. The training images are annotated by 17 experienced radiologists where at least three annotators are randomly assigned to annotate each image. Figures 1f to 1j show examples of the annotations with high bounding box and class label disagreement by the annotators. Furthermore, the testing images are annotated with the consensus of five expert radiologists through a two-stage labelling process and the test set annotations are not publicly available. The predictions are submitted to the VinBigData Kaggle [12] to obtain mean average precision (AP) at 0.4 IoU threshold on the private test set." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z.Q. Tan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 217, + 157, + 397, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: AP-4 results on VinDr-CXR [36] dataset obtained from the private score of VinBigData Kaggle competition [12]. Crowd R-CNN can only integrate FRCNN, hence there are no results reported for the two other entries corresponding to Crowd R-CNN." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 157, + 397, + 228 + ], + "lines": [ + { + "bbox": [ + 217, + 157, + 397, + 228 + ], + "spans": [ + { + "bbox": [ + 217, + 157, + 397, + 228 + ], + "type": "table", + "html": "
MethodTest AP-4
YOLOv7FRCNNEVA
NA17.417.27.8
MV13.916.38.2
Crowd R-CNN [18]-16.7-
WBF-EARL [27]16.417.08.4
BDC (ours)19.217.98.9
", + "image_path": "8c6eea403ff85663083e5499a0d405c493c5c56501ad5c2f240e7501f8dcd259.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 239, + 201, + 331 + ], + "blocks": [ + { + "bbox": [ + 133, + 239, + 201, + 331 + ], + "lines": [ + { + "bbox": [ + 133, + 239, + 201, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 201, + 331 + ], + "type": "image", + "image_path": "27d6d5e0ce6ea3eb674e5b640ea66d0fb302820d3a10fd5aa9d1791cd686ee87.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 332, + 180, + 341 + ], + "lines": [ + { + "bbox": [ + 153, + 332, + 180, + 341 + ], + "spans": [ + { + "bbox": [ + 153, + 332, + 180, + 341 + ], + "type": "text", + "content": "(a) NA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 201, + 239, + 269, + 331 + ], + "blocks": [ + { + "bbox": [ + 201, + 239, + 269, + 331 + ], + "lines": [ + { + "bbox": [ + 201, + 239, + 269, + 331 + ], + "spans": [ + { + "bbox": [ + 201, + 239, + 269, + 331 + ], + "type": "image", + "image_path": "83cedb638eeeab9d7d64990fdfdc8764761bfecc381e1887ff342da22235edd2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 332, + 250, + 341 + ], + "lines": [ + { + "bbox": [ + 221, + 332, + 250, + 341 + ], + "spans": [ + { + "bbox": [ + 221, + 332, + 250, + 341 + ], + "type": "text", + "content": "(b) MV" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 271, + 239, + 339, + 331 + ], + "blocks": [ + { + "bbox": [ + 271, + 239, + 339, + 331 + ], + "lines": [ + { + "bbox": [ + 271, + 239, + 339, + 331 + ], + "spans": [ + { + "bbox": [ + 271, + 239, + 339, + 331 + ], + "type": "image", + "image_path": "291058eff251e95d3af1d875e90d621628439a2b50e9c3bf8ae29686673dc5d7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 332, + 334, + 341 + ], + "lines": [ + { + "bbox": [ + 276, + 332, + 334, + 341 + ], + "spans": [ + { + "bbox": [ + 276, + 332, + 334, + 341 + ], + "type": "text", + "content": "(c) WBF-EARL" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 350, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 132, + 350, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 482, + 395 + ], + "type": "text", + "content": "Fig. 3: Comparison of aggregated labels from different methods on the VinDr-CXR [36]. For WBF-EARL [27], the number beside the class label is the annotators' level of agreement while for Crowd R-CNN [18] and BDC, the number indicates the class probability. For NA, the colours represent the different annotators." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 339, + 239, + 408, + 331 + ], + "blocks": [ + { + "bbox": [ + 339, + 239, + 408, + 331 + ], + "lines": [ + { + "bbox": [ + 339, + 239, + 408, + 331 + ], + "spans": [ + { + "bbox": [ + 339, + 239, + 408, + 331 + ], + "type": "image", + "image_path": "f90afadb7091e5e1cd925559bdc367793f7da1b9ae43343c7b16bda3ddec5519.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 332, + 406, + 341 + ], + "lines": [ + { + "bbox": [ + 340, + 332, + 406, + 341 + ], + "spans": [ + { + "bbox": [ + 340, + 332, + 406, + 341 + ], + "type": "text", + "content": "(d) Crowd R-CNN" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 408, + 239, + 476, + 331 + ], + "blocks": [ + { + "bbox": [ + 408, + 239, + 476, + 331 + ], + "lines": [ + { + "bbox": [ + 408, + 239, + 476, + 331 + ], + "spans": [ + { + "bbox": [ + 408, + 239, + 476, + 331 + ], + "type": "image", + "image_path": "d03061ac0e82e76322ef74896bf685d31fe6288b1f63cb222fcefe2a401f31a3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 332, + 469, + 341 + ], + "lines": [ + { + "bbox": [ + 414, + 332, + 469, + 341 + ], + "spans": [ + { + "bbox": [ + 414, + 332, + 469, + 341 + ], + "type": "text", + "content": "(e) BDC (ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 422, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 506 + ], + "type": "text", + "content": "Disaster Response Dataset. We further use the disaster response dataset [19] where the task is to detect undamaged and damaged buildings on satellite images taken after hurricane Maria in 2017. The dataset consisting of 464 satellite images of Dominica is annotated by 13 semi-professional paid crowd members to ensure high labelling quality. Despite this effort, labels from different crowd members vary significantly (Figures 1k to 1o). Consequently, the dataset does not contain ground truth labels. Thus, its qualitative results are presented in Appendix E." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 525, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 586 + ], + "type": "text", + "content": "Quantitative Results. Table 1 presents the AP.4 metric obtained from the VinBigData Kaggle competition and shows that the proposed BDC leads to higher AP.4 than other methods across different OD models on the VinDr-CXR test set. As Crowd R-CNN can only integrate FRCNN, there are no results reported for the two other entries corresponding to Crowd R-CNN." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Qualitative Results. We visualise the results of the aggregated annotations from each method on the VinDr-CXR dataset in Figure 3. Since the true ground truth of the datasets is not provided, we use 'MV' as the best approximation of the ground truth for this qualitative assessment and observe that BDC produces more accurate aggregation compared to the other methods." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "text", + "content": "Bayesian Detector Combination" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 255, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 255, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 255, + 128 + ], + "type": "text", + "content": "4.2 Synthetic Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "text", + "content": "To alleviate the limitations in the available real-world crowdsourced datasets including the lack of ground truth labels, we synthesise several synthetic datasets with crowdsourced annotations to enable the systematic evaluation of the proposed method. This makes available the ground truth annotations to allow us to obtain detailed test evaluation metrics (e.g., AP" + }, + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "text", + "content": ", AP" + }, + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "inline_equation", + "content": "^{75}" + }, + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "text", + "content": ", AP" + }, + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "inline_equation", + "content": "^{5::95}" + }, + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "text", + "content": ") and verify the performance of the proposed method by evaluating the aggregated crowdsourced annotations with the true ground truth annotations of the training dataset (indicated as 'Train Aggregation AP' in Tables 2 to 4). Two common object detection datasets are used to construct the synthetic datasets: (1) Pascal Visual Object Classes (VOC) 2007 [13], a 20-class object detection dataset which consists of 5,011 and 4,952 images for training and testing, respectively; (2) Microsoft Common Objects in Context (COCO) 2017 [29], a large-scale 80-class object detection dataset consisting of 118,287 training and 5,000 validation images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "text", + "content": "Synthetic settings. Four synthetic settings are investigated in this work: (1) VOC-FULL and (2) COCO-FULL where 10 annotators with average skill levels annotate the whole datasets; (3) VOC-MIX where the first 5 annotators are experts while the remaining 20 annotators have an average skill level. All annotators annotate the whole dataset since Pascal VOC is a small dataset; and (4) COCO-MIX where 20 experts, 550 annotators with average skill level and 330 annotators with poor skill level work together to annotate the MS COCO dataset. Additionally, the last 100 are random annotators, i.e., the annotated class labels and bounding boxes are drawn from a uniform distribution. Since MS COCO is a large-scale dataset, we set that each annotator annotates a small subset of images, i.e., the coverage for the expert, average, poor and random annotators are " + }, + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 305, + 482, + 498 + ], + "type": "text", + "content": ", respectively. Note that all annotators in these settings are simulated using deep learning models (See Appendix F for further discussions of the synthesis method and pseudocode). Additional experiments where we directly used the object detectors' predictions as synthetic crowdsourced annotations can be found in Appendix H." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Quantitative Results. Tables 2 to 4 report the experiment results of YOLOv7, FRCNN and EVA trained with different aggregation methods on all synthetic settings, respectively. Additionally, we trained each model with ground truth annotations of each dataset (denoted by 'GT') to obtain reference optimal model performance. The aggregation APs of the train dataset for NA and WBF-EARL are low because NA directly uses all the crowdsourced annotations while WBF outputs multiple bounding boxes per class per object instance, leading to a much lower precision (see Figure 4 for examples). We do not include performance from Crowd R-CNN for the COCO-MIX setting as its training time increases linearly with " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " which means that the training time is prohibitively high for this dataset where " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "K = 1,000" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". In comparison, the extra computational cost introduced by the proposed BDC is small when compared to the much heavier training cost of the OD backbone (see Section 4.4 for an analysis of the computational cost)." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z.Q. Tan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 146, + 451, + 340 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Test and train aggregation AP results for YOLOv7 model trained with different annotation aggregation methods on synthesised and ground truth datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 146, + 451, + 340 + ], + "lines": [ + { + "bbox": [ + 164, + 146, + 451, + 340 + ], + "spans": [ + { + "bbox": [ + 164, + 146, + 451, + 340 + ], + "type": "table", + "html": "
DatasetMethodTest APTrain Aggregation AP
AP.5AP.75AP.5:95AP.5AP.75AP.5:95
VOC-FULLNA59.745.340.710.29.99.9
MV72.252.347.362.048.548.5
WBF-EARL [27]69.352.345.722.316.414.5
BDC (ours)84.963.554.895.080.666.3
VOC-MIXNA46.127.627.44.44.34.3
MV77.161.053.775.770.668.6
WBF-EARL [27]61.047.240.927.924.920.0
BDC (ours)86.269.058.897.694.977.2
VOCGT87.774.065.3100100100
COCO-FULLNA53.441.437.48.68.28.2
MV61.946.442.752.544.443.3
WBF-EARL [27]55.642.538.914.68.98.6
BDC (ours)65.049.644.583.666.156.9
COCO-MIXNA53.441.237.58.57.98.0
MV60.946.442.445.842.341.1
WBF-EARL [27]56.242.639.112.17.07.0
BDC (ours)64.849.344.480.562.954.3
COCOGT67.552.848.6100100100
", + "image_path": "9c60d4a54f0850ae74defb8c521557560e8220123b9e4167bb380c0603f6fcb7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 363, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 411 + ], + "type": "text", + "content": "Tables 2 to 4 demonstrate the superior performance of BDC compared to other methods from simpler synthetic settings to more complex synthetic crowdsourced annotations, i.e. from VOC-FULL to COCO-MIX. In addition, BDC achieves the highest train aggregation AP compared to other methods across all settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 423, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 481, + 555 + ], + "type": "text", + "content": "Qualitative Results. Figure 4 visualises the results of the aggregated annotations from each method on the VOC-MIX synthetic dataset. WBF performs box fusion on a per-class basis and results in significantly more boxes compared to other methods (excluding NA) which contributed to the low train aggregation AP and noisy labels during model training. On the other hand, although both Crowd R-CNN and BDC output class probabilities, the probability distribution of Crowd R-CNN is heavily skewed which is similar to hard labels, while BDC outputs the learnt class probabilities and can benefit from these soft labels during model training [34]. Besides, BDC can cope with failure cases of its aggregator as shown by the lower class probability of incorrect aggregation. Additional qualitative results can be found in Appendix I." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 573, + 337, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 337, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 337, + 586 + ], + "type": "text", + "content": "4.3 Scalability and Robustness of BDC" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": "We analyse the scalability and robustness of the aggregation methods using the Pascal VOC dataset. YOLOv7 is selected as the OD because it is lightweight and fast to train. Figure 5a demonstrates the effect of varying numbers of annotators, " + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": ", where all annotations are synthesised using annotators of average skill level. Note that training YOLOv7 with the NA method results in a GPU out-of-memory error when " + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "K \\geq 100" + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": ". When the number of annotators increases, the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bayesian Detector Combination" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 168, + 456, + 404 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Test and train aggregation AP results for FRCNN model trained with different annotation aggregation methods on synthesised and ground truth datasets. Crowd R-CNN has a training time that increases linearly with " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": ", hence there is no result reported for COCO-MIX dataset with " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "K = 1,000" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 168, + 456, + 404 + ], + "lines": [ + { + "bbox": [ + 157, + 168, + 456, + 404 + ], + "spans": [ + { + "bbox": [ + 157, + 168, + 456, + 404 + ], + "type": "table", + "html": "
DatasetMethodTest APTrain Aggregation AP
AP·5AP·75AP·5:95AP·5AP·75AP·5:95
VOC-FULLNA47.411.319.410.29.99.9
MV79.753.548.862.048.548.5
Crowd R-CNN [18]71.633.930.158.343.031.5
WBF-EARL [27]71.837.738.722.316.414.5
BDC (ours)83.756.351.095.080.466.3
VOC-MIXNA34.14.412.24.44.34.3
MV82.057.151.975.770.668.6
Crowd R-CNN [18]70.231.229.360.644.231.0
WBF-EARL [27]69.836.537.427.924.920.0
BDC (ours)84.759.352.997.694.676.9
VOCGT84.662.155.2100100100
COCO-FULLNA39.712.017.78.68.28.2
MV55.635.733.452.544.443.3
Crowd R-CNN [18]48.527.124.249.536.233.7
WBF-EARL [27]51.928.028.414.68.98.6
BDC (ours)56.635.333.583.366.056.7
COCO-MIXNA43.917.521.48.57.98.0
MV54.834.532.945.842.341.1
Crowd R-CNN [18]------
WBF-EARL [27]51.928.028.312.17.07.0
BDC (ours)56.134.633.080.362.754.1
COCOGT58.038.535.7100100100
", + "image_path": "b354d5653d9b37de644661a35a3f573e9bc31707fd8100abc6f2120d71b04116.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 426, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 462 + ], + "type": "text", + "content": "performance of YOLOv7 increases for both BDC and MV while the performance of both NA and WBF-EARL decreases when " + }, + { + "bbox": [ + 130, + 426, + 481, + 462 + ], + "type": "inline_equation", + "content": "K > 50" + }, + { + "bbox": [ + 130, + 426, + 481, + 462 + ], + "type": "text", + "content": ". In addition, BDC has an average AP.5:95 improvement of 4.35 over MV." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "text", + "content": "We also investigate the effect of changing the percentage of noisy annotators while setting " + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "text", + "content": " to 25. Here, noisy annotators are defined as annotators with a poor skill level while the other annotators are annotators with expert skill level. As illustrated in Figure 5b, we found that AP" + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "inline_equation", + "content": "^{5:95}" + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "text", + "content": " of all aggregation methods reduce when we increase the percentage of noisy annotators. However, MV drops at a faster rate compared to the proposed method when the noisy annotator percentage is over " + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "text", + "content": ". We observe that BDC remains robust even when " + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "text", + "content": " of the annotators have low skill levels and the AP" + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "inline_equation", + "content": "^{5:95}" + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "text", + "content": " is 18.5 higher than that obtained with MV. The consistent performance of BDC (up to " + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 606 + ], + "type": "text", + "content": " noise) can be attributed to the ability of BDC to correctly capture the ability of each annotator and assign proper weights when aggregating the annotations (See Appendix J for an analysis of the posterior distributions learnt in the aggregators)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 622, + 263, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 263, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 263, + 635 + ], + "type": "text", + "content": "4.4 Computational Cost" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Table 5 presents the wall time of each annotation aggregation method spent to pre-process the annotations and train the Faster R-CNN (FRCNN) [41] model for" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z.Q. Tan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 146, + 451, + 340 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 4: Test and train aggregation AP results for EVA model trained with different annotation aggregation methods on synthesised and ground truth datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 146, + 451, + 340 + ], + "lines": [ + { + "bbox": [ + 164, + 146, + 451, + 340 + ], + "spans": [ + { + "bbox": [ + 164, + 146, + 451, + 340 + ], + "type": "table", + "html": "
DatasetMethodTest APTrain Aggregation AP
AP.5AP.75AP.5:95AP.5AP.75AP.5:95
VOC-FULLNA90.980.969.710.29.99.9
MV91.082.271.262.048.548.5
WBF-EARL [27]93.083.571.122.316.414.5
BDC (ours)94.285.671.795.381.266.7
VOC-MIXNA86.665.757.54.44.34.3
MV91.783.072.975.770.668.6
WBF-EARL [27]91.382.268.727.924.920.0
BDC (ours)94.986.573.898.895.077.3
VOCGT94.988.178.0100100100
COCO-FULLNA71.855.750.68.68.28.2
MV74.861.054.452.544.443.3
WBF-EARL [27]74.758.451.814.68.98.6
BDC (ours)78.062.755.183.365.155.3
COCO-MIXNA73.357.251.78.57.98.0
MV75.060.654.245.842.341.1
WBF-EARL [27]75.259.352.912.17.07.0
BDC (ours)77.162.554.879.962.253.2
COCOGT78.864.958.8100100100
", + "image_path": "dc0a73f645815cbf3a0953ba5622487b28ea06889d75fdac9062e1d065727273.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": "one epoch on the VOC-MIX synthetic dataset. The wall time is computed with a desktop workstation with an i9-10900X CPU and an RTX 3090 GPU. As shown in Table 5, MV, Crowd R-CNN and WBF-EARL pre-process the annotations before model training while NA and the proposed BDC require no pre-processing step. Crowd R-CNN has the longest training time which increases linearly with the number of annotators, " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": ". On the other hand, the proposed BDC performs the annotations-predictions matching step and the updating steps of the aggregators' parameters every training epoch and therefore has a slightly longer training time compared to NA, MV and WBF-EARL. However, this increase in training time is small when compared to the much higher training cost of the OD" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 510, + 191, + 590 + ], + "blocks": [ + { + "bbox": [ + 133, + 510, + 191, + 590 + ], + "lines": [ + { + "bbox": [ + 133, + 510, + 191, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 191, + 590 + ], + "type": "image", + "image_path": "093aae14ab363fa4e3b3131dd22cce74bb1972ca87d19eb8824a7afecef186e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 591, + 175, + 601 + ], + "lines": [ + { + "bbox": [ + 148, + 591, + 175, + 601 + ], + "spans": [ + { + "bbox": [ + 148, + 591, + 175, + 601 + ], + "type": "text", + "content": "(a) NA" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 192, + 510, + 249, + 590 + ], + "blocks": [ + { + "bbox": [ + 192, + 510, + 249, + 590 + ], + "lines": [ + { + "bbox": [ + 192, + 510, + 249, + 590 + ], + "spans": [ + { + "bbox": [ + 192, + 510, + 249, + 590 + ], + "type": "image", + "image_path": "def6a5bc2d50d1d72a35dfb1849b5f556020f7051bc8e9d8a28769154e8f2b06.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 591, + 234, + 600 + ], + "lines": [ + { + "bbox": [ + 205, + 591, + 234, + 600 + ], + "spans": [ + { + "bbox": [ + 205, + 591, + 234, + 600 + ], + "type": "text", + "content": "(b) MV" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 591, + 287, + 607 + ], + "lines": [ + { + "bbox": [ + 249, + 591, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 249, + 591, + 287, + 607 + ], + "type": "text", + "content": "(c) WBF-EARL" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 250, + 510, + 306, + 590 + ], + "blocks": [ + { + "bbox": [ + 250, + 510, + 306, + 590 + ], + "lines": [ + { + "bbox": [ + 250, + 510, + 306, + 590 + ], + "spans": [ + { + "bbox": [ + 250, + 510, + 306, + 590 + ], + "type": "image", + "image_path": "e6cc64f3f1cbea29b9cfa6a1a4ae3daa9a1cd47df2c99922548f9cc98e7c4836.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "type": "text", + "content": "Fig. 4: Comparison of aggregated labels from different methods on the VOC-MIX synthetic dataset. For WBF-EARL [27], the number beside the class label is the annotators' level of agreement while for Crowd R-CNN [18] and BDC, the number indicates the class probability. For NA, the colours represent the different annotators." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 510, + 365, + 590 + ], + "blocks": [ + { + "bbox": [ + 307, + 510, + 365, + 590 + ], + "lines": [ + { + "bbox": [ + 307, + 510, + 365, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 365, + 590 + ], + "type": "image", + "image_path": "1c3e1b0d2274b0cd22b2237aff5ca34a32310e431309087b4cf609ac93475bc9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 591, + 350, + 607 + ], + "lines": [ + { + "bbox": [ + 308, + 591, + 350, + 607 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 350, + 607 + ], + "type": "text", + "content": "(d) Crowd R-CNN" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 366, + 510, + 423, + 590 + ], + "blocks": [ + { + "bbox": [ + 366, + 510, + 423, + 590 + ], + "lines": [ + { + "bbox": [ + 366, + 510, + 423, + 590 + ], + "spans": [ + { + "bbox": [ + 366, + 510, + 423, + 590 + ], + "type": "image", + "image_path": "2f71fa5cd76551269c73ecefd75a9b3fb46fef874ecfd05e22d63ed71856b78a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 591, + 421, + 607 + ], + "lines": [ + { + "bbox": [ + 366, + 591, + 421, + 607 + ], + "spans": [ + { + "bbox": [ + 366, + 591, + 421, + 607 + ], + "type": "text", + "content": "(e) BDC (ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 425, + 510, + 481, + 590 + ], + "blocks": [ + { + "bbox": [ + 425, + 510, + 481, + 590 + ], + "lines": [ + { + "bbox": [ + 425, + 510, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 425, + 510, + 481, + 590 + ], + "type": "image", + "image_path": "225f8100c52b5c363bd24eda2f5c284101bf1679f99ed6103d25bf8ecd789c77.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 591, + 465, + 607 + ], + "lines": [ + { + "bbox": [ + 424, + 591, + 465, + 607 + ], + "spans": [ + { + "bbox": [ + 424, + 591, + 465, + 607 + ], + "type": "text", + "content": "(f) Ground truth" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "text", + "content": "Bayesian Detector Combination" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 114, + 306, + 245 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 306, + 245 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 306, + 245 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 306, + 245 + ], + "type": "image", + "image_path": "0364a45c8b39f31bd95e9f88021654048f93d6fc040889b4a71f36328d5aa8ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 247, + 265, + 257 + ], + "lines": [ + { + "bbox": [ + 184, + 247, + 265, + 257 + ], + "spans": [ + { + "bbox": [ + 184, + 247, + 265, + 257 + ], + "type": "text", + "content": "(a) Scalability of BDC" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 114, + 471, + 246 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 471, + 246 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 471, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 471, + 246 + ], + "type": "image", + "image_path": "e5673ad679f6f627753d37a472c6491c794e5b0a36179bf78a139e8a5537a0f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 247, + 433, + 257 + ], + "lines": [ + { + "bbox": [ + 348, + 247, + 433, + 257 + ], + "spans": [ + { + "bbox": [ + 348, + 247, + 433, + 257 + ], + "type": "text", + "content": "(b) Robustness of BDC" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "text", + "content": "Fig. 5: (a) Effect of the number of annotators, " + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "text", + "content": ", and (b) effect of varying percentage of noisy annotators (with " + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "inline_equation", + "content": "K = 25" + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "text", + "content": ") on the test AP" + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "inline_equation", + "content": "^{5:95}" + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "text", + "content": " of YOLOv7. Results for the 'NA' method are unavailable when " + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "inline_equation", + "content": "K \\geq 100" + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "text", + "content": " due to GPU out-of-memory error." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 182, + 354, + 432, + 418 + ], + "blocks": [ + { + "bbox": [ + 130, + 311, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 346 + ], + "type": "text", + "content": "Table 5: Annotations pre-processing time and FRCNN model training time per epoch of different annotation aggregation methods on the VOC-MIX dataset. The wall time is computed on a desktop workstation with an i9-10900X CPU and an RTX 3090 GPU." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 354, + 432, + 418 + ], + "lines": [ + { + "bbox": [ + 182, + 354, + 432, + 418 + ], + "spans": [ + { + "bbox": [ + 182, + 354, + 432, + 418 + ], + "type": "table", + "html": "
MethodPre-processing time (s)Training time (s)
NA0.0537.0
MV19.4505.4
Crowd R-CNN [18]15.61175.5
WBF-EARL [27]4.9518.5
BDC (ours)0.0662.9
", + "image_path": "f2912f150e357c736233d54852b23998c137b8dd7602bb7aa94aaefeb0fb615c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 441, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 466 + ], + "type": "text", + "content": "module (618.6s OD training time out of 662.9s). Note that the model inference time is unaffected by the type of annotation aggregation method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 483, + 220, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 220, + 496 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 220, + 496 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We first presented a general annotation aggregation and object detection framework, named Bayesian Detector Combination (BDC), for learning a robust object detector from noisy crowdsourced annotations. BDC infers the consensus bounding boxes and corresponding soft labels without the need for ground truth annotations. BDC is model-agnostic and we demonstrated its superior performance by integrating it with three popular object detector backbones, namely YOLOv7, FRCNN, and EVA. Furthermore, we constructed four synthetic datasets to simulate different crowdsourcing scenarios involving annotators of various skill levels, enabling the systematic evaluation of the proposed BDC. We demonstrated the improved performance from BDC on real-world datasets and all the synthetic datasets when compared to other object detectors designed for crowdsourced annotations. We showed that BDC can scale against varying numbers of annotators and is robust against annotators with low annotation reliability." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z.Q. Tan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "content": "Acknowledgements. Z.Q. T. acknowledges the support of an Industrial Cooperative Awards in Science and Engineering (ICASE) studentship from the Engineering and Physical Sciences Research Council (EPSRC) for this work. G. C. acknowledges the support of the Engineering and Physical Sciences Research Council (EPSRC) through grant EP/Y018036/1 and the Australian Research Council (ARC) through grant FT190100525. The authors would like to thank the Satellite Applications Catapult for the provision of the disaster response dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 234, + 197, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 234, + 197, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 234, + 197, + 247 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 262, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 262, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 297 + ], + "type": "text", + "content": "1. Albarqouni, S., Baur, C., Achilles, F., Belagiannis, V., Demirci, S., Navab, N.: AggNet: Deep learning from crowds for mitosis detection in breast cancer histology images. IEEE Trans. Med. Imag. 35(5), 1313-1321 (May 2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 297, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 480, + 341 + ], + "type": "text", + "content": "2. Baumgartner, C.F., Tezcan, K.C., Chaitanya, K., Hötter, A.M., Muehlematter, U.J., Schawkat, K., Becker, A.S., Donati, O., Konukoglu, E.: PHiSeg: Capturing uncertainty in medical image segmentation. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 119-127. Shenzhen, China (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "text", + "content": "3. Bernard, J.M., Smith, A.F.M.: Bayesian Theory, chap. 3, pp. 136-138. Wiley, West Sussex, UK (1993)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 364, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 408 + ], + "type": "text", + "content": "4. Bernhardt, M., Castro, D.C., Tanno, R., Schwaighofer, A., Tezcan, K.C., Monteiro, M., Bannur, S., Lungren, M.P., Nori, A., Glocker, B., Alvarez-Valle, J., Oktay, O.: Active label cleaning for improved dataset quality under resource constraints. Nature Commun. 13(1), 1161 (Mar 2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 409, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 480, + 453 + ], + "type": "text", + "content": "5. Budd, S., Day, T., Simpson, J., Lloyd, K., Matthew, J., Skelton, E., Razavi, R., Kainz, B.: Can non-specialists provide high quality gold standard labels in challenging modalities? In: Med. Imag. Comput. Comput. Assist. Interv. pp. 251-262. Strasbourg, France (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 453, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 453, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 480, + 486 + ], + "type": "text", + "content": "6. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "text", + "content": "7. Chen, Z., Wang, H., Sun, H., Chen, P., Han, T., Liu, X., Yang, J.: Structured probabilistic end-to-end learning from crowds. In: Int. Joint Conf. Artif. Intell. pp. 1512-1518 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "text", + "content": "8. Dawid, A.P., Skene, A.M.: Maximum likelihood estimation of observer error-rates using the EM algorithm. J. Roy. Statist. Soc. Ser. C (Appl. Statist.) 28(1), 20-28 (1979)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "text", + "content": "9. DungNB, Nguyen, H.Q., Elliott, J., KeepLearning, Nhan, N.T., Culliton, P.: VinBigData chest x-ray abnormalities detection (2020), https://kaggle.com/competitions/ vinbigdata-chest-xray-abnormalities-detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 587, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 587, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 587, + 480, + 620 + ], + "type": "text", + "content": "10. Everingham, M., Luc Van Gool, C.K.I.W., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2007 (VOC2007). Int. J. Comput. Vis. 88, 303-338 (Jun 2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 480, + 665 + ], + "type": "text", + "content": "11. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA: Exploring the limits of masked visual representation learning at scale. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19358-19369. Vancouver, Canada (2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "text", + "content": "Bayesian Detector Combination" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "text", + "content": "12. Goh, H.W., Tkachenko, U., Mueller, J.: CROWDLAB: Supervised learning to infer consensus labels and quality scores for data with multiple annotators (2022), arXiv:2210.06812" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "13. Guan, M.Y., Gulshan, V., Dai, A.M., Hinton, G.E.: Who said what: Modeling individual labelers improves classification. In: AAAI Conf. Artif. Intell. pp. 3109-3118. San Francisco, CA, USA (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "text", + "content": "14. Hu, Y., Meina, S.: Crowd R-CNN: An object detection model utilizing crowdsourced labels. In: Int. Conf. Vis. Image Sig. Process. pp. 1-7. Bangkok, Thailand (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "text", + "content": "15. Isupova, O., Li, Y., Kuzin, D., Roberts, S.J., Willis, K.J., Reece, S.: BCCNet: Bayesian classifier combination neural network. In: Neural Inf. Process. Syst. Workshop Mach. Learn. Develop. Montréal, Canada (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 296 + ], + "type": "text", + "content": "16. Jensen, M.H., Jorgensen, D.R., Jalaboi, R., Hansen, M.E., Olsen, M.A.: Improving uncertainty estimation in convolutional neural networks using inter-rater agreement. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 540-548. Shenzhen, China (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "text", + "content": "17. Ji, W., Yu, S., Wu, J., Ma, K., Bian, C., Bi, Q., Li, J., Liu, H., Cheng, L., Zheng, Y.: Learning calibrated medical image segmentation via multi-rater agreement modeling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12336-12346 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 374 + ], + "type": "text", + "content": "18. Jungo, A., Meier, R., Ermis, E., Blatti-Moreno, M., Herrmann, E., Wiest, R., Reyes, M.: On the effect of inter-observer variability for a reliable estimation of uncertainty of medical image segmentation. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 682-690. Granada, Spain (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 397 + ], + "type": "text", + "content": "19. Kim, H., Ghahramani, Z.: Bayesian classifier combination. In: Int. Conf. Artif. Intell. Statist. pp. 619-627. La Palma, Canary Islands (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "text", + "content": "20. Kohl, S.A.A., Romera-Paredes, B., Meyer, C., Fauw, J.D., Ledsam, J.R., Maier-Hein, K.H., Eslami, S.M.A., Rezende, D.J., Ronneberger, O.: A probabilistic u-net for segmentation of ambiguous images. In: Adv. Neural Inform. Process. Syst. pp. 6965–6975. Montréal, Canada (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 441, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 475 + ], + "type": "text", + "content": "21. Lazarus, E., Mainiero, M.B., Schepps, B., Koelliker, S.L., Livingston, L.S.: BI-RADS lexicon for US and mammography: interobserver variability and positive predictive value. Radiology 239(2), 385-391 (Mar 2006)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "text", + "content": "22. Le, K., Tran, T., Pham, H., Nguyen Trung, H., Le, T., Nguyen, H.Q.: Learning from multiple expert annotators for enhancing anomaly detection in medical image analysis. IEEE Access 11, 14105-14114 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "text", + "content": "23. LeCun, Y., Cortes, C., Burges, C.: MNIST handwritten digit database (2010), http://yann.lecun.com/exdb/mnist" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "text", + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Zurich, Switzerland (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 620 + ], + "type": "text", + "content": "25. Maier-Hein, L., Mersmann, S., Kondermann, D., Stock, C., Kenngott, H.G., Sanchez, A., Wagner, M., Preukschas, A., Wekerle, A.L., Helfert, S., Bodenstedt, S., Speidel, S.: Crowdsourcing for reference correspondence generation in endoscopic images. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 349-356. Boston, MA, USA (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "26. Müller, R., Kornblith, S., Hinton, G.: When does label smoothing help? In: Adv. Neural Inform. Process. Syst. pp. 4694-4703. Vancouver, Canada (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 644, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 666 + ], + "type": "text", + "content": "27. Murphy, K.: Conjugate bayesian analysis of the gaussian distribution (2007), https://www.cs.ubc.ca/~murphyk/Papers/bayesGauss.pdf" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z.Q. Tan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 182 + ], + "type": "text", + "content": "28. Nguyen, H.Q., Lam, K., Le, L.T., Pham, H.H., Tran, D.Q., Nguyen, D.B., Le, D.D., Pham, C.M., Tong, H.T.T., Dinh, D.H., Do, C.D., Doan, L.T., Nguyen, C.N., Nguyen, B.T., Nguyen, Q.V., Hoang, A.D., Phan, H.N., Nguyen, A.T., Ho, P.H., Ngo, D.T., Nguyen, N.T., Nguyen, N.T., Dao, M., Vu, V.: VinDr-CXR: An open dataset of chest x-rays with radiologist's annotations. Sci. Data 9(1), 429 (Jul 2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 182, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 225 + ], + "type": "text", + "content": "29. Nowruzi, F.E., Kapoor, P., Kolhatkar, D., Hassanat, F., Laganiere, R., Rebut, J.: How much real data do we actually need: Analyzing object detection performance using synthetic and real data. In: Int. Conf. Mach. Learn. Workshop AI Auton. Driving. Long Beach, CA, USA (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 481, + 257 + ], + "type": "text", + "content": "30. Raykar, V.C., Yu, S., Zhao, L.H., Valadez, G.H., Florin, C., Bogoni, L., Moy, L., Raykar, C., Valadez, G.H.: Learning from crowds. J. Mach. Learn. Res. 11(43), 1297-1322 (Aug 2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 257, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 481, + 290 + ], + "type": "text", + "content": "31. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: Adv. Neural Inform. Process. Syst. pp. 91-99. Montréal, Canada (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 481, + 332 + ], + "type": "text", + "content": "32. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 658-666. Long Beach, CA, USA (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 333, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 481, + 354 + ], + "type": "text", + "content": "33. Rodrigues, F., Pereira, F.C.: Deep learning from crowds. In: AAAI Conf. Artif. Intell. pp. 1611-1618. New Orleans, LA, USA (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 386 + ], + "type": "text", + "content": "34. Sheng, V.S., Provost, F., Ipeirotis, P.G.: Get another label? improving data quality and data mining using multiple, noisy labelers. In: ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. pp. 614-622. Las Vegas, NV, USA (2008)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "type": "text", + "content": "35. Solovyev, R., Wang, W., Gabruseva, T.: Weighted boxes fusion: Ensembling boxes from different object detection models. Image Vis. Comput. 107, 104-117 (Mar 2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 418, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 460 + ], + "type": "text", + "content": "36. Sudre, C.H., Anson, B.G., Ingala, S., Lane, C.D., Jimenez, D., Haider, L., Varsavsky, T., Tanno, R., Smith, L., Ourselin, S., Jäger, R.H., Cardoso, M.J.: Let's agree to disagree: Learning highly debatable multirater labelling. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 665-673. Shenzhen, China (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 460, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 494 + ], + "type": "text", + "content": "37. Venanzi, M., Guiver, J., Kazai, G., Kohli, P., Shokouhi, M.: Community-based bayesian aggregation models for crowdsourcing. In: Int. Conf. World Wide Web. pp. 155-164. Seoul, Korea (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 494, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 481, + 525 + ], + "type": "text", + "content": "38. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7464-7475. Vancouver, Canada (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 481, + 568 + ], + "type": "text", + "content": "39. Watadani, T., Sakai, F., Johkoh, T., Noma, S., Akira, M., Fujimoto, K., Bankier, A.A., Lee, K.S., Müller, N.L., Song, J.W., Park, J.S., Lynch, D.A., Hansell, D.M., Remy-Jardin, M., Franquet, T., Sugiyama, Y.: Interobserver variability in the CT assessment of honeycombing in the lungs. Radiology 266(3), 936-944 (Dec 2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 568, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 481, + 600 + ], + "type": "text", + "content": "40. Wei, H., Xie, R., Feng, L., Han, B., An, B.: Deep learning from multiple noisy annotators as a union. IEEE Trans. Neural Netw. Learn. Syst. pp. 1-11 (Apr 2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 481, + 632 + ], + "type": "text", + "content": "41. Wei, J., Zhu, Z., Luo, T., Amid, E., Kumar, A., Liu, Y.: To aggregate or not? learning with separate noisy labels. In: ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. pp. 2523-2535. Long Beach, CA, USA (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 665 + ], + "type": "text", + "content": "42. Whitehill, J., Wu, T.f., Bergsma, J., Movellan, J., Ruvolo, P.: Whose vote should count more: Optimal integration of labels from labelers of unknown expertise. In: Adv. Neural Inform. Process. Syst. pp. 2035-2043. Vancouver, Canada (2009)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 101 + ], + "type": "text", + "content": "Bayesian Detector Combination" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 238 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "43. Wu, J., Fang, H., Yang, Y., Liu, Y., Gao, J., Duan, L., Yang, W., Xu, Y.: Multi-rater prism: Learning self-calibrated medical image segmentation from multiple raters (2022), arXiv:2212.00601" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "44. Yu, S., Zhou, H.Y., Ma, K., Bian, C., Chu, C., Liu, H., Zheng, Y.: Difficulty-aware glaucoma classification with multi-rater consensus modeling. In: Med. Imag. Comput. Comput. Assist. Interv. pp. 741-750 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "45. Zhang, L., Tanno, R., Xu, M.C., Jin, C., Jacob, J., Ciccarelli, O., Barkhof, F., Alexander, D.C.: Disentangling human error from the ground truth in segmentation of medical images. In: Adv. Neural Inform. Process. Syst. pp. 15750-15762 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 238 + ], + "type": "text", + "content": "46. Zhu, X., Vondrick, C., Ramanan, D., Fowlkes, C.: Do we need more training data or better models for object detection? In: Brit. Mach. Vis. Conf. Surrey, UK (2012)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z.Q. Tan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bayesian Evidential Deep Learning for Online Action Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_content_list.json b/2024/Bayesian Evidential Deep Learning for Online Action Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dbaf6cd2c89a3d170ba770c1ef62fab2912519fc --- /dev/null +++ b/2024/Bayesian Evidential Deep Learning for Online Action Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_content_list.json @@ -0,0 +1,1914 @@ +[ + { + "type": "text", + "text": "Bayesian Evidential Deep Learning for Online Action Detection", + "text_level": 1, + "bbox": [ + 240, + 140, + 763, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongji Guo, Hanjing Wang, and Qiang Ji", + "bbox": [ + 349, + 212, + 651, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rensselaer Polytechnic Institute, Troy NY 12180, USA", + "bbox": [ + 316, + 239, + 683, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guoh11, wangh36, jiq}@rpi.edu", + "bbox": [ + 385, + 255, + 617, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Online action detection aims at identifying the ongoing action in a streaming video without seeing the future. Timely and reliable response is critical for real-world applications. In this paper, we introduce Bayesian Evidential Deep Learning (BEDL), an efficient and generalizable framework for online action detection and uncertainty quantification. Specifically, we combine Bayesian neural networks and evidential deep learning by a teacher-student architecture. The teacher model is built in a Bayesian manner and transfers its mutual information and distribution to the student model through evidential deep learning. In this way, the student model can make accurate online inference while efficiently quantifying the uncertainty. Compared to existing evidential deep learning methods, BEDL estimates uncertainty more accurately by leveraging the Bayesian teacher model. In addition, we designed an attention module for active OAD, which actively selects important features based on the Bayesian mutual information instead of using all the features. We evaluated BEDL on benchmark datasets including THUMPS'14, TVSeries, and HDD. BEDL achieves competitive performance while keeping efficient inference. Extensive ablation studies demonstrate the effectiveness of each component. To verify the uncertainty quantification, we perform experiments of online anomaly detection with different types of uncertainties.", + "bbox": [ + 261, + 301, + 738, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Online action detection $\\cdot$ Evidential deep learning $\\cdot$ Bayesian neural networks $\\cdot$ Knowledge distillation $\\cdot$ Uncertainty quantification", + "bbox": [ + 261, + 604, + 740, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 657, + 374, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Online action detection (OAD) [12] aims at identifying the ongoing action in a streaming video based on the historical observations. Different from the offline setting [54, 69], it only uses the past information and makes the prediction as soon as the action takes place. It has many important applications such as autonomous driving [9], visual surveillance [55], and human-robot interaction [25]. OAD is challenging due to the incomplete observations of actions and redundant information among inputs such as background and irrelevant actions. In addition, capturing the predictive uncertainty and generalizing to unseen environments are difficult, which are required by most safety-critical applications such as autonomous driving.", + "bbox": [ + 217, + 688, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we introduce Bayesian Evidential Deep Learning (BEDL) for active online action detection and uncertainty quantification. Specifically, we combine Bayesian neural networks and evidential deep learning by a teacher-student architecture. Firstly, combining Bayesian neural networks and evidential deep learning enables accurate and efficient uncertainty quantification. Secondly, we introduce Bayesian mutual information (BMI) for OAD to actively select important features with minimal redundancy and irrelevance.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evidential deep learning (EDL) [1,52,57] is a technique developed for identifying the unknowns by quantifying uncertainty. Specifically, it uses deep neural networks to predict a Dirichlet distribution over the class probabilities, which is treated as an evidence collection process. Then the learned evidence are used to quantify the uncertainty by a single forward pass. EDL is computationally efficient and it has been widely used for open-set recognition [3,4,71] as it enables efficient uncertainty quantification without sampling. However, existing EDL methods cannot accurately model the epistemic and aleatoric uncertainty, which are critical as they provide different perspectives and information about the prediction. On the other hand, Bayesian neural networks (BNN) [20,38,47] can accurately estimate both epistemic and aleatoric uncertainty. But BNN confronts intractability of exact posterior inference and expensive sampling process, which prevents it from practical applications.", + "bbox": [ + 212, + 255, + 787, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we combine EDL and BNN in this work to take advantages of their respective strengths, i.e. to efficiently and accurately estimate both epistemic and aleatoric uncertainty by a single forward pass. Specifically, we use BNN to augment EDL for uncertainty estimation by a teacher-student model. First, BNN model (teacher model) is built in a Bayesian manner to model the posterior distribution of model parameters. Then we transfer the knowledge of the teacher model to the EDL model (student model) by knowledge distillation (KD). Different from existing KD approaches that distill prediction scores or features, our BEDL distills both the uncertainties and mutual information of teacher model to student model. In this way, the BEDL model inherits the merits of Bayesian neural networks for accurately quantifying both epistemic and aleatoric uncertainty, while preserving its computational efficiency. As a result, the BEDL model can accurately and efficiently estimate both types of uncertainties by a single forward pass. The distilled mutual information as used as for feature selection as below.", + "bbox": [ + 212, + 455, + 787, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the background and irrelevant contents issue, we also designed an attention module for the BEDL model based on the distilled Bayesian mutual information (MI). We measure the dependency between the ongoing action and historical features by MI. Then the attention module is learned to generate an attention mask that can select informative features with high MI. In this way, the prediction benefits from better features and the online action detection can be improved.", + "bbox": [ + 212, + 685, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluated BEDL's accuracy and efficiency on benchmark datasets including THUMOS'14 [31], TVSeries [12], and HDD [50]. We also demonstrated BEDL's other properties such as data-efficiency and generalization by extensive", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ablation studies. In addition, we validated the effectiveness of EDL by uncertainty quantification and online anomaly detection.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the main contributions of this paper are as follows:", + "bbox": [ + 238, + 176, + 702, + 191 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce BEDL based on combining Bayesian neural newtorks and evidential deep learning, which enables efficient and accurate online inference and uncertainty quantification.", + "- We design a mutual information based attention module for the BEDL model to select informative features that are more discriminant and reliable.", + "- Our proposed BEDL achieves competitive performance on online action detection benchmark datasets in terms of accuracy and efficiency. And we validated its uncertainty quantification capability." + ], + "bbox": [ + 225, + 199, + 782, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 340, + 387, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Uncertainty quantification. Bayesian neural networks (BNNs) have traditionally been employed for quantifying uncertainty by considering neural network parameters as probabilistic variables and analyzing their posterior distributions. Techniques for uncertainty quantification (UQ) primarily involve Markov Chain Monte Carlo (MCMC) methods [11, 63, 70] and variational inference (VI) methods [19, 42, 43]. Ensemble approaches [26, 27, 40, 58, 59, 64] also serve as effective alternatives for precise UQ. While traditional BNNs offer several benefits, their need for a wide range of parameter samples for Bayesian inference results in inefficiency in UQ. This inefficiency is due to the requirement for several forward passes through the neural network, each utilizing uniquely sampled parameters from the posterior distribution. To address this issue, evidential deep learning approaches assume the parameters of the target distribution follow a conjugate prior distribution, enabling a closed-form expression for uncertainty that facilitates fast inference. Yet, these methods need additional knowledge to accurately learn the parameters of the prior distribution, including out-of-distribution (OOD) information [44, 45], ensemble models [46], and density estimations [7]. Even with these resources, point-estimation-based evidential networks may not effectively measure epistemic uncertainty, as indicated by [56]. Our strategy, therefore, involves utilizing BNNs to guide the training of evidential networks, aiming at enhancing the accuracy of evidential deep learning models while ensuring high efficiency for inference.", + "bbox": [ + 212, + 371, + 787, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evidential deep learning (EDL). EDL [3, 4, 71] is proposed to address the challenge of making the model know the unknown when the input data is unfamiliar. In addition, it allows uncertainty estimation in a single model and forward pass [57]. Recently, EDL has been widely used for open-set recognition tasks. Bao et al. introduced DEAR [3] for open-set action recognition. By formulating action recognition in an EDL framework, the unknown actions can be recognized by computing the predictive uncertainty. Zhao et al. proposed multi-label evidential learning (MULE) [71] for open-set action recognition and novelty detection. The proposed evidence debiasing constraint enables the model to handle", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BEDL", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "general problems of single or multiple actors in the same scene, with simultaneous action(s) by any actor. Later, EDL has been applied for open-set action localization [4,10,23]. By quantifying the frame-level uncertainty, the model can distinguish unknown actions from background video framesw. Park et al. proposed MEH-HUA [48] for object detection. By aggregating the uncertainty from EDL in a bottom-up order, the context within the image can be better captured for object detection. In this paper, we introduce EDL for online action detection and anomaly detection, which has not been explored yet.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Online action detection. For the model architecture, RNN-based designs [8, 13, 15, 16, 21, 24, 28, 36, 41, 65, 68] are widely adopted because of RNN's temporal modeling capability. Typically, Xu et al. [65] proposed temporal recurrent network (TRN) that leverages both the historical information and predicted future features to detect the ongoing action. Thanks to the self-attention mechanism and the parallel computing property, Transformer-based methods [5, 29, 34, 51, 60, 62, 66, 67, 73] become the mainstream for online action detection. Wang [62] proposed OadTR that makes use of both historical information and future prediction. Xu [62] proposed long short-term Transformer (LSTR) that captures both the long-range and short-term dependencies by two memory units. To overcome the latency of feature extraction, [5] proposed E2E-LOAD for end-to-end online action detection. Besides RNN and Transformer, graph modeling is also studied for online action detection [14]. To leverage the video-level annotations instead of the dense frame-level annotations, weakly-supervised methods detection [24, 68] are also explored for OAD.", + "bbox": [ + 212, + 282, + 789, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 556, + 346, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first give an overview of the proposed Bayesian evidential deep learning (§ 3.1) and formulate the online action detection problem (§ 3.2). Then we introduce BEDL's Bayesian teacher model (§ 3.3) and evidential student model (§ 3.4). In the end, we show the training procedures (§ 3.5).", + "bbox": [ + 212, + 595, + 787, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 686, + 341, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall framework of BEDL is shown in Figure 1. BEDL is built as a teacher-student framework, which is composed of a Bayesian neural network as the teacher model and an evidential neural network as the student model. Given a streaming video as input, a pretrained backbone first extracts features. Then the Bayesian tearcher model outputs the action predictions and mutual information based on the features, which are used to train the evidential student model by knowledge distillation. During the testing, only the student model is kept to perform fast inference and efficient uncertainty quantification.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/18d1e96f08f1fde4ae2292121e3305402a2ced7754ff4f1a8c59202a4813e5ae.jpg", + "image_caption": [ + "Fig. 1: Overall framework of Bayesian evidential deep learning (BEDL). The input of model is a streaming video. A pretrained backbone is used to extract the features for each frame. During the training, the Bayesian teacher model generates the mutual information (MI) and multiple sets of predictions. The evidential student model leverages MI to train the attention module for feature selection and predictions for uncertainty quantification. During testing, the student model actively selects informative features and quantifies predictive uncertainty accurately and efficiently by a single forward pass." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 779, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Problem formulation", + "text_level": 1, + "bbox": [ + 215, + 465, + 433, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Online action detection(OAD) aims at recognizing the ongoing action in a streaming video with only the past and current observations. Denote the input video as $\\mathbf{V} = [I_1, I_2, \\dots, I_T]$ , where $T$ is the length of video and $I_t$ denotes the frame at current time $t$ . The online action detection is formulated as a classification problem: $y_t^* = \\operatorname{argmax}_c p(\\hat{y}_t = c | \\mathbf{V}_t)$ , where $\\hat{y}_t$ is the prediction, $c$ is the class label, and $\\mathbf{V}_t = \\{I_1, \\dots, I_t\\}$ is the available frame set at time $t$ . A feature extractor is used to process each frame and generate the corresponding feature vector. Denote the feature set at time $t$ as $\\mathbf{F}^t = \\{F_1^t, \\dots, F_t^t\\}$ . The feature at time $i$ is $F_i^t \\in \\mathbb{R}^J$ , where $J$ is the feature dimension of each frame.", + "bbox": [ + 212, + 491, + 787, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Bayesian teacher model", + "text_level": 1, + "bbox": [ + 215, + 648, + 455, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To take advantages of Bayesian neural networks for modeling both epistemic and aleatoric uncertainty, we build a Bayesian teacher model (BTM). BTM's objectives include: 1) generating mutual information to train the attention module for feature selection. 2) providing distribution knowledge for the student model for uncertainty quantification.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mutual information (MI) between past features and ongoing action indicates the relevance of features. Denote a past feature as $F_{ij}^{t}$ , where $i \\in \\{1, \\dots, t\\}$ is the time index, $j \\in \\{1, \\dots, J\\}$ is the feature index within each frame, and $t$ denotes the current time. We aim to obtain the mutual information between $F_{ij}^{t}$ and the ongoing action $y_{t}$ . An illustration is shown in Figure 2. To compute mutual information, we take advantage of Bayesian modeling. Different from", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BEDL", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "point estimation, Bayesian method constructs a posterior distribution of model parameters. By integrating predictions from multiple models, it is less likely to be overfitting and the predictive uncertainty can be accurately quantified. Additionally, Bayesian method is more robust when training data is insufficient. We term the mutual information computed using the Bayesian method Bayesian Mutual Information (BMI).", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/92938c5935a29a1b8dfbf2e51f3a1ddbe6a3039f4d958ae64bb1452c7e896e1d.jpg", + "image_caption": [ + "Fig. 2: Illustration of mutual information in OAD." + ], + "image_footnote": [], + "bbox": [ + 217, + 263, + 390, + 330 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0cc5d72efccb59edd2751d2a23d041443ed306c1f065d4204ac43a16478112c.jpg", + "image_caption": [ + "Fig. 3: Illustration of distribution distillation by Bayesian evidential deep learning." + ], + "image_footnote": [], + "bbox": [ + 426, + 263, + 771, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Denote the model parameters of the teacher model as $\\theta$ and we treat them as probability distributions. Then the BMI between a past feature $F_{ij}^{t}$ and the ongoing action $y_{t}$ can be written as:", + "bbox": [ + 212, + 404, + 787, + 450 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} \\left[ y _ {t}; F _ {i j} ^ {t} \\mid \\mathcal {D} \\right] = \\mathcal {H} \\left[ y _ {t} \\mid \\boldsymbol {F} _ {- i j} ^ {t}, \\mathcal {D} \\right] - \\mathcal {H} \\left[ y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 460, + 785, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ denotes the training data, $\\mathcal{H}$ denotes the entropy, and $F_{-ij}^{t}$ is the feature set at time $t$ excluding $F_{ij}^{t}$ , i.e. $F_{-ij}^{t} = F^{t} / F_{ij}^{t}$ . By definition, the entropy term is written as:", + "bbox": [ + 212, + 491, + 787, + 539 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} \\left[ y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D} \\right] = - \\sum_ {y _ {t} \\in \\mathcal {Y}} p \\left(y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D}\\right) \\log p \\left(y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 547, + 784, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{Y} = \\{0,1,\\dots,C\\}$ is the action class set. 0 represents background class and $C$ is the number of action classes. The prediction distribution in Eq. (2) can be computed as:", + "bbox": [ + 212, + 592, + 787, + 638 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np (y _ {t} | \\boldsymbol {F} ^ {t}, \\mathcal {D}) = \\int p (y _ {t} | \\boldsymbol {F} ^ {t}, \\theta) p (\\theta | \\mathcal {D}) d \\theta \\approx \\frac {1}{K} p (y _ {t} | \\boldsymbol {F} ^ {t}, \\theta_ {k}), w h e r e \\theta_ {k} \\sim p (\\theta | \\mathcal {D}) (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 661, + 785, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Eq. (3), we use the sample average to approximate it since it is impractical to integrate over all the possible parameters. Similarly, we can compute $\\mathcal{H}[y_t|F_{-ij}^t,\\mathcal{D}]$ .", + "bbox": [ + 212, + 704, + 785, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain the posterior distribution $p(\\theta | \\mathcal{D})$ , we perform the Laplace approximation (LA). To reduce the computation cost, we adopt a last-layer Bayesian setting [39]. Only the posterior distribution of last-layer parameters is modeled, while keeping the remaining parameters deterministic. Later, we will show this is efficient and effective. Specifically, we assume the last-layer parameters follow a Gaussian distribution.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Firstly, we train the teacher model under the deterministic setting by a cross entropy loss:", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta^ {*} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\mathcal {L} (\\mathcal {D}; \\theta) = \\underset {\\theta} {\\operatorname {a r g m i n}} \\left(\\sum_ {n = 1} ^ {N} l \\left(x _ {n}, y _ {n}; \\theta\\right) + r (\\theta)\\right) \\tag {4} \\\\ = \\operatorname * {a r g m i n} _ {\\theta} \\left(\\sum_ {n = 1} ^ {N} - \\log p \\left(y _ {n} \\mid f _ {\\theta} \\left(x _ {n}\\right)\\right) - \\log p (\\theta)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 189, + 784, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{D} = \\{x_n, y_n\\}_{n=1}^N$ is the training set and $r(\\theta)$ is a regularizer such as weigh regularizer (a.k.a. weight decay). So $\\theta^*$ is indeed a maximum a posteriori (MAP) estimate.", + "bbox": [ + 212, + 282, + 782, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Laplace approximation uses a second-order Taylor expansion of $\\mathcal{L}(\\mathcal{D};\\theta)$ around $\\theta^{*}$ to construct a Gaussian approximation to $p(\\theta |\\mathcal{D})$ :", + "bbox": [ + 214, + 329, + 782, + 359 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathcal {D}; \\theta) \\approx \\mathcal {L} (\\mathcal {D}; \\theta^ {*}) + \\frac {1}{2} \\left(\\theta - \\theta^ {*}\\right) ^ {T} \\left(\\nabla_ {\\theta} ^ {2} \\mathcal {L} (\\mathcal {D}; \\theta) | _ {\\theta^ {*}}\\right) \\left(\\theta - \\theta^ {*}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 371, + 784, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the first-order term vanishes at $\\theta^{*}$ . So the Laplace posterior approximation can be obtained as:", + "bbox": [ + 214, + 407, + 782, + 436 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np (\\theta | \\mathcal {D}) \\approx \\mathcal {N} \\left(\\theta^ {*}, \\Sigma\\right), \\text {w i t h} \\Sigma := \\left(\\nabla_ {\\theta} ^ {2} \\mathcal {L} (\\mathcal {D}; \\theta) | _ {\\theta^ {*}}\\right) ^ {- 1} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 448, + 784, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After the LA, we sample $K$ times from $p(\\theta|\\mathcal{D})$ to obtain $K$ sets of parameters $\\{\\theta_1,\\dots,\\theta_K\\}$ . Then we approximate the prediction distribution in Eq. (3) and further compute the mutual information of each feature. The BMI we computed are based on multiple models from the teacher model. We use BMI to supervise the attention module of student model as the importance of past feature is measured by MI. Different from the traditional MI estimated from probabilistic models such as Uncertainty-OAD [27], BEDL uses BMI and hence is more robust and accurate.", + "bbox": [ + 214, + 476, + 784, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Evidential student model", + "text_level": 1, + "bbox": [ + 215, + 619, + 470, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evidential deep learning (EDL). Softmax-based deep neural networks are unable to accurately estimate the predictive uncertainty since the softmax probability is a point estimation of the predictive distribution and it tends to be over-confident in wrong predictions. To address such issue, EDL is proposed by combing Dempster-Shafer Theory [53] and subjective logic [33]. Specifically, it assumes the classification probability follows a prior Dirichlet distribution. The training is considered as a process of collecting evidence. The learned parameters of Dirichlet distribution are used as evidence to compute the predictive uncertainty. In this way, the uncertainty can be estimated in a single forward pass. In this work, we build the evidential student model under the EDL framework so that it can output both epistemic and aleatoric uncertainty in a forward pass.", + "bbox": [ + 212, + 643, + 782, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distribution distillation. Common knowledge distillation (KD) methods aim at transferring prediction scores, features, or evidence [22], so the student model", + "bbox": [ + 214, + 809, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BEDL", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cannot estimate the predictive uncertainty. Differently, we transfer the uncertainty of the Bayesian teacher model to the evidential student model, i.e. distilling the distribution. In this way, the student model can learn from predictions of the teacher model as well as the capability of capturing the uncertainty.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Under the classification setting, the target output of student model $y$ follows a categorical distribution with parameter $\\lambda$ , $y \\sim p(y|\\lambda) = Cat(\\lambda)$ . For example, $\\lambda$ represents the probability after the final softmax layer. We treat $\\lambda$ as a random variable and assume it follows Dirichlet distribution, i.e. $\\lambda \\sim p(\\lambda|\\alpha(x,\\psi)) = Dir(\\alpha(x,\\theta))$ , where $\\alpha$ denotes the parameters of the Dirichlet distribution and $\\psi$ denotes student model parameters. Similarly, the teacher model has posterior distribution $p(\\lambda|x,\\mathcal{D}) = \\int p(\\lambda|x,\\theta)p(\\theta|\\mathcal{D})d\\theta$ . During the distillation, we transfer teacher posterior distribution $p(\\lambda|x,\\mathcal{D})$ to the student posterior distribution $p(\\lambda|\\alpha(x,\\psi))$ . An illustration is shown in Figure 3. Specifically, we minimize the KL-divergence between these two distributions as below:", + "bbox": [ + 212, + 210, + 789, + 363 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {d i s} = K L (p (\\lambda | x, \\mathcal {D}) | | p (\\lambda | \\alpha (x, \\psi))) \\\\ \\propto - \\sum_ {c = 1} ^ {C} \\log (\\Gamma (\\alpha_ {c})) + \\log \\Gamma (\\sum_ {c = 1} ^ {C} \\alpha_ {c}) - \\mathbb {E} _ {p (\\theta | \\mathcal {D})} [ \\sum_ {c = 1} ^ {C} (\\alpha_ {c} - 1) \\log \\lambda_ {c} (x, \\theta) ] \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 383, + 785, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $C$ is the number of action classes. Detailed derivation of $\\mathcal{L}_{dis}$ can be found in Appendix. Since the different predictions of the teacher model are combined, the model is more robust. The complete distribution distillation algorithm is available in the supplementary.", + "bbox": [ + 212, + 464, + 787, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mutual information based attention module. The attention module actively selects informative features from the inputs by generating a spatial-temporal attention mask $A_{t}$ using a fully-connected network and applying to the original features by element-wise product. In this way, irrelevant features are masked out since they have low mutual information to the ongoing action. The training of the attention module is supervised by the BMI $\\mathcal{I}_t$ from the teacher model. In this way, the BMI from the Bayesian teacher model is distilled to the student model. By minimizing the mean squared error (MSE) loss $\\mathcal{L}_{att}$ between $A_{t}$ and $\\mathcal{I}_t$ , the attention module can directly generate the BMI-aware attention mask without computing the BMI. $\\mathcal{L}_{att}$ can be written as below:", + "bbox": [ + 212, + 529, + 792, + 681 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a t t} = M S E \\left(\\boldsymbol {A} _ {t}, \\sigma \\left(\\mathcal {I} _ {t}\\right)\\right) = \\frac {1}{t J} \\sum_ {i, j} \\left| \\left| A _ {i j} ^ {t} - \\sigma \\left(\\mathcal {I} \\left[ y _ {t}; F _ {i j} ^ {t} \\mid \\mathcal {D} \\right]\\right) \\right| \\right| ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 702, + 785, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is the sigmoid function.", + "bbox": [ + 214, + 758, + 467, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Uncertainty quantification. After the distillation through evidential deep learning, the Dirichlet distribution of the student model gains the knowledge of the Bayesian teacher model. The uncertainty can be computed efficiently in a closed-form solution by a single forward pass. Specifically, the total uncertainty", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and epistemic uncertainty can be quantified as:", + "bbox": [ + 215, + 146, + 555, + 161 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} [ p (y | x, \\theta) ] = \\sum_ {c = 1} ^ {C} \\frac {\\alpha_ {c}}{\\alpha_ {0}} \\log \\frac {\\alpha_ {c}}{\\alpha_ {0}}, \\text {w h e r e} \\alpha_ {0} = \\sum_ {c = 1} ^ {C} \\alpha_ {c} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 172, + 785, + 220 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} [ y; \\lambda | \\alpha ] = - \\sum_ {c = 1} ^ {C} \\frac {\\alpha_ {c}}{\\alpha_ {0}} \\big (\\ln \\frac {\\alpha_ {c}}{\\alpha_ {0}} \\varPsi (\\alpha_ {c} + 1) + \\varPsi (\\alpha_ {0} + 1) \\big)\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 218, + 681, + 256 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\varPsi(\\cdot)$ is the dgamma function. To be simple, the aleatoric uncertainty can be computed by taking the difference between total uncertainty and epistemic uncertainty. The uncertainty can be also used to detect the anomaly. Details derivations are available in the supplementary.", + "bbox": [ + 214, + 267, + 787, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Training", + "text_level": 1, + "bbox": [ + 215, + 349, + 333, + 366 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The evidential student model is jointly trained for online action detection, attention distillation, and distribution distillation. Specifically, we adopt a two-strategy. We first train the attention module using the attention loss $\\mathcal{L}_{att}$ in Eq. 8. Then we jointly train the evidential student model by total loss $\\mathcal{L}$ below:", + "bbox": [ + 214, + 375, + 782, + 436 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\lambda_ {1} \\mathcal {L} _ {a t t} + \\lambda_ {2} \\mathcal {L} _ {d i s} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 448, + 784, + 463 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{ce}$ is the cross-entropy loss for online action detection. $\\lambda_{1}$ and $\\lambda_{2}$ are hyper-parameters that emphasize the attention module and distribution distillation. Although the distribution distillation can make the student model perform online action detection, we still perform joint training because the Bayesian teacher model is not perfect and its posterior distribution is approximated, which may introduce error to the student model. The joint training brings two benefits: 1) the online action detection can be improved with the supervision of ground-truth label; 2) the negative log-likelihood $\\mathcal{L}_{ce}$ makes the training faster and more stable as indicated in [18]. We demonstrate the two-stage training works better in Sec. 4.6.", + "bbox": [ + 214, + 473, + 787, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 648, + 375, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and evaluation metrics", + "text_level": 1, + "bbox": [ + 215, + 679, + 526, + 694 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "THUMOS'14 [31] is a dataset for video-based temporal action localization. We use the validation set with 200 videos for training and test set with 213 videos for evaluation. There are 20 action classes and a background class. We ignore the frames with ambiguous labels. Following the settings in [21, 65], we adopt the mean Average Precision (mAP) as the evaluation metric.", + "bbox": [ + 212, + 703, + 782, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "TVSeries [12] is a dataset collected from real TV series. It contains 27 episodes with 30 daily actions. It is a challenging dataset due to the viewpoints changing and occlusions in the videos. To counter the imbalanced data distribution, we adopt the mean calibrated Average Precision (mcAP) [12] as the", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BEDL", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "evaluation metric. It is computed as $cAP = \\sum_{k}cPrec(k)\\times \\mathbb{1}(k) / P$ , where $cPrec = TP / (TP + FP / \\omega)$ , $P$ is the total number of positive frames and $\\mathbb{1}(k)$ is an indicator function that is equal to 1 if frame $k$ is a true positive. The mcAP is the mean of calibrated average precision of all action classes.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "HRI Driving Dataset (HDD) [50] is a dataset for learning driver behavior in real-life environments. It contains 104 hours of real human driving in the San Francisco Bay Area collected by an instrumented vehicle with different sensors. There are 11 goal-oriented driving actions. Following the settings in [50], we use 100 sessions for training and 37 sessions for testing. And only the sensor data is used as the input. mAP is used as the evaluation metric for this dataset.", + "bbox": [ + 212, + 207, + 787, + 297 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/0f6d65c26725b8507a24106b2f4a5e923868eff77f8e8f66f61f215b87a7ece6.jpg", + "table_caption": [ + "Table 1: Online anomaly detection results on THUMOS'14 and TVSeries." + ], + "table_footnote": [], + "table_body": "
UncertaintyTHUMOS'14 (%)TVSeries (%)
EDL [3]64.7242.39
Total81.0465.16
Aleatoric62.5139.83
Epistemic86.4869.44
", + "bbox": [ + 241, + 351, + 468, + 407 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d22c62f3510fa1f0ab7cdc8e34823a591c97e539a89f733259d974415cb4b09c.jpg", + "image_caption": [ + "Fig.4: ROC curves of online anomaly detection using epistemic uncertainty on THUMOS'14 (left) and TVSeries (right)." + ], + "image_footnote": [], + "bbox": [ + 490, + 309, + 625, + 381 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8db17517566f77e7b997ae52726f6bcff7d6804d82eb914ebacf74a50822add9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 309, + 771, + 381 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 454, + 455, + 469 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Feature extraction. Following the settings in [15, 65, 66], we use TSN [61] to extract the features for THUMOS'14 and TVSeries. Video frames are extracted at 24 fps and the chunk size is set to 6. To better capture the spatial-temporal dependencies, we adopt the multi-scale vision Transformer [17] to extract RGB features. The optical flow features are extracted with BN-Inception [32]. The backbone is pretrained on ActivityNet [30] and Kinetics-400 [6] separately for evaluation. For HDD dataset, the sensor data is used as the input. More details can be found in the supplementary.", + "bbox": [ + 212, + 478, + 787, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Settings. The BEDL is implemented in PyTorch [49]. The model is trained by the Adam optimizer [37] with a learning rate of $10^{-4}$ and a weight decay of $5 \\times 10^{-5}$ . The batch size is set to 32. The experiments were conducted on two Nvidia RTX 3090 Ti GPUs. The number of epochs is set to 20. Ablation studies on existing methods are based on their officially released codes. Detailed model architectures are available in the supplementary. And ccode will be made publicly available.", + "bbox": [ + 212, + 599, + 787, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Online action detection results", + "text_level": 1, + "bbox": [ + 215, + 726, + 514, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate BEDL on three benchmark datasets and make a comparison with other methods in Table 2. From the results, our proposed BEDL achieves state-of-the-art performance on most benchmarks. On THUMOS'14, BEDL achieves $70.6\\%$ and $72.7\\%$ mAP using ActivityNet and Kinetics pretrained features. On TVSeries, BEDL achieves $88.9\\%$ and $90.1\\%$ mAP using ActivityNet and Kinetics features respectively. On HDD, we only use the sensor data as the input and", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "achieve $33.0\\%$ mAP, which is higher than the SOTA MAT [60]. The OAD performance demonstrate the effectiveness of BEDL for accurate online inference. In the following parts, we show other superior properties of BEDL while keeping accurate inference.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2387b23fc4233f5a7e95be8ff411e042856c7e956fded5f1920b2b17049a781e.jpg", + "table_caption": [ + "Table 2: Experiment results on THUMOS'14, TVSeries and HDD. The results on THUMOS'14 and HDD are reported as mAP (\\%). The results on TVSeries are reported as mCAP (\\%). For HDD, $\\star$ indicates RGB data is used as the input." + ], + "table_footnote": [], + "table_body": "
MethodArchitectureTHUMOS'14TVSeriesHDD
ANetKineticsANetKineticsSensor
RED [21]RNN45.3-79.2-27.4
FATS [35]-59.081.784.6-
TRN [65]47.262.183.786.229.2
IDN [15]50.060.384.786.1-
PKD [72]-64.5-86.4-
WOAD [24]-67.1---
OadTR [62]Transformer58.365.285.487.229.8
CoOadTR [29]56.164.287.687.730.6
Colar [67]59.466.986.088.130.6
LSTR [66]65.369.588.189.1-
Uncertainty-OAD [27]66.069.988.389.330.1
TeSTra [73]68.271.2---
GateHUB [8]69.170.788.489.632.1
MiniROAD [2]69.371.888.589.6-
MAT [60]70.471.688.689.732.7
E2E-LOAD [5]-72.4-90.348.1*
BEDL (ours)70.672.788.990.133.0
", + "bbox": [ + 254, + 287, + 743, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Uncertainty quantification and online anomaly detection", + "text_level": 1, + "bbox": [ + 212, + 587, + 725, + 603 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To verify the uncertainty quantification of BEDL, we quantified different types of uncertainties and perform online anomaly detection on THUMOS'14 and TVSeries. Specially, we divide the data into known classes and unknown classes (anomaly). Unknown classes do not appear in the training set. During testing, both known and unknown data will appear and the model is required to identify if the input belong to known or unknown classes. For THUMOS'14, we treat the \"ambiguous\" class that are difficult to identify during labeling process as anomaly. For TVSeries, we randomly select 5 out of 30 classes as the anomaly and repeat the process for ten times. Inputs that lead to high predictive uncertainty above the pre-defined threshold are declared as anomalies.", + "bbox": [ + 212, + 612, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The experiment results are shown in Table 1 and ROC curves are plotted in Figure 4. From the results, BEDL can accurately identify the unknown classes during the testing, which validates the effectiveness of its uncertainty quantification. Since epistemic uncertainty represents the lack of knowledge of the model and it is inversely proportional to the data density, it has better anomaly detec", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BEDL", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "tion accuracy compared to aleatoric uncertainty, which represents the noise level of the data. Detailed experiment procedures can be found in the supplementary.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Computational efficiency and model complexity", + "text_level": 1, + "bbox": [ + 214, + 202, + 653, + 218 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We leverage evidential deep learning for BEDL in order to make it more practical for real-time applications. A comparison of computation efficiency and model complexity is shown in Table 3. While keeping a high detection accuracy, our proposed BEDL has much less model complexity and computational cost. And the inference speed of the model is much higher compared to other methods. Recently, [5] proposed an end-to-end OAD framework to avoid the latency from the frame and feature extraction parts, which leads to higher overall inference speed. To make fair comparisons, we used the same feature extraction procedures as prior works and we mainly compare the model inference speed since it only depends on the model architecture and pipeline.", + "bbox": [ + 212, + 229, + 787, + 382 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e896122a41c2243c4f936a0bec1676e9959e636ce9526aaf707f2aef0f77df77.jpg", + "table_caption": [ + "Table 3: Comparison of computation efficiency and model complexity. Our proposed BEDL has less model complexity and computational cost. And the inference speed is much faster than other methods." + ], + "table_footnote": [], + "table_body": "
MethodModalityModelInference Speed (FPS)
Parameter CountGFLOPsOptical Flow ComputationRGB Feature ExtractionFlow Feature ExtractionModel
TRN402.9M1.468.170.514.6123.3
OadTR75.8M2.548.170.514.6110.0
LSTRRGB + Flow58.0M7.538.170.514.691.6
GateHUB45.2M6.988.170.514.671.2
MAT94.6M-8.170.514.672.6
BEDL (ours)19.4M0.488.170.514.6163.2
", + "bbox": [ + 218, + 463, + 782, + 560 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Ablation studies", + "text_level": 1, + "bbox": [ + 214, + 614, + 397, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Training strategies. To demonstrate the effectiveness of two-stage training strategy, we compared it with two other training strategies. One is jointly training the attention module and EDL model by the total loss in Eq. 10. The other is firstly training the attention module by the attention loss and then we freezing the weights of attention module to train the EDL model by the $\\mathcal{L}_{att}$ and $\\mathcal{L}_{ce}$ . We refer the first one as joint training and the second one as separate training. The performance comparison is shown in Table 4. The results show that the two-stage training outperforms the other two strategies on both THUMOS'14 and TVSeries with different features, which demonstrates its effectiveness.", + "bbox": [ + 212, + 643, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Number of historical frames. At time $t$ , BEDL takes a certain number of past frames to predict the ongoing action. To study the long-range and short-term modeling capability of BEDL, we vary the number of past frames as the input. The experiment results are shown in Table 5. THUMOS'14 and TVSeries", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a432d8ff557991301f8124ed3216beb5e6320ca5b97bb5baf08752f72500d148.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainingTHUMOS'14TVSeries
ActivityNetKineticsActivityNetKinetics
Joint69.570.785.488.2
Separate64.967.881.684.0
Two-stage70.672.788.990.1
", + "bbox": [ + 215, + 198, + 478, + 255 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/908f53cb7441ae3cc3963ffaf44d880cacc438db8817be61f149b59651d697d4.jpg", + "table_caption": [ + "Table 4: Ablation study of train- Table 5: Ablation study of different number ing strategies. The two-stage training of historical frames. gives the best performance." + ], + "table_footnote": [], + "table_body": "
DatasetFeatureNumber of past frames
8163264128
THUMOS'14ANet41.659.268.770.666.5
Kinetics49.361.671.472.770.0
TVSeriesANet63.576.787.688.984.9
Kinetics68.279.388.590.188.2
HDDSensor30.431.833.032.130.2
", + "bbox": [ + 486, + 185, + 782, + 265 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a1e0d246b7a52ae2864a2570bd2140b1dc14dd20888656d300313a8da421d687.jpg", + "image_caption": [ + "(a) THUMOS'14-ANet" + ], + "image_footnote": [], + "bbox": [ + 218, + 290, + 357, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f04f3bb779ca52acda0d62d70a42a05395dc275c9f8011716f60d4b53f738cee.jpg", + "image_caption": [ + "Fig. 5: Experiment results of training with small-scale data. We reduce the training data from $100\\%$ to $10\\%$ and compared with LSTR [66], MAT [60], and E2E-LOAD [5]. The results are plotted for THUMOS'14 and TVSeries with both ActivityNet and Kinetics pretrained features. Our proposed BEDL outperforms all others when training data is limited." + ], + "image_footnote": [], + "bbox": [ + 362, + 291, + 500, + 366 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7c87091e9d976c2932f37508ac7104febef7bdcd6abf9e75fb0cd528d16236eb.jpg", + "image_caption": [ + "(b) THUMOS'14-K400", + "(c) TVSeries-ANet" + ], + "image_footnote": [], + "bbox": [ + 504, + 291, + 638, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8f15ea1d3a424f9ea7cfd709fd551d3c3a421fa027fde272f47ac97b1c132998.jpg", + "image_caption": [ + "(d) TVSeries-K400" + ], + "image_footnote": [], + "bbox": [ + 643, + 291, + 779, + 364 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "are extracted at 6 fps. HDD is extracted at 3 fps. So the optimal number of frames on HDD is 32.", + "bbox": [ + 212, + 503, + 787, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Data-efficiency. By applying the mutual information based attention mechanism, we expect BEDL to be more data-efficient when the amount of training data is limited. We reduce the training data from $100\\%$ to $10\\%$ and compare with other methods. The results on THUMOS'14 and TVSeries are plotted in Figure 5. When training data is reduced, BEDL has less performance decay and outperforms other methods, which demonstrates that BEDL is more data-efficient.", + "bbox": [ + 212, + 540, + 787, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Loss function. In the second training stage of BEDL, the total loss in Eq. 10 is composed of three terms. $\\lambda_{1}$ and $\\lambda_{2}$ are the weights for attention loss and distribution distillation loss. To study the balance of all three terms, we train the model with different $\\lambda$ s on THUMOS'14 and TVSeries. The comparisons are plotted in Figure 6. We empirically choose $\\lambda_{1} = 0.4$ and $\\lambda_{2} = 6$ since they give the best performance.", + "bbox": [ + 212, + 652, + 787, + 743 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generalization. To test the generalization capability of the model, we perform the Cross-View and under occlusion experiments on TVSeries dataset. Following the settings in [27], the training set and test set are from different view angles. For the occlusion, the training data does not contain occlusion and the testing data is occluded. The experiment results and comparison are shown in Table 6. The results show that BEDL generalizes better under different conditions.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BEDL", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/96b91907d4a432f34c898ed4b6e463277eaf0d0446288aa404cdae8c4f228ac4.jpg", + "image_caption": [ + "Fig.6: Ablation study of loss function for $\\lambda_{1}$ and $\\lambda_{2}$ on THUMOS'14 with ActivityNet and Kinetics features." + ], + "image_footnote": [], + "bbox": [ + 241, + 156, + 393, + 238 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d45048dd6147d77466578e377d6eed62b429cbd80a0e6f2f7759f290e41729fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 156, + 542, + 238 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/834909f7b94606d713c5fae445cd7a4027f69aef9621b6e48b31c5e4cd1e83c0.jpg", + "table_caption": [ + "Table 6: Cross-view and occlusion experiments results on TVSeries." + ], + "table_footnote": [], + "table_body": "
MethodCV (%)Occ. (%)
TRN [65]65.885.2
OadTR [62]66.287.7
U-OAD [27]67.389.5
Colar [67]66.788.3
LSTR [66]69.589.4
TeSTra [73]70.289.9
BEDL (ours)74.391.8
", + "bbox": [ + 555, + 185, + 781, + 287 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Attention module. To verify the effectiveness Bayesian mutual information, we first compare with the model without BMI. Specifically, we trained the student model with the same evidential neural network without the attention module. The total loss function is $\\mathcal{L} = \\mathcal{L}_{ce} + \\lambda_2\\mathcal{L}_{dis}$ and we grid-searched $\\lambda_{2}$ to obtain the best accuracy. In addition, we compute the mutual information based on a single prediction instead of BMI and trained the model by the same loss function in Eq. 10. The results are shown in Table 7. From the results, the BEDL with BMI outperforms other two methods, which demonstrates the effectiveness of attention mechanism using BMI.", + "bbox": [ + 212, + 292, + 787, + 428 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Last-layer Bayesian. For the teacher model, we adopt the last-layer Bayesian [39] method to reduce the difficulty of training. We also performed the Laplace approximation over all the model parameters. The comparison is shown in Table 8. The full-Bayesian method has consistent improvement on THUMOS-14 and TVSeries. But the LA in training takes much longer time than the last-layer Bayesian and the model needs careful tuning, so we adopt the last-layer Bayesian in the LA process.", + "bbox": [ + 212, + 429, + 795, + 536 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/6ed9851b3660edb4708cf41dd5c5d6aa7d094b9d6d94b12bfc204f0121d5c3e6.jpg", + "table_caption": [ + "Table 7: Effectiveness of BMI. BEDL-No-BMI denotes the model without attention module and BEDL-MI denotes the model with non-Bayesian MI." + ], + "table_footnote": [], + "table_body": "
ModelTHUMOS'14TVSeries
ActivityNetKineticsActivityNetKinetics
BEDL-No-BMI63.164.886.587.4
BEDL-MI67.668.488.088.5
BEDL70.672.788.990.1
", + "bbox": [ + 215, + 631, + 500, + 686 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/f1545d51f234161e481fc143ac7cc25a02245264590e871caa52464b6afe5e79.jpg", + "table_caption": [ + "Table 8: Comparison of full-Bayesian and last-layer Bayesian. Full-Bayesian improves the performance since it models the distribution of all model parameters." + ], + "table_footnote": [], + "table_body": "
MethodTHUMOS-14 (mAP %)TVSeries (mcAP %)
ActivityNetKineticsActivityNetKinetics
Last-layer70.672.788.990.1
Full-Bayesian71.173.090.290.3
", + "bbox": [ + 509, + 645, + 784, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 743, + 359, + 758 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusions. In this paper, we introduce Bayesian evidential deep learning for active online action detection and uncertainty quantification. By combining Bayesian neural networks and evidential deep learning, BEDL can accurately and efficiently quantify predictive uncertainty and make online inference.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Amini, A., Schwarting, W., Soleimany, A., Rus, D.: Deep evidential regression. Advances in Neural Information Processing Systems 33, 14927-14937 (2020) 2", + "2. An, J., Kang, H., Han, S.H., Yang, M.H., Kim, S.J.: Miniroad: Minimal rnn framework for online action detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10341-10350 (2023) 11", + "3. Bao, W., Yu, Q., Kong, Y.: Evidential deep learning for open set action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13349-13358 (2021) 2, 3, 10", + "4. Bao, W., Yu, Q., Kong, Y.: Opental: Towards open set temporal action localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2979-2989 (2022) 2, 3, 4", + "5. Cao, S., Luo, W., Wang, B., Zhang, W., Ma, L.: E2e-load: End-to-end long-form online action detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10422-10432 (October 2023) 4, 11, 12, 13", + "6. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) 10", + "7. Charpentier, B., Zügner, D., Gunnemann, S.: Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in Neural Information Processing Systems 33, 1356-1367 (2020) 3", + "8. Chen, J., Mittal, G., Yu, Y., Kong, Y., Chen, M.: Gatehub: Gated history unit with background suppression for online action detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19925-19934 (2022) 4, 11", + "9. Chen, L., Ma, N., Wang, P., Li, J., Wang, P., Pang, G., Shi, X.: Survey of pedestrian action recognition techniques for autonomous driving. Tsinghua Science and Technology 25(4), 458-470 (2020) 1", + "0. Chen, M., Gao, J., Yang, S., Xu, C.: Dual-evidential learning for weakly-supervised temporal action localization. In: European Conference on Computer Vision. pp. 192-208. Springer (2022) 4", + "1. Chen, T., Fox, E., Guestrin, C.: Stochastic gradient hamiltonian monte carlo. In: International conference on machine learning. pp. 1683-1691. PMLR (2014) 3", + "2. De Geest, R., Gavves, E., Ghodrati, A., Li, Z., Snoek, C., Tuytelaars, T.: Online action detection. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 269-284. Springer (2016) 1, 2, 9", + "3. De Geest, R., Tuytelaars, T.: Modeling temporal structure with lstm for online action detection. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1549-1557. IEEE (2018) 4", + "4. Elahi, G.M.E., Yang, Y.H.: Online temporal classification of human action using action inference graph. Pattern Recognition 132, 108972 (2022) 4", + "5. Eun, H., Moon, J., Park, J., Jung, C., Kim, C.: Learning to discriminate information for online action detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 809-818 (2020) 4, 10, 11", + "6. Eun, H., Moon, J., Park, J., Jung, C., Kim, C.: Temporal filtering networks for online action detection. Pattern Recognition 111, 107695 (2021) 4", + "7. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 10" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BEDL", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Fathullah, Y., Xia, G., Gales, M.J.: Logit-based ensemble distribution distillation for robust autoregressive sequence uncertainties. In: Uncertainty in Artificial Intelligence. pp. 582-591. PMLR (2023) 9", + "19. Franchi, G., Bursuc, A., Aldea, E., Dubuisson, S., Bloch, I.: Encoding the latent posterior of bayesian neural networks for uncertainty quantification. arXiv preprint arXiv:2012.02818 (2020) 3", + "20. Gal, Y., Ghahramani, Z.: Bayesian convolutional neural networks with bernoulli approximate variational inference. arXiv preprint arXiv:1506.02158 (2015) 2", + "21. Gao, J., Yang, Z., Nevatia, R.: Red: Reinforced encoder-decoder networks for action anticipation. arXiv preprint arXiv:1707.04818 (2017) 4, 9, 11", + "22. Gao, J., Chen, M., Xu, C.: Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18827-18836 (2023) 7", + "23. Gao, J., Chen, M., Xu, C.: Vectorized evidential learning for weakly-supervised temporal action localization. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 4", + "24. Gao, M., Zhou, Y., Xu, R., Socher, R., Xiong, C.: Woad: Weakly supervised online action detection in untrimmed videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1915-1923 (2021) 4, 11", + "25. Goodrich, M.A., Schultz, A.C., et al.: Human-robot interaction: a survey. Foundations and Trends in Human-Computer Interaction 1(3), 203-275 (2008) 1", + "26. Guo, H., Aved, A., Roller, C., Ardiles-Cruz, E., Ji, Q.: Video-based complex human event recognition with a probabilistic transformer. In: Geospatial Informatics XIII. vol. 12525, pp. 184-192. SPIE (2023) 3", + "27. Guo, H., Ren, Z., Wu, Y., Hua, G., Ji, Q.: Uncertainty-based spatial-temporal attention for online action detection. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 69-86. Springer (2022) 3, 7, 11, 13, 14", + "28. Han, Y., Tan, S.: Twinlstm: Two-channel LSTM network for online action detection. In: 2022 26th International Conference on Pattern Recognition (ICPR). pp. 3310-3317. IEEE (2022) 4", + "29. Hedegaard, L., Bakhtiarnia, A., Iosifidis, A.: Continual transformers: Redundancy-free attention for online inference. arXiv preprint arXiv:2201.06268 (2022) 4, 11", + "30. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: 2015 IEEE conference on computer vision and pattern recognition (CVPR). pp. 961-970. IEEE (2015) 10", + "31. Idrees, H., Zamir, A.R., Jiang, Y.G., Gorban, A., Laptev, I., Sukthankar, R., Shah, M.: The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding 155, 1-23 (2017) 2, 9", + "32. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015) 10", + "33. Jøsang, A.: Subjective logic, vol. 3. Springer (2016) 7", + "34. Kim, Y.H., Kang, H., Kim, S.J.: A sliding window scheme for online temporal action localization. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV. pp. 653-669. Springer (2022) 4", + "35. Kim, Y.H., Nam, S., Kim, S.J.: Temporally smooth online action detection using cycle-consistent future anticipation. Pattern Recognition 116, 107954 (2021) 11" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Kim, Y.H., Nam, S., Kim, S.J.: 2pesnet: Towards online processing of temporal action localization. Pattern Recognition 131, 108871 (2022) 4", + "37. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10", + "38. Kingma, D.P., Salimans, T., Welling, M.: Variational dropout and the local reparameterization trick. Advances in neural information processing systems 28 (2015) 2", + "39. Kristiadi, A., Hein, M., Hennig, P.: Being bayesian, even just a bit, fixes overconfidence in relu networks. In: International conference on machine learning. pp. 5436-5446. PMLR (2020) 6, 14", + "40. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2016), http://arxiv.org/abs/1612.014743", + "41. Li, Y., Lan, C., Xing, J., Zeng, W., Yuan, C., Liu, J.: Online human action detection using joint classification-regression recurrent neural networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14. pp. 203-220. Springer (2016) 4", + "42. Louizos, C., Welling, M.: Multiplicative normalizing flows for variational bayesian neural networks. In: International Conference on Machine Learning. pp. 2218-2227. PMLR (2017) 3", + "43. Maddox, W.J., Izmailov, P., Garipov, T., Vetrov, D.P., Wilson, A.G.: A simple baseline for bayesian uncertainty in deep learning. In: Wallach, H., Larochelle, H., Beygelzimer, A., e-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32, pp. 13132-13143. Curran Associates, Inc. (2019), http://papers.nips.cc/paper/9472-a-simple-baseline-for-bayesian-uncertainty-in-deep-learning.pdf 3", + "44. Malinin, A., Gales, M.: Predictive uncertainty estimation via prior networks. In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems 31, pp. 7047-7058. Curran Associates, Inc. (2018), http://papers.nips.cc/paper/7936-predictive-uncertainty-estimation-via-prior-networks.pdf 3", + "45. Malinin, A., Gales, M.: Reverse kl-divergence training of prior networks: Improved uncertainty and adversarial robustness. Advances in Neural Information Processing Systems 32 (2019) 3", + "46. Malinin, A., Mlodozeniec, B., Gales, M.: Ensemble distribution distillation. arXiv preprint arXiv:1905.00076 (2019) 3", + "47. Molchanov, D., Ashukha, A., Vetrov, D.: Variational dropout sparsifies deep neural networks. In: International Conference on Machine Learning. pp. 2498-2507. PMLR (2017) 2", + "48. Park, Y., Choi, W., Kim, S., Han, D.J., Moon, J.: Active learning for object detection with evidential deep learning and hierarchical uncertainty aggregation. In: The Eleventh International Conference on Learning Representations (2022) 4", + "49. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017) 10", + "50. Ramanishka, V., Chen, Y.T., Misu, T., Saenko, K.: Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7699-7707 (2018) 2, 10" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BEDL", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Rangrej, S.B., Liang, K.J., Hassner, T., Clark, J.J.: Glitr: Glimpse transformers with spatiotemporal consistency for online action prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3413-3423 (2023) 4", + "52. Sensoy, M., Kaplan, L., Kandemir, M.: Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems 31 (2018) 2", + "53. Sentz, K., Ferson, S.: Combination of evidence in dampster-shafer theory (2002) 7", + "54. Shou, Z., Wang, D., Chang, S.F.: Temporal action localization in untrimmed videos via multi-stage cnns. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1049-1058 (2016) 1", + "55. Sultani, W., Chen, C., Shah, M.: Real-world anomaly detection in surveillance videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6479-6488 (2018) 1", + "56. Ulmer, D.: A survey on evidential deep learning for single-pass uncertainty estimation. arXiv preprint arXiv:2110.03051 (2021) 3", + "57. Ulmer, D., Hardmeier, C., Frellsen, J.: Prior and posterior networks: A survey on evidential deep learning methods for uncertainty estimation. Transactions on Machine Learning Research (2023) 2, 3", + "58. Valdenegro-Toro, M.: Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168 (2019) 3", + "59. Wang, H., Ji, Q.: Diversity-enhanced probabilistic ensemble for uncertainty estimation. In: Uncertainty in Artificial Intelligence. pp. 2214-2225. PMLR (2023) 3", + "60. Wang, J., Chen, G., Huang, Y., Wang, L., Lu, T.: Memory-and-anticipation transformer for online action understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13824-13835 (October 2023) 4, 11, 13", + "61. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016) 10", + "62. Wang, X., Zhang, S., Qing, Z., Shao, Y., Zuo, Z., Gao, C., Sang, N.: Oadtr: Online action detection with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7565-7575 (2021) 4, 11, 14", + "63. Welling, M., Teh, Y.W.: Bayesian learning via stochastic gradient Langevin dynamics pp. 681-688 (2011) 3", + "64. Wen, Y., Tran, D., Ba, J.: Batchsemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715 (2020) 3", + "65. Xu, M., Gao, M., Chen, Y.T., Davis, L.S., Crandall, D.J.: Temporal recurrent networks for online action detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5532-5541 (2019) 4, 9, 10, 11, 14", + "66. Xu, M., Xiong, Y., Chen, H., Li, X., Xia, W., Tu, Z., Soatto, S.: Long short-term transformer for online action detection. Advances in Neural Information Processing Systems 34, 1086-1099 (2021) 4, 10, 11, 13, 14", + "67. Yang, L., Han, J., Zhang, D.: Colar: Effective and efficient online action detection by consulting exemplars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3160-3169 (2022) 4, 11, 14", + "68. Ye, N., Zhang, X., Yan, D., Dong, W., Yan, Q.: Scoad: Single-frame click supervision for online action detection. In: Proceedings of the Asian Conference on Computer Vision. pp. 2156-2171 (2022) 4" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "69. Zeng, R., Huang, W., Tan, M., Rong, Y., Zhao, P., Huang, J., Gan, C.: Graph convolutional networks for temporal action localization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7094-7103 (2019) 1", + "70. Zhang, R., Li, C., Zhang, J., Chen, C., Wilson, A.G.: Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932 (2019) 3", + "71. Zhao, C., Du, D., Hoogs, A., Funk, C.: Open set action recognition via multi-label evidential learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22982-22991 (2023) 2, 3", + "72. Zhao, P., Xie, L., Zhang, Y., Wang, Y., Tian, Q.: Privileged knowledge distillation for online action detection. arXiv preprint arXiv:2011.09158 (2020) 11", + "73. Zhao, Y., Krahenbuhl, P.: Real-time online video detection with temporal smoothing transformers. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV. pp. 485-502. Springer (2022) 4, 11, 14" + ], + "bbox": [ + 212, + 146, + 787, + 340 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BEDL", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Bayesian Evidential Deep Learning for Online Action Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_model.json b/2024/Bayesian Evidential Deep Learning for Online Action Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f382a9551d49fe123e2273bfe7babc44d0f7c992 --- /dev/null +++ b/2024/Bayesian Evidential Deep Learning for Online Action Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_model.json @@ -0,0 +1,2779 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.241, + 0.141, + 0.764, + 0.186 + ], + "angle": 0, + "content": "Bayesian Evidential Deep Learning for Online Action Detection" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.213, + 0.652, + 0.229 + ], + "angle": 0, + "content": "Hongji Guo, Hanjing Wang, and Qiang Ji" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.24, + 0.684, + 0.254 + ], + "angle": 0, + "content": "Rensselaer Polytechnic Institute, Troy NY 12180, USA" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.256, + 0.618, + 0.269 + ], + "angle": 0, + "content": "{guoh11, wangh36, jiq}@rpi.edu" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.302, + 0.74, + 0.593 + ], + "angle": 0, + "content": "Abstract. Online action detection aims at identifying the ongoing action in a streaming video without seeing the future. Timely and reliable response is critical for real-world applications. In this paper, we introduce Bayesian Evidential Deep Learning (BEDL), an efficient and generalizable framework for online action detection and uncertainty quantification. Specifically, we combine Bayesian neural networks and evidential deep learning by a teacher-student architecture. The teacher model is built in a Bayesian manner and transfers its mutual information and distribution to the student model through evidential deep learning. In this way, the student model can make accurate online inference while efficiently quantifying the uncertainty. Compared to existing evidential deep learning methods, BEDL estimates uncertainty more accurately by leveraging the Bayesian teacher model. In addition, we designed an attention module for active OAD, which actively selects important features based on the Bayesian mutual information instead of using all the features. We evaluated BEDL on benchmark datasets including THUMPS'14, TVSeries, and HDD. BEDL achieves competitive performance while keeping efficient inference. Extensive ablation studies demonstrate the effectiveness of each component. To verify the uncertainty quantification, we perform experiments of online anomaly detection with different types of uncertainties." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.606, + 0.741, + 0.634 + ], + "angle": 0, + "content": "Keywords: Online action detection \\(\\cdot\\) Evidential deep learning \\(\\cdot\\) Bayesian neural networks \\(\\cdot\\) Knowledge distillation \\(\\cdot\\) Uncertainty quantification" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.659, + 0.375, + 0.675 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Online action detection (OAD) [12] aims at identifying the ongoing action in a streaming video based on the historical observations. Different from the offline setting [54, 69], it only uses the past information and makes the prediction as soon as the action takes place. It has many important applications such as autonomous driving [9], visual surveillance [55], and human-robot interaction [25]. OAD is challenging due to the incomplete observations of actions and redundant information among inputs such as background and irrelevant actions. In addition, capturing the predictive uncertainty and generalizing to unseen environments are difficult, which are required by most safety-critical applications such as autonomous driving." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "To address these challenges, we introduce Bayesian Evidential Deep Learning (BEDL) for active online action detection and uncertainty quantification. Specifically, we combine Bayesian neural networks and evidential deep learning by a teacher-student architecture. Firstly, combining Bayesian neural networks and evidential deep learning enables accurate and efficient uncertainty quantification. Secondly, we introduce Bayesian mutual information (BMI) for OAD to actively select important features with minimal redundancy and irrelevance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Evidential deep learning (EDL) [1,52,57] is a technique developed for identifying the unknowns by quantifying uncertainty. Specifically, it uses deep neural networks to predict a Dirichlet distribution over the class probabilities, which is treated as an evidence collection process. Then the learned evidence are used to quantify the uncertainty by a single forward pass. EDL is computationally efficient and it has been widely used for open-set recognition [3,4,71] as it enables efficient uncertainty quantification without sampling. However, existing EDL methods cannot accurately model the epistemic and aleatoric uncertainty, which are critical as they provide different perspectives and information about the prediction. On the other hand, Bayesian neural networks (BNN) [20,38,47] can accurately estimate both epistemic and aleatoric uncertainty. But BNN confronts intractability of exact posterior inference and expensive sampling process, which prevents it from practical applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.681 + ], + "angle": 0, + "content": "Therefore, we combine EDL and BNN in this work to take advantages of their respective strengths, i.e. to efficiently and accurately estimate both epistemic and aleatoric uncertainty by a single forward pass. Specifically, we use BNN to augment EDL for uncertainty estimation by a teacher-student model. First, BNN model (teacher model) is built in a Bayesian manner to model the posterior distribution of model parameters. Then we transfer the knowledge of the teacher model to the EDL model (student model) by knowledge distillation (KD). Different from existing KD approaches that distill prediction scores or features, our BEDL distills both the uncertainties and mutual information of teacher model to student model. In this way, the BEDL model inherits the merits of Bayesian neural networks for accurately quantifying both epistemic and aleatoric uncertainty, while preserving its computational efficiency. As a result, the BEDL model can accurately and efficiently estimate both types of uncertainties by a single forward pass. The distilled mutual information as used as for feature selection as below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.793 + ], + "angle": 0, + "content": "To address the background and irrelevant contents issue, we also designed an attention module for the BEDL model based on the distilled Bayesian mutual information (MI). We measure the dependency between the ongoing action and historical features by MI. Then the attention module is learned to generate an attention mask that can select informative features with high MI. In this way, the prediction benefits from better features and the online action detection can be improved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We evaluated BEDL's accuracy and efficiency on benchmark datasets including THUMOS'14 [31], TVSeries [12], and HDD [50]. We also demonstrated BEDL's other properties such as data-efficiency and generalization by extensive" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEDL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "ablation studies. In addition, we validated the effectiveness of EDL by uncertainty quantification and online anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.178, + 0.703, + 0.192 + ], + "angle": 0, + "content": "In summary, the main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.2, + 0.784, + 0.243 + ], + "angle": 0, + "content": "- We introduce BEDL based on combining Bayesian neural newtorks and evidential deep learning, which enables efficient and accurate online inference and uncertainty quantification." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.245, + 0.784, + 0.272 + ], + "angle": 0, + "content": "- We design a mutual information based attention module for the BEDL model to select informative features that are more discriminant and reliable." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.275, + 0.784, + 0.318 + ], + "angle": 0, + "content": "- Our proposed BEDL achieves competitive performance on online action detection benchmark datasets in terms of accuracy and efficiency. And we validated its uncertainty quantification capability." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.2, + 0.784, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.342, + 0.388, + 0.358 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Uncertainty quantification. Bayesian neural networks (BNNs) have traditionally been employed for quantifying uncertainty by considering neural network parameters as probabilistic variables and analyzing their posterior distributions. Techniques for uncertainty quantification (UQ) primarily involve Markov Chain Monte Carlo (MCMC) methods [11, 63, 70] and variational inference (VI) methods [19, 42, 43]. Ensemble approaches [26, 27, 40, 58, 59, 64] also serve as effective alternatives for precise UQ. While traditional BNNs offer several benefits, their need for a wide range of parameter samples for Bayesian inference results in inefficiency in UQ. This inefficiency is due to the requirement for several forward passes through the neural network, each utilizing uniquely sampled parameters from the posterior distribution. To address this issue, evidential deep learning approaches assume the parameters of the target distribution follow a conjugate prior distribution, enabling a closed-form expression for uncertainty that facilitates fast inference. Yet, these methods need additional knowledge to accurately learn the parameters of the prior distribution, including out-of-distribution (OOD) information [44, 45], ensemble models [46], and density estimations [7]. Even with these resources, point-estimation-based evidential networks may not effectively measure epistemic uncertainty, as indicated by [56]. Our strategy, therefore, involves utilizing BNNs to guide the training of evidential networks, aiming at enhancing the accuracy of evidential deep learning models while ensuring high efficiency for inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evidential deep learning (EDL). EDL [3, 4, 71] is proposed to address the challenge of making the model know the unknown when the input data is unfamiliar. In addition, it allows uncertainty estimation in a single model and forward pass [57]. Recently, EDL has been widely used for open-set recognition tasks. Bao et al. introduced DEAR [3] for open-set action recognition. By formulating action recognition in an EDL framework, the unknown actions can be recognized by computing the predictive uncertainty. Zhao et al. proposed multi-label evidential learning (MULE) [71] for open-set action recognition and novelty detection. The proposed evidence debiasing constraint enables the model to handle" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "general problems of single or multiple actors in the same scene, with simultaneous action(s) by any actor. Later, EDL has been applied for open-set action localization [4,10,23]. By quantifying the frame-level uncertainty, the model can distinguish unknown actions from background video framesw. Park et al. proposed MEH-HUA [48] for object detection. By aggregating the uncertainty from EDL in a bottom-up order, the context within the image can be better captured for object detection. In this paper, we introduce EDL for online action detection and anomaly detection, which has not been explored yet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.79, + 0.513 + ], + "angle": 0, + "content": "Online action detection. For the model architecture, RNN-based designs [8, 13, 15, 16, 21, 24, 28, 36, 41, 65, 68] are widely adopted because of RNN's temporal modeling capability. Typically, Xu et al. [65] proposed temporal recurrent network (TRN) that leverages both the historical information and predicted future features to detect the ongoing action. Thanks to the self-attention mechanism and the parallel computing property, Transformer-based methods [5, 29, 34, 51, 60, 62, 66, 67, 73] become the mainstream for online action detection. Wang [62] proposed OadTR that makes use of both historical information and future prediction. Xu [62] proposed long short-term Transformer (LSTR) that captures both the long-range and short-term dependencies by two memory units. To overcome the latency of feature extraction, [5] proposed E2E-LOAD for end-to-end online action detection. Besides RNN and Transformer, graph modeling is also studied for online action detection [14]. To leverage the video-level annotations instead of the dense frame-level annotations, weakly-supervised methods detection [24, 68] are also explored for OAD." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.348, + 0.575 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.658 + ], + "angle": 0, + "content": "In this section, we first give an overview of the proposed Bayesian evidential deep learning (§ 3.1) and formulate the online action detection problem (§ 3.2). Then we introduce BEDL's Bayesian teacher model (§ 3.3) and evidential student model (§ 3.4). In the end, we show the training procedures (§ 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.687, + 0.342, + 0.701 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The overall framework of BEDL is shown in Figure 1. BEDL is built as a teacher-student framework, which is composed of a Bayesian neural network as the teacher model and an evidential neural network as the student model. Given a streaming video as input, a pretrained backbone first extracts features. Then the Bayesian tearcher model outputs the action predictions and mutual information based on the features, which are used to train the evidential student model by knowledge distillation. During the testing, only the student model is kept to perform fast inference and efficient uncertainty quantification." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEDL" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.147, + 0.78, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.438 + ], + "angle": 0, + "content": "Fig. 1: Overall framework of Bayesian evidential deep learning (BEDL). The input of model is a streaming video. A pretrained backbone is used to extract the features for each frame. During the training, the Bayesian teacher model generates the mutual information (MI) and multiple sets of predictions. The evidential student model leverages MI to train the attention module for feature selection and predictions for uncertainty quantification. During testing, the student model actively selects informative features and quantifies predictive uncertainty accurately and efficiently by a single forward pass." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.434, + 0.481 + ], + "angle": 0, + "content": "3.2 Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.789, + 0.629 + ], + "angle": 0, + "content": "Online action detection(OAD) aims at recognizing the ongoing action in a streaming video with only the past and current observations. Denote the input video as \\(\\mathbf{V} = [I_1, I_2, \\dots, I_T]\\), where \\(T\\) is the length of video and \\(I_t\\) denotes the frame at current time \\(t\\). The online action detection is formulated as a classification problem: \\(y_t^* = \\operatorname{argmax}_c p(\\hat{y}_t = c | \\mathbf{V}_t)\\), where \\(\\hat{y}_t\\) is the prediction, \\(c\\) is the class label, and \\(\\mathbf{V}_t = \\{I_1, \\dots, I_t\\}\\) is the available frame set at time \\(t\\). A feature extractor is used to process each frame and generate the corresponding feature vector. Denote the feature set at time \\(t\\) as \\(\\mathbf{F}^t = \\{F_1^t, \\dots, F_t^t\\}\\). The feature at time \\(i\\) is \\(F_i^t \\in \\mathbb{R}^J\\), where \\(J\\) is the feature dimension of each frame." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.457, + 0.665 + ], + "angle": 0, + "content": "3.3 Bayesian teacher model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "To take advantages of Bayesian neural networks for modeling both epistemic and aleatoric uncertainty, we build a Bayesian teacher model (BTM). BTM's objectives include: 1) generating mutual information to train the attention module for feature selection. 2) providing distribution knowledge for the student model for uncertainty quantification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The mutual information (MI) between past features and ongoing action indicates the relevance of features. Denote a past feature as \\( F_{ij}^{t} \\), where \\( i \\in \\{1, \\dots, t\\} \\) is the time index, \\( j \\in \\{1, \\dots, J\\} \\) is the feature index within each frame, and \\( t \\) denotes the current time. We aim to obtain the mutual information between \\( F_{ij}^{t} \\) and the ongoing action \\( y_{t} \\). An illustration is shown in Figure 2. To compute mutual information, we take advantage of Bayesian modeling. Different from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "point estimation, Bayesian method constructs a posterior distribution of model parameters. By integrating predictions from multiple models, it is less likely to be overfitting and the predictive uncertainty can be accurately quantified. Additionally, Bayesian method is more robust when training data is insufficient. We term the mutual information computed using the Bayesian method Bayesian Mutual Information (BMI)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.265, + 0.391, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.345, + 0.392, + 0.373 + ], + "angle": 0, + "content": "Fig. 2: Illustration of mutual information in OAD." + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.264, + 0.772, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.346, + 0.771, + 0.377 + ], + "angle": 0, + "content": "Fig. 3: Illustration of distribution distillation by Bayesian evidential deep learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.789, + 0.451 + ], + "angle": 0, + "content": "Denote the model parameters of the teacher model as \\(\\theta\\) and we treat them as probability distributions. Then the BMI between a past feature \\(F_{ij}^{t}\\) and the ongoing action \\(y_{t}\\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.461, + 0.787, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} \\left[ y _ {t}; F _ {i j} ^ {t} \\mid \\mathcal {D} \\right] = \\mathcal {H} \\left[ y _ {t} \\mid \\boldsymbol {F} _ {- i j} ^ {t}, \\mathcal {D} \\right] - \\mathcal {H} \\left[ y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.789, + 0.54 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) denotes the training data, \\(\\mathcal{H}\\) denotes the entropy, and \\(F_{-ij}^{t}\\) is the feature set at time \\(t\\) excluding \\(F_{ij}^{t}\\), i.e. \\(F_{-ij}^{t} = F^{t} / F_{ij}^{t}\\). By definition, the entropy term is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.549, + 0.785, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} \\left[ y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D} \\right] = - \\sum_ {y _ {t} \\in \\mathcal {Y}} p \\left(y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D}\\right) \\log p \\left(y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.64 + ], + "angle": 0, + "content": "where \\(\\mathcal{Y} = \\{0,1,\\dots,C\\}\\) is the action class set. 0 represents background class and \\(C\\) is the number of action classes. The prediction distribution in Eq. (2) can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.662, + 0.787, + 0.695 + ], + "angle": 0, + "content": "\\[\np (y _ {t} | \\boldsymbol {F} ^ {t}, \\mathcal {D}) = \\int p (y _ {t} | \\boldsymbol {F} ^ {t}, \\theta) p (\\theta | \\mathcal {D}) d \\theta \\approx \\frac {1}{K} p (y _ {t} | \\boldsymbol {F} ^ {t}, \\theta_ {k}), w h e r e \\theta_ {k} \\sim p (\\theta | \\mathcal {D}) (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.751 + ], + "angle": 0, + "content": "In Eq. (3), we use the sample average to approximate it since it is impractical to integrate over all the possible parameters. Similarly, we can compute \\(\\mathcal{H}[y_t|F_{-ij}^t,\\mathcal{D}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.751, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To obtain the posterior distribution \\( p(\\theta | \\mathcal{D}) \\), we perform the Laplace approximation (LA). To reduce the computation cost, we adopt a last-layer Bayesian setting [39]. Only the posterior distribution of last-layer parameters is modeled, while keeping the remaining parameters deterministic. Later, we will show this is efficient and effective. Specifically, we assume the last-layer parameters follow a Gaussian distribution." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEDL" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Firstly, we train the teacher model under the deterministic setting by a cross entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.19, + 0.785, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta^ {*} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\mathcal {L} (\\mathcal {D}; \\theta) = \\underset {\\theta} {\\operatorname {a r g m i n}} \\left(\\sum_ {n = 1} ^ {N} l \\left(x _ {n}, y _ {n}; \\theta\\right) + r (\\theta)\\right) \\tag {4} \\\\ = \\operatorname * {a r g m i n} _ {\\theta} \\left(\\sum_ {n = 1} ^ {N} - \\log p \\left(y _ {n} \\mid f _ {\\theta} \\left(x _ {n}\\right)\\right) - \\log p (\\theta)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.784, + 0.327 + ], + "angle": 0, + "content": "where \\(\\mathcal{D} = \\{x_n, y_n\\}_{n=1}^N\\) is the training set and \\(r(\\theta)\\) is a regularizer such as weigh regularizer (a.k.a. weight decay). So \\(\\theta^*\\) is indeed a maximum a posteriori (MAP) estimate." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.33, + 0.784, + 0.361 + ], + "angle": 0, + "content": "The Laplace approximation uses a second-order Taylor expansion of \\(\\mathcal{L}(\\mathcal{D};\\theta)\\) around \\(\\theta^{*}\\) to construct a Gaussian approximation to \\(p(\\theta |\\mathcal{D})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.372, + 0.785, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathcal {D}; \\theta) \\approx \\mathcal {L} (\\mathcal {D}; \\theta^ {*}) + \\frac {1}{2} \\left(\\theta - \\theta^ {*}\\right) ^ {T} \\left(\\nabla_ {\\theta} ^ {2} \\mathcal {L} (\\mathcal {D}; \\theta) | _ {\\theta^ {*}}\\right) \\left(\\theta - \\theta^ {*}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.408, + 0.784, + 0.438 + ], + "angle": 0, + "content": "where the first-order term vanishes at \\(\\theta^{*}\\). So the Laplace posterior approximation can be obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.449, + 0.785, + 0.467 + ], + "angle": 0, + "content": "\\[\np (\\theta | \\mathcal {D}) \\approx \\mathcal {N} \\left(\\theta^ {*}, \\Sigma\\right), \\text {w i t h} \\Sigma := \\left(\\nabla_ {\\theta} ^ {2} \\mathcal {L} (\\mathcal {D}; \\theta) | _ {\\theta^ {*}}\\right) ^ {- 1} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.477, + 0.785, + 0.596 + ], + "angle": 0, + "content": "After the LA, we sample \\(K\\) times from \\(p(\\theta|\\mathcal{D})\\) to obtain \\(K\\) sets of parameters \\(\\{\\theta_1,\\dots,\\theta_K\\}\\). Then we approximate the prediction distribution in Eq. (3) and further compute the mutual information of each feature. The BMI we computed are based on multiple models from the teacher model. We use BMI to supervise the attention module of student model as the importance of past feature is measured by MI. Different from the traditional MI estimated from probabilistic models such as Uncertainty-OAD [27], BEDL uses BMI and hence is more robust and accurate." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.471, + 0.633 + ], + "angle": 0, + "content": "3.4 Evidential student model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.784, + 0.81 + ], + "angle": 0, + "content": "Evidential deep learning (EDL). Softmax-based deep neural networks are unable to accurately estimate the predictive uncertainty since the softmax probability is a point estimation of the predictive distribution and it tends to be over-confident in wrong predictions. To address such issue, EDL is proposed by combing Dempster-Shafer Theory [53] and subjective logic [33]. Specifically, it assumes the classification probability follows a prior Dirichlet distribution. The training is considered as a process of collecting evidence. The learned parameters of Dirichlet distribution are used as evidence to compute the predictive uncertainty. In this way, the uncertainty can be estimated in a single forward pass. In this work, we build the evidential student model under the EDL framework so that it can output both epistemic and aleatoric uncertainty in a forward pass." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Distribution distillation. Common knowledge distillation (KD) methods aim at transferring prediction scores, features, or evidence [22], so the student model" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "cannot estimate the predictive uncertainty. Differently, we transfer the uncertainty of the Bayesian teacher model to the evidential student model, i.e. distilling the distribution. In this way, the student model can learn from predictions of the teacher model as well as the capability of capturing the uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.212, + 0.79, + 0.364 + ], + "angle": 0, + "content": "Under the classification setting, the target output of student model \\(y\\) follows a categorical distribution with parameter \\(\\lambda\\), \\(y \\sim p(y|\\lambda) = Cat(\\lambda)\\). For example, \\(\\lambda\\) represents the probability after the final softmax layer. We treat \\(\\lambda\\) as a random variable and assume it follows Dirichlet distribution, i.e. \\(\\lambda \\sim p(\\lambda|\\alpha(x,\\psi)) = Dir(\\alpha(x,\\theta))\\), where \\(\\alpha\\) denotes the parameters of the Dirichlet distribution and \\(\\psi\\) denotes student model parameters. Similarly, the teacher model has posterior distribution \\(p(\\lambda|x,\\mathcal{D}) = \\int p(\\lambda|x,\\theta)p(\\theta|\\mathcal{D})d\\theta\\). During the distillation, we transfer teacher posterior distribution \\(p(\\lambda|x,\\mathcal{D})\\) to the student posterior distribution \\(p(\\lambda|\\alpha(x,\\psi))\\). An illustration is shown in Figure 3. Specifically, we minimize the KL-divergence between these two distributions as below:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.385, + 0.786, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {d i s} = K L (p (\\lambda | x, \\mathcal {D}) | | p (\\lambda | \\alpha (x, \\psi))) \\\\ \\propto - \\sum_ {c = 1} ^ {C} \\log (\\Gamma (\\alpha_ {c})) + \\log \\Gamma (\\sum_ {c = 1} ^ {C} \\alpha_ {c}) - \\mathbb {E} _ {p (\\theta | \\mathcal {D})} [ \\sum_ {c = 1} ^ {C} (\\alpha_ {c} - 1) \\log \\lambda_ {c} (x, \\theta) ] \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.527 + ], + "angle": 0, + "content": "where \\( C \\) is the number of action classes. Detailed derivation of \\( \\mathcal{L}_{dis} \\) can be found in Appendix. Since the different predictions of the teacher model are combined, the model is more robust. The complete distribution distillation algorithm is available in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.794, + 0.683 + ], + "angle": 0, + "content": "Mutual information based attention module. The attention module actively selects informative features from the inputs by generating a spatial-temporal attention mask \\( A_{t} \\) using a fully-connected network and applying to the original features by element-wise product. In this way, irrelevant features are masked out since they have low mutual information to the ongoing action. The training of the attention module is supervised by the BMI \\( \\mathcal{I}_t \\) from the teacher model. In this way, the BMI from the Bayesian teacher model is distilled to the student model. By minimizing the mean squared error (MSE) loss \\( \\mathcal{L}_{att} \\) between \\( A_{t} \\) and \\( \\mathcal{I}_t \\), the attention module can directly generate the BMI-aware attention mask without computing the BMI. \\( \\mathcal{L}_{att} \\) can be written as below:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.703, + 0.786, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a t t} = M S E \\left(\\boldsymbol {A} _ {t}, \\sigma \\left(\\mathcal {I} _ {t}\\right)\\right) = \\frac {1}{t J} \\sum_ {i, j} \\left| \\left| A _ {i j} ^ {t} - \\sigma \\left(\\mathcal {I} \\left[ y _ {t}; F _ {i j} ^ {t} \\mid \\mathcal {D} \\right]\\right) \\right| \\right| ^ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.76, + 0.468, + 0.776 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is the sigmoid function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Uncertainty quantification. After the distillation through evidential deep learning, the Dirichlet distribution of the student model gains the knowledge of the Bayesian teacher model. The uncertainty can be computed efficiently in a closed-form solution by a single forward pass. Specifically, the total uncertainty" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEDL" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.557, + 0.162 + ], + "angle": 0, + "content": "and epistemic uncertainty can be quantified as:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.174, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} [ p (y | x, \\theta) ] = \\sum_ {c = 1} ^ {C} \\frac {\\alpha_ {c}}{\\alpha_ {0}} \\log \\frac {\\alpha_ {c}}{\\alpha_ {0}}, \\text {w h e r e} \\alpha_ {0} = \\sum_ {c = 1} ^ {C} \\alpha_ {c} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.219, + 0.682, + 0.257 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} [ y; \\lambda | \\alpha ] = - \\sum_ {c = 1} ^ {C} \\frac {\\alpha_ {c}}{\\alpha_ {0}} \\big (\\ln \\frac {\\alpha_ {c}}{\\alpha_ {0}} \\varPsi (\\alpha_ {c} + 1) + \\varPsi (\\alpha_ {0} + 1) \\big)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.268, + 0.788, + 0.33 + ], + "angle": 0, + "content": "where \\(\\varPsi(\\cdot)\\) is the dgamma function. To be simple, the aleatoric uncertainty can be computed by taking the difference between total uncertainty and epistemic uncertainty. The uncertainty can be also used to detect the anomaly. Details derivations are available in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.334, + 0.367 + ], + "angle": 0, + "content": "3.5 Training" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.376, + 0.784, + 0.437 + ], + "angle": 0, + "content": "The evidential student model is jointly trained for online action detection, attention distillation, and distribution distillation. Specifically, we adopt a two-strategy. We first train the attention module using the attention loss \\(\\mathcal{L}_{att}\\) in Eq. 8. Then we jointly train the evidential student model by total loss \\(\\mathcal{L}\\) below:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.449, + 0.785, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\lambda_ {1} \\mathcal {L} _ {a t t} + \\lambda_ {2} \\mathcal {L} _ {d i s} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.474, + 0.788, + 0.625 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{ce}\\) is the cross-entropy loss for online action detection. \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyper-parameters that emphasize the attention module and distribution distillation. Although the distribution distillation can make the student model perform online action detection, we still perform joint training because the Bayesian teacher model is not perfect and its posterior distribution is approximated, which may introduce error to the student model. The joint training brings two benefits: 1) the online action detection can be improved with the supervision of ground-truth label; 2) the negative log-likelihood \\(\\mathcal{L}_{ce}\\) makes the training faster and more stable as indicated in [18]. We demonstrate the two-stage training works better in Sec. 4.6." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.376, + 0.667 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.527, + 0.695 + ], + "angle": 0, + "content": "4.1 Datasets and evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.784, + 0.779 + ], + "angle": 0, + "content": "THUMOS'14 [31] is a dataset for video-based temporal action localization. We use the validation set with 200 videos for training and test set with 213 videos for evaluation. There are 20 action classes and a background class. We ignore the frames with ambiguous labels. Following the settings in [21, 65], we adopt the mean Average Precision (mAP) as the evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "TVSeries [12] is a dataset collected from real TV series. It contains 27 episodes with 30 daily actions. It is a challenging dataset due to the viewpoints changing and occlusions in the videos. To counter the imbalanced data distribution, we adopt the mean calibrated Average Precision (mcAP) [12] as the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "evaluation metric. It is computed as \\( cAP = \\sum_{k}cPrec(k)\\times \\mathbb{1}(k) / P \\), where \\( cPrec = TP / (TP + FP / \\omega) \\), \\( P \\) is the total number of positive frames and \\( \\mathbb{1}(k) \\) is an indicator function that is equal to 1 if frame \\( k \\) is a true positive. The mcAP is the mean of calibrated average precision of all action classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.299 + ], + "angle": 0, + "content": "HRI Driving Dataset (HDD) [50] is a dataset for learning driver behavior in real-life environments. It contains 104 hours of real human driving in the San Francisco Bay Area collected by an instrumented vehicle with different sensors. There are 11 goal-oriented driving actions. Following the settings in [50], we use 100 sessions for training and 37 sessions for testing. And only the sensor data is used as the input. mAP is used as the evaluation metric for this dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.24, + 0.31, + 0.472, + 0.35 + ], + "angle": 0, + "content": "Table 1: Online anomaly detection results on THUMOS'14 and TVSeries." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.352, + 0.469, + 0.408 + ], + "angle": 0, + "content": "
UncertaintyTHUMOS'14 (%)TVSeries (%)
EDL [3]64.7242.39
Total81.0465.16
Aleatoric62.5139.83
Epistemic86.4869.44
" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.31, + 0.626, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.31, + 0.772, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.488, + 0.394, + 0.778, + 0.437 + ], + "angle": 0, + "content": "Fig.4: ROC curves of online anomaly detection using epistemic uncertainty on THUMOS'14 (left) and TVSeries (right)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.457, + 0.47 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Feature extraction. Following the settings in [15, 65, 66], we use TSN [61] to extract the features for THUMOS'14 and TVSeries. Video frames are extracted at 24 fps and the chunk size is set to 6. To better capture the spatial-temporal dependencies, we adopt the multi-scale vision Transformer [17] to extract RGB features. The optical flow features are extracted with BN-Inception [32]. The backbone is pretrained on ActivityNet [30] and Kinetics-400 [6] separately for evaluation. For HDD dataset, the sensor data is used as the input. More details can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.706 + ], + "angle": 0, + "content": "Settings. The BEDL is implemented in PyTorch [49]. The model is trained by the Adam optimizer [37] with a learning rate of \\(10^{-4}\\) and a weight decay of \\(5 \\times 10^{-5}\\). The batch size is set to 32. The experiments were conducted on two Nvidia RTX 3090 Ti GPUs. The number of epochs is set to 20. Ablation studies on existing methods are based on their officially released codes. Detailed model architectures are available in the supplementary. And ccode will be made publicly available." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.515, + 0.74 + ], + "angle": 0, + "content": "4.3 Online action detection results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We evaluate BEDL on three benchmark datasets and make a comparison with other methods in Table 2. From the results, our proposed BEDL achieves state-of-the-art performance on most benchmarks. On THUMOS'14, BEDL achieves \\(70.6\\%\\) and \\(72.7\\%\\) mAP using ActivityNet and Kinetics pretrained features. On TVSeries, BEDL achieves \\(88.9\\%\\) and \\(90.1\\%\\) mAP using ActivityNet and Kinetics features respectively. On HDD, we only use the sensor data as the input and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEDL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "achieve \\(33.0\\%\\) mAP, which is higher than the SOTA MAT [60]. The OAD performance demonstrate the effectiveness of BEDL for accurate online inference. In the following parts, we show other superior properties of BEDL while keeping accurate inference." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.276 + ], + "angle": 0, + "content": "Table 2: Experiment results on THUMOS'14, TVSeries and HDD. The results on THUMOS'14 and HDD are reported as mAP (\\%). The results on TVSeries are reported as mCAP (\\%). For HDD, \\(\\star\\) indicates RGB data is used as the input." + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.288, + 0.744, + 0.557 + ], + "angle": 0, + "content": "
MethodArchitectureTHUMOS'14TVSeriesHDD
ANetKineticsANetKineticsSensor
RED [21]RNN45.3-79.2-27.4
FATS [35]-59.081.784.6-
TRN [65]47.262.183.786.229.2
IDN [15]50.060.384.786.1-
PKD [72]-64.5-86.4-
WOAD [24]-67.1---
OadTR [62]Transformer58.365.285.487.229.8
CoOadTR [29]56.164.287.687.730.6
Colar [67]59.466.986.088.130.6
LSTR [66]65.369.588.189.1-
Uncertainty-OAD [27]66.069.988.389.330.1
TeSTra [73]68.271.2---
GateHUB [8]69.170.788.489.632.1
MiniROAD [2]69.371.888.589.6-
MAT [60]70.471.688.689.732.7
E2E-LOAD [5]-72.4-90.348.1*
BEDL (ours)70.672.788.990.133.0
" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.588, + 0.727, + 0.604 + ], + "angle": 0, + "content": "4.4 Uncertainty quantification and online anomaly detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.764 + ], + "angle": 0, + "content": "To verify the uncertainty quantification of BEDL, we quantified different types of uncertainties and perform online anomaly detection on THUMOS'14 and TVSeries. Specially, we divide the data into known classes and unknown classes (anomaly). Unknown classes do not appear in the training set. During testing, both known and unknown data will appear and the model is required to identify if the input belong to known or unknown classes. For THUMOS'14, we treat the \"ambiguous\" class that are difficult to identify during labeling process as anomaly. For TVSeries, we randomly select 5 out of 30 classes as the anomaly and repeat the process for ten times. Inputs that lead to high predictive uncertainty above the pre-defined threshold are declared as anomalies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The experiment results are shown in Table 1 and ROC curves are plotted in Figure 4. From the results, BEDL can accurately identify the unknown classes during the testing, which validates the effectiveness of its uncertainty quantification. Since epistemic uncertainty represents the lack of knowledge of the model and it is inversely proportional to the data density, it has better anomaly detec" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "tion accuracy compared to aleatoric uncertainty, which represents the noise level of the data. Detailed experiment procedures can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.203, + 0.655, + 0.219 + ], + "angle": 0, + "content": "4.5 Computational efficiency and model complexity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.383 + ], + "angle": 0, + "content": "We leverage evidential deep learning for BEDL in order to make it more practical for real-time applications. A comparison of computation efficiency and model complexity is shown in Table 3. While keeping a high detection accuracy, our proposed BEDL has much less model complexity and computational cost. And the inference speed of the model is much higher compared to other methods. Recently, [5] proposed an end-to-end OAD framework to avoid the latency from the frame and feature extraction parts, which leads to higher overall inference speed. To make fair comparisons, we used the same feature extraction procedures as prior works and we mainly compare the model inference speed since it only depends on the model architecture and pipeline." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.41, + 0.787, + 0.452 + ], + "angle": 0, + "content": "Table 3: Comparison of computation efficiency and model complexity. Our proposed BEDL has less model complexity and computational cost. And the inference speed is much faster than other methods." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.464, + 0.784, + 0.561 + ], + "angle": 0, + "content": "
MethodModalityModelInference Speed (FPS)
Parameter CountGFLOPsOptical Flow ComputationRGB Feature ExtractionFlow Feature ExtractionModel
TRN402.9M1.468.170.514.6123.3
OadTR75.8M2.548.170.514.6110.0
LSTRRGB + Flow58.0M7.538.170.514.691.6
GateHUB45.2M6.988.170.514.671.2
MAT94.6M-8.170.514.672.6
BEDL (ours)19.4M0.488.170.514.6163.2
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.615, + 0.398, + 0.629 + ], + "angle": 0, + "content": "4.6 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Training strategies. To demonstrate the effectiveness of two-stage training strategy, we compared it with two other training strategies. One is jointly training the attention module and EDL model by the total loss in Eq. 10. The other is firstly training the attention module by the attention loss and then we freezing the weights of attention module to train the EDL model by the \\(\\mathcal{L}_{att}\\) and \\(\\mathcal{L}_{ce}\\). We refer the first one as joint training and the second one as separate training. The performance comparison is shown in Table 4. The results show that the two-stage training outperforms the other two strategies on both THUMOS'14 and TVSeries with different features, which demonstrates its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Number of historical frames. At time \\( t \\), BEDL takes a certain number of past frames to predict the ongoing action. To study the long-range and short-term modeling capability of BEDL, we vary the number of past frames as the input. The experiment results are shown in Table 5. THUMOS'14 and TVSeries" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEDL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Ablation study of train- Table 5: Ablation study of different number ing strategies. The two-stage training of historical frames. gives the best performance." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.199, + 0.48, + 0.256 + ], + "angle": 0, + "content": "
TrainingTHUMOS'14TVSeries
ActivityNetKineticsActivityNetKinetics
Joint69.570.785.488.2
Separate64.967.881.684.0
Two-stage70.672.788.990.1
" + }, + { + "type": "table", + "bbox": [ + 0.488, + 0.186, + 0.783, + 0.266 + ], + "angle": 0, + "content": "
DatasetFeatureNumber of past frames
8163264128
THUMOS'14ANet41.659.268.770.666.5
Kinetics49.361.671.472.770.0
TVSeriesANet63.576.787.688.984.9
Kinetics68.279.388.590.188.2
HDDSensor30.431.833.032.130.2
" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.291, + 0.359, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.369, + 0.357, + 0.381 + ], + "angle": 0, + "content": "(a) THUMOS'14-ANet" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.292, + 0.501, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.369, + 0.5, + 0.381 + ], + "angle": 0, + "content": "(b) THUMOS'14-K400" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.292, + 0.64, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.369, + 0.629, + 0.381 + ], + "angle": 0, + "content": "(c) TVSeries-ANet" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.292, + 0.781, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.369, + 0.77, + 0.381 + ], + "angle": 0, + "content": "(d) TVSeries-K400" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.393, + 0.789, + 0.462 + ], + "angle": 0, + "content": "Fig. 5: Experiment results of training with small-scale data. We reduce the training data from \\(100\\%\\) to \\(10\\%\\) and compared with LSTR [66], MAT [60], and E2E-LOAD [5]. The results are plotted for THUMOS'14 and TVSeries with both ActivityNet and Kinetics pretrained features. Our proposed BEDL outperforms all others when training data is limited." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.789, + 0.534 + ], + "angle": 0, + "content": "are extracted at 6 fps. HDD is extracted at 3 fps. So the optimal number of frames on HDD is 32." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Data-efficiency. By applying the mutual information based attention mechanism, we expect BEDL to be more data-efficient when the amount of training data is limited. We reduce the training data from \\(100\\%\\) to \\(10\\%\\) and compare with other methods. The results on THUMOS'14 and TVSeries are plotted in Figure 5. When training data is reduced, BEDL has less performance decay and outperforms other methods, which demonstrates that BEDL is more data-efficient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Loss function. In the second training stage of BEDL, the total loss in Eq. 10 is composed of three terms. \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are the weights for attention loss and distribution distillation loss. To study the balance of all three terms, we train the model with different \\(\\lambda\\)s on THUMOS'14 and TVSeries. The comparisons are plotted in Figure 6. We empirically choose \\(\\lambda_{1} = 0.4\\) and \\(\\lambda_{2} = 6\\) since they give the best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Generalization. To test the generalization capability of the model, we perform the Cross-View and under occlusion experiments on TVSeries dataset. Following the settings in [27], the training set and test set are from different view angles. For the occlusion, the training data does not contain occlusion and the testing data is occluded. The experiment results and comparison are shown in Table 6. The results show that BEDL generalizes better under different conditions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.157, + 0.394, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.157, + 0.544, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.252, + 0.529, + 0.293 + ], + "angle": 0, + "content": "Fig.6: Ablation study of loss function for \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) on THUMOS'14 with ActivityNet and Kinetics features." + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.158, + 0.785, + 0.185 + ], + "angle": 0, + "content": "Table 6: Cross-view and occlusion experiments results on TVSeries." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.186, + 0.782, + 0.289 + ], + "angle": 0, + "content": "
MethodCV (%)Occ. (%)
TRN [65]65.885.2
OadTR [62]66.287.7
U-OAD [27]67.389.5
Colar [67]66.788.3
LSTR [66]69.589.4
TeSTra [73]70.289.9
BEDL (ours)74.391.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.789, + 0.429 + ], + "angle": 0, + "content": "Attention module. To verify the effectiveness Bayesian mutual information, we first compare with the model without BMI. Specifically, we trained the student model with the same evidential neural network without the attention module. The total loss function is \\(\\mathcal{L} = \\mathcal{L}_{ce} + \\lambda_2\\mathcal{L}_{dis}\\) and we grid-searched \\(\\lambda_{2}\\) to obtain the best accuracy. In addition, we compute the mutual information based on a single prediction instead of BMI and trained the model by the same loss function in Eq. 10. The results are shown in Table 7. From the results, the BEDL with BMI outperforms other two methods, which demonstrates the effectiveness of attention mechanism using BMI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.796, + 0.537 + ], + "angle": 0, + "content": "Last-layer Bayesian. For the teacher model, we adopt the last-layer Bayesian [39] method to reduce the difficulty of training. We also performed the Laplace approximation over all the model parameters. The comparison is shown in Table 8. The full-Bayesian method has consistent improvement on THUMOS-14 and TVSeries. But the LA in training takes much longer time than the last-layer Bayesian and the model needs careful tuning, so we adopt the last-layer Bayesian in the LA process." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.563, + 0.505, + 0.62 + ], + "angle": 0, + "content": "Table 7: Effectiveness of BMI. BEDL-No-BMI denotes the model without attention module and BEDL-MI denotes the model with non-Bayesian MI." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.632, + 0.501, + 0.687 + ], + "angle": 0, + "content": "
ModelTHUMOS'14TVSeries
ActivityNetKineticsActivityNetKinetics
BEDL-No-BMI63.164.886.587.4
BEDL-MI67.668.488.088.5
BEDL70.672.788.990.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.563, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Table 8: Comparison of full-Bayesian and last-layer Bayesian. Full-Bayesian improves the performance since it models the distribution of all model parameters." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.646, + 0.785, + 0.688 + ], + "angle": 0, + "content": "
MethodTHUMOS-14 (mAP %)TVSeries (mcAP %)
ActivityNetKineticsActivityNetKinetics
Last-layer70.672.788.990.1
Full-Bayesian71.173.090.290.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.744, + 0.36, + 0.76 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Conclusions. In this paper, we introduce Bayesian evidential deep learning for active online action detection and uncertainty quantification. By combining Bayesian neural networks and evidential deep learning, BEDL can accurately and efficiently quantify predictive uncertainty and make online inference." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEDL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.201 + ], + "angle": 0, + "content": "1. Amini, A., Schwarting, W., Soleimany, A., Rus, D.: Deep evidential regression. Advances in Neural Information Processing Systems 33, 14927-14937 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.202, + 0.785, + 0.242 + ], + "angle": 0, + "content": "2. An, J., Kang, H., Han, S.H., Yang, M.H., Kim, S.J.: Miniroad: Minimal rnn framework for online action detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10341-10350 (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.282 + ], + "angle": 0, + "content": "3. Bao, W., Yu, Q., Kong, Y.: Evidential deep learning for open set action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13349-13358 (2021) 2, 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.323 + ], + "angle": 0, + "content": "4. Bao, W., Yu, Q., Kong, Y.: Opental: Towards open set temporal action localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2979-2989 (2022) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.365 + ], + "angle": 0, + "content": "5. Cao, S., Luo, W., Wang, B., Zhang, W., Ma, L.: E2e-load: End-to-end long-form online action detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10422-10432 (October 2023) 4, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "6. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.445 + ], + "angle": 0, + "content": "7. Charpentier, B., Zügner, D., Gunnemann, S.: Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in Neural Information Processing Systems 33, 1356-1367 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.5 + ], + "angle": 0, + "content": "8. Chen, J., Mittal, G., Yu, Y., Kong, Y., Chen, M.: Gatehub: Gated history unit with background suppression for online action detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19925-19934 (2022) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "9. Chen, L., Ma, N., Wang, P., Li, J., Wang, P., Pang, G., Shi, X.: Survey of pedestrian action recognition techniques for autonomous driving. Tsinghua Science and Technology 25(4), 458-470 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "0. Chen, M., Gao, J., Yang, S., Xu, C.: Dual-evidential learning for weakly-supervised temporal action localization. In: European Conference on Computer Vision. pp. 192-208. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "1. Chen, T., Fox, E., Guestrin, C.: Stochastic gradient hamiltonian monte carlo. In: International conference on machine learning. pp. 1683-1691. PMLR (2014) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.664 + ], + "angle": 0, + "content": "2. De Geest, R., Gavves, E., Ghodrati, A., Li, Z., Snoek, C., Tuytelaars, T.: Online action detection. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 269-284. Springer (2016) 1, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "3. De Geest, R., Tuytelaars, T.: Modeling temporal structure with lstm for online action detection. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1549-1557. IEEE (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.785, + 0.732 + ], + "angle": 0, + "content": "4. Elahi, G.M.E., Yang, Y.H.: Online temporal classification of human action using action inference graph. Pattern Recognition 132, 108972 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "5. Eun, H., Moon, J., Park, J., Jung, C., Kim, C.: Learning to discriminate information for online action detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 809-818 (2020) 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Eun, H., Moon, J., Park, J., Jung, C., Kim, C.: Temporal filtering networks for online action detection. Pattern Recognition 111, 107695 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 10" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Fathullah, Y., Xia, G., Gales, M.J.: Logit-based ensemble distribution distillation for robust autoregressive sequence uncertainties. In: Uncertainty in Artificial Intelligence. pp. 582-591. PMLR (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "19. Franchi, G., Bursuc, A., Aldea, E., Dubuisson, S., Bloch, I.: Encoding the latent posterior of bayesian neural networks for uncertainty quantification. arXiv preprint arXiv:2012.02818 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.257 + ], + "angle": 0, + "content": "20. Gal, Y., Ghahramani, Z.: Bayesian convolutional neural networks with bernoulli approximate variational inference. arXiv preprint arXiv:1506.02158 (2015) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "21. Gao, J., Yang, Z., Nevatia, R.: Red: Reinforced encoder-decoder networks for action anticipation. arXiv preprint arXiv:1707.04818 (2017) 4, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.336 + ], + "angle": 0, + "content": "22. Gao, J., Chen, M., Xu, C.: Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18827-18836 (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.337, + 0.785, + 0.379 + ], + "angle": 0, + "content": "23. Gao, J., Chen, M., Xu, C.: Vectorized evidential learning for weakly-supervised temporal action localization. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.433 + ], + "angle": 0, + "content": "24. Gao, M., Zhou, Y., Xu, R., Socher, R., Xiong, C.: Woad: Weakly supervised online action detection in untrimmed videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1915-1923 (2021) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.461 + ], + "angle": 0, + "content": "25. Goodrich, M.A., Schultz, A.C., et al.: Human-robot interaction: a survey. Foundations and Trends in Human-Computer Interaction 1(3), 203-275 (2008) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.502 + ], + "angle": 0, + "content": "26. Guo, H., Aved, A., Roller, C., Ardiles-Cruz, E., Ji, Q.: Video-based complex human event recognition with a probabilistic transformer. In: Geospatial Informatics XIII. vol. 12525, pp. 184-192. SPIE (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.556 + ], + "angle": 0, + "content": "27. Guo, H., Ren, Z., Wu, Y., Hua, G., Ji, Q.: Uncertainty-based spatial-temporal attention for online action detection. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 69-86. Springer (2022) 3, 7, 11, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.597 + ], + "angle": 0, + "content": "28. Han, Y., Tan, S.: Twinlstm: Two-channel LSTM network for online action detection. In: 2022 26th International Conference on Pattern Recognition (ICPR). pp. 3310-3317. IEEE (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.624 + ], + "angle": 0, + "content": "29. Hedegaard, L., Bakhtiarnia, A., Iosifidis, A.: Continual transformers: Redundancy-free attention for online inference. arXiv preprint arXiv:2201.06268 (2022) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.665 + ], + "angle": 0, + "content": "30. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: 2015 IEEE conference on computer vision and pattern recognition (CVPR). pp. 961-970. IEEE (2015) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.705 + ], + "angle": 0, + "content": "31. Idrees, H., Zamir, A.R., Jiang, Y.G., Gorban, A., Laptev, I., Sukthankar, R., Shah, M.: The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding 155, 1-23 (2017) 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.746 + ], + "angle": 0, + "content": "32. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.601, + 0.759 + ], + "angle": 0, + "content": "33. Jøsang, A.: Subjective logic, vol. 3. Springer (2016) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.813 + ], + "angle": 0, + "content": "34. Kim, Y.H., Kang, H., Kim, S.J.: A sliding window scheme for online temporal action localization. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV. pp. 653-669. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "35. Kim, Y.H., Nam, S., Kim, S.J.: Temporally smooth online action detection using cycle-consistent future anticipation. Pattern Recognition 116, 107954 (2021) 11" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEDL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.176 + ], + "angle": 0, + "content": "36. Kim, Y.H., Nam, S., Kim, S.J.: 2pesnet: Towards online processing of temporal action localization. Pattern Recognition 131, 108871 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.204 + ], + "angle": 0, + "content": "37. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.206, + 0.787, + 0.245 + ], + "angle": 0, + "content": "38. Kingma, D.P., Salimans, T., Welling, M.: Variational dropout and the local reparameterization trick. Advances in neural information processing systems 28 (2015) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.248, + 0.787, + 0.289 + ], + "angle": 0, + "content": "39. Kristiadi, A., Hein, M., Hennig, P.: Being bayesian, even just a bit, fixes overconfidence in relu networks. In: International conference on machine learning. pp. 5436-5446. PMLR (2020) 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.291, + 0.787, + 0.33 + ], + "angle": 0, + "content": "40. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2016), http://arxiv.org/abs/1612.014743" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.333, + 0.787, + 0.389 + ], + "angle": 0, + "content": "41. Li, Y., Lan, C., Xing, J., Zeng, W., Yuan, C., Liu, J.: Online human action detection using joint classification-regression recurrent neural networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14. pp. 203-220. Springer (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.39, + 0.787, + 0.431 + ], + "angle": 0, + "content": "42. Louizos, C., Welling, M.: Multiplicative normalizing flows for variational bayesian neural networks. In: International Conference on Machine Learning. pp. 2218-2227. PMLR (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.432, + 0.787, + 0.515 + ], + "angle": 0, + "content": "43. Maddox, W.J., Izmailov, P., Garipov, T., Vetrov, D.P., Wilson, A.G.: A simple baseline for bayesian uncertainty in deep learning. In: Wallach, H., Larochelle, H., Beygelzimer, A., e-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32, pp. 13132-13143. Curran Associates, Inc. (2019), http://papers.nips.cc/paper/9472-a-simple-baseline-for-bayesian-uncertainty-in-deep-learning.pdf 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.517, + 0.787, + 0.584 + ], + "angle": 0, + "content": "44. Malinin, A., Gales, M.: Predictive uncertainty estimation via prior networks. In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems 31, pp. 7047-7058. Curran Associates, Inc. (2018), http://papers.nips.cc/paper/7936-predictive-uncertainty-estimation-via-prior-networks.pdf 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.586, + 0.787, + 0.627 + ], + "angle": 0, + "content": "45. Malinin, A., Gales, M.: Reverse kl-divergence training of prior networks: Improved uncertainty and adversarial robustness. Advances in Neural Information Processing Systems 32 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.629, + 0.787, + 0.656 + ], + "angle": 0, + "content": "46. Malinin, A., Mlodozeniec, B., Gales, M.: Ensemble distribution distillation. arXiv preprint arXiv:1905.00076 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.658, + 0.787, + 0.698 + ], + "angle": 0, + "content": "47. Molchanov, D., Ashukha, A., Vetrov, D.: Variational dropout sparsifies deep neural networks. In: International Conference on Machine Learning. pp. 2498-2507. PMLR (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "48. Park, Y., Choi, W., Kim, S., Han, D.J., Moon, J.: Active learning for object detection with evidential deep learning and hierarchical uncertainty aggregation. In: The Eleventh International Conference on Learning Representations (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.743, + 0.787, + 0.782 + ], + "angle": 0, + "content": "49. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "50. Ramanishka, V., Chen, Y.T., Misu, T., Saenko, K.: Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7699-7707 (2018) 2, 10" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "51. Rangrej, S.B., Liang, K.J., Hassner, T., Clark, J.J.: Glitr: Glimpse transformers with spatiotemporal consistency for online action prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3413-3423 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.243 + ], + "angle": 0, + "content": "52. Sensoy, M., Kaplan, L., Kandemir, M.: Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems 31 (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.259 + ], + "angle": 0, + "content": "53. Sentz, K., Ferson, S.: Combination of evidence in dampster-shafer theory (2002) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.786, + 0.3 + ], + "angle": 0, + "content": "54. Shou, Z., Wang, D., Chang, S.F.: Temporal action localization in untrimmed videos via multi-stage cnns. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1049-1058 (2016) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.786, + 0.342 + ], + "angle": 0, + "content": "55. Sultani, W., Chen, C., Shah, M.: Real-world anomaly detection in surveillance videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6479-6488 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.37 + ], + "angle": 0, + "content": "56. Ulmer, D.: A survey on evidential deep learning for single-pass uncertainty estimation. arXiv preprint arXiv:2110.03051 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.412 + ], + "angle": 0, + "content": "57. Ulmer, D., Hardmeier, C., Frellsen, J.: Prior and posterior networks: A survey on evidential deep learning methods for uncertainty estimation. Transactions on Machine Learning Research (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.439 + ], + "angle": 0, + "content": "58. Valdenegro-Toro, M.: Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.786, + 0.478 + ], + "angle": 0, + "content": "59. Wang, H., Ji, Q.: Diversity-enhanced probabilistic ensemble for uncertainty estimation. In: Uncertainty in Artificial Intelligence. pp. 2214-2225. PMLR (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.786, + 0.535 + ], + "angle": 0, + "content": "60. Wang, J., Chen, G., Huang, Y., Wang, L., Lu, T.: Memory-and-anticipation transformer for online action understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13824-13835 (October 2023) 4, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.578 + ], + "angle": 0, + "content": "61. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.579, + 0.786, + 0.619 + ], + "angle": 0, + "content": "62. Wang, X., Zhang, S., Qing, Z., Shao, Y., Zuo, Z., Gao, C., Sang, N.: Oadtr: Online action detection with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7565-7575 (2021) 4, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.786, + 0.646 + ], + "angle": 0, + "content": "63. Welling, M., Teh, Y.W.: Bayesian learning via stochastic gradient Langevin dynamics pp. 681-688 (2011) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "64. Wen, Y., Tran, D., Ba, J.: Batchsemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.716 + ], + "angle": 0, + "content": "65. Xu, M., Gao, M., Chen, Y.T., Davis, L.S., Crandall, D.J.: Temporal recurrent networks for online action detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5532-5541 (2019) 4, 9, 10, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.786, + 0.758 + ], + "angle": 0, + "content": "66. Xu, M., Xiong, Y., Chen, H., Li, X., Xia, W., Tu, Z., Soatto, S.: Long short-term transformer for online action detection. Advances in Neural Information Processing Systems 34, 1086-1099 (2021) 4, 10, 11, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "67. Yang, L., Han, J., Zhang, D.: Colar: Effective and efficient online action detection by consulting exemplars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3160-3169 (2022) 4, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "68. Ye, N., Zhang, X., Yan, D., Dong, W., Yan, Q.: Scoad: Single-frame click supervision for online action detection. In: Proceedings of the Asian Conference on Computer Vision. pp. 2156-2171 (2022) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BEDL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "69. Zeng, R., Huang, W., Tan, M., Rong, Y., Zhao, P., Huang, J., Gan, C.: Graph convolutional networks for temporal action localization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7094-7103 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.218 + ], + "angle": 0, + "content": "70. Zhang, R., Li, C., Zhang, J., Chen, C., Wilson, A.G.: Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.259 + ], + "angle": 0, + "content": "71. Zhao, C., Du, D., Hoogs, A., Funk, C.: Open set action recognition via multi-label evidential learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22982-22991 (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.286 + ], + "angle": 0, + "content": "72. Zhao, P., Xie, L., Zhang, Y., Wang, Y., Tian, Q.: Privileged knowledge distillation for online action detection. arXiv preprint arXiv:2011.09158 (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.342 + ], + "angle": 0, + "content": "73. Zhao, Y., Krahenbuhl, P.: Real-time online video detection with temporal smoothing transformers. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV. pp. 485-502. Springer (2022) 4, 11, 14" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bayesian Evidential Deep Learning for Online Action Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_origin.pdf b/2024/Bayesian Evidential Deep Learning for Online Action Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e675f5e387fdb4abaf1afa76430364babf51794c --- /dev/null +++ b/2024/Bayesian Evidential Deep Learning for Online Action Detection/40969b27-7ac3-4e17-9156-d26dffe50c83_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c891a40d61e61067d5a52f907dab2180b0aeed4f11f49c8c9f903f5f7700939 +size 1747606 diff --git a/2024/Bayesian Evidential Deep Learning for Online Action Detection/full.md b/2024/Bayesian Evidential Deep Learning for Online Action Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e5362c827edeb3dc8c1b92306fb531147b040512 --- /dev/null +++ b/2024/Bayesian Evidential Deep Learning for Online Action Detection/full.md @@ -0,0 +1,355 @@ +# Bayesian Evidential Deep Learning for Online Action Detection + +Hongji Guo, Hanjing Wang, and Qiang Ji + +Rensselaer Polytechnic Institute, Troy NY 12180, USA + +{guoh11, wangh36, jiq}@rpi.edu + +Abstract. Online action detection aims at identifying the ongoing action in a streaming video without seeing the future. Timely and reliable response is critical for real-world applications. In this paper, we introduce Bayesian Evidential Deep Learning (BEDL), an efficient and generalizable framework for online action detection and uncertainty quantification. Specifically, we combine Bayesian neural networks and evidential deep learning by a teacher-student architecture. The teacher model is built in a Bayesian manner and transfers its mutual information and distribution to the student model through evidential deep learning. In this way, the student model can make accurate online inference while efficiently quantifying the uncertainty. Compared to existing evidential deep learning methods, BEDL estimates uncertainty more accurately by leveraging the Bayesian teacher model. In addition, we designed an attention module for active OAD, which actively selects important features based on the Bayesian mutual information instead of using all the features. We evaluated BEDL on benchmark datasets including THUMPS'14, TVSeries, and HDD. BEDL achieves competitive performance while keeping efficient inference. Extensive ablation studies demonstrate the effectiveness of each component. To verify the uncertainty quantification, we perform experiments of online anomaly detection with different types of uncertainties. + +Keywords: Online action detection $\cdot$ Evidential deep learning $\cdot$ Bayesian neural networks $\cdot$ Knowledge distillation $\cdot$ Uncertainty quantification + +# 1 Introduction + +Online action detection (OAD) [12] aims at identifying the ongoing action in a streaming video based on the historical observations. Different from the offline setting [54, 69], it only uses the past information and makes the prediction as soon as the action takes place. It has many important applications such as autonomous driving [9], visual surveillance [55], and human-robot interaction [25]. OAD is challenging due to the incomplete observations of actions and redundant information among inputs such as background and irrelevant actions. In addition, capturing the predictive uncertainty and generalizing to unseen environments are difficult, which are required by most safety-critical applications such as autonomous driving. + +To address these challenges, we introduce Bayesian Evidential Deep Learning (BEDL) for active online action detection and uncertainty quantification. Specifically, we combine Bayesian neural networks and evidential deep learning by a teacher-student architecture. Firstly, combining Bayesian neural networks and evidential deep learning enables accurate and efficient uncertainty quantification. Secondly, we introduce Bayesian mutual information (BMI) for OAD to actively select important features with minimal redundancy and irrelevance. + +Evidential deep learning (EDL) [1,52,57] is a technique developed for identifying the unknowns by quantifying uncertainty. Specifically, it uses deep neural networks to predict a Dirichlet distribution over the class probabilities, which is treated as an evidence collection process. Then the learned evidence are used to quantify the uncertainty by a single forward pass. EDL is computationally efficient and it has been widely used for open-set recognition [3,4,71] as it enables efficient uncertainty quantification without sampling. However, existing EDL methods cannot accurately model the epistemic and aleatoric uncertainty, which are critical as they provide different perspectives and information about the prediction. On the other hand, Bayesian neural networks (BNN) [20,38,47] can accurately estimate both epistemic and aleatoric uncertainty. But BNN confronts intractability of exact posterior inference and expensive sampling process, which prevents it from practical applications. + +Therefore, we combine EDL and BNN in this work to take advantages of their respective strengths, i.e. to efficiently and accurately estimate both epistemic and aleatoric uncertainty by a single forward pass. Specifically, we use BNN to augment EDL for uncertainty estimation by a teacher-student model. First, BNN model (teacher model) is built in a Bayesian manner to model the posterior distribution of model parameters. Then we transfer the knowledge of the teacher model to the EDL model (student model) by knowledge distillation (KD). Different from existing KD approaches that distill prediction scores or features, our BEDL distills both the uncertainties and mutual information of teacher model to student model. In this way, the BEDL model inherits the merits of Bayesian neural networks for accurately quantifying both epistemic and aleatoric uncertainty, while preserving its computational efficiency. As a result, the BEDL model can accurately and efficiently estimate both types of uncertainties by a single forward pass. The distilled mutual information as used as for feature selection as below. + +To address the background and irrelevant contents issue, we also designed an attention module for the BEDL model based on the distilled Bayesian mutual information (MI). We measure the dependency between the ongoing action and historical features by MI. Then the attention module is learned to generate an attention mask that can select informative features with high MI. In this way, the prediction benefits from better features and the online action detection can be improved. + +We evaluated BEDL's accuracy and efficiency on benchmark datasets including THUMOS'14 [31], TVSeries [12], and HDD [50]. We also demonstrated BEDL's other properties such as data-efficiency and generalization by extensive + +ablation studies. In addition, we validated the effectiveness of EDL by uncertainty quantification and online anomaly detection. + +In summary, the main contributions of this paper are as follows: + +- We introduce BEDL based on combining Bayesian neural newtorks and evidential deep learning, which enables efficient and accurate online inference and uncertainty quantification. +- We design a mutual information based attention module for the BEDL model to select informative features that are more discriminant and reliable. +- Our proposed BEDL achieves competitive performance on online action detection benchmark datasets in terms of accuracy and efficiency. And we validated its uncertainty quantification capability. + +# 2 Related Work + +Uncertainty quantification. Bayesian neural networks (BNNs) have traditionally been employed for quantifying uncertainty by considering neural network parameters as probabilistic variables and analyzing their posterior distributions. Techniques for uncertainty quantification (UQ) primarily involve Markov Chain Monte Carlo (MCMC) methods [11, 63, 70] and variational inference (VI) methods [19, 42, 43]. Ensemble approaches [26, 27, 40, 58, 59, 64] also serve as effective alternatives for precise UQ. While traditional BNNs offer several benefits, their need for a wide range of parameter samples for Bayesian inference results in inefficiency in UQ. This inefficiency is due to the requirement for several forward passes through the neural network, each utilizing uniquely sampled parameters from the posterior distribution. To address this issue, evidential deep learning approaches assume the parameters of the target distribution follow a conjugate prior distribution, enabling a closed-form expression for uncertainty that facilitates fast inference. Yet, these methods need additional knowledge to accurately learn the parameters of the prior distribution, including out-of-distribution (OOD) information [44, 45], ensemble models [46], and density estimations [7]. Even with these resources, point-estimation-based evidential networks may not effectively measure epistemic uncertainty, as indicated by [56]. Our strategy, therefore, involves utilizing BNNs to guide the training of evidential networks, aiming at enhancing the accuracy of evidential deep learning models while ensuring high efficiency for inference. + +Evidential deep learning (EDL). EDL [3, 4, 71] is proposed to address the challenge of making the model know the unknown when the input data is unfamiliar. In addition, it allows uncertainty estimation in a single model and forward pass [57]. Recently, EDL has been widely used for open-set recognition tasks. Bao et al. introduced DEAR [3] for open-set action recognition. By formulating action recognition in an EDL framework, the unknown actions can be recognized by computing the predictive uncertainty. Zhao et al. proposed multi-label evidential learning (MULE) [71] for open-set action recognition and novelty detection. The proposed evidence debiasing constraint enables the model to handle + +general problems of single or multiple actors in the same scene, with simultaneous action(s) by any actor. Later, EDL has been applied for open-set action localization [4,10,23]. By quantifying the frame-level uncertainty, the model can distinguish unknown actions from background video framesw. Park et al. proposed MEH-HUA [48] for object detection. By aggregating the uncertainty from EDL in a bottom-up order, the context within the image can be better captured for object detection. In this paper, we introduce EDL for online action detection and anomaly detection, which has not been explored yet. + +Online action detection. For the model architecture, RNN-based designs [8, 13, 15, 16, 21, 24, 28, 36, 41, 65, 68] are widely adopted because of RNN's temporal modeling capability. Typically, Xu et al. [65] proposed temporal recurrent network (TRN) that leverages both the historical information and predicted future features to detect the ongoing action. Thanks to the self-attention mechanism and the parallel computing property, Transformer-based methods [5, 29, 34, 51, 60, 62, 66, 67, 73] become the mainstream for online action detection. Wang [62] proposed OadTR that makes use of both historical information and future prediction. Xu [62] proposed long short-term Transformer (LSTR) that captures both the long-range and short-term dependencies by two memory units. To overcome the latency of feature extraction, [5] proposed E2E-LOAD for end-to-end online action detection. Besides RNN and Transformer, graph modeling is also studied for online action detection [14]. To leverage the video-level annotations instead of the dense frame-level annotations, weakly-supervised methods detection [24, 68] are also explored for OAD. + +# 3 Approach + +In this section, we first give an overview of the proposed Bayesian evidential deep learning (§ 3.1) and formulate the online action detection problem (§ 3.2). Then we introduce BEDL's Bayesian teacher model (§ 3.3) and evidential student model (§ 3.4). In the end, we show the training procedures (§ 3.5). + +# 3.1 Overview + +The overall framework of BEDL is shown in Figure 1. BEDL is built as a teacher-student framework, which is composed of a Bayesian neural network as the teacher model and an evidential neural network as the student model. Given a streaming video as input, a pretrained backbone first extracts features. Then the Bayesian tearcher model outputs the action predictions and mutual information based on the features, which are used to train the evidential student model by knowledge distillation. During the testing, only the student model is kept to perform fast inference and efficient uncertainty quantification. + +![](images/18d1e96f08f1fde4ae2292121e3305402a2ced7754ff4f1a8c59202a4813e5ae.jpg) +Fig. 1: Overall framework of Bayesian evidential deep learning (BEDL). The input of model is a streaming video. A pretrained backbone is used to extract the features for each frame. During the training, the Bayesian teacher model generates the mutual information (MI) and multiple sets of predictions. The evidential student model leverages MI to train the attention module for feature selection and predictions for uncertainty quantification. During testing, the student model actively selects informative features and quantifies predictive uncertainty accurately and efficiently by a single forward pass. + +# 3.2 Problem formulation + +Online action detection(OAD) aims at recognizing the ongoing action in a streaming video with only the past and current observations. Denote the input video as $\mathbf{V} = [I_1, I_2, \dots, I_T]$ , where $T$ is the length of video and $I_t$ denotes the frame at current time $t$ . The online action detection is formulated as a classification problem: $y_t^* = \operatorname{argmax}_c p(\hat{y}_t = c | \mathbf{V}_t)$ , where $\hat{y}_t$ is the prediction, $c$ is the class label, and $\mathbf{V}_t = \{I_1, \dots, I_t\}$ is the available frame set at time $t$ . A feature extractor is used to process each frame and generate the corresponding feature vector. Denote the feature set at time $t$ as $\mathbf{F}^t = \{F_1^t, \dots, F_t^t\}$ . The feature at time $i$ is $F_i^t \in \mathbb{R}^J$ , where $J$ is the feature dimension of each frame. + +# 3.3 Bayesian teacher model + +To take advantages of Bayesian neural networks for modeling both epistemic and aleatoric uncertainty, we build a Bayesian teacher model (BTM). BTM's objectives include: 1) generating mutual information to train the attention module for feature selection. 2) providing distribution knowledge for the student model for uncertainty quantification. + +The mutual information (MI) between past features and ongoing action indicates the relevance of features. Denote a past feature as $F_{ij}^{t}$ , where $i \in \{1, \dots, t\}$ is the time index, $j \in \{1, \dots, J\}$ is the feature index within each frame, and $t$ denotes the current time. We aim to obtain the mutual information between $F_{ij}^{t}$ and the ongoing action $y_{t}$ . An illustration is shown in Figure 2. To compute mutual information, we take advantage of Bayesian modeling. Different from + +point estimation, Bayesian method constructs a posterior distribution of model parameters. By integrating predictions from multiple models, it is less likely to be overfitting and the predictive uncertainty can be accurately quantified. Additionally, Bayesian method is more robust when training data is insufficient. We term the mutual information computed using the Bayesian method Bayesian Mutual Information (BMI). + +![](images/92938c5935a29a1b8dfbf2e51f3a1ddbe6a3039f4d958ae64bb1452c7e896e1d.jpg) +Fig. 2: Illustration of mutual information in OAD. + +![](images/e0cc5d72efccb59edd2751d2a23d041443ed306c1f065d4204ac43a16478112c.jpg) +Fig. 3: Illustration of distribution distillation by Bayesian evidential deep learning. + +Denote the model parameters of the teacher model as $\theta$ and we treat them as probability distributions. Then the BMI between a past feature $F_{ij}^{t}$ and the ongoing action $y_{t}$ can be written as: + +$$ +\mathcal {I} \left[ y _ {t}; F _ {i j} ^ {t} \mid \mathcal {D} \right] = \mathcal {H} \left[ y _ {t} \mid \boldsymbol {F} _ {- i j} ^ {t}, \mathcal {D} \right] - \mathcal {H} \left[ y _ {t} \mid \boldsymbol {F} ^ {t}, \mathcal {D} \right] \tag {1} +$$ + +where $\mathcal{D}$ denotes the training data, $\mathcal{H}$ denotes the entropy, and $F_{-ij}^{t}$ is the feature set at time $t$ excluding $F_{ij}^{t}$ , i.e. $F_{-ij}^{t} = F^{t} / F_{ij}^{t}$ . By definition, the entropy term is written as: + +$$ +\mathcal {H} \left[ y _ {t} \mid \boldsymbol {F} ^ {t}, \mathcal {D} \right] = - \sum_ {y _ {t} \in \mathcal {Y}} p \left(y _ {t} \mid \boldsymbol {F} ^ {t}, \mathcal {D}\right) \log p \left(y _ {t} \mid \boldsymbol {F} ^ {t}, \mathcal {D}\right) \tag {2} +$$ + +where $\mathcal{Y} = \{0,1,\dots,C\}$ is the action class set. 0 represents background class and $C$ is the number of action classes. The prediction distribution in Eq. (2) can be computed as: + +$$ +p (y _ {t} | \boldsymbol {F} ^ {t}, \mathcal {D}) = \int p (y _ {t} | \boldsymbol {F} ^ {t}, \theta) p (\theta | \mathcal {D}) d \theta \approx \frac {1}{K} p (y _ {t} | \boldsymbol {F} ^ {t}, \theta_ {k}), w h e r e \theta_ {k} \sim p (\theta | \mathcal {D}) (3) +$$ + +In Eq. (3), we use the sample average to approximate it since it is impractical to integrate over all the possible parameters. Similarly, we can compute $\mathcal{H}[y_t|F_{-ij}^t,\mathcal{D}]$ . + +To obtain the posterior distribution $p(\theta | \mathcal{D})$ , we perform the Laplace approximation (LA). To reduce the computation cost, we adopt a last-layer Bayesian setting [39]. Only the posterior distribution of last-layer parameters is modeled, while keeping the remaining parameters deterministic. Later, we will show this is efficient and effective. Specifically, we assume the last-layer parameters follow a Gaussian distribution. + +Firstly, we train the teacher model under the deterministic setting by a cross entropy loss: + +$$ +\begin{array}{l} \theta^ {*} = \underset {\theta} {\operatorname {a r g m i n}} \mathcal {L} (\mathcal {D}; \theta) = \underset {\theta} {\operatorname {a r g m i n}} \left(\sum_ {n = 1} ^ {N} l \left(x _ {n}, y _ {n}; \theta\right) + r (\theta)\right) \tag {4} \\ = \operatorname * {a r g m i n} _ {\theta} \left(\sum_ {n = 1} ^ {N} - \log p \left(y _ {n} \mid f _ {\theta} \left(x _ {n}\right)\right) - \log p (\theta)\right) \\ \end{array} +$$ + +where $\mathcal{D} = \{x_n, y_n\}_{n=1}^N$ is the training set and $r(\theta)$ is a regularizer such as weigh regularizer (a.k.a. weight decay). So $\theta^*$ is indeed a maximum a posteriori (MAP) estimate. + +The Laplace approximation uses a second-order Taylor expansion of $\mathcal{L}(\mathcal{D};\theta)$ around $\theta^{*}$ to construct a Gaussian approximation to $p(\theta |\mathcal{D})$ : + +$$ +\mathcal {L} (\mathcal {D}; \theta) \approx \mathcal {L} (\mathcal {D}; \theta^ {*}) + \frac {1}{2} \left(\theta - \theta^ {*}\right) ^ {T} \left(\nabla_ {\theta} ^ {2} \mathcal {L} (\mathcal {D}; \theta) | _ {\theta^ {*}}\right) \left(\theta - \theta^ {*}\right) \tag {5} +$$ + +where the first-order term vanishes at $\theta^{*}$ . So the Laplace posterior approximation can be obtained as: + +$$ +p (\theta | \mathcal {D}) \approx \mathcal {N} \left(\theta^ {*}, \Sigma\right), \text {w i t h} \Sigma := \left(\nabla_ {\theta} ^ {2} \mathcal {L} (\mathcal {D}; \theta) | _ {\theta^ {*}}\right) ^ {- 1} \tag {6} +$$ + +After the LA, we sample $K$ times from $p(\theta|\mathcal{D})$ to obtain $K$ sets of parameters $\{\theta_1,\dots,\theta_K\}$ . Then we approximate the prediction distribution in Eq. (3) and further compute the mutual information of each feature. The BMI we computed are based on multiple models from the teacher model. We use BMI to supervise the attention module of student model as the importance of past feature is measured by MI. Different from the traditional MI estimated from probabilistic models such as Uncertainty-OAD [27], BEDL uses BMI and hence is more robust and accurate. + +# 3.4 Evidential student model + +Evidential deep learning (EDL). Softmax-based deep neural networks are unable to accurately estimate the predictive uncertainty since the softmax probability is a point estimation of the predictive distribution and it tends to be over-confident in wrong predictions. To address such issue, EDL is proposed by combing Dempster-Shafer Theory [53] and subjective logic [33]. Specifically, it assumes the classification probability follows a prior Dirichlet distribution. The training is considered as a process of collecting evidence. The learned parameters of Dirichlet distribution are used as evidence to compute the predictive uncertainty. In this way, the uncertainty can be estimated in a single forward pass. In this work, we build the evidential student model under the EDL framework so that it can output both epistemic and aleatoric uncertainty in a forward pass. + +Distribution distillation. Common knowledge distillation (KD) methods aim at transferring prediction scores, features, or evidence [22], so the student model + +cannot estimate the predictive uncertainty. Differently, we transfer the uncertainty of the Bayesian teacher model to the evidential student model, i.e. distilling the distribution. In this way, the student model can learn from predictions of the teacher model as well as the capability of capturing the uncertainty. + +Under the classification setting, the target output of student model $y$ follows a categorical distribution with parameter $\lambda$ , $y \sim p(y|\lambda) = Cat(\lambda)$ . For example, $\lambda$ represents the probability after the final softmax layer. We treat $\lambda$ as a random variable and assume it follows Dirichlet distribution, i.e. $\lambda \sim p(\lambda|\alpha(x,\psi)) = Dir(\alpha(x,\theta))$ , where $\alpha$ denotes the parameters of the Dirichlet distribution and $\psi$ denotes student model parameters. Similarly, the teacher model has posterior distribution $p(\lambda|x,\mathcal{D}) = \int p(\lambda|x,\theta)p(\theta|\mathcal{D})d\theta$ . During the distillation, we transfer teacher posterior distribution $p(\lambda|x,\mathcal{D})$ to the student posterior distribution $p(\lambda|\alpha(x,\psi))$ . An illustration is shown in Figure 3. Specifically, we minimize the KL-divergence between these two distributions as below: + +$$ +\begin{array}{l} \mathcal {L} _ {d i s} = K L (p (\lambda | x, \mathcal {D}) | | p (\lambda | \alpha (x, \psi))) \\ \propto - \sum_ {c = 1} ^ {C} \log (\Gamma (\alpha_ {c})) + \log \Gamma (\sum_ {c = 1} ^ {C} \alpha_ {c}) - \mathbb {E} _ {p (\theta | \mathcal {D})} [ \sum_ {c = 1} ^ {C} (\alpha_ {c} - 1) \log \lambda_ {c} (x, \theta) ] \tag {7} \\ \end{array} +$$ + +where $C$ is the number of action classes. Detailed derivation of $\mathcal{L}_{dis}$ can be found in Appendix. Since the different predictions of the teacher model are combined, the model is more robust. The complete distribution distillation algorithm is available in the supplementary. + +Mutual information based attention module. The attention module actively selects informative features from the inputs by generating a spatial-temporal attention mask $A_{t}$ using a fully-connected network and applying to the original features by element-wise product. In this way, irrelevant features are masked out since they have low mutual information to the ongoing action. The training of the attention module is supervised by the BMI $\mathcal{I}_t$ from the teacher model. In this way, the BMI from the Bayesian teacher model is distilled to the student model. By minimizing the mean squared error (MSE) loss $\mathcal{L}_{att}$ between $A_{t}$ and $\mathcal{I}_t$ , the attention module can directly generate the BMI-aware attention mask without computing the BMI. $\mathcal{L}_{att}$ can be written as below: + +$$ +\mathcal {L} _ {a t t} = M S E \left(\boldsymbol {A} _ {t}, \sigma \left(\mathcal {I} _ {t}\right)\right) = \frac {1}{t J} \sum_ {i, j} \left| \left| A _ {i j} ^ {t} - \sigma \left(\mathcal {I} \left[ y _ {t}; F _ {i j} ^ {t} \mid \mathcal {D} \right]\right) \right| \right| ^ {2} \tag {8} +$$ + +where $\sigma (\cdot)$ is the sigmoid function. + +Uncertainty quantification. After the distillation through evidential deep learning, the Dirichlet distribution of the student model gains the knowledge of the Bayesian teacher model. The uncertainty can be computed efficiently in a closed-form solution by a single forward pass. Specifically, the total uncertainty + +and epistemic uncertainty can be quantified as: + +$$ +\mathcal {H} [ p (y | x, \theta) ] = \sum_ {c = 1} ^ {C} \frac {\alpha_ {c}}{\alpha_ {0}} \log \frac {\alpha_ {c}}{\alpha_ {0}}, \text {w h e r e} \alpha_ {0} = \sum_ {c = 1} ^ {C} \alpha_ {c} \tag {9} +$$ + +$$ +\mathcal {I} [ y; \lambda | \alpha ] = - \sum_ {c = 1} ^ {C} \frac {\alpha_ {c}}{\alpha_ {0}} \big (\ln \frac {\alpha_ {c}}{\alpha_ {0}} \varPsi (\alpha_ {c} + 1) + \varPsi (\alpha_ {0} + 1) \big) +$$ + +where $\varPsi(\cdot)$ is the dgamma function. To be simple, the aleatoric uncertainty can be computed by taking the difference between total uncertainty and epistemic uncertainty. The uncertainty can be also used to detect the anomaly. Details derivations are available in the supplementary. + +# 3.5 Training + +The evidential student model is jointly trained for online action detection, attention distillation, and distribution distillation. Specifically, we adopt a two-strategy. We first train the attention module using the attention loss $\mathcal{L}_{att}$ in Eq. 8. Then we jointly train the evidential student model by total loss $\mathcal{L}$ below: + +$$ +\mathcal {L} = \mathcal {L} _ {c e} + \lambda_ {1} \mathcal {L} _ {a t t} + \lambda_ {2} \mathcal {L} _ {d i s} \tag {10} +$$ + +where $\mathcal{L}_{ce}$ is the cross-entropy loss for online action detection. $\lambda_{1}$ and $\lambda_{2}$ are hyper-parameters that emphasize the attention module and distribution distillation. Although the distribution distillation can make the student model perform online action detection, we still perform joint training because the Bayesian teacher model is not perfect and its posterior distribution is approximated, which may introduce error to the student model. The joint training brings two benefits: 1) the online action detection can be improved with the supervision of ground-truth label; 2) the negative log-likelihood $\mathcal{L}_{ce}$ makes the training faster and more stable as indicated in [18]. We demonstrate the two-stage training works better in Sec. 4.6. + +# 4 Experiments + +# 4.1 Datasets and evaluation metrics + +THUMOS'14 [31] is a dataset for video-based temporal action localization. We use the validation set with 200 videos for training and test set with 213 videos for evaluation. There are 20 action classes and a background class. We ignore the frames with ambiguous labels. Following the settings in [21, 65], we adopt the mean Average Precision (mAP) as the evaluation metric. + +TVSeries [12] is a dataset collected from real TV series. It contains 27 episodes with 30 daily actions. It is a challenging dataset due to the viewpoints changing and occlusions in the videos. To counter the imbalanced data distribution, we adopt the mean calibrated Average Precision (mcAP) [12] as the + +evaluation metric. It is computed as $cAP = \sum_{k}cPrec(k)\times \mathbb{1}(k) / P$ , where $cPrec = TP / (TP + FP / \omega)$ , $P$ is the total number of positive frames and $\mathbb{1}(k)$ is an indicator function that is equal to 1 if frame $k$ is a true positive. The mcAP is the mean of calibrated average precision of all action classes. + +HRI Driving Dataset (HDD) [50] is a dataset for learning driver behavior in real-life environments. It contains 104 hours of real human driving in the San Francisco Bay Area collected by an instrumented vehicle with different sensors. There are 11 goal-oriented driving actions. Following the settings in [50], we use 100 sessions for training and 37 sessions for testing. And only the sensor data is used as the input. mAP is used as the evaluation metric for this dataset. + +Table 1: Online anomaly detection results on THUMOS'14 and TVSeries. + +
UncertaintyTHUMOS'14 (%)TVSeries (%)
EDL [3]64.7242.39
Total81.0465.16
Aleatoric62.5139.83
Epistemic86.4869.44
+ +![](images/d22c62f3510fa1f0ab7cdc8e34823a591c97e539a89f733259d974415cb4b09c.jpg) +Fig.4: ROC curves of online anomaly detection using epistemic uncertainty on THUMOS'14 (left) and TVSeries (right). + +![](images/8db17517566f77e7b997ae52726f6bcff7d6804d82eb914ebacf74a50822add9.jpg) + +# 4.2 Implementation Details + +Feature extraction. Following the settings in [15, 65, 66], we use TSN [61] to extract the features for THUMOS'14 and TVSeries. Video frames are extracted at 24 fps and the chunk size is set to 6. To better capture the spatial-temporal dependencies, we adopt the multi-scale vision Transformer [17] to extract RGB features. The optical flow features are extracted with BN-Inception [32]. The backbone is pretrained on ActivityNet [30] and Kinetics-400 [6] separately for evaluation. For HDD dataset, the sensor data is used as the input. More details can be found in the supplementary. + +Settings. The BEDL is implemented in PyTorch [49]. The model is trained by the Adam optimizer [37] with a learning rate of $10^{-4}$ and a weight decay of $5 \times 10^{-5}$ . The batch size is set to 32. The experiments were conducted on two Nvidia RTX 3090 Ti GPUs. The number of epochs is set to 20. Ablation studies on existing methods are based on their officially released codes. Detailed model architectures are available in the supplementary. And ccode will be made publicly available. + +# 4.3 Online action detection results + +We evaluate BEDL on three benchmark datasets and make a comparison with other methods in Table 2. From the results, our proposed BEDL achieves state-of-the-art performance on most benchmarks. On THUMOS'14, BEDL achieves $70.6\%$ and $72.7\%$ mAP using ActivityNet and Kinetics pretrained features. On TVSeries, BEDL achieves $88.9\%$ and $90.1\%$ mAP using ActivityNet and Kinetics features respectively. On HDD, we only use the sensor data as the input and + +achieve $33.0\%$ mAP, which is higher than the SOTA MAT [60]. The OAD performance demonstrate the effectiveness of BEDL for accurate online inference. In the following parts, we show other superior properties of BEDL while keeping accurate inference. + +Table 2: Experiment results on THUMOS'14, TVSeries and HDD. The results on THUMOS'14 and HDD are reported as mAP (\%). The results on TVSeries are reported as mCAP (\%). For HDD, $\star$ indicates RGB data is used as the input. + +
MethodArchitectureTHUMOS'14TVSeriesHDD
ANetKineticsANetKineticsSensor
RED [21]RNN45.3-79.2-27.4
FATS [35]-59.081.784.6-
TRN [65]47.262.183.786.229.2
IDN [15]50.060.384.786.1-
PKD [72]-64.5-86.4-
WOAD [24]-67.1---
OadTR [62]Transformer58.365.285.487.229.8
CoOadTR [29]56.164.287.687.730.6
Colar [67]59.466.986.088.130.6
LSTR [66]65.369.588.189.1-
Uncertainty-OAD [27]66.069.988.389.330.1
TeSTra [73]68.271.2---
GateHUB [8]69.170.788.489.632.1
MiniROAD [2]69.371.888.589.6-
MAT [60]70.471.688.689.732.7
E2E-LOAD [5]-72.4-90.348.1*
BEDL (ours)70.672.788.990.133.0
+ +# 4.4 Uncertainty quantification and online anomaly detection + +To verify the uncertainty quantification of BEDL, we quantified different types of uncertainties and perform online anomaly detection on THUMOS'14 and TVSeries. Specially, we divide the data into known classes and unknown classes (anomaly). Unknown classes do not appear in the training set. During testing, both known and unknown data will appear and the model is required to identify if the input belong to known or unknown classes. For THUMOS'14, we treat the "ambiguous" class that are difficult to identify during labeling process as anomaly. For TVSeries, we randomly select 5 out of 30 classes as the anomaly and repeat the process for ten times. Inputs that lead to high predictive uncertainty above the pre-defined threshold are declared as anomalies. + +The experiment results are shown in Table 1 and ROC curves are plotted in Figure 4. From the results, BEDL can accurately identify the unknown classes during the testing, which validates the effectiveness of its uncertainty quantification. Since epistemic uncertainty represents the lack of knowledge of the model and it is inversely proportional to the data density, it has better anomaly detec + +tion accuracy compared to aleatoric uncertainty, which represents the noise level of the data. Detailed experiment procedures can be found in the supplementary. + +# 4.5 Computational efficiency and model complexity + +We leverage evidential deep learning for BEDL in order to make it more practical for real-time applications. A comparison of computation efficiency and model complexity is shown in Table 3. While keeping a high detection accuracy, our proposed BEDL has much less model complexity and computational cost. And the inference speed of the model is much higher compared to other methods. Recently, [5] proposed an end-to-end OAD framework to avoid the latency from the frame and feature extraction parts, which leads to higher overall inference speed. To make fair comparisons, we used the same feature extraction procedures as prior works and we mainly compare the model inference speed since it only depends on the model architecture and pipeline. + +Table 3: Comparison of computation efficiency and model complexity. Our proposed BEDL has less model complexity and computational cost. And the inference speed is much faster than other methods. + +
MethodModalityModelInference Speed (FPS)
Parameter CountGFLOPsOptical Flow ComputationRGB Feature ExtractionFlow Feature ExtractionModel
TRN402.9M1.468.170.514.6123.3
OadTR75.8M2.548.170.514.6110.0
LSTRRGB + Flow58.0M7.538.170.514.691.6
GateHUB45.2M6.988.170.514.671.2
MAT94.6M-8.170.514.672.6
BEDL (ours)19.4M0.488.170.514.6163.2
+ +# 4.6 Ablation studies + +Training strategies. To demonstrate the effectiveness of two-stage training strategy, we compared it with two other training strategies. One is jointly training the attention module and EDL model by the total loss in Eq. 10. The other is firstly training the attention module by the attention loss and then we freezing the weights of attention module to train the EDL model by the $\mathcal{L}_{att}$ and $\mathcal{L}_{ce}$ . We refer the first one as joint training and the second one as separate training. The performance comparison is shown in Table 4. The results show that the two-stage training outperforms the other two strategies on both THUMOS'14 and TVSeries with different features, which demonstrates its effectiveness. + +Number of historical frames. At time $t$ , BEDL takes a certain number of past frames to predict the ongoing action. To study the long-range and short-term modeling capability of BEDL, we vary the number of past frames as the input. The experiment results are shown in Table 5. THUMOS'14 and TVSeries + +
TrainingTHUMOS'14TVSeries
ActivityNetKineticsActivityNetKinetics
Joint69.570.785.488.2
Separate64.967.881.684.0
Two-stage70.672.788.990.1
+ +Table 4: Ablation study of train- Table 5: Ablation study of different number ing strategies. The two-stage training of historical frames. gives the best performance. + +
DatasetFeatureNumber of past frames
8163264128
THUMOS'14ANet41.659.268.770.666.5
Kinetics49.361.671.472.770.0
TVSeriesANet63.576.787.688.984.9
Kinetics68.279.388.590.188.2
HDDSensor30.431.833.032.130.2
+ +![](images/a1e0d246b7a52ae2864a2570bd2140b1dc14dd20888656d300313a8da421d687.jpg) +(a) THUMOS'14-ANet + +![](images/f04f3bb779ca52acda0d62d70a42a05395dc275c9f8011716f60d4b53f738cee.jpg) +Fig. 5: Experiment results of training with small-scale data. We reduce the training data from $100\%$ to $10\%$ and compared with LSTR [66], MAT [60], and E2E-LOAD [5]. The results are plotted for THUMOS'14 and TVSeries with both ActivityNet and Kinetics pretrained features. Our proposed BEDL outperforms all others when training data is limited. + +![](images/7c87091e9d976c2932f37508ac7104febef7bdcd6abf9e75fb0cd528d16236eb.jpg) +(b) THUMOS'14-K400 +(c) TVSeries-ANet + +![](images/8f15ea1d3a424f9ea7cfd709fd551d3c3a421fa027fde272f47ac97b1c132998.jpg) +(d) TVSeries-K400 + +are extracted at 6 fps. HDD is extracted at 3 fps. So the optimal number of frames on HDD is 32. + +Data-efficiency. By applying the mutual information based attention mechanism, we expect BEDL to be more data-efficient when the amount of training data is limited. We reduce the training data from $100\%$ to $10\%$ and compare with other methods. The results on THUMOS'14 and TVSeries are plotted in Figure 5. When training data is reduced, BEDL has less performance decay and outperforms other methods, which demonstrates that BEDL is more data-efficient. + +Loss function. In the second training stage of BEDL, the total loss in Eq. 10 is composed of three terms. $\lambda_{1}$ and $\lambda_{2}$ are the weights for attention loss and distribution distillation loss. To study the balance of all three terms, we train the model with different $\lambda$ s on THUMOS'14 and TVSeries. The comparisons are plotted in Figure 6. We empirically choose $\lambda_{1} = 0.4$ and $\lambda_{2} = 6$ since they give the best performance. + +Generalization. To test the generalization capability of the model, we perform the Cross-View and under occlusion experiments on TVSeries dataset. Following the settings in [27], the training set and test set are from different view angles. For the occlusion, the training data does not contain occlusion and the testing data is occluded. The experiment results and comparison are shown in Table 6. The results show that BEDL generalizes better under different conditions. + +![](images/96b91907d4a432f34c898ed4b6e463277eaf0d0446288aa404cdae8c4f228ac4.jpg) +Fig.6: Ablation study of loss function for $\lambda_{1}$ and $\lambda_{2}$ on THUMOS'14 with ActivityNet and Kinetics features. + +![](images/d45048dd6147d77466578e377d6eed62b429cbd80a0e6f2f7759f290e41729fb.jpg) + +Table 6: Cross-view and occlusion experiments results on TVSeries. + +
MethodCV (%)Occ. (%)
TRN [65]65.885.2
OadTR [62]66.287.7
U-OAD [27]67.389.5
Colar [67]66.788.3
LSTR [66]69.589.4
TeSTra [73]70.289.9
BEDL (ours)74.391.8
+ +Attention module. To verify the effectiveness Bayesian mutual information, we first compare with the model without BMI. Specifically, we trained the student model with the same evidential neural network without the attention module. The total loss function is $\mathcal{L} = \mathcal{L}_{ce} + \lambda_2\mathcal{L}_{dis}$ and we grid-searched $\lambda_{2}$ to obtain the best accuracy. In addition, we compute the mutual information based on a single prediction instead of BMI and trained the model by the same loss function in Eq. 10. The results are shown in Table 7. From the results, the BEDL with BMI outperforms other two methods, which demonstrates the effectiveness of attention mechanism using BMI. + +Last-layer Bayesian. For the teacher model, we adopt the last-layer Bayesian [39] method to reduce the difficulty of training. We also performed the Laplace approximation over all the model parameters. The comparison is shown in Table 8. The full-Bayesian method has consistent improvement on THUMOS-14 and TVSeries. But the LA in training takes much longer time than the last-layer Bayesian and the model needs careful tuning, so we adopt the last-layer Bayesian in the LA process. + +Table 7: Effectiveness of BMI. BEDL-No-BMI denotes the model without attention module and BEDL-MI denotes the model with non-Bayesian MI. + +
ModelTHUMOS'14TVSeries
ActivityNetKineticsActivityNetKinetics
BEDL-No-BMI63.164.886.587.4
BEDL-MI67.668.488.088.5
BEDL70.672.788.990.1
+ +Table 8: Comparison of full-Bayesian and last-layer Bayesian. Full-Bayesian improves the performance since it models the distribution of all model parameters. + +
MethodTHUMOS-14 (mAP %)TVSeries (mcAP %)
ActivityNetKineticsActivityNetKinetics
Last-layer70.672.788.990.1
Full-Bayesian71.173.090.290.3
+ +# 5 Conclusion + +Conclusions. In this paper, we introduce Bayesian evidential deep learning for active online action detection and uncertainty quantification. By combining Bayesian neural networks and evidential deep learning, BEDL can accurately and efficiently quantify predictive uncertainty and make online inference. + +# References + +1. Amini, A., Schwarting, W., Soleimany, A., Rus, D.: Deep evidential regression. Advances in Neural Information Processing Systems 33, 14927-14937 (2020) 2 +2. An, J., Kang, H., Han, S.H., Yang, M.H., Kim, S.J.: Miniroad: Minimal rnn framework for online action detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10341-10350 (2023) 11 +3. Bao, W., Yu, Q., Kong, Y.: Evidential deep learning for open set action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13349-13358 (2021) 2, 3, 10 +4. Bao, W., Yu, Q., Kong, Y.: Opental: Towards open set temporal action localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2979-2989 (2022) 2, 3, 4 +5. Cao, S., Luo, W., Wang, B., Zhang, W., Ma, L.: E2e-load: End-to-end long-form online action detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10422-10432 (October 2023) 4, 11, 12, 13 +6. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) 10 +7. Charpentier, B., Zügner, D., Gunnemann, S.: Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in Neural Information Processing Systems 33, 1356-1367 (2020) 3 +8. Chen, J., Mittal, G., Yu, Y., Kong, Y., Chen, M.: Gatehub: Gated history unit with background suppression for online action detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19925-19934 (2022) 4, 11 +9. Chen, L., Ma, N., Wang, P., Li, J., Wang, P., Pang, G., Shi, X.: Survey of pedestrian action recognition techniques for autonomous driving. Tsinghua Science and Technology 25(4), 458-470 (2020) 1 +0. Chen, M., Gao, J., Yang, S., Xu, C.: Dual-evidential learning for weakly-supervised temporal action localization. In: European Conference on Computer Vision. pp. 192-208. Springer (2022) 4 +1. Chen, T., Fox, E., Guestrin, C.: Stochastic gradient hamiltonian monte carlo. In: International conference on machine learning. pp. 1683-1691. PMLR (2014) 3 +2. De Geest, R., Gavves, E., Ghodrati, A., Li, Z., Snoek, C., Tuytelaars, T.: Online action detection. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 269-284. Springer (2016) 1, 2, 9 +3. De Geest, R., Tuytelaars, T.: Modeling temporal structure with lstm for online action detection. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1549-1557. IEEE (2018) 4 +4. Elahi, G.M.E., Yang, Y.H.: Online temporal classification of human action using action inference graph. Pattern Recognition 132, 108972 (2022) 4 +5. Eun, H., Moon, J., Park, J., Jung, C., Kim, C.: Learning to discriminate information for online action detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 809-818 (2020) 4, 10, 11 +6. Eun, H., Moon, J., Park, J., Jung, C., Kim, C.: Temporal filtering networks for online action detection. Pattern Recognition 111, 107695 (2021) 4 +7. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 10 + +18. Fathullah, Y., Xia, G., Gales, M.J.: Logit-based ensemble distribution distillation for robust autoregressive sequence uncertainties. In: Uncertainty in Artificial Intelligence. pp. 582-591. PMLR (2023) 9 +19. Franchi, G., Bursuc, A., Aldea, E., Dubuisson, S., Bloch, I.: Encoding the latent posterior of bayesian neural networks for uncertainty quantification. arXiv preprint arXiv:2012.02818 (2020) 3 +20. Gal, Y., Ghahramani, Z.: Bayesian convolutional neural networks with bernoulli approximate variational inference. arXiv preprint arXiv:1506.02158 (2015) 2 +21. Gao, J., Yang, Z., Nevatia, R.: Red: Reinforced encoder-decoder networks for action anticipation. arXiv preprint arXiv:1707.04818 (2017) 4, 9, 11 +22. Gao, J., Chen, M., Xu, C.: Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18827-18836 (2023) 7 +23. Gao, J., Chen, M., Xu, C.: Vectorized evidential learning for weakly-supervised temporal action localization. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 4 +24. Gao, M., Zhou, Y., Xu, R., Socher, R., Xiong, C.: Woad: Weakly supervised online action detection in untrimmed videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1915-1923 (2021) 4, 11 +25. Goodrich, M.A., Schultz, A.C., et al.: Human-robot interaction: a survey. Foundations and Trends in Human-Computer Interaction 1(3), 203-275 (2008) 1 +26. Guo, H., Aved, A., Roller, C., Ardiles-Cruz, E., Ji, Q.: Video-based complex human event recognition with a probabilistic transformer. In: Geospatial Informatics XIII. vol. 12525, pp. 184-192. SPIE (2023) 3 +27. Guo, H., Ren, Z., Wu, Y., Hua, G., Ji, Q.: Uncertainty-based spatial-temporal attention for online action detection. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 69-86. Springer (2022) 3, 7, 11, 13, 14 +28. Han, Y., Tan, S.: Twinlstm: Two-channel LSTM network for online action detection. In: 2022 26th International Conference on Pattern Recognition (ICPR). pp. 3310-3317. IEEE (2022) 4 +29. Hedegaard, L., Bakhtiarnia, A., Iosifidis, A.: Continual transformers: Redundancy-free attention for online inference. arXiv preprint arXiv:2201.06268 (2022) 4, 11 +30. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: 2015 IEEE conference on computer vision and pattern recognition (CVPR). pp. 961-970. IEEE (2015) 10 +31. Idrees, H., Zamir, A.R., Jiang, Y.G., Gorban, A., Laptev, I., Sukthankar, R., Shah, M.: The thumos challenge on action recognition for videos "in the wild". Computer Vision and Image Understanding 155, 1-23 (2017) 2, 9 +32. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015) 10 +33. Jøsang, A.: Subjective logic, vol. 3. Springer (2016) 7 +34. Kim, Y.H., Kang, H., Kim, S.J.: A sliding window scheme for online temporal action localization. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV. pp. 653-669. Springer (2022) 4 +35. Kim, Y.H., Nam, S., Kim, S.J.: Temporally smooth online action detection using cycle-consistent future anticipation. Pattern Recognition 116, 107954 (2021) 11 + +36. Kim, Y.H., Nam, S., Kim, S.J.: 2pesnet: Towards online processing of temporal action localization. Pattern Recognition 131, 108871 (2022) 4 +37. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10 +38. Kingma, D.P., Salimans, T., Welling, M.: Variational dropout and the local reparameterization trick. Advances in neural information processing systems 28 (2015) 2 +39. Kristiadi, A., Hein, M., Hennig, P.: Being bayesian, even just a bit, fixes overconfidence in relu networks. In: International conference on machine learning. pp. 5436-5446. PMLR (2020) 6, 14 +40. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2016), http://arxiv.org/abs/1612.014743 +41. Li, Y., Lan, C., Xing, J., Zeng, W., Yuan, C., Liu, J.: Online human action detection using joint classification-regression recurrent neural networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14. pp. 203-220. Springer (2016) 4 +42. Louizos, C., Welling, M.: Multiplicative normalizing flows for variational bayesian neural networks. In: International Conference on Machine Learning. pp. 2218-2227. PMLR (2017) 3 +43. Maddox, W.J., Izmailov, P., Garipov, T., Vetrov, D.P., Wilson, A.G.: A simple baseline for bayesian uncertainty in deep learning. In: Wallach, H., Larochelle, H., Beygelzimer, A., e-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32, pp. 13132-13143. Curran Associates, Inc. (2019), http://papers.nips.cc/paper/9472-a-simple-baseline-for-bayesian-uncertainty-in-deep-learning.pdf 3 +44. Malinin, A., Gales, M.: Predictive uncertainty estimation via prior networks. In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems 31, pp. 7047-7058. Curran Associates, Inc. (2018), http://papers.nips.cc/paper/7936-predictive-uncertainty-estimation-via-prior-networks.pdf 3 +45. Malinin, A., Gales, M.: Reverse kl-divergence training of prior networks: Improved uncertainty and adversarial robustness. Advances in Neural Information Processing Systems 32 (2019) 3 +46. Malinin, A., Mlodozeniec, B., Gales, M.: Ensemble distribution distillation. arXiv preprint arXiv:1905.00076 (2019) 3 +47. Molchanov, D., Ashukha, A., Vetrov, D.: Variational dropout sparsifies deep neural networks. In: International Conference on Machine Learning. pp. 2498-2507. PMLR (2017) 2 +48. Park, Y., Choi, W., Kim, S., Han, D.J., Moon, J.: Active learning for object detection with evidential deep learning and hierarchical uncertainty aggregation. In: The Eleventh International Conference on Learning Representations (2022) 4 +49. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017) 10 +50. Ramanishka, V., Chen, Y.T., Misu, T., Saenko, K.: Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7699-7707 (2018) 2, 10 + +51. Rangrej, S.B., Liang, K.J., Hassner, T., Clark, J.J.: Glitr: Glimpse transformers with spatiotemporal consistency for online action prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3413-3423 (2023) 4 +52. Sensoy, M., Kaplan, L., Kandemir, M.: Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems 31 (2018) 2 +53. Sentz, K., Ferson, S.: Combination of evidence in dampster-shafer theory (2002) 7 +54. Shou, Z., Wang, D., Chang, S.F.: Temporal action localization in untrimmed videos via multi-stage cnns. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1049-1058 (2016) 1 +55. Sultani, W., Chen, C., Shah, M.: Real-world anomaly detection in surveillance videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6479-6488 (2018) 1 +56. Ulmer, D.: A survey on evidential deep learning for single-pass uncertainty estimation. arXiv preprint arXiv:2110.03051 (2021) 3 +57. Ulmer, D., Hardmeier, C., Frellsen, J.: Prior and posterior networks: A survey on evidential deep learning methods for uncertainty estimation. Transactions on Machine Learning Research (2023) 2, 3 +58. Valdenegro-Toro, M.: Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168 (2019) 3 +59. Wang, H., Ji, Q.: Diversity-enhanced probabilistic ensemble for uncertainty estimation. In: Uncertainty in Artificial Intelligence. pp. 2214-2225. PMLR (2023) 3 +60. Wang, J., Chen, G., Huang, Y., Wang, L., Lu, T.: Memory-and-anticipation transformer for online action understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13824-13835 (October 2023) 4, 11, 13 +61. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016) 10 +62. Wang, X., Zhang, S., Qing, Z., Shao, Y., Zuo, Z., Gao, C., Sang, N.: Oadtr: Online action detection with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7565-7575 (2021) 4, 11, 14 +63. Welling, M., Teh, Y.W.: Bayesian learning via stochastic gradient Langevin dynamics pp. 681-688 (2011) 3 +64. Wen, Y., Tran, D., Ba, J.: Batchsemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715 (2020) 3 +65. Xu, M., Gao, M., Chen, Y.T., Davis, L.S., Crandall, D.J.: Temporal recurrent networks for online action detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5532-5541 (2019) 4, 9, 10, 11, 14 +66. Xu, M., Xiong, Y., Chen, H., Li, X., Xia, W., Tu, Z., Soatto, S.: Long short-term transformer for online action detection. Advances in Neural Information Processing Systems 34, 1086-1099 (2021) 4, 10, 11, 13, 14 +67. Yang, L., Han, J., Zhang, D.: Colar: Effective and efficient online action detection by consulting exemplars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3160-3169 (2022) 4, 11, 14 +68. Ye, N., Zhang, X., Yan, D., Dong, W., Yan, Q.: Scoad: Single-frame click supervision for online action detection. In: Proceedings of the Asian Conference on Computer Vision. pp. 2156-2171 (2022) 4 + +69. Zeng, R., Huang, W., Tan, M., Rong, Y., Zhao, P., Huang, J., Gan, C.: Graph convolutional networks for temporal action localization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7094-7103 (2019) 1 +70. Zhang, R., Li, C., Zhang, J., Chen, C., Wilson, A.G.: Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932 (2019) 3 +71. Zhao, C., Du, D., Hoogs, A., Funk, C.: Open set action recognition via multi-label evidential learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22982-22991 (2023) 2, 3 +72. Zhao, P., Xie, L., Zhang, Y., Wang, Y., Tian, Q.: Privileged knowledge distillation for online action detection. arXiv preprint arXiv:2011.09158 (2020) 11 +73. Zhao, Y., Krahenbuhl, P.: Real-time online video detection with temporal smoothing transformers. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV. pp. 485-502. Springer (2022) 4, 11, 14 \ No newline at end of file diff --git a/2024/Bayesian Evidential Deep Learning for Online Action Detection/images.zip b/2024/Bayesian Evidential Deep Learning for Online Action Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..469625b42a564a3bf077d0bcce3d150e84c6aa7b --- /dev/null +++ b/2024/Bayesian Evidential Deep Learning for Online Action Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2339c0bb5827c8f23ee60e85233e13f1c4fbf8208edd5f57ad98cef0bd2c5000 +size 411837 diff --git a/2024/Bayesian Evidential Deep Learning for Online Action Detection/layout.json b/2024/Bayesian Evidential Deep Learning for Online Action Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9bcb0ea5945f1c65cddca98b08b501e367385914 --- /dev/null +++ b/2024/Bayesian Evidential Deep Learning for Online Action Detection/layout.json @@ -0,0 +1,10644 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 147, + 111, + 467, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 111, + 467, + 147 + ], + "spans": [ + { + "bbox": [ + 147, + 111, + 467, + 147 + ], + "type": "text", + "content": "Bayesian Evidential Deep Learning for Online Action Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 214, + 168, + 399, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 168, + 399, + 181 + ], + "spans": [ + { + "bbox": [ + 214, + 168, + 399, + 181 + ], + "type": "text", + "content": "Hongji Guo, Hanjing Wang, and Qiang Ji" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 190, + 418, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 190, + 418, + 201 + ], + "spans": [ + { + "bbox": [ + 194, + 190, + 418, + 201 + ], + "type": "text", + "content": "Rensselaer Polytechnic Institute, Troy NY 12180, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 236, + 202, + 378, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 202, + 378, + 213 + ], + "spans": [ + { + "bbox": [ + 236, + 202, + 378, + 213 + ], + "type": "text", + "content": "{guoh11, wangh36, jiq}@rpi.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 239, + 452, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 239, + 452, + 469 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 452, + 469 + ], + "type": "text", + "content": "Abstract. Online action detection aims at identifying the ongoing action in a streaming video without seeing the future. Timely and reliable response is critical for real-world applications. In this paper, we introduce Bayesian Evidential Deep Learning (BEDL), an efficient and generalizable framework for online action detection and uncertainty quantification. Specifically, we combine Bayesian neural networks and evidential deep learning by a teacher-student architecture. The teacher model is built in a Bayesian manner and transfers its mutual information and distribution to the student model through evidential deep learning. In this way, the student model can make accurate online inference while efficiently quantifying the uncertainty. Compared to existing evidential deep learning methods, BEDL estimates uncertainty more accurately by leveraging the Bayesian teacher model. In addition, we designed an attention module for active OAD, which actively selects important features based on the Bayesian mutual information instead of using all the features. We evaluated BEDL on benchmark datasets including THUMPS'14, TVSeries, and HDD. BEDL achieves competitive performance while keeping efficient inference. Extensive ablation studies demonstrate the effectiveness of each component. To verify the uncertainty quantification, we perform experiments of online anomaly detection with different types of uncertainties." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "spans": [ + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "text", + "content": "Keywords: Online action detection " + }, + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "text", + "content": " Evidential deep learning " + }, + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "text", + "content": " Bayesian neural networks " + }, + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "text", + "content": " Knowledge distillation " + }, + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 453, + 502 + ], + "type": "text", + "content": " Uncertainty quantification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 521, + 229, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 229, + 534 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 229, + 534 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": "Online action detection (OAD) [12] aims at identifying the ongoing action in a streaming video based on the historical observations. Different from the offline setting [54, 69], it only uses the past information and makes the prediction as soon as the action takes place. It has many important applications such as autonomous driving [9], visual surveillance [55], and human-robot interaction [25]. OAD is challenging due to the incomplete observations of actions and redundant information among inputs such as background and irrelevant actions. In addition, capturing the predictive uncertainty and generalizing to unseen environments are difficult, which are required by most safety-critical applications such as autonomous driving." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "To address these challenges, we introduce Bayesian Evidential Deep Learning (BEDL) for active online action detection and uncertainty quantification. Specifically, we combine Bayesian neural networks and evidential deep learning by a teacher-student architecture. Firstly, combining Bayesian neural networks and evidential deep learning enables accurate and efficient uncertainty quantification. Secondly, we introduce Bayesian mutual information (BMI) for OAD to actively select important features with minimal redundancy and irrelevance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 202, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 357 + ], + "type": "text", + "content": "Evidential deep learning (EDL) [1,52,57] is a technique developed for identifying the unknowns by quantifying uncertainty. Specifically, it uses deep neural networks to predict a Dirichlet distribution over the class probabilities, which is treated as an evidence collection process. Then the learned evidence are used to quantify the uncertainty by a single forward pass. EDL is computationally efficient and it has been widely used for open-set recognition [3,4,71] as it enables efficient uncertainty quantification without sampling. However, existing EDL methods cannot accurately model the epistemic and aleatoric uncertainty, which are critical as they provide different perspectives and information about the prediction. On the other hand, Bayesian neural networks (BNN) [20,38,47] can accurately estimate both epistemic and aleatoric uncertainty. But BNN confronts intractability of exact posterior inference and expensive sampling process, which prevents it from practical applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 539 + ], + "type": "text", + "content": "Therefore, we combine EDL and BNN in this work to take advantages of their respective strengths, i.e. to efficiently and accurately estimate both epistemic and aleatoric uncertainty by a single forward pass. Specifically, we use BNN to augment EDL for uncertainty estimation by a teacher-student model. First, BNN model (teacher model) is built in a Bayesian manner to model the posterior distribution of model parameters. Then we transfer the knowledge of the teacher model to the EDL model (student model) by knowledge distillation (KD). Different from existing KD approaches that distill prediction scores or features, our BEDL distills both the uncertainties and mutual information of teacher model to student model. In this way, the BEDL model inherits the merits of Bayesian neural networks for accurately quantifying both epistemic and aleatoric uncertainty, while preserving its computational efficiency. As a result, the BEDL model can accurately and efficiently estimate both types of uncertainties by a single forward pass. The distilled mutual information as used as for feature selection as below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 628 + ], + "type": "text", + "content": "To address the background and irrelevant contents issue, we also designed an attention module for the BEDL model based on the distilled Bayesian mutual information (MI). We measure the dependency between the ongoing action and historical features by MI. Then the attention module is learned to generate an attention mask that can select informative features with high MI. In this way, the prediction benefits from better features and the online action detection can be improved." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "We evaluated BEDL's accuracy and efficiency on benchmark datasets including THUMOS'14 [31], TVSeries [12], and HDD [50]. We also demonstrated BEDL's other properties such as data-efficiency and generalization by extensive" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "ablation studies. In addition, we validated the effectiveness of EDL by uncertainty quantification and online anomaly detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 140, + 430, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 140, + 430, + 152 + ], + "spans": [ + { + "bbox": [ + 146, + 140, + 430, + 152 + ], + "type": "text", + "content": "In summary, the main contributions of this paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 158, + 479, + 251 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 158, + 479, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 158, + 479, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 158, + 479, + 192 + ], + "type": "text", + "content": "- We introduce BEDL based on combining Bayesian neural newtorks and evidential deep learning, which enables efficient and accurate online inference and uncertainty quantification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 479, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 479, + 215 + ], + "type": "text", + "content": "- We design a mutual information based attention module for the BEDL model to select informative features that are more discriminant and reliable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 479, + 251 + ], + "type": "text", + "content": "- Our proposed BEDL achieves competitive performance on online action detection benchmark datasets in terms of accuracy and efficiency. And we validated its uncertainty quantification capability." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 270, + 237, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 237, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 237, + 283 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 294, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 547 + ], + "type": "text", + "content": "Uncertainty quantification. Bayesian neural networks (BNNs) have traditionally been employed for quantifying uncertainty by considering neural network parameters as probabilistic variables and analyzing their posterior distributions. Techniques for uncertainty quantification (UQ) primarily involve Markov Chain Monte Carlo (MCMC) methods [11, 63, 70] and variational inference (VI) methods [19, 42, 43]. Ensemble approaches [26, 27, 40, 58, 59, 64] also serve as effective alternatives for precise UQ. While traditional BNNs offer several benefits, their need for a wide range of parameter samples for Bayesian inference results in inefficiency in UQ. This inefficiency is due to the requirement for several forward passes through the neural network, each utilizing uniquely sampled parameters from the posterior distribution. To address this issue, evidential deep learning approaches assume the parameters of the target distribution follow a conjugate prior distribution, enabling a closed-form expression for uncertainty that facilitates fast inference. Yet, these methods need additional knowledge to accurately learn the parameters of the prior distribution, including out-of-distribution (OOD) information [44, 45], ensemble models [46], and density estimations [7]. Even with these resources, point-estimation-based evidential networks may not effectively measure epistemic uncertainty, as indicated by [56]. Our strategy, therefore, involves utilizing BNNs to guide the training of evidential networks, aiming at enhancing the accuracy of evidential deep learning models while ensuring high efficiency for inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Evidential deep learning (EDL). EDL [3, 4, 71] is proposed to address the challenge of making the model know the unknown when the input data is unfamiliar. In addition, it allows uncertainty estimation in a single model and forward pass [57]. Recently, EDL has been widely used for open-set recognition tasks. Bao et al. introduced DEAR [3] for open-set action recognition. By formulating action recognition in an EDL framework, the unknown actions can be recognized by computing the predictive uncertainty. Zhao et al. proposed multi-label evidential learning (MULE) [71] for open-set action recognition and novelty detection. The proposed evidence debiasing constraint enables the model to handle" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEDL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "general problems of single or multiple actors in the same scene, with simultaneous action(s) by any actor. Later, EDL has been applied for open-set action localization [4,10,23]. By quantifying the frame-level uncertainty, the model can distinguish unknown actions from background video framesw. Park et al. proposed MEH-HUA [48] for object detection. By aggregating the uncertainty from EDL in a bottom-up order, the context within the image can be better captured for object detection. In this paper, we introduce EDL for online action detection and anomaly detection, which has not been explored yet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 483, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 483, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 483, + 406 + ], + "type": "text", + "content": "Online action detection. For the model architecture, RNN-based designs [8, 13, 15, 16, 21, 24, 28, 36, 41, 65, 68] are widely adopted because of RNN's temporal modeling capability. Typically, Xu et al. [65] proposed temporal recurrent network (TRN) that leverages both the historical information and predicted future features to detect the ongoing action. Thanks to the self-attention mechanism and the parallel computing property, Transformer-based methods [5, 29, 34, 51, 60, 62, 66, 67, 73] become the mainstream for online action detection. Wang [62] proposed OadTR that makes use of both historical information and future prediction. Xu [62] proposed long short-term Transformer (LSTR) that captures both the long-range and short-term dependencies by two memory units. To overcome the latency of feature extraction, [5] proposed E2E-LOAD for end-to-end online action detection. Besides RNN and Transformer, graph modeling is also studied for online action detection [14]. To leverage the video-level annotations instead of the dense frame-level annotations, weakly-supervised methods detection [24, 68] are also explored for OAD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 441, + 212, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 212, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 212, + 455 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 521 + ], + "type": "text", + "content": "In this section, we first give an overview of the proposed Bayesian evidential deep learning (§ 3.1) and formulate the online action detection problem (§ 3.2). Then we introduce BEDL's Bayesian teacher model (§ 3.3) and evidential student model (§ 3.4). In the end, we show the training procedures (§ 3.5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 544, + 209, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 209, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 209, + 555 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The overall framework of BEDL is shown in Figure 1. BEDL is built as a teacher-student framework, which is composed of a Bayesian neural network as the teacher model and an evidential neural network as the student model. Given a streaming video as input, a pretrained backbone first extracts features. Then the Bayesian tearcher model outputs the action predictions and mutual information based on the features, which are used to train the evidential student model by knowledge distillation. During the testing, only the student model is kept to perform fast inference and efficient uncertainty quantification." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 477, + 237 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 477, + 237 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 477, + 237 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 477, + 237 + ], + "type": "image", + "image_path": "18d1e96f08f1fde4ae2292121e3305402a2ced7754ff4f1a8c59202a4813e5ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 346 + ], + "type": "text", + "content": "Fig. 1: Overall framework of Bayesian evidential deep learning (BEDL). The input of model is a streaming video. A pretrained backbone is used to extract the features for each frame. During the training, the Bayesian teacher model generates the mutual information (MI) and multiple sets of predictions. The evidential student model leverages MI to train the attention module for feature selection and predictions for uncertainty quantification. During testing, the student model actively selects informative features and quantifies predictive uncertainty accurately and efficiently by a single forward pass." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 369, + 265, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 265, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 265, + 380 + ], + "type": "text", + "content": "3.2 Problem formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": "Online action detection(OAD) aims at recognizing the ongoing action in a streaming video with only the past and current observations. Denote the input video as " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = [I_1, I_2, \\dots, I_T]" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " is the length of video and " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " denotes the frame at current time " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": ". The online action detection is formulated as a classification problem: " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "y_t^* = \\operatorname{argmax}_c p(\\hat{y}_t = c | \\mathbf{V}_t)" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " is the prediction, " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " is the class label, and " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_t = \\{I_1, \\dots, I_t\\}" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " is the available frame set at time " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": ". A feature extractor is used to process each frame and generate the corresponding feature vector. Denote the feature set at time " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^t = \\{F_1^t, \\dots, F_t^t\\}" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": ". The feature at time " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "F_i^t \\in \\mathbb{R}^J" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " is the feature dimension of each frame." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 279, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 279, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 279, + 526 + ], + "type": "text", + "content": "3.3 Bayesian teacher model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "To take advantages of Bayesian neural networks for modeling both epistemic and aleatoric uncertainty, we build a Bayesian teacher model (BTM). BTM's objectives include: 1) generating mutual information to train the attention module for feature selection. 2) providing distribution knowledge for the student model for uncertainty quantification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The mutual information (MI) between past features and ongoing action indicates the relevance of features. Denote a past feature as " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "F_{ij}^{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, t\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is the time index, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, J\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is the feature index within each frame, and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " denotes the current time. We aim to obtain the mutual information between " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "F_{ij}^{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and the ongoing action " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". An illustration is shown in Figure 2. To compute mutual information, we take advantage of Bayesian modeling. Different from" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEDL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "point estimation, Bayesian method constructs a posterior distribution of model parameters. By integrating predictions from multiple models, it is less likely to be overfitting and the predictive uncertainty can be accurately quantified. Additionally, Bayesian method is more robust when training data is insufficient. We term the mutual information computed using the Bayesian method Bayesian Mutual Information (BMI)." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 209, + 239, + 262 + ], + "blocks": [ + { + "bbox": [ + 133, + 209, + 239, + 262 + ], + "lines": [ + { + "bbox": [ + 133, + 209, + 239, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 209, + 239, + 262 + ], + "type": "image", + "image_path": "92938c5935a29a1b8dfbf2e51f3a1ddbe6a3039f4d958ae64bb1452c7e896e1d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 273, + 239, + 295 + ], + "lines": [ + { + "bbox": [ + 132, + 273, + 239, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 239, + 295 + ], + "type": "text", + "content": "Fig. 2: Illustration of mutual information in OAD." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 261, + 209, + 472, + 266 + ], + "blocks": [ + { + "bbox": [ + 261, + 209, + 472, + 266 + ], + "lines": [ + { + "bbox": [ + 261, + 209, + 472, + 266 + ], + "spans": [ + { + "bbox": [ + 261, + 209, + 472, + 266 + ], + "type": "image", + "image_path": "e0cc5d72efccb59edd2751d2a23d041443ed306c1f065d4204ac43a16478112c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 274, + 471, + 298 + ], + "lines": [ + { + "bbox": [ + 260, + 274, + 471, + 298 + ], + "spans": [ + { + "bbox": [ + 260, + 274, + 471, + 298 + ], + "type": "text", + "content": "Fig. 3: Illustration of distribution distillation by Bayesian evidential deep learning." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "content": "Denote the model parameters of the teacher model as " + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "content": " and we treat them as probability distributions. Then the BMI between a past feature " + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "inline_equation", + "content": "F_{ij}^{t}" + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "content": " and the ongoing action " + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 365, + 481, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 365, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 212, + 365, + 481, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {I} \\left[ y _ {t}; F _ {i j} ^ {t} \\mid \\mathcal {D} \\right] = \\mathcal {H} \\left[ y _ {t} \\mid \\boldsymbol {F} _ {- i j} ^ {t}, \\mathcal {D} \\right] - \\mathcal {H} \\left[ y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D} \\right] \\tag {1}", + "image_path": "ad7493443d722615163e590d7107d8c78cc1d9972db22aabfd00e56fa2cfdc23.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "text", + "content": " denotes the training data, " + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "text", + "content": " denotes the entropy, and " + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "inline_equation", + "content": "F_{-ij}^{t}" + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "text", + "content": " is the feature set at time " + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "text", + "content": " excluding " + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "inline_equation", + "content": "F_{ij}^{t}" + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "inline_equation", + "content": "F_{-ij}^{t} = F^{t} / F_{ij}^{t}" + }, + { + "bbox": [ + 130, + 389, + 482, + 427 + ], + "type": "text", + "content": ". By definition, the entropy term is written as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 201, + 434, + 480, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 434, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 201, + 434, + 480, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {H} \\left[ y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D} \\right] = - \\sum_ {y _ {t} \\in \\mathcal {Y}} p \\left(y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D}\\right) \\log p \\left(y _ {t} \\mid \\boldsymbol {F} ^ {t}, \\mathcal {D}\\right) \\tag {2}", + "image_path": "adbcf79d69233ee48f821a9e77f544cdbe6ab49ca71c3c85ba6faaaeba92b075.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 469, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 469, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{0,1,\\dots,C\\}" + }, + { + "bbox": [ + 130, + 469, + 482, + 506 + ], + "type": "text", + "content": " is the action class set. 0 represents background class and " + }, + { + "bbox": [ + 130, + 469, + 482, + 506 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 469, + 482, + 506 + ], + "type": "text", + "content": " is the number of action classes. The prediction distribution in Eq. (2) can be computed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 524, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 524, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 136, + 524, + 481, + 550 + ], + "type": "interline_equation", + "content": "p (y _ {t} | \\boldsymbol {F} ^ {t}, \\mathcal {D}) = \\int p (y _ {t} | \\boldsymbol {F} ^ {t}, \\theta) p (\\theta | \\mathcal {D}) d \\theta \\approx \\frac {1}{K} p (y _ {t} | \\boldsymbol {F} ^ {t}, \\theta_ {k}), w h e r e \\theta_ {k} \\sim p (\\theta | \\mathcal {D}) (3)", + "image_path": "68cd7ffcca90ce2647b3ddeee44325577254bdc7212dfc62976ba13f93d7b221.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "type": "text", + "content": "In Eq. (3), we use the sample average to approximate it since it is impractical to integrate over all the possible parameters. Similarly, we can compute " + }, + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{H}[y_t|F_{-ij}^t,\\mathcal{D}]" + }, + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "To obtain the posterior distribution " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "p(\\theta | \\mathcal{D})" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ", we perform the Laplace approximation (LA). To reduce the computation cost, we adopt a last-layer Bayesian setting [39]. Only the posterior distribution of last-layer parameters is modeled, while keeping the remaining parameters deterministic. Later, we will show this is efficient and effective. Specifically, we assume the last-layer parameters follow a Gaussian distribution." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "Firstly, we train the teacher model under the deterministic setting by a cross entropy loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 188, + 150, + 480, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 150, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 188, + 150, + 480, + 215 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta^ {*} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\mathcal {L} (\\mathcal {D}; \\theta) = \\underset {\\theta} {\\operatorname {a r g m i n}} \\left(\\sum_ {n = 1} ^ {N} l \\left(x _ {n}, y _ {n}; \\theta\\right) + r (\\theta)\\right) \\tag {4} \\\\ = \\operatorname * {a r g m i n} _ {\\theta} \\left(\\sum_ {n = 1} ^ {N} - \\log p \\left(y _ {n} \\mid f _ {\\theta} \\left(x _ {n}\\right)\\right) - \\log p (\\theta)\\right) \\\\ \\end{array}", + "image_path": "792ad08ec5dca98f60c2945f73faa3ab86a3971a0eefbf4f5f053bc77313f944.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 479, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 479, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 224, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{x_n, y_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 224, + 479, + 258 + ], + "type": "text", + "content": " is the training set and " + }, + { + "bbox": [ + 130, + 224, + 479, + 258 + ], + "type": "inline_equation", + "content": "r(\\theta)" + }, + { + "bbox": [ + 130, + 224, + 479, + 258 + ], + "type": "text", + "content": " is a regularizer such as weigh regularizer (a.k.a. weight decay). So " + }, + { + "bbox": [ + 130, + 224, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\theta^*" + }, + { + "bbox": [ + 130, + 224, + 479, + 258 + ], + "type": "text", + "content": " is indeed a maximum a posteriori (MAP) estimate." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 261, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 261, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 261, + 479, + 285 + ], + "type": "text", + "content": "The Laplace approximation uses a second-order Taylor expansion of " + }, + { + "bbox": [ + 131, + 261, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathcal{D};\\theta)" + }, + { + "bbox": [ + 131, + 261, + 479, + 285 + ], + "type": "text", + "content": " around " + }, + { + "bbox": [ + 131, + 261, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 131, + 261, + 479, + 285 + ], + "type": "text", + "content": " to construct a Gaussian approximation to " + }, + { + "bbox": [ + 131, + 261, + 479, + 285 + ], + "type": "inline_equation", + "content": "p(\\theta |\\mathcal{D})" + }, + { + "bbox": [ + 131, + 261, + 479, + 285 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 187, + 294, + 480, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 294, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 187, + 294, + 480, + 315 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathcal {D}; \\theta) \\approx \\mathcal {L} (\\mathcal {D}; \\theta^ {*}) + \\frac {1}{2} \\left(\\theta - \\theta^ {*}\\right) ^ {T} \\left(\\nabla_ {\\theta} ^ {2} \\mathcal {L} (\\mathcal {D}; \\theta) | _ {\\theta^ {*}}\\right) \\left(\\theta - \\theta^ {*}\\right) \\tag {5}", + "image_path": "1d303a212ed237fd796fbdb287fdcb773e737e9758fed8225a0673d1010c93ba.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 323, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 479, + 346 + ], + "type": "text", + "content": "where the first-order term vanishes at " + }, + { + "bbox": [ + 131, + 323, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 131, + 323, + 479, + 346 + ], + "type": "text", + "content": ". So the Laplace posterior approximation can be obtained as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 355, + 480, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 355, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 200, + 355, + 480, + 369 + ], + "type": "interline_equation", + "content": "p (\\theta | \\mathcal {D}) \\approx \\mathcal {N} \\left(\\theta^ {*}, \\Sigma\\right), \\text {w i t h} \\Sigma := \\left(\\nabla_ {\\theta} ^ {2} \\mathcal {L} (\\mathcal {D}; \\theta) | _ {\\theta^ {*}}\\right) ^ {- 1} \\tag {6}", + "image_path": "1c829b995e0e359a2260064aa61935069db2992f5d46d12855173f62f1636b13.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "text", + "content": "After the LA, we sample " + }, + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "text", + "content": " times from " + }, + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "inline_equation", + "content": "p(\\theta|\\mathcal{D})" + }, + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "text", + "content": " sets of parameters " + }, + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\{\\theta_1,\\dots,\\theta_K\\}" + }, + { + "bbox": [ + 131, + 377, + 480, + 472 + ], + "type": "text", + "content": ". Then we approximate the prediction distribution in Eq. (3) and further compute the mutual information of each feature. The BMI we computed are based on multiple models from the teacher model. We use BMI to supervise the attention module of student model as the importance of past feature is measured by MI. Different from the traditional MI estimated from probabilistic models such as Uncertainty-OAD [27], BEDL uses BMI and hence is more robust and accurate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 491, + 288, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 288, + 501 + ], + "type": "text", + "content": "3.4 Evidential student model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 479, + 641 + ], + "type": "text", + "content": "Evidential deep learning (EDL). Softmax-based deep neural networks are unable to accurately estimate the predictive uncertainty since the softmax probability is a point estimation of the predictive distribution and it tends to be over-confident in wrong predictions. To address such issue, EDL is proposed by combing Dempster-Shafer Theory [53] and subjective logic [33]. Specifically, it assumes the classification probability follows a prior Dirichlet distribution. The training is considered as a process of collecting evidence. The learned parameters of Dirichlet distribution are used as evidence to compute the predictive uncertainty. In this way, the uncertainty can be estimated in a single forward pass. In this work, we build the evidential student model under the EDL framework so that it can output both epistemic and aleatoric uncertainty in a forward pass." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": "Distribution distillation. Common knowledge distillation (KD) methods aim at transferring prediction scores, features, or evidence [22], so the student model" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEDL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "cannot estimate the predictive uncertainty. Differently, we transfer the uncertainty of the Bayesian teacher model to the evidential student model, i.e. distilling the distribution. In this way, the student model can learn from predictions of the teacher model as well as the capability of capturing the uncertainty." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": "Under the classification setting, the target output of student model " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": " follows a categorical distribution with parameter " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "y \\sim p(y|\\lambda) = Cat(\\lambda)" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": ". For example, " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": " represents the probability after the final softmax layer. We treat " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": " as a random variable and assume it follows Dirichlet distribution, i.e. " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "\\lambda \\sim p(\\lambda|\\alpha(x,\\psi)) = Dir(\\alpha(x,\\theta))" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": " denotes the parameters of the Dirichlet distribution and " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": " denotes student model parameters. Similarly, the teacher model has posterior distribution " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "p(\\lambda|x,\\mathcal{D}) = \\int p(\\lambda|x,\\theta)p(\\theta|\\mathcal{D})d\\theta" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": ". During the distillation, we transfer teacher posterior distribution " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "p(\\lambda|x,\\mathcal{D})" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": " to the student posterior distribution " + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "inline_equation", + "content": "p(\\lambda|\\alpha(x,\\psi))" + }, + { + "bbox": [ + 130, + 167, + 483, + 288 + ], + "type": "text", + "content": ". An illustration is shown in Figure 3. Specifically, we minimize the KL-divergence between these two distributions as below:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 304, + 481, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 304, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 481, + 353 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {d i s} = K L (p (\\lambda | x, \\mathcal {D}) | | p (\\lambda | \\alpha (x, \\psi))) \\\\ \\propto - \\sum_ {c = 1} ^ {C} \\log (\\Gamma (\\alpha_ {c})) + \\log \\Gamma (\\sum_ {c = 1} ^ {C} \\alpha_ {c}) - \\mathbb {E} _ {p (\\theta | \\mathcal {D})} [ \\sum_ {c = 1} ^ {C} (\\alpha_ {c} - 1) \\log \\lambda_ {c} (x, \\theta) ] \\tag {7} \\\\ \\end{array}", + "image_path": "3b046e256afb311ba87d14f23bc58d4191a12bc37780505987ad1b73904f8340.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "type": "text", + "content": " is the number of action classes. Detailed derivation of " + }, + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "type": "text", + "content": " can be found in Appendix. Since the different predictions of the teacher model are combined, the model is more robust. The complete distribution distillation algorithm is available in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "text", + "content": "Mutual information based attention module. The attention module actively selects informative features from the inputs by generating a spatial-temporal attention mask " + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "text", + "content": " using a fully-connected network and applying to the original features by element-wise product. In this way, irrelevant features are masked out since they have low mutual information to the ongoing action. The training of the attention module is supervised by the BMI " + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "text", + "content": " from the teacher model. In this way, the BMI from the Bayesian teacher model is distilled to the student model. By minimizing the mean squared error (MSE) loss " + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{att}" + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "text", + "content": ", the attention module can directly generate the BMI-aware attention mask without computing the BMI. " + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{att}" + }, + { + "bbox": [ + 130, + 419, + 485, + 540 + ], + "type": "text", + "content": " can be written as below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 183, + 556, + 481, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 556, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 183, + 556, + 481, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a t t} = M S E \\left(\\boldsymbol {A} _ {t}, \\sigma \\left(\\mathcal {I} _ {t}\\right)\\right) = \\frac {1}{t J} \\sum_ {i, j} \\left| \\left| A _ {i j} ^ {t} - \\sigma \\left(\\mathcal {I} \\left[ y _ {t}; F _ {i j} ^ {t} \\mid \\mathcal {D} \\right]\\right) \\right| \\right| ^ {2} \\tag {8}", + "image_path": "8aca7f60286139c0f093c745100fe5ebecbf66c7fefb7caa6eabd4aed9955367.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 601, + 286, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 286, + 614 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 286, + 614 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 601, + 286, + 614 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 131, + 601, + 286, + 614 + ], + "type": "text", + "content": " is the sigmoid function." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Uncertainty quantification. After the distillation through evidential deep learning, the Dirichlet distribution of the student model gains the knowledge of the Bayesian teacher model. The uncertainty can be computed efficiently in a closed-form solution by a single forward pass. Specifically, the total uncertainty" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 340, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 340, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 340, + 128 + ], + "type": "text", + "content": "and epistemic uncertainty can be quantified as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 207, + 137, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 137, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 207, + 137, + 481, + 175 + ], + "type": "interline_equation", + "content": "\\mathcal {H} [ p (y | x, \\theta) ] = \\sum_ {c = 1} ^ {C} \\frac {\\alpha_ {c}}{\\alpha_ {0}} \\log \\frac {\\alpha_ {c}}{\\alpha_ {0}}, \\text {w h e r e} \\alpha_ {0} = \\sum_ {c = 1} ^ {C} \\alpha_ {c} \\tag {9}", + "image_path": "fc5f6266e5245a8960f9fcf991b725e974f693f534493883beeb6fa2bf4c3a89.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 173, + 417, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 173, + 417, + 203 + ], + "spans": [ + { + "bbox": [ + 196, + 173, + 417, + 203 + ], + "type": "interline_equation", + "content": "\\mathcal {I} [ y; \\lambda | \\alpha ] = - \\sum_ {c = 1} ^ {C} \\frac {\\alpha_ {c}}{\\alpha_ {0}} \\big (\\ln \\frac {\\alpha_ {c}}{\\alpha_ {0}} \\varPsi (\\alpha_ {c} + 1) + \\varPsi (\\alpha_ {0} + 1) \\big)", + "image_path": "d8891089bcb18d7e578a4a04ccb14544fc4732bd20d752a28eb18b31a418e4f3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 212, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 212, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 482, + 261 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 212, + 482, + 261 + ], + "type": "inline_equation", + "content": "\\varPsi(\\cdot)" + }, + { + "bbox": [ + 131, + 212, + 482, + 261 + ], + "type": "text", + "content": " is the dgamma function. To be simple, the aleatoric uncertainty can be computed by taking the difference between total uncertainty and epistemic uncertainty. The uncertainty can be also used to detect the anomaly. Details derivations are available in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 277, + 204, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 204, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 204, + 290 + ], + "type": "text", + "content": "3.5 Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": "The evidential student model is jointly trained for online action detection, attention distillation, and distribution distillation. Specifically, we adopt a two-strategy. We first train the attention module using the attention loss " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{att}" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": " in Eq. 8. Then we jointly train the evidential student model by total loss " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": " below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 249, + 355, + 480, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 355, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 249, + 355, + 480, + 367 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c e} + \\lambda_ {1} \\mathcal {L} _ {a t t} + \\lambda_ {2} \\mathcal {L} _ {d i s} \\tag {10}", + "image_path": "75fbf7d19b6c9d3a8fe97f4fe97f69ac8b53eb539797f61eb01d4cb49e9547b2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "text", + "content": " is the cross-entropy loss for online action detection. " + }, + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "text", + "content": " are hyper-parameters that emphasize the attention module and distribution distillation. Although the distribution distillation can make the student model perform online action detection, we still perform joint training because the Bayesian teacher model is not perfect and its posterior distribution is approximated, which may introduce error to the student model. The joint training brings two benefits: 1) the online action detection can be improved with the supervision of ground-truth label; 2) the negative log-likelihood " + }, + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 131, + 375, + 482, + 495 + ], + "type": "text", + "content": " makes the training faster and more stable as indicated in [18]. We demonstrate the two-stage training works better in Sec. 4.6." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 538, + 322, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 322, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 322, + 550 + ], + "type": "text", + "content": "4.1 Datasets and evaluation metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 479, + 616 + ], + "type": "text", + "content": "THUMOS'14 [31] is a dataset for video-based temporal action localization. We use the validation set with 200 videos for training and test set with 213 videos for evaluation. There are 20 action classes and a background class. We ignore the frames with ambiguous labels. Following the settings in [21, 65], we adopt the mean Average Precision (mAP) as the evaluation metric." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "TVSeries [12] is a dataset collected from real TV series. It contains 27 episodes with 30 daily actions. It is a challenging dataset due to the viewpoints changing and occlusions in the videos. To counter the imbalanced data distribution, we adopt the mean calibrated Average Precision (mcAP) [12] as the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEDL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "evaluation metric. It is computed as " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "cAP = \\sum_{k}cPrec(k)\\times \\mathbb{1}(k) / P" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "cPrec = TP / (TP + FP / \\omega)" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " is the total number of positive frames and " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(k)" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " is an indicator function that is equal to 1 if frame " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " is a true positive. The mcAP is the mean of calibrated average precision of all action classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 236 + ], + "type": "text", + "content": "HRI Driving Dataset (HDD) [50] is a dataset for learning driver behavior in real-life environments. It contains 104 hours of real human driving in the San Francisco Bay Area collected by an instrumented vehicle with different sensors. There are 11 goal-oriented driving actions. Following the settings in [50], we use 100 sessions for training and 37 sessions for testing. And only the sensor data is used as the input. mAP is used as the evaluation metric for this dataset." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 148, + 278, + 287, + 323 + ], + "blocks": [ + { + "bbox": [ + 146, + 245, + 288, + 277 + ], + "lines": [ + { + "bbox": [ + 146, + 245, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 146, + 245, + 288, + 277 + ], + "type": "text", + "content": "Table 1: Online anomaly detection results on THUMOS'14 and TVSeries." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 278, + 287, + 323 + ], + "lines": [ + { + "bbox": [ + 148, + 278, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 148, + 278, + 287, + 323 + ], + "type": "table", + "html": "
UncertaintyTHUMOS'14 (%)TVSeries (%)
EDL [3]64.7242.39
Total81.0465.16
Aleatoric62.5139.83
Epistemic86.4869.44
", + "image_path": "0f6d65c26725b8507a24106b2f4a5e923868eff77f8e8f66f61f215b87a7ece6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 300, + 245, + 383, + 302 + ], + "blocks": [ + { + "bbox": [ + 300, + 245, + 383, + 302 + ], + "lines": [ + { + "bbox": [ + 300, + 245, + 383, + 302 + ], + "spans": [ + { + "bbox": [ + 300, + 245, + 383, + 302 + ], + "type": "image", + "image_path": "d22c62f3510fa1f0ab7cdc8e34823a591c97e539a89f733259d974415cb4b09c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 312, + 476, + 346 + ], + "lines": [ + { + "bbox": [ + 298, + 312, + 476, + 346 + ], + "spans": [ + { + "bbox": [ + 298, + 312, + 476, + 346 + ], + "type": "text", + "content": "Fig.4: ROC curves of online anomaly detection using epistemic uncertainty on THUMOS'14 (left) and TVSeries (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 245, + 472, + 302 + ], + "blocks": [ + { + "bbox": [ + 391, + 245, + 472, + 302 + ], + "lines": [ + { + "bbox": [ + 391, + 245, + 472, + 302 + ], + "spans": [ + { + "bbox": [ + 391, + 245, + 472, + 302 + ], + "type": "image", + "image_path": "8db17517566f77e7b997ae52726f6bcff7d6804d82eb914ebacf74a50822add9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 360, + 279, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 279, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 279, + 372 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 379, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 475 + ], + "type": "text", + "content": "Feature extraction. Following the settings in [15, 65, 66], we use TSN [61] to extract the features for THUMOS'14 and TVSeries. Video frames are extracted at 24 fps and the chunk size is set to 6. To better capture the spatial-temporal dependencies, we adopt the multi-scale vision Transformer [17] to extract RGB features. The optical flow features are extracted with BN-Inception [32]. The backbone is pretrained on ActivityNet [30] and Kinetics-400 [6] separately for evaluation. For HDD dataset, the sensor data is used as the input. More details can be found in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": "Settings. The BEDL is implemented in PyTorch [49]. The model is trained by the Adam optimizer [37] with a learning rate of " + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": ". The batch size is set to 32. The experiments were conducted on two Nvidia RTX 3090 Ti GPUs. The number of epochs is set to 20. Ablation studies on existing methods are based on their officially released codes. Detailed model architectures are available in the supplementary. And ccode will be made publicly available." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 575, + 315, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 315, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 315, + 586 + ], + "type": "text", + "content": "4.3 Online action detection results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We evaluate BEDL on three benchmark datasets and make a comparison with other methods in Table 2. From the results, our proposed BEDL achieves state-of-the-art performance on most benchmarks. On THUMOS'14, BEDL achieves " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "70.6\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "72.7\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " mAP using ActivityNet and Kinetics pretrained features. On TVSeries, BEDL achieves " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "88.9\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "90.1\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " mAP using ActivityNet and Kinetics features respectively. On HDD, we only use the sensor data as the input and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "achieve " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "33.0\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " mAP, which is higher than the SOTA MAT [60]. The OAD performance demonstrate the effectiveness of BEDL for accurate online inference. In the following parts, we show other superior properties of BEDL while keeping accurate inference." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 156, + 228, + 455, + 441 + ], + "blocks": [ + { + "bbox": [ + 130, + 184, + 482, + 218 + ], + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 218 + ], + "type": "text", + "content": "Table 2: Experiment results on THUMOS'14, TVSeries and HDD. The results on THUMOS'14 and HDD are reported as mAP (\\%). The results on TVSeries are reported as mCAP (\\%). For HDD, " + }, + { + "bbox": [ + 130, + 184, + 482, + 218 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 130, + 184, + 482, + 218 + ], + "type": "text", + "content": " indicates RGB data is used as the input." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 228, + 455, + 441 + ], + "lines": [ + { + "bbox": [ + 156, + 228, + 455, + 441 + ], + "spans": [ + { + "bbox": [ + 156, + 228, + 455, + 441 + ], + "type": "table", + "html": "
MethodArchitectureTHUMOS'14TVSeriesHDD
ANetKineticsANetKineticsSensor
RED [21]RNN45.3-79.2-27.4
FATS [35]-59.081.784.6-
TRN [65]47.262.183.786.229.2
IDN [15]50.060.384.786.1-
PKD [72]-64.5-86.4-
WOAD [24]-67.1---
OadTR [62]Transformer58.365.285.487.229.8
CoOadTR [29]56.164.287.687.730.6
Colar [67]59.466.986.088.130.6
LSTR [66]65.369.588.189.1-
Uncertainty-OAD [27]66.069.988.389.330.1
TeSTra [73]68.271.2---
GateHUB [8]69.170.788.489.632.1
MiniROAD [2]69.371.888.589.6-
MAT [60]70.471.688.689.732.7
E2E-LOAD [5]-72.4-90.348.1*
BEDL (ours)70.672.788.990.133.0
", + "image_path": "2387b23fc4233f5a7e95be8ff411e042856c7e956fded5f1920b2b17049a781e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 465, + 444, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 444, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 444, + 478 + ], + "type": "text", + "content": "4.4 Uncertainty quantification and online anomaly detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "content": "To verify the uncertainty quantification of BEDL, we quantified different types of uncertainties and perform online anomaly detection on THUMOS'14 and TVSeries. Specially, we divide the data into known classes and unknown classes (anomaly). Unknown classes do not appear in the training set. During testing, both known and unknown data will appear and the model is required to identify if the input belong to known or unknown classes. For THUMOS'14, we treat the \"ambiguous\" class that are difficult to identify during labeling process as anomaly. For TVSeries, we randomly select 5 out of 30 classes as the anomaly and repeat the process for ten times. Inputs that lead to high predictive uncertainty above the pre-defined threshold are declared as anomalies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "The experiment results are shown in Table 1 and ROC curves are plotted in Figure 4. From the results, BEDL can accurately identify the unknown classes during the testing, which validates the effectiveness of its uncertainty quantification. Since epistemic uncertainty represents the lack of knowledge of the model and it is inversely proportional to the data density, it has better anomaly detec" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEDL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "tion accuracy compared to aleatoric uncertainty, which represents the noise level of the data. Detailed experiment procedures can be found in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 160, + 400, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 160, + 400, + 173 + ], + "spans": [ + { + "bbox": [ + 131, + 160, + 400, + 173 + ], + "type": "text", + "content": "4.5 Computational efficiency and model complexity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 303 + ], + "type": "text", + "content": "We leverage evidential deep learning for BEDL in order to make it more practical for real-time applications. A comparison of computation efficiency and model complexity is shown in Table 3. While keeping a high detection accuracy, our proposed BEDL has much less model complexity and computational cost. And the inference speed of the model is much higher compared to other methods. Recently, [5] proposed an end-to-end OAD framework to avoid the latency from the frame and feature extraction parts, which leads to higher overall inference speed. To make fair comparisons, we used the same feature extraction procedures as prior works and we mainly compare the model inference speed since it only depends on the model architecture and pipeline." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 134, + 367, + 479, + 444 + ], + "blocks": [ + { + "bbox": [ + 130, + 324, + 481, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 324, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 481, + 357 + ], + "type": "text", + "content": "Table 3: Comparison of computation efficiency and model complexity. Our proposed BEDL has less model complexity and computational cost. And the inference speed is much faster than other methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 367, + 479, + 444 + ], + "lines": [ + { + "bbox": [ + 134, + 367, + 479, + 444 + ], + "spans": [ + { + "bbox": [ + 134, + 367, + 479, + 444 + ], + "type": "table", + "html": "
MethodModalityModelInference Speed (FPS)
Parameter CountGFLOPsOptical Flow ComputationRGB Feature ExtractionFlow Feature ExtractionModel
TRN402.9M1.468.170.514.6123.3
OadTR75.8M2.548.170.514.6110.0
LSTRRGB + Flow58.0M7.538.170.514.691.6
GateHUB45.2M6.988.170.514.671.2
MAT94.6M-8.170.514.672.6
BEDL (ours)19.4M0.488.170.514.6163.2
", + "image_path": "e896122a41c2243c4f936a0bec1676e9959e636ce9526aaf707f2aef0f77df77.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 487, + 243, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 243, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 243, + 498 + ], + "type": "text", + "content": "4.6 Ablation studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "text", + "content": "Training strategies. To demonstrate the effectiveness of two-stage training strategy, we compared it with two other training strategies. One is jointly training the attention module and EDL model by the total loss in Eq. 10. The other is firstly training the attention module by the attention loss and then we freezing the weights of attention module to train the EDL model by the " + }, + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{att}" + }, + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "text", + "content": ". We refer the first one as joint training and the second one as separate training. The performance comparison is shown in Table 4. The results show that the two-stage training outperforms the other two strategies on both THUMOS'14 and TVSeries with different features, which demonstrates its effectiveness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Number of historical frames. At time " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", BEDL takes a certain number of past frames to predict the ongoing action. To study the long-range and short-term modeling capability of BEDL, we vary the number of past frames as the input. The experiment results are shown in Table 5. THUMOS'14 and TVSeries" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 293, + 202 + ], + "blocks": [ + { + "bbox": [ + 132, + 157, + 293, + 202 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 293, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 293, + 202 + ], + "type": "table", + "html": "
TrainingTHUMOS'14TVSeries
ActivityNetKineticsActivityNetKinetics
Joint69.570.785.488.2
Separate64.967.881.684.0
Two-stage70.672.788.990.1
", + "image_path": "a432d8ff557991301f8124ed3216beb5e6320ca5b97bb5baf08752f72500d148.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 298, + 147, + 479, + 210 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Ablation study of train- Table 5: Ablation study of different number ing strategies. The two-stage training of historical frames. gives the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 147, + 479, + 210 + ], + "lines": [ + { + "bbox": [ + 298, + 147, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 298, + 147, + 479, + 210 + ], + "type": "table", + "html": "
DatasetFeatureNumber of past frames
8163264128
THUMOS'14ANet41.659.268.770.666.5
Kinetics49.361.671.472.770.0
TVSeriesANet63.576.787.688.984.9
Kinetics68.279.388.590.188.2
HDDSensor30.431.833.032.130.2
", + "image_path": "908f53cb7441ae3cc3963ffaf44d880cacc438db8817be61f149b59651d697d4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 230, + 219, + 289 + ], + "blocks": [ + { + "bbox": [ + 134, + 230, + 219, + 289 + ], + "lines": [ + { + "bbox": [ + 134, + 230, + 219, + 289 + ], + "spans": [ + { + "bbox": [ + 134, + 230, + 219, + 289 + ], + "type": "image", + "image_path": "a1e0d246b7a52ae2864a2570bd2140b1dc14dd20888656d300313a8da421d687.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 292, + 218, + 301 + ], + "lines": [ + { + "bbox": [ + 136, + 292, + 218, + 301 + ], + "spans": [ + { + "bbox": [ + 136, + 292, + 218, + 301 + ], + "type": "text", + "content": "(a) THUMOS'14-ANet" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 222, + 231, + 306, + 290 + ], + "blocks": [ + { + "bbox": [ + 222, + 231, + 306, + 290 + ], + "lines": [ + { + "bbox": [ + 222, + 231, + 306, + 290 + ], + "spans": [ + { + "bbox": [ + 222, + 231, + 306, + 290 + ], + "type": "image", + "image_path": "f04f3bb779ca52acda0d62d70a42a05395dc275c9f8011716f60d4b53f738cee.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 311, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 131, + 311, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 311, + 482, + 365 + ], + "type": "text", + "content": "Fig. 5: Experiment results of training with small-scale data. We reduce the training data from " + }, + { + "bbox": [ + 131, + 311, + 482, + 365 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 131, + 311, + 482, + 365 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 311, + 482, + 365 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 131, + 311, + 482, + 365 + ], + "type": "text", + "content": " and compared with LSTR [66], MAT [60], and E2E-LOAD [5]. The results are plotted for THUMOS'14 and TVSeries with both ActivityNet and Kinetics pretrained features. Our proposed BEDL outperforms all others when training data is limited." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 231, + 391, + 289 + ], + "blocks": [ + { + "bbox": [ + 223, + 292, + 306, + 301 + ], + "lines": [ + { + "bbox": [ + 223, + 292, + 306, + 301 + ], + "spans": [ + { + "bbox": [ + 223, + 292, + 306, + 301 + ], + "type": "text", + "content": "(b) THUMOS'14-K400" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 231, + 391, + 289 + ], + "lines": [ + { + "bbox": [ + 309, + 231, + 391, + 289 + ], + "spans": [ + { + "bbox": [ + 309, + 231, + 391, + 289 + ], + "type": "image", + "image_path": "7c87091e9d976c2932f37508ac7104febef7bdcd6abf9e75fb0cd528d16236eb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 292, + 384, + 301 + ], + "lines": [ + { + "bbox": [ + 317, + 292, + 384, + 301 + ], + "spans": [ + { + "bbox": [ + 317, + 292, + 384, + 301 + ], + "type": "text", + "content": "(c) TVSeries-ANet" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 394, + 231, + 477, + 289 + ], + "blocks": [ + { + "bbox": [ + 394, + 231, + 477, + 289 + ], + "lines": [ + { + "bbox": [ + 394, + 231, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 394, + 231, + 477, + 289 + ], + "type": "image", + "image_path": "8f15ea1d3a424f9ea7cfd709fd551d3c3a421fa027fde272f47ac97b1c132998.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 292, + 471, + 301 + ], + "lines": [ + { + "bbox": [ + 403, + 292, + 471, + 301 + ], + "spans": [ + { + "bbox": [ + 403, + 292, + 471, + 301 + ], + "type": "text", + "content": "(d) TVSeries-K400" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 399, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 422 + ], + "type": "text", + "content": "are extracted at 6 fps. HDD is extracted at 3 fps. So the optimal number of frames on HDD is 32." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 428, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 511 + ], + "type": "text", + "content": "Data-efficiency. By applying the mutual information based attention mechanism, we expect BEDL to be more data-efficient when the amount of training data is limited. We reduce the training data from " + }, + { + "bbox": [ + 130, + 428, + 482, + 511 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 428, + 482, + 511 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 428, + 482, + 511 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 428, + 482, + 511 + ], + "type": "text", + "content": " and compare with other methods. The results on THUMOS'14 and TVSeries are plotted in Figure 5. When training data is reduced, BEDL has less performance decay and outperforms other methods, which demonstrates that BEDL is more data-efficient." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "text", + "content": "Loss function. In the second training stage of BEDL, the total loss in Eq. 10 is composed of three terms. " + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "text", + "content": " are the weights for attention loss and distribution distillation loss. To study the balance of all three terms, we train the model with different " + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "text", + "content": "s on THUMOS'14 and TVSeries. The comparisons are plotted in Figure 6. We empirically choose " + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.4" + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 6" + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "text", + "content": " since they give the best performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Generalization. To test the generalization capability of the model, we perform the Cross-View and under occlusion experiments on TVSeries dataset. Following the settings in [27], the training set and test set are from different view angles. For the occlusion, the training data does not contain occlusion and the testing data is occluded. The experiment results and comparison are shown in Table 6. The results show that BEDL generalizes better under different conditions." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEDL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 124, + 241, + 189 + ], + "blocks": [ + { + "bbox": [ + 148, + 124, + 241, + 189 + ], + "lines": [ + { + "bbox": [ + 148, + 124, + 241, + 189 + ], + "spans": [ + { + "bbox": [ + 148, + 124, + 241, + 189 + ], + "type": "image", + "image_path": "96b91907d4a432f34c898ed4b6e463277eaf0d0446288aa404cdae8c4f228ac4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 199, + 323, + 232 + ], + "lines": [ + { + "bbox": [ + 146, + 199, + 323, + 232 + ], + "spans": [ + { + "bbox": [ + 146, + 199, + 323, + 232 + ], + "type": "text", + "content": "Fig.6: Ablation study of loss function for " + }, + { + "bbox": [ + 146, + 199, + 323, + 232 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 146, + 199, + 323, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 146, + 199, + 323, + 232 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 146, + 199, + 323, + 232 + ], + "type": "text", + "content": " on THUMOS'14 with ActivityNet and Kinetics features." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 240, + 124, + 332, + 189 + ], + "blocks": [ + { + "bbox": [ + 240, + 124, + 332, + 189 + ], + "lines": [ + { + "bbox": [ + 240, + 124, + 332, + 189 + ], + "spans": [ + { + "bbox": [ + 240, + 124, + 332, + 189 + ], + "type": "image", + "image_path": "d45048dd6147d77466578e377d6eed62b429cbd80a0e6f2f7759f290e41729fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 340, + 147, + 478, + 228 + ], + "blocks": [ + { + "bbox": [ + 338, + 125, + 480, + 146 + ], + "lines": [ + { + "bbox": [ + 338, + 125, + 480, + 146 + ], + "spans": [ + { + "bbox": [ + 338, + 125, + 480, + 146 + ], + "type": "text", + "content": "Table 6: Cross-view and occlusion experiments results on TVSeries." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 147, + 478, + 228 + ], + "lines": [ + { + "bbox": [ + 340, + 147, + 478, + 228 + ], + "spans": [ + { + "bbox": [ + 340, + 147, + 478, + 228 + ], + "type": "table", + "html": "
MethodCV (%)Occ. (%)
TRN [65]65.885.2
OadTR [62]66.287.7
U-OAD [27]67.389.5
Colar [67]66.788.3
LSTR [66]69.589.4
TeSTra [73]70.289.9
BEDL (ours)74.391.8
", + "image_path": "834909f7b94606d713c5fae445cd7a4027f69aef9621b6e48b31c5e4cd1e83c0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 232, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 339 + ], + "type": "text", + "content": "Attention module. To verify the effectiveness Bayesian mutual information, we first compare with the model without BMI. Specifically, we trained the student model with the same evidential neural network without the attention module. The total loss function is " + }, + { + "bbox": [ + 130, + 232, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{ce} + \\lambda_2\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 130, + 232, + 482, + 339 + ], + "type": "text", + "content": " and we grid-searched " + }, + { + "bbox": [ + 130, + 232, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 232, + 482, + 339 + ], + "type": "text", + "content": " to obtain the best accuracy. In addition, we compute the mutual information based on a single prediction instead of BMI and trained the model by the same loss function in Eq. 10. The results are shown in Table 7. From the results, the BEDL with BMI outperforms other two methods, which demonstrates the effectiveness of attention mechanism using BMI." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 340, + 487, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 487, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 487, + 425 + ], + "type": "text", + "content": "Last-layer Bayesian. For the teacher model, we adopt the last-layer Bayesian [39] method to reduce the difficulty of training. We also performed the Laplace approximation over all the model parameters. The comparison is shown in Table 8. The full-Bayesian method has consistent improvement on THUMOS-14 and TVSeries. But the LA in training takes much longer time than the last-layer Bayesian and the model needs careful tuning, so we adopt the last-layer Bayesian in the LA process." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 132, + 500, + 306, + 544 + ], + "blocks": [ + { + "bbox": [ + 131, + 445, + 309, + 491 + ], + "lines": [ + { + "bbox": [ + 131, + 445, + 309, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 445, + 309, + 491 + ], + "type": "text", + "content": "Table 7: Effectiveness of BMI. BEDL-No-BMI denotes the model without attention module and BEDL-MI denotes the model with non-Bayesian MI." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 500, + 306, + 544 + ], + "lines": [ + { + "bbox": [ + 132, + 500, + 306, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 306, + 544 + ], + "type": "table", + "html": "
ModelTHUMOS'14TVSeries
ActivityNetKineticsActivityNetKinetics
BEDL-No-BMI63.164.886.587.4
BEDL-MI67.668.488.088.5
BEDL70.672.788.990.1
", + "image_path": "6ed9851b3660edb4708cf41dd5c5d6aa7d094b9d6d94b12bfc204f0121d5c3e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 312, + 511, + 480, + 544 + ], + "blocks": [ + { + "bbox": [ + 312, + 445, + 482, + 501 + ], + "lines": [ + { + "bbox": [ + 312, + 445, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 482, + 501 + ], + "type": "text", + "content": "Table 8: Comparison of full-Bayesian and last-layer Bayesian. Full-Bayesian improves the performance since it models the distribution of all model parameters." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 511, + 480, + 544 + ], + "lines": [ + { + "bbox": [ + 312, + 511, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 511, + 480, + 544 + ], + "type": "table", + "html": "
MethodTHUMOS-14 (mAP %)TVSeries (mcAP %)
ActivityNetKineticsActivityNetKinetics
Last-layer70.672.788.990.1
Full-Bayesian71.173.090.290.3
", + "image_path": "f1545d51f234161e481fc143ac7cc25a02245264590e871caa52464b6afe5e79.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 589, + 220, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 220, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 220, + 601 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Conclusions. In this paper, we introduce Bayesian evidential deep learning for active online action detection and uncertainty quantification. By combining Bayesian neural networks and evidential deep learning, BEDL can accurately and efficiently quantify predictive uncertainty and make online inference." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "text", + "content": "1. Amini, A., Schwarting, W., Soleimany, A., Rus, D.: Deep evidential regression. Advances in Neural Information Processing Systems 33, 14927-14937 (2020) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 191 + ], + "type": "text", + "content": "2. An, J., Kang, H., Han, S.H., Yang, M.H., Kim, S.J.: Miniroad: Minimal rnn framework for online action detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10341-10350 (2023) 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 223 + ], + "type": "text", + "content": "3. Bao, W., Yu, Q., Kong, Y.: Evidential deep learning for open set action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13349-13358 (2021) 2, 3, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "text", + "content": "4. Bao, W., Yu, Q., Kong, Y.: Opental: Towards open set temporal action localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2979-2989 (2022) 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 289 + ], + "type": "text", + "content": "5. Cao, S., Luo, W., Wang, B., Zhang, W., Ma, L.: E2e-load: End-to-end long-form online action detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10422-10432 (October 2023) 4, 11, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "text", + "content": "6. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "text", + "content": "7. Charpentier, B., Zügner, D., Gunnemann, S.: Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in Neural Information Processing Systems 33, 1356-1367 (2020) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 353, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 396 + ], + "type": "text", + "content": "8. Chen, J., Mittal, G., Yu, Y., Kong, Y., Chen, M.: Gatehub: Gated history unit with background suppression for online action detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19925-19934 (2022) 4, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "text", + "content": "9. Chen, L., Ma, N., Wang, P., Li, J., Wang, P., Pang, G., Shi, X.: Survey of pedestrian action recognition techniques for autonomous driving. Tsinghua Science and Technology 25(4), 458-470 (2020) 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "text", + "content": "0. Chen, M., Gao, J., Yang, S., Xu, C.: Dual-evidential learning for weakly-supervised temporal action localization. In: European Conference on Computer Vision. pp. 192-208. Springer (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 480, + 482 + ], + "type": "text", + "content": "1. Chen, T., Fox, E., Guestrin, C.: Stochastic gradient hamiltonian monte carlo. In: International conference on machine learning. pp. 1683-1691. PMLR (2014) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 483, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 525 + ], + "type": "text", + "content": "2. De Geest, R., Gavves, E., Ghodrati, A., Li, Z., Snoek, C., Tuytelaars, T.: Online action detection. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 269-284. Springer (2016) 1, 2, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 557 + ], + "type": "text", + "content": "3. De Geest, R., Tuytelaars, T.: Modeling temporal structure with lstm for online action detection. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1549-1557. IEEE (2018) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 558, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 480, + 579 + ], + "type": "text", + "content": "4. Elahi, G.M.E., Yang, Y.H.: Online temporal classification of human action using action inference graph. Pattern Recognition 132, 108972 (2022) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "text", + "content": "5. Eun, H., Moon, J., Park, J., Jung, C., Kim, C.: Learning to discriminate information for online action detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 809-818 (2020) 4, 10, 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "6. Eun, H., Moon, J., Park, J., Jung, C., Kim, C.: Temporal filtering networks for online action detection. Pattern Recognition 111, 107695 (2021) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 10" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEDL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "18. Fathullah, Y., Xia, G., Gales, M.J.: Logit-based ensemble distribution distillation for robust autoregressive sequence uncertainties. In: Uncertainty in Artificial Intelligence. pp. 582-591. PMLR (2023) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "19. Franchi, G., Bursuc, A., Aldea, E., Dubuisson, S., Bloch, I.: Encoding the latent posterior of bayesian neural networks for uncertainty quantification. arXiv preprint arXiv:2012.02818 (2020) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "text", + "content": "20. Gal, Y., Ghahramani, Z.: Bayesian convolutional neural networks with bernoulli approximate variational inference. arXiv preprint arXiv:1506.02158 (2015) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "21. Gao, J., Yang, Z., Nevatia, R.: Red: Reinforced encoder-decoder networks for action anticipation. arXiv preprint arXiv:1707.04818 (2017) 4, 9, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 266 + ], + "type": "text", + "content": "22. Gao, J., Chen, M., Xu, C.: Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18827-18836 (2023) 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 266, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 480, + 300 + ], + "type": "text", + "content": "23. Gao, J., Chen, M., Xu, C.: Vectorized evidential learning for weakly-supervised temporal action localization. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "text", + "content": "24. Gao, M., Zhou, Y., Xu, R., Socher, R., Xiong, C.: Woad: Weakly supervised online action detection in untrimmed videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1915-1923 (2021) 4, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 365 + ], + "type": "text", + "content": "25. Goodrich, M.A., Schultz, A.C., et al.: Human-robot interaction: a survey. Foundations and Trends in Human-Computer Interaction 1(3), 203-275 (2008) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "text", + "content": "26. Guo, H., Aved, A., Roller, C., Ardiles-Cruz, E., Ji, Q.: Video-based complex human event recognition with a probabilistic transformer. In: Geospatial Informatics XIII. vol. 12525, pp. 184-192. SPIE (2023) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 440 + ], + "type": "text", + "content": "27. Guo, H., Ren, Z., Wu, Y., Hua, G., Ji, Q.: Uncertainty-based spatial-temporal attention for online action detection. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 69-86. Springer (2022) 3, 7, 11, 13, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 440, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 472 + ], + "type": "text", + "content": "28. Han, Y., Tan, S.: Twinlstm: Two-channel LSTM network for online action detection. In: 2022 26th International Conference on Pattern Recognition (ICPR). pp. 3310-3317. IEEE (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 494 + ], + "type": "text", + "content": "29. Hedegaard, L., Bakhtiarnia, A., Iosifidis, A.: Continual transformers: Redundancy-free attention for online inference. arXiv preprint arXiv:2201.06268 (2022) 4, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 494, + 480, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 526 + ], + "type": "text", + "content": "30. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: 2015 IEEE conference on computer vision and pattern recognition (CVPR). pp. 961-970. IEEE (2015) 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 525, + 480, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 558 + ], + "type": "text", + "content": "31. Idrees, H., Zamir, A.R., Jiang, Y.G., Gorban, A., Laptev, I., Sukthankar, R., Shah, M.: The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding 155, 1-23 (2017) 2, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "text", + "content": "32. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015) 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 590, + 367, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 367, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 367, + 601 + ], + "type": "text", + "content": "33. Jøsang, A.: Subjective logic, vol. 3. Springer (2016) 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 643 + ], + "type": "text", + "content": "34. Kim, Y.H., Kang, H., Kim, S.J.: A sliding window scheme for online temporal action localization. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV. pp. 653-669. Springer (2022) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "35. Kim, Y.H., Nam, S., Kim, S.J.: Temporally smooth online action detection using cycle-consistent future anticipation. Pattern Recognition 116, 107954 (2021) 11" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "text", + "content": "36. Kim, Y.H., Nam, S., Kim, S.J.: 2pesnet: Towards online processing of temporal action localization. Pattern Recognition 131, 108871 (2022) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 161 + ], + "type": "text", + "content": "37. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 163, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 481, + 194 + ], + "type": "text", + "content": "38. Kingma, D.P., Salimans, T., Welling, M.: Variational dropout and the local reparameterization trick. Advances in neural information processing systems 28 (2015) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "text", + "content": "39. Kristiadi, A., Hein, M., Hennig, P.: Being bayesian, even just a bit, fixes overconfidence in relu networks. In: International conference on machine learning. pp. 5436-5446. PMLR (2020) 6, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 230, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 481, + 261 + ], + "type": "text", + "content": "40. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2016), http://arxiv.org/abs/1612.014743" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 308 + ], + "type": "text", + "content": "41. Li, Y., Lan, C., Xing, J., Zeng, W., Yuan, C., Liu, J.: Online human action detection using joint classification-regression recurrent neural networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14. pp. 203-220. Springer (2016) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "42. Louizos, C., Welling, M.: Multiplicative normalizing flows for variational bayesian neural networks. In: International Conference on Machine Learning. pp. 2218-2227. PMLR (2017) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 407 + ], + "type": "text", + "content": "43. Maddox, W.J., Izmailov, P., Garipov, T., Vetrov, D.P., Wilson, A.G.: A simple baseline for bayesian uncertainty in deep learning. In: Wallach, H., Larochelle, H., Beygelzimer, A., e-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32, pp. 13132-13143. Curran Associates, Inc. (2019), http://papers.nips.cc/paper/9472-a-simple-baseline-for-bayesian-uncertainty-in-deep-learning.pdf 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 409, + 481, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 462 + ], + "type": "text", + "content": "44. Malinin, A., Gales, M.: Predictive uncertainty estimation via prior networks. In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems 31, pp. 7047-7058. Curran Associates, Inc. (2018), http://papers.nips.cc/paper/7936-predictive-uncertainty-estimation-via-prior-networks.pdf 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 496 + ], + "type": "text", + "content": "45. Malinin, A., Gales, M.: Reverse kl-divergence training of prior networks: Improved uncertainty and adversarial robustness. Advances in Neural Information Processing Systems 32 (2019) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 498, + 481, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 519 + ], + "type": "text", + "content": "46. Malinin, A., Mlodozeniec, B., Gales, M.: Ensemble distribution distillation. arXiv preprint arXiv:1905.00076 (2019) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 521, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 552 + ], + "type": "text", + "content": "47. Molchanov, D., Ashukha, A., Vetrov, D.: Variational dropout sparsifies deep neural networks. In: International Conference on Machine Learning. pp. 2498-2507. PMLR (2017) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "text", + "content": "48. Park, Y., Choi, W., Kim, S., Han, D.J., Moon, J.: Active learning for object detection with evidential deep learning and hierarchical uncertainty aggregation. In: The Eleventh International Conference on Learning Representations (2022) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 588, + 481, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 619 + ], + "type": "text", + "content": "49. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017) 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "50. Ramanishka, V., Chen, Y.T., Misu, T., Saenko, K.: Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7699-7707 (2018) 2, 10" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEDL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "51. Rangrej, S.B., Liang, K.J., Hassner, T., Clark, J.J.: Glitr: Glimpse transformers with spatiotemporal consistency for online action prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3413-3423 (2023) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 192 + ], + "type": "text", + "content": "52. Sensoy, M., Kaplan, L., Kandemir, M.: Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems 31 (2018) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 205 + ], + "type": "text", + "content": "53. Sentz, K., Ferson, S.: Combination of evidence in dampster-shafer theory (2002) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 481, + 237 + ], + "type": "text", + "content": "54. Shou, Z., Wang, D., Chang, S.F.: Temporal action localization in untrimmed videos via multi-stage cnns. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1049-1058 (2016) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "55. Sultani, W., Chen, C., Shah, M.: Real-world anomaly detection in surveillance videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6479-6488 (2018) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "56. Ulmer, D.: A survey on evidential deep learning for single-pass uncertainty estimation. arXiv preprint arXiv:2110.03051 (2021) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "57. Ulmer, D., Hardmeier, C., Frellsen, J.: Prior and posterior networks: A survey on evidential deep learning methods for uncertainty estimation. Transactions on Machine Learning Research (2023) 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "58. Valdenegro-Toro, M.: Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168 (2019) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 481, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 378 + ], + "type": "text", + "content": "59. Wang, H., Ji, Q.: Diversity-enhanced probabilistic ensemble for uncertainty estimation. In: Uncertainty in Artificial Intelligence. pp. 2214-2225. PMLR (2023) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "text", + "content": "60. Wang, J., Chen, G., Huang, Y., Wang, L., Lu, T.: Memory-and-anticipation transformer for online action understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13824-13835 (October 2023) 4, 11, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "text", + "content": "61. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 458, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 481, + 490 + ], + "type": "text", + "content": "62. Wang, X., Zhang, S., Qing, Z., Shao, Y., Zuo, Z., Gao, C., Sang, N.: Oadtr: Online action detection with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7565-7575 (2021) 4, 11, 14" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "text", + "content": "63. Welling, M., Teh, Y.W.: Bayesian learning via stochastic gradient Langevin dynamics pp. 681-688 (2011) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "64. Wen, Y., Tran, D., Ba, J.: Batchsemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715 (2020) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "65. Xu, M., Gao, M., Chen, Y.T., Davis, L.S., Crandall, D.J.: Temporal recurrent networks for online action detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5532-5541 (2019) 4, 9, 10, 11, 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "66. Xu, M., Xiong, Y., Chen, H., Li, X., Xia, W., Tu, Z., Soatto, S.: Long short-term transformer for online action detection. Advances in Neural Information Processing Systems 34, 1086-1099 (2021) 4, 10, 11, 13, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "text", + "content": "67. Yang, L., Han, J., Zhang, D.: Colar: Effective and efficient online action detection by consulting exemplars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3160-3169 (2022) 4, 11, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "68. Ye, N., Zhang, X., Yan, D., Dong, W., Yan, Q.: Scoad: Single-frame click supervision for online action detection. In: Proceedings of the Asian Conference on Computer Vision. pp. 2156-2171 (2022) 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "69. Zeng, R., Huang, W., Tan, M., Rong, Y., Zhao, P., Huang, J., Gan, C.: Graph convolutional networks for temporal action localization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7094-7103 (2019) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 172 + ], + "type": "text", + "content": "70. Zhang, R., Li, C., Zhang, J., Chen, C., Wilson, A.G.: Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932 (2019) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "type": "text", + "content": "71. Zhao, C., Du, D., Hoogs, A., Funk, C.: Open set action recognition via multi-label evidential learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22982-22991 (2023) 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "text", + "content": "72. Zhao, P., Xie, L., Zhang, Y., Wang, Y., Tian, Q.: Privileged knowledge distillation for online action detection. arXiv preprint arXiv:2011.09158 (2020) 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "text", + "content": "73. Zhao, Y., Krahenbuhl, P.: Real-time online video detection with temporal smoothing transformers. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV. pp. 485-502. Springer (2022) 4, 11, 14" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "BEDL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/546e4382-dd78-463e-8d89-435110642a17_content_list.json b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/546e4382-dd78-463e-8d89-435110642a17_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d9b3704e15965866d192825c721a6c5dc28353af --- /dev/null +++ b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/546e4382-dd78-463e-8d89-435110642a17_content_list.json @@ -0,0 +1,1719 @@ +[ + { + "type": "text", + "text": "Bayesian Self-Training for Semi-Supervised 3D Segmentation", + "text_level": 1, + "bbox": [ + 256, + 140, + 746, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ozan Unal $^{1,2}$ , Christos Sakaridis $^{1}$ , and Luc Van Gool $^{1,3,4}$", + "bbox": [ + 289, + 210, + 710, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ETH Zurich, $^{2}$ Huawei Technologies, $^{3}$ KU Leuven, $^{4}$ INSAIT {ozan.unal, csakarid, vangool}@vision.ee.ethz.ch", + "bbox": [ + 299, + 238, + 702, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D segmentation is a core problem in computer vision and, similarly to many other dense prediction tasks, it requires large amounts of annotated data for adequate training. However, densely labeling 3D point clouds to employ fully-supervised training remains too labor intensive and expensive. Semi-supervised training provides a more practical alternative, where only a small set of labeled data is given, accompanied by a larger unlabeled set. This area thus studies the effective use of unlabeled data to reduce the performance gap that arises due to the lack of annotations. In this work, inspired by Bayesian deep learning, we first propose a Bayesian self-training framework for semi-supervised 3D semantic segmentation. Employing stochastic inference, we generate an initial set of pseudo-labels and then filter these based on estimated point-wise uncertainty. By constructing a heuristic $n$ -partite matching algorithm, we extend the method to semi-supervised 3D instance segmentation, and finally, with the same building blocks, to dense 3D visual grounding. We demonstrate state-of-the-art results for our semi-supervised method on SemanticKITTI and ScribbleKITTI for 3D semantic segmentation and on ScanNet and S3DIS for 3D instance segmentation. We further achieve substantial improvements in dense 3D visual grounding over supervised-only baselines on ScanRefer. Our project page is available at ouenal.github.io/bst/.", + "bbox": [ + 261, + 301, + 738, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Semi-supervised $\\cdot$ Self-training $\\cdot$ 3D semantic segmentation $\\cdot$ 3D instance segmentation $\\cdot$ 3D visual grounding", + "bbox": [ + 261, + 604, + 738, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 657, + 374, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense 3D semantic perception at the level of points is pivotal for several applications that are based on fine-grained processing and reasoning in 3D, e.g. autonomous cars, robotics, and augmented reality. Such perception requires a segmentation of the unstructured input 3D point cloud into different semantic classes, i.e. 3D semantic segmentation [60], or distinct instances of these classes, i.e., 3D instance segmentation [30]. The input point cloud is often coupled with human utterances that refer to individual objects in the former in cases where a visual AI system interacts with human users in a common 3D environment, necessitating a 3D instance-level segmentation of these objects based on the aforementioned referrals, a task which is known as dense 3D visual grounding or", + "bbox": [ + 217, + 688, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80c238c525cec2778e79dc167ed82162c7a232dfce5551928bdea7297789d1d6.jpg", + "image_caption": [ + "Fig. 1: Illustration of our Bayesian pseudo-labeling pipeline for semi-supervised a) 3D semantic segmentation; b) 3D instance segmentation; and c) dense 3D visual grounding. With only slight adjustments, using the same building blocks, our method can be adapted to each of these tasks to achieve SOTA results." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 782, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "referral-based 3D instance segmentation [48, 59]. In the majority of works that address these diverse 3D segmentation tasks, it is assumed that the training data come with full 3D semantic and/or verbal annotations, which creates the pressing need for large-scale labeled 3D datasets [1, 2, 6, 11] so that large end-to-end models can be adequately fitted on these data.", + "bbox": [ + 212, + 378, + 782, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the creation of such fine-level 3D annotations and verbal referrals for 3D scans requires human labor to be fairly reliable. Such labor is time-consuming and expensive, especially due to the high dimensionality of the scanned point clouds, which makes it hard for human annotators to navigate through them, visually inspect them, and verbally describe them. These practical difficulties render learning frameworks for 3D segmentation that operate with partial labels essential, semi-supervised learning being a primary case. Answering this need, several recent works address semi-supervised 3D semantic segmentation [9,25,29], most of which focus on effective pseudo-labeling techniques for self-training on the larger, unlabeled portion of the training data. However, the generation of pseudo-labels in these works typically focuses on some grouping mechanism over 3D points, such as superpoints, to refine and propagate pseudo-labels and otherwise relies on naïve core techniques such as confidence thresholding [66] for generating the initial point-level pseudo-labels that are then post-processed by their proposed modules. Adding to this, there is a relative scarcity in works on semi-supervised 3D instance segmentation [10, 33], which we attribute to the relative difficulty in pseudo-labeling instances compared to creating plain semantic pseudo-labels. Last but not least, to our knowledge, no work has considered the setting of semi-supervised dense 3D visual grounding.", + "bbox": [ + 212, + 458, + 787, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We jointly address these shortcomings of previous works by proposing a common self-training framework for semi-supervised 3D segmentation which is based on Bayesian deep learning and applies to all three aforementioned segmentation tasks. At the core of our self-training framework lies a variational Bayesian approximation of the uncertainty of the respective network's predictions via dropout-based Monte Carlo integration [15] over multiple augmented versions of", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "O. Unal et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the original 3D input. This approximation delivers a more informed uncertainty estimate than simple confidence thresholding techniques [66], which is crucial for eliminating confirmation bias from the pseudo-labeling process that we adopt for self-training. In particular, we utilize both the estimated Shannon entropy from the sampled predictions and the predicted label from each sample to filter out uncertain or inconsistent points from the generated set of pseudo-labels.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While this formulation is sufficient for the task of 3D semantic segmentation, it does not work out of the box for instance-level segmentation, as Monte Carlo integration cannot be applied directly to instance predictions. To account for this issue, we propose a novel and efficient, heuristic matching strategy across instances coming from different stochastic dropout forward passes based on linear sum assignment. This strategy enables the application of the same filtering steps as described above for semantic pseudo-labels to instance pseudo-labels and thus extends the scope of our approach to 3D instance segmentation. Going one step further, we observe that the same matching strategy that we devise for instance segmentation applies to the selection stage of dense 3D grounding, in which the correct instance from the set of segmented candidate instances needs to be selected based on the input verbal description. Thus, for dense 3D grounding, we apply our heuristic linear sum assignment twice, first for the purely 3D-based candidate instance proposal generation and afterward for the verbo-visual instance selection. Our complete self-training pipeline for each of the three 3D segmentation tasks to which our method applies is summarized in Fig. 1.", + "bbox": [ + 212, + 243, + 787, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We test our method on leading benchmarks across a wide range of domains and datasets. Specifically, our Bayesian self-training framework achieves state-of-the-art results for semi-supervised 3D semantic segmentation on SemanticKITTI [3] and ScribbleKITTI [47] and semi-supervised 3D instance segmentation on ScanNet [11] and S3DIS [2]. Furthermore, we are the first to explore a semi-supervised setting on ScanRefer [6] for 3D visual grounding and obtain consistent significant improvements over supervised-only baselines.", + "bbox": [ + 212, + 491, + 787, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 214, + 604, + 542, + 618 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Following established uncertainty estimation techniques from active learning, we construct a Bayesian self-training framework for semi-supervised 3D semantic segmentation by generating reliable pseudo-labels through stochastic inference and entropy-based uncertainty filtering.", + "2. We introduce a novel heuristic matching algorithm to extend the proposed framework to 3D instance segmentation.", + "3. We further extend the method to dense 3D visual grounding utilizing the same building blocks as in the above tasks and are the first to explore this task in a semi-supervised setting.", + "4. We obtain state-of-the-art results for semi-supervised 3D segmentation on SemanticKITTI [3], ScribbleKITTI [47], ScanNet [11] and S3DIS [2]." + ], + "bbox": [ + 222, + 654, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bayesian Self-Training for Semi-Supervised 3D Segmentation", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semi-supervised 3D semantic segmentation has recently grown in popularity in the area of 3D understanding along with weakly supervised 3D semantic segmentation [47,52,57,61], due to the difficulty in large-scale 3D labeling that is necessary for the supervised setup [60]. In this work, we tackle the task of semi-supervised segmentation where a large portion of data is completely unlabeled. As opposed to methods developed for weakly-supervised settings, a semi-supervised method cannot leverage the availability of partial true label information within each data sample, e.g. via the ArcPoint loss [31] or a PLS descriptor [47]. A common technique for leveraging unlabeled data under this restrictive scheme is explored under pseudo-labeling. Semantic pseudo-labels are utilized by Jiang et al. [25] to guide a contrastive loss and in SSPC [9] to propagate labels on a superpoint graph. Superpoints are also leveraged by Deng et al. [13] to constrain pseudo-label propagation. In the 2D counterpart task, Chen et al. [8] enforces consistency of semantic pseudo-labels across perturbed versions of the 2D segmentation network. By contrast, we perturb both the network weights and the 3D inputs across different forward passes, and we consider both the entropy and the consistency among predictions in our pseudo-labels. Another related line of 2D semi-supervised works generates pseudo-labels based on direct thresholding of max-softmax scores to reduce confirmation bias [19, 66], while we threshold the entropy of the distribution of softmax scores. Active learning is combined with a 3D semi-supervised approach in DiAL [46] to streamline practical 3D segmentation annotation pipelines. LaserMix [29] introduces variability through cylindrical and range-view partitioning and mixing of 3D LiDAR scans, encouraging consistency across different mixes.", + "bbox": [ + 212, + 175, + 787, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semi-supervised 3D instance segmentation is still in its infancy compared to its instance-agnostic semantic segmentation counterpart reviewed above. In particular, most object-level semi-supervised 3D understanding works focus on the coarser task of 3D object detection [17,35,37,38,42,50,51,63], which outputs 3D bounding boxes that are not fine enough for precision-critical downstream tasks such as robotic grasping. For 3D instance/object segmentation, approaches that depart from the standard fully supervised setup have mostly considered the cases of unsupervised pre-training [56], weak supervision via one-point-one-instance [20, 43], or unsupervised generic 3D object segmentation [32]. To our knowledge, the only two semi-supervised works that specifically consider 3D instance segmentation are TWIST [10] and WS3D [33]. The former work proposes a pseudo-label generation strategy tailored for the instance segmentation task, for which the technique of confidence thresholding that is ubiquitously used to the same end in semantic segmentation is not applicable. It thus introduces pseudo offset vectors, besides semantic proposal-level pseudo-labels, for performing self-training. The latter work proposes a region-level contrastive loss, which performs unsupervised instance discrimination, and leverages boundary information via a region-level unsupervised energy-based loss. On the contrary, we propose a unified semi-supervised Bayesian self-training framework that spans multiple 3D segmentation tasks including 3D instance segmentation, and han", + "bbox": [ + 212, + 537, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "O. Unal et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dles the additional matching required between instances generated in different forward passes through the network via an efficient heuristic optimization step.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D visual grounding is a central 3D task in the area of vision and language [4, 14, 28, 34, 36, 62]. Relevant models accept a 3D point cloud as visual input and perform grounding of the accompanying verbal description to the referred object. While most previous works [7, 23, 24, 34, 39] focus on coarse 3D bounding box outputs, the task of dense 3D visual grounding or referral-based 3D instance segmentation has been explored far less in comparison, in which a fine-grained point-level mask is predicted [22, 48, 59]. Regardless, all aforementioned methods invariably assume a supervised setting, where full ground-truth instance annotations are available for all samples at training time. By contrast, the formulation of our method makes it fit for also handling the more practical semi-supervised setting, in which only a part of the training data is labeled.", + "bbox": [ + 212, + 181, + 787, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Uncertainty estimation is crucial in learning setups that lack full supervision as ours, given that such estimates can provide information about the reliability of the predictions of the model on unlabeled samples [21]. In the more recent deep learning era, Bayesian neural networks form the core for uncertainty estimation by modeling network weights as stochastic distributions, where test-time-dropout [41] is utilized as a variational Bayesian approximation [15] and Monte Carlo integration is computed via multiple stochastic forward passes to approximate prediction uncertainty [16]. This Bayesian deep learning framework has been applied successfully to supervised 2D dense prediction tasks [26, 27]. We focus on the aggregate uncertainty in the predictions of a general 3D segmentation model and build on the Bayesian deep learning approach of Gal et al. [16] which examines the approximate predictive entropy as an acquisition function for active learning. However, contrary to Gal et al. [16], in our self-training setup, we use low rather than high entropy as a positive indication for considering a point as a source of supervision. To the best of our knowledge, our work is the first to introduce such a Bayesian self-training framework for general 3D point cloud segmentation, specifically including 3D instance segmentation and referral-based 3D instance segmentation.", + "bbox": [ + 212, + 352, + 787, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Related uncertainty-based approaches to ours have been proposed for the semantic segmentation of 3D grid volume data in the field of medical image computing [12,54,58]. In particular, Yu et al. [58] also use Monte Carlo dropout to estimate predictive entropy but only applies dropout to a single network input, whereas we vary both network weights and input views. Moreover, that work leverages the estimated entropy merely as the weight in a consistency loss between a teacher and a student network, while we filter pseudo-labels based on our entropy estimates. Besides, Xia et al. [54] perform semi-supervised semantic segmentation with multi-view co-training and applies Bayesian deep learning across views for uncertainty-based weighting of pseudo-labels. We only train one network instead of multiple networks employed in Xia et al. [54] and we derive the uncertainty estimates by feeding all input views to that single network, which is more parameter-efficient. Both of these approaches only address 3D semantic segmentation, and cannot be utilized in instance-level dense prediction tasks.", + "bbox": [ + 212, + 628, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bayesian Self-Training for Semi-Supervised 3D Segmentation", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/33aa6741a249313c163f3dde3e2408db3f6c1b7f521ff2908ff9085a622b7eda.jpg", + "image_caption": [ + "Fig. 2: Illustration of Bayesian pseudo-labeling for semi-supervised instance segmentation. We initialize a seed prediction via a forward pass using a non-augmented input. Then our heuristic $n$ -partite matching algorithm is employed to pair each seed instance with the best matching predicted instance from each of the $K$ stochastic forward passes. For each aligned object, we compute the aggregated label through unanimous voting and filter it based on uncertainty to obtain the final pseudo mask." + ], + "image_footnote": [], + "bbox": [ + 228, + 145, + 774, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Bayesian Self-Training", + "text_level": 1, + "bbox": [ + 215, + 411, + 470, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Sec. 3.1 we start by formulating a self-training approach for semi-supervised 3D semantic segmentation based on established uncertainty estimation techniques from the field of active learning [15,46]. We then extend our formulation to include instance labels in Sec. 3.2 and finally build up to also take natural language prompts as input, i.e., towards semi-supervised dense 3D visual grounding in Sec. 3.3.", + "bbox": [ + 212, + 446, + 787, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Semi-Supervised 3D Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 563, + 627, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In a semi-supervised setting, a dataset $D$ consists of a labeled set $S$ and an unlabeled set $U$ . The baseline approach when working with such data is to treat the problem similarly to any fully supervised task and employ a semantic loss $\\mathcal{L}_{sem}$ on the available labeled scenes $S$ , typically stated as cross-entropy. Formally, the objective function over the model parameters $\\omega$ is formulated as:", + "bbox": [ + 212, + 590, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\omega} \\sum_ {i = 1} ^ {| S |} \\mathcal {L} _ {s e m} (\\hat {\\mathbf {y}}, \\mathbf {y}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 678, + 785, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Starting from this baseline formulation, a semi-supervised training strategy aims to leverage the unlabeled scenes $U$ to further improve model performance. One such common technique is self-training, where the model is iteratively trained on its own predictions, i.e. on generated per-point one-hot pseudo-label vectors $\\hat{\\mathbf{y}} = [\\hat{\\mathbf{y}}^{(1)},\\dots,\\hat{\\mathbf{y}}^{(C)}]$ , with $C$ denoting the number of classes. However, the task of generating pseudo-labels is not trivial, as the model may suffer from confirmation bias that occurs when repeatedly overfitting to incorrect labels.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "O. Unal et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A simple but effective solution is explored under the umbrella of thresholding [5, 19, 55, 65, 66]. In essence, instead of training on all model predictions, a pseudo-label set $L$ is formed by predictions that exceed a confidence of a determined threshold $\\tau$ to ensure a higher level of accuracy. Formally, the objective function of threshold-based self-training for semantic segmentation can be formulated as:", + "bbox": [ + 214, + 146, + 787, + 236 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\omega , \\hat {\\mathbf {y}}} \\sum_ {d = 1} ^ {| D |} \\left(\\mathcal {L} _ {s e m} (\\hat {\\mathbf {y}}, \\mathbf {y}) - (\\log \\hat {\\mathbf {y}} _ {d} - \\tau) \\cdot \\hat {\\mathbf {y}} _ {d}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 241, + 785, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with $\\hat{\\mathbf{y}}_d$ denoting the softmax model predictions. Eq. 2 is optimized in two steps that are repeated until convergence:", + "bbox": [ + 214, + 289, + 787, + 319 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Training: First, $\\hat{\\mathbf{y}}$ is fixed, and the objective function is optimized with respect to $\\omega$ .", + "2. Pseudo-labeling: Then, $\\omega$ is fixed and the objective function is optimized with respect to $\\hat{\\mathbf{y}}$ . Consequently, $L$ is updated given $\\hat{\\mathbf{y}}$ ." + ], + "bbox": [ + 220, + 323, + 785, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The process is initialized by setting the latent variable $\\hat{\\mathbf{y}} = \\mathbf{0}$ for all points, i.e. by reverting to the naive solution of Eq. 1 and only training on the labeled set $S$ . The two steps can be repeated to take advantage of the improved representation capability of the model thanks to pseudo-labeling.", + "bbox": [ + 214, + 388, + 785, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While naïve thresholding relies on the softmax scores of the model to determine confidence, these scores may be unreliable and often skewed when facing long-tailed distributions, resulting in overconfident predictions [18]. Inspired by uncertainty estimation techniques commonly employed in active learning (AL) pipelines [21,46], we introduce a Bayesian self-training strategy that leverages an uncertainty-aware filtering technique for generating more reliable pseudo-labels.", + "bbox": [ + 214, + 449, + 787, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first start by decomposing the goal of 3D semantic segmentation, which is to discover the dependency of the point-wise distribution over the labels $y \\in Y$ on an input sample $x \\in X$ via the model weights $\\omega$ . Formally, the task can be reduced to per-point classification, where $y$ can be of class $c \\in \\{1, \\dots, C\\}$ . The conditional probability is therefore defined by:", + "bbox": [ + 214, + 540, + 787, + 614 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np (y = c | x, D) = \\int p (y = c | a (x), \\omega) p (\\omega | D) p (a (x) | x) d \\omega d x \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 621, + 785, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for the dataset $D = (X, Y)$ , under randomly applied affine transformations modeled through $a(\\cdot)$ .", + "bbox": [ + 214, + 657, + 785, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We set $p(a(x)|x)$ as a constant, restate the intractable $p(\\omega |D)$ via the variational approximation $q(\\omega)$ and minimize $\\mathrm{KL}(q(\\omega)|p(\\omega |X,Y))$ . This objective can be approximated through Monte Carlo integration by employing variational inference via Bayesian networks, i.e. by applying stochastic forward passes with weights $\\omega_{k}$ :", + "bbox": [ + 214, + 686, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p (y = c | x, D) \\approx \\int p (y = c | a (x), \\omega) q (\\omega) d \\omega d x \\\\ \\approx \\frac {1}{K} \\sum_ {k = 1} ^ {K} p (y = c | a (x), \\omega_ {k}). \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 768, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bayesian Self-Training for Semi-Supervised 3D Segmentation", + "bbox": [ + 325, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to the common decomposition of BALD [21], the inclusion of the affine transformations via $a(\\cdot)$ allows a higher deviation amongst stochastic forward passes which in return allows better estimation of predictive uncertainty [49]. We capture the uncertainty via the Shannon entropy [40]:", + "bbox": [ + 215, + 146, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {H} [ y | x, D ] = - \\sum_ {c = 1} ^ {C} p (y = c | x, D) \\log p (y = c | x, D) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 219, + 785, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and the accumulated predicted label $\\hat{\\mathbf{y}}^*$ is given by the unanimous voting principle:", + "bbox": [ + 214, + 271, + 785, + 300 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} ^ {* (c)} = \\left\\{ \\begin{array}{l} 1, \\text {i f} K = \\sum_ {k} \\mathbb {1} (\\operatorname {a r g m a x} (p (y | a (x), \\omega_ {k})) = c) \\\\ 0, \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 301, + 785, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $\\mathbb{1}$ denoting the indicator function.", + "bbox": [ + 215, + 345, + 498, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike what is commonly seen in active learning settings, our goal is not to maximize the total entropy but rather to minimize it in order to ensure a high-quality pseudo-label set. We therefore redefine the self-training objective (Eq. 2) for stochastic pseudo-labeling as:", + "bbox": [ + 214, + 361, + 785, + 421 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\omega , \\hat {\\mathbf {y}}} \\sum_ {d = 1} ^ {D} \\left(\\mathcal {L} _ {s e m} (\\hat {\\mathbf {y}}, \\mathbf {y}) + \\left(\\mathbb {H} [ y | x, D ] - \\tau\\right) \\left(\\hat {\\mathbf {y}} ^ {*} \\cdot \\hat {\\mathbf {y}}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 433, + 785, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $\\tau \\in (0,1)$ determining the cut-off entropy threshold. To solve the nonlinear integer optimization task, we employ the following solver:", + "bbox": [ + 214, + 484, + 785, + 516 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {y}} ^ {(c)} = \\left\\{ \\begin{array}{l} 1, \\text {i f} \\mathbb {H} [ y | x, D ] < \\tau \\text {a n d} \\\\ c = \\operatorname {a r g m a x} \\hat {\\mathcal {y}} ^ {*} \\\\ 0, \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 527, + 785, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where labels $\\hat{\\mathbf{y}}^{(c)} = 0\\forall c$ are ignored during training.", + "bbox": [ + 215, + 594, + 598, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In practice, following precedent [47,66], we set $\\tau$ based on $p_{\\tau}$ , with $p_{\\tau}$ determining the percentage of pseudo-labels to be sampled amongst the accumulated predicted labels.", + "bbox": [ + 212, + 611, + 785, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, we first determine a set of labels that are agreed upon by all stochastic forward passes, i.e. by forward passes with active dropout [15, 41]. We then compute the Shannon entropy for all predictions, filtering those that exhibit high entropy. This surprisingly simple but effective approach allows us to reduce the incorrect targets that we introduce to our pseudo-label set and to avoid the overconfidence issue of naive thresholding-based approaches.", + "bbox": [ + 214, + 656, + 785, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Semi-Supervised 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 215, + 768, + 620, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed Bayesian self-training framework provides a strong basis for any semi-supervised 3D dense prediction task, one such example being 3D instance segmentation. Yet, compared to semantic segmentation where accumulation of", + "bbox": [ + 214, + 794, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "O. Unal et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Bayesian pseudo-label generation for instance segmentation." + ], + "code_body": "Input: Point cloud $P\\in \\mathbb{R}^{N\\times 3}$ , trained model with dropout $\\omega$ , number of passes $K$ , threshold $\\tau$ \n1 seed_output $= \\omega (P)\\in \\mathbb{R}^{M\\times N}$ \n2 $\\pmb{i} =$ seed_output.copy() \n3 seed_mask $=$ to_mask(i) $\\in \\{0,1\\}^{M\\times N}$ \n4 unanimous $=$ seed_mask.copy() \n5 for k in 1:K do \n6 output $= \\omega_{k}(a(P))\\in \\mathbb{R}^{M^{\\prime}\\times N}$ \n7 mask $=$ to_mask(output) $\\in \\{0,1\\}^{M^{\\prime}\\times N}$ \n8 cost $=$ -iou(seed_mask, mask) \n9 row,col $=$ LSA(cost) \n10 unanimous[col] &= mask[row] \n11 i[col] $+ =$ output[row] \n12 $\\mathbb{H} = (K - i)\\cdot \\log (K - i) + i\\log i$ \n13 pseudo_label $= (\\mathbb{H} < \\tau)$ & unanimous \nReturn:pseudo_label $\\in \\{0,1\\}^{M\\times N}$", + "bbox": [ + 217, + 179, + 759, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "labels can be easily handled across multiple stochastic forward passes, accumulation of predicted instance masks is not trivial and requires $n$ -partite matching.", + "bbox": [ + 212, + 453, + 784, + 483 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "However, the $n$ -partite matching problem is NP-hard with a complexity of $|I|!(K-1)$ with $K$ passes and $|I|$ number of instances assuming full bijection. To match predicted instance masks across various stochastic forward passes, we propose a heuristic method based on the Hungarian algorithm. First, we generate an initial prediction via an unaugmented forward pass (Alg. 1 L1&3). The goal is to establish a seed that is likely of a higher quality with minimal perturbations of the data. We then iteratively assign predicted masks of each stochastic forward pass via linear sum assignment, with the cost matrix formed by the negative pairwise IoUs (Alg. 1 L5-10).", + "bbox": [ + 212, + 484, + 785, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Having matched predicted binary instance masks from individual stochastic forward passes, each resulting set of masks can then be treated as a set of 2-class semantic segmentation outputs, and thus our proposed Bayesian pseudo-labeling approach from Eq. 8 can be trivially applied (Alg. 1 L11-15). Our pipeline for semi-supervised 3D instance segmentation is illustrated in Fig. 2.", + "bbox": [ + 212, + 619, + 785, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Semi-Supervised Dense 3D Visual Grounding", + "text_level": 1, + "bbox": [ + 214, + 720, + 635, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Similar to both semi-supervised 3D semantic and 3D instance segmentation, our proposed Bayesian self-training formulation can easily be extended to include natural language prompts as inputs and thus tackle semi-supervised dense 3D visual grounding, i.e., referral-based 3D instance segmentation.", + "bbox": [ + 212, + 748, + 784, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A common technique when tackling dense 3D visual grounding is to employ a grounding-by-selection strategy which can be summarised in two steps [6,48,62]:", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bayesian Self-Training for Semi-Supervised 3D Segmentation", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Visual: First, a 3D backbone computes object candidates $O$ from the scene.", + "2. Verbo-visual: Then, a verbo-visual fusion module $\\omega_{sel}$ selects the correct instance candidate from predicted instances $I$ based on the natural language description $W$ , i.e., it predicts the index:" + ], + "bbox": [ + 220, + 146, + 784, + 205 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\ni ^ {*} = \\operatorname {a r g m a x} G, \\text {w i t h} G = \\omega_ {s e l} (O, W) \\in \\mathbb {R} ^ {| I |}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 215, + 784, + 233 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To extend our Bayesian self-training approach to this task, we follow the above structure and individually adapt each of the two steps, with the adaptation of the visual step simply following Sec. 3.2.", + "bbox": [ + 212, + 243, + 782, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For the verbal side of 3D visual grounding, i.e. referral-based instance selection, we observe that the pseudo-label generation can be reduced to the same formulation as the pseudo-label generation for 3D instance segmentation. Utilizing the results from the heuristic $n$ -partite matching (Alg. 1 L9), vector $G$ can be reordered to match across stochastic forward passes. Again, having matched predicted probability vectors, the method presented in Sec. 3.1 can be trivially applied (Alg. 1 L11-15).", + "bbox": [ + 212, + 289, + 784, + 393 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 416, + 375, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate our method on all three semi-supervised 3D semantic perception tasks: 3D semantic segmentation, 3D instance segmentation, and dense 3D visual grounding. The dataset and implementation details regarding each task can be found in the supplementary materials.", + "bbox": [ + 212, + 446, + 782, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 3D Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 529, + 485, + 544 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The comparison to the state of the art for semi-supervised 3D semantic segmentation is shown in Tab. 1. As seen, our Bayesian self-training strategy outperforms existing work in both SemanticKITTI [3] and ScribbleKITTI [47] on splits of $10\\%$ , $20\\%$ and $50\\%$ , while performing on par for $1\\%$ . Specifically, compared to the baseline Cylinder3D [64] model denoted as Sup-only, our strategy allows significant improvements ranging from $4.4\\%$ to $10.9\\%$ mIoU. In Fig. 3 (a-c) we showcase qualitative results on SemanticKITTI $10\\%$ .", + "bbox": [ + 212, + 553, + 784, + 657 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 215, + 680, + 478, + 695 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For semi-supervised 3D instance segmentation, we report results for ScanNet [11] in Tab. 2 and for S3DIS [2] in Tab. 3. As seen, we observe a similar behavior to semi-supervised 3D semantic segmentation, where at higher thresholds of $10\\%$ , $20\\%$ for ScanNet and $20\\%$ , $50\\%$ for S3DIS, our simple yet effective Bayesian self-training strategy significantly outperforms state-of-the-art methods that require highly engineered solutions. In contrast, at lower thresholds, we see a significant gap in performance (especially for ScanNet). Still, compared to the baseline Sup-only, we see significant gains across the board for all semi-splits. Additionally in Fig. 3 (d-f), we show qualitative results on the ScanNet val-set.", + "bbox": [ + 212, + 703, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "O. Unal et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/0c1e3aca2b707086eaa2c4a75b30a25a72dbc58fbfdf01b07c8f43ae6d91bed1.jpg", + "table_caption": [ + "Table 1: Comparison of state-of-the-art methods for semi-supervised LiDAR semantic segmentation (mIoU [%]). For fair comparison, all methods use the same backbone given by Sup-only, which at $100\\%$ labels achieves $64.3\\%$ mIoU on SemanticKITTI and $57.0\\%$ mIoU on ScribbleKITTI." + ], + "table_footnote": [], + "table_body": "
Labeled1%10%20%50%
SemantickKITTI [3]Sup-only [64]45.456.157.858.7
DiAL [46]45.457.159.260.0
CBST [66]48.858.359.459.7
CPS [8]46.758.759.660.5
LaserMix [29]50.660.061.962.3
IGNet [45]49.061.363.164.8
Ours49.861.763.764.1
Δ Sup-only+4.4+5.6+5.9+5.4
ScribbleKITTI [47]Sup-only [64]39.248.052.153.8
DiAL [46]41.050.152.853.9
CBST [66]41.550.653.354.5
CPS [8]41.451.853.954.8
LaserMix [29]44.253.755.156.8
IGNet [45]44.457.759.660.8
Ours43.958.960.161.2
Δ Sup-only+4.7+10.9+8.0+7.4
", + "bbox": [ + 290, + 205, + 702, + 448 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5a9510bcb23fed4d595725be4d2b078029c69cb34debc971e4d15ac876b6915c.jpg", + "table_caption": [ + "Table 2: Comparison to state-of-the-art methods for semi-supervised 3D instance segmentation on the ScanNet Limited Reconstruction valid-set. For a fair comparison, all methods apart from that marked with * use the same backbone [20,53] as Sup-only." + ], + "table_footnote": [], + "table_body": "
Labeled Metric1% mAPAP50AP255% mAPAP50AP2510% mAPAP50AP2520% mAPAP50AP25
Sup-only [10]5.19.817.618.232.047.026.742.858.929.347.963.0
PointCont. [56]7.212.520.319.435.448.527.043.959.530.249.563.6
CSC [20]7.113.021.220.936.750.627.345.060.230.650.364.1
TWIST [10]9.617.126.227.044.156.230.649.763.032.852.966.8
WS3D* [33]-32.5--45.6--49.2--51.3-
Ours7.214.223.024.239.351.432.751.965.537.456.168.9
Δ Sup-only+2.1+4.4+5.4+6.0+7.3+4.4+6.0+9.1+6.6+8.1+8.2+5.9
", + "bbox": [ + 217, + 503, + 781, + 636 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Dense 3D Visual Grounding", + "text_level": 1, + "bbox": [ + 215, + 646, + 496, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For semi-supervised 3D visual grounding, we report results on the ScanRefer [6] dataset. As seen in Tab. 4 (Vi), our self-training approach shows improvements across the board, with significant gains on the \"unique\" category.", + "bbox": [ + 212, + 672, + 785, + 718 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "While natural language prompts still require human effort to generate, they are considerably cheaper and easier to acquire than per-point high-quality panoptic labels. In Tab. 4 (ViVe), we show that the inclusion of such prompts in the unlabeled dataset further benefits the model's accuracy. As seen, we observe even larger gains over the baseline method with the overall accuracy (OA) showing $+4.7\\%$ improvement with a $50\\%$ IoU threshold.", + "bbox": [ + 212, + 719, + 787, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Furthermore, in Fig. 3 (g) we qualitatively compare our method to the baseline Sup-only on ScanRefer.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bayesian Self-Training for Semi-Supervised 3D Segmentation", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/17435439459c2fdcb7ba5035be734083ab12ecacf852b75174414faf883fc9be.jpg", + "image_caption": [ + "Fig.3: Qualitative results from the SemanticKITTI (10%) val-set, comparing a) the ground truth, b) Sup-only and c) ours; on ScanNet (10%), comparing d) the ground-truth instance masks, e) Sup-only and f) ours; and finally g) ScanRefer (10%)." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 482, + 324 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a3d0b6e442955374224e691faa8599fd07a346a34ff644a93aa7b0f97345270d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 146, + 630, + 234 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ac34a78f41280f5d2af4764e42562d5eb617bf59bb496b79827d6b48c7e07b94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 234, + 630, + 321 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/234c063d3d34d436477c3eeb1231824c17a3ae258907b07a84541d5debd887ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 148, + 779, + 234 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/59ed4a2de850b2b756a79bfaca50b8c0abf650d965fe1226a435cba4a9343518.jpg", + "image_caption": [ + "The table is square. There is a chair to the left, and a chair to the right of it." + ], + "image_footnote": [], + "bbox": [ + 637, + 236, + 779, + 303 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/525d8a8ab28cddfd8e88d2ebaeb664ed073ac65d7009594b39368d6ce59c8463.jpg", + "table_caption": [ + "Table 3: Comparison to state-of-the-art methods for semi-supervised 3D instance segmentation on S3DIS Area 5 via AP50 [%]. For a fair comparison, all methods use the same backbone Sparse U-Net [20,53] as Sup-only." + ], + "table_footnote": [], + "table_body": "
Labeled5%10%20%50%
Sup-only [10]30.436.841.246.5
PointContrast [56]33.638.743.148.9
CSC [20]34.241.044.750.4
TWIST [10]37.145.648.4-
Ours35.042.951.156.6
Δ Sup-only+4.6+6.1+9.9+10.1
", + "bbox": [ + 287, + 439, + 705, + 541 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 561, + 401, + 575 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct ablation studies on $10\\%$ splits of SemanticKITTI [3], ScanNet [11] and ScanRefer [6] for semi-supervised 3D semantic segmentation, instance segmentation and visual grounding respectively.", + "bbox": [ + 212, + 595, + 785, + 641 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Entropy Thresholding: In Tab. 5, we show an analysis of the effectiveness of our proposed uncertainty thresholding via Shannon entropy, comparing it to the baseline unanimous voting results as well as established thresholding techniques of softmax thresholding [55] and class-balanced thresholding [66]. Alongside the performance for each individual task, we also report the accuracy of the generated pseudo-labels. For instance segmentation, a pseudo-mask is considered correct if its intersection with the associated ground-truth mask is identical to the former. For visual grounding, we reduce the accuracy threshold to 0.5 following the benchmark metric used. As seen, our uncertainty-based filtering improves pseudo-label quality across the board. We also showcase the performance implications of uncertainty filtering, where it can be observed that entropy thresholding outperforms both unanimous voting and current thresholding strategies across all tasks. We further provide a qualitative analysis of the uncertainty met-", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "O. Unal et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2d66b242a3f41c7b1c0c582a90bfc737e1450cb26559fb9eb29a0c8051b8b53c.jpg", + "table_caption": [ + "Table 4: Results for semi-supervised 3D dense visual grounding while assuming (Vi) that the model is given no natural language descriptions for the unlabeled split and (ViVe) when the model has access to natural language descriptions for the unlabeled split. Reported values are given as Acc@25/50 for the $10\\%$ split." + ], + "table_footnote": [], + "table_body": "
CategoryUniqueMultipleOverall
Sup-only [48]51.9 / 43.619.3 / 14.125.4 / 19.6
ViMe ViOurs58.8 / 48.822.0 / 18.228.8 / 23.9
Δ Sup-only+6.9 / +5.2+2.7 / +4.1+3.4 / +4.3
Ours61.0 / 51.223.4 / 18.130.4 / 24.3
Δ Sup-only+9.1 / +7.6+4.1 / +4.0+5.0 / +4.7
", + "bbox": [ + 299, + 205, + 692, + 292 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/49900eaa5474a7613075770dbe8494e6182581e3bd069675fc976ebb3a8cebab.jpg", + "table_caption": [ + "Table 5: Comparison of the pseudo-label quality of current thresholding strategies and their effects on performance on SemanticKITTI $10\\%$ , ScanNet $10\\%$ , ScanRefer $10\\%$ (with visual and verbal) for each respective task. Pseudo-label thresholding is applied after unanimous voting for a fair comparison. The letters indicate each pseudolabel type: $S-$ for semantic, $I-$ for instance, and $G-$ for grounding." + ], + "table_footnote": [], + "table_body": "
SemanticInstanceVisual Grounding
MethodS-AccmIoUS-AccI-AccAP@50S-AccI-AccG-AccOA@50
Sup-only [48]-56.1--42.8---19.6
Unanimous96.460.079.375.347.072.969.961.516.9
Naïve [5,55,65]98.360.882.977.750.282.073.068.621.9
Class-balanced [66]96.661.181.270.144.480.969.060.616.2
Entropy-based Δ Unanimous98.661.784.877.851.983.875.275.624.3
+2.2+1.7+5.5+2.5+4.9+5.8+2.3+5.7+7.4
", + "bbox": [ + 217, + 378, + 781, + 494 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3d92ef1074863bd45a8f330dfeef17aac404341158192cfb45aabb8d9535afd4.jpg", + "image_caption": [ + "Fig. 4: Qualitative analysis of the uncertainty estimation. We compare softmax confidence to our MC-derived Shannon entropy." + ], + "image_footnote": [], + "bbox": [ + 217, + 511, + 785, + 583 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ric in Fig. 4 to illustrate the better correlation of our MC-derived uncertainty with pseudo-label accuracy compared to softmax based uncertainty.", + "bbox": [ + 212, + 640, + 782, + 670 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Labeling Percentage $(p_{\\tau})$ : We compare various set labeling percentages $p_{\\tau}$ across all tasks in Tab. 6. While we obtain higher accuracy pseudo-labels at $50\\%$ , it can be seen that the reduced supervision negatively impacts performance.", + "bbox": [ + 212, + 671, + 784, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Limitations and Discussion", + "text_level": 1, + "bbox": [ + 214, + 743, + 513, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The main advantage of the proposed Bayesian self-training is its simplicity and ease of implementation compared to existing SOTA approaches that utilize highly engineered and often costly modules. However, the method has a limitation. As observed from the results, the performance is comparably inferior to", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bayesian Self-Training for Semi-Supervised 3D Segmentation", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/31e8a4c7f17156783fc9846cea32572c4d4ff658b64302e45449260d9ebaa6c0.jpg", + "table_caption": [ + "Table 6: Ablation study analyzing the effect of the hyperparameter $p_{\\tau}$ for each task. We vary the labeling percentage and show the performance on SemanticKITTI 10%, ScanNet 10%, and ScanRefer 10% (incl. verbal input) for each respective task." + ], + "table_footnote": [], + "table_body": "
ThresholdmIoUAP@50OA@50
50%61.450.022.8
75%61.751.924.3
100%60.047.016.9
", + "bbox": [ + 308, + 191, + 683, + 252 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9fd91a96193cd63d6d590fc52ee878daae9ec84f2415a463a9c7fb546c5e2615.jpg", + "table_caption": [ + "Table 7: Ablation study showing the performance on SemanticKITTI $10\\%$ , ScanNet $10\\%$ and ScanRefer $10\\%$ (incl. verbal input) for varying label aggregation strategies from majority rules to unanimous voting over 9 stochastic forward passes." + ], + "table_footnote": [], + "table_body": "
ThresholdmIoUAP@50OA@50
5/960.747.920.1
7/961.051.023.9
9/961.751.924.3
", + "bbox": [ + 308, + 306, + 683, + 367 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "SOTA semi-supervised approaches for highly sparse training splits. This behavior stems from the reliance on unanimous voting (Eq. 6). With low amounts of labeled data, the initial naive model performance suffers, causing a reduction in the number of unanimously voted points. This effect is further pronounced when tackling instance segmentation, where fused instance masks may affect multiple predictions. In Tab. 7, we show that relaxation does not aid the performance as it further introduces incorrect predictions into the pseudo-label set.", + "bbox": [ + 212, + 382, + 784, + 487 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While semi-supervised training methods that enhance the robustness of the pseudo-labels either via a teacher model (DiAL [44, 46]), mix-augmentations (LaserMix [29]) or multi-stage refinement (TWIST [10]) are presented as competing to ours, we argue that as a solution, they can in fact be utilized within our framework in place of the 3D model to improve the quality of the initial predictions. This would allow a larger set of unanimously voted samples and would thus re-enable the benefits of our proposed method on highly sparse splits.", + "bbox": [ + 212, + 488, + 785, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 617, + 359, + 633 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce a simple yet effective Bayesian self-training approach for semi-supervised 3D semantic segmentation by utilizing uncertainty estimation strategies from the field of active learning. Specifically, through multiple stochastic forward passes, we generate an initial pseudo-label set and then filter it based on estimated predictive entropy. We further extend our semi-supervised 3D method to newly include both instance segmentation and referral-based instance segmentation via an efficient, heuristic $n$ -partite matching algorithm. We conduct extensive experiments on multiple datasets, covering both indoor and outdoor domains, showing that our method performs on par if not better than current domain- and task-specific methods.", + "bbox": [ + 212, + 648, + 787, + 800 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments: This work is funded by Toyota Motor Europe via the research project TRACE-Zürich.", + "bbox": [ + 215, + 806, + 784, + 837 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "O. Unal et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: ReferIt3D: Neural listeners for fine-grained 3D object identification in real-world scenes. In: ECCV (2020)", + "2. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3D semantic parsing of large-scale indoor spaces. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "3. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)", + "4. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3DJCG: A unified framework for joint dense captioning and visual grounding on 3D point clouds. In: CVPR (2022)", + "5. Cascante-Bonilla, P., Tan, F., Qi, Y., Ordonez, V.: Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 6912-6920 (2021)", + "6. Chen, D.Z., Chang, A.X., Nießner, M.: ScanRefer: 3D object localization in RGB-D scans using natural language. In: ECCV (2020)", + "7. Chen, J., Luo, W., Wei, X., Ma, L., Zhang, W.: HAM: Hierarchical attention model with high performance for 3D visual grounding. arXiv preprint arXiv:2210.12513 (2022)", + "8. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "9. Cheng, M., Hui, L., Xie, J., Yang, J.: SSPC-Net: Semi-supervised semantic 3D point cloud segmentation network. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)", + "0. Chu, R., Ye, X., Liu, Z., Tan, X., Qi, X., Fu, C.W., Jia, J.: TWIST: Two-way inter-label self-training for semi-supervised 3D instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "1. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: ICCV (2017)", + "2. Dai, W., Li, X., Cheng, K.T.: Semi-supervised deep regression with uncertainty consistency and variational model ensembling via bayesian neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7304-7313 (2023)", + "3. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3D point clouds. In: International Conference on Robotics and Automation (ICRA) (2022)", + "4. Feng, M., Li, Z., Li, Q., Zhang, L., Zhang, X., Zhu, G., Zhang, H., Wang, Y., Mian, A.: Free-form description guided 3D visual graph network for object grounding in point cloud. In: ICCV (2021)", + "5. Gal, Y., Ghahramani, Z.: Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In: International Conference on Machine Learning (2016)", + "6. Gal, Y., Islam, R., Ghahramani, Z.: Deep Bayesian active learning with image data. In: International Conference on Machine Learning (2017)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bayesian Self-Training for Semi-Supervised 3D Segmentation", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Gao, H.a., Tian, B., Li, P., Zhao, H., Zhou, G.: DQS3D: Densely-matched quantization-aware semi-supervised 3D detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "18. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International Conference on Machine Learning (2017)", + "19. He, R., Yang, J., Qi, X.: Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "20. Hou, J., Graham, B., Niessner, M., Xie, S.: Exploring data-efficient 3D scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "21. Houlsby, N., Huszar, F., Ghahramani, Z., Lengyel, M.: Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745 (2011)", + "22. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3D instance segmentation. In: AAAI (2021)", + "23. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3D visual grounding. In: CVPR (2022)", + "24. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: ECCV (2022)", + "25. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "26. Kendall, A., Badrinarayanan, V., Cipolla, R.: Bayesian SegNet: Model uncertainty in deep convolutional encoder-decoder architectures for scene understanding. In: Proceedings of the British Machine Vision Conference (BMVC) (2017)", + "27. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "28. Kong, C., Lin, D., Bansal, M., Urtasun, R., Fidler, S.: What are you talking about? text-to-image coreference. In: ICCV (2014)", + "29. Kong, L., Ren, J., Pan, L., Liu, Z.: LaserMix for semi-supervised LiDAR semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "30. Lai, X., Yuan, Y., Chu, R., Chen, Y., Hu, H., Jia, J.: Mask-attention-free transformer for 3D instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "31. Lee, M.S., Yang, S.W., Han, S.W.: Gaia: Graphical information gain based attention network for weakly supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 582-591 (2023)", + "32. Lei, J., Deng, C., Schmeckpeper, K., Guibas, L., Daniilidis, K.: EFEM: Equivariant neural field expectation maximization for 3D object segmentation without scene supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "33. Liu, K., Zhao, Y., Nie, Q., Gao, Z., Chen, B.M.: Weakly supervised 3D scene segmentation with region-level boundary awareness and instance discrimination. In: European Conference on Computer Vision (ECCV) (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "O. Unal et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3D-SPS: Single-stage 3D visual grounding via referred point progressive selection. In: CVPR (2022)", + "35. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Van Gool, L.: Towards a weakly supervised framework for 3D point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4454-4468 (2021)", + "36. Prabhudesai, M., Tung, H.Y.F., Javed, S.A., Sieb, M., Harley, A.W., Fragkiadaki, K.: Embodied language grounding with 3D visual feature representations. In: CVPR (2020)", + "37. Qin, Z., Wang, J., Lu, Y.: Weakly supervised 3D object detection from point clouds. In: Proceedings of the ACM International Conference on Multimedia (2020)", + "38. Ren, Z., Misra, I., Schwing, A.G., Girdhar, R.: 3D spatial recognition without spatially labeled 3D. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "39. Roh, J., Desingh, K., Farhadi, A., Fox, D.: LanguageRefer: Spatial-language model for 3D visual grounding. In: CoRL (2022)", + "40. Shannon, C.E.: A mathematical theory of communication. The Bell system technical journal 27(3) (1948)", + "41. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research 15(1), 1929-1958 (2014)", + "42. Tang, Y.S., Lee, G.H.: Transferable semi-supervised 3D object detection from rgb-d data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019)", + "43. Tao, A., Duan, Y., Wei, Y., Lu, J., Zhou, J.: SegGroup: Seg-level supervision for 3D instance and semantic segmentation. IEEE Transactions on Image Processing 31, 4952-4965 (2022)", + "44. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems (2017)", + "45. Unal, O., Dai, D., Hoyer, L., Can, Y.B., Van Gool, L.: 2d feature distillation for weakly- and semi-supervised 3d semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7336-7345 (January 2024)", + "46. Unal, O., Dai, D., Unal, A.T., Van Gool, L.: Discwise active learning for LiDAR semantic segmentation. IEEE Robotics and Automation Letters (2023)", + "47. Unal, O., Dai, D., Van Gool, L.: Scribble-supervised LiDAR semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "48. Unal, O., Sakaridis, C., Saha, S., Gool, L.V.: Four ways to improve verbo-visual fusion for dense 3d visual grounding (2024), https://arxiv.org/abs/2309.04561", + "49. Wang, G., Li, W., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T.: Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing 338, 34-45 (2019)", + "50. Wang, H., Cong, Y., Litany, O., Gao, Y., Guibas, L.J.: 3D IoUMatch: Leveraging IoU prediction for semi-supervised 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bayesian Self-Training for Semi-Supervised 3D Segmentation", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Wang, J., Gang, H., Ancha, S., Chen, Y.T., Held, D.: Semi-supervised 3D object detection via temporal graph neural networks. In: International Conference on 3D Vision (3DV) (2021)", + "52. Wei, J., Lin, G., Yap, K.H., Hung, T.Y., Xie, L.: Multi-path region mining for weakly supervised 3D semantic segmentation on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "53. Wu, Y., Shi, M., Du, S., Lu, H., Cao, Z., Zhong, W.: 3d instances as 1d kernels. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 235-252. Springer (2022)", + "54. Xia, Y., Liu, F., Yang, D., Cai, J., Yu, L., Zhu, Z., Xu, D., Yuille, A., Roth, H.: 3D semi-supervised learning with uncertainty-aware multi-view co-training. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (March 2020)", + "55. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10687-10698 (2020)", + "56. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3D point cloud understanding. In: ECCV (2020)", + "57. Xu, X., Lee, G.H.: Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "58. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmentation. In: MICCAI (2019)", + "59. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: InstanceRefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: ICCV (2021)", + "60. Zhang, F., Fang, J., Wah, B., Torr, P.: Deep fusionnet for point cloud semantic segmentation. In: ECCV (2020)", + "61. Zhang, Y., Li, Z., Xie, Y., Qu, Y., Li, C., Mei, T.: Weakly supervised semantic segmentation for large-scale point cloud. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)", + "62. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3DVG-Transformer: Relation modeling for visual grounding on point clouds. In: ICCV (2021)", + "63. Zhao, N., Chua, T.S., Lee, G.H.: SESS: Self-ensembling semi-supervised 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "64. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "65. Zoph, B., Ghiasi, G., Lin, T.Y., Cui, Y., Liu, H., Cubuk, E.D., Le, Q.V.: Rethinking pre-training and self-training. arXiv preprint arXiv:2006.06882 (2020)", + "66. Zou, Y., Yu, Z., Vijaya Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: The European Conference on Computer Vision (ECCV) (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 784 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "O. Unal et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/546e4382-dd78-463e-8d89-435110642a17_model.json b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/546e4382-dd78-463e-8d89-435110642a17_model.json new file mode 100644 index 0000000000000000000000000000000000000000..005abab3d97d1ebfcf2996d4aae0c7504966ad2a --- /dev/null +++ b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/546e4382-dd78-463e-8d89-435110642a17_model.json @@ -0,0 +1,2535 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.258, + 0.141, + 0.747, + 0.186 + ], + "angle": 0, + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.212, + 0.712, + 0.227 + ], + "angle": 0, + "content": "Ozan Unal\\(^{1,2}\\), Christos Sakaridis\\(^{1}\\), and Luc Van Gool\\(^{1,3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.239, + 0.703, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\)ETH Zurich, \\(^{2}\\)Huawei Technologies, \\(^{3}\\)KU Leuven, \\(^{4}\\)INSAIT {ozan.unal, csakarid, vangool}@vision.ee.ethz.ch" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.302, + 0.74, + 0.593 + ], + "angle": 0, + "content": "Abstract. 3D segmentation is a core problem in computer vision and, similarly to many other dense prediction tasks, it requires large amounts of annotated data for adequate training. However, densely labeling 3D point clouds to employ fully-supervised training remains too labor intensive and expensive. Semi-supervised training provides a more practical alternative, where only a small set of labeled data is given, accompanied by a larger unlabeled set. This area thus studies the effective use of unlabeled data to reduce the performance gap that arises due to the lack of annotations. In this work, inspired by Bayesian deep learning, we first propose a Bayesian self-training framework for semi-supervised 3D semantic segmentation. Employing stochastic inference, we generate an initial set of pseudo-labels and then filter these based on estimated point-wise uncertainty. By constructing a heuristic \\( n \\)-partite matching algorithm, we extend the method to semi-supervised 3D instance segmentation, and finally, with the same building blocks, to dense 3D visual grounding. We demonstrate state-of-the-art results for our semi-supervised method on SemanticKITTI and ScribbleKITTI for 3D semantic segmentation and on ScanNet and S3DIS for 3D instance segmentation. We further achieve substantial improvements in dense 3D visual grounding over supervised-only baselines on ScanRefer. Our project page is available at ouenal.github.io/bst/." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.606, + 0.74, + 0.634 + ], + "angle": 0, + "content": "Keywords: Semi-supervised \\(\\cdot\\) Self-training \\(\\cdot\\) 3D semantic segmentation \\(\\cdot\\) 3D instance segmentation \\(\\cdot\\) 3D visual grounding" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.659, + 0.375, + 0.674 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Dense 3D semantic perception at the level of points is pivotal for several applications that are based on fine-grained processing and reasoning in 3D, e.g. autonomous cars, robotics, and augmented reality. Such perception requires a segmentation of the unstructured input 3D point cloud into different semantic classes, i.e. 3D semantic segmentation [60], or distinct instances of these classes, i.e., 3D instance segmentation [30]. The input point cloud is often coupled with human utterances that refer to individual objects in the former in cases where a visual AI system interacts with human users in a common 3D environment, necessitating a 3D instance-level segmentation of these objects based on the aforementioned referrals, a task which is known as dense 3D visual grounding or" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "O. Unal et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.148, + 0.784, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig. 1: Illustration of our Bayesian pseudo-labeling pipeline for semi-supervised a) 3D semantic segmentation; b) 3D instance segmentation; and c) dense 3D visual grounding. With only slight adjustments, using the same building blocks, our method can be adapted to each of these tasks to achieve SOTA results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.784, + 0.456 + ], + "angle": 0, + "content": "referral-based 3D instance segmentation [48, 59]. In the majority of works that address these diverse 3D segmentation tasks, it is assumed that the training data come with full 3D semantic and/or verbal annotations, which creates the pressing need for large-scale labeled 3D datasets [1, 2, 6, 11] so that large end-to-end models can be adequately fitted on these data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.746 + ], + "angle": 0, + "content": "However, the creation of such fine-level 3D annotations and verbal referrals for 3D scans requires human labor to be fairly reliable. Such labor is time-consuming and expensive, especially due to the high dimensionality of the scanned point clouds, which makes it hard for human annotators to navigate through them, visually inspect them, and verbally describe them. These practical difficulties render learning frameworks for 3D segmentation that operate with partial labels essential, semi-supervised learning being a primary case. Answering this need, several recent works address semi-supervised 3D semantic segmentation [9,25,29], most of which focus on effective pseudo-labeling techniques for self-training on the larger, unlabeled portion of the training data. However, the generation of pseudo-labels in these works typically focuses on some grouping mechanism over 3D points, such as superpoints, to refine and propagate pseudo-labels and otherwise relies on naïve core techniques such as confidence thresholding [66] for generating the initial point-level pseudo-labels that are then post-processed by their proposed modules. Adding to this, there is a relative scarcity in works on semi-supervised 3D instance segmentation [10, 33], which we attribute to the relative difficulty in pseudo-labeling instances compared to creating plain semantic pseudo-labels. Last but not least, to our knowledge, no work has considered the setting of semi-supervised dense 3D visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We jointly address these shortcomings of previous works by proposing a common self-training framework for semi-supervised 3D segmentation which is based on Bayesian deep learning and applies to all three aforementioned segmentation tasks. At the core of our self-training framework lies a variational Bayesian approximation of the uncertainty of the respective network's predictions via dropout-based Monte Carlo integration [15] over multiple augmented versions of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "the original 3D input. This approximation delivers a more informed uncertainty estimate than simple confidence thresholding techniques [66], which is crucial for eliminating confirmation bias from the pseudo-labeling process that we adopt for self-training. In particular, we utilize both the estimated Shannon entropy from the sampled predictions and the predicted label from each sample to filter out uncertain or inconsistent points from the generated set of pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.486 + ], + "angle": 0, + "content": "While this formulation is sufficient for the task of 3D semantic segmentation, it does not work out of the box for instance-level segmentation, as Monte Carlo integration cannot be applied directly to instance predictions. To account for this issue, we propose a novel and efficient, heuristic matching strategy across instances coming from different stochastic dropout forward passes based on linear sum assignment. This strategy enables the application of the same filtering steps as described above for semantic pseudo-labels to instance pseudo-labels and thus extends the scope of our approach to 3D instance segmentation. Going one step further, we observe that the same matching strategy that we devise for instance segmentation applies to the selection stage of dense 3D grounding, in which the correct instance from the set of segmented candidate instances needs to be selected based on the input verbal description. Thus, for dense 3D grounding, we apply our heuristic linear sum assignment twice, first for the purely 3D-based candidate instance proposal generation and afterward for the verbo-visual instance selection. Our complete self-training pipeline for each of the three 3D segmentation tasks to which our method applies is summarized in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.6 + ], + "angle": 0, + "content": "We test our method on leading benchmarks across a wide range of domains and datasets. Specifically, our Bayesian self-training framework achieves state-of-the-art results for semi-supervised 3D semantic segmentation on SemanticKITTI [3] and ScribbleKITTI [47] and semi-supervised 3D instance segmentation on ScanNet [11] and S3DIS [2]. Furthermore, we are the first to explore a semi-supervised setting on ScanRefer [6] for 3D visual grounding and obtain consistent significant improvements over supervised-only baselines." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.605, + 0.544, + 0.619 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.655, + 0.784, + 0.716 + ], + "angle": 0, + "content": "1. Following established uncertainty estimation techniques from active learning, we construct a Bayesian self-training framework for semi-supervised 3D semantic segmentation by generating reliable pseudo-labels through stochastic inference and entropy-based uncertainty filtering." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.722, + 0.785, + 0.751 + ], + "angle": 0, + "content": "2. We introduce a novel heuristic matching algorithm to extend the proposed framework to 3D instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.758, + 0.784, + 0.803 + ], + "angle": 0, + "content": "3. We further extend the method to dense 3D visual grounding utilizing the same building blocks as in the above tasks and are the first to explore this task in a semi-supervised setting." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "4. We obtain state-of-the-art results for semi-supervised 3D segmentation on SemanticKITTI [3], ScribbleKITTI [47], ScanNet [11] and S3DIS [2]." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.655, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "O. Unal et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.537 + ], + "angle": 0, + "content": "Semi-supervised 3D semantic segmentation has recently grown in popularity in the area of 3D understanding along with weakly supervised 3D semantic segmentation [47,52,57,61], due to the difficulty in large-scale 3D labeling that is necessary for the supervised setup [60]. In this work, we tackle the task of semi-supervised segmentation where a large portion of data is completely unlabeled. As opposed to methods developed for weakly-supervised settings, a semi-supervised method cannot leverage the availability of partial true label information within each data sample, e.g. via the ArcPoint loss [31] or a PLS descriptor [47]. A common technique for leveraging unlabeled data under this restrictive scheme is explored under pseudo-labeling. Semantic pseudo-labels are utilized by Jiang et al. [25] to guide a contrastive loss and in SSPC [9] to propagate labels on a superpoint graph. Superpoints are also leveraged by Deng et al. [13] to constrain pseudo-label propagation. In the 2D counterpart task, Chen et al. [8] enforces consistency of semantic pseudo-labels across perturbed versions of the 2D segmentation network. By contrast, we perturb both the network weights and the 3D inputs across different forward passes, and we consider both the entropy and the consistency among predictions in our pseudo-labels. Another related line of 2D semi-supervised works generates pseudo-labels based on direct thresholding of max-softmax scores to reduce confirmation bias [19, 66], while we threshold the entropy of the distribution of softmax scores. Active learning is combined with a 3D semi-supervised approach in DiAL [46] to streamline practical 3D segmentation annotation pipelines. LaserMix [29] introduces variability through cylindrical and range-view partitioning and mixing of 3D LiDAR scans, encouraging consistency across different mixes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Semi-supervised 3D instance segmentation is still in its infancy compared to its instance-agnostic semantic segmentation counterpart reviewed above. In particular, most object-level semi-supervised 3D understanding works focus on the coarser task of 3D object detection [17,35,37,38,42,50,51,63], which outputs 3D bounding boxes that are not fine enough for precision-critical downstream tasks such as robotic grasping. For 3D instance/object segmentation, approaches that depart from the standard fully supervised setup have mostly considered the cases of unsupervised pre-training [56], weak supervision via one-point-one-instance [20, 43], or unsupervised generic 3D object segmentation [32]. To our knowledge, the only two semi-supervised works that specifically consider 3D instance segmentation are TWIST [10] and WS3D [33]. The former work proposes a pseudo-label generation strategy tailored for the instance segmentation task, for which the technique of confidence thresholding that is ubiquitously used to the same end in semantic segmentation is not applicable. It thus introduces pseudo offset vectors, besides semantic proposal-level pseudo-labels, for performing self-training. The latter work proposes a region-level contrastive loss, which performs unsupervised instance discrimination, and leverages boundary information via a region-level unsupervised energy-based loss. On the contrary, we propose a unified semi-supervised Bayesian self-training framework that spans multiple 3D segmentation tasks including 3D instance segmentation, and han" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "dles the additional matching required between instances generated in different forward passes through the network via an efficient heuristic optimization step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.182, + 0.788, + 0.349 + ], + "angle": 0, + "content": "3D visual grounding is a central 3D task in the area of vision and language [4, 14, 28, 34, 36, 62]. Relevant models accept a 3D point cloud as visual input and perform grounding of the accompanying verbal description to the referred object. While most previous works [7, 23, 24, 34, 39] focus on coarse 3D bounding box outputs, the task of dense 3D visual grounding or referral-based 3D instance segmentation has been explored far less in comparison, in which a fine-grained point-level mask is predicted [22, 48, 59]. Regardless, all aforementioned methods invariably assume a supervised setting, where full ground-truth instance annotations are available for all samples at training time. By contrast, the formulation of our method makes it fit for also handling the more practical semi-supervised setting, in which only a part of the training data is labeled." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Uncertainty estimation is crucial in learning setups that lack full supervision as ours, given that such estimates can provide information about the reliability of the predictions of the model on unlabeled samples [21]. In the more recent deep learning era, Bayesian neural networks form the core for uncertainty estimation by modeling network weights as stochastic distributions, where test-time-dropout [41] is utilized as a variational Bayesian approximation [15] and Monte Carlo integration is computed via multiple stochastic forward passes to approximate prediction uncertainty [16]. This Bayesian deep learning framework has been applied successfully to supervised 2D dense prediction tasks [26, 27]. We focus on the aggregate uncertainty in the predictions of a general 3D segmentation model and build on the Bayesian deep learning approach of Gal et al. [16] which examines the approximate predictive entropy as an acquisition function for active learning. However, contrary to Gal et al. [16], in our self-training setup, we use low rather than high entropy as a positive indication for considering a point as a source of supervision. To the best of our knowledge, our work is the first to introduce such a Bayesian self-training framework for general 3D point cloud segmentation, specifically including 3D instance segmentation and referral-based 3D instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Related uncertainty-based approaches to ours have been proposed for the semantic segmentation of 3D grid volume data in the field of medical image computing [12,54,58]. In particular, Yu et al. [58] also use Monte Carlo dropout to estimate predictive entropy but only applies dropout to a single network input, whereas we vary both network weights and input views. Moreover, that work leverages the estimated entropy merely as the weight in a consistency loss between a teacher and a student network, while we filter pseudo-labels based on our entropy estimates. Besides, Xia et al. [54] perform semi-supervised semantic segmentation with multi-view co-training and applies Bayesian deep learning across views for uncertainty-based weighting of pseudo-labels. We only train one network instead of multiple networks employed in Xia et al. [54] and we derive the uncertainty estimates by feeding all input views to that single network, which is more parameter-efficient. Both of these approaches only address 3D semantic segmentation, and cannot be utilized in instance-level dense prediction tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "O. Unal et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.146, + 0.776, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.397 + ], + "angle": 0, + "content": "Fig. 2: Illustration of Bayesian pseudo-labeling for semi-supervised instance segmentation. We initialize a seed prediction via a forward pass using a non-augmented input. Then our heuristic \\( n \\)-partite matching algorithm is employed to pair each seed instance with the best matching predicted instance from each of the \\( K \\) stochastic forward passes. For each aligned object, we compute the aggregated label through unanimous voting and filter it based on uncertainty to obtain the final pseudo mask." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.412, + 0.471, + 0.431 + ], + "angle": 0, + "content": "3 Bayesian Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.537 + ], + "angle": 0, + "content": "In Sec. 3.1 we start by formulating a self-training approach for semi-supervised 3D semantic segmentation based on established uncertainty estimation techniques from the field of active learning [15,46]. We then extend our formulation to include instance labels in Sec. 3.2 and finally build up to also take natural language prompts as input, i.e., towards semi-supervised dense 3D visual grounding in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.564, + 0.628, + 0.58 + ], + "angle": 0, + "content": "3.1 Semi-Supervised 3D Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.669 + ], + "angle": 0, + "content": "In a semi-supervised setting, a dataset \\( D \\) consists of a labeled set \\( S \\) and an unlabeled set \\( U \\). The baseline approach when working with such data is to treat the problem similarly to any fully supervised task and employ a semantic loss \\( \\mathcal{L}_{sem} \\) on the available labeled scenes \\( S \\), typically stated as cross-entropy. Formally, the objective function over the model parameters \\( \\omega \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.679, + 0.787, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\omega} \\sum_ {i = 1} ^ {| S |} \\mathcal {L} _ {s e m} (\\hat {\\mathbf {y}}, \\mathbf {y}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Starting from this baseline formulation, a semi-supervised training strategy aims to leverage the unlabeled scenes \\(U\\) to further improve model performance. One such common technique is self-training, where the model is iteratively trained on its own predictions, i.e. on generated per-point one-hot pseudo-label vectors \\(\\hat{\\mathbf{y}} = [\\hat{\\mathbf{y}}^{(1)},\\dots,\\hat{\\mathbf{y}}^{(C)}]\\), with \\(C\\) denoting the number of classes. However, the task of generating pseudo-labels is not trivial, as the model may suffer from confirmation bias that occurs when repeatedly overfitting to incorrect labels." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "A simple but effective solution is explored under the umbrella of thresholding [5, 19, 55, 65, 66]. In essence, instead of training on all model predictions, a pseudo-label set \\( L \\) is formed by predictions that exceed a confidence of a determined threshold \\( \\tau \\) to ensure a higher level of accuracy. Formally, the objective function of threshold-based self-training for semantic segmentation can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.242, + 0.786, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\omega , \\hat {\\mathbf {y}}} \\sum_ {d = 1} ^ {| D |} \\left(\\mathcal {L} _ {s e m} (\\hat {\\mathbf {y}}, \\mathbf {y}) - (\\log \\hat {\\mathbf {y}} _ {d} - \\tau) \\cdot \\hat {\\mathbf {y}} _ {d}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.29, + 0.789, + 0.32 + ], + "angle": 0, + "content": "with \\(\\hat{\\mathbf{y}}_d\\) denoting the softmax model predictions. Eq. 2 is optimized in two steps that are repeated until convergence:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.324, + 0.787, + 0.354 + ], + "angle": 0, + "content": "1. Training: First, \\(\\hat{\\mathbf{y}}\\) is fixed, and the objective function is optimized with respect to \\(\\omega\\)." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.354, + 0.787, + 0.385 + ], + "angle": 0, + "content": "2. Pseudo-labeling: Then, \\(\\omega\\) is fixed and the objective function is optimized with respect to \\(\\hat{\\mathbf{y}}\\). Consequently, \\(L\\) is updated given \\(\\hat{\\mathbf{y}}\\)." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.324, + 0.787, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.389, + 0.787, + 0.45 + ], + "angle": 0, + "content": "The process is initialized by setting the latent variable \\(\\hat{\\mathbf{y}} = \\mathbf{0}\\) for all points, i.e. by reverting to the naive solution of Eq. 1 and only training on the labeled set \\(S\\). The two steps can be repeated to take advantage of the improved representation capability of the model thanks to pseudo-labeling." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.45, + 0.788, + 0.541 + ], + "angle": 0, + "content": "While naïve thresholding relies on the softmax scores of the model to determine confidence, these scores may be unreliable and often skewed when facing long-tailed distributions, resulting in overconfident predictions [18]. Inspired by uncertainty estimation techniques commonly employed in active learning (AL) pipelines [21,46], we introduce a Bayesian self-training strategy that leverages an uncertainty-aware filtering technique for generating more reliable pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.541, + 0.788, + 0.616 + ], + "angle": 0, + "content": "We first start by decomposing the goal of 3D semantic segmentation, which is to discover the dependency of the point-wise distribution over the labels \\( y \\in Y \\) on an input sample \\( x \\in X \\) via the model weights \\( \\omega \\). Formally, the task can be reduced to per-point classification, where \\( y \\) can be of class \\( c \\in \\{1, \\dots, C\\} \\). The conditional probability is therefore defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.622, + 0.787, + 0.653 + ], + "angle": 0, + "content": "\\[\np (y = c | x, D) = \\int p (y = c | a (x), \\omega) p (\\omega | D) p (a (x) | x) d \\omega d x \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.658, + 0.787, + 0.688 + ], + "angle": 0, + "content": "for the dataset \\( D = (X, Y) \\), under randomly applied affine transformations modeled through \\( a(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.688, + 0.788, + 0.764 + ], + "angle": 0, + "content": "We set \\( p(a(x)|x) \\) as a constant, restate the intractable \\( p(\\omega |D) \\) via the variational approximation \\( q(\\omega) \\) and minimize \\( \\mathrm{KL}(q(\\omega)|p(\\omega |X,Y)) \\). This objective can be approximated through Monte Carlo integration by employing variational inference via Bayesian networks, i.e. by applying stochastic forward passes with weights \\( \\omega_{k} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.77, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p (y = c | x, D) \\approx \\int p (y = c | a (x), \\omega) q (\\omega) d \\omega d x \\\\ \\approx \\frac {1}{K} \\sum_ {k = 1} ^ {K} p (y = c | a (x), \\omega_ {k}). \\tag {4} \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "O. Unal et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "Compared to the common decomposition of BALD [21], the inclusion of the affine transformations via \\( a(\\cdot) \\) allows a higher deviation amongst stochastic forward passes which in return allows better estimation of predictive uncertainty [49]. We capture the uncertainty via the Shannon entropy [40]:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.22, + 0.786, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\mathbb {H} [ y | x, D ] = - \\sum_ {c = 1} ^ {C} p (y = c | x, D) \\log p (y = c | x, D) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.272, + 0.786, + 0.301 + ], + "angle": 0, + "content": "and the accumulated predicted label \\(\\hat{\\mathbf{y}}^*\\) is given by the unanimous voting principle:" + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.302, + 0.786, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\hat {y} ^ {* (c)} = \\left\\{ \\begin{array}{l} 1, \\text {i f} K = \\sum_ {k} \\mathbb {1} (\\operatorname {a r g m a x} (p (y | a (x), \\omega_ {k})) = c) \\\\ 0, \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.346, + 0.499, + 0.361 + ], + "angle": 0, + "content": "with \\(\\mathbb{1}\\) denoting the indicator function." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.362, + 0.787, + 0.422 + ], + "angle": 0, + "content": "Unlike what is commonly seen in active learning settings, our goal is not to maximize the total entropy but rather to minimize it in order to ensure a high-quality pseudo-label set. We therefore redefine the self-training objective (Eq. 2) for stochastic pseudo-labeling as:" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.434, + 0.786, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\omega , \\hat {\\mathbf {y}}} \\sum_ {d = 1} ^ {D} \\left(\\mathcal {L} _ {s e m} (\\hat {\\mathbf {y}}, \\mathbf {y}) + \\left(\\mathbb {H} [ y | x, D ] - \\tau\\right) \\left(\\hat {\\mathbf {y}} ^ {*} \\cdot \\hat {\\mathbf {y}}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.485, + 0.786, + 0.517 + ], + "angle": 0, + "content": "with \\(\\tau \\in (0,1)\\) determining the cut-off entropy threshold. To solve the nonlinear integer optimization task, we employ the following solver:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.528, + 0.786, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {y}} ^ {(c)} = \\left\\{ \\begin{array}{l} 1, \\text {i f} \\mathbb {H} [ y | x, D ] < \\tau \\text {a n d} \\\\ c = \\operatorname {a r g m a x} \\hat {\\mathcal {y}} ^ {*} \\\\ 0, \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.595, + 0.599, + 0.611 + ], + "angle": 0, + "content": "where labels \\(\\hat{\\mathbf{y}}^{(c)} = 0\\forall c\\) are ignored during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.786, + 0.656 + ], + "angle": 0, + "content": "In practice, following precedent [47,66], we set \\(\\tau\\) based on \\(p_{\\tau}\\), with \\(p_{\\tau}\\) determining the percentage of pseudo-labels to be sampled amongst the accumulated predicted labels." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.657, + 0.787, + 0.747 + ], + "angle": 0, + "content": "In summary, we first determine a set of labels that are agreed upon by all stochastic forward passes, i.e. by forward passes with active dropout [15, 41]. We then compute the Shannon entropy for all predictions, filtering those that exhibit high entropy. This surprisingly simple but effective approach allows us to reduce the incorrect targets that we introduce to our pseudo-label set and to avoid the overconfidence issue of naive thresholding-based approaches." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.622, + 0.787 + ], + "angle": 0, + "content": "3.2 Semi-Supervised 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The proposed Bayesian self-training framework provides a strong basis for any semi-supervised 3D dense prediction task, one such example being 3D instance segmentation. Yet, compared to semantic segmentation where accumulation of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.227, + 0.15, + 0.739, + 0.179 + ], + "angle": 0, + "content": "Algorithm 1: Bayesian pseudo-label generation for instance segmentation." + }, + { + "type": "algorithm", + "bbox": [ + 0.218, + 0.18, + 0.761, + 0.421 + ], + "angle": 0, + "content": "Input: Point cloud \\(P\\in \\mathbb{R}^{N\\times 3}\\) , trained model with dropout \\(\\omega\\) , number of passes \\(K\\) , threshold \\(\\tau\\) \n1 seed_output \\(= \\omega (P)\\in \\mathbb{R}^{M\\times N}\\) \n2 \\(\\pmb{i} =\\) seed_output.copy() \n3 seed_mask \\(=\\) to_mask(i) \\(\\in \\{0,1\\}^{M\\times N}\\) \n4 unanimous \\(=\\) seed_mask.copy() \n5 for k in 1:K do \n6 output \\(= \\omega_{k}(a(P))\\in \\mathbb{R}^{M^{\\prime}\\times N}\\) \n7 mask \\(=\\) to_mask(output) \\(\\in \\{0,1\\}^{M^{\\prime}\\times N}\\) \n8 cost \\(=\\) -iou(seed_mask, mask) \n9 row,col \\(=\\) LSA(cost) \n10 unanimous[col] &= mask[row] \n11 i[col] \\(+ =\\) output[row] \n12 \\(\\mathbb{H} = (K - i)\\cdot \\log (K - i) + i\\log i\\) \n13 pseudo_label \\(= (\\mathbb{H} < \\tau)\\) & unanimous \nReturn:pseudo_label \\(\\in \\{0,1\\}^{M\\times N}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.785, + 0.484 + ], + "angle": 0, + "content": "labels can be easily handled across multiple stochastic forward passes, accumulation of predicted instance masks is not trivial and requires \\( n \\)-partite matching." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.787, + 0.621 + ], + "angle": 0, + "content": "However, the \\(n\\)-partite matching problem is NP-hard with a complexity of \\(|I|!(K-1)\\) with \\(K\\) passes and \\(|I|\\) number of instances assuming full bijection. To match predicted instance masks across various stochastic forward passes, we propose a heuristic method based on the Hungarian algorithm. First, we generate an initial prediction via an unaugmented forward pass (Alg. 1 L1&3). The goal is to establish a seed that is likely of a higher quality with minimal perturbations of the data. We then iteratively assign predicted masks of each stochastic forward pass via linear sum assignment, with the cost matrix formed by the negative pairwise IoUs (Alg. 1 L5-10)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.787, + 0.697 + ], + "angle": 0, + "content": "Having matched predicted binary instance masks from individual stochastic forward passes, each resulting set of masks can then be treated as a set of 2-class semantic segmentation outputs, and thus our proposed Bayesian pseudo-labeling approach from Eq. 8 can be trivially applied (Alg. 1 L11-15). Our pipeline for semi-supervised 3D instance segmentation is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.722, + 0.636, + 0.737 + ], + "angle": 0, + "content": "3.3 Semi-Supervised Dense 3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Similar to both semi-supervised 3D semantic and 3D instance segmentation, our proposed Bayesian self-training formulation can easily be extended to include natural language prompts as inputs and thus tackle semi-supervised dense 3D visual grounding, i.e., referral-based 3D instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "A common technique when tackling dense 3D visual grounding is to employ a grounding-by-selection strategy which can be summarised in two steps [6,48,62]:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "O. Unal et al." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.147, + 0.784, + 0.161 + ], + "angle": 0, + "content": "1. Visual: First, a 3D backbone computes object candidates \\( O \\) from the scene." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.162, + 0.785, + 0.207 + ], + "angle": 0, + "content": "2. Verbo-visual: Then, a verbo-visual fusion module \\(\\omega_{sel}\\) selects the correct instance candidate from predicted instances \\(I\\) based on the natural language description \\(W\\), i.e., it predicts the index:" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.217, + 0.785, + 0.234 + ], + "angle": 0, + "content": "\\[\ni ^ {*} = \\operatorname {a r g m a x} G, \\text {w i t h} G = \\omega_ {s e l} (O, W) \\in \\mathbb {R} ^ {| I |}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.784, + 0.289 + ], + "angle": 0, + "content": "To extend our Bayesian self-training approach to this task, we follow the above structure and individually adapt each of the two steps, with the adaptation of the visual step simply following Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.785, + 0.395 + ], + "angle": 0, + "content": "For the verbal side of 3D visual grounding, i.e. referral-based instance selection, we observe that the pseudo-label generation can be reduced to the same formulation as the pseudo-label generation for 3D instance segmentation. Utilizing the results from the heuristic \\( n \\)-partite matching (Alg. 1 L9), vector \\( G \\) can be reordered to match across stochastic forward passes. Again, having matched predicted probability vectors, the method presented in Sec. 3.1 can be trivially applied (Alg. 1 L11-15)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.417, + 0.376, + 0.433 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.784, + 0.508 + ], + "angle": 0, + "content": "We evaluate our method on all three semi-supervised 3D semantic perception tasks: 3D semantic segmentation, 3D instance segmentation, and dense 3D visual grounding. The dataset and implementation details regarding each task can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.53, + 0.486, + 0.545 + ], + "angle": 0, + "content": "4.1 3D Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.785, + 0.659 + ], + "angle": 0, + "content": "The comparison to the state of the art for semi-supervised 3D semantic segmentation is shown in Tab. 1. As seen, our Bayesian self-training strategy outperforms existing work in both SemanticKITTI [3] and ScribbleKITTI [47] on splits of \\(10\\%\\), \\(20\\%\\) and \\(50\\%\\), while performing on par for \\(1\\%\\). Specifically, compared to the baseline Cylinder3D [64] model denoted as Sup-only, our strategy allows significant improvements ranging from \\(4.4\\%\\) to \\(10.9\\%\\) mIoU. In Fig. 3 (a-c) we showcase qualitative results on SemanticKITTI \\(10\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.48, + 0.696 + ], + "angle": 0, + "content": "4.2 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.84 + ], + "angle": 0, + "content": "For semi-supervised 3D instance segmentation, we report results for ScanNet [11] in Tab. 2 and for S3DIS [2] in Tab. 3. As seen, we observe a similar behavior to semi-supervised 3D semantic segmentation, where at higher thresholds of \\(10\\%\\), \\(20\\%\\) for ScanNet and \\(20\\%\\), \\(50\\%\\) for S3DIS, our simple yet effective Bayesian self-training strategy significantly outperforms state-of-the-art methods that require highly engineered solutions. In contrast, at lower thresholds, we see a significant gap in performance (especially for ScanNet). Still, compared to the baseline Sup-only, we see significant gains across the board for all semi-splits. Additionally in Fig. 3 (d-f), we show qualitative results on the ScanNet val-set." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: Comparison of state-of-the-art methods for semi-supervised LiDAR semantic segmentation (mIoU [%]). For fair comparison, all methods use the same backbone given by Sup-only, which at \\(100\\%\\) labels achieves \\(64.3\\%\\) mIoU on SemanticKITTI and \\(57.0\\%\\) mIoU on ScribbleKITTI." + }, + { + "type": "table", + "bbox": [ + 0.291, + 0.207, + 0.703, + 0.449 + ], + "angle": 0, + "content": "
Labeled1%10%20%50%
SemantickKITTI [3]Sup-only [64]45.456.157.858.7
DiAL [46]45.457.159.260.0
CBST [66]48.858.359.459.7
CPS [8]46.758.759.660.5
LaserMix [29]50.660.061.962.3
IGNet [45]49.061.363.164.8
Ours49.861.763.764.1
Δ Sup-only+4.4+5.6+5.9+5.4
ScribbleKITTI [47]Sup-only [64]39.248.052.153.8
DiAL [46]41.050.152.853.9
CBST [66]41.550.653.354.5
CPS [8]41.451.853.954.8
LaserMix [29]44.253.755.156.8
IGNet [45]44.457.759.660.8
Ours43.958.960.161.2
Δ Sup-only+4.7+10.9+8.0+7.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Table 2: Comparison to state-of-the-art methods for semi-supervised 3D instance segmentation on the ScanNet Limited Reconstruction valid-set. For a fair comparison, all methods apart from that marked with * use the same backbone [20,53] as Sup-only." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.505, + 0.782, + 0.637 + ], + "angle": 0, + "content": "
Labeled Metric1% mAPAP50AP255% mAPAP50AP2510% mAPAP50AP2520% mAPAP50AP25
Sup-only [10]5.19.817.618.232.047.026.742.858.929.347.963.0
PointCont. [56]7.212.520.319.435.448.527.043.959.530.249.563.6
CSC [20]7.113.021.220.936.750.627.345.060.230.650.364.1
TWIST [10]9.617.126.227.044.156.230.649.763.032.852.966.8
WS3D* [33]-32.5--45.6--49.2--51.3-
Ours7.214.223.024.239.351.432.751.965.537.456.168.9
Δ Sup-only+2.1+4.4+5.4+6.0+7.3+4.4+6.0+9.1+6.6+8.1+8.2+5.9
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.497, + 0.663 + ], + "angle": 0, + "content": "4.3 Dense 3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.787, + 0.719 + ], + "angle": 0, + "content": "For semi-supervised 3D visual grounding, we report results on the ScanRefer [6] dataset. As seen in Tab. 4 (Vi), our self-training approach shows improvements across the board, with significant gains on the \"unique\" category." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.809 + ], + "angle": 0, + "content": "While natural language prompts still require human effort to generate, they are considerably cheaper and easier to acquire than per-point high-quality panoptic labels. In Tab. 4 (ViVe), we show that the inclusion of such prompts in the unlabeled dataset further benefits the model's accuracy. As seen, we observe even larger gains over the baseline method with the overall accuracy (OA) showing \\(+4.7\\%\\) improvement with a \\(50\\%\\) IoU threshold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Furthermore, in Fig. 3 (g) we qualitatively compare our method to the baseline Sup-only on ScanRefer." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "O. Unal et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.483, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.147, + 0.631, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.236, + 0.631, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.149, + 0.78, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.237, + 0.78, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.305, + 0.776, + 0.318 + ], + "angle": 0, + "content": "The table is square. There is a chair to the left, and a chair to the right of it." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Fig.3: Qualitative results from the SemanticKITTI (10%) val-set, comparing a) the ground truth, b) Sup-only and c) ours; on ScanNet (10%), comparing d) the ground-truth instance masks, e) Sup-only and f) ours; and finally g) ScanRefer (10%)." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.391, + 0.789, + 0.434 + ], + "angle": 0, + "content": "Table 3: Comparison to state-of-the-art methods for semi-supervised 3D instance segmentation on S3DIS Area 5 via AP50 [%]. For a fair comparison, all methods use the same backbone Sparse U-Net [20,53] as Sup-only." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.44, + 0.706, + 0.542 + ], + "angle": 0, + "content": "
Labeled5%10%20%50%
Sup-only [10]30.436.841.246.5
PointContrast [56]33.638.743.148.9
CSC [20]34.241.044.750.4
TWIST [10]37.145.648.4-
Ours35.042.951.156.6
Δ Sup-only+4.6+6.1+9.9+10.1
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.562, + 0.402, + 0.576 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.787, + 0.642 + ], + "angle": 0, + "content": "We conduct ablation studies on \\(10\\%\\) splits of SemanticKITTI [3], ScanNet [11] and ScanRefer [6] for semi-supervised 3D semantic segmentation, instance segmentation and visual grounding respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Entropy Thresholding: In Tab. 5, we show an analysis of the effectiveness of our proposed uncertainty thresholding via Shannon entropy, comparing it to the baseline unanimous voting results as well as established thresholding techniques of softmax thresholding [55] and class-balanced thresholding [66]. Alongside the performance for each individual task, we also report the accuracy of the generated pseudo-labels. For instance segmentation, a pseudo-mask is considered correct if its intersection with the associated ground-truth mask is identical to the former. For visual grounding, we reduce the accuracy threshold to 0.5 following the benchmark metric used. As seen, our uncertainty-based filtering improves pseudo-label quality across the board. We also showcase the performance implications of uncertainty filtering, where it can be observed that entropy thresholding outperforms both unanimous voting and current thresholding strategies across all tasks. We further provide a qualitative analysis of the uncertainty met-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: Results for semi-supervised 3D dense visual grounding while assuming (Vi) that the model is given no natural language descriptions for the unlabeled split and (ViVe) when the model has access to natural language descriptions for the unlabeled split. Reported values are given as Acc@25/50 for the \\(10\\%\\) split." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.207, + 0.694, + 0.294 + ], + "angle": 0, + "content": "
CategoryUniqueMultipleOverall
Sup-only [48]51.9 / 43.619.3 / 14.125.4 / 19.6
ViMe ViOurs58.8 / 48.822.0 / 18.228.8 / 23.9
Δ Sup-only+6.9 / +5.2+2.7 / +4.1+3.4 / +4.3
Ours61.0 / 51.223.4 / 18.130.4 / 24.3
Δ Sup-only+9.1 / +7.6+4.1 / +4.0+5.0 / +4.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.303, + 0.784, + 0.372 + ], + "angle": 0, + "content": "Table 5: Comparison of the pseudo-label quality of current thresholding strategies and their effects on performance on SemanticKITTI \\(10\\%\\), ScanNet \\(10\\%\\), ScanRefer \\(10\\%\\) (with visual and verbal) for each respective task. Pseudo-label thresholding is applied after unanimous voting for a fair comparison. The letters indicate each pseudolabel type: \\(S-\\) for semantic, \\(I-\\) for instance, and \\(G-\\) for grounding." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.379, + 0.782, + 0.496 + ], + "angle": 0, + "content": "
SemanticInstanceVisual Grounding
MethodS-AccmIoUS-AccI-AccAP@50S-AccI-AccG-AccOA@50
Sup-only [48]-56.1--42.8---19.6
Unanimous96.460.079.375.347.072.969.961.516.9
Naïve [5,55,65]98.360.882.977.750.282.073.068.621.9
Class-balanced [66]96.661.181.270.144.480.969.060.616.2
Entropy-based Δ Unanimous98.661.784.877.851.983.875.275.624.3
+2.2+1.7+5.5+2.5+4.9+5.8+2.3+5.7+7.4
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.512, + 0.787, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.593, + 0.784, + 0.622 + ], + "angle": 0, + "content": "Fig. 4: Qualitative analysis of the uncertainty estimation. We compare softmax confidence to our MC-derived Shannon entropy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.784, + 0.671 + ], + "angle": 0, + "content": "ric in Fig. 4 to illustrate the better correlation of our MC-derived uncertainty with pseudo-label accuracy compared to softmax based uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.785, + 0.717 + ], + "angle": 0, + "content": "Labeling Percentage \\((p_{\\tau})\\): We compare various set labeling percentages \\(p_{\\tau}\\) across all tasks in Tab. 6. While we obtain higher accuracy pseudo-labels at \\(50\\%\\), it can be seen that the reduced supervision negatively impacts performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.744, + 0.514, + 0.76 + ], + "angle": 0, + "content": "5 Limitations and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The main advantage of the proposed Bayesian self-training is its simplicity and ease of implementation compared to existing SOTA approaches that utilize highly engineered and often costly modules. However, the method has a limitation. As observed from the results, the performance is comparably inferior to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "O. Unal et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 6: Ablation study analyzing the effect of the hyperparameter \\( p_{\\tau} \\) for each task. We vary the labeling percentage and show the performance on SemanticKITTI 10%, ScanNet 10%, and ScanRefer 10% (incl. verbal input) for each respective task." + }, + { + "type": "table", + "bbox": [ + 0.31, + 0.193, + 0.684, + 0.253 + ], + "angle": 0, + "content": "
ThresholdmIoUAP@50OA@50
50%61.450.022.8
75%61.751.924.3
100%60.047.016.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.259, + 0.785, + 0.302 + ], + "angle": 0, + "content": "Table 7: Ablation study showing the performance on SemanticKITTI \\(10\\%\\), ScanNet \\(10\\%\\) and ScanRefer \\(10\\%\\) (incl. verbal input) for varying label aggregation strategies from majority rules to unanimous voting over 9 stochastic forward passes." + }, + { + "type": "table", + "bbox": [ + 0.31, + 0.308, + 0.684, + 0.368 + ], + "angle": 0, + "content": "
ThresholdmIoUAP@50OA@50
5/960.747.920.1
7/961.051.023.9
9/961.751.924.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.785, + 0.488 + ], + "angle": 0, + "content": "SOTA semi-supervised approaches for highly sparse training splits. This behavior stems from the reliance on unanimous voting (Eq. 6). With low amounts of labeled data, the initial naive model performance suffers, causing a reduction in the number of unanimously voted points. This effect is further pronounced when tackling instance segmentation, where fused instance masks may affect multiple predictions. In Tab. 7, we show that relaxation does not aid the performance as it further introduces incorrect predictions into the pseudo-label set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.787, + 0.596 + ], + "angle": 0, + "content": "While semi-supervised training methods that enhance the robustness of the pseudo-labels either via a teacher model (DiAL [44, 46]), mix-augmentations (LaserMix [29]) or multi-stage refinement (TWIST [10]) are presented as competing to ours, we argue that as a solution, they can in fact be utilized within our framework in place of the 3D model to improve the quality of the initial predictions. This would allow a larger set of unanimously voted samples and would thus re-enable the benefits of our proposed method on highly sparse splits." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.36, + 0.634 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.801 + ], + "angle": 0, + "content": "In this work, we introduce a simple yet effective Bayesian self-training approach for semi-supervised 3D semantic segmentation by utilizing uncertainty estimation strategies from the field of active learning. Specifically, through multiple stochastic forward passes, we generate an initial pseudo-label set and then filter it based on estimated predictive entropy. We further extend our semi-supervised 3D method to newly include both instance segmentation and referral-based instance segmentation via an efficient, heuristic \\( n \\)-partite matching algorithm. We conduct extensive experiments on multiple datasets, covering both indoor and outdoor domains, showing that our method performs on par if not better than current domain- and task-specific methods." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.808, + 0.785, + 0.838 + ], + "angle": 0, + "content": "Acknowledgments: This work is funded by Toyota Motor Europe via the research project TRACE-Zürich." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: ReferIt3D: Neural listeners for fine-grained 3D object identification in real-world scenes. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.263 + ], + "angle": 0, + "content": "2. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3D semantic parsing of large-scale indoor spaces. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.32 + ], + "angle": 0, + "content": "3. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.785, + 0.349 + ], + "angle": 0, + "content": "4. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3DJCG: A unified framework for joint dense captioning and visual grounding on 3D point clouds. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Cascante-Bonilla, P., Tan, F., Qi, Y., Ordonez, V.: Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 6912-6920 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.418 + ], + "angle": 0, + "content": "6. Chen, D.Z., Chang, A.X., Nießner, M.: ScanRefer: 3D object localization in RGB-D scans using natural language. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "7. Chen, J., Luo, W., Wei, X., Ma, L., Zhang, W.: HAM: Hierarchical attention model with high performance for 3D visual grounding. arXiv preprint arXiv:2210.12513 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.503 + ], + "angle": 0, + "content": "8. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.544 + ], + "angle": 0, + "content": "9. Cheng, M., Hui, L., Xie, J., Yang, J.: SSPC-Net: Semi-supervised semantic 3D point cloud segmentation network. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.602 + ], + "angle": 0, + "content": "0. Chu, R., Ye, X., Liu, Z., Tan, X., Qi, X., Fu, C.W., Jia, J.: TWIST: Two-way inter-label self-training for semi-supervised 3D instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.629 + ], + "angle": 0, + "content": "1. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.685 + ], + "angle": 0, + "content": "2. Dai, W., Li, X., Cheng, K.T.: Semi-supervised deep regression with uncertainty consistency and variational model ensembling via bayesian neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7304-7313 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.728 + ], + "angle": 0, + "content": "3. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3D point clouds. In: International Conference on Robotics and Automation (ICRA) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.77 + ], + "angle": 0, + "content": "4. Feng, M., Li, Z., Li, Q., Zhang, L., Zhang, X., Zhu, G., Zhang, H., Wang, Y., Mian, A.: Free-form description guided 3D visual graph network for object grounding in point cloud. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "5. Gal, Y., Ghahramani, Z.: Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In: International Conference on Machine Learning (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Gal, Y., Islam, R., Ghahramani, Z.: Deep Bayesian active learning with image data. In: International Conference on Machine Learning (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "O. Unal et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "17. Gao, H.a., Tian, B., Li, P., Zhao, H., Zhou, G.: DQS3D: Densely-matched quantization-aware semi-supervised 3D detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "18. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International Conference on Machine Learning (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.787, + 0.261 + ], + "angle": 0, + "content": "19. He, R., Yang, J., Qi, X.: Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.303 + ], + "angle": 0, + "content": "20. Hou, J., Graham, B., Niessner, M., Xie, S.: Exploring data-efficient 3D scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.332 + ], + "angle": 0, + "content": "21. Houlsby, N., Huszar, F., Ghahramani, Z., Lengyel, M.: Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.787, + 0.36 + ], + "angle": 0, + "content": "22. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3D instance segmentation. In: AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.389 + ], + "angle": 0, + "content": "23. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3D visual grounding. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.787, + 0.431 + ], + "angle": 0, + "content": "24. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.787, + 0.487 + ], + "angle": 0, + "content": "25. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.53 + ], + "angle": 0, + "content": "26. Kendall, A., Badrinarayanan, V., Cipolla, R.: Bayesian SegNet: Model uncertainty in deep convolutional encoder-decoder architectures for scene understanding. In: Proceedings of the British Machine Vision Conference (BMVC) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.787, + 0.572 + ], + "angle": 0, + "content": "27. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.601 + ], + "angle": 0, + "content": "28. Kong, C., Lin, D., Bansal, M., Urtasun, R., Fidler, S.: What are you talking about? text-to-image coreference. In: ICCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.787, + 0.643 + ], + "angle": 0, + "content": "29. Kong, L., Ren, J., Pan, L., Liu, Z.: LaserMix for semi-supervised LiDAR semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.787, + 0.685 + ], + "angle": 0, + "content": "30. Lai, X., Yuan, Y., Chu, R., Chen, Y., Hu, H., Jia, J.: Mask-attention-free transformer for 3D instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.787, + 0.741 + ], + "angle": 0, + "content": "31. Lee, M.S., Yang, S.W., Han, S.W.: Gaia: Graphical information gain based attention network for weakly supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 582-591 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.798 + ], + "angle": 0, + "content": "32. Lei, J., Deng, C., Schmeckpeper, K., Guibas, L., Daniilidis, K.: EFEM: Equivariant neural field expectation maximization for 3D object segmentation without scene supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "33. Liu, K., Zhao, Y., Nie, Q., Gao, Z., Chen, B.M.: Weakly supervised 3D scene segmentation with region-level boundary awareness and instance discrimination. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3D-SPS: Single-stage 3D visual grounding via referred point progressive selection. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "35. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Van Gool, L.: Towards a weakly supervised framework for 3D point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4454-4468 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.288 + ], + "angle": 0, + "content": "36. Prabhudesai, M., Tung, H.Y.F., Javed, S.A., Sieb, M., Harley, A.W., Fragkiadaki, K.: Embodied language grounding with 3D visual feature representations. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.289, + 0.786, + 0.317 + ], + "angle": 0, + "content": "37. Qin, Z., Wang, J., Lu, Y.: Weakly supervised 3D object detection from point clouds. In: Proceedings of the ACM International Conference on Multimedia (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.318, + 0.787, + 0.359 + ], + "angle": 0, + "content": "38. Ren, Z., Misra, I., Schwing, A.G., Girdhar, R.: 3D spatial recognition without spatially labeled 3D. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.36, + 0.786, + 0.388 + ], + "angle": 0, + "content": "39. Roh, J., Desingh, K., Farhadi, A., Fox, D.: LanguageRefer: Spatial-language model for 3D visual grounding. In: CoRL (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.389, + 0.786, + 0.416 + ], + "angle": 0, + "content": "40. Shannon, C.E.: A mathematical theory of communication. The Bell system technical journal 27(3) (1948)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.417, + 0.787, + 0.459 + ], + "angle": 0, + "content": "41. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research 15(1), 1929-1958 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "42. Tang, Y.S., Lee, G.H.: Transferable semi-supervised 3D object detection from rgb-d data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.502, + 0.787, + 0.543 + ], + "angle": 0, + "content": "43. Tao, A., Duan, Y., Wei, Y., Lu, J., Zhou, J.: SegGroup: Seg-level supervision for 3D instance and semantic segmentation. IEEE Transactions on Image Processing 31, 4952-4965 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.545, + 0.787, + 0.585 + ], + "angle": 0, + "content": "44. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.587, + 0.787, + 0.642 + ], + "angle": 0, + "content": "45. Unal, O., Dai, D., Hoyer, L., Can, Y.B., Van Gool, L.: 2d feature distillation for weakly- and semi-supervised 3d semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7336-7345 (January 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.643, + 0.787, + 0.671 + ], + "angle": 0, + "content": "46. Unal, O., Dai, D., Unal, A.T., Van Gool, L.: Discwise active learning for LiDAR semantic segmentation. IEEE Robotics and Automation Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.672, + 0.787, + 0.713 + ], + "angle": 0, + "content": "47. Unal, O., Dai, D., Van Gool, L.: Scribble-supervised LiDAR semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.714, + 0.787, + 0.742 + ], + "angle": 0, + "content": "48. Unal, O., Sakaridis, C., Saha, S., Gool, L.V.: Four ways to improve verbo-visual fusion for dense 3d visual grounding (2024), https://arxiv.org/abs/2309.04561" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.743, + 0.787, + 0.784 + ], + "angle": 0, + "content": "49. Wang, G., Li, W., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T.: Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing 338, 34-45 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "50. Wang, H., Cong, Y., Litany, O., Gao, Y., Guibas, L.J.: 3D IoUMatch: Leveraging IoU prediction for semi-supervised 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "O. Unal et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "51. Wang, J., Gang, H., Ancha, S., Chen, Y.T., Held, D.: Semi-supervised 3D object detection via temporal graph neural networks. In: International Conference on 3D Vision (3DV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "52. Wei, J., Lin, G., Yap, K.H., Hung, T.Y., Xie, L.: Multi-path region mining for weakly supervised 3D semantic segmentation on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "53. Wu, Y., Shi, M., Du, S., Lu, H., Cao, Z., Zhong, W.: 3d instances as 1d kernels. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 235-252. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "54. Xia, Y., Liu, F., Yang, D., Cai, J., Yu, L., Zhu, Z., Xu, D., Yuille, A., Roth, H.: 3D semi-supervised learning with uncertainty-aware multi-view co-training. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (March 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "55. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10687-10698 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "56. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3D point cloud understanding. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "57. Xu, X., Lee, G.H.: Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "58. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmentation. In: MICCAI (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.522 + ], + "angle": 0, + "content": "59. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: InstanceRefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "60. Zhang, F., Fang, J., Wah, B., Torr, P.: Deep fusionnet for point cloud semantic segmentation. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "61. Zhang, Y., Li, Z., Xie, Y., Qu, Y., Li, C., Mei, T.: Weakly supervised semantic segmentation for large-scale point cloud. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "62. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3DVG-Transformer: Relation modeling for visual grounding on point clouds. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "63. Zhao, N., Chua, T.S., Lee, G.H.: SESS: Self-ensembling semi-supervised 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "64. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.744 + ], + "angle": 0, + "content": "65. Zoph, B., Ghiasi, G., Lin, T.Y., Cui, Y., Liu, H., Cubuk, E.D., Le, Q.V.: Rethinking pre-training and self-training. arXiv preprint arXiv:2006.06882 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "66. Zou, Y., Yu, Z., Vijaya Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: The European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.785 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/546e4382-dd78-463e-8d89-435110642a17_origin.pdf b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/546e4382-dd78-463e-8d89-435110642a17_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8405f824789cd149a48098c0ef4be80b65d20fe6 --- /dev/null +++ b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/546e4382-dd78-463e-8d89-435110642a17_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e831af5ba78d0b13083651ecdf2ddc40405cdbc13e2331563bad570155eeb37 +size 9479198 diff --git a/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/full.md b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ec52ddec68897dea2c7db131c278311dc5d9440c --- /dev/null +++ b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/full.md @@ -0,0 +1,330 @@ +# Bayesian Self-Training for Semi-Supervised 3D Segmentation + +Ozan Unal $^{1,2}$ , Christos Sakaridis $^{1}$ , and Luc Van Gool $^{1,3,4}$ + +$^{1}$ ETH Zurich, $^{2}$ Huawei Technologies, $^{3}$ KU Leuven, $^{4}$ INSAIT {ozan.unal, csakarid, vangool}@vision.ee.ethz.ch + +Abstract. 3D segmentation is a core problem in computer vision and, similarly to many other dense prediction tasks, it requires large amounts of annotated data for adequate training. However, densely labeling 3D point clouds to employ fully-supervised training remains too labor intensive and expensive. Semi-supervised training provides a more practical alternative, where only a small set of labeled data is given, accompanied by a larger unlabeled set. This area thus studies the effective use of unlabeled data to reduce the performance gap that arises due to the lack of annotations. In this work, inspired by Bayesian deep learning, we first propose a Bayesian self-training framework for semi-supervised 3D semantic segmentation. Employing stochastic inference, we generate an initial set of pseudo-labels and then filter these based on estimated point-wise uncertainty. By constructing a heuristic $n$ -partite matching algorithm, we extend the method to semi-supervised 3D instance segmentation, and finally, with the same building blocks, to dense 3D visual grounding. We demonstrate state-of-the-art results for our semi-supervised method on SemanticKITTI and ScribbleKITTI for 3D semantic segmentation and on ScanNet and S3DIS for 3D instance segmentation. We further achieve substantial improvements in dense 3D visual grounding over supervised-only baselines on ScanRefer. Our project page is available at ouenal.github.io/bst/. + +Keywords: Semi-supervised $\cdot$ Self-training $\cdot$ 3D semantic segmentation $\cdot$ 3D instance segmentation $\cdot$ 3D visual grounding + +# 1 Introduction + +Dense 3D semantic perception at the level of points is pivotal for several applications that are based on fine-grained processing and reasoning in 3D, e.g. autonomous cars, robotics, and augmented reality. Such perception requires a segmentation of the unstructured input 3D point cloud into different semantic classes, i.e. 3D semantic segmentation [60], or distinct instances of these classes, i.e., 3D instance segmentation [30]. The input point cloud is often coupled with human utterances that refer to individual objects in the former in cases where a visual AI system interacts with human users in a common 3D environment, necessitating a 3D instance-level segmentation of these objects based on the aforementioned referrals, a task which is known as dense 3D visual grounding or + +![](images/80c238c525cec2778e79dc167ed82162c7a232dfce5551928bdea7297789d1d6.jpg) +Fig. 1: Illustration of our Bayesian pseudo-labeling pipeline for semi-supervised a) 3D semantic segmentation; b) 3D instance segmentation; and c) dense 3D visual grounding. With only slight adjustments, using the same building blocks, our method can be adapted to each of these tasks to achieve SOTA results. + +referral-based 3D instance segmentation [48, 59]. In the majority of works that address these diverse 3D segmentation tasks, it is assumed that the training data come with full 3D semantic and/or verbal annotations, which creates the pressing need for large-scale labeled 3D datasets [1, 2, 6, 11] so that large end-to-end models can be adequately fitted on these data. + +However, the creation of such fine-level 3D annotations and verbal referrals for 3D scans requires human labor to be fairly reliable. Such labor is time-consuming and expensive, especially due to the high dimensionality of the scanned point clouds, which makes it hard for human annotators to navigate through them, visually inspect them, and verbally describe them. These practical difficulties render learning frameworks for 3D segmentation that operate with partial labels essential, semi-supervised learning being a primary case. Answering this need, several recent works address semi-supervised 3D semantic segmentation [9,25,29], most of which focus on effective pseudo-labeling techniques for self-training on the larger, unlabeled portion of the training data. However, the generation of pseudo-labels in these works typically focuses on some grouping mechanism over 3D points, such as superpoints, to refine and propagate pseudo-labels and otherwise relies on naïve core techniques such as confidence thresholding [66] for generating the initial point-level pseudo-labels that are then post-processed by their proposed modules. Adding to this, there is a relative scarcity in works on semi-supervised 3D instance segmentation [10, 33], which we attribute to the relative difficulty in pseudo-labeling instances compared to creating plain semantic pseudo-labels. Last but not least, to our knowledge, no work has considered the setting of semi-supervised dense 3D visual grounding. + +We jointly address these shortcomings of previous works by proposing a common self-training framework for semi-supervised 3D segmentation which is based on Bayesian deep learning and applies to all three aforementioned segmentation tasks. At the core of our self-training framework lies a variational Bayesian approximation of the uncertainty of the respective network's predictions via dropout-based Monte Carlo integration [15] over multiple augmented versions of + +the original 3D input. This approximation delivers a more informed uncertainty estimate than simple confidence thresholding techniques [66], which is crucial for eliminating confirmation bias from the pseudo-labeling process that we adopt for self-training. In particular, we utilize both the estimated Shannon entropy from the sampled predictions and the predicted label from each sample to filter out uncertain or inconsistent points from the generated set of pseudo-labels. + +While this formulation is sufficient for the task of 3D semantic segmentation, it does not work out of the box for instance-level segmentation, as Monte Carlo integration cannot be applied directly to instance predictions. To account for this issue, we propose a novel and efficient, heuristic matching strategy across instances coming from different stochastic dropout forward passes based on linear sum assignment. This strategy enables the application of the same filtering steps as described above for semantic pseudo-labels to instance pseudo-labels and thus extends the scope of our approach to 3D instance segmentation. Going one step further, we observe that the same matching strategy that we devise for instance segmentation applies to the selection stage of dense 3D grounding, in which the correct instance from the set of segmented candidate instances needs to be selected based on the input verbal description. Thus, for dense 3D grounding, we apply our heuristic linear sum assignment twice, first for the purely 3D-based candidate instance proposal generation and afterward for the verbo-visual instance selection. Our complete self-training pipeline for each of the three 3D segmentation tasks to which our method applies is summarized in Fig. 1. + +We test our method on leading benchmarks across a wide range of domains and datasets. Specifically, our Bayesian self-training framework achieves state-of-the-art results for semi-supervised 3D semantic segmentation on SemanticKITTI [3] and ScribbleKITTI [47] and semi-supervised 3D instance segmentation on ScanNet [11] and S3DIS [2]. Furthermore, we are the first to explore a semi-supervised setting on ScanRefer [6] for 3D visual grounding and obtain consistent significant improvements over supervised-only baselines. + +In summary, our contributions are as follows: + +1. Following established uncertainty estimation techniques from active learning, we construct a Bayesian self-training framework for semi-supervised 3D semantic segmentation by generating reliable pseudo-labels through stochastic inference and entropy-based uncertainty filtering. +2. We introduce a novel heuristic matching algorithm to extend the proposed framework to 3D instance segmentation. +3. We further extend the method to dense 3D visual grounding utilizing the same building blocks as in the above tasks and are the first to explore this task in a semi-supervised setting. +4. We obtain state-of-the-art results for semi-supervised 3D segmentation on SemanticKITTI [3], ScribbleKITTI [47], ScanNet [11] and S3DIS [2]. + +# 2 Related Work + +Semi-supervised 3D semantic segmentation has recently grown in popularity in the area of 3D understanding along with weakly supervised 3D semantic segmentation [47,52,57,61], due to the difficulty in large-scale 3D labeling that is necessary for the supervised setup [60]. In this work, we tackle the task of semi-supervised segmentation where a large portion of data is completely unlabeled. As opposed to methods developed for weakly-supervised settings, a semi-supervised method cannot leverage the availability of partial true label information within each data sample, e.g. via the ArcPoint loss [31] or a PLS descriptor [47]. A common technique for leveraging unlabeled data under this restrictive scheme is explored under pseudo-labeling. Semantic pseudo-labels are utilized by Jiang et al. [25] to guide a contrastive loss and in SSPC [9] to propagate labels on a superpoint graph. Superpoints are also leveraged by Deng et al. [13] to constrain pseudo-label propagation. In the 2D counterpart task, Chen et al. [8] enforces consistency of semantic pseudo-labels across perturbed versions of the 2D segmentation network. By contrast, we perturb both the network weights and the 3D inputs across different forward passes, and we consider both the entropy and the consistency among predictions in our pseudo-labels. Another related line of 2D semi-supervised works generates pseudo-labels based on direct thresholding of max-softmax scores to reduce confirmation bias [19, 66], while we threshold the entropy of the distribution of softmax scores. Active learning is combined with a 3D semi-supervised approach in DiAL [46] to streamline practical 3D segmentation annotation pipelines. LaserMix [29] introduces variability through cylindrical and range-view partitioning and mixing of 3D LiDAR scans, encouraging consistency across different mixes. + +Semi-supervised 3D instance segmentation is still in its infancy compared to its instance-agnostic semantic segmentation counterpart reviewed above. In particular, most object-level semi-supervised 3D understanding works focus on the coarser task of 3D object detection [17,35,37,38,42,50,51,63], which outputs 3D bounding boxes that are not fine enough for precision-critical downstream tasks such as robotic grasping. For 3D instance/object segmentation, approaches that depart from the standard fully supervised setup have mostly considered the cases of unsupervised pre-training [56], weak supervision via one-point-one-instance [20, 43], or unsupervised generic 3D object segmentation [32]. To our knowledge, the only two semi-supervised works that specifically consider 3D instance segmentation are TWIST [10] and WS3D [33]. The former work proposes a pseudo-label generation strategy tailored for the instance segmentation task, for which the technique of confidence thresholding that is ubiquitously used to the same end in semantic segmentation is not applicable. It thus introduces pseudo offset vectors, besides semantic proposal-level pseudo-labels, for performing self-training. The latter work proposes a region-level contrastive loss, which performs unsupervised instance discrimination, and leverages boundary information via a region-level unsupervised energy-based loss. On the contrary, we propose a unified semi-supervised Bayesian self-training framework that spans multiple 3D segmentation tasks including 3D instance segmentation, and han + +dles the additional matching required between instances generated in different forward passes through the network via an efficient heuristic optimization step. + +3D visual grounding is a central 3D task in the area of vision and language [4, 14, 28, 34, 36, 62]. Relevant models accept a 3D point cloud as visual input and perform grounding of the accompanying verbal description to the referred object. While most previous works [7, 23, 24, 34, 39] focus on coarse 3D bounding box outputs, the task of dense 3D visual grounding or referral-based 3D instance segmentation has been explored far less in comparison, in which a fine-grained point-level mask is predicted [22, 48, 59]. Regardless, all aforementioned methods invariably assume a supervised setting, where full ground-truth instance annotations are available for all samples at training time. By contrast, the formulation of our method makes it fit for also handling the more practical semi-supervised setting, in which only a part of the training data is labeled. + +Uncertainty estimation is crucial in learning setups that lack full supervision as ours, given that such estimates can provide information about the reliability of the predictions of the model on unlabeled samples [21]. In the more recent deep learning era, Bayesian neural networks form the core for uncertainty estimation by modeling network weights as stochastic distributions, where test-time-dropout [41] is utilized as a variational Bayesian approximation [15] and Monte Carlo integration is computed via multiple stochastic forward passes to approximate prediction uncertainty [16]. This Bayesian deep learning framework has been applied successfully to supervised 2D dense prediction tasks [26, 27]. We focus on the aggregate uncertainty in the predictions of a general 3D segmentation model and build on the Bayesian deep learning approach of Gal et al. [16] which examines the approximate predictive entropy as an acquisition function for active learning. However, contrary to Gal et al. [16], in our self-training setup, we use low rather than high entropy as a positive indication for considering a point as a source of supervision. To the best of our knowledge, our work is the first to introduce such a Bayesian self-training framework for general 3D point cloud segmentation, specifically including 3D instance segmentation and referral-based 3D instance segmentation. + +Related uncertainty-based approaches to ours have been proposed for the semantic segmentation of 3D grid volume data in the field of medical image computing [12,54,58]. In particular, Yu et al. [58] also use Monte Carlo dropout to estimate predictive entropy but only applies dropout to a single network input, whereas we vary both network weights and input views. Moreover, that work leverages the estimated entropy merely as the weight in a consistency loss between a teacher and a student network, while we filter pseudo-labels based on our entropy estimates. Besides, Xia et al. [54] perform semi-supervised semantic segmentation with multi-view co-training and applies Bayesian deep learning across views for uncertainty-based weighting of pseudo-labels. We only train one network instead of multiple networks employed in Xia et al. [54] and we derive the uncertainty estimates by feeding all input views to that single network, which is more parameter-efficient. Both of these approaches only address 3D semantic segmentation, and cannot be utilized in instance-level dense prediction tasks. + +![](images/33aa6741a249313c163f3dde3e2408db3f6c1b7f521ff2908ff9085a622b7eda.jpg) +Fig. 2: Illustration of Bayesian pseudo-labeling for semi-supervised instance segmentation. We initialize a seed prediction via a forward pass using a non-augmented input. Then our heuristic $n$ -partite matching algorithm is employed to pair each seed instance with the best matching predicted instance from each of the $K$ stochastic forward passes. For each aligned object, we compute the aggregated label through unanimous voting and filter it based on uncertainty to obtain the final pseudo mask. + +# 3 Bayesian Self-Training + +In Sec. 3.1 we start by formulating a self-training approach for semi-supervised 3D semantic segmentation based on established uncertainty estimation techniques from the field of active learning [15,46]. We then extend our formulation to include instance labels in Sec. 3.2 and finally build up to also take natural language prompts as input, i.e., towards semi-supervised dense 3D visual grounding in Sec. 3.3. + +# 3.1 Semi-Supervised 3D Semantic Segmentation + +In a semi-supervised setting, a dataset $D$ consists of a labeled set $S$ and an unlabeled set $U$ . The baseline approach when working with such data is to treat the problem similarly to any fully supervised task and employ a semantic loss $\mathcal{L}_{sem}$ on the available labeled scenes $S$ , typically stated as cross-entropy. Formally, the objective function over the model parameters $\omega$ is formulated as: + +$$ +\min _ {\omega} \sum_ {i = 1} ^ {| S |} \mathcal {L} _ {s e m} (\hat {\mathbf {y}}, \mathbf {y}). \tag {1} +$$ + +Starting from this baseline formulation, a semi-supervised training strategy aims to leverage the unlabeled scenes $U$ to further improve model performance. One such common technique is self-training, where the model is iteratively trained on its own predictions, i.e. on generated per-point one-hot pseudo-label vectors $\hat{\mathbf{y}} = [\hat{\mathbf{y}}^{(1)},\dots,\hat{\mathbf{y}}^{(C)}]$ , with $C$ denoting the number of classes. However, the task of generating pseudo-labels is not trivial, as the model may suffer from confirmation bias that occurs when repeatedly overfitting to incorrect labels. + +A simple but effective solution is explored under the umbrella of thresholding [5, 19, 55, 65, 66]. In essence, instead of training on all model predictions, a pseudo-label set $L$ is formed by predictions that exceed a confidence of a determined threshold $\tau$ to ensure a higher level of accuracy. Formally, the objective function of threshold-based self-training for semantic segmentation can be formulated as: + +$$ +\min _ {\omega , \hat {\mathbf {y}}} \sum_ {d = 1} ^ {| D |} \left(\mathcal {L} _ {s e m} (\hat {\mathbf {y}}, \mathbf {y}) - (\log \hat {\mathbf {y}} _ {d} - \tau) \cdot \hat {\mathbf {y}} _ {d}\right) \tag {2} +$$ + +with $\hat{\mathbf{y}}_d$ denoting the softmax model predictions. Eq. 2 is optimized in two steps that are repeated until convergence: + +1. Training: First, $\hat{\mathbf{y}}$ is fixed, and the objective function is optimized with respect to $\omega$ . +2. Pseudo-labeling: Then, $\omega$ is fixed and the objective function is optimized with respect to $\hat{\mathbf{y}}$ . Consequently, $L$ is updated given $\hat{\mathbf{y}}$ . + +The process is initialized by setting the latent variable $\hat{\mathbf{y}} = \mathbf{0}$ for all points, i.e. by reverting to the naive solution of Eq. 1 and only training on the labeled set $S$ . The two steps can be repeated to take advantage of the improved representation capability of the model thanks to pseudo-labeling. + +While naïve thresholding relies on the softmax scores of the model to determine confidence, these scores may be unreliable and often skewed when facing long-tailed distributions, resulting in overconfident predictions [18]. Inspired by uncertainty estimation techniques commonly employed in active learning (AL) pipelines [21,46], we introduce a Bayesian self-training strategy that leverages an uncertainty-aware filtering technique for generating more reliable pseudo-labels. + +We first start by decomposing the goal of 3D semantic segmentation, which is to discover the dependency of the point-wise distribution over the labels $y \in Y$ on an input sample $x \in X$ via the model weights $\omega$ . Formally, the task can be reduced to per-point classification, where $y$ can be of class $c \in \{1, \dots, C\}$ . The conditional probability is therefore defined by: + +$$ +p (y = c | x, D) = \int p (y = c | a (x), \omega) p (\omega | D) p (a (x) | x) d \omega d x \tag {3} +$$ + +for the dataset $D = (X, Y)$ , under randomly applied affine transformations modeled through $a(\cdot)$ . + +We set $p(a(x)|x)$ as a constant, restate the intractable $p(\omega |D)$ via the variational approximation $q(\omega)$ and minimize $\mathrm{KL}(q(\omega)|p(\omega |X,Y))$ . This objective can be approximated through Monte Carlo integration by employing variational inference via Bayesian networks, i.e. by applying stochastic forward passes with weights $\omega_{k}$ : + +$$ +\begin{array}{l} p (y = c | x, D) \approx \int p (y = c | a (x), \omega) q (\omega) d \omega d x \\ \approx \frac {1}{K} \sum_ {k = 1} ^ {K} p (y = c | a (x), \omega_ {k}). \tag {4} \\ \end{array} +$$ + +Compared to the common decomposition of BALD [21], the inclusion of the affine transformations via $a(\cdot)$ allows a higher deviation amongst stochastic forward passes which in return allows better estimation of predictive uncertainty [49]. We capture the uncertainty via the Shannon entropy [40]: + +$$ +\mathbb {H} [ y | x, D ] = - \sum_ {c = 1} ^ {C} p (y = c | x, D) \log p (y = c | x, D) \tag {5} +$$ + +and the accumulated predicted label $\hat{\mathbf{y}}^*$ is given by the unanimous voting principle: + +$$ +\hat {y} ^ {* (c)} = \left\{ \begin{array}{l} 1, \text {i f} K = \sum_ {k} \mathbb {1} (\operatorname {a r g m a x} (p (y | a (x), \omega_ {k})) = c) \\ 0, \text {o t h e r w i s e}, \end{array} \right. \tag {6} +$$ + +with $\mathbb{1}$ denoting the indicator function. + +Unlike what is commonly seen in active learning settings, our goal is not to maximize the total entropy but rather to minimize it in order to ensure a high-quality pseudo-label set. We therefore redefine the self-training objective (Eq. 2) for stochastic pseudo-labeling as: + +$$ +\min _ {\omega , \hat {\mathbf {y}}} \sum_ {d = 1} ^ {D} \left(\mathcal {L} _ {s e m} (\hat {\mathbf {y}}, \mathbf {y}) + \left(\mathbb {H} [ y | x, D ] - \tau\right) \left(\hat {\mathbf {y}} ^ {*} \cdot \hat {\mathbf {y}}\right)\right) \tag {7} +$$ + +with $\tau \in (0,1)$ determining the cut-off entropy threshold. To solve the nonlinear integer optimization task, we employ the following solver: + +$$ +\hat {\mathcal {y}} ^ {(c)} = \left\{ \begin{array}{l} 1, \text {i f} \mathbb {H} [ y | x, D ] < \tau \text {a n d} \\ c = \operatorname {a r g m a x} \hat {\mathcal {y}} ^ {*} \\ 0, \text {o t h e r w i s e ,} \end{array} \right. \tag {8} +$$ + +where labels $\hat{\mathbf{y}}^{(c)} = 0\forall c$ are ignored during training. + +In practice, following precedent [47,66], we set $\tau$ based on $p_{\tau}$ , with $p_{\tau}$ determining the percentage of pseudo-labels to be sampled amongst the accumulated predicted labels. + +In summary, we first determine a set of labels that are agreed upon by all stochastic forward passes, i.e. by forward passes with active dropout [15, 41]. We then compute the Shannon entropy for all predictions, filtering those that exhibit high entropy. This surprisingly simple but effective approach allows us to reduce the incorrect targets that we introduce to our pseudo-label set and to avoid the overconfidence issue of naive thresholding-based approaches. + +# 3.2 Semi-Supervised 3D Instance Segmentation + +The proposed Bayesian self-training framework provides a strong basis for any semi-supervised 3D dense prediction task, one such example being 3D instance segmentation. Yet, compared to semantic segmentation where accumulation of + +Algorithm 1: Bayesian pseudo-label generation for instance segmentation. +Input: Point cloud $P\in \mathbb{R}^{N\times 3}$ , trained model with dropout $\omega$ , number of passes $K$ , threshold $\tau$ +1 seed_output $= \omega (P)\in \mathbb{R}^{M\times N}$ +2 $\pmb{i} =$ seed_output.copy() +3 seed_mask $=$ to_mask(i) $\in \{0,1\}^{M\times N}$ +4 unanimous $=$ seed_mask.copy() +5 for k in 1:K do +6 output $= \omega_{k}(a(P))\in \mathbb{R}^{M^{\prime}\times N}$ +7 mask $=$ to_mask(output) $\in \{0,1\}^{M^{\prime}\times N}$ +8 cost $=$ -iou(seed_mask, mask) +9 row,col $=$ LSA(cost) +10 unanimous[col] &= mask[row] +11 i[col] $+ =$ output[row] +12 $\mathbb{H} = (K - i)\cdot \log (K - i) + i\log i$ +13 pseudo_label $= (\mathbb{H} < \tau)$ & unanimous +Return:pseudo_label $\in \{0,1\}^{M\times N}$ + +labels can be easily handled across multiple stochastic forward passes, accumulation of predicted instance masks is not trivial and requires $n$ -partite matching. + +However, the $n$ -partite matching problem is NP-hard with a complexity of $|I|!(K-1)$ with $K$ passes and $|I|$ number of instances assuming full bijection. To match predicted instance masks across various stochastic forward passes, we propose a heuristic method based on the Hungarian algorithm. First, we generate an initial prediction via an unaugmented forward pass (Alg. 1 L1&3). The goal is to establish a seed that is likely of a higher quality with minimal perturbations of the data. We then iteratively assign predicted masks of each stochastic forward pass via linear sum assignment, with the cost matrix formed by the negative pairwise IoUs (Alg. 1 L5-10). + +Having matched predicted binary instance masks from individual stochastic forward passes, each resulting set of masks can then be treated as a set of 2-class semantic segmentation outputs, and thus our proposed Bayesian pseudo-labeling approach from Eq. 8 can be trivially applied (Alg. 1 L11-15). Our pipeline for semi-supervised 3D instance segmentation is illustrated in Fig. 2. + +# 3.3 Semi-Supervised Dense 3D Visual Grounding + +Similar to both semi-supervised 3D semantic and 3D instance segmentation, our proposed Bayesian self-training formulation can easily be extended to include natural language prompts as inputs and thus tackle semi-supervised dense 3D visual grounding, i.e., referral-based 3D instance segmentation. + +A common technique when tackling dense 3D visual grounding is to employ a grounding-by-selection strategy which can be summarised in two steps [6,48,62]: + +1. Visual: First, a 3D backbone computes object candidates $O$ from the scene. +2. Verbo-visual: Then, a verbo-visual fusion module $\omega_{sel}$ selects the correct instance candidate from predicted instances $I$ based on the natural language description $W$ , i.e., it predicts the index: + +$$ +i ^ {*} = \operatorname {a r g m a x} G, \text {w i t h} G = \omega_ {s e l} (O, W) \in \mathbb {R} ^ {| I |}, \tag {9} +$$ + +To extend our Bayesian self-training approach to this task, we follow the above structure and individually adapt each of the two steps, with the adaptation of the visual step simply following Sec. 3.2. + +For the verbal side of 3D visual grounding, i.e. referral-based instance selection, we observe that the pseudo-label generation can be reduced to the same formulation as the pseudo-label generation for 3D instance segmentation. Utilizing the results from the heuristic $n$ -partite matching (Alg. 1 L9), vector $G$ can be reordered to match across stochastic forward passes. Again, having matched predicted probability vectors, the method presented in Sec. 3.1 can be trivially applied (Alg. 1 L11-15). + +# 4 Experiments + +We evaluate our method on all three semi-supervised 3D semantic perception tasks: 3D semantic segmentation, 3D instance segmentation, and dense 3D visual grounding. The dataset and implementation details regarding each task can be found in the supplementary materials. + +# 4.1 3D Semantic Segmentation + +The comparison to the state of the art for semi-supervised 3D semantic segmentation is shown in Tab. 1. As seen, our Bayesian self-training strategy outperforms existing work in both SemanticKITTI [3] and ScribbleKITTI [47] on splits of $10\%$ , $20\%$ and $50\%$ , while performing on par for $1\%$ . Specifically, compared to the baseline Cylinder3D [64] model denoted as Sup-only, our strategy allows significant improvements ranging from $4.4\%$ to $10.9\%$ mIoU. In Fig. 3 (a-c) we showcase qualitative results on SemanticKITTI $10\%$ . + +# 4.2 3D Instance Segmentation + +For semi-supervised 3D instance segmentation, we report results for ScanNet [11] in Tab. 2 and for S3DIS [2] in Tab. 3. As seen, we observe a similar behavior to semi-supervised 3D semantic segmentation, where at higher thresholds of $10\%$ , $20\%$ for ScanNet and $20\%$ , $50\%$ for S3DIS, our simple yet effective Bayesian self-training strategy significantly outperforms state-of-the-art methods that require highly engineered solutions. In contrast, at lower thresholds, we see a significant gap in performance (especially for ScanNet). Still, compared to the baseline Sup-only, we see significant gains across the board for all semi-splits. Additionally in Fig. 3 (d-f), we show qualitative results on the ScanNet val-set. + +Table 1: Comparison of state-of-the-art methods for semi-supervised LiDAR semantic segmentation (mIoU [%]). For fair comparison, all methods use the same backbone given by Sup-only, which at $100\%$ labels achieves $64.3\%$ mIoU on SemanticKITTI and $57.0\%$ mIoU on ScribbleKITTI. + +
Labeled1%10%20%50%
SemantickKITTI [3]Sup-only [64]45.456.157.858.7
DiAL [46]45.457.159.260.0
CBST [66]48.858.359.459.7
CPS [8]46.758.759.660.5
LaserMix [29]50.660.061.962.3
IGNet [45]49.061.363.164.8
Ours49.861.763.764.1
Δ Sup-only+4.4+5.6+5.9+5.4
ScribbleKITTI [47]Sup-only [64]39.248.052.153.8
DiAL [46]41.050.152.853.9
CBST [66]41.550.653.354.5
CPS [8]41.451.853.954.8
LaserMix [29]44.253.755.156.8
IGNet [45]44.457.759.660.8
Ours43.958.960.161.2
Δ Sup-only+4.7+10.9+8.0+7.4
+ +Table 2: Comparison to state-of-the-art methods for semi-supervised 3D instance segmentation on the ScanNet Limited Reconstruction valid-set. For a fair comparison, all methods apart from that marked with * use the same backbone [20,53] as Sup-only. + +
Labeled Metric1% mAPAP50AP255% mAPAP50AP2510% mAPAP50AP2520% mAPAP50AP25
Sup-only [10]5.19.817.618.232.047.026.742.858.929.347.963.0
PointCont. [56]7.212.520.319.435.448.527.043.959.530.249.563.6
CSC [20]7.113.021.220.936.750.627.345.060.230.650.364.1
TWIST [10]9.617.126.227.044.156.230.649.763.032.852.966.8
WS3D* [33]-32.5--45.6--49.2--51.3-
Ours7.214.223.024.239.351.432.751.965.537.456.168.9
Δ Sup-only+2.1+4.4+5.4+6.0+7.3+4.4+6.0+9.1+6.6+8.1+8.2+5.9
+ +# 4.3 Dense 3D Visual Grounding + +For semi-supervised 3D visual grounding, we report results on the ScanRefer [6] dataset. As seen in Tab. 4 (Vi), our self-training approach shows improvements across the board, with significant gains on the "unique" category. + +While natural language prompts still require human effort to generate, they are considerably cheaper and easier to acquire than per-point high-quality panoptic labels. In Tab. 4 (ViVe), we show that the inclusion of such prompts in the unlabeled dataset further benefits the model's accuracy. As seen, we observe even larger gains over the baseline method with the overall accuracy (OA) showing $+4.7\%$ improvement with a $50\%$ IoU threshold. + +Furthermore, in Fig. 3 (g) we qualitatively compare our method to the baseline Sup-only on ScanRefer. + +![](images/17435439459c2fdcb7ba5035be734083ab12ecacf852b75174414faf883fc9be.jpg) +Fig.3: Qualitative results from the SemanticKITTI (10%) val-set, comparing a) the ground truth, b) Sup-only and c) ours; on ScanNet (10%), comparing d) the ground-truth instance masks, e) Sup-only and f) ours; and finally g) ScanRefer (10%). + +![](images/a3d0b6e442955374224e691faa8599fd07a346a34ff644a93aa7b0f97345270d.jpg) + +![](images/ac34a78f41280f5d2af4764e42562d5eb617bf59bb496b79827d6b48c7e07b94.jpg) + +![](images/234c063d3d34d436477c3eeb1231824c17a3ae258907b07a84541d5debd887ad.jpg) + +![](images/59ed4a2de850b2b756a79bfaca50b8c0abf650d965fe1226a435cba4a9343518.jpg) +The table is square. There is a chair to the left, and a chair to the right of it. + +Table 3: Comparison to state-of-the-art methods for semi-supervised 3D instance segmentation on S3DIS Area 5 via AP50 [%]. For a fair comparison, all methods use the same backbone Sparse U-Net [20,53] as Sup-only. + +
Labeled5%10%20%50%
Sup-only [10]30.436.841.246.5
PointContrast [56]33.638.743.148.9
CSC [20]34.241.044.750.4
TWIST [10]37.145.648.4-
Ours35.042.951.156.6
Δ Sup-only+4.6+6.1+9.9+10.1
+ +# 4.4 Ablation Studies + +We conduct ablation studies on $10\%$ splits of SemanticKITTI [3], ScanNet [11] and ScanRefer [6] for semi-supervised 3D semantic segmentation, instance segmentation and visual grounding respectively. + +Entropy Thresholding: In Tab. 5, we show an analysis of the effectiveness of our proposed uncertainty thresholding via Shannon entropy, comparing it to the baseline unanimous voting results as well as established thresholding techniques of softmax thresholding [55] and class-balanced thresholding [66]. Alongside the performance for each individual task, we also report the accuracy of the generated pseudo-labels. For instance segmentation, a pseudo-mask is considered correct if its intersection with the associated ground-truth mask is identical to the former. For visual grounding, we reduce the accuracy threshold to 0.5 following the benchmark metric used. As seen, our uncertainty-based filtering improves pseudo-label quality across the board. We also showcase the performance implications of uncertainty filtering, where it can be observed that entropy thresholding outperforms both unanimous voting and current thresholding strategies across all tasks. We further provide a qualitative analysis of the uncertainty met- + +Table 4: Results for semi-supervised 3D dense visual grounding while assuming (Vi) that the model is given no natural language descriptions for the unlabeled split and (ViVe) when the model has access to natural language descriptions for the unlabeled split. Reported values are given as Acc@25/50 for the $10\%$ split. + +
CategoryUniqueMultipleOverall
Sup-only [48]51.9 / 43.619.3 / 14.125.4 / 19.6
ViMe ViOurs58.8 / 48.822.0 / 18.228.8 / 23.9
Δ Sup-only+6.9 / +5.2+2.7 / +4.1+3.4 / +4.3
Ours61.0 / 51.223.4 / 18.130.4 / 24.3
Δ Sup-only+9.1 / +7.6+4.1 / +4.0+5.0 / +4.7
+ +Table 5: Comparison of the pseudo-label quality of current thresholding strategies and their effects on performance on SemanticKITTI $10\%$ , ScanNet $10\%$ , ScanRefer $10\%$ (with visual and verbal) for each respective task. Pseudo-label thresholding is applied after unanimous voting for a fair comparison. The letters indicate each pseudolabel type: $S-$ for semantic, $I-$ for instance, and $G-$ for grounding. + +
SemanticInstanceVisual Grounding
MethodS-AccmIoUS-AccI-AccAP@50S-AccI-AccG-AccOA@50
Sup-only [48]-56.1--42.8---19.6
Unanimous96.460.079.375.347.072.969.961.516.9
Naïve [5,55,65]98.360.882.977.750.282.073.068.621.9
Class-balanced [66]96.661.181.270.144.480.969.060.616.2
Entropy-based Δ Unanimous98.661.784.877.851.983.875.275.624.3
+2.2+1.7+5.5+2.5+4.9+5.8+2.3+5.7+7.4
+ +![](images/3d92ef1074863bd45a8f330dfeef17aac404341158192cfb45aabb8d9535afd4.jpg) +Fig. 4: Qualitative analysis of the uncertainty estimation. We compare softmax confidence to our MC-derived Shannon entropy. + +ric in Fig. 4 to illustrate the better correlation of our MC-derived uncertainty with pseudo-label accuracy compared to softmax based uncertainty. + +Labeling Percentage $(p_{\tau})$ : We compare various set labeling percentages $p_{\tau}$ across all tasks in Tab. 6. While we obtain higher accuracy pseudo-labels at $50\%$ , it can be seen that the reduced supervision negatively impacts performance. + +# 5 Limitations and Discussion + +The main advantage of the proposed Bayesian self-training is its simplicity and ease of implementation compared to existing SOTA approaches that utilize highly engineered and often costly modules. However, the method has a limitation. As observed from the results, the performance is comparably inferior to + +Table 6: Ablation study analyzing the effect of the hyperparameter $p_{\tau}$ for each task. We vary the labeling percentage and show the performance on SemanticKITTI 10%, ScanNet 10%, and ScanRefer 10% (incl. verbal input) for each respective task. + +
ThresholdmIoUAP@50OA@50
50%61.450.022.8
75%61.751.924.3
100%60.047.016.9
+ +Table 7: Ablation study showing the performance on SemanticKITTI $10\%$ , ScanNet $10\%$ and ScanRefer $10\%$ (incl. verbal input) for varying label aggregation strategies from majority rules to unanimous voting over 9 stochastic forward passes. + +
ThresholdmIoUAP@50OA@50
5/960.747.920.1
7/961.051.023.9
9/961.751.924.3
+ +SOTA semi-supervised approaches for highly sparse training splits. This behavior stems from the reliance on unanimous voting (Eq. 6). With low amounts of labeled data, the initial naive model performance suffers, causing a reduction in the number of unanimously voted points. This effect is further pronounced when tackling instance segmentation, where fused instance masks may affect multiple predictions. In Tab. 7, we show that relaxation does not aid the performance as it further introduces incorrect predictions into the pseudo-label set. + +While semi-supervised training methods that enhance the robustness of the pseudo-labels either via a teacher model (DiAL [44, 46]), mix-augmentations (LaserMix [29]) or multi-stage refinement (TWIST [10]) are presented as competing to ours, we argue that as a solution, they can in fact be utilized within our framework in place of the 3D model to improve the quality of the initial predictions. This would allow a larger set of unanimously voted samples and would thus re-enable the benefits of our proposed method on highly sparse splits. + +# 6 Conclusion + +In this work, we introduce a simple yet effective Bayesian self-training approach for semi-supervised 3D semantic segmentation by utilizing uncertainty estimation strategies from the field of active learning. Specifically, through multiple stochastic forward passes, we generate an initial pseudo-label set and then filter it based on estimated predictive entropy. We further extend our semi-supervised 3D method to newly include both instance segmentation and referral-based instance segmentation via an efficient, heuristic $n$ -partite matching algorithm. We conduct extensive experiments on multiple datasets, covering both indoor and outdoor domains, showing that our method performs on par if not better than current domain- and task-specific methods. + +Acknowledgments: This work is funded by Toyota Motor Europe via the research project TRACE-Zürich. + +# References + +1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: ReferIt3D: Neural listeners for fine-grained 3D object identification in real-world scenes. In: ECCV (2020) +2. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3D semantic parsing of large-scale indoor spaces. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2016) +3. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019) +4. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3DJCG: A unified framework for joint dense captioning and visual grounding on 3D point clouds. In: CVPR (2022) +5. Cascante-Bonilla, P., Tan, F., Qi, Y., Ordonez, V.: Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 6912-6920 (2021) +6. Chen, D.Z., Chang, A.X., Nießner, M.: ScanRefer: 3D object localization in RGB-D scans using natural language. In: ECCV (2020) +7. Chen, J., Luo, W., Wei, X., Ma, L., Zhang, W.: HAM: Hierarchical attention model with high performance for 3D visual grounding. arXiv preprint arXiv:2210.12513 (2022) +8. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +9. Cheng, M., Hui, L., Xie, J., Yang, J.: SSPC-Net: Semi-supervised semantic 3D point cloud segmentation network. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021) +0. Chu, R., Ye, X., Liu, Z., Tan, X., Qi, X., Fu, C.W., Jia, J.: TWIST: Two-way inter-label self-training for semi-supervised 3D instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +1. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: ICCV (2017) +2. Dai, W., Li, X., Cheng, K.T.: Semi-supervised deep regression with uncertainty consistency and variational model ensembling via bayesian neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7304-7313 (2023) +3. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3D point clouds. In: International Conference on Robotics and Automation (ICRA) (2022) +4. Feng, M., Li, Z., Li, Q., Zhang, L., Zhang, X., Zhu, G., Zhang, H., Wang, Y., Mian, A.: Free-form description guided 3D visual graph network for object grounding in point cloud. In: ICCV (2021) +5. Gal, Y., Ghahramani, Z.: Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In: International Conference on Machine Learning (2016) +6. Gal, Y., Islam, R., Ghahramani, Z.: Deep Bayesian active learning with image data. In: International Conference on Machine Learning (2017) + +17. Gao, H.a., Tian, B., Li, P., Zhao, H., Zhou, G.: DQS3D: Densely-matched quantization-aware semi-supervised 3D detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +18. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International Conference on Machine Learning (2017) +19. He, R., Yang, J., Qi, X.: Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +20. Hou, J., Graham, B., Niessner, M., Xie, S.: Exploring data-efficient 3D scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +21. Houlsby, N., Huszar, F., Ghahramani, Z., Lengyel, M.: Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745 (2011) +22. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3D instance segmentation. In: AAAI (2021) +23. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3D visual grounding. In: CVPR (2022) +24. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: ECCV (2022) +25. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +26. Kendall, A., Badrinarayanan, V., Cipolla, R.: Bayesian SegNet: Model uncertainty in deep convolutional encoder-decoder architectures for scene understanding. In: Proceedings of the British Machine Vision Conference (BMVC) (2017) +27. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +28. Kong, C., Lin, D., Bansal, M., Urtasun, R., Fidler, S.: What are you talking about? text-to-image coreference. In: ICCV (2014) +29. Kong, L., Ren, J., Pan, L., Liu, Z.: LaserMix for semi-supervised LiDAR semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +30. Lai, X., Yuan, Y., Chu, R., Chen, Y., Hu, H., Jia, J.: Mask-attention-free transformer for 3D instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +31. Lee, M.S., Yang, S.W., Han, S.W.: Gaia: Graphical information gain based attention network for weakly supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 582-591 (2023) +32. Lei, J., Deng, C., Schmeckpeper, K., Guibas, L., Daniilidis, K.: EFEM: Equivariant neural field expectation maximization for 3D object segmentation without scene supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +33. Liu, K., Zhao, Y., Nie, Q., Gao, Z., Chen, B.M.: Weakly supervised 3D scene segmentation with region-level boundary awareness and instance discrimination. In: European Conference on Computer Vision (ECCV) (2022) + +34. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3D-SPS: Single-stage 3D visual grounding via referred point progressive selection. In: CVPR (2022) +35. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Van Gool, L.: Towards a weakly supervised framework for 3D point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4454-4468 (2021) +36. Prabhudesai, M., Tung, H.Y.F., Javed, S.A., Sieb, M., Harley, A.W., Fragkiadaki, K.: Embodied language grounding with 3D visual feature representations. In: CVPR (2020) +37. Qin, Z., Wang, J., Lu, Y.: Weakly supervised 3D object detection from point clouds. In: Proceedings of the ACM International Conference on Multimedia (2020) +38. Ren, Z., Misra, I., Schwing, A.G., Girdhar, R.: 3D spatial recognition without spatially labeled 3D. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +39. Roh, J., Desingh, K., Farhadi, A., Fox, D.: LanguageRefer: Spatial-language model for 3D visual grounding. In: CoRL (2022) +40. Shannon, C.E.: A mathematical theory of communication. The Bell system technical journal 27(3) (1948) +41. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research 15(1), 1929-1958 (2014) +42. Tang, Y.S., Lee, G.H.: Transferable semi-supervised 3D object detection from rgb-d data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019) +43. Tao, A., Duan, Y., Wei, Y., Lu, J., Zhou, J.: SegGroup: Seg-level supervision for 3D instance and semantic segmentation. IEEE Transactions on Image Processing 31, 4952-4965 (2022) +44. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems (2017) +45. Unal, O., Dai, D., Hoyer, L., Can, Y.B., Van Gool, L.: 2d feature distillation for weakly- and semi-supervised 3d semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7336-7345 (January 2024) +46. Unal, O., Dai, D., Unal, A.T., Van Gool, L.: Discwise active learning for LiDAR semantic segmentation. IEEE Robotics and Automation Letters (2023) +47. Unal, O., Dai, D., Van Gool, L.: Scribble-supervised LiDAR semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +48. Unal, O., Sakaridis, C., Saha, S., Gool, L.V.: Four ways to improve verbo-visual fusion for dense 3d visual grounding (2024), https://arxiv.org/abs/2309.04561 +49. Wang, G., Li, W., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T.: Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing 338, 34-45 (2019) +50. Wang, H., Cong, Y., Litany, O., Gao, Y., Guibas, L.J.: 3D IoUMatch: Leveraging IoU prediction for semi-supervised 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) + +51. Wang, J., Gang, H., Ancha, S., Chen, Y.T., Held, D.: Semi-supervised 3D object detection via temporal graph neural networks. In: International Conference on 3D Vision (3DV) (2021) +52. Wei, J., Lin, G., Yap, K.H., Hung, T.Y., Xie, L.: Multi-path region mining for weakly supervised 3D semantic segmentation on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +53. Wu, Y., Shi, M., Du, S., Lu, H., Cao, Z., Zhong, W.: 3d instances as 1d kernels. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 235-252. Springer (2022) +54. Xia, Y., Liu, F., Yang, D., Cai, J., Yu, L., Zhu, Z., Xu, D., Yuille, A., Roth, H.: 3D semi-supervised learning with uncertainty-aware multi-view co-training. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (March 2020) +55. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10687-10698 (2020) +56. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3D point cloud understanding. In: ECCV (2020) +57. Xu, X., Lee, G.H.: Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +58. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmentation. In: MICCAI (2019) +59. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: InstanceRefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: ICCV (2021) +60. Zhang, F., Fang, J., Wah, B., Torr, P.: Deep fusionnet for point cloud semantic segmentation. In: ECCV (2020) +61. Zhang, Y., Li, Z., Xie, Y., Qu, Y., Li, C., Mei, T.: Weakly supervised semantic segmentation for large-scale point cloud. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021) +62. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3DVG-Transformer: Relation modeling for visual grounding on point clouds. In: ICCV (2021) +63. Zhao, N., Chua, T.S., Lee, G.H.: SESS: Self-ensembling semi-supervised 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +64. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +65. Zoph, B., Ghiasi, G., Lin, T.Y., Cui, Y., Liu, H., Cubuk, E.D., Le, Q.V.: Rethinking pre-training and self-training. arXiv preprint arXiv:2006.06882 (2020) +66. Zou, Y., Yu, Z., Vijaya Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: The European Conference on Computer Vision (ECCV) (2018) \ No newline at end of file diff --git a/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/images.zip b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fee0ab9557341b157815ddf850ee6eae3b2e38c8 --- /dev/null +++ b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3efc81770f9ed94f717d72d5bd5e9260713efb907bc504018e570108bad0e49d +size 572824 diff --git a/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/layout.json b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f896b84d8218fb6030e264e6f335a4f3321fde0f --- /dev/null +++ b/2024/Bayesian Self-Training for Semi-Supervised 3D Segmentation/layout.json @@ -0,0 +1,10040 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 457, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 457, + 147 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 457, + 147 + ], + "type": "text", + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 177, + 167, + 435, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 167, + 435, + 179 + ], + "spans": [ + { + "bbox": [ + 177, + 167, + 435, + 179 + ], + "type": "text", + "content": "Ozan Unal" + }, + { + "bbox": [ + 177, + 167, + 435, + 179 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 177, + 167, + 435, + 179 + ], + "type": "text", + "content": ", Christos Sakaridis" + }, + { + "bbox": [ + 177, + 167, + 435, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 177, + 167, + 435, + 179 + ], + "type": "text", + "content": ", and Luc Van Gool" + }, + { + "bbox": [ + 177, + 167, + 435, + 179 + ], + "type": "inline_equation", + "content": "^{1,3,4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "spans": [ + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "type": "text", + "content": "ETH Zurich, " + }, + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "type": "text", + "content": "Huawei Technologies, " + }, + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "type": "text", + "content": "KU Leuven, " + }, + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 183, + 189, + 430, + 213 + ], + "type": "text", + "content": "INSAIT {ozan.unal, csakarid, vangool}@vision.ee.ethz.ch" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 239, + 452, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 239, + 452, + 469 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 452, + 469 + ], + "type": "text", + "content": "Abstract. 3D segmentation is a core problem in computer vision and, similarly to many other dense prediction tasks, it requires large amounts of annotated data for adequate training. However, densely labeling 3D point clouds to employ fully-supervised training remains too labor intensive and expensive. Semi-supervised training provides a more practical alternative, where only a small set of labeled data is given, accompanied by a larger unlabeled set. This area thus studies the effective use of unlabeled data to reduce the performance gap that arises due to the lack of annotations. In this work, inspired by Bayesian deep learning, we first propose a Bayesian self-training framework for semi-supervised 3D semantic segmentation. Employing stochastic inference, we generate an initial set of pseudo-labels and then filter these based on estimated point-wise uncertainty. By constructing a heuristic " + }, + { + "bbox": [ + 160, + 239, + 452, + 469 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 160, + 239, + 452, + 469 + ], + "type": "text", + "content": "-partite matching algorithm, we extend the method to semi-supervised 3D instance segmentation, and finally, with the same building blocks, to dense 3D visual grounding. We demonstrate state-of-the-art results for our semi-supervised method on SemanticKITTI and ScribbleKITTI for 3D semantic segmentation and on ScanNet and S3DIS for 3D instance segmentation. We further achieve substantial improvements in dense 3D visual grounding over supervised-only baselines on ScanRefer. Our project page is available at ouenal.github.io/bst/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "spans": [ + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "text", + "content": "Keywords: Semi-supervised " + }, + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "text", + "content": " Self-training " + }, + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "text", + "content": " 3D semantic segmentation " + }, + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "text", + "content": " 3D instance segmentation " + }, + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 479, + 452, + 502 + ], + "type": "text", + "content": " 3D visual grounding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 521, + 229, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 229, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 229, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": "Dense 3D semantic perception at the level of points is pivotal for several applications that are based on fine-grained processing and reasoning in 3D, e.g. autonomous cars, robotics, and augmented reality. Such perception requires a segmentation of the unstructured input 3D point cloud into different semantic classes, i.e. 3D semantic segmentation [60], or distinct instances of these classes, i.e., 3D instance segmentation [30]. The input point cloud is often coupled with human utterances that refer to individual objects in the former in cases where a visual AI system interacts with human users in a common 3D environment, necessitating a 3D instance-level segmentation of these objects based on the aforementioned referrals, a task which is known as dense 3D visual grounding or" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 479, + 218 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 479, + 218 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 479, + 218 + ], + "type": "image", + "image_path": "80c238c525cec2778e79dc167ed82162c7a232dfce5551928bdea7297789d1d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "text", + "content": "Fig. 1: Illustration of our Bayesian pseudo-labeling pipeline for semi-supervised a) 3D semantic segmentation; b) 3D instance segmentation; and c) dense 3D visual grounding. With only slight adjustments, using the same building blocks, our method can be adapted to each of these tasks to achieve SOTA results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 479, + 361 + ], + "type": "text", + "content": "referral-based 3D instance segmentation [48, 59]. In the majority of works that address these diverse 3D segmentation tasks, it is assumed that the training data come with full 3D semantic and/or verbal annotations, which creates the pressing need for large-scale labeled 3D datasets [1, 2, 6, 11] so that large end-to-end models can be adequately fitted on these data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 590 + ], + "type": "text", + "content": "However, the creation of such fine-level 3D annotations and verbal referrals for 3D scans requires human labor to be fairly reliable. Such labor is time-consuming and expensive, especially due to the high dimensionality of the scanned point clouds, which makes it hard for human annotators to navigate through them, visually inspect them, and verbally describe them. These practical difficulties render learning frameworks for 3D segmentation that operate with partial labels essential, semi-supervised learning being a primary case. Answering this need, several recent works address semi-supervised 3D semantic segmentation [9,25,29], most of which focus on effective pseudo-labeling techniques for self-training on the larger, unlabeled portion of the training data. However, the generation of pseudo-labels in these works typically focuses on some grouping mechanism over 3D points, such as superpoints, to refine and propagate pseudo-labels and otherwise relies on naïve core techniques such as confidence thresholding [66] for generating the initial point-level pseudo-labels that are then post-processed by their proposed modules. Adding to this, there is a relative scarcity in works on semi-supervised 3D instance segmentation [10, 33], which we attribute to the relative difficulty in pseudo-labeling instances compared to creating plain semantic pseudo-labels. Last but not least, to our knowledge, no work has considered the setting of semi-supervised dense 3D visual grounding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We jointly address these shortcomings of previous works by proposing a common self-training framework for semi-supervised 3D segmentation which is based on Bayesian deep learning and applies to all three aforementioned segmentation tasks. At the core of our self-training framework lies a variational Bayesian approximation of the uncertainty of the respective network's predictions via dropout-based Monte Carlo integration [15] over multiple augmented versions of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "O. Unal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "the original 3D input. This approximation delivers a more informed uncertainty estimate than simple confidence thresholding techniques [66], which is crucial for eliminating confirmation bias from the pseudo-labeling process that we adopt for self-training. In particular, we utilize both the estimated Shannon entropy from the sampled predictions and the predicted label from each sample to filter out uncertain or inconsistent points from the generated set of pseudo-labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 193, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 384 + ], + "type": "text", + "content": "While this formulation is sufficient for the task of 3D semantic segmentation, it does not work out of the box for instance-level segmentation, as Monte Carlo integration cannot be applied directly to instance predictions. To account for this issue, we propose a novel and efficient, heuristic matching strategy across instances coming from different stochastic dropout forward passes based on linear sum assignment. This strategy enables the application of the same filtering steps as described above for semantic pseudo-labels to instance pseudo-labels and thus extends the scope of our approach to 3D instance segmentation. Going one step further, we observe that the same matching strategy that we devise for instance segmentation applies to the selection stage of dense 3D grounding, in which the correct instance from the set of segmented candidate instances needs to be selected based on the input verbal description. Thus, for dense 3D grounding, we apply our heuristic linear sum assignment twice, first for the purely 3D-based candidate instance proposal generation and afterward for the verbo-visual instance selection. Our complete self-training pipeline for each of the three 3D segmentation tasks to which our method applies is summarized in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 389, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 475 + ], + "type": "text", + "content": "We test our method on leading benchmarks across a wide range of domains and datasets. Specifically, our Bayesian self-training framework achieves state-of-the-art results for semi-supervised 3D semantic segmentation on SemanticKITTI [3] and ScribbleKITTI [47] and semi-supervised 3D instance segmentation on ScanNet [11] and S3DIS [2]. Furthermore, we are the first to explore a semi-supervised setting on ScanRefer [6] for 3D visual grounding and obtain consistent significant improvements over supervised-only baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 479, + 332, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 332, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 332, + 490 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 518, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 518, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 518, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 136, + 518, + 479, + 567 + ], + "type": "text", + "content": "1. Following established uncertainty estimation techniques from active learning, we construct a Bayesian self-training framework for semi-supervised 3D semantic segmentation by generating reliable pseudo-labels through stochastic inference and entropy-based uncertainty filtering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 571, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 571, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 136, + 571, + 480, + 594 + ], + "type": "text", + "content": "2. We introduce a novel heuristic matching algorithm to extend the proposed framework to 3D instance segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 600, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 600, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 136, + 600, + 479, + 635 + ], + "type": "text", + "content": "3. We further extend the method to dense 3D visual grounding utilizing the same building blocks as in the above tasks and are the first to explore this task in a semi-supervised setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 641, + 479, + 665 + ], + "type": "text", + "content": "4. We obtain state-of-the-art results for semi-supervised 3D segmentation on SemanticKITTI [3], ScribbleKITTI [47], ScanNet [11] and S3DIS [2]." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 425 + ], + "type": "text", + "content": "Semi-supervised 3D semantic segmentation has recently grown in popularity in the area of 3D understanding along with weakly supervised 3D semantic segmentation [47,52,57,61], due to the difficulty in large-scale 3D labeling that is necessary for the supervised setup [60]. In this work, we tackle the task of semi-supervised segmentation where a large portion of data is completely unlabeled. As opposed to methods developed for weakly-supervised settings, a semi-supervised method cannot leverage the availability of partial true label information within each data sample, e.g. via the ArcPoint loss [31] or a PLS descriptor [47]. A common technique for leveraging unlabeled data under this restrictive scheme is explored under pseudo-labeling. Semantic pseudo-labels are utilized by Jiang et al. [25] to guide a contrastive loss and in SSPC [9] to propagate labels on a superpoint graph. Superpoints are also leveraged by Deng et al. [13] to constrain pseudo-label propagation. In the 2D counterpart task, Chen et al. [8] enforces consistency of semantic pseudo-labels across perturbed versions of the 2D segmentation network. By contrast, we perturb both the network weights and the 3D inputs across different forward passes, and we consider both the entropy and the consistency among predictions in our pseudo-labels. Another related line of 2D semi-supervised works generates pseudo-labels based on direct thresholding of max-softmax scores to reduce confirmation bias [19, 66], while we threshold the entropy of the distribution of softmax scores. Active learning is combined with a 3D semi-supervised approach in DiAL [46] to streamline practical 3D segmentation annotation pipelines. LaserMix [29] introduces variability through cylindrical and range-view partitioning and mixing of 3D LiDAR scans, encouraging consistency across different mixes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": "Semi-supervised 3D instance segmentation is still in its infancy compared to its instance-agnostic semantic segmentation counterpart reviewed above. In particular, most object-level semi-supervised 3D understanding works focus on the coarser task of 3D object detection [17,35,37,38,42,50,51,63], which outputs 3D bounding boxes that are not fine enough for precision-critical downstream tasks such as robotic grasping. For 3D instance/object segmentation, approaches that depart from the standard fully supervised setup have mostly considered the cases of unsupervised pre-training [56], weak supervision via one-point-one-instance [20, 43], or unsupervised generic 3D object segmentation [32]. To our knowledge, the only two semi-supervised works that specifically consider 3D instance segmentation are TWIST [10] and WS3D [33]. The former work proposes a pseudo-label generation strategy tailored for the instance segmentation task, for which the technique of confidence thresholding that is ubiquitously used to the same end in semantic segmentation is not applicable. It thus introduces pseudo offset vectors, besides semantic proposal-level pseudo-labels, for performing self-training. The latter work proposes a region-level contrastive loss, which performs unsupervised instance discrimination, and leverages boundary information via a region-level unsupervised energy-based loss. On the contrary, we propose a unified semi-supervised Bayesian self-training framework that spans multiple 3D segmentation tasks including 3D instance segmentation, and han" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "O. Unal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "dles the additional matching required between instances generated in different forward passes through the network via an efficient heuristic optimization step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 482, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 482, + 276 + ], + "type": "text", + "content": "3D visual grounding is a central 3D task in the area of vision and language [4, 14, 28, 34, 36, 62]. Relevant models accept a 3D point cloud as visual input and perform grounding of the accompanying verbal description to the referred object. While most previous works [7, 23, 24, 34, 39] focus on coarse 3D bounding box outputs, the task of dense 3D visual grounding or referral-based 3D instance segmentation has been explored far less in comparison, in which a fine-grained point-level mask is predicted [22, 48, 59]. Regardless, all aforementioned methods invariably assume a supervised setting, where full ground-truth instance annotations are available for all samples at training time. By contrast, the formulation of our method makes it fit for also handling the more practical semi-supervised setting, in which only a part of the training data is labeled." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 495 + ], + "type": "text", + "content": "Uncertainty estimation is crucial in learning setups that lack full supervision as ours, given that such estimates can provide information about the reliability of the predictions of the model on unlabeled samples [21]. In the more recent deep learning era, Bayesian neural networks form the core for uncertainty estimation by modeling network weights as stochastic distributions, where test-time-dropout [41] is utilized as a variational Bayesian approximation [15] and Monte Carlo integration is computed via multiple stochastic forward passes to approximate prediction uncertainty [16]. This Bayesian deep learning framework has been applied successfully to supervised 2D dense prediction tasks [26, 27]. We focus on the aggregate uncertainty in the predictions of a general 3D segmentation model and build on the Bayesian deep learning approach of Gal et al. [16] which examines the approximate predictive entropy as an acquisition function for active learning. However, contrary to Gal et al. [16], in our self-training setup, we use low rather than high entropy as a positive indication for considering a point as a source of supervision. To the best of our knowledge, our work is the first to introduce such a Bayesian self-training framework for general 3D point cloud segmentation, specifically including 3D instance segmentation and referral-based 3D instance segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "content": "Related uncertainty-based approaches to ours have been proposed for the semantic segmentation of 3D grid volume data in the field of medical image computing [12,54,58]. In particular, Yu et al. [58] also use Monte Carlo dropout to estimate predictive entropy but only applies dropout to a single network input, whereas we vary both network weights and input views. Moreover, that work leverages the estimated entropy merely as the weight in a consistency loss between a teacher and a student network, while we filter pseudo-labels based on our entropy estimates. Besides, Xia et al. [54] perform semi-supervised semantic segmentation with multi-view co-training and applies Bayesian deep learning across views for uncertainty-based weighting of pseudo-labels. We only train one network instead of multiple networks employed in Xia et al. [54] and we derive the uncertainty estimates by feeding all input views to that single network, which is more parameter-efficient. Both of these approaches only address 3D semantic segmentation, and cannot be utilized in instance-level dense prediction tasks." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 115, + 474, + 238 + ], + "blocks": [ + { + "bbox": [ + 140, + 115, + 474, + 238 + ], + "lines": [ + { + "bbox": [ + 140, + 115, + 474, + 238 + ], + "spans": [ + { + "bbox": [ + 140, + 115, + 474, + 238 + ], + "type": "image", + "image_path": "33aa6741a249313c163f3dde3e2408db3f6c1b7f521ff2908ff9085a622b7eda.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "text", + "content": "Fig. 2: Illustration of Bayesian pseudo-labeling for semi-supervised instance segmentation. We initialize a seed prediction via a forward pass using a non-augmented input. Then our heuristic " + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "text", + "content": "-partite matching algorithm is employed to pair each seed instance with the best matching predicted instance from each of the " + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "text", + "content": " stochastic forward passes. For each aligned object, we compute the aggregated label through unanimous voting and filter it based on uncertainty to obtain the final pseudo mask." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 326, + 288, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 288, + 341 + ], + "type": "text", + "content": "3 Bayesian Self-Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 425 + ], + "type": "text", + "content": "In Sec. 3.1 we start by formulating a self-training approach for semi-supervised 3D semantic segmentation based on established uncertainty estimation techniques from the field of active learning [15,46]. We then extend our formulation to include instance labels in Sec. 3.2 and finally build up to also take natural language prompts as input, i.e., towards semi-supervised dense 3D visual grounding in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 446, + 384, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 384, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 384, + 459 + ], + "type": "text", + "content": "3.1 Semi-Supervised 3D Semantic Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "text", + "content": "In a semi-supervised setting, a dataset " + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "text", + "content": " consists of a labeled set " + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "text", + "content": " and an unlabeled set " + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "text", + "content": ". The baseline approach when working with such data is to treat the problem similarly to any fully supervised task and employ a semantic loss " + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sem}" + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "text", + "content": " on the available labeled scenes " + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "text", + "content": ", typically stated as cross-entropy. Formally, the objective function over the model parameters " + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 468, + 482, + 529 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 262, + 537, + 481, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 537, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 262, + 537, + 481, + 572 + ], + "type": "interline_equation", + "content": "\\min _ {\\omega} \\sum_ {i = 1} ^ {| S |} \\mathcal {L} _ {s e m} (\\hat {\\mathbf {y}}, \\mathbf {y}). \\tag {1}", + "image_path": "ef469c6f2bd225059cdad7b1c65ab265908aa7380e4d83b8a221e8e76663364b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Starting from this baseline formulation, a semi-supervised training strategy aims to leverage the unlabeled scenes " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " to further improve model performance. One such common technique is self-training, where the model is iteratively trained on its own predictions, i.e. on generated per-point one-hot pseudo-label vectors " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}} = [\\hat{\\mathbf{y}}^{(1)},\\dots,\\hat{\\mathbf{y}}^{(C)}]" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " denoting the number of classes. However, the task of generating pseudo-labels is not trivial, as the model may suffer from confirmation bias that occurs when repeatedly overfitting to incorrect labels." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "O. Unal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 187 + ], + "type": "text", + "content": "A simple but effective solution is explored under the umbrella of thresholding [5, 19, 55, 65, 66]. In essence, instead of training on all model predictions, a pseudo-label set " + }, + { + "bbox": [ + 131, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 116, + 482, + 187 + ], + "type": "text", + "content": " is formed by predictions that exceed a confidence of a determined threshold " + }, + { + "bbox": [ + 131, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 116, + 482, + 187 + ], + "type": "text", + "content": " to ensure a higher level of accuracy. Formally, the objective function of threshold-based self-training for semantic segmentation can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 218, + 191, + 481, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 191, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 218, + 191, + 481, + 225 + ], + "type": "interline_equation", + "content": "\\min _ {\\omega , \\hat {\\mathbf {y}}} \\sum_ {d = 1} ^ {| D |} \\left(\\mathcal {L} _ {s e m} (\\hat {\\mathbf {y}}, \\mathbf {y}) - (\\log \\hat {\\mathbf {y}} _ {d} - \\tau) \\cdot \\hat {\\mathbf {y}} _ {d}\\right) \\tag {2}", + "image_path": "53fe7c448f9e8b995e0612c5be7d1c3ac818e8ce9e501c3d5c5d121ad1c0e861.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 229, + 482, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 482, + 253 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 229, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_d" + }, + { + "bbox": [ + 131, + 229, + 482, + 253 + ], + "type": "text", + "content": " denoting the softmax model predictions. Eq. 2 is optimized in two steps that are repeated until convergence:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 256, + 481, + 304 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 256, + 481, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 256, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 136, + 256, + 481, + 280 + ], + "type": "text", + "content": "1. Training: First, " + }, + { + "bbox": [ + 136, + 256, + 481, + 280 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}" + }, + { + "bbox": [ + 136, + 256, + 481, + 280 + ], + "type": "text", + "content": " is fixed, and the objective function is optimized with respect to " + }, + { + "bbox": [ + 136, + 256, + 481, + 280 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 136, + 256, + 481, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "text", + "content": "2. Pseudo-labeling: Then, " + }, + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "text", + "content": " is fixed and the objective function is optimized with respect to " + }, + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}" + }, + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "text", + "content": ". Consequently, " + }, + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "text", + "content": " is updated given " + }, + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}" + }, + { + "bbox": [ + 135, + 280, + 481, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 308, + 481, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 481, + 356 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 481, + 356 + ], + "type": "text", + "content": "The process is initialized by setting the latent variable " + }, + { + "bbox": [ + 131, + 308, + 481, + 356 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}} = \\mathbf{0}" + }, + { + "bbox": [ + 131, + 308, + 481, + 356 + ], + "type": "text", + "content": " for all points, i.e. by reverting to the naive solution of Eq. 1 and only training on the labeled set " + }, + { + "bbox": [ + 131, + 308, + 481, + 356 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 308, + 481, + 356 + ], + "type": "text", + "content": ". The two steps can be repeated to take advantage of the improved representation capability of the model thanks to pseudo-labeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 356, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 356, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 356, + 482, + 428 + ], + "type": "text", + "content": "While naïve thresholding relies on the softmax scores of the model to determine confidence, these scores may be unreliable and often skewed when facing long-tailed distributions, resulting in overconfident predictions [18]. Inspired by uncertainty estimation techniques commonly employed in active learning (AL) pipelines [21,46], we introduce a Bayesian self-training strategy that leverages an uncertainty-aware filtering technique for generating more reliable pseudo-labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "text", + "content": "We first start by decomposing the goal of 3D semantic segmentation, which is to discover the dependency of the point-wise distribution over the labels " + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "inline_equation", + "content": "y \\in Y" + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "text", + "content": " on an input sample " + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "inline_equation", + "content": "x \\in X" + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "text", + "content": " via the model weights " + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "text", + "content": ". Formally, the task can be reduced to per-point classification, where " + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "text", + "content": " can be of class " + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "inline_equation", + "content": "c \\in \\{1, \\dots, C\\}" + }, + { + "bbox": [ + 131, + 428, + 482, + 487 + ], + "type": "text", + "content": ". The conditional probability is therefore defined by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 188, + 492, + 481, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 492, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 188, + 492, + 481, + 517 + ], + "type": "interline_equation", + "content": "p (y = c | x, D) = \\int p (y = c | a (x), \\omega) p (\\omega | D) p (a (x) | x) d \\omega d x \\tag {3}", + "image_path": "32ebff61b56c174a1d3d4c761147a026565537bbea49fac6bc9d31e6cf8d2aeb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 521, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 481, + 544 + ], + "type": "text", + "content": "for the dataset " + }, + { + "bbox": [ + 131, + 521, + 481, + 544 + ], + "type": "inline_equation", + "content": "D = (X, Y)" + }, + { + "bbox": [ + 131, + 521, + 481, + 544 + ], + "type": "text", + "content": ", under randomly applied affine transformations modeled through " + }, + { + "bbox": [ + 131, + 521, + 481, + 544 + ], + "type": "inline_equation", + "content": "a(\\cdot)" + }, + { + "bbox": [ + 131, + 521, + 481, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "p(a(x)|x)" + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "text", + "content": " as a constant, restate the intractable " + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "p(\\omega |D)" + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "text", + "content": " via the variational approximation " + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "q(\\omega)" + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "text", + "content": " and minimize " + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(q(\\omega)|p(\\omega |X,Y))" + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "text", + "content": ". This objective can be approximated through Monte Carlo integration by employing variational inference via Bayesian networks, i.e. by applying stochastic forward passes with weights " + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\omega_{k}" + }, + { + "bbox": [ + 131, + 544, + 482, + 605 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 211, + 609, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 609, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 211, + 609, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p (y = c | x, D) \\approx \\int p (y = c | a (x), \\omega) q (\\omega) d \\omega d x \\\\ \\approx \\frac {1}{K} \\sum_ {k = 1} ^ {K} p (y = c | a (x), \\omega_ {k}). \\tag {4} \\\\ \\end{array}", + "image_path": "5806730727b05efa6656f4b75f0e772ba885082aa4ab2afc8d9ef7a3d38396aa.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "text", + "content": "Compared to the common decomposition of BALD [21], the inclusion of the affine transformations via " + }, + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "a(\\cdot)" + }, + { + "bbox": [ + 132, + 116, + 482, + 165 + ], + "type": "text", + "content": " allows a higher deviation amongst stochastic forward passes which in return allows better estimation of predictive uncertainty [49]. We capture the uncertainty via the Shannon entropy [40]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 174, + 481, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 174, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 202, + 174, + 481, + 206 + ], + "type": "interline_equation", + "content": "\\mathbb {H} [ y | x, D ] = - \\sum_ {c = 1} ^ {C} p (y = c | x, D) \\log p (y = c | x, D) \\tag {5}", + "image_path": "0f8fa5ad5b4dfd2932b96114ede724491700685d80166ef66e7c3f6186266c50.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 215, + 481, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 215, + 481, + 238 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 481, + 238 + ], + "type": "text", + "content": "and the accumulated predicted label " + }, + { + "bbox": [ + 131, + 215, + 481, + 238 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^*" + }, + { + "bbox": [ + 131, + 215, + 481, + 238 + ], + "type": "text", + "content": " is given by the unanimous voting principle:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 187, + 239, + 481, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 239, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 187, + 239, + 481, + 270 + ], + "type": "interline_equation", + "content": "\\hat {y} ^ {* (c)} = \\left\\{ \\begin{array}{l} 1, \\text {i f} K = \\sum_ {k} \\mathbb {1} (\\operatorname {a r g m a x} (p (y | a (x), \\omega_ {k})) = c) \\\\ 0, \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {6}", + "image_path": "84020855b76b2314ebf13528b6b2f66ae7f2442eb9601e6a34f1283e3c559914.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 305, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 305, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 305, + 285 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 132, + 274, + 305, + 285 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 132, + 274, + 305, + 285 + ], + "type": "text", + "content": " denoting the indicator function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 286, + 481, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 286, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 481, + 334 + ], + "type": "text", + "content": "Unlike what is commonly seen in active learning settings, our goal is not to maximize the total entropy but rather to minimize it in order to ensure a high-quality pseudo-label set. We therefore redefine the self-training objective (Eq. 2) for stochastic pseudo-labeling as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 204, + 343, + 481, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 343, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 204, + 343, + 481, + 376 + ], + "type": "interline_equation", + "content": "\\min _ {\\omega , \\hat {\\mathbf {y}}} \\sum_ {d = 1} ^ {D} \\left(\\mathcal {L} _ {s e m} (\\hat {\\mathbf {y}}, \\mathbf {y}) + \\left(\\mathbb {H} [ y | x, D ] - \\tau\\right) \\left(\\hat {\\mathbf {y}} ^ {*} \\cdot \\hat {\\mathbf {y}}\\right)\\right) \\tag {7}", + "image_path": "659304d97fd1a14b44a6524fe4889516d631397b5895137866c2166ef3afd90e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 384, + 481, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 481, + 409 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 384, + 481, + 409 + ], + "type": "inline_equation", + "content": "\\tau \\in (0,1)" + }, + { + "bbox": [ + 131, + 384, + 481, + 409 + ], + "type": "text", + "content": " determining the cut-off entropy threshold. To solve the nonlinear integer optimization task, we employ the following solver:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 418, + 481, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 418, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 233, + 418, + 481, + 462 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {y}} ^ {(c)} = \\left\\{ \\begin{array}{l} 1, \\text {i f} \\mathbb {H} [ y | x, D ] < \\tau \\text {a n d} \\\\ c = \\operatorname {a r g m a x} \\hat {\\mathcal {y}} ^ {*} \\\\ 0, \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {8}", + "image_path": "308fbceb0705786131183ffb31abf07cef15a715ce2d894c7cb6eddfcf7bdb74.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 471, + 366, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 366, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 366, + 483 + ], + "type": "text", + "content": "where labels " + }, + { + "bbox": [ + 132, + 471, + 366, + 483 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^{(c)} = 0\\forall c" + }, + { + "bbox": [ + 132, + 471, + 366, + 483 + ], + "type": "text", + "content": " are ignored during training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 484, + 481, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 481, + 519 + ], + "type": "text", + "content": "In practice, following precedent [47,66], we set " + }, + { + "bbox": [ + 130, + 484, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 484, + 481, + 519 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 130, + 484, + 481, + 519 + ], + "type": "inline_equation", + "content": "p_{\\tau}" + }, + { + "bbox": [ + 130, + 484, + 481, + 519 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 484, + 481, + 519 + ], + "type": "inline_equation", + "content": "p_{\\tau}" + }, + { + "bbox": [ + 130, + 484, + 481, + 519 + ], + "type": "text", + "content": " determining the percentage of pseudo-labels to be sampled amongst the accumulated predicted labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 520, + 481, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 520, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 131, + 520, + 481, + 591 + ], + "type": "text", + "content": "In summary, we first determine a set of labels that are agreed upon by all stochastic forward passes, i.e. by forward passes with active dropout [15, 41]. We then compute the Shannon entropy for all predictions, filtering those that exhibit high entropy. This surprisingly simple but effective approach allows us to reduce the incorrect targets that we introduce to our pseudo-label set and to avoid the overconfidence issue of naive thresholding-based approaches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 609, + 380, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 380, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 380, + 623 + ], + "type": "text", + "content": "3.2 Semi-Supervised 3D Instance Segmentation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": "The proposed Bayesian self-training framework provides a strong basis for any semi-supervised 3D dense prediction task, one such example being 3D instance segmentation. Yet, compared to semantic segmentation where accumulation of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "O. Unal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 133, + 142, + 465, + 333 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 452, + 141 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 452, + 141 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 452, + 141 + ], + "type": "text", + "content": "Algorithm 1: Bayesian pseudo-label generation for instance segmentation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "lines": [ + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "spans": [ + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": "Input: Point cloud " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " , trained model with dropout " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " , number of passes " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " , threshold " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " \n1 seed_output " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "= \\omega (P)\\in \\mathbb{R}^{M\\times N}" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " \n2 " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "\\pmb{i} =" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " seed_output.copy() \n3 seed_mask " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " to_mask(i) " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "\\in \\{0,1\\}^{M\\times N}" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " \n4 unanimous " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " seed_mask.copy() \n5 for k in 1:K do \n6 output " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "= \\omega_{k}(a(P))\\in \\mathbb{R}^{M^{\\prime}\\times N}" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " \n7 mask " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " to_mask(output) " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "\\in \\{0,1\\}^{M^{\\prime}\\times N}" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " \n8 cost " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " -iou(seed_mask, mask) \n9 row,col " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " LSA(cost) \n10 unanimous[col] &= mask[row] \n11 i[col] " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "+ =" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " output[row] \n12 " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "\\mathbb{H} = (K - i)\\cdot \\log (K - i) + i\\log i" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " \n13 pseudo_label " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "= (\\mathbb{H} < \\tau)" + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "text", + "content": " & unanimous \nReturn:pseudo_label " + }, + { + "bbox": [ + 133, + 142, + 465, + 333 + ], + "type": "inline_equation", + "content": "\\in \\{0,1\\}^{M\\times N}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 359, + 480, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 480, + 383 + ], + "type": "text", + "content": "labels can be easily handled across multiple stochastic forward passes, accumulation of predicted instance masks is not trivial and requires " + }, + { + "bbox": [ + 130, + 359, + 480, + 383 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 359, + 480, + 383 + ], + "type": "text", + "content": "-partite matching." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "text", + "content": "However, the " + }, + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "text", + "content": "-partite matching problem is NP-hard with a complexity of " + }, + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "inline_equation", + "content": "|I|!(K-1)" + }, + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "text", + "content": " passes and " + }, + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "inline_equation", + "content": "|I|" + }, + { + "bbox": [ + 130, + 384, + 481, + 491 + ], + "type": "text", + "content": " number of instances assuming full bijection. To match predicted instance masks across various stochastic forward passes, we propose a heuristic method based on the Hungarian algorithm. First, we generate an initial prediction via an unaugmented forward pass (Alg. 1 L1&3). The goal is to establish a seed that is likely of a higher quality with minimal perturbations of the data. We then iteratively assign predicted masks of each stochastic forward pass via linear sum assignment, with the cost matrix formed by the negative pairwise IoUs (Alg. 1 L5-10)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 481, + 552 + ], + "type": "text", + "content": "Having matched predicted binary instance masks from individual stochastic forward passes, each resulting set of masks can then be treated as a set of 2-class semantic segmentation outputs, and thus our proposed Bayesian pseudo-labeling approach from Eq. 8 can be trivially applied (Alg. 1 L11-15). Our pipeline for semi-supervised 3D instance segmentation is illustrated in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 571, + 389, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 389, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 389, + 583 + ], + "type": "text", + "content": "3.3 Semi-Supervised Dense 3D Visual Grounding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 593, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 480, + 640 + ], + "type": "text", + "content": "Similar to both semi-supervised 3D semantic and 3D instance segmentation, our proposed Bayesian self-training formulation can easily be extended to include natural language prompts as inputs and thus tackle semi-supervised dense 3D visual grounding, i.e., referral-based 3D instance segmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "A common technique when tackling dense 3D visual grounding is to employ a grounding-by-selection strategy which can be summarised in two steps [6,48,62]:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 135, + 116, + 480, + 163 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 135, + 116, + 479, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 116, + 479, + 127 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 479, + 127 + ], + "type": "text", + "content": "1. Visual: First, a 3D backbone computes object candidates " + }, + { + "bbox": [ + 135, + 116, + 479, + 127 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 135, + 116, + 479, + 127 + ], + "type": "text", + "content": " from the scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 128, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 128, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 135, + 128, + 480, + 163 + ], + "type": "text", + "content": "2. Verbo-visual: Then, a verbo-visual fusion module " + }, + { + "bbox": [ + 135, + 128, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\omega_{sel}" + }, + { + "bbox": [ + 135, + 128, + 480, + 163 + ], + "type": "text", + "content": " selects the correct instance candidate from predicted instances " + }, + { + "bbox": [ + 135, + 128, + 480, + 163 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 135, + 128, + 480, + 163 + ], + "type": "text", + "content": " based on the natural language description " + }, + { + "bbox": [ + 135, + 128, + 480, + 163 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 135, + 128, + 480, + 163 + ], + "type": "text", + "content": ", i.e., it predicts the index:" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 218, + 171, + 480, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 171, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 218, + 171, + 480, + 185 + ], + "type": "interline_equation", + "content": "i ^ {*} = \\operatorname {a r g m a x} G, \\text {w i t h} G = \\omega_ {s e l} (O, W) \\in \\mathbb {R} ^ {| I |}, \\tag {9}", + "image_path": "83f4dee5659bab60d6737c67fbe80e3f2b8889b7af96c7d40ffa9c588999603f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 193, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 479, + 228 + ], + "type": "text", + "content": "To extend our Bayesian self-training approach to this task, we follow the above structure and individually adapt each of the two steps, with the adaptation of the visual step simply following Sec. 3.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 229, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 480, + 312 + ], + "type": "text", + "content": "For the verbal side of 3D visual grounding, i.e. referral-based instance selection, we observe that the pseudo-label generation can be reduced to the same formulation as the pseudo-label generation for 3D instance segmentation. Utilizing the results from the heuristic " + }, + { + "bbox": [ + 130, + 229, + 480, + 312 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 229, + 480, + 312 + ], + "type": "text", + "content": "-partite matching (Alg. 1 L9), vector " + }, + { + "bbox": [ + 130, + 229, + 480, + 312 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 229, + 480, + 312 + ], + "type": "text", + "content": " can be reordered to match across stochastic forward passes. Again, having matched predicted probability vectors, the method presented in Sec. 3.1 can be trivially applied (Alg. 1 L11-15)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 230, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 230, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 230, + 342 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": "We evaluate our method on all three semi-supervised 3D semantic perception tasks: 3D semantic segmentation, 3D instance segmentation, and dense 3D visual grounding. The dataset and implementation details regarding each task can be found in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 297, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 297, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 297, + 431 + ], + "type": "text", + "content": "4.1 3D Semantic Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "content": "The comparison to the state of the art for semi-supervised 3D semantic segmentation is shown in Tab. 1. As seen, our Bayesian self-training strategy outperforms existing work in both SemanticKITTI [3] and ScribbleKITTI [47] on splits of " + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "content": ", while performing on par for " + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "content": ". Specifically, compared to the baseline Cylinder3D [64] model denoted as Sup-only, our strategy allows significant improvements ranging from " + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "inline_equation", + "content": "10.9\\%" + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "content": " mIoU. In Fig. 3 (a-c) we showcase qualitative results on SemanticKITTI " + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 539, + 293, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 293, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 293, + 551 + ], + "type": "text", + "content": "4.2 3D Instance Segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "text", + "content": "For semi-supervised 3D instance segmentation, we report results for ScanNet [11] in Tab. 2 and for S3DIS [2] in Tab. 3. As seen, we observe a similar behavior to semi-supervised 3D semantic segmentation, where at higher thresholds of " + }, + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "text", + "content": " for ScanNet and " + }, + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 557, + 480, + 665 + ], + "type": "text", + "content": " for S3DIS, our simple yet effective Bayesian self-training strategy significantly outperforms state-of-the-art methods that require highly engineered solutions. In contrast, at lower thresholds, we see a significant gap in performance (especially for ScanNet). Still, compared to the baseline Sup-only, we see significant gains across the board for all semi-splits. Additionally in Fig. 3 (d-f), we show qualitative results on the ScanNet val-set." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "O. Unal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 178, + 163, + 430, + 355 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 1: Comparison of state-of-the-art methods for semi-supervised LiDAR semantic segmentation (mIoU [%]). For fair comparison, all methods use the same backbone given by Sup-only, which at " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": " labels achieves " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "64.3\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": " mIoU on SemanticKITTI and " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "57.0\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": " mIoU on ScribbleKITTI." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 178, + 163, + 430, + 355 + ], + "lines": [ + { + "bbox": [ + 178, + 163, + 430, + 355 + ], + "spans": [ + { + "bbox": [ + 178, + 163, + 430, + 355 + ], + "type": "table", + "html": "
Labeled1%10%20%50%
SemantickKITTI [3]Sup-only [64]45.456.157.858.7
DiAL [46]45.457.159.260.0
CBST [66]48.858.359.459.7
CPS [8]46.758.759.660.5
LaserMix [29]50.660.061.962.3
IGNet [45]49.061.363.164.8
Ours49.861.763.764.1
Δ Sup-only+4.4+5.6+5.9+5.4
ScribbleKITTI [47]Sup-only [64]39.248.052.153.8
DiAL [46]41.050.152.853.9
CBST [66]41.550.653.354.5
CPS [8]41.451.853.954.8
LaserMix [29]44.253.755.156.8
IGNet [45]44.457.759.660.8
Ours43.958.960.161.2
Δ Sup-only+4.7+10.9+8.0+7.4
", + "image_path": "0c1e3aca2b707086eaa2c4a75b30a25a72dbc58fbfdf01b07c8f43ae6d91bed1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 399, + 478, + 504 + ], + "blocks": [ + { + "bbox": [ + 130, + 361, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 396 + ], + "type": "text", + "content": "Table 2: Comparison to state-of-the-art methods for semi-supervised 3D instance segmentation on the ScanNet Limited Reconstruction valid-set. For a fair comparison, all methods apart from that marked with * use the same backbone [20,53] as Sup-only." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 399, + 478, + 504 + ], + "lines": [ + { + "bbox": [ + 133, + 399, + 478, + 504 + ], + "spans": [ + { + "bbox": [ + 133, + 399, + 478, + 504 + ], + "type": "table", + "html": "
Labeled Metric1% mAPAP50AP255% mAPAP50AP2510% mAPAP50AP2520% mAPAP50AP25
Sup-only [10]5.19.817.618.232.047.026.742.858.929.347.963.0
PointCont. [56]7.212.520.319.435.448.527.043.959.530.249.563.6
CSC [20]7.113.021.220.936.750.627.345.060.230.650.364.1
TWIST [10]9.617.126.227.044.156.230.649.763.032.852.966.8
WS3D* [33]-32.5--45.6--49.2--51.3-
Ours7.214.223.024.239.351.432.751.965.537.456.168.9
Δ Sup-only+2.1+4.4+5.4+6.0+7.3+4.4+6.0+9.1+6.6+8.1+8.2+5.9
", + "image_path": "5a9510bcb23fed4d595725be4d2b078029c69cb34debc971e4d15ac876b6915c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 512, + 304, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 304, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 304, + 525 + ], + "type": "text", + "content": "4.3 Dense 3D Visual Grounding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 569 + ], + "type": "text", + "content": "For semi-supervised 3D visual grounding, we report results on the ScanRefer [6] dataset. As seen in Tab. 4 (Vi), our self-training approach shows improvements across the board, with significant gains on the \"unique\" category." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "content": "While natural language prompts still require human effort to generate, they are considerably cheaper and easier to acquire than per-point high-quality panoptic labels. In Tab. 4 (ViVe), we show that the inclusion of such prompts in the unlabeled dataset further benefits the model's accuracy. As seen, we observe even larger gains over the baseline method with the overall accuracy (OA) showing " + }, + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "inline_equation", + "content": "+4.7\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "content": " improvement with a " + }, + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "content": " IoU threshold." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Furthermore, in Fig. 3 (g) we qualitatively compare our method to the baseline Sup-only on ScanRefer." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 295, + 257 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 295, + 257 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 295, + 257 + ], + "type": "image", + "image_path": "17435439459c2fdcb7ba5035be734083ab12ecacf852b75174414faf883fc9be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 266, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 301 + ], + "type": "text", + "content": "Fig.3: Qualitative results from the SemanticKITTI (10%) val-set, comparing a) the ground truth, b) Sup-only and c) ours; on ScanNet (10%), comparing d) the ground-truth instance masks, e) Sup-only and f) ours; and finally g) ScanRefer (10%)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 116, + 386, + 186 + ], + "blocks": [ + { + "bbox": [ + 298, + 116, + 386, + 186 + ], + "lines": [ + { + "bbox": [ + 298, + 116, + 386, + 186 + ], + "spans": [ + { + "bbox": [ + 298, + 116, + 386, + 186 + ], + "type": "image", + "image_path": "a3d0b6e442955374224e691faa8599fd07a346a34ff644a93aa7b0f97345270d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 299, + 186, + 386, + 255 + ], + "blocks": [ + { + "bbox": [ + 299, + 186, + 386, + 255 + ], + "lines": [ + { + "bbox": [ + 299, + 186, + 386, + 255 + ], + "spans": [ + { + "bbox": [ + 299, + 186, + 386, + 255 + ], + "type": "image", + "image_path": "ac34a78f41280f5d2af4764e42562d5eb617bf59bb496b79827d6b48c7e07b94.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 389, + 118, + 477, + 186 + ], + "blocks": [ + { + "bbox": [ + 389, + 118, + 477, + 186 + ], + "lines": [ + { + "bbox": [ + 389, + 118, + 477, + 186 + ], + "spans": [ + { + "bbox": [ + 389, + 118, + 477, + 186 + ], + "type": "image", + "image_path": "234c063d3d34d436477c3eeb1231824c17a3ae258907b07a84541d5debd887ad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 390, + 187, + 477, + 240 + ], + "blocks": [ + { + "bbox": [ + 390, + 187, + 477, + 240 + ], + "lines": [ + { + "bbox": [ + 390, + 187, + 477, + 240 + ], + "spans": [ + { + "bbox": [ + 390, + 187, + 477, + 240 + ], + "type": "image", + "image_path": "59ed4a2de850b2b756a79bfaca50b8c0abf650d965fe1226a435cba4a9343518.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 241, + 474, + 251 + ], + "lines": [ + { + "bbox": [ + 391, + 241, + 474, + 251 + ], + "spans": [ + { + "bbox": [ + 391, + 241, + 474, + 251 + ], + "type": "text", + "content": "The table is square. There is a chair to the left, and a chair to the right of it." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 176, + 348, + 432, + 429 + ], + "blocks": [ + { + "bbox": [ + 130, + 309, + 482, + 343 + ], + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 343 + ], + "type": "text", + "content": "Table 3: Comparison to state-of-the-art methods for semi-supervised 3D instance segmentation on S3DIS Area 5 via AP50 [%]. For a fair comparison, all methods use the same backbone Sparse U-Net [20,53] as Sup-only." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 348, + 432, + 429 + ], + "lines": [ + { + "bbox": [ + 176, + 348, + 432, + 429 + ], + "spans": [ + { + "bbox": [ + 176, + 348, + 432, + 429 + ], + "type": "table", + "html": "
Labeled5%10%20%50%
Sup-only [10]30.436.841.246.5
PointContrast [56]33.638.743.148.9
CSC [20]34.241.044.750.4
TWIST [10]37.145.648.4-
Ours35.042.951.156.6
Δ Sup-only+4.6+6.1+9.9+10.1
", + "image_path": "525d8a8ab28cddfd8e88d2ebaeb664ed073ac65d7009594b39368d6ce59c8463.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 445, + 246, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 246, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 246, + 456 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 472, + 481, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 508 + ], + "type": "text", + "content": "We conduct ablation studies on " + }, + { + "bbox": [ + 130, + 472, + 481, + 508 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 472, + 481, + 508 + ], + "type": "text", + "content": " splits of SemanticKITTI [3], ScanNet [11] and ScanRefer [6] for semi-supervised 3D semantic segmentation, instance segmentation and visual grounding respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Entropy Thresholding: In Tab. 5, we show an analysis of the effectiveness of our proposed uncertainty thresholding via Shannon entropy, comparing it to the baseline unanimous voting results as well as established thresholding techniques of softmax thresholding [55] and class-balanced thresholding [66]. Alongside the performance for each individual task, we also report the accuracy of the generated pseudo-labels. For instance segmentation, a pseudo-mask is considered correct if its intersection with the associated ground-truth mask is identical to the former. For visual grounding, we reduce the accuracy threshold to 0.5 following the benchmark metric used. As seen, our uncertainty-based filtering improves pseudo-label quality across the board. We also showcase the performance implications of uncertainty filtering, where it can be observed that entropy thresholding outperforms both unanimous voting and current thresholding strategies across all tasks. We further provide a qualitative analysis of the uncertainty met-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "O. Unal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 183, + 163, + 424, + 232 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: Results for semi-supervised 3D dense visual grounding while assuming (Vi) that the model is given no natural language descriptions for the unlabeled split and (ViVe) when the model has access to natural language descriptions for the unlabeled split. Reported values are given as Acc@25/50 for the " + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": " split." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 163, + 424, + 232 + ], + "lines": [ + { + "bbox": [ + 183, + 163, + 424, + 232 + ], + "spans": [ + { + "bbox": [ + 183, + 163, + 424, + 232 + ], + "type": "table", + "html": "
CategoryUniqueMultipleOverall
Sup-only [48]51.9 / 43.619.3 / 14.125.4 / 19.6
ViMe ViOurs58.8 / 48.822.0 / 18.228.8 / 23.9
Δ Sup-only+6.9 / +5.2+2.7 / +4.1+3.4 / +4.3
Ours61.0 / 51.223.4 / 18.130.4 / 24.3
Δ Sup-only+9.1 / +7.6+4.1 / +4.0+5.0 / +4.7
", + "image_path": "2d66b242a3f41c7b1c0c582a90bfc737e1450cb26559fb9eb29a0c8051b8b53c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 300, + 478, + 392 + ], + "blocks": [ + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "text", + "content": "Table 5: Comparison of the pseudo-label quality of current thresholding strategies and their effects on performance on SemanticKITTI " + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "text", + "content": ", ScanNet " + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "text", + "content": ", ScanRefer " + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "text", + "content": " (with visual and verbal) for each respective task. Pseudo-label thresholding is applied after unanimous voting for a fair comparison. The letters indicate each pseudolabel type: " + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "inline_equation", + "content": "S-" + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "text", + "content": " for semantic, " + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "inline_equation", + "content": "I-" + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "text", + "content": " for instance, and " + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "inline_equation", + "content": "G-" + }, + { + "bbox": [ + 131, + 239, + 479, + 294 + ], + "type": "text", + "content": " for grounding." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 300, + 478, + 392 + ], + "lines": [ + { + "bbox": [ + 133, + 300, + 478, + 392 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 478, + 392 + ], + "type": "table", + "html": "
SemanticInstanceVisual Grounding
MethodS-AccmIoUS-AccI-AccAP@50S-AccI-AccG-AccOA@50
Sup-only [48]-56.1--42.8---19.6
Unanimous96.460.079.375.347.072.969.961.516.9
Naïve [5,55,65]98.360.882.977.750.282.073.068.621.9
Class-balanced [66]96.661.181.270.144.480.969.060.616.2
Entropy-based Δ Unanimous98.661.784.877.851.983.875.275.624.3
+2.2+1.7+5.5+2.5+4.9+5.8+2.3+5.7+7.4
", + "image_path": "49900eaa5474a7613075770dbe8494e6182581e3bd069675fc976ebb3a8cebab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 405, + 481, + 462 + ], + "blocks": [ + { + "bbox": [ + 133, + 405, + 481, + 462 + ], + "lines": [ + { + "bbox": [ + 133, + 405, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 133, + 405, + 481, + 462 + ], + "type": "image", + "image_path": "3d92ef1074863bd45a8f330dfeef17aac404341158192cfb45aabb8d9535afd4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 469, + 479, + 492 + ], + "lines": [ + { + "bbox": [ + 131, + 469, + 479, + 492 + ], + "spans": [ + { + "bbox": [ + 131, + 469, + 479, + 492 + ], + "type": "text", + "content": "Fig. 4: Qualitative analysis of the uncertainty estimation. We compare softmax confidence to our MC-derived Shannon entropy." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 507, + 479, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 479, + 531 + ], + "type": "text", + "content": "ric in Fig. 4 to illustrate the better correlation of our MC-derived uncertainty with pseudo-label accuracy compared to softmax based uncertainty." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 532, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 480, + 567 + ], + "type": "text", + "content": "Labeling Percentage " + }, + { + "bbox": [ + 130, + 532, + 480, + 567 + ], + "type": "inline_equation", + "content": "(p_{\\tau})" + }, + { + "bbox": [ + 130, + 532, + 480, + 567 + ], + "type": "text", + "content": ": We compare various set labeling percentages " + }, + { + "bbox": [ + 130, + 532, + 480, + 567 + ], + "type": "inline_equation", + "content": "p_{\\tau}" + }, + { + "bbox": [ + 130, + 532, + 480, + 567 + ], + "type": "text", + "content": " across all tasks in Tab. 6. While we obtain higher accuracy pseudo-labels at " + }, + { + "bbox": [ + 130, + 532, + 480, + 567 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 532, + 480, + 567 + ], + "type": "text", + "content": ", it can be seen that the reduced supervision negatively impacts performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 589, + 314, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 589, + 314, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 589, + 314, + 601 + ], + "type": "text", + "content": "5 Limitations and Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "The main advantage of the proposed Bayesian self-training is its simplicity and ease of implementation compared to existing SOTA approaches that utilize highly engineered and often costly modules. However, the method has a limitation. As observed from the results, the performance is comparably inferior to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 189, + 152, + 418, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 6: Ablation study analyzing the effect of the hyperparameter " + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "inline_equation", + "content": "p_{\\tau}" + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": " for each task. We vary the labeling percentage and show the performance on SemanticKITTI 10%, ScanNet 10%, and ScanRefer 10% (incl. verbal input) for each respective task." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 152, + 418, + 200 + ], + "lines": [ + { + "bbox": [ + 189, + 152, + 418, + 200 + ], + "spans": [ + { + "bbox": [ + 189, + 152, + 418, + 200 + ], + "type": "table", + "html": "
ThresholdmIoUAP@50OA@50
50%61.450.022.8
75%61.751.924.3
100%60.047.016.9
", + "image_path": "31e8a4c7f17156783fc9846cea32572c4d4ff658b64302e45449260d9ebaa6c0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 189, + 243, + 418, + 291 + ], + "blocks": [ + { + "bbox": [ + 132, + 205, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 239 + ], + "type": "text", + "content": "Table 7: Ablation study showing the performance on SemanticKITTI " + }, + { + "bbox": [ + 132, + 205, + 480, + 239 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 132, + 205, + 480, + 239 + ], + "type": "text", + "content": ", ScanNet " + }, + { + "bbox": [ + 132, + 205, + 480, + 239 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 132, + 205, + 480, + 239 + ], + "type": "text", + "content": " and ScanRefer " + }, + { + "bbox": [ + 132, + 205, + 480, + 239 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 132, + 205, + 480, + 239 + ], + "type": "text", + "content": " (incl. verbal input) for varying label aggregation strategies from majority rules to unanimous voting over 9 stochastic forward passes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 243, + 418, + 291 + ], + "lines": [ + { + "bbox": [ + 189, + 243, + 418, + 291 + ], + "spans": [ + { + "bbox": [ + 189, + 243, + 418, + 291 + ], + "type": "table", + "html": "
ThresholdmIoUAP@50OA@50
5/960.747.920.1
7/961.051.023.9
9/961.751.924.3
", + "image_path": "9fd91a96193cd63d6d590fc52ee878daae9ec84f2415a463a9c7fb546c5e2615.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 303, + 480, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 480, + 386 + ], + "type": "text", + "content": "SOTA semi-supervised approaches for highly sparse training splits. This behavior stems from the reliance on unanimous voting (Eq. 6). With low amounts of labeled data, the initial naive model performance suffers, causing a reduction in the number of unanimously voted points. This effect is further pronounced when tackling instance segmentation, where fused instance masks may affect multiple predictions. In Tab. 7, we show that relaxation does not aid the performance as it further introduces incorrect predictions into the pseudo-label set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 387, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 481, + 472 + ], + "type": "text", + "content": "While semi-supervised training methods that enhance the robustness of the pseudo-labels either via a teacher model (DiAL [44, 46]), mix-augmentations (LaserMix [29]) or multi-stage refinement (TWIST [10]) are presented as competing to ours, we argue that as a solution, they can in fact be utilized within our framework in place of the 3D model to improve the quality of the initial predictions. This would allow a larger set of unanimously voted samples and would thus re-enable the benefits of our proposed method on highly sparse splits." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 489, + 220, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 220, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 220, + 502 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 514, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 634 + ], + "type": "text", + "content": "In this work, we introduce a simple yet effective Bayesian self-training approach for semi-supervised 3D semantic segmentation by utilizing uncertainty estimation strategies from the field of active learning. Specifically, through multiple stochastic forward passes, we generate an initial pseudo-label set and then filter it based on estimated predictive entropy. We further extend our semi-supervised 3D method to newly include both instance segmentation and referral-based instance segmentation via an efficient, heuristic " + }, + { + "bbox": [ + 130, + 514, + 482, + 634 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 514, + 482, + 634 + ], + "type": "text", + "content": "-partite matching algorithm. We conduct extensive experiments on multiple datasets, covering both indoor and outdoor domains, showing that our method performs on par if not better than current domain- and task-specific methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 639, + 480, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 639, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 639, + 480, + 663 + ], + "type": "text", + "content": "Acknowledgments: This work is funded by Toyota Motor Europe via the research project TRACE-Zürich." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "O. Unal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "text", + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: ReferIt3D: Neural listeners for fine-grained 3D object identification in real-world scenes. In: ECCV (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "text", + "content": "2. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3D semantic parsing of large-scale indoor spaces. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 253 + ], + "type": "text", + "content": "3. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 480, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 480, + 276 + ], + "type": "text", + "content": "4. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3DJCG: A unified framework for joint dense captioning and visual grounding on 3D point clouds. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 308 + ], + "type": "text", + "content": "5. Cascante-Bonilla, P., Tan, F., Qi, Y., Ordonez, V.: Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 6912-6920 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "type": "text", + "content": "6. Chen, D.Z., Chang, A.X., Nießner, M.: ScanRefer: 3D object localization in RGB-D scans using natural language. In: ECCV (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "text", + "content": "7. Chen, J., Luo, W., Wei, X., Ma, L., Zhang, W.: HAM: Hierarchical attention model with high performance for 3D visual grounding. arXiv preprint arXiv:2210.12513 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "type": "text", + "content": "8. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 430 + ], + "type": "text", + "content": "9. Cheng, M., Hui, L., Xie, J., Yang, J.: SSPC-Net: Semi-supervised semantic 3D point cloud segmentation network. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 432, + 480, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 476 + ], + "type": "text", + "content": "0. Chu, R., Ye, X., Liu, Z., Tan, X., Qi, X., Fu, C.W., Jia, J.: TWIST: Two-way inter-label self-training for semi-supervised 3D instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "type": "text", + "content": "1. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: ICCV (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 498, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 542 + ], + "type": "text", + "content": "2. Dai, W., Li, X., Cheng, K.T.: Semi-supervised deep regression with uncertainty consistency and variational model ensembling via bayesian neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7304-7313 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "text", + "content": "3. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3D point clouds. In: International Conference on Robotics and Automation (ICRA) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "text", + "content": "4. Feng, M., Li, Z., Li, Q., Zhang, L., Zhang, X., Zhu, G., Zhang, H., Wang, Y., Mian, A.: Free-form description guided 3D visual graph network for object grounding in point cloud. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "5. Gal, Y., Ghahramani, Z.: Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In: International Conference on Machine Learning (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "6. Gal, Y., Islam, R., Ghahramani, Z.: Deep Bayesian active learning with image data. In: International Conference on Machine Learning (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "17. Gao, H.a., Tian, B., Li, P., Zhao, H., Zhou, G.: DQS3D: Densely-matched quantization-aware semi-supervised 3D detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "18. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International Conference on Machine Learning (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 481, + 206 + ], + "type": "text", + "content": "19. He, R., Yang, J., Qi, X.: Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "text", + "content": "20. Hou, J., Graham, B., Niessner, M., Xie, S.: Exploring data-efficient 3D scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "text", + "content": "21. Houlsby, N., Huszar, F., Ghahramani, Z., Lengyel, M.: Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745 (2011)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 285 + ], + "type": "text", + "content": "22. Huang, P.H., Lee, H.H., Chen, H.T., Liu, T.L.: Text-guided graph neural networks for referring 3D instance segmentation. In: AAAI (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 308 + ], + "type": "text", + "content": "23. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3D visual grounding. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "24. Jain, A., Gkanatsios, N., Mediratta, I., Fragkiadaki, K.: Bottom up top down detection transformers for language grounding in images and point clouds. In: ECCV (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 385 + ], + "type": "text", + "content": "25. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "text", + "content": "26. Kendall, A., Badrinarayanan, V., Cipolla, R.: Bayesian SegNet: Model uncertainty in deep convolutional encoder-decoder architectures for scene understanding. In: Proceedings of the British Machine Vision Conference (BMVC) (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "text", + "content": "27. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "text", + "content": "28. Kong, C., Lin, D., Bansal, M., Urtasun, R., Fidler, S.: What are you talking about? text-to-image coreference. In: ICCV (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 476, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 481, + 509 + ], + "type": "text", + "content": "29. Kong, L., Ren, J., Pan, L., Liu, Z.: LaserMix for semi-supervised LiDAR semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "type": "text", + "content": "30. Lai, X., Yuan, Y., Chu, R., Chen, Y., Hu, H., Jia, J.: Mask-attention-free transformer for 3D instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 586 + ], + "type": "text", + "content": "31. Lee, M.S., Yang, S.W., Han, S.W.: Gaia: Graphical information gain based attention network for weakly supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 582-591 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "text", + "content": "32. Lei, J., Deng, C., Schmeckpeper, K., Guibas, L., Daniilidis, K.: EFEM: Equivariant neural field expectation maximization for 3D object segmentation without scene supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "33. Liu, K., Zhao, Y., Nie, Q., Gao, Z., Chen, B.M.: Weakly supervised 3D scene segmentation with region-level boundary awareness and instance discrimination. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "O. Unal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3D-SPS: Single-stage 3D visual grounding via referred point progressive selection. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "35. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Van Gool, L.: Towards a weakly supervised framework for 3D point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4454-4468 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 228 + ], + "type": "text", + "content": "36. Prabhudesai, M., Tung, H.Y.F., Javed, S.A., Sieb, M., Harley, A.W., Fragkiadaki, K.: Embodied language grounding with 3D visual feature representations. In: CVPR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "type": "text", + "content": "37. Qin, Z., Wang, J., Lu, Y.: Weakly supervised 3D object detection from point clouds. In: Proceedings of the ACM International Conference on Multimedia (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "text", + "content": "38. Ren, Z., Misra, I., Schwing, A.G., Girdhar, R.: 3D spatial recognition without spatially labeled 3D. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "text", + "content": "39. Roh, J., Desingh, K., Farhadi, A., Fox, D.: LanguageRefer: Spatial-language model for 3D visual grounding. In: CoRL (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 329 + ], + "type": "text", + "content": "40. Shannon, C.E.: A mathematical theory of communication. The Bell system technical journal 27(3) (1948)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "text", + "content": "41. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research 15(1), 1929-1958 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "text", + "content": "42. Tang, Y.S., Lee, G.H.: Transferable semi-supervised 3D object detection from rgb-d data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "text", + "content": "43. Tao, A., Duan, Y., Wei, Y., Lu, J., Zhou, J.: SegGroup: Seg-level supervision for 3D instance and semantic segmentation. IEEE Transactions on Image Processing 31, 4952-4965 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "text", + "content": "44. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 508 + ], + "type": "text", + "content": "45. Unal, O., Dai, D., Hoyer, L., Can, Y.B., Van Gool, L.: 2d feature distillation for weakly- and semi-supervised 3d semantic segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7336-7345 (January 2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 509, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 531 + ], + "type": "text", + "content": "46. Unal, O., Dai, D., Unal, A.T., Van Gool, L.: Discwise active learning for LiDAR semantic segmentation. IEEE Robotics and Automation Letters (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "type": "text", + "content": "47. Unal, O., Dai, D., Van Gool, L.: Scribble-supervised LiDAR semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 565, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 587 + ], + "type": "text", + "content": "48. Unal, O., Sakaridis, C., Saha, S., Gool, L.V.: Four ways to improve verbo-visual fusion for dense 3d visual grounding (2024), https://arxiv.org/abs/2309.04561" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 620 + ], + "type": "text", + "content": "49. Wang, G., Li, W., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T.: Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing 338, 34-45 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "50. Wang, H., Cong, Y., Litany, O., Gao, Y., Guibas, L.J.: 3D IoUMatch: Leveraging IoU prediction for semi-supervised 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bayesian Self-Training for Semi-Supervised 3D Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 621 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "51. Wang, J., Gang, H., Ancha, S., Chen, Y.T., Held, D.: Semi-supervised 3D object detection via temporal graph neural networks. In: International Conference on 3D Vision (3DV) (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "52. Wei, J., Lin, G., Yap, K.H., Hung, T.Y., Xie, L.: Multi-path region mining for weakly supervised 3D semantic segmentation on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "53. Wu, Y., Shi, M., Du, S., Lu, H., Cao, Z., Zhong, W.: 3d instances as 1d kernels. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 235-252. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "54. Xia, Y., Liu, F., Yang, D., Cai, J., Yu, L., Zhu, Z., Xu, D., Yuille, A., Roth, H.: 3D semi-supervised learning with uncertainty-aware multi-view co-training. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (March 2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "55. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10687-10698 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "56. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3D point cloud understanding. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "57. Xu, X., Lee, G.H.: Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "58. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmentation. In: MICCAI (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "59. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: InstanceRefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: ICCV (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "text", + "content": "60. Zhang, F., Fang, J., Wah, B., Torr, P.: Deep fusionnet for point cloud semantic segmentation. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "61. Zhang, Y., Li, Z., Xie, Y., Qu, Y., Li, C., Mei, T.: Weakly supervised semantic segmentation for large-scale point cloud. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "62. Zhao, L., Cai, D., Sheng, L., Xu, D.: 3DVG-Transformer: Relation modeling for visual grounding on point clouds. In: ICCV (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "63. Zhao, N., Chua, T.S., Lee, G.H.: SESS: Self-ensembling semi-supervised 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "text", + "content": "64. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "type": "text", + "content": "65. Zoph, B., Ghiasi, G., Lin, T.Y., Cui, Y., Liu, H., Cubuk, E.D., Le, Q.V.: Rethinking pre-training and self-training. arXiv preprint arXiv:2006.06882 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "66. Zou, Y., Yu, Z., Vijaya Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: The European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "O. Unal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_content_list.json b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2541e1be2913a114a03c0b610768f618e9399696 --- /dev/null +++ b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_content_list.json @@ -0,0 +1,1826 @@ +[ + { + "type": "text", + "text": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 269, + 141, + 733, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Omer Dahary1, Or Patashnik1,2, Kfir Aberman2, and Daniel Cohen-Or1,2", + "bbox": [ + 215, + 210, + 785, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Tel Aviv University", + "bbox": [ + 426, + 238, + 575, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Snap Research", + "bbox": [ + 444, + 252, + 557, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 ginger kittens and 2 gray kittens...", + "bbox": [ + 362, + 290, + 640, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/458d9d6bc89265b09d53cdac1017211b3ad601bd8f74190ec9496699c6819862.jpg", + "image_caption": [ + "\"... on the stairs\"" + ], + "image_footnote": [], + "bbox": [ + 217, + 308, + 354, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9332232f51f07a12917bb8a8cedba40642fa194b4a49e63bd7781902d1baaf00.jpg", + "image_caption": [ + "\"...on the street\"" + ], + "image_footnote": [], + "bbox": [ + 361, + 309, + 498, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f1ca0dbef871026421dd1dabc665c1ee1879bf09633ac24e7c14f0d84867441.jpg", + "image_caption": [ + "\"...in a basket\"", + "Fig. 1: Our method bounds the attention to enable layout control over a pre-trained text-to-image diffusion model. Bounded Attention effectively reduces the impact of the innate semantic leakage during denoising, encouraging each subject to be itself. Our method can faithfully generate challenging layouts featuring multiple similar subjects with different modifiers (e.g., ginger and gray kittens)." + ], + "image_footnote": [], + "bbox": [ + 504, + 309, + 640, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc74ce6b4fc1619b6e9df50dcd7d20364af2f282ec035ac5477471e887318c45.jpg", + "image_caption": [ + "\"... on a tree\"" + ], + "image_footnote": [], + "bbox": [ + 648, + 309, + 784, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-to-image diffusion models have an unprecedented ability to generate diverse and high-quality images. However, they often struggle to faithfully capture the intended semantics of complex input prompts that include multiple subjects. Recently, numerous layout-to-image extensions have been introduced to improve user control, aiming to localize subjects represented by specific tokens. Yet, these methods often produce semantically inaccurate images, especially when dealing with multiple semantically or visually similar subjects. In this work, we study and analyze the causes of these limitations. Our exploration reveals that the primary issue stems from inadvertent semantic leakage between subjects in the denoising process. This leakage is attributed to the diffusion model's attention layers, which tend to blend the visual features of different subjects. To address these issues, we introduce Bounded Attention, a training-free method for bounding the information flow in the sampling process. Bounded Attention prevents detrimental leakage among subjects and enables guiding the generation to promote each subject's individuality, even with complex multi-subject conditioning. Through extensive experimentation, we demonstrate that our method empowers the generation of multiple subjects that better align with given prompts and layouts.", + "bbox": [ + 259, + 535, + 740, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image Generation $\\cdot$ Diffusion Models $\\cdot$ Semantic Alignment", + "bbox": [ + 261, + 825, + 740, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, text-to-image generation has undergone a significant shift with the integration of conditional diffusion models $[?,?,?,?]$ , allowing for the facile generation of high-quality and diverse images. The use of attention layers in architectures of such generative models has been identified as a major factor contributing to improved quality of generated images $[?,?]$ . However, these models struggle to accurately generate scenes containing multiple subjects, especially when they are semantically or visually similar.", + "bbox": [ + 212, + 175, + 785, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we study the problem of multi-subject image generation in attention-based diffusion models. Our contribution is twofold. First, we recognize the underlying reasons for the difficulty in generating images containing multiple subjects, especially those sharing semantic similarities. Second, building on our insights, we present a method aimed at mitigating semantic leakage in the generated images, allowing control over the generation of multiple subjects (see Figure 1).", + "bbox": [ + 212, + 282, + 785, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate an innate bias within the common attention-based architectures utilized in diffusion models, which predisposes them to leak visual features between subjects. In particular, the functionality of attention layers is designed to blend features across the image. Therefore, they inherently lead to information leakage between subjects. This phenomenon is particularly noticeable when subjects are semantically similar and, therefore, attend to each other (Figure 2).", + "bbox": [ + 212, + 387, + 785, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A plethora of works tries to mitigate the cross-subject leakage issue, either by modifying the sampling process to better follow different subjects in the prompt $[\\?,?]$ , or by coupling the global prompt with layout information via segmentation maps or bounding boxes labeled with subject classes or local prompts $[\\?,?,?,?]$ . However, the majority of these methods still encounter difficulties in accurately aligning to input layouts, particularly in scenarios involving two or more semantically similar subjects.", + "bbox": [ + 212, + 478, + 785, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our approach, we guide the image generation with a spatial layout $[?,?]$ . To address cross-subject leakage, we introduce the Bounded Attention mechanism, utilized during the denoising process to generate an image. This method bounds the influence of irrelevant visual and textual tokens on each pixel, which otherwise promotes leakage. By applying this mechanism, we encourage each subject to be itself, in the sense that it hinders the borrowing of features from other subjects in the scene. We show that bounding the attention is needed both in the cross- and self-attention layers. Moreover, we find additional architectural components that amplify leakage, modify their operation, and present remedies to them.", + "bbox": [ + 212, + 583, + 785, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that our method succeeds in facilitating control over multiple subjects within complex, coarse-grained layouts comprising numerous bounding boxes with similar semantics. Particularly challenging examples are demonstrated in Figure 1, where we successfully generate five kittens with a mix of adjectives. We conduct experiments on both Stable Diffusion [?] and SDXL [?] architectures and demonstrate the advantage of our method compared to previous ones, both supervised and unsupervised.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "O. Dahary et al.", + "bbox": [ + 271, + 114, + 382, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/14e4ea5dbb99eed94036ac648061a206307e46c5eadf11efc71ba276ca218eb5.jpg", + "image_caption": [ + "Fig.2: Misalignment in layout-to-image generation include (i) catastrophic neglect [?] where the model fails to include one or more subjects mentioned in the prompt within the generated image, (ii) incorrect attribute binding [?,?] where attributes are not correctly matched to their corresponding subjects, and (iii) subject fusion [?] where the model merges multiple subjects into a single, larger subject." + ], + "image_footnote": [], + "bbox": [ + 527, + 160, + 687, + 279 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ab0f05d519ab46ce516e55615ffdd6925ac0cb61fa077cfe3cad7a62060939a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 159, + 774, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 321, + 382, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-image diffusion models. Diffusion models, trained on huge datasets [?], have demonstrated their power in learning the complex distributions of diverse natural images [?,?,?,?,?] Augmenting attention layers into diffusion models and conditioning them on textual prompts by coupling them with visually-aligned text encoders [?] leads to powerful text-to-image models [?] In this work, we specifically examine two such open-source text-to-image diffusion models: Stable Diffusion [?], and the more recent SDXL [?]", + "bbox": [ + 212, + 354, + 787, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic alignment in text-to-image synthesis. A critical drawback of current text-to-image models pertains to their limited ability to faithfully represent the precise semantics of input prompts. Various studies have identified common semantic failures and proposed mitigating strategies, such as adjusting text embeddings $[?,?,?]$ , or optimizing noisy signals to strengthen or align cross-attention maps $[?,?]$ . Nevertheless, these methods often fall short in generating multiple subjects, and do not adhere to positional semantics, such as subjects' number or location.", + "bbox": [ + 212, + 468, + 787, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layout-guided image synthesis. Addressing the semantic alignment concern, alternative approaches advocate for conditioning the diffusion process on layout information, either by training new models from scratch $[?,?]$ or fine-tuning an existing one $[?,?,?,?]$ . Despite their promise, these methods demand extensive computational resources and prolonged training times. Moreover, they are constrained by the layout distribution of the training data and the models' architectural bias to blend subject features, a limitation that our Bounded Attention aims to overcome.", + "bbox": [ + 212, + 598, + 787, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To circumvent these challenges, numerous researchers explore training-free techniques, where the generation process itself is modified to enforce layout constraints. Several optimization-based works employ techniques similar to classifier-free guidance to localize the cross-attention $[\\?,?,?,?]$ and/or self-attention maps $[\\?]$ . While effective in aligning random noise with the intended layout, guiding attention maps to distinct regions may inadvertently lead to undesired behavior, particularly when different subjects share similar semantics and visual features. Furthermore, these methods often exhibit a deteriorating effect on", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "visual quality, thereby limiting their applicability to only the initial denoising steps and neglecting finer control over shape and visual details, which are determined only in later stages [?]. Bounded Attention addresses these shortcomings by regulating attention computation throughout the entire denoising process.", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another approach involves generating each subject separately in its own denoising process $[\\?,?]$ . While these methods inherently address catastrophic neglect, they tend to generate disharmonious images, and remain susceptible to leakage when merging subjects in subsequent stages. In contrast, masking attention maps to input bounding boxes $[\\?]$ or attenuating attention in specific segments $[\\?]$ represents a milder variant of this strategy, aimed at avoiding visible stitching. However, these methods often fall short of fully mitigating subject leakage, compromising semantic alignment. In comparison, Bounded Attention is able to carefully govern information propagation among subjects in a single denoising process.", + "bbox": [ + 212, + 208, + 784, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While both the trained models and training-free techniques aim to generate numerous objects, they do not mitigate the inherent leakage caused by attention mechanisms. Unlike our Bounded Attention technique, these methods encounter challenges in effectively generating and controlling a multitude of subjects, especially when they share semantic similarity. Notably, existing techniques struggle to accurately generate even two semantically similar subjects, whereas our method, as demonstrated succeeds in generating five and even more subjects.", + "bbox": [ + 212, + 359, + 785, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 488, + 382, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent diffusion models. In this work, we examine Stable Diffusion [?] and SDXL [?], which are both publicly available latent diffusion models. These models operate in the latent space of a pretrained image autoencoder, and are thus tasked with denoising a latent representation of the image, where each latent pixel corresponds to a patch in the generated image. Starting from pure random noise $\\mathbf{z}_T$ , at each timestep $t$ , the current noisy latent $\\mathbf{z}_t$ is passed to a denoising UNet $\\epsilon_{\\theta}$ , trained to predict the current noise estimate $\\epsilon_{\\theta}(\\mathbf{z}_t,y,t)$ using the guiding prompt encoding $y$ .", + "bbox": [ + 212, + 517, + 787, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attention layers. At each block, the UNet utilizes residual convolution layers, producing intermediate features $\\phi^{(l)}(\\mathbf{z}_t)$ , where $l$ denotes the layer's index. These, in turn, are passed to attention layers, which essentially average different values $\\mathbf{V}_{\\mathbf{t}}^{(1)}$ according to pixel-specific weights:", + "bbox": [ + 214, + 648, + 787, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi^ {(l + 1)} (\\mathbf {z} _ {t}) = \\mathbf {A} _ {t} ^ {(l)} \\mathbf {V} _ {t} ^ {(l)}, \\text {w h e r e} \\mathbf {A} _ {t} ^ {(l)} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} _ {t} ^ {(l)} \\mathbf {K} _ {t} ^ {(l) ^ {\\intercal}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 724, + 785, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the keys $\\mathbf{K}_t^{(l)} = f_K^{(l)}\\left(\\mathbf{C}_t^{(l)}\\right)$ and values $\\mathbf{V}_t^{(l)} = f_V^{(l)}\\left(\\mathbf{C}_{\\mathbf{t}}^{(\\mathbf{1})}\\right)$ are linear projections of context vectors $\\mathbf{C}_t^{(l)}$ . In the cross-attention layers we inject semantic context from the prompt encoding $\\mathbf{C}_t^{(l)} \\equiv y$ , while self-attention layers utilize global information from the latent itself $\\mathbf{C}_t^{(l)} \\equiv \\phi^{(l)}(\\mathbf{z}_t)$ .", + "bbox": [ + 212, + 762, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "O. Dahary et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The weighting scheme is determined by the attention map $\\mathbf{A}_t^{(l)}$ which represents the probability of each pixel to semantically associate with a key-value pair. This is done by linearly projecting the latent noisy pixels to queries $\\mathbf{Q}_t^{(l)} = f_Q^{(l)}\\left(\\phi^{(l)}(\\mathbf{z}_t)\\right)$ and computing their inner product with the keys. It has been widely demonstrated that the cross-attention maps are highly indicative of the semantic association between the image layout and the prompt tokens [?]. Meanwhile, the self-attention maps govern the correspondence between pixels, and thus form the image's structure [?].", + "bbox": [ + 212, + 145, + 787, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Semantic Leakage", + "text_level": 1, + "bbox": [ + 214, + 284, + 426, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by studying the causes of semantic leakage in Stable Diffusion [?], and examine the limitations of existing layout-to-image approaches.", + "bbox": [ + 212, + 306, + 785, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 On Subject Similarity", + "text_level": 1, + "bbox": [ + 214, + 347, + 444, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 illustrates various misalignment failures observed in state-of-the-art layout-to-image training-free methods. As we shall show, these failures are more prevalent for subjects that share semantic or visual similarity.", + "bbox": [ + 212, + 361, + 784, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $x_{s_1}, x_{s_2} \\in \\mathbb{R}^2$ be 2D latent coordinates corresponding to two semantically similar subjects $s_1, s_2$ in the generated image. Intuitively, we expect that along the denoising process, the queries corresponding to these pixels, $\\mathbf{Q}_t^{(l)}[x_{s_1}], \\mathbf{Q}_t^{(l)}[x_{s_2}]$ , will be similar and hence also their attention responses. This, in turn, also implies that they will share semantic information from the token embeddings through the cross-attention layers or visual information via the self-attention layers.", + "bbox": [ + 212, + 406, + 785, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To explore this hypothesis, we investigate the model's behavior when tasked with generating two subjects and analyze their attention features in both cross- and self-attention layers. Subsequently, we meticulously examine these features and demonstrate how their behavior sheds light on the leakage observed in generated images.", + "bbox": [ + 212, + 513, + 785, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 Cross-Attention Leakage", + "text_level": 1, + "bbox": [ + 214, + 599, + 465, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To analyze the leakage caused by cross-attention layers, we examine the cross-attention queries. We depict these queries in the plots in Figure 3, where each point corresponds to a single query projected to 2D with PCA. To label each point with its corresponding subject, we compute the subject masks by averaging cross-attention maps [?] and color each projected query according to the subject's text color. The leftmost plot, in which the two subjects were generated separately, serves as a reference point to the relation between the queries when there is no leakage. For comparative analysis, we also present results for Layout-guidance [?], as a simple representative of current training-free layout-guided approaches, and Bounded Attention, which we shall cover in the sequel.", + "bbox": [ + 212, + 614, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider the cross-attention queries in two examples: \"a kitten\" and \"a puppy\", and \"a hamster\" and a \"squirrel\". As can be seen in the reference plots, the kitten and puppy queries share some small overlap, and the hamster and squirrel queries are mixed together. The level of separation between the red and blue dots in the plots reveals the semantic similarity of the two forming subjects.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3800d6710180dd23f74519b9113dfeda3ab809d5a409e47f3a5864df3163d88d.jpg", + "image_caption": [ + "Fig. 3: Cross-Attention Leakage. We demonstrate the emergence of semantic leakage at the cross-attention layers. We show two examples: a puppy a kitten, and a hamster and a squirrel. In the two leftmost columns, the subjects were generated separately using Stable Diffusion (SD). In the right three columns, we generate a single image with the two subjects using three different methods: Stable Diffusion (SD), Layout Guidance (LG), and Bounded Attention (BA, ours). Under each row, we plot the two first principal components of the cross-attention queries. As can be seen, the separation of the queries (blue and red) reflects the leakage between the subjects in the generated images." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 485, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ed8d39cfdcbe7b23060e6178933c2a07a3efbcb6f493db72ce21b5c183ce560.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 143, + 781, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clearly, vanilla Stable Diffusion struggles to adequately generate the two subjects within the same image. This is apparent by the visual leakage between the two subjects, as the model cannot avoid averaging their distinct visual properties. For example, the puppy has the visual features of a kitten, like raised ears and a triangular nose, while the squirrel loses its distinct ear shape and its feet are now pink like the hamster. Respectively, it can be seen that the queries of the two subjects are mixed, even for the kitten and the puppy which are more separated in the reference plot.", + "bbox": [ + 212, + 426, + 787, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meanwhile, Layout Guidance (LG), which optimizes $\\mathbf{z}_t$ to have attention peaks for each noun token at their corresponding region, exhibits interesting results. Its optimization objective implicitly encourages the separation between subjects' cross-attention queries. This can have positive effects, like the hamster and squirrel having unique colors, but at the same time yields unwanted artifacts, like the puppy losing its face altogether. Moreover, it can inadvertently push the latent signal out of distribution, causing quality degradation, as evident by the hamster's and squirrel's cartoonish texture. When it overly pushes the latent out of distribution, it leads to the catastrophic neglect phenomenon (Figure 2).", + "bbox": [ + 212, + 547, + 787, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In comparison, when examining the plots of our method alongside the reference plots, our approach preserves the feature distribution of the subjects' queries, and successfully generates the two subjects, even when the queries are as mixed as in the hamster and the squirrel.", + "bbox": [ + 212, + 686, + 787, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above analysis yields two immediate conclusions: (i) Semantic similarity between subjects is reflected by their queries proximity, and leads to mixed queries when the subjects are generated together. This in turn leads to leakage between the subjects in the generated images, and (ii) enforcing semantic separation by modifying the semantic meaning of the cross-attention queries is harmful. The former observation represents a crucial architectural limitation in", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "O. Dahary et al.", + "bbox": [ + 271, + 114, + 382, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e6b310983195244e63448ea3138f357dd58613f5ee290d0cc5ad227fe6172c1d.jpg", + "image_caption": [ + "SD", + "Fig. 4: Self-Attention Leakage. We demonstrate the emergence of semantic leakage at the self-attention maps of two generally semantically dissimilar subjects: a crab and a frog. The images are generated by Stable Diffusion (SD) and Layout-guidance (LG). We highlight specific pixels, such as those of a subject's eye or leg, and present their respective self-attention maps." + ], + "image_footnote": [], + "bbox": [ + 217, + 154, + 772, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "current diffusion models, and the latter pinpoints to a previously unexplored weakness in the widely used latent optimization methodology. Bounded Attention is designed to overcome these limitations.", + "bbox": [ + 212, + 321, + 784, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Self-Attention Leakage", + "text_level": 1, + "bbox": [ + 214, + 388, + 450, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now turn to analyze the leakage caused by self-attention layers. It has been shown that self-attention features exhibit dense correspondences within the same subject $[\\ref{e}]$ and across semantic similar ones $[\\ref{e}]$ . Hence, they are suspected as another source of leakage, that we shall study next.", + "bbox": [ + 212, + 402, + 784, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we choose to examine the self-attention maps as a means to understand the leakage. In Figure 4 we focus on representative pixels (marked in yellow) associated with the subjects' eyes and legs, where visual leakage is most pronounced. As expected, features from one subject's eye or leg attend to the semantic similar body parts of the other. As a result, the features of each of the yellow points are directly affected by the features of the counterpart subject, causing leakage. In both images, the crab and the frog have similar appearances. In vanilla SD, both have a crab-like color and limbs with frog's toe pads. In LG, both have frog-like eyes and crab legs.", + "bbox": [ + 212, + 464, + 785, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, this tendency to rely on similar patches aids the model in denoising the latent signal and is essential for achieving coherent images with properly blended subjects and backgrounds. Nevertheless, it has the drawback of leaking features between disjointed subjects. Therefore, completely disjointing the visual features during denoising $[?]$ , or naively pushing self-attention queries apart through optimization $[?]$ , can lead to subpar results. Consequently, we introduce the Bounded Attention mechanism to mitigate leakage and guide the latent signal towards subject separability, while avoiding detrimental artifacts.", + "bbox": [ + 212, + 599, + 785, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It's important to highlight that the leakage resulting from both cross- and self-attention layers is intertwined and mutually reinforced. Therefore, addressing the leakage caused by only one of these layers is insufficient to prevent the leakage in the generated images.", + "bbox": [ + 212, + 720, + 785, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 Levels of Similarity", + "text_level": 1, + "bbox": [ + 214, + 795, + 421, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the two previous sections, our focus was primarily on subjects that share semantic similarity. Building upon the observation that the UNet's inner layers", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation", + "bbox": [ + 218, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A kitten A puppy $64 \\times 64$ $32 \\times 32$ $16 \\times 16$", + "bbox": [ + 217, + 143, + 478, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7643c0359d0c3a9cb8ccfc6c73c19521535ea823ea1b8000c49bf05bb12663b1.jpg", + "image_caption": [ + "Fig. 5: We generate different subjects, and plot the first two principal components of the cross-attention queries at different layers of the UNet, where each layer is of different resolution. The high semantic similarity between the kitten and the puppy is expressed by the proximity of their queries through all layers. Meanwhile, the lizard and fruit share similar texture, and hence only their high-resolution queries are entangled." + ], + "image_footnote": [], + "bbox": [ + 217, + 155, + 478, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A lizard A fruit $64 \\times 64$ $32 \\times 32$ $16 \\times 16$", + "bbox": [ + 503, + 143, + 759, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a656c60d3d9a5eb3ea2cebf3978c9a9ba84d4486d3b92b505aadd150f86ddfac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 155, + 763, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dictate the subject's semantics and shape, while the outer layers control its style and appearance [?], we analyzed the inner UNet layers. However, leakage can also occur prominently when generating subjects that share visual similarity rather than semantic. We now turn to explore this scenario and demonstrate that, in such cases, the leakage originates from the UNet's outer layers.", + "bbox": [ + 212, + 313, + 782, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 5, we visualize cross-attention queries at different decoder's layers, when generating the kitten, puppy, lizard, and fruit in isolation. As can be seen, the queries of the kitten and the puppy are mixed across all UNet's layers, aligning with the visual and semantic similarity between these animals. On the other hand, the queries of the lizard and the fruit are overlapped only in the highest-resolution layer, aligning with the lack of semantic similarity between them. A closer look reveals that the lizard and the fruit share surprisingly similar textures, which explains the overlap of the queries in the highest-resolution layer. As explained in the previous sections, this overlap causes leakage between the two subjects, in this case, a visual rather than a semantic leakage (see Figure 2).", + "bbox": [ + 212, + 390, + 787, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Bounded Attention", + "text_level": 1, + "bbox": [ + 215, + 566, + 439, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method takes as input $n$ distinct textual subjects $S = \\{s_i\\}_{i=1}^n$ contained within a global prompt $y$ , along with their corresponding bounding boxes $B = \\{b_i\\}_{i=1}^n$ . Our objective is to condition the generation process on $y$ , $S$ , and $B$ , while preserving the intended semantics of each subject, all without requiring any training or fine-tuning.", + "bbox": [ + 212, + 595, + 787, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 illustrates an overview of our method. There, the input prompt $y$ is \"A kitten and a puppy\", $S = \\{kitten\\}, \\{puppy\\}$ , and the two corresponding bounding boxes $\\{b_1, b_2\\}$ are illustrated in the top left corner. Bounded Attention operates in two modes: Bounded Guidance and Bounded Denoising. Specifically, at the beginning of the denoising process, for $t \\in [T, T_{\\text{guidance}}]$ , we perform a Bounded Guidance step followed by a Bounded Denoising step. In the guidance step, we utilize the Bounded Guidance loss. This interval of timesteps constitutes the optimization phase. Then, for $t \\in [T_{\\text{guidance}}, 0]$ , we apply only Bounded Denoising steps. In both modes, we manipulate the model's forward pass by adopting an augmented weighting scheme in the attention layers, that safeguard the flow of information between the queries and the keys:", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "O. Dahary et al.", + "bbox": [ + 271, + 114, + 382, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37551b515abbb0158346d77723a9ad88db6cb9ae1223715f381023d2a7bad39d.jpg", + "image_caption": [ + "Fig. 6: Bounded Attention operates in two modes: guidance and denoising, each imposing strict constraints to bound the attention of each subject solely to itself and possibly to the background, preventing any influence from other subjects' features. In guidance mode, a loss is minimized to encourage each subject's attention to concentrate within its corresponding bounding box, with other bounding boxes masked to prevent artifacts. During the denoising step, we confine the attention of each subject solely to its bounding box, along with the background in the self-attention. This strategy effectively prevents feature leakage while maintaining the natural immersion of the subject within the image. Attention maps of a specific key, marked with $\\star$ , are shown for each mode: the cross-attention map displays the key corresponding to the \"kitten\" token, and the self-attention map shows a key within the kitten's target bounding box." + ], + "image_footnote": [], + "bbox": [ + 271, + 147, + 733, + 356 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {t} ^ {(l)} = \\mathrm {s o f t m a x} \\left(\\mathbf {Q} _ {t} ^ {(l)} \\mathbf {K} _ {t} ^ {(l) ^ {\\intercal}} + \\mathbf {M} _ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 563, + 785, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $l$ represents the layer index, $t$ represents the diffusion timestep, and $\\mathbf{Q}_t^{(l)},\\mathbf{K}_t^{(l)}$ are the queries and keys of the $l$ -th attention layer. $\\mathbf{M}_t$ represents time-specific masks composed of zeros and $-\\infty$ elements. We refer to $\\mathbf{A}_t^{(l)}$ as the Bounded Attention map. When indexing $\\mathbf{A}_t^{(l)}$ , we use pixel coordinates $\\mathbf{x}$ for rows, and attention-type-specific context vectors $\\mathbf{c}$ for columns. In locations $[\\mathbf{x},\\mathbf{c}]$ , where $\\mathbf{M}_t[\\mathbf{x},\\mathbf{c}] = -\\infty$ , it holds that $\\mathbf{A}_t^{(l)}[\\mathbf{x},\\mathbf{c}] = 0$ . Therefore, these masks prevent harmful information flow between pixels in self-attention layers, and between pixels and token embeddings in cross-attention layers.", + "bbox": [ + 212, + 595, + 787, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Bounded Guidance", + "text_level": 1, + "bbox": [ + 214, + 732, + 419, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Bounded Guidance, we backpropagate through the diffusion model to steer the latent signal toward the desired layout, using Gradient Descent. Our Bounded Guidance loss encourages the Bounded Attention map of each key corresponding to subject $s_i$ , to be within the $b_i$ bounding box. To this end, for each subject key we consider the ratio between the attention within the corresponding bounding box, to the entire Bounded Attention map (see Figure 6). Formally, we aggregate", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation", + "bbox": [ + 220, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the following loss on the different subjects:", + "bbox": [ + 215, + 146, + 524, + 162 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} = 1 - \\frac {\\sum_ {\\mathbf {x} \\in b _ {i} , \\mathbf {c} \\in C _ {i}} \\hat {\\mathbf {A}} [ \\mathbf {x} , \\mathbf {c} ]}{\\sum_ {\\mathbf {x} \\in b _ {i} , \\mathbf {c} \\in C _ {i}} \\hat {\\mathbf {A}} [ \\mathbf {x} , \\mathbf {c} ] + \\alpha \\sum_ {\\mathbf {x} \\notin b _ {i} , \\mathbf {c} \\in C _ {i}} \\hat {\\mathbf {A}} [ \\mathbf {x} , \\mathbf {c} ]}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 169, + 785, + 209 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $i$ denotes the index of subject $s_i$ , $\\hat{\\mathbf{A}}$ is the mean Bounded Attention map, averaged across heads and layers, and $\\alpha$ is a hyperparameter that magnifies the significance of disregarding attention towards the background, as we explain later. Similarly to the above, When indexing $\\hat{\\mathbf{A}}$ , pixel coordinates $\\mathbf{x}$ represent rows, and attention-type-specific context vectors $\\mathbf{c}$ represent columns. We designate $C_i$ as the set of all $s_i$ -related context vectors, i.e., pixel coordinates in $b_i$ for self-attention layers, and the token identifiers of $s_i$ in cross-attention layers. Additionally, for cross-attention layers, we include the first padding token [EoT] in $C_i$ to enhance layout alignment [?].", + "bbox": [ + 212, + 212, + 782, + 347 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For each subject $s_i$ , the mask $\\mathbf{M}_t$ should block the influence of opposing keys (tokens in $s_j$ and pixels in $b_j$ for $j \\neq i$ ), to avoid the artifacts illustrated in Figure 3. This fosters that the queries of different subjects, including similar ones, are not erroneously forced to be far apart. Utilizing this loss, our Bounded Guidance step is defined as $z_t^{\\mathrm{opt}} = z_t - \\beta \\nabla_{z_t} \\sum_i \\mathcal{L}_i^2$ .", + "bbox": [ + 212, + 348, + 784, + 424 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Integrating this loss within the cross-attention layers encourages the localization of each subject's semantics within its bounding boxes [?]. However, as cross-attention responses tend to peak around more typical patches associated with the subject's semantics (e.g., the face of a human, the legs of a crab, etc.), it may lack control over the subject's boundaries. By applying the loss within the self-attention layers, we encourage each subject to establish its own boundaries close to its bounding box, thereby discouraging subject fusion (see Figure 2).", + "bbox": [ + 212, + 424, + 784, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the computation of the loss, we also introduce a hyperparameter $\\alpha$ to reinforce attention to the background. This adjustment aids in preventing subject amalgamation, where a redundant subject is realized from different subject semantics in the background. To preserve image quality, we limit the application of this mode to an initial time interval $[T, T_{\\text{guidance}}]$ , following similar works [?,?].", + "bbox": [ + 212, + 529, + 784, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Bounded Denoising", + "text_level": 1, + "bbox": [ + 215, + 614, + 423, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In Bounded Denoising, we compute the model's output and use it as the next latent in the series. Here, the masks aim to reduce semantic leakage between subjects, as detailed in Section 4, and to prevent unintended semantics from leaking to the background. Unlike Bounded Guidance and typical attention-based guidance approaches, Bounded Denoising can be applied throughout all time steps to mitigate leaks in fine details, which emerge only in later stages [?].", + "bbox": [ + 212, + 628, + 784, + 719 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "However, coarse masking in later stages may degrade image quality and result in noticeable stitching. To address this, after the optimization phase, for $t \\in [T_{\\text{guidance}}, 0]$ , we replace each bounding box with a fine segmentation mask obtained by clustering the self-attention maps [?]. Since the subject outlines are roughly determined in the initial time steps and evolve gradually thereafter [?], we refine the masks periodically.", + "bbox": [ + 212, + 719, + 784, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Notably, this mechanism also addresses imperfect alignments between subjects and bounding boxes after the guidance phase, which are more common", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "O. Dahary et al.", + "bbox": [ + 271, + 114, + 382, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7f7a4c772b48c2000383a0f93bdcc2553bd7ce60eac9bb384abbb9c2b92ca02e.jpg", + "image_caption": [ + "Fig. 7: Our results with SDXL. See the supplemental for more examples." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 746, + 320 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "when generating numerous subjects. Thus, employing this technique enhances the robustness of our method to seed selection, ensuring proper semantics even when subjects extend beyond their initial confines (see Figure 1,7). Compared to methods that necessitate strict input masks yet remain susceptible to leakage $[?,?,?]$ , our method offers greater user control with simpler inputs and more satisfactory outcomes.", + "bbox": [ + 212, + 380, + 787, + 470 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Method Details. Further details on the adaptation of Bounded Attention to the cross- and self-attention layers, along with a description of the subject mask refinement process, are provided in the supplementary materials.", + "bbox": [ + 214, + 481, + 787, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 214, + 547, + 375, + 565 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we conduct both qualitative and quantitative experiments to assess the effectiveness of our Bounded Attention method. We compare our approach with three training-free baseline methods: Layout-guidance (LG) [?], BoxDiff (BD) [?], and MultiDiffusion (MD) [?]. Additionally, we include comparisons with GLIGEN [?] and ReCo [?], which necessitate training. Since Attention-refocusing (AR) [?] is based on GLIGEN, we categorize it as a trained method for the purpose of our evaluation. For fairness, when comparing our method with other methods, we use Stable Diffusion.", + "bbox": [ + 212, + 578, + 787, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.1 Qualitative Results", + "text_level": 1, + "bbox": [ + 214, + 720, + 421, + 736 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SDXL results. We begin our experiments by demonstrating the efficacy of our method in challenging scenarios, particularly when tasked with generating multiple semantically similar subjects using SDXL. As can be seen in Figure 7, Vanilla SDXL fails to follow the prompt due to semantic leakage. It generates inaccurate number of dogs and mixes the features of each breed. In comparison, our approach adeptly generates each dog with its unique characteristics.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "\"A gray kitten and a ginger kitten and black puppy in a yard.\"", + "bbox": [ + 316, + 142, + 681, + 154 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7632eac4e548eea099c73f413c5ea32ce25f2d217f7d841b46de947426cd0c99.jpg", + "image_caption": [ + "Fig. 8: Comparison of the first six images generated from the seed 0." + ], + "image_footnote": [], + "bbox": [ + 230, + 155, + 769, + 320 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Non-curated results. Next, we conduct a non-curated comparison with the training-free baseline methods and present the results in Figure 8. We showcase the initial six images sampled from seed 0 for each method. We anticipate that in the absence of Bounded Attention, semantic leakage may freely blend subject features, hindering the intended layout's formation.", + "bbox": [ + 215, + 378, + 785, + 452 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "It is evident from the results that none of the competing methods is able to consistently construct the input layout. Layout Guidance [?] frequently neglects one of the subjects, and even when it generates three subjects, it struggles to avoid leakage, resulting in puppies with kitten-like features or incorrect color assignments. BoxDiff [?] often generates the correct number of subjects but suffers from artifacts in the form of blobs. Similar to Layout Guidance, it encounters difficulties in properly constructing the puppy. Surprisingly, even MultiDiffusion [?], which generates the subjects separately, faces challenges in generating them all, with some disappearing or merging together in its bootstrapping phase.", + "bbox": [ + 212, + 454, + 787, + 590 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In contrast, our method consistently outperforms these approaches, producing three subjects that align with the both prompt and layout in all six images.", + "bbox": [ + 215, + 592, + 785, + 622 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparisons with baselines. We present a qualitative comparison in Figure 9. All competing methods, including those trained specifically for the layout-to-image task, exhibit significant visual and semantic leakage. The training-free methods perform the worst: MultiDiffusion produces disharmonious, low-quality images, while optimization-based methods often result in object fusion, combining different semantics without adhering to the layout.", + "bbox": [ + 215, + 641, + 785, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The training-based approaches closely follow the layout but fail to convey the correct semantics. In the first row, they neglect the corduroy jacket, leaking the denim texture into the other jacket, or even fusing them together. In the second row, the elephant's features leak into the rhino.", + "bbox": [ + 215, + 733, + 785, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In comparison, our method generates images that align with the input layout and prompt, ensuring each subject retains its unique attributes, semantics, and appearance.", + "bbox": [ + 215, + 795, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "O. Dahary et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "\"A denim jacket and a corduroy jacket and a leather handbag in a closet.\"", + "bbox": [ + 256, + 143, + 743, + 157 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1bad5bcf3921fd21d36b4c9872dd3e87124038ceb0f0426b0dd2ea796b2a46c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 157, + 784, + 222 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "\"A big red elephant and a far away rhino in a jungle.\"", + "bbox": [ + 320, + 223, + 679, + 237 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9a947536fbd0f4c66b6701a2b8d231a27646ef426f02a6f3c4daa2af170fc0a4.jpg", + "image_caption": [ + "Fig. 9: Qualitative comparison of our method with baseline methods: Above each row, we display the input prompt, where each subject's color matches the color of its corresponding bounding box. We compare with both training-free methods (2nd-4th columns) and trained models (5th-7th columns). See supplemental for more results." + ], + "image_footnote": [], + "bbox": [ + 217, + 237, + 784, + 320 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 453, + 431, + 467 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We evaluate our method's effectiveness using the Draw-Bench dataset [?], known for its challenging prompts designed to test a model's ability to compose multiple subjects with specific quantities and relations. We use the evaluation procedure from previous work [?, ?]. Our re", + "bbox": [ + 212, + 467, + 431, + 618 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e6df5d0aa843fb7a14335a4e0fed3ef88fa7f2245e083d64ca67dcfeb2bc8fd3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCountingSpatial
PrecisionRecallF1Accuracy
Stable Diffusion [?]0.740.780.730.19
Layout-guidance [?]0.720.780.720.35
BoxDiff [?]0.810.780.760.28
MultiDiffusion [?]0.700.550.570.15
Ours0.810.910.820.43
", + "bbox": [ + 464, + 472, + 766, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 1: Evaluation on the DrawBench dataset.", + "bbox": [ + 450, + 575, + 774, + 590 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "sults, alongside those of other training-free methods, are summarized in Table 1. Unlike other approaches that do not account for semantic leakage, our method demonstrates notable improvements in both the counting and the spatial categories. Notably, other methods struggle to surpass the counting recall rates of vanilla SD, highlighting the effectiveness of Bounded Attention in generating multiple subjects.", + "bbox": [ + 212, + 618, + 787, + 708 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We further conduct a quantitative comparison of semantic leakage in training-free methods through a user study. Further details are in the supplemental.", + "bbox": [ + 212, + 708, + 785, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 750, + 401, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To assess the significance of each component, we conduct an ablation study where we systematically vary our method's configurations by omitting one component in each setting. We show the results in Figure 10.", + "bbox": [ + 212, + 763, + 785, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Guidance is crucial for aligning the latent signal with the intended layout. Neglecting it can lead to subject generation in the background, as demonstrated", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "\"A red lizard and a turtle on the grass.\"", + "bbox": [ + 383, + 143, + 614, + 154 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/591aeb0f9a2447109048e47333aeb97c7bbed21b31e4210a15ef7b9cd1d89c15.jpg", + "image_caption": [ + "w/o G" + ], + "image_footnote": [], + "bbox": [ + 223, + 154, + 302, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/19a6b8330b6947168d23ccc642327b532875c79758a434328cf599019c195e50.jpg", + "image_caption": [ + "w/o BG" + ], + "image_footnote": [], + "bbox": [ + 303, + 155, + 380, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3f73670f251fa60fce3425654f92c11a874c4c1233551299f1d7d337be678dea.jpg", + "image_caption": [ + "w/o BD", + "Fig. 10: Qualitative ablation. We ablate our method by skipping the guidance step (G), performing guidance without Bounded Guidance (BG), not applying Bounded Denoising (BD), and not performing Mask Refinement (MR). We show zoomed-in images of the two rightmost configurations. More examples can be found in the supplemental." + ], + "image_footnote": [], + "bbox": [ + 382, + 155, + 460, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/25bd0bd466b74bc1f55d42a895f017273a814097b8e9e2bb87f36241a7caf1bc.jpg", + "image_caption": [ + "w/o MR" + ], + "image_footnote": [], + "bbox": [ + 460, + 155, + 537, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/446db0dd9c92ae76d87114eb4082cde8f14cdddaf98c919d3a9e7bc50b4c1148.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 155, + 614, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/96ca2fd3d8a4586ad16dd6d1bead476365fa8c728a49d5a5d5d4b7d4a9efb473.jpg", + "image_caption": [ + "Full Method" + ], + "image_footnote": [], + "bbox": [ + 616, + 154, + 694, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/596f51289e632be8f3f0a3dd6581aa9c978a1d5b730263be74db7b0effb50674.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 154, + 774, + 214 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "by the emergence of a turtle-lizard hybrid. However, attempting to guide the latent signal without our Bounded Guidance mechanism may cause distortions in subject appearances, such as the turtle's neck growing out of its leg and inconsistent artifacts appearing on the lizard's head and back.", + "bbox": [ + 212, + 309, + 784, + 368 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Meanwhile, forgoing Bounded Denoising results in noticeable semantic leakage, with the lizard being replaced by a turtle, and the \"red\" attribute erroneously leaking to the wrong subject.", + "bbox": [ + 212, + 369, + 784, + 414 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Lastly, incorporating mask refinement in the later stages prevents fine-details from leaking. Without mask refinement, the lizard exhibits shell-like contours on its back.", + "bbox": [ + 212, + 415, + 784, + 458 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 470, + 367, + 486 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce Bounded Attention, a technique designed to regulate the accurate generation of multiple subjects within an image. This approach encourages each subject to \"be yourself\", emphasizing the importance of preserving individuality and uniqueness without being excessively influenced by other subjects present in the image. Our development of the Bounded Attention technique stemmed from an in-depth analysis of the root causes behind the misalignment observed between the provided prompt and the resulting generated image. Our investigation revealed that this misalignment primarily arises due to semantic leakage among the generated subjects, a phenomenon observed in both the cross and self-attention layers.", + "bbox": [ + 212, + 492, + 785, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While Bounded Attention effectively mitigates a significant portion of semantic leakage, it does not entirely eliminate it. Our findings demonstrate a marked improvement in performance compared to other methods that seek to achieve semantic alignment. However, residual leakage persists, which we attribute to imperfect optimization during the guidance mode and inaccurate segmentation of the subject prior to the second phase.", + "bbox": [ + 212, + 643, + 785, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While Bounded Attention excels in generating multiple subjects with plausible semantic alignment, its performance may vary across different layouts. Achieving success with Bounded Attention hinges on a strong match between the seed and the layout. Moving forward, we aim to explore methods for generating well-suited seeds tailored to specific layouts. One potential avenue is to introduce noise to the layout image, thereby creating a seed that aligns more closely with the desired outcomes.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "O. Dahary et al.", + "bbox": [ + 271, + 114, + 382, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 145, + 393, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Guy Tevet for fruitful discussions and useful suggestions. This work is partly supported by a research gift from Snap Inc.", + "bbox": [ + 215, + 176, + 784, + 207 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_model.json b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fedece00a9a8b1fda8d5cd4e0895a3c143079ead --- /dev/null +++ b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_model.json @@ -0,0 +1,1968 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.27, + 0.142, + 0.734, + 0.187 + ], + "angle": 0, + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.787, + 0.228 + ], + "angle": 0, + "content": "Omer Dahary1, Or Patashnik1,2, Kfir Aberman2, and Daniel Cohen-Or1,2" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.239, + 0.576, + 0.253 + ], + "angle": 0, + "content": "1 Tel Aviv University" + }, + { + "type": "text", + "bbox": [ + 0.445, + 0.253, + 0.558, + 0.268 + ], + "angle": 0, + "content": "2 Snap Research" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.291, + 0.641, + 0.307 + ], + "angle": 0, + "content": "3 ginger kittens and 2 gray kittens..." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.309, + 0.356, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.417, + 0.34, + 0.429 + ], + "angle": 0, + "content": "\"... on the stairs\"" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.31, + 0.499, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.417, + 0.485, + 0.428 + ], + "angle": 0, + "content": "\"...on the street\"" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.31, + 0.642, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.417, + 0.621, + 0.428 + ], + "angle": 0, + "content": "\"...in a basket\"" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.31, + 0.785, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.417, + 0.758, + 0.428 + ], + "angle": 0, + "content": "\"... on a tree\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.444, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Fig. 1: Our method bounds the attention to enable layout control over a pre-trained text-to-image diffusion model. Bounded Attention effectively reduces the impact of the innate semantic leakage during denoising, encouraging each subject to be itself. Our method can faithfully generate challenging layouts featuring multiple similar subjects with different modifiers (e.g., ginger and gray kittens)." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.536, + 0.741, + 0.813 + ], + "angle": 0, + "content": "Abstract. Text-to-image diffusion models have an unprecedented ability to generate diverse and high-quality images. However, they often struggle to faithfully capture the intended semantics of complex input prompts that include multiple subjects. Recently, numerous layout-to-image extensions have been introduced to improve user control, aiming to localize subjects represented by specific tokens. Yet, these methods often produce semantically inaccurate images, especially when dealing with multiple semantically or visually similar subjects. In this work, we study and analyze the causes of these limitations. Our exploration reveals that the primary issue stems from inadvertent semantic leakage between subjects in the denoising process. This leakage is attributed to the diffusion model's attention layers, which tend to blend the visual features of different subjects. To address these issues, we introduce Bounded Attention, a training-free method for bounding the information flow in the sampling process. Bounded Attention prevents detrimental leakage among subjects and enables guiding the generation to promote each subject's individuality, even with complex multi-subject conditioning. Through extensive experimentation, we demonstrate that our method empowers the generation of multiple subjects that better align with given prompts and layouts." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.826, + 0.741, + 0.84 + ], + "angle": 0, + "content": "Keywords: Image Generation \\(\\cdot\\) Diffusion Models \\(\\cdot\\) Semantic Alignment" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.129 + ], + "angle": 0, + "content": "O. Dahary et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.376, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.282 + ], + "angle": 0, + "content": "In recent years, text-to-image generation has undergone a significant shift with the integration of conditional diffusion models \\([?,?,?,?]\\), allowing for the facile generation of high-quality and diverse images. The use of attention layers in architectures of such generative models has been identified as a major factor contributing to improved quality of generated images \\([?,?]\\). However, these models struggle to accurately generate scenes containing multiple subjects, especially when they are semantically or visually similar." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.787, + 0.387 + ], + "angle": 0, + "content": "In this work, we study the problem of multi-subject image generation in attention-based diffusion models. Our contribution is twofold. First, we recognize the underlying reasons for the difficulty in generating images containing multiple subjects, especially those sharing semantic similarities. Second, building on our insights, we present a method aimed at mitigating semantic leakage in the generated images, allowing control over the generation of multiple subjects (see Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.787, + 0.478 + ], + "angle": 0, + "content": "We demonstrate an innate bias within the common attention-based architectures utilized in diffusion models, which predisposes them to leak visual features between subjects. In particular, the functionality of attention layers is designed to blend features across the image. Therefore, they inherently lead to information leakage between subjects. This phenomenon is particularly noticeable when subjects are semantically similar and, therefore, attend to each other (Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.787, + 0.583 + ], + "angle": 0, + "content": "A plethora of works tries to mitigate the cross-subject leakage issue, either by modifying the sampling process to better follow different subjects in the prompt \\([\\?,?]\\), or by coupling the global prompt with layout information via segmentation maps or bounding boxes labeled with subject classes or local prompts \\([\\?,?,?,?]\\). However, the majority of these methods still encounter difficulties in accurately aligning to input layouts, particularly in scenarios involving two or more semantically similar subjects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.734 + ], + "angle": 0, + "content": "In our approach, we guide the image generation with a spatial layout \\([?,?]\\). To address cross-subject leakage, we introduce the Bounded Attention mechanism, utilized during the denoising process to generate an image. This method bounds the influence of irrelevant visual and textual tokens on each pixel, which otherwise promotes leakage. By applying this mechanism, we encourage each subject to be itself, in the sense that it hinders the borrowing of features from other subjects in the scene. We show that bounding the attention is needed both in the cross- and self-attention layers. Moreover, we find additional architectural components that amplify leakage, modify their operation, and present remedies to them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We show that our method succeeds in facilitating control over multiple subjects within complex, coarse-grained layouts comprising numerous bounding boxes with similar semantics. Particularly challenging examples are demonstrated in Figure 1, where we successfully generate five kittens with a mix of adjectives. We conduct experiments on both Stable Diffusion [?] and SDXL [?] architectures and demonstrate the advantage of our method compared to previous ones, both supervised and unsupervised." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.145, + 0.505, + 0.284 + ], + "angle": 0, + "content": "Fig.2: Misalignment in layout-to-image generation include (i) catastrophic neglect [?] where the model fails to include one or more subjects mentioned in the prompt within the generated image, (ii) incorrect attribute binding [?,?] where attributes are not correctly matched to their corresponding subjects, and (iii) subject fusion [?] where the model merges multiple subjects into a single, larger subject." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.161, + 0.688, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.16, + 0.776, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.383, + 0.339 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Text-to-image diffusion models. Diffusion models, trained on huge datasets [?], have demonstrated their power in learning the complex distributions of diverse natural images [?,?,?,?,?] Augmenting attention layers into diffusion models and conditioning them on textual prompts by coupling them with visually-aligned text encoders [?] leads to powerful text-to-image models [?] In this work, we specifically examine two such open-source text-to-image diffusion models: Stable Diffusion [?], and the more recent SDXL [?]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.789, + 0.59 + ], + "angle": 0, + "content": "Semantic alignment in text-to-image synthesis. A critical drawback of current text-to-image models pertains to their limited ability to faithfully represent the precise semantics of input prompts. Various studies have identified common semantic failures and proposed mitigating strategies, such as adjusting text embeddings \\([?,?,?]\\), or optimizing noisy signals to strengthen or align cross-attention maps \\([?,?]\\). Nevertheless, these methods often fall short in generating multiple subjects, and do not adhere to positional semantics, such as subjects' number or location." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Layout-guided image synthesis. Addressing the semantic alignment concern, alternative approaches advocate for conditioning the diffusion process on layout information, either by training new models from scratch \\([?,?]\\) or fine-tuning an existing one \\([?,?,?,?]\\). Despite their promise, these methods demand extensive computational resources and prolonged training times. Moreover, they are constrained by the layout distribution of the training data and the models' architectural bias to blend subject features, a limitation that our Bounded Attention aims to overcome." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "To circumvent these challenges, numerous researchers explore training-free techniques, where the generation process itself is modified to enforce layout constraints. Several optimization-based works employ techniques similar to classifier-free guidance to localize the cross-attention \\([\\?,?,?,?]\\) and/or self-attention maps \\([\\?]\\). While effective in aligning random noise with the intended layout, guiding attention maps to distinct regions may inadvertently lead to undesired behavior, particularly when different subjects share similar semantics and visual features. Furthermore, these methods often exhibit a deteriorating effect on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.129 + ], + "angle": 0, + "content": "O. Dahary et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "visual quality, thereby limiting their applicability to only the initial denoising steps and neglecting finer control over shape and visual details, which are determined only in later stages [?]. Bounded Attention addresses these shortcomings by regulating attention computation throughout the entire denoising process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.785, + 0.358 + ], + "angle": 0, + "content": "Another approach involves generating each subject separately in its own denoising process \\([\\?,?]\\). While these methods inherently address catastrophic neglect, they tend to generate disharmonious images, and remain susceptible to leakage when merging subjects in subsequent stages. In contrast, masking attention maps to input bounding boxes \\([\\?]\\) or attenuating attention in specific segments \\([\\?]\\) represents a milder variant of this strategy, aimed at avoiding visible stitching. However, these methods often fall short of fully mitigating subject leakage, compromising semantic alignment. In comparison, Bounded Attention is able to carefully govern information propagation among subjects in a single denoising process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.787, + 0.466 + ], + "angle": 0, + "content": "While both the trained models and training-free techniques aim to generate numerous objects, they do not mitigate the inherent leakage caused by attention mechanisms. Unlike our Bounded Attention technique, these methods encounter challenges in effectively generating and controlling a multitude of subjects, especially when they share semantic similarity. Notably, existing techniques struggle to accurately generate even two semantically similar subjects, whereas our method, as demonstrated succeeds in generating five and even more subjects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.383, + 0.505 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.64 + ], + "angle": 0, + "content": "Latent diffusion models. In this work, we examine Stable Diffusion [?] and SDXL [?], which are both publicly available latent diffusion models. These models operate in the latent space of a pretrained image autoencoder, and are thus tasked with denoising a latent representation of the image, where each latent pixel corresponds to a patch in the generated image. Starting from pure random noise \\(\\mathbf{z}_T\\), at each timestep \\(t\\), the current noisy latent \\(\\mathbf{z}_t\\) is passed to a denoising UNet \\(\\epsilon_{\\theta}\\), trained to predict the current noise estimate \\(\\epsilon_{\\theta}(\\mathbf{z}_t,y,t)\\) using the guiding prompt encoding \\(y\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.649, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Attention layers. At each block, the UNet utilizes residual convolution layers, producing intermediate features \\(\\phi^{(l)}(\\mathbf{z}_t)\\), where \\(l\\) denotes the layer's index. These, in turn, are passed to attention layers, which essentially average different values \\(\\mathbf{V}_{\\mathbf{t}}^{(1)}\\) according to pixel-specific weights:" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.725, + 0.786, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\phi^ {(l + 1)} (\\mathbf {z} _ {t}) = \\mathbf {A} _ {t} ^ {(l)} \\mathbf {V} _ {t} ^ {(l)}, \\text {w h e r e} \\mathbf {A} _ {t} ^ {(l)} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} _ {t} ^ {(l)} \\mathbf {K} _ {t} ^ {(l) ^ {\\intercal}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Here, the keys \\(\\mathbf{K}_t^{(l)} = f_K^{(l)}\\left(\\mathbf{C}_t^{(l)}\\right)\\) and values \\(\\mathbf{V}_t^{(l)} = f_V^{(l)}\\left(\\mathbf{C}_{\\mathbf{t}}^{(\\mathbf{1})}\\right)\\) are linear projections of context vectors \\(\\mathbf{C}_t^{(l)}\\). In the cross-attention layers we inject semantic context from the prompt encoding \\(\\mathbf{C}_t^{(l)} \\equiv y\\), while self-attention layers utilize global information from the latent itself \\(\\mathbf{C}_t^{(l)} \\equiv \\phi^{(l)}(\\mathbf{z}_t)\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.273 + ], + "angle": 0, + "content": "The weighting scheme is determined by the attention map \\(\\mathbf{A}_t^{(l)}\\) which represents the probability of each pixel to semantically associate with a key-value pair. This is done by linearly projecting the latent noisy pixels to queries \\(\\mathbf{Q}_t^{(l)} = f_Q^{(l)}\\left(\\phi^{(l)}(\\mathbf{z}_t)\\right)\\) and computing their inner product with the keys. It has been widely demonstrated that the cross-attention maps are highly indicative of the semantic association between the image layout and the prompt tokens [?]. Meanwhile, the self-attention maps govern the correspondence between pixels, and thus form the image's structure [?]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.285, + 0.427, + 0.302 + ], + "angle": 0, + "content": "4 Semantic Leakage" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.787, + 0.339 + ], + "angle": 0, + "content": "We begin by studying the causes of semantic leakage in Stable Diffusion [?], and examine the limitations of existing layout-to-image approaches." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.348, + 0.445, + 0.363 + ], + "angle": 0, + "content": "4.1 On Subject Similarity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.785, + 0.406 + ], + "angle": 0, + "content": "Figure 2 illustrates various misalignment failures observed in state-of-the-art layout-to-image training-free methods. As we shall show, these failures are more prevalent for subjects that share semantic or visual similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.787, + 0.514 + ], + "angle": 0, + "content": "Let \\(x_{s_1}, x_{s_2} \\in \\mathbb{R}^2\\) be 2D latent coordinates corresponding to two semantically similar subjects \\(s_1, s_2\\) in the generated image. Intuitively, we expect that along the denoising process, the queries corresponding to these pixels, \\(\\mathbf{Q}_t^{(l)}[x_{s_1}], \\mathbf{Q}_t^{(l)}[x_{s_2}]\\), will be similar and hence also their attention responses. This, in turn, also implies that they will share semantic information from the token embeddings through the cross-attention layers or visual information via the self-attention layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.787, + 0.59 + ], + "angle": 0, + "content": "To explore this hypothesis, we investigate the model's behavior when tasked with generating two subjects and analyze their attention features in both cross- and self-attention layers. Subsequently, we meticulously examine these features and demonstrate how their behavior sheds light on the leakage observed in generated images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.601, + 0.466, + 0.615 + ], + "angle": 0, + "content": "4.2 Cross-Attention Leakage" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.764 + ], + "angle": 0, + "content": "To analyze the leakage caused by cross-attention layers, we examine the cross-attention queries. We depict these queries in the plots in Figure 3, where each point corresponds to a single query projected to 2D with PCA. To label each point with its corresponding subject, we compute the subject masks by averaging cross-attention maps [?] and color each projected query according to the subject's text color. The leftmost plot, in which the two subjects were generated separately, serves as a reference point to the relation between the queries when there is no leakage. For comparative analysis, we also present results for Layout-guidance [?], as a simple representative of current training-free layout-guided approaches, and Bounded Attention, which we shall cover in the sequel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We consider the cross-attention queries in two examples: \"a kitten\" and \"a puppy\", and \"a hamster\" and a \"squirrel\". As can be seen in the reference plots, the kitten and puppy queries share some small overlap, and the hamster and squirrel queries are mixed together. The level of separation between the red and blue dots in the plots reveals the semantic similarity of the two forming subjects." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.129 + ], + "angle": 0, + "content": "O. Dahary et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.486, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.144, + 0.782, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Fig. 3: Cross-Attention Leakage. We demonstrate the emergence of semantic leakage at the cross-attention layers. We show two examples: a puppy a kitten, and a hamster and a squirrel. In the two leftmost columns, the subjects were generated separately using Stable Diffusion (SD). In the right three columns, we generate a single image with the two subjects using three different methods: Stable Diffusion (SD), Layout Guidance (LG), and Bounded Attention (BA, ours). Under each row, we plot the two first principal components of the cross-attention queries. As can be seen, the separation of the queries (blue and red) reflects the leakage between the subjects in the generated images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.789, + 0.547 + ], + "angle": 0, + "content": "Clearly, vanilla Stable Diffusion struggles to adequately generate the two subjects within the same image. This is apparent by the visual leakage between the two subjects, as the model cannot avoid averaging their distinct visual properties. For example, the puppy has the visual features of a kitten, like raised ears and a triangular nose, while the squirrel loses its distinct ear shape and its feet are now pink like the hamster. Respectively, it can be seen that the queries of the two subjects are mixed, even for the kitten and the puppy which are more separated in the reference plot." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.789, + 0.686 + ], + "angle": 0, + "content": "Meanwhile, Layout Guidance (LG), which optimizes \\(\\mathbf{z}_t\\) to have attention peaks for each noun token at their corresponding region, exhibits interesting results. Its optimization objective implicitly encourages the separation between subjects' cross-attention queries. This can have positive effects, like the hamster and squirrel having unique colors, but at the same time yields unwanted artifacts, like the puppy losing its face altogether. Moreover, it can inadvertently push the latent signal out of distribution, causing quality degradation, as evident by the hamster's and squirrel's cartoonish texture. When it overly pushes the latent out of distribution, it leads to the catastrophic neglect phenomenon (Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.789, + 0.748 + ], + "angle": 0, + "content": "In comparison, when examining the plots of our method alongside the reference plots, our approach preserves the feature distribution of the subjects' queries, and successfully generates the two subjects, even when the queries are as mixed as in the hamster and the squirrel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.84 + ], + "angle": 0, + "content": "The above analysis yields two immediate conclusions: (i) Semantic similarity between subjects is reflected by their queries proximity, and leads to mixed queries when the subjects are generated together. This in turn leads to leakage between the subjects in the generated images, and (ii) enforcing semantic separation by modifying the semantic meaning of the cross-attention queries is harmful. The former observation represents a crucial architectural limitation in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.145, + 0.368, + 0.154 + ], + "angle": 0, + "content": "SD" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.155, + 0.774, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.294 + ], + "angle": 0, + "content": "Fig. 4: Self-Attention Leakage. We demonstrate the emergence of semantic leakage at the self-attention maps of two generally semantically dissimilar subjects: a crab and a frog. The images are generated by Stable Diffusion (SD) and Layout-guidance (LG). We highlight specific pixels, such as those of a subject's eye or leg, and present their respective self-attention maps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.785, + 0.368 + ], + "angle": 0, + "content": "current diffusion models, and the latter pinpoints to a previously unexplored weakness in the widely used latent optimization methodology. Bounded Attention is designed to overcome these limitations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.39, + 0.451, + 0.404 + ], + "angle": 0, + "content": "4.3 Self-Attention Leakage" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.785, + 0.464 + ], + "angle": 0, + "content": "We now turn to analyze the leakage caused by self-attention layers. It has been shown that self-attention features exhibit dense correspondences within the same subject \\([\\ref{e}]\\) and across semantic similar ones \\([\\ref{e}]\\). Hence, they are suspected as another source of leakage, that we shall study next." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.786, + 0.6 + ], + "angle": 0, + "content": "Here, we choose to examine the self-attention maps as a means to understand the leakage. In Figure 4 we focus on representative pixels (marked in yellow) associated with the subjects' eyes and legs, where visual leakage is most pronounced. As expected, features from one subject's eye or leg attend to the semantic similar body parts of the other. As a result, the features of each of the yellow points are directly affected by the features of the counterpart subject, causing leakage. In both images, the crab and the frog have similar appearances. In vanilla SD, both have a crab-like color and limbs with frog's toe pads. In LG, both have frog-like eyes and crab legs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.786, + 0.72 + ], + "angle": 0, + "content": "Notably, this tendency to rely on similar patches aids the model in denoising the latent signal and is essential for achieving coherent images with properly blended subjects and backgrounds. Nevertheless, it has the drawback of leaking features between disjointed subjects. Therefore, completely disjointing the visual features during denoising \\([?]\\), or naively pushing self-attention queries apart through optimization \\([?]\\), can lead to subpar results. Consequently, we introduce the Bounded Attention mechanism to mitigate leakage and guide the latent signal towards subject separability, while avoiding detrimental artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.721, + 0.787, + 0.782 + ], + "angle": 0, + "content": "It's important to highlight that the leakage resulting from both cross- and self-attention layers is intertwined and mutually reinforced. Therefore, addressing the leakage caused by only one of these layers is insufficient to prevent the leakage in the generated images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.796, + 0.422, + 0.81 + ], + "angle": 0, + "content": "4.4 Levels of Similarity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the two previous sections, our focus was primarily on subjects that share semantic similarity. Building upon the observation that the UNet's inner layers" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.129 + ], + "angle": 0, + "content": "O. Dahary et al." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.144, + 0.479, + 0.156 + ], + "angle": 0, + "content": "A kitten A puppy \\(64 \\times 64\\) \\(32 \\times 32\\) \\(16 \\times 16\\)" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.156, + 0.479, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.144, + 0.761, + 0.156 + ], + "angle": 0, + "content": "A lizard A fruit \\(64 \\times 64\\) \\(32 \\times 32\\) \\(16 \\times 16\\)" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.156, + 0.764, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.212, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Fig. 5: We generate different subjects, and plot the first two principal components of the cross-attention queries at different layers of the UNet, where each layer is of different resolution. The high semantic similarity between the kitten and the puppy is expressed by the proximity of their queries through all layers. Meanwhile, the lizard and fruit share similar texture, and hence only their high-resolution queries are entangled." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.784, + 0.389 + ], + "angle": 0, + "content": "dictate the subject's semantics and shape, while the outer layers control its style and appearance [?], we analyzed the inner UNet layers. However, leakage can also occur prominently when generating subjects that share visual similarity rather than semantic. We now turn to explore this scenario and demonstrate that, in such cases, the leakage originates from the UNet's outer layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.543 + ], + "angle": 0, + "content": "In Figure 5, we visualize cross-attention queries at different decoder's layers, when generating the kitten, puppy, lizard, and fruit in isolation. As can be seen, the queries of the kitten and the puppy are mixed across all UNet's layers, aligning with the visual and semantic similarity between these animals. On the other hand, the queries of the lizard and the fruit are overlapped only in the highest-resolution layer, aligning with the lack of semantic similarity between them. A closer look reveals that the lizard and the fruit share surprisingly similar textures, which explains the overlap of the queries in the highest-resolution layer. As explained in the previous sections, this overlap causes leakage between the two subjects, in this case, a visual rather than a semantic leakage (see Figure 2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.441, + 0.584 + ], + "angle": 0, + "content": "5 Bounded Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Our method takes as input \\( n \\) distinct textual subjects \\( S = \\{s_i\\}_{i=1}^n \\) contained within a global prompt \\( y \\), along with their corresponding bounding boxes \\( B = \\{b_i\\}_{i=1}^n \\). Our objective is to condition the generation process on \\( y \\), \\( S \\), and \\( B \\), while preserving the intended semantics of each subject, all without requiring any training or fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Figure 6 illustrates an overview of our method. There, the input prompt \\( y \\) is \"A kitten and a puppy\", \\( S = \\{kitten\\}, \\{puppy\\} \\), and the two corresponding bounding boxes \\( \\{b_1, b_2\\} \\) are illustrated in the top left corner. Bounded Attention operates in two modes: Bounded Guidance and Bounded Denoising. Specifically, at the beginning of the denoising process, for \\( t \\in [T, T_{\\text{guidance}}] \\), we perform a Bounded Guidance step followed by a Bounded Denoising step. In the guidance step, we utilize the Bounded Guidance loss. This interval of timesteps constitutes the optimization phase. Then, for \\( t \\in [T_{\\text{guidance}}, 0] \\), we apply only Bounded Denoising steps. In both modes, we manipulate the model's forward pass by adopting an augmented weighting scheme in the attention layers, that safeguard the flow of information between the queries and the keys:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.148, + 0.734, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig. 6: Bounded Attention operates in two modes: guidance and denoising, each imposing strict constraints to bound the attention of each subject solely to itself and possibly to the background, preventing any influence from other subjects' features. In guidance mode, a loss is minimized to encourage each subject's attention to concentrate within its corresponding bounding box, with other bounding boxes masked to prevent artifacts. During the denoising step, we confine the attention of each subject solely to its bounding box, along with the background in the self-attention. This strategy effectively prevents feature leakage while maintaining the natural immersion of the subject within the image. Attention maps of a specific key, marked with \\(\\star\\), are shown for each mode: the cross-attention map displays the key corresponding to the \"kitten\" token, and the self-attention map shows a key within the kitten's target bounding box." + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.564, + 0.787, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {t} ^ {(l)} = \\mathrm {s o f t m a x} \\left(\\mathbf {Q} _ {t} ^ {(l)} \\mathbf {K} _ {t} ^ {(l) ^ {\\intercal}} + \\mathbf {M} _ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.727 + ], + "angle": 0, + "content": "where \\(l\\) represents the layer index, \\(t\\) represents the diffusion timestep, and \\(\\mathbf{Q}_t^{(l)},\\mathbf{K}_t^{(l)}\\) are the queries and keys of the \\(l\\)-th attention layer. \\(\\mathbf{M}_t\\) represents time-specific masks composed of zeros and \\(-\\infty\\) elements. We refer to \\(\\mathbf{A}_t^{(l)}\\) as the Bounded Attention map. When indexing \\(\\mathbf{A}_t^{(l)}\\), we use pixel coordinates \\(\\mathbf{x}\\) for rows, and attention-type-specific context vectors \\(\\mathbf{c}\\) for columns. In locations \\([\\mathbf{x},\\mathbf{c}]\\), where \\(\\mathbf{M}_t[\\mathbf{x},\\mathbf{c}] = -\\infty\\), it holds that \\(\\mathbf{A}_t^{(l)}[\\mathbf{x},\\mathbf{c}] = 0\\). Therefore, these masks prevent harmful information flow between pixels in self-attention layers, and between pixels and token embeddings in cross-attention layers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.733, + 0.421, + 0.746 + ], + "angle": 0, + "content": "5.1 Bounded Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Bounded Guidance, we backpropagate through the diffusion model to steer the latent signal toward the desired layout, using Gradient Descent. Our Bounded Guidance loss encourages the Bounded Attention map of each key corresponding to subject \\( s_i \\), to be within the \\( b_i \\) bounding box. To this end, for each subject key we consider the ratio between the attention within the corresponding bounding box, to the entire Bounded Attention map (see Figure 6). Formally, we aggregate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.129 + ], + "angle": 0, + "content": "O. Dahary et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.525, + 0.163 + ], + "angle": 0, + "content": "the following loss on the different subjects:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.17, + 0.786, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} = 1 - \\frac {\\sum_ {\\mathbf {x} \\in b _ {i} , \\mathbf {c} \\in C _ {i}} \\hat {\\mathbf {A}} [ \\mathbf {x} , \\mathbf {c} ]}{\\sum_ {\\mathbf {x} \\in b _ {i} , \\mathbf {c} \\in C _ {i}} \\hat {\\mathbf {A}} [ \\mathbf {x} , \\mathbf {c} ] + \\alpha \\sum_ {\\mathbf {x} \\notin b _ {i} , \\mathbf {c} \\in C _ {i}} \\hat {\\mathbf {A}} [ \\mathbf {x} , \\mathbf {c} ]}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.213, + 0.784, + 0.348 + ], + "angle": 0, + "content": "where \\( i \\) denotes the index of subject \\( s_i \\), \\( \\hat{\\mathbf{A}} \\) is the mean Bounded Attention map, averaged across heads and layers, and \\( \\alpha \\) is a hyperparameter that magnifies the significance of disregarding attention towards the background, as we explain later. Similarly to the above, When indexing \\( \\hat{\\mathbf{A}} \\), pixel coordinates \\( \\mathbf{x} \\) represent rows, and attention-type-specific context vectors \\( \\mathbf{c} \\) represent columns. We designate \\( C_i \\) as the set of all \\( s_i \\)-related context vectors, i.e., pixel coordinates in \\( b_i \\) for self-attention layers, and the token identifiers of \\( s_i \\) in cross-attention layers. Additionally, for cross-attention layers, we include the first padding token [EoT] in \\( C_i \\) to enhance layout alignment [?]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.785, + 0.425 + ], + "angle": 0, + "content": "For each subject \\( s_i \\), the mask \\( \\mathbf{M}_t \\) should block the influence of opposing keys (tokens in \\( s_j \\) and pixels in \\( b_j \\) for \\( j \\neq i \\)), to avoid the artifacts illustrated in Figure 3. This fosters that the queries of different subjects, including similar ones, are not erroneously forced to be far apart. Utilizing this loss, our Bounded Guidance step is defined as \\( z_t^{\\mathrm{opt}} = z_t - \\beta \\nabla_{z_t} \\sum_i \\mathcal{L}_i^2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.785, + 0.529 + ], + "angle": 0, + "content": "Integrating this loss within the cross-attention layers encourages the localization of each subject's semantics within its bounding boxes [?]. However, as cross-attention responses tend to peak around more typical patches associated with the subject's semantics (e.g., the face of a human, the legs of a crab, etc.), it may lack control over the subject's boundaries. By applying the loss within the self-attention layers, we encourage each subject to establish its own boundaries close to its bounding box, thereby discouraging subject fusion (see Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.785, + 0.606 + ], + "angle": 0, + "content": "In the computation of the loss, we also introduce a hyperparameter \\(\\alpha\\) to reinforce attention to the background. This adjustment aids in preventing subject amalgamation, where a redundant subject is realized from different subject semantics in the background. To preserve image quality, we limit the application of this mode to an initial time interval \\([T, T_{\\text{guidance}}]\\), following similar works [?,?]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.424, + 0.629 + ], + "angle": 0, + "content": "5.2 Bounded Denoising" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.785, + 0.72 + ], + "angle": 0, + "content": "In Bounded Denoising, we compute the model's output and use it as the next latent in the series. Here, the masks aim to reduce semantic leakage between subjects, as detailed in Section 4, and to prevent unintended semantics from leaking to the background. Unlike Bounded Guidance and typical attention-based guidance approaches, Bounded Denoising can be applied throughout all time steps to mitigate leaks in fine details, which emerge only in later stages [?]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.809 + ], + "angle": 0, + "content": "However, coarse masking in later stages may degrade image quality and result in noticeable stitching. To address this, after the optimization phase, for \\( t \\in [T_{\\text{guidance}}, 0] \\), we replace each bounding box with a fine segmentation mask obtained by clustering the self-attention maps [?]. Since the subject outlines are roughly determined in the initial time steps and evolve gradually thereafter [?], we refine the masks periodically." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Notably, this mechanism also addresses imperfect alignments between subjects and bounding boxes after the guidance phase, which are more common" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.747, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.334, + 0.743, + 0.348 + ], + "angle": 0, + "content": "Fig. 7: Our results with SDXL. See the supplemental for more examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.472 + ], + "angle": 0, + "content": "when generating numerous subjects. Thus, employing this technique enhances the robustness of our method to seed selection, ensuring proper semantics even when subjects extend beyond their initial confines (see Figure 1,7). Compared to methods that necessitate strict input masks yet remain susceptible to leakage \\([?,?,?]\\), our method offers greater user control with simpler inputs and more satisfactory outcomes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.482, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Method Details. Further details on the adaptation of Bounded Attention to the cross- and self-attention layers, along with a description of the subject mask refinement process, are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.548, + 0.377, + 0.566 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.701 + ], + "angle": 0, + "content": "In this section, we conduct both qualitative and quantitative experiments to assess the effectiveness of our Bounded Attention method. We compare our approach with three training-free baseline methods: Layout-guidance (LG) [?], BoxDiff (BD) [?], and MultiDiffusion (MD) [?]. Additionally, we include comparisons with GLIGEN [?] and ReCo [?], which necessitate training. Since Attention-refocusing (AR) [?] is based on GLIGEN, we categorize it as a trained method for the purpose of our evaluation. For fairness, when comparing our method with other methods, we use Stable Diffusion." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.722, + 0.422, + 0.737 + ], + "angle": 0, + "content": "6.1 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "SDXL results. We begin our experiments by demonstrating the efficacy of our method in challenging scenarios, particularly when tasked with generating multiple semantically similar subjects using SDXL. As can be seen in Figure 7, Vanilla SDXL fails to follow the prompt due to semantic leakage. It generates inaccurate number of dogs and mixes the features of each breed. In comparison, our approach adeptly generates each dog with its unique characteristics." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "O. Dahary et al." + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.143, + 0.683, + 0.155 + ], + "angle": 0, + "content": "\"A gray kitten and a ginger kitten and black puppy in a yard.\"" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.156, + 0.77, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.332, + 0.731, + 0.346 + ], + "angle": 0, + "content": "Fig. 8: Comparison of the first six images generated from the seed 0." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.379, + 0.787, + 0.453 + ], + "angle": 0, + "content": "Non-curated results. Next, we conduct a non-curated comparison with the training-free baseline methods and present the results in Figure 8. We showcase the initial six images sampled from seed 0 for each method. We anticipate that in the absence of Bounded Attention, semantic leakage may freely blend subject features, hindering the intended layout's formation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.592 + ], + "angle": 0, + "content": "It is evident from the results that none of the competing methods is able to consistently construct the input layout. Layout Guidance [?] frequently neglects one of the subjects, and even when it generates three subjects, it struggles to avoid leakage, resulting in puppies with kitten-like features or incorrect color assignments. BoxDiff [?] often generates the correct number of subjects but suffers from artifacts in the form of blobs. Similar to Layout Guidance, it encounters difficulties in properly constructing the puppy. Surprisingly, even MultiDiffusion [?], which generates the subjects separately, faces challenges in generating them all, with some disappearing or merging together in its bootstrapping phase." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.593, + 0.787, + 0.623 + ], + "angle": 0, + "content": "In contrast, our method consistently outperforms these approaches, producing three subjects that align with the both prompt and layout in all six images." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.642, + 0.787, + 0.734 + ], + "angle": 0, + "content": "Comparisons with baselines. We present a qualitative comparison in Figure 9. All competing methods, including those trained specifically for the layout-to-image task, exhibit significant visual and semantic leakage. The training-free methods perform the worst: MultiDiffusion produces disharmonious, low-quality images, while optimization-based methods often result in object fusion, combining different semantics without adhering to the layout." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.734, + 0.787, + 0.794 + ], + "angle": 0, + "content": "The training-based approaches closely follow the layout but fail to convey the correct semantics. In the first row, they neglect the corduroy jacket, leaking the denim texture into the other jacket, or even fusing them together. In the second row, the elephant's features leak into the rhino." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In comparison, our method generates images that align with the input layout and prompt, ensuring each subject retains its unique attributes, semantics, and appearance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.145, + 0.744, + 0.159 + ], + "angle": 0, + "content": "\"A denim jacket and a corduroy jacket and a leather handbag in a closet.\"" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.159, + 0.785, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.224, + 0.68, + 0.238 + ], + "angle": 0, + "content": "\"A big red elephant and a far away rhino in a jungle.\"" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.238, + 0.785, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 9: Qualitative comparison of our method with baseline methods: Above each row, we display the input prompt, where each subject's color matches the color of its corresponding bounding box. We compare with both training-free methods (2nd-4th columns) and trained models (5th-7th columns). See supplemental for more results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.454, + 0.432, + 0.468 + ], + "angle": 0, + "content": "6.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.432, + 0.619 + ], + "angle": 0, + "content": "We evaluate our method's effectiveness using the Draw-Bench dataset [?], known for its challenging prompts designed to test a model's ability to compose multiple subjects with specific quantities and relations. We use the evaluation procedure from previous work [?, ?]. Our re" + }, + { + "type": "table", + "bbox": [ + 0.465, + 0.473, + 0.767, + 0.567 + ], + "angle": 0, + "content": "
MethodCountingSpatial
PrecisionRecallF1Accuracy
Stable Diffusion [?]0.740.780.730.19
Layout-guidance [?]0.720.780.720.35
BoxDiff [?]0.810.780.760.28
MultiDiffusion [?]0.700.550.570.15
Ours0.810.910.820.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.451, + 0.577, + 0.776, + 0.591 + ], + "angle": 0, + "content": "Table 1: Evaluation on the DrawBench dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.709 + ], + "angle": 0, + "content": "sults, alongside those of other training-free methods, are summarized in Table 1. Unlike other approaches that do not account for semantic leakage, our method demonstrates notable improvements in both the counting and the spatial categories. Notably, other methods struggle to surpass the counting recall rates of vanilla SD, highlighting the effectiveness of Bounded Attention in generating multiple subjects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.709, + 0.787, + 0.74 + ], + "angle": 0, + "content": "We further conduct a quantitative comparison of semantic leakage in training-free methods through a user study. Further details are in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.751, + 0.402, + 0.765 + ], + "angle": 0, + "content": "6.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.81 + ], + "angle": 0, + "content": "To assess the significance of each component, we conduct an ablation study where we systematically vary our method's configurations by omitting one component in each setting. We show the results in Figure 10." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Guidance is crucial for aligning the latent signal with the intended layout. Neglecting it can lead to subject generation in the background, as demonstrated" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.129 + ], + "angle": 0, + "content": "O. Dahary et al." + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.144, + 0.615, + 0.155 + ], + "angle": 0, + "content": "\"A red lizard and a turtle on the grass.\"" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.155, + 0.303, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.217, + 0.284, + 0.228 + ], + "angle": 0, + "content": "w/o G" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.156, + 0.382, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.217, + 0.367, + 0.228 + ], + "angle": 0, + "content": "w/o BG" + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.156, + 0.461, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.217, + 0.445, + 0.228 + ], + "angle": 0, + "content": "w/o BD" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.156, + 0.538, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.217, + 0.564, + 0.228 + ], + "angle": 0, + "content": "w/o MR" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.156, + 0.616, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.155, + 0.695, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.155, + 0.775, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.217, + 0.732, + 0.228 + ], + "angle": 0, + "content": "Full Method" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.239, + 0.787, + 0.296 + ], + "angle": 0, + "content": "Fig. 10: Qualitative ablation. We ablate our method by skipping the guidance step (G), performing guidance without Bounded Guidance (BG), not applying Bounded Denoising (BD), and not performing Mask Refinement (MR). We show zoomed-in images of the two rightmost configurations. More examples can be found in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.369 + ], + "angle": 0, + "content": "by the emergence of a turtle-lizard hybrid. However, attempting to guide the latent signal without our Bounded Guidance mechanism may cause distortions in subject appearances, such as the turtle's neck growing out of its leg and inconsistent artifacts appearing on the lizard's head and back." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.415 + ], + "angle": 0, + "content": "Meanwhile, forgoing Bounded Denoising results in noticeable semantic leakage, with the lizard being replaced by a turtle, and the \"red\" attribute erroneously leaking to the wrong subject." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.459 + ], + "angle": 0, + "content": "Lastly, incorporating mask refinement in the later stages prevents fine-details from leaking. Without mask refinement, the lizard exhibits shell-like contours on its back." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.471, + 0.369, + 0.487 + ], + "angle": 0, + "content": "7 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.644 + ], + "angle": 0, + "content": "We introduce Bounded Attention, a technique designed to regulate the accurate generation of multiple subjects within an image. This approach encourages each subject to \"be yourself\", emphasizing the importance of preserving individuality and uniqueness without being excessively influenced by other subjects present in the image. Our development of the Bounded Attention technique stemmed from an in-depth analysis of the root causes behind the misalignment observed between the provided prompt and the resulting generated image. Our investigation revealed that this misalignment primarily arises due to semantic leakage among the generated subjects, a phenomenon observed in both the cross and self-attention layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.735 + ], + "angle": 0, + "content": "While Bounded Attention effectively mitigates a significant portion of semantic leakage, it does not entirely eliminate it. Our findings demonstrate a marked improvement in performance compared to other methods that seek to achieve semantic alignment. However, residual leakage persists, which we attribute to imperfect optimization during the guidance mode and inaccurate segmentation of the subject prior to the second phase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "While Bounded Attention excels in generating multiple subjects with plausible semantic alignment, its performance may vary across different layouts. Achieving success with Bounded Attention hinges on a strong match between the seed and the layout. Moving forward, we aim to explore methods for generating well-suited seeds tailored to specific layouts. One potential avenue is to introduce noise to the layout image, thereby creating a seed that aligns more closely with the desired outcomes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.395, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.178, + 0.785, + 0.208 + ], + "angle": 0, + "content": "We thank Guy Tevet for fruitful discussions and useful suggestions. This work is partly supported by a research gift from Snap Inc." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.231, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + } + ] +] \ No newline at end of file diff --git a/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_origin.pdf b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0ed8d481282f95227ccf50d7ffd268be20b67d6f --- /dev/null +++ b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/4961345c-6650-4e30-b5d4-2d25eabb4019_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4bd327f6bae10046a52038d184716da2645175647d69d0829db797af4fe9c75 +size 2334042 diff --git a/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/full.md b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..423146961ae8f33587bf4fe29e2cbb22919b0a4c --- /dev/null +++ b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/full.md @@ -0,0 +1,281 @@ +# Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation + +Omer Dahary1, Or Patashnik1,2, Kfir Aberman2, and Daniel Cohen-Or1,2 + +1 Tel Aviv University + +2 Snap Research + +3 ginger kittens and 2 gray kittens... + +![](images/458d9d6bc89265b09d53cdac1017211b3ad601bd8f74190ec9496699c6819862.jpg) +"... on the stairs" + +![](images/9332232f51f07a12917bb8a8cedba40642fa194b4a49e63bd7781902d1baaf00.jpg) +"...on the street" + +![](images/0f1ca0dbef871026421dd1dabc665c1ee1879bf09633ac24e7c14f0d84867441.jpg) +"...in a basket" +Fig. 1: Our method bounds the attention to enable layout control over a pre-trained text-to-image diffusion model. Bounded Attention effectively reduces the impact of the innate semantic leakage during denoising, encouraging each subject to be itself. Our method can faithfully generate challenging layouts featuring multiple similar subjects with different modifiers (e.g., ginger and gray kittens). + +![](images/bc74ce6b4fc1619b6e9df50dcd7d20364af2f282ec035ac5477471e887318c45.jpg) +"... on a tree" + +Abstract. Text-to-image diffusion models have an unprecedented ability to generate diverse and high-quality images. However, they often struggle to faithfully capture the intended semantics of complex input prompts that include multiple subjects. Recently, numerous layout-to-image extensions have been introduced to improve user control, aiming to localize subjects represented by specific tokens. Yet, these methods often produce semantically inaccurate images, especially when dealing with multiple semantically or visually similar subjects. In this work, we study and analyze the causes of these limitations. Our exploration reveals that the primary issue stems from inadvertent semantic leakage between subjects in the denoising process. This leakage is attributed to the diffusion model's attention layers, which tend to blend the visual features of different subjects. To address these issues, we introduce Bounded Attention, a training-free method for bounding the information flow in the sampling process. Bounded Attention prevents detrimental leakage among subjects and enables guiding the generation to promote each subject's individuality, even with complex multi-subject conditioning. Through extensive experimentation, we demonstrate that our method empowers the generation of multiple subjects that better align with given prompts and layouts. + +Keywords: Image Generation $\cdot$ Diffusion Models $\cdot$ Semantic Alignment + +# 1 Introduction + +In recent years, text-to-image generation has undergone a significant shift with the integration of conditional diffusion models $[?,?,?,?]$ , allowing for the facile generation of high-quality and diverse images. The use of attention layers in architectures of such generative models has been identified as a major factor contributing to improved quality of generated images $[?,?]$ . However, these models struggle to accurately generate scenes containing multiple subjects, especially when they are semantically or visually similar. + +In this work, we study the problem of multi-subject image generation in attention-based diffusion models. Our contribution is twofold. First, we recognize the underlying reasons for the difficulty in generating images containing multiple subjects, especially those sharing semantic similarities. Second, building on our insights, we present a method aimed at mitigating semantic leakage in the generated images, allowing control over the generation of multiple subjects (see Figure 1). + +We demonstrate an innate bias within the common attention-based architectures utilized in diffusion models, which predisposes them to leak visual features between subjects. In particular, the functionality of attention layers is designed to blend features across the image. Therefore, they inherently lead to information leakage between subjects. This phenomenon is particularly noticeable when subjects are semantically similar and, therefore, attend to each other (Figure 2). + +A plethora of works tries to mitigate the cross-subject leakage issue, either by modifying the sampling process to better follow different subjects in the prompt $[\?,?]$ , or by coupling the global prompt with layout information via segmentation maps or bounding boxes labeled with subject classes or local prompts $[\?,?,?,?]$ . However, the majority of these methods still encounter difficulties in accurately aligning to input layouts, particularly in scenarios involving two or more semantically similar subjects. + +In our approach, we guide the image generation with a spatial layout $[?,?]$ . To address cross-subject leakage, we introduce the Bounded Attention mechanism, utilized during the denoising process to generate an image. This method bounds the influence of irrelevant visual and textual tokens on each pixel, which otherwise promotes leakage. By applying this mechanism, we encourage each subject to be itself, in the sense that it hinders the borrowing of features from other subjects in the scene. We show that bounding the attention is needed both in the cross- and self-attention layers. Moreover, we find additional architectural components that amplify leakage, modify their operation, and present remedies to them. + +We show that our method succeeds in facilitating control over multiple subjects within complex, coarse-grained layouts comprising numerous bounding boxes with similar semantics. Particularly challenging examples are demonstrated in Figure 1, where we successfully generate five kittens with a mix of adjectives. We conduct experiments on both Stable Diffusion [?] and SDXL [?] architectures and demonstrate the advantage of our method compared to previous ones, both supervised and unsupervised. + +![](images/14e4ea5dbb99eed94036ac648061a206307e46c5eadf11efc71ba276ca218eb5.jpg) +Fig.2: Misalignment in layout-to-image generation include (i) catastrophic neglect [?] where the model fails to include one or more subjects mentioned in the prompt within the generated image, (ii) incorrect attribute binding [?,?] where attributes are not correctly matched to their corresponding subjects, and (iii) subject fusion [?] where the model merges multiple subjects into a single, larger subject. + +![](images/ab0f05d519ab46ce516e55615ffdd6925ac0cb61fa077cfe3cad7a62060939a0.jpg) + +# 2 Related work + +Text-to-image diffusion models. Diffusion models, trained on huge datasets [?], have demonstrated their power in learning the complex distributions of diverse natural images [?,?,?,?,?] Augmenting attention layers into diffusion models and conditioning them on textual prompts by coupling them with visually-aligned text encoders [?] leads to powerful text-to-image models [?] In this work, we specifically examine two such open-source text-to-image diffusion models: Stable Diffusion [?], and the more recent SDXL [?] + +Semantic alignment in text-to-image synthesis. A critical drawback of current text-to-image models pertains to their limited ability to faithfully represent the precise semantics of input prompts. Various studies have identified common semantic failures and proposed mitigating strategies, such as adjusting text embeddings $[?,?,?]$ , or optimizing noisy signals to strengthen or align cross-attention maps $[?,?]$ . Nevertheless, these methods often fall short in generating multiple subjects, and do not adhere to positional semantics, such as subjects' number or location. + +Layout-guided image synthesis. Addressing the semantic alignment concern, alternative approaches advocate for conditioning the diffusion process on layout information, either by training new models from scratch $[?,?]$ or fine-tuning an existing one $[?,?,?,?]$ . Despite their promise, these methods demand extensive computational resources and prolonged training times. Moreover, they are constrained by the layout distribution of the training data and the models' architectural bias to blend subject features, a limitation that our Bounded Attention aims to overcome. + +To circumvent these challenges, numerous researchers explore training-free techniques, where the generation process itself is modified to enforce layout constraints. Several optimization-based works employ techniques similar to classifier-free guidance to localize the cross-attention $[\?,?,?,?]$ and/or self-attention maps $[\?]$ . While effective in aligning random noise with the intended layout, guiding attention maps to distinct regions may inadvertently lead to undesired behavior, particularly when different subjects share similar semantics and visual features. Furthermore, these methods often exhibit a deteriorating effect on + +visual quality, thereby limiting their applicability to only the initial denoising steps and neglecting finer control over shape and visual details, which are determined only in later stages [?]. Bounded Attention addresses these shortcomings by regulating attention computation throughout the entire denoising process. + +Another approach involves generating each subject separately in its own denoising process $[\?,?]$ . While these methods inherently address catastrophic neglect, they tend to generate disharmonious images, and remain susceptible to leakage when merging subjects in subsequent stages. In contrast, masking attention maps to input bounding boxes $[\?]$ or attenuating attention in specific segments $[\?]$ represents a milder variant of this strategy, aimed at avoiding visible stitching. However, these methods often fall short of fully mitigating subject leakage, compromising semantic alignment. In comparison, Bounded Attention is able to carefully govern information propagation among subjects in a single denoising process. + +While both the trained models and training-free techniques aim to generate numerous objects, they do not mitigate the inherent leakage caused by attention mechanisms. Unlike our Bounded Attention technique, these methods encounter challenges in effectively generating and controlling a multitude of subjects, especially when they share semantic similarity. Notably, existing techniques struggle to accurately generate even two semantically similar subjects, whereas our method, as demonstrated succeeds in generating five and even more subjects. + +# 3 Preliminaries + +Latent diffusion models. In this work, we examine Stable Diffusion [?] and SDXL [?], which are both publicly available latent diffusion models. These models operate in the latent space of a pretrained image autoencoder, and are thus tasked with denoising a latent representation of the image, where each latent pixel corresponds to a patch in the generated image. Starting from pure random noise $\mathbf{z}_T$ , at each timestep $t$ , the current noisy latent $\mathbf{z}_t$ is passed to a denoising UNet $\epsilon_{\theta}$ , trained to predict the current noise estimate $\epsilon_{\theta}(\mathbf{z}_t,y,t)$ using the guiding prompt encoding $y$ . + +Attention layers. At each block, the UNet utilizes residual convolution layers, producing intermediate features $\phi^{(l)}(\mathbf{z}_t)$ , where $l$ denotes the layer's index. These, in turn, are passed to attention layers, which essentially average different values $\mathbf{V}_{\mathbf{t}}^{(1)}$ according to pixel-specific weights: + +$$ +\phi^ {(l + 1)} (\mathbf {z} _ {t}) = \mathbf {A} _ {t} ^ {(l)} \mathbf {V} _ {t} ^ {(l)}, \text {w h e r e} \mathbf {A} _ {t} ^ {(l)} = \operatorname {s o f t m a x} \left(\mathbf {Q} _ {t} ^ {(l)} \mathbf {K} _ {t} ^ {(l) ^ {\intercal}}\right). \tag {1} +$$ + +Here, the keys $\mathbf{K}_t^{(l)} = f_K^{(l)}\left(\mathbf{C}_t^{(l)}\right)$ and values $\mathbf{V}_t^{(l)} = f_V^{(l)}\left(\mathbf{C}_{\mathbf{t}}^{(\mathbf{1})}\right)$ are linear projections of context vectors $\mathbf{C}_t^{(l)}$ . In the cross-attention layers we inject semantic context from the prompt encoding $\mathbf{C}_t^{(l)} \equiv y$ , while self-attention layers utilize global information from the latent itself $\mathbf{C}_t^{(l)} \equiv \phi^{(l)}(\mathbf{z}_t)$ . + +The weighting scheme is determined by the attention map $\mathbf{A}_t^{(l)}$ which represents the probability of each pixel to semantically associate with a key-value pair. This is done by linearly projecting the latent noisy pixels to queries $\mathbf{Q}_t^{(l)} = f_Q^{(l)}\left(\phi^{(l)}(\mathbf{z}_t)\right)$ and computing their inner product with the keys. It has been widely demonstrated that the cross-attention maps are highly indicative of the semantic association between the image layout and the prompt tokens [?]. Meanwhile, the self-attention maps govern the correspondence between pixels, and thus form the image's structure [?]. + +# 4 Semantic Leakage + +We begin by studying the causes of semantic leakage in Stable Diffusion [?], and examine the limitations of existing layout-to-image approaches. + +# 4.1 On Subject Similarity + +Figure 2 illustrates various misalignment failures observed in state-of-the-art layout-to-image training-free methods. As we shall show, these failures are more prevalent for subjects that share semantic or visual similarity. + +Let $x_{s_1}, x_{s_2} \in \mathbb{R}^2$ be 2D latent coordinates corresponding to two semantically similar subjects $s_1, s_2$ in the generated image. Intuitively, we expect that along the denoising process, the queries corresponding to these pixels, $\mathbf{Q}_t^{(l)}[x_{s_1}], \mathbf{Q}_t^{(l)}[x_{s_2}]$ , will be similar and hence also their attention responses. This, in turn, also implies that they will share semantic information from the token embeddings through the cross-attention layers or visual information via the self-attention layers. + +To explore this hypothesis, we investigate the model's behavior when tasked with generating two subjects and analyze their attention features in both cross- and self-attention layers. Subsequently, we meticulously examine these features and demonstrate how their behavior sheds light on the leakage observed in generated images. + +# 4.2 Cross-Attention Leakage + +To analyze the leakage caused by cross-attention layers, we examine the cross-attention queries. We depict these queries in the plots in Figure 3, where each point corresponds to a single query projected to 2D with PCA. To label each point with its corresponding subject, we compute the subject masks by averaging cross-attention maps [?] and color each projected query according to the subject's text color. The leftmost plot, in which the two subjects were generated separately, serves as a reference point to the relation between the queries when there is no leakage. For comparative analysis, we also present results for Layout-guidance [?], as a simple representative of current training-free layout-guided approaches, and Bounded Attention, which we shall cover in the sequel. + +We consider the cross-attention queries in two examples: "a kitten" and "a puppy", and "a hamster" and a "squirrel". As can be seen in the reference plots, the kitten and puppy queries share some small overlap, and the hamster and squirrel queries are mixed together. The level of separation between the red and blue dots in the plots reveals the semantic similarity of the two forming subjects. + +![](images/3800d6710180dd23f74519b9113dfeda3ab809d5a409e47f3a5864df3163d88d.jpg) +Fig. 3: Cross-Attention Leakage. We demonstrate the emergence of semantic leakage at the cross-attention layers. We show two examples: a puppy a kitten, and a hamster and a squirrel. In the two leftmost columns, the subjects were generated separately using Stable Diffusion (SD). In the right three columns, we generate a single image with the two subjects using three different methods: Stable Diffusion (SD), Layout Guidance (LG), and Bounded Attention (BA, ours). Under each row, we plot the two first principal components of the cross-attention queries. As can be seen, the separation of the queries (blue and red) reflects the leakage between the subjects in the generated images. + +![](images/5ed8d39cfdcbe7b23060e6178933c2a07a3efbcb6f493db72ce21b5c183ce560.jpg) + +Clearly, vanilla Stable Diffusion struggles to adequately generate the two subjects within the same image. This is apparent by the visual leakage between the two subjects, as the model cannot avoid averaging their distinct visual properties. For example, the puppy has the visual features of a kitten, like raised ears and a triangular nose, while the squirrel loses its distinct ear shape and its feet are now pink like the hamster. Respectively, it can be seen that the queries of the two subjects are mixed, even for the kitten and the puppy which are more separated in the reference plot. + +Meanwhile, Layout Guidance (LG), which optimizes $\mathbf{z}_t$ to have attention peaks for each noun token at their corresponding region, exhibits interesting results. Its optimization objective implicitly encourages the separation between subjects' cross-attention queries. This can have positive effects, like the hamster and squirrel having unique colors, but at the same time yields unwanted artifacts, like the puppy losing its face altogether. Moreover, it can inadvertently push the latent signal out of distribution, causing quality degradation, as evident by the hamster's and squirrel's cartoonish texture. When it overly pushes the latent out of distribution, it leads to the catastrophic neglect phenomenon (Figure 2). + +In comparison, when examining the plots of our method alongside the reference plots, our approach preserves the feature distribution of the subjects' queries, and successfully generates the two subjects, even when the queries are as mixed as in the hamster and the squirrel. + +The above analysis yields two immediate conclusions: (i) Semantic similarity between subjects is reflected by their queries proximity, and leads to mixed queries when the subjects are generated together. This in turn leads to leakage between the subjects in the generated images, and (ii) enforcing semantic separation by modifying the semantic meaning of the cross-attention queries is harmful. The former observation represents a crucial architectural limitation in + +![](images/e6b310983195244e63448ea3138f357dd58613f5ee290d0cc5ad227fe6172c1d.jpg) +SD +Fig. 4: Self-Attention Leakage. We demonstrate the emergence of semantic leakage at the self-attention maps of two generally semantically dissimilar subjects: a crab and a frog. The images are generated by Stable Diffusion (SD) and Layout-guidance (LG). We highlight specific pixels, such as those of a subject's eye or leg, and present their respective self-attention maps. + +current diffusion models, and the latter pinpoints to a previously unexplored weakness in the widely used latent optimization methodology. Bounded Attention is designed to overcome these limitations. + +# 4.3 Self-Attention Leakage + +We now turn to analyze the leakage caused by self-attention layers. It has been shown that self-attention features exhibit dense correspondences within the same subject $[\ref{e}]$ and across semantic similar ones $[\ref{e}]$ . Hence, they are suspected as another source of leakage, that we shall study next. + +Here, we choose to examine the self-attention maps as a means to understand the leakage. In Figure 4 we focus on representative pixels (marked in yellow) associated with the subjects' eyes and legs, where visual leakage is most pronounced. As expected, features from one subject's eye or leg attend to the semantic similar body parts of the other. As a result, the features of each of the yellow points are directly affected by the features of the counterpart subject, causing leakage. In both images, the crab and the frog have similar appearances. In vanilla SD, both have a crab-like color and limbs with frog's toe pads. In LG, both have frog-like eyes and crab legs. + +Notably, this tendency to rely on similar patches aids the model in denoising the latent signal and is essential for achieving coherent images with properly blended subjects and backgrounds. Nevertheless, it has the drawback of leaking features between disjointed subjects. Therefore, completely disjointing the visual features during denoising $[?]$ , or naively pushing self-attention queries apart through optimization $[?]$ , can lead to subpar results. Consequently, we introduce the Bounded Attention mechanism to mitigate leakage and guide the latent signal towards subject separability, while avoiding detrimental artifacts. + +It's important to highlight that the leakage resulting from both cross- and self-attention layers is intertwined and mutually reinforced. Therefore, addressing the leakage caused by only one of these layers is insufficient to prevent the leakage in the generated images. + +# 4.4 Levels of Similarity + +In the two previous sections, our focus was primarily on subjects that share semantic similarity. Building upon the observation that the UNet's inner layers + +A kitten A puppy $64 \times 64$ $32 \times 32$ $16 \times 16$ + +![](images/7643c0359d0c3a9cb8ccfc6c73c19521535ea823ea1b8000c49bf05bb12663b1.jpg) +Fig. 5: We generate different subjects, and plot the first two principal components of the cross-attention queries at different layers of the UNet, where each layer is of different resolution. The high semantic similarity between the kitten and the puppy is expressed by the proximity of their queries through all layers. Meanwhile, the lizard and fruit share similar texture, and hence only their high-resolution queries are entangled. + +A lizard A fruit $64 \times 64$ $32 \times 32$ $16 \times 16$ + +![](images/a656c60d3d9a5eb3ea2cebf3978c9a9ba84d4486d3b92b505aadd150f86ddfac.jpg) + +dictate the subject's semantics and shape, while the outer layers control its style and appearance [?], we analyzed the inner UNet layers. However, leakage can also occur prominently when generating subjects that share visual similarity rather than semantic. We now turn to explore this scenario and demonstrate that, in such cases, the leakage originates from the UNet's outer layers. + +In Figure 5, we visualize cross-attention queries at different decoder's layers, when generating the kitten, puppy, lizard, and fruit in isolation. As can be seen, the queries of the kitten and the puppy are mixed across all UNet's layers, aligning with the visual and semantic similarity between these animals. On the other hand, the queries of the lizard and the fruit are overlapped only in the highest-resolution layer, aligning with the lack of semantic similarity between them. A closer look reveals that the lizard and the fruit share surprisingly similar textures, which explains the overlap of the queries in the highest-resolution layer. As explained in the previous sections, this overlap causes leakage between the two subjects, in this case, a visual rather than a semantic leakage (see Figure 2). + +# 5 Bounded Attention + +Our method takes as input $n$ distinct textual subjects $S = \{s_i\}_{i=1}^n$ contained within a global prompt $y$ , along with their corresponding bounding boxes $B = \{b_i\}_{i=1}^n$ . Our objective is to condition the generation process on $y$ , $S$ , and $B$ , while preserving the intended semantics of each subject, all without requiring any training or fine-tuning. + +Figure 6 illustrates an overview of our method. There, the input prompt $y$ is "A kitten and a puppy", $S = \{kitten\}, \{puppy\}$ , and the two corresponding bounding boxes $\{b_1, b_2\}$ are illustrated in the top left corner. Bounded Attention operates in two modes: Bounded Guidance and Bounded Denoising. Specifically, at the beginning of the denoising process, for $t \in [T, T_{\text{guidance}}]$ , we perform a Bounded Guidance step followed by a Bounded Denoising step. In the guidance step, we utilize the Bounded Guidance loss. This interval of timesteps constitutes the optimization phase. Then, for $t \in [T_{\text{guidance}}, 0]$ , we apply only Bounded Denoising steps. In both modes, we manipulate the model's forward pass by adopting an augmented weighting scheme in the attention layers, that safeguard the flow of information between the queries and the keys: + +![](images/37551b515abbb0158346d77723a9ad88db6cb9ae1223715f381023d2a7bad39d.jpg) +Fig. 6: Bounded Attention operates in two modes: guidance and denoising, each imposing strict constraints to bound the attention of each subject solely to itself and possibly to the background, preventing any influence from other subjects' features. In guidance mode, a loss is minimized to encourage each subject's attention to concentrate within its corresponding bounding box, with other bounding boxes masked to prevent artifacts. During the denoising step, we confine the attention of each subject solely to its bounding box, along with the background in the self-attention. This strategy effectively prevents feature leakage while maintaining the natural immersion of the subject within the image. Attention maps of a specific key, marked with $\star$ , are shown for each mode: the cross-attention map displays the key corresponding to the "kitten" token, and the self-attention map shows a key within the kitten's target bounding box. + +$$ +\mathbf {A} _ {t} ^ {(l)} = \mathrm {s o f t m a x} \left(\mathbf {Q} _ {t} ^ {(l)} \mathbf {K} _ {t} ^ {(l) ^ {\intercal}} + \mathbf {M} _ {t}\right), \tag {2} +$$ + +where $l$ represents the layer index, $t$ represents the diffusion timestep, and $\mathbf{Q}_t^{(l)},\mathbf{K}_t^{(l)}$ are the queries and keys of the $l$ -th attention layer. $\mathbf{M}_t$ represents time-specific masks composed of zeros and $-\infty$ elements. We refer to $\mathbf{A}_t^{(l)}$ as the Bounded Attention map. When indexing $\mathbf{A}_t^{(l)}$ , we use pixel coordinates $\mathbf{x}$ for rows, and attention-type-specific context vectors $\mathbf{c}$ for columns. In locations $[\mathbf{x},\mathbf{c}]$ , where $\mathbf{M}_t[\mathbf{x},\mathbf{c}] = -\infty$ , it holds that $\mathbf{A}_t^{(l)}[\mathbf{x},\mathbf{c}] = 0$ . Therefore, these masks prevent harmful information flow between pixels in self-attention layers, and between pixels and token embeddings in cross-attention layers. + +# 5.1 Bounded Guidance + +In Bounded Guidance, we backpropagate through the diffusion model to steer the latent signal toward the desired layout, using Gradient Descent. Our Bounded Guidance loss encourages the Bounded Attention map of each key corresponding to subject $s_i$ , to be within the $b_i$ bounding box. To this end, for each subject key we consider the ratio between the attention within the corresponding bounding box, to the entire Bounded Attention map (see Figure 6). Formally, we aggregate + +the following loss on the different subjects: + +$$ +\mathcal {L} _ {i} = 1 - \frac {\sum_ {\mathbf {x} \in b _ {i} , \mathbf {c} \in C _ {i}} \hat {\mathbf {A}} [ \mathbf {x} , \mathbf {c} ]}{\sum_ {\mathbf {x} \in b _ {i} , \mathbf {c} \in C _ {i}} \hat {\mathbf {A}} [ \mathbf {x} , \mathbf {c} ] + \alpha \sum_ {\mathbf {x} \notin b _ {i} , \mathbf {c} \in C _ {i}} \hat {\mathbf {A}} [ \mathbf {x} , \mathbf {c} ]}, \tag {3} +$$ + +where $i$ denotes the index of subject $s_i$ , $\hat{\mathbf{A}}$ is the mean Bounded Attention map, averaged across heads and layers, and $\alpha$ is a hyperparameter that magnifies the significance of disregarding attention towards the background, as we explain later. Similarly to the above, When indexing $\hat{\mathbf{A}}$ , pixel coordinates $\mathbf{x}$ represent rows, and attention-type-specific context vectors $\mathbf{c}$ represent columns. We designate $C_i$ as the set of all $s_i$ -related context vectors, i.e., pixel coordinates in $b_i$ for self-attention layers, and the token identifiers of $s_i$ in cross-attention layers. Additionally, for cross-attention layers, we include the first padding token [EoT] in $C_i$ to enhance layout alignment [?]. + +For each subject $s_i$ , the mask $\mathbf{M}_t$ should block the influence of opposing keys (tokens in $s_j$ and pixels in $b_j$ for $j \neq i$ ), to avoid the artifacts illustrated in Figure 3. This fosters that the queries of different subjects, including similar ones, are not erroneously forced to be far apart. Utilizing this loss, our Bounded Guidance step is defined as $z_t^{\mathrm{opt}} = z_t - \beta \nabla_{z_t} \sum_i \mathcal{L}_i^2$ . + +Integrating this loss within the cross-attention layers encourages the localization of each subject's semantics within its bounding boxes [?]. However, as cross-attention responses tend to peak around more typical patches associated with the subject's semantics (e.g., the face of a human, the legs of a crab, etc.), it may lack control over the subject's boundaries. By applying the loss within the self-attention layers, we encourage each subject to establish its own boundaries close to its bounding box, thereby discouraging subject fusion (see Figure 2). + +In the computation of the loss, we also introduce a hyperparameter $\alpha$ to reinforce attention to the background. This adjustment aids in preventing subject amalgamation, where a redundant subject is realized from different subject semantics in the background. To preserve image quality, we limit the application of this mode to an initial time interval $[T, T_{\text{guidance}}]$ , following similar works [?,?]. + +# 5.2 Bounded Denoising + +In Bounded Denoising, we compute the model's output and use it as the next latent in the series. Here, the masks aim to reduce semantic leakage between subjects, as detailed in Section 4, and to prevent unintended semantics from leaking to the background. Unlike Bounded Guidance and typical attention-based guidance approaches, Bounded Denoising can be applied throughout all time steps to mitigate leaks in fine details, which emerge only in later stages [?]. + +However, coarse masking in later stages may degrade image quality and result in noticeable stitching. To address this, after the optimization phase, for $t \in [T_{\text{guidance}}, 0]$ , we replace each bounding box with a fine segmentation mask obtained by clustering the self-attention maps [?]. Since the subject outlines are roughly determined in the initial time steps and evolve gradually thereafter [?], we refine the masks periodically. + +Notably, this mechanism also addresses imperfect alignments between subjects and bounding boxes after the guidance phase, which are more common + +![](images/7f7a4c772b48c2000383a0f93bdcc2553bd7ce60eac9bb384abbb9c2b92ca02e.jpg) +Fig. 7: Our results with SDXL. See the supplemental for more examples. + +when generating numerous subjects. Thus, employing this technique enhances the robustness of our method to seed selection, ensuring proper semantics even when subjects extend beyond their initial confines (see Figure 1,7). Compared to methods that necessitate strict input masks yet remain susceptible to leakage $[?,?,?]$ , our method offers greater user control with simpler inputs and more satisfactory outcomes. + +Method Details. Further details on the adaptation of Bounded Attention to the cross- and self-attention layers, along with a description of the subject mask refinement process, are provided in the supplementary materials. + +# 6 Experiments + +In this section, we conduct both qualitative and quantitative experiments to assess the effectiveness of our Bounded Attention method. We compare our approach with three training-free baseline methods: Layout-guidance (LG) [?], BoxDiff (BD) [?], and MultiDiffusion (MD) [?]. Additionally, we include comparisons with GLIGEN [?] and ReCo [?], which necessitate training. Since Attention-refocusing (AR) [?] is based on GLIGEN, we categorize it as a trained method for the purpose of our evaluation. For fairness, when comparing our method with other methods, we use Stable Diffusion. + +# 6.1 Qualitative Results + +SDXL results. We begin our experiments by demonstrating the efficacy of our method in challenging scenarios, particularly when tasked with generating multiple semantically similar subjects using SDXL. As can be seen in Figure 7, Vanilla SDXL fails to follow the prompt due to semantic leakage. It generates inaccurate number of dogs and mixes the features of each breed. In comparison, our approach adeptly generates each dog with its unique characteristics. + +"A gray kitten and a ginger kitten and black puppy in a yard." + +![](images/7632eac4e548eea099c73f413c5ea32ce25f2d217f7d841b46de947426cd0c99.jpg) +Fig. 8: Comparison of the first six images generated from the seed 0. + +Non-curated results. Next, we conduct a non-curated comparison with the training-free baseline methods and present the results in Figure 8. We showcase the initial six images sampled from seed 0 for each method. We anticipate that in the absence of Bounded Attention, semantic leakage may freely blend subject features, hindering the intended layout's formation. + +It is evident from the results that none of the competing methods is able to consistently construct the input layout. Layout Guidance [?] frequently neglects one of the subjects, and even when it generates three subjects, it struggles to avoid leakage, resulting in puppies with kitten-like features or incorrect color assignments. BoxDiff [?] often generates the correct number of subjects but suffers from artifacts in the form of blobs. Similar to Layout Guidance, it encounters difficulties in properly constructing the puppy. Surprisingly, even MultiDiffusion [?], which generates the subjects separately, faces challenges in generating them all, with some disappearing or merging together in its bootstrapping phase. + +In contrast, our method consistently outperforms these approaches, producing three subjects that align with the both prompt and layout in all six images. + +Comparisons with baselines. We present a qualitative comparison in Figure 9. All competing methods, including those trained specifically for the layout-to-image task, exhibit significant visual and semantic leakage. The training-free methods perform the worst: MultiDiffusion produces disharmonious, low-quality images, while optimization-based methods often result in object fusion, combining different semantics without adhering to the layout. + +The training-based approaches closely follow the layout but fail to convey the correct semantics. In the first row, they neglect the corduroy jacket, leaking the denim texture into the other jacket, or even fusing them together. In the second row, the elephant's features leak into the rhino. + +In comparison, our method generates images that align with the input layout and prompt, ensuring each subject retains its unique attributes, semantics, and appearance. + +"A denim jacket and a corduroy jacket and a leather handbag in a closet." + +![](images/1bad5bcf3921fd21d36b4c9872dd3e87124038ceb0f0426b0dd2ea796b2a46c1.jpg) + +"A big red elephant and a far away rhino in a jungle." + +![](images/9a947536fbd0f4c66b6701a2b8d231a27646ef426f02a6f3c4daa2af170fc0a4.jpg) +Fig. 9: Qualitative comparison of our method with baseline methods: Above each row, we display the input prompt, where each subject's color matches the color of its corresponding bounding box. We compare with both training-free methods (2nd-4th columns) and trained models (5th-7th columns). See supplemental for more results. + +# 6.2 Quantitative Results + +We evaluate our method's effectiveness using the Draw-Bench dataset [?], known for its challenging prompts designed to test a model's ability to compose multiple subjects with specific quantities and relations. We use the evaluation procedure from previous work [?, ?]. Our re + +
MethodCountingSpatial
PrecisionRecallF1Accuracy
Stable Diffusion [?]0.740.780.730.19
Layout-guidance [?]0.720.780.720.35
BoxDiff [?]0.810.780.760.28
MultiDiffusion [?]0.700.550.570.15
Ours0.810.910.820.43
+ +Table 1: Evaluation on the DrawBench dataset. + +sults, alongside those of other training-free methods, are summarized in Table 1. Unlike other approaches that do not account for semantic leakage, our method demonstrates notable improvements in both the counting and the spatial categories. Notably, other methods struggle to surpass the counting recall rates of vanilla SD, highlighting the effectiveness of Bounded Attention in generating multiple subjects. + +We further conduct a quantitative comparison of semantic leakage in training-free methods through a user study. Further details are in the supplemental. + +# 6.3 Ablation Studies + +To assess the significance of each component, we conduct an ablation study where we systematically vary our method's configurations by omitting one component in each setting. We show the results in Figure 10. + +Guidance is crucial for aligning the latent signal with the intended layout. Neglecting it can lead to subject generation in the background, as demonstrated + +"A red lizard and a turtle on the grass." + +![](images/591aeb0f9a2447109048e47333aeb97c7bbed21b31e4210a15ef7b9cd1d89c15.jpg) +w/o G + +![](images/19a6b8330b6947168d23ccc642327b532875c79758a434328cf599019c195e50.jpg) +w/o BG + +![](images/3f73670f251fa60fce3425654f92c11a874c4c1233551299f1d7d337be678dea.jpg) +w/o BD +Fig. 10: Qualitative ablation. We ablate our method by skipping the guidance step (G), performing guidance without Bounded Guidance (BG), not applying Bounded Denoising (BD), and not performing Mask Refinement (MR). We show zoomed-in images of the two rightmost configurations. More examples can be found in the supplemental. + +![](images/25bd0bd466b74bc1f55d42a895f017273a814097b8e9e2bb87f36241a7caf1bc.jpg) +w/o MR + +![](images/446db0dd9c92ae76d87114eb4082cde8f14cdddaf98c919d3a9e7bc50b4c1148.jpg) + +![](images/96ca2fd3d8a4586ad16dd6d1bead476365fa8c728a49d5a5d5d4b7d4a9efb473.jpg) +Full Method + +![](images/596f51289e632be8f3f0a3dd6581aa9c978a1d5b730263be74db7b0effb50674.jpg) + +by the emergence of a turtle-lizard hybrid. However, attempting to guide the latent signal without our Bounded Guidance mechanism may cause distortions in subject appearances, such as the turtle's neck growing out of its leg and inconsistent artifacts appearing on the lizard's head and back. + +Meanwhile, forgoing Bounded Denoising results in noticeable semantic leakage, with the lizard being replaced by a turtle, and the "red" attribute erroneously leaking to the wrong subject. + +Lastly, incorporating mask refinement in the later stages prevents fine-details from leaking. Without mask refinement, the lizard exhibits shell-like contours on its back. + +# 7 Conclusions + +We introduce Bounded Attention, a technique designed to regulate the accurate generation of multiple subjects within an image. This approach encourages each subject to "be yourself", emphasizing the importance of preserving individuality and uniqueness without being excessively influenced by other subjects present in the image. Our development of the Bounded Attention technique stemmed from an in-depth analysis of the root causes behind the misalignment observed between the provided prompt and the resulting generated image. Our investigation revealed that this misalignment primarily arises due to semantic leakage among the generated subjects, a phenomenon observed in both the cross and self-attention layers. + +While Bounded Attention effectively mitigates a significant portion of semantic leakage, it does not entirely eliminate it. Our findings demonstrate a marked improvement in performance compared to other methods that seek to achieve semantic alignment. However, residual leakage persists, which we attribute to imperfect optimization during the guidance mode and inaccurate segmentation of the subject prior to the second phase. + +While Bounded Attention excels in generating multiple subjects with plausible semantic alignment, its performance may vary across different layouts. Achieving success with Bounded Attention hinges on a strong match between the seed and the layout. Moving forward, we aim to explore methods for generating well-suited seeds tailored to specific layouts. One potential avenue is to introduce noise to the layout image, thereby creating a seed that aligns more closely with the desired outcomes. + +# Acknowledgement + +We thank Guy Tevet for fruitful discussions and useful suggestions. This work is partly supported by a research gift from Snap Inc. + +# References \ No newline at end of file diff --git a/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/images.zip b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ae8276cbe4bc03e9f9c4593ce912f0bd019995d8 --- /dev/null +++ b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e034d86b4868edf8662bbc6c41ffea8dc5b2243e0f88c4d3f899e0c57b780b28 +size 542408 diff --git a/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/layout.json b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..948c7d8ba061de5eff3e2823977f672123048276 --- /dev/null +++ b/2024/Be Yourself_ Bounded Attention for Multi-Subject Text-to-Image Generation/layout.json @@ -0,0 +1,8130 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 112, + 449, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 112, + 449, + 148 + ], + "spans": [ + { + "bbox": [ + 165, + 112, + 449, + 148 + ], + "type": "text", + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 180 + ], + "type": "text", + "content": "Omer Dahary1, Or Patashnik1,2, Kfir Aberman2, and Daniel Cohen-Or1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 261, + 189, + 352, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 189, + 352, + 200 + ], + "spans": [ + { + "bbox": [ + 261, + 189, + 352, + 200 + ], + "type": "text", + "content": "1 Tel Aviv University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 272, + 200, + 341, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 200, + 341, + 212 + ], + "spans": [ + { + "bbox": [ + 272, + 200, + 341, + 212 + ], + "type": "text", + "content": "2 Snap Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 222, + 230, + 392, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 230, + 392, + 243 + ], + "spans": [ + { + "bbox": [ + 222, + 230, + 392, + 243 + ], + "type": "text", + "content": "3 ginger kittens and 2 gray kittens..." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 244, + 217, + 329 + ], + "blocks": [ + { + "bbox": [ + 133, + 244, + 217, + 329 + ], + "lines": [ + { + "bbox": [ + 133, + 244, + 217, + 329 + ], + "spans": [ + { + "bbox": [ + 133, + 244, + 217, + 329 + ], + "type": "image", + "image_path": "458d9d6bc89265b09d53cdac1017211b3ad601bd8f74190ec9496699c6819862.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 330, + 208, + 339 + ], + "lines": [ + { + "bbox": [ + 143, + 330, + 208, + 339 + ], + "spans": [ + { + "bbox": [ + 143, + 330, + 208, + 339 + ], + "type": "text", + "content": "\"... on the stairs\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 221, + 245, + 305, + 328 + ], + "blocks": [ + { + "bbox": [ + 221, + 245, + 305, + 328 + ], + "lines": [ + { + "bbox": [ + 221, + 245, + 305, + 328 + ], + "spans": [ + { + "bbox": [ + 221, + 245, + 305, + 328 + ], + "type": "image", + "image_path": "9332232f51f07a12917bb8a8cedba40642fa194b4a49e63bd7781902d1baaf00.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 330, + 296, + 338 + ], + "lines": [ + { + "bbox": [ + 230, + 330, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 230, + 330, + 296, + 338 + ], + "type": "text", + "content": "\"...on the street\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 245, + 392, + 328 + ], + "blocks": [ + { + "bbox": [ + 309, + 245, + 392, + 328 + ], + "lines": [ + { + "bbox": [ + 309, + 245, + 392, + 328 + ], + "spans": [ + { + "bbox": [ + 309, + 245, + 392, + 328 + ], + "type": "image", + "image_path": "0f1ca0dbef871026421dd1dabc665c1ee1879bf09633ac24e7c14f0d84867441.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 330, + 380, + 338 + ], + "lines": [ + { + "bbox": [ + 322, + 330, + 380, + 338 + ], + "spans": [ + { + "bbox": [ + 322, + 330, + 380, + 338 + ], + "type": "text", + "content": "\"...in a basket\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 351, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 132, + 351, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 482, + 407 + ], + "type": "text", + "content": "Fig. 1: Our method bounds the attention to enable layout control over a pre-trained text-to-image diffusion model. Bounded Attention effectively reduces the impact of the innate semantic leakage during denoising, encouraging each subject to be itself. Our method can faithfully generate challenging layouts featuring multiple similar subjects with different modifiers (e.g., ginger and gray kittens)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 397, + 245, + 480, + 328 + ], + "blocks": [ + { + "bbox": [ + 397, + 245, + 480, + 328 + ], + "lines": [ + { + "bbox": [ + 397, + 245, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 397, + 245, + 480, + 328 + ], + "type": "image", + "image_path": "bc74ce6b4fc1619b6e9df50dcd7d20364af2f282ec035ac5477471e887318c45.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 330, + 463, + 338 + ], + "lines": [ + { + "bbox": [ + 414, + 330, + 463, + 338 + ], + "spans": [ + { + "bbox": [ + 414, + 330, + 463, + 338 + ], + "type": "text", + "content": "\"... on a tree\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 159, + 424, + 453, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 424, + 453, + 643 + ], + "spans": [ + { + "bbox": [ + 159, + 424, + 453, + 643 + ], + "type": "text", + "content": "Abstract. Text-to-image diffusion models have an unprecedented ability to generate diverse and high-quality images. However, they often struggle to faithfully capture the intended semantics of complex input prompts that include multiple subjects. Recently, numerous layout-to-image extensions have been introduced to improve user control, aiming to localize subjects represented by specific tokens. Yet, these methods often produce semantically inaccurate images, especially when dealing with multiple semantically or visually similar subjects. In this work, we study and analyze the causes of these limitations. Our exploration reveals that the primary issue stems from inadvertent semantic leakage between subjects in the denoising process. This leakage is attributed to the diffusion model's attention layers, which tend to blend the visual features of different subjects. To address these issues, we introduce Bounded Attention, a training-free method for bounding the information flow in the sampling process. Bounded Attention prevents detrimental leakage among subjects and enables guiding the generation to promote each subject's individuality, even with complex multi-subject conditioning. Through extensive experimentation, we demonstrate that our method empowers the generation of multiple subjects that better align with given prompts and layouts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 160, + 654, + 453, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 654, + 453, + 665 + ], + "spans": [ + { + "bbox": [ + 160, + 654, + 453, + 665 + ], + "type": "text", + "content": "Keywords: Image Generation " + }, + { + "bbox": [ + 160, + 654, + 453, + 665 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 654, + 453, + 665 + ], + "type": "text", + "content": " Diffusion Models " + }, + { + "bbox": [ + 160, + 654, + 453, + 665 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 654, + 453, + 665 + ], + "type": "text", + "content": " Semantic Alignment" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 223 + ], + "type": "text", + "content": "In recent years, text-to-image generation has undergone a significant shift with the integration of conditional diffusion models " + }, + { + "bbox": [ + 130, + 139, + 481, + 223 + ], + "type": "inline_equation", + "content": "[?,?,?,?]" + }, + { + "bbox": [ + 130, + 139, + 481, + 223 + ], + "type": "text", + "content": ", allowing for the facile generation of high-quality and diverse images. The use of attention layers in architectures of such generative models has been identified as a major factor contributing to improved quality of generated images " + }, + { + "bbox": [ + 130, + 139, + 481, + 223 + ], + "type": "inline_equation", + "content": "[?,?]" + }, + { + "bbox": [ + 130, + 139, + 481, + 223 + ], + "type": "text", + "content": ". However, these models struggle to accurately generate scenes containing multiple subjects, especially when they are semantically or visually similar." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 306 + ], + "type": "text", + "content": "In this work, we study the problem of multi-subject image generation in attention-based diffusion models. Our contribution is twofold. First, we recognize the underlying reasons for the difficulty in generating images containing multiple subjects, especially those sharing semantic similarities. Second, building on our insights, we present a method aimed at mitigating semantic leakage in the generated images, allowing control over the generation of multiple subjects (see Figure 1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 481, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 481, + 378 + ], + "type": "text", + "content": "We demonstrate an innate bias within the common attention-based architectures utilized in diffusion models, which predisposes them to leak visual features between subjects. In particular, the functionality of attention layers is designed to blend features across the image. Therefore, they inherently lead to information leakage between subjects. This phenomenon is particularly noticeable when subjects are semantically similar and, therefore, attend to each other (Figure 2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 481, + 461 + ], + "type": "text", + "content": "A plethora of works tries to mitigate the cross-subject leakage issue, either by modifying the sampling process to better follow different subjects in the prompt " + }, + { + "bbox": [ + 130, + 379, + 481, + 461 + ], + "type": "inline_equation", + "content": "[\\?,?]" + }, + { + "bbox": [ + 130, + 379, + 481, + 461 + ], + "type": "text", + "content": ", or by coupling the global prompt with layout information via segmentation maps or bounding boxes labeled with subject classes or local prompts " + }, + { + "bbox": [ + 130, + 379, + 481, + 461 + ], + "type": "inline_equation", + "content": "[\\?,?,?,?]" + }, + { + "bbox": [ + 130, + 379, + 481, + 461 + ], + "type": "text", + "content": ". However, the majority of these methods still encounter difficulties in accurately aligning to input layouts, particularly in scenarios involving two or more semantically similar subjects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 581 + ], + "type": "text", + "content": "In our approach, we guide the image generation with a spatial layout " + }, + { + "bbox": [ + 130, + 462, + 481, + 581 + ], + "type": "inline_equation", + "content": "[?,?]" + }, + { + "bbox": [ + 130, + 462, + 481, + 581 + ], + "type": "text", + "content": ". To address cross-subject leakage, we introduce the Bounded Attention mechanism, utilized during the denoising process to generate an image. This method bounds the influence of irrelevant visual and textual tokens on each pixel, which otherwise promotes leakage. By applying this mechanism, we encourage each subject to be itself, in the sense that it hinders the borrowing of features from other subjects in the scene. We show that bounding the attention is needed both in the cross- and self-attention layers. Moreover, we find additional architectural components that amplify leakage, modify their operation, and present remedies to them." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "We show that our method succeeds in facilitating control over multiple subjects within complex, coarse-grained layouts comprising numerous bounding boxes with similar semantics. Particularly challenging examples are demonstrated in Figure 1, where we successfully generate five kittens with a mix of adjectives. We conduct experiments on both Stable Diffusion [?] and SDXL [?] architectures and demonstrate the advantage of our method compared to previous ones, both supervised and unsupervised." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "text", + "content": "O. Dahary et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 323, + 127, + 421, + 221 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 309, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 309, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 309, + 224 + ], + "type": "text", + "content": "Fig.2: Misalignment in layout-to-image generation include (i) catastrophic neglect [?] where the model fails to include one or more subjects mentioned in the prompt within the generated image, (ii) incorrect attribute binding [?,?] where attributes are not correctly matched to their corresponding subjects, and (iii) subject fusion [?] where the model merges multiple subjects into a single, larger subject." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 127, + 421, + 221 + ], + "lines": [ + { + "bbox": [ + 323, + 127, + 421, + 221 + ], + "spans": [ + { + "bbox": [ + 323, + 127, + 421, + 221 + ], + "type": "image", + "image_path": "14e4ea5dbb99eed94036ac648061a206307e46c5eadf11efc71ba276ca218eb5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 422, + 126, + 474, + 219 + ], + "blocks": [ + { + "bbox": [ + 422, + 126, + 474, + 219 + ], + "lines": [ + { + "bbox": [ + 422, + 126, + 474, + 219 + ], + "spans": [ + { + "bbox": [ + 422, + 126, + 474, + 219 + ], + "type": "image", + "image_path": "ab0f05d519ab46ce516e55615ffdd6925ac0cb61fa077cfe3cad7a62060939a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 255, + 234, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 234, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 234, + 268 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "type": "text", + "content": "Text-to-image diffusion models. Diffusion models, trained on huge datasets [?], have demonstrated their power in learning the complex distributions of diverse natural images [?,?,?,?,?] Augmenting attention layers into diffusion models and conditioning them on textual prompts by coupling them with visually-aligned text encoders [?] leads to powerful text-to-image models [?] In this work, we specifically examine two such open-source text-to-image diffusion models: Stable Diffusion [?], and the more recent SDXL [?]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 371, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 467 + ], + "type": "text", + "content": "Semantic alignment in text-to-image synthesis. A critical drawback of current text-to-image models pertains to their limited ability to faithfully represent the precise semantics of input prompts. Various studies have identified common semantic failures and proposed mitigating strategies, such as adjusting text embeddings " + }, + { + "bbox": [ + 130, + 371, + 482, + 467 + ], + "type": "inline_equation", + "content": "[?,?,?]" + }, + { + "bbox": [ + 130, + 371, + 482, + 467 + ], + "type": "text", + "content": ", or optimizing noisy signals to strengthen or align cross-attention maps " + }, + { + "bbox": [ + 130, + 371, + 482, + 467 + ], + "type": "inline_equation", + "content": "[?,?]" + }, + { + "bbox": [ + 130, + 371, + 482, + 467 + ], + "type": "text", + "content": ". Nevertheless, these methods often fall short in generating multiple subjects, and do not adhere to positional semantics, such as subjects' number or location." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "text", + "content": "Layout-guided image synthesis. Addressing the semantic alignment concern, alternative approaches advocate for conditioning the diffusion process on layout information, either by training new models from scratch " + }, + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "inline_equation", + "content": "[?,?]" + }, + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "text", + "content": " or fine-tuning an existing one " + }, + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "inline_equation", + "content": "[?,?,?,?]" + }, + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "text", + "content": ". Despite their promise, these methods demand extensive computational resources and prolonged training times. Moreover, they are constrained by the layout distribution of the training data and the models' architectural bias to blend subject features, a limitation that our Bounded Attention aims to overcome." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "To circumvent these challenges, numerous researchers explore training-free techniques, where the generation process itself is modified to enforce layout constraints. Several optimization-based works employ techniques similar to classifier-free guidance to localize the cross-attention " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "[\\?,?,?,?]" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and/or self-attention maps " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "[\\?]" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". While effective in aligning random noise with the intended layout, guiding attention maps to distinct regions may inadvertently lead to undesired behavior, particularly when different subjects share similar semantics and visual features. Furthermore, these methods often exhibit a deteriorating effect on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "visual quality, thereby limiting their applicability to only the initial denoising steps and neglecting finer control over shape and visual details, which are determined only in later stages [?]. Bounded Attention addresses these shortcomings by regulating attention computation throughout the entire denoising process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 480, + 283 + ], + "type": "text", + "content": "Another approach involves generating each subject separately in its own denoising process " + }, + { + "bbox": [ + 130, + 165, + 480, + 283 + ], + "type": "inline_equation", + "content": "[\\?,?]" + }, + { + "bbox": [ + 130, + 165, + 480, + 283 + ], + "type": "text", + "content": ". While these methods inherently address catastrophic neglect, they tend to generate disharmonious images, and remain susceptible to leakage when merging subjects in subsequent stages. In contrast, masking attention maps to input bounding boxes " + }, + { + "bbox": [ + 130, + 165, + 480, + 283 + ], + "type": "inline_equation", + "content": "[\\?]" + }, + { + "bbox": [ + 130, + 165, + 480, + 283 + ], + "type": "text", + "content": " or attenuating attention in specific segments " + }, + { + "bbox": [ + 130, + 165, + 480, + 283 + ], + "type": "inline_equation", + "content": "[\\?]" + }, + { + "bbox": [ + 130, + 165, + 480, + 283 + ], + "type": "text", + "content": " represents a milder variant of this strategy, aimed at avoiding visible stitching. However, these methods often fall short of fully mitigating subject leakage, compromising semantic alignment. In comparison, Bounded Attention is able to carefully govern information propagation among subjects in a single denoising process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 481, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 481, + 369 + ], + "type": "text", + "content": "While both the trained models and training-free techniques aim to generate numerous objects, they do not mitigate the inherent leakage caused by attention mechanisms. Unlike our Bounded Attention technique, these methods encounter challenges in effectively generating and controlling a multitude of subjects, especially when they share semantic similarity. Notably, existing techniques struggle to accurately generate even two semantically similar subjects, whereas our method, as demonstrated succeeds in generating five and even more subjects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 387, + 234, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 234, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 234, + 399 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": "Latent diffusion models. In this work, we examine Stable Diffusion [?] and SDXL [?], which are both publicly available latent diffusion models. These models operate in the latent space of a pretrained image autoencoder, and are thus tasked with denoising a latent representation of the image, where each latent pixel corresponds to a patch in the generated image. Starting from pure random noise " + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T" + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": ", at each timestep " + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": ", the current noisy latent " + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": " is passed to a denoising UNet " + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": ", trained to predict the current noise estimate " + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t,y,t)" + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": " using the guiding prompt encoding " + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 514, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 482, + 564 + ], + "type": "text", + "content": "Attention layers. At each block, the UNet utilizes residual convolution layers, producing intermediate features " + }, + { + "bbox": [ + 131, + 514, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\phi^{(l)}(\\mathbf{z}_t)" + }, + { + "bbox": [ + 131, + 514, + 482, + 564 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 514, + 482, + 564 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 514, + 482, + 564 + ], + "type": "text", + "content": " denotes the layer's index. These, in turn, are passed to attention layers, which essentially average different values " + }, + { + "bbox": [ + 131, + 514, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{t}}^{(1)}" + }, + { + "bbox": [ + 131, + 514, + 482, + 564 + ], + "type": "text", + "content": " according to pixel-specific weights:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 180, + 574, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 574, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 180, + 574, + 481, + 594 + ], + "type": "interline_equation", + "content": "\\phi^ {(l + 1)} (\\mathbf {z} _ {t}) = \\mathbf {A} _ {t} ^ {(l)} \\mathbf {V} _ {t} ^ {(l)}, \\text {w h e r e} \\mathbf {A} _ {t} ^ {(l)} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} _ {t} ^ {(l)} \\mathbf {K} _ {t} ^ {(l) ^ {\\intercal}}\\right). \\tag {1}", + "image_path": "ebd3e88b78d7e9c9f7353223e94959a9674cda82f9382e105dbbdf07ede184e2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "text", + "content": "Here, the keys " + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_t^{(l)} = f_K^{(l)}\\left(\\mathbf{C}_t^{(l)}\\right)" + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_t^{(l)} = f_V^{(l)}\\left(\\mathbf{C}_{\\mathbf{t}}^{(\\mathbf{1})}\\right)" + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "text", + "content": " are linear projections of context vectors " + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_t^{(l)}" + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "text", + "content": ". In the cross-attention layers we inject semantic context from the prompt encoding " + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_t^{(l)} \\equiv y" + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "text", + "content": ", while self-attention layers utilize global information from the latent itself " + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_t^{(l)} \\equiv \\phi^{(l)}(\\mathbf{z}_t)" + }, + { + "bbox": [ + 130, + 604, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "O. Dahary et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": "The weighting scheme is determined by the attention map " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t^{(l)}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " which represents the probability of each pixel to semantically associate with a key-value pair. This is done by linearly projecting the latent noisy pixels to queries " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t^{(l)} = f_Q^{(l)}\\left(\\phi^{(l)}(\\mathbf{z}_t)\\right)" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " and computing their inner product with the keys. It has been widely demonstrated that the cross-attention maps are highly indicative of the semantic association between the image layout and the prompt tokens [?]. Meanwhile, the self-attention maps govern the correspondence between pixels, and thus form the image's structure [?]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 225, + 261, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 225, + 261, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 261, + 239 + ], + "type": "text", + "content": "4 Semantic Leakage" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 481, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 481, + 268 + ], + "type": "text", + "content": "We begin by studying the causes of semantic leakage in Stable Diffusion [?], and examine the limitations of existing layout-to-image approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 275, + 272, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 272, + 287 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 272, + 287 + ], + "type": "text", + "content": "4.1 On Subject Similarity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 286, + 480, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 480, + 321 + ], + "type": "text", + "content": "Figure 2 illustrates various misalignment failures observed in state-of-the-art layout-to-image training-free methods. As we shall show, these failures are more prevalent for subjects that share semantic or visual similarity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 322, + 481, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 481, + 407 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 322, + 481, + 407 + ], + "type": "inline_equation", + "content": "x_{s_1}, x_{s_2} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 322, + 481, + 407 + ], + "type": "text", + "content": " be 2D latent coordinates corresponding to two semantically similar subjects " + }, + { + "bbox": [ + 130, + 322, + 481, + 407 + ], + "type": "inline_equation", + "content": "s_1, s_2" + }, + { + "bbox": [ + 130, + 322, + 481, + 407 + ], + "type": "text", + "content": " in the generated image. Intuitively, we expect that along the denoising process, the queries corresponding to these pixels, " + }, + { + "bbox": [ + 130, + 322, + 481, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t^{(l)}[x_{s_1}], \\mathbf{Q}_t^{(l)}[x_{s_2}]" + }, + { + "bbox": [ + 130, + 322, + 481, + 407 + ], + "type": "text", + "content": ", will be similar and hence also their attention responses. This, in turn, also implies that they will share semantic information from the token embeddings through the cross-attention layers or visual information via the self-attention layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 481, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 467 + ], + "type": "text", + "content": "To explore this hypothesis, we investigate the model's behavior when tasked with generating two subjects and analyze their attention features in both cross- and self-attention layers. Subsequently, we meticulously examine these features and demonstrate how their behavior sheds light on the leakage observed in generated images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 475, + 285, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 285, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 285, + 487 + ], + "type": "text", + "content": "4.2 Cross-Attention Leakage" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 487, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 605 + ], + "type": "text", + "content": "To analyze the leakage caused by cross-attention layers, we examine the cross-attention queries. We depict these queries in the plots in Figure 3, where each point corresponds to a single query projected to 2D with PCA. To label each point with its corresponding subject, we compute the subject masks by averaging cross-attention maps [?] and color each projected query according to the subject's text color. The leftmost plot, in which the two subjects were generated separately, serves as a reference point to the relation between the queries when there is no leakage. For comparative analysis, we also present results for Layout-guidance [?], as a simple representative of current training-free layout-guided approaches, and Bounded Attention, which we shall cover in the sequel." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We consider the cross-attention queries in two examples: \"a kitten\" and \"a puppy\", and \"a hamster\" and a \"squirrel\". As can be seen in the reference plots, the kitten and puppy queries share some small overlap, and the hamster and squirrel queries are mixed together. The level of separation between the red and blue dots in the plots reveals the semantic similarity of the two forming subjects." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 297, + 202 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 297, + 202 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 297, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 297, + 202 + ], + "type": "image", + "image_path": "3800d6710180dd23f74519b9113dfeda3ab809d5a409e47f3a5864df3163d88d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 312 + ], + "type": "text", + "content": "Fig. 3: Cross-Attention Leakage. We demonstrate the emergence of semantic leakage at the cross-attention layers. We show two examples: a puppy a kitten, and a hamster and a squirrel. In the two leftmost columns, the subjects were generated separately using Stable Diffusion (SD). In the right three columns, we generate a single image with the two subjects using three different methods: Stable Diffusion (SD), Layout Guidance (LG), and Bounded Attention (BA, ours). Under each row, we plot the two first principal components of the cross-attention queries. As can be seen, the separation of the queries (blue and red) reflects the leakage between the subjects in the generated images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 308, + 114, + 478, + 202 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 478, + 202 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 478, + 202 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 478, + 202 + ], + "type": "image", + "image_path": "5ed8d39cfdcbe7b23060e6178933c2a07a3efbcb6f493db72ce21b5c183ce560.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 338, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 433 + ], + "type": "text", + "content": "Clearly, vanilla Stable Diffusion struggles to adequately generate the two subjects within the same image. This is apparent by the visual leakage between the two subjects, as the model cannot avoid averaging their distinct visual properties. For example, the puppy has the visual features of a kitten, like raised ears and a triangular nose, while the squirrel loses its distinct ear shape and its feet are now pink like the hamster. Respectively, it can be seen that the queries of the two subjects are mixed, even for the kitten and the puppy which are more separated in the reference plot." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 543 + ], + "type": "text", + "content": "Meanwhile, Layout Guidance (LG), which optimizes " + }, + { + "bbox": [ + 130, + 434, + 482, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 434, + 482, + 543 + ], + "type": "text", + "content": " to have attention peaks for each noun token at their corresponding region, exhibits interesting results. Its optimization objective implicitly encourages the separation between subjects' cross-attention queries. This can have positive effects, like the hamster and squirrel having unique colors, but at the same time yields unwanted artifacts, like the puppy losing its face altogether. Moreover, it can inadvertently push the latent signal out of distribution, causing quality degradation, as evident by the hamster's and squirrel's cartoonish texture. When it overly pushes the latent out of distribution, it leads to the catastrophic neglect phenomenon (Figure 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 592 + ], + "type": "text", + "content": "In comparison, when examining the plots of our method alongside the reference plots, our approach preserves the feature distribution of the subjects' queries, and successfully generates the two subjects, even when the queries are as mixed as in the hamster and the squirrel." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "The above analysis yields two immediate conclusions: (i) Semantic similarity between subjects is reflected by their queries proximity, and leads to mixed queries when the subjects are generated together. This in turn leads to leakage between the subjects in the generated images, and (ii) enforcing semantic separation by modifying the semantic meaning of the cross-attention queries is harmful. The former observation represents a crucial architectural limitation in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "text", + "content": "O. Dahary et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 122, + 473, + 166 + ], + "blocks": [ + { + "bbox": [ + 212, + 114, + 225, + 121 + ], + "lines": [ + { + "bbox": [ + 212, + 114, + 225, + 121 + ], + "spans": [ + { + "bbox": [ + 212, + 114, + 225, + 121 + ], + "type": "text", + "content": "SD" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 122, + 473, + 166 + ], + "lines": [ + { + "bbox": [ + 133, + 122, + 473, + 166 + ], + "spans": [ + { + "bbox": [ + 133, + 122, + 473, + 166 + ], + "type": "image", + "image_path": "e6b310983195244e63448ea3138f357dd58613f5ee290d0cc5ad227fe6172c1d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 176, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 232 + ], + "type": "text", + "content": "Fig. 4: Self-Attention Leakage. We demonstrate the emergence of semantic leakage at the self-attention maps of two generally semantically dissimilar subjects: a crab and a frog. The images are generated by Stable Diffusion (SD) and Layout-guidance (LG). We highlight specific pixels, such as those of a subject's eye or leg, and present their respective self-attention maps." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 480, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 291 + ], + "type": "text", + "content": "current diffusion models, and the latter pinpoints to a previously unexplored weakness in the widely used latent optimization methodology. Bounded Attention is designed to overcome these limitations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 308, + 276, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 276, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 276, + 319 + ], + "type": "text", + "content": "4.3 Self-Attention Leakage" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 319, + 480, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 367 + ], + "type": "text", + "content": "We now turn to analyze the leakage caused by self-attention layers. It has been shown that self-attention features exhibit dense correspondences within the same subject " + }, + { + "bbox": [ + 130, + 319, + 480, + 367 + ], + "type": "inline_equation", + "content": "[\\ref{e}]" + }, + { + "bbox": [ + 130, + 319, + 480, + 367 + ], + "type": "text", + "content": " and across semantic similar ones " + }, + { + "bbox": [ + 130, + 319, + 480, + 367 + ], + "type": "inline_equation", + "content": "[\\ref{e}]" + }, + { + "bbox": [ + 130, + 319, + 480, + 367 + ], + "type": "text", + "content": ". Hence, they are suspected as another source of leakage, that we shall study next." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 368, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 481, + 475 + ], + "type": "text", + "content": "Here, we choose to examine the self-attention maps as a means to understand the leakage. In Figure 4 we focus on representative pixels (marked in yellow) associated with the subjects' eyes and legs, where visual leakage is most pronounced. As expected, features from one subject's eye or leg attend to the semantic similar body parts of the other. As a result, the features of each of the yellow points are directly affected by the features of the counterpart subject, causing leakage. In both images, the crab and the frog have similar appearances. In vanilla SD, both have a crab-like color and limbs with frog's toe pads. In LG, both have frog-like eyes and crab legs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 475, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 570 + ], + "type": "text", + "content": "Notably, this tendency to rely on similar patches aids the model in denoising the latent signal and is essential for achieving coherent images with properly blended subjects and backgrounds. Nevertheless, it has the drawback of leaking features between disjointed subjects. Therefore, completely disjointing the visual features during denoising " + }, + { + "bbox": [ + 130, + 475, + 481, + 570 + ], + "type": "inline_equation", + "content": "[?]" + }, + { + "bbox": [ + 130, + 475, + 481, + 570 + ], + "type": "text", + "content": ", or naively pushing self-attention queries apart through optimization " + }, + { + "bbox": [ + 130, + 475, + 481, + 570 + ], + "type": "inline_equation", + "content": "[?]" + }, + { + "bbox": [ + 130, + 475, + 481, + 570 + ], + "type": "text", + "content": ", can lead to subpar results. Consequently, we introduce the Bounded Attention mechanism to mitigate leakage and guide the latent signal towards subject separability, while avoiding detrimental artifacts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 571, + 481, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 481, + 619 + ], + "type": "text", + "content": "It's important to highlight that the leakage resulting from both cross- and self-attention layers is intertwined and mutually reinforced. Therefore, addressing the leakage caused by only one of these layers is insufficient to prevent the leakage in the generated images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 630, + 258, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 258, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 258, + 641 + ], + "type": "text", + "content": "4.4 Levels of Similarity" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In the two previous sections, our focus was primarily on subjects that share semantic similarity. Building upon the observation that the UNet's inner layers" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 102 + ], + "type": "text", + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 293, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 293, + 123 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 293, + 123 + ], + "type": "text", + "content": "A kitten A puppy " + }, + { + "bbox": [ + 133, + 114, + 293, + 123 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 133, + 114, + 293, + 123 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 133, + 114, + 293, + 123 + ], + "type": "inline_equation", + "content": "16 \\times 16" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 123, + 293, + 156 + ], + "blocks": [ + { + "bbox": [ + 133, + 123, + 293, + 156 + ], + "lines": [ + { + "bbox": [ + 133, + 123, + 293, + 156 + ], + "spans": [ + { + "bbox": [ + 133, + 123, + 293, + 156 + ], + "type": "image", + "image_path": "7643c0359d0c3a9cb8ccfc6c73c19521535ea823ea1b8000c49bf05bb12663b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 167, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 130, + 167, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 482, + 224 + ], + "type": "text", + "content": "Fig. 5: We generate different subjects, and plot the first two principal components of the cross-attention queries at different layers of the UNet, where each layer is of different resolution. The high semantic similarity between the kitten and the puppy is expressed by the proximity of their queries through all layers. Meanwhile, the lizard and fruit share similar texture, and hence only their high-resolution queries are entangled." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 114, + 465, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 114, + 465, + 123 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 465, + 123 + ], + "type": "text", + "content": "A lizard A fruit " + }, + { + "bbox": [ + 308, + 114, + 465, + 123 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 308, + 114, + 465, + 123 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 308, + 114, + 465, + 123 + ], + "type": "inline_equation", + "content": "16 \\times 16" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 123, + 467, + 156 + ], + "blocks": [ + { + "bbox": [ + 308, + 123, + 467, + 156 + ], + "lines": [ + { + "bbox": [ + 308, + 123, + 467, + 156 + ], + "spans": [ + { + "bbox": [ + 308, + 123, + 467, + 156 + ], + "type": "image", + "image_path": "a656c60d3d9a5eb3ea2cebf3978c9a9ba84d4486d3b92b505aadd150f86ddfac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "type": "text", + "content": "dictate the subject's semantics and shape, while the outer layers control its style and appearance [?], we analyzed the inner UNet layers. However, leakage can also occur prominently when generating subjects that share visual similarity rather than semantic. We now turn to explore this scenario and demonstrate that, in such cases, the leakage originates from the UNet's outer layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 309, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 430 + ], + "type": "text", + "content": "In Figure 5, we visualize cross-attention queries at different decoder's layers, when generating the kitten, puppy, lizard, and fruit in isolation. As can be seen, the queries of the kitten and the puppy are mixed across all UNet's layers, aligning with the visual and semantic similarity between these animals. On the other hand, the queries of the lizard and the fruit are overlapped only in the highest-resolution layer, aligning with the lack of semantic similarity between them. A closer look reveals that the lizard and the fruit share surprisingly similar textures, which explains the overlap of the queries in the highest-resolution layer. As explained in the previous sections, this overlap causes leakage between the two subjects, in this case, a visual rather than a semantic leakage (see Figure 2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 449, + 269, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 269, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 269, + 462 + ], + "type": "text", + "content": "5 Bounded Attention" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "content": "Our method takes as input " + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "content": " distinct textual subjects " + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "inline_equation", + "content": "S = \\{s_i\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "content": " contained within a global prompt " + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "content": ", along with their corresponding bounding boxes " + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "inline_equation", + "content": "B = \\{b_i\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "content": ". Our objective is to condition the generation process on " + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "content": ", while preserving the intended semantics of each subject, all without requiring any training or fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Figure 6 illustrates an overview of our method. There, the input prompt " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " is \"A kitten and a puppy\", " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "S = \\{kitten\\}, \\{puppy\\}" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ", and the two corresponding bounding boxes " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{b_1, b_2\\}" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " are illustrated in the top left corner. Bounded Attention operates in two modes: Bounded Guidance and Bounded Denoising. Specifically, at the beginning of the denoising process, for " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "t \\in [T, T_{\\text{guidance}}]" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ", we perform a Bounded Guidance step followed by a Bounded Denoising step. In the guidance step, we utilize the Bounded Guidance loss. This interval of timesteps constitutes the optimization phase. Then, for " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "t \\in [T_{\\text{guidance}}, 0]" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ", we apply only Bounded Denoising steps. In both modes, we manipulate the model's forward pass by adopting an augmented weighting scheme in the attention layers, that safeguard the flow of information between the queries and the keys:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "text", + "content": "O. Dahary et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 117, + 449, + 282 + ], + "blocks": [ + { + "bbox": [ + 166, + 117, + 449, + 282 + ], + "lines": [ + { + "bbox": [ + 166, + 117, + 449, + 282 + ], + "spans": [ + { + "bbox": [ + 166, + 117, + 449, + 282 + ], + "type": "image", + "image_path": "37551b515abbb0158346d77723a9ad88db6cb9ae1223715f381023d2a7bad39d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 292, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 413 + ], + "type": "text", + "content": "Fig. 6: Bounded Attention operates in two modes: guidance and denoising, each imposing strict constraints to bound the attention of each subject solely to itself and possibly to the background, preventing any influence from other subjects' features. In guidance mode, a loss is minimized to encourage each subject's attention to concentrate within its corresponding bounding box, with other bounding boxes masked to prevent artifacts. During the denoising step, we confine the attention of each subject solely to its bounding box, along with the background in the self-attention. This strategy effectively prevents feature leakage while maintaining the natural immersion of the subject within the image. Attention maps of a specific key, marked with " + }, + { + "bbox": [ + 130, + 292, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 130, + 292, + 482, + 413 + ], + "type": "text", + "content": ", are shown for each mode: the cross-attention map displays the key corresponding to the \"kitten\" token, and the self-attention map shows a key within the kitten's target bounding box." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 229, + 446, + 481, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 446, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 229, + 446, + 481, + 467 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {t} ^ {(l)} = \\mathrm {s o f t m a x} \\left(\\mathbf {Q} _ {t} ^ {(l)} \\mathbf {K} _ {t} ^ {(l) ^ {\\intercal}} + \\mathbf {M} _ {t}\\right), \\tag {2}", + "image_path": "da00db9702717f76941807518efc300e3dbb45e07b2255275949d2fac321762f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": " represents the layer index, " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": " represents the diffusion timestep, and " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t^{(l)},\\mathbf{K}_t^{(l)}" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": " are the queries and keys of the " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": "-th attention layer. " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_t" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": " represents time-specific masks composed of zeros and " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "-\\infty" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": " elements. We refer to " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t^{(l)}" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": " as the Bounded Attention map. When indexing " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t^{(l)}" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": ", we use pixel coordinates " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": " for rows, and attention-type-specific context vectors " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": " for columns. In locations " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "[\\mathbf{x},\\mathbf{c}]" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_t[\\mathbf{x},\\mathbf{c}] = -\\infty" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": ", it holds that " + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t^{(l)}[\\mathbf{x},\\mathbf{c}] = 0" + }, + { + "bbox": [ + 130, + 472, + 482, + 575 + ], + "type": "text", + "content": ". Therefore, these masks prevent harmful information flow between pixels in self-attention layers, and between pixels and token embeddings in cross-attention layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 580, + 257, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 580, + 257, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 580, + 257, + 590 + ], + "type": "text", + "content": "5.1 Bounded Guidance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In Bounded Guidance, we backpropagate through the diffusion model to steer the latent signal toward the desired layout, using Gradient Descent. Our Bounded Guidance loss encourages the Bounded Attention map of each key corresponding to subject " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", to be within the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " bounding box. To this end, for each subject key we consider the ratio between the attention within the corresponding bounding box, to the entire Bounded Attention map (see Figure 6). Formally, we aggregate" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "text", + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 321, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 321, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 321, + 129 + ], + "type": "text", + "content": "the following loss on the different subjects:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 134, + 481, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 134, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 192, + 134, + 481, + 166 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} = 1 - \\frac {\\sum_ {\\mathbf {x} \\in b _ {i} , \\mathbf {c} \\in C _ {i}} \\hat {\\mathbf {A}} [ \\mathbf {x} , \\mathbf {c} ]}{\\sum_ {\\mathbf {x} \\in b _ {i} , \\mathbf {c} \\in C _ {i}} \\hat {\\mathbf {A}} [ \\mathbf {x} , \\mathbf {c} ] + \\alpha \\sum_ {\\mathbf {x} \\notin b _ {i} , \\mathbf {c} \\in C _ {i}} \\hat {\\mathbf {A}} [ \\mathbf {x} , \\mathbf {c} ]}, \\tag {3}", + "image_path": "a0fd7e96b57cf96163268ed3a089504ed4de4bf0cc6fa3b1d8d6feac943968bb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": " denotes the index of subject " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": " is the mean Bounded Attention map, averaged across heads and layers, and " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": " is a hyperparameter that magnifies the significance of disregarding attention towards the background, as we explain later. Similarly to the above, When indexing " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": ", pixel coordinates " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": " represent rows, and attention-type-specific context vectors " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": " represent columns. We designate " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": " as the set of all " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": "-related context vectors, i.e., pixel coordinates in " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": " for self-attention layers, and the token identifiers of " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": " in cross-attention layers. Additionally, for cross-attention layers, we include the first padding token [EoT] in " + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 275 + ], + "type": "text", + "content": " to enhance layout alignment [?]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "text", + "content": "For each subject " + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "text", + "content": ", the mask " + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_t" + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "text", + "content": " should block the influence of opposing keys (tokens in " + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "text", + "content": " and pixels in " + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "inline_equation", + "content": "b_j" + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "inline_equation", + "content": "j \\neq i" + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "text", + "content": "), to avoid the artifacts illustrated in Figure 3. This fosters that the queries of different subjects, including similar ones, are not erroneously forced to be far apart. Utilizing this loss, our Bounded Guidance step is defined as " + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "inline_equation", + "content": "z_t^{\\mathrm{opt}} = z_t - \\beta \\nabla_{z_t} \\sum_i \\mathcal{L}_i^2" + }, + { + "bbox": [ + 130, + 276, + 480, + 336 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 336, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 480, + 418 + ], + "type": "text", + "content": "Integrating this loss within the cross-attention layers encourages the localization of each subject's semantics within its bounding boxes [?]. However, as cross-attention responses tend to peak around more typical patches associated with the subject's semantics (e.g., the face of a human, the legs of a crab, etc.), it may lack control over the subject's boundaries. By applying the loss within the self-attention layers, we encourage each subject to establish its own boundaries close to its bounding box, thereby discouraging subject fusion (see Figure 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 419, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 480, + 479 + ], + "type": "text", + "content": "In the computation of the loss, we also introduce a hyperparameter " + }, + { + "bbox": [ + 130, + 419, + 480, + 479 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 419, + 480, + 479 + ], + "type": "text", + "content": " to reinforce attention to the background. This adjustment aids in preventing subject amalgamation, where a redundant subject is realized from different subject semantics in the background. To preserve image quality, we limit the application of this mode to an initial time interval " + }, + { + "bbox": [ + 130, + 419, + 480, + 479 + ], + "type": "inline_equation", + "content": "[T, T_{\\text{guidance}}]" + }, + { + "bbox": [ + 130, + 419, + 480, + 479 + ], + "type": "text", + "content": ", following similar works [?,?]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 487, + 259, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 259, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 259, + 498 + ], + "type": "text", + "content": "5.2 Bounded Denoising" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 480, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 570 + ], + "type": "text", + "content": "In Bounded Denoising, we compute the model's output and use it as the next latent in the series. Here, the masks aim to reduce semantic leakage between subjects, as detailed in Section 4, and to prevent unintended semantics from leaking to the background. Unlike Bounded Guidance and typical attention-based guidance approaches, Bounded Denoising can be applied throughout all time steps to mitigate leaks in fine details, which emerge only in later stages [?]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 640 + ], + "type": "text", + "content": "However, coarse masking in later stages may degrade image quality and result in noticeable stitching. To address this, after the optimization phase, for " + }, + { + "bbox": [ + 130, + 570, + 480, + 640 + ], + "type": "inline_equation", + "content": "t \\in [T_{\\text{guidance}}, 0]" + }, + { + "bbox": [ + 130, + 570, + 480, + 640 + ], + "type": "text", + "content": ", we replace each bounding box with a fine segmentation mask obtained by clustering the self-attention maps [?]. Since the subject outlines are roughly determined in the initial time steps and evolve gradually thereafter [?], we refine the masks periodically." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Notably, this mechanism also addresses imperfect alignments between subjects and bounding boxes after the guidance phase, which are more common" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "text", + "content": "O. Dahary et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 457, + 254 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 457, + 254 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 457, + 254 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 457, + 254 + ], + "type": "image", + "image_path": "7f7a4c772b48c2000383a0f93bdcc2553bd7ce60eac9bb384abbb9c2b92ca02e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 264, + 454, + 275 + ], + "lines": [ + { + "bbox": [ + 155, + 264, + 454, + 275 + ], + "spans": [ + { + "bbox": [ + 155, + 264, + 454, + 275 + ], + "type": "text", + "content": "Fig. 7: Our results with SDXL. See the supplemental for more examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 373 + ], + "type": "text", + "content": "when generating numerous subjects. Thus, employing this technique enhances the robustness of our method to seed selection, ensuring proper semantics even when subjects extend beyond their initial confines (see Figure 1,7). Compared to methods that necessitate strict input masks yet remain susceptible to leakage " + }, + { + "bbox": [ + 130, + 301, + 482, + 373 + ], + "type": "inline_equation", + "content": "[?,?,?]" + }, + { + "bbox": [ + 130, + 301, + 482, + 373 + ], + "type": "text", + "content": ", our method offers greater user control with simpler inputs and more satisfactory outcomes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 381, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 482, + 418 + ], + "type": "text", + "content": "Method Details. Further details on the adaptation of Bounded Attention to the cross- and self-attention layers, along with a description of the subject mask refinement process, are provided in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 434, + 230, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 230, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 230, + 448 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 458, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 555 + ], + "type": "text", + "content": "In this section, we conduct both qualitative and quantitative experiments to assess the effectiveness of our Bounded Attention method. We compare our approach with three training-free baseline methods: Layout-guidance (LG) [?], BoxDiff (BD) [?], and MultiDiffusion (MD) [?]. Additionally, we include comparisons with GLIGEN [?] and ReCo [?], which necessitate training. Since Attention-refocusing (AR) [?] is based on GLIGEN, we categorize it as a trained method for the purpose of our evaluation. For fairness, when comparing our method with other methods, we use Stable Diffusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 571, + 258, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 258, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 258, + 583 + ], + "type": "text", + "content": "6.1 Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "SDXL results. We begin our experiments by demonstrating the efficacy of our method in challenging scenarios, particularly when tasked with generating multiple semantically similar subjects using SDXL. As can be seen in Figure 7, Vanilla SDXL fails to follow the prompt due to semantic leakage. It generates inaccurate number of dogs and mixes the features of each breed. In comparison, our approach adeptly generates each dog with its unique characteristics." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 194, + 113, + 417, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 113, + 417, + 122 + ], + "spans": [ + { + "bbox": [ + 194, + 113, + 417, + 122 + ], + "type": "text", + "content": "\"A gray kitten and a ginger kitten and black puppy in a yard.\"" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 123, + 471, + 254 + ], + "blocks": [ + { + "bbox": [ + 141, + 123, + 471, + 254 + ], + "lines": [ + { + "bbox": [ + 141, + 123, + 471, + 254 + ], + "spans": [ + { + "bbox": [ + 141, + 123, + 471, + 254 + ], + "type": "image", + "image_path": "7632eac4e548eea099c73f413c5ea32ce25f2d217f7d841b46de947426cd0c99.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 262, + 447, + 274 + ], + "lines": [ + { + "bbox": [ + 165, + 262, + 447, + 274 + ], + "spans": [ + { + "bbox": [ + 165, + 262, + 447, + 274 + ], + "type": "text", + "content": "Fig. 8: Comparison of the first six images generated from the seed 0." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 481, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 358 + ], + "type": "text", + "content": "Non-curated results. Next, we conduct a non-curated comparison with the training-free baseline methods and present the results in Figure 8. We showcase the initial six images sampled from seed 0 for each method. We anticipate that in the absence of Bounded Attention, semantic leakage may freely blend subject features, hindering the intended layout's formation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "content": "It is evident from the results that none of the competing methods is able to consistently construct the input layout. Layout Guidance [?] frequently neglects one of the subjects, and even when it generates three subjects, it struggles to avoid leakage, resulting in puppies with kitten-like features or incorrect color assignments. BoxDiff [?] often generates the correct number of subjects but suffers from artifacts in the form of blobs. Similar to Layout Guidance, it encounters difficulties in properly constructing the puppy. Surprisingly, even MultiDiffusion [?], which generates the subjects separately, faces challenges in generating them all, with some disappearing or merging together in its bootstrapping phase." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 469, + 481, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 469, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 481, + 493 + ], + "type": "text", + "content": "In contrast, our method consistently outperforms these approaches, producing three subjects that align with the both prompt and layout in all six images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 508, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 481, + 581 + ], + "type": "text", + "content": "Comparisons with baselines. We present a qualitative comparison in Figure 9. All competing methods, including those trained specifically for the layout-to-image task, exhibit significant visual and semantic leakage. The training-free methods perform the worst: MultiDiffusion produces disharmonious, low-quality images, while optimization-based methods often result in object fusion, combining different semantics without adhering to the layout." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 581, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 481, + 628 + ], + "type": "text", + "content": "The training-based approaches closely follow the layout but fail to convey the correct semantics. In the first row, they neglect the corduroy jacket, leaking the denim texture into the other jacket, or even fusing them together. In the second row, the elephant's features leak into the rhino." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 630, + 481, + 665 + ], + "type": "text", + "content": "In comparison, our method generates images that align with the input layout and prompt, ensuring each subject retains its unique attributes, semantics, and appearance." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "O. Dahary et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 157, + 114, + 455, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 114, + 455, + 125 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 455, + 125 + ], + "type": "text", + "content": "\"A denim jacket and a corduroy jacket and a leather handbag in a closet.\"" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 125, + 480, + 176 + ], + "blocks": [ + { + "bbox": [ + 134, + 125, + 480, + 176 + ], + "lines": [ + { + "bbox": [ + 134, + 125, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 134, + 125, + 480, + 176 + ], + "type": "image", + "image_path": "1bad5bcf3921fd21d36b4c9872dd3e87124038ceb0f0426b0dd2ea796b2a46c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 177, + 416, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 177, + 416, + 188 + ], + "spans": [ + { + "bbox": [ + 196, + 177, + 416, + 188 + ], + "type": "text", + "content": "\"A big red elephant and a far away rhino in a jungle.\"" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 188, + 480, + 254 + ], + "blocks": [ + { + "bbox": [ + 133, + 188, + 480, + 254 + ], + "lines": [ + { + "bbox": [ + 133, + 188, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 133, + 188, + 480, + 254 + ], + "type": "image", + "image_path": "9a947536fbd0f4c66b6701a2b8d231a27646ef426f02a6f3c4daa2af170fc0a4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "type": "text", + "content": "Fig. 9: Qualitative comparison of our method with baseline methods: Above each row, we display the input prompt, where each subject's color matches the color of its corresponding bounding box. We compare with both training-free methods (2nd-4th columns) and trained models (5th-7th columns). See supplemental for more results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 359, + 264, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 264, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 264, + 370 + ], + "type": "text", + "content": "6.2 Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 370, + 264, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 264, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 264, + 490 + ], + "type": "text", + "content": "We evaluate our method's effectiveness using the Draw-Bench dataset [?], known for its challenging prompts designed to test a model's ability to compose multiple subjects with specific quantities and relations. We use the evaluation procedure from previous work [?, ?]. Our re" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 284, + 374, + 469, + 449 + ], + "blocks": [ + { + "bbox": [ + 284, + 374, + 469, + 449 + ], + "lines": [ + { + "bbox": [ + 284, + 374, + 469, + 449 + ], + "spans": [ + { + "bbox": [ + 284, + 374, + 469, + 449 + ], + "type": "table", + "html": "
MethodCountingSpatial
PrecisionRecallF1Accuracy
Stable Diffusion [?]0.740.780.730.19
Layout-guidance [?]0.720.780.720.35
BoxDiff [?]0.810.780.760.28
MultiDiffusion [?]0.700.550.570.15
Ours0.810.910.820.43
", + "image_path": "e6df5d0aa843fb7a14335a4e0fed3ef88fa7f2245e083d64ca67dcfeb2bc8fd3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 276, + 456, + 474, + 468 + ], + "lines": [ + { + "bbox": [ + 276, + 456, + 474, + 468 + ], + "spans": [ + { + "bbox": [ + 276, + 456, + 474, + 468 + ], + "type": "text", + "content": "Table 1: Evaluation on the DrawBench dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 490, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 561 + ], + "type": "text", + "content": "sults, alongside those of other training-free methods, are summarized in Table 1. Unlike other approaches that do not account for semantic leakage, our method demonstrates notable improvements in both the counting and the spatial categories. Notably, other methods struggle to surpass the counting recall rates of vanilla SD, highlighting the effectiveness of Bounded Attention in generating multiple subjects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 561, + 481, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 481, + 586 + ], + "type": "text", + "content": "We further conduct a quantitative comparison of semantic leakage in training-free methods through a user study. Further details are in the supplemental." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 594, + 246, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 246, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 246, + 605 + ], + "type": "text", + "content": "6.3 Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "content": "To assess the significance of each component, we conduct an ablation study where we systematically vary our method's configurations by omitting one component in each setting. We show the results in Figure 10." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Guidance is crucial for aligning the latent signal with the intended layout. Neglecting it can lead to subject generation in the background, as demonstrated" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 235, + 114, + 376, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 114, + 376, + 122 + ], + "spans": [ + { + "bbox": [ + 235, + 114, + 376, + 122 + ], + "type": "text", + "content": "\"A red lizard and a turtle on the grass.\"" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 122, + 185, + 170 + ], + "blocks": [ + { + "bbox": [ + 137, + 122, + 185, + 170 + ], + "lines": [ + { + "bbox": [ + 137, + 122, + 185, + 170 + ], + "spans": [ + { + "bbox": [ + 137, + 122, + 185, + 170 + ], + "type": "image", + "image_path": "591aeb0f9a2447109048e47333aeb97c7bbed21b31e4210a15ef7b9cd1d89c15.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 171, + 173, + 180 + ], + "lines": [ + { + "bbox": [ + 149, + 171, + 173, + 180 + ], + "spans": [ + { + "bbox": [ + 149, + 171, + 173, + 180 + ], + "type": "text", + "content": "w/o G" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 186, + 123, + 233, + 170 + ], + "blocks": [ + { + "bbox": [ + 186, + 123, + 233, + 170 + ], + "lines": [ + { + "bbox": [ + 186, + 123, + 233, + 170 + ], + "spans": [ + { + "bbox": [ + 186, + 123, + 233, + 170 + ], + "type": "image", + "image_path": "19a6b8330b6947168d23ccc642327b532875c79758a434328cf599019c195e50.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 171, + 224, + 180 + ], + "lines": [ + { + "bbox": [ + 194, + 171, + 224, + 180 + ], + "spans": [ + { + "bbox": [ + 194, + 171, + 224, + 180 + ], + "type": "text", + "content": "w/o BG" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 234, + 123, + 282, + 170 + ], + "blocks": [ + { + "bbox": [ + 234, + 123, + 282, + 170 + ], + "lines": [ + { + "bbox": [ + 234, + 123, + 282, + 170 + ], + "spans": [ + { + "bbox": [ + 234, + 123, + 282, + 170 + ], + "type": "image", + "image_path": "3f73670f251fa60fce3425654f92c11a874c4c1233551299f1d7d337be678dea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 171, + 272, + 180 + ], + "lines": [ + { + "bbox": [ + 242, + 171, + 272, + 180 + ], + "spans": [ + { + "bbox": [ + 242, + 171, + 272, + 180 + ], + "type": "text", + "content": "w/o BD" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 189, + 481, + 234 + ], + "lines": [ + { + "bbox": [ + 131, + 189, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 131, + 189, + 481, + 234 + ], + "type": "text", + "content": "Fig. 10: Qualitative ablation. We ablate our method by skipping the guidance step (G), performing guidance without Bounded Guidance (BG), not applying Bounded Denoising (BD), and not performing Mask Refinement (MR). We show zoomed-in images of the two rightmost configurations. More examples can be found in the supplemental." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 282, + 123, + 329, + 170 + ], + "blocks": [ + { + "bbox": [ + 282, + 123, + 329, + 170 + ], + "lines": [ + { + "bbox": [ + 282, + 123, + 329, + 170 + ], + "spans": [ + { + "bbox": [ + 282, + 123, + 329, + 170 + ], + "type": "image", + "image_path": "25bd0bd466b74bc1f55d42a895f017273a814097b8e9e2bb87f36241a7caf1bc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 171, + 345, + 180 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 345, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 345, + 180 + ], + "type": "text", + "content": "w/o MR" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 123, + 376, + 170 + ], + "blocks": [ + { + "bbox": [ + 329, + 123, + 376, + 170 + ], + "lines": [ + { + "bbox": [ + 329, + 123, + 376, + 170 + ], + "spans": [ + { + "bbox": [ + 329, + 123, + 376, + 170 + ], + "type": "image", + "image_path": "446db0dd9c92ae76d87114eb4082cde8f14cdddaf98c919d3a9e7bc50b4c1148.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 377, + 122, + 425, + 170 + ], + "blocks": [ + { + "bbox": [ + 377, + 122, + 425, + 170 + ], + "lines": [ + { + "bbox": [ + 377, + 122, + 425, + 170 + ], + "spans": [ + { + "bbox": [ + 377, + 122, + 425, + 170 + ], + "type": "image", + "image_path": "96ca2fd3d8a4586ad16dd6d1bead476365fa8c728a49d5a5d5d4b7d4a9efb473.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 171, + 447, + 180 + ], + "lines": [ + { + "bbox": [ + 403, + 171, + 447, + 180 + ], + "spans": [ + { + "bbox": [ + 403, + 171, + 447, + 180 + ], + "type": "text", + "content": "Full Method" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 122, + 474, + 170 + ], + "blocks": [ + { + "bbox": [ + 425, + 122, + 474, + 170 + ], + "lines": [ + { + "bbox": [ + 425, + 122, + 474, + 170 + ], + "spans": [ + { + "bbox": [ + 425, + 122, + 474, + 170 + ], + "type": "image", + "image_path": "596f51289e632be8f3f0a3dd6581aa9c978a1d5b730263be74db7b0effb50674.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 245, + 480, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 292 + ], + "type": "text", + "content": "by the emergence of a turtle-lizard hybrid. However, attempting to guide the latent signal without our Bounded Guidance mechanism may cause distortions in subject appearances, such as the turtle's neck growing out of its leg and inconsistent artifacts appearing on the lizard's head and back." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 293, + 480, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 328 + ], + "type": "text", + "content": "Meanwhile, forgoing Bounded Denoising results in noticeable semantic leakage, with the lizard being replaced by a turtle, and the \"red\" attribute erroneously leaking to the wrong subject." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 329, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 363 + ], + "type": "text", + "content": "Lastly, incorporating mask refinement in the later stages prevents fine-details from leaking. Without mask refinement, the lizard exhibits shell-like contours on its back." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 373, + 225, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 225, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 225, + 385 + ], + "type": "text", + "content": "7 Conclusions" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": "We introduce Bounded Attention, a technique designed to regulate the accurate generation of multiple subjects within an image. This approach encourages each subject to \"be yourself\", emphasizing the importance of preserving individuality and uniqueness without being excessively influenced by other subjects present in the image. Our development of the Bounded Attention technique stemmed from an in-depth analysis of the root causes behind the misalignment observed between the provided prompt and the resulting generated image. Our investigation revealed that this misalignment primarily arises due to semantic leakage among the generated subjects, a phenomenon observed in both the cross and self-attention layers." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "text", + "content": "While Bounded Attention effectively mitigates a significant portion of semantic leakage, it does not entirely eliminate it. Our findings demonstrate a marked improvement in performance compared to other methods that seek to achieve semantic alignment. However, residual leakage persists, which we attribute to imperfect optimization during the guidance mode and inaccurate segmentation of the subject prior to the second phase." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "While Bounded Attention excels in generating multiple subjects with plausible semantic alignment, its performance may vary across different layouts. Achieving success with Bounded Attention hinges on a strong match between the seed and the layout. Moving forward, we aim to explore methods for generating well-suited seeds tailored to specific layouts. One potential avenue is to introduce noise to the layout image, thereby creating a seed that aligns more closely with the desired outcomes." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "text", + "content": "O. Dahary et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 241, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 241, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 241, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 164 + ], + "type": "text", + "content": "We thank Guy Tevet for fruitful discussions and useful suggestions. This work is partly supported by a research gift from Snap Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_content_list.json b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e5586d75f819215215b43f3b2100ddc4d088d4d --- /dev/null +++ b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_content_list.json @@ -0,0 +1,1726 @@ +[ + { + "type": "text", + "text": "Be-Your-Outpainter: Mastering Video Outpainting through Input-Specific Adaptation", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fu-Yun Wang $^{1}$ Xiaoshi Wu $^{1}$ Zhaoyang Huang $^{2}$", + "bbox": [ + 303, + 210, + 702, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyu Shi $^{1}$ Dazhong Shen $^{3}$ Guanglu Song $^{4}$", + "bbox": [ + 316, + 227, + 700, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Liu $^{3}$ Hongsheng Li $^{1,3,5}$", + "bbox": [ + 392, + 243, + 611, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MMLab, CUHK $^{2}$ Avolution AI", + "bbox": [ + 387, + 268, + 614, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Shanghai AI Lab $^{4}$ SenseTime Research $^{5}$ CPII under InnoHK", + "bbox": [ + 284, + 282, + 732, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fywang@link, hsli@ee}.cuhk.edu.hk", + "bbox": [ + 370, + 299, + 633, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Video outpainting is a challenging task, aiming at generating video content outside the viewpoint of the input video while maintaining inter-frame and intra-frame consistency. Existing methods fall short in either generation quality or flexibility. We introduce MOTIA (Mastering Video Outpainting Through Input-Specific Adaptation), a diffusion-based pipeline that leverages both the intrinsic data-specific patterns of the source video and the image/video generative prior for effective outpainting. MOTIA comprises two main phases: input-specific adaptation and pattern-aware outpainting. The input-specific adaptation phase involves conducting efficient and effective pseudo outpainting learning on the single-shot source video. This process encourages the model to identify and learn patterns within the source video, as well as bridging the gap between standard generative processes and outpainting. The subsequent phase, pattern-aware outpainting, is dedicated to the generalization of these learned patterns to generate outpainting outcomes. Additional strategies are proposed to better leverage the diffusion model's generative prior and the acquired video patterns from source videos for inference. Extensive evaluations underscore MOTIA's superiority, outperforming existing state-of-the-art methods in widely recognized benchmarks. Notably, these advancements are achieved without necessitating extensive, task-specific tuning. More details are available at https://be-your-outpainter.github.io/.", + "bbox": [ + 259, + 344, + 743, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video outpainting aims to expand the visual content out of the spatial boundaries of videos, which has important real-world applications [3,5,6]. For instance, in practice, videos are usually recorded with a fixed aspect ratio, such as in movies or short clips. This becomes an issue when viewing these videos on smartphones, which often have varying aspect ratios, resulting in unsightly black bars on the screen that detract from the viewing experience. Proper ways for video outpainting are crucial in solving this issue. By expanding the visual content beyond the original frame, it adapts the video to fit various screen sizes seamlessly. This process ensures that the audience enjoys a full-screen experience without any", + "bbox": [ + 212, + 703, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/050d0281f73a46062e6c49285c721d0512498e66ab73eb53909bc5814d6f2094.jpg", + "image_caption": [ + "Fig. 1: MOTIA is a high-quality flexible video outpainting pipeline, leveraging the intrinsic data-specific patterns of source videos and image/video generative prior for state-of-the-art performance. Quantitative metric improvement of MOTIA is significant (Table 1)." + ], + "image_footnote": [], + "bbox": [ + 250, + 143, + 424, + 446 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/870d2ca7013cd6a21e83dd46eb593a8b798574f79b7c2171f48cd60ddb98b540.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 143, + 754, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "compromise in visual integrity. However, the challenges associated with video outpainting are significant. It requires not just the expansion of each frame's content but also the preservation of temporal (inter-frame) and spatial (intra-frame) consistency across the video.", + "bbox": [ + 212, + 547, + 787, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, there are two primary approaches to video outpainting. The first employs optical flows and specialized warping techniques to extend video frames, involving complex computations and carefully tailored hyperparameters to ensure the added content remains consistent [5, 7]. However, their results are far from satisfactory, suffering from blurred content. The other type of approach in video outpainting revolves around training specialized models tailored for video inpainting and outpainting with extensive datasets [6, 35]. However, they have two notable limitations: 1) An obvious drawback of these models is their dependency on the types of masks and the resolutions of videos they can handle, which significantly constrains their versatility and effectiveness in real-world applications, as they may not be adequately equipped to deal with the diverse range of video formats and resolutions commonly encountered in practical scenarios. 2) The other drawback is their inability to out-domain video outpainting, even intensively trained on massive video data. Fig. 5 shows a failure example of most advanced previous work [6] that the model faces complete outpainting failure,", + "bbox": [ + 212, + 613, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Fu-Yun Wang et al.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with only blurred corners. We show that a crucial reason behind this is that the model fails at capturing the intrinsic data-specific patterns from out-domain source (input) videos.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose MOTIA: Mastering Video Outpainting Through Input-Specific Adaptation, a diffusion-based method for open-domain video outpainting with arbitrary types of mask, arbitrary video resolutions and lengths, and arbitrary styles. At the heart of MOTIA is treating the source video itself as a rich source of information [18, 22], which contains key motion and content patterns (intrinsic data-specific patterns) necessary for effective outpainting. We demonstrate that the patterns learned from the source video, combined with the generative capabilities of diffusion models, can achieve surprisingly great outpainting performance.", + "bbox": [ + 212, + 196, + 787, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the workflow of MOTIA. MOTIA consists of two stages: input-specific adaptation and pattern-aware outpainting. During the input-specific adaptation stage, we conduct pseudo video outpainting learning on the source video (videos to be outpainted) itself. Specifically, at each iteration, we heuristically add random masks to the source video and prompt the base diffusion model to recover the masked regions by learning to denoise the video corrupted by white noise, relying on the extracted information from unmasked regions. This process not only allows the model to capture essential data-specific patterns from the source video but also narrows the gap between standard generation and outpainting. We insert trainable lightweight adapters into the diffusion model for tuning to keep the efficiency and stability. In the pattern-aware outpainting stage, we combine the learned patterns from the source video and the generation prior of the diffusion model for effective outpainting. To better leverage the generation ability of the pretrained diffusion model and the learned pattern from the source video, we propose spatial-aware insertion (SA-Insertion) of the tuned adapters for outpainting. Specifically, the insertion weights of adapters gradually decay as the spatial position of features away from the known regions. In this way, the outpainting of pixels near the known regions is more influenced by the learned patterns, while the outpainting of pixels far from the known regions relies more on the original generative prior of diffusion model. To further mitigate potential denoising conflicts and enhance the knowledge transfer between known regions and unknown regions, we incorporate noise regret that we add noise and denoise periodically at early inference steps, which works for more harmonious outpainting results.", + "bbox": [ + 212, + 335, + 789, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensively quantitative and qualitative experiments verify the effectiveness of our proposed method. MOTIA overcomes many limitations of previous methods and outperforms the state-of-the-art intensively trained outpainting method in standard widely used benchmarks. In summary, our contribution is three-fold: 1) We show that the data-specific patterns of source videos are crucial for effective outpainting, which is neglected by previous work. 2) We introduce an adaptation strategy to effectively capture the data-specific patterns and then propose novel strategies to better leverage the captured patterns and pretrained image/video generative prior for better results. 3) Vast experiments verify that", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Be-Your-Outpainter", + "bbox": [ + 596, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2e51e50cd12449630c71e6942a4f9b5dfec56676fde7061c119d5a82fe5e07af.jpg", + "image_caption": [ + "Fig. 2: Workflow of MOTIA. Blue lines represent the workflow of input-specific adaptation, and green lines represent the workflow of pattern-aware outpainting." + ], + "image_footnote": [], + "bbox": [ + 256, + 143, + 754, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "our performance in video outpainting is great, significantly outperforming previous state-of-the-art methods in both quantitative metrics and user studies.", + "bbox": [ + 214, + 402, + 784, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 458, + 395, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we discuss related diffusion models and outpainting methods. Diffusion models. Diffusion models (a.k.a., score-based models) [9,10,17,20,21] have gained increasing attention due to its amazing ability to generate highly detailed images. Current video diffusion models [4, 9, 15, 16, 24, 25, 28, 30] are generally built upon the extension of image diffusion models through insertion of temporal layers. They are either trained with image-video joint tuning [12, 22] or trained with spatial weights frozen [4] to mitigate the negative influence of the poor captions and visual quality of video data.", + "bbox": [ + 212, + 491, + 766, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Outpainting methods. Video outpainting is largely built upon the advancements in image outpainting, where techniques ranged from patch-based methods (e.g., PatchMatch [3]) to more recent deep learning approaches like GANs [1,34]. Diffusion models [2,14], benefiting from the learned priors on synthesis tasks and the process of iterative refinement, achieve state-of-the-art performance on image outpainting tasks. The research focusing on video outpainting is comparatively few. Previous works typically apply optical flow for outpainting, which warps content from adjacent frames to the outside corners, but their results are far from satisfactory. Recently, M3DDM [6] trained a large 3D diffusion models with specially designed architecture for outpainting on massive video data, achieving much better results compared to previous methods, showcasing the huge potential of diffusion models on video outpainting. However, as we claimed, they have two main limitations: 1) The inflexibility for mask types and video resolutions. They can only outpaint video with resolution $256 \\times 256$ with square type of masking. 2) Inability for out-domain video outpainting. As shown in", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Fu-Yun Wang et al.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 5, they encounter outpainting failure when processing out-domain videos even intensively trained on massive video data.", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 204, + 382, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion models [10] add noise to data through a Markov chain process. Given initial data $\\pmb{x}_0 \\sim q(\\pmb{x}_0)$ , this process is formulated as:", + "bbox": [ + 214, + 239, + 782, + 271 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {x} _ {1: T} \\mid \\boldsymbol {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right), \\quad q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 284, + 785, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ is the noise schedule and $\\alpha_{t} = 1 - \\beta_{t}$ . The data reconstruction, or denoising process, is accomplished by the reverse transition modeled by $p_{\\theta}(\\boldsymbol{x}_{t-1}|\\boldsymbol{x}_t)$ :", + "bbox": [ + 214, + 337, + 782, + 368 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1}; \\tilde {\\boldsymbol {\\mu}} _ {t} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 383, + 785, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\tilde{\\pmb{\\mu}}_t(\\pmb{x}_t,\\pmb{x}_0) = \\frac{1}{\\sqrt{\\alpha_t}}\\pmb{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t\\sqrt{\\alpha_t}}}\\pmb{\\epsilon}, \\bar{\\alpha}_t = \\prod_{s=1}^t\\alpha_s, \\tilde{\\beta}_t = \\frac{1 - \\bar{\\alpha}_{t-1}}{1 - \\bar{\\alpha}_t}\\beta_t$ , and $\\pmb{\\epsilon}$ is the noise added to $\\pmb{x}_0$ to obtain $\\pmb{x}_t$ .", + "bbox": [ + 214, + 409, + 782, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion-based outpainting aims to predict missing pixels at the corners of the masked region with the pre-trained diffusion models. We denote the ground truth as $\\pmb{x}$ , mask as $\\pmb{m}$ , known region as $(1 - m) \\odot \\pmb{x}$ , and unknown region as $m \\odot x$ . At each reverse step in the denoising process, we modify the known regions by incorporating the intermediate noisy state of the source data from the corresponding timestep in the forward diffusion process (which adds noise), provided that this maintains the correct distribution of correspondences. Specifically, each reverse step can be denoted as the following formulas:", + "bbox": [ + 214, + 444, + 784, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t - 1} ^ {\\text {k n o w n}} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\quad \\boldsymbol {x} _ {t - 1} ^ {\\text {u n k n o w n}} \\sim \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {\\theta} (\\boldsymbol {x} _ {t}, t), \\Sigma_ {\\theta} (\\boldsymbol {x} _ {t}, t)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 579, + 785, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t - 1} = \\boldsymbol {m} \\odot \\boldsymbol {x} _ {t - 1} ^ {\\text {k n o w n}} + (\\boldsymbol {1} - \\boldsymbol {m}) \\odot \\boldsymbol {x} _ {t - 1} ^ {\\text {u n k n o w n}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 628, + 785, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\pmb{x}_{t - 1}^{\\mathrm{known}}$ is padded to the target resolution before the masked merging.", + "bbox": [ + 214, + 655, + 782, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 696, + 380, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents MOTIA, a method demonstrating exceptional performance in video outpainting tasks. We begin by defining the concept of video outpainting and describing the foundational model in Section 4.1. and Section 4.2. Subsequently, we delve into the specifics of input-specific adaptation and pattern-aware outpainting in Sections 4.3 and 4.4, respectively. Moreover, we show that our approach has great promise in extending its application to long video outpainting, which will be explored in Section 4.5.", + "bbox": [ + 214, + 734, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Be-Your-Outpainter", + "bbox": [ + 596, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/998d63b3de30b762524377e7f4ca953c13b7ee17ff387a654343256b1558191c.jpg", + "image_caption": [ + "Fig. 3: Sample results of quantitative experiments. All videos are outpainted with a horizontal mask ratio of 0.66. Contents outside the yellow lines are outpainted by MOTIA." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 759, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 571, + 437, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a video represented as $\\pmb{v} \\in \\mathbb{R}^{t \\times h \\times w \\times d}$ , where $t$ denotes the number of frames, $h$ denotes frame height, $w$ denotes frame width, and $d$ denotes channel depth. Video outpainting model $f(\\pmb{v})$ is designed to generate a spatially expanded video $\\pmb{v}' \\in \\mathbb{R}^{t \\times h' \\times w' \\times d}$ . This process not only increases the spatial dimensions ( $h' > h$ , $w' > w$ ), but also requires to ensure continuity and harmony between the newly expanded regions and the known regions. The transformation maintains the known regions unchanged, with $f(\\pmb{v})$ acting as an identity in these regions.", + "bbox": [ + 212, + 595, + 787, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Network Expansion", + "text_level": 1, + "bbox": [ + 215, + 724, + 423, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network inflation. MOTIA leverages the pre-trained text-to-image (T2I) model, Stable Diffusion. In line with previous video editing techniques [32], we transform 2D convolutions into pseudo 3D convolutions and adapt 2D group normalizations into 3D group normalizations to process video latent features. Specifically, the $3 \\times 3$ kernels in convolutions are replaced by $1 \\times 3 \\times 3$ kernels, maintaining identical weights. Group normalizations are executed across both temporal", + "bbox": [ + 212, + 750, + 795, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Fu-Yun Wang et al.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and spatial dimensions, meaning that all 3D features within the same group are normalized simultaneously, followed by scaling and shifting.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Masked video as conditional input. Additionally, we incorporate a ControlNet [36], initially trained for image inpainting, to manage additional mask inputs. Apart from noise input, ControlNet can also process masked videos to extract effective information for more controllable denoising. In these masked videos, known regions have pixel values ranging from 0 to 1, while values of masked regions are set to $-1$ .", + "bbox": [ + 212, + 176, + 785, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal consistency prior. To infuse the model with temporal consistency priors, we integrate temporal modules pre-trained on text-to-video (T2V) tasks. Note that although MOTIA relies on pre-trained video diffusion modules, applying these pre-trained temporal modules directly for video outpainting yields rather bad results, significantly inferior to all baseline methods (Table. 2). However, when equipped with our proposed MOTIA, the model demonstrates superior performance even in comparison to models specifically designed and trained for video outpainting, underscoring the efficacy of MOTIA.", + "bbox": [ + 212, + 268, + 787, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Input-Specific Adaptation", + "text_level": 1, + "bbox": [ + 215, + 415, + 475, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The input-specific adaptation phase is crucial in our video outpainting method, aiming to tailor the model for the specific challenges of outpainting. This phase involves training on the source video with a pseudo-outpainting task, importantly, enabling the model to learn intrinsic content and motion patterns (data-specific patterns) within the source video as well as narrowing the gap between the standard generation process and outpainting.", + "bbox": [ + 212, + 443, + 782, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video augmentation. Initially, we augment the source video. Transformations like identity transformation, random flipping, cropping, and resizing can be employed. This step can potentially help the model better learn and adapt to diverse changes in video content. For longer video outpainting, as we will discuss later, instead of taking it as a whole, we randomly sample short video clips from it to reduce the cost of the adaptation phase.", + "bbox": [ + 212, + 534, + 785, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video masking. We then add random masks to the video. We adopt a heuristic approach that uniformly samples edge boundaries of 4 sides within given limits. The area enclosed by these boundaries is considered the known region, while the rest is the unknown region. This masked video serves as the conditional input for the ControlNet, simulating the distribution of known and unknown areas in actual outpainting scenarios.", + "bbox": [ + 212, + 625, + 785, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video noising. Additionally, we apply noise to the video following the DDPM [10] by randomly selecting diffusion timesteps. This noisy video serves as an input for both the ControlNet and the Stable Diffusion model.", + "bbox": [ + 212, + 715, + 794, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimization. Finally, we optimize the model. To ensure efficiency, low-rank adapters are inserted into the layers of the diffusion model. We optimize only the parameters of these adapters while keeping the other parameters frozen. The loss function is", + "bbox": [ + 212, + 762, + 785, + 821 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\| \\boldsymbol {\\epsilon} - \\hat {\\boldsymbol {\\epsilon}} _ {\\bar {\\theta} _ {l}, \\bar {\\theta} _ {c}, \\boldsymbol {\\theta} _ {a}} (\\boldsymbol {v} _ {\\mathrm {n o i s y}}, \\boldsymbol {v} _ {\\mathrm {m a s k e d}}, t) \\right\\| _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 823, + 785, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Be-Your-Outpainter", + "bbox": [ + 596, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81d223586e6d6f8f65400938a9f42153d292da8e22227d435f3d96639c29818c.jpg", + "image_caption": [ + "Fig.4: Spatial-aware insertion scales the insertion weights of adapters for better leveraging of learned patterns and generative prior." + ], + "image_footnote": [], + "bbox": [ + 217, + 169, + 431, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f8e59ec18a3fc9a843bde6c668506280c2883f48460e5e041a0d17da4dba2391.jpg", + "image_caption": [ + "Fig. 5: Failure example of previous methods. Many previous methods including the intensively trained models on video outpainting still might suffer from generation failure, that the model simply generates blurred corners. MOTIA never encounters this failure." + ], + "image_footnote": [], + "bbox": [ + 454, + 145, + 785, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $t$ represents the timestep in the process, $\\epsilon$ is the added noise, $\\pmb{v}_{\\mathrm{noisy}}$ refers to the video perturbed by $\\epsilon$ , and $\\pmb{v}_{\\mathrm{masked}}$ denotes the masked video. The parameters $\\theta_{l}$ , $\\theta_{c}$ , and $\\theta_{a}$ correspond to the Diffusion Model, ControlNet, and adapters, respectively. The bar over these parameters indicates they are frozen during the optimization. This optimization process, including the steps of augmentation, masking, and noising, is repeated to update the lightweight adapters to capture the data-specific patterns from the source video.", + "bbox": [ + 212, + 378, + 787, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 Pattern-Aware Outpainting", + "text_level": 1, + "bbox": [ + 215, + 516, + 488, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following the initial phase of input-specific adaptation, our model shows promising results in video outpainting using basic pipelines as outlined in Eq. 3 and Eq. 4, achieving commendable quality. However, we here introduce additional inference strategies that can be combined to better leverage the learned data-specific patterns from the input-specific adaptation phase for better outpainting results. We call the outpainting process that incorporates these strategies pattern-aware outpainting.", + "bbox": [ + 212, + 551, + 787, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial-aware insertion. It is important to acknowledge that in the input-specific adaptation phase, the model is fine-tuned through learning outpainting within the source video. However, at the outpainting phase, the model is expected to treat the entire source video as known regions and then fill the unknown regions at edges (i.e., generating a video with a larger viewpoint and resolution). This specificity may lead to a noticeable training-inference gap during outpainting, potentially affecting the outpainting quality. To balance the fine-tuned patterns with the diffusion model's inherent generative prior, we introduce the concept of spatial-aware insertion (SA-Insertion) of adapters as shown in Fig. 4. The adaptation involves adjusting the insertion weight of tuned low-rank adapters based on the feature's spatial position. We increase insertion weight near known areas to utilize the learned patterns while decreasing it in farther regions to rely", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Fu-Yun Wang et al.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "more on the original generative capacity of the diffusion model. To be specific,", + "bbox": [ + 215, + 146, + 779, + 161 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {\\mathrm {a d a p t e d}} ^ {\\top} \\boldsymbol {x} _ {\\boldsymbol {p}} = \\mathbf {W} ^ {\\top} \\boldsymbol {x} _ {\\boldsymbol {p}} + \\alpha (\\boldsymbol {p}) \\left(\\mathbf {W} _ {\\mathrm {u p}} \\mathbf {W} _ {\\mathrm {d o w n}}\\right) ^ {\\top} \\boldsymbol {x} _ {\\boldsymbol {p}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 172, + 785, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $\\pmb{p}$ signifies the spatial position of $\\pmb{x}$ , $\\mathbf{W} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times d_{\\mathrm{out}}}$ denotes the linear transformation in layers of diffusion model, $\\mathbf{W}_{\\mathrm{down}} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times r}$ and $\\mathbf{W}_{\\mathrm{up}} \\in \\mathbb{R}^{r \\times d_{\\mathrm{out}}}$ are the linear components of the adapter with rank $r \\ll \\min(d_{\\mathrm{in}}, d_{\\mathrm{out}})$ . The function $\\alpha(\\pmb{p})$ is defined as:", + "bbox": [ + 214, + 204, + 782, + 266 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha (\\boldsymbol {p}) = \\exp \\left(- \\frac {K \\| \\boldsymbol {p} - \\boldsymbol {p} _ {c} \\|}{\\operatorname* {m a x} _ {\\bar {\\boldsymbol {p}}} \\| \\bar {\\boldsymbol {p}} - \\boldsymbol {p} _ {c} \\|}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 277, + 785, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $K$ is a constant for controlling decay speed, and $\\pmb{p}_c$ represents the nearest side of the known region to $\\pmb{p}$ .", + "bbox": [ + 214, + 321, + 782, + 349 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Noise regret. In the context of Eq. 3, merging noisy states from known and unknown regions in video outpainting tasks is similar to sampling from two different vectors, which can potentially disrupt the denoising direction. Given the significant impact of early steps on the generation's structure, later denoising may not rectify these initial discrepancies. Inspired by DDPM-based image inpainting methods [14,21], we propose to re-propagate the noisy state into a noisier state by adding noise when denoising and then provide the model a second chance for re-denoising. This helps integrate known region data more effectively and reduces denoising direction conflicts. In detail, after obtaining $\\mathbf{v}_t$ during denoising, we conduct", + "bbox": [ + 214, + 352, + 784, + 500 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {t + L} = \\sqrt {\\Pi_ {i = t + 1} ^ {t + L} \\alpha_ {i}} \\boldsymbol {v} _ {t} + \\sqrt {1 - \\Pi_ {i = t + 1} ^ {t + L} \\alpha_ {i}} \\boldsymbol {\\epsilon}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 501, + 785, + 526 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\alpha_{i} = 1 - \\beta_{i}$ and $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . Then we restart the denoising process. We repeat this progress for $M$ times. We only conduct it in the early denoising steps.", + "bbox": [ + 214, + 532, + 782, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5 Extension to Long Video Outpainting", + "text_level": 1, + "bbox": [ + 215, + 588, + 571, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We show that our method can be easily extended for long video outpainting. Specifically, for the stage of input-specific adaptation, instead of taking the long video as a whole for adaptation (Direct adaptation on long videos is costly and does not align with the video generation prior of the pretrained modules), we randomly sample short video clips from the long video for tuning to learn global patterns without requiring more GPU memory cost. For the stage of pattern-aware outpainting, we split the long video into short video clips with temporal overlapping (i.e., some frames are shared by different short video clips), and then conduct temporal co-denoising following Gen-L [29]. Specifically, the denoising result for $j^{th}$ frame of the long video at timestep $t$ is approximated by the weighted sum of all the corresponding frames in short video clips that contain it,", + "bbox": [ + 214, + 614, + 782, + 795 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {t - 1, j} = \\frac {\\sum_ {i \\in \\mathcal {I} ^ {j}} \\left(\\left(W _ {i , j ^ {*}}\\right) ^ {2} \\otimes \\boldsymbol {v} _ {t - 1 , j ^ {*}} ^ {i}\\right)}{\\sum_ {i \\in \\mathcal {I} ^ {j}} \\left(W _ {i , j ^ {*}} ^ {2}\\right) ^ {2}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 797, + 785, + 844 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Be-Your-Outpainter", + "bbox": [ + 596, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a9eb8b1e7124d465cde9d8f138a9a0a1feccd55ca0fb9a4d7dc3b6b9783ed94b.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparison. Other methods outpainting the source video with a mask ratio of 0.6. MOTIA outpainting the source video with a larger mask ratio of 0.66 while achieving obviously better outpainting results." + ], + "image_footnote": [], + "bbox": [ + 263, + 145, + 743, + 310 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes element-wise multiplication, $\\pmb{v}_{t-1,j}$ denotes the noisy state of the $j^{th}$ frame at timestep $t$ , $\\pmb{v}_{t-1,j^*}^i$ is the noisy state of $j^{th}$ frame predicted with only information from the $i^{th}$ video clip at timestep $t$ , $W_{i,j^*}$ is the per-pixel weight, which is as 1 as default.", + "bbox": [ + 212, + 395, + 787, + 458 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 481, + 375, + 498 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 512, + 426, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmarks. To verify the effectiveness of MOTIA, we conduct evaluations on DAVIS [19] and YouTube-VOS [33], which are widely used benchmarks for video outpainting. Following M3DDM [6], we compare the results of different methods in the horizontal direction, using mask ratios of 0.25 and 0.66.", + "bbox": [ + 212, + 537, + 785, + 597 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. Our evaluation approach utilizes four well-established metrics: Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [31], Learned Perceptual Image Patch Similarity (LPIPS) [37], and Frechet Video Distance (FVD) [27]. For assessing PSNR, SSIM, and FVD, the generated videos are converted into frames within a normalized value range of $[0, 1]$ . LPIPS is evaluated over a range of $[-1, 1]$ . About the FVD metric, we adopt a uniform frame sampling, with 16 frames per video for evaluation following M3DDM.", + "bbox": [ + 212, + 598, + 785, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. Our method is built upon Stable Diffusion v1-5. We add the ControlNet pretrained on image inpainting to enable the model to accept additional masked image inputs. The temporal modules are initialized with the weights from pretrained motion modules [8] to obtain additional motion priors. The motion modules are naive transformer blocks trained with solely text-to-video tasks on WebVid. For the input-specific adaptation, the low-rank adapters are trained using the Adam optimizer. We set the learning rate to $10^{-4}$ , and set the weight decay to $10^{-2}$ . The LoRA rank and $\\alpha_{\\mathrm{lora}}$ are set to 16 and,", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Fu-Yun Wang et al.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c5cfdf1bd461884d3d45c3494161373a26d177fefeede2f936b84e55171001e5.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of video outpainting methods on DAVIS and YouTube-VOS datasets. $\\uparrow$ means 'better when higher', and $\\downarrow$ indicates 'better when lower'." + ], + "table_footnote": [], + "table_body": "
MethodDAVISYouTube-VOS
PSNR↑ SSIM↑ LPIPS↓ FVD↓PSNR↑ SSIM↑ LPIPS↓ FVD
VideoOutpainting [5]17.960.62720.2331363.118.250.71950.2278149.7
SDM [6]20.020.70780.2165334.619.910.72770.200194.81
M3DDM [6]20.260.70820.2026300.020.200.73120.185466.62
MOTIA20.360.75780.1595286.320.250.76360.172758.99
", + "bbox": [ + 241, + 198, + 756, + 276 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "8, respectively. The number of training steps is set to 1000. We do not apply augmentation for simplicity. For both mask ratios of 0.66 and 0.25, we simply apply the same random mask strategy, which uniformly crops a square in the middle as the known regions. For the pattern-aware outpainting, the diffusion steps are set to 25 and the classifier-free guidance (CFG) [11, 23] scale is set to 7.5 and we only apply CFG at the first 15 inference steps. When adding noise regret to further improve the results, we set jump length $L = 3$ , and repeat time $M = 4$ . We only apply noise regret in the first half inference steps. Note that our method is built upon LDM, which requires text-conditional inputs. For a fair comparison and to remove the influence of the choice of text prompt, we apply Blip [13] to select the prompt automatically. We observe dozens of prompt mistakes but do not modify them to avoid man-made influence.", + "bbox": [ + 212, + 303, + 787, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Qualitative Comparison", + "text_level": 1, + "bbox": [ + 214, + 507, + 459, + 523 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 6 showcases a qualitative comparison of MOTIA against other methods. Outpainting a narrow video into a square format. MOTIA employs a mask ratio of 0.66, surpassing the 0.6 ratio utilized by other methods, and demonstrates superior performance even with this higher mask ratio. The SDM method only manages to blur the extremities of the video's background, egregiously overlooking the primary subject and resulting in the outpainting failure as previously highlighted in Fig. 5. Dehan's approach effectively outpaints the background but utterly fails to address the foreground, leading to notable distortions. In contrast, the M3DDM method adeptly handles both subject and background integration but is marred by considerable deviations in subject characteristics, such as incorrect brown coloration in the dog's fur across several frames. Our method stands out by achieving optimal results, ensuring a harmonious and consistent outpainting of both the foreground and background.", + "bbox": [ + 212, + 534, + 787, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Quantitative Comparison", + "text_level": 1, + "bbox": [ + 214, + 753, + 470, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1 summarizes the evaluation metrics of our method compared to other approaches. Our method achieves comparable results to the best method in PSNR. It shows significant improvements in video quality (SSIM), perceptual metric (LPIPS), and distribution similarity (FVD). Specifically, our SSIM, LPIPS,", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Be-Your-Outpainter", + "bbox": [ + 596, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9b3e1be9256d00d767c65f89419a7a65f45b503849d5caf919ab908ad9c514f6.jpg", + "image_caption": [ + "$SD + T$" + ], + "image_footnote": [], + "bbox": [ + 285, + 143, + 390, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6167695fb83bfd4ce1b7d9aef173f38a4c90f078a3210a414066c017fb5ffa5d.jpg", + "image_caption": [ + "$SD + T + C$" + ], + "image_footnote": [], + "bbox": [ + 395, + 145, + 500, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8f6d01ebfdc090075a8da7817199060153f0d5da6b92ea64f899dca46a7ea6d0.jpg", + "image_caption": [ + "Direct-tune", + "Input-Specific Adaptation" + ], + "image_footnote": [], + "bbox": [ + 504, + 145, + 607, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1c63d74782cf2f2447a94af9ad250041469872d1597d57522ce75ed9ae2898ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 145, + 718, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/87b4dad41111d6cb1172867e35621992308b67e104485b01d3d236c9a7cd623b.jpg", + "image_caption": [ + "Fig. 7: Visual examples of ablation study on the proposed input-specific adaptation.", + "Direct (Ours $w / o$ SA + NR)", + "SA (Ours w/o NR)", + "Pattern-Aware Outpainting", + "Fig. 8: Visual examples of ablation study on pattern-aware outpainting." + ], + "image_footnote": [], + "bbox": [ + 274, + 286, + 727, + 345 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and FVD metrics show improvements of $7.00\\%$ , $21.27\\%$ , and $4.57\\%$ respectively on the DAVIS dataset, and $4.43\\%$ , $6.85\\%$ , and $11.45\\%$ on the YouTube-VOS dataset compared to the best-performing method.", + "bbox": [ + 212, + 415, + 784, + 460 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 483, + 388, + 498 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation study on input-specific adaptation. We conducted the ablation study on input-specific adaptation with the DAVIS dataset to verify its effectiveness, as shown in Fig. 7 and Table 2. \"SD+T\" represents the result of directly combining the temporal module with Stable Diffusion, which led to a complete outpainting failure. \"SD+T+C\" indicates the additional use of ControlNet, resulting in similarly poor outcomes. \"Direct-tune\" refers to the approach of directly fitting the original video without outpainting training; in this case, we observed a very noticeable color discrepancy between the outpainted and known areas. In contrast, our method achieved the best results, ensuring consistency in both the visual and temporal aspects. The metrics shown in Table 2 also support this observation, with MOTIA significantly outperforming the other baselines.", + "bbox": [ + 212, + 507, + 784, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation study on pattern-aware outpainting. Table 3 summarizes our ablation experiments for the pattern-aware outpainting part. We conducted extensive validation on the YouTube-VOS dataset. \"Direct\" refers to performing outpainting according to Eq. 3 directly after input-specific adaptation. \"SA\" denotes spatially-aware insertion, and \"SA+NR\" indicates the combined use of spatially-aware insertion and noise regret. The experimental results demonstrate that each of our components effectively enhances performance. Specifically, Combining both SA-Insertion and Noise regret, the PSNR, SSIM, LPIPS, and FVD metrics show improvements of $2.69\\%$ , $0.90\\%$ , $3.95\\%$ , and $11.32\\%$ respectively than directly applying Eq. 3. Fig. 8 presents the visual examples of ablation study on our proposed pattern-aware outpainting part. When removing NR, it", + "bbox": [ + 212, + 674, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Fu-Yun Wang et al.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3dbf9997f164bd3b9566408a353b3c41c5c61feab2a7c2b1fade71e1ae958b3c.jpg", + "table_caption": [ + "Table 2: Ablation study on input-specific adaptation." + ], + "table_footnote": [], + "table_body": "
ComponentPSNR↑SSIM↑LPIPS↓FVD ↓
SD+T15.590.66400.2979672.7
SD+T+C16.810.69610.2338515.4
Direct-tune19.420.73750.1784312.1
MOTIA20.360.75780.1595286.3
", + "bbox": [ + 215, + 184, + 504, + 272 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/218302381bb7b0dff041a475caba83f74ddb2d6bfc24988b0d09ad46d7daff22.jpg", + "table_caption": [ + "Table 3: Ablation study on the proposed pattern-aware outpainting." + ], + "table_footnote": [], + "table_body": "
ComponentPSNR↑SSIM↑LPIPS↓FVD ↓
Direct19.720.75680.179866.52
SA19.970.76080.175258.40
SA+NR20.250.76360.172758.99
", + "bbox": [ + 509, + 184, + 799, + 258 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "might fail to align the texture colors or produce unreasonable details (e.g., arms in the middle of Fig. 8). When further removing SA, it could potentially generate unrealistic results caused by the overfitting to the target video (e.g., the white collar on the left of Fig. 8). Note that even though the FVD degrades in a very slight manner, all the other metrics increase and we qualitatively find it to be helpful for improving results.", + "bbox": [ + 212, + 297, + 787, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Discussions", + "text_level": 1, + "bbox": [ + 215, + 409, + 356, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Model and computation complexity. Model Complexity: The original model has 1.79 billion (including the auto-encoder and text encoder) parameters in total, while the added adapters contain 7.49 million parameters, leading to an increase of $0.42\\%$ in memory usage. Computation Complexity: We report the peak GPU VRAM and the time required for outpainting a target video from $512 \\times 512$ to $512 \\times 1024$ with 16 frames at two stages in Table 4. For longer videos, as described in Section 4.5, instead of processing the long video as a whole, we adapt only to short video clips sampled from the long video. This approach does not require additional time or GPU VRAM during the input-specific adaptation phase. Additionally, with temporal co-denoising [29], the GPU VRAM usage remains the same as that for short video during the pattern-aware outpainting phase, while the required time increases linearly with the video length.", + "bbox": [ + 212, + 431, + 787, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "User study. We conducted a user study between MOTIA and M3DDM, utilizing the DAVIS dataset with a horizontal mask of 0.66 as source videos. Preferences were collected from 10 volunteers, each evaluating 50 randomly selected sets of results based on visual quality (such as clarity, color fidelity, and texture detail) and realism (including motion consistency, object continuity, and integration with the background). Table 2 demonstrates that the outputs from MOTIA are preferred over those from M3DDM in both visual quality and realism.", + "bbox": [ + 212, + 613, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Why MOTIA outperforms (Why previous methods fail). 1) Flexibility. Current video diffusion models are mostly trained with fixed resolution and length, lacking the ability to tackle videos with various aspect ratios and lengths. In contrast, the adaptation phase of MOTIA allows the model to better capture the size, length, and style distribution of the source video, greatly narrowing the gap between pretrained weights and the source video. 2) Ability for capturing intrinsic patterns from source video. A crucial point for successful outpainting is the predicted score of diffusion models should be well-compatible with the", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Be-Your-Outpainter", + "bbox": [ + 596, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b4d8191f70eaa241395c039dff944f17d932e5b42568860d49c2fd51c9ae39af.jpg", + "table_caption": [ + "Table 4: Computation complexity." + ], + "table_footnote": [], + "table_body": "
PhaseVRAM↓Time↓
Adaptation12.70 GB309 Seconds
Outpainting5.80 GB58 Seconds
MOTIA (In total)12.70 GB367 Seconds
", + "bbox": [ + 215, + 171, + 491, + 250 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/324fe69e918e2419ab7ced5b623bdb634c452f1a3d3ff48637db352969c55ed0.jpg", + "table_caption": [ + "Table 5: User study comparison between M3DDM and MOTIA." + ], + "table_footnote": [], + "table_body": "
MethodVisual-QualityRealism
M3DDM27.4%42.8%
MOTIA72.6%57.2%
", + "bbox": [ + 545, + 184, + 767, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "original known regions of the source video. To achieve this, the model should effectively extract useful information from the source video for denoising. For instance, M3DDM concatenates local frames of source video at the input layers and incorporates the global frames through the cross-attention mechanism after passing light encoders. However, the information might not be properly handled especially for out-domain inputs, thus leading to outpainting failure. Instead, by conducting input-specific adaptation on the source video, the model can effectively capture the data-specific patterns in the source videos through gradient. Through this, MOTIA better leverage the data-specific patterns of the source video and image/video generative prior for outpainting. We hope this work inspires following research to exploit more from the source video itself instead of purely relying on the generative prior from intensive training on videos.", + "bbox": [ + 212, + 279, + 787, + 462 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 488, + 357, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present MOTIA, relying on a combination of input-specific adaptation for capturing inner video patterns and pattern-aware outpainting to generalize these patterns for effective outpainting. Extensive experiments validate the effectiveness.", + "bbox": [ + 212, + 523, + 785, + 584 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations: MOTIA requires learning necessary patterns from the source video, it faces limitation for outpainting videos with limited information. We provide an example of failure case. In Fig. 9, the original source video just contains grass and a woman with a blue shirt. We find our method produces a woman with very similar appearance when outpainting, which might be regarded as a \"copy\" phenomenon.", + "bbox": [ + 212, + 585, + 787, + 676 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/88f31870890111efaa882087367b78e385d684c45c3eb90966c0a182986cd2bc.jpg", + "image_caption": [ + "Fig. 9: Example of failure case. Contents outside the red lines are outpainted." + ], + "image_footnote": [], + "bbox": [ + 287, + 704, + 426, + 811 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/06eb5ae5a9ca15710ed566732a9752e4649bf0cea7ad81e32b85908cf56b775c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 704, + 570, + 811 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a52662512af143c2459994acbfb8d0c2101461270dd2a98ec61b0d9992497269.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 704, + 714, + 811 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Fu-Yun Wang et al.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment:", + "text_level": 1, + "bbox": [ + 217, + 143, + 351, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR, by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK.", + "bbox": [ + 215, + 176, + 785, + 252 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Be-Your-Outpainter", + "bbox": [ + 596, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arora, R., Lee, Y.J.: Singan-gif: Learning a generative video model from a single gif. In: CVPR. pp. 1310-1319 (2021)", + "2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: CVPR. pp. 18208-18218 (2022)", + "3. Barnes, C., Shechtman, E., Finkelstein, A., Goldman, D.B.: Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph. 28(3), 24 (2009)", + "4. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR. pp. 22563-22575 (2023)", + "5. Dehan, L., Van Ranst, W., Vandewalle, P., Goedemé, T.: Complete and temporally consistent video outpainting. In: CVPRW. pp. 687-695 (2022)", + "6. Fan, F., Guo, C., Gong, L., Wang, B., Ge, T., Jiang, Y., Luo, C., Zhan, J.: Hierarchical masked 3d diffusion model for video outpainting. In: ACM MM. pp. 7890-7900 (2023)", + "7. Gao, C., Saraf, A., Huang, J.B., Kopf, J.: Flow-edge guided video completion. In: ECCV. pp. 713-729. Springer (2020)", + "8. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: Animatediff: animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023)", + "9. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020)", + "1. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "2. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. arXiv:2204.03458 (2022)", + "3. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900. PMLR (2022)", + "4. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR. pp. 11461-11471 (2022)", + "5. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186 (2023)", + "6. Ma, Y., He, Y., Wang, H., Wang, A., Qi, C., Cai, C., Li, X., Li, Z., Shum, H.Y., Liu, W., et al.: Follow-your-click: Open-domain regional image animation via short prompts. arXiv preprint arXiv:2403.08268 (2024)", + "7. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "8. Nikankin, Y., Haim, N., Irani, M.: Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743 (2022)", + "9. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Fu-Yun Wang et al.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)", + "21. Rout, L., Parulekar, A., Caramanis, C., Shakkottai, S.: A theoretical justification for image inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2302.01217 (2023)", + "22. Shaham, T.R., Dekel, T., Michaeli, T.: Singan: Learning a generative model from a single natural image. In: ICCV. pp. 4570-4580 (2019)", + "23. Shen, D., Song, G., Xue, Z., Wang, F.Y., Liu, Y.: Rethinking the spatial inconsistency in classifier-free diffusion guidance. In: CVPR. pp. 9370-9379 (June 2024)", + "24. Shi, X., Huang, Z., Wang, F.Y., Bian, W., Li, D., Zhang, Y., Zhang, M., Cheung, K.C., See, S., Qin, H., et al.: Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In: ACM SIGGRAPH 2024 Conference Papers. pp. 1-11 (2024)", + "25. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)", + "26. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "27. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)", + "28. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. NeurIPS (2022)", + "29. Wang, F.Y., Chen, W., Song, G., Ye, H.J., Liu, Y., Li, H.: Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264 (2023)", + "30. Wang, F.Y., Huang, Z., Shi, X., Bian, W., Song, G., Liu, Y., Li, H.: Animatelem: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning. arXiv preprint arXiv:2402.00769 (2024)", + "31. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "32. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565 (2022)", + "33. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)", + "34. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In: CVPR. pp. 1316-1324 (2018)", + "35. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: CVPR. pp. 10459-10469 (2023)", + "36. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)", + "37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 813 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Be-Your-Outpainter", + "bbox": [ + 596, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_model.json b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4a8996087cf2fe240a1640d680712a6ee5a040ee --- /dev/null +++ b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_model.json @@ -0,0 +1,2247 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.232, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Be-Your-Outpainter: Mastering Video Outpainting through Input-Specific Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.212, + 0.703, + 0.228 + ], + "angle": 0, + "content": "Fu-Yun Wang\\(^{1}\\) Xiaoshi Wu\\(^{1}\\) Zhaoyang Huang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.228, + 0.702, + 0.244 + ], + "angle": 0, + "content": "Xiaoyu Shi\\(^{1}\\) Dazhong Shen\\(^{3}\\) Guanglu Song\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.244, + 0.612, + 0.258 + ], + "angle": 0, + "content": "Yu Liu\\(^{3}\\) Hongsheng Li\\(^{1,3,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.269, + 0.616, + 0.284 + ], + "angle": 0, + "content": "\\(^{1}\\)MMLab, CUHK \\(^{2}\\)Avolution AI" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.284, + 0.733, + 0.299 + ], + "angle": 0, + "content": "\\(^{3}\\)Shanghai AI Lab \\(^{4}\\)SenseTime Research \\(^{5}\\)CPII under InnoHK" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.3, + 0.635, + 0.312 + ], + "angle": 0, + "content": "{fywang@link, hsli@ee}.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.345, + 0.744, + 0.651 + ], + "angle": 0, + "content": "Abstract. Video outpainting is a challenging task, aiming at generating video content outside the viewpoint of the input video while maintaining inter-frame and intra-frame consistency. Existing methods fall short in either generation quality or flexibility. We introduce MOTIA (Mastering Video Outpainting Through Input-Specific Adaptation), a diffusion-based pipeline that leverages both the intrinsic data-specific patterns of the source video and the image/video generative prior for effective outpainting. MOTIA comprises two main phases: input-specific adaptation and pattern-aware outpainting. The input-specific adaptation phase involves conducting efficient and effective pseudo outpainting learning on the single-shot source video. This process encourages the model to identify and learn patterns within the source video, as well as bridging the gap between standard generative processes and outpainting. The subsequent phase, pattern-aware outpainting, is dedicated to the generalization of these learned patterns to generate outpainting outcomes. Additional strategies are proposed to better leverage the diffusion model's generative prior and the acquired video patterns from source videos for inference. Extensive evaluations underscore MOTIA's superiority, outperforming existing state-of-the-art methods in widely recognized benchmarks. Notably, these advancements are achieved without necessitating extensive, task-specific tuning. More details are available at https://be-your-outpainter.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Video outpainting aims to expand the visual content out of the spatial boundaries of videos, which has important real-world applications [3,5,6]. For instance, in practice, videos are usually recorded with a fixed aspect ratio, such as in movies or short clips. This becomes an issue when viewing these videos on smartphones, which often have varying aspect ratios, resulting in unsightly black bars on the screen that detract from the viewing experience. Proper ways for video outpainting are crucial in solving this issue. By expanding the visual content beyond the original frame, it adapts the video to fit various screen sizes seamlessly. This process ensures that the audience enjoys a full-screen experience without any" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "Fu-Yun Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.145, + 0.425, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.144, + 0.756, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Fig. 1: MOTIA is a high-quality flexible video outpainting pipeline, leveraging the intrinsic data-specific patterns of source videos and image/video generative prior for state-of-the-art performance. Quantitative metric improvement of MOTIA is significant (Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.61 + ], + "angle": 0, + "content": "compromise in visual integrity. However, the challenges associated with video outpainting are significant. It requires not just the expansion of each frame's content but also the preservation of temporal (inter-frame) and spatial (intra-frame) consistency across the video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Currently, there are two primary approaches to video outpainting. The first employs optical flows and specialized warping techniques to extend video frames, involving complex computations and carefully tailored hyperparameters to ensure the added content remains consistent [5, 7]. However, their results are far from satisfactory, suffering from blurred content. The other type of approach in video outpainting revolves around training specialized models tailored for video inpainting and outpainting with extensive datasets [6, 35]. However, they have two notable limitations: 1) An obvious drawback of these models is their dependency on the types of masks and the resolutions of videos they can handle, which significantly constrains their versatility and effectiveness in real-world applications, as they may not be adequately equipped to deal with the diverse range of video formats and resolutions commonly encountered in practical scenarios. 2) The other drawback is their inability to out-domain video outpainting, even intensively trained on massive video data. Fig. 5 shows a failure example of most advanced previous work [6] that the model faces complete outpainting failure," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be-Your-Outpainter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "with only blurred corners. We show that a crucial reason behind this is that the model fails at capturing the intrinsic data-specific patterns from out-domain source (input) videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.333 + ], + "angle": 0, + "content": "In this work, we propose MOTIA: Mastering Video Outpainting Through Input-Specific Adaptation, a diffusion-based method for open-domain video outpainting with arbitrary types of mask, arbitrary video resolutions and lengths, and arbitrary styles. At the heart of MOTIA is treating the source video itself as a rich source of information [18, 22], which contains key motion and content patterns (intrinsic data-specific patterns) necessary for effective outpainting. We demonstrate that the patterns learned from the source video, combined with the generative capabilities of diffusion models, can achieve surprisingly great outpainting performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.79, + 0.7 + ], + "angle": 0, + "content": "Fig. 2 illustrates the workflow of MOTIA. MOTIA consists of two stages: input-specific adaptation and pattern-aware outpainting. During the input-specific adaptation stage, we conduct pseudo video outpainting learning on the source video (videos to be outpainted) itself. Specifically, at each iteration, we heuristically add random masks to the source video and prompt the base diffusion model to recover the masked regions by learning to denoise the video corrupted by white noise, relying on the extracted information from unmasked regions. This process not only allows the model to capture essential data-specific patterns from the source video but also narrows the gap between standard generation and outpainting. We insert trainable lightweight adapters into the diffusion model for tuning to keep the efficiency and stability. In the pattern-aware outpainting stage, we combine the learned patterns from the source video and the generation prior of the diffusion model for effective outpainting. To better leverage the generation ability of the pretrained diffusion model and the learned pattern from the source video, we propose spatial-aware insertion (SA-Insertion) of the tuned adapters for outpainting. Specifically, the insertion weights of adapters gradually decay as the spatial position of features away from the known regions. In this way, the outpainting of pixels near the known regions is more influenced by the learned patterns, while the outpainting of pixels far from the known regions relies more on the original generative prior of diffusion model. To further mitigate potential denoising conflicts and enhance the knowledge transfer between known regions and unknown regions, we incorporate noise regret that we add noise and denoise periodically at early inference steps, which works for more harmonious outpainting results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Extensively quantitative and qualitative experiments verify the effectiveness of our proposed method. MOTIA overcomes many limitations of previous methods and outperforms the state-of-the-art intensively trained outpainting method in standard widely used benchmarks. In summary, our contribution is three-fold: 1) We show that the data-specific patterns of source videos are crucial for effective outpainting, which is neglected by previous work. 2) We introduce an adaptation strategy to effectively capture the data-specific patterns and then propose novel strategies to better leverage the captured patterns and pretrained image/video generative prior for better results. 3) Vast experiments verify that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "Fu-Yun Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.144, + 0.756, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.344, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Fig. 2: Workflow of MOTIA. Blue lines represent the workflow of input-specific adaptation, and green lines represent the workflow of pattern-aware outpainting." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.404, + 0.785, + 0.434 + ], + "angle": 0, + "content": "our performance in video outpainting is great, significantly outperforming previous state-of-the-art methods in both quantitative metrics and user studies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.459, + 0.396, + 0.475 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.767, + 0.613 + ], + "angle": 0, + "content": "In this section, we discuss related diffusion models and outpainting methods. Diffusion models. Diffusion models (a.k.a., score-based models) [9,10,17,20,21] have gained increasing attention due to its amazing ability to generate highly detailed images. Current video diffusion models [4, 9, 15, 16, 24, 25, 28, 30] are generally built upon the extension of image diffusion models through insertion of temporal layers. They are either trained with image-video joint tuning [12, 22] or trained with spatial weights frozen [4] to mitigate the negative influence of the poor captions and visual quality of video data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Outpainting methods. Video outpainting is largely built upon the advancements in image outpainting, where techniques ranged from patch-based methods (e.g., PatchMatch [3]) to more recent deep learning approaches like GANs [1,34]. Diffusion models [2,14], benefiting from the learned priors on synthesis tasks and the process of iterative refinement, achieve state-of-the-art performance on image outpainting tasks. The research focusing on video outpainting is comparatively few. Previous works typically apply optical flow for outpainting, which warps content from adjacent frames to the outside corners, but their results are far from satisfactory. Recently, M3DDM [6] trained a large 3D diffusion models with specially designed architecture for outpainting on massive video data, achieving much better results compared to previous methods, showcasing the huge potential of diffusion models on video outpainting. However, as we claimed, they have two main limitations: 1) The inflexibility for mask types and video resolutions. They can only outpaint video with resolution \\(256 \\times 256\\) with square type of masking. 2) Inability for out-domain video outpainting. As shown in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be-Your-Outpainter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Fig. 5, they encounter outpainting failure when processing out-domain videos even intensively trained on massive video data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.205, + 0.383, + 0.22 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.241, + 0.784, + 0.272 + ], + "angle": 0, + "content": "Diffusion models [10] add noise to data through a Markov chain process. Given initial data \\( \\pmb{x}_0 \\sim q(\\pmb{x}_0) \\), this process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.285, + 0.786, + 0.327 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {x} _ {1: T} \\mid \\boldsymbol {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right), \\quad q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.338, + 0.784, + 0.369 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) is the noise schedule and \\(\\alpha_{t} = 1 - \\beta_{t}\\). The data reconstruction, or denoising process, is accomplished by the reverse transition modeled by \\(p_{\\theta}(\\boldsymbol{x}_{t-1}|\\boldsymbol{x}_t)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.385, + 0.786, + 0.403 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1}; \\tilde {\\boldsymbol {\\mu}} _ {t} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.41, + 0.784, + 0.444 + ], + "angle": 0, + "content": "with \\(\\tilde{\\pmb{\\mu}}_t(\\pmb{x}_t,\\pmb{x}_0) = \\frac{1}{\\sqrt{\\alpha_t}}\\pmb{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t\\sqrt{\\alpha_t}}}\\pmb{\\epsilon}, \\bar{\\alpha}_t = \\prod_{s=1}^t\\alpha_s, \\tilde{\\beta}_t = \\frac{1 - \\bar{\\alpha}_{t-1}}{1 - \\bar{\\alpha}_t}\\beta_t\\), and \\(\\pmb{\\epsilon}\\) is the noise added to \\(\\pmb{x}_0\\) to obtain \\(\\pmb{x}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.445, + 0.785, + 0.567 + ], + "angle": 0, + "content": "Diffusion-based outpainting aims to predict missing pixels at the corners of the masked region with the pre-trained diffusion models. We denote the ground truth as \\( \\pmb{x} \\), mask as \\( \\pmb{m} \\), known region as \\( (1 - m) \\odot \\pmb{x} \\), and unknown region as \\( m \\odot x \\). At each reverse step in the denoising process, we modify the known regions by incorporating the intermediate noisy state of the source data from the corresponding timestep in the forward diffusion process (which adds noise), provided that this maintains the correct distribution of correspondences. Specifically, each reverse step can be denoted as the following formulas:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.58, + 0.786, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t - 1} ^ {\\text {k n o w n}} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\quad \\boldsymbol {x} _ {t - 1} ^ {\\text {u n k n o w n}} \\sim \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {\\theta} (\\boldsymbol {x} _ {t}, t), \\Sigma_ {\\theta} (\\boldsymbol {x} _ {t}, t)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.629, + 0.786, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t - 1} = \\boldsymbol {m} \\odot \\boldsymbol {x} _ {t - 1} ^ {\\text {k n o w n}} + (\\boldsymbol {1} - \\boldsymbol {m}) \\odot \\boldsymbol {x} _ {t - 1} ^ {\\text {u n k n o w n}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.656, + 0.784, + 0.672 + ], + "angle": 0, + "content": "where the \\( \\pmb{x}_{t - 1}^{\\mathrm{known}} \\) is padded to the target resolution before the masked merging." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.698, + 0.381, + 0.716 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.784, + 0.84 + ], + "angle": 0, + "content": "This section presents MOTIA, a method demonstrating exceptional performance in video outpainting tasks. We begin by defining the concept of video outpainting and describing the foundational model in Section 4.1. and Section 4.2. Subsequently, we delve into the specifics of input-specific adaptation and pattern-aware outpainting in Sections 4.3 and 4.4, respectively. Moreover, we show that our approach has great promise in extending its application to long video outpainting, which will be explored in Section 4.5." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "Fu-Yun Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.761, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Fig. 3: Sample results of quantitative experiments. All videos are outpainted with a horizontal mask ratio of 0.66. Contents outside the yellow lines are outpainted by MOTIA." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.439, + 0.587 + ], + "angle": 0, + "content": "4.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.704 + ], + "angle": 0, + "content": "For a video represented as \\(\\pmb{v} \\in \\mathbb{R}^{t \\times h \\times w \\times d}\\), where \\(t\\) denotes the number of frames, \\(h\\) denotes frame height, \\(w\\) denotes frame width, and \\(d\\) denotes channel depth. Video outpainting model \\(f(\\pmb{v})\\) is designed to generate a spatially expanded video \\(\\pmb{v}' \\in \\mathbb{R}^{t \\times h' \\times w' \\times d}\\). This process not only increases the spatial dimensions (\\(h' > h\\), \\(w' > w\\)), but also requires to ensure continuity and harmony between the newly expanded regions and the known regions. The transformation maintains the known regions unchanged, with \\(f(\\pmb{v})\\) acting as an identity in these regions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.424, + 0.74 + ], + "angle": 0, + "content": "4.2 Network Expansion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.796, + 0.841 + ], + "angle": 0, + "content": "Network inflation. MOTIA leverages the pre-trained text-to-image (T2I) model, Stable Diffusion. In line with previous video editing techniques [32], we transform 2D convolutions into pseudo 3D convolutions and adapt 2D group normalizations into 3D group normalizations to process video latent features. Specifically, the \\(3 \\times 3\\) kernels in convolutions are replaced by \\(1 \\times 3 \\times 3\\) kernels, maintaining identical weights. Group normalizations are executed across both temporal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be-Your-Outpainter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "and spatial dimensions, meaning that all 3D features within the same group are normalized simultaneously, followed by scaling and shifting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.268 + ], + "angle": 0, + "content": "Masked video as conditional input. Additionally, we incorporate a ControlNet [36], initially trained for image inpainting, to manage additional mask inputs. Apart from noise input, ControlNet can also process masked videos to extract effective information for more controllable denoising. In these masked videos, known regions have pixel values ranging from 0 to 1, while values of masked regions are set to \\(-1\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.391 + ], + "angle": 0, + "content": "Temporal consistency prior. To infuse the model with temporal consistency priors, we integrate temporal modules pre-trained on text-to-video (T2V) tasks. Note that although MOTIA relies on pre-trained video diffusion modules, applying these pre-trained temporal modules directly for video outpainting yields rather bad results, significantly inferior to all baseline methods (Table. 2). However, when equipped with our proposed MOTIA, the model demonstrates superior performance even in comparison to models specifically designed and trained for video outpainting, underscoring the efficacy of MOTIA." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.416, + 0.476, + 0.431 + ], + "angle": 0, + "content": "4.3 Input-Specific Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.784, + 0.534 + ], + "angle": 0, + "content": "The input-specific adaptation phase is crucial in our video outpainting method, aiming to tailor the model for the specific challenges of outpainting. This phase involves training on the source video with a pseudo-outpainting task, importantly, enabling the model to learn intrinsic content and motion patterns (data-specific patterns) within the source video as well as narrowing the gap between the standard generation process and outpainting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.786, + 0.626 + ], + "angle": 0, + "content": "Video augmentation. Initially, we augment the source video. Transformations like identity transformation, random flipping, cropping, and resizing can be employed. This step can potentially help the model better learn and adapt to diverse changes in video content. For longer video outpainting, as we will discuss later, instead of taking it as a whole, we randomly sample short video clips from it to reduce the cost of the adaptation phase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.786, + 0.716 + ], + "angle": 0, + "content": "Video masking. We then add random masks to the video. We adopt a heuristic approach that uniformly samples edge boundaries of 4 sides within given limits. The area enclosed by these boundaries is considered the known region, while the rest is the unknown region. This masked video serves as the conditional input for the ControlNet, simulating the distribution of known and unknown areas in actual outpainting scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.795, + 0.762 + ], + "angle": 0, + "content": "Video noising. Additionally, we apply noise to the video following the DDPM [10] by randomly selecting diffusion timesteps. This noisy video serves as an input for both the ControlNet and the Stable Diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.787, + 0.822 + ], + "angle": 0, + "content": "Optimization. Finally, we optimize the model. To ensure efficiency, low-rank adapters are inserted into the layers of the diffusion model. We optimize only the parameters of these adapters while keeping the other parameters frozen. The loss function is" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.824, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\| \\boldsymbol {\\epsilon} - \\hat {\\boldsymbol {\\epsilon}} _ {\\bar {\\theta} _ {l}, \\bar {\\theta} _ {c}, \\boldsymbol {\\theta} _ {a}} (\\boldsymbol {v} _ {\\mathrm {n o i s y}}, \\boldsymbol {v} _ {\\mathrm {m a s k e d}}, t) \\right\\| _ {2}, \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "Fu-Yun Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.17, + 0.432, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.263, + 0.436, + 0.332 + ], + "angle": 0, + "content": "Fig.4: Spatial-aware insertion scales the insertion weights of adapters for better leveraging of learned patterns and generative prior." + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.146, + 0.787, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.262, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Fig. 5: Failure example of previous methods. Many previous methods including the intensively trained models on video outpainting still might suffer from generation failure, that the model simply generates blurred corners. MOTIA never encounters this failure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.487 + ], + "angle": 0, + "content": "where \\(t\\) represents the timestep in the process, \\(\\epsilon\\) is the added noise, \\(\\pmb{v}_{\\mathrm{noisy}}\\) refers to the video perturbed by \\(\\epsilon\\), and \\(\\pmb{v}_{\\mathrm{masked}}\\) denotes the masked video. The parameters \\(\\theta_{l}\\), \\(\\theta_{c}\\), and \\(\\theta_{a}\\) correspond to the Diffusion Model, ControlNet, and adapters, respectively. The bar over these parameters indicates they are frozen during the optimization. This optimization process, including the steps of augmentation, masking, and noising, is repeated to update the lightweight adapters to capture the data-specific patterns from the source video." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.517, + 0.49, + 0.534 + ], + "angle": 0, + "content": "4.4 Pattern-Aware Outpainting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Following the initial phase of input-specific adaptation, our model shows promising results in video outpainting using basic pipelines as outlined in Eq. 3 and Eq. 4, achieving commendable quality. However, we here introduce additional inference strategies that can be combined to better leverage the learned data-specific patterns from the input-specific adaptation phase for better outpainting results. We call the outpainting process that incorporates these strategies pattern-aware outpainting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Spatial-aware insertion. It is important to acknowledge that in the input-specific adaptation phase, the model is fine-tuned through learning outpainting within the source video. However, at the outpainting phase, the model is expected to treat the entire source video as known regions and then fill the unknown regions at edges (i.e., generating a video with a larger viewpoint and resolution). This specificity may lead to a noticeable training-inference gap during outpainting, potentially affecting the outpainting quality. To balance the fine-tuned patterns with the diffusion model's inherent generative prior, we introduce the concept of spatial-aware insertion (SA-Insertion) of adapters as shown in Fig. 4. The adaptation involves adjusting the insertion weight of tuned low-rank adapters based on the feature's spatial position. We increase insertion weight near known areas to utilize the learned patterns while decreasing it in farther regions to rely" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be-Your-Outpainter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.78, + 0.162 + ], + "angle": 0, + "content": "more on the original generative capacity of the diffusion model. To be specific," + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.173, + 0.786, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {\\mathrm {a d a p t e d}} ^ {\\top} \\boldsymbol {x} _ {\\boldsymbol {p}} = \\mathbf {W} ^ {\\top} \\boldsymbol {x} _ {\\boldsymbol {p}} + \\alpha (\\boldsymbol {p}) \\left(\\mathbf {W} _ {\\mathrm {u p}} \\mathbf {W} _ {\\mathrm {d o w n}}\\right) ^ {\\top} \\boldsymbol {x} _ {\\boldsymbol {p}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.205, + 0.784, + 0.267 + ], + "angle": 0, + "content": "Here, \\(\\pmb{p}\\) signifies the spatial position of \\(\\pmb{x}\\), \\(\\mathbf{W} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times d_{\\mathrm{out}}}\\) denotes the linear transformation in layers of diffusion model, \\(\\mathbf{W}_{\\mathrm{down}} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times r}\\) and \\(\\mathbf{W}_{\\mathrm{up}} \\in \\mathbb{R}^{r \\times d_{\\mathrm{out}}}\\) are the linear components of the adapter with rank \\(r \\ll \\min(d_{\\mathrm{in}}, d_{\\mathrm{out}})\\). The function \\(\\alpha(\\pmb{p})\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.279, + 0.786, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\alpha (\\boldsymbol {p}) = \\exp \\left(- \\frac {K \\| \\boldsymbol {p} - \\boldsymbol {p} _ {c} \\|}{\\operatorname* {m a x} _ {\\bar {\\boldsymbol {p}}} \\| \\bar {\\boldsymbol {p}} - \\boldsymbol {p} _ {c} \\|}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.322, + 0.784, + 0.351 + ], + "angle": 0, + "content": "where \\(K\\) is a constant for controlling decay speed, and \\(\\pmb{p}_c\\) represents the nearest side of the known region to \\(\\pmb{p}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.353, + 0.785, + 0.501 + ], + "angle": 0, + "content": "Noise regret. In the context of Eq. 3, merging noisy states from known and unknown regions in video outpainting tasks is similar to sampling from two different vectors, which can potentially disrupt the denoising direction. Given the significant impact of early steps on the generation's structure, later denoising may not rectify these initial discrepancies. Inspired by DDPM-based image inpainting methods [14,21], we propose to re-propagate the noisy state into a noisier state by adding noise when denoising and then provide the model a second chance for re-denoising. This helps integrate known region data more effectively and reduces denoising direction conflicts. In detail, after obtaining \\(\\mathbf{v}_t\\) during denoising, we conduct" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.502, + 0.786, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {t + L} = \\sqrt {\\Pi_ {i = t + 1} ^ {t + L} \\alpha_ {i}} \\boldsymbol {v} _ {t} + \\sqrt {1 - \\Pi_ {i = t + 1} ^ {t + L} \\alpha_ {i}} \\boldsymbol {\\epsilon}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.534, + 0.784, + 0.565 + ], + "angle": 0, + "content": "where \\(\\alpha_{i} = 1 - \\beta_{i}\\) and \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). Then we restart the denoising process. We repeat this progress for \\(M\\) times. We only conduct it in the early denoising steps." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.573, + 0.604 + ], + "angle": 0, + "content": "4.5 Extension to Long Video Outpainting" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.616, + 0.784, + 0.796 + ], + "angle": 0, + "content": "We show that our method can be easily extended for long video outpainting. Specifically, for the stage of input-specific adaptation, instead of taking the long video as a whole for adaptation (Direct adaptation on long videos is costly and does not align with the video generation prior of the pretrained modules), we randomly sample short video clips from the long video for tuning to learn global patterns without requiring more GPU memory cost. For the stage of pattern-aware outpainting, we split the long video into short video clips with temporal overlapping (i.e., some frames are shared by different short video clips), and then conduct temporal co-denoising following Gen-L [29]. Specifically, the denoising result for \\( j^{th} \\) frame of the long video at timestep \\( t \\) is approximated by the weighted sum of all the corresponding frames in short video clips that contain it," + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.798, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {t - 1, j} = \\frac {\\sum_ {i \\in \\mathcal {I} ^ {j}} \\left(\\left(W _ {i , j ^ {*}}\\right) ^ {2} \\otimes \\boldsymbol {v} _ {t - 1 , j ^ {*}} ^ {i}\\right)}{\\sum_ {i \\in \\mathcal {I} ^ {j}} \\left(W _ {i , j ^ {*}} ^ {2}\\right) ^ {2}}, \\tag {9}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "Fu-Yun Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.146, + 0.744, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.324, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison. Other methods outpainting the source video with a mask ratio of 0.6. MOTIA outpainting the source video with a larger mask ratio of 0.66 while achieving obviously better outpainting results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.789, + 0.459 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes element-wise multiplication, \\(\\pmb{v}_{t-1,j}\\) denotes the noisy state of the \\(j^{th}\\) frame at timestep \\(t\\), \\(\\pmb{v}_{t-1,j^*}^i\\) is the noisy state of \\(j^{th}\\) frame predicted with only information from the \\(i^{th}\\) video clip at timestep \\(t\\), \\(W_{i,j^*}\\) is the per-pixel weight, which is as 1 as default." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.482, + 0.376, + 0.499 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.427, + 0.529 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.787, + 0.598 + ], + "angle": 0, + "content": "Benchmarks. To verify the effectiveness of MOTIA, we conduct evaluations on DAVIS [19] and YouTube-VOS [33], which are widely used benchmarks for video outpainting. Following M3DDM [6], we compare the results of different methods in the horizontal direction, using mask ratios of 0.25 and 0.66." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Evaluation metrics. Our evaluation approach utilizes four well-established metrics: Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [31], Learned Perceptual Image Patch Similarity (LPIPS) [37], and Frechet Video Distance (FVD) [27]. For assessing PSNR, SSIM, and FVD, the generated videos are converted into frames within a normalized value range of \\([0, 1]\\). LPIPS is evaluated over a range of \\([-1, 1]\\). About the FVD metric, we adopt a uniform frame sampling, with 16 frames per video for evaluation following M3DDM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Implementation details. Our method is built upon Stable Diffusion v1-5. We add the ControlNet pretrained on image inpainting to enable the model to accept additional masked image inputs. The temporal modules are initialized with the weights from pretrained motion modules [8] to obtain additional motion priors. The motion modules are naive transformer blocks trained with solely text-to-video tasks on WebVid. For the input-specific adaptation, the low-rank adapters are trained using the Adam optimizer. We set the learning rate to \\( 10^{-4} \\), and set the weight decay to \\( 10^{-2} \\). The LoRA rank and \\( \\alpha_{\\mathrm{lora}} \\) are set to 16 and," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be-Your-Outpainter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of video outpainting methods on DAVIS and YouTube-VOS datasets. \\(\\uparrow\\) means 'better when higher', and \\(\\downarrow\\) indicates 'better when lower'." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.199, + 0.758, + 0.277 + ], + "angle": 0, + "content": "
MethodDAVISYouTube-VOS
PSNR↑ SSIM↑ LPIPS↓ FVD↓PSNR↑ SSIM↑ LPIPS↓ FVD
VideoOutpainting [5]17.960.62720.2331363.118.250.71950.2278149.7
SDM [6]20.020.70780.2165334.619.910.72770.200194.81
M3DDM [6]20.260.70820.2026300.020.200.73120.185466.62
MOTIA20.360.75780.1595286.320.250.76360.172758.99
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.485 + ], + "angle": 0, + "content": "8, respectively. The number of training steps is set to 1000. We do not apply augmentation for simplicity. For both mask ratios of 0.66 and 0.25, we simply apply the same random mask strategy, which uniformly crops a square in the middle as the known regions. For the pattern-aware outpainting, the diffusion steps are set to 25 and the classifier-free guidance (CFG) [11, 23] scale is set to 7.5 and we only apply CFG at the first 15 inference steps. When adding noise regret to further improve the results, we set jump length \\( L = 3 \\), and repeat time \\( M = 4 \\). We only apply noise regret in the first half inference steps. Note that our method is built upon LDM, which requires text-conditional inputs. For a fair comparison and to remove the influence of the choice of text prompt, we apply Blip [13] to select the prompt automatically. We observe dozens of prompt mistakes but do not modify them to avoid man-made influence." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.508, + 0.46, + 0.525 + ], + "angle": 0, + "content": "5.2 Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Fig. 6 showcases a qualitative comparison of MOTIA against other methods. Outpainting a narrow video into a square format. MOTIA employs a mask ratio of 0.66, surpassing the 0.6 ratio utilized by other methods, and demonstrates superior performance even with this higher mask ratio. The SDM method only manages to blur the extremities of the video's background, egregiously overlooking the primary subject and resulting in the outpainting failure as previously highlighted in Fig. 5. Dehan's approach effectively outpaints the background but utterly fails to address the foreground, leading to notable distortions. In contrast, the M3DDM method adeptly handles both subject and background integration but is marred by considerable deviations in subject characteristics, such as incorrect brown coloration in the dog's fur across several frames. Our method stands out by achieving optimal results, ensuring a harmonious and consistent outpainting of both the foreground and background." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.754, + 0.472, + 0.77 + ], + "angle": 0, + "content": "5.3 Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Table 1 summarizes the evaluation metrics of our method compared to other approaches. Our method achieves comparable results to the best method in PSNR. It shows significant improvements in video quality (SSIM), perceptual metric (LPIPS), and distribution similarity (FVD). Specifically, our SSIM, LPIPS," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "Fu-Yun Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.145, + 0.391, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.229, + 0.357, + 0.239 + ], + "angle": 0, + "content": "\\(SD + T\\)" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.146, + 0.5, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.229, + 0.469, + 0.239 + ], + "angle": 0, + "content": "\\(SD + T + C\\)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.609, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.229, + 0.588, + 0.24 + ], + "angle": 0, + "content": "Direct-tune" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.146, + 0.72, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.229, + 0.728, + 0.24 + ], + "angle": 0, + "content": "Input-Specific Adaptation" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.255, + 0.784, + 0.269 + ], + "angle": 0, + "content": "Fig. 7: Visual examples of ablation study on the proposed input-specific adaptation." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.287, + 0.728, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.348, + 0.41, + 0.357 + ], + "angle": 0, + "content": "Direct (Ours \\( w / o \\) SA + NR)" + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.348, + 0.543, + 0.358 + ], + "angle": 0, + "content": "SA (Ours w/o NR)" + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.348, + 0.714, + 0.359 + ], + "angle": 0, + "content": "Pattern-Aware Outpainting" + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.372, + 0.748, + 0.387 + ], + "angle": 0, + "content": "Fig. 8: Visual examples of ablation study on pattern-aware outpainting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.462 + ], + "angle": 0, + "content": "and FVD metrics show improvements of \\(7.00\\%\\), \\(21.27\\%\\), and \\(4.57\\%\\) respectively on the DAVIS dataset, and \\(4.43\\%\\), \\(6.85\\%\\), and \\(11.45\\%\\) on the YouTube-VOS dataset compared to the best-performing method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.484, + 0.389, + 0.499 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.674 + ], + "angle": 0, + "content": "Ablation study on input-specific adaptation. We conducted the ablation study on input-specific adaptation with the DAVIS dataset to verify its effectiveness, as shown in Fig. 7 and Table 2. \"SD+T\" represents the result of directly combining the temporal module with Stable Diffusion, which led to a complete outpainting failure. \"SD+T+C\" indicates the additional use of ControlNet, resulting in similarly poor outcomes. \"Direct-tune\" refers to the approach of directly fitting the original video without outpainting training; in this case, we observed a very noticeable color discrepancy between the outpainted and known areas. In contrast, our method achieved the best results, ensuring consistency in both the visual and temporal aspects. The metrics shown in Table 2 also support this observation, with MOTIA significantly outperforming the other baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Ablation study on pattern-aware outpainting. Table 3 summarizes our ablation experiments for the pattern-aware outpainting part. We conducted extensive validation on the YouTube-VOS dataset. \"Direct\" refers to performing outpainting according to Eq. 3 directly after input-specific adaptation. \"SA\" denotes spatially-aware insertion, and \"SA+NR\" indicates the combined use of spatially-aware insertion and noise regret. The experimental results demonstrate that each of our components effectively enhances performance. Specifically, Combining both SA-Insertion and Noise regret, the PSNR, SSIM, LPIPS, and FVD metrics show improvements of \\(2.69\\%\\), \\(0.90\\%\\), \\(3.95\\%\\), and \\(11.32\\%\\) respectively than directly applying Eq. 3. Fig. 8 presents the visual examples of ablation study on our proposed pattern-aware outpainting part. When removing NR, it" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be-Your-Outpainter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.477, + 0.173 + ], + "angle": 0, + "content": "Table 2: Ablation study on input-specific adaptation." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.185, + 0.505, + 0.273 + ], + "angle": 0, + "content": "
ComponentPSNR↑SSIM↑LPIPS↓FVD ↓
SD+T15.590.66400.2979672.7
SD+T+C16.810.69610.2338515.4
Direct-tune19.420.73750.1784312.1
MOTIA20.360.75780.1595286.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 3: Ablation study on the proposed pattern-aware outpainting." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.185, + 0.8, + 0.26 + ], + "angle": 0, + "content": "
ComponentPSNR↑SSIM↑LPIPS↓FVD ↓
Direct19.720.75680.179866.52
SA19.970.76080.175258.40
SA+NR20.250.76360.172758.99
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.39 + ], + "angle": 0, + "content": "might fail to align the texture colors or produce unreasonable details (e.g., arms in the middle of Fig. 8). When further removing SA, it could potentially generate unrealistic results caused by the overfitting to the target video (e.g., the white collar on the left of Fig. 8). Note that even though the FVD degrades in a very slight manner, all the other metrics increase and we qualitatively find it to be helpful for improving results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.357, + 0.423 + ], + "angle": 0, + "content": "5.5 Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.789, + 0.613 + ], + "angle": 0, + "content": "Model and computation complexity. Model Complexity: The original model has 1.79 billion (including the auto-encoder and text encoder) parameters in total, while the added adapters contain 7.49 million parameters, leading to an increase of \\(0.42\\%\\) in memory usage. Computation Complexity: We report the peak GPU VRAM and the time required for outpainting a target video from \\(512 \\times 512\\) to \\(512 \\times 1024\\) with 16 frames at two stages in Table 4. For longer videos, as described in Section 4.5, instead of processing the long video as a whole, we adapt only to short video clips sampled from the long video. This approach does not require additional time or GPU VRAM during the input-specific adaptation phase. Additionally, with temporal co-denoising [29], the GPU VRAM usage remains the same as that for short video during the pattern-aware outpainting phase, while the required time increases linearly with the video length." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.719 + ], + "angle": 0, + "content": "User study. We conducted a user study between MOTIA and M3DDM, utilizing the DAVIS dataset with a horizontal mask of 0.66 as source videos. Preferences were collected from 10 volunteers, each evaluating 50 randomly selected sets of results based on visual quality (such as clarity, color fidelity, and texture detail) and realism (including motion consistency, object continuity, and integration with the background). Table 2 demonstrates that the outputs from MOTIA are preferred over those from M3DDM in both visual quality and realism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Why MOTIA outperforms (Why previous methods fail). 1) Flexibility. Current video diffusion models are mostly trained with fixed resolution and length, lacking the ability to tackle videos with various aspect ratios and lengths. In contrast, the adaptation phase of MOTIA allows the model to better capture the size, length, and style distribution of the source video, greatly narrowing the gap between pretrained weights and the source video. 2) Ability for capturing intrinsic patterns from source video. A crucial point for successful outpainting is the predicted score of diffusion models should be well-compatible with the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "Fu-Yun Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.145, + 0.475, + 0.16 + ], + "angle": 0, + "content": "Table 4: Computation complexity." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.172, + 0.493, + 0.251 + ], + "angle": 0, + "content": "
PhaseVRAM↓Time↓
Adaptation12.70 GB309 Seconds
Outpainting5.80 GB58 Seconds
MOTIA (In total)12.70 GB367 Seconds
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.145, + 0.785, + 0.172 + ], + "angle": 0, + "content": "Table 5: User study comparison between M3DDM and MOTIA." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.185, + 0.768, + 0.251 + ], + "angle": 0, + "content": "
MethodVisual-QualityRealism
M3DDM27.4%42.8%
MOTIA72.6%57.2%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.463 + ], + "angle": 0, + "content": "original known regions of the source video. To achieve this, the model should effectively extract useful information from the source video for denoising. For instance, M3DDM concatenates local frames of source video at the input layers and incorporates the global frames through the cross-attention mechanism after passing light encoders. However, the information might not be properly handled especially for out-domain inputs, thus leading to outpainting failure. Instead, by conducting input-specific adaptation on the source video, the model can effectively capture the data-specific patterns in the source videos through gradient. Through this, MOTIA better leverage the data-specific patterns of the source video and image/video generative prior for outpainting. We hope this work inspires following research to exploit more from the source video itself instead of purely relying on the generative prior from intensive training on videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.359, + 0.505 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.585 + ], + "angle": 0, + "content": "We present MOTIA, relying on a combination of input-specific adaptation for capturing inner video patterns and pattern-aware outpainting to generalize these patterns for effective outpainting. Extensive experiments validate the effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Limitations: MOTIA requires learning necessary patterns from the source video, it faces limitation for outpainting videos with limited information. We provide an example of failure case. In Fig. 9, the original source video just contains grass and a woman with a blue shirt. We find our method produces a woman with very similar appearance when outpainting, which might be regarded as a \"copy\" phenomenon." + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.705, + 0.427, + 0.812 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.705, + 0.571, + 0.812 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.705, + 0.715, + 0.812 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.822, + 0.76, + 0.837 + ], + "angle": 0, + "content": "Fig. 9: Example of failure case. Contents outside the red lines are outpainted." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be-Your-Outpainter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.352, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.253 + ], + "angle": 0, + "content": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR, by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "Fu-Yun Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Arora, R., Lee, Y.J.: Singan-gif: Learning a generative video model from a single gif. In: CVPR. pp. 1310-1319 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.227 + ], + "angle": 0, + "content": "2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: CVPR. pp. 18208-18218 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.228, + 0.785, + 0.269 + ], + "angle": 0, + "content": "3. Barnes, C., Shechtman, E., Finkelstein, A., Goldman, D.B.: Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph. 28(3), 24 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.27, + 0.785, + 0.31 + ], + "angle": 0, + "content": "4. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR. pp. 22563-22575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.337 + ], + "angle": 0, + "content": "5. Dehan, L., Van Ranst, W., Vandewalle, P., Goedemé, T.: Complete and temporally consistent video outpainting. In: CVPRW. pp. 687-695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.377 + ], + "angle": 0, + "content": "6. Fan, F., Guo, C., Gong, L., Wang, B., Ge, T., Jiang, Y., Luo, C., Zhan, J.: Hierarchical masked 3d diffusion model for video outpainting. In: ACM MM. pp. 7890-7900 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.404 + ], + "angle": 0, + "content": "7. Gao, C., Saraf, A., Huang, J.B., Kopf, J.: Flow-edge guided video completion. In: ECCV. pp. 713-729. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.445 + ], + "angle": 0, + "content": "8. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: Animatediff: animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "9. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.513 + ], + "angle": 0, + "content": "0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.54 + ], + "angle": 0, + "content": "1. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.785, + 0.567 + ], + "angle": 0, + "content": "2. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. arXiv:2204.03458 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.785, + 0.609 + ], + "angle": 0, + "content": "3. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.648 + ], + "angle": 0, + "content": "4. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR. pp. 11461-11471 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.69 + ], + "angle": 0, + "content": "5. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "6. Ma, Y., He, Y., Wang, H., Wang, A., Qi, C., Cai, C., Li, X., Li, Z., Shum, H.Y., Liu, W., et al.: Follow-your-click: Open-domain regional image animation via short prompts. arXiv preprint arXiv:2403.08268 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "7. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "8. Nikankin, Y., Haim, N., Irani, M.: Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Be-Your-Outpainter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "20. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.217 + ], + "angle": 0, + "content": "21. Rout, L., Parulekar, A., Caramanis, C., Shakkottai, S.: A theoretical justification for image inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2302.01217 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "22. Shaham, T.R., Dekel, T., Michaeli, T.: Singan: Learning a generative model from a single natural image. In: ICCV. pp. 4570-4580 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "23. Shen, D., Song, G., Xue, Z., Wang, F.Y., Liu, Y.: Rethinking the spatial inconsistency in classifier-free diffusion guidance. In: CVPR. pp. 9370-9379 (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.328 + ], + "angle": 0, + "content": "24. Shi, X., Huang, Z., Wang, F.Y., Bian, W., Li, D., Zhang, Y., Zhang, M., Cheung, K.C., See, S., Qin, H., et al.: Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In: ACM SIGGRAPH 2024 Conference Papers. pp. 1-11 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "25. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.787, + 0.397 + ], + "angle": 0, + "content": "26. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "27. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.439, + 0.787, + 0.467 + ], + "angle": 0, + "content": "28. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "29. Wang, F.Y., Chen, W., Song, G., Ye, H.J., Liu, Y., Li, H.: Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.508, + 0.787, + 0.55 + ], + "angle": 0, + "content": "30. Wang, F.Y., Huang, Z., Shi, X., Bian, W., Song, G., Liu, Y., Li, H.: Animatelem: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning. arXiv preprint arXiv:2402.00769 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "31. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "32. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "33. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.674, + 0.787, + 0.715 + ], + "angle": 0, + "content": "34. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In: CVPR. pp. 1316-1324 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "35. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: CVPR. pp. 10459-10469 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.758, + 0.787, + 0.785 + ], + "angle": 0, + "content": "36. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.785, + 0.787, + 0.814 + ], + "angle": 0, + "content": "37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.814 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_origin.pdf b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8f2c6db331549bb8275a9b926817643559b23525 --- /dev/null +++ b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/ba54e466-10d4-4f1f-b94b-4523eb49157e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f92bf0fc2906df2ce31bc01be68806d86a88bca9b32c6dd7b04af9566aa42a0 +size 10650636 diff --git a/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/full.md b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1975a399529b84ab673ee27a2ea59f8c95d733d --- /dev/null +++ b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/full.md @@ -0,0 +1,300 @@ +# Be-Your-Outpainter: Mastering Video Outpainting through Input-Specific Adaptation + +Fu-Yun Wang $^{1}$ Xiaoshi Wu $^{1}$ Zhaoyang Huang $^{2}$ + +Xiaoyu Shi $^{1}$ Dazhong Shen $^{3}$ Guanglu Song $^{4}$ + +Yu Liu $^{3}$ Hongsheng Li $^{1,3,5}$ + +$^{1}$ MMLab, CUHK $^{2}$ Avolution AI + +$^{3}$ Shanghai AI Lab $^{4}$ SenseTime Research $^{5}$ CPII under InnoHK + +{fywang@link, hsli@ee}.cuhk.edu.hk + +Abstract. Video outpainting is a challenging task, aiming at generating video content outside the viewpoint of the input video while maintaining inter-frame and intra-frame consistency. Existing methods fall short in either generation quality or flexibility. We introduce MOTIA (Mastering Video Outpainting Through Input-Specific Adaptation), a diffusion-based pipeline that leverages both the intrinsic data-specific patterns of the source video and the image/video generative prior for effective outpainting. MOTIA comprises two main phases: input-specific adaptation and pattern-aware outpainting. The input-specific adaptation phase involves conducting efficient and effective pseudo outpainting learning on the single-shot source video. This process encourages the model to identify and learn patterns within the source video, as well as bridging the gap between standard generative processes and outpainting. The subsequent phase, pattern-aware outpainting, is dedicated to the generalization of these learned patterns to generate outpainting outcomes. Additional strategies are proposed to better leverage the diffusion model's generative prior and the acquired video patterns from source videos for inference. Extensive evaluations underscore MOTIA's superiority, outperforming existing state-of-the-art methods in widely recognized benchmarks. Notably, these advancements are achieved without necessitating extensive, task-specific tuning. More details are available at https://be-your-outpainter.github.io/. + +# 1 Introduction + +Video outpainting aims to expand the visual content out of the spatial boundaries of videos, which has important real-world applications [3,5,6]. For instance, in practice, videos are usually recorded with a fixed aspect ratio, such as in movies or short clips. This becomes an issue when viewing these videos on smartphones, which often have varying aspect ratios, resulting in unsightly black bars on the screen that detract from the viewing experience. Proper ways for video outpainting are crucial in solving this issue. By expanding the visual content beyond the original frame, it adapts the video to fit various screen sizes seamlessly. This process ensures that the audience enjoys a full-screen experience without any + +![](images/050d0281f73a46062e6c49285c721d0512498e66ab73eb53909bc5814d6f2094.jpg) +Fig. 1: MOTIA is a high-quality flexible video outpainting pipeline, leveraging the intrinsic data-specific patterns of source videos and image/video generative prior for state-of-the-art performance. Quantitative metric improvement of MOTIA is significant (Table 1). + +![](images/870d2ca7013cd6a21e83dd46eb593a8b798574f79b7c2171f48cd60ddb98b540.jpg) + +compromise in visual integrity. However, the challenges associated with video outpainting are significant. It requires not just the expansion of each frame's content but also the preservation of temporal (inter-frame) and spatial (intra-frame) consistency across the video. + +Currently, there are two primary approaches to video outpainting. The first employs optical flows and specialized warping techniques to extend video frames, involving complex computations and carefully tailored hyperparameters to ensure the added content remains consistent [5, 7]. However, their results are far from satisfactory, suffering from blurred content. The other type of approach in video outpainting revolves around training specialized models tailored for video inpainting and outpainting with extensive datasets [6, 35]. However, they have two notable limitations: 1) An obvious drawback of these models is their dependency on the types of masks and the resolutions of videos they can handle, which significantly constrains their versatility and effectiveness in real-world applications, as they may not be adequately equipped to deal with the diverse range of video formats and resolutions commonly encountered in practical scenarios. 2) The other drawback is their inability to out-domain video outpainting, even intensively trained on massive video data. Fig. 5 shows a failure example of most advanced previous work [6] that the model faces complete outpainting failure, + +with only blurred corners. We show that a crucial reason behind this is that the model fails at capturing the intrinsic data-specific patterns from out-domain source (input) videos. + +In this work, we propose MOTIA: Mastering Video Outpainting Through Input-Specific Adaptation, a diffusion-based method for open-domain video outpainting with arbitrary types of mask, arbitrary video resolutions and lengths, and arbitrary styles. At the heart of MOTIA is treating the source video itself as a rich source of information [18, 22], which contains key motion and content patterns (intrinsic data-specific patterns) necessary for effective outpainting. We demonstrate that the patterns learned from the source video, combined with the generative capabilities of diffusion models, can achieve surprisingly great outpainting performance. + +Fig. 2 illustrates the workflow of MOTIA. MOTIA consists of two stages: input-specific adaptation and pattern-aware outpainting. During the input-specific adaptation stage, we conduct pseudo video outpainting learning on the source video (videos to be outpainted) itself. Specifically, at each iteration, we heuristically add random masks to the source video and prompt the base diffusion model to recover the masked regions by learning to denoise the video corrupted by white noise, relying on the extracted information from unmasked regions. This process not only allows the model to capture essential data-specific patterns from the source video but also narrows the gap between standard generation and outpainting. We insert trainable lightweight adapters into the diffusion model for tuning to keep the efficiency and stability. In the pattern-aware outpainting stage, we combine the learned patterns from the source video and the generation prior of the diffusion model for effective outpainting. To better leverage the generation ability of the pretrained diffusion model and the learned pattern from the source video, we propose spatial-aware insertion (SA-Insertion) of the tuned adapters for outpainting. Specifically, the insertion weights of adapters gradually decay as the spatial position of features away from the known regions. In this way, the outpainting of pixels near the known regions is more influenced by the learned patterns, while the outpainting of pixels far from the known regions relies more on the original generative prior of diffusion model. To further mitigate potential denoising conflicts and enhance the knowledge transfer between known regions and unknown regions, we incorporate noise regret that we add noise and denoise periodically at early inference steps, which works for more harmonious outpainting results. + +Extensively quantitative and qualitative experiments verify the effectiveness of our proposed method. MOTIA overcomes many limitations of previous methods and outperforms the state-of-the-art intensively trained outpainting method in standard widely used benchmarks. In summary, our contribution is three-fold: 1) We show that the data-specific patterns of source videos are crucial for effective outpainting, which is neglected by previous work. 2) We introduce an adaptation strategy to effectively capture the data-specific patterns and then propose novel strategies to better leverage the captured patterns and pretrained image/video generative prior for better results. 3) Vast experiments verify that + +![](images/2e51e50cd12449630c71e6942a4f9b5dfec56676fde7061c119d5a82fe5e07af.jpg) +Fig. 2: Workflow of MOTIA. Blue lines represent the workflow of input-specific adaptation, and green lines represent the workflow of pattern-aware outpainting. + +our performance in video outpainting is great, significantly outperforming previous state-of-the-art methods in both quantitative metrics and user studies. + +# 2 Related Works + +In this section, we discuss related diffusion models and outpainting methods. Diffusion models. Diffusion models (a.k.a., score-based models) [9,10,17,20,21] have gained increasing attention due to its amazing ability to generate highly detailed images. Current video diffusion models [4, 9, 15, 16, 24, 25, 28, 30] are generally built upon the extension of image diffusion models through insertion of temporal layers. They are either trained with image-video joint tuning [12, 22] or trained with spatial weights frozen [4] to mitigate the negative influence of the poor captions and visual quality of video data. + +Outpainting methods. Video outpainting is largely built upon the advancements in image outpainting, where techniques ranged from patch-based methods (e.g., PatchMatch [3]) to more recent deep learning approaches like GANs [1,34]. Diffusion models [2,14], benefiting from the learned priors on synthesis tasks and the process of iterative refinement, achieve state-of-the-art performance on image outpainting tasks. The research focusing on video outpainting is comparatively few. Previous works typically apply optical flow for outpainting, which warps content from adjacent frames to the outside corners, but their results are far from satisfactory. Recently, M3DDM [6] trained a large 3D diffusion models with specially designed architecture for outpainting on massive video data, achieving much better results compared to previous methods, showcasing the huge potential of diffusion models on video outpainting. However, as we claimed, they have two main limitations: 1) The inflexibility for mask types and video resolutions. They can only outpaint video with resolution $256 \times 256$ with square type of masking. 2) Inability for out-domain video outpainting. As shown in + +Fig. 5, they encounter outpainting failure when processing out-domain videos even intensively trained on massive video data. + +# 3 Preliminaries + +Diffusion models [10] add noise to data through a Markov chain process. Given initial data $\pmb{x}_0 \sim q(\pmb{x}_0)$ , this process is formulated as: + +$$ +q \left(\boldsymbol {x} _ {1: T} \mid \boldsymbol {x} _ {0}\right) = \prod_ {t = 1} ^ {T} q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1}\right), \quad q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t} \mid \sqrt {\alpha_ {t}} \boldsymbol {x} _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where $\beta_{t}$ is the noise schedule and $\alpha_{t} = 1 - \beta_{t}$ . The data reconstruction, or denoising process, is accomplished by the reverse transition modeled by $p_{\theta}(\boldsymbol{x}_{t-1}|\boldsymbol{x}_t)$ : + +$$ +q \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}, \boldsymbol {x} _ {0}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t - 1}; \tilde {\boldsymbol {\mu}} _ {t} \left(\boldsymbol {x} _ {t}, \boldsymbol {x} _ {0}\right), \tilde {\beta} _ {t} \mathbf {I}\right), \tag {2} +$$ + +with $\tilde{\pmb{\mu}}_t(\pmb{x}_t,\pmb{x}_0) = \frac{1}{\sqrt{\alpha_t}}\pmb{x}_t - \frac{1 - \alpha_t}{\sqrt{1 - \bar{\alpha}_t\sqrt{\alpha_t}}}\pmb{\epsilon}, \bar{\alpha}_t = \prod_{s=1}^t\alpha_s, \tilde{\beta}_t = \frac{1 - \bar{\alpha}_{t-1}}{1 - \bar{\alpha}_t}\beta_t$ , and $\pmb{\epsilon}$ is the noise added to $\pmb{x}_0$ to obtain $\pmb{x}_t$ . + +Diffusion-based outpainting aims to predict missing pixels at the corners of the masked region with the pre-trained diffusion models. We denote the ground truth as $\pmb{x}$ , mask as $\pmb{m}$ , known region as $(1 - m) \odot \pmb{x}$ , and unknown region as $m \odot x$ . At each reverse step in the denoising process, we modify the known regions by incorporating the intermediate noisy state of the source data from the corresponding timestep in the forward diffusion process (which adds noise), provided that this maintains the correct distribution of correspondences. Specifically, each reverse step can be denoted as the following formulas: + +$$ +\boldsymbol {x} _ {t - 1} ^ {\text {k n o w n}} \sim \mathcal {N} \left(\sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right), \quad \boldsymbol {x} _ {t - 1} ^ {\text {u n k n o w n}} \sim \mathcal {N} \left(\boldsymbol {\mu} _ {\theta} (\boldsymbol {x} _ {t}, t), \Sigma_ {\theta} (\boldsymbol {x} _ {t}, t)\right), \tag {3} +$$ + +$$ +\boldsymbol {x} _ {t - 1} = \boldsymbol {m} \odot \boldsymbol {x} _ {t - 1} ^ {\text {k n o w n}} + (\boldsymbol {1} - \boldsymbol {m}) \odot \boldsymbol {x} _ {t - 1} ^ {\text {u n k n o w n}}, \tag {4} +$$ + +where the $\pmb{x}_{t - 1}^{\mathrm{known}}$ is padded to the target resolution before the masked merging. + +# 4 Methodology + +This section presents MOTIA, a method demonstrating exceptional performance in video outpainting tasks. We begin by defining the concept of video outpainting and describing the foundational model in Section 4.1. and Section 4.2. Subsequently, we delve into the specifics of input-specific adaptation and pattern-aware outpainting in Sections 4.3 and 4.4, respectively. Moreover, we show that our approach has great promise in extending its application to long video outpainting, which will be explored in Section 4.5. + +![](images/998d63b3de30b762524377e7f4ca953c13b7ee17ff387a654343256b1558191c.jpg) +Fig. 3: Sample results of quantitative experiments. All videos are outpainted with a horizontal mask ratio of 0.66. Contents outside the yellow lines are outpainted by MOTIA. + +# 4.1 Problem Formulation + +For a video represented as $\pmb{v} \in \mathbb{R}^{t \times h \times w \times d}$ , where $t$ denotes the number of frames, $h$ denotes frame height, $w$ denotes frame width, and $d$ denotes channel depth. Video outpainting model $f(\pmb{v})$ is designed to generate a spatially expanded video $\pmb{v}' \in \mathbb{R}^{t \times h' \times w' \times d}$ . This process not only increases the spatial dimensions ( $h' > h$ , $w' > w$ ), but also requires to ensure continuity and harmony between the newly expanded regions and the known regions. The transformation maintains the known regions unchanged, with $f(\pmb{v})$ acting as an identity in these regions. + +# 4.2 Network Expansion + +Network inflation. MOTIA leverages the pre-trained text-to-image (T2I) model, Stable Diffusion. In line with previous video editing techniques [32], we transform 2D convolutions into pseudo 3D convolutions and adapt 2D group normalizations into 3D group normalizations to process video latent features. Specifically, the $3 \times 3$ kernels in convolutions are replaced by $1 \times 3 \times 3$ kernels, maintaining identical weights. Group normalizations are executed across both temporal + +and spatial dimensions, meaning that all 3D features within the same group are normalized simultaneously, followed by scaling and shifting. + +Masked video as conditional input. Additionally, we incorporate a ControlNet [36], initially trained for image inpainting, to manage additional mask inputs. Apart from noise input, ControlNet can also process masked videos to extract effective information for more controllable denoising. In these masked videos, known regions have pixel values ranging from 0 to 1, while values of masked regions are set to $-1$ . + +Temporal consistency prior. To infuse the model with temporal consistency priors, we integrate temporal modules pre-trained on text-to-video (T2V) tasks. Note that although MOTIA relies on pre-trained video diffusion modules, applying these pre-trained temporal modules directly for video outpainting yields rather bad results, significantly inferior to all baseline methods (Table. 2). However, when equipped with our proposed MOTIA, the model demonstrates superior performance even in comparison to models specifically designed and trained for video outpainting, underscoring the efficacy of MOTIA. + +# 4.3 Input-Specific Adaptation + +The input-specific adaptation phase is crucial in our video outpainting method, aiming to tailor the model for the specific challenges of outpainting. This phase involves training on the source video with a pseudo-outpainting task, importantly, enabling the model to learn intrinsic content and motion patterns (data-specific patterns) within the source video as well as narrowing the gap between the standard generation process and outpainting. + +Video augmentation. Initially, we augment the source video. Transformations like identity transformation, random flipping, cropping, and resizing can be employed. This step can potentially help the model better learn and adapt to diverse changes in video content. For longer video outpainting, as we will discuss later, instead of taking it as a whole, we randomly sample short video clips from it to reduce the cost of the adaptation phase. + +Video masking. We then add random masks to the video. We adopt a heuristic approach that uniformly samples edge boundaries of 4 sides within given limits. The area enclosed by these boundaries is considered the known region, while the rest is the unknown region. This masked video serves as the conditional input for the ControlNet, simulating the distribution of known and unknown areas in actual outpainting scenarios. + +Video noising. Additionally, we apply noise to the video following the DDPM [10] by randomly selecting diffusion timesteps. This noisy video serves as an input for both the ControlNet and the Stable Diffusion model. + +Optimization. Finally, we optimize the model. To ensure efficiency, low-rank adapters are inserted into the layers of the diffusion model. We optimize only the parameters of these adapters while keeping the other parameters frozen. The loss function is + +$$ +\mathcal {L} = \left\| \boldsymbol {\epsilon} - \hat {\boldsymbol {\epsilon}} _ {\bar {\theta} _ {l}, \bar {\theta} _ {c}, \boldsymbol {\theta} _ {a}} (\boldsymbol {v} _ {\mathrm {n o i s y}}, \boldsymbol {v} _ {\mathrm {m a s k e d}}, t) \right\| _ {2}, \tag {5} +$$ + +![](images/81d223586e6d6f8f65400938a9f42153d292da8e22227d435f3d96639c29818c.jpg) +Fig.4: Spatial-aware insertion scales the insertion weights of adapters for better leveraging of learned patterns and generative prior. + +![](images/f8e59ec18a3fc9a843bde6c668506280c2883f48460e5e041a0d17da4dba2391.jpg) +Fig. 5: Failure example of previous methods. Many previous methods including the intensively trained models on video outpainting still might suffer from generation failure, that the model simply generates blurred corners. MOTIA never encounters this failure. + +where $t$ represents the timestep in the process, $\epsilon$ is the added noise, $\pmb{v}_{\mathrm{noisy}}$ refers to the video perturbed by $\epsilon$ , and $\pmb{v}_{\mathrm{masked}}$ denotes the masked video. The parameters $\theta_{l}$ , $\theta_{c}$ , and $\theta_{a}$ correspond to the Diffusion Model, ControlNet, and adapters, respectively. The bar over these parameters indicates they are frozen during the optimization. This optimization process, including the steps of augmentation, masking, and noising, is repeated to update the lightweight adapters to capture the data-specific patterns from the source video. + +# 4.4 Pattern-Aware Outpainting + +Following the initial phase of input-specific adaptation, our model shows promising results in video outpainting using basic pipelines as outlined in Eq. 3 and Eq. 4, achieving commendable quality. However, we here introduce additional inference strategies that can be combined to better leverage the learned data-specific patterns from the input-specific adaptation phase for better outpainting results. We call the outpainting process that incorporates these strategies pattern-aware outpainting. + +Spatial-aware insertion. It is important to acknowledge that in the input-specific adaptation phase, the model is fine-tuned through learning outpainting within the source video. However, at the outpainting phase, the model is expected to treat the entire source video as known regions and then fill the unknown regions at edges (i.e., generating a video with a larger viewpoint and resolution). This specificity may lead to a noticeable training-inference gap during outpainting, potentially affecting the outpainting quality. To balance the fine-tuned patterns with the diffusion model's inherent generative prior, we introduce the concept of spatial-aware insertion (SA-Insertion) of adapters as shown in Fig. 4. The adaptation involves adjusting the insertion weight of tuned low-rank adapters based on the feature's spatial position. We increase insertion weight near known areas to utilize the learned patterns while decreasing it in farther regions to rely + +more on the original generative capacity of the diffusion model. To be specific, + +$$ +\mathbf {W} _ {\mathrm {a d a p t e d}} ^ {\top} \boldsymbol {x} _ {\boldsymbol {p}} = \mathbf {W} ^ {\top} \boldsymbol {x} _ {\boldsymbol {p}} + \alpha (\boldsymbol {p}) \left(\mathbf {W} _ {\mathrm {u p}} \mathbf {W} _ {\mathrm {d o w n}}\right) ^ {\top} \boldsymbol {x} _ {\boldsymbol {p}}. \tag {6} +$$ + +Here, $\pmb{p}$ signifies the spatial position of $\pmb{x}$ , $\mathbf{W} \in \mathbb{R}^{d_{\mathrm{in}} \times d_{\mathrm{out}}}$ denotes the linear transformation in layers of diffusion model, $\mathbf{W}_{\mathrm{down}} \in \mathbb{R}^{d_{\mathrm{in}} \times r}$ and $\mathbf{W}_{\mathrm{up}} \in \mathbb{R}^{r \times d_{\mathrm{out}}}$ are the linear components of the adapter with rank $r \ll \min(d_{\mathrm{in}}, d_{\mathrm{out}})$ . The function $\alpha(\pmb{p})$ is defined as: + +$$ +\alpha (\boldsymbol {p}) = \exp \left(- \frac {K \| \boldsymbol {p} - \boldsymbol {p} _ {c} \|}{\operatorname* {m a x} _ {\bar {\boldsymbol {p}}} \| \bar {\boldsymbol {p}} - \boldsymbol {p} _ {c} \|}\right), \tag {7} +$$ + +where $K$ is a constant for controlling decay speed, and $\pmb{p}_c$ represents the nearest side of the known region to $\pmb{p}$ . + +Noise regret. In the context of Eq. 3, merging noisy states from known and unknown regions in video outpainting tasks is similar to sampling from two different vectors, which can potentially disrupt the denoising direction. Given the significant impact of early steps on the generation's structure, later denoising may not rectify these initial discrepancies. Inspired by DDPM-based image inpainting methods [14,21], we propose to re-propagate the noisy state into a noisier state by adding noise when denoising and then provide the model a second chance for re-denoising. This helps integrate known region data more effectively and reduces denoising direction conflicts. In detail, after obtaining $\mathbf{v}_t$ during denoising, we conduct + +$$ +\boldsymbol {v} _ {t + L} = \sqrt {\Pi_ {i = t + 1} ^ {t + L} \alpha_ {i}} \boldsymbol {v} _ {t} + \sqrt {1 - \Pi_ {i = t + 1} ^ {t + L} \alpha_ {i}} \boldsymbol {\epsilon}, \tag {8} +$$ + +where $\alpha_{i} = 1 - \beta_{i}$ and $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . Then we restart the denoising process. We repeat this progress for $M$ times. We only conduct it in the early denoising steps. + +# 4.5 Extension to Long Video Outpainting + +We show that our method can be easily extended for long video outpainting. Specifically, for the stage of input-specific adaptation, instead of taking the long video as a whole for adaptation (Direct adaptation on long videos is costly and does not align with the video generation prior of the pretrained modules), we randomly sample short video clips from the long video for tuning to learn global patterns without requiring more GPU memory cost. For the stage of pattern-aware outpainting, we split the long video into short video clips with temporal overlapping (i.e., some frames are shared by different short video clips), and then conduct temporal co-denoising following Gen-L [29]. Specifically, the denoising result for $j^{th}$ frame of the long video at timestep $t$ is approximated by the weighted sum of all the corresponding frames in short video clips that contain it, + +$$ +\boldsymbol {v} _ {t - 1, j} = \frac {\sum_ {i \in \mathcal {I} ^ {j}} \left(\left(W _ {i , j ^ {*}}\right) ^ {2} \otimes \boldsymbol {v} _ {t - 1 , j ^ {*}} ^ {i}\right)}{\sum_ {i \in \mathcal {I} ^ {j}} \left(W _ {i , j ^ {*}} ^ {2}\right) ^ {2}}, \tag {9} +$$ + +![](images/a9eb8b1e7124d465cde9d8f138a9a0a1feccd55ca0fb9a4d7dc3b6b9783ed94b.jpg) +Fig. 6: Qualitative comparison. Other methods outpainting the source video with a mask ratio of 0.6. MOTIA outpainting the source video with a larger mask ratio of 0.66 while achieving obviously better outpainting results. + +where $\otimes$ denotes element-wise multiplication, $\pmb{v}_{t-1,j}$ denotes the noisy state of the $j^{th}$ frame at timestep $t$ , $\pmb{v}_{t-1,j^*}^i$ is the noisy state of $j^{th}$ frame predicted with only information from the $i^{th}$ video clip at timestep $t$ , $W_{i,j^*}$ is the per-pixel weight, which is as 1 as default. + +# 5 Experiments + +# 5.1 Experimental Setup + +Benchmarks. To verify the effectiveness of MOTIA, we conduct evaluations on DAVIS [19] and YouTube-VOS [33], which are widely used benchmarks for video outpainting. Following M3DDM [6], we compare the results of different methods in the horizontal direction, using mask ratios of 0.25 and 0.66. + +Evaluation metrics. Our evaluation approach utilizes four well-established metrics: Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [31], Learned Perceptual Image Patch Similarity (LPIPS) [37], and Frechet Video Distance (FVD) [27]. For assessing PSNR, SSIM, and FVD, the generated videos are converted into frames within a normalized value range of $[0, 1]$ . LPIPS is evaluated over a range of $[-1, 1]$ . About the FVD metric, we adopt a uniform frame sampling, with 16 frames per video for evaluation following M3DDM. + +Implementation details. Our method is built upon Stable Diffusion v1-5. We add the ControlNet pretrained on image inpainting to enable the model to accept additional masked image inputs. The temporal modules are initialized with the weights from pretrained motion modules [8] to obtain additional motion priors. The motion modules are naive transformer blocks trained with solely text-to-video tasks on WebVid. For the input-specific adaptation, the low-rank adapters are trained using the Adam optimizer. We set the learning rate to $10^{-4}$ , and set the weight decay to $10^{-2}$ . The LoRA rank and $\alpha_{\mathrm{lora}}$ are set to 16 and, + +Table 1: Quantitative comparison of video outpainting methods on DAVIS and YouTube-VOS datasets. $\uparrow$ means 'better when higher', and $\downarrow$ indicates 'better when lower'. + +
MethodDAVISYouTube-VOS
PSNR↑ SSIM↑ LPIPS↓ FVD↓PSNR↑ SSIM↑ LPIPS↓ FVD
VideoOutpainting [5]17.960.62720.2331363.118.250.71950.2278149.7
SDM [6]20.020.70780.2165334.619.910.72770.200194.81
M3DDM [6]20.260.70820.2026300.020.200.73120.185466.62
MOTIA20.360.75780.1595286.320.250.76360.172758.99
+ +8, respectively. The number of training steps is set to 1000. We do not apply augmentation for simplicity. For both mask ratios of 0.66 and 0.25, we simply apply the same random mask strategy, which uniformly crops a square in the middle as the known regions. For the pattern-aware outpainting, the diffusion steps are set to 25 and the classifier-free guidance (CFG) [11, 23] scale is set to 7.5 and we only apply CFG at the first 15 inference steps. When adding noise regret to further improve the results, we set jump length $L = 3$ , and repeat time $M = 4$ . We only apply noise regret in the first half inference steps. Note that our method is built upon LDM, which requires text-conditional inputs. For a fair comparison and to remove the influence of the choice of text prompt, we apply Blip [13] to select the prompt automatically. We observe dozens of prompt mistakes but do not modify them to avoid man-made influence. + +# 5.2 Qualitative Comparison + +Fig. 6 showcases a qualitative comparison of MOTIA against other methods. Outpainting a narrow video into a square format. MOTIA employs a mask ratio of 0.66, surpassing the 0.6 ratio utilized by other methods, and demonstrates superior performance even with this higher mask ratio. The SDM method only manages to blur the extremities of the video's background, egregiously overlooking the primary subject and resulting in the outpainting failure as previously highlighted in Fig. 5. Dehan's approach effectively outpaints the background but utterly fails to address the foreground, leading to notable distortions. In contrast, the M3DDM method adeptly handles both subject and background integration but is marred by considerable deviations in subject characteristics, such as incorrect brown coloration in the dog's fur across several frames. Our method stands out by achieving optimal results, ensuring a harmonious and consistent outpainting of both the foreground and background. + +# 5.3 Quantitative Comparison + +Table 1 summarizes the evaluation metrics of our method compared to other approaches. Our method achieves comparable results to the best method in PSNR. It shows significant improvements in video quality (SSIM), perceptual metric (LPIPS), and distribution similarity (FVD). Specifically, our SSIM, LPIPS, + +![](images/9b3e1be9256d00d767c65f89419a7a65f45b503849d5caf919ab908ad9c514f6.jpg) +$SD + T$ + +![](images/6167695fb83bfd4ce1b7d9aef173f38a4c90f078a3210a414066c017fb5ffa5d.jpg) +$SD + T + C$ + +![](images/8f6d01ebfdc090075a8da7817199060153f0d5da6b92ea64f899dca46a7ea6d0.jpg) +Direct-tune +Input-Specific Adaptation + +![](images/1c63d74782cf2f2447a94af9ad250041469872d1597d57522ce75ed9ae2898ce.jpg) + +![](images/87b4dad41111d6cb1172867e35621992308b67e104485b01d3d236c9a7cd623b.jpg) +Fig. 7: Visual examples of ablation study on the proposed input-specific adaptation. +Direct (Ours $w / o$ SA + NR) +SA (Ours w/o NR) +Pattern-Aware Outpainting +Fig. 8: Visual examples of ablation study on pattern-aware outpainting. + +and FVD metrics show improvements of $7.00\%$ , $21.27\%$ , and $4.57\%$ respectively on the DAVIS dataset, and $4.43\%$ , $6.85\%$ , and $11.45\%$ on the YouTube-VOS dataset compared to the best-performing method. + +# 5.4 Ablation Study + +Ablation study on input-specific adaptation. We conducted the ablation study on input-specific adaptation with the DAVIS dataset to verify its effectiveness, as shown in Fig. 7 and Table 2. "SD+T" represents the result of directly combining the temporal module with Stable Diffusion, which led to a complete outpainting failure. "SD+T+C" indicates the additional use of ControlNet, resulting in similarly poor outcomes. "Direct-tune" refers to the approach of directly fitting the original video without outpainting training; in this case, we observed a very noticeable color discrepancy between the outpainted and known areas. In contrast, our method achieved the best results, ensuring consistency in both the visual and temporal aspects. The metrics shown in Table 2 also support this observation, with MOTIA significantly outperforming the other baselines. + +Ablation study on pattern-aware outpainting. Table 3 summarizes our ablation experiments for the pattern-aware outpainting part. We conducted extensive validation on the YouTube-VOS dataset. "Direct" refers to performing outpainting according to Eq. 3 directly after input-specific adaptation. "SA" denotes spatially-aware insertion, and "SA+NR" indicates the combined use of spatially-aware insertion and noise regret. The experimental results demonstrate that each of our components effectively enhances performance. Specifically, Combining both SA-Insertion and Noise regret, the PSNR, SSIM, LPIPS, and FVD metrics show improvements of $2.69\%$ , $0.90\%$ , $3.95\%$ , and $11.32\%$ respectively than directly applying Eq. 3. Fig. 8 presents the visual examples of ablation study on our proposed pattern-aware outpainting part. When removing NR, it + +Table 2: Ablation study on input-specific adaptation. + +
ComponentPSNR↑SSIM↑LPIPS↓FVD ↓
SD+T15.590.66400.2979672.7
SD+T+C16.810.69610.2338515.4
Direct-tune19.420.73750.1784312.1
MOTIA20.360.75780.1595286.3
+ +Table 3: Ablation study on the proposed pattern-aware outpainting. + +
ComponentPSNR↑SSIM↑LPIPS↓FVD ↓
Direct19.720.75680.179866.52
SA19.970.76080.175258.40
SA+NR20.250.76360.172758.99
+ +might fail to align the texture colors or produce unreasonable details (e.g., arms in the middle of Fig. 8). When further removing SA, it could potentially generate unrealistic results caused by the overfitting to the target video (e.g., the white collar on the left of Fig. 8). Note that even though the FVD degrades in a very slight manner, all the other metrics increase and we qualitatively find it to be helpful for improving results. + +# 5.5 Discussions + +Model and computation complexity. Model Complexity: The original model has 1.79 billion (including the auto-encoder and text encoder) parameters in total, while the added adapters contain 7.49 million parameters, leading to an increase of $0.42\%$ in memory usage. Computation Complexity: We report the peak GPU VRAM and the time required for outpainting a target video from $512 \times 512$ to $512 \times 1024$ with 16 frames at two stages in Table 4. For longer videos, as described in Section 4.5, instead of processing the long video as a whole, we adapt only to short video clips sampled from the long video. This approach does not require additional time or GPU VRAM during the input-specific adaptation phase. Additionally, with temporal co-denoising [29], the GPU VRAM usage remains the same as that for short video during the pattern-aware outpainting phase, while the required time increases linearly with the video length. + +User study. We conducted a user study between MOTIA and M3DDM, utilizing the DAVIS dataset with a horizontal mask of 0.66 as source videos. Preferences were collected from 10 volunteers, each evaluating 50 randomly selected sets of results based on visual quality (such as clarity, color fidelity, and texture detail) and realism (including motion consistency, object continuity, and integration with the background). Table 2 demonstrates that the outputs from MOTIA are preferred over those from M3DDM in both visual quality and realism. + +Why MOTIA outperforms (Why previous methods fail). 1) Flexibility. Current video diffusion models are mostly trained with fixed resolution and length, lacking the ability to tackle videos with various aspect ratios and lengths. In contrast, the adaptation phase of MOTIA allows the model to better capture the size, length, and style distribution of the source video, greatly narrowing the gap between pretrained weights and the source video. 2) Ability for capturing intrinsic patterns from source video. A crucial point for successful outpainting is the predicted score of diffusion models should be well-compatible with the + +Table 4: Computation complexity. + +
PhaseVRAM↓Time↓
Adaptation12.70 GB309 Seconds
Outpainting5.80 GB58 Seconds
MOTIA (In total)12.70 GB367 Seconds
+ +Table 5: User study comparison between M3DDM and MOTIA. + +
MethodVisual-QualityRealism
M3DDM27.4%42.8%
MOTIA72.6%57.2%
+ +original known regions of the source video. To achieve this, the model should effectively extract useful information from the source video for denoising. For instance, M3DDM concatenates local frames of source video at the input layers and incorporates the global frames through the cross-attention mechanism after passing light encoders. However, the information might not be properly handled especially for out-domain inputs, thus leading to outpainting failure. Instead, by conducting input-specific adaptation on the source video, the model can effectively capture the data-specific patterns in the source videos through gradient. Through this, MOTIA better leverage the data-specific patterns of the source video and image/video generative prior for outpainting. We hope this work inspires following research to exploit more from the source video itself instead of purely relying on the generative prior from intensive training on videos. + +# 6 Conclusion + +We present MOTIA, relying on a combination of input-specific adaptation for capturing inner video patterns and pattern-aware outpainting to generalize these patterns for effective outpainting. Extensive experiments validate the effectiveness. + +Limitations: MOTIA requires learning necessary patterns from the source video, it faces limitation for outpainting videos with limited information. We provide an example of failure case. In Fig. 9, the original source video just contains grass and a woman with a blue shirt. We find our method produces a woman with very similar appearance when outpainting, which might be regarded as a "copy" phenomenon. + +![](images/88f31870890111efaa882087367b78e385d684c45c3eb90966c0a182986cd2bc.jpg) +Fig. 9: Example of failure case. Contents outside the red lines are outpainted. + +![](images/06eb5ae5a9ca15710ed566732a9752e4649bf0cea7ad81e32b85908cf56b775c.jpg) + +![](images/a52662512af143c2459994acbfb8d0c2101461270dd2a98ec61b0d9992497269.jpg) + +# Acknowledgment: + +This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR, by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. + +# References + +1. Arora, R., Lee, Y.J.: Singan-gif: Learning a generative video model from a single gif. In: CVPR. pp. 1310-1319 (2021) +2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: CVPR. pp. 18208-18218 (2022) +3. Barnes, C., Shechtman, E., Finkelstein, A., Goldman, D.B.: Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph. 28(3), 24 (2009) +4. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR. pp. 22563-22575 (2023) +5. Dehan, L., Van Ranst, W., Vandewalle, P., Goedemé, T.: Complete and temporally consistent video outpainting. In: CVPRW. pp. 687-695 (2022) +6. Fan, F., Guo, C., Gong, L., Wang, B., Ge, T., Jiang, Y., Luo, C., Zhan, J.: Hierarchical masked 3d diffusion model for video outpainting. In: ACM MM. pp. 7890-7900 (2023) +7. Gao, C., Saraf, A., Huang, J.B., Kopf, J.: Flow-edge guided video completion. In: ECCV. pp. 713-729. Springer (2020) +8. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: Animatediff: animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023) +9. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020) +1. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +2. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. arXiv:2204.03458 (2022) +3. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900. PMLR (2022) +4. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR. pp. 11461-11471 (2022) +5. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186 (2023) +6. Ma, Y., He, Y., Wang, H., Wang, A., Qi, C., Cai, C., Li, X., Li, Z., Shum, H.Y., Liu, W., et al.: Follow-your-click: Open-domain regional image animation via short prompts. arXiv preprint arXiv:2403.08268 (2024) +7. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +8. Nikankin, Y., Haim, N., Irani, M.: Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743 (2022) +9. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) + +20. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) +21. Rout, L., Parulekar, A., Caramanis, C., Shakkottai, S.: A theoretical justification for image inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2302.01217 (2023) +22. Shaham, T.R., Dekel, T., Michaeli, T.: Singan: Learning a generative model from a single natural image. In: ICCV. pp. 4570-4580 (2019) +23. Shen, D., Song, G., Xue, Z., Wang, F.Y., Liu, Y.: Rethinking the spatial inconsistency in classifier-free diffusion guidance. In: CVPR. pp. 9370-9379 (June 2024) +24. Shi, X., Huang, Z., Wang, F.Y., Bian, W., Li, D., Zhang, Y., Zhang, M., Cheung, K.C., See, S., Qin, H., et al.: Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In: ACM SIGGRAPH 2024 Conference Papers. pp. 1-11 (2024) +25. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022) +26. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +27. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018) +28. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. NeurIPS (2022) +29. Wang, F.Y., Chen, W., Song, G., Ye, H.J., Liu, Y., Li, H.: Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264 (2023) +30. Wang, F.Y., Huang, Z., Shi, X., Bian, W., Song, G., Liu, Y., Li, H.: Animatelem: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning. arXiv preprint arXiv:2402.00769 (2024) +31. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +32. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565 (2022) +33. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018) +34. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In: CVPR. pp. 1316-1324 (2018) +35. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: CVPR. pp. 10459-10469 (2023) +36. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023) +37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018) \ No newline at end of file diff --git a/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/images.zip b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..23742fb813d07e555aa33e8622690495f0e9dd38 --- /dev/null +++ b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0502f6448d7d863133c5fe87546ecffe763ab0d2f8b5dae214c0d779209f5ffc +size 689275 diff --git a/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/layout.json b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..382e6747ca32a7e6eb108efd4aa33fb7ba6f0ec2 --- /dev/null +++ b/2024/Be-Your-Outpainter_ Mastering Video Outpainting through Input-Specific Adaptation/layout.json @@ -0,0 +1,8911 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "text", + "content": "Be-Your-Outpainter: Mastering Video Outpainting through Input-Specific Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 186, + 167, + 430, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 167, + 430, + 180 + ], + "spans": [ + { + "bbox": [ + 186, + 167, + 430, + 180 + ], + "type": "text", + "content": "Fu-Yun Wang" + }, + { + "bbox": [ + 186, + 167, + 430, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 430, + 180 + ], + "type": "text", + "content": " Xiaoshi Wu" + }, + { + "bbox": [ + 186, + 167, + 430, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 430, + 180 + ], + "type": "text", + "content": " Zhaoyang Huang" + }, + { + "bbox": [ + 186, + 167, + 430, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 180, + 429, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 180, + 429, + 193 + ], + "spans": [ + { + "bbox": [ + 194, + 180, + 429, + 193 + ], + "type": "text", + "content": "Xiaoyu Shi" + }, + { + "bbox": [ + 194, + 180, + 429, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 180, + 429, + 193 + ], + "type": "text", + "content": " Dazhong Shen" + }, + { + "bbox": [ + 194, + 180, + 429, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 194, + 180, + 429, + 193 + ], + "type": "text", + "content": " Guanglu Song" + }, + { + "bbox": [ + 194, + 180, + 429, + 193 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 240, + 193, + 374, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 193, + 374, + 204 + ], + "spans": [ + { + "bbox": [ + 240, + 193, + 374, + 204 + ], + "type": "text", + "content": "Yu Liu" + }, + { + "bbox": [ + 240, + 193, + 374, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 240, + 193, + 374, + 204 + ], + "type": "text", + "content": " Hongsheng Li" + }, + { + "bbox": [ + 240, + 193, + 374, + 204 + ], + "type": "inline_equation", + "content": "^{1,3,5}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 237, + 213, + 376, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 213, + 376, + 224 + ], + "spans": [ + { + "bbox": [ + 237, + 213, + 376, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 237, + 213, + 376, + 224 + ], + "type": "text", + "content": "MMLab, CUHK " + }, + { + "bbox": [ + 237, + 213, + 376, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 237, + 213, + 376, + 224 + ], + "type": "text", + "content": "Avolution AI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 174, + 224, + 448, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 224, + 448, + 236 + ], + "spans": [ + { + "bbox": [ + 174, + 224, + 448, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 174, + 224, + 448, + 236 + ], + "type": "text", + "content": "Shanghai AI Lab " + }, + { + "bbox": [ + 174, + 224, + 448, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 174, + 224, + 448, + 236 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 174, + 224, + 448, + 236 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 174, + 224, + 448, + 236 + ], + "type": "text", + "content": "CPII under InnoHK" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 237, + 388, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 237, + 388, + 247 + ], + "spans": [ + { + "bbox": [ + 227, + 237, + 388, + 247 + ], + "type": "text", + "content": "{fywang@link, hsli@ee}.cuhk.edu.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 273, + 455, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 273, + 455, + 515 + ], + "spans": [ + { + "bbox": [ + 159, + 273, + 455, + 515 + ], + "type": "text", + "content": "Abstract. Video outpainting is a challenging task, aiming at generating video content outside the viewpoint of the input video while maintaining inter-frame and intra-frame consistency. Existing methods fall short in either generation quality or flexibility. We introduce MOTIA (Mastering Video Outpainting Through Input-Specific Adaptation), a diffusion-based pipeline that leverages both the intrinsic data-specific patterns of the source video and the image/video generative prior for effective outpainting. MOTIA comprises two main phases: input-specific adaptation and pattern-aware outpainting. The input-specific adaptation phase involves conducting efficient and effective pseudo outpainting learning on the single-shot source video. This process encourages the model to identify and learn patterns within the source video, as well as bridging the gap between standard generative processes and outpainting. The subsequent phase, pattern-aware outpainting, is dedicated to the generalization of these learned patterns to generate outpainting outcomes. Additional strategies are proposed to better leverage the diffusion model's generative prior and the acquired video patterns from source videos for inference. Extensive evaluations underscore MOTIA's superiority, outperforming existing state-of-the-art methods in widely recognized benchmarks. Notably, these advancements are achieved without necessitating extensive, task-specific tuning. More details are available at https://be-your-outpainter.github.io/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 668 + ], + "type": "text", + "content": "Video outpainting aims to expand the visual content out of the spatial boundaries of videos, which has important real-world applications [3,5,6]. For instance, in practice, videos are usually recorded with a fixed aspect ratio, such as in movies or short clips. This becomes an issue when viewing these videos on smartphones, which often have varying aspect ratios, resulting in unsightly black bars on the screen that detract from the viewing experience. Proper ways for video outpainting are crucial in solving this issue. By expanding the visual content beyond the original frame, it adapts the video to fit various screen sizes seamlessly. This process ensures that the audience enjoys a full-screen experience without any" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 114, + 260, + 354 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 260, + 354 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 260, + 354 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 260, + 354 + ], + "type": "image", + "image_path": "050d0281f73a46062e6c49285c721d0512498e66ab73eb53909bc5814d6f2094.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 361, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 407 + ], + "type": "text", + "content": "Fig. 1: MOTIA is a high-quality flexible video outpainting pipeline, leveraging the intrinsic data-specific patterns of source videos and image/video generative prior for state-of-the-art performance. Quantitative metric improvement of MOTIA is significant (Table 1)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 265, + 114, + 462, + 354 + ], + "blocks": [ + { + "bbox": [ + 265, + 114, + 462, + 354 + ], + "lines": [ + { + "bbox": [ + 265, + 114, + 462, + 354 + ], + "spans": [ + { + "bbox": [ + 265, + 114, + 462, + 354 + ], + "type": "image", + "image_path": "870d2ca7013cd6a21e83dd46eb593a8b798574f79b7c2171f48cd60ddb98b540.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 434, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 483 + ], + "type": "text", + "content": "compromise in visual integrity. However, the challenges associated with video outpainting are significant. It requires not just the expansion of each frame's content but also the preservation of temporal (inter-frame) and spatial (intra-frame) consistency across the video." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": "Currently, there are two primary approaches to video outpainting. The first employs optical flows and specialized warping techniques to extend video frames, involving complex computations and carefully tailored hyperparameters to ensure the added content remains consistent [5, 7]. However, their results are far from satisfactory, suffering from blurred content. The other type of approach in video outpainting revolves around training specialized models tailored for video inpainting and outpainting with extensive datasets [6, 35]. However, they have two notable limitations: 1) An obvious drawback of these models is their dependency on the types of masks and the resolutions of videos they can handle, which significantly constrains their versatility and effectiveness in real-world applications, as they may not be adequately equipped to deal with the diverse range of video formats and resolutions commonly encountered in practical scenarios. 2) The other drawback is their inability to out-domain video outpainting, even intensively trained on massive video data. Fig. 5 shows a failure example of most advanced previous work [6] that the model faces complete outpainting failure," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "Fu-Yun Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "with only blurred corners. We show that a crucial reason behind this is that the model fails at capturing the intrinsic data-specific patterns from out-domain source (input) videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "text", + "content": "In this work, we propose MOTIA: Mastering Video Outpainting Through Input-Specific Adaptation, a diffusion-based method for open-domain video outpainting with arbitrary types of mask, arbitrary video resolutions and lengths, and arbitrary styles. At the heart of MOTIA is treating the source video itself as a rich source of information [18, 22], which contains key motion and content patterns (intrinsic data-specific patterns) necessary for effective outpainting. We demonstrate that the patterns learned from the source video, combined with the generative capabilities of diffusion models, can achieve surprisingly great outpainting performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 483, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 483, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 483, + 554 + ], + "type": "text", + "content": "Fig. 2 illustrates the workflow of MOTIA. MOTIA consists of two stages: input-specific adaptation and pattern-aware outpainting. During the input-specific adaptation stage, we conduct pseudo video outpainting learning on the source video (videos to be outpainted) itself. Specifically, at each iteration, we heuristically add random masks to the source video and prompt the base diffusion model to recover the masked regions by learning to denoise the video corrupted by white noise, relying on the extracted information from unmasked regions. This process not only allows the model to capture essential data-specific patterns from the source video but also narrows the gap between standard generation and outpainting. We insert trainable lightweight adapters into the diffusion model for tuning to keep the efficiency and stability. In the pattern-aware outpainting stage, we combine the learned patterns from the source video and the generation prior of the diffusion model for effective outpainting. To better leverage the generation ability of the pretrained diffusion model and the learned pattern from the source video, we propose spatial-aware insertion (SA-Insertion) of the tuned adapters for outpainting. Specifically, the insertion weights of adapters gradually decay as the spatial position of features away from the known regions. In this way, the outpainting of pixels near the known regions is more influenced by the learned patterns, while the outpainting of pixels far from the known regions relies more on the original generative prior of diffusion model. To further mitigate potential denoising conflicts and enhance the knowledge transfer between known regions and unknown regions, we incorporate noise regret that we add noise and denoise periodically at early inference steps, which works for more harmonious outpainting results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Extensively quantitative and qualitative experiments verify the effectiveness of our proposed method. MOTIA overcomes many limitations of previous methods and outperforms the state-of-the-art intensively trained outpainting method in standard widely used benchmarks. In summary, our contribution is three-fold: 1) We show that the data-specific patterns of source videos are crucial for effective outpainting, which is neglected by previous work. 2) We introduce an adaptation strategy to effectively capture the data-specific patterns and then propose novel strategies to better leverage the captured patterns and pretrained image/video generative prior for better results. 3) Vast experiments verify that" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be-Your-Outpainter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 462, + 264 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 462, + 264 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 462, + 264 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 462, + 264 + ], + "type": "image", + "image_path": "2e51e50cd12449630c71e6942a4f9b5dfec56676fde7061c119d5a82fe5e07af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 272, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 131, + 272, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 131, + 272, + 482, + 296 + ], + "type": "text", + "content": "Fig. 2: Workflow of MOTIA. Blue lines represent the workflow of input-specific adaptation, and green lines represent the workflow of pattern-aware outpainting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 319, + 480, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 480, + 343 + ], + "type": "text", + "content": "our performance in video outpainting is great, significantly outperforming previous state-of-the-art methods in both quantitative metrics and user studies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 363, + 242, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 242, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 242, + 376 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 469, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 469, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 469, + 485 + ], + "type": "text", + "content": "In this section, we discuss related diffusion models and outpainting methods. Diffusion models. Diffusion models (a.k.a., score-based models) [9,10,17,20,21] have gained increasing attention due to its amazing ability to generate highly detailed images. Current video diffusion models [4, 9, 15, 16, 24, 25, 28, 30] are generally built upon the extension of image diffusion models through insertion of temporal layers. They are either trained with image-video joint tuning [12, 22] or trained with spatial weights frozen [4] to mitigate the negative influence of the poor captions and visual quality of video data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Outpainting methods. Video outpainting is largely built upon the advancements in image outpainting, where techniques ranged from patch-based methods (e.g., PatchMatch [3]) to more recent deep learning approaches like GANs [1,34]. Diffusion models [2,14], benefiting from the learned priors on synthesis tasks and the process of iterative refinement, achieve state-of-the-art performance on image outpainting tasks. The research focusing on video outpainting is comparatively few. Previous works typically apply optical flow for outpainting, which warps content from adjacent frames to the outside corners, but their results are far from satisfactory. Recently, M3DDM [6] trained a large 3D diffusion models with specially designed architecture for outpainting on massive video data, achieving much better results compared to previous methods, showcasing the huge potential of diffusion models on video outpainting. However, as we claimed, they have two main limitations: 1) The inflexibility for mask types and video resolutions. They can only outpaint video with resolution " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " with square type of masking. 2) Inability for out-domain video outpainting. As shown in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "Fu-Yun Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "Fig. 5, they encounter outpainting failure when processing out-domain videos even intensively trained on massive video data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 234, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 234, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 234, + 174 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 190, + 479, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 190, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 479, + 215 + ], + "type": "text", + "content": "Diffusion models [10] add noise to data through a Markov chain process. Given initial data " + }, + { + "bbox": [ + 131, + 190, + 479, + 215 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 \\sim q(\\pmb{x}_0)" + }, + { + "bbox": [ + 131, + 190, + 479, + 215 + ], + "type": "text", + "content": ", this process is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 225, + 481, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 225, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 164, + 225, + 481, + 258 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {x} _ {1: T} \\mid \\boldsymbol {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right), \\quad q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "55734b0ad186efdcfdd14646e03b95f72842886e7f09eac46ab9563ff26decfc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 267, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 267, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 131, + 267, + 479, + 292 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 267, + 479, + 292 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 131, + 267, + 479, + 292 + ], + "type": "text", + "content": " is the noise schedule and " + }, + { + "bbox": [ + 131, + 267, + 479, + 292 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 131, + 267, + 479, + 292 + ], + "type": "text", + "content": ". The data reconstruction, or denoising process, is accomplished by the reverse transition modeled by " + }, + { + "bbox": [ + 131, + 267, + 479, + 292 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}_{t-1}|\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 131, + 267, + 479, + 292 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 304, + 481, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 304, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 214, + 304, + 481, + 319 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1}; \\tilde {\\boldsymbol {\\mu}} _ {t} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right), \\tag {2}", + "image_path": "a464ce04af2dad73f9a7e30b850ddf3d30a28c6bb8877590c61a82c14b7fc975.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{\\mu}}_t(\\pmb{x}_t,\\pmb{x}_0) = \\frac{1}{\\sqrt{\\alpha_t}}\\pmb{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t\\sqrt{\\alpha_t}}}\\pmb{\\epsilon}, \\bar{\\alpha}_t = \\prod_{s=1}^t\\alpha_s, \\tilde{\\beta}_t = \\frac{1 - \\bar{\\alpha}_{t-1}}{1 - \\bar{\\alpha}_t}\\beta_t" + }, + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "text", + "content": " is the noise added to " + }, + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 131, + 324, + 479, + 351 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "text", + "content": "Diffusion-based outpainting aims to predict missing pixels at the corners of the masked region with the pre-trained diffusion models. We denote the ground truth as " + }, + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "text", + "content": ", mask as " + }, + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "inline_equation", + "content": "\\pmb{m}" + }, + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "text", + "content": ", known region as " + }, + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "inline_equation", + "content": "(1 - m) \\odot \\pmb{x}" + }, + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "text", + "content": ", and unknown region as " + }, + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "inline_equation", + "content": "m \\odot x" + }, + { + "bbox": [ + 131, + 352, + 480, + 449 + ], + "type": "text", + "content": ". At each reverse step in the denoising process, we modify the known regions by incorporating the intermediate noisy state of the source data from the corresponding timestep in the forward diffusion process (which adds noise), provided that this maintains the correct distribution of correspondences. Specifically, each reverse step can be denoted as the following formulas:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 459, + 481, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 459, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 141, + 459, + 481, + 474 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t - 1} ^ {\\text {k n o w n}} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\quad \\boldsymbol {x} _ {t - 1} ^ {\\text {u n k n o w n}} \\sim \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {\\theta} (\\boldsymbol {x} _ {t}, t), \\Sigma_ {\\theta} (\\boldsymbol {x} _ {t}, t)\\right), \\tag {3}", + "image_path": "4015f46fb5d4f7fbac83ba488bbb720e1b9fd0d05dc70f6786bbf719345b8d5e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 208, + 498, + 481, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 498, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 208, + 498, + 481, + 512 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t - 1} = \\boldsymbol {m} \\odot \\boldsymbol {x} _ {t - 1} ^ {\\text {k n o w n}} + (\\boldsymbol {1} - \\boldsymbol {m}) \\odot \\boldsymbol {x} _ {t - 1} ^ {\\text {u n k n o w n}}, \\tag {4}", + "image_path": "61a91283bf318b370164370e2fa665fd9214c8742d6fd443841f54946eecae28.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 519, + 479, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 519, + 479, + 532 + ], + "spans": [ + { + "bbox": [ + 131, + 519, + 479, + 532 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 131, + 519, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t - 1}^{\\mathrm{known}}" + }, + { + "bbox": [ + 131, + 519, + 479, + 532 + ], + "type": "text", + "content": " is padded to the target resolution before the masked merging." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 552, + 233, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 233, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 233, + 567 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 582, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 479, + 665 + ], + "type": "text", + "content": "This section presents MOTIA, a method demonstrating exceptional performance in video outpainting tasks. We begin by defining the concept of video outpainting and describing the foundational model in Section 4.1. and Section 4.2. Subsequently, we delve into the specifics of input-specific adaptation and pattern-aware outpainting in Sections 4.3 and 4.4, respectively. Moreover, we show that our approach has great promise in extending its application to long video outpainting, which will be explored in Section 4.5." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be-Your-Outpainter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 465, + 388 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 465, + 388 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 465, + 388 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 465, + 388 + ], + "type": "image", + "image_path": "998d63b3de30b762524377e7f4ca953c13b7ee17ff387a654343256b1558191c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 396, + 482, + 430 + ], + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 430 + ], + "type": "text", + "content": "Fig. 3: Sample results of quantitative experiments. All videos are outpainted with a horizontal mask ratio of 0.66. Contents outside the yellow lines are outpainted by MOTIA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 453, + 268, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 268, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 268, + 464 + ], + "type": "text", + "content": "4.1 Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": "For a video represented as " + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\pmb{v} \\in \\mathbb{R}^{t \\times h \\times w \\times d}" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": " denotes the number of frames, " + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": " denotes frame height, " + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": " denotes frame width, and " + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": " denotes channel depth. Video outpainting model " + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "f(\\pmb{v})" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": " is designed to generate a spatially expanded video " + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\pmb{v}' \\in \\mathbb{R}^{t \\times h' \\times w' \\times d}" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": ". This process not only increases the spatial dimensions (" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "h' > h" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "w' > w" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": "), but also requires to ensure continuity and harmony between the newly expanded regions and the known regions. The transformation maintains the known regions unchanged, with " + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "inline_equation", + "content": "f(\\pmb{v})" + }, + { + "bbox": [ + 130, + 472, + 482, + 557 + ], + "type": "text", + "content": " acting as an identity in these regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 574, + 259, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 259, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 259, + 586 + ], + "type": "text", + "content": "4.2 Network Expansion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 487, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 487, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 487, + 666 + ], + "type": "text", + "content": "Network inflation. MOTIA leverages the pre-trained text-to-image (T2I) model, Stable Diffusion. In line with previous video editing techniques [32], we transform 2D convolutions into pseudo 3D convolutions and adapt 2D group normalizations into 3D group normalizations to process video latent features. Specifically, the " + }, + { + "bbox": [ + 130, + 594, + 487, + 666 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 594, + 487, + 666 + ], + "type": "text", + "content": " kernels in convolutions are replaced by " + }, + { + "bbox": [ + 130, + 594, + 487, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 3 \\times 3" + }, + { + "bbox": [ + 130, + 594, + 487, + 666 + ], + "type": "text", + "content": " kernels, maintaining identical weights. Group normalizations are executed across both temporal" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "Fu-Yun Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "and spatial dimensions, meaning that all 3D features within the same group are normalized simultaneously, followed by scaling and shifting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "type": "text", + "content": "Masked video as conditional input. Additionally, we incorporate a ControlNet [36], initially trained for image inpainting, to manage additional mask inputs. Apart from noise input, ControlNet can also process masked videos to extract effective information for more controllable denoising. In these masked videos, known regions have pixel values ranging from 0 to 1, while values of masked regions are set to " + }, + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 309 + ], + "type": "text", + "content": "Temporal consistency prior. To infuse the model with temporal consistency priors, we integrate temporal modules pre-trained on text-to-video (T2V) tasks. Note that although MOTIA relies on pre-trained video diffusion modules, applying these pre-trained temporal modules directly for video outpainting yields rather bad results, significantly inferior to all baseline methods (Table. 2). However, when equipped with our proposed MOTIA, the model demonstrates superior performance even in comparison to models specifically designed and trained for video outpainting, underscoring the efficacy of MOTIA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 329, + 291, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 291, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 291, + 341 + ], + "type": "text", + "content": "4.3 Input-Specific Adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 479, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 479, + 422 + ], + "type": "text", + "content": "The input-specific adaptation phase is crucial in our video outpainting method, aiming to tailor the model for the specific challenges of outpainting. This phase involves training on the source video with a pseudo-outpainting task, importantly, enabling the model to learn intrinsic content and motion patterns (data-specific patterns) within the source video as well as narrowing the gap between the standard generation process and outpainting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 481, + 495 + ], + "type": "text", + "content": "Video augmentation. Initially, we augment the source video. Transformations like identity transformation, random flipping, cropping, and resizing can be employed. This step can potentially help the model better learn and adapt to diverse changes in video content. For longer video outpainting, as we will discuss later, instead of taking it as a whole, we randomly sample short video clips from it to reduce the cost of the adaptation phase." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 481, + 567 + ], + "type": "text", + "content": "Video masking. We then add random masks to the video. We adopt a heuristic approach that uniformly samples edge boundaries of 4 sides within given limits. The area enclosed by these boundaries is considered the known region, while the rest is the unknown region. This masked video serves as the conditional input for the ControlNet, simulating the distribution of known and unknown areas in actual outpainting scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 567, + 486, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 486, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 486, + 603 + ], + "type": "text", + "content": "Video noising. Additionally, we apply noise to the video following the DDPM [10] by randomly selecting diffusion timesteps. This noisy video serves as an input for both the ControlNet and the Stable Diffusion model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 604, + 481, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 481, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 481, + 651 + ], + "type": "text", + "content": "Optimization. Finally, we optimize the model. To ensure efficiency, low-rank adapters are inserted into the layers of the diffusion model. We optimize only the parameters of these adapters while keeping the other parameters frozen. The loss function is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 220, + 652, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 652, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 220, + 652, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\| \\boldsymbol {\\epsilon} - \\hat {\\boldsymbol {\\epsilon}} _ {\\bar {\\theta} _ {l}, \\bar {\\theta} _ {c}, \\boldsymbol {\\theta} _ {a}} (\\boldsymbol {v} _ {\\mathrm {n o i s y}}, \\boldsymbol {v} _ {\\mathrm {m a s k e d}}, t) \\right\\| _ {2}, \\tag {5}", + "image_path": "b521f1ffcbc801bd7ded893d9b8654b30160fc4e80b38cedf7f46127350d3ae7.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be-Your-Outpainter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 134, + 264, + 199 + ], + "blocks": [ + { + "bbox": [ + 133, + 134, + 264, + 199 + ], + "lines": [ + { + "bbox": [ + 133, + 134, + 264, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 134, + 264, + 199 + ], + "type": "image", + "image_path": "81d223586e6d6f8f65400938a9f42153d292da8e22227d435f3d96639c29818c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 208, + 266, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 208, + 266, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 208, + 266, + 262 + ], + "type": "text", + "content": "Fig.4: Spatial-aware insertion scales the insertion weights of adapters for better leveraging of learned patterns and generative prior." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 278, + 115, + 481, + 198 + ], + "blocks": [ + { + "bbox": [ + 278, + 115, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 278, + 115, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 278, + 115, + 481, + 198 + ], + "type": "image", + "image_path": "f8e59ec18a3fc9a843bde6c668506280c2883f48460e5e041a0d17da4dba2391.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 207, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 276, + 207, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 276, + 207, + 482, + 273 + ], + "type": "text", + "content": "Fig. 5: Failure example of previous methods. Many previous methods including the intensively trained models on video outpainting still might suffer from generation failure, that the model simply generates blurred corners. MOTIA never encounters this failure." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": " represents the timestep in the process, " + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": " is the added noise, " + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{\\mathrm{noisy}}" + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": " refers to the video perturbed by " + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{\\mathrm{masked}}" + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": " denotes the masked video. The parameters " + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\theta_{c}" + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\theta_{a}" + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": " correspond to the Diffusion Model, ControlNet, and adapters, respectively. The bar over these parameters indicates they are frozen during the optimization. This optimization process, including the steps of augmentation, masking, and noising, is repeated to update the lightweight adapters to capture the data-specific patterns from the source video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 409, + 299, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 299, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 299, + 422 + ], + "type": "text", + "content": "4.4 Pattern-Aware Outpainting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "text", + "content": "Following the initial phase of input-specific adaptation, our model shows promising results in video outpainting using basic pipelines as outlined in Eq. 3 and Eq. 4, achieving commendable quality. However, we here introduce additional inference strategies that can be combined to better leverage the learned data-specific patterns from the input-specific adaptation phase for better outpainting results. We call the outpainting process that incorporates these strategies pattern-aware outpainting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Spatial-aware insertion. It is important to acknowledge that in the input-specific adaptation phase, the model is fine-tuned through learning outpainting within the source video. However, at the outpainting phase, the model is expected to treat the entire source video as known regions and then fill the unknown regions at edges (i.e., generating a video with a larger viewpoint and resolution). This specificity may lead to a noticeable training-inference gap during outpainting, potentially affecting the outpainting quality. To balance the fine-tuned patterns with the diffusion model's inherent generative prior, we introduce the concept of spatial-aware insertion (SA-Insertion) of adapters as shown in Fig. 4. The adaptation involves adjusting the insertion weight of tuned low-rank adapters based on the feature's spatial position. We increase insertion weight near known areas to utilize the learned patterns while decreasing it in farther regions to rely" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "Fu-Yun Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 477, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 477, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 477, + 128 + ], + "type": "text", + "content": "more on the original generative capacity of the diffusion model. To be specific," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 200, + 137, + 481, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 137, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 200, + 137, + 481, + 152 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {\\mathrm {a d a p t e d}} ^ {\\top} \\boldsymbol {x} _ {\\boldsymbol {p}} = \\mathbf {W} ^ {\\top} \\boldsymbol {x} _ {\\boldsymbol {p}} + \\alpha (\\boldsymbol {p}) \\left(\\mathbf {W} _ {\\mathrm {u p}} \\mathbf {W} _ {\\mathrm {d o w n}}\\right) ^ {\\top} \\boldsymbol {x} _ {\\boldsymbol {p}}. \\tag {6}", + "image_path": "edcc0549b3513816e891cc7a37459bf0fdfd5bac7b1e514838fac1a47ecb7143.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "spans": [ + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "text", + "content": " signifies the spatial position of " + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times d_{\\mathrm{out}}}" + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "text", + "content": " denotes the linear transformation in layers of diffusion model, " + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{down}} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times r}" + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{up}} \\in \\mathbb{R}^{r \\times d_{\\mathrm{out}}}" + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "text", + "content": " are the linear components of the adapter with rank " + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d_{\\mathrm{in}}, d_{\\mathrm{out}})" + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "inline_equation", + "content": "\\alpha(\\pmb{p})" + }, + { + "bbox": [ + 131, + 162, + 479, + 211 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 239, + 220, + 481, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 220, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 239, + 220, + 481, + 247 + ], + "type": "interline_equation", + "content": "\\alpha (\\boldsymbol {p}) = \\exp \\left(- \\frac {K \\| \\boldsymbol {p} - \\boldsymbol {p} _ {c} \\|}{\\operatorname* {m a x} _ {\\bar {\\boldsymbol {p}}} \\| \\bar {\\boldsymbol {p}} - \\boldsymbol {p} _ {c} \\|}\\right), \\tag {7}", + "image_path": "62d4e93f857f2cb1f3accb40b737e691fdb0c0fd7f62e36649691fedf2a65cc6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 255, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 479, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 255, + 479, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 255, + 479, + 277 + ], + "type": "text", + "content": " is a constant for controlling decay speed, and " + }, + { + "bbox": [ + 131, + 255, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 131, + 255, + 479, + 277 + ], + "type": "text", + "content": " represents the nearest side of the known region to " + }, + { + "bbox": [ + 131, + 255, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 131, + 255, + 479, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 279, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 279, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 480, + 396 + ], + "type": "text", + "content": "Noise regret. In the context of Eq. 3, merging noisy states from known and unknown regions in video outpainting tasks is similar to sampling from two different vectors, which can potentially disrupt the denoising direction. Given the significant impact of early steps on the generation's structure, later denoising may not rectify these initial discrepancies. Inspired by DDPM-based image inpainting methods [14,21], we propose to re-propagate the noisy state into a noisier state by adding noise when denoising and then provide the model a second chance for re-denoising. This helps integrate known region data more effectively and reduces denoising direction conflicts. In detail, after obtaining " + }, + { + "bbox": [ + 131, + 279, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 131, + 279, + 480, + 396 + ], + "type": "text", + "content": " during denoising, we conduct" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 214, + 397, + 481, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 397, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 214, + 397, + 481, + 417 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {t + L} = \\sqrt {\\Pi_ {i = t + 1} ^ {t + L} \\alpha_ {i}} \\boldsymbol {v} _ {t} + \\sqrt {1 - \\Pi_ {i = t + 1} ^ {t + L} \\alpha_ {i}} \\boldsymbol {\\epsilon}, \\tag {8}", + "image_path": "11ed1c72ee23d0605b1b75b6b4325a987913d91d8449cb5f7cda01bf41152730.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 422, + 479, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 479, + 447 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 479, + 447 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 422, + 479, + 447 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = 1 - \\beta_{i}" + }, + { + "bbox": [ + 131, + 422, + 479, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 422, + 479, + 447 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 131, + 422, + 479, + 447 + ], + "type": "text", + "content": ". Then we restart the denoising process. We repeat this progress for " + }, + { + "bbox": [ + 131, + 422, + 479, + 447 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 422, + 479, + 447 + ], + "type": "text", + "content": " times. We only conduct it in the early denoising steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 466, + 350, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 350, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 350, + 478 + ], + "type": "text", + "content": "4.5 Extension to Long Video Outpainting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 487, + 479, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 479, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 479, + 630 + ], + "type": "text", + "content": "We show that our method can be easily extended for long video outpainting. Specifically, for the stage of input-specific adaptation, instead of taking the long video as a whole for adaptation (Direct adaptation on long videos is costly and does not align with the video generation prior of the pretrained modules), we randomly sample short video clips from the long video for tuning to learn global patterns without requiring more GPU memory cost. For the stage of pattern-aware outpainting, we split the long video into short video clips with temporal overlapping (i.e., some frames are shared by different short video clips), and then conduct temporal co-denoising following Gen-L [29]. Specifically, the denoising result for " + }, + { + "bbox": [ + 131, + 487, + 479, + 630 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 131, + 487, + 479, + 630 + ], + "type": "text", + "content": " frame of the long video at timestep " + }, + { + "bbox": [ + 131, + 487, + 479, + 630 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 487, + 479, + 630 + ], + "type": "text", + "content": " is approximated by the weighted sum of all the corresponding frames in short video clips that contain it," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 225, + 632, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 632, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 225, + 632, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {t - 1, j} = \\frac {\\sum_ {i \\in \\mathcal {I} ^ {j}} \\left(\\left(W _ {i , j ^ {*}}\\right) ^ {2} \\otimes \\boldsymbol {v} _ {t - 1 , j ^ {*}} ^ {i}\\right)}{\\sum_ {i \\in \\mathcal {I} ^ {j}} \\left(W _ {i , j ^ {*}} ^ {2}\\right) ^ {2}}, \\tag {9}", + "image_path": "37c8b0eefb1fcaa4ee4fef7f1a04275f1d3270f03a471733349143cfd8f4c615.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be-Your-Outpainter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 115, + 455, + 246 + ], + "blocks": [ + { + "bbox": [ + 161, + 115, + 455, + 246 + ], + "lines": [ + { + "bbox": [ + 161, + 115, + 455, + 246 + ], + "spans": [ + { + "bbox": [ + 161, + 115, + 455, + 246 + ], + "type": "image", + "image_path": "a9eb8b1e7124d465cde9d8f138a9a0a1feccd55ca0fb9a4d7dc3b6b9783ed94b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 256, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 256, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 482, + 291 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison. Other methods outpainting the source video with a mask ratio of 0.6. MOTIA outpainting the source video with a larger mask ratio of 0.66 while achieving obviously better outpainting results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": " denotes element-wise multiplication, " + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{t-1,j}" + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": " denotes the noisy state of the " + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": " frame at timestep " + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{t-1,j^*}^i" + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": " is the noisy state of " + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": " frame predicted with only information from the " + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": " video clip at timestep " + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "inline_equation", + "content": "W_{i,j^*}" + }, + { + "bbox": [ + 130, + 313, + 482, + 363 + ], + "type": "text", + "content": " is the per-pixel weight, which is as 1 as default." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 381, + 230, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 230, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 230, + 395 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 406, + 261, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 261, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 261, + 418 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 481, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 473 + ], + "type": "text", + "content": "Benchmarks. To verify the effectiveness of MOTIA, we conduct evaluations on DAVIS [19] and YouTube-VOS [33], which are widely used benchmarks for video outpainting. Following M3DDM [6], we compare the results of different methods in the horizontal direction, using mask ratios of 0.25 and 0.66." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 569 + ], + "type": "text", + "content": "Evaluation metrics. Our evaluation approach utilizes four well-established metrics: Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [31], Learned Perceptual Image Patch Similarity (LPIPS) [37], and Frechet Video Distance (FVD) [27]. For assessing PSNR, SSIM, and FVD, the generated videos are converted into frames within a normalized value range of " + }, + { + "bbox": [ + 130, + 474, + 481, + 569 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 130, + 474, + 481, + 569 + ], + "type": "text", + "content": ". LPIPS is evaluated over a range of " + }, + { + "bbox": [ + 130, + 474, + 481, + 569 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 130, + 474, + 481, + 569 + ], + "type": "text", + "content": ". About the FVD metric, we adopt a uniform frame sampling, with 16 frames per video for evaluation following M3DDM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "Implementation details. Our method is built upon Stable Diffusion v1-5. We add the ControlNet pretrained on image inpainting to enable the model to accept additional masked image inputs. The temporal modules are initialized with the weights from pretrained motion modules [8] to obtain additional motion priors. The motion modules are naive transformer blocks trained with solely text-to-video tasks on WebVid. For the input-specific adaptation, the low-rank adapters are trained using the Adam optimizer. We set the learning rate to " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": ", and set the weight decay to " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": ". The LoRA rank and " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{lora}}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " are set to 16 and," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "Fu-Yun Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 157, + 463, + 219 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of video outpainting methods on DAVIS and YouTube-VOS datasets. " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " means 'better when higher', and " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " indicates 'better when lower'." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 157, + 463, + 219 + ], + "lines": [ + { + "bbox": [ + 148, + 157, + 463, + 219 + ], + "spans": [ + { + "bbox": [ + 148, + 157, + 463, + 219 + ], + "type": "table", + "html": "
MethodDAVISYouTube-VOS
PSNR↑ SSIM↑ LPIPS↓ FVD↓PSNR↑ SSIM↑ LPIPS↓ FVD
VideoOutpainting [5]17.960.62720.2331363.118.250.71950.2278149.7
SDM [6]20.020.70780.2165334.619.910.72770.200194.81
M3DDM [6]20.260.70820.2026300.020.200.73120.185466.62
MOTIA20.360.75780.1595286.320.250.76360.172758.99
", + "image_path": "c5cfdf1bd461884d3d45c3494161373a26d177fefeede2f936b84e55171001e5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "text", + "content": "8, respectively. The number of training steps is set to 1000. We do not apply augmentation for simplicity. For both mask ratios of 0.66 and 0.25, we simply apply the same random mask strategy, which uniformly crops a square in the middle as the known regions. For the pattern-aware outpainting, the diffusion steps are set to 25 and the classifier-free guidance (CFG) [11, 23] scale is set to 7.5 and we only apply CFG at the first 15 inference steps. When adding noise regret to further improve the results, we set jump length " + }, + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "text", + "content": ", and repeat time " + }, + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "text", + "content": ". We only apply noise regret in the first half inference steps. Note that our method is built upon LDM, which requires text-conditional inputs. For a fair comparison and to remove the influence of the choice of text prompt, we apply Blip [13] to select the prompt automatically. We observe dozens of prompt mistakes but do not modify them to avoid man-made influence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 402, + 281, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 281, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 281, + 415 + ], + "type": "text", + "content": "5.2 Qualitative Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": "Fig. 6 showcases a qualitative comparison of MOTIA against other methods. Outpainting a narrow video into a square format. MOTIA employs a mask ratio of 0.66, surpassing the 0.6 ratio utilized by other methods, and demonstrates superior performance even with this higher mask ratio. The SDM method only manages to blur the extremities of the video's background, egregiously overlooking the primary subject and resulting in the outpainting failure as previously highlighted in Fig. 5. Dehan's approach effectively outpaints the background but utterly fails to address the foreground, leading to notable distortions. In contrast, the M3DDM method adeptly handles both subject and background integration but is marred by considerable deviations in subject characteristics, such as incorrect brown coloration in the dog's fur across several frames. Our method stands out by achieving optimal results, ensuring a harmonious and consistent outpainting of both the foreground and background." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 597, + 288, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 288, + 609 + ], + "type": "text", + "content": "5.3 Quantitative Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Table 1 summarizes the evaluation metrics of our method compared to other approaches. Our method achieves comparable results to the best method in PSNR. It shows significant improvements in video quality (SSIM), perceptual metric (LPIPS), and distribution similarity (FVD). Specifically, our SSIM, LPIPS," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be-Your-Outpainter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 175, + 114, + 239, + 178 + ], + "blocks": [ + { + "bbox": [ + 175, + 114, + 239, + 178 + ], + "lines": [ + { + "bbox": [ + 175, + 114, + 239, + 178 + ], + "spans": [ + { + "bbox": [ + 175, + 114, + 239, + 178 + ], + "type": "image", + "image_path": "9b3e1be9256d00d767c65f89419a7a65f45b503849d5caf919ab908ad9c514f6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 181, + 218, + 189 + ], + "lines": [ + { + "bbox": [ + 197, + 181, + 218, + 189 + ], + "spans": [ + { + "bbox": [ + 197, + 181, + 218, + 189 + ], + "type": "inline_equation", + "content": "SD + T" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 115, + 306, + 178 + ], + "blocks": [ + { + "bbox": [ + 242, + 115, + 306, + 178 + ], + "lines": [ + { + "bbox": [ + 242, + 115, + 306, + 178 + ], + "spans": [ + { + "bbox": [ + 242, + 115, + 306, + 178 + ], + "type": "image", + "image_path": "6167695fb83bfd4ce1b7d9aef173f38a4c90f078a3210a414066c017fb5ffa5d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 181, + 287, + 189 + ], + "lines": [ + { + "bbox": [ + 257, + 181, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 257, + 181, + 287, + 189 + ], + "type": "inline_equation", + "content": "SD + T + C" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 372, + 178 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 372, + 178 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 372, + 178 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 372, + 178 + ], + "type": "image", + "image_path": "8f6d01ebfdc090075a8da7817199060153f0d5da6b92ea64f899dca46a7ea6d0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 181, + 359, + 190 + ], + "lines": [ + { + "bbox": [ + 318, + 181, + 359, + 190 + ], + "spans": [ + { + "bbox": [ + 318, + 181, + 359, + 190 + ], + "type": "text", + "content": "Direct-tune" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 181, + 445, + 190 + ], + "lines": [ + { + "bbox": [ + 369, + 181, + 445, + 190 + ], + "spans": [ + { + "bbox": [ + 369, + 181, + 445, + 190 + ], + "type": "text", + "content": "Input-Specific Adaptation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 376, + 115, + 440, + 178 + ], + "blocks": [ + { + "bbox": [ + 376, + 115, + 440, + 178 + ], + "lines": [ + { + "bbox": [ + 376, + 115, + 440, + 178 + ], + "spans": [ + { + "bbox": [ + 376, + 115, + 440, + 178 + ], + "type": "image", + "image_path": "1c63d74782cf2f2447a94af9ad250041469872d1597d57522ce75ed9ae2898ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 168, + 227, + 445, + 274 + ], + "blocks": [ + { + "bbox": [ + 132, + 201, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 201, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 479, + 213 + ], + "type": "text", + "content": "Fig. 7: Visual examples of ablation study on the proposed input-specific adaptation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 227, + 445, + 274 + ], + "lines": [ + { + "bbox": [ + 168, + 227, + 445, + 274 + ], + "spans": [ + { + "bbox": [ + 168, + 227, + 445, + 274 + ], + "type": "image", + "image_path": "87b4dad41111d6cb1172867e35621992308b67e104485b01d3d236c9a7cd623b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 275, + 250, + 282 + ], + "lines": [ + { + "bbox": [ + 176, + 275, + 250, + 282 + ], + "spans": [ + { + "bbox": [ + 176, + 275, + 250, + 282 + ], + "type": "text", + "content": "Direct (Ours " + }, + { + "bbox": [ + 176, + 275, + 250, + 282 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 176, + 275, + 250, + 282 + ], + "type": "text", + "content": " SA + NR)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 280, + 275, + 332, + 283 + ], + "lines": [ + { + "bbox": [ + 280, + 275, + 332, + 283 + ], + "spans": [ + { + "bbox": [ + 280, + 275, + 332, + 283 + ], + "type": "text", + "content": "SA (Ours w/o NR)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 275, + 436, + 284 + ], + "lines": [ + { + "bbox": [ + 358, + 275, + 436, + 284 + ], + "spans": [ + { + "bbox": [ + 358, + 275, + 436, + 284 + ], + "type": "text", + "content": "Pattern-Aware Outpainting" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 294, + 457, + 306 + ], + "lines": [ + { + "bbox": [ + 153, + 294, + 457, + 306 + ], + "spans": [ + { + "bbox": [ + 153, + 294, + 457, + 306 + ], + "type": "text", + "content": "Fig. 8: Visual examples of ablation study on pattern-aware outpainting." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": "and FVD metrics show improvements of " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "7.00\\%" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "21.27\\%" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "4.57\\%" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": " respectively on the DAVIS dataset, and " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "4.43\\%" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "6.85\\%" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "11.45\\%" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": " on the YouTube-VOS dataset compared to the best-performing method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 383, + 238, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 383, + 238, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 383, + 238, + 395 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 402, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 533 + ], + "type": "text", + "content": "Ablation study on input-specific adaptation. We conducted the ablation study on input-specific adaptation with the DAVIS dataset to verify its effectiveness, as shown in Fig. 7 and Table 2. \"SD+T\" represents the result of directly combining the temporal module with Stable Diffusion, which led to a complete outpainting failure. \"SD+T+C\" indicates the additional use of ControlNet, resulting in similarly poor outcomes. \"Direct-tune\" refers to the approach of directly fitting the original video without outpainting training; in this case, we observed a very noticeable color discrepancy between the outpainted and known areas. In contrast, our method achieved the best results, ensuring consistency in both the visual and temporal aspects. The metrics shown in Table 2 also support this observation, with MOTIA significantly outperforming the other baselines." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": "Ablation study on pattern-aware outpainting. Table 3 summarizes our ablation experiments for the pattern-aware outpainting part. We conducted extensive validation on the YouTube-VOS dataset. \"Direct\" refers to performing outpainting according to Eq. 3 directly after input-specific adaptation. \"SA\" denotes spatially-aware insertion, and \"SA+NR\" indicates the combined use of spatially-aware insertion and noise regret. The experimental results demonstrate that each of our components effectively enhances performance. Specifically, Combining both SA-Insertion and Noise regret, the PSNR, SSIM, LPIPS, and FVD metrics show improvements of " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "2.69\\%" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "0.90\\%" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "3.95\\%" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "11.32\\%" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": " respectively than directly applying Eq. 3. Fig. 8 presents the visual examples of ablation study on our proposed pattern-aware outpainting part. When removing NR, it" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "Fu-Yun Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 309, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 291, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 291, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 291, + 137 + ], + "type": "text", + "content": "Table 2: Ablation study on input-specific adaptation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 309, + 216 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 309, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 309, + 216 + ], + "type": "table", + "html": "
ComponentPSNR↑SSIM↑LPIPS↓FVD ↓
SD+T15.590.66400.2979672.7
SD+T+C16.810.69610.2338515.4
Direct-tune19.420.73750.1784312.1
MOTIA20.360.75780.1595286.3
", + "image_path": "3dbf9997f164bd3b9566408a353b3c41c5c61feab2a7c2b1fade71e1ae958b3c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 312, + 146, + 489, + 205 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 3: Ablation study on the proposed pattern-aware outpainting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 146, + 489, + 205 + ], + "lines": [ + { + "bbox": [ + 312, + 146, + 489, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 146, + 489, + 205 + ], + "type": "table", + "html": "
ComponentPSNR↑SSIM↑LPIPS↓FVD ↓
Direct19.720.75680.179866.52
SA19.970.76080.175258.40
SA+NR20.250.76360.172758.99
", + "image_path": "218302381bb7b0dff041a475caba83f74ddb2d6bfc24988b0d09ad46d7daff22.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "content": "might fail to align the texture colors or produce unreasonable details (e.g., arms in the middle of Fig. 8). When further removing SA, it could potentially generate unrealistic results caused by the overfitting to the target video (e.g., the white collar on the left of Fig. 8). Note that even though the FVD degrades in a very slight manner, all the other metrics increase and we qualitatively find it to be helpful for improving results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 324, + 218, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 218, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 218, + 335 + ], + "type": "text", + "content": "5.5 Discussions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 485 + ], + "type": "text", + "content": "Model and computation complexity. Model Complexity: The original model has 1.79 billion (including the auto-encoder and text encoder) parameters in total, while the added adapters contain 7.49 million parameters, leading to an increase of " + }, + { + "bbox": [ + 130, + 342, + 482, + 485 + ], + "type": "inline_equation", + "content": "0.42\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 485 + ], + "type": "text", + "content": " in memory usage. Computation Complexity: We report the peak GPU VRAM and the time required for outpainting a target video from " + }, + { + "bbox": [ + 130, + 342, + 482, + 485 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 342, + 482, + 485 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 342, + 482, + 485 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 130, + 342, + 482, + 485 + ], + "type": "text", + "content": " with 16 frames at two stages in Table 4. For longer videos, as described in Section 4.5, instead of processing the long video as a whole, we adapt only to short video clips sampled from the long video. This approach does not require additional time or GPU VRAM during the input-specific adaptation phase. Additionally, with temporal co-denoising [29], the GPU VRAM usage remains the same as that for short video during the pattern-aware outpainting phase, while the required time increases linearly with the video length." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "text", + "content": "User study. We conducted a user study between MOTIA and M3DDM, utilizing the DAVIS dataset with a horizontal mask of 0.66 as source videos. Preferences were collected from 10 volunteers, each evaluating 50 randomly selected sets of results based on visual quality (such as clarity, color fidelity, and texture detail) and realism (including motion consistency, object continuity, and integration with the background). Table 2 demonstrates that the outputs from MOTIA are preferred over those from M3DDM in both visual quality and realism." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Why MOTIA outperforms (Why previous methods fail). 1) Flexibility. Current video diffusion models are mostly trained with fixed resolution and length, lacking the ability to tackle videos with various aspect ratios and lengths. In contrast, the adaptation phase of MOTIA allows the model to better capture the size, length, and style distribution of the source video, greatly narrowing the gap between pretrained weights and the source video. 2) Ability for capturing intrinsic patterns from source video. A crucial point for successful outpainting is the predicted score of diffusion models should be well-compatible with the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be-Your-Outpainter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 136, + 301, + 198 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 290, + 126 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 290, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 290, + 126 + ], + "type": "text", + "content": "Table 4: Computation complexity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 136, + 301, + 198 + ], + "lines": [ + { + "bbox": [ + 132, + 136, + 301, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 301, + 198 + ], + "type": "table", + "html": "
PhaseVRAM↓Time↓
Adaptation12.70 GB309 Seconds
Outpainting5.80 GB58 Seconds
MOTIA (In total)12.70 GB367 Seconds
", + "image_path": "b4d8191f70eaa241395c039dff944f17d932e5b42568860d49c2fd51c9ae39af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 334, + 146, + 470, + 198 + ], + "blocks": [ + { + "bbox": [ + 322, + 114, + 480, + 136 + ], + "lines": [ + { + "bbox": [ + 322, + 114, + 480, + 136 + ], + "spans": [ + { + "bbox": [ + 322, + 114, + 480, + 136 + ], + "type": "text", + "content": "Table 5: User study comparison between M3DDM and MOTIA." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 146, + 470, + 198 + ], + "lines": [ + { + "bbox": [ + 334, + 146, + 470, + 198 + ], + "spans": [ + { + "bbox": [ + 334, + 146, + 470, + 198 + ], + "type": "table", + "html": "
MethodVisual-QualityRealism
M3DDM27.4%42.8%
MOTIA72.6%57.2%
", + "image_path": "324fe69e918e2419ab7ced5b623bdb634c452f1a3d3ff48637db352969c55ed0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 221, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 366 + ], + "type": "text", + "content": "original known regions of the source video. To achieve this, the model should effectively extract useful information from the source video for denoising. For instance, M3DDM concatenates local frames of source video at the input layers and incorporates the global frames through the cross-attention mechanism after passing light encoders. However, the information might not be properly handled especially for out-domain inputs, thus leading to outpainting failure. Instead, by conducting input-specific adaptation on the source video, the model can effectively capture the data-specific patterns in the source videos through gradient. Through this, MOTIA better leverage the data-specific patterns of the source video and image/video generative prior for outpainting. We hope this work inspires following research to exploit more from the source video itself instead of purely relying on the generative prior from intensive training on videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 387, + 219, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 219, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 219, + 399 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 481, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 463 + ], + "type": "text", + "content": "We present MOTIA, relying on a combination of input-specific adaptation for capturing inner video patterns and pattern-aware outpainting to generalize these patterns for effective outpainting. Extensive experiments validate the effectiveness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 464, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 536 + ], + "type": "text", + "content": "Limitations: MOTIA requires learning necessary patterns from the source video, it faces limitation for outpainting videos with limited information. We provide an example of failure case. In Fig. 9, the original source video just contains grass and a woman with a blue shirt. We find our method produces a woman with very similar appearance when outpainting, which might be regarded as a \"copy\" phenomenon." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 176, + 558, + 261, + 643 + ], + "blocks": [ + { + "bbox": [ + 176, + 558, + 261, + 643 + ], + "lines": [ + { + "bbox": [ + 176, + 558, + 261, + 643 + ], + "spans": [ + { + "bbox": [ + 176, + 558, + 261, + 643 + ], + "type": "image", + "image_path": "88f31870890111efaa882087367b78e385d684c45c3eb90966c0a182986cd2bc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 651, + 465, + 662 + ], + "lines": [ + { + "bbox": [ + 146, + 651, + 465, + 662 + ], + "spans": [ + { + "bbox": [ + 146, + 651, + 465, + 662 + ], + "type": "text", + "content": "Fig. 9: Example of failure case. Contents outside the red lines are outpainted." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 264, + 558, + 349, + 643 + ], + "blocks": [ + { + "bbox": [ + 264, + 558, + 349, + 643 + ], + "lines": [ + { + "bbox": [ + 264, + 558, + 349, + 643 + ], + "spans": [ + { + "bbox": [ + 264, + 558, + 349, + 643 + ], + "type": "image", + "image_path": "06eb5ae5a9ca15710ed566732a9752e4649bf0cea7ad81e32b85908cf56b775c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 353, + 558, + 437, + 643 + ], + "blocks": [ + { + "bbox": [ + 353, + 558, + 437, + 643 + ], + "lines": [ + { + "bbox": [ + 353, + 558, + 437, + 643 + ], + "spans": [ + { + "bbox": [ + 353, + 558, + 437, + 643 + ], + "type": "image", + "image_path": "a52662512af143c2459994acbfb8d0c2101461270dd2a98ec61b0d9992497269.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "Fu-Yun Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 215, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 215, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 215, + 129 + ], + "type": "text", + "content": "Acknowledgment:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "type": "text", + "content": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR, by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be-Your-Outpainter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Arora, R., Lee, Y.J.: Singan-gif: Learning a generative video model from a single gif. In: CVPR. pp. 1310-1319 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "text", + "content": "2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: CVPR. pp. 18208-18218 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 480, + 213 + ], + "type": "text", + "content": "3. Barnes, C., Shechtman, E., Finkelstein, A., Goldman, D.B.: Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph. 28(3), 24 (2009)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "text", + "content": "4. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR. pp. 22563-22575 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 246, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 266 + ], + "type": "text", + "content": "5. Dehan, L., Van Ranst, W., Vandewalle, P., Goedemé, T.: Complete and temporally consistent video outpainting. In: CVPRW. pp. 687-695 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 267, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 298 + ], + "type": "text", + "content": "6. Fan, F., Guo, C., Gong, L., Wang, B., Ge, T., Jiang, Y., Luo, C., Zhan, J.: Hierarchical masked 3d diffusion model for video outpainting. In: ACM MM. pp. 7890-7900 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "text", + "content": "7. Gao, C., Saraf, A., Huang, J.B., Kopf, J.: Flow-edge guided video completion. In: ECCV. pp. 713-729. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "text", + "content": "8. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: Animatediff: animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "type": "text", + "content": "9. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 386, + 480, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 406 + ], + "type": "text", + "content": "0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "type": "text", + "content": "1. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 428, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 449 + ], + "type": "text", + "content": "2. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. arXiv:2204.03458 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 449, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 482 + ], + "type": "text", + "content": "3. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 483, + 480, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 513 + ], + "type": "text", + "content": "4. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR. pp. 11461-11471 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "text", + "content": "5. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 578 + ], + "type": "text", + "content": "6. Ma, Y., He, Y., Wang, H., Wang, A., Qi, C., Cai, C., Li, X., Li, Z., Shum, H.Y., Liu, W., et al.: Follow-your-click: Open-domain regional image animation via short prompts. arXiv preprint arXiv:2403.08268 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "text", + "content": "7. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 612, + 480, + 632 + ], + "type": "text", + "content": "8. Nikankin, Y., Haim, N., Irani, M.: Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "9. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "Fu-Yun Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 644 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "20. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "text", + "content": "21. Rout, L., Parulekar, A., Caramanis, C., Shakkottai, S.: A theoretical justification for image inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2302.01217 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "22. Shaham, T.R., Dekel, T., Michaeli, T.: Singan: Learning a generative model from a single natural image. In: ICCV. pp. 4570-4580 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "23. Shen, D., Song, G., Xue, Z., Wang, F.Y., Liu, Y.: Rethinking the spatial inconsistency in classifier-free diffusion guidance. In: CVPR. pp. 9370-9379 (June 2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 259 + ], + "type": "text", + "content": "24. Shi, X., Huang, Z., Wang, F.Y., Bian, W., Li, D., Zhang, Y., Zhang, M., Cheung, K.C., See, S., Qin, H., et al.: Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In: ACM SIGGRAPH 2024 Conference Papers. pp. 1-11 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "25. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "26. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "27. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "28. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. NeurIPS (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "29. Wang, F.Y., Chen, W., Song, G., Ye, H.J., Liu, Y., Li, H.: Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "30. Wang, F.Y., Huang, Z., Shi, X., Bian, W., Song, G., Liu, Y., Li, H.: Animatelem: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning. arXiv preprint arXiv:2402.00769 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "31. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "32. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "33. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "text", + "content": "34. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In: CVPR. pp. 1316-1324 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "35. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: CVPR. pp. 10459-10469 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "text", + "content": "36. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 621, + 481, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 644 + ], + "type": "text", + "content": "37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 102 + ], + "type": "text", + "content": "Be-Your-Outpainter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_content_list.json b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f05a1e8343336729748587ef67174adba62dcd7d --- /dev/null +++ b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_content_list.json @@ -0,0 +1,1637 @@ +[ + { + "type": "text", + "text": "BeNeRF: Neural Radiance Fields from a Single Blurry Image and Event Stream", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenpu Li $^{1,5*}$ , Pian Wan $^{1,2*}$ , Peng Wang $^{1,3*}$ , Jinghang Li $^{4}$ , Yi Zhou $^{4}$ , and Peidong Liu $^{1\\dagger}$", + "bbox": [ + 222, + 210, + 781, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Westlake University, 2EPFL, 3Zhejiang University, 4Hunan University, 5Guangdong University of Technology", + "bbox": [ + 241, + 253, + 759, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/wu-cvgl/BeNeRF", + "bbox": [ + 372, + 296, + 627, + 311 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da8780e08dedb3b058109ef09b8272e1f41c44c86471cf9d6d501ad222a46689.jpg", + "image_caption": [ + "Fig. 1: Given a single blurry image and its corresponding event stream, BeNeRF can synthesize high-quality novel images along the camera trajectory, recovering a sharp and coherent video from the single blurry image." + ], + "image_footnote": [], + "bbox": [ + 233, + 325, + 764, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Implicit scene representation has attracted a lot of attention in recent research of computer vision and graphics. Most prior methods focus on how to reconstruct 3D scene representation from a set of images. In this work, we demonstrate the possibility to recover the neural radiance fields (NeRF) from a single blurry image and its corresponding event stream. To eliminate motion blur, we introduce event stream to regularize the learning process of NeRF by accumulating it into an image. We model the camera motion with a cubic B-Spline in SE(3) space. Both the blurry image and the brightness change within a time interval, can then be synthesized from the NeRF given the 6-DoF poses interpolated from the cubic B-Spline. Our method can jointly learn both the implicit scene representation and the camera motion by minimizing the differences between the synthesized data and the real measurements without any prior knowledge of camera poses. We evaluate the proposed method with both synthetic and real datasets. The experimental results demonstrate that we are able to render view-consistent latent sharp images from the learned NeRF and bring a blurry image alive in high quality.", + "bbox": [ + 259, + 530, + 740, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Neural Radiance Fields $\\cdot$ Event Stream $\\cdot$ Pose Estimation $\\cdot$ Deblurring $\\cdot$ Novel View Synthesis $\\cdot$ 3D from a Single Image", + "bbox": [ + 259, + 777, + 740, + 806 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution: {liwenpu, wangpeng}@westlake.edu.cn, pianwan@outlook.com", + "bbox": [ + 218, + 811, + 777, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: Peidong Liu(liupeidong@westlake.edu.cn).", + "bbox": [ + 220, + 825, + 666, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) [28] has drawn much attention due to its extraordinary ability in representing 3D scenes and synthesizing novel views. Given multi-view sharp RGB images and calibrated camera poses from COLMAP [40], NeRF takes corresponding 3D spatial location and 2D view direction as input, and optimizes a multi-layer perception (MLP) to represent the 3D scene. More recent advanced methods also exploit explicit octree [52], multi-resolution hash encoding [30] etc., to represent the 3D scene to improve both the training and rendering efficiency.", + "bbox": [ + 212, + 174, + 785, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior methods usually rely on multi-view images to learn the 3D representation. Several pioneering works recently attempt to exploit a single image to learn the underlying neural radiance fields [3,36,38,53]. They usually rely on a large dataset to pre-train the networks to learn priors to address the ill-posed problem. A blurry image further aggravates the problem due to the image quality degradation. Although motion blur is usually not preferred by most vision algorithms, they actually encode additional camera motion trajectory and more structural information compared to a sharp image. In this paper, we explore the possibility of recovering the neural radiance fields and camera motion trajectory from a single blurry image. Instead of learning priors from a large dataset as in previous works, we exploit the usage of an additional event stream to better constrain the problem.", + "bbox": [ + 212, + 296, + 785, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event stream can be acquired by an event camera [23] which captures pixel intensity changes caused by the relative motion between the static scene and camera. Unlike standard frame-based cameras, event camera captures asynchronous events with very low latency, leading to extremely high temporal resolution [8]. This characteristic compensates with the image formation process of a blurry image (i.e. integral of photon measurements across time). Several prior works thus take advantage of both modalities for high quality single image deblurring [33,44,47]. However, these methods are unable to recover the camera motion trajectory and extract structural details from a single blurry image, thereby limiting their applicability in 3D computer vision tasks. Some NeRF-based methods that incorporate event stream [10,16,25,34,39] demonstrate the capability to achieve image deblurring and accurate reconstruction of neural radiance fields. Nonetheless, these methods necessitate input images from multiple viewpoints alongside event data. In contrast, we explore the usage of only a single blurry image for the NeRF recovery with unknown camera motions in this work.", + "bbox": [ + 212, + 478, + 785, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We represent the continuous camera motion with a cubic B-Spline in SE(3) space and define it as the trajectory of both frame-based camera and event camera. Given the neural 3D representation and interpolated poses from the cubic B-Spline, we can synthesize both the blurry image and the brightness change within a time interval via the physical image formation process. The NeRF and motion trajectory can then be jointly optimized by minimizing the difference between the synthesized data and the real measurements. To evaluate the performance of our method, we conduct experiments with both synthetic and real datasets. The experimental results demonstrate that our method is able to", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W. Li, P. Wan, P. Wang et al.", + "bbox": [ + 271, + 113, + 473, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "recover the neural radiance fields from a blurry image and its corresponding event stream without knowing prior knowledge of poses. We are thus able to render view-consistent latent sharp images encoded in a single blurred image from learned NeRF, effectively enhancing the quality of the blurry image. The experimental results further demonstrate that our method is even able to reach same performance as $\\mathrm{E}^2\\mathrm{NeRF}$ [34], which targets for the same problem, but with multi-view images and longer event data. In summary, our key contributions are as follows:", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a NeRF-based method that can recover the neural 3D representation from a single blurry image and its corresponding event stream, without knowing any ground truth poses;", + "- Our method is able to estimate the complex continuous trajectory of camera motion during the imaging process from a single blurry image, providing multi-view geometric information;", + "- We experimentally validate that our approach is able to recover high quality latent sharp images and high frame-rate video from a single blurry image, without any generalization issue. Furthermore, we are able to reach same level of performance as $\\mathrm{E}^2\\mathrm{NeRF}$ [34], which requires multi-view images and longer event data." + ], + "bbox": [ + 225, + 277, + 785, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 469, + 387, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We roughly categorize our related works into three main areas: neural implicit scene representation, single image deblurring and event-enhanced image deblurring.", + "bbox": [ + 212, + 503, + 785, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Implicit Scene Representation. NeRF has attracted lots of attention due to its powerful ability of implicit 3D scene representation and novel view synthesis [28]. Many following works have been proposed to improve NeRF's performance or extend NeRF to other fields. [2,24,49] jointly trained NeRF with inaccurate camera poses. [6,22,26,27,32,53] improved the performance of NeRF with degraded images, including noisy or few images etc. Recently, several event based NeRF [10,17,34,39] have also been proposed.", + "bbox": [ + 212, + 551, + 787, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will mainly focus on the methods that are most related to ours. BAD-NeRF [48] aims to recover true underlying 3D scene representation from multiview blurry inputs and inaccurate camera poses estimated from COLMAP [40]. However, BAD-NeRF and its variants [20,58] struggle to address situations where the input is limited to a single image, primarily due to the severe illness of the problem. Event-Enhanced NeRF (E²NeRF) [34] aims to recover the 3D scene representation from multiple blurry images and event streams. To train NeRF without optimizing the camera poses, E²NeRF manually segments the event stream using preset parameters, aiming to recover sharp images individually through EDI [33]. However, in cases of serve camera motion, the recovered images from event stream may still be blurry, leading to inaccurate pose estimation in COLMAP [40]. This two-stage approach introduces errors (i.e. either from", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BeNeRF", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "EDI [33] or COLMAP [40]) and fails to accurately model the continuous camera trajectory. In contrast, we are able to train both the NeRF and the camera motion trajectory jointly from a single blurry image and its corresponding event stream, which is harder than prior methods.", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single Image Deblurring. A blurred image can be formulated as the convolution result of a sharp image and a kernel. Therefore, classical approaches [11,14,21,42,51] generally regard the deblurring problem as a joint optimization of the blur kernel and the latent virtual sharp images. With the development of deep learning, many learning based end-to-end deblurring methods have been proposed [13, 18, 19, 31, 43, 46, 54]. These methods usually demonstrate better qualitative and quantitative deblurring results. However, learning based deblurring methods are usually trained on a large dataset which contains paired blurry-sharp images. They would thus usually have limited generalization performance to domain-shifted images. Since NeRF is a test-time optimization approach, our method does not have the generalization performance issue.", + "bbox": [ + 212, + 209, + 785, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event Enhanced Image Deblurring. Since event camera is able to capture high dynamic temporal information [8], prior methods usually exploit event measurements to enhance the image deblurring performance [12,33,44,45,47,50,57]. EDI [33] is a simple and effective model, which is able to generate a sharp video under various types of blur by solving a single variable non-convex optimization problem. Different from EDI [33], [44,45,47] design end-to-end neural networks for event enhanced image deblurring/frame interpolation. The critical distinction lies in our method's capacity to extract potential camera motion trajectories from the event stream, thereby enhancing subsequent 3D vision tasks with additional geometric insights.", + "bbox": [ + 212, + 378, + 787, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 553, + 380, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a single blurry image and its corresponding event stream, our method recovers the underlying 3D scene representation and the camera motion trajectory jointly. The details of our method are shown in Fig. 2. In particular, we represent the 3D scene with neural radiance fields and the camera motion trajectory with a cubic B-Spline in SE(3) space. Both the blurry image and accumulated events within a time interval can thus be synthesized from the 3D scene representation providing the camera poses. The camera trajectory, NeRF, are then optimized by minimizing the difference between the synthesized data and the real measurements. The details are as follows.", + "bbox": [ + 212, + 583, + 787, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Neural Implicit Representation", + "text_level": 1, + "bbox": [ + 215, + 739, + 519, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt Multi-layer Perceptron (MLP) to represent the 3D scene as the original NeRF [28]. More advanced representations, such as multi-resolution hash encoding [30], can also be exploited to further improve its performance. In particular, the scene model is represented by a learnable mapping function $\\mathbf{F}_{\\theta}:(\\mathbf{x},\\mathbf{d})\\to (\\mathbf{c},\\sigma)$ , which requires a Cartesian coordinates $\\mathbf{x}\\in \\mathbb{R}^3$ and a", + "bbox": [ + 212, + 763, + 785, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W. Li, P. Wan, P. Wang et al.", + "bbox": [ + 271, + 113, + 473, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03fb71c5155deb9571c00c01aa6482d3257aafb8ee7efc28cb8d894c30c4234f.jpg", + "image_caption": [ + "Fig. 2: The pipeline of our method. Given a motion blurry image and its corresponding event stream, we aim to recover both the implicit sharp scene representation and its camera motion trajectory within exposure time. We exploit a continuous time representation for the motion trajectory, and maximize the coherence between both the real measurements and synthesized data for the recovery." + ], + "image_footnote": [], + "bbox": [ + 223, + 148, + 782, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "viewing direction $\\mathbf{d} \\in \\mathbb{S}^2$ as input, and outputs the corresponding volume density $\\sigma \\in \\mathbb{R}$ and color $\\mathbf{c} \\in \\mathbb{R}^3$ . Both the 3D point $\\mathbf{x}$ and viewing direction $\\mathbf{d}$ are defined in the world coordinate frame. They are a function of the pixel coordinate, camera pose and the corresponding intrinsic parameters. To query the intensity of a pixel, we can apply volume rendering by sampling 3D points along the ray, which originates from the camera center and passes through the pixel. The volume rendering can be formally defined as follows:", + "bbox": [ + 212, + 492, + 787, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} (\\mathbf {x}) = \\sum_ {i = 1} ^ {n} T _ {i} \\left(1 - \\exp \\left(- \\sigma_ {i} \\delta_ {i}\\right)\\right) \\boldsymbol {c} _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 604, + 785, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n$ is the number of sampled points along the ray, $c_{i}$ and $\\sigma_{i}$ are the predicted color and volume density of the $i^{th}$ sampled 3D point via $\\mathbf{F}_{\\theta}$ , $\\delta_{i}$ is the distance between the $i^{th}$ and $(i + 1)^{th}$ sampled point, $T_{i}$ is the transmittance which represents the probability that the ray does not hit any particle until the $i^{th}$ sampled point. $T_{i}$ can be computed via:", + "bbox": [ + 212, + 648, + 787, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} = \\exp \\left(- \\sum_ {k = 1} ^ {i - 1} \\sigma_ {k} \\delta_ {k}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 731, + 785, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Camera Motion Trajectory Modeling", + "text_level": 1, + "bbox": [ + 214, + 787, + 570, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a differentiable cubic B-Spline in SE(3) space to better model the camera motion trajectory. The spline is represented by a set of learnable control knots", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BeNeRF", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\pmb{T}_{c_i}^w \\in \\mathbb{S}\\mathbb{E}(3)$ for $i = 0,1,\\dots,n$ . $\\pmb{T}_{c_i}^w$ represents the $i^{th}$ control knot, which is defined as a transformation matrix from the camera coordinate frame to world frame. For brevity, we denote $\\pmb{T}_{c_i}^w$ with $\\pmb{T}_i$ for subsequent derivations. We assume the control knots are sampled with a uniform time interval $\\Delta t$ and the trajectory starts from $t_0$ . Spline with a smaller $\\Delta t$ can represent a smoother motion, with an expense of more control knots to optimize. Since four consecutive control knots determine the value of the spline curve at a particular timestamp, we can thus compute the starting index of the four control knots for time $t$ by:", + "bbox": [ + 212, + 145, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nk = \\left\\lfloor \\frac {t - t _ {0}}{\\Delta t} \\right\\rfloor , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 454, + 272, + 785, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lfloor *\\rfloor$ is the floor operator. Then we can obtain the four control knots responsible for time $t$ as $\\pmb{T}_k,\\pmb{T}_{k + 1},\\pmb{T}_{k + 2}$ and $\\pmb{T}_{k + 3}$ . We can further define $u = \\frac{t - t_0}{\\Delta t} -k$ , where $u\\in [0,1)$ to transform $t$ into a uniform time representation. Using this time representation and based on the matrix representation for the De Boor-Cox formula [35], we can write the matrix representation of a cumulative basis $\\mathcal{B}(\\mathrm{u})$ as", + "bbox": [ + 212, + 304, + 787, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} (u) = \\mathcal {M} \\left[ \\begin{array}{l} 1 \\\\ u \\\\ u ^ {2} \\\\ u ^ {3} \\end{array} \\right], \\quad \\mathcal {M} = \\frac {1}{6} \\left[ \\begin{array}{c c c c} 6 & 0 & 0 & 0 \\\\ 5 & 3 & - 3 & 1 \\\\ 1 & 3 & 3 & - 2 \\\\ 0 & 0 & 0 & 1 \\end{array} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 391, + 785, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pose at time $t$ can be computed as:", + "bbox": [ + 215, + 455, + 504, + 470 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} (u) = \\boldsymbol {T} _ {k} \\cdot \\prod_ {j = 0} ^ {2} \\exp (\\mathcal {B} (u) _ {j + 1} \\cdot \\boldsymbol {\\Omega} _ {k + j}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 477, + 785, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{B}(u)_{j+1}$ denotes the $(j+1)^{th}$ element of the vector $\\mathcal{B}(u)$ , $\\pmb{\\Omega}_{k+j} = \\log(\\pmb{T}_{k+j}^{-1} \\cdot \\pmb{T}_{k+j+1})$ .", + "bbox": [ + 212, + 523, + 785, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since we only consider the continuous camera motion corresponding to a single blurry image, the time interval is thus usually short. We found that four control knots would be sufficient to deliver satisfying results. Therefore, we use the minimal configuration for the following experiments and they are initialized randomly around the identity pose.", + "bbox": [ + 212, + 555, + 787, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Blurry Image Formation Model", + "text_level": 1, + "bbox": [ + 215, + 648, + 522, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A motion blurred image $\\mathbf{B}(\\mathbf{x})\\in \\mathbb{R}^{\\mathrm{W}\\times \\mathrm{H}\\times 3}$ is physically formed by collecting photons during the exposure time and it can be mathematically modeled as:", + "bbox": [ + 212, + 667, + 787, + 700 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} (\\mathbf {x}) \\approx \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\mathbf {I} _ {\\mathrm {i}} (\\mathbf {x}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 705, + 785, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where both W and H are the width and height of the image respectively, $n$ is the number of sampled images, $\\mathbf{x} \\in \\mathbb{R}^2$ represents the pixel location, $\\mathbf{I}_i(\\mathbf{x}) \\in \\mathbb{R}^{W \\times H \\times 3}$ is the $i^{th}$ virtual sharp image sampled within the exposure time. The virtual sharp images can be rendered from the neural 3D scene representation along the previous defined camera trajectory. It can be seen that $\\mathbf{B}(\\mathbf{x})$ is differentiable with respect to the parameters of NeRF and the motion trajectory.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W. Li, P. Wan, P. Wang et al.", + "bbox": [ + 271, + 114, + 473, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Event Data Formation Model", + "text_level": 1, + "bbox": [ + 215, + 146, + 504, + 160 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "An event camera records changes of the brightness as a stream of events asynchronously. Every time a pixel brightness change reaches a contrast threshold (i.e. $|L(\\mathbf{x}, t_i + \\delta t) - L(\\mathbf{x}, t_i)| \\geq C$ ), the camera will trigger an event $e_i = (\\mathbf{x}, t_i, p_i)$ , where $p_i \\in (-1, +1)$ is the polarity of the event, $L(\\mathbf{x}, t_i) = \\log(\\mathbf{I}(\\mathbf{x}, t_i))$ is the brightness logarithm of pixel $\\mathbf{x}$ at timestamp $t_i$ , $C$ is the contrast threshold.", + "bbox": [ + 212, + 176, + 782, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To relate NeRF representation with the event stream, we accumulate the real measured events within a time interval $\\Delta t$ to an image $\\mathbf{E}(\\mathbf{x})$ . The accumulation is defined as:", + "bbox": [ + 212, + 252, + 782, + 296 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} (\\mathbf {x}) = C \\left\\{e _ {i} \\left(\\mathbf {x}, t _ {i}, p _ {i}\\right) \\right\\} _ {t _ {k} < t _ {i} < t _ {k} + \\Delta t}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 301, + 784, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $e(\\mathbf{x}, t_i, p_i)$ is the $i^{th}$ event within the defined time interval corresponding to pixel $\\mathbf{x}$ . For real event cameras, the contrast threshold $C$ changes over time and varies pixel by pixel. We therefore normalize the accumulated event as in [9, 17] to eliminate the effect of unknown $C$ :", + "bbox": [ + 212, + 325, + 782, + 385 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {n} (\\mathbf {x}) = \\frac {\\mathbf {E} (\\mathbf {x})}{\\| \\mathbf {E} (\\mathbf {x}) \\| _ {2}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 398, + 784, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the interpolated start pose and end pose corresponding to the time interval $\\Delta t$ from the spline, we are able to render two gray-scale images (i.e. $\\mathbf{I}_{start}$ and $\\mathbf{I}_{end}$ ) from NeRF. The synthesized accumulated event image $\\hat{\\mathbf{E}}$ can then be computed as:", + "bbox": [ + 214, + 441, + 782, + 500 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} (\\mathbf {x}) = \\log \\left(\\mathbf {I} _ {\\text {e n d}} (\\mathbf {x})\\right) - \\log \\left(\\mathbf {I} _ {\\text {s t a r t}} (\\mathbf {x})\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 505, + 784, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{E}}(\\mathbf{x})$ depends on the parameters of both the cubic spline and NeRF, and is differentiable with respect to them. We can also normalize $\\hat{\\mathbf{E}}(\\mathbf{x})$ to $\\hat{\\mathbf{E}}_n(\\mathbf{x})$ similarly as in Eq. (8) for loss computation.", + "bbox": [ + 212, + 530, + 782, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Loss Functions", + "text_level": 1, + "bbox": [ + 215, + 604, + 383, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We minimize the sum of a photo-metric loss $\\mathcal{L}_p$ and an event loss $\\mathcal{L}_e$ :", + "bbox": [ + 214, + 633, + 715, + 650 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {p} + \\beta \\mathcal {L} _ {e}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 665, + 784, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_p$ represents the loss for the frame-based camera, $\\mathcal{L}_e$ represents the loss for the accumulated events within a randomly sampled time interval, and $\\beta$ is a hyper-parameter. Both losses are defined as follows:", + "bbox": [ + 212, + 693, + 782, + 738 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\left\\| \\mathbf {B} (\\mathbf {x}) - \\hat {\\mathbf {B}} (\\mathbf {x}) \\right\\| ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 752, + 784, + 779 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e} = \\left\\| \\mathbf {E} _ {n} (\\mathbf {x}) - \\hat {\\mathbf {E}} _ {n} (\\mathbf {x}) \\right\\| ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 782, + 784, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{B}} (\\mathbf{x})$ is the real captured blurry image.", + "bbox": [ + 214, + 824, + 547, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BeNeRF", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 217, + 184, + 426, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Synthetic datasets. We generate synthetic datasets for both quantitative and qualitative evaluations via Unreal Engine [1] and Blender [7]. To have a more realistic synthesis, we interpolate the real camera motion trajectories from ETH3D [41] to render high frame-rate images. In total, we generate three sequences (i.e. livingroom, whiteroom and pinkcastle) via Unreal Engine and two sequences (i.e. tanabata and outdoorpool) via Blender. For thorough evaluations, we synthesized twenty blurry images and corresponding event streams for each sequence. The event streams are generated via ESIM [37] from high frame-rate video. Furthermore, we additionally employed the synthetic dataset proposed by $\\mathrm{E}^2\\mathrm{NeRF}$ to compare our method with NeRF-based image-deblur methods which require multi-view training data. This dataset includes synthesized blurry images paired with their respective event streams, which expands upon the six scenes (i.e. chair, ficus, hotdog, lego, materials and mic).", + "bbox": [ + 217, + 220, + 785, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real-world datasets. We utilized the real-world datasets proposed by $\\mathrm{E}^2\\mathrm{NeRF}$ , captured using the DAVIS346 color event camera in real-world scenarios. The dataset encompasses five challenging scenes (i.e. letter, lego, camera, plant and toys). The exposure time for RGB frames was set to 100ms, resulting in occurrences of complex camera motion and severe motion blur within the time interval.", + "bbox": [ + 217, + 421, + 789, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baseline methods and evaluation metrics. To evaluate the performance of our method in terms of image deblurring, we compare it against state-of-the-art deep learning-based single image deblur methods, i.e. SRN-Deblur [46], HINet [5], DeblurGANv2 [19], MPRNet [55], NAFNet [4] and Restormer [54], as well as event-enhanced single image-based deblur methods, i.e. EDI [33], eSLNet [47]. We also compared our method with NeRF-based image deblur method using multi-view information. The quality of the rendered image is evaluated with the commonly used PSNR, SSIM and LPIPS [56] metrics. Since the lack of sharp reference images in real-world datasets, we conducted quantitative analysis experiments on five real scenes using the no-reference image quality assessment metrics BRISQUE [29].", + "bbox": [ + 217, + 517, + 785, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation details. We implement our method with PyTorch. The implicit representation of the scene from MLP (i.e. $F_{\\theta}$ ) is built from NeRF [28] without any modification. We randomly initialize trajectory control knots for cameras within a range of $(0, 0.01)$ . We use two separate Adam optimizers [15] for the scene model (i.e. $F_{\\theta}$ ) and camera motion (i.e. $\\mathbf{P}_i$ ). The learning rate for the scene model and poses decay from $5 \\times 10^{-4}$ with a rate of 0.1 exponentially. In each training step, 1024 pixels for brightness and 1024 pixels for color are sampled. The weight of the event loss $\\beta$ is selected to be 0.1 for synthetic datasets and 2 for real-world datasets, respectively. We train our model for 80K iterations for each image and its corresponding event stream.", + "bbox": [ + 217, + 688, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W. Li, P. Wan, P. Wang et al.", + "bbox": [ + 271, + 114, + 473, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7270b13ee06e607f84dedd5277e8fc2edca8902d8ec917ac4d1750d917513d75.jpg", + "table_caption": [ + "Table 1: Ablation studies on the number of virtual sharp images. The results demonstrate that the image quality gradually saturates as the number of virtual sharp images increases. By compromising the image quality and computational efficiency, we select $n = 19$ for all our experiments." + ], + "table_footnote": [], + "table_body": "
LivingroomTanabata
nPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
736.16.9291.071129.28.8533.0712
1136.84.9353.065930.77.8788.0622
1537.03.9368.063531.90.8979.0529
1937.11.9370.063232.14.9015.0515
2337.11.9375.062932.35.9042.0506
", + "bbox": [ + 246, + 212, + 751, + 306 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2d5f2477f619776eaf3ce92a027a5e04dacb953615f581ad852d8305ec695531.jpg", + "table_caption": [ + "Table 2: Ablation studies on event accumulation time lengths. The experimental results demonstrate that the image quality can be affected by the time length. For all experiments, we select $\\alpha = 0.1$ for event data accumulation." + ], + "table_footnote": [], + "table_body": "
LivingroomTanabata
αPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
0.0536.66.9325.072431.86.8977.0538
0.1037.11.9370.063232.14.9015.0515
0.1537.21.9376.060132.19.9019.0509
0.2037.16.9369.059632.20.9016.0499
0.2537.11.9358.058932.21.9015.0496
", + "bbox": [ + 246, + 375, + 751, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 496, + 388, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate the performance of our method under various configurations. To quantify the differences, we exploit two synthetic datasets (i.e. Livingroom and Tanabata) for the experiments.", + "bbox": [ + 212, + 525, + 785, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effect of the number of virtual sharp images. We evaluate the effect of different numbers of the interpolated virtual images as mentioned in Eq. (6). The experimental results are presented in Table 1. It demonstrates that more virtual images deliver better image quality, while requiring more computational resources. By compromising the image quality and computation requirement, we choose $n = 19$ for our experiments.", + "bbox": [ + 212, + 574, + 787, + 665 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effect of time lengths for event accumulation. We study the effect of different time lengths $\\Delta t$ for event accumulation as in Eq. (7). The timestamps of the event stream are normalized to a range of $(0,1)$ by its total time length. We choose different values from 0.05 to 0.25 for the ablation study. The experimental results are presented in Table 2, showing that the performance on Tanabata dataset gradually saturates as $\\alpha$ increases until to 0.25, whereas on Livingroom dataset, performance initially improves with increasing $\\alpha$ but subsequently declines. This may depend on the noise level in event stream of different time lengths. In all experiments, we select $\\alpha = 0.1$ .", + "bbox": [ + 212, + 669, + 787, + 805 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effect of trajectory representations. We explore the difference between linear interpolation and cubic B-Spline to represent the motion trajectory. The", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BeNeRF", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d0ce49f993f5908c3754072faad634fecc30820f96c94f5b0746ad7d24f4cf85.jpg", + "table_caption": [ + "Table 3: Ablation studies on trajectory representations. The results demonstrate that cubic B-spline can deliver better performance than linear interpolation." + ], + "table_footnote": [], + "table_body": "
LivingroomTanabata
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
linear interpolation34.16.9035.113327.42.8164.1117
cubic B-Spline37.11.9370.063232.14.9015.0515
", + "bbox": [ + 246, + 184, + 751, + 239 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e51eaddcf767706b39542b85ae12dbbe6d880c073a91b0dd5ff64dd765c152e2.jpg", + "table_caption": [ + "Table 4: Quantitative comparisons on single image deblurring with synthetic datasets. The results demonstrate that our method performs better than prior learning-based methods in terms of image quality. For HINet and NAFNet, we tests pre-trained weights from both GoPro and REDS datasets(\\*). Due to page limits, the results of the SSIM metric can be found in the supplementary materials." + ], + "table_footnote": [], + "table_body": "
PSNR ↑
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
DeblurGANv2 [19]29.2627.6423.1620.0926.8925.41
SRN-deblur [46]30.8627.5923.1219.8927.7925.85
MPRNet [55]28.5726.4921.6018.2027.0224.38
HINet [5]28.5626.2721.9118.5926.7024.41
HINet* [5]27.5522.8920.2518.1527.1423.20
NAFNet [4]29.9228.1622.4118.9626.7525.24
NAFNet* [4]28.1823.6720.8518.3827.5223.72
Restormer [54]29.4827.3922.2218.8227.3525.05
BeNeRF37.1132.9529.6832.1436.3833.65
LPIPS ↓
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
DeblurGANv2 [19].2087.1989.2608.3934.3100.2744
SRN-deblur [46].2529.2503.3245.4260.3594.3226
MPRNet [55].2621.2564.3586.4173.3679.3325
HINet [5].2468.2620.3500.4024.3355.3193
HINet* [5].3327.3602.3789.5265.4397.4076
NAFNet [4].2268.1991.3058.3908.3280.2901
NAFNet* [4].3182.3566.3943.5271.4257.4044
Restormer [54].2391.2493.3373.4248.3664.3234
BeNeRF.0632.0788.0761.0515.0677.0675
", + "bbox": [ + 246, + 335, + 751, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "experimental results shown in Table 3 demonstrate that cubic B-Spline deliver better performance on complex motions than that of the linear interpolation. We exploit cubic B-Spline for the experiments.", + "bbox": [ + 212, + 643, + 784, + 688 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Quantitative evaluations", + "text_level": 1, + "bbox": [ + 215, + 709, + 462, + 724 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To evaluate the performance of our method, we compare it against single-image deblurring methods, event-enhanced single-image deblurring methods, and NeRF-based image deblurring methods requiring multi-view information on both synthetic datasets and real datasets. The experimental results are presented in Table 4, Table 5, Table 6 and Table 7.", + "bbox": [ + 212, + 734, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In particular, we compare against SRN-Deblur [46], DeblurGANv2 [19], MPR-Net [55], HINet [5], NAFNet [4], Restormer [54] in terms of single image deblur-", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W. Li, P. Wan, P. Wang et al.", + "bbox": [ + 271, + 114, + 473, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8fb0681f766cd09096ef964edf457aaabce8c41600ff66b28eea2361a90741fd.jpg", + "table_caption": [ + "Table 5: Quantitative comparisons on event-enhanced single image deblurring with synthetic datasets. The results demonstrate that our method performs better than both EDI and eSLNet. Due to page limits, the results of the SSIM metric can be found in the supplementary materials." + ], + "table_footnote": [], + "table_body": "
PSNR ↑
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
eSLNet [47]14.2210.8110.498.8611.8011.24
EDI [33]32.6130.3327.2424.8731.6429.34
BeNeRF37.1132.9529.6832.1436.3833.65
LPIPS ↓
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
eSLNet [47].3981.4236.4902.5067.4676.4572
EDI [33].0904.1020.0779.1039.1409.1030
BeNeRF.0632.0788.0761.0515.0677.0675
", + "bbox": [ + 246, + 212, + 753, + 351 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ring. The experimental results shown in Table 4 demonstrates that our method significantly outperforms prior state-of-the-art methods. It shows that prior learning-based methods have limited generalization performance on domain-shifted images, especially with large motion blurs.", + "bbox": [ + 212, + 378, + 782, + 441 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/102b7cf71884f6b04d9a64bd910c73c08d5344b46a858bc5b80afee2019136db.jpg", + "table_caption": [ + "Table 6: Quantitative comparisons on NeRF-based image deblurring with synthetic datasets from $\\mathbf{E}^2\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}$ . The results indicate that our method outperforms both NeRF and Deblur-NeRF, and exhibits performance comparable to $\\mathrm{E}^2\\mathrm{N}\\mathrm{e}\\mathrm{R}\\mathrm{F}$ in terms of the PSNR metric. Moreover, our method even surpasses $\\mathrm{E}^2\\mathrm{N}\\mathrm{e}\\mathrm{R}\\mathrm{F}$ with the LPIPS metric. Due to page limits, the results of the SSIM metric can be found in the supplementary materials." + ], + "table_footnote": [], + "table_body": "
PSNR ↑
ChairFicusHotdogLegoMaterialsMicAverage
NeRF [28]24.2922.9827.7521.9519.9920.5022.91
Deblur-NeRF [26]25.8722.8624.6224.4720.5411.9221.71
E²NeRF [34]31.2830.0034.3428.1127.2727.6029.77
BeNeRF31.1730.8134.3128.0927.4426.1329.66
ChairFicusHotdogLPIPS ↓
LegoMaterialsMicAverage
NeRF [28].1254.1037.1158.2103.1512.1579.1441
Deblur-NeRF [26].2185.1541.2138.2053.2562.3706.2364
E²NeRF [34].0608.0362.0660.1078.0919.0724.0725
BeNeRF.0500.0299.0539.0745.0708.0738.0588
", + "bbox": [ + 246, + 561, + 753, + 720 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also compare against prior event-enhanced single image deblurring methods, such as EDI [33] and eSLNet [47]. The results in Table 5 demonstrate that our method has superior performance when compared to them. eSLNet demonstrates poor generalization performance, since we are unable to fine-tune it on our evaluation datasets. It demonstrates the benefit on incorporating event streams to enhance single image deblurring task under the framework of NeRF.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BeNeRF", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Furthermore, we conducted detailed comparisons with NeRF-based image deblurring methods that require multi-view information on the synthetic dataset proposed by $\\mathrm{E}^2\\mathrm{NeRF}$ . We compared against NeRF [28], Deblur-NeRF [26] and $\\mathrm{E}^2\\mathrm{NeRF}$ [34]. The experimental results in Table 6 demonstrate that despite utilizing only a single blurred image and event stream of a small time interval, our method achieves performance comparable to $\\mathrm{E}^2\\mathrm{NeRF}$ [34], which utilizes multiview images and a longer event stream, in terms of PSNR metric. Moreover, our method even surpasses $\\mathrm{E}^2\\mathrm{NeRF}$ [34] in terms of the LPIPS metric.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Finally, we select the best-performing algorithms on the synthetic dataset, excluding our method, from single-image deblurring methods, event-enhanced single-image deblurring methods, and NeRF-based image deblurring methods, which are SRN-deblur [46], EDI [33], and $\\mathrm{E}^2\\mathrm{NeRF}$ [34], respectively. We compare against with these methods on the real dataset and provide the BRISQUE [29] metric. The results in Table 7 indicate a significant improvement of our method over the aforementioned methods. This is attributed to our method's incorporation of a physical model for the imaging process of blurry images, enabling better performance on real-world datasets.", + "bbox": [ + 212, + 268, + 787, + 402 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7bc8628882ca3c7084fb110c1307c4d47595b2dcbd402c87f38b67a44b87e0cd.jpg", + "table_caption": [ + "Table 7: Quantitative comparisons on real-world datasets from $\\mathbf{E}^2\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}$ . We exploit the used real-world dataset proposed by E2NeRF [34] for the evaluations, which is collected via a DAVIS event camera. The results indicate that our method outperforms EDI, SRN-Deblur and even $\\mathrm{E}^2\\mathrm{NeRF}$ on the real-world datasets. Note that $\\mathrm{E}^2\\mathrm{NeRF}$ [34] requires multi-view images while ours only need a single image. Since $\\mathrm{E}^2\\mathrm{NeRF}$ does not provide the trained model and the code for the metric computation, we re-trained $\\mathrm{E}^2\\mathrm{NeRF}$ for this experiment and compute the metric with the MATLAB implementation of the BRISQUE metric for fair comparisons." + ], + "table_footnote": [], + "table_body": "
BRISQUE ↓
CameraLegoLetterPlantToysAverage
EDI [33]29.7429.3528.7431.0937.0931.20
SRN-Deblur [46]32.2034.9140.8237.4546.1038.30
E²NeRF [34]33.4033.8537.4132.0243.0035.94
BeNeRF19.4725.8627.3721.4625.2023.87
", + "bbox": [ + 246, + 555, + 751, + 633 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Qualitative evaluations", + "text_level": 1, + "bbox": [ + 215, + 690, + 450, + 705 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The qualitative evaluation results are shown in Fig. 3 and Fig. 4 for both synthetic and real datasets respectively. The experimental results demonstrate that our method delivers better performance than prior methods even when the image is severely blurred. In particular, Fig. 3 shows that prior learning-based methods struggle to generalize to domain-shifted images. Notably, EDI [33] performs well on synthetic datasets due to the high quality of event data. Fig. 4 shows that our method outperforms all prior methods (even trained with multi-view images) on real noisy dataset, which demonstrates the advantage of our method", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W. Li, P. Wan, P. Wang et al.", + "bbox": [ + 271, + 114, + 472, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6d63bf5c6f1577901a422ff754c0e292ef5fdb27c236317a80193f518da1ab33.jpg", + "image_caption": [ + "Fig. 3: Qualitative results of different methods with synthetic datasets. It demonstrates that our method delivers better performance compared to prior approaches. The learning based methods fail to generalize on severely blurry images." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 776, + 536 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "and the necessity to jointly optimize the camera motion and the implicit 3D representation.", + "bbox": [ + 212, + 619, + 787, + 651 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 671, + 359, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In conclusion, we present a novel method to jointly recover the underlying 3D scene representation and camera motion trajectory from a single blurry image and its corresponding event stream. Extensive experimental evaluations with both synthetic and real datasets demonstrate the superior performance of our method over prior works, even for those requiring multi-view images and longer event streams.", + "bbox": [ + 212, + 700, + 787, + 789 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BeNeRF", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8ef558275be268dc07a1cbdf2ceb1beed7d81b6dd1db2b410843409e52773d24.jpg", + "image_caption": [ + "Fig. 4: Qualitative results of different methods on the real datasets. The experimental results demonstrate that our method delivers superior performance on the real DAVIS datasets from $\\mathrm{E}^2\\mathrm{NeRF}$ . We are even achieve better performance than prior methods requiring multi-view images and longer event stream." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 787 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W. Li, P. Wan, P. Wang et al.", + "bbox": [ + 271, + 114, + 473, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Unreal Engine: The most powerful real-time 3D creation tool. https://www.unrealengine.com/en-US/", + "2. Bian, W., Wang, Z., Li, K., Bian, J., Prisacariu, V.A.: NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "3. Cai, S., Obukhov, A., Dai, D., Van Gool, L.: Pix2NeRF: Unsupervised Conditional pi-GAN for Single Image to Neural Radiance Fields Translation. In: CVPR (2022)", + "4. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple Baselines for Image Restoration. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 17-33. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-20071-7_2", + "5. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: HINet: Half Instance Normalization Network for Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021)", + "6. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12882-12891 (2022)", + "7. Foundation, B.: Blender.org - Home of the Blender project - Free and Open 3D Creation Software", + "8. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (Jan 2022). https://doi.org/10.1109/TPAMI.2020.3008413", + "9. Hidalgo-Carrió, J., Gallego, G., Scaramuzza, D.: Event-aided Direct Sparse Odometry (Apr 2022). https://doi.org/10.48550/arXiv.2204.07640", + "10. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 837-847 (2023)", + "1. Jia, J.: Single image motion deblurring using transparency. In: 2007 IEEE Conference on computer vision and pattern recognition. pp. 1-8. IEEE (2007)", + "2. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning Event-Based Motion Deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)", + "3. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018)", + "4. Joshi, N., Zitnick, C.L., Szeliski, R., Kriegman, D.J.: Image deblurring and denoising using color priors. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1550-1557. IEEE (2009)", + "5. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization (Jan 2017). https://doi.org/10.48550/arXiv.1412.6980", + "6. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023)", + "7. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-NeRF: Neural Radiance Fields From a Moving Event Camera. IEEE Robotics and Automation Letters 8(3), 1587–1594 (Mar 2023). https://doi.org/10.1109/LRA.2023.3240646" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BeNeRF", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8183-8192 (2018)", + "19. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8878-8887 (2019)", + "20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "21. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)", + "22. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields. In: International Conference on Learning Representations (ICLR) (2024)", + "23. Lichtsteiner, P., Posch, C., Delbruck, T.: A 128 $\\times$ times 128 120 dB 15 Ms Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits 43(2), 566-576 (Feb 2008). https://doi.org/10.1109/JSSC.2007.914337", + "24. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: BARF: Bundle-Adjusting Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5741-5751 (2021)", + "25. Low, W.F., Lee, G.H.: Robust e-nerf: Nerf from sparse & noisy events under nonuniform motion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "26. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural Radiance Fields From Blurry Images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)", + "27. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. CVPR (2022)", + "28. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis (Aug 2020). https://doi.org/10.48550/arXiv.2003.08934", + "29. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing 21(12), 4695-4708 (2012)", + "30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (Jul 2022). https://doi.org/10.1145/3528223.3530127", + "31. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)", + "32. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)", + "33. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a Blurry Frame Alive at High Frame-Rate With an Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W. Li, P. Wan, P. Wang et al.", + "bbox": [ + 271, + 114, + 473, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Qi, Y., Zhu, L., Zhang, Y., Li, J.: E2NeRF: Event Enhanced Neural Radiance Fields from Blurry Images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13254-13264 (2023)", + "35. Qin, K.: General matrix representations for B-splines. In: Proceedings Pacific Graphics '98. Sixth Pacific Conference on Computer Graphics and Applications (Cat. No.98EX208). pp. 37-43 (Oct 1998). https://doi.org/10.1109/PCCGA.1998.731996", + "36. Rebain, D., Matthews, M., Yi, K.M., Lagun, D., Tagliasacchi, A.: LOLNeRF: Learn from One Look. In: CVPR (2022)", + "37. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: An Open Event Camera Simulator. In: Proceedings of The 2nd Conference on Robot Learning. pp. 969-982. PMLR (Oct 2018)", + "38. Rematas, K., Martin-Brualla, R., Ferrari, V.: ShaRF: Shape-conditioned Radiance Fields from a Single View. In: ICML (2021)", + "39. Rudnev, V., Elgharib, M., Theobalt, C., Golyanik, V.: EventNeRF: Neural Radiance Fields From a Single Colour Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4992-5002 (2023)", + "40. Schonberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "41. Schops, T., Sattler, T., Pollefeys, M.: BAD SLAM: Bundle Adjusted Direct RGB-D SLAM. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 134-144 (2019)", + "42. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. Acm transactions on graphics (tog) 27(3), 1-10 (2008)", + "43. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 769-777 (2015)", + "44. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-Based Fusion for Motion Deblurring with Cross-modal Attention. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 412-428. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19797-0_24", + "45. Sun, L., Sakaridis, C., Liang, J., Sun, P., Cao, J., Zhang, K., Jiang, Q., Wang, K., Van Gool, L.: Event-based frame interpolation with ad-hoc deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18043-18052 (2023)", + "46. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-Recurrent Network for Deep Image Deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8174–8182 (2018)", + "47. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event Enhanced High-Quality Image Recovery. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 155-171. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58601-0_10", + "48. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)", + "49. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF-: Neural Radiance Fields Without Known Camera Parameters (Apr 2022). https://doi.org/10.48550/arXiv.2102.07064" + ], + "bbox": [ + 212, + 146, + 787, + 838 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BeNeRF", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion Deblurring With Real Events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021)", + "51. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 157-170. Springer (2010)", + "52. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)", + "53. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2021)", + "54. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient Transformer for High-Resolution Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5728-5739 (2022)", + "55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-Stage Progressive Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14821-14831 (2021)", + "56. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018)", + "57. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing Event-Based Motion Deblurring in Real-World Scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)", + "58. Zhao, L., Wang, P., Liu, P.: BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting. In: European Conference on Computer Vision (ECCV). Springer (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 535 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "W. Li, P. Wan, P. Wang et al.", + "bbox": [ + 271, + 114, + 473, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_model.json b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4cb6257496e89d7b2421affb71b0064428a7308d --- /dev/null +++ b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_model.json @@ -0,0 +1,2326 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.231, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "BeNeRF: Neural Radiance Fields from a Single Blurry Image and Event Stream" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.211, + 0.782, + 0.245 + ], + "angle": 0, + "content": "Wenpu Li\\(^{1,5*}\\), Pian Wan\\(^{1,2*}\\), Peng Wang\\(^{1,3*}\\), Jinghang Li\\(^{4}\\), Yi Zhou\\(^{4}\\), and Peidong Liu\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.254, + 0.761, + 0.285 + ], + "angle": 0, + "content": "1Westlake University, 2EPFL, 3Zhejiang University, 4Hunan University, 5Guangdong University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.297, + 0.628, + 0.312 + ], + "angle": 0, + "content": "https://github.com/wu-cvgl/BeNeRF" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.327, + 0.765, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.513 + ], + "angle": 0, + "content": "Fig. 1: Given a single blurry image and its corresponding event stream, BeNeRF can synthesize high-quality novel images along the camera trajectory, recovering a sharp and coherent video from the single blurry image." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.531, + 0.741, + 0.768 + ], + "angle": 0, + "content": "Abstract. Implicit scene representation has attracted a lot of attention in recent research of computer vision and graphics. Most prior methods focus on how to reconstruct 3D scene representation from a set of images. In this work, we demonstrate the possibility to recover the neural radiance fields (NeRF) from a single blurry image and its corresponding event stream. To eliminate motion blur, we introduce event stream to regularize the learning process of NeRF by accumulating it into an image. We model the camera motion with a cubic B-Spline in SE(3) space. Both the blurry image and the brightness change within a time interval, can then be synthesized from the NeRF given the 6-DoF poses interpolated from the cubic B-Spline. Our method can jointly learn both the implicit scene representation and the camera motion by minimizing the differences between the synthesized data and the real measurements without any prior knowledge of camera poses. We evaluate the proposed method with both synthetic and real datasets. The experimental results demonstrate that we are able to render view-consistent latent sharp images from the learned NeRF and bring a blurry image alive in high quality." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.779, + 0.741, + 0.808 + ], + "angle": 0, + "content": "Keywords: Neural Radiance Fields \\(\\cdot\\) Event Stream \\(\\cdot\\) Pose Estimation \\(\\cdot\\) Deblurring \\(\\cdot\\) Novel View Synthesis \\(\\cdot\\) 3D from a Single Image" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.812, + 0.778, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution: {liwenpu, wangpeng}@westlake.edu.cn, pianwan@outlook.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.826, + 0.667, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author: Peidong Liu(liupeidong@westlake.edu.cn)." + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.812, + 0.778, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.474, + 0.129 + ], + "angle": 0, + "content": "W. Li, P. Wan, P. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.162 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.296 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) [28] has drawn much attention due to its extraordinary ability in representing 3D scenes and synthesizing novel views. Given multi-view sharp RGB images and calibrated camera poses from COLMAP [40], NeRF takes corresponding 3D spatial location and 2D view direction as input, and optimizes a multi-layer perception (MLP) to represent the 3D scene. More recent advanced methods also exploit explicit octree [52], multi-resolution hash encoding [30] etc., to represent the 3D scene to improve both the training and rendering efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.787, + 0.476 + ], + "angle": 0, + "content": "Prior methods usually rely on multi-view images to learn the 3D representation. Several pioneering works recently attempt to exploit a single image to learn the underlying neural radiance fields [3,36,38,53]. They usually rely on a large dataset to pre-train the networks to learn priors to address the ill-posed problem. A blurry image further aggravates the problem due to the image quality degradation. Although motion blur is usually not preferred by most vision algorithms, they actually encode additional camera motion trajectory and more structural information compared to a sharp image. In this paper, we explore the possibility of recovering the neural radiance fields and camera motion trajectory from a single blurry image. Instead of learning priors from a large dataset as in previous works, we exploit the usage of an additional event stream to better constrain the problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.787, + 0.703 + ], + "angle": 0, + "content": "Event stream can be acquired by an event camera [23] which captures pixel intensity changes caused by the relative motion between the static scene and camera. Unlike standard frame-based cameras, event camera captures asynchronous events with very low latency, leading to extremely high temporal resolution [8]. This characteristic compensates with the image formation process of a blurry image (i.e. integral of photon measurements across time). Several prior works thus take advantage of both modalities for high quality single image deblurring [33,44,47]. However, these methods are unable to recover the camera motion trajectory and extract structural details from a single blurry image, thereby limiting their applicability in 3D computer vision tasks. Some NeRF-based methods that incorporate event stream [10,16,25,34,39] demonstrate the capability to achieve image deblurring and accurate reconstruction of neural radiance fields. Nonetheless, these methods necessitate input images from multiple viewpoints alongside event data. In contrast, we explore the usage of only a single blurry image for the NeRF recovery with unknown camera motions in this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We represent the continuous camera motion with a cubic B-Spline in SE(3) space and define it as the trajectory of both frame-based camera and event camera. Given the neural 3D representation and interpolated poses from the cubic B-Spline, we can synthesize both the blurry image and the brightness change within a time interval via the physical image formation process. The NeRF and motion trajectory can then be jointly optimized by minimizing the difference between the synthesized data and the real measurements. To evaluate the performance of our method, we conduct experiments with both synthetic and real datasets. The experimental results demonstrate that our method is able to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BeNeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "recover the neural radiance fields from a blurry image and its corresponding event stream without knowing prior knowledge of poses. We are thus able to render view-consistent latent sharp images encoded in a single blurred image from learned NeRF, effectively enhancing the quality of the blurry image. The experimental results further demonstrate that our method is even able to reach same performance as \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [34], which targets for the same problem, but with multi-view images and longer event data. In summary, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.279, + 0.785, + 0.325 + ], + "angle": 0, + "content": "- We propose a NeRF-based method that can recover the neural 3D representation from a single blurry image and its corresponding event stream, without knowing any ground truth poses;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.37 + ], + "angle": 0, + "content": "- Our method is able to estimate the complex continuous trajectory of camera motion during the imaging process from a single blurry image, providing multi-view geometric information;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.37, + 0.787, + 0.446 + ], + "angle": 0, + "content": "- We experimentally validate that our approach is able to recover high quality latent sharp images and high frame-rate video from a single blurry image, without any generalization issue. Furthermore, we are able to reach same level of performance as \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [34], which requires multi-view images and longer event data." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.279, + 0.787, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.388, + 0.487 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.787, + 0.55 + ], + "angle": 0, + "content": "We roughly categorize our related works into three main areas: neural implicit scene representation, single image deblurring and event-enhanced image deblurring." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Neural Implicit Scene Representation. NeRF has attracted lots of attention due to its powerful ability of implicit 3D scene representation and novel view synthesis [28]. Many following works have been proposed to improve NeRF's performance or extend NeRF to other fields. [2,24,49] jointly trained NeRF with inaccurate camera poses. [6,22,26,27,32,53] improved the performance of NeRF with degraded images, including noisy or few images etc. Recently, several event based NeRF [10,17,34,39] have also been proposed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We will mainly focus on the methods that are most related to ours. BAD-NeRF [48] aims to recover true underlying 3D scene representation from multiview blurry inputs and inaccurate camera poses estimated from COLMAP [40]. However, BAD-NeRF and its variants [20,58] struggle to address situations where the input is limited to a single image, primarily due to the severe illness of the problem. Event-Enhanced NeRF (E²NeRF) [34] aims to recover the 3D scene representation from multiple blurry images and event streams. To train NeRF without optimizing the camera poses, E²NeRF manually segments the event stream using preset parameters, aiming to recover sharp images individually through EDI [33]. However, in cases of serve camera motion, the recovered images from event stream may still be blurry, leading to inaccurate pose estimation in COLMAP [40]. This two-stage approach introduces errors (i.e. either from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.474, + 0.129 + ], + "angle": 0, + "content": "W. Li, P. Wan, P. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.783, + 0.207 + ], + "angle": 0, + "content": "EDI [33] or COLMAP [40]) and fails to accurately model the continuous camera trajectory. In contrast, we are able to train both the NeRF and the camera motion trajectory jointly from a single blurry image and its corresponding event stream, which is harder than prior methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.787, + 0.377 + ], + "angle": 0, + "content": "Single Image Deblurring. A blurred image can be formulated as the convolution result of a sharp image and a kernel. Therefore, classical approaches [11,14,21,42,51] generally regard the deblurring problem as a joint optimization of the blur kernel and the latent virtual sharp images. With the development of deep learning, many learning based end-to-end deblurring methods have been proposed [13, 18, 19, 31, 43, 46, 54]. These methods usually demonstrate better qualitative and quantitative deblurring results. However, learning based deblurring methods are usually trained on a large dataset which contains paired blurry-sharp images. They would thus usually have limited generalization performance to domain-shifted images. Since NeRF is a test-time optimization approach, our method does not have the generalization performance issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Event Enhanced Image Deblurring. Since event camera is able to capture high dynamic temporal information [8], prior methods usually exploit event measurements to enhance the image deblurring performance [12,33,44,45,47,50,57]. EDI [33] is a simple and effective model, which is able to generate a sharp video under various types of blur by solving a single variable non-convex optimization problem. Different from EDI [33], [44,45,47] design end-to-end neural networks for event enhanced image deblurring/frame interpolation. The critical distinction lies in our method's capacity to extract potential camera motion trajectories from the event stream, thereby enhancing subsequent 3D vision tasks with additional geometric insights." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.381, + 0.571 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Given a single blurry image and its corresponding event stream, our method recovers the underlying 3D scene representation and the camera motion trajectory jointly. The details of our method are shown in Fig. 2. In particular, we represent the 3D scene with neural radiance fields and the camera motion trajectory with a cubic B-Spline in SE(3) space. Both the blurry image and accumulated events within a time interval can thus be synthesized from the 3D scene representation providing the camera poses. The camera trajectory, NeRF, are then optimized by minimizing the difference between the synthesized data and the real measurements. The details are as follows." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.52, + 0.756 + ], + "angle": 0, + "content": "3.1 Neural Implicit Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.843 + ], + "angle": 0, + "content": "We adopt Multi-layer Perceptron (MLP) to represent the 3D scene as the original NeRF [28]. More advanced representations, such as multi-resolution hash encoding [30], can also be exploited to further improve its performance. In particular, the scene model is represented by a learnable mapping function \\(\\mathbf{F}_{\\theta}:(\\mathbf{x},\\mathbf{d})\\to (\\mathbf{c},\\sigma)\\), which requires a Cartesian coordinates \\(\\mathbf{x}\\in \\mathbb{R}^3\\) and a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BeNeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.149, + 0.784, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Fig. 2: The pipeline of our method. Given a motion blurry image and its corresponding event stream, we aim to recover both the implicit sharp scene representation and its camera motion trajectory within exposure time. We exploit a continuous time representation for the motion trajectory, and maximize the coherence between both the real measurements and synthesized data for the recovery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.601 + ], + "angle": 0, + "content": "viewing direction \\(\\mathbf{d} \\in \\mathbb{S}^2\\) as input, and outputs the corresponding volume density \\(\\sigma \\in \\mathbb{R}\\) and color \\(\\mathbf{c} \\in \\mathbb{R}^3\\). Both the 3D point \\(\\mathbf{x}\\) and viewing direction \\(\\mathbf{d}\\) are defined in the world coordinate frame. They are a function of the pixel coordinate, camera pose and the corresponding intrinsic parameters. To query the intensity of a pixel, we can apply volume rendering by sampling 3D points along the ray, which originates from the camera center and passes through the pixel. The volume rendering can be formally defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.606, + 0.787, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} (\\mathbf {x}) = \\sum_ {i = 1} ^ {n} T _ {i} \\left(1 - \\exp \\left(- \\sigma_ {i} \\delta_ {i}\\right)\\right) \\boldsymbol {c} _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.727 + ], + "angle": 0, + "content": "where \\( n \\) is the number of sampled points along the ray, \\( c_{i} \\) and \\( \\sigma_{i} \\) are the predicted color and volume density of the \\( i^{th} \\) sampled 3D point via \\( \\mathbf{F}_{\\theta} \\), \\( \\delta_{i} \\) is the distance between the \\( i^{th} \\) and \\( (i + 1)^{th} \\) sampled point, \\( T_{i} \\) is the transmittance which represents the probability that the ray does not hit any particle until the \\( i^{th} \\) sampled point. \\( T_{i} \\) can be computed via:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.732, + 0.787, + 0.773 + ], + "angle": 0, + "content": "\\[\nT _ {i} = \\exp \\left(- \\sum_ {k = 1} ^ {i - 1} \\sigma_ {k} \\delta_ {k}\\right). \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.571, + 0.806 + ], + "angle": 0, + "content": "3.2 Camera Motion Trajectory Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We use a differentiable cubic B-Spline in SE(3) space to better model the camera motion trajectory. The spline is represented by a set of learnable control knots" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.474, + 0.129 + ], + "angle": 0, + "content": "W. Li, P. Wan, P. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.269 + ], + "angle": 0, + "content": "\\(\\pmb{T}_{c_i}^w \\in \\mathbb{S}\\mathbb{E}(3)\\) for \\(i = 0,1,\\dots,n\\). \\(\\pmb{T}_{c_i}^w\\) represents the \\(i^{th}\\) control knot, which is defined as a transformation matrix from the camera coordinate frame to world frame. For brevity, we denote \\(\\pmb{T}_{c_i}^w\\) with \\(\\pmb{T}_i\\) for subsequent derivations. We assume the control knots are sampled with a uniform time interval \\(\\Delta t\\) and the trajectory starts from \\(t_0\\). Spline with a smaller \\(\\Delta t\\) can represent a smoother motion, with an expense of more control knots to optimize. Since four consecutive control knots determine the value of the spline curve at a particular timestamp, we can thus compute the starting index of the four control knots for time \\(t\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.455, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "\\[\nk = \\left\\lfloor \\frac {t - t _ {0}}{\\Delta t} \\right\\rfloor , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.788, + 0.392 + ], + "angle": 0, + "content": "where \\(\\lfloor *\\rfloor\\) is the floor operator. Then we can obtain the four control knots responsible for time \\(t\\) as \\(\\pmb{T}_k,\\pmb{T}_{k + 1},\\pmb{T}_{k + 2}\\) and \\(\\pmb{T}_{k + 3}\\). We can further define \\(u = \\frac{t - t_0}{\\Delta t} -k\\), where \\(u\\in [0,1)\\) to transform \\(t\\) into a uniform time representation. Using this time representation and based on the matrix representation for the De Boor-Cox formula [35], we can write the matrix representation of a cumulative basis \\(\\mathcal{B}(\\mathrm{u})\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.392, + 0.786, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} (u) = \\mathcal {M} \\left[ \\begin{array}{l} 1 \\\\ u \\\\ u ^ {2} \\\\ u ^ {3} \\end{array} \\right], \\quad \\mathcal {M} = \\frac {1}{6} \\left[ \\begin{array}{c c c c} 6 & 0 & 0 & 0 \\\\ 5 & 3 & - 3 & 1 \\\\ 1 & 3 & 3 & - 2 \\\\ 0 & 0 & 0 & 1 \\end{array} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.456, + 0.506, + 0.471 + ], + "angle": 0, + "content": "The pose at time \\(t\\) can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.478, + 0.786, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} (u) = \\boldsymbol {T} _ {k} \\cdot \\prod_ {j = 0} ^ {2} \\exp (\\mathcal {B} (u) _ {j + 1} \\cdot \\boldsymbol {\\Omega} _ {k + j}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.786, + 0.556 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}(u)_{j+1}\\) denotes the \\((j+1)^{th}\\) element of the vector \\(\\mathcal{B}(u)\\), \\(\\pmb{\\Omega}_{k+j} = \\log(\\pmb{T}_{k+j}^{-1} \\cdot \\pmb{T}_{k+j+1})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.632 + ], + "angle": 0, + "content": "Since we only consider the continuous camera motion corresponding to a single blurry image, the time interval is thus usually short. We found that four control knots would be sufficient to deliver satisfying results. Therefore, we use the minimal configuration for the following experiments and they are initialized randomly around the identity pose." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.524, + 0.664 + ], + "angle": 0, + "content": "3.3 Blurry Image Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.788, + 0.701 + ], + "angle": 0, + "content": "A motion blurred image \\(\\mathbf{B}(\\mathbf{x})\\in \\mathbb{R}^{\\mathrm{W}\\times \\mathrm{H}\\times 3}\\) is physically formed by collecting photons during the exposure time and it can be mathematically modeled as:" + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.707, + 0.786, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} (\\mathbf {x}) \\approx \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\mathbf {I} _ {\\mathrm {i}} (\\mathbf {x}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where both W and H are the width and height of the image respectively, \\(n\\) is the number of sampled images, \\(\\mathbf{x} \\in \\mathbb{R}^2\\) represents the pixel location, \\(\\mathbf{I}_i(\\mathbf{x}) \\in \\mathbb{R}^{W \\times H \\times 3}\\) is the \\(i^{th}\\) virtual sharp image sampled within the exposure time. The virtual sharp images can be rendered from the neural 3D scene representation along the previous defined camera trajectory. It can be seen that \\(\\mathbf{B}(\\mathbf{x})\\) is differentiable with respect to the parameters of NeRF and the motion trajectory." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BeNeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.506, + 0.161 + ], + "angle": 0, + "content": "3.4 Event Data Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.784, + 0.252 + ], + "angle": 0, + "content": "An event camera records changes of the brightness as a stream of events asynchronously. Every time a pixel brightness change reaches a contrast threshold (i.e. \\( |L(\\mathbf{x}, t_i + \\delta t) - L(\\mathbf{x}, t_i)| \\geq C \\)), the camera will trigger an event \\( e_i = (\\mathbf{x}, t_i, p_i) \\), where \\( p_i \\in (-1, +1) \\) is the polarity of the event, \\( L(\\mathbf{x}, t_i) = \\log(\\mathbf{I}(\\mathbf{x}, t_i)) \\) is the brightness logarithm of pixel \\( \\mathbf{x} \\) at timestamp \\( t_i \\), \\( C \\) is the contrast threshold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.784, + 0.297 + ], + "angle": 0, + "content": "To relate NeRF representation with the event stream, we accumulate the real measured events within a time interval \\(\\Delta t\\) to an image \\(\\mathbf{E}(\\mathbf{x})\\). The accumulation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.302, + 0.785, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} (\\mathbf {x}) = C \\left\\{e _ {i} \\left(\\mathbf {x}, t _ {i}, p _ {i}\\right) \\right\\} _ {t _ {k} < t _ {i} < t _ {k} + \\Delta t}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.784, + 0.386 + ], + "angle": 0, + "content": "where \\( e(\\mathbf{x}, t_i, p_i) \\) is the \\( i^{th} \\) event within the defined time interval corresponding to pixel \\( \\mathbf{x} \\). For real event cameras, the contrast threshold \\( C \\) changes over time and varies pixel by pixel. We therefore normalize the accumulated event as in [9, 17] to eliminate the effect of unknown \\( C \\):" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.399, + 0.785, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {n} (\\mathbf {x}) = \\frac {\\mathbf {E} (\\mathbf {x})}{\\| \\mathbf {E} (\\mathbf {x}) \\| _ {2}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.442, + 0.784, + 0.501 + ], + "angle": 0, + "content": "Given the interpolated start pose and end pose corresponding to the time interval \\(\\Delta t\\) from the spline, we are able to render two gray-scale images (i.e. \\(\\mathbf{I}_{start}\\) and \\(\\mathbf{I}_{end}\\)) from NeRF. The synthesized accumulated event image \\(\\hat{\\mathbf{E}}\\) can then be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.506, + 0.785, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} (\\mathbf {x}) = \\log \\left(\\mathbf {I} _ {\\text {e n d}} (\\mathbf {x})\\right) - \\log \\left(\\mathbf {I} _ {\\text {s t a r t}} (\\mathbf {x})\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.784, + 0.579 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{E}}(\\mathbf{x})\\) depends on the parameters of both the cubic spline and NeRF, and is differentiable with respect to them. We can also normalize \\(\\hat{\\mathbf{E}}(\\mathbf{x})\\) to \\(\\hat{\\mathbf{E}}_n(\\mathbf{x})\\) similarly as in Eq. (8) for loss computation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.384, + 0.619 + ], + "angle": 0, + "content": "3.5 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.635, + 0.716, + 0.651 + ], + "angle": 0, + "content": "We minimize the sum of a photo-metric loss \\(\\mathcal{L}_p\\) and an event loss \\(\\mathcal{L}_e\\):" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.666, + 0.785, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {p} + \\beta \\mathcal {L} _ {e}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.784, + 0.739 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_p\\) represents the loss for the frame-based camera, \\(\\mathcal{L}_e\\) represents the loss for the accumulated events within a randomly sampled time interval, and \\(\\beta\\) is a hyper-parameter. Both losses are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.753, + 0.785, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\left\\| \\mathbf {B} (\\mathbf {x}) - \\hat {\\mathbf {B}} (\\mathbf {x}) \\right\\| ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.783, + 0.785, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e} = \\left\\| \\mathbf {E} _ {n} (\\mathbf {x}) - \\hat {\\mathbf {E}} _ {n} (\\mathbf {x}) \\right\\| ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.548, + 0.841 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{B}} (\\mathbf{x})\\) is the real captured blurry image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.474, + 0.129 + ], + "angle": 0, + "content": "W. Li, P. Wan, P. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.185, + 0.427, + 0.201 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.221, + 0.786, + 0.418 + ], + "angle": 0, + "content": "Synthetic datasets. We generate synthetic datasets for both quantitative and qualitative evaluations via Unreal Engine [1] and Blender [7]. To have a more realistic synthesis, we interpolate the real camera motion trajectories from ETH3D [41] to render high frame-rate images. In total, we generate three sequences (i.e. livingroom, whiteroom and pinkcastle) via Unreal Engine and two sequences (i.e. tanabata and outdoorpool) via Blender. For thorough evaluations, we synthesized twenty blurry images and corresponding event streams for each sequence. The event streams are generated via ESIM [37] from high frame-rate video. Furthermore, we additionally employed the synthetic dataset proposed by \\(\\mathrm{E}^2\\mathrm{NeRF}\\) to compare our method with NeRF-based image-deblur methods which require multi-view training data. This dataset includes synthesized blurry images paired with their respective event streams, which expands upon the six scenes (i.e. chair, ficus, hotdog, lego, materials and mic)." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.422, + 0.79, + 0.512 + ], + "angle": 0, + "content": "Real-world datasets. We utilized the real-world datasets proposed by \\(\\mathrm{E}^2\\mathrm{NeRF}\\), captured using the DAVIS346 color event camera in real-world scenarios. The dataset encompasses five challenging scenes (i.e. letter, lego, camera, plant and toys). The exposure time for RGB frames was set to 100ms, resulting in occurrences of complex camera motion and severe motion blur within the time interval." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.518, + 0.786, + 0.685 + ], + "angle": 0, + "content": "Baseline methods and evaluation metrics. To evaluate the performance of our method in terms of image deblurring, we compare it against state-of-the-art deep learning-based single image deblur methods, i.e. SRN-Deblur [46], HINet [5], DeblurGANv2 [19], MPRNet [55], NAFNet [4] and Restormer [54], as well as event-enhanced single image-based deblur methods, i.e. EDI [33], eSLNet [47]. We also compared our method with NeRF-based image deblur method using multi-view information. The quality of the rendered image is evaluated with the commonly used PSNR, SSIM and LPIPS [56] metrics. Since the lack of sharp reference images in real-world datasets, we conducted quantitative analysis experiments on five real scenes using the no-reference image quality assessment metrics BRISQUE [29]." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.689, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Implementation details. We implement our method with PyTorch. The implicit representation of the scene from MLP (i.e. \\( F_{\\theta} \\)) is built from NeRF [28] without any modification. We randomly initialize trajectory control knots for cameras within a range of \\((0, 0.01)\\). We use two separate Adam optimizers [15] for the scene model (i.e. \\( F_{\\theta} \\)) and camera motion (i.e. \\( \\mathbf{P}_i \\)). The learning rate for the scene model and poses decay from \\( 5 \\times 10^{-4} \\) with a rate of 0.1 exponentially. In each training step, 1024 pixels for brightness and 1024 pixels for color are sampled. The weight of the event loss \\( \\beta \\) is selected to be 0.1 for synthetic datasets and 2 for real-world datasets, respectively. We train our model for 80K iterations for each image and its corresponding event stream." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BeNeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Ablation studies on the number of virtual sharp images. The results demonstrate that the image quality gradually saturates as the number of virtual sharp images increases. By compromising the image quality and computational efficiency, we select \\( n = 19 \\) for all our experiments." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.213, + 0.753, + 0.307 + ], + "angle": 0, + "content": "
LivingroomTanabata
nPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
736.16.9291.071129.28.8533.0712
1136.84.9353.065930.77.8788.0622
1537.03.9368.063531.90.8979.0529
1937.11.9370.063232.14.9015.0515
2337.11.9375.062932.35.9042.0506
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Table 2: Ablation studies on event accumulation time lengths. The experimental results demonstrate that the image quality can be affected by the time length. For all experiments, we select \\(\\alpha = 0.1\\) for event data accumulation." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.376, + 0.753, + 0.468 + ], + "angle": 0, + "content": "
LivingroomTanabata
αPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
0.0536.66.9325.072431.86.8977.0538
0.1037.11.9370.063232.14.9015.0515
0.1537.21.9376.060132.19.9019.0509
0.2037.16.9369.059632.20.9016.0499
0.2537.11.9358.058932.21.9015.0496
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.497, + 0.39, + 0.513 + ], + "angle": 0, + "content": "4.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.787, + 0.571 + ], + "angle": 0, + "content": "We evaluate the performance of our method under various configurations. To quantify the differences, we exploit two synthetic datasets (i.e. Livingroom and Tanabata) for the experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.666 + ], + "angle": 0, + "content": "Effect of the number of virtual sharp images. We evaluate the effect of different numbers of the interpolated virtual images as mentioned in Eq. (6). The experimental results are presented in Table 1. It demonstrates that more virtual images deliver better image quality, while requiring more computational resources. By compromising the image quality and computation requirement, we choose \\( n = 19 \\) for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Effect of time lengths for event accumulation. We study the effect of different time lengths \\(\\Delta t\\) for event accumulation as in Eq. (7). The timestamps of the event stream are normalized to a range of \\((0,1)\\) by its total time length. We choose different values from 0.05 to 0.25 for the ablation study. The experimental results are presented in Table 2, showing that the performance on Tanabata dataset gradually saturates as \\(\\alpha\\) increases until to 0.25, whereas on Livingroom dataset, performance initially improves with increasing \\(\\alpha\\) but subsequently declines. This may depend on the noise level in event stream of different time lengths. In all experiments, we select \\(\\alpha = 0.1\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effect of trajectory representations. We explore the difference between linear interpolation and cubic B-Spline to represent the motion trajectory. The" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.474, + 0.129 + ], + "angle": 0, + "content": "W. Li, P. Wan, P. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 3: Ablation studies on trajectory representations. The results demonstrate that cubic B-spline can deliver better performance than linear interpolation." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.185, + 0.752, + 0.24 + ], + "angle": 0, + "content": "
LivingroomTanabata
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
linear interpolation34.16.9035.113327.42.8164.1117
cubic B-Spline37.11.9370.063232.14.9015.0515
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.323 + ], + "angle": 0, + "content": "Table 4: Quantitative comparisons on single image deblurring with synthetic datasets. The results demonstrate that our method performs better than prior learning-based methods in terms of image quality. For HINet and NAFNet, we tests pre-trained weights from both GoPro and REDS datasets(\\*). Due to page limits, the results of the SSIM metric can be found in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.336, + 0.752, + 0.615 + ], + "angle": 0, + "content": "
PSNR ↑
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
DeblurGANv2 [19]29.2627.6423.1620.0926.8925.41
SRN-deblur [46]30.8627.5923.1219.8927.7925.85
MPRNet [55]28.5726.4921.6018.2027.0224.38
HINet [5]28.5626.2721.9118.5926.7024.41
HINet* [5]27.5522.8920.2518.1527.1423.20
NAFNet [4]29.9228.1622.4118.9626.7525.24
NAFNet* [4]28.1823.6720.8518.3827.5223.72
Restormer [54]29.4827.3922.2218.8227.3525.05
BeNeRF37.1132.9529.6832.1436.3833.65
LPIPS ↓
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
DeblurGANv2 [19].2087.1989.2608.3934.3100.2744
SRN-deblur [46].2529.2503.3245.4260.3594.3226
MPRNet [55].2621.2564.3586.4173.3679.3325
HINet [5].2468.2620.3500.4024.3355.3193
HINet* [5].3327.3602.3789.5265.4397.4076
NAFNet [4].2268.1991.3058.3908.3280.2901
NAFNet* [4].3182.3566.3943.5271.4257.4044
Restormer [54].2391.2493.3373.4248.3664.3234
BeNeRF.0632.0788.0761.0515.0677.0675
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.689 + ], + "angle": 0, + "content": "experimental results shown in Table 3 demonstrate that cubic B-Spline deliver better performance on complex motions than that of the linear interpolation. We exploit cubic B-Spline for the experiments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.464, + 0.726 + ], + "angle": 0, + "content": "4.3 Quantitative evaluations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.809 + ], + "angle": 0, + "content": "To evaluate the performance of our method, we compare it against single-image deblurring methods, event-enhanced single-image deblurring methods, and NeRF-based image deblurring methods requiring multi-view information on both synthetic datasets and real datasets. The experimental results are presented in Table 4, Table 5, Table 6 and Table 7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In particular, we compare against SRN-Deblur [46], DeblurGANv2 [19], MPR-Net [55], HINet [5], NAFNet [4], Restormer [54] in terms of single image deblur-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BeNeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 5: Quantitative comparisons on event-enhanced single image deblurring with synthetic datasets. The results demonstrate that our method performs better than both EDI and eSLNet. Due to page limits, the results of the SSIM metric can be found in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.213, + 0.754, + 0.352 + ], + "angle": 0, + "content": "
PSNR ↑
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
eSLNet [47]14.2210.8110.498.8611.8011.24
EDI [33]32.6130.3327.2424.8731.6429.34
BeNeRF37.1132.9529.6832.1436.3833.65
LPIPS ↓
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
eSLNet [47].3981.4236.4902.5067.4676.4572
EDI [33].0904.1020.0779.1039.1409.1030
BeNeRF.0632.0788.0761.0515.0677.0675
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.784, + 0.442 + ], + "angle": 0, + "content": "ring. The experimental results shown in Table 4 demonstrates that our method significantly outperforms prior state-of-the-art methods. It shows that prior learning-based methods have limited generalization performance on domain-shifted images, especially with large motion blurs." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Table 6: Quantitative comparisons on NeRF-based image deblurring with synthetic datasets from \\(\\mathbf{E}^2\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}\\). The results indicate that our method outperforms both NeRF and Deblur-NeRF, and exhibits performance comparable to \\(\\mathrm{E}^2\\mathrm{N}\\mathrm{e}\\mathrm{R}\\mathrm{F}\\) in terms of the PSNR metric. Moreover, our method even surpasses \\(\\mathrm{E}^2\\mathrm{N}\\mathrm{e}\\mathrm{R}\\mathrm{F}\\) with the LPIPS metric. Due to page limits, the results of the SSIM metric can be found in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.563, + 0.754, + 0.722 + ], + "angle": 0, + "content": "
PSNR ↑
ChairFicusHotdogLegoMaterialsMicAverage
NeRF [28]24.2922.9827.7521.9519.9920.5022.91
Deblur-NeRF [26]25.8722.8624.6224.4720.5411.9221.71
E²NeRF [34]31.2830.0034.3428.1127.2727.6029.77
BeNeRF31.1730.8134.3128.0927.4426.1329.66
ChairFicusHotdogLPIPS ↓
LegoMaterialsMicAverage
NeRF [28].1254.1037.1158.2103.1512.1579.1441
Deblur-NeRF [26].2185.1541.2138.2053.2562.3706.2364
E²NeRF [34].0608.0362.0660.1078.0919.0724.0725
BeNeRF.0500.0299.0539.0745.0708.0738.0588
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We also compare against prior event-enhanced single image deblurring methods, such as EDI [33] and eSLNet [47]. The results in Table 5 demonstrate that our method has superior performance when compared to them. eSLNet demonstrates poor generalization performance, since we are unable to fine-tune it on our evaluation datasets. It demonstrates the benefit on incorporating event streams to enhance single image deblurring task under the framework of NeRF." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.473, + 0.129 + ], + "angle": 0, + "content": "W. Li, P. Wan, P. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "Furthermore, we conducted detailed comparisons with NeRF-based image deblurring methods that require multi-view information on the synthetic dataset proposed by \\(\\mathrm{E}^2\\mathrm{NeRF}\\). We compared against NeRF [28], Deblur-NeRF [26] and \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [34]. The experimental results in Table 6 demonstrate that despite utilizing only a single blurred image and event stream of a small time interval, our method achieves performance comparable to \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [34], which utilizes multiview images and a longer event stream, in terms of PSNR metric. Moreover, our method even surpasses \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [34] in terms of the LPIPS metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Finally, we select the best-performing algorithms on the synthetic dataset, excluding our method, from single-image deblurring methods, event-enhanced single-image deblurring methods, and NeRF-based image deblurring methods, which are SRN-deblur [46], EDI [33], and \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [34], respectively. We compare against with these methods on the real dataset and provide the BRISQUE [29] metric. The results in Table 7 indicate a significant improvement of our method over the aforementioned methods. This is attributed to our method's incorporation of a physical model for the imaging process of blurry images, enabling better performance on real-world datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Table 7: Quantitative comparisons on real-world datasets from \\(\\mathbf{E}^2\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}\\). We exploit the used real-world dataset proposed by E2NeRF [34] for the evaluations, which is collected via a DAVIS event camera. The results indicate that our method outperforms EDI, SRN-Deblur and even \\(\\mathrm{E}^2\\mathrm{NeRF}\\) on the real-world datasets. Note that \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [34] requires multi-view images while ours only need a single image. Since \\(\\mathrm{E}^2\\mathrm{NeRF}\\) does not provide the trained model and the code for the metric computation, we re-trained \\(\\mathrm{E}^2\\mathrm{NeRF}\\) for this experiment and compute the metric with the MATLAB implementation of the BRISQUE metric for fair comparisons." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.556, + 0.752, + 0.635 + ], + "angle": 0, + "content": "
BRISQUE ↓
CameraLegoLetterPlantToysAverage
EDI [33]29.7429.3528.7431.0937.0931.20
SRN-Deblur [46]32.2034.9140.8237.4546.1038.30
E²NeRF [34]33.4033.8537.4132.0243.0035.94
BeNeRF19.4725.8627.3721.4625.2023.87
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.691, + 0.452, + 0.706 + ], + "angle": 0, + "content": "4.4 Qualitative evaluations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The qualitative evaluation results are shown in Fig. 3 and Fig. 4 for both synthetic and real datasets respectively. The experimental results demonstrate that our method delivers better performance than prior methods even when the image is severely blurred. In particular, Fig. 3 shows that prior learning-based methods struggle to generalize to domain-shifted images. Notably, EDI [33] performs well on synthetic datasets due to the high quality of event data. Fig. 4 shows that our method outperforms all prior methods (even trained with multi-view images) on real noisy dataset, which demonstrates the advantage of our method" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BeNeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.144, + 0.777, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.551, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Fig. 3: Qualitative results of different methods with synthetic datasets. It demonstrates that our method delivers better performance compared to prior approaches. The learning based methods fail to generalize on severely blurry images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.652 + ], + "angle": 0, + "content": "and the necessity to jointly optimize the camera motion and the implicit 3D representation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.36, + 0.687 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.79 + ], + "angle": 0, + "content": "In conclusion, we present a novel method to jointly recover the underlying 3D scene representation and camera motion trajectory from a single blurry image and its corresponding event stream. Extensive experimental evaluations with both synthetic and real datasets demonstrate the superior performance of our method over prior works, even for those requiring multi-view images and longer event streams." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.474, + 0.129 + ], + "angle": 0, + "content": "W. Li, P. Wan, P. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.773, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.797, + 0.788, + 0.854 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results of different methods on the real datasets. The experimental results demonstrate that our method delivers superior performance on the real DAVIS datasets from \\(\\mathrm{E}^2\\mathrm{NeRF}\\). We are even achieve better performance than prior methods requiring multi-view images and longer event stream." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BeNeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Unreal Engine: The most powerful real-time 3D creation tool. https://www.unrealengine.com/en-US/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Bian, W., Wang, Z., Li, K., Bian, J., Prisacariu, V.A.: NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.273 + ], + "angle": 0, + "content": "3. Cai, S., Obukhov, A., Dai, D., Van Gool, L.: Pix2NeRF: Unsupervised Conditional pi-GAN for Single Image to Neural Radiance Fields Translation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple Baselines for Image Restoration. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 17-33. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-20071-7_2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "5. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: HINet: Half Instance Normalization Network for Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.425 + ], + "angle": 0, + "content": "6. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12882-12891 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "7. Foundation, B.: Blender.org - Home of the Blender project - Free and Open 3D Creation Software" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.52 + ], + "angle": 0, + "content": "8. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (Jan 2022). https://doi.org/10.1109/TPAMI.2020.3008413" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "9. Hidalgo-Carrió, J., Gallego, G., Scaramuzza, D.: Event-aided Direct Sparse Odometry (Apr 2022). https://doi.org/10.48550/arXiv.2204.07640" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "10. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 837-847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "1. Jia, J.: Single image motion deblurring using transparency. In: 2007 IEEE Conference on computer vision and pattern recognition. pp. 1-8. IEEE (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "2. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning Event-Based Motion Deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "3. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "4. Joshi, N., Zitnick, C.L., Szeliski, R., Kriegman, D.J.: Image deblurring and denoising using color priors. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1550-1557. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "5. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization (Jan 2017). https://doi.org/10.48550/arXiv.1412.6980" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-NeRF: Neural Radiance Fields From a Moving Event Camera. IEEE Robotics and Automation Letters 8(3), 1587–1594 (Mar 2023). https://doi.org/10.1109/LRA.2023.3240646" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.474, + 0.129 + ], + "angle": 0, + "content": "W. Li, P. Wan, P. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8183-8192 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "19. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8878-8887 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "21. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "22. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields. In: International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "23. Lichtsteiner, P., Posch, C., Delbruck, T.: A 128\\(\\times\\) times 128 120 dB 15 Ms Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits 43(2), 566-576 (Feb 2008). https://doi.org/10.1109/JSSC.2007.914337" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "24. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: BARF: Bundle-Adjusting Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5741-5751 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "25. Low, W.F., Lee, G.H.: Robust e-nerf: Nerf from sparse & noisy events under nonuniform motion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "26. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural Radiance Fields From Blurry Images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "27. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "28. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis (Aug 2020). https://doi.org/10.48550/arXiv.2003.08934" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "29. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing 21(12), 4695-4708 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (Jul 2022). https://doi.org/10.1145/3528223.3530127" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "31. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.785 + ], + "angle": 0, + "content": "32. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "33. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a Blurry Frame Alive at High Frame-Rate With an Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BeNeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Qi, Y., Zhu, L., Zhang, Y., Li, J.: E2NeRF: Event Enhanced Neural Radiance Fields from Blurry Images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13254-13264 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.242 + ], + "angle": 0, + "content": "35. Qin, K.: General matrix representations for B-splines. In: Proceedings Pacific Graphics '98. Sixth Pacific Conference on Computer Graphics and Applications (Cat. No.98EX208). pp. 37-43 (Oct 1998). https://doi.org/10.1109/PCCGA.1998.731996" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.271 + ], + "angle": 0, + "content": "36. Rebain, D., Matthews, M., Yi, K.M., Lagun, D., Tagliasacchi, A.: LOLNeRF: Learn from One Look. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.311 + ], + "angle": 0, + "content": "37. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: An Open Event Camera Simulator. In: Proceedings of The 2nd Conference on Robot Learning. pp. 969-982. PMLR (Oct 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.338 + ], + "angle": 0, + "content": "38. Rematas, K., Martin-Brualla, R., Ferrari, V.: ShaRF: Shape-conditioned Radiance Fields from a Single View. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.379 + ], + "angle": 0, + "content": "39. Rudnev, V., Elgharib, M., Theobalt, C., Golyanik, V.: EventNeRF: Neural Radiance Fields From a Single Colour Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4992-5002 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.405 + ], + "angle": 0, + "content": "40. Schonberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.446 + ], + "angle": 0, + "content": "41. Schops, T., Sattler, T., Pollefeys, M.: BAD SLAM: Bundle Adjusted Direct RGB-D SLAM. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 134-144 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.473 + ], + "angle": 0, + "content": "42. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. Acm transactions on graphics (tog) 27(3), 1-10 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.513 + ], + "angle": 0, + "content": "43. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 769-777 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.594 + ], + "angle": 0, + "content": "44. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-Based Fusion for Motion Deblurring with Cross-modal Attention. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 412-428. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19797-0_24" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.65 + ], + "angle": 0, + "content": "45. Sun, L., Sakaridis, C., Liang, J., Sun, P., Cao, J., Zhang, K., Jiang, Q., Wang, K., Van Gool, L.: Event-based frame interpolation with ad-hoc deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18043-18052 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.69 + ], + "angle": 0, + "content": "46. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-Recurrent Network for Deep Image Deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8174–8182 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.758 + ], + "angle": 0, + "content": "47. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event Enhanced High-Quality Image Recovery. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 155-171. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58601-0_10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "48. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.839 + ], + "angle": 0, + "content": "49. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF-: Neural Radiance Fields Without Known Camera Parameters (Apr 2022). https://doi.org/10.48550/arXiv.2102.07064" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.474, + 0.129 + ], + "angle": 0, + "content": "W. Li, P. Wan, P. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "50. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion Deblurring With Real Events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.245 + ], + "angle": 0, + "content": "51. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 157-170. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "52. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "53. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.383 + ], + "angle": 0, + "content": "54. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient Transformer for High-Resolution Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5728-5739 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-Stage Progressive Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14821-14831 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "56. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.508 + ], + "angle": 0, + "content": "57. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing Event-Based Motion Deblurring in Real-World Scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "58. Zhao, L., Wang, P., Liu, P.: BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting. In: European Conference on Computer Vision (ECCV). Springer (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.536 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_origin.pdf b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82a503191e43ff1d3dfaaaec79a1ab7398f090b4 --- /dev/null +++ b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/15bbfcb3-e561-4706-81eb-bff4256e4add_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85914c5fe2183dc26300886dfad887a66a7b38a6fa08ea2f6be693947f35711a +size 3610247 diff --git a/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/full.md b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7f4397f8751e51a61ac332b1a954e2484baa0c18 --- /dev/null +++ b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/full.md @@ -0,0 +1,295 @@ +# BeNeRF: Neural Radiance Fields from a Single Blurry Image and Event Stream + +Wenpu Li $^{1,5*}$ , Pian Wan $^{1,2*}$ , Peng Wang $^{1,3*}$ , Jinghang Li $^{4}$ , Yi Zhou $^{4}$ , and Peidong Liu $^{1\dagger}$ + +1Westlake University, 2EPFL, 3Zhejiang University, 4Hunan University, 5Guangdong University of Technology + +https://github.com/wu-cvgl/BeNeRF + +![](images/da8780e08dedb3b058109ef09b8272e1f41c44c86471cf9d6d501ad222a46689.jpg) +Fig. 1: Given a single blurry image and its corresponding event stream, BeNeRF can synthesize high-quality novel images along the camera trajectory, recovering a sharp and coherent video from the single blurry image. + +Abstract. Implicit scene representation has attracted a lot of attention in recent research of computer vision and graphics. Most prior methods focus on how to reconstruct 3D scene representation from a set of images. In this work, we demonstrate the possibility to recover the neural radiance fields (NeRF) from a single blurry image and its corresponding event stream. To eliminate motion blur, we introduce event stream to regularize the learning process of NeRF by accumulating it into an image. We model the camera motion with a cubic B-Spline in SE(3) space. Both the blurry image and the brightness change within a time interval, can then be synthesized from the NeRF given the 6-DoF poses interpolated from the cubic B-Spline. Our method can jointly learn both the implicit scene representation and the camera motion by minimizing the differences between the synthesized data and the real measurements without any prior knowledge of camera poses. We evaluate the proposed method with both synthetic and real datasets. The experimental results demonstrate that we are able to render view-consistent latent sharp images from the learned NeRF and bring a blurry image alive in high quality. + +Keywords: Neural Radiance Fields $\cdot$ Event Stream $\cdot$ Pose Estimation $\cdot$ Deblurring $\cdot$ Novel View Synthesis $\cdot$ 3D from a Single Image + +# 1 Introduction + +Neural Radiance Fields (NeRF) [28] has drawn much attention due to its extraordinary ability in representing 3D scenes and synthesizing novel views. Given multi-view sharp RGB images and calibrated camera poses from COLMAP [40], NeRF takes corresponding 3D spatial location and 2D view direction as input, and optimizes a multi-layer perception (MLP) to represent the 3D scene. More recent advanced methods also exploit explicit octree [52], multi-resolution hash encoding [30] etc., to represent the 3D scene to improve both the training and rendering efficiency. + +Prior methods usually rely on multi-view images to learn the 3D representation. Several pioneering works recently attempt to exploit a single image to learn the underlying neural radiance fields [3,36,38,53]. They usually rely on a large dataset to pre-train the networks to learn priors to address the ill-posed problem. A blurry image further aggravates the problem due to the image quality degradation. Although motion blur is usually not preferred by most vision algorithms, they actually encode additional camera motion trajectory and more structural information compared to a sharp image. In this paper, we explore the possibility of recovering the neural radiance fields and camera motion trajectory from a single blurry image. Instead of learning priors from a large dataset as in previous works, we exploit the usage of an additional event stream to better constrain the problem. + +Event stream can be acquired by an event camera [23] which captures pixel intensity changes caused by the relative motion between the static scene and camera. Unlike standard frame-based cameras, event camera captures asynchronous events with very low latency, leading to extremely high temporal resolution [8]. This characteristic compensates with the image formation process of a blurry image (i.e. integral of photon measurements across time). Several prior works thus take advantage of both modalities for high quality single image deblurring [33,44,47]. However, these methods are unable to recover the camera motion trajectory and extract structural details from a single blurry image, thereby limiting their applicability in 3D computer vision tasks. Some NeRF-based methods that incorporate event stream [10,16,25,34,39] demonstrate the capability to achieve image deblurring and accurate reconstruction of neural radiance fields. Nonetheless, these methods necessitate input images from multiple viewpoints alongside event data. In contrast, we explore the usage of only a single blurry image for the NeRF recovery with unknown camera motions in this work. + +We represent the continuous camera motion with a cubic B-Spline in SE(3) space and define it as the trajectory of both frame-based camera and event camera. Given the neural 3D representation and interpolated poses from the cubic B-Spline, we can synthesize both the blurry image and the brightness change within a time interval via the physical image formation process. The NeRF and motion trajectory can then be jointly optimized by minimizing the difference between the synthesized data and the real measurements. To evaluate the performance of our method, we conduct experiments with both synthetic and real datasets. The experimental results demonstrate that our method is able to + +recover the neural radiance fields from a blurry image and its corresponding event stream without knowing prior knowledge of poses. We are thus able to render view-consistent latent sharp images encoded in a single blurred image from learned NeRF, effectively enhancing the quality of the blurry image. The experimental results further demonstrate that our method is even able to reach same performance as $\mathrm{E}^2\mathrm{NeRF}$ [34], which targets for the same problem, but with multi-view images and longer event data. In summary, our key contributions are as follows: + +- We propose a NeRF-based method that can recover the neural 3D representation from a single blurry image and its corresponding event stream, without knowing any ground truth poses; +- Our method is able to estimate the complex continuous trajectory of camera motion during the imaging process from a single blurry image, providing multi-view geometric information; +- We experimentally validate that our approach is able to recover high quality latent sharp images and high frame-rate video from a single blurry image, without any generalization issue. Furthermore, we are able to reach same level of performance as $\mathrm{E}^2\mathrm{NeRF}$ [34], which requires multi-view images and longer event data. + +# 2 Related Work + +We roughly categorize our related works into three main areas: neural implicit scene representation, single image deblurring and event-enhanced image deblurring. + +Neural Implicit Scene Representation. NeRF has attracted lots of attention due to its powerful ability of implicit 3D scene representation and novel view synthesis [28]. Many following works have been proposed to improve NeRF's performance or extend NeRF to other fields. [2,24,49] jointly trained NeRF with inaccurate camera poses. [6,22,26,27,32,53] improved the performance of NeRF with degraded images, including noisy or few images etc. Recently, several event based NeRF [10,17,34,39] have also been proposed. + +We will mainly focus on the methods that are most related to ours. BAD-NeRF [48] aims to recover true underlying 3D scene representation from multiview blurry inputs and inaccurate camera poses estimated from COLMAP [40]. However, BAD-NeRF and its variants [20,58] struggle to address situations where the input is limited to a single image, primarily due to the severe illness of the problem. Event-Enhanced NeRF (E²NeRF) [34] aims to recover the 3D scene representation from multiple blurry images and event streams. To train NeRF without optimizing the camera poses, E²NeRF manually segments the event stream using preset parameters, aiming to recover sharp images individually through EDI [33]. However, in cases of serve camera motion, the recovered images from event stream may still be blurry, leading to inaccurate pose estimation in COLMAP [40]. This two-stage approach introduces errors (i.e. either from + +EDI [33] or COLMAP [40]) and fails to accurately model the continuous camera trajectory. In contrast, we are able to train both the NeRF and the camera motion trajectory jointly from a single blurry image and its corresponding event stream, which is harder than prior methods. + +Single Image Deblurring. A blurred image can be formulated as the convolution result of a sharp image and a kernel. Therefore, classical approaches [11,14,21,42,51] generally regard the deblurring problem as a joint optimization of the blur kernel and the latent virtual sharp images. With the development of deep learning, many learning based end-to-end deblurring methods have been proposed [13, 18, 19, 31, 43, 46, 54]. These methods usually demonstrate better qualitative and quantitative deblurring results. However, learning based deblurring methods are usually trained on a large dataset which contains paired blurry-sharp images. They would thus usually have limited generalization performance to domain-shifted images. Since NeRF is a test-time optimization approach, our method does not have the generalization performance issue. + +Event Enhanced Image Deblurring. Since event camera is able to capture high dynamic temporal information [8], prior methods usually exploit event measurements to enhance the image deblurring performance [12,33,44,45,47,50,57]. EDI [33] is a simple and effective model, which is able to generate a sharp video under various types of blur by solving a single variable non-convex optimization problem. Different from EDI [33], [44,45,47] design end-to-end neural networks for event enhanced image deblurring/frame interpolation. The critical distinction lies in our method's capacity to extract potential camera motion trajectories from the event stream, thereby enhancing subsequent 3D vision tasks with additional geometric insights. + +# 3 Methodology + +Given a single blurry image and its corresponding event stream, our method recovers the underlying 3D scene representation and the camera motion trajectory jointly. The details of our method are shown in Fig. 2. In particular, we represent the 3D scene with neural radiance fields and the camera motion trajectory with a cubic B-Spline in SE(3) space. Both the blurry image and accumulated events within a time interval can thus be synthesized from the 3D scene representation providing the camera poses. The camera trajectory, NeRF, are then optimized by minimizing the difference between the synthesized data and the real measurements. The details are as follows. + +# 3.1 Neural Implicit Representation + +We adopt Multi-layer Perceptron (MLP) to represent the 3D scene as the original NeRF [28]. More advanced representations, such as multi-resolution hash encoding [30], can also be exploited to further improve its performance. In particular, the scene model is represented by a learnable mapping function $\mathbf{F}_{\theta}:(\mathbf{x},\mathbf{d})\to (\mathbf{c},\sigma)$ , which requires a Cartesian coordinates $\mathbf{x}\in \mathbb{R}^3$ and a + +![](images/03fb71c5155deb9571c00c01aa6482d3257aafb8ee7efc28cb8d894c30c4234f.jpg) +Fig. 2: The pipeline of our method. Given a motion blurry image and its corresponding event stream, we aim to recover both the implicit sharp scene representation and its camera motion trajectory within exposure time. We exploit a continuous time representation for the motion trajectory, and maximize the coherence between both the real measurements and synthesized data for the recovery. + +viewing direction $\mathbf{d} \in \mathbb{S}^2$ as input, and outputs the corresponding volume density $\sigma \in \mathbb{R}$ and color $\mathbf{c} \in \mathbb{R}^3$ . Both the 3D point $\mathbf{x}$ and viewing direction $\mathbf{d}$ are defined in the world coordinate frame. They are a function of the pixel coordinate, camera pose and the corresponding intrinsic parameters. To query the intensity of a pixel, we can apply volume rendering by sampling 3D points along the ray, which originates from the camera center and passes through the pixel. The volume rendering can be formally defined as follows: + +$$ +\mathbf {I} (\mathbf {x}) = \sum_ {i = 1} ^ {n} T _ {i} \left(1 - \exp \left(- \sigma_ {i} \delta_ {i}\right)\right) \boldsymbol {c} _ {i}, \tag {1} +$$ + +where $n$ is the number of sampled points along the ray, $c_{i}$ and $\sigma_{i}$ are the predicted color and volume density of the $i^{th}$ sampled 3D point via $\mathbf{F}_{\theta}$ , $\delta_{i}$ is the distance between the $i^{th}$ and $(i + 1)^{th}$ sampled point, $T_{i}$ is the transmittance which represents the probability that the ray does not hit any particle until the $i^{th}$ sampled point. $T_{i}$ can be computed via: + +$$ +T _ {i} = \exp \left(- \sum_ {k = 1} ^ {i - 1} \sigma_ {k} \delta_ {k}\right). \tag {2} +$$ + +# 3.2 Camera Motion Trajectory Modeling + +We use a differentiable cubic B-Spline in SE(3) space to better model the camera motion trajectory. The spline is represented by a set of learnable control knots + +$\pmb{T}_{c_i}^w \in \mathbb{S}\mathbb{E}(3)$ for $i = 0,1,\dots,n$ . $\pmb{T}_{c_i}^w$ represents the $i^{th}$ control knot, which is defined as a transformation matrix from the camera coordinate frame to world frame. For brevity, we denote $\pmb{T}_{c_i}^w$ with $\pmb{T}_i$ for subsequent derivations. We assume the control knots are sampled with a uniform time interval $\Delta t$ and the trajectory starts from $t_0$ . Spline with a smaller $\Delta t$ can represent a smoother motion, with an expense of more control knots to optimize. Since four consecutive control knots determine the value of the spline curve at a particular timestamp, we can thus compute the starting index of the four control knots for time $t$ by: + +$$ +k = \left\lfloor \frac {t - t _ {0}}{\Delta t} \right\rfloor , \tag {3} +$$ + +where $\lfloor *\rfloor$ is the floor operator. Then we can obtain the four control knots responsible for time $t$ as $\pmb{T}_k,\pmb{T}_{k + 1},\pmb{T}_{k + 2}$ and $\pmb{T}_{k + 3}$ . We can further define $u = \frac{t - t_0}{\Delta t} -k$ , where $u\in [0,1)$ to transform $t$ into a uniform time representation. Using this time representation and based on the matrix representation for the De Boor-Cox formula [35], we can write the matrix representation of a cumulative basis $\mathcal{B}(\mathrm{u})$ as + +$$ +\mathcal {B} (u) = \mathcal {M} \left[ \begin{array}{l} 1 \\ u \\ u ^ {2} \\ u ^ {3} \end{array} \right], \quad \mathcal {M} = \frac {1}{6} \left[ \begin{array}{c c c c} 6 & 0 & 0 & 0 \\ 5 & 3 & - 3 & 1 \\ 1 & 3 & 3 & - 2 \\ 0 & 0 & 0 & 1 \end{array} \right]. \tag {4} +$$ + +The pose at time $t$ can be computed as: + +$$ +\boldsymbol {T} (u) = \boldsymbol {T} _ {k} \cdot \prod_ {j = 0} ^ {2} \exp (\mathcal {B} (u) _ {j + 1} \cdot \boldsymbol {\Omega} _ {k + j}), \tag {5} +$$ + +where $\mathcal{B}(u)_{j+1}$ denotes the $(j+1)^{th}$ element of the vector $\mathcal{B}(u)$ , $\pmb{\Omega}_{k+j} = \log(\pmb{T}_{k+j}^{-1} \cdot \pmb{T}_{k+j+1})$ . + +Since we only consider the continuous camera motion corresponding to a single blurry image, the time interval is thus usually short. We found that four control knots would be sufficient to deliver satisfying results. Therefore, we use the minimal configuration for the following experiments and they are initialized randomly around the identity pose. + +# 3.3 Blurry Image Formation Model + +A motion blurred image $\mathbf{B}(\mathbf{x})\in \mathbb{R}^{\mathrm{W}\times \mathrm{H}\times 3}$ is physically formed by collecting photons during the exposure time and it can be mathematically modeled as: + +$$ +\mathbf {B} (\mathbf {x}) \approx \frac {1}{n} \sum_ {i = 0} ^ {n - 1} \mathbf {I} _ {\mathrm {i}} (\mathbf {x}), \tag {6} +$$ + +where both W and H are the width and height of the image respectively, $n$ is the number of sampled images, $\mathbf{x} \in \mathbb{R}^2$ represents the pixel location, $\mathbf{I}_i(\mathbf{x}) \in \mathbb{R}^{W \times H \times 3}$ is the $i^{th}$ virtual sharp image sampled within the exposure time. The virtual sharp images can be rendered from the neural 3D scene representation along the previous defined camera trajectory. It can be seen that $\mathbf{B}(\mathbf{x})$ is differentiable with respect to the parameters of NeRF and the motion trajectory. + +# 3.4 Event Data Formation Model + +An event camera records changes of the brightness as a stream of events asynchronously. Every time a pixel brightness change reaches a contrast threshold (i.e. $|L(\mathbf{x}, t_i + \delta t) - L(\mathbf{x}, t_i)| \geq C$ ), the camera will trigger an event $e_i = (\mathbf{x}, t_i, p_i)$ , where $p_i \in (-1, +1)$ is the polarity of the event, $L(\mathbf{x}, t_i) = \log(\mathbf{I}(\mathbf{x}, t_i))$ is the brightness logarithm of pixel $\mathbf{x}$ at timestamp $t_i$ , $C$ is the contrast threshold. + +To relate NeRF representation with the event stream, we accumulate the real measured events within a time interval $\Delta t$ to an image $\mathbf{E}(\mathbf{x})$ . The accumulation is defined as: + +$$ +\mathbf {E} (\mathbf {x}) = C \left\{e _ {i} \left(\mathbf {x}, t _ {i}, p _ {i}\right) \right\} _ {t _ {k} < t _ {i} < t _ {k} + \Delta t}, \tag {7} +$$ + +where $e(\mathbf{x}, t_i, p_i)$ is the $i^{th}$ event within the defined time interval corresponding to pixel $\mathbf{x}$ . For real event cameras, the contrast threshold $C$ changes over time and varies pixel by pixel. We therefore normalize the accumulated event as in [9, 17] to eliminate the effect of unknown $C$ : + +$$ +\mathbf {E} _ {n} (\mathbf {x}) = \frac {\mathbf {E} (\mathbf {x})}{\| \mathbf {E} (\mathbf {x}) \| _ {2}}, \tag {8} +$$ + +Given the interpolated start pose and end pose corresponding to the time interval $\Delta t$ from the spline, we are able to render two gray-scale images (i.e. $\mathbf{I}_{start}$ and $\mathbf{I}_{end}$ ) from NeRF. The synthesized accumulated event image $\hat{\mathbf{E}}$ can then be computed as: + +$$ +\mathbf {E} (\mathbf {x}) = \log \left(\mathbf {I} _ {\text {e n d}} (\mathbf {x})\right) - \log \left(\mathbf {I} _ {\text {s t a r t}} (\mathbf {x})\right), \tag {9} +$$ + +where $\hat{\mathbf{E}}(\mathbf{x})$ depends on the parameters of both the cubic spline and NeRF, and is differentiable with respect to them. We can also normalize $\hat{\mathbf{E}}(\mathbf{x})$ to $\hat{\mathbf{E}}_n(\mathbf{x})$ similarly as in Eq. (8) for loss computation. + +# 3.5 Loss Functions + +We minimize the sum of a photo-metric loss $\mathcal{L}_p$ and an event loss $\mathcal{L}_e$ : + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {p} + \beta \mathcal {L} _ {e}, \tag {10} +$$ + +where $\mathcal{L}_p$ represents the loss for the frame-based camera, $\mathcal{L}_e$ represents the loss for the accumulated events within a randomly sampled time interval, and $\beta$ is a hyper-parameter. Both losses are defined as follows: + +$$ +\mathcal {L} _ {p} = \left\| \mathbf {B} (\mathbf {x}) - \hat {\mathbf {B}} (\mathbf {x}) \right\| ^ {2}, \tag {11} +$$ + +$$ +\mathcal {L} _ {e} = \left\| \mathbf {E} _ {n} (\mathbf {x}) - \hat {\mathbf {E}} _ {n} (\mathbf {x}) \right\| ^ {2}, \tag {12} +$$ + +where $\hat{\mathbf{B}} (\mathbf{x})$ is the real captured blurry image. + +# 4 Experiments + +# 4.1 Experimental Setup + +Synthetic datasets. We generate synthetic datasets for both quantitative and qualitative evaluations via Unreal Engine [1] and Blender [7]. To have a more realistic synthesis, we interpolate the real camera motion trajectories from ETH3D [41] to render high frame-rate images. In total, we generate three sequences (i.e. livingroom, whiteroom and pinkcastle) via Unreal Engine and two sequences (i.e. tanabata and outdoorpool) via Blender. For thorough evaluations, we synthesized twenty blurry images and corresponding event streams for each sequence. The event streams are generated via ESIM [37] from high frame-rate video. Furthermore, we additionally employed the synthetic dataset proposed by $\mathrm{E}^2\mathrm{NeRF}$ to compare our method with NeRF-based image-deblur methods which require multi-view training data. This dataset includes synthesized blurry images paired with their respective event streams, which expands upon the six scenes (i.e. chair, ficus, hotdog, lego, materials and mic). + +Real-world datasets. We utilized the real-world datasets proposed by $\mathrm{E}^2\mathrm{NeRF}$ , captured using the DAVIS346 color event camera in real-world scenarios. The dataset encompasses five challenging scenes (i.e. letter, lego, camera, plant and toys). The exposure time for RGB frames was set to 100ms, resulting in occurrences of complex camera motion and severe motion blur within the time interval. + +Baseline methods and evaluation metrics. To evaluate the performance of our method in terms of image deblurring, we compare it against state-of-the-art deep learning-based single image deblur methods, i.e. SRN-Deblur [46], HINet [5], DeblurGANv2 [19], MPRNet [55], NAFNet [4] and Restormer [54], as well as event-enhanced single image-based deblur methods, i.e. EDI [33], eSLNet [47]. We also compared our method with NeRF-based image deblur method using multi-view information. The quality of the rendered image is evaluated with the commonly used PSNR, SSIM and LPIPS [56] metrics. Since the lack of sharp reference images in real-world datasets, we conducted quantitative analysis experiments on five real scenes using the no-reference image quality assessment metrics BRISQUE [29]. + +Implementation details. We implement our method with PyTorch. The implicit representation of the scene from MLP (i.e. $F_{\theta}$ ) is built from NeRF [28] without any modification. We randomly initialize trajectory control knots for cameras within a range of $(0, 0.01)$ . We use two separate Adam optimizers [15] for the scene model (i.e. $F_{\theta}$ ) and camera motion (i.e. $\mathbf{P}_i$ ). The learning rate for the scene model and poses decay from $5 \times 10^{-4}$ with a rate of 0.1 exponentially. In each training step, 1024 pixels for brightness and 1024 pixels for color are sampled. The weight of the event loss $\beta$ is selected to be 0.1 for synthetic datasets and 2 for real-world datasets, respectively. We train our model for 80K iterations for each image and its corresponding event stream. + +Table 1: Ablation studies on the number of virtual sharp images. The results demonstrate that the image quality gradually saturates as the number of virtual sharp images increases. By compromising the image quality and computational efficiency, we select $n = 19$ for all our experiments. + +
LivingroomTanabata
nPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
736.16.9291.071129.28.8533.0712
1136.84.9353.065930.77.8788.0622
1537.03.9368.063531.90.8979.0529
1937.11.9370.063232.14.9015.0515
2337.11.9375.062932.35.9042.0506
+ +Table 2: Ablation studies on event accumulation time lengths. The experimental results demonstrate that the image quality can be affected by the time length. For all experiments, we select $\alpha = 0.1$ for event data accumulation. + +
LivingroomTanabata
αPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
0.0536.66.9325.072431.86.8977.0538
0.1037.11.9370.063232.14.9015.0515
0.1537.21.9376.060132.19.9019.0509
0.2037.16.9369.059632.20.9016.0499
0.2537.11.9358.058932.21.9015.0496
+ +# 4.2 Ablation Study + +We evaluate the performance of our method under various configurations. To quantify the differences, we exploit two synthetic datasets (i.e. Livingroom and Tanabata) for the experiments. + +Effect of the number of virtual sharp images. We evaluate the effect of different numbers of the interpolated virtual images as mentioned in Eq. (6). The experimental results are presented in Table 1. It demonstrates that more virtual images deliver better image quality, while requiring more computational resources. By compromising the image quality and computation requirement, we choose $n = 19$ for our experiments. + +Effect of time lengths for event accumulation. We study the effect of different time lengths $\Delta t$ for event accumulation as in Eq. (7). The timestamps of the event stream are normalized to a range of $(0,1)$ by its total time length. We choose different values from 0.05 to 0.25 for the ablation study. The experimental results are presented in Table 2, showing that the performance on Tanabata dataset gradually saturates as $\alpha$ increases until to 0.25, whereas on Livingroom dataset, performance initially improves with increasing $\alpha$ but subsequently declines. This may depend on the noise level in event stream of different time lengths. In all experiments, we select $\alpha = 0.1$ . + +Effect of trajectory representations. We explore the difference between linear interpolation and cubic B-Spline to represent the motion trajectory. The + +Table 3: Ablation studies on trajectory representations. The results demonstrate that cubic B-spline can deliver better performance than linear interpolation. + +
LivingroomTanabata
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
linear interpolation34.16.9035.113327.42.8164.1117
cubic B-Spline37.11.9370.063232.14.9015.0515
+ +Table 4: Quantitative comparisons on single image deblurring with synthetic datasets. The results demonstrate that our method performs better than prior learning-based methods in terms of image quality. For HINet and NAFNet, we tests pre-trained weights from both GoPro and REDS datasets(\*). Due to page limits, the results of the SSIM metric can be found in the supplementary materials. + +
PSNR ↑
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
DeblurGANv2 [19]29.2627.6423.1620.0926.8925.41
SRN-deblur [46]30.8627.5923.1219.8927.7925.85
MPRNet [55]28.5726.4921.6018.2027.0224.38
HINet [5]28.5626.2721.9118.5926.7024.41
HINet* [5]27.5522.8920.2518.1527.1423.20
NAFNet [4]29.9228.1622.4118.9626.7525.24
NAFNet* [4]28.1823.6720.8518.3827.5223.72
Restormer [54]29.4827.3922.2218.8227.3525.05
BeNeRF37.1132.9529.6832.1436.3833.65
LPIPS ↓
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
DeblurGANv2 [19].2087.1989.2608.3934.3100.2744
SRN-deblur [46].2529.2503.3245.4260.3594.3226
MPRNet [55].2621.2564.3586.4173.3679.3325
HINet [5].2468.2620.3500.4024.3355.3193
HINet* [5].3327.3602.3789.5265.4397.4076
NAFNet [4].2268.1991.3058.3908.3280.2901
NAFNet* [4].3182.3566.3943.5271.4257.4044
Restormer [54].2391.2493.3373.4248.3664.3234
BeNeRF.0632.0788.0761.0515.0677.0675
+ +experimental results shown in Table 3 demonstrate that cubic B-Spline deliver better performance on complex motions than that of the linear interpolation. We exploit cubic B-Spline for the experiments. + +# 4.3 Quantitative evaluations + +To evaluate the performance of our method, we compare it against single-image deblurring methods, event-enhanced single-image deblurring methods, and NeRF-based image deblurring methods requiring multi-view information on both synthetic datasets and real datasets. The experimental results are presented in Table 4, Table 5, Table 6 and Table 7. + +In particular, we compare against SRN-Deblur [46], DeblurGANv2 [19], MPR-Net [55], HINet [5], NAFNet [4], Restormer [54] in terms of single image deblur- + +Table 5: Quantitative comparisons on event-enhanced single image deblurring with synthetic datasets. The results demonstrate that our method performs better than both EDI and eSLNet. Due to page limits, the results of the SSIM metric can be found in the supplementary materials. + +
PSNR ↑
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
eSLNet [47]14.2210.8110.498.8611.8011.24
EDI [33]32.6130.3327.2424.8731.6429.34
BeNeRF37.1132.9529.6832.1436.3833.65
LPIPS ↓
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
eSLNet [47].3981.4236.4902.5067.4676.4572
EDI [33].0904.1020.0779.1039.1409.1030
BeNeRF.0632.0788.0761.0515.0677.0675
+ +ring. The experimental results shown in Table 4 demonstrates that our method significantly outperforms prior state-of-the-art methods. It shows that prior learning-based methods have limited generalization performance on domain-shifted images, especially with large motion blurs. + +Table 6: Quantitative comparisons on NeRF-based image deblurring with synthetic datasets from $\mathbf{E}^2\mathbf{N}\mathbf{e}\mathbf{R}\mathbf{F}$ . The results indicate that our method outperforms both NeRF and Deblur-NeRF, and exhibits performance comparable to $\mathrm{E}^2\mathrm{N}\mathrm{e}\mathrm{R}\mathrm{F}$ in terms of the PSNR metric. Moreover, our method even surpasses $\mathrm{E}^2\mathrm{N}\mathrm{e}\mathrm{R}\mathrm{F}$ with the LPIPS metric. Due to page limits, the results of the SSIM metric can be found in the supplementary materials. + +
PSNR ↑
ChairFicusHotdogLegoMaterialsMicAverage
NeRF [28]24.2922.9827.7521.9519.9920.5022.91
Deblur-NeRF [26]25.8722.8624.6224.4720.5411.9221.71
E²NeRF [34]31.2830.0034.3428.1127.2727.6029.77
BeNeRF31.1730.8134.3128.0927.4426.1329.66
ChairFicusHotdogLPIPS ↓
LegoMaterialsMicAverage
NeRF [28].1254.1037.1158.2103.1512.1579.1441
Deblur-NeRF [26].2185.1541.2138.2053.2562.3706.2364
E²NeRF [34].0608.0362.0660.1078.0919.0724.0725
BeNeRF.0500.0299.0539.0745.0708.0738.0588
+ +We also compare against prior event-enhanced single image deblurring methods, such as EDI [33] and eSLNet [47]. The results in Table 5 demonstrate that our method has superior performance when compared to them. eSLNet demonstrates poor generalization performance, since we are unable to fine-tune it on our evaluation datasets. It demonstrates the benefit on incorporating event streams to enhance single image deblurring task under the framework of NeRF. + +Furthermore, we conducted detailed comparisons with NeRF-based image deblurring methods that require multi-view information on the synthetic dataset proposed by $\mathrm{E}^2\mathrm{NeRF}$ . We compared against NeRF [28], Deblur-NeRF [26] and $\mathrm{E}^2\mathrm{NeRF}$ [34]. The experimental results in Table 6 demonstrate that despite utilizing only a single blurred image and event stream of a small time interval, our method achieves performance comparable to $\mathrm{E}^2\mathrm{NeRF}$ [34], which utilizes multiview images and a longer event stream, in terms of PSNR metric. Moreover, our method even surpasses $\mathrm{E}^2\mathrm{NeRF}$ [34] in terms of the LPIPS metric. + +Finally, we select the best-performing algorithms on the synthetic dataset, excluding our method, from single-image deblurring methods, event-enhanced single-image deblurring methods, and NeRF-based image deblurring methods, which are SRN-deblur [46], EDI [33], and $\mathrm{E}^2\mathrm{NeRF}$ [34], respectively. We compare against with these methods on the real dataset and provide the BRISQUE [29] metric. The results in Table 7 indicate a significant improvement of our method over the aforementioned methods. This is attributed to our method's incorporation of a physical model for the imaging process of blurry images, enabling better performance on real-world datasets. + +Table 7: Quantitative comparisons on real-world datasets from $\mathbf{E}^2\mathbf{N}\mathbf{e}\mathbf{R}\mathbf{F}$ . We exploit the used real-world dataset proposed by E2NeRF [34] for the evaluations, which is collected via a DAVIS event camera. The results indicate that our method outperforms EDI, SRN-Deblur and even $\mathrm{E}^2\mathrm{NeRF}$ on the real-world datasets. Note that $\mathrm{E}^2\mathrm{NeRF}$ [34] requires multi-view images while ours only need a single image. Since $\mathrm{E}^2\mathrm{NeRF}$ does not provide the trained model and the code for the metric computation, we re-trained $\mathrm{E}^2\mathrm{NeRF}$ for this experiment and compute the metric with the MATLAB implementation of the BRISQUE metric for fair comparisons. + +
BRISQUE ↓
CameraLegoLetterPlantToysAverage
EDI [33]29.7429.3528.7431.0937.0931.20
SRN-Deblur [46]32.2034.9140.8237.4546.1038.30
E²NeRF [34]33.4033.8537.4132.0243.0035.94
BeNeRF19.4725.8627.3721.4625.2023.87
+ +# 4.4 Qualitative evaluations + +The qualitative evaluation results are shown in Fig. 3 and Fig. 4 for both synthetic and real datasets respectively. The experimental results demonstrate that our method delivers better performance than prior methods even when the image is severely blurred. In particular, Fig. 3 shows that prior learning-based methods struggle to generalize to domain-shifted images. Notably, EDI [33] performs well on synthetic datasets due to the high quality of event data. Fig. 4 shows that our method outperforms all prior methods (even trained with multi-view images) on real noisy dataset, which demonstrates the advantage of our method + +![](images/6d63bf5c6f1577901a422ff754c0e292ef5fdb27c236317a80193f518da1ab33.jpg) +Fig. 3: Qualitative results of different methods with synthetic datasets. It demonstrates that our method delivers better performance compared to prior approaches. The learning based methods fail to generalize on severely blurry images. + +and the necessity to jointly optimize the camera motion and the implicit 3D representation. + +# 5 Conclusion + +In conclusion, we present a novel method to jointly recover the underlying 3D scene representation and camera motion trajectory from a single blurry image and its corresponding event stream. Extensive experimental evaluations with both synthetic and real datasets demonstrate the superior performance of our method over prior works, even for those requiring multi-view images and longer event streams. + +Acknowledgements. This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation. + +![](images/8ef558275be268dc07a1cbdf2ceb1beed7d81b6dd1db2b410843409e52773d24.jpg) +Fig. 4: Qualitative results of different methods on the real datasets. The experimental results demonstrate that our method delivers superior performance on the real DAVIS datasets from $\mathrm{E}^2\mathrm{NeRF}$ . We are even achieve better performance than prior methods requiring multi-view images and longer event stream. + +# References + +1. Unreal Engine: The most powerful real-time 3D creation tool. https://www.unrealengine.com/en-US/ +2. Bian, W., Wang, Z., Li, K., Bian, J., Prisacariu, V.A.: NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +3. Cai, S., Obukhov, A., Dai, D., Van Gool, L.: Pix2NeRF: Unsupervised Conditional pi-GAN for Single Image to Neural Radiance Fields Translation. In: CVPR (2022) +4. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple Baselines for Image Restoration. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 17-33. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-20071-7_2 +5. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: HINet: Half Instance Normalization Network for Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021) +6. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12882-12891 (2022) +7. Foundation, B.: Blender.org - Home of the Blender project - Free and Open 3D Creation Software +8. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (Jan 2022). https://doi.org/10.1109/TPAMI.2020.3008413 +9. Hidalgo-Carrió, J., Gallego, G., Scaramuzza, D.: Event-aided Direct Sparse Odometry (Apr 2022). https://doi.org/10.48550/arXiv.2204.07640 +10. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 837-847 (2023) +1. Jia, J.: Single image motion deblurring using transparency. In: 2007 IEEE Conference on computer vision and pattern recognition. pp. 1-8. IEEE (2007) +2. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning Event-Based Motion Deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020) +3. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018) +4. Joshi, N., Zitnick, C.L., Szeliski, R., Kriegman, D.J.: Image deblurring and denoising using color priors. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1550-1557. IEEE (2009) +5. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization (Jan 2017). https://doi.org/10.48550/arXiv.1412.6980 +6. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023) +7. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-NeRF: Neural Radiance Fields From a Moving Event Camera. IEEE Robotics and Automation Letters 8(3), 1587–1594 (Mar 2023). https://doi.org/10.1109/LRA.2023.3240646 + +18. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8183-8192 (2018) +19. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8878-8887 (2019) +20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +21. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009) +22. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields. In: International Conference on Learning Representations (ICLR) (2024) +23. Lichtsteiner, P., Posch, C., Delbruck, T.: A 128 $\times$ times 128 120 dB 15 Ms Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits 43(2), 566-576 (Feb 2008). https://doi.org/10.1109/JSSC.2007.914337 +24. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: BARF: Bundle-Adjusting Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5741-5751 (2021) +25. Low, W.F., Lee, G.H.: Robust e-nerf: Nerf from sparse & noisy events under nonuniform motion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +26. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural Radiance Fields From Blurry Images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022) +27. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. CVPR (2022) +28. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis (Aug 2020). https://doi.org/10.48550/arXiv.2003.08934 +29. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing 21(12), 4695-4708 (2012) +30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (Jul 2022). https://doi.org/10.1145/3528223.3530127 +31. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017) +32. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022) +33. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a Blurry Frame Alive at High Frame-Rate With an Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019) + +34. Qi, Y., Zhu, L., Zhang, Y., Li, J.: E2NeRF: Event Enhanced Neural Radiance Fields from Blurry Images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13254-13264 (2023) +35. Qin, K.: General matrix representations for B-splines. In: Proceedings Pacific Graphics '98. Sixth Pacific Conference on Computer Graphics and Applications (Cat. No.98EX208). pp. 37-43 (Oct 1998). https://doi.org/10.1109/PCCGA.1998.731996 +36. Rebain, D., Matthews, M., Yi, K.M., Lagun, D., Tagliasacchi, A.: LOLNeRF: Learn from One Look. In: CVPR (2022) +37. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: An Open Event Camera Simulator. In: Proceedings of The 2nd Conference on Robot Learning. pp. 969-982. PMLR (Oct 2018) +38. Rematas, K., Martin-Brualla, R., Ferrari, V.: ShaRF: Shape-conditioned Radiance Fields from a Single View. In: ICML (2021) +39. Rudnev, V., Elgharib, M., Theobalt, C., Golyanik, V.: EventNeRF: Neural Radiance Fields From a Single Colour Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4992-5002 (2023) +40. Schonberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +41. Schops, T., Sattler, T., Pollefeys, M.: BAD SLAM: Bundle Adjusted Direct RGB-D SLAM. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 134-144 (2019) +42. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. Acm transactions on graphics (tog) 27(3), 1-10 (2008) +43. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 769-777 (2015) +44. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-Based Fusion for Motion Deblurring with Cross-modal Attention. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 412-428. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19797-0_24 +45. Sun, L., Sakaridis, C., Liang, J., Sun, P., Cao, J., Zhang, K., Jiang, Q., Wang, K., Van Gool, L.: Event-based frame interpolation with ad-hoc deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18043-18052 (2023) +46. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-Recurrent Network for Deep Image Deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8174–8182 (2018) +47. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event Enhanced High-Quality Image Recovery. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 155-171. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58601-0_10 +48. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023) +49. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF-: Neural Radiance Fields Without Known Camera Parameters (Apr 2022). https://doi.org/10.48550/arXiv.2102.07064 + +50. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion Deblurring With Real Events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021) +51. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 157-170. Springer (2010) +52. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021) +53. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2021) +54. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient Transformer for High-Resolution Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5728-5739 (2022) +55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-Stage Progressive Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14821-14831 (2021) +56. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018) +57. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing Event-Based Motion Deblurring in Real-World Scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023) +58. Zhao, L., Wang, P., Liu, P.: BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting. In: European Conference on Computer Vision (ECCV). Springer (2024) \ No newline at end of file diff --git a/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/images.zip b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5608e1420686337a3397fd5b5f70e4d2a6d87cf6 --- /dev/null +++ b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13566d4cdef44d84b1f04867efd708542f85e1c36d2c9a5f6986ada6c67578af +size 823692 diff --git a/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/layout.json b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2e350269856891627ae85e5eb9ed1f10d1615af6 --- /dev/null +++ b/2024/BeNeRF_Neural Radiance Fields from a Single Blurry Image and Event Stream/layout.json @@ -0,0 +1,9431 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "text", + "content": "BeNeRF: Neural Radiance Fields from a Single Blurry Image and Event Stream" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": "Wenpu Li" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1,5*}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Pian Wan" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Peng Wang" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Jinghang Li" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Yi Zhou" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", and Peidong Liu" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 201, + 465, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 201, + 465, + 225 + ], + "spans": [ + { + "bbox": [ + 148, + 201, + 465, + 225 + ], + "type": "text", + "content": "1Westlake University, 2EPFL, 3Zhejiang University, 4Hunan University, 5Guangdong University of Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 235, + 384, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 235, + 384, + 247 + ], + "spans": [ + { + "bbox": [ + 228, + 235, + 384, + 247 + ], + "type": "text", + "content": "https://github.com/wu-cvgl/BeNeRF" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 258, + 468, + 364 + ], + "blocks": [ + { + "bbox": [ + 143, + 258, + 468, + 364 + ], + "lines": [ + { + "bbox": [ + 143, + 258, + 468, + 364 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 468, + 364 + ], + "type": "image", + "image_path": "da8780e08dedb3b058109ef09b8272e1f41c44c86471cf9d6d501ad222a46689.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 371, + 482, + 406 + ], + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 406 + ], + "type": "text", + "content": "Fig. 1: Given a single blurry image and its corresponding event stream, BeNeRF can synthesize high-quality novel images along the camera trajectory, recovering a sharp and coherent video from the single blurry image." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 420, + 453, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 420, + 453, + 608 + ], + "spans": [ + { + "bbox": [ + 159, + 420, + 453, + 608 + ], + "type": "text", + "content": "Abstract. Implicit scene representation has attracted a lot of attention in recent research of computer vision and graphics. Most prior methods focus on how to reconstruct 3D scene representation from a set of images. In this work, we demonstrate the possibility to recover the neural radiance fields (NeRF) from a single blurry image and its corresponding event stream. To eliminate motion blur, we introduce event stream to regularize the learning process of NeRF by accumulating it into an image. We model the camera motion with a cubic B-Spline in SE(3) space. Both the blurry image and the brightness change within a time interval, can then be synthesized from the NeRF given the 6-DoF poses interpolated from the cubic B-Spline. Our method can jointly learn both the implicit scene representation and the camera motion by minimizing the differences between the synthesized data and the real measurements without any prior knowledge of camera poses. We evaluate the proposed method with both synthetic and real datasets. The experimental results demonstrate that we are able to render view-consistent latent sharp images from the learned NeRF and bring a blurry image alive in high quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "spans": [ + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": "Keywords: Neural Radiance Fields " + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": " Event Stream " + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": " Pose Estimation " + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": " Deblurring " + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": " Novel View Synthesis " + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 616, + 453, + 639 + ], + "type": "text", + "content": " 3D from a Single Image" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 643, + 476, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 476, + 653 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 476, + 653 + ], + "type": "text", + "content": "* Equal contribution: {liwenpu, wangpeng}@westlake.edu.cn, pianwan@outlook.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 654, + 408, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 408, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 408, + 665 + ], + "type": "text", + "content": "† Corresponding author: Peidong Liu(liupeidong@westlake.edu.cn)." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 234 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) [28] has drawn much attention due to its extraordinary ability in representing 3D scenes and synthesizing novel views. Given multi-view sharp RGB images and calibrated camera poses from COLMAP [40], NeRF takes corresponding 3D spatial location and 2D view direction as input, and optimizes a multi-layer perception (MLP) to represent the 3D scene. More recent advanced methods also exploit explicit octree [52], multi-resolution hash encoding [30] etc., to represent the 3D scene to improve both the training and rendering efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 235, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 481, + 376 + ], + "type": "text", + "content": "Prior methods usually rely on multi-view images to learn the 3D representation. Several pioneering works recently attempt to exploit a single image to learn the underlying neural radiance fields [3,36,38,53]. They usually rely on a large dataset to pre-train the networks to learn priors to address the ill-posed problem. A blurry image further aggravates the problem due to the image quality degradation. Although motion blur is usually not preferred by most vision algorithms, they actually encode additional camera motion trajectory and more structural information compared to a sharp image. In this paper, we explore the possibility of recovering the neural radiance fields and camera motion trajectory from a single blurry image. Instead of learning priors from a large dataset as in previous works, we exploit the usage of an additional event stream to better constrain the problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 379, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 481, + 556 + ], + "type": "text", + "content": "Event stream can be acquired by an event camera [23] which captures pixel intensity changes caused by the relative motion between the static scene and camera. Unlike standard frame-based cameras, event camera captures asynchronous events with very low latency, leading to extremely high temporal resolution [8]. This characteristic compensates with the image formation process of a blurry image (i.e. integral of photon measurements across time). Several prior works thus take advantage of both modalities for high quality single image deblurring [33,44,47]. However, these methods are unable to recover the camera motion trajectory and extract structural details from a single blurry image, thereby limiting their applicability in 3D computer vision tasks. Some NeRF-based methods that incorporate event stream [10,16,25,34,39] demonstrate the capability to achieve image deblurring and accurate reconstruction of neural radiance fields. Nonetheless, these methods necessitate input images from multiple viewpoints alongside event data. In contrast, we explore the usage of only a single blurry image for the NeRF recovery with unknown camera motions in this work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "We represent the continuous camera motion with a cubic B-Spline in SE(3) space and define it as the trajectory of both frame-based camera and event camera. Given the neural 3D representation and interpolated poses from the cubic B-Spline, we can synthesize both the blurry image and the brightness change within a time interval via the physical image formation process. The NeRF and motion trajectory can then be jointly optimized by minimizing the difference between the synthesized data and the real measurements. To evaluate the performance of our method, we conduct experiments with both synthetic and real datasets. The experimental results demonstrate that our method is able to" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 290, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 290, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 290, + 102 + ], + "type": "text", + "content": "W. Li, P. Wan, P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "recover the neural radiance fields from a blurry image and its corresponding event stream without knowing prior knowledge of poses. We are thus able to render view-consistent latent sharp images encoded in a single blurred image from learned NeRF, effectively enhancing the quality of the blurry image. The experimental results further demonstrate that our method is even able to reach same performance as " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " [34], which targets for the same problem, but with multi-view images and longer event data. In summary, our key contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 220, + 481, + 353 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 220, + 480, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 257 + ], + "type": "text", + "content": "- We propose a NeRF-based method that can recover the neural 3D representation from a single blurry image and its corresponding event stream, without knowing any ground truth poses;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 257, + 480, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 293 + ], + "type": "text", + "content": "- Our method is able to estimate the complex continuous trajectory of camera motion during the imaging process from a single blurry image, providing multi-view geometric information;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 293, + 481, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 353 + ], + "type": "text", + "content": "- We experimentally validate that our approach is able to recover high quality latent sharp images and high frame-rate video from a single blurry image, without any generalization issue. Furthermore, we are able to reach same level of performance as " + }, + { + "bbox": [ + 138, + 293, + 481, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 138, + 293, + 481, + 353 + ], + "type": "text", + "content": " [34], which requires multi-view images and longer event data." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 372, + 237, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 237, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 237, + 385 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 399, + 481, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 435 + ], + "type": "text", + "content": "We roughly categorize our related works into three main areas: neural implicit scene representation, single image deblurring and event-enhanced image deblurring." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "type": "text", + "content": "Neural Implicit Scene Representation. NeRF has attracted lots of attention due to its powerful ability of implicit 3D scene representation and novel view synthesis [28]. Many following works have been proposed to improve NeRF's performance or extend NeRF to other fields. [2,24,49] jointly trained NeRF with inaccurate camera poses. [6,22,26,27,32,53] improved the performance of NeRF with degraded images, including noisy or few images etc. Recently, several event based NeRF [10,17,34,39] have also been proposed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "We will mainly focus on the methods that are most related to ours. BAD-NeRF [48] aims to recover true underlying 3D scene representation from multiview blurry inputs and inaccurate camera poses estimated from COLMAP [40]. However, BAD-NeRF and its variants [20,58] struggle to address situations where the input is limited to a single image, primarily due to the severe illness of the problem. Event-Enhanced NeRF (E²NeRF) [34] aims to recover the 3D scene representation from multiple blurry images and event streams. To train NeRF without optimizing the camera poses, E²NeRF manually segments the event stream using preset parameters, aiming to recover sharp images individually through EDI [33]. However, in cases of serve camera motion, the recovered images from event stream may still be blurry, leading to inaccurate pose estimation in COLMAP [40]. This two-stage approach introduces errors (i.e. either from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "BeNeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "EDI [33] or COLMAP [40]) and fails to accurately model the continuous camera trajectory. In contrast, we are able to train both the NeRF and the camera motion trajectory jointly from a single blurry image and its corresponding event stream, which is harder than prior methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 166, + 481, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 481, + 298 + ], + "type": "text", + "content": "Single Image Deblurring. A blurred image can be formulated as the convolution result of a sharp image and a kernel. Therefore, classical approaches [11,14,21,42,51] generally regard the deblurring problem as a joint optimization of the blur kernel and the latent virtual sharp images. With the development of deep learning, many learning based end-to-end deblurring methods have been proposed [13, 18, 19, 31, 43, 46, 54]. These methods usually demonstrate better qualitative and quantitative deblurring results. However, learning based deblurring methods are usually trained on a large dataset which contains paired blurry-sharp images. They would thus usually have limited generalization performance to domain-shifted images. Since NeRF is a test-time optimization approach, our method does not have the generalization performance issue." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 421 + ], + "type": "text", + "content": "Event Enhanced Image Deblurring. Since event camera is able to capture high dynamic temporal information [8], prior methods usually exploit event measurements to enhance the image deblurring performance [12,33,44,45,47,50,57]. EDI [33] is a simple and effective model, which is able to generate a sharp video under various types of blur by solving a single variable non-convex optimization problem. Different from EDI [33], [44,45,47] design end-to-end neural networks for event enhanced image deblurring/frame interpolation. The critical distinction lies in our method's capacity to extract potential camera motion trajectories from the event stream, thereby enhancing subsequent 3D vision tasks with additional geometric insights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 438, + 233, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 233, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 233, + 452 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 570 + ], + "type": "text", + "content": "Given a single blurry image and its corresponding event stream, our method recovers the underlying 3D scene representation and the camera motion trajectory jointly. The details of our method are shown in Fig. 2. In particular, we represent the 3D scene with neural radiance fields and the camera motion trajectory with a cubic B-Spline in SE(3) space. Both the blurry image and accumulated events within a time interval can thus be synthesized from the 3D scene representation providing the camera poses. The camera trajectory, NeRF, are then optimized by minimizing the difference between the synthesized data and the real measurements. The details are as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 586, + 318, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 318, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 318, + 598 + ], + "type": "text", + "content": "3.1 Neural Implicit Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "text", + "content": "We adopt Multi-layer Perceptron (MLP) to represent the 3D scene as the original NeRF [28]. More advanced representations, such as multi-resolution hash encoding [30], can also be exploited to further improve its performance. In particular, the scene model is represented by a learnable mapping function " + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\theta}:(\\mathbf{x},\\mathbf{d})\\to (\\mathbf{c},\\sigma)" + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "text", + "content": ", which requires a Cartesian coordinates " + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 605, + 481, + 667 + ], + "type": "text", + "content": " and a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 290, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 290, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 290, + 102 + ], + "type": "text", + "content": "W. Li, P. Wan, P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 118, + 479, + 308 + ], + "blocks": [ + { + "bbox": [ + 137, + 118, + 479, + 308 + ], + "lines": [ + { + "bbox": [ + 137, + 118, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 479, + 308 + ], + "type": "image", + "image_path": "03fb71c5155deb9571c00c01aa6482d3257aafb8ee7efc28cb8d894c30c4234f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 315, + 482, + 370 + ], + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 370 + ], + "type": "text", + "content": "Fig. 2: The pipeline of our method. Given a motion blurry image and its corresponding event stream, we aim to recover both the implicit sharp scene representation and its camera motion trajectory within exposure time. We exploit a continuous time representation for the motion trajectory, and maximize the coherence between both the real measurements and synthesized data for the recovery." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "content": "viewing direction " + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "content": " as input, and outputs the corresponding volume density " + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "content": ". Both the 3D point " + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "content": " and viewing direction " + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "content": " are defined in the world coordinate frame. They are a function of the pixel coordinate, camera pose and the corresponding intrinsic parameters. To query the intensity of a pixel, we can apply volume rendering by sampling 3D points along the ray, which originates from the camera center and passes through the pixel. The volume rendering can be formally defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 233, + 479, + 481, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 479, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 233, + 479, + 481, + 510 + ], + "type": "interline_equation", + "content": "\\mathbf {I} (\\mathbf {x}) = \\sum_ {i = 1} ^ {n} T _ {i} \\left(1 - \\exp \\left(- \\sigma_ {i} \\delta_ {i}\\right)\\right) \\boldsymbol {c} _ {i}, \\tag {1}", + "image_path": "01d63d3d51be4e32f2d4ccffd4149f0e5bc5672cb558f029ab6ad6eb2ad02620.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " is the number of sampled points along the ray, " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " are the predicted color and volume density of the " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " sampled 3D point via " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\theta}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " is the distance between the " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "(i + 1)^{th}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " sampled point, " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " is the transmittance which represents the probability that the ray does not hit any particle until the " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " sampled point. " + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 514, + 482, + 575 + ], + "type": "text", + "content": " can be computed via:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 257, + 579, + 481, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 579, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 257, + 579, + 481, + 612 + ], + "type": "interline_equation", + "content": "T _ {i} = \\exp \\left(- \\sum_ {k = 1} ^ {i - 1} \\sigma_ {k} \\delta_ {k}\\right). \\tag {2}", + "image_path": "cf133f61fb417f720122e60b4c7a0c9bd70a7708fc6b1c73ebd60a31f7b74e74.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 624, + 349, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 349, + 638 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 349, + 638 + ], + "type": "text", + "content": "3.2 Camera Motion Trajectory Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We use a differentiable cubic B-Spline in SE(3) space to better model the camera motion trajectory. The spline is represented by a set of learnable control knots" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "BeNeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\pmb{T}_{c_i}^w \\in \\mathbb{S}\\mathbb{E}(3)" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "i = 0,1,\\dots,n" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\pmb{T}_{c_i}^w" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": " control knot, which is defined as a transformation matrix from the camera coordinate frame to world frame. For brevity, we denote " + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\pmb{T}_{c_i}^w" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\pmb{T}_i" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": " for subsequent derivations. We assume the control knots are sampled with a uniform time interval " + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": " and the trajectory starts from " + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": ". Spline with a smaller " + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": " can represent a smoother motion, with an expense of more control knots to optimize. Since four consecutive control knots determine the value of the spline curve at a particular timestamp, we can thus compute the starting index of the four control knots for time " + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 115, + 482, + 213 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 278, + 216, + 481, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 278, + 216, + 481, + 237 + ], + "type": "interline_equation", + "content": "k = \\left\\lfloor \\frac {t - t _ {0}}{\\Delta t} \\right\\rfloor , \\tag {3}", + "image_path": "8dcaa929389a42b0298994a96088a1219b9fb25ed211871e280aaa3b78e2230d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\lfloor *\\rfloor" + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "content": " is the floor operator. Then we can obtain the four control knots responsible for time " + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\pmb{T}_k,\\pmb{T}_{k + 1},\\pmb{T}_{k + 2}" + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\pmb{T}_{k + 3}" + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "content": ". We can further define " + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "inline_equation", + "content": "u = \\frac{t - t_0}{\\Delta t} -k" + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "inline_equation", + "content": "u\\in [0,1)" + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "content": " to transform " + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "content": " into a uniform time representation. Using this time representation and based on the matrix representation for the De Boor-Cox formula [35], we can write the matrix representation of a cumulative basis " + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\mathrm{u})" + }, + { + "bbox": [ + 130, + 241, + 482, + 310 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 207, + 310, + 481, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 310, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 207, + 310, + 481, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {B} (u) = \\mathcal {M} \\left[ \\begin{array}{l} 1 \\\\ u \\\\ u ^ {2} \\\\ u ^ {3} \\end{array} \\right], \\quad \\mathcal {M} = \\frac {1}{6} \\left[ \\begin{array}{c c c c} 6 & 0 & 0 & 0 \\\\ 5 & 3 & - 3 & 1 \\\\ 1 & 3 & 3 & - 2 \\\\ 0 & 0 & 0 & 1 \\end{array} \\right]. \\tag {4}", + "image_path": "74eddc96511a32ce701e25c9230a01c84e9d95690e5f4c2d274c2d79f2684a10.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 361, + 309, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 309, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 309, + 373 + ], + "type": "text", + "content": "The pose at time " + }, + { + "bbox": [ + 132, + 361, + 309, + 373 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 361, + 309, + 373 + ], + "type": "text", + "content": " can be computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 225, + 378, + 481, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 378, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 225, + 378, + 481, + 410 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} (u) = \\boldsymbol {T} _ {k} \\cdot \\prod_ {j = 0} ^ {2} \\exp (\\mathcal {B} (u) _ {j + 1} \\cdot \\boldsymbol {\\Omega} _ {k + j}), \\tag {5}", + "image_path": "7964a9129272dfb9cb15fb2c6ea5e1f8a772b19a5dbf0420a67784dfc8b9a056.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(u)_{j+1}" + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "inline_equation", + "content": "(j+1)^{th}" + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "text", + "content": " element of the vector " + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(u)" + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "inline_equation", + "content": "\\pmb{\\Omega}_{k+j} = \\log(\\pmb{T}_{k+j}^{-1} \\cdot \\pmb{T}_{k+j+1})" + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 440, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 500 + ], + "type": "text", + "content": "Since we only consider the continuous camera motion corresponding to a single blurry image, the time interval is thus usually short. We found that four control knots would be sufficient to deliver satisfying results. Therefore, we use the minimal configuration for the following experiments and they are initialized randomly around the identity pose." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 514, + 320, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 320, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 320, + 525 + ], + "type": "text", + "content": "3.3 Blurry Image Formation Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 529, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 555 + ], + "type": "text", + "content": "A motion blurred image " + }, + { + "bbox": [ + 130, + 529, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{B}(\\mathbf{x})\\in \\mathbb{R}^{\\mathrm{W}\\times \\mathrm{H}\\times 3}" + }, + { + "bbox": [ + 130, + 529, + 482, + 555 + ], + "type": "text", + "content": " is physically formed by collecting photons during the exposure time and it can be mathematically modeled as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 264, + 559, + 481, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 559, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 264, + 559, + 481, + 590 + ], + "type": "interline_equation", + "content": "\\mathbf {B} (\\mathbf {x}) \\approx \\frac {1}{n} \\sum_ {i = 0} ^ {n - 1} \\mathbf {I} _ {\\mathrm {i}} (\\mathbf {x}), \\tag {6}", + "image_path": "204545a7f0aaa062597da97e7667de2ee7659be4a7528a34d5f05538def22f68.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "where both W and H are the width and height of the image respectively, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is the number of sampled images, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " represents the pixel location, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i(\\mathbf{x}) \\in \\mathbb{R}^{W \\times H \\times 3}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " virtual sharp image sampled within the exposure time. The virtual sharp images can be rendered from the neural 3D scene representation along the previous defined camera trajectory. It can be seen that " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{B}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is differentiable with respect to the parameters of NeRF and the motion trajectory." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "text", + "content": "W. Li, P. Wan, P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 309, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 309, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 309, + 127 + ], + "type": "text", + "content": "3.4 Event Data Formation Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "text", + "content": "An event camera records changes of the brightness as a stream of events asynchronously. Every time a pixel brightness change reaches a contrast threshold (i.e. " + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "inline_equation", + "content": "|L(\\mathbf{x}, t_i + \\delta t) - L(\\mathbf{x}, t_i)| \\geq C" + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "text", + "content": "), the camera will trigger an event " + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "inline_equation", + "content": "e_i = (\\mathbf{x}, t_i, p_i)" + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "inline_equation", + "content": "p_i \\in (-1, +1)" + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "text", + "content": " is the polarity of the event, " + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "inline_equation", + "content": "L(\\mathbf{x}, t_i) = \\log(\\mathbf{I}(\\mathbf{x}, t_i))" + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "text", + "content": " is the brightness logarithm of pixel " + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "text", + "content": " at timestamp " + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 140, + 479, + 199 + ], + "type": "text", + "content": " is the contrast threshold." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "type": "text", + "content": "To relate NeRF representation with the event stream, we accumulate the real measured events within a time interval " + }, + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "type": "text", + "content": " to an image " + }, + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "type": "text", + "content": ". The accumulation is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 239, + 480, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 239, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 231, + 239, + 480, + 251 + ], + "type": "interline_equation", + "content": "\\mathbf {E} (\\mathbf {x}) = C \\left\\{e _ {i} \\left(\\mathbf {x}, t _ {i}, p _ {i}\\right) \\right\\} _ {t _ {k} < t _ {i} < t _ {k} + \\Delta t}, \\tag {7}", + "image_path": "6c790e405a259989919462b406fae09f854e984f77560bff51da073ebd21bce7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "inline_equation", + "content": "e(\\mathbf{x}, t_i, p_i)" + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "text", + "content": " event within the defined time interval corresponding to pixel " + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "text", + "content": ". For real event cameras, the contrast threshold " + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "text", + "content": " changes over time and varies pixel by pixel. We therefore normalize the accumulated event as in [9, 17] to eliminate the effect of unknown " + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 266, + 316, + 480, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 316, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 266, + 316, + 480, + 341 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {n} (\\mathbf {x}) = \\frac {\\mathbf {E} (\\mathbf {x})}{\\| \\mathbf {E} (\\mathbf {x}) \\| _ {2}}, \\tag {8}", + "image_path": "5b96c2a0ca23f464a8646d8256ff87f3c01145a948dae0445bb665fec03dabe8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "text", + "content": "Given the interpolated start pose and end pose corresponding to the time interval " + }, + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "text", + "content": " from the spline, we are able to render two gray-scale images (i.e. " + }, + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{start}" + }, + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{end}" + }, + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "text", + "content": ") from NeRF. The synthesized accumulated event image " + }, + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{E}}" + }, + { + "bbox": [ + 131, + 350, + 479, + 396 + ], + "type": "text", + "content": " can then be computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 400, + 480, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 400, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 228, + 400, + 480, + 412 + ], + "type": "interline_equation", + "content": "\\mathbf {E} (\\mathbf {x}) = \\log \\left(\\mathbf {I} _ {\\text {e n d}} (\\mathbf {x})\\right) - \\log \\left(\\mathbf {I} _ {\\text {s t a r t}} (\\mathbf {x})\\right), \\tag {9}", + "image_path": "d3d9cc397b016eec382f88246cff70af9384acd3bf883de6ca2c020e75d8128b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 420, + 479, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 479, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 479, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 420, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{E}}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 420, + 479, + 458 + ], + "type": "text", + "content": " depends on the parameters of both the cubic spline and NeRF, and is differentiable with respect to them. We can also normalize " + }, + { + "bbox": [ + 130, + 420, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{E}}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 420, + 479, + 458 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 420, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{E}}_n(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 420, + 479, + 458 + ], + "type": "text", + "content": " similarly as in Eq. (8) for loss computation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 235, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 235, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 235, + 490 + ], + "type": "text", + "content": "3.5 Loss Functions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 502, + 438, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 438, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 438, + 515 + ], + "type": "text", + "content": "We minimize the sum of a photo-metric loss " + }, + { + "bbox": [ + 131, + 502, + 438, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 131, + 502, + 438, + 515 + ], + "type": "text", + "content": " and an event loss " + }, + { + "bbox": [ + 131, + 502, + 438, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 131, + 502, + 438, + 515 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 267, + 527, + 480, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 527, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 267, + 527, + 480, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {p} + \\beta \\mathcal {L} _ {e}, \\tag {10}", + "image_path": "2619eae25a375a2f6787fff39056306ba67eb81ae2a44d76c9ffb3b70ddbc930.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 549, + 479, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 479, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 479, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 549, + 479, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 130, + 549, + 479, + 585 + ], + "type": "text", + "content": " represents the loss for the frame-based camera, " + }, + { + "bbox": [ + 130, + 549, + 479, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 130, + 549, + 479, + 585 + ], + "type": "text", + "content": " represents the loss for the accumulated events within a randomly sampled time interval, and " + }, + { + "bbox": [ + 130, + 549, + 479, + 585 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 549, + 479, + 585 + ], + "type": "text", + "content": " is a hyper-parameter. Both losses are defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 252, + 596, + 480, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 596, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 252, + 596, + 480, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\left\\| \\mathbf {B} (\\mathbf {x}) - \\hat {\\mathbf {B}} (\\mathbf {x}) \\right\\| ^ {2}, \\tag {11}", + "image_path": "d85e8c782cafcf0fe4df46ad040c54947d9a6d1f7f143b7318efb05526bb2017.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 252, + 620, + 480, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 620, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 252, + 620, + 480, + 641 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e} = \\left\\| \\mathbf {E} _ {n} (\\mathbf {x}) - \\hat {\\mathbf {E}} _ {n} (\\mathbf {x}) \\right\\| ^ {2}, \\tag {12}", + "image_path": "b4f94ae8579cb8f5f26166ab56fbd134a3237551640eea2ca8fe5e4377172d34.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 653, + 335, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 335, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 335, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 653, + 335, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}} (\\mathbf{x})" + }, + { + "bbox": [ + 131, + 653, + 335, + 666 + ], + "type": "text", + "content": " is the real captured blurry image." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "BeNeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 146, + 261, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 146, + 261, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 261, + 159 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 175, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 175, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 481, + 331 + ], + "type": "text", + "content": "Synthetic datasets. We generate synthetic datasets for both quantitative and qualitative evaluations via Unreal Engine [1] and Blender [7]. To have a more realistic synthesis, we interpolate the real camera motion trajectories from ETH3D [41] to render high frame-rate images. In total, we generate three sequences (i.e. livingroom, whiteroom and pinkcastle) via Unreal Engine and two sequences (i.e. tanabata and outdoorpool) via Blender. For thorough evaluations, we synthesized twenty blurry images and corresponding event streams for each sequence. The event streams are generated via ESIM [37] from high frame-rate video. Furthermore, we additionally employed the synthetic dataset proposed by " + }, + { + "bbox": [ + 133, + 175, + 481, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 133, + 175, + 481, + 331 + ], + "type": "text", + "content": " to compare our method with NeRF-based image-deblur methods which require multi-view training data. This dataset includes synthesized blurry images paired with their respective event streams, which expands upon the six scenes (i.e. chair, ficus, hotdog, lego, materials and mic)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 334, + 483, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 334, + 483, + 405 + ], + "spans": [ + { + "bbox": [ + 133, + 334, + 483, + 405 + ], + "type": "text", + "content": "Real-world datasets. We utilized the real-world datasets proposed by " + }, + { + "bbox": [ + 133, + 334, + 483, + 405 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 133, + 334, + 483, + 405 + ], + "type": "text", + "content": ", captured using the DAVIS346 color event camera in real-world scenarios. The dataset encompasses five challenging scenes (i.e. letter, lego, camera, plant and toys). The exposure time for RGB frames was set to 100ms, resulting in occurrences of complex camera motion and severe motion blur within the time interval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 410, + 481, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 410, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 410, + 481, + 542 + ], + "type": "text", + "content": "Baseline methods and evaluation metrics. To evaluate the performance of our method in terms of image deblurring, we compare it against state-of-the-art deep learning-based single image deblur methods, i.e. SRN-Deblur [46], HINet [5], DeblurGANv2 [19], MPRNet [55], NAFNet [4] and Restormer [54], as well as event-enhanced single image-based deblur methods, i.e. EDI [33], eSLNet [47]. We also compared our method with NeRF-based image deblur method using multi-view information. The quality of the rendered image is evaluated with the commonly used PSNR, SSIM and LPIPS [56] metrics. Since the lack of sharp reference images in real-world datasets, we conducted quantitative analysis experiments on five real scenes using the no-reference image quality assessment metrics BRISQUE [29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": "Implementation details. We implement our method with PyTorch. The implicit representation of the scene from MLP (i.e. " + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": ") is built from NeRF [28] without any modification. We randomly initialize trajectory control knots for cameras within a range of " + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "(0, 0.01)" + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": ". We use two separate Adam optimizers [15] for the scene model (i.e. " + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": ") and camera motion (i.e. " + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": "). The learning rate for the scene model and poses decay from " + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": " with a rate of 0.1 exponentially. In each training step, 1024 pixels for brightness and 1024 pixels for color are sampled. The weight of the event loss " + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": " is selected to be 0.1 for synthetic datasets and 2 for real-world datasets, respectively. We train our model for 80K iterations for each image and its corresponding event stream." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "text", + "content": "W. Li, P. Wan, P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 168, + 460, + 243 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Ablation studies on the number of virtual sharp images. The results demonstrate that the image quality gradually saturates as the number of virtual sharp images increases. By compromising the image quality and computational efficiency, we select " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "n = 19" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " for all our experiments." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 168, + 460, + 243 + ], + "lines": [ + { + "bbox": [ + 151, + 168, + 460, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 168, + 460, + 243 + ], + "type": "table", + "html": "
LivingroomTanabata
nPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
736.16.9291.071129.28.8533.0712
1136.84.9353.065930.77.8788.0622
1537.03.9368.063531.90.8979.0529
1937.11.9370.063232.14.9015.0515
2337.11.9375.062932.35.9042.0506
", + "image_path": "7270b13ee06e607f84dedd5277e8fc2edca8902d8ec917ac4d1750d917513d75.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 151, + 297, + 460, + 370 + ], + "blocks": [ + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "text", + "content": "Table 2: Ablation studies on event accumulation time lengths. The experimental results demonstrate that the image quality can be affected by the time length. For all experiments, we select " + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "text", + "content": " for event data accumulation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 297, + 460, + 370 + ], + "lines": [ + { + "bbox": [ + 151, + 297, + 460, + 370 + ], + "spans": [ + { + "bbox": [ + 151, + 297, + 460, + 370 + ], + "type": "table", + "html": "
LivingroomTanabata
αPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
0.0536.66.9325.072431.86.8977.0538
0.1037.11.9370.063232.14.9015.0515
0.1537.21.9376.060132.19.9019.0509
0.2037.16.9369.059632.20.9016.0499
0.2537.11.9358.058932.21.9015.0496
", + "image_path": "2d5f2477f619776eaf3ce92a027a5e04dacb953615f581ad852d8305ec695531.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 393, + 238, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 393, + 238, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 393, + 238, + 406 + ], + "type": "text", + "content": "4.2 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 416, + 481, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 481, + 452 + ], + "type": "text", + "content": "We evaluate the performance of our method under various configurations. To quantify the differences, we exploit two synthetic datasets (i.e. Livingroom and Tanabata) for the experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 455, + 482, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 527 + ], + "type": "text", + "content": "Effect of the number of virtual sharp images. We evaluate the effect of different numbers of the interpolated virtual images as mentioned in Eq. (6). The experimental results are presented in Table 1. It demonstrates that more virtual images deliver better image quality, while requiring more computational resources. By compromising the image quality and computation requirement, we choose " + }, + { + "bbox": [ + 130, + 455, + 482, + 527 + ], + "type": "inline_equation", + "content": "n = 19" + }, + { + "bbox": [ + 130, + 455, + 482, + 527 + ], + "type": "text", + "content": " for our experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "text", + "content": "Effect of time lengths for event accumulation. We study the effect of different time lengths " + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "text", + "content": " for event accumulation as in Eq. (7). The timestamps of the event stream are normalized to a range of " + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "text", + "content": " by its total time length. We choose different values from 0.05 to 0.25 for the ablation study. The experimental results are presented in Table 2, showing that the performance on Tanabata dataset gradually saturates as " + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "text", + "content": " increases until to 0.25, whereas on Livingroom dataset, performance initially improves with increasing " + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "text", + "content": " but subsequently declines. This may depend on the noise level in event stream of different time lengths. In all experiments, we select " + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 130, + 530, + 482, + 638 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Effect of trajectory representations. We explore the difference between linear interpolation and cubic B-Spline to represent the motion trajectory. The" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "BeNeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 146, + 460, + 190 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Ablation studies on trajectory representations. The results demonstrate that cubic B-spline can deliver better performance than linear interpolation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 146, + 460, + 190 + ], + "lines": [ + { + "bbox": [ + 151, + 146, + 460, + 190 + ], + "spans": [ + { + "bbox": [ + 151, + 146, + 460, + 190 + ], + "type": "table", + "html": "
LivingroomTanabata
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
linear interpolation34.16.9035.113327.42.8164.1117
cubic B-Spline37.11.9370.063232.14.9015.0515
", + "image_path": "d0ce49f993f5908c3754072faad634fecc30820f96c94f5b0746ad7d24f4cf85.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 151, + 266, + 460, + 487 + ], + "blocks": [ + { + "bbox": [ + 130, + 200, + 480, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 255 + ], + "type": "text", + "content": "Table 4: Quantitative comparisons on single image deblurring with synthetic datasets. The results demonstrate that our method performs better than prior learning-based methods in terms of image quality. For HINet and NAFNet, we tests pre-trained weights from both GoPro and REDS datasets(\\*). Due to page limits, the results of the SSIM metric can be found in the supplementary materials." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 266, + 460, + 487 + ], + "lines": [ + { + "bbox": [ + 151, + 266, + 460, + 487 + ], + "spans": [ + { + "bbox": [ + 151, + 266, + 460, + 487 + ], + "type": "table", + "html": "
PSNR ↑
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
DeblurGANv2 [19]29.2627.6423.1620.0926.8925.41
SRN-deblur [46]30.8627.5923.1219.8927.7925.85
MPRNet [55]28.5726.4921.6018.2027.0224.38
HINet [5]28.5626.2721.9118.5926.7024.41
HINet* [5]27.5522.8920.2518.1527.1423.20
NAFNet [4]29.9228.1622.4118.9626.7525.24
NAFNet* [4]28.1823.6720.8518.3827.5223.72
Restormer [54]29.4827.3922.2218.8227.3525.05
BeNeRF37.1132.9529.6832.1436.3833.65
LPIPS ↓
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
DeblurGANv2 [19].2087.1989.2608.3934.3100.2744
SRN-deblur [46].2529.2503.3245.4260.3594.3226
MPRNet [55].2621.2564.3586.4173.3679.3325
HINet [5].2468.2620.3500.4024.3355.3193
HINet* [5].3327.3602.3789.5265.4397.4076
NAFNet [4].2268.1991.3058.3908.3280.2901
NAFNet* [4].3182.3566.3943.5271.4257.4044
Restormer [54].2391.2493.3373.4248.3664.3234
BeNeRF.0632.0788.0761.0515.0677.0675
", + "image_path": "e51eaddcf767706b39542b85ae12dbbe6d880c073a91b0dd5ff64dd765c152e2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 545 + ], + "type": "text", + "content": "experimental results shown in Table 3 demonstrate that cubic B-Spline deliver better performance on complex motions than that of the linear interpolation. We exploit cubic B-Spline for the experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 562, + 283, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 283, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 283, + 574 + ], + "type": "text", + "content": "4.3 Quantitative evaluations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 640 + ], + "type": "text", + "content": "To evaluate the performance of our method, we compare it against single-image deblurring methods, event-enhanced single-image deblurring methods, and NeRF-based image deblurring methods requiring multi-view information on both synthetic datasets and real datasets. The experimental results are presented in Table 4, Table 5, Table 6 and Table 7." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "In particular, we compare against SRN-Deblur [46], DeblurGANv2 [19], MPR-Net [55], HINet [5], NAFNet [4], Restormer [54] in terms of single image deblur-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "text", + "content": "W. Li, P. Wan, P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 168, + 461, + 278 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 5: Quantitative comparisons on event-enhanced single image deblurring with synthetic datasets. The results demonstrate that our method performs better than both EDI and eSLNet. Due to page limits, the results of the SSIM metric can be found in the supplementary materials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 168, + 461, + 278 + ], + "lines": [ + { + "bbox": [ + 151, + 168, + 461, + 278 + ], + "spans": [ + { + "bbox": [ + 151, + 168, + 461, + 278 + ], + "type": "table", + "html": "
PSNR ↑
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
eSLNet [47]14.2210.8110.498.8611.8011.24
EDI [33]32.6130.3327.2424.8731.6429.34
BeNeRF37.1132.9529.6832.1436.3833.65
LPIPS ↓
LivingroomWhiteroomPinkcastleTanabataOutdoorpoolAverage
eSLNet [47].3981.4236.4902.5067.4676.4572
EDI [33].0904.1020.0779.1039.1409.1030
BeNeRF.0632.0788.0761.0515.0677.0675
", + "image_path": "8fb0681f766cd09096ef964edf457aaabce8c41600ff66b28eea2361a90741fd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 479, + 350 + ], + "type": "text", + "content": "ring. The experimental results shown in Table 4 demonstrates that our method significantly outperforms prior state-of-the-art methods. It shows that prior learning-based methods have limited generalization performance on domain-shifted images, especially with large motion blurs." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 151, + 445, + 461, + 571 + ], + "blocks": [ + { + "bbox": [ + 130, + 369, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 437 + ], + "type": "text", + "content": "Table 6: Quantitative comparisons on NeRF-based image deblurring with synthetic datasets from " + }, + { + "bbox": [ + 130, + 369, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^2\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}" + }, + { + "bbox": [ + 130, + 369, + 482, + 437 + ], + "type": "text", + "content": ". The results indicate that our method outperforms both NeRF and Deblur-NeRF, and exhibits performance comparable to " + }, + { + "bbox": [ + 130, + 369, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{N}\\mathrm{e}\\mathrm{R}\\mathrm{F}" + }, + { + "bbox": [ + 130, + 369, + 482, + 437 + ], + "type": "text", + "content": " in terms of the PSNR metric. Moreover, our method even surpasses " + }, + { + "bbox": [ + 130, + 369, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{N}\\mathrm{e}\\mathrm{R}\\mathrm{F}" + }, + { + "bbox": [ + 130, + 369, + 482, + 437 + ], + "type": "text", + "content": " with the LPIPS metric. Due to page limits, the results of the SSIM metric can be found in the supplementary materials." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 445, + 461, + 571 + ], + "lines": [ + { + "bbox": [ + 151, + 445, + 461, + 571 + ], + "spans": [ + { + "bbox": [ + 151, + 445, + 461, + 571 + ], + "type": "table", + "html": "
PSNR ↑
ChairFicusHotdogLegoMaterialsMicAverage
NeRF [28]24.2922.9827.7521.9519.9920.5022.91
Deblur-NeRF [26]25.8722.8624.6224.4720.5411.9221.71
E²NeRF [34]31.2830.0034.3428.1127.2727.6029.77
BeNeRF31.1730.8134.3128.0927.4426.1329.66
ChairFicusHotdogLPIPS ↓
LegoMaterialsMicAverage
NeRF [28].1254.1037.1158.2103.1512.1579.1441
Deblur-NeRF [26].2185.1541.2138.2053.2562.3706.2364
E²NeRF [34].0608.0362.0660.1078.0919.0724.0725
BeNeRF.0500.0299.0539.0745.0708.0738.0588
", + "image_path": "102b7cf71884f6b04d9a64bd910c73c08d5344b46a858bc5b80afee2019136db.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We also compare against prior event-enhanced single image deblurring methods, such as EDI [33] and eSLNet [47]. The results in Table 5 demonstrate that our method has superior performance when compared to them. eSLNet demonstrates poor generalization performance, since we are unable to fine-tune it on our evaluation datasets. It demonstrates the benefit on incorporating event streams to enhance single image deblurring task under the framework of NeRF." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "BeNeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "Furthermore, we conducted detailed comparisons with NeRF-based image deblurring methods that require multi-view information on the synthetic dataset proposed by " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ". We compared against NeRF [28], Deblur-NeRF [26] and " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " [34]. The experimental results in Table 6 demonstrate that despite utilizing only a single blurred image and event stream of a small time interval, our method achieves performance comparable to " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " [34], which utilizes multiview images and a longer event stream, in terms of PSNR metric. Moreover, our method even surpasses " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " [34] in terms of the LPIPS metric." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 319 + ], + "type": "text", + "content": "Finally, we select the best-performing algorithms on the synthetic dataset, excluding our method, from single-image deblurring methods, event-enhanced single-image deblurring methods, and NeRF-based image deblurring methods, which are SRN-deblur [46], EDI [33], and " + }, + { + "bbox": [ + 130, + 213, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 213, + 482, + 319 + ], + "type": "text", + "content": " [34], respectively. We compare against with these methods on the real dataset and provide the BRISQUE [29] metric. The results in Table 7 indicate a significant improvement of our method over the aforementioned methods. This is attributed to our method's incorporation of a physical model for the imaging process of blurry images, enabling better performance on real-world datasets." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 151, + 440, + 460, + 502 + ], + "blocks": [ + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "text", + "content": "Table 7: Quantitative comparisons on real-world datasets from " + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^2\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}" + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "text", + "content": ". We exploit the used real-world dataset proposed by E2NeRF [34] for the evaluations, which is collected via a DAVIS event camera. The results indicate that our method outperforms EDI, SRN-Deblur and even " + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "text", + "content": " on the real-world datasets. Note that " + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "text", + "content": " [34] requires multi-view images while ours only need a single image. Since " + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "text", + "content": " does not provide the trained model and the code for the metric computation, we re-trained " + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 342, + 482, + 430 + ], + "type": "text", + "content": " for this experiment and compute the metric with the MATLAB implementation of the BRISQUE metric for fair comparisons." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 440, + 460, + 502 + ], + "lines": [ + { + "bbox": [ + 151, + 440, + 460, + 502 + ], + "spans": [ + { + "bbox": [ + 151, + 440, + 460, + 502 + ], + "type": "table", + "html": "
BRISQUE ↓
CameraLegoLetterPlantToysAverage
EDI [33]29.7429.3528.7431.0937.0931.20
SRN-Deblur [46]32.2034.9140.8237.4546.1038.30
E²NeRF [34]33.4033.8537.4132.0243.0035.94
BeNeRF19.4725.8627.3721.4625.2023.87
", + "image_path": "7bc8628882ca3c7084fb110c1307c4d47595b2dcbd402c87f38b67a44b87e0cd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 547, + 276, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 276, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 276, + 559 + ], + "type": "text", + "content": "4.4 Qualitative evaluations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The qualitative evaluation results are shown in Fig. 3 and Fig. 4 for both synthetic and real datasets respectively. The experimental results demonstrate that our method delivers better performance than prior methods even when the image is severely blurred. In particular, Fig. 3 shows that prior learning-based methods struggle to generalize to domain-shifted images. Notably, EDI [33] performs well on synthetic datasets due to the high quality of event data. Fig. 4 shows that our method outperforms all prior methods (even trained with multi-view images) on real noisy dataset, which demonstrates the advantage of our method" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 289, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 289, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 289, + 102 + ], + "type": "text", + "content": "W. Li, P. Wan, P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 475, + 425 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 475, + 425 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 475, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 475, + 425 + ], + "type": "image", + "image_path": "6d63bf5c6f1577901a422ff754c0e292ef5fdb27c236317a80193f518da1ab33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 436, + 482, + 470 + ], + "lines": [ + { + "bbox": [ + 132, + 436, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 482, + 470 + ], + "type": "text", + "content": "Fig. 3: Qualitative results of different methods with synthetic datasets. It demonstrates that our method delivers better performance compared to prior approaches. The learning based methods fail to generalize on severely blurry images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 491, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 516 + ], + "type": "text", + "content": "and the necessity to jointly optimize the camera motion and the implicit 3D representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 555, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 625 + ], + "type": "text", + "content": "In conclusion, we present a novel method to jointly recover the underlying 3D scene representation and camera motion trajectory from a single blurry image and its corresponding event stream. Extensive experimental evaluations with both synthetic and real datasets demonstrate the superior performance of our method over prior works, even for those requiring multi-view images and longer event streams." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "BeNeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 624 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 624 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 624 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 624 + ], + "type": "image", + "image_path": "8ef558275be268dc07a1cbdf2ceb1beed7d81b6dd1db2b410843409e52773d24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 631, + 482, + 676 + ], + "lines": [ + { + "bbox": [ + 130, + 631, + 482, + 676 + ], + "spans": [ + { + "bbox": [ + 130, + 631, + 482, + 676 + ], + "type": "text", + "content": "Fig. 4: Qualitative results of different methods on the real datasets. The experimental results demonstrate that our method delivers superior performance on the real DAVIS datasets from " + }, + { + "bbox": [ + 130, + 631, + 482, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 130, + 631, + 482, + 676 + ], + "type": "text", + "content": ". We are even achieve better performance than prior methods requiring multi-view images and longer event stream." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "text", + "content": "W. Li, P. Wan, P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Unreal Engine: The most powerful real-time 3D creation tool. https://www.unrealengine.com/en-US/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "2. Bian, W., Wang, Z., Li, K., Bian, J., Prisacariu, V.A.: NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "type": "text", + "content": "3. Cai, S., Obukhov, A., Dai, D., Van Gool, L.: Pix2NeRF: Unsupervised Conditional pi-GAN for Single Image to Neural Radiance Fields Translation. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 270 + ], + "type": "text", + "content": "4. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple Baselines for Image Restoration. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 17-33. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-20071-7_2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "text", + "content": "5. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: HINet: Half Instance Normalization Network for Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "text", + "content": "6. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12882-12891 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "type": "text", + "content": "7. Foundation, B.: Blender.org - Home of the Blender project - Free and Open 3D Creation Software" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 358, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 411 + ], + "type": "text", + "content": "8. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (Jan 2022). https://doi.org/10.1109/TPAMI.2020.3008413" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 435 + ], + "type": "text", + "content": "9. Hidalgo-Carrió, J., Gallego, G., Scaramuzza, D.: Event-aided Direct Sparse Odometry (Apr 2022). https://doi.org/10.48550/arXiv.2204.07640" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "text", + "content": "10. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 837-847 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "type": "text", + "content": "1. Jia, J.: Single image motion deblurring using transparency. In: 2007 IEEE Conference on computer vision and pattern recognition. pp. 1-8. IEEE (2007)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "2. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning Event-Based Motion Deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "3. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 588 + ], + "type": "text", + "content": "4. Joshi, N., Zitnick, C.L., Szeliski, R., Kriegman, D.J.: Image deblurring and denoising using color priors. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1550-1557. IEEE (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "text", + "content": "5. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization (Jan 2017). https://doi.org/10.48550/arXiv.1412.6980" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "6. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "7. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-NeRF: Neural Radiance Fields From a Moving Event Camera. IEEE Robotics and Automation Letters 8(3), 1587–1594 (Mar 2023). https://doi.org/10.1109/LRA.2023.3240646" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "BeNeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8183-8192 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8878-8887 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "21. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "22. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields. In: International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "23. Lichtsteiner, P., Posch, C., Delbruck, T.: A 128" + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": " times 128 120 dB 15 Ms Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits 43(2), 566-576 (Feb 2008). https://doi.org/10.1109/JSSC.2007.914337" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "24. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: BARF: Bundle-Adjusting Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5741-5751 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "25. Low, W.F., Lee, G.H.: Robust e-nerf: Nerf from sparse & noisy events under nonuniform motion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "26. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural Radiance Fields From Blurry Images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "27. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "28. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis (Aug 2020). https://doi.org/10.48550/arXiv.2003.08934" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "29. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing 21(12), 4695-4708 (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (Jul 2022). https://doi.org/10.1145/3528223.3530127" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "31. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "text", + "content": "32. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "33. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a Blurry Frame Alive at High Frame-Rate With an Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "text", + "content": "W. Li, P. Wan, P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 664 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Qi, Y., Zhu, L., Zhang, Y., Li, J.: E2NeRF: Event Enhanced Neural Radiance Fields from Blurry Images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13254-13264 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 191 + ], + "type": "text", + "content": "35. Qin, K.: General matrix representations for B-splines. In: Proceedings Pacific Graphics '98. Sixth Pacific Conference on Computer Graphics and Applications (Cat. No.98EX208). pp. 37-43 (Oct 1998). https://doi.org/10.1109/PCCGA.1998.731996" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 192, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 214 + ], + "type": "text", + "content": "36. Rebain, D., Matthews, M., Yi, K.M., Lagun, D., Tagliasacchi, A.: LOLNeRF: Learn from One Look. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "text", + "content": "37. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: An Open Event Camera Simulator. In: Proceedings of The 2nd Conference on Robot Learning. pp. 969-982. PMLR (Oct 2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "type": "text", + "content": "38. Rematas, K., Martin-Brualla, R., Ferrari, V.: ShaRF: Shape-conditioned Radiance Fields from a Single View. In: ICML (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "type": "text", + "content": "39. Rudnev, V., Elgharib, M., Theobalt, C., Golyanik, V.: EventNeRF: Neural Radiance Fields From a Single Colour Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4992-5002 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 300, + 482, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 320 + ], + "type": "text", + "content": "40. Schonberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 320, + 482, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 353 + ], + "type": "text", + "content": "41. Schops, T., Sattler, T., Pollefeys, M.: BAD SLAM: Bundle Adjusted Direct RGB-D SLAM. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 134-144 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 353, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 374 + ], + "type": "text", + "content": "42. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. Acm transactions on graphics (tog) 27(3), 1-10 (2008)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 374, + 482, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 406 + ], + "type": "text", + "content": "43. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 769-777 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 406, + 482, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 470 + ], + "type": "text", + "content": "44. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-Based Fusion for Motion Deblurring with Cross-modal Attention. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 412-428. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19797-0_24" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 471, + 482, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 514 + ], + "type": "text", + "content": "45. Sun, L., Sakaridis, C., Liang, J., Sun, P., Cao, J., Zhang, K., Jiang, Q., Wang, K., Van Gool, L.: Event-based frame interpolation with ad-hoc deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18043-18052 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 514, + 482, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 546 + ], + "type": "text", + "content": "46. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-Recurrent Network for Deep Image Deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8174–8182 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 546, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 600 + ], + "type": "text", + "content": "47. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event Enhanced High-Quality Image Recovery. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 155-171. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58601-0_10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "48. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 633, + 482, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 664 + ], + "type": "text", + "content": "49. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF-: Neural Radiance Fields Without Known Camera Parameters (Apr 2022). https://doi.org/10.48550/arXiv.2102.07064" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "BeNeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 424 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "50. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion Deblurring With Real Events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 194 + ], + "type": "text", + "content": "51. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 157-170. Springer (2010)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "52. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "53. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "text", + "content": "54. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient Transformer for High-Resolution Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-Stage Progressive Image Restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14821-14831 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "56. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 402 + ], + "type": "text", + "content": "57. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing Event-Based Motion Deblurring in Real-World Scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "text", + "content": "58. Zhao, L., Wang, P., Liu, P.: BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting. In: European Conference on Computer Vision (ECCV). Springer (2024)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 290, + 102 + ], + "type": "text", + "content": "W. Li, P. Wan, P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_content_list.json b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4cc3d41547fddb2e18321b8ac21897940a7e7c50 --- /dev/null +++ b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_content_list.json @@ -0,0 +1,1640 @@ +[ + { + "type": "text", + "text": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zikai Huang $^{1}$ , Xuemiao $\\mathrm{Xu}^{1,3,4(\\boxtimes)}$ , Cheng $\\mathrm{Xu}^{2(\\boxtimes)}$ , Huaidong Zhang $^{1,3}$ , Chenxi Zheng $^{1}$ , Jing Qin $^{2}$ , and Shengfeng He $^{5}$", + "bbox": [ + 225, + 210, + 779, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ South China University of Technology, China xuemx@scut.edu.cn", + "bbox": [ + 339, + 253, + 658, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The Hong Kong Polytechnic University, Hong Kong SAR, China \ncschengxu@gmail.com", + "bbox": [ + 277, + 282, + 722, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guangdong Engineering Center for Large Model and GenAI Technology", + "bbox": [ + 277, + 309, + 746, + 324 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "4 Guangdong Provincial Key Lab of Computational Intelligence and Cyberspace Information", + "5 Singapore Management University, Singapore https://zikaihuangscut.github.io/Beat-It/" + ], + "bbox": [ + 232, + 324, + 767, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db6755f06941388ecccbfa2bb7c69e65b47f8c6fd7368897d48248abded99334.jpg", + "image_caption": [ + "Fig. 1: We introduce Beat-It, a novel method for generating 3D dance motions with beat alignment and motion controllability. Our approach explicitly injects beat awareness and seamlessly integrates multiple conditions to guide the generation process, leading to beat-synchronized, key pose-guided dance generation." + ], + "image_footnote": [], + "bbox": [ + 217, + 417, + 787, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Dance, as an art form, fundamentally hinges on the precise synchronization with musical beats. However, achieving aesthetically pleasing dance sequences from music is challenging, with existing methods often falling short in controllability and beat alignment. To address these shortcomings, this paper introduces Beat-It, a novel framework for beat-specific, key pose-guided dance generation. Unlike prior approaches, Beat-It uniquely integrates explicit beat awareness and key pose guidance, effectively resolving two main issues: the misalignment of generated dance motions with musical beats, and the inability to map key poses to specific beats, critical for practical choreography. Our approach disentangles beat conditions from music using a nearest beat distance representation and employs a hierarchical multi-condition fusion mechanism. This mechanism", + "bbox": [ + 259, + 672, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "seamlessly integrates key poses, beats, and music features, mitigating condition conflicts and offering rich, multi-conditioned guidance for dance generation. Additionally, a specially designed beat alignment loss ensures the generated dance movements remain in sync with the designated beats. Extensive experiments confirm Beat-It's superiority over existing state-of-the-art methods in terms of beat alignment and motion controllability.", + "bbox": [ + 259, + 147, + 743, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Dance generation $\\cdot$ Beat synchronization $\\cdot$ Multi-condition diffusion generation", + "bbox": [ + 261, + 243, + 740, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 295, + 375, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dance, a time-honored art form, resonates across various ages and cultures due to its artistic and aesthetic significance, playing a vital role in the cultural and entertainment sectors. Traditional dance choreography creation, which demands an intricate balance of aesthetic movements, emotional expression, and precise synchronization with musical beats, often presents itself as a costly, laborious, and time-intensive endeavor, even for seasoned artists.", + "bbox": [ + 212, + 325, + 787, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in deep learning, particularly in generative models, have catalyzed efforts to automate the choreography process. Initial approaches predominantly focused on single-condition dance generation, learning a direct mapping from music to dance motions [1, 15, 21-23, 35, 39, 43]. The advent of diffusion models, renowned for their exceptional content generation capabilities, has marked a paradigm shift in this domain. For instance, Tseng et al. [42] introduced a transformer-based diffusion model that sets a new benchmark in music-to-dance conversion, showcasing the potential of these models in elevating the quality of generated dance sequences.", + "bbox": [ + 212, + 417, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notwithstanding the remarkable progress, achieving precise synchronization between dance motions and musical beats, coupled with flexible motion controllability in dance generation, remains a formidable challenge. In real-world choreography, the goal is often to assign specific key poses to particular musical beats and then fill in the connecting movements to form the complete dance sequence. This process requires not only the creation of designated key poses but also their harmonious synchronization with the musical beats. While few methods have explored key pose conditions [42, 52], none of them has considered explicit beat synchronization and controllability, rendering inferior dance motion generation. Furthermore, previous methods typically merge keyframe features with music features through simple concatenation or temporal/spatial blending. This approach, however, is flawed due to the sparse nature of keyframes. Consequently, these methods disproportionately emphasize dense music features at the expense of key pose conditions, leading to a notable misalignment between the generated dance motions and the specified keyframes.", + "bbox": [ + 212, + 553, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To combat the above issues, we aim to completely disentangle beats from input music and simultaneously incorporate explicit beat and key pose guidance to enable beat-controllable, key pose-guided dance generation. We introduce BeatIt, a diffusion-based, multi-condition framework designed for this purpose (see", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1). Our approach begins by formulating beat condition as a nearest beat distance representation, which is a straightforward yet effective method to serve as beat guidance within our framework. To effectively harness and balance the multiple conditions involved, we propose a hierarchical multi-condition fusion mechanism. Initially, we integrate sparse key pose condition with dense beat and music conditions using a tailored beat-aware dilated cross-attention strategy. We further blend refined beat and music features to generate comprehensive multi-condition features. This strategy not only enhances beat and motion controllability in dance generation but also maintains the realism and coherence of the produced dance motions. Additionally, leveraging our beat representation, we introduce a novel beat alignment loss, explicitly ensuring synchronization between dance motions and the given beat condition, thereby significantly enhancing the overall quality of generation. Extensive experiments confirm that our method outperforms current state-of-the-art approaches in terms of beat alignment and motion controllability. Beyond the advantages above, our framework also supports arbitrary beat designation and flexible frame assignments of key poses.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our main contributions are fourfold:", + "bbox": [ + 238, + 388, + 594, + 402 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a multi-condition dance generation framework that achieves beat synchronization and enhanced motion controllability. To the best of our knowledge, our method makes the first attempt at beat-controllable key pose-guided dance generation.", + "- We present a hierarchical multi-condition fusion mechanism to effectively suppress the conflicts and fully exploit the complementary information among different conditions.", + "- We delve into the property of beats and formulate it as a nearest beat distance representation. A beat alignment loss is further tailored to offer explicit supervision signals to the generated dance motions, largely elevating the synchronization between the generated motions and the given beat conditions", + "- Extensive experiments demonstrate our method performs favorably against the state-of-the-art approaches, especially in motion controllability and motion-beat alignment." + ], + "bbox": [ + 223, + 412, + 785, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 647, + 387, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Human Motion Generation", + "text_level": 1, + "bbox": [ + 215, + 679, + 485, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human motion generation aims to generate realistic human motion automatically. Previous works can be generally categorized into three genres according to the input condition: text-conditioned [5,18,24,56,61], audio-conditioned, and scene-conditioned [4, 10, 16, 45]. Our work belongs to audio-conditioned human motion generation. But different from the primary audio-conditioned human motion generation, music-to-dance generation requires a comprehensive consideration of aesthetic movements, emotional expression, and precise synchronization with musical beats, making it extremely challenging to render convincing results. Recently, researchers have widely explored the application of diffusion models", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation", + "bbox": [ + 289, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in human motion generation [2,3,5,16,18,61,62], which shows unprecedented performance than traditional generative models. In this paper, we get a deep insight into the beat synchronization problem and investigate a multi-condition diffusion-based scheme for controllable dance generation.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Dance Generation", + "text_level": 1, + "bbox": [ + 215, + 231, + 413, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the emergence of deep learning [47-49], dance generation has attracted lots of research interest in recent years. Previous methods have explored various frameworks with different types of backbones to achieve single-condition dance generation, including CNN [14], RNN [1, 15, 39, 51], GCN [7, 32], VAE [9, 38], GAN [19, 20, 37], and Transformer [21-23, 35, 36, 43, 53]. FACT [23] proposes a full-attention cross-modal transformer model to capture the correlations between music and dance. Danceformer [21] presents a two-stage deterministic framework for audio-driven single-condition dance generation. It first generates key poses from the input music and then performs interpolation between the key poses. However, it falls short of arbitrary beat choreography and flexible key pose controllability due to its single-condition nature. Bailando [35] utilizes separate VQ-VAEs on upper/lower half bodies and a motion GPT to map the music and seed poses to dance sequences. Several recent attempts have investigated the diffusion models for dance generation [29, 42, 55], which significantly pushes the boundaries of music-to-dance performance. Compared to applying only a single input condition, multi-condition dance generation incorporates additional conditions beyond music into the generation process for better controllability, such as dance style labels [2], text [11], and keyframes [42, 52]. For example, LDA [2] introduced dance style labels as auxiliary conditions. However, it is limited to generating pre-defined styles and lacks flexibility. TM2D [11] adopts a VQ-VAE based model to support dual-modal music and text-driven dance motion generation. Yang et al. [52] employ normalizing flows to model dance motion probability distributions and utilize time embedding to achieve keyframe control. EDGE [42] adopts a diffusion-based framework and enables key pose control via temporal/spatial blending during inference. Although the existing methods have investigated additional conditions for better dance generation control, they typically neglect explicit guidance and supervision to align the generated motion dances with specified beats. Consequently, these methods struggle to produce both beat-specific and key pose-guided dance sequences. In contrast, we propose to explicitly disentangle the beat condition from music and inject flexible beat and key pose controllability into the entire generation process, significantly enhancing the beat alignment while supporting flexible frame assignments of key poses.", + "bbox": [ + 212, + 258, + 789, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Multi Condition Diffusion Generation", + "text_level": 1, + "bbox": [ + 215, + 767, + 576, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-condition diffusion generative models seek to inject multi-condition guidance into the generation process for versatile controllability [6,8,25,28,31,34, 44,50,54,59,60]. Benefiting from large-scale training data, Stable Diffusion [33]", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "constructs a powerful foundation for image generation. Building upon this, ControlNet [57] and T2I-Adapter [27] efficiently integrate trainable parameters into existing models, enabling the inclusion of spatially localized input conditions in a pre-trained text-to-image diffusion model. Uni-ControlNet [58] proposes a framework equipped with two additional control modules for multi-condition control. Qin et al. [30] present a unified model capable of handling various visual conditions, resulting in a more streamlined scheme. Codi [40] acquire multi-modal synthesis skills through the mapping of diverse modalities into a unified space, training on lists of different multi-modal generation tasks. The aforementioned works typically controllable generation using pre-trained models under diverse conditions. Our work, conversely, focuses on the effective and comprehensive fusion of multiple conditions by presenting a hierarchical multi-condition fusion mechanism aimed at better integrating sparse keyframe control conditions with dense music and beat control conditions, resulting in more controllable dance generation.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 396, + 330, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Definition", + "text_level": 1, + "bbox": [ + 215, + 429, + 419, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dance generation aims to create realistic dance motion sequences $\\mathcal{X} = \\{\\mathbf{x}^1, \\mathbf{x}^2, \\mathbf{x}^3, \\dots, \\mathbf{x}^L\\}$ from a given piece of music $\\mathcal{C}$ with a duration of $L$ . Here, $\\mathbf{x}^i \\in \\mathbb{R}^D$ represents a human pose at the $i$ -th frame, which is denoted as a $D$ -dimensional vector. In this paper, we endeavor to disentangle beat condition from the input music and simultaneously incorporate explicit beat and key pose guidance to achieve beat-synchronized, key pose-guided dance generation. Given a music condition $\\mathcal{C}$ , a sparse set of keyframes $\\mathcal{X}^{ref} = \\mathcal{X} \\odot \\mathbf{M}$ , where $\\mathbf{M} \\in \\{0, 1\\}^L$ is a temporal binary mask that specifies the assigned locations of the keyframes, and a beat condition $\\mathcal{B}$ , our goal is to generate a beat-specific, and key pose-guided dance motion sequence $\\hat{\\mathcal{X}} = \\{\\hat{\\mathbf{x}}^1, \\hat{\\mathbf{x}}^2, \\hat{\\mathbf{x}}^3, \\dots, \\hat{\\mathbf{x}}^L\\}$ that conforms to the music $\\mathcal{C}$ , consistent with reference frames $\\mathcal{X}^{ref}$ at positions specified by $\\mathbf{M}$ , and synchronized with the beat condition $\\mathcal{B}$ .", + "bbox": [ + 212, + 454, + 787, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 659, + 372, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we adopt diffusion model [13, 28] as the backbone of our framework. Diffusion models establish a consistent Markovian forward process that incrementally introduces noise into clean sample data $\\mathbf{x}_0^{1:L} \\in q(\\mathbf{x}_0)$ , and a corresponding reverse process that progressively eliminates noise from noisy samples. For brevity, we use $\\mathbf{x}_t$ to represent the entire sequence. During the forward process, a predefined noise variance schedule $\\beta_t$ is employed to regulate the noise increments at each step. The forward process can be formulated as follows:", + "bbox": [ + 212, + 684, + 787, + 791 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 801, + 787, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d98af7a52b18e9bf64aa37787f267782767498df9ff1a0e7be5db58eb38f0032.jpg", + "image_caption": [ + "Fig.2: Overview of our proposed method, Beat-It. We generate a beat-synchronized dance sequence utilizing music, keyframes, and beat conditions. Conditional embeddings are derived and subsequently fused in a two-stage process: initially integrating sparse keyframe condition with other dense conditions, followed by the fusion of these dense conditions. The final fused condition is then processed by the conditional diffusion module. To ensure precise beat control, a beat alignment loss is employed to explicitly supervise the generated motions at the beat level." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 787, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After $T$ steps, the sample data progressively transforms into a noise distribution $q(\\mathbf{x}_T)$ , which is usually a standard Gaussian distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . In the reverse process, the noise is gradually removed from the noisy sample $\\mathbf{x}_T$ to produce the clean sample $\\mathbf{x}_0$ . In our task, multi-conditional dance generation aims to model the distribution $p(\\mathbf{x}_0|\\mathbf{C})$ with a set of conditions $\\mathbf{C}$ . Following Ho et al. [13], we directly predict the clean sample $\\mathbf{x}_0$ from the noise distribution $q(\\mathbf{x}_T)$ with the following objective:", + "bbox": [ + 212, + 474, + 787, + 582 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} | \\mathbf {C}), t \\sim [ 1, T ]} \\left[ \\| \\mathbf {x} _ {0} - G (\\mathbf {x} _ {t}, t, \\mathbf {C}) \\| _ {2} ^ {2} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 590, + 787, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A spatial or temporal constraint $\\mathbf{x}^c$ with positions specified by a binary mask $m$ can be directly incorporated into the denoising process without additional training. This can be achieved by simply performing the following operation at every timestep:", + "bbox": [ + 212, + 619, + 787, + 680 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = \\mathbf {x} _ {t} ^ {c} \\odot m + \\mathbf {x} _ {t} \\odot (1 - m), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 680, + 787, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the Hadamard product. By doing so, regions under constraint can be replaced with forward-diffused samples of the constraint, while the remaining regions are left unchanged. However, it is worth noting that this simplistic approach tends to render inferior results when given sparse constraints.", + "bbox": [ + 212, + 702, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Overview", + "text_level": 1, + "bbox": [ + 215, + 785, + 341, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pipeline of our proposed model is illustrated in Fig. 2. To begin with, we feed the music condition $\\mathcal{C}$ , keyframe condition $\\mathcal{X}^{ref}$ , and beat condition $\\mathcal{B}$ into", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "three encoders respectively: the music encoder $\\mathcal{E}_m$ , keyframe encoder $\\mathcal{E}_r$ , and beat encoder $\\mathcal{E}_b$ . This process yields the corresponding music embedding $\\mathbf{e}_m$ , keyframe embedding $\\mathbf{e}_r$ , and beat embedding $\\mathbf{e}_b$ , where $\\mathbf{e}_m, \\mathbf{e}_r, \\mathbf{e}_b \\in \\mathbb{R}^{L \\times d}$ . Then, we forward these condition embeddings to a hierarchical multi-condition fusion module to produce comprehensive multi-condition features $\\mathbf{c}_f$ , which are finally sent to the conditional diffusion denoising module to render the denoised dance motion sequence $\\hat{\\mathcal{X}}$ .", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Beat Representation", + "text_level": 1, + "bbox": [ + 214, + 273, + 431, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned in previous studies [23, 35, 42], there exists a strong consistency between musical beats and dance motion beats. In practical dance choreography, the arrangement of motion beats is relatively flexible and does not require a strict point-to-point alignment with musical beats. Therefore, we introduce an independent beat condition disentangled from the music, to allow for flexible motion beat controllability in dance generation. For an input music sequence of duration $L$ , each frame can be categorized as either a beat or non-beat frame. A straightforward way to represent the beat condition is using a binary mask, where 1 and 0 denote the beat and non-beat frames respectively. However, such a simple binary presentation exhibits high sparsity as the beat frames only account for a small portion of the whole music sequence. This makes it less informative and, at times, perceived as noise or disregarded by the model. As an alternative, we propose to represent the beat condition as a vector $b$ , with each entry $b^i \\in \\mathbb{N}$ denoting the distance between the current frame and the nearest beat frame. This representation not only mitigates the sparsity caused by binary formulations but also provides the model with local temporal context. This significantly facilitates the precise acquisition of rhythmic characteristics within the choreography and injects more effective beat controllability into the learning process. The beat embedding $\\mathbf{e}_b \\in \\mathbb{R}^{L \\times d}$ is obtained using a separate embedding layer, followed by the transformer-based encoder $\\mathcal{E}_b$ . Please refer to the supplementary materials for detailed representations of music and dance.", + "bbox": [ + 212, + 296, + 787, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Hierarchical Multi-Condition Fusion", + "text_level": 1, + "bbox": [ + 214, + 633, + 562, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although the keyframe constraint offers precise pose specifications for specific frames, it is still too sparse which significantly increases the learning difficulty of the model. Naively incorporating this condition with others introduces a vast of padding, leading to excessive noises that may impair the effectiveness of keyframe constraints. By contrast, music condition and beat condition are much more denser, with valid values for each frame. To harmoniously combine these three conditions, we introduce a novel multi-condition hierarchical fusion mechanism.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, in the first stage, we inject sparse keyframe condition into other dense conditions through a beat-aware dilation attention strategy, which shares a similar spirit of DiffKFC [46]. We observe that dance movements on motion-beat frames are often more iconic in the whole sequence, distinguished by their higher emphasis and significance. As these movements typically serve as guiding", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc014b8417fc5fb840cdaf4d584fa6004918cb17f28e891618e36e834d2824a9.jpg", + "image_caption": [ + "Fig. 3: Illustration of the beat-aware mask dilation scheme. The first row visualizes the keyframe mask, with deep green indicating valid control constraints and gray indicating invalid ones. The second row presents the dilation step curve with red lines marking beat frames. The third row is a heatmap of the dilation step. The fourth row shows the neighborhood range of keyframes, with light green indicating the expanded valid region from beat-aware mask dilation. The final row displays the beat-aware dilated keyframe mask." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "elements, influencing the choreography and style of the entire dance, they exert a more substantial influence on the surrounding frames. To utilize this property, we design a beat-aware mask dilation scheme. Unlike DiffKFC which uses identical dilation steps for all keyframes, treating them uniformly, we opt for different dilation weights based on the distance between the keyframe and the nearest designated beat frame. The dilation step $n$ is calculated as $n = \\lceil s \\cdot e^{-2\\frac{b^i}{d^i}} \\rceil$ , where $b^i$ and $d^i$ are the ground truth beat distance at frame $i$ and the distance between frame $i$ 's adjacent motion-beat frames, respectively. Base dilation step $s$ is a hyper-parameter controlling overall dilation speed. The closer the keyframe is to the nearest beat frame, the larger the dilation step. This strategy enables the model to gradually propagate the keyframe condition to the surrounding frames with beat awareness. The beat-aware mask dilation scheme is illustrated in Fig. 3. Given the keyframe binary mask $\\mathbf{M}$ and dilation step $n$ , the beat-aware dilated mask $\\mathbf{M}^d$ is obtained as follows:", + "bbox": [ + 212, + 411, + 789, + 627 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {d} [ i ] = \\left\\{ \\begin{array}{l l} \\max _ {j} \\mathbf {M} [ i - j ], j \\in \\{- n, n + 1, \\dots , n - 1, n \\} & \\text {i f} \\mathbf {M} _ {i} = 1, \\\\ \\mathbf {M} _ {i}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 642, + 787, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The beat-aware dilated attention mask is then calculated as $\\mathbf{M}_{attn} = \\mathbf{MM}_d^{\\mathsf{T}}$ . This beat-aware mask dilation scheme enables beat-aware keyframe condition propagation, thereby amplifying the synergy among conditions characterized by varying degrees of sparsity, thus significantly boosting the generation quality.", + "bbox": [ + 212, + 700, + 787, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the second phase, we concatenate two refined dense conditions, namely keyframe-fused music embedding and keyframe-fused beat embedding. These concatenated conditions are fed into a transformer condition fusion module, yielding the final fused features that can be subsequently sent to the conditional diffusion denoising block.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 Beat Alignment Loss", + "text_level": 1, + "bbox": [ + 215, + 146, + 434, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For controllable dance generation, it is crucial to ensure precise beat alignment between the generated dance motions and the given beat conditions. Current methods typically ignore the explicit constraints on the beats of the generated dance sequences. As a consequence, these approaches lack necessary beat controllability and struggle to guarantee beat synchronization between the generated dance motions and musical beats, which severely deteriorates the overall generation quality. To conquer this issue, building upon our beat representation, we present a beat alignment loss to provide explicit supervision that enforces the beats of the generated motions to precisely align with the given beat conditions. To this end, we first pre-train a beat distance estimator, which accepts a motion sequence as input and predicts the distance from the current frame to the nearest motion beat frame for each frame in the sequence. During the training process, the pre-trained beat distance estimator is responsible for providing supervision signals on the beats of the generated motions. The beat alignment loss $\\mathcal{L}_{beat}$ can be formulated as follows:", + "bbox": [ + 212, + 172, + 787, + 398 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b e a t}} = \\sum_ {i = 1} ^ {L} w _ {s} ^ {i} \\cdot w _ {b} ^ {i} \\cdot \\mathbf {M S E} \\left(b ^ {i}, \\hat {b} ^ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 407, + 785, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $w_{s}^{i} = \\frac{1}{1 + e^{a\\cdot(c - \\|b^{i} - \\hat{b}^{i}\\| / b^{i})}}$ , and $w_{b}^{i} = e^{-\\frac{2b^{i}}{d^{i}}}$ . Here, $\\hat{b}^i$ represents the predicted beat distance at frame $i$ . $w_{s}^{i}$ denotes an adaptive weight, which is derived from [26]. It aims to accentuate supervision in regions where beat alignment is imprecise. This is achieved by suppressing the loss magnitude in well-aligned beat regions. This strategy enables the model to focus more attention on challenging instances. The hyper-parameters $a$ and $c$ are used for controlling the degree of penalization and the threshold for beat alignment accuracy, respectively. $\\mathbf{MSE}(\\cdot ,\\cdot)$ denotes the mean squared error. $w_{b}^{i}$ is an adaptive weight that enhances the supervision of frames closely aligned with the beats while diminishing the supervision of frames distantly related to the beats. Benefiting from the beat alignment loss, our model can yield dance motions with superior beat alignment and controllability, largely elevating the quality of the generated results.", + "bbox": [ + 212, + 463, + 787, + 652 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.7 Other Losses", + "text_level": 1, + "bbox": [ + 215, + 674, + 370, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Apart from the proposed beat alignment loss $\\mathcal{L}_{\\text {beat }}$ , we also employ the basic diffusion loss $\\mathcal{L}_{\\text {simple }}$ (Eq. 2) and several additional auxiliary losses to govern the training.", + "bbox": [ + 212, + 700, + 785, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kinematic Loss. We adopt the same auxiliary losses as [41, 42]. The joint positions loss $\\mathcal{L}_{joint}$ (Eq. 6) and velocities loss $\\mathcal{L}_{vel}$ (Eq. 7) are used to improve the overall physical plausibility of the generated dance motions.", + "bbox": [ + 212, + 744, + 785, + 792 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| F K \\left(\\mathbf {x} _ {0} ^ {i}\\right) - F K \\left(\\hat {\\mathbf {x}} _ {0} ^ {i}\\right) \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 804, + 785, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v e l} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| \\mathbf {x} _ {0} ^ {i ^ {\\prime}} - \\hat {\\mathbf {x}} _ {0} ^ {i ^ {\\prime}} \\| _ {2} ^ {2} + \\| F K (\\mathbf {x} _ {0} ^ {i}) ^ {\\prime} - F K (\\hat {\\mathbf {x}} _ {0} ^ {i}) ^ {\\prime} \\| _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 157, + 787, + 198 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The contact consistency loss $\\mathcal{L}_{\\text {contact }}$ (Eq. 8) proposed in EDGE [42] is employed to alleviating the foot-sliding.", + "bbox": [ + 214, + 204, + 785, + 236 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t a c t}} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| F K _ {\\text {f o o t}} \\left(\\hat {\\mathbf {x}} _ {0} ^ {i}\\right) ^ {\\prime} \\cdot \\hat {g} ^ {i} \\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 244, + 787, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $i$ denotes the index of the frame, and $FK(\\cdot)$ is the forward kinematics function that converts the generated 6-DOF rotation representation into 3D key points in Cartesian space. $\\hat{g}^i$ denotes the predicted binary foot contact label. In addition, we employ acceleration loss $\\mathcal{L}_{acc}$ to further improve the quality of generated dance motions, which is suggested in previous work [35], preventing jitters in the generated dance motions.", + "bbox": [ + 214, + 297, + 787, + 388 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a c c} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| \\mathbf {x} _ {0} ^ {i ^ {\\prime \\prime}} - \\hat {\\mathbf {x}} _ {0} ^ {i ^ {\\prime \\prime}} \\| _ {2} ^ {2} + \\| F K (\\mathbf {x} _ {0} ^ {i}) ^ {\\prime \\prime} - F K (\\hat {\\mathbf {x}} _ {0} ^ {i}) ^ {\\prime \\prime} \\| _ {2} ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 398, + 787, + 440 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The overall kinematic loss is formulated as follows:", + "bbox": [ + 215, + 450, + 584, + 463 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {k i n}} = \\lambda_ {\\text {j o i n t}} \\mathcal {L} _ {\\text {j o i n t}} + \\lambda_ {\\text {v e l}} \\mathcal {L} _ {\\text {v e l}} + \\lambda_ {\\text {c o n t a c t}} \\mathcal {L} _ {\\text {c o n t a c t}} + \\lambda_ {\\text {a c c}} \\mathcal {L} _ {\\text {a c c}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 477, + 787, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda_{joint}$ , $\\lambda_{vel}$ , $\\lambda_{contact}$ , and $\\lambda_{acc}$ are hyper-parameters controlling the weights of the corresponding losses. We set $\\lambda_{joint} = 1$ , $\\lambda_{vel} = 2.5$ , $\\lambda_{contact} = 10$ , and $\\lambda_{acc} = 0.1$ empirically. The overall loss function is formulated as follows:", + "bbox": [ + 214, + 503, + 785, + 550 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {\\text {k i n}} \\mathcal {L} _ {\\text {k i n}} + \\lambda_ {\\text {b e a t}} \\mathcal {L} _ {\\text {b e a t}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 561, + 787, + 579 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda_{kin}$ and $\\lambda_{beat}$ are hyper-parameters controlling the weights of the corresponding losses. $\\lambda_{kin}$ and $\\lambda_{beat}$ are set as 1 and 0.5, respectively.", + "bbox": [ + 214, + 589, + 787, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 642, + 375, + 659 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. Following previous works [35, 42], we choose the most widely used dataset, AIST++ [23] for both training and testing. This dataset comprises 1,408 dance sequences encompassing 10 distinct street dance genres. Each dance genre incorporates 6 musical pieces with different BPMs. $85\\%$ of these sequences are categorized as basic choreographies, where performers execute the same choreography across all 6 musical pieces, each featuring varying tempos to synchronize with distinct musical beats. With its high choreographic and beat diversity, this dataset is ideal for our controllable dance generation task.", + "bbox": [ + 214, + 672, + 787, + 794 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. For the hierarchical multi-condition fusion module, we implement six sparse-dense fusion blocks and two dense-dense fusion blocks, utilizing a base dilation step $s$ set to 4, 8, 12, 16, 20, 24. For diffusion", + "bbox": [ + 214, + 795, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "process, we adopt the cosine schedule as beat schedule and set diffusion steps $t$ as 1000 diffusion following [41]. The classifier-free guidance scale is set to 2. The diffusion module aligns with a transformer-based architecture similar to [42]. We adopt Adam optimizer with the learning rate set as 2e-4 and the training batch size is set to 64. During training, we randomly sample between $1\\%$ and $30\\%$ of the ground truth (GT) motion frames as the keyframe conditions. This training strategy can effectively improve the generalization capability of our method on different numbers of keyframes. We use the GT motion beats as beat conditions.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To ensure fair comparisons, we follow EDGE [42] to crop all data into 5-second clips at 30 frames per second, with 2.5 seconds of overlapping. During testing, we randomly sample $10\\%$ of keyframes from unpaired samples in the testing set as keyframe conditions and subsets of testing musical beats as beat conditions to prevent data leakage. The musical beats used for experiments are extracted by the off-the-shelf audio toolkit Librosa [17]. All experiments are conducted on 4 NVIDIA RTX3090 GPUs.", + "bbox": [ + 212, + 267, + 787, + 371 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. For quantitative evaluation, we measure the generated dance from three aspects: generation quality, diversity, and controllability. Due to the limited samples in the testing set of AIST++ [23], the prior work EDGE [42] has demonstrated that the Fréchet Inception Distance (FID) is not reliable for measuring the generation quality. Therefore, we do not adopt FID as the evaluation metric. Instead, we assess generation quality from two different perspectives: music-dance correlation and kinematic plausibility. In terms of music-dance correlation, we utilize the Beat Alignment Score (BAS) metric, following the approach presented in [35]. BAS quantifies the synchronization quality between the generated dance and the musical beat. For the evaluation of kinematic plausibility, we adopt the Physical Foot Contact score (PFC) metric proposed in [42]. PFC provides a metric for measuring the physical plausibility of the generated dance from a kinematic perspective. As for the diversity, we measure the average feature distance of kinetic $(\\mathrm{Div}_k)$ and geometric $(\\mathrm{Div}_g)$ features extracted by fairmotion [12] following the previous works [23, 35, 42].", + "bbox": [ + 212, + 373, + 787, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For assessing controllability, we employ the Key Pose Distance (KPD) and Beat Assignment Precision (BAP) metrics to evaluate keyframe control and beat control, respectively. KPD quantifies the average mean square error of local joint positions in Cartesian space at the keyframes. To evaluate beat control, we measure the alignment between the generated dance and the designated beat condition. BAP is defined as the percentage of beats in the generated dance that are correctly assigned to the designated beat frames. Note that this metric differs from the BAS, as a dance sequence with a high BAS does not necessarily adhere to the specific motion beat choreography constraint we intend to impose.", + "bbox": [ + 212, + 599, + 787, + 736 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Comparison to Existing Methods", + "text_level": 1, + "bbox": [ + 214, + 756, + 535, + 771 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Currently, only three music-to-dance methods are publicly accessible: EDGE [42], Bailando [35], and FACT [23]. Of these, EDGE [42] is most directly aligned with our approach, as it similarly accommodates keyframes as inputs for generating dance sequences. We therefore assess EDGE using the same keyframe conditions", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2f5a94ebd3313a0ad8e8943299325b761a69daa1770a323d12dc1219c5aac1f3.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons among different methods on AIST++. Bold indicates best result. ↓ means lower is better, ↑ means higher is better and → means closer to the ground truth is better." + ], + "table_footnote": [], + "table_body": "
MethodsQualityDiversityControllability
PFC ↓ BAS ↑Divk→Divm→KPD ↓ BAP ↑
Ground Truth1.3380.3849.7737.212--
FACT [23]2.6980.2029.7047.342--
Bailando [35]1.5780.2159.6227.175--
EDGE [42](keyframes)1.0840.2359.7437.2740.859-
Ours(beat & keyframes)0.9660.6619.6607.2480.3060.793
", + "bbox": [ + 259, + 198, + 741, + 333 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "as our method to ensure a fair comparison. Conversely, certain approaches, like DanceFormer [21], are not open-source, thus hindering any direct evaluative comparison. We use the original test split of AIST++ [23] for evaluation, which has no overlap with the training split in terms of both the music and dance choreography.", + "bbox": [ + 212, + 359, + 789, + 421 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/48f4518a6c760932164bf2f4041f3393e367b7ec615012bb8004c34e4e7e76a1.jpg", + "table_caption": [ + "Table 2: Results of user study." + ], + "table_footnote": [], + "table_body": "
User StudyBeat-It Win Rate
FACT92.2%
Bailando78.8%
EDGE \n(keyframes)QualityControllability \n(Keyframe)
60.3%86.9%
", + "bbox": [ + 228, + 458, + 470, + 569 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6a43d171843f2ee9b4dbf5734ffb28dcfc4d992f0ad8c07056a3864a06eb7187.jpg", + "table_caption": [ + "Table 3: Ablation study on AIST++." + ], + "table_footnote": [], + "table_body": "
AblationsQualityControllability
PFC ↓BAS ↑KPD ↓BAP ↑
w/o HF25.6260.3220.4770.323
w/o BD1.6320.3580.3890.371
w/o Lbeat1.3420.3970.3430.411
Ours0.9660.6610.3060.793
", + "bbox": [ + 511, + 458, + 766, + 555 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Comparisons. The quantitative results are shown in Tab. 1. As is revealed, our method surpasses all the competitors in terms of generation quality, diversity, and controllability. Particularly, our method shows considerable advantages over the compared methods for BAS metrics, notably with an improvement of 0.426 compared with the SOTA method EDGE [42]. This demonstrates the effectiveness of our method in generating beat-synchronized dance sequences while maintaining diversity. Furthermore, in the aspect of controllability, our method also significantly excels EDGE [42] with a notable improvement of 0.553 in KPD. Fig. 4 shows the mean joint velocities over time for a specific dance motion concerning both the GT and the existing methods. Notably, compared with others, our approach demonstrates more motion beats (local minimum point of velocities) precisely at the beat frames. This contributes to a better visual alignment with the audio when compared to the other methods. The BAP further validates our method's superior ability to produce dance sequences precisely aligned with the beat condition. Due to limited space, we present additional quantitative experiments in the supplementary materials.", + "bbox": [ + 212, + 569, + 787, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "User Study. We also conduct a user study to assess the generation quality and controllability of our method. We randomly select 20 generated dance sequences", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4260c2e47c89f273f0e0a170133747059998e2458f470ae96ed7d18041a0ef2f.jpg", + "image_caption": [ + "Fig. 4: Visualization comparison on beat alignment among different methods. The motion generated by our method shows precise beat alignment with the given beat condition, demonstrating the superiority of our method in beat control." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 782, + 248 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "from the testing set and ask 18 participants to pick the best one among our method and the other compared methods. In particular, participants are required to choose the superior dance according to several criteria, namely, the overall visual plausibility of movements, synchronization with the dance beats, and keyframe controllability. The results of the user study are tabulated in Tab. 2. Our method shows noticeable advantages in terms of generation quality compared to other approaches, particularly in beat alignment. For qualitative evaluations on keyframe-conditioned controllability, we also provide the participants with the keyframe references for intuitive comparison between our method against EDGE [42]. The results show that our approach owns better motion controllability than EDGE [42].", + "bbox": [ + 212, + 328, + 787, + 494 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 513, + 401, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To get a more comprehensive insight into our main contributions, we conduct ablation studies on the hierarchical multi-condition fusion module, the beat-aware dilation scheme, and the beat alignment loss to evaluate their efficacy. The quantitative results are shown in Tab. 3. The qualitative results can be also found in the supplementary video.", + "bbox": [ + 212, + 537, + 787, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hierarchical Multi-condition Fusion Module. To verify the significance of the hierarchical multi-condition fusion module, we remove it by simply concatenating all the conditions and feeding them into a one-stage encoder (w/o HF). As shown in Tab. 3, the simple one-stage fusion module can lead to inferior results in both quality and controllability. This is because the direct concatenation of multiple conditions can lead to inevitable condition conflicts, giving rise to suboptimal optimization of the model. Even worse, the naive padding of the sparse keyframe condition introduces severe noises, further complicating the learning process and resulting in unsatisfactory generation performance. By decently fusing multiple conditions gradually, our hierarchical multi-condition fusion mechanism effectively mitigates the conflicts among different conditions and yields more informative conditions, allowing for higher-quality generation with better controllability.", + "bbox": [ + 212, + 613, + 787, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Beat-aware Dilation Scheme. To validate the effectiveness of our beat-aware dilation scheme, we replace it with a vanilla masked attention strategy (w/o", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation", + "bbox": [ + 289, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "BD). As shown in Tab. 3, the absence of the beat-aware dilation scheme notably impairs both keyframe and beat choreography controllability. The reason behind this is that beat-aware dilation can dynamically adjust the expansion degree of the attention mask, contingent on the distance between the keyframe and the beat frames. This tactical approach significantly improves the synchronization between keyframes and their associated beat frames, thereby empowering the model to more effectively harness prior knowledge of beats. Consequently, this enhancement considerably bolsters beat control while augmenting keyframe controllability.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Beat Alignment Loss. To further validate the effectiveness of the beat alignment loss, we assess the model's performance by comparing it to its counterpart without it (w/o $\\mathcal{L}_{\\text {beat }}$ ). As shown in Tab. 3, the ablated variant without the beat alignment loss demonstrated performance degradation in beat choreography, showcasing decreases of 0.264 and 0.382 in BAS and BAP, respectively. This decline in performance can be attributed to the advantageous impact of direct supervision during training, particularly in guiding the model to comprehend beat-level dance choreography through the proposed beat alignment loss. Consequently, this enhancement significantly improves the synchronization between motion sequences and designated beats.", + "bbox": [ + 212, + 267, + 789, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 215, + 439, + 509, + 457 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we propose a novel multi-condition diffusion-based framework, BeatIt, for beat-synchronized and key pose-guided dance generation. The proposed framework presents a hierarchical multi-condition fusion mechanism equipped with a beat-aware mask dilation scheme to integrate conditions with different information sparsity. To achieve precise beat synchronization, we explicitly disentangle the beats from the music and inject beat controllability throughout the entire generation process. Additionally, we introduce a specifically designed beat alignment loss to provide explicit guidance and supervision on motion beats. Both qualitative and quantitative experimental results on the AIST++ dataset validate the superiority of our method in producing high-quality dance sequences with precise beat synchronization and flexible keyframe control. Further limitations and broader impact are discussed in supplementary materials.", + "bbox": [ + 212, + 470, + 789, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 672, + 406, + 691 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The work is supported by China National Key R&D Program (No. 2023YFE0202700), Key-Area Research and Development Program of Guangzhou City (No. 2023B01J0022), Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (No. 2024B1515040010), Health and Medical Research Fund (HMRF) of Hong Kong Health Bureau (No. 10211516), Innovation and Technology Fund - Innovation and Technology Support Programme (ITF-ITSP) (No. ITS/127/22), Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2023B1515020097), National Research Foundation Singapore under the AI Singapore Programme (No. AISG3-GV-2023-011).", + "bbox": [ + 212, + 703, + 789, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alemi, O., Françoise, J., Pasquier, P.: Groovenet: Real-time music-driven dance movement generation using artificial neural networks. networks 8(17), 26 (2017) 2, 4", + "2. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM TOG 42(4), 1-20 (2023) 4", + "3. Ao, T., Zhang, Z., Liu, L.: Gesturediffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613 (2023) 4", + "4. Araújo, J.P., Li, J., Vetrivel, K., Agarwal, R., Wu, J., Gopinath, D., Clegg, A.W., Liu, K.: Circle: Capture in rich contextual environments. In: CVPR. pp. 21211-21221 (2023) 3", + "5. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR. pp. 9760-9770 (2023) 3, 4", + "6. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. NeurIPS 34, 19822-19835 (2021) 4", + "7. Ferreira, J.P., Coutinho, T.M., Gomes, T.L., Neto, J.F., Azevedo, R., Martins, R., Nascimento, E.R.: Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics 94, 11-21 (2021) 4", + "8. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV. pp. 89-106. Springer (2022) 4", + "9. Gao, X., Hu, L., Zhang, P., Zhang, B., Bo, L.: Dancemeld: Unraveling dance phrases with hierarchical latent codes for music-to-dance synthesis. arXiv preprint arXiv: 2401.10242 (2023) 4", + "10. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: Imos: Intent-driven full-body motion synthesis for human-object interactions. In: CGF. vol. 42, pp. 1-12. Wiley Online Library (2023) 3", + "1. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: ICCV. pp. 9942-9952 (2023) 4", + "2. Gopinath, D., Won, J.: fairmotion - tools to load, process and visualize motion capture data. Github (2020), https://github.com/facebookresearch/fairmotion11", + "3. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020) 5, 6", + "4. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. ACM TOG 35(4), 1-11 (2016) 4", + "5. Huang, R., Hu, H., Wu, W., Sawada, K., Zhang, M., Jiang, D.: Dance revolution: Long-term dance generation with music via curriculum learning. arXiv preprint arXiv:2006.06119 (2020) 2, 4", + "6. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR. pp. 16750-16761 (2023) 3, 4", + "7. Jin, Y., Zhang, J., Li, M., Tian, Y., Zhu, H., Fang, Z.: Towards the automatic anime characters creation with generative adversarial networks. arXiv preprint arXiv:1708.05509 (2017) 11" + ], + "bbox": [ + 225, + 176, + 785, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI. vol. 37, pp. 8255-8263 (2023) 3, 4", + "19. Kim, J., Oh, H., Kim, S., Tong, H., Lee, S.: A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In: CVPR. pp. 3490-3500 (2022) 4", + "20. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. NeurIPS 32 (2019) 4", + "21. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI. vol. 36, pp. 1272-1279 (2022) 2, 4, 12", + "22. Li, J., Yin, Y., Chu, H., Zhou, Y., Wang, T., Fidler, S., Li, H.: Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171 (2020) 2, 4", + "23. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV. pp. 13401-13412 (2021) 2, 4, 7, 10, 11, 12", + "24. Lin, J., Chang, J., Liu, L., Li, G., Lin, L., Tian, Q., Chen, C.w.: Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In: CVPR. pp. 23222-23231 (2023) 3", + "25. Liu, H., Xu, C., Yang, Y., Zeng, L., He, S.: Drag your noise: Interactive point-based editing via diffusion semantic propagation. In: CVPR. pp. 6743-6752 (2024) 4", + "26. Lu, X., Ma, C., Shen, J., Yang, X., Reid, I., Yang, M.H.: Deep object tracking with shrinkage loss. IEEE TPAMI 44(5), 2386-2401 (2020) 9", + "27. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) 5", + "28. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) 4, 5", + "29. Qi, Q., Zhuo, L., Zhang, A., Liao, Y., Fang, F., Liu, S., Yan, S.: Diffdance: Cascaded human motion diffusion model for dance generation. In: ACM MM. pp. 1374-1382 (2023) 4", + "30. Qin, C., Zhang, S., Yu, N., Feng, Y., Yang, X., Zhou, Y., Wang, H., Niebles, J.C., Xiong, C., Savarese, S., et al.: Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147 (2023) 5", + "31. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML. pp. 8821-8831. PMLR (2021) 4", + "32. Ren, X., Li, H., Huang, Z., Chen, Q.: Self-supervised dance video synthesis conditioned on music. In: ACM MM. pp. 46-54 (2020) 4", + "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) 4", + "34. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS 35, 36479-36494 (2022) 4", + "35. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR. pp. 11050-11059 (2022) 2, 4, 7, 10, 11, 12", + "36. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando++: 3d dance gpt with choreographic memory. IEEE TPAMI (2023) 4" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Sun, G., Wong, Y., Cheng, Z., Kankanhalli, M.S., Geng, W., Li, X.: Deepdance: music-to-dance motion choreography with adversarial learning. IEEE TMM 23, 497-509 (2020) 4", + "38. Sun, J., Wang, C., Hu, H., Lai, H., Jin, Z., Hu, J.F.: You never stop dancing: Non-freezing dance generation via bank-constrained manifold projection. NeurIPS 35, 9995-10007 (2022) 4", + "39. Tang, T., Jia, J., Mao, H.: Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In: ACM MM. pp. 1598-1606 (2018) 2, 4", + "40. Tang, Z., Yang, Z., Zhu, C., Zeng, M., Bansal, M.: Any-to-any generation via composable diffusion. arXiv preprint arXiv: 2305.11846 (2023) 5", + "41. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) 9, 11", + "42. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: CVPR. pp. 448-458 (2023) 2, 4, 7, 9, 10, 11, 12, 13", + "43. Valle-Pérez, G., Henter, G.E., Beskow, J., Holzapfel, A., Oudeyer, P.Y., Alexanderson, S.: Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM TOG 40(6), 1-14 (2021) 2, 4", + "44. Wang, T., Zhang, T., Zhang, B., Ouyang, H., Chen, D., Chen, Q., Wen, F.: Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952 (2022) 4", + "45. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS 35, 14959-14971 (2022) 3", + "46. Wei, D., Sun, X., Sun, H., Li, B., Hu, S., Li, W., Lu, J.: Understanding text-driven motion synthesis with keyframe collaboration via diffusion models. arXiv preprint arXiv:2305.13773 (2023) 7", + "47. Xie, Y., Lin, Y., Cai, W., Xu, X., Zhang, H., Du, Y., He, S.: D3still: Decoupled differential distillation for asymmetric image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17181-17190 (2024) 4", + "48. Xu, C., Chen, Z., Mai, J., Xu, X., He, S.: Pose-and attribute-consistent person image synthesis. ACM TOMM 19(2s), 1-21 (2023) 4", + "49. Xu, C., Li, K., Luo, X., Xu, X., He, S., Zhang, K.: Fully deformable network for multiview face image synthesis. IEEE TNNLS (1-15) (2022) 4", + "50. Xu, C., Xu, Y., Zhang, H., Xu, X., He, S.: Dreamanime: Learning style-identity textual disentanglement for anime and beyond. IEEE TVCG (2024) 4", + "51. Yalta, N., Watanabe, S., Nakadai, K., Ogata, T.: Weakly-supervised deep recurrent neural networks for basic dance step generation. In: IJCNN. pp. 1-8. IEEE (2019) 4", + "52. Yang, Z., Wen, Y.H., Chen, S.Y., Liu, X., Gao, Y., Liu, Y.J., Gao, L., Fu, H.: Keyframe control of music-driven 3d dance generation. IEEE TVCG (2023) 2, 4", + "53. Yin, W., Yin, H., Baraka, K., Kragic, D., Björkman, M.: Dance style transfer with cross-modal transformer. In: WACV. pp. 5058-5067 (2023) 4", + "54. Yu, Y., Liu, B., Zheng, C., Xu, X., Zhang, H., He, S.: Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In: CVPR. pp. 7109-7118 (2024) 4", + "55. Zhang, C., Tang, Y., Zhang, N., Lin, R.S., Han, M., Xiao, J., Wang, S.: Bidirectional autoregressive diffusion model for dance generation. arXiv preprint arXiv: 2402.04356 (2024) 4" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation", + "bbox": [ + 287, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023) 3", + "57. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023) 5", + "58. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023) 5", + "59. Zheng, C., Liu, B., Xu, X., Zhang, H., He, S.: Learning an interpretable stylized subspace for 3d-aware animatable artforms. IEEE TVCG (2024) 4", + "60. Zheng, C., Liu, B., Zhang, H., Xu, X., He, S.: Where is my spot? few-shot image generation via latent subspace optimization. In: CVPR. pp. 3272-3281 (2023) 4", + "61. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: CVPR. pp. 5632-5641 (2023) 3, 4", + "62. Zhu, L., Liu, X., Liu, X., Qian, R., Liu, Z., Yu, L.: Taming diffusion models for audio-driven co-speech gesture generation. In: CVPR. pp. 10544-10553 (2023) 4" + ], + "bbox": [ + 215, + 146, + 787, + 369 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_model.json b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_model.json new file mode 100644 index 0000000000000000000000000000000000000000..339268cbb9bef81cc74ba68c355c113c49574c19 --- /dev/null +++ b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_model.json @@ -0,0 +1,2359 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.776, + 0.185 + ], + "angle": 0, + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.212, + 0.78, + 0.245 + ], + "angle": 0, + "content": "Zikai Huang\\(^{1}\\), Xuemiao \\(\\mathrm{Xu}^{1,3,4(\\boxtimes)}\\), Cheng \\(\\mathrm{Xu}^{2(\\boxtimes)}\\), Huaidong Zhang\\(^{1,3}\\), Chenxi Zheng\\(^{1}\\), Jing Qin\\(^{2}\\), and Shengfeng He\\(^{5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.255, + 0.66, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) South China University of Technology, China xuemx@scut.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.283, + 0.723, + 0.31 + ], + "angle": 0, + "content": "2 The Hong Kong Polytechnic University, Hong Kong SAR, China \ncschengxu@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.31, + 0.748, + 0.325 + ], + "angle": 0, + "content": "Guangdong Engineering Center for Large Model and GenAI Technology" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.325, + 0.769, + 0.351 + ], + "angle": 0, + "content": "4 Guangdong Provincial Key Lab of Computational Intelligence and Cyberspace Information" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.352, + 0.661, + 0.38 + ], + "angle": 0, + "content": "5 Singapore Management University, Singapore https://zikaihuangscut.github.io/Beat-It/" + }, + { + "type": "list", + "bbox": [ + 0.233, + 0.325, + 0.769, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.419, + 0.789, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.564, + 0.79, + 0.622 + ], + "angle": 0, + "content": "Fig. 1: We introduce Beat-It, a novel method for generating 3D dance motions with beat alignment and motion controllability. Our approach explicitly injects beat awareness and seamlessly integrates multiple conditions to guide the generation process, leading to beat-synchronized, key pose-guided dance generation." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.674, + 0.744, + 0.842 + ], + "angle": 0, + "content": "Abstract. Dance, as an art form, fundamentally hinges on the precise synchronization with musical beats. However, achieving aesthetically pleasing dance sequences from music is challenging, with existing methods often falling short in controllability and beat alignment. To address these shortcomings, this paper introduces Beat-It, a novel framework for beat-specific, key pose-guided dance generation. Unlike prior approaches, Beat-It uniquely integrates explicit beat awareness and key pose guidance, effectively resolving two main issues: the misalignment of generated dance motions with musical beats, and the inability to map key poses to specific beats, critical for practical choreography. Our approach disentangles beat conditions from music using a nearest beat distance representation and employs a hierarchical multi-condition fusion mechanism. This mechanism" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Z. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.232 + ], + "angle": 0, + "content": "seamlessly integrates key poses, beats, and music features, mitigating condition conflicts and offering rich, multi-conditioned guidance for dance generation. Additionally, a specially designed beat alignment loss ensures the generated dance movements remain in sync with the designated beats. Extensive experiments confirm Beat-It's superiority over existing state-of-the-art methods in terms of beat alignment and motion controllability." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.244, + 0.741, + 0.273 + ], + "angle": 0, + "content": "Keywords: Dance generation \\(\\cdot\\) Beat synchronization \\(\\cdot\\) Multi-condition diffusion generation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.296, + 0.377, + 0.312 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Dance, a time-honored art form, resonates across various ages and cultures due to its artistic and aesthetic significance, playing a vital role in the cultural and entertainment sectors. Traditional dance choreography creation, which demands an intricate balance of aesthetic movements, emotional expression, and precise synchronization with musical beats, often presents itself as a costly, laborious, and time-intensive endeavor, even for seasoned artists." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.789, + 0.553 + ], + "angle": 0, + "content": "Recent advances in deep learning, particularly in generative models, have catalyzed efforts to automate the choreography process. Initial approaches predominantly focused on single-condition dance generation, learning a direct mapping from music to dance motions [1, 15, 21-23, 35, 39, 43]. The advent of diffusion models, renowned for their exceptional content generation capabilities, has marked a paradigm shift in this domain. For instance, Tseng et al. [42] introduced a transformer-based diffusion model that sets a new benchmark in music-to-dance conversion, showcasing the potential of these models in elevating the quality of generated dance sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.789, + 0.78 + ], + "angle": 0, + "content": "Notwithstanding the remarkable progress, achieving precise synchronization between dance motions and musical beats, coupled with flexible motion controllability in dance generation, remains a formidable challenge. In real-world choreography, the goal is often to assign specific key poses to particular musical beats and then fill in the connecting movements to form the complete dance sequence. This process requires not only the creation of designated key poses but also their harmonious synchronization with the musical beats. While few methods have explored key pose conditions [42, 52], none of them has considered explicit beat synchronization and controllability, rendering inferior dance motion generation. Furthermore, previous methods typically merge keyframe features with music features through simple concatenation or temporal/spatial blending. This approach, however, is flawed due to the sparse nature of keyframes. Consequently, these methods disproportionately emphasize dense music features at the expense of key pose conditions, leading to a notable misalignment between the generated dance motions and the specified keyframes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "To combat the above issues, we aim to completely disentangle beats from input music and simultaneously incorporate explicit beat and key pose guidance to enable beat-controllable, key pose-guided dance generation. We introduce BeatIt, a diffusion-based, multi-condition framework designed for this purpose (see" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Fig. 1). Our approach begins by formulating beat condition as a nearest beat distance representation, which is a straightforward yet effective method to serve as beat guidance within our framework. To effectively harness and balance the multiple conditions involved, we propose a hierarchical multi-condition fusion mechanism. Initially, we integrate sparse key pose condition with dense beat and music conditions using a tailored beat-aware dilated cross-attention strategy. We further blend refined beat and music features to generate comprehensive multi-condition features. This strategy not only enhances beat and motion controllability in dance generation but also maintains the realism and coherence of the produced dance motions. Additionally, leveraging our beat representation, we introduce a novel beat alignment loss, explicitly ensuring synchronization between dance motions and the given beat condition, thereby significantly enhancing the overall quality of generation. Extensive experiments confirm that our method outperforms current state-of-the-art approaches in terms of beat alignment and motion controllability. Beyond the advantages above, our framework also supports arbitrary beat designation and flexible frame assignments of key poses." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.39, + 0.595, + 0.404 + ], + "angle": 0, + "content": "In summary, our main contributions are fourfold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.414, + 0.786, + 0.471 + ], + "angle": 0, + "content": "- We introduce a multi-condition dance generation framework that achieves beat synchronization and enhanced motion controllability. To the best of our knowledge, our method makes the first attempt at beat-controllable key pose-guided dance generation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.474, + 0.785, + 0.515 + ], + "angle": 0, + "content": "- We present a hierarchical multi-condition fusion mechanism to effectively suppress the conflicts and fully exploit the complementary information among different conditions." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.519, + 0.785, + 0.576 + ], + "angle": 0, + "content": "- We delve into the property of beats and formulate it as a nearest beat distance representation. A beat alignment loss is further tailored to offer explicit supervision signals to the generated dance motions, largely elevating the synchronization between the generated motions and the given beat conditions" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.579, + 0.785, + 0.622 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate our method performs favorably against the state-of-the-art approaches, especially in motion controllability and motion-beat alignment." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.414, + 0.786, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.388, + 0.663 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.486, + 0.694 + ], + "angle": 0, + "content": "2.1 Human Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Human motion generation aims to generate realistic human motion automatically. Previous works can be generally categorized into three genres according to the input condition: text-conditioned [5,18,24,56,61], audio-conditioned, and scene-conditioned [4, 10, 16, 45]. Our work belongs to audio-conditioned human motion generation. But different from the primary audio-conditioned human motion generation, music-to-dance generation requires a comprehensive consideration of aesthetic movements, emotional expression, and precise synchronization with musical beats, making it extremely challenging to render convincing results. Recently, researchers have widely explored the application of diffusion models" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Z. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "in human motion generation [2,3,5,16,18,61,62], which shows unprecedented performance than traditional generative models. In this paper, we get a deep insight into the beat synchronization problem and investigate a multi-condition diffusion-based scheme for controllable dance generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.232, + 0.414, + 0.246 + ], + "angle": 0, + "content": "2.2 Dance Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.79, + 0.744 + ], + "angle": 0, + "content": "With the emergence of deep learning [47-49], dance generation has attracted lots of research interest in recent years. Previous methods have explored various frameworks with different types of backbones to achieve single-condition dance generation, including CNN [14], RNN [1, 15, 39, 51], GCN [7, 32], VAE [9, 38], GAN [19, 20, 37], and Transformer [21-23, 35, 36, 43, 53]. FACT [23] proposes a full-attention cross-modal transformer model to capture the correlations between music and dance. Danceformer [21] presents a two-stage deterministic framework for audio-driven single-condition dance generation. It first generates key poses from the input music and then performs interpolation between the key poses. However, it falls short of arbitrary beat choreography and flexible key pose controllability due to its single-condition nature. Bailando [35] utilizes separate VQ-VAEs on upper/lower half bodies and a motion GPT to map the music and seed poses to dance sequences. Several recent attempts have investigated the diffusion models for dance generation [29, 42, 55], which significantly pushes the boundaries of music-to-dance performance. Compared to applying only a single input condition, multi-condition dance generation incorporates additional conditions beyond music into the generation process for better controllability, such as dance style labels [2], text [11], and keyframes [42, 52]. For example, LDA [2] introduced dance style labels as auxiliary conditions. However, it is limited to generating pre-defined styles and lacks flexibility. TM2D [11] adopts a VQ-VAE based model to support dual-modal music and text-driven dance motion generation. Yang et al. [52] employ normalizing flows to model dance motion probability distributions and utilize time embedding to achieve keyframe control. EDGE [42] adopts a diffusion-based framework and enables key pose control via temporal/spatial blending during inference. Although the existing methods have investigated additional conditions for better dance generation control, they typically neglect explicit guidance and supervision to align the generated motion dances with specified beats. Consequently, these methods struggle to produce both beat-specific and key pose-guided dance sequences. In contrast, we propose to explicitly disentangle the beat condition from music and inject flexible beat and key pose controllability into the entire generation process, significantly enhancing the beat alignment while supporting flexible frame assignments of key poses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.768, + 0.577, + 0.782 + ], + "angle": 0, + "content": "2.3 Multi Condition Diffusion Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Multi-condition diffusion generative models seek to inject multi-condition guidance into the generation process for versatile controllability [6,8,25,28,31,34, 44,50,54,59,60]. Benefiting from large-scale training data, Stable Diffusion [33]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "constructs a powerful foundation for image generation. Building upon this, ControlNet [57] and T2I-Adapter [27] efficiently integrate trainable parameters into existing models, enabling the inclusion of spatially localized input conditions in a pre-trained text-to-image diffusion model. Uni-ControlNet [58] proposes a framework equipped with two additional control modules for multi-condition control. Qin et al. [30] present a unified model capable of handling various visual conditions, resulting in a more streamlined scheme. Codi [40] acquire multi-modal synthesis skills through the mapping of diverse modalities into a unified space, training on lists of different multi-modal generation tasks. The aforementioned works typically controllable generation using pre-trained models under diverse conditions. Our work, conversely, focuses on the effective and comprehensive fusion of multiple conditions by presenting a hierarchical multi-condition fusion mechanism aimed at better integrating sparse keyframe control conditions with dense music and beat control conditions, resulting in more controllable dance generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.397, + 0.331, + 0.413 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.43, + 0.421, + 0.444 + ], + "angle": 0, + "content": "3.1 Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.789, + 0.637 + ], + "angle": 0, + "content": "Dance generation aims to create realistic dance motion sequences \\(\\mathcal{X} = \\{\\mathbf{x}^1, \\mathbf{x}^2, \\mathbf{x}^3, \\dots, \\mathbf{x}^L\\}\\) from a given piece of music \\(\\mathcal{C}\\) with a duration of \\(L\\). Here, \\(\\mathbf{x}^i \\in \\mathbb{R}^D\\) represents a human pose at the \\(i\\)-th frame, which is denoted as a \\(D\\)-dimensional vector. In this paper, we endeavor to disentangle beat condition from the input music and simultaneously incorporate explicit beat and key pose guidance to achieve beat-synchronized, key pose-guided dance generation. Given a music condition \\(\\mathcal{C}\\), a sparse set of keyframes \\(\\mathcal{X}^{ref} = \\mathcal{X} \\odot \\mathbf{M}\\), where \\(\\mathbf{M} \\in \\{0, 1\\}^L\\) is a temporal binary mask that specifies the assigned locations of the keyframes, and a beat condition \\(\\mathcal{B}\\), our goal is to generate a beat-specific, and key pose-guided dance motion sequence \\(\\hat{\\mathcal{X}} = \\{\\hat{\\mathbf{x}}^1, \\hat{\\mathbf{x}}^2, \\hat{\\mathbf{x}}^3, \\dots, \\hat{\\mathbf{x}}^L\\}\\) that conforms to the music \\(\\mathcal{C}\\), consistent with reference frames \\(\\mathcal{X}^{ref}\\) at positions specified by \\(\\mathbf{M}\\), and synchronized with the beat condition \\(\\mathcal{B}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.66, + 0.373, + 0.673 + ], + "angle": 0, + "content": "3.2 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.789, + 0.792 + ], + "angle": 0, + "content": "In this work, we adopt diffusion model [13, 28] as the backbone of our framework. Diffusion models establish a consistent Markovian forward process that incrementally introduces noise into clean sample data \\(\\mathbf{x}_0^{1:L} \\in q(\\mathbf{x}_0)\\), and a corresponding reverse process that progressively eliminates noise from noisy samples. For brevity, we use \\(\\mathbf{x}_t\\) to represent the entire sequence. During the forward process, a predefined noise variance schedule \\(\\beta_t\\) is employed to regulate the noise increments at each step. The forward process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.263, + 0.803, + 0.788, + 0.845 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Z. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.789, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.349, + 0.79, + 0.447 + ], + "angle": 0, + "content": "Fig.2: Overview of our proposed method, Beat-It. We generate a beat-synchronized dance sequence utilizing music, keyframes, and beat conditions. Conditional embeddings are derived and subsequently fused in a two-stage process: initially integrating sparse keyframe condition with other dense conditions, followed by the fusion of these dense conditions. The final fused condition is then processed by the conditional diffusion module. To ensure precise beat control, a beat alignment loss is employed to explicitly supervise the generated motions at the beat level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.789, + 0.583 + ], + "angle": 0, + "content": "After \\(T\\) steps, the sample data progressively transforms into a noise distribution \\(q(\\mathbf{x}_T)\\), which is usually a standard Gaussian distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). In the reverse process, the noise is gradually removed from the noisy sample \\(\\mathbf{x}_T\\) to produce the clean sample \\(\\mathbf{x}_0\\). In our task, multi-conditional dance generation aims to model the distribution \\(p(\\mathbf{x}_0|\\mathbf{C})\\) with a set of conditions \\(\\mathbf{C}\\). Following Ho et al. [13], we directly predict the clean sample \\(\\mathbf{x}_0\\) from the noise distribution \\(q(\\mathbf{x}_T)\\) with the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.592, + 0.788, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} | \\mathbf {C}), t \\sim [ 1, T ]} \\left[ \\| \\mathbf {x} _ {0} - G (\\mathbf {x} _ {t}, t, \\mathbf {C}) \\| _ {2} ^ {2} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.681 + ], + "angle": 0, + "content": "A spatial or temporal constraint \\(\\mathbf{x}^c\\) with positions specified by a binary mask \\(m\\) can be directly incorporated into the denoising process without additional training. This can be achieved by simply performing the following operation at every timestep:" + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.681, + 0.788, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = \\mathbf {x} _ {t} ^ {c} \\odot m + \\mathbf {x} _ {t} \\odot (1 - m), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.765 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the Hadamard product. By doing so, regions under constraint can be replaced with forward-diffused samples of the constraint, while the remaining regions are left unchanged. However, it is worth noting that this simplistic approach tends to render inferior results when given sparse constraints." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.342, + 0.8 + ], + "angle": 0, + "content": "3.3 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "The pipeline of our proposed model is illustrated in Fig. 2. To begin with, we feed the music condition \\(\\mathcal{C}\\), keyframe condition \\(\\mathcal{X}^{ref}\\), and beat condition \\(\\mathcal{B}\\) into" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "three encoders respectively: the music encoder \\(\\mathcal{E}_m\\), keyframe encoder \\(\\mathcal{E}_r\\), and beat encoder \\(\\mathcal{E}_b\\). This process yields the corresponding music embedding \\(\\mathbf{e}_m\\), keyframe embedding \\(\\mathbf{e}_r\\), and beat embedding \\(\\mathbf{e}_b\\), where \\(\\mathbf{e}_m, \\mathbf{e}_r, \\mathbf{e}_b \\in \\mathbb{R}^{L \\times d}\\). Then, we forward these condition embeddings to a hierarchical multi-condition fusion module to produce comprehensive multi-condition features \\(\\mathbf{c}_f\\), which are finally sent to the conditional diffusion denoising module to render the denoised dance motion sequence \\(\\hat{\\mathcal{X}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.274, + 0.432, + 0.289 + ], + "angle": 0, + "content": "3.4 Beat Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.789, + 0.614 + ], + "angle": 0, + "content": "As mentioned in previous studies [23, 35, 42], there exists a strong consistency between musical beats and dance motion beats. In practical dance choreography, the arrangement of motion beats is relatively flexible and does not require a strict point-to-point alignment with musical beats. Therefore, we introduce an independent beat condition disentangled from the music, to allow for flexible motion beat controllability in dance generation. For an input music sequence of duration \\( L \\), each frame can be categorized as either a beat or non-beat frame. A straightforward way to represent the beat condition is using a binary mask, where 1 and 0 denote the beat and non-beat frames respectively. However, such a simple binary presentation exhibits high sparsity as the beat frames only account for a small portion of the whole music sequence. This makes it less informative and, at times, perceived as noise or disregarded by the model. As an alternative, we propose to represent the beat condition as a vector \\( b \\), with each entry \\( b^i \\in \\mathbb{N} \\) denoting the distance between the current frame and the nearest beat frame. This representation not only mitigates the sparsity caused by binary formulations but also provides the model with local temporal context. This significantly facilitates the precise acquisition of rhythmic characteristics within the choreography and injects more effective beat controllability into the learning process. The beat embedding \\( \\mathbf{e}_b \\in \\mathbb{R}^{L \\times d} \\) is obtained using a separate embedding layer, followed by the transformer-based encoder \\( \\mathcal{E}_b \\). Please refer to the supplementary materials for detailed representations of music and dance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.635, + 0.563, + 0.65 + ], + "angle": 0, + "content": "3.5 Hierarchical Multi-Condition Fusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Although the keyframe constraint offers precise pose specifications for specific frames, it is still too sparse which significantly increases the learning difficulty of the model. Naively incorporating this condition with others introduces a vast of padding, leading to excessive noises that may impair the effectiveness of keyframe constraints. By contrast, music condition and beat condition are much more denser, with valid values for each frame. To harmoniously combine these three conditions, we introduce a novel multi-condition hierarchical fusion mechanism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Specifically, in the first stage, we inject sparse keyframe condition into other dense conditions through a beat-aware dilation attention strategy, which shares a similar spirit of DiffKFC [46]. We observe that dance movements on motion-beat frames are often more iconic in the whole sequence, distinguished by their higher emphasis and significance. As these movements typically serve as guiding" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Z. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.278, + 0.79, + 0.377 + ], + "angle": 0, + "content": "Fig. 3: Illustration of the beat-aware mask dilation scheme. The first row visualizes the keyframe mask, with deep green indicating valid control constraints and gray indicating invalid ones. The second row presents the dilation step curve with red lines marking beat frames. The third row is a heatmap of the dilation step. The fourth row shows the neighborhood range of keyframes, with light green indicating the expanded valid region from beat-aware mask dilation. The final row displays the beat-aware dilated keyframe mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.79, + 0.628 + ], + "angle": 0, + "content": "elements, influencing the choreography and style of the entire dance, they exert a more substantial influence on the surrounding frames. To utilize this property, we design a beat-aware mask dilation scheme. Unlike DiffKFC which uses identical dilation steps for all keyframes, treating them uniformly, we opt for different dilation weights based on the distance between the keyframe and the nearest designated beat frame. The dilation step \\( n \\) is calculated as \\( n = \\lceil s \\cdot e^{-2\\frac{b^i}{d^i}} \\rceil \\), where \\( b^i \\) and \\( d^i \\) are the ground truth beat distance at frame \\( i \\) and the distance between frame \\( i \\)'s adjacent motion-beat frames, respectively. Base dilation step \\( s \\) is a hyper-parameter controlling overall dilation speed. The closer the keyframe is to the nearest beat frame, the larger the dilation step. This strategy enables the model to gradually propagate the keyframe condition to the surrounding frames with beat awareness. The beat-aware mask dilation scheme is illustrated in Fig. 3. Given the keyframe binary mask \\( \\mathbf{M} \\) and dilation step \\( n \\), the beat-aware dilated mask \\( \\mathbf{M}^d \\) is obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.26, + 0.643, + 0.788, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {d} [ i ] = \\left\\{ \\begin{array}{l l} \\max _ {j} \\mathbf {M} [ i - j ], j \\in \\{- n, n + 1, \\dots , n - 1, n \\} & \\text {i f} \\mathbf {M} _ {i} = 1, \\\\ \\mathbf {M} _ {i}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.789, + 0.763 + ], + "angle": 0, + "content": "The beat-aware dilated attention mask is then calculated as \\(\\mathbf{M}_{attn} = \\mathbf{MM}_d^{\\mathsf{T}}\\). This beat-aware mask dilation scheme enables beat-aware keyframe condition propagation, thereby amplifying the synergy among conditions characterized by varying degrees of sparsity, thus significantly boosting the generation quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In the second phase, we concatenate two refined dense conditions, namely keyframe-fused music embedding and keyframe-fused beat embedding. These concatenated conditions are fed into a transformer condition fusion module, yielding the final fused features that can be subsequently sent to the conditional diffusion denoising block." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.436, + 0.162 + ], + "angle": 0, + "content": "3.6 Beat Alignment Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.399 + ], + "angle": 0, + "content": "For controllable dance generation, it is crucial to ensure precise beat alignment between the generated dance motions and the given beat conditions. Current methods typically ignore the explicit constraints on the beats of the generated dance sequences. As a consequence, these approaches lack necessary beat controllability and struggle to guarantee beat synchronization between the generated dance motions and musical beats, which severely deteriorates the overall generation quality. To conquer this issue, building upon our beat representation, we present a beat alignment loss to provide explicit supervision that enforces the beats of the generated motions to precisely align with the given beat conditions. To this end, we first pre-train a beat distance estimator, which accepts a motion sequence as input and predicts the distance from the current frame to the nearest motion beat frame for each frame in the sequence. During the training process, the pre-trained beat distance estimator is responsible for providing supervision signals on the beats of the generated motions. The beat alignment loss \\(\\mathcal{L}_{beat}\\) can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.409, + 0.787, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b e a t}} = \\sum_ {i = 1} ^ {L} w _ {s} ^ {i} \\cdot w _ {b} ^ {i} \\cdot \\mathbf {M S E} \\left(b ^ {i}, \\hat {b} ^ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.789, + 0.653 + ], + "angle": 0, + "content": "where \\( w_{s}^{i} = \\frac{1}{1 + e^{a\\cdot(c - \\|b^{i} - \\hat{b}^{i}\\| / b^{i})}} \\), and \\( w_{b}^{i} = e^{-\\frac{2b^{i}}{d^{i}}} \\). Here, \\( \\hat{b}^i \\) represents the predicted beat distance at frame \\( i \\). \\( w_{s}^{i} \\) denotes an adaptive weight, which is derived from [26]. It aims to accentuate supervision in regions where beat alignment is imprecise. This is achieved by suppressing the loss magnitude in well-aligned beat regions. This strategy enables the model to focus more attention on challenging instances. The hyper-parameters \\( a \\) and \\( c \\) are used for controlling the degree of penalization and the threshold for beat alignment accuracy, respectively. \\( \\mathbf{MSE}(\\cdot ,\\cdot) \\) denotes the mean squared error. \\( w_{b}^{i} \\) is an adaptive weight that enhances the supervision of frames closely aligned with the beats while diminishing the supervision of frames distantly related to the beats. Benefiting from the beat alignment loss, our model can yield dance motions with superior beat alignment and controllability, largely elevating the quality of the generated results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.371, + 0.689 + ], + "angle": 0, + "content": "3.7 Other Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.787, + 0.744 + ], + "angle": 0, + "content": "Apart from the proposed beat alignment loss \\(\\mathcal{L}_{\\text {beat }}\\), we also employ the basic diffusion loss \\(\\mathcal{L}_{\\text {simple }}\\) (Eq. 2) and several additional auxiliary losses to govern the training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.746, + 0.787, + 0.793 + ], + "angle": 0, + "content": "Kinematic Loss. We adopt the same auxiliary losses as [41, 42]. The joint positions loss \\(\\mathcal{L}_{joint}\\) (Eq. 6) and velocities loss \\(\\mathcal{L}_{vel}\\) (Eq. 7) are used to improve the overall physical plausibility of the generated dance motions." + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.805, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| F K \\left(\\mathbf {x} _ {0} ^ {i}\\right) - F K \\left(\\hat {\\mathbf {x}} _ {0} ^ {i}\\right) \\| _ {2} ^ {2}, \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Z. Huang et al." + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.159, + 0.788, + 0.199 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v e l} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| \\mathbf {x} _ {0} ^ {i ^ {\\prime}} - \\hat {\\mathbf {x}} _ {0} ^ {i ^ {\\prime}} \\| _ {2} ^ {2} + \\| F K (\\mathbf {x} _ {0} ^ {i}) ^ {\\prime} - F K (\\hat {\\mathbf {x}} _ {0} ^ {i}) ^ {\\prime} \\| _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.205, + 0.786, + 0.237 + ], + "angle": 0, + "content": "The contact consistency loss \\(\\mathcal{L}_{\\text {contact }}\\) (Eq. 8) proposed in EDGE [42] is employed to alleviating the foot-sliding." + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.246, + 0.788, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t a c t}} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| F K _ {\\text {f o o t}} \\left(\\hat {\\mathbf {x}} _ {0} ^ {i}\\right) ^ {\\prime} \\cdot \\hat {g} ^ {i} \\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.298, + 0.789, + 0.389 + ], + "angle": 0, + "content": "where \\( i \\) denotes the index of the frame, and \\( FK(\\cdot) \\) is the forward kinematics function that converts the generated 6-DOF rotation representation into 3D key points in Cartesian space. \\( \\hat{g}^i \\) denotes the predicted binary foot contact label. In addition, we employ acceleration loss \\( \\mathcal{L}_{acc} \\) to further improve the quality of generated dance motions, which is suggested in previous work [35], preventing jitters in the generated dance motions." + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.4, + 0.788, + 0.441 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a c c} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| \\mathbf {x} _ {0} ^ {i ^ {\\prime \\prime}} - \\hat {\\mathbf {x}} _ {0} ^ {i ^ {\\prime \\prime}} \\| _ {2} ^ {2} + \\| F K (\\mathbf {x} _ {0} ^ {i}) ^ {\\prime \\prime} - F K (\\hat {\\mathbf {x}} _ {0} ^ {i}) ^ {\\prime \\prime} \\| _ {2} ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.451, + 0.585, + 0.464 + ], + "angle": 0, + "content": "The overall kinematic loss is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.286, + 0.478, + 0.788, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {k i n}} = \\lambda_ {\\text {j o i n t}} \\mathcal {L} _ {\\text {j o i n t}} + \\lambda_ {\\text {v e l}} \\mathcal {L} _ {\\text {v e l}} + \\lambda_ {\\text {c o n t a c t}} \\mathcal {L} _ {\\text {c o n t a c t}} + \\lambda_ {\\text {a c c}} \\mathcal {L} _ {\\text {a c c}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.505, + 0.787, + 0.551 + ], + "angle": 0, + "content": "where \\(\\lambda_{joint}\\), \\(\\lambda_{vel}\\), \\(\\lambda_{contact}\\), and \\(\\lambda_{acc}\\) are hyper-parameters controlling the weights of the corresponding losses. We set \\(\\lambda_{joint} = 1\\), \\(\\lambda_{vel} = 2.5\\), \\(\\lambda_{contact} = 10\\), and \\(\\lambda_{acc} = 0.1\\) empirically. The overall loss function is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.562, + 0.788, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {\\text {k i n}} \\mathcal {L} _ {\\text {k i n}} + \\lambda_ {\\text {b e a t}} \\mathcal {L} _ {\\text {b e a t}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.59, + 0.788, + 0.621 + ], + "angle": 0, + "content": "where \\(\\lambda_{kin}\\) and \\(\\lambda_{beat}\\) are hyper-parameters controlling the weights of the corresponding losses. \\(\\lambda_{kin}\\) and \\(\\lambda_{beat}\\) are set as 1 and 0.5, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.377, + 0.66 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Dataset. Following previous works [35, 42], we choose the most widely used dataset, AIST++ [23] for both training and testing. This dataset comprises 1,408 dance sequences encompassing 10 distinct street dance genres. Each dance genre incorporates 6 musical pieces with different BPMs. \\(85\\%\\) of these sequences are categorized as basic choreographies, where performers execute the same choreography across all 6 musical pieces, each featuring varying tempos to synchronize with distinct musical beats. With its high choreographic and beat diversity, this dataset is ideal for our controllable dance generation task." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.796, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation Details. For the hierarchical multi-condition fusion module, we implement six sparse-dense fusion blocks and two dense-dense fusion blocks, utilizing a base dilation step \\(s\\) set to 4, 8, 12, 16, 20, 24. For diffusion" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "process, we adopt the cosine schedule as beat schedule and set diffusion steps \\( t \\) as 1000 diffusion following [41]. The classifier-free guidance scale is set to 2. The diffusion module aligns with a transformer-based architecture similar to [42]. We adopt Adam optimizer with the learning rate set as 2e-4 and the training batch size is set to 64. During training, we randomly sample between \\( 1\\% \\) and \\( 30\\% \\) of the ground truth (GT) motion frames as the keyframe conditions. This training strategy can effectively improve the generalization capability of our method on different numbers of keyframes. We use the GT motion beats as beat conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.372 + ], + "angle": 0, + "content": "To ensure fair comparisons, we follow EDGE [42] to crop all data into 5-second clips at 30 frames per second, with 2.5 seconds of overlapping. During testing, we randomly sample \\(10\\%\\) of keyframes from unpaired samples in the testing set as keyframe conditions and subsets of testing musical beats as beat conditions to prevent data leakage. The musical beats used for experiments are extracted by the off-the-shelf audio toolkit Librosa [17]. All experiments are conducted on 4 NVIDIA RTX3090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Evaluation Metrics. For quantitative evaluation, we measure the generated dance from three aspects: generation quality, diversity, and controllability. Due to the limited samples in the testing set of AIST++ [23], the prior work EDGE [42] has demonstrated that the Fréchet Inception Distance (FID) is not reliable for measuring the generation quality. Therefore, we do not adopt FID as the evaluation metric. Instead, we assess generation quality from two different perspectives: music-dance correlation and kinematic plausibility. In terms of music-dance correlation, we utilize the Beat Alignment Score (BAS) metric, following the approach presented in [35]. BAS quantifies the synchronization quality between the generated dance and the musical beat. For the evaluation of kinematic plausibility, we adopt the Physical Foot Contact score (PFC) metric proposed in [42]. PFC provides a metric for measuring the physical plausibility of the generated dance from a kinematic perspective. As for the diversity, we measure the average feature distance of kinetic \\((\\mathrm{Div}_k)\\) and geometric \\((\\mathrm{Div}_g)\\) features extracted by fairmotion [12] following the previous works [23, 35, 42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.737 + ], + "angle": 0, + "content": "For assessing controllability, we employ the Key Pose Distance (KPD) and Beat Assignment Precision (BAP) metrics to evaluate keyframe control and beat control, respectively. KPD quantifies the average mean square error of local joint positions in Cartesian space at the keyframes. To evaluate beat control, we measure the alignment between the generated dance and the designated beat condition. BAP is defined as the percentage of beats in the generated dance that are correctly assigned to the designated beat frames. Note that this metric differs from the BAS, as a dance sequence with a high BAS does not necessarily adhere to the specific motion beat choreography constraint we intend to impose." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.757, + 0.537, + 0.772 + ], + "angle": 0, + "content": "4.1 Comparison to Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Currently, only three music-to-dance methods are publicly accessible: EDGE [42], Bailando [35], and FACT [23]. Of these, EDGE [42] is most directly aligned with our approach, as it similarly accommodates keyframes as inputs for generating dance sequences. We therefore assess EDGE using the same keyframe conditions" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Z. Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons among different methods on AIST++. Bold indicates best result. ↓ means lower is better, ↑ means higher is better and → means closer to the ground truth is better." + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.199, + 0.742, + 0.334 + ], + "angle": 0, + "content": "
MethodsQualityDiversityControllability
PFC ↓ BAS ↑Divk→Divm→KPD ↓ BAP ↑
Ground Truth1.3380.3849.7737.212--
FACT [23]2.6980.2029.7047.342--
Bailando [35]1.5780.2159.6227.175--
EDGE [42](keyframes)1.0840.2359.7437.2740.859-
Ours(beat & keyframes)0.9660.6619.6607.2480.3060.793
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.79, + 0.422 + ], + "angle": 0, + "content": "as our method to ensure a fair comparison. Conversely, certain approaches, like DanceFormer [21], are not open-source, thus hindering any direct evaluative comparison. We use the original test split of AIST++ [23] for evaluation, which has no overlap with the training split in terms of both the music and dance choreography." + }, + { + "type": "table_caption", + "bbox": [ + 0.248, + 0.444, + 0.46, + 0.457 + ], + "angle": 0, + "content": "Table 2: Results of user study." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.459, + 0.472, + 0.57 + ], + "angle": 0, + "content": "
User StudyBeat-It Win Rate
FACT92.2%
Bailando78.8%
EDGE \n(keyframes)QualityControllability \n(Keyframe)
60.3%86.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.507, + 0.444, + 0.765, + 0.457 + ], + "angle": 0, + "content": "Table 3: Ablation study on AIST++." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.459, + 0.767, + 0.556 + ], + "angle": 0, + "content": "
AblationsQualityControllability
PFC ↓BAS ↑KPD ↓BAP ↑
w/o HF25.6260.3220.4770.323
w/o BD1.6320.3580.3890.371
w/o Lbeat1.3420.3970.3430.411
Ours0.9660.6610.3060.793
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Quantitative Comparisons. The quantitative results are shown in Tab. 1. As is revealed, our method surpasses all the competitors in terms of generation quality, diversity, and controllability. Particularly, our method shows considerable advantages over the compared methods for BAS metrics, notably with an improvement of 0.426 compared with the SOTA method EDGE [42]. This demonstrates the effectiveness of our method in generating beat-synchronized dance sequences while maintaining diversity. Furthermore, in the aspect of controllability, our method also significantly excels EDGE [42] with a notable improvement of 0.553 in KPD. Fig. 4 shows the mean joint velocities over time for a specific dance motion concerning both the GT and the existing methods. Notably, compared with others, our approach demonstrates more motion beats (local minimum point of velocities) precisely at the beat frames. This contributes to a better visual alignment with the audio when compared to the other methods. The BAP further validates our method's superior ability to produce dance sequences precisely aligned with the beat condition. Due to limited space, we present additional quantitative experiments in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "User Study. We also conduct a user study to assess the generation quality and controllability of our method. We randomly select 20 generated dance sequences" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Fig. 4: Visualization comparison on beat alignment among different methods. The motion generated by our method shows precise beat alignment with the given beat condition, demonstrating the superiority of our method in beat control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.789, + 0.495 + ], + "angle": 0, + "content": "from the testing set and ask 18 participants to pick the best one among our method and the other compared methods. In particular, participants are required to choose the superior dance according to several criteria, namely, the overall visual plausibility of movements, synchronization with the dance beats, and keyframe controllability. The results of the user study are tabulated in Tab. 2. Our method shows noticeable advantages in terms of generation quality compared to other approaches, particularly in beat alignment. For qualitative evaluations on keyframe-conditioned controllability, we also provide the participants with the keyframe references for intuitive comparison between our method against EDGE [42]. The results show that our approach owns better motion controllability than EDGE [42]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.515, + 0.402, + 0.529 + ], + "angle": 0, + "content": "4.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.613 + ], + "angle": 0, + "content": "To get a more comprehensive insight into our main contributions, we conduct ablation studies on the hierarchical multi-condition fusion module, the beat-aware dilation scheme, and the beat alignment loss to evaluate their efficacy. The quantitative results are shown in Tab. 3. The qualitative results can be also found in the supplementary video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.809 + ], + "angle": 0, + "content": "Hierarchical Multi-condition Fusion Module. To verify the significance of the hierarchical multi-condition fusion module, we remove it by simply concatenating all the conditions and feeding them into a one-stage encoder (w/o HF). As shown in Tab. 3, the simple one-stage fusion module can lead to inferior results in both quality and controllability. This is because the direct concatenation of multiple conditions can lead to inevitable condition conflicts, giving rise to suboptimal optimization of the model. Even worse, the naive padding of the sparse keyframe condition introduces severe noises, further complicating the learning process and resulting in unsatisfactory generation performance. By decently fusing multiple conditions gradually, our hierarchical multi-condition fusion mechanism effectively mitigates the conflicts among different conditions and yields more informative conditions, allowing for higher-quality generation with better controllability." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Beat-aware Dilation Scheme. To validate the effectiveness of our beat-aware dilation scheme, we replace it with a vanilla masked attention strategy (w/o" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Z. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "BD). As shown in Tab. 3, the absence of the beat-aware dilation scheme notably impairs both keyframe and beat choreography controllability. The reason behind this is that beat-aware dilation can dynamically adjust the expansion degree of the attention mask, contingent on the distance between the keyframe and the beat frames. This tactical approach significantly improves the synchronization between keyframes and their associated beat frames, thereby empowering the model to more effectively harness prior knowledge of beats. Consequently, this enhancement considerably bolsters beat control while augmenting keyframe controllability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.79, + 0.419 + ], + "angle": 0, + "content": "Beat Alignment Loss. To further validate the effectiveness of the beat alignment loss, we assess the model's performance by comparing it to its counterpart without it (w/o \\(\\mathcal{L}_{\\text {beat }}\\)). As shown in Tab. 3, the ablated variant without the beat alignment loss demonstrated performance degradation in beat choreography, showcasing decreases of 0.264 and 0.382 in BAS and BAP, respectively. This decline in performance can be attributed to the advantageous impact of direct supervision during training, particularly in guiding the model to comprehend beat-level dance choreography through the proposed beat alignment loss. Consequently, this enhancement significantly improves the synchronization between motion sequences and designated beats." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.44, + 0.511, + 0.458 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.79, + 0.653 + ], + "angle": 0, + "content": "In this work, we propose a novel multi-condition diffusion-based framework, BeatIt, for beat-synchronized and key pose-guided dance generation. The proposed framework presents a hierarchical multi-condition fusion mechanism equipped with a beat-aware mask dilation scheme to integrate conditions with different information sparsity. To achieve precise beat synchronization, we explicitly disentangle the beats from the music and inject beat controllability throughout the entire generation process. Additionally, we introduce a specifically designed beat alignment loss to provide explicit guidance and supervision on motion beats. Both qualitative and quantitative experimental results on the AIST++ dataset validate the superiority of our method in producing high-quality dance sequences with precise beat synchronization and flexible keyframe control. Further limitations and broader impact are discussed in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.407, + 0.692 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.79, + 0.841 + ], + "angle": 0, + "content": "The work is supported by China National Key R&D Program (No. 2023YFE0202700), Key-Area Research and Development Program of Guangzhou City (No. 2023B01J0022), Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (No. 2024B1515040010), Health and Medical Research Fund (HMRF) of Hong Kong Health Bureau (No. 10211516), Innovation and Technology Fund - Innovation and Technology Support Programme (ITF-ITSP) (No. ITS/127/22), Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2023B1515020097), National Research Foundation Singapore under the AI Singapore Programme (No. AISG3-GV-2023-011)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.787, + 0.216 + ], + "angle": 0, + "content": "1. Alemi, O., Françoise, J., Pasquier, P.: Groovenet: Real-time music-driven dance movement generation using artificial neural networks. networks 8(17), 26 (2017) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.787, + 0.258 + ], + "angle": 0, + "content": "2. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM TOG 42(4), 1-20 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.787, + 0.286 + ], + "angle": 0, + "content": "3. Ao, T., Zhang, Z., Liu, L.: Gesturediffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.787, + 0.327 + ], + "angle": 0, + "content": "4. Araújo, J.P., Li, J., Vetrivel, K., Agarwal, R., Wu, J., Gopinath, D., Clegg, A.W., Liu, K.: Circle: Capture in rich contextual environments. In: CVPR. pp. 21211-21221 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.787, + 0.356 + ], + "angle": 0, + "content": "5. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR. pp. 9760-9770 (2023) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.787, + 0.397 + ], + "angle": 0, + "content": "6. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. NeurIPS 34, 19822-19835 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.787, + 0.453 + ], + "angle": 0, + "content": "7. Ferreira, J.P., Coutinho, T.M., Gomes, T.L., Neto, J.F., Azevedo, R., Martins, R., Nascimento, E.R.: Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics 94, 11-21 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.787, + 0.494 + ], + "angle": 0, + "content": "8. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV. pp. 89-106. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.787, + 0.535 + ], + "angle": 0, + "content": "9. Gao, X., Hu, L., Zhang, P., Zhang, B., Bo, L.: Dancemeld: Unraveling dance phrases with hierarchical latent codes for music-to-dance synthesis. arXiv preprint arXiv: 2401.10242 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.787, + 0.577 + ], + "angle": 0, + "content": "10. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: Imos: Intent-driven full-body motion synthesis for human-object interactions. In: CGF. vol. 42, pp. 1-12. Wiley Online Library (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.787, + 0.618 + ], + "angle": 0, + "content": "1. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: ICCV. pp. 9942-9952 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.787, + 0.658 + ], + "angle": 0, + "content": "2. Gopinath, D., Won, J.: fairmotion - tools to load, process and visualize motion capture data. Github (2020), https://github.com/facebookresearch/fairmotion11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.787, + 0.687 + ], + "angle": 0, + "content": "3. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.715 + ], + "angle": 0, + "content": "4. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. ACM TOG 35(4), 1-11 (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.787, + 0.756 + ], + "angle": 0, + "content": "5. Huang, R., Hu, H., Wu, W., Sawada, K., Zhang, M., Jiang, D.: Dance revolution: Long-term dance generation with music via curriculum learning. arXiv preprint arXiv:2006.06119 (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.787, + 0.797 + ], + "angle": 0, + "content": "6. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR. pp. 16750-16761 (2023) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.839 + ], + "angle": 0, + "content": "7. Jin, Y., Zhang, J., Li, M., Tian, Y., Zhu, H., Fang, Z.: Towards the automatic anime characters creation with generative adversarial networks. arXiv preprint arXiv:1708.05509 (2017) 11" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Z. Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI. vol. 37, pp. 8255-8263 (2023) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.215 + ], + "angle": 0, + "content": "19. Kim, J., Oh, H., Kim, S., Tong, H., Lee, S.: A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In: CVPR. pp. 3490-3500 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.215, + 0.787, + 0.242 + ], + "angle": 0, + "content": "20. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. NeurIPS 32 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.242, + 0.786, + 0.283 + ], + "angle": 0, + "content": "21. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI. vol. 36, pp. 1272-1279 (2022) 2, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.283, + 0.787, + 0.323 + ], + "angle": 0, + "content": "22. Li, J., Yin, Y., Chu, H., Zhou, Y., Wang, T., Fidler, S., Li, H.: Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171 (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.324, + 0.787, + 0.364 + ], + "angle": 0, + "content": "23. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV. pp. 13401-13412 (2021) 2, 4, 7, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.364, + 0.787, + 0.405 + ], + "angle": 0, + "content": "24. Lin, J., Chang, J., Liu, L., Li, G., Lin, L., Tian, Q., Chen, C.w.: Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In: CVPR. pp. 23222-23231 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.787, + 0.433 + ], + "angle": 0, + "content": "25. Liu, H., Xu, C., Yang, Y., Zeng, L., He, S.: Drag your noise: Interactive point-based editing via diffusion semantic propagation. In: CVPR. pp. 6743-6752 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.459 + ], + "angle": 0, + "content": "26. Lu, X., Ma, C., Shen, J., Yang, X., Reid, I., Yang, M.H.: Deep object tracking with shrinkage loss. IEEE TPAMI 44(5), 2386-2401 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.5 + ], + "angle": 0, + "content": "27. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.5, + 0.787, + 0.542 + ], + "angle": 0, + "content": "28. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "29. Qi, Q., Zhuo, L., Zhang, A., Liao, Y., Fang, F., Liu, S., Yan, S.: Diffdance: Cascaded human motion diffusion model for dance generation. In: ACM MM. pp. 1374-1382 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "30. Qin, C., Zhang, S., Yu, N., Feng, Y., Yang, X., Zhou, Y., Wang, H., Niebles, J.C., Xiong, C., Savarese, S., et al.: Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.663 + ], + "angle": 0, + "content": "31. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML. pp. 8821-8831. PMLR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.787, + 0.691 + ], + "angle": 0, + "content": "32. Ren, X., Li, H., Huang, Z., Chen, Q.: Self-supervised dance video synthesis conditioned on music. In: ACM MM. pp. 46-54 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.718 + ], + "angle": 0, + "content": "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.771 + ], + "angle": 0, + "content": "34. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS 35, 36479-36494 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "35. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR. pp. 11050-11059 (2022) 2, 4, 7, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "36. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando++: 3d dance gpt with choreographic memory. IEEE TPAMI (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.289, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Sun, G., Wong, Y., Cheng, Z., Kankanhalli, M.S., Geng, W., Li, X.: Deepdance: music-to-dance motion choreography with adversarial learning. IEEE TMM 23, 497-509 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "38. Sun, J., Wang, C., Hu, H., Lai, H., Jin, Z., Hu, J.F.: You never stop dancing: Non-freezing dance generation via bank-constrained manifold projection. NeurIPS 35, 9995-10007 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.233, + 0.786, + 0.261 + ], + "angle": 0, + "content": "39. Tang, T., Jia, J., Mao, H.: Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In: ACM MM. pp. 1598-1606 (2018) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.289 + ], + "angle": 0, + "content": "40. Tang, Z., Yang, Z., Zhu, C., Zeng, M., Bansal, M.: Any-to-any generation via composable diffusion. arXiv preprint arXiv: 2305.11846 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.318 + ], + "angle": 0, + "content": "41. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.787, + 0.346 + ], + "angle": 0, + "content": "42. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: CVPR. pp. 448-458 (2023) 2, 4, 7, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.347, + 0.787, + 0.389 + ], + "angle": 0, + "content": "43. Valle-Pérez, G., Henter, G.E., Beskow, J., Holzapfel, A., Oudeyer, P.Y., Alexanderson, S.: Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM TOG 40(6), 1-14 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.39, + 0.787, + 0.431 + ], + "angle": 0, + "content": "44. Wang, T., Zhang, T., Zhang, B., Ouyang, H., Chen, D., Chen, Q., Wen, F.: Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.432, + 0.787, + 0.472 + ], + "angle": 0, + "content": "45. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS 35, 14959-14971 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.474, + 0.787, + 0.514 + ], + "angle": 0, + "content": "46. Wei, D., Sun, X., Sun, H., Li, B., Hu, S., Li, W., Lu, J.: Understanding text-driven motion synthesis with keyframe collaboration via diffusion models. arXiv preprint arXiv:2305.13773 (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.516, + 0.787, + 0.571 + ], + "angle": 0, + "content": "47. Xie, Y., Lin, Y., Cai, W., Xu, X., Zhang, H., Du, Y., He, S.: D3still: Decoupled differential distillation for asymmetric image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17181-17190 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.573, + 0.787, + 0.6 + ], + "angle": 0, + "content": "48. Xu, C., Chen, Z., Mai, J., Xu, X., He, S.: Pose-and attribute-consistent person image synthesis. ACM TOMM 19(2s), 1-21 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.601, + 0.787, + 0.628 + ], + "angle": 0, + "content": "49. Xu, C., Li, K., Luo, X., Xu, X., He, S., Zhang, K.: Fully deformable network for multiview face image synthesis. IEEE TNNLS (1-15) (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.63, + 0.787, + 0.656 + ], + "angle": 0, + "content": "50. Xu, C., Xu, Y., Zhang, H., Xu, X., He, S.: Dreamanime: Learning style-identity textual disentanglement for anime and beyond. IEEE TVCG (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.658, + 0.787, + 0.697 + ], + "angle": 0, + "content": "51. Yalta, N., Watanabe, S., Nakadai, K., Ogata, T.: Weakly-supervised deep recurrent neural networks for basic dance step generation. In: IJCNN. pp. 1-8. IEEE (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.7, + 0.787, + 0.728 + ], + "angle": 0, + "content": "52. Yang, Z., Wen, Y.H., Chen, S.Y., Liu, X., Gao, Y., Liu, Y.J., Gao, L., Fu, H.: Keyframe control of music-driven 3d dance generation. IEEE TVCG (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.729, + 0.787, + 0.756 + ], + "angle": 0, + "content": "53. Yin, W., Yin, H., Baraka, K., Kragic, D., Björkman, M.: Dance style transfer with cross-modal transformer. In: WACV. pp. 5058-5067 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "54. Yu, Y., Liu, B., Zheng, C., Xu, X., Zhang, H., He, S.: Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In: CVPR. pp. 7109-7118 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "55. Zhang, C., Tang, Y., Zhang, N., Lin, R.S., Han, M., Xiao, J., Wang, S.: Bidirectional autoregressive diffusion model for dance generation. arXiv preprint arXiv: 2402.04356 (2024) 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Z. Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "56. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "57. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "58. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.286 + ], + "angle": 0, + "content": "59. Zheng, C., Liu, B., Xu, X., Zhang, H., He, S.: Learning an interpretable stylized subspace for 3d-aware animatable artforms. IEEE TVCG (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "60. Zheng, C., Liu, B., Zhang, H., Xu, X., He, S.: Where is my spot? few-shot image generation via latent subspace optimization. In: CVPR. pp. 3272-3281 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "61. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: CVPR. pp. 5632-5641 (2023) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.371 + ], + "angle": 0, + "content": "62. Zhu, L., Liu, X., Liu, X., Qian, R., Liu, Z., Yu, L.: Taming diffusion models for audio-driven co-speech gesture generation. In: CVPR. pp. 10544-10553 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.371 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_origin.pdf b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ea6e8eb96cf58de108b4dde18aa70a2bb83cfcd6 --- /dev/null +++ b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/65595e86-aeb0-427c-bb7b-72a9a31a0745_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0107766a37cf88f250e8ec3c93472fea6e826f2b88e8f1a85b191c2bd698025 +size 4367758 diff --git a/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/full.md b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..902350d88ba20c560500b0d422bda6f771bdcfdf --- /dev/null +++ b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/full.md @@ -0,0 +1,300 @@ +# Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation + +Zikai Huang $^{1}$ , Xuemiao $\mathrm{Xu}^{1,3,4(\boxtimes)}$ , Cheng $\mathrm{Xu}^{2(\boxtimes)}$ , Huaidong Zhang $^{1,3}$ , Chenxi Zheng $^{1}$ , Jing Qin $^{2}$ , and Shengfeng He $^{5}$ + +$^{1}$ South China University of Technology, China xuemx@scut.edu.cn + +2 The Hong Kong Polytechnic University, Hong Kong SAR, China +cschengxu@gmail.com + +Guangdong Engineering Center for Large Model and GenAI Technology + +4 Guangdong Provincial Key Lab of Computational Intelligence and Cyberspace Information +5 Singapore Management University, Singapore https://zikaihuangscut.github.io/Beat-It/ + +![](images/db6755f06941388ecccbfa2bb7c69e65b47f8c6fd7368897d48248abded99334.jpg) +Fig. 1: We introduce Beat-It, a novel method for generating 3D dance motions with beat alignment and motion controllability. Our approach explicitly injects beat awareness and seamlessly integrates multiple conditions to guide the generation process, leading to beat-synchronized, key pose-guided dance generation. + +Abstract. Dance, as an art form, fundamentally hinges on the precise synchronization with musical beats. However, achieving aesthetically pleasing dance sequences from music is challenging, with existing methods often falling short in controllability and beat alignment. To address these shortcomings, this paper introduces Beat-It, a novel framework for beat-specific, key pose-guided dance generation. Unlike prior approaches, Beat-It uniquely integrates explicit beat awareness and key pose guidance, effectively resolving two main issues: the misalignment of generated dance motions with musical beats, and the inability to map key poses to specific beats, critical for practical choreography. Our approach disentangles beat conditions from music using a nearest beat distance representation and employs a hierarchical multi-condition fusion mechanism. This mechanism + +seamlessly integrates key poses, beats, and music features, mitigating condition conflicts and offering rich, multi-conditioned guidance for dance generation. Additionally, a specially designed beat alignment loss ensures the generated dance movements remain in sync with the designated beats. Extensive experiments confirm Beat-It's superiority over existing state-of-the-art methods in terms of beat alignment and motion controllability. + +Keywords: Dance generation $\cdot$ Beat synchronization $\cdot$ Multi-condition diffusion generation + +# 1 Introduction + +Dance, a time-honored art form, resonates across various ages and cultures due to its artistic and aesthetic significance, playing a vital role in the cultural and entertainment sectors. Traditional dance choreography creation, which demands an intricate balance of aesthetic movements, emotional expression, and precise synchronization with musical beats, often presents itself as a costly, laborious, and time-intensive endeavor, even for seasoned artists. + +Recent advances in deep learning, particularly in generative models, have catalyzed efforts to automate the choreography process. Initial approaches predominantly focused on single-condition dance generation, learning a direct mapping from music to dance motions [1, 15, 21-23, 35, 39, 43]. The advent of diffusion models, renowned for their exceptional content generation capabilities, has marked a paradigm shift in this domain. For instance, Tseng et al. [42] introduced a transformer-based diffusion model that sets a new benchmark in music-to-dance conversion, showcasing the potential of these models in elevating the quality of generated dance sequences. + +Notwithstanding the remarkable progress, achieving precise synchronization between dance motions and musical beats, coupled with flexible motion controllability in dance generation, remains a formidable challenge. In real-world choreography, the goal is often to assign specific key poses to particular musical beats and then fill in the connecting movements to form the complete dance sequence. This process requires not only the creation of designated key poses but also their harmonious synchronization with the musical beats. While few methods have explored key pose conditions [42, 52], none of them has considered explicit beat synchronization and controllability, rendering inferior dance motion generation. Furthermore, previous methods typically merge keyframe features with music features through simple concatenation or temporal/spatial blending. This approach, however, is flawed due to the sparse nature of keyframes. Consequently, these methods disproportionately emphasize dense music features at the expense of key pose conditions, leading to a notable misalignment between the generated dance motions and the specified keyframes. + +To combat the above issues, we aim to completely disentangle beats from input music and simultaneously incorporate explicit beat and key pose guidance to enable beat-controllable, key pose-guided dance generation. We introduce BeatIt, a diffusion-based, multi-condition framework designed for this purpose (see + +Fig. 1). Our approach begins by formulating beat condition as a nearest beat distance representation, which is a straightforward yet effective method to serve as beat guidance within our framework. To effectively harness and balance the multiple conditions involved, we propose a hierarchical multi-condition fusion mechanism. Initially, we integrate sparse key pose condition with dense beat and music conditions using a tailored beat-aware dilated cross-attention strategy. We further blend refined beat and music features to generate comprehensive multi-condition features. This strategy not only enhances beat and motion controllability in dance generation but also maintains the realism and coherence of the produced dance motions. Additionally, leveraging our beat representation, we introduce a novel beat alignment loss, explicitly ensuring synchronization between dance motions and the given beat condition, thereby significantly enhancing the overall quality of generation. Extensive experiments confirm that our method outperforms current state-of-the-art approaches in terms of beat alignment and motion controllability. Beyond the advantages above, our framework also supports arbitrary beat designation and flexible frame assignments of key poses. + +In summary, our main contributions are fourfold: + +- We introduce a multi-condition dance generation framework that achieves beat synchronization and enhanced motion controllability. To the best of our knowledge, our method makes the first attempt at beat-controllable key pose-guided dance generation. +- We present a hierarchical multi-condition fusion mechanism to effectively suppress the conflicts and fully exploit the complementary information among different conditions. +- We delve into the property of beats and formulate it as a nearest beat distance representation. A beat alignment loss is further tailored to offer explicit supervision signals to the generated dance motions, largely elevating the synchronization between the generated motions and the given beat conditions +- Extensive experiments demonstrate our method performs favorably against the state-of-the-art approaches, especially in motion controllability and motion-beat alignment. + +# 2 Related Work + +# 2.1 Human Motion Generation + +Human motion generation aims to generate realistic human motion automatically. Previous works can be generally categorized into three genres according to the input condition: text-conditioned [5,18,24,56,61], audio-conditioned, and scene-conditioned [4, 10, 16, 45]. Our work belongs to audio-conditioned human motion generation. But different from the primary audio-conditioned human motion generation, music-to-dance generation requires a comprehensive consideration of aesthetic movements, emotional expression, and precise synchronization with musical beats, making it extremely challenging to render convincing results. Recently, researchers have widely explored the application of diffusion models + +in human motion generation [2,3,5,16,18,61,62], which shows unprecedented performance than traditional generative models. In this paper, we get a deep insight into the beat synchronization problem and investigate a multi-condition diffusion-based scheme for controllable dance generation. + +# 2.2 Dance Generation + +With the emergence of deep learning [47-49], dance generation has attracted lots of research interest in recent years. Previous methods have explored various frameworks with different types of backbones to achieve single-condition dance generation, including CNN [14], RNN [1, 15, 39, 51], GCN [7, 32], VAE [9, 38], GAN [19, 20, 37], and Transformer [21-23, 35, 36, 43, 53]. FACT [23] proposes a full-attention cross-modal transformer model to capture the correlations between music and dance. Danceformer [21] presents a two-stage deterministic framework for audio-driven single-condition dance generation. It first generates key poses from the input music and then performs interpolation between the key poses. However, it falls short of arbitrary beat choreography and flexible key pose controllability due to its single-condition nature. Bailando [35] utilizes separate VQ-VAEs on upper/lower half bodies and a motion GPT to map the music and seed poses to dance sequences. Several recent attempts have investigated the diffusion models for dance generation [29, 42, 55], which significantly pushes the boundaries of music-to-dance performance. Compared to applying only a single input condition, multi-condition dance generation incorporates additional conditions beyond music into the generation process for better controllability, such as dance style labels [2], text [11], and keyframes [42, 52]. For example, LDA [2] introduced dance style labels as auxiliary conditions. However, it is limited to generating pre-defined styles and lacks flexibility. TM2D [11] adopts a VQ-VAE based model to support dual-modal music and text-driven dance motion generation. Yang et al. [52] employ normalizing flows to model dance motion probability distributions and utilize time embedding to achieve keyframe control. EDGE [42] adopts a diffusion-based framework and enables key pose control via temporal/spatial blending during inference. Although the existing methods have investigated additional conditions for better dance generation control, they typically neglect explicit guidance and supervision to align the generated motion dances with specified beats. Consequently, these methods struggle to produce both beat-specific and key pose-guided dance sequences. In contrast, we propose to explicitly disentangle the beat condition from music and inject flexible beat and key pose controllability into the entire generation process, significantly enhancing the beat alignment while supporting flexible frame assignments of key poses. + +# 2.3 Multi Condition Diffusion Generation + +Multi-condition diffusion generative models seek to inject multi-condition guidance into the generation process for versatile controllability [6,8,25,28,31,34, 44,50,54,59,60]. Benefiting from large-scale training data, Stable Diffusion [33] + +constructs a powerful foundation for image generation. Building upon this, ControlNet [57] and T2I-Adapter [27] efficiently integrate trainable parameters into existing models, enabling the inclusion of spatially localized input conditions in a pre-trained text-to-image diffusion model. Uni-ControlNet [58] proposes a framework equipped with two additional control modules for multi-condition control. Qin et al. [30] present a unified model capable of handling various visual conditions, resulting in a more streamlined scheme. Codi [40] acquire multi-modal synthesis skills through the mapping of diverse modalities into a unified space, training on lists of different multi-modal generation tasks. The aforementioned works typically controllable generation using pre-trained models under diverse conditions. Our work, conversely, focuses on the effective and comprehensive fusion of multiple conditions by presenting a hierarchical multi-condition fusion mechanism aimed at better integrating sparse keyframe control conditions with dense music and beat control conditions, resulting in more controllable dance generation. + +# 3 Method + +# 3.1 Problem Definition + +Dance generation aims to create realistic dance motion sequences $\mathcal{X} = \{\mathbf{x}^1, \mathbf{x}^2, \mathbf{x}^3, \dots, \mathbf{x}^L\}$ from a given piece of music $\mathcal{C}$ with a duration of $L$ . Here, $\mathbf{x}^i \in \mathbb{R}^D$ represents a human pose at the $i$ -th frame, which is denoted as a $D$ -dimensional vector. In this paper, we endeavor to disentangle beat condition from the input music and simultaneously incorporate explicit beat and key pose guidance to achieve beat-synchronized, key pose-guided dance generation. Given a music condition $\mathcal{C}$ , a sparse set of keyframes $\mathcal{X}^{ref} = \mathcal{X} \odot \mathbf{M}$ , where $\mathbf{M} \in \{0, 1\}^L$ is a temporal binary mask that specifies the assigned locations of the keyframes, and a beat condition $\mathcal{B}$ , our goal is to generate a beat-specific, and key pose-guided dance motion sequence $\hat{\mathcal{X}} = \{\hat{\mathbf{x}}^1, \hat{\mathbf{x}}^2, \hat{\mathbf{x}}^3, \dots, \hat{\mathbf{x}}^L\}$ that conforms to the music $\mathcal{C}$ , consistent with reference frames $\mathcal{X}^{ref}$ at positions specified by $\mathbf{M}$ , and synchronized with the beat condition $\mathcal{B}$ . + +# 3.2 Preliminaries + +In this work, we adopt diffusion model [13, 28] as the backbone of our framework. Diffusion models establish a consistent Markovian forward process that incrementally introduces noise into clean sample data $\mathbf{x}_0^{1:L} \in q(\mathbf{x}_0)$ , and a corresponding reverse process that progressively eliminates noise from noisy samples. For brevity, we use $\mathbf{x}_t$ to represent the entire sequence. During the forward process, a predefined noise variance schedule $\beta_t$ is employed to regulate the noise increments at each step. The forward process can be formulated as follows: + +$$ +q \left(\mathbf {x} _ {1: T} \mid \mathbf {x} _ {0}\right) = \prod_ {t = 1} ^ {T} q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right), q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}\right). \tag {1} +$$ + +![](images/d98af7a52b18e9bf64aa37787f267782767498df9ff1a0e7be5db58eb38f0032.jpg) +Fig.2: Overview of our proposed method, Beat-It. We generate a beat-synchronized dance sequence utilizing music, keyframes, and beat conditions. Conditional embeddings are derived and subsequently fused in a two-stage process: initially integrating sparse keyframe condition with other dense conditions, followed by the fusion of these dense conditions. The final fused condition is then processed by the conditional diffusion module. To ensure precise beat control, a beat alignment loss is employed to explicitly supervise the generated motions at the beat level. + +After $T$ steps, the sample data progressively transforms into a noise distribution $q(\mathbf{x}_T)$ , which is usually a standard Gaussian distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ . In the reverse process, the noise is gradually removed from the noisy sample $\mathbf{x}_T$ to produce the clean sample $\mathbf{x}_0$ . In our task, multi-conditional dance generation aims to model the distribution $p(\mathbf{x}_0|\mathbf{C})$ with a set of conditions $\mathbf{C}$ . Following Ho et al. [13], we directly predict the clean sample $\mathbf{x}_0$ from the noise distribution $q(\mathbf{x}_T)$ with the following objective: + +$$ +\mathcal {L} _ {\text {s i m p l e}} = \mathbb {E} _ {\mathbf {x} _ {0} \sim q (\mathbf {x} | \mathbf {C}), t \sim [ 1, T ]} \left[ \| \mathbf {x} _ {0} - G (\mathbf {x} _ {t}, t, \mathbf {C}) \| _ {2} ^ {2} \right]. \tag {2} +$$ + +A spatial or temporal constraint $\mathbf{x}^c$ with positions specified by a binary mask $m$ can be directly incorporated into the denoising process without additional training. This can be achieved by simply performing the following operation at every timestep: + +$$ +\mathbf {x} _ {t} = \mathbf {x} _ {t} ^ {c} \odot m + \mathbf {x} _ {t} \odot (1 - m), \tag {3} +$$ + +where $\odot$ denotes the Hadamard product. By doing so, regions under constraint can be replaced with forward-diffused samples of the constraint, while the remaining regions are left unchanged. However, it is worth noting that this simplistic approach tends to render inferior results when given sparse constraints. + +# 3.3 Overview + +The pipeline of our proposed model is illustrated in Fig. 2. To begin with, we feed the music condition $\mathcal{C}$ , keyframe condition $\mathcal{X}^{ref}$ , and beat condition $\mathcal{B}$ into + +three encoders respectively: the music encoder $\mathcal{E}_m$ , keyframe encoder $\mathcal{E}_r$ , and beat encoder $\mathcal{E}_b$ . This process yields the corresponding music embedding $\mathbf{e}_m$ , keyframe embedding $\mathbf{e}_r$ , and beat embedding $\mathbf{e}_b$ , where $\mathbf{e}_m, \mathbf{e}_r, \mathbf{e}_b \in \mathbb{R}^{L \times d}$ . Then, we forward these condition embeddings to a hierarchical multi-condition fusion module to produce comprehensive multi-condition features $\mathbf{c}_f$ , which are finally sent to the conditional diffusion denoising module to render the denoised dance motion sequence $\hat{\mathcal{X}}$ . + +# 3.4 Beat Representation + +As mentioned in previous studies [23, 35, 42], there exists a strong consistency between musical beats and dance motion beats. In practical dance choreography, the arrangement of motion beats is relatively flexible and does not require a strict point-to-point alignment with musical beats. Therefore, we introduce an independent beat condition disentangled from the music, to allow for flexible motion beat controllability in dance generation. For an input music sequence of duration $L$ , each frame can be categorized as either a beat or non-beat frame. A straightforward way to represent the beat condition is using a binary mask, where 1 and 0 denote the beat and non-beat frames respectively. However, such a simple binary presentation exhibits high sparsity as the beat frames only account for a small portion of the whole music sequence. This makes it less informative and, at times, perceived as noise or disregarded by the model. As an alternative, we propose to represent the beat condition as a vector $b$ , with each entry $b^i \in \mathbb{N}$ denoting the distance between the current frame and the nearest beat frame. This representation not only mitigates the sparsity caused by binary formulations but also provides the model with local temporal context. This significantly facilitates the precise acquisition of rhythmic characteristics within the choreography and injects more effective beat controllability into the learning process. The beat embedding $\mathbf{e}_b \in \mathbb{R}^{L \times d}$ is obtained using a separate embedding layer, followed by the transformer-based encoder $\mathcal{E}_b$ . Please refer to the supplementary materials for detailed representations of music and dance. + +# 3.5 Hierarchical Multi-Condition Fusion + +Although the keyframe constraint offers precise pose specifications for specific frames, it is still too sparse which significantly increases the learning difficulty of the model. Naively incorporating this condition with others introduces a vast of padding, leading to excessive noises that may impair the effectiveness of keyframe constraints. By contrast, music condition and beat condition are much more denser, with valid values for each frame. To harmoniously combine these three conditions, we introduce a novel multi-condition hierarchical fusion mechanism. + +Specifically, in the first stage, we inject sparse keyframe condition into other dense conditions through a beat-aware dilation attention strategy, which shares a similar spirit of DiffKFC [46]. We observe that dance movements on motion-beat frames are often more iconic in the whole sequence, distinguished by their higher emphasis and significance. As these movements typically serve as guiding + +![](images/bc014b8417fc5fb840cdaf4d584fa6004918cb17f28e891618e36e834d2824a9.jpg) +Fig. 3: Illustration of the beat-aware mask dilation scheme. The first row visualizes the keyframe mask, with deep green indicating valid control constraints and gray indicating invalid ones. The second row presents the dilation step curve with red lines marking beat frames. The third row is a heatmap of the dilation step. The fourth row shows the neighborhood range of keyframes, with light green indicating the expanded valid region from beat-aware mask dilation. The final row displays the beat-aware dilated keyframe mask. + +elements, influencing the choreography and style of the entire dance, they exert a more substantial influence on the surrounding frames. To utilize this property, we design a beat-aware mask dilation scheme. Unlike DiffKFC which uses identical dilation steps for all keyframes, treating them uniformly, we opt for different dilation weights based on the distance between the keyframe and the nearest designated beat frame. The dilation step $n$ is calculated as $n = \lceil s \cdot e^{-2\frac{b^i}{d^i}} \rceil$ , where $b^i$ and $d^i$ are the ground truth beat distance at frame $i$ and the distance between frame $i$ 's adjacent motion-beat frames, respectively. Base dilation step $s$ is a hyper-parameter controlling overall dilation speed. The closer the keyframe is to the nearest beat frame, the larger the dilation step. This strategy enables the model to gradually propagate the keyframe condition to the surrounding frames with beat awareness. The beat-aware mask dilation scheme is illustrated in Fig. 3. Given the keyframe binary mask $\mathbf{M}$ and dilation step $n$ , the beat-aware dilated mask $\mathbf{M}^d$ is obtained as follows: + +$$ +\mathbf {M} _ {d} [ i ] = \left\{ \begin{array}{l l} \max _ {j} \mathbf {M} [ i - j ], j \in \{- n, n + 1, \dots , n - 1, n \} & \text {i f} \mathbf {M} _ {i} = 1, \\ \mathbf {M} _ {i}, & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +The beat-aware dilated attention mask is then calculated as $\mathbf{M}_{attn} = \mathbf{MM}_d^{\mathsf{T}}$ . This beat-aware mask dilation scheme enables beat-aware keyframe condition propagation, thereby amplifying the synergy among conditions characterized by varying degrees of sparsity, thus significantly boosting the generation quality. + +In the second phase, we concatenate two refined dense conditions, namely keyframe-fused music embedding and keyframe-fused beat embedding. These concatenated conditions are fed into a transformer condition fusion module, yielding the final fused features that can be subsequently sent to the conditional diffusion denoising block. + +# 3.6 Beat Alignment Loss + +For controllable dance generation, it is crucial to ensure precise beat alignment between the generated dance motions and the given beat conditions. Current methods typically ignore the explicit constraints on the beats of the generated dance sequences. As a consequence, these approaches lack necessary beat controllability and struggle to guarantee beat synchronization between the generated dance motions and musical beats, which severely deteriorates the overall generation quality. To conquer this issue, building upon our beat representation, we present a beat alignment loss to provide explicit supervision that enforces the beats of the generated motions to precisely align with the given beat conditions. To this end, we first pre-train a beat distance estimator, which accepts a motion sequence as input and predicts the distance from the current frame to the nearest motion beat frame for each frame in the sequence. During the training process, the pre-trained beat distance estimator is responsible for providing supervision signals on the beats of the generated motions. The beat alignment loss $\mathcal{L}_{beat}$ can be formulated as follows: + +$$ +\mathcal {L} _ {\text {b e a t}} = \sum_ {i = 1} ^ {L} w _ {s} ^ {i} \cdot w _ {b} ^ {i} \cdot \mathbf {M S E} \left(b ^ {i}, \hat {b} ^ {i}\right), \tag {5} +$$ + +where $w_{s}^{i} = \frac{1}{1 + e^{a\cdot(c - \|b^{i} - \hat{b}^{i}\| / b^{i})}}$ , and $w_{b}^{i} = e^{-\frac{2b^{i}}{d^{i}}}$ . Here, $\hat{b}^i$ represents the predicted beat distance at frame $i$ . $w_{s}^{i}$ denotes an adaptive weight, which is derived from [26]. It aims to accentuate supervision in regions where beat alignment is imprecise. This is achieved by suppressing the loss magnitude in well-aligned beat regions. This strategy enables the model to focus more attention on challenging instances. The hyper-parameters $a$ and $c$ are used for controlling the degree of penalization and the threshold for beat alignment accuracy, respectively. $\mathbf{MSE}(\cdot ,\cdot)$ denotes the mean squared error. $w_{b}^{i}$ is an adaptive weight that enhances the supervision of frames closely aligned with the beats while diminishing the supervision of frames distantly related to the beats. Benefiting from the beat alignment loss, our model can yield dance motions with superior beat alignment and controllability, largely elevating the quality of the generated results. + +# 3.7 Other Losses + +Apart from the proposed beat alignment loss $\mathcal{L}_{\text {beat }}$ , we also employ the basic diffusion loss $\mathcal{L}_{\text {simple }}$ (Eq. 2) and several additional auxiliary losses to govern the training. + +Kinematic Loss. We adopt the same auxiliary losses as [41, 42]. The joint positions loss $\mathcal{L}_{joint}$ (Eq. 6) and velocities loss $\mathcal{L}_{vel}$ (Eq. 7) are used to improve the overall physical plausibility of the generated dance motions. + +$$ +\mathcal {L} _ {\text {j o i n t}} = \frac {1}{L} \sum_ {i = 1} ^ {L} \| F K \left(\mathbf {x} _ {0} ^ {i}\right) - F K \left(\hat {\mathbf {x}} _ {0} ^ {i}\right) \| _ {2} ^ {2}, \tag {6} +$$ + +$$ +\mathcal {L} _ {v e l} = \frac {1}{L} \sum_ {i = 1} ^ {L} \| \mathbf {x} _ {0} ^ {i ^ {\prime}} - \hat {\mathbf {x}} _ {0} ^ {i ^ {\prime}} \| _ {2} ^ {2} + \| F K (\mathbf {x} _ {0} ^ {i}) ^ {\prime} - F K (\hat {\mathbf {x}} _ {0} ^ {i}) ^ {\prime} \| _ {2} ^ {2}. \tag {7} +$$ + +The contact consistency loss $\mathcal{L}_{\text {contact }}$ (Eq. 8) proposed in EDGE [42] is employed to alleviating the foot-sliding. + +$$ +\mathcal {L} _ {\text {c o n t a c t}} = \frac {1}{L} \sum_ {i = 1} ^ {L} \| F K _ {\text {f o o t}} \left(\hat {\mathbf {x}} _ {0} ^ {i}\right) ^ {\prime} \cdot \hat {g} ^ {i} \| _ {2} ^ {2}, \tag {8} +$$ + +where $i$ denotes the index of the frame, and $FK(\cdot)$ is the forward kinematics function that converts the generated 6-DOF rotation representation into 3D key points in Cartesian space. $\hat{g}^i$ denotes the predicted binary foot contact label. In addition, we employ acceleration loss $\mathcal{L}_{acc}$ to further improve the quality of generated dance motions, which is suggested in previous work [35], preventing jitters in the generated dance motions. + +$$ +\mathcal {L} _ {a c c} = \frac {1}{L} \sum_ {i = 1} ^ {L} \| \mathbf {x} _ {0} ^ {i ^ {\prime \prime}} - \hat {\mathbf {x}} _ {0} ^ {i ^ {\prime \prime}} \| _ {2} ^ {2} + \| F K (\mathbf {x} _ {0} ^ {i}) ^ {\prime \prime} - F K (\hat {\mathbf {x}} _ {0} ^ {i}) ^ {\prime \prime} \| _ {2} ^ {2}. \tag {9} +$$ + +The overall kinematic loss is formulated as follows: + +$$ +\mathcal {L} _ {\text {k i n}} = \lambda_ {\text {j o i n t}} \mathcal {L} _ {\text {j o i n t}} + \lambda_ {\text {v e l}} \mathcal {L} _ {\text {v e l}} + \lambda_ {\text {c o n t a c t}} \mathcal {L} _ {\text {c o n t a c t}} + \lambda_ {\text {a c c}} \mathcal {L} _ {\text {a c c}}, \tag {10} +$$ + +where $\lambda_{joint}$ , $\lambda_{vel}$ , $\lambda_{contact}$ , and $\lambda_{acc}$ are hyper-parameters controlling the weights of the corresponding losses. We set $\lambda_{joint} = 1$ , $\lambda_{vel} = 2.5$ , $\lambda_{contact} = 10$ , and $\lambda_{acc} = 0.1$ empirically. The overall loss function is formulated as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s i m p l e}} + \lambda_ {\text {k i n}} \mathcal {L} _ {\text {k i n}} + \lambda_ {\text {b e a t}} \mathcal {L} _ {\text {b e a t}}, \tag {11} +$$ + +where $\lambda_{kin}$ and $\lambda_{beat}$ are hyper-parameters controlling the weights of the corresponding losses. $\lambda_{kin}$ and $\lambda_{beat}$ are set as 1 and 0.5, respectively. + +# 4 Experiments + +Dataset. Following previous works [35, 42], we choose the most widely used dataset, AIST++ [23] for both training and testing. This dataset comprises 1,408 dance sequences encompassing 10 distinct street dance genres. Each dance genre incorporates 6 musical pieces with different BPMs. $85\%$ of these sequences are categorized as basic choreographies, where performers execute the same choreography across all 6 musical pieces, each featuring varying tempos to synchronize with distinct musical beats. With its high choreographic and beat diversity, this dataset is ideal for our controllable dance generation task. + +Implementation Details. For the hierarchical multi-condition fusion module, we implement six sparse-dense fusion blocks and two dense-dense fusion blocks, utilizing a base dilation step $s$ set to 4, 8, 12, 16, 20, 24. For diffusion + +process, we adopt the cosine schedule as beat schedule and set diffusion steps $t$ as 1000 diffusion following [41]. The classifier-free guidance scale is set to 2. The diffusion module aligns with a transformer-based architecture similar to [42]. We adopt Adam optimizer with the learning rate set as 2e-4 and the training batch size is set to 64. During training, we randomly sample between $1\%$ and $30\%$ of the ground truth (GT) motion frames as the keyframe conditions. This training strategy can effectively improve the generalization capability of our method on different numbers of keyframes. We use the GT motion beats as beat conditions. + +To ensure fair comparisons, we follow EDGE [42] to crop all data into 5-second clips at 30 frames per second, with 2.5 seconds of overlapping. During testing, we randomly sample $10\%$ of keyframes from unpaired samples in the testing set as keyframe conditions and subsets of testing musical beats as beat conditions to prevent data leakage. The musical beats used for experiments are extracted by the off-the-shelf audio toolkit Librosa [17]. All experiments are conducted on 4 NVIDIA RTX3090 GPUs. + +Evaluation Metrics. For quantitative evaluation, we measure the generated dance from three aspects: generation quality, diversity, and controllability. Due to the limited samples in the testing set of AIST++ [23], the prior work EDGE [42] has demonstrated that the Fréchet Inception Distance (FID) is not reliable for measuring the generation quality. Therefore, we do not adopt FID as the evaluation metric. Instead, we assess generation quality from two different perspectives: music-dance correlation and kinematic plausibility. In terms of music-dance correlation, we utilize the Beat Alignment Score (BAS) metric, following the approach presented in [35]. BAS quantifies the synchronization quality between the generated dance and the musical beat. For the evaluation of kinematic plausibility, we adopt the Physical Foot Contact score (PFC) metric proposed in [42]. PFC provides a metric for measuring the physical plausibility of the generated dance from a kinematic perspective. As for the diversity, we measure the average feature distance of kinetic $(\mathrm{Div}_k)$ and geometric $(\mathrm{Div}_g)$ features extracted by fairmotion [12] following the previous works [23, 35, 42]. + +For assessing controllability, we employ the Key Pose Distance (KPD) and Beat Assignment Precision (BAP) metrics to evaluate keyframe control and beat control, respectively. KPD quantifies the average mean square error of local joint positions in Cartesian space at the keyframes. To evaluate beat control, we measure the alignment between the generated dance and the designated beat condition. BAP is defined as the percentage of beats in the generated dance that are correctly assigned to the designated beat frames. Note that this metric differs from the BAS, as a dance sequence with a high BAS does not necessarily adhere to the specific motion beat choreography constraint we intend to impose. + +# 4.1 Comparison to Existing Methods + +Currently, only three music-to-dance methods are publicly accessible: EDGE [42], Bailando [35], and FACT [23]. Of these, EDGE [42] is most directly aligned with our approach, as it similarly accommodates keyframes as inputs for generating dance sequences. We therefore assess EDGE using the same keyframe conditions + +Table 1: Quantitative comparisons among different methods on AIST++. Bold indicates best result. ↓ means lower is better, ↑ means higher is better and → means closer to the ground truth is better. + +
MethodsQualityDiversityControllability
PFC ↓ BAS ↑Divk→Divm→KPD ↓ BAP ↑
Ground Truth1.3380.3849.7737.212--
FACT [23]2.6980.2029.7047.342--
Bailando [35]1.5780.2159.6227.175--
EDGE [42](keyframes)1.0840.2359.7437.2740.859-
Ours(beat & keyframes)0.9660.6619.6607.2480.3060.793
+ +as our method to ensure a fair comparison. Conversely, certain approaches, like DanceFormer [21], are not open-source, thus hindering any direct evaluative comparison. We use the original test split of AIST++ [23] for evaluation, which has no overlap with the training split in terms of both the music and dance choreography. + +Table 2: Results of user study. + +
User StudyBeat-It Win Rate
FACT92.2%
Bailando78.8%
EDGE +(keyframes)QualityControllability +(Keyframe)
60.3%86.9%
+ +Table 3: Ablation study on AIST++. + +
AblationsQualityControllability
PFC ↓BAS ↑KPD ↓BAP ↑
w/o HF25.6260.3220.4770.323
w/o BD1.6320.3580.3890.371
w/o Lbeat1.3420.3970.3430.411
Ours0.9660.6610.3060.793
+ +Quantitative Comparisons. The quantitative results are shown in Tab. 1. As is revealed, our method surpasses all the competitors in terms of generation quality, diversity, and controllability. Particularly, our method shows considerable advantages over the compared methods for BAS metrics, notably with an improvement of 0.426 compared with the SOTA method EDGE [42]. This demonstrates the effectiveness of our method in generating beat-synchronized dance sequences while maintaining diversity. Furthermore, in the aspect of controllability, our method also significantly excels EDGE [42] with a notable improvement of 0.553 in KPD. Fig. 4 shows the mean joint velocities over time for a specific dance motion concerning both the GT and the existing methods. Notably, compared with others, our approach demonstrates more motion beats (local minimum point of velocities) precisely at the beat frames. This contributes to a better visual alignment with the audio when compared to the other methods. The BAP further validates our method's superior ability to produce dance sequences precisely aligned with the beat condition. Due to limited space, we present additional quantitative experiments in the supplementary materials. + +User Study. We also conduct a user study to assess the generation quality and controllability of our method. We randomly select 20 generated dance sequences + +![](images/4260c2e47c89f273f0e0a170133747059998e2458f470ae96ed7d18041a0ef2f.jpg) +Fig. 4: Visualization comparison on beat alignment among different methods. The motion generated by our method shows precise beat alignment with the given beat condition, demonstrating the superiority of our method in beat control. + +from the testing set and ask 18 participants to pick the best one among our method and the other compared methods. In particular, participants are required to choose the superior dance according to several criteria, namely, the overall visual plausibility of movements, synchronization with the dance beats, and keyframe controllability. The results of the user study are tabulated in Tab. 2. Our method shows noticeable advantages in terms of generation quality compared to other approaches, particularly in beat alignment. For qualitative evaluations on keyframe-conditioned controllability, we also provide the participants with the keyframe references for intuitive comparison between our method against EDGE [42]. The results show that our approach owns better motion controllability than EDGE [42]. + +# 4.2 Ablation Studies + +To get a more comprehensive insight into our main contributions, we conduct ablation studies on the hierarchical multi-condition fusion module, the beat-aware dilation scheme, and the beat alignment loss to evaluate their efficacy. The quantitative results are shown in Tab. 3. The qualitative results can be also found in the supplementary video. + +Hierarchical Multi-condition Fusion Module. To verify the significance of the hierarchical multi-condition fusion module, we remove it by simply concatenating all the conditions and feeding them into a one-stage encoder (w/o HF). As shown in Tab. 3, the simple one-stage fusion module can lead to inferior results in both quality and controllability. This is because the direct concatenation of multiple conditions can lead to inevitable condition conflicts, giving rise to suboptimal optimization of the model. Even worse, the naive padding of the sparse keyframe condition introduces severe noises, further complicating the learning process and resulting in unsatisfactory generation performance. By decently fusing multiple conditions gradually, our hierarchical multi-condition fusion mechanism effectively mitigates the conflicts among different conditions and yields more informative conditions, allowing for higher-quality generation with better controllability. + +Beat-aware Dilation Scheme. To validate the effectiveness of our beat-aware dilation scheme, we replace it with a vanilla masked attention strategy (w/o + +BD). As shown in Tab. 3, the absence of the beat-aware dilation scheme notably impairs both keyframe and beat choreography controllability. The reason behind this is that beat-aware dilation can dynamically adjust the expansion degree of the attention mask, contingent on the distance between the keyframe and the beat frames. This tactical approach significantly improves the synchronization between keyframes and their associated beat frames, thereby empowering the model to more effectively harness prior knowledge of beats. Consequently, this enhancement considerably bolsters beat control while augmenting keyframe controllability. + +Beat Alignment Loss. To further validate the effectiveness of the beat alignment loss, we assess the model's performance by comparing it to its counterpart without it (w/o $\mathcal{L}_{\text {beat }}$ ). As shown in Tab. 3, the ablated variant without the beat alignment loss demonstrated performance degradation in beat choreography, showcasing decreases of 0.264 and 0.382 in BAS and BAP, respectively. This decline in performance can be attributed to the advantageous impact of direct supervision during training, particularly in guiding the model to comprehend beat-level dance choreography through the proposed beat alignment loss. Consequently, this enhancement significantly improves the synchronization between motion sequences and designated beats. + +# 5 Discussion and Conclusion + +In this work, we propose a novel multi-condition diffusion-based framework, BeatIt, for beat-synchronized and key pose-guided dance generation. The proposed framework presents a hierarchical multi-condition fusion mechanism equipped with a beat-aware mask dilation scheme to integrate conditions with different information sparsity. To achieve precise beat synchronization, we explicitly disentangle the beats from the music and inject beat controllability throughout the entire generation process. Additionally, we introduce a specifically designed beat alignment loss to provide explicit guidance and supervision on motion beats. Both qualitative and quantitative experimental results on the AIST++ dataset validate the superiority of our method in producing high-quality dance sequences with precise beat synchronization and flexible keyframe control. Further limitations and broader impact are discussed in supplementary materials. + +# Acknowledgements + +The work is supported by China National Key R&D Program (No. 2023YFE0202700), Key-Area Research and Development Program of Guangzhou City (No. 2023B01J0022), Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (No. 2024B1515040010), Health and Medical Research Fund (HMRF) of Hong Kong Health Bureau (No. 10211516), Innovation and Technology Fund - Innovation and Technology Support Programme (ITF-ITSP) (No. ITS/127/22), Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2023B1515020097), National Research Foundation Singapore under the AI Singapore Programme (No. AISG3-GV-2023-011). + +# References + +1. Alemi, O., Françoise, J., Pasquier, P.: Groovenet: Real-time music-driven dance movement generation using artificial neural networks. networks 8(17), 26 (2017) 2, 4 +2. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM TOG 42(4), 1-20 (2023) 4 +3. Ao, T., Zhang, Z., Liu, L.: Gesturediffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613 (2023) 4 +4. Araújo, J.P., Li, J., Vetrivel, K., Agarwal, R., Wu, J., Gopinath, D., Clegg, A.W., Liu, K.: Circle: Capture in rich contextual environments. In: CVPR. pp. 21211-21221 (2023) 3 +5. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR. pp. 9760-9770 (2023) 3, 4 +6. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. NeurIPS 34, 19822-19835 (2021) 4 +7. Ferreira, J.P., Coutinho, T.M., Gomes, T.L., Neto, J.F., Azevedo, R., Martins, R., Nascimento, E.R.: Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics 94, 11-21 (2021) 4 +8. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV. pp. 89-106. Springer (2022) 4 +9. Gao, X., Hu, L., Zhang, P., Zhang, B., Bo, L.: Dancemeld: Unraveling dance phrases with hierarchical latent codes for music-to-dance synthesis. arXiv preprint arXiv: 2401.10242 (2023) 4 +10. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: Imos: Intent-driven full-body motion synthesis for human-object interactions. In: CGF. vol. 42, pp. 1-12. Wiley Online Library (2023) 3 +1. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: ICCV. pp. 9942-9952 (2023) 4 +2. Gopinath, D., Won, J.: fairmotion - tools to load, process and visualize motion capture data. Github (2020), https://github.com/facebookresearch/fairmotion11 +3. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020) 5, 6 +4. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. ACM TOG 35(4), 1-11 (2016) 4 +5. Huang, R., Hu, H., Wu, W., Sawada, K., Zhang, M., Jiang, D.: Dance revolution: Long-term dance generation with music via curriculum learning. arXiv preprint arXiv:2006.06119 (2020) 2, 4 +6. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR. pp. 16750-16761 (2023) 3, 4 +7. Jin, Y., Zhang, J., Li, M., Tian, Y., Zhu, H., Fang, Z.: Towards the automatic anime characters creation with generative adversarial networks. arXiv preprint arXiv:1708.05509 (2017) 11 + +18. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI. vol. 37, pp. 8255-8263 (2023) 3, 4 +19. Kim, J., Oh, H., Kim, S., Tong, H., Lee, S.: A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In: CVPR. pp. 3490-3500 (2022) 4 +20. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. NeurIPS 32 (2019) 4 +21. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI. vol. 36, pp. 1272-1279 (2022) 2, 4, 12 +22. Li, J., Yin, Y., Chu, H., Zhou, Y., Wang, T., Fidler, S., Li, H.: Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171 (2020) 2, 4 +23. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV. pp. 13401-13412 (2021) 2, 4, 7, 10, 11, 12 +24. Lin, J., Chang, J., Liu, L., Li, G., Lin, L., Tian, Q., Chen, C.w.: Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In: CVPR. pp. 23222-23231 (2023) 3 +25. Liu, H., Xu, C., Yang, Y., Zeng, L., He, S.: Drag your noise: Interactive point-based editing via diffusion semantic propagation. In: CVPR. pp. 6743-6752 (2024) 4 +26. Lu, X., Ma, C., Shen, J., Yang, X., Reid, I., Yang, M.H.: Deep object tracking with shrinkage loss. IEEE TPAMI 44(5), 2386-2401 (2020) 9 +27. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) 5 +28. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) 4, 5 +29. Qi, Q., Zhuo, L., Zhang, A., Liao, Y., Fang, F., Liu, S., Yan, S.: Diffdance: Cascaded human motion diffusion model for dance generation. In: ACM MM. pp. 1374-1382 (2023) 4 +30. Qin, C., Zhang, S., Yu, N., Feng, Y., Yang, X., Zhou, Y., Wang, H., Niebles, J.C., Xiong, C., Savarese, S., et al.: Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147 (2023) 5 +31. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML. pp. 8821-8831. PMLR (2021) 4 +32. Ren, X., Li, H., Huang, Z., Chen, Q.: Self-supervised dance video synthesis conditioned on music. In: ACM MM. pp. 46-54 (2020) 4 +33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) 4 +34. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS 35, 36479-36494 (2022) 4 +35. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR. pp. 11050-11059 (2022) 2, 4, 7, 10, 11, 12 +36. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando++: 3d dance gpt with choreographic memory. IEEE TPAMI (2023) 4 + +37. Sun, G., Wong, Y., Cheng, Z., Kankanhalli, M.S., Geng, W., Li, X.: Deepdance: music-to-dance motion choreography with adversarial learning. IEEE TMM 23, 497-509 (2020) 4 +38. Sun, J., Wang, C., Hu, H., Lai, H., Jin, Z., Hu, J.F.: You never stop dancing: Non-freezing dance generation via bank-constrained manifold projection. NeurIPS 35, 9995-10007 (2022) 4 +39. Tang, T., Jia, J., Mao, H.: Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In: ACM MM. pp. 1598-1606 (2018) 2, 4 +40. Tang, Z., Yang, Z., Zhu, C., Zeng, M., Bansal, M.: Any-to-any generation via composable diffusion. arXiv preprint arXiv: 2305.11846 (2023) 5 +41. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) 9, 11 +42. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: CVPR. pp. 448-458 (2023) 2, 4, 7, 9, 10, 11, 12, 13 +43. Valle-Pérez, G., Henter, G.E., Beskow, J., Holzapfel, A., Oudeyer, P.Y., Alexanderson, S.: Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM TOG 40(6), 1-14 (2021) 2, 4 +44. Wang, T., Zhang, T., Zhang, B., Ouyang, H., Chen, D., Chen, Q., Wen, F.: Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952 (2022) 4 +45. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS 35, 14959-14971 (2022) 3 +46. Wei, D., Sun, X., Sun, H., Li, B., Hu, S., Li, W., Lu, J.: Understanding text-driven motion synthesis with keyframe collaboration via diffusion models. arXiv preprint arXiv:2305.13773 (2023) 7 +47. Xie, Y., Lin, Y., Cai, W., Xu, X., Zhang, H., Du, Y., He, S.: D3still: Decoupled differential distillation for asymmetric image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17181-17190 (2024) 4 +48. Xu, C., Chen, Z., Mai, J., Xu, X., He, S.: Pose-and attribute-consistent person image synthesis. ACM TOMM 19(2s), 1-21 (2023) 4 +49. Xu, C., Li, K., Luo, X., Xu, X., He, S., Zhang, K.: Fully deformable network for multiview face image synthesis. IEEE TNNLS (1-15) (2022) 4 +50. Xu, C., Xu, Y., Zhang, H., Xu, X., He, S.: Dreamanime: Learning style-identity textual disentanglement for anime and beyond. IEEE TVCG (2024) 4 +51. Yalta, N., Watanabe, S., Nakadai, K., Ogata, T.: Weakly-supervised deep recurrent neural networks for basic dance step generation. In: IJCNN. pp. 1-8. IEEE (2019) 4 +52. Yang, Z., Wen, Y.H., Chen, S.Y., Liu, X., Gao, Y., Liu, Y.J., Gao, L., Fu, H.: Keyframe control of music-driven 3d dance generation. IEEE TVCG (2023) 2, 4 +53. Yin, W., Yin, H., Baraka, K., Kragic, D., Björkman, M.: Dance style transfer with cross-modal transformer. In: WACV. pp. 5058-5067 (2023) 4 +54. Yu, Y., Liu, B., Zheng, C., Xu, X., Zhang, H., He, S.: Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In: CVPR. pp. 7109-7118 (2024) 4 +55. Zhang, C., Tang, Y., Zhang, N., Lin, R.S., Han, M., Xiao, J., Wang, S.: Bidirectional autoregressive diffusion model for dance generation. arXiv preprint arXiv: 2402.04356 (2024) 4 + +56. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023) 3 +57. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023) 5 +58. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023) 5 +59. Zheng, C., Liu, B., Xu, X., Zhang, H., He, S.: Learning an interpretable stylized subspace for 3d-aware animatable artforms. IEEE TVCG (2024) 4 +60. Zheng, C., Liu, B., Zhang, H., Xu, X., He, S.: Where is my spot? few-shot image generation via latent subspace optimization. In: CVPR. pp. 3272-3281 (2023) 4 +61. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: CVPR. pp. 5632-5641 (2023) 3, 4 +62. Zhu, L., Liu, X., Liu, X., Qian, R., Liu, Z., Yu, L.: Taming diffusion models for audio-driven co-speech gesture generation. In: CVPR. pp. 10544-10553 (2023) 4 \ No newline at end of file diff --git a/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/images.zip b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2998cf570659090185c6f0f1ccd7375ea35a6bad --- /dev/null +++ b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:812fcc08b6ba3bb6273a2c04d892ad31512ec0dbd9291f892547037c3fbaaaf5 +size 318981 diff --git a/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/layout.json b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf921c47088fd9c3ed75e3a5d537b4eda5effacb --- /dev/null +++ b/2024/Beat-It_ Beat-Synchronized Multi-Condition 3D Dance Generation/layout.json @@ -0,0 +1,9290 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 146 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 146 + ], + "type": "text", + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": "Zikai Huang" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Xuemiao " + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{1,3,4(\\boxtimes)}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Cheng " + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{2(\\boxtimes)}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Huaidong Zhang" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Chenxi Zheng" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Jing Qin" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": ", and Shengfeng He" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 208, + 201, + 403, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 201, + 403, + 223 + ], + "spans": [ + { + "bbox": [ + 208, + 201, + 403, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 208, + 201, + 403, + 223 + ], + "type": "text", + "content": " South China University of Technology, China xuemx@scut.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 224, + 442, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 224, + 442, + 245 + ], + "spans": [ + { + "bbox": [ + 170, + 224, + 442, + 245 + ], + "type": "text", + "content": "2 The Hong Kong Polytechnic University, Hong Kong SAR, China \ncschengxu@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 245, + 457, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 245, + 457, + 257 + ], + "spans": [ + { + "bbox": [ + 170, + 245, + 457, + 257 + ], + "type": "text", + "content": "Guangdong Engineering Center for Large Model and GenAI Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 257, + 470, + 300 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 142, + 257, + 470, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 257, + 470, + 277 + ], + "spans": [ + { + "bbox": [ + 142, + 257, + 470, + 277 + ], + "type": "text", + "content": "4 Guangdong Provincial Key Lab of Computational Intelligence and Cyberspace Information" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 209, + 278, + 404, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 278, + 404, + 300 + ], + "spans": [ + { + "bbox": [ + 209, + 278, + 404, + 300 + ], + "type": "text", + "content": "5 Singapore Management University, Singapore https://zikaihuangscut.github.io/Beat-It/" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 331, + 482, + 439 + ], + "blocks": [ + { + "bbox": [ + 133, + 331, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 133, + 331, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 331, + 482, + 439 + ], + "type": "image", + "image_path": "db6755f06941388ecccbfa2bb7c69e65b47f8c6fd7368897d48248abded99334.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 446, + 483, + 492 + ], + "lines": [ + { + "bbox": [ + 130, + 446, + 483, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 483, + 492 + ], + "type": "text", + "content": "Fig. 1: We introduce Beat-It, a novel method for generating 3D dance motions with beat alignment and motion controllability. Our approach explicitly injects beat awareness and seamlessly integrates multiple conditions to guide the generation process, leading to beat-synchronized, key pose-guided dance generation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 533, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 533, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 533, + 455, + 666 + ], + "type": "text", + "content": "Abstract. Dance, as an art form, fundamentally hinges on the precise synchronization with musical beats. However, achieving aesthetically pleasing dance sequences from music is challenging, with existing methods often falling short in controllability and beat alignment. To address these shortcomings, this paper introduces Beat-It, a novel framework for beat-specific, key pose-guided dance generation. Unlike prior approaches, Beat-It uniquely integrates explicit beat awareness and key pose guidance, effectively resolving two main issues: the misalignment of generated dance motions with musical beats, and the inability to map key poses to specific beats, critical for practical choreography. Our approach disentangles beat conditions from music using a nearest beat distance representation and employs a hierarchical multi-condition fusion mechanism. This mechanism" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "text", + "content": "seamlessly integrates key poses, beats, and music features, mitigating condition conflicts and offering rich, multi-conditioned guidance for dance generation. Additionally, a specially designed beat alignment loss ensures the generated dance movements remain in sync with the designated beats. Extensive experiments confirm Beat-It's superiority over existing state-of-the-art methods in terms of beat alignment and motion controllability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 193, + 453, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 193, + 453, + 216 + ], + "spans": [ + { + "bbox": [ + 160, + 193, + 453, + 216 + ], + "type": "text", + "content": "Keywords: Dance generation " + }, + { + "bbox": [ + 160, + 193, + 453, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 193, + 453, + 216 + ], + "type": "text", + "content": " Beat synchronization " + }, + { + "bbox": [ + 160, + 193, + 453, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 193, + 453, + 216 + ], + "type": "text", + "content": " Multi-condition diffusion generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 234, + 230, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 234, + 230, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 234, + 230, + 247 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 482, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 330 + ], + "type": "text", + "content": "Dance, a time-honored art form, resonates across various ages and cultures due to its artistic and aesthetic significance, playing a vital role in the cultural and entertainment sectors. Traditional dance choreography creation, which demands an intricate balance of aesthetic movements, emotional expression, and precise synchronization with musical beats, often presents itself as a costly, laborious, and time-intensive endeavor, even for seasoned artists." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 437 + ], + "type": "text", + "content": "Recent advances in deep learning, particularly in generative models, have catalyzed efforts to automate the choreography process. Initial approaches predominantly focused on single-condition dance generation, learning a direct mapping from music to dance motions [1, 15, 21-23, 35, 39, 43]. The advent of diffusion models, renowned for their exceptional content generation capabilities, has marked a paradigm shift in this domain. For instance, Tseng et al. [42] introduced a transformer-based diffusion model that sets a new benchmark in music-to-dance conversion, showcasing the potential of these models in elevating the quality of generated dance sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 438, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 617 + ], + "type": "text", + "content": "Notwithstanding the remarkable progress, achieving precise synchronization between dance motions and musical beats, coupled with flexible motion controllability in dance generation, remains a formidable challenge. In real-world choreography, the goal is often to assign specific key poses to particular musical beats and then fill in the connecting movements to form the complete dance sequence. This process requires not only the creation of designated key poses but also their harmonious synchronization with the musical beats. While few methods have explored key pose conditions [42, 52], none of them has considered explicit beat synchronization and controllability, rendering inferior dance motion generation. Furthermore, previous methods typically merge keyframe features with music features through simple concatenation or temporal/spatial blending. This approach, however, is flawed due to the sparse nature of keyframes. Consequently, these methods disproportionately emphasize dense music features at the expense of key pose conditions, leading to a notable misalignment between the generated dance motions and the specified keyframes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "To combat the above issues, we aim to completely disentangle beats from input music and simultaneously incorporate explicit beat and key pose guidance to enable beat-controllable, key pose-guided dance generation. We introduce BeatIt, a diffusion-based, multi-condition framework designed for this purpose (see" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Z. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "Fig. 1). Our approach begins by formulating beat condition as a nearest beat distance representation, which is a straightforward yet effective method to serve as beat guidance within our framework. To effectively harness and balance the multiple conditions involved, we propose a hierarchical multi-condition fusion mechanism. Initially, we integrate sparse key pose condition with dense beat and music conditions using a tailored beat-aware dilated cross-attention strategy. We further blend refined beat and music features to generate comprehensive multi-condition features. This strategy not only enhances beat and motion controllability in dance generation but also maintains the realism and coherence of the produced dance motions. Additionally, leveraging our beat representation, we introduce a novel beat alignment loss, explicitly ensuring synchronization between dance motions and the given beat condition, thereby significantly enhancing the overall quality of generation. Extensive experiments confirm that our method outperforms current state-of-the-art approaches in terms of beat alignment and motion controllability. Beyond the advantages above, our framework also supports arbitrary beat designation and flexible frame assignments of key poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 308, + 364, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 308, + 364, + 319 + ], + "spans": [ + { + "bbox": [ + 146, + 308, + 364, + 319 + ], + "type": "text", + "content": "In summary, our main contributions are fourfold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 327, + 481, + 492 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 327, + 481, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 481, + 373 + ], + "type": "text", + "content": "- We introduce a multi-condition dance generation framework that achieves beat synchronization and enhanced motion controllability. To the best of our knowledge, our method makes the first attempt at beat-controllable key pose-guided dance generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "text", + "content": "- We present a hierarchical multi-condition fusion mechanism to effectively suppress the conflicts and fully exploit the complementary information among different conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 411, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 411, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 137, + 411, + 480, + 456 + ], + "type": "text", + "content": "- We delve into the property of beats and formulate it as a nearest beat distance representation. A beat alignment loss is further tailored to offer explicit supervision signals to the generated dance motions, largely elevating the synchronization between the generated motions and the given beat conditions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 458, + 480, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 458, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 137, + 458, + 480, + 492 + ], + "type": "text", + "content": "- Extensive experiments demonstrate our method performs favorably against the state-of-the-art approaches, especially in motion controllability and motion-beat alignment." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 513, + 237, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 237, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 237, + 525 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 538, + 297, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 297, + 549 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 297, + 549 + ], + "type": "text", + "content": "2.1 Human Motion Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Human motion generation aims to generate realistic human motion automatically. Previous works can be generally categorized into three genres according to the input condition: text-conditioned [5,18,24,56,61], audio-conditioned, and scene-conditioned [4, 10, 16, 45]. Our work belongs to audio-conditioned human motion generation. But different from the primary audio-conditioned human motion generation, music-to-dance generation requires a comprehensive consideration of aesthetic movements, emotional expression, and precise synchronization with musical beats, making it extremely challenging to render convincing results. Recently, researchers have widely explored the application of diffusion models" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "in human motion generation [2,3,5,16,18,61,62], which shows unprecedented performance than traditional generative models. In this paper, we get a deep insight into the beat synchronization problem and investigate a multi-condition diffusion-based scheme for controllable dance generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 183, + 253, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 253, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 253, + 194 + ], + "type": "text", + "content": "2.2 Dance Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 483, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 483, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 483, + 589 + ], + "type": "text", + "content": "With the emergence of deep learning [47-49], dance generation has attracted lots of research interest in recent years. Previous methods have explored various frameworks with different types of backbones to achieve single-condition dance generation, including CNN [14], RNN [1, 15, 39, 51], GCN [7, 32], VAE [9, 38], GAN [19, 20, 37], and Transformer [21-23, 35, 36, 43, 53]. FACT [23] proposes a full-attention cross-modal transformer model to capture the correlations between music and dance. Danceformer [21] presents a two-stage deterministic framework for audio-driven single-condition dance generation. It first generates key poses from the input music and then performs interpolation between the key poses. However, it falls short of arbitrary beat choreography and flexible key pose controllability due to its single-condition nature. Bailando [35] utilizes separate VQ-VAEs on upper/lower half bodies and a motion GPT to map the music and seed poses to dance sequences. Several recent attempts have investigated the diffusion models for dance generation [29, 42, 55], which significantly pushes the boundaries of music-to-dance performance. Compared to applying only a single input condition, multi-condition dance generation incorporates additional conditions beyond music into the generation process for better controllability, such as dance style labels [2], text [11], and keyframes [42, 52]. For example, LDA [2] introduced dance style labels as auxiliary conditions. However, it is limited to generating pre-defined styles and lacks flexibility. TM2D [11] adopts a VQ-VAE based model to support dual-modal music and text-driven dance motion generation. Yang et al. [52] employ normalizing flows to model dance motion probability distributions and utilize time embedding to achieve keyframe control. EDGE [42] adopts a diffusion-based framework and enables key pose control via temporal/spatial blending during inference. Although the existing methods have investigated additional conditions for better dance generation control, they typically neglect explicit guidance and supervision to align the generated motion dances with specified beats. Consequently, these methods struggle to produce both beat-specific and key pose-guided dance sequences. In contrast, we propose to explicitly disentangle the beat condition from music and inject flexible beat and key pose controllability into the entire generation process, significantly enhancing the beat alignment while supporting flexible frame assignments of key poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 608, + 353, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 608, + 353, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 608, + 353, + 619 + ], + "type": "text", + "content": "2.3 Multi Condition Diffusion Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Multi-condition diffusion generative models seek to inject multi-condition guidance into the generation process for versatile controllability [6,8,25,28,31,34, 44,50,54,59,60]. Benefiting from large-scale training data, Stable Diffusion [33]" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Z. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "constructs a powerful foundation for image generation. Building upon this, ControlNet [57] and T2I-Adapter [27] efficiently integrate trainable parameters into existing models, enabling the inclusion of spatially localized input conditions in a pre-trained text-to-image diffusion model. Uni-ControlNet [58] proposes a framework equipped with two additional control modules for multi-condition control. Qin et al. [30] present a unified model capable of handling various visual conditions, resulting in a more streamlined scheme. Codi [40] acquire multi-modal synthesis skills through the mapping of diverse modalities into a unified space, training on lists of different multi-modal generation tasks. The aforementioned works typically controllable generation using pre-trained models under diverse conditions. Our work, conversely, focuses on the effective and comprehensive fusion of multiple conditions by presenting a hierarchical multi-condition fusion mechanism aimed at better integrating sparse keyframe control conditions with dense music and beat control conditions, resulting in more controllable dance generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 314, + 202, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 202, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 202, + 327 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 340, + 257, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 340, + 257, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 257, + 351 + ], + "type": "text", + "content": "3.1 Problem Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": "Dance generation aims to create realistic dance motion sequences " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathbf{x}^1, \\mathbf{x}^2, \\mathbf{x}^3, \\dots, \\mathbf{x}^L\\}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": " from a given piece of music " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": " with a duration of " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^i \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": " represents a human pose at the " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": "-th frame, which is denoted as a " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": "-dimensional vector. In this paper, we endeavor to disentangle beat condition from the input music and simultaneously incorporate explicit beat and key pose guidance to achieve beat-synchronized, key pose-guided dance generation. Given a music condition " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": ", a sparse set of keyframes " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{ref} = \\mathcal{X} \\odot \\mathbf{M}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\{0, 1\\}^L" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": " is a temporal binary mask that specifies the assigned locations of the keyframes, and a beat condition " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": ", our goal is to generate a beat-specific, and key pose-guided dance motion sequence " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{X}} = \\{\\hat{\\mathbf{x}}^1, \\hat{\\mathbf{x}}^2, \\hat{\\mathbf{x}}^3, \\dots, \\hat{\\mathbf{x}}^L\\}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": " that conforms to the music " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": ", consistent with reference frames " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{ref}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": " at positions specified by " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": ", and synchronized with the beat condition " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 522, + 228, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 228, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 228, + 533 + ], + "type": "text", + "content": "3.2 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 542, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 627 + ], + "type": "text", + "content": "In this work, we adopt diffusion model [13, 28] as the backbone of our framework. Diffusion models establish a consistent Markovian forward process that incrementally introduces noise into clean sample data " + }, + { + "bbox": [ + 130, + 542, + 482, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{1:L} \\in q(\\mathbf{x}_0)" + }, + { + "bbox": [ + 130, + 542, + 482, + 627 + ], + "type": "text", + "content": ", and a corresponding reverse process that progressively eliminates noise from noisy samples. For brevity, we use " + }, + { + "bbox": [ + 130, + 542, + 482, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 542, + 482, + 627 + ], + "type": "text", + "content": " to represent the entire sequence. During the forward process, a predefined noise variance schedule " + }, + { + "bbox": [ + 130, + 542, + 482, + 627 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 130, + 542, + 482, + 627 + ], + "type": "text", + "content": " is employed to regulate the noise increments at each step. The forward process can be formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 635, + 482, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 635, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 160, + 635, + 482, + 669 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}", + "image_path": "0baab420dec633cefe5e6be33d3d5f6c4b7dfc73e36798650d1399c31ed7b451.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 482, + 266 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 482, + 266 + ], + "type": "image", + "image_path": "d98af7a52b18e9bf64aa37787f267782767498df9ff1a0e7be5db58eb38f0032.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 276, + 483, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 276, + 483, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 483, + 354 + ], + "type": "text", + "content": "Fig.2: Overview of our proposed method, Beat-It. We generate a beat-synchronized dance sequence utilizing music, keyframes, and beat conditions. Conditional embeddings are derived and subsequently fused in a two-stage process: initially integrating sparse keyframe condition with other dense conditions, followed by the fusion of these dense conditions. The final fused condition is then processed by the conditional diffusion module. To ensure precise beat control, a beat alignment loss is employed to explicitly supervise the generated motions at the beat level." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": "After " + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": " steps, the sample data progressively transforms into a noise distribution " + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_T)" + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": ", which is usually a standard Gaussian distribution " + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": ". In the reverse process, the noise is gradually removed from the noisy sample " + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": " to produce the clean sample " + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": ". In our task, multi-conditional dance generation aims to model the distribution " + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0|\\mathbf{C})" + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": " with a set of conditions " + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": ". Following Ho et al. [13], we directly predict the clean sample " + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": " from the noise distribution " + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_T)" + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": " with the following objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 199, + 468, + 482, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 468, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 199, + 468, + 482, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} | \\mathbf {C}), t \\sim [ 1, T ]} \\left[ \\| \\mathbf {x} _ {0} - G (\\mathbf {x} _ {t}, t, \\mathbf {C}) \\| _ {2} ^ {2} \\right]. \\tag {2}", + "image_path": "9412c7c9974cb7bd87a63997500a3e45b99b768f010e424114bcaa48db4b6cef.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "content": "A spatial or temporal constraint " + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^c" + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "content": " with positions specified by a binary mask " + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "content": " can be directly incorporated into the denoising process without additional training. This can be achieved by simply performing the following operation at every timestep:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 539, + 482, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 539, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 242, + 539, + 482, + 552 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = \\mathbf {x} _ {t} ^ {c} \\odot m + \\mathbf {x} _ {t} \\odot (1 - m), \\tag {3}", + "image_path": "8aab81cac7e47062a0d96df6610fa8aa775f328aa29438a58cd76417d438dc00.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 556, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 556, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 556, + 482, + 605 + ], + "type": "text", + "content": " denotes the Hadamard product. By doing so, regions under constraint can be replaced with forward-diffused samples of the constraint, while the remaining regions are left unchanged. However, it is worth noting that this simplistic approach tends to render inferior results when given sparse constraints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 622, + 209, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 209, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 209, + 633 + ], + "type": "text", + "content": "3.3 Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "The pipeline of our proposed model is illustrated in Fig. 2. To begin with, we feed the music condition " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", keyframe condition " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{ref}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", and beat condition " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " into" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Z. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "three encoders respectively: the music encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", keyframe encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", and beat encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_b" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". This process yields the corresponding music embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_m" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", keyframe embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", and beat embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_b" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_m, \\mathbf{e}_r, \\mathbf{e}_b \\in \\mathbb{R}^{L \\times d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". Then, we forward these condition embeddings to a hierarchical multi-condition fusion module to produce comprehensive multi-condition features " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", which are finally sent to the conditional diffusion denoising module to render the denoised dance motion sequence " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{X}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 217, + 264, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 264, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 264, + 228 + ], + "type": "text", + "content": "3.4 Beat Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "text", + "content": "As mentioned in previous studies [23, 35, 42], there exists a strong consistency between musical beats and dance motion beats. In practical dance choreography, the arrangement of motion beats is relatively flexible and does not require a strict point-to-point alignment with musical beats. Therefore, we introduce an independent beat condition disentangled from the music, to allow for flexible motion beat controllability in dance generation. For an input music sequence of duration " + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "text", + "content": ", each frame can be categorized as either a beat or non-beat frame. A straightforward way to represent the beat condition is using a binary mask, where 1 and 0 denote the beat and non-beat frames respectively. However, such a simple binary presentation exhibits high sparsity as the beat frames only account for a small portion of the whole music sequence. This makes it less informative and, at times, perceived as noise or disregarded by the model. As an alternative, we propose to represent the beat condition as a vector " + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "text", + "content": ", with each entry " + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "inline_equation", + "content": "b^i \\in \\mathbb{N}" + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "text", + "content": " denoting the distance between the current frame and the nearest beat frame. This representation not only mitigates the sparsity caused by binary formulations but also provides the model with local temporal context. This significantly facilitates the precise acquisition of rhythmic characteristics within the choreography and injects more effective beat controllability into the learning process. The beat embedding " + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_b \\in \\mathbb{R}^{L \\times d}" + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "text", + "content": " is obtained using a separate embedding layer, followed by the transformer-based encoder " + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_b" + }, + { + "bbox": [ + 130, + 235, + 482, + 486 + ], + "type": "text", + "content": ". Please refer to the supplementary materials for detailed representations of music and dance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 502, + 344, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 344, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 344, + 514 + ], + "type": "text", + "content": "3.5 Hierarchical Multi-Condition Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Although the keyframe constraint offers precise pose specifications for specific frames, it is still too sparse which significantly increases the learning difficulty of the model. Naively incorporating this condition with others introduces a vast of padding, leading to excessive noises that may impair the effectiveness of keyframe constraints. By contrast, music condition and beat condition are much more denser, with valid values for each frame. To harmoniously combine these three conditions, we introduce a novel multi-condition hierarchical fusion mechanism." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Specifically, in the first stage, we inject sparse keyframe condition into other dense conditions through a beat-aware dilation attention strategy, which shares a similar spirit of DiffKFC [46]. We observe that dance movements on motion-beat frames are often more iconic in the whole sequence, distinguished by their higher emphasis and significance. As these movements typically serve as guiding" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 211 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 211 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 211 + ], + "type": "image", + "image_path": "bc014b8417fc5fb840cdaf4d584fa6004918cb17f28e891618e36e834d2824a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": "Fig. 3: Illustration of the beat-aware mask dilation scheme. The first row visualizes the keyframe mask, with deep green indicating valid control constraints and gray indicating invalid ones. The second row presents the dilation step curve with red lines marking beat frames. The third row is a heatmap of the dilation step. The fourth row shows the neighborhood range of keyframes, with light green indicating the expanded valid region from beat-aware mask dilation. The final row displays the beat-aware dilated keyframe mask." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": "elements, influencing the choreography and style of the entire dance, they exert a more substantial influence on the surrounding frames. To utilize this property, we design a beat-aware mask dilation scheme. Unlike DiffKFC which uses identical dilation steps for all keyframes, treating them uniformly, we opt for different dilation weights based on the distance between the keyframe and the nearest designated beat frame. The dilation step " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": " is calculated as " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "n = \\lceil s \\cdot e^{-2\\frac{b^i}{d^i}} \\rceil" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "b^i" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "d^i" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": " are the ground truth beat distance at frame " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": " and the distance between frame " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": "'s adjacent motion-beat frames, respectively. Base dilation step " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": " is a hyper-parameter controlling overall dilation speed. The closer the keyframe is to the nearest beat frame, the larger the dilation step. This strategy enables the model to gradually propagate the keyframe condition to the surrounding frames with beat awareness. The beat-aware mask dilation scheme is illustrated in Fig. 3. Given the keyframe binary mask " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": " and dilation step " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": ", the beat-aware dilated mask " + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^d" + }, + { + "bbox": [ + 130, + 326, + 483, + 497 + ], + "type": "text", + "content": " is obtained as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 509, + 482, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 509, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 159, + 509, + 482, + 541 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {d} [ i ] = \\left\\{ \\begin{array}{l l} \\max _ {j} \\mathbf {M} [ i - j ], j \\in \\{- n, n + 1, \\dots , n - 1, n \\} & \\text {i f} \\mathbf {M} _ {i} = 1, \\\\ \\mathbf {M} _ {i}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}", + "image_path": "6baa3fedce741cc6168d904b1a64612bb86359034265eccbf149607cc23cada2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": "The beat-aware dilated attention mask is then calculated as " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{attn} = \\mathbf{MM}_d^{\\mathsf{T}}" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": ". This beat-aware mask dilation scheme enables beat-aware keyframe condition propagation, thereby amplifying the synergy among conditions characterized by varying degrees of sparsity, thus significantly boosting the generation quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In the second phase, we concatenate two refined dense conditions, namely keyframe-fused music embedding and keyframe-fused beat embedding. These concatenated conditions are fed into a transformer condition fusion module, yielding the final fused features that can be subsequently sent to the conditional diffusion denoising block." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Z. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 266, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 266, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 266, + 128 + ], + "type": "text", + "content": "3.6 Beat Alignment Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "type": "text", + "content": "For controllable dance generation, it is crucial to ensure precise beat alignment between the generated dance motions and the given beat conditions. Current methods typically ignore the explicit constraints on the beats of the generated dance sequences. As a consequence, these approaches lack necessary beat controllability and struggle to guarantee beat synchronization between the generated dance motions and musical beats, which severely deteriorates the overall generation quality. To conquer this issue, building upon our beat representation, we present a beat alignment loss to provide explicit supervision that enforces the beats of the generated motions to precisely align with the given beat conditions. To this end, we first pre-train a beat distance estimator, which accepts a motion sequence as input and predicts the distance from the current frame to the nearest motion beat frame for each frame in the sequence. During the training process, the pre-trained beat distance estimator is responsible for providing supervision signals on the beats of the generated motions. The beat alignment loss " + }, + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{beat}" + }, + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "type": "text", + "content": " can be formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 323, + 481, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 323, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 235, + 323, + 481, + 355 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b e a t}} = \\sum_ {i = 1} ^ {L} w _ {s} ^ {i} \\cdot w _ {b} ^ {i} \\cdot \\mathbf {M S E} \\left(b ^ {i}, \\hat {b} ^ {i}\\right), \\tag {5}", + "image_path": "2c4e3c0d2ee9a33670ae6cb6a3c9fe86b5c0f7b2d3105fd35ae8f660eed68099.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "inline_equation", + "content": "w_{s}^{i} = \\frac{1}{1 + e^{a\\cdot(c - \\|b^{i} - \\hat{b}^{i}\\| / b^{i})}}" + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "inline_equation", + "content": "w_{b}^{i} = e^{-\\frac{2b^{i}}{d^{i}}}" + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\hat{b}^i" + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": " represents the predicted beat distance at frame " + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "inline_equation", + "content": "w_{s}^{i}" + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": " denotes an adaptive weight, which is derived from [26]. It aims to accentuate supervision in regions where beat alignment is imprecise. This is achieved by suppressing the loss magnitude in well-aligned beat regions. This strategy enables the model to focus more attention on challenging instances. The hyper-parameters " + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": " are used for controlling the degree of penalization and the threshold for beat alignment accuracy, respectively. " + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{MSE}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": " denotes the mean squared error. " + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "inline_equation", + "content": "w_{b}^{i}" + }, + { + "bbox": [ + 130, + 367, + 482, + 517 + ], + "type": "text", + "content": " is an adaptive weight that enhances the supervision of frames closely aligned with the beats while diminishing the supervision of frames distantly related to the beats. Benefiting from the beat alignment loss, our model can yield dance motions with superior beat alignment and controllability, largely elevating the quality of the generated results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 534, + 227, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 227, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 227, + 545 + ], + "type": "text", + "content": "3.7 Other Losses" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "type": "text", + "content": "Apart from the proposed beat alignment loss " + }, + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {beat }}" + }, + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "type": "text", + "content": ", we also employ the basic diffusion loss " + }, + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {simple }}" + }, + { + "bbox": [ + 130, + 555, + 481, + 589 + ], + "type": "text", + "content": " (Eq. 2) and several additional auxiliary losses to govern the training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 590, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 481, + 628 + ], + "type": "text", + "content": "Kinematic Loss. We adopt the same auxiliary losses as [41, 42]. The joint positions loss " + }, + { + "bbox": [ + 130, + 590, + 481, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{joint}" + }, + { + "bbox": [ + 130, + 590, + 481, + 628 + ], + "type": "text", + "content": " (Eq. 6) and velocities loss " + }, + { + "bbox": [ + 130, + 590, + 481, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vel}" + }, + { + "bbox": [ + 130, + 590, + 481, + 628 + ], + "type": "text", + "content": " (Eq. 7) are used to improve the overall physical plausibility of the generated dance motions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 224, + 637, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 637, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 224, + 637, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| F K \\left(\\mathbf {x} _ {0} ^ {i}\\right) - F K \\left(\\hat {\\mathbf {x}} _ {0} ^ {i}\\right) \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "d163a61cf0678de806ee9d7db6e3b51812dac8f2307af0bb9dcc9e6607587e71.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 190, + 125, + 482, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 125, + 482, + 157 + ], + "spans": [ + { + "bbox": [ + 190, + 125, + 482, + 157 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v e l} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| \\mathbf {x} _ {0} ^ {i ^ {\\prime}} - \\hat {\\mathbf {x}} _ {0} ^ {i ^ {\\prime}} \\| _ {2} ^ {2} + \\| F K (\\mathbf {x} _ {0} ^ {i}) ^ {\\prime} - F K (\\hat {\\mathbf {x}} _ {0} ^ {i}) ^ {\\prime} \\| _ {2} ^ {2}. \\tag {7}", + "image_path": "4325742d825edb634c9a467d231ff3e97e0a1efebc21e2decaf77de8e1b9993a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 162, + 481, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 162, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 131, + 162, + 481, + 187 + ], + "type": "text", + "content": "The contact consistency loss " + }, + { + "bbox": [ + 131, + 162, + 481, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {contact }}" + }, + { + "bbox": [ + 131, + 162, + 481, + 187 + ], + "type": "text", + "content": " (Eq. 8) proposed in EDGE [42] is employed to alleviating the foot-sliding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 194, + 482, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 194, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 225, + 194, + 482, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t a c t}} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| F K _ {\\text {f o o t}} \\left(\\hat {\\mathbf {x}} _ {0} ^ {i}\\right) ^ {\\prime} \\cdot \\hat {g} ^ {i} \\| _ {2} ^ {2}, \\tag {8}", + "image_path": "b7abab16b4e326698d1d5f2fa2f66beaa17a2e59480db05dab02a8756587541c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "text", + "content": " denotes the index of the frame, and " + }, + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "inline_equation", + "content": "FK(\\cdot)" + }, + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "text", + "content": " is the forward kinematics function that converts the generated 6-DOF rotation representation into 3D key points in Cartesian space. " + }, + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\hat{g}^i" + }, + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "text", + "content": " denotes the predicted binary foot contact label. In addition, we employ acceleration loss " + }, + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{acc}" + }, + { + "bbox": [ + 131, + 236, + 482, + 308 + ], + "type": "text", + "content": " to further improve the quality of generated dance motions, which is suggested in previous work [35], preventing jitters in the generated dance motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 185, + 316, + 482, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 316, + 482, + 349 + ], + "spans": [ + { + "bbox": [ + 185, + 316, + 482, + 349 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a c c} = \\frac {1}{L} \\sum_ {i = 1} ^ {L} \\| \\mathbf {x} _ {0} ^ {i ^ {\\prime \\prime}} - \\hat {\\mathbf {x}} _ {0} ^ {i ^ {\\prime \\prime}} \\| _ {2} ^ {2} + \\| F K (\\mathbf {x} _ {0} ^ {i}) ^ {\\prime \\prime} - F K (\\hat {\\mathbf {x}} _ {0} ^ {i}) ^ {\\prime \\prime} \\| _ {2} ^ {2}. \\tag {9}", + "image_path": "9ffbb756d12f65b4a6f1a4cd3bba80c65df4a2bc4d27e5d55ec83cd9e50cedcd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 357, + 358, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 358, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 358, + 367 + ], + "type": "text", + "content": "The overall kinematic loss is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 175, + 378, + 482, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 378, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 175, + 378, + 482, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {k i n}} = \\lambda_ {\\text {j o i n t}} \\mathcal {L} _ {\\text {j o i n t}} + \\lambda_ {\\text {v e l}} \\mathcal {L} _ {\\text {v e l}} + \\lambda_ {\\text {c o n t a c t}} \\mathcal {L} _ {\\text {c o n t a c t}} + \\lambda_ {\\text {a c c}} \\mathcal {L} _ {\\text {a c c}}, \\tag {10}", + "image_path": "fd2302873da16017d537e9031cf0573f66fb8b3b3698662cda39dfbae9e036fe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "inline_equation", + "content": "\\lambda_{joint}" + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "inline_equation", + "content": "\\lambda_{vel}" + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "inline_equation", + "content": "\\lambda_{contact}" + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "inline_equation", + "content": "\\lambda_{acc}" + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "content": " are hyper-parameters controlling the weights of the corresponding losses. We set " + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "inline_equation", + "content": "\\lambda_{joint} = 1" + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "inline_equation", + "content": "\\lambda_{vel} = 2.5" + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "inline_equation", + "content": "\\lambda_{contact} = 10" + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "inline_equation", + "content": "\\lambda_{acc} = 0.1" + }, + { + "bbox": [ + 131, + 399, + 481, + 436 + ], + "type": "text", + "content": " empirically. The overall loss function is formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 225, + 445, + 482, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 445, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 225, + 445, + 482, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {\\text {k i n}} \\mathcal {L} _ {\\text {k i n}} + \\lambda_ {\\text {b e a t}} \\mathcal {L} _ {\\text {b e a t}}, \\tag {11}", + "image_path": "32fb4f159bcc106fdc567b50d70305653fce0b267234e4db4f6d1ea8722f481a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{kin}" + }, + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{beat}" + }, + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "text", + "content": " are hyper-parameters controlling the weights of the corresponding losses. " + }, + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{kin}" + }, + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{beat}" + }, + { + "bbox": [ + 131, + 467, + 482, + 491 + ], + "type": "text", + "content": " are set as 1 and 0.5, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 509, + 230, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 230, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 230, + 522 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 482, + 629 + ], + "type": "text", + "content": "Dataset. Following previous works [35, 42], we choose the most widely used dataset, AIST++ [23] for both training and testing. This dataset comprises 1,408 dance sequences encompassing 10 distinct street dance genres. Each dance genre incorporates 6 musical pieces with different BPMs. " + }, + { + "bbox": [ + 131, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 131, + 533, + 482, + 629 + ], + "type": "text", + "content": " of these sequences are categorized as basic choreographies, where performers execute the same choreography across all 6 musical pieces, each featuring varying tempos to synchronize with distinct musical beats. With its high choreographic and beat diversity, this dataset is ideal for our controllable dance generation task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. For the hierarchical multi-condition fusion module, we implement six sparse-dense fusion blocks and two dense-dense fusion blocks, utilizing a base dilation step " + }, + { + "bbox": [ + 131, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 630, + 482, + 666 + ], + "type": "text", + "content": " set to 4, 8, 12, 16, 20, 24. For diffusion" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Z. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "process, we adopt the cosine schedule as beat schedule and set diffusion steps " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " as 1000 diffusion following [41]. The classifier-free guidance scale is set to 2. The diffusion module aligns with a transformer-based architecture similar to [42]. We adopt Adam optimizer with the learning rate set as 2e-4 and the training batch size is set to 64. During training, we randomly sample between " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " of the ground truth (GT) motion frames as the keyframe conditions. This training strategy can effectively improve the generalization capability of our method on different numbers of keyframes. We use the GT motion beats as beat conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 294 + ], + "type": "text", + "content": "To ensure fair comparisons, we follow EDGE [42] to crop all data into 5-second clips at 30 frames per second, with 2.5 seconds of overlapping. During testing, we randomly sample " + }, + { + "bbox": [ + 130, + 212, + 482, + 294 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 212, + 482, + 294 + ], + "type": "text", + "content": " of keyframes from unpaired samples in the testing set as keyframe conditions and subsets of testing musical beats as beat conditions to prevent data leakage. The musical beats used for experiments are extracted by the off-the-shelf audio toolkit Librosa [17]. All experiments are conducted on 4 NVIDIA RTX3090 GPUs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "content": "Evaluation Metrics. For quantitative evaluation, we measure the generated dance from three aspects: generation quality, diversity, and controllability. Due to the limited samples in the testing set of AIST++ [23], the prior work EDGE [42] has demonstrated that the Fréchet Inception Distance (FID) is not reliable for measuring the generation quality. Therefore, we do not adopt FID as the evaluation metric. Instead, we assess generation quality from two different perspectives: music-dance correlation and kinematic plausibility. In terms of music-dance correlation, we utilize the Beat Alignment Score (BAS) metric, following the approach presented in [35]. BAS quantifies the synchronization quality between the generated dance and the musical beat. For the evaluation of kinematic plausibility, we adopt the Physical Foot Contact score (PFC) metric proposed in [42]. PFC provides a metric for measuring the physical plausibility of the generated dance from a kinematic perspective. As for the diversity, we measure the average feature distance of kinetic " + }, + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "inline_equation", + "content": "(\\mathrm{Div}_k)" + }, + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "content": " and geometric " + }, + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "inline_equation", + "content": "(\\mathrm{Div}_g)" + }, + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "content": " features extracted by fairmotion [12] following the previous works [23, 35, 42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "content": "For assessing controllability, we employ the Key Pose Distance (KPD) and Beat Assignment Precision (BAP) metrics to evaluate keyframe control and beat control, respectively. KPD quantifies the average mean square error of local joint positions in Cartesian space at the keyframes. To evaluate beat control, we measure the alignment between the generated dance and the designated beat condition. BAP is defined as the percentage of beats in the generated dance that are correctly assigned to the designated beat frames. Note that this metric differs from the BAS, as a dance sequence with a high BAS does not necessarily adhere to the specific motion beat choreography constraint we intend to impose." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 599, + 328, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 599, + 328, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 599, + 328, + 611 + ], + "type": "text", + "content": "4.1 Comparison to Existing Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Currently, only three music-to-dance methods are publicly accessible: EDGE [42], Bailando [35], and FACT [23]. Of these, EDGE [42] is most directly aligned with our approach, as it similarly accommodates keyframes as inputs for generating dance sequences. We therefore assess EDGE using the same keyframe conditions" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 157, + 454, + 264 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons among different methods on AIST++. Bold indicates best result. ↓ means lower is better, ↑ means higher is better and → means closer to the ground truth is better." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 157, + 454, + 264 + ], + "lines": [ + { + "bbox": [ + 159, + 157, + 454, + 264 + ], + "spans": [ + { + "bbox": [ + 159, + 157, + 454, + 264 + ], + "type": "table", + "html": "
MethodsQualityDiversityControllability
PFC ↓ BAS ↑Divk→Divm→KPD ↓ BAP ↑
Ground Truth1.3380.3849.7737.212--
FACT [23]2.6980.2029.7047.342--
Bailando [35]1.5780.2159.6227.175--
EDGE [42](keyframes)1.0840.2359.7437.2740.859-
Ours(beat & keyframes)0.9660.6619.6607.2480.3060.793
", + "image_path": "2f5a94ebd3313a0ad8e8943299325b761a69daa1770a323d12dc1219c5aac1f3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 483, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 483, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 483, + 334 + ], + "type": "text", + "content": "as our method to ensure a fair comparison. Conversely, certain approaches, like DanceFormer [21], are not open-source, thus hindering any direct evaluative comparison. We use the original test split of AIST++ [23] for evaluation, which has no overlap with the training split in terms of both the music and dance choreography." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 140, + 363, + 288, + 451 + ], + "blocks": [ + { + "bbox": [ + 151, + 351, + 281, + 361 + ], + "lines": [ + { + "bbox": [ + 151, + 351, + 281, + 361 + ], + "spans": [ + { + "bbox": [ + 151, + 351, + 281, + 361 + ], + "type": "text", + "content": "Table 2: Results of user study." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 363, + 288, + 451 + ], + "lines": [ + { + "bbox": [ + 140, + 363, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 140, + 363, + 288, + 451 + ], + "type": "table", + "html": "
User StudyBeat-It Win Rate
FACT92.2%
Bailando78.8%
EDGE \n(keyframes)QualityControllability \n(Keyframe)
60.3%86.9%
", + "image_path": "48f4518a6c760932164bf2f4041f3393e367b7ec615012bb8004c34e4e7e76a1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 363, + 469, + 440 + ], + "blocks": [ + { + "bbox": [ + 310, + 351, + 468, + 361 + ], + "lines": [ + { + "bbox": [ + 310, + 351, + 468, + 361 + ], + "spans": [ + { + "bbox": [ + 310, + 351, + 468, + 361 + ], + "type": "text", + "content": "Table 3: Ablation study on AIST++." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 363, + 469, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 363, + 469, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 469, + 440 + ], + "type": "table", + "html": "
AblationsQualityControllability
PFC ↓BAS ↑KPD ↓BAP ↑
w/o HF25.6260.3220.4770.323
w/o BD1.6320.3580.3890.371
w/o Lbeat1.3420.3970.3430.411
Ours0.9660.6610.3060.793
", + "image_path": "6a43d171843f2ee9b4dbf5734ffb28dcfc4d992f0ad8c07056a3864a06eb7187.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 451, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 641 + ], + "type": "text", + "content": "Quantitative Comparisons. The quantitative results are shown in Tab. 1. As is revealed, our method surpasses all the competitors in terms of generation quality, diversity, and controllability. Particularly, our method shows considerable advantages over the compared methods for BAS metrics, notably with an improvement of 0.426 compared with the SOTA method EDGE [42]. This demonstrates the effectiveness of our method in generating beat-synchronized dance sequences while maintaining diversity. Furthermore, in the aspect of controllability, our method also significantly excels EDGE [42] with a notable improvement of 0.553 in KPD. Fig. 4 shows the mean joint velocities over time for a specific dance motion concerning both the GT and the existing methods. Notably, compared with others, our approach demonstrates more motion beats (local minimum point of velocities) precisely at the beat frames. This contributes to a better visual alignment with the audio when compared to the other methods. The BAP further validates our method's superior ability to produce dance sequences precisely aligned with the beat condition. Due to limited space, we present additional quantitative experiments in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "User Study. We also conduct a user study to assess the generation quality and controllability of our method. We randomly select 20 generated dance sequences" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Z. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 479, + 197 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 197 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 197 + ], + "type": "image", + "image_path": "4260c2e47c89f273f0e0a170133747059998e2458f470ae96ed7d18041a0ef2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 204, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 236 + ], + "type": "text", + "content": "Fig. 4: Visualization comparison on beat alignment among different methods. The motion generated by our method shows precise beat alignment with the given beat condition, demonstrating the superiority of our method in beat control." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "content": "from the testing set and ask 18 participants to pick the best one among our method and the other compared methods. In particular, participants are required to choose the superior dance according to several criteria, namely, the overall visual plausibility of movements, synchronization with the dance beats, and keyframe controllability. The results of the user study are tabulated in Tab. 2. Our method shows noticeable advantages in terms of generation quality compared to other approaches, particularly in beat alignment. For qualitative evaluations on keyframe-conditioned controllability, we also provide the participants with the keyframe references for intuitive comparison between our method against EDGE [42]. The results show that our approach owns better motion controllability than EDGE [42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 407, + 246, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 246, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 246, + 418 + ], + "type": "text", + "content": "4.2 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "content": "To get a more comprehensive insight into our main contributions, we conduct ablation studies on the hierarchical multi-condition fusion module, the beat-aware dilation scheme, and the beat alignment loss to evaluate their efficacy. The quantitative results are shown in Tab. 3. The qualitative results can be also found in the supplementary video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 640 + ], + "type": "text", + "content": "Hierarchical Multi-condition Fusion Module. To verify the significance of the hierarchical multi-condition fusion module, we remove it by simply concatenating all the conditions and feeding them into a one-stage encoder (w/o HF). As shown in Tab. 3, the simple one-stage fusion module can lead to inferior results in both quality and controllability. This is because the direct concatenation of multiple conditions can lead to inevitable condition conflicts, giving rise to suboptimal optimization of the model. Even worse, the naive padding of the sparse keyframe condition introduces severe noises, further complicating the learning process and resulting in unsatisfactory generation performance. By decently fusing multiple conditions gradually, our hierarchical multi-condition fusion mechanism effectively mitigates the conflicts among different conditions and yields more informative conditions, allowing for higher-quality generation with better controllability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Beat-aware Dilation Scheme. To validate the effectiveness of our beat-aware dilation scheme, we replace it with a vanilla masked attention strategy (w/o" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "BD). As shown in Tab. 3, the absence of the beat-aware dilation scheme notably impairs both keyframe and beat choreography controllability. The reason behind this is that beat-aware dilation can dynamically adjust the expansion degree of the attention mask, contingent on the distance between the keyframe and the beat frames. This tactical approach significantly improves the synchronization between keyframes and their associated beat frames, thereby empowering the model to more effectively harness prior knowledge of beats. Consequently, this enhancement considerably bolsters beat control while augmenting keyframe controllability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 483, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 483, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 483, + 331 + ], + "type": "text", + "content": "Beat Alignment Loss. To further validate the effectiveness of the beat alignment loss, we assess the model's performance by comparing it to its counterpart without it (w/o " + }, + { + "bbox": [ + 130, + 212, + 483, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {beat }}" + }, + { + "bbox": [ + 130, + 212, + 483, + 331 + ], + "type": "text", + "content": "). As shown in Tab. 3, the ablated variant without the beat alignment loss demonstrated performance degradation in beat choreography, showcasing decreases of 0.264 and 0.382 in BAS and BAP, respectively. This decline in performance can be attributed to the advantageous impact of direct supervision during training, particularly in guiding the model to comprehend beat-level dance choreography through the proposed beat alignment loss. Consequently, this enhancement significantly improves the synchronization between motion sequences and designated beats." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 348, + 312, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 312, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 312, + 362 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 373, + 483, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 483, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 483, + 517 + ], + "type": "text", + "content": "In this work, we propose a novel multi-condition diffusion-based framework, BeatIt, for beat-synchronized and key pose-guided dance generation. The proposed framework presents a hierarchical multi-condition fusion mechanism equipped with a beat-aware mask dilation scheme to integrate conditions with different information sparsity. To achieve precise beat synchronization, we explicitly disentangle the beats from the music and inject beat controllability throughout the entire generation process. Additionally, we introduce a specifically designed beat alignment loss to provide explicit guidance and supervision on motion beats. Both qualitative and quantitative experimental results on the AIST++ dataset validate the superiority of our method in producing high-quality dance sequences with precise beat synchronization and flexible keyframe control. Further limitations and broader impact are discussed in supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 533, + 249, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 249, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 249, + 548 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": "The work is supported by China National Key R&D Program (No. 2023YFE0202700), Key-Area Research and Development Program of Guangzhou City (No. 2023B01J0022), Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (No. 2024B1515040010), Health and Medical Research Fund (HMRF) of Hong Kong Health Bureau (No. 10211516), Innovation and Technology Fund - Innovation and Technology Support Programme (ITF-ITSP) (No. ITS/127/22), Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2023B1515020097), National Research Foundation Singapore under the AI Singapore Programme (No. AISG3-GV-2023-011)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Z. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 171 + ], + "type": "text", + "content": "1. Alemi, O., Françoise, J., Pasquier, P.: Groovenet: Real-time music-driven dance movement generation using artificial neural networks. networks 8(17), 26 (2017) 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 204 + ], + "type": "text", + "content": "2. Alexanderson, S., Nagy, R., Beskow, J., Henter, G.E.: Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM TOG 42(4), 1-20 (2023) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 226 + ], + "type": "text", + "content": "3. Ao, T., Zhang, Z., Liu, L.: Gesturediffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613 (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 258 + ], + "type": "text", + "content": "4. Araújo, J.P., Li, J., Vetrivel, K., Agarwal, R., Wu, J., Gopinath, D., Clegg, A.W., Liu, K.: Circle: Capture in rich contextual environments. In: CVPR. pp. 21211-21221 (2023) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 260, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 481, + 281 + ], + "type": "text", + "content": "5. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR. pp. 9760-9770 (2023) 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 283, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 481, + 314 + ], + "type": "text", + "content": "6. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. NeurIPS 34, 19822-19835 (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 358 + ], + "type": "text", + "content": "7. Ferreira, J.P., Coutinho, T.M., Gomes, T.L., Neto, J.F., Azevedo, R., Martins, R., Nascimento, E.R.: Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics 94, 11-21 (2021) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 360, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 481, + 391 + ], + "type": "text", + "content": "8. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV. pp. 89-106. Springer (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 423 + ], + "type": "text", + "content": "9. Gao, X., Hu, L., Zhang, P., Zhang, B., Bo, L.: Dancemeld: Unraveling dance phrases with hierarchical latent codes for music-to-dance synthesis. arXiv preprint arXiv: 2401.10242 (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 425, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 481, + 456 + ], + "type": "text", + "content": "10. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: Imos: Intent-driven full-body motion synthesis for human-object interactions. In: CGF. vol. 42, pp. 1-12. Wiley Online Library (2023) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 458, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 481, + 489 + ], + "type": "text", + "content": "1. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: ICCV. pp. 9942-9952 (2023) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 521 + ], + "type": "text", + "content": "2. Gopinath, D., Won, J.: fairmotion - tools to load, process and visualize motion capture data. Github (2020), https://github.com/facebookresearch/fairmotion11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "text", + "content": "3. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020) 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 566 + ], + "type": "text", + "content": "4. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. ACM TOG 35(4), 1-11 (2016) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 598 + ], + "type": "text", + "content": "5. Huang, R., Hu, H., Wu, W., Sawada, K., Zhang, M., Jiang, D.: Dance revolution: Long-term dance generation with music via curriculum learning. arXiv preprint arXiv:2006.06119 (2020) 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 600, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 631 + ], + "type": "text", + "content": "6. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR. pp. 16750-16761 (2023) 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 664 + ], + "type": "text", + "content": "7. Jin, Y., Zhang, J., Li, M., Tian, Y., Zhu, H., Fang, Z.: Towards the automatic anime characters creation with generative adversarial networks. arXiv preprint arXiv:1708.05509 (2017) 11" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI. vol. 37, pp. 8255-8263 (2023) 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "text", + "content": "19. Kim, J., Oh, H., Kim, S., Tong, H., Lee, S.: A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In: CVPR. pp. 3490-3500 (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 170, + 481, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 481, + 191 + ], + "type": "text", + "content": "20. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. NeurIPS 32 (2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 191, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 191, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 481, + 224 + ], + "type": "text", + "content": "21. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI. vol. 36, pp. 1272-1279 (2022) 2, 4, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 255 + ], + "type": "text", + "content": "22. Li, J., Yin, Y., Chu, H., Zhou, Y., Wang, T., Fidler, S., Li, H.: Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171 (2020) 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 256, + 481, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 481, + 288 + ], + "type": "text", + "content": "23. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV. pp. 13401-13412 (2021) 2, 4, 7, 10, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 288, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 288, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 288, + 481, + 320 + ], + "type": "text", + "content": "24. Lin, J., Chang, J., Liu, L., Li, G., Lin, L., Tian, Q., Chen, C.w.: Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In: CVPR. pp. 23222-23231 (2023) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 320, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 481, + 342 + ], + "type": "text", + "content": "25. Liu, H., Xu, C., Yang, Y., Zeng, L., He, S.: Drag your noise: Interactive point-based editing via diffusion semantic propagation. In: CVPR. pp. 6743-6752 (2024) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 363 + ], + "type": "text", + "content": "26. Lu, X., Ma, C., Shen, J., Yang, X., Reid, I., Yang, M.H.: Deep object tracking with shrinkage loss. IEEE TPAMI 44(5), 2386-2401 (2020) 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "text", + "content": "27. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "text", + "content": "28. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "text", + "content": "29. Qi, Q., Zhuo, L., Zhang, A., Liao, Y., Fang, F., Liu, S., Yan, S.: Diffdance: Cascaded human motion diffusion model for dance generation. In: ACM MM. pp. 1374-1382 (2023) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "text", + "content": "30. Qin, C., Zhang, S., Yu, N., Feng, Y., Yang, X., Zhou, Y., Wang, H., Niebles, J.C., Xiong, C., Savarese, S., et al.: Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147 (2023) 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "31. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML. pp. 8821-8831. PMLR (2021) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 547 + ], + "type": "text", + "content": "32. Ren, X., Li, H., Huang, Z., Chen, Q.: Self-supervised dance video synthesis conditioned on music. In: ACM MM. pp. 46-54 (2020) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "text", + "content": "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 568, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 610 + ], + "type": "text", + "content": "34. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS 35, 36479-36494 (2022) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "35. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR. pp. 11050-11059 (2022) 2, 4, 7, 10, 11, 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "36. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando++: 3d dance gpt with choreographic memory. IEEE TPAMI (2023) 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Z. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Sun, G., Wong, Y., Cheng, Z., Kankanhalli, M.S., Geng, W., Li, X.: Deepdance: music-to-dance motion choreography with adversarial learning. IEEE TMM 23, 497-509 (2020) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "38. Sun, J., Wang, C., Hu, H., Lai, H., Jin, Z., Hu, J.F.: You never stop dancing: Non-freezing dance generation via bank-constrained manifold projection. NeurIPS 35, 9995-10007 (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "text", + "content": "39. Tang, T., Jia, J., Mao, H.: Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In: ACM MM. pp. 1598-1606 (2018) 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "type": "text", + "content": "40. Tang, Z., Yang, Z., Zhu, C., Zeng, M., Bansal, M.: Any-to-any generation via composable diffusion. arXiv preprint arXiv: 2305.11846 (2023) 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "text", + "content": "41. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) 9, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 252, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 481, + 274 + ], + "type": "text", + "content": "42. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: CVPR. pp. 448-458 (2023) 2, 4, 7, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 133, + 274, + 481, + 308 + ], + "type": "text", + "content": "43. Valle-Pérez, G., Henter, G.E., Beskow, J., Holzapfel, A., Oudeyer, P.Y., Alexanderson, S.: Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM TOG 40(6), 1-14 (2021) 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 481, + 341 + ], + "type": "text", + "content": "44. Wang, T., Zhang, T., Zhang, B., Ouyang, H., Chen, D., Chen, Q., Wen, F.: Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952 (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 342, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 481, + 373 + ], + "type": "text", + "content": "45. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS 35, 14959-14971 (2022) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 375, + 481, + 407 + ], + "type": "text", + "content": "46. Wei, D., Sun, X., Sun, H., Li, B., Hu, S., Li, W., Lu, J.: Understanding text-driven motion synthesis with keyframe collaboration via diffusion models. arXiv preprint arXiv:2305.13773 (2023) 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 408, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 408, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 133, + 408, + 481, + 452 + ], + "type": "text", + "content": "47. Xie, Y., Lin, Y., Cai, W., Xu, X., Zhang, H., Du, Y., He, S.: D3still: Decoupled differential distillation for asymmetric image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17181-17190 (2024) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 453, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 453, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 453, + 481, + 475 + ], + "type": "text", + "content": "48. Xu, C., Chen, Z., Mai, J., Xu, X., He, S.: Pose-and attribute-consistent person image synthesis. ACM TOMM 19(2s), 1-21 (2023) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 475, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 481, + 497 + ], + "type": "text", + "content": "49. Xu, C., Li, K., Luo, X., Xu, X., He, S., Zhang, K.: Fully deformable network for multiview face image synthesis. IEEE TNNLS (1-15) (2022) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 498, + 481, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 481, + 519 + ], + "type": "text", + "content": "50. Xu, C., Xu, Y., Zhang, H., Xu, X., He, S.: Dreamanime: Learning style-identity textual disentanglement for anime and beyond. IEEE TVCG (2024) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 521, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 481, + 552 + ], + "type": "text", + "content": "51. Yalta, N., Watanabe, S., Nakadai, K., Ogata, T.: Weakly-supervised deep recurrent neural networks for basic dance step generation. In: IJCNN. pp. 1-8. IEEE (2019) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 554, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 554, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 133, + 554, + 481, + 576 + ], + "type": "text", + "content": "52. Yang, Z., Wen, Y.H., Chen, S.Y., Liu, X., Gao, Y., Liu, Y.J., Gao, L., Fu, H.: Keyframe control of music-driven 3d dance generation. IEEE TVCG (2023) 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 577, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 577, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 133, + 577, + 481, + 598 + ], + "type": "text", + "content": "53. Yin, W., Yin, H., Baraka, K., Kragic, D., Björkman, M.: Dance style transfer with cross-modal transformer. In: WACV. pp. 5058-5067 (2023) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "text", + "content": "54. Yu, Y., Liu, B., Zheng, C., Xu, X., Zhang, H., He, S.: Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In: CVPR. pp. 7109-7118 (2024) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "text", + "content": "55. Zhang, C., Tang, Y., Zhang, N., Lin, R.S., Han, M., Xiao, J., Wang, S.: Bidirectional autoregressive diffusion model for dance generation. arXiv preprint arXiv: 2402.04356 (2024) 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 101 + ], + "type": "text", + "content": "Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 293 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "56. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "57. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023) 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "58. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "59. Zheng, C., Liu, B., Xu, X., Zhang, H., He, S.: Learning an interpretable stylized subspace for 3d-aware animatable artforms. IEEE TVCG (2024) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "60. Zheng, C., Liu, B., Zhang, H., Xu, X., He, S.: Where is my spot? few-shot image generation via latent subspace optimization. In: CVPR. pp. 3272-3281 (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "61. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: CVPR. pp. 5632-5641 (2023) 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "62. Zhu, L., Liu, X., Liu, X., Qian, R., Liu, Z., Yu, L.: Taming diffusion models for audio-driven co-speech gesture generation. In: CVPR. pp. 10544-10553 (2023) 4" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Z. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_content_list.json b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e1f50d6d7bc13ebbf9543d329d06956616ad5379 --- /dev/null +++ b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_content_list.json @@ -0,0 +1,1789 @@ +[ + { + "type": "text", + "text": "BenchLMM: Benchmarking Cross-style Visual Capability of Large Multimodal Models", + "text_level": 1, + "bbox": [ + 238, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rizhao Cai $^{1}$ , Zirui Song $^{1,2}$ , Dayan Guan $^{3}$ , Zhenhao Chen $^{4}$ , Yaohang Li $^{2}$ , Xing Luo $^{5}$ , Chenyu Yi $^{1}$ , and Alex Kot $^{1}$", + "bbox": [ + 259, + 210, + 743, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Nanyang Technological University, Singapore", + "bbox": [ + 341, + 253, + 658, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of Technology Sydney, Australia; Northeastern University, China;", + "bbox": [ + 238, + 268, + 761, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Harbin Institute of Technology, China", + "bbox": [ + 367, + 281, + 635, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates", + "bbox": [ + 230, + 295, + 771, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Zhejiang University, China", + "bbox": [ + 403, + 309, + 599, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† Equal Contribution Corresponding author (guandayan@hit.edu.cn)", + "bbox": [ + 299, + 325, + 700, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large Multimodal Models (LMMs) such as GPT-4V and LLaVA have shown remarkable capabilities in visual reasoning on data in common image styles. However, their robustness against diverse style shifts, crucial for practical applications, remains largely unexplored. In this paper, we propose a new benchmark, BenchLMM, to assess the robustness of LMMs toward three different styles: artistic image style, imaging sensor style, and application style. Utilizing BenchLMM, we comprehensively evaluate state-of-the-art LMMs and reveal: 1) LMMs generally suffer performance degradation when working with other styles; 2) An LMM performs better than another model in common style does not guarantee its superior performance in other styles; 3) LMMs' reasoning capability can be enhanced by prompting LMMs to predict the style first, based on which we propose a versatile and training-free method for improving LMMs; 4) An intelligent LMM is expected to interpret the causes of its errors when facing stylistic variations. We hope that our benchmark and analysis can shed new light on developing more intelligent and versatile LMMs. The benchmark and evaluation have been released on https://github.com/AIFEG/BenchLMM.", + "bbox": [ + 259, + 369, + 743, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 643, + 375, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The dynamic landscape of computer vision has recently been shaped by the rise of Large Multimodal Models (LMMs). These models [10, 14, 32, 39, 71], incorporating visual and textual data, are pivotal in developing general-purpose assistants for visual reasoning, a widely-studied task that involves reasoning and answering questions based on images. Their growing prominence and the enthusiasm surrounding visual reasoning are evident from numerous studies such as GPT-4V [62]. As the interest in LMMs swells within the research community, understanding their depth, breadth, and limitations becomes imperative.", + "bbox": [ + 212, + 672, + 787, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To comprehensively understand the visual reasoning capabilities of LMMs, numerous benchmarks [17, 21, 33, 34, 41, 67] have been proposed to systematically evaluate the linguistic capabilities of LMMs. These linguistic capabilities", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/404bb180824da96571d9ab6fe665a23fdf682a8ebbbf6904cd560edb5cdd35fa.jpg", + "image_caption": [ + "Q: How many forks? \nA (GT): Two \nA (GPT-4V): Two \nA (LLaVA-1: 5) Two" + ], + "image_footnote": [], + "bbox": [ + 225, + 145, + 359, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/22939106675c9ad1454f1a74756e33be497feb2114334d0298ca783a0f8ea8d4.jpg", + "image_caption": [ + "[Painting]Q:How many elephants? A (GT): Seven A (GPT-4V):Five A(LLAVA-1.5):Five" + ], + "image_footnote": [], + "bbox": [ + 362, + 145, + 500, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dcee0d27a1f39563afdbacaf526c3d584184758197c7ab7e9d3002689d3b8e16.jpg", + "image_caption": [ + "[Xray] Q: What are prohibited objects? \nA (GT): None \nA (GPT-4V): Gun \nA (LLVaA-1.5): A knife and a scissor", + "Fig. 1: The motivation of our work. State-of-the-art LMMs, such as GPT-4V and LLaVA-1.5 can understand the images in a common style. However, these models have difficulties in reasoning images in other styles and deliver wrong answers." + ], + "image_footnote": [], + "bbox": [ + 501, + 145, + 638, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/34c9bd7e9dfad720e12e11ec50f92198090d351c476e418930f531b04b805486.jpg", + "image_caption": [ + "[Driving] Q: What is the action for the ego-vehicle during autonomous driving? A (GT): Go straight A (GPT-4V): Stop at traffic light A (LLAvA-1.5): Stop" + ], + "image_footnote": [], + "bbox": [ + 642, + 146, + 779, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "are assessed through various types of responses for visual reasoning tasks, including Yes/No answers [17], multi-choice answers [33,41], single word or phrase answers [21,24,44]. However, while these benchmarks investigate textual diversity, they may fall short in terms of visual diversity as most images in these benchmarks are photographic images sourced from the internet. Despite providing valuable insights into LMMs' performance, there remains a gap in understanding how LMMs handle shifts in visual distribution. Although the recently released GPT-4V [62] claims remarkable capabilities across diverse application scenarios, only a few samples of qualitative results are presented in the technical report [62], and the quantitative evaluations across different scenarios are missing. Moreover, according to our preliminary study as illustrated in Fig. 1, we reveal that LMMs, even the state-of-the-art GPT-4V, tend to make mistakes in reasoning images with non-common styles.", + "bbox": [ + 212, + 339, + 787, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further investigate LMMs' cross-style visual reasoning capability, we propose a novel benchmark BenchLMM, which systematically assesses existing LMMs in three distinct styles of distribution shifts: artistic styles, sensor styles, and application styles. Firstly, images of the common style are realistic, but images can be in other artistic styles, such as Cartoon, Painting, Sketch, etc. Besides, there are different imaging sensors other than RGB cameras, such as infrared and X-ray. The discrepancy of sensors induces the distribution shifts of images. Moreover, we also identify the shift of knowledge required for different application styles. For example, the image recognition application requires knowledge about object shape and appearance, while the application of decision-making for robot action requires application-specific knowledge.", + "bbox": [ + 212, + 537, + 789, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With our BenchLMM, we conduct in-depth analysis through experiments of evaluating advanced LMMs. The comprehensive experiments reveal that while these cutting-edge LMMs excel with commonly styled data, their performance significantly declines with data in other styles. Interestingly, we also discover that the superior performance of an LMM in common data styles does not necessarily predict better performance across other styles. Furthermore, our investigation shows that prompting an LMM to predict the style of the data before answering a question can significantly boost its reasoning abilities. Based on this insight, we introduce the Style Prompt Enhancement (SPE) technique. SPE is a novel", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "approach that improves LMMs' performance without additional training and operates effectively in a zero-shot manner. Moreover, we analyze the failure cases by conducting an in-depth error-reflection analysis by asking an LMM to interpret why it made mistakes in the Q&A. We find that a more intelligent LMM can understand the detailed reason why it made mistakes and can learn to derive the correct answer, which is an important capability that deserves attention when developing LMMs. We hope that our benchmark and analysis can shed new light on developing more intelligent and versatile LMMs.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized below: 1) BenchchLMM: We introduce BenchLMM, a pioneering benchmark specifically designed to quantitatively evaluate the performance of LMMs' visual reasoning capability across visual styles in artistic, sensor, and application shifts. 2) Comprehensive evaluation of cutting-edge LMMs: We conduct extensive quantitative benchmark experiments on LMMs, including the commercial GPT-4V. Our rigorous and quantitative evaluation reveals LMMs' limitations in versatility and performance across different data styles. This finding challenges the prevailing assumption of universal applicability and highlights the need for specialized strategies to enhance LMM adaptability. 3) Insightful Analysis and Enhancement Strategy: Beyond mere performance assessment, we conduct an in-depth error-reflection analysis, where LMMs are prompted to identify the reasons behind their errors in question-and-answer tasks. This study reveals that more sophisticated LMMs can recognize their mistakes and also learn from them to improve future responses. Moreover, we introduce a novel SPE technique, as a training-free and zero-shot strategy, to significantly boost LMMs' reasoning capabilities across varied data styles, marking a substantial advancement in LMM development.", + "bbox": [ + 212, + 267, + 787, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 549, + 392, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Reasoning Visual reasoning, also known as visual question answering (VQA), aims to reason and answer questions about visual information [3,27,28]. Early developments [5, 18, 30] focus on designing more complex fusion mechanisms instead of simple summation fusion [3]. By including modern attention mechanism [58], numerous works [8, 29, 46, 53] contributed to the creation of transformer-based vision-language models. Recent methods have proposed to improve other aspects of visual question answering, including avoiding shortcut learning [15, 23], addressing the linguistic bias [9, 55], improving transformer-based vision-language models [61, 70], external knowledge integration [16, 19], and consistency regularization [49, 54]. Lastly, the advent of large multimodal models has markedly enhanced the capabilities of visual reasoning.", + "bbox": [ + 212, + 582, + 787, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Multimodal Models Recent advancements in large language models (LLMs) [7, 13, 48, 56] have inspired the development of LMMs that enhance LLMs with vision-language capabilities. Existing LMMs can be categorized into three main classes: multimodal instruction tuning [10, 14, 32, 38, 39, 71] that involves the finetuning of pre-trained LLMs on multi-modal instruction-following data; multimodal in-context learning [2, 32, 57, 63] that derives insights from", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BenchLMM", + "bbox": [ + 650, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "few-shot task-specific samples; multimodal chain-of-thought [20,69] that encourages LMMs to articulate both the final answer and the underlying reasoning process. Additionally, some commercial LMMs like Google's Bard [42] and Microsoft's GPT-4V [62], with detailed information about their architecture remaining undisclosed, have shown impressive visual capabilities.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation of LMMs To evaluate their capabilities, existing LMMs have been tested on a number of vision-language datasets. These datasets focus on specific visual reasoning tasks by designing different kinds of Q&A task, and the specific task includes visual recognition [21, 44], image description [1, 12], scene text understanding [51, 52, 64], object hallucination evaluation [35], complex reasoning [39, 67], and commonsense reasoning [66, 68]. To facilitate a comprehensive comparison of LMMs, several benchmarks [17, 36, 60] have been established by encompassing various visual reasoning tasks with common-style images and diverse Q&A pairs. However, the above benchmarks do not consider different styles of visual distribution shift, which motivates us to propose our BenchLMM.", + "bbox": [ + 212, + 222, + 784, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 BenchLMM", + "text_level": 1, + "bbox": [ + 215, + 398, + 367, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Motivation", + "text_level": 1, + "bbox": [ + 215, + 433, + 352, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LMMs (e.g., GPT-4V) are claimed to excel in various scenarios [62], shown with qualitative results. Nevertheless, we provide comprehensive and quantitative results experiments with our BenchLMM reveal a paradox. We will elaborate on the benchmark that encompasses different cross-style data as follows.", + "bbox": [ + 212, + 460, + 782, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-artistic Style Existing benchmarks for LMMs commonly use photo images for evaluation, which we name as 'Common' style. Meanwhile, there are diverse image styles, including Cartoon, Sketch, etc., as shown and compared in Fig. 2. The object in other artistic styles have different distributions of object characteristics. For instance, a colorful zebra, common in painting and tattoo styles, is rare in 'Common' style images. While humans easily interpret such characteristics, the extent of LMMs' understanding of the other styles remains uncertain. This uncertainty arises from potential insufficient training on datasets encompassing these diverse styles. Therefore, we propose an evaluation to assess LMMs' robustness in analyzing images across various artistic styles.", + "bbox": [ + 212, + 522, + 784, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-sensor Style Photographic images are commonly obtained through RGB cameras for information sensing. Alongside RGB cameras, other sensor types are also important for capturing images. For instance, infrared cameras are frequently utilized in surveillance systems to capture images during nighttime, as RGB cameras rely on adequate illumination for optimal image quality. Additionally, X-ray sensors are used in airport security checks, enabling the detection of prohibited items within luggage. These sensors feature distinct imaging processes compared to RGB cameras. Consequently, images acquired through various sensors exhibit unique intrinsic properties. The question of whether the performance of LMMs remains remarkable with other sensor data has not been thoroughly investigated. This issue will be addressed in our benchmark study.", + "bbox": [ + 212, + 672, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f2d2babc524998a0bf475f18c4582ac1e28a99862186cc4d0d6561d5f93d6239.jpg", + "image_caption": [ + "Q: What time is it? A: 2:57:39" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 336, + 207 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4be0ed448c216202254e1066ace484dd64c0887d12b1f9b8e610c1bbb95e8c2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 145, + 447, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4424394d2d32f6650034b12110bceeec2d7b594dc316c94b2164dadc46ce58e9.jpg", + "image_caption": [ + "Q: Who is sitting on the couch? A: A woman" + ], + "image_footnote": [], + "bbox": [ + 450, + 145, + 558, + 207 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/698a9a8b89db27df8fff7a464e3d07f55ce4c5feea13f35d8baa49b2accf3164.jpg", + "image_caption": [ + "Q: What kind of sport? A: football" + ], + "image_footnote": [], + "bbox": [ + 562, + 145, + 669, + 207 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cabfe674ef11cc3714c16781c111d0578268257bf16c11bc2781ce0e842aad14.jpg", + "image_caption": [ + "Tattoo", + "Q: What is in the image? A: A colorful zebra" + ], + "image_footnote": [], + "bbox": [ + 673, + 152, + 782, + 207 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c031735a6bd5d0bea0fd295827d09b7be8f5e1a38c91fbf4c699d0b4ebf8cd68.jpg", + "image_caption": [ + "Q: How many pedestrians? A: 4" + ], + "image_footnote": [], + "bbox": [ + 217, + 224, + 336, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/708b1202d956d4b16cd5625bc8484eed65f62761e9035bfc4036438e4210fdce.jpg", + "image_caption": [ + "A: A bronze sculpture of a person with a very large head", + "Q: What are prohibited objects? \nA: Gun, knife" + ], + "image_footnote": [], + "bbox": [ + 341, + 224, + 447, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/288027131b791a85b0e6d07efea4d486e0636593c8828a7ca31e99089c59981b.jpg", + "image_caption": [ + "Q: What is the main organ? \nA: Lung" + ], + "image_footnote": [], + "bbox": [ + 459, + 224, + 550, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3d1c1f07e0627ce2b5b668d903b5a0a36320db83ab59f8440ada08f192aaf07f.jpg", + "image_caption": [ + "Q: What kind of disease? A: Brain edema, brain tumor" + ], + "image_footnote": [], + "bbox": [ + 562, + 224, + 669, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/27c2cd8ad1a0c8faba2c2eb7310deb1b207c4f5b4897dd164a688650dfc11e81.jpg", + "image_caption": [ + "Q: What kind of disease? \nA: Lung cancer" + ], + "image_footnote": [], + "bbox": [ + 674, + 224, + 782, + 289 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bc3aca5efe34effc91a59b408995d2f6d8215a42242c2b2c256840fe86e8889d.jpg", + "image_caption": [ + "Q: How many sport facilities? A: 3", + "Fig. 2: Benchmark examples. For each specific field, one example image and its corresponding Q&A have been shown. Note that, for a simple presentation, the questions in Domestic Robot and Open Game have been simplified from multiple-choice format. Please see the Appendix for more examples and detailed questions." + ], + "image_footnote": [], + "bbox": [ + 217, + 306, + 336, + 380 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fcfe2f66f387cebd09afbfa3f4132f468b81318ffc366443c59cb035be9ee25b.jpg", + "image_caption": [ + "Q: What is the intended action for the main vehicle in an autonomous driving scenario? A: Go straight" + ], + "image_footnote": [], + "bbox": [ + 341, + 306, + 447, + 364 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6c1e9b620e6aa0a1779dd5688dbf400fe7ee8f0560c89be275330b62a4c20f5.jpg", + "image_caption": [ + "Q: Your goal is to repair the broken clock in your hands. What action should you do next? A: Look for repair tools" + ], + "image_footnote": [], + "bbox": [ + 450, + 306, + 558, + 364 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/70a1048286a32200221a0411d837f2edab99e2038afea1d54ccf1601a7916c4d.jpg", + "image_caption": [ + "Q: Your task is to harvest beef using the iron sword. Which action should you do next? A: Find cow" + ], + "image_footnote": [], + "bbox": [ + 562, + 311, + 669, + 364 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c43a8182b60d8a4722ea56871e70ed174d8c8d4083b159c15d4cff869937cbea.jpg", + "image_caption": [ + "Q: What is the defect type? A: Wire combined" + ], + "image_footnote": [], + "bbox": [ + 674, + 311, + 782, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-application Style Existing LMMs are mainly assessed in general application scenarios where questions can be answered using common knowledge [17, 21]. For instance, identifying the object on a table in an image can be accomplished with general knowledge. However, there exist specific application scenarios that demand specialized domain knowledge to address. For example, when an autonomous driving control center processes the road image to decide the next action: going straight or turning left, specific driving knowledge is required.", + "bbox": [ + 212, + 494, + 784, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As existing LMMs are primarily evaluated with questions and answers (Q&A) that rely on common knowledge, the extent to which an LMM can remain effective in situations requiring domain-specific knowledge has not been comprehensively explored. We also investigate this aspect in our proposed benchmark.", + "bbox": [ + 212, + 602, + 784, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Benchmark Construction", + "text_level": 1, + "bbox": [ + 215, + 689, + 468, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Artistic Style Data To construct a benchmark characterized by distributional shifts pertaining to artistic styles, we curated a diverse set of images encompassing Cartoon, Painting, Sketch, Handmade, and Tattoo styles from the COCO-O dataset [43]. We systematically sampled 100 images for each stylistic category, resulting in a total of 500 images extracted from the COCO-O dataset. This dataset comprises a rich array of object types, providing a comprehensive evaluation ground for assessing the image understanding capabilities of the existing LMMs. Accordingly, the questions are prepared as asking the models to reason", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BenchLMM", + "bbox": [ + 650, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and answer questions about images, and the sample images with Q&A are shown in the first row of Fig. 2.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sensor-style Data Other than RGB cameras, Infrared sensors, X-ray sensors, Magnetic Resonance Imaging (MRI) sensors, and Computed Tomography (CT) sensors are often used to collect image data for human beings to perceive and analyze. Our benchmark examines whether the LMMs can still have the comparative reasoning capability toward these images that are intrinsically different from RGB photo images. In the case of Infrared data, we collected a set of 200 images from the datasets that focus on street scenes analysis toward with pedestrians, buildings, and vehicles [22, 25, 26, 40]. Questions for LMMs mainly revolve around analyzing the contents of these street scenes. Specifically, we design some thermal-related questions such as determining the hottest content. An illustrative example is provided in the second row of Fig. 2. For Low-energy X-ray (L-Xray) data from the SiXray-D dataset [47], we gathered 200 images, including 160 with prohibited objects (knife, scissors, wrench, gun, hammer) and 40 without any prohibited objects. Questions are designed to assess the ability to detect security risks in X-ray images, such as identifying forbidden objects. A sample question is presented in the second row of Fig. 2. In the medical application, high-energy X-ray (H-Xray), MRI and CT scanning images (200 each) were extracted from the [37]. Questions for LMMs in this context involve tasks like recognizing organs or identifying diseases in the respective images. By comparing inference results between different sensor types, our benchmark aims to uncover challenges associated with cross-sensor analysis.", + "bbox": [ + 212, + 181, + 787, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Application-style Data In order to establish a comprehensive cross-application-style benchmark, we systematically integrated diverse task domains into our evaluation framework. We consider four representative applications: remote sensing, autonomous driving, agent action prediction, and defect detection. Specifically, for the remote sensing application, we utilized 200 images sourced from the DOTA dataset [59] and annotated them in the VQA format from a satellite perspective. For the autonomous driving task, we extracted 250 images from the BDD100K dataset [65], featuring various weather conditions. Our primary focus was on designing questions and answers aimed at predicting driving actions such as going straight, turning left, or turning right. An illustrative example is provided in the third row of Fig. 2.", + "bbox": [ + 212, + 503, + 789, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To introduce the agent action prediction application, we meticulously selected 100 images from the Domestic Robot dataset [11] and an additional 117 images from the Open Game dataset [11]. The associated questions center around predicting the next action to be taken under the current circumstances. Furthermore, for evaluating defect detection capabilities, we curated 100 images from the Defect Detection dataset [6], with questions focused on identifying defect types and determining the presence of defects. This diverse set of applications and datasets collectively forms a diverse benchmark, allowing for a comprehensive assessment of the proposed approach across various application domains. More examples and Q&A of different artistic styles, sensor styles, application styles, and the constrain of answers can be found in the Appendix.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Style Prompt Enhancement", + "text_level": 1, + "bbox": [ + 215, + 146, + 488, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is generally expected that additional prior information to LMMs' inputs can enhance performance. Nevertheless, such contextual priors (e.g. knowledge of artistic styles) might be unavailable in automated inference scenarios. We are inspired by human cognitive behavior that humans often comprehend the style of an image before delving into object recognition, akin to appreciating a painting by Van Gogh, where recognition of the impressionist style precedes identification of objects within the artwork. Other contextual priors, like sensor or application styles, can also enhance LMMs' performance. For example, identifying people in infrared images is easier when LMMs know that the images are captured by infrared thermal cameras since humans typically show consistent temperatures.", + "bbox": [ + 212, + 167, + 787, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With this insight, we investigate the efficacy of prompting LMMs to predict the style of an image before addressing associated questions. To this end, we devised Style Prompt Enhancement by introducing a text prompt, instructing the model to \"Determine the style of the image before answering the question.\" SPE aims to leverage stylistic cues as a precursor to visual question answering, potentially enhancing the model's capability of visual reasoning. Despite its simplicity, SPE's versatility allows it to be applied directly to images of various styles without any additional prompt tuning or model training, clearly boosting the overall performance of LMMs in a zero-shot manner, as detailed in Section 4.3.", + "bbox": [ + 212, + 319, + 787, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 474, + 366, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 503, + 428, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This paper evaluates a selection of existing open-source LMMs, which include LLaVA [38,39], MiniGPT4 [10,71], InstructBLIP [14], Otter [32], and OpenFlamingo-7B [4]. These models are initialized by pre-trained weights and evaluated on our BenchLMM. In addition, we use the variants of these models, where the language encoders and vision encoders are changed to investigate how different encoders have impacts on the cross-field capabilities. Moreover, we perform a quantitative evaluation of GPT-4V, the latest commercial LMM, with our dataset and compare its performance with the aforementioned open-source LMMs, and such a quantitative evaluation is the first time.", + "bbox": [ + 212, + 526, + 813, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our performance evaluation, we adhere to previous works [38,39] methodology, employing the ChatGPT API to gauge the proximity of answers predicted by the LMMs to ground-truth answers. Here, we designate the output of the LMM as $y$ , and the ground-truth label as $\\hat{y}$ . The evaluation utilizes the prompt \"Compare the similarity between $y$ and $\\hat{y}$ with a correctness score ranging from 0.0 (totally wrong) to 1.0 (totally right). The middle score provides the percentage of correctness.\"", + "bbox": [ + 212, + 662, + 787, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Evaluations of public LMMs", + "text_level": 1, + "bbox": [ + 215, + 787, + 496, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-artistic-style Capability In most existing works [17,21,33,34,41,67], LMMs are predominantly evaluated using images in the 'Photo' style, leading", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BenchLMM", + "bbox": [ + 650, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7d1163ce2f891bd85640af0d0a4a2fcc0622195f15f14b9fc35c4912f27e3434.jpg", + "table_caption": [ + "Table 1: Experimental results (\\%) of public LMMs on BenchLMM's cross-artistic-style evaluation. Note that Average* represents the average accuracy computed over five artistic-style benchmarks. The best and the second best results are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
LMMsVisual EncoderLanguage ModelCommonCartoonHandmakePaintingSketchTattooAverage*
GPT-4V [62]--81.563.358.758.264.768.362.6
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B74.662.056.657.663.357.059.3
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B72.759.057.159.357.561.458.9
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B56.649.945.448.053.747.048.8
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B56.957.828.437.842.632.340.0
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B71.251.953.343.462.054.653.0
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B73.958.152.055.855.255.555.3
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B66.428.737.840.941.728.635.5
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B45.635.037.038.536.437.736.9
Otter-7B [32]CLIP-ViT-L/14MPT-7B64.845.149.039.844.142.444.1
", + "bbox": [ + 217, + 212, + 784, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to a gap in understanding their performance across diverse artistic styles. We extend the evaluation scope by examining LMMs' performance with various artistic styles beyond the common 'Photo' style. The results detailed in Table 1 reveal a notable decline in LMMs' effectiveness when processing these artistic styles. This trend suggests a potential overfitting of LMMs to the 'Photo' style, highlighting their limited adaptability to varied artistic styles, a capability that humans typically possess. Interestingly, GPT-4V, as an outstanding commercial model, exhibits similar limitations in handling diverse styles.", + "bbox": [ + 212, + 411, + 784, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides, LLaVA-1.5-13B [38] is improved from LLaVA-13B [39] by introducing more training tricks and data, and LLaVA-1.5-13B outperforms LLaVA-13B by a clear margin on the 'Common' and other artistic styles. This comparison reveals the importance of data and training strategy as LLaVA-1.5-13B and LLaVA-13B use the same visual encoder and language encoder. However, we also find that a better performance on 'Common' does not necessarily indicate a better performance on other styles. For example, compared with InstructBLIP-7B, InstructBLIP-13B demonstrates better performance across five artistic-style benchmarks in terms of average accuracy, though it shows comparatively lower performance in the 'Common' style. Therefore, such results convey that it is important to conduct comprehensive evaluations to fully validate an LMM's capability rather than only focus on the 'Common' style, which shows the importance of our proposed benchmark.", + "bbox": [ + 212, + 534, + 787, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-sensor Capability We benchmark the LMM models' cross-sensor capability in Table 2. Because GPT-4V denied to provide medical diagnostics or identify potential abnormalities in medical images like H-Xray/MRI/CT, GPT-4V is only evaluated on the Infrared and L-Xray sensor data and achieves the results of $55.0\\%$ and $50.0\\%$ respectively, which are far lower than the 'Common' style $(81.5\\%)$ . Meanwhile, other LMM models' performance on images captured by X-ray, MRI, CT, and Infrared dropped significantly compared to", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e38eb3c5a4ca2099a97fdc2721a964d08718652238077114e31beb44da21541b.jpg", + "table_caption": [ + "Table 2: Experimental results (\\%) of public LMMs on BenchLMM's cross-sensor evaluation. Note that GPT-4V denied to test medical images (H-Xray/MRI/CT)." + ], + "table_footnote": [], + "table_body": "
LMMsVisual EncoderLanguage ModelInfraredL-XrayH-XrayMRICTAverage
GPT-4V [62]--55.050.0----
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B52.151.555.933.743.147.3
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B40.824.656.620.943.437.3
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B53.746.039.128.439.741.4
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B32.947.224.625.323.230.6
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B46.553.249.429.646.845.3
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B42.235.459.519.438.539.0
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B30.143.031.517.417.828.2
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B35.144.123.716.815.327.0
Otter-7B [32]CLIP-ViT-L/14MPT-7B30.246.439.147.846.342.0
", + "bbox": [ + 218, + 185, + 782, + 357 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f7f000a2f997f7438533b1f5bb4116b97815426f785f7b809a8493fb06e1fa81.jpg", + "table_caption": [ + "Table 3: Results of public LMMs on BenchLMM's cross-application experiments." + ], + "table_footnote": [], + "table_body": "
LMMsVisual EncoderLanguage ModelRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
GPT-4V [62]--69.735.856.074.064.460.0
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B65.628.649.021.460.345.2
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B63.632.632.019.547.038.9
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B38.731.759.043.828.240.3
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B33.322.648.034.231.734.2
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B61.731.344.015.459.642.2
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B63.536.742.024.830.839.6
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B34.517.640.022.218.826.6
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B27.920.841.040.113.628.7
Otter-7B [32]CLIP-ViT-L/14MPT-7B44.829.736.018.827.331.3
", + "bbox": [ + 218, + 398, + 782, + 550 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the performance on 'Common' captured by the RGB camera sensor. Therefore, the intrinsic difference in the imaging process between RGB cameras and other sensors is also a crucial challenge to LMMs. With respect to the MRI, H-Xray, and CT images, the questions may require specific medical knowledge to answer, and it may be argued that the performance drop is due to the knowledge gap. It is also observed from the L-Xray, where the objects (i.e., knife, gun, hammer, etc.) can be recognized with only common knowledge, but the LMMs' performance on the L-Xray data is significantly lower than 'Common', which further validates the challenge brought by the sensor difference.", + "bbox": [ + 212, + 582, + 784, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cross-application Capability In Table 3, we evaluate the different LMMs on various applications, finding that GPT-4V generally outperforms others, except in autonomous driving where its accuracy drops to $35.8\\%$ . This decrease is likely due to the challenges in the Autonomous Driving dataset, such as recognizing low-resolution traffic signs and requiring specific driving knowledge, where both GPT-4 and open-source LMMs have poor performance. We also observe that LLaVA-1.5-13B achieves comparable performance as GPT-4V on the Remote Sensing and Defect Detection datasets, which seems to have comparable cross", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BenchLMM", + "bbox": [ + 650, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0380a6083d16477095a416ccc400a78abf498111c561d5e00f8b6ab19fe86115.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 147, + 413, + 349 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/33142e0e9e08c940bc339e886bfcaeb3340de61d4137574d2f0538bb87d58cc8.jpg", + "image_caption": [ + "Fig. 3: Error-reflection capability comparisons: Unlike LLaVA-1.5-13B, GPT-4V possesses the ability to analyze its errors when provided with the correct answers." + ], + "image_footnote": [], + "bbox": [ + 251, + 351, + 413, + 551 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6c1d652eb59e200fbdb12562d24478fcce6e20ee556b1a42bc31d72aec8d565d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 147, + 584, + 349 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/99c115e4e8592c9be0f536820e8b711492a2e77aae19d67dd415162189b0dd35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 351, + 584, + 551 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8014eb06bb66a4f626987b8bb87350790aee9126fa7e82f4f6af747800e7a82c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 147, + 754, + 349 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4bb714b1b0881729cff988e04b0136ebc29c150e77483af4875789a93dcd5f8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 351, + 754, + 551 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "application capability. However, we surprisingly find that LLaVA-1.5-13B merely achieves $21.4\\%$ but GPT-4V achieves up to $74.0\\%$ . As GPT-4V is close-source, the technical details are not disclosed and we conjecture that the training data for GPT-4V may involve data similar to the Open Game dataset.", + "bbox": [ + 212, + 632, + 787, + 694 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conclude from the experiments that: 1) Existing LMMs often degrade in performance with cross-style images; 2) Better performance on 'Common' data doesn't ensure better results on other styles; 3) Comprehensive evaluations across different dimensions are crucial, as errors in tasks like medical or security can lead to significant losses", + "bbox": [ + 212, + 698, + 787, + 773 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Error-reflection Capability To understand the LMMs further, we let the LMM do an error reflection that uses itself to parse the error information. When a model gives wrong answers, we warp up the conversation: the previous question input and predicted answer output. Specifically, we add \"The correct answer is", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/335f61c177c312e9339a388afc68644e84cd032d9afe64d52b928fea9b40aee6.jpg", + "table_caption": [ + "Table 4: Experimental results (%) of our proposed SPE on the BenchLMM benchmark." + ], + "table_footnote": [], + "table_body": "
LMMsCartoonHandmakePaintingSketchTattooAverage
LLaVA-1.5-13B [38]62.056.657.663.357.059.3
+ SPE (Ours)67.657.767.968.863.565.1 (+5.8)
InstructBLIP-13B [14]59.057.159.357.561.458.9
+ SPE (Ours)61.162.364.860.963.362.5 (+3.6)
LMMsInfraredL-XrayH-XrayMRICTAverage
LLaVA-1.5-13B [38]52.151.555.933.743.147.3
+ SPE (Ours)58.762.856.233.647.751.8 (+4.5)
InstructBLIP-13B [14]40.824.656.620.943.437.3
+ SPE (Ours)43.524.857.725.042.638.7 (+1.4)
LMMsRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
LLaVA-1.5-13B [38]65.628.649.021.460.345.2
+ SPE (Ours)68.029.451.022.761.746.6 (+1.4)
InstructBLIP-13B [14]63.632.632.019.547.038.9
+ SPE (Ours)63.538.831.019.049.440.3 (+1.4)
", + "bbox": [ + 246, + 171, + 754, + 431 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[GT Answer]. Explain why your answer is incorrect. Use a single sentence.\" to the conversation (GT means Ground-truth), as the new input of LMMs.", + "bbox": [ + 212, + 459, + 784, + 489 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Although both GPT-4V and LLaVA-1.5-13B produce incorrect answers, GPT-4V demonstrates the ability to reason the correct answer through a detailed inference process. As illustrated in Fig. 3, GPT-4V can offer insightful analysis beyond the ground-truth answer, including the positions of clock hands for estimating time, the location of prohibited objects, and the traffic signals for driving planning. Moreover, from the second row of Fig. 3, we can see that the LLaVA model can just simply repeat the answer or say \"I was wrong\". By contrast, GPT-4V can do inference itself to provide the correct answer, such as recognizing the toy animals (Handmake) or the scratch on the capsule (Defect Detection). Through this study, we can see that the error-reflection capability of LLaVA-1.5-13B is inferior to that of GPT-4V. Therefore, we reveal that another important capability of a large multimodal model is whether the model can reflect how to derive the correct answer. Studying such capability is useful to common users. For example, when users use LMMs for self-learning, the user can provide reference answers and LMMs can provide a more detailed inference process, which can improve the learning effect. Moreover, the answer of GPT-4V can also be useful materials used to train open-source LMMs in the future.", + "bbox": [ + 212, + 489, + 787, + 746 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Experiments of Style Prompt Enhancement", + "text_level": 1, + "bbox": [ + 215, + 768, + 620, + 784 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison to Baseline Models In this section, we aim to study if a large multimodal model can be enhanced by prompting the model to estimate the style first and then answer the question, based on which we propose the Style Prompt", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BenchLMM", + "bbox": [ + 650, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/017ea27847e70f4b877715c1df774394501c381fd4163d5fd2b95f86f64aaf2c.jpg", + "table_caption": [ + "Table 5: Complementary studies of our SPE with recent zero-shot chain-of-thought method ZS-COT [31] using LLaVA-1.5-13B [38] on the BenchLMM" + ], + "table_footnote": [], + "table_body": "
MethodCartoonHandmakePaintingSketchTattooAverage
LLaVA-1.5-13B [38]62.056.657.663.357.059.3
+ZS-COT [31]64.359.770.662.871.065.7(+6.4)
+ZS-COT [31]+SPE (Ours)66.064.466.565.468.366.1(+6.8)
MethodInfraredL-XrayH-XrayMRICTAverage
LLaVA-1.5-13B [38]52.151.555.933.743.147.3
+ZS-COT [31]49.355.957.635.644.748.6(+1.3)
+ZS-COT [31]+SPE (Ours)54.470.456.735.345.452.4(+5.1)
MethodRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
LLaVA-1.5-13B [38]65.628.649.021.460.345.2
+ZS-COT [31]65.527.043.034.263.746.7(+1.5)
+ZS-COT [31]+SPE (Ours)76.024.049.052.160.152.2(+7.0)
", + "bbox": [ + 236, + 184, + 764, + 378 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Enhancement (SPE) method and conduct experiments with our BenchLMM. Specifically, we examined our SPE with LLaVA-1.5-13B [38] and InstructBLIP-13B [14] because these two models are representative of better average performance than MiniGPT4 [71] and Otter [32].", + "bbox": [ + 212, + 420, + 782, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The experimental results are presented in Table 4. As shown in the table, our proposed SPE can benefit the LMMs by achieving general improvement over different styles, which means that reasoning the style first is an effective prompting method to improve LMMs' visual reasoning capability. Besides, in the experiments of crossing to different applications (Remote Sensing, Autonomous Driving, etc.), we observe that its benefit is smaller than others. For example, LLaVA-1.5-13B gains $5.8\\%$ and $4.5\\%$ improvement in the experiments of cross-artistic and cross-sensor styles, while the improvement is merely $1.4\\%$ when crossing applications. We conjecture that specific application knowledge could inferred from only the style, which limits the benefits of SPE. Moreover, we observe that LLaVA-1.5-13 generally obtains more benefits than InstructBLIP-13B, which means that the stronger the LMM, the more benefits from SPE, which is because a stronger LMM has stronger inference capability and can better utilize prior information.", + "bbox": [ + 212, + 486, + 787, + 698 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison to Other Prompting Methods Our proposed SPE method enhances LMMs' performance by prompting LMMs to estimate the input image's style first, a technique akin to chain-of-thought prompting approaches [31, 45]. In Table 5, we compare our proposed SPE with the state-of-the-art prompting Chain-of-Thought (COT) [31], which is a zero-shot, to demonstrate the efficacy of SPE. As shown, the COT method can improve LMM's performance, and our SPE is compatible with the COT methods and further enhances the model's capability, which is demonstrated by the general improvement brought by “+ZS-COT+SPE” as shown in Table 5. We also notice that in some cases, using SPE", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e340eaf9b742dc6c007d5b6aba0d1180d31010fd95faf817f5e5c20210003519.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons between LLaVA-1.5-13B and LLaVA-1.5-13B + SPE (Ours) on style-transfer images." + ], + "image_footnote": [], + "bbox": [ + 263, + 143, + 743, + 244 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7680074573adb13986bd03cd08807ada04811168e525ea129afe69db818081fa.jpg", + "table_caption": [ + "Table 6: Quantitative Results of our proposed SPE on style-transfer images." + ], + "table_footnote": [], + "table_body": "
LMMs|Cartoon Handmake Painting Sketch Tattoo | Average
LLaVA-1.5-13B [38] + SPE (Ours)62.161.152.468.841.257.1
67.462.159.469.445.560.8 (+3.7)
InstructBLIP-13B [14] + SPE (Ours)69.370.861.353.865.064.0
73.071.664.262.165.367.2 (+3.2)
", + "bbox": [ + 240, + 325, + 761, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "with COT could lead to performance decline. The investigation is beyond the scope of this paper, which will be explored in our future work.", + "bbox": [ + 212, + 439, + 784, + 468 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results on the Style-transfer Images We employ the latest commercial style transfer model, DALL-E3 [50], to transfer the 'Common' style images into different artistic styles. As depicted in Fig. 4, these images share similar structures and layouts. However, they are just in different artistic styles, and the tested LLaVA model also makes mistakes. The significant performance degradation on the transferred images is also proved by the quantitative results presented in Table 6, which validates that LMMs suffer from difficulties in recognizing images of different styles. Nevertheless, our SPE significantly enhances the cross-style capability of LMMs in this challenging task, as shown in Fig. 4 and Table 6.", + "bbox": [ + 212, + 470, + 784, + 604 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In-depth Analysis of SPE Besides quantitative assessments, we explore the inner workings of SPE to understand how it improves LMMs. As illustrated in Fig. 5, LMMs are adept at accurately estimating the style of images. For instance, in the figure's first row and second column, the depicted style is 'Handmake,' and the model accurately recognizes this style as akin to a toy, indicating a correct match. Moreover, when estimating the style, the model can also understand the context first and thus better answer the final question, as shown in the examples of Remote Sensing and Defect Detection. Our findings suggest that prompting LMMs to first identify the style of images serves as an effective, training-free method to improve the models' proficiency in handling cross-style images.", + "bbox": [ + 212, + 606, + 784, + 757 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 779, + 357, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this work, we introduce BenchLMM, a novel benchmark for quantitatively evaluating large multimodal models (LMMs) across varied visual distribution", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BenchLMM", + "bbox": [ + 650, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1135ef199bff7c35cebd55bcacb9ad3c2f3ea7ba6ad9e4f23730986f4e29f609.jpg", + "image_caption": [ + "Fig. 5: Evaluation of LLaVA-1.5-13B with SPE on the BenchLMM." + ], + "image_footnote": [], + "bbox": [ + 243, + 146, + 754, + 621 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "shifts, including artistic, sensor, and application styles. Our evaluations of numerous existing LMMs highlight two key findings: 1) LMMs typically underperform with non-common image styles. 2) Our style prompt enhancement approach, inspired by human perception of art, significantly improves LMMs' visual reasoning without extra fine-tuning. Additionally, our error-reflection study shows that stronger LMMs can self-diagnose errors by providing insights not given by humans, whereas less capable models merely restate correct answers without new insights. Our benchmark offers a comprehensive tool for LMM research, emphasizing the importance of error-reflection capabilities in the future.", + "bbox": [ + 212, + 688, + 787, + 825 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was carried out at Rapid-Rich Object Search (ROSE) Lab, School of Electrical & Electronic Engineering, Nanyang Technological University. This research is supported by the NTU-PKU Joint Research Institute (a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation.", + "bbox": [ + 212, + 175, + 787, + 251 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 273, + 321, + 287 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8948-8957 (2019)", + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015)", + "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)", + "5. Ben-Younes, H., Cadene, R., Cord, M., Thome, N.: Multimodal tucker fusion for visual question answering. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2612-2620 (2017)", + "6. Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., Steger, C.: The mvtec anomaly detection dataset: a comprehensive real-world dataset for unsupervised anomaly detection. International Journal of Computer Vision 129(4), 1038-1059 (2021)", + "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "8. Cadene, R., Ben-Younes, H., Cord, M., Thome, N.: Murel: Multimodal relational reasoning for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1989-1998 (2019)", + "9. Cao, Q., Wan, W., Wang, K., Liang, X., Lin, L.: Linguistically routing capsule network for out-of-distribution visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1614-1623 (2021)", + "10. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning (2023)", + "11. Chen, L., Zhang, Y., Ren, S., Zhao, H., Cai, Z., Wang, Y., Wang, P., Liu, T., Chang, B.: Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4-vision and beyond. arXiv preprint arXiv:2310.02071 (2023)" + ], + "bbox": [ + 217, + 301, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BenchLMM", + "bbox": [ + 650, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollár, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)", + "13. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)", + "14. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)", + "15. Dancette, C., Cadene, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1574-1583 (2021)", + "16. Ding, Y., Yu, J., Liu, B., Hu, Y., Cui, M., Wu, Q.: Mukea: Multimodal knowledge extraction and accumulation for knowledge-based visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5089-5098 (2022)", + "17. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models. arXiv preprint arXiv:2306.13394 (2023)", + "18. Fukui, A., Park, D.H., Yang, D., Rohrbach, A., Darrell, T., Rohrbach, M.: Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847 (2016)", + "19. Gao, F., Ping, Q., Thattai, G., Reganti, A., Wu, Y.N., Natarajan, P.: Transform-retrieve-produce: Natural language-centric outside-knowledge visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5067-5077 (2022)", + "20. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhan, S.: Chain of thought prompt tuning in vision language models. arXiv preprint arXiv:2304.07919 (2023)", + "21. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6904-6913 (2017)", + "22. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)", + "23. Han, X., Wang, S., Su, C., Huang, Q., Tian, Q.: Greedy gradient ensemble for robust visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1584-1593 (2021)", + "24. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019)", + "25. Hwang, S., Park, J., Kim, N., Choi, Y., So Kweon, I.: Multispectral pedestrian detection: Benchmark dataset and baseline. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1037-1045 (2015)", + "26. Jia, X., Zhu, C., Li, M., Tang, W., Zhou, W.: LLVIP: A visible-infrared paired dataset for low-light vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3496-3504 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2901-2910 (2017)", + "28. Johnson, J., Hariharan, B., Van Der Maaten, L., Hoffman, J., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Inferring and executing programs for visual reasoning. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2989-2998 (2017)", + "29. Kim, J.H., Jun, J., Zhang, B.T.: Bilinear attention networks. Advances in neural information processing systems 31 (2018)", + "30. Kim, J.H., On, K.W., Lim, W., Kim, J., Ha, J.W., Zhang, B.T.: Hadamard product for low-rank bilinear pooling. arXiv preprint arXiv:1610.04325 (2016)", + "31. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. Advances in neural information processing systems 35, 22199-22213 (2022)", + "32. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)", + "33. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)", + "34. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)", + "35. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)", + "36. Liang, P.P., Lyu, Y., Fan, X., Wu, Z., Cheng, Y., Wu, J., Chen, L., Wu, P., Lee, M.A., Zhu, Y., et al.: Multibench: Multiscale benchmarks for multimodal representation learning. arXiv preprint arXiv:2107.07502 (2021)", + "37. Liu, B., Zhan, L.M., Xu, L., Ma, L., Yang, Y., Wu, X.M.: Slake: A semantically-labeled knowledge-enhanced dataset for medical visual question answering. In: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). pp. 1650-1654. IEEE (2021)", + "38. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)", + "39. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "40. Liu, J., Fan, X., Huang, Z., Wu, G., Liu, R., Zhong, W., Luo, Z.: Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5802-5811 (June 2022)", + "41. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: MMBench: Is Your Multi-modal Model an All-around Player? arXiv preprint arXiv:2307.06281 (2023)", + "42. Manyika, J.: An overview of bard: an early experiment with generative ai. AI. Google Static Documents (2023)", + "43. Mao, X., Chen, Y., Zhu, Y., Chen, D., Su, H., Zhang, R., Xue, H.: COCO-O: A Benchmark for Object Detectors under Natural Distribution Shifts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6339-6350 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BenchLMM", + "bbox": [ + 650, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3195-3204 (2019)", + "45. Mitra, C., Huang, B., Darrell, T., Herzig, R.: Compositional chain-of-thought prompting for large multimodal models. arXiv preprint arXiv:2311.17076 (2023)", + "46. Nam, H., Ha, J.W., Kim, J.: Dual attention networks for multimodal reasoning and matching. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 299-307 (2017)", + "47. Nguyen, H.D., Cai, R., Zhao, H., Kot, A.C., Wen, B.: Towards more efficient security inspection via deep learning: a task-driven x-ray image cropping scheme. Micromachines 13(4), 565 (2022)", + "48. OpenAI, R.: Gpt-4 technical report. arxiv 2303.08774. View in Article (2023)", + "49. Shah, M., Chen, X., Rohrbach, M., Parikh, D.: Cycle-consistency for robust visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6649-6658 (2019)", + "50. Shi, Z., Zhou, X., Qiu, X., Zhu, X.: Improving image captioning with better use of captions. arXiv preprint arXiv:2006.11807 (2020)", + "51. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 742-758. Springer (2020)", + "52. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards VQA models that can read. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8317-8326 (2019)", + "53. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019)", + "54. Tascon-Morales, S., Márquez-Neila, P., Sznitman, R.: Logical implications for visual question answering consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6725-6735 (2023)", + "55. Teney, D., Abbasnejad, E., van den Hengel, A.: Unshuffling data for improved generalization in visual question answering. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1417-1427 (2021)", + "56. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "57. Tsimpoukelli, M., Menick, J.L., Cabi, S., Eslami, S., Vinyals, O., Hill, F.: Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems 34, 200-212 (2021)", + "58. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "59. Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: Dota: A large-scale dataset for object detection in aerial images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3974-3983 (2018)", + "60. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Yang, X., Gao, C., Zhang, H., Cai, J.: Auto-parsing network for image captioning and visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2197-2207 (2021)", + "62. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of LMMs: Preliminary explorations with GPT-4V (ision). arXiv preprint arXiv:2309.17421 (2023)", + "63. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023)", + "64. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2633-2642 (2020). https://doi.org/10.1109/CVPR42600.2020.00271", + "66. Yu, S., Wu, P., Liang, P.P., Salakhutdinov, R., Morency, L.P.: Pacs: A dataset for physical audiovisual commonsense reasoning. In: European Conference on Computer Vision. pp. 292-309. Springer (2022)", + "67. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)", + "68. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6720-6731 (2019)", + "69. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023)", + "70. Zhou, Y., Ren, T., Zhu, C., Sun, X., Liu, J., Ding, X., Xu, M., Ji, R.: Trar: Routing the attention spans in transformer for visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2074-2084 (2021)", + "71. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 604 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BenchLMM", + "bbox": [ + 650, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_model.json b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_model.json new file mode 100644 index 0000000000000000000000000000000000000000..865e2cdaa7e7b4f464ed38bfcdbb7e4b54104124 --- /dev/null +++ b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_model.json @@ -0,0 +1,2724 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "BenchLMM: Benchmarking Cross-style Visual Capability of Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.212, + 0.744, + 0.245 + ], + "angle": 0, + "content": "Rizhao Cai\\(^{1}\\), Zirui Song\\(^{1,2}\\), Dayan Guan\\(^{3}\\), Zhenhao Chen\\(^{4}\\), Yaohang Li\\(^{2}\\), Xing Luo\\(^{5}\\), Chenyu Yi\\(^{1}\\), and Alex Kot\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.255, + 0.66, + 0.27 + ], + "angle": 0, + "content": "1 Nanyang Technological University, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.269, + 0.762, + 0.283 + ], + "angle": 0, + "content": "2 University of Technology Sydney, Australia; Northeastern University, China;" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.282, + 0.636, + 0.297 + ], + "angle": 0, + "content": "3 Harbin Institute of Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.296, + 0.772, + 0.311 + ], + "angle": 0, + "content": "4 Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.31, + 0.6, + 0.325 + ], + "angle": 0, + "content": "5 Zhejiang University, China" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.327, + 0.702, + 0.339 + ], + "angle": 0, + "content": "† Equal Contribution Corresponding author (guandayan@hit.edu.cn)" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.371, + 0.744, + 0.621 + ], + "angle": 0, + "content": "Abstract. Large Multimodal Models (LMMs) such as GPT-4V and LLaVA have shown remarkable capabilities in visual reasoning on data in common image styles. However, their robustness against diverse style shifts, crucial for practical applications, remains largely unexplored. In this paper, we propose a new benchmark, BenchLMM, to assess the robustness of LMMs toward three different styles: artistic image style, imaging sensor style, and application style. Utilizing BenchLMM, we comprehensively evaluate state-of-the-art LMMs and reveal: 1) LMMs generally suffer performance degradation when working with other styles; 2) An LMM performs better than another model in common style does not guarantee its superior performance in other styles; 3) LMMs' reasoning capability can be enhanced by prompting LMMs to predict the style first, based on which we propose a versatile and training-free method for improving LMMs; 4) An intelligent LMM is expected to interpret the causes of its errors when facing stylistic variations. We hope that our benchmark and analysis can shed new light on developing more intelligent and versatile LMMs. The benchmark and evaluation have been released on https://github.com/AIFEG/BenchLMM." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.644, + 0.377, + 0.66 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.795 + ], + "angle": 0, + "content": "The dynamic landscape of computer vision has recently been shaped by the rise of Large Multimodal Models (LMMs). These models [10, 14, 32, 39, 71], incorporating visual and textual data, are pivotal in developing general-purpose assistants for visual reasoning, a widely-studied task that involves reasoning and answering questions based on images. Their growing prominence and the enthusiasm surrounding visual reasoning are evident from numerous studies such as GPT-4V [62]. As the interest in LMMs swells within the research community, understanding their depth, breadth, and limitations becomes imperative." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To comprehensively understand the visual reasoning capabilities of LMMs, numerous benchmarks [17, 21, 33, 34, 41, 67] have been proposed to systematically evaluate the linguistic capabilities of LMMs. These linguistic capabilities" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.146, + 0.361, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.226, + 0.296, + 0.253 + ], + "angle": 0, + "content": "Q: How many forks? \nA (GT): Two \nA (GPT-4V): Two \nA (LLaVA-1: 5) Two" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.146, + 0.501, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.226, + 0.483, + 0.253 + ], + "angle": 0, + "content": "[Painting]Q:How many elephants? A (GT): Seven A (GPT-4V):Five A(LLAVA-1.5):Five" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.146, + 0.64, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.226, + 0.637, + 0.254 + ], + "angle": 0, + "content": "[Xray] Q: What are prohibited objects? \nA (GT): None \nA (GPT-4V): Gun \nA (LLVaA-1.5): A knife and a scissor" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.147, + 0.78, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.22, + 0.778, + 0.255 + ], + "angle": 0, + "content": "[Driving] Q: What is the action for the ego-vehicle during autonomous driving? A (GT): Go straight A (GPT-4V): Stop at traffic light A (LLAvA-1.5): Stop" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.267, + 0.787, + 0.309 + ], + "angle": 0, + "content": "Fig. 1: The motivation of our work. State-of-the-art LMMs, such as GPT-4V and LLaVA-1.5 can understand the images in a common style. However, these models have difficulties in reasoning images in other styles and deliver wrong answers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.537 + ], + "angle": 0, + "content": "are assessed through various types of responses for visual reasoning tasks, including Yes/No answers [17], multi-choice answers [33,41], single word or phrase answers [21,24,44]. However, while these benchmarks investigate textual diversity, they may fall short in terms of visual diversity as most images in these benchmarks are photographic images sourced from the internet. Despite providing valuable insights into LMMs' performance, there remains a gap in understanding how LMMs handle shifts in visual distribution. Although the recently released GPT-4V [62] claims remarkable capabilities across diverse application scenarios, only a few samples of qualitative results are presented in the technical report [62], and the quantitative evaluations across different scenarios are missing. Moreover, according to our preliminary study as illustrated in Fig. 1, we reveal that LMMs, even the state-of-the-art GPT-4V, tend to make mistakes in reasoning images with non-common styles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.79, + 0.704 + ], + "angle": 0, + "content": "To further investigate LMMs' cross-style visual reasoning capability, we propose a novel benchmark BenchLMM, which systematically assesses existing LMMs in three distinct styles of distribution shifts: artistic styles, sensor styles, and application styles. Firstly, images of the common style are realistic, but images can be in other artistic styles, such as Cartoon, Painting, Sketch, etc. Besides, there are different imaging sensors other than RGB cameras, such as infrared and X-ray. The discrepancy of sensors induces the distribution shifts of images. Moreover, we also identify the shift of knowledge required for different application styles. For example, the image recognition application requires knowledge about object shape and appearance, while the application of decision-making for robot action requires application-specific knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.84 + ], + "angle": 0, + "content": "With our BenchLMM, we conduct in-depth analysis through experiments of evaluating advanced LMMs. The comprehensive experiments reveal that while these cutting-edge LMMs excel with commonly styled data, their performance significantly declines with data in other styles. Interestingly, we also discover that the superior performance of an LMM in common data styles does not necessarily predict better performance across other styles. Furthermore, our investigation shows that prompting an LMM to predict the style of the data before answering a question can significantly boost its reasoning abilities. Based on this insight, we introduce the Style Prompt Enhancement (SPE) technique. SPE is a novel" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.651, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BenchLMM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "approach that improves LMMs' performance without additional training and operates effectively in a zero-shot manner. Moreover, we analyze the failure cases by conducting an in-depth error-reflection analysis by asking an LMM to interpret why it made mistakes in the Q&A. We find that a more intelligent LMM can understand the detailed reason why it made mistakes and can learn to derive the correct answer, which is an important capability that deserves attention when developing LMMs. We hope that our benchmark and analysis can shed new light on developing more intelligent and versatile LMMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Our contributions can be summarized below: 1) BenchchLMM: We introduce BenchLMM, a pioneering benchmark specifically designed to quantitatively evaluate the performance of LMMs' visual reasoning capability across visual styles in artistic, sensor, and application shifts. 2) Comprehensive evaluation of cutting-edge LMMs: We conduct extensive quantitative benchmark experiments on LMMs, including the commercial GPT-4V. Our rigorous and quantitative evaluation reveals LMMs' limitations in versatility and performance across different data styles. This finding challenges the prevailing assumption of universal applicability and highlights the need for specialized strategies to enhance LMM adaptability. 3) Insightful Analysis and Enhancement Strategy: Beyond mere performance assessment, we conduct an in-depth error-reflection analysis, where LMMs are prompted to identify the reasons behind their errors in question-and-answer tasks. This study reveals that more sophisticated LMMs can recognize their mistakes and also learn from them to improve future responses. Moreover, we introduce a novel SPE technique, as a training-free and zero-shot strategy, to significantly boost LMMs' reasoning capabilities across varied data styles, marking a substantial advancement in LMM development." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.393, + 0.567 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Visual Reasoning Visual reasoning, also known as visual question answering (VQA), aims to reason and answer questions about visual information [3,27,28]. Early developments [5, 18, 30] focus on designing more complex fusion mechanisms instead of simple summation fusion [3]. By including modern attention mechanism [58], numerous works [8, 29, 46, 53] contributed to the creation of transformer-based vision-language models. Recent methods have proposed to improve other aspects of visual question answering, including avoiding shortcut learning [15, 23], addressing the linguistic bias [9, 55], improving transformer-based vision-language models [61, 70], external knowledge integration [16, 19], and consistency regularization [49, 54]. Lastly, the advent of large multimodal models has markedly enhanced the capabilities of visual reasoning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Large Multimodal Models Recent advancements in large language models (LLMs) [7, 13, 48, 56] have inspired the development of LMMs that enhance LLMs with vision-language capabilities. Existing LMMs can be categorized into three main classes: multimodal instruction tuning [10, 14, 32, 38, 39, 71] that involves the finetuning of pre-trained LLMs on multi-modal instruction-following data; multimodal in-context learning [2, 32, 57, 63] that derives insights from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Cai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "few-shot task-specific samples; multimodal chain-of-thought [20,69] that encourages LMMs to articulate both the final answer and the underlying reasoning process. Additionally, some commercial LMMs like Google's Bard [42] and Microsoft's GPT-4V [62], with detailed information about their architecture remaining undisclosed, have shown impressive visual capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.374 + ], + "angle": 0, + "content": "Evaluation of LMMs To evaluate their capabilities, existing LMMs have been tested on a number of vision-language datasets. These datasets focus on specific visual reasoning tasks by designing different kinds of Q&A task, and the specific task includes visual recognition [21, 44], image description [1, 12], scene text understanding [51, 52, 64], object hallucination evaluation [35], complex reasoning [39, 67], and commonsense reasoning [66, 68]. To facilitate a comprehensive comparison of LMMs, several benchmarks [17, 36, 60] have been established by encompassing various visual reasoning tasks with common-style images and diverse Q&A pairs. However, the above benchmarks do not consider different styles of visual distribution shift, which motivates us to propose our BenchLMM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.4, + 0.368, + 0.416 + ], + "angle": 0, + "content": "3 BenchLMM" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.434, + 0.353, + 0.449 + ], + "angle": 0, + "content": "3.1 Motivation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.784, + 0.52 + ], + "angle": 0, + "content": "LMMs (e.g., GPT-4V) are claimed to excel in various scenarios [62], shown with qualitative results. Nevertheless, we provide comprehensive and quantitative results experiments with our BenchLMM reveal a paradox. We will elaborate on the benchmark that encompasses different cross-style data as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.785, + 0.673 + ], + "angle": 0, + "content": "Cross-artistic Style Existing benchmarks for LMMs commonly use photo images for evaluation, which we name as 'Common' style. Meanwhile, there are diverse image styles, including Cartoon, Sketch, etc., as shown and compared in Fig. 2. The object in other artistic styles have different distributions of object characteristics. For instance, a colorful zebra, common in painting and tattoo styles, is rare in 'Common' style images. While humans easily interpret such characteristics, the extent of LMMs' understanding of the other styles remains uncertain. This uncertainty arises from potential insufficient training on datasets encompassing these diverse styles. Therefore, we propose an evaluation to assess LMMs' robustness in analyzing images across various artistic styles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Cross-sensor Style Photographic images are commonly obtained through RGB cameras for information sensing. Alongside RGB cameras, other sensor types are also important for capturing images. For instance, infrared cameras are frequently utilized in surveillance systems to capture images during nighttime, as RGB cameras rely on adequate illumination for optimal image quality. Additionally, X-ray sensors are used in airport security checks, enabling the detection of prohibited items within luggage. These sensors feature distinct imaging processes compared to RGB cameras. Consequently, images acquired through various sensors exhibit unique intrinsic properties. The question of whether the performance of LMMs remains remarkable with other sensor data has not been thoroughly investigated. This issue will be addressed in our benchmark study." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.651, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BenchLMM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.338, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.209, + 0.296, + 0.223 + ], + "angle": 0, + "content": "Q: What time is it? A: 2:57:39" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.146, + 0.449, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.208, + 0.439, + 0.225 + ], + "angle": 0, + "content": "A: A bronze sculpture of a person with a very large head" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.146, + 0.56, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.209, + 0.556, + 0.223 + ], + "angle": 0, + "content": "Q: Who is sitting on the couch? A: A woman" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.146, + 0.671, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.209, + 0.641, + 0.223 + ], + "angle": 0, + "content": "Q: What kind of sport? A: football" + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.146, + 0.74, + 0.153 + ], + "angle": 0, + "content": "Tattoo" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.153, + 0.783, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.209, + 0.757, + 0.223 + ], + "angle": 0, + "content": "Q: What is in the image? A: A colorful zebra" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.226, + 0.338, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.29, + 0.321, + 0.304 + ], + "angle": 0, + "content": "Q: How many pedestrians? A: 4" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.226, + 0.449, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.29, + 0.448, + 0.305 + ], + "angle": 0, + "content": "Q: What are prohibited objects? \nA: Gun, knife" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.226, + 0.552, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.29, + 0.543, + 0.305 + ], + "angle": 0, + "content": "Q: What is the main organ? \nA: Lung" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.226, + 0.671, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.29, + 0.658, + 0.305 + ], + "angle": 0, + "content": "Q: What kind of disease? A: Brain edema, brain tumor" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.226, + 0.783, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.291, + 0.759, + 0.305 + ], + "angle": 0, + "content": "Q: What kind of disease? \nA: Lung cancer" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.307, + 0.338, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.382, + 0.331, + 0.395 + ], + "angle": 0, + "content": "Q: How many sport facilities? A: 3" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.307, + 0.449, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.367, + 0.442, + 0.396 + ], + "angle": 0, + "content": "Q: What is the intended action for the main vehicle in an autonomous driving scenario? A: Go straight" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.307, + 0.56, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.367, + 0.56, + 0.396 + ], + "angle": 0, + "content": "Q: Your goal is to repair the broken clock in your hands. What action should you do next? A: Look for repair tools" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.312, + 0.671, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.367, + 0.666, + 0.395 + ], + "angle": 0, + "content": "Q: Your task is to harvest beef using the iron sword. Which action should you do next? A: Find cow" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.312, + 0.783, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.381, + 0.765, + 0.395 + ], + "angle": 0, + "content": "Q: What is the defect type? A: Wire combined" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.408, + 0.785, + 0.464 + ], + "angle": 0, + "content": "Fig. 2: Benchmark examples. For each specific field, one example image and its corresponding Q&A have been shown. Note that, for a simple presentation, the questions in Domestic Robot and Open Game have been simplified from multiple-choice format. Please see the Appendix for more examples and detailed questions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.785, + 0.602 + ], + "angle": 0, + "content": "Cross-application Style Existing LMMs are mainly assessed in general application scenarios where questions can be answered using common knowledge [17, 21]. For instance, identifying the object on a table in an image can be accomplished with general knowledge. However, there exist specific application scenarios that demand specialized domain knowledge to address. For example, when an autonomous driving control center processes the road image to decide the next action: going straight or turning left, specific driving knowledge is required." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.785, + 0.663 + ], + "angle": 0, + "content": "As existing LMMs are primarily evaluated with questions and answers (Q&A) that rely on common knowledge, the extent to which an LMM can remain effective in situations requiring domain-specific knowledge has not been comprehensively explored. We also investigate this aspect in our proposed benchmark." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.469, + 0.705 + ], + "angle": 0, + "content": "3.2 Benchmark Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Artistic Style Data To construct a benchmark characterized by distributional shifts pertaining to artistic styles, we curated a diverse set of images encompassing Cartoon, Painting, Sketch, Handmade, and Tattoo styles from the COCO-O dataset [43]. We systematically sampled 100 images for each stylistic category, resulting in a total of 500 images extracted from the COCO-O dataset. This dataset comprises a rich array of object types, providing a comprehensive evaluation ground for assessing the image understanding capabilities of the existing LMMs. Accordingly, the questions are prepared as asking the models to reason" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Cai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "and answer questions about images, and the sample images with Q&A are shown in the first row of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.182, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Sensor-style Data Other than RGB cameras, Infrared sensors, X-ray sensors, Magnetic Resonance Imaging (MRI) sensors, and Computed Tomography (CT) sensors are often used to collect image data for human beings to perceive and analyze. Our benchmark examines whether the LMMs can still have the comparative reasoning capability toward these images that are intrinsically different from RGB photo images. In the case of Infrared data, we collected a set of 200 images from the datasets that focus on street scenes analysis toward with pedestrians, buildings, and vehicles [22, 25, 26, 40]. Questions for LMMs mainly revolve around analyzing the contents of these street scenes. Specifically, we design some thermal-related questions such as determining the hottest content. An illustrative example is provided in the second row of Fig. 2. For Low-energy X-ray (L-Xray) data from the SiXray-D dataset [47], we gathered 200 images, including 160 with prohibited objects (knife, scissors, wrench, gun, hammer) and 40 without any prohibited objects. Questions are designed to assess the ability to detect security risks in X-ray images, such as identifying forbidden objects. A sample question is presented in the second row of Fig. 2. In the medical application, high-energy X-ray (H-Xray), MRI and CT scanning images (200 each) were extracted from the [37]. Questions for LMMs in this context involve tasks like recognizing organs or identifying diseases in the respective images. By comparing inference results between different sensor types, our benchmark aims to uncover challenges associated with cross-sensor analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.79, + 0.671 + ], + "angle": 0, + "content": "Application-style Data In order to establish a comprehensive cross-application-style benchmark, we systematically integrated diverse task domains into our evaluation framework. We consider four representative applications: remote sensing, autonomous driving, agent action prediction, and defect detection. Specifically, for the remote sensing application, we utilized 200 images sourced from the DOTA dataset [59] and annotated them in the VQA format from a satellite perspective. For the autonomous driving task, we extracted 250 images from the BDD100K dataset [65], featuring various weather conditions. Our primary focus was on designing questions and answers aimed at predicting driving actions such as going straight, turning left, or turning right. An illustrative example is provided in the third row of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To introduce the agent action prediction application, we meticulously selected 100 images from the Domestic Robot dataset [11] and an additional 117 images from the Open Game dataset [11]. The associated questions center around predicting the next action to be taken under the current circumstances. Furthermore, for evaluating defect detection capabilities, we curated 100 images from the Defect Detection dataset [6], with questions focused on identifying defect types and determining the presence of defects. This diverse set of applications and datasets collectively forms a diverse benchmark, allowing for a comprehensive assessment of the proposed approach across various application domains. More examples and Q&A of different artistic styles, sensor styles, application styles, and the constrain of answers can be found in the Appendix." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.651, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BenchLMM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.49, + 0.162 + ], + "angle": 0, + "content": "3.3 Style Prompt Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.788, + 0.32 + ], + "angle": 0, + "content": "It is generally expected that additional prior information to LMMs' inputs can enhance performance. Nevertheless, such contextual priors (e.g. knowledge of artistic styles) might be unavailable in automated inference scenarios. We are inspired by human cognitive behavior that humans often comprehend the style of an image before delving into object recognition, akin to appreciating a painting by Van Gogh, where recognition of the impressionist style precedes identification of objects within the artwork. Other contextual priors, like sensor or application styles, can also enhance LMMs' performance. For example, identifying people in infrared images is easier when LMMs know that the images are captured by infrared thermal cameras since humans typically show consistent temperatures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.456 + ], + "angle": 0, + "content": "With this insight, we investigate the efficacy of prompting LMMs to predict the style of an image before addressing associated questions. To this end, we devised Style Prompt Enhancement by introducing a text prompt, instructing the model to \"Determine the style of the image before answering the question.\" SPE aims to leverage stylistic cues as a precursor to visual question answering, potentially enhancing the model's capability of visual reasoning. Despite its simplicity, SPE's versatility allows it to be applied directly to images of various styles without any additional prompt tuning or model training, clearly boosting the overall performance of LMMs in a zero-shot manner, as detailed in Section 4.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.476, + 0.367, + 0.494 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.505, + 0.429, + 0.521 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.814, + 0.662 + ], + "angle": 0, + "content": "This paper evaluates a selection of existing open-source LMMs, which include LLaVA [38,39], MiniGPT4 [10,71], InstructBLIP [14], Otter [32], and OpenFlamingo-7B [4]. These models are initialized by pre-trained weights and evaluated on our BenchLMM. In addition, we use the variants of these models, where the language encoders and vision encoders are changed to investigate how different encoders have impacts on the cross-field capabilities. Moreover, we perform a quantitative evaluation of GPT-4V, the latest commercial LMM, with our dataset and compare its performance with the aforementioned open-source LMMs, and such a quantitative evaluation is the first time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.769 + ], + "angle": 0, + "content": "In our performance evaluation, we adhere to previous works [38,39] methodology, employing the ChatGPT API to gauge the proximity of answers predicted by the LMMs to ground-truth answers. Here, we designate the output of the LMM as \\( y \\), and the ground-truth label as \\( \\hat{y} \\). The evaluation utilizes the prompt \"Compare the similarity between \\( y \\) and \\( \\hat{y} \\) with a correctness score ranging from 0.0 (totally wrong) to 1.0 (totally right). The middle score provides the percentage of correctness.\"" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.498, + 0.803 + ], + "angle": 0, + "content": "4.2 Evaluations of public LMMs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Cross-artistic-style Capability In most existing works [17,21,33,34,41,67], LMMs are predominantly evaluated using images in the 'Photo' style, leading" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Cai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Experimental results (\\%) of public LMMs on BenchLMM's cross-artistic-style evaluation. Note that Average* represents the average accuracy computed over five artistic-style benchmarks. The best and the second best results are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.213, + 0.785, + 0.378 + ], + "angle": 0, + "content": "
LMMsVisual EncoderLanguage ModelCommonCartoonHandmakePaintingSketchTattooAverage*
GPT-4V [62]--81.563.358.758.264.768.362.6
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B74.662.056.657.663.357.059.3
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B72.759.057.159.357.561.458.9
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B56.649.945.448.053.747.048.8
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B56.957.828.437.842.632.340.0
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B71.251.953.343.462.054.653.0
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B73.958.152.055.855.255.555.3
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B66.428.737.840.941.728.635.5
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B45.635.037.038.536.437.736.9
Otter-7B [32]CLIP-ViT-L/14MPT-7B64.845.149.039.844.142.444.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.533 + ], + "angle": 0, + "content": "to a gap in understanding their performance across diverse artistic styles. We extend the evaluation scope by examining LMMs' performance with various artistic styles beyond the common 'Photo' style. The results detailed in Table 1 reveal a notable decline in LMMs' effectiveness when processing these artistic styles. This trend suggests a potential overfitting of LMMs to the 'Photo' style, highlighting their limited adaptability to varied artistic styles, a capability that humans typically possess. Interestingly, GPT-4V, as an outstanding commercial model, exhibits similar limitations in handling diverse styles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Besides, LLaVA-1.5-13B [38] is improved from LLaVA-13B [39] by introducing more training tricks and data, and LLaVA-1.5-13B outperforms LLaVA-13B by a clear margin on the 'Common' and other artistic styles. This comparison reveals the importance of data and training strategy as LLaVA-1.5-13B and LLaVA-13B use the same visual encoder and language encoder. However, we also find that a better performance on 'Common' does not necessarily indicate a better performance on other styles. For example, compared with InstructBLIP-7B, InstructBLIP-13B demonstrates better performance across five artistic-style benchmarks in terms of average accuracy, though it shows comparatively lower performance in the 'Common' style. Therefore, such results convey that it is important to conduct comprehensive evaluations to fully validate an LMM's capability rather than only focus on the 'Common' style, which shows the importance of our proposed benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Cross-sensor Capability We benchmark the LMM models' cross-sensor capability in Table 2. Because GPT-4V denied to provide medical diagnostics or identify potential abnormalities in medical images like H-Xray/MRI/CT, GPT-4V is only evaluated on the Infrared and L-Xray sensor data and achieves the results of \\(55.0\\%\\) and \\(50.0\\%\\) respectively, which are far lower than the 'Common' style \\((81.5\\%)\\). Meanwhile, other LMM models' performance on images captured by X-ray, MRI, CT, and Infrared dropped significantly compared to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.651, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BenchLMM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 2: Experimental results (\\%) of public LMMs on BenchLMM's cross-sensor evaluation. Note that GPT-4V denied to test medical images (H-Xray/MRI/CT)." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.186, + 0.783, + 0.358 + ], + "angle": 0, + "content": "
LMMsVisual EncoderLanguage ModelInfraredL-XrayH-XrayMRICTAverage
GPT-4V [62]--55.050.0----
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B52.151.555.933.743.147.3
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B40.824.656.620.943.437.3
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B53.746.039.128.439.741.4
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B32.947.224.625.323.230.6
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B46.553.249.429.646.845.3
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B42.235.459.519.438.539.0
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B30.143.031.517.417.828.2
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B35.144.123.716.815.327.0
Otter-7B [32]CLIP-ViT-L/14MPT-7B30.246.439.147.846.342.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.373, + 0.773, + 0.388 + ], + "angle": 0, + "content": "Table 3: Results of public LMMs on BenchLMM's cross-application experiments." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.4, + 0.783, + 0.551 + ], + "angle": 0, + "content": "
LMMsVisual EncoderLanguage ModelRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
GPT-4V [62]--69.735.856.074.064.460.0
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B65.628.649.021.460.345.2
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B63.632.632.019.547.038.9
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B38.731.759.043.828.240.3
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B33.322.648.034.231.734.2
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B61.731.344.015.459.642.2
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B63.536.742.024.830.839.6
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B34.517.640.022.218.826.6
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B27.920.841.040.113.628.7
Otter-7B [32]CLIP-ViT-L/14MPT-7B44.829.736.018.827.331.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.785, + 0.718 + ], + "angle": 0, + "content": "the performance on 'Common' captured by the RGB camera sensor. Therefore, the intrinsic difference in the imaging process between RGB cameras and other sensors is also a crucial challenge to LMMs. With respect to the MRI, H-Xray, and CT images, the questions may require specific medical knowledge to answer, and it may be argued that the performance drop is due to the knowledge gap. It is also observed from the L-Xray, where the objects (i.e., knife, gun, hammer, etc.) can be recognized with only common knowledge, but the LMMs' performance on the L-Xray data is significantly lower than 'Common', which further validates the challenge brought by the sensor difference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Cross-application Capability In Table 3, we evaluate the different LMMs on various applications, finding that GPT-4V generally outperforms others, except in autonomous driving where its accuracy drops to \\(35.8\\%\\). This decrease is likely due to the challenges in the Autonomous Driving dataset, such as recognizing low-resolution traffic signs and requiring specific driving knowledge, where both GPT-4 and open-source LMMs have poor performance. We also observe that LLaVA-1.5-13B achieves comparable performance as GPT-4V on the Remote Sensing and Defect Detection datasets, which seems to have comparable cross" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.148, + 0.415, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.352, + 0.415, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.148, + 0.586, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.352, + 0.586, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.148, + 0.756, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.352, + 0.756, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.564, + 0.785, + 0.593 + ], + "angle": 0, + "content": "Fig. 3: Error-reflection capability comparisons: Unlike LLaVA-1.5-13B, GPT-4V possesses the ability to analyze its errors when provided with the correct answers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.695 + ], + "angle": 0, + "content": "application capability. However, we surprisingly find that LLaVA-1.5-13B merely achieves \\(21.4\\%\\) but GPT-4V achieves up to \\(74.0\\%\\). As GPT-4V is close-source, the technical details are not disclosed and we conjecture that the training data for GPT-4V may involve data similar to the Open Game dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.775 + ], + "angle": 0, + "content": "We conclude from the experiments that: 1) Existing LMMs often degrade in performance with cross-style images; 2) Better performance on 'Common' data doesn't ensure better results on other styles; 3) Comprehensive evaluations across different dimensions are crucial, as errors in tasks like medical or security can lead to significant losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Error-reflection Capability To understand the LMMs further, we let the LMM do an error reflection that uses itself to parse the error information. When a model gives wrong answers, we warp up the conversation: the previous question input and predicted answer output. Specifically, we add \"The correct answer is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.651, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BenchLMM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.159 + ], + "angle": 0, + "content": "Table 4: Experimental results (%) of our proposed SPE on the BenchLMM benchmark." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.172, + 0.756, + 0.433 + ], + "angle": 0, + "content": "
LMMsCartoonHandmakePaintingSketchTattooAverage
LLaVA-1.5-13B [38]62.056.657.663.357.059.3
+ SPE (Ours)67.657.767.968.863.565.1 (+5.8)
InstructBLIP-13B [14]59.057.159.357.561.458.9
+ SPE (Ours)61.162.364.860.963.362.5 (+3.6)
LMMsInfraredL-XrayH-XrayMRICTAverage
LLaVA-1.5-13B [38]52.151.555.933.743.147.3
+ SPE (Ours)58.762.856.233.647.751.8 (+4.5)
InstructBLIP-13B [14]40.824.656.620.943.437.3
+ SPE (Ours)43.524.857.725.042.638.7 (+1.4)
LMMsRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
LLaVA-1.5-13B [38]65.628.649.021.460.345.2
+ SPE (Ours)68.029.451.022.761.746.6 (+1.4)
InstructBLIP-13B [14]63.632.632.019.547.038.9
+ SPE (Ours)63.538.831.019.049.440.3 (+1.4)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.785, + 0.49 + ], + "angle": 0, + "content": "[GT Answer]. Explain why your answer is incorrect. Use a single sentence.\" to the conversation (GT means Ground-truth), as the new input of LMMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Although both GPT-4V and LLaVA-1.5-13B produce incorrect answers, GPT-4V demonstrates the ability to reason the correct answer through a detailed inference process. As illustrated in Fig. 3, GPT-4V can offer insightful analysis beyond the ground-truth answer, including the positions of clock hands for estimating time, the location of prohibited objects, and the traffic signals for driving planning. Moreover, from the second row of Fig. 3, we can see that the LLaVA model can just simply repeat the answer or say \"I was wrong\". By contrast, GPT-4V can do inference itself to provide the correct answer, such as recognizing the toy animals (Handmake) or the scratch on the capsule (Defect Detection). Through this study, we can see that the error-reflection capability of LLaVA-1.5-13B is inferior to that of GPT-4V. Therefore, we reveal that another important capability of a large multimodal model is whether the model can reflect how to derive the correct answer. Studying such capability is useful to common users. For example, when users use LMMs for self-learning, the user can provide reference answers and LMMs can provide a more detailed inference process, which can improve the learning effect. Moreover, the answer of GPT-4V can also be useful materials used to train open-source LMMs in the future." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.622, + 0.785 + ], + "angle": 0, + "content": "4.3 Experiments of Style Prompt Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Comparison to Baseline Models In this section, we aim to study if a large multimodal model can be enhanced by prompting the model to estimate the style first and then answer the question, based on which we propose the Style Prompt" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Cai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 5: Complementary studies of our SPE with recent zero-shot chain-of-thought method ZS-COT [31] using LLaVA-1.5-13B [38] on the BenchLMM" + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.185, + 0.765, + 0.38 + ], + "angle": 0, + "content": "
MethodCartoonHandmakePaintingSketchTattooAverage
LLaVA-1.5-13B [38]62.056.657.663.357.059.3
+ZS-COT [31]64.359.770.662.871.065.7(+6.4)
+ZS-COT [31]+SPE (Ours)66.064.466.565.468.366.1(+6.8)
MethodInfraredL-XrayH-XrayMRICTAverage
LLaVA-1.5-13B [38]52.151.555.933.743.147.3
+ZS-COT [31]49.355.957.635.644.748.6(+1.3)
+ZS-COT [31]+SPE (Ours)54.470.456.735.345.452.4(+5.1)
MethodRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
LLaVA-1.5-13B [38]65.628.649.021.460.345.2
+ZS-COT [31]65.527.043.034.263.746.7(+1.5)
+ZS-COT [31]+SPE (Ours)76.024.049.052.160.152.2(+7.0)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.481 + ], + "angle": 0, + "content": "Enhancement (SPE) method and conduct experiments with our BenchLMM. Specifically, we examined our SPE with LLaVA-1.5-13B [38] and InstructBLIP-13B [14] because these two models are representative of better average performance than MiniGPT4 [71] and Otter [32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.699 + ], + "angle": 0, + "content": "The experimental results are presented in Table 4. As shown in the table, our proposed SPE can benefit the LMMs by achieving general improvement over different styles, which means that reasoning the style first is an effective prompting method to improve LMMs' visual reasoning capability. Besides, in the experiments of crossing to different applications (Remote Sensing, Autonomous Driving, etc.), we observe that its benefit is smaller than others. For example, LLaVA-1.5-13B gains \\(5.8\\%\\) and \\(4.5\\%\\) improvement in the experiments of cross-artistic and cross-sensor styles, while the improvement is merely \\(1.4\\%\\) when crossing applications. We conjecture that specific application knowledge could inferred from only the style, which limits the benefits of SPE. Moreover, we observe that LLaVA-1.5-13 generally obtains more benefits than InstructBLIP-13B, which means that the stronger the LMM, the more benefits from SPE, which is because a stronger LMM has stronger inference capability and can better utilize prior information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison to Other Prompting Methods Our proposed SPE method enhances LMMs' performance by prompting LMMs to estimate the input image's style first, a technique akin to chain-of-thought prompting approaches [31, 45]. In Table 5, we compare our proposed SPE with the state-of-the-art prompting Chain-of-Thought (COT) [31], which is a zero-shot, to demonstrate the efficacy of SPE. As shown, the COT method can improve LMM's performance, and our SPE is compatible with the COT methods and further enhances the model's capability, which is demonstrated by the general improvement brought by “+ZS-COT+SPE” as shown in Table 5. We also notice that in some cases, using SPE" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.651, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BenchLMM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.145, + 0.744, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.258, + 0.785, + 0.286 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons between LLaVA-1.5-13B and LLaVA-1.5-13B + SPE (Ours) on style-transfer images." + }, + { + "type": "table_caption", + "bbox": [ + 0.243, + 0.301, + 0.757, + 0.315 + ], + "angle": 0, + "content": "Table 6: Quantitative Results of our proposed SPE on style-transfer images." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.327, + 0.763, + 0.414 + ], + "angle": 0, + "content": "
LMMs|Cartoon Handmake Painting Sketch Tattoo | Average
LLaVA-1.5-13B [38] + SPE (Ours)62.161.152.468.841.257.1
67.462.159.469.445.560.8 (+3.7)
InstructBLIP-13B [14] + SPE (Ours)69.370.861.353.865.064.0
73.071.664.262.165.367.2 (+3.2)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.469 + ], + "angle": 0, + "content": "with COT could lead to performance decline. The investigation is beyond the scope of this paper, which will be explored in our future work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.785, + 0.606 + ], + "angle": 0, + "content": "Results on the Style-transfer Images We employ the latest commercial style transfer model, DALL-E3 [50], to transfer the 'Common' style images into different artistic styles. As depicted in Fig. 4, these images share similar structures and layouts. However, they are just in different artistic styles, and the tested LLaVA model also makes mistakes. The significant performance degradation on the transferred images is also proved by the quantitative results presented in Table 6, which validates that LMMs suffer from difficulties in recognizing images of different styles. Nevertheless, our SPE significantly enhances the cross-style capability of LMMs in this challenging task, as shown in Fig. 4 and Table 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.785, + 0.758 + ], + "angle": 0, + "content": "In-depth Analysis of SPE Besides quantitative assessments, we explore the inner workings of SPE to understand how it improves LMMs. As illustrated in Fig. 5, LMMs are adept at accurately estimating the style of images. For instance, in the figure's first row and second column, the depicted style is 'Handmake,' and the model accurately recognizes this style as akin to a toy, indicating a correct match. Moreover, when estimating the style, the model can also understand the context first and thus better answer the final question, as shown in the examples of Remote Sensing and Defect Detection. Our findings suggest that prompting LMMs to first identify the style of images serves as an effective, training-free method to improve the models' proficiency in handling cross-style images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.359, + 0.795 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this work, we introduce BenchLMM, a novel benchmark for quantitatively evaluating large multimodal models (LMMs) across varied visual distribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.147, + 0.756, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.635, + 0.726, + 0.649 + ], + "angle": 0, + "content": "Fig. 5: Evaluation of LLaVA-1.5-13B with SPE on the BenchLMM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.827 + ], + "angle": 0, + "content": "shifts, including artistic, sensor, and application styles. Our evaluations of numerous existing LMMs highlight two key findings: 1) LMMs typically underperform with non-common image styles. 2) Our style prompt enhancement approach, inspired by human perception of art, significantly improves LMMs' visual reasoning without extra fine-tuning. Additionally, our error-reflection study shows that stronger LMMs can self-diagnose errors by providing insights not given by humans, whereas less capable models merely restate correct answers without new insights. Our benchmark offers a comprehensive tool for LMM research, emphasizing the importance of error-reflection capabilities in the future." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.651, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BenchLMM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.252 + ], + "angle": 0, + "content": "This work was carried out at Rapid-Rich Object Search (ROSE) Lab, School of Electrical & Electronic Engineering, Nanyang Technological University. This research is supported by the NTU-PKU Joint Research Institute (a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.274, + 0.323, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.785, + 0.358 + ], + "angle": 0, + "content": "1. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8948-8957 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.413 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.455 + ], + "angle": 0, + "content": "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.523 + ], + "angle": 0, + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.564 + ], + "angle": 0, + "content": "5. Ben-Younes, H., Cadene, R., Cord, M., Thome, N.: Multimodal tucker fusion for visual question answering. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2612-2620 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.785, + 0.619 + ], + "angle": 0, + "content": "6. Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., Steger, C.: The mvtec anomaly detection dataset: a comprehensive real-world dataset for unsupervised anomaly detection. International Journal of Computer Vision 129(4), 1038-1059 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.661 + ], + "angle": 0, + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.702 + ], + "angle": 0, + "content": "8. Cadene, R., Ben-Younes, H., Cord, M., Thome, N.: Murel: Multimodal relational reasoning for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1989-1998 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.743 + ], + "angle": 0, + "content": "9. Cao, Q., Wan, W., Wang, K., Liang, X., Lin, L.: Linguistically routing capsule network for out-of-distribution visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1614-1623 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.745, + 0.785, + 0.784 + ], + "angle": 0, + "content": "10. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "11. Chen, L., Zhang, Y., Ren, S., Zhao, H., Cai, Z., Wang, Y., Wang, P., Liu, T., Chang, B.: Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4-vision and beyond. arXiv preprint arXiv:2310.02071 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.303, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Cai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "12. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollár, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.233 + ], + "angle": 0, + "content": "13. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.274 + ], + "angle": 0, + "content": "14. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.331 + ], + "angle": 0, + "content": "15. Dancette, C., Cadene, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1574-1583 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.387 + ], + "angle": 0, + "content": "16. Ding, Y., Yu, J., Liu, B., Hu, Y., Cui, M., Wu, Q.: Mukea: Multimodal knowledge extraction and accumulation for knowledge-based visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5089-5098 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.785, + 0.43 + ], + "angle": 0, + "content": "17. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models. arXiv preprint arXiv:2306.13394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.785, + 0.472 + ], + "angle": 0, + "content": "18. Fukui, A., Park, D.H., Yang, D., Rohrbach, A., Darrell, T., Rohrbach, M.: Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.528 + ], + "angle": 0, + "content": "19. Gao, F., Ping, Q., Thattai, G., Reganti, A., Wu, Y.N., Natarajan, P.: Transform-retrieve-produce: Natural language-centric outside-knowledge visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5067-5077 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.557 + ], + "angle": 0, + "content": "20. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhan, S.: Chain of thought prompt tuning in vision language models. arXiv preprint arXiv:2304.07919 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.613 + ], + "angle": 0, + "content": "21. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6904-6913 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.67 + ], + "angle": 0, + "content": "22. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.785, + 0.713 + ], + "angle": 0, + "content": "23. Han, X., Wang, S., Su, C., Huang, Q., Tian, Q.: Greedy gradient ensemble for robust visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1584-1593 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "24. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "25. Hwang, S., Park, J., Kim, N., Choi, Y., So Kweon, I.: Multispectral pedestrian detection: Benchmark dataset and baseline. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1037-1045 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "26. Jia, X., Zhu, C., Li, M., Tang, W., Zhou, W.: LLVIP: A visible-infrared paired dataset for low-light vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3496-3504 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.651, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BenchLMM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "27. Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2901-2910 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.26 + ], + "angle": 0, + "content": "28. Johnson, J., Hariharan, B., Van Der Maaten, L., Hoffman, J., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Inferring and executing programs for visual reasoning. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2989-2998 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.288 + ], + "angle": 0, + "content": "29. Kim, J.H., Jun, J., Zhang, B.T.: Bilinear attention networks. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.317 + ], + "angle": 0, + "content": "30. Kim, J.H., On, K.W., Lim, W., Kim, J., Ha, J.W., Zhang, B.T.: Hadamard product for low-rank bilinear pooling. arXiv preprint arXiv:1610.04325 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.358 + ], + "angle": 0, + "content": "31. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. Advances in neural information processing systems 35, 22199-22213 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.388 + ], + "angle": 0, + "content": "32. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.787, + 0.43 + ], + "angle": 0, + "content": "33. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.787, + 0.459 + ], + "angle": 0, + "content": "34. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.487 + ], + "angle": 0, + "content": "35. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.787, + 0.529 + ], + "angle": 0, + "content": "36. Liang, P.P., Lyu, Y., Fan, X., Wu, Z., Cheng, Y., Wu, J., Chen, L., Wu, P., Lee, M.A., Zhu, Y., et al.: Multibench: Multiscale benchmarks for multimodal representation learning. arXiv preprint arXiv:2107.07502 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.787, + 0.585 + ], + "angle": 0, + "content": "37. Liu, B., Zhan, L.M., Xu, L., Ma, L., Yang, Y., Wu, X.M.: Slake: A semantically-labeled knowledge-enhanced dataset for medical visual question answering. In: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). pp. 1650-1654. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.614 + ], + "angle": 0, + "content": "38. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.787, + 0.643 + ], + "angle": 0, + "content": "39. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.645, + 0.787, + 0.713 + ], + "angle": 0, + "content": "40. Liu, J., Fan, X., Huang, Z., Wu, G., Liu, R., Zhong, W., Luo, Z.: Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5802-5811 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.787, + 0.755 + ], + "angle": 0, + "content": "41. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: MMBench: Is Your Multi-modal Model an All-around Player? arXiv preprint arXiv:2307.06281 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.784 + ], + "angle": 0, + "content": "42. Manyika, J.: An overview of bard: an early experiment with generative ai. AI. Google Static Documents (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "43. Mao, X., Chen, Y., Zhu, Y., Chen, D., Su, H., Zhang, R., Xue, H.: COCO-O: A Benchmark for Object Detectors under Natural Distribution Shifts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6339-6350 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Cai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.203 + ], + "angle": 0, + "content": "44. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3195-3204 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.786, + 0.231 + ], + "angle": 0, + "content": "45. Mitra, C., Huang, B., Darrell, T., Herzig, R.: Compositional chain-of-thought prompting for large multimodal models. arXiv preprint arXiv:2311.17076 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "46. Nam, H., Ha, J.W., Kim, J.: Dual attention networks for multimodal reasoning and matching. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 299-307 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "47. Nguyen, H.D., Cai, R., Zhao, H., Kot, A.C., Wen, B.: Towards more efficient security inspection via deep learning: a task-driven x-ray image cropping scheme. Micromachines 13(4), 565 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.756, + 0.328 + ], + "angle": 0, + "content": "48. OpenAI, R.: Gpt-4 technical report. arxiv 2303.08774. View in Article (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.329, + 0.786, + 0.37 + ], + "angle": 0, + "content": "49. Shah, M., Chen, X., Rohrbach, M., Parikh, D.: Cycle-consistency for robust visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6649-6658 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.397 + ], + "angle": 0, + "content": "50. Shi, Z., Zhou, X., Qiu, X., Zhu, X.: Improving image captioning with better use of captions. arXiv preprint arXiv:2006.11807 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.453 + ], + "angle": 0, + "content": "51. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 742-758. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.508 + ], + "angle": 0, + "content": "52. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards VQA models that can read. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8317-8326 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.536 + ], + "angle": 0, + "content": "53. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.786, + 0.577 + ], + "angle": 0, + "content": "54. Tascon-Morales, S., Márquez-Neila, P., Sznitman, R.: Logical implications for visual question answering consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6725-6735 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.619 + ], + "angle": 0, + "content": "55. Teney, D., Abbasnejad, E., van den Hengel, A.: Unshuffling data for improved generalization in visual question answering. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1417-1427 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.786, + 0.66 + ], + "angle": 0, + "content": "56. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "57. Tsimpoukelli, M., Menick, J.L., Cabi, S., Eslami, S., Vinyals, O., Hill, F.: Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems 34, 200-212 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.743 + ], + "angle": 0, + "content": "58. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.799 + ], + "angle": 0, + "content": "59. Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: Dota: A large-scale dataset for object detection in aerial images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3974-3983 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "60. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.651, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BenchLMM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "61. Yang, X., Gao, C., Zhang, H., Cai, J.: Auto-parsing network for image captioning and visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2197-2207 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "62. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of LMMs: Preliminary explorations with GPT-4V (ision). arXiv preprint arXiv:2309.17421 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "63. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.314 + ], + "angle": 0, + "content": "64. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.37 + ], + "angle": 0, + "content": "65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2633-2642 (2020). https://doi.org/10.1109/CVPR42600.2020.00271" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.411 + ], + "angle": 0, + "content": "66. Yu, S., Wu, P., Liang, P.P., Salakhutdinov, R., Morency, L.P.: Pacs: A dataset for physical audiovisual commonsense reasoning. In: European Conference on Computer Vision. pp. 292-309. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.453 + ], + "angle": 0, + "content": "67. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.495 + ], + "angle": 0, + "content": "68. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6720-6731 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.522 + ], + "angle": 0, + "content": "69. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.577 + ], + "angle": 0, + "content": "70. Zhou, Y., Ren, T., Zhu, C., Sun, X., Liu, J., Ding, X., Xu, M., Ji, R.: Trar: Routing the attention spans in transformer for visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2074-2084 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.606 + ], + "angle": 0, + "content": "71. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.606 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_origin.pdf b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9cf5c620fb573efc372539ab28bfb3d2ab82abd3 --- /dev/null +++ b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/dd120aa9-9823-44e2-b7b7-9e3a376cc7da_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa7f565f9a4292fa5e5b221b00a78debd205caebfaf603ba2d35282115c3cfd7 +size 6626824 diff --git a/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/full.md b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d1a4aeb48e5cabe7a075931e5c1d5b7967feddfe --- /dev/null +++ b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/full.md @@ -0,0 +1,323 @@ +# BenchLMM: Benchmarking Cross-style Visual Capability of Large Multimodal Models + +Rizhao Cai $^{1}$ , Zirui Song $^{1,2}$ , Dayan Guan $^{3}$ , Zhenhao Chen $^{4}$ , Yaohang Li $^{2}$ , Xing Luo $^{5}$ , Chenyu Yi $^{1}$ , and Alex Kot $^{1}$ + +1 Nanyang Technological University, Singapore + +2 University of Technology Sydney, Australia; Northeastern University, China; + +3 Harbin Institute of Technology, China + +4 Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates + +5 Zhejiang University, China + +† Equal Contribution Corresponding author (guandayan@hit.edu.cn) + +Abstract. Large Multimodal Models (LMMs) such as GPT-4V and LLaVA have shown remarkable capabilities in visual reasoning on data in common image styles. However, their robustness against diverse style shifts, crucial for practical applications, remains largely unexplored. In this paper, we propose a new benchmark, BenchLMM, to assess the robustness of LMMs toward three different styles: artistic image style, imaging sensor style, and application style. Utilizing BenchLMM, we comprehensively evaluate state-of-the-art LMMs and reveal: 1) LMMs generally suffer performance degradation when working with other styles; 2) An LMM performs better than another model in common style does not guarantee its superior performance in other styles; 3) LMMs' reasoning capability can be enhanced by prompting LMMs to predict the style first, based on which we propose a versatile and training-free method for improving LMMs; 4) An intelligent LMM is expected to interpret the causes of its errors when facing stylistic variations. We hope that our benchmark and analysis can shed new light on developing more intelligent and versatile LMMs. The benchmark and evaluation have been released on https://github.com/AIFEG/BenchLMM. + +# 1 Introduction + +The dynamic landscape of computer vision has recently been shaped by the rise of Large Multimodal Models (LMMs). These models [10, 14, 32, 39, 71], incorporating visual and textual data, are pivotal in developing general-purpose assistants for visual reasoning, a widely-studied task that involves reasoning and answering questions based on images. Their growing prominence and the enthusiasm surrounding visual reasoning are evident from numerous studies such as GPT-4V [62]. As the interest in LMMs swells within the research community, understanding their depth, breadth, and limitations becomes imperative. + +To comprehensively understand the visual reasoning capabilities of LMMs, numerous benchmarks [17, 21, 33, 34, 41, 67] have been proposed to systematically evaluate the linguistic capabilities of LMMs. These linguistic capabilities + +![](images/404bb180824da96571d9ab6fe665a23fdf682a8ebbbf6904cd560edb5cdd35fa.jpg) +Q: How many forks? +A (GT): Two +A (GPT-4V): Two +A (LLaVA-1: 5) Two + +![](images/22939106675c9ad1454f1a74756e33be497feb2114334d0298ca783a0f8ea8d4.jpg) +[Painting]Q:How many elephants? A (GT): Seven A (GPT-4V):Five A(LLAVA-1.5):Five + +![](images/dcee0d27a1f39563afdbacaf526c3d584184758197c7ab7e9d3002689d3b8e16.jpg) +[Xray] Q: What are prohibited objects? +A (GT): None +A (GPT-4V): Gun +A (LLVaA-1.5): A knife and a scissor +Fig. 1: The motivation of our work. State-of-the-art LMMs, such as GPT-4V and LLaVA-1.5 can understand the images in a common style. However, these models have difficulties in reasoning images in other styles and deliver wrong answers. + +![](images/34c9bd7e9dfad720e12e11ec50f92198090d351c476e418930f531b04b805486.jpg) +[Driving] Q: What is the action for the ego-vehicle during autonomous driving? A (GT): Go straight A (GPT-4V): Stop at traffic light A (LLAvA-1.5): Stop + +are assessed through various types of responses for visual reasoning tasks, including Yes/No answers [17], multi-choice answers [33,41], single word or phrase answers [21,24,44]. However, while these benchmarks investigate textual diversity, they may fall short in terms of visual diversity as most images in these benchmarks are photographic images sourced from the internet. Despite providing valuable insights into LMMs' performance, there remains a gap in understanding how LMMs handle shifts in visual distribution. Although the recently released GPT-4V [62] claims remarkable capabilities across diverse application scenarios, only a few samples of qualitative results are presented in the technical report [62], and the quantitative evaluations across different scenarios are missing. Moreover, according to our preliminary study as illustrated in Fig. 1, we reveal that LMMs, even the state-of-the-art GPT-4V, tend to make mistakes in reasoning images with non-common styles. + +To further investigate LMMs' cross-style visual reasoning capability, we propose a novel benchmark BenchLMM, which systematically assesses existing LMMs in three distinct styles of distribution shifts: artistic styles, sensor styles, and application styles. Firstly, images of the common style are realistic, but images can be in other artistic styles, such as Cartoon, Painting, Sketch, etc. Besides, there are different imaging sensors other than RGB cameras, such as infrared and X-ray. The discrepancy of sensors induces the distribution shifts of images. Moreover, we also identify the shift of knowledge required for different application styles. For example, the image recognition application requires knowledge about object shape and appearance, while the application of decision-making for robot action requires application-specific knowledge. + +With our BenchLMM, we conduct in-depth analysis through experiments of evaluating advanced LMMs. The comprehensive experiments reveal that while these cutting-edge LMMs excel with commonly styled data, their performance significantly declines with data in other styles. Interestingly, we also discover that the superior performance of an LMM in common data styles does not necessarily predict better performance across other styles. Furthermore, our investigation shows that prompting an LMM to predict the style of the data before answering a question can significantly boost its reasoning abilities. Based on this insight, we introduce the Style Prompt Enhancement (SPE) technique. SPE is a novel + +approach that improves LMMs' performance without additional training and operates effectively in a zero-shot manner. Moreover, we analyze the failure cases by conducting an in-depth error-reflection analysis by asking an LMM to interpret why it made mistakes in the Q&A. We find that a more intelligent LMM can understand the detailed reason why it made mistakes and can learn to derive the correct answer, which is an important capability that deserves attention when developing LMMs. We hope that our benchmark and analysis can shed new light on developing more intelligent and versatile LMMs. + +Our contributions can be summarized below: 1) BenchchLMM: We introduce BenchLMM, a pioneering benchmark specifically designed to quantitatively evaluate the performance of LMMs' visual reasoning capability across visual styles in artistic, sensor, and application shifts. 2) Comprehensive evaluation of cutting-edge LMMs: We conduct extensive quantitative benchmark experiments on LMMs, including the commercial GPT-4V. Our rigorous and quantitative evaluation reveals LMMs' limitations in versatility and performance across different data styles. This finding challenges the prevailing assumption of universal applicability and highlights the need for specialized strategies to enhance LMM adaptability. 3) Insightful Analysis and Enhancement Strategy: Beyond mere performance assessment, we conduct an in-depth error-reflection analysis, where LMMs are prompted to identify the reasons behind their errors in question-and-answer tasks. This study reveals that more sophisticated LMMs can recognize their mistakes and also learn from them to improve future responses. Moreover, we introduce a novel SPE technique, as a training-free and zero-shot strategy, to significantly boost LMMs' reasoning capabilities across varied data styles, marking a substantial advancement in LMM development. + +# 2 Related works + +Visual Reasoning Visual reasoning, also known as visual question answering (VQA), aims to reason and answer questions about visual information [3,27,28]. Early developments [5, 18, 30] focus on designing more complex fusion mechanisms instead of simple summation fusion [3]. By including modern attention mechanism [58], numerous works [8, 29, 46, 53] contributed to the creation of transformer-based vision-language models. Recent methods have proposed to improve other aspects of visual question answering, including avoiding shortcut learning [15, 23], addressing the linguistic bias [9, 55], improving transformer-based vision-language models [61, 70], external knowledge integration [16, 19], and consistency regularization [49, 54]. Lastly, the advent of large multimodal models has markedly enhanced the capabilities of visual reasoning. + +Large Multimodal Models Recent advancements in large language models (LLMs) [7, 13, 48, 56] have inspired the development of LMMs that enhance LLMs with vision-language capabilities. Existing LMMs can be categorized into three main classes: multimodal instruction tuning [10, 14, 32, 38, 39, 71] that involves the finetuning of pre-trained LLMs on multi-modal instruction-following data; multimodal in-context learning [2, 32, 57, 63] that derives insights from + +few-shot task-specific samples; multimodal chain-of-thought [20,69] that encourages LMMs to articulate both the final answer and the underlying reasoning process. Additionally, some commercial LMMs like Google's Bard [42] and Microsoft's GPT-4V [62], with detailed information about their architecture remaining undisclosed, have shown impressive visual capabilities. + +Evaluation of LMMs To evaluate their capabilities, existing LMMs have been tested on a number of vision-language datasets. These datasets focus on specific visual reasoning tasks by designing different kinds of Q&A task, and the specific task includes visual recognition [21, 44], image description [1, 12], scene text understanding [51, 52, 64], object hallucination evaluation [35], complex reasoning [39, 67], and commonsense reasoning [66, 68]. To facilitate a comprehensive comparison of LMMs, several benchmarks [17, 36, 60] have been established by encompassing various visual reasoning tasks with common-style images and diverse Q&A pairs. However, the above benchmarks do not consider different styles of visual distribution shift, which motivates us to propose our BenchLMM. + +# 3 BenchLMM + +# 3.1 Motivation + +LMMs (e.g., GPT-4V) are claimed to excel in various scenarios [62], shown with qualitative results. Nevertheless, we provide comprehensive and quantitative results experiments with our BenchLMM reveal a paradox. We will elaborate on the benchmark that encompasses different cross-style data as follows. + +Cross-artistic Style Existing benchmarks for LMMs commonly use photo images for evaluation, which we name as 'Common' style. Meanwhile, there are diverse image styles, including Cartoon, Sketch, etc., as shown and compared in Fig. 2. The object in other artistic styles have different distributions of object characteristics. For instance, a colorful zebra, common in painting and tattoo styles, is rare in 'Common' style images. While humans easily interpret such characteristics, the extent of LMMs' understanding of the other styles remains uncertain. This uncertainty arises from potential insufficient training on datasets encompassing these diverse styles. Therefore, we propose an evaluation to assess LMMs' robustness in analyzing images across various artistic styles. + +Cross-sensor Style Photographic images are commonly obtained through RGB cameras for information sensing. Alongside RGB cameras, other sensor types are also important for capturing images. For instance, infrared cameras are frequently utilized in surveillance systems to capture images during nighttime, as RGB cameras rely on adequate illumination for optimal image quality. Additionally, X-ray sensors are used in airport security checks, enabling the detection of prohibited items within luggage. These sensors feature distinct imaging processes compared to RGB cameras. Consequently, images acquired through various sensors exhibit unique intrinsic properties. The question of whether the performance of LMMs remains remarkable with other sensor data has not been thoroughly investigated. This issue will be addressed in our benchmark study. + +![](images/f2d2babc524998a0bf475f18c4582ac1e28a99862186cc4d0d6561d5f93d6239.jpg) +Q: What time is it? A: 2:57:39 + +![](images/4be0ed448c216202254e1066ace484dd64c0887d12b1f9b8e610c1bbb95e8c2d.jpg) + +![](images/4424394d2d32f6650034b12110bceeec2d7b594dc316c94b2164dadc46ce58e9.jpg) +Q: Who is sitting on the couch? A: A woman + +![](images/698a9a8b89db27df8fff7a464e3d07f55ce4c5feea13f35d8baa49b2accf3164.jpg) +Q: What kind of sport? A: football + +![](images/cabfe674ef11cc3714c16781c111d0578268257bf16c11bc2781ce0e842aad14.jpg) +Tattoo +Q: What is in the image? A: A colorful zebra + +![](images/c031735a6bd5d0bea0fd295827d09b7be8f5e1a38c91fbf4c699d0b4ebf8cd68.jpg) +Q: How many pedestrians? A: 4 + +![](images/708b1202d956d4b16cd5625bc8484eed65f62761e9035bfc4036438e4210fdce.jpg) +A: A bronze sculpture of a person with a very large head +Q: What are prohibited objects? +A: Gun, knife + +![](images/288027131b791a85b0e6d07efea4d486e0636593c8828a7ca31e99089c59981b.jpg) +Q: What is the main organ? +A: Lung + +![](images/3d1c1f07e0627ce2b5b668d903b5a0a36320db83ab59f8440ada08f192aaf07f.jpg) +Q: What kind of disease? A: Brain edema, brain tumor + +![](images/27c2cd8ad1a0c8faba2c2eb7310deb1b207c4f5b4897dd164a688650dfc11e81.jpg) +Q: What kind of disease? +A: Lung cancer + +![](images/bc3aca5efe34effc91a59b408995d2f6d8215a42242c2b2c256840fe86e8889d.jpg) +Q: How many sport facilities? A: 3 +Fig. 2: Benchmark examples. For each specific field, one example image and its corresponding Q&A have been shown. Note that, for a simple presentation, the questions in Domestic Robot and Open Game have been simplified from multiple-choice format. Please see the Appendix for more examples and detailed questions. + +![](images/fcfe2f66f387cebd09afbfa3f4132f468b81318ffc366443c59cb035be9ee25b.jpg) +Q: What is the intended action for the main vehicle in an autonomous driving scenario? A: Go straight + +![](images/f6c1e9b620e6aa0a1779dd5688dbf400fe7ee8f0560c89be275330b62a4c20f5.jpg) +Q: Your goal is to repair the broken clock in your hands. What action should you do next? A: Look for repair tools + +![](images/70a1048286a32200221a0411d837f2edab99e2038afea1d54ccf1601a7916c4d.jpg) +Q: Your task is to harvest beef using the iron sword. Which action should you do next? A: Find cow + +![](images/c43a8182b60d8a4722ea56871e70ed174d8c8d4083b159c15d4cff869937cbea.jpg) +Q: What is the defect type? A: Wire combined + +Cross-application Style Existing LMMs are mainly assessed in general application scenarios where questions can be answered using common knowledge [17, 21]. For instance, identifying the object on a table in an image can be accomplished with general knowledge. However, there exist specific application scenarios that demand specialized domain knowledge to address. For example, when an autonomous driving control center processes the road image to decide the next action: going straight or turning left, specific driving knowledge is required. + +As existing LMMs are primarily evaluated with questions and answers (Q&A) that rely on common knowledge, the extent to which an LMM can remain effective in situations requiring domain-specific knowledge has not been comprehensively explored. We also investigate this aspect in our proposed benchmark. + +# 3.2 Benchmark Construction + +Artistic Style Data To construct a benchmark characterized by distributional shifts pertaining to artistic styles, we curated a diverse set of images encompassing Cartoon, Painting, Sketch, Handmade, and Tattoo styles from the COCO-O dataset [43]. We systematically sampled 100 images for each stylistic category, resulting in a total of 500 images extracted from the COCO-O dataset. This dataset comprises a rich array of object types, providing a comprehensive evaluation ground for assessing the image understanding capabilities of the existing LMMs. Accordingly, the questions are prepared as asking the models to reason + +and answer questions about images, and the sample images with Q&A are shown in the first row of Fig. 2. + +Sensor-style Data Other than RGB cameras, Infrared sensors, X-ray sensors, Magnetic Resonance Imaging (MRI) sensors, and Computed Tomography (CT) sensors are often used to collect image data for human beings to perceive and analyze. Our benchmark examines whether the LMMs can still have the comparative reasoning capability toward these images that are intrinsically different from RGB photo images. In the case of Infrared data, we collected a set of 200 images from the datasets that focus on street scenes analysis toward with pedestrians, buildings, and vehicles [22, 25, 26, 40]. Questions for LMMs mainly revolve around analyzing the contents of these street scenes. Specifically, we design some thermal-related questions such as determining the hottest content. An illustrative example is provided in the second row of Fig. 2. For Low-energy X-ray (L-Xray) data from the SiXray-D dataset [47], we gathered 200 images, including 160 with prohibited objects (knife, scissors, wrench, gun, hammer) and 40 without any prohibited objects. Questions are designed to assess the ability to detect security risks in X-ray images, such as identifying forbidden objects. A sample question is presented in the second row of Fig. 2. In the medical application, high-energy X-ray (H-Xray), MRI and CT scanning images (200 each) were extracted from the [37]. Questions for LMMs in this context involve tasks like recognizing organs or identifying diseases in the respective images. By comparing inference results between different sensor types, our benchmark aims to uncover challenges associated with cross-sensor analysis. + +Application-style Data In order to establish a comprehensive cross-application-style benchmark, we systematically integrated diverse task domains into our evaluation framework. We consider four representative applications: remote sensing, autonomous driving, agent action prediction, and defect detection. Specifically, for the remote sensing application, we utilized 200 images sourced from the DOTA dataset [59] and annotated them in the VQA format from a satellite perspective. For the autonomous driving task, we extracted 250 images from the BDD100K dataset [65], featuring various weather conditions. Our primary focus was on designing questions and answers aimed at predicting driving actions such as going straight, turning left, or turning right. An illustrative example is provided in the third row of Fig. 2. + +To introduce the agent action prediction application, we meticulously selected 100 images from the Domestic Robot dataset [11] and an additional 117 images from the Open Game dataset [11]. The associated questions center around predicting the next action to be taken under the current circumstances. Furthermore, for evaluating defect detection capabilities, we curated 100 images from the Defect Detection dataset [6], with questions focused on identifying defect types and determining the presence of defects. This diverse set of applications and datasets collectively forms a diverse benchmark, allowing for a comprehensive assessment of the proposed approach across various application domains. More examples and Q&A of different artistic styles, sensor styles, application styles, and the constrain of answers can be found in the Appendix. + +# 3.3 Style Prompt Enhancement + +It is generally expected that additional prior information to LMMs' inputs can enhance performance. Nevertheless, such contextual priors (e.g. knowledge of artistic styles) might be unavailable in automated inference scenarios. We are inspired by human cognitive behavior that humans often comprehend the style of an image before delving into object recognition, akin to appreciating a painting by Van Gogh, where recognition of the impressionist style precedes identification of objects within the artwork. Other contextual priors, like sensor or application styles, can also enhance LMMs' performance. For example, identifying people in infrared images is easier when LMMs know that the images are captured by infrared thermal cameras since humans typically show consistent temperatures. + +With this insight, we investigate the efficacy of prompting LMMs to predict the style of an image before addressing associated questions. To this end, we devised Style Prompt Enhancement by introducing a text prompt, instructing the model to "Determine the style of the image before answering the question." SPE aims to leverage stylistic cues as a precursor to visual question answering, potentially enhancing the model's capability of visual reasoning. Despite its simplicity, SPE's versatility allows it to be applied directly to images of various styles without any additional prompt tuning or model training, clearly boosting the overall performance of LMMs in a zero-shot manner, as detailed in Section 4.3. + +# 4 Experiment + +# 4.1 Experimental Setup + +This paper evaluates a selection of existing open-source LMMs, which include LLaVA [38,39], MiniGPT4 [10,71], InstructBLIP [14], Otter [32], and OpenFlamingo-7B [4]. These models are initialized by pre-trained weights and evaluated on our BenchLMM. In addition, we use the variants of these models, where the language encoders and vision encoders are changed to investigate how different encoders have impacts on the cross-field capabilities. Moreover, we perform a quantitative evaluation of GPT-4V, the latest commercial LMM, with our dataset and compare its performance with the aforementioned open-source LMMs, and such a quantitative evaluation is the first time. + +In our performance evaluation, we adhere to previous works [38,39] methodology, employing the ChatGPT API to gauge the proximity of answers predicted by the LMMs to ground-truth answers. Here, we designate the output of the LMM as $y$ , and the ground-truth label as $\hat{y}$ . The evaluation utilizes the prompt "Compare the similarity between $y$ and $\hat{y}$ with a correctness score ranging from 0.0 (totally wrong) to 1.0 (totally right). The middle score provides the percentage of correctness." + +# 4.2 Evaluations of public LMMs + +Cross-artistic-style Capability In most existing works [17,21,33,34,41,67], LMMs are predominantly evaluated using images in the 'Photo' style, leading + +Table 1: Experimental results (\%) of public LMMs on BenchLMM's cross-artistic-style evaluation. Note that Average* represents the average accuracy computed over five artistic-style benchmarks. The best and the second best results are highlighted in bold and underline, respectively. + +
LMMsVisual EncoderLanguage ModelCommonCartoonHandmakePaintingSketchTattooAverage*
GPT-4V [62]--81.563.358.758.264.768.362.6
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B74.662.056.657.663.357.059.3
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B72.759.057.159.357.561.458.9
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B56.649.945.448.053.747.048.8
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B56.957.828.437.842.632.340.0
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B71.251.953.343.462.054.653.0
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B73.958.152.055.855.255.555.3
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B66.428.737.840.941.728.635.5
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B45.635.037.038.536.437.736.9
Otter-7B [32]CLIP-ViT-L/14MPT-7B64.845.149.039.844.142.444.1
+ +to a gap in understanding their performance across diverse artistic styles. We extend the evaluation scope by examining LMMs' performance with various artistic styles beyond the common 'Photo' style. The results detailed in Table 1 reveal a notable decline in LMMs' effectiveness when processing these artistic styles. This trend suggests a potential overfitting of LMMs to the 'Photo' style, highlighting their limited adaptability to varied artistic styles, a capability that humans typically possess. Interestingly, GPT-4V, as an outstanding commercial model, exhibits similar limitations in handling diverse styles. + +Besides, LLaVA-1.5-13B [38] is improved from LLaVA-13B [39] by introducing more training tricks and data, and LLaVA-1.5-13B outperforms LLaVA-13B by a clear margin on the 'Common' and other artistic styles. This comparison reveals the importance of data and training strategy as LLaVA-1.5-13B and LLaVA-13B use the same visual encoder and language encoder. However, we also find that a better performance on 'Common' does not necessarily indicate a better performance on other styles. For example, compared with InstructBLIP-7B, InstructBLIP-13B demonstrates better performance across five artistic-style benchmarks in terms of average accuracy, though it shows comparatively lower performance in the 'Common' style. Therefore, such results convey that it is important to conduct comprehensive evaluations to fully validate an LMM's capability rather than only focus on the 'Common' style, which shows the importance of our proposed benchmark. + +Cross-sensor Capability We benchmark the LMM models' cross-sensor capability in Table 2. Because GPT-4V denied to provide medical diagnostics or identify potential abnormalities in medical images like H-Xray/MRI/CT, GPT-4V is only evaluated on the Infrared and L-Xray sensor data and achieves the results of $55.0\%$ and $50.0\%$ respectively, which are far lower than the 'Common' style $(81.5\%)$ . Meanwhile, other LMM models' performance on images captured by X-ray, MRI, CT, and Infrared dropped significantly compared to + +Table 2: Experimental results (\%) of public LMMs on BenchLMM's cross-sensor evaluation. Note that GPT-4V denied to test medical images (H-Xray/MRI/CT). + +
LMMsVisual EncoderLanguage ModelInfraredL-XrayH-XrayMRICTAverage
GPT-4V [62]--55.050.0----
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B52.151.555.933.743.147.3
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B40.824.656.620.943.437.3
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B53.746.039.128.439.741.4
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B32.947.224.625.323.230.6
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B46.553.249.429.646.845.3
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B42.235.459.519.438.539.0
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B30.143.031.517.417.828.2
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B35.144.123.716.815.327.0
Otter-7B [32]CLIP-ViT-L/14MPT-7B30.246.439.147.846.342.0
+ +Table 3: Results of public LMMs on BenchLMM's cross-application experiments. + +
LMMsVisual EncoderLanguage ModelRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
GPT-4V [62]--69.735.856.074.064.460.0
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B65.628.649.021.460.345.2
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B63.632.632.019.547.038.9
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B38.731.759.043.828.240.3
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B33.322.648.034.231.734.2
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B61.731.344.015.459.642.2
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B63.536.742.024.830.839.6
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B34.517.640.022.218.826.6
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B27.920.841.040.113.628.7
Otter-7B [32]CLIP-ViT-L/14MPT-7B44.829.736.018.827.331.3
+ +the performance on 'Common' captured by the RGB camera sensor. Therefore, the intrinsic difference in the imaging process between RGB cameras and other sensors is also a crucial challenge to LMMs. With respect to the MRI, H-Xray, and CT images, the questions may require specific medical knowledge to answer, and it may be argued that the performance drop is due to the knowledge gap. It is also observed from the L-Xray, where the objects (i.e., knife, gun, hammer, etc.) can be recognized with only common knowledge, but the LMMs' performance on the L-Xray data is significantly lower than 'Common', which further validates the challenge brought by the sensor difference. + +Cross-application Capability In Table 3, we evaluate the different LMMs on various applications, finding that GPT-4V generally outperforms others, except in autonomous driving where its accuracy drops to $35.8\%$ . This decrease is likely due to the challenges in the Autonomous Driving dataset, such as recognizing low-resolution traffic signs and requiring specific driving knowledge, where both GPT-4 and open-source LMMs have poor performance. We also observe that LLaVA-1.5-13B achieves comparable performance as GPT-4V on the Remote Sensing and Defect Detection datasets, which seems to have comparable cross + +![](images/0380a6083d16477095a416ccc400a78abf498111c561d5e00f8b6ab19fe86115.jpg) + +![](images/33142e0e9e08c940bc339e886bfcaeb3340de61d4137574d2f0538bb87d58cc8.jpg) +Fig. 3: Error-reflection capability comparisons: Unlike LLaVA-1.5-13B, GPT-4V possesses the ability to analyze its errors when provided with the correct answers. + +![](images/6c1d652eb59e200fbdb12562d24478fcce6e20ee556b1a42bc31d72aec8d565d.jpg) + +![](images/99c115e4e8592c9be0f536820e8b711492a2e77aae19d67dd415162189b0dd35.jpg) + +![](images/8014eb06bb66a4f626987b8bb87350790aee9126fa7e82f4f6af747800e7a82c.jpg) + +![](images/4bb714b1b0881729cff988e04b0136ebc29c150e77483af4875789a93dcd5f8a.jpg) + +application capability. However, we surprisingly find that LLaVA-1.5-13B merely achieves $21.4\%$ but GPT-4V achieves up to $74.0\%$ . As GPT-4V is close-source, the technical details are not disclosed and we conjecture that the training data for GPT-4V may involve data similar to the Open Game dataset. + +We conclude from the experiments that: 1) Existing LMMs often degrade in performance with cross-style images; 2) Better performance on 'Common' data doesn't ensure better results on other styles; 3) Comprehensive evaluations across different dimensions are crucial, as errors in tasks like medical or security can lead to significant losses + +Error-reflection Capability To understand the LMMs further, we let the LMM do an error reflection that uses itself to parse the error information. When a model gives wrong answers, we warp up the conversation: the previous question input and predicted answer output. Specifically, we add "The correct answer is + +Table 4: Experimental results (%) of our proposed SPE on the BenchLMM benchmark. + +
LMMsCartoonHandmakePaintingSketchTattooAverage
LLaVA-1.5-13B [38]62.056.657.663.357.059.3
+ SPE (Ours)67.657.767.968.863.565.1 (+5.8)
InstructBLIP-13B [14]59.057.159.357.561.458.9
+ SPE (Ours)61.162.364.860.963.362.5 (+3.6)
LMMsInfraredL-XrayH-XrayMRICTAverage
LLaVA-1.5-13B [38]52.151.555.933.743.147.3
+ SPE (Ours)58.762.856.233.647.751.8 (+4.5)
InstructBLIP-13B [14]40.824.656.620.943.437.3
+ SPE (Ours)43.524.857.725.042.638.7 (+1.4)
LMMsRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
LLaVA-1.5-13B [38]65.628.649.021.460.345.2
+ SPE (Ours)68.029.451.022.761.746.6 (+1.4)
InstructBLIP-13B [14]63.632.632.019.547.038.9
+ SPE (Ours)63.538.831.019.049.440.3 (+1.4)
+ +[GT Answer]. Explain why your answer is incorrect. Use a single sentence." to the conversation (GT means Ground-truth), as the new input of LMMs. + +Although both GPT-4V and LLaVA-1.5-13B produce incorrect answers, GPT-4V demonstrates the ability to reason the correct answer through a detailed inference process. As illustrated in Fig. 3, GPT-4V can offer insightful analysis beyond the ground-truth answer, including the positions of clock hands for estimating time, the location of prohibited objects, and the traffic signals for driving planning. Moreover, from the second row of Fig. 3, we can see that the LLaVA model can just simply repeat the answer or say "I was wrong". By contrast, GPT-4V can do inference itself to provide the correct answer, such as recognizing the toy animals (Handmake) or the scratch on the capsule (Defect Detection). Through this study, we can see that the error-reflection capability of LLaVA-1.5-13B is inferior to that of GPT-4V. Therefore, we reveal that another important capability of a large multimodal model is whether the model can reflect how to derive the correct answer. Studying such capability is useful to common users. For example, when users use LMMs for self-learning, the user can provide reference answers and LMMs can provide a more detailed inference process, which can improve the learning effect. Moreover, the answer of GPT-4V can also be useful materials used to train open-source LMMs in the future. + +# 4.3 Experiments of Style Prompt Enhancement + +Comparison to Baseline Models In this section, we aim to study if a large multimodal model can be enhanced by prompting the model to estimate the style first and then answer the question, based on which we propose the Style Prompt + +Table 5: Complementary studies of our SPE with recent zero-shot chain-of-thought method ZS-COT [31] using LLaVA-1.5-13B [38] on the BenchLMM + +
MethodCartoonHandmakePaintingSketchTattooAverage
LLaVA-1.5-13B [38]62.056.657.663.357.059.3
+ZS-COT [31]64.359.770.662.871.065.7(+6.4)
+ZS-COT [31]+SPE (Ours)66.064.466.565.468.366.1(+6.8)
MethodInfraredL-XrayH-XrayMRICTAverage
LLaVA-1.5-13B [38]52.151.555.933.743.147.3
+ZS-COT [31]49.355.957.635.644.748.6(+1.3)
+ZS-COT [31]+SPE (Ours)54.470.456.735.345.452.4(+5.1)
MethodRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
LLaVA-1.5-13B [38]65.628.649.021.460.345.2
+ZS-COT [31]65.527.043.034.263.746.7(+1.5)
+ZS-COT [31]+SPE (Ours)76.024.049.052.160.152.2(+7.0)
+ +Enhancement (SPE) method and conduct experiments with our BenchLMM. Specifically, we examined our SPE with LLaVA-1.5-13B [38] and InstructBLIP-13B [14] because these two models are representative of better average performance than MiniGPT4 [71] and Otter [32]. + +The experimental results are presented in Table 4. As shown in the table, our proposed SPE can benefit the LMMs by achieving general improvement over different styles, which means that reasoning the style first is an effective prompting method to improve LMMs' visual reasoning capability. Besides, in the experiments of crossing to different applications (Remote Sensing, Autonomous Driving, etc.), we observe that its benefit is smaller than others. For example, LLaVA-1.5-13B gains $5.8\%$ and $4.5\%$ improvement in the experiments of cross-artistic and cross-sensor styles, while the improvement is merely $1.4\%$ when crossing applications. We conjecture that specific application knowledge could inferred from only the style, which limits the benefits of SPE. Moreover, we observe that LLaVA-1.5-13 generally obtains more benefits than InstructBLIP-13B, which means that the stronger the LMM, the more benefits from SPE, which is because a stronger LMM has stronger inference capability and can better utilize prior information. + +Comparison to Other Prompting Methods Our proposed SPE method enhances LMMs' performance by prompting LMMs to estimate the input image's style first, a technique akin to chain-of-thought prompting approaches [31, 45]. In Table 5, we compare our proposed SPE with the state-of-the-art prompting Chain-of-Thought (COT) [31], which is a zero-shot, to demonstrate the efficacy of SPE. As shown, the COT method can improve LMM's performance, and our SPE is compatible with the COT methods and further enhances the model's capability, which is demonstrated by the general improvement brought by “+ZS-COT+SPE” as shown in Table 5. We also notice that in some cases, using SPE + +![](images/e340eaf9b742dc6c007d5b6aba0d1180d31010fd95faf817f5e5c20210003519.jpg) +Fig. 4: Qualitative comparisons between LLaVA-1.5-13B and LLaVA-1.5-13B + SPE (Ours) on style-transfer images. + +Table 6: Quantitative Results of our proposed SPE on style-transfer images. + +
LMMs|Cartoon Handmake Painting Sketch Tattoo | Average
LLaVA-1.5-13B [38] + SPE (Ours)62.161.152.468.841.257.1
67.462.159.469.445.560.8 (+3.7)
InstructBLIP-13B [14] + SPE (Ours)69.370.861.353.865.064.0
73.071.664.262.165.367.2 (+3.2)
+ +with COT could lead to performance decline. The investigation is beyond the scope of this paper, which will be explored in our future work. + +Results on the Style-transfer Images We employ the latest commercial style transfer model, DALL-E3 [50], to transfer the 'Common' style images into different artistic styles. As depicted in Fig. 4, these images share similar structures and layouts. However, they are just in different artistic styles, and the tested LLaVA model also makes mistakes. The significant performance degradation on the transferred images is also proved by the quantitative results presented in Table 6, which validates that LMMs suffer from difficulties in recognizing images of different styles. Nevertheless, our SPE significantly enhances the cross-style capability of LMMs in this challenging task, as shown in Fig. 4 and Table 6. + +In-depth Analysis of SPE Besides quantitative assessments, we explore the inner workings of SPE to understand how it improves LMMs. As illustrated in Fig. 5, LMMs are adept at accurately estimating the style of images. For instance, in the figure's first row and second column, the depicted style is 'Handmake,' and the model accurately recognizes this style as akin to a toy, indicating a correct match. Moreover, when estimating the style, the model can also understand the context first and thus better answer the final question, as shown in the examples of Remote Sensing and Defect Detection. Our findings suggest that prompting LMMs to first identify the style of images serves as an effective, training-free method to improve the models' proficiency in handling cross-style images. + +# 5 Conclusion + +In this work, we introduce BenchLMM, a novel benchmark for quantitatively evaluating large multimodal models (LMMs) across varied visual distribution + +![](images/1135ef199bff7c35cebd55bcacb9ad3c2f3ea7ba6ad9e4f23730986f4e29f609.jpg) +Fig. 5: Evaluation of LLaVA-1.5-13B with SPE on the BenchLMM. + +shifts, including artistic, sensor, and application styles. Our evaluations of numerous existing LMMs highlight two key findings: 1) LMMs typically underperform with non-common image styles. 2) Our style prompt enhancement approach, inspired by human perception of art, significantly improves LMMs' visual reasoning without extra fine-tuning. Additionally, our error-reflection study shows that stronger LMMs can self-diagnose errors by providing insights not given by humans, whereas less capable models merely restate correct answers without new insights. Our benchmark offers a comprehensive tool for LMM research, emphasizing the importance of error-reflection capabilities in the future. + +# Acknowledgement + +This work was carried out at Rapid-Rich Object Search (ROSE) Lab, School of Electrical & Electronic Engineering, Nanyang Technological University. This research is supported by the NTU-PKU Joint Research Institute (a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation. + +# References + +1. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8948-8957 (2019) +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015) +4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023) +5. Ben-Younes, H., Cadene, R., Cord, M., Thome, N.: Multimodal tucker fusion for visual question answering. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2612-2620 (2017) +6. Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., Steger, C.: The mvtec anomaly detection dataset: a comprehensive real-world dataset for unsupervised anomaly detection. International Journal of Computer Vision 129(4), 1038-1059 (2021) +7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +8. Cadene, R., Ben-Younes, H., Cord, M., Thome, N.: Murel: Multimodal relational reasoning for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1989-1998 (2019) +9. Cao, Q., Wan, W., Wang, K., Liang, X., Lin, L.: Linguistically routing capsule network for out-of-distribution visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1614-1623 (2021) +10. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning (2023) +11. Chen, L., Zhang, Y., Ren, S., Zhao, H., Cai, Z., Wang, Y., Wang, P., Liu, T., Chang, B.: Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4-vision and beyond. arXiv preprint arXiv:2310.02071 (2023) + +12. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollár, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) +13. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022) +14. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023) +15. Dancette, C., Cadene, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1574-1583 (2021) +16. Ding, Y., Yu, J., Liu, B., Hu, Y., Cui, M., Wu, Q.: Mukea: Multimodal knowledge extraction and accumulation for knowledge-based visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5089-5098 (2022) +17. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models. arXiv preprint arXiv:2306.13394 (2023) +18. Fukui, A., Park, D.H., Yang, D., Rohrbach, A., Darrell, T., Rohrbach, M.: Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847 (2016) +19. Gao, F., Ping, Q., Thattai, G., Reganti, A., Wu, Y.N., Natarajan, P.: Transform-retrieve-produce: Natural language-centric outside-knowledge visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5067-5077 (2022) +20. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhan, S.: Chain of thought prompt tuning in vision language models. arXiv preprint arXiv:2304.07919 (2023) +21. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6904-6913 (2017) +22. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017) +23. Han, X., Wang, S., Su, C., Huang, Q., Tian, Q.: Greedy gradient ensemble for robust visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1584-1593 (2021) +24. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019) +25. Hwang, S., Park, J., Kim, N., Choi, Y., So Kweon, I.: Multispectral pedestrian detection: Benchmark dataset and baseline. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1037-1045 (2015) +26. Jia, X., Zhu, C., Li, M., Tang, W., Zhou, W.: LLVIP: A visible-infrared paired dataset for low-light vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3496-3504 (2021) + +27. Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2901-2910 (2017) +28. Johnson, J., Hariharan, B., Van Der Maaten, L., Hoffman, J., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Inferring and executing programs for visual reasoning. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2989-2998 (2017) +29. Kim, J.H., Jun, J., Zhang, B.T.: Bilinear attention networks. Advances in neural information processing systems 31 (2018) +30. Kim, J.H., On, K.W., Lim, W., Kim, J., Ha, J.W., Zhang, B.T.: Hadamard product for low-rank bilinear pooling. arXiv preprint arXiv:1610.04325 (2016) +31. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. Advances in neural information processing systems 35, 22199-22213 (2022) +32. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) +33. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023) +34. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023) +35. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023) +36. Liang, P.P., Lyu, Y., Fan, X., Wu, Z., Cheng, Y., Wu, J., Chen, L., Wu, P., Lee, M.A., Zhu, Y., et al.: Multibench: Multiscale benchmarks for multimodal representation learning. arXiv preprint arXiv:2107.07502 (2021) +37. Liu, B., Zhan, L.M., Xu, L., Ma, L., Yang, Y., Wu, X.M.: Slake: A semantically-labeled knowledge-enhanced dataset for medical visual question answering. In: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). pp. 1650-1654. IEEE (2021) +38. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) +39. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +40. Liu, J., Fan, X., Huang, Z., Wu, G., Liu, R., Zhong, W., Luo, Z.: Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5802-5811 (June 2022) +41. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: MMBench: Is Your Multi-modal Model an All-around Player? arXiv preprint arXiv:2307.06281 (2023) +42. Manyika, J.: An overview of bard: an early experiment with generative ai. AI. Google Static Documents (2023) +43. Mao, X., Chen, Y., Zhu, Y., Chen, D., Su, H., Zhang, R., Xue, H.: COCO-O: A Benchmark for Object Detectors under Natural Distribution Shifts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6339-6350 (2023) + +44. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3195-3204 (2019) +45. Mitra, C., Huang, B., Darrell, T., Herzig, R.: Compositional chain-of-thought prompting for large multimodal models. arXiv preprint arXiv:2311.17076 (2023) +46. Nam, H., Ha, J.W., Kim, J.: Dual attention networks for multimodal reasoning and matching. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 299-307 (2017) +47. Nguyen, H.D., Cai, R., Zhao, H., Kot, A.C., Wen, B.: Towards more efficient security inspection via deep learning: a task-driven x-ray image cropping scheme. Micromachines 13(4), 565 (2022) +48. OpenAI, R.: Gpt-4 technical report. arxiv 2303.08774. View in Article (2023) +49. Shah, M., Chen, X., Rohrbach, M., Parikh, D.: Cycle-consistency for robust visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6649-6658 (2019) +50. Shi, Z., Zhou, X., Qiu, X., Zhu, X.: Improving image captioning with better use of captions. arXiv preprint arXiv:2006.11807 (2020) +51. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 742-758. Springer (2020) +52. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards VQA models that can read. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8317-8326 (2019) +53. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019) +54. Tascon-Morales, S., Márquez-Neila, P., Sznitman, R.: Logical implications for visual question answering consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6725-6735 (2023) +55. Teney, D., Abbasnejad, E., van den Hengel, A.: Unshuffling data for improved generalization in visual question answering. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1417-1427 (2021) +56. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +57. Tsimpoukelli, M., Menick, J.L., Cabi, S., Eslami, S., Vinyals, O., Hill, F.: Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems 34, 200-212 (2021) +58. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +59. Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: Dota: A large-scale dataset for object detection in aerial images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3974-3983 (2018) +60. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023) + +61. Yang, X., Gao, C., Zhang, H., Cai, J.: Auto-parsing network for image captioning and visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2197-2207 (2021) +62. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of LMMs: Preliminary explorations with GPT-4V (ision). arXiv preprint arXiv:2309.17421 (2023) +63. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023) +64. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2633-2642 (2020). https://doi.org/10.1109/CVPR42600.2020.00271 +66. Yu, S., Wu, P., Liang, P.P., Salakhutdinov, R., Morency, L.P.: Pacs: A dataset for physical audiovisual commonsense reasoning. In: European Conference on Computer Vision. pp. 292-309. Springer (2022) +67. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023) +68. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6720-6731 (2019) +69. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) +70. Zhou, Y., Ren, T., Zhu, C., Sun, X., Liu, J., Ding, X., Xu, M., Ji, R.: Trar: Routing the attention spans in transformer for visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2074-2084 (2021) +71. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models (2023) \ No newline at end of file diff --git a/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/images.zip b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..352cb99a316deb66b4a6a056be73fc0846e8295f --- /dev/null +++ b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08dda56259a24409f4a4fe7b55984fa8568e5953aead72ed98c33a31761011e2 +size 917585 diff --git a/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/layout.json b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a991213fa998cbad98a57b1110b36d13041c72be --- /dev/null +++ b/2024/BenchLMM_ Benchmarking Cross-style Visual Capability of Large Multimodal Models/layout.json @@ -0,0 +1,8983 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "text", + "content": "BenchLMM: Benchmarking Cross-style Visual Capability of Large Multimodal Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "spans": [ + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "text", + "content": "Rizhao Cai" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Zirui Song" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Dayan Guan" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Zhenhao Chen" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Yaohang Li" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Xing Luo" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Chenyu Yi" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "text", + "content": ", and Alex Kot" + }, + { + "bbox": [ + 159, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 209, + 201, + 403, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 201, + 403, + 213 + ], + "spans": [ + { + "bbox": [ + 209, + 201, + 403, + 213 + ], + "type": "text", + "content": "1 Nanyang Technological University, Singapore" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 213, + 466, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 213, + 466, + 224 + ], + "spans": [ + { + "bbox": [ + 146, + 213, + 466, + 224 + ], + "type": "text", + "content": "2 University of Technology Sydney, Australia; Northeastern University, China;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 223, + 389, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 223, + 389, + 235 + ], + "spans": [ + { + "bbox": [ + 225, + 223, + 389, + 235 + ], + "type": "text", + "content": "3 Harbin Institute of Technology, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 234, + 472, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 234, + 472, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 472, + 246 + ], + "type": "text", + "content": "4 Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 245, + 367, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 245, + 367, + 257 + ], + "spans": [ + { + "bbox": [ + 247, + 245, + 367, + 257 + ], + "type": "text", + "content": "5 Zhejiang University, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 183, + 258, + 429, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 258, + 429, + 268 + ], + "spans": [ + { + "bbox": [ + 183, + 258, + 429, + 268 + ], + "type": "text", + "content": "† Equal Contribution Corresponding author (guandayan@hit.edu.cn)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 293, + 455, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 293, + 455, + 491 + ], + "spans": [ + { + "bbox": [ + 159, + 293, + 455, + 491 + ], + "type": "text", + "content": "Abstract. Large Multimodal Models (LMMs) such as GPT-4V and LLaVA have shown remarkable capabilities in visual reasoning on data in common image styles. However, their robustness against diverse style shifts, crucial for practical applications, remains largely unexplored. In this paper, we propose a new benchmark, BenchLMM, to assess the robustness of LMMs toward three different styles: artistic image style, imaging sensor style, and application style. Utilizing BenchLMM, we comprehensively evaluate state-of-the-art LMMs and reveal: 1) LMMs generally suffer performance degradation when working with other styles; 2) An LMM performs better than another model in common style does not guarantee its superior performance in other styles; 3) LMMs' reasoning capability can be enhanced by prompting LMMs to predict the style first, based on which we propose a versatile and training-free method for improving LMMs; 4) An intelligent LMM is expected to interpret the causes of its errors when facing stylistic variations. We hope that our benchmark and analysis can shed new light on developing more intelligent and versatile LMMs. The benchmark and evaluation have been released on https://github.com/AIFEG/BenchLMM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 510, + 230, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 230, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 230, + 522 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": "The dynamic landscape of computer vision has recently been shaped by the rise of Large Multimodal Models (LMMs). These models [10, 14, 32, 39, 71], incorporating visual and textual data, are pivotal in developing general-purpose assistants for visual reasoning, a widely-studied task that involves reasoning and answering questions based on images. Their growing prominence and the enthusiasm surrounding visual reasoning are evident from numerous studies such as GPT-4V [62]. As the interest in LMMs swells within the research community, understanding their depth, breadth, and limitations becomes imperative." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "To comprehensively understand the visual reasoning capabilities of LMMs, numerous benchmarks [17, 21, 33, 34, 41, 67] have been proposed to systematically evaluate the linguistic capabilities of LMMs. These linguistic capabilities" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 220, + 177 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 220, + 177 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 220, + 177 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 220, + 177 + ], + "type": "image", + "image_path": "404bb180824da96571d9ab6fe665a23fdf682a8ebbbf6904cd560edb5cdd35fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 178, + 181, + 200 + ], + "lines": [ + { + "bbox": [ + 139, + 178, + 181, + 200 + ], + "spans": [ + { + "bbox": [ + 139, + 178, + 181, + 200 + ], + "type": "text", + "content": "Q: How many forks? \nA (GT): Two \nA (GPT-4V): Two \nA (LLaVA-1: 5) Two" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 115, + 306, + 178 + ], + "blocks": [ + { + "bbox": [ + 222, + 115, + 306, + 178 + ], + "lines": [ + { + "bbox": [ + 222, + 115, + 306, + 178 + ], + "spans": [ + { + "bbox": [ + 222, + 115, + 306, + 178 + ], + "type": "image", + "image_path": "22939106675c9ad1454f1a74756e33be497feb2114334d0298ca783a0f8ea8d4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 178, + 295, + 200 + ], + "lines": [ + { + "bbox": [ + 224, + 178, + 295, + 200 + ], + "spans": [ + { + "bbox": [ + 224, + 178, + 295, + 200 + ], + "type": "text", + "content": "[Painting]Q:How many elephants? A (GT): Seven A (GPT-4V):Five A(LLAVA-1.5):Five" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 115, + 391, + 178 + ], + "blocks": [ + { + "bbox": [ + 307, + 115, + 391, + 178 + ], + "lines": [ + { + "bbox": [ + 307, + 115, + 391, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 115, + 391, + 178 + ], + "type": "image", + "image_path": "dcee0d27a1f39563afdbacaf526c3d584184758197c7ab7e9d3002689d3b8e16.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 178, + 389, + 201 + ], + "lines": [ + { + "bbox": [ + 309, + 178, + 389, + 201 + ], + "spans": [ + { + "bbox": [ + 309, + 178, + 389, + 201 + ], + "type": "text", + "content": "[Xray] Q: What are prohibited objects? \nA (GT): None \nA (GPT-4V): Gun \nA (LLVaA-1.5): A knife and a scissor" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 211, + 481, + 244 + ], + "lines": [ + { + "bbox": [ + 131, + 211, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 481, + 244 + ], + "type": "text", + "content": "Fig. 1: The motivation of our work. State-of-the-art LMMs, such as GPT-4V and LLaVA-1.5 can understand the images in a common style. However, these models have difficulties in reasoning images in other styles and deliver wrong answers." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 393, + 116, + 477, + 173 + ], + "blocks": [ + { + "bbox": [ + 393, + 116, + 477, + 173 + ], + "lines": [ + { + "bbox": [ + 393, + 116, + 477, + 173 + ], + "spans": [ + { + "bbox": [ + 393, + 116, + 477, + 173 + ], + "type": "image", + "image_path": "34c9bd7e9dfad720e12e11ec50f92198090d351c476e418930f531b04b805486.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 174, + 476, + 201 + ], + "lines": [ + { + "bbox": [ + 394, + 174, + 476, + 201 + ], + "spans": [ + { + "bbox": [ + 394, + 174, + 476, + 201 + ], + "type": "text", + "content": "[Driving] Q: What is the action for the ego-vehicle during autonomous driving? A (GT): Go straight A (GPT-4V): Stop at traffic light A (LLAvA-1.5): Stop" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 269, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 425 + ], + "type": "text", + "content": "are assessed through various types of responses for visual reasoning tasks, including Yes/No answers [17], multi-choice answers [33,41], single word or phrase answers [21,24,44]. However, while these benchmarks investigate textual diversity, they may fall short in terms of visual diversity as most images in these benchmarks are photographic images sourced from the internet. Despite providing valuable insights into LMMs' performance, there remains a gap in understanding how LMMs handle shifts in visual distribution. Although the recently released GPT-4V [62] claims remarkable capabilities across diverse application scenarios, only a few samples of qualitative results are presented in the technical report [62], and the quantitative evaluations across different scenarios are missing. Moreover, according to our preliminary study as illustrated in Fig. 1, we reveal that LMMs, even the state-of-the-art GPT-4V, tend to make mistakes in reasoning images with non-common styles." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 426, + 483, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 483, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 483, + 557 + ], + "type": "text", + "content": "To further investigate LMMs' cross-style visual reasoning capability, we propose a novel benchmark BenchLMM, which systematically assesses existing LMMs in three distinct styles of distribution shifts: artistic styles, sensor styles, and application styles. Firstly, images of the common style are realistic, but images can be in other artistic styles, such as Cartoon, Painting, Sketch, etc. Besides, there are different imaging sensors other than RGB cameras, such as infrared and X-ray. The discrepancy of sensors induces the distribution shifts of images. Moreover, we also identify the shift of knowledge required for different application styles. For example, the image recognition application requires knowledge about object shape and appearance, while the application of decision-making for robot action requires application-specific knowledge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "With our BenchLMM, we conduct in-depth analysis through experiments of evaluating advanced LMMs. The comprehensive experiments reveal that while these cutting-edge LMMs excel with commonly styled data, their performance significantly declines with data in other styles. Interestingly, we also discover that the superior performance of an LMM in common data styles does not necessarily predict better performance across other styles. Furthermore, our investigation shows that prompting an LMM to predict the style of the data before answering a question can significantly boost its reasoning abilities. Based on this insight, we introduce the Style Prompt Enhancement (SPE) technique. SPE is a novel" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "approach that improves LMMs' performance without additional training and operates effectively in a zero-shot manner. Moreover, we analyze the failure cases by conducting an in-depth error-reflection analysis by asking an LMM to interpret why it made mistakes in the Q&A. We find that a more intelligent LMM can understand the detailed reason why it made mistakes and can learn to derive the correct answer, which is an important capability that deserves attention when developing LMMs. We hope that our benchmark and analysis can shed new light on developing more intelligent and versatile LMMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 416 + ], + "type": "text", + "content": "Our contributions can be summarized below: 1) BenchchLMM: We introduce BenchLMM, a pioneering benchmark specifically designed to quantitatively evaluate the performance of LMMs' visual reasoning capability across visual styles in artistic, sensor, and application shifts. 2) Comprehensive evaluation of cutting-edge LMMs: We conduct extensive quantitative benchmark experiments on LMMs, including the commercial GPT-4V. Our rigorous and quantitative evaluation reveals LMMs' limitations in versatility and performance across different data styles. This finding challenges the prevailing assumption of universal applicability and highlights the need for specialized strategies to enhance LMM adaptability. 3) Insightful Analysis and Enhancement Strategy: Beyond mere performance assessment, we conduct an in-depth error-reflection analysis, where LMMs are prompted to identify the reasons behind their errors in question-and-answer tasks. This study reveals that more sophisticated LMMs can recognize their mistakes and also learn from them to improve future responses. Moreover, we introduce a novel SPE technique, as a training-free and zero-shot strategy, to significantly boost LMMs' reasoning capabilities across varied data styles, marking a substantial advancement in LMM development." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 435, + 240, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 240, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 240, + 449 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 593 + ], + "type": "text", + "content": "Visual Reasoning Visual reasoning, also known as visual question answering (VQA), aims to reason and answer questions about visual information [3,27,28]. Early developments [5, 18, 30] focus on designing more complex fusion mechanisms instead of simple summation fusion [3]. By including modern attention mechanism [58], numerous works [8, 29, 46, 53] contributed to the creation of transformer-based vision-language models. Recent methods have proposed to improve other aspects of visual question answering, including avoiding shortcut learning [15, 23], addressing the linguistic bias [9, 55], improving transformer-based vision-language models [61, 70], external knowledge integration [16, 19], and consistency regularization [49, 54]. Lastly, the advent of large multimodal models has markedly enhanced the capabilities of visual reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Large Multimodal Models Recent advancements in large language models (LLMs) [7, 13, 48, 56] have inspired the development of LMMs that enhance LLMs with vision-language capabilities. Existing LMMs can be categorized into three main classes: multimodal instruction tuning [10, 14, 32, 38, 39, 71] that involves the finetuning of pre-trained LLMs on multi-modal instruction-following data; multimodal in-context learning [2, 32, 57, 63] that derives insights from" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "text", + "content": "BenchLMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "few-shot task-specific samples; multimodal chain-of-thought [20,69] that encourages LMMs to articulate both the final answer and the underlying reasoning process. Additionally, some commercial LMMs like Google's Bard [42] and Microsoft's GPT-4V [62], with detailed information about their architecture remaining undisclosed, have shown impressive visual capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 296 + ], + "type": "text", + "content": "Evaluation of LMMs To evaluate their capabilities, existing LMMs have been tested on a number of vision-language datasets. These datasets focus on specific visual reasoning tasks by designing different kinds of Q&A task, and the specific task includes visual recognition [21, 44], image description [1, 12], scene text understanding [51, 52, 64], object hallucination evaluation [35], complex reasoning [39, 67], and commonsense reasoning [66, 68]. To facilitate a comprehensive comparison of LMMs, several benchmarks [17, 36, 60] have been established by encompassing various visual reasoning tasks with common-style images and diverse Q&A pairs. However, the above benchmarks do not consider different styles of visual distribution shift, which motivates us to propose our BenchLMM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 316, + 225, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 225, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 225, + 329 + ], + "type": "text", + "content": "3 BenchLMM" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 343, + 216, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 216, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 216, + 355 + ], + "type": "text", + "content": "3.1 Motivation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 479, + 411 + ], + "type": "text", + "content": "LMMs (e.g., GPT-4V) are claimed to excel in various scenarios [62], shown with qualitative results. Nevertheless, we provide comprehensive and quantitative results experiments with our BenchLMM reveal a paradox. We will elaborate on the benchmark that encompasses different cross-style data as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 480, + 533 + ], + "type": "text", + "content": "Cross-artistic Style Existing benchmarks for LMMs commonly use photo images for evaluation, which we name as 'Common' style. Meanwhile, there are diverse image styles, including Cartoon, Sketch, etc., as shown and compared in Fig. 2. The object in other artistic styles have different distributions of object characteristics. For instance, a colorful zebra, common in painting and tattoo styles, is rare in 'Common' style images. While humans easily interpret such characteristics, the extent of LMMs' understanding of the other styles remains uncertain. This uncertainty arises from potential insufficient training on datasets encompassing these diverse styles. Therefore, we propose an evaluation to assess LMMs' robustness in analyzing images across various artistic styles." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "content": "Cross-sensor Style Photographic images are commonly obtained through RGB cameras for information sensing. Alongside RGB cameras, other sensor types are also important for capturing images. For instance, infrared cameras are frequently utilized in surveillance systems to capture images during nighttime, as RGB cameras rely on adequate illumination for optimal image quality. Additionally, X-ray sensors are used in airport security checks, enabling the detection of prohibited items within luggage. These sensors feature distinct imaging processes compared to RGB cameras. Consequently, images acquired through various sensors exhibit unique intrinsic properties. The question of whether the performance of LMMs remains remarkable with other sensor data has not been thoroughly investigated. This issue will be addressed in our benchmark study." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 206, + 164 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 206, + 164 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 206, + 164 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 206, + 164 + ], + "type": "image", + "image_path": "f2d2babc524998a0bf475f18c4582ac1e28a99862186cc4d0d6561d5f93d6239.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 165, + 181, + 176 + ], + "lines": [ + { + "bbox": [ + 141, + 165, + 181, + 176 + ], + "spans": [ + { + "bbox": [ + 141, + 165, + 181, + 176 + ], + "type": "text", + "content": "Q: What time is it? A: 2:57:39" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 115, + 274, + 163 + ], + "blocks": [ + { + "bbox": [ + 209, + 115, + 274, + 163 + ], + "lines": [ + { + "bbox": [ + 209, + 115, + 274, + 163 + ], + "spans": [ + { + "bbox": [ + 209, + 115, + 274, + 163 + ], + "type": "image", + "image_path": "4be0ed448c216202254e1066ace484dd64c0887d12b1f9b8e610c1bbb95e8c2d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 276, + 115, + 342, + 164 + ], + "blocks": [ + { + "bbox": [ + 276, + 115, + 342, + 164 + ], + "lines": [ + { + "bbox": [ + 276, + 115, + 342, + 164 + ], + "spans": [ + { + "bbox": [ + 276, + 115, + 342, + 164 + ], + "type": "image", + "image_path": "4424394d2d32f6650034b12110bceeec2d7b594dc316c94b2164dadc46ce58e9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 165, + 340, + 176 + ], + "lines": [ + { + "bbox": [ + 277, + 165, + 340, + 176 + ], + "spans": [ + { + "bbox": [ + 277, + 165, + 340, + 176 + ], + "type": "text", + "content": "Q: Who is sitting on the couch? A: A woman" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 344, + 115, + 410, + 164 + ], + "blocks": [ + { + "bbox": [ + 344, + 115, + 410, + 164 + ], + "lines": [ + { + "bbox": [ + 344, + 115, + 410, + 164 + ], + "spans": [ + { + "bbox": [ + 344, + 115, + 410, + 164 + ], + "type": "image", + "image_path": "698a9a8b89db27df8fff7a464e3d07f55ce4c5feea13f35d8baa49b2accf3164.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 165, + 392, + 176 + ], + "lines": [ + { + "bbox": [ + 345, + 165, + 392, + 176 + ], + "spans": [ + { + "bbox": [ + 345, + 165, + 392, + 176 + ], + "type": "text", + "content": "Q: What kind of sport? A: football" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 412, + 121, + 479, + 164 + ], + "blocks": [ + { + "bbox": [ + 438, + 115, + 452, + 121 + ], + "lines": [ + { + "bbox": [ + 438, + 115, + 452, + 121 + ], + "spans": [ + { + "bbox": [ + 438, + 115, + 452, + 121 + ], + "type": "text", + "content": "Tattoo" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 412, + 121, + 479, + 164 + ], + "lines": [ + { + "bbox": [ + 412, + 121, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 412, + 121, + 479, + 164 + ], + "type": "image", + "image_path": "cabfe674ef11cc3714c16781c111d0578268257bf16c11bc2781ce0e842aad14.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 165, + 463, + 176 + ], + "lines": [ + { + "bbox": [ + 414, + 165, + 463, + 176 + ], + "spans": [ + { + "bbox": [ + 414, + 165, + 463, + 176 + ], + "type": "text", + "content": "Q: What is in the image? A: A colorful zebra" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 133, + 178, + 206, + 227 + ], + "blocks": [ + { + "bbox": [ + 133, + 178, + 206, + 227 + ], + "lines": [ + { + "bbox": [ + 133, + 178, + 206, + 227 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 206, + 227 + ], + "type": "image", + "image_path": "c031735a6bd5d0bea0fd295827d09b7be8f5e1a38c91fbf4c699d0b4ebf8cd68.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 229, + 196, + 240 + ], + "lines": [ + { + "bbox": [ + 141, + 229, + 196, + 240 + ], + "spans": [ + { + "bbox": [ + 141, + 229, + 196, + 240 + ], + "type": "text", + "content": "Q: How many pedestrians? A: 4" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 209, + 178, + 274, + 228 + ], + "blocks": [ + { + "bbox": [ + 209, + 164, + 268, + 178 + ], + "lines": [ + { + "bbox": [ + 209, + 164, + 268, + 178 + ], + "spans": [ + { + "bbox": [ + 209, + 164, + 268, + 178 + ], + "type": "text", + "content": "A: A bronze sculpture of a person with a very large head" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 178, + 274, + 228 + ], + "lines": [ + { + "bbox": [ + 209, + 178, + 274, + 228 + ], + "spans": [ + { + "bbox": [ + 209, + 178, + 274, + 228 + ], + "type": "image", + "image_path": "708b1202d956d4b16cd5625bc8484eed65f62761e9035bfc4036438e4210fdce.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 229, + 274, + 241 + ], + "lines": [ + { + "bbox": [ + 209, + 229, + 274, + 241 + ], + "spans": [ + { + "bbox": [ + 209, + 229, + 274, + 241 + ], + "type": "text", + "content": "Q: What are prohibited objects? \nA: Gun, knife" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 281, + 178, + 337, + 228 + ], + "blocks": [ + { + "bbox": [ + 281, + 178, + 337, + 228 + ], + "lines": [ + { + "bbox": [ + 281, + 178, + 337, + 228 + ], + "spans": [ + { + "bbox": [ + 281, + 178, + 337, + 228 + ], + "type": "image", + "image_path": "288027131b791a85b0e6d07efea4d486e0636593c8828a7ca31e99089c59981b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 229, + 332, + 241 + ], + "lines": [ + { + "bbox": [ + 278, + 229, + 332, + 241 + ], + "spans": [ + { + "bbox": [ + 278, + 229, + 332, + 241 + ], + "type": "text", + "content": "Q: What is the main organ? \nA: Lung" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 344, + 178, + 410, + 228 + ], + "blocks": [ + { + "bbox": [ + 344, + 178, + 410, + 228 + ], + "lines": [ + { + "bbox": [ + 344, + 178, + 410, + 228 + ], + "spans": [ + { + "bbox": [ + 344, + 178, + 410, + 228 + ], + "type": "image", + "image_path": "3d1c1f07e0627ce2b5b668d903b5a0a36320db83ab59f8440ada08f192aaf07f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 229, + 402, + 241 + ], + "lines": [ + { + "bbox": [ + 346, + 229, + 402, + 241 + ], + "spans": [ + { + "bbox": [ + 346, + 229, + 402, + 241 + ], + "type": "text", + "content": "Q: What kind of disease? A: Brain edema, brain tumor" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 413, + 178, + 479, + 229 + ], + "blocks": [ + { + "bbox": [ + 413, + 178, + 479, + 229 + ], + "lines": [ + { + "bbox": [ + 413, + 178, + 479, + 229 + ], + "spans": [ + { + "bbox": [ + 413, + 178, + 479, + 229 + ], + "type": "image", + "image_path": "27c2cd8ad1a0c8faba2c2eb7310deb1b207c4f5b4897dd164a688650dfc11e81.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 230, + 464, + 241 + ], + "lines": [ + { + "bbox": [ + 413, + 230, + 464, + 241 + ], + "spans": [ + { + "bbox": [ + 413, + 230, + 464, + 241 + ], + "type": "text", + "content": "Q: What kind of disease? \nA: Lung cancer" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 133, + 243, + 206, + 301 + ], + "blocks": [ + { + "bbox": [ + 133, + 243, + 206, + 301 + ], + "lines": [ + { + "bbox": [ + 133, + 243, + 206, + 301 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 206, + 301 + ], + "type": "image", + "image_path": "bc3aca5efe34effc91a59b408995d2f6d8215a42242c2b2c256840fe86e8889d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 302, + 202, + 312 + ], + "lines": [ + { + "bbox": [ + 141, + 302, + 202, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 202, + 312 + ], + "type": "text", + "content": "Q: How many sport facilities? A: 3" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 323, + 480, + 367 + ], + "lines": [ + { + "bbox": [ + 131, + 323, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 480, + 367 + ], + "type": "text", + "content": "Fig. 2: Benchmark examples. For each specific field, one example image and its corresponding Q&A have been shown. Note that, for a simple presentation, the questions in Domestic Robot and Open Game have been simplified from multiple-choice format. Please see the Appendix for more examples and detailed questions." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 209, + 243, + 274, + 289 + ], + "blocks": [ + { + "bbox": [ + 209, + 243, + 274, + 289 + ], + "lines": [ + { + "bbox": [ + 209, + 243, + 274, + 289 + ], + "spans": [ + { + "bbox": [ + 209, + 243, + 274, + 289 + ], + "type": "image", + "image_path": "fcfe2f66f387cebd09afbfa3f4132f468b81318ffc366443c59cb035be9ee25b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 290, + 270, + 313 + ], + "lines": [ + { + "bbox": [ + 209, + 290, + 270, + 313 + ], + "spans": [ + { + "bbox": [ + 209, + 290, + 270, + 313 + ], + "type": "text", + "content": "Q: What is the intended action for the main vehicle in an autonomous driving scenario? A: Go straight" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 276, + 243, + 342, + 289 + ], + "blocks": [ + { + "bbox": [ + 276, + 243, + 342, + 289 + ], + "lines": [ + { + "bbox": [ + 276, + 243, + 342, + 289 + ], + "spans": [ + { + "bbox": [ + 276, + 243, + 342, + 289 + ], + "type": "image", + "image_path": "f6c1e9b620e6aa0a1779dd5688dbf400fe7ee8f0560c89be275330b62a4c20f5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 290, + 342, + 313 + ], + "lines": [ + { + "bbox": [ + 277, + 290, + 342, + 313 + ], + "spans": [ + { + "bbox": [ + 277, + 290, + 342, + 313 + ], + "type": "text", + "content": "Q: Your goal is to repair the broken clock in your hands. What action should you do next? A: Look for repair tools" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 344, + 247, + 410, + 289 + ], + "blocks": [ + { + "bbox": [ + 344, + 247, + 410, + 289 + ], + "lines": [ + { + "bbox": [ + 344, + 247, + 410, + 289 + ], + "spans": [ + { + "bbox": [ + 344, + 247, + 410, + 289 + ], + "type": "image", + "image_path": "70a1048286a32200221a0411d837f2edab99e2038afea1d54ccf1601a7916c4d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 290, + 407, + 312 + ], + "lines": [ + { + "bbox": [ + 346, + 290, + 407, + 312 + ], + "spans": [ + { + "bbox": [ + 346, + 290, + 407, + 312 + ], + "type": "text", + "content": "Q: Your task is to harvest beef using the iron sword. Which action should you do next? A: Find cow" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 413, + 247, + 479, + 300 + ], + "blocks": [ + { + "bbox": [ + 413, + 247, + 479, + 300 + ], + "lines": [ + { + "bbox": [ + 413, + 247, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 413, + 247, + 479, + 300 + ], + "type": "image", + "image_path": "c43a8182b60d8a4722ea56871e70ed174d8c8d4083b159c15d4cff869937cbea.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 301, + 468, + 312 + ], + "lines": [ + { + "bbox": [ + 413, + 301, + 468, + 312 + ], + "spans": [ + { + "bbox": [ + 413, + 301, + 468, + 312 + ], + "type": "text", + "content": "Q: What is the defect type? A: Wire combined" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 392, + 480, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 476 + ], + "type": "text", + "content": "Cross-application Style Existing LMMs are mainly assessed in general application scenarios where questions can be answered using common knowledge [17, 21]. For instance, identifying the object on a table in an image can be accomplished with general knowledge. However, there exist specific application scenarios that demand specialized domain knowledge to address. For example, when an autonomous driving control center processes the road image to decide the next action: going straight or turning left, specific driving knowledge is required." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 130, + 477, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 480, + 525 + ], + "type": "text", + "content": "As existing LMMs are primarily evaluated with questions and answers (Q&A) that rely on common knowledge, the extent to which an LMM can remain effective in situations requiring domain-specific knowledge has not been comprehensively explored. We also investigate this aspect in our proposed benchmark." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 546, + 287, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 287, + 558 + ], + "type": "text", + "content": "3.2 Benchmark Construction" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Artistic Style Data To construct a benchmark characterized by distributional shifts pertaining to artistic styles, we curated a diverse set of images encompassing Cartoon, Painting, Sketch, Handmade, and Tattoo styles from the COCO-O dataset [43]. We systematically sampled 100 images for each stylistic category, resulting in a total of 500 images extracted from the COCO-O dataset. This dataset comprises a rich array of object types, providing a comprehensive evaluation ground for assessing the image understanding capabilities of the existing LMMs. Accordingly, the questions are prepared as asking the models to reason" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "text", + "content": "BenchLMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "and answer questions about images, and the sample images with Q&A are shown in the first row of Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 482, + 395 + ], + "type": "text", + "content": "Sensor-style Data Other than RGB cameras, Infrared sensors, X-ray sensors, Magnetic Resonance Imaging (MRI) sensors, and Computed Tomography (CT) sensors are often used to collect image data for human beings to perceive and analyze. Our benchmark examines whether the LMMs can still have the comparative reasoning capability toward these images that are intrinsically different from RGB photo images. In the case of Infrared data, we collected a set of 200 images from the datasets that focus on street scenes analysis toward with pedestrians, buildings, and vehicles [22, 25, 26, 40]. Questions for LMMs mainly revolve around analyzing the contents of these street scenes. Specifically, we design some thermal-related questions such as determining the hottest content. An illustrative example is provided in the second row of Fig. 2. For Low-energy X-ray (L-Xray) data from the SiXray-D dataset [47], we gathered 200 images, including 160 with prohibited objects (knife, scissors, wrench, gun, hammer) and 40 without any prohibited objects. Questions are designed to assess the ability to detect security risks in X-ray images, such as identifying forbidden objects. A sample question is presented in the second row of Fig. 2. In the medical application, high-energy X-ray (H-Xray), MRI and CT scanning images (200 each) were extracted from the [37]. Questions for LMMs in this context involve tasks like recognizing organs or identifying diseases in the respective images. By comparing inference results between different sensor types, our benchmark aims to uncover challenges associated with cross-sensor analysis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 399, + 483, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 483, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 483, + 531 + ], + "type": "text", + "content": "Application-style Data In order to establish a comprehensive cross-application-style benchmark, we systematically integrated diverse task domains into our evaluation framework. We consider four representative applications: remote sensing, autonomous driving, agent action prediction, and defect detection. Specifically, for the remote sensing application, we utilized 200 images sourced from the DOTA dataset [59] and annotated them in the VQA format from a satellite perspective. For the autonomous driving task, we extracted 250 images from the BDD100K dataset [65], featuring various weather conditions. Our primary focus was on designing questions and answers aimed at predicting driving actions such as going straight, turning left, or turning right. An illustrative example is provided in the third row of Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "To introduce the agent action prediction application, we meticulously selected 100 images from the Domestic Robot dataset [11] and an additional 117 images from the Open Game dataset [11]. The associated questions center around predicting the next action to be taken under the current circumstances. Furthermore, for evaluating defect detection capabilities, we curated 100 images from the Defect Detection dataset [6], with questions focused on identifying defect types and determining the presence of defects. This diverse set of applications and datasets collectively forms a diverse benchmark, allowing for a comprehensive assessment of the proposed approach across various application domains. More examples and Q&A of different artistic styles, sensor styles, application styles, and the constrain of answers can be found in the Appendix." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 299, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 299, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 299, + 128 + ], + "type": "text", + "content": "3.3 Style Prompt Enhancement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 253 + ], + "type": "text", + "content": "It is generally expected that additional prior information to LMMs' inputs can enhance performance. Nevertheless, such contextual priors (e.g. knowledge of artistic styles) might be unavailable in automated inference scenarios. We are inspired by human cognitive behavior that humans often comprehend the style of an image before delving into object recognition, akin to appreciating a painting by Van Gogh, where recognition of the impressionist style precedes identification of objects within the artwork. Other contextual priors, like sensor or application styles, can also enhance LMMs' performance. For example, identifying people in infrared images is easier when LMMs know that the images are captured by infrared thermal cameras since humans typically show consistent temperatures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 361 + ], + "type": "text", + "content": "With this insight, we investigate the efficacy of prompting LMMs to predict the style of an image before addressing associated questions. To this end, we devised Style Prompt Enhancement by introducing a text prompt, instructing the model to \"Determine the style of the image before answering the question.\" SPE aims to leverage stylistic cues as a precursor to visual question answering, potentially enhancing the model's capability of visual reasoning. Despite its simplicity, SPE's versatility allows it to be applied directly to images of various styles without any additional prompt tuning or model training, clearly boosting the overall performance of LMMs in a zero-shot manner, as detailed in Section 4.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 376, + 224, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 224, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 224, + 391 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 399, + 262, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 262, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 262, + 412 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 417, + 498, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 498, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 498, + 524 + ], + "type": "text", + "content": "This paper evaluates a selection of existing open-source LMMs, which include LLaVA [38,39], MiniGPT4 [10,71], InstructBLIP [14], Otter [32], and OpenFlamingo-7B [4]. These models are initialized by pre-trained weights and evaluated on our BenchLMM. In addition, we use the variants of these models, where the language encoders and vision encoders are changed to investigate how different encoders have impacts on the cross-field capabilities. Moreover, we perform a quantitative evaluation of GPT-4V, the latest commercial LMM, with our dataset and compare its performance with the aforementioned open-source LMMs, and such a quantitative evaluation is the first time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "text", + "content": "In our performance evaluation, we adhere to previous works [38,39] methodology, employing the ChatGPT API to gauge the proximity of answers predicted by the LMMs to ground-truth answers. Here, we designate the output of the LMM as " + }, + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "text", + "content": ", and the ground-truth label as " + }, + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "text", + "content": ". The evaluation utilizes the prompt \"Compare the similarity between " + }, + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 525, + 482, + 609 + ], + "type": "text", + "content": " with a correctness score ranging from 0.0 (totally wrong) to 1.0 (totally right). The middle score provides the percentage of correctness.\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 304, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 304, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 304, + 635 + ], + "type": "text", + "content": "4.2 Evaluations of public LMMs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Cross-artistic-style Capability In most existing works [17,21,33,34,41,67], LMMs are predominantly evaluated using images in the 'Photo' style, leading" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "text", + "content": "BenchLMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 168, + 480, + 299 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Experimental results (\\%) of public LMMs on BenchLMM's cross-artistic-style evaluation. Note that Average* represents the average accuracy computed over five artistic-style benchmarks. The best and the second best results are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 168, + 480, + 299 + ], + "lines": [ + { + "bbox": [ + 133, + 168, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 480, + 299 + ], + "type": "table", + "html": "
LMMsVisual EncoderLanguage ModelCommonCartoonHandmakePaintingSketchTattooAverage*
GPT-4V [62]--81.563.358.758.264.768.362.6
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B74.662.056.657.663.357.059.3
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B72.759.057.159.357.561.458.9
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B56.649.945.448.053.747.048.8
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B56.957.828.437.842.632.340.0
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B71.251.953.343.462.054.653.0
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B73.958.152.055.855.255.555.3
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B66.428.737.840.941.728.635.5
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B45.635.037.038.536.437.736.9
Otter-7B [32]CLIP-ViT-L/14MPT-7B64.845.149.039.844.142.444.1
", + "image_path": "7d1163ce2f891bd85640af0d0a4a2fcc0622195f15f14b9fc35c4912f27e3434.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 326, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 422 + ], + "type": "text", + "content": "to a gap in understanding their performance across diverse artistic styles. We extend the evaluation scope by examining LMMs' performance with various artistic styles beyond the common 'Photo' style. The results detailed in Table 1 reveal a notable decline in LMMs' effectiveness when processing these artistic styles. This trend suggests a potential overfitting of LMMs to the 'Photo' style, highlighting their limited adaptability to varied artistic styles, a capability that humans typically possess. Interestingly, GPT-4V, as an outstanding commercial model, exhibits similar limitations in handling diverse styles." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": "Besides, LLaVA-1.5-13B [38] is improved from LLaVA-13B [39] by introducing more training tricks and data, and LLaVA-1.5-13B outperforms LLaVA-13B by a clear margin on the 'Common' and other artistic styles. This comparison reveals the importance of data and training strategy as LLaVA-1.5-13B and LLaVA-13B use the same visual encoder and language encoder. However, we also find that a better performance on 'Common' does not necessarily indicate a better performance on other styles. For example, compared with InstructBLIP-7B, InstructBLIP-13B demonstrates better performance across five artistic-style benchmarks in terms of average accuracy, though it shows comparatively lower performance in the 'Common' style. Therefore, such results convey that it is important to conduct comprehensive evaluations to fully validate an LMM's capability rather than only focus on the 'Common' style, which shows the importance of our proposed benchmark." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Cross-sensor Capability We benchmark the LMM models' cross-sensor capability in Table 2. Because GPT-4V denied to provide medical diagnostics or identify potential abnormalities in medical images like H-Xray/MRI/CT, GPT-4V is only evaluated on the Infrared and L-Xray sensor data and achieves the results of " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "55.0\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "50.0\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " respectively, which are far lower than the 'Common' style " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "(81.5\\%)" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". Meanwhile, other LMM models' performance on images captured by X-ray, MRI, CT, and Infrared dropped significantly compared to" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 147, + 479, + 283 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Experimental results (\\%) of public LMMs on BenchLMM's cross-sensor evaluation. Note that GPT-4V denied to test medical images (H-Xray/MRI/CT)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 147, + 479, + 283 + ], + "lines": [ + { + "bbox": [ + 134, + 147, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 134, + 147, + 479, + 283 + ], + "type": "table", + "html": "
LMMsVisual EncoderLanguage ModelInfraredL-XrayH-XrayMRICTAverage
GPT-4V [62]--55.050.0----
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B52.151.555.933.743.147.3
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B40.824.656.620.943.437.3
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B53.746.039.128.439.741.4
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B32.947.224.625.323.230.6
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B46.553.249.429.646.845.3
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B42.235.459.519.438.539.0
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B30.143.031.517.417.828.2
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B35.144.123.716.815.327.0
Otter-7B [32]CLIP-ViT-L/14MPT-7B30.246.439.147.846.342.0
", + "image_path": "e38eb3c5a4ca2099a97fdc2721a964d08718652238077114e31beb44da21541b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 316, + 479, + 436 + ], + "blocks": [ + { + "bbox": [ + 138, + 295, + 473, + 307 + ], + "lines": [ + { + "bbox": [ + 138, + 295, + 473, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 473, + 307 + ], + "type": "text", + "content": "Table 3: Results of public LMMs on BenchLMM's cross-application experiments." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 316, + 479, + 436 + ], + "lines": [ + { + "bbox": [ + 134, + 316, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 134, + 316, + 479, + 436 + ], + "type": "table", + "html": "
LMMsVisual EncoderLanguage ModelRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
GPT-4V [62]--69.735.856.074.064.460.0
LLaVA-1.5-13B [38]CLIP-ViT-L/14Vicuna-13B65.628.649.021.460.345.2
InstructBLIP-13B [14]EVA-ViT-GVicuna-13B63.632.632.019.547.038.9
LLaVA-13B [39]CLIP-ViT-L/14Vicuna-13B38.731.759.043.828.240.3
MiniGPT4-13B [71]CLIP-ViT-L/14Vicuna-13B33.322.648.034.231.734.2
LLaVA-1.5-7B [39]CLIP-ViT-L/14Vicuna-7B61.731.344.015.459.642.2
InstructBLIP-7B [14]EVA-ViT-LVicuna-7B63.536.742.024.830.839.6
MiniGPT4-v2-7B [10]EVA-ViT-GLLaMA2-7B34.517.640.022.218.826.6
MiniGPT4-7B [71]CLIP-ViT-L/14Vicuna-7B27.920.841.040.113.628.7
Otter-7B [32]CLIP-ViT-L/14MPT-7B44.829.736.018.827.331.3
", + "image_path": "f7f000a2f997f7438533b1f5bb4116b97815426f785f7b809a8493fb06e1fa81.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 480, + 568 + ], + "type": "text", + "content": "the performance on 'Common' captured by the RGB camera sensor. Therefore, the intrinsic difference in the imaging process between RGB cameras and other sensors is also a crucial challenge to LMMs. With respect to the MRI, H-Xray, and CT images, the questions may require specific medical knowledge to answer, and it may be argued that the performance drop is due to the knowledge gap. It is also observed from the L-Xray, where the objects (i.e., knife, gun, hammer, etc.) can be recognized with only common knowledge, but the LMMs' performance on the L-Xray data is significantly lower than 'Common', which further validates the challenge brought by the sensor difference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Cross-application Capability In Table 3, we evaluate the different LMMs on various applications, finding that GPT-4V generally outperforms others, except in autonomous driving where its accuracy drops to " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "35.8\\%" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": ". This decrease is likely due to the challenges in the Autonomous Driving dataset, such as recognizing low-resolution traffic signs and requiring specific driving knowledge, where both GPT-4 and open-source LMMs have poor performance. We also observe that LLaVA-1.5-13B achieves comparable performance as GPT-4V on the Remote Sensing and Defect Detection datasets, which seems to have comparable cross" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "text", + "content": "BenchLMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 117, + 253, + 277 + ], + "blocks": [ + { + "bbox": [ + 153, + 117, + 253, + 277 + ], + "lines": [ + { + "bbox": [ + 153, + 117, + 253, + 277 + ], + "spans": [ + { + "bbox": [ + 153, + 117, + 253, + 277 + ], + "type": "image", + "image_path": "0380a6083d16477095a416ccc400a78abf498111c561d5e00f8b6ab19fe86115.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 154, + 278, + 253, + 437 + ], + "blocks": [ + { + "bbox": [ + 154, + 278, + 253, + 437 + ], + "lines": [ + { + "bbox": [ + 154, + 278, + 253, + 437 + ], + "spans": [ + { + "bbox": [ + 154, + 278, + 253, + 437 + ], + "type": "image", + "image_path": "33142e0e9e08c940bc339e886bfcaeb3340de61d4137574d2f0538bb87d58cc8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 446, + 480, + 469 + ], + "lines": [ + { + "bbox": [ + 130, + 446, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 480, + 469 + ], + "type": "text", + "content": "Fig. 3: Error-reflection capability comparisons: Unlike LLaVA-1.5-13B, GPT-4V possesses the ability to analyze its errors when provided with the correct answers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 257, + 117, + 358, + 277 + ], + "blocks": [ + { + "bbox": [ + 257, + 117, + 358, + 277 + ], + "lines": [ + { + "bbox": [ + 257, + 117, + 358, + 277 + ], + "spans": [ + { + "bbox": [ + 257, + 117, + 358, + 277 + ], + "type": "image", + "image_path": "6c1d652eb59e200fbdb12562d24478fcce6e20ee556b1a42bc31d72aec8d565d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 257, + 278, + 358, + 437 + ], + "blocks": [ + { + "bbox": [ + 257, + 278, + 358, + 437 + ], + "lines": [ + { + "bbox": [ + 257, + 278, + 358, + 437 + ], + "spans": [ + { + "bbox": [ + 257, + 278, + 358, + 437 + ], + "type": "image", + "image_path": "99c115e4e8592c9be0f536820e8b711492a2e77aae19d67dd415162189b0dd35.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 361, + 117, + 462, + 277 + ], + "blocks": [ + { + "bbox": [ + 361, + 117, + 462, + 277 + ], + "lines": [ + { + "bbox": [ + 361, + 117, + 462, + 277 + ], + "spans": [ + { + "bbox": [ + 361, + 117, + 462, + 277 + ], + "type": "image", + "image_path": "8014eb06bb66a4f626987b8bb87350790aee9126fa7e82f4f6af747800e7a82c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 361, + 278, + 462, + 437 + ], + "blocks": [ + { + "bbox": [ + 361, + 278, + 462, + 437 + ], + "lines": [ + { + "bbox": [ + 361, + 278, + 462, + 437 + ], + "spans": [ + { + "bbox": [ + 361, + 278, + 462, + 437 + ], + "type": "image", + "image_path": "4bb714b1b0881729cff988e04b0136ebc29c150e77483af4875789a93dcd5f8a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 501, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 550 + ], + "type": "text", + "content": "application capability. However, we surprisingly find that LLaVA-1.5-13B merely achieves " + }, + { + "bbox": [ + 130, + 501, + 482, + 550 + ], + "type": "inline_equation", + "content": "21.4\\%" + }, + { + "bbox": [ + 130, + 501, + 482, + 550 + ], + "type": "text", + "content": " but GPT-4V achieves up to " + }, + { + "bbox": [ + 130, + 501, + 482, + 550 + ], + "type": "inline_equation", + "content": "74.0\\%" + }, + { + "bbox": [ + 130, + 501, + 482, + 550 + ], + "type": "text", + "content": ". As GPT-4V is close-source, the technical details are not disclosed and we conjecture that the training data for GPT-4V may involve data similar to the Open Game dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 553, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 613 + ], + "type": "text", + "content": "We conclude from the experiments that: 1) Existing LMMs often degrade in performance with cross-style images; 2) Better performance on 'Common' data doesn't ensure better results on other styles; 3) Comprehensive evaluations across different dimensions are crucial, as errors in tasks like medical or security can lead to significant losses" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Error-reflection Capability To understand the LMMs further, we let the LMM do an error reflection that uses itself to parse the error information. When a model gives wrong answers, we warp up the conversation: the previous question input and predicted answer output. Specifically, we add \"The correct answer is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 136, + 462, + 342 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "type": "text", + "content": "Table 4: Experimental results (%) of our proposed SPE on the BenchLMM benchmark." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 136, + 462, + 342 + ], + "lines": [ + { + "bbox": [ + 151, + 136, + 462, + 342 + ], + "spans": [ + { + "bbox": [ + 151, + 136, + 462, + 342 + ], + "type": "table", + "html": "
LMMsCartoonHandmakePaintingSketchTattooAverage
LLaVA-1.5-13B [38]62.056.657.663.357.059.3
+ SPE (Ours)67.657.767.968.863.565.1 (+5.8)
InstructBLIP-13B [14]59.057.159.357.561.458.9
+ SPE (Ours)61.162.364.860.963.362.5 (+3.6)
LMMsInfraredL-XrayH-XrayMRICTAverage
LLaVA-1.5-13B [38]52.151.555.933.743.147.3
+ SPE (Ours)58.762.856.233.647.751.8 (+4.5)
InstructBLIP-13B [14]40.824.656.620.943.437.3
+ SPE (Ours)43.524.857.725.042.638.7 (+1.4)
LMMsRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
LLaVA-1.5-13B [38]65.628.649.021.460.345.2
+ SPE (Ours)68.029.451.022.761.746.6 (+1.4)
InstructBLIP-13B [14]63.632.632.019.547.038.9
+ SPE (Ours)63.538.831.019.049.440.3 (+1.4)
", + "image_path": "335f61c177c312e9339a388afc68644e84cd032d9afe64d52b928fea9b40aee6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 364, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 480, + 388 + ], + "type": "text", + "content": "[GT Answer]. Explain why your answer is incorrect. Use a single sentence.\" to the conversation (GT means Ground-truth), as the new input of LMMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 591 + ], + "type": "text", + "content": "Although both GPT-4V and LLaVA-1.5-13B produce incorrect answers, GPT-4V demonstrates the ability to reason the correct answer through a detailed inference process. As illustrated in Fig. 3, GPT-4V can offer insightful analysis beyond the ground-truth answer, including the positions of clock hands for estimating time, the location of prohibited objects, and the traffic signals for driving planning. Moreover, from the second row of Fig. 3, we can see that the LLaVA model can just simply repeat the answer or say \"I was wrong\". By contrast, GPT-4V can do inference itself to provide the correct answer, such as recognizing the toy animals (Handmake) or the scratch on the capsule (Defect Detection). Through this study, we can see that the error-reflection capability of LLaVA-1.5-13B is inferior to that of GPT-4V. Therefore, we reveal that another important capability of a large multimodal model is whether the model can reflect how to derive the correct answer. Studying such capability is useful to common users. For example, when users use LMMs for self-learning, the user can provide reference answers and LMMs can provide a more detailed inference process, which can improve the learning effect. Moreover, the answer of GPT-4V can also be useful materials used to train open-source LMMs in the future." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 609, + 380, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 380, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 380, + 621 + ], + "type": "text", + "content": "4.3 Experiments of Style Prompt Enhancement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Comparison to Baseline Models In this section, we aim to study if a large multimodal model can be enhanced by prompting the model to estimate the style first and then answer the question, based on which we propose the Style Prompt" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "text", + "content": "BenchLMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 146, + 468, + 300 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 5: Complementary studies of our SPE with recent zero-shot chain-of-thought method ZS-COT [31] using LLaVA-1.5-13B [38] on the BenchLMM" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 146, + 468, + 300 + ], + "lines": [ + { + "bbox": [ + 145, + 146, + 468, + 300 + ], + "spans": [ + { + "bbox": [ + 145, + 146, + 468, + 300 + ], + "type": "table", + "html": "
MethodCartoonHandmakePaintingSketchTattooAverage
LLaVA-1.5-13B [38]62.056.657.663.357.059.3
+ZS-COT [31]64.359.770.662.871.065.7(+6.4)
+ZS-COT [31]+SPE (Ours)66.064.466.565.468.366.1(+6.8)
MethodInfraredL-XrayH-XrayMRICTAverage
LLaVA-1.5-13B [38]52.151.555.933.743.147.3
+ZS-COT [31]49.355.957.635.644.748.6(+1.3)
+ZS-COT [31]+SPE (Ours)54.470.456.735.345.452.4(+5.1)
MethodRemote SensingAutonomous DrivingDomestic RobotOpen GameDefect DetectionAverage
LLaVA-1.5-13B [38]65.628.649.021.460.345.2
+ZS-COT [31]65.527.043.034.263.746.7(+1.5)
+ZS-COT [31]+SPE (Ours)76.024.049.052.160.152.2(+7.0)
", + "image_path": "017ea27847e70f4b877715c1df774394501c381fd4163d5fd2b95f86f64aaf2c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 380 + ], + "type": "text", + "content": "Enhancement (SPE) method and conduct experiments with our BenchLMM. Specifically, we examined our SPE with LLaVA-1.5-13B [38] and InstructBLIP-13B [14] because these two models are representative of better average performance than MiniGPT4 [71] and Otter [32]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 553 + ], + "type": "text", + "content": "The experimental results are presented in Table 4. As shown in the table, our proposed SPE can benefit the LMMs by achieving general improvement over different styles, which means that reasoning the style first is an effective prompting method to improve LMMs' visual reasoning capability. Besides, in the experiments of crossing to different applications (Remote Sensing, Autonomous Driving, etc.), we observe that its benefit is smaller than others. For example, LLaVA-1.5-13B gains " + }, + { + "bbox": [ + 130, + 385, + 482, + 553 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 130, + 385, + 482, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 385, + 482, + 553 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 130, + 385, + 482, + 553 + ], + "type": "text", + "content": " improvement in the experiments of cross-artistic and cross-sensor styles, while the improvement is merely " + }, + { + "bbox": [ + 130, + 385, + 482, + 553 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 385, + 482, + 553 + ], + "type": "text", + "content": " when crossing applications. We conjecture that specific application knowledge could inferred from only the style, which limits the benefits of SPE. Moreover, we observe that LLaVA-1.5-13 generally obtains more benefits than InstructBLIP-13B, which means that the stronger the LMM, the more benefits from SPE, which is because a stronger LMM has stronger inference capability and can better utilize prior information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Comparison to Other Prompting Methods Our proposed SPE method enhances LMMs' performance by prompting LMMs to estimate the input image's style first, a technique akin to chain-of-thought prompting approaches [31, 45]. In Table 5, we compare our proposed SPE with the state-of-the-art prompting Chain-of-Thought (COT) [31], which is a zero-shot, to demonstrate the efficacy of SPE. As shown, the COT method can improve LMM's performance, and our SPE is compatible with the COT methods and further enhances the model's capability, which is demonstrated by the general improvement brought by “+ZS-COT+SPE” as shown in Table 5. We also notice that in some cases, using SPE" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 114, + 455, + 194 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 455, + 194 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 455, + 194 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 455, + 194 + ], + "type": "image", + "image_path": "e340eaf9b742dc6c007d5b6aba0d1180d31010fd95faf817f5e5c20210003519.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 204, + 480, + 226 + ], + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 226 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons between LLaVA-1.5-13B and LLaVA-1.5-13B + SPE (Ours) on style-transfer images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 147, + 258, + 466, + 327 + ], + "blocks": [ + { + "bbox": [ + 148, + 238, + 463, + 249 + ], + "lines": [ + { + "bbox": [ + 148, + 238, + 463, + 249 + ], + "spans": [ + { + "bbox": [ + 148, + 238, + 463, + 249 + ], + "type": "text", + "content": "Table 6: Quantitative Results of our proposed SPE on style-transfer images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 258, + 466, + 327 + ], + "lines": [ + { + "bbox": [ + 147, + 258, + 466, + 327 + ], + "spans": [ + { + "bbox": [ + 147, + 258, + 466, + 327 + ], + "type": "table", + "html": "
LMMs|Cartoon Handmake Painting Sketch Tattoo | Average
LLaVA-1.5-13B [38] + SPE (Ours)62.161.152.468.841.257.1
67.462.159.469.445.560.8 (+3.7)
InstructBLIP-13B [14] + SPE (Ours)69.370.861.353.865.064.0
73.071.664.262.165.367.2 (+3.2)
", + "image_path": "7680074573adb13986bd03cd08807ada04811168e525ea129afe69db818081fa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "type": "text", + "content": "with COT could lead to performance decline. The investigation is beyond the scope of this paper, which will be explored in our future work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 373, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 480, + 479 + ], + "type": "text", + "content": "Results on the Style-transfer Images We employ the latest commercial style transfer model, DALL-E3 [50], to transfer the 'Common' style images into different artistic styles. As depicted in Fig. 4, these images share similar structures and layouts. However, they are just in different artistic styles, and the tested LLaVA model also makes mistakes. The significant performance degradation on the transferred images is also proved by the quantitative results presented in Table 6, which validates that LMMs suffer from difficulties in recognizing images of different styles. Nevertheless, our SPE significantly enhances the cross-style capability of LMMs in this challenging task, as shown in Fig. 4 and Table 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 480, + 480, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 480, + 600 + ], + "type": "text", + "content": "In-depth Analysis of SPE Besides quantitative assessments, we explore the inner workings of SPE to understand how it improves LMMs. As illustrated in Fig. 5, LMMs are adept at accurately estimating the style of images. For instance, in the figure's first row and second column, the depicted style is 'Handmake,' and the model accurately recognizes this style as akin to a toy, indicating a correct match. Moreover, when estimating the style, the model can also understand the context first and thus better answer the final question, as shown in the examples of Remote Sensing and Defect Detection. Our findings suggest that prompting LMMs to first identify the style of images serves as an effective, training-free method to improve the models' proficiency in handling cross-style images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 617, + 219, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 219, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 219, + 629 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "In this work, we introduce BenchLMM, a novel benchmark for quantitatively evaluating large multimodal models (LMMs) across varied visual distribution" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "text", + "content": "BenchLMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 116, + 462, + 492 + ], + "blocks": [ + { + "bbox": [ + 149, + 116, + 462, + 492 + ], + "lines": [ + { + "bbox": [ + 149, + 116, + 462, + 492 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 462, + 492 + ], + "type": "image", + "image_path": "1135ef199bff7c35cebd55bcacb9ad3c2f3ea7ba6ad9e4f23730986f4e29f609.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 502, + 444, + 514 + ], + "lines": [ + { + "bbox": [ + 168, + 502, + 444, + 514 + ], + "spans": [ + { + "bbox": [ + 168, + 502, + 444, + 514 + ], + "type": "text", + "content": "Fig. 5: Evaluation of LLaVA-1.5-13B with SPE on the BenchLMM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 545, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 654 + ], + "type": "text", + "content": "shifts, including artistic, sensor, and application styles. Our evaluations of numerous existing LMMs highlight two key findings: 1) LMMs typically underperform with non-common image styles. 2) Our style prompt enhancement approach, inspired by human perception of art, significantly improves LMMs' visual reasoning without extra fine-tuning. Additionally, our error-reflection study shows that stronger LMMs can self-diagnose errors by providing insights not given by humans, whereas less capable models merely restate correct answers without new insights. Our benchmark offers a comprehensive tool for LMM research, emphasizing the importance of error-reflection capabilities in the future." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "content": "This work was carried out at Rapid-Rich Object Search (ROSE) Lab, School of Electrical & Electronic Engineering, Nanyang Technological University. This research is supported by the NTU-PKU Joint Research Institute (a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 217, + 197, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 217, + 197, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 217, + 197, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 239, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 239, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 283 + ], + "type": "text", + "content": "1. Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8948-8957 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 327 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "type": "text", + "content": "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 361, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 414 + ], + "type": "text", + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "text", + "content": "5. Ben-Younes, H., Cadene, R., Cord, M., Thome, N.: Multimodal tucker fusion for visual question answering. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2612-2620 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 449, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 490 + ], + "type": "text", + "content": "6. Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., Steger, C.: The mvtec anomaly detection dataset: a comprehensive real-world dataset for unsupervised anomaly detection. International Journal of Computer Vision 129(4), 1038-1059 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "text", + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 525, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 555 + ], + "type": "text", + "content": "8. Cadene, R., Ben-Younes, H., Cord, M., Thome, N.: Murel: Multimodal relational reasoning for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1989-1998 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 557, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 588 + ], + "type": "text", + "content": "9. Cao, Q., Wan, W., Wang, K., Liang, X., Lin, L.: Linguistically routing capsule network for out-of-distribution visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1614-1623 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 590, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 620 + ], + "type": "text", + "content": "10. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "text", + "content": "11. Chen, L., Zhang, Y., Ren, S., Zhao, H., Cai, Z., Wang, Y., Wang, P., Liu, T., Chang, B.: Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4-vision and beyond. arXiv preprint arXiv:2310.02071 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "text", + "content": "BenchLMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "12. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollár, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 184 + ], + "type": "text", + "content": "13. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "text", + "content": "14. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 262 + ], + "type": "text", + "content": "15. Dancette, C., Cadene, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1574-1583 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 306 + ], + "type": "text", + "content": "16. Ding, Y., Yu, J., Liu, B., Hu, Y., Cui, M., Wu, Q.: Mukea: Multimodal knowledge extraction and accumulation for knowledge-based visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5089-5098 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 307, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 480, + 340 + ], + "type": "text", + "content": "17. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models. arXiv preprint arXiv:2306.13394 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "text", + "content": "18. Fukui, A., Park, D.H., Yang, D., Rohrbach, A., Darrell, T., Rohrbach, M.: Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 375, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 418 + ], + "type": "text", + "content": "19. Gao, F., Ping, Q., Thattai, G., Reganti, A., Wu, Y.N., Natarajan, P.: Transform-retrieve-produce: Natural language-centric outside-knowledge visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5067-5077 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "text", + "content": "20. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhan, S.: Chain of thought prompt tuning in vision language models. arXiv preprint arXiv:2304.07919 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 442, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 485 + ], + "type": "text", + "content": "21. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6904-6913 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 487, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 530 + ], + "type": "text", + "content": "22. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 531, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 480, + 564 + ], + "type": "text", + "content": "23. Han, X., Wang, S., Su, C., Huang, Q., Tian, Q.: Greedy gradient ensemble for robust visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1584-1593 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "text", + "content": "24. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "25. Hwang, S., Park, J., Kim, N., Choi, Y., So Kweon, I.: Multispectral pedestrian detection: Benchmark dataset and baseline. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1037-1045 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "26. Jia, X., Zhu, C., Li, M., Tang, W., Zhou, W.: LLVIP: A visible-infrared paired dataset for low-light vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3496-3504 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "27. Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2901-2910 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 205 + ], + "type": "text", + "content": "28. Johnson, J., Hariharan, B., Van Der Maaten, L., Hoffman, J., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Inferring and executing programs for visual reasoning. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2989-2998 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "text", + "content": "29. Kim, J.H., Jun, J., Zhang, B.T.: Bilinear attention networks. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "text", + "content": "30. Kim, J.H., On, K.W., Lim, W., Kim, J., Ha, J.W., Zhang, B.T.: Hadamard product for low-rank bilinear pooling. arXiv preprint arXiv:1610.04325 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "type": "text", + "content": "31. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. Advances in neural information processing systems 35, 22199-22213 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "text", + "content": "32. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "type": "text", + "content": "33. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 363 + ], + "type": "text", + "content": "34. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "text", + "content": "35. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 387, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 481, + 418 + ], + "type": "text", + "content": "36. Liang, P.P., Lyu, Y., Fan, X., Wu, Z., Cheng, Y., Wu, J., Chen, L., Wu, P., Lee, M.A., Zhu, Y., et al.: Multibench: Multiscale benchmarks for multimodal representation learning. arXiv preprint arXiv:2107.07502 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 420, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 463 + ], + "type": "text", + "content": "37. Liu, B., Zhan, L.M., Xu, L., Ma, L., Yang, Y., Wu, X.M.: Slake: A semantically-labeled knowledge-enhanced dataset for medical visual question answering. In: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). pp. 1650-1654. IEEE (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "type": "text", + "content": "38. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 487, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 509 + ], + "type": "text", + "content": "39. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 510, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 564 + ], + "type": "text", + "content": "40. Liu, J., Fan, X., Huang, Z., Wu, G., Liu, R., Zhong, W., Luo, Z.: Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5802-5811 (June 2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 566, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 481, + 597 + ], + "type": "text", + "content": "41. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: MMBench: Is Your Multi-modal Model an All-around Player? arXiv preprint arXiv:2307.06281 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 620 + ], + "type": "text", + "content": "42. Manyika, J.: An overview of bard: an early experiment with generative ai. AI. Google Static Documents (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "43. Mao, X., Chen, Y., Zhu, Y., Chen, D., Su, H., Zhang, R., Xue, H.: COCO-O: A Benchmark for Object Detectors under Natural Distribution Shifts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6339-6350 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "text", + "content": "BenchLMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "text", + "content": "44. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3195-3204 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "text", + "content": "45. Mitra, C., Huang, B., Darrell, T., Herzig, R.: Compositional chain-of-thought prompting for large multimodal models. arXiv preprint arXiv:2311.17076 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "46. Nam, H., Ha, J.W., Kim, J.: Dual attention networks for multimodal reasoning and matching. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 299-307 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "47. Nguyen, H.D., Cai, R., Zhao, H., Kot, A.C., Wen, B.: Towards more efficient security inspection via deep learning: a task-driven x-ray image cropping scheme. Micromachines 13(4), 565 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 462, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 462, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 462, + 259 + ], + "type": "text", + "content": "48. OpenAI, R.: Gpt-4 technical report. arxiv 2303.08774. View in Article (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 260, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 260, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 260, + 481, + 293 + ], + "type": "text", + "content": "49. Shah, M., Chen, X., Rohrbach, M., Parikh, D.: Cycle-consistency for robust visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6649-6658 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "50. Shi, Z., Zhou, X., Qiu, X., Zhu, X.: Improving image captioning with better use of captions. arXiv preprint arXiv:2006.11807 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "text", + "content": "51. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 742-758. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "52. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards VQA models that can read. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8317-8326 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "53. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "text", + "content": "54. Tascon-Morales, S., Márquez-Neila, P., Sznitman, R.: Logical implications for visual question answering consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6725-6735 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "55. Teney, D., Abbasnejad, E., van den Hengel, A.: Unshuffling data for improved generalization in visual question answering. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1417-1427 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "56. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "57. Tsimpoukelli, M., Menick, J.L., Cabi, S., Eslami, S., Vinyals, O., Hill, F.: Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems 34, 200-212 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "58. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "59. Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: Dota: A large-scale dataset for object detection in aerial images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3974-3983 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "60. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 479 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "61. Yang, X., Gao, C., Zhang, H., Cai, J.: Auto-parsing network for image captioning and visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2197-2207 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "62. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of LMMs: Preliminary explorations with GPT-4V (ision). arXiv preprint arXiv:2309.17421 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "63. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "text", + "content": "64. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "type": "text", + "content": "65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2633-2642 (2020). https://doi.org/10.1109/CVPR42600.2020.00271" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "66. Yu, S., Wu, P., Liang, P.P., Salakhutdinov, R., Morency, L.P.: Pacs: A dataset for physical audiovisual commonsense reasoning. In: European Conference on Computer Vision. pp. 292-309. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "type": "text", + "content": "67. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 392 + ], + "type": "text", + "content": "68. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6720-6731 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 391, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 413 + ], + "type": "text", + "content": "69. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "type": "text", + "content": "70. Zhou, Y., Ren, T., Zhu, C., Sun, X., Liu, J., Ding, X., Xu, M., Ji, R.: Trar: Routing the attention spans in transformer for visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2074-2084 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 456, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 479 + ], + "type": "text", + "content": "71. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models (2023)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 398, + 91, + 447, + 100 + ], + "type": "text", + "content": "BenchLMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/64864c4f-d7d3-4bce-8017-e14998347c20_content_list.json b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/64864c4f-d7d3-4bce-8017-e14998347c20_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d1a188f45e612160dc2bb4aff5ddadc92fb321f7 --- /dev/null +++ b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/64864c4f-d7d3-4bce-8017-e14998347c20_content_list.json @@ -0,0 +1,2000 @@ +[ + { + "type": "text", + "text": "Benchmarking Object Detectors with COCO: A New Path Forward", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shweta Singh1* Aayan Yadav1* Jitesh Jain2", + "bbox": [ + 323, + 210, + 676, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humphrey Shi² Justin Johnson³ Karan Desai³", + "bbox": [ + 316, + 227, + 684, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ IIT Roorkee $^{2}$ Georgia Tech $^{3}$ University of Michigan", + "bbox": [ + 300, + 241, + 700, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The Common Objects in Context (COCO) dataset has been instrumental in benchmarking object detectors over the past decade. Like every dataset, COCO contains subtle errors and imperfections stemming from its annotation procedure. With the advent of high-performing models, we ask whether these errors of COCO are hindering its utility in reliably benchmarking further progress. In search for an answer, we inspect thousands of masks from COCO (2017 version) and uncover different types of errors such as imprecise mask boundaries, non-exhaustively annotated instances, and mislabeled masks. Due to the prevalence of COCO, we choose to correct these errors to maintain continuity with prior research. We develop COCO-ReM (Refined Masks), a cleaner set of annotations with visibly better mask quality than COCO-2017. We evaluate fifty object detectors and find that models that predict visually sharper masks score higher on COCO-ReM, affirming that they were being incorrectly penalized due to errors in COCO-2017. Moreover, our models trained using COCO-ReM converge faster and score higher than their larger variants trained using COCO-2017, highlighting the importance of data quality in improving object detectors. With these findings, we advocate using COCO-ReM for future object detection research. Our dataset is available at https://cocorem.xyz", + "bbox": [ + 259, + 287, + 743, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: object detection, segmentation, datasets and benchmarks", + "bbox": [ + 261, + 578, + 732, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 636, + 375, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rigorous benchmarking of computer vision models relies on task-specific datasets annotated by humans [9, 10, 22, 25]. Understandably, human-annotated data is prone to errors and imperfections due to factors such as ambiguous instructions to annotators, differences in perspective of multiple annotators, etc. Errors in ground-truth training data provide a noisy learning signal for models, resulting in sub-par capabilities. Even more problematic is the presence of errors in ground-truth evaluation data, which can cause a misalignment between evaluation metrics and human judgment when comparing model performances. In this work, we comprehensively study the extent of this issue for the canonical computer vision task of object detection and segmentation.", + "bbox": [ + 212, + 666, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Denotes equal contribution. Correspondence to: Karan Desai (kdexd@umich.edu)", + "bbox": [ + 217, + 824, + 766, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/574a8021e3ae671cea135c567d69b3a6cba6a871bc830aa174b6e3b4bb0b0354.jpg", + "image_caption": [ + "COCO-2017" + ], + "image_footnote": [], + "bbox": [ + 217, + 156, + 313, + 262 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/71d3b68b4be4dc387a967ec666f23ce8385c698499951c0195fb9dc493d0e876.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 157, + 413, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a3a57ce1ee21d5fada8d68af59c2b8623efe7943f968c74ac516902a666f9c14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 157, + 495, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ca0c32cfc84361a6ab439be6b5dd0c959d89393dec0db9ba2b1e297fac71795.jpg", + "image_caption": [ + "COCO-ReM (ours)" + ], + "image_footnote": [], + "bbox": [ + 503, + 157, + 598, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2aad59070ba17f7fd23e798c7f2bb474602f38c834f7f1d4dd2ac0c445007cb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 157, + 696, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7061296a8b7bb6524378490b51226fae9d518b468d47f46d5bb5236ff8ddf1c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 157, + 779, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3e9526155ddb54d2a6c92a17aebf7c245012d8582485b0f34201d9827a3e4c5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 266, + 320, + 344 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c878476568b68c6f5b4e2ebf2705e58c56fa4ba9c52e529e960ed16d305ff941.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 266, + 450, + 344 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/692a973549fc58d9a355daa99d7f32cc1677912ec2eb55b39d93428b89e92d2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 267, + 553, + 344 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b201818a6f93b8b07e35907fe2c875e13fb1035ce4d442bdbb20316d45bb4cc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 267, + 642, + 344 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae1207e334a76ec80287f2438435d8ec7be7e0f3aa08f91c9d675225e7123be1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 267, + 779, + 422 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/26062413e2c70ce9e4237df4b68488f9df7c2a8652e905a920b9f22fbd4ef011.jpg", + "image_caption": [ + "Fig. 1: COCO with Refined Masks (COCO-ReM). Top: COCO-ReM improves upon the quality of COCO-2017 masks by handling occlusions consistently (dining table) and annotating instances exhaustively (banana). Bottom: Random examples from COCO-ReM showing the mask quality (single category per image, for clarity)." + ], + "image_footnote": [], + "bbox": [ + 218, + 344, + 352, + 422 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/52fff3971e82fde5b1533a07b70bb4d8c15a6d1b807f2a89a8dea9e602861892.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 344, + 503, + 422 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6f303e59aaef9a8013714213d3162bf6b38387ffe95955f912d278192195feba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 344, + 642, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Common Objects in Context (COCO) dataset [25] has been the standard benchmark for object detection and segmentation since its inception in 2014. COCO fostered the empirical study of modern object detectors by providing a large set of 123K images annotated with 896K instance masks. The Mask AP (average precision) of COCO detectors has improved dramatically, with $\\approx 70\\%$ relative gain between the pioneering Mask R-CNN [16, 17] from 2017, and recent models published in 2022-23 such as ViTDet [3, 23] and other Transformer-based models [4, 11, 33] like Mask2Former [8] and OneFormer [19].", + "bbox": [ + 212, + 503, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its popularity, COCO is not without limitations. COCO masks are known to have coarse boundaries [1, 14]. We manually inspect thousands of masks in COCO (2017 version) and observe other imperfections. As shown in Figure 1 (top), COCO-2017 masks do not handle holes and occlusions properly (dining table), and sometimes have non-exhaustively annotated instances (banana). We elaborate on these flaws in Section 2.", + "bbox": [ + 212, + 627, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From an evaluation perspective, imperfections in ground-truth data make the metrics less reliable. COCO-2017 AP would wrongfully penalize models that predict more precise masks than the imperfect ground-truth masks. Moreover, models trained with such imperfect masks may exploit unwanted biases from the training data, e.g., learning to never predict masks with holes. Recently, Kirillov et al. observed that human evaluators consistently rated the masks predicted by their Segment Anything Model (SAM) to be of higher quality than a ViTDet model [23] trained on COCO, whereas the latter scored higher on COCO AP.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Singh et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These observations raise an important question - can COCO reliably benchmark the future progress in object detection research? COCO is the de-facto benchmark for object detection - countless research papers report COCO results every year; libraries like Torchvision [29] and Detectron2 [36] readily support object detector development with COCO. Hence, to maintain continuity with prior research and tap into this rich open-source ecosystem, we choose to rectify the errors of COCO and reinforce its utility for future research.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contributions. We develop COCO-ReM (Refined Masks), a cleaner set of high-quality instance annotations for COCO images. The improved quality over COCO-2017 masks is visually palpable, as seen in Figure 1.", + "bbox": [ + 212, + 253, + 787, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our annotation pipeline (Section 3) utilizes SAM [21] to refine masks, and uses the LVIS dataset [14] and an ensemble of LVIS-trained models to exhaustively annotate grouped or missing instances. We (the authors) manually verify the entire validation set to provide a strong quality guarantee for evaluation.", + "bbox": [ + 212, + 300, + 787, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Section 4, we use COCO-ReM for benchmarking object detectors. We evaluate fifty object detectors and observe that their COCO-ReM AP produces a different model ranking than COCO-2017 AP. Surprisingly, we observe that query-based models (Mask2Former and OneFormer) score much higher on COCO-ReM than region-based models (ViTDet). Query-based models predict visually sharper masks - AP trends on COCO-ReM accurately reflect this, unlike COCO-2017. We observe that models trained using COCO-ReM converge faster and perform better than those trained using COCO-2017, highlighting the importance of mask quality in improving object detectors.", + "bbox": [ + 212, + 361, + 787, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Background: Revisiting COCO Masks", + "text_level": 1, + "bbox": [ + 214, + 518, + 619, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The annotation pipeline of a dataset determines the quality of annotations obtained. To design an effective annotation pipeline for COCO-ReM, we review the COCO annotation pipeline by Lin et al. [25] and understand its shortcomings.", + "bbox": [ + 212, + 550, + 787, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "COCO annotation pipeline. Lin et al. [25] employed a three-stage pipeline to collect instance mask annotations for COCO images:", + "bbox": [ + 212, + 598, + 787, + 628 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Category labeling: Annotators assigned one or more labels to images depending on the visible content, choosing from a fixed set of categories.", + "2. Instance spotting: Annotators locate up to ten instances per label (from the previous stage) in the corresponding image, by providing clicks (points).", + "3. Instance segmentation: Finally, annotators clicked around the boundaries of each located instance (from the previous stage), rendering a polygon mask. Lin et al. [25] also took numerous measures to improve annotation quality. The first two stages - category labeling and instance spotting - were performed by eight annotators per image to maximize recall. The instance segmentation stage required annotators to complete a training task before annotating actual instances in the dataset. Next, a different set of crowd workers verified the correctness and quality of annotated masks. Finally, the annotators used a paint-brush tool to mark remaining objects as crowd regions. These extensive measures avoided many scenarios that could have led to low-quality masks." + ], + "bbox": [ + 215, + 628, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "COCO-ReM Benchmark for Object Detection", + "bbox": [ + 424, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a85a2a36df96677fe3711cd2e8fc3dfe84022b5ae482a9f270c3b842e2068111.jpg", + "image_caption": [ + "(a) Holes" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 367, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/164ecf5f97458ae414f1efc41f0ec0f7b8261091beab2607dcd0bdcfeb0f17de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 143, + 513, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f401c01c49242f3a8a8442fa7073fdb7f8326fd126c7ad29efa731abf0d25a9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 143, + 650, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/156e5ae362c04b3fcb861b82b8154913933b9e842bf897644e42f1018563d3f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 143, + 784, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/024087c82d1006410b557aa3d76e1090adfde3fee288435f5175ca96d69e7968.jpg", + "image_caption": [ + "(c) Missing annotations and labeling multiple instances as one" + ], + "image_footnote": [], + "bbox": [ + 217, + 234, + 364, + 311 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cdb9f7686de82db74f57114d736848c7c72bd1f6b7a09001c7307c0eae433ed2.jpg", + "image_caption": [ + "Fig. 2: Imperfections in COCO-2017 masks. All masks in general have coarse boundaries. Moreover, they lack holes when required (scissors) and do not handle occlusions consistently ( $TV$ with or without figurines). Annotations are sometimes non-exhaustive, either masks are missing (apples and donuts), or multiple instances are grouped into a single mask (banana and dog)." + ], + "image_footnote": [], + "bbox": [ + 364, + 234, + 514, + 311 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5f298d7666a58ec6a6271b8cb8d177e3664461a110fda874e8d381b54f525e02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 234, + 651, + 311 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b25c53ca87e7d70e76612b23089761504113a0693a6cd32da86cf4ff83f39122.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 234, + 784, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Imperfections in COCO-2017 masks. To further understand the flaws in COCO annotations, we visualized thousands of COCO-2017 masks and inspected their quality and label correctness. Through this exercise, we observed recurring types of errors in annotations, as shown in Figure 2.", + "bbox": [ + 212, + 421, + 787, + 479 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. COCO-2017 masks have coarse boundaries. This is a natural consequence of collecting masks as polygon contours, which only coarsely approximate the pixel-precise mask boundaries through piece-wise straight edges connecting points (clicks) supplied by annotators. This shortcoming is well-known and identified by several follow-up works [1, 14, 21].", + "2. COCO-2017 masks lack holes when required. The polygon-drawing interface of COCO did not allow annotators to erase the interior regions of polygons. Consequently, certain objects that should naturally contain holes, lack them. Figure 2a shows two examples - the handles of scissors and cup.", + "3. COCO-2017 masks do not handle occlusions consistently. Masks of occluded objects are either drawn over the other occluding objects (amodal masks) or around them (modal masks). For example, observe the $TV$ masks in Figure 2b. While both, amodal and modal masks are valid, either one should be consistent in the dataset according to the task definition. We believe this inconsistency is a consequence of inadequate task instructions to annotators as well as the lack of holes in masks.", + "4. COCO-2017 masks are often non-exhaustive. This issue occurs in two modes, as shown in Figure 2c - either few instance masks of a category are missing1 (apple, donut), or multiple instances are grouped in a single mask (banana, dog). This is a recurring issue throughout COCO - we believe that certain annotators may not have deemed masks with grouped instances to have a wrong object contour, due to oversight." + ], + "bbox": [ + 215, + 482, + 787, + 813 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Singh et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 These images have all their missing masks annotated as non-crowd regions.", + "bbox": [ + 217, + 824, + 733, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. COCO-2017 contains few near-duplicate masks. We found 410 pairs of masks that overlap with intersection-over-union (IoU) greater than 0.8 in the COCO-2017 validation set ( $\\approx 2.3\\%$ of all instances). These pairs often segment the same underlying object with different labels. Examples include armchairs (chair and couch), calves (cow and sheep), and duffel bags (suitcase and backpack). We attribute this issue to the lack of a cross-category deduplication step in the COCO annotation pipeline.", + "bbox": [ + 215, + 146, + 787, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All these imperfections hurt the reliability of the benchmarking process. They provide noisy supervision during training and wrongfully penalize correct model predictions during evaluation. Having characterized these imperfections, we make targeted design choices in our COCO-ReM annotation pipeline to rectify them.", + "bbox": [ + 212, + 252, + 785, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 COCO-ReM: COCO with Refined Masks", + "text_level": 1, + "bbox": [ + 215, + 330, + 650, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we outline our annotation process for COCO-ReM. Our goal is to rectify the imperfections in COCO-2017 annotations as discussed in Section 2 to maintain continuity with prior research. We hope that higher quality ground-truth masks would improve the reliability of the benchmarking process.", + "bbox": [ + 212, + 359, + 785, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Annotation Pipeline", + "text_level": 1, + "bbox": [ + 215, + 438, + 429, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We develop a semi-automatic pipeline - first, we refine the masks using off-the-shelf models and LVIS dataset [14], followed by manual verification of the updated masks. We refine both, the training and validation set of COCO-2017, comprising $\\approx 860\\mathrm{K}$ and $\\approx 36\\mathrm{K}$ instances respectively. We thoroughly verify the validation set to ensure consistency and pixel-precise mask quality for model evaluation. As the training set is much larger, it is only processed through the automatic steps of our pipeline. Our pipeline comprises three stages (Figure 3).", + "bbox": [ + 212, + 459, + 785, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 1. Mask boundary refinement. In this stage, we aim to rectify the coarse boundary quality of COCO-2017 masks. A straightforward solution to achieve this is to re-associate all masks through a web interface with paintbrush and eraser tools instead of drawing polygon contours. Moreover, this solution naturally allows including holes in masks where necessary. While accurate, this solution would sharply increase the mask annotation time and cost.", + "bbox": [ + 212, + 568, + 785, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Along a different trend, prior works have explored model-assisted interactive segmentation as a feasible solution for collecting mask annotations with smooth boundaries [1, 14]. The key idea is to let an object detector predict a coarse initial mask and have the annotators refine it by prompting corrective mouse clicks. Inspired by the accuracy and practical feasibility of this approach, we use the Segment Anything Model (SAM [21]) to refine the COCO-2017 masks.", + "bbox": [ + 212, + 659, + 785, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Refining masks using SAM. SAM is a promptable segmentation model that predicts high-quality masks conditioned on user-provided point and box prompts. This promptable functionality makes SAM suitable for our use case. We prompted SAM using various combinations of prompts sampled using COCO-2017 masks and obtained its predictions as the refined masks. Refer to the appendix for details about this prompting methodology.", + "bbox": [ + 215, + 750, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "COCO-ReM Benchmark for Object Detection", + "bbox": [ + 424, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ef3c160383a599d8ec932245d3b44adf2318d46fb8cf14ec6a73eedf138d5e30.jpg", + "image_caption": [ + "Fig. 3: COCO-ReM annotation pipeline. We develop a semi-automatic annotation pipeline comprising three stages to rectify the imperfections in COCO-2017 masks." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Manual verification of mask quality. We visualized and manually inspected every mask created in the previous process. Through this inspection, we identified nearly 900 masks that have low quality or do not accurately cover the underlying object. These masks represent only $2.4\\%$ of 36K instances in the COCO-2017 validation set, greatly reducing our manual effort and showing the efficacy of our automated process. To correct these masks, we deployed a web browser demo of SAM and provided point prompts iteratively until obtaining a precise and accurate mask.", + "bbox": [ + 215, + 351, + 787, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through this stage, we obtain masks with more precise boundaries and better handling of occlusions, as illustrated in Figure 3 (left).", + "bbox": [ + 215, + 473, + 785, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 2. Exhaustive instance annotation. This stage addresses non-exhaustively annotated instances in COCO-2017 (Figure 2c). We note that manually re-annotating every image is the most reliable solution to ensure exhaustive instance coverage. However, the manual effort would not only require extensive instructions and verification but would also be highly cost-ineffective and may lead to redundant annotations. To mitigate these issues, we automatically import instances from the LVIS [14] dataset as a strong starting point.", + "bbox": [ + 212, + 506, + 815, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Importing instances from LVIS. The LVIS dataset comprises 1.2M high-quality instance annotations across 1203 categories for images in COCO, including all 80 COCO categories. Gupta et al. [14] took firm measures to ensure exhaustive annotations. Every image is labeled with a set of positive LVIS categories whose instances are exhaustively labeled, and a set of negative LVIS categories that are verifiably absent. The presence (or absence) of the remaining LVIS categories is unverified for the image.", + "bbox": [ + 215, + 613, + 787, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the 80 COCO categories, LVIS contains $\\approx 20\\mathrm{K}$ instances for images in the COCO-2017 validation set. To import these instances, we checked whether LVIS contained more instances of an (image, category) pair than COCO-2017. If so, we replaced all COCO-2017 instances of this pair with those from LVIS. Nearly $9\\%$ of (image, category) pairs in LVIS are flagged to have nonexhaustive instance annotations [14] - we ignored these while importing.", + "bbox": [ + 230, + 719, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Discovering instances using LVIS-trained models. Instances of a given (image, category) pair of COCO-2017 may be missing in LVIS if the category", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Singh et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2be83d315527fd14b7c33e54bd2c93437019664aec47c36aa2040a5430fba003.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSplitInstance SourceTotal
COCO-2017LVISLVIS models
COCO-2017val36,781--36,781
COCO-ReMval33,4986,135105640,689
COCO-2017train860,001--860,001
COCO-ReMtrain738,353354,674-1,093,027
", + "bbox": [ + 215, + 143, + 785, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Instance statistics in COCO-ReM. We list the exact number of annotations obtained from different sources for both train and val splits of COCO-ReM. All masks undergo refinement through SAM before being included in COCO-ReM.", + "bbox": [ + 215, + 250, + 782, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is unverified for the corresponding image. Instances of this pair may be nonexhaustive in COCO-2017, and hence remain unresolved after the previous automatic step. To resolve this, we use publicly available object detectors trained using LVIS. On the validation set, we ensemble mask predictions using the official checkpoints of three best-performing ViTDet models [23]. As model predictions are not as accurate as human-annotated data, we manually inspect $\\approx 2500$ masks predicted with high confidence and select 1056 masks. We refine these masks (Stage 1) and add them to COCO-ReM.", + "bbox": [ + 232, + 311, + 784, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This stage augments COCO-ReM with additional instances from external sources and includes more exhaustive annotations, see example in Figure 3 (middle).", + "bbox": [ + 215, + 434, + 782, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stage 3. Correction of labeling errors. This is the final stage of our pipeline, performed manually, only for the validation set. Recall from Section 2 that we found 410 near-duplicate mask pairs in the COCO-2017 validation set. We manually inspected these and retained one mask per pair with the most accurate label, such as chair for armchairs and car for SUVs (Figure 3 (right)). We also observed that $\\approx 100$ grouped instances were not resolved through Stage 2. We annotated these masks as crowd to accurately reflect their characteristics.", + "bbox": [ + 215, + 467, + 784, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Mask Characteristics", + "text_level": 1, + "bbox": [ + 215, + 595, + 434, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instance statistics. Our annotation pipeline imports instances from external sources - LVIS dataset and LVIS-trained models - to annotate instances exhaustively. As a result, every instance in COCO-ReM traces to one of these sources or COCO-2017. As shown in Table 1, COCO-ReM validation set extends COCO-2017 by including $\\approx 6\\mathrm{K}$ masks from LVIS [14] and $\\approx 1\\mathrm{K}$ masks from LVISTrained models [3, 23]. Since these masks replace a subset of COCO-2017 masks (see Stage 2), we retain $\\approx 33.5\\mathrm{K}$ masks from COCO-2017 (out of $36\\mathrm{K}$ masks). Overall, COCO-ReM validation set has $\\approx 4\\mathrm{K}$ more instances than COCO-2017.", + "bbox": [ + 215, + 625, + 784, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the other hand, the training set of COCO-ReM comprises a total of 1.09M high-quality masks, as compared to 860K masks in the COCO-2017 training set. The difference in instance counts between COCO-ReM and COCO-2017 sharply highlights the extent of non-exhaustively annotated instances in COCO-2017.", + "bbox": [ + 215, + 746, + 784, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask IoU. Figure 4 (left) shows the distribution of IoU between the (refined) masks of COCO-ReM (validation set) and their corresponding source masks in", + "bbox": [ + 215, + 809, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "COCO-ReM Benchmark for Object Detection", + "bbox": [ + 424, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3edfe2a4dce42f677abb07768602a9c87fb4a8a3575d82818686ee43417c25e0.jpg", + "image_caption": [ + "Fig. 4: Left: Mask IoU distribution. Distribution of IoU between masks in COCO-ReM validation set and their source dataset. Right: Masks with holes. Nearly 2000 masks in COCO-ReM validation set have holes, that were missing in COCO-2017." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 614, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/612ce9a12d3c9a1c22ab482be51666ca946822a59a47f41b61eff57c3c311e43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 148, + 779, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "COCO-2017 and LVIS datasets. The median IoU is slightly less than one, indicating that most masks have a proper coarse shape, but lack precise boundaries. This median estimate matches the findings of Gupta et al. [14], albeit observed from a smaller sample of 100 masks collected from human annotators.", + "bbox": [ + 212, + 342, + 782, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Holes in masks. Figure 4 (right) shows that $\\approx 2000$ masks in COCO-ReM validation set have at least one hole. By design, COCO-2017 lacks these holes. See the appendix for qualitative examples and additional analysis per category.", + "bbox": [ + 212, + 405, + 784, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 473, + 375, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our experiments, we aim to demonstrate the utility of COCO-ReM for reliable benchmarking of object detection research. Recall Section 1, we hypothesize that the imperfections in COCO-2017 may cause two issues: (1) During evaluation, models may be undesirably penalized despite predicting correct masks, yielding counter-intuitive modeling observations, and (2) During training, models receive noisy supervision and/or may learn unwanted biases (e.g., never predict masks with holes), resulting in sub-par capabilities. We claim that high-quality masks of COCO-ReM can alleviate these issues. To this end, we use COCO-ReM to evaluate and train object detectors, and compare trends with COCO-2017.", + "bbox": [ + 212, + 505, + 787, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Evaluation using COCO-ReM", + "text_level": 1, + "bbox": [ + 214, + 662, + 508, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we evaluate fifty publicly available COCO object detectors using the COCO-ReM validation set. We focus on the validation set as the test set annotations of COCO-2017 are not publicly available. We cover a wide range of models from existing literature. All models can be categorized as either region-based or query-based:", + "bbox": [ + 212, + 688, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Region-based models, following Mask R-CNN [17]. These models first predict labeled boxes, followed by binary masks inside those boxes. We also include cascade models, following Cascade R-CNN [3]. These models extend the region-based design by iteratively refining their predicted bounding boxes before making the final mask prediction.", + "bbox": [ + 215, + 763, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Singh et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e7d7f8d08f0a50f02103090b39eaa88181f4721c4dfdb9e7b48a8093b489de6.jpg", + "image_caption": [ + "Fig. 5: Evaluating fifty object detectors using COCO-ReM. All models score higher on COCO-ReM than COCO-2017. Query-based models ( $\\star$ ) score much higher than region-based models on COCO-ReM, yielding opposite trends than COCO-2017." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 781, + 284 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- **Query-based models**, inspired by the Detection Transformer design [4]. Instead of predicting masks within image regions, these models start with a set of image-independent query vectors, then decode labeled masks through a Transformer [33] by iteratively attending to image embeddings. Among query-based models, we consider Mask2Former [8] and OneFormer [19].", + "bbox": [ + 215, + 358, + 782, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We ensure diverse image backbones in our model collection. The architecture and capacity of the backbone significantly impact object detector performance. Thus, we include convolutional backbones (ResNet [16], RegNet [31], ConvNeXt [27], InternImage [35]), hierarchical vision transformers (Swin [26], MViTv2 [24], DiNAT [15]), and plain vision transformers [11]. We include backbones of different sizes for a single model, e.g., for Cascade ViTDet [23], we include ViT-B (141M parameters), ViT-L (361M parameters), ViT-H (692M parameters).", + "bbox": [ + 212, + 435, + 782, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We believe that our model collection is sufficiently diverse to draw meaningful conclusions from comparing trends between COCO-ReM and COCO-2017", + "bbox": [ + 212, + 541, + 782, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Testing protocol. We compute the mask Average Precision (AP) of all models using the official COCO evaluation API in Detectron2 [36]. For simplicity, we adopt a common testing protocol for all models - we resize the longest edge of input images to 1024 pixels (without changing the aspect ratio), and limit the number of detections per image to 100.", + "bbox": [ + 212, + 575, + 784, + 651 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Result 1. All models score higher on COCO-ReM. In Figure 5, we plot the mask AP trends for all fifty object detectors sorted by their COCO-2017 AP. All fifty models achieve higher AP on COCO-ReM despite being trained using only COCO-2017 masks. This result validates our hypothesis that the imperfections of COCO-2017 may have wrongfully penalized all these models.", + "bbox": [ + 212, + 654, + 782, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The difference between COCO-ReM AP and COCO-2017 AP is larger with recent high-performing models - the current state-of-the-art model (OneFormer InternImage-H) scores $+7.7$ points higher AP on COCO-ReM than COCO-2017. This increasing gap in AP suggests that the imperfections in ground-truth have worsened the issue of over-penalization and made COCO-2017 AP less adept at fine-grained assessment and ranking of recent models.", + "bbox": [ + 212, + 729, + 782, + 820 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Result 2. Query-based detectors score much higher on COCO-ReM.", + "bbox": [ + 215, + 824, + 777, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "COCO-ReM Benchmark for Object Detection", + "bbox": [ + 424, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b53c630ec91e3aa63f37b4738179b34be24250dda2004bca31d9c18046c316b6.jpg", + "image_caption": [ + "Cascade ViTDet-L" + ], + "image_footnote": [], + "bbox": [ + 218, + 154, + 357, + 224 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/94e64f762682d91157bff782dbef335746ff1914bd9df9ef13621a6b6370c148.jpg", + "image_caption": [ + "Fig. 6: Mask predictions from models with opposite rank order on COCO-ReM and COCO-2017. Qualitative examples of ViTDet (region-based model) and Mask2Former (query-based model) show that the latter predicts visually sharper masks. COCO-ReM AP correctly ranks Mask2Former higher, unlike COCO-2017 AP." + ], + "image_footnote": [], + "bbox": [ + 218, + 224, + 357, + 285 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c4e1e7d5b611a6329229028c8b1f3185c8d24ba16f25851e9ded27eabc9205e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 154, + 496, + 285 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f53cb8956febce8ee7ff7ee71fea5673e5b19f21bdd6340fd17ca511acbd7974.jpg", + "image_caption": [ + "Mask2Former Swin-L" + ], + "image_footnote": [], + "bbox": [ + 504, + 154, + 642, + 224 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/beb3e0aac27eb5dadcf890bbcbb155b3d5305d5e1134ce28a6939a5f0ec20521.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 224, + 642, + 285 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/49f5a704a75040d2ad1e43e4c328a755e60e471a3298ac07cf5a04ed81414926.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 154, + 781, + 285 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Interestingly, we observe that the trend in AP values calculated using COCO-ReM does not align perfectly with COCO-2017. We observe frequent flips in ranks of several model pairs on COCO-ReM. While the model rankings do not transfer between benchmarks perfectly, we observe a consistent pattern in model pairs with flipped rankings. Notice the star markers $(\\star)$ in Figure 5 - the flips are attributed to significantly higher scores of query-based models (Mask2Former [8] and OneFormer [19]) than region-based models on COCO-ReM.", + "bbox": [ + 212, + 369, + 784, + 474 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To further understand the cause of this performance trend, we observe the predicted masks of two models having opposite rank order according to COCO-ReM and COCO-2017 AP: Cascade ViTDet-L [23] and Mask2Former Swin-L [8]. The latter is a query-based model. Figure 6 shows some predicted masks by both of these models. Through a cursory glance, we can observe that Mask2Former predicts visually sharper masks whereas masks of ViTDet have coarse boundaries. This human judgment aligns better with COCO-ReM AP, which correctly rates Mask2Former higher. On the other hand, COCO-2017 AP rates ViTDet higher, yielding a counter-intuitive modeling observation.", + "bbox": [ + 212, + 476, + 785, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Guided by these trends, we reaffirm our claim that benchmarking newer models using the COCO-2017 may be misleading from a research perspective, and the community should adopt our COCO-ReM annotations moving forward. We include numerical results of Figure 5 in the appendix.", + "bbox": [ + 212, + 612, + 785, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Understanding the Difference in AP", + "text_level": 1, + "bbox": [ + 214, + 691, + 560, + 708 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Recall Section 4.1 (Result 1) - all models score higher average precision (AP) on COCO-ReM, as compared to COCO-2017. What explains the difference in $AP$ ? To understand the reasons underlying this difference, we observe the components of COCO AP for a few high-performing models from Section 4.1. For this analysis, we consider two Cascade ViTDet models [23] using ViT-L/H backbones [11], and two Mask2Former models [8] using Swin-B/L backbones [26].", + "bbox": [ + 212, + 715, + 785, + 806 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "AP per IoU threshold. The calculation of COCO AP involves an average of per-category AP scores of 80 COCO categories, each computed at ten different", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Singh et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6dce440b69244498cf17db818348e786f7e4cb6f5f02cd116959d0b0a9011b0e.jpg", + "image_caption": [ + "Fig. 7: AP per IoU threshold. While models score similarly at IoU = 0.5 on COCO-ReM and COCO-2017, they score much higher on COCO-ReM at higher IoU thresholds (0.75 and beyond), contributing to the overall difference in AP. Due to the improved mask quality of COCO-ReM, the AP at high IoU thresholds becomes more sensitive and enables better assessment of model capabilities." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 771, + 287 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b9e2e04dae014ccdf706c6b3d0ed925153d7bb8088a9c8434a502663e9b32210.jpg", + "image_caption": [ + "Fig. 8: AP per category. Models score higher AP on COCO-ReM for 69 out of 80 categories. The masks of the remaining 11 categories (left) in COCO-2017 do not handle occlusions consistently (dining table, bed), lack holes (scissors, donut), or have grouped instances (banana, apple). Models learn these biases from COCO-2017 training masks and score lower COCO-ReM AP." + ], + "image_footnote": [], + "bbox": [ + 225, + 367, + 772, + 511 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "IoU thresholds, $t \\in \\{0.5, 0.55, \\dots, 0.9, 0.95\\}$ . In Figure 7, we show components of AP at different IoU thresholds, averaged for the four models we use. On average, models score similarly at IoU = 0.5 for COCO-ReM and COCO-2017. However, COCO-ReM AP is much larger for IoU = 0.75 and beyond. Hence, higher scores on COCO-ReM can be attributed to AP at higher IoU thresholds.", + "bbox": [ + 212, + 595, + 787, + 670 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Intuitively, AP at IoU = t rewards model predictions that overlap with a ground-truth mask of the same category with IoU ≥ t. AP at IoU = 0.5 requires a very coarse overlap between the predicted and ground-truth masks, whereas AP at high IoU thresholds have stricter overlap requirements. Since IoU is symmetric, imprecise boundaries of ground-truth masks (COCO-2017) can wrongfully penalize valid predictions and lead to lower AP scores. COCO-ReM fixes this issue, as evident with higher AP at high IoU thresholds. This analysis hints that precise masks of COCO-ReM improve the sensitivity of AP and enable fine-grained assessment of high-performing object detectors.", + "bbox": [ + 212, + 671, + 787, + 806 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "AP per category. Next, we calculate AP for every COCO category by averaging across all IoU thresholds. Figure 8 shows the difference in COCO-ReM", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "COCO-ReM Benchmark for Object Detection", + "bbox": [ + 424, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d3c1adaedf9e1a609e0b3abc100b9b7b8ec9e5cb2681f8d1dfbb72fbfc0bd2ad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
COCO evaluation set ⇒2017ReM-S1ReM-S2aReM (final)
AP49.657.355.655.5
Number of instances36,78136,78141,16840,689
", + "bbox": [ + 215, + 143, + 785, + 199 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: COCO-ReM ablations. AP of four high-performing models (Cascade ViTDet-B/L and Mask2Former Swin-B/L) using intermediate versions from our COCO-ReM annotation pipeline. Trends suggest that the coarse boundary quality of COCO-2017 masks was a major source of noise in the benchmarking process.", + "bbox": [ + 212, + 200, + 787, + 257 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "AP and COCO-2017 AP per category, averaged across the same four models. Models score higher on COCO-ReM for 69 out of 80 categories, which is reflected in the higher overall COCO-ReM AP. Significant improvements are observed in categories with generally consistent appearance (carrot and orange), and fine details (fork). Intuitively speaking, the larger difference in per-category AP can be attributed to the better quality of ground-truth masks in COCO-ReM.", + "bbox": [ + 212, + 276, + 785, + 366 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For 11 categories, models score lower on COCO-ReM than COCO-2017. These categories either have holes in their natural appearance or are often heavily occluded (bowl, dining table, scissors). Masks in COCO-ReM include holes and handle occlusions more consistently than in COCO-2017. Hence, lower per-category AP with COCO-ReM implies that the imperfections of COCO-2017 training masks impart these undesirable biases to trained models – upon checking the model predictions, we observed that they often include occluding objects for dining table and lack holes for scissors. Such predictions are not rewarded by COCO-ReM, hence incur a drop in the per-category AP.", + "bbox": [ + 212, + 367, + 787, + 503 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablations", + "text_level": 1, + "bbox": [ + 215, + 521, + 344, + 535 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we conduct a simple ablation study to observe the contribution of different stages in our annotation pipeline. For the sake of simplicity, we select the same four models as chosen in Section 4.2. We evaluate these models using two versions of COCO-ReM validation set:", + "bbox": [ + 212, + 542, + 785, + 602 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. ReM-S1: This version is obtained after the Stage 1 (mask boundary refinement) in our annotation pipeline. It comprises the same number of instances as COCO-2017, each with an improved boundary quality.", + "2. ReM-S2a: This version is obtained after importing instances from the LVIS dataset – the automatic part of Stage 2 (exhaustive instance annotation). We exclude the manual part of Stage 2 from this ablation for simplicity." + ], + "bbox": [ + 215, + 603, + 784, + 693 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results are shown in Table 2. The large difference between COCO-ReM AP and COCO-2017 AP appears right after Stage 1, indicating that the coarse mask boundaries are a major source of noise that hurts the reliability of COCO-2017 AP. Subsequent stages incur a negligible drop in AP, likely due to larger of the validation set, as COCO AP is known to favor smaller evaluation sets [14].", + "bbox": [ + 212, + 694, + 785, + 771 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Training with COCO-ReM", + "text_level": 1, + "bbox": [ + 215, + 787, + 483, + 803 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our evaluation shows COCO-ReM as a useful benchmark for accurate modeling conclusions. In this section, we train strong baseline models using COCO-ReM.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Singh et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/aca81a5fb569ac865ffeb8877a78e2719cbce4efb6ed8ec55b6548b327912667.jpg", + "image_caption": [ + "Fig.9: Training with COCO-ReM. A ViTDet model trained using COCO-ReM training set converges faster and performs better than its counterpart trained using COCO-2017, and closely matches a $3 \\times$ larger ViTDet-L trained using COCO-2017." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 504, + 239 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/202992b70680db8b366a6d249176b1df967e7b795404b349a76d026ad450d029.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 146, + 777, + 239 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Implementation details. We train ViTDet models [23] using their official open-source implementation [36], leaving all hyperparameters unchanged. We train two models - Mask R-CNN ViTDet-B (111M parameters) and Cascade ViTDet-B (141M parameters). Li et al. [23] trained this model for 100 COCO epochs across $64 \\times$ A100 GPUs (batch size of 64). To manage these costs, we implement gradient accumulation for every 8 iterations to emulate a batch size of 64 using our $8 \\times$ A40 GPUs, resulting in an 11 day-long training schedule. We could only afford to train the base (ViT-B) models due to limited resources. The gorund truth annotations for test set is inaccessible so, we use the validation set solely for evaluation purposes for fair assessment of models.", + "bbox": [ + 212, + 315, + 787, + 467 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. Figure 9 shows that our ViTDet-B models trained using COCO-ReM significantly outperform the same ViTDet-B models trained using COCO-2017. Higher quality of training data enables faster convergence. Moreover, our ViTDet-B models are competitive with $3 \\times$ larger ViTDet-L models trained using COCO-2017, indicating better parameter efficiency. While numerous prior works improve object detectors through modeling innovations, our results highlight the critical role of data quality in improving object detectors.", + "bbox": [ + 212, + 470, + 789, + 578 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Related Work", + "text_level": 1, + "bbox": [ + 215, + 604, + 387, + 619 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Object detection and image segmentation. Visual perception is fundamental for any vision system to understand the world around it. At the core of a visual perception system is the task of image segmentation that aims at recognizing the entities (objects) in a given scene and densely labeling every pixel constituting the corresponding object. Image segmentation can broadly be divided into three sub-tasks: semantic, instance, and panoptic segmentation. On the one hand, semantic segmentation [5, 6, 28] targets obtaining a single amorphous binary mask covering all pixels for the corresponding category. On the other hand, instance segmentation [3, 18] involves detecting pixels for only foreground \"thing\" regions with clear boundaries and differentiating among separate instances for a category. Panoptic segmentation [7, 20, 34] is a combination of the former two tasks targeted at predicting distinct binary masks for \"thing\" regions and a single amorphous mask for the \"stuff\" regions.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "COCO-ReM Benchmark for Object Detection", + "bbox": [ + 424, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this work, we particularly look at the task of instance segmentation and contribute high-quality mask annotations to benchmark further progress.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Image segmentation benchmarks. In the last decade, considerable advancements have been observed in the classical vision task of image segmentation. A part of this progress is owed to the introduction of numerous datasets for benchmarking image segmentation models. In one of the earliest efforts, Everingham et al. [12] introduced the PASCAL-VOC-2012 challenge to recognize 20 classes in about 10K images. More challenging datasets were introduced in the following years, with ADE20K [10] covering 150 classes. Lin et al. [25] introduced the concept of \"stuff\" and \"thing\" classes in their COCO dataset with a total of 171 classes: 80 \"thing\" and 91 \"stuff\". The Cityscapes [9], Mapillary-Vistas [30], and KITTI [13] datasets introduced image segmentation datasets focused on autonomous driving applications.", + "bbox": [ + 212, + 183, + 787, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Gupta et al. [14] introduced the LVIS dataset with exhaustive annotations covering 1203 categories for COCO images. OpenImages [1] introduced 2.8M mask annotations covering 350 classes. Recently, the SA-1B dataset [21] introduced over 1.1B fine-grained, class-agnostic mask annotations for over 11M images. Despite various datasets, COCO has remained the de-facto benchmark for object detectors in instance segmentation. In this work, we examined the COCO instance annotations and proposed a new benchmark with superior mask quality.", + "bbox": [ + 212, + 353, + 787, + 460 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Re-assessing datasets for benchmarking. Human-annotated datasets are always prone to errors and imperfections, as no human is perfect. Some prior works have shared our motivation and conducted in-depth re-assessment studies, the closest to ours are works that re-assess the ImageNet dataset [2, 32]. Concurrent to our work, Zimmermann et al. [37] also re-assess COCO annotations, however they opt for recollecting polygon masks from human annotators.", + "bbox": [ + 212, + 465, + 787, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 601, + 359, + 617 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we revisit the COCO-2017 mask annotations and find major issues involving, but not limited to their boundary quality and exhaustiveness. These issues raise significant concerns about the dependability of COCO-2017 to benchmark recent and future object detectors accurately. Consequently, we introduce our COCO-ReM benchmark consisting of high-quality mask annotations obtained using an effective semi-automatic annotation pipeline.", + "bbox": [ + 212, + 654, + 787, + 744 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "On comparing the mask AP trends for existing object detectors on COCO-ReM and COCO-2017, we notice significant changes in model rankings, notably contradicting the conclusions drawn through COCO-2017. Moreover, training using COCO-ReM shows that paying attention to mask quality is crucial to advancing the capabilities of object detectors. We hope that COCO-ReM will aid the future research in object detection.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Singh et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Mohamed El Banani, Agrim Gupta, and Nilesh Kulkarni for helpful discussions and feedback. We thank Palak Jagtap for her help during manual verification of masks. We thank all the authors of the prior works for publicly releasing the pre-trained model checkpoints, which helped us maintain empirical comprehensiveness in our evaluation study. We thank Vision and Language Group, IIT Roorkee for being a constant support system throughout the process.", + "bbox": [ + 215, + 176, + 785, + 268 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "COCO-ReM Benchmark for Object Detection", + "bbox": [ + 426, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Bibliography", + "text_level": 1, + "bbox": [ + 426, + 141, + 576, + 162 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 4, 5, 14", + "[2] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aïron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020. 14", + "[3] Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN: Delving into high quality object detection. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7, 8, 13", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 2, 9", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In ICLR, 2015. 13", + "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In TPAMI, 2017. 13", + "[7] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 13", + "[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 9, 10", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 14", + "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. Semantic understanding of scenes through the ade20k dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 14", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer," + ], + "bbox": [ + 217, + 210, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 9, 10", + "[12] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2012. 14", + "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 14", + "[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3, 4, 5, 6, 7, 8, 12, 14", + "[15] Ali Hassani and Humphrey Shi. Dilated neighborhood attention transformer. arXiv:2209.15001, 2022. 9", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 9", + "[17] Kaiming He, Georgia Gkioxari, Piotr Dolkar, and Ross Girshick. Mask R-CNN. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 2, 8", + "[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 13", + "[19] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 9, 10", + "[20] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 13", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, WanYen Lo, Piotr Dolkar, and Ross Girshick. Segment Anything. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 14", + "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2012. 1", + "[23] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 7, 9, 10, 13", + "[24] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale" + ], + "bbox": [ + 217, + 145, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "COCO-ReM Benchmark for Object Detection", + "bbox": [ + 424, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision transformers for classification and detection. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 9", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 1, 2, 3, 14", + "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2021. 9, 10", + "[27] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 9", + "[28] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 13", + "[29] TorchVision maintainers and contributors. TorchVision: PyTorch's Computer Vision library. https://github.com/pytorch/vision, 2016.3", + "[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 14", + "[31] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 9", + "[32] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In Proceedings of the International Conference on Machine Learning (ICML), 2019. 14", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2, 9", + "[34] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 13", + "[35] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 9", + "[36] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. 3, 9, 13", + "[37] Eric Zimmermann, Justin Szeto, Jerome Pasquero, and Frederic Ratle. Benchmarking a benchmark: How reliable is ms-coco? arXiv preprint arXiv:2311.02709, 2023. 14" + ], + "bbox": [ + 217, + 145, + 787, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Singh et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/64864c4f-d7d3-4bce-8017-e14998347c20_model.json b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/64864c4f-d7d3-4bce-8017-e14998347c20_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9305d6e15d812114c7d1a101e26e9328c35024e3 --- /dev/null +++ b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/64864c4f-d7d3-4bce-8017-e14998347c20_model.json @@ -0,0 +1,2546 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.244, + 0.141, + 0.761, + 0.185 + ], + "angle": 0, + "content": "Benchmarking Object Detectors with COCO: A New Path Forward" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.212, + 0.677, + 0.227 + ], + "angle": 0, + "content": "Shweta Singh1* Aayan Yadav1* Jitesh Jain2" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.228, + 0.685, + 0.242 + ], + "angle": 0, + "content": "Humphrey Shi² Justin Johnson³ Karan Desai³" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.242, + 0.702, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)IIT Roorkee \\(^{2}\\)Georgia Tech \\(^{3}\\)University of Michigan" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.289, + 0.744, + 0.568 + ], + "angle": 0, + "content": "Abstract. The Common Objects in Context (COCO) dataset has been instrumental in benchmarking object detectors over the past decade. Like every dataset, COCO contains subtle errors and imperfections stemming from its annotation procedure. With the advent of high-performing models, we ask whether these errors of COCO are hindering its utility in reliably benchmarking further progress. In search for an answer, we inspect thousands of masks from COCO (2017 version) and uncover different types of errors such as imprecise mask boundaries, non-exhaustively annotated instances, and mislabeled masks. Due to the prevalence of COCO, we choose to correct these errors to maintain continuity with prior research. We develop COCO-ReM (Refined Masks), a cleaner set of annotations with visibly better mask quality than COCO-2017. We evaluate fifty object detectors and find that models that predict visually sharper masks score higher on COCO-ReM, affirming that they were being incorrectly penalized due to errors in COCO-2017. Moreover, our models trained using COCO-ReM converge faster and score higher than their larger variants trained using COCO-2017, highlighting the importance of data quality in improving object detectors. With these findings, we advocate using COCO-ReM for future object detection research. Our dataset is available at https://cocorem.xyz" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.579, + 0.733, + 0.594 + ], + "angle": 0, + "content": "Keywords: object detection, segmentation, datasets and benchmarks" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.637, + 0.376, + 0.652 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Rigorous benchmarking of computer vision models relies on task-specific datasets annotated by humans [9, 10, 22, 25]. Understandably, human-annotated data is prone to errors and imperfections due to factors such as ambiguous instructions to annotators, differences in perspective of multiple annotators, etc. Errors in ground-truth training data provide a noisy learning signal for models, resulting in sub-par capabilities. Even more problematic is the presence of errors in ground-truth evaluation data, which can cause a misalignment between evaluation metrics and human judgment when comparing model performances. In this work, we comprehensively study the extent of this issue for the canonical computer vision task of object detection and segmentation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.767, + 0.841 + ], + "angle": 0, + "content": "* Denotes equal contribution. Correspondence to: Karan Desai (kdexd@umich.edu)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.129 + ], + "angle": 0, + "content": "Singh et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.146, + 0.392, + 0.156 + ], + "angle": 0, + "content": "COCO-2017" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.157, + 0.314, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.158, + 0.414, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.158, + 0.496, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.146, + 0.695, + 0.157 + ], + "angle": 0, + "content": "COCO-ReM (ours)" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.158, + 0.599, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.158, + 0.697, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.158, + 0.781, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.267, + 0.321, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.267, + 0.452, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.268, + 0.555, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.268, + 0.643, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.268, + 0.78, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.345, + 0.353, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.345, + 0.504, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.345, + 0.643, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.429, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Fig. 1: COCO with Refined Masks (COCO-ReM). Top: COCO-ReM improves upon the quality of COCO-2017 masks by handling occlusions consistently (dining table) and annotating instances exhaustively (banana). Bottom: Random examples from COCO-ReM showing the mask quality (single category per image, for clarity)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.788, + 0.627 + ], + "angle": 0, + "content": "The Common Objects in Context (COCO) dataset [25] has been the standard benchmark for object detection and segmentation since its inception in 2014. COCO fostered the empirical study of modern object detectors by providing a large set of 123K images annotated with 896K instance masks. The Mask AP (average precision) of COCO detectors has improved dramatically, with \\(\\approx 70\\%\\) relative gain between the pioneering Mask R-CNN [16, 17] from 2017, and recent models published in 2022-23 such as ViTDet [3, 23] and other Transformer-based models [4, 11, 33] like Mask2Former [8] and OneFormer [19]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Despite its popularity, COCO is not without limitations. COCO masks are known to have coarse boundaries [1, 14]. We manually inspect thousands of masks in COCO (2017 version) and observe other imperfections. As shown in Figure 1 (top), COCO-2017 masks do not handle holes and occlusions properly (dining table), and sometimes have non-exhaustively annotated instances (banana). We elaborate on these flaws in Section 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "From an evaluation perspective, imperfections in ground-truth data make the metrics less reliable. COCO-2017 AP would wrongfully penalize models that predict more precise masks than the imperfect ground-truth masks. Moreover, models trained with such imperfect masks may exploit unwanted biases from the training data, e.g., learning to never predict masks with holes. Recently, Kirillov et al. observed that human evaluators consistently rated the masks predicted by their Segment Anything Model (SAM) to be of higher quality than a ViTDet model [23] trained on COCO, whereas the latter scored higher on COCO AP." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.426, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COCO-ReM Benchmark for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "These observations raise an important question - can COCO reliably benchmark the future progress in object detection research? COCO is the de-facto benchmark for object detection - countless research papers report COCO results every year; libraries like Torchvision [29] and Detectron2 [36] readily support object detector development with COCO. Hence, to maintain continuity with prior research and tap into this rich open-source ecosystem, we choose to rectify the errors of COCO and reinforce its utility for future research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.3 + ], + "angle": 0, + "content": "Contributions. We develop COCO-ReM (Refined Masks), a cleaner set of high-quality instance annotations for COCO images. The improved quality over COCO-2017 masks is visually palpable, as seen in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Our annotation pipeline (Section 3) utilizes SAM [21] to refine masks, and uses the LVIS dataset [14] and an ensemble of LVIS-trained models to exhaustively annotate grouped or missing instances. We (the authors) manually verify the entire validation set to provide a strong quality guarantee for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.497 + ], + "angle": 0, + "content": "In Section 4, we use COCO-ReM for benchmarking object detectors. We evaluate fifty object detectors and observe that their COCO-ReM AP produces a different model ranking than COCO-2017 AP. Surprisingly, we observe that query-based models (Mask2Former and OneFormer) score much higher on COCO-ReM than region-based models (ViTDet). Query-based models predict visually sharper masks - AP trends on COCO-ReM accurately reflect this, unlike COCO-2017. We observe that models trained using COCO-ReM converge faster and perform better than those trained using COCO-2017, highlighting the importance of mask quality in improving object detectors." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.519, + 0.62, + 0.538 + ], + "angle": 0, + "content": "2 Background: Revisiting COCO Masks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.597 + ], + "angle": 0, + "content": "The annotation pipeline of a dataset determines the quality of annotations obtained. To design an effective annotation pipeline for COCO-ReM, we review the COCO annotation pipeline by Lin et al. [25] and understand its shortcomings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.629 + ], + "angle": 0, + "content": "COCO annotation pipeline. Lin et al. [25] employed a three-stage pipeline to collect instance mask annotations for COCO images:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.658 + ], + "angle": 0, + "content": "1. Category labeling: Annotators assigned one or more labels to images depending on the visible content, choosing from a fixed set of categories." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "2. Instance spotting: Annotators locate up to ten instances per label (from the previous stage) in the corresponding image, by providing clicks (points)." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.69, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3. Instance segmentation: Finally, annotators clicked around the boundaries of each located instance (from the previous stage), rendering a polygon mask. Lin et al. [25] also took numerous measures to improve annotation quality. The first two stages - category labeling and instance spotting - were performed by eight annotators per image to maximize recall. The instance segmentation stage required annotators to complete a training task before annotating actual instances in the dataset. Next, a different set of crowd workers verified the correctness and quality of annotated masks. Finally, the annotators used a paint-brush tool to mark remaining objects as crowd regions. These extensive measures avoided many scenarios that could have led to low-quality masks." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.63, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.129 + ], + "angle": 0, + "content": "Singh et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.368, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.144, + 0.514, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.144, + 0.651, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.144, + 0.785, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.224, + 0.386, + 0.234 + ], + "angle": 0, + "content": "(a) Holes" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.235, + 0.366, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.235, + 0.515, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.235, + 0.652, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.235, + 0.785, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.315, + 0.688, + 0.327 + ], + "angle": 0, + "content": "(c) Missing annotations and labeling multiple instances as one" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 2: Imperfections in COCO-2017 masks. All masks in general have coarse boundaries. Moreover, they lack holes when required (scissors) and do not handle occlusions consistently (\\(TV\\) with or without figurines). Annotations are sometimes non-exhaustive, either masks are missing (apples and donuts), or multiple instances are grouped into a single mask (banana and dog)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.481 + ], + "angle": 0, + "content": "Imperfections in COCO-2017 masks. To further understand the flaws in COCO annotations, we visualized thousands of COCO-2017 masks and inspected their quality and label correctness. Through this exercise, we observed recurring types of errors in annotations, as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.483, + 0.788, + 0.558 + ], + "angle": 0, + "content": "1. COCO-2017 masks have coarse boundaries. This is a natural consequence of collecting masks as polygon contours, which only coarsely approximate the pixel-precise mask boundaries through piece-wise straight edges connecting points (clicks) supplied by annotators. This shortcoming is well-known and identified by several follow-up works [1, 14, 21]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.558, + 0.788, + 0.618 + ], + "angle": 0, + "content": "2. COCO-2017 masks lack holes when required. The polygon-drawing interface of COCO did not allow annotators to erase the interior regions of polygons. Consequently, certain objects that should naturally contain holes, lack them. Figure 2a shows two examples - the handles of scissors and cup." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.619, + 0.788, + 0.722 + ], + "angle": 0, + "content": "3. COCO-2017 masks do not handle occlusions consistently. Masks of occluded objects are either drawn over the other occluding objects (amodal masks) or around them (modal masks). For example, observe the \\(TV\\) masks in Figure 2b. While both, amodal and modal masks are valid, either one should be consistent in the dataset according to the task definition. We believe this inconsistency is a consequence of inadequate task instructions to annotators as well as the lack of holes in masks." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.724, + 0.788, + 0.814 + ], + "angle": 0, + "content": "4. COCO-2017 masks are often non-exhaustive. This issue occurs in two modes, as shown in Figure 2c - either few instance masks of a category are missing1 (apple, donut), or multiple instances are grouped in a single mask (banana, dog). This is a recurring issue throughout COCO - we believe that certain annotators may not have deemed masks with grouped instances to have a wrong object contour, due to oversight." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.483, + 0.788, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.735, + 0.84 + ], + "angle": 0, + "content": "1 These images have all their missing masks annotated as non-crowd regions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.426, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "COCO-ReM Benchmark for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "5. COCO-2017 contains few near-duplicate masks. We found 410 pairs of masks that overlap with intersection-over-union (IoU) greater than 0.8 in the COCO-2017 validation set (\\(\\approx 2.3\\%\\) of all instances). These pairs often segment the same underlying object with different labels. Examples include armchairs (chair and couch), calves (cow and sheep), and duffel bags (suitcase and backpack). We attribute this issue to the lack of a cross-category deduplication step in the COCO annotation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.787, + 0.314 + ], + "angle": 0, + "content": "All these imperfections hurt the reliability of the benchmarking process. They provide noisy supervision during training and wrongfully penalize correct model predictions during evaluation. Having characterized these imperfections, we make targeted design choices in our COCO-ReM annotation pipeline to rectify them." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.332, + 0.651, + 0.349 + ], + "angle": 0, + "content": "3 COCO-ReM: COCO with Refined Masks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.787, + 0.421 + ], + "angle": 0, + "content": "In this section, we outline our annotation process for COCO-ReM. Our goal is to rectify the imperfections in COCO-2017 annotations as discussed in Section 2 to maintain continuity with prior research. We hope that higher quality ground-truth masks would improve the reliability of the benchmarking process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.439, + 0.431, + 0.455 + ], + "angle": 0, + "content": "3.1 Annotation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.787, + 0.567 + ], + "angle": 0, + "content": "We develop a semi-automatic pipeline - first, we refine the masks using off-the-shelf models and LVIS dataset [14], followed by manual verification of the updated masks. We refine both, the training and validation set of COCO-2017, comprising \\(\\approx 860\\mathrm{K}\\) and \\(\\approx 36\\mathrm{K}\\) instances respectively. We thoroughly verify the validation set to ensure consistency and pixel-precise mask quality for model evaluation. As the training set is much larger, it is only processed through the automatic steps of our pipeline. Our pipeline comprises three stages (Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.659 + ], + "angle": 0, + "content": "Stage 1. Mask boundary refinement. In this stage, we aim to rectify the coarse boundary quality of COCO-2017 masks. A straightforward solution to achieve this is to re-associate all masks through a web interface with paintbrush and eraser tools instead of drawing polygon contours. Moreover, this solution naturally allows including holes in masks where necessary. While accurate, this solution would sharply increase the mask annotation time and cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.75 + ], + "angle": 0, + "content": "Along a different trend, prior works have explored model-assisted interactive segmentation as a feasible solution for collecting mask annotations with smooth boundaries [1, 14]. The key idea is to let an object detector predict a coarse initial mask and have the annotators refine it by prompting corrective mouse clicks. Inspired by the accuracy and practical feasibility of this approach, we use the Segment Anything Model (SAM [21]) to refine the COCO-2017 masks." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "- Refining masks using SAM. SAM is a promptable segmentation model that predicts high-quality masks conditioned on user-provided point and box prompts. This promptable functionality makes SAM suitable for our use case. We prompted SAM using various combinations of prompts sampled using COCO-2017 masks and obtained its predictions as the refined masks. Refer to the appendix for details about this prompting methodology." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.129 + ], + "angle": 0, + "content": "Singh et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.298, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 3: COCO-ReM annotation pipeline. We develop a semi-automatic annotation pipeline comprising three stages to rectify the imperfections in COCO-2017 masks." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.352, + 0.788, + 0.472 + ], + "angle": 0, + "content": "- Manual verification of mask quality. We visualized and manually inspected every mask created in the previous process. Through this inspection, we identified nearly 900 masks that have low quality or do not accurately cover the underlying object. These masks represent only \\(2.4\\%\\) of 36K instances in the COCO-2017 validation set, greatly reducing our manual effort and showing the efficacy of our automated process. To correct these masks, we deployed a web browser demo of SAM and provided point prompts iteratively until obtaining a precise and accurate mask." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.474, + 0.787, + 0.505 + ], + "angle": 0, + "content": "Through this stage, we obtain masks with more precise boundaries and better handling of occlusions, as illustrated in Figure 3 (left)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.816, + 0.613 + ], + "angle": 0, + "content": "Stage 2. Exhaustive instance annotation. This stage addresses non-exhaustively annotated instances in COCO-2017 (Figure 2c). We note that manually re-annotating every image is the most reliable solution to ensure exhaustive instance coverage. However, the manual effort would not only require extensive instructions and verification but would also be highly cost-ineffective and may lead to redundant annotations. To mitigate these issues, we automatically import instances from the LVIS [14] dataset as a strong starting point." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.614, + 0.788, + 0.72 + ], + "angle": 0, + "content": "- Importing instances from LVIS. The LVIS dataset comprises 1.2M high-quality instance annotations across 1203 categories for images in COCO, including all 80 COCO categories. Gupta et al. [14] took firm measures to ensure exhaustive annotations. Every image is labeled with a set of positive LVIS categories whose instances are exhaustively labeled, and a set of negative LVIS categories that are verifiably absent. The presence (or absence) of the remaining LVIS categories is unverified for the image." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.72, + 0.788, + 0.81 + ], + "angle": 0, + "content": "For the 80 COCO categories, LVIS contains \\(\\approx 20\\mathrm{K}\\) instances for images in the COCO-2017 validation set. To import these instances, we checked whether LVIS contained more instances of an (image, category) pair than COCO-2017. If so, we replaced all COCO-2017 instances of this pair with those from LVIS. Nearly \\(9\\%\\) of (image, category) pairs in LVIS are flagged to have nonexhaustive instance annotations [14] - we ignored these while importing." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "- Discovering instances using LVIS-trained models. Instances of a given (image, category) pair of COCO-2017 may be missing in LVIS if the category" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.426, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "COCO-ReM Benchmark for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.25 + ], + "angle": 0, + "content": "
DatasetSplitInstance SourceTotal
COCO-2017LVISLVIS models
COCO-2017val36,781--36,781
COCO-ReMval33,4986,135105640,689
COCO-2017train860,001--860,001
COCO-ReMtrain738,353354,674-1,093,027
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.251, + 0.784, + 0.293 + ], + "angle": 0, + "content": "Table 1: Instance statistics in COCO-ReM. We list the exact number of annotations obtained from different sources for both train and val splits of COCO-ReM. All masks undergo refinement through SAM before being included in COCO-ReM." + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.313, + 0.785, + 0.433 + ], + "angle": 0, + "content": "is unverified for the corresponding image. Instances of this pair may be nonexhaustive in COCO-2017, and hence remain unresolved after the previous automatic step. To resolve this, we use publicly available object detectors trained using LVIS. On the validation set, we ensemble mask predictions using the official checkpoints of three best-performing ViTDet models [23]. As model predictions are not as accurate as human-annotated data, we manually inspect \\(\\approx 2500\\) masks predicted with high confidence and select 1056 masks. We refine these masks (Stage 1) and add them to COCO-ReM." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.435, + 0.784, + 0.464 + ], + "angle": 0, + "content": "This stage augments COCO-ReM with additional instances from external sources and includes more exhaustive annotations, see example in Figure 3 (middle)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.468, + 0.785, + 0.572 + ], + "angle": 0, + "content": "Stage 3. Correction of labeling errors. This is the final stage of our pipeline, performed manually, only for the validation set. Recall from Section 2 that we found 410 near-duplicate mask pairs in the COCO-2017 validation set. We manually inspected these and retained one mask per pair with the most accurate label, such as chair for armchairs and car for SUVs (Figure 3 (right)). We also observed that \\(\\approx 100\\) grouped instances were not resolved through Stage 2. We annotated these masks as crowd to accurately reflect their characteristics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.436, + 0.61 + ], + "angle": 0, + "content": "3.2 Mask Characteristics" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.626, + 0.785, + 0.746 + ], + "angle": 0, + "content": "Instance statistics. Our annotation pipeline imports instances from external sources - LVIS dataset and LVIS-trained models - to annotate instances exhaustively. As a result, every instance in COCO-ReM traces to one of these sources or COCO-2017. As shown in Table 1, COCO-ReM validation set extends COCO-2017 by including \\(\\approx 6\\mathrm{K}\\) masks from LVIS [14] and \\(\\approx 1\\mathrm{K}\\) masks from LVISTrained models [3, 23]. Since these masks replace a subset of COCO-2017 masks (see Stage 2), we retain \\(\\approx 33.5\\mathrm{K}\\) masks from COCO-2017 (out of \\(36\\mathrm{K}\\) masks). Overall, COCO-ReM validation set has \\(\\approx 4\\mathrm{K}\\) more instances than COCO-2017." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.747, + 0.785, + 0.807 + ], + "angle": 0, + "content": "On the other hand, the training set of COCO-ReM comprises a total of 1.09M high-quality masks, as compared to 860K masks in the COCO-2017 training set. The difference in instance counts between COCO-ReM and COCO-2017 sharply highlights the extent of non-exhaustively annotated instances in COCO-2017." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Mask IoU. Figure 4 (left) shows the distribution of IoU between the (refined) masks of COCO-ReM (validation set) and their corresponding source masks in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Singh et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.615, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.149, + 0.78, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.277, + 0.788, + 0.32 + ], + "angle": 0, + "content": "Fig. 4: Left: Mask IoU distribution. Distribution of IoU between masks in COCO-ReM validation set and their source dataset. Right: Masks with holes. Nearly 2000 masks in COCO-ReM validation set have holes, that were missing in COCO-2017." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.784, + 0.403 + ], + "angle": 0, + "content": "COCO-2017 and LVIS datasets. The median IoU is slightly less than one, indicating that most masks have a proper coarse shape, but lack precise boundaries. This median estimate matches the findings of Gupta et al. [14], albeit observed from a smaller sample of 100 masks collected from human annotators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.785, + 0.452 + ], + "angle": 0, + "content": "Holes in masks. Figure 4 (right) shows that \\(\\approx 2000\\) masks in COCO-ReM validation set have at least one hole. By design, COCO-2017 lacks these holes. See the appendix for qualitative examples and additional analysis per category." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.474, + 0.377, + 0.492 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.643 + ], + "angle": 0, + "content": "In our experiments, we aim to demonstrate the utility of COCO-ReM for reliable benchmarking of object detection research. Recall Section 1, we hypothesize that the imperfections in COCO-2017 may cause two issues: (1) During evaluation, models may be undesirably penalized despite predicting correct masks, yielding counter-intuitive modeling observations, and (2) During training, models receive noisy supervision and/or may learn unwanted biases (e.g., never predict masks with holes), resulting in sub-par capabilities. We claim that high-quality masks of COCO-ReM can alleviate these issues. To this end, we use COCO-ReM to evaluate and train object detectors, and compare trends with COCO-2017." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.664, + 0.509, + 0.68 + ], + "angle": 0, + "content": "4.1 Evaluation using COCO-ReM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.764 + ], + "angle": 0, + "content": "In this section, we evaluate fifty publicly available COCO object detectors using the COCO-ReM validation set. We focus on the validation set as the test set annotations of COCO-2017 are not publicly available. We cover a wide range of models from existing literature. All models can be categorized as either region-based or query-based:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "- Region-based models, following Mask R-CNN [17]. These models first predict labeled boxes, followed by binary masks inside those boxes. We also include cascade models, following Cascade R-CNN [3]. These models extend the region-based design by iteratively refining their predicted bounding boxes before making the final mask prediction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.426, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "COCO-ReM Benchmark for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.782, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.788, + 0.335 + ], + "angle": 0, + "content": "Fig. 5: Evaluating fifty object detectors using COCO-ReM. All models score higher on COCO-ReM than COCO-2017. Query-based models (\\(\\star\\)) score much higher than region-based models on COCO-ReM, yielding opposite trends than COCO-2017." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.359, + 0.784, + 0.435 + ], + "angle": 0, + "content": "- **Query-based models**, inspired by the Detection Transformer design [4]. Instead of predicting masks within image regions, these models start with a set of image-independent query vectors, then decode labeled masks through a Transformer [33] by iteratively attending to image embeddings. Among query-based models, we consider Mask2Former [8] and OneFormer [19]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.784, + 0.541 + ], + "angle": 0, + "content": "We ensure diverse image backbones in our model collection. The architecture and capacity of the backbone significantly impact object detector performance. Thus, we include convolutional backbones (ResNet [16], RegNet [31], ConvNeXt [27], InternImage [35]), hierarchical vision transformers (Swin [26], MViTv2 [24], DiNAT [15]), and plain vision transformers [11]. We include backbones of different sizes for a single model, e.g., for Cascade ViTDet [23], we include ViT-B (141M parameters), ViT-L (361M parameters), ViT-H (692M parameters)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.784, + 0.572 + ], + "angle": 0, + "content": "We believe that our model collection is sufficiently diverse to draw meaningful conclusions from comparing trends between COCO-ReM and COCO-2017" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.785, + 0.652 + ], + "angle": 0, + "content": "Testing protocol. We compute the mask Average Precision (AP) of all models using the official COCO evaluation API in Detectron2 [36]. For simplicity, we adopt a common testing protocol for all models - we resize the longest edge of input images to 1024 pixels (without changing the aspect ratio), and limit the number of detections per image to 100." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.784, + 0.73 + ], + "angle": 0, + "content": "Result 1. All models score higher on COCO-ReM. In Figure 5, we plot the mask AP trends for all fifty object detectors sorted by their COCO-2017 AP. All fifty models achieve higher AP on COCO-ReM despite being trained using only COCO-2017 masks. This result validates our hypothesis that the imperfections of COCO-2017 may have wrongfully penalized all these models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.784, + 0.821 + ], + "angle": 0, + "content": "The difference between COCO-ReM AP and COCO-2017 AP is larger with recent high-performing models - the current state-of-the-art model (OneFormer InternImage-H) scores \\(+7.7\\) points higher AP on COCO-ReM than COCO-2017. This increasing gap in AP suggests that the imperfections in ground-truth have worsened the issue of over-penalization and made COCO-2017 AP less adept at fine-grained assessment and ranking of recent models." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.779, + 0.84 + ], + "angle": 0, + "content": "Result 2. Query-based detectors score much higher on COCO-ReM." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Singh et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.144, + 0.418, + 0.155 + ], + "angle": 0, + "content": "Cascade ViTDet-L" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.155, + 0.358, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.226, + 0.358, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.155, + 0.497, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.144, + 0.704, + 0.155 + ], + "angle": 0, + "content": "Mask2Former Swin-L" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.155, + 0.643, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.226, + 0.643, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.155, + 0.782, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.299, + 0.787, + 0.356 + ], + "angle": 0, + "content": "Fig. 6: Mask predictions from models with opposite rank order on COCO-ReM and COCO-2017. Qualitative examples of ViTDet (region-based model) and Mask2Former (query-based model) show that the latter predicts visually sharper masks. COCO-ReM AP correctly ranks Mask2Former higher, unlike COCO-2017 AP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.785, + 0.476 + ], + "angle": 0, + "content": "Interestingly, we observe that the trend in AP values calculated using COCO-ReM does not align perfectly with COCO-2017. We observe frequent flips in ranks of several model pairs on COCO-ReM. While the model rankings do not transfer between benchmarks perfectly, we observe a consistent pattern in model pairs with flipped rankings. Notice the star markers \\((\\star)\\) in Figure 5 - the flips are attributed to significantly higher scores of query-based models (Mask2Former [8] and OneFormer [19]) than region-based models on COCO-ReM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.787, + 0.612 + ], + "angle": 0, + "content": "To further understand the cause of this performance trend, we observe the predicted masks of two models having opposite rank order according to COCO-ReM and COCO-2017 AP: Cascade ViTDet-L [23] and Mask2Former Swin-L [8]. The latter is a query-based model. Figure 6 shows some predicted masks by both of these models. Through a cursory glance, we can observe that Mask2Former predicts visually sharper masks whereas masks of ViTDet have coarse boundaries. This human judgment aligns better with COCO-ReM AP, which correctly rates Mask2Former higher. On the other hand, COCO-2017 AP rates ViTDet higher, yielding a counter-intuitive modeling observation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.673 + ], + "angle": 0, + "content": "Guided by these trends, we reaffirm our claim that benchmarking newer models using the COCO-2017 may be misleading from a research perspective, and the community should adopt our COCO-ReM annotations moving forward. We include numerical results of Figure 5 in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.562, + 0.709 + ], + "angle": 0, + "content": "4.2 Understanding the Difference in AP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.808 + ], + "angle": 0, + "content": "Recall Section 4.1 (Result 1) - all models score higher average precision (AP) on COCO-ReM, as compared to COCO-2017. What explains the difference in \\( AP \\)? To understand the reasons underlying this difference, we observe the components of COCO AP for a few high-performing models from Section 4.1. For this analysis, we consider two Cascade ViTDet models [23] using ViT-L/H backbones [11], and two Mask2Former models [8] using Swin-B/L backbones [26]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "AP per IoU threshold. The calculation of COCO AP involves an average of per-category AP scores of 80 COCO categories, each computed at ten different" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.426, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "COCO-ReM Benchmark for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.149, + 0.772, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.289, + 0.787, + 0.359 + ], + "angle": 0, + "content": "Fig. 7: AP per IoU threshold. While models score similarly at IoU = 0.5 on COCO-ReM and COCO-2017, they score much higher on COCO-ReM at higher IoU thresholds (0.75 and beyond), contributing to the overall difference in AP. Due to the improved mask quality of COCO-ReM, the AP at high IoU thresholds becomes more sensitive and enables better assessment of model capabilities." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.368, + 0.773, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Fig. 8: AP per category. Models score higher AP on COCO-ReM for 69 out of 80 categories. The masks of the remaining 11 categories (left) in COCO-2017 do not handle occlusions consistently (dining table, bed), lack holes (scissors, donut), or have grouped instances (banana, apple). Models learn these biases from COCO-2017 training masks and score lower COCO-ReM AP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.671 + ], + "angle": 0, + "content": "IoU thresholds, \\( t \\in \\{0.5, 0.55, \\dots, 0.9, 0.95\\} \\). In Figure 7, we show components of AP at different IoU thresholds, averaged for the four models we use. On average, models score similarly at IoU = 0.5 for COCO-ReM and COCO-2017. However, COCO-ReM AP is much larger for IoU = 0.75 and beyond. Hence, higher scores on COCO-ReM can be attributed to AP at higher IoU thresholds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.808 + ], + "angle": 0, + "content": "Intuitively, AP at IoU = t rewards model predictions that overlap with a ground-truth mask of the same category with IoU ≥ t. AP at IoU = 0.5 requires a very coarse overlap between the predicted and ground-truth masks, whereas AP at high IoU thresholds have stricter overlap requirements. Since IoU is symmetric, imprecise boundaries of ground-truth masks (COCO-2017) can wrongfully penalize valid predictions and lead to lower AP scores. COCO-ReM fixes this issue, as evident with higher AP at high IoU thresholds. This analysis hints that precise masks of COCO-ReM improve the sensitivity of AP and enable fine-grained assessment of high-performing object detectors." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "AP per category. Next, we calculate AP for every COCO category by averaging across all IoU thresholds. Figure 8 shows the difference in COCO-ReM" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Singh et al." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.2 + ], + "angle": 0, + "content": "
COCO evaluation set ⇒2017ReM-S1ReM-S2aReM (final)
AP49.657.355.655.5
Number of instances36,78136,78141,16840,689
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.201, + 0.788, + 0.258 + ], + "angle": 0, + "content": "Table 2: COCO-ReM ablations. AP of four high-performing models (Cascade ViTDet-B/L and Mask2Former Swin-B/L) using intermediate versions from our COCO-ReM annotation pipeline. Trends suggest that the coarse boundary quality of COCO-2017 masks was a major source of noise in the benchmarking process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.786, + 0.367 + ], + "angle": 0, + "content": "AP and COCO-2017 AP per category, averaged across the same four models. Models score higher on COCO-ReM for 69 out of 80 categories, which is reflected in the higher overall COCO-ReM AP. Significant improvements are observed in categories with generally consistent appearance (carrot and orange), and fine details (fork). Intuitively speaking, the larger difference in per-category AP can be attributed to the better quality of ground-truth masks in COCO-ReM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.504 + ], + "angle": 0, + "content": "For 11 categories, models score lower on COCO-ReM than COCO-2017. These categories either have holes in their natural appearance or are often heavily occluded (bowl, dining table, scissors). Masks in COCO-ReM include holes and handle occlusions more consistently than in COCO-2017. Hence, lower per-category AP with COCO-ReM implies that the imperfections of COCO-2017 training masks impart these undesirable biases to trained models – upon checking the model predictions, we observed that they often include occluding objects for dining table and lack holes for scissors. Such predictions are not rewarded by COCO-ReM, hence incur a drop in the per-category AP." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.345, + 0.536 + ], + "angle": 0, + "content": "4.3 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.787, + 0.603 + ], + "angle": 0, + "content": "In this section, we conduct a simple ablation study to observe the contribution of different stages in our annotation pipeline. For the sake of simplicity, we select the same four models as chosen in Section 4.2. We evaluate these models using two versions of COCO-ReM validation set:" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.604, + 0.785, + 0.649 + ], + "angle": 0, + "content": "1. ReM-S1: This version is obtained after the Stage 1 (mask boundary refinement) in our annotation pipeline. It comprises the same number of instances as COCO-2017, each with an improved boundary quality." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.649, + 0.785, + 0.694 + ], + "angle": 0, + "content": "2. ReM-S2a: This version is obtained after importing instances from the LVIS dataset – the automatic part of Stage 2 (exhaustive instance annotation). We exclude the manual part of Stage 2 from this ablation for simplicity." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.604, + 0.785, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.787, + 0.772 + ], + "angle": 0, + "content": "Results are shown in Table 2. The large difference between COCO-ReM AP and COCO-2017 AP appears right after Stage 1, indicating that the coarse mask boundaries are a major source of noise that hurts the reliability of COCO-2017 AP. Subsequent stages incur a negligible drop in AP, likely due to larger of the validation set, as COCO AP is known to favor smaller evaluation sets [14]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.485, + 0.804 + ], + "angle": 0, + "content": "4.4 Training with COCO-ReM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our evaluation shows COCO-ReM as a useful benchmark for accurate modeling conclusions. In this section, we train strong baseline models using COCO-ReM." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.426, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COCO-ReM Benchmark for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.505, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.147, + 0.779, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Fig.9: Training with COCO-ReM. A ViTDet model trained using COCO-ReM training set converges faster and performs better than its counterpart trained using COCO-2017, and closely matches a \\(3 \\times\\) larger ViTDet-L trained using COCO-2017." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.789, + 0.468 + ], + "angle": 0, + "content": "Implementation details. We train ViTDet models [23] using their official open-source implementation [36], leaving all hyperparameters unchanged. We train two models - Mask R-CNN ViTDet-B (111M parameters) and Cascade ViTDet-B (141M parameters). Li et al. [23] trained this model for 100 COCO epochs across \\(64 \\times\\) A100 GPUs (batch size of 64). To manage these costs, we implement gradient accumulation for every 8 iterations to emulate a batch size of 64 using our \\(8 \\times\\) A40 GPUs, resulting in an 11 day-long training schedule. We could only afford to train the base (ViT-B) models due to limited resources. The gorund truth annotations for test set is inaccessible so, we use the validation set solely for evaluation purposes for fair assessment of models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.79, + 0.579 + ], + "angle": 0, + "content": "Results. Figure 9 shows that our ViTDet-B models trained using COCO-ReM significantly outperform the same ViTDet-B models trained using COCO-2017. Higher quality of training data enables faster convergence. Moreover, our ViTDet-B models are competitive with \\(3 \\times\\) larger ViTDet-L models trained using COCO-2017, indicating better parameter efficiency. While numerous prior works improve object detectors through modeling innovations, our results highlight the critical role of data quality in improving object detectors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.388, + 0.621 + ], + "angle": 0, + "content": "5 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Object detection and image segmentation. Visual perception is fundamental for any vision system to understand the world around it. At the core of a visual perception system is the task of image segmentation that aims at recognizing the entities (objects) in a given scene and densely labeling every pixel constituting the corresponding object. Image segmentation can broadly be divided into three sub-tasks: semantic, instance, and panoptic segmentation. On the one hand, semantic segmentation [5, 6, 28] targets obtaining a single amorphous binary mask covering all pixels for the corresponding category. On the other hand, instance segmentation [3, 18] involves detecting pixels for only foreground \"thing\" regions with clear boundaries and differentiating among separate instances for a category. Panoptic segmentation [7, 20, 34] is a combination of the former two tasks targeted at predicting distinct binary masks for \"thing\" regions and a single amorphous mask for the \"stuff\" regions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Singh et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "In this work, we particularly look at the task of instance segmentation and contribute high-quality mask annotations to benchmark further progress." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.184, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Image segmentation benchmarks. In the last decade, considerable advancements have been observed in the classical vision task of image segmentation. A part of this progress is owed to the introduction of numerous datasets for benchmarking image segmentation models. In one of the earliest efforts, Everingham et al. [12] introduced the PASCAL-VOC-2012 challenge to recognize 20 classes in about 10K images. More challenging datasets were introduced in the following years, with ADE20K [10] covering 150 classes. Lin et al. [25] introduced the concept of \"stuff\" and \"thing\" classes in their COCO dataset with a total of 171 classes: 80 \"thing\" and 91 \"stuff\". The Cityscapes [9], Mapillary-Vistas [30], and KITTI [13] datasets introduced image segmentation datasets focused on autonomous driving applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Gupta et al. [14] introduced the LVIS dataset with exhaustive annotations covering 1203 categories for COCO images. OpenImages [1] introduced 2.8M mask annotations covering 350 classes. Recently, the SA-1B dataset [21] introduced over 1.1B fine-grained, class-agnostic mask annotations for over 11M images. Despite various datasets, COCO has remained the de-facto benchmark for object detectors in instance segmentation. In this work, we examined the COCO instance annotations and proposed a new benchmark with superior mask quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Re-assessing datasets for benchmarking. Human-annotated datasets are always prone to errors and imperfections, as no human is perfect. Some prior works have shared our motivation and conducted in-depth re-assessment studies, the closest to ours are works that re-assess the ImageNet dataset [2, 32]. Concurrent to our work, Zimmermann et al. [37] also re-assess COCO annotations, however they opt for recollecting polygon masks from human annotators." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.36, + 0.618 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.746 + ], + "angle": 0, + "content": "In this work, we revisit the COCO-2017 mask annotations and find major issues involving, but not limited to their boundary quality and exhaustiveness. These issues raise significant concerns about the dependability of COCO-2017 to benchmark recent and future object detectors accurately. Consequently, we introduce our COCO-ReM benchmark consisting of high-quality mask annotations obtained using an effective semi-automatic annotation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "On comparing the mask AP trends for existing object detectors on COCO-ReM and COCO-2017, we notice significant changes in model rankings, notably contradicting the conclusions drawn through COCO-2017. Moreover, training using COCO-ReM shows that paying attention to mask quality is crucial to advancing the capabilities of object detectors. We hope that COCO-ReM will aid the future research in object detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.427, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COCO-ReM Benchmark for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.269 + ], + "angle": 0, + "content": "We thank Mohamed El Banani, Agrim Gupta, and Nilesh Kulkarni for helpful discussions and feedback. We thank Palak Jagtap for her help during manual verification of masks. We thank all the authors of the prior works for publicly releasing the pre-trained model checkpoints, which helped us maintain empirical comprehensiveness in our evaluation study. We thank Vision and Language Group, IIT Roorkee for being a constant support system throughout the process." + } + ], + [ + { + "type": "title", + "bbox": [ + 0.427, + 0.142, + 0.577, + 0.164 + ], + "angle": 0, + "content": "Bibliography" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.211, + 0.787, + 0.272 + ], + "angle": 0, + "content": "[1] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 4, 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.787, + 0.317 + ], + "angle": 0, + "content": "[2] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aïron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.317, + 0.787, + 0.362 + ], + "angle": 0, + "content": "[3] Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN: Delving into high quality object detection. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.787, + 0.422 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.787, + 0.467 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In ICLR, 2015. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.787, + 0.525 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In TPAMI, 2017. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.526, + 0.787, + 0.6 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.787, + 0.661 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.787, + 0.735 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.736, + 0.787, + 0.81 + ], + "angle": 0, + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. Semantic understanding of scenes through the ade20k dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer," + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.211, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.426, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COCO-ReM Benchmark for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.249, + 0.146, + 0.786, + 0.207 + ], + "angle": 0, + "content": "Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.208, + 0.787, + 0.268 + ], + "angle": 0, + "content": "[12] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2012. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.269, + 0.787, + 0.313 + ], + "angle": 0, + "content": "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.314, + 0.787, + 0.373 + ], + "angle": 0, + "content": "[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3, 4, 5, 6, 7, 8, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.374, + 0.787, + 0.403 + ], + "angle": 0, + "content": "[15] Ali Hassani and Humphrey Shi. Dilated neighborhood attention transformer. arXiv:2209.15001, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.404, + 0.787, + 0.449 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.449, + 0.787, + 0.494 + ], + "angle": 0, + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dolkar, and Ross Girshick. Mask R-CNN. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.787, + 0.539 + ], + "angle": 0, + "content": "[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.539, + 0.787, + 0.599 + ], + "angle": 0, + "content": "[19] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.6, + 0.787, + 0.644 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.645, + 0.787, + 0.718 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, WanYen Lo, Piotr Dolkar, and Ross Girshick. Segment Anything. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.72, + 0.787, + 0.765 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.766, + 0.787, + 0.81 + ], + "angle": 0, + "content": "[23] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 7, 9, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "[24] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.146, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.129 + ], + "angle": 0, + "content": "Singh et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.249, + 0.146, + 0.788, + 0.177 + ], + "angle": 0, + "content": "vision transformers for classification and detection. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.236 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 1, 2, 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.236, + 0.787, + 0.295 + ], + "angle": 0, + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2021. 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.295, + 0.787, + 0.34 + ], + "angle": 0, + "content": "[27] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.34, + 0.787, + 0.383 + ], + "angle": 0, + "content": "[28] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.383, + 0.787, + 0.412 + ], + "angle": 0, + "content": "[29] TorchVision maintainers and contributors. TorchVision: PyTorch's Computer Vision library. https://github.com/pytorch/vision, 2016.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.787, + 0.471 + ], + "angle": 0, + "content": "[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.471, + 0.787, + 0.515 + ], + "angle": 0, + "content": "[31] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.515, + 0.787, + 0.559 + ], + "angle": 0, + "content": "[32] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In Proceedings of the International Conference on Machine Learning (ICML), 2019. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.559, + 0.787, + 0.618 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.618, + 0.787, + 0.678 + ], + "angle": 0, + "content": "[34] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.787, + 0.751 + ], + "angle": 0, + "content": "[35] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.752, + 0.787, + 0.796 + ], + "angle": 0, + "content": "[36] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. 3, 9, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "[37] Eric Zimmermann, Justin Szeto, Jerome Pasquero, and Frederic Ratle. Benchmarking a benchmark: How reliable is ms-coco? arXiv preprint arXiv:2311.02709, 2023. 14" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.146, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/64864c4f-d7d3-4bce-8017-e14998347c20_origin.pdf b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/64864c4f-d7d3-4bce-8017-e14998347c20_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..03eaaeea5c0b9e9f55ffb783e9d118aa890cd096 --- /dev/null +++ b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/64864c4f-d7d3-4bce-8017-e14998347c20_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d08d3b735a9b02cb7dbc2cb0aa1f813ad510ba1cc41ed1fdd27af686ac645b03 +size 2660878 diff --git a/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/full.md b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/full.md new file mode 100644 index 0000000000000000000000000000000000000000..03d3b08967a478fe5260c455fecd14805d7fdafb --- /dev/null +++ b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/full.md @@ -0,0 +1,318 @@ +# Benchmarking Object Detectors with COCO: A New Path Forward + +Shweta Singh1* Aayan Yadav1* Jitesh Jain2 + +Humphrey Shi² Justin Johnson³ Karan Desai³ + +$^{1}$ IIT Roorkee $^{2}$ Georgia Tech $^{3}$ University of Michigan + +Abstract. The Common Objects in Context (COCO) dataset has been instrumental in benchmarking object detectors over the past decade. Like every dataset, COCO contains subtle errors and imperfections stemming from its annotation procedure. With the advent of high-performing models, we ask whether these errors of COCO are hindering its utility in reliably benchmarking further progress. In search for an answer, we inspect thousands of masks from COCO (2017 version) and uncover different types of errors such as imprecise mask boundaries, non-exhaustively annotated instances, and mislabeled masks. Due to the prevalence of COCO, we choose to correct these errors to maintain continuity with prior research. We develop COCO-ReM (Refined Masks), a cleaner set of annotations with visibly better mask quality than COCO-2017. We evaluate fifty object detectors and find that models that predict visually sharper masks score higher on COCO-ReM, affirming that they were being incorrectly penalized due to errors in COCO-2017. Moreover, our models trained using COCO-ReM converge faster and score higher than their larger variants trained using COCO-2017, highlighting the importance of data quality in improving object detectors. With these findings, we advocate using COCO-ReM for future object detection research. Our dataset is available at https://cocorem.xyz + +Keywords: object detection, segmentation, datasets and benchmarks + +# 1 Introduction + +Rigorous benchmarking of computer vision models relies on task-specific datasets annotated by humans [9, 10, 22, 25]. Understandably, human-annotated data is prone to errors and imperfections due to factors such as ambiguous instructions to annotators, differences in perspective of multiple annotators, etc. Errors in ground-truth training data provide a noisy learning signal for models, resulting in sub-par capabilities. Even more problematic is the presence of errors in ground-truth evaluation data, which can cause a misalignment between evaluation metrics and human judgment when comparing model performances. In this work, we comprehensively study the extent of this issue for the canonical computer vision task of object detection and segmentation. + +![](images/574a8021e3ae671cea135c567d69b3a6cba6a871bc830aa174b6e3b4bb0b0354.jpg) +COCO-2017 + +![](images/71d3b68b4be4dc387a967ec666f23ce8385c698499951c0195fb9dc493d0e876.jpg) + +![](images/a3a57ce1ee21d5fada8d68af59c2b8623efe7943f968c74ac516902a666f9c14.jpg) + +![](images/4ca0c32cfc84361a6ab439be6b5dd0c959d89393dec0db9ba2b1e297fac71795.jpg) +COCO-ReM (ours) + +![](images/2aad59070ba17f7fd23e798c7f2bb474602f38c834f7f1d4dd2ac0c445007cb9.jpg) + +![](images/7061296a8b7bb6524378490b51226fae9d518b468d47f46d5bb5236ff8ddf1c6.jpg) + +![](images/3e9526155ddb54d2a6c92a17aebf7c245012d8582485b0f34201d9827a3e4c5a.jpg) + +![](images/c878476568b68c6f5b4e2ebf2705e58c56fa4ba9c52e529e960ed16d305ff941.jpg) + +![](images/692a973549fc58d9a355daa99d7f32cc1677912ec2eb55b39d93428b89e92d2f.jpg) + +![](images/b201818a6f93b8b07e35907fe2c875e13fb1035ce4d442bdbb20316d45bb4cc2.jpg) + +![](images/ae1207e334a76ec80287f2438435d8ec7be7e0f3aa08f91c9d675225e7123be1.jpg) + +![](images/26062413e2c70ce9e4237df4b68488f9df7c2a8652e905a920b9f22fbd4ef011.jpg) +Fig. 1: COCO with Refined Masks (COCO-ReM). Top: COCO-ReM improves upon the quality of COCO-2017 masks by handling occlusions consistently (dining table) and annotating instances exhaustively (banana). Bottom: Random examples from COCO-ReM showing the mask quality (single category per image, for clarity). + +![](images/52fff3971e82fde5b1533a07b70bb4d8c15a6d1b807f2a89a8dea9e602861892.jpg) + +![](images/6f303e59aaef9a8013714213d3162bf6b38387ffe95955f912d278192195feba.jpg) + +The Common Objects in Context (COCO) dataset [25] has been the standard benchmark for object detection and segmentation since its inception in 2014. COCO fostered the empirical study of modern object detectors by providing a large set of 123K images annotated with 896K instance masks. The Mask AP (average precision) of COCO detectors has improved dramatically, with $\approx 70\%$ relative gain between the pioneering Mask R-CNN [16, 17] from 2017, and recent models published in 2022-23 such as ViTDet [3, 23] and other Transformer-based models [4, 11, 33] like Mask2Former [8] and OneFormer [19]. + +Despite its popularity, COCO is not without limitations. COCO masks are known to have coarse boundaries [1, 14]. We manually inspect thousands of masks in COCO (2017 version) and observe other imperfections. As shown in Figure 1 (top), COCO-2017 masks do not handle holes and occlusions properly (dining table), and sometimes have non-exhaustively annotated instances (banana). We elaborate on these flaws in Section 2. + +From an evaluation perspective, imperfections in ground-truth data make the metrics less reliable. COCO-2017 AP would wrongfully penalize models that predict more precise masks than the imperfect ground-truth masks. Moreover, models trained with such imperfect masks may exploit unwanted biases from the training data, e.g., learning to never predict masks with holes. Recently, Kirillov et al. observed that human evaluators consistently rated the masks predicted by their Segment Anything Model (SAM) to be of higher quality than a ViTDet model [23] trained on COCO, whereas the latter scored higher on COCO AP. + +These observations raise an important question - can COCO reliably benchmark the future progress in object detection research? COCO is the de-facto benchmark for object detection - countless research papers report COCO results every year; libraries like Torchvision [29] and Detectron2 [36] readily support object detector development with COCO. Hence, to maintain continuity with prior research and tap into this rich open-source ecosystem, we choose to rectify the errors of COCO and reinforce its utility for future research. + +Contributions. We develop COCO-ReM (Refined Masks), a cleaner set of high-quality instance annotations for COCO images. The improved quality over COCO-2017 masks is visually palpable, as seen in Figure 1. + +Our annotation pipeline (Section 3) utilizes SAM [21] to refine masks, and uses the LVIS dataset [14] and an ensemble of LVIS-trained models to exhaustively annotate grouped or missing instances. We (the authors) manually verify the entire validation set to provide a strong quality guarantee for evaluation. + +In Section 4, we use COCO-ReM for benchmarking object detectors. We evaluate fifty object detectors and observe that their COCO-ReM AP produces a different model ranking than COCO-2017 AP. Surprisingly, we observe that query-based models (Mask2Former and OneFormer) score much higher on COCO-ReM than region-based models (ViTDet). Query-based models predict visually sharper masks - AP trends on COCO-ReM accurately reflect this, unlike COCO-2017. We observe that models trained using COCO-ReM converge faster and perform better than those trained using COCO-2017, highlighting the importance of mask quality in improving object detectors. + +# 2 Background: Revisiting COCO Masks + +The annotation pipeline of a dataset determines the quality of annotations obtained. To design an effective annotation pipeline for COCO-ReM, we review the COCO annotation pipeline by Lin et al. [25] and understand its shortcomings. + +COCO annotation pipeline. Lin et al. [25] employed a three-stage pipeline to collect instance mask annotations for COCO images: + +1. Category labeling: Annotators assigned one or more labels to images depending on the visible content, choosing from a fixed set of categories. +2. Instance spotting: Annotators locate up to ten instances per label (from the previous stage) in the corresponding image, by providing clicks (points). +3. Instance segmentation: Finally, annotators clicked around the boundaries of each located instance (from the previous stage), rendering a polygon mask. Lin et al. [25] also took numerous measures to improve annotation quality. The first two stages - category labeling and instance spotting - were performed by eight annotators per image to maximize recall. The instance segmentation stage required annotators to complete a training task before annotating actual instances in the dataset. Next, a different set of crowd workers verified the correctness and quality of annotated masks. Finally, the annotators used a paint-brush tool to mark remaining objects as crowd regions. These extensive measures avoided many scenarios that could have led to low-quality masks. + +![](images/a85a2a36df96677fe3711cd2e8fc3dfe84022b5ae482a9f270c3b842e2068111.jpg) +(a) Holes + +![](images/164ecf5f97458ae414f1efc41f0ec0f7b8261091beab2607dcd0bdcfeb0f17de.jpg) + +![](images/f401c01c49242f3a8a8442fa7073fdb7f8326fd126c7ad29efa731abf0d25a9a.jpg) + +![](images/156e5ae362c04b3fcb861b82b8154913933b9e842bf897644e42f1018563d3f2.jpg) + +![](images/024087c82d1006410b557aa3d76e1090adfde3fee288435f5175ca96d69e7968.jpg) +(c) Missing annotations and labeling multiple instances as one + +![](images/cdb9f7686de82db74f57114d736848c7c72bd1f6b7a09001c7307c0eae433ed2.jpg) +Fig. 2: Imperfections in COCO-2017 masks. All masks in general have coarse boundaries. Moreover, they lack holes when required (scissors) and do not handle occlusions consistently ( $TV$ with or without figurines). Annotations are sometimes non-exhaustive, either masks are missing (apples and donuts), or multiple instances are grouped into a single mask (banana and dog). + +![](images/5f298d7666a58ec6a6271b8cb8d177e3664461a110fda874e8d381b54f525e02.jpg) + +![](images/b25c53ca87e7d70e76612b23089761504113a0693a6cd32da86cf4ff83f39122.jpg) + +Imperfections in COCO-2017 masks. To further understand the flaws in COCO annotations, we visualized thousands of COCO-2017 masks and inspected their quality and label correctness. Through this exercise, we observed recurring types of errors in annotations, as shown in Figure 2. + +1. COCO-2017 masks have coarse boundaries. This is a natural consequence of collecting masks as polygon contours, which only coarsely approximate the pixel-precise mask boundaries through piece-wise straight edges connecting points (clicks) supplied by annotators. This shortcoming is well-known and identified by several follow-up works [1, 14, 21]. +2. COCO-2017 masks lack holes when required. The polygon-drawing interface of COCO did not allow annotators to erase the interior regions of polygons. Consequently, certain objects that should naturally contain holes, lack them. Figure 2a shows two examples - the handles of scissors and cup. +3. COCO-2017 masks do not handle occlusions consistently. Masks of occluded objects are either drawn over the other occluding objects (amodal masks) or around them (modal masks). For example, observe the $TV$ masks in Figure 2b. While both, amodal and modal masks are valid, either one should be consistent in the dataset according to the task definition. We believe this inconsistency is a consequence of inadequate task instructions to annotators as well as the lack of holes in masks. +4. COCO-2017 masks are often non-exhaustive. This issue occurs in two modes, as shown in Figure 2c - either few instance masks of a category are missing1 (apple, donut), or multiple instances are grouped in a single mask (banana, dog). This is a recurring issue throughout COCO - we believe that certain annotators may not have deemed masks with grouped instances to have a wrong object contour, due to oversight. + +5. COCO-2017 contains few near-duplicate masks. We found 410 pairs of masks that overlap with intersection-over-union (IoU) greater than 0.8 in the COCO-2017 validation set ( $\approx 2.3\%$ of all instances). These pairs often segment the same underlying object with different labels. Examples include armchairs (chair and couch), calves (cow and sheep), and duffel bags (suitcase and backpack). We attribute this issue to the lack of a cross-category deduplication step in the COCO annotation pipeline. + +All these imperfections hurt the reliability of the benchmarking process. They provide noisy supervision during training and wrongfully penalize correct model predictions during evaluation. Having characterized these imperfections, we make targeted design choices in our COCO-ReM annotation pipeline to rectify them. + +# 3 COCO-ReM: COCO with Refined Masks + +In this section, we outline our annotation process for COCO-ReM. Our goal is to rectify the imperfections in COCO-2017 annotations as discussed in Section 2 to maintain continuity with prior research. We hope that higher quality ground-truth masks would improve the reliability of the benchmarking process. + +# 3.1 Annotation Pipeline + +We develop a semi-automatic pipeline - first, we refine the masks using off-the-shelf models and LVIS dataset [14], followed by manual verification of the updated masks. We refine both, the training and validation set of COCO-2017, comprising $\approx 860\mathrm{K}$ and $\approx 36\mathrm{K}$ instances respectively. We thoroughly verify the validation set to ensure consistency and pixel-precise mask quality for model evaluation. As the training set is much larger, it is only processed through the automatic steps of our pipeline. Our pipeline comprises three stages (Figure 3). + +Stage 1. Mask boundary refinement. In this stage, we aim to rectify the coarse boundary quality of COCO-2017 masks. A straightforward solution to achieve this is to re-associate all masks through a web interface with paintbrush and eraser tools instead of drawing polygon contours. Moreover, this solution naturally allows including holes in masks where necessary. While accurate, this solution would sharply increase the mask annotation time and cost. + +Along a different trend, prior works have explored model-assisted interactive segmentation as a feasible solution for collecting mask annotations with smooth boundaries [1, 14]. The key idea is to let an object detector predict a coarse initial mask and have the annotators refine it by prompting corrective mouse clicks. Inspired by the accuracy and practical feasibility of this approach, we use the Segment Anything Model (SAM [21]) to refine the COCO-2017 masks. + +- Refining masks using SAM. SAM is a promptable segmentation model that predicts high-quality masks conditioned on user-provided point and box prompts. This promptable functionality makes SAM suitable for our use case. We prompted SAM using various combinations of prompts sampled using COCO-2017 masks and obtained its predictions as the refined masks. Refer to the appendix for details about this prompting methodology. + +![](images/ef3c160383a599d8ec932245d3b44adf2318d46fb8cf14ec6a73eedf138d5e30.jpg) +Fig. 3: COCO-ReM annotation pipeline. We develop a semi-automatic annotation pipeline comprising three stages to rectify the imperfections in COCO-2017 masks. + +- Manual verification of mask quality. We visualized and manually inspected every mask created in the previous process. Through this inspection, we identified nearly 900 masks that have low quality or do not accurately cover the underlying object. These masks represent only $2.4\%$ of 36K instances in the COCO-2017 validation set, greatly reducing our manual effort and showing the efficacy of our automated process. To correct these masks, we deployed a web browser demo of SAM and provided point prompts iteratively until obtaining a precise and accurate mask. + +Through this stage, we obtain masks with more precise boundaries and better handling of occlusions, as illustrated in Figure 3 (left). + +Stage 2. Exhaustive instance annotation. This stage addresses non-exhaustively annotated instances in COCO-2017 (Figure 2c). We note that manually re-annotating every image is the most reliable solution to ensure exhaustive instance coverage. However, the manual effort would not only require extensive instructions and verification but would also be highly cost-ineffective and may lead to redundant annotations. To mitigate these issues, we automatically import instances from the LVIS [14] dataset as a strong starting point. + +- Importing instances from LVIS. The LVIS dataset comprises 1.2M high-quality instance annotations across 1203 categories for images in COCO, including all 80 COCO categories. Gupta et al. [14] took firm measures to ensure exhaustive annotations. Every image is labeled with a set of positive LVIS categories whose instances are exhaustively labeled, and a set of negative LVIS categories that are verifiably absent. The presence (or absence) of the remaining LVIS categories is unverified for the image. + +For the 80 COCO categories, LVIS contains $\approx 20\mathrm{K}$ instances for images in the COCO-2017 validation set. To import these instances, we checked whether LVIS contained more instances of an (image, category) pair than COCO-2017. If so, we replaced all COCO-2017 instances of this pair with those from LVIS. Nearly $9\%$ of (image, category) pairs in LVIS are flagged to have nonexhaustive instance annotations [14] - we ignored these while importing. + +- Discovering instances using LVIS-trained models. Instances of a given (image, category) pair of COCO-2017 may be missing in LVIS if the category + +
DatasetSplitInstance SourceTotal
COCO-2017LVISLVIS models
COCO-2017val36,781--36,781
COCO-ReMval33,4986,135105640,689
COCO-2017train860,001--860,001
COCO-ReMtrain738,353354,674-1,093,027
+ +Table 1: Instance statistics in COCO-ReM. We list the exact number of annotations obtained from different sources for both train and val splits of COCO-ReM. All masks undergo refinement through SAM before being included in COCO-ReM. + +is unverified for the corresponding image. Instances of this pair may be nonexhaustive in COCO-2017, and hence remain unresolved after the previous automatic step. To resolve this, we use publicly available object detectors trained using LVIS. On the validation set, we ensemble mask predictions using the official checkpoints of three best-performing ViTDet models [23]. As model predictions are not as accurate as human-annotated data, we manually inspect $\approx 2500$ masks predicted with high confidence and select 1056 masks. We refine these masks (Stage 1) and add them to COCO-ReM. + +This stage augments COCO-ReM with additional instances from external sources and includes more exhaustive annotations, see example in Figure 3 (middle). + +Stage 3. Correction of labeling errors. This is the final stage of our pipeline, performed manually, only for the validation set. Recall from Section 2 that we found 410 near-duplicate mask pairs in the COCO-2017 validation set. We manually inspected these and retained one mask per pair with the most accurate label, such as chair for armchairs and car for SUVs (Figure 3 (right)). We also observed that $\approx 100$ grouped instances were not resolved through Stage 2. We annotated these masks as crowd to accurately reflect their characteristics. + +# 3.2 Mask Characteristics + +Instance statistics. Our annotation pipeline imports instances from external sources - LVIS dataset and LVIS-trained models - to annotate instances exhaustively. As a result, every instance in COCO-ReM traces to one of these sources or COCO-2017. As shown in Table 1, COCO-ReM validation set extends COCO-2017 by including $\approx 6\mathrm{K}$ masks from LVIS [14] and $\approx 1\mathrm{K}$ masks from LVISTrained models [3, 23]. Since these masks replace a subset of COCO-2017 masks (see Stage 2), we retain $\approx 33.5\mathrm{K}$ masks from COCO-2017 (out of $36\mathrm{K}$ masks). Overall, COCO-ReM validation set has $\approx 4\mathrm{K}$ more instances than COCO-2017. + +On the other hand, the training set of COCO-ReM comprises a total of 1.09M high-quality masks, as compared to 860K masks in the COCO-2017 training set. The difference in instance counts between COCO-ReM and COCO-2017 sharply highlights the extent of non-exhaustively annotated instances in COCO-2017. + +Mask IoU. Figure 4 (left) shows the distribution of IoU between the (refined) masks of COCO-ReM (validation set) and their corresponding source masks in + +![](images/3edfe2a4dce42f677abb07768602a9c87fb4a8a3575d82818686ee43417c25e0.jpg) +Fig. 4: Left: Mask IoU distribution. Distribution of IoU between masks in COCO-ReM validation set and their source dataset. Right: Masks with holes. Nearly 2000 masks in COCO-ReM validation set have holes, that were missing in COCO-2017. + +![](images/612ce9a12d3c9a1c22ab482be51666ca946822a59a47f41b61eff57c3c311e43.jpg) + +COCO-2017 and LVIS datasets. The median IoU is slightly less than one, indicating that most masks have a proper coarse shape, but lack precise boundaries. This median estimate matches the findings of Gupta et al. [14], albeit observed from a smaller sample of 100 masks collected from human annotators. + +Holes in masks. Figure 4 (right) shows that $\approx 2000$ masks in COCO-ReM validation set have at least one hole. By design, COCO-2017 lacks these holes. See the appendix for qualitative examples and additional analysis per category. + +# 4 Experiments + +In our experiments, we aim to demonstrate the utility of COCO-ReM for reliable benchmarking of object detection research. Recall Section 1, we hypothesize that the imperfections in COCO-2017 may cause two issues: (1) During evaluation, models may be undesirably penalized despite predicting correct masks, yielding counter-intuitive modeling observations, and (2) During training, models receive noisy supervision and/or may learn unwanted biases (e.g., never predict masks with holes), resulting in sub-par capabilities. We claim that high-quality masks of COCO-ReM can alleviate these issues. To this end, we use COCO-ReM to evaluate and train object detectors, and compare trends with COCO-2017. + +# 4.1 Evaluation using COCO-ReM + +In this section, we evaluate fifty publicly available COCO object detectors using the COCO-ReM validation set. We focus on the validation set as the test set annotations of COCO-2017 are not publicly available. We cover a wide range of models from existing literature. All models can be categorized as either region-based or query-based: + +- Region-based models, following Mask R-CNN [17]. These models first predict labeled boxes, followed by binary masks inside those boxes. We also include cascade models, following Cascade R-CNN [3]. These models extend the region-based design by iteratively refining their predicted bounding boxes before making the final mask prediction. + +![](images/2e7d7f8d08f0a50f02103090b39eaa88181f4721c4dfdb9e7b48a8093b489de6.jpg) +Fig. 5: Evaluating fifty object detectors using COCO-ReM. All models score higher on COCO-ReM than COCO-2017. Query-based models ( $\star$ ) score much higher than region-based models on COCO-ReM, yielding opposite trends than COCO-2017. + +- **Query-based models**, inspired by the Detection Transformer design [4]. Instead of predicting masks within image regions, these models start with a set of image-independent query vectors, then decode labeled masks through a Transformer [33] by iteratively attending to image embeddings. Among query-based models, we consider Mask2Former [8] and OneFormer [19]. + +We ensure diverse image backbones in our model collection. The architecture and capacity of the backbone significantly impact object detector performance. Thus, we include convolutional backbones (ResNet [16], RegNet [31], ConvNeXt [27], InternImage [35]), hierarchical vision transformers (Swin [26], MViTv2 [24], DiNAT [15]), and plain vision transformers [11]. We include backbones of different sizes for a single model, e.g., for Cascade ViTDet [23], we include ViT-B (141M parameters), ViT-L (361M parameters), ViT-H (692M parameters). + +We believe that our model collection is sufficiently diverse to draw meaningful conclusions from comparing trends between COCO-ReM and COCO-2017 + +Testing protocol. We compute the mask Average Precision (AP) of all models using the official COCO evaluation API in Detectron2 [36]. For simplicity, we adopt a common testing protocol for all models - we resize the longest edge of input images to 1024 pixels (without changing the aspect ratio), and limit the number of detections per image to 100. + +Result 1. All models score higher on COCO-ReM. In Figure 5, we plot the mask AP trends for all fifty object detectors sorted by their COCO-2017 AP. All fifty models achieve higher AP on COCO-ReM despite being trained using only COCO-2017 masks. This result validates our hypothesis that the imperfections of COCO-2017 may have wrongfully penalized all these models. + +The difference between COCO-ReM AP and COCO-2017 AP is larger with recent high-performing models - the current state-of-the-art model (OneFormer InternImage-H) scores $+7.7$ points higher AP on COCO-ReM than COCO-2017. This increasing gap in AP suggests that the imperfections in ground-truth have worsened the issue of over-penalization and made COCO-2017 AP less adept at fine-grained assessment and ranking of recent models. + +Result 2. Query-based detectors score much higher on COCO-ReM. + +![](images/b53c630ec91e3aa63f37b4738179b34be24250dda2004bca31d9c18046c316b6.jpg) +Cascade ViTDet-L + +![](images/94e64f762682d91157bff782dbef335746ff1914bd9df9ef13621a6b6370c148.jpg) +Fig. 6: Mask predictions from models with opposite rank order on COCO-ReM and COCO-2017. Qualitative examples of ViTDet (region-based model) and Mask2Former (query-based model) show that the latter predicts visually sharper masks. COCO-ReM AP correctly ranks Mask2Former higher, unlike COCO-2017 AP. + +![](images/c4e1e7d5b611a6329229028c8b1f3185c8d24ba16f25851e9ded27eabc9205e3.jpg) + +![](images/f53cb8956febce8ee7ff7ee71fea5673e5b19f21bdd6340fd17ca511acbd7974.jpg) +Mask2Former Swin-L + +![](images/beb3e0aac27eb5dadcf890bbcbb155b3d5305d5e1134ce28a6939a5f0ec20521.jpg) + +![](images/49f5a704a75040d2ad1e43e4c328a755e60e471a3298ac07cf5a04ed81414926.jpg) + +Interestingly, we observe that the trend in AP values calculated using COCO-ReM does not align perfectly with COCO-2017. We observe frequent flips in ranks of several model pairs on COCO-ReM. While the model rankings do not transfer between benchmarks perfectly, we observe a consistent pattern in model pairs with flipped rankings. Notice the star markers $(\star)$ in Figure 5 - the flips are attributed to significantly higher scores of query-based models (Mask2Former [8] and OneFormer [19]) than region-based models on COCO-ReM. + +To further understand the cause of this performance trend, we observe the predicted masks of two models having opposite rank order according to COCO-ReM and COCO-2017 AP: Cascade ViTDet-L [23] and Mask2Former Swin-L [8]. The latter is a query-based model. Figure 6 shows some predicted masks by both of these models. Through a cursory glance, we can observe that Mask2Former predicts visually sharper masks whereas masks of ViTDet have coarse boundaries. This human judgment aligns better with COCO-ReM AP, which correctly rates Mask2Former higher. On the other hand, COCO-2017 AP rates ViTDet higher, yielding a counter-intuitive modeling observation. + +Guided by these trends, we reaffirm our claim that benchmarking newer models using the COCO-2017 may be misleading from a research perspective, and the community should adopt our COCO-ReM annotations moving forward. We include numerical results of Figure 5 in the appendix. + +# 4.2 Understanding the Difference in AP + +Recall Section 4.1 (Result 1) - all models score higher average precision (AP) on COCO-ReM, as compared to COCO-2017. What explains the difference in $AP$ ? To understand the reasons underlying this difference, we observe the components of COCO AP for a few high-performing models from Section 4.1. For this analysis, we consider two Cascade ViTDet models [23] using ViT-L/H backbones [11], and two Mask2Former models [8] using Swin-B/L backbones [26]. + +AP per IoU threshold. The calculation of COCO AP involves an average of per-category AP scores of 80 COCO categories, each computed at ten different + +![](images/6dce440b69244498cf17db818348e786f7e4cb6f5f02cd116959d0b0a9011b0e.jpg) +Fig. 7: AP per IoU threshold. While models score similarly at IoU = 0.5 on COCO-ReM and COCO-2017, they score much higher on COCO-ReM at higher IoU thresholds (0.75 and beyond), contributing to the overall difference in AP. Due to the improved mask quality of COCO-ReM, the AP at high IoU thresholds becomes more sensitive and enables better assessment of model capabilities. + +![](images/b9e2e04dae014ccdf706c6b3d0ed925153d7bb8088a9c8434a502663e9b32210.jpg) +Fig. 8: AP per category. Models score higher AP on COCO-ReM for 69 out of 80 categories. The masks of the remaining 11 categories (left) in COCO-2017 do not handle occlusions consistently (dining table, bed), lack holes (scissors, donut), or have grouped instances (banana, apple). Models learn these biases from COCO-2017 training masks and score lower COCO-ReM AP. + +IoU thresholds, $t \in \{0.5, 0.55, \dots, 0.9, 0.95\}$ . In Figure 7, we show components of AP at different IoU thresholds, averaged for the four models we use. On average, models score similarly at IoU = 0.5 for COCO-ReM and COCO-2017. However, COCO-ReM AP is much larger for IoU = 0.75 and beyond. Hence, higher scores on COCO-ReM can be attributed to AP at higher IoU thresholds. + +Intuitively, AP at IoU = t rewards model predictions that overlap with a ground-truth mask of the same category with IoU ≥ t. AP at IoU = 0.5 requires a very coarse overlap between the predicted and ground-truth masks, whereas AP at high IoU thresholds have stricter overlap requirements. Since IoU is symmetric, imprecise boundaries of ground-truth masks (COCO-2017) can wrongfully penalize valid predictions and lead to lower AP scores. COCO-ReM fixes this issue, as evident with higher AP at high IoU thresholds. This analysis hints that precise masks of COCO-ReM improve the sensitivity of AP and enable fine-grained assessment of high-performing object detectors. + +AP per category. Next, we calculate AP for every COCO category by averaging across all IoU thresholds. Figure 8 shows the difference in COCO-ReM + +
COCO evaluation set ⇒2017ReM-S1ReM-S2aReM (final)
AP49.657.355.655.5
Number of instances36,78136,78141,16840,689
+ +Table 2: COCO-ReM ablations. AP of four high-performing models (Cascade ViTDet-B/L and Mask2Former Swin-B/L) using intermediate versions from our COCO-ReM annotation pipeline. Trends suggest that the coarse boundary quality of COCO-2017 masks was a major source of noise in the benchmarking process. + +AP and COCO-2017 AP per category, averaged across the same four models. Models score higher on COCO-ReM for 69 out of 80 categories, which is reflected in the higher overall COCO-ReM AP. Significant improvements are observed in categories with generally consistent appearance (carrot and orange), and fine details (fork). Intuitively speaking, the larger difference in per-category AP can be attributed to the better quality of ground-truth masks in COCO-ReM. + +For 11 categories, models score lower on COCO-ReM than COCO-2017. These categories either have holes in their natural appearance or are often heavily occluded (bowl, dining table, scissors). Masks in COCO-ReM include holes and handle occlusions more consistently than in COCO-2017. Hence, lower per-category AP with COCO-ReM implies that the imperfections of COCO-2017 training masks impart these undesirable biases to trained models – upon checking the model predictions, we observed that they often include occluding objects for dining table and lack holes for scissors. Such predictions are not rewarded by COCO-ReM, hence incur a drop in the per-category AP. + +# 4.3 Ablations + +In this section, we conduct a simple ablation study to observe the contribution of different stages in our annotation pipeline. For the sake of simplicity, we select the same four models as chosen in Section 4.2. We evaluate these models using two versions of COCO-ReM validation set: + +1. ReM-S1: This version is obtained after the Stage 1 (mask boundary refinement) in our annotation pipeline. It comprises the same number of instances as COCO-2017, each with an improved boundary quality. +2. ReM-S2a: This version is obtained after importing instances from the LVIS dataset – the automatic part of Stage 2 (exhaustive instance annotation). We exclude the manual part of Stage 2 from this ablation for simplicity. + +Results are shown in Table 2. The large difference between COCO-ReM AP and COCO-2017 AP appears right after Stage 1, indicating that the coarse mask boundaries are a major source of noise that hurts the reliability of COCO-2017 AP. Subsequent stages incur a negligible drop in AP, likely due to larger of the validation set, as COCO AP is known to favor smaller evaluation sets [14]. + +# 4.4 Training with COCO-ReM + +Our evaluation shows COCO-ReM as a useful benchmark for accurate modeling conclusions. In this section, we train strong baseline models using COCO-ReM. + +![](images/aca81a5fb569ac865ffeb8877a78e2719cbce4efb6ed8ec55b6548b327912667.jpg) +Fig.9: Training with COCO-ReM. A ViTDet model trained using COCO-ReM training set converges faster and performs better than its counterpart trained using COCO-2017, and closely matches a $3 \times$ larger ViTDet-L trained using COCO-2017. + +![](images/202992b70680db8b366a6d249176b1df967e7b795404b349a76d026ad450d029.jpg) + +Implementation details. We train ViTDet models [23] using their official open-source implementation [36], leaving all hyperparameters unchanged. We train two models - Mask R-CNN ViTDet-B (111M parameters) and Cascade ViTDet-B (141M parameters). Li et al. [23] trained this model for 100 COCO epochs across $64 \times$ A100 GPUs (batch size of 64). To manage these costs, we implement gradient accumulation for every 8 iterations to emulate a batch size of 64 using our $8 \times$ A40 GPUs, resulting in an 11 day-long training schedule. We could only afford to train the base (ViT-B) models due to limited resources. The gorund truth annotations for test set is inaccessible so, we use the validation set solely for evaluation purposes for fair assessment of models. + +Results. Figure 9 shows that our ViTDet-B models trained using COCO-ReM significantly outperform the same ViTDet-B models trained using COCO-2017. Higher quality of training data enables faster convergence. Moreover, our ViTDet-B models are competitive with $3 \times$ larger ViTDet-L models trained using COCO-2017, indicating better parameter efficiency. While numerous prior works improve object detectors through modeling innovations, our results highlight the critical role of data quality in improving object detectors. + +# 5 Related Work + +Object detection and image segmentation. Visual perception is fundamental for any vision system to understand the world around it. At the core of a visual perception system is the task of image segmentation that aims at recognizing the entities (objects) in a given scene and densely labeling every pixel constituting the corresponding object. Image segmentation can broadly be divided into three sub-tasks: semantic, instance, and panoptic segmentation. On the one hand, semantic segmentation [5, 6, 28] targets obtaining a single amorphous binary mask covering all pixels for the corresponding category. On the other hand, instance segmentation [3, 18] involves detecting pixels for only foreground "thing" regions with clear boundaries and differentiating among separate instances for a category. Panoptic segmentation [7, 20, 34] is a combination of the former two tasks targeted at predicting distinct binary masks for "thing" regions and a single amorphous mask for the "stuff" regions. + +In this work, we particularly look at the task of instance segmentation and contribute high-quality mask annotations to benchmark further progress. + +Image segmentation benchmarks. In the last decade, considerable advancements have been observed in the classical vision task of image segmentation. A part of this progress is owed to the introduction of numerous datasets for benchmarking image segmentation models. In one of the earliest efforts, Everingham et al. [12] introduced the PASCAL-VOC-2012 challenge to recognize 20 classes in about 10K images. More challenging datasets were introduced in the following years, with ADE20K [10] covering 150 classes. Lin et al. [25] introduced the concept of "stuff" and "thing" classes in their COCO dataset with a total of 171 classes: 80 "thing" and 91 "stuff". The Cityscapes [9], Mapillary-Vistas [30], and KITTI [13] datasets introduced image segmentation datasets focused on autonomous driving applications. + +Gupta et al. [14] introduced the LVIS dataset with exhaustive annotations covering 1203 categories for COCO images. OpenImages [1] introduced 2.8M mask annotations covering 350 classes. Recently, the SA-1B dataset [21] introduced over 1.1B fine-grained, class-agnostic mask annotations for over 11M images. Despite various datasets, COCO has remained the de-facto benchmark for object detectors in instance segmentation. In this work, we examined the COCO instance annotations and proposed a new benchmark with superior mask quality. + +Re-assessing datasets for benchmarking. Human-annotated datasets are always prone to errors and imperfections, as no human is perfect. Some prior works have shared our motivation and conducted in-depth re-assessment studies, the closest to ours are works that re-assess the ImageNet dataset [2, 32]. Concurrent to our work, Zimmermann et al. [37] also re-assess COCO annotations, however they opt for recollecting polygon masks from human annotators. + +# 6 Conclusion + +In this work, we revisit the COCO-2017 mask annotations and find major issues involving, but not limited to their boundary quality and exhaustiveness. These issues raise significant concerns about the dependability of COCO-2017 to benchmark recent and future object detectors accurately. Consequently, we introduce our COCO-ReM benchmark consisting of high-quality mask annotations obtained using an effective semi-automatic annotation pipeline. + +On comparing the mask AP trends for existing object detectors on COCO-ReM and COCO-2017, we notice significant changes in model rankings, notably contradicting the conclusions drawn through COCO-2017. Moreover, training using COCO-ReM shows that paying attention to mask quality is crucial to advancing the capabilities of object detectors. We hope that COCO-ReM will aid the future research in object detection. + +# Acknowledgments + +We thank Mohamed El Banani, Agrim Gupta, and Nilesh Kulkarni for helpful discussions and feedback. We thank Palak Jagtap for her help during manual verification of masks. We thank all the authors of the prior works for publicly releasing the pre-trained model checkpoints, which helped us maintain empirical comprehensiveness in our evaluation study. We thank Vision and Language Group, IIT Roorkee for being a constant support system throughout the process. + +# Bibliography + +[1] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 4, 5, 14 +[2] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aïron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020. 14 +[3] Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN: Delving into high quality object detection. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7, 8, 13 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 2, 9 +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In ICLR, 2015. 13 +[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In TPAMI, 2017. 13 +[7] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 13 +[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 9, 10 +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 14 +[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. Semantic understanding of scenes through the ade20k dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 14 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, + +Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 9, 10 +[12] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2012. 14 +[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 14 +[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3, 4, 5, 6, 7, 8, 12, 14 +[15] Ali Hassani and Humphrey Shi. Dilated neighborhood attention transformer. arXiv:2209.15001, 2022. 9 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 9 +[17] Kaiming He, Georgia Gkioxari, Piotr Dolkar, and Ross Girshick. Mask R-CNN. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 2, 8 +[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 13 +[19] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 9, 10 +[20] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 13 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, WanYen Lo, Piotr Dolkar, and Ross Girshick. Segment Anything. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 14 +[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2012. 1 +[23] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 7, 9, 10, 13 +[24] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale + +vision transformers for classification and detection. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 9 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 1, 2, 3, 14 +[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2021. 9, 10 +[27] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 9 +[28] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 13 +[29] TorchVision maintainers and contributors. TorchVision: PyTorch's Computer Vision library. https://github.com/pytorch/vision, 2016.3 +[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 14 +[31] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 9 +[32] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In Proceedings of the International Conference on Machine Learning (ICML), 2019. 14 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2, 9 +[34] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 13 +[35] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 9 +[36] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. 3, 9, 13 +[37] Eric Zimmermann, Justin Szeto, Jerome Pasquero, and Frederic Ratle. Benchmarking a benchmark: How reliable is ms-coco? arXiv preprint arXiv:2311.02709, 2023. 14 \ No newline at end of file diff --git a/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/images.zip b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9cb2009f0de4bc418ec63e2af1a9b1e9f67a9f87 --- /dev/null +++ b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b5f463a0572b7afc5de179f37284b5bc13d3953a874c461e9c58fb670ee3837 +size 546354 diff --git a/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/layout.json b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a984c145c768bd69ba2fb3456a1facfde985f10 --- /dev/null +++ b/2024/Benchmarking Object Detectors with COCO_ A New Path Forward/layout.json @@ -0,0 +1,8641 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "text", + "content": "Benchmarking Object Detectors with COCO: A New Path Forward" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 198, + 167, + 414, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 167, + 414, + 179 + ], + "spans": [ + { + "bbox": [ + 198, + 167, + 414, + 179 + ], + "type": "text", + "content": "Shweta Singh1* Aayan Yadav1* Jitesh Jain2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 180, + 419, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 180, + 419, + 191 + ], + "spans": [ + { + "bbox": [ + 194, + 180, + 419, + 191 + ], + "type": "text", + "content": "Humphrey Shi² Justin Johnson³ Karan Desai³" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 184, + 191, + 429, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 191, + 429, + 203 + ], + "spans": [ + { + "bbox": [ + 184, + 191, + 429, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 191, + 429, + 203 + ], + "type": "text", + "content": "IIT Roorkee " + }, + { + "bbox": [ + 184, + 191, + 429, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 184, + 191, + 429, + 203 + ], + "type": "text", + "content": "Georgia Tech " + }, + { + "bbox": [ + 184, + 191, + 429, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 184, + 191, + 429, + 203 + ], + "type": "text", + "content": "University of Michigan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 228, + 455, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 228, + 455, + 449 + ], + "spans": [ + { + "bbox": [ + 159, + 228, + 455, + 449 + ], + "type": "text", + "content": "Abstract. The Common Objects in Context (COCO) dataset has been instrumental in benchmarking object detectors over the past decade. Like every dataset, COCO contains subtle errors and imperfections stemming from its annotation procedure. With the advent of high-performing models, we ask whether these errors of COCO are hindering its utility in reliably benchmarking further progress. In search for an answer, we inspect thousands of masks from COCO (2017 version) and uncover different types of errors such as imprecise mask boundaries, non-exhaustively annotated instances, and mislabeled masks. Due to the prevalence of COCO, we choose to correct these errors to maintain continuity with prior research. We develop COCO-ReM (Refined Masks), a cleaner set of annotations with visibly better mask quality than COCO-2017. We evaluate fifty object detectors and find that models that predict visually sharper masks score higher on COCO-ReM, affirming that they were being incorrectly penalized due to errors in COCO-2017. Moreover, our models trained using COCO-ReM converge faster and score higher than their larger variants trained using COCO-2017, highlighting the importance of data quality in improving object detectors. With these findings, we advocate using COCO-ReM for future object detection research. Our dataset is available at https://cocorem.xyz" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 458, + 448, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 458, + 448, + 470 + ], + "spans": [ + { + "bbox": [ + 160, + 458, + 448, + 470 + ], + "type": "text", + "content": "Keywords: object detection, segmentation, datasets and benchmarks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 504, + 230, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 230, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 230, + 516 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 528, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 648 + ], + "type": "text", + "content": "Rigorous benchmarking of computer vision models relies on task-specific datasets annotated by humans [9, 10, 22, 25]. Understandably, human-annotated data is prone to errors and imperfections due to factors such as ambiguous instructions to annotators, differences in perspective of multiple annotators, etc. Errors in ground-truth training data provide a noisy learning signal for models, resulting in sub-par capabilities. Even more problematic is the presence of errors in ground-truth evaluation data, which can cause a misalignment between evaluation metrics and human judgment when comparing model performances. In this work, we comprehensively study the extent of this issue for the canonical computer vision task of object detection and segmentation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 469, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 469, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 469, + 666 + ], + "type": "text", + "content": "* Denotes equal contribution. Correspondence to: Karan Desai (kdexd@umich.edu)" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 124, + 192, + 208 + ], + "blocks": [ + { + "bbox": [ + 197, + 115, + 239, + 123 + ], + "lines": [ + { + "bbox": [ + 197, + 115, + 239, + 123 + ], + "spans": [ + { + "bbox": [ + 197, + 115, + 239, + 123 + ], + "type": "text", + "content": "COCO-2017" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 124, + 192, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 124, + 192, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 124, + 192, + 208 + ], + "type": "image", + "image_path": "574a8021e3ae671cea135c567d69b3a6cba6a871bc830aa174b6e3b4bb0b0354.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 192, + 125, + 253, + 207 + ], + "blocks": [ + { + "bbox": [ + 192, + 125, + 253, + 207 + ], + "lines": [ + { + "bbox": [ + 192, + 125, + 253, + 207 + ], + "spans": [ + { + "bbox": [ + 192, + 125, + 253, + 207 + ], + "type": "image", + "image_path": "71d3b68b4be4dc387a967ec666f23ce8385c698499951c0195fb9dc493d0e876.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 125, + 303, + 207 + ], + "blocks": [ + { + "bbox": [ + 253, + 125, + 303, + 207 + ], + "lines": [ + { + "bbox": [ + 253, + 125, + 303, + 207 + ], + "spans": [ + { + "bbox": [ + 253, + 125, + 303, + 207 + ], + "type": "image", + "image_path": "a3a57ce1ee21d5fada8d68af59c2b8623efe7943f968c74ac516902a666f9c14.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 125, + 366, + 207 + ], + "blocks": [ + { + "bbox": [ + 361, + 115, + 425, + 124 + ], + "lines": [ + { + "bbox": [ + 361, + 115, + 425, + 124 + ], + "spans": [ + { + "bbox": [ + 361, + 115, + 425, + 124 + ], + "type": "text", + "content": "COCO-ReM (ours)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 125, + 366, + 207 + ], + "lines": [ + { + "bbox": [ + 308, + 125, + 366, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 125, + 366, + 207 + ], + "type": "image", + "image_path": "4ca0c32cfc84361a6ab439be6b5dd0c959d89393dec0db9ba2b1e297fac71795.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 366, + 125, + 426, + 207 + ], + "blocks": [ + { + "bbox": [ + 366, + 125, + 426, + 207 + ], + "lines": [ + { + "bbox": [ + 366, + 125, + 426, + 207 + ], + "spans": [ + { + "bbox": [ + 366, + 125, + 426, + 207 + ], + "type": "image", + "image_path": "2aad59070ba17f7fd23e798c7f2bb474602f38c834f7f1d4dd2ac0c445007cb9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 427, + 125, + 477, + 207 + ], + "blocks": [ + { + "bbox": [ + 427, + 125, + 477, + 207 + ], + "lines": [ + { + "bbox": [ + 427, + 125, + 477, + 207 + ], + "spans": [ + { + "bbox": [ + 427, + 125, + 477, + 207 + ], + "type": "image", + "image_path": "7061296a8b7bb6524378490b51226fae9d518b468d47f46d5bb5236ff8ddf1c6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 134, + 211, + 196, + 273 + ], + "blocks": [ + { + "bbox": [ + 134, + 211, + 196, + 273 + ], + "lines": [ + { + "bbox": [ + 134, + 211, + 196, + 273 + ], + "spans": [ + { + "bbox": [ + 134, + 211, + 196, + 273 + ], + "type": "image", + "image_path": "3e9526155ddb54d2a6c92a17aebf7c245012d8582485b0f34201d9827a3e4c5a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 196, + 211, + 276, + 273 + ], + "blocks": [ + { + "bbox": [ + 196, + 211, + 276, + 273 + ], + "lines": [ + { + "bbox": [ + 196, + 211, + 276, + 273 + ], + "spans": [ + { + "bbox": [ + 196, + 211, + 276, + 273 + ], + "type": "image", + "image_path": "c878476568b68c6f5b4e2ebf2705e58c56fa4ba9c52e529e960ed16d305ff941.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 277, + 212, + 339, + 273 + ], + "blocks": [ + { + "bbox": [ + 277, + 212, + 339, + 273 + ], + "lines": [ + { + "bbox": [ + 277, + 212, + 339, + 273 + ], + "spans": [ + { + "bbox": [ + 277, + 212, + 339, + 273 + ], + "type": "image", + "image_path": "692a973549fc58d9a355daa99d7f32cc1677912ec2eb55b39d93428b89e92d2f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 339, + 212, + 393, + 273 + ], + "blocks": [ + { + "bbox": [ + 339, + 212, + 393, + 273 + ], + "lines": [ + { + "bbox": [ + 339, + 212, + 393, + 273 + ], + "spans": [ + { + "bbox": [ + 339, + 212, + 393, + 273 + ], + "type": "image", + "image_path": "b201818a6f93b8b07e35907fe2c875e13fb1035ce4d442bdbb20316d45bb4cc2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 393, + 212, + 477, + 335 + ], + "blocks": [ + { + "bbox": [ + 393, + 212, + 477, + 335 + ], + "lines": [ + { + "bbox": [ + 393, + 212, + 477, + 335 + ], + "spans": [ + { + "bbox": [ + 393, + 212, + 477, + 335 + ], + "type": "image", + "image_path": "ae1207e334a76ec80287f2438435d8ec7be7e0f3aa08f91c9d675225e7123be1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 134, + 273, + 216, + 335 + ], + "blocks": [ + { + "bbox": [ + 134, + 273, + 216, + 335 + ], + "lines": [ + { + "bbox": [ + 134, + 273, + 216, + 335 + ], + "spans": [ + { + "bbox": [ + 134, + 273, + 216, + 335 + ], + "type": "image", + "image_path": "26062413e2c70ce9e4237df4b68488f9df7c2a8652e905a920b9f22fbd4ef011.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 339, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 132, + 339, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 482, + 384 + ], + "type": "text", + "content": "Fig. 1: COCO with Refined Masks (COCO-ReM). Top: COCO-ReM improves upon the quality of COCO-2017 masks by handling occlusions consistently (dining table) and annotating instances exhaustively (banana). Bottom: Random examples from COCO-ReM showing the mask quality (single category per image, for clarity)." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 216, + 273, + 308, + 335 + ], + "blocks": [ + { + "bbox": [ + 216, + 273, + 308, + 335 + ], + "lines": [ + { + "bbox": [ + 216, + 273, + 308, + 335 + ], + "spans": [ + { + "bbox": [ + 216, + 273, + 308, + 335 + ], + "type": "image", + "image_path": "52fff3971e82fde5b1533a07b70bb4d8c15a6d1b807f2a89a8dea9e602861892.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 308, + 273, + 393, + 335 + ], + "blocks": [ + { + "bbox": [ + 308, + 273, + 393, + 335 + ], + "lines": [ + { + "bbox": [ + 308, + 273, + 393, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 393, + 335 + ], + "type": "image", + "image_path": "6f303e59aaef9a8013714213d3162bf6b38387ffe95955f912d278192195feba.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 399, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 496 + ], + "type": "text", + "content": "The Common Objects in Context (COCO) dataset [25] has been the standard benchmark for object detection and segmentation since its inception in 2014. COCO fostered the empirical study of modern object detectors by providing a large set of 123K images annotated with 896K instance masks. The Mask AP (average precision) of COCO detectors has improved dramatically, with " + }, + { + "bbox": [ + 130, + 399, + 482, + 496 + ], + "type": "inline_equation", + "content": "\\approx 70\\%" + }, + { + "bbox": [ + 130, + 399, + 482, + 496 + ], + "type": "text", + "content": " relative gain between the pioneering Mask R-CNN [16, 17] from 2017, and recent models published in 2022-23 such as ViTDet [3, 23] and other Transformer-based models [4, 11, 33] like Mask2Former [8] and OneFormer [19]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 497, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 568 + ], + "type": "text", + "content": "Despite its popularity, COCO is not without limitations. COCO masks are known to have coarse boundaries [1, 14]. We manually inspect thousands of masks in COCO (2017 version) and observe other imperfections. As shown in Figure 1 (top), COCO-2017 masks do not handle holes and occlusions properly (dining table), and sometimes have non-exhaustively annotated instances (banana). We elaborate on these flaws in Section 2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "From an evaluation perspective, imperfections in ground-truth data make the metrics less reliable. COCO-2017 AP would wrongfully penalize models that predict more precise masks than the imperfect ground-truth masks. Moreover, models trained with such imperfect masks may exploit unwanted biases from the training data, e.g., learning to never predict masks with holes. Recently, Kirillov et al. observed that human evaluators consistently rated the masks predicted by their Segment Anything Model (SAM) to be of higher quality than a ViTDet model [23] trained on COCO, whereas the latter scored higher on COCO AP." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Singh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "These observations raise an important question - can COCO reliably benchmark the future progress in object detection research? COCO is the de-facto benchmark for object detection - countless research papers report COCO results every year; libraries like Torchvision [29] and Detectron2 [36] readily support object detector development with COCO. Hence, to maintain continuity with prior research and tap into this rich open-source ecosystem, we choose to rectify the errors of COCO and reinforce its utility for future research." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 482, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 237 + ], + "type": "text", + "content": "Contributions. We develop COCO-ReM (Refined Masks), a cleaner set of high-quality instance annotations for COCO images. The improved quality over COCO-2017 masks is visually palpable, as seen in Figure 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 238, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 285 + ], + "type": "text", + "content": "Our annotation pipeline (Section 3) utilizes SAM [21] to refine masks, and uses the LVIS dataset [14] and an ensemble of LVIS-trained models to exhaustively annotate grouped or missing instances. We (the authors) manually verify the entire validation set to provide a strong quality guarantee for evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 286, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 393 + ], + "type": "text", + "content": "In Section 4, we use COCO-ReM for benchmarking object detectors. We evaluate fifty object detectors and observe that their COCO-ReM AP produces a different model ranking than COCO-2017 AP. Surprisingly, we observe that query-based models (Mask2Former and OneFormer) score much higher on COCO-ReM than region-based models (ViTDet). Query-based models predict visually sharper masks - AP trends on COCO-ReM accurately reflect this, unlike COCO-2017. We observe that models trained using COCO-ReM converge faster and perform better than those trained using COCO-2017, highlighting the importance of mask quality in improving object detectors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 411, + 379, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 379, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 379, + 426 + ], + "type": "text", + "content": "2 Background: Revisiting COCO Masks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 436, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 472 + ], + "type": "text", + "content": "The annotation pipeline of a dataset determines the quality of annotations obtained. To design an effective annotation pipeline for COCO-ReM, we review the COCO annotation pipeline by Lin et al. [25] and understand its shortcomings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 498 + ], + "type": "text", + "content": "COCO annotation pipeline. Lin et al. [25] employed a three-stage pipeline to collect instance mask annotations for COCO images:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 498, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 521 + ], + "type": "text", + "content": "1. Category labeling: Annotators assigned one or more labels to images depending on the visible content, choosing from a fixed set of categories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "2. Instance spotting: Annotators locate up to ten instances per label (from the previous stage) in the corresponding image, by providing clicks (points)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": "3. Instance segmentation: Finally, annotators clicked around the boundaries of each located instance (from the previous stage), rendering a polygon mask. Lin et al. [25] also took numerous measures to improve annotation quality. The first two stages - category labeling and instance spotting - were performed by eight annotators per image to maximize recall. The instance segmentation stage required annotators to complete a training task before annotating actual instances in the dataset. Next, a different set of crowd workers verified the correctness and quality of annotated masks. Finally, the annotators used a paint-brush tool to mark remaining objects as crowd regions. These extensive measures avoided many scenarios that could have led to low-quality masks." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "text", + "content": "COCO-ReM Benchmark for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 225, + 176 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 225, + 176 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 225, + 176 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 225, + 176 + ], + "type": "image", + "image_path": "a85a2a36df96677fe3711cd2e8fc3dfe84022b5ae482a9f270c3b842e2068111.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 177, + 236, + 185 + ], + "lines": [ + { + "bbox": [ + 205, + 177, + 236, + 185 + ], + "spans": [ + { + "bbox": [ + 205, + 177, + 236, + 185 + ], + "type": "text", + "content": "(a) Holes" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 114, + 314, + 176 + ], + "blocks": [ + { + "bbox": [ + 225, + 114, + 314, + 176 + ], + "lines": [ + { + "bbox": [ + 225, + 114, + 314, + 176 + ], + "spans": [ + { + "bbox": [ + 225, + 114, + 314, + 176 + ], + "type": "image", + "image_path": "164ecf5f97458ae414f1efc41f0ec0f7b8261091beab2607dcd0bdcfeb0f17de.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 114, + 398, + 176 + ], + "blocks": [ + { + "bbox": [ + 315, + 114, + 398, + 176 + ], + "lines": [ + { + "bbox": [ + 315, + 114, + 398, + 176 + ], + "spans": [ + { + "bbox": [ + 315, + 114, + 398, + 176 + ], + "type": "image", + "image_path": "f401c01c49242f3a8a8442fa7073fdb7f8326fd126c7ad29efa731abf0d25a9a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 114, + 480, + 176 + ], + "blocks": [ + { + "bbox": [ + 399, + 114, + 480, + 176 + ], + "lines": [ + { + "bbox": [ + 399, + 114, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 399, + 114, + 480, + 176 + ], + "type": "image", + "image_path": "156e5ae362c04b3fcb861b82b8154913933b9e842bf897644e42f1018563d3f2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 186, + 223, + 247 + ], + "blocks": [ + { + "bbox": [ + 133, + 186, + 223, + 247 + ], + "lines": [ + { + "bbox": [ + 133, + 186, + 223, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 223, + 247 + ], + "type": "image", + "image_path": "024087c82d1006410b557aa3d76e1090adfde3fee288435f5175ca96d69e7968.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 249, + 421, + 258 + ], + "lines": [ + { + "bbox": [ + 206, + 249, + 421, + 258 + ], + "spans": [ + { + "bbox": [ + 206, + 249, + 421, + 258 + ], + "type": "text", + "content": "(c) Missing annotations and labeling multiple instances as one" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 223, + 186, + 315, + 247 + ], + "blocks": [ + { + "bbox": [ + 223, + 186, + 315, + 247 + ], + "lines": [ + { + "bbox": [ + 223, + 186, + 315, + 247 + ], + "spans": [ + { + "bbox": [ + 223, + 186, + 315, + 247 + ], + "type": "image", + "image_path": "cdb9f7686de82db74f57114d736848c7c72bd1f6b7a09001c7307c0eae433ed2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 263, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 319 + ], + "type": "text", + "content": "Fig. 2: Imperfections in COCO-2017 masks. All masks in general have coarse boundaries. Moreover, they lack holes when required (scissors) and do not handle occlusions consistently (" + }, + { + "bbox": [ + 130, + 263, + 482, + 319 + ], + "type": "inline_equation", + "content": "TV" + }, + { + "bbox": [ + 130, + 263, + 482, + 319 + ], + "type": "text", + "content": " with or without figurines). Annotations are sometimes non-exhaustive, either masks are missing (apples and donuts), or multiple instances are grouped into a single mask (banana and dog)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 186, + 399, + 247 + ], + "blocks": [ + { + "bbox": [ + 315, + 186, + 399, + 247 + ], + "lines": [ + { + "bbox": [ + 315, + 186, + 399, + 247 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 399, + 247 + ], + "type": "image", + "image_path": "5f298d7666a58ec6a6271b8cb8d177e3664461a110fda874e8d381b54f525e02.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 399, + 186, + 480, + 247 + ], + "blocks": [ + { + "bbox": [ + 399, + 186, + 480, + 247 + ], + "lines": [ + { + "bbox": [ + 399, + 186, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 399, + 186, + 480, + 247 + ], + "type": "image", + "image_path": "b25c53ca87e7d70e76612b23089761504113a0693a6cd32da86cf4ff83f39122.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 334, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 380 + ], + "type": "text", + "content": "Imperfections in COCO-2017 masks. To further understand the flaws in COCO annotations, we visualized thousands of COCO-2017 masks and inspected their quality and label correctness. Through this exercise, we observed recurring types of errors in annotations, as shown in Figure 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 382, + 482, + 644 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 382, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 382, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 133, + 382, + 482, + 441 + ], + "type": "text", + "content": "1. COCO-2017 masks have coarse boundaries. This is a natural consequence of collecting masks as polygon contours, which only coarsely approximate the pixel-precise mask boundaries through piece-wise straight edges connecting points (clicks) supplied by annotators. This shortcoming is well-known and identified by several follow-up works [1, 14, 21]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 441, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 482, + 489 + ], + "type": "text", + "content": "2. COCO-2017 masks lack holes when required. The polygon-drawing interface of COCO did not allow annotators to erase the interior regions of polygons. Consequently, certain objects that should naturally contain holes, lack them. Figure 2a shows two examples - the handles of scissors and cup." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 482, + 571 + ], + "type": "text", + "content": "3. COCO-2017 masks do not handle occlusions consistently. Masks of occluded objects are either drawn over the other occluding objects (amodal masks) or around them (modal masks). For example, observe the " + }, + { + "bbox": [ + 132, + 490, + 482, + 571 + ], + "type": "inline_equation", + "content": "TV" + }, + { + "bbox": [ + 132, + 490, + 482, + 571 + ], + "type": "text", + "content": " masks in Figure 2b. While both, amodal and modal masks are valid, either one should be consistent in the dataset according to the task definition. We believe this inconsistency is a consequence of inadequate task instructions to annotators as well as the lack of holes in masks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 573, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 482, + 644 + ], + "type": "text", + "content": "4. COCO-2017 masks are often non-exhaustive. This issue occurs in two modes, as shown in Figure 2c - either few instance masks of a category are missing1 (apple, donut), or multiple instances are grouped in a single mask (banana, dog). This is a recurring issue throughout COCO - we believe that certain annotators may not have deemed masks with grouped instances to have a wrong object contour, due to oversight." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Singh et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 449, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 449, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 449, + 665 + ], + "type": "text", + "content": "1 These images have all their missing masks annotated as non-crowd regions." + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "type": "text", + "content": "5. COCO-2017 contains few near-duplicate masks. We found 410 pairs of masks that overlap with intersection-over-union (IoU) greater than 0.8 in the COCO-2017 validation set (" + }, + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\approx 2.3\\%" + }, + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "type": "text", + "content": " of all instances). These pairs often segment the same underlying object with different labels. Examples include armchairs (chair and couch), calves (cow and sheep), and duffel bags (suitcase and backpack). We attribute this issue to the lack of a cross-category deduplication step in the COCO annotation pipeline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 481, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 248 + ], + "type": "text", + "content": "All these imperfections hurt the reliability of the benchmarking process. They provide noisy supervision during training and wrongfully penalize correct model predictions during evaluation. Having characterized these imperfections, we make targeted design choices in our COCO-ReM annotation pipeline to rectify them." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 262, + 398, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 398, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 398, + 276 + ], + "type": "text", + "content": "3 COCO-ReM: COCO with Refined Masks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 285, + 481, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 481, + 333 + ], + "type": "text", + "content": "In this section, we outline our annotation process for COCO-ReM. Our goal is to rectify the imperfections in COCO-2017 annotations as discussed in Section 2 to maintain continuity with prior research. We hope that higher quality ground-truth masks would improve the reliability of the benchmarking process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 347, + 263, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 263, + 360 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 263, + 360 + ], + "type": "text", + "content": "3.1 Annotation Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 364, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 481, + 449 + ], + "type": "text", + "content": "We develop a semi-automatic pipeline - first, we refine the masks using off-the-shelf models and LVIS dataset [14], followed by manual verification of the updated masks. We refine both, the training and validation set of COCO-2017, comprising " + }, + { + "bbox": [ + 130, + 364, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\approx 860\\mathrm{K}" + }, + { + "bbox": [ + 130, + 364, + 481, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 364, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\approx 36\\mathrm{K}" + }, + { + "bbox": [ + 130, + 364, + 481, + 449 + ], + "type": "text", + "content": " instances respectively. We thoroughly verify the validation set to ensure consistency and pixel-precise mask quality for model evaluation. As the training set is much larger, it is only processed through the automatic steps of our pipeline. Our pipeline comprises three stages (Figure 3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 450, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 521 + ], + "type": "text", + "content": "Stage 1. Mask boundary refinement. In this stage, we aim to rectify the coarse boundary quality of COCO-2017 masks. A straightforward solution to achieve this is to re-associate all masks through a web interface with paintbrush and eraser tools instead of drawing polygon contours. Moreover, this solution naturally allows including holes in masks where necessary. While accurate, this solution would sharply increase the mask annotation time and cost." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 522, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 594 + ], + "type": "text", + "content": "Along a different trend, prior works have explored model-assisted interactive segmentation as a feasible solution for collecting mask annotations with smooth boundaries [1, 14]. The key idea is to let an object detector predict a coarse initial mask and have the annotators refine it by prompting corrective mouse clicks. Inspired by the accuracy and practical feasibility of this approach, we use the Segment Anything Model (SAM [21]) to refine the COCO-2017 masks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 481, + 666 + ], + "type": "text", + "content": "- Refining masks using SAM. SAM is a promptable segmentation model that predicts high-quality masks conditioned on user-provided point and box prompts. This promptable functionality makes SAM suitable for our use case. We prompted SAM using various combinations of prompts sampled using COCO-2017 masks and obtained its predictions as the refined masks. Refer to the appendix for details about this prompting methodology." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "text", + "content": "COCO-ReM Benchmark for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 231 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 231 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 231 + ], + "type": "image", + "image_path": "ef3c160383a599d8ec932245d3b44adf2318d46fb8cf14ec6a73eedf138d5e30.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 236, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 482, + 258 + ], + "type": "text", + "content": "Fig. 3: COCO-ReM annotation pipeline. We develop a semi-automatic annotation pipeline comprising three stages to rectify the imperfections in COCO-2017 masks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 278, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 482, + 373 + ], + "type": "text", + "content": "- Manual verification of mask quality. We visualized and manually inspected every mask created in the previous process. Through this inspection, we identified nearly 900 masks that have low quality or do not accurately cover the underlying object. These masks represent only " + }, + { + "bbox": [ + 132, + 278, + 482, + 373 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 132, + 278, + 482, + 373 + ], + "type": "text", + "content": " of 36K instances in the COCO-2017 validation set, greatly reducing our manual effort and showing the efficacy of our automated process. To correct these masks, we deployed a web browser demo of SAM and provided point prompts iteratively until obtaining a precise and accurate mask." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 375, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 399 + ], + "type": "text", + "content": "Through this stage, we obtain masks with more precise boundaries and better handling of occlusions, as illustrated in Figure 3 (left)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 499, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 499, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 499, + 485 + ], + "type": "text", + "content": "Stage 2. Exhaustive instance annotation. This stage addresses non-exhaustively annotated instances in COCO-2017 (Figure 2c). We note that manually re-annotating every image is the most reliable solution to ensure exhaustive instance coverage. However, the manual effort would not only require extensive instructions and verification but would also be highly cost-ineffective and may lead to redundant annotations. To mitigate these issues, we automatically import instances from the LVIS [14] dataset as a strong starting point." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 486, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 482, + 570 + ], + "type": "text", + "content": "- Importing instances from LVIS. The LVIS dataset comprises 1.2M high-quality instance annotations across 1203 categories for images in COCO, including all 80 COCO categories. Gupta et al. [14] took firm measures to ensure exhaustive annotations. Every image is labeled with a set of positive LVIS categories whose instances are exhaustively labeled, and a set of negative LVIS categories that are verifiably absent. The presence (or absence) of the remaining LVIS categories is unverified for the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 570, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 570, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 141, + 570, + 482, + 641 + ], + "type": "text", + "content": "For the 80 COCO categories, LVIS contains " + }, + { + "bbox": [ + 141, + 570, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\approx 20\\mathrm{K}" + }, + { + "bbox": [ + 141, + 570, + 482, + 641 + ], + "type": "text", + "content": " instances for images in the COCO-2017 validation set. To import these instances, we checked whether LVIS contained more instances of an (image, category) pair than COCO-2017. If so, we replaced all COCO-2017 instances of this pair with those from LVIS. Nearly " + }, + { + "bbox": [ + 141, + 570, + 482, + 641 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 141, + 570, + 482, + 641 + ], + "type": "text", + "content": " of (image, category) pairs in LVIS are flagged to have nonexhaustive instance annotations [14] - we ignored these while importing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "- Discovering instances using LVIS-trained models. Instances of a given (image, category) pair of COCO-2017 may be missing in LVIS if the category" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Singh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 198 + ], + "type": "table", + "html": "
DatasetSplitInstance SourceTotal
COCO-2017LVISLVIS models
COCO-2017val36,781--36,781
COCO-ReMval33,4986,135105640,689
COCO-2017train860,001--860,001
COCO-ReMtrain738,353354,674-1,093,027
", + "image_path": "2be83d315527fd14b7c33e54bd2c93437019664aec47c36aa2040a5430fba003.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 198, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 198, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 479, + 232 + ], + "type": "text", + "content": "Table 1: Instance statistics in COCO-ReM. We list the exact number of annotations obtained from different sources for both train and val splits of COCO-ReM. All masks undergo refinement through SAM before being included in COCO-ReM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 142, + 247, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 247, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 142, + 247, + 480, + 342 + ], + "type": "text", + "content": "is unverified for the corresponding image. Instances of this pair may be nonexhaustive in COCO-2017, and hence remain unresolved after the previous automatic step. To resolve this, we use publicly available object detectors trained using LVIS. On the validation set, we ensemble mask predictions using the official checkpoints of three best-performing ViTDet models [23]. As model predictions are not as accurate as human-annotated data, we manually inspect " + }, + { + "bbox": [ + 142, + 247, + 480, + 342 + ], + "type": "inline_equation", + "content": "\\approx 2500" + }, + { + "bbox": [ + 142, + 247, + 480, + 342 + ], + "type": "text", + "content": " masks predicted with high confidence and select 1056 masks. We refine these masks (Stage 1) and add them to COCO-ReM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 344, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 479, + 367 + ], + "type": "text", + "content": "This stage augments COCO-ReM with additional instances from external sources and includes more exhaustive annotations, see example in Figure 3 (middle)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 370, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 453 + ], + "type": "text", + "content": "Stage 3. Correction of labeling errors. This is the final stage of our pipeline, performed manually, only for the validation set. Recall from Section 2 that we found 410 near-duplicate mask pairs in the COCO-2017 validation set. We manually inspected these and retained one mask per pair with the most accurate label, such as chair for armchairs and car for SUVs (Figure 3 (right)). We also observed that " + }, + { + "bbox": [ + 132, + 370, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\approx 100" + }, + { + "bbox": [ + 132, + 370, + 480, + 453 + ], + "type": "text", + "content": " grouped instances were not resolved through Stage 2. We annotated these masks as crowd to accurately reflect their characteristics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 472, + 266, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 266, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 266, + 483 + ], + "type": "text", + "content": "3.2 Mask Characteristics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "text", + "content": "Instance statistics. Our annotation pipeline imports instances from external sources - LVIS dataset and LVIS-trained models - to annotate instances exhaustively. As a result, every instance in COCO-ReM traces to one of these sources or COCO-2017. As shown in Table 1, COCO-ReM validation set extends COCO-2017 by including " + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\approx 6\\mathrm{K}" + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "text", + "content": " masks from LVIS [14] and " + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\approx 1\\mathrm{K}" + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "text", + "content": " masks from LVISTrained models [3, 23]. Since these masks replace a subset of COCO-2017 masks (see Stage 2), we retain " + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\approx 33.5\\mathrm{K}" + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "text", + "content": " masks from COCO-2017 (out of " + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "inline_equation", + "content": "36\\mathrm{K}" + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "text", + "content": " masks). Overall, COCO-ReM validation set has " + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\approx 4\\mathrm{K}" + }, + { + "bbox": [ + 132, + 495, + 480, + 590 + ], + "type": "text", + "content": " more instances than COCO-2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 591, + 480, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 480, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 480, + 639 + ], + "type": "text", + "content": "On the other hand, the training set of COCO-ReM comprises a total of 1.09M high-quality masks, as compared to 860K masks in the COCO-2017 training set. The difference in instance counts between COCO-ReM and COCO-2017 sharply highlights the extent of non-exhaustively annotated instances in COCO-2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "content": "Mask IoU. Figure 4 (left) shows the distribution of IoU between the (refined) masks of COCO-ReM (validation set) and their corresponding source masks in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "text", + "content": "COCO-ReM Benchmark for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 376, + 212 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 376, + 212 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 376, + 212 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 376, + 212 + ], + "type": "image", + "image_path": "3edfe2a4dce42f677abb07768602a9c87fb4a8a3575d82818686ee43417c25e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 253 + ], + "type": "text", + "content": "Fig. 4: Left: Mask IoU distribution. Distribution of IoU between masks in COCO-ReM validation set and their source dataset. Right: Masks with holes. Nearly 2000 masks in COCO-ReM validation set have holes, that were missing in COCO-2017." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 383, + 118, + 477, + 211 + ], + "blocks": [ + { + "bbox": [ + 383, + 118, + 477, + 211 + ], + "lines": [ + { + "bbox": [ + 383, + 118, + 477, + 211 + ], + "spans": [ + { + "bbox": [ + 383, + 118, + 477, + 211 + ], + "type": "image", + "image_path": "612ce9a12d3c9a1c22ab482be51666ca946822a59a47f41b61eff57c3c311e43.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "type": "text", + "content": "COCO-2017 and LVIS datasets. The median IoU is slightly less than one, indicating that most masks have a proper coarse shape, but lack precise boundaries. This median estimate matches the findings of Gupta et al. [14], albeit observed from a smaller sample of 100 masks collected from human annotators." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 321, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 480, + 357 + ], + "type": "text", + "content": "Holes in masks. Figure 4 (right) shows that " + }, + { + "bbox": [ + 130, + 321, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\approx 2000" + }, + { + "bbox": [ + 130, + 321, + 480, + 357 + ], + "type": "text", + "content": " masks in COCO-ReM validation set have at least one hole. By design, COCO-2017 lacks these holes. See the appendix for qualitative examples and additional analysis per category." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 375, + 230, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 375, + 230, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 230, + 389 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 400, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 509 + ], + "type": "text", + "content": "In our experiments, we aim to demonstrate the utility of COCO-ReM for reliable benchmarking of object detection research. Recall Section 1, we hypothesize that the imperfections in COCO-2017 may cause two issues: (1) During evaluation, models may be undesirably penalized despite predicting correct masks, yielding counter-intuitive modeling observations, and (2) During training, models receive noisy supervision and/or may learn unwanted biases (e.g., never predict masks with holes), resulting in sub-par capabilities. We claim that high-quality masks of COCO-ReM can alleviate these issues. To this end, we use COCO-ReM to evaluate and train object detectors, and compare trends with COCO-2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 525, + 311, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 311, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 311, + 538 + ], + "type": "text", + "content": "4.1 Evaluation using COCO-ReM" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "In this section, we evaluate fifty publicly available COCO object detectors using the COCO-ReM validation set. We focus on the validation set as the test set annotations of COCO-2017 are not publicly available. We cover a wide range of models from existing literature. All models can be categorized as either region-based or query-based:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 482, + 665 + ], + "type": "text", + "content": "- Region-based models, following Mask R-CNN [17]. These models first predict labeled boxes, followed by binary masks inside those boxes. We also include cascade models, following Cascade R-CNN [3]. These models extend the region-based design by iteratively refining their predicted bounding boxes before making the final mask prediction." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Singh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 478, + 225 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 478, + 225 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 478, + 225 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 478, + 225 + ], + "type": "image", + "image_path": "2e7d7f8d08f0a50f02103090b39eaa88181f4721c4dfdb9e7b48a8093b489de6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 231, + 482, + 265 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 265 + ], + "type": "text", + "content": "Fig. 5: Evaluating fifty object detectors using COCO-ReM. All models score higher on COCO-ReM than COCO-2017. Query-based models (" + }, + { + "bbox": [ + 130, + 231, + 482, + 265 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 130, + 231, + 482, + 265 + ], + "type": "text", + "content": ") score much higher than region-based models on COCO-ReM, yielding opposite trends than COCO-2017." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 284, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 284, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 284, + 479, + 344 + ], + "type": "text", + "content": "- **Query-based models**, inspired by the Detection Transformer design [4]. Instead of predicting masks within image regions, these models start with a set of image-independent query vectors, then decode labeled masks through a Transformer [33] by iteratively attending to image embeddings. Among query-based models, we consider Mask2Former [8] and OneFormer [19]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 479, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 479, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 479, + 428 + ], + "type": "text", + "content": "We ensure diverse image backbones in our model collection. The architecture and capacity of the backbone significantly impact object detector performance. Thus, we include convolutional backbones (ResNet [16], RegNet [31], ConvNeXt [27], InternImage [35]), hierarchical vision transformers (Swin [26], MViTv2 [24], DiNAT [15]), and plain vision transformers [11]. We include backbones of different sizes for a single model, e.g., for Cascade ViTDet [23], we include ViT-B (141M parameters), ViT-L (361M parameters), ViT-H (692M parameters)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 429, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 479, + 453 + ], + "type": "text", + "content": "We believe that our model collection is sufficiently diverse to draw meaningful conclusions from comparing trends between COCO-ReM and COCO-2017" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 456, + 480, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 480, + 516 + ], + "type": "text", + "content": "Testing protocol. We compute the mask Average Precision (AP) of all models using the official COCO evaluation API in Detectron2 [36]. For simplicity, we adopt a common testing protocol for all models - we resize the longest edge of input images to 1024 pixels (without changing the aspect ratio), and limit the number of detections per image to 100." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "text", + "content": "Result 1. All models score higher on COCO-ReM. In Figure 5, we plot the mask AP trends for all fifty object detectors sorted by their COCO-2017 AP. All fifty models achieve higher AP on COCO-ReM despite being trained using only COCO-2017 masks. This result validates our hypothesis that the imperfections of COCO-2017 may have wrongfully penalized all these models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 578, + 479, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 479, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 479, + 650 + ], + "type": "text", + "content": "The difference between COCO-ReM AP and COCO-2017 AP is larger with recent high-performing models - the current state-of-the-art model (OneFormer InternImage-H) scores " + }, + { + "bbox": [ + 130, + 578, + 479, + 650 + ], + "type": "inline_equation", + "content": "+7.7" + }, + { + "bbox": [ + 130, + 578, + 479, + 650 + ], + "type": "text", + "content": " points higher AP on COCO-ReM than COCO-2017. This increasing gap in AP suggests that the imperfections in ground-truth have worsened the issue of over-penalization and made COCO-2017 AP less adept at fine-grained assessment and ranking of recent models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 653, + 476, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 476, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 476, + 665 + ], + "type": "text", + "content": "Result 2. Query-based detectors score much higher on COCO-ReM." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "text", + "content": "COCO-ReM Benchmark for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 122, + 219, + 178 + ], + "blocks": [ + { + "bbox": [ + 189, + 114, + 255, + 122 + ], + "lines": [ + { + "bbox": [ + 189, + 114, + 255, + 122 + ], + "spans": [ + { + "bbox": [ + 189, + 114, + 255, + 122 + ], + "type": "text", + "content": "Cascade ViTDet-L" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 122, + 219, + 178 + ], + "lines": [ + { + "bbox": [ + 134, + 122, + 219, + 178 + ], + "spans": [ + { + "bbox": [ + 134, + 122, + 219, + 178 + ], + "type": "image", + "image_path": "b53c630ec91e3aa63f37b4738179b34be24250dda2004bca31d9c18046c316b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 178, + 219, + 226 + ], + "blocks": [ + { + "bbox": [ + 134, + 178, + 219, + 226 + ], + "lines": [ + { + "bbox": [ + 134, + 178, + 219, + 226 + ], + "spans": [ + { + "bbox": [ + 134, + 178, + 219, + 226 + ], + "type": "image", + "image_path": "94e64f762682d91157bff782dbef335746ff1914bd9df9ef13621a6b6370c148.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 481, + 281 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 481, + 281 + ], + "type": "text", + "content": "Fig. 6: Mask predictions from models with opposite rank order on COCO-ReM and COCO-2017. Qualitative examples of ViTDet (region-based model) and Mask2Former (query-based model) show that the latter predicts visually sharper masks. COCO-ReM AP correctly ranks Mask2Former higher, unlike COCO-2017 AP." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 122, + 304, + 226 + ], + "blocks": [ + { + "bbox": [ + 220, + 122, + 304, + 226 + ], + "lines": [ + { + "bbox": [ + 220, + 122, + 304, + 226 + ], + "spans": [ + { + "bbox": [ + 220, + 122, + 304, + 226 + ], + "type": "image", + "image_path": "c4e1e7d5b611a6329229028c8b1f3185c8d24ba16f25851e9ded27eabc9205e3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 122, + 393, + 178 + ], + "blocks": [ + { + "bbox": [ + 351, + 114, + 430, + 122 + ], + "lines": [ + { + "bbox": [ + 351, + 114, + 430, + 122 + ], + "spans": [ + { + "bbox": [ + 351, + 114, + 430, + 122 + ], + "type": "text", + "content": "Mask2Former Swin-L" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 122, + 393, + 178 + ], + "lines": [ + { + "bbox": [ + 309, + 122, + 393, + 178 + ], + "spans": [ + { + "bbox": [ + 309, + 122, + 393, + 178 + ], + "type": "image", + "image_path": "f53cb8956febce8ee7ff7ee71fea5673e5b19f21bdd6340fd17ca511acbd7974.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 178, + 393, + 226 + ], + "blocks": [ + { + "bbox": [ + 309, + 178, + 393, + 226 + ], + "lines": [ + { + "bbox": [ + 309, + 178, + 393, + 226 + ], + "spans": [ + { + "bbox": [ + 309, + 178, + 393, + 226 + ], + "type": "image", + "image_path": "beb3e0aac27eb5dadcf890bbcbb155b3d5305d5e1134ce28a6939a5f0ec20521.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 393, + 122, + 478, + 226 + ], + "blocks": [ + { + "bbox": [ + 393, + 122, + 478, + 226 + ], + "lines": [ + { + "bbox": [ + 393, + 122, + 478, + 226 + ], + "spans": [ + { + "bbox": [ + 393, + 122, + 478, + 226 + ], + "type": "image", + "image_path": "49f5a704a75040d2ad1e43e4c328a755e60e471a3298ac07cf5a04ed81414926.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 293, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 376 + ], + "type": "text", + "content": "Interestingly, we observe that the trend in AP values calculated using COCO-ReM does not align perfectly with COCO-2017. We observe frequent flips in ranks of several model pairs on COCO-ReM. While the model rankings do not transfer between benchmarks perfectly, we observe a consistent pattern in model pairs with flipped rankings. Notice the star markers " + }, + { + "bbox": [ + 130, + 293, + 480, + 376 + ], + "type": "inline_equation", + "content": "(\\star)" + }, + { + "bbox": [ + 130, + 293, + 480, + 376 + ], + "type": "text", + "content": " in Figure 5 - the flips are attributed to significantly higher scores of query-based models (Mask2Former [8] and OneFormer [19]) than region-based models on COCO-ReM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 377, + 481, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 481, + 484 + ], + "type": "text", + "content": "To further understand the cause of this performance trend, we observe the predicted masks of two models having opposite rank order according to COCO-ReM and COCO-2017 AP: Cascade ViTDet-L [23] and Mask2Former Swin-L [8]. The latter is a query-based model. Figure 6 shows some predicted masks by both of these models. Through a cursory glance, we can observe that Mask2Former predicts visually sharper masks whereas masks of ViTDet have coarse boundaries. This human judgment aligns better with COCO-ReM AP, which correctly rates Mask2Former higher. On the other hand, COCO-2017 AP rates ViTDet higher, yielding a counter-intuitive modeling observation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 485, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 533 + ], + "type": "text", + "content": "Guided by these trends, we reaffirm our claim that benchmarking newer models using the COCO-2017 may be misleading from a research perspective, and the community should adopt our COCO-ReM annotations moving forward. We include numerical results of Figure 5 in the appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 548, + 343, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 343, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 343, + 561 + ], + "type": "text", + "content": "4.2 Understanding the Difference in AP" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 567, + 481, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 639 + ], + "type": "text", + "content": "Recall Section 4.1 (Result 1) - all models score higher average precision (AP) on COCO-ReM, as compared to COCO-2017. What explains the difference in " + }, + { + "bbox": [ + 130, + 567, + 481, + 639 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 130, + 567, + 481, + 639 + ], + "type": "text", + "content": "? To understand the reasons underlying this difference, we observe the components of COCO AP for a few high-performing models from Section 4.1. For this analysis, we consider two Cascade ViTDet models [23] using ViT-L/H backbones [11], and two Mask2Former models [8] using Swin-B/L backbones [26]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "AP per IoU threshold. The calculation of COCO AP involves an average of per-category AP scores of 80 COCO categories, each computed at ten different" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Singh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 472, + 228 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 472, + 228 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 472, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 472, + 228 + ], + "type": "image", + "image_path": "6dce440b69244498cf17db818348e786f7e4cb6f5f02cd116959d0b0a9011b0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 228, + 481, + 284 + ], + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 284 + ], + "type": "text", + "content": "Fig. 7: AP per IoU threshold. While models score similarly at IoU = 0.5 on COCO-ReM and COCO-2017, they score much higher on COCO-ReM at higher IoU thresholds (0.75 and beyond), contributing to the overall difference in AP. Due to the improved mask quality of COCO-ReM, the AP at high IoU thresholds becomes more sensitive and enables better assessment of model capabilities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 291, + 473, + 405 + ], + "blocks": [ + { + "bbox": [ + 138, + 291, + 473, + 405 + ], + "lines": [ + { + "bbox": [ + 138, + 291, + 473, + 405 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 473, + 405 + ], + "type": "image", + "image_path": "b9e2e04dae014ccdf706c6b3d0ed925153d7bb8088a9c8434a502663e9b32210.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 405, + 482, + 460 + ], + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 460 + ], + "type": "text", + "content": "Fig. 8: AP per category. Models score higher AP on COCO-ReM for 69 out of 80 categories. The masks of the remaining 11 categories (left) in COCO-2017 do not handle occlusions consistently (dining table, bed), lack holes (scissors, donut), or have grouped instances (banana, apple). Models learn these biases from COCO-2017 training masks and score lower COCO-ReM AP." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 531 + ], + "type": "text", + "content": "IoU thresholds, " + }, + { + "bbox": [ + 130, + 472, + 482, + 531 + ], + "type": "inline_equation", + "content": "t \\in \\{0.5, 0.55, \\dots, 0.9, 0.95\\}" + }, + { + "bbox": [ + 130, + 472, + 482, + 531 + ], + "type": "text", + "content": ". In Figure 7, we show components of AP at different IoU thresholds, averaged for the four models we use. On average, models score similarly at IoU = 0.5 for COCO-ReM and COCO-2017. However, COCO-ReM AP is much larger for IoU = 0.75 and beyond. Hence, higher scores on COCO-ReM can be attributed to AP at higher IoU thresholds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "text", + "content": "Intuitively, AP at IoU = t rewards model predictions that overlap with a ground-truth mask of the same category with IoU ≥ t. AP at IoU = 0.5 requires a very coarse overlap between the predicted and ground-truth masks, whereas AP at high IoU thresholds have stricter overlap requirements. Since IoU is symmetric, imprecise boundaries of ground-truth masks (COCO-2017) can wrongfully penalize valid predictions and lead to lower AP scores. COCO-ReM fixes this issue, as evident with higher AP at high IoU thresholds. This analysis hints that precise masks of COCO-ReM improve the sensitivity of AP and enable fine-grained assessment of high-performing object detectors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "AP per category. Next, we calculate AP for every COCO category by averaging across all IoU thresholds. Figure 8 shows the difference in COCO-ReM" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "text", + "content": "COCO-ReM Benchmark for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 158 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 158 + ], + "type": "table", + "html": "
COCO evaluation set ⇒2017ReM-S1ReM-S2aReM (final)
AP49.657.355.655.5
Number of instances36,78136,78141,16840,689
", + "image_path": "d3c1adaedf9e1a609e0b3abc100b9b7b8ec9e5cb2681f8d1dfbb72fbfc0bd2ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 159, + 482, + 204 + ], + "lines": [ + { + "bbox": [ + 130, + 159, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 482, + 204 + ], + "type": "text", + "content": "Table 2: COCO-ReM ablations. AP of four high-performing models (Cascade ViTDet-B/L and Mask2Former Swin-B/L) using intermediate versions from our COCO-ReM annotation pipeline. Trends suggest that the coarse boundary quality of COCO-2017 masks was a major source of noise in the benchmarking process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 219, + 481, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 481, + 290 + ], + "type": "text", + "content": "AP and COCO-2017 AP per category, averaged across the same four models. Models score higher on COCO-ReM for 69 out of 80 categories, which is reflected in the higher overall COCO-ReM AP. Significant improvements are observed in categories with generally consistent appearance (carrot and orange), and fine details (fork). Intuitively speaking, the larger difference in per-category AP can be attributed to the better quality of ground-truth masks in COCO-ReM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "text", + "content": "For 11 categories, models score lower on COCO-ReM than COCO-2017. These categories either have holes in their natural appearance or are often heavily occluded (bowl, dining table, scissors). Masks in COCO-ReM include holes and handle occlusions more consistently than in COCO-2017. Hence, lower per-category AP with COCO-ReM implies that the imperfections of COCO-2017 training masks impart these undesirable biases to trained models – upon checking the model predictions, we observed that they often include occluding objects for dining table and lack holes for scissors. Such predictions are not rewarded by COCO-ReM, hence incur a drop in the per-category AP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 413, + 211, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 211, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 211, + 424 + ], + "type": "text", + "content": "4.3 Ablations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 481, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 477 + ], + "type": "text", + "content": "In this section, we conduct a simple ablation study to observe the contribution of different stages in our annotation pipeline. For the sake of simplicity, we select the same four models as chosen in Section 4.2. We evaluate these models using two versions of COCO-ReM validation set:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 478, + 480, + 549 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 133, + 478, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 478, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 478, + 480, + 514 + ], + "type": "text", + "content": "1. ReM-S1: This version is obtained after the Stage 1 (mask boundary refinement) in our annotation pipeline. It comprises the same number of instances as COCO-2017, each with an improved boundary quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 514, + 480, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 549 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 549 + ], + "type": "text", + "content": "2. ReM-S2a: This version is obtained after importing instances from the LVIS dataset – the automatic part of Stage 2 (exhaustive instance annotation). We exclude the manual part of Stage 2 from this ablation for simplicity." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 550, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 481, + 611 + ], + "type": "text", + "content": "Results are shown in Table 2. The large difference between COCO-ReM AP and COCO-2017 AP appears right after Stage 1, indicating that the coarse mask boundaries are a major source of noise that hurts the reliability of COCO-2017 AP. Subsequent stages incur a negligible drop in AP, likely due to larger of the validation set, as COCO AP is known to favor smaller evaluation sets [14]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 624, + 296, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 296, + 636 + ], + "type": "text", + "content": "4.4 Training with COCO-ReM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Our evaluation shows COCO-ReM as a useful benchmark for accurate modeling conclusions. In this section, we train strong baseline models using COCO-ReM." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Singh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 309, + 190 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 309, + 190 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 309, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 309, + 190 + ], + "type": "image", + "image_path": "aca81a5fb569ac865ffeb8877a78e2719cbce4efb6ed8ec55b6548b327912667.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "type": "text", + "content": "Fig.9: Training with COCO-ReM. A ViTDet model trained using COCO-ReM training set converges faster and performs better than its counterpart trained using COCO-2017, and closely matches a " + }, + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "type": "text", + "content": " larger ViTDet-L trained using COCO-2017." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 116, + 476, + 190 + ], + "blocks": [ + { + "bbox": [ + 310, + 116, + 476, + 190 + ], + "lines": [ + { + "bbox": [ + 310, + 116, + 476, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 116, + 476, + 190 + ], + "type": "image", + "image_path": "202992b70680db8b366a6d249176b1df967e7b795404b349a76d026ad450d029.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 250, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 370 + ], + "type": "text", + "content": "Implementation details. We train ViTDet models [23] using their official open-source implementation [36], leaving all hyperparameters unchanged. We train two models - Mask R-CNN ViTDet-B (111M parameters) and Cascade ViTDet-B (141M parameters). Li et al. [23] trained this model for 100 COCO epochs across " + }, + { + "bbox": [ + 130, + 250, + 482, + 370 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 130, + 250, + 482, + 370 + ], + "type": "text", + "content": " A100 GPUs (batch size of 64). To manage these costs, we implement gradient accumulation for every 8 iterations to emulate a batch size of 64 using our " + }, + { + "bbox": [ + 130, + 250, + 482, + 370 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 130, + 250, + 482, + 370 + ], + "type": "text", + "content": " A40 GPUs, resulting in an 11 day-long training schedule. We could only afford to train the base (ViT-B) models due to limited resources. The gorund truth annotations for test set is inaccessible so, we use the validation set solely for evaluation purposes for fair assessment of models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 483, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 483, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 483, + 458 + ], + "type": "text", + "content": "Results. Figure 9 shows that our ViTDet-B models trained using COCO-ReM significantly outperform the same ViTDet-B models trained using COCO-2017. Higher quality of training data enables faster convergence. Moreover, our ViTDet-B models are competitive with " + }, + { + "bbox": [ + 130, + 373, + 483, + 458 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 130, + 373, + 483, + 458 + ], + "type": "text", + "content": " larger ViTDet-L models trained using COCO-2017, indicating better parameter efficiency. While numerous prior works improve object detectors through modeling innovations, our results highlight the critical role of data quality in improving object detectors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 479, + 237, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 237, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 237, + 491 + ], + "type": "text", + "content": "5 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Object detection and image segmentation. Visual perception is fundamental for any vision system to understand the world around it. At the core of a visual perception system is the task of image segmentation that aims at recognizing the entities (objects) in a given scene and densely labeling every pixel constituting the corresponding object. Image segmentation can broadly be divided into three sub-tasks: semantic, instance, and panoptic segmentation. On the one hand, semantic segmentation [5, 6, 28] targets obtaining a single amorphous binary mask covering all pixels for the corresponding category. On the other hand, instance segmentation [3, 18] involves detecting pixels for only foreground \"thing\" regions with clear boundaries and differentiating among separate instances for a category. Panoptic segmentation [7, 20, 34] is a combination of the former two tasks targeted at predicting distinct binary masks for \"thing\" regions and a single amorphous mask for the \"stuff\" regions." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "text", + "content": "COCO-ReM Benchmark for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "In this work, we particularly look at the task of instance segmentation and contribute high-quality mask annotations to benchmark further progress." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 145, + 482, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 145, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 145, + 482, + 277 + ], + "type": "text", + "content": "Image segmentation benchmarks. In the last decade, considerable advancements have been observed in the classical vision task of image segmentation. A part of this progress is owed to the introduction of numerous datasets for benchmarking image segmentation models. In one of the earliest efforts, Everingham et al. [12] introduced the PASCAL-VOC-2012 challenge to recognize 20 classes in about 10K images. More challenging datasets were introduced in the following years, with ADE20K [10] covering 150 classes. Lin et al. [25] introduced the concept of \"stuff\" and \"thing\" classes in their COCO dataset with a total of 171 classes: 80 \"thing\" and 91 \"stuff\". The Cityscapes [9], Mapillary-Vistas [30], and KITTI [13] datasets introduced image segmentation datasets focused on autonomous driving applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 280, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 365 + ], + "type": "text", + "content": "Gupta et al. [14] introduced the LVIS dataset with exhaustive annotations covering 1203 categories for COCO images. OpenImages [1] introduced 2.8M mask annotations covering 350 classes. Recently, the SA-1B dataset [21] introduced over 1.1B fine-grained, class-agnostic mask annotations for over 11M images. Despite various datasets, COCO has remained the de-facto benchmark for object detectors in instance segmentation. In this work, we examined the COCO instance annotations and proposed a new benchmark with superior mask quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": "Re-assessing datasets for benchmarking. Human-annotated datasets are always prone to errors and imperfections, as no human is perfect. Some prior works have shared our motivation and conducted in-depth re-assessment studies, the closest to ours are works that re-assess the ImageNet dataset [2, 32]. Concurrent to our work, Zimmermann et al. [37] also re-assess COCO annotations, however they opt for recollecting polygon masks from human annotators." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 476, + 220, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 220, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 220, + 489 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 518, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 590 + ], + "type": "text", + "content": "In this work, we revisit the COCO-2017 mask annotations and find major issues involving, but not limited to their boundary quality and exhaustiveness. These issues raise significant concerns about the dependability of COCO-2017 to benchmark recent and future object detectors accurately. Consequently, we introduce our COCO-ReM benchmark consisting of high-quality mask annotations obtained using an effective semi-automatic annotation pipeline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "On comparing the mask AP trends for existing object detectors on COCO-ReM and COCO-2017, we notice significant changes in model rankings, notably contradicting the conclusions drawn through COCO-2017. Moreover, training using COCO-ReM shows that paying attention to mask quality is crucial to advancing the capabilities of object detectors. We hope that COCO-ReM will aid the future research in object detection." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Singh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 213 + ], + "type": "text", + "content": "We thank Mohamed El Banani, Agrim Gupta, and Nilesh Kulkarni for helpful discussions and feedback. We thank Palak Jagtap for her help during manual verification of masks. We thank all the authors of the prior works for publicly releasing the pre-trained model checkpoints, which helped us maintain empirical comprehensiveness in our evaluation study. We thank Vision and Language Group, IIT Roorkee for being a constant support system throughout the process." + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 261, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 261, + 91, + 447, + 102 + ], + "type": "text", + "content": "COCO-ReM Benchmark for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 261, + 112, + 353, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 112, + 353, + 129 + ], + "spans": [ + { + "bbox": [ + 261, + 112, + 353, + 129 + ], + "type": "text", + "content": "Bibliography" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 138, + 167, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 481, + 215 + ], + "type": "text", + "content": "[1] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 4, 5, 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 216, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 481, + 251 + ], + "type": "text", + "content": "[2] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aïron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020. 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 251, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 286 + ], + "type": "text", + "content": "[3] Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN: Delving into high quality object detection. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7, 8, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 286, + 481, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 481, + 334 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 2, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 334, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 481, + 369 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In ICLR, 2015. 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 369, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 481, + 415 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In TPAMI, 2017. 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 416, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 416, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 416, + 481, + 475 + ], + "type": "text", + "content": "[7] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 475, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 481, + 523 + ], + "type": "text", + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 9, 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 523, + 481, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 582 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 582, + 481, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 582, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 133, + 582, + 481, + 641 + ], + "type": "text", + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. Semantic understanding of scenes through the ade20k dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 152, + 115, + 481, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 115, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 481, + 163 + ], + "type": "text", + "content": "Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 9, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 164, + 481, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 164, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 133, + 164, + 481, + 212 + ], + "type": "text", + "content": "[12] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2012. 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 213, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 213, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 481, + 247 + ], + "type": "text", + "content": "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 248, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 248, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 133, + 248, + 481, + 295 + ], + "type": "text", + "content": "[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3, 4, 5, 6, 7, 8, 12, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 296, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 296, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 296, + 481, + 319 + ], + "type": "text", + "content": "[15] Ali Hassani and Humphrey Shi. Dilated neighborhood attention transformer. arXiv:2209.15001, 2022. 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 319, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 319, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 133, + 319, + 481, + 355 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 355, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 355, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 133, + 355, + 481, + 391 + ], + "type": "text", + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dolkar, and Ross Girshick. Mask R-CNN. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 392, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 481, + 426 + ], + "type": "text", + "content": "[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 426, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 426, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 133, + 426, + 481, + 474 + ], + "type": "text", + "content": "[19] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 9, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 475, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 481, + 510 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 510, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 481, + 568 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, WanYen Lo, Piotr Dolkar, and Ross Girshick. Segment Anything. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 570, + 481, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 570, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 133, + 570, + 481, + 605 + ], + "type": "text", + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2012. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 606, + 481, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 606, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 133, + 606, + 481, + 641 + ], + "type": "text", + "content": "[23] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 7, 9, 10, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "[24] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "text", + "content": "COCO-ReM Benchmark for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 152, + 115, + 482, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 115, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 482, + 140 + ], + "type": "text", + "content": "vision transformers for classification and detection. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 186 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 1, 2, 3, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 186, + 481, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 481, + 233 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 481, + 233 + ], + "type": "text", + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2021. 9, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 233, + 481, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 233, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 133, + 233, + 481, + 269 + ], + "type": "text", + "content": "[27] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 269, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 269, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 269, + 481, + 303 + ], + "type": "text", + "content": "[28] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 303, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 303, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 133, + 303, + 481, + 326 + ], + "type": "text", + "content": "[29] TorchVision maintainers and contributors. TorchVision: PyTorch's Computer Vision library. https://github.com/pytorch/vision, 2016.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 326, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 481, + 373 + ], + "type": "text", + "content": "[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2017. 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 373, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 373, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 373, + 481, + 407 + ], + "type": "text", + "content": "[31] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 407, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 481, + 442 + ], + "type": "text", + "content": "[32] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In Proceedings of the International Conference on Machine Learning (ICML), 2019. 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 442, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 442, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 442, + 481, + 489 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 489, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 489, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 489, + 481, + 536 + ], + "type": "text", + "content": "[34] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 536, + 481, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 481, + 594 + ], + "type": "text", + "content": "[35] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 595, + 481, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 595, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 133, + 595, + 481, + 630 + ], + "type": "text", + "content": "[36] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. 3, 9, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 630, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 630, + 481, + 665 + ], + "type": "text", + "content": "[37] Eric Zimmermann, Justin Szeto, Jerome Pasquero, and Frederic Ratle. Benchmarking a benchmark: How reliable is ms-coco? arXiv preprint arXiv:2311.02709, 2023. 14" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Singh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_content_list.json b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d07e1c9574d1225da36bb41a33fdc76400f4fbd2 --- /dev/null +++ b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_content_list.json @@ -0,0 +1,1871 @@ +[ + { + "type": "text", + "text": "Benchmarking Spurious Bias in Few-Shot Image Classifiers", + "text_level": 1, + "bbox": [ + 227, + 140, + 779, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guangtao Zheng1, Wenqian Ye1, and Aidong Zhang1", + "bbox": [ + 292, + 210, + 712, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Virginia, Charlottesville VA 22904, USA {gz5hp,wenqian,aidong}@virginia.edu", + "bbox": [ + 316, + 239, + 684, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Few-shot image classifiers are designed to recognize and classify new data with minimal supervision and limited data but often show reliance on spurious correlations between classes and spurious attributes, known as spurious bias. Spurious correlations commonly hold in certain samples and few-shot classifiers can suffer from spurious bias induced from them. There is an absence of an automatic benchmarking system to assess the robustness of few-shot classifiers against spurious bias. In this paper, we propose a systematic and rigorous benchmark framework, termed FewSTAB, to fairly demonstrate and quantify varied degrees of robustness of few-shot classifiers to spurious bias. Few-STAB creates few-shot evaluation tasks with biased attributes so that using them for predictions can demonstrate poor performance. To construct these tasks, we propose attribute-based sample selection strategies based on a pre-trained vision-language model, eliminating the need for manual dataset curation. This allows FewSTAB to automatically benchmark spurious bias using any existing test data. FewSTAB offers evaluation results in a new dimension along with a new design guideline for building robust classifiers. Moreover, it can benchmark spurious bias in varied degrees and enable designs for varied degrees of robustness. Its effectiveness is demonstrated through experiments on ten few-shot learning methods across three datasets. We hope our framework can inspire new designs of robust few-shot classifiers. Our code is available at https://github.com/gtzheng/FewSTAB", + "bbox": [ + 261, + 310, + 743, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Few-shot classification $\\cdot$ Spurious bias $\\cdot$ Robustness $\\cdot$ Benchmark system", + "bbox": [ + 259, + 643, + 740, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 699, + 375, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot classification [5,21,46,61,64] (FSC) has attracted great attention recently due to its promise for recognizing novel classes efficiently with limited data. Few-shot classifiers can transfer the knowledge learned from base classes to recognize novel classes with a few labeled samples. However, they face potential risks when deployed in the real world, such as data distribution shifts [29,62] and adversarial examples [8,14]. A subtle yet critical risk factor is the spurious correlations [3,4,11,13,43,47,59] between classes and spurious attributes —", + "bbox": [ + 212, + 733, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9615ac1307e59941c988adebb0cadcaf3289cd17d9f320a16aedf8d536e05498.jpg", + "image_caption": [ + "Learned path" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 498, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e57ac34f47a9e4f529ee4d44feb0c29dae1b31546867bfbab0f09309258e8b57.jpg", + "image_caption": [ + "Non-existing path", + "Fig. 1: Exploiting the spurious correlation between the class bird and the spurious attribute tree branch to predict bird leads to an incorrect prediction on the test image showing birds on a grass field. For clarity, we only show the case for one class." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 787, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attributes of inputs non-essential to the classes. In the traditional learning setting [1,2,43], deep learning models tend to rely on spurious correlations as their prediction shortcuts or exhibit spurious bias, such as predicting classes using the associated backgrounds [58] or image textures [12], leading to significant performance drops when the associated backgrounds or textures change to different ones. In the low-data regime, spurious bias becomes more evident. For example, in Fig. 1, the correlation between the class bird and the attribute tree branch in the support (training) image may form a shortcut path from tree branch to predicting the image as bird and hinder the learning of the desired one that uses class-related attributes, such as head, tail, and wing. The shortcut will fail to generalize in the query (test) image where no tree branch can be found. In general, few-shot image classifiers are susceptible to spurious bias.", + "bbox": [ + 212, + 323, + 787, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, there lacks a dedicated benchmarking framework that evaluates the robustness of few-shot classifiers to spurious bias. The standard benchmarking procedure in FSC trains a few-shot classifier on base classes from a training set with ample samples and evaluates the classifier on FSC test tasks constructed from a test set with novel classes. The problem with this procedure is the lack of explicit control over the spurious correlations in the constructed FSC tasks. Each FSC test task contains randomly sampled support and query samples. Thus, spurious correlations in the majority of the test set samples can be demonstrated in these tasks, providing unfair advantages for few-shot classifiers with high reliance on the spurious correlations.", + "bbox": [ + 212, + 506, + 789, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a systematic and rigorous benchmark framework, termed Few-Shot Tasks with Attribute Biases (FewSTAB), to fairly compare the robustness of various few-shot classifiers to spurious bias. Our framework explicitly controls spurious correlations in the support and query samples when constructing an FSC test task to reveal the robustness pitfalls caused by spurious bias. To achieve this, we propose attribute-based sample selection strategies that select support and query samples with biased attributes. These attributes together with their associated classes formulate spurious correlations such that if the support samples induce spurious bias in a few-shot classifier, i.e., the classifier learns the spurious correlations in the support samples as its prediction shortcuts, then the query samples can effectively degrade the classifier's performance, exposing its non-robustness to spurious bias.", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Zheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework exploits the spurious attributes in test data for formulating spurious correlations in FSC test tasks. Some existing datasets [15,27,43] provide spurious attribute annotations. However, they only have a few classes and cannot provide enough classes for training and testing. Many benchmark datasets for FSC do not have annotations on spurious attributes, and obtaining these annotations typically involves labor-intensive human-guided labeling [33,66]. To address this, we further propose to use a pre-trained vision-language model (VLM) to automatically identify distinct attributes in images in the high-level text format. Our attribute-based sampling methods can use the identified attributes to simulate various spurious correlations. Thus, we can reuse any existing FSC datasets for benchmarking few-shot classifiers' robustness to spurious bias, eliminating the need for the manual curation of new datasets.", + "bbox": [ + 212, + 146, + 787, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of our work are summarized as follows:", + "bbox": [ + 240, + 328, + 696, + 340 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a systematic and rigorous benchmark framework, termed Few-Shot Tasks with Attribute Biases (FewSTAB), that specifically targets spurious bias in few-shot classifiers, demonstrates their varied degrees of robustness to spurious bias, and benchmarks spurious bias in varied degrees.", + "- We propose novel attribute-based sample selection strategies using a pretrained VLM for constructing few-shot evaluation tasks, allowing us to reuse any existing few-shot benchmark datasets without manually curating new ones for the evaluation.", + "- FewSTAB provides a new dimension of evaluation on the robustness to spurious bias along with a new design guideline for building robust few-shot classifiers. We demonstrate the effectiveness of FewSTAB by applying it to models trained on three benchmark datasets with ten FSC methods." + ], + "bbox": [ + 225, + 349, + 785, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 551, + 387, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot classification. Few-shot classification [6,46,50,51,53,55] has received vast attention recently. Few-shot classifiers can be trained with meta-learning or transfer learning on base classes to learn the knowledge that can be transferred to recognize novel classes with a few labeled samples. The transfer learning approaches [6,50] first learn a good embedding model and then fine-tune the model on samples from novel classes. The meta-learning approaches can be further divided into optimization-based and metric-based methods. The optimization-based methods [9,10,22,26,54] aim to learn a good initialized model such that the model can adapt to novel classes efficiently with a few gradient update steps on a few labeled samples. The metric-based methods [5,21,35,42,49,53,61,65] aim to learn a generalizable representation space with a well-defined metric, such as Euclidean distance [46], to learn novel classes with a few labeled samples. Recently, large vision-language models [20,38,68] are used for few-shot classification. However, they have completely different training and inference pipelines from the models that we consider in this paper.", + "bbox": [ + 212, + 582, + 787, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Robustness in few-shot classification. There are several notions of robustness for few-shot classifiers. The common one requires a few-shot classifier to", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Benchmarking Spurious Bias in Few-Shot Image Classifiers", + "bbox": [ + 338, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "perform well on the in-distribution samples of novel classes in randomly sampled FSC test tasks. The robustness to adversarial perturbations further requires a few-shot classifier to perform well on samples with imperceptible perturbations [8,14]. Moreover, the cross-domain generalization [37,51,52] aims to test how robust a few-shot classifier is on samples from novel classes with domain shifts, which are typically reflected by the changes in both image styles and classes. In contrast, we focus on a new notion of robustness: the robustness to spurious bias. There is a lack of rigorous evaluation methods on the topic. We provide a new evaluation method that specifically targets spurious bias and can systematically demonstrate few-shot classifiers' varied degrees of vulnerability to spurious bias, which has not been addressed in the existing literature.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Benchmarks for spurious bias. There are some existing datasets [15,27,43] that are designed to benchmark spurious bias in image classifiers. However, these datasets are only applicable to the traditional learning setting [1,2,43] since the classes in them are not sufficient for the training and testing of few-shot classifiers. Existing benchmarks in few-shot classification are not tailored for benchmarking spurious bias in few-shot classifiers. A recent work [67] creates a large-scale few-shot classification benchmark dataset with spurious-correlation shifts. In contrast, we propose a benchmark framework that can reuse existing few-shot classification datasets and provide a new dimension of evaluation.", + "bbox": [ + 212, + 311, + 787, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discovering spurious attributes. A spurious attribute is non-essential to a class and only exists in some samples. Early works on discovering spurious attributes [33,66] require a predefined list of spurious attributes and expensive human-guided labeling of visual attributes. Recent works [31,44,45,57] greatly reduce the need for manual annotations by using the neurons of robust models to detect visual attributes. However, they still need humans to annotate the detected visual attributes. We automate this process by using a pre-trained VLM to obtain distinct attributes as words. Instead of discovering spurious correlations, we simulate them via attribute-based sampling for benchmarking.", + "bbox": [ + 212, + 449, + 787, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 609, + 369, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Few-shot classification tasks. A typical FSC task $\\mathcal{T}$ has a support set $S$ for training and a query set $\\mathcal{Q}$ for testing. In this task, there are $C$ classes $(c = 1,\\dots ,C)$ with $N_{\\mathcal{S}}$ (a small number) training samples and $N_{\\mathcal{Q}}$ test samples per class in $S$ and $\\mathcal{Q}$ , respectively. The task is called a $C$ -way $N_{\\mathcal{S}}$ -shot task.", + "bbox": [ + 212, + 642, + 787, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Few-shot classifiers. A few-shot classifier $f_{\\theta}$ with parameters $\\theta$ aims to classify the samples in $\\mathcal{Q}$ after learning from $S$ with a learning algorithm $\\mathcal{O}$ in a few-shot task $\\mathcal{T}$ . Here, $\\mathcal{O}$ could be any learning algorithms, such as the optimization method [9] or a prototype-based classifier learning method [7,46]. To acquire a good few-shot learning capability, $f_{\\theta}$ is typically meta-trained or pre-trained [25] on a base training set $\\mathcal{D}_{train} = \\{(x_n, y_n) | y_n \\in \\mathcal{C}_{train}, n = 1, \\dots, N_{train}\\}$ with $N_{train}$ sample(x)-label(y) pairs, where $\\mathcal{C}_{train}$ is a set of base classes.", + "bbox": [ + 212, + 703, + 787, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Performance metrics. The performance of a few-shot classifier is typically measured by its average classification accuracy over $N_{\\mathcal{T}}$ $C$ -way $N_{S}$ -shot tasks", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ec743651d9cbdab2e1284fae6e5f72e5f74a8e0e2ea35aa4a6e472a36053620e.jpg", + "table_caption": [ + "Table 1: Meanings of major symbols used in the paper." + ], + "table_footnote": [], + "table_body": "
SymbolMeaning
TAn FSC task
SSupport (training) set in T
QQuery (test) set in T
cA class in T
CNumber of classes per task
NsNumber of samples (shots) per class in S
NQNumber of samples per class in Q
OA few-shot adaptation algorithm
ψAn attribute detector
ΩAn automatic word selector
DtrainThe base training set
DvalThe validation set for selecting a few-shot classifier
DtestThe test set for evaluating a few-shot classifier
CtrainClasses in Dtrain
CtestClasses in Dtest
DcA set of all samples belonging to class c
aA text-format attribute
AA set of text-format attributes
", + "bbox": [ + 339, + 171, + 656, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "randomly sampled from $\\mathcal{D}_{test} = \\{(x_n, y_n) | y_n \\in \\mathcal{C}_{test}, n = 1, \\dots, N_{test}\\}$ where $N_{test}$ sample-label pairs from novel classes $\\mathcal{C}_{test}$ do not appear in $\\mathcal{D}_{train}$ , i.e., $\\mathcal{C}_{train} \\cap \\mathcal{C}_{test} = \\emptyset$ . We denote this metric as standard accuracy $\\mathrm{Acc}(f_\\theta)$ , i.e.,", + "bbox": [ + 212, + 420, + 787, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A c c} \\left(f _ {\\theta}\\right) = \\frac {1}{N _ {\\mathcal {T}}} \\sum_ {t = 1} ^ {N _ {\\mathcal {T}}} \\sum_ {c = 1} ^ {C} M _ {c} \\left(\\mathcal {T} _ {t}; f _ {\\theta}, \\mathcal {O}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 477, + 785, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{c}(\\mathcal{T}_{t};f_{\\theta},\\mathcal{O})$ denotes the classification accuracy of $f_{\\theta}$ on the query samples from the class $c$ in $\\mathcal{T}_t$ after $f_{\\theta}$ is trained on $S$ with $\\mathcal{O}$ . The metric $\\mathrm{Acc}(f_{\\theta})$ in Eq. (11) only shows the average learning capability of $f_{\\theta}$ over $C$ randomly selected novel classes. To better characterize the robustness of $f_{\\theta}$ to spurious bias, we define the class-wise worst classification accuracy over tasks as", + "bbox": [ + 212, + 529, + 787, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {w A c c} \\left(f _ {\\theta}\\right) = \\frac {1}{N _ {\\mathcal {T}}} \\sum_ {t = 1} ^ {N _ {\\mathcal {T}}} \\min _ {c = 1, \\dots , C} M _ {c} \\left(\\mathcal {T} _ {t}; f _ {\\theta}, \\mathcal {O}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 616, + 785, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A larger $\\mathrm{wAcc}(f_{\\theta})$ indicates that $f_{\\theta}$ is more robust to spurious bias.", + "bbox": [ + 212, + 667, + 702, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spurious correlations. A spurious correlation is the association between a class and an attribute of inputs that is non-essential to the class, and it only holds in some samples. We formally define it as follows.", + "bbox": [ + 212, + 684, + 787, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 1. Let $\\mathcal{D}_c$ denote a set of sample-label pairs having the label $c$ , and let $\\psi : \\mathcal{X} \\to \\mathcal{B}_A$ be an attribute detector, where $\\mathcal{X}$ is the set of all possible inputs, $\\mathcal{B}_A$ denotes all possible subsets of $\\mathcal{A}$ , and $\\mathcal{A}$ is the set of all possible attributes. The class $c$ and an attribute $a \\in \\mathcal{A}$ form a spurious correlation, denoted as $\\langle c, a \\rangle$ , if and only if the following conditions hold:", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. There exists $(x,c)\\in \\mathcal{D}_c$ that satisfies $a\\in \\psi (x)$ , and", + "bbox": [ + 222, + 824, + 617, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Benchmarking Spurious Bias in Few-Shot Image Classifiers", + "bbox": [ + 338, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3ec32a50922a691a3cb03923c1de7f2bb667cfe8e4bd69072773176976fd3941.jpg", + "image_caption": [ + "Fig. 2: FewSTAB overview. (a) Extract distinct attributes using a pre-trained VLM. (b) Generate an FSC task for the evaluation of spurious bias in few-shot classifiers." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 785, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2. There exists $(x', c) \\in \\mathcal{D}_c$ that satisfies $a \\notin \\psi(x')$ .", + "text_level": 1, + "bbox": [ + 220, + 401, + 589, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We define $a$ as the spurious attribute in $\\langle c, a \\rangle$ .", + "bbox": [ + 217, + 425, + 570, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition $\\mathbb{I}$ specifies that all the spurious correlations are based on $\\mathcal{D}_c$ . In the remainder of the paper, we define $\\mathcal{D}_c = \\{(x,c)|\\forall (x,c)\\in \\mathcal{D}_{test}\\}$ with $c\\in \\mathcal{C}_{test}$ as we focus on evaluating the robustness to spurious bias.", + "bbox": [ + 212, + 450, + 784, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We list major symbols in the paper alongside their meanings in Tab. 1.", + "bbox": [ + 238, + 496, + 751, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 534, + 379, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Attribute-Based Sample Selection", + "text_level": 1, + "bbox": [ + 215, + 565, + 540, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first propose two attribute-based sample selection methods to reveal spurious bias in a few-shot classifier. Consider a training set $\\mathcal{S}$ in a few-shot test task $\\mathcal{T}$ , which has $C$ classes with each class $c\\in \\mathcal{C}_{test}$ associating with a unique spurious attribute $a\\in \\mathcal{A}$ . We aim to discover samples that can exhibit a classifier's spurious bias on $\\langle c,a\\rangle$ induced from $\\mathcal{S}$ . Motivated by existing findings [43,59,60] that classifiers with high reliance on $\\langle c,a\\rangle$ tend to perform poorly on samples without it, we propose an attribute-based sample selection strategy below.", + "bbox": [ + 212, + 590, + 787, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intra-class attribute-based sample selection. Given $\\mathcal{D}_c$ and the training set $S$ having the spurious correlation $\\langle c, a \\rangle$ , we generate a set $\\mathcal{I}_{\\langle c, a \\rangle}$ of sample-label pairs which have class $c$ but do not contain attribute $a$ , i.e.,", + "bbox": [ + 212, + 696, + 787, + 742 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\langle c, a \\rangle} = \\{(x, c) | \\forall (x, c) \\in \\mathcal {D} _ {c}, a \\notin \\psi (x) \\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 752, + 785, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above proposed method demonstrates a few-shot classifier's robustness to individual spurious correlation $\\langle c, a \\rangle$ and does not consider a multi-class classification setting where spurious attributes from some other class $c'$ exist in samples of the class $c$ . In this case, these attributes may mislead the classifier to", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Zheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "predict those samples as the class $c'$ and severely degrade the performance on the class $c$ . For example, consider using the spurious correlations $\\langle \\text{vase}, \\text{blue} \\rangle$ and $\\langle \\text{bowl}, \\text{green} \\rangle$ for predicting vase and bowl, respectively. An image showing a vase in green is more effective in revealing the robustness to $\\langle \\text{vase}, \\text{blue} \\rangle$ as it is more likely to be misclassified as bowl than other images. Motivated by this, we propose the inter-class attribute-based sample selection below.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inter-class attribute-based sample selection. Given $\\mathcal{D}_c$ and the training set $S$ having the spurious correlations $\\langle c, a \\rangle$ and $\\langle c', a' \\rangle$ , where $c' \\neq c$ and $a' \\neq a$ , we generate a set $\\mathcal{I}_{\\langle c, a \\rangle}^{\\langle c', a' \\rangle}$ of sample-label pairs which have class $c$ , do not contain attribute $a$ , but contain attribute $a'$ from another class $c'$ :", + "bbox": [ + 212, + 237, + 782, + 301 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\langle c, a \\rangle} ^ {\\langle c ^ {\\prime}, a ^ {\\prime} \\rangle} = \\mathcal {I} _ {\\langle c, a \\rangle} \\cap \\left\\{\\left(x, c\\right) | \\forall (x, c) \\in \\mathcal {D} _ {c}, a _ {\\langle c ^ {\\prime} \\rangle} ^ {\\prime} \\in \\psi (x) \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 314, + 784, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $a_{\\langle c^{\\prime}\\rangle}^{\\prime}$ denotes $a^\\prime$ in $\\langle c',a'\\rangle$ , and $\\mathcal{I}_{\\langle c,a\\rangle}$ is defined in Eq. (3).", + "bbox": [ + 212, + 347, + 673, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Considering that there are $C$ classes in the training set $S$ with each class associating with a unique spurious attribute $a$ , to effectively demonstrate the reliance on the spurious correlation $\\langle c, a \\rangle$ with the inter-class attribute-based sample selection, we consider all the spurious correlations in $S$ . Specifically, we apply the above selection strategy to all the $C - 1$ spurious correlations in $S$ other than $\\langle c, a \\rangle$ and obtain $\\mathcal{I}_{\\langle c, a \\rangle}^{C}$ as the union of the $C - 1$ sets as follows:", + "bbox": [ + 212, + 364, + 784, + 457 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\langle c, a \\rangle} ^ {C} = \\bigcup_ {\\langle c ^ {\\prime}, a ^ {\\prime} \\rangle \\in \\mathcal {C} \\backslash c} \\mathcal {I} _ {\\langle c, a \\rangle} ^ {\\langle c ^ {\\prime}, a ^ {\\prime} \\rangle}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 469, + 784, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_{\\backslash c}$ denotes all the spurious correlations in $S$ other than $\\langle c, a \\rangle$ .", + "bbox": [ + 212, + 517, + 712, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The inter-class attribute-based sample selection is built upon the intra-class attribute-based sample selection. In the remainder of the paper, we use the inter-class method as our default sample selection strategy, which is more effective empirically (Sec. 5.6). In certain cases, however, where there are not enough desired samples during task construction, we resort to the intra-class sample selection strategy (Sec. 5.1, Implementation details).", + "bbox": [ + 212, + 532, + 784, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following, we introduce FewSTAB, a benchmark framework that uses the proposed selection strategies to construct FSC tasks containing samples with biased attributes for benchmarking spurious bias in few-shot classifiers.", + "bbox": [ + 212, + 623, + 782, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 FewSTAB (Part 1): Text-Based Attribute Detection", + "text_level": 1, + "bbox": [ + 214, + 691, + 689, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our attribute-based sample selection methods require knowing the attributes in images, which typically involves labor-intensive human labeling. To make our method scalable and applicable to few-shot classifiers trained on different datasets, we adopt a pre-trained VLM to automatically identify distinct attributes in images in text format, which includes the following two steps.", + "bbox": [ + 212, + 719, + 782, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step 1: Generating text descriptions. We use a pre-trained VLM [23, 32] $\\phi$ to automatically generate text descriptions for images in $\\mathcal{D}_{test}$ . The VLM is a model in the general domain and can produce text descriptions for various", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Benchmarking Spurious Bias in Few-Shot Image Classifiers", + "bbox": [ + 338, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "objects and patterns. For example, for the current input image in the vase class in Fig. 2(a), besides the class object vase, the VLM also detects the vase's color green, and another object table with its material wooden.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 2: Extracting informative words. From the generated text descriptions, we extract nouns and adjectives as the detected attributes via an automatic procedure $\\Omega$ . The two kinds of words are informative as a noun describes an object, and an adjective describes a property of an object. All the detected attributes form the candidate attribute set $\\mathcal{A}$ . We realize the attribute detector $\\psi$ defined in Definition 1 as $\\psi(x) = \\Omega(\\phi(x))$ .", + "bbox": [ + 212, + 191, + 784, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark 1: A VLM in general can extract many distinct attributes from the images. On some images, the VLM may detect non-relevant attributes, such as detecting a duck from a bird image. A more capable VLM could warrant a better attribute detection accuracy and benefit individual measurements on few-shot classifiers. Although being a VLM-dependent benchmark framework, FewSTAB can produce consistent and robust relative measurements among all the compared FSC methods, regardless the choice of VLMs (Sec. 5.6).", + "bbox": [ + 212, + 284, + 784, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark 2: The candidate set $\\mathcal{A}$ constructed with all the extracted words may contain attributes that represent the classes in $\\mathcal{D}_{test}$ . However, during our attribute-based sample selection, these attributes will not be used since they always correlate with classes and therefore do not satisfy the definition of spurious attributes in Definition 1. We provide details of $\\phi$ and $\\Omega$ in Sec. 5.1", + "bbox": [ + 212, + 390, + 784, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 FewSTAB (Part 2): FSC Task Construction", + "text_level": 1, + "bbox": [ + 214, + 491, + 620, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Constructing a $C$ -way $N_{S}$ -shot FSC task $\\mathcal{T}$ for benchmarking spurious bias in few-shot classifiers involves constructing a support (training) set $\\mathcal{S}$ and a query (test) $\\mathcal{Q}$ with biased attributes.", + "bbox": [ + 212, + 518, + 782, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Constructing the support set. The support set contains the spurious correlations that we aim to demonstrate to a few-shot classifier. As a fair and rigorous benchmark system, FewSTAB makes no assumptions on the few-shot classifiers being tested and randomly samples $C$ classes from $\\mathcal{C}_{test}$ . For each sampled class, it randomly selects a spurious correlation $\\langle c, a \\rangle$ in $\\mathcal{D}_{test}$ with $a \\in \\mathcal{A}$ . To effectively demonstrate the spurious correlation $\\langle c, a \\rangle$ to a few-shot classifier, we select samples of the class $c$ such that (1) they all have the spurious attribute $a$ and (2) do not have spurious attributes from the other $C - 1$ spurious correlations. We construct $S_{c}$ with $N_{S}$ samples for the class $c$ that satisfy the above two conditions. Thus, the spurious attribute $a$ becomes predictive of the class $c$ in $S_{c}$ . We take the union of all $C$ such sets to get $S = \\cup_{c=1}^{C} S_{c}$ . Fig. 2(b) demonstrates the case when $C = 3$ . Note that we have no requirements for other non-selected attributes in $A$ to ensure that we have enough samples for $S_{c}$ .", + "bbox": [ + 212, + 565, + 784, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Constructing the query set. To evaluate the robustness to the spurious correlations formulated in $\\mathcal{S}$ , we first construct a candidate set $\\mathcal{I}_{\\langle c,a\\rangle}^{C}$ in Eq. (5) for each spurious correlation $\\langle c,a\\rangle$ in $\\mathcal{S}$ . Since we have no requirements on the non-selected attributes that are not used to formulate spurious correlations in $\\mathcal{S}$ , a few-shot classifier may predict query samples via some of these attributes,", + "bbox": [ + 212, + 762, + 784, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "e.g., the yellow blocks in Fig. 2(b), bypassing the test on the formulated spurious correlations in $\\mathcal{S}$ . To address this, we propose query sample selection below.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Query sample selection: We select query samples from $\\mathcal{I}_{\\langle c,a\\rangle}^{C}$ that are least likely to have non-selected spurious attributes, such as the ones enclosed with red boxes in Fig. 2(b). To achieve this, we first calculate the fraction of sample-label pairs in $\\mathcal{I}_{\\langle c,a\\rangle}^{C}$ that have the attribute $a$ as", + "bbox": [ + 212, + 176, + 782, + 239 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\np _ {a} = | \\{x \\mid a \\in \\psi (x), \\forall (x, c) \\in \\mathcal {I} _ {\\langle c, a \\rangle} ^ {C} \\} | / | \\mathcal {I} _ {\\langle c, a \\rangle} ^ {C} |, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 250, + 785, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $|\\cdot|$ denotes the size of a set, $a \\in \\tilde{\\mathcal{A}}$ , and $\\tilde{\\mathcal{A}}$ contains all non-selected attributes. A larger $p_a$ indicates that the attribute $a$ occurs more frequently in data and is more likely to be used in formulating prediction shortcuts. We then calculate the likelihood score for each $(x,c) \\in \\mathcal{I}_{\\langle c,a\\rangle}^C$ as $s(x) = \\sum_{a \\in \\psi(x), a \\in \\tilde{\\mathcal{A}}} p_a$ , i.e., the summation of all $p_a$ of non-selected attributes in $x$ . The likelihood score will be zero if there are no non-selected attributes in $x$ . A large $s(x)$ indicates that the image $x$ can be predicted via many non-selected attributes. Therefore, we select $N_{\\mathcal{Q}}$ samples from $\\mathcal{I}_{\\langle c,a\\rangle}^C$ that have the lowest likelihood scores to construct $\\mathcal{Q}_c$ . Then, we have $\\mathcal{Q} = \\cup_{c=1}^{C} \\mathcal{Q}_c$ , which contains samples for evaluating the robustness of a few-shot classifier to the spurious correlations in $\\mathcal{S}$ .", + "bbox": [ + 212, + 277, + 784, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Complexity analysis. The text-based attribute detection only needs to use VLMs once to extract attributes for each test set of a dataset. For the task construction, in a nutshell, we analyze the attributes of samples from each of the $C$ classes and do the sampling. Thus, the computational complexity is $O(N_{\\mathcal{T}}CN_cN_{\\mathcal{A}})$ , where $N_{c}$ is the maximum number of samples per class in test data, $N_{\\mathcal{A}}$ is the number of extracted attributes. We only need to run the process once and use the generated tasks to benchmark various models.", + "bbox": [ + 212, + 433, + 782, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 560, + 375, + 578 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 590, + 428, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We used two general datasets of different scales, miniImageNet [40] and tieredImageNet [41], and one fine-grained dataset, CUB-200 [56]. Each dataset consists of $\\mathcal{D}_{train}$ , $\\mathcal{D}_{val}$ , and $\\mathcal{D}_{test}$ for training, validation, and test, respectively (see Appendix). All images were resized to $84 \\times 84$ .", + "bbox": [ + 212, + 612, + 782, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "FSC methods. We trained FSC models with ten algorithms covering three major categories. For gradient-based meta-learning algorithms, we chose ANIL [39], LEO [42], and BOIL [34]. For metric-based meta-learning algorithms, we chose ProtoNet [46], DN4 [24], R2D2 [5], CAN [16], and RENet [19]. For transfer learning algorithms, we chose Baseline++ [6] and RFS [50]. See Appendix for more details. Any backbones can be used as the feature extractor. For fair comparisons between different methods, we used the ResNet-12 backbone adopted in [35].", + "bbox": [ + 212, + 672, + 784, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Text-based attribute detection. We used a pre-trained VLM named ViT-GPT2 [32] to generate text descriptions for images in $\\mathcal{D}_{test}$ . After that, we used Spacy (https://spacy.io/) to extract nouns and adjectives from these descriptions automatically. We also used another pre-trained VLM, BLIP [23], to test", + "bbox": [ + 212, + 780, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Benchmarking Spurious Bias in Few-Shot Image Classifiers", + "bbox": [ + 338, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5c20361fd4e320960818b3046b8f12d0bff4a72d0065c59852c83eb4369634c3.jpg", + "table_caption": [ + "Table 2: Statistics of detected attributes in ${\\mathcal{D}}_{\\text{test }}$ by two VLMs." + ], + "table_footnote": [], + "table_body": "
VLMUnique attributesAvg. attributes per class
miniImageNettieredImageNetCUB-200miniImageNettieredImageNetCUB-200
ViT-GPT211112532159190.40230.9425.78
BLIP20326710247254.40310.4029.74
", + "bbox": [ + 241, + 170, + 756, + 223 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3439fbe77ecf1db06516627fe54b44d47f8e42407bec302e739ba779246afc67.jpg", + "image_caption": [ + "Fig. 3: A 5-way 1-shot task constructed by our inter-class attribute-based sample selection using samples from the miniImageNet dataset. Note that due to the limited capacity of a VLM, the attributes may not well align with human understandings." + ], + "image_footnote": [], + "bbox": [ + 217, + 233, + 352, + 406 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/11d3cc987d5fc106a29b22db9362ebbb7fe3b46aaca79791304fbe4ec5f252de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 233, + 785, + 406 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "whether FewSTAB can produce consistent results. The statistics of the detected attributes are shown in Table 2.", + "bbox": [ + 212, + 488, + 787, + 517 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. We trained FSC models with the implementation in [25]. Each model was trained on $\\mathcal{D}_{train}$ of a dataset with one of the ten FSC methods. For each meta-learning based method, we trained two models using randomly sampled 5-way 1-shot and 5-way 5-shot tasks, respectively. All the tasks have 15 samples per class in the query set. We saved the model that achieves the best validation accuracy on $\\mathcal{D}_{val}$ for evaluation. For FewSTAB, if we do not have enough desired samples to construct a support set, we redo the construction from the beginning. If there are not enough desired samples to construct a query set, we first try to use the intra-class attribute-based sample selection; if the desired samples are still not enough, we redo the construction from the beginning. We created 3000 tasks for model evaluation. All experiments were conducted on the NVIDIA RTX 8000 GPUs.", + "bbox": [ + 212, + 518, + 789, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Visualization of a Constructed Task", + "text_level": 1, + "bbox": [ + 214, + 723, + 558, + 738 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We show a 5-way 1-shot task constructed by FewSTAB in Fig. 3. Each class in the support set correlates with a unique spurious attribute. The query samples of a class do not have the spurious attribute correlated with the class and some of them have spurious attributes associated with other classes in the support set. For example, the query samples of the class lion do not have the spurious attribute water, and some of them have spurious attributes from other classes in", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Zheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5dbbd9a19d935159bc69b7007d19839678e332f4aa0ff43f5039d8b3d8c81583.jpg", + "table_caption": [ + "Table 3: Comparison between wAcc-R and wAcc-A with $95\\%$ confidence interval on the miniImageNet, tieredImageNet, and CUB datasets. Numbers in the Shot column indicate that the models are both trained (if applicable) and tested on 5-way 1- or 5-shot tasks. Darker colors indicate higher values." + ], + "table_footnote": [], + "table_body": "
ShotMethodminiImageNettieredImageNetCUB-200
wAcc-A (↑)wAcc-R (↑)wAcc-A (↑)wAcc-R (↑)wAcc-A (↑)wAcc-R (↑)
1ANIL10.38±0.3014.36±0.3311.21±0.3015.63±0.3613.78±0.4016.94±0.43
LEO14.26±0.4621.35±0.5416.00±0.5529.63±0.7128.29±0.8040.22±0.87
BOIL12.48±0.2314.93±0.2412.27±0.2114.13±0.2319.15±0.2922.50±0.29
ProtoNet14.03±0.4921.96±0.5814.50±0.5027.13±0.6934.62±0.8546.61±0.89
DN412.37±0.4619.28±0.5611.99±0.4723.62±0.6535.22±0.8647.26±0.88
R2D218.05±0.5326.50±0.6016.41±0.5430.41±0.7136.70±0.9048.82±0.88
CAN17.37±0.5325.96±0.6018.84±0.6036.29±0.7822.74±0.7231.95±0.78
RENet19.10±0.5728.85±0.6518.83±0.6135.70±0.7832.43±0.8143.98±0.81
Baseline++15.30±0.4823.18±0.5617.51±0.5431.62±0.719.17±0.4714.59±0.58
RFS18.00±0.5327.12±0.6118.35±0.6035.24±0.7732.45±0.8044.49±0.82
5ANIL14.83±0.4025.37±0.5213.72±0.3930.60±0.5131.63±0.5545.47±0.53
LEO26.31±0.5941.33±0.5929.49±0.7257.22±0.7246.62±0.8259.76±0.77
BOIL13.09±0.2215.21±0.2314.90±0.2218.55±0.2419.17±0.2821.33±0.27
ProtoNet32.07±0.5851.95±0.5230.95±0.7062.53±0.6260.06±0.7475.68±0.50
DN427.60±0.5842.68±0.6216.07±0.6240.63±0.8159.25±0.7773.58±0.56
R2D235.37±0.5950.84±0.5531.12±0.7261.08±0.6558.66±0.8275.20±0.54
CAN36.44±0.6554.23±0.5531.17±0.7664.19±0.6241.31±0.7461.61±0.58
RENet36.19±0.6356.52±0.5830.27±0.7663.49±0.6452.93±0.8271.82±0.56
Baseline++29.52±0.5744.94±0.5730.01±0.7259.06±0.6716.86±0.5229.84±0.69
RFS36.85±0.6455.66±0.5531.15±0.7662.71±0.6754.98±0.8174.33±0.53
", + "bbox": [ + 271, + 212, + 728, + 435 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/77a621976c9e3e83be3775b669456e11c5b87d0e7f3f61174569f043553fa85f.jpg", + "image_caption": [ + "Fig. 4: Accuracy gaps (wAcc-R minus wAcc-A) on the 5-way 1-shot and 5-way 5-shot tasks from the (a) miniImageNet, (b) tieredImageNet, and (c) CUB-200 datasets." + ], + "image_footnote": [], + "bbox": [ + 215, + 446, + 403, + 546 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ff8b411fada5e4286329789d29cc4a0ff250a1219ebec889c5253dfb42285f6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 446, + 594, + 545 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3881f6eb4bb4aa3e8517b8b3565d61a8367db67a916c12552c64a931f8a738fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 446, + 787, + 545 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the support set, such as man and green. FewSTAB introduces biased attributes in the task so that query samples can be easily misclassified as other classes by a few-shot classifier that relies on the spurious correlations in the support set.", + "bbox": [ + 212, + 613, + 787, + 659 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Effectiveness of FewSTAB", + "text_level": 1, + "bbox": [ + 214, + 679, + 477, + 694 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "FewSTAB can effectively reveal spurious bias in few-shot classifiers. We show in Table 3 the wAcc (Eq. 2) on 5-way 1/5-shot test tasks that are randomly sampled (wAcc-R) and are constructed with FewSTAB (wAcc-A), respectively. FewSTAB generates FSC test tasks only based on the class-attribute correlations in data. In each test setting, the FSC methods in Table 3 are evaluated with the same FSC tasks. We observe that wAcc-A is consistently lower than wAcc-R on the three datasets and on two test-shot numbers, showing that FewSTAB is more effective than the standard evaluation procedure (random task construction) in exhibiting the spurious bias in various few-shot classifiers. We", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Benchmarking Spurious Bias in Few-Shot Image Classifiers", + "bbox": [ + 339, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "additionally show that FewSTAB also works on the most recent FSC methods and can reflect the improvement made to mitigate spurious bias (see Appendix).", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "FewSTAB reveals new robustness patterns among FSC methods. In Table 4, we calculate the Spearman's rank correlation coefficients [30] between the values of wAcc-R and wAcc-A from Table 3. The coefficients are bounded from 0 to 1, with larger values indicating that the ranks of FSC methods based on wAcc-R are more similar to those based on wAcc-A. In the 1-shot setting, it is not effective to control the spurious correlations since we only have one sample per class in the support set. Hence, the coefficients are large, and", + "bbox": [ + 212, + 178, + 545, + 358 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e1f8252cff38797152a1d7d2ebb25929072dfb1d32b2328ebbcac0236704542c.jpg", + "table_caption": [ + "Table 4: Spearman's rank correlations between wAcc-A and wAcc-R in Table 3" + ], + "table_footnote": [], + "table_body": "
Dataset1-shot5-shot
miniImageNet0.960.95
tieredImageNet0.960.90
CUB-2001.000.94
", + "bbox": [ + 576, + 266, + 761, + 325 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the ranks based on wAcc-A are similar to those based on wAcc-R. In the 5-shot cases, we have more samples to demonstrate the spurious correlations. The coefficients become smaller, i.e., the ranks based on wAcc-A show different trends from those based on wAcc-R. In this case, FewSTAB reveals new information on FSC methods' varied degrees of robustness to spurious bias.", + "bbox": [ + 212, + 359, + 787, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "FewSTAB can benchmark spurious bias in varied degrees. As shown in Fig. 4, the accuracy gap, defined as wAcc-R minus wAcc-A, in general, becomes larger when we switch from 5-way 1-shot to 5-way 5-shot tasks. Compared with the random task construction, FewSTAB creates more challenging tasks in the 5-shot case for demonstrating spurious bias in few-shot classifiers. In other words, with a higher shot value in the constructed test tasks, FewSTAB aims to benchmark spurious bias in a higher degree.", + "bbox": [ + 212, + 435, + 787, + 541 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 A New Dimension of Evaluation and a New Design Guideline", + "text_level": 1, + "bbox": [ + 214, + 566, + 769, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "FewSTAB creates a new dimension of evaluation on the robustness to spurious bias. We demonstrate this with a scatter plot (Fig. 5) of Acc (Eq. 11) and wAcc-A of the ten few-shot classifiers. FewSTAB offers new information regarding different few-shot classifiers' robustness to spurious bias as we observe that Acc does not well correlate with wAcc-A. A high wAcc-A indicates that the classifier is robust to spurious bias, while a high Acc indicates that the classifier can correctly predict most of the samples. With the scatter plot, we can view tradeoffs between the two metrics on existing few-shot classifiers. A desirable few-shot classifier should appear in the top-right corner of the plot.", + "bbox": [ + 212, + 598, + 545, + 840 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/22bcc91d8e622a747deb3d0e84d3396166351e12be6c55259b2e70e05221e339.jpg", + "image_caption": [ + "Fig.5: Acc versus wAcc-A of the ten FSC methods tested on 5-way 5-shot tasks from miniImageNet." + ], + "image_footnote": [], + "bbox": [ + 566, + 628, + 774, + 732 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Zheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.5 FewSTAB Enables Designs for Varied Degrees of Robustness", + "text_level": 1, + "bbox": [ + 215, + 145, + 763, + 162 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As demonstrated in Section 5.3, Few-STAB can benchmark spurious bias in varied degrees, which in turn enables practitioners to design robust few-shot classifiers targeted for different degrees of robustness to spurious bias. The reason for differentiating designs for varied degrees of robustness is that the same design choice may not work under different robustness requirements. For example, increasing shot number in training tasks is a common strategy for improving the few-shot generalization of meta-learning based methods. We trained few-shot classifiers with 5-way 5-shot and 5-way 1-shot training tasks randomly sampled from $\\mathcal{D}_{train}$ , respectively. We then calculated the accuracy gap defined as the wAcc-A of a model trained on 5-shot tasks minus the wAcc-A of the same model trained on 1-shot tasks. A positive and large accuracy gap indicates that this strat", + "bbox": [ + 212, + 176, + 488, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "egy is effective in improving the model's robustness to spurious bias. In Fig. 6, on each of the three datasets, we give results of the eight meta-learning based FSC methods on the 5-way 1-, 5-, and 10-shot FewSTAB tasks which are used to demonstrate the strategy's robustness to increased degrees of spurious bias. This strategy does not work consistently under different test shots. For example, in Fig. 6(a) this strategy with CAN only works the best on the 5-way 5-shot FewSTAB tasks.", + "bbox": [ + 212, + 539, + 787, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.6 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 672, + 401, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Techniques used in FewSTAB. We analyze how different sample selection methods affect the effectiveness of FewSTAB in Table 5. With only intra-class attribute-based sample selection, we randomly select query samples from Eq. ③. For inter-class attribute-based sample selection and intra-class attribute-based sample selection (automatically included by Eq. ④), we randomly select query samples from Eq. ⑤. FewSTAB uses all the techniques in Table 5. We define accuracy drop as wAcc-R minus wAcc-A, and we use the drop averaged over the ten FSC methods tested on 5-way 5-shot tasks from the miniImageNet dataset as our metric. A larger average drop indicates that the corresponding", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/43d7254cf725f38714ac47e3765cc3b0384c00334b8338f89dbb13f0aae9c173.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 208, + 784, + 280 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cdff529d27acec6e59d9fbb595167f898308ae8dc75e2e5ba3bb866824d8dd52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 280, + 784, + 349 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/54d76d0f09429176594ef571243424c2bcd7ee74c8dba207014c1a8d0a032151.jpg", + "image_caption": [ + "Fig. 6: Accuracy gaps of few-shot classifiers tested on 1-shot, 5-shot, and 10-shot tasks constructed from (a) miniImageNet, (b) tieredImageNet, and (c) CUB-200 datasets." + ], + "image_footnote": [], + "bbox": [ + 498, + 351, + 785, + 434 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Benchmarking Spurious Bias in Few-Shot Image Classifiers", + "bbox": [ + 338, + 113, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f7ef8f68554fea56d7f1931c75e4797d356ee2a548e15b69ea7b33004676144f.jpg", + "table_caption": [ + "Table 5: Comparison between different techniques used by FewSTAB for constructing FSC tasks." + ], + "table_footnote": [], + "table_body": "
Attribute-based sample selectionQuery sample selectionAvg. drop (%)
Intra-classInter-class
✓5.13
✓✓13.30
✓✓✓15.05
", + "bbox": [ + 241, + 198, + 475, + 272 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/b55fe173cb225a16abb08662ced1dc7dbd76afcbae0b281214b16ba2bf666c03.jpg", + "table_caption": [ + "Table 6: Spearman's rank correlation coefficients between wAcc-A obtained using ViT-GPT2 and BLIP." + ], + "table_footnote": [], + "table_body": "
Dataset1-shot5-shot
miniImageNet0.981.00
tieredImageNet1.000.99
CUB-2001.000.98
", + "bbox": [ + 575, + 199, + 746, + 262 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "sample selection method is more effective in reflecting the spurious bias in few-shot classifiers. We observe that all proposed techniques are effective and the inter-class attribute-based sample selection is the most effective method.", + "bbox": [ + 212, + 300, + 785, + 344 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Choice of VLMs. Although our main results are based on the pre-trained ViT-GPT2 model [32], we show in Table 6 that when switching to a different VLM, i.e., BLIP [23], the relative ranks of different few-shot classifiers based on wAcc-A still hold with high correlations. In other words, FewSTAB is robust to different choices of VLMs.", + "bbox": [ + 212, + 345, + 787, + 419 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Detection accuracy of VLMs. A VLM may miss some attributes due to its limited capacity, resulting in a small detection accuracy. However, the detection accuracy of a VLM has little impact on our framework. To demonstrate this, we adopt a cross-validation strategy, i.e., we use the outputs from one VLM as the ground truth to evaluate those from another VLM, since assessing the detection accuracy of a VLM typically requires labor-intensive human labeling. On the CUB-200 dataset, we observe that the detection accuracy of ViT-GPT2 based on the BLIP's outputs is $70.12\\%$ , while the detection accuracy of BLIP based on the ViT-GPT2's outputs is $59.28\\%$ . Although the two VLMs differ significantly in the detected attributes, our framework shows almost consistent rankings of the evaluated FSC methods (Table 6).", + "bbox": [ + 212, + 421, + 787, + 587 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Additional results are presented in Appendix.", + "bbox": [ + 215, + 587, + 545, + 602 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 626, + 359, + 642 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we proposed a systematic and rigorous benchmark framework called FewSTAB for evaluating the robustness of few-shot classifiers to spurious bias. FewSTAB adopts attribute-based sample selection strategies to construct FSC test tasks with biased attributes so that the reliance on spurious correlations can be effectively revealed. FewSTAB can automatically benchmark spurious bias in few-shot classifiers on any existing test data thanks to its use of a pretrained VLM for automated attribute detection. With FewSTAB, we provided a new dimension of evaluation on the robustness of few-shot image classifiers to spurious bias and a new design guideline for building robust few-shot classifiers. FewSTAB can reveal and enable designs for varied degrees of robustness to spurious bias. We hope FewSTAB will inspire new developments on designing robust few-shot classifiers.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Zheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported in part by the US National Science Foundation under grants 2313865, 2217071, 2213700, 2106913, 2008208, 1955151.", + "bbox": [ + 215, + 178, + 784, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 234, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahmed, F., Bengio, Y., Van Seijen, H., Courville, A.: Systematic generalisation with group invariant predictions. In: International Conference on Learning Representations (2020)", + "2. Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.: Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019)", + "3. Baker, N., Lu, H., Erlikhman, G., Kellman, P.J.: Deep convolutional networks do not classify based on global object shape. PLoS Computational Biology 14(12), e1006613 (2018)", + "4. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: European Conference on Computer Vision. pp. 456-473 (2018)", + "5. Bertinetto, L., Henriques, J.F., Torr, P., Vedaldi, A.: Meta-learning with differentiable closed-form solvers. In: International Conference on Learning Representations (2019)", + "6. Chen, W.Y., Liu, Y.C., Kira, Z., Wang, Y.C.F., Huang, J.B.: A closer look at few-shot classification. In: International Conference on Learning Representations (2019)", + "7. Chen, Y., Wang, X., Liu, Z., Xu, H., Darrell, T.: A new meta-baseline for few-shot learning. arXiv preprint arXiv:2003.04390 (2020)", + "8. Dong, J., Wang, Y., Lai, J.H., Xie, X.: Improving adversarially robust few-shot image classification with generalizable representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9025-9034 (2022)", + "9. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Proceedings of the 34th International Conference on Machine Learning. vol. 70, pp. 1126–1135 (2017)", + "0. Flennerhag, S., Rusu, A.A., Pascanu, R., Visin, F., Yin, H., Hadsell, R.: Meta-learning with warped gradient descent. In: International Conference on Learning Representations (2020)", + "1. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2019)", + "12. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2019)", + "13. Ghosal, S.S., Li, Y.: Are vision transformers robust to spurious correlations? International Journal of Computer Vision 132(3), 689-709 (2024)", + "14. Goldblum, M., Fowl, L., Goldstein, T.: Adversarily robust few-shot learning: A meta-learning approach. Advances in Neural Information Processing Systems 33, 17886-17895 (2020)" + ], + "bbox": [ + 218, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Benchmarking Spurious Bias in Few-Shot Image Classifiers", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. He, Y., Shen, Z., Cui, P.: Towards non-iid image classification: A dataset and baselines. Pattern Recognition 110, 107383 (2021)", + "16. Hou, R., Chang, H., Bingpeng, M., Shan, S., Chen, X.: Cross attention network for few-shot classification. In: Advances in Neural Information Processing Systems. pp. 4003-4014 (2019)", + "17. Hu, S.X., Moreno, P.G., Xiao, Y., Shen, X., Obozinski, G., Lawrence, N., Dami-anou, A.: Empirical bayes transductive meta-learning with synthetic gradients. In: International Conference on Learning Representations (2020)", + "18. Jang, H., Lee, H., Shin, J.: Unsupervised meta-learning via few-shot pseudosupervised contrastive learning. In: International Conference on Learning Representations (2023)", + "19. Kang, D., Kwon, H., Min, J., Cho, M.: Relational embedding for few-shot classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8822-8833 (2021)", + "20. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023)", + "21. Lee, K., Maji, S., Ravichandran, A., Soatto, S.: Meta-learning with differentiable convex optimization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10657-10665 (2019)", + "22. Lee, Y., Choi, S.: Gradient-based meta-learning with learned layerwise metric and subspace. In: International Conference on Machine Learning. pp. 2927-2936. PMLR (2018)", + "23. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "24. Li, W., Wang, L., Xu, J., Huo, J., Gao, Y., Luo, J.: Revisiting local descriptor based image-to-class measure for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7260-7268 (2019)", + "25. Li, W., Wang, Z., Yang, X., Dong, C., Tian, P., Qin, T., Jing, H., Shi, Y., Wang, L., Gao, Y., Luo, J.: Libfewshot: A comprehensive library for few-shot learning. IEEE Transactions on Pattern Analysis & Machine Intelligence (01), 1-18 (2023)", + "26. Li, Z., Zhou, F., Chen, F., Li, H.: Meta-SGD: Learning to learn quickly for few shot learning. ArXiv abs/1707.09835 (2017)", + "27. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3730-3738 (2015)", + "28. Lu, Y., Wen, L., Liu, J., Liu, Y., Tian, X.: Self-supervision can be a good few-shot learner. In: European Conference on Computer Vision. pp. 740-758. Springer (2022)", + "29. Motiian, S., Jones, Q., Iranmanesh, S., Doretto, G.: Few-shot adversarial domain adaptation. Advances in Neural Information Processing systems 30 (2017)", + "30. Myers, L., Sirois, M.J.: Spearman correlation coefficients, differences between. Encyclopedia of Statistical Sciences 12 (2004)", + "31. Neuhaus, Y., Augustin, M., Boreiko, V., Hein, M.: Spurious features everywhere large-scale detection of harmful spurious features in imagenet. arXiv preprint arXiv:2212.04871 (2022)", + "32. NLP Connect: vit-gpt2-image-captioning (revision 0e334c7) (2022). https://doi.org/10.57967/hf/0222, https://huggingface.co/nlpconnect/vit-gpt2-image-captioning" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Zheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Nushi, B., Kamar, E., Horvitz, E.: Towards accountable ai: Hybrid human-machine analyses for characterizing system failure. In: Proceedings of the AAAI Conference on Human Computation and Crowdsourcing. vol. 6, pp. 126-135 (2018)", + "34. Oh, J., Yoo, H., Kim, C., Yun, S.Y.: Boil: Towards representation change for few-shot learning. In: International Conference on Learning Representations (2020)", + "35. Oreshkin, B., Rodríguez López, P., Lacoste, A.: Tadam: Task dependent adaptive metric for improved few-shot learning. In: Advances in Neural Information Processing Systems 31, pp. 721-731 (2018)", + "36. Poulakakis-Daktylidis, S., Jamali-Rad, H.: Bectr: Batch enhanced contrastive few-shot learning. In: International Conference on Learning Representations (2024)", + "37. Qin, X., Song, X., Jiang, S.: Bi-level meta-learning for few-shot domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15900-15910 (2023)", + "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)", + "39. Raghu, A., Raghu, M., Bengio, S., Vinyals, O.: Rapid learning or feature reuse? towards understanding the effectiveness of maml. In: International Conference on Learning Representations (2020)", + "40. Ravi, S., Larochelle, H.: Optimization as a model for few-shot learning. In: International Conference on Learning Representations (2017)", + "41. Ren, M., Ravi, S., Triantafillou, E., Snell, J., Swersky, K., Tenenbaum, J.B., Larochelle, H., Zemel, R.S.: Meta-learning for semi-supervised few-shot classification. In: International Conference on Learning Representations (2018)", + "42. Rusu, A.A., Rao, D., Sygnowski, J., Vinyals, O., Pascanu, R., Osindero, S., Hadsell, R.: Meta-learning with latent embedding optimization. In: International Conference on Learning Representations (2019)", + "43. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks. In: International Conference on Learning Representations (2020)", + "44. Singla, S., Feizi, S.: Salient imagenet: How to discover spurious features in deep learning? In: International Conference on Learning Representations (2022)", + "45. Singla, S., Nushi, B., Shah, S., Kamar, E., Horvitz, E.: Understanding failures of deep networks via robust feature extraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12853-12862 (2021)", + "46. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems 30, pp. 4077-4087 (2017)", + "47. Stock, P., Cisse, M.: Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases. In: European Conference on Computer Vision. pp. 498-512 (2018)", + "48. Sun, Q., Liu, Y., Chua, T.S., Schiele, B.: Meta-transfer learning for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 403-412 (2019)", + "49. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H.S., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1199-1208 (2018)", + "50. Tian, Y., Wang, Y., Krishnan, D., Tenenbaum, J.B., Isola, P.: Rethinking few-shot image classification: a good embedding is all you need? In: European Conference on Computer Vision. pp. 266-282. Springer (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Benchmarking Spurious Bias in Few-Shot Image Classifiers", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Triantafillou, E., Zhu, T., Dumoulin, V., Lamblin, P., Evci, U., Xu, K., Goroshin, R., Gelada, C., Swersky, K., Manzagol, P.A., Larochelle, H.: Meta-dataset: A dataset of datasets for learning to learn from few examples. In: International Conference on Learning Representations (2020)", + "52. Tseng, H.Y., Lee, H.Y., Huang, J.B., Yang, M.H.: Cross-domain few-shot classification via learned feature-wise transformation. In: International Conference on Learning Representations (2020)", + "53. Vinyals, O., Blundell, C., Lillicrap, T., kavukcuoglu, k., Wierstra, D.: Matching networks for one shot learning. In: Advances in Neural Information Processing Systems 29, pp. 3630-3638 (2016)", + "54. Wang, Y., Girshick, R., Hebert, M., Hariharan, B.: Low-shot learning from imaginary data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7278-7286 (2018)", + "55. Wang, Y., Yao, Q., Kwok, J.T., Ni, L.M.: Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys 53(3), 1-34 (2020)", + "56. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-UCSD Birds 200. Tech. Rep. CNS-TR-2010-001, California Institute of Technology (2010)", + "57. Wong, E., Santurkar, S., Madry, A.: Leveraging sparse linear layers for debuggable deep networks. In: International Conference on Machine Learning. pp. 11205-11216. PMLR (2021)", + "58. Xiao, K.Y., Engstrom, L., Ilyas, A., Madry, A.: Noise or signal: The role of image backgrounds in object recognition. In: International Conference on Learning Representations (2021)", + "59. Xue, Y., Payani, A., Yang, Y., Mirzasoleiman, B.: Eliminating spurious correlations from pre-trained models via data mixing. arXiv preprint arXiv:2305.14521 (2023)", + "60. Yang, Y., Nushi, B., Palangi, H., Mirzasoleiman, B.: Mitigating spurious correlations in multi-modal models during fine-tuning. arXiv preprint arXiv:2304.03916 (2023)", + "61. Ye, H.J., Hu, H., Zhan, D.C., Sha, F.: Few-shot learning via embedding adaptation with set-to-set functions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8808-8817 (2020)", + "62. Yue, X., Zheng, Z., Das, H.P., Keutzer, K., Vincentelli, A.S.: Multi-source few-shot domain adaptation. arXiv preprint arXiv:2109.12391 (2021)", + "63. Yue, Z., Zhang, H., Sun, Q., Hua, X.S.: Interventional few-shot learning. Advances in Neural Information Processing Systems 33, 2734-2746 (2020)", + "64. Zhang, C., Cai, Y., Lin, G., Shen, C.: Deepemd: Differentiable earth mover's distance for few-shot learning (2020)", + "65. Zhang, C., Cai, Y., Lin, G., Shen, C.: Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)", + "66. Zhang, J., Wang, Y., Molino, P., Li, L., Ebert, D.S.: Manifold: A model-agnostic framework for interpretation and diagnosis of machine learning models. IEEE Transactions on Visualization and Computer Graphics 25(1), 364-373 (2018)", + "67. Zhang, M., Li, H., Wu, F., Kuang, K.: Metacoco: A new few-shot classification benchmark with spurious correlation. In: The Twelfth International Conference on Learning Representations (2024)", + "68. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 830 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "G. Zheng et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_model.json b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5090ef7dd90e6dcc3743703ee4766be35cadb2d3 --- /dev/null +++ b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_model.json @@ -0,0 +1,2667 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.228, + 0.141, + 0.78, + 0.185 + ], + "angle": 0, + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.212, + 0.713, + 0.231 + ], + "angle": 0, + "content": "Guangtao Zheng1, Wenqian Ye1, and Aidong Zhang1" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.24, + 0.686, + 0.27 + ], + "angle": 0, + "content": "1University of Virginia, Charlottesville VA 22904, USA {gz5hp,wenqian,aidong}@virginia.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.311, + 0.744, + 0.631 + ], + "angle": 0, + "content": "Abstract. Few-shot image classifiers are designed to recognize and classify new data with minimal supervision and limited data but often show reliance on spurious correlations between classes and spurious attributes, known as spurious bias. Spurious correlations commonly hold in certain samples and few-shot classifiers can suffer from spurious bias induced from them. There is an absence of an automatic benchmarking system to assess the robustness of few-shot classifiers against spurious bias. In this paper, we propose a systematic and rigorous benchmark framework, termed FewSTAB, to fairly demonstrate and quantify varied degrees of robustness of few-shot classifiers to spurious bias. Few-STAB creates few-shot evaluation tasks with biased attributes so that using them for predictions can demonstrate poor performance. To construct these tasks, we propose attribute-based sample selection strategies based on a pre-trained vision-language model, eliminating the need for manual dataset curation. This allows FewSTAB to automatically benchmark spurious bias using any existing test data. FewSTAB offers evaluation results in a new dimension along with a new design guideline for building robust classifiers. Moreover, it can benchmark spurious bias in varied degrees and enable designs for varied degrees of robustness. Its effectiveness is demonstrated through experiments on ten few-shot learning methods across three datasets. We hope our framework can inspire new designs of robust few-shot classifiers. Our code is available at https://github.com/gtzheng/FewSTAB" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.644, + 0.741, + 0.672 + ], + "angle": 0, + "content": "Keywords: Few-shot classification \\(\\cdot\\) Spurious bias \\(\\cdot\\) Robustness \\(\\cdot\\) Benchmark system" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.7, + 0.377, + 0.716 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Few-shot classification [5,21,46,61,64] (FSC) has attracted great attention recently due to its promise for recognizing novel classes efficiently with limited data. Few-shot classifiers can transfer the knowledge learned from base classes to recognize novel classes with a few labeled samples. However, they face potential risks when deployed in the real world, such as data distribution shifts [29,62] and adversarial examples [8,14]. A subtle yet critical risk factor is the spurious correlations [3,4,11,13,43,47,59] between classes and spurious attributes —" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.499, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.228, + 0.424, + 0.238 + ], + "angle": 0, + "content": "Learned path" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.788, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.228, + 0.672, + 0.238 + ], + "angle": 0, + "content": "Non-existing path" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.292 + ], + "angle": 0, + "content": "Fig. 1: Exploiting the spurious correlation between the class bird and the spurious attribute tree branch to predict bird leads to an incorrect prediction on the test image showing birds on a grass field. For clarity, we only show the case for one class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.789, + 0.505 + ], + "angle": 0, + "content": "attributes of inputs non-essential to the classes. In the traditional learning setting [1,2,43], deep learning models tend to rely on spurious correlations as their prediction shortcuts or exhibit spurious bias, such as predicting classes using the associated backgrounds [58] or image textures [12], leading to significant performance drops when the associated backgrounds or textures change to different ones. In the low-data regime, spurious bias becomes more evident. For example, in Fig. 1, the correlation between the class bird and the attribute tree branch in the support (training) image may form a shortcut path from tree branch to predicting the image as bird and hinder the learning of the desired one that uses class-related attributes, such as head, tail, and wing. The shortcut will fail to generalize in the query (test) image where no tree branch can be found. In general, few-shot image classifiers are susceptible to spurious bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.79, + 0.657 + ], + "angle": 0, + "content": "However, there lacks a dedicated benchmarking framework that evaluates the robustness of few-shot classifiers to spurious bias. The standard benchmarking procedure in FSC trains a few-shot classifier on base classes from a training set with ample samples and evaluates the classifier on FSC test tasks constructed from a test set with novel classes. The problem with this procedure is the lack of explicit control over the spurious correlations in the constructed FSC tasks. Each FSC test task contains randomly sampled support and query samples. Thus, spurious correlations in the majority of the test set samples can be demonstrated in these tasks, providing unfair advantages for few-shot classifiers with high reliance on the spurious correlations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose a systematic and rigorous benchmark framework, termed Few-Shot Tasks with Attribute Biases (FewSTAB), to fairly compare the robustness of various few-shot classifiers to spurious bias. Our framework explicitly controls spurious correlations in the support and query samples when constructing an FSC test task to reveal the robustness pitfalls caused by spurious bias. To achieve this, we propose attribute-based sample selection strategies that select support and query samples with biased attributes. These attributes together with their associated classes formulate spurious correlations such that if the support samples induce spurious bias in a few-shot classifier, i.e., the classifier learns the spurious correlations in the support samples as its prediction shortcuts, then the query samples can effectively degrade the classifier's performance, exposing its non-robustness to spurious bias." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Our framework exploits the spurious attributes in test data for formulating spurious correlations in FSC test tasks. Some existing datasets [15,27,43] provide spurious attribute annotations. However, they only have a few classes and cannot provide enough classes for training and testing. Many benchmark datasets for FSC do not have annotations on spurious attributes, and obtaining these annotations typically involves labor-intensive human-guided labeling [33,66]. To address this, we further propose to use a pre-trained vision-language model (VLM) to automatically identify distinct attributes in images in the high-level text format. Our attribute-based sampling methods can use the identified attributes to simulate various spurious correlations. Thus, we can reuse any existing FSC datasets for benchmarking few-shot classifiers' robustness to spurious bias, eliminating the need for the manual curation of new datasets." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.329, + 0.697, + 0.342 + ], + "angle": 0, + "content": "The main contributions of our work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.351, + 0.786, + 0.41 + ], + "angle": 0, + "content": "- We propose a systematic and rigorous benchmark framework, termed Few-Shot Tasks with Attribute Biases (FewSTAB), that specifically targets spurious bias in few-shot classifiers, demonstrates their varied degrees of robustness to spurious bias, and benchmarks spurious bias in varied degrees." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.412, + 0.786, + 0.469 + ], + "angle": 0, + "content": "- We propose novel attribute-based sample selection strategies using a pretrained VLM for constructing few-shot evaluation tasks, allowing us to reuse any existing few-shot benchmark datasets without manually curating new ones for the evaluation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.472, + 0.786, + 0.53 + ], + "angle": 0, + "content": "- FewSTAB provides a new dimension of evaluation on the robustness to spurious bias along with a new design guideline for building robust few-shot classifiers. We demonstrate the effectiveness of FewSTAB by applying it to models trained on three benchmark datasets with ten FSC methods." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.351, + 0.786, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.388, + 0.569 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Few-shot classification. Few-shot classification [6,46,50,51,53,55] has received vast attention recently. Few-shot classifiers can be trained with meta-learning or transfer learning on base classes to learn the knowledge that can be transferred to recognize novel classes with a few labeled samples. The transfer learning approaches [6,50] first learn a good embedding model and then fine-tune the model on samples from novel classes. The meta-learning approaches can be further divided into optimization-based and metric-based methods. The optimization-based methods [9,10,22,26,54] aim to learn a good initialized model such that the model can adapt to novel classes efficiently with a few gradient update steps on a few labeled samples. The metric-based methods [5,21,35,42,49,53,61,65] aim to learn a generalizable representation space with a well-defined metric, such as Euclidean distance [46], to learn novel classes with a few labeled samples. Recently, large vision-language models [20,38,68] are used for few-shot classification. However, they have completely different training and inference pipelines from the models that we consider in this paper." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Robustness in few-shot classification. There are several notions of robustness for few-shot classifiers. The common one requires a few-shot classifier to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "G. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "perform well on the in-distribution samples of novel classes in randomly sampled FSC test tasks. The robustness to adversarial perturbations further requires a few-shot classifier to perform well on samples with imperceptible perturbations [8,14]. Moreover, the cross-domain generalization [37,51,52] aims to test how robust a few-shot classifier is on samples from novel classes with domain shifts, which are typically reflected by the changes in both image styles and classes. In contrast, we focus on a new notion of robustness: the robustness to spurious bias. There is a lack of rigorous evaluation methods on the topic. We provide a new evaluation method that specifically targets spurious bias and can systematically demonstrate few-shot classifiers' varied degrees of vulnerability to spurious bias, which has not been addressed in the existing literature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.449 + ], + "angle": 0, + "content": "Benchmarks for spurious bias. There are some existing datasets [15,27,43] that are designed to benchmark spurious bias in image classifiers. However, these datasets are only applicable to the traditional learning setting [1,2,43] since the classes in them are not sufficient for the training and testing of few-shot classifiers. Existing benchmarks in few-shot classification are not tailored for benchmarking spurious bias in few-shot classifiers. A recent work [67] creates a large-scale few-shot classification benchmark dataset with spurious-correlation shifts. In contrast, we propose a benchmark framework that can reuse existing few-shot classification datasets and provide a new dimension of evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Discovering spurious attributes. A spurious attribute is non-essential to a class and only exists in some samples. Early works on discovering spurious attributes [33,66] require a predefined list of spurious attributes and expensive human-guided labeling of visual attributes. Recent works [31,44,45,57] greatly reduce the need for manual annotations by using the neurons of robust models to detect visual attributes. However, they still need humans to annotate the detected visual attributes. We automate this process by using a pre-trained VLM to obtain distinct attributes as words. Instead of discovering spurious correlations, we simulate them via attribute-based sampling for benchmarking." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.37, + 0.628 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Few-shot classification tasks. A typical FSC task \\(\\mathcal{T}\\) has a support set \\(S\\) for training and a query set \\(\\mathcal{Q}\\) for testing. In this task, there are \\(C\\) classes \\((c = 1,\\dots ,C)\\) with \\(N_{\\mathcal{S}}\\) (a small number) training samples and \\(N_{\\mathcal{Q}}\\) test samples per class in \\(S\\) and \\(\\mathcal{Q}\\), respectively. The task is called a \\(C\\)-way \\(N_{\\mathcal{S}}\\)-shot task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Few-shot classifiers. A few-shot classifier \\( f_{\\theta} \\) with parameters \\( \\theta \\) aims to classify the samples in \\( \\mathcal{Q} \\) after learning from \\( S \\) with a learning algorithm \\( \\mathcal{O} \\) in a few-shot task \\( \\mathcal{T} \\). Here, \\( \\mathcal{O} \\) could be any learning algorithms, such as the optimization method [9] or a prototype-based classifier learning method [7,46]. To acquire a good few-shot learning capability, \\( f_{\\theta} \\) is typically meta-trained or pre-trained [25] on a base training set \\( \\mathcal{D}_{train} = \\{(x_n, y_n) | y_n \\in \\mathcal{C}_{train}, n = 1, \\dots, N_{train}\\} \\) with \\( N_{train} \\) sample(x)-label(y) pairs, where \\( \\mathcal{C}_{train} \\) is a set of base classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Performance metrics. The performance of a few-shot classifier is typically measured by its average classification accuracy over \\( N_{\\mathcal{T}} \\) \\( C \\)-way \\( N_{S} \\)-shot tasks" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.313, + 0.145, + 0.69, + 0.16 + ], + "angle": 0, + "content": "Table 1: Meanings of major symbols used in the paper." + }, + { + "type": "table", + "bbox": [ + 0.341, + 0.172, + 0.658, + 0.393 + ], + "angle": 0, + "content": "
SymbolMeaning
TAn FSC task
SSupport (training) set in T
QQuery (test) set in T
cA class in T
CNumber of classes per task
NsNumber of samples (shots) per class in S
NQNumber of samples per class in Q
OA few-shot adaptation algorithm
ψAn attribute detector
ΩAn automatic word selector
DtrainThe base training set
DvalThe validation set for selecting a few-shot classifier
DtestThe test set for evaluating a few-shot classifier
CtrainClasses in Dtrain
CtestClasses in Dtest
DcA set of all samples belonging to class c
aA text-format attribute
AA set of text-format attributes
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.468 + ], + "angle": 0, + "content": "randomly sampled from \\(\\mathcal{D}_{test} = \\{(x_n, y_n) | y_n \\in \\mathcal{C}_{test}, n = 1, \\dots, N_{test}\\}\\) where \\(N_{test}\\) sample-label pairs from novel classes \\(\\mathcal{C}_{test}\\) do not appear in \\(\\mathcal{D}_{train}\\), i.e., \\(\\mathcal{C}_{train} \\cap \\mathcal{C}_{test} = \\emptyset\\). We denote this metric as standard accuracy \\(\\mathrm{Acc}(f_\\theta)\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.478, + 0.787, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A c c} \\left(f _ {\\theta}\\right) = \\frac {1}{N _ {\\mathcal {T}}} \\sum_ {t = 1} ^ {N _ {\\mathcal {T}}} \\sum_ {c = 1} ^ {C} M _ {c} \\left(\\mathcal {T} _ {t}; f _ {\\theta}, \\mathcal {O}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.606 + ], + "angle": 0, + "content": "where \\( M_{c}(\\mathcal{T}_{t};f_{\\theta},\\mathcal{O}) \\) denotes the classification accuracy of \\( f_{\\theta} \\) on the query samples from the class \\( c \\) in \\( \\mathcal{T}_t \\) after \\( f_{\\theta} \\) is trained on \\( S \\) with \\( \\mathcal{O} \\). The metric \\( \\mathrm{Acc}(f_{\\theta}) \\) in Eq. (11) only shows the average learning capability of \\( f_{\\theta} \\) over \\( C \\) randomly selected novel classes. To better characterize the robustness of \\( f_{\\theta} \\) to spurious bias, we define the class-wise worst classification accuracy over tasks as" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.617, + 0.787, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\operatorname {w A c c} \\left(f _ {\\theta}\\right) = \\frac {1}{N _ {\\mathcal {T}}} \\sum_ {t = 1} ^ {N _ {\\mathcal {T}}} \\min _ {c = 1, \\dots , C} M _ {c} \\left(\\mathcal {T} _ {t}; f _ {\\theta}, \\mathcal {O}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.704, + 0.684 + ], + "angle": 0, + "content": "A larger \\(\\mathrm{wAcc}(f_{\\theta})\\) indicates that \\(f_{\\theta}\\) is more robust to spurious bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Spurious correlations. A spurious correlation is the association between a class and an attribute of inputs that is non-essential to the class, and it only holds in some samples. We formally define it as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Definition 1. Let \\(\\mathcal{D}_c\\) denote a set of sample-label pairs having the label \\(c\\), and let \\(\\psi : \\mathcal{X} \\to \\mathcal{B}_A\\) be an attribute detector, where \\(\\mathcal{X}\\) is the set of all possible inputs, \\(\\mathcal{B}_A\\) denotes all possible subsets of \\(\\mathcal{A}\\), and \\(\\mathcal{A}\\) is the set of all possible attributes. The class \\(c\\) and an attribute \\(a \\in \\mathcal{A}\\) form a spurious correlation, denoted as \\(\\langle c, a \\rangle\\), if and only if the following conditions hold:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.825, + 0.618, + 0.841 + ], + "angle": 0, + "content": "1. There exists \\((x,c)\\in \\mathcal{D}_c\\) that satisfies \\(a\\in \\psi (x)\\) , and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.787, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.344, + 0.785, + 0.373 + ], + "angle": 0, + "content": "Fig. 2: FewSTAB overview. (a) Extract distinct attributes using a pre-trained VLM. (b) Generate an FSC task for the evaluation of spurious bias in few-shot classifiers." + }, + { + "type": "title", + "bbox": [ + 0.222, + 0.402, + 0.591, + 0.418 + ], + "angle": 0, + "content": "2. There exists \\((x', c) \\in \\mathcal{D}_c\\) that satisfies \\(a \\notin \\psi(x')\\)." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.426, + 0.571, + 0.443 + ], + "angle": 0, + "content": "We define \\( a \\) as the spurious attribute in \\( \\langle c, a \\rangle \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.785, + 0.496 + ], + "angle": 0, + "content": "Definition \\(\\mathbb{I}\\) specifies that all the spurious correlations are based on \\(\\mathcal{D}_c\\). In the remainder of the paper, we define \\(\\mathcal{D}_c = \\{(x,c)|\\forall (x,c)\\in \\mathcal{D}_{test}\\}\\) with \\(c\\in \\mathcal{C}_{test}\\) as we focus on evaluating the robustness to spurious bias." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.497, + 0.753, + 0.514 + ], + "angle": 0, + "content": "We list major symbols in the paper alongside their meanings in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.535, + 0.38, + 0.553 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.566, + 0.542, + 0.581 + ], + "angle": 0, + "content": "4.1 Attribute-Based Sample Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.696 + ], + "angle": 0, + "content": "We first propose two attribute-based sample selection methods to reveal spurious bias in a few-shot classifier. Consider a training set \\(\\mathcal{S}\\) in a few-shot test task \\(\\mathcal{T}\\), which has \\(C\\) classes with each class \\(c\\in \\mathcal{C}_{test}\\) associating with a unique spurious attribute \\(a\\in \\mathcal{A}\\). We aim to discover samples that can exhibit a classifier's spurious bias on \\(\\langle c,a\\rangle\\) induced from \\(\\mathcal{S}\\). Motivated by existing findings [43,59,60] that classifiers with high reliance on \\(\\langle c,a\\rangle\\) tend to perform poorly on samples without it, we propose an attribute-based sample selection strategy below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Intra-class attribute-based sample selection. Given \\(\\mathcal{D}_c\\) and the training set \\(S\\) having the spurious correlation \\(\\langle c, a \\rangle\\), we generate a set \\(\\mathcal{I}_{\\langle c, a \\rangle}\\) of sample-label pairs which have class \\(c\\) but do not contain attribute \\(a\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.753, + 0.786, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\langle c, a \\rangle} = \\{(x, c) | \\forall (x, c) \\in \\mathcal {D} _ {c}, a \\notin \\psi (x) \\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The above proposed method demonstrates a few-shot classifier's robustness to individual spurious correlation \\(\\langle c, a \\rangle\\) and does not consider a multi-class classification setting where spurious attributes from some other class \\(c'\\) exist in samples of the class \\(c\\). In this case, these attributes may mislead the classifier to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "predict those samples as the class \\( c' \\) and severely degrade the performance on the class \\( c \\). For example, consider using the spurious correlations \\( \\langle \\text{vase}, \\text{blue} \\rangle \\) and \\( \\langle \\text{bowl}, \\text{green} \\rangle \\) for predicting vase and bowl, respectively. An image showing a vase in green is more effective in revealing the robustness to \\( \\langle \\text{vase}, \\text{blue} \\rangle \\) as it is more likely to be misclassified as bowl than other images. Motivated by this, we propose the inter-class attribute-based sample selection below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.303 + ], + "angle": 0, + "content": "Inter-class attribute-based sample selection. Given \\(\\mathcal{D}_c\\) and the training set \\(S\\) having the spurious correlations \\(\\langle c, a \\rangle\\) and \\(\\langle c', a' \\rangle\\), where \\(c' \\neq c\\) and \\(a' \\neq a\\), we generate a set \\(\\mathcal{I}_{\\langle c, a \\rangle}^{\\langle c', a' \\rangle}\\) of sample-label pairs which have class \\(c\\), do not contain attribute \\(a\\), but contain attribute \\(a'\\) from another class \\(c'\\):" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.315, + 0.785, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\langle c, a \\rangle} ^ {\\langle c ^ {\\prime}, a ^ {\\prime} \\rangle} = \\mathcal {I} _ {\\langle c, a \\rangle} \\cap \\left\\{\\left(x, c\\right) | \\forall (x, c) \\in \\mathcal {D} _ {c}, a _ {\\langle c ^ {\\prime} \\rangle} ^ {\\prime} \\in \\psi (x) \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.674, + 0.366 + ], + "angle": 0, + "content": "where \\(a_{\\langle c^{\\prime}\\rangle}^{\\prime}\\) denotes \\(a^\\prime\\) in \\(\\langle c',a'\\rangle\\) , and \\(\\mathcal{I}_{\\langle c,a\\rangle}\\) is defined in Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.785, + 0.458 + ], + "angle": 0, + "content": "Considering that there are \\( C \\) classes in the training set \\( S \\) with each class associating with a unique spurious attribute \\( a \\), to effectively demonstrate the reliance on the spurious correlation \\( \\langle c, a \\rangle \\) with the inter-class attribute-based sample selection, we consider all the spurious correlations in \\( S \\). Specifically, we apply the above selection strategy to all the \\( C - 1 \\) spurious correlations in \\( S \\) other than \\( \\langle c, a \\rangle \\) and obtain \\( \\mathcal{I}_{\\langle c, a \\rangle}^{C} \\) as the union of the \\( C - 1 \\) sets as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.47, + 0.785, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\langle c, a \\rangle} ^ {C} = \\bigcup_ {\\langle c ^ {\\prime}, a ^ {\\prime} \\rangle \\in \\mathcal {C} \\backslash c} \\mathcal {I} _ {\\langle c, a \\rangle} ^ {\\langle c ^ {\\prime}, a ^ {\\prime} \\rangle}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.713, + 0.534 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_{\\backslash c}\\) denotes all the spurious correlations in \\(S\\) other than \\(\\langle c, a \\rangle\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.785, + 0.623 + ], + "angle": 0, + "content": "The inter-class attribute-based sample selection is built upon the intra-class attribute-based sample selection. In the remainder of the paper, we use the inter-class method as our default sample selection strategy, which is more effective empirically (Sec. 5.6). In certain cases, however, where there are not enough desired samples during task construction, we resort to the intra-class sample selection strategy (Sec. 5.1, Implementation details)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.784, + 0.669 + ], + "angle": 0, + "content": "In the following, we introduce FewSTAB, a benchmark framework that uses the proposed selection strategies to construct FSC tasks containing samples with biased attributes for benchmarking spurious bias in few-shot classifiers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.691, + 0.708 + ], + "angle": 0, + "content": "4.2 FewSTAB (Part 1): Text-Based Attribute Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.794 + ], + "angle": 0, + "content": "Our attribute-based sample selection methods require knowing the attributes in images, which typically involves labor-intensive human labeling. To make our method scalable and applicable to few-shot classifiers trained on different datasets, we adopt a pre-trained VLM to automatically identify distinct attributes in images in text format, which includes the following two steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Step 1: Generating text descriptions. We use a pre-trained VLM [23, 32] \\(\\phi\\) to automatically generate text descriptions for images in \\(\\mathcal{D}_{test}\\). The VLM is a model in the general domain and can produce text descriptions for various" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "G. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "objects and patterns. For example, for the current input image in the vase class in Fig. 2(a), besides the class object vase, the VLM also detects the vase's color green, and another object table with its material wooden." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.284 + ], + "angle": 0, + "content": "Step 2: Extracting informative words. From the generated text descriptions, we extract nouns and adjectives as the detected attributes via an automatic procedure \\(\\Omega\\). The two kinds of words are informative as a noun describes an object, and an adjective describes a property of an object. All the detected attributes form the candidate attribute set \\(\\mathcal{A}\\). We realize the attribute detector \\(\\psi\\) defined in Definition 1 as \\(\\psi(x) = \\Omega(\\phi(x))\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.785, + 0.39 + ], + "angle": 0, + "content": "Remark 1: A VLM in general can extract many distinct attributes from the images. On some images, the VLM may detect non-relevant attributes, such as detecting a duck from a bird image. A more capable VLM could warrant a better attribute detection accuracy and benefit individual measurements on few-shot classifiers. Although being a VLM-dependent benchmark framework, FewSTAB can produce consistent and robust relative measurements among all the compared FSC methods, regardless the choice of VLMs (Sec. 5.6)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.468 + ], + "angle": 0, + "content": "Remark 2: The candidate set \\(\\mathcal{A}\\) constructed with all the extracted words may contain attributes that represent the classes in \\(\\mathcal{D}_{test}\\). However, during our attribute-based sample selection, these attributes will not be used since they always correlate with classes and therefore do not satisfy the definition of spurious attributes in Definition 1. We provide details of \\(\\phi\\) and \\(\\Omega\\) in Sec. 5.1" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.492, + 0.622, + 0.508 + ], + "angle": 0, + "content": "4.3 FewSTAB (Part 2): FSC Task Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.784, + 0.564 + ], + "angle": 0, + "content": "Constructing a \\(C\\)-way \\(N_{S}\\)-shot FSC task \\(\\mathcal{T}\\) for benchmarking spurious bias in few-shot classifiers involves constructing a support (training) set \\(\\mathcal{S}\\) and a query (test) \\(\\mathcal{Q}\\) with biased attributes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.785, + 0.762 + ], + "angle": 0, + "content": "Constructing the support set. The support set contains the spurious correlations that we aim to demonstrate to a few-shot classifier. As a fair and rigorous benchmark system, FewSTAB makes no assumptions on the few-shot classifiers being tested and randomly samples \\( C \\) classes from \\( \\mathcal{C}_{test} \\). For each sampled class, it randomly selects a spurious correlation \\( \\langle c, a \\rangle \\) in \\( \\mathcal{D}_{test} \\) with \\( a \\in \\mathcal{A} \\). To effectively demonstrate the spurious correlation \\( \\langle c, a \\rangle \\) to a few-shot classifier, we select samples of the class \\( c \\) such that (1) they all have the spurious attribute \\( a \\) and (2) do not have spurious attributes from the other \\( C - 1 \\) spurious correlations. We construct \\( S_{c} \\) with \\( N_{S} \\) samples for the class \\( c \\) that satisfy the above two conditions. Thus, the spurious attribute \\( a \\) becomes predictive of the class \\( c \\) in \\( S_{c} \\). We take the union of all \\( C \\) such sets to get \\( S = \\cup_{c=1}^{C} S_{c} \\). Fig. 2(b) demonstrates the case when \\( C = 3 \\). Note that we have no requirements for other non-selected attributes in \\( A \\) to ensure that we have enough samples for \\( S_{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Constructing the query set. To evaluate the robustness to the spurious correlations formulated in \\(\\mathcal{S}\\), we first construct a candidate set \\(\\mathcal{I}_{\\langle c,a\\rangle}^{C}\\) in Eq. (5) for each spurious correlation \\(\\langle c,a\\rangle\\) in \\(\\mathcal{S}\\). Since we have no requirements on the non-selected attributes that are not used to formulate spurious correlations in \\(\\mathcal{S}\\), a few-shot classifier may predict query samples via some of these attributes," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "e.g., the yellow blocks in Fig. 2(b), bypassing the test on the formulated spurious correlations in \\(\\mathcal{S}\\). To address this, we propose query sample selection below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.784, + 0.24 + ], + "angle": 0, + "content": "Query sample selection: We select query samples from \\(\\mathcal{I}_{\\langle c,a\\rangle}^{C}\\) that are least likely to have non-selected spurious attributes, such as the ones enclosed with red boxes in Fig. 2(b). To achieve this, we first calculate the fraction of sample-label pairs in \\(\\mathcal{I}_{\\langle c,a\\rangle}^{C}\\) that have the attribute \\(a\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.25, + 0.786, + 0.27 + ], + "angle": 0, + "content": "\\[\np _ {a} = | \\{x \\mid a \\in \\psi (x), \\forall (x, c) \\in \\mathcal {I} _ {\\langle c, a \\rangle} ^ {C} \\} | / | \\mathcal {I} _ {\\langle c, a \\rangle} ^ {C} |, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.785, + 0.433 + ], + "angle": 0, + "content": "where \\(|\\cdot|\\) denotes the size of a set, \\(a \\in \\tilde{\\mathcal{A}}\\), and \\(\\tilde{\\mathcal{A}}\\) contains all non-selected attributes. A larger \\(p_a\\) indicates that the attribute \\(a\\) occurs more frequently in data and is more likely to be used in formulating prediction shortcuts. We then calculate the likelihood score for each \\((x,c) \\in \\mathcal{I}_{\\langle c,a\\rangle}^C\\) as \\(s(x) = \\sum_{a \\in \\psi(x), a \\in \\tilde{\\mathcal{A}}} p_a\\), i.e., the summation of all \\(p_a\\) of non-selected attributes in \\(x\\). The likelihood score will be zero if there are no non-selected attributes in \\(x\\). A large \\(s(x)\\) indicates that the image \\(x\\) can be predicted via many non-selected attributes. Therefore, we select \\(N_{\\mathcal{Q}}\\) samples from \\(\\mathcal{I}_{\\langle c,a\\rangle}^C\\) that have the lowest likelihood scores to construct \\(\\mathcal{Q}_c\\). Then, we have \\(\\mathcal{Q} = \\cup_{c=1}^{C} \\mathcal{Q}_c\\), which contains samples for evaluating the robustness of a few-shot classifier to the spurious correlations in \\(\\mathcal{S}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.784, + 0.54 + ], + "angle": 0, + "content": "Complexity analysis. The text-based attribute detection only needs to use VLMs once to extract attributes for each test set of a dataset. For the task construction, in a nutshell, we analyze the attributes of samples from each of the \\(C\\) classes and do the sampling. Thus, the computational complexity is \\(O(N_{\\mathcal{T}}CN_cN_{\\mathcal{A}})\\), where \\(N_{c}\\) is the maximum number of samples per class in test data, \\(N_{\\mathcal{A}}\\) is the number of extracted attributes. We only need to run the process once and use the generated tasks to benchmark various models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.561, + 0.377, + 0.579 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.591, + 0.429, + 0.607 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.784, + 0.674 + ], + "angle": 0, + "content": "Datasets. We used two general datasets of different scales, miniImageNet [40] and tieredImageNet [41], and one fine-grained dataset, CUB-200 [56]. Each dataset consists of \\( \\mathcal{D}_{train} \\), \\( \\mathcal{D}_{val} \\), and \\( \\mathcal{D}_{test} \\) for training, validation, and test, respectively (see Appendix). All images were resized to \\( 84 \\times 84 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.781 + ], + "angle": 0, + "content": "FSC methods. We trained FSC models with ten algorithms covering three major categories. For gradient-based meta-learning algorithms, we chose ANIL [39], LEO [42], and BOIL [34]. For metric-based meta-learning algorithms, we chose ProtoNet [46], DN4 [24], R2D2 [5], CAN [16], and RENet [19]. For transfer learning algorithms, we chose Baseline++ [6] and RFS [50]. See Appendix for more details. Any backbones can be used as the feature extractor. For fair comparisons between different methods, we used the ResNet-12 backbone adopted in [35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.784, + 0.842 + ], + "angle": 0, + "content": "Text-based attribute detection. We used a pre-trained VLM named ViT-GPT2 [32] to generate text descriptions for images in \\(\\mathcal{D}_{test}\\). After that, we used Spacy (https://spacy.io/) to extract nouns and adjectives from these descriptions automatically. We also used another pre-trained VLM, BLIP [23], to test" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.284, + 0.145, + 0.719, + 0.16 + ], + "angle": 0, + "content": "Table 2: Statistics of detected attributes in \\( {\\mathcal{D}}_{\\text{test }} \\) by two VLMs." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.171, + 0.758, + 0.224 + ], + "angle": 0, + "content": "
VLMUnique attributesAvg. attributes per class
miniImageNettieredImageNetCUB-200miniImageNettieredImageNetCUB-200
ViT-GPT211112532159190.40230.9425.78
BLIP20326710247254.40310.4029.74
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.234, + 0.353, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.234, + 0.787, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.417, + 0.789, + 0.461 + ], + "angle": 0, + "content": "Fig. 3: A 5-way 1-shot task constructed by our inter-class attribute-based sample selection using samples from the miniImageNet dataset. Note that due to the limited capacity of a VLM, the attributes may not well align with human understandings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.789, + 0.518 + ], + "angle": 0, + "content": "whether FewSTAB can produce consistent results. The statistics of the detected attributes are shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.79, + 0.701 + ], + "angle": 0, + "content": "Implementation details. We trained FSC models with the implementation in [25]. Each model was trained on \\(\\mathcal{D}_{train}\\) of a dataset with one of the ten FSC methods. For each meta-learning based method, we trained two models using randomly sampled 5-way 1-shot and 5-way 5-shot tasks, respectively. All the tasks have 15 samples per class in the query set. We saved the model that achieves the best validation accuracy on \\(\\mathcal{D}_{val}\\) for evaluation. For FewSTAB, if we do not have enough desired samples to construct a support set, we redo the construction from the beginning. If there are not enough desired samples to construct a query set, we first try to use the intra-class attribute-based sample selection; if the desired samples are still not enough, we redo the construction from the beginning. We created 3000 tasks for model evaluation. All experiments were conducted on the NVIDIA RTX 8000 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.56, + 0.739 + ], + "angle": 0, + "content": "5.2 Visualization of a Constructed Task" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We show a 5-way 1-shot task constructed by FewSTAB in Fig. 3. Each class in the support set correlates with a unique spurious attribute. The query samples of a class do not have the spurious attribute correlated with the class and some of them have spurious attributes associated with other classes in the support set. For example, the query samples of the class lion do not have the spurious attribute water, and some of them have spurious attributes from other classes in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Comparison between wAcc-R and wAcc-A with \\(95\\%\\) confidence interval on the miniImageNet, tieredImageNet, and CUB datasets. Numbers in the Shot column indicate that the models are both trained (if applicable) and tested on 5-way 1- or 5-shot tasks. Darker colors indicate higher values." + }, + { + "type": "table", + "bbox": [ + 0.272, + 0.213, + 0.73, + 0.436 + ], + "angle": 0, + "content": "
ShotMethodminiImageNettieredImageNetCUB-200
wAcc-A (↑)wAcc-R (↑)wAcc-A (↑)wAcc-R (↑)wAcc-A (↑)wAcc-R (↑)
1ANIL10.38±0.3014.36±0.3311.21±0.3015.63±0.3613.78±0.4016.94±0.43
LEO14.26±0.4621.35±0.5416.00±0.5529.63±0.7128.29±0.8040.22±0.87
BOIL12.48±0.2314.93±0.2412.27±0.2114.13±0.2319.15±0.2922.50±0.29
ProtoNet14.03±0.4921.96±0.5814.50±0.5027.13±0.6934.62±0.8546.61±0.89
DN412.37±0.4619.28±0.5611.99±0.4723.62±0.6535.22±0.8647.26±0.88
R2D218.05±0.5326.50±0.6016.41±0.5430.41±0.7136.70±0.9048.82±0.88
CAN17.37±0.5325.96±0.6018.84±0.6036.29±0.7822.74±0.7231.95±0.78
RENet19.10±0.5728.85±0.6518.83±0.6135.70±0.7832.43±0.8143.98±0.81
Baseline++15.30±0.4823.18±0.5617.51±0.5431.62±0.719.17±0.4714.59±0.58
RFS18.00±0.5327.12±0.6118.35±0.6035.24±0.7732.45±0.8044.49±0.82
5ANIL14.83±0.4025.37±0.5213.72±0.3930.60±0.5131.63±0.5545.47±0.53
LEO26.31±0.5941.33±0.5929.49±0.7257.22±0.7246.62±0.8259.76±0.77
BOIL13.09±0.2215.21±0.2314.90±0.2218.55±0.2419.17±0.2821.33±0.27
ProtoNet32.07±0.5851.95±0.5230.95±0.7062.53±0.6260.06±0.7475.68±0.50
DN427.60±0.5842.68±0.6216.07±0.6240.63±0.8159.25±0.7773.58±0.56
R2D235.37±0.5950.84±0.5531.12±0.7261.08±0.6558.66±0.8275.20±0.54
CAN36.44±0.6554.23±0.5531.17±0.7664.19±0.6241.31±0.7461.61±0.58
RENet36.19±0.6356.52±0.5830.27±0.7663.49±0.6452.93±0.8271.82±0.56
Baseline++29.52±0.5744.94±0.5730.01±0.7259.06±0.6716.86±0.5229.84±0.69
RFS36.85±0.6455.66±0.5531.15±0.7662.71±0.6754.98±0.8174.33±0.53
" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.448, + 0.405, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.447, + 0.596, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.447, + 0.788, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.586 + ], + "angle": 0, + "content": "Fig. 4: Accuracy gaps (wAcc-R minus wAcc-A) on the 5-way 1-shot and 5-way 5-shot tasks from the (a) miniImageNet, (b) tieredImageNet, and (c) CUB-200 datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.66 + ], + "angle": 0, + "content": "the support set, such as man and green. FewSTAB introduces biased attributes in the task so that query samples can be easily misclassified as other classes by a few-shot classifier that relies on the spurious correlations in the support set." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.68, + 0.478, + 0.695 + ], + "angle": 0, + "content": "5.3 Effectiveness of FewSTAB" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "FewSTAB can effectively reveal spurious bias in few-shot classifiers. We show in Table 3 the wAcc (Eq. 2) on 5-way 1/5-shot test tasks that are randomly sampled (wAcc-R) and are constructed with FewSTAB (wAcc-A), respectively. FewSTAB generates FSC test tasks only based on the class-attribute correlations in data. In each test setting, the FSC methods in Table 3 are evaluated with the same FSC tasks. We observe that wAcc-A is consistently lower than wAcc-R on the three datasets and on two test-shot numbers, showing that FewSTAB is more effective than the standard evaluation procedure (random task construction) in exhibiting the spurious bias in various few-shot classifiers. We" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "additionally show that FewSTAB also works on the most recent FSC methods and can reflect the improvement made to mitigate spurious bias (see Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.546, + 0.359 + ], + "angle": 0, + "content": "FewSTAB reveals new robustness patterns among FSC methods. In Table 4, we calculate the Spearman's rank correlation coefficients [30] between the values of wAcc-R and wAcc-A from Table 3. The coefficients are bounded from 0 to 1, with larger values indicating that the ranks of FSC methods based on wAcc-R are more similar to those based on wAcc-A. In the 1-shot setting, it is not effective to control the spurious correlations since we only have one sample per class in the support set. Hence, the coefficients are large, and" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.222, + 0.788, + 0.266 + ], + "angle": 0, + "content": "Table 4: Spearman's rank correlations between wAcc-A and wAcc-R in Table 3" + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.267, + 0.763, + 0.327 + ], + "angle": 0, + "content": "
Dataset1-shot5-shot
miniImageNet0.960.95
tieredImageNet0.960.90
CUB-2001.000.94
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.435 + ], + "angle": 0, + "content": "the ranks based on wAcc-A are similar to those based on wAcc-R. In the 5-shot cases, we have more samples to demonstrate the spurious correlations. The coefficients become smaller, i.e., the ranks based on wAcc-A show different trends from those based on wAcc-R. In this case, FewSTAB reveals new information on FSC methods' varied degrees of robustness to spurious bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.542 + ], + "angle": 0, + "content": "FewSTAB can benchmark spurious bias in varied degrees. As shown in Fig. 4, the accuracy gap, defined as wAcc-R minus wAcc-A, in general, becomes larger when we switch from 5-way 1-shot to 5-way 5-shot tasks. Compared with the random task construction, FewSTAB creates more challenging tasks in the 5-shot case for demonstrating spurious bias in few-shot classifiers. In other words, with a higher shot value in the constructed test tasks, FewSTAB aims to benchmark spurious bias in a higher degree." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.568, + 0.771, + 0.584 + ], + "angle": 0, + "content": "5.4 A New Dimension of Evaluation and a New Design Guideline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.546, + 0.841 + ], + "angle": 0, + "content": "FewSTAB creates a new dimension of evaluation on the robustness to spurious bias. We demonstrate this with a scatter plot (Fig. 5) of Acc (Eq. 11) and wAcc-A of the ten few-shot classifiers. FewSTAB offers new information regarding different few-shot classifiers' robustness to spurious bias as we observe that Acc does not well correlate with wAcc-A. A high wAcc-A indicates that the classifier is robust to spurious bias, while a high Acc indicates that the classifier can correctly predict most of the samples. With the scatter plot, we can view tradeoffs between the two metrics on existing few-shot classifiers. A desirable few-shot classifier should appear in the top-right corner of the plot." + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.63, + 0.776, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.743, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Fig.5: Acc versus wAcc-A of the ten FSC methods tested on 5-way 5-shot tasks from miniImageNet." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.146, + 0.764, + 0.163 + ], + "angle": 0, + "content": "5.5 FewSTAB Enables Designs for Varied Degrees of Robustness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.49, + 0.539 + ], + "angle": 0, + "content": "As demonstrated in Section 5.3, Few-STAB can benchmark spurious bias in varied degrees, which in turn enables practitioners to design robust few-shot classifiers targeted for different degrees of robustness to spurious bias. The reason for differentiating designs for varied degrees of robustness is that the same design choice may not work under different robustness requirements. For example, increasing shot number in training tasks is a common strategy for improving the few-shot generalization of meta-learning based methods. We trained few-shot classifiers with 5-way 5-shot and 5-way 1-shot training tasks randomly sampled from \\(\\mathcal{D}_{train}\\), respectively. We then calculated the accuracy gap defined as the wAcc-A of a model trained on 5-shot tasks minus the wAcc-A of the same model trained on 1-shot tasks. A positive and large accuracy gap indicates that this strat" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.645 + ], + "angle": 0, + "content": "egy is effective in improving the model's robustness to spurious bias. In Fig. 6, on each of the three datasets, we give results of the eight meta-learning based FSC methods on the 5-way 1-, 5-, and 10-shot FewSTAB tasks which are used to demonstrate the strategy's robustness to increased degrees of spurious bias. This strategy does not work consistently under different test shots. For example, in Fig. 6(a) this strategy with CAN only works the best on the 5-way 5-shot FewSTAB tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.403, + 0.687 + ], + "angle": 0, + "content": "5.6 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Techniques used in FewSTAB. We analyze how different sample selection methods affect the effectiveness of FewSTAB in Table 5. With only intra-class attribute-based sample selection, we randomly select query samples from Eq. ③. For inter-class attribute-based sample selection and intra-class attribute-based sample selection (automatically included by Eq. ④), we randomly select query samples from Eq. ⑤. FewSTAB uses all the techniques in Table 5. We define accuracy drop as wAcc-R minus wAcc-A, and we use the drop averaged over the ten FSC methods tested on 5-way 5-shot tasks from the miniImageNet dataset as our metric. A larger average drop indicates that the corresponding" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.209, + 0.785, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.281, + 0.785, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.352, + 0.787, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.444, + 0.788, + 0.5 + ], + "angle": 0, + "content": "Fig. 6: Accuracy gaps of few-shot classifiers tested on 1-shot, 5-shot, and 10-shot tasks constructed from (a) miniImageNet, (b) tieredImageNet, and (c) CUB-200 datasets." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.505, + 0.187 + ], + "angle": 0, + "content": "Table 5: Comparison between different techniques used by FewSTAB for constructing FSC tasks." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.199, + 0.477, + 0.273 + ], + "angle": 0, + "content": "
Attribute-based sample selectionQuery sample selectionAvg. drop (%)
Intra-classInter-class
✓5.13
✓✓13.30
✓✓✓15.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 6: Spearman's rank correlation coefficients between wAcc-A obtained using ViT-GPT2 and BLIP." + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.2, + 0.747, + 0.263 + ], + "angle": 0, + "content": "
Dataset1-shot5-shot
miniImageNet0.981.00
tieredImageNet1.000.99
CUB-2001.000.98
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.786, + 0.345 + ], + "angle": 0, + "content": "sample selection method is more effective in reflecting the spurious bias in few-shot classifiers. We observe that all proposed techniques are effective and the inter-class attribute-based sample selection is the most effective method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Choice of VLMs. Although our main results are based on the pre-trained ViT-GPT2 model [32], we show in Table 6 that when switching to a different VLM, i.e., BLIP [23], the relative ranks of different few-shot classifiers based on wAcc-A still hold with high correlations. In other words, FewSTAB is robust to different choices of VLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.588 + ], + "angle": 0, + "content": "Detection accuracy of VLMs. A VLM may miss some attributes due to its limited capacity, resulting in a small detection accuracy. However, the detection accuracy of a VLM has little impact on our framework. To demonstrate this, we adopt a cross-validation strategy, i.e., we use the outputs from one VLM as the ground truth to evaluate those from another VLM, since assessing the detection accuracy of a VLM typically requires labor-intensive human labeling. On the CUB-200 dataset, we observe that the detection accuracy of ViT-GPT2 based on the BLIP's outputs is \\(70.12\\%\\), while the detection accuracy of BLIP based on the ViT-GPT2's outputs is \\(59.28\\%\\). Although the two VLMs differ significantly in the detected attributes, our framework shows almost consistent rankings of the evaluated FSC methods (Table 6)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.588, + 0.547, + 0.603 + ], + "angle": 0, + "content": "Additional results are presented in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.627, + 0.36, + 0.643 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this paper, we proposed a systematic and rigorous benchmark framework called FewSTAB for evaluating the robustness of few-shot classifiers to spurious bias. FewSTAB adopts attribute-based sample selection strategies to construct FSC test tasks with biased attributes so that the reliance on spurious correlations can be effectively revealed. FewSTAB can automatically benchmark spurious bias in few-shot classifiers on any existing test data thanks to its use of a pretrained VLM for automated attribute detection. With FewSTAB, we provided a new dimension of evaluation on the robustness of few-shot image classifiers to spurious bias and a new design guideline for building robust few-shot classifiers. FewSTAB can reveal and enable designs for varied degrees of robustness to spurious bias. We hope FewSTAB will inspire new developments on designing robust few-shot classifiers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.209 + ], + "angle": 0, + "content": "This work is supported in part by the US National Science Foundation under grants 2313865, 2217071, 2213700, 2106913, 2008208, 1955151." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.235, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.309 + ], + "angle": 0, + "content": "1. Ahmed, F., Bengio, Y., Van Seijen, H., Courville, A.: Systematic generalisation with group invariant predictions. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.786, + 0.338 + ], + "angle": 0, + "content": "2. Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.: Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.787, + 0.379 + ], + "angle": 0, + "content": "3. Baker, N., Lu, H., Erlikhman, G., Kellman, P.J.: Deep convolutional networks do not classify based on global object shape. PLoS Computational Biology 14(12), e1006613 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.786, + 0.408 + ], + "angle": 0, + "content": "4. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: European Conference on Computer Vision. pp. 456-473 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.786, + 0.449 + ], + "angle": 0, + "content": "5. Bertinetto, L., Henriques, J.F., Torr, P., Vedaldi, A.: Meta-learning with differentiable closed-form solvers. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.786, + 0.491 + ], + "angle": 0, + "content": "6. Chen, W.Y., Liu, Y.C., Kira, Z., Wang, Y.C.F., Huang, J.B.: A closer look at few-shot classification. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.786, + 0.519 + ], + "angle": 0, + "content": "7. Chen, Y., Wang, X., Liu, Z., Xu, H., Darrell, T.: A new meta-baseline for few-shot learning. arXiv preprint arXiv:2003.04390 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.786, + 0.575 + ], + "angle": 0, + "content": "8. Dong, J., Wang, Y., Lai, J.H., Xie, X.: Improving adversarially robust few-shot image classification with generalizable representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9025-9034 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.786, + 0.617 + ], + "angle": 0, + "content": "9. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Proceedings of the 34th International Conference on Machine Learning. vol. 70, pp. 1126–1135 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.786, + 0.658 + ], + "angle": 0, + "content": "0. Flennerhag, S., Rusu, A.A., Pascanu, R., Visin, F., Yin, H., Hadsell, R.: Meta-learning with warped gradient descent. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.786, + 0.714 + ], + "angle": 0, + "content": "1. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.716, + 0.786, + 0.77 + ], + "angle": 0, + "content": "12. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.786, + 0.798 + ], + "angle": 0, + "content": "13. Ghosal, S.S., Li, Y.: Are vision transformers robust to spurious correlations? International Journal of Computer Vision 132(3), 689-709 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "14. Goldblum, M., Fowl, L., Goldstein, T.: Adversarily robust few-shot learning: A meta-learning approach. Advances in Neural Information Processing Systems 33, 17886-17895 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.267, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.175 + ], + "angle": 0, + "content": "15. He, Y., Shen, Z., Cui, P.: Towards non-iid image classification: A dataset and baselines. Pattern Recognition 110, 107383 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "16. Hou, R., Chang, H., Bingpeng, M., Shan, S., Chen, X.: Cross attention network for few-shot classification. In: Advances in Neural Information Processing Systems. pp. 4003-4014 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "17. Hu, S.X., Moreno, P.G., Xiao, Y., Shen, X., Obozinski, G., Lawrence, N., Dami-anou, A.: Empirical bayes transductive meta-learning with synthetic gradients. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "18. Jang, H., Lee, H., Shin, J.: Unsupervised meta-learning via few-shot pseudosupervised contrastive learning. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "19. Kang, D., Kwon, H., Min, J., Cho, M.: Relational embedding for few-shot classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8822-8833 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "20. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "21. Lee, K., Maji, S., Ravichandran, A., Soatto, S.: Meta-learning with differentiable convex optimization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10657-10665 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.426, + 0.787, + 0.466 + ], + "angle": 0, + "content": "22. Lee, Y., Choi, S.: Gradient-based meta-learning with learned layerwise metric and subspace. In: International Conference on Machine Learning. pp. 2927-2936. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "23. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "24. Li, W., Wang, L., Xu, J., Huo, J., Gao, Y., Luo, J.: Revisiting local descriptor based image-to-class measure for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7260-7268 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.551, + 0.787, + 0.592 + ], + "angle": 0, + "content": "25. Li, W., Wang, Z., Yang, X., Dong, C., Tian, P., Qin, T., Jing, H., Shi, Y., Wang, L., Gao, Y., Luo, J.: Libfewshot: A comprehensive library for few-shot learning. IEEE Transactions on Pattern Analysis & Machine Intelligence (01), 1-18 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "26. Li, Z., Zhou, F., Chen, F., Li, H.: Meta-SGD: Learning to learn quickly for few shot learning. ArXiv abs/1707.09835 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "27. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3730-3738 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "28. Lu, Y., Wen, L., Liu, J., Liu, Y., Tian, X.: Self-supervision can be a good few-shot learner. In: European Conference on Computer Vision. pp. 740-758. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.702, + 0.787, + 0.73 + ], + "angle": 0, + "content": "29. Motiian, S., Jones, Q., Iranmanesh, S., Doretto, G.: Few-shot adversarial domain adaptation. Advances in Neural Information Processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.731, + 0.787, + 0.757 + ], + "angle": 0, + "content": "30. Myers, L., Sirois, M.J.: Spearman correlation coefficients, differences between. Encyclopedia of Statistical Sciences 12 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "31. Neuhaus, Y., Augustin, M., Boreiko, V., Hein, M.: Spurious features everywhere large-scale detection of harmful spurious features in imagenet. arXiv preprint arXiv:2212.04871 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "32. NLP Connect: vit-gpt2-image-captioning (revision 0e334c7) (2022). https://doi.org/10.57967/hf/0222, https://huggingface.co/nlpconnect/vit-gpt2-image-captioning" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Nushi, B., Kamar, E., Horvitz, E.: Towards accountable ai: Hybrid human-machine analyses for characterizing system failure. In: Proceedings of the AAAI Conference on Human Computation and Crowdsourcing. vol. 6, pp. 126-135 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.218 + ], + "angle": 0, + "content": "34. Oh, J., Yoo, H., Kim, C., Yun, S.Y.: Boil: Towards representation change for few-shot learning. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.259 + ], + "angle": 0, + "content": "35. Oreshkin, B., Rodríguez López, P., Lacoste, A.: Tadam: Task dependent adaptive metric for improved few-shot learning. In: Advances in Neural Information Processing Systems 31, pp. 721-731 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.286 + ], + "angle": 0, + "content": "36. Poulakakis-Daktylidis, S., Jamali-Rad, H.: Bectr: Batch enhanced contrastive few-shot learning. In: International Conference on Learning Representations (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "37. Qin, X., Song, X., Jiang, S.: Bi-level meta-learning for few-shot domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15900-15910 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.384 + ], + "angle": 0, + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.425 + ], + "angle": 0, + "content": "39. Raghu, A., Raghu, M., Bengio, S., Vinyals, O.: Rapid learning or feature reuse? towards understanding the effectiveness of maml. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.452 + ], + "angle": 0, + "content": "40. Ravi, S., Larochelle, H.: Optimization as a model for few-shot learning. In: International Conference on Learning Representations (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.493 + ], + "angle": 0, + "content": "41. Ren, M., Ravi, S., Triantafillou, E., Snell, J., Swersky, K., Tenenbaum, J.B., Larochelle, H., Zemel, R.S.: Meta-learning for semi-supervised few-shot classification. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.536 + ], + "angle": 0, + "content": "42. Rusu, A.A., Rao, D., Sygnowski, J., Vinyals, O., Pascanu, R., Osindero, S., Hadsell, R.: Meta-learning with latent embedding optimization. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.563 + ], + "angle": 0, + "content": "43. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.591 + ], + "angle": 0, + "content": "44. Singla, S., Feizi, S.: Salient imagenet: How to discover spurious features in deep learning? In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.633 + ], + "angle": 0, + "content": "45. Singla, S., Nushi, B., Shah, S., Kamar, E., Horvitz, E.: Understanding failures of deep networks via robust feature extraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12853-12862 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.661 + ], + "angle": 0, + "content": "46. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems 30, pp. 4077-4087 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.702 + ], + "angle": 0, + "content": "47. Stock, P., Cisse, M.: Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases. In: European Conference on Computer Vision. pp. 498-512 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "48. Sun, Q., Liu, Y., Chua, T.S., Schiele, B.: Meta-transfer learning for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 403-412 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.799 + ], + "angle": 0, + "content": "49. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H.S., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1199-1208 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "50. Tian, Y., Wang, Y., Krishnan, D., Tenenbaum, J.B., Isola, P.: Rethinking few-shot image classification: a good embedding is all you need? In: European Conference on Computer Vision. pp. 266-282. Springer (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "51. Triantafillou, E., Zhu, T., Dumoulin, V., Lamblin, P., Evci, U., Xu, K., Goroshin, R., Gelada, C., Swersky, K., Manzagol, P.A., Larochelle, H.: Meta-dataset: A dataset of datasets for learning to learn from few examples. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.788, + 0.244 + ], + "angle": 0, + "content": "52. Tseng, H.Y., Lee, H.Y., Huang, J.B., Yang, M.H.: Cross-domain few-shot classification via learned feature-wise transformation. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.788, + 0.285 + ], + "angle": 0, + "content": "53. Vinyals, O., Blundell, C., Lillicrap, T., kavukcuoglu, k., Wierstra, D.: Matching networks for one shot learning. In: Advances in Neural Information Processing Systems 29, pp. 3630-3638 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.788, + 0.326 + ], + "angle": 0, + "content": "54. Wang, Y., Girshick, R., Hebert, M., Hariharan, B.: Low-shot learning from imaginary data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7278-7286 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.326, + 0.788, + 0.354 + ], + "angle": 0, + "content": "55. Wang, Y., Yao, Q., Kwok, J.T., Ni, L.M.: Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys 53(3), 1-34 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.354, + 0.788, + 0.394 + ], + "angle": 0, + "content": "56. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-UCSD Birds 200. Tech. Rep. CNS-TR-2010-001, California Institute of Technology (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.788, + 0.435 + ], + "angle": 0, + "content": "57. Wong, E., Santurkar, S., Madry, A.: Leveraging sparse linear layers for debuggable deep networks. In: International Conference on Machine Learning. pp. 11205-11216. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.435, + 0.788, + 0.476 + ], + "angle": 0, + "content": "58. Xiao, K.Y., Engstrom, L., Ilyas, A., Madry, A.: Noise or signal: The role of image backgrounds in object recognition. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.788, + 0.503 + ], + "angle": 0, + "content": "59. Xue, Y., Payani, A., Yang, Y., Mirzasoleiman, B.: Eliminating spurious correlations from pre-trained models via data mixing. arXiv preprint arXiv:2305.14521 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.503, + 0.788, + 0.544 + ], + "angle": 0, + "content": "60. Yang, Y., Nushi, B., Palangi, H., Mirzasoleiman, B.: Mitigating spurious correlations in multi-modal models during fine-tuning. arXiv preprint arXiv:2304.03916 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.788, + 0.585 + ], + "angle": 0, + "content": "61. Ye, H.J., Hu, H., Zhan, D.C., Sha, F.: Few-shot learning via embedding adaptation with set-to-set functions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8808-8817 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.585, + 0.788, + 0.613 + ], + "angle": 0, + "content": "62. Yue, X., Zheng, Z., Das, H.P., Keutzer, K., Vincentelli, A.S.: Multi-source few-shot domain adaptation. arXiv preprint arXiv:2109.12391 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.613, + 0.788, + 0.639 + ], + "angle": 0, + "content": "63. Yue, Z., Zhang, H., Sun, Q., Hua, X.S.: Interventional few-shot learning. Advances in Neural Information Processing Systems 33, 2734-2746 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.639, + 0.788, + 0.667 + ], + "angle": 0, + "content": "64. Zhang, C., Cai, Y., Lin, G., Shen, C.: Deepemd: Differentiable earth mover's distance for few-shot learning (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.667, + 0.788, + 0.707 + ], + "angle": 0, + "content": "65. Zhang, C., Cai, Y., Lin, G., Shen, C.: Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.707, + 0.788, + 0.748 + ], + "angle": 0, + "content": "66. Zhang, J., Wang, Y., Molino, P., Li, L., Ebert, D.S.: Manifold: A model-agnostic framework for interpretation and diagnosis of machine learning models. IEEE Transactions on Visualization and Computer Graphics 25(1), 364-373 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.748, + 0.788, + 0.789 + ], + "angle": 0, + "content": "67. Zhang, M., Li, H., Wu, F., Kuang, K.: Metacoco: A new few-shot classification benchmark with spurious correlation. In: The Twelfth International Conference on Learning Representations (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.789, + 0.788, + 0.831 + ], + "angle": 0, + "content": "68. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.831 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_origin.pdf b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1aae6287409dfbb5d58123aa673d4d6f8934fc0f --- /dev/null +++ b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/3a32a91c-4112-4ab6-b169-da66f799dbb5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58be245c84fb571d81faf15955c948d716fb295157dd36928788ac679390b64c +size 2559250 diff --git a/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/full.md b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5699f32c54d6915754a74144f38e41abfb2a2e96 --- /dev/null +++ b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/full.md @@ -0,0 +1,337 @@ +# Benchmarking Spurious Bias in Few-Shot Image Classifiers + +Guangtao Zheng1, Wenqian Ye1, and Aidong Zhang1 + +1University of Virginia, Charlottesville VA 22904, USA {gz5hp,wenqian,aidong}@virginia.edu + +Abstract. Few-shot image classifiers are designed to recognize and classify new data with minimal supervision and limited data but often show reliance on spurious correlations between classes and spurious attributes, known as spurious bias. Spurious correlations commonly hold in certain samples and few-shot classifiers can suffer from spurious bias induced from them. There is an absence of an automatic benchmarking system to assess the robustness of few-shot classifiers against spurious bias. In this paper, we propose a systematic and rigorous benchmark framework, termed FewSTAB, to fairly demonstrate and quantify varied degrees of robustness of few-shot classifiers to spurious bias. Few-STAB creates few-shot evaluation tasks with biased attributes so that using them for predictions can demonstrate poor performance. To construct these tasks, we propose attribute-based sample selection strategies based on a pre-trained vision-language model, eliminating the need for manual dataset curation. This allows FewSTAB to automatically benchmark spurious bias using any existing test data. FewSTAB offers evaluation results in a new dimension along with a new design guideline for building robust classifiers. Moreover, it can benchmark spurious bias in varied degrees and enable designs for varied degrees of robustness. Its effectiveness is demonstrated through experiments on ten few-shot learning methods across three datasets. We hope our framework can inspire new designs of robust few-shot classifiers. Our code is available at https://github.com/gtzheng/FewSTAB + +Keywords: Few-shot classification $\cdot$ Spurious bias $\cdot$ Robustness $\cdot$ Benchmark system + +# 1 Introduction + +Few-shot classification [5,21,46,61,64] (FSC) has attracted great attention recently due to its promise for recognizing novel classes efficiently with limited data. Few-shot classifiers can transfer the knowledge learned from base classes to recognize novel classes with a few labeled samples. However, they face potential risks when deployed in the real world, such as data distribution shifts [29,62] and adversarial examples [8,14]. A subtle yet critical risk factor is the spurious correlations [3,4,11,13,43,47,59] between classes and spurious attributes — + +![](images/9615ac1307e59941c988adebb0cadcaf3289cd17d9f320a16aedf8d536e05498.jpg) +Learned path + +![](images/e57ac34f47a9e4f529ee4d44feb0c29dae1b31546867bfbab0f09309258e8b57.jpg) +Non-existing path +Fig. 1: Exploiting the spurious correlation between the class bird and the spurious attribute tree branch to predict bird leads to an incorrect prediction on the test image showing birds on a grass field. For clarity, we only show the case for one class. + +attributes of inputs non-essential to the classes. In the traditional learning setting [1,2,43], deep learning models tend to rely on spurious correlations as their prediction shortcuts or exhibit spurious bias, such as predicting classes using the associated backgrounds [58] or image textures [12], leading to significant performance drops when the associated backgrounds or textures change to different ones. In the low-data regime, spurious bias becomes more evident. For example, in Fig. 1, the correlation between the class bird and the attribute tree branch in the support (training) image may form a shortcut path from tree branch to predicting the image as bird and hinder the learning of the desired one that uses class-related attributes, such as head, tail, and wing. The shortcut will fail to generalize in the query (test) image where no tree branch can be found. In general, few-shot image classifiers are susceptible to spurious bias. + +However, there lacks a dedicated benchmarking framework that evaluates the robustness of few-shot classifiers to spurious bias. The standard benchmarking procedure in FSC trains a few-shot classifier on base classes from a training set with ample samples and evaluates the classifier on FSC test tasks constructed from a test set with novel classes. The problem with this procedure is the lack of explicit control over the spurious correlations in the constructed FSC tasks. Each FSC test task contains randomly sampled support and query samples. Thus, spurious correlations in the majority of the test set samples can be demonstrated in these tasks, providing unfair advantages for few-shot classifiers with high reliance on the spurious correlations. + +In this paper, we propose a systematic and rigorous benchmark framework, termed Few-Shot Tasks with Attribute Biases (FewSTAB), to fairly compare the robustness of various few-shot classifiers to spurious bias. Our framework explicitly controls spurious correlations in the support and query samples when constructing an FSC test task to reveal the robustness pitfalls caused by spurious bias. To achieve this, we propose attribute-based sample selection strategies that select support and query samples with biased attributes. These attributes together with their associated classes formulate spurious correlations such that if the support samples induce spurious bias in a few-shot classifier, i.e., the classifier learns the spurious correlations in the support samples as its prediction shortcuts, then the query samples can effectively degrade the classifier's performance, exposing its non-robustness to spurious bias. + +Our framework exploits the spurious attributes in test data for formulating spurious correlations in FSC test tasks. Some existing datasets [15,27,43] provide spurious attribute annotations. However, they only have a few classes and cannot provide enough classes for training and testing. Many benchmark datasets for FSC do not have annotations on spurious attributes, and obtaining these annotations typically involves labor-intensive human-guided labeling [33,66]. To address this, we further propose to use a pre-trained vision-language model (VLM) to automatically identify distinct attributes in images in the high-level text format. Our attribute-based sampling methods can use the identified attributes to simulate various spurious correlations. Thus, we can reuse any existing FSC datasets for benchmarking few-shot classifiers' robustness to spurious bias, eliminating the need for the manual curation of new datasets. + +The main contributions of our work are summarized as follows: + +- We propose a systematic and rigorous benchmark framework, termed Few-Shot Tasks with Attribute Biases (FewSTAB), that specifically targets spurious bias in few-shot classifiers, demonstrates their varied degrees of robustness to spurious bias, and benchmarks spurious bias in varied degrees. +- We propose novel attribute-based sample selection strategies using a pretrained VLM for constructing few-shot evaluation tasks, allowing us to reuse any existing few-shot benchmark datasets without manually curating new ones for the evaluation. +- FewSTAB provides a new dimension of evaluation on the robustness to spurious bias along with a new design guideline for building robust few-shot classifiers. We demonstrate the effectiveness of FewSTAB by applying it to models trained on three benchmark datasets with ten FSC methods. + +# 2 Related Work + +Few-shot classification. Few-shot classification [6,46,50,51,53,55] has received vast attention recently. Few-shot classifiers can be trained with meta-learning or transfer learning on base classes to learn the knowledge that can be transferred to recognize novel classes with a few labeled samples. The transfer learning approaches [6,50] first learn a good embedding model and then fine-tune the model on samples from novel classes. The meta-learning approaches can be further divided into optimization-based and metric-based methods. The optimization-based methods [9,10,22,26,54] aim to learn a good initialized model such that the model can adapt to novel classes efficiently with a few gradient update steps on a few labeled samples. The metric-based methods [5,21,35,42,49,53,61,65] aim to learn a generalizable representation space with a well-defined metric, such as Euclidean distance [46], to learn novel classes with a few labeled samples. Recently, large vision-language models [20,38,68] are used for few-shot classification. However, they have completely different training and inference pipelines from the models that we consider in this paper. + +Robustness in few-shot classification. There are several notions of robustness for few-shot classifiers. The common one requires a few-shot classifier to + +perform well on the in-distribution samples of novel classes in randomly sampled FSC test tasks. The robustness to adversarial perturbations further requires a few-shot classifier to perform well on samples with imperceptible perturbations [8,14]. Moreover, the cross-domain generalization [37,51,52] aims to test how robust a few-shot classifier is on samples from novel classes with domain shifts, which are typically reflected by the changes in both image styles and classes. In contrast, we focus on a new notion of robustness: the robustness to spurious bias. There is a lack of rigorous evaluation methods on the topic. We provide a new evaluation method that specifically targets spurious bias and can systematically demonstrate few-shot classifiers' varied degrees of vulnerability to spurious bias, which has not been addressed in the existing literature. + +Benchmarks for spurious bias. There are some existing datasets [15,27,43] that are designed to benchmark spurious bias in image classifiers. However, these datasets are only applicable to the traditional learning setting [1,2,43] since the classes in them are not sufficient for the training and testing of few-shot classifiers. Existing benchmarks in few-shot classification are not tailored for benchmarking spurious bias in few-shot classifiers. A recent work [67] creates a large-scale few-shot classification benchmark dataset with spurious-correlation shifts. In contrast, we propose a benchmark framework that can reuse existing few-shot classification datasets and provide a new dimension of evaluation. + +Discovering spurious attributes. A spurious attribute is non-essential to a class and only exists in some samples. Early works on discovering spurious attributes [33,66] require a predefined list of spurious attributes and expensive human-guided labeling of visual attributes. Recent works [31,44,45,57] greatly reduce the need for manual annotations by using the neurons of robust models to detect visual attributes. However, they still need humans to annotate the detected visual attributes. We automate this process by using a pre-trained VLM to obtain distinct attributes as words. Instead of discovering spurious correlations, we simulate them via attribute-based sampling for benchmarking. + +# 3 Preliminary + +Few-shot classification tasks. A typical FSC task $\mathcal{T}$ has a support set $S$ for training and a query set $\mathcal{Q}$ for testing. In this task, there are $C$ classes $(c = 1,\dots ,C)$ with $N_{\mathcal{S}}$ (a small number) training samples and $N_{\mathcal{Q}}$ test samples per class in $S$ and $\mathcal{Q}$ , respectively. The task is called a $C$ -way $N_{\mathcal{S}}$ -shot task. + +Few-shot classifiers. A few-shot classifier $f_{\theta}$ with parameters $\theta$ aims to classify the samples in $\mathcal{Q}$ after learning from $S$ with a learning algorithm $\mathcal{O}$ in a few-shot task $\mathcal{T}$ . Here, $\mathcal{O}$ could be any learning algorithms, such as the optimization method [9] or a prototype-based classifier learning method [7,46]. To acquire a good few-shot learning capability, $f_{\theta}$ is typically meta-trained or pre-trained [25] on a base training set $\mathcal{D}_{train} = \{(x_n, y_n) | y_n \in \mathcal{C}_{train}, n = 1, \dots, N_{train}\}$ with $N_{train}$ sample(x)-label(y) pairs, where $\mathcal{C}_{train}$ is a set of base classes. + +Performance metrics. The performance of a few-shot classifier is typically measured by its average classification accuracy over $N_{\mathcal{T}}$ $C$ -way $N_{S}$ -shot tasks + +Table 1: Meanings of major symbols used in the paper. + +
SymbolMeaning
TAn FSC task
SSupport (training) set in T
QQuery (test) set in T
cA class in T
CNumber of classes per task
NsNumber of samples (shots) per class in S
NQNumber of samples per class in Q
OA few-shot adaptation algorithm
ψAn attribute detector
ΩAn automatic word selector
DtrainThe base training set
DvalThe validation set for selecting a few-shot classifier
DtestThe test set for evaluating a few-shot classifier
CtrainClasses in Dtrain
CtestClasses in Dtest
DcA set of all samples belonging to class c
aA text-format attribute
AA set of text-format attributes
+ +randomly sampled from $\mathcal{D}_{test} = \{(x_n, y_n) | y_n \in \mathcal{C}_{test}, n = 1, \dots, N_{test}\}$ where $N_{test}$ sample-label pairs from novel classes $\mathcal{C}_{test}$ do not appear in $\mathcal{D}_{train}$ , i.e., $\mathcal{C}_{train} \cap \mathcal{C}_{test} = \emptyset$ . We denote this metric as standard accuracy $\mathrm{Acc}(f_\theta)$ , i.e., + +$$ +\operatorname {A c c} \left(f _ {\theta}\right) = \frac {1}{N _ {\mathcal {T}}} \sum_ {t = 1} ^ {N _ {\mathcal {T}}} \sum_ {c = 1} ^ {C} M _ {c} \left(\mathcal {T} _ {t}; f _ {\theta}, \mathcal {O}\right), \tag {1} +$$ + +where $M_{c}(\mathcal{T}_{t};f_{\theta},\mathcal{O})$ denotes the classification accuracy of $f_{\theta}$ on the query samples from the class $c$ in $\mathcal{T}_t$ after $f_{\theta}$ is trained on $S$ with $\mathcal{O}$ . The metric $\mathrm{Acc}(f_{\theta})$ in Eq. (11) only shows the average learning capability of $f_{\theta}$ over $C$ randomly selected novel classes. To better characterize the robustness of $f_{\theta}$ to spurious bias, we define the class-wise worst classification accuracy over tasks as + +$$ +\operatorname {w A c c} \left(f _ {\theta}\right) = \frac {1}{N _ {\mathcal {T}}} \sum_ {t = 1} ^ {N _ {\mathcal {T}}} \min _ {c = 1, \dots , C} M _ {c} \left(\mathcal {T} _ {t}; f _ {\theta}, \mathcal {O}\right). \tag {2} +$$ + +A larger $\mathrm{wAcc}(f_{\theta})$ indicates that $f_{\theta}$ is more robust to spurious bias. + +Spurious correlations. A spurious correlation is the association between a class and an attribute of inputs that is non-essential to the class, and it only holds in some samples. We formally define it as follows. + +Definition 1. Let $\mathcal{D}_c$ denote a set of sample-label pairs having the label $c$ , and let $\psi : \mathcal{X} \to \mathcal{B}_A$ be an attribute detector, where $\mathcal{X}$ is the set of all possible inputs, $\mathcal{B}_A$ denotes all possible subsets of $\mathcal{A}$ , and $\mathcal{A}$ is the set of all possible attributes. The class $c$ and an attribute $a \in \mathcal{A}$ form a spurious correlation, denoted as $\langle c, a \rangle$ , if and only if the following conditions hold: + +1. There exists $(x,c)\in \mathcal{D}_c$ that satisfies $a\in \psi (x)$ , and + +![](images/3ec32a50922a691a3cb03923c1de7f2bb667cfe8e4bd69072773176976fd3941.jpg) +Fig. 2: FewSTAB overview. (a) Extract distinct attributes using a pre-trained VLM. (b) Generate an FSC task for the evaluation of spurious bias in few-shot classifiers. + +# 2. There exists $(x', c) \in \mathcal{D}_c$ that satisfies $a \notin \psi(x')$ . + +We define $a$ as the spurious attribute in $\langle c, a \rangle$ . + +Definition $\mathbb{I}$ specifies that all the spurious correlations are based on $\mathcal{D}_c$ . In the remainder of the paper, we define $\mathcal{D}_c = \{(x,c)|\forall (x,c)\in \mathcal{D}_{test}\}$ with $c\in \mathcal{C}_{test}$ as we focus on evaluating the robustness to spurious bias. + +We list major symbols in the paper alongside their meanings in Tab. 1. + +# 4 Methodology + +# 4.1 Attribute-Based Sample Selection + +We first propose two attribute-based sample selection methods to reveal spurious bias in a few-shot classifier. Consider a training set $\mathcal{S}$ in a few-shot test task $\mathcal{T}$ , which has $C$ classes with each class $c\in \mathcal{C}_{test}$ associating with a unique spurious attribute $a\in \mathcal{A}$ . We aim to discover samples that can exhibit a classifier's spurious bias on $\langle c,a\rangle$ induced from $\mathcal{S}$ . Motivated by existing findings [43,59,60] that classifiers with high reliance on $\langle c,a\rangle$ tend to perform poorly on samples without it, we propose an attribute-based sample selection strategy below. + +Intra-class attribute-based sample selection. Given $\mathcal{D}_c$ and the training set $S$ having the spurious correlation $\langle c, a \rangle$ , we generate a set $\mathcal{I}_{\langle c, a \rangle}$ of sample-label pairs which have class $c$ but do not contain attribute $a$ , i.e., + +$$ +\mathcal {I} _ {\langle c, a \rangle} = \{(x, c) | \forall (x, c) \in \mathcal {D} _ {c}, a \notin \psi (x) \}. \tag {3} +$$ + +The above proposed method demonstrates a few-shot classifier's robustness to individual spurious correlation $\langle c, a \rangle$ and does not consider a multi-class classification setting where spurious attributes from some other class $c'$ exist in samples of the class $c$ . In this case, these attributes may mislead the classifier to + +predict those samples as the class $c'$ and severely degrade the performance on the class $c$ . For example, consider using the spurious correlations $\langle \text{vase}, \text{blue} \rangle$ and $\langle \text{bowl}, \text{green} \rangle$ for predicting vase and bowl, respectively. An image showing a vase in green is more effective in revealing the robustness to $\langle \text{vase}, \text{blue} \rangle$ as it is more likely to be misclassified as bowl than other images. Motivated by this, we propose the inter-class attribute-based sample selection below. + +Inter-class attribute-based sample selection. Given $\mathcal{D}_c$ and the training set $S$ having the spurious correlations $\langle c, a \rangle$ and $\langle c', a' \rangle$ , where $c' \neq c$ and $a' \neq a$ , we generate a set $\mathcal{I}_{\langle c, a \rangle}^{\langle c', a' \rangle}$ of sample-label pairs which have class $c$ , do not contain attribute $a$ , but contain attribute $a'$ from another class $c'$ : + +$$ +\mathcal {I} _ {\langle c, a \rangle} ^ {\langle c ^ {\prime}, a ^ {\prime} \rangle} = \mathcal {I} _ {\langle c, a \rangle} \cap \left\{\left(x, c\right) | \forall (x, c) \in \mathcal {D} _ {c}, a _ {\langle c ^ {\prime} \rangle} ^ {\prime} \in \psi (x) \right\}, \tag {4} +$$ + +where $a_{\langle c^{\prime}\rangle}^{\prime}$ denotes $a^\prime$ in $\langle c',a'\rangle$ , and $\mathcal{I}_{\langle c,a\rangle}$ is defined in Eq. (3). + +Considering that there are $C$ classes in the training set $S$ with each class associating with a unique spurious attribute $a$ , to effectively demonstrate the reliance on the spurious correlation $\langle c, a \rangle$ with the inter-class attribute-based sample selection, we consider all the spurious correlations in $S$ . Specifically, we apply the above selection strategy to all the $C - 1$ spurious correlations in $S$ other than $\langle c, a \rangle$ and obtain $\mathcal{I}_{\langle c, a \rangle}^{C}$ as the union of the $C - 1$ sets as follows: + +$$ +\mathcal {I} _ {\langle c, a \rangle} ^ {C} = \bigcup_ {\langle c ^ {\prime}, a ^ {\prime} \rangle \in \mathcal {C} \backslash c} \mathcal {I} _ {\langle c, a \rangle} ^ {\langle c ^ {\prime}, a ^ {\prime} \rangle}, \tag {5} +$$ + +where $\mathcal{C}_{\backslash c}$ denotes all the spurious correlations in $S$ other than $\langle c, a \rangle$ . + +The inter-class attribute-based sample selection is built upon the intra-class attribute-based sample selection. In the remainder of the paper, we use the inter-class method as our default sample selection strategy, which is more effective empirically (Sec. 5.6). In certain cases, however, where there are not enough desired samples during task construction, we resort to the intra-class sample selection strategy (Sec. 5.1, Implementation details). + +In the following, we introduce FewSTAB, a benchmark framework that uses the proposed selection strategies to construct FSC tasks containing samples with biased attributes for benchmarking spurious bias in few-shot classifiers. + +# 4.2 FewSTAB (Part 1): Text-Based Attribute Detection + +Our attribute-based sample selection methods require knowing the attributes in images, which typically involves labor-intensive human labeling. To make our method scalable and applicable to few-shot classifiers trained on different datasets, we adopt a pre-trained VLM to automatically identify distinct attributes in images in text format, which includes the following two steps. + +Step 1: Generating text descriptions. We use a pre-trained VLM [23, 32] $\phi$ to automatically generate text descriptions for images in $\mathcal{D}_{test}$ . The VLM is a model in the general domain and can produce text descriptions for various + +objects and patterns. For example, for the current input image in the vase class in Fig. 2(a), besides the class object vase, the VLM also detects the vase's color green, and another object table with its material wooden. + +Step 2: Extracting informative words. From the generated text descriptions, we extract nouns and adjectives as the detected attributes via an automatic procedure $\Omega$ . The two kinds of words are informative as a noun describes an object, and an adjective describes a property of an object. All the detected attributes form the candidate attribute set $\mathcal{A}$ . We realize the attribute detector $\psi$ defined in Definition 1 as $\psi(x) = \Omega(\phi(x))$ . + +Remark 1: A VLM in general can extract many distinct attributes from the images. On some images, the VLM may detect non-relevant attributes, such as detecting a duck from a bird image. A more capable VLM could warrant a better attribute detection accuracy and benefit individual measurements on few-shot classifiers. Although being a VLM-dependent benchmark framework, FewSTAB can produce consistent and robust relative measurements among all the compared FSC methods, regardless the choice of VLMs (Sec. 5.6). + +Remark 2: The candidate set $\mathcal{A}$ constructed with all the extracted words may contain attributes that represent the classes in $\mathcal{D}_{test}$ . However, during our attribute-based sample selection, these attributes will not be used since they always correlate with classes and therefore do not satisfy the definition of spurious attributes in Definition 1. We provide details of $\phi$ and $\Omega$ in Sec. 5.1 + +# 4.3 FewSTAB (Part 2): FSC Task Construction + +Constructing a $C$ -way $N_{S}$ -shot FSC task $\mathcal{T}$ for benchmarking spurious bias in few-shot classifiers involves constructing a support (training) set $\mathcal{S}$ and a query (test) $\mathcal{Q}$ with biased attributes. + +Constructing the support set. The support set contains the spurious correlations that we aim to demonstrate to a few-shot classifier. As a fair and rigorous benchmark system, FewSTAB makes no assumptions on the few-shot classifiers being tested and randomly samples $C$ classes from $\mathcal{C}_{test}$ . For each sampled class, it randomly selects a spurious correlation $\langle c, a \rangle$ in $\mathcal{D}_{test}$ with $a \in \mathcal{A}$ . To effectively demonstrate the spurious correlation $\langle c, a \rangle$ to a few-shot classifier, we select samples of the class $c$ such that (1) they all have the spurious attribute $a$ and (2) do not have spurious attributes from the other $C - 1$ spurious correlations. We construct $S_{c}$ with $N_{S}$ samples for the class $c$ that satisfy the above two conditions. Thus, the spurious attribute $a$ becomes predictive of the class $c$ in $S_{c}$ . We take the union of all $C$ such sets to get $S = \cup_{c=1}^{C} S_{c}$ . Fig. 2(b) demonstrates the case when $C = 3$ . Note that we have no requirements for other non-selected attributes in $A$ to ensure that we have enough samples for $S_{c}$ . + +Constructing the query set. To evaluate the robustness to the spurious correlations formulated in $\mathcal{S}$ , we first construct a candidate set $\mathcal{I}_{\langle c,a\rangle}^{C}$ in Eq. (5) for each spurious correlation $\langle c,a\rangle$ in $\mathcal{S}$ . Since we have no requirements on the non-selected attributes that are not used to formulate spurious correlations in $\mathcal{S}$ , a few-shot classifier may predict query samples via some of these attributes, + +e.g., the yellow blocks in Fig. 2(b), bypassing the test on the formulated spurious correlations in $\mathcal{S}$ . To address this, we propose query sample selection below. + +Query sample selection: We select query samples from $\mathcal{I}_{\langle c,a\rangle}^{C}$ that are least likely to have non-selected spurious attributes, such as the ones enclosed with red boxes in Fig. 2(b). To achieve this, we first calculate the fraction of sample-label pairs in $\mathcal{I}_{\langle c,a\rangle}^{C}$ that have the attribute $a$ as + +$$ +p _ {a} = | \{x \mid a \in \psi (x), \forall (x, c) \in \mathcal {I} _ {\langle c, a \rangle} ^ {C} \} | / | \mathcal {I} _ {\langle c, a \rangle} ^ {C} |, \tag {6} +$$ + +where $|\cdot|$ denotes the size of a set, $a \in \tilde{\mathcal{A}}$ , and $\tilde{\mathcal{A}}$ contains all non-selected attributes. A larger $p_a$ indicates that the attribute $a$ occurs more frequently in data and is more likely to be used in formulating prediction shortcuts. We then calculate the likelihood score for each $(x,c) \in \mathcal{I}_{\langle c,a\rangle}^C$ as $s(x) = \sum_{a \in \psi(x), a \in \tilde{\mathcal{A}}} p_a$ , i.e., the summation of all $p_a$ of non-selected attributes in $x$ . The likelihood score will be zero if there are no non-selected attributes in $x$ . A large $s(x)$ indicates that the image $x$ can be predicted via many non-selected attributes. Therefore, we select $N_{\mathcal{Q}}$ samples from $\mathcal{I}_{\langle c,a\rangle}^C$ that have the lowest likelihood scores to construct $\mathcal{Q}_c$ . Then, we have $\mathcal{Q} = \cup_{c=1}^{C} \mathcal{Q}_c$ , which contains samples for evaluating the robustness of a few-shot classifier to the spurious correlations in $\mathcal{S}$ . + +Complexity analysis. The text-based attribute detection only needs to use VLMs once to extract attributes for each test set of a dataset. For the task construction, in a nutshell, we analyze the attributes of samples from each of the $C$ classes and do the sampling. Thus, the computational complexity is $O(N_{\mathcal{T}}CN_cN_{\mathcal{A}})$ , where $N_{c}$ is the maximum number of samples per class in test data, $N_{\mathcal{A}}$ is the number of extracted attributes. We only need to run the process once and use the generated tasks to benchmark various models. + +# 5 Experiments + +# 5.1 Experimental Setup + +Datasets. We used two general datasets of different scales, miniImageNet [40] and tieredImageNet [41], and one fine-grained dataset, CUB-200 [56]. Each dataset consists of $\mathcal{D}_{train}$ , $\mathcal{D}_{val}$ , and $\mathcal{D}_{test}$ for training, validation, and test, respectively (see Appendix). All images were resized to $84 \times 84$ . + +FSC methods. We trained FSC models with ten algorithms covering three major categories. For gradient-based meta-learning algorithms, we chose ANIL [39], LEO [42], and BOIL [34]. For metric-based meta-learning algorithms, we chose ProtoNet [46], DN4 [24], R2D2 [5], CAN [16], and RENet [19]. For transfer learning algorithms, we chose Baseline++ [6] and RFS [50]. See Appendix for more details. Any backbones can be used as the feature extractor. For fair comparisons between different methods, we used the ResNet-12 backbone adopted in [35]. + +Text-based attribute detection. We used a pre-trained VLM named ViT-GPT2 [32] to generate text descriptions for images in $\mathcal{D}_{test}$ . After that, we used Spacy (https://spacy.io/) to extract nouns and adjectives from these descriptions automatically. We also used another pre-trained VLM, BLIP [23], to test + +Table 2: Statistics of detected attributes in ${\mathcal{D}}_{\text{test }}$ by two VLMs. + +
VLMUnique attributesAvg. attributes per class
miniImageNettieredImageNetCUB-200miniImageNettieredImageNetCUB-200
ViT-GPT211112532159190.40230.9425.78
BLIP20326710247254.40310.4029.74
+ +![](images/3439fbe77ecf1db06516627fe54b44d47f8e42407bec302e739ba779246afc67.jpg) +Fig. 3: A 5-way 1-shot task constructed by our inter-class attribute-based sample selection using samples from the miniImageNet dataset. Note that due to the limited capacity of a VLM, the attributes may not well align with human understandings. + +![](images/11d3cc987d5fc106a29b22db9362ebbb7fe3b46aaca79791304fbe4ec5f252de.jpg) + +whether FewSTAB can produce consistent results. The statistics of the detected attributes are shown in Table 2. + +Implementation details. We trained FSC models with the implementation in [25]. Each model was trained on $\mathcal{D}_{train}$ of a dataset with one of the ten FSC methods. For each meta-learning based method, we trained two models using randomly sampled 5-way 1-shot and 5-way 5-shot tasks, respectively. All the tasks have 15 samples per class in the query set. We saved the model that achieves the best validation accuracy on $\mathcal{D}_{val}$ for evaluation. For FewSTAB, if we do not have enough desired samples to construct a support set, we redo the construction from the beginning. If there are not enough desired samples to construct a query set, we first try to use the intra-class attribute-based sample selection; if the desired samples are still not enough, we redo the construction from the beginning. We created 3000 tasks for model evaluation. All experiments were conducted on the NVIDIA RTX 8000 GPUs. + +# 5.2 Visualization of a Constructed Task + +We show a 5-way 1-shot task constructed by FewSTAB in Fig. 3. Each class in the support set correlates with a unique spurious attribute. The query samples of a class do not have the spurious attribute correlated with the class and some of them have spurious attributes associated with other classes in the support set. For example, the query samples of the class lion do not have the spurious attribute water, and some of them have spurious attributes from other classes in + +Table 3: Comparison between wAcc-R and wAcc-A with $95\%$ confidence interval on the miniImageNet, tieredImageNet, and CUB datasets. Numbers in the Shot column indicate that the models are both trained (if applicable) and tested on 5-way 1- or 5-shot tasks. Darker colors indicate higher values. + +
ShotMethodminiImageNettieredImageNetCUB-200
wAcc-A (↑)wAcc-R (↑)wAcc-A (↑)wAcc-R (↑)wAcc-A (↑)wAcc-R (↑)
1ANIL10.38±0.3014.36±0.3311.21±0.3015.63±0.3613.78±0.4016.94±0.43
LEO14.26±0.4621.35±0.5416.00±0.5529.63±0.7128.29±0.8040.22±0.87
BOIL12.48±0.2314.93±0.2412.27±0.2114.13±0.2319.15±0.2922.50±0.29
ProtoNet14.03±0.4921.96±0.5814.50±0.5027.13±0.6934.62±0.8546.61±0.89
DN412.37±0.4619.28±0.5611.99±0.4723.62±0.6535.22±0.8647.26±0.88
R2D218.05±0.5326.50±0.6016.41±0.5430.41±0.7136.70±0.9048.82±0.88
CAN17.37±0.5325.96±0.6018.84±0.6036.29±0.7822.74±0.7231.95±0.78
RENet19.10±0.5728.85±0.6518.83±0.6135.70±0.7832.43±0.8143.98±0.81
Baseline++15.30±0.4823.18±0.5617.51±0.5431.62±0.719.17±0.4714.59±0.58
RFS18.00±0.5327.12±0.6118.35±0.6035.24±0.7732.45±0.8044.49±0.82
5ANIL14.83±0.4025.37±0.5213.72±0.3930.60±0.5131.63±0.5545.47±0.53
LEO26.31±0.5941.33±0.5929.49±0.7257.22±0.7246.62±0.8259.76±0.77
BOIL13.09±0.2215.21±0.2314.90±0.2218.55±0.2419.17±0.2821.33±0.27
ProtoNet32.07±0.5851.95±0.5230.95±0.7062.53±0.6260.06±0.7475.68±0.50
DN427.60±0.5842.68±0.6216.07±0.6240.63±0.8159.25±0.7773.58±0.56
R2D235.37±0.5950.84±0.5531.12±0.7261.08±0.6558.66±0.8275.20±0.54
CAN36.44±0.6554.23±0.5531.17±0.7664.19±0.6241.31±0.7461.61±0.58
RENet36.19±0.6356.52±0.5830.27±0.7663.49±0.6452.93±0.8271.82±0.56
Baseline++29.52±0.5744.94±0.5730.01±0.7259.06±0.6716.86±0.5229.84±0.69
RFS36.85±0.6455.66±0.5531.15±0.7662.71±0.6754.98±0.8174.33±0.53
+ +![](images/77a621976c9e3e83be3775b669456e11c5b87d0e7f3f61174569f043553fa85f.jpg) +Fig. 4: Accuracy gaps (wAcc-R minus wAcc-A) on the 5-way 1-shot and 5-way 5-shot tasks from the (a) miniImageNet, (b) tieredImageNet, and (c) CUB-200 datasets. + +![](images/ff8b411fada5e4286329789d29cc4a0ff250a1219ebec889c5253dfb42285f6f.jpg) + +![](images/3881f6eb4bb4aa3e8517b8b3565d61a8367db67a916c12552c64a931f8a738fc.jpg) + +the support set, such as man and green. FewSTAB introduces biased attributes in the task so that query samples can be easily misclassified as other classes by a few-shot classifier that relies on the spurious correlations in the support set. + +# 5.3 Effectiveness of FewSTAB + +FewSTAB can effectively reveal spurious bias in few-shot classifiers. We show in Table 3 the wAcc (Eq. 2) on 5-way 1/5-shot test tasks that are randomly sampled (wAcc-R) and are constructed with FewSTAB (wAcc-A), respectively. FewSTAB generates FSC test tasks only based on the class-attribute correlations in data. In each test setting, the FSC methods in Table 3 are evaluated with the same FSC tasks. We observe that wAcc-A is consistently lower than wAcc-R on the three datasets and on two test-shot numbers, showing that FewSTAB is more effective than the standard evaluation procedure (random task construction) in exhibiting the spurious bias in various few-shot classifiers. We + +additionally show that FewSTAB also works on the most recent FSC methods and can reflect the improvement made to mitigate spurious bias (see Appendix). + +FewSTAB reveals new robustness patterns among FSC methods. In Table 4, we calculate the Spearman's rank correlation coefficients [30] between the values of wAcc-R and wAcc-A from Table 3. The coefficients are bounded from 0 to 1, with larger values indicating that the ranks of FSC methods based on wAcc-R are more similar to those based on wAcc-A. In the 1-shot setting, it is not effective to control the spurious correlations since we only have one sample per class in the support set. Hence, the coefficients are large, and + +Table 4: Spearman's rank correlations between wAcc-A and wAcc-R in Table 3 + +
Dataset1-shot5-shot
miniImageNet0.960.95
tieredImageNet0.960.90
CUB-2001.000.94
+ +the ranks based on wAcc-A are similar to those based on wAcc-R. In the 5-shot cases, we have more samples to demonstrate the spurious correlations. The coefficients become smaller, i.e., the ranks based on wAcc-A show different trends from those based on wAcc-R. In this case, FewSTAB reveals new information on FSC methods' varied degrees of robustness to spurious bias. + +FewSTAB can benchmark spurious bias in varied degrees. As shown in Fig. 4, the accuracy gap, defined as wAcc-R minus wAcc-A, in general, becomes larger when we switch from 5-way 1-shot to 5-way 5-shot tasks. Compared with the random task construction, FewSTAB creates more challenging tasks in the 5-shot case for demonstrating spurious bias in few-shot classifiers. In other words, with a higher shot value in the constructed test tasks, FewSTAB aims to benchmark spurious bias in a higher degree. + +# 5.4 A New Dimension of Evaluation and a New Design Guideline + +FewSTAB creates a new dimension of evaluation on the robustness to spurious bias. We demonstrate this with a scatter plot (Fig. 5) of Acc (Eq. 11) and wAcc-A of the ten few-shot classifiers. FewSTAB offers new information regarding different few-shot classifiers' robustness to spurious bias as we observe that Acc does not well correlate with wAcc-A. A high wAcc-A indicates that the classifier is robust to spurious bias, while a high Acc indicates that the classifier can correctly predict most of the samples. With the scatter plot, we can view tradeoffs between the two metrics on existing few-shot classifiers. A desirable few-shot classifier should appear in the top-right corner of the plot. + +![](images/22bcc91d8e622a747deb3d0e84d3396166351e12be6c55259b2e70e05221e339.jpg) +Fig.5: Acc versus wAcc-A of the ten FSC methods tested on 5-way 5-shot tasks from miniImageNet. + +# 5.5 FewSTAB Enables Designs for Varied Degrees of Robustness + +As demonstrated in Section 5.3, Few-STAB can benchmark spurious bias in varied degrees, which in turn enables practitioners to design robust few-shot classifiers targeted for different degrees of robustness to spurious bias. The reason for differentiating designs for varied degrees of robustness is that the same design choice may not work under different robustness requirements. For example, increasing shot number in training tasks is a common strategy for improving the few-shot generalization of meta-learning based methods. We trained few-shot classifiers with 5-way 5-shot and 5-way 1-shot training tasks randomly sampled from $\mathcal{D}_{train}$ , respectively. We then calculated the accuracy gap defined as the wAcc-A of a model trained on 5-shot tasks minus the wAcc-A of the same model trained on 1-shot tasks. A positive and large accuracy gap indicates that this strat + +egy is effective in improving the model's robustness to spurious bias. In Fig. 6, on each of the three datasets, we give results of the eight meta-learning based FSC methods on the 5-way 1-, 5-, and 10-shot FewSTAB tasks which are used to demonstrate the strategy's robustness to increased degrees of spurious bias. This strategy does not work consistently under different test shots. For example, in Fig. 6(a) this strategy with CAN only works the best on the 5-way 5-shot FewSTAB tasks. + +# 5.6 Ablation Studies + +Techniques used in FewSTAB. We analyze how different sample selection methods affect the effectiveness of FewSTAB in Table 5. With only intra-class attribute-based sample selection, we randomly select query samples from Eq. ③. For inter-class attribute-based sample selection and intra-class attribute-based sample selection (automatically included by Eq. ④), we randomly select query samples from Eq. ⑤. FewSTAB uses all the techniques in Table 5. We define accuracy drop as wAcc-R minus wAcc-A, and we use the drop averaged over the ten FSC methods tested on 5-way 5-shot tasks from the miniImageNet dataset as our metric. A larger average drop indicates that the corresponding + +![](images/43d7254cf725f38714ac47e3765cc3b0384c00334b8338f89dbb13f0aae9c173.jpg) + +![](images/cdff529d27acec6e59d9fbb595167f898308ae8dc75e2e5ba3bb866824d8dd52.jpg) + +![](images/54d76d0f09429176594ef571243424c2bcd7ee74c8dba207014c1a8d0a032151.jpg) +Fig. 6: Accuracy gaps of few-shot classifiers tested on 1-shot, 5-shot, and 10-shot tasks constructed from (a) miniImageNet, (b) tieredImageNet, and (c) CUB-200 datasets. + +Table 5: Comparison between different techniques used by FewSTAB for constructing FSC tasks. + +
Attribute-based sample selectionQuery sample selectionAvg. drop (%)
Intra-classInter-class
✓5.13
✓✓13.30
✓✓✓15.05
+ +Table 6: Spearman's rank correlation coefficients between wAcc-A obtained using ViT-GPT2 and BLIP. + +
Dataset1-shot5-shot
miniImageNet0.981.00
tieredImageNet1.000.99
CUB-2001.000.98
+ +sample selection method is more effective in reflecting the spurious bias in few-shot classifiers. We observe that all proposed techniques are effective and the inter-class attribute-based sample selection is the most effective method. + +Choice of VLMs. Although our main results are based on the pre-trained ViT-GPT2 model [32], we show in Table 6 that when switching to a different VLM, i.e., BLIP [23], the relative ranks of different few-shot classifiers based on wAcc-A still hold with high correlations. In other words, FewSTAB is robust to different choices of VLMs. + +Detection accuracy of VLMs. A VLM may miss some attributes due to its limited capacity, resulting in a small detection accuracy. However, the detection accuracy of a VLM has little impact on our framework. To demonstrate this, we adopt a cross-validation strategy, i.e., we use the outputs from one VLM as the ground truth to evaluate those from another VLM, since assessing the detection accuracy of a VLM typically requires labor-intensive human labeling. On the CUB-200 dataset, we observe that the detection accuracy of ViT-GPT2 based on the BLIP's outputs is $70.12\%$ , while the detection accuracy of BLIP based on the ViT-GPT2's outputs is $59.28\%$ . Although the two VLMs differ significantly in the detected attributes, our framework shows almost consistent rankings of the evaluated FSC methods (Table 6). + +Additional results are presented in Appendix. + +# 6 Conclusion + +In this paper, we proposed a systematic and rigorous benchmark framework called FewSTAB for evaluating the robustness of few-shot classifiers to spurious bias. FewSTAB adopts attribute-based sample selection strategies to construct FSC test tasks with biased attributes so that the reliance on spurious correlations can be effectively revealed. FewSTAB can automatically benchmark spurious bias in few-shot classifiers on any existing test data thanks to its use of a pretrained VLM for automated attribute detection. With FewSTAB, we provided a new dimension of evaluation on the robustness of few-shot image classifiers to spurious bias and a new design guideline for building robust few-shot classifiers. FewSTAB can reveal and enable designs for varied degrees of robustness to spurious bias. We hope FewSTAB will inspire new developments on designing robust few-shot classifiers. + +# Acknowledgments + +This work is supported in part by the US National Science Foundation under grants 2313865, 2217071, 2213700, 2106913, 2008208, 1955151. + +# References + +1. Ahmed, F., Bengio, Y., Van Seijen, H., Courville, A.: Systematic generalisation with group invariant predictions. In: International Conference on Learning Representations (2020) +2. Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.: Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019) +3. Baker, N., Lu, H., Erlikhman, G., Kellman, P.J.: Deep convolutional networks do not classify based on global object shape. PLoS Computational Biology 14(12), e1006613 (2018) +4. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: European Conference on Computer Vision. pp. 456-473 (2018) +5. Bertinetto, L., Henriques, J.F., Torr, P., Vedaldi, A.: Meta-learning with differentiable closed-form solvers. In: International Conference on Learning Representations (2019) +6. Chen, W.Y., Liu, Y.C., Kira, Z., Wang, Y.C.F., Huang, J.B.: A closer look at few-shot classification. In: International Conference on Learning Representations (2019) +7. Chen, Y., Wang, X., Liu, Z., Xu, H., Darrell, T.: A new meta-baseline for few-shot learning. arXiv preprint arXiv:2003.04390 (2020) +8. Dong, J., Wang, Y., Lai, J.H., Xie, X.: Improving adversarially robust few-shot image classification with generalizable representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9025-9034 (2022) +9. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Proceedings of the 34th International Conference on Machine Learning. vol. 70, pp. 1126–1135 (2017) +0. Flennerhag, S., Rusu, A.A., Pascanu, R., Visin, F., Yin, H., Hadsell, R.: Meta-learning with warped gradient descent. In: International Conference on Learning Representations (2020) +1. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2019) +12. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2019) +13. Ghosal, S.S., Li, Y.: Are vision transformers robust to spurious correlations? International Journal of Computer Vision 132(3), 689-709 (2024) +14. Goldblum, M., Fowl, L., Goldstein, T.: Adversarily robust few-shot learning: A meta-learning approach. Advances in Neural Information Processing Systems 33, 17886-17895 (2020) + +15. He, Y., Shen, Z., Cui, P.: Towards non-iid image classification: A dataset and baselines. Pattern Recognition 110, 107383 (2021) +16. Hou, R., Chang, H., Bingpeng, M., Shan, S., Chen, X.: Cross attention network for few-shot classification. In: Advances in Neural Information Processing Systems. pp. 4003-4014 (2019) +17. Hu, S.X., Moreno, P.G., Xiao, Y., Shen, X., Obozinski, G., Lawrence, N., Dami-anou, A.: Empirical bayes transductive meta-learning with synthetic gradients. In: International Conference on Learning Representations (2020) +18. Jang, H., Lee, H., Shin, J.: Unsupervised meta-learning via few-shot pseudosupervised contrastive learning. In: International Conference on Learning Representations (2023) +19. Kang, D., Kwon, H., Min, J., Cho, M.: Relational embedding for few-shot classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8822-8833 (2021) +20. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) +21. Lee, K., Maji, S., Ravichandran, A., Soatto, S.: Meta-learning with differentiable convex optimization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10657-10665 (2019) +22. Lee, Y., Choi, S.: Gradient-based meta-learning with learned layerwise metric and subspace. In: International Conference on Machine Learning. pp. 2927-2936. PMLR (2018) +23. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +24. Li, W., Wang, L., Xu, J., Huo, J., Gao, Y., Luo, J.: Revisiting local descriptor based image-to-class measure for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7260-7268 (2019) +25. Li, W., Wang, Z., Yang, X., Dong, C., Tian, P., Qin, T., Jing, H., Shi, Y., Wang, L., Gao, Y., Luo, J.: Libfewshot: A comprehensive library for few-shot learning. IEEE Transactions on Pattern Analysis & Machine Intelligence (01), 1-18 (2023) +26. Li, Z., Zhou, F., Chen, F., Li, H.: Meta-SGD: Learning to learn quickly for few shot learning. ArXiv abs/1707.09835 (2017) +27. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3730-3738 (2015) +28. Lu, Y., Wen, L., Liu, J., Liu, Y., Tian, X.: Self-supervision can be a good few-shot learner. In: European Conference on Computer Vision. pp. 740-758. Springer (2022) +29. Motiian, S., Jones, Q., Iranmanesh, S., Doretto, G.: Few-shot adversarial domain adaptation. Advances in Neural Information Processing systems 30 (2017) +30. Myers, L., Sirois, M.J.: Spearman correlation coefficients, differences between. Encyclopedia of Statistical Sciences 12 (2004) +31. Neuhaus, Y., Augustin, M., Boreiko, V., Hein, M.: Spurious features everywhere large-scale detection of harmful spurious features in imagenet. arXiv preprint arXiv:2212.04871 (2022) +32. NLP Connect: vit-gpt2-image-captioning (revision 0e334c7) (2022). https://doi.org/10.57967/hf/0222, https://huggingface.co/nlpconnect/vit-gpt2-image-captioning + +33. Nushi, B., Kamar, E., Horvitz, E.: Towards accountable ai: Hybrid human-machine analyses for characterizing system failure. In: Proceedings of the AAAI Conference on Human Computation and Crowdsourcing. vol. 6, pp. 126-135 (2018) +34. Oh, J., Yoo, H., Kim, C., Yun, S.Y.: Boil: Towards representation change for few-shot learning. In: International Conference on Learning Representations (2020) +35. Oreshkin, B., Rodríguez López, P., Lacoste, A.: Tadam: Task dependent adaptive metric for improved few-shot learning. In: Advances in Neural Information Processing Systems 31, pp. 721-731 (2018) +36. Poulakakis-Daktylidis, S., Jamali-Rad, H.: Bectr: Batch enhanced contrastive few-shot learning. In: International Conference on Learning Representations (2024) +37. Qin, X., Song, X., Jiang, S.: Bi-level meta-learning for few-shot domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15900-15910 (2023) +38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) +39. Raghu, A., Raghu, M., Bengio, S., Vinyals, O.: Rapid learning or feature reuse? towards understanding the effectiveness of maml. In: International Conference on Learning Representations (2020) +40. Ravi, S., Larochelle, H.: Optimization as a model for few-shot learning. In: International Conference on Learning Representations (2017) +41. Ren, M., Ravi, S., Triantafillou, E., Snell, J., Swersky, K., Tenenbaum, J.B., Larochelle, H., Zemel, R.S.: Meta-learning for semi-supervised few-shot classification. In: International Conference on Learning Representations (2018) +42. Rusu, A.A., Rao, D., Sygnowski, J., Vinyals, O., Pascanu, R., Osindero, S., Hadsell, R.: Meta-learning with latent embedding optimization. In: International Conference on Learning Representations (2019) +43. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks. In: International Conference on Learning Representations (2020) +44. Singla, S., Feizi, S.: Salient imagenet: How to discover spurious features in deep learning? In: International Conference on Learning Representations (2022) +45. Singla, S., Nushi, B., Shah, S., Kamar, E., Horvitz, E.: Understanding failures of deep networks via robust feature extraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12853-12862 (2021) +46. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems 30, pp. 4077-4087 (2017) +47. Stock, P., Cisse, M.: Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases. In: European Conference on Computer Vision. pp. 498-512 (2018) +48. Sun, Q., Liu, Y., Chua, T.S., Schiele, B.: Meta-transfer learning for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 403-412 (2019) +49. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H.S., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1199-1208 (2018) +50. Tian, Y., Wang, Y., Krishnan, D., Tenenbaum, J.B., Isola, P.: Rethinking few-shot image classification: a good embedding is all you need? In: European Conference on Computer Vision. pp. 266-282. Springer (2020) + +51. Triantafillou, E., Zhu, T., Dumoulin, V., Lamblin, P., Evci, U., Xu, K., Goroshin, R., Gelada, C., Swersky, K., Manzagol, P.A., Larochelle, H.: Meta-dataset: A dataset of datasets for learning to learn from few examples. In: International Conference on Learning Representations (2020) +52. Tseng, H.Y., Lee, H.Y., Huang, J.B., Yang, M.H.: Cross-domain few-shot classification via learned feature-wise transformation. In: International Conference on Learning Representations (2020) +53. Vinyals, O., Blundell, C., Lillicrap, T., kavukcuoglu, k., Wierstra, D.: Matching networks for one shot learning. In: Advances in Neural Information Processing Systems 29, pp. 3630-3638 (2016) +54. Wang, Y., Girshick, R., Hebert, M., Hariharan, B.: Low-shot learning from imaginary data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7278-7286 (2018) +55. Wang, Y., Yao, Q., Kwok, J.T., Ni, L.M.: Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys 53(3), 1-34 (2020) +56. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-UCSD Birds 200. Tech. Rep. CNS-TR-2010-001, California Institute of Technology (2010) +57. Wong, E., Santurkar, S., Madry, A.: Leveraging sparse linear layers for debuggable deep networks. In: International Conference on Machine Learning. pp. 11205-11216. PMLR (2021) +58. Xiao, K.Y., Engstrom, L., Ilyas, A., Madry, A.: Noise or signal: The role of image backgrounds in object recognition. In: International Conference on Learning Representations (2021) +59. Xue, Y., Payani, A., Yang, Y., Mirzasoleiman, B.: Eliminating spurious correlations from pre-trained models via data mixing. arXiv preprint arXiv:2305.14521 (2023) +60. Yang, Y., Nushi, B., Palangi, H., Mirzasoleiman, B.: Mitigating spurious correlations in multi-modal models during fine-tuning. arXiv preprint arXiv:2304.03916 (2023) +61. Ye, H.J., Hu, H., Zhan, D.C., Sha, F.: Few-shot learning via embedding adaptation with set-to-set functions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8808-8817 (2020) +62. Yue, X., Zheng, Z., Das, H.P., Keutzer, K., Vincentelli, A.S.: Multi-source few-shot domain adaptation. arXiv preprint arXiv:2109.12391 (2021) +63. Yue, Z., Zhang, H., Sun, Q., Hua, X.S.: Interventional few-shot learning. Advances in Neural Information Processing Systems 33, 2734-2746 (2020) +64. Zhang, C., Cai, Y., Lin, G., Shen, C.: Deepemd: Differentiable earth mover's distance for few-shot learning (2020) +65. Zhang, C., Cai, Y., Lin, G., Shen, C.: Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) +66. Zhang, J., Wang, Y., Molino, P., Li, L., Ebert, D.S.: Manifold: A model-agnostic framework for interpretation and diagnosis of machine learning models. IEEE Transactions on Visualization and Computer Graphics 25(1), 364-373 (2018) +67. Zhang, M., Li, H., Wu, F., Kuang, K.: Metacoco: A new few-shot classification benchmark with spurious correlation. In: The Twelfth International Conference on Learning Representations (2024) +68. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023) \ No newline at end of file diff --git a/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/images.zip b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d49614567b51326f0b82c90f49691fd8a3fa25bc --- /dev/null +++ b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dfbf8cf30eb274a56a40333935d2e641526eeb725aa0903191c99cf0c0893d3 +size 470732 diff --git a/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/layout.json b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5f8e134f55035d92fc6df7c06fc2cc146c514f6e --- /dev/null +++ b/2024/Benchmarking Spurious Bias in Few-Shot Image Classifiers/layout.json @@ -0,0 +1,12093 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 111, + 477, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 111, + 477, + 146 + ], + "spans": [ + { + "bbox": [ + 139, + 111, + 477, + 146 + ], + "type": "text", + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 167, + 436, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 167, + 436, + 182 + ], + "spans": [ + { + "bbox": [ + 179, + 167, + 436, + 182 + ], + "type": "text", + "content": "Guangtao Zheng1, Wenqian Ye1, and Aidong Zhang1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 190, + 419, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 190, + 419, + 213 + ], + "spans": [ + { + "bbox": [ + 194, + 190, + 419, + 213 + ], + "type": "text", + "content": "1University of Virginia, Charlottesville VA 22904, USA {gz5hp,wenqian,aidong}@virginia.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 246, + 455, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 246, + 455, + 499 + ], + "spans": [ + { + "bbox": [ + 160, + 246, + 455, + 499 + ], + "type": "text", + "content": "Abstract. Few-shot image classifiers are designed to recognize and classify new data with minimal supervision and limited data but often show reliance on spurious correlations between classes and spurious attributes, known as spurious bias. Spurious correlations commonly hold in certain samples and few-shot classifiers can suffer from spurious bias induced from them. There is an absence of an automatic benchmarking system to assess the robustness of few-shot classifiers against spurious bias. In this paper, we propose a systematic and rigorous benchmark framework, termed FewSTAB, to fairly demonstrate and quantify varied degrees of robustness of few-shot classifiers to spurious bias. Few-STAB creates few-shot evaluation tasks with biased attributes so that using them for predictions can demonstrate poor performance. To construct these tasks, we propose attribute-based sample selection strategies based on a pre-trained vision-language model, eliminating the need for manual dataset curation. This allows FewSTAB to automatically benchmark spurious bias using any existing test data. FewSTAB offers evaluation results in a new dimension along with a new design guideline for building robust classifiers. Moreover, it can benchmark spurious bias in varied degrees and enable designs for varied degrees of robustness. Its effectiveness is demonstrated through experiments on ten few-shot learning methods across three datasets. We hope our framework can inspire new designs of robust few-shot classifiers. Our code is available at https://github.com/gtzheng/FewSTAB" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "spans": [ + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "text", + "content": "Keywords: Few-shot classification " + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "text", + "content": " Spurious bias " + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "text", + "content": " Robustness " + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 510, + 453, + 532 + ], + "type": "text", + "content": " Benchmark system" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "content": "Few-shot classification [5,21,46,61,64] (FSC) has attracted great attention recently due to its promise for recognizing novel classes efficiently with limited data. Few-shot classifiers can transfer the knowledge learned from base classes to recognize novel classes with a few labeled samples. However, they face potential risks when deployed in the real world, such as data distribution shifts [29,62] and adversarial examples [8,14]. A subtle yet critical risk factor is the spurious correlations [3,4,11,13,43,47,59] between classes and spurious attributes —" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 305, + 178 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 305, + 178 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 305, + 178 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 305, + 178 + ], + "type": "image", + "image_path": "9615ac1307e59941c988adebb0cadcaf3289cd17d9f320a16aedf8d536e05498.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 180, + 259, + 188 + ], + "lines": [ + { + "bbox": [ + 212, + 180, + 259, + 188 + ], + "spans": [ + { + "bbox": [ + 212, + 180, + 259, + 188 + ], + "type": "text", + "content": "Learned path" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 482, + 178 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 482, + 178 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 482, + 178 + ], + "type": "image", + "image_path": "e57ac34f47a9e4f529ee4d44feb0c29dae1b31546867bfbab0f09309258e8b57.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 180, + 411, + 188 + ], + "lines": [ + { + "bbox": [ + 351, + 180, + 411, + 188 + ], + "spans": [ + { + "bbox": [ + 351, + 180, + 411, + 188 + ], + "type": "text", + "content": "Non-existing path" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 197, + 482, + 231 + ], + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 231 + ], + "type": "text", + "content": "Fig. 1: Exploiting the spurious correlation between the class bird and the spurious attribute tree branch to predict bird leads to an incorrect prediction on the test image showing birds on a grass field. For clarity, we only show the case for one class." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 256, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 399 + ], + "type": "text", + "content": "attributes of inputs non-essential to the classes. In the traditional learning setting [1,2,43], deep learning models tend to rely on spurious correlations as their prediction shortcuts or exhibit spurious bias, such as predicting classes using the associated backgrounds [58] or image textures [12], leading to significant performance drops when the associated backgrounds or textures change to different ones. In the low-data regime, spurious bias becomes more evident. For example, in Fig. 1, the correlation between the class bird and the attribute tree branch in the support (training) image may form a shortcut path from tree branch to predicting the image as bird and hinder the learning of the desired one that uses class-related attributes, such as head, tail, and wing. The shortcut will fail to generalize in the query (test) image where no tree branch can be found. In general, few-shot image classifiers are susceptible to spurious bias." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 401, + 483, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 483, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 483, + 520 + ], + "type": "text", + "content": "However, there lacks a dedicated benchmarking framework that evaluates the robustness of few-shot classifiers to spurious bias. The standard benchmarking procedure in FSC trains a few-shot classifier on base classes from a training set with ample samples and evaluates the classifier on FSC test tasks constructed from a test set with novel classes. The problem with this procedure is the lack of explicit control over the spurious correlations in the constructed FSC tasks. Each FSC test task contains randomly sampled support and query samples. Thus, spurious correlations in the majority of the test set samples can be demonstrated in these tasks, providing unfair advantages for few-shot classifiers with high reliance on the spurious correlations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a systematic and rigorous benchmark framework, termed Few-Shot Tasks with Attribute Biases (FewSTAB), to fairly compare the robustness of various few-shot classifiers to spurious bias. Our framework explicitly controls spurious correlations in the support and query samples when constructing an FSC test task to reveal the robustness pitfalls caused by spurious bias. To achieve this, we propose attribute-based sample selection strategies that select support and query samples with biased attributes. These attributes together with their associated classes formulate spurious correlations such that if the support samples induce spurious bias in a few-shot classifier, i.e., the classifier learns the spurious correlations in the support samples as its prediction shortcuts, then the query samples can effectively degrade the classifier's performance, exposing its non-robustness to spurious bias." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": "Our framework exploits the spurious attributes in test data for formulating spurious correlations in FSC test tasks. Some existing datasets [15,27,43] provide spurious attribute annotations. However, they only have a few classes and cannot provide enough classes for training and testing. Many benchmark datasets for FSC do not have annotations on spurious attributes, and obtaining these annotations typically involves labor-intensive human-guided labeling [33,66]. To address this, we further propose to use a pre-trained vision-language model (VLM) to automatically identify distinct attributes in images in the high-level text format. Our attribute-based sampling methods can use the identified attributes to simulate various spurious correlations. Thus, we can reuse any existing FSC datasets for benchmarking few-shot classifiers' robustness to spurious bias, eliminating the need for the manual curation of new datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 260, + 426, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 260, + 426, + 270 + ], + "spans": [ + { + "bbox": [ + 147, + 260, + 426, + 270 + ], + "type": "text", + "content": "The main contributions of our work are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 277, + 481, + 419 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 277, + 481, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 324 + ], + "type": "text", + "content": "- We propose a systematic and rigorous benchmark framework, termed Few-Shot Tasks with Attribute Biases (FewSTAB), that specifically targets spurious bias in few-shot classifiers, demonstrates their varied degrees of robustness to spurious bias, and benchmarks spurious bias in varied degrees." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 326, + 481, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 371 + ], + "type": "text", + "content": "- We propose novel attribute-based sample selection strategies using a pretrained VLM for constructing few-shot evaluation tasks, allowing us to reuse any existing few-shot benchmark datasets without manually curating new ones for the evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 373, + 481, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 419 + ], + "type": "text", + "content": "- FewSTAB provides a new dimension of evaluation on the robustness to spurious bias along with a new design guideline for building robust few-shot classifiers. We demonstrate the effectiveness of FewSTAB by applying it to models trained on three benchmark datasets with ten FSC methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 437, + 237, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 237, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 237, + 450 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 641 + ], + "type": "text", + "content": "Few-shot classification. Few-shot classification [6,46,50,51,53,55] has received vast attention recently. Few-shot classifiers can be trained with meta-learning or transfer learning on base classes to learn the knowledge that can be transferred to recognize novel classes with a few labeled samples. The transfer learning approaches [6,50] first learn a good embedding model and then fine-tune the model on samples from novel classes. The meta-learning approaches can be further divided into optimization-based and metric-based methods. The optimization-based methods [9,10,22,26,54] aim to learn a good initialized model such that the model can adapt to novel classes efficiently with a few gradient update steps on a few labeled samples. The metric-based methods [5,21,35,42,49,53,61,65] aim to learn a generalizable representation space with a well-defined metric, such as Euclidean distance [46], to learn novel classes with a few labeled samples. Recently, large vision-language models [20,38,68] are used for few-shot classification. However, they have completely different training and inference pipelines from the models that we consider in this paper." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Robustness in few-shot classification. There are several notions of robustness for few-shot classifiers. The common one requires a few-shot classifier to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 448, + 103 + ], + "type": "text", + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "perform well on the in-distribution samples of novel classes in randomly sampled FSC test tasks. The robustness to adversarial perturbations further requires a few-shot classifier to perform well on samples with imperceptible perturbations [8,14]. Moreover, the cross-domain generalization [37,51,52] aims to test how robust a few-shot classifier is on samples from novel classes with domain shifts, which are typically reflected by the changes in both image styles and classes. In contrast, we focus on a new notion of robustness: the robustness to spurious bias. There is a lack of rigorous evaluation methods on the topic. We provide a new evaluation method that specifically targets spurious bias and can systematically demonstrate few-shot classifiers' varied degrees of vulnerability to spurious bias, which has not been addressed in the existing literature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "type": "text", + "content": "Benchmarks for spurious bias. There are some existing datasets [15,27,43] that are designed to benchmark spurious bias in image classifiers. However, these datasets are only applicable to the traditional learning setting [1,2,43] since the classes in them are not sufficient for the training and testing of few-shot classifiers. Existing benchmarks in few-shot classification are not tailored for benchmarking spurious bias in few-shot classifiers. A recent work [67] creates a large-scale few-shot classification benchmark dataset with spurious-correlation shifts. In contrast, we propose a benchmark framework that can reuse existing few-shot classification datasets and provide a new dimension of evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "type": "text", + "content": "Discovering spurious attributes. A spurious attribute is non-essential to a class and only exists in some samples. Early works on discovering spurious attributes [33,66] require a predefined list of spurious attributes and expensive human-guided labeling of visual attributes. Recent works [31,44,45,57] greatly reduce the need for manual annotations by using the neurons of robust models to detect visual attributes. However, they still need humans to annotate the detected visual attributes. We automate this process by using a pre-trained VLM to obtain distinct attributes as words. Instead of discovering spurious correlations, we simulate them via attribute-based sampling for benchmarking." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 483, + 226, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 226, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 226, + 497 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": "Few-shot classification tasks. A typical FSC task " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": " has a support set " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": " for training and a query set " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": " for testing. In this task, there are " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": " classes " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "(c = 1,\\dots ,C)" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{S}}" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": " (a small number) training samples and " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": " test samples per class in " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": ", respectively. The task is called a " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{S}}" + }, + { + "bbox": [ + 130, + 509, + 482, + 557 + ], + "type": "text", + "content": "-shot task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": "Few-shot classifiers. A few-shot classifier " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " aims to classify the samples in " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " after learning from " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " with a learning algorithm " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " in a few-shot task " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " could be any learning algorithms, such as the optimization method [9] or a prototype-based classifier learning method [7,46]. To acquire a good few-shot learning capability, " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " is typically meta-trained or pre-trained [25] on a base training set " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train} = \\{(x_n, y_n) | y_n \\in \\mathcal{C}_{train}, n = 1, \\dots, N_{train}\\}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "N_{train}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " sample(x)-label(y) pairs, where " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{train}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " is a set of base classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Performance metrics. The performance of a few-shot classifier is typically measured by its average classification accuracy over " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "N_{S}" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "-shot tasks" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "G. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 208, + 136, + 402, + 311 + ], + "blocks": [ + { + "bbox": [ + 191, + 114, + 422, + 126 + ], + "lines": [ + { + "bbox": [ + 191, + 114, + 422, + 126 + ], + "spans": [ + { + "bbox": [ + 191, + 114, + 422, + 126 + ], + "type": "text", + "content": "Table 1: Meanings of major symbols used in the paper." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 208, + 136, + 402, + 311 + ], + "lines": [ + { + "bbox": [ + 208, + 136, + 402, + 311 + ], + "spans": [ + { + "bbox": [ + 208, + 136, + 402, + 311 + ], + "type": "table", + "html": "
SymbolMeaning
TAn FSC task
SSupport (training) set in T
QQuery (test) set in T
cA class in T
CNumber of classes per task
NsNumber of samples (shots) per class in S
NQNumber of samples per class in Q
OA few-shot adaptation algorithm
ψAn attribute detector
ΩAn automatic word selector
DtrainThe base training set
DvalThe validation set for selecting a few-shot classifier
DtestThe test set for evaluating a few-shot classifier
CtrainClasses in Dtrain
CtestClasses in Dtest
DcA set of all samples belonging to class c
aA text-format attribute
AA set of text-format attributes
", + "image_path": "ec743651d9cbdab2e1284fae6e5f72e5f74a8e0e2ea35aa4a6e472a36053620e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "text", + "content": "randomly sampled from " + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test} = \\{(x_n, y_n) | y_n \\in \\mathcal{C}_{test}, n = 1, \\dots, N_{test}\\}" + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "inline_equation", + "content": "N_{test}" + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "text", + "content": " sample-label pairs from novel classes " + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{test}" + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "text", + "content": " do not appear in " + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{train} \\cap \\mathcal{C}_{test} = \\emptyset" + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "text", + "content": ". We denote this metric as standard accuracy " + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}(f_\\theta)" + }, + { + "bbox": [ + 130, + 333, + 482, + 370 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 226, + 378, + 481, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 378, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 226, + 378, + 481, + 411 + ], + "type": "interline_equation", + "content": "\\operatorname {A c c} \\left(f _ {\\theta}\\right) = \\frac {1}{N _ {\\mathcal {T}}} \\sum_ {t = 1} ^ {N _ {\\mathcal {T}}} \\sum_ {c = 1} ^ {C} M _ {c} \\left(\\mathcal {T} _ {t}; f _ {\\theta}, \\mathcal {O}\\right), \\tag {1}", + "image_path": "67324d212d17a2821209acded5532ba3c784b888311f274a3768903fde7f1806.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "M_{c}(\\mathcal{T}_{t};f_{\\theta},\\mathcal{O})" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " denotes the classification accuracy of " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " on the query samples from the class " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " is trained on " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": ". The metric " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " in Eq. (11) only shows the average learning capability of " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " randomly selected novel classes. To better characterize the robustness of " + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": " to spurious bias, we define the class-wise worst classification accuracy over tasks as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 488, + 481, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 488, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 214, + 488, + 481, + 520 + ], + "type": "interline_equation", + "content": "\\operatorname {w A c c} \\left(f _ {\\theta}\\right) = \\frac {1}{N _ {\\mathcal {T}}} \\sum_ {t = 1} ^ {N _ {\\mathcal {T}}} \\min _ {c = 1, \\dots , C} M _ {c} \\left(\\mathcal {T} _ {t}; f _ {\\theta}, \\mathcal {O}\\right). \\tag {2}", + "image_path": "cccd3361e15fdf1631be9f78631aa9f59743e2d615f4d2b3b81120f70164087c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 529, + 430, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 430, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 430, + 541 + ], + "type": "text", + "content": "A larger " + }, + { + "bbox": [ + 130, + 529, + 430, + 541 + ], + "type": "inline_equation", + "content": "\\mathrm{wAcc}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 529, + 430, + 541 + ], + "type": "text", + "content": " indicates that " + }, + { + "bbox": [ + 130, + 529, + 430, + 541 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 529, + 430, + 541 + ], + "type": "text", + "content": " is more robust to spurious bias." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 542, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 578 + ], + "type": "text", + "content": "Spurious correlations. A spurious correlation is the association between a class and an attribute of inputs that is non-essential to the class, and it only holds in some samples. We formally define it as follows." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Definition 1. Let " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " denote a set of sample-label pairs having the label " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\psi : \\mathcal{X} \\to \\mathcal{B}_A" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " be an attribute detector, where " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " is the set of all possible inputs, " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_A" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " denotes all possible subsets of " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " is the set of all possible attributes. The class " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " and an attribute " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "a \\in \\mathcal{A}" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " form a spurious correlation, denoted as " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": ", if and only if the following conditions hold:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 653, + 378, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 653, + 378, + 666 + ], + "spans": [ + { + "bbox": [ + 136, + 653, + 378, + 666 + ], + "type": "text", + "content": "1. There exists " + }, + { + "bbox": [ + 136, + 653, + 378, + 666 + ], + "type": "inline_equation", + "content": "(x,c)\\in \\mathcal{D}_c" + }, + { + "bbox": [ + 136, + 653, + 378, + 666 + ], + "type": "text", + "content": " that satisfies " + }, + { + "bbox": [ + 136, + 653, + 378, + 666 + ], + "type": "inline_equation", + "content": "a\\in \\psi (x)" + }, + { + "bbox": [ + 136, + 653, + 378, + 666 + ], + "type": "text", + "content": " , and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "text", + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 481, + 266 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 266 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 266 + ], + "type": "image", + "image_path": "3ec32a50922a691a3cb03923c1de7f2bb667cfe8e4bd69072773176976fd3941.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 272, + 480, + 295 + ], + "lines": [ + { + "bbox": [ + 131, + 272, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 131, + 272, + 480, + 295 + ], + "type": "text", + "content": "Fig. 2: FewSTAB overview. (a) Extract distinct attributes using a pre-trained VLM. (b) Generate an FSC task for the evaluation of spurious bias in few-shot classifiers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 318, + 361, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 318, + 361, + 331 + ], + "spans": [ + { + "bbox": [ + 135, + 318, + 361, + 331 + ], + "type": "text", + "content": "2. There exists " + }, + { + "bbox": [ + 135, + 318, + 361, + 331 + ], + "type": "inline_equation", + "content": "(x', c) \\in \\mathcal{D}_c" + }, + { + "bbox": [ + 135, + 318, + 361, + 331 + ], + "type": "text", + "content": " that satisfies " + }, + { + "bbox": [ + 135, + 318, + 361, + 331 + ], + "type": "inline_equation", + "content": "a \\notin \\psi(x')" + }, + { + "bbox": [ + 135, + 318, + 361, + 331 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 337, + 349, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 337, + 349, + 350 + ], + "spans": [ + { + "bbox": [ + 133, + 337, + 349, + 350 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 133, + 337, + 349, + 350 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 133, + 337, + 349, + 350 + ], + "type": "text", + "content": " as the spurious attribute in " + }, + { + "bbox": [ + 133, + 337, + 349, + 350 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 133, + 337, + 349, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "text", + "content": "Definition " + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "text", + "content": " specifies that all the spurious correlations are based on " + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "text", + "content": ". In the remainder of the paper, we define " + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c = \\{(x,c)|\\forall (x,c)\\in \\mathcal{D}_{test}\\}" + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "inline_equation", + "content": "c\\in \\mathcal{C}_{test}" + }, + { + "bbox": [ + 130, + 357, + 480, + 392 + ], + "type": "text", + "content": " as we focus on evaluating the robustness to spurious bias." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 393, + 460, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 393, + 460, + 407 + ], + "spans": [ + { + "bbox": [ + 146, + 393, + 460, + 407 + ], + "type": "text", + "content": "We list major symbols in the paper alongside their meanings in Tab. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 423, + 232, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 232, + 437 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 232, + 437 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 448, + 331, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 331, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 331, + 460 + ], + "type": "text", + "content": "4.1 Attribute-Based Sample Selection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "content": "We first propose two attribute-based sample selection methods to reveal spurious bias in a few-shot classifier. Consider a training set " + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "content": " in a few-shot test task " + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "content": ", which has " + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "content": " classes with each class " + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "inline_equation", + "content": "c\\in \\mathcal{C}_{test}" + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "content": " associating with a unique spurious attribute " + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "inline_equation", + "content": "a\\in \\mathcal{A}" + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "content": ". We aim to discover samples that can exhibit a classifier's spurious bias on " + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "inline_equation", + "content": "\\langle c,a\\rangle" + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "content": " induced from " + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "content": ". Motivated by existing findings [43,59,60] that classifiers with high reliance on " + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "inline_equation", + "content": "\\langle c,a\\rangle" + }, + { + "bbox": [ + 130, + 468, + 482, + 551 + ], + "type": "text", + "content": " tend to perform poorly on samples without it, we propose an attribute-based sample selection strategy below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": "Intra-class attribute-based sample selection. Given " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": " and the training set " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": " having the spurious correlation " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": ", we generate a set " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\langle c, a \\rangle}" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": " of sample-label pairs which have class " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": " but do not contain attribute " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 219, + 596, + 481, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 596, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 219, + 596, + 481, + 610 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\langle c, a \\rangle} = \\{(x, c) | \\forall (x, c) \\in \\mathcal {D} _ {c}, a \\notin \\psi (x) \\}. \\tag {3}", + "image_path": "e78b37e076d3d108d45ea4fd4125b2ed0df91568ccb9fbc933d39ce2c1fd51c8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The above proposed method demonstrates a few-shot classifier's robustness to individual spurious correlation " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and does not consider a multi-class classification setting where spurious attributes from some other class " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " exist in samples of the class " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". In this case, these attributes may mislead the classifier to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "predict those samples as the class " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " and severely degrade the performance on the class " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": ". For example, consider using the spurious correlations " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\langle \\text{vase}, \\text{blue} \\rangle" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\langle \\text{bowl}, \\text{green} \\rangle" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " for predicting vase and bowl, respectively. An image showing a vase in green is more effective in revealing the robustness to " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\langle \\text{vase}, \\text{blue} \\rangle" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " as it is more likely to be misclassified as bowl than other images. Motivated by this, we propose the inter-class attribute-based sample selection below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": "Inter-class attribute-based sample selection. Given " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": " and the training set " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": " having the spurious correlations " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\langle c', a' \\rangle" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "c' \\neq c" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "a' \\neq a" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": ", we generate a set " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\langle c, a \\rangle}^{\\langle c', a' \\rangle}" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": " of sample-label pairs which have class " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": ", do not contain attribute " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": ", but contain attribute " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "a'" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": " from another class " + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 188, + 479, + 239 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 249, + 480, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 249, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 196, + 249, + 480, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\langle c, a \\rangle} ^ {\\langle c ^ {\\prime}, a ^ {\\prime} \\rangle} = \\mathcal {I} _ {\\langle c, a \\rangle} \\cap \\left\\{\\left(x, c\\right) | \\forall (x, c) \\in \\mathcal {D} _ {c}, a _ {\\langle c ^ {\\prime} \\rangle} ^ {\\prime} \\in \\psi (x) \\right\\}, \\tag {4}", + "image_path": "7792d75f6278433b36f1f1bfd2e28897cdb14104cd0a8dbcf9f908d20fe34dc5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "inline_equation", + "content": "a_{\\langle c^{\\prime}\\rangle}^{\\prime}" + }, + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "inline_equation", + "content": "a^\\prime" + }, + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "inline_equation", + "content": "\\langle c',a'\\rangle" + }, + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\langle c,a\\rangle}" + }, + { + "bbox": [ + 130, + 275, + 412, + 289 + ], + "type": "text", + "content": " is defined in Eq. (3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": "Considering that there are " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": " classes in the training set " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": " with each class associating with a unique spurious attribute " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": ", to effectively demonstrate the reliance on the spurious correlation " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": " with the inter-class attribute-based sample selection, we consider all the spurious correlations in " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": ". Specifically, we apply the above selection strategy to all the " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "C - 1" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": " spurious correlations in " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": " other than " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\langle c, a \\rangle}^{C}" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": " as the union of the " + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "inline_equation", + "content": "C - 1" + }, + { + "bbox": [ + 130, + 289, + 480, + 362 + ], + "type": "text", + "content": " sets as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 372, + 480, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 372, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 252, + 372, + 480, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\langle c, a \\rangle} ^ {C} = \\bigcup_ {\\langle c ^ {\\prime}, a ^ {\\prime} \\rangle \\in \\mathcal {C} \\backslash c} \\mathcal {I} _ {\\langle c, a \\rangle} ^ {\\langle c ^ {\\prime}, a ^ {\\prime} \\rangle}, \\tag {5}", + "image_path": "150210f39698daf2c38f0cdd23ce82b7a77080f48b2780c93f63f606277016b5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 410, + 436, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 436, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 436, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 410, + 436, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\backslash c}" + }, + { + "bbox": [ + 130, + 410, + 436, + 422 + ], + "type": "text", + "content": " denotes all the spurious correlations in " + }, + { + "bbox": [ + 130, + 410, + 436, + 422 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 410, + 436, + 422 + ], + "type": "text", + "content": " other than " + }, + { + "bbox": [ + 130, + 410, + 436, + 422 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 130, + 410, + 436, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 422, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 493 + ], + "type": "text", + "content": "The inter-class attribute-based sample selection is built upon the intra-class attribute-based sample selection. In the remainder of the paper, we use the inter-class method as our default sample selection strategy, which is more effective empirically (Sec. 5.6). In certain cases, however, where there are not enough desired samples during task construction, we resort to the intra-class sample selection strategy (Sec. 5.1, Implementation details)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "content": "In the following, we introduce FewSTAB, a benchmark framework that uses the proposed selection strategies to construct FSC tasks containing samples with biased attributes for benchmarking spurious bias in few-shot classifiers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 548, + 422, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 422, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 422, + 560 + ], + "type": "text", + "content": "4.2 FewSTAB (Part 1): Text-Based Attribute Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 628 + ], + "type": "text", + "content": "Our attribute-based sample selection methods require knowing the attributes in images, which typically involves labor-intensive human labeling. To make our method scalable and applicable to few-shot classifiers trained on different datasets, we adopt a pre-trained VLM to automatically identify distinct attributes in images in text format, which includes the following two steps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Step 1: Generating text descriptions. We use a pre-trained VLM [23, 32] " + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": " to automatically generate text descriptions for images in " + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": ". The VLM is a model in the general domain and can produce text descriptions for various" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 448, + 103 + ], + "type": "text", + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "objects and patterns. For example, for the current input image in the vase class in Fig. 2(a), besides the class object vase, the VLM also detects the vase's color green, and another object table with its material wooden." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": "Step 2: Extracting informative words. From the generated text descriptions, we extract nouns and adjectives as the detected attributes via an automatic procedure " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": ". The two kinds of words are informative as a noun describes an object, and an adjective describes a property of an object. All the detected attributes form the candidate attribute set " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": ". We realize the attribute detector " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " defined in Definition 1 as " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\psi(x) = \\Omega(\\phi(x))" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 480, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 480, + 308 + ], + "type": "text", + "content": "Remark 1: A VLM in general can extract many distinct attributes from the images. On some images, the VLM may detect non-relevant attributes, such as detecting a duck from a bird image. A more capable VLM could warrant a better attribute detection accuracy and benefit individual measurements on few-shot classifiers. Although being a VLM-dependent benchmark framework, FewSTAB can produce consistent and robust relative measurements among all the compared FSC methods, regardless the choice of VLMs (Sec. 5.6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "text", + "content": "Remark 2: The candidate set " + }, + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "text", + "content": " constructed with all the extracted words may contain attributes that represent the classes in " + }, + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "text", + "content": ". However, during our attribute-based sample selection, these attributes will not be used since they always correlate with classes and therefore do not satisfy the definition of spurious attributes in Definition 1. We provide details of " + }, + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 309, + 480, + 370 + ], + "type": "text", + "content": " in Sec. 5.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 389, + 380, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 389, + 380, + 402 + ], + "spans": [ + { + "bbox": [ + 131, + 389, + 380, + 402 + ], + "type": "text", + "content": "4.3 FewSTAB (Part 2): FSC Task Construction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "text", + "content": "Constructing a " + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "inline_equation", + "content": "N_{S}" + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "text", + "content": "-shot FSC task " + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "text", + "content": " for benchmarking spurious bias in few-shot classifiers involves constructing a support (training) set " + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "text", + "content": " and a query (test) " + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 411, + 479, + 446 + ], + "type": "text", + "content": " with biased attributes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": "Constructing the support set. The support set contains the spurious correlations that we aim to demonstrate to a few-shot classifier. As a fair and rigorous benchmark system, FewSTAB makes no assumptions on the few-shot classifiers being tested and randomly samples " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " classes from " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{test}" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": ". For each sampled class, it randomly selects a spurious correlation " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "a \\in \\mathcal{A}" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": ". To effectively demonstrate the spurious correlation " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "\\langle c, a \\rangle" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " to a few-shot classifier, we select samples of the class " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " such that (1) they all have the spurious attribute " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " and (2) do not have spurious attributes from the other " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "C - 1" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " spurious correlations. We construct " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "N_{S}" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " samples for the class " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " that satisfy the above two conditions. Thus, the spurious attribute " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " becomes predictive of the class " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": ". We take the union of all " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " such sets to get " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "S = \\cup_{c=1}^{C} S_{c}" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": ". Fig. 2(b) demonstrates the case when " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "C = 3" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": ". Note that we have no requirements for other non-selected attributes in " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": " to ensure that we have enough samples for " + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 130, + 448, + 480, + 603 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "text", + "content": "Constructing the query set. To evaluate the robustness to the spurious correlations formulated in " + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "text", + "content": ", we first construct a candidate set " + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\langle c,a\\rangle}^{C}" + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "text", + "content": " in Eq. (5) for each spurious correlation " + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\langle c,a\\rangle" + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "text", + "content": ". Since we have no requirements on the non-selected attributes that are not used to formulate spurious correlations in " + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 604, + 480, + 665 + ], + "type": "text", + "content": ", a few-shot classifier may predict query samples via some of these attributes," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "G. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "e.g., the yellow blocks in Fig. 2(b), bypassing the test on the formulated spurious correlations in " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": ". To address this, we propose query sample selection below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 190 + ], + "type": "text", + "content": "Query sample selection: We select query samples from " + }, + { + "bbox": [ + 130, + 140, + 479, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\langle c,a\\rangle}^{C}" + }, + { + "bbox": [ + 130, + 140, + 479, + 190 + ], + "type": "text", + "content": " that are least likely to have non-selected spurious attributes, such as the ones enclosed with red boxes in Fig. 2(b). To achieve this, we first calculate the fraction of sample-label pairs in " + }, + { + "bbox": [ + 130, + 140, + 479, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\langle c,a\\rangle}^{C}" + }, + { + "bbox": [ + 130, + 140, + 479, + 190 + ], + "type": "text", + "content": " that have the attribute " + }, + { + "bbox": [ + 130, + 140, + 479, + 190 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 140, + 479, + 190 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 198, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 198, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 212, + 198, + 481, + 213 + ], + "type": "interline_equation", + "content": "p _ {a} = | \\{x \\mid a \\in \\psi (x), \\forall (x, c) \\in \\mathcal {I} _ {\\langle c, a \\rangle} ^ {C} \\} | / | \\mathcal {I} _ {\\langle c, a \\rangle} ^ {C} |, \\tag {6}", + "image_path": "10f69e6c87f9f2a44d3cf8a54199b6cd802a79c8cb6f633b8d9297a4b41b196c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " denotes the size of a set, " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "a \\in \\tilde{\\mathcal{A}}" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{A}}" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " contains all non-selected attributes. A larger " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "p_a" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " indicates that the attribute " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " occurs more frequently in data and is more likely to be used in formulating prediction shortcuts. We then calculate the likelihood score for each " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "(x,c) \\in \\mathcal{I}_{\\langle c,a\\rangle}^C" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "s(x) = \\sum_{a \\in \\psi(x), a \\in \\tilde{\\mathcal{A}}} p_a" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": ", i.e., the summation of all " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "p_a" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " of non-selected attributes in " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": ". The likelihood score will be zero if there are no non-selected attributes in " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": ". A large " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "s(x)" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " indicates that the image " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " can be predicted via many non-selected attributes. Therefore, we select " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " samples from " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\langle c,a\\rangle}^C" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": " that have the lowest likelihood scores to construct " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_c" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": ". Then, we have " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\cup_{c=1}^{C} \\mathcal{Q}_c" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": ", which contains samples for evaluating the robustness of a few-shot classifier to the spurious correlations in " + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 220, + 480, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "text", + "content": "Complexity analysis. The text-based attribute detection only needs to use VLMs once to extract attributes for each test set of a dataset. For the task construction, in a nutshell, we analyze the attributes of samples from each of the " + }, + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "text", + "content": " classes and do the sampling. Thus, the computational complexity is " + }, + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "inline_equation", + "content": "O(N_{\\mathcal{T}}CN_cN_{\\mathcal{A}})" + }, + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "text", + "content": " is the maximum number of samples per class in test data, " + }, + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{A}}" + }, + { + "bbox": [ + 130, + 343, + 479, + 427 + ], + "type": "text", + "content": " is the number of extracted attributes. We only need to run the process once and use the generated tasks to benchmark various models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 444, + 230, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 230, + 458 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 230, + 458 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 468, + 262, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 262, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 262, + 480 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "content": "Datasets. We used two general datasets of different scales, miniImageNet [40] and tieredImageNet [41], and one fine-grained dataset, CUB-200 [56]. Each dataset consists of " + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{val}" + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "content": " for training, validation, and test, respectively (see Appendix). All images were resized to " + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "inline_equation", + "content": "84 \\times 84" + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 480, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 618 + ], + "type": "text", + "content": "FSC methods. We trained FSC models with ten algorithms covering three major categories. For gradient-based meta-learning algorithms, we chose ANIL [39], LEO [42], and BOIL [34]. For metric-based meta-learning algorithms, we chose ProtoNet [46], DN4 [24], R2D2 [5], CAN [16], and RENet [19]. For transfer learning algorithms, we chose Baseline++ [6] and RFS [50]. See Appendix for more details. Any backbones can be used as the feature extractor. For fair comparisons between different methods, we used the ResNet-12 backbone adopted in [35]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 618, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 479, + 666 + ], + "type": "text", + "content": "Text-based attribute detection. We used a pre-trained VLM named ViT-GPT2 [32] to generate text descriptions for images in " + }, + { + "bbox": [ + 130, + 618, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 130, + 618, + 479, + 666 + ], + "type": "text", + "content": ". After that, we used Spacy (https://spacy.io/) to extract nouns and adjectives from these descriptions automatically. We also used another pre-trained VLM, BLIP [23], to test" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 448, + 103 + ], + "type": "text", + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 135, + 463, + 177 + ], + "blocks": [ + { + "bbox": [ + 173, + 114, + 440, + 126 + ], + "lines": [ + { + "bbox": [ + 173, + 114, + 440, + 126 + ], + "spans": [ + { + "bbox": [ + 173, + 114, + 440, + 126 + ], + "type": "text", + "content": "Table 2: Statistics of detected attributes in " + }, + { + "bbox": [ + 173, + 114, + 440, + 126 + ], + "type": "inline_equation", + "content": "{\\mathcal{D}}_{\\text{test }}" + }, + { + "bbox": [ + 173, + 114, + 440, + 126 + ], + "type": "text", + "content": " by two VLMs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 135, + 463, + 177 + ], + "lines": [ + { + "bbox": [ + 148, + 135, + 463, + 177 + ], + "spans": [ + { + "bbox": [ + 148, + 135, + 463, + 177 + ], + "type": "table", + "html": "
VLMUnique attributesAvg. attributes per class
miniImageNettieredImageNetCUB-200miniImageNettieredImageNetCUB-200
ViT-GPT211112532159190.40230.9425.78
BLIP20326710247254.40310.4029.74
", + "image_path": "5c20361fd4e320960818b3046b8f12d0bff4a72d0065c59852c83eb4369634c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 185, + 216, + 322 + ], + "blocks": [ + { + "bbox": [ + 133, + 185, + 216, + 322 + ], + "lines": [ + { + "bbox": [ + 133, + 185, + 216, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 216, + 322 + ], + "type": "image", + "image_path": "3439fbe77ecf1db06516627fe54b44d47f8e42407bec302e739ba779246afc67.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 330, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 365 + ], + "type": "text", + "content": "Fig. 3: A 5-way 1-shot task constructed by our inter-class attribute-based sample selection using samples from the miniImageNet dataset. Note that due to the limited capacity of a VLM, the attributes may not well align with human understandings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 185, + 481, + 322 + ], + "blocks": [ + { + "bbox": [ + 217, + 185, + 481, + 322 + ], + "lines": [ + { + "bbox": [ + 217, + 185, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 217, + 185, + 481, + 322 + ], + "type": "image", + "image_path": "11d3cc987d5fc106a29b22db9362ebbb7fe3b46aaca79791304fbe4ec5f252de.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 410 + ], + "type": "text", + "content": "whether FewSTAB can produce consistent results. The statistics of the detected attributes are shown in Table 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 483, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 483, + 555 + ], + "type": "text", + "content": "Implementation details. We trained FSC models with the implementation in [25]. Each model was trained on " + }, + { + "bbox": [ + 130, + 411, + 483, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 130, + 411, + 483, + 555 + ], + "type": "text", + "content": " of a dataset with one of the ten FSC methods. For each meta-learning based method, we trained two models using randomly sampled 5-way 1-shot and 5-way 5-shot tasks, respectively. All the tasks have 15 samples per class in the query set. We saved the model that achieves the best validation accuracy on " + }, + { + "bbox": [ + 130, + 411, + 483, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{val}" + }, + { + "bbox": [ + 130, + 411, + 483, + 555 + ], + "type": "text", + "content": " for evaluation. For FewSTAB, if we do not have enough desired samples to construct a support set, we redo the construction from the beginning. If there are not enough desired samples to construct a query set, we first try to use the intra-class attribute-based sample selection; if the desired samples are still not enough, we redo the construction from the beginning. We created 3000 tasks for model evaluation. All experiments were conducted on the NVIDIA RTX 8000 GPUs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 573, + 342, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 342, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 342, + 585 + ], + "type": "text", + "content": "5.2 Visualization of a Constructed Task" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We show a 5-way 1-shot task constructed by FewSTAB in Fig. 3. Each class in the support set correlates with a unique spurious attribute. The query samples of a class do not have the spurious attribute correlated with the class and some of them have spurious attributes associated with other classes in the support set. For example, the query samples of the class lion do not have the spurious attribute water, and some of them have spurious attributes from other classes in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 166, + 168, + 446, + 345 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Comparison between wAcc-R and wAcc-A with " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " confidence interval on the miniImageNet, tieredImageNet, and CUB datasets. Numbers in the Shot column indicate that the models are both trained (if applicable) and tested on 5-way 1- or 5-shot tasks. Darker colors indicate higher values." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 166, + 168, + 446, + 345 + ], + "lines": [ + { + "bbox": [ + 166, + 168, + 446, + 345 + ], + "spans": [ + { + "bbox": [ + 166, + 168, + 446, + 345 + ], + "type": "table", + "html": "
ShotMethodminiImageNettieredImageNetCUB-200
wAcc-A (↑)wAcc-R (↑)wAcc-A (↑)wAcc-R (↑)wAcc-A (↑)wAcc-R (↑)
1ANIL10.38±0.3014.36±0.3311.21±0.3015.63±0.3613.78±0.4016.94±0.43
LEO14.26±0.4621.35±0.5416.00±0.5529.63±0.7128.29±0.8040.22±0.87
BOIL12.48±0.2314.93±0.2412.27±0.2114.13±0.2319.15±0.2922.50±0.29
ProtoNet14.03±0.4921.96±0.5814.50±0.5027.13±0.6934.62±0.8546.61±0.89
DN412.37±0.4619.28±0.5611.99±0.4723.62±0.6535.22±0.8647.26±0.88
R2D218.05±0.5326.50±0.6016.41±0.5430.41±0.7136.70±0.9048.82±0.88
CAN17.37±0.5325.96±0.6018.84±0.6036.29±0.7822.74±0.7231.95±0.78
RENet19.10±0.5728.85±0.6518.83±0.6135.70±0.7832.43±0.8143.98±0.81
Baseline++15.30±0.4823.18±0.5617.51±0.5431.62±0.719.17±0.4714.59±0.58
RFS18.00±0.5327.12±0.6118.35±0.6035.24±0.7732.45±0.8044.49±0.82
5ANIL14.83±0.4025.37±0.5213.72±0.3930.60±0.5131.63±0.5545.47±0.53
LEO26.31±0.5941.33±0.5929.49±0.7257.22±0.7246.62±0.8259.76±0.77
BOIL13.09±0.2215.21±0.2314.90±0.2218.55±0.2419.17±0.2821.33±0.27
ProtoNet32.07±0.5851.95±0.5230.95±0.7062.53±0.6260.06±0.7475.68±0.50
DN427.60±0.5842.68±0.6216.07±0.6240.63±0.8159.25±0.7773.58±0.56
R2D235.37±0.5950.84±0.5531.12±0.7261.08±0.6558.66±0.8275.20±0.54
CAN36.44±0.6554.23±0.5531.17±0.7664.19±0.6241.31±0.7461.61±0.58
RENet36.19±0.6356.52±0.5830.27±0.7663.49±0.6452.93±0.8271.82±0.56
Baseline++29.52±0.5744.94±0.5730.01±0.7259.06±0.6716.86±0.5229.84±0.69
RFS36.85±0.6455.66±0.5531.15±0.7662.71±0.6754.98±0.8174.33±0.53
", + "image_path": "5dbbd9a19d935159bc69b7007d19839678e332f4aa0ff43f5039d8b3d8c81583.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 132, + 354, + 247, + 433 + ], + "blocks": [ + { + "bbox": [ + 132, + 354, + 247, + 433 + ], + "lines": [ + { + "bbox": [ + 132, + 354, + 247, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 247, + 433 + ], + "type": "image", + "image_path": "77a621976c9e3e83be3775b669456e11c5b87d0e7f3f61174569f043553fa85f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 441, + 482, + 464 + ], + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 464 + ], + "type": "text", + "content": "Fig. 4: Accuracy gaps (wAcc-R minus wAcc-A) on the 5-way 1-shot and 5-way 5-shot tasks from the (a) miniImageNet, (b) tieredImageNet, and (c) CUB-200 datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 248, + 354, + 364, + 432 + ], + "blocks": [ + { + "bbox": [ + 248, + 354, + 364, + 432 + ], + "lines": [ + { + "bbox": [ + 248, + 354, + 364, + 432 + ], + "spans": [ + { + "bbox": [ + 248, + 354, + 364, + 432 + ], + "type": "image", + "image_path": "ff8b411fada5e4286329789d29cc4a0ff250a1219ebec889c5253dfb42285f6f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 354, + 482, + 432 + ], + "blocks": [ + { + "bbox": [ + 365, + 354, + 482, + 432 + ], + "lines": [ + { + "bbox": [ + 365, + 354, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 365, + 354, + 482, + 432 + ], + "type": "image", + "image_path": "3881f6eb4bb4aa3e8517b8b3565d61a8367db67a916c12552c64a931f8a738fc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 522 + ], + "type": "text", + "content": "the support set, such as man and green. FewSTAB introduces biased attributes in the task so that query samples can be easily misclassified as other classes by a few-shot classifier that relies on the spurious correlations in the support set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 538, + 292, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 538, + 292, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 538, + 292, + 550 + ], + "type": "text", + "content": "5.3 Effectiveness of FewSTAB" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "FewSTAB can effectively reveal spurious bias in few-shot classifiers. We show in Table 3 the wAcc (Eq. 2) on 5-way 1/5-shot test tasks that are randomly sampled (wAcc-R) and are constructed with FewSTAB (wAcc-A), respectively. FewSTAB generates FSC test tasks only based on the class-attribute correlations in data. In each test setting, the FSC methods in Table 3 are evaluated with the same FSC tasks. We observe that wAcc-A is consistently lower than wAcc-R on the three datasets and on two test-shot numbers, showing that FewSTAB is more effective than the standard evaluation procedure (random task construction) in exhibiting the spurious bias in various few-shot classifiers. We" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 448, + 102 + ], + "type": "text", + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "additionally show that FewSTAB also works on the most recent FSC methods and can reflect the improvement made to mitigate spurious bias (see Appendix)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 334, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 334, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 334, + 284 + ], + "type": "text", + "content": "FewSTAB reveals new robustness patterns among FSC methods. In Table 4, we calculate the Spearman's rank correlation coefficients [30] between the values of wAcc-R and wAcc-A from Table 3. The coefficients are bounded from 0 to 1, with larger values indicating that the ranks of FSC methods based on wAcc-R are more similar to those based on wAcc-A. In the 1-shot setting, it is not effective to control the spurious correlations since we only have one sample per class in the support set. Hence, the coefficients are large, and" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 353, + 211, + 466, + 258 + ], + "blocks": [ + { + "bbox": [ + 339, + 175, + 482, + 210 + ], + "lines": [ + { + "bbox": [ + 339, + 175, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 339, + 175, + 482, + 210 + ], + "type": "text", + "content": "Table 4: Spearman's rank correlations between wAcc-A and wAcc-R in Table 3" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 211, + 466, + 258 + ], + "lines": [ + { + "bbox": [ + 353, + 211, + 466, + 258 + ], + "spans": [ + { + "bbox": [ + 353, + 211, + 466, + 258 + ], + "type": "table", + "html": "
Dataset1-shot5-shot
miniImageNet0.960.95
tieredImageNet0.960.90
CUB-2001.000.94
", + "image_path": "e1f8252cff38797152a1d7d2ebb25929072dfb1d32b2328ebbcac0236704542c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 344 + ], + "type": "text", + "content": "the ranks based on wAcc-A are similar to those based on wAcc-R. In the 5-shot cases, we have more samples to demonstrate the spurious correlations. The coefficients become smaller, i.e., the ranks based on wAcc-A show different trends from those based on wAcc-R. In this case, FewSTAB reveals new information on FSC methods' varied degrees of robustness to spurious bias." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 345, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 429 + ], + "type": "text", + "content": "FewSTAB can benchmark spurious bias in varied degrees. As shown in Fig. 4, the accuracy gap, defined as wAcc-R minus wAcc-A, in general, becomes larger when we switch from 5-way 1-shot to 5-way 5-shot tasks. Compared with the random task construction, FewSTAB creates more challenging tasks in the 5-shot case for demonstrating spurious bias in few-shot classifiers. In other words, with a higher shot value in the constructed test tasks, FewSTAB aims to benchmark spurious bias in a higher degree." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 449, + 471, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 471, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 471, + 462 + ], + "type": "text", + "content": "5.4 A New Dimension of Evaluation and a New Design Guideline" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 334, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 334, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 334, + 666 + ], + "type": "text", + "content": "FewSTAB creates a new dimension of evaluation on the robustness to spurious bias. We demonstrate this with a scatter plot (Fig. 5) of Acc (Eq. 11) and wAcc-A of the ten few-shot classifiers. FewSTAB offers new information regarding different few-shot classifiers' robustness to spurious bias as we observe that Acc does not well correlate with wAcc-A. A high wAcc-A indicates that the classifier is robust to spurious bias, while a high Acc indicates that the classifier can correctly predict most of the samples. With the scatter plot, we can view tradeoffs between the two metrics on existing few-shot classifiers. A desirable few-shot classifier should appear in the top-right corner of the plot." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 347, + 498, + 474, + 580 + ], + "blocks": [ + { + "bbox": [ + 347, + 498, + 474, + 580 + ], + "lines": [ + { + "bbox": [ + 347, + 498, + 474, + 580 + ], + "spans": [ + { + "bbox": [ + 347, + 498, + 474, + 580 + ], + "type": "image", + "image_path": "22bcc91d8e622a747deb3d0e84d3396166351e12be6c55259b2e70e05221e339.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 588, + 482, + 622 + ], + "lines": [ + { + "bbox": [ + 338, + 588, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 338, + 588, + 482, + 622 + ], + "type": "text", + "content": "Fig.5: Acc versus wAcc-A of the ten FSC methods tested on 5-way 5-shot tasks from miniImageNet." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 467, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 467, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 467, + 129 + ], + "type": "text", + "content": "5.5 FewSTAB Enables Designs for Varied Degrees of Robustness" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 299, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 299, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 299, + 426 + ], + "type": "text", + "content": "As demonstrated in Section 5.3, Few-STAB can benchmark spurious bias in varied degrees, which in turn enables practitioners to design robust few-shot classifiers targeted for different degrees of robustness to spurious bias. The reason for differentiating designs for varied degrees of robustness is that the same design choice may not work under different robustness requirements. For example, increasing shot number in training tasks is a common strategy for improving the few-shot generalization of meta-learning based methods. We trained few-shot classifiers with 5-way 5-shot and 5-way 1-shot training tasks randomly sampled from " + }, + { + "bbox": [ + 130, + 140, + 299, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 130, + 140, + 299, + 426 + ], + "type": "text", + "content": ", respectively. We then calculated the accuracy gap defined as the wAcc-A of a model trained on 5-shot tasks minus the wAcc-A of the same model trained on 1-shot tasks. A positive and large accuracy gap indicates that this strat" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 427, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 510 + ], + "type": "text", + "content": "egy is effective in improving the model's robustness to spurious bias. In Fig. 6, on each of the three datasets, we give results of the eight meta-learning based FSC methods on the 5-way 1-, 5-, and 10-shot FewSTAB tasks which are used to demonstrate the strategy's robustness to increased degrees of spurious bias. This strategy does not work consistently under different test shots. For example, in Fig. 6(a) this strategy with CAN only works the best on the 5-way 5-shot FewSTAB tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 533, + 246, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 246, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 246, + 544 + ], + "type": "text", + "content": "5.6 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Techniques used in FewSTAB. We analyze how different sample selection methods affect the effectiveness of FewSTAB in Table 5. With only intra-class attribute-based sample selection, we randomly select query samples from Eq. ③. For inter-class attribute-based sample selection and intra-class attribute-based sample selection (automatically included by Eq. ④), we randomly select query samples from Eq. ⑤. FewSTAB uses all the techniques in Table 5. We define accuracy drop as wAcc-R minus wAcc-A, and we use the drop averaged over the ten FSC methods tested on 5-way 5-shot tasks from the miniImageNet dataset as our metric. A larger average drop indicates that the corresponding" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 305, + 165, + 480, + 222 + ], + "blocks": [ + { + "bbox": [ + 305, + 165, + 480, + 222 + ], + "lines": [ + { + "bbox": [ + 305, + 165, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 480, + 222 + ], + "type": "image", + "image_path": "43d7254cf725f38714ac47e3765cc3b0384c00334b8338f89dbb13f0aae9c173.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 222, + 480, + 277 + ], + "blocks": [ + { + "bbox": [ + 305, + 222, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 305, + 222, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 222, + 480, + 277 + ], + "type": "image", + "image_path": "cdff529d27acec6e59d9fbb595167f898308ae8dc75e2e5ba3bb866824d8dd52.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 278, + 481, + 344 + ], + "blocks": [ + { + "bbox": [ + 305, + 278, + 481, + 344 + ], + "lines": [ + { + "bbox": [ + 305, + 278, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 481, + 344 + ], + "type": "image", + "image_path": "54d76d0f09429176594ef571243424c2bcd7ee74c8dba207014c1a8d0a032151.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 351, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 304, + 351, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 482, + 396 + ], + "type": "text", + "content": "Fig. 6: Accuracy gaps of few-shot classifiers tested on 1-shot, 5-shot, and 10-shot tasks constructed from (a) miniImageNet, (b) tieredImageNet, and (c) CUB-200 datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 90, + 448, + 102 + ], + "type": "text", + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 157, + 291, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 309, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 309, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 309, + 148 + ], + "type": "text", + "content": "Table 5: Comparison between different techniques used by FewSTAB for constructing FSC tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 157, + 291, + 216 + ], + "lines": [ + { + "bbox": [ + 148, + 157, + 291, + 216 + ], + "spans": [ + { + "bbox": [ + 148, + 157, + 291, + 216 + ], + "type": "table", + "html": "
Attribute-based sample selectionQuery sample selectionAvg. drop (%)
Intra-classInter-class
✓5.13
✓✓13.30
✓✓✓15.05
", + "image_path": "f7ef8f68554fea56d7f1931c75e4797d356ee2a548e15b69ea7b33004676144f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 352, + 158, + 457, + 208 + ], + "blocks": [ + { + "bbox": [ + 329, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 329, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 329, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 6: Spearman's rank correlation coefficients between wAcc-A obtained using ViT-GPT2 and BLIP." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 158, + 457, + 208 + ], + "lines": [ + { + "bbox": [ + 352, + 158, + 457, + 208 + ], + "spans": [ + { + "bbox": [ + 352, + 158, + 457, + 208 + ], + "type": "table", + "html": "
Dataset1-shot5-shot
miniImageNet0.981.00
tieredImageNet1.000.99
CUB-2001.000.98
", + "image_path": "b55fe173cb225a16abb08662ced1dc7dbd76afcbae0b281214b16ba2bf666c03.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 238, + 481, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 481, + 273 + ], + "type": "text", + "content": "sample selection method is more effective in reflecting the spurious bias in few-shot classifiers. We observe that all proposed techniques are effective and the inter-class attribute-based sample selection is the most effective method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 274, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 332 + ], + "type": "text", + "content": "Choice of VLMs. Although our main results are based on the pre-trained ViT-GPT2 model [32], we show in Table 6 that when switching to a different VLM, i.e., BLIP [23], the relative ranks of different few-shot classifiers based on wAcc-A still hold with high correlations. In other words, FewSTAB is robust to different choices of VLMs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": "Detection accuracy of VLMs. A VLM may miss some attributes due to its limited capacity, resulting in a small detection accuracy. However, the detection accuracy of a VLM has little impact on our framework. To demonstrate this, we adopt a cross-validation strategy, i.e., we use the outputs from one VLM as the ground truth to evaluate those from another VLM, since assessing the detection accuracy of a VLM typically requires labor-intensive human labeling. On the CUB-200 dataset, we observe that the detection accuracy of ViT-GPT2 based on the BLIP's outputs is " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "70.12\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": ", while the detection accuracy of BLIP based on the ViT-GPT2's outputs is " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "59.28\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": ". Although the two VLMs differ significantly in the detected attributes, our framework shows almost consistent rankings of the evaluated FSC methods (Table 6)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 465, + 334, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 334, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 334, + 477 + ], + "type": "text", + "content": "Additional results are presented in Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 496, + 220, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 496, + 220, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 496, + 220, + 509 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we proposed a systematic and rigorous benchmark framework called FewSTAB for evaluating the robustness of few-shot classifiers to spurious bias. FewSTAB adopts attribute-based sample selection strategies to construct FSC test tasks with biased attributes so that the reliance on spurious correlations can be effectively revealed. FewSTAB can automatically benchmark spurious bias in few-shot classifiers on any existing test data thanks to its use of a pretrained VLM for automated attribute detection. With FewSTAB, we provided a new dimension of evaluation on the robustness of few-shot image classifiers to spurious bias and a new design guideline for building robust few-shot classifiers. FewSTAB can reveal and enable designs for varied degrees of robustness to spurious bias. We hope FewSTAB will inspire new developments on designing robust few-shot classifiers." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "content": "This work is supported in part by the US National Science Foundation under grants 2313865, 2217071, 2213700, 2106913, 2008208, 1955151." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "text", + "content": "1. Ahmed, F., Bengio, Y., Van Seijen, H., Courville, A.: Systematic generalisation with group invariant predictions. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "text", + "content": "2. Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.: Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "text", + "content": "3. Baker, N., Lu, H., Erlikhman, G., Kellman, P.J.: Deep convolutional networks do not classify based on global object shape. PLoS Computational Biology 14(12), e1006613 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "type": "text", + "content": "4. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: European Conference on Computer Vision. pp. 456-473 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "text", + "content": "5. Bertinetto, L., Henriques, J.F., Torr, P., Vedaldi, A.: Meta-learning with differentiable closed-form solvers. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "type": "text", + "content": "6. Chen, W.Y., Liu, Y.C., Kira, Z., Wang, Y.C.F., Huang, J.B.: A closer look at few-shot classification. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "type": "text", + "content": "7. Chen, Y., Wang, X., Liu, Z., Xu, H., Darrell, T.: A new meta-baseline for few-shot learning. arXiv preprint arXiv:2003.04390 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 412, + 481, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 481, + 455 + ], + "type": "text", + "content": "8. Dong, J., Wang, Y., Lai, J.H., Xie, X.: Improving adversarially robust few-shot image classification with generalizable representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9025-9034 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 456, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 488 + ], + "type": "text", + "content": "9. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Proceedings of the 34th International Conference on Machine Learning. vol. 70, pp. 1126–1135 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 490, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 481, + 521 + ], + "type": "text", + "content": "0. Flennerhag, S., Rusu, A.A., Pascanu, R., Visin, F., Yin, H., Hadsell, R.: Meta-learning with warped gradient descent. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 565 + ], + "type": "text", + "content": "1. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 567, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 481, + 609 + ], + "type": "text", + "content": "12. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 481, + 632 + ], + "type": "text", + "content": "13. Ghosal, S.S., Li, Y.: Are vision transformers robust to spurious correlations? International Journal of Computer Vision 132(3), 689-709 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "14. Goldblum, M., Fowl, L., Goldstein, T.: Adversarily robust few-shot learning: A meta-learning approach. Advances in Neural Information Processing Systems 33, 17886-17895 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "text", + "content": "15. He, Y., Shen, Z., Cui, P.: Towards non-iid image classification: A dataset and baselines. Pattern Recognition 110, 107383 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "16. Hou, R., Chang, H., Bingpeng, M., Shan, S., Chen, X.: Cross attention network for few-shot classification. In: Advances in Neural Information Processing Systems. pp. 4003-4014 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "17. Hu, S.X., Moreno, P.G., Xiao, Y., Shen, X., Obozinski, G., Lawrence, N., Dami-anou, A.: Empirical bayes transductive meta-learning with synthetic gradients. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "text", + "content": "18. Jang, H., Lee, H., Shin, J.: Unsupervised meta-learning via few-shot pseudosupervised contrastive learning. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 238, + 481, + 270 + ], + "type": "text", + "content": "19. Kang, D., Kwon, H., Min, J., Cho, M.: Relational embedding for few-shot classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8822-8833 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 481, + 304 + ], + "type": "text", + "content": "20. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 134, + 304, + 481, + 336 + ], + "type": "text", + "content": "21. Lee, K., Maji, S., Ravichandran, A., Soatto, S.: Meta-learning with differentiable convex optimization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10657-10665 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 337, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 337, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 337, + 481, + 369 + ], + "type": "text", + "content": "22. Lee, Y., Choi, S.: Gradient-based meta-learning with learned layerwise metric and subspace. In: International Conference on Machine Learning. pp. 2927-2936. PMLR (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 134, + 369, + 481, + 402 + ], + "type": "text", + "content": "23. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 134, + 403, + 481, + 435 + ], + "type": "text", + "content": "24. Li, W., Wang, L., Xu, J., Huo, J., Gao, Y., Luo, J.: Revisiting local descriptor based image-to-class measure for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7260-7268 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 134, + 436, + 481, + 468 + ], + "type": "text", + "content": "25. Li, W., Wang, Z., Yang, X., Dong, C., Tian, P., Qin, T., Jing, H., Shi, Y., Wang, L., Gao, Y., Luo, J.: Libfewshot: A comprehensive library for few-shot learning. IEEE Transactions on Pattern Analysis & Machine Intelligence (01), 1-18 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 481, + 490 + ], + "type": "text", + "content": "26. Li, Z., Zhou, F., Chen, F., Li, H.: Meta-SGD: Learning to learn quickly for few shot learning. ArXiv abs/1707.09835 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 134, + 491, + 481, + 522 + ], + "type": "text", + "content": "27. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3730-3738 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 523, + 481, + 555 + ], + "type": "text", + "content": "28. Lu, Y., Wen, L., Liu, J., Liu, Y., Tian, X.: Self-supervision can be a good few-shot learner. In: European Conference on Computer Vision. pp. 740-758. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 481, + 578 + ], + "type": "text", + "content": "29. Motiian, S., Jones, Q., Iranmanesh, S., Doretto, G.: Few-shot adversarial domain adaptation. Advances in Neural Information Processing systems 30 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 481, + 599 + ], + "type": "text", + "content": "30. Myers, L., Sirois, M.J.: Spearman correlation coefficients, differences between. Encyclopedia of Statistical Sciences 12 (2004)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "text", + "content": "31. Neuhaus, Y., Augustin, M., Boreiko, V., Hein, M.: Spurious features everywhere large-scale detection of harmful spurious features in imagenet. arXiv preprint arXiv:2212.04871 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "text", + "content": "32. NLP Connect: vit-gpt2-image-captioning (revision 0e334c7) (2022). https://doi.org/10.57967/hf/0222, https://huggingface.co/nlpconnect/vit-gpt2-image-captioning" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Nushi, B., Kamar, E., Horvitz, E.: Towards accountable ai: Hybrid human-machine analyses for characterizing system failure. In: Proceedings of the AAAI Conference on Human Computation and Crowdsourcing. vol. 6, pp. 126-135 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 172 + ], + "type": "text", + "content": "34. Oh, J., Yoo, H., Kim, C., Yun, S.Y.: Boil: Towards representation change for few-shot learning. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "type": "text", + "content": "35. Oreshkin, B., Rodríguez López, P., Lacoste, A.: Tadam: Task dependent adaptive metric for improved few-shot learning. In: Advances in Neural Information Processing Systems 31, pp. 721-731 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "text", + "content": "36. Poulakakis-Daktylidis, S., Jamali-Rad, H.: Bectr: Batch enhanced contrastive few-shot learning. In: International Conference on Learning Representations (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "37. Qin, X., Song, X., Jiang, S.: Bi-level meta-learning for few-shot domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15900-15910 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 304 + ], + "type": "text", + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "text", + "content": "39. Raghu, A., Raghu, M., Bengio, S., Vinyals, O.: Rapid learning or feature reuse? towards understanding the effectiveness of maml. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "text", + "content": "40. Ravi, S., Larochelle, H.: Optimization as a model for few-shot learning. In: International Conference on Learning Representations (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "text", + "content": "41. Ren, M., Ravi, S., Triantafillou, E., Snell, J., Swersky, K., Tenenbaum, J.B., Larochelle, H., Zemel, R.S.: Meta-learning for semi-supervised few-shot classification. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 390, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 424 + ], + "type": "text", + "content": "42. Rusu, A.A., Rao, D., Sygnowski, J., Vinyals, O., Pascanu, R., Osindero, S., Hadsell, R.: Meta-learning with latent embedding optimization. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 424, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 445 + ], + "type": "text", + "content": "43. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "type": "text", + "content": "44. Singla, S., Feizi, S.: Salient imagenet: How to discover spurious features in deep learning? In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "45. Singla, S., Nushi, B., Shah, S., Kamar, E., Horvitz, E.: Understanding failures of deep networks via robust feature extraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12853-12862 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 501, + 482, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 523 + ], + "type": "text", + "content": "46. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems 30, pp. 4077-4087 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 523, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 555 + ], + "type": "text", + "content": "47. Stock, P., Cisse, M.: Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases. In: European Conference on Computer Vision. pp. 498-512 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "text", + "content": "48. Sun, Q., Liu, Y., Chua, T.S., Schiele, B.: Meta-transfer learning for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 403-412 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "text", + "content": "49. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H.S., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1199-1208 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "50. Tian, Y., Wang, Y., Krishnan, D., Tenenbaum, J.B., Isola, P.: Rethinking few-shot image classification: a good embedding is all you need? In: European Conference on Computer Vision. pp. 266-282. Springer (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarking Spurious Bias in Few-Shot Image Classifiers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 658 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "51. Triantafillou, E., Zhu, T., Dumoulin, V., Lamblin, P., Evci, U., Xu, K., Goroshin, R., Gelada, C., Swersky, K., Manzagol, P.A., Larochelle, H.: Meta-dataset: A dataset of datasets for learning to learn from few examples. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 482, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 482, + 193 + ], + "type": "text", + "content": "52. Tseng, H.Y., Lee, H.Y., Huang, J.B., Yang, M.H.: Cross-domain few-shot classification via learned feature-wise transformation. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 482, + 225 + ], + "type": "text", + "content": "53. Vinyals, O., Blundell, C., Lillicrap, T., kavukcuoglu, k., Wierstra, D.: Matching networks for one shot learning. In: Advances in Neural Information Processing Systems 29, pp. 3630-3638 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 482, + 258 + ], + "type": "text", + "content": "54. Wang, Y., Girshick, R., Hebert, M., Hariharan, B.: Low-shot learning from imaginary data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7278-7286 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 258, + 482, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 482, + 280 + ], + "type": "text", + "content": "55. Wang, Y., Yao, Q., Kwok, J.T., Ni, L.M.: Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys 53(3), 1-34 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 280, + 482, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 482, + 312 + ], + "type": "text", + "content": "56. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-UCSD Birds 200. Tech. Rep. CNS-TR-2010-001, California Institute of Technology (2010)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 312, + 482, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 482, + 344 + ], + "type": "text", + "content": "57. Wong, E., Santurkar, S., Madry, A.: Leveraging sparse linear layers for debuggable deep networks. In: International Conference on Machine Learning. pp. 11205-11216. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 344, + 482, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 482, + 376 + ], + "type": "text", + "content": "58. Xiao, K.Y., Engstrom, L., Ilyas, A., Madry, A.: Noise or signal: The role of image backgrounds in object recognition. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 376, + 482, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 482, + 398 + ], + "type": "text", + "content": "59. Xue, Y., Payani, A., Yang, Y., Mirzasoleiman, B.: Eliminating spurious correlations from pre-trained models via data mixing. arXiv preprint arXiv:2305.14521 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 398, + 482, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 482, + 430 + ], + "type": "text", + "content": "60. Yang, Y., Nushi, B., Palangi, H., Mirzasoleiman, B.: Mitigating spurious correlations in multi-modal models during fine-tuning. arXiv preprint arXiv:2304.03916 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 482, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 482, + 463 + ], + "type": "text", + "content": "61. Ye, H.J., Hu, H., Zhan, D.C., Sha, F.: Few-shot learning via embedding adaptation with set-to-set functions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8808-8817 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 463, + 482, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 482, + 485 + ], + "type": "text", + "content": "62. Yue, X., Zheng, Z., Das, H.P., Keutzer, K., Vincentelli, A.S.: Multi-source few-shot domain adaptation. arXiv preprint arXiv:2109.12391 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 485, + 482, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 482, + 506 + ], + "type": "text", + "content": "63. Yue, Z., Zhang, H., Sun, Q., Hua, X.S.: Interventional few-shot learning. Advances in Neural Information Processing Systems 33, 2734-2746 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 506, + 482, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 482, + 528 + ], + "type": "text", + "content": "64. Zhang, C., Cai, Y., Lin, G., Shen, C.: Deepemd: Differentiable earth mover's distance for few-shot learning (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 528, + 482, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 482, + 559 + ], + "type": "text", + "content": "65. Zhang, C., Cai, Y., Lin, G., Shen, C.: Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 559, + 482, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 482, + 592 + ], + "type": "text", + "content": "66. Zhang, J., Wang, Y., Molino, P., Li, L., Ebert, D.S.: Manifold: A model-agnostic framework for interpretation and diagnosis of machine learning models. IEEE Transactions on Visualization and Computer Graphics 25(1), 364-373 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 592, + 482, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 482, + 624 + ], + "type": "text", + "content": "67. Zhang, M., Li, H., Wu, F., Kuang, K.: Metacoco: A new few-shot classification benchmark with spurious correlation. In: The Twelfth International Conference on Learning Representations (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 624, + 482, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 482, + 658 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 482, + 658 + ], + "type": "text", + "content": "68. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_content_list.json b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..92209eb3681d6e81c969471a66e52d06152b9fad --- /dev/null +++ b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_content_list.json @@ -0,0 +1,1745 @@ +[ + { + "type": "text", + "text": "Benchmarking the Robustness of Cross-view Geo-localization Models", + "text_level": 1, + "bbox": [ + 248, + 140, + 754, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingwang Zhang and Yingying Zhu", + "bbox": [ + 349, + 212, + 653, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "College of Computer Science and Software Engineering, Shenzhen University,", + "bbox": [ + 243, + 239, + 756, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenzhen, China", + "bbox": [ + 444, + 253, + 557, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhangqingwang2022@email.szu.edu.cn，zhuyy@szu.edu.cn", + "bbox": [ + 300, + 268, + 700, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://zqwlearning.github.io/CrossViewRobustness", + "bbox": [ + 308, + 282, + 691, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Cross-view geo-localization serves as a viable alternative to providing geographical location information when GPS signals are unstable or unavailable by matching ground images with geo-tagged aerial image databases. While significant progress has been made on some common benchmarks like CVUSA and CVACT, there remains a lack of comprehensive consideration for robustness against real-world environmental challenges such as adverse weather or sensor noise. This deficiency poses a significant challenge for deploying this technology in safety-critical domains like autonomous driving and robot navigation. To the best of our knowledge, there is currently no specialized benchmark for evaluating the robustness of cross-view geo-localization. To comprehensively and fairly evaluate the robustness of cross-view geo-localization models in real-world scenarios, we introduce 16 common types of data corruption. By synthesizing these corruptions on public datasets, we establish two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL), covering approximately 1.5 million corrupted images. Subsequently, we conduct large-scale experiments on various cross-view geo-localization models to evaluate their robustness in corrupted environments and derive novel insights. Finally, we explore two data augmentation strategies as potential solutions to enhance model robustness. Combined with the training strategies proposed, these approaches effectively enhance the robustness of multiple models.", + "bbox": [ + 261, + 332, + 743, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Cross-view geo-localization $\\cdot$ Benchmarking $\\cdot$ Robustness", + "bbox": [ + 261, + 678, + 730, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 718, + 375, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Geo-localization is fundamental for autonomous driving [26], robot navigation [24], 3D reconstruction [14], augmented reality [27], and road maintenance [2], etc. However, in the realistic scenario, a GPS-degraded or GPS-denied environment is not uncommon, including urban canyon effects, forest areas, mountainous terrain, and interference or jamming regions. The absence of geo-tags in image metadata, often resulting from GPS signal obstruction, signal attenuation, or", + "bbox": [ + 212, + 748, + 792, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06bd3fd4b0605c8f789f4ebd564a7bd8d03c092fb68b5f03881678912ad5b12c.jpg", + "image_caption": [ + "Fig. 1: Existing cross-view geo-localization models fail when the ground query image is corrupted. When the query image is clean, almost all existing methods retrieve the ground truth correctly (marked by the red box and the red circle). Once the image is corrupted, almost all existing methods have missed the correct result. Incorrect retrieve and location results are marked in green and blue." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lack of a camera localization module, leads to only $4.3\\%$ of Internet images being geo-tagged [7]. Cross-view geo-localization can determine the location where a ground-level image is taken by comparing it with a geo-tagged database of aerial images (e.g., satellite images), which can serve as an alternative way or an effective complement to provide geo-location information in the case of GPS-degraded or GPS-denied. Therefore, cross-view geo-localization has recently garnered considerable attention due to its wide range of practical applications.", + "bbox": [ + 212, + 380, + 787, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Two favorite benchmarks have been introduced to evaluate cross-view geo-localization performance, e.g., CVUSA [40] and CVACT [21]. Existing methods are all evaluated on both datasets and state-of-the-art approaches even achieve near-perfect results [42,43,45]. Nevertheless, these benchmarks only present ideal settings and do not reflect real-world scenarios. Almost all their ground images were collected on clean days, with little coverage in conditions such as rain, snow, and sensor noise corruption. Inevitably, due to the significant differences between the distributions of the training data and the real-world data, the performance of current data-driven models degrades drastically or even fails when the models trained on \"clean\" data are applied to \"corrupted\" data. Fig. 1 shows examples of localization failures that can result from the ground-level query under different weather conditions. This poses a significant challenge to applying this task to safety-critical applications such as autonomous driving. Thus, prior to deployment, it has become imperative to create an easy-to-use corruption robustness benchmark and to comprehensively evaluate the robustness of existing models with respect to corrupted data. To the best of our knowledge, we first create cross-view geo-localization robustness benchmarks and provide an exhaustive assessment of the robustness of existing mainstream cross-view geo-localization models when faced with corrupted data.", + "bbox": [ + 212, + 489, + 787, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the cross-view geo-localization task setup, the aerial images are cached by the system in advance with high-definition and stable characteristics, while the ground-view images are collected in real-time, which is more flexible and versatile. Therefore, in this paper, we only corrupt the ground images to consider", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/16fcf319507002768594b158dcce7f6092c42045997bab15964a51a7b60c7406.jpg", + "image_caption": [ + "Clean" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 328, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/86a29d270ec50067d1d63210cde1cab99d587fa047bcad0105ed588710224b93.jpg", + "image_caption": [ + "Weather - Snow" + ], + "image_footnote": [], + "bbox": [ + 331, + 143, + 442, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8be61a441409db8a515cfc624ed746383bdbb1d963a241a47462ac7f5fdff23d.jpg", + "image_caption": [ + "Blur - Defocus" + ], + "image_footnote": [], + "bbox": [ + 444, + 143, + 555, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/04c907550e4dbe1e9d4bec543738ad329a3ad1385e785af42128e4012e887666.jpg", + "image_caption": [ + "Noise-Gaussian" + ], + "image_footnote": [], + "bbox": [ + 558, + 143, + 669, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3b8089f5e83a3868d7a34237f1ec89e5c5f4395f410ab109a3744533afbb1924.jpg", + "image_caption": [ + "Digital-JPEG" + ], + "image_footnote": [], + "bbox": [ + 671, + 143, + 784, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/93bb8548180c491d9b84d1f3d77a634d93ccdea4f83982a79f49fb821d3a868d.jpg", + "image_caption": [ + "Severity $= 1$", + "(b) Comparison of different severity levels, taking snow as an example.", + "Fig. 2: Various types and severities of common corruption in our robustness benchmarks, after cropping are provided (best viewed when zoomed in on the screen). Visualizations of all corruption types can be found in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 218, + 217, + 328, + 260 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/53542b972cfd266f41ebd48dd936501018af1b0da302fe523729f00e45d853e0.jpg", + "image_caption": [ + "Severity $= 2$" + ], + "image_footnote": [], + "bbox": [ + 331, + 217, + 442, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/60f90392f7d73aa66daa8e25e0e469a25d8aca5722912552a7372c7d9d07d3e1.jpg", + "image_caption": [ + "(a) Comparison between clean and corrupted images.", + "Severity $= 3$" + ], + "image_footnote": [], + "bbox": [ + 444, + 215, + 555, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fb0a4cf89e82329d1a452b8ea15aa9537e1455000858c51c412b0588ae8cb366.jpg", + "image_caption": [ + "Severity $= 4$" + ], + "image_footnote": [], + "bbox": [ + 558, + 215, + 669, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0212ff984b351a754bf813d509db881005f0dc4a8d5efa7283d309c178dda7dc.jpg", + "image_caption": [ + "Severity $= 5$" + ], + "image_footnote": [], + "bbox": [ + 673, + 215, + 784, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the robustness of the existing models, which is more practical while alleviating the evaluation cost. We systematically design 16 common types of corruption that are commonly found in the cross-view geo-localization task and provide a comprehensive evaluation of the corruption robustness of existing models. These corruptions include Weather, Blur, Noise, and Digital, covering a wide range of real-world situations. Furthermore, each type of corruption has 5 severity levels, totaling 80 different types of corruption involved. We have applied these corruptions on CVUSA [40] and CVACT [21] to create two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL), which cover a total of about 1.5 million corrupted images. Fig. 2 shows some examples of corrupted images. We expect these datasets to serve as common benchmarks for comprehensively and fairly evaluating the corruption robustness of cross-view geo-localization models and to facilitate future research to advance the cross-view geo-localization task.", + "bbox": [ + 212, + 373, + 787, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions made to this work are summed up as follows:", + "bbox": [ + 238, + 601, + 699, + 616 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we have benchmarked the robustness of state-of-the-art cross-view geo-localization models against real-world data corruption challenges for the first time. We have generated about 1.5 million corrupted images based on the CVUSA and CVACT datasets to establish benchmarks for assessing the robustness of cross-view geo-localization models.", + "- Based on the benchmark study, we have several new insights: (1) in most cases, the clean performance $(\\mathrm{R}@\\mathcal{K}_{\\mathrm{clean}})$ of a model is positively, but not absolutely, correlated with its robustness; (2) snow, spatter, and zoom blur more significantly affect the robustness of various models compared to other corruptions; (3) models trained on more intricate scenarios (e.g., CVACT) exhibit better robustness.", + "- Introducing stylization and histogram equalization as data augmentation techniques, along with our proposed training strategy, significantly enhances the robustness of various cross-view geo-localization models." + ], + "bbox": [ + 223, + 627, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Cross-view Geo-localization Models", + "bbox": [ + 272, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Cross-view Geo-localization", + "text_level": 1, + "bbox": [ + 215, + 180, + 491, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-view geo-localization has witnessed significant advancements in recent years. Initially, [40] introduced Convolutional Neural Networks (CNN) to the cross-view matching task, resulting in notable performance improvements. Subsequently, [17] employed a VGG [36] backbone network with two branches, combined with the NetVlad [1], and proposed the weighted soft-margin ranking loss, achieving state-of-the-art performance. To further enhance network performance, [21] emphasized the importance of orientation in cross-view geo-localization and devised a method to provide orientation information to the neural network explicitly. [33] adopted spatial-aware feature aggregation modules to aggregate information-rich and diverse feature maps, while introducing polar transformation for pre-processing to narrow the geometric gap between center-aligned satellites and ground-level images. Recently, [42] explored the integration of Vision Transformer [5] in cross-view geo-localization, proposing a novel layer-to-layer Transformer with self-cross attention mechanism that highlighted the significance of considering global dependencies to reduce visual ambiguity. [45] introduced an attention-guided pure Transformer approach, which further improved the resolution of satellite images through additional training, thus advancing the performance of the task. [43] introduced a novel geometric layout extraction module that explicitly decouples geometric information from original features and proposed two types of data augmentation methods. However, all these methods ignore the performance of the model on corrupted images, i.e., robustness, which is the focus of this paper.", + "bbox": [ + 212, + 209, + 787, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Robustness Benchmarks", + "text_level": 1, + "bbox": [ + 215, + 568, + 465, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robustness benchmarks are essential in the field of computer vision to enhance the stability and reliability of computer vision systems when facing uncertainties and noise. For instance, IMAGENET-C and IMAGENET-P were proposed to evaluate robustness in classification tasks [16]. In the context of autonomous driving, PASCAL-C, COCO-C, and Cityscapes-C were introduced as benchmarks for evaluating the robustness of object detection tasks [13, 25]. Similarly, robustness benchmarks for semantic segmentation [23] were established using PASCAL VOC 2012 [6], Cityscapes [3], and ADE20K [18, 44]. Furthermore, KITTI-C, nuScenes-C, and Waymo-C were devised as robustness benchmarks for 3D object detection [4]. These benchmarks aid in evaluating the robustness of models under various complex environmental conditions, thus facilitating their application in the real world. However, to the best of our knowledge, no researcher has proposed relevant robustness benchmarks specifically for cross-view geo-localization. Therefore, this paper addresses this gap by introducing the benchmarks CVUSA-C, CVACT_val-C, CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL, thus filling this void in the field.", + "bbox": [ + 212, + 598, + 789, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Zhang et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Robustness Enhancement", + "text_level": 1, + "bbox": [ + 215, + 146, + 472, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Several strategies have been proposed to address the impact of corruption in computer vision. For example, [20] and [8] present methods that utilize patch-based priors for background and rain layers, along with a deep CNN to effectively remove rain streaks from images. In a similar vein, [15] introduced a simple yet powerful technique employing the dark channel before eliminating haze from single-input images. Moreover, [22] devised the DesnowNet, a multi-stage network specifically designed to remove snow particles from images. However, a limitation of these methods lies in their reliance on specific designs tailored to particular types of corruption. This hinders their ability to generalize and handle other types of corruption effectively. As an alternative, some approaches have sought to enhance model performance through data augmentation, incorporating corrupted data during the training process. Although fine-tuning on specific corruption data has shown promise in boosting performance for those particular corruption types [12, 39] discovered that fine-tuning on one type of corruption often struggles to generalize to other corruption types. How to enhance robustness remains an open question. In this work, we employ the stylization [11] and histogram equalization [28] techniques to the train set, and we observe improvements in robustness.", + "bbox": [ + 212, + 167, + 787, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Robustness in Cross-view Geo-localization", + "text_level": 1, + "bbox": [ + 215, + 460, + 660, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 489, + 437, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Robustness. First, we consider a set of ground-to-aerial image pairs $\\{\\mathbf{I}_i^g,\\mathbf{I}_i^a\\}_{i = 1}^N$ where $N$ represents the number of image pairs. The superscripts $g$ and $a$ denote ground and aerial views, respectively. For a ground query image indexed as $q$ , we assume the existence of ground image encoder $f_{g}$ and aerial image encoder $f_{a}$ , which have been trained on samples from the distribution $\\mathcal{D}$ . We let $\\mathbb{P}_{\\mathcal{C}}(c)$ approximate the frequency of real-world corruption. Existing models mostly evaluate performance when samples are drawn from the distribution $\\mathcal{D}$ , denoted as $\\mathbb{P}_{(\\mathbf{I}^g,\\mathbf{I}^a)\\sim \\mathcal{D}}(d(f_g(\\mathbf{I}_q^g),f_a(\\mathbf{I}_q^a)) < \\{d(f_g(\\mathbf{I}_q^g),f_a(\\mathbf{I}_i^a))|\\forall i\\in \\{1,\\dots,N\\},i\\neq q\\})$ , $d(\\cdot ,\\cdot)$ representing the $L_{2}$ distance. However, in practical deployments, systems often need to operate on low-quality or corrupted images. Hence, we construct an evaluation of the corruption robustness of the model, denoted as:", + "bbox": [ + 212, + 508, + 787, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {c \\sim \\mathcal {C}} \\left[ \\mathbb {P} _ {\\left(\\mathbf {I} ^ {g}, \\mathbf {I} ^ {a}\\right) \\sim \\mathcal {D}} \\left(d \\left(f _ {g} \\left(c \\left(\\mathbf {I} _ {q} ^ {g}\\right)\\right), f _ {a} \\left(\\mathbf {I} _ {q} ^ {a}\\right)\\right) < \\left\\{d \\left(f _ {g} \\left(c \\left(\\mathbf {I} _ {q} ^ {g}\\right)\\right), f _ {a} \\left(\\mathbf {I} _ {i} ^ {a}\\right)\\right) | \\forall i \\in \\{1, \\dots , N \\}, i \\neq q \\right\\}\\right) \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 670, + 772, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, in order to approximate the real-world corruption $\\mathcal{C}$ , we designed a set of common corruptions (Weather, Blur, Noise, and Digital) in the cross-view geo-localization task. The types of corruption already provide a good overview of corruption that street view images may face. From these common corruptions, we create our cross-view geo-localization robustness evaluation benchmarks.", + "bbox": [ + 212, + 691, + 785, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Image Corruption", + "text_level": 1, + "bbox": [ + 215, + 787, + 411, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Common corruptions are categorized based on [16], and ground-level image corruptions are divided into four major categories with a total of 16 subcategories.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Cross-view Geo-localization Models", + "bbox": [ + 271, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first major category is weather-related corruptions, which encompasses snow, frost, fog, brightness, and spatter. The second major category consists of blur-related corruptions, including defocus, glass, motion, and zoom blur. Defocus blur occurs when an image is out of focus, while glass blur arises when images are captured through frosted glass windows of sensors. Motion and zoom blurring occur in scenes with rapid camera movement or quick approaches toward objects, and they are particularly prone to occur in image collections from vehicular devices. The third major category involves noise-related corruptions, which cover Gaussian, shot, impulse, and speckle noise. Gaussian noise emerges under low-light conditions, attributed to the discrete nature of light and causing electronic noise. Impulse noise results from bit errors, resembling the color analog of salt-and-pepper noise. Speckle noise, which occurs due to light interference, leads to the appearance of bright and dark speckles. Lastly, the fourth category includes digital corruptions, including contrast, pixelation, and JPEG corruption. Contrast corruption is dependent on lighting conditions and the colors of the objects during capture. Pixelation occurs during upscaling low-resolution images, and JPEG corruption arises when images undergo lossy compression, introducing compression artifacts. To better simulate real-world corruption, each type of corruption also contains 5 severity levels. We present visualizations of various types and severities of common corruption in Fig. 2.", + "bbox": [ + 212, + 146, + 787, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Robustness Enhancement Methods", + "text_level": 1, + "bbox": [ + 214, + 468, + 552, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enhance the robustness of existing models against corruption, we investigate two data augmentation approaches: stylization and histogram equalization. By applying either stylization or histogram equalization to the training data and employing our proposed training strategy, we effectively bolstered the corruption robustness of multiple cross-view geo-localization models.", + "bbox": [ + 212, + 492, + 785, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stylization. Style transfer, introduced by [9], merges the content and style of two different images to generate a novel image that retains the content of the original image while adopting the style of the target image, which is also known as stylization. In the context of image classification and object detection tasks, style transfer has demonstrated its effectiveness in enhancing robustness [10].", + "bbox": [ + 212, + 568, + 785, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Histogram Equalization. Histogram equalization is a fundamental technique used for enhancing image contrast. It operates by redistributing pixel intensity values within an image, stretching or compressing the brightness range, thus achieving a more uniform distribution of pixel intensity values across the entire brightness range of the image. In this study, we employ Contrast Limited Adaptive Histogram Equalization (CLAHE) [29] to enhance the robustness of existing methods.", + "bbox": [ + 212, + 643, + 787, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Strategy. In this paper, we apply Stylization / CLAHE to the cross-view geo-localization training set, testing 3 settings: (1) training using the standard (raw) training data; (2) replacing all training images with Stylization / CLAHE images, thereby eliminating the use of standard images during training; (3) having Stylization / CLAHE images participate in the training process alongside the standard images with equal probability at each iteration. Fig. 3", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8e94643ccb2b9ee0d9fd5fdad02f27ce85203c2e21478727f18368cc12886479.jpg", + "image_caption": [ + "Fig. 3: Visualization of Stylization / CLAHE applied to the CVUSA training set. The illustration depicts standard images (middle row), stylization images (top row), and CLAHE images (bottom row). The rounded rectangles on both sides represent different training strategies. The red dashed boxes indicate the style images sourced from https://www.kaggle.com/c/painter-by-numbers/. Inspired by [25]." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 784, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shows examples of applying Stylization / CLAHE, and it illustrates 3 training strategies with the same training complexity.", + "bbox": [ + 212, + 378, + 784, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Corruption Robustness Benchmarks", + "text_level": 1, + "bbox": [ + 215, + 431, + 599, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We create two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) from the classic cross-view geo-localization datasets CVUSA and CVACT. See Table 1 for more detailed information on these benchmarks.", + "bbox": [ + 212, + 460, + 792, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Fine-grained Corruption Robustness Benchmarks", + "text_level": 1, + "bbox": [ + 215, + 559, + 668, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We create fine-grained robustness benchmarks comprising 16 corruption types, each corresponding to 5 severity levels denoted by integers from 1 to 5, where higher numbers indicate more severe corruption. Consequently, one single original ground query image can yield 80 $(16 \\times 5)$ corrupted images. In the fine-grained corruption robustness benchmarks, each type of corrupted image forms an independent evaluation subset with no overlap. The suffix \"-C\" designates all fine-grained robustness benchmarks (CVUSA-C and CVACT_val-C).", + "bbox": [ + 212, + 583, + 784, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CVUSA-C. The CVUSA [40] dataset is one of the earliest cross-view geo-localization datasets, primarily collected from suburban areas in the United States. It encompasses a total of 35,532 pairs of ground-aerial images for training and 8,884 pairs for testing. We have employed the CVUSA test set to generate the fine-grained corruption robustness benchmark, CVUSA-C. To be specific, CVUSA-C comprises all types and severities of image corruptions, each residing independently within individual subsets of CVUSA-C. Given that there are 16 distinct corruption types and 5 levels of corruption severity, CVUSA-C effectively comprises 80 evaluation subsets. Each of these subsets contains 8,884 ground images for testing, resulting in an aggregate of 710,720 corrupted images.", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Cross-view Geo-localization Models", + "bbox": [ + 272, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/107ebb215b14fea7b660bd7c7598441c9705be06fadbc8a4439b97c8e2d8d24f.jpg", + "table_caption": [ + "Table 1: Detailed information on the proposed corruption robustness benchmarks." + ], + "table_footnote": [], + "table_body": "
Fine-grainedComprehensive
CVUSA-CCVACT_val-CCVUSA-C-ALLCVACT_val-C-ALLCVACT_test-C-ALL
Number of original validation / test ground images8,8848,8848,8848,88492,802
Whether or not evaluation subsets are generated for each corruption✓✓✗✗✗
Whether all corruptions are included in an independent subset✗✗✓✓✓
Number of validation / test ground images for our benchmark8,884 × 16 × 58,884 × 16 × 58,8848,88492,802
Storage space~39 GB~178 GB~0.5 GB~2 GB~21 GB
", + "bbox": [ + 218, + 159, + 784, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CVACT_val-C. The CVACT [21] dataset was collected by [21] and densely covers an urban area of Australia (Canberra). Similar to the CVUSA dataset, this dataset comprises 35,532 pairs of ground-aerial images for training and 8,884 image pairs for validation (referred to as CVACT_val). We create a fine-grained robustness benchmark from CVACT_val, as CVUSA does. The benchmark includes various types and severities, resulting in a total of 80 validation subsets. These subsets consist of 710,720 corrupted images and are collectively referred to as CVACT_val-C.", + "bbox": [ + 212, + 244, + 782, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metrics. The original evaluation metrics for cross-view geolocation primarily revolved around $\\mathrm{R}@\\mathcal{K}$ ( $K\\in \\{1,5,10,1\\% \\}$ ), which represents the probability of correctly identifying the matching image within the top $K$ retrieved reference images based on the query image [17,21,33,42,45]. We denote the performance of models on the original validation or test sets as $\\mathrm{R}@\\mathcal{K}_{\\mathrm{clean}}$ . For each corruption type $c$ and each severity level $s$ , we employ $\\mathrm{R}@\\mathcal{K}_{c,s}$ to gauge the performance of models under corruption conditions. Finally, by averaging across all corruption types and severity levels, we compute the average performance of model $\\mathrm{R}@\\mathcal{K}_{\\mathrm{cor}}$ under corruption conditions. In this context, $\\mathcal{C}$ represents the set of corruptions in evaluation. Additionally, we calculate the Relative Corruption Error (RCE) by measuring the percentage of performance degradation [4]. A higher RCE implies poorer robustness.", + "bbox": [ + 212, + 366, + 784, + 547 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nR @ K _ {\\text {c o r}} = \\frac {1}{| \\mathcal {C} |} \\sum_ {c \\in \\mathcal {C}} \\frac {1}{5} \\sum_ {s = 1} ^ {5} R @ K _ {c, s} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 561, + 784, + 595 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R C E} _ {c, s} = \\frac {\\mathrm {R @ K} _ {\\text {c l e a n}} - \\mathrm {R @ K} _ {c , s}}{\\mathrm {R @ K} _ {\\text {c l e a n}}}; \\mathrm {R C E} = \\frac {\\mathrm {R @ K} _ {\\text {c l e a n}} - \\mathrm {R @ K} _ {\\text {c o r}}}{\\mathrm {R @ K} _ {\\text {c l e a n}}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 609, + 784, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Comprehensive Corruption Robustness Benchmarks", + "text_level": 1, + "bbox": [ + 215, + 660, + 689, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we introduce comprehensive corruption robustness benchmarks. Prior efforts [4,16,25] often generated separate test sets for each corruption type and its corresponding 5 severity levels. While this approach allows for a finer-grained evaluation of each corruption type's impact on the model, it imposes significant storage and computational costs (80 evaluation subsets). Recognizing that real-world concerns may prioritize a model's overall performance against diverse corruptions, for each corruption type and severity level, we aggregate all forms of corruption into a single evaluation set, creating comprehensive benchmarks. The suffix \"-C-ALL\" designates all comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL).", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/809b3f9149d68c4a21693a59694559f4a543cef90dd9f70e45fb1680b53bb2b7.jpg", + "image_caption": [ + "Fig. 4: The proposed fine-grained and comprehensive robustness benchmarks. Each corruption category encompasses 5 severity levels. The primary distinction between fine-grained and comprehensive robustness benchmarks lies in whether separate evaluation subsets are created for each corruption category and severity level." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 233 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CVUSA-C-ALL and CVACT_val-C-ALL. For the 8,884-pair test set of the CVUSA dataset and the 8,884-pair validation set of the CVACT_val dataset, we designed comprehensive corruption robustness benchmarks. These benchmarks are denoted as CVUSA-C-ALL and CVACT_val-C-ALL, respectively. Each dataset contains 8,884 corrupted images and systematically covers all corruption types and severity levels.", + "bbox": [ + 212, + 330, + 782, + 421 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CVACT_test-C-ALL. The CVACT dataset additionally provides 92,802 pairs of images for testing purposes (referred to as CVACT_test), showcasing dense coverage of urban areas. Similar to CVUSA-C-ALL and CVACT_val-C-ALL, we create a comprehensive corruption robustness benchmark from CVACT_test, named CVACT_test-C-ALL, containing a total of 92,802 corrupted images, covering all corruption types and severity levels.", + "bbox": [ + 212, + 422, + 821, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. Since the comprehensive corruption robustness benchmarks represent independent subsets for testing, the recall accuracy $(\\mathbb{R}@\\mathcal{K}_{\\mathrm{all}}, K_{\\mathrm{all}} \\in \\{1, 5, 10, 1\\% \\})$ can be directly employed for evaluation.", + "bbox": [ + 214, + 515, + 784, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The comprehensive corruption robustness benchmarks serve as valuable tools for evaluating the overall robustness of models under all corruption scenarios, and they effectively reduce the cost of robustness evaluation.", + "bbox": [ + 212, + 561, + 784, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In summary, fine-grained robustness benchmarks provide a more detailed performance evaluation, while comprehensive robustness benchmarks provide an overall performance evaluation and reduce the evaluation cost. An illustrative diagram of the corruption robustness benchmarks is presented in Fig. 4.", + "bbox": [ + 212, + 607, + 784, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Benchmarking Results", + "text_level": 1, + "bbox": [ + 214, + 696, + 470, + 713 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have selected classical cross-view geo-localization models to benchmark on the proposed benchmarks. The experimental results on fine-grained corruption robustness benchmarks are shown in Section 5.1. Section 5.2 presents the experimental results on comprehensive corruption robustness benchmarks. In Section 5.3, we report the experimental results demonstrating the influence of stylization and histogram equalization on the robustness of models. Supplementary material provides additional experimental results.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Cross-view Geo-localization Models", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/099cc9235da12671bd9e825bf9e84e66dfed5cd85e53b7205caa6cef09e845ad.jpg", + "table_caption": [ + "Table 2: Experimental results of 8 cross-view geo-localization methods on the CVUSA-C benchmark. We report the R@1 performance of each method under different corruptions (obtained by averaging the 5 corruption severities), as well as the average R@1cor under all corruption types." + ], + "table_footnote": [], + "table_body": "
MethodCleanCVUSA-C
WeatherBlurNoiseDigital
SnowFrostFogBrightSpatterDefocusGlassMotionZoomGaussianShotImpulseSpeckleContrastPixelJPEG
CVM-Net [17]22.470.868.428.3713.756.111.064.811.470.231.821.181.282.324.756.896.23
OriCNN [21]40.797.366.517.5721.6922.0120.4626.1019.6010.3217.2413.9519.4014.097.9428.5127.27
SAFA [33]89.8419.3260.4267.6381.9649.8651.2480.5655.4911.4433.0428.5130.3737.5931.6788.0581.15
CVFT [35]61.438.0030.7947.4647.5427.6324.5544.9334.898.1721.8319.1920.5626.2538.2857.2547.11
DSM [34]91.9624.2464.4484.0882.4457.5862.4884.5266.0225.1549.5546.4048.8460.8372.1190.2085.56
L2LTR [42]94.0567.1985.0092.6491.6175.2488.3593.1389.3342.0781.3280.2982.8886.5486.3693.6490.56
TransGeo [45]94.0829.3969.5070.8985.0164.2680.9792.1685.9640.9772.9570.2774.3283.9943.0193.7490.13
GeoDTR [43]95.4344.2084.9592.8093.5573.1482.6493.2976.8027.1968.4064.4568.5378.2874.8094.4590.20
", + "bbox": [ + 218, + 200, + 782, + 280 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Fine-grained Corruption Robustness Benchmarking Results", + "text_level": 1, + "bbox": [ + 215, + 313, + 751, + 328 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmarking Results on CVUSA-C Table 2 presents the results of 8 cross-view geo-localization methods on the fine-grained corruption robustness benchmark CVUSA-C. In general, the robustness of the model is positively correlated with its accuracy on clean data. These models with higher $\\mathrm{R@1_{clean}}$ , such as L2LTR [42], TransGeo [45], and GeoDTR [43], also achieve higher $\\mathrm{R@1_{cor}}$ which is understandable as different models show consistent performance degradation on corrupted images. It's worth noting that L2LTR outperforms TransGeo and GeoDTR in $\\mathrm{R@1_{cor}}$ , even though TransGeo and GeoDTR exhibit higher $\\mathrm{R@1_{clean}}$ . This observation differs from the results obtained in [18] and [4]. The discrepancy might be attributed to its deeper network architecture and more parameters, which make it more robust to input variations. This finding emphasizes the importance of considering the robustness of the model independently, especially when its performance on clean data is high, as higher clean performance does not necessarily indicate stronger robustness. In Fig. 5a, we plot the RCE of models under various corruption types (according to major categories). Based on our experimental results, we derive the following insights.", + "bbox": [ + 212, + 351, + 787, + 592 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Impact of Corruption Types. Based on the results from Table 2 and Fig. 5a, we observed that snow, spatter, and zoom corruptions have a significant impact on the performance of various cross-view geo-localization models in the CVUSA-C benchmark, resulting in RCE values exceeding $28\\%$ for all models. Conversely, the effects of glass blur, JPEG, and pixelation on performance are relatively minor. These findings demonstrate the threats posed by adverse weather conditions and fast motion to cross-view geo-localization models. In contrast, most models exhibited less performance degradation under the influence of glass blur, JPEG compression, and pixelation, possibly due to similar corruptions being present in the training dataset, allowing the models to learn prior knowledge about these types of corruptions.", + "bbox": [ + 212, + 595, + 787, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Performance of Different Models. Among all the evaluated models, L2LTR [42] exhibited the most outstanding $\\mathrm{R@1_{cor}}$ performance. Additionally, we observed a synchronized growth trend between $\\mathrm{R@1_{clean}}$ and $\\mathrm{R@1_{cor}}$ for all methods except L2LTR. Notably, CVM-Net [17] demonstrated the weakest corruption robustness. All of the methods show an average performance degradation", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b33b3ecfa1fbd791481f8ad1851726969a82a86762fd6a60741c0783f74d16db.jpg", + "image_caption": [ + "(a) CVUSA-C" + ], + "image_footnote": [], + "bbox": [ + 223, + 148, + 485, + 295 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8056ba5dea64649a21ee3c4253f247f71e3f343212a589d3937012402131d9ef.jpg", + "image_caption": [ + "(b) CVACT_val-C", + "Fig. 5: The Relative Corruption Error (RCE) of different cross-view geo-localization models on CVUSA-C and CVACT_val-C benchmarks. The results are shown for each major category of corruptions, including Weather, Blur, Noise, and Digital, as well as the overall performance across all types of corruptions (Overall)." + ], + "image_footnote": [], + "bbox": [ + 516, + 148, + 779, + 296 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fac464576d046fac9d24cd92c45f47f30071eda3284080942206e0d89173bb07.jpg", + "table_caption": [ + "Table 3: Experimental results of 7 cross-view geo-localization methods on the CVACT_val-C benchmark. We report the R@1 performance of each method under different corruptions (obtained by averaging the 5 corruption severities), as well as the average R@1cor under all corruption types." + ], + "table_footnote": [], + "table_body": "
MethodCleanCVACT_val-C
WeatherBlurNoiseDigitalR@1cor
SnowFrostFogBrightSpatterDefocusGlassMotionZoomGaussianShotImpulseSpeckleContrastPixelJPEG
OriCNN [21]46.9613.946.133.7829.4540.5431.7139.9937.5824.8934.2432.2739.0133.284.5644.3842.5628.65
SAFA [33]81.0320.0331.6633.1966.9945.6039.8372.8749.864.6248.6643.6848.8251.6115.9176.9075.8345.38
CVFT [35]61.0515.0022.3242.5347.6037.2531.3053.8836.914.1035.6830.8036.3236.8431.7958.2157.9736.16
DSM [34]82.4931.9551.7070.4369.4852.3557.3580.1667.3815.3458.3453.0558.1863.0652.7981.7280.5558.99
L2LTR [42]84.8971.0377.9383.5081.1773.7883.9885.0784.0049.7982.2081.1982.9882.2379.1585.0783.4079.15
TransGeo [45]84.9547.6558.5132.9172.6767.1381.4384.8381.8036.3481.9680.8682.8483.0122.1884.9283.7467.68
GeoDTR [43]86.2148.2471.7483.2684.6061.3979.1185.5173.448.2675.4473.9977.0680.2355.4886.0185.1970.56
", + "bbox": [ + 218, + 454, + 782, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "of $24.24\\%$ on the CVUSA-C benchmark, with SAFA [33] degrading by up to $39.32\\%$ and L2LTR [42] degrading by a minimum of $11.17\\%$ .", + "bbox": [ + 212, + 555, + 784, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Benchmarking Results on CVACT_val-C Table 3 presents the results of 7 cross-view geo-localization methods on the fine-grained corruption robustness benchmark CVACT_val-C. Fig. 5b shows the RCE of models across various corruption types. Similar to the experimental results for the CVUSA-C benchmark, L2LTR [42] exhibits the best robustness. Meanwhile, snow, spatter, and zoom corruptions hurt the performance of the various models the most, and the various models have better robustness to glass blur, JPEG, and pixelation corruptions. However, it is worth noting that all methods show an average performance degradation of $20.14\\%$ on the CVACT_val-C benchmark.", + "bbox": [ + 212, + 612, + 785, + 747 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effects of different training sets. From the average results of all methods of CVUSA-C and CVACT_val-C benchmarks, the R@1Clean to R@1cor drop of CVACT_val-C benchmark is $4.1\\%$ less than CVUSA-C. We believe that this may be due to the fact that the CVACT dataset is collected in urban areas, while the CVUSA dataset is collected in the suburbs, and thus the CVACT dataset has richer scene information and so brings more robustness to the model.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Cross-view Geo-localization Models", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/16b7bd95ae18dcb257a22a4e18648185884aaf14f754783663dad2b614569a80.jpg", + "table_caption": [ + "Table 4: Experimental results of cross-view geo-localization methods on CVUSA-C-ALL, CVACT val-C-ALL, and CVACT test-C-ALL benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodCVUSA-C-ALLCVACT_val-C-ALLCVACT_test-C-ALL
R@1_allR@5_allR@10_allR@1%_allR@1_allR@5_allR@10_allR@1%_allR@1_allR@5_allR@10_allR@1%_all
CVM-Net [17]6.0916.0523.1452.51--------
OriCNN [21]9.3822.2630.0458.9915.3128.3135.2158.393.698.3311.0443.93
SAFA [33]63.6878.0882.8293.9156.7273.6078.5991.3231.1852.0658.6090.41
CVFT [35]41.0564.0172.6491.3745.6966.4572.9788.3822.8243.4851.0788.99
DSM [34]75.2786.2689.4295.0770.0482.8185.8693.5147.1368.4173.5293.18
L2LTR [42]87.9395.4597.0199.0182.1393.3494.9398.1057.2082.5987.2398.09
TransGeo [45]82.7291.9594.0397.9274.0486.1989.1094.9852.1874.3578.9995.03
GeoDTR [43]84.6493.2995.0198.2477.4088.9591.2895.9152.8778.8483.1795.84
", + "bbox": [ + 218, + 172, + 784, + 268 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Comprehensive Corruption Robustness Benchmarking Results", + "text_level": 1, + "bbox": [ + 215, + 297, + 774, + 313 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The performances of different models on comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) are shown in Table 4. It is clear from the experimental results that the performance degradation is closely and positively correlated with the original performance when evaluated using the comprehensive corruption robustness benchmarks, except for L2LTR [42]. The L2LTR exhibits the highest level of robustness, albeit at the expense of increased computational cost and a greater number of trainable parameters. The results obtained by all methods for the comprehensive robustness benchmarks follow the same trend as the $\\mathrm{R@1_{cor}}$ results for the fine-grained robustness benchmarks. This demonstrates the benefit of the comprehensive robustness benchmarks when considering overall robustness; they allow the best robustness model to be selected with only a small evaluation cost (as in the normal evaluation process).", + "bbox": [ + 212, + 323, + 787, + 521 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Robustness Enhancement Results", + "text_level": 1, + "bbox": [ + 215, + 542, + 539, + 556 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/57895e5a55403f00188a7d8444dbe8f3e15fed7c7da3e933c8f0f50ac7e6a2a4.jpg", + "image_caption": [ + "(a) SAFA" + ], + "image_footnote": [], + "bbox": [ + 230, + 595, + 403, + 695 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4711de5cc3c6a7cc362cb0b02d9eb62158b1b860f39521be035b466e11ab41cf.jpg", + "image_caption": [ + "(b) L2LTR", + "Fig. 6: Stylization improves the robustness of SAFA [33], L2LTR [42], and Trans-Geo [45] on the CVUSA-C benchmark, with each severity level representing the average across all 16 corruption types. Severity $= 0$ corresponds to clean images for testing. The Standard denotes the original, unaltered training data, while Stylization denotes training exclusively on images subjected to stylization. Stylization & Standard denotes stylization and origin training data are equally interleaved during the training process. Notably, the three different training strategies require identical training complexity, and experimental configurations and model structures remain consistent throughout." + ], + "image_footnote": [], + "bbox": [ + 415, + 595, + 589, + 695 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/99dfa3e77744bb355b93bb7f00cc3901dc7179f8f34f9001e16a92b7d4633f39.jpg", + "image_caption": [ + "(c) TransGeo" + ], + "image_footnote": [], + "bbox": [ + 601, + 595, + 774, + 695 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/15963e5e485b97e7b892d4289cfacd5d97780620f7484d24e2a8ae436f10815a.jpg", + "image_caption": [ + "(a) SAFA" + ], + "image_footnote": [], + "bbox": [ + 230, + 146, + 401, + 244 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ae609c7562a9bdfcf1728217fea9fd0cd05d6857879f93772a9709f58ee63efb.jpg", + "image_caption": [ + "(b) L2LTR", + "Fig. 7: CLAHE improves the robustness of SAFA [33], L2LTR [42], and TransGeo [45] on the CVUSA-C benchmark, with each severity level representing the average across all 16 corruption types. Notations are similar to Fig. 6." + ], + "image_footnote": [], + "bbox": [ + 415, + 147, + 589, + 244 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b2fcb0cd50e1e7bf810a2e42c2f8df73a0f3a6c192658f78fcfa896a7cc2aa82.jpg", + "image_caption": [ + "(c) TransGeo" + ], + "image_footnote": [], + "bbox": [ + 601, + 147, + 774, + 244 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Stylization for Robustness Enhancement We investigated whether the use of stylization enhances the robustness of various cross-view geo-localization models. We evaluated the performance of 3 classical cross-view geo-localization models on the CVUSA-C benchmark using 3 different training strategies, as depicted in Section 3.3 and Fig. 3. Fig. 6 shows the experimental results.", + "bbox": [ + 212, + 348, + 782, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Training on stylized images indeed resulted in stronger robustness compared to models trained only on clean images, showing less performance degradation as the severity of corruption increased. However, the performance on the original clean images (severity $= 0$ ) is noticeably poorer. This can be attributed to the fact that stylized data alters the distribution of the original data. By equally combining stylized and clean data during training (Stylization & Standard), a trade-off between clean and corrupted performance was achieved, leading to both high performance akin to the standard data on clean instances and significantly improved performance on corrupted data. Results from Fig. 6 indicate that both Stylization and Stylization & Standard training enhanced $\\mathrm{R@1_{cor}}$ under corruption, especially when corruption level is greater (e.g., Severity $= 4 / 5$ ).", + "bbox": [ + 212, + 424, + 784, + 589 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Histogram Equalization for Robustness Enhancement We have further examined the impact of employing histogram equalization on the robustness of cross-view geo-localization models. In our study, we employ Contrast Limited Adaptive Histogram Equalization (CLAHE) [29] to enhance the robustness of existing methods. Fig. 7 shows the experimental results.", + "bbox": [ + 212, + 613, + 784, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It is evident that training models only on data that has undergone CLAHE does not greatly improve their robustness. On the other hand, our training strategy 3, which combines CLAHE and clean data in equal proportions, can somewhat increase the robustness of these models, though the improvement is less than that of stylization-based methods. It's also noteworthy that different models respond differently to the same training set of data. This emphasizes the importance of model robustness and the challenge of attracting researchers to invest in model robustness.", + "bbox": [ + 212, + 688, + 784, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We must emphasize that how to improve model robustness is still an open topic, which requires the efforts of more researchers.", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Cross-view Geo-localization Models", + "bbox": [ + 272, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Further Discussion", + "text_level": 1, + "bbox": [ + 215, + 143, + 433, + 160 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Real-world corruption can be much more extensive and complex. Although we cannot enumerate all possible real-world corruptions, we have systematically designed 16 corruption types, each with 5 severity levels. These types can provide a large amount of rich data to support machine learning and computer vision research. Compared to images that must be captured in the real world, synthetic imagery can provide highly controllable and diverse data for research by adjusting weather conditions, severity, and scene types as needed. At the same time, it circumvents the security risks that may be associated with capturing images of real scenes under adverse weather conditions, as well as the privacy and ethical issues that may be involved in capturing in specific environments. Therefore, we strongly believe that our benchmarks can serve as a practical testbed to perform controllable robustness evaluation, which in turn will advance research on cross-view geo-localization robustness.", + "bbox": [ + 212, + 178, + 787, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our focus is on evaluating how well cross-view geo-localization models withstand the corruption of input images. However, the corruption robustness benchmarks we present can be applied to more research topics. For example, robustness evaluations and enhancements on cross-view image synthesis [30,31,37,38], cross-view camera pose estimation [19, 32, 41] and autonomous driving [14, 26] can also benefit from our benchmarks.", + "bbox": [ + 212, + 375, + 787, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 489, + 359, + 506 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper systematically investigates the impact of corruption data on cross-view geo-localization models, which is a challenge previously overlooked in the context of cross-view geo-localization studies. In contrast to \"clean\" data, \"corrupted\" data is more common and closer to the realistic world, and exhaustive experiments have shown these models invariably experience significant performance degradation in the face of corrupted data. This further validates the necessity and importance of robustness benchmarking for cross-view geo-localization. To address this crucial issue and track future developments, we propose two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) for the cross-view geo-localization task. Extensive experiments are conducted to evaluate existing classical methods on these corruption robustness benchmarks, revealing new insights. Furthermore, we introduce two simple techniques (stylization and histogram equalization) and the training strategy to effectively enhance robustness, without requiring any model adjustments or introducing additional training complexity. We hope that the corruption robustness benchmarks, in-depth analysis, and insightful findings presented in this paper will raise awareness within the community regarding the robustness of cross-view geo-localization models and contribute to enhancing their robustness to challenges in complex real-world environments in the future.", + "bbox": [ + 212, + 522, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62072318 and Grant U22A2079, in part by the Key Project of Department of Education of Guangdong Province under Grant 2023ZDZX1016, and in part by Shenzhen Science and Technology Program under Grant 20220810142553001.", + "bbox": [ + 212, + 180, + 787, + 256 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 282, + 321, + 299 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016)", + "2. Chaabane, M., Gueguen, L., Trabelsi, A., Beveridge, R., O'Hara, S.: End-to-end learning improves static object geo-localization from video. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2063-2072 (2021)", + "3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)", + "4. Dong, Y., Kang, C., Zhang, J., Zhu, Z., Wang, Y., Yang, X., Su, H., Wei, X., Zhu, J.: Benchmarking robustness of 3d object detection to common corruptions in autonomous driving. arXiv preprint arXiv:2303.11040 (2023)", + "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)", + "6. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)", + "7. Friedland, G., Vinyals, O., Darrell, T.: Multimodal location estimation. In: Proceedings of the 18th ACM international conference on Multimedia. pp. 1245-1252 (2010)", + "8. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3855-3863 (2017)", + "9. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016)", + "0. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)", + "11. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. International Conference on Learning Representations, International Conference on Learning Representations (Sep 2018)" + ], + "bbox": [ + 217, + 316, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Cross-view Geo-localization Models", + "bbox": [ + 271, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Geirhos, R., Temme, C.R., Rauber, J., Schütt, H.H., Bethge, M., Wichmann, F.A.: Generalisation in humans and deep neural networks. Advances in neural information processing systems 31 (2018)", + "13. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 580-587 (2014)", + "14. Hane, C., Heng, L., Lee, G.H., Fraundorfer, F., Furgale, P., Sattler, T., Pollefeys, M.: 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image and Vision Computing 68, 14-27 (2017)", + "15. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence 33(12), 2341-2353 (2010)", + "16. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)", + "17. Hu, S., Feng, M., Nguyen, R.M., Lee, G.H.: Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7258-7267 (2018)", + "18. Kamann, C., Rother, C.: Benchmarking the robustness of semantic segmentation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8828-8838 (2020)", + "19. Lentsch, T., Xia, Z., Caesar, H., Kooij, J.F.: Slicematch: Geometry-guided aggregation for cross-view pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17225-17234 (2023)", + "20. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2736-2744 (2016)", + "21. Liu, L., Li, H.: Lending orientation to neural networks for cross-view geolocation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5624-5633 (2019)", + "22. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing 27(6), 3064-3073 (2018)", + "23. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)", + "24. McManus, C., Churchill, W., Maddern, W., Stewart, A.D., Newman, P.: Shady dealings: Robust, long-term visual localisation using illumination invariance. In: 2014 IEEE international conference on robotics and automation (ICRA). pp. 901-906. IEEE (2014)", + "25. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019)", + "26. Middelberg, S., Sattler, T., Untzelmann, O., Kobbelt, L.: Scalable 6-dof localization on mobile devices. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13. pp. 268-283. Springer (2014)", + "27. Mithun, N.C., Minhas, K.S., Chiu, H.P., Oskiper, T., Sizintsev, M., Samarasekera, S., Kumar, R.: Cross-view visual geo-localization for outdoor augmented reality. In: 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR). pp. 493-502. IEEE (2023)" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A., Greer, T., ter Haar Romeny, B., Zimmerman, J.B., Zuiderveld, K.: Adaptive histogram equalization and its variations. Computer Vision, Graphics, and Image Processing p. 355-368 (Aug 1987). https://doi.org/10.1016/s0734-189x(87)80186-x, http://dx.doi.org/10.1016/s0734-189x(87)80186-x", + "29. Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A., Greer, T., ter Haar Romeny, B., Zimmerman, J.B., Zuiderveld, K.: Adaptive histogram equalization and its variations. Computer vision, graphics, and image processing 39(3), 355-368 (1987)", + "30. Regmi, K., Borji, A.: Cross-view image synthesis using conditional gans. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 3501-3510 (2018)", + "31. Shi, Y., Campbell, D., Yu, X., Li, H.: Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 10009-10022 (2022)", + "32. Shi, Y., Li, H.: Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17010-17020 (2022)", + "33. Shi, Y., Liu, L., Yu, X., Li, H.: Spatial-aware feature aggregation for image based cross-view geo-localization. Advances in Neural Information Processing Systems 32 (2019)", + "34. Shi, Y., Yu, X., Campbell, D., Li, H.: Where am i looking at? joint location and orientation estimation by cross-view matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4064-4072 (2020)", + "35. Shi, Y., Yu, X., Liu, L., Zhang, T., Li, H.: Optimal feature transport for cross-view image geo-localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11990-11997 (2020)", + "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (Jan 2015)", + "37. Tang, H., Xu, D., Yan, Y., Torr, P.H., Sebe, N.: Local class-specific and global image-level generative adversarial networks for semantic-guided scene generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7870-7879 (2020)", + "38. Token, A., Zhou, Q., Maximov, M., Leal-Taixe, L.: Coming down to earth: Satellite-to-street view synthesis for geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6488-6497 (2021)", + "39. Vasiljevic, I., Chakrabarti, A., Shakhnarovich, G.: Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760 (2016)", + "40. Workman, S., Souvenir, R., Jacobs, N.: Wide-area image geolocation with aerial reference imagery. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3961-3969 (2015)", + "41. Xia, Z., Booij, O., Manfredi, M., Kooij, J.F.: Visual cross-view metric localization with dense uncertainty estimates. In: European Conference on Computer Vision. pp. 90-106. Springer (2022)", + "42. Yang, H., Lu, X., Zhu, Y.: Cross-view geo-localization with layer-to-layer transformer. Advances in Neural Information Processing Systems 34, 29009-29020 (2021)", + "43. Zhang, X., Li, X., Sultani, W., Zhou, Y., Wshah, S.: Cross-view geo-localization via learning disentangled geometric layout correspondence. arXiv preprint arXiv:2212.04074 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Cross-view Geo-localization Models", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)", + "45. Zhu, S., Shah, M., Chen, C.: Transgeo: Transformer is all you need for cross-view image geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1162-1171 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 231 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Q. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_model.json b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb0f696db2e983ea6e3b10f59bb1f39a8932b565 --- /dev/null +++ b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_model.json @@ -0,0 +1,2458 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.141, + 0.756, + 0.186 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.213, + 0.654, + 0.229 + ], + "angle": 0, + "content": "Qingwang Zhang and Yingying Zhu" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.24, + 0.758, + 0.255 + ], + "angle": 0, + "content": "College of Computer Science and Software Engineering, Shenzhen University," + }, + { + "type": "text", + "bbox": [ + 0.445, + 0.255, + 0.558, + 0.268 + ], + "angle": 0, + "content": "Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.269, + 0.702, + 0.282 + ], + "angle": 0, + "content": "zhangqingwang2022@email.szu.edu.cn，zhuyy@szu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.283, + 0.692, + 0.296 + ], + "angle": 0, + "content": "https://zqwlearning.github.io/CrossViewRobustness" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.333, + 0.744, + 0.665 + ], + "angle": 0, + "content": "Abstract. Cross-view geo-localization serves as a viable alternative to providing geographical location information when GPS signals are unstable or unavailable by matching ground images with geo-tagged aerial image databases. While significant progress has been made on some common benchmarks like CVUSA and CVACT, there remains a lack of comprehensive consideration for robustness against real-world environmental challenges such as adverse weather or sensor noise. This deficiency poses a significant challenge for deploying this technology in safety-critical domains like autonomous driving and robot navigation. To the best of our knowledge, there is currently no specialized benchmark for evaluating the robustness of cross-view geo-localization. To comprehensively and fairly evaluate the robustness of cross-view geo-localization models in real-world scenarios, we introduce 16 common types of data corruption. By synthesizing these corruptions on public datasets, we establish two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL), covering approximately 1.5 million corrupted images. Subsequently, we conduct large-scale experiments on various cross-view geo-localization models to evaluate their robustness in corrupted environments and derive novel insights. Finally, we explore two data augmentation strategies as potential solutions to enhance model robustness. Combined with the training strategies proposed, these approaches effectively enhance the robustness of multiple models." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.679, + 0.731, + 0.693 + ], + "angle": 0, + "content": "Keywords: Cross-view geo-localization \\(\\cdot\\) Benchmarking \\(\\cdot\\) Robustness" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.719, + 0.377, + 0.735 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.794, + 0.841 + ], + "angle": 0, + "content": "Geo-localization is fundamental for autonomous driving [26], robot navigation [24], 3D reconstruction [14], augmented reality [27], and road maintenance [2], etc. However, in the realistic scenario, a GPS-degraded or GPS-denied environment is not uncommon, including urban canyon effects, forest areas, mountainous terrain, and interference or jamming regions. The absence of geo-tags in image metadata, often resulting from GPS signal obstruction, signal attenuation, or" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Fig. 1: Existing cross-view geo-localization models fail when the ground query image is corrupted. When the query image is clean, almost all existing methods retrieve the ground truth correctly (marked by the red box and the red circle). Once the image is corrupted, almost all existing methods have missed the correct result. Incorrect retrieve and location results are marked in green and blue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.488 + ], + "angle": 0, + "content": "lack of a camera localization module, leads to only \\(4.3\\%\\) of Internet images being geo-tagged [7]. Cross-view geo-localization can determine the location where a ground-level image is taken by comparing it with a geo-tagged database of aerial images (e.g., satellite images), which can serve as an alternative way or an effective complement to provide geo-location information in the case of GPS-degraded or GPS-denied. Therefore, cross-view geo-localization has recently garnered considerable attention due to its wide range of practical applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.777 + ], + "angle": 0, + "content": "Two favorite benchmarks have been introduced to evaluate cross-view geo-localization performance, e.g., CVUSA [40] and CVACT [21]. Existing methods are all evaluated on both datasets and state-of-the-art approaches even achieve near-perfect results [42,43,45]. Nevertheless, these benchmarks only present ideal settings and do not reflect real-world scenarios. Almost all their ground images were collected on clean days, with little coverage in conditions such as rain, snow, and sensor noise corruption. Inevitably, due to the significant differences between the distributions of the training data and the real-world data, the performance of current data-driven models degrades drastically or even fails when the models trained on \"clean\" data are applied to \"corrupted\" data. Fig. 1 shows examples of localization failures that can result from the ground-level query under different weather conditions. This poses a significant challenge to applying this task to safety-critical applications such as autonomous driving. Thus, prior to deployment, it has become imperative to create an easy-to-use corruption robustness benchmark and to comprehensively evaluate the robustness of existing models with respect to corrupted data. To the best of our knowledge, we first create cross-view geo-localization robustness benchmarks and provide an exhaustive assessment of the robustness of existing mainstream cross-view geo-localization models when faced with corrupted data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In the cross-view geo-localization task setup, the aerial images are cached by the system in advance with high-definition and stable characteristics, while the ground-view images are collected in real-time, which is more flexible and versatile. Therefore, in this paper, we only corrupt the ground images to consider" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.329, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.189, + 0.29, + 0.198 + ], + "angle": 0, + "content": "Clean" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.144, + 0.444, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.189, + 0.427, + 0.198 + ], + "angle": 0, + "content": "Weather - Snow" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.144, + 0.557, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.189, + 0.537, + 0.198 + ], + "angle": 0, + "content": "Blur - Defocus" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.144, + 0.671, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.189, + 0.658, + 0.198 + ], + "angle": 0, + "content": "Noise-Gaussian" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.144, + 0.785, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.189, + 0.765, + 0.198 + ], + "angle": 0, + "content": "Digital-JPEG" + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.206, + 0.655, + 0.217 + ], + "angle": 0, + "content": "(a) Comparison between clean and corrupted images." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.218, + 0.33, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.262, + 0.304, + 0.271 + ], + "angle": 0, + "content": "Severity \\(= 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.218, + 0.444, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.262, + 0.419, + 0.271 + ], + "angle": 0, + "content": "Severity \\(= 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.217, + 0.557, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.262, + 0.532, + 0.271 + ], + "angle": 0, + "content": "Severity \\(= 3\\)" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.217, + 0.671, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.262, + 0.645, + 0.271 + ], + "angle": 0, + "content": "Severity \\(= 4\\)" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.217, + 0.785, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.262, + 0.758, + 0.271 + ], + "angle": 0, + "content": "Severity \\(= 5\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.277, + 0.706, + 0.29 + ], + "angle": 0, + "content": "(b) Comparison of different severity levels, taking snow as an example." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.301, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Fig. 2: Various types and severities of common corruption in our robustness benchmarks, after cropping are provided (best viewed when zoomed in on the screen). Visualizations of all corruption types can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.6 + ], + "angle": 0, + "content": "the robustness of the existing models, which is more practical while alleviating the evaluation cost. We systematically design 16 common types of corruption that are commonly found in the cross-view geo-localization task and provide a comprehensive evaluation of the corruption robustness of existing models. These corruptions include Weather, Blur, Noise, and Digital, covering a wide range of real-world situations. Furthermore, each type of corruption has 5 severity levels, totaling 80 different types of corruption involved. We have applied these corruptions on CVUSA [40] and CVACT [21] to create two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL), which cover a total of about 1.5 million corrupted images. Fig. 2 shows some examples of corrupted images. We expect these datasets to serve as common benchmarks for comprehensively and fairly evaluating the corruption robustness of cross-view geo-localization models and to facilitate future research to advance the cross-view geo-localization task." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.602, + 0.7, + 0.617 + ], + "angle": 0, + "content": "The contributions made to this work are summed up as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.628, + 0.785, + 0.703 + ], + "angle": 0, + "content": "- To the best of our knowledge, we have benchmarked the robustness of state-of-the-art cross-view geo-localization models against real-world data corruption challenges for the first time. We have generated about 1.5 million corrupted images based on the CVUSA and CVACT datasets to establish benchmarks for assessing the robustness of cross-view geo-localization models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.793 + ], + "angle": 0, + "content": "- Based on the benchmark study, we have several new insights: (1) in most cases, the clean performance \\((\\mathrm{R}@\\mathcal{K}_{\\mathrm{clean}})\\) of a model is positively, but not absolutely, correlated with its robustness; (2) snow, spatter, and zoom blur more significantly affect the robustness of various models compared to other corruptions; (3) models trained on more intricate scenarios (e.g., CVACT) exhibit better robustness." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- Introducing stylization and histogram equalization as data augmentation techniques, along with our proposed training strategy, significantly enhances the robustness of various cross-view geo-localization models." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.628, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.181, + 0.492, + 0.196 + ], + "angle": 0, + "content": "2.1 Cross-view Geo-localization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.789, + 0.543 + ], + "angle": 0, + "content": "Cross-view geo-localization has witnessed significant advancements in recent years. Initially, [40] introduced Convolutional Neural Networks (CNN) to the cross-view matching task, resulting in notable performance improvements. Subsequently, [17] employed a VGG [36] backbone network with two branches, combined with the NetVlad [1], and proposed the weighted soft-margin ranking loss, achieving state-of-the-art performance. To further enhance network performance, [21] emphasized the importance of orientation in cross-view geo-localization and devised a method to provide orientation information to the neural network explicitly. [33] adopted spatial-aware feature aggregation modules to aggregate information-rich and diverse feature maps, while introducing polar transformation for pre-processing to narrow the geometric gap between center-aligned satellites and ground-level images. Recently, [42] explored the integration of Vision Transformer [5] in cross-view geo-localization, proposing a novel layer-to-layer Transformer with self-cross attention mechanism that highlighted the significance of considering global dependencies to reduce visual ambiguity. [45] introduced an attention-guided pure Transformer approach, which further improved the resolution of satellite images through additional training, thus advancing the performance of the task. [43] introduced a novel geometric layout extraction module that explicitly decouples geometric information from original features and proposed two types of data augmentation methods. However, all these methods ignore the performance of the model on corrupted images, i.e., robustness, which is the focus of this paper." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.466, + 0.584 + ], + "angle": 0, + "content": "2.2 Robustness Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.79, + 0.844 + ], + "angle": 0, + "content": "Robustness benchmarks are essential in the field of computer vision to enhance the stability and reliability of computer vision systems when facing uncertainties and noise. For instance, IMAGENET-C and IMAGENET-P were proposed to evaluate robustness in classification tasks [16]. In the context of autonomous driving, PASCAL-C, COCO-C, and Cityscapes-C were introduced as benchmarks for evaluating the robustness of object detection tasks [13, 25]. Similarly, robustness benchmarks for semantic segmentation [23] were established using PASCAL VOC 2012 [6], Cityscapes [3], and ADE20K [18, 44]. Furthermore, KITTI-C, nuScenes-C, and Waymo-C were devised as robustness benchmarks for 3D object detection [4]. These benchmarks aid in evaluating the robustness of models under various complex environmental conditions, thus facilitating their application in the real world. However, to the best of our knowledge, no researcher has proposed relevant robustness benchmarks specifically for cross-view geo-localization. Therefore, this paper addresses this gap by introducing the benchmarks CVUSA-C, CVACT_val-C, CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL, thus filling this void in the field." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.473, + 0.16 + ], + "angle": 0, + "content": "2.3 Robustness Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.788, + 0.44 + ], + "angle": 0, + "content": "Several strategies have been proposed to address the impact of corruption in computer vision. For example, [20] and [8] present methods that utilize patch-based priors for background and rain layers, along with a deep CNN to effectively remove rain streaks from images. In a similar vein, [15] introduced a simple yet powerful technique employing the dark channel before eliminating haze from single-input images. Moreover, [22] devised the DesnowNet, a multi-stage network specifically designed to remove snow particles from images. However, a limitation of these methods lies in their reliance on specific designs tailored to particular types of corruption. This hinders their ability to generalize and handle other types of corruption effectively. As an alternative, some approaches have sought to enhance model performance through data augmentation, incorporating corrupted data during the training process. Although fine-tuning on specific corruption data has shown promise in boosting performance for those particular corruption types [12, 39] discovered that fine-tuning on one type of corruption often struggles to generalize to other corruption types. How to enhance robustness remains an open question. In this work, we employ the stylization [11] and histogram equalization [28] techniques to the train set, and we observe improvements in robustness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.461, + 0.661, + 0.477 + ], + "angle": 0, + "content": "3 Robustness in Cross-view Geo-localization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.49, + 0.438, + 0.504 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Robustness. First, we consider a set of ground-to-aerial image pairs \\(\\{\\mathbf{I}_i^g,\\mathbf{I}_i^a\\}_{i = 1}^N\\) where \\(N\\) represents the number of image pairs. The superscripts \\(g\\) and \\(a\\) denote ground and aerial views, respectively. For a ground query image indexed as \\(q\\), we assume the existence of ground image encoder \\(f_{g}\\) and aerial image encoder \\(f_{a}\\), which have been trained on samples from the distribution \\(\\mathcal{D}\\). We let \\(\\mathbb{P}_{\\mathcal{C}}(c)\\) approximate the frequency of real-world corruption. Existing models mostly evaluate performance when samples are drawn from the distribution \\(\\mathcal{D}\\), denoted as \\(\\mathbb{P}_{(\\mathbf{I}^g,\\mathbf{I}^a)\\sim \\mathcal{D}}(d(f_g(\\mathbf{I}_q^g),f_a(\\mathbf{I}_q^a)) < \\{d(f_g(\\mathbf{I}_q^g),f_a(\\mathbf{I}_i^a))|\\forall i\\in \\{1,\\dots,N\\},i\\neq q\\})\\), \\(d(\\cdot ,\\cdot)\\) representing the \\(L_{2}\\) distance. However, in practical deployments, systems often need to operate on low-quality or corrupted images. Hence, we construct an evaluation of the corruption robustness of the model, denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.671, + 0.773, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {c \\sim \\mathcal {C}} \\left[ \\mathbb {P} _ {\\left(\\mathbf {I} ^ {g}, \\mathbf {I} ^ {a}\\right) \\sim \\mathcal {D}} \\left(d \\left(f _ {g} \\left(c \\left(\\mathbf {I} _ {q} ^ {g}\\right)\\right), f _ {a} \\left(\\mathbf {I} _ {q} ^ {a}\\right)\\right) < \\left\\{d \\left(f _ {g} \\left(c \\left(\\mathbf {I} _ {q} ^ {g}\\right)\\right), f _ {a} \\left(\\mathbf {I} _ {i} ^ {a}\\right)\\right) | \\forall i \\in \\{1, \\dots , N \\}, i \\neq q \\right\\}\\right) \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.787, + 0.769 + ], + "angle": 0, + "content": "In this work, in order to approximate the real-world corruption \\(\\mathcal{C}\\), we designed a set of common corruptions (Weather, Blur, Noise, and Digital) in the cross-view geo-localization task. The types of corruption already provide a good overview of corruption that street view images may face. From these common corruptions, we create our cross-view geo-localization robustness evaluation benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.412, + 0.804 + ], + "angle": 0, + "content": "3.2 Image Corruption" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Common corruptions are categorized based on [16], and ground-level image corruptions are divided into four major categories with a total of 16 subcategories." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.451 + ], + "angle": 0, + "content": "The first major category is weather-related corruptions, which encompasses snow, frost, fog, brightness, and spatter. The second major category consists of blur-related corruptions, including defocus, glass, motion, and zoom blur. Defocus blur occurs when an image is out of focus, while glass blur arises when images are captured through frosted glass windows of sensors. Motion and zoom blurring occur in scenes with rapid camera movement or quick approaches toward objects, and they are particularly prone to occur in image collections from vehicular devices. The third major category involves noise-related corruptions, which cover Gaussian, shot, impulse, and speckle noise. Gaussian noise emerges under low-light conditions, attributed to the discrete nature of light and causing electronic noise. Impulse noise results from bit errors, resembling the color analog of salt-and-pepper noise. Speckle noise, which occurs due to light interference, leads to the appearance of bright and dark speckles. Lastly, the fourth category includes digital corruptions, including contrast, pixelation, and JPEG corruption. Contrast corruption is dependent on lighting conditions and the colors of the objects during capture. Pixelation occurs during upscaling low-resolution images, and JPEG corruption arises when images undergo lossy compression, introducing compression artifacts. To better simulate real-world corruption, each type of corruption also contains 5 severity levels. We present visualizations of various types and severities of common corruption in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.469, + 0.553, + 0.485 + ], + "angle": 0, + "content": "3.3 Robustness Enhancement Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.569 + ], + "angle": 0, + "content": "To enhance the robustness of existing models against corruption, we investigate two data augmentation approaches: stylization and histogram equalization. By applying either stylization or histogram equalization to the training data and employing our proposed training strategy, we effectively bolstered the corruption robustness of multiple cross-view geo-localization models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.645 + ], + "angle": 0, + "content": "Stylization. Style transfer, introduced by [9], merges the content and style of two different images to generate a novel image that retains the content of the original image while adopting the style of the target image, which is also known as stylization. In the context of image classification and object detection tasks, style transfer has demonstrated its effectiveness in enhancing robustness [10]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Histogram Equalization. Histogram equalization is a fundamental technique used for enhancing image contrast. It operates by redistributing pixel intensity values within an image, stretching or compressing the brightness range, thus achieving a more uniform distribution of pixel intensity values across the entire brightness range of the image. In this study, we employ Contrast Limited Adaptive Histogram Equalization (CLAHE) [29] to enhance the robustness of existing methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Training Strategy. In this paper, we apply Stylization / CLAHE to the cross-view geo-localization training set, testing 3 settings: (1) training using the standard (raw) training data; (2) replacing all training images with Stylization / CLAHE images, thereby eliminating the use of standard images during training; (3) having Stylization / CLAHE images participate in the training process alongside the standard images with equal probability at each iteration. Fig. 3" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.785, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.281, + 0.785, + 0.351 + ], + "angle": 0, + "content": "Fig. 3: Visualization of Stylization / CLAHE applied to the CVUSA training set. The illustration depicts standard images (middle row), stylization images (top row), and CLAHE images (bottom row). The rounded rectangles on both sides represent different training strategies. The red dashed boxes indicate the style images sourced from https://www.kaggle.com/c/painter-by-numbers/. Inspired by [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.41 + ], + "angle": 0, + "content": "shows examples of applying Stylization / CLAHE, and it illustrates 3 training strategies with the same training complexity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.601, + 0.449 + ], + "angle": 0, + "content": "4 Corruption Robustness Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.794, + 0.537 + ], + "angle": 0, + "content": "We create two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) from the classic cross-view geo-localization datasets CVUSA and CVACT. See Table 1 for more detailed information on these benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.56, + 0.669, + 0.575 + ], + "angle": 0, + "content": "4.1 Fine-grained Corruption Robustness Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.689 + ], + "angle": 0, + "content": "We create fine-grained robustness benchmarks comprising 16 corruption types, each corresponding to 5 severity levels denoted by integers from 1 to 5, where higher numbers indicate more severe corruption. Consequently, one single original ground query image can yield 80 \\((16 \\times 5)\\) corrupted images. In the fine-grained corruption robustness benchmarks, each type of corrupted image forms an independent evaluation subset with no overlap. The suffix \"-C\" designates all fine-grained robustness benchmarks (CVUSA-C and CVACT_val-C)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.786, + 0.841 + ], + "angle": 0, + "content": "CVUSA-C. The CVUSA [40] dataset is one of the earliest cross-view geo-localization datasets, primarily collected from suburban areas in the United States. It encompasses a total of 35,532 pairs of ground-aerial images for training and 8,884 pairs for testing. We have employed the CVUSA test set to generate the fine-grained corruption robustness benchmark, CVUSA-C. To be specific, CVUSA-C comprises all types and severities of image corruptions, each residing independently within individual subsets of CVUSA-C. Given that there are 16 distinct corruption types and 5 levels of corruption severity, CVUSA-C effectively comprises 80 evaluation subsets. Each of these subsets contains 8,884 ground images for testing, resulting in an aggregate of 710,720 corrupted images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.145, + 0.776, + 0.159 + ], + "angle": 0, + "content": "Table 1: Detailed information on the proposed corruption robustness benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.16, + 0.785, + 0.216 + ], + "angle": 0, + "content": "
Fine-grainedComprehensive
CVUSA-CCVACT_val-CCVUSA-C-ALLCVACT_val-C-ALLCVACT_test-C-ALL
Number of original validation / test ground images8,8848,8848,8848,88492,802
Whether or not evaluation subsets are generated for each corruption✓✓✗✗✗
Whether all corruptions are included in an independent subset✗✗✓✓✓
Number of validation / test ground images for our benchmark8,884 × 16 × 58,884 × 16 × 58,8848,88492,802
Storage space~39 GB~178 GB~0.5 GB~2 GB~21 GB
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.245, + 0.784, + 0.365 + ], + "angle": 0, + "content": "CVACT_val-C. The CVACT [21] dataset was collected by [21] and densely covers an urban area of Australia (Canberra). Similar to the CVUSA dataset, this dataset comprises 35,532 pairs of ground-aerial images for training and 8,884 image pairs for validation (referred to as CVACT_val). We create a fine-grained robustness benchmark from CVACT_val, as CVUSA does. The benchmark includes various types and severities, resulting in a total of 80 validation subsets. These subsets consist of 710,720 corrupted images and are collectively referred to as CVACT_val-C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.549 + ], + "angle": 0, + "content": "Evaluation Metrics. The original evaluation metrics for cross-view geolocation primarily revolved around \\(\\mathrm{R}@\\mathcal{K}\\) (\\(K\\in \\{1,5,10,1\\% \\}\\)), which represents the probability of correctly identifying the matching image within the top \\(K\\) retrieved reference images based on the query image [17,21,33,42,45]. We denote the performance of models on the original validation or test sets as \\(\\mathrm{R}@\\mathcal{K}_{\\mathrm{clean}}\\). For each corruption type \\(c\\) and each severity level \\(s\\), we employ \\(\\mathrm{R}@\\mathcal{K}_{c,s}\\) to gauge the performance of models under corruption conditions. Finally, by averaging across all corruption types and severity levels, we compute the average performance of model \\(\\mathrm{R}@\\mathcal{K}_{\\mathrm{cor}}\\) under corruption conditions. In this context, \\(\\mathcal{C}\\) represents the set of corruptions in evaluation. Additionally, we calculate the Relative Corruption Error (RCE) by measuring the percentage of performance degradation [4]. A higher RCE implies poorer robustness." + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.562, + 0.785, + 0.597 + ], + "angle": 0, + "content": "\\[\nR @ K _ {\\text {c o r}} = \\frac {1}{| \\mathcal {C} |} \\sum_ {c \\in \\mathcal {C}} \\frac {1}{5} \\sum_ {s = 1} ^ {5} R @ K _ {c, s} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.611, + 0.785, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R C E} _ {c, s} = \\frac {\\mathrm {R @ K} _ {\\text {c l e a n}} - \\mathrm {R @ K} _ {c , s}}{\\mathrm {R @ K} _ {\\text {c l e a n}}}; \\mathrm {R C E} = \\frac {\\mathrm {R @ K} _ {\\text {c l e a n}} - \\mathrm {R @ K} _ {\\text {c o r}}}{\\mathrm {R @ K} _ {\\text {c l e a n}}} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.691, + 0.677 + ], + "angle": 0, + "content": "4.2 Comprehensive Corruption Robustness Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Furthermore, we introduce comprehensive corruption robustness benchmarks. Prior efforts [4,16,25] often generated separate test sets for each corruption type and its corresponding 5 severity levels. While this approach allows for a finer-grained evaluation of each corruption type's impact on the model, it imposes significant storage and computational costs (80 evaluation subsets). Recognizing that real-world concerns may prioritize a model's overall performance against diverse corruptions, for each corruption type and severity level, we aggregate all forms of corruption into a single evaluation set, creating comprehensive benchmarks. The suffix \"-C-ALL\" designates all comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.3 + ], + "angle": 0, + "content": "Fig. 4: The proposed fine-grained and comprehensive robustness benchmarks. Each corruption category encompasses 5 severity levels. The primary distinction between fine-grained and comprehensive robustness benchmarks lies in whether separate evaluation subsets are created for each corruption category and severity level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.784, + 0.422 + ], + "angle": 0, + "content": "CVUSA-C-ALL and CVACT_val-C-ALL. For the 8,884-pair test set of the CVUSA dataset and the 8,884-pair validation set of the CVACT_val dataset, we designed comprehensive corruption robustness benchmarks. These benchmarks are denoted as CVUSA-C-ALL and CVACT_val-C-ALL, respectively. Each dataset contains 8,884 corrupted images and systematically covers all corruption types and severity levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.822, + 0.514 + ], + "angle": 0, + "content": "CVACT_test-C-ALL. The CVACT dataset additionally provides 92,802 pairs of images for testing purposes (referred to as CVACT_test), showcasing dense coverage of urban areas. Similar to CVUSA-C-ALL and CVACT_val-C-ALL, we create a comprehensive corruption robustness benchmark from CVACT_test, named CVACT_test-C-ALL, containing a total of 92,802 corrupted images, covering all corruption types and severity levels." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.516, + 0.785, + 0.56 + ], + "angle": 0, + "content": "Evaluation Metrics. Since the comprehensive corruption robustness benchmarks represent independent subsets for testing, the recall accuracy \\((\\mathbb{R}@\\mathcal{K}_{\\mathrm{all}}, K_{\\mathrm{all}} \\in \\{1, 5, 10, 1\\% \\})\\) can be directly employed for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.785, + 0.606 + ], + "angle": 0, + "content": "The comprehensive corruption robustness benchmarks serve as valuable tools for evaluating the overall robustness of models under all corruption scenarios, and they effectively reduce the cost of robustness evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.785, + 0.669 + ], + "angle": 0, + "content": "In summary, fine-grained robustness benchmarks provide a more detailed performance evaluation, while comprehensive robustness benchmarks provide an overall performance evaluation and reduce the evaluation cost. An illustrative diagram of the corruption robustness benchmarks is presented in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.471, + 0.714 + ], + "angle": 0, + "content": "5 Benchmarking Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We have selected classical cross-view geo-localization models to benchmark on the proposed benchmarks. The experimental results on fine-grained corruption robustness benchmarks are shown in Section 5.1. Section 5.2 presents the experimental results on comprehensive corruption robustness benchmarks. In Section 5.3, we report the experimental results demonstrating the influence of stylization and histogram equalization on the robustness of models. Supplementary material provides additional experimental results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.2 + ], + "angle": 0, + "content": "Table 2: Experimental results of 8 cross-view geo-localization methods on the CVUSA-C benchmark. We report the R@1 performance of each method under different corruptions (obtained by averaging the 5 corruption severities), as well as the average R@1cor under all corruption types." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.202, + 0.784, + 0.281 + ], + "angle": 0, + "content": "
MethodCleanCVUSA-C
WeatherBlurNoiseDigital
SnowFrostFogBrightSpatterDefocusGlassMotionZoomGaussianShotImpulseSpeckleContrastPixelJPEG
CVM-Net [17]22.470.868.428.3713.756.111.064.811.470.231.821.181.282.324.756.896.23
OriCNN [21]40.797.366.517.5721.6922.0120.4626.1019.6010.3217.2413.9519.4014.097.9428.5127.27
SAFA [33]89.8419.3260.4267.6381.9649.8651.2480.5655.4911.4433.0428.5130.3737.5931.6788.0581.15
CVFT [35]61.438.0030.7947.4647.5427.6324.5544.9334.898.1721.8319.1920.5626.2538.2857.2547.11
DSM [34]91.9624.2464.4484.0882.4457.5862.4884.5266.0225.1549.5546.4048.8460.8372.1190.2085.56
L2LTR [42]94.0567.1985.0092.6491.6175.2488.3593.1389.3342.0781.3280.2982.8886.5486.3693.6490.56
TransGeo [45]94.0829.3969.5070.8985.0164.2680.9792.1685.9640.9772.9570.2774.3283.9943.0193.7490.13
GeoDTR [43]95.4344.2084.9592.8093.5573.1482.6493.2976.8027.1968.4064.4568.5378.2874.8094.4590.20
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.314, + 0.753, + 0.329 + ], + "angle": 0, + "content": "5.1 Fine-grained Corruption Robustness Benchmarking Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Benchmarking Results on CVUSA-C Table 2 presents the results of 8 cross-view geo-localization methods on the fine-grained corruption robustness benchmark CVUSA-C. In general, the robustness of the model is positively correlated with its accuracy on clean data. These models with higher \\( \\mathrm{R@1_{clean}} \\), such as L2LTR [42], TransGeo [45], and GeoDTR [43], also achieve higher \\( \\mathrm{R@1_{cor}} \\) which is understandable as different models show consistent performance degradation on corrupted images. It's worth noting that L2LTR outperforms TransGeo and GeoDTR in \\( \\mathrm{R@1_{cor}} \\), even though TransGeo and GeoDTR exhibit higher \\( \\mathrm{R@1_{clean}} \\). This observation differs from the results obtained in [18] and [4]. The discrepancy might be attributed to its deeper network architecture and more parameters, which make it more robust to input variations. This finding emphasizes the importance of considering the robustness of the model independently, especially when its performance on clean data is high, as higher clean performance does not necessarily indicate stronger robustness. In Fig. 5a, we plot the RCE of models under various corruption types (according to major categories). Based on our experimental results, we derive the following insights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Impact of Corruption Types. Based on the results from Table 2 and Fig. 5a, we observed that snow, spatter, and zoom corruptions have a significant impact on the performance of various cross-view geo-localization models in the CVUSA-C benchmark, resulting in RCE values exceeding \\(28\\%\\) for all models. Conversely, the effects of glass blur, JPEG, and pixelation on performance are relatively minor. These findings demonstrate the threats posed by adverse weather conditions and fast motion to cross-view geo-localization models. In contrast, most models exhibited less performance degradation under the influence of glass blur, JPEG compression, and pixelation, possibly due to similar corruptions being present in the training dataset, allowing the models to learn prior knowledge about these types of corruptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Performance of Different Models. Among all the evaluated models, L2LTR [42] exhibited the most outstanding \\(\\mathrm{R@1_{cor}}\\) performance. Additionally, we observed a synchronized growth trend between \\(\\mathrm{R@1_{clean}}\\) and \\(\\mathrm{R@1_{cor}}\\) for all methods except L2LTR. Notably, CVM-Net [17] demonstrated the weakest corruption robustness. All of the methods show an average performance degradation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.149, + 0.486, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.304, + 0.398, + 0.316 + ], + "angle": 0, + "content": "(a) CVUSA-C" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.149, + 0.78, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.304, + 0.705, + 0.316 + ], + "angle": 0, + "content": "(b) CVACT_val-C" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.326, + 0.785, + 0.385 + ], + "angle": 0, + "content": "Fig. 5: The Relative Corruption Error (RCE) of different cross-view geo-localization models on CVUSA-C and CVACT_val-C benchmarks. The results are shown for each major category of corruptions, including Weather, Blur, Noise, and Digital, as well as the overall performance across all types of corruptions (Overall)." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.398, + 0.784, + 0.454 + ], + "angle": 0, + "content": "Table 3: Experimental results of 7 cross-view geo-localization methods on the CVACT_val-C benchmark. We report the R@1 performance of each method under different corruptions (obtained by averaging the 5 corruption severities), as well as the average R@1cor under all corruption types." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.455, + 0.784, + 0.528 + ], + "angle": 0, + "content": "
MethodCleanCVACT_val-C
WeatherBlurNoiseDigitalR@1cor
SnowFrostFogBrightSpatterDefocusGlassMotionZoomGaussianShotImpulseSpeckleContrastPixelJPEG
OriCNN [21]46.9613.946.133.7829.4540.5431.7139.9937.5824.8934.2432.2739.0133.284.5644.3842.5628.65
SAFA [33]81.0320.0331.6633.1966.9945.6039.8372.8749.864.6248.6643.6848.8251.6115.9176.9075.8345.38
CVFT [35]61.0515.0022.3242.5347.6037.2531.3053.8836.914.1035.6830.8036.3236.8431.7958.2157.9736.16
DSM [34]82.4931.9551.7070.4369.4852.3557.3580.1667.3815.3458.3453.0558.1863.0652.7981.7280.5558.99
L2LTR [42]84.8971.0377.9383.5081.1773.7883.9885.0784.0049.7982.2081.1982.9882.2379.1585.0783.4079.15
TransGeo [45]84.9547.6558.5132.9172.6767.1381.4384.8381.8036.3481.9680.8682.8483.0122.1884.9283.7467.68
GeoDTR [43]86.2148.2471.7483.2684.6061.3979.1185.5173.448.2675.4473.9977.0680.2355.4886.0185.1970.56
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.785, + 0.587 + ], + "angle": 0, + "content": "of \\(24.24\\%\\) on the CVUSA-C benchmark, with SAFA [33] degrading by up to \\(39.32\\%\\) and L2LTR [42] degrading by a minimum of \\(11.17\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.748 + ], + "angle": 0, + "content": "Benchmarking Results on CVACT_val-C Table 3 presents the results of 7 cross-view geo-localization methods on the fine-grained corruption robustness benchmark CVACT_val-C. Fig. 5b shows the RCE of models across various corruption types. Similar to the experimental results for the CVUSA-C benchmark, L2LTR [42] exhibits the best robustness. Meanwhile, snow, spatter, and zoom corruptions hurt the performance of the various models the most, and the various models have better robustness to glass blur, JPEG, and pixelation corruptions. However, it is worth noting that all methods show an average performance degradation of \\(20.14\\%\\) on the CVACT_val-C benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Effects of different training sets. From the average results of all methods of CVUSA-C and CVACT_val-C benchmarks, the R@1Clean to R@1cor drop of CVACT_val-C benchmark is \\(4.1\\%\\) less than CVUSA-C. We believe that this may be due to the fact that the CVACT dataset is collected in urban areas, while the CVUSA dataset is collected in the suburbs, and thus the CVACT dataset has richer scene information and so brings more robustness to the model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 4: Experimental results of cross-view geo-localization methods on CVUSA-C-ALL, CVACT val-C-ALL, and CVACT test-C-ALL benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.173, + 0.785, + 0.27 + ], + "angle": 0, + "content": "
MethodCVUSA-C-ALLCVACT_val-C-ALLCVACT_test-C-ALL
R@1_allR@5_allR@10_allR@1%_allR@1_allR@5_allR@10_allR@1%_allR@1_allR@5_allR@10_allR@1%_all
CVM-Net [17]6.0916.0523.1452.51--------
OriCNN [21]9.3822.2630.0458.9915.3128.3135.2158.393.698.3311.0443.93
SAFA [33]63.6878.0882.8293.9156.7273.6078.5991.3231.1852.0658.6090.41
CVFT [35]41.0564.0172.6491.3745.6966.4572.9788.3822.8243.4851.0788.99
DSM [34]75.2786.2689.4295.0770.0482.8185.8693.5147.1368.4173.5293.18
L2LTR [42]87.9395.4597.0199.0182.1393.3494.9398.1057.2082.5987.2398.09
TransGeo [45]82.7291.9594.0397.9274.0486.1989.1094.9852.1874.3578.9995.03
GeoDTR [43]84.6493.2995.0198.2477.4088.9591.2895.9152.8778.8483.1795.84
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.298, + 0.776, + 0.314 + ], + "angle": 0, + "content": "5.2 Comprehensive Corruption Robustness Benchmarking Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.522 + ], + "angle": 0, + "content": "The performances of different models on comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) are shown in Table 4. It is clear from the experimental results that the performance degradation is closely and positively correlated with the original performance when evaluated using the comprehensive corruption robustness benchmarks, except for L2LTR [42]. The L2LTR exhibits the highest level of robustness, albeit at the expense of increased computational cost and a greater number of trainable parameters. The results obtained by all methods for the comprehensive robustness benchmarks follow the same trend as the \\(\\mathrm{R@1_{cor}}\\) results for the fine-grained robustness benchmarks. This demonstrates the benefit of the comprehensive robustness benchmarks when considering overall robustness; they allow the best robustness model to be selected with only a small evaluation cost (as in the normal evaluation process)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.54, + 0.558 + ], + "angle": 0, + "content": "5.3 Robustness Enhancement Results" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.596, + 0.404, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.702, + 0.346, + 0.714 + ], + "angle": 0, + "content": "(a) SAFA" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.597, + 0.59, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.701, + 0.536, + 0.713 + ], + "angle": 0, + "content": "(b) L2LTR" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.597, + 0.776, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.702, + 0.728, + 0.714 + ], + "angle": 0, + "content": "(c) TransGeo" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 6: Stylization improves the robustness of SAFA [33], L2LTR [42], and Trans-Geo [45] on the CVUSA-C benchmark, with each severity level representing the average across all 16 corruption types. Severity \\(= 0\\) corresponds to clean images for testing. The Standard denotes the original, unaltered training data, while Stylization denotes training exclusively on images subjected to stylization. Stylization & Standard denotes stylization and origin training data are equally interleaved during the training process. Notably, the three different training strategies require identical training complexity, and experimental configurations and model structures remain consistent throughout." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.147, + 0.403, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.252, + 0.346, + 0.264 + ], + "angle": 0, + "content": "(a) SAFA" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.148, + 0.59, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.253, + 0.536, + 0.264 + ], + "angle": 0, + "content": "(b) L2LTR" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.148, + 0.776, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.253, + 0.728, + 0.265 + ], + "angle": 0, + "content": "(c) TransGeo" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.785, + 0.318 + ], + "angle": 0, + "content": "Fig. 7: CLAHE improves the robustness of SAFA [33], L2LTR [42], and TransGeo [45] on the CVUSA-C benchmark, with each severity level representing the average across all 16 corruption types. Notations are similar to Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.784, + 0.424 + ], + "angle": 0, + "content": "Stylization for Robustness Enhancement We investigated whether the use of stylization enhances the robustness of various cross-view geo-localization models. We evaluated the performance of 3 classical cross-view geo-localization models on the CVUSA-C benchmark using 3 different training strategies, as depicted in Section 3.3 and Fig. 3. Fig. 6 shows the experimental results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.785, + 0.59 + ], + "angle": 0, + "content": "Training on stylized images indeed resulted in stronger robustness compared to models trained only on clean images, showing less performance degradation as the severity of corruption increased. However, the performance on the original clean images (severity \\(= 0\\)) is noticeably poorer. This can be attributed to the fact that stylized data alters the distribution of the original data. By equally combining stylized and clean data during training (Stylization & Standard), a trade-off between clean and corrupted performance was achieved, leading to both high performance akin to the standard data on clean instances and significantly improved performance on corrupted data. Results from Fig. 6 indicate that both Stylization and Stylization & Standard training enhanced \\(\\mathrm{R@1_{cor}}\\) under corruption, especially when corruption level is greater (e.g., Severity \\(= 4 / 5\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.688 + ], + "angle": 0, + "content": "Histogram Equalization for Robustness Enhancement We have further examined the impact of employing histogram equalization on the robustness of cross-view geo-localization models. In our study, we employ Contrast Limited Adaptive Histogram Equalization (CLAHE) [29] to enhance the robustness of existing methods. Fig. 7 shows the experimental results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.809 + ], + "angle": 0, + "content": "It is evident that training models only on data that has undergone CLAHE does not greatly improve their robustness. On the other hand, our training strategy 3, which combines CLAHE and clean data in equal proportions, can somewhat increase the robustness of these models, though the improvement is less than that of stylization-based methods. It's also noteworthy that different models respond differently to the same training set of data. This emphasizes the importance of model robustness and the challenge of attracting researchers to invest in model robustness." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We must emphasize that how to improve model robustness is still an open topic, which requires the efforts of more researchers." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.434, + 0.161 + ], + "angle": 0, + "content": "6 Further Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Real-world corruption can be much more extensive and complex. Although we cannot enumerate all possible real-world corruptions, we have systematically designed 16 corruption types, each with 5 severity levels. These types can provide a large amount of rich data to support machine learning and computer vision research. Compared to images that must be captured in the real world, synthetic imagery can provide highly controllable and diverse data for research by adjusting weather conditions, severity, and scene types as needed. At the same time, it circumvents the security risks that may be associated with capturing images of real scenes under adverse weather conditions, as well as the privacy and ethical issues that may be involved in capturing in specific environments. Therefore, we strongly believe that our benchmarks can serve as a practical testbed to perform controllable robustness evaluation, which in turn will advance research on cross-view geo-localization robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Our focus is on evaluating how well cross-view geo-localization models withstand the corruption of input images. However, the corruption robustness benchmarks we present can be applied to more research topics. For example, robustness evaluations and enhancements on cross-view image synthesis [30,31,37,38], cross-view camera pose estimation [19, 32, 41] and autonomous driving [14, 26] can also benefit from our benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.49, + 0.36, + 0.507 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.841 + ], + "angle": 0, + "content": "This paper systematically investigates the impact of corruption data on cross-view geo-localization models, which is a challenge previously overlooked in the context of cross-view geo-localization studies. In contrast to \"clean\" data, \"corrupted\" data is more common and closer to the realistic world, and exhaustive experiments have shown these models invariably experience significant performance degradation in the face of corrupted data. This further validates the necessity and importance of robustness benchmarking for cross-view geo-localization. To address this crucial issue and track future developments, we propose two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) for the cross-view geo-localization task. Extensive experiments are conducted to evaluate existing classical methods on these corruption robustness benchmarks, revealing new insights. Furthermore, we introduce two simple techniques (stylization and histogram equalization) and the training strategy to effectively enhance robustness, without requiring any model adjustments or introducing additional training complexity. We hope that the corruption robustness benchmarks, in-depth analysis, and insightful findings presented in this paper will raise awareness within the community regarding the robustness of cross-view geo-localization models and contribute to enhancing their robustness to challenges in complex real-world environments in the future." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.257 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62072318 and Grant U22A2079, in part by the Key Project of Department of Education of Guangdong Province under Grant 2023ZDZX1016, and in part by Shenzhen Science and Technology Program under Grant 20220810142553001." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.284, + 0.323, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.362 + ], + "angle": 0, + "content": "1. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.787, + 0.417 + ], + "angle": 0, + "content": "2. Chaabane, M., Gueguen, L., Trabelsi, A., Beveridge, R., O'Hara, S.: End-to-end learning improves static object geo-localization from video. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2063-2072 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.787, + 0.473 + ], + "angle": 0, + "content": "3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.787, + 0.516 + ], + "angle": 0, + "content": "4. Dong, Y., Kang, C., Zhang, J., Zhu, Z., Wang, Y., Yang, X., Su, H., Wei, X., Zhu, J.: Benchmarking robustness of 3d object detection to common corruptions in autonomous driving. arXiv preprint arXiv:2303.11040 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.787, + 0.572 + ], + "angle": 0, + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.573, + 0.787, + 0.614 + ], + "angle": 0, + "content": "6. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.615, + 0.787, + 0.657 + ], + "angle": 0, + "content": "7. Friedland, G., Vinyals, O., Darrell, T.: Multimodal location estimation. In: Proceedings of the 18th ACM international conference on Multimedia. pp. 1245-1252 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "8. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3855-3863 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.787, + 0.742 + ], + "angle": 0, + "content": "9. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.787, + 0.784 + ], + "angle": 0, + "content": "0. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "11. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. International Conference on Learning Representations, International Conference on Learning Representations (Sep 2018)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.318, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "12. Geirhos, R., Temme, C.R., Rauber, J., Schütt, H.H., Bethge, M., Wichmann, F.A.: Generalisation in humans and deep neural networks. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "13. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 580-587 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.285 + ], + "angle": 0, + "content": "14. Hane, C., Heng, L., Lee, G.H., Fraundorfer, F., Furgale, P., Sattler, T., Pollefeys, M.: 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image and Vision Computing 68, 14-27 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "15. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence 33(12), 2341-2353 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "16. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "17. Hu, S., Feng, M., Nguyen, R.M., Lee, G.H.: Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7258-7267 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.393, + 0.785, + 0.433 + ], + "angle": 0, + "content": "18. Kamann, C., Rother, C.: Benchmarking the robustness of semantic segmentation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8828-8838 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "19. Lentsch, T., Xia, Z., Caesar, H., Kooij, J.F.: Slicematch: Geometry-guided aggregation for cross-view pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17225-17234 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "20. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2736-2744 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "21. Liu, L., Li, H.: Lending orientation to neural networks for cross-view geolocation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5624-5633 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "22. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing 27(6), 3064-3073 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.595, + 0.785, + 0.636 + ], + "angle": 0, + "content": "23. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.636, + 0.785, + 0.69 + ], + "angle": 0, + "content": "24. McManus, C., Churchill, W., Maddern, W., Stewart, A.D., Newman, P.: Shady dealings: Robust, long-term visual localisation using illumination invariance. In: 2014 IEEE international conference on robotics and automation (ICRA). pp. 901-906. IEEE (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.69, + 0.785, + 0.732 + ], + "angle": 0, + "content": "25. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.732, + 0.785, + 0.785 + ], + "angle": 0, + "content": "26. Middelberg, S., Sattler, T., Untzelmann, O., Kobbelt, L.: Scalable 6-dof localization on mobile devices. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13. pp. 268-283. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "27. Mithun, N.C., Minhas, K.S., Chiu, H.P., Oskiper, T., Sizintsev, M., Samarasekera, S., Kumar, R.: Cross-view visual geo-localization for outdoor augmented reality. In: 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR). pp. 493-502. IEEE (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.217 + ], + "angle": 0, + "content": "28. Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A., Greer, T., ter Haar Romeny, B., Zimmerman, J.B., Zuiderveld, K.: Adaptive histogram equalization and its variations. Computer Vision, Graphics, and Image Processing p. 355-368 (Aug 1987). https://doi.org/10.1016/s0734-189x(87)80186-x, http://dx.doi.org/10.1016/s0734-189x(87)80186-x" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "29. Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A., Greer, T., ter Haar Romeny, B., Zimmerman, J.B., Zuiderveld, K.: Adaptive histogram equalization and its variations. Computer vision, graphics, and image processing 39(3), 355-368 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.313 + ], + "angle": 0, + "content": "30. Regmi, K., Borji, A.: Cross-view image synthesis using conditional gans. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 3501-3510 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "31. Shi, Y., Campbell, D., Yu, X., Li, H.: Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 10009-10022 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "32. Shi, Y., Li, H.: Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17010-17020 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "33. Shi, Y., Liu, L., Yu, X., Li, H.: Spatial-aware feature aggregation for image based cross-view geo-localization. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "34. Shi, Y., Yu, X., Campbell, D., Li, H.: Where am i looking at? joint location and orientation estimation by cross-view matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4064-4072 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "35. Shi, Y., Yu, X., Liu, L., Zhang, T., Li, H.: Optimal feature transport for cross-view image geo-localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11990-11997 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.549 + ], + "angle": 0, + "content": "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (Jan 2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "37. Tang, H., Xu, D., Yan, Y., Torr, P.H., Sebe, N.: Local class-specific and global image-level generative adversarial networks for semantic-guided scene generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7870-7879 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.647 + ], + "angle": 0, + "content": "38. Token, A., Zhou, Q., Maximov, M., Leal-Taixe, L.: Coming down to earth: Satellite-to-street view synthesis for geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6488-6497 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.648, + 0.787, + 0.674 + ], + "angle": 0, + "content": "39. Vasiljevic, I., Chakrabarti, A., Shakhnarovich, G.: Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "40. Workman, S., Souvenir, R., Jacobs, N.: Wide-area image geolocation with aerial reference imagery. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3961-3969 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "41. Xia, Z., Booij, O., Manfredi, M., Kooij, J.F.: Visual cross-view metric localization with dense uncertainty estimates. In: European Conference on Computer Vision. pp. 90-106. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "42. Yang, H., Lu, X., Zhu, Y.: Cross-view geo-localization with layer-to-layer transformer. Advances in Neural Information Processing Systems 34, 29009-29020 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "43. Zhang, X., Li, X., Sultani, W., Zhou, Y., Wshah, S.: Cross-view geo-localization via learning disentangled geometric layout correspondence. arXiv preprint arXiv:2212.04074 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "44. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "45. Zhu, S., Shah, M., Chen, C.: Transgeo: Transformer is all you need for cross-view image geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1162-1171 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.232 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_origin.pdf b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..006e0ffd743ff6d4aae5520a8757a8aee7c385a8 --- /dev/null +++ b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/4b1c92cb-d705-4da9-b833-74d509b5e4fe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:791cd8938a0d33c35b6e05d5db659a0f1da2ba59f4d7454b5c6be58ff684b317 +size 10358583 diff --git a/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/full.md b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0e49ac27b842bc6b06a95f8a5dfbbb33c9e8db17 --- /dev/null +++ b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/full.md @@ -0,0 +1,297 @@ +# Benchmarking the Robustness of Cross-view Geo-localization Models + +Qingwang Zhang and Yingying Zhu + +College of Computer Science and Software Engineering, Shenzhen University, + +Shenzhen, China + +zhangqingwang2022@email.szu.edu.cn，zhuyy@szu.edu.cn + +https://zqwlearning.github.io/CrossViewRobustness + +Abstract. Cross-view geo-localization serves as a viable alternative to providing geographical location information when GPS signals are unstable or unavailable by matching ground images with geo-tagged aerial image databases. While significant progress has been made on some common benchmarks like CVUSA and CVACT, there remains a lack of comprehensive consideration for robustness against real-world environmental challenges such as adverse weather or sensor noise. This deficiency poses a significant challenge for deploying this technology in safety-critical domains like autonomous driving and robot navigation. To the best of our knowledge, there is currently no specialized benchmark for evaluating the robustness of cross-view geo-localization. To comprehensively and fairly evaluate the robustness of cross-view geo-localization models in real-world scenarios, we introduce 16 common types of data corruption. By synthesizing these corruptions on public datasets, we establish two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL), covering approximately 1.5 million corrupted images. Subsequently, we conduct large-scale experiments on various cross-view geo-localization models to evaluate their robustness in corrupted environments and derive novel insights. Finally, we explore two data augmentation strategies as potential solutions to enhance model robustness. Combined with the training strategies proposed, these approaches effectively enhance the robustness of multiple models. + +Keywords: Cross-view geo-localization $\cdot$ Benchmarking $\cdot$ Robustness + +# 1 Introduction + +Geo-localization is fundamental for autonomous driving [26], robot navigation [24], 3D reconstruction [14], augmented reality [27], and road maintenance [2], etc. However, in the realistic scenario, a GPS-degraded or GPS-denied environment is not uncommon, including urban canyon effects, forest areas, mountainous terrain, and interference or jamming regions. The absence of geo-tags in image metadata, often resulting from GPS signal obstruction, signal attenuation, or + +![](images/06bd3fd4b0605c8f789f4ebd564a7bd8d03c092fb68b5f03881678912ad5b12c.jpg) +Fig. 1: Existing cross-view geo-localization models fail when the ground query image is corrupted. When the query image is clean, almost all existing methods retrieve the ground truth correctly (marked by the red box and the red circle). Once the image is corrupted, almost all existing methods have missed the correct result. Incorrect retrieve and location results are marked in green and blue. + +lack of a camera localization module, leads to only $4.3\%$ of Internet images being geo-tagged [7]. Cross-view geo-localization can determine the location where a ground-level image is taken by comparing it with a geo-tagged database of aerial images (e.g., satellite images), which can serve as an alternative way or an effective complement to provide geo-location information in the case of GPS-degraded or GPS-denied. Therefore, cross-view geo-localization has recently garnered considerable attention due to its wide range of practical applications. + +Two favorite benchmarks have been introduced to evaluate cross-view geo-localization performance, e.g., CVUSA [40] and CVACT [21]. Existing methods are all evaluated on both datasets and state-of-the-art approaches even achieve near-perfect results [42,43,45]. Nevertheless, these benchmarks only present ideal settings and do not reflect real-world scenarios. Almost all their ground images were collected on clean days, with little coverage in conditions such as rain, snow, and sensor noise corruption. Inevitably, due to the significant differences between the distributions of the training data and the real-world data, the performance of current data-driven models degrades drastically or even fails when the models trained on "clean" data are applied to "corrupted" data. Fig. 1 shows examples of localization failures that can result from the ground-level query under different weather conditions. This poses a significant challenge to applying this task to safety-critical applications such as autonomous driving. Thus, prior to deployment, it has become imperative to create an easy-to-use corruption robustness benchmark and to comprehensively evaluate the robustness of existing models with respect to corrupted data. To the best of our knowledge, we first create cross-view geo-localization robustness benchmarks and provide an exhaustive assessment of the robustness of existing mainstream cross-view geo-localization models when faced with corrupted data. + +In the cross-view geo-localization task setup, the aerial images are cached by the system in advance with high-definition and stable characteristics, while the ground-view images are collected in real-time, which is more flexible and versatile. Therefore, in this paper, we only corrupt the ground images to consider + +![](images/16fcf319507002768594b158dcce7f6092c42045997bab15964a51a7b60c7406.jpg) +Clean + +![](images/86a29d270ec50067d1d63210cde1cab99d587fa047bcad0105ed588710224b93.jpg) +Weather - Snow + +![](images/8be61a441409db8a515cfc624ed746383bdbb1d963a241a47462ac7f5fdff23d.jpg) +Blur - Defocus + +![](images/04c907550e4dbe1e9d4bec543738ad329a3ad1385e785af42128e4012e887666.jpg) +Noise-Gaussian + +![](images/3b8089f5e83a3868d7a34237f1ec89e5c5f4395f410ab109a3744533afbb1924.jpg) +Digital-JPEG + +![](images/93bb8548180c491d9b84d1f3d77a634d93ccdea4f83982a79f49fb821d3a868d.jpg) +Severity $= 1$ +(b) Comparison of different severity levels, taking snow as an example. +Fig. 2: Various types and severities of common corruption in our robustness benchmarks, after cropping are provided (best viewed when zoomed in on the screen). Visualizations of all corruption types can be found in the supplementary material. + +![](images/53542b972cfd266f41ebd48dd936501018af1b0da302fe523729f00e45d853e0.jpg) +Severity $= 2$ + +![](images/60f90392f7d73aa66daa8e25e0e469a25d8aca5722912552a7372c7d9d07d3e1.jpg) +(a) Comparison between clean and corrupted images. +Severity $= 3$ + +![](images/fb0a4cf89e82329d1a452b8ea15aa9537e1455000858c51c412b0588ae8cb366.jpg) +Severity $= 4$ + +![](images/0212ff984b351a754bf813d509db881005f0dc4a8d5efa7283d309c178dda7dc.jpg) +Severity $= 5$ + +the robustness of the existing models, which is more practical while alleviating the evaluation cost. We systematically design 16 common types of corruption that are commonly found in the cross-view geo-localization task and provide a comprehensive evaluation of the corruption robustness of existing models. These corruptions include Weather, Blur, Noise, and Digital, covering a wide range of real-world situations. Furthermore, each type of corruption has 5 severity levels, totaling 80 different types of corruption involved. We have applied these corruptions on CVUSA [40] and CVACT [21] to create two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL), which cover a total of about 1.5 million corrupted images. Fig. 2 shows some examples of corrupted images. We expect these datasets to serve as common benchmarks for comprehensively and fairly evaluating the corruption robustness of cross-view geo-localization models and to facilitate future research to advance the cross-view geo-localization task. + +The contributions made to this work are summed up as follows: + +- To the best of our knowledge, we have benchmarked the robustness of state-of-the-art cross-view geo-localization models against real-world data corruption challenges for the first time. We have generated about 1.5 million corrupted images based on the CVUSA and CVACT datasets to establish benchmarks for assessing the robustness of cross-view geo-localization models. +- Based on the benchmark study, we have several new insights: (1) in most cases, the clean performance $(\mathrm{R}@\mathcal{K}_{\mathrm{clean}})$ of a model is positively, but not absolutely, correlated with its robustness; (2) snow, spatter, and zoom blur more significantly affect the robustness of various models compared to other corruptions; (3) models trained on more intricate scenarios (e.g., CVACT) exhibit better robustness. +- Introducing stylization and histogram equalization as data augmentation techniques, along with our proposed training strategy, significantly enhances the robustness of various cross-view geo-localization models. + +# 2 Related Work + +# 2.1 Cross-view Geo-localization + +Cross-view geo-localization has witnessed significant advancements in recent years. Initially, [40] introduced Convolutional Neural Networks (CNN) to the cross-view matching task, resulting in notable performance improvements. Subsequently, [17] employed a VGG [36] backbone network with two branches, combined with the NetVlad [1], and proposed the weighted soft-margin ranking loss, achieving state-of-the-art performance. To further enhance network performance, [21] emphasized the importance of orientation in cross-view geo-localization and devised a method to provide orientation information to the neural network explicitly. [33] adopted spatial-aware feature aggregation modules to aggregate information-rich and diverse feature maps, while introducing polar transformation for pre-processing to narrow the geometric gap between center-aligned satellites and ground-level images. Recently, [42] explored the integration of Vision Transformer [5] in cross-view geo-localization, proposing a novel layer-to-layer Transformer with self-cross attention mechanism that highlighted the significance of considering global dependencies to reduce visual ambiguity. [45] introduced an attention-guided pure Transformer approach, which further improved the resolution of satellite images through additional training, thus advancing the performance of the task. [43] introduced a novel geometric layout extraction module that explicitly decouples geometric information from original features and proposed two types of data augmentation methods. However, all these methods ignore the performance of the model on corrupted images, i.e., robustness, which is the focus of this paper. + +# 2.2 Robustness Benchmarks + +Robustness benchmarks are essential in the field of computer vision to enhance the stability and reliability of computer vision systems when facing uncertainties and noise. For instance, IMAGENET-C and IMAGENET-P were proposed to evaluate robustness in classification tasks [16]. In the context of autonomous driving, PASCAL-C, COCO-C, and Cityscapes-C were introduced as benchmarks for evaluating the robustness of object detection tasks [13, 25]. Similarly, robustness benchmarks for semantic segmentation [23] were established using PASCAL VOC 2012 [6], Cityscapes [3], and ADE20K [18, 44]. Furthermore, KITTI-C, nuScenes-C, and Waymo-C were devised as robustness benchmarks for 3D object detection [4]. These benchmarks aid in evaluating the robustness of models under various complex environmental conditions, thus facilitating their application in the real world. However, to the best of our knowledge, no researcher has proposed relevant robustness benchmarks specifically for cross-view geo-localization. Therefore, this paper addresses this gap by introducing the benchmarks CVUSA-C, CVACT_val-C, CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL, thus filling this void in the field. + +# 2.3 Robustness Enhancement + +Several strategies have been proposed to address the impact of corruption in computer vision. For example, [20] and [8] present methods that utilize patch-based priors for background and rain layers, along with a deep CNN to effectively remove rain streaks from images. In a similar vein, [15] introduced a simple yet powerful technique employing the dark channel before eliminating haze from single-input images. Moreover, [22] devised the DesnowNet, a multi-stage network specifically designed to remove snow particles from images. However, a limitation of these methods lies in their reliance on specific designs tailored to particular types of corruption. This hinders their ability to generalize and handle other types of corruption effectively. As an alternative, some approaches have sought to enhance model performance through data augmentation, incorporating corrupted data during the training process. Although fine-tuning on specific corruption data has shown promise in boosting performance for those particular corruption types [12, 39] discovered that fine-tuning on one type of corruption often struggles to generalize to other corruption types. How to enhance robustness remains an open question. In this work, we employ the stylization [11] and histogram equalization [28] techniques to the train set, and we observe improvements in robustness. + +# 3 Robustness in Cross-view Geo-localization + +# 3.1 Problem Formulation + +Robustness. First, we consider a set of ground-to-aerial image pairs $\{\mathbf{I}_i^g,\mathbf{I}_i^a\}_{i = 1}^N$ where $N$ represents the number of image pairs. The superscripts $g$ and $a$ denote ground and aerial views, respectively. For a ground query image indexed as $q$ , we assume the existence of ground image encoder $f_{g}$ and aerial image encoder $f_{a}$ , which have been trained on samples from the distribution $\mathcal{D}$ . We let $\mathbb{P}_{\mathcal{C}}(c)$ approximate the frequency of real-world corruption. Existing models mostly evaluate performance when samples are drawn from the distribution $\mathcal{D}$ , denoted as $\mathbb{P}_{(\mathbf{I}^g,\mathbf{I}^a)\sim \mathcal{D}}(d(f_g(\mathbf{I}_q^g),f_a(\mathbf{I}_q^a)) < \{d(f_g(\mathbf{I}_q^g),f_a(\mathbf{I}_i^a))|\forall i\in \{1,\dots,N\},i\neq q\})$ , $d(\cdot ,\cdot)$ representing the $L_{2}$ distance. However, in practical deployments, systems often need to operate on low-quality or corrupted images. Hence, we construct an evaluation of the corruption robustness of the model, denoted as: + +$$ +\mathbb {E} _ {c \sim \mathcal {C}} \left[ \mathbb {P} _ {\left(\mathbf {I} ^ {g}, \mathbf {I} ^ {a}\right) \sim \mathcal {D}} \left(d \left(f _ {g} \left(c \left(\mathbf {I} _ {q} ^ {g}\right)\right), f _ {a} \left(\mathbf {I} _ {q} ^ {a}\right)\right) < \left\{d \left(f _ {g} \left(c \left(\mathbf {I} _ {q} ^ {g}\right)\right), f _ {a} \left(\mathbf {I} _ {i} ^ {a}\right)\right) | \forall i \in \{1, \dots , N \}, i \neq q \right\}\right) \right] +$$ + +In this work, in order to approximate the real-world corruption $\mathcal{C}$ , we designed a set of common corruptions (Weather, Blur, Noise, and Digital) in the cross-view geo-localization task. The types of corruption already provide a good overview of corruption that street view images may face. From these common corruptions, we create our cross-view geo-localization robustness evaluation benchmarks. + +# 3.2 Image Corruption + +Common corruptions are categorized based on [16], and ground-level image corruptions are divided into four major categories with a total of 16 subcategories. + +The first major category is weather-related corruptions, which encompasses snow, frost, fog, brightness, and spatter. The second major category consists of blur-related corruptions, including defocus, glass, motion, and zoom blur. Defocus blur occurs when an image is out of focus, while glass blur arises when images are captured through frosted glass windows of sensors. Motion and zoom blurring occur in scenes with rapid camera movement or quick approaches toward objects, and they are particularly prone to occur in image collections from vehicular devices. The third major category involves noise-related corruptions, which cover Gaussian, shot, impulse, and speckle noise. Gaussian noise emerges under low-light conditions, attributed to the discrete nature of light and causing electronic noise. Impulse noise results from bit errors, resembling the color analog of salt-and-pepper noise. Speckle noise, which occurs due to light interference, leads to the appearance of bright and dark speckles. Lastly, the fourth category includes digital corruptions, including contrast, pixelation, and JPEG corruption. Contrast corruption is dependent on lighting conditions and the colors of the objects during capture. Pixelation occurs during upscaling low-resolution images, and JPEG corruption arises when images undergo lossy compression, introducing compression artifacts. To better simulate real-world corruption, each type of corruption also contains 5 severity levels. We present visualizations of various types and severities of common corruption in Fig. 2. + +# 3.3 Robustness Enhancement Methods + +To enhance the robustness of existing models against corruption, we investigate two data augmentation approaches: stylization and histogram equalization. By applying either stylization or histogram equalization to the training data and employing our proposed training strategy, we effectively bolstered the corruption robustness of multiple cross-view geo-localization models. + +Stylization. Style transfer, introduced by [9], merges the content and style of two different images to generate a novel image that retains the content of the original image while adopting the style of the target image, which is also known as stylization. In the context of image classification and object detection tasks, style transfer has demonstrated its effectiveness in enhancing robustness [10]. + +Histogram Equalization. Histogram equalization is a fundamental technique used for enhancing image contrast. It operates by redistributing pixel intensity values within an image, stretching or compressing the brightness range, thus achieving a more uniform distribution of pixel intensity values across the entire brightness range of the image. In this study, we employ Contrast Limited Adaptive Histogram Equalization (CLAHE) [29] to enhance the robustness of existing methods. + +Training Strategy. In this paper, we apply Stylization / CLAHE to the cross-view geo-localization training set, testing 3 settings: (1) training using the standard (raw) training data; (2) replacing all training images with Stylization / CLAHE images, thereby eliminating the use of standard images during training; (3) having Stylization / CLAHE images participate in the training process alongside the standard images with equal probability at each iteration. Fig. 3 + +![](images/8e94643ccb2b9ee0d9fd5fdad02f27ce85203c2e21478727f18368cc12886479.jpg) +Fig. 3: Visualization of Stylization / CLAHE applied to the CVUSA training set. The illustration depicts standard images (middle row), stylization images (top row), and CLAHE images (bottom row). The rounded rectangles on both sides represent different training strategies. The red dashed boxes indicate the style images sourced from https://www.kaggle.com/c/painter-by-numbers/. Inspired by [25]. + +shows examples of applying Stylization / CLAHE, and it illustrates 3 training strategies with the same training complexity. + +# 4 Corruption Robustness Benchmarks + +We create two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) from the classic cross-view geo-localization datasets CVUSA and CVACT. See Table 1 for more detailed information on these benchmarks. + +# 4.1 Fine-grained Corruption Robustness Benchmarks + +We create fine-grained robustness benchmarks comprising 16 corruption types, each corresponding to 5 severity levels denoted by integers from 1 to 5, where higher numbers indicate more severe corruption. Consequently, one single original ground query image can yield 80 $(16 \times 5)$ corrupted images. In the fine-grained corruption robustness benchmarks, each type of corrupted image forms an independent evaluation subset with no overlap. The suffix "-C" designates all fine-grained robustness benchmarks (CVUSA-C and CVACT_val-C). + +CVUSA-C. The CVUSA [40] dataset is one of the earliest cross-view geo-localization datasets, primarily collected from suburban areas in the United States. It encompasses a total of 35,532 pairs of ground-aerial images for training and 8,884 pairs for testing. We have employed the CVUSA test set to generate the fine-grained corruption robustness benchmark, CVUSA-C. To be specific, CVUSA-C comprises all types and severities of image corruptions, each residing independently within individual subsets of CVUSA-C. Given that there are 16 distinct corruption types and 5 levels of corruption severity, CVUSA-C effectively comprises 80 evaluation subsets. Each of these subsets contains 8,884 ground images for testing, resulting in an aggregate of 710,720 corrupted images. + +Table 1: Detailed information on the proposed corruption robustness benchmarks. + +
Fine-grainedComprehensive
CVUSA-CCVACT_val-CCVUSA-C-ALLCVACT_val-C-ALLCVACT_test-C-ALL
Number of original validation / test ground images8,8848,8848,8848,88492,802
Whether or not evaluation subsets are generated for each corruption✓✓✗✗✗
Whether all corruptions are included in an independent subset✗✗✓✓✓
Number of validation / test ground images for our benchmark8,884 × 16 × 58,884 × 16 × 58,8848,88492,802
Storage space~39 GB~178 GB~0.5 GB~2 GB~21 GB
+ +CVACT_val-C. The CVACT [21] dataset was collected by [21] and densely covers an urban area of Australia (Canberra). Similar to the CVUSA dataset, this dataset comprises 35,532 pairs of ground-aerial images for training and 8,884 image pairs for validation (referred to as CVACT_val). We create a fine-grained robustness benchmark from CVACT_val, as CVUSA does. The benchmark includes various types and severities, resulting in a total of 80 validation subsets. These subsets consist of 710,720 corrupted images and are collectively referred to as CVACT_val-C. + +Evaluation Metrics. The original evaluation metrics for cross-view geolocation primarily revolved around $\mathrm{R}@\mathcal{K}$ ( $K\in \{1,5,10,1\% \}$ ), which represents the probability of correctly identifying the matching image within the top $K$ retrieved reference images based on the query image [17,21,33,42,45]. We denote the performance of models on the original validation or test sets as $\mathrm{R}@\mathcal{K}_{\mathrm{clean}}$ . For each corruption type $c$ and each severity level $s$ , we employ $\mathrm{R}@\mathcal{K}_{c,s}$ to gauge the performance of models under corruption conditions. Finally, by averaging across all corruption types and severity levels, we compute the average performance of model $\mathrm{R}@\mathcal{K}_{\mathrm{cor}}$ under corruption conditions. In this context, $\mathcal{C}$ represents the set of corruptions in evaluation. Additionally, we calculate the Relative Corruption Error (RCE) by measuring the percentage of performance degradation [4]. A higher RCE implies poorer robustness. + +$$ +R @ K _ {\text {c o r}} = \frac {1}{| \mathcal {C} |} \sum_ {c \in \mathcal {C}} \frac {1}{5} \sum_ {s = 1} ^ {5} R @ K _ {c, s} \tag {1} +$$ + +$$ +\mathrm {R C E} _ {c, s} = \frac {\mathrm {R @ K} _ {\text {c l e a n}} - \mathrm {R @ K} _ {c , s}}{\mathrm {R @ K} _ {\text {c l e a n}}}; \mathrm {R C E} = \frac {\mathrm {R @ K} _ {\text {c l e a n}} - \mathrm {R @ K} _ {\text {c o r}}}{\mathrm {R @ K} _ {\text {c l e a n}}} \tag {2} +$$ + +# 4.2 Comprehensive Corruption Robustness Benchmarks + +Furthermore, we introduce comprehensive corruption robustness benchmarks. Prior efforts [4,16,25] often generated separate test sets for each corruption type and its corresponding 5 severity levels. While this approach allows for a finer-grained evaluation of each corruption type's impact on the model, it imposes significant storage and computational costs (80 evaluation subsets). Recognizing that real-world concerns may prioritize a model's overall performance against diverse corruptions, for each corruption type and severity level, we aggregate all forms of corruption into a single evaluation set, creating comprehensive benchmarks. The suffix "-C-ALL" designates all comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL). + +![](images/809b3f9149d68c4a21693a59694559f4a543cef90dd9f70e45fb1680b53bb2b7.jpg) +Fig. 4: The proposed fine-grained and comprehensive robustness benchmarks. Each corruption category encompasses 5 severity levels. The primary distinction between fine-grained and comprehensive robustness benchmarks lies in whether separate evaluation subsets are created for each corruption category and severity level. + +CVUSA-C-ALL and CVACT_val-C-ALL. For the 8,884-pair test set of the CVUSA dataset and the 8,884-pair validation set of the CVACT_val dataset, we designed comprehensive corruption robustness benchmarks. These benchmarks are denoted as CVUSA-C-ALL and CVACT_val-C-ALL, respectively. Each dataset contains 8,884 corrupted images and systematically covers all corruption types and severity levels. + +CVACT_test-C-ALL. The CVACT dataset additionally provides 92,802 pairs of images for testing purposes (referred to as CVACT_test), showcasing dense coverage of urban areas. Similar to CVUSA-C-ALL and CVACT_val-C-ALL, we create a comprehensive corruption robustness benchmark from CVACT_test, named CVACT_test-C-ALL, containing a total of 92,802 corrupted images, covering all corruption types and severity levels. + +Evaluation Metrics. Since the comprehensive corruption robustness benchmarks represent independent subsets for testing, the recall accuracy $(\mathbb{R}@\mathcal{K}_{\mathrm{all}}, K_{\mathrm{all}} \in \{1, 5, 10, 1\% \})$ can be directly employed for evaluation. + +The comprehensive corruption robustness benchmarks serve as valuable tools for evaluating the overall robustness of models under all corruption scenarios, and they effectively reduce the cost of robustness evaluation. + +In summary, fine-grained robustness benchmarks provide a more detailed performance evaluation, while comprehensive robustness benchmarks provide an overall performance evaluation and reduce the evaluation cost. An illustrative diagram of the corruption robustness benchmarks is presented in Fig. 4. + +# 5 Benchmarking Results + +We have selected classical cross-view geo-localization models to benchmark on the proposed benchmarks. The experimental results on fine-grained corruption robustness benchmarks are shown in Section 5.1. Section 5.2 presents the experimental results on comprehensive corruption robustness benchmarks. In Section 5.3, we report the experimental results demonstrating the influence of stylization and histogram equalization on the robustness of models. Supplementary material provides additional experimental results. + +Table 2: Experimental results of 8 cross-view geo-localization methods on the CVUSA-C benchmark. We report the R@1 performance of each method under different corruptions (obtained by averaging the 5 corruption severities), as well as the average R@1cor under all corruption types. + +
MethodCleanCVUSA-C
WeatherBlurNoiseDigital
SnowFrostFogBrightSpatterDefocusGlassMotionZoomGaussianShotImpulseSpeckleContrastPixelJPEG
CVM-Net [17]22.470.868.428.3713.756.111.064.811.470.231.821.181.282.324.756.896.23
OriCNN [21]40.797.366.517.5721.6922.0120.4626.1019.6010.3217.2413.9519.4014.097.9428.5127.27
SAFA [33]89.8419.3260.4267.6381.9649.8651.2480.5655.4911.4433.0428.5130.3737.5931.6788.0581.15
CVFT [35]61.438.0030.7947.4647.5427.6324.5544.9334.898.1721.8319.1920.5626.2538.2857.2547.11
DSM [34]91.9624.2464.4484.0882.4457.5862.4884.5266.0225.1549.5546.4048.8460.8372.1190.2085.56
L2LTR [42]94.0567.1985.0092.6491.6175.2488.3593.1389.3342.0781.3280.2982.8886.5486.3693.6490.56
TransGeo [45]94.0829.3969.5070.8985.0164.2680.9792.1685.9640.9772.9570.2774.3283.9943.0193.7490.13
GeoDTR [43]95.4344.2084.9592.8093.5573.1482.6493.2976.8027.1968.4064.4568.5378.2874.8094.4590.20
+ +# 5.1 Fine-grained Corruption Robustness Benchmarking Results + +Benchmarking Results on CVUSA-C Table 2 presents the results of 8 cross-view geo-localization methods on the fine-grained corruption robustness benchmark CVUSA-C. In general, the robustness of the model is positively correlated with its accuracy on clean data. These models with higher $\mathrm{R@1_{clean}}$ , such as L2LTR [42], TransGeo [45], and GeoDTR [43], also achieve higher $\mathrm{R@1_{cor}}$ which is understandable as different models show consistent performance degradation on corrupted images. It's worth noting that L2LTR outperforms TransGeo and GeoDTR in $\mathrm{R@1_{cor}}$ , even though TransGeo and GeoDTR exhibit higher $\mathrm{R@1_{clean}}$ . This observation differs from the results obtained in [18] and [4]. The discrepancy might be attributed to its deeper network architecture and more parameters, which make it more robust to input variations. This finding emphasizes the importance of considering the robustness of the model independently, especially when its performance on clean data is high, as higher clean performance does not necessarily indicate stronger robustness. In Fig. 5a, we plot the RCE of models under various corruption types (according to major categories). Based on our experimental results, we derive the following insights. + +Impact of Corruption Types. Based on the results from Table 2 and Fig. 5a, we observed that snow, spatter, and zoom corruptions have a significant impact on the performance of various cross-view geo-localization models in the CVUSA-C benchmark, resulting in RCE values exceeding $28\%$ for all models. Conversely, the effects of glass blur, JPEG, and pixelation on performance are relatively minor. These findings demonstrate the threats posed by adverse weather conditions and fast motion to cross-view geo-localization models. In contrast, most models exhibited less performance degradation under the influence of glass blur, JPEG compression, and pixelation, possibly due to similar corruptions being present in the training dataset, allowing the models to learn prior knowledge about these types of corruptions. + +Performance of Different Models. Among all the evaluated models, L2LTR [42] exhibited the most outstanding $\mathrm{R@1_{cor}}$ performance. Additionally, we observed a synchronized growth trend between $\mathrm{R@1_{clean}}$ and $\mathrm{R@1_{cor}}$ for all methods except L2LTR. Notably, CVM-Net [17] demonstrated the weakest corruption robustness. All of the methods show an average performance degradation + +![](images/b33b3ecfa1fbd791481f8ad1851726969a82a86762fd6a60741c0783f74d16db.jpg) +(a) CVUSA-C + +![](images/8056ba5dea64649a21ee3c4253f247f71e3f343212a589d3937012402131d9ef.jpg) +(b) CVACT_val-C +Fig. 5: The Relative Corruption Error (RCE) of different cross-view geo-localization models on CVUSA-C and CVACT_val-C benchmarks. The results are shown for each major category of corruptions, including Weather, Blur, Noise, and Digital, as well as the overall performance across all types of corruptions (Overall). + +Table 3: Experimental results of 7 cross-view geo-localization methods on the CVACT_val-C benchmark. We report the R@1 performance of each method under different corruptions (obtained by averaging the 5 corruption severities), as well as the average R@1cor under all corruption types. + +
MethodCleanCVACT_val-C
WeatherBlurNoiseDigitalR@1cor
SnowFrostFogBrightSpatterDefocusGlassMotionZoomGaussianShotImpulseSpeckleContrastPixelJPEG
OriCNN [21]46.9613.946.133.7829.4540.5431.7139.9937.5824.8934.2432.2739.0133.284.5644.3842.5628.65
SAFA [33]81.0320.0331.6633.1966.9945.6039.8372.8749.864.6248.6643.6848.8251.6115.9176.9075.8345.38
CVFT [35]61.0515.0022.3242.5347.6037.2531.3053.8836.914.1035.6830.8036.3236.8431.7958.2157.9736.16
DSM [34]82.4931.9551.7070.4369.4852.3557.3580.1667.3815.3458.3453.0558.1863.0652.7981.7280.5558.99
L2LTR [42]84.8971.0377.9383.5081.1773.7883.9885.0784.0049.7982.2081.1982.9882.2379.1585.0783.4079.15
TransGeo [45]84.9547.6558.5132.9172.6767.1381.4384.8381.8036.3481.9680.8682.8483.0122.1884.9283.7467.68
GeoDTR [43]86.2148.2471.7483.2684.6061.3979.1185.5173.448.2675.4473.9977.0680.2355.4886.0185.1970.56
+ +of $24.24\%$ on the CVUSA-C benchmark, with SAFA [33] degrading by up to $39.32\%$ and L2LTR [42] degrading by a minimum of $11.17\%$ . + +Benchmarking Results on CVACT_val-C Table 3 presents the results of 7 cross-view geo-localization methods on the fine-grained corruption robustness benchmark CVACT_val-C. Fig. 5b shows the RCE of models across various corruption types. Similar to the experimental results for the CVUSA-C benchmark, L2LTR [42] exhibits the best robustness. Meanwhile, snow, spatter, and zoom corruptions hurt the performance of the various models the most, and the various models have better robustness to glass blur, JPEG, and pixelation corruptions. However, it is worth noting that all methods show an average performance degradation of $20.14\%$ on the CVACT_val-C benchmark. + +Effects of different training sets. From the average results of all methods of CVUSA-C and CVACT_val-C benchmarks, the R@1Clean to R@1cor drop of CVACT_val-C benchmark is $4.1\%$ less than CVUSA-C. We believe that this may be due to the fact that the CVACT dataset is collected in urban areas, while the CVUSA dataset is collected in the suburbs, and thus the CVACT dataset has richer scene information and so brings more robustness to the model. + +Table 4: Experimental results of cross-view geo-localization methods on CVUSA-C-ALL, CVACT val-C-ALL, and CVACT test-C-ALL benchmarks. + +
MethodCVUSA-C-ALLCVACT_val-C-ALLCVACT_test-C-ALL
R@1_allR@5_allR@10_allR@1%_allR@1_allR@5_allR@10_allR@1%_allR@1_allR@5_allR@10_allR@1%_all
CVM-Net [17]6.0916.0523.1452.51--------
OriCNN [21]9.3822.2630.0458.9915.3128.3135.2158.393.698.3311.0443.93
SAFA [33]63.6878.0882.8293.9156.7273.6078.5991.3231.1852.0658.6090.41
CVFT [35]41.0564.0172.6491.3745.6966.4572.9788.3822.8243.4851.0788.99
DSM [34]75.2786.2689.4295.0770.0482.8185.8693.5147.1368.4173.5293.18
L2LTR [42]87.9395.4597.0199.0182.1393.3494.9398.1057.2082.5987.2398.09
TransGeo [45]82.7291.9594.0397.9274.0486.1989.1094.9852.1874.3578.9995.03
GeoDTR [43]84.6493.2995.0198.2477.4088.9591.2895.9152.8778.8483.1795.84
+ +# 5.2 Comprehensive Corruption Robustness Benchmarking Results + +The performances of different models on comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) are shown in Table 4. It is clear from the experimental results that the performance degradation is closely and positively correlated with the original performance when evaluated using the comprehensive corruption robustness benchmarks, except for L2LTR [42]. The L2LTR exhibits the highest level of robustness, albeit at the expense of increased computational cost and a greater number of trainable parameters. The results obtained by all methods for the comprehensive robustness benchmarks follow the same trend as the $\mathrm{R@1_{cor}}$ results for the fine-grained robustness benchmarks. This demonstrates the benefit of the comprehensive robustness benchmarks when considering overall robustness; they allow the best robustness model to be selected with only a small evaluation cost (as in the normal evaluation process). + +# 5.3 Robustness Enhancement Results + +![](images/57895e5a55403f00188a7d8444dbe8f3e15fed7c7da3e933c8f0f50ac7e6a2a4.jpg) +(a) SAFA + +![](images/4711de5cc3c6a7cc362cb0b02d9eb62158b1b860f39521be035b466e11ab41cf.jpg) +(b) L2LTR +Fig. 6: Stylization improves the robustness of SAFA [33], L2LTR [42], and Trans-Geo [45] on the CVUSA-C benchmark, with each severity level representing the average across all 16 corruption types. Severity $= 0$ corresponds to clean images for testing. The Standard denotes the original, unaltered training data, while Stylization denotes training exclusively on images subjected to stylization. Stylization & Standard denotes stylization and origin training data are equally interleaved during the training process. Notably, the three different training strategies require identical training complexity, and experimental configurations and model structures remain consistent throughout. + +![](images/99dfa3e77744bb355b93bb7f00cc3901dc7179f8f34f9001e16a92b7d4633f39.jpg) +(c) TransGeo + +![](images/15963e5e485b97e7b892d4289cfacd5d97780620f7484d24e2a8ae436f10815a.jpg) +(a) SAFA + +![](images/ae609c7562a9bdfcf1728217fea9fd0cd05d6857879f93772a9709f58ee63efb.jpg) +(b) L2LTR +Fig. 7: CLAHE improves the robustness of SAFA [33], L2LTR [42], and TransGeo [45] on the CVUSA-C benchmark, with each severity level representing the average across all 16 corruption types. Notations are similar to Fig. 6. + +![](images/b2fcb0cd50e1e7bf810a2e42c2f8df73a0f3a6c192658f78fcfa896a7cc2aa82.jpg) +(c) TransGeo + +Stylization for Robustness Enhancement We investigated whether the use of stylization enhances the robustness of various cross-view geo-localization models. We evaluated the performance of 3 classical cross-view geo-localization models on the CVUSA-C benchmark using 3 different training strategies, as depicted in Section 3.3 and Fig. 3. Fig. 6 shows the experimental results. + +Training on stylized images indeed resulted in stronger robustness compared to models trained only on clean images, showing less performance degradation as the severity of corruption increased. However, the performance on the original clean images (severity $= 0$ ) is noticeably poorer. This can be attributed to the fact that stylized data alters the distribution of the original data. By equally combining stylized and clean data during training (Stylization & Standard), a trade-off between clean and corrupted performance was achieved, leading to both high performance akin to the standard data on clean instances and significantly improved performance on corrupted data. Results from Fig. 6 indicate that both Stylization and Stylization & Standard training enhanced $\mathrm{R@1_{cor}}$ under corruption, especially when corruption level is greater (e.g., Severity $= 4 / 5$ ). + +Histogram Equalization for Robustness Enhancement We have further examined the impact of employing histogram equalization on the robustness of cross-view geo-localization models. In our study, we employ Contrast Limited Adaptive Histogram Equalization (CLAHE) [29] to enhance the robustness of existing methods. Fig. 7 shows the experimental results. + +It is evident that training models only on data that has undergone CLAHE does not greatly improve their robustness. On the other hand, our training strategy 3, which combines CLAHE and clean data in equal proportions, can somewhat increase the robustness of these models, though the improvement is less than that of stylization-based methods. It's also noteworthy that different models respond differently to the same training set of data. This emphasizes the importance of model robustness and the challenge of attracting researchers to invest in model robustness. + +We must emphasize that how to improve model robustness is still an open topic, which requires the efforts of more researchers. + +# 6 Further Discussion + +Real-world corruption can be much more extensive and complex. Although we cannot enumerate all possible real-world corruptions, we have systematically designed 16 corruption types, each with 5 severity levels. These types can provide a large amount of rich data to support machine learning and computer vision research. Compared to images that must be captured in the real world, synthetic imagery can provide highly controllable and diverse data for research by adjusting weather conditions, severity, and scene types as needed. At the same time, it circumvents the security risks that may be associated with capturing images of real scenes under adverse weather conditions, as well as the privacy and ethical issues that may be involved in capturing in specific environments. Therefore, we strongly believe that our benchmarks can serve as a practical testbed to perform controllable robustness evaluation, which in turn will advance research on cross-view geo-localization robustness. + +Our focus is on evaluating how well cross-view geo-localization models withstand the corruption of input images. However, the corruption robustness benchmarks we present can be applied to more research topics. For example, robustness evaluations and enhancements on cross-view image synthesis [30,31,37,38], cross-view camera pose estimation [19, 32, 41] and autonomous driving [14, 26] can also benefit from our benchmarks. + +# 7 Conclusion + +This paper systematically investigates the impact of corruption data on cross-view geo-localization models, which is a challenge previously overlooked in the context of cross-view geo-localization studies. In contrast to "clean" data, "corrupted" data is more common and closer to the realistic world, and exhaustive experiments have shown these models invariably experience significant performance degradation in the face of corrupted data. This further validates the necessity and importance of robustness benchmarking for cross-view geo-localization. To address this crucial issue and track future developments, we propose two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) for the cross-view geo-localization task. Extensive experiments are conducted to evaluate existing classical methods on these corruption robustness benchmarks, revealing new insights. Furthermore, we introduce two simple techniques (stylization and histogram equalization) and the training strategy to effectively enhance robustness, without requiring any model adjustments or introducing additional training complexity. We hope that the corruption robustness benchmarks, in-depth analysis, and insightful findings presented in this paper will raise awareness within the community regarding the robustness of cross-view geo-localization models and contribute to enhancing their robustness to challenges in complex real-world environments in the future. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China under Grant 62072318 and Grant U22A2079, in part by the Key Project of Department of Education of Guangdong Province under Grant 2023ZDZX1016, and in part by Shenzhen Science and Technology Program under Grant 20220810142553001. + +# References + +1. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016) +2. Chaabane, M., Gueguen, L., Trabelsi, A., Beveridge, R., O'Hara, S.: End-to-end learning improves static object geo-localization from video. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2063-2072 (2021) +3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016) +4. Dong, Y., Kang, C., Zhang, J., Zhu, Z., Wang, Y., Yang, X., Su, H., Wei, X., Zhu, J.: Benchmarking robustness of 3d object detection to common corruptions in autonomous driving. arXiv preprint arXiv:2303.11040 (2023) +5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) +6. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010) +7. Friedland, G., Vinyals, O., Darrell, T.: Multimodal location estimation. In: Proceedings of the 18th ACM international conference on Multimedia. pp. 1245-1252 (2010) +8. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3855-3863 (2017) +9. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016) +0. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018) +11. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. International Conference on Learning Representations, International Conference on Learning Representations (Sep 2018) + +12. Geirhos, R., Temme, C.R., Rauber, J., Schütt, H.H., Bethge, M., Wichmann, F.A.: Generalisation in humans and deep neural networks. Advances in neural information processing systems 31 (2018) +13. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 580-587 (2014) +14. Hane, C., Heng, L., Lee, G.H., Fraundorfer, F., Furgale, P., Sattler, T., Pollefeys, M.: 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image and Vision Computing 68, 14-27 (2017) +15. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence 33(12), 2341-2353 (2010) +16. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) +17. Hu, S., Feng, M., Nguyen, R.M., Lee, G.H.: Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7258-7267 (2018) +18. Kamann, C., Rother, C.: Benchmarking the robustness of semantic segmentation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8828-8838 (2020) +19. Lentsch, T., Xia, Z., Caesar, H., Kooij, J.F.: Slicematch: Geometry-guided aggregation for cross-view pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17225-17234 (2023) +20. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2736-2744 (2016) +21. Liu, L., Li, H.: Lending orientation to neural networks for cross-view geolocation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5624-5633 (2019) +22. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing 27(6), 3064-3073 (2018) +23. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015) +24. McManus, C., Churchill, W., Maddern, W., Stewart, A.D., Newman, P.: Shady dealings: Robust, long-term visual localisation using illumination invariance. In: 2014 IEEE international conference on robotics and automation (ICRA). pp. 901-906. IEEE (2014) +25. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019) +26. Middelberg, S., Sattler, T., Untzelmann, O., Kobbelt, L.: Scalable 6-dof localization on mobile devices. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13. pp. 268-283. Springer (2014) +27. Mithun, N.C., Minhas, K.S., Chiu, H.P., Oskiper, T., Sizintsev, M., Samarasekera, S., Kumar, R.: Cross-view visual geo-localization for outdoor augmented reality. In: 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR). pp. 493-502. IEEE (2023) + +28. Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A., Greer, T., ter Haar Romeny, B., Zimmerman, J.B., Zuiderveld, K.: Adaptive histogram equalization and its variations. Computer Vision, Graphics, and Image Processing p. 355-368 (Aug 1987). https://doi.org/10.1016/s0734-189x(87)80186-x, http://dx.doi.org/10.1016/s0734-189x(87)80186-x +29. Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A., Greer, T., ter Haar Romeny, B., Zimmerman, J.B., Zuiderveld, K.: Adaptive histogram equalization and its variations. Computer vision, graphics, and image processing 39(3), 355-368 (1987) +30. Regmi, K., Borji, A.: Cross-view image synthesis using conditional gans. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 3501-3510 (2018) +31. Shi, Y., Campbell, D., Yu, X., Li, H.: Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 10009-10022 (2022) +32. Shi, Y., Li, H.: Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17010-17020 (2022) +33. Shi, Y., Liu, L., Yu, X., Li, H.: Spatial-aware feature aggregation for image based cross-view geo-localization. Advances in Neural Information Processing Systems 32 (2019) +34. Shi, Y., Yu, X., Campbell, D., Li, H.: Where am i looking at? joint location and orientation estimation by cross-view matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4064-4072 (2020) +35. Shi, Y., Yu, X., Liu, L., Zhang, T., Li, H.: Optimal feature transport for cross-view image geo-localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11990-11997 (2020) +36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (Jan 2015) +37. Tang, H., Xu, D., Yan, Y., Torr, P.H., Sebe, N.: Local class-specific and global image-level generative adversarial networks for semantic-guided scene generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7870-7879 (2020) +38. Token, A., Zhou, Q., Maximov, M., Leal-Taixe, L.: Coming down to earth: Satellite-to-street view synthesis for geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6488-6497 (2021) +39. Vasiljevic, I., Chakrabarti, A., Shakhnarovich, G.: Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760 (2016) +40. Workman, S., Souvenir, R., Jacobs, N.: Wide-area image geolocation with aerial reference imagery. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3961-3969 (2015) +41. Xia, Z., Booij, O., Manfredi, M., Kooij, J.F.: Visual cross-view metric localization with dense uncertainty estimates. In: European Conference on Computer Vision. pp. 90-106. Springer (2022) +42. Yang, H., Lu, X., Zhu, Y.: Cross-view geo-localization with layer-to-layer transformer. Advances in Neural Information Processing Systems 34, 29009-29020 (2021) +43. Zhang, X., Li, X., Sultani, W., Zhou, Y., Wshah, S.: Cross-view geo-localization via learning disentangled geometric layout correspondence. arXiv preprint arXiv:2212.04074 (2022) + +44. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019) +45. Zhu, S., Shah, M., Chen, C.: Transgeo: Transformer is all you need for cross-view image geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1162-1171 (2022) \ No newline at end of file diff --git a/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/images.zip b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fe19e5325b5a2d9bd78f165a9ebf76afce5c0a1e --- /dev/null +++ b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1b4526240a1118af2eb6f28c9aeadfb19602f46d6a6b38e060aba8a4b76f0ed +size 489442 diff --git a/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/layout.json b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..661c876844222baec2b02dca7e6e4772557fb39c --- /dev/null +++ b/2024/Benchmarking the Robustness of Cross-view Geo-localization Models/layout.json @@ -0,0 +1,8665 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 111, + 462, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 111, + 462, + 147 + ], + "spans": [ + { + "bbox": [ + 152, + 111, + 462, + 147 + ], + "type": "text", + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 214, + 168, + 400, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 168, + 400, + 181 + ], + "spans": [ + { + "bbox": [ + 214, + 168, + 400, + 181 + ], + "type": "text", + "content": "Qingwang Zhang and Yingying Zhu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 190, + 463, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 190, + 463, + 201 + ], + "spans": [ + { + "bbox": [ + 149, + 190, + 463, + 201 + ], + "type": "text", + "content": "College of Computer Science and Software Engineering, Shenzhen University," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 272, + 201, + 341, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 201, + 341, + 212 + ], + "spans": [ + { + "bbox": [ + 272, + 201, + 341, + 212 + ], + "type": "text", + "content": "Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 213, + 429, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 213, + 429, + 223 + ], + "spans": [ + { + "bbox": [ + 184, + 213, + 429, + 223 + ], + "type": "text", + "content": "zhangqingwang2022@email.szu.edu.cn，zhuyy@szu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 189, + 224, + 423, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 224, + 423, + 234 + ], + "spans": [ + { + "bbox": [ + 189, + 224, + 423, + 234 + ], + "type": "text", + "content": "https://zqwlearning.github.io/CrossViewRobustness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 263, + 455, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 263, + 455, + 526 + ], + "spans": [ + { + "bbox": [ + 160, + 263, + 455, + 526 + ], + "type": "text", + "content": "Abstract. Cross-view geo-localization serves as a viable alternative to providing geographical location information when GPS signals are unstable or unavailable by matching ground images with geo-tagged aerial image databases. While significant progress has been made on some common benchmarks like CVUSA and CVACT, there remains a lack of comprehensive consideration for robustness against real-world environmental challenges such as adverse weather or sensor noise. This deficiency poses a significant challenge for deploying this technology in safety-critical domains like autonomous driving and robot navigation. To the best of our knowledge, there is currently no specialized benchmark for evaluating the robustness of cross-view geo-localization. To comprehensively and fairly evaluate the robustness of cross-view geo-localization models in real-world scenarios, we introduce 16 common types of data corruption. By synthesizing these corruptions on public datasets, we establish two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL), covering approximately 1.5 million corrupted images. Subsequently, we conduct large-scale experiments on various cross-view geo-localization models to evaluate their robustness in corrupted environments and derive novel insights. Finally, we explore two data augmentation strategies as potential solutions to enhance model robustness. Combined with the training strategies proposed, these approaches effectively enhance the robustness of multiple models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 537, + 447, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 537, + 447, + 548 + ], + "spans": [ + { + "bbox": [ + 160, + 537, + 447, + 548 + ], + "type": "text", + "content": "Keywords: Cross-view geo-localization " + }, + { + "bbox": [ + 160, + 537, + 447, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 537, + 447, + 548 + ], + "type": "text", + "content": " Benchmarking " + }, + { + "bbox": [ + 160, + 537, + 447, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 537, + 447, + 548 + ], + "type": "text", + "content": " Robustness" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "content": "Geo-localization is fundamental for autonomous driving [26], robot navigation [24], 3D reconstruction [14], augmented reality [27], and road maintenance [2], etc. However, in the realistic scenario, a GPS-degraded or GPS-denied environment is not uncommon, including urban canyon effects, forest areas, mountainous terrain, and interference or jamming regions. The absence of geo-tags in image metadata, often resulting from GPS signal obstruction, signal attenuation, or" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 209 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 209 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 209 + ], + "type": "image", + "image_path": "06bd3fd4b0605c8f789f4ebd564a7bd8d03c092fb68b5f03881678912ad5b12c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 218, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 274 + ], + "type": "text", + "content": "Fig. 1: Existing cross-view geo-localization models fail when the ground query image is corrupted. When the query image is clean, almost all existing methods retrieve the ground truth correctly (marked by the red box and the red circle). Once the image is corrupted, almost all existing methods have missed the correct result. Incorrect retrieve and location results are marked in green and blue." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 386 + ], + "type": "text", + "content": "lack of a camera localization module, leads to only " + }, + { + "bbox": [ + 130, + 301, + 482, + 386 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 130, + 301, + 482, + 386 + ], + "type": "text", + "content": " of Internet images being geo-tagged [7]. Cross-view geo-localization can determine the location where a ground-level image is taken by comparing it with a geo-tagged database of aerial images (e.g., satellite images), which can serve as an alternative way or an effective complement to provide geo-location information in the case of GPS-degraded or GPS-denied. Therefore, cross-view geo-localization has recently garnered considerable attention due to its wide range of practical applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 615 + ], + "type": "text", + "content": "Two favorite benchmarks have been introduced to evaluate cross-view geo-localization performance, e.g., CVUSA [40] and CVACT [21]. Existing methods are all evaluated on both datasets and state-of-the-art approaches even achieve near-perfect results [42,43,45]. Nevertheless, these benchmarks only present ideal settings and do not reflect real-world scenarios. Almost all their ground images were collected on clean days, with little coverage in conditions such as rain, snow, and sensor noise corruption. Inevitably, due to the significant differences between the distributions of the training data and the real-world data, the performance of current data-driven models degrades drastically or even fails when the models trained on \"clean\" data are applied to \"corrupted\" data. Fig. 1 shows examples of localization failures that can result from the ground-level query under different weather conditions. This poses a significant challenge to applying this task to safety-critical applications such as autonomous driving. Thus, prior to deployment, it has become imperative to create an easy-to-use corruption robustness benchmark and to comprehensively evaluate the robustness of existing models with respect to corrupted data. To the best of our knowledge, we first create cross-view geo-localization robustness benchmarks and provide an exhaustive assessment of the robustness of existing mainstream cross-view geo-localization models when faced with corrupted data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In the cross-view geo-localization task setup, the aerial images are cached by the system in advance with high-definition and stable characteristics, while the ground-view images are collected in real-time, which is more flexible and versatile. Therefore, in this paper, we only corrupt the ground images to consider" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Q. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 201, + 148 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 201, + 148 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 201, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 201, + 148 + ], + "type": "image", + "image_path": "16fcf319507002768594b158dcce7f6092c42045997bab15964a51a7b60c7406.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 149, + 177, + 156 + ], + "lines": [ + { + "bbox": [ + 159, + 149, + 177, + 156 + ], + "spans": [ + { + "bbox": [ + 159, + 149, + 177, + 156 + ], + "type": "text", + "content": "Clean" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 114, + 271, + 148 + ], + "blocks": [ + { + "bbox": [ + 203, + 114, + 271, + 148 + ], + "lines": [ + { + "bbox": [ + 203, + 114, + 271, + 148 + ], + "spans": [ + { + "bbox": [ + 203, + 114, + 271, + 148 + ], + "type": "image", + "image_path": "86a29d270ec50067d1d63210cde1cab99d587fa047bcad0105ed588710224b93.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 149, + 261, + 156 + ], + "lines": [ + { + "bbox": [ + 214, + 149, + 261, + 156 + ], + "spans": [ + { + "bbox": [ + 214, + 149, + 261, + 156 + ], + "type": "text", + "content": "Weather - Snow" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 272, + 114, + 340, + 148 + ], + "blocks": [ + { + "bbox": [ + 272, + 114, + 340, + 148 + ], + "lines": [ + { + "bbox": [ + 272, + 114, + 340, + 148 + ], + "spans": [ + { + "bbox": [ + 272, + 114, + 340, + 148 + ], + "type": "image", + "image_path": "8be61a441409db8a515cfc624ed746383bdbb1d963a241a47462ac7f5fdff23d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 149, + 328, + 156 + ], + "lines": [ + { + "bbox": [ + 285, + 149, + 328, + 156 + ], + "spans": [ + { + "bbox": [ + 285, + 149, + 328, + 156 + ], + "type": "text", + "content": "Blur - Defocus" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 410, + 148 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 410, + 148 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 410, + 148 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 410, + 148 + ], + "type": "image", + "image_path": "04c907550e4dbe1e9d4bec543738ad329a3ad1385e785af42128e4012e887666.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 149, + 402, + 156 + ], + "lines": [ + { + "bbox": [ + 353, + 149, + 402, + 156 + ], + "spans": [ + { + "bbox": [ + 353, + 149, + 402, + 156 + ], + "type": "text", + "content": "Noise-Gaussian" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 411, + 114, + 480, + 148 + ], + "blocks": [ + { + "bbox": [ + 411, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 411, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 411, + 114, + 480, + 148 + ], + "type": "image", + "image_path": "3b8089f5e83a3868d7a34237f1ec89e5c5f4395f410ab109a3744533afbb1924.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 149, + 468, + 156 + ], + "lines": [ + { + "bbox": [ + 425, + 149, + 468, + 156 + ], + "spans": [ + { + "bbox": [ + 425, + 149, + 468, + 156 + ], + "type": "text", + "content": "Digital-JPEG" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 134, + 172, + 201, + 206 + ], + "blocks": [ + { + "bbox": [ + 134, + 172, + 201, + 206 + ], + "lines": [ + { + "bbox": [ + 134, + 172, + 201, + 206 + ], + "spans": [ + { + "bbox": [ + 134, + 172, + 201, + 206 + ], + "type": "image", + "image_path": "93bb8548180c491d9b84d1f3d77a634d93ccdea4f83982a79f49fb821d3a868d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 207, + 186, + 214 + ], + "lines": [ + { + "bbox": [ + 149, + 207, + 186, + 214 + ], + "spans": [ + { + "bbox": [ + 149, + 207, + 186, + 214 + ], + "type": "text", + "content": "Severity " + }, + { + "bbox": [ + 149, + 207, + 186, + 214 + ], + "type": "inline_equation", + "content": "= 1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 219, + 432, + 229 + ], + "lines": [ + { + "bbox": [ + 181, + 219, + 432, + 229 + ], + "spans": [ + { + "bbox": [ + 181, + 219, + 432, + 229 + ], + "type": "text", + "content": "(b) Comparison of different severity levels, taking snow as an example." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 238, + 480, + 271 + ], + "lines": [ + { + "bbox": [ + 131, + 238, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 480, + 271 + ], + "type": "text", + "content": "Fig. 2: Various types and severities of common corruption in our robustness benchmarks, after cropping are provided (best viewed when zoomed in on the screen). Visualizations of all corruption types can be found in the supplementary material." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 203, + 172, + 271, + 205 + ], + "blocks": [ + { + "bbox": [ + 203, + 172, + 271, + 205 + ], + "lines": [ + { + "bbox": [ + 203, + 172, + 271, + 205 + ], + "spans": [ + { + "bbox": [ + 203, + 172, + 271, + 205 + ], + "type": "image", + "image_path": "53542b972cfd266f41ebd48dd936501018af1b0da302fe523729f00e45d853e0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 207, + 256, + 214 + ], + "lines": [ + { + "bbox": [ + 219, + 207, + 256, + 214 + ], + "spans": [ + { + "bbox": [ + 219, + 207, + 256, + 214 + ], + "type": "text", + "content": "Severity " + }, + { + "bbox": [ + 219, + 207, + 256, + 214 + ], + "type": "inline_equation", + "content": "= 2" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 272, + 171, + 340, + 205 + ], + "blocks": [ + { + "bbox": [ + 212, + 163, + 400, + 171 + ], + "lines": [ + { + "bbox": [ + 212, + 163, + 400, + 171 + ], + "spans": [ + { + "bbox": [ + 212, + 163, + 400, + 171 + ], + "type": "text", + "content": "(a) Comparison between clean and corrupted images." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 171, + 340, + 205 + ], + "lines": [ + { + "bbox": [ + 272, + 171, + 340, + 205 + ], + "spans": [ + { + "bbox": [ + 272, + 171, + 340, + 205 + ], + "type": "image", + "image_path": "60f90392f7d73aa66daa8e25e0e469a25d8aca5722912552a7372c7d9d07d3e1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 207, + 325, + 214 + ], + "lines": [ + { + "bbox": [ + 288, + 207, + 325, + 214 + ], + "spans": [ + { + "bbox": [ + 288, + 207, + 325, + 214 + ], + "type": "text", + "content": "Severity " + }, + { + "bbox": [ + 288, + 207, + 325, + 214 + ], + "type": "inline_equation", + "content": "= 3" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 342, + 171, + 410, + 205 + ], + "blocks": [ + { + "bbox": [ + 342, + 171, + 410, + 205 + ], + "lines": [ + { + "bbox": [ + 342, + 171, + 410, + 205 + ], + "spans": [ + { + "bbox": [ + 342, + 171, + 410, + 205 + ], + "type": "image", + "image_path": "fb0a4cf89e82329d1a452b8ea15aa9537e1455000858c51c412b0588ae8cb366.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 207, + 394, + 214 + ], + "lines": [ + { + "bbox": [ + 358, + 207, + 394, + 214 + ], + "spans": [ + { + "bbox": [ + 358, + 207, + 394, + 214 + ], + "type": "text", + "content": "Severity " + }, + { + "bbox": [ + 358, + 207, + 394, + 214 + ], + "type": "inline_equation", + "content": "= 4" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 412, + 171, + 480, + 205 + ], + "blocks": [ + { + "bbox": [ + 412, + 171, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 412, + 171, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 412, + 171, + 480, + 205 + ], + "type": "image", + "image_path": "0212ff984b351a754bf813d509db881005f0dc4a8d5efa7283d309c178dda7dc.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 207, + 463, + 214 + ], + "lines": [ + { + "bbox": [ + 428, + 207, + 463, + 214 + ], + "spans": [ + { + "bbox": [ + 428, + 207, + 463, + 214 + ], + "type": "text", + "content": "Severity " + }, + { + "bbox": [ + 428, + 207, + 463, + 214 + ], + "type": "inline_equation", + "content": "= 5" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "content": "the robustness of the existing models, which is more practical while alleviating the evaluation cost. We systematically design 16 common types of corruption that are commonly found in the cross-view geo-localization task and provide a comprehensive evaluation of the corruption robustness of existing models. These corruptions include Weather, Blur, Noise, and Digital, covering a wide range of real-world situations. Furthermore, each type of corruption has 5 severity levels, totaling 80 different types of corruption involved. We have applied these corruptions on CVUSA [40] and CVACT [21] to create two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL), which cover a total of about 1.5 million corrupted images. Fig. 2 shows some examples of corrupted images. We expect these datasets to serve as common benchmarks for comprehensively and fairly evaluating the corruption robustness of cross-view geo-localization models and to facilitate future research to advance the cross-view geo-localization task." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 146, + 476, + 428, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 476, + 428, + 488 + ], + "spans": [ + { + "bbox": [ + 146, + 476, + 428, + 488 + ], + "type": "text", + "content": "The contributions made to this work are summed up as follows:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 137, + 497, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 138, + 497, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 497, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 497, + 480, + 556 + ], + "type": "text", + "content": "- To the best of our knowledge, we have benchmarked the robustness of state-of-the-art cross-view geo-localization models against real-world data corruption challenges for the first time. We have generated about 1.5 million corrupted images based on the CVUSA and CVACT datasets to establish benchmarks for assessing the robustness of cross-view geo-localization models." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 138, + 557, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 628 + ], + "type": "text", + "content": "- Based on the benchmark study, we have several new insights: (1) in most cases, the clean performance " + }, + { + "bbox": [ + 138, + 557, + 480, + 628 + ], + "type": "inline_equation", + "content": "(\\mathrm{R}@\\mathcal{K}_{\\mathrm{clean}})" + }, + { + "bbox": [ + 138, + 557, + 480, + 628 + ], + "type": "text", + "content": " of a model is positively, but not absolutely, correlated with its robustness; (2) snow, spatter, and zoom blur more significantly affect the robustness of various models compared to other corruptions; (3) models trained on more intricate scenarios (e.g., CVACT) exhibit better robustness." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 137, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 480, + 665 + ], + "type": "text", + "content": "- Introducing stylization and histogram equalization as data augmentation techniques, along with our proposed training strategy, significantly enhances the robustness of various cross-view geo-localization models." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 448, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 143, + 301, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 143, + 301, + 155 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 301, + 155 + ], + "type": "text", + "content": "2.1 Cross-view Geo-localization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 166, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 482, + 430 + ], + "type": "text", + "content": "Cross-view geo-localization has witnessed significant advancements in recent years. Initially, [40] introduced Convolutional Neural Networks (CNN) to the cross-view matching task, resulting in notable performance improvements. Subsequently, [17] employed a VGG [36] backbone network with two branches, combined with the NetVlad [1], and proposed the weighted soft-margin ranking loss, achieving state-of-the-art performance. To further enhance network performance, [21] emphasized the importance of orientation in cross-view geo-localization and devised a method to provide orientation information to the neural network explicitly. [33] adopted spatial-aware feature aggregation modules to aggregate information-rich and diverse feature maps, while introducing polar transformation for pre-processing to narrow the geometric gap between center-aligned satellites and ground-level images. Recently, [42] explored the integration of Vision Transformer [5] in cross-view geo-localization, proposing a novel layer-to-layer Transformer with self-cross attention mechanism that highlighted the significance of considering global dependencies to reduce visual ambiguity. [45] introduced an attention-guided pure Transformer approach, which further improved the resolution of satellite images through additional training, thus advancing the performance of the task. [43] introduced a novel geometric layout extraction module that explicitly decouples geometric information from original features and proposed two types of data augmentation methods. However, all these methods ignore the performance of the model on corrupted images, i.e., robustness, which is the focus of this paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 450, + 285, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 285, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 285, + 462 + ], + "type": "text", + "content": "2.2 Robustness Benchmarks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 483, + 668 + ], + "type": "text", + "content": "Robustness benchmarks are essential in the field of computer vision to enhance the stability and reliability of computer vision systems when facing uncertainties and noise. For instance, IMAGENET-C and IMAGENET-P were proposed to evaluate robustness in classification tasks [16]. In the context of autonomous driving, PASCAL-C, COCO-C, and Cityscapes-C were introduced as benchmarks for evaluating the robustness of object detection tasks [13, 25]. Similarly, robustness benchmarks for semantic segmentation [23] were established using PASCAL VOC 2012 [6], Cityscapes [3], and ADE20K [18, 44]. Furthermore, KITTI-C, nuScenes-C, and Waymo-C were devised as robustness benchmarks for 3D object detection [4]. These benchmarks aid in evaluating the robustness of models under various complex environmental conditions, thus facilitating their application in the real world. However, to the best of our knowledge, no researcher has proposed relevant robustness benchmarks specifically for cross-view geo-localization. Therefore, this paper addresses this gap by introducing the benchmarks CVUSA-C, CVACT_val-C, CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL, thus filling this void in the field." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "Q. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 289, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 289, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 289, + 126 + ], + "type": "text", + "content": "2.3 Robustness Enhancement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 348 + ], + "type": "text", + "content": "Several strategies have been proposed to address the impact of corruption in computer vision. For example, [20] and [8] present methods that utilize patch-based priors for background and rain layers, along with a deep CNN to effectively remove rain streaks from images. In a similar vein, [15] introduced a simple yet powerful technique employing the dark channel before eliminating haze from single-input images. Moreover, [22] devised the DesnowNet, a multi-stage network specifically designed to remove snow particles from images. However, a limitation of these methods lies in their reliance on specific designs tailored to particular types of corruption. This hinders their ability to generalize and handle other types of corruption effectively. As an alternative, some approaches have sought to enhance model performance through data augmentation, incorporating corrupted data during the training process. Although fine-tuning on specific corruption data has shown promise in boosting performance for those particular corruption types [12, 39] discovered that fine-tuning on one type of corruption often struggles to generalize to other corruption types. How to enhance robustness remains an open question. In this work, we employ the stylization [11] and histogram equalization [28] techniques to the train set, and we observe improvements in robustness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 365, + 404, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 404, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 404, + 377 + ], + "type": "text", + "content": "3 Robustness in Cross-view Geo-localization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 388, + 268, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 268, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 268, + 399 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": "Robustness. First, we consider a set of ground-to-aerial image pairs " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i^g,\\mathbf{I}_i^a\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": " represents the number of image pairs. The superscripts " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": " denote ground and aerial views, respectively. For a ground query image indexed as " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": ", we assume the existence of ground image encoder " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": " and aerial image encoder " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": ", which have been trained on samples from the distribution " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": ". We let " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathcal{C}}(c)" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": " approximate the frequency of real-world corruption. Existing models mostly evaluate performance when samples are drawn from the distribution " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{(\\mathbf{I}^g,\\mathbf{I}^a)\\sim \\mathcal{D}}(d(f_g(\\mathbf{I}_q^g),f_a(\\mathbf{I}_q^a)) < \\{d(f_g(\\mathbf{I}_q^g),f_a(\\mathbf{I}_i^a))|\\forall i\\in \\{1,\\dots,N\\},i\\neq q\\})" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "d(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": " representing the " + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 403, + 482, + 525 + ], + "type": "text", + "content": " distance. However, in practical deployments, systems often need to operate on low-quality or corrupted images. Hence, we construct an evaluation of the corruption robustness of the model, denoted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 140, + 531, + 473, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 531, + 473, + 544 + ], + "spans": [ + { + "bbox": [ + 140, + 531, + 473, + 544 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {c \\sim \\mathcal {C}} \\left[ \\mathbb {P} _ {\\left(\\mathbf {I} ^ {g}, \\mathbf {I} ^ {a}\\right) \\sim \\mathcal {D}} \\left(d \\left(f _ {g} \\left(c \\left(\\mathbf {I} _ {q} ^ {g}\\right)\\right), f _ {a} \\left(\\mathbf {I} _ {q} ^ {a}\\right)\\right) < \\left\\{d \\left(f _ {g} \\left(c \\left(\\mathbf {I} _ {q} ^ {g}\\right)\\right), f _ {a} \\left(\\mathbf {I} _ {i} ^ {a}\\right)\\right) | \\forall i \\in \\{1, \\dots , N \\}, i \\neq q \\right\\}\\right) \\right]", + "image_path": "738140a617b7ee3b001e8a6a40f4c824d0f7ce6286770375bbc5918ec538acb8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 548, + 481, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 481, + 609 + ], + "type": "text", + "content": "In this work, in order to approximate the real-world corruption " + }, + { + "bbox": [ + 130, + 548, + 481, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 548, + 481, + 609 + ], + "type": "text", + "content": ", we designed a set of common corruptions (Weather, Blur, Noise, and Digital) in the cross-view geo-localization task. The types of corruption already provide a good overview of corruption that street view images may face. From these common corruptions, we create our cross-view geo-localization robustness evaluation benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 252, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 252, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 252, + 636 + ], + "type": "text", + "content": "3.2 Image Corruption" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Common corruptions are categorized based on [16], and ground-level image corruptions are divided into four major categories with a total of 16 subcategories." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 166, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 448, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 357 + ], + "type": "text", + "content": "The first major category is weather-related corruptions, which encompasses snow, frost, fog, brightness, and spatter. The second major category consists of blur-related corruptions, including defocus, glass, motion, and zoom blur. Defocus blur occurs when an image is out of focus, while glass blur arises when images are captured through frosted glass windows of sensors. Motion and zoom blurring occur in scenes with rapid camera movement or quick approaches toward objects, and they are particularly prone to occur in image collections from vehicular devices. The third major category involves noise-related corruptions, which cover Gaussian, shot, impulse, and speckle noise. Gaussian noise emerges under low-light conditions, attributed to the discrete nature of light and causing electronic noise. Impulse noise results from bit errors, resembling the color analog of salt-and-pepper noise. Speckle noise, which occurs due to light interference, leads to the appearance of bright and dark speckles. Lastly, the fourth category includes digital corruptions, including contrast, pixelation, and JPEG corruption. Contrast corruption is dependent on lighting conditions and the colors of the objects during capture. Pixelation occurs during upscaling low-resolution images, and JPEG corruption arises when images undergo lossy compression, introducing compression artifacts. To better simulate real-world corruption, each type of corruption also contains 5 severity levels. We present visualizations of various types and severities of common corruption in Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 371, + 338, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 371, + 338, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 371, + 338, + 384 + ], + "type": "text", + "content": "3.3 Robustness Enhancement Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 390, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 450 + ], + "type": "text", + "content": "To enhance the robustness of existing models against corruption, we investigate two data augmentation approaches: stylization and histogram equalization. By applying either stylization or histogram equalization to the training data and employing our proposed training strategy, we effectively bolstered the corruption robustness of multiple cross-view geo-localization models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 510 + ], + "type": "text", + "content": "Stylization. Style transfer, introduced by [9], merges the content and style of two different images to generate a novel image that retains the content of the original image while adopting the style of the target image, which is also known as stylization. In the context of image classification and object detection tasks, style transfer has demonstrated its effectiveness in enhancing robustness [10]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": "Histogram Equalization. Histogram equalization is a fundamental technique used for enhancing image contrast. It operates by redistributing pixel intensity values within an image, stretching or compressing the brightness range, thus achieving a more uniform distribution of pixel intensity values across the entire brightness range of the image. In this study, we employ Contrast Limited Adaptive Histogram Equalization (CLAHE) [29] to enhance the robustness of existing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Training Strategy. In this paper, we apply Stylization / CLAHE to the cross-view geo-localization training set, testing 3 settings: (1) training using the standard (raw) training data; (2) replacing all training images with Stylization / CLAHE images, thereby eliminating the use of standard images during training; (3) having Stylization / CLAHE images participate in the training process alongside the standard images with equal probability at each iteration. Fig. 3" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Q. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 480, + 214 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 480, + 214 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 480, + 214 + ], + "type": "image", + "image_path": "8e94643ccb2b9ee0d9fd5fdad02f27ce85203c2e21478727f18368cc12886479.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 222, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 222, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 480, + 277 + ], + "type": "text", + "content": "Fig. 3: Visualization of Stylization / CLAHE applied to the CVUSA training set. The illustration depicts standard images (middle row), stylization images (top row), and CLAHE images (bottom row). The rounded rectangles on both sides represent different training strategies. The red dashed boxes indicate the style images sourced from https://www.kaggle.com/c/painter-by-numbers/. Inspired by [25]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 324 + ], + "type": "text", + "content": "shows examples of applying Stylization / CLAHE, and it illustrates 3 training strategies with the same training complexity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 342, + 367, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 367, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 367, + 355 + ], + "type": "text", + "content": "4 Corruption Robustness Benchmarks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 485, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 485, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 485, + 425 + ], + "type": "text", + "content": "We create two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) from the classic cross-view geo-localization datasets CVUSA and CVACT. See Table 1 for more detailed information on these benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 443, + 409, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 409, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 409, + 455 + ], + "type": "text", + "content": "4.1 Fine-grained Corruption Robustness Benchmarks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 545 + ], + "type": "text", + "content": "We create fine-grained robustness benchmarks comprising 16 corruption types, each corresponding to 5 severity levels denoted by integers from 1 to 5, where higher numbers indicate more severe corruption. Consequently, one single original ground query image can yield 80 " + }, + { + "bbox": [ + 130, + 462, + 480, + 545 + ], + "type": "inline_equation", + "content": "(16 \\times 5)" + }, + { + "bbox": [ + 130, + 462, + 480, + 545 + ], + "type": "text", + "content": " corrupted images. In the fine-grained corruption robustness benchmarks, each type of corrupted image forms an independent evaluation subset with no overlap. The suffix \"-C\" designates all fine-grained robustness benchmarks (CVUSA-C and CVACT_val-C)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "CVUSA-C. The CVUSA [40] dataset is one of the earliest cross-view geo-localization datasets, primarily collected from suburban areas in the United States. It encompasses a total of 35,532 pairs of ground-aerial images for training and 8,884 pairs for testing. We have employed the CVUSA test set to generate the fine-grained corruption robustness benchmark, CVUSA-C. To be specific, CVUSA-C comprises all types and severities of image corruptions, each residing independently within individual subsets of CVUSA-C. Given that there are 16 distinct corruption types and 5 levels of corruption severity, CVUSA-C effectively comprises 80 evaluation subsets. Each of these subsets contains 8,884 ground images for testing, resulting in an aggregate of 710,720 corrupted images." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 101 + ], + "type": "text", + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 126, + 480, + 171 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 474, + 125 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 474, + 125 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 474, + 125 + ], + "type": "text", + "content": "Table 1: Detailed information on the proposed corruption robustness benchmarks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 126, + 480, + 171 + ], + "lines": [ + { + "bbox": [ + 134, + 126, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 134, + 126, + 480, + 171 + ], + "type": "table", + "html": "
Fine-grainedComprehensive
CVUSA-CCVACT_val-CCVUSA-C-ALLCVACT_val-C-ALLCVACT_test-C-ALL
Number of original validation / test ground images8,8848,8848,8848,88492,802
Whether or not evaluation subsets are generated for each corruption✓✓✗✗✗
Whether all corruptions are included in an independent subset✗✗✓✓✓
Number of validation / test ground images for our benchmark8,884 × 16 × 58,884 × 16 × 58,8848,88492,802
Storage space~39 GB~178 GB~0.5 GB~2 GB~21 GB
", + "image_path": "107ebb215b14fea7b660bd7c7598441c9705be06fadbc8a4439b97c8e2d8d24f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 479, + 289 + ], + "type": "text", + "content": "CVACT_val-C. The CVACT [21] dataset was collected by [21] and densely covers an urban area of Australia (Canberra). Similar to the CVUSA dataset, this dataset comprises 35,532 pairs of ground-aerial images for training and 8,884 image pairs for validation (referred to as CVACT_val). We create a fine-grained robustness benchmark from CVACT_val, as CVUSA does. The benchmark includes various types and severities, resulting in a total of 80 validation subsets. These subsets consist of 710,720 corrupted images and are collectively referred to as CVACT_val-C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": "Evaluation Metrics. The original evaluation metrics for cross-view geolocation primarily revolved around " + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathcal{K}" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "inline_equation", + "content": "K\\in \\{1,5,10,1\\% \\}" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": "), which represents the probability of correctly identifying the matching image within the top " + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": " retrieved reference images based on the query image [17,21,33,42,45]. We denote the performance of models on the original validation or test sets as " + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathcal{K}_{\\mathrm{clean}}" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": ". For each corruption type " + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": " and each severity level " + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": ", we employ " + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathcal{K}_{c,s}" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": " to gauge the performance of models under corruption conditions. Finally, by averaging across all corruption types and severity levels, we compute the average performance of model " + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathcal{K}_{\\mathrm{cor}}" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": " under corruption conditions. In this context, " + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 290, + 480, + 434 + ], + "type": "text", + "content": " represents the set of corruptions in evaluation. Additionally, we calculate the Relative Corruption Error (RCE) by measuring the percentage of performance degradation [4]. A higher RCE implies poorer robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 242, + 445, + 480, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 445, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 242, + 445, + 480, + 472 + ], + "type": "interline_equation", + "content": "R @ K _ {\\text {c o r}} = \\frac {1}{| \\mathcal {C} |} \\sum_ {c \\in \\mathcal {C}} \\frac {1}{5} \\sum_ {s = 1} ^ {5} R @ K _ {c, s} \\tag {1}", + "image_path": "79b587796f5a1931f57269f9c53f4d54c8e0cb8d36a8cf4f9b890a018e6fc7b9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 187, + 483, + 480, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 483, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 187, + 483, + 480, + 506 + ], + "type": "interline_equation", + "content": "\\mathrm {R C E} _ {c, s} = \\frac {\\mathrm {R @ K} _ {\\text {c l e a n}} - \\mathrm {R @ K} _ {c , s}}{\\mathrm {R @ K} _ {\\text {c l e a n}}}; \\mathrm {R C E} = \\frac {\\mathrm {R @ K} _ {\\text {c l e a n}} - \\mathrm {R @ K} _ {\\text {c o r}}}{\\mathrm {R @ K} _ {\\text {c l e a n}}} \\tag {2}", + "image_path": "13cf8d37e9e56e377825c3656beeac749af8515a760401d16316be67b61f5e1e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 523, + 422, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 422, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 422, + 536 + ], + "type": "text", + "content": "4.2 Comprehensive Corruption Robustness Benchmarks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, we introduce comprehensive corruption robustness benchmarks. Prior efforts [4,16,25] often generated separate test sets for each corruption type and its corresponding 5 severity levels. While this approach allows for a finer-grained evaluation of each corruption type's impact on the model, it imposes significant storage and computational costs (80 evaluation subsets). Recognizing that real-world concerns may prioritize a model's overall performance against diverse corruptions, for each corruption type and severity level, we aggregate all forms of corruption into a single evaluation set, creating comprehensive benchmarks. The suffix \"-C-ALL\" designates all comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Q. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 185 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 185 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 185 + ], + "type": "image", + "image_path": "809b3f9149d68c4a21693a59694559f4a543cef90dd9f70e45fb1680b53bb2b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 193, + 482, + 237 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 237 + ], + "type": "text", + "content": "Fig. 4: The proposed fine-grained and comprehensive robustness benchmarks. Each corruption category encompasses 5 severity levels. The primary distinction between fine-grained and comprehensive robustness benchmarks lies in whether separate evaluation subsets are created for each corruption category and severity level." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 479, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 334 + ], + "type": "text", + "content": "CVUSA-C-ALL and CVACT_val-C-ALL. For the 8,884-pair test set of the CVUSA dataset and the 8,884-pair validation set of the CVACT_val dataset, we designed comprehensive corruption robustness benchmarks. These benchmarks are denoted as CVUSA-C-ALL and CVACT_val-C-ALL, respectively. Each dataset contains 8,884 corrupted images and systematically covers all corruption types and severity levels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 503, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 503, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 503, + 407 + ], + "type": "text", + "content": "CVACT_test-C-ALL. The CVACT dataset additionally provides 92,802 pairs of images for testing purposes (referred to as CVACT_test), showcasing dense coverage of urban areas. Similar to CVUSA-C-ALL and CVACT_val-C-ALL, we create a comprehensive corruption robustness benchmark from CVACT_test, named CVACT_test-C-ALL, containing a total of 92,802 corrupted images, covering all corruption types and severity levels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 408, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 408, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 408, + 480, + 443 + ], + "type": "text", + "content": "Evaluation Metrics. Since the comprehensive corruption robustness benchmarks represent independent subsets for testing, the recall accuracy " + }, + { + "bbox": [ + 131, + 408, + 480, + 443 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}@\\mathcal{K}_{\\mathrm{all}}, K_{\\mathrm{all}} \\in \\{1, 5, 10, 1\\% \\})" + }, + { + "bbox": [ + 131, + 408, + 480, + 443 + ], + "type": "text", + "content": " can be directly employed for evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 480, + 479 + ], + "type": "text", + "content": "The comprehensive corruption robustness benchmarks serve as valuable tools for evaluating the overall robustness of models under all corruption scenarios, and they effectively reduce the cost of robustness evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 481, + 480, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 480, + 529 + ], + "type": "text", + "content": "In summary, fine-grained robustness benchmarks provide a more detailed performance evaluation, while comprehensive robustness benchmarks provide an overall performance evaluation and reduce the evaluation cost. An illustrative diagram of the corruption robustness benchmarks is presented in Fig. 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 552, + 288, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 288, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 288, + 565 + ], + "type": "text", + "content": "5 Benchmarking Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "We have selected classical cross-view geo-localization models to benchmark on the proposed benchmarks. The experimental results on fine-grained corruption robustness benchmarks are shown in Section 5.1. Section 5.2 presents the experimental results on comprehensive corruption robustness benchmarks. In Section 5.3, we report the experimental results demonstrating the influence of stylization and histogram equalization on the robustness of models. Supplementary material provides additional experimental results." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 159, + 479, + 222 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "type": "text", + "content": "Table 2: Experimental results of 8 cross-view geo-localization methods on the CVUSA-C benchmark. We report the R@1 performance of each method under different corruptions (obtained by averaging the 5 corruption severities), as well as the average R@1cor under all corruption types." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 159, + 479, + 222 + ], + "lines": [ + { + "bbox": [ + 134, + 159, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 134, + 159, + 479, + 222 + ], + "type": "table", + "html": "
MethodCleanCVUSA-C
WeatherBlurNoiseDigital
SnowFrostFogBrightSpatterDefocusGlassMotionZoomGaussianShotImpulseSpeckleContrastPixelJPEG
CVM-Net [17]22.470.868.428.3713.756.111.064.811.470.231.821.181.282.324.756.896.23
OriCNN [21]40.797.366.517.5721.6922.0120.4626.1019.6010.3217.2413.9519.4014.097.9428.5127.27
SAFA [33]89.8419.3260.4267.6381.9649.8651.2480.5655.4911.4433.0428.5130.3737.5931.6788.0581.15
CVFT [35]61.438.0030.7947.4647.5427.6324.5544.9334.898.1721.8319.1920.5626.2538.2857.2547.11
DSM [34]91.9624.2464.4484.0882.4457.5862.4884.5266.0225.1549.5546.4048.8460.8372.1190.2085.56
L2LTR [42]94.0567.1985.0092.6491.6175.2488.3593.1389.3342.0781.3280.2982.8886.5486.3693.6490.56
TransGeo [45]94.0829.3969.5070.8985.0164.2680.9792.1685.9640.9772.9570.2774.3283.9943.0193.7490.13
GeoDTR [43]95.4344.2084.9592.8093.5573.1482.6493.2976.8027.1968.4064.4568.5378.2874.8094.4590.20
", + "image_path": "099cc9235da12671bd9e825bf9e84e66dfed5cd85e53b7205caa6cef09e845ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 248, + 460, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 460, + 260 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 460, + 260 + ], + "type": "text", + "content": "5.1 Fine-grained Corruption Robustness Benchmarking Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "text", + "content": "Benchmarking Results on CVUSA-C Table 2 presents the results of 8 cross-view geo-localization methods on the fine-grained corruption robustness benchmark CVUSA-C. In general, the robustness of the model is positively correlated with its accuracy on clean data. These models with higher " + }, + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1_{clean}}" + }, + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "text", + "content": ", such as L2LTR [42], TransGeo [45], and GeoDTR [43], also achieve higher " + }, + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1_{cor}}" + }, + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "text", + "content": " which is understandable as different models show consistent performance degradation on corrupted images. It's worth noting that L2LTR outperforms TransGeo and GeoDTR in " + }, + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1_{cor}}" + }, + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "text", + "content": ", even though TransGeo and GeoDTR exhibit higher " + }, + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1_{clean}}" + }, + { + "bbox": [ + 130, + 278, + 482, + 469 + ], + "type": "text", + "content": ". This observation differs from the results obtained in [18] and [4]. The discrepancy might be attributed to its deeper network architecture and more parameters, which make it more robust to input variations. This finding emphasizes the importance of considering the robustness of the model independently, especially when its performance on clean data is high, as higher clean performance does not necessarily indicate stronger robustness. In Fig. 5a, we plot the RCE of models under various corruption types (according to major categories). Based on our experimental results, we derive the following insights." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": "Impact of Corruption Types. Based on the results from Table 2 and Fig. 5a, we observed that snow, spatter, and zoom corruptions have a significant impact on the performance of various cross-view geo-localization models in the CVUSA-C benchmark, resulting in RCE values exceeding " + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": " for all models. Conversely, the effects of glass blur, JPEG, and pixelation on performance are relatively minor. These findings demonstrate the threats posed by adverse weather conditions and fast motion to cross-view geo-localization models. In contrast, most models exhibited less performance degradation under the influence of glass blur, JPEG compression, and pixelation, possibly due to similar corruptions being present in the training dataset, allowing the models to learn prior knowledge about these types of corruptions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Performance of Different Models. Among all the evaluated models, L2LTR [42] exhibited the most outstanding " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1_{cor}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " performance. Additionally, we observed a synchronized growth trend between " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1_{clean}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1_{cor}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " for all methods except L2LTR. Notably, CVM-Net [17] demonstrated the weakest corruption robustness. All of the methods show an average performance degradation" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Q. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 118, + 297, + 234 + ], + "blocks": [ + { + "bbox": [ + 137, + 118, + 297, + 234 + ], + "lines": [ + { + "bbox": [ + 137, + 118, + 297, + 234 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 297, + 234 + ], + "type": "image", + "image_path": "b33b3ecfa1fbd791481f8ad1851726969a82a86762fd6a60741c0783f74d16db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 240, + 243, + 250 + ], + "lines": [ + { + "bbox": [ + 192, + 240, + 243, + 250 + ], + "spans": [ + { + "bbox": [ + 192, + 240, + 243, + 250 + ], + "type": "text", + "content": "(a) CVUSA-C" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 118, + 477, + 235 + ], + "blocks": [ + { + "bbox": [ + 316, + 118, + 477, + 235 + ], + "lines": [ + { + "bbox": [ + 316, + 118, + 477, + 235 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 477, + 235 + ], + "type": "image", + "image_path": "8056ba5dea64649a21ee3c4253f247f71e3f343212a589d3937012402131d9ef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 240, + 431, + 250 + ], + "lines": [ + { + "bbox": [ + 362, + 240, + 431, + 250 + ], + "spans": [ + { + "bbox": [ + 362, + 240, + 431, + 250 + ], + "type": "text", + "content": "(b) CVACT_val-C" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 258, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 304 + ], + "type": "text", + "content": "Fig. 5: The Relative Corruption Error (RCE) of different cross-view geo-localization models on CVUSA-C and CVACT_val-C benchmarks. The results are shown for each major category of corruptions, including Weather, Blur, Noise, and Digital, as well as the overall performance across all types of corruptions (Overall)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 134, + 360, + 479, + 418 + ], + "blocks": [ + { + "bbox": [ + 130, + 315, + 479, + 359 + ], + "lines": [ + { + "bbox": [ + 130, + 315, + 479, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 479, + 359 + ], + "type": "text", + "content": "Table 3: Experimental results of 7 cross-view geo-localization methods on the CVACT_val-C benchmark. We report the R@1 performance of each method under different corruptions (obtained by averaging the 5 corruption severities), as well as the average R@1cor under all corruption types." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 360, + 479, + 418 + ], + "lines": [ + { + "bbox": [ + 134, + 360, + 479, + 418 + ], + "spans": [ + { + "bbox": [ + 134, + 360, + 479, + 418 + ], + "type": "table", + "html": "
MethodCleanCVACT_val-C
WeatherBlurNoiseDigitalR@1cor
SnowFrostFogBrightSpatterDefocusGlassMotionZoomGaussianShotImpulseSpeckleContrastPixelJPEG
OriCNN [21]46.9613.946.133.7829.4540.5431.7139.9937.5824.8934.2432.2739.0133.284.5644.3842.5628.65
SAFA [33]81.0320.0331.6633.1966.9945.6039.8372.8749.864.6248.6643.6848.8251.6115.9176.9075.8345.38
CVFT [35]61.0515.0022.3242.5347.6037.2531.3053.8836.914.1035.6830.8036.3236.8431.7958.2157.9736.16
DSM [34]82.4931.9551.7070.4369.4852.3557.3580.1667.3815.3458.3453.0558.1863.0652.7981.7280.5558.99
L2LTR [42]84.8971.0377.9383.5081.1773.7883.9885.0784.0049.7982.2081.1982.9882.2379.1585.0783.4079.15
TransGeo [45]84.9547.6558.5132.9172.6767.1381.4384.8381.8036.3481.9680.8682.8483.0122.1884.9283.7467.68
GeoDTR [43]86.2148.2471.7483.2684.6061.3979.1185.5173.448.2675.4473.9977.0680.2355.4886.0185.1970.56
", + "image_path": "fac464576d046fac9d24cd92c45f47f30071eda3284080942206e0d89173bb07.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "inline_equation", + "content": "24.24\\%" + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": " on the CVUSA-C benchmark, with SAFA [33] degrading by up to " + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "inline_equation", + "content": "39.32\\%" + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": " and L2LTR [42] degrading by a minimum of " + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "inline_equation", + "content": "11.17\\%" + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 485, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 592 + ], + "type": "text", + "content": "Benchmarking Results on CVACT_val-C Table 3 presents the results of 7 cross-view geo-localization methods on the fine-grained corruption robustness benchmark CVACT_val-C. Fig. 5b shows the RCE of models across various corruption types. Similar to the experimental results for the CVUSA-C benchmark, L2LTR [42] exhibits the best robustness. Meanwhile, snow, spatter, and zoom corruptions hurt the performance of the various models the most, and the various models have better robustness to glass blur, JPEG, and pixelation corruptions. However, it is worth noting that all methods show an average performance degradation of " + }, + { + "bbox": [ + 130, + 485, + 481, + 592 + ], + "type": "inline_equation", + "content": "20.14\\%" + }, + { + "bbox": [ + 130, + 485, + 481, + 592 + ], + "type": "text", + "content": " on the CVACT_val-C benchmark." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Effects of different training sets. From the average results of all methods of CVUSA-C and CVACT_val-C benchmarks, the R@1Clean to R@1cor drop of CVACT_val-C benchmark is " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " less than CVUSA-C. We believe that this may be due to the fact that the CVACT dataset is collected in urban areas, while the CVUSA dataset is collected in the suburbs, and thus the CVACT dataset has richer scene information and so brings more robustness to the model." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 137, + 480, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 4: Experimental results of cross-view geo-localization methods on CVUSA-C-ALL, CVACT val-C-ALL, and CVACT test-C-ALL benchmarks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 137, + 480, + 213 + ], + "lines": [ + { + "bbox": [ + 134, + 137, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 134, + 137, + 480, + 213 + ], + "type": "table", + "html": "
MethodCVUSA-C-ALLCVACT_val-C-ALLCVACT_test-C-ALL
R@1_allR@5_allR@10_allR@1%_allR@1_allR@5_allR@10_allR@1%_allR@1_allR@5_allR@10_allR@1%_all
CVM-Net [17]6.0916.0523.1452.51--------
OriCNN [21]9.3822.2630.0458.9915.3128.3135.2158.393.698.3311.0443.93
SAFA [33]63.6878.0882.8293.9156.7273.6078.5991.3231.1852.0658.6090.41
CVFT [35]41.0564.0172.6491.3745.6966.4572.9788.3822.8243.4851.0788.99
DSM [34]75.2786.2689.4295.0770.0482.8185.8693.5147.1368.4173.5293.18
L2LTR [42]87.9395.4597.0199.0182.1393.3494.9398.1057.2082.5987.2398.09
TransGeo [45]82.7291.9594.0397.9274.0486.1989.1094.9852.1874.3578.9995.03
GeoDTR [43]84.6493.2995.0198.2477.4088.9591.2895.9152.8778.8483.1795.84
", + "image_path": "16b7bd95ae18dcb257a22a4e18648185884aaf14f754783663dad2b614569a80.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 236, + 474, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 474, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 474, + 248 + ], + "type": "text", + "content": "5.2 Comprehensive Corruption Robustness Benchmarking Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 256, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 413 + ], + "type": "text", + "content": "The performances of different models on comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) are shown in Table 4. It is clear from the experimental results that the performance degradation is closely and positively correlated with the original performance when evaluated using the comprehensive corruption robustness benchmarks, except for L2LTR [42]. The L2LTR exhibits the highest level of robustness, albeit at the expense of increased computational cost and a greater number of trainable parameters. The results obtained by all methods for the comprehensive robustness benchmarks follow the same trend as the " + }, + { + "bbox": [ + 130, + 256, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1_{cor}}" + }, + { + "bbox": [ + 130, + 256, + 482, + 413 + ], + "type": "text", + "content": " results for the fine-grained robustness benchmarks. This demonstrates the benefit of the comprehensive robustness benchmarks when considering overall robustness; they allow the best robustness model to be selected with only a small evaluation cost (as in the normal evaluation process)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 430, + 330, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 330, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 330, + 441 + ], + "type": "text", + "content": "5.3 Robustness Enhancement Results" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 141, + 472, + 247, + 551 + ], + "blocks": [ + { + "bbox": [ + 141, + 472, + 247, + 551 + ], + "lines": [ + { + "bbox": [ + 141, + 472, + 247, + 551 + ], + "spans": [ + { + "bbox": [ + 141, + 472, + 247, + 551 + ], + "type": "image", + "image_path": "57895e5a55403f00188a7d8444dbe8f3e15fed7c7da3e933c8f0f50ac7e6a2a4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 555, + 211, + 565 + ], + "lines": [ + { + "bbox": [ + 175, + 555, + 211, + 565 + ], + "spans": [ + { + "bbox": [ + 175, + 555, + 211, + 565 + ], + "type": "text", + "content": "(a) SAFA" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 254, + 472, + 361, + 551 + ], + "blocks": [ + { + "bbox": [ + 254, + 472, + 361, + 551 + ], + "lines": [ + { + "bbox": [ + 254, + 472, + 361, + 551 + ], + "spans": [ + { + "bbox": [ + 254, + 472, + 361, + 551 + ], + "type": "image", + "image_path": "4711de5cc3c6a7cc362cb0b02d9eb62158b1b860f39521be035b466e11ab41cf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 555, + 328, + 564 + ], + "lines": [ + { + "bbox": [ + 286, + 555, + 328, + 564 + ], + "spans": [ + { + "bbox": [ + 286, + 555, + 328, + 564 + ], + "type": "text", + "content": "(b) L2LTR" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 574, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 662 + ], + "type": "text", + "content": "Fig. 6: Stylization improves the robustness of SAFA [33], L2LTR [42], and Trans-Geo [45] on the CVUSA-C benchmark, with each severity level representing the average across all 16 corruption types. Severity " + }, + { + "bbox": [ + 130, + 574, + 482, + 662 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 130, + 574, + 482, + 662 + ], + "type": "text", + "content": " corresponds to clean images for testing. The Standard denotes the original, unaltered training data, while Stylization denotes training exclusively on images subjected to stylization. Stylization & Standard denotes stylization and origin training data are equally interleaved during the training process. Notably, the three different training strategies require identical training complexity, and experimental configurations and model structures remain consistent throughout." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 368, + 472, + 474, + 551 + ], + "blocks": [ + { + "bbox": [ + 368, + 472, + 474, + 551 + ], + "lines": [ + { + "bbox": [ + 368, + 472, + 474, + 551 + ], + "spans": [ + { + "bbox": [ + 368, + 472, + 474, + 551 + ], + "type": "image", + "image_path": "99dfa3e77744bb355b93bb7f00cc3901dc7179f8f34f9001e16a92b7d4633f39.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 555, + 445, + 565 + ], + "lines": [ + { + "bbox": [ + 396, + 555, + 445, + 565 + ], + "spans": [ + { + "bbox": [ + 396, + 555, + 445, + 565 + ], + "type": "text", + "content": "(c) TransGeo" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Q. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 116, + 246, + 194 + ], + "blocks": [ + { + "bbox": [ + 141, + 116, + 246, + 194 + ], + "lines": [ + { + "bbox": [ + 141, + 116, + 246, + 194 + ], + "spans": [ + { + "bbox": [ + 141, + 116, + 246, + 194 + ], + "type": "image", + "image_path": "15963e5e485b97e7b892d4289cfacd5d97780620f7484d24e2a8ae436f10815a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 199, + 211, + 209 + ], + "lines": [ + { + "bbox": [ + 175, + 199, + 211, + 209 + ], + "spans": [ + { + "bbox": [ + 175, + 199, + 211, + 209 + ], + "type": "text", + "content": "(a) SAFA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 254, + 117, + 361, + 194 + ], + "blocks": [ + { + "bbox": [ + 254, + 117, + 361, + 194 + ], + "lines": [ + { + "bbox": [ + 254, + 117, + 361, + 194 + ], + "spans": [ + { + "bbox": [ + 254, + 117, + 361, + 194 + ], + "type": "image", + "image_path": "ae609c7562a9bdfcf1728217fea9fd0cd05d6857879f93772a9709f58ee63efb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 200, + 328, + 209 + ], + "lines": [ + { + "bbox": [ + 286, + 200, + 328, + 209 + ], + "spans": [ + { + "bbox": [ + 286, + 200, + 328, + 209 + ], + "type": "text", + "content": "(b) L2LTR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 218, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 480, + 251 + ], + "type": "text", + "content": "Fig. 7: CLAHE improves the robustness of SAFA [33], L2LTR [42], and TransGeo [45] on the CVUSA-C benchmark, with each severity level representing the average across all 16 corruption types. Notations are similar to Fig. 6." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 368, + 117, + 474, + 194 + ], + "blocks": [ + { + "bbox": [ + 368, + 117, + 474, + 194 + ], + "lines": [ + { + "bbox": [ + 368, + 117, + 474, + 194 + ], + "spans": [ + { + "bbox": [ + 368, + 117, + 474, + 194 + ], + "type": "image", + "image_path": "b2fcb0cd50e1e7bf810a2e42c2f8df73a0f3a6c192658f78fcfa896a7cc2aa82.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 200, + 445, + 209 + ], + "lines": [ + { + "bbox": [ + 396, + 200, + 445, + 209 + ], + "spans": [ + { + "bbox": [ + 396, + 200, + 445, + 209 + ], + "type": "text", + "content": "(c) TransGeo" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "text", + "content": "Stylization for Robustness Enhancement We investigated whether the use of stylization enhances the robustness of various cross-view geo-localization models. We evaluated the performance of 3 classical cross-view geo-localization models on the CVUSA-C benchmark using 3 different training strategies, as depicted in Section 3.3 and Fig. 3. Fig. 6 shows the experimental results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 336, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 480, + 467 + ], + "type": "text", + "content": "Training on stylized images indeed resulted in stronger robustness compared to models trained only on clean images, showing less performance degradation as the severity of corruption increased. However, the performance on the original clean images (severity " + }, + { + "bbox": [ + 130, + 336, + 480, + 467 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 130, + 336, + 480, + 467 + ], + "type": "text", + "content": ") is noticeably poorer. This can be attributed to the fact that stylized data alters the distribution of the original data. By equally combining stylized and clean data during training (Stylization & Standard), a trade-off between clean and corrupted performance was achieved, leading to both high performance akin to the standard data on clean instances and significantly improved performance on corrupted data. Results from Fig. 6 indicate that both Stylization and Stylization & Standard training enhanced " + }, + { + "bbox": [ + 130, + 336, + 480, + 467 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1_{cor}}" + }, + { + "bbox": [ + 130, + 336, + 480, + 467 + ], + "type": "text", + "content": " under corruption, especially when corruption level is greater (e.g., Severity " + }, + { + "bbox": [ + 130, + 336, + 480, + 467 + ], + "type": "inline_equation", + "content": "= 4 / 5" + }, + { + "bbox": [ + 130, + 336, + 480, + 467 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 486, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 544 + ], + "type": "text", + "content": "Histogram Equalization for Robustness Enhancement We have further examined the impact of employing histogram equalization on the robustness of cross-view geo-localization models. In our study, we employ Contrast Limited Adaptive Histogram Equalization (CLAHE) [29] to enhance the robustness of existing methods. Fig. 7 shows the experimental results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 640 + ], + "type": "text", + "content": "It is evident that training models only on data that has undergone CLAHE does not greatly improve their robustness. On the other hand, our training strategy 3, which combines CLAHE and clean data in equal proportions, can somewhat increase the robustness of these models, though the improvement is less than that of stylization-based methods. It's also noteworthy that different models respond differently to the same training set of data. This emphasizes the importance of model robustness and the challenge of attracting researchers to invest in model robustness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "We must emphasize that how to improve model robustness is still an open topic, which requires the efforts of more researchers." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 448, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 265, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 265, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 265, + 127 + ], + "type": "text", + "content": "6 Further Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 297 + ], + "type": "text", + "content": "Real-world corruption can be much more extensive and complex. Although we cannot enumerate all possible real-world corruptions, we have systematically designed 16 corruption types, each with 5 severity levels. These types can provide a large amount of rich data to support machine learning and computer vision research. Compared to images that must be captured in the real world, synthetic imagery can provide highly controllable and diverse data for research by adjusting weather conditions, severity, and scene types as needed. At the same time, it circumvents the security risks that may be associated with capturing images of real scenes under adverse weather conditions, as well as the privacy and ethical issues that may be involved in capturing in specific environments. Therefore, we strongly believe that our benchmarks can serve as a practical testbed to perform controllable robustness evaluation, which in turn will advance research on cross-view geo-localization robustness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 369 + ], + "type": "text", + "content": "Our focus is on evaluating how well cross-view geo-localization models withstand the corruption of input images. However, the corruption robustness benchmarks we present can be applied to more research topics. For example, robustness evaluations and enhancements on cross-view image synthesis [30,31,37,38], cross-view camera pose estimation [19, 32, 41] and autonomous driving [14, 26] can also benefit from our benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 388, + 220, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 220, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 220, + 401 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "type": "text", + "content": "This paper systematically investigates the impact of corruption data on cross-view geo-localization models, which is a challenge previously overlooked in the context of cross-view geo-localization studies. In contrast to \"clean\" data, \"corrupted\" data is more common and closer to the realistic world, and exhaustive experiments have shown these models invariably experience significant performance degradation in the face of corrupted data. This further validates the necessity and importance of robustness benchmarking for cross-view geo-localization. To address this crucial issue and track future developments, we propose two fine-grained corruption robustness benchmarks (CVUSA-C and CVACT_val-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACT_val-C-ALL, and CVACT_test-C-ALL) for the cross-view geo-localization task. Extensive experiments are conducted to evaluate existing classical methods on these corruption robustness benchmarks, revealing new insights. Furthermore, we introduce two simple techniques (stylization and histogram equalization) and the training strategy to effectively enhance robustness, without requiring any model adjustments or introducing additional training complexity. We hope that the corruption robustness benchmarks, in-depth analysis, and insightful findings presented in this paper will raise awareness within the community regarding the robustness of cross-view geo-localization models and contribute to enhancing their robustness to challenges in complex real-world environments in the future." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Q. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62072318 and Grant U22A2079, in part by the Key Project of Department of Education of Guangdong Province under Grant 2023ZDZX1016, and in part by Shenzhen Science and Technology Program under Grant 20220810142553001." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 251, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 251, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 286 + ], + "type": "text", + "content": "1. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 287, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 330 + ], + "type": "text", + "content": "2. Chaabane, M., Gueguen, L., Trabelsi, A., Beveridge, R., O'Hara, S.: End-to-end learning improves static object geo-localization from video. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2063-2072 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 331, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 374 + ], + "type": "text", + "content": "3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "type": "text", + "content": "4. Dong, Y., Kang, C., Zhang, J., Zhu, Z., Wang, Y., Yang, X., Su, H., Wei, X., Zhu, J.: Benchmarking robustness of 3d object detection to common corruptions in autonomous driving. arXiv preprint arXiv:2303.11040 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 409, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 481, + 453 + ], + "type": "text", + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 481, + 486 + ], + "type": "text", + "content": "6. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "text", + "content": "7. Friedland, G., Vinyals, O., Darrell, T.: Multimodal location estimation. In: Proceedings of the 18th ACM international conference on Multimedia. pp. 1245-1252 (2010)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "text", + "content": "8. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3855-3863 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "type": "text", + "content": "9. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 587, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 481, + 620 + ], + "type": "text", + "content": "0. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "text", + "content": "11. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. International Conference on Learning Representations, International Conference on Learning Representations (Sep 2018)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 166, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "12. Geirhos, R., Temme, C.R., Rauber, J., Schütt, H.H., Bethge, M., Wichmann, F.A.: Generalisation in humans and deep neural networks. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "13. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 580-587 (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "text", + "content": "14. Hane, C., Heng, L., Lee, G.H., Fraundorfer, F., Furgale, P., Sattler, T., Pollefeys, M.: 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image and Vision Computing 68, 14-27 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 480, + 257 + ], + "type": "text", + "content": "15. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence 33(12), 2341-2353 (2010)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 480, + 278 + ], + "type": "text", + "content": "16. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 480, + 311 + ], + "type": "text", + "content": "17. Hu, S., Feng, M., Nguyen, R.M., Lee, G.H.: Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7258-7267 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 480, + 342 + ], + "type": "text", + "content": "18. Kamann, C., Rother, C.: Benchmarking the robustness of semantic segmentation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8828-8838 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 480, + 374 + ], + "type": "text", + "content": "19. Lentsch, T., Xia, Z., Caesar, H., Kooij, J.F.: Slicematch: Geometry-guided aggregation for cross-view pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17225-17234 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 480, + 407 + ], + "type": "text", + "content": "20. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2736-2744 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 480, + 439 + ], + "type": "text", + "content": "21. Liu, L., Li, H.: Lending orientation to neural networks for cross-view geolocation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5624-5633 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 133, + 439, + 480, + 471 + ], + "type": "text", + "content": "22. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing 27(6), 3064-3073 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 471, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 471, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 133, + 471, + 480, + 503 + ], + "type": "text", + "content": "23. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 503, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 503, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 133, + 503, + 480, + 546 + ], + "type": "text", + "content": "24. McManus, C., Churchill, W., Maddern, W., Stewart, A.D., Newman, P.: Shady dealings: Robust, long-term visual localisation using illumination invariance. In: 2014 IEEE international conference on robotics and automation (ICRA). pp. 901-906. IEEE (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 546, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 546, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 133, + 546, + 480, + 579 + ], + "type": "text", + "content": "25. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 579, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 579, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 579, + 480, + 621 + ], + "type": "text", + "content": "26. Middelberg, S., Sattler, T., Untzelmann, O., Kobbelt, L.: Scalable 6-dof localization on mobile devices. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13. pp. 268-283. Springer (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 480, + 665 + ], + "type": "text", + "content": "27. Mithun, N.C., Minhas, K.S., Chiu, H.P., Oskiper, T., Sizintsev, M., Samarasekera, S., Kumar, R.: Cross-view visual geo-localization for outdoor augmented reality. In: 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR). pp. 493-502. IEEE (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Q. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 171 + ], + "type": "text", + "content": "28. Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A., Greer, T., ter Haar Romeny, B., Zimmerman, J.B., Zuiderveld, K.: Adaptive histogram equalization and its variations. Computer Vision, Graphics, and Image Processing p. 355-368 (Aug 1987). https://doi.org/10.1016/s0734-189x(87)80186-x, http://dx.doi.org/10.1016/s0734-189x(87)80186-x" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "29. Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A., Greer, T., ter Haar Romeny, B., Zimmerman, J.B., Zuiderveld, K.: Adaptive histogram equalization and its variations. Computer vision, graphics, and image processing 39(3), 355-368 (1987)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "30. Regmi, K., Borji, A.: Cross-view image synthesis using conditional gans. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 3501-3510 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "31. Shi, Y., Campbell, D., Yu, X., Li, H.: Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 10009-10022 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "text", + "content": "32. Shi, Y., Li, H.: Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17010-17020 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "33. Shi, Y., Liu, L., Yu, X., Li, H.: Spatial-aware feature aggregation for image based cross-view geo-localization. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "34. Shi, Y., Yu, X., Campbell, D., Li, H.: Where am i looking at? joint location and orientation estimation by cross-view matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4064-4072 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "35. Shi, Y., Yu, X., Liu, L., Zhang, T., Li, H.: Optimal feature transport for cross-view image geo-localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11990-11997 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 414, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 434 + ], + "type": "text", + "content": "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (Jan 2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "37. Tang, H., Xu, D., Yan, Y., Torr, P.H., Sebe, N.: Local class-specific and global image-level generative adversarial networks for semantic-guided scene generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7870-7879 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "38. Token, A., Zhou, Q., Maximov, M., Leal-Taixe, L.: Coming down to earth: Satellite-to-street view synthesis for geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6488-6497 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 513, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 481, + 533 + ], + "type": "text", + "content": "39. Vasiljevic, I., Chakrabarti, A., Shakhnarovich, G.: Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "40. Workman, S., Souvenir, R., Jacobs, N.: Wide-area image geolocation with aerial reference imagery. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3961-3969 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "41. Xia, Z., Booij, O., Manfredi, M., Kooij, J.F.: Visual cross-view metric localization with dense uncertainty estimates. In: European Conference on Computer Vision. pp. 90-106. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "42. Yang, H., Lu, X., Zhu, Y.: Cross-view geo-localization with layer-to-layer transformer. Advances in Neural Information Processing Systems 34, 29009-29020 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "43. Zhang, X., Li, X., Sultani, W., Zhou, Y., Wshah, S.: Cross-view geo-localization via learning disentangled geometric layout correspondence. arXiv preprint arXiv:2212.04074 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Cross-view Geo-localization Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 183 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "44. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "45. Zhu, S., Shah, M., Chen, C.: Transgeo: Transformer is all you need for cross-view image geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1162-1171 (2022)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Q. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_content_list.json b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..50dbe87ca63b95afea63745ffb70d1c2d1754cce --- /dev/null +++ b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zicong Fan $^{1,2*}$ , Takehiko Ohkawa $^{3*}$ , Linlin Yang $^{4*}$ , Nie Lin $^{3}$ , Zhishan Zhou $^{5}$ , Shihao Zhou $^{5}$ , Jiajun Liang $^{5}$ , Zhong Gao $^{6}$ , Xuanyang Zhang $^{6}$ , Xue Zhang $^{7}$ , Fei Li $^{7}$ , Liu Zheng $^{8}$ , Feng Lu $^{8}$ , Karim Abou Zeid $^{9}$ , Bastian Leibe $^{9}$ , Jeongwan On $^{10}$ , Seungryul Baek $^{10}$ , Aditya Prakash $^{11}$ , Saurabh Gupta $^{11}$ , Kun He $^{12}$ , Yoichi Sato $^{3}$ , Otmar Hilliges $^{1}$ , Hyung Jin Chang $^{13}$ , Angela Yao $^{14}$", + "bbox": [ + 259, + 210, + 746, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We interact with the world with our hands and see it through our own (egocentric) perspective. A holistic 3D understanding of such interactions from egocentric views is important for tasks in robotics, AR/VR, action recognition and motion generation. Accurately reconstructing such interactions in 3D is challenging due to heavy occlusion, viewpoint bias, camera distortion, and motion blur from the head movement. To this end, we designed the HANDS23 challenge based on the AssemblyHands and ARCTIC datasets with carefully designed training and testing splits. Based on the results of the top submitted methods and more recent baselines on the leaderboards, we perform a thorough analysis on 3D hand(-object) reconstruction tasks. Our analysis demonstrates the effectiveness of addressing distortion specific to egocentric cameras, adopting high-capacity transformers to learn complex hand-object interactions, and fusing predictions from different views. Our study further reveals challenging scenarios intractable with state-of-the-art methods, such as fast hand motion, object reconstruction from narrow egocentric views, and close contact between two hands and objects. Our efforts will enrich the community's knowledge foundation and facilitate future hand studies on egocentric hand-object interactions.", + "bbox": [ + 261, + 340, + 743, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 667, + 375, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We interact with the world with our hands and see it through our eyes: we wake up and grab our phone to check the time; we use tools when assembling parts of a car; we open the microwave door to heat food, to name a few. An egocentric 3D understanding of our hand interactions with objects will fundamentally impact", + "bbox": [ + 212, + 700, + 787, + 762 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{*}$ Equal contribution, $^{1}$ ETH Zürich, $^{2}$ Max Planck Institute for Intelligent Systems, Tübingen, Germany, $^{3}$ The University of Tokyo, $^{4}$ Communication University of China, $^{5}$ Jiiov, $^{6}$ Bytedance, $^{7}$ Fujitsu Research & Development Center Co., Ltd., $^{8}$ Beihang University, $^{9}$ RWTH Aachen University, $^{10}$ UNIST, $^{11}$ UIUC, $^{12}$ Meta, $^{13}$ University of Birmingham, $^{14}$ National University of Singapore.", + "bbox": [ + 228, + 768, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/467e6a201dc213513725c5c616426b8a315bc6ff1a58b39b1f4df0e30a92146d.jpg", + "image_caption": [ + "Fig. 1: Tasks in HANDS23. In AssemblyHands, from its multi-view headset (a), we estimate 3D hand poses from images (b); In ARCTIC, we estimate the poses of two hands and articulated objects from an image (c)." + ], + "image_footnote": [], + "bbox": [ + 307, + 143, + 699, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "areas such as robotics grasping [11,80], augmented and virtual reality [25], action recognition [7,19,72] and motion generation [79,80].", + "bbox": [ + 212, + 306, + 782, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, it is non-trivial to accurately reconstruct 3D hands and or objects due to its high degree of freedoms [65, 73], ambiguous texture [16], and heavy occlusions. These challenges are intensified in an egocentric view [50, 51], particularly with object interactions, due to significant camera distortion, rapid and varied changes in viewpoint caused by head movements and hand-object occlusion. To better understand these challenges, we introduce a public challenge in conjunction with ICCV 2023 (i.e., HANDS23) based on recent egocentric hand datasets, AssemblyHands [49] and ARCTIC [17] (see Figure 1). These two datasets are large-scale, multi-view, and provide monochrome or RGB egocentric videos of the hands dexterously manipulating objects. Accordingly, we host two tasks: 1) egocentric 3D hand pose estimation from a single-view image based on AssemblyHands, and 2) consistent motion reconstruction based on ARCTIC.", + "bbox": [ + 212, + 338, + 784, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce new HANDS23 methods and recent dataset leaderboard baselines that substantially outperform initial baselines for both tasks, setting new benchmarks for subsequent comparisons on the datasets. The two datasets are significantly larger and include a wider variety of bimanual manipulations compared to earlier datasets [6, 23], enabling a more authentic assessment of real-world interactions. With these benchmarks, we thoroughly analyze factors such as viewpoint, action types, hand position, model size, and object variations to determine their impact on 3D hand(-object) reconstruction.", + "bbox": [ + 212, + 520, + 784, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our findings show the success of addressing the distortion of egocentric cameras with explicit perspective cropping or implicit learning for the distortion bias. In addition, recent high-capacity vision transformers can learn complex hand-object interactions. Adaptive fusion techniques for multi-view predictions further boost performance. We also analyze the remaining challenges that are still difficult to handle with the recent methods, e.g., fast hand motion, object reconstruction from narrow and moving views, and intricate interactions and close contact between two hands and objects.", + "bbox": [ + 212, + 642, + 785, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we contribute state-of-the-art baselines and gather the submitted methods for AssemblyHands and ARCTIC to foster future research on egocentric hand-object interactions. Furthermore, we thoroughly analyze the two benchmarks to provide insights for future directions in egocentric hand pose estimation and consistent motion reconstruction.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 269, + 114, + 498, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D hand pose estimation: Reconstructing 3D hand poses has a long history [14,48] ever since the important work led by Rehg and Kanade [57]. A large body of research in this area focuses on single-hand reconstruction [3,4,6,9,13, 16,18,20,28,32,37,46,63-66,71,81,87,88]. For example, while a popular OpenPose library features 2D hand keypoints [63], Zimmermann et al. [88] initially extend to estimate 3D hand poses using deep convolutional networks. Ever since the release of InterHand2.6M [45] dataset, the community has increased focus on strongly interacting two hands reconstruction [22,33-36,42-45,49]. For example, Li et al. [35] and Moon et al. [44] use relighting techniques to augment InterHand2.6M with more natural lighting and diverse backgrounds.", + "bbox": [ + 212, + 172, + 787, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hand-object reconstruction: The holistic reconstruction of hands and objects have increased interest in the hand community in recent years [6,12,17,21, 26-28,33,39,69,70,74,82]. Methods in this area mostly assume a given object template and jointly estimate the hand poses and the object rigid poses [12, 21,26-28,39,69,70,74,82] or articulated poses [17]. For example, Cao et al. [5] fits object templates to in-the-wild interaction videos. Liu et al. [39] introduce a semi-supervised learning framework via pseudo-groundtruth from temporal data to improve hand-object reconstruction.", + "bbox": [ + 212, + 325, + 787, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recent methods do not assume object templates for hand-object reconstruction [10, 15, 28, 31, 68, 75, 76]. For example, Fan et al. [15] introduced the first category-agnostic method that reconstructs an articulated hand and object jointly from a video. Nonetheless, our challenges and insights on hand-object occlusions and camera distortions are still applicable to these more challenging template-free reconstruction settings.", + "bbox": [ + 212, + 446, + 787, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Public reports for the previous challenges (HANDS17 [77] and HANDS19 [2]) distilled the insights from individual review papers and practical techniques into comprehensive summaries to enrich the community's knowledge base. These past challenges use benchmarks which include depth-based hand pose estimation from egocentric views. Instead of depth sensors, the HANDS23 benchmarks are based on affordable and widely applicable image sensors, i.e., RGB and monochrome images. This paper further advances the analysis with unique insights, such as multi-view egocentric cameras, object reconstruction in contact, and modeling with recent transformers beyond conventional CNNs.", + "bbox": [ + 212, + 536, + 787, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 HANDS23 challenge overview", + "text_level": 1, + "bbox": [ + 215, + 691, + 542, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The workshop contains two hand-related 3D reconstruction challenges in hand-object strongly interacting settings. In this section, we introduce the two challenges and their evaluation criteria.", + "bbox": [ + 212, + 720, + 785, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Workshop challenges", + "text_level": 1, + "bbox": [ + 215, + 787, + 434, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D hand pose estimation in AssemblyHands: As illustrated in Figure 1, this task focuses on egocentric 3D hand pose estimation from a single-view im", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Benchmarks and Challenges for Egocentric Hand Interactions", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "age based on AssemblyHands. The dataset provides multi-view captured videos of hand-object interaction while assembling and disassembling toy vehicles. In particular, it provides allocentric (fixed-view) and egocentric recordings and auxiliary cues like action, object, or context information for hand pose estimation. We refer readers to [49] for more dataset details. The training, validation, and testing sets contain 383K, 32K, and 62K monochrome images, respectively, all captured from egocentric cameras. During training, 3D hand keypoint coordinates, hand bounding boxes and camera intrinsic and extrinsic matrices for the four egocentric cameras attached to the headset are provided. The same information is provided during testing, minus the 3D keypoints. Unlike [49], given the availability of multi-view egocentric images, this task lets participants develop multi-view fusion based on the corresponding multi-view images.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consistent motion reconstruction in ARCTIC: Given an RGB image, the goal of this task is to estimate poses of hands and articulated objects to recover the 3D surfaces of the interaction (see Figure 1). We refer readers to [17] for more details. The ARCTIC dataset contains data of hands dexterously manipulating articulated objects and videos from $8 \\times$ allocentric views and $1 \\times$ egocentric views. The official splits of the ARCTIC dataset are used for training, validation, and testing. There are two sub-tasks: allocentric task and egocentric task. In the former, only allocentric images can be used for training and evaluation. For the latter, all images from the training set can be used for training while only the egocentric view images are used during evaluation.", + "bbox": [ + 212, + 329, + 787, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Evaluation criteria", + "text_level": 1, + "bbox": [ + 215, + 508, + 415, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "AssemblyHands evaluation: We use MPJPE as an evaluation metric in millimeters, comparing the model predictions against the ground-truth in world coordinates. We provide the intrinsic and extrinsic of the egocentric cameras to construct submission results defined in the world coordinates. Assuming that the human hand has a total $N_J$ joints, we denote wrist-relative coordinates of the prediction and ground-truth as $\\hat{J} \\in \\mathbb{R}^{N_J \\times 3}$ and $J \\in \\mathbb{R}^{N_J \\times 3}$ , respectively. Given a joint visibility indicator $\\gamma_i$ per joint $J_i$ , we compute the Euclidean distance between predicted and ground-truth joints as $\\frac{1}{\\sum_{i=1}^{N_J} \\gamma_i} \\sum_{i=1}^{N_J} \\gamma_i \\left\\| \\hat{J}_i - J_i \\right\\|_2$ . The visibility indicator offers per-joint binary labels, representing whether the annotated keypoints are visible from the given egocentric view.", + "bbox": [ + 212, + 540, + 787, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ARCTIC evaluation: Since the original ARCTIC paper [17] has a heavy focus on the quality of hand-object contact in the reconstructed hand and object meshes, we use Contact Deviation (CDev) introduced in the ARCTIC dataset as the main metric for the competition. In particular, this metric measures the extent to which a hand vertex deviates from the supposed contact object vertex in the prediction. Concretely, suppose that for a given frame, $\\{(\\mathbf{h}_i,\\mathbf{o}_i)\\}_{i = 1}^C$ are $C$ pairs of in-contact ( $< 3mm$ distance) hand-object vertices according to ground-truth, and $\\{(\\hat{\\mathbf{h}}_i,\\hat{\\mathbf{o}}_i)\\}_{i = 1}^C$ are the predictions respectively. The CDev metric is the average distance between $\\hat{\\mathbf{h}}_i$ and $\\hat{\\mathbf{o}}_i$ in millimeters, $\\frac{1}{C}\\sum_{i = 1}^{C}||\\hat{\\mathbf{h}}_i - \\hat{\\mathbf{o}}_i||$ .", + "bbox": [ + 212, + 700, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 271, + 114, + 496, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For completeness, we report all metrics introduced in ARCTIC. In particular, the task requires the reconstructed meshes to have accurate hand-object contact (CDev) and smooth motion (ACC). Additionally, during articulation or when carrying an object, it is crucial that the vertices of the hand and object which are in contact maintain synchronized movement (MDev). Moreover, we assess hand and object poses, alongside their relative movements, using metrics like MPJPE, AAE, Success Rate, and MRRPE. For detailed information, see [17].", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Methods", + "text_level": 1, + "bbox": [ + 215, + 276, + 339, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents the methods in the two challenges and other competing methods on the leaderboards. Four methods outperform the baseline in both AssemblyHands and ARCTIC.", + "bbox": [ + 212, + 308, + 785, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 AssemblyHands methods", + "text_level": 1, + "bbox": [ + 215, + 376, + 470, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Participants develop methods that learn the mapping from egocentric images to 3D keypoints. The methods are categorized into: heatmap-based and regression-based approaches. Given the presence of complex hand-object interactions in the egocentric scenes, high-capacity transformer models and attention mechanisms addressing occluded regions have been proposed as the backbone networks. Table 1 summarizes the methods based on the learning, preprocessing, multi-view fusion, and post-processing approaches.", + "bbox": [ + 212, + 402, + 785, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base: This method uses a heatmap-based framework based on heatmaps [45] with 2.5D representations [32] and a ResNet50 [29] backbone. The implementation can be found in [47].", + "bbox": [ + 212, + 507, + 785, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "JHands [85]: This method employs a regression-based approach with simple MLP heads for regressing 2D keypoints, root-relative 3D keypoints, and the global root depth. The regression training is empowered by a recent fast and strong vision transformer, Hiera [60], pre-trained with masked auto-encoder [30]. A multi-level feature fusion that concatenates the features of different layers is adopted for better feature extraction at different scales. The method additionally uses other publicly available datasets for training, namely FreiHAND [88], DexYCB [6], and CompHand [8]. The implementation is available in [83].", + "bbox": [ + 212, + 551, + 785, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PICO-AI: This method proposes a heatmap voting scheme in addition to the 2.5D heatmaps. Due to their sparsity, the conventional heatmaps pose an imbalance problem between positive and negative samples in the loss function. Hence, the proposed voting mechanism aims to spread the loss evenly across the entire heatmaps. Given the initial guess of keypoints obtained from the heatmaps, the method defines a local region centered on the joint position and operates the soft-argmax within the region to obtain refined keypoint coordinates. This restricts the impact of background points, leading to more reliable optimization. The training is facilitated by CNN-based RegNety320 [56].", + "bbox": [ + 212, + 672, + 785, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phi-AI: While following the heatmap-based approach, this method adapts MinimalHand [82] with the ResNet50 backbone, where 2D heatmaps and 3D location", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Benchmarks and Challenges for Egocentric Hand Interactions", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e9ee2538b0bb1b72a08da74aab36af1314922c7962eb13d5b35a4afce06e96ee.jpg", + "table_caption": [ + "Table 1: Method and preprocessing summary in AssemblyHands. We summarize submitted methods in terms of learning methods, architecture, preprocessing, and multi-view fusion techniques. The tuple (views, phase) indicates the number of views used in either train or test time." + ], + "table_footnote": [], + "table_body": "
MethodLearning methodsArchitecturePreprocessingMulti-view fusion (views, phase)
Base2.5D heatmaps [45]ResNet50 [29]-Simple average (4, test)
JHands [85]RegressionHiera [60]Warp perspective, color jitter, random maskAdaptive view selection and average (2, test)
PICO-AI2.5D heatmaps [45] Heatmap votingRegNety320 [56]Scale, rotate, flip, translateAdaptive view selection FTL [58] (2, train)
FRDCRegression 2D heatmapsHandOccNet [53] with ConvNeXt [41]Scale, rotate, color jitterWeighted average (4, test)
Phi-AI2D heatmaps and 3D location maps [82]ResNet50 [29]Scale, rotate, translate, color jitter, gaussian blurWeighted average (4, test)
", + "bbox": [ + 220, + 212, + 784, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "maps are regressed. Instead of selecting 3D keypoint coordinates from the location maps, the proposed method modifies it by using heatmap values to weight 3D keypoint coordinates, achieving a more robust estimation. Moreover, the method adds a residual structured layer after the original three-tier cascade networks to refine the calculated location maps. The method further applies the ensemble of final keypoint outputs combined with the Base.", + "bbox": [ + 212, + 407, + 782, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FRDC: This method adopts a hybrid approach by combining regression with heatmap for training. HandOccNet [53] is modified to regress 3D keypoint coordinates and integrated with an additional branch of 2D heatmap regression. HandOccNet enriches feature extraction with spatial attention mechanisms for occluded regions, making it robust under hand-object occlusions. The method further utilizes a stronger ConvNeXt [41] backbone and feature fusion from the 2D keypoint regressor.", + "bbox": [ + 212, + 500, + 784, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Preprocessing of egocentric images: Compared to conventional static camera setups, egocentric images exhibit unique properties and biases, such as distortion, head camera motion, and different color representations. Thus, it is vital to preprocess egocentric images to alleviate these effects during training. Augmentation techniques are detailed in Table 1.", + "bbox": [ + 212, + 608, + 784, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The method JHands addresses the distortion issue with a warp perspective operation to make the hands near the edge less stretched. While AssemblyHands provides rectified images converted from fisheye cameras to a pinhole camera model, they often include excessively stretched areas near the edges. To address this, the method calculates a virtual camera and corresponding perspective transformation matrix based on the pixel coordinates of the crop and the camera parameters. The generated crops can be found in the analysis of Figure 3.", + "bbox": [ + 212, + 686, + 785, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-view fusion: Since AssemblyHands offers multi-view egocentric videos, participants can optionally use the constraint of multi-view geometry and fusion techniques during training or inference.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 271, + 114, + 498, + 128 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3578f4de50aa661167cf6ab06e4fa3e95a49b28c649cf3ecedbcdd12531b4d95.jpg", + "table_caption": [ + "Table 2: Method and preprocessing summary in ARCTIC. We summarize baselines on ARCTIC in terms of input dimensions, image backbones, learning rate scheduling, training epochs, batch size and the cropping used for input. *Method trains 50 epochs for decoder and 36 for backbone. +Learning rate is 1e-7 to 1e-4 with linear warmup for first $5\\%$ step, and 1e-4 to 1e-7 with cosine decay for rest." + ], + "table_footnote": [], + "table_body": "
MethodInput sizeBackbonesLearning rate scheduleTraining epochsBatch sizeCropping
ArcticNet-SF [17]224 × 224ResNet501e-5allocentric: 20 \negocentric: 5064object
DIGIT [16]224 × 224HRNet-W321e-5allocentric: 20 \negocentric: 5064object
AmbiguousHands224 × 224ResNet501e-5allocentric: 20 \negocentric: 10032hand \nobject
UVHand384 × 384Swin-L2e-4 (backbone) \n1e-7 (others)allocentric: N/A \negocentric: 50/36*48object
JointTransformer [1]224 × 224ViT-G1e-7/1e-5+allocentric: 20 \negocentric: 10064object
", + "bbox": [ + 220, + 226, + 782, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While Base uses a simple average of predicted keypoints from all four camera views in the test time, PICO-AI proposes multi-view feature fusion during training using Feature Transform Layers (FTL) [58]. This FTL training requires fusing two out of four views; thus, the method chooses the most suitable views for every frame. In cases with multiple candidates, the Intersection over Union (IoU) is computed between hand boxes from per-view predictions and 2D keypoints from previous 3D predictions. The two views with the highest IoUs are selected for their superior prediction reliability.", + "bbox": [ + 212, + 385, + 782, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The methods JHands, FRDC, and Phi-AI apply adaptive fusion in predicted keypoints during testing. The method JHands computes the MPJPE with each other view and selects two results of views with the lowest MPJPE, excluding noisy predictions in the average. If the MPJPE is lower than a threshold, the mean of the two results is calculated as the final result. Otherwise, the result with a lower PA-MPJPE with the predictions in the previous frame is chosen. The methods FRDC and Phi-AI use a weighted average for each view prediction, assigning weights based on each view's validation performance.", + "bbox": [ + 212, + 506, + 784, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Postprocessing: Several postprocessing techniques, including test-time augmentation, smoothing, and model ensembling are used to enhance inference outcomes. In particular, the method JHands applies an offline smooth (Savitzky-Golay) filter on each video sequence.", + "bbox": [ + 212, + 628, + 784, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 ARCTIC methods", + "text_level": 1, + "bbox": [ + 214, + 709, + 415, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 summarizes the details for each method in terms of the input image dimensions, image backbones, learning rate scheduling, the number of training epochs, batch size, and cropping strategies.", + "bbox": [ + 212, + 733, + 785, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Preliminary: All methods below are regression-based and predict two-hand MANO [59] parameters $\\Theta = \\{\\theta, \\beta\\}$ and articulated object parameters $\\Omega$ . In particular, with the MANO pose and shape parameters $\\theta, \\beta$ , the MANO model $\\mathcal{H}$ returns a mesh with vertices via $\\mathcal{H}(\\theta, \\beta) \\in \\mathbb{R}^{778 \\times 3}$ . 3D joints are obtained via", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Benchmarks and Challenges for Egocentric Hand Interactions", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a linear regressor. The articulated object model $\\mathcal{O}$ was introduced in ARCTIC to provide an articulated mesh with vertices via $\\mathcal{O}(\\Omega)\\in \\mathbb{R}^{\\mathrm{V}\\times 3}$ , where $\\Omega \\in \\mathbb{R}^{7}$ contains the global orientation, global translation, and object articulation.", + "bbox": [ + 215, + 145, + 784, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ArcticNet-SF [17]: Introduced in ARCTIC, it is a single-frame baseline. It first extracts an image feature vector from the input image; then, it regresses hand and object parameters with simple MLPs. The hand and object meshes can then be extracted via $\\mathcal{H}(\\cdot)$ and $\\mathcal{O}(\\cdot)$ . For more details, see [17].", + "bbox": [ + 215, + 196, + 784, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "JointTransformer [1]: JointTransformer enhances ArcticNet-SF by integrating a transformer decoder instead of the MLP regressors for hand and object parameter estimation. The decoder employs learned queries for the angle of each joint, the shape and translation of each hand, and the translation, rotation, and articulation of the object. It alternates between self-attention between queries and cross-attention of queries to the elements of the backbone feature map, followed by linear layers that regress the final parameters. Specifically, separate linear layers are dedicated to regressing joint angles, hand shape, hand translation, object translation, object orientation, and object articulation. The best model uses a ViT-G [78] backbone with frozen DINOv2 weights [52].", + "bbox": [ + 215, + 261, + 784, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AmbiguousHands [55]: The method addresses scale ambiguity, resulting from bounding box cropping in data augmentation and camera intrinsics, by employing positional encoding of these elements to mitigate scale issues. This leads to improved spatial alignment. Subsequently, the approach enhances network visibility by integrating local features through distinct hand and object crops. They follow the general approach of ArcticNet-SF to regress hand/object parameters.", + "bbox": [ + 215, + 417, + 784, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "UVHand: Since ArcticNet-SF only leverages a global feature vector to estimate hand and object parameters, the image features lack local context. To address this, UVHand leverages Swin-L transformer [40] to extract image features. They then further leverage Deformable DETR [86] to encode the multiple-scale feature maps. The encoded feature maps are then aggregated via self- and cross-attention before regressing hand and object parameters.", + "bbox": [ + 215, + 512, + 784, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DIGIT [16]: The method was introduced to estimate strongly interacting hands in [16]. Since ArcticNet-SF is sensitive when the hands are interacting with objects, DIGIT was extended to the ARCTIC setting. Given an image, it first estimates hand part-wise segmentation masks and object masks. The mask predictions are fused with the image features to perform parameter estimation.", + "bbox": [ + 215, + 608, + 784, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation details: Table 2 shows that all methods use the default cropping as in ARCTIC to crop around the object, while AmbiguousHands performs three crops (around two hands and the object). DIGIT uses the HRNet-W32 backbone [67] and trains with a batch size of 64 with the same learning rate for all iterations. UVHand takes as input a $384 \\times 384$ image cropped around the object and encodes it with the Swin-L transformer [40] backbone. It was trained with a batch size of 48 with a learning rate of 2e-4 for the backbone and 1e-7 for other weights. Due to computational cost, they train 50 epochs for the decoder and 36 for the backbone. JointTransformer uses ViT-G [78] backbone with frozen DINOv2 [52] weights to train with a batch size of 64. It performs a", + "bbox": [ + 215, + 688, + 784, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 271, + 114, + 496, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dedce2be5cc15fd6c00504010db757ecdeac83d0c142b42b81ab8f451b5b9aa8.jpg", + "table_caption": [ + "Table 3: Method performance in AssemblyHands. We compare AssemblyHands method performance on egocentric test data. We show the final MPJPE on the test set as the metrics (lower better). We also provide detailed evaluations, regarding the varying distances of hand position from the image center and different verb action categories. The hand distance is computed by the distance from the image center to the hand center position per image, and averaged over the lower two views of the headset. Verb classes of \"attempt to X\" are merged to \"X\" for simplicity. The higher and lower three verbs are color-coded in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
MethodScoreHand distance (px)Verb class
-200200-250250-clapinspectpasspick uppositionposition screw onpull
Base20.6920.3121.9724.8519.8822.8922.4821.8522.6521.6218.74
JHands12.2112.3511.9813.7210.6516.2712.8613.6714.5812.811.06
PICO-AI12.4612.5111.6212.9512.9815.311.3713.213.1811.3915.13
FRDC16.4816.3915.8918.6915.2421.3317.8618.2619.2118.0312.83
Phi-AI17.2617.2415.8119.5119.8620.9317.9119.0119.719.3517.3
Verb class (continue)
pushput downremoveremove screw fromrotatescrewtilt downtilt upunscrewnone
Base19.2920.2619.9916.4722.7122.9513.1215.1120.7819.82
JHands13.9613.7213.1111.7212.2614.119.619.9212.2510.99
PICO-AI12.2913.4112.839.9913.713.7210.4411.5612.8711.71
FRDC19.5217.8718.5514.4716.4418.8114.0313.3716.4115.01
Phi-AI19.1218.2918.1813.9518.3519.7813.1315.3617.2915.8
", + "bbox": [ + 220, + 268, + 782, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "linear warmup from 1e-7 to 1e-4 in the first $5\\%$ steps and uses cosine decay from 1e-4 to 1e-7 for the rest of the steps.", + "bbox": [ + 212, + 494, + 782, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Results and analysis", + "text_level": 1, + "bbox": [ + 214, + 551, + 449, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Results", + "text_level": 1, + "bbox": [ + 214, + 585, + 326, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, we benchmark results of valid submissions for state-of-the-art comparison in AssemblyHands and ARCTIC and other more recent baselines. In particular, for AssemblyHands, we report egocentric hand pose estimation results. For ARCTIC, we report results for the allocentric and egocentric test sets.", + "bbox": [ + 212, + 612, + 785, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "AssemblyHands benchmark: Table 3 shows the final test scores on the AssemblyHands dataset. The methods in the table exceed the baseline (Base) with a test score of 20.69 MPJPE. Notably, the methods JHands and PICO-AI achieve a nearly $40\\%$ reduction over the baseline. The methods FRDC and Phi-AI improve the test score by $20.3\\%$ and $16.5\\%$ against the baseline, respectively.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ARCTIC benchmark: Table 4 presents the comparative performance of methods in the ARCTIC dataset, where ArcticNet-SF serves as the initial benchmark. The majority surpass ArcticNet-SF in both allocentric and egocentric views, except for UVHand, which underperforms due to incomplete training. In the egocentric view, AmbiguousHands excels in creating smooth, consistent mesh motions (refer to MDev and $\\mathrm{ACC}_{h / o}$ metrics). Notably, JointTransformer stands", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Benchmarks and Challenges for Egocentric Hand Interactions", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4e14a9ddd492c48dd605f2a881b3b3787ee20ff48e54e1cfd635d92794654e10.jpg", + "table_caption": [ + "Table 4: Method performance in ARCTIC. We compare performance in both allocentric (top half) and egocentric (bottom half) views. We evaluate using metrics for contact and relative position (measuring hand-object contact and prediction of relative root position), motion (assessing temporally-consistent contact and smoothness), and hand and object metrics (indicating root-relative reconstruction error). We use the CDev score as the main metric for this competition. We denote left and right hands as $l$ and $r$ , and the object as $o$ ." + ], + "table_footnote": [], + "table_body": "
Contact and Relative PositionsMotionHandObject
MethodCDev [mm]↓ MRRPEa/ro [mm]↓ MDev [mm]↓ ACCh/α [m/s2]↓ MPJPE [mm]↓ AAE [°] ↓ Success Rate [%] ↑
AllocentricArcticNet-SF41.5652.39/37.4710.405.72/7.5721.455.3771.39
DIGIT34.9244.19/35.438.374.86/6.6317.925.2476.52
UVHand64.1584.68/70.3114.127.05/12.0440.9912.3631.47
AmbiguousHands33.2545.78/34.5610.126.37/6.4018.024.6481.94
JointTransformer27.9736.17/28.188.936.08/5.7917.123.9589.79
Ego-centricArcticNet-SF44.7128.31/36.1611.805.03/9.1519.186.3953.89
DIGIT41.3125.49/32.619.484.01/8.3216.746.6053.33
UVHand40.4340.93/36.889.965.32/8.3324.537.3257.28
AmbiguousHands35.9323.07/27.539.513.95/6.7616.264.8668.36
JointTransformer32.5626.07/26.2211.345.52/8.6816.334.4474.07
", + "bbox": [ + 218, + 255, + 784, + 369 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f70b708de24ac2a015506912d092aab7fbd6af0ae3f5fb3de06d9c9f40476a16.jpg", + "image_caption": [ + "Fig. 2: Qualitative results per action in AssemblyHands. We show Base results with \"verb (noun)\" actions. The left three figures are lower error situations while the right four ones are failure cases. The red boxes denote the area where the action occurs." + ], + "image_footnote": [], + "bbox": [ + 218, + 385, + 784, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "out by significantly lowering CDev errors by $32.7\\%$ in allocentric and $27.2\\%$ in egocentric settings compared to the baseline.", + "bbox": [ + 212, + 536, + 787, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 AssemblyHands analysis", + "text_level": 1, + "bbox": [ + 214, + 594, + 464, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We provide analysis, regarding action-wise evaluation, distortion effect in training, and the effect of multi-view fusion. See SupMat for additional results.", + "bbox": [ + 212, + 626, + 785, + 656 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Action-wise evaluation: To analyze errors related to hand-object occlusions and interactions, we show pose evaluation according to fine verb action classes in Table 3. We use the verb classes annotated by Assembly101 [61], spanning every few seconds in a video. Figure 2 shows qualitative results of representative verb classes with the top and bottom error cases.", + "bbox": [ + 212, + 657, + 787, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We observe that the performance varies among different verb actions. The verbs \"tilt down/up\" and \"remove screw from\" exhibit lower errors among the submitted methods, because hands are less occluded and their movement is relatively stable. The \"tilt\" action holds a small part of the toy and turns it around alternately, leading to less overlap between the hand and the object (lib). The \"remove screw from\" action takes a screw out from the toy vehicle by their hand where observed hand poses do not change drastically.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 269, + 114, + 498, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bfcce42240a3fc25f8f9024cb50c2ba7df097d8e93de6e8993ef4e333b395161.jpg", + "image_caption": [ + "Fig. 3: Effect of distortion in AssemblyHands. The officially released images in the dataset have highly stretched areas near the edges (original crop). The method JHands uses a perspective crop with a virtual camera to correct this distortion." + ], + "image_footnote": [], + "bbox": [ + 230, + 145, + 570, + 271 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f4ec049a039d28e352c7815ffb0a36202d3361245380132eb81e7629bb4a4272.jpg", + "table_caption": [ + "Table 5: Multi-view fusion in AssemblyHands. We use the Base result to show performance before and after fusion. Missing instances per view are denoted as \"Miss(\\%)\"." + ], + "table_footnote": [], + "table_body": "
ViewMPJPEMiss(%)
cam137.9770.8
cam225.7188.3
cam322.190.92
cam422.290.74
cam3+421.520.08
all four20.690
", + "bbox": [ + 602, + 255, + 781, + 352 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Higher error classes, such as \"inspect\", \"screw\", \"rotate\", and \"position\", contain heavy occlusions, fast hand motion, complex two hands and object interactions. The \"inspect\" action brings the toy close to the human eyes where the toy occupies a large portion of the image causing heavy object occlusions. The \"screw\" action involves intricate fingertip movements to rotate the screwdriver quickly. The \"rotate\" and \"position\" actions are performed so that the two hands and the object interact in close contact, which complicates the estimation. We observe that the top two methods JHands and PICO-AI significantly correct the results of these higher error actions compared to the other submitted methods.", + "bbox": [ + 212, + 383, + 787, + 520 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Bias of hand position in an image: Hands near image edges are highly distorted due to the fish-eye cameras. Directly using these noisy images in training will degrade performance [84]; thus, some methods create new crops with less distortion, select training instances, or adaptively fuse predictions during the inference. Specifically, Figure 3 shows that the method JHands reformulates the perspective during cropping and creates less-distorted (perspective) crops.", + "bbox": [ + 212, + 521, + 787, + 611 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To study this effect in the final performance, we split the evaluation instances into classes with different 2D distances between the hand center and the image center in Table 3. Higher distances (250- pixels) indicate closer hand crops to image edges. The method, Base, without any training instance selection and distortion correction, has higher error as the crops approach image edges (20.31 $\\rightarrow$ 24.85). In contrast, the newly proposed methods are more robust and have a lower error, particularly in the 200-250 range. We observe that the ranges 200-250 and 250- occupy $10\\%$ and $5\\%$ of the test images, respectively, thus the improvement in the 200-250 range helps the lowering of the overall score.", + "bbox": [ + 212, + 612, + 787, + 747 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of multi-view fusion: The multi-view egocentric camera setup is unique to the dataset. We show the statistics and performance of multi-view fusion in Table 5. Note that Table 3 shows the final results after multi-view fusion.", + "bbox": [ + 212, + 750, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We found that samples captured from the lower cameras (cam3 and cam4; see Table 1 for the layout) are numerous (fewer missing samples) and their errors are lower as they are faced toward the area occurring hand interactions.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Benchmarks and Challenges for Egocentric Hand Interactions", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/de923b27e131452ecd823d6759203356406cb7483ac08fa0d6cf68d936c8f67e.jpg", + "image_caption": [ + "(a) Egocentric-allocentric comparison", + "Fig. 4: Performance comparison: Egocentric vs Allocentric. (a) Comparative difficulty ratio of egocentric to allocentric views. (b) Egocentric view performance by method across objects. (c) Allocentric view performance by method across objects." + ], + "image_footnote": [], + "bbox": [ + 272, + 155, + 500, + 337 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a3f0d36d71f802fa0824cba0f0261952046b8a21b15aec361ed4d8a11f4db8c6.jpg", + "image_caption": [ + "(b) Egocentric view" + ], + "image_footnote": [], + "bbox": [ + 501, + 155, + 728, + 239 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7d4745864fdcfb2f47fee1c2fa7d7b487c45155f8cda94e7f978014bc72387a4.jpg", + "image_caption": [ + "(c) Allocentric view" + ], + "image_footnote": [], + "bbox": [ + 501, + 250, + 727, + 335 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Conversely, the samples from cam1 and cam2 are fewer and unbalanced as their cameras often fail to capture hands due to the camera layout. For instance, cam1 (top-left) tends to capture more hand region than cam2 as the participants are mostly right-handed and bring up the object with the right hand, which can be better observed from cam1. Given this uneven sample distribution, the proposed adaptive view selection methods in either training or testing are essential to perform effective multi-view fusion, and outperform the Base's test-time average using all views all the time (see Multi-view fusion in Section 4.1).", + "bbox": [ + 212, + 421, + 784, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We further study the performance gain before and after multi-view fusion using Base's results. While per-view performance achieves 22.19 and 22.29 in cam3 and cam4, respectively, their fused results with simple average reduce the error to 21.52. Merging all four views has shown to be more effective than two-view fusion (20.69 vs. 21.52), indicating a $6.5\\%$ reduction compared to the single camera setup (cam3). This suggests predictions from the top views (cam1 and cam2) are informative in averaging even when they are prone to be erroneous.", + "bbox": [ + 212, + 542, + 785, + 650 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 ARCTIC analysis", + "text_level": 1, + "bbox": [ + 215, + 674, + 408, + 690 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Egocentric-allocentric comparison: Figure 4a compares the performance between egocentric and allocentric views. In particular, we compute a ratio between the metric values of the egocentric and allocentric view to measure the extent of difficulty for the egocentric view compared to the allocentric view. Since success rate is a metric whose value is positively correlated to performance, we take its reciprocal ratio. We average the ratios across methods and actions.", + "bbox": [ + 212, + 703, + 784, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We observe that hand pose-related metrics such as MPJPE and $\\mathrm{ACC}_h$ are less than 1.0 on average (see blue color cells), meaning the the egocentric view is easier than allocentric view. This is because most allocentric cameras in ARCTIC", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 269, + 114, + 496, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e6b68b50c2cc66b47deb604456e31050bfb31943521ba650e2124c8b0a40ddcd.jpg", + "image_caption": [ + "Fig. 5: Egocentric reconstruction by top method in ARCTIC. In the egocentric view, object reconstruction struggle when the object is partially observed on the image boundaries, as well as when heavy hand/arm occlusion occurs." + ], + "image_footnote": [], + "bbox": [ + 272, + 141, + 733, + 223 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "are meters away from the subject while the egocentric camera is often close-up, offering higher hand visibility. Relative translation metrics between hand and object such as $\\mathrm{MRRPE}_{rl}$ are also easier in the egocentric view because estimating translation is more difficult from further cameras.", + "bbox": [ + 212, + 306, + 782, + 366 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Object reconstruction performance faces unique challenges, as highlighted by the red cells. In the egocentric view, it is notably more difficult to reconstruct accurate object surfaces, articulation (AAE), and hand-object contact (CDev). This increased difficulty arises because objects are often positioned at the image edges and obscured by human arms. Additionally, object poses exhibit greater diversity in the egocentric view due to varying camera angles and occlusions. While a static camera maintains consistent camera extrinsics across a sequence, an egocentric camera's extrinsics change with each frame, resulting in higher diversity in camera-view object 6D poses. This diversity complicates object pose estimation in egocentric views. Figure 5 illustrates these challenges using the best-performing method, JointTransformer. Despite achieving reasonable hand poses, object poses are significantly impacted by occlusions from hands and arms, and the egocentric view undergoes substantial changes throughout a sequence.", + "bbox": [ + 212, + 368, + 787, + 564 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Object-wise evaluation: Figure 4b and 4c break down performances on different objects in the egocentric and allocentric view test sets. The best method in both cases is JointTransformer. The hardest object to reconstruct with good contact consistency (see CDev) is the microwave in both settings due to global rotation sensitivity (see Fig. 5), though this can be mitigated by a keypoint-based approach [24]. Estimating objects in the egocentric view is more difficult than in the allocentric view, which is indicated by higher errors for all methods.", + "bbox": [ + 212, + 566, + 787, + 671 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Action-wise evaluation: Figure 6 compares performance of different methods in \"grab\" and \"use\" actions. In ARCTIC, there are sequences to interact with the object with two types of actions by either not articulating the object, or allowing object articulation. Interestingly, the \"grab\" motion is more challenging in egocentric and allocentric views. We hypothesize that this is because there are more diverse object poses for \"grab\" motions since during object articulation, the participants often focus on articulation instead of object manipulation.", + "bbox": [ + 212, + 672, + 787, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of model size: Figure 7 illustrates the impact of model size on handobject contact performance, measured by CDev, for reconstruction on the allocentric validation set. Most methods utilize ResNet50, with JointTransformer being the top performer. As the trainable parameters in the backbone increase,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Benchmarks and Challenges for Egocentric Hand Interactions", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/55f0d09062b13f1d742c0f0e963fe0c3739330130d02190cec3aca6a252904bd.jpg", + "image_caption": [ + "Fig.6: Hand-object contact quality for reconstructed results per action. We evaluate the contact quality of the 3D reconstruction results from all methods for each action (i.e., grab or use), using Contact Deviation (CDev) in mm as the metric, where lower values indicate better quality." + ], + "image_footnote": [], + "bbox": [ + 217, + 164, + 500, + 311 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/607b42b7cc78a95cfcacfeb6cfc77bece5b5b1d9e1c697ac7d6c6cf3c6d29189.jpg", + "image_caption": [ + "Fig. 7: Contact deviation vs. model size. We assess the contact quality of the reconstruction results, varying by the number of parameters in each model. Contact quality is measured using Contact Deviation (CDev) in mm, with lower values indicating superior results." + ], + "image_footnote": [], + "bbox": [ + 517, + 146, + 781, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "JointTransformer consistently reduces the CDev error. Note that the x-axis is in log-scale. JointTransformer achieved a CDev error of 30.5mm with ViT-L and 29.0mm with ViT-G, which has ten times more parameters than ViT-L. Interestingly, JointTransformer uses frozen weights in the large-scale ViT-L and ViT-G backbones, yet achieves the best results. This suggests a potential direction for leveraging large-scale foundational backbones for hand-object reconstruction.", + "bbox": [ + 212, + 452, + 787, + 544 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 566, + 359, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce the HANDS23 challenge and provide analysis based on the results of the top submitted methods and more recent baselines on the leaderboards. We organize and compare the submissions and their implementation details based on the learning methods, architecture, pre- and post-processing techniques, and training configurations. We thoroughly analyze various aspects, such as hand-object occlusions, action and object-wise evaluation, distortion correction, multi-view fusion, egocentric-allocentric comparison, and performance gain of large transformer models.", + "bbox": [ + 212, + 597, + 785, + 718 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future directions: There are several future directions that the community can take. For example, one can explore more efficient training using multi-view egocentric cameras, leveraging 3D foundation priors [38,54] to regularize template-free hand-object reconstruction [15], estimating hand-object poses with more expressive representations (e.g., heatmap-based approaches [24]), incorporating motion and temporal modeling [18], featuring more diverse egocentric interaction scenarios, recognizing actions through captured hand poses [62], learning robotic grasping from reconstructed hand-object pose sequences, and so forth.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 271, + 114, + 496, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the Ministry of Science and ICT, Korea, under the ITRC program (IITP-2024-2020-0-01789). supervised by the IITP. UTokyo is supported by JST ACT-X Grant Number JPMJAX2007, JSPS KAKENHI Grant Number JP22KJ0999, JST Adopting Sustainable Partnerships for Innovative Research Ecosystem (ASPIRE) Grant Number JPMJAP2303. NUS is supported by the Ministry of Education, Singapore, under its MOE Academic Research Fund Tier 2 (STEM RIE2025 MOE-T2EP20220-0015).", + "bbox": [ + 212, + 174, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 301, + 321, + 316 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abou Zeid, K.: JointTransformer: Winner of the HANDS'2023 ARCTIC Challenge @ ICCV. https://github.com/kabouzeid/JointTransformer (2023) 7, 8", + "2. Armagan, A., Garcia-Hernando, G., Baek, S., Hampali, S., Rad, M., Zhang, Z., Xie, S., Chen, M., Zhang, B., Xiong, F., et al.: Measuring generalisation to unseen viewpoints, articulations, shapes and objects for 3d hand pose estimation under hand-object interaction. In: ECCV. pp. 85-101. Springer (2020) 3", + "3. Boukhayma, A., de Bem, R., Torr, P.H.S.: 3D hand shape and pose from images in the wild. In: Computer Vision and Pattern Recognition (CVPR). pp. 10843-10852 (2019) 3", + "4. Cai, Y., Ge, L., Cai, J., Yuan, J.: Weakly-supervised 3D hand pose estimation from monocular RGB images. In: European Conference on Computer Vision (ECCV). pp. 678-694 (2018) 3", + "5. Cao, Z., Radosavovic, I., Kanazawa, A., Malik, J.: Reconstructing hand-object interactions in the wild. In: International Conference on Computer Vision (ICCV). pp. 12417-12426 (2021) 3", + "6. Chao, Y.W., Yang, W., Xiang, Y., Molchanov, P., Handa, A., Tremblay, J., Narang, Y.S., Van Wyk, K., Iqbal, U., Birchfield, S., Kautz, J., Fox, D.: DexYCB: A benchmark for capturing hand grasping of objects. In: Computer Vision and Pattern Recognition (CVPR). pp. 9044-9053 (2021) 2, 3, 5", + "7. Chatterjee, D., Sener, F., Ma, S., Yao, A.: Opening the vocabulary of egocentric actions. Conference on Neural Information Processing Systems (NeurIPS) 36 (2024) 2", + "8. Chen, X., Liu, Y., Dong, Y., Zhang, X., Ma, C., Xiong, Y., Zhang, Y., Guo, X.: MobRecon: mobile-friendly hand mesh reconstruction from monocular image. In: Computer Vision and Pattern Recognition (CVPR). pp. 20512-20522 (2022) 5", + "9. Chen, X., Wang, B., Shum, H.Y.: Hand avatar: Free-pose hand animation and rendering from monocular video. In: Computer Vision and Pattern Recognition (CVPR) (2023) 3", + "0. Chen, Z., Chen, S., Schmid, C., Laptev, I.: gSDF: Geometry-driven signed distance functions for 3d hand-object reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 12890-12900 (2023) 3", + "1. Christen, S., Kocabas, M., Aksan, E., Hwangbo, J., Song, J., Hilliges, O.: D-Grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In: Computer Vision and Pattern Recognition (CVPR). pp. 20545–20554 (2022) 2", + "2. Corona, E., Pumarola, A., Alenyà, G., Moreno-Noguer, F., Rogez, G.: GanHand: Predicting human grasp affordances in multi-object scenes. In: Computer Vision and Pattern Recognition (CVPR). pp. 5030-5040 (2020) 3" + ], + "bbox": [ + 225, + 330, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Benchmarks and Challenges for Egocentric Hand Interactions", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Duran, E., Kocabas, M., Choutas, V., Fan, Z., Black, M.J.: HMP: Hand motion priors for pose and shape estimation from video. In: Winter Conference on Applications of Computer Vision (WACV) (2024) 3", + "14. Erol, A., Bebis, G., Nicoescu, M., Boyle, R.D., Twombly, X.: Vision-based hand pose estimation: A review. Computer Vision and Image Understanding (CVIU) 108(1-2), 52-73 (2007) 3", + "15. Fan, Z., Parelli, M., Kadoglou, M.E., Kocabas, M., Chen, X., Black, M.J., Hilliges, O.: HOLD: Category-agnostic 3d reconstruction of interacting hands and objects from video. In: Computer Vision and Pattern Recognition (CVPR) (2024) 3, 14", + "16. Fan, Z., Spurr, A., Kocabas, M., Tang, S., Black, M.J., Hilliges, O.: Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In: International Conference on 3D Vision (3DV). pp. 1-10 (2021) 2, 3, 7, 8", + "17. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 4, 5, 7, 8", + "18. Fu, Q., Liu, X., Xu, R., Niebles, J.C., Kitani, K.M.: Deformer: Dynamic fusion transformer for robust hand pose estimation. In: International Conference on Computer Vision (ICCV). pp. 23600-23611 (October 2023) 3, 14", + "19. Garcia-Hernando, G., Yuan, S., Baek, S., Kim, T.K.: First-person hand action benchmark with rgb-d videos and 3d hand pose annotations. In: Computer Vision and Pattern Recognition (CVPR) (2018) 2", + "20. Ge, L., Ren, Z., Li, Y., Xue, Z., Wang, Y., Cai, J., Yuan, J.: 3D hand shape and pose estimation from a single rgb image. In: Computer Vision and Pattern Recognition (CVPR). pp. 10833-10842 (2019) 3", + "21. Grady, P., Tang, C., Twigg, C.D., Vo, M., Brahmbhatt, S., Kemp, C.C.: ContactOpt: Optimizing contact to improve grasps. In: Computer Vision and Pattern Recognition (CVPR). pp. 1471-1481 (2021) 3", + "22. Guo, Z., Zhou, W., Wang, M., Li, L., Li, H.: HandNeRF: Neural radiance fields for animatable interacting hands. In: Computer Vision and Pattern Recognition (CVPR). pp. 21078-21087 (2023) 3", + "23. Hampali, S., Rad, M., Oberweger, M., Lepetit, V.: HOnnotate: A method for 3D annotation of hand and object poses. In: Computer Vision and Pattern Recognition (CVPR). pp. 3193-3203 (2020) 2", + "24. Hampali, S., Sarkar, S.D., Rad, M., Lepetit, V.: Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3D pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 11090-11100 (2022) 13, 14", + "25. Han, S., Wu, P.C., Zhang, Y., Liu, B., Zhang, L., Wang, Z., Si, W., Zhang, P., Cai, Y., Hodan, T., Cabezas, R., Tran, L., Akbay, M., Yu, T.H., Keskin, C., Wang, R.: Umetrack: Unified multi-view end-to-end hand tracking for VR. In: International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH). pp. 50:1-50:9. ACM (2022) 2", + "26. Hasson, Y., Tekin, B., Bogo, F., Laptev, I., Pollefeys, M., Schmid, C.: Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 568-577 (2020) 3", + "27. Hasson, Y., Varol, G., Schmid, C., Laptev, I.: Towards unconstrained joint handobject reconstruction from rgb videos. In: International Conference on 3D Vision (3DV). pp. 659-668. IEEE (2021) 3" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 271, + 114, + 496, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Hasson, Y., Varol, G., Tzionas, D., Kalevatykh, I., Black, M.J., Laptev, I., Schmid, C.: Learning joint reconstruction of hands and manipulated objects. In: Computer Vision and Pattern Recognition (CVPR). pp. 11807-11816 (2019) 3", + "29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) 5, 6", + "30. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: Computer Vision and Pattern Recognition (CVPR). pp. 15979-15988 (2022) 5", + "31. Huang, D., Ji, X., He, X., Sun, J., He, T., Shuai, Q., Ouyang, W., Zhou, X.: Reconstructing hand-held objects from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) 3", + "32. Iqbal, U., Molchanov, P., Gall, T.B.J., Kautz, J.: Hand pose estimation via latent 2.5D heatmap regression. In: European Conference on Computer Vision (ECCV). pp. 118-134 (2018) 3, 5", + "33. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2O: Two hands manipulating objects for first person interaction recognition. In: International Conference on Computer Vision (ICCV). pp. 10138-10148 (2021) 3", + "34. Lee, J., Sung, M., Choi, H., Kim, T.K.: Im2hands: Learning attentive implicit representation of interacting two-hand shapes. In: Computer Vision and Pattern Recognition (CVPR). pp. 21169-21178 (2023) 3", + "35. Li, L., Tian, L., Zhang, X., Wang, Q., Zhang, B., Bo, L., Liu, M., Chen, C.: Renderih: A large-scale synthetic dataset for 3d interacting hand pose estimation. In: International Conference on Computer Vision (ICCV). pp. 20395-20405 (2023) 3", + "36. Li, M., An, L., Zhang, H., Wu, L., Chen, F., Yu, T., Liu, Y.: Interacting attention graph for single image two-hand reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 2761-2770 (2022) 3", + "37. Liu, R., Ohkawa, T., Zhang, M., Sato, Y.: Single-to-dual-view adaptation for egocentric 3d hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 677-686 (2024) 3", + "38. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023) 14", + "39. Liu, S., Jiang, H., Xu, J., Liu, S., Wang, X.: Semi-supervised 3D hand-object poses estimation with interactions in time. In: Computer Vision and Pattern Recognition (CVPR). pp. 14687-14697 (2021) 3", + "40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: International Conference on Computer Vision (ICCV). pp. 10012-10022 (2021) 8", + "41. Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Computer Vision and Pattern Recognition (CVPR). pp. 11966-11976 (2022) 6", + "42. Meng, H., Jin, S., Liu, W., Qian, C., Lin, M., Ouyang, W., Luo, P.: 3D interacting hand pose estimation by hand de-occlusion and removal. In: European Conference on Computer Vision (ECCV). pp. 380–397. Springer (2022) 3", + "43. Moon, G.: Bringing inputs to shared domains for 3d interacting hands recovery in the wild. In: Computer Vision and Pattern Recognition (CVPR). pp. 17028-17037 (2023) 3", + "44. Moon, G., Saito, S., Xu, W., Joshi, R., Buffalini, J., Bellan, H., Rosen, N., Richardson, J., Mize, M., De Bree, P., et al.: A dataset of relighted 3d interacting hands. Conference on Neural Information Processing Systems (NeurIPS) 36 (2024) 3" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Benchmarks and Challenges for Egocentric Hand Interactions", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Moon, G., Yu, S., Wen, H., Shiratori, T., Lee, K.M.: Interhand2.6m: A dataset and baseline for 3d interacting hand pose estimation from a single RGB image. In: European Conference on Computer Vision (ECCV). vol. 12365, pp. 548-564 (2020) 3, 5, 6", + "46. Mueller, F., Bernard, F., Sotnychenko, O., Mehta, D., Sridhar, S., Casas, D., Theobalt, C.: GANerated hands for real-time 3D hand tracking from monocular RGB. In: Computer Vision and Pattern Recognition (CVPR). pp. 49-59 (2018) 3", + "47. Ohkawa, T.: AssemblyHands Toolkit. https://github.com/facebookresearch/assemblyhands-toolkit (2023) 5", + "48. Ohkawa, T., Furuta, R., Sato, Y.: Efficient annotation and learning for 3d hand pose estimation: A survey. International Journal of Computer Vision (IJCV) 131, 3193-3206 (2023) 3", + "49. Ohkawa, T., He, K., Sener, F., Hodan, T., Tran, L., Keskin, C.: AssemblyHands: towards egocentric activity understanding via 3d hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 12999-13008 (2023) 2, 3, 4", + "50. Ohkawa, T., Li, Y.J., Fu, Q., Furuta, R., Kitani, K.M., Sato, Y.: Domain adaptive hand keypoint and pixel localization in the wild. In: European Conference on Computer Vision (ECCV). pp. 68—87 (2022) 2", + "51. Ohkawa, T., Yagi, T., Hashimoto, A., Ushiku, Y., Sato, Y.: Foreground-aware stylization and consensus pseudo-labeling for domain adaptation of first-person hand segmentation. IEEE Access 9, 94644-94655 (2021) 2", + "52. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) 8", + "53. Park, J., Oh, Y., Moon, G., Choi, H., Lee, K.M.: HandOccNet: Occlusion-robust 3d hand mesh estimation network. In: Computer Vision and Pattern Recognition (CVPR). pp. 1496–1505 (2022) 6", + "54. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022) 14", + "55. Prakash, A., Tu, R., Chang, M., Gupta, S.: 3d hand pose estimation in everyday egocentric images. In: European Conference on Computer Vision (ECCV) (2024) 8", + "56. Radosavovic, I., Kosaraju, R.P., Girshick, R.B., He, K., Dollár, P.: Designing network design spaces. In: Computer Vision and Pattern Recognition (CVPR). pp. 10425-10433 (2020) 5, 6", + "57. Rehg, J.M., Kanade, T.: Visual tracking of high DOF articulated structures: An application to human hand tracking. In: European Conference on Computer Vision (ECCV). vol. 801, pp. 35-46 (1994) 3", + "58. Remelli, E., Han, S., Honari, S., Fua, P., Wang, R.: Lightweight multi-view 3d pose estimation through camera-disentangled representation. In: Computer Vision and Pattern Recognition (CVPR). pp. 6039-6048 (2020) 6, 7", + "59. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG) 36(6), 245:1-245:17 (2017) 7", + "60. Ryali, C., Hu, Y., Bolya, D., Wei, C., Fan, H., Huang, P., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: International Conference on Machine Learning (ICML). vol. 202, pp. 29441-29454 (2023) 5, 6" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Computer Vision and Pattern Recognition (CVPR). pp. 21064-21074 (2022) 10", + "62. Shamil, M.S., Chatterjee, D., Sener, F., Ma, S., Yao, A.: On the utility of 3d hand poses for action recognition. In: European Conference on Computer Vision (ECCV) (2024) 14", + "63. Simon, T., Joo, H., Matthews, I., Sheikh, Y.: Hand keypoint detection in single images using multiview bootstrapping. In: Computer Vision and Pattern Recognition (CVPR). pp. 4645-4653 (2017) 3", + "64. Spurr, A., Dahiya, A., Wang, X., Zhang, X., Hilliges, O.: Self-supervised 3D hand pose estimation from monocular RGB via contrastive learning. In: International Conference on Computer Vision (ICCV). pp. 11210-11219 (2021) 3", + "65. Spurr, A., Iqbal, U., Molchanov, P., Hilliges, O., Kautz, J.: Weakly supervised 3D hand pose estimation via biomechanical constraints. In: European Conference on Computer Vision (ECCV). vol. 12362, pp. 211-228 (2020) 2, 3", + "66. Spurr, A., Song, J., Park, S., Hilliges, O.: Cross-modal deep variational hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 89-98 (2018) 3", + "67. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Computer Vision and Pattern Recognition (CVPR) (2019) 8", + "68. Swamy, A., Leroy, V., Weinzaepfel, P., Baradel, F., Galaoui, S., Brégier, R., Armando, M., Franco, J.S., Rogez, G.: SHOWMe: Benchmarking object-agnostic hand-object 3d reconstruction. In: International Conference on Computer Vision (ICCV). pp. 1935–1944 (2023) 3", + "69. Tekin, B., Bogo, F., Pollefeys, M.: $\\mathrm{H} + \\mathrm{O}$ : Unified egocentric recognition of 3D hand-object poses and interactions. In: Computer Vision and Pattern Recognition (CVPR). pp. 4511-4520 (2019) 3", + "70. Tse, T.H.E., Kim, K.I., Leonardis, A., Chang, H.J.: Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In: Computer Vision and Pattern Recognition (CVPR). pp. 1664-1674 (2022) 3", + "71. Tzionas, D., Gall, J.: A comparison of directional distances for hand pose estimation. In: German Conference on Pattern Recognition (GCPR). vol. 8142, pp. 131-141 (2013) 3", + "72. Wen, Y., Pan, H., Ohkawa, T., Yang, L., Pan, J., Sato, Y., Komura, T., Wang, W.: Generative hierarchical temporal transformer for hand action recognition and motion prediction. CoRR abs/2311.17366 (2023) 2", + "73. Yang, L., Chen, S., Yao, A.: Semihand: Semi-supervised hand pose estimation with consistency. In: International Conference on Computer Vision (ICCV). pp. 11364-11373 (2021) 2", + "74. Yang, L., Zhan, X., Li, K., Xu, W., Li, J., Lu, C.: CPF: Learning a contact potential field to model the hand-object interaction. In: International Conference on Computer Vision (ICCV) (2021) 3", + "75. Ye, Y., Gupta, A., Tulsiani, S.: What's in your hands? 3d reconstruction of generic objects in hands. In: Computer Vision and Pattern Recognition (CVPR) (2022) 3", + "76. Ye, Y., Hebbar, P., Gupta, A., Tulsiani, S.: Diffusion-guided reconstruction of everyday hand-object interaction clips. In: International Conference on Computer Vision (ICCV) (2023) 3" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Benchmarks and Challenges for Egocentric Hand Interactions", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "77. Yuan, S., Garcia-Hernando, G., Stenger, B., Moon, G., Chang, J.Y., Lee, K.M., Molchanov, P., Kautz, J., Honari, S., Ge, L., Yuan, J., Chen, X., Wang, G., Yang, F., Akiyama, K., Wu, Y., Wan, Q., Madadi, M., Escalera, S., Li, S., Lee, D., Oikonomidis, I., Argyros, A.A., Kim, T.: Depth-based 3D hand pose estimation: From current achievements to future goals. In: Computer Vision and Pattern Recognition (CVPR). pp. 2636-2645 (2018) 3", + "78. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Computer Vision and Pattern Recognition (CVPR). pp. 12104-12113 (2022) 8", + "79. Zhang, H., Christen, S., Fan, Z., Hilliges, O., Song, J.: GraspXL: Generating grasping motions for diverse objects at scale. European Conference on Computer Vision (ECCV) (2024) 2", + "80. Zhang, H., Christen, S., Fan, Z., Zheng, L., Hwangbo, J., Song, J., Hilliges, O.: ArtiGrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. In: International Conference on 3D Vision (3DV) (2024) 2", + "81. Zhang, X., Li, Q., Mo, H., Zhang, W., Zheng, W.: End-to-end hand mesh recovery from a monocular RGB image. In: International Conference on Computer Vision (ICCV). pp. 2354-2364 (2019) 3", + "82. Zhou, Y., Habermann, M., Xu, W., Habibie, I., Theobalt, C., Xu, F.: Monocular real-time hand shape and motion capture using multi-modal data. In: Computer Vision and Pattern Recognition (CVPR). pp. 5345-5354 (2020) 3, 5, 6", + "83. Zhou, Z.: SimpleHand: Winner of the HANDS'2023 AssemblyHands Challenge @ ICCV. https://github.com/patienceFromZhou/simpleHand (2024) 5", + "84. Zhou, Z., Lv, Z., Zhou, S., Zou, M., Wu, T., Yu, M., Tang, Y., Liang, J.: 1st place solution of egocentric 3d hand pose estimation challenge 2023 technical report: A concise pipeline for egocentric hand pose reconstruction. CoRR abs/2310.04769 (2023) 11", + "85. Zhou, Z., Zhou, S., Lv, Z., Zou, M., Tang, Y., Liang, J.: A simple baseline for efficient hand mesh reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 1367-1376 (2024) 5, 6", + "86. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (ICLR). OpenReview.net (2021), https://openreview.net/forum?id=gZ9hCDWe6ke 8", + "87. Ziani, A., Fan, Z., Kocabas, M., Christen, S., Hilliges, O.: TempCLR: Reconstructing hands via time-coherent contrastive learning. In: International Conference on 3D Vision (3DV). pp. 627-636 (2022) 3", + "88. Zimmermann, C., Brox, T.: Learning to estimate 3D hand pose from single RGB images. In: International Conference on Computer Vision (ICCV). pp. 4913-4921 (2017) 3, 5" + ], + "bbox": [ + 215, + 146, + 785, + 686 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Z. Fan, T. Ohkawa, L. Yang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_model.json b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..02e77e1ad8d553f7b8c0eb202842ca66b207918d --- /dev/null +++ b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_model.json @@ -0,0 +1,2748 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.212, + 0.747, + 0.306 + ], + "angle": 0, + "content": "Zicong Fan\\(^{1,2*}\\), Takehiko Ohkawa\\(^{3*}\\), Linlin Yang\\(^{4*}\\), Nie Lin\\(^{3}\\), Zhishan Zhou\\(^{5}\\), Shihao Zhou\\(^{5}\\), Jiajun Liang\\(^{5}\\), Zhong Gao\\(^{6}\\), Xuanyang Zhang\\(^{6}\\), Xue Zhang\\(^{7}\\), Fei Li\\(^{7}\\), Liu Zheng\\(^{8}\\), Feng Lu\\(^{8}\\), Karim Abou Zeid\\(^{9}\\), Bastian Leibe\\(^{9}\\), Jeongwan On\\(^{10}\\), Seungryul Baek\\(^{10}\\), Aditya Prakash\\(^{11}\\), Saurabh Gupta\\(^{11}\\), Kun He\\(^{12}\\), Yoichi Sato\\(^{3}\\), Otmar Hilliges\\(^{1}\\), Hyung Jin Chang\\(^{13}\\), Angela Yao\\(^{14}\\)" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.341, + 0.744, + 0.606 + ], + "angle": 0, + "content": "Abstract. We interact with the world with our hands and see it through our own (egocentric) perspective. A holistic 3D understanding of such interactions from egocentric views is important for tasks in robotics, AR/VR, action recognition and motion generation. Accurately reconstructing such interactions in 3D is challenging due to heavy occlusion, viewpoint bias, camera distortion, and motion blur from the head movement. To this end, we designed the HANDS23 challenge based on the AssemblyHands and ARCTIC datasets with carefully designed training and testing splits. Based on the results of the top submitted methods and more recent baselines on the leaderboards, we perform a thorough analysis on 3D hand(-object) reconstruction tasks. Our analysis demonstrates the effectiveness of addressing distortion specific to egocentric cameras, adopting high-capacity transformers to learn complex hand-object interactions, and fusing predictions from different views. Our study further reveals challenging scenarios intractable with state-of-the-art methods, such as fast hand motion, object reconstruction from narrow egocentric views, and close contact between two hands and objects. Our efforts will enrich the community's knowledge foundation and facilitate future hand studies on egocentric hand-object interactions." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.669, + 0.377, + 0.685 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.789, + 0.763 + ], + "angle": 0, + "content": "We interact with the world with our hands and see it through our eyes: we wake up and grab our phone to check the time; we use tools when assembling parts of a car; we open the microwave door to heat food, to name a few. An egocentric 3D understanding of our hand interactions with objects will fundamentally impact" + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.77, + 0.788, + 0.842 + ], + "angle": 0, + "content": "\\(^{*}\\)Equal contribution, \\(^{1}\\)ETH Zürich, \\(^{2}\\)Max Planck Institute for Intelligent Systems, Tübingen, Germany, \\(^{3}\\)The University of Tokyo, \\(^{4}\\)Communication University of China, \\(^{5}\\)Jiiov, \\(^{6}\\)Bytedance, \\(^{7}\\)Fujitsu Research & Development Center Co., Ltd., \\(^{8}\\)Beihang University, \\(^{9}\\)RWTH Aachen University, \\(^{10}\\)UNIST, \\(^{11}\\)UIUC, \\(^{12}\\)Meta, \\(^{13}\\)University of Birmingham, \\(^{14}\\)National University of Singapore." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.499, + 0.129 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.145, + 0.7, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.276 + ], + "angle": 0, + "content": "Fig. 1: Tasks in HANDS23. In AssemblyHands, from its multi-view headset (a), we estimate 3D hand poses from images (b); In ARCTIC, we estimate the poses of two hands and articulated objects from an image (c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.337 + ], + "angle": 0, + "content": "areas such as robotics grasping [11,80], augmented and virtual reality [25], action recognition [7,19,72] and motion generation [79,80]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.785, + 0.519 + ], + "angle": 0, + "content": "However, it is non-trivial to accurately reconstruct 3D hands and or objects due to its high degree of freedoms [65, 73], ambiguous texture [16], and heavy occlusions. These challenges are intensified in an egocentric view [50, 51], particularly with object interactions, due to significant camera distortion, rapid and varied changes in viewpoint caused by head movements and hand-object occlusion. To better understand these challenges, we introduce a public challenge in conjunction with ICCV 2023 (i.e., HANDS23) based on recent egocentric hand datasets, AssemblyHands [49] and ARCTIC [17] (see Figure 1). These two datasets are large-scale, multi-view, and provide monochrome or RGB egocentric videos of the hands dexterously manipulating objects. Accordingly, we host two tasks: 1) egocentric 3D hand pose estimation from a single-view image based on AssemblyHands, and 2) consistent motion reconstruction based on ARCTIC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.785, + 0.642 + ], + "angle": 0, + "content": "We introduce new HANDS23 methods and recent dataset leaderboard baselines that substantially outperform initial baselines for both tasks, setting new benchmarks for subsequent comparisons on the datasets. The two datasets are significantly larger and include a wider variety of bimanual manipulations compared to earlier datasets [6, 23], enabling a more authentic assessment of real-world interactions. With these benchmarks, we thoroughly analyze factors such as viewpoint, action types, hand position, model size, and object variations to determine their impact on 3D hand(-object) reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.786, + 0.764 + ], + "angle": 0, + "content": "Our findings show the success of addressing the distortion of egocentric cameras with explicit perspective cropping or implicit learning for the distortion bias. In addition, recent high-capacity vision transformers can learn complex hand-object interactions. Adaptive fusion techniques for multi-view predictions further boost performance. We also analyze the remaining challenges that are still difficult to handle with the recent methods, e.g., fast hand motion, object reconstruction from narrow and moving views, and intricate interactions and close contact between two hands and objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.786, + 0.84 + ], + "angle": 0, + "content": "To summarize, we contribute state-of-the-art baselines and gather the submitted methods for AssemblyHands and ARCTIC to foster future research on egocentric hand-object interactions. Furthermore, we thoroughly analyze the two benchmarks to provide insights for future directions in egocentric hand pose estimation and consistent motion reconstruction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.326 + ], + "angle": 0, + "content": "3D hand pose estimation: Reconstructing 3D hand poses has a long history [14,48] ever since the important work led by Rehg and Kanade [57]. A large body of research in this area focuses on single-hand reconstruction [3,4,6,9,13, 16,18,20,28,32,37,46,63-66,71,81,87,88]. For example, while a popular OpenPose library features 2D hand keypoints [63], Zimmermann et al. [88] initially extend to estimate 3D hand poses using deep convolutional networks. Ever since the release of InterHand2.6M [45] dataset, the community has increased focus on strongly interacting two hands reconstruction [22,33-36,42-45,49]. For example, Li et al. [35] and Moon et al. [44] use relighting techniques to augment InterHand2.6M with more natural lighting and diverse backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Hand-object reconstruction: The holistic reconstruction of hands and objects have increased interest in the hand community in recent years [6,12,17,21, 26-28,33,39,69,70,74,82]. Methods in this area mostly assume a given object template and jointly estimate the hand poses and the object rigid poses [12, 21,26-28,39,69,70,74,82] or articulated poses [17]. For example, Cao et al. [5] fits object templates to in-the-wild interaction videos. Liu et al. [39] introduce a semi-supervised learning framework via pseudo-groundtruth from temporal data to improve hand-object reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.537 + ], + "angle": 0, + "content": "More recent methods do not assume object templates for hand-object reconstruction [10, 15, 28, 31, 68, 75, 76]. For example, Fan et al. [15] introduced the first category-agnostic method that reconstructs an articulated hand and object jointly from a video. Nonetheless, our challenges and insights on hand-object occlusions and camera distortions are still applicable to these more challenging template-free reconstruction settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Public reports for the previous challenges (HANDS17 [77] and HANDS19 [2]) distilled the insights from individual review papers and practical techniques into comprehensive summaries to enrich the community's knowledge base. These past challenges use benchmarks which include depth-based hand pose estimation from egocentric views. Instead of depth sensors, the HANDS23 benchmarks are based on affordable and widely applicable image sensors, i.e., RGB and monochrome images. This paper further advances the analysis with unique insights, such as multi-view egocentric cameras, object reconstruction in contact, and modeling with recent transformers beyond conventional CNNs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.543, + 0.712 + ], + "angle": 0, + "content": "3 HANDS23 challenge overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.787, + 0.768 + ], + "angle": 0, + "content": "The workshop contains two hand-related 3D reconstruction challenges in hand-object strongly interacting settings. In this section, we introduce the two challenges and their evaluation criteria." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.436, + 0.803 + ], + "angle": 0, + "content": "3.1 Workshop challenges" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "3D hand pose estimation in AssemblyHands: As illustrated in Figure 1, this task focuses on egocentric 3D hand pose estimation from a single-view im" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.128 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "age based on AssemblyHands. The dataset provides multi-view captured videos of hand-object interaction while assembling and disassembling toy vehicles. In particular, it provides allocentric (fixed-view) and egocentric recordings and auxiliary cues like action, object, or context information for hand pose estimation. We refer readers to [49] for more dataset details. The training, validation, and testing sets contain 383K, 32K, and 62K monochrome images, respectively, all captured from egocentric cameras. During training, 3D hand keypoint coordinates, hand bounding boxes and camera intrinsic and extrinsic matrices for the four egocentric cameras attached to the headset are provided. The same information is provided during testing, minus the 3D keypoints. Unlike [49], given the availability of multi-view egocentric images, this task lets participants develop multi-view fusion based on the corresponding multi-view images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.482 + ], + "angle": 0, + "content": "Consistent motion reconstruction in ARCTIC: Given an RGB image, the goal of this task is to estimate poses of hands and articulated objects to recover the 3D surfaces of the interaction (see Figure 1). We refer readers to [17] for more details. The ARCTIC dataset contains data of hands dexterously manipulating articulated objects and videos from \\(8 \\times\\) allocentric views and \\(1 \\times\\) egocentric views. The official splits of the ARCTIC dataset are used for training, validation, and testing. There are two sub-tasks: allocentric task and egocentric task. In the former, only allocentric images can be used for training and evaluation. For the latter, all images from the training set can be used for training while only the egocentric view images are used during evaluation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.51, + 0.416, + 0.523 + ], + "angle": 0, + "content": "3.2 Evaluation criteria" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.7 + ], + "angle": 0, + "content": "AssemblyHands evaluation: We use MPJPE as an evaluation metric in millimeters, comparing the model predictions against the ground-truth in world coordinates. We provide the intrinsic and extrinsic of the egocentric cameras to construct submission results defined in the world coordinates. Assuming that the human hand has a total \\( N_J \\) joints, we denote wrist-relative coordinates of the prediction and ground-truth as \\( \\hat{J} \\in \\mathbb{R}^{N_J \\times 3} \\) and \\( J \\in \\mathbb{R}^{N_J \\times 3} \\), respectively. Given a joint visibility indicator \\( \\gamma_i \\) per joint \\( J_i \\), we compute the Euclidean distance between predicted and ground-truth joints as \\( \\frac{1}{\\sum_{i=1}^{N_J} \\gamma_i} \\sum_{i=1}^{N_J} \\gamma_i \\left\\| \\hat{J}_i - J_i \\right\\|_2 \\). The visibility indicator offers per-joint binary labels, representing whether the annotated keypoints are visible from the given egocentric view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.843 + ], + "angle": 0, + "content": "ARCTIC evaluation: Since the original ARCTIC paper [17] has a heavy focus on the quality of hand-object contact in the reconstructed hand and object meshes, we use Contact Deviation (CDev) introduced in the ARCTIC dataset as the main metric for the competition. In particular, this metric measures the extent to which a hand vertex deviates from the supposed contact object vertex in the prediction. Concretely, suppose that for a given frame, \\(\\{(\\mathbf{h}_i,\\mathbf{o}_i)\\}_{i = 1}^C\\) are \\(C\\) pairs of in-contact (\\(< 3mm\\) distance) hand-object vertices according to ground-truth, and \\(\\{(\\hat{\\mathbf{h}}_i,\\hat{\\mathbf{o}}_i)\\}_{i = 1}^C\\) are the predictions respectively. The CDev metric is the average distance between \\(\\hat{\\mathbf{h}}_i\\) and \\(\\hat{\\mathbf{o}}_i\\) in millimeters, \\(\\frac{1}{C}\\sum_{i = 1}^{C}||\\hat{\\mathbf{h}}_i - \\hat{\\mathbf{o}}_i||\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "For completeness, we report all metrics introduced in ARCTIC. In particular, the task requires the reconstructed meshes to have accurate hand-object contact (CDev) and smooth motion (ACC). Additionally, during articulation or when carrying an object, it is crucial that the vertices of the hand and object which are in contact maintain synchronized movement (MDev). Moreover, we assess hand and object poses, alongside their relative movements, using metrics like MPJPE, AAE, Success Rate, and MRRPE. For detailed information, see [17]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.277, + 0.341, + 0.293 + ], + "angle": 0, + "content": "4 Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.787, + 0.355 + ], + "angle": 0, + "content": "This section presents the methods in the two challenges and other competing methods on the leaderboards. Four methods outperform the baseline in both AssemblyHands and ARCTIC." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.377, + 0.472, + 0.392 + ], + "angle": 0, + "content": "4.1 AssemblyHands methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.508 + ], + "angle": 0, + "content": "Participants develop methods that learn the mapping from egocentric images to 3D keypoints. The methods are categorized into: heatmap-based and regression-based approaches. Given the presence of complex hand-object interactions in the egocentric scenes, high-capacity transformer models and attention mechanisms addressing occluded regions have been proposed as the backbone networks. Table 1 summarizes the methods based on the learning, preprocessing, multi-view fusion, and post-processing approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.553 + ], + "angle": 0, + "content": "Base: This method uses a heatmap-based framework based on heatmaps [45] with 2.5D representations [32] and a ResNet50 [29] backbone. The implementation can be found in [47]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.787, + 0.674 + ], + "angle": 0, + "content": "JHands [85]: This method employs a regression-based approach with simple MLP heads for regressing 2D keypoints, root-relative 3D keypoints, and the global root depth. The regression training is empowered by a recent fast and strong vision transformer, Hiera [60], pre-trained with masked auto-encoder [30]. A multi-level feature fusion that concatenates the features of different layers is adopted for better feature extraction at different scales. The method additionally uses other publicly available datasets for training, namely FreiHAND [88], DexYCB [6], and CompHand [8]. The implementation is available in [83]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.81 + ], + "angle": 0, + "content": "PICO-AI: This method proposes a heatmap voting scheme in addition to the 2.5D heatmaps. Due to their sparsity, the conventional heatmaps pose an imbalance problem between positive and negative samples in the loss function. Hence, the proposed voting mechanism aims to spread the loss evenly across the entire heatmaps. Given the initial guess of keypoints obtained from the heatmaps, the method defines a local region centered on the joint position and operates the soft-argmax within the region to obtain refined keypoint coordinates. This restricts the impact of background points, leading to more reliable optimization. The training is facilitated by CNN-based RegNety320 [56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Phi-AI: While following the heatmap-based approach, this method adapts MinimalHand [82] with the ResNet50 backbone, where 2D heatmaps and 3D location" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.499, + 0.129 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: Method and preprocessing summary in AssemblyHands. We summarize submitted methods in terms of learning methods, architecture, preprocessing, and multi-view fusion techniques. The tuple (views, phase) indicates the number of views used in either train or test time." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.213, + 0.785, + 0.376 + ], + "angle": 0, + "content": "
MethodLearning methodsArchitecturePreprocessingMulti-view fusion (views, phase)
Base2.5D heatmaps [45]ResNet50 [29]-Simple average (4, test)
JHands [85]RegressionHiera [60]Warp perspective, color jitter, random maskAdaptive view selection and average (2, test)
PICO-AI2.5D heatmaps [45] Heatmap votingRegNety320 [56]Scale, rotate, flip, translateAdaptive view selection FTL [58] (2, train)
FRDCRegression 2D heatmapsHandOccNet [53] with ConvNeXt [41]Scale, rotate, color jitterWeighted average (4, test)
Phi-AI2D heatmaps and 3D location maps [82]ResNet50 [29]Scale, rotate, translate, color jitter, gaussian blurWeighted average (4, test)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.784, + 0.499 + ], + "angle": 0, + "content": "maps are regressed. Instead of selecting 3D keypoint coordinates from the location maps, the proposed method modifies it by using heatmap values to weight 3D keypoint coordinates, achieving a more robust estimation. Moreover, the method adds a residual structured layer after the original three-tier cascade networks to refine the calculated location maps. The method further applies the ensemble of final keypoint outputs combined with the Base." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.785, + 0.607 + ], + "angle": 0, + "content": "FRDC: This method adopts a hybrid approach by combining regression with heatmap for training. HandOccNet [53] is modified to regress 3D keypoint coordinates and integrated with an additional branch of 2D heatmap regression. HandOccNet enriches feature extraction with spatial attention mechanisms for occluded regions, making it robust under hand-object occlusions. The method further utilizes a stronger ConvNeXt [41] backbone and feature fusion from the 2D keypoint regressor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.785, + 0.684 + ], + "angle": 0, + "content": "Preprocessing of egocentric images: Compared to conventional static camera setups, egocentric images exhibit unique properties and biases, such as distortion, head camera motion, and different color representations. Thus, it is vital to preprocess egocentric images to alleviate these effects during training. Augmentation techniques are detailed in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.787, + 0.793 + ], + "angle": 0, + "content": "The method JHands addresses the distortion issue with a warp perspective operation to make the hands near the edge less stretched. While AssemblyHands provides rectified images converted from fisheye cameras to a pinhole camera model, they often include excessively stretched areas near the edges. To address this, the method calculates a virtual camera and corresponding perspective transformation matrix based on the pixel coordinates of the crop and the camera parameters. The generated crops can be found in the analysis of Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Multi-view fusion: Since AssemblyHands offers multi-view egocentric videos, participants can optionally use the constraint of multi-view geometry and fusion techniques during training or inference." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Method and preprocessing summary in ARCTIC. We summarize baselines on ARCTIC in terms of input dimensions, image backbones, learning rate scheduling, training epochs, batch size and the cropping used for input. *Method trains 50 epochs for decoder and 36 for backbone. +Learning rate is 1e-7 to 1e-4 with linear warmup for first \\(5\\%\\) step, and 1e-4 to 1e-7 with cosine decay for rest." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.227, + 0.784, + 0.358 + ], + "angle": 0, + "content": "
MethodInput sizeBackbonesLearning rate scheduleTraining epochsBatch sizeCropping
ArcticNet-SF [17]224 × 224ResNet501e-5allocentric: 20 \negocentric: 5064object
DIGIT [16]224 × 224HRNet-W321e-5allocentric: 20 \negocentric: 5064object
AmbiguousHands224 × 224ResNet501e-5allocentric: 20 \negocentric: 10032hand \nobject
UVHand384 × 384Swin-L2e-4 (backbone) \n1e-7 (others)allocentric: N/A \negocentric: 50/36*48object
JointTransformer [1]224 × 224ViT-G1e-7/1e-5+allocentric: 20 \negocentric: 10064object
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.784, + 0.506 + ], + "angle": 0, + "content": "While Base uses a simple average of predicted keypoints from all four camera views in the test time, PICO-AI proposes multi-view feature fusion during training using Feature Transform Layers (FTL) [58]. This FTL training requires fusing two out of four views; thus, the method chooses the most suitable views for every frame. In cases with multiple candidates, the Intersection over Union (IoU) is computed between hand boxes from per-view predictions and 2D keypoints from previous 3D predictions. The two views with the highest IoUs are selected for their superior prediction reliability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.627 + ], + "angle": 0, + "content": "The methods JHands, FRDC, and Phi-AI apply adaptive fusion in predicted keypoints during testing. The method JHands computes the MPJPE with each other view and selects two results of views with the lowest MPJPE, excluding noisy predictions in the average. If the MPJPE is lower than a threshold, the mean of the two results is calculated as the final result. Otherwise, the result with a lower PA-MPJPE with the predictions in the previous frame is chosen. The methods FRDC and Phi-AI use a weighted average for each view prediction, assigning weights based on each view's validation performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.688 + ], + "angle": 0, + "content": "Postprocessing: Several postprocessing techniques, including test-time augmentation, smoothing, and model ensembling are used to enhance inference outcomes. In particular, the method JHands applies an offline smooth (Savitzky-Golay) filter on each video sequence." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.416, + 0.725 + ], + "angle": 0, + "content": "4.2 ARCTIC methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Table 2 summarizes the details for each method in terms of the input image dimensions, image backbones, learning rate scheduling, the number of training epochs, batch size, and cropping strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Preliminary: All methods below are regression-based and predict two-hand MANO [59] parameters \\(\\Theta = \\{\\theta, \\beta\\}\\) and articulated object parameters \\(\\Omega\\). In particular, with the MANO pose and shape parameters \\(\\theta, \\beta\\), the MANO model \\(\\mathcal{H}\\) returns a mesh with vertices via \\(\\mathcal{H}(\\theta, \\beta) \\in \\mathbb{R}^{778 \\times 3}\\). 3D joints are obtained via" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.128 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.785, + 0.193 + ], + "angle": 0, + "content": "a linear regressor. The articulated object model \\(\\mathcal{O}\\) was introduced in ARCTIC to provide an articulated mesh with vertices via \\(\\mathcal{O}(\\Omega)\\in \\mathbb{R}^{\\mathrm{V}\\times 3}\\), where \\(\\Omega \\in \\mathbb{R}^{7}\\) contains the global orientation, global translation, and object articulation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.197, + 0.785, + 0.258 + ], + "angle": 0, + "content": "ArcticNet-SF [17]: Introduced in ARCTIC, it is a single-frame baseline. It first extracts an image feature vector from the input image; then, it regresses hand and object parameters with simple MLPs. The hand and object meshes can then be extracted via \\(\\mathcal{H}(\\cdot)\\) and \\(\\mathcal{O}(\\cdot)\\). For more details, see [17]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.414 + ], + "angle": 0, + "content": "JointTransformer [1]: JointTransformer enhances ArcticNet-SF by integrating a transformer decoder instead of the MLP regressors for hand and object parameter estimation. The decoder employs learned queries for the angle of each joint, the shape and translation of each hand, and the translation, rotation, and articulation of the object. It alternates between self-attention between queries and cross-attention of queries to the elements of the backbone feature map, followed by linear layers that regress the final parameters. Specifically, separate linear layers are dedicated to regressing joint angles, hand shape, hand translation, object translation, object orientation, and object articulation. The best model uses a ViT-G [78] backbone with frozen DINOv2 weights [52]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.509 + ], + "angle": 0, + "content": "AmbiguousHands [55]: The method addresses scale ambiguity, resulting from bounding box cropping in data augmentation and camera intrinsics, by employing positional encoding of these elements to mitigate scale issues. This leads to improved spatial alignment. Subsequently, the approach enhances network visibility by integrating local features through distinct hand and object crops. They follow the general approach of ArcticNet-SF to regress hand/object parameters." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.513, + 0.785, + 0.604 + ], + "angle": 0, + "content": "UVHand: Since ArcticNet-SF only leverages a global feature vector to estimate hand and object parameters, the image features lack local context. To address this, UVHand leverages Swin-L transformer [40] to extract image features. They then further leverage Deformable DETR [86] to encode the multiple-scale feature maps. The encoded feature maps are then aggregated via self- and cross-attention before regressing hand and object parameters." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.685 + ], + "angle": 0, + "content": "DIGIT [16]: The method was introduced to estimate strongly interacting hands in [16]. Since ArcticNet-SF is sensitive when the hands are interacting with objects, DIGIT was extended to the ARCTIC setting. Given an image, it first estimates hand part-wise segmentation masks and object masks. The mask predictions are fused with the image features to perform parameter estimation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Implementation details: Table 2 shows that all methods use the default cropping as in ARCTIC to crop around the object, while AmbiguousHands performs three crops (around two hands and the object). DIGIT uses the HRNet-W32 backbone [67] and trains with a batch size of 64 with the same learning rate for all iterations. UVHand takes as input a \\(384 \\times 384\\) image cropped around the object and encodes it with the Swin-L transformer [40] backbone. It was trained with a batch size of 48 with a learning rate of 2e-4 for the backbone and 1e-7 for other weights. Due to computational cost, they train 50 epochs for the decoder and 36 for the backbone. JointTransformer uses ViT-G [78] backbone with frozen DINOv2 [52] weights to train with a batch size of 64. It performs a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.256 + ], + "angle": 0, + "content": "Table 3: Method performance in AssemblyHands. We compare AssemblyHands method performance on egocentric test data. We show the final MPJPE on the test set as the metrics (lower better). We also provide detailed evaluations, regarding the varying distances of hand position from the image center and different verb action categories. The hand distance is computed by the distance from the image center to the hand center position per image, and averaged over the lower two views of the headset. Verb classes of \"attempt to X\" are merged to \"X\" for simplicity. The higher and lower three verbs are color-coded in red and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.269, + 0.784, + 0.468 + ], + "angle": 0, + "content": "
MethodScoreHand distance (px)Verb class
-200200-250250-clapinspectpasspick uppositionposition screw onpull
Base20.6920.3121.9724.8519.8822.8922.4821.8522.6521.6218.74
JHands12.2112.3511.9813.7210.6516.2712.8613.6714.5812.811.06
PICO-AI12.4612.5111.6212.9512.9815.311.3713.213.1811.3915.13
FRDC16.4816.3915.8918.6915.2421.3317.8618.2619.2118.0312.83
Phi-AI17.2617.2415.8119.5119.8620.9317.9119.0119.719.3517.3
Verb class (continue)
pushput downremoveremove screw fromrotatescrewtilt downtilt upunscrewnone
Base19.2920.2619.9916.4722.7122.9513.1215.1120.7819.82
JHands13.9613.7213.1111.7212.2614.119.619.9212.2510.99
PICO-AI12.2913.4112.839.9913.713.7210.4411.5612.8711.71
FRDC19.5217.8718.5514.4716.4418.8114.0313.3716.4115.01
Phi-AI19.1218.2918.1813.9518.3519.7813.1315.3617.2915.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.784, + 0.526 + ], + "angle": 0, + "content": "linear warmup from 1e-7 to 1e-4 in the first \\(5\\%\\) steps and uses cosine decay from 1e-4 to 1e-7 for the rest of the steps." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.552, + 0.45, + 0.57 + ], + "angle": 0, + "content": "5 Results and analysis" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.586, + 0.327, + 0.601 + ], + "angle": 0, + "content": "5.1 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.673 + ], + "angle": 0, + "content": "Here, we benchmark results of valid submissions for state-of-the-art comparison in AssemblyHands and ARCTIC and other more recent baselines. In particular, for AssemblyHands, we report egocentric hand pose estimation results. For ARCTIC, we report results for the allocentric and egocentric test sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "AssemblyHands benchmark: Table 3 shows the final test scores on the AssemblyHands dataset. The methods in the table exceed the baseline (Base) with a test score of 20.69 MPJPE. Notably, the methods JHands and PICO-AI achieve a nearly \\(40\\%\\) reduction over the baseline. The methods FRDC and Phi-AI improve the test score by \\(20.3\\%\\) and \\(16.5\\%\\) against the baseline, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "ARCTIC benchmark: Table 4 presents the comparative performance of methods in the ARCTIC dataset, where ArcticNet-SF serves as the initial benchmark. The majority surpass ArcticNet-SF in both allocentric and egocentric views, except for UVHand, which underperforms due to incomplete training. In the egocentric view, AmbiguousHands excels in creating smooth, consistent mesh motions (refer to MDev and \\(\\mathrm{ACC}_{h / o}\\) metrics). Notably, JointTransformer stands" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.499, + 0.129 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.243 + ], + "angle": 0, + "content": "Table 4: Method performance in ARCTIC. We compare performance in both allocentric (top half) and egocentric (bottom half) views. We evaluate using metrics for contact and relative position (measuring hand-object contact and prediction of relative root position), motion (assessing temporally-consistent contact and smoothness), and hand and object metrics (indicating root-relative reconstruction error). We use the CDev score as the main metric for this competition. We denote left and right hands as \\( l \\) and \\( r \\), and the object as \\( o \\)." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.256, + 0.785, + 0.37 + ], + "angle": 0, + "content": "
Contact and Relative PositionsMotionHandObject
MethodCDev [mm]↓ MRRPEa/ro [mm]↓ MDev [mm]↓ ACCh/α [m/s2]↓ MPJPE [mm]↓ AAE [°] ↓ Success Rate [%] ↑
AllocentricArcticNet-SF41.5652.39/37.4710.405.72/7.5721.455.3771.39
DIGIT34.9244.19/35.438.374.86/6.6317.925.2476.52
UVHand64.1584.68/70.3114.127.05/12.0440.9912.3631.47
AmbiguousHands33.2545.78/34.5610.126.37/6.4018.024.6481.94
JointTransformer27.9736.17/28.188.936.08/5.7917.123.9589.79
Ego-centricArcticNet-SF44.7128.31/36.1611.805.03/9.1519.186.3953.89
DIGIT41.3125.49/32.619.484.01/8.3216.746.6053.33
UVHand40.4340.93/36.889.965.32/8.3324.537.3257.28
AmbiguousHands35.9323.07/27.539.513.95/6.7616.264.8668.36
JointTransformer32.5626.07/26.2211.345.52/8.6816.334.4474.07
" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.386, + 0.785, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.505 + ], + "angle": 0, + "content": "Fig. 2: Qualitative results per action in AssemblyHands. We show Base results with \"verb (noun)\" actions. The left three figures are lower error situations while the right four ones are failure cases. The red boxes denote the area where the action occurs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.568 + ], + "angle": 0, + "content": "out by significantly lowering CDev errors by \\(32.7\\%\\) in allocentric and \\(27.2\\%\\) in egocentric settings compared to the baseline." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.595, + 0.465, + 0.611 + ], + "angle": 0, + "content": "5.2 AssemblyHands analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.787, + 0.657 + ], + "angle": 0, + "content": "We provide analysis, regarding action-wise evaluation, distortion effect in training, and the effect of multi-view fusion. See SupMat for additional results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Action-wise evaluation: To analyze errors related to hand-object occlusions and interactions, we show pose evaluation according to fine verb action classes in Table 3. We use the verb classes annotated by Assembly101 [61], spanning every few seconds in a video. Figure 2 shows qualitative results of representative verb classes with the top and bottom error cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We observe that the performance varies among different verb actions. The verbs \"tilt down/up\" and \"remove screw from\" exhibit lower errors among the submitted methods, because hands are less occluded and their movement is relatively stable. The \"tilt\" action holds a small part of the toy and turns it around alternately, leading to less overlap between the hand and the object (lib). The \"remove screw from\" action takes a screw out from the toy vehicle by their hand where observed hand poses do not change drastically." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.146, + 0.571, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.284, + 0.581, + 0.352 + ], + "angle": 0, + "content": "Fig. 3: Effect of distortion in AssemblyHands. The officially released images in the dataset have highly stretched areas near the edges (original crop). The method JHands uses a perspective crop with a virtual camera to correct this distortion." + }, + { + "type": "table_caption", + "bbox": [ + 0.6, + 0.158, + 0.788, + 0.254 + ], + "angle": 0, + "content": "Table 5: Multi-view fusion in AssemblyHands. We use the Base result to show performance before and after fusion. Missing instances per view are denoted as \"Miss(\\%)\"." + }, + { + "type": "table", + "bbox": [ + 0.603, + 0.256, + 0.782, + 0.353 + ], + "angle": 0, + "content": "
ViewMPJPEMiss(%)
cam137.9770.8
cam225.7188.3
cam322.190.92
cam422.290.74
cam3+421.520.08
all four20.690
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.521 + ], + "angle": 0, + "content": "Higher error classes, such as \"inspect\", \"screw\", \"rotate\", and \"position\", contain heavy occlusions, fast hand motion, complex two hands and object interactions. The \"inspect\" action brings the toy close to the human eyes where the toy occupies a large portion of the image causing heavy object occlusions. The \"screw\" action involves intricate fingertip movements to rotate the screwdriver quickly. The \"rotate\" and \"position\" actions are performed so that the two hands and the object interact in close contact, which complicates the estimation. We observe that the top two methods JHands and PICO-AI significantly correct the results of these higher error actions compared to the other submitted methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Bias of hand position in an image: Hands near image edges are highly distorted due to the fish-eye cameras. Directly using these noisy images in training will degrade performance [84]; thus, some methods create new crops with less distortion, select training instances, or adaptively fuse predictions during the inference. Specifically, Figure 3 shows that the method JHands reformulates the perspective during cropping and creates less-distorted (perspective) crops." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.748 + ], + "angle": 0, + "content": "To study this effect in the final performance, we split the evaluation instances into classes with different 2D distances between the hand center and the image center in Table 3. Higher distances (250- pixels) indicate closer hand crops to image edges. The method, Base, without any training instance selection and distortion correction, has higher error as the crops approach image edges (20.31 \\(\\rightarrow\\) 24.85). In contrast, the newly proposed methods are more robust and have a lower error, particularly in the 200-250 range. We observe that the ranges 200-250 and 250- occupy \\(10\\%\\) and \\(5\\%\\) of the test images, respectively, thus the improvement in the 200-250 range helps the lowering of the overall score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Effect of multi-view fusion: The multi-view egocentric camera setup is unique to the dataset. We show the statistics and performance of multi-view fusion in Table 5. Note that Table 3 shows the final results after multi-view fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We found that samples captured from the lower cameras (cam3 and cam4; see Table 1 for the layout) are numerous (fewer missing samples) and their errors are lower as they are faced toward the area occurring hand interactions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.498, + 0.128 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.147, + 0.497, + 0.156 + ], + "angle": 0, + "content": "(a) Egocentric-allocentric comparison" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.156, + 0.501, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.147, + 0.671, + 0.156 + ], + "angle": 0, + "content": "(b) Egocentric view" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.156, + 0.729, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.242, + 0.671, + 0.251 + ], + "angle": 0, + "content": "(c) Allocentric view" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.251, + 0.728, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.349, + 0.785, + 0.392 + ], + "angle": 0, + "content": "Fig. 4: Performance comparison: Egocentric vs Allocentric. (a) Comparative difficulty ratio of egocentric to allocentric views. (b) Egocentric view performance by method across objects. (c) Allocentric view performance by method across objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.785, + 0.543 + ], + "angle": 0, + "content": "Conversely, the samples from cam1 and cam2 are fewer and unbalanced as their cameras often fail to capture hands due to the camera layout. For instance, cam1 (top-left) tends to capture more hand region than cam2 as the participants are mostly right-handed and bring up the object with the right hand, which can be better observed from cam1. Given this uneven sample distribution, the proposed adaptive view selection methods in either training or testing are essential to perform effective multi-view fusion, and outperform the Base's test-time average using all views all the time (see Multi-view fusion in Section 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.787, + 0.651 + ], + "angle": 0, + "content": "We further study the performance gain before and after multi-view fusion using Base's results. While per-view performance achieves 22.19 and 22.29 in cam3 and cam4, respectively, their fused results with simple average reduce the error to 21.52. Merging all four views has shown to be more effective than two-view fusion (20.69 vs. 21.52), indicating a \\(6.5\\%\\) reduction compared to the single camera setup (cam3). This suggests predictions from the top views (cam1 and cam2) are informative in averaging even when they are prone to be erroneous." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.409, + 0.691 + ], + "angle": 0, + "content": "5.3 ARCTIC analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Egocentric-allocentric comparison: Figure 4a compares the performance between egocentric and allocentric views. In particular, we compute a ratio between the metric values of the egocentric and allocentric view to measure the extent of difficulty for the egocentric view compared to the allocentric view. Since success rate is a metric whose value is positively correlated to performance, we take its reciprocal ratio. We average the ratios across methods and actions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We observe that hand pose-related metrics such as MPJPE and \\(\\mathrm{ACC}_h\\) are less than 1.0 on average (see blue color cells), meaning the the egocentric view is easier than allocentric view. This is because most allocentric cameras in ARCTIC" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.142, + 0.734, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.275 + ], + "angle": 0, + "content": "Fig. 5: Egocentric reconstruction by top method in ARCTIC. In the egocentric view, object reconstruction struggle when the object is partially observed on the image boundaries, as well as when heavy hand/arm occlusion occurs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.784, + 0.367 + ], + "angle": 0, + "content": "are meters away from the subject while the egocentric camera is often close-up, offering higher hand visibility. Relative translation metrics between hand and object such as \\(\\mathrm{MRRPE}_{rl}\\) are also easier in the egocentric view because estimating translation is more difficult from further cameras." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.789, + 0.565 + ], + "angle": 0, + "content": "Object reconstruction performance faces unique challenges, as highlighted by the red cells. In the egocentric view, it is notably more difficult to reconstruct accurate object surfaces, articulation (AAE), and hand-object contact (CDev). This increased difficulty arises because objects are often positioned at the image edges and obscured by human arms. Additionally, object poses exhibit greater diversity in the egocentric view due to varying camera angles and occlusions. While a static camera maintains consistent camera extrinsics across a sequence, an egocentric camera's extrinsics change with each frame, resulting in higher diversity in camera-view object 6D poses. This diversity complicates object pose estimation in egocentric views. Figure 5 illustrates these challenges using the best-performing method, JointTransformer. Despite achieving reasonable hand poses, object poses are significantly impacted by occlusions from hands and arms, and the egocentric view undergoes substantial changes throughout a sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.672 + ], + "angle": 0, + "content": "Object-wise evaluation: Figure 4b and 4c break down performances on different objects in the egocentric and allocentric view test sets. The best method in both cases is JointTransformer. The hardest object to reconstruct with good contact consistency (see CDev) is the microwave in both settings due to global rotation sensitivity (see Fig. 5), though this can be mitigated by a keypoint-based approach [24]. Estimating objects in the egocentric view is more difficult than in the allocentric view, which is indicated by higher errors for all methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.789, + 0.779 + ], + "angle": 0, + "content": "Action-wise evaluation: Figure 6 compares performance of different methods in \"grab\" and \"use\" actions. In ARCTIC, there are sequences to interact with the object with two types of actions by either not articulating the object, or allowing object articulation. Interestingly, the \"grab\" motion is more challenging in egocentric and allocentric views. We hypothesize that this is because there are more diverse object poses for \"grab\" motions since during object articulation, the participants often focus on articulation instead of object manipulation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Effect of model size: Figure 7 illustrates the impact of model size on handobject contact performance, measured by CDev, for reconstruction on the allocentric validation set. Most methods utilize ResNet50, with JointTransformer being the top performer. As the trainable parameters in the backbone increase," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.128 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.165, + 0.501, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.325, + 0.504, + 0.424 + ], + "angle": 0, + "content": "Fig.6: Hand-object contact quality for reconstructed results per action. We evaluate the contact quality of the 3D reconstruction results from all methods for each action (i.e., grab or use), using Contact Deviation (CDev) in mm as the metric, where lower values indicate better quality." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.147, + 0.782, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.325, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Fig. 7: Contact deviation vs. model size. We assess the contact quality of the reconstruction results, varying by the number of parameters in each model. Contact quality is measured using Contact Deviation (CDev) in mm, with lower values indicating superior results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.545 + ], + "angle": 0, + "content": "JointTransformer consistently reduces the CDev error. Note that the x-axis is in log-scale. JointTransformer achieved a CDev error of 30.5mm with ViT-L and 29.0mm with ViT-G, which has ten times more parameters than ViT-L. Interestingly, JointTransformer uses frozen weights in the large-scale ViT-L and ViT-G backbones, yet achieves the best results. This suggests a potential direction for leveraging large-scale foundational backbones for hand-object reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.36, + 0.584 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.719 + ], + "angle": 0, + "content": "In this paper, we introduce the HANDS23 challenge and provide analysis based on the results of the top submitted methods and more recent baselines on the leaderboards. We organize and compare the submissions and their implementation details based on the learning methods, architecture, pre- and post-processing techniques, and training configurations. We thoroughly analyze various aspects, such as hand-object occlusions, action and object-wise evaluation, distortion correction, multi-view fusion, egocentric-allocentric comparison, and performance gain of large transformer models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Future directions: There are several future directions that the community can take. For example, one can explore more efficient training using multi-view egocentric cameras, leveraging 3D foundation priors [38,54] to regularize template-free hand-object reconstruction [15], estimating hand-object poses with more expressive representations (e.g., heatmap-based approaches [24]), incorporating motion and temporal modeling [18], featuring more diverse egocentric interaction scenarios, recognizing actions through captured hand poses [62], learning robotic grasping from reconstructed hand-object pose sequences, and so forth." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.283 + ], + "angle": 0, + "content": "This work is supported by the Ministry of Science and ICT, Korea, under the ITRC program (IITP-2024-2020-0-01789). supervised by the IITP. UTokyo is supported by JST ACT-X Grant Number JPMJAX2007, JSPS KAKENHI Grant Number JP22KJ0999, JST Adopting Sustainable Partnerships for Innovative Research Ecosystem (ASPIRE) Grant Number JPMJAP2303. NUS is supported by the Ministry of Education, Singapore, under its MOE Academic Research Fund Tier 2 (STEM RIE2025 MOE-T2EP20220-0015)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.303, + 0.323, + 0.318 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Abou Zeid, K.: JointTransformer: Winner of the HANDS'2023 ARCTIC Challenge @ ICCV. https://github.com/kabouzeid/JointTransformer (2023) 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.785, + 0.415 + ], + "angle": 0, + "content": "2. Armagan, A., Garcia-Hernando, G., Baek, S., Hampali, S., Rad, M., Zhang, Z., Xie, S., Chen, M., Zhang, B., Xiong, F., et al.: Measuring generalisation to unseen viewpoints, articulations, shapes and objects for 3d hand pose estimation under hand-object interaction. In: ECCV. pp. 85-101. Springer (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.416, + 0.785, + 0.455 + ], + "angle": 0, + "content": "3. Boukhayma, A., de Bem, R., Torr, P.H.S.: 3D hand shape and pose from images in the wild. In: Computer Vision and Pattern Recognition (CVPR). pp. 10843-10852 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.497 + ], + "angle": 0, + "content": "4. Cai, Y., Ge, L., Cai, J., Yuan, J.: Weakly-supervised 3D hand pose estimation from monocular RGB images. In: European Conference on Computer Vision (ECCV). pp. 678-694 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.785, + 0.538 + ], + "angle": 0, + "content": "5. Cao, Z., Radosavovic, I., Kanazawa, A., Malik, J.: Reconstructing hand-object interactions in the wild. In: International Conference on Computer Vision (ICCV). pp. 12417-12426 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.594 + ], + "angle": 0, + "content": "6. Chao, Y.W., Yang, W., Xiang, Y., Molchanov, P., Handa, A., Tremblay, J., Narang, Y.S., Van Wyk, K., Iqbal, U., Birchfield, S., Kautz, J., Fox, D.: DexYCB: A benchmark for capturing hand grasping of objects. In: Computer Vision and Pattern Recognition (CVPR). pp. 9044-9053 (2021) 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.632 + ], + "angle": 0, + "content": "7. Chatterjee, D., Sener, F., Ma, S., Yao, A.: Opening the vocabulary of egocentric actions. Conference on Neural Information Processing Systems (NeurIPS) 36 (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.676 + ], + "angle": 0, + "content": "8. Chen, X., Liu, Y., Dong, Y., Zhang, X., Ma, C., Xiong, Y., Zhang, Y., Guo, X.: MobRecon: mobile-friendly hand mesh reconstruction from monocular image. In: Computer Vision and Pattern Recognition (CVPR). pp. 20512-20522 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "9. Chen, X., Wang, B., Shum, H.Y.: Hand avatar: Free-pose hand animation and rendering from monocular video. In: Computer Vision and Pattern Recognition (CVPR) (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "0. Chen, Z., Chen, S., Schmid, C., Laptev, I.: gSDF: Geometry-driven signed distance functions for 3d hand-object reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 12890-12900 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "1. Christen, S., Kocabas, M., Aksan, E., Hwangbo, J., Song, J., Hilliges, O.: D-Grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In: Computer Vision and Pattern Recognition (CVPR). pp. 20545–20554 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "2. Corona, E., Pumarola, A., Alenyà, G., Moreno-Noguer, F., Rogez, G.: GanHand: Predicting human grasp affordances in multi-object scenes. In: Computer Vision and Pattern Recognition (CVPR). pp. 5030-5040 (2020) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.128 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Duran, E., Kocabas, M., Choutas, V., Fan, Z., Black, M.J.: HMP: Hand motion priors for pose and shape estimation from video. In: Winter Conference on Applications of Computer Vision (WACV) (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "14. Erol, A., Bebis, G., Nicoescu, M., Boyle, R.D., Twombly, X.: Vision-based hand pose estimation: A review. Computer Vision and Image Understanding (CVIU) 108(1-2), 52-73 (2007) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.229, + 0.785, + 0.271 + ], + "angle": 0, + "content": "15. Fan, Z., Parelli, M., Kadoglou, M.E., Kocabas, M., Chen, X., Black, M.J., Hilliges, O.: HOLD: Category-agnostic 3d reconstruction of interacting hands and objects from video. In: Computer Vision and Pattern Recognition (CVPR) (2024) 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.323 + ], + "angle": 0, + "content": "16. Fan, Z., Spurr, A., Kocabas, M., Tang, S., Black, M.J., Hilliges, O.: Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In: International Conference on 3D Vision (3DV). pp. 1-10 (2021) 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.324, + 0.785, + 0.379 + ], + "angle": 0, + "content": "17. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "18. Fu, Q., Liu, X., Xu, R., Niebles, J.C., Kitani, K.M.: Deformer: Dynamic fusion transformer for robust hand pose estimation. In: International Conference on Computer Vision (ICCV). pp. 23600-23611 (October 2023) 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "19. Garcia-Hernando, G., Yuan, S., Baek, S., Kim, T.K.: First-person hand action benchmark with rgb-d videos and 3d hand pose annotations. In: Computer Vision and Pattern Recognition (CVPR) (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "20. Ge, L., Ren, Z., Li, Y., Xue, Z., Wang, Y., Cai, J., Yuan, J.: 3D hand shape and pose estimation from a single rgb image. In: Computer Vision and Pattern Recognition (CVPR). pp. 10833-10842 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "21. Grady, P., Tang, C., Twigg, C.D., Vo, M., Brahmbhatt, S., Kemp, C.C.: ContactOpt: Optimizing contact to improve grasps. In: Computer Vision and Pattern Recognition (CVPR). pp. 1471-1481 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "22. Guo, Z., Zhou, W., Wang, M., Li, L., Li, H.: HandNeRF: Neural radiance fields for animatable interacting hands. In: Computer Vision and Pattern Recognition (CVPR). pp. 21078-21087 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.582, + 0.785, + 0.623 + ], + "angle": 0, + "content": "23. Hampali, S., Rad, M., Oberweger, M., Lepetit, V.: HOnnotate: A method for 3D annotation of hand and object poses. In: Computer Vision and Pattern Recognition (CVPR). pp. 3193-3203 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.622, + 0.785, + 0.677 + ], + "angle": 0, + "content": "24. Hampali, S., Sarkar, S.D., Rad, M., Lepetit, V.: Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3D pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 11090-11100 (2022) 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.677, + 0.785, + 0.745 + ], + "angle": 0, + "content": "25. Han, S., Wu, P.C., Zhang, Y., Liu, B., Zhang, L., Wang, Z., Si, W., Zhang, P., Cai, Y., Hodan, T., Cabezas, R., Tran, L., Akbay, M., Yu, T.H., Keskin, C., Wang, R.: Umetrack: Unified multi-view end-to-end hand tracking for VR. In: International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH). pp. 50:1-50:9. ACM (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.745, + 0.785, + 0.798 + ], + "angle": 0, + "content": "26. Hasson, Y., Tekin, B., Bogo, F., Laptev, I., Pollefeys, M., Schmid, C.: Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 568-577 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "27. Hasson, Y., Varol, G., Schmid, C., Laptev, I.: Towards unconstrained joint handobject reconstruction from rgb videos. In: International Conference on 3D Vision (3DV). pp. 659-668. IEEE (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "28. Hasson, Y., Varol, G., Tzionas, D., Kalevatykh, I., Black, M.J., Laptev, I., Schmid, C.: Learning joint reconstruction of hands and manipulated objects. In: Computer Vision and Pattern Recognition (CVPR). pp. 11807-11816 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "30. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: Computer Vision and Pattern Recognition (CVPR). pp. 15979-15988 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "31. Huang, D., Ji, X., He, X., Sun, J., He, T., Shuai, Q., Ouyang, W., Zhou, X.: Reconstructing hand-held objects from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "32. Iqbal, U., Molchanov, P., Gall, T.B.J., Kautz, J.: Hand pose estimation via latent 2.5D heatmap regression. In: European Conference on Computer Vision (ECCV). pp. 118-134 (2018) 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.384 + ], + "angle": 0, + "content": "33. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2O: Two hands manipulating objects for first person interaction recognition. In: International Conference on Computer Vision (ICCV). pp. 10138-10148 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "34. Lee, J., Sung, M., Choi, H., Kim, T.K.: Im2hands: Learning attentive implicit representation of interacting two-hand shapes. In: Computer Vision and Pattern Recognition (CVPR). pp. 21169-21178 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.478 + ], + "angle": 0, + "content": "35. Li, L., Tian, L., Zhang, X., Wang, Q., Zhang, B., Bo, L., Liu, M., Chen, C.: Renderih: A large-scale synthetic dataset for 3d interacting hand pose estimation. In: International Conference on Computer Vision (ICCV). pp. 20395-20405 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.522 + ], + "angle": 0, + "content": "36. Li, M., An, L., Zhang, H., Wu, L., Chen, F., Yu, T., Liu, Y.: Interacting attention graph for single image two-hand reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 2761-2770 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "37. Liu, R., Ohkawa, T., Zhang, M., Sato, Y.: Single-to-dual-view adaptation for egocentric 3d hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 677-686 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "38. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.633 + ], + "angle": 0, + "content": "39. Liu, S., Jiang, H., Xu, J., Liu, S., Wang, X.: Semi-supervised 3D hand-object poses estimation with interactions in time. In: Computer Vision and Pattern Recognition (CVPR). pp. 14687-14697 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: International Conference on Computer Vision (ICCV). pp. 10012-10022 (2021) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "41. Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Computer Vision and Pattern Recognition (CVPR). pp. 11966-11976 (2022) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "42. Meng, H., Jin, S., Liu, W., Qian, C., Lin, M., Ouyang, W., Luo, P.: 3D interacting hand pose estimation by hand de-occlusion and removal. In: European Conference on Computer Vision (ECCV). pp. 380–397. Springer (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "43. Moon, G.: Bringing inputs to shared domains for 3d interacting hands recovery in the wild. In: Computer Vision and Pattern Recognition (CVPR). pp. 17028-17037 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "44. Moon, G., Saito, S., Xu, W., Joshi, R., Buffalini, J., Bellan, H., Rosen, N., Richardson, J., Mize, M., De Bree, P., et al.: A dataset of relighted 3d interacting hands. Conference on Neural Information Processing Systems (NeurIPS) 36 (2024) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "45. Moon, G., Yu, S., Wen, H., Shiratori, T., Lee, K.M.: Interhand2.6m: A dataset and baseline for 3d interacting hand pose estimation from a single RGB image. In: European Conference on Computer Vision (ECCV). vol. 12365, pp. 548-564 (2020) 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.786, + 0.245 + ], + "angle": 0, + "content": "46. Mueller, F., Bernard, F., Sotnychenko, O., Mehta, D., Sridhar, S., Casas, D., Theobalt, C.: GANerated hands for real-time 3D hand tracking from monocular RGB. In: Computer Vision and Pattern Recognition (CVPR). pp. 49-59 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "47. Ohkawa, T.: AssemblyHands Toolkit. https://github.com/facebookresearch/assemblyhands-toolkit (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.314 + ], + "angle": 0, + "content": "48. Ohkawa, T., Furuta, R., Sato, Y.: Efficient annotation and learning for 3d hand pose estimation: A survey. International Journal of Computer Vision (IJCV) 131, 3193-3206 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.357 + ], + "angle": 0, + "content": "49. Ohkawa, T., He, K., Sener, F., Hodan, T., Tran, L., Keskin, C.: AssemblyHands: towards egocentric activity understanding via 3d hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 12999-13008 (2023) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.786, + 0.398 + ], + "angle": 0, + "content": "50. Ohkawa, T., Li, Y.J., Fu, Q., Furuta, R., Kitani, K.M., Sato, Y.: Domain adaptive hand keypoint and pixel localization in the wild. In: European Conference on Computer Vision (ECCV). pp. 68—87 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.439 + ], + "angle": 0, + "content": "51. Ohkawa, T., Yagi, T., Hashimoto, A., Ushiku, Y., Sato, Y.: Foreground-aware stylization and consensus pseudo-labeling for domain adaptation of first-person hand segmentation. IEEE Access 9, 94644-94655 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.508 + ], + "angle": 0, + "content": "52. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.55 + ], + "angle": 0, + "content": "53. Park, J., Oh, Y., Moon, G., Choi, H., Lee, K.M.: HandOccNet: Occlusion-robust 3d hand mesh estimation network. In: Computer Vision and Pattern Recognition (CVPR). pp. 1496–1505 (2022) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.577 + ], + "angle": 0, + "content": "54. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.786, + 0.618 + ], + "angle": 0, + "content": "55. Prakash, A., Tu, R., Chang, M., Gupta, S.: 3d hand pose estimation in everyday egocentric images. In: European Conference on Computer Vision (ECCV) (2024) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.786, + 0.66 + ], + "angle": 0, + "content": "56. Radosavovic, I., Kosaraju, R.P., Girshick, R.B., He, K., Dollár, P.: Designing network design spaces. In: Computer Vision and Pattern Recognition (CVPR). pp. 10425-10433 (2020) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.786, + 0.702 + ], + "angle": 0, + "content": "57. Rehg, J.M., Kanade, T.: Visual tracking of high DOF articulated structures: An application to human hand tracking. In: European Conference on Computer Vision (ECCV). vol. 801, pp. 35-46 (1994) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.744 + ], + "angle": 0, + "content": "58. Remelli, E., Han, S., Honari, S., Fua, P., Wang, R.: Lightweight multi-view 3d pose estimation through camera-disentangled representation. In: Computer Vision and Pattern Recognition (CVPR). pp. 6039-6048 (2020) 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.785 + ], + "angle": 0, + "content": "59. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG) 36(6), 245:1-245:17 (2017) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.786, + 0.842 + ], + "angle": 0, + "content": "60. Ryali, C., Hu, Y., Bolya, D., Wei, C., Fan, H., Huang, P., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: International Conference on Machine Learning (ICML). vol. 202, pp. 29441-29454 (2023) 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.205 + ], + "angle": 0, + "content": "61. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Computer Vision and Pattern Recognition (CVPR). pp. 21064-21074 (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "62. Shamil, M.S., Chatterjee, D., Sener, F., Ma, S., Yao, A.: On the utility of 3d hand poses for action recognition. In: European Conference on Computer Vision (ECCV) (2024) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.247, + 0.786, + 0.288 + ], + "angle": 0, + "content": "63. Simon, T., Joo, H., Matthews, I., Sheikh, Y.: Hand keypoint detection in single images using multiview bootstrapping. In: Computer Vision and Pattern Recognition (CVPR). pp. 4645-4653 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.289, + 0.787, + 0.331 + ], + "angle": 0, + "content": "64. Spurr, A., Dahiya, A., Wang, X., Zhang, X., Hilliges, O.: Self-supervised 3D hand pose estimation from monocular RGB via contrastive learning. In: International Conference on Computer Vision (ICCV). pp. 11210-11219 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "65. Spurr, A., Iqbal, U., Molchanov, P., Hilliges, O., Kautz, J.: Weakly supervised 3D hand pose estimation via biomechanical constraints. In: European Conference on Computer Vision (ECCV). vol. 12362, pp. 211-228 (2020) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "66. Spurr, A., Song, J., Park, S., Hilliges, O.: Cross-modal deep variational hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 89-98 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.417, + 0.787, + 0.459 + ], + "angle": 0, + "content": "67. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Computer Vision and Pattern Recognition (CVPR) (2019) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.459, + 0.787, + 0.514 + ], + "angle": 0, + "content": "68. Swamy, A., Leroy, V., Weinzaepfel, P., Baradel, F., Galaoui, S., Brégier, R., Armando, M., Franco, J.S., Rogez, G.: SHOWMe: Benchmarking object-agnostic hand-object 3d reconstruction. In: International Conference on Computer Vision (ICCV). pp. 1935–1944 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.515, + 0.787, + 0.557 + ], + "angle": 0, + "content": "69. Tekin, B., Bogo, F., Pollefeys, M.: \\(\\mathrm{H} + \\mathrm{O}\\): Unified egocentric recognition of 3D hand-object poses and interactions. In: Computer Vision and Pattern Recognition (CVPR). pp. 4511-4520 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.558, + 0.787, + 0.6 + ], + "angle": 0, + "content": "70. Tse, T.H.E., Kim, K.I., Leonardis, A., Chang, H.J.: Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In: Computer Vision and Pattern Recognition (CVPR). pp. 1664-1674 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "71. Tzionas, D., Gall, J.: A comparison of directional distances for hand pose estimation. In: German Conference on Pattern Recognition (GCPR). vol. 8142, pp. 131-141 (2013) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "72. Wen, Y., Pan, H., Ohkawa, T., Yang, L., Pan, J., Sato, Y., Komura, T., Wang, W.: Generative hierarchical temporal transformer for hand action recognition and motion prediction. CoRR abs/2311.17366 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "73. Yang, L., Chen, S., Yao, A.: Semihand: Semi-supervised hand pose estimation with consistency. In: International Conference on Computer Vision (ICCV). pp. 11364-11373 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.728, + 0.787, + 0.77 + ], + "angle": 0, + "content": "74. Yang, L., Zhan, X., Li, K., Xu, W., Li, J., Lu, C.: CPF: Learning a contact potential field to model the hand-object interaction. In: International Conference on Computer Vision (ICCV) (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.771, + 0.787, + 0.798 + ], + "angle": 0, + "content": "75. Ye, Y., Gupta, A., Tulsiani, S.: What's in your hands? 3d reconstruction of generic objects in hands. In: Computer Vision and Pattern Recognition (CVPR) (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "76. Ye, Y., Hebbar, P., Gupta, A., Tulsiani, S.: Diffusion-guided reconstruction of everyday hand-object interaction clips. In: International Conference on Computer Vision (ICCV) (2023) 3" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.231 + ], + "angle": 0, + "content": "77. Yuan, S., Garcia-Hernando, G., Stenger, B., Moon, G., Chang, J.Y., Lee, K.M., Molchanov, P., Kautz, J., Honari, S., Ge, L., Yuan, J., Chen, X., Wang, G., Yang, F., Akiyama, K., Wu, Y., Wan, Q., Madadi, M., Escalera, S., Li, S., Lee, D., Oikonomidis, I., Argyros, A.A., Kim, T.: Depth-based 3D hand pose estimation: From current achievements to future goals. In: Computer Vision and Pattern Recognition (CVPR). pp. 2636-2645 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "78. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Computer Vision and Pattern Recognition (CVPR). pp. 12104-12113 (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "79. Zhang, H., Christen, S., Fan, Z., Hilliges, O., Song, J.: GraspXL: Generating grasping motions for diverse objects at scale. European Conference on Computer Vision (ECCV) (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.786, + 0.342 + ], + "angle": 0, + "content": "80. Zhang, H., Christen, S., Fan, Z., Zheng, L., Hwangbo, J., Song, J., Hilliges, O.: ArtiGrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. In: International Conference on 3D Vision (3DV) (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "81. Zhang, X., Li, Q., Mo, H., Zhang, W., Zheng, W.: End-to-end hand mesh recovery from a monocular RGB image. In: International Conference on Computer Vision (ICCV). pp. 2354-2364 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.425 + ], + "angle": 0, + "content": "82. Zhou, Y., Habermann, M., Xu, W., Habibie, I., Theobalt, C., Xu, F.: Monocular real-time hand shape and motion capture using multi-modal data. In: Computer Vision and Pattern Recognition (CVPR). pp. 5345-5354 (2020) 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.786, + 0.452 + ], + "angle": 0, + "content": "83. Zhou, Z.: SimpleHand: Winner of the HANDS'2023 AssemblyHands Challenge @ ICCV. https://github.com/patienceFromZhou/simpleHand (2024) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "84. Zhou, Z., Lv, Z., Zhou, S., Zou, M., Wu, T., Yu, M., Tang, Y., Liang, J.: 1st place solution of egocentric 3d hand pose estimation challenge 2023 technical report: A concise pipeline for egocentric hand pose reconstruction. CoRR abs/2310.04769 (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.55 + ], + "angle": 0, + "content": "85. Zhou, Z., Zhou, S., Lv, Z., Zou, M., Tang, Y., Liang, J.: A simple baseline for efficient hand mesh reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 1367-1376 (2024) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "86. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (ICLR). OpenReview.net (2021), https://openreview.net/forum?id=gZ9hCDWe6ke 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "87. Ziani, A., Fan, Z., Kocabas, M., Christen, S., Hilliges, O.: TempCLR: Reconstructing hands via time-coherent contrastive learning. In: International Conference on 3D Vision (3DV). pp. 627-636 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "88. Zimmermann, C., Brox, T.: Learning to estimate 3D hand pose from single RGB images. In: International Conference on Computer Vision (ICCV). pp. 4913-4921 (2017) 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_origin.pdf b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..21a944014b8f649a0b32f2b58740c5d1f4386ab0 --- /dev/null +++ b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/add6d04e-f1bb-4e0c-8828-53e7d24bd3ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9a2cde455a8e9078ab48d81d46843e26198ed8c6e7a8985476495224ccf4061 +size 2538882 diff --git a/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/full.md b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c3b1812d48700b8842870b0a10d44c681a96e4ba --- /dev/null +++ b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/full.md @@ -0,0 +1,316 @@ +# Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects + +Zicong Fan $^{1,2*}$ , Takehiko Ohkawa $^{3*}$ , Linlin Yang $^{4*}$ , Nie Lin $^{3}$ , Zhishan Zhou $^{5}$ , Shihao Zhou $^{5}$ , Jiajun Liang $^{5}$ , Zhong Gao $^{6}$ , Xuanyang Zhang $^{6}$ , Xue Zhang $^{7}$ , Fei Li $^{7}$ , Liu Zheng $^{8}$ , Feng Lu $^{8}$ , Karim Abou Zeid $^{9}$ , Bastian Leibe $^{9}$ , Jeongwan On $^{10}$ , Seungryul Baek $^{10}$ , Aditya Prakash $^{11}$ , Saurabh Gupta $^{11}$ , Kun He $^{12}$ , Yoichi Sato $^{3}$ , Otmar Hilliges $^{1}$ , Hyung Jin Chang $^{13}$ , Angela Yao $^{14}$ + +Abstract. We interact with the world with our hands and see it through our own (egocentric) perspective. A holistic 3D understanding of such interactions from egocentric views is important for tasks in robotics, AR/VR, action recognition and motion generation. Accurately reconstructing such interactions in 3D is challenging due to heavy occlusion, viewpoint bias, camera distortion, and motion blur from the head movement. To this end, we designed the HANDS23 challenge based on the AssemblyHands and ARCTIC datasets with carefully designed training and testing splits. Based on the results of the top submitted methods and more recent baselines on the leaderboards, we perform a thorough analysis on 3D hand(-object) reconstruction tasks. Our analysis demonstrates the effectiveness of addressing distortion specific to egocentric cameras, adopting high-capacity transformers to learn complex hand-object interactions, and fusing predictions from different views. Our study further reveals challenging scenarios intractable with state-of-the-art methods, such as fast hand motion, object reconstruction from narrow egocentric views, and close contact between two hands and objects. Our efforts will enrich the community's knowledge foundation and facilitate future hand studies on egocentric hand-object interactions. + +# 1 Introduction + +We interact with the world with our hands and see it through our eyes: we wake up and grab our phone to check the time; we use tools when assembling parts of a car; we open the microwave door to heat food, to name a few. An egocentric 3D understanding of our hand interactions with objects will fundamentally impact + +![](images/467e6a201dc213513725c5c616426b8a315bc6ff1a58b39b1f4df0e30a92146d.jpg) +Fig. 1: Tasks in HANDS23. In AssemblyHands, from its multi-view headset (a), we estimate 3D hand poses from images (b); In ARCTIC, we estimate the poses of two hands and articulated objects from an image (c). + +areas such as robotics grasping [11,80], augmented and virtual reality [25], action recognition [7,19,72] and motion generation [79,80]. + +However, it is non-trivial to accurately reconstruct 3D hands and or objects due to its high degree of freedoms [65, 73], ambiguous texture [16], and heavy occlusions. These challenges are intensified in an egocentric view [50, 51], particularly with object interactions, due to significant camera distortion, rapid and varied changes in viewpoint caused by head movements and hand-object occlusion. To better understand these challenges, we introduce a public challenge in conjunction with ICCV 2023 (i.e., HANDS23) based on recent egocentric hand datasets, AssemblyHands [49] and ARCTIC [17] (see Figure 1). These two datasets are large-scale, multi-view, and provide monochrome or RGB egocentric videos of the hands dexterously manipulating objects. Accordingly, we host two tasks: 1) egocentric 3D hand pose estimation from a single-view image based on AssemblyHands, and 2) consistent motion reconstruction based on ARCTIC. + +We introduce new HANDS23 methods and recent dataset leaderboard baselines that substantially outperform initial baselines for both tasks, setting new benchmarks for subsequent comparisons on the datasets. The two datasets are significantly larger and include a wider variety of bimanual manipulations compared to earlier datasets [6, 23], enabling a more authentic assessment of real-world interactions. With these benchmarks, we thoroughly analyze factors such as viewpoint, action types, hand position, model size, and object variations to determine their impact on 3D hand(-object) reconstruction. + +Our findings show the success of addressing the distortion of egocentric cameras with explicit perspective cropping or implicit learning for the distortion bias. In addition, recent high-capacity vision transformers can learn complex hand-object interactions. Adaptive fusion techniques for multi-view predictions further boost performance. We also analyze the remaining challenges that are still difficult to handle with the recent methods, e.g., fast hand motion, object reconstruction from narrow and moving views, and intricate interactions and close contact between two hands and objects. + +To summarize, we contribute state-of-the-art baselines and gather the submitted methods for AssemblyHands and ARCTIC to foster future research on egocentric hand-object interactions. Furthermore, we thoroughly analyze the two benchmarks to provide insights for future directions in egocentric hand pose estimation and consistent motion reconstruction. + +# 2 Related work + +3D hand pose estimation: Reconstructing 3D hand poses has a long history [14,48] ever since the important work led by Rehg and Kanade [57]. A large body of research in this area focuses on single-hand reconstruction [3,4,6,9,13, 16,18,20,28,32,37,46,63-66,71,81,87,88]. For example, while a popular OpenPose library features 2D hand keypoints [63], Zimmermann et al. [88] initially extend to estimate 3D hand poses using deep convolutional networks. Ever since the release of InterHand2.6M [45] dataset, the community has increased focus on strongly interacting two hands reconstruction [22,33-36,42-45,49]. For example, Li et al. [35] and Moon et al. [44] use relighting techniques to augment InterHand2.6M with more natural lighting and diverse backgrounds. + +Hand-object reconstruction: The holistic reconstruction of hands and objects have increased interest in the hand community in recent years [6,12,17,21, 26-28,33,39,69,70,74,82]. Methods in this area mostly assume a given object template and jointly estimate the hand poses and the object rigid poses [12, 21,26-28,39,69,70,74,82] or articulated poses [17]. For example, Cao et al. [5] fits object templates to in-the-wild interaction videos. Liu et al. [39] introduce a semi-supervised learning framework via pseudo-groundtruth from temporal data to improve hand-object reconstruction. + +More recent methods do not assume object templates for hand-object reconstruction [10, 15, 28, 31, 68, 75, 76]. For example, Fan et al. [15] introduced the first category-agnostic method that reconstructs an articulated hand and object jointly from a video. Nonetheless, our challenges and insights on hand-object occlusions and camera distortions are still applicable to these more challenging template-free reconstruction settings. + +Public reports for the previous challenges (HANDS17 [77] and HANDS19 [2]) distilled the insights from individual review papers and practical techniques into comprehensive summaries to enrich the community's knowledge base. These past challenges use benchmarks which include depth-based hand pose estimation from egocentric views. Instead of depth sensors, the HANDS23 benchmarks are based on affordable and widely applicable image sensors, i.e., RGB and monochrome images. This paper further advances the analysis with unique insights, such as multi-view egocentric cameras, object reconstruction in contact, and modeling with recent transformers beyond conventional CNNs. + +# 3 HANDS23 challenge overview + +The workshop contains two hand-related 3D reconstruction challenges in hand-object strongly interacting settings. In this section, we introduce the two challenges and their evaluation criteria. + +# 3.1 Workshop challenges + +3D hand pose estimation in AssemblyHands: As illustrated in Figure 1, this task focuses on egocentric 3D hand pose estimation from a single-view im + +age based on AssemblyHands. The dataset provides multi-view captured videos of hand-object interaction while assembling and disassembling toy vehicles. In particular, it provides allocentric (fixed-view) and egocentric recordings and auxiliary cues like action, object, or context information for hand pose estimation. We refer readers to [49] for more dataset details. The training, validation, and testing sets contain 383K, 32K, and 62K monochrome images, respectively, all captured from egocentric cameras. During training, 3D hand keypoint coordinates, hand bounding boxes and camera intrinsic and extrinsic matrices for the four egocentric cameras attached to the headset are provided. The same information is provided during testing, minus the 3D keypoints. Unlike [49], given the availability of multi-view egocentric images, this task lets participants develop multi-view fusion based on the corresponding multi-view images. + +Consistent motion reconstruction in ARCTIC: Given an RGB image, the goal of this task is to estimate poses of hands and articulated objects to recover the 3D surfaces of the interaction (see Figure 1). We refer readers to [17] for more details. The ARCTIC dataset contains data of hands dexterously manipulating articulated objects and videos from $8 \times$ allocentric views and $1 \times$ egocentric views. The official splits of the ARCTIC dataset are used for training, validation, and testing. There are two sub-tasks: allocentric task and egocentric task. In the former, only allocentric images can be used for training and evaluation. For the latter, all images from the training set can be used for training while only the egocentric view images are used during evaluation. + +# 3.2 Evaluation criteria + +AssemblyHands evaluation: We use MPJPE as an evaluation metric in millimeters, comparing the model predictions against the ground-truth in world coordinates. We provide the intrinsic and extrinsic of the egocentric cameras to construct submission results defined in the world coordinates. Assuming that the human hand has a total $N_J$ joints, we denote wrist-relative coordinates of the prediction and ground-truth as $\hat{J} \in \mathbb{R}^{N_J \times 3}$ and $J \in \mathbb{R}^{N_J \times 3}$ , respectively. Given a joint visibility indicator $\gamma_i$ per joint $J_i$ , we compute the Euclidean distance between predicted and ground-truth joints as $\frac{1}{\sum_{i=1}^{N_J} \gamma_i} \sum_{i=1}^{N_J} \gamma_i \left\| \hat{J}_i - J_i \right\|_2$ . The visibility indicator offers per-joint binary labels, representing whether the annotated keypoints are visible from the given egocentric view. + +ARCTIC evaluation: Since the original ARCTIC paper [17] has a heavy focus on the quality of hand-object contact in the reconstructed hand and object meshes, we use Contact Deviation (CDev) introduced in the ARCTIC dataset as the main metric for the competition. In particular, this metric measures the extent to which a hand vertex deviates from the supposed contact object vertex in the prediction. Concretely, suppose that for a given frame, $\{(\mathbf{h}_i,\mathbf{o}_i)\}_{i = 1}^C$ are $C$ pairs of in-contact ( $< 3mm$ distance) hand-object vertices according to ground-truth, and $\{(\hat{\mathbf{h}}_i,\hat{\mathbf{o}}_i)\}_{i = 1}^C$ are the predictions respectively. The CDev metric is the average distance between $\hat{\mathbf{h}}_i$ and $\hat{\mathbf{o}}_i$ in millimeters, $\frac{1}{C}\sum_{i = 1}^{C}||\hat{\mathbf{h}}_i - \hat{\mathbf{o}}_i||$ . + +For completeness, we report all metrics introduced in ARCTIC. In particular, the task requires the reconstructed meshes to have accurate hand-object contact (CDev) and smooth motion (ACC). Additionally, during articulation or when carrying an object, it is crucial that the vertices of the hand and object which are in contact maintain synchronized movement (MDev). Moreover, we assess hand and object poses, alongside their relative movements, using metrics like MPJPE, AAE, Success Rate, and MRRPE. For detailed information, see [17]. + +# 4 Methods + +This section presents the methods in the two challenges and other competing methods on the leaderboards. Four methods outperform the baseline in both AssemblyHands and ARCTIC. + +# 4.1 AssemblyHands methods + +Participants develop methods that learn the mapping from egocentric images to 3D keypoints. The methods are categorized into: heatmap-based and regression-based approaches. Given the presence of complex hand-object interactions in the egocentric scenes, high-capacity transformer models and attention mechanisms addressing occluded regions have been proposed as the backbone networks. Table 1 summarizes the methods based on the learning, preprocessing, multi-view fusion, and post-processing approaches. + +Base: This method uses a heatmap-based framework based on heatmaps [45] with 2.5D representations [32] and a ResNet50 [29] backbone. The implementation can be found in [47]. + +JHands [85]: This method employs a regression-based approach with simple MLP heads for regressing 2D keypoints, root-relative 3D keypoints, and the global root depth. The regression training is empowered by a recent fast and strong vision transformer, Hiera [60], pre-trained with masked auto-encoder [30]. A multi-level feature fusion that concatenates the features of different layers is adopted for better feature extraction at different scales. The method additionally uses other publicly available datasets for training, namely FreiHAND [88], DexYCB [6], and CompHand [8]. The implementation is available in [83]. + +PICO-AI: This method proposes a heatmap voting scheme in addition to the 2.5D heatmaps. Due to their sparsity, the conventional heatmaps pose an imbalance problem between positive and negative samples in the loss function. Hence, the proposed voting mechanism aims to spread the loss evenly across the entire heatmaps. Given the initial guess of keypoints obtained from the heatmaps, the method defines a local region centered on the joint position and operates the soft-argmax within the region to obtain refined keypoint coordinates. This restricts the impact of background points, leading to more reliable optimization. The training is facilitated by CNN-based RegNety320 [56]. + +Phi-AI: While following the heatmap-based approach, this method adapts MinimalHand [82] with the ResNet50 backbone, where 2D heatmaps and 3D location + +Table 1: Method and preprocessing summary in AssemblyHands. We summarize submitted methods in terms of learning methods, architecture, preprocessing, and multi-view fusion techniques. The tuple (views, phase) indicates the number of views used in either train or test time. + +
MethodLearning methodsArchitecturePreprocessingMulti-view fusion (views, phase)
Base2.5D heatmaps [45]ResNet50 [29]-Simple average (4, test)
JHands [85]RegressionHiera [60]Warp perspective, color jitter, random maskAdaptive view selection and average (2, test)
PICO-AI2.5D heatmaps [45] Heatmap votingRegNety320 [56]Scale, rotate, flip, translateAdaptive view selection FTL [58] (2, train)
FRDCRegression 2D heatmapsHandOccNet [53] with ConvNeXt [41]Scale, rotate, color jitterWeighted average (4, test)
Phi-AI2D heatmaps and 3D location maps [82]ResNet50 [29]Scale, rotate, translate, color jitter, gaussian blurWeighted average (4, test)
+ +maps are regressed. Instead of selecting 3D keypoint coordinates from the location maps, the proposed method modifies it by using heatmap values to weight 3D keypoint coordinates, achieving a more robust estimation. Moreover, the method adds a residual structured layer after the original three-tier cascade networks to refine the calculated location maps. The method further applies the ensemble of final keypoint outputs combined with the Base. + +FRDC: This method adopts a hybrid approach by combining regression with heatmap for training. HandOccNet [53] is modified to regress 3D keypoint coordinates and integrated with an additional branch of 2D heatmap regression. HandOccNet enriches feature extraction with spatial attention mechanisms for occluded regions, making it robust under hand-object occlusions. The method further utilizes a stronger ConvNeXt [41] backbone and feature fusion from the 2D keypoint regressor. + +Preprocessing of egocentric images: Compared to conventional static camera setups, egocentric images exhibit unique properties and biases, such as distortion, head camera motion, and different color representations. Thus, it is vital to preprocess egocentric images to alleviate these effects during training. Augmentation techniques are detailed in Table 1. + +The method JHands addresses the distortion issue with a warp perspective operation to make the hands near the edge less stretched. While AssemblyHands provides rectified images converted from fisheye cameras to a pinhole camera model, they often include excessively stretched areas near the edges. To address this, the method calculates a virtual camera and corresponding perspective transformation matrix based on the pixel coordinates of the crop and the camera parameters. The generated crops can be found in the analysis of Figure 3. + +Multi-view fusion: Since AssemblyHands offers multi-view egocentric videos, participants can optionally use the constraint of multi-view geometry and fusion techniques during training or inference. + +Table 2: Method and preprocessing summary in ARCTIC. We summarize baselines on ARCTIC in terms of input dimensions, image backbones, learning rate scheduling, training epochs, batch size and the cropping used for input. *Method trains 50 epochs for decoder and 36 for backbone. +Learning rate is 1e-7 to 1e-4 with linear warmup for first $5\%$ step, and 1e-4 to 1e-7 with cosine decay for rest. + +
MethodInput sizeBackbonesLearning rate scheduleTraining epochsBatch sizeCropping
ArcticNet-SF [17]224 × 224ResNet501e-5allocentric: 20 +egocentric: 5064object
DIGIT [16]224 × 224HRNet-W321e-5allocentric: 20 +egocentric: 5064object
AmbiguousHands224 × 224ResNet501e-5allocentric: 20 +egocentric: 10032hand +object
UVHand384 × 384Swin-L2e-4 (backbone) +1e-7 (others)allocentric: N/A +egocentric: 50/36*48object
JointTransformer [1]224 × 224ViT-G1e-7/1e-5+allocentric: 20 +egocentric: 10064object
+ +While Base uses a simple average of predicted keypoints from all four camera views in the test time, PICO-AI proposes multi-view feature fusion during training using Feature Transform Layers (FTL) [58]. This FTL training requires fusing two out of four views; thus, the method chooses the most suitable views for every frame. In cases with multiple candidates, the Intersection over Union (IoU) is computed between hand boxes from per-view predictions and 2D keypoints from previous 3D predictions. The two views with the highest IoUs are selected for their superior prediction reliability. + +The methods JHands, FRDC, and Phi-AI apply adaptive fusion in predicted keypoints during testing. The method JHands computes the MPJPE with each other view and selects two results of views with the lowest MPJPE, excluding noisy predictions in the average. If the MPJPE is lower than a threshold, the mean of the two results is calculated as the final result. Otherwise, the result with a lower PA-MPJPE with the predictions in the previous frame is chosen. The methods FRDC and Phi-AI use a weighted average for each view prediction, assigning weights based on each view's validation performance. + +Postprocessing: Several postprocessing techniques, including test-time augmentation, smoothing, and model ensembling are used to enhance inference outcomes. In particular, the method JHands applies an offline smooth (Savitzky-Golay) filter on each video sequence. + +# 4.2 ARCTIC methods + +Table 2 summarizes the details for each method in terms of the input image dimensions, image backbones, learning rate scheduling, the number of training epochs, batch size, and cropping strategies. + +Preliminary: All methods below are regression-based and predict two-hand MANO [59] parameters $\Theta = \{\theta, \beta\}$ and articulated object parameters $\Omega$ . In particular, with the MANO pose and shape parameters $\theta, \beta$ , the MANO model $\mathcal{H}$ returns a mesh with vertices via $\mathcal{H}(\theta, \beta) \in \mathbb{R}^{778 \times 3}$ . 3D joints are obtained via + +a linear regressor. The articulated object model $\mathcal{O}$ was introduced in ARCTIC to provide an articulated mesh with vertices via $\mathcal{O}(\Omega)\in \mathbb{R}^{\mathrm{V}\times 3}$ , where $\Omega \in \mathbb{R}^{7}$ contains the global orientation, global translation, and object articulation. + +ArcticNet-SF [17]: Introduced in ARCTIC, it is a single-frame baseline. It first extracts an image feature vector from the input image; then, it regresses hand and object parameters with simple MLPs. The hand and object meshes can then be extracted via $\mathcal{H}(\cdot)$ and $\mathcal{O}(\cdot)$ . For more details, see [17]. + +JointTransformer [1]: JointTransformer enhances ArcticNet-SF by integrating a transformer decoder instead of the MLP regressors for hand and object parameter estimation. The decoder employs learned queries for the angle of each joint, the shape and translation of each hand, and the translation, rotation, and articulation of the object. It alternates between self-attention between queries and cross-attention of queries to the elements of the backbone feature map, followed by linear layers that regress the final parameters. Specifically, separate linear layers are dedicated to regressing joint angles, hand shape, hand translation, object translation, object orientation, and object articulation. The best model uses a ViT-G [78] backbone with frozen DINOv2 weights [52]. + +AmbiguousHands [55]: The method addresses scale ambiguity, resulting from bounding box cropping in data augmentation and camera intrinsics, by employing positional encoding of these elements to mitigate scale issues. This leads to improved spatial alignment. Subsequently, the approach enhances network visibility by integrating local features through distinct hand and object crops. They follow the general approach of ArcticNet-SF to regress hand/object parameters. + +UVHand: Since ArcticNet-SF only leverages a global feature vector to estimate hand and object parameters, the image features lack local context. To address this, UVHand leverages Swin-L transformer [40] to extract image features. They then further leverage Deformable DETR [86] to encode the multiple-scale feature maps. The encoded feature maps are then aggregated via self- and cross-attention before regressing hand and object parameters. + +DIGIT [16]: The method was introduced to estimate strongly interacting hands in [16]. Since ArcticNet-SF is sensitive when the hands are interacting with objects, DIGIT was extended to the ARCTIC setting. Given an image, it first estimates hand part-wise segmentation masks and object masks. The mask predictions are fused with the image features to perform parameter estimation. + +Implementation details: Table 2 shows that all methods use the default cropping as in ARCTIC to crop around the object, while AmbiguousHands performs three crops (around two hands and the object). DIGIT uses the HRNet-W32 backbone [67] and trains with a batch size of 64 with the same learning rate for all iterations. UVHand takes as input a $384 \times 384$ image cropped around the object and encodes it with the Swin-L transformer [40] backbone. It was trained with a batch size of 48 with a learning rate of 2e-4 for the backbone and 1e-7 for other weights. Due to computational cost, they train 50 epochs for the decoder and 36 for the backbone. JointTransformer uses ViT-G [78] backbone with frozen DINOv2 [52] weights to train with a batch size of 64. It performs a + +Table 3: Method performance in AssemblyHands. We compare AssemblyHands method performance on egocentric test data. We show the final MPJPE on the test set as the metrics (lower better). We also provide detailed evaluations, regarding the varying distances of hand position from the image center and different verb action categories. The hand distance is computed by the distance from the image center to the hand center position per image, and averaged over the lower two views of the headset. Verb classes of "attempt to X" are merged to "X" for simplicity. The higher and lower three verbs are color-coded in red and blue, respectively. + +
MethodScoreHand distance (px)Verb class
-200200-250250-clapinspectpasspick uppositionposition screw onpull
Base20.6920.3121.9724.8519.8822.8922.4821.8522.6521.6218.74
JHands12.2112.3511.9813.7210.6516.2712.8613.6714.5812.811.06
PICO-AI12.4612.5111.6212.9512.9815.311.3713.213.1811.3915.13
FRDC16.4816.3915.8918.6915.2421.3317.8618.2619.2118.0312.83
Phi-AI17.2617.2415.8119.5119.8620.9317.9119.0119.719.3517.3
Verb class (continue)
pushput downremoveremove screw fromrotatescrewtilt downtilt upunscrewnone
Base19.2920.2619.9916.4722.7122.9513.1215.1120.7819.82
JHands13.9613.7213.1111.7212.2614.119.619.9212.2510.99
PICO-AI12.2913.4112.839.9913.713.7210.4411.5612.8711.71
FRDC19.5217.8718.5514.4716.4418.8114.0313.3716.4115.01
Phi-AI19.1218.2918.1813.9518.3519.7813.1315.3617.2915.8
+ +linear warmup from 1e-7 to 1e-4 in the first $5\%$ steps and uses cosine decay from 1e-4 to 1e-7 for the rest of the steps. + +# 5 Results and analysis + +# 5.1 Results + +Here, we benchmark results of valid submissions for state-of-the-art comparison in AssemblyHands and ARCTIC and other more recent baselines. In particular, for AssemblyHands, we report egocentric hand pose estimation results. For ARCTIC, we report results for the allocentric and egocentric test sets. + +AssemblyHands benchmark: Table 3 shows the final test scores on the AssemblyHands dataset. The methods in the table exceed the baseline (Base) with a test score of 20.69 MPJPE. Notably, the methods JHands and PICO-AI achieve a nearly $40\%$ reduction over the baseline. The methods FRDC and Phi-AI improve the test score by $20.3\%$ and $16.5\%$ against the baseline, respectively. + +ARCTIC benchmark: Table 4 presents the comparative performance of methods in the ARCTIC dataset, where ArcticNet-SF serves as the initial benchmark. The majority surpass ArcticNet-SF in both allocentric and egocentric views, except for UVHand, which underperforms due to incomplete training. In the egocentric view, AmbiguousHands excels in creating smooth, consistent mesh motions (refer to MDev and $\mathrm{ACC}_{h / o}$ metrics). Notably, JointTransformer stands + +Table 4: Method performance in ARCTIC. We compare performance in both allocentric (top half) and egocentric (bottom half) views. We evaluate using metrics for contact and relative position (measuring hand-object contact and prediction of relative root position), motion (assessing temporally-consistent contact and smoothness), and hand and object metrics (indicating root-relative reconstruction error). We use the CDev score as the main metric for this competition. We denote left and right hands as $l$ and $r$ , and the object as $o$ . + +
Contact and Relative PositionsMotionHandObject
MethodCDev [mm]↓ MRRPEa/ro [mm]↓ MDev [mm]↓ ACCh/α [m/s2]↓ MPJPE [mm]↓ AAE [°] ↓ Success Rate [%] ↑
AllocentricArcticNet-SF41.5652.39/37.4710.405.72/7.5721.455.3771.39
DIGIT34.9244.19/35.438.374.86/6.6317.925.2476.52
UVHand64.1584.68/70.3114.127.05/12.0440.9912.3631.47
AmbiguousHands33.2545.78/34.5610.126.37/6.4018.024.6481.94
JointTransformer27.9736.17/28.188.936.08/5.7917.123.9589.79
Ego-centricArcticNet-SF44.7128.31/36.1611.805.03/9.1519.186.3953.89
DIGIT41.3125.49/32.619.484.01/8.3216.746.6053.33
UVHand40.4340.93/36.889.965.32/8.3324.537.3257.28
AmbiguousHands35.9323.07/27.539.513.95/6.7616.264.8668.36
JointTransformer32.5626.07/26.2211.345.52/8.6816.334.4474.07
+ +![](images/f70b708de24ac2a015506912d092aab7fbd6af0ae3f5fb3de06d9c9f40476a16.jpg) +Fig. 2: Qualitative results per action in AssemblyHands. We show Base results with "verb (noun)" actions. The left three figures are lower error situations while the right four ones are failure cases. The red boxes denote the area where the action occurs. + +out by significantly lowering CDev errors by $32.7\%$ in allocentric and $27.2\%$ in egocentric settings compared to the baseline. + +# 5.2 AssemblyHands analysis + +We provide analysis, regarding action-wise evaluation, distortion effect in training, and the effect of multi-view fusion. See SupMat for additional results. + +Action-wise evaluation: To analyze errors related to hand-object occlusions and interactions, we show pose evaluation according to fine verb action classes in Table 3. We use the verb classes annotated by Assembly101 [61], spanning every few seconds in a video. Figure 2 shows qualitative results of representative verb classes with the top and bottom error cases. + +We observe that the performance varies among different verb actions. The verbs "tilt down/up" and "remove screw from" exhibit lower errors among the submitted methods, because hands are less occluded and their movement is relatively stable. The "tilt" action holds a small part of the toy and turns it around alternately, leading to less overlap between the hand and the object (lib). The "remove screw from" action takes a screw out from the toy vehicle by their hand where observed hand poses do not change drastically. + +![](images/bfcce42240a3fc25f8f9024cb50c2ba7df097d8e93de6e8993ef4e333b395161.jpg) +Fig. 3: Effect of distortion in AssemblyHands. The officially released images in the dataset have highly stretched areas near the edges (original crop). The method JHands uses a perspective crop with a virtual camera to correct this distortion. + +Table 5: Multi-view fusion in AssemblyHands. We use the Base result to show performance before and after fusion. Missing instances per view are denoted as "Miss(\%)". + +
ViewMPJPEMiss(%)
cam137.9770.8
cam225.7188.3
cam322.190.92
cam422.290.74
cam3+421.520.08
all four20.690
+ +Higher error classes, such as "inspect", "screw", "rotate", and "position", contain heavy occlusions, fast hand motion, complex two hands and object interactions. The "inspect" action brings the toy close to the human eyes where the toy occupies a large portion of the image causing heavy object occlusions. The "screw" action involves intricate fingertip movements to rotate the screwdriver quickly. The "rotate" and "position" actions are performed so that the two hands and the object interact in close contact, which complicates the estimation. We observe that the top two methods JHands and PICO-AI significantly correct the results of these higher error actions compared to the other submitted methods. + +Bias of hand position in an image: Hands near image edges are highly distorted due to the fish-eye cameras. Directly using these noisy images in training will degrade performance [84]; thus, some methods create new crops with less distortion, select training instances, or adaptively fuse predictions during the inference. Specifically, Figure 3 shows that the method JHands reformulates the perspective during cropping and creates less-distorted (perspective) crops. + +To study this effect in the final performance, we split the evaluation instances into classes with different 2D distances between the hand center and the image center in Table 3. Higher distances (250- pixels) indicate closer hand crops to image edges. The method, Base, without any training instance selection and distortion correction, has higher error as the crops approach image edges (20.31 $\rightarrow$ 24.85). In contrast, the newly proposed methods are more robust and have a lower error, particularly in the 200-250 range. We observe that the ranges 200-250 and 250- occupy $10\%$ and $5\%$ of the test images, respectively, thus the improvement in the 200-250 range helps the lowering of the overall score. + +Effect of multi-view fusion: The multi-view egocentric camera setup is unique to the dataset. We show the statistics and performance of multi-view fusion in Table 5. Note that Table 3 shows the final results after multi-view fusion. + +We found that samples captured from the lower cameras (cam3 and cam4; see Table 1 for the layout) are numerous (fewer missing samples) and their errors are lower as they are faced toward the area occurring hand interactions. + +![](images/de923b27e131452ecd823d6759203356406cb7483ac08fa0d6cf68d936c8f67e.jpg) +(a) Egocentric-allocentric comparison +Fig. 4: Performance comparison: Egocentric vs Allocentric. (a) Comparative difficulty ratio of egocentric to allocentric views. (b) Egocentric view performance by method across objects. (c) Allocentric view performance by method across objects. + +![](images/a3f0d36d71f802fa0824cba0f0261952046b8a21b15aec361ed4d8a11f4db8c6.jpg) +(b) Egocentric view + +![](images/7d4745864fdcfb2f47fee1c2fa7d7b487c45155f8cda94e7f978014bc72387a4.jpg) +(c) Allocentric view + +Conversely, the samples from cam1 and cam2 are fewer and unbalanced as their cameras often fail to capture hands due to the camera layout. For instance, cam1 (top-left) tends to capture more hand region than cam2 as the participants are mostly right-handed and bring up the object with the right hand, which can be better observed from cam1. Given this uneven sample distribution, the proposed adaptive view selection methods in either training or testing are essential to perform effective multi-view fusion, and outperform the Base's test-time average using all views all the time (see Multi-view fusion in Section 4.1). + +We further study the performance gain before and after multi-view fusion using Base's results. While per-view performance achieves 22.19 and 22.29 in cam3 and cam4, respectively, their fused results with simple average reduce the error to 21.52. Merging all four views has shown to be more effective than two-view fusion (20.69 vs. 21.52), indicating a $6.5\%$ reduction compared to the single camera setup (cam3). This suggests predictions from the top views (cam1 and cam2) are informative in averaging even when they are prone to be erroneous. + +# 5.3 ARCTIC analysis + +Egocentric-allocentric comparison: Figure 4a compares the performance between egocentric and allocentric views. In particular, we compute a ratio between the metric values of the egocentric and allocentric view to measure the extent of difficulty for the egocentric view compared to the allocentric view. Since success rate is a metric whose value is positively correlated to performance, we take its reciprocal ratio. We average the ratios across methods and actions. + +We observe that hand pose-related metrics such as MPJPE and $\mathrm{ACC}_h$ are less than 1.0 on average (see blue color cells), meaning the the egocentric view is easier than allocentric view. This is because most allocentric cameras in ARCTIC + +![](images/e6b68b50c2cc66b47deb604456e31050bfb31943521ba650e2124c8b0a40ddcd.jpg) +Fig. 5: Egocentric reconstruction by top method in ARCTIC. In the egocentric view, object reconstruction struggle when the object is partially observed on the image boundaries, as well as when heavy hand/arm occlusion occurs. + +are meters away from the subject while the egocentric camera is often close-up, offering higher hand visibility. Relative translation metrics between hand and object such as $\mathrm{MRRPE}_{rl}$ are also easier in the egocentric view because estimating translation is more difficult from further cameras. + +Object reconstruction performance faces unique challenges, as highlighted by the red cells. In the egocentric view, it is notably more difficult to reconstruct accurate object surfaces, articulation (AAE), and hand-object contact (CDev). This increased difficulty arises because objects are often positioned at the image edges and obscured by human arms. Additionally, object poses exhibit greater diversity in the egocentric view due to varying camera angles and occlusions. While a static camera maintains consistent camera extrinsics across a sequence, an egocentric camera's extrinsics change with each frame, resulting in higher diversity in camera-view object 6D poses. This diversity complicates object pose estimation in egocentric views. Figure 5 illustrates these challenges using the best-performing method, JointTransformer. Despite achieving reasonable hand poses, object poses are significantly impacted by occlusions from hands and arms, and the egocentric view undergoes substantial changes throughout a sequence. + +Object-wise evaluation: Figure 4b and 4c break down performances on different objects in the egocentric and allocentric view test sets. The best method in both cases is JointTransformer. The hardest object to reconstruct with good contact consistency (see CDev) is the microwave in both settings due to global rotation sensitivity (see Fig. 5), though this can be mitigated by a keypoint-based approach [24]. Estimating objects in the egocentric view is more difficult than in the allocentric view, which is indicated by higher errors for all methods. + +Action-wise evaluation: Figure 6 compares performance of different methods in "grab" and "use" actions. In ARCTIC, there are sequences to interact with the object with two types of actions by either not articulating the object, or allowing object articulation. Interestingly, the "grab" motion is more challenging in egocentric and allocentric views. We hypothesize that this is because there are more diverse object poses for "grab" motions since during object articulation, the participants often focus on articulation instead of object manipulation. + +Effect of model size: Figure 7 illustrates the impact of model size on handobject contact performance, measured by CDev, for reconstruction on the allocentric validation set. Most methods utilize ResNet50, with JointTransformer being the top performer. As the trainable parameters in the backbone increase, + +![](images/55f0d09062b13f1d742c0f0e963fe0c3739330130d02190cec3aca6a252904bd.jpg) +Fig.6: Hand-object contact quality for reconstructed results per action. We evaluate the contact quality of the 3D reconstruction results from all methods for each action (i.e., grab or use), using Contact Deviation (CDev) in mm as the metric, where lower values indicate better quality. + +![](images/607b42b7cc78a95cfcacfeb6cfc77bece5b5b1d9e1c697ac7d6c6cf3c6d29189.jpg) +Fig. 7: Contact deviation vs. model size. We assess the contact quality of the reconstruction results, varying by the number of parameters in each model. Contact quality is measured using Contact Deviation (CDev) in mm, with lower values indicating superior results. + +JointTransformer consistently reduces the CDev error. Note that the x-axis is in log-scale. JointTransformer achieved a CDev error of 30.5mm with ViT-L and 29.0mm with ViT-G, which has ten times more parameters than ViT-L. Interestingly, JointTransformer uses frozen weights in the large-scale ViT-L and ViT-G backbones, yet achieves the best results. This suggests a potential direction for leveraging large-scale foundational backbones for hand-object reconstruction. + +# 6 Conclusion + +In this paper, we introduce the HANDS23 challenge and provide analysis based on the results of the top submitted methods and more recent baselines on the leaderboards. We organize and compare the submissions and their implementation details based on the learning methods, architecture, pre- and post-processing techniques, and training configurations. We thoroughly analyze various aspects, such as hand-object occlusions, action and object-wise evaluation, distortion correction, multi-view fusion, egocentric-allocentric comparison, and performance gain of large transformer models. + +Future directions: There are several future directions that the community can take. For example, one can explore more efficient training using multi-view egocentric cameras, leveraging 3D foundation priors [38,54] to regularize template-free hand-object reconstruction [15], estimating hand-object poses with more expressive representations (e.g., heatmap-based approaches [24]), incorporating motion and temporal modeling [18], featuring more diverse egocentric interaction scenarios, recognizing actions through captured hand poses [62], learning robotic grasping from reconstructed hand-object pose sequences, and so forth. + +# Acknowledgements + +This work is supported by the Ministry of Science and ICT, Korea, under the ITRC program (IITP-2024-2020-0-01789). supervised by the IITP. UTokyo is supported by JST ACT-X Grant Number JPMJAX2007, JSPS KAKENHI Grant Number JP22KJ0999, JST Adopting Sustainable Partnerships for Innovative Research Ecosystem (ASPIRE) Grant Number JPMJAP2303. NUS is supported by the Ministry of Education, Singapore, under its MOE Academic Research Fund Tier 2 (STEM RIE2025 MOE-T2EP20220-0015). + +# References + +1. Abou Zeid, K.: JointTransformer: Winner of the HANDS'2023 ARCTIC Challenge @ ICCV. https://github.com/kabouzeid/JointTransformer (2023) 7, 8 +2. Armagan, A., Garcia-Hernando, G., Baek, S., Hampali, S., Rad, M., Zhang, Z., Xie, S., Chen, M., Zhang, B., Xiong, F., et al.: Measuring generalisation to unseen viewpoints, articulations, shapes and objects for 3d hand pose estimation under hand-object interaction. In: ECCV. pp. 85-101. Springer (2020) 3 +3. Boukhayma, A., de Bem, R., Torr, P.H.S.: 3D hand shape and pose from images in the wild. In: Computer Vision and Pattern Recognition (CVPR). pp. 10843-10852 (2019) 3 +4. Cai, Y., Ge, L., Cai, J., Yuan, J.: Weakly-supervised 3D hand pose estimation from monocular RGB images. In: European Conference on Computer Vision (ECCV). pp. 678-694 (2018) 3 +5. Cao, Z., Radosavovic, I., Kanazawa, A., Malik, J.: Reconstructing hand-object interactions in the wild. In: International Conference on Computer Vision (ICCV). pp. 12417-12426 (2021) 3 +6. Chao, Y.W., Yang, W., Xiang, Y., Molchanov, P., Handa, A., Tremblay, J., Narang, Y.S., Van Wyk, K., Iqbal, U., Birchfield, S., Kautz, J., Fox, D.: DexYCB: A benchmark for capturing hand grasping of objects. In: Computer Vision and Pattern Recognition (CVPR). pp. 9044-9053 (2021) 2, 3, 5 +7. Chatterjee, D., Sener, F., Ma, S., Yao, A.: Opening the vocabulary of egocentric actions. Conference on Neural Information Processing Systems (NeurIPS) 36 (2024) 2 +8. Chen, X., Liu, Y., Dong, Y., Zhang, X., Ma, C., Xiong, Y., Zhang, Y., Guo, X.: MobRecon: mobile-friendly hand mesh reconstruction from monocular image. In: Computer Vision and Pattern Recognition (CVPR). pp. 20512-20522 (2022) 5 +9. Chen, X., Wang, B., Shum, H.Y.: Hand avatar: Free-pose hand animation and rendering from monocular video. In: Computer Vision and Pattern Recognition (CVPR) (2023) 3 +0. Chen, Z., Chen, S., Schmid, C., Laptev, I.: gSDF: Geometry-driven signed distance functions for 3d hand-object reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 12890-12900 (2023) 3 +1. Christen, S., Kocabas, M., Aksan, E., Hwangbo, J., Song, J., Hilliges, O.: D-Grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In: Computer Vision and Pattern Recognition (CVPR). pp. 20545–20554 (2022) 2 +2. Corona, E., Pumarola, A., Alenyà, G., Moreno-Noguer, F., Rogez, G.: GanHand: Predicting human grasp affordances in multi-object scenes. In: Computer Vision and Pattern Recognition (CVPR). pp. 5030-5040 (2020) 3 + +13. Duran, E., Kocabas, M., Choutas, V., Fan, Z., Black, M.J.: HMP: Hand motion priors for pose and shape estimation from video. In: Winter Conference on Applications of Computer Vision (WACV) (2024) 3 +14. Erol, A., Bebis, G., Nicoescu, M., Boyle, R.D., Twombly, X.: Vision-based hand pose estimation: A review. Computer Vision and Image Understanding (CVIU) 108(1-2), 52-73 (2007) 3 +15. Fan, Z., Parelli, M., Kadoglou, M.E., Kocabas, M., Chen, X., Black, M.J., Hilliges, O.: HOLD: Category-agnostic 3d reconstruction of interacting hands and objects from video. In: Computer Vision and Pattern Recognition (CVPR) (2024) 3, 14 +16. Fan, Z., Spurr, A., Kocabas, M., Tang, S., Black, M.J., Hilliges, O.: Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In: International Conference on 3D Vision (3DV). pp. 1-10 (2021) 2, 3, 7, 8 +17. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 4, 5, 7, 8 +18. Fu, Q., Liu, X., Xu, R., Niebles, J.C., Kitani, K.M.: Deformer: Dynamic fusion transformer for robust hand pose estimation. In: International Conference on Computer Vision (ICCV). pp. 23600-23611 (October 2023) 3, 14 +19. Garcia-Hernando, G., Yuan, S., Baek, S., Kim, T.K.: First-person hand action benchmark with rgb-d videos and 3d hand pose annotations. In: Computer Vision and Pattern Recognition (CVPR) (2018) 2 +20. Ge, L., Ren, Z., Li, Y., Xue, Z., Wang, Y., Cai, J., Yuan, J.: 3D hand shape and pose estimation from a single rgb image. In: Computer Vision and Pattern Recognition (CVPR). pp. 10833-10842 (2019) 3 +21. Grady, P., Tang, C., Twigg, C.D., Vo, M., Brahmbhatt, S., Kemp, C.C.: ContactOpt: Optimizing contact to improve grasps. In: Computer Vision and Pattern Recognition (CVPR). pp. 1471-1481 (2021) 3 +22. Guo, Z., Zhou, W., Wang, M., Li, L., Li, H.: HandNeRF: Neural radiance fields for animatable interacting hands. In: Computer Vision and Pattern Recognition (CVPR). pp. 21078-21087 (2023) 3 +23. Hampali, S., Rad, M., Oberweger, M., Lepetit, V.: HOnnotate: A method for 3D annotation of hand and object poses. In: Computer Vision and Pattern Recognition (CVPR). pp. 3193-3203 (2020) 2 +24. Hampali, S., Sarkar, S.D., Rad, M., Lepetit, V.: Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3D pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 11090-11100 (2022) 13, 14 +25. Han, S., Wu, P.C., Zhang, Y., Liu, B., Zhang, L., Wang, Z., Si, W., Zhang, P., Cai, Y., Hodan, T., Cabezas, R., Tran, L., Akbay, M., Yu, T.H., Keskin, C., Wang, R.: Umetrack: Unified multi-view end-to-end hand tracking for VR. In: International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH). pp. 50:1-50:9. ACM (2022) 2 +26. Hasson, Y., Tekin, B., Bogo, F., Laptev, I., Pollefeys, M., Schmid, C.: Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 568-577 (2020) 3 +27. Hasson, Y., Varol, G., Schmid, C., Laptev, I.: Towards unconstrained joint handobject reconstruction from rgb videos. In: International Conference on 3D Vision (3DV). pp. 659-668. IEEE (2021) 3 + +28. Hasson, Y., Varol, G., Tzionas, D., Kalevatykh, I., Black, M.J., Laptev, I., Schmid, C.: Learning joint reconstruction of hands and manipulated objects. In: Computer Vision and Pattern Recognition (CVPR). pp. 11807-11816 (2019) 3 +29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) 5, 6 +30. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: Computer Vision and Pattern Recognition (CVPR). pp. 15979-15988 (2022) 5 +31. Huang, D., Ji, X., He, X., Sun, J., He, T., Shuai, Q., Ouyang, W., Zhou, X.: Reconstructing hand-held objects from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) 3 +32. Iqbal, U., Molchanov, P., Gall, T.B.J., Kautz, J.: Hand pose estimation via latent 2.5D heatmap regression. In: European Conference on Computer Vision (ECCV). pp. 118-134 (2018) 3, 5 +33. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2O: Two hands manipulating objects for first person interaction recognition. In: International Conference on Computer Vision (ICCV). pp. 10138-10148 (2021) 3 +34. Lee, J., Sung, M., Choi, H., Kim, T.K.: Im2hands: Learning attentive implicit representation of interacting two-hand shapes. In: Computer Vision and Pattern Recognition (CVPR). pp. 21169-21178 (2023) 3 +35. Li, L., Tian, L., Zhang, X., Wang, Q., Zhang, B., Bo, L., Liu, M., Chen, C.: Renderih: A large-scale synthetic dataset for 3d interacting hand pose estimation. In: International Conference on Computer Vision (ICCV). pp. 20395-20405 (2023) 3 +36. Li, M., An, L., Zhang, H., Wu, L., Chen, F., Yu, T., Liu, Y.: Interacting attention graph for single image two-hand reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 2761-2770 (2022) 3 +37. Liu, R., Ohkawa, T., Zhang, M., Sato, Y.: Single-to-dual-view adaptation for egocentric 3d hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 677-686 (2024) 3 +38. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023) 14 +39. Liu, S., Jiang, H., Xu, J., Liu, S., Wang, X.: Semi-supervised 3D hand-object poses estimation with interactions in time. In: Computer Vision and Pattern Recognition (CVPR). pp. 14687-14697 (2021) 3 +40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: International Conference on Computer Vision (ICCV). pp. 10012-10022 (2021) 8 +41. Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Computer Vision and Pattern Recognition (CVPR). pp. 11966-11976 (2022) 6 +42. Meng, H., Jin, S., Liu, W., Qian, C., Lin, M., Ouyang, W., Luo, P.: 3D interacting hand pose estimation by hand de-occlusion and removal. In: European Conference on Computer Vision (ECCV). pp. 380–397. Springer (2022) 3 +43. Moon, G.: Bringing inputs to shared domains for 3d interacting hands recovery in the wild. In: Computer Vision and Pattern Recognition (CVPR). pp. 17028-17037 (2023) 3 +44. Moon, G., Saito, S., Xu, W., Joshi, R., Buffalini, J., Bellan, H., Rosen, N., Richardson, J., Mize, M., De Bree, P., et al.: A dataset of relighted 3d interacting hands. Conference on Neural Information Processing Systems (NeurIPS) 36 (2024) 3 + +45. Moon, G., Yu, S., Wen, H., Shiratori, T., Lee, K.M.: Interhand2.6m: A dataset and baseline for 3d interacting hand pose estimation from a single RGB image. In: European Conference on Computer Vision (ECCV). vol. 12365, pp. 548-564 (2020) 3, 5, 6 +46. Mueller, F., Bernard, F., Sotnychenko, O., Mehta, D., Sridhar, S., Casas, D., Theobalt, C.: GANerated hands for real-time 3D hand tracking from monocular RGB. In: Computer Vision and Pattern Recognition (CVPR). pp. 49-59 (2018) 3 +47. Ohkawa, T.: AssemblyHands Toolkit. https://github.com/facebookresearch/assemblyhands-toolkit (2023) 5 +48. Ohkawa, T., Furuta, R., Sato, Y.: Efficient annotation and learning for 3d hand pose estimation: A survey. International Journal of Computer Vision (IJCV) 131, 3193-3206 (2023) 3 +49. Ohkawa, T., He, K., Sener, F., Hodan, T., Tran, L., Keskin, C.: AssemblyHands: towards egocentric activity understanding via 3d hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 12999-13008 (2023) 2, 3, 4 +50. Ohkawa, T., Li, Y.J., Fu, Q., Furuta, R., Kitani, K.M., Sato, Y.: Domain adaptive hand keypoint and pixel localization in the wild. In: European Conference on Computer Vision (ECCV). pp. 68—87 (2022) 2 +51. Ohkawa, T., Yagi, T., Hashimoto, A., Ushiku, Y., Sato, Y.: Foreground-aware stylization and consensus pseudo-labeling for domain adaptation of first-person hand segmentation. IEEE Access 9, 94644-94655 (2021) 2 +52. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) 8 +53. Park, J., Oh, Y., Moon, G., Choi, H., Lee, K.M.: HandOccNet: Occlusion-robust 3d hand mesh estimation network. In: Computer Vision and Pattern Recognition (CVPR). pp. 1496–1505 (2022) 6 +54. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022) 14 +55. Prakash, A., Tu, R., Chang, M., Gupta, S.: 3d hand pose estimation in everyday egocentric images. In: European Conference on Computer Vision (ECCV) (2024) 8 +56. Radosavovic, I., Kosaraju, R.P., Girshick, R.B., He, K., Dollár, P.: Designing network design spaces. In: Computer Vision and Pattern Recognition (CVPR). pp. 10425-10433 (2020) 5, 6 +57. Rehg, J.M., Kanade, T.: Visual tracking of high DOF articulated structures: An application to human hand tracking. In: European Conference on Computer Vision (ECCV). vol. 801, pp. 35-46 (1994) 3 +58. Remelli, E., Han, S., Honari, S., Fua, P., Wang, R.: Lightweight multi-view 3d pose estimation through camera-disentangled representation. In: Computer Vision and Pattern Recognition (CVPR). pp. 6039-6048 (2020) 6, 7 +59. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG) 36(6), 245:1-245:17 (2017) 7 +60. Ryali, C., Hu, Y., Bolya, D., Wei, C., Fan, H., Huang, P., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: International Conference on Machine Learning (ICML). vol. 202, pp. 29441-29454 (2023) 5, 6 + +61. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Computer Vision and Pattern Recognition (CVPR). pp. 21064-21074 (2022) 10 +62. Shamil, M.S., Chatterjee, D., Sener, F., Ma, S., Yao, A.: On the utility of 3d hand poses for action recognition. In: European Conference on Computer Vision (ECCV) (2024) 14 +63. Simon, T., Joo, H., Matthews, I., Sheikh, Y.: Hand keypoint detection in single images using multiview bootstrapping. In: Computer Vision and Pattern Recognition (CVPR). pp. 4645-4653 (2017) 3 +64. Spurr, A., Dahiya, A., Wang, X., Zhang, X., Hilliges, O.: Self-supervised 3D hand pose estimation from monocular RGB via contrastive learning. In: International Conference on Computer Vision (ICCV). pp. 11210-11219 (2021) 3 +65. Spurr, A., Iqbal, U., Molchanov, P., Hilliges, O., Kautz, J.: Weakly supervised 3D hand pose estimation via biomechanical constraints. In: European Conference on Computer Vision (ECCV). vol. 12362, pp. 211-228 (2020) 2, 3 +66. Spurr, A., Song, J., Park, S., Hilliges, O.: Cross-modal deep variational hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 89-98 (2018) 3 +67. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Computer Vision and Pattern Recognition (CVPR) (2019) 8 +68. Swamy, A., Leroy, V., Weinzaepfel, P., Baradel, F., Galaoui, S., Brégier, R., Armando, M., Franco, J.S., Rogez, G.: SHOWMe: Benchmarking object-agnostic hand-object 3d reconstruction. In: International Conference on Computer Vision (ICCV). pp. 1935–1944 (2023) 3 +69. Tekin, B., Bogo, F., Pollefeys, M.: $\mathrm{H} + \mathrm{O}$ : Unified egocentric recognition of 3D hand-object poses and interactions. In: Computer Vision and Pattern Recognition (CVPR). pp. 4511-4520 (2019) 3 +70. Tse, T.H.E., Kim, K.I., Leonardis, A., Chang, H.J.: Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In: Computer Vision and Pattern Recognition (CVPR). pp. 1664-1674 (2022) 3 +71. Tzionas, D., Gall, J.: A comparison of directional distances for hand pose estimation. In: German Conference on Pattern Recognition (GCPR). vol. 8142, pp. 131-141 (2013) 3 +72. Wen, Y., Pan, H., Ohkawa, T., Yang, L., Pan, J., Sato, Y., Komura, T., Wang, W.: Generative hierarchical temporal transformer for hand action recognition and motion prediction. CoRR abs/2311.17366 (2023) 2 +73. Yang, L., Chen, S., Yao, A.: Semihand: Semi-supervised hand pose estimation with consistency. In: International Conference on Computer Vision (ICCV). pp. 11364-11373 (2021) 2 +74. Yang, L., Zhan, X., Li, K., Xu, W., Li, J., Lu, C.: CPF: Learning a contact potential field to model the hand-object interaction. In: International Conference on Computer Vision (ICCV) (2021) 3 +75. Ye, Y., Gupta, A., Tulsiani, S.: What's in your hands? 3d reconstruction of generic objects in hands. In: Computer Vision and Pattern Recognition (CVPR) (2022) 3 +76. Ye, Y., Hebbar, P., Gupta, A., Tulsiani, S.: Diffusion-guided reconstruction of everyday hand-object interaction clips. In: International Conference on Computer Vision (ICCV) (2023) 3 + +77. Yuan, S., Garcia-Hernando, G., Stenger, B., Moon, G., Chang, J.Y., Lee, K.M., Molchanov, P., Kautz, J., Honari, S., Ge, L., Yuan, J., Chen, X., Wang, G., Yang, F., Akiyama, K., Wu, Y., Wan, Q., Madadi, M., Escalera, S., Li, S., Lee, D., Oikonomidis, I., Argyros, A.A., Kim, T.: Depth-based 3D hand pose estimation: From current achievements to future goals. In: Computer Vision and Pattern Recognition (CVPR). pp. 2636-2645 (2018) 3 +78. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Computer Vision and Pattern Recognition (CVPR). pp. 12104-12113 (2022) 8 +79. Zhang, H., Christen, S., Fan, Z., Hilliges, O., Song, J.: GraspXL: Generating grasping motions for diverse objects at scale. European Conference on Computer Vision (ECCV) (2024) 2 +80. Zhang, H., Christen, S., Fan, Z., Zheng, L., Hwangbo, J., Song, J., Hilliges, O.: ArtiGrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. In: International Conference on 3D Vision (3DV) (2024) 2 +81. Zhang, X., Li, Q., Mo, H., Zhang, W., Zheng, W.: End-to-end hand mesh recovery from a monocular RGB image. In: International Conference on Computer Vision (ICCV). pp. 2354-2364 (2019) 3 +82. Zhou, Y., Habermann, M., Xu, W., Habibie, I., Theobalt, C., Xu, F.: Monocular real-time hand shape and motion capture using multi-modal data. In: Computer Vision and Pattern Recognition (CVPR). pp. 5345-5354 (2020) 3, 5, 6 +83. Zhou, Z.: SimpleHand: Winner of the HANDS'2023 AssemblyHands Challenge @ ICCV. https://github.com/patienceFromZhou/simpleHand (2024) 5 +84. Zhou, Z., Lv, Z., Zhou, S., Zou, M., Wu, T., Yu, M., Tang, Y., Liang, J.: 1st place solution of egocentric 3d hand pose estimation challenge 2023 technical report: A concise pipeline for egocentric hand pose reconstruction. CoRR abs/2310.04769 (2023) 11 +85. Zhou, Z., Zhou, S., Lv, Z., Zou, M., Tang, Y., Liang, J.: A simple baseline for efficient hand mesh reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 1367-1376 (2024) 5, 6 +86. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (ICLR). OpenReview.net (2021), https://openreview.net/forum?id=gZ9hCDWe6ke 8 +87. Ziani, A., Fan, Z., Kocabas, M., Christen, S., Hilliges, O.: TempCLR: Reconstructing hands via time-coherent contrastive learning. In: International Conference on 3D Vision (3DV). pp. 627-636 (2022) 3 +88. Zimmermann, C., Brox, T.: Learning to estimate 3D hand pose from single RGB images. In: International Conference on Computer Vision (ICCV). pp. 4913-4921 (2017) 3, 5 \ No newline at end of file diff --git a/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/images.zip b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5c6d5aebc97a757fc81ca9ef821ecaee2ad6138c --- /dev/null +++ b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5396ff2c23e068b06b979831302b1fc9e0b755678ade078e297b1a3bba581229 +size 458133 diff --git a/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/layout.json b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..efe33eb58d15f025cd28b428599d8109ea3cd3b0 --- /dev/null +++ b/2024/Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects/layout.json @@ -0,0 +1,10043 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "spans": [ + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": "Zicong Fan" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Takehiko Ohkawa" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Linlin Yang" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Nie Lin" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Zhishan Zhou" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Shihao Zhou" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Jiajun Liang" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Zhong Gao" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Xuanyang Zhang" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Xue Zhang" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Fei Li" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Liu Zheng" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Feng Lu" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Karim Abou Zeid" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Bastian Leibe" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Jeongwan On" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Seungryul Baek" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Aditya Prakash" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{11}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Saurabh Gupta" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{11}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Kun He" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Yoichi Sato" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Otmar Hilliges" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Hyung Jin Chang" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "text", + "content": ", Angela Yao" + }, + { + "bbox": [ + 159, + 167, + 457, + 242 + ], + "type": "inline_equation", + "content": "^{14}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 270, + 455, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 270, + 455, + 479 + ], + "spans": [ + { + "bbox": [ + 160, + 270, + 455, + 479 + ], + "type": "text", + "content": "Abstract. We interact with the world with our hands and see it through our own (egocentric) perspective. A holistic 3D understanding of such interactions from egocentric views is important for tasks in robotics, AR/VR, action recognition and motion generation. Accurately reconstructing such interactions in 3D is challenging due to heavy occlusion, viewpoint bias, camera distortion, and motion blur from the head movement. To this end, we designed the HANDS23 challenge based on the AssemblyHands and ARCTIC datasets with carefully designed training and testing splits. Based on the results of the top submitted methods and more recent baselines on the leaderboards, we perform a thorough analysis on 3D hand(-object) reconstruction tasks. Our analysis demonstrates the effectiveness of addressing distortion specific to egocentric cameras, adopting high-capacity transformers to learn complex hand-object interactions, and fusing predictions from different views. Our study further reveals challenging scenarios intractable with state-of-the-art methods, such as fast hand motion, object reconstruction from narrow egocentric views, and close contact between two hands and objects. Our efforts will enrich the community's knowledge foundation and facilitate future hand studies on egocentric hand-object interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 529, + 230, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 230, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 230, + 542 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": "We interact with the world with our hands and see it through our eyes: we wake up and grab our phone to check the time; we use tools when assembling parts of a car; we open the microwave door to heat food, to name a few. An egocentric 3D understanding of our hand interactions with objects will fundamentally impact" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{*}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "Equal contribution, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "ETH Zürich, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent Systems, Tübingen, Germany, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "The University of Tokyo, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "Communication University of China, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "Jiiov, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "Bytedance, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "Fujitsu Research & Development Center Co., Ltd., " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "Beihang University, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "RWTH Aachen University, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "UNIST, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{11}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "UIUC, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "Meta, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "University of Birmingham, " + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{14}" + }, + { + "bbox": [ + 140, + 609, + 482, + 666 + ], + "type": "text", + "content": "National University of Singapore." + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 188, + 114, + 428, + 176 + ], + "blocks": [ + { + "bbox": [ + 188, + 114, + 428, + 176 + ], + "lines": [ + { + "bbox": [ + 188, + 114, + 428, + 176 + ], + "spans": [ + { + "bbox": [ + 188, + 114, + 428, + 176 + ], + "type": "image", + "image_path": "467e6a201dc213513725c5c616426b8a315bc6ff1a58b39b1f4df0e30a92146d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 183, + 482, + 218 + ], + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 218 + ], + "type": "text", + "content": "Fig. 1: Tasks in HANDS23. In AssemblyHands, from its multi-view headset (a), we estimate 3D hand poses from images (b); In ARCTIC, we estimate the poses of two hands and articulated objects from an image (c)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "type": "text", + "content": "areas such as robotics grasping [11,80], augmented and virtual reality [25], action recognition [7,19,72] and motion generation [79,80]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "text", + "content": "However, it is non-trivial to accurately reconstruct 3D hands and or objects due to its high degree of freedoms [65, 73], ambiguous texture [16], and heavy occlusions. These challenges are intensified in an egocentric view [50, 51], particularly with object interactions, due to significant camera distortion, rapid and varied changes in viewpoint caused by head movements and hand-object occlusion. To better understand these challenges, we introduce a public challenge in conjunction with ICCV 2023 (i.e., HANDS23) based on recent egocentric hand datasets, AssemblyHands [49] and ARCTIC [17] (see Figure 1). These two datasets are large-scale, multi-view, and provide monochrome or RGB egocentric videos of the hands dexterously manipulating objects. Accordingly, we host two tasks: 1) egocentric 3D hand pose estimation from a single-view image based on AssemblyHands, and 2) consistent motion reconstruction based on ARCTIC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 480, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 480, + 508 + ], + "type": "text", + "content": "We introduce new HANDS23 methods and recent dataset leaderboard baselines that substantially outperform initial baselines for both tasks, setting new benchmarks for subsequent comparisons on the datasets. The two datasets are significantly larger and include a wider variety of bimanual manipulations compared to earlier datasets [6, 23], enabling a more authentic assessment of real-world interactions. With these benchmarks, we thoroughly analyze factors such as viewpoint, action types, hand position, model size, and object variations to determine their impact on 3D hand(-object) reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 605 + ], + "type": "text", + "content": "Our findings show the success of addressing the distortion of egocentric cameras with explicit perspective cropping or implicit learning for the distortion bias. In addition, recent high-capacity vision transformers can learn complex hand-object interactions. Adaptive fusion techniques for multi-view predictions further boost performance. We also analyze the remaining challenges that are still difficult to handle with the recent methods, e.g., fast hand motion, object reconstruction from narrow and moving views, and intricate interactions and close contact between two hands and objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "To summarize, we contribute state-of-the-art baselines and gather the submitted methods for AssemblyHands and ARCTIC to foster future research on egocentric hand-object interactions. Furthermore, we thoroughly analyze the two benchmarks to provide insights for future directions in egocentric hand pose estimation and consistent motion reconstruction." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 305, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 305, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 305, + 102 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "type": "text", + "content": "3D hand pose estimation: Reconstructing 3D hand poses has a long history [14,48] ever since the important work led by Rehg and Kanade [57]. A large body of research in this area focuses on single-hand reconstruction [3,4,6,9,13, 16,18,20,28,32,37,46,63-66,71,81,87,88]. For example, while a popular OpenPose library features 2D hand keypoints [63], Zimmermann et al. [88] initially extend to estimate 3D hand poses using deep convolutional networks. Ever since the release of InterHand2.6M [45] dataset, the community has increased focus on strongly interacting two hands reconstruction [22,33-36,42-45,49]. For example, Li et al. [35] and Moon et al. [44] use relighting techniques to augment InterHand2.6M with more natural lighting and diverse backgrounds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 353 + ], + "type": "text", + "content": "Hand-object reconstruction: The holistic reconstruction of hands and objects have increased interest in the hand community in recent years [6,12,17,21, 26-28,33,39,69,70,74,82]. Methods in this area mostly assume a given object template and jointly estimate the hand poses and the object rigid poses [12, 21,26-28,39,69,70,74,82] or articulated poses [17]. For example, Cao et al. [5] fits object templates to in-the-wild interaction videos. Liu et al. [39] introduce a semi-supervised learning framework via pseudo-groundtruth from temporal data to improve hand-object reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 425 + ], + "type": "text", + "content": "More recent methods do not assume object templates for hand-object reconstruction [10, 15, 28, 31, 68, 75, 76]. For example, Fan et al. [15] introduced the first category-agnostic method that reconstructs an articulated hand and object jointly from a video. Nonetheless, our challenges and insights on hand-object occlusions and camera distortions are still applicable to these more challenging template-free reconstruction settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 533 + ], + "type": "text", + "content": "Public reports for the previous challenges (HANDS17 [77] and HANDS19 [2]) distilled the insights from individual review papers and practical techniques into comprehensive summaries to enrich the community's knowledge base. These past challenges use benchmarks which include depth-based hand pose estimation from egocentric views. Instead of depth sensors, the HANDS23 benchmarks are based on affordable and widely applicable image sensors, i.e., RGB and monochrome images. This paper further advances the analysis with unique insights, such as multi-view egocentric cameras, object reconstruction in contact, and modeling with recent transformers beyond conventional CNNs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 548, + 332, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 332, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 332, + 563 + ], + "type": "text", + "content": "3 HANDS23 challenge overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 571, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 481, + 608 + ], + "type": "text", + "content": "The workshop contains two hand-related 3D reconstruction challenges in hand-object strongly interacting settings. In this section, we introduce the two challenges and their evaluation criteria." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 266, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 266, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 266, + 635 + ], + "type": "text", + "content": "3.1 Workshop challenges" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "3D hand pose estimation in AssemblyHands: As illustrated in Figure 1, this task focuses on egocentric 3D hand pose estimation from a single-view im" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "age based on AssemblyHands. The dataset provides multi-view captured videos of hand-object interaction while assembling and disassembling toy vehicles. In particular, it provides allocentric (fixed-view) and egocentric recordings and auxiliary cues like action, object, or context information for hand pose estimation. We refer readers to [49] for more dataset details. The training, validation, and testing sets contain 383K, 32K, and 62K monochrome images, respectively, all captured from egocentric cameras. During training, 3D hand keypoint coordinates, hand bounding boxes and camera intrinsic and extrinsic matrices for the four egocentric cameras attached to the headset are provided. The same information is provided during testing, minus the 3D keypoints. Unlike [49], given the availability of multi-view egocentric images, this task lets participants develop multi-view fusion based on the corresponding multi-view images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 381 + ], + "type": "text", + "content": "Consistent motion reconstruction in ARCTIC: Given an RGB image, the goal of this task is to estimate poses of hands and articulated objects to recover the 3D surfaces of the interaction (see Figure 1). We refer readers to [17] for more details. The ARCTIC dataset contains data of hands dexterously manipulating articulated objects and videos from " + }, + { + "bbox": [ + 130, + 261, + 482, + 381 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 130, + 261, + 482, + 381 + ], + "type": "text", + "content": " allocentric views and " + }, + { + "bbox": [ + 130, + 261, + 482, + 381 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 130, + 261, + 482, + 381 + ], + "type": "text", + "content": " egocentric views. The official splits of the ARCTIC dataset are used for training, validation, and testing. There are two sub-tasks: allocentric task and egocentric task. In the former, only allocentric images can be used for training and evaluation. For the latter, all images from the training set can be used for training while only the egocentric view images are used during evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 403, + 254, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 254, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 254, + 414 + ], + "type": "text", + "content": "3.2 Evaluation criteria" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "text", + "content": "AssemblyHands evaluation: We use MPJPE as an evaluation metric in millimeters, comparing the model predictions against the ground-truth in world coordinates. We provide the intrinsic and extrinsic of the egocentric cameras to construct submission results defined in the world coordinates. Assuming that the human hand has a total " + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "inline_equation", + "content": "N_J" + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "text", + "content": " joints, we denote wrist-relative coordinates of the prediction and ground-truth as " + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\hat{J} \\in \\mathbb{R}^{N_J \\times 3}" + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "inline_equation", + "content": "J \\in \\mathbb{R}^{N_J \\times 3}" + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "text", + "content": ", respectively. Given a joint visibility indicator " + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\gamma_i" + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "text", + "content": " per joint " + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "inline_equation", + "content": "J_i" + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "text", + "content": ", we compute the Euclidean distance between predicted and ground-truth joints as " + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\sum_{i=1}^{N_J} \\gamma_i} \\sum_{i=1}^{N_J} \\gamma_i \\left\\| \\hat{J}_i - J_i \\right\\|_2" + }, + { + "bbox": [ + 130, + 428, + 482, + 554 + ], + "type": "text", + "content": ". The visibility indicator offers per-joint binary labels, representing whether the annotated keypoints are visible from the given egocentric view." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "text", + "content": "ARCTIC evaluation: Since the original ARCTIC paper [17] has a heavy focus on the quality of hand-object contact in the reconstructed hand and object meshes, we use Contact Deviation (CDev) introduced in the ARCTIC dataset as the main metric for the competition. In particular, this metric measures the extent to which a hand vertex deviates from the supposed contact object vertex in the prediction. Concretely, suppose that for a given frame, " + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{h}_i,\\mathbf{o}_i)\\}_{i = 1}^C" + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "text", + "content": " pairs of in-contact (" + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "inline_equation", + "content": "< 3mm" + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "text", + "content": " distance) hand-object vertices according to ground-truth, and " + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{(\\hat{\\mathbf{h}}_i,\\hat{\\mathbf{o}}_i)\\}_{i = 1}^C" + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "text", + "content": " are the predictions respectively. The CDev metric is the average distance between " + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{h}}_i" + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{o}}_i" + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "text", + "content": " in millimeters, " + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\frac{1}{C}\\sum_{i = 1}^{C}||\\hat{\\mathbf{h}}_i - \\hat{\\mathbf{o}}_i||" + }, + { + "bbox": [ + 130, + 555, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "For completeness, we report all metrics introduced in ARCTIC. In particular, the task requires the reconstructed meshes to have accurate hand-object contact (CDev) and smooth motion (ACC). Additionally, during articulation or when carrying an object, it is crucial that the vertices of the hand and object which are in contact maintain synchronized movement (MDev). Moreover, we assess hand and object poses, alongside their relative movements, using metrics like MPJPE, AAE, Success Rate, and MRRPE. For detailed information, see [17]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 219, + 208, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 208, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 208, + 232 + ], + "type": "text", + "content": "4 Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 244, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 481, + 281 + ], + "type": "text", + "content": "This section presents the methods in the two challenges and other competing methods on the leaderboards. Four methods outperform the baseline in both AssemblyHands and ARCTIC." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 298, + 288, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 298, + 288, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 298, + 288, + 310 + ], + "type": "text", + "content": "4.1 AssemblyHands methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 402 + ], + "type": "text", + "content": "Participants develop methods that learn the mapping from egocentric images to 3D keypoints. The methods are categorized into: heatmap-based and regression-based approaches. Given the presence of complex hand-object interactions in the egocentric scenes, high-capacity transformer models and attention mechanisms addressing occluded regions have been proposed as the backbone networks. Table 1 summarizes the methods based on the learning, preprocessing, multi-view fusion, and post-processing approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 437 + ], + "type": "text", + "content": "Base: This method uses a heatmap-based framework based on heatmaps [45] with 2.5D representations [32] and a ResNet50 [29] backbone. The implementation can be found in [47]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 533 + ], + "type": "text", + "content": "JHands [85]: This method employs a regression-based approach with simple MLP heads for regressing 2D keypoints, root-relative 3D keypoints, and the global root depth. The regression training is empowered by a recent fast and strong vision transformer, Hiera [60], pre-trained with masked auto-encoder [30]. A multi-level feature fusion that concatenates the features of different layers is adopted for better feature extraction at different scales. The method additionally uses other publicly available datasets for training, namely FreiHAND [88], DexYCB [6], and CompHand [8]. The implementation is available in [83]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 641 + ], + "type": "text", + "content": "PICO-AI: This method proposes a heatmap voting scheme in addition to the 2.5D heatmaps. Due to their sparsity, the conventional heatmaps pose an imbalance problem between positive and negative samples in the loss function. Hence, the proposed voting mechanism aims to spread the loss evenly across the entire heatmaps. Given the initial guess of keypoints obtained from the heatmaps, the method defines a local region centered on the joint position and operates the soft-argmax within the region to obtain refined keypoint coordinates. This restricts the impact of background points, leading to more reliable optimization. The training is facilitated by CNN-based RegNety320 [56]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Phi-AI: While following the heatmap-based approach, this method adapts MinimalHand [82] with the ResNet50 backbone, where 2D heatmaps and 3D location" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 480, + 297 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 1: Method and preprocessing summary in AssemblyHands. We summarize submitted methods in terms of learning methods, architecture, preprocessing, and multi-view fusion techniques. The tuple (views, phase) indicates the number of views used in either train or test time." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 480, + 297 + ], + "type": "table", + "html": "
MethodLearning methodsArchitecturePreprocessingMulti-view fusion (views, phase)
Base2.5D heatmaps [45]ResNet50 [29]-Simple average (4, test)
JHands [85]RegressionHiera [60]Warp perspective, color jitter, random maskAdaptive view selection and average (2, test)
PICO-AI2.5D heatmaps [45] Heatmap votingRegNety320 [56]Scale, rotate, flip, translateAdaptive view selection FTL [58] (2, train)
FRDCRegression 2D heatmapsHandOccNet [53] with ConvNeXt [41]Scale, rotate, color jitterWeighted average (4, test)
Phi-AI2D heatmaps and 3D location maps [82]ResNet50 [29]Scale, rotate, translate, color jitter, gaussian blurWeighted average (4, test)
", + "image_path": "e9ee2538b0bb1b72a08da74aab36af1314922c7962eb13d5b35a4afce06e96ee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "text", + "content": "maps are regressed. Instead of selecting 3D keypoint coordinates from the location maps, the proposed method modifies it by using heatmap values to weight 3D keypoint coordinates, achieving a more robust estimation. Moreover, the method adds a residual structured layer after the original three-tier cascade networks to refine the calculated location maps. The method further applies the ensemble of final keypoint outputs combined with the Base." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 396, + 480, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 480, + 480 + ], + "type": "text", + "content": "FRDC: This method adopts a hybrid approach by combining regression with heatmap for training. HandOccNet [53] is modified to regress 3D keypoint coordinates and integrated with an additional branch of 2D heatmap regression. HandOccNet enriches feature extraction with spatial attention mechanisms for occluded regions, making it robust under hand-object occlusions. The method further utilizes a stronger ConvNeXt [41] backbone and feature fusion from the 2D keypoint regressor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 482, + 480, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 480, + 541 + ], + "type": "text", + "content": "Preprocessing of egocentric images: Compared to conventional static camera setups, egocentric images exhibit unique properties and biases, such as distortion, head camera motion, and different color representations. Thus, it is vital to preprocess egocentric images to alleviate these effects during training. Augmentation techniques are detailed in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 628 + ], + "type": "text", + "content": "The method JHands addresses the distortion issue with a warp perspective operation to make the hands near the edge less stretched. While AssemblyHands provides rectified images converted from fisheye cameras to a pinhole camera model, they often include excessively stretched areas near the edges. To address this, the method calculates a virtual camera and corresponding perspective transformation matrix based on the pixel coordinates of the crop and the camera parameters. The generated crops can be found in the analysis of Figure 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Multi-view fusion: Since AssemblyHands offers multi-view egocentric videos, participants can optionally use the constraint of multi-view geometry and fusion techniques during training or inference." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 305, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 305, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 305, + 102 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 179, + 479, + 283 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Method and preprocessing summary in ARCTIC. We summarize baselines on ARCTIC in terms of input dimensions, image backbones, learning rate scheduling, training epochs, batch size and the cropping used for input. *Method trains 50 epochs for decoder and 36 for backbone. +Learning rate is 1e-7 to 1e-4 with linear warmup for first " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " step, and 1e-4 to 1e-7 with cosine decay for rest." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 179, + 479, + 283 + ], + "lines": [ + { + "bbox": [ + 135, + 179, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 135, + 179, + 479, + 283 + ], + "type": "table", + "html": "
MethodInput sizeBackbonesLearning rate scheduleTraining epochsBatch sizeCropping
ArcticNet-SF [17]224 × 224ResNet501e-5allocentric: 20 \negocentric: 5064object
DIGIT [16]224 × 224HRNet-W321e-5allocentric: 20 \negocentric: 5064object
AmbiguousHands224 × 224ResNet501e-5allocentric: 20 \negocentric: 10032hand \nobject
UVHand384 × 384Swin-L2e-4 (backbone) \n1e-7 (others)allocentric: N/A \negocentric: 50/36*48object
JointTransformer [1]224 × 224ViT-G1e-7/1e-5+allocentric: 20 \negocentric: 10064object
", + "image_path": "3578f4de50aa661167cf6ab06e4fa3e95a49b28c649cf3ecedbcdd12531b4d95.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 305, + 479, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 479, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 479, + 400 + ], + "type": "text", + "content": "While Base uses a simple average of predicted keypoints from all four camera views in the test time, PICO-AI proposes multi-view feature fusion during training using Feature Transform Layers (FTL) [58]. This FTL training requires fusing two out of four views; thus, the method chooses the most suitable views for every frame. In cases with multiple candidates, the Intersection over Union (IoU) is computed between hand boxes from per-view predictions and 2D keypoints from previous 3D predictions. The two views with the highest IoUs are selected for their superior prediction reliability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 496 + ], + "type": "text", + "content": "The methods JHands, FRDC, and Phi-AI apply adaptive fusion in predicted keypoints during testing. The method JHands computes the MPJPE with each other view and selects two results of views with the lowest MPJPE, excluding noisy predictions in the average. If the MPJPE is lower than a threshold, the mean of the two results is calculated as the final result. Otherwise, the result with a lower PA-MPJPE with the predictions in the previous frame is chosen. The methods FRDC and Phi-AI use a weighted average for each view prediction, assigning weights based on each view's validation performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 544 + ], + "type": "text", + "content": "Postprocessing: Several postprocessing techniques, including test-time augmentation, smoothing, and model ensembling are used to enhance inference outcomes. In particular, the method JHands applies an offline smooth (Savitzky-Golay) filter on each video sequence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 562, + 254, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 254, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 254, + 574 + ], + "type": "text", + "content": "4.2 ARCTIC methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 616 + ], + "type": "text", + "content": "Table 2 summarizes the details for each method in terms of the input image dimensions, image backbones, learning rate scheduling, the number of training epochs, batch size, and cropping strategies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Preliminary: All methods below are regression-based and predict two-hand MANO [59] parameters " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\theta, \\beta\\}" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " and articulated object parameters " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": ". In particular, with the MANO pose and shape parameters " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\theta, \\beta" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": ", the MANO model " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " returns a mesh with vertices via " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\theta, \\beta) \\in \\mathbb{R}^{778 \\times 3}" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": ". 3D joints are obtained via" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "content": "a linear regressor. The articulated object model " + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "content": " was introduced in ARCTIC to provide an articulated mesh with vertices via " + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\Omega)\\in \\mathbb{R}^{\\mathrm{V}\\times 3}" + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\Omega \\in \\mathbb{R}^{7}" + }, + { + "bbox": [ + 132, + 115, + 480, + 152 + ], + "type": "text", + "content": " contains the global orientation, global translation, and object articulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 480, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 480, + 204 + ], + "type": "text", + "content": "ArcticNet-SF [17]: Introduced in ARCTIC, it is a single-frame baseline. It first extracts an image feature vector from the input image; then, it regresses hand and object parameters with simple MLPs. The hand and object meshes can then be extracted via " + }, + { + "bbox": [ + 132, + 156, + 480, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot)" + }, + { + "bbox": [ + 132, + 156, + 480, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 156, + 480, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\cdot)" + }, + { + "bbox": [ + 132, + 156, + 480, + 204 + ], + "type": "text", + "content": ". For more details, see [17]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 207, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 327 + ], + "type": "text", + "content": "JointTransformer [1]: JointTransformer enhances ArcticNet-SF by integrating a transformer decoder instead of the MLP regressors for hand and object parameter estimation. The decoder employs learned queries for the angle of each joint, the shape and translation of each hand, and the translation, rotation, and articulation of the object. It alternates between self-attention between queries and cross-attention of queries to the elements of the backbone feature map, followed by linear layers that regress the final parameters. Specifically, separate linear layers are dedicated to regressing joint angles, hand shape, hand translation, object translation, object orientation, and object articulation. The best model uses a ViT-G [78] backbone with frozen DINOv2 weights [52]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 331, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 403 + ], + "type": "text", + "content": "AmbiguousHands [55]: The method addresses scale ambiguity, resulting from bounding box cropping in data augmentation and camera intrinsics, by employing positional encoding of these elements to mitigate scale issues. This leads to improved spatial alignment. Subsequently, the approach enhances network visibility by integrating local features through distinct hand and object crops. They follow the general approach of ArcticNet-SF to regress hand/object parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 406, + 480, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 480, + 478 + ], + "type": "text", + "content": "UVHand: Since ArcticNet-SF only leverages a global feature vector to estimate hand and object parameters, the image features lack local context. To address this, UVHand leverages Swin-L transformer [40] to extract image features. They then further leverage Deformable DETR [86] to encode the multiple-scale feature maps. The encoded feature maps are then aggregated via self- and cross-attention before regressing hand and object parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 482, + 480, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 542 + ], + "type": "text", + "content": "DIGIT [16]: The method was introduced to estimate strongly interacting hands in [16]. Since ArcticNet-SF is sensitive when the hands are interacting with objects, DIGIT was extended to the ARCTIC setting. Given an image, it first estimates hand part-wise segmentation masks and object masks. The mask predictions are fused with the image features to perform parameter estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 545, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 665 + ], + "type": "text", + "content": "Implementation details: Table 2 shows that all methods use the default cropping as in ARCTIC to crop around the object, while AmbiguousHands performs three crops (around two hands and the object). DIGIT uses the HRNet-W32 backbone [67] and trains with a batch size of 64 with the same learning rate for all iterations. UVHand takes as input a " + }, + { + "bbox": [ + 132, + 545, + 480, + 665 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 132, + 545, + 480, + 665 + ], + "type": "text", + "content": " image cropped around the object and encodes it with the Swin-L transformer [40] backbone. It was trained with a batch size of 48 with a learning rate of 2e-4 for the backbone and 1e-7 for other weights. Due to computational cost, they train 50 epochs for the decoder and 36 for the backbone. JointTransformer uses ViT-G [78] backbone with frozen DINOv2 [52] weights to train with a batch size of 64. It performs a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 213, + 479, + 370 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "type": "text", + "content": "Table 3: Method performance in AssemblyHands. We compare AssemblyHands method performance on egocentric test data. We show the final MPJPE on the test set as the metrics (lower better). We also provide detailed evaluations, regarding the varying distances of hand position from the image center and different verb action categories. The hand distance is computed by the distance from the image center to the hand center position per image, and averaged over the lower two views of the headset. Verb classes of \"attempt to X\" are merged to \"X\" for simplicity. The higher and lower three verbs are color-coded in red and blue, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 213, + 479, + 370 + ], + "lines": [ + { + "bbox": [ + 135, + 213, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 135, + 213, + 479, + 370 + ], + "type": "table", + "html": "
MethodScoreHand distance (px)Verb class
-200200-250250-clapinspectpasspick uppositionposition screw onpull
Base20.6920.3121.9724.8519.8822.8922.4821.8522.6521.6218.74
JHands12.2112.3511.9813.7210.6516.2712.8613.6714.5812.811.06
PICO-AI12.4612.5111.6212.9512.9815.311.3713.213.1811.3915.13
FRDC16.4816.3915.8918.6915.2421.3317.8618.2619.2118.0312.83
Phi-AI17.2617.2415.8119.5119.8620.9317.9119.0119.719.3517.3
Verb class (continue)
pushput downremoveremove screw fromrotatescrewtilt downtilt upunscrewnone
Base19.2920.2619.9916.4722.7122.9513.1215.1120.7819.82
JHands13.9613.7213.1111.7212.2614.119.619.9212.2510.99
PICO-AI12.2913.4112.839.9913.713.7210.4411.5612.8711.71
FRDC19.5217.8718.5514.4716.4418.8114.0313.3716.4115.01
Phi-AI19.1218.2918.1813.9518.3519.7813.1315.3617.2915.8
", + "image_path": "dedce2be5cc15fd6c00504010db757ecdeac83d0c142b42b81ab8f451b5b9aa8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "text", + "content": "linear warmup from 1e-7 to 1e-4 in the first " + }, + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "text", + "content": " steps and uses cosine decay from 1e-4 to 1e-7 for the rest of the steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 437, + 275, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 275, + 451 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 275, + 451 + ], + "type": "text", + "content": "5 Results and analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 464, + 200, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 200, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 200, + 475 + ], + "type": "text", + "content": "5.1 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 533 + ], + "type": "text", + "content": "Here, we benchmark results of valid submissions for state-of-the-art comparison in AssemblyHands and ARCTIC and other more recent baselines. In particular, for AssemblyHands, we report egocentric hand pose estimation results. For ARCTIC, we report results for the allocentric and egocentric test sets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "AssemblyHands benchmark: Table 3 shows the final test scores on the AssemblyHands dataset. The methods in the table exceed the baseline (Base) with a test score of 20.69 MPJPE. Notably, the methods JHands and PICO-AI achieve a nearly " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " reduction over the baseline. The methods FRDC and Phi-AI improve the test score by " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "20.3\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "16.5\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " against the baseline, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "ARCTIC benchmark: Table 4 presents the comparative performance of methods in the ARCTIC dataset, where ArcticNet-SF serves as the initial benchmark. The majority surpass ArcticNet-SF in both allocentric and egocentric views, except for UVHand, which underperforms due to incomplete training. In the egocentric view, AmbiguousHands excels in creating smooth, consistent mesh motions (refer to MDev and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{ACC}_{h / o}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " metrics). Notably, JointTransformer stands" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 202, + 480, + 293 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": "Table 4: Method performance in ARCTIC. We compare performance in both allocentric (top half) and egocentric (bottom half) views. We evaluate using metrics for contact and relative position (measuring hand-object contact and prediction of relative root position), motion (assessing temporally-consistent contact and smoothness), and hand and object metrics (indicating root-relative reconstruction error). We use the CDev score as the main metric for this competition. We denote left and right hands as " + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": ", and the object as " + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 202, + 480, + 293 + ], + "lines": [ + { + "bbox": [ + 134, + 202, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 134, + 202, + 480, + 293 + ], + "type": "table", + "html": "
Contact and Relative PositionsMotionHandObject
MethodCDev [mm]↓ MRRPEa/ro [mm]↓ MDev [mm]↓ ACCh/α [m/s2]↓ MPJPE [mm]↓ AAE [°] ↓ Success Rate [%] ↑
AllocentricArcticNet-SF41.5652.39/37.4710.405.72/7.5721.455.3771.39
DIGIT34.9244.19/35.438.374.86/6.6317.925.2476.52
UVHand64.1584.68/70.3114.127.05/12.0440.9912.3631.47
AmbiguousHands33.2545.78/34.5610.126.37/6.4018.024.6481.94
JointTransformer27.9736.17/28.188.936.08/5.7917.123.9589.79
Ego-centricArcticNet-SF44.7128.31/36.1611.805.03/9.1519.186.3953.89
DIGIT41.3125.49/32.619.484.01/8.3216.746.6053.33
UVHand40.4340.93/36.889.965.32/8.3324.537.3257.28
AmbiguousHands35.9323.07/27.539.513.95/6.7616.264.8668.36
JointTransformer32.5626.07/26.2211.345.52/8.6816.334.4474.07
", + "image_path": "4e14a9ddd492c48dd605f2a881b3b3787ee20ff48e54e1cfd635d92794654e10.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 305, + 480, + 357 + ], + "blocks": [ + { + "bbox": [ + 134, + 305, + 480, + 357 + ], + "lines": [ + { + "bbox": [ + 134, + 305, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 134, + 305, + 480, + 357 + ], + "type": "image", + "image_path": "f70b708de24ac2a015506912d092aab7fbd6af0ae3f5fb3de06d9c9f40476a16.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": "Fig. 2: Qualitative results per action in AssemblyHands. We show Base results with \"verb (noun)\" actions. The left three figures are lower error situations while the right four ones are failure cases. The red boxes denote the area where the action occurs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "text", + "content": "out by significantly lowering CDev errors by " + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "inline_equation", + "content": "32.7\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "text", + "content": " in allocentric and " + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "inline_equation", + "content": "27.2\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "text", + "content": " in egocentric settings compared to the baseline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 471, + 284, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 471, + 284, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 471, + 284, + 483 + ], + "type": "text", + "content": "5.2 AssemblyHands analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 496, + 481, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 481, + 520 + ], + "type": "text", + "content": "We provide analysis, regarding action-wise evaluation, distortion effect in training, and the effect of multi-view fusion. See SupMat for additional results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": "Action-wise evaluation: To analyze errors related to hand-object occlusions and interactions, we show pose evaluation according to fine verb action classes in Table 3. We use the verb classes annotated by Assembly101 [61], spanning every few seconds in a video. Figure 2 shows qualitative results of representative verb classes with the top and bottom error cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "We observe that the performance varies among different verb actions. The verbs \"tilt down/up\" and \"remove screw from\" exhibit lower errors among the submitted methods, because hands are less occluded and their movement is relatively stable. The \"tilt\" action holds a small part of the toy and turns it around alternately, leading to less overlap between the hand and the object (lib). The \"remove screw from\" action takes a screw out from the toy vehicle by their hand where observed hand poses do not change drastically." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 305, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 305, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 305, + 102 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 115, + 349, + 215 + ], + "blocks": [ + { + "bbox": [ + 141, + 115, + 349, + 215 + ], + "lines": [ + { + "bbox": [ + 141, + 115, + 349, + 215 + ], + "spans": [ + { + "bbox": [ + 141, + 115, + 349, + 215 + ], + "type": "image", + "image_path": "bfcce42240a3fc25f8f9024cb50c2ba7df097d8e93de6e8993ef4e333b395161.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 224, + 355, + 278 + ], + "lines": [ + { + "bbox": [ + 132, + 224, + 355, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 355, + 278 + ], + "type": "text", + "content": "Fig. 3: Effect of distortion in AssemblyHands. The officially released images in the dataset have highly stretched areas near the edges (original crop). The method JHands uses a perspective crop with a virtual camera to correct this distortion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 369, + 202, + 478, + 279 + ], + "blocks": [ + { + "bbox": [ + 367, + 125, + 482, + 201 + ], + "lines": [ + { + "bbox": [ + 367, + 125, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 367, + 125, + 482, + 201 + ], + "type": "text", + "content": "Table 5: Multi-view fusion in AssemblyHands. We use the Base result to show performance before and after fusion. Missing instances per view are denoted as \"Miss(\\%)\"." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 369, + 202, + 478, + 279 + ], + "lines": [ + { + "bbox": [ + 369, + 202, + 478, + 279 + ], + "spans": [ + { + "bbox": [ + 369, + 202, + 478, + 279 + ], + "type": "table", + "html": "
ViewMPJPEMiss(%)
cam137.9770.8
cam225.7188.3
cam322.190.92
cam422.290.74
cam3+421.520.08
all four20.690
", + "image_path": "f4ec049a039d28e352c7815ffb0a36202d3361245380132eb81e7629bb4a4272.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 482, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 412 + ], + "type": "text", + "content": "Higher error classes, such as \"inspect\", \"screw\", \"rotate\", and \"position\", contain heavy occlusions, fast hand motion, complex two hands and object interactions. The \"inspect\" action brings the toy close to the human eyes where the toy occupies a large portion of the image causing heavy object occlusions. The \"screw\" action involves intricate fingertip movements to rotate the screwdriver quickly. The \"rotate\" and \"position\" actions are performed so that the two hands and the object interact in close contact, which complicates the estimation. We observe that the top two methods JHands and PICO-AI significantly correct the results of these higher error actions compared to the other submitted methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 484 + ], + "type": "text", + "content": "Bias of hand position in an image: Hands near image edges are highly distorted due to the fish-eye cameras. Directly using these noisy images in training will degrade performance [84]; thus, some methods create new crops with less distortion, select training instances, or adaptively fuse predictions during the inference. Specifically, Figure 3 shows that the method JHands reformulates the perspective during cropping and creates less-distorted (perspective) crops." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "text", + "content": "To study this effect in the final performance, we split the evaluation instances into classes with different 2D distances between the hand center and the image center in Table 3. Higher distances (250- pixels) indicate closer hand crops to image edges. The method, Base, without any training instance selection and distortion correction, has higher error as the crops approach image edges (20.31 " + }, + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "text", + "content": " 24.85). In contrast, the newly proposed methods are more robust and have a lower error, particularly in the 200-250 range. We observe that the ranges 200-250 and 250- occupy " + }, + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "text", + "content": " of the test images, respectively, thus the improvement in the 200-250 range helps the lowering of the overall score." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 628 + ], + "type": "text", + "content": "Effect of multi-view fusion: The multi-view egocentric camera setup is unique to the dataset. We show the statistics and performance of multi-view fusion in Table 5. Note that Table 3 shows the final results after multi-view fusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "We found that samples captured from the lower cameras (cam3 and cam4; see Table 1 for the layout) are numerous (fewer missing samples) and their errors are lower as they are faced toward the area occurring hand interactions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 123, + 306, + 267 + ], + "blocks": [ + { + "bbox": [ + 200, + 116, + 304, + 123 + ], + "lines": [ + { + "bbox": [ + 200, + 116, + 304, + 123 + ], + "spans": [ + { + "bbox": [ + 200, + 116, + 304, + 123 + ], + "type": "text", + "content": "(a) Egocentric-allocentric comparison" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 167, + 123, + 306, + 267 + ], + "lines": [ + { + "bbox": [ + 167, + 123, + 306, + 267 + ], + "spans": [ + { + "bbox": [ + 167, + 123, + 306, + 267 + ], + "type": "image", + "image_path": "de923b27e131452ecd823d6759203356406cb7483ac08fa0d6cf68d936c8f67e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 276, + 480, + 310 + ], + "lines": [ + { + "bbox": [ + 131, + 276, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 480, + 310 + ], + "type": "text", + "content": "Fig. 4: Performance comparison: Egocentric vs Allocentric. (a) Comparative difficulty ratio of egocentric to allocentric views. (b) Egocentric view performance by method across objects. (c) Allocentric view performance by method across objects." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 123, + 446, + 190 + ], + "blocks": [ + { + "bbox": [ + 357, + 116, + 410, + 123 + ], + "lines": [ + { + "bbox": [ + 357, + 116, + 410, + 123 + ], + "spans": [ + { + "bbox": [ + 357, + 116, + 410, + 123 + ], + "type": "text", + "content": "(b) Egocentric view" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 123, + 446, + 190 + ], + "lines": [ + { + "bbox": [ + 307, + 123, + 446, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 123, + 446, + 190 + ], + "type": "image", + "image_path": "a3f0d36d71f802fa0824cba0f0261952046b8a21b15aec361ed4d8a11f4db8c6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 198, + 445, + 266 + ], + "blocks": [ + { + "bbox": [ + 358, + 191, + 410, + 198 + ], + "lines": [ + { + "bbox": [ + 358, + 191, + 410, + 198 + ], + "spans": [ + { + "bbox": [ + 358, + 191, + 410, + 198 + ], + "type": "text", + "content": "(c) Allocentric view" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 198, + 445, + 266 + ], + "lines": [ + { + "bbox": [ + 307, + 198, + 445, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 445, + 266 + ], + "type": "image", + "image_path": "7d4745864fdcfb2f47fee1c2fa7d7b487c45155f8cda94e7f978014bc72387a4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 334, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 480, + 430 + ], + "type": "text", + "content": "Conversely, the samples from cam1 and cam2 are fewer and unbalanced as their cameras often fail to capture hands due to the camera layout. For instance, cam1 (top-left) tends to capture more hand region than cam2 as the participants are mostly right-handed and bring up the object with the right hand, which can be better observed from cam1. Given this uneven sample distribution, the proposed adaptive view selection methods in either training or testing are essential to perform effective multi-view fusion, and outperform the Base's test-time average using all views all the time (see Multi-view fusion in Section 4.1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 430, + 481, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 515 + ], + "type": "text", + "content": "We further study the performance gain before and after multi-view fusion using Base's results. While per-view performance achieves 22.19 and 22.29 in cam3 and cam4, respectively, their fused results with simple average reduce the error to 21.52. Merging all four views has shown to be more effective than two-view fusion (20.69 vs. 21.52), indicating a " + }, + { + "bbox": [ + 130, + 430, + 481, + 515 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 130, + 430, + 481, + 515 + ], + "type": "text", + "content": " reduction compared to the single camera setup (cam3). This suggests predictions from the top views (cam1 and cam2) are informative in averaging even when they are prone to be erroneous." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 534, + 250, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 250, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 250, + 547 + ], + "type": "text", + "content": "5.3 ARCTIC analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 628 + ], + "type": "text", + "content": "Egocentric-allocentric comparison: Figure 4a compares the performance between egocentric and allocentric views. In particular, we compute a ratio between the metric values of the egocentric and allocentric view to measure the extent of difficulty for the egocentric view compared to the allocentric view. Since success rate is a metric whose value is positively correlated to performance, we take its reciprocal ratio. We average the ratios across methods and actions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We observe that hand pose-related metrics such as MPJPE and " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{ACC}_h" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " are less than 1.0 on average (see blue color cells), meaning the the egocentric view is easier than allocentric view. This is because most allocentric cameras in ARCTIC" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 304, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 304, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 304, + 101 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 112, + 449, + 177 + ], + "blocks": [ + { + "bbox": [ + 167, + 112, + 449, + 177 + ], + "lines": [ + { + "bbox": [ + 167, + 112, + 449, + 177 + ], + "spans": [ + { + "bbox": [ + 167, + 112, + 449, + 177 + ], + "type": "image", + "image_path": "e6b68b50c2cc66b47deb604456e31050bfb31943521ba650e2124c8b0a40ddcd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 184, + 482, + 217 + ], + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 217 + ], + "type": "text", + "content": "Fig. 5: Egocentric reconstruction by top method in ARCTIC. In the egocentric view, object reconstruction struggle when the object is partially observed on the image boundaries, as well as when heavy hand/arm occlusion occurs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 243, + 479, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 290 + ], + "type": "text", + "content": "are meters away from the subject while the egocentric camera is often close-up, offering higher hand visibility. Relative translation metrics between hand and object such as " + }, + { + "bbox": [ + 130, + 243, + 479, + 290 + ], + "type": "inline_equation", + "content": "\\mathrm{MRRPE}_{rl}" + }, + { + "bbox": [ + 130, + 243, + 479, + 290 + ], + "type": "text", + "content": " are also easier in the egocentric view because estimating translation is more difficult from further cameras." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 447 + ], + "type": "text", + "content": "Object reconstruction performance faces unique challenges, as highlighted by the red cells. In the egocentric view, it is notably more difficult to reconstruct accurate object surfaces, articulation (AAE), and hand-object contact (CDev). This increased difficulty arises because objects are often positioned at the image edges and obscured by human arms. Additionally, object poses exhibit greater diversity in the egocentric view due to varying camera angles and occlusions. While a static camera maintains consistent camera extrinsics across a sequence, an egocentric camera's extrinsics change with each frame, resulting in higher diversity in camera-view object 6D poses. This diversity complicates object pose estimation in egocentric views. Figure 5 illustrates these challenges using the best-performing method, JointTransformer. Despite achieving reasonable hand poses, object poses are significantly impacted by occlusions from hands and arms, and the egocentric view undergoes substantial changes throughout a sequence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 532 + ], + "type": "text", + "content": "Object-wise evaluation: Figure 4b and 4c break down performances on different objects in the egocentric and allocentric view test sets. The best method in both cases is JointTransformer. The hardest object to reconstruct with good contact consistency (see CDev) is the microwave in both settings due to global rotation sensitivity (see Fig. 5), though this can be mitigated by a keypoint-based approach [24]. Estimating objects in the egocentric view is more difficult than in the allocentric view, which is indicated by higher errors for all methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": "Action-wise evaluation: Figure 6 compares performance of different methods in \"grab\" and \"use\" actions. In ARCTIC, there are sequences to interact with the object with two types of actions by either not articulating the object, or allowing object articulation. Interestingly, the \"grab\" motion is more challenging in egocentric and allocentric views. We hypothesize that this is because there are more diverse object poses for \"grab\" motions since during object articulation, the participants often focus on articulation instead of object manipulation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Effect of model size: Figure 7 illustrates the impact of model size on handobject contact performance, measured by CDev, for reconstruction on the allocentric validation set. Most methods utilize ResNet50, with JointTransformer being the top performer. As the trainable parameters in the backbone increase," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 130, + 306, + 247 + ], + "blocks": [ + { + "bbox": [ + 133, + 130, + 306, + 247 + ], + "lines": [ + { + "bbox": [ + 133, + 130, + 306, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 130, + 306, + 247 + ], + "type": "image", + "image_path": "55f0d09062b13f1d742c0f0e963fe0c3739330130d02190cec3aca6a252904bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 257, + 308, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 257, + 308, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 308, + 335 + ], + "type": "text", + "content": "Fig.6: Hand-object contact quality for reconstructed results per action. We evaluate the contact quality of the 3D reconstruction results from all methods for each action (i.e., grab or use), using Contact Deviation (CDev) in mm as the metric, where lower values indicate better quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 317, + 116, + 478, + 247 + ], + "blocks": [ + { + "bbox": [ + 317, + 116, + 478, + 247 + ], + "lines": [ + { + "bbox": [ + 317, + 116, + 478, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 478, + 247 + ], + "type": "image", + "image_path": "607b42b7cc78a95cfcacfeb6cfc77bece5b5b1d9e1c697ac7d6c6cf3c6d29189.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 257, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 314, + 257, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 257, + 482, + 335 + ], + "type": "text", + "content": "Fig. 7: Contact deviation vs. model size. We assess the contact quality of the reconstruction results, varying by the number of parameters in each model. Contact quality is measured using Contact Deviation (CDev) in mm, with lower values indicating superior results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 358, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 431 + ], + "type": "text", + "content": "JointTransformer consistently reduces the CDev error. Note that the x-axis is in log-scale. JointTransformer achieved a CDev error of 30.5mm with ViT-L and 29.0mm with ViT-G, which has ten times more parameters than ViT-L. Interestingly, JointTransformer uses frozen weights in the large-scale ViT-L and ViT-G backbones, yet achieves the best results. This suggests a potential direction for leveraging large-scale foundational backbones for hand-object reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 569 + ], + "type": "text", + "content": "In this paper, we introduce the HANDS23 challenge and provide analysis based on the results of the top submitted methods and more recent baselines on the leaderboards. We organize and compare the submissions and their implementation details based on the learning methods, architecture, pre- and post-processing techniques, and training configurations. We thoroughly analyze various aspects, such as hand-object occlusions, action and object-wise evaluation, distortion correction, multi-view fusion, egocentric-allocentric comparison, and performance gain of large transformer models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Future directions: There are several future directions that the community can take. For example, one can explore more efficient training using multi-view egocentric cameras, leveraging 3D foundation priors [38,54] to regularize template-free hand-object reconstruction [15], estimating hand-object poses with more expressive representations (e.g., heatmap-based approaches [24]), incorporating motion and temporal modeling [18], featuring more diverse egocentric interaction scenarios, recognizing actions through captured hand poses [62], learning robotic grasping from reconstructed hand-object pose sequences, and so forth." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "content": "This work is supported by the Ministry of Science and ICT, Korea, under the ITRC program (IITP-2024-2020-0-01789). supervised by the IITP. UTokyo is supported by JST ACT-X Grant Number JPMJAX2007, JSPS KAKENHI Grant Number JP22KJ0999, JST Adopting Sustainable Partnerships for Innovative Research Ecosystem (ASPIRE) Grant Number JPMJAP2303. NUS is supported by the Ministry of Education, Singapore, under its MOE Academic Research Fund Tier 2 (STEM RIE2025 MOE-T2EP20220-0015)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 239, + 197, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 197, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 197, + 251 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 262, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 262, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 285 + ], + "type": "text", + "content": "1. Abou Zeid, K.: JointTransformer: Winner of the HANDS'2023 ARCTIC Challenge @ ICCV. https://github.com/kabouzeid/JointTransformer (2023) 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 328 + ], + "type": "text", + "content": "2. Armagan, A., Garcia-Hernando, G., Baek, S., Hampali, S., Rad, M., Zhang, Z., Xie, S., Chen, M., Zhang, B., Xiong, F., et al.: Measuring generalisation to unseen viewpoints, articulations, shapes and objects for 3d hand pose estimation under hand-object interaction. In: ECCV. pp. 85-101. Springer (2020) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 329, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 329, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 329, + 480, + 360 + ], + "type": "text", + "content": "3. Boukhayma, A., de Bem, R., Torr, P.H.S.: 3D hand shape and pose from images in the wild. In: Computer Vision and Pattern Recognition (CVPR). pp. 10843-10852 (2019) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 361, + 480, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 393 + ], + "type": "text", + "content": "4. Cai, Y., Ge, L., Cai, J., Yuan, J.: Weakly-supervised 3D hand pose estimation from monocular RGB images. In: European Conference on Computer Vision (ECCV). pp. 678-694 (2018) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 394, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 480, + 426 + ], + "type": "text", + "content": "5. Cao, Z., Radosavovic, I., Kanazawa, A., Malik, J.: Reconstructing hand-object interactions in the wild. In: International Conference on Computer Vision (ICCV). pp. 12417-12426 (2021) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 426, + 480, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 470 + ], + "type": "text", + "content": "6. Chao, Y.W., Yang, W., Xiang, Y., Molchanov, P., Handa, A., Tremblay, J., Narang, Y.S., Van Wyk, K., Iqbal, U., Birchfield, S., Kautz, J., Fox, D.: DexYCB: A benchmark for capturing hand grasping of objects. In: Computer Vision and Pattern Recognition (CVPR). pp. 9044-9053 (2021) 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 471, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 500 + ], + "type": "text", + "content": "7. Chatterjee, D., Sener, F., Ma, S., Yao, A.: Opening the vocabulary of egocentric actions. Conference on Neural Information Processing Systems (NeurIPS) 36 (2024) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 502, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 535 + ], + "type": "text", + "content": "8. Chen, X., Liu, Y., Dong, Y., Zhang, X., Ma, C., Xiong, Y., Zhang, Y., Guo, X.: MobRecon: mobile-friendly hand mesh reconstruction from monocular image. In: Computer Vision and Pattern Recognition (CVPR). pp. 20512-20522 (2022) 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "9. Chen, X., Wang, B., Shum, H.Y.: Hand avatar: Free-pose hand animation and rendering from monocular video. In: Computer Vision and Pattern Recognition (CVPR) (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "0. Chen, Z., Chen, S., Schmid, C., Laptev, I.: gSDF: Geometry-driven signed distance functions for 3d hand-object reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 12890-12900 (2023) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "1. Christen, S., Kocabas, M., Aksan, E., Hwangbo, J., Song, J., Hilliges, O.: D-Grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In: Computer Vision and Pattern Recognition (CVPR). pp. 20545–20554 (2022) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "2. Corona, E., Pumarola, A., Alenyà, G., Moreno-Noguer, F., Rogez, G.: GanHand: Predicting human grasp affordances in multi-object scenes. In: Computer Vision and Pattern Recognition (CVPR). pp. 5030-5040 (2020) 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "13. Duran, E., Kocabas, M., Choutas, V., Fan, Z., Black, M.J.: HMP: Hand motion priors for pose and shape estimation from video. In: Winter Conference on Applications of Computer Vision (WACV) (2024) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "text", + "content": "14. Erol, A., Bebis, G., Nicoescu, M., Boyle, R.D., Twombly, X.: Vision-based hand pose estimation: A review. Computer Vision and Image Understanding (CVIU) 108(1-2), 52-73 (2007) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 181, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 181, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 181, + 480, + 214 + ], + "type": "text", + "content": "15. Fan, Z., Parelli, M., Kadoglou, M.E., Kocabas, M., Chen, X., Black, M.J., Hilliges, O.: HOLD: Category-agnostic 3d reconstruction of interacting hands and objects from video. In: Computer Vision and Pattern Recognition (CVPR) (2024) 3, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 255 + ], + "type": "text", + "content": "16. Fan, Z., Spurr, A., Kocabas, M., Tang, S., Black, M.J., Hilliges, O.: Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In: International Conference on 3D Vision (3DV). pp. 1-10 (2021) 2, 3, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 256, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 256, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 256, + 480, + 300 + ], + "type": "text", + "content": "17. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 480, + 332 + ], + "type": "text", + "content": "18. Fu, Q., Liu, X., Xu, R., Niebles, J.C., Kitani, K.M.: Deformer: Dynamic fusion transformer for robust hand pose estimation. In: International Conference on Computer Vision (ICCV). pp. 23600-23611 (October 2023) 3, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "type": "text", + "content": "19. Garcia-Hernando, G., Yuan, S., Baek, S., Kim, T.K.: First-person hand action benchmark with rgb-d videos and 3d hand pose annotations. In: Computer Vision and Pattern Recognition (CVPR) (2018) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 364, + 480, + 396 + ], + "type": "text", + "content": "20. Ge, L., Ren, Z., Li, Y., Xue, Z., Wang, Y., Cai, J., Yuan, J.: 3D hand shape and pose estimation from a single rgb image. In: Computer Vision and Pattern Recognition (CVPR). pp. 10833-10842 (2019) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 480, + 429 + ], + "type": "text", + "content": "21. Grady, P., Tang, C., Twigg, C.D., Vo, M., Brahmbhatt, S., Kemp, C.C.: ContactOpt: Optimizing contact to improve grasps. In: Computer Vision and Pattern Recognition (CVPR). pp. 1471-1481 (2021) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "type": "text", + "content": "22. Guo, Z., Zhou, W., Wang, M., Li, L., Li, H.: HandNeRF: Neural radiance fields for animatable interacting hands. In: Computer Vision and Pattern Recognition (CVPR). pp. 21078-21087 (2023) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 460, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 460, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 133, + 460, + 480, + 493 + ], + "type": "text", + "content": "23. Hampali, S., Rad, M., Oberweger, M., Lepetit, V.: HOnnotate: A method for 3D annotation of hand and object poses. In: Computer Vision and Pattern Recognition (CVPR). pp. 3193-3203 (2020) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 492, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 492, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 492, + 480, + 536 + ], + "type": "text", + "content": "24. Hampali, S., Sarkar, S.D., Rad, M., Lepetit, V.: Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3D pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 11090-11100 (2022) 13, 14" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 536, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 590 + ], + "type": "text", + "content": "25. Han, S., Wu, P.C., Zhang, Y., Liu, B., Zhang, L., Wang, Z., Si, W., Zhang, P., Cai, Y., Hodan, T., Cabezas, R., Tran, L., Akbay, M., Yu, T.H., Keskin, C., Wang, R.: Umetrack: Unified multi-view end-to-end hand tracking for VR. In: International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH). pp. 50:1-50:9. ACM (2022) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 480, + 632 + ], + "type": "text", + "content": "26. Hasson, Y., Tekin, B., Bogo, F., Laptev, I., Pollefeys, M., Schmid, C.: Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 568-577 (2020) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "27. Hasson, Y., Varol, G., Schmid, C., Laptev, I.: Towards unconstrained joint handobject reconstruction from rgb videos. In: International Conference on 3D Vision (3DV). pp. 659-668. IEEE (2021) 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 101 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "28. Hasson, Y., Varol, G., Tzionas, D., Kalevatykh, I., Black, M.J., Laptev, I., Schmid, C.: Learning joint reconstruction of hands and manipulated objects. In: Computer Vision and Pattern Recognition (CVPR). pp. 11807-11816 (2019) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "30. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: Computer Vision and Pattern Recognition (CVPR). pp. 15979-15988 (2022) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "31. Huang, D., Ji, X., He, X., Sun, J., He, T., Shuai, Q., Ouyang, W., Zhou, X.: Reconstructing hand-held objects from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "32. Iqbal, U., Molchanov, P., Gall, T.B.J., Kautz, J.: Hand pose estimation via latent 2.5D heatmap regression. In: European Conference on Computer Vision (ECCV). pp. 118-134 (2018) 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "text", + "content": "33. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2O: Two hands manipulating objects for first person interaction recognition. In: International Conference on Computer Vision (ICCV). pp. 10138-10148 (2021) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "text", + "content": "34. Lee, J., Sung, M., Choi, H., Kim, T.K.: Im2hands: Learning attentive implicit representation of interacting two-hand shapes. In: Computer Vision and Pattern Recognition (CVPR). pp. 21169-21178 (2023) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 378 + ], + "type": "text", + "content": "35. Li, L., Tian, L., Zhang, X., Wang, Q., Zhang, B., Bo, L., Liu, M., Chen, C.: Renderih: A large-scale synthetic dataset for 3d interacting hand pose estimation. In: International Conference on Computer Vision (ICCV). pp. 20395-20405 (2023) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "36. Li, M., An, L., Zhang, H., Wu, L., Chen, F., Yu, T., Liu, Y.: Interacting attention graph for single image two-hand reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 2761-2770 (2022) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "37. Liu, R., Ohkawa, T., Zhang, M., Sato, Y.: Single-to-dual-view adaptation for egocentric 3d hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 677-686 (2024) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "38. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023) 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "39. Liu, S., Jiang, H., Xu, J., Liu, S., Wang, X.: Semi-supervised 3D hand-object poses estimation with interactions in time. In: Computer Vision and Pattern Recognition (CVPR). pp. 14687-14697 (2021) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: International Conference on Computer Vision (ICCV). pp. 10012-10022 (2021) 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "text", + "content": "41. Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Computer Vision and Pattern Recognition (CVPR). pp. 11966-11976 (2022) 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "42. Meng, H., Jin, S., Liu, W., Qian, C., Lin, M., Ouyang, W., Luo, P.: 3D interacting hand pose estimation by hand de-occlusion and removal. In: European Conference on Computer Vision (ECCV). pp. 380–397. Springer (2022) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "43. Moon, G.: Bringing inputs to shared domains for 3d interacting hands recovery in the wild. In: Computer Vision and Pattern Recognition (CVPR). pp. 17028-17037 (2023) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "44. Moon, G., Saito, S., Xu, W., Joshi, R., Buffalini, J., Bellan, H., Rosen, N., Richardson, J., Mize, M., De Bree, P., et al.: A dataset of relighted 3d interacting hands. Conference on Neural Information Processing Systems (NeurIPS) 36 (2024) 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "45. Moon, G., Yu, S., Wen, H., Shiratori, T., Lee, K.M.: Interhand2.6m: A dataset and baseline for 3d interacting hand pose estimation from a single RGB image. In: European Conference on Computer Vision (ECCV). vol. 12365, pp. 548-564 (2020) 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "46. Mueller, F., Bernard, F., Sotnychenko, O., Mehta, D., Sridhar, S., Casas, D., Theobalt, C.: GANerated hands for real-time 3D hand tracking from monocular RGB. In: Computer Vision and Pattern Recognition (CVPR). pp. 49-59 (2018) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "47. Ohkawa, T.: AssemblyHands Toolkit. https://github.com/facebookresearch/assemblyhands-toolkit (2023) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "48. Ohkawa, T., Furuta, R., Sato, Y.: Efficient annotation and learning for 3d hand pose estimation: A survey. International Journal of Computer Vision (IJCV) 131, 3193-3206 (2023) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "49. Ohkawa, T., He, K., Sener, F., Hodan, T., Tran, L., Keskin, C.: AssemblyHands: towards egocentric activity understanding via 3d hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 12999-13008 (2023) 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "50. Ohkawa, T., Li, Y.J., Fu, Q., Furuta, R., Kitani, K.M., Sato, Y.: Domain adaptive hand keypoint and pixel localization in the wild. In: European Conference on Computer Vision (ECCV). pp. 68—87 (2022) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "51. Ohkawa, T., Yagi, T., Hashimoto, A., Ushiku, Y., Sato, Y.: Foreground-aware stylization and consensus pseudo-labeling for domain adaptation of first-person hand segmentation. IEEE Access 9, 94644-94655 (2021) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 402 + ], + "type": "text", + "content": "52. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "53. Park, J., Oh, Y., Moon, G., Choi, H., Lee, K.M.: HandOccNet: Occlusion-robust 3d hand mesh estimation network. In: Computer Vision and Pattern Recognition (CVPR). pp. 1496–1505 (2022) 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "54. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022) 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "type": "text", + "content": "55. Prakash, A., Tu, R., Chang, M., Gupta, S.: 3d hand pose estimation in everyday egocentric images. In: European Conference on Computer Vision (ECCV) (2024) 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "56. Radosavovic, I., Kosaraju, R.P., Girshick, R.B., He, K., Dollár, P.: Designing network design spaces. In: Computer Vision and Pattern Recognition (CVPR). pp. 10425-10433 (2020) 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "text", + "content": "57. Rehg, J.M., Kanade, T.: Visual tracking of high DOF articulated structures: An application to human hand tracking. In: European Conference on Computer Vision (ECCV). vol. 801, pp. 35-46 (1994) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "type": "text", + "content": "58. Remelli, E., Han, S., Honari, S., Fua, P., Wang, R.: Lightweight multi-view 3d pose estimation through camera-disentangled representation. In: Computer Vision and Pattern Recognition (CVPR). pp. 6039-6048 (2020) 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "59. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG) 36(6), 245:1-245:17 (2017) 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "text", + "content": "60. Ryali, C., Hu, Y., Bolya, D., Wei, C., Fan, H., Huang, P., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: International Conference on Machine Learning (ICML). vol. 202, pp. 29441-29454 (2023) 5, 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "text", + "content": "61. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Computer Vision and Pattern Recognition (CVPR). pp. 21064-21074 (2022) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 195 + ], + "type": "text", + "content": "62. Shamil, M.S., Chatterjee, D., Sener, F., Ma, S., Yao, A.: On the utility of 3d hand poses for action recognition. In: European Conference on Computer Vision (ECCV) (2024) 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "63. Simon, T., Joo, H., Matthews, I., Sheikh, Y.: Hand keypoint detection in single images using multiview bootstrapping. In: Computer Vision and Pattern Recognition (CVPR). pp. 4645-4653 (2017) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "text", + "content": "64. Spurr, A., Dahiya, A., Wang, X., Zhang, X., Hilliges, O.: Self-supervised 3D hand pose estimation from monocular RGB via contrastive learning. In: International Conference on Computer Vision (ICCV). pp. 11210-11219 (2021) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "65. Spurr, A., Iqbal, U., Molchanov, P., Hilliges, O., Kautz, J.: Weakly supervised 3D hand pose estimation via biomechanical constraints. In: European Conference on Computer Vision (ECCV). vol. 12362, pp. 211-228 (2020) 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "66. Spurr, A., Song, J., Park, S., Hilliges, O.: Cross-modal deep variational hand pose estimation. In: Computer Vision and Pattern Recognition (CVPR). pp. 89-98 (2018) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "text", + "content": "67. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Computer Vision and Pattern Recognition (CVPR) (2019) 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 363, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 407 + ], + "type": "text", + "content": "68. Swamy, A., Leroy, V., Weinzaepfel, P., Baradel, F., Galaoui, S., Brégier, R., Armando, M., Franco, J.S., Rogez, G.: SHOWMe: Benchmarking object-agnostic hand-object 3d reconstruction. In: International Conference on Computer Vision (ICCV). pp. 1935–1944 (2023) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "type": "text", + "content": "69. Tekin, B., Bogo, F., Pollefeys, M.: " + }, + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{H} + \\mathrm{O}" + }, + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "type": "text", + "content": ": Unified egocentric recognition of 3D hand-object poses and interactions. In: Computer Vision and Pattern Recognition (CVPR). pp. 4511-4520 (2019) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 441, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 475 + ], + "type": "text", + "content": "70. Tse, T.H.E., Kim, K.I., Leonardis, A., Chang, H.J.: Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In: Computer Vision and Pattern Recognition (CVPR). pp. 1664-1674 (2022) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "71. Tzionas, D., Gall, J.: A comparison of directional distances for hand pose estimation. In: German Conference on Pattern Recognition (GCPR). vol. 8142, pp. 131-141 (2013) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "72. Wen, Y., Pan, H., Ohkawa, T., Yang, L., Pan, J., Sato, Y., Komura, T., Wang, W.: Generative hierarchical temporal transformer for hand action recognition and motion prediction. CoRR abs/2311.17366 (2023) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "text", + "content": "73. Yang, L., Chen, S., Yao, A.: Semihand: Semi-supervised hand pose estimation with consistency. In: International Conference on Computer Vision (ICCV). pp. 11364-11373 (2021) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "text", + "content": "74. Yang, L., Zhan, X., Li, K., Xu, W., Li, J., Lu, C.: CPF: Learning a contact potential field to model the hand-object interaction. In: International Conference on Computer Vision (ICCV) (2021) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "75. Ye, Y., Gupta, A., Tulsiani, S.: What's in your hands? 3d reconstruction of generic objects in hands. In: Computer Vision and Pattern Recognition (CVPR) (2022) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "76. Ye, Y., Hebbar, P., Gupta, A., Tulsiani, S.: Diffusion-guided reconstruction of everyday hand-object interaction clips. In: International Conference on Computer Vision (ICCV) (2023) 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Benchmarks and Challenges for Egocentric Hand Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 544 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 182 + ], + "type": "text", + "content": "77. Yuan, S., Garcia-Hernando, G., Stenger, B., Moon, G., Chang, J.Y., Lee, K.M., Molchanov, P., Kautz, J., Honari, S., Ge, L., Yuan, J., Chen, X., Wang, G., Yang, F., Akiyama, K., Wu, Y., Wan, Q., Madadi, M., Escalera, S., Li, S., Lee, D., Oikonomidis, I., Argyros, A.A., Kim, T.: Depth-based 3D hand pose estimation: From current achievements to future goals. In: Computer Vision and Pattern Recognition (CVPR). pp. 2636-2645 (2018) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "78. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Computer Vision and Pattern Recognition (CVPR). pp. 12104-12113 (2022) 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 481, + 237 + ], + "type": "text", + "content": "79. Zhang, H., Christen, S., Fan, Z., Hilliges, O., Song, J.: GraspXL: Generating grasping motions for diverse objects at scale. European Conference on Computer Vision (ECCV) (2024) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "80. Zhang, H., Christen, S., Fan, Z., Zheng, L., Hwangbo, J., Song, J., Hilliges, O.: ArtiGrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. In: International Conference on 3D Vision (3DV) (2024) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "81. Zhang, X., Li, Q., Mo, H., Zhang, W., Zheng, W.: End-to-end hand mesh recovery from a monocular RGB image. In: International Conference on Computer Vision (ICCV). pp. 2354-2364 (2019) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "82. Zhou, Y., Habermann, M., Xu, W., Habibie, I., Theobalt, C., Xu, F.: Monocular real-time hand shape and motion capture using multi-modal data. In: Computer Vision and Pattern Recognition (CVPR). pp. 5345-5354 (2020) 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "text", + "content": "83. Zhou, Z.: SimpleHand: Winner of the HANDS'2023 AssemblyHands Challenge @ ICCV. https://github.com/patienceFromZhou/simpleHand (2024) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "84. Zhou, Z., Lv, Z., Zhou, S., Zou, M., Wu, T., Yu, M., Tang, Y., Liang, J.: 1st place solution of egocentric 3d hand pose estimation challenge 2023 technical report: A concise pipeline for egocentric hand pose reconstruction. CoRR abs/2310.04769 (2023) 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "85. Zhou, Z., Zhou, S., Lv, Z., Zou, M., Tang, Y., Liang, J.: A simple baseline for efficient hand mesh reconstruction. In: Computer Vision and Pattern Recognition (CVPR). pp. 1367-1376 (2024) 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "86. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (ICLR). OpenReview.net (2021), https://openreview.net/forum?id=gZ9hCDWe6ke 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "87. Ziani, A., Fan, Z., Kocabas, M., Christen, S., Hilliges, O.: TempCLR: Reconstructing hands via time-coherent contrastive learning. In: International Conference on 3D Vision (3DV). pp. 627-636 (2022) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "88. Zimmermann, C., Brox, T.: Learning to estimate 3D hand pose from single RGB images. In: International Conference on Computer Vision (ICCV). pp. 4913-4921 (2017) 3, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "Z. Fan, T. Ohkawa, L. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Beta-Tuned Timestep Diffusion Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_content_list.json b/2024/Beta-Tuned Timestep Diffusion Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3400f8db469da719732c64c966c1bb5ed7b22e2a --- /dev/null +++ b/2024/Beta-Tuned Timestep Diffusion Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_content_list.json @@ -0,0 +1,2009 @@ +[ + { + "type": "text", + "text": "Beta-Tuned Timestep Diffusion Model", + "text_level": 1, + "bbox": [ + 282, + 140, + 722, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyi Zheng $^{1,2}$ , Peng-Tao Jiang $^{2}$ , Ben Wan $^{1}$ , Hao Zhang $^{2}$ , Jinwei Chen $^{2}$ , Jia Wang $^{1}$ $\\boxtimes$ , and Bo Li $^{2}$ $\\boxtimes$", + "bbox": [ + 282, + 188, + 722, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University, Shanghai, China {tyzheng, burn-w, jiawang}@sjtu.edu.cn $^{2}$ vivo Mobile Communication Co., Ltd, China {pt.jiang,haozhang,jinwei.chen,libra}@vivo.com", + "bbox": [ + 321, + 229, + 679, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Diffusion models have received a lot of attention in the field of generation due to their ability to produce high-quality samples. However, several recent studies indicate that treating all distributions equally in diffusion model training is sub-optimal. In this paper, we conduct an in-depth theoretical analysis of the forward process of diffusion models. Our findings reveal that the distribution variations are non-uniform throughout the diffusion process and the most drastic variations in distribution occur in the initial stages. Consequently, simple uniform timestep sampling strategy fail to align with these properties, potentially leading to suboptimal training of diffusion models. To address this, we propose the Beta-Tuned Timestep Diffusion Model (B-TTDM), which devises a timestep sampling strategy based on the beta distribution. By choosing the correct parameters, B-TTDM aligns the timestep sampling distribution with the properties of the forward diffusion process. Extensive experiments on different benchmark datasets validate the effectiveness of B-TTDM.", + "bbox": [ + 259, + 340, + 743, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Generative Models $\\cdot$ Diffusion Models $\\cdot$ Beta Distribution", + "bbox": [ + 261, + 563, + 733, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 603, + 375, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have gained widespread application across various domains in recent years, such as text-to-image generation [5,43,50], text generation [31,52], video synthesis [21] and AI security [6,12-14,32], primarily due to their ability to produce high-quality samples. During the training stage, diffusion models introduce varying levels of Gaussian noise into the data distribution, creating a range of altered distributions. The model is then trained by minimizing denoising score-matching losses across these diverse distributions. In the inference stage, the trained model reverses this noise addition process to generate new samples.", + "bbox": [ + 212, + 636, + 787, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While DDPM [22], iDDPM [37] and ADM [9] have achieved significant success in image generation tasks, several studies [7, 17, 19, 48] identify that treating the distribution equally across different timesteps in the training stage could be", + "bbox": [ + 212, + 756, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\boxtimes$ Corresponding authors. This work was done during Tianyi Zheng's internship at vivo.", + "bbox": [ + 228, + 809, + 643, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sub-optimal. For instance, the P2-Weight [7] method discover through empirical analysis that distributions at various timesteps in forward diffusion encapsulate distinct semantic information. Building on this insight, P2-Weight propose a re-weighted loss function to elevate the quality of generated samples. Meanwhile, Min-SNR [19] and ANT [17] find inconsistencies in optimization gradients across various timesteps and recommend a re-weighted loss strategy to mitigate these issues. Recently, E-TSDM [48] uncover that the issue of numerical instability caused by high Lipschitz constants near the zero point. Therefore, E-TSDM assigns a weight of 0 to certain timesteps near zero to mitigate the issue of numerical instability. Despite the improvement in sample quality achieved by these methods through empirical adjustments to the training objective's weight, they still utilize DDPM's uniform timestep sampling strategy. As another influencing factor, the choice of timestep sampling strategy is also closely related to the training objective but has been overlooked in previous work. In this paper, we conduct a theoretical analysis of DDPM's forward diffusion process and design a novel strategy for timestep sampling, aiming to address this overlooked aspect and further improve the quality of generated samples.", + "bbox": [ + 212, + 145, + 787, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our theoretical analysis in Theorem 2 demonstrates that distribution variations throughout the diffusion process are non-uniform, with the most significant variations occurring at the initial stages. This is experimentally supported by Figure 1, which shows non-uniform variations in both Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) [51] metrics, especially pronounced in the initial stages. Both our theoretical insights from Theorem 2 and the experimental evidence in Figure 1 underscore the non-uniform properties of forward diffusion, indicating that a uniform timestep sampling strategy may not be the most effective approach in the training stage.", + "bbox": [ + 212, + 406, + 789, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these findings, we propose a novel timestep sampling strategy that utilizes the beta distribution, named the Beta-Tuned Timestep Diffusion Model (B-TTDM). By choosing the correct parameters of the beta distribution, B-TTDM aligns the timestep sampling distribution with the observed properties of the diffusion process and is specifically tuned to avoid singularities at $t = 0$ timesteps. Extensive experiments on various datasets demonstrate that B-TTDM not only improves the quality of the generated samples but also expedites the training process. For instance, B-TTDM achieves similar FID scores on the ImageNet dataset with about $20\\%$ of the training iterations needed by the original approach and significantly improves generation quality after convergence. Moreover, B-TTDM can be compatible with other advanced diffusion model training methods, such as P2-Weight [7], Min-SNR [19], Debias [19], and DDPM-IP [39]. Through integration with these methods, B-TTDM can combine different advantages and generate higher quality samples. In summary, our contributions are:", + "bbox": [ + 212, + 545, + 789, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Through theoretical and experimental verification, we confirm that distribution variations at different timesteps during forward diffusion are non-uniform, with the most significant shifts occurring in the initial stages. This indicates that the uniform time sampling in training is sub-optimal.", + "bbox": [ + 225, + 779, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a novel timestep sampling strategy leveraging the Beta distribution to align with the non-uniform distribution shifts observed during forward diffusion. Furthermore, our B-TTDM is compatible with other enhancement methods, allowing for flexible integration into existing improved methods.", + "- Extensive experiments on different benchmark datasets demonstrate that the B-TTDM method significantly improves generation quality and drastically speeds up training of the diffusion models." + ], + "bbox": [ + 223, + 146, + 784, + 251 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac860bafc98cfff77481754b1e8e2308e330b625403463731bbf7bfd38393657.jpg", + "image_caption": [ + "(a) PSNR at each forward diffusion timestep." + ], + "image_footnote": [], + "bbox": [ + 223, + 287, + 491, + 445 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cecff69cba16f4b21ecc8a2341fae4bfb0627e85784655fc797f5de07440ab81.jpg", + "image_caption": [ + "(b) LPIPS at each forward diffusion timestep.", + "Fig. 1: The PSNR and LPIPS metrics between the distribution at timestep $t$ and the original distribution in forward diffusion under different noise schedules." + ], + "image_footnote": [], + "bbox": [ + 511, + 287, + 777, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 551, + 387, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models [15,22,37] are a family of generative models that generate samples from Gaussian noise via a learned denoising process. Recently, diffusion models have outperformed flow models [10,27] and variational autoencoders [16, 28] in generating higher quality samples. Furthermore, the diffusion models also outperform GANs [18,47] in certain cases [9]. This superior performance has led to the widespread adoption of diffusion models across diverse applications. For example, the Diffusion model [43] implements the generation of images based on textual inputs [4,11], image to image translation [35,40] and 3D Avatar generation [41]. Despite their success, the speed of sample generation by diffusion models is unsatisfactory. Therefore, DPM-Solver [34] and DDIM [44] design new generation methods based on the framework of ordinary differential equation (ODE), which have greatly improved the generation speed of the diffusion model. Meanwhile, EDM [25] design a higher order Runge-Kutta sampling method to further improve the generation speed and quality. These methods significantly enhance the speed of sample generation in diffusion models, with a primary emphasis on the generative aspect. Yet, they rely on pre-trained diffusion models without addressing training optimizations. Our proposed method presents an", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Beta-Tuned Timestep Diffusion Model", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "orthogonal strategy, aiming to improve the training stage and thereby further elevate the generate quality of diffusion models.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Improved Diffusion Models. To further improve the quality of generative samples in the training stage, ADM [9] and U-ViT [2] design new advanced architectures for diffusion models. Based on these advanced architectures, the quality of generation is significantly improved. Additionally, P2-Weight [7], Min-SNR [19] and Debias [49] methods design the re-weight loss function based on SNR in the training stage. These can reduce the bias problem in the diffusion models. To make the neural network smoother, the E-TSDM [48] sets the weight of the loss function corresponding to timesteps close to the zero to 0. Additional, ANT [17] propose multi-task learning methods to set the loss weight of different timestep in the training stage. Although these improved methods can produce better samples, they overlook the non-uniform distribution variations in the diffusion process and continue to use uniform timestep sampling during training. Our theoretical analysis indicates that this uniform time sampling strategy in the training stage is not optimal, suggesting that a more tailored timestep sampling approach could yield significant improvements in generated samples.", + "bbox": [ + 212, + 176, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 425, + 331, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Background", + "text_level": 1, + "bbox": [ + 215, + 455, + 362, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $q(x_0)$ represent the data distribution over $\\mathbb{R}^n$ . During the forward diffusion, a sample $x_0$ is drawn from $q(x_0)$ and subsequently perturbed using a stochastic differential equation (SDE) [46].", + "bbox": [ + 212, + 478, + 787, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} x _ {t} = f \\left(x _ {t}, t\\right) \\mathrm {d} t + g (t) \\mathrm {d} w. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 535, + 787, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The purpose of the forward diffusion is to transform the data distribution into simple normal Gaussian distribution $\\mathcal{N}(0, I)$ . Meanwhile, we define $q_t(x_t)$ to represent the distribution at each time $t$ during the forward diffusion process. For any time $t$ , the conditional distribution $q_t(x_t | x_0)$ is [9, 22, 37]", + "bbox": [ + 212, + 560, + 789, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {t} \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t} \\mid \\alpha_ {t} x _ {0}, \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 636, + 787, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Equation (2), $\\alpha_{t}$ and $\\sigma_{t}$ satisfy $\\alpha_{t}^{2} + \\sigma_{t}^{2} = 1$ . This forward diffusion process is known as the Ornstein-Uhlenbeck (OU) process. For more general expressions, we have the following conditional distribution", + "bbox": [ + 212, + 657, + 787, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {t} \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t} \\mid e ^ {- \\frac {t}{2}} x _ {0}, (1 - e ^ {- t}) \\mathbf {I}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 715, + 787, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The reverse diffusion transforms normal Gaussian distribution into the data distribution $q(x_0)$ . This process satisfies the following reverse SDE equation [1].", + "bbox": [ + 212, + 742, + 787, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} x _ {t} = \\left(f \\left(x _ {t}, t\\right) - g (t) ^ {2} \\nabla_ {x} \\log q _ {t} \\left(x _ {t}\\right)\\right) \\mathrm {d} t + g (t) \\mathrm {d} w. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 787, + 787, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specially, $\\nabla_x\\log q_t(x_t)$ is the score function of the probability density $q_{t}(x_{t})$ Motivated by this correspondence, diffusion models train a neural network to", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimate the $-\\sigma_t\\nabla_x\\log q_t(x_t)$ . In the inference stage, we can simulate Equation (4) for sampling and generate new samples.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Non-uniform variation in forward diffusion", + "text_level": 1, + "bbox": [ + 214, + 200, + 612, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first start by introducing and discussing our main assumptions. The only assumption we consider is that in the vicinity of high-density data, the score of the data distribution is bounded. Based on the definition of score [23, 45, 46], this assumption is quite common used [3, 30].", + "bbox": [ + 212, + 226, + 782, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assumption 1 For the diffusion process described by Equation (3), assume there exists a constant $\\delta >0$ . For any $t\\in [0,T]$ and any point $x_{t}$ in high-density regions, the score $\\| \\nabla_x\\log q_t(x_t)\\|$ is bounded by $\\delta$ , i.e.,", + "bbox": [ + 212, + 297, + 784, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\nabla_ {x} \\log q _ {t} (x _ {t}) \\right\\| \\leq \\delta .\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 356, + 571, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This ensures that the score of the data distribution is bounded in the vicinity of high-density data throughout the diffusion process.", + "bbox": [ + 212, + 383, + 782, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we examine the Fokker-Planck equation [42] corresponding to the Ornstein-Uhlenbeck process, as depicted in Equation (5). The parameter $\\theta$ is drift coefficient and the parameter $D$ is diffusion coefficient.", + "bbox": [ + 212, + 425, + 782, + 469 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial}{\\partial t} q \\left(x _ {t}\\right) = \\theta \\nabla_ {x _ {t}} q \\left(x _ {t}\\right) + D \\nabla_ {x _ {t}} ^ {2} q \\left(x _ {t}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 482, + 785, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the $\\theta$ and $D$ are both constants independent of $x_{t}$ , to simplify the analysis that follows, we can let both of them be 1. Therefore, we have the following simplified Fokker-Planck equation.", + "bbox": [ + 212, + 516, + 782, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial}{\\partial t} q \\left(x _ {t}\\right) = \\nabla_ {x _ {t}} q \\left(x _ {t}\\right) + \\nabla_ {x _ {t}} ^ {2} q \\left(x _ {t}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 574, + 785, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Fokker-Planck equation (6) sheds light on how a distribution's evolution is influenced by both time and spatial variables in the forward diffusion stage of DDPM. Our analysis further reveals that in the initial stages of diffusion, the changes in the distribution are primarily driven by the second-order derivative term, indicating a significant diffusion effect at this stage. Therefore, we have the following Lemma:", + "bbox": [ + 212, + 609, + 782, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 1. Consider the forward diffusion process as described by Equation (3) When $t\\to 0^+$ , the time derivative of the distribution $q_{t}(x_{t})$ is predominantly governed by the second-order derivative term, encapsulated in the relationship:", + "bbox": [ + 212, + 712, + 782, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right) = \\nabla_ {x _ {t}} ^ {2} q \\left(x _ {t}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 768, + 785, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nabla_{x_t}^2$ denotes the Laplacian operator with respect to $x_{t}$ , indicating the diffusion term's dominance in the evolution of $q_{t}(x_{t})$ when $t\\to 0^{+}$ .", + "bbox": [ + 212, + 809, + 782, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Beta-Tuned Timestep Diffusion Model", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The detailed proof of the Lemma (1) is shown in Appendix Section 2.2. Lemma (1) suggests that the alteration in the distribution primarily stems from the $\\nabla_{x_t}^2 q(x_t)$ . Expanding on this insight, we can establish the following theorem, providing an upper bound on the alteration in distribution.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 2. Consider the forward diffusion process described by Equation (7). The norm of the time derivative of the distribution $q_{t}(x_{t})$ is bounded by the following inequality:", + "bbox": [ + 212, + 217, + 787, + 262 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right) \\right\\| \\leq \\left\\| \\frac {q _ {t} \\left(x _ {t}\\right)}{2 t} \\right\\|. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 260, + 787, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As $t \\to 0^{+}$ , the norm of the derivative of the distribution $q_{t}(x_{t})$ may become unbounded. This is indicated by the limit superior:", + "bbox": [ + 212, + 300, + 785, + 330 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {t \\rightarrow 0 ^ {+}} \\left\\| \\frac {q _ {t} \\left(x _ {t}\\right)}{2 t} \\right\\| = \\infty . \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 340, + 787, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The detailed proof of the Theorem (2) is shown in Appendix Section 2.3. Theorem (2) suggests that the bound on the norm of the time derivative of $q_{t}(x_{t})$ grows indefinitely as $t \\to 0^{+}$ , indicating that the norm of the derivative could be unbounded. We present a straightforward example to illustrate this behavior.", + "bbox": [ + 212, + 402, + 785, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Simple case illustration. Let $q_{t}(x_{t}) = \\frac{1}{\\sqrt{4\\pi t}} e^{-\\frac{x^{2}}{4t}}$ be a solution to differential equation (7). Meanwhile, we notice that $\\int_{-\\infty}^{+\\infty} q_{t}(x_{t}) dx = 1$ and $q_{t}(x) > 0$ , which means $q_{t}(x)$ is a probability density function. For every $t > 0$ , the norm of the time derivative of $q_{t}(x)$ is given by:", + "bbox": [ + 212, + 465, + 785, + 532 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right) \\right\\| = \\left\\| \\frac {\\sqrt {t} (- 2 t + x ^ {2}) e ^ {- \\frac {x ^ {2}}{4 t}}}{8 \\sqrt {\\pi} t ^ {3}} \\right\\|. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 542, + 787, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This expression leads to the conclusion that:", + "bbox": [ + 214, + 595, + 539, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {t \\rightarrow 0 ^ {+}} \\left\\| \\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right)\\right\\| = \\infty , \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 619, + 787, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "indicating that the norm of the time derivative of $q_{t}(x)$ becomes unbounded as $t \\to 0^{+}$ . This result indicates an infinite rate of variation for the distribution $q_{t}(x_{t})$ when $t \\to 0^{+}$ .", + "bbox": [ + 212, + 665, + 785, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem (2) highlights the non-uniform properties of distribution variations during diffusion, with the initial stages experiencing the most significant shifts. Furthermore, Proposition (1) clarifies that the distribution variations $q_{t}(x_{t})$ are non-uniform across the diffusion process, and these variations tend to flatten as the process progresses to later stages.", + "bbox": [ + 212, + 710, + 785, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 1. Consider the forward diffusion process in Equation (3). The upper bound of the KL divergence between the distribution $q_{t}(x_{t})$ and the standard Gaussian distribution converges exponentially. i.e.,", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K L} \\left(q _ {t} \\left(x _ {t}\\right) \\| \\mathcal {N} (0, 1)\\right) \\leq e ^ {- t} \\operatorname {K L} \\left(q _ {0} \\left(x _ {0}\\right) \\| \\mathcal {N} (0, 1)\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 161, + 787, + 179 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the Theorem (2) and Proposition (1), we notice that the variations in the distribution of forward diffusion is non-uniform, with the sharpest change at the beginning of the diffusion and converging to a Gaussian distribution at an exponential rate. These theoretical insights guide the design of our timestep sampling strategy.", + "bbox": [ + 212, + 209, + 784, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Beta distribution timestep sampling", + "text_level": 1, + "bbox": [ + 214, + 308, + 560, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To align the non-uniform trend of variation, we introduce a time sampling strategy that utilizes the Beta distribution for training the diffusion model. This strategy leverages the Beta density function's capacity to take on a diverse array of shapes, controlled by the parameters $\\alpha$ and $\\beta$ . The probability density function (pdf) of the Beta distribution is defined as:", + "bbox": [ + 212, + 334, + 784, + 409 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nf (t; \\alpha , \\beta) = \\frac {\\Gamma (\\alpha + \\beta)}{\\Gamma (\\alpha) \\Gamma (\\beta)} t ^ {\\alpha - 1} (1 - t) ^ {\\beta - 1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 419, + 787, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Where $\\Gamma(z)$ is the gamma function. Setting $\\alpha = 1$ and $\\beta = 1$ simplifies the Beta distribution to a uniform distribution, illustrating the flexibility of Beta distribution. By adjusting the values of $\\alpha$ and $\\beta$ , we can tailor the time sampling distribution within the diffusion model training to better align the desired variation properties.", + "bbox": [ + 212, + 460, + 784, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the beta distribution is a family of continuous probability distributions defined on the interval [0, 1]. We can use $\\frac{t}{T}$ to map the diffusion timestep into the [0, 1] interval of the beta distribution. Then we compare the probability density function of beta distribution for different $\\alpha$ and $\\beta$ in Figure 2. We find that when we set $\\alpha > 1$ and $\\beta > 1$ , The probability density function is U-shaped. Meanwhile, we notice that $\\alpha = 1$ and $\\beta > 1$ and $\\alpha < 1$ and $\\beta = 1$ , the probability density function conforms to our expectation of a non-uniformly decreasing. However, when $\\alpha = 1$ and $\\beta > 1$ , the probability density function exhibits a positively skewed shape, lacking alignment with the most significant varies in the initial diffusion stage. We provide ablation study of different $\\alpha$ and $\\beta$ in the section 4.4.", + "bbox": [ + 212, + 537, + 787, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To align with the previously identified properties, we choose to set $\\alpha < 1$ and $\\beta = 1$ . With these parameters, the distribution adopts a reverse J-shaped profile. Consequently, the probability density function under these conditions becomes a strictly decreasing function, exhibiting a sharp change as it approaches 0. This specific shape and behavior of the pdf are instrumental in aligning the rapid initial changes and the convergence properties observed in the forward diffusion process. Therefore, we replace the uniform time sampling in the original diffusion model with this time sampling method based on the beta distribution. In addition, to prevent the probability density function from reaching infinite values as time t approaches 0, we adopt $\\frac{1}{T}$ as the time sampling weight at the initial", + "bbox": [ + 212, + 688, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Beta-Tuned Timestep Diffusion Model", + "bbox": [ + 473, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47b782961c05d521e76e9e8d6c4c37ae95c5618af0e96ec98c2b7f0d681b09fc.jpg", + "image_caption": [ + "Fig. 2: Probability density function of beta distribution with different parameters." + ], + "image_footnote": [], + "bbox": [ + 318, + 146, + 684, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "moment. Therefore, our method is called Beta-Tuned Timestep Diffusion Model (B-TTDM). Next, we analyse the errors introduced by this tuned approach.", + "bbox": [ + 212, + 373, + 784, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 2. Given a timestep sampling method utilizing the B-TTDM with a Beta distribution parameterized by $\\alpha < 1$ and $\\beta = 1$ , the tuned error at the initial time position $t$ is $\\left(\\frac{1}{T}\\right)^{\\alpha}$ , where $T$ is the total diffusion steps.", + "bbox": [ + 212, + 411, + 785, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The detailed proof of the Proposition (2) is shown in Appendix Section 2.4. Since the value of $T$ is usually large in the training of diffusion models (e.g., $T = 1000$ ), this implies that the tuned error is quite small. Therefore, we can ignore it and still consider this as a beta distribution. Finally, the corresponding training loss can be written as", + "bbox": [ + 212, + 463, + 787, + 537 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\epsilon_ {\\theta}\\right) := \\mathbb {E} _ {t \\sim \\operatorname {B e t a} (\\alpha , \\beta), \\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0}), \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\left[ \\| \\epsilon_ {\\theta} \\left(\\alpha_ {t} \\mathbf {x} _ {0} + \\sigma_ {t} \\epsilon , t\\right) - \\epsilon \\| _ {2} ^ {2} \\right]. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 250, + 542, + 787, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we present a comprehensive outline of the B-TTDM method in Algorithm (1). The key difference between B-TTDM and DDPM is the strategy to timestep sampling in the training stage. Thus, B-TTDM is adaptable to most fast sampling methods and improved training methods designed for DDPM without the need for specific modifications.", + "bbox": [ + 212, + 573, + 787, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Algorithm 1 Training of B-TTDM", + "text_level": 1, + "bbox": [ + 217, + 675, + 478, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Require: Beta distribution parameter $\\alpha$ and $\\beta$ . Total diffusion timestep $T$ .", + "bbox": [ + 215, + 694, + 720, + 708 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Generate a beta distribution with parameters $\\alpha$ and $\\beta$ .", + "2: Tuned $f(0; \\alpha, \\beta) = f\\left(\\frac{1}{T}; \\alpha, \\beta\\right)$", + "3: repeat", + "4: $x_0\\sim q(x_0)$", + "5: $t\\sim \\mathrm{Beta}(\\alpha ,\\beta),\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$", + "6: Take a gradient descent step on $\\nabla_{\\pmb{\\theta}}\\| \\pmb {\\epsilon} - \\pmb {\\epsilon}_{\\pmb{\\theta}}(\\pmb {x}_t,t)\\|^2$", + "7: until converged" + ], + "bbox": [ + 222, + 708, + 607, + 805 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/263b89ac5d26157f296b9bc30c9d92b3632d15dafa117c711debc073a7dd3b6c.jpg", + "image_caption": [ + "(a) Comparison of FID on low-resolution." + ], + "image_footnote": [], + "bbox": [ + 241, + 152, + 483, + 287 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a95b254369b24c330635592f984919eed9eef2932b8ccde8db896b6f65c4d5a9.jpg", + "image_caption": [ + "(b) Comparison of FID on high-resolution.", + "Fig. 3: Quantitative Comparison. Comparing FID of different resolution datasets." + ], + "image_footnote": [], + "bbox": [ + 516, + 151, + 759, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 364, + 364, + 382 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we first provide an implementation details of the experimental setup in section 4.1. Then, we provide quantitative comparison of B-TTDM with other state-of-the-art methods in section 4.2. In Section 4.3, we assess the robustness of B-TTDM to different fast sampling algorithms and its training efficiency in comparison to DDPM. Subsequent ablation studies are discussed in Section 4.4. Lastly, qualitative comparisons and visualization results are showcased in Section 4.5.", + "bbox": [ + 212, + 396, + 784, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 215, + 523, + 411, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. All of our experiments utilize the same settings of DDPM [22]. Meanwhile we employ the same model architecture as ADM [9] since it is more advanced compared to DDPM baseline. Throughout the training stage, we maintain $T = 1,000$ for all experiments. We conduct comparative experiments using the CIFAR-10 $(32 \\times 32)$ , ImageNet $(32 \\times 32)$ , CelebA $(64 \\times 64)$ [33], FFHQ $(128 \\times 128)$ [26], AFHQ-D $(256 \\times 256)$ [8] and CelebAHQ $(256 \\times 256)$ [24] datasets for unconditional image generation. To evaluate the sampling quality, we use the Frechet inception distance (FID) [20], sFID [36], improved Precision and Recall [29] as our evaluation metrics. To be more specific, we generate 50K samples for CIFAR-10, ImageNet, CelebA, FFHQ and 10K samples for AFHQ-D, CelebAHQ. In the inference stage, we use 1000 steps for CIFAR-10 and ImageNet, 300 steps for CelebA and FFHQ, and 100 steps for AFHQ-D and CelebAHQ. More details can be found in the Appendix Section 3.", + "bbox": [ + 212, + 550, + 787, + 746 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Quantitative Comparison", + "text_level": 1, + "bbox": [ + 215, + 768, + 470, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison to Baseline. We compare our method with DDPM baseline using various benchmark datasets. The results of the experiments are shown in Figure 3, where we find that the FID of our method significantly outperforms the DDPM", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Beta-Tuned Timestep Diffusion Model", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "on datasets of different resolutions. For instance, on the FFHQ dataset, B-TTDM achieved a FID improvement of 3.79. This underscores our method's efficacy and suggests that the uniform time sampling approach previously employed for training diffusion models is not the most efficient. Notably, B-TTDM reaches an FID of 2.87 on the ImageNet dataset with a larger data size, significantly enhancing DDPM's performance and highlighting B-TTDM's effectiveness.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f07fbcfb907a367ab89f71713732433570a62a7ddf341da9ea36559a08bb2a4a.jpg", + "table_caption": [ + "Table 1: Quantitative Comparison. Comparison of FID and sFID with other state-of-the-art methods on different resolution datasets." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR 32×32FFHQ 128×128AFHQ-D 256×256CelebAHQ 256×256
FIDsFIDFIDsFIDFIDsFIDFIDsFID
DDPM3.014.6914.0712.8917.2149.0316.3225.20
Debias3.174.5611.8611.6515.9146.7714.2321.98
Min-SNR3.234.8012.4612.8517.0347.9414.3724.87
P2-Weight3.194.7811.9612.7416.1847.8814.3822.14
B-TTDM2.894.2210.2811.3813.8145.2513.6221.93
", + "bbox": [ + 267, + 303, + 736, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison to other methods. Since the B-TTDM method is a non-uniform treatment of different time intervals, we contrast it with other start-of-the-art methods that employ non-uniformly re-weighted loss functions in the training stage. The result shown in Table 1 indicates that our B-TTDM perform best across multiple different resolution datasets. In addition, we find that most of these re-weighted loss methods have negative effect on low-resolution CIFAR datasets. Despite this, B-TTDM consistently achieves the best performance, underscoring its broad applicability and generalizability.", + "bbox": [ + 212, + 429, + 787, + 549 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Plug-and-Enhance. Our B-TTDM is designed to align the time sampling and the non-uniform variation observed in the data distribution during forward diffusion, setting it apart from other techniques aimed at enhancing diffusion model performance. Therefore, our method is complementary and can be integrated as a plug-in enhancement within existing strategies for optimizing the training of diffusion models. We combine our B-TTDM with other different improved training methods during the training stage following the same setting. The results are shown in Table 2, we find that our B-TTDM is beneficial for other improved methods like P2-Weight, Min-SNR and DDPM-IP. Notably, the integration of B-TTDM with DDPM-IP leads to a significant achievement, such as an FID score of 2.46 for the CIFAR dataset and a 6.69 FID score for FFHQ.", + "bbox": [ + 212, + 550, + 787, + 717 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Fast Sampling and Training", + "text_level": 1, + "bbox": [ + 214, + 739, + 493, + 755 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fast Sampling. While diffusion models produce high-quality samples, their slow inference speed constrains their applicability in real-world settings. To improve this, various methods like DDIM [44], Time Respacing [9], DPM-Slover [34] and Exposure Bias [38] have been introduced. Fortunately, our B-TTDM method maintains the reliance of these acceleration techniques on the training process,", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/420a25da1e00a6f8a2a625f7c93c24096b18190acb0e4fe2266b17d0441839c5.jpg", + "table_caption": [ + "Table 2: Quantitative Comparison. Comparison of FID and sFID when combining B-TTDM with other improved training methods." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR FID sFIDFFHQ FID sFIDMethodAFHQ-D FID sFIDMethodAFHQ-D FID sFID
DDPM-IP2.724.547.078.35Min-SNR17.03
+ B-TTDM2.463.956.698.10+ B-TTDM13.34
46.13
", + "bbox": [ + 215, + 184, + 785, + 247 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "allowing for their straightforward integration into our approach. The fast sampling outcomes of B-TTDM and DDPM models are contrasted in Table 3. We find that the B-TTDM method significantly outperforms the DDPM method under each different fast samplers and different NFE. Thus, B-TTDM can be seamlessly applied with various fast samplers for DDPM without requiring specific tuning. Fast Training. In addition to yielding higher-quality samples, B-TTDM also accelerates the training process of diffusion models. In Figure 4, we present a comparison between B-TTDM and DDPM in terms of performance and the training iterations on the large scale ImageNet and high resolution AFHQ-D datasets. Specifically, for ImageNet, B-TTDM achieves a $4.71 \\times$ acceleration in training speed to attain equivalent FID scores. Once training converges, B-TTDM's FID score significantly exceeds that of DDPM. Similarly, on the large resolution AFHQ-D dataset, B-TTDM also achieves a $3.33 \\times$ training speed-up and perform better than DDPM upon convergence. We attribute this acceleration to our beta-tuned timestep sampling, which aligns the timestep sampling distribution with the properties of the diffusion process. This alignment not only speeds up the training of diffusion models but also improves the generated samples quality.", + "bbox": [ + 212, + 275, + 787, + 532 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bf6bea86414e64997251de612b1d257549ed0fc9e32116358ef068952b454553.jpg", + "table_caption": [ + "Table 3: Quantitative comparison. Comparison of FID score across various Fast Samplers on AFHQ-D dataset. NFE means the number of function evaluations." + ], + "table_footnote": [], + "table_body": "
Fast SamplersEpsilonScalingDDIMRespacingDPM-Solver
NFE2550255025502550
MethodDDPM23.6318.5024.2218.2523.6419.8318.4917.49
B-TTDM20.8115.3421.7616.7721.7916.6315.8214.79
", + "bbox": [ + 284, + 597, + 718, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 710, + 388, + 726 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Parameters of the beta distribution. The B-TTDM utilizes the Beta-tuned distribution to devise a novel time sampling strategy. This distribution is defined by two parameters, $\\alpha$ and $\\beta$ , where $\\alpha$ predominantly influences the left side of the distribution's shape. As $\\alpha$ increases, the bulk of the probability density function shifts towards the right. Conversely, $\\beta$ primarily affects the right side of the distribution's shape. To further explore their impact on diffusion model training, we experiment with various $\\alpha$ and $\\beta$ combinations, with the experimental", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Beta-Tuned Timestep Diffusion Model", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9c00b3831333798d092e5d2909f06ff13a61e9d136722dfaff608b9f537c5a87.jpg", + "image_caption": [ + "(a) ImageNet $(32\\times 32)$" + ], + "image_footnote": [], + "bbox": [ + 245, + 146, + 480, + 284 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3094ff394f11932ea854c19fd44a35d41db6e0e3a0a09668f7fe10c397be7b8e.jpg", + "image_caption": [ + "(b) AFHQ-D $(256\\times 256)$", + "Fig. 4: FID scores concerning the number of training iterations on ImageNet $(32 \\times 32)$ and AFHQ-D $(256 \\times 256)$ dataset." + ], + "image_footnote": [], + "bbox": [ + 521, + 145, + 756, + 284 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "outcomes presented in Table 4. We find that beta-tuned timestep sampling leads to significant improvements in all metrics and achieves the best performance at $\\alpha = 0.8, \\beta = 1$ . This also implies that beta-tuned timestep sampling is beneficial for generating quality improvements in diffusion models. More experimental results of different parameters can be found in the Appendix Section 3.", + "bbox": [ + 212, + 372, + 787, + 448 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Noise schedules. To ensure the robustness of the B-TTDM method to different noises schedules, we try two different commonly used noise schedules, cosine and linear noise schedules [9] on the CIFAR10 dataset. We follow the previous setting and each FID and sFID score is computed using $T' = 1000$ inference steps. The results of the different noise schedules experiments are presented in Table 5. The B-TTDM is significantly better than the Baseline in both FID and sFID scores, which indicates that the B-TTDM method is robust to different noise schedules.", + "bbox": [ + 212, + 449, + 789, + 555 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ce3fccab5d676d38d5777fa9bc1c3c7dcacaaaf33280205dde4618618c2c2c9e.jpg", + "table_caption": [ + "Table 4: Ablation study on the parameters of Beta Distribution." + ], + "table_footnote": [], + "table_body": "
Beta DistributionAFHQ-D (256×256)
αβFIDsFIDRecallPrecision
1117.2149.030.5530.738
0.95114.1246.460.5620.805
0.90113.8545.470.6120.752
0.80113.8145.250.6120.819
0.70114.7847.560.5640.817
", + "bbox": [ + 248, + 621, + 500, + 700 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5c9dc9dcbbfcf9eb20380241abb2c45bd3df315b4158b1651123205acc6b023c.jpg", + "table_caption": [ + "Table 5: Ablation study on the Noise Schedule on CIFAR." + ], + "table_footnote": [], + "table_body": "
NoiseScheduleDDPMB-TTDM
LinearFID3.533.28
sFID4.784.41
CosineFID3.012.89
sFID4.694.22
", + "bbox": [ + 558, + 625, + 763, + 696 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Other timestep adjust method. In this section, we compare various methods for adjusting the timestep on CIFAR and AFHQ datasets. For instance, the ANT-UW [17] method employs Uncertainty Weighting for different timestep based on multi-task learning with varying variances. Additionally, we compare timestep sampling methods based on alternative non-uniform distributions, such as exponential distributions (DDPM-E) and those where $\\beta > 1$ and $\\alpha = 1$ (B-TTDM $^{\\alpha}$ ). The experimental results are presented in Table 6. It is evident that", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b652c6cc019836197a328ba4d2e676722032fc17e0d1b9e4d9193dbbd17538cf.jpg", + "table_caption": [ + "Table 6: Quantitative comparison. Ablation study on different non-uniform timestep sampling method in the training stage of diffusion model." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR (32×32)AFHQ (256×256)
FIDsFIDRecallPrecisionFIDsFIDRecallPrecision
DDPM3.014.690.6000.68217.2149.030.5530.738
DDPM-E2.954.660.5890.69515.1248.380.5090.703
ANT-UW2.934.450.5960.68714.9647.020.5980.794
B-TTDMα2.984.530.6000.69215.0046.980.6020.768
B-TTDM2.894.220.6020.69513.8145.250.6120.819
", + "bbox": [ + 272, + 184, + 733, + 282 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "all these enhanced methods exhibit some degree of improvement in generation quality compare to DDPM. However, the B-TTDM method demonstrates the most significant enhancement in all evaluation metrics, underscoring its effectiveness.", + "bbox": [ + 212, + 311, + 784, + 357 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3f02b6a6fcc46ef9beab80dd8088504d8725ebef752a5cfaa491a008947f24e7.jpg", + "image_caption": [ + "Longer Training", + "Fig. 5: Qualitative comparison. The generation results from DDPM and B-TTDM on FFHQ dataset. Images in each column are sampled from 50K, 100K, 300K and 500K training iterations. B-TTDM generates significantly better quality than DDPM with the same number of training iterations." + ], + "image_footnote": [], + "bbox": [ + 246, + 400, + 754, + 496 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Qualitative Comparison", + "text_level": 1, + "bbox": [ + 214, + 628, + 457, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generation results. We present a comparison of the generation results between various DDPMs and B-TTDMs in Figure 5. As depicted, images generated by B-TTDM exhibit markedly superior detail compared to those generated by DDPM. Furthermore, B-TTDMs manage to produce clear images by the 100K training iterations, whereas the quality of DDPM generated images at this stage remains relatively poor. Furthermore, we find that after 500K training iterations, B-TTDM still have better generation quality than DDPM in details.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Reconstruct results. Building upon previous analyses, which suggest that the distribution exhibits greater variability in the initial stages of diffusion, we compare the generation capabilities of B-TTDM and DDPM for this significant difference. To achieve this, we input identical noise-containing samples to models trained for the same iterations and reconstructed the outputs. The results in", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Beta-Tuned Timestep Diffusion Model", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fc446349999942ea5161653ad6aaa1cdde305a09055731a4375e05ea82b116cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 145, + 330, + 202 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7b585033f94b8a86bef404694d364aea1ba29e54d0bdb9801ab1e372040e8fd7.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 256, + 203, + 330, + 260 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/dac7b92ff7f2d854c06f98b88cf13ed1f2ea7a84b65586f2b7570dbfbfdb3e8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 146, + 406, + 202 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/daa05f69e53819155fd131d7058b436b5eecffd3afb02c78e3c8e5fe3200610c.jpg", + "image_caption": [ + "DDPM" + ], + "image_footnote": [], + "bbox": [ + 333, + 203, + 408, + 260 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9e2bb0d32fbedb681b7d9aed33e95eb1c5bef4129f0a316ace40a954df8fcaee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 146, + 485, + 202 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0ff623cf41352abca1a0cbe19f516a6dcc1e6bfee19a8085e1c03a9220c47839.jpg", + "image_caption": [ + "B-TTDM", + "Fig. 6: Qualitative comparison. Reconstruction results based on the same noisy inputs on FFHQ dataset." + ], + "image_footnote": [], + "bbox": [ + 411, + 203, + 485, + 260 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4748509463d6cde1ec5aa5eb8a3735813b9586c1e45e58abbd4278d168f8cd0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 146, + 591, + 202 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5102a8bbdd1d8fea7303520407baa8c230b3dab391942e4663c1d1bac0059f53.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 517, + 203, + 589, + 260 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/871fad37211a72a1a924c03dcbc5ccd65c4f860248a334d20058640af0dc0640.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 146, + 671, + 202 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/55f0a861aac0a3c4bb8de2a6decc57b3d22f729683bebac2b8ba7fd470876a1c.jpg", + "image_caption": [ + "DDPM" + ], + "image_footnote": [], + "bbox": [ + 598, + 203, + 671, + 260 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/89b049e5525748b55a3f7680a4e40d95baeef9dd18c379ccb1a7fa446ee57cdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 146, + 751, + 202 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2ae3c5d992445768451eba9bf81b00a14de5462dc651f946413c0a6eac592903.jpg", + "image_caption": [ + "B-TTDM" + ], + "image_footnote": [], + "bbox": [ + 678, + 203, + 750, + 260 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Figure 6 indicate that the reconstructions generated by B-TTDM closely resemble the original samples, whereas those generated by DDPM exhibit more differences. Unconditional generation results. Figure 7 provides a qualitative comparison of the unconditional image generation capabilities of B-TTDM and DDPM across the different datasets. The Figure 7 reveals that samples generated by DDPM often exhibit noticeable color biases, whereas samples generated by B-TTDM are free from such biases and show a marked improvement in quality compared to DDPM. This underscores the effectiveness of B-TTDM. More unconditional generation results can be found in the Appendix Section 4.", + "bbox": [ + 212, + 338, + 787, + 474 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d47ffde84d45c1552ab68a2f17eed8a533b2c9cc80dbdc8fbfbf00ae01006744.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 502, + 424, + 611 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/20ec56947c0b59e9b9885edb1b78ec3bbaa5a30ec98a4deca63499e59866d972.jpg", + "image_caption": [ + "Fig. 7: Qualitative comparison. Unconditional generation results." + ], + "image_footnote": [], + "bbox": [ + 437, + 502, + 633, + 612 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a42300ea8f58c508ea46d71ab7a027a65411765abe966156e0e8db30e22c1672.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 502, + 774, + 612 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 690, + 359, + 707 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, through both experimental exploration and theoretical analysis, we conclude that the distribution varies non-uniformly in forward diffusion. Moreover, the most drastic variations occur in the initial stages of the forward diffusion process. These insights suggest that the uniform timestep sampling is sub-optimal for training diffusion models. Building on these findings, we design the B-TTDM method, which aligns timestep sampling with these inherent properties. Our extensive experiments demonstrate that B-TTDM significantly enhances the quality of model generation while also accelerating the training process.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by NSFC under Grant 61927809 and in part by STCSM under Grant 22DZ2229005.", + "bbox": [ + 212, + 174, + 787, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 224, + 323, + 242 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anderson, B.D.: Reverse-time diffusion equation models. Stochastic Processes and their Applications 12(3), 313-326 (1982)", + "2. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR. pp. 22669-22679. IEEE (2023)", + "3. Block, A., Mroueh, Y., Rakhlin, A., Ross, J.: Fast mixing of multi-scale Langevin dynamics under the manifold hypothesis. CoRR abs/2006.11166 (2020)", + "4. Cai, R., Song, Z., Guan, D., Chen, Z., Luo, X., Yi, C., Kot, A.: Benchlmm: Benchmarking cross-style visual capability of large multimodal models. ECCV (2024)", + "5. Chen, Z., Li, B., Wu, S., Jiang, K., Ding, S., Zhang, W.: Content-based unrestricted adversarial attack. NeurIPS 36 (2024)", + "6. Chen, Z., Li, B., Xu, J., Wu, S., Ding, S., Zhang, W.: Towards practical certifiable patch defense with vision transformer. In: CVPR. pp. 15148-15158 (2022)", + "7. Choi, J., Lee, J., Shin, C., Kim, S., Kim, H., Yoon, S.: Perception prioritized training of diffusion models. In: CVPR. pp. 11462-11471. IEEE (2022)", + "8. Choi, Y., Uh, Y., Yoo, J., Ha, J.: Stargan v2: Diverse image synthesis for multiple domains. In: CVPR. pp. 8185-8194. Computer Vision Foundation / IEEE (2020)", + "9. Dhariwal, P., Nichol, A.Q.: Diffusion models beat gans on image synthesis. In: NeurIPS. pp. 8780-8794 (2021)", + "10. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real NVP. In: ICLR. OpenReview.net (2017)", + "1. Fan, K., Tang, J., Cao, W., Yi, R., Li, M., Gong, J., Zhang, J., Wang, Y., Wang, C., Ma, L.: Freemotion: A unified framework for number-free text-to-motion synthesis ECCV (2024)", + "2. Fang, B., Li, B., Wu, S., Yi, R., Ding, S., Ma, L.: Re-thinking data availability attacks against deep neural networks. CVPR (2024)", + "3. Fang, B., Li, B., Wu, S., Zheng, T., Ding, S., Yi, R., Ma, L.: Towards generalizable data protection with transferable unlearnable examples. CoRR abs/2305.11191 (2023)", + "4. Ge, X., Liu, X., Yu, Z., Shi, J., Qi, C., Li, J., Kalviainen, H.: Diffas: Face anti-spoofing via generative diffusion models. In: ECCV (2024)", + "5. Geng, C., Han, T., Jiang, P.T., Zhang, H., Chen, J., Hauberg, S., Li, B.: Improving adversarial energy-based model via diffusion process. ICML (2024)", + "6. Geng, C., Wang, J., Gao, Z., Frellsen, J., Hauberg, S.: Bounds all around: training energy-based models with bidirectional bounds. NeurIPS (2021)", + "7. Go, H., Lee, Y., Lee, S., Oh, S., Moon, H., Choi, S.: Addressing negative transfer in diffusion models. NeurIPS 36 (2024)", + "8. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.C., Bengio, Y.: Generative adversarial nets. In: NIPS. pp. 2672–2680 (2014)", + "9. Hang, T., Gu, S., Li, C., Bao, J., Chen, D., Hu, H., Geng, X., Guo, B.: Efficient diffusion training via min-snr weighting strategy. In: ICCV. pp. 7407-7417. IEEE (2023)" + ], + "bbox": [ + 225, + 253, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Beta-Tuned Timestep Diffusion Model", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS. pp. 6626-6637 (2017)", + "21. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A.A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models. CoRR abs/2210.02303 (2022)", + "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "23. Hyvarinen, A.: Estimation of non-normalized statistical models by score matching. J. Mach. Learn. Res. 6, 695-709 (2005)", + "24. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. CoRR abs/1710.10196 (2017)", + "25. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. In: NeurIPS (2022)", + "26. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. PAMI 43(12), 4217-4228 (2021)", + "27. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. In: NeurIPS. pp. 10236-10245 (2018)", + "28. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)", + "29. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. In: NeurIPS. pp. 3929-3938 (2019)", + "30. Lee, H., Lu, J., Tan, Y.: Convergence for score-based generative modeling with polynomial complexity. In: NeurIPS (2022)", + "31. Li, X., Thickstun, J., Gulrajani, I., Liang, P., Hashimoto, T.B.: Diffusion-lm improves controllable text generation. In: NeurIPS (2022)", + "32. Liu, Y., Chen, Y., Dai, W., Gou, M., Huang, C.T., Xiong, H.: Source-free domain adaptation with domain generalized pretraining for face anti-spoofing. PAMI (2024)", + "33. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: ICCV. pp. 3730-3738. IEEE Computer Society (2015)", + "34. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In: NeurIPS (2022)", + "35. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)", + "36. Nash, C., Menick, J., Dieleman, S., Battaglia, P.W.: Generating images with sparse representations. In: ICML. vol. 139, pp. 7958-7968. PMLR (2021)", + "37. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. vol. 139, pp. 8162-8171. PMLR (2021)", + "38. Ning, M., Li, M., Su, J., Salah, A.A., Ertugrul, I.O.: Elucidating the exposure bias in diffusion models. ICLR (2024)", + "39. Ning, M., Sangineto, E., Porrello, A., Calderara, S., Cucchiara, R.: Input perturbation reduces exposure bias in diffusion models. In: ICML (2023)", + "40. Parmar, G., Singh, K.K., Zhang, R., Li, Y., Lu, J., Zhu, J.: Zero-shot image-to-image translation. In: SIGGRAPH (Conference Paper Track). pp. 11:1-11:11. ACM (2023)", + "41. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023)", + "42. Risken, H., Risken, H.: Fokker-planck equation. Springer (1996)", + "43. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10674-10685. IEEE (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Zheng et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR. OpenReview.net (2021)", + "45. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS. pp. 11895-11907 (2019)", + "46. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR. OpenReview.net (2021)", + "47. Tao, S., Wang, J.: Alleviation of gradient exploding in gans: Fake can be real. In: CVPR. pp. 1188-1197. Computer Vision Foundation / IEEE (2020)", + "48. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., Cheng, F.: Eliminating lipschitz singularities in diffusion models. ICLR (2024)", + "49. Yu, H., Shen, L., Huang, J., Zhou, M., Li, H., Zhao, F.: Debias the training of diffusion models. CoRR abs/2310.08442 (2023)", + "50. Zeng, W., Yan, Y., Zhu, Q., Chen, Z., Chu, P., Zhao, W., Yang, X.: Infusion: Preventing customized text-to-image diffusion from overfitting (2024)", + "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "52. Zhu, Z., Wei, Y., Wang, J., Gan, Z., Zhang, Z., Wang, L., Hua, G., Wang, L., Liu, Z., Hu, H.: Exploring discrete diffusion models for image captioning. CoRR abs/2211.11694 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 438 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Beta-Tuned Timestep Diffusion Model", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Beta-Tuned Timestep Diffusion Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_model.json b/2024/Beta-Tuned Timestep Diffusion Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99bbd639aa455c337fe247aad617358d29628aff --- /dev/null +++ b/2024/Beta-Tuned Timestep Diffusion Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_model.json @@ -0,0 +1,2786 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.283, + 0.141, + 0.723, + 0.163 + ], + "angle": 0, + "content": "Beta-Tuned Timestep Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.189, + 0.723, + 0.221 + ], + "angle": 0, + "content": "Tianyi Zheng\\(^{1,2}\\), Peng-Tao Jiang\\(^{2}\\), Ben Wan\\(^{1}\\), Hao Zhang\\(^{2}\\), Jinwei Chen\\(^{2}\\), Jia Wang\\(^{1}\\) \\(\\boxtimes\\), and Bo Li\\(^{2}\\) \\(\\boxtimes\\)" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.231, + 0.68, + 0.291 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Jiao Tong University, Shanghai, China {tyzheng, burn-w, jiawang}@sjtu.edu.cn \\(^{2}\\) vivo Mobile Communication Co., Ltd, China {pt.jiang,haozhang,jinwei.chen,libra}@vivo.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.341, + 0.744, + 0.55 + ], + "angle": 0, + "content": "Abstract. Diffusion models have received a lot of attention in the field of generation due to their ability to produce high-quality samples. However, several recent studies indicate that treating all distributions equally in diffusion model training is sub-optimal. In this paper, we conduct an in-depth theoretical analysis of the forward process of diffusion models. Our findings reveal that the distribution variations are non-uniform throughout the diffusion process and the most drastic variations in distribution occur in the initial stages. Consequently, simple uniform timestep sampling strategy fail to align with these properties, potentially leading to suboptimal training of diffusion models. To address this, we propose the Beta-Tuned Timestep Diffusion Model (B-TTDM), which devises a timestep sampling strategy based on the beta distribution. By choosing the correct parameters, B-TTDM aligns the timestep sampling distribution with the properties of the forward diffusion process. Extensive experiments on different benchmark datasets validate the effectiveness of B-TTDM." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.564, + 0.734, + 0.578 + ], + "angle": 0, + "content": "Keywords: Generative Models \\(\\cdot\\) Diffusion Models \\(\\cdot\\) Beta Distribution" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.377, + 0.62 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Diffusion models have gained widespread application across various domains in recent years, such as text-to-image generation [5,43,50], text generation [31,52], video synthesis [21] and AI security [6,12-14,32], primarily due to their ability to produce high-quality samples. During the training stage, diffusion models introduce varying levels of Gaussian noise into the data distribution, creating a range of altered distributions. The model is then trained by minimizing denoising score-matching losses across these diverse distributions. In the inference stage, the trained model reverses this noise addition process to generate new samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.803 + ], + "angle": 0, + "content": "While DDPM [22], iDDPM [37] and ADM [9] have achieved significant success in image generation tasks, several studies [7, 17, 19, 48] identify that treating the distribution equally across different timesteps in the training stage could be" + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.81, + 0.644, + 0.84 + ], + "angle": 0, + "content": "\\(\\boxtimes\\) Corresponding authors. This work was done during Tianyi Zheng's internship at vivo." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "T. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.405 + ], + "angle": 0, + "content": "sub-optimal. For instance, the P2-Weight [7] method discover through empirical analysis that distributions at various timesteps in forward diffusion encapsulate distinct semantic information. Building on this insight, P2-Weight propose a re-weighted loss function to elevate the quality of generated samples. Meanwhile, Min-SNR [19] and ANT [17] find inconsistencies in optimization gradients across various timesteps and recommend a re-weighted loss strategy to mitigate these issues. Recently, E-TSDM [48] uncover that the issue of numerical instability caused by high Lipschitz constants near the zero point. Therefore, E-TSDM assigns a weight of 0 to certain timesteps near zero to mitigate the issue of numerical instability. Despite the improvement in sample quality achieved by these methods through empirical adjustments to the training objective's weight, they still utilize DDPM's uniform timestep sampling strategy. As another influencing factor, the choice of timestep sampling strategy is also closely related to the training objective but has been overlooked in previous work. In this paper, we conduct a theoretical analysis of DDPM's forward diffusion process and design a novel strategy for timestep sampling, aiming to address this overlooked aspect and further improve the quality of generated samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.79, + 0.543 + ], + "angle": 0, + "content": "Our theoretical analysis in Theorem 2 demonstrates that distribution variations throughout the diffusion process are non-uniform, with the most significant variations occurring at the initial stages. This is experimentally supported by Figure 1, which shows non-uniform variations in both Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) [51] metrics, especially pronounced in the initial stages. Both our theoretical insights from Theorem 2 and the experimental evidence in Figure 1 underscore the non-uniform properties of forward diffusion, indicating that a uniform timestep sampling strategy may not be the most effective approach in the training stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.79, + 0.759 + ], + "angle": 0, + "content": "Building on these findings, we propose a novel timestep sampling strategy that utilizes the beta distribution, named the Beta-Tuned Timestep Diffusion Model (B-TTDM). By choosing the correct parameters of the beta distribution, B-TTDM aligns the timestep sampling distribution with the observed properties of the diffusion process and is specifically tuned to avoid singularities at \\( t = 0 \\) timesteps. Extensive experiments on various datasets demonstrate that B-TTDM not only improves the quality of the generated samples but also expedites the training process. For instance, B-TTDM achieves similar FID scores on the ImageNet dataset with about \\( 20\\% \\) of the training iterations needed by the original approach and significantly improves generation quality after convergence. Moreover, B-TTDM can be compatible with other advanced diffusion model training methods, such as P2-Weight [7], Min-SNR [19], Debias [19], and DDPM-IP [39]. Through integration with these methods, B-TTDM can combine different advantages and generate higher quality samples. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "- Through theoretical and experimental verification, we confirm that distribution variations at different timesteps during forward diffusion are non-uniform, with the most significant shifts occurring in the initial stages. This indicates that the uniform time sampling in training is sub-optimal." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beta-Tuned Timestep Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "- We design a novel timestep sampling strategy leveraging the Beta distribution to align with the non-uniform distribution shifts observed during forward diffusion. Furthermore, our B-TTDM is compatible with other enhancement methods, allowing for flexible integration into existing improved methods." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.208, + 0.785, + 0.252 + ], + "angle": 0, + "content": "- Extensive experiments on different benchmark datasets demonstrate that the B-TTDM method significantly improves generation quality and drastically speeds up training of the diffusion models." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.288, + 0.492, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.45, + 0.488, + 0.462 + ], + "angle": 0, + "content": "(a) PSNR at each forward diffusion timestep." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.289, + 0.779, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.45, + 0.777, + 0.462 + ], + "angle": 0, + "content": "(b) LPIPS at each forward diffusion timestep." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.473, + 0.785, + 0.501 + ], + "angle": 0, + "content": "Fig. 1: The PSNR and LPIPS metrics between the distribution at timestep \\( t \\) and the original distribution in forward diffusion under different noise schedules." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.388, + 0.569 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models [15,22,37] are a family of generative models that generate samples from Gaussian noise via a learned denoising process. Recently, diffusion models have outperformed flow models [10,27] and variational autoencoders [16, 28] in generating higher quality samples. Furthermore, the diffusion models also outperform GANs [18,47] in certain cases [9]. This superior performance has led to the widespread adoption of diffusion models across diverse applications. For example, the Diffusion model [43] implements the generation of images based on textual inputs [4,11], image to image translation [35,40] and 3D Avatar generation [41]. Despite their success, the speed of sample generation by diffusion models is unsatisfactory. Therefore, DPM-Solver [34] and DDIM [44] design new generation methods based on the framework of ordinary differential equation (ODE), which have greatly improved the generation speed of the diffusion model. Meanwhile, EDM [25] design a higher order Runge-Kutta sampling method to further improve the generation speed and quality. These methods significantly enhance the speed of sample generation in diffusion models, with a primary emphasis on the generative aspect. Yet, they rely on pre-trained diffusion models without addressing training optimizations. Our proposed method presents an" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "T. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "orthogonal strategy, aiming to improve the training stage and thereby further elevate the generate quality of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Improved Diffusion Models. To further improve the quality of generative samples in the training stage, ADM [9] and U-ViT [2] design new advanced architectures for diffusion models. Based on these advanced architectures, the quality of generation is significantly improved. Additionally, P2-Weight [7], Min-SNR [19] and Debias [49] methods design the re-weight loss function based on SNR in the training stage. These can reduce the bias problem in the diffusion models. To make the neural network smoother, the E-TSDM [48] sets the weight of the loss function corresponding to timesteps close to the zero to 0. Additional, ANT [17] propose multi-task learning methods to set the loss weight of different timestep in the training stage. Although these improved methods can produce better samples, they overlook the non-uniform distribution variations in the diffusion process and continue to use uniform timestep sampling during training. Our theoretical analysis indicates that this uniform time sampling strategy in the training stage is not optimal, suggesting that a more tailored timestep sampling approach could yield significant improvements in generated samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.426, + 0.332, + 0.441 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.456, + 0.364, + 0.472 + ], + "angle": 0, + "content": "3.1 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.789, + 0.526 + ], + "angle": 0, + "content": "Let \\( q(x_0) \\) represent the data distribution over \\( \\mathbb{R}^n \\). During the forward diffusion, a sample \\( x_0 \\) is drawn from \\( q(x_0) \\) and subsequently perturbed using a stochastic differential equation (SDE) [46]." + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.536, + 0.788, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} x _ {t} = f \\left(x _ {t}, t\\right) \\mathrm {d} t + g (t) \\mathrm {d} w. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.79, + 0.625 + ], + "angle": 0, + "content": "The purpose of the forward diffusion is to transform the data distribution into simple normal Gaussian distribution \\(\\mathcal{N}(0, I)\\). Meanwhile, we define \\(q_t(x_t)\\) to represent the distribution at each time \\(t\\) during the forward diffusion process. For any time \\(t\\), the conditional distribution \\(q_t(x_t | x_0)\\) is [9, 22, 37]" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.637, + 0.788, + 0.654 + ], + "angle": 0, + "content": "\\[\nq _ {t} \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t} \\mid \\alpha_ {t} x _ {0}, \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.705 + ], + "angle": 0, + "content": "In Equation (2), \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) satisfy \\(\\alpha_{t}^{2} + \\sigma_{t}^{2} = 1\\). This forward diffusion process is known as the Ornstein-Uhlenbeck (OU) process. For more general expressions, we have the following conditional distribution" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.716, + 0.788, + 0.741 + ], + "angle": 0, + "content": "\\[\nq _ {t} \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t} \\mid e ^ {- \\frac {t}{2}} x _ {0}, (1 - e ^ {- t}) \\mathbf {I}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.775 + ], + "angle": 0, + "content": "The reverse diffusion transforms normal Gaussian distribution into the data distribution \\( q(x_0) \\). This process satisfies the following reverse SDE equation [1]." + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.789, + 0.788, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} x _ {t} = \\left(f \\left(x _ {t}, t\\right) - g (t) ^ {2} \\nabla_ {x} \\log q _ {t} \\left(x _ {t}\\right)\\right) \\mathrm {d} t + g (t) \\mathrm {d} w. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Specially, \\(\\nabla_x\\log q_t(x_t)\\) is the score function of the probability density \\(q_{t}(x_{t})\\) Motivated by this correspondence, diffusion models train a neural network to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beta-Tuned Timestep Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "estimate the \\(-\\sigma_t\\nabla_x\\log q_t(x_t)\\). In the inference stage, we can simulate Equation (4) for sampling and generate new samples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.201, + 0.614, + 0.215 + ], + "angle": 0, + "content": "3.2 Non-uniform variation in forward diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.784, + 0.288 + ], + "angle": 0, + "content": "We first start by introducing and discussing our main assumptions. The only assumption we consider is that in the vicinity of high-density data, the score of the data distribution is bounded. Based on the definition of score [23, 45, 46], this assumption is quite common used [3, 30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.785, + 0.346 + ], + "angle": 0, + "content": "Assumption 1 For the diffusion process described by Equation (3), assume there exists a constant \\(\\delta >0\\). For any \\(t\\in [0,T]\\) and any point \\(x_{t}\\) in high-density regions, the score \\(\\| \\nabla_x\\log q_t(x_t)\\|\\) is bounded by \\(\\delta\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.357, + 0.573, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\nabla_ {x} \\log q _ {t} (x _ {t}) \\right\\| \\leq \\delta .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.784, + 0.416 + ], + "angle": 0, + "content": "This ensures that the score of the data distribution is bounded in the vicinity of high-density data throughout the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.784, + 0.47 + ], + "angle": 0, + "content": "Next, we examine the Fokker-Planck equation [42] corresponding to the Ornstein-Uhlenbeck process, as depicted in Equation (5). The parameter \\(\\theta\\) is drift coefficient and the parameter \\(D\\) is diffusion coefficient." + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.483, + 0.787, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial}{\\partial t} q \\left(x _ {t}\\right) = \\theta \\nabla_ {x _ {t}} q \\left(x _ {t}\\right) + D \\nabla_ {x _ {t}} ^ {2} q \\left(x _ {t}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.784, + 0.563 + ], + "angle": 0, + "content": "Since the \\(\\theta\\) and \\(D\\) are both constants independent of \\(x_{t}\\), to simplify the analysis that follows, we can let both of them be 1. Therefore, we have the following simplified Fokker-Planck equation." + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.575, + 0.787, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial}{\\partial t} q \\left(x _ {t}\\right) = \\nabla_ {x _ {t}} q \\left(x _ {t}\\right) + \\nabla_ {x _ {t}} ^ {2} q \\left(x _ {t}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.784, + 0.701 + ], + "angle": 0, + "content": "The Fokker-Planck equation (6) sheds light on how a distribution's evolution is influenced by both time and spatial variables in the forward diffusion stage of DDPM. Our analysis further reveals that in the initial stages of diffusion, the changes in the distribution are primarily driven by the second-order derivative term, indicating a significant diffusion effect at this stage. Therefore, we have the following Lemma:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.784, + 0.759 + ], + "angle": 0, + "content": "Lemma 1. Consider the forward diffusion process as described by Equation (3) When \\(t\\to 0^+\\), the time derivative of the distribution \\(q_{t}(x_{t})\\) is predominantly governed by the second-order derivative term, encapsulated in the relationship:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.77, + 0.787, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right) = \\nabla_ {x _ {t}} ^ {2} q \\left(x _ {t}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "where \\(\\nabla_{x_t}^2\\) denotes the Laplacian operator with respect to \\(x_{t}\\), indicating the diffusion term's dominance in the evolution of \\(q_{t}(x_{t})\\) when \\(t\\to 0^{+}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "T. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "The detailed proof of the Lemma (1) is shown in Appendix Section 2.2. Lemma (1) suggests that the alteration in the distribution primarily stems from the \\(\\nabla_{x_t}^2 q(x_t)\\). Expanding on this insight, we can establish the following theorem, providing an upper bound on the alteration in distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.263 + ], + "angle": 0, + "content": "Theorem 2. Consider the forward diffusion process described by Equation (7). The norm of the time derivative of the distribution \\( q_{t}(x_{t}) \\) is bounded by the following inequality:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.261, + 0.788, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right) \\right\\| \\leq \\left\\| \\frac {q _ {t} \\left(x _ {t}\\right)}{2 t} \\right\\|. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.787, + 0.332 + ], + "angle": 0, + "content": "As \\( t \\to 0^{+} \\), the norm of the derivative of the distribution \\( q_{t}(x_{t}) \\) may become unbounded. This is indicated by the limit superior:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.341, + 0.788, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\lim _ {t \\rightarrow 0 ^ {+}} \\left\\| \\frac {q _ {t} \\left(x _ {t}\\right)}{2 t} \\right\\| = \\infty . \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.464 + ], + "angle": 0, + "content": "The detailed proof of the Theorem (2) is shown in Appendix Section 2.3. Theorem (2) suggests that the bound on the norm of the time derivative of \\( q_{t}(x_{t}) \\) grows indefinitely as \\( t \\to 0^{+} \\), indicating that the norm of the derivative could be unbounded. We present a straightforward example to illustrate this behavior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.787, + 0.534 + ], + "angle": 0, + "content": "Simple case illustration. Let \\( q_{t}(x_{t}) = \\frac{1}{\\sqrt{4\\pi t}} e^{-\\frac{x^{2}}{4t}} \\) be a solution to differential equation (7). Meanwhile, we notice that \\( \\int_{-\\infty}^{+\\infty} q_{t}(x_{t}) dx = 1 \\) and \\( q_{t}(x) > 0 \\), which means \\( q_{t}(x) \\) is a probability density function. For every \\( t > 0 \\), the norm of the time derivative of \\( q_{t}(x) \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.544, + 0.788, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right) \\right\\| = \\left\\| \\frac {\\sqrt {t} (- 2 t + x ^ {2}) e ^ {- \\frac {x ^ {2}}{4 t}}}{8 \\sqrt {\\pi} t ^ {3}} \\right\\|. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.596, + 0.54, + 0.611 + ], + "angle": 0, + "content": "This expression leads to the conclusion that:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.621, + 0.788, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\lim _ {t \\rightarrow 0 ^ {+}} \\left\\| \\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right)\\right\\| = \\infty , \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.787, + 0.71 + ], + "angle": 0, + "content": "indicating that the norm of the time derivative of \\( q_{t}(x) \\) becomes unbounded as \\( t \\to 0^{+} \\). This result indicates an infinite rate of variation for the distribution \\( q_{t}(x_{t}) \\) when \\( t \\to 0^{+} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.787, + 0.787 + ], + "angle": 0, + "content": "Theorem (2) highlights the non-uniform properties of distribution variations during diffusion, with the initial stages experiencing the most significant shifts. Furthermore, Proposition (1) clarifies that the distribution variations \\( q_{t}(x_{t}) \\) are non-uniform across the diffusion process, and these variations tend to flatten as the process progresses to later stages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Proposition 1. Consider the forward diffusion process in Equation (3). The upper bound of the KL divergence between the distribution \\( q_{t}(x_{t}) \\) and the standard Gaussian distribution converges exponentially. i.e.," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Beta-Tuned Timestep Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.162, + 0.788, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K L} \\left(q _ {t} \\left(x _ {t}\\right) \\| \\mathcal {N} (0, 1)\\right) \\leq e ^ {- t} \\operatorname {K L} \\left(q _ {0} \\left(x _ {0}\\right) \\| \\mathcal {N} (0, 1)\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.785, + 0.288 + ], + "angle": 0, + "content": "Based on the Theorem (2) and Proposition (1), we notice that the variations in the distribution of forward diffusion is non-uniform, with the sharpest change at the beginning of the diffusion and converging to a Gaussian distribution at an exponential rate. These theoretical insights guide the design of our timestep sampling strategy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.309, + 0.561, + 0.325 + ], + "angle": 0, + "content": "3.3 Beta distribution timestep sampling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.785, + 0.41 + ], + "angle": 0, + "content": "To align the non-uniform trend of variation, we introduce a time sampling strategy that utilizes the Beta distribution for training the diffusion model. This strategy leverages the Beta density function's capacity to take on a diverse array of shapes, controlled by the parameters \\(\\alpha\\) and \\(\\beta\\). The probability density function (pdf) of the Beta distribution is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.42, + 0.788, + 0.452 + ], + "angle": 0, + "content": "\\[\nf (t; \\alpha , \\beta) = \\frac {\\Gamma (\\alpha + \\beta)}{\\Gamma (\\alpha) \\Gamma (\\beta)} t ^ {\\alpha - 1} (1 - t) ^ {\\beta - 1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.538 + ], + "angle": 0, + "content": "Where \\(\\Gamma(z)\\) is the gamma function. Setting \\(\\alpha = 1\\) and \\(\\beta = 1\\) simplifies the Beta distribution to a uniform distribution, illustrating the flexibility of Beta distribution. By adjusting the values of \\(\\alpha\\) and \\(\\beta\\), we can tailor the time sampling distribution within the diffusion model training to better align the desired variation properties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Since the beta distribution is a family of continuous probability distributions defined on the interval [0, 1]. We can use \\( \\frac{t}{T} \\) to map the diffusion timestep into the [0, 1] interval of the beta distribution. Then we compare the probability density function of beta distribution for different \\( \\alpha \\) and \\( \\beta \\) in Figure 2. We find that when we set \\( \\alpha > 1 \\) and \\( \\beta > 1 \\), The probability density function is U-shaped. Meanwhile, we notice that \\( \\alpha = 1 \\) and \\( \\beta > 1 \\) and \\( \\alpha < 1 \\) and \\( \\beta = 1 \\), the probability density function conforms to our expectation of a non-uniformly decreasing. However, when \\( \\alpha = 1 \\) and \\( \\beta > 1 \\), the probability density function exhibits a positively skewed shape, lacking alignment with the most significant varies in the initial diffusion stage. We provide ablation study of different \\( \\alpha \\) and \\( \\beta \\) in the section 4.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To align with the previously identified properties, we choose to set \\(\\alpha < 1\\) and \\(\\beta = 1\\). With these parameters, the distribution adopts a reverse J-shaped profile. Consequently, the probability density function under these conditions becomes a strictly decreasing function, exhibiting a sharp change as it approaches 0. This specific shape and behavior of the pdf are instrumental in aligning the rapid initial changes and the convergence properties observed in the forward diffusion process. Therefore, we replace the uniform time sampling in the original diffusion model with this time sampling method based on the beta distribution. In addition, to prevent the probability density function from reaching infinite values as time t approaches 0, we adopt \\(\\frac{1}{T}\\) as the time sampling weight at the initial" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "T. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.147, + 0.685, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.333, + 0.776, + 0.349 + ], + "angle": 0, + "content": "Fig. 2: Probability density function of beta distribution with different parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.785, + 0.406 + ], + "angle": 0, + "content": "moment. Therefore, our method is called Beta-Tuned Timestep Diffusion Model (B-TTDM). Next, we analyse the errors introduced by this tuned approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.459 + ], + "angle": 0, + "content": "Proposition 2. Given a timestep sampling method utilizing the B-TTDM with a Beta distribution parameterized by \\(\\alpha < 1\\) and \\(\\beta = 1\\), the tuned error at the initial time position \\(t\\) is \\(\\left(\\frac{1}{T}\\right)^{\\alpha}\\), where \\(T\\) is the total diffusion steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.539 + ], + "angle": 0, + "content": "The detailed proof of the Proposition (2) is shown in Appendix Section 2.4. Since the value of \\( T \\) is usually large in the training of diffusion models (e.g., \\( T = 1000 \\)), this implies that the tuned error is quite small. Therefore, we can ignore it and still consider this as a beta distribution. Finally, the corresponding training loss can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.251, + 0.543, + 0.788, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\epsilon_ {\\theta}\\right) := \\mathbb {E} _ {t \\sim \\operatorname {B e t a} (\\alpha , \\beta), \\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0}), \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\left[ \\| \\epsilon_ {\\theta} \\left(\\alpha_ {t} \\mathbf {x} _ {0} + \\sigma_ {t} \\epsilon , t\\right) - \\epsilon \\| _ {2} ^ {2} \\right]. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Furthermore, we present a comprehensive outline of the B-TTDM method in Algorithm (1). The key difference between B-TTDM and DDPM is the strategy to timestep sampling in the training stage. Thus, B-TTDM is adaptable to most fast sampling methods and improved training methods designed for DDPM without the need for specific modifications." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.676, + 0.479, + 0.691 + ], + "angle": 0, + "content": "Algorithm 1 Training of B-TTDM" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.695, + 0.722, + 0.709 + ], + "angle": 0, + "content": "Require: Beta distribution parameter \\(\\alpha\\) and \\(\\beta\\). Total diffusion timestep \\(T\\)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.709, + 0.609, + 0.723 + ], + "angle": 0, + "content": "1: Generate a beta distribution with parameters \\(\\alpha\\) and \\(\\beta\\)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.723, + 0.439, + 0.738 + ], + "angle": 0, + "content": "2: Tuned \\( f(0; \\alpha, \\beta) = f\\left(\\frac{1}{T}; \\alpha, \\beta\\right) \\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.738, + 0.295, + 0.75 + ], + "angle": 0, + "content": "3: repeat" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.751, + 0.332, + 0.765 + ], + "angle": 0, + "content": "4: \\(x_0\\sim q(x_0)\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.765, + 0.439, + 0.778 + ], + "angle": 0, + "content": "5: \\(t\\sim \\mathrm{Beta}(\\alpha ,\\beta),\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.778, + 0.603, + 0.793 + ], + "angle": 0, + "content": "6: Take a gradient descent step on \\(\\nabla_{\\pmb{\\theta}}\\| \\pmb {\\epsilon} - \\pmb {\\epsilon}_{\\pmb{\\theta}}(\\pmb {x}_t,t)\\|^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.793, + 0.352, + 0.806 + ], + "angle": 0, + "content": "7: until converged" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.709, + 0.609, + 0.806 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beta-Tuned Timestep Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.153, + 0.485, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.298, + 0.484, + 0.31 + ], + "angle": 0, + "content": "(a) Comparison of FID on low-resolution." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.152, + 0.761, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.298, + 0.762, + 0.31 + ], + "angle": 0, + "content": "(b) Comparison of FID on high-resolution." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.322, + 0.784, + 0.336 + ], + "angle": 0, + "content": "Fig. 3: Quantitative Comparison. Comparing FID of different resolution datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.365, + 0.365, + 0.383 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.785, + 0.501 + ], + "angle": 0, + "content": "In this section, we first provide an implementation details of the experimental setup in section 4.1. Then, we provide quantitative comparison of B-TTDM with other state-of-the-art methods in section 4.2. In Section 4.3, we assess the robustness of B-TTDM to different fast sampling algorithms and its training efficiency in comparison to DDPM. Subsequent ablation studies are discussed in Section 4.4. Lastly, qualitative comparisons and visualization results are showcased in Section 4.5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.525, + 0.412, + 0.541 + ], + "angle": 0, + "content": "4.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Implementation details. All of our experiments utilize the same settings of DDPM [22]. Meanwhile we employ the same model architecture as ADM [9] since it is more advanced compared to DDPM baseline. Throughout the training stage, we maintain \\( T = 1,000 \\) for all experiments. We conduct comparative experiments using the CIFAR-10 \\( (32 \\times 32) \\), ImageNet \\( (32 \\times 32) \\), CelebA \\( (64 \\times 64) \\) [33], FFHQ \\( (128 \\times 128) \\) [26], AFHQ-D \\( (256 \\times 256) \\) [8] and CelebAHQ \\( (256 \\times 256) \\) [24] datasets for unconditional image generation. To evaluate the sampling quality, we use the Frechet inception distance (FID) [20], sFID [36], improved Precision and Recall [29] as our evaluation metrics. To be more specific, we generate 50K samples for CIFAR-10, ImageNet, CelebA, FFHQ and 10K samples for AFHQ-D, CelebAHQ. In the inference stage, we use 1000 steps for CIFAR-10 and ImageNet, 300 steps for CelebA and FFHQ, and 100 steps for AFHQ-D and CelebAHQ. More details can be found in the Appendix Section 3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.471, + 0.785 + ], + "angle": 0, + "content": "4.2 Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Comparison to Baseline. We compare our method with DDPM baseline using various benchmark datasets. The results of the experiments are shown in Figure 3, where we find that the FID of our method significantly outperforms the DDPM" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "T. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "on datasets of different resolutions. For instance, on the FFHQ dataset, B-TTDM achieved a FID improvement of 3.79. This underscores our method's efficacy and suggests that the uniform time sampling approach previously employed for training diffusion models is not the most efficient. Notably, B-TTDM reaches an FID of 2.87 on the ImageNet dataset with a larger data size, significantly enhancing DDPM's performance and highlighting B-TTDM's effectiveness." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.264, + 0.79, + 0.293 + ], + "angle": 0, + "content": "Table 1: Quantitative Comparison. Comparison of FID and sFID with other state-of-the-art methods on different resolution datasets." + }, + { + "type": "table", + "bbox": [ + 0.268, + 0.304, + 0.737, + 0.404 + ], + "angle": 0, + "content": "
MethodCIFAR 32×32FFHQ 128×128AFHQ-D 256×256CelebAHQ 256×256
FIDsFIDFIDsFIDFIDsFIDFIDsFID
DDPM3.014.6914.0712.8917.2149.0316.3225.20
Debias3.174.5611.8611.6515.9146.7714.2321.98
Min-SNR3.234.8012.4612.8517.0347.9414.3724.87
P2-Weight3.194.7811.9612.7416.1847.8814.3822.14
B-TTDM2.894.2210.2811.3813.8145.2513.6221.93
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Comparison to other methods. Since the B-TTDM method is a non-uniform treatment of different time intervals, we contrast it with other start-of-the-art methods that employ non-uniformly re-weighted loss functions in the training stage. The result shown in Table 1 indicates that our B-TTDM perform best across multiple different resolution datasets. In addition, we find that most of these re-weighted loss methods have negative effect on low-resolution CIFAR datasets. Despite this, B-TTDM consistently achieves the best performance, underscoring its broad applicability and generalizability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Plug-and-Enhance. Our B-TTDM is designed to align the time sampling and the non-uniform variation observed in the data distribution during forward diffusion, setting it apart from other techniques aimed at enhancing diffusion model performance. Therefore, our method is complementary and can be integrated as a plug-in enhancement within existing strategies for optimizing the training of diffusion models. We combine our B-TTDM with other different improved training methods during the training stage following the same setting. The results are shown in Table 2, we find that our B-TTDM is beneficial for other improved methods like P2-Weight, Min-SNR and DDPM-IP. Notably, the integration of B-TTDM with DDPM-IP leads to a significant achievement, such as an FID score of 2.46 for the CIFAR dataset and a 6.69 FID score for FFHQ." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.494, + 0.756 + ], + "angle": 0, + "content": "4.3 Fast Sampling and Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Fast Sampling. While diffusion models produce high-quality samples, their slow inference speed constrains their applicability in real-world settings. To improve this, various methods like DDIM [44], Time Respacing [9], DPM-Slover [34] and Exposure Bias [38] have been introduced. Fortunately, our B-TTDM method maintains the reliance of these acceleration techniques on the training process," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beta-Tuned Timestep Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 2: Quantitative Comparison. Comparison of FID and sFID when combining B-TTDM with other improved training methods." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.787, + 0.248 + ], + "angle": 0, + "content": "
MethodCIFAR FID sFIDFFHQ FID sFIDMethodAFHQ-D FID sFIDMethodAFHQ-D FID sFID
DDPM-IP2.724.547.078.35Min-SNR17.03
+ B-TTDM2.463.956.698.10+ B-TTDM13.34
46.13
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.534 + ], + "angle": 0, + "content": "allowing for their straightforward integration into our approach. The fast sampling outcomes of B-TTDM and DDPM models are contrasted in Table 3. We find that the B-TTDM method significantly outperforms the DDPM method under each different fast samplers and different NFE. Thus, B-TTDM can be seamlessly applied with various fast samplers for DDPM without requiring specific tuning. Fast Training. In addition to yielding higher-quality samples, B-TTDM also accelerates the training process of diffusion models. In Figure 4, we present a comparison between B-TTDM and DDPM in terms of performance and the training iterations on the large scale ImageNet and high resolution AFHQ-D datasets. Specifically, for ImageNet, B-TTDM achieves a \\(4.71 \\times\\) acceleration in training speed to attain equivalent FID scores. Once training converges, B-TTDM's FID score significantly exceeds that of DDPM. Similarly, on the large resolution AFHQ-D dataset, B-TTDM also achieves a \\(3.33 \\times\\) training speed-up and perform better than DDPM upon convergence. We attribute this acceleration to our beta-tuned timestep sampling, which aligns the timestep sampling distribution with the properties of the diffusion process. This alignment not only speeds up the training of diffusion models but also improves the generated samples quality." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison. Comparison of FID score across various Fast Samplers on AFHQ-D dataset. NFE means the number of function evaluations." + }, + { + "type": "table", + "bbox": [ + 0.285, + 0.598, + 0.72, + 0.663 + ], + "angle": 0, + "content": "
Fast SamplersEpsilonScalingDDIMRespacingDPM-Solver
NFE2550255025502550
MethodDDPM23.6318.5024.2218.2523.6419.8318.4917.49
B-TTDM20.8115.3421.7616.7721.7916.6315.8214.79
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.389, + 0.727 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Parameters of the beta distribution. The B-TTDM utilizes the Beta-tuned distribution to devise a novel time sampling strategy. This distribution is defined by two parameters, \\(\\alpha\\) and \\(\\beta\\), where \\(\\alpha\\) predominantly influences the left side of the distribution's shape. As \\(\\alpha\\) increases, the bulk of the probability density function shifts towards the right. Conversely, \\(\\beta\\) primarily affects the right side of the distribution's shape. To further explore their impact on diffusion model training, we experiment with various \\(\\alpha\\) and \\(\\beta\\) combinations, with the experimental" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "T. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.147, + 0.482, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.289, + 0.429, + 0.302 + ], + "angle": 0, + "content": "(a) ImageNet \\((32\\times 32)\\)" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.146, + 0.758, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.289, + 0.71, + 0.302 + ], + "angle": 0, + "content": "(b) AFHQ-D \\((256\\times 256)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.313, + 0.789, + 0.34 + ], + "angle": 0, + "content": "Fig. 4: FID scores concerning the number of training iterations on ImageNet \\((32 \\times 32)\\) and AFHQ-D \\((256 \\times 256)\\) dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.789, + 0.449 + ], + "angle": 0, + "content": "outcomes presented in Table 4. We find that beta-tuned timestep sampling leads to significant improvements in all metrics and achieves the best performance at \\(\\alpha = 0.8, \\beta = 1\\). This also implies that beta-tuned timestep sampling is beneficial for generating quality improvements in diffusion models. More experimental results of different parameters can be found in the Appendix Section 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.79, + 0.556 + ], + "angle": 0, + "content": "Noise schedules. To ensure the robustness of the B-TTDM method to different noises schedules, we try two different commonly used noise schedules, cosine and linear noise schedules [9] on the CIFAR10 dataset. We follow the previous setting and each FID and sFID score is computed using \\( T' = 1000 \\) inference steps. The results of the different noise schedules experiments are presented in Table 5. The B-TTDM is significantly better than the Baseline in both FID and sFID scores, which indicates that the B-TTDM method is robust to different noise schedules." + }, + { + "type": "table_caption", + "bbox": [ + 0.245, + 0.582, + 0.479, + 0.61 + ], + "angle": 0, + "content": "Table 4: Ablation study on the parameters of Beta Distribution." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.622, + 0.501, + 0.702 + ], + "angle": 0, + "content": "
Beta DistributionAFHQ-D (256×256)
αβFIDsFIDRecallPrecision
1117.2149.030.5530.738
0.95114.1246.460.5620.805
0.90113.8545.470.6120.752
0.80113.8145.250.6120.819
0.70114.7847.560.5640.817
" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.585, + 0.757, + 0.612 + ], + "angle": 0, + "content": "Table 5: Ablation study on the Noise Schedule on CIFAR." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.626, + 0.764, + 0.698 + ], + "angle": 0, + "content": "
NoiseScheduleDDPMB-TTDM
LinearFID3.533.28
sFID4.784.41
CosineFID3.012.89
sFID4.694.22
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Other timestep adjust method. In this section, we compare various methods for adjusting the timestep on CIFAR and AFHQ datasets. For instance, the ANT-UW [17] method employs Uncertainty Weighting for different timestep based on multi-task learning with varying variances. Additionally, we compare timestep sampling methods based on alternative non-uniform distributions, such as exponential distributions (DDPM-E) and those where \\(\\beta > 1\\) and \\(\\alpha = 1\\) (B-TTDM\\(^{\\alpha}\\)). The experimental results are presented in Table 6. It is evident that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beta-Tuned Timestep Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 6: Quantitative comparison. Ablation study on different non-uniform timestep sampling method in the training stage of diffusion model." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.185, + 0.734, + 0.283 + ], + "angle": 0, + "content": "
MethodCIFAR (32×32)AFHQ (256×256)
FIDsFIDRecallPrecisionFIDsFIDRecallPrecision
DDPM3.014.690.6000.68217.2149.030.5530.738
DDPM-E2.954.660.5890.69515.1248.380.5090.703
ANT-UW2.934.450.5960.68714.9647.020.5980.794
B-TTDMα2.984.530.6000.69215.0046.980.6020.768
B-TTDM2.894.220.6020.69513.8145.250.6120.819
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.785, + 0.358 + ], + "angle": 0, + "content": "all these enhanced methods exhibit some degree of improvement in generation quality compare to DDPM. However, the B-TTDM method demonstrates the most significant enhancement in all evaluation metrics, underscoring its effectiveness." + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.387, + 0.415, + 0.397 + ], + "angle": 0, + "content": "Longer Training" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.401, + 0.755, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison. The generation results from DDPM and B-TTDM on FFHQ dataset. Images in each column are sampled from 50K, 100K, 300K and 500K training iterations. B-TTDM generates significantly better quality than DDPM with the same number of training iterations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.459, + 0.646 + ], + "angle": 0, + "content": "4.5 Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Generation results. We present a comparison of the generation results between various DDPMs and B-TTDMs in Figure 5. As depicted, images generated by B-TTDM exhibit markedly superior detail compared to those generated by DDPM. Furthermore, B-TTDMs manage to produce clear images by the 100K training iterations, whereas the quality of DDPM generated images at this stage remains relatively poor. Furthermore, we find that after 500K training iterations, B-TTDM still have better generation quality than DDPM in details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Reconstruct results. Building upon previous analyses, which suggest that the distribution exhibits greater variability in the initial stages of diffusion, we compare the generation capabilities of B-TTDM and DDPM for this significant difference. To achieve this, we input identical noise-containing samples to models trained for the same iterations and reconstructed the outputs. The results in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "T. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.146, + 0.331, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.204, + 0.331, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.262, + 0.304, + 0.271 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.147, + 0.408, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.204, + 0.409, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.262, + 0.392, + 0.271 + ], + "angle": 0, + "content": "DDPM" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.147, + 0.486, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.204, + 0.486, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.424, + 0.262, + 0.476, + 0.272 + ], + "angle": 0, + "content": "B-TTDM" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.147, + 0.592, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.204, + 0.591, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.262, + 0.565, + 0.272 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.147, + 0.672, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.204, + 0.672, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.262, + 0.655, + 0.272 + ], + "angle": 0, + "content": "DDPM" + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.147, + 0.752, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.204, + 0.75, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.262, + 0.74, + 0.272 + ], + "angle": 0, + "content": "B-TTDM" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.285, + 0.787, + 0.313 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison. Reconstruction results based on the same noisy inputs on FFHQ dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.789, + 0.476 + ], + "angle": 0, + "content": "Figure 6 indicate that the reconstructions generated by B-TTDM closely resemble the original samples, whereas those generated by DDPM exhibit more differences. Unconditional generation results. Figure 7 provides a qualitative comparison of the unconditional image generation capabilities of B-TTDM and DDPM across the different datasets. The Figure 7 reveals that samples generated by DDPM often exhibit noticeable color biases, whereas samples generated by B-TTDM are free from such biases and show a marked improvement in quality compared to DDPM. This underscores the effectiveness of B-TTDM. More unconditional generation results can be found in the Appendix Section 4." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.503, + 0.426, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.503, + 0.634, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.503, + 0.776, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.629, + 0.731, + 0.644 + ], + "angle": 0, + "content": "Fig. 7: Qualitative comparison. Unconditional generation results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.691, + 0.36, + 0.708 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, through both experimental exploration and theoretical analysis, we conclude that the distribution varies non-uniformly in forward diffusion. Moreover, the most drastic variations occur in the initial stages of the forward diffusion process. These insights suggest that the uniform timestep sampling is sub-optimal for training diffusion models. Building on these findings, we design the B-TTDM method, which aligns timestep sampling with these inherent properties. Our extensive experiments demonstrate that B-TTDM significantly enhances the quality of model generation while also accelerating the training process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beta-Tuned Timestep Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.205 + ], + "angle": 0, + "content": "This work was supported in part by NSFC under Grant 61927809 and in part by STCSM under Grant 22DZ2229005." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.226, + 0.324, + 0.243 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.255, + 0.785, + 0.283 + ], + "angle": 0, + "content": "1. Anderson, B.D.: Reverse-time diffusion equation models. Stochastic Processes and their Applications 12(3), 313-326 (1982)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.31 + ], + "angle": 0, + "content": "2. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR. pp. 22669-22679. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.337 + ], + "angle": 0, + "content": "3. Block, A., Mroueh, Y., Rakhlin, A., Ross, J.: Fast mixing of multi-scale Langevin dynamics under the manifold hypothesis. CoRR abs/2006.11166 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.378 + ], + "angle": 0, + "content": "4. Cai, R., Song, Z., Guan, D., Chen, Z., Luo, X., Yi, C., Kot, A.: Benchlmm: Benchmarking cross-style visual capability of large multimodal models. ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.404 + ], + "angle": 0, + "content": "5. Chen, Z., Li, B., Wu, S., Jiang, K., Ding, S., Zhang, W.: Content-based unrestricted adversarial attack. NeurIPS 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.433 + ], + "angle": 0, + "content": "6. Chen, Z., Li, B., Xu, J., Wu, S., Ding, S., Zhang, W.: Towards practical certifiable patch defense with vision transformer. In: CVPR. pp. 15148-15158 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.459 + ], + "angle": 0, + "content": "7. Choi, J., Lee, J., Shin, C., Kim, S., Kim, H., Yoon, S.: Perception prioritized training of diffusion models. In: CVPR. pp. 11462-11471. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.487 + ], + "angle": 0, + "content": "8. Choi, Y., Uh, Y., Yoo, J., Ha, J.: Stargan v2: Diverse image synthesis for multiple domains. In: CVPR. pp. 8185-8194. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.785, + 0.514 + ], + "angle": 0, + "content": "9. Dhariwal, P., Nichol, A.Q.: Diffusion models beat gans on image synthesis. In: NeurIPS. pp. 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.541 + ], + "angle": 0, + "content": "10. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real NVP. In: ICLR. OpenReview.net (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.785, + 0.581 + ], + "angle": 0, + "content": "1. Fan, K., Tang, J., Cao, W., Yi, R., Li, M., Gong, J., Zhang, J., Wang, Y., Wang, C., Ma, L.: Freemotion: A unified framework for number-free text-to-motion synthesis ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.581, + 0.785, + 0.609 + ], + "angle": 0, + "content": "2. Fang, B., Li, B., Wu, S., Yi, R., Ding, S., Ma, L.: Re-thinking data availability attacks against deep neural networks. CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "3. Fang, B., Li, B., Wu, S., Zheng, T., Ding, S., Yi, R., Ma, L.: Towards generalizable data protection with transferable unlearnable examples. CoRR abs/2305.11191 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.677 + ], + "angle": 0, + "content": "4. Ge, X., Liu, X., Yu, Z., Shi, J., Qi, C., Li, J., Kalviainen, H.: Diffas: Face anti-spoofing via generative diffusion models. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.704 + ], + "angle": 0, + "content": "5. Geng, C., Han, T., Jiang, P.T., Zhang, H., Chen, J., Hauberg, S., Li, B.: Improving adversarial energy-based model via diffusion process. ICML (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.732 + ], + "angle": 0, + "content": "6. Geng, C., Wang, J., Gao, Z., Frellsen, J., Hauberg, S.: Bounds all around: training energy-based models with bidirectional bounds. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.758 + ], + "angle": 0, + "content": "7. Go, H., Lee, Y., Lee, S., Oh, S., Moon, H., Choi, S.: Addressing negative transfer in diffusion models. NeurIPS 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "8. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.C., Bengio, Y.: Generative adversarial nets. In: NIPS. pp. 2672–2680 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Hang, T., Gu, S., Li, C., Bao, J., Chen, D., Hu, H., Geng, X., Guo, B.: Efficient diffusion training via min-snr weighting strategy. In: ICCV. pp. 7407-7417. IEEE (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.255, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "T. Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS. pp. 6626-6637 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "21. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A.A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models. CoRR abs/2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.286 + ], + "angle": 0, + "content": "23. Hyvarinen, A.: Estimation of non-normalized statistical models by score matching. J. Mach. Learn. Res. 6, 695-709 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "24. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. CoRR abs/1710.10196 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "25. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "26. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. PAMI 43(12), 4217-4228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.398 + ], + "angle": 0, + "content": "27. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. In: NeurIPS. pp. 10236-10245 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.399, + 0.787, + 0.412 + ], + "angle": 0, + "content": "28. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "29. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. In: NeurIPS. pp. 3929-3938 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.466 + ], + "angle": 0, + "content": "30. Lee, H., Lu, J., Tan, Y.: Convergence for score-based generative modeling with polynomial complexity. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "31. Li, X., Thickstun, J., Gulrajani, I., Liang, P., Hashimoto, T.B.: Diffusion-lm improves controllable text generation. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "32. Liu, Y., Chen, Y., Dai, W., Gou, M., Huang, C.T., Xiong, H.: Source-free domain adaptation with domain generalized pretraining for face anti-spoofing. PAMI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.55 + ], + "angle": 0, + "content": "33. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: ICCV. pp. 3730-3738. IEEE Computer Society (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "34. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "35. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "36. Nash, C., Menick, J., Dieleman, S., Battaglia, P.W.: Generating images with sparse representations. In: ICML. vol. 139, pp. 7958-7968. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.787, + 0.66 + ], + "angle": 0, + "content": "37. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. vol. 139, pp. 8162-8171. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "38. Ning, M., Li, M., Su, J., Salah, A.A., Ertugrul, I.O.: Elucidating the exposure bias in diffusion models. ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "39. Ning, M., Sangineto, E., Porrello, A., Calderara, S., Cucchiara, R.: Input perturbation reduces exposure bias in diffusion models. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.787, + 0.757 + ], + "angle": 0, + "content": "40. Parmar, G., Singh, K.K., Zhang, R., Li, Y., Lu, J., Zhu, J.: Zero-shot image-to-image translation. In: SIGGRAPH (Conference Paper Track). pp. 11:1-11:11. ACM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.784 + ], + "angle": 0, + "content": "41. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.668, + 0.799 + ], + "angle": 0, + "content": "42. Risken, H., Risken, H.: Fokker-planck equation. Springer (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "43. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10674-10685. IEEE (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beta-Tuned Timestep Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "44. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR. OpenReview.net (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "45. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS. pp. 11895-11907 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.245 + ], + "angle": 0, + "content": "46. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR. OpenReview.net (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.273 + ], + "angle": 0, + "content": "47. Tao, S., Wang, J.: Alleviation of gradient exploding in gans: Fake can be real. In: CVPR. pp. 1188-1197. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.313 + ], + "angle": 0, + "content": "48. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., Cheng, F.: Eliminating lipschitz singularities in diffusion models. ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.342 + ], + "angle": 0, + "content": "49. Yu, H., Shen, L., Huang, J., Zhou, M., Li, H., Zhao, F.: Debias the training of diffusion models. CoRR abs/2310.08442 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.37 + ], + "angle": 0, + "content": "50. Zeng, W., Yan, Y., Zhu, Q., Chen, Z., Chu, P., Zhao, W., Yang, X.: Infusion: Preventing customized text-to-image diffusion from overfitting (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.397 + ], + "angle": 0, + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.439 + ], + "angle": 0, + "content": "52. Zhu, Z., Wei, Y., Wang, J., Gan, Z., Zhang, Z., Wang, L., Hua, G., Wang, L., Liu, Z., Hu, H.: Exploring discrete diffusion models for image captioning. CoRR abs/2211.11694 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Beta-Tuned Timestep Diffusion Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_origin.pdf b/2024/Beta-Tuned Timestep Diffusion Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..57a5223b6327cbd792b9817e81acc6c6b98401f9 --- /dev/null +++ b/2024/Beta-Tuned Timestep Diffusion Model/360a561c-fdcc-427e-bd4f-3884d7f2debd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9e458dd9ba44d4f32e0e40b92d58ee5068b24a46a7c3724c846e0ea8e18b38d +size 936113 diff --git a/2024/Beta-Tuned Timestep Diffusion Model/full.md b/2024/Beta-Tuned Timestep Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9f7afbb96ce3e148c6e44a0993938338233fe109 --- /dev/null +++ b/2024/Beta-Tuned Timestep Diffusion Model/full.md @@ -0,0 +1,377 @@ +# Beta-Tuned Timestep Diffusion Model + +Tianyi Zheng $^{1,2}$ , Peng-Tao Jiang $^{2}$ , Ben Wan $^{1}$ , Hao Zhang $^{2}$ , Jinwei Chen $^{2}$ , Jia Wang $^{1}$ $\boxtimes$ , and Bo Li $^{2}$ $\boxtimes$ + +$^{1}$ Shanghai Jiao Tong University, Shanghai, China {tyzheng, burn-w, jiawang}@sjtu.edu.cn $^{2}$ vivo Mobile Communication Co., Ltd, China {pt.jiang,haozhang,jinwei.chen,libra}@vivo.com + +Abstract. Diffusion models have received a lot of attention in the field of generation due to their ability to produce high-quality samples. However, several recent studies indicate that treating all distributions equally in diffusion model training is sub-optimal. In this paper, we conduct an in-depth theoretical analysis of the forward process of diffusion models. Our findings reveal that the distribution variations are non-uniform throughout the diffusion process and the most drastic variations in distribution occur in the initial stages. Consequently, simple uniform timestep sampling strategy fail to align with these properties, potentially leading to suboptimal training of diffusion models. To address this, we propose the Beta-Tuned Timestep Diffusion Model (B-TTDM), which devises a timestep sampling strategy based on the beta distribution. By choosing the correct parameters, B-TTDM aligns the timestep sampling distribution with the properties of the forward diffusion process. Extensive experiments on different benchmark datasets validate the effectiveness of B-TTDM. + +Keywords: Generative Models $\cdot$ Diffusion Models $\cdot$ Beta Distribution + +# 1 Introduction + +Diffusion models have gained widespread application across various domains in recent years, such as text-to-image generation [5,43,50], text generation [31,52], video synthesis [21] and AI security [6,12-14,32], primarily due to their ability to produce high-quality samples. During the training stage, diffusion models introduce varying levels of Gaussian noise into the data distribution, creating a range of altered distributions. The model is then trained by minimizing denoising score-matching losses across these diverse distributions. In the inference stage, the trained model reverses this noise addition process to generate new samples. + +While DDPM [22], iDDPM [37] and ADM [9] have achieved significant success in image generation tasks, several studies [7, 17, 19, 48] identify that treating the distribution equally across different timesteps in the training stage could be + +sub-optimal. For instance, the P2-Weight [7] method discover through empirical analysis that distributions at various timesteps in forward diffusion encapsulate distinct semantic information. Building on this insight, P2-Weight propose a re-weighted loss function to elevate the quality of generated samples. Meanwhile, Min-SNR [19] and ANT [17] find inconsistencies in optimization gradients across various timesteps and recommend a re-weighted loss strategy to mitigate these issues. Recently, E-TSDM [48] uncover that the issue of numerical instability caused by high Lipschitz constants near the zero point. Therefore, E-TSDM assigns a weight of 0 to certain timesteps near zero to mitigate the issue of numerical instability. Despite the improvement in sample quality achieved by these methods through empirical adjustments to the training objective's weight, they still utilize DDPM's uniform timestep sampling strategy. As another influencing factor, the choice of timestep sampling strategy is also closely related to the training objective but has been overlooked in previous work. In this paper, we conduct a theoretical analysis of DDPM's forward diffusion process and design a novel strategy for timestep sampling, aiming to address this overlooked aspect and further improve the quality of generated samples. + +Our theoretical analysis in Theorem 2 demonstrates that distribution variations throughout the diffusion process are non-uniform, with the most significant variations occurring at the initial stages. This is experimentally supported by Figure 1, which shows non-uniform variations in both Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) [51] metrics, especially pronounced in the initial stages. Both our theoretical insights from Theorem 2 and the experimental evidence in Figure 1 underscore the non-uniform properties of forward diffusion, indicating that a uniform timestep sampling strategy may not be the most effective approach in the training stage. + +Building on these findings, we propose a novel timestep sampling strategy that utilizes the beta distribution, named the Beta-Tuned Timestep Diffusion Model (B-TTDM). By choosing the correct parameters of the beta distribution, B-TTDM aligns the timestep sampling distribution with the observed properties of the diffusion process and is specifically tuned to avoid singularities at $t = 0$ timesteps. Extensive experiments on various datasets demonstrate that B-TTDM not only improves the quality of the generated samples but also expedites the training process. For instance, B-TTDM achieves similar FID scores on the ImageNet dataset with about $20\%$ of the training iterations needed by the original approach and significantly improves generation quality after convergence. Moreover, B-TTDM can be compatible with other advanced diffusion model training methods, such as P2-Weight [7], Min-SNR [19], Debias [19], and DDPM-IP [39]. Through integration with these methods, B-TTDM can combine different advantages and generate higher quality samples. In summary, our contributions are: + +- Through theoretical and experimental verification, we confirm that distribution variations at different timesteps during forward diffusion are non-uniform, with the most significant shifts occurring in the initial stages. This indicates that the uniform time sampling in training is sub-optimal. + +- We design a novel timestep sampling strategy leveraging the Beta distribution to align with the non-uniform distribution shifts observed during forward diffusion. Furthermore, our B-TTDM is compatible with other enhancement methods, allowing for flexible integration into existing improved methods. +- Extensive experiments on different benchmark datasets demonstrate that the B-TTDM method significantly improves generation quality and drastically speeds up training of the diffusion models. + +![](images/ac860bafc98cfff77481754b1e8e2308e330b625403463731bbf7bfd38393657.jpg) +(a) PSNR at each forward diffusion timestep. + +![](images/cecff69cba16f4b21ecc8a2341fae4bfb0627e85784655fc797f5de07440ab81.jpg) +(b) LPIPS at each forward diffusion timestep. +Fig. 1: The PSNR and LPIPS metrics between the distribution at timestep $t$ and the original distribution in forward diffusion under different noise schedules. + +# 2 Related Work + +Diffusion Models. Diffusion models [15,22,37] are a family of generative models that generate samples from Gaussian noise via a learned denoising process. Recently, diffusion models have outperformed flow models [10,27] and variational autoencoders [16, 28] in generating higher quality samples. Furthermore, the diffusion models also outperform GANs [18,47] in certain cases [9]. This superior performance has led to the widespread adoption of diffusion models across diverse applications. For example, the Diffusion model [43] implements the generation of images based on textual inputs [4,11], image to image translation [35,40] and 3D Avatar generation [41]. Despite their success, the speed of sample generation by diffusion models is unsatisfactory. Therefore, DPM-Solver [34] and DDIM [44] design new generation methods based on the framework of ordinary differential equation (ODE), which have greatly improved the generation speed of the diffusion model. Meanwhile, EDM [25] design a higher order Runge-Kutta sampling method to further improve the generation speed and quality. These methods significantly enhance the speed of sample generation in diffusion models, with a primary emphasis on the generative aspect. Yet, they rely on pre-trained diffusion models without addressing training optimizations. Our proposed method presents an + +orthogonal strategy, aiming to improve the training stage and thereby further elevate the generate quality of diffusion models. + +Improved Diffusion Models. To further improve the quality of generative samples in the training stage, ADM [9] and U-ViT [2] design new advanced architectures for diffusion models. Based on these advanced architectures, the quality of generation is significantly improved. Additionally, P2-Weight [7], Min-SNR [19] and Debias [49] methods design the re-weight loss function based on SNR in the training stage. These can reduce the bias problem in the diffusion models. To make the neural network smoother, the E-TSDM [48] sets the weight of the loss function corresponding to timesteps close to the zero to 0. Additional, ANT [17] propose multi-task learning methods to set the loss weight of different timestep in the training stage. Although these improved methods can produce better samples, they overlook the non-uniform distribution variations in the diffusion process and continue to use uniform timestep sampling during training. Our theoretical analysis indicates that this uniform time sampling strategy in the training stage is not optimal, suggesting that a more tailored timestep sampling approach could yield significant improvements in generated samples. + +# 3 Method + +# 3.1 Background + +Let $q(x_0)$ represent the data distribution over $\mathbb{R}^n$ . During the forward diffusion, a sample $x_0$ is drawn from $q(x_0)$ and subsequently perturbed using a stochastic differential equation (SDE) [46]. + +$$ +\mathrm {d} x _ {t} = f \left(x _ {t}, t\right) \mathrm {d} t + g (t) \mathrm {d} w. \tag {1} +$$ + +The purpose of the forward diffusion is to transform the data distribution into simple normal Gaussian distribution $\mathcal{N}(0, I)$ . Meanwhile, we define $q_t(x_t)$ to represent the distribution at each time $t$ during the forward diffusion process. For any time $t$ , the conditional distribution $q_t(x_t | x_0)$ is [9, 22, 37] + +$$ +q _ {t} \left(x _ {t} \mid x _ {0}\right) = \mathcal {N} \left(x _ {t} \mid \alpha_ {t} x _ {0}, \sigma_ {t} ^ {2} \mathbf {I}\right). \tag {2} +$$ + +In Equation (2), $\alpha_{t}$ and $\sigma_{t}$ satisfy $\alpha_{t}^{2} + \sigma_{t}^{2} = 1$ . This forward diffusion process is known as the Ornstein-Uhlenbeck (OU) process. For more general expressions, we have the following conditional distribution + +$$ +q _ {t} \left(x _ {t} \mid x _ {0}\right) = \mathcal {N} \left(x _ {t} \mid e ^ {- \frac {t}{2}} x _ {0}, (1 - e ^ {- t}) \mathbf {I}\right). \tag {3} +$$ + +The reverse diffusion transforms normal Gaussian distribution into the data distribution $q(x_0)$ . This process satisfies the following reverse SDE equation [1]. + +$$ +\mathrm {d} x _ {t} = \left(f \left(x _ {t}, t\right) - g (t) ^ {2} \nabla_ {x} \log q _ {t} \left(x _ {t}\right)\right) \mathrm {d} t + g (t) \mathrm {d} w. \tag {4} +$$ + +Specially, $\nabla_x\log q_t(x_t)$ is the score function of the probability density $q_{t}(x_{t})$ Motivated by this correspondence, diffusion models train a neural network to + +estimate the $-\sigma_t\nabla_x\log q_t(x_t)$ . In the inference stage, we can simulate Equation (4) for sampling and generate new samples. + +# 3.2 Non-uniform variation in forward diffusion + +We first start by introducing and discussing our main assumptions. The only assumption we consider is that in the vicinity of high-density data, the score of the data distribution is bounded. Based on the definition of score [23, 45, 46], this assumption is quite common used [3, 30]. + +Assumption 1 For the diffusion process described by Equation (3), assume there exists a constant $\delta >0$ . For any $t\in [0,T]$ and any point $x_{t}$ in high-density regions, the score $\| \nabla_x\log q_t(x_t)\|$ is bounded by $\delta$ , i.e., + +$$ +\left\| \nabla_ {x} \log q _ {t} (x _ {t}) \right\| \leq \delta . +$$ + +This ensures that the score of the data distribution is bounded in the vicinity of high-density data throughout the diffusion process. + +Next, we examine the Fokker-Planck equation [42] corresponding to the Ornstein-Uhlenbeck process, as depicted in Equation (5). The parameter $\theta$ is drift coefficient and the parameter $D$ is diffusion coefficient. + +$$ +\frac {\partial}{\partial t} q \left(x _ {t}\right) = \theta \nabla_ {x _ {t}} q \left(x _ {t}\right) + D \nabla_ {x _ {t}} ^ {2} q \left(x _ {t}\right). \tag {5} +$$ + +Since the $\theta$ and $D$ are both constants independent of $x_{t}$ , to simplify the analysis that follows, we can let both of them be 1. Therefore, we have the following simplified Fokker-Planck equation. + +$$ +\frac {\partial}{\partial t} q \left(x _ {t}\right) = \nabla_ {x _ {t}} q \left(x _ {t}\right) + \nabla_ {x _ {t}} ^ {2} q \left(x _ {t}\right). \tag {6} +$$ + +The Fokker-Planck equation (6) sheds light on how a distribution's evolution is influenced by both time and spatial variables in the forward diffusion stage of DDPM. Our analysis further reveals that in the initial stages of diffusion, the changes in the distribution are primarily driven by the second-order derivative term, indicating a significant diffusion effect at this stage. Therefore, we have the following Lemma: + +Lemma 1. Consider the forward diffusion process as described by Equation (3) When $t\to 0^+$ , the time derivative of the distribution $q_{t}(x_{t})$ is predominantly governed by the second-order derivative term, encapsulated in the relationship: + +$$ +\frac {\partial}{\partial t} q _ {t} \left(x _ {t}\right) = \nabla_ {x _ {t}} ^ {2} q \left(x _ {t}\right), \tag {7} +$$ + +where $\nabla_{x_t}^2$ denotes the Laplacian operator with respect to $x_{t}$ , indicating the diffusion term's dominance in the evolution of $q_{t}(x_{t})$ when $t\to 0^{+}$ . + +The detailed proof of the Lemma (1) is shown in Appendix Section 2.2. Lemma (1) suggests that the alteration in the distribution primarily stems from the $\nabla_{x_t}^2 q(x_t)$ . Expanding on this insight, we can establish the following theorem, providing an upper bound on the alteration in distribution. + +Theorem 2. Consider the forward diffusion process described by Equation (7). The norm of the time derivative of the distribution $q_{t}(x_{t})$ is bounded by the following inequality: + +$$ +\left\| \frac {\partial}{\partial t} q _ {t} \left(x _ {t}\right) \right\| \leq \left\| \frac {q _ {t} \left(x _ {t}\right)}{2 t} \right\|. \tag {8} +$$ + +As $t \to 0^{+}$ , the norm of the derivative of the distribution $q_{t}(x_{t})$ may become unbounded. This is indicated by the limit superior: + +$$ +\lim _ {t \rightarrow 0 ^ {+}} \left\| \frac {q _ {t} \left(x _ {t}\right)}{2 t} \right\| = \infty . \tag {9} +$$ + +The detailed proof of the Theorem (2) is shown in Appendix Section 2.3. Theorem (2) suggests that the bound on the norm of the time derivative of $q_{t}(x_{t})$ grows indefinitely as $t \to 0^{+}$ , indicating that the norm of the derivative could be unbounded. We present a straightforward example to illustrate this behavior. + +Simple case illustration. Let $q_{t}(x_{t}) = \frac{1}{\sqrt{4\pi t}} e^{-\frac{x^{2}}{4t}}$ be a solution to differential equation (7). Meanwhile, we notice that $\int_{-\infty}^{+\infty} q_{t}(x_{t}) dx = 1$ and $q_{t}(x) > 0$ , which means $q_{t}(x)$ is a probability density function. For every $t > 0$ , the norm of the time derivative of $q_{t}(x)$ is given by: + +$$ +\left\| \frac {\partial}{\partial t} q _ {t} \left(x _ {t}\right) \right\| = \left\| \frac {\sqrt {t} (- 2 t + x ^ {2}) e ^ {- \frac {x ^ {2}}{4 t}}}{8 \sqrt {\pi} t ^ {3}} \right\|. \tag {10} +$$ + +This expression leads to the conclusion that: + +$$ +\lim _ {t \rightarrow 0 ^ {+}} \left\| \frac {\partial}{\partial t} q _ {t} \left(x _ {t}\right)\right\| = \infty , \tag {11} +$$ + +indicating that the norm of the time derivative of $q_{t}(x)$ becomes unbounded as $t \to 0^{+}$ . This result indicates an infinite rate of variation for the distribution $q_{t}(x_{t})$ when $t \to 0^{+}$ . + +Theorem (2) highlights the non-uniform properties of distribution variations during diffusion, with the initial stages experiencing the most significant shifts. Furthermore, Proposition (1) clarifies that the distribution variations $q_{t}(x_{t})$ are non-uniform across the diffusion process, and these variations tend to flatten as the process progresses to later stages. + +Proposition 1. Consider the forward diffusion process in Equation (3). The upper bound of the KL divergence between the distribution $q_{t}(x_{t})$ and the standard Gaussian distribution converges exponentially. i.e., + +$$ +\operatorname {K L} \left(q _ {t} \left(x _ {t}\right) \| \mathcal {N} (0, 1)\right) \leq e ^ {- t} \operatorname {K L} \left(q _ {0} \left(x _ {0}\right) \| \mathcal {N} (0, 1)\right). \tag {12} +$$ + +Based on the Theorem (2) and Proposition (1), we notice that the variations in the distribution of forward diffusion is non-uniform, with the sharpest change at the beginning of the diffusion and converging to a Gaussian distribution at an exponential rate. These theoretical insights guide the design of our timestep sampling strategy. + +# 3.3 Beta distribution timestep sampling + +To align the non-uniform trend of variation, we introduce a time sampling strategy that utilizes the Beta distribution for training the diffusion model. This strategy leverages the Beta density function's capacity to take on a diverse array of shapes, controlled by the parameters $\alpha$ and $\beta$ . The probability density function (pdf) of the Beta distribution is defined as: + +$$ +f (t; \alpha , \beta) = \frac {\Gamma (\alpha + \beta)}{\Gamma (\alpha) \Gamma (\beta)} t ^ {\alpha - 1} (1 - t) ^ {\beta - 1}, \tag {13} +$$ + +Where $\Gamma(z)$ is the gamma function. Setting $\alpha = 1$ and $\beta = 1$ simplifies the Beta distribution to a uniform distribution, illustrating the flexibility of Beta distribution. By adjusting the values of $\alpha$ and $\beta$ , we can tailor the time sampling distribution within the diffusion model training to better align the desired variation properties. + +Since the beta distribution is a family of continuous probability distributions defined on the interval [0, 1]. We can use $\frac{t}{T}$ to map the diffusion timestep into the [0, 1] interval of the beta distribution. Then we compare the probability density function of beta distribution for different $\alpha$ and $\beta$ in Figure 2. We find that when we set $\alpha > 1$ and $\beta > 1$ , The probability density function is U-shaped. Meanwhile, we notice that $\alpha = 1$ and $\beta > 1$ and $\alpha < 1$ and $\beta = 1$ , the probability density function conforms to our expectation of a non-uniformly decreasing. However, when $\alpha = 1$ and $\beta > 1$ , the probability density function exhibits a positively skewed shape, lacking alignment with the most significant varies in the initial diffusion stage. We provide ablation study of different $\alpha$ and $\beta$ in the section 4.4. + +To align with the previously identified properties, we choose to set $\alpha < 1$ and $\beta = 1$ . With these parameters, the distribution adopts a reverse J-shaped profile. Consequently, the probability density function under these conditions becomes a strictly decreasing function, exhibiting a sharp change as it approaches 0. This specific shape and behavior of the pdf are instrumental in aligning the rapid initial changes and the convergence properties observed in the forward diffusion process. Therefore, we replace the uniform time sampling in the original diffusion model with this time sampling method based on the beta distribution. In addition, to prevent the probability density function from reaching infinite values as time t approaches 0, we adopt $\frac{1}{T}$ as the time sampling weight at the initial + +![](images/47b782961c05d521e76e9e8d6c4c37ae95c5618af0e96ec98c2b7f0d681b09fc.jpg) +Fig. 2: Probability density function of beta distribution with different parameters. + +moment. Therefore, our method is called Beta-Tuned Timestep Diffusion Model (B-TTDM). Next, we analyse the errors introduced by this tuned approach. + +Proposition 2. Given a timestep sampling method utilizing the B-TTDM with a Beta distribution parameterized by $\alpha < 1$ and $\beta = 1$ , the tuned error at the initial time position $t$ is $\left(\frac{1}{T}\right)^{\alpha}$ , where $T$ is the total diffusion steps. + +The detailed proof of the Proposition (2) is shown in Appendix Section 2.4. Since the value of $T$ is usually large in the training of diffusion models (e.g., $T = 1000$ ), this implies that the tuned error is quite small. Therefore, we can ignore it and still consider this as a beta distribution. Finally, the corresponding training loss can be written as + +$$ +\mathcal {L} \left(\epsilon_ {\theta}\right) := \mathbb {E} _ {t \sim \operatorname {B e t a} (\alpha , \beta), \mathbf {x} _ {0} \sim q (\mathbf {x} _ {0}), \epsilon \sim \mathcal {N} (0, \mathbf {I})} \left[ \| \epsilon_ {\theta} \left(\alpha_ {t} \mathbf {x} _ {0} + \sigma_ {t} \epsilon , t\right) - \epsilon \| _ {2} ^ {2} \right]. \tag {14} +$$ + +Furthermore, we present a comprehensive outline of the B-TTDM method in Algorithm (1). The key difference between B-TTDM and DDPM is the strategy to timestep sampling in the training stage. Thus, B-TTDM is adaptable to most fast sampling methods and improved training methods designed for DDPM without the need for specific modifications. + +# Algorithm 1 Training of B-TTDM + +Require: Beta distribution parameter $\alpha$ and $\beta$ . Total diffusion timestep $T$ . + +1: Generate a beta distribution with parameters $\alpha$ and $\beta$ . +2: Tuned $f(0; \alpha, \beta) = f\left(\frac{1}{T}; \alpha, \beta\right)$ +3: repeat +4: $x_0\sim q(x_0)$ +5: $t\sim \mathrm{Beta}(\alpha ,\beta),\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ +6: Take a gradient descent step on $\nabla_{\pmb{\theta}}\| \pmb {\epsilon} - \pmb {\epsilon}_{\pmb{\theta}}(\pmb {x}_t,t)\|^2$ +7: until converged + +![](images/263b89ac5d26157f296b9bc30c9d92b3632d15dafa117c711debc073a7dd3b6c.jpg) +(a) Comparison of FID on low-resolution. + +![](images/a95b254369b24c330635592f984919eed9eef2932b8ccde8db896b6f65c4d5a9.jpg) +(b) Comparison of FID on high-resolution. +Fig. 3: Quantitative Comparison. Comparing FID of different resolution datasets. + +# 4 Experiment + +In this section, we first provide an implementation details of the experimental setup in section 4.1. Then, we provide quantitative comparison of B-TTDM with other state-of-the-art methods in section 4.2. In Section 4.3, we assess the robustness of B-TTDM to different fast sampling algorithms and its training efficiency in comparison to DDPM. Subsequent ablation studies are discussed in Section 4.4. Lastly, qualitative comparisons and visualization results are showcased in Section 4.5. + +# 4.1 Experiment Setup + +Implementation details. All of our experiments utilize the same settings of DDPM [22]. Meanwhile we employ the same model architecture as ADM [9] since it is more advanced compared to DDPM baseline. Throughout the training stage, we maintain $T = 1,000$ for all experiments. We conduct comparative experiments using the CIFAR-10 $(32 \times 32)$ , ImageNet $(32 \times 32)$ , CelebA $(64 \times 64)$ [33], FFHQ $(128 \times 128)$ [26], AFHQ-D $(256 \times 256)$ [8] and CelebAHQ $(256 \times 256)$ [24] datasets for unconditional image generation. To evaluate the sampling quality, we use the Frechet inception distance (FID) [20], sFID [36], improved Precision and Recall [29] as our evaluation metrics. To be more specific, we generate 50K samples for CIFAR-10, ImageNet, CelebA, FFHQ and 10K samples for AFHQ-D, CelebAHQ. In the inference stage, we use 1000 steps for CIFAR-10 and ImageNet, 300 steps for CelebA and FFHQ, and 100 steps for AFHQ-D and CelebAHQ. More details can be found in the Appendix Section 3. + +# 4.2 Quantitative Comparison + +Comparison to Baseline. We compare our method with DDPM baseline using various benchmark datasets. The results of the experiments are shown in Figure 3, where we find that the FID of our method significantly outperforms the DDPM + +on datasets of different resolutions. For instance, on the FFHQ dataset, B-TTDM achieved a FID improvement of 3.79. This underscores our method's efficacy and suggests that the uniform time sampling approach previously employed for training diffusion models is not the most efficient. Notably, B-TTDM reaches an FID of 2.87 on the ImageNet dataset with a larger data size, significantly enhancing DDPM's performance and highlighting B-TTDM's effectiveness. + +Table 1: Quantitative Comparison. Comparison of FID and sFID with other state-of-the-art methods on different resolution datasets. + +
MethodCIFAR 32×32FFHQ 128×128AFHQ-D 256×256CelebAHQ 256×256
FIDsFIDFIDsFIDFIDsFIDFIDsFID
DDPM3.014.6914.0712.8917.2149.0316.3225.20
Debias3.174.5611.8611.6515.9146.7714.2321.98
Min-SNR3.234.8012.4612.8517.0347.9414.3724.87
P2-Weight3.194.7811.9612.7416.1847.8814.3822.14
B-TTDM2.894.2210.2811.3813.8145.2513.6221.93
+ +Comparison to other methods. Since the B-TTDM method is a non-uniform treatment of different time intervals, we contrast it with other start-of-the-art methods that employ non-uniformly re-weighted loss functions in the training stage. The result shown in Table 1 indicates that our B-TTDM perform best across multiple different resolution datasets. In addition, we find that most of these re-weighted loss methods have negative effect on low-resolution CIFAR datasets. Despite this, B-TTDM consistently achieves the best performance, underscoring its broad applicability and generalizability. + +Plug-and-Enhance. Our B-TTDM is designed to align the time sampling and the non-uniform variation observed in the data distribution during forward diffusion, setting it apart from other techniques aimed at enhancing diffusion model performance. Therefore, our method is complementary and can be integrated as a plug-in enhancement within existing strategies for optimizing the training of diffusion models. We combine our B-TTDM with other different improved training methods during the training stage following the same setting. The results are shown in Table 2, we find that our B-TTDM is beneficial for other improved methods like P2-Weight, Min-SNR and DDPM-IP. Notably, the integration of B-TTDM with DDPM-IP leads to a significant achievement, such as an FID score of 2.46 for the CIFAR dataset and a 6.69 FID score for FFHQ. + +# 4.3 Fast Sampling and Training + +Fast Sampling. While diffusion models produce high-quality samples, their slow inference speed constrains their applicability in real-world settings. To improve this, various methods like DDIM [44], Time Respacing [9], DPM-Slover [34] and Exposure Bias [38] have been introduced. Fortunately, our B-TTDM method maintains the reliance of these acceleration techniques on the training process, + +Table 2: Quantitative Comparison. Comparison of FID and sFID when combining B-TTDM with other improved training methods. + +
MethodCIFAR FID sFIDFFHQ FID sFIDMethodAFHQ-D FID sFIDMethodAFHQ-D FID sFID
DDPM-IP2.724.547.078.35Min-SNR17.03
+ B-TTDM2.463.956.698.10+ B-TTDM13.34
46.13
+ +allowing for their straightforward integration into our approach. The fast sampling outcomes of B-TTDM and DDPM models are contrasted in Table 3. We find that the B-TTDM method significantly outperforms the DDPM method under each different fast samplers and different NFE. Thus, B-TTDM can be seamlessly applied with various fast samplers for DDPM without requiring specific tuning. Fast Training. In addition to yielding higher-quality samples, B-TTDM also accelerates the training process of diffusion models. In Figure 4, we present a comparison between B-TTDM and DDPM in terms of performance and the training iterations on the large scale ImageNet and high resolution AFHQ-D datasets. Specifically, for ImageNet, B-TTDM achieves a $4.71 \times$ acceleration in training speed to attain equivalent FID scores. Once training converges, B-TTDM's FID score significantly exceeds that of DDPM. Similarly, on the large resolution AFHQ-D dataset, B-TTDM also achieves a $3.33 \times$ training speed-up and perform better than DDPM upon convergence. We attribute this acceleration to our beta-tuned timestep sampling, which aligns the timestep sampling distribution with the properties of the diffusion process. This alignment not only speeds up the training of diffusion models but also improves the generated samples quality. + +Table 3: Quantitative comparison. Comparison of FID score across various Fast Samplers on AFHQ-D dataset. NFE means the number of function evaluations. + +
Fast SamplersEpsilonScalingDDIMRespacingDPM-Solver
NFE2550255025502550
MethodDDPM23.6318.5024.2218.2523.6419.8318.4917.49
B-TTDM20.8115.3421.7616.7721.7916.6315.8214.79
+ +# 4.4 Ablation Study + +Parameters of the beta distribution. The B-TTDM utilizes the Beta-tuned distribution to devise a novel time sampling strategy. This distribution is defined by two parameters, $\alpha$ and $\beta$ , where $\alpha$ predominantly influences the left side of the distribution's shape. As $\alpha$ increases, the bulk of the probability density function shifts towards the right. Conversely, $\beta$ primarily affects the right side of the distribution's shape. To further explore their impact on diffusion model training, we experiment with various $\alpha$ and $\beta$ combinations, with the experimental + +![](images/9c00b3831333798d092e5d2909f06ff13a61e9d136722dfaff608b9f537c5a87.jpg) +(a) ImageNet $(32\times 32)$ + +![](images/3094ff394f11932ea854c19fd44a35d41db6e0e3a0a09668f7fe10c397be7b8e.jpg) +(b) AFHQ-D $(256\times 256)$ +Fig. 4: FID scores concerning the number of training iterations on ImageNet $(32 \times 32)$ and AFHQ-D $(256 \times 256)$ dataset. + +outcomes presented in Table 4. We find that beta-tuned timestep sampling leads to significant improvements in all metrics and achieves the best performance at $\alpha = 0.8, \beta = 1$ . This also implies that beta-tuned timestep sampling is beneficial for generating quality improvements in diffusion models. More experimental results of different parameters can be found in the Appendix Section 3. + +Noise schedules. To ensure the robustness of the B-TTDM method to different noises schedules, we try two different commonly used noise schedules, cosine and linear noise schedules [9] on the CIFAR10 dataset. We follow the previous setting and each FID and sFID score is computed using $T' = 1000$ inference steps. The results of the different noise schedules experiments are presented in Table 5. The B-TTDM is significantly better than the Baseline in both FID and sFID scores, which indicates that the B-TTDM method is robust to different noise schedules. + +Table 4: Ablation study on the parameters of Beta Distribution. + +
Beta DistributionAFHQ-D (256×256)
αβFIDsFIDRecallPrecision
1117.2149.030.5530.738
0.95114.1246.460.5620.805
0.90113.8545.470.6120.752
0.80113.8145.250.6120.819
0.70114.7847.560.5640.817
+ +Table 5: Ablation study on the Noise Schedule on CIFAR. + +
NoiseScheduleDDPMB-TTDM
LinearFID3.533.28
sFID4.784.41
CosineFID3.012.89
sFID4.694.22
+ +Other timestep adjust method. In this section, we compare various methods for adjusting the timestep on CIFAR and AFHQ datasets. For instance, the ANT-UW [17] method employs Uncertainty Weighting for different timestep based on multi-task learning with varying variances. Additionally, we compare timestep sampling methods based on alternative non-uniform distributions, such as exponential distributions (DDPM-E) and those where $\beta > 1$ and $\alpha = 1$ (B-TTDM $^{\alpha}$ ). The experimental results are presented in Table 6. It is evident that + +Table 6: Quantitative comparison. Ablation study on different non-uniform timestep sampling method in the training stage of diffusion model. + +
MethodCIFAR (32×32)AFHQ (256×256)
FIDsFIDRecallPrecisionFIDsFIDRecallPrecision
DDPM3.014.690.6000.68217.2149.030.5530.738
DDPM-E2.954.660.5890.69515.1248.380.5090.703
ANT-UW2.934.450.5960.68714.9647.020.5980.794
B-TTDMα2.984.530.6000.69215.0046.980.6020.768
B-TTDM2.894.220.6020.69513.8145.250.6120.819
+ +all these enhanced methods exhibit some degree of improvement in generation quality compare to DDPM. However, the B-TTDM method demonstrates the most significant enhancement in all evaluation metrics, underscoring its effectiveness. + +![](images/3f02b6a6fcc46ef9beab80dd8088504d8725ebef752a5cfaa491a008947f24e7.jpg) +Longer Training +Fig. 5: Qualitative comparison. The generation results from DDPM and B-TTDM on FFHQ dataset. Images in each column are sampled from 50K, 100K, 300K and 500K training iterations. B-TTDM generates significantly better quality than DDPM with the same number of training iterations. + +# 4.5 Qualitative Comparison + +Generation results. We present a comparison of the generation results between various DDPMs and B-TTDMs in Figure 5. As depicted, images generated by B-TTDM exhibit markedly superior detail compared to those generated by DDPM. Furthermore, B-TTDMs manage to produce clear images by the 100K training iterations, whereas the quality of DDPM generated images at this stage remains relatively poor. Furthermore, we find that after 500K training iterations, B-TTDM still have better generation quality than DDPM in details. + +Reconstruct results. Building upon previous analyses, which suggest that the distribution exhibits greater variability in the initial stages of diffusion, we compare the generation capabilities of B-TTDM and DDPM for this significant difference. To achieve this, we input identical noise-containing samples to models trained for the same iterations and reconstructed the outputs. The results in + +![](images/fc446349999942ea5161653ad6aaa1cdde305a09055731a4375e05ea82b116cc.jpg) + +![](images/7b585033f94b8a86bef404694d364aea1ba29e54d0bdb9801ab1e372040e8fd7.jpg) +GT + +![](images/dac7b92ff7f2d854c06f98b88cf13ed1f2ea7a84b65586f2b7570dbfbfdb3e8f.jpg) + +![](images/daa05f69e53819155fd131d7058b436b5eecffd3afb02c78e3c8e5fe3200610c.jpg) +DDPM + +![](images/9e2bb0d32fbedb681b7d9aed33e95eb1c5bef4129f0a316ace40a954df8fcaee.jpg) + +![](images/0ff623cf41352abca1a0cbe19f516a6dcc1e6bfee19a8085e1c03a9220c47839.jpg) +B-TTDM +Fig. 6: Qualitative comparison. Reconstruction results based on the same noisy inputs on FFHQ dataset. + +![](images/4748509463d6cde1ec5aa5eb8a3735813b9586c1e45e58abbd4278d168f8cd0b.jpg) + +![](images/5102a8bbdd1d8fea7303520407baa8c230b3dab391942e4663c1d1bac0059f53.jpg) +GT + +![](images/871fad37211a72a1a924c03dcbc5ccd65c4f860248a334d20058640af0dc0640.jpg) + +![](images/55f0a861aac0a3c4bb8de2a6decc57b3d22f729683bebac2b8ba7fd470876a1c.jpg) +DDPM + +![](images/89b049e5525748b55a3f7680a4e40d95baeef9dd18c379ccb1a7fa446ee57cdd.jpg) + +![](images/2ae3c5d992445768451eba9bf81b00a14de5462dc651f946413c0a6eac592903.jpg) +B-TTDM + +Figure 6 indicate that the reconstructions generated by B-TTDM closely resemble the original samples, whereas those generated by DDPM exhibit more differences. Unconditional generation results. Figure 7 provides a qualitative comparison of the unconditional image generation capabilities of B-TTDM and DDPM across the different datasets. The Figure 7 reveals that samples generated by DDPM often exhibit noticeable color biases, whereas samples generated by B-TTDM are free from such biases and show a marked improvement in quality compared to DDPM. This underscores the effectiveness of B-TTDM. More unconditional generation results can be found in the Appendix Section 4. + +![](images/d47ffde84d45c1552ab68a2f17eed8a533b2c9cc80dbdc8fbfbf00ae01006744.jpg) + +![](images/20ec56947c0b59e9b9885edb1b78ec3bbaa5a30ec98a4deca63499e59866d972.jpg) +Fig. 7: Qualitative comparison. Unconditional generation results. + +![](images/a42300ea8f58c508ea46d71ab7a027a65411765abe966156e0e8db30e22c1672.jpg) + +# 5 Conclusion + +In this paper, through both experimental exploration and theoretical analysis, we conclude that the distribution varies non-uniformly in forward diffusion. Moreover, the most drastic variations occur in the initial stages of the forward diffusion process. These insights suggest that the uniform timestep sampling is sub-optimal for training diffusion models. Building on these findings, we design the B-TTDM method, which aligns timestep sampling with these inherent properties. Our extensive experiments demonstrate that B-TTDM significantly enhances the quality of model generation while also accelerating the training process. + +# Acknowledgments + +This work was supported in part by NSFC under Grant 61927809 and in part by STCSM under Grant 22DZ2229005. + +# References + +1. Anderson, B.D.: Reverse-time diffusion equation models. Stochastic Processes and their Applications 12(3), 313-326 (1982) +2. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR. pp. 22669-22679. IEEE (2023) +3. Block, A., Mroueh, Y., Rakhlin, A., Ross, J.: Fast mixing of multi-scale Langevin dynamics under the manifold hypothesis. CoRR abs/2006.11166 (2020) +4. Cai, R., Song, Z., Guan, D., Chen, Z., Luo, X., Yi, C., Kot, A.: Benchlmm: Benchmarking cross-style visual capability of large multimodal models. ECCV (2024) +5. Chen, Z., Li, B., Wu, S., Jiang, K., Ding, S., Zhang, W.: Content-based unrestricted adversarial attack. NeurIPS 36 (2024) +6. Chen, Z., Li, B., Xu, J., Wu, S., Ding, S., Zhang, W.: Towards practical certifiable patch defense with vision transformer. In: CVPR. pp. 15148-15158 (2022) +7. Choi, J., Lee, J., Shin, C., Kim, S., Kim, H., Yoon, S.: Perception prioritized training of diffusion models. In: CVPR. pp. 11462-11471. IEEE (2022) +8. Choi, Y., Uh, Y., Yoo, J., Ha, J.: Stargan v2: Diverse image synthesis for multiple domains. In: CVPR. pp. 8185-8194. Computer Vision Foundation / IEEE (2020) +9. Dhariwal, P., Nichol, A.Q.: Diffusion models beat gans on image synthesis. In: NeurIPS. pp. 8780-8794 (2021) +10. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real NVP. In: ICLR. OpenReview.net (2017) +1. Fan, K., Tang, J., Cao, W., Yi, R., Li, M., Gong, J., Zhang, J., Wang, Y., Wang, C., Ma, L.: Freemotion: A unified framework for number-free text-to-motion synthesis ECCV (2024) +2. Fang, B., Li, B., Wu, S., Yi, R., Ding, S., Ma, L.: Re-thinking data availability attacks against deep neural networks. CVPR (2024) +3. Fang, B., Li, B., Wu, S., Zheng, T., Ding, S., Yi, R., Ma, L.: Towards generalizable data protection with transferable unlearnable examples. CoRR abs/2305.11191 (2023) +4. Ge, X., Liu, X., Yu, Z., Shi, J., Qi, C., Li, J., Kalviainen, H.: Diffas: Face anti-spoofing via generative diffusion models. In: ECCV (2024) +5. Geng, C., Han, T., Jiang, P.T., Zhang, H., Chen, J., Hauberg, S., Li, B.: Improving adversarial energy-based model via diffusion process. ICML (2024) +6. Geng, C., Wang, J., Gao, Z., Frellsen, J., Hauberg, S.: Bounds all around: training energy-based models with bidirectional bounds. NeurIPS (2021) +7. Go, H., Lee, Y., Lee, S., Oh, S., Moon, H., Choi, S.: Addressing negative transfer in diffusion models. NeurIPS 36 (2024) +8. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.C., Bengio, Y.: Generative adversarial nets. In: NIPS. pp. 2672–2680 (2014) +9. Hang, T., Gu, S., Li, C., Bao, J., Chen, D., Hu, H., Geng, X., Guo, B.: Efficient diffusion training via min-snr weighting strategy. In: ICCV. pp. 7407-7417. IEEE (2023) + +20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS. pp. 6626-6637 (2017) +21. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A.A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models. CoRR abs/2210.02303 (2022) +22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +23. Hyvarinen, A.: Estimation of non-normalized statistical models by score matching. J. Mach. Learn. Res. 6, 695-709 (2005) +24. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. CoRR abs/1710.10196 (2017) +25. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. In: NeurIPS (2022) +26. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. PAMI 43(12), 4217-4228 (2021) +27. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. In: NeurIPS. pp. 10236-10245 (2018) +28. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014) +29. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. In: NeurIPS. pp. 3929-3938 (2019) +30. Lee, H., Lu, J., Tan, Y.: Convergence for score-based generative modeling with polynomial complexity. In: NeurIPS (2022) +31. Li, X., Thickstun, J., Gulrajani, I., Liang, P., Hashimoto, T.B.: Diffusion-lm improves controllable text generation. In: NeurIPS (2022) +32. Liu, Y., Chen, Y., Dai, W., Gou, M., Huang, C.T., Xiong, H.: Source-free domain adaptation with domain generalized pretraining for face anti-spoofing. PAMI (2024) +33. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: ICCV. pp. 3730-3738. IEEE Computer Society (2015) +34. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In: NeurIPS (2022) +35. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022) +36. Nash, C., Menick, J., Dieleman, S., Battaglia, P.W.: Generating images with sparse representations. In: ICML. vol. 139, pp. 7958-7968. PMLR (2021) +37. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. vol. 139, pp. 8162-8171. PMLR (2021) +38. Ning, M., Li, M., Su, J., Salah, A.A., Ertugrul, I.O.: Elucidating the exposure bias in diffusion models. ICLR (2024) +39. Ning, M., Sangineto, E., Porrello, A., Calderara, S., Cucchiara, R.: Input perturbation reduces exposure bias in diffusion models. In: ICML (2023) +40. Parmar, G., Singh, K.K., Zhang, R., Li, Y., Lu, J., Zhu, J.: Zero-shot image-to-image translation. In: SIGGRAPH (Conference Paper Track). pp. 11:1-11:11. ACM (2023) +41. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023) +42. Risken, H., Risken, H.: Fokker-planck equation. Springer (1996) +43. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10674-10685. IEEE (2022) + +44. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR. OpenReview.net (2021) +45. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS. pp. 11895-11907 (2019) +46. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR. OpenReview.net (2021) +47. Tao, S., Wang, J.: Alleviation of gradient exploding in gans: Fake can be real. In: CVPR. pp. 1188-1197. Computer Vision Foundation / IEEE (2020) +48. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., Cheng, F.: Eliminating lipschitz singularities in diffusion models. ICLR (2024) +49. Yu, H., Shen, L., Huang, J., Zhou, M., Li, H., Zhao, F.: Debias the training of diffusion models. CoRR abs/2310.08442 (2023) +50. Zeng, W., Yan, Y., Zhu, Q., Chen, Z., Chu, P., Zhao, W., Yang, X.: Infusion: Preventing customized text-to-image diffusion from overfitting (2024) +51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +52. Zhu, Z., Wei, Y., Wang, J., Gan, Z., Zhang, Z., Wang, L., Hua, G., Wang, L., Liu, Z., Hu, H.: Exploring discrete diffusion models for image captioning. CoRR abs/2211.11694 (2022) \ No newline at end of file diff --git a/2024/Beta-Tuned Timestep Diffusion Model/images.zip b/2024/Beta-Tuned Timestep Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..126f311fcd7a04dc70fd19328fe25d9adca6b9cc --- /dev/null +++ b/2024/Beta-Tuned Timestep Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:491b4dd0d5a43efb543f577db55f26bd8a1d0d00f516cc539104e2d682ca3dfa +size 491457 diff --git a/2024/Beta-Tuned Timestep Diffusion Model/layout.json b/2024/Beta-Tuned Timestep Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..69c684c7fe4ebc7acd5e63f4d19d5cd4ac29996c --- /dev/null +++ b/2024/Beta-Tuned Timestep Diffusion Model/layout.json @@ -0,0 +1,11227 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 111, + 442, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 111, + 442, + 129 + ], + "spans": [ + { + "bbox": [ + 173, + 111, + 442, + 129 + ], + "type": "text", + "content": "Beta-Tuned Timestep Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "spans": [ + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "text", + "content": "Tianyi Zheng" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "text", + "content": ", Peng-Tao Jiang" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "text", + "content": ", Ben Wan" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "text", + "content": ", Hao Zhang" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "text", + "content": ", Jinwei Chen" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "text", + "content": ", Jia Wang" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "text", + "content": ", and Bo Li" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 173, + 149, + 442, + 175 + ], + "type": "inline_equation", + "content": "\\boxtimes" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 197, + 182, + 416, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 182, + 416, + 230 + ], + "spans": [ + { + "bbox": [ + 197, + 182, + 416, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 182, + 416, + 230 + ], + "type": "text", + "content": " Shanghai Jiao Tong University, Shanghai, China {tyzheng, burn-w, jiawang}@sjtu.edu.cn " + }, + { + "bbox": [ + 197, + 182, + 416, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 197, + 182, + 416, + 230 + ], + "type": "text", + "content": " vivo Mobile Communication Co., Ltd, China {pt.jiang,haozhang,jinwei.chen,libra}@vivo.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 270, + 455, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 270, + 455, + 435 + ], + "spans": [ + { + "bbox": [ + 159, + 270, + 455, + 435 + ], + "type": "text", + "content": "Abstract. Diffusion models have received a lot of attention in the field of generation due to their ability to produce high-quality samples. However, several recent studies indicate that treating all distributions equally in diffusion model training is sub-optimal. In this paper, we conduct an in-depth theoretical analysis of the forward process of diffusion models. Our findings reveal that the distribution variations are non-uniform throughout the diffusion process and the most drastic variations in distribution occur in the initial stages. Consequently, simple uniform timestep sampling strategy fail to align with these properties, potentially leading to suboptimal training of diffusion models. To address this, we propose the Beta-Tuned Timestep Diffusion Model (B-TTDM), which devises a timestep sampling strategy based on the beta distribution. By choosing the correct parameters, B-TTDM aligns the timestep sampling distribution with the properties of the forward diffusion process. Extensive experiments on different benchmark datasets validate the effectiveness of B-TTDM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 446, + 449, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 446, + 449, + 457 + ], + "spans": [ + { + "bbox": [ + 160, + 446, + 449, + 457 + ], + "type": "text", + "content": "Keywords: Generative Models " + }, + { + "bbox": [ + 160, + 446, + 449, + 457 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 446, + 449, + 457 + ], + "type": "text", + "content": " Diffusion Models " + }, + { + "bbox": [ + 160, + 446, + 449, + 457 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 446, + 449, + 457 + ], + "type": "text", + "content": " Beta Distribution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 478, + 230, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 230, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 504, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 598 + ], + "type": "text", + "content": "Diffusion models have gained widespread application across various domains in recent years, such as text-to-image generation [5,43,50], text generation [31,52], video synthesis [21] and AI security [6,12-14,32], primarily due to their ability to produce high-quality samples. During the training stage, diffusion models introduce varying levels of Gaussian noise into the data distribution, creating a range of altered distributions. The model is then trained by minimizing denoising score-matching losses across these diverse distributions. In the inference stage, the trained model reverses this noise addition process to generate new samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 599, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 635 + ], + "type": "text", + "content": "While DDPM [22], iDDPM [37] and ADM [9] have achieved significant success in image generation tasks, several studies [7, 17, 19, 48] identify that treating the distribution equally across different timesteps in the training stage could be" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 641, + 394, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 641, + 394, + 665 + ], + "spans": [ + { + "bbox": [ + 140, + 641, + 394, + 665 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 140, + 641, + 394, + 665 + ], + "type": "text", + "content": " Corresponding authors. This work was done during Tianyi Zheng's internship at vivo." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 320 + ], + "type": "text", + "content": "sub-optimal. For instance, the P2-Weight [7] method discover through empirical analysis that distributions at various timesteps in forward diffusion encapsulate distinct semantic information. Building on this insight, P2-Weight propose a re-weighted loss function to elevate the quality of generated samples. Meanwhile, Min-SNR [19] and ANT [17] find inconsistencies in optimization gradients across various timesteps and recommend a re-weighted loss strategy to mitigate these issues. Recently, E-TSDM [48] uncover that the issue of numerical instability caused by high Lipschitz constants near the zero point. Therefore, E-TSDM assigns a weight of 0 to certain timesteps near zero to mitigate the issue of numerical instability. Despite the improvement in sample quality achieved by these methods through empirical adjustments to the training objective's weight, they still utilize DDPM's uniform timestep sampling strategy. As another influencing factor, the choice of timestep sampling strategy is also closely related to the training objective but has been overlooked in previous work. In this paper, we conduct a theoretical analysis of DDPM's forward diffusion process and design a novel strategy for timestep sampling, aiming to address this overlooked aspect and further improve the quality of generated samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 322, + 483, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 483, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 483, + 430 + ], + "type": "text", + "content": "Our theoretical analysis in Theorem 2 demonstrates that distribution variations throughout the diffusion process are non-uniform, with the most significant variations occurring at the initial stages. This is experimentally supported by Figure 1, which shows non-uniform variations in both Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) [51] metrics, especially pronounced in the initial stages. Both our theoretical insights from Theorem 2 and the experimental evidence in Figure 1 underscore the non-uniform properties of forward diffusion, indicating that a uniform timestep sampling strategy may not be the most effective approach in the training stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 432, + 483, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 483, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 483, + 601 + ], + "type": "text", + "content": "Building on these findings, we propose a novel timestep sampling strategy that utilizes the beta distribution, named the Beta-Tuned Timestep Diffusion Model (B-TTDM). By choosing the correct parameters of the beta distribution, B-TTDM aligns the timestep sampling distribution with the observed properties of the diffusion process and is specifically tuned to avoid singularities at " + }, + { + "bbox": [ + 130, + 432, + 483, + 601 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 130, + 432, + 483, + 601 + ], + "type": "text", + "content": " timesteps. Extensive experiments on various datasets demonstrate that B-TTDM not only improves the quality of the generated samples but also expedites the training process. For instance, B-TTDM achieves similar FID scores on the ImageNet dataset with about " + }, + { + "bbox": [ + 130, + 432, + 483, + 601 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 432, + 483, + 601 + ], + "type": "text", + "content": " of the training iterations needed by the original approach and significantly improves generation quality after convergence. Moreover, B-TTDM can be compatible with other advanced diffusion model training methods, such as P2-Weight [7], Min-SNR [19], Debias [19], and DDPM-IP [39]. Through integration with these methods, B-TTDM can combine different advantages and generate higher quality samples. In summary, our contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 483, + 666 + ], + "type": "text", + "content": "- Through theoretical and experimental verification, we confirm that distribution variations at different timesteps during forward diffusion are non-uniform, with the most significant shifts occurring in the initial stages. This indicates that the uniform time sampling in training is sub-optimal." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "T. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 199 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "type": "text", + "content": "- We design a novel timestep sampling strategy leveraging the Beta distribution to align with the non-uniform distribution shifts observed during forward diffusion. Furthermore, our B-TTDM is compatible with other enhancement methods, allowing for flexible integration into existing improved methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 164, + 480, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 164, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 164, + 480, + 199 + ], + "type": "text", + "content": "- Extensive experiments on different benchmark datasets demonstrate that the B-TTDM method significantly improves generation quality and drastically speeds up training of the diffusion models." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 137, + 228, + 301, + 353 + ], + "blocks": [ + { + "bbox": [ + 137, + 228, + 301, + 353 + ], + "lines": [ + { + "bbox": [ + 137, + 228, + 301, + 353 + ], + "spans": [ + { + "bbox": [ + 137, + 228, + 301, + 353 + ], + "type": "image", + "image_path": "ac860bafc98cfff77481754b1e8e2308e330b625403463731bbf7bfd38393657.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 356, + 298, + 365 + ], + "lines": [ + { + "bbox": [ + 138, + 356, + 298, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 298, + 365 + ], + "type": "text", + "content": "(a) PSNR at each forward diffusion timestep." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 228, + 476, + 353 + ], + "blocks": [ + { + "bbox": [ + 313, + 228, + 476, + 353 + ], + "lines": [ + { + "bbox": [ + 313, + 228, + 476, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 476, + 353 + ], + "type": "image", + "image_path": "cecff69cba16f4b21ecc8a2341fae4bfb0627e85784655fc797f5de07440ab81.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 356, + 475, + 365 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 475, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 475, + 365 + ], + "type": "text", + "content": "(b) LPIPS at each forward diffusion timestep." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "type": "text", + "content": "Fig. 1: The PSNR and LPIPS metrics between the distribution at timestep " + }, + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "type": "text", + "content": " and the original distribution in forward diffusion under different noise schedules." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 437, + 237, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 237, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 237, + 450 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models [15,22,37] are a family of generative models that generate samples from Gaussian noise via a learned denoising process. Recently, diffusion models have outperformed flow models [10,27] and variational autoencoders [16, 28] in generating higher quality samples. Furthermore, the diffusion models also outperform GANs [18,47] in certain cases [9]. This superior performance has led to the widespread adoption of diffusion models across diverse applications. For example, the Diffusion model [43] implements the generation of images based on textual inputs [4,11], image to image translation [35,40] and 3D Avatar generation [41]. Despite their success, the speed of sample generation by diffusion models is unsatisfactory. Therefore, DPM-Solver [34] and DDIM [44] design new generation methods based on the framework of ordinary differential equation (ODE), which have greatly improved the generation speed of the diffusion model. Meanwhile, EDM [25] design a higher order Runge-Kutta sampling method to further improve the generation speed and quality. These methods significantly enhance the speed of sample generation in diffusion models, with a primary emphasis on the generative aspect. Yet, they rely on pre-trained diffusion models without addressing training optimizations. Our proposed method presents an" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beta-Tuned Timestep Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "orthogonal strategy, aiming to improve the training stage and thereby further elevate the generate quality of diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 319 + ], + "type": "text", + "content": "Improved Diffusion Models. To further improve the quality of generative samples in the training stage, ADM [9] and U-ViT [2] design new advanced architectures for diffusion models. Based on these advanced architectures, the quality of generation is significantly improved. Additionally, P2-Weight [7], Min-SNR [19] and Debias [49] methods design the re-weight loss function based on SNR in the training stage. These can reduce the bias problem in the diffusion models. To make the neural network smoother, the E-TSDM [48] sets the weight of the loss function corresponding to timesteps close to the zero to 0. Additional, ANT [17] propose multi-task learning methods to set the loss weight of different timestep in the training stage. Although these improved methods can produce better samples, they overlook the non-uniform distribution variations in the diffusion process and continue to use uniform timestep sampling during training. Our theoretical analysis indicates that this uniform time sampling strategy in the training stage is not optimal, suggesting that a more tailored timestep sampling approach could yield significant improvements in generated samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 337, + 203, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 203, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 203, + 349 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 222, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 222, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 222, + 373 + ], + "type": "text", + "content": "3.1 Background" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "text", + "content": " represent the data distribution over " + }, + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "text", + "content": ". During the forward diffusion, a sample " + }, + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "text", + "content": " is drawn from " + }, + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 130, + 379, + 482, + 416 + ], + "type": "text", + "content": " and subsequently perturbed using a stochastic differential equation (SDE) [46]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 250, + 424, + 482, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 424, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 250, + 424, + 482, + 437 + ], + "type": "interline_equation", + "content": "\\mathrm {d} x _ {t} = f \\left(x _ {t}, t\\right) \\mathrm {d} t + g (t) \\mathrm {d} w. \\tag {1}", + "image_path": "0bd37257cc9641bd0ebdba8367a1a76f4f03290ff29cfced501864a7e27fb9cf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "text", + "content": "The purpose of the forward diffusion is to transform the data distribution into simple normal Gaussian distribution " + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "text", + "content": ". Meanwhile, we define " + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "inline_equation", + "content": "q_t(x_t)" + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "text", + "content": " to represent the distribution at each time " + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "text", + "content": " during the forward diffusion process. For any time " + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "text", + "content": ", the conditional distribution " + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "inline_equation", + "content": "q_t(x_t | x_0)" + }, + { + "bbox": [ + 130, + 444, + 483, + 495 + ], + "type": "text", + "content": " is [9, 22, 37]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 237, + 504, + 482, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 504, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 237, + 504, + 482, + 517 + ], + "type": "interline_equation", + "content": "q _ {t} \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t} \\mid \\alpha_ {t} x _ {0}, \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {2}", + "image_path": "a6d979d6b98f2b45f43b200ddf37fe664aba6413e2877d2d635a89c8f0889ce7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 558 + ], + "type": "text", + "content": "In Equation (2), " + }, + { + "bbox": [ + 130, + 521, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 521, + 482, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 130, + 521, + 482, + 558 + ], + "type": "text", + "content": " satisfy " + }, + { + "bbox": [ + 130, + 521, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\alpha_{t}^{2} + \\sigma_{t}^{2} = 1" + }, + { + "bbox": [ + 130, + 521, + 482, + 558 + ], + "type": "text", + "content": ". This forward diffusion process is known as the Ornstein-Uhlenbeck (OU) process. For more general expressions, we have the following conditional distribution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 217, + 567, + 482, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 567, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 217, + 567, + 482, + 586 + ], + "type": "interline_equation", + "content": "q _ {t} \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t} \\mid e ^ {- \\frac {t}{2}} x _ {0}, (1 - e ^ {- t}) \\mathbf {I}\\right). \\tag {3}", + "image_path": "247633a6249950118005d7f3ca93e157d5ad4f13990745d456545970d71ed723.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 588, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 613 + ], + "type": "text", + "content": "The reverse diffusion transforms normal Gaussian distribution into the data distribution " + }, + { + "bbox": [ + 130, + 588, + 482, + 613 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 130, + 588, + 482, + 613 + ], + "type": "text", + "content": ". This process satisfies the following reverse SDE equation [1]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 200, + 624, + 482, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 624, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 200, + 624, + 482, + 638 + ], + "type": "interline_equation", + "content": "\\mathrm {d} x _ {t} = \\left(f \\left(x _ {t}, t\\right) - g (t) ^ {2} \\nabla_ {x} \\log q _ {t} \\left(x _ {t}\\right)\\right) \\mathrm {d} t + g (t) \\mathrm {d} w. \\tag {4}", + "image_path": "027337318c7400c049730727cdc52e7a4e1a5664dc772133a2d30fdaa5046b05.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Specially, " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\nabla_x\\log q_t(x_t)" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the score function of the probability density " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t})" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " Motivated by this correspondence, diffusion models train a neural network to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "T. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "estimate the " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "-\\sigma_t\\nabla_x\\log q_t(x_t)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ". In the inference stage, we can simulate Equation (4) for sampling and generate new samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 159, + 375, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 159, + 375, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 159, + 375, + 170 + ], + "type": "text", + "content": "3.2 Non-uniform variation in forward diffusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 179, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 479, + 228 + ], + "type": "text", + "content": "We first start by introducing and discussing our main assumptions. The only assumption we consider is that in the vicinity of high-density data, the score of the data distribution is bounded. Based on the definition of score [23, 45, 46], this assumption is quite common used [3, 30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "text", + "content": "Assumption 1 For the diffusion process described by Equation (3), assume there exists a constant " + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "inline_equation", + "content": "\\delta >0" + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "text", + "content": ". For any " + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "text", + "content": " and any point " + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "text", + "content": " in high-density regions, the score " + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "inline_equation", + "content": "\\| \\nabla_x\\log q_t(x_t)\\|" + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "text", + "content": " is bounded by " + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 236, + 480, + 274 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 264, + 282, + 350, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 282, + 350, + 295 + ], + "spans": [ + { + "bbox": [ + 264, + 282, + 350, + 295 + ], + "type": "interline_equation", + "content": "\\left\\| \\nabla_ {x} \\log q _ {t} (x _ {t}) \\right\\| \\leq \\delta .", + "image_path": "5fc64970d410c3830065bb7807effb8dd13d8b519e9c786e3878e6feb7ede83a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 479, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 479, + 329 + ], + "type": "text", + "content": "This ensures that the score of the data distribution is bounded in the vicinity of high-density data throughout the diffusion process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 337, + 479, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 479, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 479, + 372 + ], + "type": "text", + "content": "Next, we examine the Fokker-Planck equation [42] corresponding to the Ornstein-Uhlenbeck process, as depicted in Equation (5). The parameter " + }, + { + "bbox": [ + 130, + 337, + 479, + 372 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 337, + 479, + 372 + ], + "type": "text", + "content": " is drift coefficient and the parameter " + }, + { + "bbox": [ + 130, + 337, + 479, + 372 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 337, + 479, + 372 + ], + "type": "text", + "content": " is diffusion coefficient." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 382, + 481, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 382, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 231, + 382, + 481, + 406 + ], + "type": "interline_equation", + "content": "\\frac {\\partial}{\\partial t} q \\left(x _ {t}\\right) = \\theta \\nabla_ {x _ {t}} q \\left(x _ {t}\\right) + D \\nabla_ {x _ {t}} ^ {2} q \\left(x _ {t}\\right). \\tag {5}", + "image_path": "acc39d92ccf47cf65fbacef672aac99b2ce9b20705e552de90f1e817e71832db.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 409, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 479, + 445 + ], + "type": "text", + "content": "Since the " + }, + { + "bbox": [ + 130, + 409, + 479, + 445 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 409, + 479, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 409, + 479, + 445 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 409, + 479, + 445 + ], + "type": "text", + "content": " are both constants independent of " + }, + { + "bbox": [ + 130, + 409, + 479, + 445 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 409, + 479, + 445 + ], + "type": "text", + "content": ", to simplify the analysis that follows, we can let both of them be 1. Therefore, we have the following simplified Fokker-Planck equation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 238, + 455, + 481, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 455, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 238, + 455, + 481, + 479 + ], + "type": "interline_equation", + "content": "\\frac {\\partial}{\\partial t} q \\left(x _ {t}\\right) = \\nabla_ {x _ {t}} q \\left(x _ {t}\\right) + \\nabla_ {x _ {t}} ^ {2} q \\left(x _ {t}\\right). \\tag {6}", + "image_path": "71cdcad1e2326a71fafa798c68685bf46a76aee14d716f6c024552fc3b24be94.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "content": "The Fokker-Planck equation (6) sheds light on how a distribution's evolution is influenced by both time and spatial variables in the forward diffusion stage of DDPM. Our analysis further reveals that in the initial stages of diffusion, the changes in the distribution are primarily driven by the second-order derivative term, indicating a significant diffusion effect at this stage. Therefore, we have the following Lemma:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 564, + 479, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 479, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 479, + 601 + ], + "type": "text", + "content": "Lemma 1. Consider the forward diffusion process as described by Equation (3) When " + }, + { + "bbox": [ + 130, + 564, + 479, + 601 + ], + "type": "inline_equation", + "content": "t\\to 0^+" + }, + { + "bbox": [ + 130, + 564, + 479, + 601 + ], + "type": "text", + "content": ", the time derivative of the distribution " + }, + { + "bbox": [ + 130, + 564, + 479, + 601 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t})" + }, + { + "bbox": [ + 130, + 564, + 479, + 601 + ], + "type": "text", + "content": " is predominantly governed by the second-order derivative term, encapsulated in the relationship:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 261, + 609, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 609, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 261, + 609, + 481, + 633 + ], + "type": "interline_equation", + "content": "\\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right) = \\nabla_ {x _ {t}} ^ {2} q \\left(x _ {t}\\right), \\tag {7}", + "image_path": "00742a830a0f9c9219b6bd0f0d3cd184303db73fa4a3ea2824ae974844eff73e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}^2" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": " denotes the Laplacian operator with respect to " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": ", indicating the diffusion term's dominance in the evolution of " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t})" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "t\\to 0^{+}" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beta-Tuned Timestep Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "The detailed proof of the Lemma (1) is shown in Appendix Section 2.2. Lemma (1) suggests that the alteration in the distribution primarily stems from the " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}^2 q(x_t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ". Expanding on this insight, we can establish the following theorem, providing an upper bound on the alteration in distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 172, + 482, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 208 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 208 + ], + "type": "text", + "content": "Theorem 2. Consider the forward diffusion process described by Equation (7). The norm of the time derivative of the distribution " + }, + { + "bbox": [ + 130, + 172, + 482, + 208 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t})" + }, + { + "bbox": [ + 130, + 172, + 482, + 208 + ], + "type": "text", + "content": " is bounded by the following inequality:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 206, + 482, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 206, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 253, + 206, + 482, + 235 + ], + "type": "interline_equation", + "content": "\\left\\| \\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right) \\right\\| \\leq \\left\\| \\frac {q _ {t} \\left(x _ {t}\\right)}{2 t} \\right\\|. \\tag {8}", + "image_path": "b3f05b4a8008ec07b2745e7754ff165ceb46135d81c86220ab4c848bccc27e83.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 238, + 481, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 481, + 262 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 130, + 238, + 481, + 262 + ], + "type": "inline_equation", + "content": "t \\to 0^{+}" + }, + { + "bbox": [ + 130, + 238, + 481, + 262 + ], + "type": "text", + "content": ", the norm of the derivative of the distribution " + }, + { + "bbox": [ + 130, + 238, + 481, + 262 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t})" + }, + { + "bbox": [ + 130, + 238, + 481, + 262 + ], + "type": "text", + "content": " may become unbounded. This is indicated by the limit superior:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 255, + 270, + 482, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 270, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 255, + 270, + 482, + 298 + ], + "type": "interline_equation", + "content": "\\lim _ {t \\rightarrow 0 ^ {+}} \\left\\| \\frac {q _ {t} \\left(x _ {t}\\right)}{2 t} \\right\\| = \\infty . \\tag {9}", + "image_path": "a0d634d9383bb7d56aee9c879957ea44b7f702e838c0a8a8299284d331b554a9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "type": "text", + "content": "The detailed proof of the Theorem (2) is shown in Appendix Section 2.3. Theorem (2) suggests that the bound on the norm of the time derivative of " + }, + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t})" + }, + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "type": "text", + "content": " grows indefinitely as " + }, + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "type": "inline_equation", + "content": "t \\to 0^{+}" + }, + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "type": "text", + "content": ", indicating that the norm of the derivative could be unbounded. We present a straightforward example to illustrate this behavior." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "text", + "content": "Simple case illustration. Let " + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t}) = \\frac{1}{\\sqrt{4\\pi t}} e^{-\\frac{x^{2}}{4t}}" + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "text", + "content": " be a solution to differential equation (7). Meanwhile, we notice that " + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\int_{-\\infty}^{+\\infty} q_{t}(x_{t}) dx = 1" + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "inline_equation", + "content": "q_{t}(x) > 0" + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "text", + "content": ", which means " + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "inline_equation", + "content": "q_{t}(x)" + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "text", + "content": " is a probability density function. For every " + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "text", + "content": ", the norm of the time derivative of " + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "inline_equation", + "content": "q_{t}(x)" + }, + { + "bbox": [ + 130, + 369, + 481, + 422 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 225, + 430, + 482, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 430, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 225, + 430, + 482, + 464 + ], + "type": "interline_equation", + "content": "\\left\\| \\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right) \\right\\| = \\left\\| \\frac {\\sqrt {t} (- 2 t + x ^ {2}) e ^ {- \\frac {x ^ {2}}{4 t}}}{8 \\sqrt {\\pi} t ^ {3}} \\right\\|. \\tag {10}", + "image_path": "8433310bf97ce2d50d029b9bd3a79061bc92a2073e9b5e7cb8d999eb15a1c2e0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 472, + 330, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 330, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 330, + 483 + ], + "type": "text", + "content": "This expression leads to the conclusion that:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 250, + 491, + 482, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 491, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 250, + 491, + 482, + 519 + ], + "type": "interline_equation", + "content": "\\lim _ {t \\rightarrow 0 ^ {+}} \\left\\| \\frac {\\partial}{\\partial t} q _ {t} \\left(x _ {t}\\right)\\right\\| = \\infty , \\tag {11}", + "image_path": "25b62e0e29d5e80cb65debf7d536c415220186c85c926945237fae2c12328866.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "text", + "content": "indicating that the norm of the time derivative of " + }, + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "inline_equation", + "content": "q_{t}(x)" + }, + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "text", + "content": " becomes unbounded as " + }, + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "inline_equation", + "content": "t \\to 0^{+}" + }, + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "text", + "content": ". This result indicates an infinite rate of variation for the distribution " + }, + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t})" + }, + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "inline_equation", + "content": "t \\to 0^{+}" + }, + { + "bbox": [ + 130, + 527, + 481, + 562 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 563, + 481, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 481, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 481, + 623 + ], + "type": "text", + "content": "Theorem (2) highlights the non-uniform properties of distribution variations during diffusion, with the initial stages experiencing the most significant shifts. Furthermore, Proposition (1) clarifies that the distribution variations " + }, + { + "bbox": [ + 130, + 563, + 481, + 623 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t})" + }, + { + "bbox": [ + 130, + 563, + 481, + 623 + ], + "type": "text", + "content": " are non-uniform across the diffusion process, and these variations tend to flatten as the process progresses to later stages." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Proposition 1. Consider the forward diffusion process in Equation (3). The upper bound of the KL divergence between the distribution " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "q_{t}(x_{t})" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and the standard Gaussian distribution converges exponentially. i.e.," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "T. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 203, + 128, + 482, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 128, + 482, + 142 + ], + "spans": [ + { + "bbox": [ + 203, + 128, + 482, + 142 + ], + "type": "interline_equation", + "content": "\\operatorname {K L} \\left(q _ {t} \\left(x _ {t}\\right) \\| \\mathcal {N} (0, 1)\\right) \\leq e ^ {- t} \\operatorname {K L} \\left(q _ {0} \\left(x _ {0}\\right) \\| \\mathcal {N} (0, 1)\\right). \\tag {12}", + "image_path": "c2b165760ff5aeeb153c83660ec61817ab010bfd87c4a332c57f6128516cd42a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 166, + 480, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 480, + 228 + ], + "type": "text", + "content": "Based on the Theorem (2) and Proposition (1), we notice that the variations in the distribution of forward diffusion is non-uniform, with the sharpest change at the beginning of the diffusion and converging to a Gaussian distribution at an exponential rate. These theoretical insights guide the design of our timestep sampling strategy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 244, + 343, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 244, + 343, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 244, + 343, + 257 + ], + "type": "text", + "content": "3.3 Beta distribution timestep sampling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 480, + 324 + ], + "type": "text", + "content": "To align the non-uniform trend of variation, we introduce a time sampling strategy that utilizes the Beta distribution for training the diffusion model. This strategy leverages the Beta density function's capacity to take on a diverse array of shapes, controlled by the parameters " + }, + { + "bbox": [ + 130, + 265, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 265, + 480, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 265, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 265, + 480, + 324 + ], + "type": "text", + "content": ". The probability density function (pdf) of the Beta distribution is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 332, + 482, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 332, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 225, + 332, + 482, + 357 + ], + "type": "interline_equation", + "content": "f (t; \\alpha , \\beta) = \\frac {\\Gamma (\\alpha + \\beta)}{\\Gamma (\\alpha) \\Gamma (\\beta)} t ^ {\\alpha - 1} (1 - t) ^ {\\beta - 1}, \\tag {13}", + "image_path": "ae771ffae4bac518407509b0b58af982c510ab2bc4f7238adee1bd4286a4ffa1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\Gamma(z)" + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "text", + "content": " is the gamma function. Setting " + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "text", + "content": " simplifies the Beta distribution to a uniform distribution, illustrating the flexibility of Beta distribution. By adjusting the values of " + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "text", + "content": ", we can tailor the time sampling distribution within the diffusion model training to better align the desired variation properties." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "Since the beta distribution is a family of continuous probability distributions defined on the interval [0, 1]. We can use " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\frac{t}{T}" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " to map the diffusion timestep into the [0, 1] interval of the beta distribution. Then we compare the probability density function of beta distribution for different " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " in Figure 2. We find that when we set " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\alpha > 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\beta > 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ", The probability density function is U-shaped. Meanwhile, we notice that " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\beta > 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\alpha < 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ", the probability density function conforms to our expectation of a non-uniformly decreasing. However, when " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\beta > 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ", the probability density function exhibits a positively skewed shape, lacking alignment with the most significant varies in the initial diffusion stage. We provide ablation study of different " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " in the section 4.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "text", + "content": "To align with the previously identified properties, we choose to set " + }, + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\alpha < 1" + }, + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "text", + "content": ". With these parameters, the distribution adopts a reverse J-shaped profile. Consequently, the probability density function under these conditions becomes a strictly decreasing function, exhibiting a sharp change as it approaches 0. This specific shape and behavior of the pdf are instrumental in aligning the rapid initial changes and the convergence properties observed in the forward diffusion process. Therefore, we replace the uniform time sampling in the original diffusion model with this time sampling method based on the beta distribution. In addition, to prevent the probability density function from reaching infinite values as time t approaches 0, we adopt " + }, + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\frac{1}{T}" + }, + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "text", + "content": " as the time sampling weight at the initial" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "type": "text", + "content": "Beta-Tuned Timestep Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 195, + 116, + 419, + 254 + ], + "blocks": [ + { + "bbox": [ + 195, + 116, + 419, + 254 + ], + "lines": [ + { + "bbox": [ + 195, + 116, + 419, + 254 + ], + "spans": [ + { + "bbox": [ + 195, + 116, + 419, + 254 + ], + "type": "image", + "image_path": "47b782961c05d521e76e9e8d6c4c37ae95c5618af0e96ec98c2b7f0d681b09fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 263, + 474, + 276 + ], + "lines": [ + { + "bbox": [ + 137, + 263, + 474, + 276 + ], + "spans": [ + { + "bbox": [ + 137, + 263, + 474, + 276 + ], + "type": "text", + "content": "Fig. 2: Probability density function of beta distribution with different parameters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 480, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 480, + 321 + ], + "type": "text", + "content": "moment. Therefore, our method is called Beta-Tuned Timestep Diffusion Model (B-TTDM). Next, we analyse the errors introduced by this tuned approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "text", + "content": "Proposition 2. Given a timestep sampling method utilizing the B-TTDM with a Beta distribution parameterized by " + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "inline_equation", + "content": "\\alpha < 1" + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "text", + "content": ", the tuned error at the initial time position " + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "inline_equation", + "content": "\\left(\\frac{1}{T}\\right)^{\\alpha}" + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 326, + 481, + 363 + ], + "type": "text", + "content": " is the total diffusion steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 367, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 426 + ], + "type": "text", + "content": "The detailed proof of the Proposition (2) is shown in Appendix Section 2.4. Since the value of " + }, + { + "bbox": [ + 130, + 367, + 482, + 426 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 367, + 482, + 426 + ], + "type": "text", + "content": " is usually large in the training of diffusion models (e.g., " + }, + { + "bbox": [ + 130, + 367, + 482, + 426 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 130, + 367, + 482, + 426 + ], + "type": "text", + "content": "), this implies that the tuned error is quite small. Therefore, we can ignore it and still consider this as a beta distribution. Finally, the corresponding training loss can be written as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 153, + 430, + 482, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 430, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 153, + 430, + 482, + 450 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\epsilon_ {\\theta}\\right) := \\mathbb {E} _ {t \\sim \\operatorname {B e t a} (\\alpha , \\beta), \\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0}), \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\left[ \\| \\epsilon_ {\\theta} \\left(\\alpha_ {t} \\mathbf {x} _ {0} + \\sigma_ {t} \\epsilon , t\\right) - \\epsilon \\| _ {2} ^ {2} \\right]. \\tag {14}", + "image_path": "e33428da6d83c58a608ece64e341d91b0056bf2d3afcb9c18e23a3d0fe1f233e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 454, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 515 + ], + "type": "text", + "content": "Furthermore, we present a comprehensive outline of the B-TTDM method in Algorithm (1). The key difference between B-TTDM and DDPM is the strategy to timestep sampling in the training stage. Thus, B-TTDM is adaptable to most fast sampling methods and improved training methods designed for DDPM without the need for specific modifications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 535, + 293, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 535, + 293, + 547 + ], + "spans": [ + { + "bbox": [ + 133, + 535, + 293, + 547 + ], + "type": "text", + "content": "Algorithm 1 Training of B-TTDM" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 550, + 441, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 441, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 441, + 561 + ], + "type": "text", + "content": "Require: Beta distribution parameter " + }, + { + "bbox": [ + 132, + 550, + 441, + 561 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 550, + 441, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 550, + 441, + 561 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 132, + 550, + 441, + 561 + ], + "type": "text", + "content": ". Total diffusion timestep " + }, + { + "bbox": [ + 132, + 550, + 441, + 561 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 550, + 441, + 561 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 561, + 372, + 638 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 136, + 561, + 372, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 561, + 372, + 572 + ], + "spans": [ + { + "bbox": [ + 136, + 561, + 372, + 572 + ], + "type": "text", + "content": "1: Generate a beta distribution with parameters " + }, + { + "bbox": [ + 136, + 561, + 372, + 572 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 136, + 561, + 372, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 561, + 372, + 572 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 136, + 561, + 372, + 572 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 572, + 268, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 572, + 268, + 584 + ], + "spans": [ + { + "bbox": [ + 136, + 572, + 268, + 584 + ], + "type": "text", + "content": "2: Tuned " + }, + { + "bbox": [ + 136, + 572, + 268, + 584 + ], + "type": "inline_equation", + "content": "f(0; \\alpha, \\beta) = f\\left(\\frac{1}{T}; \\alpha, \\beta\\right)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 136, + 584, + 180, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 584, + 180, + 594 + ], + "spans": [ + { + "bbox": [ + 136, + 584, + 180, + 594 + ], + "type": "text", + "content": "3: repeat" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 136, + 594, + 203, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 594, + 203, + 605 + ], + "spans": [ + { + "bbox": [ + 136, + 594, + 203, + 605 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 136, + 594, + 203, + 605 + ], + "type": "inline_equation", + "content": "x_0\\sim q(x_0)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 136, + 605, + 268, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 605, + 268, + 616 + ], + "spans": [ + { + "bbox": [ + 136, + 605, + 268, + 616 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 136, + 605, + 268, + 616 + ], + "type": "inline_equation", + "content": "t\\sim \\mathrm{Beta}(\\alpha ,\\beta),\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 136, + 616, + 369, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 616, + 369, + 628 + ], + "spans": [ + { + "bbox": [ + 136, + 616, + 369, + 628 + ], + "type": "text", + "content": "6: Take a gradient descent step on " + }, + { + "bbox": [ + 136, + 616, + 369, + 628 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{\\theta}}\\| \\pmb {\\epsilon} - \\pmb {\\epsilon}_{\\pmb{\\theta}}(\\pmb {x}_t,t)\\|^2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 136, + 628, + 215, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 628, + 215, + 638 + ], + "spans": [ + { + "bbox": [ + 136, + 628, + 215, + 638 + ], + "type": "text", + "content": "7: until converged" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "T. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 121, + 296, + 228 + ], + "blocks": [ + { + "bbox": [ + 148, + 121, + 296, + 228 + ], + "lines": [ + { + "bbox": [ + 148, + 121, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 148, + 121, + 296, + 228 + ], + "type": "image", + "image_path": "263b89ac5d26157f296b9bc30c9d92b3632d15dafa117c711debc073a7dd3b6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 236, + 296, + 245 + ], + "lines": [ + { + "bbox": [ + 148, + 236, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 148, + 236, + 296, + 245 + ], + "type": "text", + "content": "(a) Comparison of FID on low-resolution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 120, + 465, + 228 + ], + "blocks": [ + { + "bbox": [ + 316, + 120, + 465, + 228 + ], + "lines": [ + { + "bbox": [ + 316, + 120, + 465, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 120, + 465, + 228 + ], + "type": "image", + "image_path": "a95b254369b24c330635592f984919eed9eef2932b8ccde8db896b6f65c4d5a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 236, + 466, + 245 + ], + "lines": [ + { + "bbox": [ + 315, + 236, + 466, + 245 + ], + "spans": [ + { + "bbox": [ + 315, + 236, + 466, + 245 + ], + "type": "text", + "content": "(b) Comparison of FID on high-resolution." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 255, + 479, + 266 + ], + "lines": [ + { + "bbox": [ + 132, + 255, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 479, + 266 + ], + "type": "text", + "content": "Fig. 3: Quantitative Comparison. Comparing FID of different resolution datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 289, + 223, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 223, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 223, + 303 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 480, + 396 + ], + "type": "text", + "content": "In this section, we first provide an implementation details of the experimental setup in section 4.1. Then, we provide quantitative comparison of B-TTDM with other state-of-the-art methods in section 4.2. In Section 4.3, we assess the robustness of B-TTDM to different fast sampling algorithms and its training efficiency in comparison to DDPM. Subsequent ablation studies are discussed in Section 4.4. Lastly, qualitative comparisons and visualization results are showcased in Section 4.5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 415, + 252, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 252, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 252, + 428 + ], + "type": "text", + "content": "4.1 Experiment Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "text", + "content": "Implementation details. All of our experiments utilize the same settings of DDPM [22]. Meanwhile we employ the same model architecture as ADM [9] since it is more advanced compared to DDPM baseline. Throughout the training stage, we maintain " + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "inline_equation", + "content": "T = 1,000" + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "text", + "content": " for all experiments. We conduct comparative experiments using the CIFAR-10 " + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "text", + "content": ", ImageNet " + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "text", + "content": ", CelebA " + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "inline_equation", + "content": "(64 \\times 64)" + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "text", + "content": " [33], FFHQ " + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "inline_equation", + "content": "(128 \\times 128)" + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "text", + "content": " [26], AFHQ-D " + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "text", + "content": " [8] and CelebAHQ " + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 130, + 436, + 482, + 591 + ], + "type": "text", + "content": " [24] datasets for unconditional image generation. To evaluate the sampling quality, we use the Frechet inception distance (FID) [20], sFID [36], improved Precision and Recall [29] as our evaluation metrics. To be more specific, we generate 50K samples for CIFAR-10, ImageNet, CelebA, FFHQ and 10K samples for AFHQ-D, CelebAHQ. In the inference stage, we use 1000 steps for CIFAR-10 and ImageNet, 300 steps for CelebA and FFHQ, and 100 steps for AFHQ-D and CelebAHQ. More details can be found in the Appendix Section 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 609, + 288, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 288, + 621 + ], + "type": "text", + "content": "4.2 Quantitative Comparison" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Comparison to Baseline. We compare our method with DDPM baseline using various benchmark datasets. The results of the experiments are shown in Figure 3, where we find that the FID of our method significantly outperforms the DDPM" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beta-Tuned Timestep Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "on datasets of different resolutions. For instance, on the FFHQ dataset, B-TTDM achieved a FID improvement of 3.79. This underscores our method's efficacy and suggests that the uniform time sampling approach previously employed for training diffusion models is not the most efficient. Notably, B-TTDM reaches an FID of 2.87 on the ImageNet dataset with a larger data size, significantly enhancing DDPM's performance and highlighting B-TTDM's effectiveness." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 164, + 240, + 451, + 319 + ], + "blocks": [ + { + "bbox": [ + 130, + 209, + 483, + 232 + ], + "lines": [ + { + "bbox": [ + 130, + 209, + 483, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 483, + 232 + ], + "type": "text", + "content": "Table 1: Quantitative Comparison. Comparison of FID and sFID with other state-of-the-art methods on different resolution datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 240, + 451, + 319 + ], + "lines": [ + { + "bbox": [ + 164, + 240, + 451, + 319 + ], + "spans": [ + { + "bbox": [ + 164, + 240, + 451, + 319 + ], + "type": "table", + "html": "
MethodCIFAR 32×32FFHQ 128×128AFHQ-D 256×256CelebAHQ 256×256
FIDsFIDFIDsFIDFIDsFIDFIDsFID
DDPM3.014.6914.0712.8917.2149.0316.3225.20
Debias3.174.5611.8611.6515.9146.7714.2321.98
Min-SNR3.234.8012.4612.8517.0347.9414.3724.87
P2-Weight3.194.7811.9612.7416.1847.8814.3822.14
B-TTDM2.894.2210.2811.3813.8145.2513.6221.93
", + "image_path": "f07fbcfb907a367ab89f71713732433570a62a7ddf341da9ea36559a08bb2a4a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "type": "text", + "content": "Comparison to other methods. Since the B-TTDM method is a non-uniform treatment of different time intervals, we contrast it with other start-of-the-art methods that employ non-uniformly re-weighted loss functions in the training stage. The result shown in Table 1 indicates that our B-TTDM perform best across multiple different resolution datasets. In addition, we find that most of these re-weighted loss methods have negative effect on low-resolution CIFAR datasets. Despite this, B-TTDM consistently achieves the best performance, underscoring its broad applicability and generalizability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 436, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 568 + ], + "type": "text", + "content": "Plug-and-Enhance. Our B-TTDM is designed to align the time sampling and the non-uniform variation observed in the data distribution during forward diffusion, setting it apart from other techniques aimed at enhancing diffusion model performance. Therefore, our method is complementary and can be integrated as a plug-in enhancement within existing strategies for optimizing the training of diffusion models. We combine our B-TTDM with other different improved training methods during the training stage following the same setting. The results are shown in Table 2, we find that our B-TTDM is beneficial for other improved methods like P2-Weight, Min-SNR and DDPM-IP. Notably, the integration of B-TTDM with DDPM-IP leads to a significant achievement, such as an FID score of 2.46 for the CIFAR dataset and a 6.69 FID score for FFHQ." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 586, + 302, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 302, + 598 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 302, + 598 + ], + "type": "text", + "content": "4.3 Fast Sampling and Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Fast Sampling. While diffusion models produce high-quality samples, their slow inference speed constrains their applicability in real-world settings. To improve this, various methods like DDIM [44], Time Respacing [9], DPM-Slover [34] and Exposure Bias [38] have been introduced. Fortunately, our B-TTDM method maintains the reliance of these acceleration techniques on the training process," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "T. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 481, + 196 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Quantitative Comparison. Comparison of FID and sFID when combining B-TTDM with other improved training methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 481, + 196 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 481, + 196 + ], + "type": "table", + "html": "
MethodCIFAR FID sFIDFFHQ FID sFIDMethodAFHQ-D FID sFIDMethodAFHQ-D FID sFID
DDPM-IP2.724.547.078.35Min-SNR17.03
+ B-TTDM2.463.956.698.10+ B-TTDM13.34
46.13
", + "image_path": "420a25da1e00a6f8a2a625f7c93c24096b18190acb0e4fe2266b17d0441839c5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 218, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 422 + ], + "type": "text", + "content": "allowing for their straightforward integration into our approach. The fast sampling outcomes of B-TTDM and DDPM models are contrasted in Table 3. We find that the B-TTDM method significantly outperforms the DDPM method under each different fast samplers and different NFE. Thus, B-TTDM can be seamlessly applied with various fast samplers for DDPM without requiring specific tuning. Fast Training. In addition to yielding higher-quality samples, B-TTDM also accelerates the training process of diffusion models. In Figure 4, we present a comparison between B-TTDM and DDPM in terms of performance and the training iterations on the large scale ImageNet and high resolution AFHQ-D datasets. Specifically, for ImageNet, B-TTDM achieves a " + }, + { + "bbox": [ + 130, + 218, + 482, + 422 + ], + "type": "inline_equation", + "content": "4.71 \\times" + }, + { + "bbox": [ + 130, + 218, + 482, + 422 + ], + "type": "text", + "content": " acceleration in training speed to attain equivalent FID scores. Once training converges, B-TTDM's FID score significantly exceeds that of DDPM. Similarly, on the large resolution AFHQ-D dataset, B-TTDM also achieves a " + }, + { + "bbox": [ + 130, + 218, + 482, + 422 + ], + "type": "inline_equation", + "content": "3.33 \\times" + }, + { + "bbox": [ + 130, + 218, + 482, + 422 + ], + "type": "text", + "content": " training speed-up and perform better than DDPM upon convergence. We attribute this acceleration to our beta-tuned timestep sampling, which aligns the timestep sampling distribution with the properties of the diffusion process. This alignment not only speeds up the training of diffusion models but also improves the generated samples quality." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 174, + 473, + 440, + 525 + ], + "blocks": [ + { + "bbox": [ + 130, + 441, + 482, + 464 + ], + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 464 + ], + "type": "text", + "content": "Table 3: Quantitative comparison. Comparison of FID score across various Fast Samplers on AFHQ-D dataset. NFE means the number of function evaluations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 174, + 473, + 440, + 525 + ], + "lines": [ + { + "bbox": [ + 174, + 473, + 440, + 525 + ], + "spans": [ + { + "bbox": [ + 174, + 473, + 440, + 525 + ], + "type": "table", + "html": "
Fast SamplersEpsilonScalingDDIMRespacingDPM-Solver
NFE2550255025502550
MethodDDPM23.6318.5024.2218.2523.6419.8318.4917.49
B-TTDM20.8115.3421.7616.7721.7916.6315.8214.79
", + "image_path": "bf6bea86414e64997251de612b1d257549ed0fc9e32116358ef068952b454553.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 563, + 238, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 238, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 238, + 575 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Parameters of the beta distribution. The B-TTDM utilizes the Beta-tuned distribution to devise a novel time sampling strategy. This distribution is defined by two parameters, " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " predominantly influences the left side of the distribution's shape. As " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " increases, the bulk of the probability density function shifts towards the right. Conversely, " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " primarily affects the right side of the distribution's shape. To further explore their impact on diffusion model training, we experiment with various " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " combinations, with the experimental" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beta-Tuned Timestep Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 116, + 294, + 225 + ], + "blocks": [ + { + "bbox": [ + 150, + 116, + 294, + 225 + ], + "lines": [ + { + "bbox": [ + 150, + 116, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 150, + 116, + 294, + 225 + ], + "type": "image", + "image_path": "9c00b3831333798d092e5d2909f06ff13a61e9d136722dfaff608b9f537c5a87.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 228, + 262, + 239 + ], + "lines": [ + { + "bbox": [ + 181, + 228, + 262, + 239 + ], + "spans": [ + { + "bbox": [ + 181, + 228, + 262, + 239 + ], + "type": "text", + "content": "(a) ImageNet " + }, + { + "bbox": [ + 181, + 228, + 262, + 239 + ], + "type": "inline_equation", + "content": "(32\\times 32)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 115, + 463, + 225 + ], + "blocks": [ + { + "bbox": [ + 319, + 115, + 463, + 225 + ], + "lines": [ + { + "bbox": [ + 319, + 115, + 463, + 225 + ], + "spans": [ + { + "bbox": [ + 319, + 115, + 463, + 225 + ], + "type": "image", + "image_path": "3094ff394f11932ea854c19fd44a35d41db6e0e3a0a09668f7fe10c397be7b8e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 228, + 434, + 239 + ], + "lines": [ + { + "bbox": [ + 347, + 228, + 434, + 239 + ], + "spans": [ + { + "bbox": [ + 347, + 228, + 434, + 239 + ], + "type": "text", + "content": "(b) AFHQ-D " + }, + { + "bbox": [ + 347, + 228, + 434, + 239 + ], + "type": "inline_equation", + "content": "(256\\times 256)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "type": "text", + "content": "Fig. 4: FID scores concerning the number of training iterations on ImageNet " + }, + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "type": "text", + "content": " and AFHQ-D " + }, + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "type": "text", + "content": " dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 355 + ], + "type": "text", + "content": "outcomes presented in Table 4. We find that beta-tuned timestep sampling leads to significant improvements in all metrics and achieves the best performance at " + }, + { + "bbox": [ + 130, + 295, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\alpha = 0.8, \\beta = 1" + }, + { + "bbox": [ + 130, + 295, + 482, + 355 + ], + "type": "text", + "content": ". This also implies that beta-tuned timestep sampling is beneficial for generating quality improvements in diffusion models. More experimental results of different parameters can be found in the Appendix Section 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 356, + 483, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 483, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 483, + 440 + ], + "type": "text", + "content": "Noise schedules. To ensure the robustness of the B-TTDM method to different noises schedules, we try two different commonly used noise schedules, cosine and linear noise schedules [9] on the CIFAR10 dataset. We follow the previous setting and each FID and sFID score is computed using " + }, + { + "bbox": [ + 130, + 356, + 483, + 440 + ], + "type": "inline_equation", + "content": "T' = 1000" + }, + { + "bbox": [ + 130, + 356, + 483, + 440 + ], + "type": "text", + "content": " inference steps. The results of the different noise schedules experiments are presented in Table 5. The B-TTDM is significantly better than the Baseline in both FID and sFID scores, which indicates that the B-TTDM method is robust to different noise schedules." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 152, + 492, + 306, + 555 + ], + "blocks": [ + { + "bbox": [ + 149, + 460, + 293, + 483 + ], + "lines": [ + { + "bbox": [ + 149, + 460, + 293, + 483 + ], + "spans": [ + { + "bbox": [ + 149, + 460, + 293, + 483 + ], + "type": "text", + "content": "Table 4: Ablation study on the parameters of Beta Distribution." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 492, + 306, + 555 + ], + "lines": [ + { + "bbox": [ + 152, + 492, + 306, + 555 + ], + "spans": [ + { + "bbox": [ + 152, + 492, + 306, + 555 + ], + "type": "table", + "html": "
Beta DistributionAFHQ-D (256×256)
αβFIDsFIDRecallPrecision
1117.2149.030.5530.738
0.95114.1246.460.5620.805
0.90113.8545.470.6120.752
0.80113.8145.250.6120.819
0.70114.7847.560.5640.817
", + "image_path": "ce3fccab5d676d38d5777fa9bc1c3c7dcacaaaf33280205dde4618618c2c2c9e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 342, + 495, + 467, + 552 + ], + "blocks": [ + { + "bbox": [ + 339, + 463, + 463, + 484 + ], + "lines": [ + { + "bbox": [ + 339, + 463, + 463, + 484 + ], + "spans": [ + { + "bbox": [ + 339, + 463, + 463, + 484 + ], + "type": "text", + "content": "Table 5: Ablation study on the Noise Schedule on CIFAR." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 495, + 467, + 552 + ], + "lines": [ + { + "bbox": [ + 342, + 495, + 467, + 552 + ], + "spans": [ + { + "bbox": [ + 342, + 495, + 467, + 552 + ], + "type": "table", + "html": "
NoiseScheduleDDPMB-TTDM
LinearFID3.533.28
sFID4.784.41
CosineFID3.012.89
sFID4.694.22
", + "image_path": "5c9dc9dcbbfcf9eb20380241abb2c45bd3df315b4158b1651123205acc6b023c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Other timestep adjust method. In this section, we compare various methods for adjusting the timestep on CIFAR and AFHQ datasets. For instance, the ANT-UW [17] method employs Uncertainty Weighting for different timestep based on multi-task learning with varying variances. Additionally, we compare timestep sampling methods based on alternative non-uniform distributions, such as exponential distributions (DDPM-E) and those where " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta > 1" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " (B-TTDM" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{\\alpha}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "). The experimental results are presented in Table 6. It is evident that" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "T. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 146, + 449, + 224 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 6: Quantitative comparison. Ablation study on different non-uniform timestep sampling method in the training stage of diffusion model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 146, + 449, + 224 + ], + "lines": [ + { + "bbox": [ + 167, + 146, + 449, + 224 + ], + "spans": [ + { + "bbox": [ + 167, + 146, + 449, + 224 + ], + "type": "table", + "html": "
MethodCIFAR (32×32)AFHQ (256×256)
FIDsFIDRecallPrecisionFIDsFIDRecallPrecision
DDPM3.014.690.6000.68217.2149.030.5530.738
DDPM-E2.954.660.5890.69515.1248.380.5090.703
ANT-UW2.934.450.5960.68714.9647.020.5980.794
B-TTDMα2.984.530.6000.69215.0046.980.6020.768
B-TTDM2.894.220.6020.69513.8145.250.6120.819
", + "image_path": "b652c6cc019836197a328ba4d2e676722032fc17e0d1b9e4d9193dbbd17538cf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 283 + ], + "type": "text", + "content": "all these enhanced methods exhibit some degree of improvement in generation quality compare to DDPM. However, the B-TTDM method demonstrates the most significant enhancement in all evaluation metrics, underscoring its effectiveness." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 151, + 317, + 462, + 393 + ], + "blocks": [ + { + "bbox": [ + 206, + 306, + 253, + 314 + ], + "lines": [ + { + "bbox": [ + 206, + 306, + 253, + 314 + ], + "spans": [ + { + "bbox": [ + 206, + 306, + 253, + 314 + ], + "type": "text", + "content": "Longer Training" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 317, + 462, + 393 + ], + "lines": [ + { + "bbox": [ + 151, + 317, + 462, + 393 + ], + "spans": [ + { + "bbox": [ + 151, + 317, + 462, + 393 + ], + "type": "image", + "image_path": "3f02b6a6fcc46ef9beab80dd8088504d8725ebef752a5cfaa491a008947f24e7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 410, + 482, + 455 + ], + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 455 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison. The generation results from DDPM and B-TTDM on FFHQ dataset. Images in each column are sampled from 50K, 100K, 300K and 500K training iterations. B-TTDM generates significantly better quality than DDPM with the same number of training iterations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 498, + 280, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 280, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 280, + 511 + ], + "type": "text", + "content": "4.5 Qualitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Generation results. We present a comparison of the generation results between various DDPMs and B-TTDMs in Figure 5. As depicted, images generated by B-TTDM exhibit markedly superior detail compared to those generated by DDPM. Furthermore, B-TTDMs manage to produce clear images by the 100K training iterations, whereas the quality of DDPM generated images at this stage remains relatively poor. Furthermore, we find that after 500K training iterations, B-TTDM still have better generation quality than DDPM in details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Reconstruct results. Building upon previous analyses, which suggest that the distribution exhibits greater variability in the initial stages of diffusion, we compare the generation capabilities of B-TTDM and DDPM for this significant difference. To achieve this, we input identical noise-containing samples to models trained for the same iterations and reconstructed the outputs. The results in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beta-Tuned Timestep Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 115, + 202, + 160 + ], + "blocks": [ + { + "bbox": [ + 157, + 115, + 202, + 160 + ], + "lines": [ + { + "bbox": [ + 157, + 115, + 202, + 160 + ], + "spans": [ + { + "bbox": [ + 157, + 115, + 202, + 160 + ], + "type": "image", + "image_path": "fc446349999942ea5161653ad6aaa1cdde305a09055731a4375e05ea82b116cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 161, + 202, + 206 + ], + "blocks": [ + { + "bbox": [ + 157, + 161, + 202, + 206 + ], + "lines": [ + { + "bbox": [ + 157, + 161, + 202, + 206 + ], + "spans": [ + { + "bbox": [ + 157, + 161, + 202, + 206 + ], + "type": "image", + "image_path": "7b585033f94b8a86bef404694d364aea1ba29e54d0bdb9801ab1e372040e8fd7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 207, + 186, + 214 + ], + "lines": [ + { + "bbox": [ + 174, + 207, + 186, + 214 + ], + "spans": [ + { + "bbox": [ + 174, + 207, + 186, + 214 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 204, + 116, + 249, + 160 + ], + "blocks": [ + { + "bbox": [ + 204, + 116, + 249, + 160 + ], + "lines": [ + { + "bbox": [ + 204, + 116, + 249, + 160 + ], + "spans": [ + { + "bbox": [ + 204, + 116, + 249, + 160 + ], + "type": "image", + "image_path": "dac7b92ff7f2d854c06f98b88cf13ed1f2ea7a84b65586f2b7570dbfbfdb3e8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 204, + 161, + 250, + 206 + ], + "blocks": [ + { + "bbox": [ + 204, + 161, + 250, + 206 + ], + "lines": [ + { + "bbox": [ + 204, + 161, + 250, + 206 + ], + "spans": [ + { + "bbox": [ + 204, + 161, + 250, + 206 + ], + "type": "image", + "image_path": "daa05f69e53819155fd131d7058b436b5eecffd3afb02c78e3c8e5fe3200610c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 207, + 239, + 214 + ], + "lines": [ + { + "bbox": [ + 215, + 207, + 239, + 214 + ], + "spans": [ + { + "bbox": [ + 215, + 207, + 239, + 214 + ], + "type": "text", + "content": "DDPM" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 252, + 116, + 297, + 160 + ], + "blocks": [ + { + "bbox": [ + 252, + 116, + 297, + 160 + ], + "lines": [ + { + "bbox": [ + 252, + 116, + 297, + 160 + ], + "spans": [ + { + "bbox": [ + 252, + 116, + 297, + 160 + ], + "type": "image", + "image_path": "9e2bb0d32fbedb681b7d9aed33e95eb1c5bef4129f0a316ace40a954df8fcaee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 252, + 161, + 297, + 206 + ], + "blocks": [ + { + "bbox": [ + 252, + 161, + 297, + 206 + ], + "lines": [ + { + "bbox": [ + 252, + 161, + 297, + 206 + ], + "spans": [ + { + "bbox": [ + 252, + 161, + 297, + 206 + ], + "type": "image", + "image_path": "0ff623cf41352abca1a0cbe19f516a6dcc1e6bfee19a8085e1c03a9220c47839.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 259, + 207, + 291, + 215 + ], + "lines": [ + { + "bbox": [ + 259, + 207, + 291, + 215 + ], + "spans": [ + { + "bbox": [ + 259, + 207, + 291, + 215 + ], + "type": "text", + "content": "B-TTDM" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 225, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 247 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison. Reconstruction results based on the same noisy inputs on FFHQ dataset." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 116, + 362, + 160 + ], + "blocks": [ + { + "bbox": [ + 317, + 116, + 362, + 160 + ], + "lines": [ + { + "bbox": [ + 317, + 116, + 362, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 362, + 160 + ], + "type": "image", + "image_path": "4748509463d6cde1ec5aa5eb8a3735813b9586c1e45e58abbd4278d168f8cd0b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 161, + 361, + 206 + ], + "blocks": [ + { + "bbox": [ + 317, + 161, + 361, + 206 + ], + "lines": [ + { + "bbox": [ + 317, + 161, + 361, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 161, + 361, + 206 + ], + "type": "image", + "image_path": "5102a8bbdd1d8fea7303520407baa8c230b3dab391942e4663c1d1bac0059f53.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 207, + 345, + 215 + ], + "lines": [ + { + "bbox": [ + 333, + 207, + 345, + 215 + ], + "spans": [ + { + "bbox": [ + 333, + 207, + 345, + 215 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 366, + 116, + 411, + 160 + ], + "blocks": [ + { + "bbox": [ + 366, + 116, + 411, + 160 + ], + "lines": [ + { + "bbox": [ + 366, + 116, + 411, + 160 + ], + "spans": [ + { + "bbox": [ + 366, + 116, + 411, + 160 + ], + "type": "image", + "image_path": "871fad37211a72a1a924c03dcbc5ccd65c4f860248a334d20058640af0dc0640.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 366, + 161, + 411, + 206 + ], + "blocks": [ + { + "bbox": [ + 366, + 161, + 411, + 206 + ], + "lines": [ + { + "bbox": [ + 366, + 161, + 411, + 206 + ], + "spans": [ + { + "bbox": [ + 366, + 161, + 411, + 206 + ], + "type": "image", + "image_path": "55f0a861aac0a3c4bb8de2a6decc57b3d22f729683bebac2b8ba7fd470876a1c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 207, + 400, + 215 + ], + "lines": [ + { + "bbox": [ + 375, + 207, + 400, + 215 + ], + "spans": [ + { + "bbox": [ + 375, + 207, + 400, + 215 + ], + "type": "text", + "content": "DDPM" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 415, + 116, + 460, + 160 + ], + "blocks": [ + { + "bbox": [ + 415, + 116, + 460, + 160 + ], + "lines": [ + { + "bbox": [ + 415, + 116, + 460, + 160 + ], + "spans": [ + { + "bbox": [ + 415, + 116, + 460, + 160 + ], + "type": "image", + "image_path": "89b049e5525748b55a3f7680a4e40d95baeef9dd18c379ccb1a7fa446ee57cdd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 415, + 161, + 459, + 206 + ], + "blocks": [ + { + "bbox": [ + 415, + 161, + 459, + 206 + ], + "lines": [ + { + "bbox": [ + 415, + 161, + 459, + 206 + ], + "spans": [ + { + "bbox": [ + 415, + 161, + 459, + 206 + ], + "type": "image", + "image_path": "2ae3c5d992445768451eba9bf81b00a14de5462dc651f946413c0a6eac592903.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 207, + 452, + 215 + ], + "lines": [ + { + "bbox": [ + 421, + 207, + 452, + 215 + ], + "spans": [ + { + "bbox": [ + 421, + 207, + 452, + 215 + ], + "type": "text", + "content": "B-TTDM" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "type": "text", + "content": "Figure 6 indicate that the reconstructions generated by B-TTDM closely resemble the original samples, whereas those generated by DDPM exhibit more differences. Unconditional generation results. Figure 7 provides a qualitative comparison of the unconditional image generation capabilities of B-TTDM and DDPM across the different datasets. The Figure 7 reveals that samples generated by DDPM often exhibit noticeable color biases, whereas samples generated by B-TTDM are free from such biases and show a marked improvement in quality compared to DDPM. This underscores the effectiveness of B-TTDM. More unconditional generation results can be found in the Appendix Section 4." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 136, + 398, + 260, + 484 + ], + "blocks": [ + { + "bbox": [ + 136, + 398, + 260, + 484 + ], + "lines": [ + { + "bbox": [ + 136, + 398, + 260, + 484 + ], + "spans": [ + { + "bbox": [ + 136, + 398, + 260, + 484 + ], + "type": "image", + "image_path": "d47ffde84d45c1552ab68a2f17eed8a533b2c9cc80dbdc8fbfbf00ae01006744.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 268, + 398, + 388, + 485 + ], + "blocks": [ + { + "bbox": [ + 268, + 398, + 388, + 485 + ], + "lines": [ + { + "bbox": [ + 268, + 398, + 388, + 485 + ], + "spans": [ + { + "bbox": [ + 268, + 398, + 388, + 485 + ], + "type": "image", + "image_path": "20ec56947c0b59e9b9885edb1b78ec3bbaa5a30ec98a4deca63499e59866d972.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 498, + 447, + 510 + ], + "lines": [ + { + "bbox": [ + 165, + 498, + 447, + 510 + ], + "spans": [ + { + "bbox": [ + 165, + 498, + 447, + 510 + ], + "type": "text", + "content": "Fig. 7: Qualitative comparison. Unconditional generation results." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 395, + 398, + 474, + 485 + ], + "blocks": [ + { + "bbox": [ + 395, + 398, + 474, + 485 + ], + "lines": [ + { + "bbox": [ + 395, + 398, + 474, + 485 + ], + "spans": [ + { + "bbox": [ + 395, + 398, + 474, + 485 + ], + "type": "image", + "image_path": "a42300ea8f58c508ea46d71ab7a027a65411765abe966156e0e8db30e22c1672.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 547, + 220, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 220, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 220, + 560 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this paper, through both experimental exploration and theoretical analysis, we conclude that the distribution varies non-uniformly in forward diffusion. Moreover, the most drastic variations occur in the initial stages of the forward diffusion process. These insights suggest that the uniform timestep sampling is sub-optimal for training diffusion models. Building on these findings, we design the B-TTDM method, which aligns timestep sampling with these inherent properties. Our extensive experiments demonstrate that B-TTDM significantly enhances the quality of model generation while also accelerating the training process." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "T. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 162 + ], + "type": "text", + "content": "This work was supported in part by NSFC under Grant 61927809 and in part by STCSM under Grant 22DZ2229005." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 198, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 198, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 198, + 192 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 201, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 201, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 480, + 224 + ], + "type": "text", + "content": "1. Anderson, B.D.: Reverse-time diffusion equation models. Stochastic Processes and their Applications 12(3), 313-326 (1982)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "text", + "content": "2. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR. pp. 22669-22679. IEEE (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "3. Block, A., Mroueh, Y., Rakhlin, A., Ross, J.: Fast mixing of multi-scale Langevin dynamics under the manifold hypothesis. CoRR abs/2006.11166 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "text", + "content": "4. Cai, R., Song, Z., Guan, D., Chen, Z., Luo, X., Yi, C., Kot, A.: Benchlmm: Benchmarking cross-style visual capability of large multimodal models. ECCV (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "text", + "content": "5. Chen, Z., Li, B., Wu, S., Jiang, K., Ding, S., Zhang, W.: Content-based unrestricted adversarial attack. NeurIPS 36 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 320, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 342 + ], + "type": "text", + "content": "6. Chen, Z., Li, B., Xu, J., Wu, S., Ding, S., Zhang, W.: Towards practical certifiable patch defense with vision transformer. In: CVPR. pp. 15148-15158 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "text", + "content": "7. Choi, J., Lee, J., Shin, C., Kim, S., Kim, H., Yoon, S.: Perception prioritized training of diffusion models. In: CVPR. pp. 11462-11471. IEEE (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 363, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 385 + ], + "type": "text", + "content": "8. Choi, Y., Uh, Y., Yoo, J., Ha, J.: Stargan v2: Diverse image synthesis for multiple domains. In: CVPR. pp. 8185-8194. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 385, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 480, + 407 + ], + "type": "text", + "content": "9. Dhariwal, P., Nichol, A.Q.: Diffusion models beat gans on image synthesis. In: NeurIPS. pp. 8780-8794 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 407, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 428 + ], + "type": "text", + "content": "10. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real NVP. In: ICLR. OpenReview.net (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "text", + "content": "1. Fan, K., Tang, J., Cao, W., Yi, R., Li, M., Gong, J., Zhang, J., Wang, Y., Wang, C., Ma, L.: Freemotion: A unified framework for number-free text-to-motion synthesis ECCV (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "text", + "content": "2. Fang, B., Li, B., Wu, S., Yi, R., Ding, S., Ma, L.: Re-thinking data availability attacks against deep neural networks. CVPR (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 480, + 514 + ], + "type": "text", + "content": "3. Fang, B., Li, B., Wu, S., Zheng, T., Ding, S., Yi, R., Ma, L.: Towards generalizable data protection with transferable unlearnable examples. CoRR abs/2305.11191 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 514, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 536 + ], + "type": "text", + "content": "4. Ge, X., Liu, X., Yu, Z., Shi, J., Qi, C., Li, J., Kalviainen, H.: Diffas: Face anti-spoofing via generative diffusion models. In: ECCV (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 557 + ], + "type": "text", + "content": "5. Geng, C., Han, T., Jiang, P.T., Zhang, H., Chen, J., Hauberg, S., Li, B.: Improving adversarial energy-based model via diffusion process. ICML (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 557, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 579 + ], + "type": "text", + "content": "6. Geng, C., Wang, J., Gao, Z., Frellsen, J., Hauberg, S.: Bounds all around: training energy-based models with bidirectional bounds. NeurIPS (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "type": "text", + "content": "7. Go, H., Lee, Y., Lee, S., Oh, S., Moon, H., Choi, S.: Addressing negative transfer in diffusion models. NeurIPS 36 (2024)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "8. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.C., Bengio, Y.: Generative adversarial nets. In: NIPS. pp. 2672–2680 (2014)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "9. Hang, T., Gu, S., Li, C., Bao, J., Chen, D., Hu, H., Geng, X., Guo, B.: Efficient diffusion training via min-snr weighting strategy. In: ICCV. pp. 7407-7417. IEEE (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beta-Tuned Timestep Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS. pp. 6626-6637 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "21. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A.A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models. CoRR abs/2210.02303 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "23. Hyvarinen, A.: Estimation of non-normalized statistical models by score matching. J. Mach. Learn. Res. 6, 695-709 (2005)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "24. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. CoRR abs/1710.10196 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "25. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. In: NeurIPS (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "26. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. PAMI 43(12), 4217-4228 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "text", + "content": "27. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. In: NeurIPS. pp. 10236-10245 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 316, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 481, + 326 + ], + "type": "text", + "content": "28. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "29. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. In: NeurIPS. pp. 3929-3938 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "30. Lee, H., Lu, J., Tan, Y.: Convergence for score-based generative modeling with polynomial complexity. In: NeurIPS (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "31. Li, X., Thickstun, J., Gulrajani, I., Liang, P., Hashimoto, T.B.: Diffusion-lm improves controllable text generation. In: NeurIPS (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "32. Liu, Y., Chen, Y., Dai, W., Gou, M., Huang, C.T., Xiong, H.: Source-free domain adaptation with domain generalized pretraining for face anti-spoofing. PAMI (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "text", + "content": "33. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: ICCV. pp. 3730-3738. IEEE Computer Society (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "34. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In: NeurIPS (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "35. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "36. Nash, C., Menick, J., Dieleman, S., Battaglia, P.W.: Generating images with sparse representations. In: ICML. vol. 139, pp. 7958-7968. PMLR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 502, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 522 + ], + "type": "text", + "content": "37. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. vol. 139, pp. 8162-8171. PMLR (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "38. Ning, M., Li, M., Su, J., Salah, A.A., Ertugrul, I.O.: Elucidating the exposure bias in diffusion models. ICLR (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "39. Ning, M., Sangineto, E., Porrello, A., Calderara, S., Cucchiara, R.: Input perturbation reduces exposure bias in diffusion models. In: ICML (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "40. Parmar, G., Singh, K.K., Zhang, R., Li, Y., Lu, J., Zhu, J.: Zero-shot image-to-image translation. In: SIGGRAPH (Conference Paper Track). pp. 11:1-11:11. ACM (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 620 + ], + "type": "text", + "content": "41. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 621, + 408, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 408, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 408, + 632 + ], + "type": "text", + "content": "42. Risken, H., Risken, H.: Fokker-planck equation. Springer (1996)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "43. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10674-10685. IEEE (2022)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "T. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 347 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "44. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR. OpenReview.net (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "45. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS. pp. 11895-11907 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "text", + "content": "46. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR. OpenReview.net (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 216 + ], + "type": "text", + "content": "47. Tao, S., Wang, J.: Alleviation of gradient exploding in gans: Fake can be real. In: CVPR. pp. 1188-1197. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "text", + "content": "48. Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., Cheng, F.: Eliminating lipschitz singularities in diffusion models. ICLR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "text", + "content": "49. Yu, H., Shen, L., Huang, J., Zhou, M., Li, H., Zhao, F.: Debias the training of diffusion models. CoRR abs/2310.08442 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "50. Zeng, W., Yan, Y., Zhu, Q., Chen, Z., Chu, P., Zhao, W., Yang, X.: Infusion: Preventing customized text-to-image diffusion from overfitting (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "text", + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "type": "text", + "content": "52. Zhu, Z., Wei, Y., Wang, J., Gan, Z., Zhang, Z., Wang, L., Hua, G., Wang, L., Liu, Z., Hu, H.: Exploring discrete diffusion models for image captioning. CoRR abs/2211.11694 (2022)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beta-Tuned Timestep Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_content_list.json b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0efd3f63c3368b545a19f7c6aba596fb6754ec6e --- /dev/null +++ b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_content_list.json @@ -0,0 +1,1766 @@ +[ + { + "type": "text", + "text": "Betrayed by Attention: A Simple yet Effective Approach for Self-supervised Video Object Segmentation", + "text_level": 1, + "bbox": [ + 238, + 140, + 767, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuangrui Ding $^{1*}$ , Rui Qian $^{1*}$ , Haohang Xu $^{2}$ , Dahua Lin $^{1,3,4\\dagger}$ , and Hongkai Xiong $^{2}$", + "bbox": [ + 295, + 236, + 705, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Chinese University of Hong Kong, Hong Kong, China", + "bbox": [ + 297, + 277, + 702, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shanghai Jiao Tong University, Shanghai, China", + "bbox": [ + 333, + 292, + 669, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shanghai Artificial Intelligence Laboratory, Shanghai, China", + "bbox": [ + 294, + 306, + 709, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Centre for Perceptual and Interactive Intelligence (CPII), Hong Kong, China {ds023, qr021}@ie.cuhk.edu.hk", + "bbox": [ + 238, + 321, + 764, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we propose a simple yet effective approach for self-supervised video object segmentation (VOS). Previous self-supervised VOS techniques majorly resort to auxiliary modalities or utilize iterative slot attention to assist in object discovery, which restricts their general applicability. To deal with these challenges, we develop a simplified architecture that capitalizes on the emerging objectness from DINO-pretrained Transformers, bypassing the need for additional modalities or slot attention. Our key insight is that the inherent structural dependencies present in DINO-pretrained Transformers can be leveraged to establish robust spatio-temporal correspondences in videos. Furthermore, simple clustering on this correspondence cue is sufficient to yield competitive segmentation results. Specifically, we first introduce a single spatio-temporal Transformer block to process the frame-wise DINO features and establish spatio-temporal dependencies in the form of self-attention. Subsequently, utilizing these attention maps, we implement hierarchical clustering to generate object segmentation masks. To train the spatio-temporal block in a fully self-supervised manner, we employ semantic and dynamic motion consistency coupled with entropy normalization. Our method demonstrates state-of-the-art performance across three multi-object video segmentation tasks. Specifically, we achieve over 5 points of improvement in terms of FG-ARI on complex real-world DAVIS-17-Unsupervised and YouTube-VIS-19 compared to the previous best result. The code and checkpoint are released at https://github.com/shvdiwnkozbw/SSL-UVOS.", + "bbox": [ + 259, + 401, + 743, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Video Object Segmentation $\\cdot$ Self-supervised Learning", + "bbox": [ + 261, + 750, + 712, + 763 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution", + "bbox": [ + 217, + 810, + 367, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author. Email: dhlin@ie.cuhk.edu.hk", + "bbox": [ + 220, + 824, + 580, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/617556ba29e3068a86f00013c0bda63f615c151c7e85f5ca448fc00d6db67c48.jpg", + "image_caption": [ + "(a) Video Frame" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 312, + 260 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/94dc76dc03c4752b3232035c31e681a3ca1c088b99fa37e08eeb34db2d5fb825.jpg", + "image_caption": [ + "(b) DINO Attention" + ], + "image_footnote": [], + "bbox": [ + 313, + 143, + 406, + 260 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/118cf47bd5adbaa47eee8f770320b14d420d09017c790a417f11d48bc22635a8.jpg", + "image_caption": [ + "Fig. 1: Attention leaks the object's position! We visualize the self-attention maps of different queries (■ prompt) from the video sequence (a). The frame-wise DINO attention maps (b) highlight image regions corresponding to the queried object. A randomly initialized spatio-temporal Transformer block on top of DINO produces noisy spatio-temporal attention maps (c) that coarsely track objects over time. Our method diminishes noise in the learned spatio-temporal attention maps (d) which derive temporally coherent object segmentation." + ], + "image_footnote": [], + "bbox": [ + 410, + 143, + 503, + 260 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7881613c9aab4294a133be65b35c520741e98fd9e1e522999a8c976d6c41debd.jpg", + "image_caption": [ + "(c) Random Initialized Spatio-temporal Attention" + ], + "image_footnote": [], + "bbox": [ + 503, + 143, + 596, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ratio-temporal Attention", + "bbox": [ + 509, + 262, + 581, + 271 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/484b88b39daec6bdba722480e537bb1d81e7180df283dbb26e11cc62c3a253f0.jpg", + "image_caption": [ + "(d) Learned Spatio-temporal Attention" + ], + "image_footnote": [], + "bbox": [ + 598, + 143, + 692, + 260 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4560323e676730cf9c4d3f8dc1c02873cb478f00c885c5f3948551d46a00285a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 143, + 785, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 412, + 375, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Representing the visual scene with objects as the basic elements has been long acknowledged as a core cognitive capability of an intelligent agent. In the realm of computer vision, Video Object Segmentation (VOS) tasks require a model to segment and track specified objects within a video sequence, striving to emulate this foundational function. This task holds significant importance in various real-world vision systems, including but not limited to, autonomous driving [18] and surveillance security [43]. However, traditional methods for VOS [6,44] typically follow a fully supervised paradigm and entail substantial costs for obtaining pixel-level per-frame annotations. This factor significantly limits their feasibility for large-scale applications. Thus, the researchers turn their attention to self-supervised learning [8,9,11,12,23,38,47,50,66], a more efficient approach that leverages unlabeled data.", + "bbox": [ + 212, + 446, + 787, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its promising potential, self-supervised VOS poses significant challenges. Firstly, previous methods tend to incorporate additional signals such as optical flow [36, 68, 71], or depth cues [16, 29], to furnish object-related clues. However, these auxiliary signals within videos can often be difficult to extract and unreliable when it comes to multi-object scenarios, thus limiting their availability. Secondly, most existing models [1,48,74] adopt slot attention [39] to parse the frame into objects. Nevertheless, the requirement for a predefined quantity of learnable slot queries limits its adaptability to in-the-wild data, especially when the number of objects is unknown. To address these limitations, we recall the recent self-supervised learning technique DINO [7, 42], which demonstrates emerging objectness in the attention maps of pretrained Vision Transformer (ViT) [14]. As depicted in Fig. 1(b), DINO attention maps encode spatial dependencies between diverse patches. These dependencies present different patterns for various objects, providing abundant cues for object segmentation.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 361, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This phenomenon motivates us with an intriguing possibility: Can we utilize the object-aware attributes of DINO to learn robust spatio-temporal dependencies and thereby produce coherent object segmentations in videos?", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we delve into the potential of this idea with a simple architecture. Based on DINO pretrained ViT, we introduce a single spatio-temporal Transformer block to process frame-wise DINO features and calculate spatiotemporal attention maps. Remarkably, these attention maps in Fig. 1(c) showcase the ability to coarsely discriminate and track object parts even with random initialization. Based on this observation, we can roughly estimate the patches belonging to the same object over time. Then, we encourage both semantic and motion alignment among patches belonging to the same object while distinguishing those from different objects in a fully self-supervised manner. The training significantly diminishes noise in the learned spatio-temporal dependencies as displayed in Fig. 1(d). In inference, we directly apply Hierarchical Clustering [28] to these spatio-temporal attention maps to derive object segmentation masks within a sequence. Interestingly, this naive clustering method yields unexpectedly competitive results. Our method greatly outperforms slot attention-based methods on various datasets, e.g., 9 points improvement over SOLV [1] on DAVIS-17, 12 points improvement over SMTC [48] on YTVIS-19. We argue that clustering on the whole spatio-temporal attention maps, instead of frame-by-frame feature clustering as in slot attention ensures the temporal coherence of produced segmentations and guarantees the object's permanence.", + "bbox": [ + 212, + 191, + 785, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our proposed method BA (short for Betrayed by Attention) offers three distinct advantages: (1) BA solely requires RGB frames and operates entirely under self-supervision without the need for external modality. This characteristic renders our method highly accessible and adaptable in real-world scenarios; (2) The architecture is remarkably simple and systematically efficient. We only introduce a single learnable Transformer block and harness a parameter-free clustering algorithm to discover the objects. It facilitates easy generalization to multi-object scenes, without the constraints imposed by slot attention; (3) Our method achieves state-of-the-art results across a bunch of unsupervised video multi-object segmentation benchmarks, such as MOVi-E [19], DAVIS-17-Unsupervised [46], and YouTube-VIS-19 [69]. Notably, we surpass the previous state-of-the-art method, TimeT [52], by 6.6 points on YTVIS-19 and by 6.1 points on DAVIS-17 in terms of the FG-ARI metric.", + "bbox": [ + 212, + 479, + 787, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 700, + 387, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Object Segmentation aims to segment objects coherently in a video sequence [2,6,15,17,20,21,25,27,33,44]. In video object segmentation (VOS), there are two prevalent protocols for evaluating the learned models: Semi-supervised VOS and Unsupervised VOS. In semi-supervised VOS, the algorithm is provided with the object masks in the first frame, and required to track them in subsequent frames. In contrast, unsupervised VOS aims to identify and segment salient objects from the background without any specific reference. In this paper, we fo", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Betrayed by Attention", + "bbox": [ + 580, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cus on the more challenging unsupervised VOS setting without using any kind of manual annotations in either training or inference. Recently, there emerges a line of self-supervised algorithms for unsupervised VOS [10, 29, 34, 58, 64, 68, 70-72]. CIS [71] facilitates fully unsupervised motion segmentation, discarding object mask supervision during training. By formulating a min-max game of mutual information, the generator is motivated to create segmentations that effectively distinguish foreground objects from the background. AMD [37] minimizes the warping synthesis error to train appearance and motion pathways without any supervision. GWM [10] uses a single RGB image as input with optical flow acting as supervision to highlight moving areas. MG [68] and OCLR [65] solely leverage optical flow as input to generate object-centric layered representations, with each layer indicating a potential object. Despite the impressive performance of state-of-the-art self-supervised VOS methods [1, 48, 52] in discovering single objects, they still face limitations when it comes to solving multi-object discovery tasks. In contrast, by leveraging the rich object cues in attention maps, our method can discern multiple objects in real-world scenarios.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-supervised Spatio-temporal Correspondence learning usually leverages free temporal supervision signals in videos to learn representations that facilitate accurate pixel or object tracking across space and time [4, 7, 13, 24, 26, 30, 31, 35, 57, 59, 67]. Vondrick et al. [57] employ the natural temporal color coherence to train a colorization model on grayscale videos, thereby establishing fine-grained correspondence between current and future frames. CRW [26] presents a self-supervised learning approach for dense correspondence in raw videos. It uses space-time graph-based random walks and cycle consistency to implicitly supervise chains of comparisons. Hu et al. [24] independently learn semantics and temporal correspondence from two pathways and fuse them at a later stage. UME [32] respectively designs short-term appearance and long-term semantic consistency to learn generalizable correspondence. Taking a step further by integrating high-level semantics with low-level temporal correlation, SMTC [48] develops two-stage slot attention to establish dense correspondence with more emphasis on foreground objects. In this work, we explicitly initialize the spatio-temporal correspondence on top of the DINO pretrained Transformer. This initialization effectively guides the learning process with a self-supervised pixel-level consistency.", + "bbox": [ + 212, + 393, + 787, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object-aware DINO Features have shown efficacy in object localization tasks, particularly in image domain [5, 22, 41, 54, 60, 63, 73, 76]. DINO's original paper [7] reveals that the learned representations and attention maps carry substantial object cues that can be harnessed for object discovery. The groundbreaking work, LOST [54], utilizes these features from DINO to perform object segmentation by constructing a graph wherein objects are segmented using the inverse degrees of nodes. TokenCut [62] employs DINO features for applying the Normalized Cut algorithm [53], thereby obtaining foreground segments in an image. Furthermore, CutLER [60] and MOST [51] can localize multiple objects without any supervision based on the DINO features. Motivated by the successful application of DINO features in image settings, recent works have expanded", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 361, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "their use to video segmentation tasks [1,48,52,61,62]. SMTC [48] utilizes the rich semantics and correspondence cues in the DINO features to provide a reliable reference for object decomposition in consecutive frames. TimeT [52] introduces a feature forwarding process to propagate dense DINO features across time and establish temporal consistency. Most of the existing works focus on leveraging DINO feature vectors for optimization and adaptation to videos. While in our work, we lay more emphasis on the self-attention maps. We extend the framework-wise DINO attention maps to spatio-temporal maps, which explicitly encode space-time structural dependencies and help coherent object discovery in videos.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 304, + 330, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary on DINO Attention Maps", + "text_level": 1, + "bbox": [ + 214, + 335, + 578, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The self-supervised ViT, as achieved by DINO [7], has exhibited a rich emergence of objectness in the attention maps. To elucidate this, we visualize the self-attention map from the final Transformer block of the DINO pretrained ViT-S/8 as a representative. As illustrated in Fig. 1(b), an image patch exhibits a high degree of attention dependencies with patches of the same object. Conversely, image patches of distinct objects tend to present different attention distributions. Encouraged by this observation, we pose the following question: Could it be possible to directly harness these attention maps to achieve object segmentation without relying on annotations?", + "bbox": [ + 212, + 358, + 787, + 494 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2b829db8e5c8dc451d0e4114155b96f855c1fe1f5cb886391f19d19931f2e784.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Clustering MetricCosine similarityKL divergence
FAAvAv
DAVIS-1647.460.351.175.4
DAVIS-1714.726.720.939.2
", + "bbox": [ + 217, + 520, + 531, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Preliminary video object segmentation results using varying clustering metrics. $F$ represents features pretrained with DINO. $A$ denotes per-frame attention maps from DINO, while $A_{v}$ and $\\tilde{A}_v$ denote spatio-temporal attention maps produced by a randomly initialized and a final learned temporal correlator respectively, on top of DINO. The IoU ( $\\mathcal{I}$ score) is reported for both DAVIS-16 and DAVIS-17-Unsupervised datasets.", + "bbox": [ + 215, + 597, + 532, + 736 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c7ed16abeed06e7fec0e373b28216579cd1e477319c01b41e4bbe0cb7b01eeac.jpg", + "image_caption": [ + "(a) DINO Feature $F$", + "Fig.2: Visualizations of the clustering results. The left column is the results of DINO features $F$ , the right column is the results of DINO attention $A$ . $F$ results in much noisier clusters, while $A$ distinguishes different classes of objects." + ], + "image_footnote": [], + "bbox": [ + 545, + 520, + 665, + 625 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2472acf7534ce171972bb6f951e4e70ce1c6f3b69b910271f54c98ea5c10f94b.jpg", + "image_caption": [ + "(b) DINO Attention A" + ], + "image_footnote": [], + "bbox": [ + 666, + 520, + 787, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In response to this question, we conduct a straightforward initial exploration with these attention maps. We denote the attention map of the last Transformer block as $A \\in \\mathbb{R}^{HW \\times HW}$ and its output features $F \\in \\mathbb{R}^{HW \\times C}$ , where $H$ , $W$ and $C$ respectively denote the height, width, and channel dimension of the feature", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Betrayed by Attention", + "bbox": [ + 580, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0315186df5a53734e34c6da14f8a914d056d5cecd0dc4ac38ee4527f324623d.jpg", + "image_caption": [ + "Fig. 3: Our architecture BA overview. Given the video frames and a DINO pretrained Transformer, we first use a temporal correlator to construct the spatio-temporal correspondence. We then utilize these attention maps as a clustering metric and apply hierarchical clustering across all frames to generate segmentation masks. During training, for each patch, we sample a positive/negative set and assign an importance weight based on its corresponding attention map. We promote alignment within the positive set while differentiating the representations from the negative set. The final loss, normalized with importance weights, only trains the temporal correlator with the DINO ViT remaining frozen." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 782, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "map5. For an intuitive comparison, we apply unsupervised clustering on both attention maps $A$ and features $F$ to yield the object segmentation masks. For the detailed procedure, please refer to the Sec. 3.2. Then, we report frame-wise evaluation results on the video object segmentation datasets DAVIS-16 [45] and DAVIS-17-Unsupervised [46] in Table 1. Interestingly, we observe that the attention maps significantly excel in decoupling object components compared to features, exhibiting an improvement of over 10 points on both datasets in terms of IoU. Surprisingly, a spatio-temporal attention map $A_v$ generated by a randomly initialized Transformer block on top of DINO also outperforms the frame-wise DINO feature $F$ . And it is further demonstrated in Fig. 2. The DINO features lead to very noisy clusters, while attention maps can discriminate objects of different categories. We speculate this is because the attention maps preserve intact correspondences, which are more suitable for fine-grained tasks such as object grouping and discrimination. Therefore, creating spatio-temporal structural dependencies that expand beyond spatial DINO attention maps appears promising for video object segmentation tasks.", + "bbox": [ + 212, + 421, + 787, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Our Simple Architecture for VOS", + "text_level": 1, + "bbox": [ + 215, + 686, + 540, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motivated by the frame-wise preliminary result, we introduce a temporal correlator instantiated with a spatio-temporal Transformer block [3] on top of the DINO pretrained ViT backbone, to extend the self-attention calculation to the spatiotemporal domain. Specifically, given the frame-wise DINO features $\\{F_1,\\dots ,F_T\\}$ of $T$ frames, we process them through a spatio-temporal Transformer block $\\varPhi_{\\mathrm{st}}$ . It consists of one standard ViT encoder layer, with self-attention computed over", + "bbox": [ + 212, + 710, + 787, + 803 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 361, + 128 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "Here we omit the CLS token, and average the multi-head attention along the head dimension.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the spatio-temporal grids. Specifically, we concatenate the frame-wise DINO features along the temporal dimension with the addition of learnable temporal positional embeddings $\\mathbf{pe}$ , i.e., $F = \\{F_{1} + \\mathbf{pe}_{1},\\dots ,F_{T} + \\mathbf{pe}_{T}\\} \\in \\mathbb{R}^{THW\\times C}$ and then linearly project $F$ to obtain the queries $Q = f_{Q}(F)$ , keys $K = f_{K}(F)$ , and values $V = f_{V}(F)$ . Then, we can obtain the self-attention map $A_{v}\\in \\mathbb{R}^{THW\\times THW}$ and temporally fused features $F_{v}\\in \\mathbb{R}^{THW\\times C}$ :", + "bbox": [ + 214, + 146, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA _ {v} = \\mathrm {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {C}}\\right), \\quad F _ {v} = A _ {v} V = \\{\\tilde {F} _ {1}, \\dots , \\tilde {F} _ {T} \\}. \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 247, + 785, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this way, we establish spatio-temporal structural dependencies beyond frameworkwise DINO features and have the potential to improve object permanence across the entire time span.", + "bbox": [ + 214, + 291, + 784, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since $A_v[i] \\in \\mathbb{R}^{THW}$ is a proximity distribution that describes the structural dependencies between the $i$ -th patch and all other positions, we adopt symmetric KL divergence to measure the distance $\\mathcal{M} \\in \\mathbb{R}^{THW \\times THW}$ between any two patches as:", + "bbox": [ + 214, + 335, + 785, + 397 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} [ i, j ] = \\mathcal {D} _ {K L} \\left(A _ {v} [ i ] | | A [ j ]\\right) + \\mathcal {D} _ {K L} \\left(A _ {v} [ j ] | | A _ {v} [ i ]\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 410, + 785, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{KL}$ denotes KL divergence. Based on this distance matrix $\\mathcal{M}$ , we then apply Hierarchical Clustering [28] to obtain the object segmentation masks. Please consult our appendix for a detailed implementation of the hierarchical clustering algorithm. In this way, we iteratively merge $THW$ patches into $N$ centroids $A_{c} \\in \\mathbb{R}^{N \\times THW}$ . Since each centroid likely represents the spatial distribution of an object, we consider the cluster assignments as potential candidates for the object segmentation mask. Specifically, we calculate the KL divergence distance between the $A_{v}$ and $A_{c}$ as $D \\in \\mathbb{R}^{THW \\times N}$ , and use argmin operation to produce the cluster assignments $Z$ within the video sequence:", + "bbox": [ + 214, + 436, + 787, + 574 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nZ = \\operatorname {a r g m i n} (D, \\dim = 1) \\in \\{1, \\dots , N \\} ^ {T H W}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 584, + 785, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, we employ a bipartite matching mechanism, following the standard evaluation protocol in DAVIS-17 Challenge [46], to match the predicted masks with ground truth for evaluation.", + "bbox": [ + 214, + 613, + 784, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A recent study, DiffSeg [56], used hierarchical clustering to process attention maps from a stable diffusion model, but it operated solely in the image domain. Our work extends the powerful image foundation model to the spatio-temporal domain, demonstrating that accurate spatio-temporal correspondence can be learned without supervision. Additionally, our method does not rely on hierarchical clustering. As shown in Table 4, using K-means clustering also achieves competitive performance, indicating that the learned spatio-temporal correspondence is effective across various clustering strategies.", + "bbox": [ + 214, + 657, + 787, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, there are two noteworthy aspects in our clustering stage. First, in the clustering process, we only set a distance threshold hyper-parameter without specifying the number of cluster centroids and the hierarchical clustering algorithm can adaptively generate different numbers of clusters according to", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Betrayed by Attention", + "bbox": [ + 578, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the complexity of the video. Second, our reference to the entire spatio-temporal distribution enables coherent object segmentation across multiple frames. As a result, the generated segmentation masks effortlessly track corresponding objects over time without extra association procedures.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Training", + "text_level": 1, + "bbox": [ + 215, + 228, + 333, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Noticing that even a randomly initialized attention map $A_v$ shown in Figure 1(c), exhibits an initial ability to differentiate parts and track objects over time, we leverage these maps as indicators for positive and negative sets sampling. Mathematically, given the attention maps $A_v \\in \\mathbb{R}^{THW \\times THW}$ , we first reshape $A_v$ into shape $THW \\times T \\times HW$ then use top-k operation to retrieve $K_p$ position indexes in each frame with highest attention scores:", + "bbox": [ + 212, + 253, + 787, + 344 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI _ {p} = \\operatorname {t o p - k} (\\text {r e s h a p e} (A _ {v}), K _ {p}, \\dim = - 1). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 356, + 785, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this way, we obtain a position index set $I_{p} \\in \\mathbb{R}^{THW \\times T \\times K_{p}}$ , which represents positions potentially associated with the same object as the $i$ -th patch. The reshape operation ensures that we consistently extract the same number of patches in each frame. This, in turn, guarantees effective object tracking within consecutive frames and promotes consistent temporal correspondence learning. Similarly, we perform this operation to find $K_{n}$ positions in each frame with the lowest attention scores $I_{n} \\in \\mathbb{R}^{THW \\times T \\times K_{n}}$ . Each $I_{n}[i]$ represents positions least likely to belong to the same object as the $i$ -th patch. After obtaining the positive and negative index sets, we apply self-supervised correspondence learning in two aspects: semantic consistency and dynamic motion consistency.", + "bbox": [ + 212, + 382, + 787, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic consistency. Intuitively, we take the temporally fused features $F_{v} = \\{\\tilde{F}_{1},\\dots,\\tilde{F}_{T}\\} \\in \\mathbb{R}^{THW\\times C}$ as the semantic representation. We promote the semantic alignment between the features of the same object and distinguish features from different objects. Given the $i$ -th patch, we first refer to the positive pair indexes $I_{p}[i]$ in Eq. 4 to gather the features that are likely from the same object. Then we calculate the cosine distance between the query feature and all positive feature vectors, and produce the distance matrix $S_{p}[i]\\in \\mathbb{R}^{TK_{p}}$ . Similarly, we gather distinct features using $I_{n}[i]$ , and form negative distance matrix $S_{n}[i]$ . We employ a simple margin loss to enforce higher consistency between corresponding object areas:", + "bbox": [ + 212, + 534, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s}} [ i ] = \\sum_ {j = 1} ^ {T K _ {p}} \\sum_ {k = 1} ^ {T K _ {n}} \\max \\left(\\mathcal {S} _ {p} [ i, j ] - \\mathcal {S} _ {n} [ i, k ] + \\lambda_ {1}, 0\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 696, + 785, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ is the margin hyper-parameter and we set $\\lambda_{1} = 0.8$ as default.", + "bbox": [ + 212, + 750, + 732, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic motion consistency. Regarding dynamic motions, rather than calculating optical flow, inter-frame feature correlations, or latent cost volumes as motion representations, we turn to highly accessible attention maps $A_v$ . These maps exhibit spatio-temporal correlations of specific patches, serving as an effective latent representation of temporal dynamics. Therefore, it is feasible to", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "directly employ these attention maps to bolster motion consistency between corresponding objects. Similar to the semantic consistency learning, given the attention map of $i$ -th patch $A_v[i]$ as a query, we also leverage the positive (negative) pair indexes $I_p[i] \\left(I_n[i]\\right)$ to gather the attention maps of the corresponding (distinct) objects. Then we calculate the symmetric KL divergence as in Eq. 2 between the query and all positive (negative) attention maps, obtaining the distance matrix $\\mathcal{M}_p[i] \\in \\mathbb{R}^{TK_p}$ and $\\mathcal{M}_n[i] \\in \\mathbb{R}^{TK_n}$ . We utilize the same margin loss for optimization, with $\\lambda_2 = 1.0$ as the margin hyper-parameter:", + "bbox": [ + 214, + 146, + 787, + 268 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m}} [ i ] = \\sum_ {j = 1} ^ {T K _ {p}} \\sum_ {k = 1} ^ {T K _ {n}} \\max \\left(\\mathcal {M} _ {p} [ i, j ] - \\mathcal {M} _ {n} [ i, k ] + \\lambda_ {2}, 0\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 276, + 785, + 319 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Overall objectives. Besides, we evaluate the importance of each spatio-temporal patch and lay more emphasis on the informative regions. Formally, we employ the entropy of the attention maps as a measure of the information content within each patch:", + "bbox": [ + 214, + 327, + 792, + 388 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ne [ i ] = \\sum_ {k = 1} ^ {T H W} - A _ {v} [ i, k ] \\log A _ {v} [ i, k ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 397, + 785, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "with $e[i]$ denoting the entropy of $i$ -th patch. A higher entropy value indicates more ambiguous spatio-temporal correlations, which provide less information. Hence, we perform softmax normalization on the reversed entropy to produce the importance weight $w$ for each patch as:", + "bbox": [ + 214, + 445, + 787, + 506 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nw [ i ] = \\frac {\\exp (- e [ i ])}{\\sum_ {j = 1} ^ {T H W} \\exp (- e [ j ])}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 513, + 785, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Then, the overall learning objectives on semantic and motion consistency can be formulated as the weighted summation over all spatio-temporal patches:", + "bbox": [ + 214, + 558, + 787, + 590 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {T H W} w [ i ] \\left(\\mathcal {L} _ {\\mathrm {s}} [ i ] + \\mathcal {L} _ {\\mathrm {m}} [ i ]\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 598, + 785, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since we only introduce one additional temporal correlator and an entirely parameter-free clustering strategy, we exclusively tune a single standard spatiotemporal Transformer block with 1.6M parameters and freeze the DINO encoder throughout the training process. This straightforward method not only maintains a minimal number of learnable parameters but also yields remarkable results in video object segmentation.", + "bbox": [ + 214, + 645, + 787, + 737 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 757, + 375, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 786, + 334, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For training, we adopt the challenging video dataset YouTube-VIS-19 [69] for video instance segmentation. YouTube-VIS-19 consists of 2,883 high-resolution", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Betrayed by Attention", + "bbox": [ + 578, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "YouTube videos and each video usually contains multiple object instances of distinct semantics. During the inference, we evaluate our method on unsupervised multiple object segmentation. We use the synthetic MOVi-E [19] dataset and real-world datasets like DAVIS-17-Unsupervised [46] and YouTube-VIS-19 [69], reporting Foreground Adjusted Rand Index (FG-ARI) and mIoU. For MOVi-E, we integrate its training set into our training due to significant distribution shifts. For the rest, we maintain solely training on the YouTube-VIS-19 set. Note that we do not present any mask annotation to the model during either the training or inference stages.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 306, + 457, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In training, we sample $T = 3$ frames with temporal stride 4 as the input clip. Each frame is augmented with random crop and color jitter and resized to $192 \\times 384$ . We adopt DINO pretrained ViT-S/8 [7] as the frame encoder, which is then followed by a single spatio-temporal Transformer Encoder block with 8 heads as temporal correlator $\\varPhi_{\\mathrm{st}}$ . We set the number of sampled positive and negative pairs to $K_{p} = 10$ and $K_{n} = 50$ in default. For optimization, we adopt AdamW [40] with a learning rate $1 \\times 10^{-4}$ , and train the model for a total of 30k iterations with a batch size of 16. In the inference stage, our model can be flexibly applied to video sequences of arbitrary lengths. For inference, we set the distance threshold in hierarchical clustering to 1.0. We follow the standard evaluation protocol outlined in the DAVIS-17 challenge [46] to match the predicted masks with the ground truth.", + "bbox": [ + 212, + 334, + 787, + 515 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2dabd9ea438ac67f3350a874c40f922c4ff89c80db0b7387c6f57a1fedcf24fe.jpg", + "table_caption": [ + "Table 2: Quantitative results on multiple object video segmentation. For MOVi-E and YouTube-VIS-19 datasets, we report FG-ARI and mIOU. Besides, we report Region Similarity $\\mathcal{J}$ and Contour Accuracy $\\mathcal{F}$ on DAVIS-17-Unsupervised. $\\star$ The original paper of TimeT [52] includes the IoU of the background category, we rerun the evaluation to exclude background here. $\\dagger$ We use DINOv2 ViT-S/14 as our frozen frame encoder." + ], + "table_footnote": [], + "table_body": "
ModelMOVi-EYTVIS-19DAVIS-17
FG-ARImIoUFG-ARImIoUFG-ARIJ&FJF
SAVi [29]42.816.011.112.7----
STEVE [55]50.626.620.020.9----
OCLR [65]--15.932.514.7-34.6-
VideoSAUR [74]73.935.639.529.1----
SOLV [1]80.8-29.145.332.2-30.2-
SMTC [48]--31.438.833.340.536.444.6
TimeT* [52]--37.940.435.540.035.844.2
BA (ours)83.440.244.350.140.143.939.248.6
BA† (ours)84.440.744.550.141.643.739.448.0
", + "bbox": [ + 215, + 625, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 361, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparison with State-of-the-art", + "text_level": 1, + "bbox": [ + 215, + 146, + 535, + 161 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluate our method on challenging multi-object segmentation, which aims to segment different objects in a temporally coherent manner. Among the compared methods, SAVi [29] and OCLR [65] employ optical flow as supervision to discriminate motions of different objects. STEVE [55], VideoSAUR [74], SOLV [1] and SMTC [48] develop variants of slot attention with temporal constraints to bind to distinct instances. TimeT [52] propagates dense DINO features across time to enhance temporal consistency and applies clustering on learned features to produce object segmentations. In contrast, our method directly uses spatiotemporal attention maps for object segmentation. As shown in Table 2, BA achieves state-of-the-art results on both synthetic (MOVi-E) and realistic video datasets (YTVIS-19 and DAVIS-17), indicating promising generalization ability. Specifically, comparing to the prior arts, we gain an approximately 5-point advantage over both VideoSAUR [74] on YTVIS-19 and SMTC [48] and TimeT [52] on DAVIS-17 in terms of FG-ARI. The superior results under our simple architecture reveal the feasibility of utilizing the spatio-temporal dependencies to mine object cues and derive reliable video object segmentation results. Besides, we also report the results using DINOv2 ViT-S/14 as the frame encoder. Despite the larger patch size, our method still achieves competitive performance. This demonstrates the robustness of our approach to pretraining models.", + "bbox": [ + 212, + 179, + 787, + 467 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/90ec1e8bd5bfcaccaee153d5d1c1bd46daa30fa48bed3a6d416b24161f0c692b.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison with DINO [7] and SMTC [48] on DAVIS-17-Unsupervised. Different color denotes different clusters." + ], + "image_footnote": [], + "bbox": [ + 222, + 494, + 787, + 797 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Betrayed by Attention", + "bbox": [ + 580, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Comparison. To visually demonstrate the effectiveness of our proposed method, we contrast our segmentation results with those of the DINO counterpart and prior arts SMTC [48] in Fig. 4. For the DINO baseline, we cluster the patches through the attention map $A$ . As can be observed, our method is able to segment with significantly less noise. For instance, in the motocross-jump sequence, BA nearly outlines the entire motorbike and the rider, whereas DINO misses most of the bike's body and only identifies its tyre. This shows that inferring the masks by associating a collection of RGB features and constructing spatio-temporal correspondences greatly enhances the segmentation result. Additionally, with the benefit of temporal correspondence, our model can differentiate between distinct object instances with the same semantics. Consider the soapbox sequence displayed in the right column; our method distinguishes two different people - one running behind the soapbox and another sitting in the soapbox throughout the video. This level of distinction cannot be obtained from DINO's spatial attention map, further emphasizing the effectiveness of introducing rich spatio-temporal relations in video segmentation tasks.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 409, + 388, + 425 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In the ablation study, we utilize DINO ViT-S/8 as the frozen backbone. We report FG-ARI and IoU on YouTube-VIS-2019 and $\\mathcal{J}\\&\\mathcal{F}$ on DAVIS-17-Unsupervised for multi-object segmentation.", + "bbox": [ + 212, + 431, + 800, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Training objective $\\mathcal{L}_{\\mathrm{s}}$ , $\\mathcal{L}_{\\mathrm{m}}$ , and $w$ . To understand the contribution of each component to the VOS performance, we individually evaluate each loss item as depicted in Table 3. Compared to the baseline without training, the inclusion of semantic consistency loss $\\mathcal{L}_{\\mathrm{s}}$ with entropy-based weighting mechanism $w$ (second row) elevates FG-ARI from 23.4 to 30.1 on YouTube-VIS-19. Furthermore, the motion consistency loss $\\mathcal{L}_{\\mathrm{m}}$ (third row) yields even more substantial improvements, resulting in FG-ARI reaching 39.9. This aligns with our preliminary observations that dynamic motions play a more significant role than semantics in fine-grained object segmentation. Combining all components (fourth row) provides the best results, highlighting the synergy between the components. Additionally, removing the weighting function $w$ (fifth row) slightly decreases the performance on both two benchmarks, affirming the value of entropy normalization to prioritize informative regions.", + "bbox": [ + 212, + 478, + 787, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Clustering metrics and strategy. To further explore whether our clustering strategy optimally leverages the learned spatio-temporal properties, we select four types of clustering metrics for comparison. Specifically, we adopt both feature vectors and attention maps directly from DINO and those from our learned spatio-temporal Transformer block. Table 4 reveals three key observations. First, our learning paradigm consistently outperforms the off-the-shelf DINO in terms of both features and attention maps, improving all metrics by over 10 points. This strongly supports our motivation that establishing spatio-temporal correspondence is crucial for video object segmentation. Second, utilizing the intermediate attention maps directly for clustering proves to be more effective than employing the final features. We speculate this occurs because the attention maps encode", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "fine-grained structural dependencies between various positions, whereas feature vectors primarily capture high-level semantics. Third, the hierarchical clustering strategy is not mandatory for success. We also achieve similar performance by replacing it with K-means clustering, where we set the number of centroids to 8, indicating that BA distills general object-aware representations across time.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bece968fb4c424ba6dae50edd949e48d49c43c1dfb87a44eb57e6b1e7ac560f2.jpg", + "table_caption": [ + "Table 3: Ablation on different combinations of semantic consistency loss $\\mathcal{L}_{\\mathrm{s}}$ , motion consistency loss $\\mathcal{L}_{\\mathrm{m}}$ and entropy weight $w$ . The first row represents the results obtained from a randomly initialized spatio-temporal Transformer block." + ], + "table_footnote": [], + "table_body": "
LsmYTVIS-19DAVIS-17
wFG-ARImIoUJ&FFG-ARI
--23.425.523.118.4
✓X30.132.429.627.5
X✓39.946.138.637.3
✓✓44.350.143.940.1
✓✓43.148.142.538.1
", + "bbox": [ + 215, + 358, + 485, + 478 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f396aa383a64a12f0029f239e21a9da4233498c8fb80b5913bf2bd5f20916ffb.jpg", + "table_caption": [ + "Table 4: Ablation on different clustering metrics. Features $F$ and attention maps $A$ are directly obtained from DINO. $F_{v}$ and $A_{v}$ are generated from our learned spatiotemporal Transformer block. $A_{v}$ and $A_{v}^{*}$ respectively denote the hierarchical clustering and K-means clustering results." + ], + "table_footnote": [], + "table_body": "
Clustering MetricYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
F22.123.816.119.4
Fv33.931.433.432.1
A30.134.829.330.5
Av44.350.143.940.1
Av*42.548.541.339.3
", + "bbox": [ + 493, + 358, + 772, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Number of frames $T$ and temporal stride. To investigate whether a wider temporal receptive field can reinforce temporal coherence and enhance object segmentations, we experiment by varying the number of input frames and temporal stride in training. Our results, displayed in Table 5, reveal that the performance significantly deteriorates when the stride is very small, such as a temporal stride of 1 in the first row. However, it begins to plateau upon reaching 4. This is because a small temporal stride leads to very subtle temporal dynamics in consecutive frames, resulting in a trivial learning task while a larger temporal stride provides richer dynamics, effectively encouraging the model to capture temporal consistency and maintain object permanence. Moreover, adding more frames results in diminishing returns, indicating that a modest temporal receptive field is adequate for sufficient temporal context in video object segmentation.", + "bbox": [ + 212, + 507, + 787, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Motion representations. In training, we use the spatio-temporal attention maps as latent motion representations for dynamic motion alignment $\\mathcal{L}_{\\mathrm{m}}$ . There are some alternatives to formulate this motion representation. Here we explore two variants in Table 6. The one is simple global feature correlation, which directly calculates the dot product between spatio-temporal feature vectors, i.e., $C_g = \\tilde{F}\\tilde{F}^\\top \\in \\mathbb{R}^{THW\\times THW}$ . The other is local feature correlation, where we sample a sliding local window with size $T_{p}\\times H_{p}\\times W_{p}$ for each grid, and obtain the local correlation matrix $C_l\\in \\mathbb{R}^{THW\\times T_pH_pW_p}$ . It is obvious that using the self-attention maps as motion representation substantially outperforms the performance of the other two methods. Compared to the global correlation, the", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Betrayed by Attention", + "bbox": [ + 580, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/dc486cf859072a5bdef078e6e385b1d1da48831a03e4be1a787cb1b396bc777a.jpg", + "table_caption": [ + "Table 5: Ablation on the number of frames $T$ and temporal stride. When sampling more than 3 frames with a larger stride, the method exhibits a diminishing effect." + ], + "table_footnote": [], + "table_body": "
TStrideYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
3137.443.637.133.6
3444.350.143.940.1
3844.250.243.340.7
5444.850.644.341.5
", + "bbox": [ + 215, + 224, + 480, + 333 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/ec33a81abf97155f93105eaccd6926b3f97e58613bc34a3b7141e043d051407d.jpg", + "table_caption": [ + "Table 6: Ablation on different motion representations in calculating motion consistency loss ${\\mathcal{L}}_{\\mathrm{m}}$ . We compare using the self-attention matrix,simple local and global feature correlation as the latent motion representation." + ], + "table_footnote": [], + "table_body": "
MotionRepYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
Local Corr.30.533.829.627.7
Global Corr.39.149.236.433.9
Self-Attention44.350.143.940.1
", + "bbox": [ + 491, + 239, + 794, + 333 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "learnable parameters in attention layers enhance the modeling capacity, which enables the self-attention maps to capture more comprehensive spatio-temporal dependencies and serve as a better motion representation. The local correlation only captures temporal dynamics in limited reception fields, restricting the ability for long-term perception.", + "bbox": [ + 212, + 362, + 787, + 439 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 460, + 359, + 477 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce a simple yet powerful approach for self-supervised video object segmentation. Specifically, our proposed architecture introduces only a single spatio-temporal Transformer block that ingests frame-wise DINO features and constructs the spatio-temporal correspondence through the attention layer. During training, we adopt a dual self-supervised consistency objective that encompasses both semantic consistency and dynamic motion consistency, supplemented with entropy normalization. In inference, we utilize hierarchical clustering on spatio-temporal attention maps to generate temporally coherent object segmentation masks. This extremely straightforward approach nevertheless delivers state-of-the-art results on challenging real-world multi-object segmentation tasks such as DAVIS-17-Unsupervised and YouTube-VIS-19.", + "bbox": [ + 212, + 492, + 787, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation. While promising, BA still has room for improvement. Firstly, the current spatio-temporal attention block cannot accommodate very long window-size frames due to the quadratic memory cost. Potentially, adopting memory mechanisms [49, 75] that are capable of handling long-duration data may prove more suitable. Secondly, neither pixel-level constraints nor spatio-temporal dependencies consider the granularity in the video. Incorporating multi-scale features [4] may present a promising direction.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Social Negative Impact. Although our method produces coherent video object segmentation without annotation, it could also enhance deepfake video technologies, making it easier to produce false videos for malicious intent. It is important to regulate the use of such technology and ensure it is leveraged for beneficial purposes.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 361, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 217, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 61931023, Grant 61932022, Grant 62371288, Grant 62320106003, Grant 62301299, Grant T2122024, Grant 62120106007, the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK.", + "bbox": [ + 212, + 180, + 787, + 272 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 297, + 321, + 313 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aydemir, G., Xie, W., Guney, F.: Self-supervised object-centric learning for videos. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=919tWtJPXe", + "2. Bao, L., Wu, B., Liu, W.: Cnn in mrf: Video object segmentation via inference in a cnn-based higher-order spatio-temporal mrf. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5977-5986 (2018)", + "3. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)", + "4. Bian, Z., Jabri, A., Efros, A.A., Owens, A.: Learning pixel trajectories with multiscale contrastive random walks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6508-6519 (2022)", + "5. Bielski, A., Favaro, P.: Move: Unsupervised movable object segmentation and detection. Advances in Neural Information Processing Systems 35, 33371-33386 (2022)", + "6. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017)", + "7. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650-9660 (2021)", + "8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)", + "9. Chen, Y., Liu, Y., Jiang, D., Zhang, X., Dai, W., Xiong, H., Tian, Q.: Sdae: Self-distillated masked autoencoder. In: European conference on computer vision. pp. 108-124. Springer (2022)", + "0. Choudhury, S., Karazija, L., Laina, I., Vedaldi, A., Rupprecht, C.: Guess What Moves: Unsupervised Video and Image Segmentation by Anticipating Motion. In: British Machine Vision Conference (BMVC) (2022)", + "11. Ding, S., Li, M., Yang, T., Qian, R., Xu, H., Chen, Q., Wang, J., Xiong, H.: Motion-aware contrastive video representation learning via foreground-background merging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9716-9726 (2022)", + "12. Ding, S., Qian, R., Xiong, H.: Dual contrastive learning for spatio-temporal representation. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5649-5658 (2022)" + ], + "bbox": [ + 218, + 332, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Betrayed by Attention", + "bbox": [ + 580, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Ding, S., Xie, W., Chen, Y., Qian, R., Zhang, X., Xiong, H., Tian, Q.: Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221 (2022)", + "14. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy", + "15. Dutt Jain, S., Xiong, B., Grauman, K.: Fusionseg: Learning to combine motion and appearance for fully automatic segmentation of generic objects in videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3664-3673 (2017)", + "16. Elsayed, G., Mahendran, A., van Steenkiste, S., Greff, K., Mozer, M.C., Kipf, T.: Savi++: Towards end-to-end object-centric learning from real-world videos. Advances in Neural Information Processing Systems 35, 28940-28954 (2022)", + "17. Fan, D.P., Wang, W., Cheng, M.M., Shen, J.: Shifting more attention to video salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8554-8564 (2019)", + "18. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020)", + "19. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., Kipf, T., Kundu, A., Lagun, D., Laradji, I., Liu, H.T.D., Meyer, H., Miao, Y., Nowrouzezahrai, D., Oztireli, C., Pot, E., Radwan, N., Rebain, D., Sabour, S., Sajjadi, M.S.M., Sela, M., Sitzmann, V., Stone, A., Sun, D., Vora, S., Wang, Z., Wu, T., Yi, K.M., Zhong, F., Tagliasacchi, A.: Kubric: a scalable dataset generator (2022)", + "20. Guo, Y., Yang, C., Rao, A., Agrawala, M., Lin, D., Dai, B.: Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933 (2023)", + "21. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=Fx2SbBgcte", + "22. Hamilton, M., Zhang, Z., Hariharan, B., Snavely, N., Freeman, W.T.: Unsupervised semantic segmentation by distilling feature correspondences. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=SaK06z6H10c", + "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)", + "24. Hu, Y., Wang, R., Zhang, K., Gao, Y.: Semantic-aware fine-grained correspondence. In: European Conference on Computer Vision. pp. 97-115. Springer (2022)", + "25. Hu, Y.T., Huang, J.B., Schwing, A.G.: Videomatch: Matching based video object segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 54-70 (2018)", + "26. Jabri, A., Owens, A., Efros, A.: Space-time correspondence as a contrastive random walk. Advances in neural information processing systems 33, 19545-19560 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Johnander, J., Danelljan, M., Brissman, E., Khan, F.S., Felsberg, M.: A generative appearance model for end-to-end video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8953-8962 (2019)", + "28. Johnson, S.C.: Hierarchical clustering schemes. Psychometrika 32(3), 241-254 (1967)", + "29. Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jonschkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=aD7uesX1GF_", + "30. Lai, Z., Xie, W.: Self-supervised learning for video correspondence flow. In: BMVC (2019)", + "31. Lai, Z., Lu, E., Xie, W.: Mast: A memory-augmented self-supervised tracker. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6479-6488 (2020)", + "32. Li, L., Wang, W., Zhou, T., Li, J., Yang, Y.: Unified mask embedding and correspondence learning for self-supervised video segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18706-18716 (2023)", + "33. Li, X., Loy, C.C.: Video object segmentation with joint re-identification and attention-aware mask propagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 90-105 (2018)", + "34. Li, X., Liu, S., De Mello, S., Wang, X., Kautz, J., Yang, M.H.: Joint-task self-supervised learning for temporal correspondence. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32. Curran Associates, Inc. (2019), https://proceedings.neurips.cc/paper/2019/file/140f6969d5213fd0ece03148e62e461e-Paper.pdf", + "35. Li, X., Liu, S., De Mello, S., Wang, X., Kautz, J., Yang, M.H.: Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems 32 (2019)", + "36. Lian, L., Wu, Z., Yu, S.X.: Bootstrapping objectness from videos by relaxed common fate and visual grouping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14582-14591 (2023)", + "37. Liu, R., Wu, Z., Yu, S., Lin, S.: The emergence of objectness: Learning zero-shot segmentation from videos. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 13137-13152. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper/2021/file/6d9cb7de5e8ac30bd5e8734bc96a35c1-Paper.pdf", + "38. Liu, Y., Wang, Y., Chen, Y., Dai, W., Li, C., Zou, J., Xiong, H.: Promoting semantic connectivity: Dual nearest neighbors contrastive learning for unsupervised domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3510-3519 (June 2023)", + "39. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)", + "40. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)", + "41. Melas-Kyriazi, L., Rupprecht, C., Laina, I., Vedaldi, A.: Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and local-" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Betrayed by Attention", + "bbox": [ + 580, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8364-8375 (2022)", + "42. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "43. Patil, P.W., Dudhane, A., Kulkarni, A., Murala, S., Gonde, A.B., Gupta, S.: An unified recurrent video object segmentation framework for various surveillance environments. IEEE Transactions on Image Processing 30, 7889-7902 (2021)", + "44. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017)", + "45. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)", + "46. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)", + "47. Qian, R., Ding, S., Liu, X., Lin, D.: Static and dynamic concepts for self-supervised video representation learning. In: European Conference on Computer Vision. pp. 145-164. Springer (2022)", + "48. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023)", + "49. Qian, R., Dong, X., Zhang, P., Zang, Y., Ding, S., Lin, D., Wang, J.: Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009 (2024)", + "50. Qian, R., Li, Y., Liu, H., See, J., Ding, S., Liu, X., Li, D., Lin, W.: Enhancing self-supervised video representation learning via multi-level feature optimization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7990-8001 (2021)", + "51. Rambhatla, S.S., Misra, I., Chellappa, R., Shrivastava, A.: Most: Multiple object localization with self-supervised transformers for object discovery. arXiv preprint arXiv:2304.05387 (2023)", + "52. Salehi, M., Gavves, E., Snoek, C.G., Asano, Y.M.: Time does tell: Self-supervised time-tuning of dense image representations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16536-16547 (2023)", + "53. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence 22(8), 888-905 (2000)", + "54. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. In: BMVC 2021-32nd British Machine Vision Conference (2021)", + "55. Singh, G., Wu, Y.F., Ahn, S.: Simple unsupervised object-centric learning for complex and naturalistic videos. Advances in Neural Information Processing Systems 35, 18181-18196 (2022)", + "56. Tian, J., Aggarwal, L., Colaco, A., Kira, Z., Gonzalez-Franco, M.: Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv preprint arXiv:2308.12469 (2023)", + "57. Vondrick, C., Shrivastava, A., Fathi, A., Guadarrama, S., Murphy, K.: Tracking emerges by colorizing videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 391-408 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "58. Wang, N., Song, Y., Ma, C., Zhou, W., Liu, W., Li, H.: Unsupervised deep tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1308-1317 (2019)", + "59. Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle-consistency of time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2566-2576 (2019)", + "60. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3124-3134 (June 2023)", + "61. Wang, X., Misra, I., Zeng, Z., Girdhar, R., Darrell, T.: Videocutler: Surprisingly simple unsupervised video instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22755-22764 (2024)", + "62. Wang, Y., Shen, X., Yuan, Y., Du, Y., Li, M., Hu, S.X., Crowley, J.L., Vaufreydaz, D.: Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "63. Wang, Y., Li, J., ZHANG, X., Shi, B., Li, C., Dai, W., Xiong, H., Tian, Q.: BarleRiA: An efficient tuning framework for referring image segmentation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=wHLDHRkmEu", + "64. Xie, C., Xiang, Y., Harchaoui, Z., Fox, D.: Object discovery in videos as foreground motion clustering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9994-10003 (2019)", + "65. Xie, J., Xie, W., Zisserman, A.: Segmenting moving objects via an object-centric layered representation. In: Advances in Neural Information Processing Systems (2022)", + "66. Xu, H., Ding, S., Zhang, X., Xiong, H., Tian, Q.: Masked autoencoders are robust data augmentors. arXiv preprint arXiv:2206.04846 (2022)", + "67. Xu, J., Wang, X.: Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10075-10085 (2021)", + "68. Yang, C., Lamdouar, H., Lu, E., Zisserman, A., Xie, W.: Self-supervised video object segmentation by motion grouping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7177-7188 (October 2021)", + "69. Yang, L., Fan, Y., Xu, N.: Video instance segmentation. In: ICCV (2019)", + "70. Yang, Y., Lai, B., Soatto, S.: Dystab: Unsupervised object segmentation via dynamic-static bootstrapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2826-2836 (June 2021)", + "71. Yang, Y., Loquercio, A., Scaramuzza, D., Soatto, S.: Unsupervised moving object detection via contextual information separation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "72. Ye, V., Li, Z., Tucker, R., Kanazawa, A., Snavely, N.: Deformable sprites for unsupervised video decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2657-2666 (2022)", + "73. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=1_jFneF07YC" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Betrayed by Attention", + "bbox": [ + 580, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "74. Zadaianchuk, A., Seitzer, M., Martius, G.: Object-centric learning for real-world videos by predicting temporal feature similarities. In: Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023) (2023)", + "75. Zhou, J., Pang, Z., Wang, Y.X.: Rmem: Restricted memory banks improve video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18602-18611 (2024)", + "76. Ziegler, A., Asano, Y.M.: Self-supervised learning of object parts for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14502-14511 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 272 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "S. Ding et al.", + "bbox": [ + 271, + 114, + 361, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_model.json b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ca785513d92fa5aa670ecab0f3047cf2216f107b --- /dev/null +++ b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_model.json @@ -0,0 +1,2638 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.768, + 0.21 + ], + "angle": 0, + "content": "Betrayed by Attention: A Simple yet Effective Approach for Self-supervised Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.237, + 0.707, + 0.269 + ], + "angle": 0, + "content": "Shuangrui Ding\\(^{1*}\\), Rui Qian\\(^{1*}\\), Haohang Xu\\(^{2}\\), Dahua Lin\\(^{1,3,4\\dagger}\\), and Hongkai Xiong\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.279, + 0.703, + 0.293 + ], + "angle": 0, + "content": "1 The Chinese University of Hong Kong, Hong Kong, China" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.293, + 0.67, + 0.308 + ], + "angle": 0, + "content": "2 Shanghai Jiao Tong University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.308, + 0.71, + 0.322 + ], + "angle": 0, + "content": "3 Shanghai Artificial Intelligence Laboratory, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.322, + 0.765, + 0.35 + ], + "angle": 0, + "content": "4 Centre for Perceptual and Interactive Intelligence (CPII), Hong Kong, China {ds023, qr021}@ie.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.402, + 0.744, + 0.735 + ], + "angle": 0, + "content": "Abstract. In this paper, we propose a simple yet effective approach for self-supervised video object segmentation (VOS). Previous self-supervised VOS techniques majorly resort to auxiliary modalities or utilize iterative slot attention to assist in object discovery, which restricts their general applicability. To deal with these challenges, we develop a simplified architecture that capitalizes on the emerging objectness from DINO-pretrained Transformers, bypassing the need for additional modalities or slot attention. Our key insight is that the inherent structural dependencies present in DINO-pretrained Transformers can be leveraged to establish robust spatio-temporal correspondences in videos. Furthermore, simple clustering on this correspondence cue is sufficient to yield competitive segmentation results. Specifically, we first introduce a single spatio-temporal Transformer block to process the frame-wise DINO features and establish spatio-temporal dependencies in the form of self-attention. Subsequently, utilizing these attention maps, we implement hierarchical clustering to generate object segmentation masks. To train the spatio-temporal block in a fully self-supervised manner, we employ semantic and dynamic motion consistency coupled with entropy normalization. Our method demonstrates state-of-the-art performance across three multi-object video segmentation tasks. Specifically, we achieve over 5 points of improvement in terms of FG-ARI on complex real-world DAVIS-17-Unsupervised and YouTube-VIS-19 compared to the previous best result. The code and checkpoint are released at https://github.com/shvdiwnkozbw/SSL-UVOS." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.75, + 0.714, + 0.765 + ], + "angle": 0, + "content": "Keywords: Video Object Segmentation \\(\\cdot\\) Self-supervised Learning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.368, + 0.825 + ], + "angle": 0, + "content": "* Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.825, + 0.581, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author. Email: dhlin@ie.cuhk.edu.hk" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.129 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.313, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.263, + 0.292, + 0.272 + ], + "angle": 0, + "content": "(a) Video Frame" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.144, + 0.407, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.263, + 0.393, + 0.271 + ], + "angle": 0, + "content": "(b) DINO Attention" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.144, + 0.504, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.424, + 0.263, + 0.582, + 0.272 + ], + "angle": 0, + "content": "(c) Random Initialized Spatio-temporal Attention" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.144, + 0.597, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.263, + 0.582, + 0.272 + ], + "angle": 0, + "content": "Ratio-temporal Attention" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.144, + 0.693, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.263, + 0.753, + 0.272 + ], + "angle": 0, + "content": "(d) Learned Spatio-temporal Attention" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.144, + 0.786, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.384 + ], + "angle": 0, + "content": "Fig. 1: Attention leaks the object's position! We visualize the self-attention maps of different queries (■ prompt) from the video sequence (a). The frame-wise DINO attention maps (b) highlight image regions corresponding to the queried object. A randomly initialized spatio-temporal Transformer block on top of DINO produces noisy spatio-temporal attention maps (c) that coarsely track objects over time. Our method diminishes noise in the learned spatio-temporal attention maps (d) which derive temporally coherent object segmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.376, + 0.43 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Representing the visual scene with objects as the basic elements has been long acknowledged as a core cognitive capability of an intelligent agent. In the realm of computer vision, Video Object Segmentation (VOS) tasks require a model to segment and track specified objects within a video sequence, striving to emulate this foundational function. This task holds significant importance in various real-world vision systems, including but not limited to, autonomous driving [18] and surveillance security [43]. However, traditional methods for VOS [6,44] typically follow a fully supervised paradigm and entail substantial costs for obtaining pixel-level per-frame annotations. This factor significantly limits their feasibility for large-scale applications. Thus, the researchers turn their attention to self-supervised learning [8,9,11,12,23,38,47,50,66], a more efficient approach that leverages unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Despite its promising potential, self-supervised VOS poses significant challenges. Firstly, previous methods tend to incorporate additional signals such as optical flow [36, 68, 71], or depth cues [16, 29], to furnish object-related clues. However, these auxiliary signals within videos can often be difficult to extract and unreliable when it comes to multi-object scenarios, thus limiting their availability. Secondly, most existing models [1,48,74] adopt slot attention [39] to parse the frame into objects. Nevertheless, the requirement for a predefined quantity of learnable slot queries limits its adaptability to in-the-wild data, especially when the number of objects is unknown. To address these limitations, we recall the recent self-supervised learning technique DINO [7, 42], which demonstrates emerging objectness in the attention maps of pretrained Vision Transformer (ViT) [14]. As depicted in Fig. 1(b), DINO attention maps encode spatial dependencies between diverse patches. These dependencies present different patterns for various objects, providing abundant cues for object segmentation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.581, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Betrayed by Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "This phenomenon motivates us with an intriguing possibility: Can we utilize the object-aware attributes of DINO to learn robust spatio-temporal dependencies and thereby produce coherent object segmentations in videos?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.786, + 0.479 + ], + "angle": 0, + "content": "In this paper, we delve into the potential of this idea with a simple architecture. Based on DINO pretrained ViT, we introduce a single spatio-temporal Transformer block to process frame-wise DINO features and calculate spatiotemporal attention maps. Remarkably, these attention maps in Fig. 1(c) showcase the ability to coarsely discriminate and track object parts even with random initialization. Based on this observation, we can roughly estimate the patches belonging to the same object over time. Then, we encourage both semantic and motion alignment among patches belonging to the same object while distinguishing those from different objects in a fully self-supervised manner. The training significantly diminishes noise in the learned spatio-temporal dependencies as displayed in Fig. 1(d). In inference, we directly apply Hierarchical Clustering [28] to these spatio-temporal attention maps to derive object segmentation masks within a sequence. Interestingly, this naive clustering method yields unexpectedly competitive results. Our method greatly outperforms slot attention-based methods on various datasets, e.g., 9 points improvement over SOLV [1] on DAVIS-17, 12 points improvement over SMTC [48] on YTVIS-19. We argue that clustering on the whole spatio-temporal attention maps, instead of frame-by-frame feature clustering as in slot attention ensures the temporal coherence of produced segmentations and guarantees the object's permanence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.678 + ], + "angle": 0, + "content": "In summary, our proposed method BA (short for Betrayed by Attention) offers three distinct advantages: (1) BA solely requires RGB frames and operates entirely under self-supervision without the need for external modality. This characteristic renders our method highly accessible and adaptable in real-world scenarios; (2) The architecture is remarkably simple and systematically efficient. We only introduce a single learnable Transformer block and harness a parameter-free clustering algorithm to discover the objects. It facilitates easy generalization to multi-object scenes, without the constraints imposed by slot attention; (3) Our method achieves state-of-the-art results across a bunch of unsupervised video multi-object segmentation benchmarks, such as MOVi-E [19], DAVIS-17-Unsupervised [46], and YouTube-VIS-19 [69]. Notably, we surpass the previous state-of-the-art method, TimeT [52], by 6.6 points on YTVIS-19 and by 6.1 points on DAVIS-17 in terms of the FG-ARI metric." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.388, + 0.718 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Video Object Segmentation aims to segment objects coherently in a video sequence [2,6,15,17,20,21,25,27,33,44]. In video object segmentation (VOS), there are two prevalent protocols for evaluating the learned models: Semi-supervised VOS and Unsupervised VOS. In semi-supervised VOS, the algorithm is provided with the object masks in the first frame, and required to track them in subsequent frames. In contrast, unsupervised VOS aims to identify and segment salient objects from the background without any specific reference. In this paper, we fo" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.129 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.389 + ], + "angle": 0, + "content": "cus on the more challenging unsupervised VOS setting without using any kind of manual annotations in either training or inference. Recently, there emerges a line of self-supervised algorithms for unsupervised VOS [10, 29, 34, 58, 64, 68, 70-72]. CIS [71] facilitates fully unsupervised motion segmentation, discarding object mask supervision during training. By formulating a min-max game of mutual information, the generator is motivated to create segmentations that effectively distinguish foreground objects from the background. AMD [37] minimizes the warping synthesis error to train appearance and motion pathways without any supervision. GWM [10] uses a single RGB image as input with optical flow acting as supervision to highlight moving areas. MG [68] and OCLR [65] solely leverage optical flow as input to generate object-centric layered representations, with each layer indicating a potential object. Despite the impressive performance of state-of-the-art self-supervised VOS methods [1, 48, 52] in discovering single objects, they still face limitations when it comes to solving multi-object discovery tasks. In contrast, by leveraging the rich object cues in attention maps, our method can discern multiple objects in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Self-supervised Spatio-temporal Correspondence learning usually leverages free temporal supervision signals in videos to learn representations that facilitate accurate pixel or object tracking across space and time [4, 7, 13, 24, 26, 30, 31, 35, 57, 59, 67]. Vondrick et al. [57] employ the natural temporal color coherence to train a colorization model on grayscale videos, thereby establishing fine-grained correspondence between current and future frames. CRW [26] presents a self-supervised learning approach for dense correspondence in raw videos. It uses space-time graph-based random walks and cycle consistency to implicitly supervise chains of comparisons. Hu et al. [24] independently learn semantics and temporal correspondence from two pathways and fuse them at a later stage. UME [32] respectively designs short-term appearance and long-term semantic consistency to learn generalizable correspondence. Taking a step further by integrating high-level semantics with low-level temporal correlation, SMTC [48] develops two-stage slot attention to establish dense correspondence with more emphasis on foreground objects. In this work, we explicitly initialize the spatio-temporal correspondence on top of the DINO pretrained Transformer. This initialization effectively guides the learning process with a self-supervised pixel-level consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Object-aware DINO Features have shown efficacy in object localization tasks, particularly in image domain [5, 22, 41, 54, 60, 63, 73, 76]. DINO's original paper [7] reveals that the learned representations and attention maps carry substantial object cues that can be harnessed for object discovery. The groundbreaking work, LOST [54], utilizes these features from DINO to perform object segmentation by constructing a graph wherein objects are segmented using the inverse degrees of nodes. TokenCut [62] employs DINO features for applying the Normalized Cut algorithm [53], thereby obtaining foreground segments in an image. Furthermore, CutLER [60] and MOST [51] can localize multiple objects without any supervision based on the DINO features. Motivated by the successful application of DINO features in image settings, recent works have expanded" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.581, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Betrayed by Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "their use to video segmentation tasks [1,48,52,61,62]. SMTC [48] utilizes the rich semantics and correspondence cues in the DINO features to provide a reliable reference for object decomposition in consecutive frames. TimeT [52] introduces a feature forwarding process to propagate dense DINO features across time and establish temporal consistency. Most of the existing works focus on leveraging DINO feature vectors for optimization and adaptation to videos. While in our work, we lay more emphasis on the self-attention maps. We extend the framework-wise DINO attention maps to spatio-temporal maps, which explicitly encode space-time structural dependencies and help coherent object discovery in videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.305, + 0.331, + 0.321 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.336, + 0.58, + 0.352 + ], + "angle": 0, + "content": "3.1 Preliminary on DINO Attention Maps" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.496 + ], + "angle": 0, + "content": "The self-supervised ViT, as achieved by DINO [7], has exhibited a rich emergence of objectness in the attention maps. To elucidate this, we visualize the self-attention map from the final Transformer block of the DINO pretrained ViT-S/8 as a representative. As illustrated in Fig. 1(b), an image patch exhibits a high degree of attention dependencies with patches of the same object. Conversely, image patches of distinct objects tend to present different attention distributions. Encouraged by this observation, we pose the following question: Could it be possible to directly harness these attention maps to achieve object segmentation without relying on annotations?" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.521, + 0.532, + 0.596 + ], + "angle": 0, + "content": "
Clustering MetricCosine similarityKL divergence
FAAvAv
DAVIS-1647.460.351.175.4
DAVIS-1714.726.720.939.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.598, + 0.534, + 0.737 + ], + "angle": 0, + "content": "Table 1: Preliminary video object segmentation results using varying clustering metrics. \\(F\\) represents features pretrained with DINO. \\(A\\) denotes per-frame attention maps from DINO, while \\(A_{v}\\) and \\(\\tilde{A}_v\\) denote spatio-temporal attention maps produced by a randomly initialized and a final learned temporal correlator respectively, on top of DINO. The IoU (\\(\\mathcal{I}\\) score) is reported for both DAVIS-16 and DAVIS-17-Unsupervised datasets." + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.521, + 0.666, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.626, + 0.651, + 0.636 + ], + "angle": 0, + "content": "(a) DINO Feature \\(F\\)" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.521, + 0.788, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.626, + 0.776, + 0.636 + ], + "angle": 0, + "content": "(b) DINO Attention A" + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.64, + 0.788, + 0.738 + ], + "angle": 0, + "content": "Fig.2: Visualizations of the clustering results. The left column is the results of DINO features \\(F\\), the right column is the results of DINO attention \\(A\\). \\(F\\) results in much noisier clusters, while \\(A\\) distinguishes different classes of objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In response to this question, we conduct a straightforward initial exploration with these attention maps. We denote the attention map of the last Transformer block as \\( A \\in \\mathbb{R}^{HW \\times HW} \\) and its output features \\( F \\in \\mathbb{R}^{HW \\times C} \\), where \\( H \\), \\( W \\) and \\( C \\) respectively denote the height, width, and channel dimension of the feature" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.129 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 3: Our architecture BA overview. Given the video frames and a DINO pretrained Transformer, we first use a temporal correlator to construct the spatio-temporal correspondence. We then utilize these attention maps as a clustering metric and apply hierarchical clustering across all frames to generate segmentation masks. During training, for each patch, we sample a positive/negative set and assign an importance weight based on its corresponding attention map. We promote alignment within the positive set while differentiating the representations from the negative set. The final loss, normalized with importance weights, only trains the temporal correlator with the DINO ViT remaining frozen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.665 + ], + "angle": 0, + "content": "map5. For an intuitive comparison, we apply unsupervised clustering on both attention maps \\( A \\) and features \\( F \\) to yield the object segmentation masks. For the detailed procedure, please refer to the Sec. 3.2. Then, we report frame-wise evaluation results on the video object segmentation datasets DAVIS-16 [45] and DAVIS-17-Unsupervised [46] in Table 1. Interestingly, we observe that the attention maps significantly excel in decoupling object components compared to features, exhibiting an improvement of over 10 points on both datasets in terms of IoU. Surprisingly, a spatio-temporal attention map \\( A_v \\) generated by a randomly initialized Transformer block on top of DINO also outperforms the frame-wise DINO feature \\( F \\). And it is further demonstrated in Fig. 2. The DINO features lead to very noisy clusters, while attention maps can discriminate objects of different categories. We speculate this is because the attention maps preserve intact correspondences, which are more suitable for fine-grained tasks such as object grouping and discrimination. Therefore, creating spatio-temporal structural dependencies that expand beyond spatial DINO attention maps appears promising for video object segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.687, + 0.542, + 0.703 + ], + "angle": 0, + "content": "3.2 Our Simple Architecture for VOS" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Motivated by the frame-wise preliminary result, we introduce a temporal correlator instantiated with a spatio-temporal Transformer block [3] on top of the DINO pretrained ViT backbone, to extend the self-attention calculation to the spatiotemporal domain. Specifically, given the frame-wise DINO features \\(\\{F_1,\\dots ,F_T\\}\\) of \\(T\\) frames, we process them through a spatio-temporal Transformer block \\(\\varPhi_{\\mathrm{st}}\\). It consists of one standard ViT encoder layer, with self-attention computed over" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Here we omit the CLS token, and average the multi-head attention along the head dimension." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.58, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Betrayed by Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "the spatio-temporal grids. Specifically, we concatenate the frame-wise DINO features along the temporal dimension with the addition of learnable temporal positional embeddings \\(\\mathbf{pe}\\), i.e., \\(F = \\{F_{1} + \\mathbf{pe}_{1},\\dots ,F_{T} + \\mathbf{pe}_{T}\\} \\in \\mathbb{R}^{THW\\times C}\\) and then linearly project \\(F\\) to obtain the queries \\(Q = f_{Q}(F)\\), keys \\(K = f_{K}(F)\\), and values \\(V = f_{V}(F)\\). Then, we can obtain the self-attention map \\(A_{v}\\in \\mathbb{R}^{THW\\times THW}\\) and temporally fused features \\(F_{v}\\in \\mathbb{R}^{THW\\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.248, + 0.786, + 0.284 + ], + "angle": 0, + "content": "\\[\nA _ {v} = \\mathrm {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {C}}\\right), \\quad F _ {v} = A _ {v} V = \\{\\tilde {F} _ {1}, \\dots , \\tilde {F} _ {T} \\}. \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.292, + 0.785, + 0.336 + ], + "angle": 0, + "content": "In this way, we establish spatio-temporal structural dependencies beyond frameworkwise DINO features and have the potential to improve object permanence across the entire time span." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.337, + 0.787, + 0.398 + ], + "angle": 0, + "content": "Since \\( A_v[i] \\in \\mathbb{R}^{THW} \\) is a proximity distribution that describes the structural dependencies between the \\( i \\)-th patch and all other positions, we adopt symmetric KL divergence to measure the distance \\( \\mathcal{M} \\in \\mathbb{R}^{THW \\times THW} \\) between any two patches as:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.411, + 0.786, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} [ i, j ] = \\mathcal {D} _ {K L} \\left(A _ {v} [ i ] | | A [ j ]\\right) + \\mathcal {D} _ {K L} \\left(A _ {v} [ j ] | | A _ {v} [ i ]\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.438, + 0.788, + 0.575 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{KL}\\) denotes KL divergence. Based on this distance matrix \\(\\mathcal{M}\\), we then apply Hierarchical Clustering [28] to obtain the object segmentation masks. Please consult our appendix for a detailed implementation of the hierarchical clustering algorithm. In this way, we iteratively merge \\(THW\\) patches into \\(N\\) centroids \\(A_{c} \\in \\mathbb{R}^{N \\times THW}\\). Since each centroid likely represents the spatial distribution of an object, we consider the cluster assignments as potential candidates for the object segmentation mask. Specifically, we calculate the KL divergence distance between the \\(A_{v}\\) and \\(A_{c}\\) as \\(D \\in \\mathbb{R}^{THW \\times N}\\), and use argmin operation to produce the cluster assignments \\(Z\\) within the video sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.585, + 0.786, + 0.603 + ], + "angle": 0, + "content": "\\[\nZ = \\operatorname {a r g m i n} (D, \\dim = 1) \\in \\{1, \\dots , N \\} ^ {T H W}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.614, + 0.785, + 0.658 + ], + "angle": 0, + "content": "Then, we employ a bipartite matching mechanism, following the standard evaluation protocol in DAVIS-17 Challenge [46], to match the predicted masks with ground truth for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.788, + 0.78 + ], + "angle": 0, + "content": "A recent study, DiffSeg [56], used hierarchical clustering to process attention maps from a stable diffusion model, but it operated solely in the image domain. Our work extends the powerful image foundation model to the spatio-temporal domain, demonstrating that accurate spatio-temporal correspondence can be learned without supervision. Additionally, our method does not rely on hierarchical clustering. As shown in Table 4, using K-means clustering also achieves competitive performance, indicating that the learned spatio-temporal correspondence is effective across various clustering strategies." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Besides, there are two noteworthy aspects in our clustering stage. First, in the clustering process, we only set a distance threshold hyper-parameter without specifying the number of cluster centroids and the hierarchical clustering algorithm can adaptively generate different numbers of clusters according to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.129 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "the complexity of the video. Second, our reference to the entire spatio-temporal distribution enables coherent object segmentation across multiple frames. As a result, the generated segmentation masks effortlessly track corresponding objects over time without extra association procedures." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.229, + 0.334, + 0.244 + ], + "angle": 0, + "content": "3.3 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Noticing that even a randomly initialized attention map \\( A_v \\) shown in Figure 1(c), exhibits an initial ability to differentiate parts and track objects over time, we leverage these maps as indicators for positive and negative sets sampling. Mathematically, given the attention maps \\( A_v \\in \\mathbb{R}^{THW \\times THW} \\), we first reshape \\( A_v \\) into shape \\( THW \\times T \\times HW \\) then use top-k operation to retrieve \\( K_p \\) position indexes in each frame with highest attention scores:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.357, + 0.786, + 0.373 + ], + "angle": 0, + "content": "\\[\nI _ {p} = \\operatorname {t o p - k} (\\text {r e s h a p e} (A _ {v}), K _ {p}, \\dim = - 1). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.789, + 0.534 + ], + "angle": 0, + "content": "In this way, we obtain a position index set \\( I_{p} \\in \\mathbb{R}^{THW \\times T \\times K_{p}} \\), which represents positions potentially associated with the same object as the \\( i \\)-th patch. The reshape operation ensures that we consistently extract the same number of patches in each frame. This, in turn, guarantees effective object tracking within consecutive frames and promotes consistent temporal correspondence learning. Similarly, we perform this operation to find \\( K_{n} \\) positions in each frame with the lowest attention scores \\( I_{n} \\in \\mathbb{R}^{THW \\times T \\times K_{n}} \\). Each \\( I_{n}[i] \\) represents positions least likely to belong to the same object as the \\( i \\)-th patch. After obtaining the positive and negative index sets, we apply self-supervised correspondence learning in two aspects: semantic consistency and dynamic motion consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.789, + 0.687 + ], + "angle": 0, + "content": "Semantic consistency. Intuitively, we take the temporally fused features \\( F_{v} = \\{\\tilde{F}_{1},\\dots,\\tilde{F}_{T}\\} \\in \\mathbb{R}^{THW\\times C} \\) as the semantic representation. We promote the semantic alignment between the features of the same object and distinguish features from different objects. Given the \\( i \\)-th patch, we first refer to the positive pair indexes \\( I_{p}[i] \\) in Eq. 4 to gather the features that are likely from the same object. Then we calculate the cosine distance between the query feature and all positive feature vectors, and produce the distance matrix \\( S_{p}[i]\\in \\mathbb{R}^{TK_{p}} \\). Similarly, we gather distinct features using \\( I_{n}[i] \\), and form negative distance matrix \\( S_{n}[i] \\). We employ a simple margin loss to enforce higher consistency between corresponding object areas:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.697, + 0.787, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s}} [ i ] = \\sum_ {j = 1} ^ {T K _ {p}} \\sum_ {k = 1} ^ {T K _ {n}} \\max \\left(\\mathcal {S} _ {p} [ i, j ] - \\mathcal {S} _ {n} [ i, k ] + \\lambda_ {1}, 0\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.733, + 0.764 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) is the margin hyper-parameter and we set \\(\\lambda_{1} = 0.8\\) as default." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dynamic motion consistency. Regarding dynamic motions, rather than calculating optical flow, inter-frame feature correlations, or latent cost volumes as motion representations, we turn to highly accessible attention maps \\( A_v \\). These maps exhibit spatio-temporal correlations of specific patches, serving as an effective latent representation of temporal dynamics. Therefore, it is feasible to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.58, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Betrayed by Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "directly employ these attention maps to bolster motion consistency between corresponding objects. Similar to the semantic consistency learning, given the attention map of \\(i\\)-th patch \\(A_v[i]\\) as a query, we also leverage the positive (negative) pair indexes \\(I_p[i] \\left(I_n[i]\\right)\\) to gather the attention maps of the corresponding (distinct) objects. Then we calculate the symmetric KL divergence as in Eq. 2 between the query and all positive (negative) attention maps, obtaining the distance matrix \\(\\mathcal{M}_p[i] \\in \\mathbb{R}^{TK_p}\\) and \\(\\mathcal{M}_n[i] \\in \\mathbb{R}^{TK_n}\\). We utilize the same margin loss for optimization, with \\(\\lambda_2 = 1.0\\) as the margin hyper-parameter:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.277, + 0.786, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m}} [ i ] = \\sum_ {j = 1} ^ {T K _ {p}} \\sum_ {k = 1} ^ {T K _ {n}} \\max \\left(\\mathcal {M} _ {p} [ i, j ] - \\mathcal {M} _ {n} [ i, k ] + \\lambda_ {2}, 0\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.328, + 0.793, + 0.39 + ], + "angle": 0, + "content": "Overall objectives. Besides, we evaluate the importance of each spatio-temporal patch and lay more emphasis on the informative regions. Formally, we employ the entropy of the attention maps as a measure of the information content within each patch:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.398, + 0.786, + 0.438 + ], + "angle": 0, + "content": "\\[\ne [ i ] = \\sum_ {k = 1} ^ {T H W} - A _ {v} [ i, k ] \\log A _ {v} [ i, k ], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.446, + 0.788, + 0.507 + ], + "angle": 0, + "content": "with \\( e[i] \\) denoting the entropy of \\( i \\)-th patch. A higher entropy value indicates more ambiguous spatio-temporal correlations, which provide less information. Hence, we perform softmax normalization on the reversed entropy to produce the importance weight \\( w \\) for each patch as:" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.515, + 0.786, + 0.552 + ], + "angle": 0, + "content": "\\[\nw [ i ] = \\frac {\\exp (- e [ i ])}{\\sum_ {j = 1} ^ {T H W} \\exp (- e [ j ])}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.559, + 0.788, + 0.591 + ], + "angle": 0, + "content": "Then, the overall learning objectives on semantic and motion consistency can be formulated as the weighted summation over all spatio-temporal patches:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.599, + 0.786, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {T H W} w [ i ] \\left(\\mathcal {L} _ {\\mathrm {s}} [ i ] + \\mathcal {L} _ {\\mathrm {m}} [ i ]\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.646, + 0.788, + 0.738 + ], + "angle": 0, + "content": "Since we only introduce one additional temporal correlator and an entirely parameter-free clustering strategy, we exclusively tune a single standard spatiotemporal Transformer block with 1.6M parameters and freeze the DINO encoder throughout the training process. This straightforward method not only maintains a minimal number of learnable parameters but also yields remarkable results in video object segmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.758, + 0.377, + 0.775 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.336, + 0.801 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "For training, we adopt the challenging video dataset YouTube-VIS-19 [69] for video instance segmentation. YouTube-VIS-19 consists of 2,883 high-resolution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.129 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "YouTube videos and each video usually contains multiple object instances of distinct semantics. During the inference, we evaluate our method on unsupervised multiple object segmentation. We use the synthetic MOVi-E [19] dataset and real-world datasets like DAVIS-17-Unsupervised [46] and YouTube-VIS-19 [69], reporting Foreground Adjusted Rand Index (FG-ARI) and mIoU. For MOVi-E, we integrate its training set into our training due to significant distribution shifts. For the rest, we maintain solely training on the YouTube-VIS-19 set. Note that we do not present any mask annotation to the model during either the training or inference stages." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.308, + 0.458, + 0.323 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.789, + 0.516 + ], + "angle": 0, + "content": "In training, we sample \\( T = 3 \\) frames with temporal stride 4 as the input clip. Each frame is augmented with random crop and color jitter and resized to \\( 192 \\times 384 \\). We adopt DINO pretrained ViT-S/8 [7] as the frame encoder, which is then followed by a single spatio-temporal Transformer Encoder block with 8 heads as temporal correlator \\( \\varPhi_{\\mathrm{st}} \\). We set the number of sampled positive and negative pairs to \\( K_{p} = 10 \\) and \\( K_{n} = 50 \\) in default. For optimization, we adopt AdamW [40] with a learning rate \\( 1 \\times 10^{-4} \\), and train the model for a total of 30k iterations with a batch size of 16. In the inference stage, our model can be flexibly applied to video sequences of arbitrary lengths. For inference, we set the distance threshold in hierarchical clustering to 1.0. We follow the standard evaluation protocol outlined in the DAVIS-17 challenge [46] to match the predicted masks with the ground truth." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Table 2: Quantitative results on multiple object video segmentation. For MOVi-E and YouTube-VIS-19 datasets, we report FG-ARI and mIOU. Besides, we report Region Similarity \\(\\mathcal{J}\\) and Contour Accuracy \\(\\mathcal{F}\\) on DAVIS-17-Unsupervised. \\(\\star\\) The original paper of TimeT [52] includes the IoU of the background category, we rerun the evaluation to exclude background here. \\(\\dagger\\) We use DINOv2 ViT-S/14 as our frozen frame encoder." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.625, + 0.788, + 0.809 + ], + "angle": 0, + "content": "
ModelMOVi-EYTVIS-19DAVIS-17
FG-ARImIoUFG-ARImIoUFG-ARIJ&FJF
SAVi [29]42.816.011.112.7----
STEVE [55]50.626.620.020.9----
OCLR [65]--15.932.514.7-34.6-
VideoSAUR [74]73.935.639.529.1----
SOLV [1]80.8-29.145.332.2-30.2-
SMTC [48]--31.438.833.340.536.444.6
TimeT* [52]--37.940.435.540.035.844.2
BA (ours)83.440.244.350.140.143.939.248.6
BA† (ours)84.440.744.550.141.643.739.448.0
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.581, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Betrayed by Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.536, + 0.162 + ], + "angle": 0, + "content": "4.3 Comparison with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.468 + ], + "angle": 0, + "content": "We evaluate our method on challenging multi-object segmentation, which aims to segment different objects in a temporally coherent manner. Among the compared methods, SAVi [29] and OCLR [65] employ optical flow as supervision to discriminate motions of different objects. STEVE [55], VideoSAUR [74], SOLV [1] and SMTC [48] develop variants of slot attention with temporal constraints to bind to distinct instances. TimeT [52] propagates dense DINO features across time to enhance temporal consistency and applies clustering on learned features to produce object segmentations. In contrast, our method directly uses spatiotemporal attention maps for object segmentation. As shown in Table 2, BA achieves state-of-the-art results on both synthetic (MOVi-E) and realistic video datasets (YTVIS-19 and DAVIS-17), indicating promising generalization ability. Specifically, comparing to the prior arts, we gain an approximately 5-point advantage over both VideoSAUR [74] on YTVIS-19 and SMTC [48] and TimeT [52] on DAVIS-17 in terms of FG-ARI. The superior results under our simple architecture reveal the feasibility of utilizing the spatio-temporal dependencies to mine object cues and derive reliable video object segmentation results. Besides, we also report the results using DINOv2 ViT-S/14 as the frame encoder. Despite the larger patch size, our method still achieves competitive performance. This demonstrates the robustness of our approach to pretraining models." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.496, + 0.789, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.807, + 0.787, + 0.836 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison with DINO [7] and SMTC [48] on DAVIS-17-Unsupervised. Different color denotes different clusters." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Qualitative Comparison. To visually demonstrate the effectiveness of our proposed method, we contrast our segmentation results with those of the DINO counterpart and prior arts SMTC [48] in Fig. 4. For the DINO baseline, we cluster the patches through the attention map \\( A \\). As can be observed, our method is able to segment with significantly less noise. For instance, in the motocross-jump sequence, BA nearly outlines the entire motorbike and the rider, whereas DINO misses most of the bike's body and only identifies its tyre. This shows that inferring the masks by associating a collection of RGB features and constructing spatio-temporal correspondences greatly enhances the segmentation result. Additionally, with the benefit of temporal correspondence, our model can differentiate between distinct object instances with the same semantics. Consider the soapbox sequence displayed in the right column; our method distinguishes two different people - one running behind the soapbox and another sitting in the soapbox throughout the video. This level of distinction cannot be obtained from DINO's spatial attention map, further emphasizing the effectiveness of introducing rich spatio-temporal relations in video segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.389, + 0.426 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.802, + 0.478 + ], + "angle": 0, + "content": "In the ablation study, we utilize DINO ViT-S/8 as the frozen backbone. We report FG-ARI and IoU on YouTube-VIS-2019 and \\(\\mathcal{J}\\&\\mathcal{F}\\) on DAVIS-17-Unsupervised for multi-object segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Training objective \\(\\mathcal{L}_{\\mathrm{s}}\\), \\(\\mathcal{L}_{\\mathrm{m}}\\), and \\(w\\). To understand the contribution of each component to the VOS performance, we individually evaluate each loss item as depicted in Table 3. Compared to the baseline without training, the inclusion of semantic consistency loss \\(\\mathcal{L}_{\\mathrm{s}}\\) with entropy-based weighting mechanism \\(w\\) (second row) elevates FG-ARI from 23.4 to 30.1 on YouTube-VIS-19. Furthermore, the motion consistency loss \\(\\mathcal{L}_{\\mathrm{m}}\\) (third row) yields even more substantial improvements, resulting in FG-ARI reaching 39.9. This aligns with our preliminary observations that dynamic motions play a more significant role than semantics in fine-grained object segmentation. Combining all components (fourth row) provides the best results, highlighting the synergy between the components. Additionally, removing the weighting function \\(w\\) (fifth row) slightly decreases the performance on both two benchmarks, affirming the value of entropy normalization to prioritize informative regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Clustering metrics and strategy. To further explore whether our clustering strategy optimally leverages the learned spatio-temporal properties, we select four types of clustering metrics for comparison. Specifically, we adopt both feature vectors and attention maps directly from DINO and those from our learned spatio-temporal Transformer block. Table 4 reveals three key observations. First, our learning paradigm consistently outperforms the off-the-shelf DINO in terms of both features and attention maps, improving all metrics by over 10 points. This strongly supports our motivation that establishing spatio-temporal correspondence is crucial for video object segmentation. Second, utilizing the intermediate attention maps directly for clustering proves to be more effective than employing the final features. We speculate this occurs because the attention maps encode" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.581, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Betrayed by Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "fine-grained structural dependencies between various positions, whereas feature vectors primarily capture high-level semantics. Third, the hierarchical clustering strategy is not mandatory for success. We also achieve similar performance by replacing it with K-means clustering, where we set the number of centroids to 8, indicating that BA distills general object-aware representations across time." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.248, + 0.479, + 0.346 + ], + "angle": 0, + "content": "Table 3: Ablation on different combinations of semantic consistency loss \\(\\mathcal{L}_{\\mathrm{s}}\\), motion consistency loss \\(\\mathcal{L}_{\\mathrm{m}}\\) and entropy weight \\(w\\). The first row represents the results obtained from a randomly initialized spatio-temporal Transformer block." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.359, + 0.486, + 0.479 + ], + "angle": 0, + "content": "
LsmYTVIS-19DAVIS-17
wFG-ARImIoUJ&FFG-ARI
--23.425.523.118.4
✓X30.132.429.627.5
X✓39.946.138.637.3
✓✓44.350.143.940.1
✓✓43.148.142.538.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.491, + 0.249, + 0.782, + 0.348 + ], + "angle": 0, + "content": "Table 4: Ablation on different clustering metrics. Features \\(F\\) and attention maps \\(A\\) are directly obtained from DINO. \\(F_{v}\\) and \\(A_{v}\\) are generated from our learned spatiotemporal Transformer block. \\(A_{v}\\) and \\(A_{v}^{*}\\) respectively denote the hierarchical clustering and K-means clustering results." + }, + { + "type": "table", + "bbox": [ + 0.495, + 0.359, + 0.774, + 0.48 + ], + "angle": 0, + "content": "
Clustering MetricYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
F22.123.816.119.4
Fv33.931.433.432.1
A30.134.829.330.5
Av44.350.143.940.1
Av*42.548.541.339.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Number of frames \\( T \\) and temporal stride. To investigate whether a wider temporal receptive field can reinforce temporal coherence and enhance object segmentations, we experiment by varying the number of input frames and temporal stride in training. Our results, displayed in Table 5, reveal that the performance significantly deteriorates when the stride is very small, such as a temporal stride of 1 in the first row. However, it begins to plateau upon reaching 4. This is because a small temporal stride leads to very subtle temporal dynamics in consecutive frames, resulting in a trivial learning task while a larger temporal stride provides richer dynamics, effectively encouraging the model to capture temporal consistency and maintain object permanence. Moreover, adding more frames results in diminishing returns, indicating that a modest temporal receptive field is adequate for sufficient temporal context in video object segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Motion representations. In training, we use the spatio-temporal attention maps as latent motion representations for dynamic motion alignment \\(\\mathcal{L}_{\\mathrm{m}}\\). There are some alternatives to formulate this motion representation. Here we explore two variants in Table 6. The one is simple global feature correlation, which directly calculates the dot product between spatio-temporal feature vectors, i.e., \\(C_g = \\tilde{F}\\tilde{F}^\\top \\in \\mathbb{R}^{THW\\times THW}\\). The other is local feature correlation, where we sample a sliding local window with size \\(T_{p}\\times H_{p}\\times W_{p}\\) for each grid, and obtain the local correlation matrix \\(C_l\\in \\mathbb{R}^{THW\\times T_pH_pW_p}\\). It is obvious that using the self-attention maps as motion representation substantially outperforms the performance of the other two methods. Compared to the global correlation, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.129 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.479, + 0.215 + ], + "angle": 0, + "content": "Table 5: Ablation on the number of frames \\(T\\) and temporal stride. When sampling more than 3 frames with a larger stride, the method exhibits a diminishing effect." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.226, + 0.481, + 0.334 + ], + "angle": 0, + "content": "
TStrideYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
3137.443.637.133.6
3444.350.143.940.1
3844.250.243.340.7
5444.850.644.341.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.491, + 0.145, + 0.782, + 0.228 + ], + "angle": 0, + "content": "Table 6: Ablation on different motion representations in calculating motion consistency loss \\( {\\mathcal{L}}_{\\mathrm{m}} \\) . We compare using the self-attention matrix,simple local and global feature correlation as the latent motion representation." + }, + { + "type": "table", + "bbox": [ + 0.492, + 0.24, + 0.795, + 0.334 + ], + "angle": 0, + "content": "
MotionRepYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
Local Corr.30.533.829.627.7
Global Corr.39.149.236.433.9
Self-Attention44.350.143.940.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.44 + ], + "angle": 0, + "content": "learnable parameters in attention layers enhance the modeling capacity, which enables the self-attention maps to capture more comprehensive spatio-temporal dependencies and serve as a better motion representation. The local correlation only captures temporal dynamics in limited reception fields, restricting the ability for long-term perception." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.461, + 0.36, + 0.478 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.658 + ], + "angle": 0, + "content": "In this paper, we introduce a simple yet powerful approach for self-supervised video object segmentation. Specifically, our proposed architecture introduces only a single spatio-temporal Transformer block that ingests frame-wise DINO features and constructs the spatio-temporal correspondence through the attention layer. During training, we adopt a dual self-supervised consistency objective that encompasses both semantic consistency and dynamic motion consistency, supplemented with entropy normalization. In inference, we utilize hierarchical clustering on spatio-temporal attention maps to generate temporally coherent object segmentation masks. This extremely straightforward approach nevertheless delivers state-of-the-art results on challenging real-world multi-object segmentation tasks such as DAVIS-17-Unsupervised and YouTube-VIS-19." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Limitation. While promising, BA still has room for improvement. Firstly, the current spatio-temporal attention block cannot accommodate very long window-size frames due to the quadratic memory cost. Potentially, adopting memory mechanisms [49, 75] that are capable of handling long-duration data may prove more suitable. Secondly, neither pixel-level constraints nor spatio-temporal dependencies consider the granularity in the video. Incorporating multi-scale features [4] may present a promising direction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Social Negative Impact. Although our method produces coherent video object segmentation without annotation, it could also enhance deepfake video technologies, making it easier to produce false videos for malicious intent. It is important to regulate the use of such technology and ensure it is leveraged for beneficial purposes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.581, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Betrayed by Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.273 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 61931023, Grant 61932022, Grant 62371288, Grant 62320106003, Grant 62301299, Grant T2122024, Grant 62120106007, the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.298, + 0.323, + 0.314 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.375 + ], + "angle": 0, + "content": "1. Aydemir, G., Xie, W., Guney, F.: Self-supervised object-centric learning for videos. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=919tWtJPXe" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.417 + ], + "angle": 0, + "content": "2. Bao, L., Wu, B., Liu, W.: Cnn in mrf: Video object segmentation via inference in a cnn-based higher-order spatio-temporal mrf. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5977-5986 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.446 + ], + "angle": 0, + "content": "3. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.489 + ], + "angle": 0, + "content": "4. Bian, Z., Jabri, A., Efros, A.A., Owens, A.: Learning pixel trajectories with multiscale contrastive random walks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6508-6519 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.53 + ], + "angle": 0, + "content": "5. Bielski, A., Favaro, P.: Move: Unsupervised movable object segmentation and detection. Advances in Neural Information Processing Systems 35, 33371-33386 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.573 + ], + "angle": 0, + "content": "6. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.615 + ], + "angle": 0, + "content": "7. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.658, + 0.785, + 0.699 + ], + "angle": 0, + "content": "9. Chen, Y., Liu, Y., Jiang, D., Zhang, X., Dai, W., Xiong, H., Tian, Q.: Sdae: Self-distillated masked autoencoder. In: European conference on computer vision. pp. 108-124. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.742 + ], + "angle": 0, + "content": "0. Choudhury, S., Karazija, L., Laina, I., Vedaldi, A., Rupprecht, C.: Guess What Moves: Unsupervised Video and Image Segmentation by Anticipating Motion. In: British Machine Vision Conference (BMVC) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.785, + 0.798 + ], + "angle": 0, + "content": "11. Ding, S., Li, M., Yang, T., Qian, R., Xu, H., Chen, Q., Wang, J., Xiong, H.: Motion-aware contrastive video representation learning via foreground-background merging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9716-9726 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Ding, S., Qian, R., Xiong, H.: Dual contrastive learning for spatio-temporal representation. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5649-5658 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.333, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Ding, S., Xie, W., Chen, Y., Qian, R., Zhang, X., Xiong, H., Tian, Q.: Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.26 + ], + "angle": 0, + "content": "14. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.317 + ], + "angle": 0, + "content": "15. Dutt Jain, S., Xiong, B., Grauman, K.: Fusionseg: Learning to combine motion and appearance for fully automatic segmentation of generic objects in videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3664-3673 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.359 + ], + "angle": 0, + "content": "16. Elsayed, G., Mahendran, A., van Steenkiste, S., Greff, K., Mozer, M.C., Kipf, T.: Savi++: Towards end-to-end object-centric learning from real-world videos. Advances in Neural Information Processing Systems 35, 28940-28954 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.402 + ], + "angle": 0, + "content": "17. Fan, D.P., Wang, W., Cheng, M.M., Shen, J.: Shifting more attention to video salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8554-8564 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.787, + 0.459 + ], + "angle": 0, + "content": "18. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.543 + ], + "angle": 0, + "content": "19. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., Kipf, T., Kundu, A., Lagun, D., Laradji, I., Liu, H.T.D., Meyer, H., Miao, Y., Nowrouzezahrai, D., Oztireli, C., Pot, E., Radwan, N., Rebain, D., Sabour, S., Sajjadi, M.S.M., Sela, M., Sitzmann, V., Stone, A., Sun, D., Vora, S., Wang, Z., Wu, T., Yi, K.M., Zhong, F., Tagliasacchi, A.: Kubric: a scalable dataset generator (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.787, + 0.584 + ], + "angle": 0, + "content": "20. Guo, Y., Yang, C., Rao, A., Agrawala, M., Lin, D., Dai, B.: Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.585, + 0.787, + 0.642 + ], + "angle": 0, + "content": "21. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=Fx2SbBgcte" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.642, + 0.787, + 0.696 + ], + "angle": 0, + "content": "22. Hamilton, M., Zhang, Z., Hariharan, B., Snavely, N., Freeman, W.T.: Unsupervised semantic segmentation by distilling feature correspondences. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=SaK06z6H10c" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.698, + 0.787, + 0.74 + ], + "angle": 0, + "content": "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.741, + 0.787, + 0.77 + ], + "angle": 0, + "content": "24. Hu, Y., Wang, R., Zhang, K., Gao, Y.: Semantic-aware fine-grained correspondence. In: European Conference on Computer Vision. pp. 97-115. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.812 + ], + "angle": 0, + "content": "25. Hu, Y.T., Huang, J.B., Schwing, A.G.: Videomatch: Matching based video object segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 54-70 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "26. Jabri, A., Owens, A., Efros, A.: Space-time correspondence as a contrastive random walk. Advances in neural information processing systems 33, 19545-19560 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.581, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Betrayed by Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "27. Johnander, J., Danelljan, M., Brissman, E., Khan, F.S., Felsberg, M.: A generative appearance model for end-to-end video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8953-8962 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.232 + ], + "angle": 0, + "content": "28. Johnson, S.C.: Hierarchical clustering schemes. Psychometrika 32(3), 241-254 (1967)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.233, + 0.787, + 0.286 + ], + "angle": 0, + "content": "29. Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jonschkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=aD7uesX1GF_" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "30. Lai, Z., Xie, W.: Self-supervised learning for video correspondence flow. In: BMVC (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "31. Lai, Z., Lu, E., Xie, W.: Mast: A memory-augmented self-supervised tracker. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6479-6488 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.411 + ], + "angle": 0, + "content": "32. Li, L., Wang, W., Zhou, T., Li, J., Yang, Y.: Unified mask embedding and correspondence learning for self-supervised video segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18706-18716 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.452 + ], + "angle": 0, + "content": "33. Li, X., Loy, C.C.: Video object segmentation with joint re-identification and attention-aware mask propagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 90-105 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.535 + ], + "angle": 0, + "content": "34. Li, X., Liu, S., De Mello, S., Wang, X., Kautz, J., Yang, M.H.: Joint-task self-supervised learning for temporal correspondence. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32. Curran Associates, Inc. (2019), https://proceedings.neurips.cc/paper/2019/file/140f6969d5213fd0ece03148e62e461e-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "35. Li, X., Liu, S., De Mello, S., Wang, X., Kautz, J., Yang, M.H.: Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "36. Lian, L., Wu, Z., Yu, S.X.: Bootstrapping objectness from videos by relaxed common fate and visual grouping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14582-14591 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.62, + 0.787, + 0.687 + ], + "angle": 0, + "content": "37. Liu, R., Wu, Z., Yu, S., Lin, S.: The emergence of objectness: Learning zero-shot segmentation from videos. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 13137-13152. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper/2021/file/6d9cb7de5e8ac30bd5e8734bc96a35c1-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "38. Liu, Y., Wang, Y., Chen, Y., Dai, W., Li, C., Zou, J., Xiong, H.: Promoting semantic connectivity: Dual nearest neighbors contrastive learning for unsupervised domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3510-3519 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "39. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "40. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "41. Melas-Kyriazi, L., Rupprecht, C., Laina, I., Vedaldi, A.: Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and local-" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "ization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8364-8375 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.216 + ], + "angle": 0, + "content": "42. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.787, + 0.257 + ], + "angle": 0, + "content": "43. Patil, P.W., Dudhane, A., Kulkarni, A., Murala, S., Gonde, A.B., Gupta, S.: An unified recurrent video object segmentation framework for various surveillance environments. IEEE Transactions on Image Processing 30, 7889-7902 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "44. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.352 + ], + "angle": 0, + "content": "45. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.393 + ], + "angle": 0, + "content": "46. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.434 + ], + "angle": 0, + "content": "47. Qian, R., Ding, S., Liu, X., Lin, D.: Static and dynamic concepts for self-supervised video representation learning. In: European Conference on Computer Vision. pp. 145-164. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "48. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "49. Qian, R., Dong, X., Zhang, P., Zang, Y., Ding, S., Lin, D., Wang, J.: Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.569 + ], + "angle": 0, + "content": "50. Qian, R., Li, Y., Liu, H., See, J., Ding, S., Liu, X., Li, D., Lin, W.: Enhancing self-supervised video representation learning via multi-level feature optimization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7990-8001 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "51. Rambhatla, S.S., Misra, I., Chellappa, R., Shrivastava, A.: Most: Multiple object localization with self-supervised transformers for object discovery. arXiv preprint arXiv:2304.05387 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.651 + ], + "angle": 0, + "content": "52. Salehi, M., Gavves, E., Snoek, C.G., Asano, Y.M.: Time does tell: Self-supervised time-tuning of dense image representations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16536-16547 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.677 + ], + "angle": 0, + "content": "53. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence 22(8), 888-905 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.717 + ], + "angle": 0, + "content": "54. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. In: BMVC 2021-32nd British Machine Vision Conference (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "55. Singh, G., Wu, Y.F., Ahn, S.: Simple unsupervised object-centric learning for complex and naturalistic videos. Advances in Neural Information Processing Systems 35, 18181-18196 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "56. Tian, J., Aggarwal, L., Colaco, A., Kira, Z., Gonzalez-Franco, M.: Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv preprint arXiv:2308.12469 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "57. Vondrick, C., Shrivastava, A., Fathi, A., Guadarrama, S., Murphy, K.: Tracking emerges by colorizing videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 391-408 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.581, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Betrayed by Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "58. Wang, N., Song, Y., Ma, C., Zhou, W., Liu, W., Li, H.: Unsupervised deep tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1308-1317 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "59. Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle-consistency of time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2566-2576 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.234, + 0.787, + 0.275 + ], + "angle": 0, + "content": "60. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3124-3134 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.787, + 0.33 + ], + "angle": 0, + "content": "61. Wang, X., Misra, I., Zeng, Z., Girdhar, R., Darrell, T.: Videocutler: Surprisingly simple unsupervised video instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22755-22764 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.332, + 0.787, + 0.387 + ], + "angle": 0, + "content": "62. Wang, Y., Shen, X., Yuan, Y., Du, Y., Li, M., Hu, S.X., Crowley, J.L., Vaufreydaz, D.: Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.388, + 0.787, + 0.443 + ], + "angle": 0, + "content": "63. Wang, Y., Li, J., ZHANG, X., Shi, B., Li, C., Dai, W., Xiong, H., Tian, Q.: BarleRiA: An efficient tuning framework for referring image segmentation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=wHLDHRkmEu" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.445, + 0.787, + 0.486 + ], + "angle": 0, + "content": "64. Xie, C., Xiang, Y., Harchaoui, Z., Fox, D.: Object discovery in videos as foreground motion clustering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9994-10003 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.787, + 0.528 + ], + "angle": 0, + "content": "65. Xie, J., Xie, W., Zisserman, A.: Segmenting moving objects via an object-centric layered representation. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.53, + 0.787, + 0.556 + ], + "angle": 0, + "content": "66. Xu, H., Ding, S., Zhang, X., Xiong, H., Tian, Q.: Masked autoencoders are robust data augmentors. arXiv preprint arXiv:2206.04846 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.558, + 0.787, + 0.599 + ], + "angle": 0, + "content": "67. Xu, J., Wang, X.: Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10075-10085 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "68. Yang, C., Lamdouar, H., Lu, E., Zisserman, A., Xie, W.: Self-supervised video object segmentation by motion grouping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7177-7188 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.643, + 0.733, + 0.657 + ], + "angle": 0, + "content": "69. Yang, L., Fan, Y., Xu, N.: Video instance segmentation. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "70. Yang, Y., Lai, B., Soatto, S.: Dystab: Unsupervised object segmentation via dynamic-static bootstrapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2826-2836 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "71. Yang, Y., Loquercio, A., Scaramuzza, D., Soatto, S.: Unsupervised moving object detection via contextual information separation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.742, + 0.787, + 0.784 + ], + "angle": 0, + "content": "72. Ye, V., Li, Z., Tucker, R., Kanazawa, A., Snavely, N.: Deformable sprites for unsupervised video decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2657-2666 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "73. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=1_jFneF07YC" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.129 + ], + "angle": 0, + "content": "S. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "74. Zadaianchuk, A., Seitzer, M., Martius, G.: Object-centric learning for real-world videos by predicting temporal feature similarities. In: Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "75. Zhou, J., Pang, Z., Wang, Y.X.: Rmem: Restricted memory banks improve video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18602-18611 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "76. Ziegler, A., Asano, Y.M.: Self-supervised learning of object parts for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14502-14511 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_origin.pdf b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82ef4bbdd0753da97aee365b4c362a50f421bf7d --- /dev/null +++ b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/c2e87780-6d0f-4d74-9df7-fb50decc3b02_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:add7602ed281cccb8aaf3c1871b82fd24cd393d2ccf29d078a8c8cfff4da3f1e +size 3761636 diff --git a/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/full.md b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..365a21587540b689f6d5dd192bc7b6a354513e1c --- /dev/null +++ b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/full.md @@ -0,0 +1,314 @@ +# Betrayed by Attention: A Simple yet Effective Approach for Self-supervised Video Object Segmentation + +Shuangrui Ding $^{1*}$ , Rui Qian $^{1*}$ , Haohang Xu $^{2}$ , Dahua Lin $^{1,3,4\dagger}$ , and Hongkai Xiong $^{2}$ + +1 The Chinese University of Hong Kong, Hong Kong, China + +2 Shanghai Jiao Tong University, Shanghai, China + +3 Shanghai Artificial Intelligence Laboratory, Shanghai, China + +4 Centre for Perceptual and Interactive Intelligence (CPII), Hong Kong, China {ds023, qr021}@ie.cuhk.edu.hk + +Abstract. In this paper, we propose a simple yet effective approach for self-supervised video object segmentation (VOS). Previous self-supervised VOS techniques majorly resort to auxiliary modalities or utilize iterative slot attention to assist in object discovery, which restricts their general applicability. To deal with these challenges, we develop a simplified architecture that capitalizes on the emerging objectness from DINO-pretrained Transformers, bypassing the need for additional modalities or slot attention. Our key insight is that the inherent structural dependencies present in DINO-pretrained Transformers can be leveraged to establish robust spatio-temporal correspondences in videos. Furthermore, simple clustering on this correspondence cue is sufficient to yield competitive segmentation results. Specifically, we first introduce a single spatio-temporal Transformer block to process the frame-wise DINO features and establish spatio-temporal dependencies in the form of self-attention. Subsequently, utilizing these attention maps, we implement hierarchical clustering to generate object segmentation masks. To train the spatio-temporal block in a fully self-supervised manner, we employ semantic and dynamic motion consistency coupled with entropy normalization. Our method demonstrates state-of-the-art performance across three multi-object video segmentation tasks. Specifically, we achieve over 5 points of improvement in terms of FG-ARI on complex real-world DAVIS-17-Unsupervised and YouTube-VIS-19 compared to the previous best result. The code and checkpoint are released at https://github.com/shvdiwnkozbw/SSL-UVOS. + +Keywords: Video Object Segmentation $\cdot$ Self-supervised Learning + +![](images/617556ba29e3068a86f00013c0bda63f615c151c7e85f5ca448fc00d6db67c48.jpg) +(a) Video Frame + +![](images/94dc76dc03c4752b3232035c31e681a3ca1c088b99fa37e08eeb34db2d5fb825.jpg) +(b) DINO Attention + +![](images/118cf47bd5adbaa47eee8f770320b14d420d09017c790a417f11d48bc22635a8.jpg) +Fig. 1: Attention leaks the object's position! We visualize the self-attention maps of different queries (■ prompt) from the video sequence (a). The frame-wise DINO attention maps (b) highlight image regions corresponding to the queried object. A randomly initialized spatio-temporal Transformer block on top of DINO produces noisy spatio-temporal attention maps (c) that coarsely track objects over time. Our method diminishes noise in the learned spatio-temporal attention maps (d) which derive temporally coherent object segmentation. + +![](images/7881613c9aab4294a133be65b35c520741e98fd9e1e522999a8c976d6c41debd.jpg) +(c) Random Initialized Spatio-temporal Attention + +Ratio-temporal Attention + +![](images/484b88b39daec6bdba722480e537bb1d81e7180df283dbb26e11cc62c3a253f0.jpg) +(d) Learned Spatio-temporal Attention + +![](images/4560323e676730cf9c4d3f8dc1c02873cb478f00c885c5f3948551d46a00285a.jpg) + +# 1 Introduction + +Representing the visual scene with objects as the basic elements has been long acknowledged as a core cognitive capability of an intelligent agent. In the realm of computer vision, Video Object Segmentation (VOS) tasks require a model to segment and track specified objects within a video sequence, striving to emulate this foundational function. This task holds significant importance in various real-world vision systems, including but not limited to, autonomous driving [18] and surveillance security [43]. However, traditional methods for VOS [6,44] typically follow a fully supervised paradigm and entail substantial costs for obtaining pixel-level per-frame annotations. This factor significantly limits their feasibility for large-scale applications. Thus, the researchers turn their attention to self-supervised learning [8,9,11,12,23,38,47,50,66], a more efficient approach that leverages unlabeled data. + +Despite its promising potential, self-supervised VOS poses significant challenges. Firstly, previous methods tend to incorporate additional signals such as optical flow [36, 68, 71], or depth cues [16, 29], to furnish object-related clues. However, these auxiliary signals within videos can often be difficult to extract and unreliable when it comes to multi-object scenarios, thus limiting their availability. Secondly, most existing models [1,48,74] adopt slot attention [39] to parse the frame into objects. Nevertheless, the requirement for a predefined quantity of learnable slot queries limits its adaptability to in-the-wild data, especially when the number of objects is unknown. To address these limitations, we recall the recent self-supervised learning technique DINO [7, 42], which demonstrates emerging objectness in the attention maps of pretrained Vision Transformer (ViT) [14]. As depicted in Fig. 1(b), DINO attention maps encode spatial dependencies between diverse patches. These dependencies present different patterns for various objects, providing abundant cues for object segmentation. + +This phenomenon motivates us with an intriguing possibility: Can we utilize the object-aware attributes of DINO to learn robust spatio-temporal dependencies and thereby produce coherent object segmentations in videos? + +In this paper, we delve into the potential of this idea with a simple architecture. Based on DINO pretrained ViT, we introduce a single spatio-temporal Transformer block to process frame-wise DINO features and calculate spatiotemporal attention maps. Remarkably, these attention maps in Fig. 1(c) showcase the ability to coarsely discriminate and track object parts even with random initialization. Based on this observation, we can roughly estimate the patches belonging to the same object over time. Then, we encourage both semantic and motion alignment among patches belonging to the same object while distinguishing those from different objects in a fully self-supervised manner. The training significantly diminishes noise in the learned spatio-temporal dependencies as displayed in Fig. 1(d). In inference, we directly apply Hierarchical Clustering [28] to these spatio-temporal attention maps to derive object segmentation masks within a sequence. Interestingly, this naive clustering method yields unexpectedly competitive results. Our method greatly outperforms slot attention-based methods on various datasets, e.g., 9 points improvement over SOLV [1] on DAVIS-17, 12 points improvement over SMTC [48] on YTVIS-19. We argue that clustering on the whole spatio-temporal attention maps, instead of frame-by-frame feature clustering as in slot attention ensures the temporal coherence of produced segmentations and guarantees the object's permanence. + +In summary, our proposed method BA (short for Betrayed by Attention) offers three distinct advantages: (1) BA solely requires RGB frames and operates entirely under self-supervision without the need for external modality. This characteristic renders our method highly accessible and adaptable in real-world scenarios; (2) The architecture is remarkably simple and systematically efficient. We only introduce a single learnable Transformer block and harness a parameter-free clustering algorithm to discover the objects. It facilitates easy generalization to multi-object scenes, without the constraints imposed by slot attention; (3) Our method achieves state-of-the-art results across a bunch of unsupervised video multi-object segmentation benchmarks, such as MOVi-E [19], DAVIS-17-Unsupervised [46], and YouTube-VIS-19 [69]. Notably, we surpass the previous state-of-the-art method, TimeT [52], by 6.6 points on YTVIS-19 and by 6.1 points on DAVIS-17 in terms of the FG-ARI metric. + +# 2 Related Work + +Video Object Segmentation aims to segment objects coherently in a video sequence [2,6,15,17,20,21,25,27,33,44]. In video object segmentation (VOS), there are two prevalent protocols for evaluating the learned models: Semi-supervised VOS and Unsupervised VOS. In semi-supervised VOS, the algorithm is provided with the object masks in the first frame, and required to track them in subsequent frames. In contrast, unsupervised VOS aims to identify and segment salient objects from the background without any specific reference. In this paper, we fo + +cus on the more challenging unsupervised VOS setting without using any kind of manual annotations in either training or inference. Recently, there emerges a line of self-supervised algorithms for unsupervised VOS [10, 29, 34, 58, 64, 68, 70-72]. CIS [71] facilitates fully unsupervised motion segmentation, discarding object mask supervision during training. By formulating a min-max game of mutual information, the generator is motivated to create segmentations that effectively distinguish foreground objects from the background. AMD [37] minimizes the warping synthesis error to train appearance and motion pathways without any supervision. GWM [10] uses a single RGB image as input with optical flow acting as supervision to highlight moving areas. MG [68] and OCLR [65] solely leverage optical flow as input to generate object-centric layered representations, with each layer indicating a potential object. Despite the impressive performance of state-of-the-art self-supervised VOS methods [1, 48, 52] in discovering single objects, they still face limitations when it comes to solving multi-object discovery tasks. In contrast, by leveraging the rich object cues in attention maps, our method can discern multiple objects in real-world scenarios. + +Self-supervised Spatio-temporal Correspondence learning usually leverages free temporal supervision signals in videos to learn representations that facilitate accurate pixel or object tracking across space and time [4, 7, 13, 24, 26, 30, 31, 35, 57, 59, 67]. Vondrick et al. [57] employ the natural temporal color coherence to train a colorization model on grayscale videos, thereby establishing fine-grained correspondence between current and future frames. CRW [26] presents a self-supervised learning approach for dense correspondence in raw videos. It uses space-time graph-based random walks and cycle consistency to implicitly supervise chains of comparisons. Hu et al. [24] independently learn semantics and temporal correspondence from two pathways and fuse them at a later stage. UME [32] respectively designs short-term appearance and long-term semantic consistency to learn generalizable correspondence. Taking a step further by integrating high-level semantics with low-level temporal correlation, SMTC [48] develops two-stage slot attention to establish dense correspondence with more emphasis on foreground objects. In this work, we explicitly initialize the spatio-temporal correspondence on top of the DINO pretrained Transformer. This initialization effectively guides the learning process with a self-supervised pixel-level consistency. + +Object-aware DINO Features have shown efficacy in object localization tasks, particularly in image domain [5, 22, 41, 54, 60, 63, 73, 76]. DINO's original paper [7] reveals that the learned representations and attention maps carry substantial object cues that can be harnessed for object discovery. The groundbreaking work, LOST [54], utilizes these features from DINO to perform object segmentation by constructing a graph wherein objects are segmented using the inverse degrees of nodes. TokenCut [62] employs DINO features for applying the Normalized Cut algorithm [53], thereby obtaining foreground segments in an image. Furthermore, CutLER [60] and MOST [51] can localize multiple objects without any supervision based on the DINO features. Motivated by the successful application of DINO features in image settings, recent works have expanded + +their use to video segmentation tasks [1,48,52,61,62]. SMTC [48] utilizes the rich semantics and correspondence cues in the DINO features to provide a reliable reference for object decomposition in consecutive frames. TimeT [52] introduces a feature forwarding process to propagate dense DINO features across time and establish temporal consistency. Most of the existing works focus on leveraging DINO feature vectors for optimization and adaptation to videos. While in our work, we lay more emphasis on the self-attention maps. We extend the framework-wise DINO attention maps to spatio-temporal maps, which explicitly encode space-time structural dependencies and help coherent object discovery in videos. + +# 3 Method + +# 3.1 Preliminary on DINO Attention Maps + +The self-supervised ViT, as achieved by DINO [7], has exhibited a rich emergence of objectness in the attention maps. To elucidate this, we visualize the self-attention map from the final Transformer block of the DINO pretrained ViT-S/8 as a representative. As illustrated in Fig. 1(b), an image patch exhibits a high degree of attention dependencies with patches of the same object. Conversely, image patches of distinct objects tend to present different attention distributions. Encouraged by this observation, we pose the following question: Could it be possible to directly harness these attention maps to achieve object segmentation without relying on annotations? + +
Clustering MetricCosine similarityKL divergence
FAAvAv
DAVIS-1647.460.351.175.4
DAVIS-1714.726.720.939.2
+ +Table 1: Preliminary video object segmentation results using varying clustering metrics. $F$ represents features pretrained with DINO. $A$ denotes per-frame attention maps from DINO, while $A_{v}$ and $\tilde{A}_v$ denote spatio-temporal attention maps produced by a randomly initialized and a final learned temporal correlator respectively, on top of DINO. The IoU ( $\mathcal{I}$ score) is reported for both DAVIS-16 and DAVIS-17-Unsupervised datasets. + +![](images/c7ed16abeed06e7fec0e373b28216579cd1e477319c01b41e4bbe0cb7b01eeac.jpg) +(a) DINO Feature $F$ +Fig.2: Visualizations of the clustering results. The left column is the results of DINO features $F$ , the right column is the results of DINO attention $A$ . $F$ results in much noisier clusters, while $A$ distinguishes different classes of objects. + +![](images/2472acf7534ce171972bb6f951e4e70ce1c6f3b69b910271f54c98ea5c10f94b.jpg) +(b) DINO Attention A + +In response to this question, we conduct a straightforward initial exploration with these attention maps. We denote the attention map of the last Transformer block as $A \in \mathbb{R}^{HW \times HW}$ and its output features $F \in \mathbb{R}^{HW \times C}$ , where $H$ , $W$ and $C$ respectively denote the height, width, and channel dimension of the feature + +![](images/b0315186df5a53734e34c6da14f8a914d056d5cecd0dc4ac38ee4527f324623d.jpg) +Fig. 3: Our architecture BA overview. Given the video frames and a DINO pretrained Transformer, we first use a temporal correlator to construct the spatio-temporal correspondence. We then utilize these attention maps as a clustering metric and apply hierarchical clustering across all frames to generate segmentation masks. During training, for each patch, we sample a positive/negative set and assign an importance weight based on its corresponding attention map. We promote alignment within the positive set while differentiating the representations from the negative set. The final loss, normalized with importance weights, only trains the temporal correlator with the DINO ViT remaining frozen. + +map5. For an intuitive comparison, we apply unsupervised clustering on both attention maps $A$ and features $F$ to yield the object segmentation masks. For the detailed procedure, please refer to the Sec. 3.2. Then, we report frame-wise evaluation results on the video object segmentation datasets DAVIS-16 [45] and DAVIS-17-Unsupervised [46] in Table 1. Interestingly, we observe that the attention maps significantly excel in decoupling object components compared to features, exhibiting an improvement of over 10 points on both datasets in terms of IoU. Surprisingly, a spatio-temporal attention map $A_v$ generated by a randomly initialized Transformer block on top of DINO also outperforms the frame-wise DINO feature $F$ . And it is further demonstrated in Fig. 2. The DINO features lead to very noisy clusters, while attention maps can discriminate objects of different categories. We speculate this is because the attention maps preserve intact correspondences, which are more suitable for fine-grained tasks such as object grouping and discrimination. Therefore, creating spatio-temporal structural dependencies that expand beyond spatial DINO attention maps appears promising for video object segmentation tasks. + +# 3.2 Our Simple Architecture for VOS + +Motivated by the frame-wise preliminary result, we introduce a temporal correlator instantiated with a spatio-temporal Transformer block [3] on top of the DINO pretrained ViT backbone, to extend the self-attention calculation to the spatiotemporal domain. Specifically, given the frame-wise DINO features $\{F_1,\dots ,F_T\}$ of $T$ frames, we process them through a spatio-temporal Transformer block $\varPhi_{\mathrm{st}}$ . It consists of one standard ViT encoder layer, with self-attention computed over + +the spatio-temporal grids. Specifically, we concatenate the frame-wise DINO features along the temporal dimension with the addition of learnable temporal positional embeddings $\mathbf{pe}$ , i.e., $F = \{F_{1} + \mathbf{pe}_{1},\dots ,F_{T} + \mathbf{pe}_{T}\} \in \mathbb{R}^{THW\times C}$ and then linearly project $F$ to obtain the queries $Q = f_{Q}(F)$ , keys $K = f_{K}(F)$ , and values $V = f_{V}(F)$ . Then, we can obtain the self-attention map $A_{v}\in \mathbb{R}^{THW\times THW}$ and temporally fused features $F_{v}\in \mathbb{R}^{THW\times C}$ : + +$$ +A _ {v} = \mathrm {s o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {C}}\right), \quad F _ {v} = A _ {v} V = \{\tilde {F} _ {1}, \dots , \tilde {F} _ {T} \}. \qquad (1) +$$ + +In this way, we establish spatio-temporal structural dependencies beyond frameworkwise DINO features and have the potential to improve object permanence across the entire time span. + +Since $A_v[i] \in \mathbb{R}^{THW}$ is a proximity distribution that describes the structural dependencies between the $i$ -th patch and all other positions, we adopt symmetric KL divergence to measure the distance $\mathcal{M} \in \mathbb{R}^{THW \times THW}$ between any two patches as: + +$$ +\mathcal {M} [ i, j ] = \mathcal {D} _ {K L} \left(A _ {v} [ i ] | | A [ j ]\right) + \mathcal {D} _ {K L} \left(A _ {v} [ j ] | | A _ {v} [ i ]\right), \tag {2} +$$ + +where $\mathcal{D}_{KL}$ denotes KL divergence. Based on this distance matrix $\mathcal{M}$ , we then apply Hierarchical Clustering [28] to obtain the object segmentation masks. Please consult our appendix for a detailed implementation of the hierarchical clustering algorithm. In this way, we iteratively merge $THW$ patches into $N$ centroids $A_{c} \in \mathbb{R}^{N \times THW}$ . Since each centroid likely represents the spatial distribution of an object, we consider the cluster assignments as potential candidates for the object segmentation mask. Specifically, we calculate the KL divergence distance between the $A_{v}$ and $A_{c}$ as $D \in \mathbb{R}^{THW \times N}$ , and use argmin operation to produce the cluster assignments $Z$ within the video sequence: + +$$ +Z = \operatorname {a r g m i n} (D, \dim = 1) \in \{1, \dots , N \} ^ {T H W}. \tag {3} +$$ + +Then, we employ a bipartite matching mechanism, following the standard evaluation protocol in DAVIS-17 Challenge [46], to match the predicted masks with ground truth for evaluation. + +A recent study, DiffSeg [56], used hierarchical clustering to process attention maps from a stable diffusion model, but it operated solely in the image domain. Our work extends the powerful image foundation model to the spatio-temporal domain, demonstrating that accurate spatio-temporal correspondence can be learned without supervision. Additionally, our method does not rely on hierarchical clustering. As shown in Table 4, using K-means clustering also achieves competitive performance, indicating that the learned spatio-temporal correspondence is effective across various clustering strategies. + +Besides, there are two noteworthy aspects in our clustering stage. First, in the clustering process, we only set a distance threshold hyper-parameter without specifying the number of cluster centroids and the hierarchical clustering algorithm can adaptively generate different numbers of clusters according to + +the complexity of the video. Second, our reference to the entire spatio-temporal distribution enables coherent object segmentation across multiple frames. As a result, the generated segmentation masks effortlessly track corresponding objects over time without extra association procedures. + +# 3.3 Training + +Noticing that even a randomly initialized attention map $A_v$ shown in Figure 1(c), exhibits an initial ability to differentiate parts and track objects over time, we leverage these maps as indicators for positive and negative sets sampling. Mathematically, given the attention maps $A_v \in \mathbb{R}^{THW \times THW}$ , we first reshape $A_v$ into shape $THW \times T \times HW$ then use top-k operation to retrieve $K_p$ position indexes in each frame with highest attention scores: + +$$ +I _ {p} = \operatorname {t o p - k} (\text {r e s h a p e} (A _ {v}), K _ {p}, \dim = - 1). \tag {4} +$$ + +In this way, we obtain a position index set $I_{p} \in \mathbb{R}^{THW \times T \times K_{p}}$ , which represents positions potentially associated with the same object as the $i$ -th patch. The reshape operation ensures that we consistently extract the same number of patches in each frame. This, in turn, guarantees effective object tracking within consecutive frames and promotes consistent temporal correspondence learning. Similarly, we perform this operation to find $K_{n}$ positions in each frame with the lowest attention scores $I_{n} \in \mathbb{R}^{THW \times T \times K_{n}}$ . Each $I_{n}[i]$ represents positions least likely to belong to the same object as the $i$ -th patch. After obtaining the positive and negative index sets, we apply self-supervised correspondence learning in two aspects: semantic consistency and dynamic motion consistency. + +Semantic consistency. Intuitively, we take the temporally fused features $F_{v} = \{\tilde{F}_{1},\dots,\tilde{F}_{T}\} \in \mathbb{R}^{THW\times C}$ as the semantic representation. We promote the semantic alignment between the features of the same object and distinguish features from different objects. Given the $i$ -th patch, we first refer to the positive pair indexes $I_{p}[i]$ in Eq. 4 to gather the features that are likely from the same object. Then we calculate the cosine distance between the query feature and all positive feature vectors, and produce the distance matrix $S_{p}[i]\in \mathbb{R}^{TK_{p}}$ . Similarly, we gather distinct features using $I_{n}[i]$ , and form negative distance matrix $S_{n}[i]$ . We employ a simple margin loss to enforce higher consistency between corresponding object areas: + +$$ +\mathcal {L} _ {\mathrm {s}} [ i ] = \sum_ {j = 1} ^ {T K _ {p}} \sum_ {k = 1} ^ {T K _ {n}} \max \left(\mathcal {S} _ {p} [ i, j ] - \mathcal {S} _ {n} [ i, k ] + \lambda_ {1}, 0\right), \tag {5} +$$ + +where $\lambda_{1}$ is the margin hyper-parameter and we set $\lambda_{1} = 0.8$ as default. + +Dynamic motion consistency. Regarding dynamic motions, rather than calculating optical flow, inter-frame feature correlations, or latent cost volumes as motion representations, we turn to highly accessible attention maps $A_v$ . These maps exhibit spatio-temporal correlations of specific patches, serving as an effective latent representation of temporal dynamics. Therefore, it is feasible to + +directly employ these attention maps to bolster motion consistency between corresponding objects. Similar to the semantic consistency learning, given the attention map of $i$ -th patch $A_v[i]$ as a query, we also leverage the positive (negative) pair indexes $I_p[i] \left(I_n[i]\right)$ to gather the attention maps of the corresponding (distinct) objects. Then we calculate the symmetric KL divergence as in Eq. 2 between the query and all positive (negative) attention maps, obtaining the distance matrix $\mathcal{M}_p[i] \in \mathbb{R}^{TK_p}$ and $\mathcal{M}_n[i] \in \mathbb{R}^{TK_n}$ . We utilize the same margin loss for optimization, with $\lambda_2 = 1.0$ as the margin hyper-parameter: + +$$ +\mathcal {L} _ {\mathrm {m}} [ i ] = \sum_ {j = 1} ^ {T K _ {p}} \sum_ {k = 1} ^ {T K _ {n}} \max \left(\mathcal {M} _ {p} [ i, j ] - \mathcal {M} _ {n} [ i, k ] + \lambda_ {2}, 0\right). \tag {6} +$$ + +Overall objectives. Besides, we evaluate the importance of each spatio-temporal patch and lay more emphasis on the informative regions. Formally, we employ the entropy of the attention maps as a measure of the information content within each patch: + +$$ +e [ i ] = \sum_ {k = 1} ^ {T H W} - A _ {v} [ i, k ] \log A _ {v} [ i, k ], \tag {7} +$$ + +with $e[i]$ denoting the entropy of $i$ -th patch. A higher entropy value indicates more ambiguous spatio-temporal correlations, which provide less information. Hence, we perform softmax normalization on the reversed entropy to produce the importance weight $w$ for each patch as: + +$$ +w [ i ] = \frac {\exp (- e [ i ])}{\sum_ {j = 1} ^ {T H W} \exp (- e [ j ])}. \tag {8} +$$ + +Then, the overall learning objectives on semantic and motion consistency can be formulated as the weighted summation over all spatio-temporal patches: + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {T H W} w [ i ] \left(\mathcal {L} _ {\mathrm {s}} [ i ] + \mathcal {L} _ {\mathrm {m}} [ i ]\right). \tag {9} +$$ + +Since we only introduce one additional temporal correlator and an entirely parameter-free clustering strategy, we exclusively tune a single standard spatiotemporal Transformer block with 1.6M parameters and freeze the DINO encoder throughout the training process. This straightforward method not only maintains a minimal number of learnable parameters but also yields remarkable results in video object segmentation. + +# 4 Experiments + +# 4.1 Datasets + +For training, we adopt the challenging video dataset YouTube-VIS-19 [69] for video instance segmentation. YouTube-VIS-19 consists of 2,883 high-resolution + +YouTube videos and each video usually contains multiple object instances of distinct semantics. During the inference, we evaluate our method on unsupervised multiple object segmentation. We use the synthetic MOVi-E [19] dataset and real-world datasets like DAVIS-17-Unsupervised [46] and YouTube-VIS-19 [69], reporting Foreground Adjusted Rand Index (FG-ARI) and mIoU. For MOVi-E, we integrate its training set into our training due to significant distribution shifts. For the rest, we maintain solely training on the YouTube-VIS-19 set. Note that we do not present any mask annotation to the model during either the training or inference stages. + +# 4.2 Implementation Details + +In training, we sample $T = 3$ frames with temporal stride 4 as the input clip. Each frame is augmented with random crop and color jitter and resized to $192 \times 384$ . We adopt DINO pretrained ViT-S/8 [7] as the frame encoder, which is then followed by a single spatio-temporal Transformer Encoder block with 8 heads as temporal correlator $\varPhi_{\mathrm{st}}$ . We set the number of sampled positive and negative pairs to $K_{p} = 10$ and $K_{n} = 50$ in default. For optimization, we adopt AdamW [40] with a learning rate $1 \times 10^{-4}$ , and train the model for a total of 30k iterations with a batch size of 16. In the inference stage, our model can be flexibly applied to video sequences of arbitrary lengths. For inference, we set the distance threshold in hierarchical clustering to 1.0. We follow the standard evaluation protocol outlined in the DAVIS-17 challenge [46] to match the predicted masks with the ground truth. + +Table 2: Quantitative results on multiple object video segmentation. For MOVi-E and YouTube-VIS-19 datasets, we report FG-ARI and mIOU. Besides, we report Region Similarity $\mathcal{J}$ and Contour Accuracy $\mathcal{F}$ on DAVIS-17-Unsupervised. $\star$ The original paper of TimeT [52] includes the IoU of the background category, we rerun the evaluation to exclude background here. $\dagger$ We use DINOv2 ViT-S/14 as our frozen frame encoder. + +
ModelMOVi-EYTVIS-19DAVIS-17
FG-ARImIoUFG-ARImIoUFG-ARIJ&FJF
SAVi [29]42.816.011.112.7----
STEVE [55]50.626.620.020.9----
OCLR [65]--15.932.514.7-34.6-
VideoSAUR [74]73.935.639.529.1----
SOLV [1]80.8-29.145.332.2-30.2-
SMTC [48]--31.438.833.340.536.444.6
TimeT* [52]--37.940.435.540.035.844.2
BA (ours)83.440.244.350.140.143.939.248.6
BA† (ours)84.440.744.550.141.643.739.448.0
+ +# 4.3 Comparison with State-of-the-art + +We evaluate our method on challenging multi-object segmentation, which aims to segment different objects in a temporally coherent manner. Among the compared methods, SAVi [29] and OCLR [65] employ optical flow as supervision to discriminate motions of different objects. STEVE [55], VideoSAUR [74], SOLV [1] and SMTC [48] develop variants of slot attention with temporal constraints to bind to distinct instances. TimeT [52] propagates dense DINO features across time to enhance temporal consistency and applies clustering on learned features to produce object segmentations. In contrast, our method directly uses spatiotemporal attention maps for object segmentation. As shown in Table 2, BA achieves state-of-the-art results on both synthetic (MOVi-E) and realistic video datasets (YTVIS-19 and DAVIS-17), indicating promising generalization ability. Specifically, comparing to the prior arts, we gain an approximately 5-point advantage over both VideoSAUR [74] on YTVIS-19 and SMTC [48] and TimeT [52] on DAVIS-17 in terms of FG-ARI. The superior results under our simple architecture reveal the feasibility of utilizing the spatio-temporal dependencies to mine object cues and derive reliable video object segmentation results. Besides, we also report the results using DINOv2 ViT-S/14 as the frame encoder. Despite the larger patch size, our method still achieves competitive performance. This demonstrates the robustness of our approach to pretraining models. + +![](images/90ec1e8bd5bfcaccaee153d5d1c1bd46daa30fa48bed3a6d416b24161f0c692b.jpg) +Fig. 4: Qualitative comparison with DINO [7] and SMTC [48] on DAVIS-17-Unsupervised. Different color denotes different clusters. + +Qualitative Comparison. To visually demonstrate the effectiveness of our proposed method, we contrast our segmentation results with those of the DINO counterpart and prior arts SMTC [48] in Fig. 4. For the DINO baseline, we cluster the patches through the attention map $A$ . As can be observed, our method is able to segment with significantly less noise. For instance, in the motocross-jump sequence, BA nearly outlines the entire motorbike and the rider, whereas DINO misses most of the bike's body and only identifies its tyre. This shows that inferring the masks by associating a collection of RGB features and constructing spatio-temporal correspondences greatly enhances the segmentation result. Additionally, with the benefit of temporal correspondence, our model can differentiate between distinct object instances with the same semantics. Consider the soapbox sequence displayed in the right column; our method distinguishes two different people - one running behind the soapbox and another sitting in the soapbox throughout the video. This level of distinction cannot be obtained from DINO's spatial attention map, further emphasizing the effectiveness of introducing rich spatio-temporal relations in video segmentation tasks. + +# 4.4 Ablation Study + +In the ablation study, we utilize DINO ViT-S/8 as the frozen backbone. We report FG-ARI and IoU on YouTube-VIS-2019 and $\mathcal{J}\&\mathcal{F}$ on DAVIS-17-Unsupervised for multi-object segmentation. + +Training objective $\mathcal{L}_{\mathrm{s}}$ , $\mathcal{L}_{\mathrm{m}}$ , and $w$ . To understand the contribution of each component to the VOS performance, we individually evaluate each loss item as depicted in Table 3. Compared to the baseline without training, the inclusion of semantic consistency loss $\mathcal{L}_{\mathrm{s}}$ with entropy-based weighting mechanism $w$ (second row) elevates FG-ARI from 23.4 to 30.1 on YouTube-VIS-19. Furthermore, the motion consistency loss $\mathcal{L}_{\mathrm{m}}$ (third row) yields even more substantial improvements, resulting in FG-ARI reaching 39.9. This aligns with our preliminary observations that dynamic motions play a more significant role than semantics in fine-grained object segmentation. Combining all components (fourth row) provides the best results, highlighting the synergy between the components. Additionally, removing the weighting function $w$ (fifth row) slightly decreases the performance on both two benchmarks, affirming the value of entropy normalization to prioritize informative regions. + +Clustering metrics and strategy. To further explore whether our clustering strategy optimally leverages the learned spatio-temporal properties, we select four types of clustering metrics for comparison. Specifically, we adopt both feature vectors and attention maps directly from DINO and those from our learned spatio-temporal Transformer block. Table 4 reveals three key observations. First, our learning paradigm consistently outperforms the off-the-shelf DINO in terms of both features and attention maps, improving all metrics by over 10 points. This strongly supports our motivation that establishing spatio-temporal correspondence is crucial for video object segmentation. Second, utilizing the intermediate attention maps directly for clustering proves to be more effective than employing the final features. We speculate this occurs because the attention maps encode + +fine-grained structural dependencies between various positions, whereas feature vectors primarily capture high-level semantics. Third, the hierarchical clustering strategy is not mandatory for success. We also achieve similar performance by replacing it with K-means clustering, where we set the number of centroids to 8, indicating that BA distills general object-aware representations across time. + +Table 3: Ablation on different combinations of semantic consistency loss $\mathcal{L}_{\mathrm{s}}$ , motion consistency loss $\mathcal{L}_{\mathrm{m}}$ and entropy weight $w$ . The first row represents the results obtained from a randomly initialized spatio-temporal Transformer block. + +
LsmYTVIS-19DAVIS-17
wFG-ARImIoUJ&FFG-ARI
--23.425.523.118.4
✓X30.132.429.627.5
X✓39.946.138.637.3
✓✓44.350.143.940.1
✓✓43.148.142.538.1
+ +Table 4: Ablation on different clustering metrics. Features $F$ and attention maps $A$ are directly obtained from DINO. $F_{v}$ and $A_{v}$ are generated from our learned spatiotemporal Transformer block. $A_{v}$ and $A_{v}^{*}$ respectively denote the hierarchical clustering and K-means clustering results. + +
Clustering MetricYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
F22.123.816.119.4
Fv33.931.433.432.1
A30.134.829.330.5
Av44.350.143.940.1
Av*42.548.541.339.3
+ +Number of frames $T$ and temporal stride. To investigate whether a wider temporal receptive field can reinforce temporal coherence and enhance object segmentations, we experiment by varying the number of input frames and temporal stride in training. Our results, displayed in Table 5, reveal that the performance significantly deteriorates when the stride is very small, such as a temporal stride of 1 in the first row. However, it begins to plateau upon reaching 4. This is because a small temporal stride leads to very subtle temporal dynamics in consecutive frames, resulting in a trivial learning task while a larger temporal stride provides richer dynamics, effectively encouraging the model to capture temporal consistency and maintain object permanence. Moreover, adding more frames results in diminishing returns, indicating that a modest temporal receptive field is adequate for sufficient temporal context in video object segmentation. + +Motion representations. In training, we use the spatio-temporal attention maps as latent motion representations for dynamic motion alignment $\mathcal{L}_{\mathrm{m}}$ . There are some alternatives to formulate this motion representation. Here we explore two variants in Table 6. The one is simple global feature correlation, which directly calculates the dot product between spatio-temporal feature vectors, i.e., $C_g = \tilde{F}\tilde{F}^\top \in \mathbb{R}^{THW\times THW}$ . The other is local feature correlation, where we sample a sliding local window with size $T_{p}\times H_{p}\times W_{p}$ for each grid, and obtain the local correlation matrix $C_l\in \mathbb{R}^{THW\times T_pH_pW_p}$ . It is obvious that using the self-attention maps as motion representation substantially outperforms the performance of the other two methods. Compared to the global correlation, the + +Table 5: Ablation on the number of frames $T$ and temporal stride. When sampling more than 3 frames with a larger stride, the method exhibits a diminishing effect. + +
TStrideYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
3137.443.637.133.6
3444.350.143.940.1
3844.250.243.340.7
5444.850.644.341.5
+ +Table 6: Ablation on different motion representations in calculating motion consistency loss ${\mathcal{L}}_{\mathrm{m}}$ . We compare using the self-attention matrix,simple local and global feature correlation as the latent motion representation. + +
MotionRepYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
Local Corr.30.533.829.627.7
Global Corr.39.149.236.433.9
Self-Attention44.350.143.940.1
+ +learnable parameters in attention layers enhance the modeling capacity, which enables the self-attention maps to capture more comprehensive spatio-temporal dependencies and serve as a better motion representation. The local correlation only captures temporal dynamics in limited reception fields, restricting the ability for long-term perception. + +# 5 Conclusion + +In this paper, we introduce a simple yet powerful approach for self-supervised video object segmentation. Specifically, our proposed architecture introduces only a single spatio-temporal Transformer block that ingests frame-wise DINO features and constructs the spatio-temporal correspondence through the attention layer. During training, we adopt a dual self-supervised consistency objective that encompasses both semantic consistency and dynamic motion consistency, supplemented with entropy normalization. In inference, we utilize hierarchical clustering on spatio-temporal attention maps to generate temporally coherent object segmentation masks. This extremely straightforward approach nevertheless delivers state-of-the-art results on challenging real-world multi-object segmentation tasks such as DAVIS-17-Unsupervised and YouTube-VIS-19. + +Limitation. While promising, BA still has room for improvement. Firstly, the current spatio-temporal attention block cannot accommodate very long window-size frames due to the quadratic memory cost. Potentially, adopting memory mechanisms [49, 75] that are capable of handling long-duration data may prove more suitable. Secondly, neither pixel-level constraints nor spatio-temporal dependencies consider the granularity in the video. Incorporating multi-scale features [4] may present a promising direction. + +Social Negative Impact. Although our method produces coherent video object segmentation without annotation, it could also enhance deepfake video technologies, making it easier to produce false videos for malicious intent. It is important to regulate the use of such technology and ensure it is leveraged for beneficial purposes. + +# Acknowledgment + +This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 61931023, Grant 61932022, Grant 62371288, Grant 62320106003, Grant 62301299, Grant T2122024, Grant 62120106007, the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. + +# References + +1. Aydemir, G., Xie, W., Guney, F.: Self-supervised object-centric learning for videos. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=919tWtJPXe +2. Bao, L., Wu, B., Liu, W.: Cnn in mrf: Video object segmentation via inference in a cnn-based higher-order spatio-temporal mrf. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5977-5986 (2018) +3. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021) +4. Bian, Z., Jabri, A., Efros, A.A., Owens, A.: Learning pixel trajectories with multiscale contrastive random walks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6508-6519 (2022) +5. Bielski, A., Favaro, P.: Move: Unsupervised movable object segmentation and detection. Advances in Neural Information Processing Systems 35, 33371-33386 (2022) +6. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017) +7. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650-9660 (2021) +8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) +9. Chen, Y., Liu, Y., Jiang, D., Zhang, X., Dai, W., Xiong, H., Tian, Q.: Sdae: Self-distillated masked autoencoder. In: European conference on computer vision. pp. 108-124. Springer (2022) +0. Choudhury, S., Karazija, L., Laina, I., Vedaldi, A., Rupprecht, C.: Guess What Moves: Unsupervised Video and Image Segmentation by Anticipating Motion. In: British Machine Vision Conference (BMVC) (2022) +11. Ding, S., Li, M., Yang, T., Qian, R., Xu, H., Chen, Q., Wang, J., Xiong, H.: Motion-aware contrastive video representation learning via foreground-background merging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9716-9726 (2022) +12. Ding, S., Qian, R., Xiong, H.: Dual contrastive learning for spatio-temporal representation. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5649-5658 (2022) + +13. Ding, S., Xie, W., Chen, Y., Qian, R., Zhang, X., Xiong, H., Tian, Q.: Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221 (2022) +14. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy +15. Dutt Jain, S., Xiong, B., Grauman, K.: Fusionseg: Learning to combine motion and appearance for fully automatic segmentation of generic objects in videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3664-3673 (2017) +16. Elsayed, G., Mahendran, A., van Steenkiste, S., Greff, K., Mozer, M.C., Kipf, T.: Savi++: Towards end-to-end object-centric learning from real-world videos. Advances in Neural Information Processing Systems 35, 28940-28954 (2022) +17. Fan, D.P., Wang, W., Cheng, M.M., Shen, J.: Shifting more attention to video salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8554-8564 (2019) +18. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020) +19. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., Kipf, T., Kundu, A., Lagun, D., Laradji, I., Liu, H.T.D., Meyer, H., Miao, Y., Nowrouzezahrai, D., Oztireli, C., Pot, E., Radwan, N., Rebain, D., Sabour, S., Sajjadi, M.S.M., Sela, M., Sitzmann, V., Stone, A., Sun, D., Vora, S., Wang, Z., Wu, T., Yi, K.M., Zhong, F., Tagliasacchi, A.: Kubric: a scalable dataset generator (2022) +20. Guo, Y., Yang, C., Rao, A., Agrawala, M., Lin, D., Dai, B.: Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933 (2023) +21. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=Fx2SbBgcte +22. Hamilton, M., Zhang, Z., Hariharan, B., Snavely, N., Freeman, W.T.: Unsupervised semantic segmentation by distilling feature correspondences. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=SaK06z6H10c +23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) +24. Hu, Y., Wang, R., Zhang, K., Gao, Y.: Semantic-aware fine-grained correspondence. In: European Conference on Computer Vision. pp. 97-115. Springer (2022) +25. Hu, Y.T., Huang, J.B., Schwing, A.G.: Videomatch: Matching based video object segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 54-70 (2018) +26. Jabri, A., Owens, A., Efros, A.: Space-time correspondence as a contrastive random walk. Advances in neural information processing systems 33, 19545-19560 (2020) + +27. Johnander, J., Danelljan, M., Brissman, E., Khan, F.S., Felsberg, M.: A generative appearance model for end-to-end video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8953-8962 (2019) +28. Johnson, S.C.: Hierarchical clustering schemes. Psychometrika 32(3), 241-254 (1967) +29. Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jonschkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=aD7uesX1GF_ +30. Lai, Z., Xie, W.: Self-supervised learning for video correspondence flow. In: BMVC (2019) +31. Lai, Z., Lu, E., Xie, W.: Mast: A memory-augmented self-supervised tracker. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6479-6488 (2020) +32. Li, L., Wang, W., Zhou, T., Li, J., Yang, Y.: Unified mask embedding and correspondence learning for self-supervised video segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18706-18716 (2023) +33. Li, X., Loy, C.C.: Video object segmentation with joint re-identification and attention-aware mask propagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 90-105 (2018) +34. Li, X., Liu, S., De Mello, S., Wang, X., Kautz, J., Yang, M.H.: Joint-task self-supervised learning for temporal correspondence. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32. Curran Associates, Inc. (2019), https://proceedings.neurips.cc/paper/2019/file/140f6969d5213fd0ece03148e62e461e-Paper.pdf +35. Li, X., Liu, S., De Mello, S., Wang, X., Kautz, J., Yang, M.H.: Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems 32 (2019) +36. Lian, L., Wu, Z., Yu, S.X.: Bootstrapping objectness from videos by relaxed common fate and visual grouping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14582-14591 (2023) +37. Liu, R., Wu, Z., Yu, S., Lin, S.: The emergence of objectness: Learning zero-shot segmentation from videos. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 13137-13152. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper/2021/file/6d9cb7de5e8ac30bd5e8734bc96a35c1-Paper.pdf +38. Liu, Y., Wang, Y., Chen, Y., Dai, W., Li, C., Zou, J., Xiong, H.: Promoting semantic connectivity: Dual nearest neighbors contrastive learning for unsupervised domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3510-3519 (June 2023) +39. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020) +40. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018) +41. Melas-Kyriazi, L., Rupprecht, C., Laina, I., Vedaldi, A.: Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and local- + +ization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8364-8375 (2022) +42. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +43. Patil, P.W., Dudhane, A., Kulkarni, A., Murala, S., Gonde, A.B., Gupta, S.: An unified recurrent video object segmentation framework for various surveillance environments. IEEE Transactions on Image Processing 30, 7889-7902 (2021) +44. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017) +45. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016) +46. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) +47. Qian, R., Ding, S., Liu, X., Lin, D.: Static and dynamic concepts for self-supervised video representation learning. In: European Conference on Computer Vision. pp. 145-164. Springer (2022) +48. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023) +49. Qian, R., Dong, X., Zhang, P., Zang, Y., Ding, S., Lin, D., Wang, J.: Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009 (2024) +50. Qian, R., Li, Y., Liu, H., See, J., Ding, S., Liu, X., Li, D., Lin, W.: Enhancing self-supervised video representation learning via multi-level feature optimization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7990-8001 (2021) +51. Rambhatla, S.S., Misra, I., Chellappa, R., Shrivastava, A.: Most: Multiple object localization with self-supervised transformers for object discovery. arXiv preprint arXiv:2304.05387 (2023) +52. Salehi, M., Gavves, E., Snoek, C.G., Asano, Y.M.: Time does tell: Self-supervised time-tuning of dense image representations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16536-16547 (2023) +53. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence 22(8), 888-905 (2000) +54. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. In: BMVC 2021-32nd British Machine Vision Conference (2021) +55. Singh, G., Wu, Y.F., Ahn, S.: Simple unsupervised object-centric learning for complex and naturalistic videos. Advances in Neural Information Processing Systems 35, 18181-18196 (2022) +56. Tian, J., Aggarwal, L., Colaco, A., Kira, Z., Gonzalez-Franco, M.: Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv preprint arXiv:2308.12469 (2023) +57. Vondrick, C., Shrivastava, A., Fathi, A., Guadarrama, S., Murphy, K.: Tracking emerges by colorizing videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 391-408 (2018) + +58. Wang, N., Song, Y., Ma, C., Zhou, W., Liu, W., Li, H.: Unsupervised deep tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1308-1317 (2019) +59. Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle-consistency of time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2566-2576 (2019) +60. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3124-3134 (June 2023) +61. Wang, X., Misra, I., Zeng, Z., Girdhar, R., Darrell, T.: Videocutler: Surprisingly simple unsupervised video instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22755-22764 (2024) +62. Wang, Y., Shen, X., Yuan, Y., Du, Y., Li, M., Hu, S.X., Crowley, J.L., Vaufreydaz, D.: Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +63. Wang, Y., Li, J., ZHANG, X., Shi, B., Li, C., Dai, W., Xiong, H., Tian, Q.: BarleRiA: An efficient tuning framework for referring image segmentation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=wHLDHRkmEu +64. Xie, C., Xiang, Y., Harchaoui, Z., Fox, D.: Object discovery in videos as foreground motion clustering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9994-10003 (2019) +65. Xie, J., Xie, W., Zisserman, A.: Segmenting moving objects via an object-centric layered representation. In: Advances in Neural Information Processing Systems (2022) +66. Xu, H., Ding, S., Zhang, X., Xiong, H., Tian, Q.: Masked autoencoders are robust data augmentors. arXiv preprint arXiv:2206.04846 (2022) +67. Xu, J., Wang, X.: Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10075-10085 (2021) +68. Yang, C., Lamdouar, H., Lu, E., Zisserman, A., Xie, W.: Self-supervised video object segmentation by motion grouping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7177-7188 (October 2021) +69. Yang, L., Fan, Y., Xu, N.: Video instance segmentation. In: ICCV (2019) +70. Yang, Y., Lai, B., Soatto, S.: Dystab: Unsupervised object segmentation via dynamic-static bootstrapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2826-2836 (June 2021) +71. Yang, Y., Loquercio, A., Scaramuzza, D., Soatto, S.: Unsupervised moving object detection via contextual information separation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +72. Ye, V., Li, Z., Tucker, R., Kanazawa, A., Snavely, N.: Deformable sprites for unsupervised video decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2657-2666 (2022) +73. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=1_jFneF07YC + +74. Zadaianchuk, A., Seitzer, M., Martius, G.: Object-centric learning for real-world videos by predicting temporal feature similarities. In: Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023) (2023) +75. Zhou, J., Pang, Z., Wang, Y.X.: Rmem: Restricted memory banks improve video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18602-18611 (2024) +76. Ziegler, A., Asano, Y.M.: Self-supervised learning of object parts for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14502-14511 (2022) \ No newline at end of file diff --git a/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/images.zip b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ee434442668ae0baed2115129c2a065250855f92 --- /dev/null +++ b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0de14cc1d293110133df129302e22a6b150e0ef25f3939923ac9e3ef92ba7a06 +size 424002 diff --git a/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/layout.json b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2e15f6519c4a7aa751256036aabfe11402da1648 --- /dev/null +++ b/2024/Betrayed by Attention_ A Simple yet Effective Approach for Self-supervised Video Object Segmentation/layout.json @@ -0,0 +1,10327 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 470, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 470, + 166 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 470, + 166 + ], + "type": "text", + "content": "Betrayed by Attention: A Simple yet Effective Approach for Self-supervised Video Object Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "spans": [ + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "text", + "content": "Shuangrui Ding" + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "text", + "content": ", Rui Qian" + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "text", + "content": ", Haohang Xu" + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "text", + "content": ", Dahua Lin" + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "inline_equation", + "content": "^{1,3,4\\dagger}" + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "text", + "content": ", and Hongkai Xiong" + }, + { + "bbox": [ + 181, + 187, + 432, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 182, + 220, + 430, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 220, + 430, + 232 + ], + "spans": [ + { + "bbox": [ + 182, + 220, + 430, + 232 + ], + "type": "text", + "content": "1 The Chinese University of Hong Kong, Hong Kong, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 204, + 232, + 410, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 232, + 410, + 243 + ], + "spans": [ + { + "bbox": [ + 204, + 232, + 410, + 243 + ], + "type": "text", + "content": "2 Shanghai Jiao Tong University, Shanghai, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 243, + 434, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 243, + 434, + 255 + ], + "spans": [ + { + "bbox": [ + 180, + 243, + 434, + 255 + ], + "type": "text", + "content": "3 Shanghai Artificial Intelligence Laboratory, Shanghai, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 255, + 468, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 255, + 468, + 277 + ], + "spans": [ + { + "bbox": [ + 146, + 255, + 468, + 277 + ], + "type": "text", + "content": "4 Centre for Perceptual and Interactive Intelligence (CPII), Hong Kong, China {ds023, qr021}@ie.cuhk.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 318, + 455, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 318, + 455, + 582 + ], + "spans": [ + { + "bbox": [ + 159, + 318, + 455, + 582 + ], + "type": "text", + "content": "Abstract. In this paper, we propose a simple yet effective approach for self-supervised video object segmentation (VOS). Previous self-supervised VOS techniques majorly resort to auxiliary modalities or utilize iterative slot attention to assist in object discovery, which restricts their general applicability. To deal with these challenges, we develop a simplified architecture that capitalizes on the emerging objectness from DINO-pretrained Transformers, bypassing the need for additional modalities or slot attention. Our key insight is that the inherent structural dependencies present in DINO-pretrained Transformers can be leveraged to establish robust spatio-temporal correspondences in videos. Furthermore, simple clustering on this correspondence cue is sufficient to yield competitive segmentation results. Specifically, we first introduce a single spatio-temporal Transformer block to process the frame-wise DINO features and establish spatio-temporal dependencies in the form of self-attention. Subsequently, utilizing these attention maps, we implement hierarchical clustering to generate object segmentation masks. To train the spatio-temporal block in a fully self-supervised manner, we employ semantic and dynamic motion consistency coupled with entropy normalization. Our method demonstrates state-of-the-art performance across three multi-object video segmentation tasks. Specifically, we achieve over 5 points of improvement in terms of FG-ARI on complex real-world DAVIS-17-Unsupervised and YouTube-VIS-19 compared to the previous best result. The code and checkpoint are released at https://github.com/shvdiwnkozbw/SSL-UVOS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 594, + 436, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 594, + 436, + 605 + ], + "spans": [ + { + "bbox": [ + 160, + 594, + 436, + 605 + ], + "type": "text", + "content": "Keywords: Video Object Segmentation " + }, + { + "bbox": [ + 160, + 594, + 436, + 605 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 594, + 436, + 605 + ], + "type": "text", + "content": " Self-supervised Learning" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 225, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 225, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 225, + 653 + ], + "type": "text", + "content": "* Equal Contribution" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 653, + 355, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 355, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 355, + 665 + ], + "type": "text", + "content": "† Corresponding author. Email: dhlin@ie.cuhk.edu.hk" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 191, + 206 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 191, + 206 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 191, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 191, + 206 + ], + "type": "image", + "image_path": "617556ba29e3068a86f00013c0bda63f615c151c7e85f5ca448fc00d6db67c48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 208, + 178, + 215 + ], + "lines": [ + { + "bbox": [ + 145, + 208, + 178, + 215 + ], + "spans": [ + { + "bbox": [ + 145, + 208, + 178, + 215 + ], + "type": "text", + "content": "(a) Video Frame" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 249, + 206 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 249, + 206 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 249, + 206 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 249, + 206 + ], + "type": "image", + "image_path": "94dc76dc03c4752b3232035c31e681a3ca1c088b99fa37e08eeb34db2d5fb825.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 208, + 240, + 214 + ], + "lines": [ + { + "bbox": [ + 200, + 208, + 240, + 214 + ], + "spans": [ + { + "bbox": [ + 200, + 208, + 240, + 214 + ], + "type": "text", + "content": "(b) DINO Attention" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 251, + 114, + 308, + 206 + ], + "blocks": [ + { + "bbox": [ + 251, + 114, + 308, + 206 + ], + "lines": [ + { + "bbox": [ + 251, + 114, + 308, + 206 + ], + "spans": [ + { + "bbox": [ + 251, + 114, + 308, + 206 + ], + "type": "image", + "image_path": "118cf47bd5adbaa47eee8f770320b14d420d09017c790a417f11d48bc22635a8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 304 + ], + "type": "text", + "content": "Fig. 1: Attention leaks the object's position! We visualize the self-attention maps of different queries (■ prompt) from the video sequence (a). The frame-wise DINO attention maps (b) highlight image regions corresponding to the queried object. A randomly initialized spatio-temporal Transformer block on top of DINO produces noisy spatio-temporal attention maps (c) that coarsely track objects over time. Our method diminishes noise in the learned spatio-temporal attention maps (d) which derive temporally coherent object segmentation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 114, + 365, + 206 + ], + "blocks": [ + { + "bbox": [ + 259, + 208, + 356, + 215 + ], + "lines": [ + { + "bbox": [ + 259, + 208, + 356, + 215 + ], + "spans": [ + { + "bbox": [ + 259, + 208, + 356, + 215 + ], + "type": "text", + "content": "(c) Random Initialized Spatio-temporal Attention" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 114, + 365, + 206 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 365, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 365, + 206 + ], + "type": "image", + "image_path": "7881613c9aab4294a133be65b35c520741e98fd9e1e522999a8c976d6c41debd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 208, + 356, + 215 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 208, + 356, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 208, + 356, + 215 + ], + "type": "text", + "content": "Ratio-temporal Attention" + } + ] + } + ], + "index": 9, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 424, + 206 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 424, + 206 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 424, + 206 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 424, + 206 + ], + "type": "image", + "image_path": "484b88b39daec6bdba722480e537bb1d81e7180df283dbb26e11cc62c3a253f0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 208, + 460, + 215 + ], + "lines": [ + { + "bbox": [ + 386, + 208, + 460, + 215 + ], + "spans": [ + { + "bbox": [ + 386, + 208, + 460, + 215 + ], + "type": "text", + "content": "(d) Learned Spatio-temporal Attention" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 114, + 481, + 206 + ], + "blocks": [ + { + "bbox": [ + 425, + 114, + 481, + 206 + ], + "lines": [ + { + "bbox": [ + 425, + 114, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 425, + 114, + 481, + 206 + ], + "type": "image", + "image_path": "4560323e676730cf9c4d3f8dc1c02873cb478f00c885c5f3948551d46a00285a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 327, + 230, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 230, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 230, + 340 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "content": "Representing the visual scene with objects as the basic elements has been long acknowledged as a core cognitive capability of an intelligent agent. In the realm of computer vision, Video Object Segmentation (VOS) tasks require a model to segment and track specified objects within a video sequence, striving to emulate this foundational function. This task holds significant importance in various real-world vision systems, including but not limited to, autonomous driving [18] and surveillance security [43]. However, traditional methods for VOS [6,44] typically follow a fully supervised paradigm and entail substantial costs for obtaining pixel-level per-frame annotations. This factor significantly limits their feasibility for large-scale applications. Thus, the researchers turn their attention to self-supervised learning [8,9,11,12,23,38,47,50,66], a more efficient approach that leverages unlabeled data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Despite its promising potential, self-supervised VOS poses significant challenges. Firstly, previous methods tend to incorporate additional signals such as optical flow [36, 68, 71], or depth cues [16, 29], to furnish object-related clues. However, these auxiliary signals within videos can often be difficult to extract and unreliable when it comes to multi-object scenarios, thus limiting their availability. Secondly, most existing models [1,48,74] adopt slot attention [39] to parse the frame into objects. Nevertheless, the requirement for a predefined quantity of learnable slot queries limits its adaptability to in-the-wild data, especially when the number of objects is unknown. To address these limitations, we recall the recent self-supervised learning technique DINO [7, 42], which demonstrates emerging objectness in the attention maps of pretrained Vision Transformer (ViT) [14]. As depicted in Fig. 1(b), DINO attention maps encode spatial dependencies between diverse patches. These dependencies present different patterns for various objects, providing abundant cues for object segmentation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "This phenomenon motivates us with an intriguing possibility: Can we utilize the object-aware attributes of DINO to learn robust spatio-temporal dependencies and thereby produce coherent object segmentations in videos?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 379 + ], + "type": "text", + "content": "In this paper, we delve into the potential of this idea with a simple architecture. Based on DINO pretrained ViT, we introduce a single spatio-temporal Transformer block to process frame-wise DINO features and calculate spatiotemporal attention maps. Remarkably, these attention maps in Fig. 1(c) showcase the ability to coarsely discriminate and track object parts even with random initialization. Based on this observation, we can roughly estimate the patches belonging to the same object over time. Then, we encourage both semantic and motion alignment among patches belonging to the same object while distinguishing those from different objects in a fully self-supervised manner. The training significantly diminishes noise in the learned spatio-temporal dependencies as displayed in Fig. 1(d). In inference, we directly apply Hierarchical Clustering [28] to these spatio-temporal attention maps to derive object segmentation masks within a sequence. Interestingly, this naive clustering method yields unexpectedly competitive results. Our method greatly outperforms slot attention-based methods on various datasets, e.g., 9 points improvement over SOLV [1] on DAVIS-17, 12 points improvement over SMTC [48] on YTVIS-19. We argue that clustering on the whole spatio-temporal attention maps, instead of frame-by-frame feature clustering as in slot attention ensures the temporal coherence of produced segmentations and guarantees the object's permanence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 536 + ], + "type": "text", + "content": "In summary, our proposed method BA (short for Betrayed by Attention) offers three distinct advantages: (1) BA solely requires RGB frames and operates entirely under self-supervision without the need for external modality. This characteristic renders our method highly accessible and adaptable in real-world scenarios; (2) The architecture is remarkably simple and systematically efficient. We only introduce a single learnable Transformer block and harness a parameter-free clustering algorithm to discover the objects. It facilitates easy generalization to multi-object scenes, without the constraints imposed by slot attention; (3) Our method achieves state-of-the-art results across a bunch of unsupervised video multi-object segmentation benchmarks, such as MOVi-E [19], DAVIS-17-Unsupervised [46], and YouTube-VIS-19 [69]. Notably, we surpass the previous state-of-the-art method, TimeT [52], by 6.6 points on YTVIS-19 and by 6.1 points on DAVIS-17 in terms of the FG-ARI metric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 555, + 237, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 237, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 237, + 568 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Video Object Segmentation aims to segment objects coherently in a video sequence [2,6,15,17,20,21,25,27,33,44]. In video object segmentation (VOS), there are two prevalent protocols for evaluating the learned models: Semi-supervised VOS and Unsupervised VOS. In semi-supervised VOS, the algorithm is provided with the object masks in the first frame, and required to track them in subsequent frames. In contrast, unsupervised VOS aims to identify and segment salient objects from the background without any specific reference. In this paper, we fo" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "text", + "content": "Betrayed by Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "cus on the more challenging unsupervised VOS setting without using any kind of manual annotations in either training or inference. Recently, there emerges a line of self-supervised algorithms for unsupervised VOS [10, 29, 34, 58, 64, 68, 70-72]. CIS [71] facilitates fully unsupervised motion segmentation, discarding object mask supervision during training. By formulating a min-max game of mutual information, the generator is motivated to create segmentations that effectively distinguish foreground objects from the background. AMD [37] minimizes the warping synthesis error to train appearance and motion pathways without any supervision. GWM [10] uses a single RGB image as input with optical flow acting as supervision to highlight moving areas. MG [68] and OCLR [65] solely leverage optical flow as input to generate object-centric layered representations, with each layer indicating a potential object. Despite the impressive performance of state-of-the-art self-supervised VOS methods [1, 48, 52] in discovering single objects, they still face limitations when it comes to solving multi-object discovery tasks. In contrast, by leveraging the rich object cues in attention maps, our method can discern multiple objects in real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 529 + ], + "type": "text", + "content": "Self-supervised Spatio-temporal Correspondence learning usually leverages free temporal supervision signals in videos to learn representations that facilitate accurate pixel or object tracking across space and time [4, 7, 13, 24, 26, 30, 31, 35, 57, 59, 67]. Vondrick et al. [57] employ the natural temporal color coherence to train a colorization model on grayscale videos, thereby establishing fine-grained correspondence between current and future frames. CRW [26] presents a self-supervised learning approach for dense correspondence in raw videos. It uses space-time graph-based random walks and cycle consistency to implicitly supervise chains of comparisons. Hu et al. [24] independently learn semantics and temporal correspondence from two pathways and fuse them at a later stage. UME [32] respectively designs short-term appearance and long-term semantic consistency to learn generalizable correspondence. Taking a step further by integrating high-level semantics with low-level temporal correlation, SMTC [48] develops two-stage slot attention to establish dense correspondence with more emphasis on foreground objects. In this work, we explicitly initialize the spatio-temporal correspondence on top of the DINO pretrained Transformer. This initialization effectively guides the learning process with a self-supervised pixel-level consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Object-aware DINO Features have shown efficacy in object localization tasks, particularly in image domain [5, 22, 41, 54, 60, 63, 73, 76]. DINO's original paper [7] reveals that the learned representations and attention maps carry substantial object cues that can be harnessed for object discovery. The groundbreaking work, LOST [54], utilizes these features from DINO to perform object segmentation by constructing a graph wherein objects are segmented using the inverse degrees of nodes. TokenCut [62] employs DINO features for applying the Normalized Cut algorithm [53], thereby obtaining foreground segments in an image. Furthermore, CutLER [60] and MOST [51] can localize multiple objects without any supervision based on the DINO features. Motivated by the successful application of DINO features in image settings, recent works have expanded" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "their use to video segmentation tasks [1,48,52,61,62]. SMTC [48] utilizes the rich semantics and correspondence cues in the DINO features to provide a reliable reference for object decomposition in consecutive frames. TimeT [52] introduces a feature forwarding process to propagate dense DINO features across time and establish temporal consistency. Most of the existing works focus on leveraging DINO feature vectors for optimization and adaptation to videos. While in our work, we lay more emphasis on the self-attention maps. We extend the framework-wise DINO attention maps to spatio-temporal maps, which explicitly encode space-time structural dependencies and help coherent object discovery in videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 241, + 202, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 202, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 202, + 254 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 266, + 354, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 266, + 354, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 354, + 278 + ], + "type": "text", + "content": "3.1 Preliminary on DINO Attention Maps" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 392 + ], + "type": "text", + "content": "The self-supervised ViT, as achieved by DINO [7], has exhibited a rich emergence of objectness in the attention maps. To elucidate this, we visualize the self-attention map from the final Transformer block of the DINO pretrained ViT-S/8 as a representative. As illustrated in Fig. 1(b), an image patch exhibits a high degree of attention dependencies with patches of the same object. Conversely, image patches of distinct objects tend to present different attention distributions. Encouraged by this observation, we pose the following question: Could it be possible to directly harness these attention maps to achieve object segmentation without relying on annotations?" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 133, + 412, + 325, + 472 + ], + "blocks": [ + { + "bbox": [ + 133, + 412, + 325, + 472 + ], + "lines": [ + { + "bbox": [ + 133, + 412, + 325, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 412, + 325, + 472 + ], + "type": "table", + "html": "
Clustering MetricCosine similarityKL divergence
FAAvAv
DAVIS-1647.460.351.175.4
DAVIS-1714.726.720.939.2
", + "image_path": "2b829db8e5c8dc451d0e4114155b96f855c1fe1f5cb886391f19d19931f2e784.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "lines": [ + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "text", + "content": "Table 1: Preliminary video object segmentation results using varying clustering metrics. " + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "text", + "content": " represents features pretrained with DINO. " + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "text", + "content": " denotes per-frame attention maps from DINO, while " + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "inline_equation", + "content": "A_{v}" + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "inline_equation", + "content": "\\tilde{A}_v" + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "text", + "content": " denote spatio-temporal attention maps produced by a randomly initialized and a final learned temporal correlator respectively, on top of DINO. The IoU (" + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 132, + 473, + 326, + 583 + ], + "type": "text", + "content": " score) is reported for both DAVIS-16 and DAVIS-17-Unsupervised datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 334, + 412, + 407, + 495 + ], + "blocks": [ + { + "bbox": [ + 334, + 412, + 407, + 495 + ], + "lines": [ + { + "bbox": [ + 334, + 412, + 407, + 495 + ], + "spans": [ + { + "bbox": [ + 334, + 412, + 407, + 495 + ], + "type": "image", + "image_path": "c7ed16abeed06e7fec0e373b28216579cd1e477319c01b41e4bbe0cb7b01eeac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 495, + 398, + 503 + ], + "lines": [ + { + "bbox": [ + 343, + 495, + 398, + 503 + ], + "spans": [ + { + "bbox": [ + 343, + 495, + 398, + 503 + ], + "type": "text", + "content": "(a) DINO Feature " + }, + { + "bbox": [ + 343, + 495, + 398, + 503 + ], + "type": "inline_equation", + "content": "F" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "lines": [ + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "type": "text", + "content": "Fig.2: Visualizations of the clustering results. The left column is the results of DINO features " + }, + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "type": "text", + "content": ", the right column is the results of DINO attention " + }, + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "type": "text", + "content": " results in much noisier clusters, while " + }, + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 332, + 506, + 482, + 584 + ], + "type": "text", + "content": " distinguishes different classes of objects." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 408, + 412, + 482, + 495 + ], + "blocks": [ + { + "bbox": [ + 408, + 412, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 408, + 412, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 408, + 412, + 482, + 495 + ], + "type": "image", + "image_path": "2472acf7534ce171972bb6f951e4e70ce1c6f3b69b910271f54c98ea5c10f94b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 495, + 474, + 503 + ], + "lines": [ + { + "bbox": [ + 414, + 495, + 474, + 503 + ], + "spans": [ + { + "bbox": [ + 414, + 495, + 474, + 503 + ], + "type": "text", + "content": "(b) DINO Attention A" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In response to this question, we conduct a straightforward initial exploration with these attention maps. We denote the attention map of the last Transformer block as " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{HW \\times HW}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and its output features " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " respectively denote the height, width, and channel dimension of the feature" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "text", + "content": "Betrayed by Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 479, + 203 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 203 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 203 + ], + "type": "image", + "image_path": "b0315186df5a53734e34c6da14f8a914d056d5cecd0dc4ac38ee4527f324623d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 211, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 310 + ], + "type": "text", + "content": "Fig. 3: Our architecture BA overview. Given the video frames and a DINO pretrained Transformer, we first use a temporal correlator to construct the spatio-temporal correspondence. We then utilize these attention maps as a clustering metric and apply hierarchical clustering across all frames to generate segmentation masks. During training, for each patch, we sample a positive/negative set and assign an importance weight based on its corresponding attention map. We promote alignment within the positive set while differentiating the representations from the negative set. The final loss, normalized with importance weights, only trains the temporal correlator with the DINO ViT remaining frozen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "text", + "content": "map5. For an intuitive comparison, we apply unsupervised clustering on both attention maps " + }, + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "text", + "content": " to yield the object segmentation masks. For the detailed procedure, please refer to the Sec. 3.2. Then, we report frame-wise evaluation results on the video object segmentation datasets DAVIS-16 [45] and DAVIS-17-Unsupervised [46] in Table 1. Interestingly, we observe that the attention maps significantly excel in decoupling object components compared to features, exhibiting an improvement of over 10 points on both datasets in terms of IoU. Surprisingly, a spatio-temporal attention map " + }, + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "inline_equation", + "content": "A_v" + }, + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "text", + "content": " generated by a randomly initialized Transformer block on top of DINO also outperforms the frame-wise DINO feature " + }, + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 334, + 482, + 526 + ], + "type": "text", + "content": ". And it is further demonstrated in Fig. 2. The DINO features lead to very noisy clusters, while attention maps can discriminate objects of different categories. We speculate this is because the attention maps preserve intact correspondences, which are more suitable for fine-grained tasks such as object grouping and discrimination. Therefore, creating spatio-temporal structural dependencies that expand beyond spatial DINO attention maps appears promising for video object segmentation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 544, + 331, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 331, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 331, + 556 + ], + "type": "text", + "content": "3.2 Our Simple Architecture for VOS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "text", + "content": "Motivated by the frame-wise preliminary result, we introduce a temporal correlator instantiated with a spatio-temporal Transformer block [3] on top of the DINO pretrained ViT backbone, to extend the self-attention calculation to the spatiotemporal domain. Specifically, given the frame-wise DINO features " + }, + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\{F_1,\\dots ,F_T\\}" + }, + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "text", + "content": " frames, we process them through a spatio-temporal Transformer block " + }, + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{st}}" + }, + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "text", + "content": ". It consists of one standard ViT encoder layer, with self-attention computed over" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "Here we omit the CLS token, and average the multi-head attention along the head dimension." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": "the spatio-temporal grids. Specifically, we concatenate the frame-wise DINO features along the temporal dimension with the addition of learnable temporal positional embeddings " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{pe}" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "F = \\{F_{1} + \\mathbf{pe}_{1},\\dots ,F_{T} + \\mathbf{pe}_{T}\\} \\in \\mathbb{R}^{THW\\times C}" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " and then linearly project " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " to obtain the queries " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "Q = f_{Q}(F)" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "K = f_{K}(F)" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ", and values " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "V = f_{V}(F)" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ". Then, we can obtain the self-attention map " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "A_{v}\\in \\mathbb{R}^{THW\\times THW}" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " and temporally fused features " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "F_{v}\\in \\mathbb{R}^{THW\\times C}" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 196, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 196, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 189, + 196, + 481, + 224 + ], + "type": "interline_equation", + "content": "A _ {v} = \\mathrm {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {C}}\\right), \\quad F _ {v} = A _ {v} V = \\{\\tilde {F} _ {1}, \\dots , \\tilde {F} _ {T} \\}. \\qquad (1)", + "image_path": "e615e307ded4790026375688f3f4a4790068e72c5a5966fc5541ff94c77babaa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 231, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 231, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 480, + 266 + ], + "type": "text", + "content": "In this way, we establish spatio-temporal structural dependencies beyond frameworkwise DINO features and have the potential to improve object permanence across the entire time span." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 266, + 481, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 266, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 481, + 315 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 131, + 266, + 481, + 315 + ], + "type": "inline_equation", + "content": "A_v[i] \\in \\mathbb{R}^{THW}" + }, + { + "bbox": [ + 131, + 266, + 481, + 315 + ], + "type": "text", + "content": " is a proximity distribution that describes the structural dependencies between the " + }, + { + "bbox": [ + 131, + 266, + 481, + 315 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 266, + 481, + 315 + ], + "type": "text", + "content": "-th patch and all other positions, we adopt symmetric KL divergence to measure the distance " + }, + { + "bbox": [ + 131, + 266, + 481, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{THW \\times THW}" + }, + { + "bbox": [ + 131, + 266, + 481, + 315 + ], + "type": "text", + "content": " between any two patches as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 202, + 325, + 481, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 325, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 202, + 325, + 481, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {M} [ i, j ] = \\mathcal {D} _ {K L} \\left(A _ {v} [ i ] | | A [ j ]\\right) + \\mathcal {D} _ {K L} \\left(A _ {v} [ j ] | | A _ {v} [ i ]\\right), \\tag {2}", + "image_path": "ace8405cf02c70a515fd80d791337e50bf648276d51aca78570923451620f4f3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{KL}" + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": " denotes KL divergence. Based on this distance matrix " + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": ", we then apply Hierarchical Clustering [28] to obtain the object segmentation masks. Please consult our appendix for a detailed implementation of the hierarchical clustering algorithm. In this way, we iteratively merge " + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "inline_equation", + "content": "THW" + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": " patches into " + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": " centroids " + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "inline_equation", + "content": "A_{c} \\in \\mathbb{R}^{N \\times THW}" + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": ". Since each centroid likely represents the spatial distribution of an object, we consider the cluster assignments as potential candidates for the object segmentation mask. Specifically, we calculate the KL divergence distance between the " + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "inline_equation", + "content": "A_{v}" + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "inline_equation", + "content": "A_{c}" + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}^{THW \\times N}" + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": ", and use argmin operation to produce the cluster assignments " + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 131, + 346, + 482, + 455 + ], + "type": "text", + "content": " within the video sequence:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 463, + 481, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 463, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 219, + 463, + 481, + 477 + ], + "type": "interline_equation", + "content": "Z = \\operatorname {a r g m i n} (D, \\dim = 1) \\in \\{1, \\dots , N \\} ^ {T H W}. \\tag {3}", + "image_path": "47d9083d30a7995d95a5838b5c5f057e46cd5eeadb5b1afeff41e46d38dfb7cc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 486, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 480, + 521 + ], + "type": "text", + "content": "Then, we employ a bipartite matching mechanism, following the standard evaluation protocol in DAVIS-17 Challenge [46], to match the predicted masks with ground truth for evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 521, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 482, + 617 + ], + "type": "text", + "content": "A recent study, DiffSeg [56], used hierarchical clustering to process attention maps from a stable diffusion model, but it operated solely in the image domain. Our work extends the powerful image foundation model to the spatio-temporal domain, demonstrating that accurate spatio-temporal correspondence can be learned without supervision. Additionally, our method does not rely on hierarchical clustering. As shown in Table 4, using K-means clustering also achieves competitive performance, indicating that the learned spatio-temporal correspondence is effective across various clustering strategies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "Besides, there are two noteworthy aspects in our clustering stage. First, in the clustering process, we only set a distance threshold hyper-parameter without specifying the number of cluster centroids and the hierarchical clustering algorithm can adaptively generate different numbers of clusters according to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 354, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 354, + 91, + 447, + 102 + ], + "type": "text", + "content": "Betrayed by Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "the complexity of the video. Second, our reference to the entire spatio-temporal distribution enables coherent object segmentation across multiple frames. As a result, the generated segmentation masks effortlessly track corresponding objects over time without extra association procedures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 181, + 204, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 204, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 204, + 193 + ], + "type": "text", + "content": "3.3 Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "text", + "content": "Noticing that even a randomly initialized attention map " + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "inline_equation", + "content": "A_v" + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "text", + "content": " shown in Figure 1(c), exhibits an initial ability to differentiate parts and track objects over time, we leverage these maps as indicators for positive and negative sets sampling. Mathematically, given the attention maps " + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "inline_equation", + "content": "A_v \\in \\mathbb{R}^{THW \\times THW}" + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "text", + "content": ", we first reshape " + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "inline_equation", + "content": "A_v" + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "text", + "content": " into shape " + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "inline_equation", + "content": "THW \\times T \\times HW" + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "text", + "content": " then use top-k operation to retrieve " + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "inline_equation", + "content": "K_p" + }, + { + "bbox": [ + 130, + 201, + 482, + 273 + ], + "type": "text", + "content": " position indexes in each frame with highest attention scores:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 282, + 481, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 282, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 223, + 282, + 481, + 295 + ], + "type": "interline_equation", + "content": "I _ {p} = \\operatorname {t o p - k} (\\text {r e s h a p e} (A _ {v}), K _ {p}, \\dim = - 1). \\tag {4}", + "image_path": "3d00cf3d79375c28ee8ffdafc35fe382502695876a6d50a83ef2dd591eb59e3d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": "In this way, we obtain a position index set " + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "inline_equation", + "content": "I_{p} \\in \\mathbb{R}^{THW \\times T \\times K_{p}}" + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": ", which represents positions potentially associated with the same object as the " + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": "-th patch. The reshape operation ensures that we consistently extract the same number of patches in each frame. This, in turn, guarantees effective object tracking within consecutive frames and promotes consistent temporal correspondence learning. Similarly, we perform this operation to find " + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "inline_equation", + "content": "K_{n}" + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": " positions in each frame with the lowest attention scores " + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "inline_equation", + "content": "I_{n} \\in \\mathbb{R}^{THW \\times T \\times K_{n}}" + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "inline_equation", + "content": "I_{n}[i]" + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": " represents positions least likely to belong to the same object as the " + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": "-th patch. After obtaining the positive and negative index sets, we apply self-supervised correspondence learning in two aspects: semantic consistency and dynamic motion consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": "Semantic consistency. Intuitively, we take the temporally fused features " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "F_{v} = \\{\\tilde{F}_{1},\\dots,\\tilde{F}_{T}\\} \\in \\mathbb{R}^{THW\\times C}" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " as the semantic representation. We promote the semantic alignment between the features of the same object and distinguish features from different objects. Given the " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": "-th patch, we first refer to the positive pair indexes " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "I_{p}[i]" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " in Eq. 4 to gather the features that are likely from the same object. Then we calculate the cosine distance between the query feature and all positive feature vectors, and produce the distance matrix " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "S_{p}[i]\\in \\mathbb{R}^{TK_{p}}" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ". Similarly, we gather distinct features using " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "I_{n}[i]" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ", and form negative distance matrix " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "S_{n}[i]" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ". We employ a simple margin loss to enforce higher consistency between corresponding object areas:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 206, + 552, + 481, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 552, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 206, + 552, + 481, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s}} [ i ] = \\sum_ {j = 1} ^ {T K _ {p}} \\sum_ {k = 1} ^ {T K _ {n}} \\max \\left(\\mathcal {S} _ {p} [ i, j ] - \\mathcal {S} _ {n} [ i, k ] + \\lambda_ {1}, 0\\right), \\tag {5}", + "image_path": "b3fc41308fd5bf5c221ddcb7be0fffa75fc2fb21fc969fdf8f32d7b363943be0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 448, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 448, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 448, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 594, + 448, + 605 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 594, + 448, + 605 + ], + "type": "text", + "content": " is the margin hyper-parameter and we set " + }, + { + "bbox": [ + 130, + 594, + 448, + 605 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.8" + }, + { + "bbox": [ + 130, + 594, + 448, + 605 + ], + "type": "text", + "content": " as default." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Dynamic motion consistency. Regarding dynamic motions, rather than calculating optical flow, inter-frame feature correlations, or latent cost volumes as motion representations, we turn to highly accessible attention maps " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_v" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": ". These maps exhibit spatio-temporal correlations of specific patches, serving as an effective latent representation of temporal dynamics. Therefore, it is feasible to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": "directly employ these attention maps to bolster motion consistency between corresponding objects. Similar to the semantic consistency learning, given the attention map of " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": "-th patch " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "A_v[i]" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": " as a query, we also leverage the positive (negative) pair indexes " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "I_p[i] \\left(I_n[i]\\right)" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": " to gather the attention maps of the corresponding (distinct) objects. Then we calculate the symmetric KL divergence as in Eq. 2 between the query and all positive (negative) attention maps, obtaining the distance matrix " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_p[i] \\in \\mathbb{R}^{TK_p}" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_n[i] \\in \\mathbb{R}^{TK_n}" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": ". We utilize the same margin loss for optimization, with " + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\lambda_2 = 1.0" + }, + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": " as the margin hyper-parameter:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 219, + 481, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 219, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 198, + 219, + 481, + 253 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m}} [ i ] = \\sum_ {j = 1} ^ {T K _ {p}} \\sum_ {k = 1} ^ {T K _ {n}} \\max \\left(\\mathcal {M} _ {p} [ i, j ] - \\mathcal {M} _ {n} [ i, k ] + \\lambda_ {2}, 0\\right). \\tag {6}", + "image_path": "c3345b74d4ad3a6dd112fb29a2f478988783154bbcb7499d47e5143fa35caa6b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 259, + 485, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 259, + 485, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 259, + 485, + 308 + ], + "type": "text", + "content": "Overall objectives. Besides, we evaluate the importance of each spatio-temporal patch and lay more emphasis on the informative regions. Formally, we employ the entropy of the attention maps as a measure of the information content within each patch:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 238, + 315, + 481, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 238, + 315, + 481, + 346 + ], + "type": "interline_equation", + "content": "e [ i ] = \\sum_ {k = 1} ^ {T H W} - A _ {v} [ i, k ] \\log A _ {v} [ i, k ], \\tag {7}", + "image_path": "a4bbaa829ecb931e03ac5e2dae6f4140de0699ade02c309de78992fc219b1ff2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 353, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 353, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 131, + 353, + 482, + 401 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 353, + 482, + 401 + ], + "type": "inline_equation", + "content": "e[i]" + }, + { + "bbox": [ + 131, + 353, + 482, + 401 + ], + "type": "text", + "content": " denoting the entropy of " + }, + { + "bbox": [ + 131, + 353, + 482, + 401 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 353, + 482, + 401 + ], + "type": "text", + "content": "-th patch. A higher entropy value indicates more ambiguous spatio-temporal correlations, which provide less information. Hence, we perform softmax normalization on the reversed entropy to produce the importance weight " + }, + { + "bbox": [ + 131, + 353, + 482, + 401 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 131, + 353, + 482, + 401 + ], + "type": "text", + "content": " for each patch as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 250, + 407, + 481, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 407, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 250, + 407, + 481, + 437 + ], + "type": "interline_equation", + "content": "w [ i ] = \\frac {\\exp (- e [ i ])}{\\sum_ {j = 1} ^ {T H W} \\exp (- e [ j ])}. \\tag {8}", + "image_path": "13ba517f13c20848c9826d50b94495675e4e003ede72c2974d134fe1f414183f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 442, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 482, + 468 + ], + "type": "text", + "content": "Then, the overall learning objectives on semantic and motion consistency can be formulated as the weighted summation over all spatio-temporal patches:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 245, + 474, + 481, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 474, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 245, + 474, + 481, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {T H W} w [ i ] \\left(\\mathcal {L} _ {\\mathrm {s}} [ i ] + \\mathcal {L} _ {\\mathrm {m}} [ i ]\\right). \\tag {9}", + "image_path": "9643c941d7191e5074134fb8ba65a34e7fd8a595b77a0010a5d5e2ecbe43fad9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 511, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 511, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 511, + 482, + 584 + ], + "type": "text", + "content": "Since we only introduce one additional temporal correlator and an entirely parameter-free clustering strategy, we exclusively tune a single standard spatiotemporal Transformer block with 1.6M parameters and freeze the DINO encoder throughout the training process. This straightforward method not only maintains a minimal number of learnable parameters but also yields remarkable results in video object segmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 600, + 230, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 230, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 230, + 613 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 623, + 205, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 205, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 205, + 634 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "For training, we adopt the challenging video dataset YouTube-VIS-19 [69] for video instance segmentation. YouTube-VIS-19 consists of 2,883 high-resolution" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 354, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 354, + 91, + 447, + 102 + ], + "type": "text", + "content": "Betrayed by Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "YouTube videos and each video usually contains multiple object instances of distinct semantics. During the inference, we evaluate our method on unsupervised multiple object segmentation. We use the synthetic MOVi-E [19] dataset and real-world datasets like DAVIS-17-Unsupervised [46] and YouTube-VIS-19 [69], reporting Foreground Adjusted Rand Index (FG-ARI) and mIoU. For MOVi-E, we integrate its training set into our training due to significant distribution shifts. For the rest, we maintain solely training on the YouTube-VIS-19 set. Note that we do not present any mask annotation to the model during either the training or inference stages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 243, + 280, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 280, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 280, + 255 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "text", + "content": "In training, we sample " + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "text", + "content": " frames with temporal stride 4 as the input clip. Each frame is augmented with random crop and color jitter and resized to " + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "inline_equation", + "content": "192 \\times 384" + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "text", + "content": ". We adopt DINO pretrained ViT-S/8 [7] as the frame encoder, which is then followed by a single spatio-temporal Transformer Encoder block with 8 heads as temporal correlator " + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{st}}" + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "text", + "content": ". We set the number of sampled positive and negative pairs to " + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "inline_equation", + "content": "K_{p} = 10" + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "inline_equation", + "content": "K_{n} = 50" + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "text", + "content": " in default. For optimization, we adopt AdamW [40] with a learning rate " + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 265, + 482, + 408 + ], + "type": "text", + "content": ", and train the model for a total of 30k iterations with a batch size of 16. In the inference stage, our model can be flexibly applied to video sequences of arbitrary lengths. For inference, we set the distance threshold in hierarchical clustering to 1.0. We follow the standard evaluation protocol outlined in the DAVIS-17 challenge [46] to match the predicted masks with the ground truth." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 132, + 495, + 482, + 640 + ], + "blocks": [ + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "text", + "content": "Table 2: Quantitative results on multiple object video segmentation. For MOVi-E and YouTube-VIS-19 datasets, we report FG-ARI and mIOU. Besides, we report Region Similarity " + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "text", + "content": " and Contour Accuracy " + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "text", + "content": " on DAVIS-17-Unsupervised. " + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "text", + "content": " The original paper of TimeT [52] includes the IoU of the background category, we rerun the evaluation to exclude background here. " + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "text", + "content": " We use DINOv2 ViT-S/14 as our frozen frame encoder." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 495, + 482, + 640 + ], + "lines": [ + { + "bbox": [ + 132, + 495, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 482, + 640 + ], + "type": "table", + "html": "
ModelMOVi-EYTVIS-19DAVIS-17
FG-ARImIoUFG-ARImIoUFG-ARIJ&FJF
SAVi [29]42.816.011.112.7----
STEVE [55]50.626.620.020.9----
OCLR [65]--15.932.514.7-34.6-
VideoSAUR [74]73.935.639.529.1----
SOLV [1]80.8-29.145.332.2-30.2-
SMTC [48]--31.438.833.340.536.444.6
TimeT* [52]--37.940.435.540.035.844.2
BA (ours)83.440.244.350.140.143.939.248.6
BA† (ours)84.440.744.550.141.643.739.448.0
", + "image_path": "2dabd9ea438ac67f3350a874c40f922c4ff89c80db0b7387c6f57a1fedcf24fe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 328, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 328, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 328, + 128 + ], + "type": "text", + "content": "4.3 Comparison with State-of-the-art" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 370 + ], + "type": "text", + "content": "We evaluate our method on challenging multi-object segmentation, which aims to segment different objects in a temporally coherent manner. Among the compared methods, SAVi [29] and OCLR [65] employ optical flow as supervision to discriminate motions of different objects. STEVE [55], VideoSAUR [74], SOLV [1] and SMTC [48] develop variants of slot attention with temporal constraints to bind to distinct instances. TimeT [52] propagates dense DINO features across time to enhance temporal consistency and applies clustering on learned features to produce object segmentations. In contrast, our method directly uses spatiotemporal attention maps for object segmentation. As shown in Table 2, BA achieves state-of-the-art results on both synthetic (MOVi-E) and realistic video datasets (YTVIS-19 and DAVIS-17), indicating promising generalization ability. Specifically, comparing to the prior arts, we gain an approximately 5-point advantage over both VideoSAUR [74] on YTVIS-19 and SMTC [48] and TimeT [52] on DAVIS-17 in terms of FG-ARI. The superior results under our simple architecture reveal the feasibility of utilizing the spatio-temporal dependencies to mine object cues and derive reliable video object segmentation results. Besides, we also report the results using DINOv2 ViT-S/14 as the frame encoder. Despite the larger patch size, our method still achieves competitive performance. This demonstrates the robustness of our approach to pretraining models." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 392, + 482, + 632 + ], + "blocks": [ + { + "bbox": [ + 136, + 392, + 482, + 632 + ], + "lines": [ + { + "bbox": [ + 136, + 392, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 136, + 392, + 482, + 632 + ], + "type": "image", + "image_path": "90ec1e8bd5bfcaccaee153d5d1c1bd46daa30fa48bed3a6d416b24161f0c692b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 639, + 481, + 662 + ], + "lines": [ + { + "bbox": [ + 132, + 639, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 132, + 639, + 481, + 662 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison with DINO [7] and SMTC [48] on DAVIS-17-Unsupervised. Different color denotes different clusters." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "text", + "content": "Betrayed by Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "Qualitative Comparison. To visually demonstrate the effectiveness of our proposed method, we contrast our segmentation results with those of the DINO counterpart and prior arts SMTC [48] in Fig. 4. For the DINO baseline, we cluster the patches through the attention map " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ". As can be observed, our method is able to segment with significantly less noise. For instance, in the motocross-jump sequence, BA nearly outlines the entire motorbike and the rider, whereas DINO misses most of the bike's body and only identifies its tyre. This shows that inferring the masks by associating a collection of RGB features and constructing spatio-temporal correspondences greatly enhances the segmentation result. Additionally, with the benefit of temporal correspondence, our model can differentiate between distinct object instances with the same semantics. Consider the soapbox sequence displayed in the right column; our method distinguishes two different people - one running behind the soapbox and another sitting in the soapbox throughout the video. This level of distinction cannot be obtained from DINO's spatial attention map, further emphasizing the effectiveness of introducing rich spatio-temporal relations in video segmentation tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 324, + 238, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 238, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 238, + 337 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 490, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 490, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 490, + 378 + ], + "type": "text", + "content": "In the ablation study, we utilize DINO ViT-S/8 as the frozen backbone. We report FG-ARI and IoU on YouTube-VIS-2019 and " + }, + { + "bbox": [ + 130, + 342, + 490, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 130, + 342, + 490, + 378 + ], + "type": "text", + "content": " on DAVIS-17-Unsupervised for multi-object segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "content": "Training objective " + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "content": ". To understand the contribution of each component to the VOS performance, we individually evaluate each loss item as depicted in Table 3. Compared to the baseline without training, the inclusion of semantic consistency loss " + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "content": " with entropy-based weighting mechanism " + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "content": " (second row) elevates FG-ARI from 23.4 to 30.1 on YouTube-VIS-19. Furthermore, the motion consistency loss " + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "content": " (third row) yields even more substantial improvements, resulting in FG-ARI reaching 39.9. This aligns with our preliminary observations that dynamic motions play a more significant role than semantics in fine-grained object segmentation. Combining all components (fourth row) provides the best results, highlighting the synergy between the components. Additionally, removing the weighting function " + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "content": " (fifth row) slightly decreases the performance on both two benchmarks, affirming the value of entropy normalization to prioritize informative regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Clustering metrics and strategy. To further explore whether our clustering strategy optimally leverages the learned spatio-temporal properties, we select four types of clustering metrics for comparison. Specifically, we adopt both feature vectors and attention maps directly from DINO and those from our learned spatio-temporal Transformer block. Table 4 reveals three key observations. First, our learning paradigm consistently outperforms the off-the-shelf DINO in terms of both features and attention maps, improving all metrics by over 10 points. This strongly supports our motivation that establishing spatio-temporal correspondence is crucial for video object segmentation. Second, utilizing the intermediate attention maps directly for clustering proves to be more effective than employing the final features. We speculate this occurs because the attention maps encode" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "fine-grained structural dependencies between various positions, whereas feature vectors primarily capture high-level semantics. Third, the hierarchical clustering strategy is not mandatory for success. We also achieve similar performance by replacing it with K-means clustering, where we set the number of centroids to 8, indicating that BA distills general object-aware representations across time." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 284, + 297, + 379 + ], + "blocks": [ + { + "bbox": [ + 130, + 196, + 293, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 196, + 293, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 293, + 274 + ], + "type": "text", + "content": "Table 3: Ablation on different combinations of semantic consistency loss " + }, + { + "bbox": [ + 130, + 196, + 293, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 196, + 293, + 274 + ], + "type": "text", + "content": ", motion consistency loss " + }, + { + "bbox": [ + 130, + 196, + 293, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 196, + 293, + 274 + ], + "type": "text", + "content": " and entropy weight " + }, + { + "bbox": [ + 130, + 196, + 293, + 274 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 196, + 293, + 274 + ], + "type": "text", + "content": ". The first row represents the results obtained from a randomly initialized spatio-temporal Transformer block." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 284, + 297, + 379 + ], + "lines": [ + { + "bbox": [ + 132, + 284, + 297, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 284, + 297, + 379 + ], + "type": "table", + "html": "
LsmYTVIS-19DAVIS-17
wFG-ARImIoUJ&FFG-ARI
--23.425.523.118.4
✓X30.132.429.627.5
X✓39.946.138.637.3
✓✓44.350.143.940.1
✓✓43.148.142.538.1
", + "image_path": "bece968fb4c424ba6dae50edd949e48d49c43c1dfb87a44eb57e6b1e7ac560f2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 302, + 284, + 473, + 380 + ], + "blocks": [ + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "lines": [ + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "spans": [ + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "text", + "content": "Table 4: Ablation on different clustering metrics. Features " + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "text", + "content": " and attention maps " + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "text", + "content": " are directly obtained from DINO. " + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "inline_equation", + "content": "A_{v}" + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "text", + "content": " are generated from our learned spatiotemporal Transformer block. " + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "inline_equation", + "content": "A_{v}" + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "inline_equation", + "content": "A_{v}^{*}" + }, + { + "bbox": [ + 300, + 197, + 478, + 275 + ], + "type": "text", + "content": " respectively denote the hierarchical clustering and K-means clustering results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 302, + 284, + 473, + 380 + ], + "lines": [ + { + "bbox": [ + 302, + 284, + 473, + 380 + ], + "spans": [ + { + "bbox": [ + 302, + 284, + 473, + 380 + ], + "type": "table", + "html": "
Clustering MetricYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
F22.123.816.119.4
Fv33.931.433.432.1
A30.134.829.330.5
Av44.350.143.940.1
Av*42.548.541.339.3
", + "image_path": "f396aa383a64a12f0029f239e21a9da4233498c8fb80b5913bf2bd5f20916ffb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": "Number of frames " + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": " and temporal stride. To investigate whether a wider temporal receptive field can reinforce temporal coherence and enhance object segmentations, we experiment by varying the number of input frames and temporal stride in training. Our results, displayed in Table 5, reveal that the performance significantly deteriorates when the stride is very small, such as a temporal stride of 1 in the first row. However, it begins to plateau upon reaching 4. This is because a small temporal stride leads to very subtle temporal dynamics in consecutive frames, resulting in a trivial learning task while a larger temporal stride provides richer dynamics, effectively encouraging the model to capture temporal consistency and maintain object permanence. Moreover, adding more frames results in diminishing returns, indicating that a modest temporal receptive field is adequate for sufficient temporal context in video object segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Motion representations. In training, we use the spatio-temporal attention maps as latent motion representations for dynamic motion alignment " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". There are some alternatives to formulate this motion representation. Here we explore two variants in Table 6. The one is simple global feature correlation, which directly calculates the dot product between spatio-temporal feature vectors, i.e., " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_g = \\tilde{F}\\tilde{F}^\\top \\in \\mathbb{R}^{THW\\times THW}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". The other is local feature correlation, where we sample a sliding local window with size " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{p}\\times H_{p}\\times W_{p}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " for each grid, and obtain the local correlation matrix " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_l\\in \\mathbb{R}^{THW\\times T_pH_pW_p}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". It is obvious that using the self-attention maps as motion representation substantially outperforms the performance of the other two methods. Compared to the global correlation, the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "text", + "content": "Betrayed by Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 178, + 294, + 264 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 293, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 293, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 293, + 170 + ], + "type": "text", + "content": "Table 5: Ablation on the number of frames " + }, + { + "bbox": [ + 132, + 114, + 293, + 170 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 114, + 293, + 170 + ], + "type": "text", + "content": " and temporal stride. When sampling more than 3 frames with a larger stride, the method exhibits a diminishing effect." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 178, + 294, + 264 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 294, + 264 + ], + "type": "table", + "html": "
TStrideYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
3137.443.637.133.6
3444.350.143.940.1
3844.250.243.340.7
5444.850.644.341.5
", + "image_path": "dc486cf859072a5bdef078e6e385b1d1da48831a03e4be1a787cb1b396bc777a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 301, + 190, + 486, + 264 + ], + "blocks": [ + { + "bbox": [ + 300, + 114, + 478, + 180 + ], + "lines": [ + { + "bbox": [ + 300, + 114, + 478, + 180 + ], + "spans": [ + { + "bbox": [ + 300, + 114, + 478, + 180 + ], + "type": "text", + "content": "Table 6: Ablation on different motion representations in calculating motion consistency loss " + }, + { + "bbox": [ + 300, + 114, + 478, + 180 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\mathrm{m}}" + }, + { + "bbox": [ + 300, + 114, + 478, + 180 + ], + "type": "text", + "content": " . We compare using the self-attention matrix,simple local and global feature correlation as the latent motion representation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 301, + 190, + 486, + 264 + ], + "lines": [ + { + "bbox": [ + 301, + 190, + 486, + 264 + ], + "spans": [ + { + "bbox": [ + 301, + 190, + 486, + 264 + ], + "type": "table", + "html": "
MotionRepYTVIS-19DAVIS-17
FG-ARImIoUJ&FFG-ARI
Local Corr.30.533.829.627.7
Global Corr.39.149.236.433.9
Self-Attention44.350.143.940.1
", + "image_path": "ec33a81abf97155f93105eaccd6926b3f97e58613bc34a3b7141e043d051407d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 287, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 348 + ], + "type": "text", + "content": "learnable parameters in attention layers enhance the modeling capacity, which enables the self-attention maps to capture more comprehensive spatio-temporal dependencies and serve as a better motion representation. The local correlation only captures temporal dynamics in limited reception fields, restricting the ability for long-term perception." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 365, + 220, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 220, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 220, + 378 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": "In this paper, we introduce a simple yet powerful approach for self-supervised video object segmentation. Specifically, our proposed architecture introduces only a single spatio-temporal Transformer block that ingests frame-wise DINO features and constructs the spatio-temporal correspondence through the attention layer. During training, we adopt a dual self-supervised consistency objective that encompasses both semantic consistency and dynamic motion consistency, supplemented with entropy normalization. In inference, we utilize hierarchical clustering on spatio-temporal attention maps to generate temporally coherent object segmentation masks. This extremely straightforward approach nevertheless delivers state-of-the-art results on challenging real-world multi-object segmentation tasks such as DAVIS-17-Unsupervised and YouTube-VIS-19." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Limitation. While promising, BA still has room for improvement. Firstly, the current spatio-temporal attention block cannot accommodate very long window-size frames due to the quadratic memory cost. Potentially, adopting memory mechanisms [49, 75] that are capable of handling long-duration data may prove more suitable. Secondly, neither pixel-level constraints nor spatio-temporal dependencies consider the granularity in the video. Incorporating multi-scale features [4] may present a promising direction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Social Negative Impact. Although our method produces coherent video object segmentation without annotation, it could also enhance deepfake video technologies, making it easier to produce false videos for malicious intent. It is important to regulate the use of such technology and ensure it is leveraged for beneficial purposes." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 61931023, Grant 61932022, Grant 62371288, Grant 62320106003, Grant 62301299, Grant T2122024, Grant 62120106007, the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 263, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 263, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 297 + ], + "type": "text", + "content": "1. Aydemir, G., Xie, W., Guney, F.: Self-supervised object-centric learning for videos. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=919tWtJPXe" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "text", + "content": "2. Bao, L., Wu, B., Liu, W.: Cnn in mrf: Video object segmentation via inference in a cnn-based higher-order spatio-temporal mrf. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5977-5986 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "type": "text", + "content": "3. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 354, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 387 + ], + "type": "text", + "content": "4. Bian, Z., Jabri, A., Efros, A.A., Owens, A.: Learning pixel trajectories with multiscale contrastive random walks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6508-6519 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 388, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 419 + ], + "type": "text", + "content": "5. Bielski, A., Favaro, P.: Move: Unsupervised movable object segmentation and detection. Advances in Neural Information Processing Systems 35, 33371-33386 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "text", + "content": "6. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "type": "text", + "content": "7. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "text", + "content": "8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "text", + "content": "9. Chen, Y., Liu, Y., Jiang, D., Zhang, X., Dai, W., Xiong, H., Tian, Q.: Sdae: Self-distillated masked autoencoder. In: European conference on computer vision. pp. 108-124. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "text", + "content": "0. Choudhury, S., Karazija, L., Laina, I., Vedaldi, A., Rupprecht, C.: Guess What Moves: Unsupervised Video and Image Segmentation by Anticipating Motion. In: British Machine Vision Conference (BMVC) (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 480, + 632 + ], + "type": "text", + "content": "11. Ding, S., Li, M., Yang, T., Qian, R., Xu, H., Chen, Q., Wang, J., Xiong, H.: Motion-aware contrastive video representation learning via foreground-background merging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9716-9726 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "12. Ding, S., Qian, R., Xiong, H.: Dual contrastive learning for spatio-temporal representation. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5649-5658 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "text", + "content": "Betrayed by Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "13. Ding, S., Xie, W., Chen, Y., Qian, R., Zhang, X., Xiong, H., Tian, Q.: Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "text", + "content": "14. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 251 + ], + "type": "text", + "content": "15. Dutt Jain, S., Xiong, B., Grauman, K.: Fusionseg: Learning to combine motion and appearance for fully automatic segmentation of generic objects in videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3664-3673 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "text", + "content": "16. Elsayed, G., Mahendran, A., van Steenkiste, S., Greff, K., Mozer, M.C., Kipf, T.: Savi++: Towards end-to-end object-centric learning from real-world videos. Advances in Neural Information Processing Systems 35, 28940-28954 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "17. Fan, D.P., Wang, W., Cheng, M.M., Shen, J.: Shifting more attention to video salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8554-8564 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 318, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 363 + ], + "type": "text", + "content": "18. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 363, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 430 + ], + "type": "text", + "content": "19. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., Kipf, T., Kundu, A., Lagun, D., Laradji, I., Liu, H.T.D., Meyer, H., Miao, Y., Nowrouzezahrai, D., Oztireli, C., Pot, E., Radwan, N., Rebain, D., Sabour, S., Sajjadi, M.S.M., Sela, M., Sitzmann, V., Stone, A., Sun, D., Vora, S., Wang, Z., Wu, T., Yi, K.M., Zhong, F., Tagliasacchi, A.: Kubric: a scalable dataset generator (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 430, + 481, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 462 + ], + "type": "text", + "content": "20. Guo, Y., Yang, C., Rao, A., Agrawala, M., Lin, D., Dai, B.: Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 463, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 481, + 508 + ], + "type": "text", + "content": "21. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=Fx2SbBgcte" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 508, + 481, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 481, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 481, + 551 + ], + "type": "text", + "content": "22. Hamilton, M., Zhang, Z., Hariharan, B., Snavely, N., Freeman, W.T.: Unsupervised semantic segmentation by distilling feature correspondences. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=SaK06z6H10c" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 552, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 481, + 586 + ], + "type": "text", + "content": "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "type": "text", + "content": "24. Hu, Y., Wang, R., Zhang, K., Gao, Y.: Semantic-aware fine-grained correspondence. In: European Conference on Computer Vision. pp. 97-115. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "text", + "content": "25. Hu, Y.T., Huang, J.B., Schwing, A.G.: Videomatch: Matching based video object segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 54-70 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "26. Jabri, A., Owens, A., Efros, A.: Space-time correspondence as a contrastive random walk. Advances in neural information processing systems 33, 19545-19560 (2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "27. Johnander, J., Danelljan, M., Brissman, E., Khan, F.S., Felsberg, M.: A generative appearance model for end-to-end video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8953-8962 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 183 + ], + "type": "text", + "content": "28. Johnson, S.C.: Hierarchical clustering schemes. Psychometrika 32(3), 241-254 (1967)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 226 + ], + "type": "text", + "content": "29. Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jonschkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=aD7uesX1GF_" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "30. Lai, Z., Xie, W.: Self-supervised learning for video correspondence flow. In: BMVC (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "31. Lai, Z., Lu, E., Xie, W.: Mast: A memory-augmented self-supervised tracker. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6479-6488 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "text", + "content": "32. Li, L., Wang, W., Zhou, T., Li, J., Yang, Y.: Unified mask embedding and correspondence learning for self-supervised video segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18706-18716 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "33. Li, X., Loy, C.C.: Video object segmentation with joint re-identification and attention-aware mask propagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 90-105 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 423 + ], + "type": "text", + "content": "34. Li, X., Liu, S., De Mello, S., Wang, X., Kautz, J., Yang, M.H.: Joint-task self-supervised learning for temporal correspondence. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32. Curran Associates, Inc. (2019), https://proceedings.neurips.cc/paper/2019/file/140f6969d5213fd0ece03148e62e461e-Paper.pdf" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "35. Li, X., Liu, S., De Mello, S., Wang, X., Kautz, J., Yang, M.H.: Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "36. Lian, L., Wu, Z., Yu, S.X.: Bootstrapping objectness from videos by relaxed common fate and visual grouping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14582-14591 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "type": "text", + "content": "37. Liu, R., Wu, Z., Yu, S., Lin, S.: The emergence of objectness: Learning zero-shot segmentation from videos. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 13137-13152. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper/2021/file/6d9cb7de5e8ac30bd5e8734bc96a35c1-Paper.pdf" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "38. Liu, Y., Wang, Y., Chen, Y., Dai, W., Li, C., Zou, J., Xiong, H.: Promoting semantic connectivity: Dual nearest neighbors contrastive learning for unsupervised domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3510-3519 (June 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "39. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "40. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "41. Melas-Kyriazi, L., Rupprecht, C., Laina, I., Vedaldi, A.: Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and local-" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "text", + "content": "Betrayed by Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "ization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8364-8375 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "text", + "content": "42. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "text", + "content": "43. Patil, P.W., Dudhane, A., Kulkarni, A., Murala, S., Gonde, A.B., Gupta, S.: An unified recurrent video object segmentation framework for various surveillance environments. IEEE Transactions on Image Processing 30, 7889-7902 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "44. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 278 + ], + "type": "text", + "content": "45. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "text", + "content": "46. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "text", + "content": "47. Qian, R., Ding, S., Liu, X., Lin, D.: Static and dynamic concepts for self-supervised video representation learning. In: European Conference on Computer Vision. pp. 145-164. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "text", + "content": "48. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "49. Qian, R., Dong, X., Zhang, P., Zang, Y., Ding, S., Lin, D., Wang, J.: Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "type": "text", + "content": "50. Qian, R., Li, Y., Liu, H., See, J., Ding, S., Liu, X., Li, D., Lin, W.: Enhancing self-supervised video representation learning via multi-level feature optimization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7990-8001 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "text", + "content": "51. Rambhatla, S.S., Misra, I., Chellappa, R., Shrivastava, A.: Most: Multiple object localization with self-supervised transformers for object discovery. arXiv preprint arXiv:2304.05387 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "type": "text", + "content": "52. Salehi, M., Gavves, E., Snoek, C.G., Asano, Y.M.: Time does tell: Self-supervised time-tuning of dense image representations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16536-16547 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "text", + "content": "53. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence 22(8), 888-905 (2000)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "text", + "content": "54. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. In: BMVC 2021-32nd British Machine Vision Conference (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "55. Singh, G., Wu, Y.F., Ahn, S.: Simple unsupervised object-centric learning for complex and naturalistic videos. Advances in Neural Information Processing Systems 35, 18181-18196 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "56. Tian, J., Aggarwal, L., Colaco, A., Kira, Z., Gonzalez-Franco, M.: Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv preprint arXiv:2308.12469 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "57. Vondrick, C., Shrivastava, A., Fathi, A., Guadarrama, S., Murphy, K.: Tracking emerges by colorizing videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 391-408 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "58. Wang, N., Song, Y., Ma, C., Zhou, W., Liu, W., Li, H.: Unsupervised deep tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1308-1317 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "text", + "content": "59. Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle-consistency of time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2566-2576 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 481, + 217 + ], + "type": "text", + "content": "60. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3124-3134 (June 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 481, + 261 + ], + "type": "text", + "content": "61. Wang, X., Misra, I., Zeng, Z., Girdhar, R., Darrell, T.: Videocutler: Surprisingly simple unsupervised video instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22755-22764 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 262, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 481, + 306 + ], + "type": "text", + "content": "62. Wang, Y., Shen, X., Yuan, Y., Du, Y., Li, M., Hu, S.X., Crowley, J.L., Vaufreydaz, D.: Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 307, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 481, + 350 + ], + "type": "text", + "content": "63. Wang, Y., Li, J., ZHANG, X., Shi, B., Li, C., Dai, W., Xiong, H., Tian, Q.: BarleRiA: An efficient tuning framework for referring image segmentation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=wHLDHRkmEu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 352, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 481, + 384 + ], + "type": "text", + "content": "64. Xie, C., Xiang, Y., Harchaoui, Z., Fox, D.: Object discovery in videos as foreground motion clustering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9994-10003 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 481, + 418 + ], + "type": "text", + "content": "65. Xie, J., Xie, W., Zisserman, A.: Segmenting moving objects via an object-centric layered representation. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 419, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 481, + 440 + ], + "type": "text", + "content": "66. Xu, H., Ding, S., Zhang, X., Xiong, H., Tian, Q.: Masked autoencoders are robust data augmentors. arXiv preprint arXiv:2206.04846 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 441, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 481, + 474 + ], + "type": "text", + "content": "67. Xu, J., Wang, X.: Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10075-10085 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 508 + ], + "type": "text", + "content": "68. Yang, C., Lamdouar, H., Lu, E., Zisserman, A., Xie, W.: Self-supervised video object segmentation by motion grouping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7177-7188 (October 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 509, + 448, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 448, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 448, + 520 + ], + "type": "text", + "content": "69. Yang, L., Fan, Y., Xu, N.: Video instance segmentation. In: ICCV (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 553 + ], + "type": "text", + "content": "70. Yang, Y., Lai, B., Soatto, S.: Dystab: Unsupervised object segmentation via dynamic-static bootstrapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2826-2836 (June 2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 481, + 586 + ], + "type": "text", + "content": "71. Yang, Y., Loquercio, A., Scaramuzza, D., Soatto, S.: Unsupervised moving object detection via contextual information separation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 587, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 481, + 620 + ], + "type": "text", + "content": "72. Ye, V., Li, Z., Tucker, R., Kanazawa, A., Snavely, N.: Deformable sprites for unsupervised video decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2657-2666 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "type": "text", + "content": "73. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=1_jFneF07YC" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 355, + 91, + 447, + 102 + ], + "type": "text", + "content": "Betrayed by Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "74. Zadaianchuk, A., Seitzer, M., Martius, G.: Object-centric learning for real-world videos by predicting temporal feature similarities. In: Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "75. Zhou, J., Pang, Z., Wang, Y.X.: Rmem: Restricted memory banks improve video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18602-18611 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "type": "text", + "content": "76. Ziegler, A., Asano, Y.M.: Self-supervised learning of object parts for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14502-14511 (2022)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "text", + "content": "S. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_content_list.json b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d9b311532a9ad7fc51eed543cf10b3046c73a95b --- /dev/null +++ b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_content_list.json @@ -0,0 +1,1823 @@ +[ + { + "type": "text", + "text": "Better Call SAL: Towards Learning to Segment Anything in Lidar", + "text_level": 1, + "bbox": [ + 282, + 140, + 722, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aljosa Osep1*", + "bbox": [ + 243, + 212, + 344, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tim Meinhardt $^{1*}$", + "bbox": [ + 362, + 213, + 490, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Francesco Ferroni", + "bbox": [ + 509, + 213, + 643, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neehar Peri²", + "bbox": [ + 663, + 213, + 759, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deva Ramanan", + "bbox": [ + 362, + 229, + 483, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Laura Leal-Taixe1", + "bbox": [ + 501, + 229, + 632, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Equal contribution", + "bbox": [ + 434, + 250, + 568, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NVIDIA $^{2}$ CMU", + "bbox": [ + 428, + 273, + 573, + 287 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ff034901fcb4faedefe67f0f01f1518971ef5a39c4d50b3d34aa76493c5910c.jpg", + "image_caption": [ + "Class-agnostic instance segmentation" + ], + "image_footnote": [], + "bbox": [ + 223, + 333, + 356, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6bc6d52bfa739a8ef74b704c25dc5ba34fdd0ae6ed9335ab53532e9a9985f403.jpg", + "image_caption": [ + "Text prompts: {car, building,...}" + ], + "image_footnote": [], + "bbox": [ + 362, + 333, + 496, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20cf5d720d00f4ba55a989460329df6254c01ef76753b7495edf7d090aeb0969.jpg", + "image_caption": [ + "Text prompts: {trash bin}", + "Fig. 1: The SAL (Segment Anything in Lidar) model performs class-agnostic instance segmentation (i) and zero-shot classification via text prompting. This allows us to not only predict panoptic segmentation (ii) for fixed class vocabularies but segment any object (iii and iv) in a given Lidar scan." + ], + "image_footnote": [], + "bbox": [ + 503, + 333, + 637, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1512bc5e8da8666c27fc3de232e224f6db5e4c3d742bdc5c88a4d2855cbbd9db.jpg", + "image_caption": [ + "Text prompts: {streetcar}" + ], + "image_footnote": [], + "bbox": [ + 645, + 333, + 779, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We propose the SAL (Segment Anything in Lidar) method consisting of a text-promptable zero-shot model for segmenting and classifying any object in Lidar, and a pseudo-labeling engine that facilitates model training without manual supervision. While the established paradigm for Lidar Panoptic Segmentation (LPS) relies on manual supervision for a handful of object classes defined a priori, we utilize 2D vision foundation models to generate 3D supervision \"for free\". Our pseudolabels consist of instance masks and corresponding CLIP tokens, which we lift to Lidar using calibrated multi-modal data. By training our model on these labels, we distill the 2D foundation models into our Lidar SAL model. Even without manual labels, our model achieves $91\\%$ in terms of class-agnostic segmentation and $54\\%$ in terms of zero-shot LPS of the fully supervised state-of-the-art. Furthermore, we outperform several baselines that do not distill but only lift image features to 3D. More importantly, we demonstrate that SAL supports arbitrary class prompts, can be easily extended to new datasets, and shows significant potential to improve with increasing amounts of self-labeled data. We release all models and the code.", + "bbox": [ + 259, + 590, + 743, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We tackle segmentation and recognition of objects in Lidar point clouds, a task commonly tackled via Lidar Panoptic Segmentation (LPS).", + "bbox": [ + 215, + 189, + 784, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Status quo. LPS has been gaining significant attention in the community due to its role in scene understanding, which is vital for safe autonomous navigation. Thanks to the availability of labeled datasets [6, 18] and advances in learning representations from unordered point sets [12, 63, 64, 77], the Lidar community made significant progress in learning to segment and classify instances of predefined and manually-labeled classes. While this progress has been impressive, existing models [2, 26, 34, 43, 68, 72, 96] cannot adapt to continually evolving class ontologies [38] that may even vary in different geographic regions [80].", + "bbox": [ + 215, + 227, + 785, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stirring the pot. We challenge this well-established approach and investigate how we can train general LPS models that can be prompted to segment point clouds according to any object class vocabulary. Towards such a promptable Lidar segmentation approach, we first need to be able to segment any object, a capability that remains elusive in the Lidar domain.", + "bbox": [ + 215, + 356, + 785, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards segmenting anything in Lidar. To address these challenges, we propose SAL, which consists of a text-promptable zero-shot model (Fig. 2, left) for panoptic segmentation of arbitrary objects (Fig. 1), and pseudo-labeling engine (Fig. 2, right) that facilitates model training directly from raw sensory data. The SAL pseudo-label engine automatically labels Lidar sequences using image segmentation [30] and vision-language models [65]. We utilize SAM [30] to generate class-agnostic masks in images, and CLIP [65] to generate per-mask tokens that connect visual features to language, and finally, transfer both to Lidar using a calibrated sensory setup. Even though generated pseudo-labels only partially cover Lidar scans and are inherently noisy due to errors in the image-level generation process and imperfect sensory calibration, we demonstrate their effectiveness as self-supervised training data for the SAL zero-shot model. In contrast to prior work in LPS, SAL is the first model that can be prompted with any new vocabulary without re-training. Different from recent advances [56] that lift image features to 3D without distillation into a model, SAL does not require image features during inference and segments full $360^{\\circ}$ point clouds.", + "bbox": [ + 215, + 439, + 785, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Talk to your SAL. Our model predicts Lidar segmentation masks and their corresponding CLIP tokens, which allow us to perform zero-shot classification of segmented objects using arbitrary text prompts (Fig. 1). We evaluate SAL by prompting our zero-shot model on standard benchmarks for LPS. For class-agnostic segmentation, we reach $91\\%$ of the performance of manually supervised baselines. By classifying segmented objects in a zero-shot manner, we report the first (and very encouraging) results for zero-shot LPS and reach $42\\%$ and $54\\%$ of the supervised baselines trained on SemanticKITTI and nuScenes, respectively. Beyond that, as shown in Fig. 1, we can prompt SAL model to segment objects outside of existing Lidar dataset class vocabularies.", + "bbox": [ + 215, + 688, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5aeb54162997a1b95043f6c9a6f3d87021c2e9daf025cad3f5bfd29f5191cfa3.jpg", + "image_caption": [ + "Fig. 2: SAL overview: Given a Lidar scan and a class vocabulary prompt, specified as a list of per-class free-form text descriptions (left), SAL segments and classifiers objects (things and stuff classes). As labeled data for training such a model does not exist, we supervise SAL by distilling off-the-shelf vision foundation models to Lidar (right)." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 434, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3cc983aafba28d114fb4f567e6d9819b781280699abf68937ee548c90c2382f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 146, + 779, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contributions. As our main contribution, we re-think the established approach to Lidar Panoptic Segmentation and (i) present SAL for segmentation and classification of any object in a Lidar scan. As we do not utilize labeled Lidar data, we propose (ii) a pseudo-label engine that distills vision foundation models to the Lidar domain. The resulting pseudo-labels allow us to train (iii) a zero-shot LPS model without any human supervision. We demonstrate (iv) encouraging results on standard LPS benchmarks and outline a clear path towards universal, promptable segmentation foundation models for Lidar data.", + "bbox": [ + 212, + 340, + 784, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 482, + 385, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lidar has played a pivotal role since the dawn of embodied navigation [60,78,79]. Recent data-driven efforts in Lidar perception have been pushing boundaries in semantic segmentation [3,12,35,47,69,75,82,83,85,98], object detection [32,40, 42,57-59,88,90,95], and tracking [25,49,76].", + "bbox": [ + 212, + 512, + 782, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lidar panoptic segmentation. Recently, Lidar Panoptic Segmentation [4,6, 7,18] has emerged as a holistic approach to Lidar-based dynamic scene understanding. Prior works [2,4,19,26,28,31,34,43-45,68,89,96,97] learn to group and classify points specified in the training data according to the target class vocabularies, while methods for open-set instance [24,49,76] and panoptic segmentation [81] rely on bottom-up grouping based on Euclidean distance between points. While these developments have been impressive, end-to-end methods remain limited to specific target classes that appear in the training data, while bottom-up, clustering-based approaches [24,27,49,76,81] are sensitive to a particular choice of clustering parameters and are unable to improve their performance in a data-driven fashion. In contrast, the SAL zero-shot model not only learns to segment objects but is also capable of zero-shot classification.", + "bbox": [ + 212, + 578, + 784, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised learning for Lidar. As labeled data in the Lidar domain is scarce, several works investigate how contrastive learning, proven effective in the image domain [10, 23], can reduce the need for labeled data. SegContrast [52] learns to align representations of point clouds and their augmented views, utilizing density-based clustering (DBSCAN [17]) to pool and contrast", + "bbox": [ + 212, + 763, + 784, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Segment Anything in Lidar", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "features. Rather than using DBSCAN, [39, 70] leverage image-based methods [1] and foundation models [30] to perform contrastive alignment between image and corresponding Lidar features. The aforementioned methods only provide pretraining recipes for potential segmentation downstream tasks. Our SAL method goes a step further by training a full zero-shot Lidar Panoptic Segmentation model that could benefit from such pre-training recipes.", + "bbox": [ + 212, + 145, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-supervised (3D) object detection. Several methods (self-)supervise object detectors with (RGB-D) videos [22, 53-55, 61, 62], or Lidar sequences [50, 71, 92] to discover objects in sensory streams. However, such motion-based supervision provides no semantic information. Therefore, [51] distills per-point (PCA quantized) CLIP [65] features to Lidar for zero-shot classification of detected objects. Significantly different from [51], we tackle the more general and challenging task of LPS, which segments and classifies both moving and stationary things and stuff classes. To this end, we distill CLIP features not per point but per object instance. This provides a more holistic object-centric semantic distillation and allows us to supervise our model with non-quantized CLIP features.", + "bbox": [ + 212, + 242, + 787, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zero-shot recognition. Zero-shot recognition (ZSR) [84] methods (recently also referred to as \"open-vocabulary recognition\") tackle the recognition of distinct, yet related, semantic concepts that are either not labeled in the training data (transductive) or not observed at all (inductive). Such methods assume a dataset with labels for some classes, whereas unseen instances are supervised via attributes or class names. The latter is often used in conjunction with word embeddings [46]. By learning to align image-language features, ZSR methods can infer class labels for unseen objects [65]. ZSR has also been explored in the context of object detection, semantic segmentation, and panoptic segmentation [16,86] using word embedding models [5,8,48,66], or CLIP vision-language embeddings [20,21,33,36,67,87,91,93,94]. In the context of semantic segmentation in RGB-D and Lidar data, [41,56] augment image and language embeddings with features extracted from point clouds. Both rely on lifting dense image features from images to 3D during inference, which limits their applicability to setups with dense camera coverage. Similarly, [74] segments instances in accumulated RGB-D point clouds (ScanNet [14]) and relies on manual supervision in conjunction with image-based dense CLIP features for zero-shot classification. Moreover, datasets such as ScanNet [14] were recorded with RGB-D sensors in static environments and fused into 3D reconstructions based on scans, taken from multiple viewpoints. In contrast, SAL directly applies to full Lidar point clouds while requiring no image features at the inference time.", + "bbox": [ + 212, + 398, + 787, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 SAL: Segment Anything in Lidar", + "text_level": 1, + "bbox": [ + 214, + 743, + 571, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section outlines the key challenges and components towards a Lidar model that segments and classifies any object. We first define the underlying problem formation in Sec. 3.1 and then present our (Segmenting Anything in Lidar) SAL model in Sec. 3.2.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2c58dbd436e1694c79a318974d52cd36f7a1e35f1f775233fbc35fde74144a5f.jpg", + "image_caption": [ + "(a) Pseudo-label engine" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 433, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f37f165dfcb9f08cd9cb6920825715429cf2640b099b3c8d3b7ff2958784bc1a.jpg", + "image_caption": [ + "(b) Zero-shot model", + "Fig. 3: Our pseudo-label engine (Fig. 3a) utilizes SAM [30] to estimate segmentation masks in images, MaskCLIP [16] to estimate corresponding per-mask CLIP features, and a calibrated sensory setup to transfer them to the Lidar domain. We distill these pseudo-labels to our zero-shot model (Fig. 3b), which segments and classifies Lidar point clouds. The SAL zero-shot model employs a sparse-convolutional backbone [12], followed by a Transformer decoder that predicts objectness scores, segmentation masks, and CLIP tokens for each query. To (optionally) perform zero-shot classification, we forward the dataset class vocabulary through the CLIP text encoder and match the encoded vocabulary with predicted CLIP tokens. Our model requires no retraining for different vocabularies and no image features at inference time." + ], + "image_footnote": [], + "bbox": [ + 441, + 143, + 781, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Zero-Shot Lidar Panoptic Segmentation", + "text_level": 1, + "bbox": [ + 214, + 416, + 589, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The task. We discuss and evaluate SAL in the context of Lidar Panoptic Segmentation (LPS), which tackles both instance segmentation and object recognition. LPS methods take as input an unordered point cloud $P \\in \\mathbb{R}^{N \\times 4}$ encoding spatial coordinates and their sensor intensity. The currently established problem setting [7,18] assumes supervision and task outputs as per-point semantic class and instance identity labels. The class labels are confined to a pre-defined and fixed class vocabulary $\\mathcal{V}$ . In contrast, we tackle LPS in a generalized, zero-shot setting, where the class vocabulary is only provided at inference, not training time. Such a vocabulary $\\mathcal{V}$ specifies target classes as a list of $C$ class prompts. Each prompt $c_{i} \\in \\mathcal{V}$ is specified via free-form text, e.g., with a class name and optional class description. To perform zero-shot LPS a model must be designed to segment and classify any object in a Lidar scan.", + "bbox": [ + 212, + 444, + 787, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The challenge. Where do we get the data to train such a model? Image-based models [16,21,30,91,93] rely on large and diverse datasets for pre-training [15], dense prediction [13,37], and foundation models that align images with textual descriptions [65] - a commodity not available in the Lidar domain.", + "bbox": [ + 212, + 628, + 787, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 SAL Overview", + "text_level": 1, + "bbox": [ + 214, + 710, + 372, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our 2D-to-Lidar distillation method consists of two core components: (i) The pseudo-label engine transfers 2D vision foundation models into Lidar pseudolabels using multi-modal inputs from a calibrated sensory setup, shown in Fig. 3a. (ii) Our zero-shot model is trained on the generated pseudo-labels and is able to perform class-agnostic segmentation and zero-shot classification via text prompts. Given a semantic dataset vocabulary, this allows us to tackle zero-shot LPS as illustrated in Fig. 3b.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Segment Anything in Lidar", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 SAL Pseudo-Label Engine", + "text_level": 1, + "bbox": [ + 215, + 146, + 468, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The SAL pseudo-label engine relies on a calibrated multi-modal sensory setup with $k \\geq 1$ RGB cameras. Furthermore, a sufficient overlap between the Lidar sensing area and each camera view is paramount.", + "bbox": [ + 212, + 172, + 784, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mask generation: To pseudo-label object instances, we utilize the Segment Anything (SAM) model [30], which generates an overlapping set of segmentation masks. We flatten SAM's output mask hierarchy by non-maxima suppression (NMS) with a minimal overlap threshold to ensure mutually exclusive masks, suppressing object parts and subparts in favor of objects. This way, we obtain a set of non-overlapping binary masks $m_{i}^{k} \\in \\{0,1\\}^{W \\times H}$ for each camera view $k$ with image plane of size $W \\times H$ .", + "bbox": [ + 212, + 223, + 785, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP image token generation: As shown by CLIP [65], aligning image and text features allows vision-language foundation models to perform zero-shot image classification based on arbitrary text prompts. To transfer this capability to the Lidar domain, we generate a localized CLIP image feature token $f_{i}^{k} \\in \\mathbb{R}^{C_{t}}$ for each binary SAM mask $m_{i}^{k}$ . To obtain image tokens for a masked region of the input image, we utilize [16] and their relative mask attention in the CLIP image encoder feature space. Note that we never classify segments in the image domain but merely distill CLIP image feature tokens to Lidar to facilitate zero-shot classification during inference where we do not use any image features.", + "bbox": [ + 212, + 334, + 787, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image-to-Lidar unprojection: From the Lidar perspective, we unproject each image mask $m_i^k \\in \\{0,1\\}^{W \\times H}$ to a binary Lidar segmentation mask $\\tilde{m}_i \\in \\{0,1\\}^N$ by transforming the respective camera coordinate frame to the Lidar space. For datasets with multiple cameras, such as nuScenes [18], we process each image independently, followed by a cross-camera fusion of masks with a sufficient IoU overlap and averaging of their CLIP image tokens. The unprojection yields pairs $\\{\\tilde{m}_i^k, f_i^k\\}$ of Lidar masks and their corresponding CLIP features.", + "bbox": [ + 212, + 476, + 787, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Refinement via density-based clustering: Image-to-Lidar transformation (Fig. 4a→Fig. 4b) is inherently noisy due to imperfect calibration and issues with synchronization and Lidar rolling shutter. We improve our pseudo-label quality by creating an ensemble of DBSCAN [17] clusters $\\tilde{m}_l \\in \\{0,1\\}^N$ , obtained by varying the density threshold to compensate for varying density in Lidar point clouds (details in the appendix). We replace each $\\tilde{m}_i$ with its best-matching $\\tilde{m}_l$ in case their IoU exceeds a minimal overlap threshold and retain the original mask otherwise to obtain a refined set of pseudo-labels (Fig. 4c) that retain their original cardinality and associated CLIP features.", + "bbox": [ + 212, + 587, + 787, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 SAL Zero-Shot Model", + "text_level": 1, + "bbox": [ + 215, + 747, + 434, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The universal SAL zero-shot model deconstructs LPS into class-agnostic segmentation and zero-shot classification via text prompts.", + "bbox": [ + 212, + 773, + 784, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Universal architecture. Instead of relying on highly specialized and engineered LPS models, we base the SAL model on a universal Transformer decoder", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b54c146f43ff079276c0b5fa2bd9ea445b35766b2aaee6e933c85664775d83ce.jpg", + "image_caption": [ + "(a) SAM masks" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 401, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e738e9d95690a48a4678403483432bed70b632f7484385335f3b48b3778b380.jpg", + "image_caption": [ + "(b) SAM unprojected", + "Fig. 4: Refinement via clustering. After transferring image masks (Fig. 4a) to Lidar (Fig. 4b), we obtain pseudo-labels that suffer from sensory misalignment-related issues. Our geometric refinement (Fig. 4c) improves localization." + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 594, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e948e3f8612409e66dc5ff7741d878924f11a71b688833788f4fba3f7ee0688f.jpg", + "image_caption": [ + "(c) SAM + DBSCAN" + ], + "image_footnote": [], + "bbox": [ + 601, + 143, + 787, + 207 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "architecture (Fig. 3b), similar to [43]. Its 2D counterparts [9, 11] provide top-performing results across semantic, instance, and panoptic segmentation for images. With increasing amounts of (pseudo-)labeled data, Transformer decoders have the potential to achieve a similar impact in the Lidar domain. Following [43], we deploy a Minkowski U-Net [12] backbone for feature extraction followed by a Transformer decoder architecture with object query to point/voxel feature crossattention. Training our model to segment and classify any object and perform in a zero-shot evaluation setting requires a unique design of the final task heads.", + "bbox": [ + 212, + 281, + 784, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Class-agnostic segmentation. To localize segments in a point cloud, we rely on two model heads that predict segmentation masks and objectness scores for each query. The architecture of the former follows [43] and predicts binary output masks $\\hat{m}_i\\in \\{0,1\\} ^N$ by computing the dot product between queries and point features. The objectness head reduces the multi-class problem to a binary object or no-object decision.", + "bbox": [ + 212, + 407, + 784, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot classification. To equip our model with zero-shot classification capabilities without relying on any image input, we learn to predict CLIP [65] Lidar tokens, i.e., features in the CLIP space obtained from Lidar inputs, for each query. The token head consists of a three-layer MLP that directly regresses tokens $\\hat{f}_i^j$ . At inference, the predicted tokens are matched to text prompts via the CLIP text encoder. The matching yields a probability distribution over the prompts. To perform zero-shot LPS on a pre-defined dataset, we use its class vocabulary as input text prompts (Fig. 3b).", + "bbox": [ + 212, + 505, + 784, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train our model jointly for both class-agnostic segmentation and zero-shot classification with the following loss:", + "bbox": [ + 212, + 626, + 784, + 656 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S A L}} = \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {\\text {t o k e n}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 670, + 784, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with binary cross-entropy loss $\\mathcal{L}_{obj}$ and a cosine distance loss $\\mathcal{L}_{token}$ . The segmentation loss $\\mathcal{L}_{seg}$ follows [43] and consists of a binary mask cross-entropy and dice loss. The $\\mathcal{L}_{\\mathrm{SAL}}$ loss is supervised by pseudo-label pairs $\\{\\tilde{m}_i^k, f_i^k\\}$ obtained from our label engine and thereby distills both 2D foundation models (SAM and CLIP) into our LPS model.", + "bbox": [ + 212, + 698, + 784, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How to train with partial labels? Partially (pseudo-)labeled point clouds present a challenge for SAL model training. For example, in SemanticKITTI, only $14\\%$ of points are pseudo-labeled due to low camera coverage (see appendix). If we naively train SAL on partially (pseudo-)labeled point clouds, the objectness", + "bbox": [ + 214, + 779, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Segment Anything in Lidar", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "loss $\\mathcal{L}_{obj}$ would penalize any segmentation in these regions as a false positive, thereby teaching the model to ignore these points entirely. Empirically, we determined that the most effective training strategy is to (i) remove unlabeled points from point clouds, (ii) utilize standard data augmentations (rotations, flipping, scaling, and translations), in conjunction with proposed FrankenFruustum augmentation (detailed in the appendix) that mimics fully-labeled point clouds during training by randomly removing unlabeled points and replicating labeled frustum regions around the vertical axis. This augmentation does not increase the overall label coverage, but our ablations in Sec. 4.2 show its effectiveness in reducing the domain gap between training and inference input clouds.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 332, + 375, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the following section, we outline our experimental setup (Sec. 4.1) used to ablate the SAL model and its training strategies (Sec. 4.2). Since SAL presents the first Zero-Shot Lidar Panoptic Segmentation (ZS-LPS) model, we design and compare with multiple hand-crafted baselines that implement concepts from related works. We conclude our analysis by comparing our zero-shot performance (Sec. 4.4) to fully-supervised LPS methods.", + "bbox": [ + 212, + 359, + 787, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 470, + 428, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We evaluate SAL on two public Lidar Panoptic Segmentation (LPS) datasets, SemanticKITTI [6, 7] and nuScenes [18]. We utilize provided, human-annotated ground-truth (GT) panoptic segmentation labels solely for evaluation purposes, linear probing ablations, and baselines. We train SAL model only using our self-supervised pseudo-labels (Sec. 3.3). To demonstrate the general applicability of SAL to other domains and datasets, we show qualitative results of SAL trained on Waymo Open [73] dataset, which contains no GT LPS labels.", + "bbox": [ + 212, + 492, + 787, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics. For the evaluation, we utilize the standard Panoptic Segmentation [29] metrics. The Panoptic Quality $PQ = RQ \\times SQ$ combines Segmentation Quality (SQ) and Recognition Quality (RQ), i.e., F-1 score. True positives (TPs) are mask predictions with sufficient intersection-over-union (IoU) overlap with GT masks of the same class.", + "bbox": [ + 212, + 602, + 787, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic Oracle (SO). The aforementioned metrics assume GT labels and predictions with instance IDs and semantic classes for each point. Therefore, they are not suitable for evaluating class-agnostic segmentation. To assess the class-agnostic segmentation independently of semantics, we apply SO during the evaluation by assigning ground truth semantic classes $c_{i}$ to predicted masks $m_{i}$ via majority voting. Evaluation with SO is only needed for evaluation of class-agnostic segmentation - our zero-shot models provide both instance and semantic class predictions.", + "bbox": [ + 212, + 683, + 787, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stuff Merging (SM). SO allows us to assess class-agnostic instance segmentation for things classes. However, this approach is unsuitable for assessing", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the performance of stuff classes, as existing Lidar datasets merge instances of stuff classes into a single instance. The things- stuff separation can be unintuitive and inconsistent across datasets (e.g., SemanticKITTI merges instances of traffic signs into a single stuff class). In contrast, we fully embrace the philosophy that all classes can be segmented into individual instances (e.g., segmentation models should localize individual trees/bushes in the vegetation class or distinguish individual buildings). To evaluate our models on the target datasets in the SO regime on Lidar datasets that do not provide instance-level annotations for all classes, we additionally report results using a merge strategy (SM), which merges all instances of a particular stuff class into a single mask to ensure outputs of our models are consistent with the format of target datasets.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 SAL Ablations", + "text_level": 1, + "bbox": [ + 215, + 337, + 374, + 351 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we discuss how to train SAL zero-shot segmentation model using self-generated pseudo-labels (Tab. 1) and discuss design decisions behind our instance (Tab. 2) and semantic distillation (Tab. 3).", + "bbox": [ + 212, + 366, + 785, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Learning with partial labels. For the single-camera setup [6,7], our pseudolabels only cover $14\\%$ of the input point clouds (see appendix). Training on partial label coverage presents a challenge for our model. In Tab. 1, we only evaluate class-agnostic segmentation by applying SO and SM (see Sec. 4.1). We compare SAL models trained using pseudo and GT labels. The Frustum Filter removes all unlabeled points not visible in the camera frustum during training and/or evaluation. As shown in Tab. 1, removing unlabeled points from pseudo-labeled point clouds during training has a significant benefit on the performance ( $3^{rd}$ row, 59.3 PQ) as compared to the variant, where we merely ignore unlabeled region during the training ( $2^{nd}$ row, 22.2 PQ). This is likely due to data imbalance in a single-camera setup, leading to a class imbalance between labeled/unlabeled points, incentivizing SAL to suppress predictions.", + "bbox": [ + 212, + 416, + 787, + 598 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While utilizing standard augmentations (rotations, flipping, scaling, and translations) is crucial, we observe that performing our FrankenFrustum augmentation, which concatenates the visible portion of the point cloud around the z-axis (see Sec. 3.4), significantly improves results ( $4^{th}$ row, 62.5 PQ). Mixing point clouds from different scans ( $5^{th}$ row, 62.8 PQ) further boosts results. Overall, we obtain 69 PQ from GT supervision (1st row) and 62.8 PQ when using our pseudo-labels (5th row). Remarkably, our labels yield $91\\%$ of the GT performance while covering only $14\\%$ of the input point cloud (see appendix). To further contextualize these results, we note that this evaluation is performed only on classes for which GT instance labels are available. As shown in Fig. 1, our model learns to segment a much larger variety of classes. Evaluating only the subset of points visible in the camera ( $14\\%$ of all points, rows 6&7), the gap between the GT-supervised model (71.8 PQ) and pseudo-supervised model (70.7 PQ) shrinks further. We conclude that by training SAL model using pseudo-labels, we distill the notion of objectness from SAM [30] into our model and obtain segmentation capabilities, similar to the model trained on GT data.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Segment Anything in Lidar", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b24378f03d874df4f99196aeb219411d2749539c1c572431a9907bbf2480fba2.jpg", + "table_caption": [ + "Table 1: Class-agnostic segmentation. By cropping unlabeled points and performing data augmentations (in combination with our FrankenFrostum), SAL successfully learns to segment full point clouds even when only $14\\%$ of points are (pseudo)-labeled." + ], + "table_footnote": [], + "table_body": "
LabelsFrust.-Filter TrainEvalFranken Frust.PQSQ\\( PQ^{Th} \\)\\( PQ^{St} \\)
GT69.083.581.659.7
Pseu.22.267.545.15.5
Pseu.✓59.378.265.854.5
Pseu.✓✓62.579.167.658.8
Pseu.✓✓(mix)62.879.069.058.3
GT✓71.884.884.662.5
Pseu.✓✓✓(mix)70.781.975.467.3
", + "bbox": [ + 215, + 239, + 517, + 366 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2a939d7f05942b6f906506af9c6f7e2892a0b21a39d7446d8a5ec0379e98953b.jpg", + "table_caption": [ + "Table 2: Scaling. In contrast to predefined GT data, increasing the number of queries SAL improves performance on stuff classes. By increasing the amount of labeled data, we further improve segmentation performance." + ], + "table_footnote": [], + "table_body": "
TrainNum.PQSQ\\( PQ^{Th} \\)\\( PQ^{St} \\)
LabelsSetQuer.
GTtrain10067.882.479.759.2
GTtrain20067.983.679.859.3
GTtrain30069.083.581.659.7
Pseu.train10053.975.362.247.8
Pseu.train20060.877.066.756.5
Pseu.train30062.879.069.058.3
Pseu.bigtrain30065.379.571.960.5
", + "bbox": [ + 522, + 239, + 779, + 358 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Scaling queries and data. Analogous to the grid inference of SAM, which operates with hundreds of spatial queries (prompts), we analyze the effect of different numbers of decoder queries. We subsample instances during training to train models with fewer queries than the maximum number of segments per scan. As visualized in Tab. 2, increasing the number of decoder queries improves the recognition of things for both training with pseudo and GT labels. However, we observe a significant performance boost with pseudo labels for stuff classes $(58.3\\mathrm{PQ}^{St}, +10.8)$ , while the improvement for GT labels is marginal $(59.7\\mathrm{PQ}^{St}, +0.5)$ . Our model benefits from a larger number of queries for stuff classes as it learns a fine-grained segmentation model that is not limited to a prefixed set of things classes. In the last row, we concatenate pseudo-labeled train and test sets (neither used for the validation) into a bigtrain set and observe an improvement of $+2.5$ PQ, suggesting that scaling the amount of training data has the potential to improve the performance of our models further.", + "bbox": [ + 212, + 378, + 787, + 590 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c61cf55e92c0284e2b30610e115595330dda41318a1dadd0e2f5ddb7a3c1b21b.jpg", + "table_caption": [ + "Table 3: Semantic distillation. We linearly probe the SAL model, trained with and without semantic distillation loss $\\mathcal{L}_{token}$ . We train a linear classifier with GT labels while keeping backbone and decoder features frozen. As can be seen, $\\mathcal{L}_{token}$ successfully distills a notion of semantics into our model." + ], + "table_footnote": [], + "table_body": "
Linear prob. LtokenPQRQSQmIoU
×20.026.355.223.4
××33.141.968.340.0
×24.832.366.829.7
", + "bbox": [ + 215, + 739, + 454, + 801 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Semantic distillation. At inference, SAL predicts a binary segmentation mask and objectness score for each query. To perform zero-shot classification, our model additionally predicts CLIP image tokens. This allows us to not rely on image features during inference but prompt our predicted tokens with arbitrary texts, i.e., object classes. Does the token head loss $\\mathcal{L}_{token}$ successfully distill a notion of object semantics from CLIP into our model? To verify this, we perform linear probing experiments: we replace the token distillation head of a trained model with a linear classifier (backbone and instance de", + "bbox": [ + 467, + 598, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "coder remain frozen) and train it using GT labels. As can be seen in Tab. 3, distilling CLIP features into our model significantly improves linear probing", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/144cadf95b7313c8a13e48091d9cddd53459183980961368daf392f02b875354.jpg", + "table_caption": [ + "Table 4: Zero-shot panoptic segmentation. We utilize prior efforts in the image domain [30] and Lidar [56] domain to craft multiple baselines that only unproject segmentation masks and lift image features to Lidar. By contrast, SAL distills outputs of such baselines (pseudo-labels) into a stronger Lidar segmentation model. With ImageFeat, we denote methods that require image features at inference time, and Frust. Eval. denotes the evaluation of a subset of points visible in the camera." + ], + "table_footnote": [], + "table_body": "
MethodFrost. ImageDefault classesSuper classes
EvalFeat.PQSQPQThPQStmIoUPQSQPQThPQStmIoU
Class-agnostic Segmentation (Semantic Oracle)
SAM✓✓46.072.149.743.4------
SAM+Erosion✓✓42.269.445.639.6------
SAM+DBS (filter)✓✓46.770.376.824.8------
SAM+DBS (replace)✓✓48.773.753.145.4------
SAL✓X70.781.975.467.3------
Zero-Shot Lidar Panoptic Segmentation
SAM+DBS+CLIP✓✓27.571.531.724.530.651.177.571.241.054.3
SAL✓X33.171.422.840.533.563.984.288.351.766.4
SAM+DBS+CLIPXX8.256.418.60.67.511.547.60.017.311.2
SALXX24.866.817.430.228.748.578.880.432.652.8
", + "bbox": [ + 215, + 239, + 784, + 439 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "performance (33.1 PQ, 40 mIoU) compared to the baseline, not trained with the semantic distillation loss $\\mathcal{L}_{token}$ (20.0 PQ, 23.4 mIoU). This confirms that by distilling CLIP features, we inject a notion of semantics into our model, even though our model is not explicitly supervised with any labels containing semantic information - the notion of semantics is learned implicitly via CLIP feature distillation. Without this loss, our zero-shot model (24.8 PQ, 29.7 mIoU) outperforms the linearly probed model tuned using GT semantic labels.", + "bbox": [ + 212, + 449, + 787, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Class-Agnostic and Zero-Shot Lidar Panoptic Segmentation", + "text_level": 1, + "bbox": [ + 214, + 585, + 751, + 603 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As the first study tackling Zero-Shot Lidar Panoptic Segmentation (ZS-LPS), we devise several strong baselines inspired by prior works in RGB-D semantic segmentation. During inference, these methods utilize image-based models and lift image features to 3D [56]. To ensure these are evaluated fairly, we report results in Tab. 4 on the subset of points visible in at least one camera (Frost. Eval.). Our SAL model does not have this limitation. Therefore, we additionally report results evaluated on full point clouds in the bottom of Tab. 4.", + "bbox": [ + 212, + 621, + 787, + 728 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Class-agnostic segmentation. Our first baseline (SAM [30]) generates masks in images and unprojects them to Lidar (see Sec. 3.3), and leads to 46 PQ. We observe bleeding edges (see Fig. 4b) after unprojecting from the cameras to Lidar caused by imperfect calibrations. To mitigate these artifacts, we experiment with slightly eroding SAM predictions in the image domain. However, this leads to a performance drop (42.2 PQ), likely because smaller segments after erosion fail to pass $>0.5$ overlap with labeled instances.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Segment Anything in Lidar", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Density-based filtering: Alternatively, to mitigate unprojection errors, we experiment with density-based clustering methods. We generate a large pool of DBSCAN (DBS) clusters and test two strategies to improve our SAM-based segmentation labels: (i) filtering segments without a sufficient overlap with any DBSCAN cluster or (ii) replacing segments with large overlaps.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SAM+DBS (filter): Intuitively, the filter approach removed Lidar segments only segmentable in the image domain, e.g., a SAM mask that segments a shadow on a flat wall. While this strategy improves precision for things classes (76.8 PQTh, +27.1), it significantly degrades performance on stuff classes (24.8 PQSt -18.6) and overall degrades PQ to 46.7.", + "bbox": [ + 212, + 227, + 787, + 304 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SAM+DBS (replace): The replace strategy (that replaces SAM masks with DBSCAN clusters with sufficient mutual overlap), on the other hand, improves both things and stuff classes to an overall PQ of 48.7 (+3.9) and SQ 73.7 (+1.1), respectively. This improvement can be visually verified in Fig. 4c.", + "bbox": [ + 212, + 308, + 787, + 371 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SAL: We train our SAL model using top-performing $SAM+DBS$ (replace) and observe a significant improvement (70.7 PQ) in terms of both things and stuff classes. Even though SAL obtains significantly higher PQ than hand-crafted baselines, it also improves in terms of SQ (81.9 vs. 73.7 closest competitor), suggesting that the distilled model is robust to projection artifacts.", + "bbox": [ + 212, + 375, + 787, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Do we need a vision segmentation foundation model? To generate a set with a sufficiently high recall, we need to cluster points with varying DBSCAN radius parameters (details in the appendix). This leads to a large set of predicted masks per scan (5,413 avg.). While most of the correct segments are in this set, separating the signal from the noise is difficult. Prior works, such as [50], are limited to moving thing classes as they rely on motion cues to filter static point cloud regions and DBSCAN to group remaining points. Our core insight is that we can utilize vision foundation models that already learned a notion of objectness: SAM generates only 171 masks per image, reduced to 45 via flattening, and to 39 masks after the unprojection, achieving a high recall with only a few instances, or in other words, a high signal-to-noise ratio.", + "bbox": [ + 212, + 455, + 787, + 623 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Zero-shot segmentation. So far, we discussed class-agnostic segmentation. To obtain semantic class predictions per mask in a zero-shot fashion, we devise a simple baseline inspired by [56]. Once masks are unprojected from the image to Lidar space, we utilize associated CLIP features (directly extracted from images using [16]) to perform zero-shot classification via dot product between encoded class prompts and (lifted) CLIP features [65]. This baseline (SAM+DBS+CLIP) can be understood as an approach that directly lifts image features from state-of-the-art image-based models [16] to Lidar. We detail prompts in the appendix.", + "bbox": [ + 212, + 627, + 787, + 750 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "There is a significant performance drop between $SAM+DBS$ (replace) baseline, evaluated with Semantic Oracle ( $5^{th}$ row, 48.7 PQ), and zero-shot baseline ( $SAM+DBS$ (replace)+CLIP), $7^{th}$ row, 27.5 PQ. This is not surprising, as zero-shot segmentation is a challenging problem in both image [16] and Lidar domains. To this end, even when transferring features of state-of-the-art image-based models to Lidar we observe a significant performance drop. Remarkably,", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bc047352df77dc94781bfa399903d099fdfac172b540eceabd4ca2efe8b619f4.jpg", + "table_caption": [ + "Table 5: Lidar Panoptic Segmentation (LPS) on SemanticKITTI and nuScenes validation sets. We prompt our zero-shot SAL model with the respective class vocabularies and compare its performance to fully-supervised baselines. On SemanticKITTI and nuScenes, we reach $42\\%$ and $54\\%$ of the fully-supervised model, respectively. This gap reduces significantly when we evaluate super classes." + ], + "table_footnote": [], + "table_body": "
MethodSupervisionDefault classesSuper classes
PQRQSQ\\( PQ^{Th} \\)\\( PQ^{St} \\)mIoUPQRQSQmIoU
SemanticKITT1DS-Net [26]Full57.768.077.661.854.863.5---
PolarSeg [96]Full59.170.278.365.754.364.5---
EfficientLPS [72]Full59.269.875.058.060.964.9---
GP-S3Net [68]Full63.375.981.470.258.373.0---
MaskPLS [43]Full59.869.076.3---78.487.188.2
SALFull59.569.275.762.357.463.881.790.089.2
SALZero-shot24.832.366.817.430.228.748.559.478.8
mScenesPHNet [34]Full74.784.288.274.075.979.7---
DS-Net [26]Full51.259.086.138.472.373.5---
GP-S3Net [68]Full61.072.084.156.066.075.8---
EfficientLPS [72]Full62.073.983.456.870.665.6---
PolarSeg [96]Full63.475.383.959.270.466.9---
MaskPLS [43]Full57.766.071.864.452.262.571.581.086.2
SALFull70.580.885.979.461.772.874.282.787.1
SALZero-shot38.447.877.247.529.233.952.663.577.3
", + "bbox": [ + 246, + 226, + 751, + 439 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SAL distills such noisy image-based features to a stronger model ( $8^{th}$ row, 33.1 PQ, +5.6 PQ). Finally, to emphasize the inherent limitation of baselines, relying on image features, we evaluate on full $360^{\\circ}$ point clouds. SAL performs similarly in both settings. The baseline, on the other hand, relies on image features and is, therefore, unable to predict labels out of the camera frustum. This problem can be mitigated by utilizing a setup with denser camera coverage.", + "bbox": [ + 212, + 446, + 787, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Evaluation on super-classes. We observe that the CLIP classifier often confuses related classes, e.g., car and other vehicle. We additionally evaluate our semantics by prompting with super-classes (as defined in [6]). This alone increases PQ to 63.9 (48.5 when evaluating on full point clouds) and suggests significant potential for improving prompting and text-to-image alignment.", + "bbox": [ + 212, + 542, + 787, + 619 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Lidar Panoptic Segmentation Evaluation", + "text_level": 1, + "bbox": [ + 214, + 642, + 596, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Finally, we compare SAL with state-of-the-art LPS methods on standard benchmarks [6,18]. We focus this discussion on comparing SAL, trained with labeled data, and our SAL zero-shot model that specifies classes in benchmarks [6,18] as text prompts. Importantly, zero-shot results are obtained without training on any labeled data. In contrast, all baselines are trained using human-labeled data, and are constrained to their respective, pre-defined class vocabularies.", + "bbox": [ + 212, + 667, + 787, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SemanticKITTI. When evaluating with the default class vocabulary (Tab. 5), our zero-shot model reaches 24.8 PQ or $42\\%$ of the fully-supervised model (59.5 PQ). As discussed in Sec. 4.2, our method is mainly limited by the quality of the generated CLIP features. In practice, our zero-shot semantic classifier mainly confuses objects within super classes, e.g., car vs. other-vehicle. Therefore,", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Segment Anything in Lidar", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "without any retraining, we prompt our zero-shot model with super classes and obtain 48.5 PQ, $59\\%$ of the supervised model (81.7 PQ),", + "bbox": [ + 215, + 146, + 785, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "nuScenes. We report our result in Tab. 5 and reach similar conclusions. SAL reaches 38.4 PQ on default classes, $54\\%$ of the supervised model (70.5 PQ). As with SemanticKITTI, results consistently improve when evaluated with superclasses. In this setting, zero-shot model yields 52.6 PQ. Note that SAL model, supervised with labeled data, significantly outperforms MaskPLS.", + "bbox": [ + 215, + 181, + 785, + 257 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Towards closing the gap. Compared to SemanticKITTI (24.8 PQ, $42\\%$ ), we observe a significantly smaller gap between the model trained on pseudo- and ground truth labels for nuScenes (38.4 PQ, $54\\%$ ). Despite its sparse Lidar signal and more diverse input scenarios, SAL performs better on this challenging Lidar dataset. We attribute this to a larger pseudo-label coverage of nuScenes ( $48\\%$ , compared to $14\\%$ coverage on SemanticKITTI), and, therefore, effectively more labeled data, which is also consistent with our observations on SemanticKITTI w.r.t. increasing the amount of labeled data (Tab. 2). This trend suggests a clear path forward: applying SAL to pseudo-label more data with zero annotation cost.", + "bbox": [ + 215, + 261, + 785, + 398 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Beyond labeled datasets. Datasets, such as SemanticKITTI and nuScenes, provide dense, point-level panoptic segmentation labels, which we use for analysis and the evaluation of our method on a small set of classes that were labeled in the respective dataset (14/11 things/stuff in SemanticKITTI, respectively, and 23/6 in nuScenes). To demonstrate the versatility of SAL, we additionally pseudolabel and train a model on Waymo Open [73] dataset and show qualitative results in Fig. 1 and more in the appendix. This effort takes less than a week of pseudolabeling and model training efforts (we provide details in the appendix).", + "bbox": [ + 215, + 402, + 785, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The good, the (breaking) bad, and the ugly. While we are thrilled about class-agnostic segmentation performance, our experimental analysis reveals a substantial gap between full supervision and our zero-shot model. Finally, at the moment, SAL is limited to training on specific datasets. As in fully-supervised Lidar perception, cross-sensor generalization remains a challenge for SAL. This could be addressed by utilizing temporal context, data scaling efforts and investigating sensor-agnostic backbone networks.", + "bbox": [ + 215, + 527, + 785, + 633 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 657, + 366, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We proposed SAL, a method for Zero-Shot Lidar Panoptic Segmentation. The key SAL components are a pseudo-label engine that utilizes vision foundation models and a zero-shot model trained via self-supervision. While SAL is the first step in the direction of learning general Lidar segmentation models by distilling image-based foundation to Lidar, we believe we just scratched the surface and opened the door for training Lidar segmentation models without manual supervision.", + "bbox": [ + 215, + 688, + 785, + 779 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This project was funded, in parts, by ERC Starting Grant DynAI (ERC-101043189). We are grateful to Žan Gojčić, Guillem Braso, Cristiano Saltori, Sérgio Agostinho, and Jonas Schult for their feedback on the paper and their insightful comments. Special thanks to Maxim Maximov for his help on figures.", + "bbox": [ + 215, + 782, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell. 34(11), 2274-2282 (2012)", + "2. Agarwalla, A., Huang, X., Ziglar, J., Ferroni, F., Laura, L.T., Hays, J., Osep, A., Ramanan, D.: Lidar panoptic segmentation and tracking without bells and whistles. In: Int. Conf. Intel. Rob. Sys. (2023)", + "3. Aksoy, E.E., Baci, S., Cavidar, S.: Salsanet: Fast road and vehicle segmentation in lidar point clouds for autonomous driving. In: Intel. Veh. Symp. (2020)", + "4. Aygiun, M., Osep, A., Weber, M., Maximov, M., Stachniss, C., Behley, J., Leal-Taixe, L.: 4d panoptic lidar segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "5. Bansal, A., Sikka, K., Sharma, G., Chellappa, R., Divakaran, A.: Zero-shot object detection. In: Eur. Conf. Comput. Vis. (2018)", + "6. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In: Int. Conf. Comput. Vis. (2019)", + "7. Behley, J., Milioto, A., Stachniss, C.: A Benchmark for LiDAR-based Panoptic Segmentation based on KITTI. In: Int. Conf. Rob. Automat. (2021)", + "8. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. Adv. Neural Inform. Process. Syst. (2019)", + "9. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. (2020)", + "0. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Adv. Neural Inform. Process. Syst. (2020)", + "1. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "2. Choy, C., Gwak, J., Savarese, S.: 4D spatio-temporal convnets: Minkowski convolutional neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)", + "3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)", + "4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)", + "5. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. (2009)", + "6. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip. In: Int. Conf. Mach. Learn. (2023)", + "7. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Rob. Sci. Sys. (1996)", + "8. Fong, W.K., Mohan, R., Hurtado, J.V., Zhou, L., Caesar, H., Beijbom, O., Valada, A.: Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Rob. Automat. Letters (2021)", + "9. Gasperini, S., Mahani, M.A.N., Marcos-Ramiro, A., Navab, N., Tombari, F.: Panost-ter: End-to-end panoptic segmentation of lidar point clouds. IEEE Rob. Automat. Letters (2021)" + ], + "bbox": [ + 225, + 176, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Segment Anything in Lidar", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Eur. Conf. Comput. Vis. (2022)", + "21. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)", + "22. Harley, A.W., Zuo, Y., Wen, J., Mangal, A., Potdar, S., Chaudhry, R., Fragkiadaki, K.: Track, check, repeat: An em approach to unsupervised tracking. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "24. Held, D., Guillory, D., Rebsamen, B., Thrun, S., Savarese, S.: A probabilistic framework for real-time 3d segmentation using spatial, temporal, and semantic cues. In: Rob. Sci. Sys. (2016)", + "25. Held, D., Levinson, J., Thrun, S., Savarese, S.: Combining 3d shape, color, and motion for robust anytime tracking. In: Rob. Sci. Sys. (2014)", + "26. Hong, F., Zhou, H., Zhu, X., Li, H., Liu, Z.: Lidar-based panoptic segmentation via dynamic shifting network. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "27. Hu, P., Held, D., Ramanan, D.: Learning to optimally segment point clouds. IEEE Robotics and Automation Letters 5(2), 875-882 (2020)", + "28. Hurtado, J.V., Mohan, R., Valada, A.: Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189 (2020)", + "29. Kirillov, A., He, K., Girshick, R.B., Rother, C., Dollar, P.: Panoptic segmentation. IEEE Conf. Comput. Vis. Pattern Recog. (2018)", + "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Int. Conf. Comput. Vis. (2023)", + "31. Kreuzberg, L., Zulfikar, I.E., Mahadevan, S., Engelmann, F., Leibe, B.: 4d-stop: Panoptic segmentation of 4d lidar using spatio-temporal object proposal generation and aggregation. In: ECCV AVVision Workshop (2022)", + "32. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)", + "33. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: Int. Conf. Learn. Represent. (2022)", + "34. Li, J., He, X., Wen, Y., Gao, Y., Cheng, Y., Zhang, D.: Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "35. Li, S., Chen, X., Liu, Y., Dai, D., Stachniss, C., Gall, J.: Multi-scale interaction for real-time lidar data segmentation on an embedded platform. IEEE Rob. Automat. Letters 7(2), 738-745 (2021)", + "36. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "37. Lin, T., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. (2014)", + "38. Lin, Z., Pathak, D., Wang, Y.X., Ramanan, D., Kong, S.: Continual learning with evolving class ontologies. Adv. Neural Inform. Process. Syst. (2022)", + "39. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Liu, Z., Zhang, Z., Cao, Y., Hu, H., Tong, X.: Group-free 3d object detection via transformers. In: Int. Conf. Comput. Vis. (2021)", + "41. Lu, Y., Jiang, Q., Chen, R., Hou, Y., Zhu, X., Ma, Y.: See more and know more: Zero-shot point cloud segmentation via multi-modal visual data. In: Int. Conf. Comput. Vis. (2023)", + "42. Ma, Y., Peri, N., Wei, S., Hua, W., Ramanan, D., Li, Y., Kong, S.: Long-tailed 3d detection via 2d late fusion. arXiv preprint arXiv:2312.10986 (2023)", + "43. Marcuzzi, R., Nunes, L., Wiesmann, L., Behley, J., Stachniss, C.: Mask-based panoptic lidar segmentation for autonomous driving. IEEE Rob. Automat. Letters 8(2), 1141-1148 (2023)", + "44. Marcuzzi, R., Nunes, L., Wiesmann, L., Marks, E., Behley, J., Stachniss, C.: Mask4d: End-to-end mask-based 4d panoptic segmentation for lidar sequences. IEEE Rob. Automat. Letters (2023)", + "45. Marcuzzi, R., Nunes, L., Wiesmann, L., Vizzo, I., Behley, J., Stachniss, C.: Contrastive instance association for 4d panoptic segmentation using sequences of 3d lidar scans. IEEE Rob. Automat. Letters (2022)", + "46. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013)", + "47. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: RangeNet++: Fast and Accurate LiDAR Semantic Segmentation. In: Int. Conf. Intel. Rob. Sys. (2019)", + "48. Miller, D., Nicholson, L., Dayoub, F., Sünderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: Int. Conf. Rob. Automat. (2018)", + "49. Moosmann, F., Stiller, C.: Joint self-localization and tracking of generic objects in 3d range data. In: Int. Conf. Rob. Automat. (2013)", + "50. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: Eur. Conf. Comput. Vis. (2022)", + "51. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: Int. Conf. Comput. Vis. (2023)", + "52. Nunes, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Segcontrast: 3d point cloud feature representation learning through self-supervised segment discrimination. IEEE Rob. Automat. Letters 7(2), 2116-2123 (2022)", + "53. Osep, A., Voigtlaender, P., Luiten, J., Breuers, S., Leibe, B.: Towards large-scale video video object mining. In: ECCV Workshop on Interactive and Adaptive Learning in an Open World (2018)", + "54. Osep, A., Mehner, W., Voigtlaender, P., Leibe, B.: Track, then decide: Category-agnostic vision-based multi-object tracking. In: Int. Conf. Rob. Automat. (2018)", + "55. Osep, A., Voigtlaender, P., Luiten, J., Breuers, S., Leibe, B.: Large-scale object mining for object discovery from unlabeled video. In: Int. Conf. Rob. Automat. (2019)", + "56. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "57. Peri, N., Dave, A., Ramanan, D., Kong, S.: Towards long-tailed 3d detection. In: Conf. Rob. Learn. (2023)", + "58. Peri, N., Li, M., Wilson, B., Wang, Y.X., Hays, J., Ramanan, D.: An empirical analysis of range for 3d object detection. In: ICCV Workshops (2023)", + "59. Peri, N., Luiten, J., Li, M., Osep, A., Leal-Taixe, L., Ramanan, D.: Forecasting from lidar via future object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Segment Anything in Lidar", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Petrovskaya, A., Thrun, S.: Model based vehicle detection and tracking for autonomous urban driving. Aut. Rob. 26, 123-139 (2009)", + "61. Pot, E., Toshev, A., Kosecka, J.: Self-supervisory signals for object discovery and detection. arXiv preprint arXiv:1806.03370 (2018)", + "62. Prest, A., Leistner, C., Civera, J., Schmid, C., Ferrari, V.: Learning object class detectors from weakly annotated video. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012)", + "63. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)", + "64. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Adv. Neural Inform. Process. Syst. (2017)", + "65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021)", + "66. Rahman, S., Khan, S.H., Porikli, F.: Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. Asian Conf. Comput. Vis. (2018)", + "67. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "68. Razani, R., Cheng, R., Li, E., Taghavi, E., Ren, Y., Bingbing, L.: Gp-s3net: Graph-based panoptic sparse semantic segmentation network. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "69. Razani, R., Cheng, R., Taghavi, E., Bingbing, L.: Lite-hdseg: Lidar semantic segmentation using lite harmonic dense convolutions. In: Int. Conf. Rob. Automat. (2021)", + "70. Sautier, C., Puy, G., Gidaris, S., Boulch, A., Bursuc, A., Marlet, R.: Image-to-lidar self-supervised distillation for autonomous driving data. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "71. Seidenschwarz, J., Osep, A., Ferroni, F., Lucey, S., Leal-Taixe, L.: Semoli: What moves together belongs together. IEEE Conf. Comput. Vis. Pattern Recog. (2024)", + "72. Sirohi, K., Mohan, R., Bülicher, D., Burgard, W., Valada, A.: Efficientlps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics (2021)", + "73. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "74. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023)", + "75. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: Eur. Conf. Comput. Vis. (2020)", + "76. Teichman, A., Levinson, J., Thrun, S.: Towards 3D object recognition via classification of arbitrary object tracks. In: Int. Conf. Rob. Automat. (2011)", + "77. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Int. Conf. Comput. Vis. (2019)", + "78. Thorpe, C., Herbert, M., Kanade, T., Shafer, S.: Toward autonomous driving: the cmu navlab. i. perception. IEEE expert 6(4), 31-42 (1991)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "79. Thrun, S., Montemerlo, M., Dahlkamp, H., Stavens, D., Aron, A., Diebel, J., Fong, P., Gale, J., Halpenny, M., Hoffmann, G.: Stanley: The robot that won the darpa grand challenge. Journal of field Robotics (2006)", + "80. Wang, Y., Chen, X., You, Y., Li, L., Hariharan, B., Campbell, M., Weinberger, K., Chao, W.: Train in germany, test in the usa: Making 3d object detectors generalize. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "81. Wong, K., Wang, S., Ren, M., Liang, M., Urtasun, R.: Identifying unknown instances for autonomous driving. In: Conference on Robot Learning. pp. 384-393. PMLR (2020)", + "82. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: Int. Conf. Rob. Automat. (2018)", + "83. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: Int. Conf. Rob. Automat. (2019)", + "84. Xian, Y., Lampert, C.H., Schiele, B., Akata, Z.: Zero-shot learning - a comprehensive evaluation of the good, the bad and the ugly. IEEE Trans. Pattern Anal. Mach. Intell. (2018)", + "85. Xiong, X., Munoz, D., Bagnell, J.A., Hebert, M.: 3-D Scene Analysis via Sequenced Predictions over Points and Regions. In: Int. Conf. Rob. Automat. pp. 2609-2616 (2011)", + "86. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "87. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "88. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)", + "89. Yilmaz, K., Schult, J., Nekrasov, A., Leibe, B.: Mask4d: Mask transformer for 4d panoptic segmentation. arXiv preprint arXiv:2309.16133 (2023)", + "90. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "91. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "92. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "93. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "94. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: Eur. Conf. Comput. Vis. (2022)", + "95. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)", + "96. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "97. Zhu, M., Han, S., Cai, H., Borse, S., Ghaffari, M., Porikli, F.: 4d panoptic segmentation as invariant and equivariant field prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Segment Anything in Lidar", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "ref_text", + "text": "98. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "bbox": [ + 217, + 146, + 785, + 189 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "A. Osep et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_model.json b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..350d8f71d7b830c5aa19eec910d5363d942b1ce3 --- /dev/null +++ b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_model.json @@ -0,0 +1,2935 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.284, + 0.141, + 0.723, + 0.187 + ], + "angle": 0, + "content": "Better Call SAL: Towards Learning to Segment Anything in Lidar" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.213, + 0.346, + 0.23 + ], + "angle": 0, + "content": "Aljosa Osep1*" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.214, + 0.491, + 0.229 + ], + "angle": 0, + "content": "Tim Meinhardt\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.214, + 0.645, + 0.229 + ], + "angle": 0, + "content": "Francesco Ferroni" + }, + { + "type": "text", + "bbox": [ + 0.665, + 0.214, + 0.76, + 0.229 + ], + "angle": 0, + "content": "Neehar Peri²" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.23, + 0.484, + 0.244 + ], + "angle": 0, + "content": "Deva Ramanan" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.23, + 0.633, + 0.244 + ], + "angle": 0, + "content": "Laura Leal-Taixe1" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.25, + 0.569, + 0.265 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.274, + 0.575, + 0.289 + ], + "angle": 0, + "content": "\\(^{1}\\)NVIDIA \\(^{2}\\)CMU" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.334, + 0.357, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.436, + 0.356, + 0.456 + ], + "angle": 0, + "content": "Class-agnostic instance segmentation" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.334, + 0.498, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.436, + 0.487, + 0.457 + ], + "angle": 0, + "content": "Text prompts: {car, building,...}" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.334, + 0.638, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.436, + 0.611, + 0.456 + ], + "angle": 0, + "content": "Text prompts: {trash bin}" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.334, + 0.781, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.436, + 0.753, + 0.456 + ], + "angle": 0, + "content": "Text prompts: {streetcar}" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.468, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Fig. 1: The SAL (Segment Anything in Lidar) model performs class-agnostic instance segmentation (i) and zero-shot classification via text prompting. This allows us to not only predict panoptic segmentation (ii) for fixed class vocabularies but segment any object (iii and iv) in a given Lidar scan." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.591, + 0.744, + 0.84 + ], + "angle": 0, + "content": "Abstract. We propose the SAL (Segment Anything in Lidar) method consisting of a text-promptable zero-shot model for segmenting and classifying any object in Lidar, and a pseudo-labeling engine that facilitates model training without manual supervision. While the established paradigm for Lidar Panoptic Segmentation (LPS) relies on manual supervision for a handful of object classes defined a priori, we utilize 2D vision foundation models to generate 3D supervision \"for free\". Our pseudolabels consist of instance masks and corresponding CLIP tokens, which we lift to Lidar using calibrated multi-modal data. By training our model on these labels, we distill the 2D foundation models into our Lidar SAL model. Even without manual labels, our model achieves \\(91\\%\\) in terms of class-agnostic segmentation and \\(54\\%\\) in terms of zero-shot LPS of the fully supervised state-of-the-art. Furthermore, we outperform several baselines that do not distill but only lift image features to 3D. More importantly, we demonstrate that SAL supports arbitrary class prompts, can be easily extended to new datasets, and shows significant potential to improve with increasing amounts of self-labeled data. We release all models and the code." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.162 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.221 + ], + "angle": 0, + "content": "We tackle segmentation and recognition of objects in Lidar point clouds, a task commonly tackled via Lidar Panoptic Segmentation (LPS)." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.228, + 0.787, + 0.35 + ], + "angle": 0, + "content": "Status quo. LPS has been gaining significant attention in the community due to its role in scene understanding, which is vital for safe autonomous navigation. Thanks to the availability of labeled datasets [6, 18] and advances in learning representations from unordered point sets [12, 63, 64, 77], the Lidar community made significant progress in learning to segment and classify instances of predefined and manually-labeled classes. While this progress has been impressive, existing models [2, 26, 34, 43, 68, 72, 96] cannot adapt to continually evolving class ontologies [38] that may even vary in different geographic regions [80]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.432 + ], + "angle": 0, + "content": "Stirring the pot. We challenge this well-established approach and investigate how we can train general LPS models that can be prompted to segment point clouds according to any object class vocabulary. Towards such a promptable Lidar segmentation approach, we first need to be able to segment any object, a capability that remains elusive in the Lidar domain." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.681 + ], + "angle": 0, + "content": "Towards segmenting anything in Lidar. To address these challenges, we propose SAL, which consists of a text-promptable zero-shot model (Fig. 2, left) for panoptic segmentation of arbitrary objects (Fig. 1), and pseudo-labeling engine (Fig. 2, right) that facilitates model training directly from raw sensory data. The SAL pseudo-label engine automatically labels Lidar sequences using image segmentation [30] and vision-language models [65]. We utilize SAM [30] to generate class-agnostic masks in images, and CLIP [65] to generate per-mask tokens that connect visual features to language, and finally, transfer both to Lidar using a calibrated sensory setup. Even though generated pseudo-labels only partially cover Lidar scans and are inherently noisy due to errors in the image-level generation process and imperfect sensory calibration, we demonstrate their effectiveness as self-supervised training data for the SAL zero-shot model. In contrast to prior work in LPS, SAL is the first model that can be prompted with any new vocabulary without re-training. Different from recent advances [56] that lift image features to 3D without distillation into a model, SAL does not require image features during inference and segments full \\(360^{\\circ}\\) point clouds." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Talk to your SAL. Our model predicts Lidar segmentation masks and their corresponding CLIP tokens, which allow us to perform zero-shot classification of segmented objects using arbitrary text prompts (Fig. 1). We evaluate SAL by prompting our zero-shot model on standard benchmarks for LPS. For class-agnostic segmentation, we reach \\(91\\%\\) of the performance of manually supervised baselines. By classifying segmented objects in a zero-shot manner, we report the first (and very encouraging) results for zero-shot LPS and reach \\(42\\%\\) and \\(54\\%\\) of the supervised baselines trained on SemanticKITTI and nuScenes, respectively. Beyond that, as shown in Fig. 1, we can prompt SAL model to segment objects outside of existing Lidar dataset class vocabularies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment Anything in Lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.436, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.147, + 0.78, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.279, + 0.785, + 0.336 + ], + "angle": 0, + "content": "Fig. 2: SAL overview: Given a Lidar scan and a class vocabulary prompt, specified as a list of per-class free-form text descriptions (left), SAL segments and classifiers objects (things and stuff classes). As labeled data for training such a model does not exist, we supervise SAL by distilling off-the-shelf vision foundation models to Lidar (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.462 + ], + "angle": 0, + "content": "Contributions. As our main contribution, we re-think the established approach to Lidar Panoptic Segmentation and (i) present SAL for segmentation and classification of any object in a Lidar scan. As we do not utilize labeled Lidar data, we propose (ii) a pseudo-label engine that distills vision foundation models to the Lidar domain. The resulting pseudo-labels allow us to train (iii) a zero-shot LPS model without any human supervision. We demonstrate (iv) encouraging results on standard LPS benchmarks and outline a clear path towards universal, promptable segmentation foundation models for Lidar data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.387, + 0.499 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.784, + 0.575 + ], + "angle": 0, + "content": "Lidar has played a pivotal role since the dawn of embodied navigation [60,78,79]. Recent data-driven efforts in Lidar perception have been pushing boundaries in semantic segmentation [3,12,35,47,69,75,82,83,85,98], object detection [32,40, 42,57-59,88,90,95], and tracking [25,49,76]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.785, + 0.76 + ], + "angle": 0, + "content": "Lidar panoptic segmentation. Recently, Lidar Panoptic Segmentation [4,6, 7,18] has emerged as a holistic approach to Lidar-based dynamic scene understanding. Prior works [2,4,19,26,28,31,34,43-45,68,89,96,97] learn to group and classify points specified in the training data according to the target class vocabularies, while methods for open-set instance [24,49,76] and panoptic segmentation [81] rely on bottom-up grouping based on Euclidean distance between points. While these developments have been impressive, end-to-end methods remain limited to specific target classes that appear in the training data, while bottom-up, clustering-based approaches [24,27,49,76,81] are sensitive to a particular choice of clustering parameters and are unable to improve their performance in a data-driven fashion. In contrast, the SAL zero-shot model not only learns to segment objects but is also capable of zero-shot classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Self-supervised learning for Lidar. As labeled data in the Lidar domain is scarce, several works investigate how contrastive learning, proven effective in the image domain [10, 23], can reduce the need for labeled data. SegContrast [52] learns to align representations of point clouds and their augmented views, utilizing density-based clustering (DBSCAN [17]) to pool and contrast" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.239 + ], + "angle": 0, + "content": "features. Rather than using DBSCAN, [39, 70] leverage image-based methods [1] and foundation models [30] to perform contrastive alignment between image and corresponding Lidar features. The aforementioned methods only provide pretraining recipes for potential segmentation downstream tasks. Our SAL method goes a step further by training a full zero-shot Lidar Panoptic Segmentation model that could benefit from such pre-training recipes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Self-supervised (3D) object detection. Several methods (self-)supervise object detectors with (RGB-D) videos [22, 53-55, 61, 62], or Lidar sequences [50, 71, 92] to discover objects in sensory streams. However, such motion-based supervision provides no semantic information. Therefore, [51] distills per-point (PCA quantized) CLIP [65] features to Lidar for zero-shot classification of detected objects. Significantly different from [51], we tackle the more general and challenging task of LPS, which segments and classifies both moving and stationary things and stuff classes. To this end, we distill CLIP features not per point but per object instance. This provides a more holistic object-centric semantic distillation and allows us to supervise our model with non-quantized CLIP features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Zero-shot recognition. Zero-shot recognition (ZSR) [84] methods (recently also referred to as \"open-vocabulary recognition\") tackle the recognition of distinct, yet related, semantic concepts that are either not labeled in the training data (transductive) or not observed at all (inductive). Such methods assume a dataset with labels for some classes, whereas unseen instances are supervised via attributes or class names. The latter is often used in conjunction with word embeddings [46]. By learning to align image-language features, ZSR methods can infer class labels for unseen objects [65]. ZSR has also been explored in the context of object detection, semantic segmentation, and panoptic segmentation [16,86] using word embedding models [5,8,48,66], or CLIP vision-language embeddings [20,21,33,36,67,87,91,93,94]. In the context of semantic segmentation in RGB-D and Lidar data, [41,56] augment image and language embeddings with features extracted from point clouds. Both rely on lifting dense image features from images to 3D during inference, which limits their applicability to setups with dense camera coverage. Similarly, [74] segments instances in accumulated RGB-D point clouds (ScanNet [14]) and relies on manual supervision in conjunction with image-based dense CLIP features for zero-shot classification. Moreover, datasets such as ScanNet [14] were recorded with RGB-D sensors in static environments and fused into 3D reconstructions based on scans, taken from multiple viewpoints. In contrast, SAL directly applies to full Lidar point clouds while requiring no image features at the inference time." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.744, + 0.573, + 0.763 + ], + "angle": 0, + "content": "3 SAL: Segment Anything in Lidar" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This section outlines the key challenges and components towards a Lidar model that segments and classifies any object. We first define the underlying problem formation in Sec. 3.1 and then present our (Segmenting Anything in Lidar) SAL model in Sec. 3.2." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment Anything in Lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.434, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.248, + 0.397, + 0.26 + ], + "angle": 0, + "content": "(a) Pseudo-label engine" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.145, + 0.782, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.248, + 0.673, + 0.26 + ], + "angle": 0, + "content": "(b) Zero-shot model" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 3: Our pseudo-label engine (Fig. 3a) utilizes SAM [30] to estimate segmentation masks in images, MaskCLIP [16] to estimate corresponding per-mask CLIP features, and a calibrated sensory setup to transfer them to the Lidar domain. We distill these pseudo-labels to our zero-shot model (Fig. 3b), which segments and classifies Lidar point clouds. The SAL zero-shot model employs a sparse-convolutional backbone [12], followed by a Transformer decoder that predicts objectness scores, segmentation masks, and CLIP tokens for each query. To (optionally) perform zero-shot classification, we forward the dataset class vocabulary through the CLIP text encoder and match the encoded vocabulary with predicted CLIP tokens. Our model requires no retraining for different vocabularies and no image features at inference time." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.417, + 0.59, + 0.432 + ], + "angle": 0, + "content": "3.1 Zero-Shot Lidar Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.626 + ], + "angle": 0, + "content": "The task. We discuss and evaluate SAL in the context of Lidar Panoptic Segmentation (LPS), which tackles both instance segmentation and object recognition. LPS methods take as input an unordered point cloud \\( P \\in \\mathbb{R}^{N \\times 4} \\) encoding spatial coordinates and their sensor intensity. The currently established problem setting [7,18] assumes supervision and task outputs as per-point semantic class and instance identity labels. The class labels are confined to a pre-defined and fixed class vocabulary \\( \\mathcal{V} \\). In contrast, we tackle LPS in a generalized, zero-shot setting, where the class vocabulary is only provided at inference, not training time. Such a vocabulary \\( \\mathcal{V} \\) specifies target classes as a list of \\( C \\) class prompts. Each prompt \\( c_{i} \\in \\mathcal{V} \\) is specified via free-form text, e.g., with a class name and optional class description. To perform zero-shot LPS a model must be designed to segment and classify any object in a Lidar scan." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.692 + ], + "angle": 0, + "content": "The challenge. Where do we get the data to train such a model? Image-based models [16,21,30,91,93] rely on large and diverse datasets for pre-training [15], dense prediction [13,37], and foundation models that align images with textual descriptions [65] - a commodity not available in the Lidar domain." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.712, + 0.373, + 0.725 + ], + "angle": 0, + "content": "3.2 SAL Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our 2D-to-Lidar distillation method consists of two core components: (i) The pseudo-label engine transfers 2D vision foundation models into Lidar pseudolabels using multi-modal inputs from a calibrated sensory setup, shown in Fig. 3a. (ii) Our zero-shot model is trained on the generated pseudo-labels and is able to perform class-agnostic segmentation and zero-shot classification via text prompts. Given a semantic dataset vocabulary, this allows us to tackle zero-shot LPS as illustrated in Fig. 3b." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.469, + 0.162 + ], + "angle": 0, + "content": "3.3 SAL Pseudo-Label Engine" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.785, + 0.219 + ], + "angle": 0, + "content": "The SAL pseudo-label engine relies on a calibrated multi-modal sensory setup with \\( k \\geq 1 \\) RGB cameras. Furthermore, a sufficient overlap between the Lidar sensing area and each camera view is paramount." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.787, + 0.33 + ], + "angle": 0, + "content": "Mask generation: To pseudo-label object instances, we utilize the Segment Anything (SAM) model [30], which generates an overlapping set of segmentation masks. We flatten SAM's output mask hierarchy by non-maxima suppression (NMS) with a minimal overlap threshold to ensure mutually exclusive masks, suppressing object parts and subparts in favor of objects. This way, we obtain a set of non-overlapping binary masks \\( m_{i}^{k} \\in \\{0,1\\}^{W \\times H} \\) for each camera view \\( k \\) with image plane of size \\( W \\times H \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.472 + ], + "angle": 0, + "content": "CLIP image token generation: As shown by CLIP [65], aligning image and text features allows vision-language foundation models to perform zero-shot image classification based on arbitrary text prompts. To transfer this capability to the Lidar domain, we generate a localized CLIP image feature token \\( f_{i}^{k} \\in \\mathbb{R}^{C_{t}} \\) for each binary SAM mask \\( m_{i}^{k} \\). To obtain image tokens for a masked region of the input image, we utilize [16] and their relative mask attention in the CLIP image encoder feature space. Note that we never classify segments in the image domain but merely distill CLIP image feature tokens to Lidar to facilitate zero-shot classification during inference where we do not use any image features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Image-to-Lidar unprojection: From the Lidar perspective, we unproject each image mask \\(m_i^k \\in \\{0,1\\}^{W \\times H}\\) to a binary Lidar segmentation mask \\(\\tilde{m}_i \\in \\{0,1\\}^N\\) by transforming the respective camera coordinate frame to the Lidar space. For datasets with multiple cameras, such as nuScenes [18], we process each image independently, followed by a cross-camera fusion of masks with a sufficient IoU overlap and averaging of their CLIP image tokens. The unprojection yields pairs \\(\\{\\tilde{m}_i^k, f_i^k\\}\\) of Lidar masks and their corresponding CLIP features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.725 + ], + "angle": 0, + "content": "Refinement via density-based clustering: Image-to-Lidar transformation (Fig. 4a→Fig. 4b) is inherently noisy due to imperfect calibration and issues with synchronization and Lidar rolling shutter. We improve our pseudo-label quality by creating an ensemble of DBSCAN [17] clusters \\(\\tilde{m}_l \\in \\{0,1\\}^N\\), obtained by varying the density threshold to compensate for varying density in Lidar point clouds (details in the appendix). We replace each \\(\\tilde{m}_i\\) with its best-matching \\(\\tilde{m}_l\\) in case their IoU exceeds a minimal overlap threshold and retain the original mask otherwise to obtain a refined set of pseudo-labels (Fig. 4c) that retain their original cardinality and associated CLIP features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.436, + 0.763 + ], + "angle": 0, + "content": "3.4 SAL Zero-Shot Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.785, + 0.805 + ], + "angle": 0, + "content": "The universal SAL zero-shot model deconstructs LPS into class-agnostic segmentation and zero-shot classification via text prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Universal architecture. Instead of relying on highly specialized and engineered LPS models, we base the SAL model on a universal Transformer decoder" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment Anything in Lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.402, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.209, + 0.357, + 0.221 + ], + "angle": 0, + "content": "(a) SAM masks" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.145, + 0.596, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.209, + 0.565, + 0.221 + ], + "angle": 0, + "content": "(b) SAM unprojected" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.145, + 0.788, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.209, + 0.758, + 0.221 + ], + "angle": 0, + "content": "(c) SAM + DBSCAN" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.232, + 0.785, + 0.274 + ], + "angle": 0, + "content": "Fig. 4: Refinement via clustering. After transferring image masks (Fig. 4a) to Lidar (Fig. 4b), we obtain pseudo-labels that suffer from sensory misalignment-related issues. Our geometric refinement (Fig. 4c) improves localization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.785, + 0.404 + ], + "angle": 0, + "content": "architecture (Fig. 3b), similar to [43]. Its 2D counterparts [9, 11] provide top-performing results across semantic, instance, and panoptic segmentation for images. With increasing amounts of (pseudo-)labeled data, Transformer decoders have the potential to achieve a similar impact in the Lidar domain. Following [43], we deploy a Minkowski U-Net [12] backbone for feature extraction followed by a Transformer decoder architecture with object query to point/voxel feature crossattention. Training our model to segment and classify any object and perform in a zero-shot evaluation setting requires a unique design of the final task heads." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.785, + 0.5 + ], + "angle": 0, + "content": "Class-agnostic segmentation. To localize segments in a point cloud, we rely on two model heads that predict segmentation masks and objectness scores for each query. The architecture of the former follows [43] and predicts binary output masks \\(\\hat{m}_i\\in \\{0,1\\} ^N\\) by computing the dot product between queries and point features. The objectness head reduces the multi-class problem to a binary object or no-object decision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.785, + 0.626 + ], + "angle": 0, + "content": "Zero-shot classification. To equip our model with zero-shot classification capabilities without relying on any image input, we learn to predict CLIP [65] Lidar tokens, i.e., features in the CLIP space obtained from Lidar inputs, for each query. The token head consists of a three-layer MLP that directly regresses tokens \\(\\hat{f}_i^j\\). At inference, the predicted tokens are matched to text prompts via the CLIP text encoder. The matching yields a probability distribution over the prompts. To perform zero-shot LPS on a pre-defined dataset, we use its class vocabulary as input text prompts (Fig. 3b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.785, + 0.657 + ], + "angle": 0, + "content": "We train our model jointly for both class-agnostic segmentation and zero-shot classification with the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.671, + 0.785, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S A L}} = \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {\\text {t o k e n}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.785, + 0.774 + ], + "angle": 0, + "content": "with binary cross-entropy loss \\(\\mathcal{L}_{obj}\\) and a cosine distance loss \\(\\mathcal{L}_{token}\\). The segmentation loss \\(\\mathcal{L}_{seg}\\) follows [43] and consists of a binary mask cross-entropy and dice loss. The \\(\\mathcal{L}_{\\mathrm{SAL}}\\) loss is supervised by pseudo-label pairs \\(\\{\\tilde{m}_i^k, f_i^k\\}\\) obtained from our label engine and thereby distills both 2D foundation models (SAM and CLIP) into our LPS model." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "How to train with partial labels? Partially (pseudo-)labeled point clouds present a challenge for SAL model training. For example, in SemanticKITTI, only \\(14\\%\\) of points are pseudo-labeled due to low camera coverage (see appendix). If we naively train SAL on partially (pseudo-)labeled point clouds, the objectness" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "loss \\(\\mathcal{L}_{obj}\\) would penalize any segmentation in these regions as a false positive, thereby teaching the model to ignore these points entirely. Empirically, we determined that the most effective training strategy is to (i) remove unlabeled points from point clouds, (ii) utilize standard data augmentations (rotations, flipping, scaling, and translations), in conjunction with proposed FrankenFruustum augmentation (detailed in the appendix) that mimics fully-labeled point clouds during training by randomly removing unlabeled points and replicating labeled frustum regions around the vertical axis. This augmentation does not increase the overall label coverage, but our ablations in Sec. 4.2 show its effectiveness in reducing the domain gap between training and inference input clouds." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.333, + 0.377, + 0.351 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.455 + ], + "angle": 0, + "content": "In the following section, we outline our experimental setup (Sec. 4.1) used to ablate the SAL model and its training strategies (Sec. 4.2). Since SAL presents the first Zero-Shot Lidar Panoptic Segmentation (ZS-LPS) model, we design and compare with multiple hand-crafted baselines that implement concepts from related works. We conclude our analysis by comparing our zero-shot performance (Sec. 4.4) to fully-supervised LPS methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.471, + 0.429, + 0.487 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Datasets. We evaluate SAL on two public Lidar Panoptic Segmentation (LPS) datasets, SemanticKITTI [6, 7] and nuScenes [18]. We utilize provided, human-annotated ground-truth (GT) panoptic segmentation labels solely for evaluation purposes, linear probing ablations, and baselines. We train SAL model only using our self-supervised pseudo-labels (Sec. 3.3). To demonstrate the general applicability of SAL to other domains and datasets, we show qualitative results of SAL trained on Waymo Open [73] dataset, which contains no GT LPS labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Metrics. For the evaluation, we utilize the standard Panoptic Segmentation [29] metrics. The Panoptic Quality \\( PQ = RQ \\times SQ \\) combines Segmentation Quality (SQ) and Recognition Quality (RQ), i.e., F-1 score. True positives (TPs) are mask predictions with sufficient intersection-over-union (IoU) overlap with GT masks of the same class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Semantic Oracle (SO). The aforementioned metrics assume GT labels and predictions with instance IDs and semantic classes for each point. Therefore, they are not suitable for evaluating class-agnostic segmentation. To assess the class-agnostic segmentation independently of semantics, we apply SO during the evaluation by assigning ground truth semantic classes \\( c_{i} \\) to predicted masks \\( m_{i} \\) via majority voting. Evaluation with SO is only needed for evaluation of class-agnostic segmentation - our zero-shot models provide both instance and semantic class predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Stuff Merging (SM). SO allows us to assess class-agnostic instance segmentation for things classes. However, this approach is unsuitable for assessing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment Anything in Lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "the performance of stuff classes, as existing Lidar datasets merge instances of stuff classes into a single instance. The things- stuff separation can be unintuitive and inconsistent across datasets (e.g., SemanticKITTI merges instances of traffic signs into a single stuff class). In contrast, we fully embrace the philosophy that all classes can be segmented into individual instances (e.g., segmentation models should localize individual trees/bushes in the vegetation class or distinguish individual buildings). To evaluate our models on the target datasets in the SO regime on Lidar datasets that do not provide instance-level annotations for all classes, we additionally report results using a merge strategy (SM), which merges all instances of a particular stuff class into a single mask to ensure outputs of our models are consistent with the format of target datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.338, + 0.375, + 0.352 + ], + "angle": 0, + "content": "4.2 SAL Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.787, + 0.412 + ], + "angle": 0, + "content": "In this section, we discuss how to train SAL zero-shot segmentation model using self-generated pseudo-labels (Tab. 1) and discuss design decisions behind our instance (Tab. 2) and semantic distillation (Tab. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Learning with partial labels. For the single-camera setup [6,7], our pseudolabels only cover \\(14\\%\\) of the input point clouds (see appendix). Training on partial label coverage presents a challenge for our model. In Tab. 1, we only evaluate class-agnostic segmentation by applying SO and SM (see Sec. 4.1). We compare SAL models trained using pseudo and GT labels. The Frustum Filter removes all unlabeled points not visible in the camera frustum during training and/or evaluation. As shown in Tab. 1, removing unlabeled points from pseudo-labeled point clouds during training has a significant benefit on the performance (\\(3^{rd}\\) row, 59.3 PQ) as compared to the variant, where we merely ignore unlabeled region during the training (\\(2^{nd}\\) row, 22.2 PQ). This is likely due to data imbalance in a single-camera setup, leading to a class imbalance between labeled/unlabeled points, incentivizing SAL to suppress predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.841 + ], + "angle": 0, + "content": "While utilizing standard augmentations (rotations, flipping, scaling, and translations) is crucial, we observe that performing our FrankenFrustum augmentation, which concatenates the visible portion of the point cloud around the z-axis (see Sec. 3.4), significantly improves results (\\(4^{th}\\) row, 62.5 PQ). Mixing point clouds from different scans (\\(5^{th}\\) row, 62.8 PQ) further boosts results. Overall, we obtain 69 PQ from GT supervision (1st row) and 62.8 PQ when using our pseudo-labels (5th row). Remarkably, our labels yield \\(91\\%\\) of the GT performance while covering only \\(14\\%\\) of the input point cloud (see appendix). To further contextualize these results, we note that this evaluation is performed only on classes for which GT instance labels are available. As shown in Fig. 1, our model learns to segment a much larger variety of classes. Evaluating only the subset of points visible in the camera (\\(14\\%\\) of all points, rows 6&7), the gap between the GT-supervised model (71.8 PQ) and pseudo-supervised model (70.7 PQ) shrinks further. We conclude that by training SAL model using pseudo-labels, we distill the notion of objectness from SAM [30] into our model and obtain segmentation capabilities, similar to the model trained on GT data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.52, + 0.228 + ], + "angle": 0, + "content": "Table 1: Class-agnostic segmentation. By cropping unlabeled points and performing data augmentations (in combination with our FrankenFrostum), SAL successfully learns to segment full point clouds even when only \\(14\\%\\) of points are (pseudo)-labeled." + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.145, + 0.783, + 0.229 + ], + "angle": 0, + "content": "Table 2: Scaling. In contrast to predefined GT data, increasing the number of queries SAL improves performance on stuff classes. By increasing the amount of labeled data, we further improve segmentation performance." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.24, + 0.518, + 0.367 + ], + "angle": 0, + "content": "
LabelsFrust.-Filter TrainEvalFranken Frust.PQSQ\\( PQ^{Th} \\)\\( PQ^{St} \\)
GT69.083.581.659.7
Pseu.22.267.545.15.5
Pseu.✓59.378.265.854.5
Pseu.✓✓62.579.167.658.8
Pseu.✓✓(mix)62.879.069.058.3
GT✓71.884.884.662.5
Pseu.✓✓✓(mix)70.781.975.467.3
" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.24, + 0.781, + 0.359 + ], + "angle": 0, + "content": "
TrainNum.PQSQ\\( PQ^{Th} \\)\\( PQ^{St} \\)
LabelsSetQuer.
GTtrain10067.882.479.759.2
GTtrain20067.983.679.859.3
GTtrain30069.083.581.659.7
Pseu.train10053.975.362.247.8
Pseu.train20060.877.066.756.5
Pseu.train30062.879.069.058.3
Pseu.bigtrain30065.379.571.960.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Scaling queries and data. Analogous to the grid inference of SAM, which operates with hundreds of spatial queries (prompts), we analyze the effect of different numbers of decoder queries. We subsample instances during training to train models with fewer queries than the maximum number of segments per scan. As visualized in Tab. 2, increasing the number of decoder queries improves the recognition of things for both training with pseudo and GT labels. However, we observe a significant performance boost with pseudo labels for stuff classes \\((58.3\\mathrm{PQ}^{St}, +10.8)\\), while the improvement for GT labels is marginal \\((59.7\\mathrm{PQ}^{St}, +0.5)\\). Our model benefits from a larger number of queries for stuff classes as it learns a fine-grained segmentation model that is not limited to a prefixed set of things classes. In the last row, we concatenate pseudo-labeled train and test sets (neither used for the validation) into a bigtrain set and observe an improvement of \\(+2.5\\) PQ, suggesting that scaling the amount of training data has the potential to improve the performance of our models further." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.615, + 0.459, + 0.739 + ], + "angle": 0, + "content": "Table 3: Semantic distillation. We linearly probe the SAL model, trained with and without semantic distillation loss \\(\\mathcal{L}_{token}\\). We train a linear classifier with GT labels while keeping backbone and decoder features frozen. As can be seen, \\(\\mathcal{L}_{token}\\) successfully distills a notion of semantics into our model." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.741, + 0.455, + 0.803 + ], + "angle": 0, + "content": "
Linear prob. LtokenPQRQSQmIoU
×20.026.355.223.4
××33.141.968.340.0
×24.832.366.829.7
" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.599, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Semantic distillation. At inference, SAL predicts a binary segmentation mask and objectness score for each query. To perform zero-shot classification, our model additionally predicts CLIP image tokens. This allows us to not rely on image features during inference but prompt our predicted tokens with arbitrary texts, i.e., object classes. Does the token head loss \\(\\mathcal{L}_{token}\\) successfully distill a notion of object semantics from CLIP into our model? To verify this, we perform linear probing experiments: we replace the token distillation head of a trained model with a linear classifier (backbone and instance de" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "coder remain frozen) and train it using GT labels. As can be seen in Tab. 3, distilling CLIP features into our model significantly improves linear probing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment Anything in Lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 4: Zero-shot panoptic segmentation. We utilize prior efforts in the image domain [30] and Lidar [56] domain to craft multiple baselines that only unproject segmentation masks and lift image features to Lidar. By contrast, SAL distills outputs of such baselines (pseudo-labels) into a stronger Lidar segmentation model. With ImageFeat, we denote methods that require image features at inference time, and Frust. Eval. denotes the evaluation of a subset of points visible in the camera." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.24, + 0.785, + 0.44 + ], + "angle": 0, + "content": "
MethodFrost. ImageDefault classesSuper classes
EvalFeat.PQSQPQThPQStmIoUPQSQPQThPQStmIoU
Class-agnostic Segmentation (Semantic Oracle)
SAM✓✓46.072.149.743.4------
SAM+Erosion✓✓42.269.445.639.6------
SAM+DBS (filter)✓✓46.770.376.824.8------
SAM+DBS (replace)✓✓48.773.753.145.4------
SAL✓X70.781.975.467.3------
Zero-Shot Lidar Panoptic Segmentation
SAM+DBS+CLIP✓✓27.571.531.724.530.651.177.571.241.054.3
SAL✓X33.171.422.840.533.563.984.288.351.766.4
SAM+DBS+CLIPXX8.256.418.60.67.511.547.60.017.311.2
SALXX24.866.817.430.228.748.578.880.432.652.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.557 + ], + "angle": 0, + "content": "performance (33.1 PQ, 40 mIoU) compared to the baseline, not trained with the semantic distillation loss \\(\\mathcal{L}_{token}\\) (20.0 PQ, 23.4 mIoU). This confirms that by distilling CLIP features, we inject a notion of semantics into our model, even though our model is not explicitly supervised with any labels containing semantic information - the notion of semantics is learned implicitly via CLIP feature distillation. Without this loss, our zero-shot model (24.8 PQ, 29.7 mIoU) outperforms the linearly probed model tuned using GT semantic labels." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.587, + 0.753, + 0.604 + ], + "angle": 0, + "content": "4.3 Class-Agnostic and Zero-Shot Lidar Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.729 + ], + "angle": 0, + "content": "As the first study tackling Zero-Shot Lidar Panoptic Segmentation (ZS-LPS), we devise several strong baselines inspired by prior works in RGB-D semantic segmentation. During inference, these methods utilize image-based models and lift image features to 3D [56]. To ensure these are evaluated fairly, we report results in Tab. 4 on the subset of points visible in at least one camera (Frost. Eval.). Our SAL model does not have this limitation. Therefore, we additionally report results evaluated on full point clouds in the bottom of Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Class-agnostic segmentation. Our first baseline (SAM [30]) generates masks in images and unprojects them to Lidar (see Sec. 3.3), and leads to 46 PQ. We observe bleeding edges (see Fig. 4b) after unprojecting from the cameras to Lidar caused by imperfect calibrations. To mitigate these artifacts, we experiment with slightly eroding SAM predictions in the image domain. However, this leads to a performance drop (42.2 PQ), likely because smaller segments after erosion fail to pass \\(>0.5\\) overlap with labeled instances." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Density-based filtering: Alternatively, to mitigate unprojection errors, we experiment with density-based clustering methods. We generate a large pool of DBSCAN (DBS) clusters and test two strategies to improve our SAM-based segmentation labels: (i) filtering segments without a sufficient overlap with any DBSCAN cluster or (ii) replacing segments with large overlaps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.788, + 0.305 + ], + "angle": 0, + "content": "SAM+DBS (filter): Intuitively, the filter approach removed Lidar segments only segmentable in the image domain, e.g., a SAM mask that segments a shadow on a flat wall. While this strategy improves precision for things classes (76.8 PQTh, +27.1), it significantly degrades performance on stuff classes (24.8 PQSt -18.6) and overall degrades PQ to 46.7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.372 + ], + "angle": 0, + "content": "SAM+DBS (replace): The replace strategy (that replaces SAM masks with DBSCAN clusters with sufficient mutual overlap), on the other hand, improves both things and stuff classes to an overall PQ of 48.7 (+3.9) and SQ 73.7 (+1.1), respectively. This improvement can be visually verified in Fig. 4c." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.452 + ], + "angle": 0, + "content": "SAL: We train our SAL model using top-performing \\( SAM+DBS \\) (replace) and observe a significant improvement (70.7 PQ) in terms of both things and stuff classes. Even though SAL obtains significantly higher PQ than hand-crafted baselines, it also improves in terms of SQ (81.9 vs. 73.7 closest competitor), suggesting that the distilled model is robust to projection artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.624 + ], + "angle": 0, + "content": "Do we need a vision segmentation foundation model? To generate a set with a sufficiently high recall, we need to cluster points with varying DBSCAN radius parameters (details in the appendix). This leads to a large set of predicted masks per scan (5,413 avg.). While most of the correct segments are in this set, separating the signal from the noise is difficult. Prior works, such as [50], are limited to moving thing classes as they rely on motion cues to filter static point cloud regions and DBSCAN to group remaining points. Our core insight is that we can utilize vision foundation models that already learned a notion of objectness: SAM generates only 171 masks per image, reduced to 45 via flattening, and to 39 masks after the unprojection, achieving a high recall with only a few instances, or in other words, a high signal-to-noise ratio." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Zero-shot segmentation. So far, we discussed class-agnostic segmentation. To obtain semantic class predictions per mask in a zero-shot fashion, we devise a simple baseline inspired by [56]. Once masks are unprojected from the image to Lidar space, we utilize associated CLIP features (directly extracted from images using [16]) to perform zero-shot classification via dot product between encoded class prompts and (lifted) CLIP features [65]. This baseline (SAM+DBS+CLIP) can be understood as an approach that directly lifts image features from state-of-the-art image-based models [16] to Lidar. We detail prompts in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "There is a significant performance drop between \\( SAM+DBS \\) (replace) baseline, evaluated with Semantic Oracle (\\( 5^{th} \\) row, 48.7 PQ), and zero-shot baseline (\\( SAM+DBS \\) (replace)+CLIP), \\( 7^{th} \\) row, 27.5 PQ. This is not surprising, as zero-shot segmentation is a challenging problem in both image [16] and Lidar domains. To this end, even when transferring features of state-of-the-art image-based models to Lidar we observe a significant performance drop. Remarkably," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment Anything in Lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 5: Lidar Panoptic Segmentation (LPS) on SemanticKITTI and nuScenes validation sets. We prompt our zero-shot SAL model with the respective class vocabularies and compare its performance to fully-supervised baselines. On SemanticKITTI and nuScenes, we reach \\(42\\%\\) and \\(54\\%\\) of the fully-supervised model, respectively. This gap reduces significantly when we evaluate super classes." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.227, + 0.753, + 0.44 + ], + "angle": 0, + "content": "
MethodSupervisionDefault classesSuper classes
PQRQSQ\\( PQ^{Th} \\)\\( PQ^{St} \\)mIoUPQRQSQmIoU
SemanticKITT1DS-Net [26]Full57.768.077.661.854.863.5---
PolarSeg [96]Full59.170.278.365.754.364.5---
EfficientLPS [72]Full59.269.875.058.060.964.9---
GP-S3Net [68]Full63.375.981.470.258.373.0---
MaskPLS [43]Full59.869.076.3---78.487.188.2
SALFull59.569.275.762.357.463.881.790.089.2
SALZero-shot24.832.366.817.430.228.748.559.478.8
mScenesPHNet [34]Full74.784.288.274.075.979.7---
DS-Net [26]Full51.259.086.138.472.373.5---
GP-S3Net [68]Full61.072.084.156.066.075.8---
EfficientLPS [72]Full62.073.983.456.870.665.6---
PolarSeg [96]Full63.475.383.959.270.466.9---
MaskPLS [43]Full57.766.071.864.452.262.571.581.086.2
SALFull70.580.885.979.461.772.874.282.787.1
SALZero-shot38.447.877.247.529.233.952.663.577.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.538 + ], + "angle": 0, + "content": "SAL distills such noisy image-based features to a stronger model (\\(8^{th}\\) row, 33.1 PQ, +5.6 PQ). Finally, to emphasize the inherent limitation of baselines, relying on image features, we evaluate on full \\(360^{\\circ}\\) point clouds. SAL performs similarly in both settings. The baseline, on the other hand, relies on image features and is, therefore, unable to predict labels out of the camera frustum. This problem can be mitigated by utilizing a setup with denser camera coverage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.62 + ], + "angle": 0, + "content": "Evaluation on super-classes. We observe that the CLIP classifier often confuses related classes, e.g., car and other vehicle. We additionally evaluate our semantics by prompting with super-classes (as defined in [6]). This alone increases PQ to 63.9 (48.5 when evaluating on full point clouds) and suggests significant potential for improving prompting and text-to-image alignment." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.643, + 0.597, + 0.658 + ], + "angle": 0, + "content": "4.4 Lidar Panoptic Segmentation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.788, + 0.76 + ], + "angle": 0, + "content": "Finally, we compare SAL with state-of-the-art LPS methods on standard benchmarks [6,18]. We focus this discussion on comparing SAL, trained with labeled data, and our SAL zero-shot model that specifies classes in benchmarks [6,18] as text prompts. Importantly, zero-shot results are obtained without training on any labeled data. In contrast, all baselines are trained using human-labeled data, and are constrained to their respective, pre-defined class vocabularies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "SemanticKITTI. When evaluating with the default class vocabulary (Tab. 5), our zero-shot model reaches 24.8 PQ or \\(42\\%\\) of the fully-supervised model (59.5 PQ). As discussed in Sec. 4.2, our method is mainly limited by the quality of the generated CLIP features. In practice, our zero-shot semantic classifier mainly confuses objects within super classes, e.g., car vs. other-vehicle. Therefore," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.178 + ], + "angle": 0, + "content": "without any retraining, we prompt our zero-shot model with super classes and obtain 48.5 PQ, \\(59\\%\\) of the supervised model (81.7 PQ)," + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.182, + 0.787, + 0.258 + ], + "angle": 0, + "content": "nuScenes. We report our result in Tab. 5 and reach similar conclusions. SAL reaches 38.4 PQ on default classes, \\(54\\%\\) of the supervised model (70.5 PQ). As with SemanticKITTI, results consistently improve when evaluated with superclasses. In this setting, zero-shot model yields 52.6 PQ. Note that SAL model, supervised with labeled data, significantly outperforms MaskPLS." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.399 + ], + "angle": 0, + "content": "Towards closing the gap. Compared to SemanticKITTI (24.8 PQ, \\(42\\%\\)), we observe a significantly smaller gap between the model trained on pseudo- and ground truth labels for nuScenes (38.4 PQ, \\(54\\%\\)). Despite its sparse Lidar signal and more diverse input scenarios, SAL performs better on this challenging Lidar dataset. We attribute this to a larger pseudo-label coverage of nuScenes (\\(48\\%\\), compared to \\(14\\%\\) coverage on SemanticKITTI), and, therefore, effectively more labeled data, which is also consistent with our observations on SemanticKITTI w.r.t. increasing the amount of labeled data (Tab. 2). This trend suggests a clear path forward: applying SAL to pseudo-label more data with zero annotation cost." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.525 + ], + "angle": 0, + "content": "Beyond labeled datasets. Datasets, such as SemanticKITTI and nuScenes, provide dense, point-level panoptic segmentation labels, which we use for analysis and the evaluation of our method on a small set of classes that were labeled in the respective dataset (14/11 things/stuff in SemanticKITTI, respectively, and 23/6 in nuScenes). To demonstrate the versatility of SAL, we additionally pseudolabel and train a model on Waymo Open [73] dataset and show qualitative results in Fig. 1 and more in the appendix. This effort takes less than a week of pseudolabeling and model training efforts (we provide details in the appendix)." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.635 + ], + "angle": 0, + "content": "The good, the (breaking) bad, and the ugly. While we are thrilled about class-agnostic segmentation performance, our experimental analysis reveals a substantial gap between full supervision and our zero-shot model. Finally, at the moment, SAL is limited to training on specific datasets. As in fully-supervised Lidar perception, cross-sensor generalization remains a challenge for SAL. This could be addressed by utilizing temporal context, data scaling efforts and investigating sensor-agnostic backbone networks." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.658, + 0.367, + 0.673 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.78 + ], + "angle": 0, + "content": "We proposed SAL, a method for Zero-Shot Lidar Panoptic Segmentation. The key SAL components are a pseudo-label engine that utilizes vision foundation models and a zero-shot model trained via self-supervision. While SAL is the first step in the direction of learning general Lidar segmentation models by distilling image-based foundation to Lidar, we believe we just scratched the surface and opened the door for training Lidar segmentation models without manual supervision." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.784, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Acknowledgments. This project was funded, in parts, by ERC Starting Grant DynAI (ERC-101043189). We are grateful to Žan Gojčić, Guillem Braso, Cristiano Saltori, Sérgio Agostinho, and Jonas Schult for their feedback on the paper and their insightful comments. Special thanks to Maxim Maximov for his help on figures." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment Anything in Lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell. 34(11), 2274-2282 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.259 + ], + "angle": 0, + "content": "2. Agarwalla, A., Huang, X., Ziglar, J., Ferroni, F., Laura, L.T., Hays, J., Osep, A., Ramanan, D.: Lidar panoptic segmentation and tracking without bells and whistles. In: Int. Conf. Intel. Rob. Sys. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Aksoy, E.E., Baci, S., Cavidar, S.: Salsanet: Fast road and vehicle segmentation in lidar point clouds for autonomous driving. In: Intel. Veh. Symp. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "4. Aygiun, M., Osep, A., Weber, M., Maximov, M., Stachniss, C., Behley, J., Leal-Taixe, L.: 4d panoptic lidar segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.785, + 0.355 + ], + "angle": 0, + "content": "5. Bansal, A., Sikka, K., Sharma, G., Chellappa, R., Divakaran, A.: Zero-shot object detection. In: Eur. Conf. Comput. Vis. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.396 + ], + "angle": 0, + "content": "6. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In: Int. Conf. Comput. Vis. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "7. Behley, J., Milioto, A., Stachniss, C.: A Benchmark for LiDAR-based Panoptic Segmentation based on KITTI. In: Int. Conf. Rob. Automat. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "8. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. Adv. Neural Inform. Process. Syst. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "9. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.521 + ], + "angle": 0, + "content": "0. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Adv. Neural Inform. Process. Syst. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.562 + ], + "angle": 0, + "content": "1. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "2. Choy, C., Gwak, J., Savarese, S.: 4D spatio-temporal convnets: Minkowski convolutional neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.673 + ], + "angle": 0, + "content": "4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.701 + ], + "angle": 0, + "content": "5. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.728 + ], + "angle": 0, + "content": "6. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip. In: Int. Conf. Mach. Learn. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.756 + ], + "angle": 0, + "content": "7. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Rob. Sci. Sys. (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.797 + ], + "angle": 0, + "content": "8. Fong, W.K., Mohan, R., Hurtado, J.V., Zhou, L., Caesar, H., Beijbom, O., Valada, A.: Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Rob. Automat. Letters (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.839 + ], + "angle": 0, + "content": "9. Gasperini, S., Mahani, M.A.N., Marcos-Ramiro, A., Navab, N., Tombari, F.: Panost-ter: End-to-end panoptic segmentation of lidar point clouds. IEEE Rob. Automat. Letters (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Eur. Conf. Comput. Vis. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "21. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.244 + ], + "angle": 0, + "content": "22. Harley, A.W., Zuo, Y., Wen, J., Mangal, A., Potdar, S., Chaudhry, R., Fragkiadaki, K.: Track, check, repeat: An em approach to unsupervised tracking. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.272 + ], + "angle": 0, + "content": "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "24. Held, D., Guillory, D., Rebsamen, B., Thrun, S., Savarese, S.: A probabilistic framework for real-time 3d segmentation using spatial, temporal, and semantic cues. In: Rob. Sci. Sys. (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "25. Held, D., Levinson, J., Thrun, S., Savarese, S.: Combining 3d shape, color, and motion for robust anytime tracking. In: Rob. Sci. Sys. (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "26. Hong, F., Zhou, H., Zhu, X., Li, H., Liu, Z.: Lidar-based panoptic segmentation via dynamic shifting network. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "27. Hu, P., Held, D., Ramanan, D.: Learning to optimally segment point clouds. IEEE Robotics and Automation Letters 5(2), 875-882 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "28. Hurtado, J.V., Mohan, R., Valada, A.: Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "29. Kirillov, A., He, K., Girshick, R.B., Rother, C., Dollar, P.: Panoptic segmentation. IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "31. Kreuzberg, L., Zulfikar, I.E., Mahadevan, S., Engelmann, F., Leibe, B.: 4d-stop: Panoptic segmentation of 4d lidar using spatio-temporal object proposal generation and aggregation. In: ECCV AVVision Workshop (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "32. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "33. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.646 + ], + "angle": 0, + "content": "34. Li, J., He, X., Wen, Y., Gao, Y., Cheng, Y., Zhang, D.: Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "35. Li, S., Chen, X., Liu, Y., Dai, D., Stachniss, C., Gall, J.: Multi-scale interaction for real-time lidar data segmentation on an embedded platform. IEEE Rob. Automat. Letters 7(2), 738-745 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "36. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.77 + ], + "angle": 0, + "content": "37. Lin, T., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "38. Lin, Z., Pathak, D., Wang, Y.X., Ramanan, D., Kong, S.: Continual learning with evolving class ontologies. Adv. Neural Inform. Process. Syst. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment Anything in Lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "40. Liu, Z., Zhang, Z., Cao, Y., Hu, H., Tong, X.: Group-free 3d object detection via transformers. In: Int. Conf. Comput. Vis. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "41. Lu, Y., Jiang, Q., Chen, R., Hou, Y., Zhu, X., Ma, Y.: See more and know more: Zero-shot point cloud segmentation via multi-modal visual data. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "42. Ma, Y., Peri, N., Wei, S., Hua, W., Ramanan, D., Li, Y., Kong, S.: Long-tailed 3d detection via 2d late fusion. arXiv preprint arXiv:2312.10986 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "43. Marcuzzi, R., Nunes, L., Wiesmann, L., Behley, J., Stachniss, C.: Mask-based panoptic lidar segmentation for autonomous driving. IEEE Rob. Automat. Letters 8(2), 1141-1148 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.324 + ], + "angle": 0, + "content": "44. Marcuzzi, R., Nunes, L., Wiesmann, L., Marks, E., Behley, J., Stachniss, C.: Mask4d: End-to-end mask-based 4d panoptic segmentation for lidar sequences. IEEE Rob. Automat. Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "45. Marcuzzi, R., Nunes, L., Wiesmann, L., Vizzo, I., Behley, J., Stachniss, C.: Contrastive instance association for 4d panoptic segmentation using sequences of 3d lidar scans. IEEE Rob. Automat. Letters (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.393 + ], + "angle": 0, + "content": "46. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.42 + ], + "angle": 0, + "content": "47. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: RangeNet++: Fast and Accurate LiDAR Semantic Segmentation. In: Int. Conf. Intel. Rob. Sys. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.447 + ], + "angle": 0, + "content": "48. Miller, D., Nicholson, L., Dayoub, F., Sünderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: Int. Conf. Rob. Automat. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "49. Moosmann, F., Stiller, C.: Joint self-localization and tracking of generic objects in 3d range data. In: Int. Conf. Rob. Automat. (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "50. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: Eur. Conf. Comput. Vis. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "51. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "52. Nunes, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Segcontrast: 3d point cloud feature representation learning through self-supervised segment discrimination. IEEE Rob. Automat. Letters 7(2), 2116-2123 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "53. Osep, A., Voigtlaender, P., Luiten, J., Breuers, S., Leibe, B.: Towards large-scale video video object mining. In: ECCV Workshop on Interactive and Adaptive Learning in an Open World (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.664 + ], + "angle": 0, + "content": "54. Osep, A., Mehner, W., Voigtlaender, P., Leibe, B.: Track, then decide: Category-agnostic vision-based multi-object tracking. In: Int. Conf. Rob. Automat. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "55. Osep, A., Voigtlaender, P., Luiten, J., Breuers, S., Leibe, B.: Large-scale object mining for object discovery from unlabeled video. In: Int. Conf. Rob. Automat. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.745 + ], + "angle": 0, + "content": "56. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.772 + ], + "angle": 0, + "content": "57. Peri, N., Dave, A., Ramanan, D., Kong, S.: Towards long-tailed 3d detection. In: Conf. Rob. Learn. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "58. Peri, N., Li, M., Wilson, B., Wang, Y.X., Hays, J., Ramanan, D.: An empirical analysis of range for 3d object detection. In: ICCV Workshops (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "59. Peri, N., Luiten, J., Li, M., Osep, A., Leal-Taixe, L., Ramanan, D.: Forecasting from lidar via future object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "60. Petrovskaya, A., Thrun, S.: Model based vehicle detection and tracking for autonomous urban driving. Aut. Rob. 26, 123-139 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "61. Pot, E., Toshev, A., Kosecka, J.: Self-supervisory signals for object discovery and detection. arXiv preprint arXiv:1806.03370 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.785, + 0.247 + ], + "angle": 0, + "content": "62. Prest, A., Leistner, C., Civera, J., Schmid, C., Ferrari, V.: Learning object class detectors from weakly annotated video. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.785, + 0.288 + ], + "angle": 0, + "content": "63. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.318 + ], + "angle": 0, + "content": "64. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Adv. Neural Inform. Process. Syst. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.785, + 0.36 + ], + "angle": 0, + "content": "65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.388 + ], + "angle": 0, + "content": "66. Rahman, S., Khan, S.H., Porikli, F.: Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. Asian Conf. Comput. Vis. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.431 + ], + "angle": 0, + "content": "67. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.472 + ], + "angle": 0, + "content": "68. Razani, R., Cheng, R., Li, E., Taghavi, E., Ren, Y., Bingbing, L.: Gp-s3net: Graph-based panoptic sparse semantic segmentation network. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.515 + ], + "angle": 0, + "content": "69. Razani, R., Cheng, R., Taghavi, E., Bingbing, L.: Lite-hdseg: Lidar semantic segmentation using lite harmonic dense convolutions. In: Int. Conf. Rob. Automat. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.558 + ], + "angle": 0, + "content": "70. Sautier, C., Puy, G., Gidaris, S., Boulch, A., Bursuc, A., Marlet, R.: Image-to-lidar self-supervised distillation for autonomous driving data. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.586 + ], + "angle": 0, + "content": "71. Seidenschwarz, J., Osep, A., Ferroni, F., Lucey, S., Leal-Taixe, L.: Semoli: What moves together belongs together. IEEE Conf. Comput. Vis. Pattern Recog. (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.614 + ], + "angle": 0, + "content": "72. Sirohi, K., Mohan, R., Bülicher, D., Burgard, W., Valada, A.: Efficientlps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.657 + ], + "angle": 0, + "content": "73. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.699 + ], + "angle": 0, + "content": "74. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "75. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: Eur. Conf. Comput. Vis. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.77 + ], + "angle": 0, + "content": "76. Teichman, A., Levinson, J., Thrun, S.: Towards 3D object recognition via classification of arbitrary object tracks. In: Int. Conf. Rob. Automat. (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "77. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Int. Conf. Comput. Vis. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "78. Thorpe, C., Herbert, M., Kanade, T., Shafer, S.: Toward autonomous driving: the cmu navlab. i. perception. IEEE expert 6(4), 31-42 (1991)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Segment Anything in Lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "79. Thrun, S., Montemerlo, M., Dahlkamp, H., Stavens, D., Aron, A., Diebel, J., Fong, P., Gale, J., Halpenny, M., Hoffmann, G.: Stanley: The robot that won the darpa grand challenge. Journal of field Robotics (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "80. Wang, Y., Chen, X., You, Y., Li, L., Hariharan, B., Campbell, M., Weinberger, K., Chao, W.: Train in germany, test in the usa: Making 3d object detectors generalize. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "81. Wong, K., Wang, S., Ren, M., Liang, M., Urtasun, R.: Identifying unknown instances for autonomous driving. In: Conference on Robot Learning. pp. 384-393. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "82. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: Int. Conf. Rob. Automat. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.357 + ], + "angle": 0, + "content": "83. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: Int. Conf. Rob. Automat. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.397 + ], + "angle": 0, + "content": "84. Xian, Y., Lampert, C.H., Schiele, B., Akata, Z.: Zero-shot learning - a comprehensive evaluation of the good, the bad and the ugly. IEEE Trans. Pattern Anal. Mach. Intell. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "85. Xiong, X., Munoz, D., Bagnell, J.A., Hebert, M.: 3-D Scene Analysis via Sequenced Predictions over Points and Regions. In: Int. Conf. Rob. Automat. pp. 2609-2616 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.481 + ], + "angle": 0, + "content": "86. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "87. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.55 + ], + "angle": 0, + "content": "88. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.578 + ], + "angle": 0, + "content": "89. Yilmaz, K., Schult, J., Nekrasov, A., Leibe, B.: Mask4d: Mask transformer for 4d panoptic segmentation. arXiv preprint arXiv:2309.16133 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.605 + ], + "angle": 0, + "content": "90. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.633 + ], + "angle": 0, + "content": "91. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.675 + ], + "angle": 0, + "content": "92. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.716 + ], + "angle": 0, + "content": "93. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.744 + ], + "angle": 0, + "content": "94. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: Eur. Conf. Comput. Vis. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.744, + 0.788, + 0.772 + ], + "angle": 0, + "content": "95. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.8 + ], + "angle": 0, + "content": "96. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.841 + ], + "angle": 0, + "content": "97. Zhu, M., Han, S., Cai, H., Borse, S., Ghaffari, M., Porikli, F.: 4d panoptic segmentation as invariant and equivariant field prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.365, + 0.129 + ], + "angle": 0, + "content": "A. Osep et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.19 + ], + "angle": 0, + "content": "98. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] +] \ No newline at end of file diff --git a/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_origin.pdf b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d11e4e20f8520c890a257348a51ca86c4a8ec263 --- /dev/null +++ b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/41b81b1c-0240-4e15-8742-99aeed4b713d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b0b628d2b3aa2d940cf705bf70387e094b8b562204eec6a5a10a8e7b64b22a8 +size 9229795 diff --git a/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/full.md b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fc01f996fe567b57db31bd7ad4ec9f251f85d8db --- /dev/null +++ b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/full.md @@ -0,0 +1,336 @@ +# Better Call SAL: Towards Learning to Segment Anything in Lidar + +Aljosa Osep1* + +Tim Meinhardt $^{1*}$ + +Francesco Ferroni + +Neehar Peri² + +Deva Ramanan + +Laura Leal-Taixe1 + +*Equal contribution + +$^{1}$ NVIDIA $^{2}$ CMU + +![](images/8ff034901fcb4faedefe67f0f01f1518971ef5a39c4d50b3d34aa76493c5910c.jpg) +Class-agnostic instance segmentation + +![](images/6bc6d52bfa739a8ef74b704c25dc5ba34fdd0ae6ed9335ab53532e9a9985f403.jpg) +Text prompts: {car, building,...} + +![](images/20cf5d720d00f4ba55a989460329df6254c01ef76753b7495edf7d090aeb0969.jpg) +Text prompts: {trash bin} +Fig. 1: The SAL (Segment Anything in Lidar) model performs class-agnostic instance segmentation (i) and zero-shot classification via text prompting. This allows us to not only predict panoptic segmentation (ii) for fixed class vocabularies but segment any object (iii and iv) in a given Lidar scan. + +![](images/1512bc5e8da8666c27fc3de232e224f6db5e4c3d742bdc5c88a4d2855cbbd9db.jpg) +Text prompts: {streetcar} + +Abstract. We propose the SAL (Segment Anything in Lidar) method consisting of a text-promptable zero-shot model for segmenting and classifying any object in Lidar, and a pseudo-labeling engine that facilitates model training without manual supervision. While the established paradigm for Lidar Panoptic Segmentation (LPS) relies on manual supervision for a handful of object classes defined a priori, we utilize 2D vision foundation models to generate 3D supervision "for free". Our pseudolabels consist of instance masks and corresponding CLIP tokens, which we lift to Lidar using calibrated multi-modal data. By training our model on these labels, we distill the 2D foundation models into our Lidar SAL model. Even without manual labels, our model achieves $91\%$ in terms of class-agnostic segmentation and $54\%$ in terms of zero-shot LPS of the fully supervised state-of-the-art. Furthermore, we outperform several baselines that do not distill but only lift image features to 3D. More importantly, we demonstrate that SAL supports arbitrary class prompts, can be easily extended to new datasets, and shows significant potential to improve with increasing amounts of self-labeled data. We release all models and the code. + +# 1 Introduction + +We tackle segmentation and recognition of objects in Lidar point clouds, a task commonly tackled via Lidar Panoptic Segmentation (LPS). + +Status quo. LPS has been gaining significant attention in the community due to its role in scene understanding, which is vital for safe autonomous navigation. Thanks to the availability of labeled datasets [6, 18] and advances in learning representations from unordered point sets [12, 63, 64, 77], the Lidar community made significant progress in learning to segment and classify instances of predefined and manually-labeled classes. While this progress has been impressive, existing models [2, 26, 34, 43, 68, 72, 96] cannot adapt to continually evolving class ontologies [38] that may even vary in different geographic regions [80]. + +Stirring the pot. We challenge this well-established approach and investigate how we can train general LPS models that can be prompted to segment point clouds according to any object class vocabulary. Towards such a promptable Lidar segmentation approach, we first need to be able to segment any object, a capability that remains elusive in the Lidar domain. + +Towards segmenting anything in Lidar. To address these challenges, we propose SAL, which consists of a text-promptable zero-shot model (Fig. 2, left) for panoptic segmentation of arbitrary objects (Fig. 1), and pseudo-labeling engine (Fig. 2, right) that facilitates model training directly from raw sensory data. The SAL pseudo-label engine automatically labels Lidar sequences using image segmentation [30] and vision-language models [65]. We utilize SAM [30] to generate class-agnostic masks in images, and CLIP [65] to generate per-mask tokens that connect visual features to language, and finally, transfer both to Lidar using a calibrated sensory setup. Even though generated pseudo-labels only partially cover Lidar scans and are inherently noisy due to errors in the image-level generation process and imperfect sensory calibration, we demonstrate their effectiveness as self-supervised training data for the SAL zero-shot model. In contrast to prior work in LPS, SAL is the first model that can be prompted with any new vocabulary without re-training. Different from recent advances [56] that lift image features to 3D without distillation into a model, SAL does not require image features during inference and segments full $360^{\circ}$ point clouds. + +Talk to your SAL. Our model predicts Lidar segmentation masks and their corresponding CLIP tokens, which allow us to perform zero-shot classification of segmented objects using arbitrary text prompts (Fig. 1). We evaluate SAL by prompting our zero-shot model on standard benchmarks for LPS. For class-agnostic segmentation, we reach $91\%$ of the performance of manually supervised baselines. By classifying segmented objects in a zero-shot manner, we report the first (and very encouraging) results for zero-shot LPS and reach $42\%$ and $54\%$ of the supervised baselines trained on SemanticKITTI and nuScenes, respectively. Beyond that, as shown in Fig. 1, we can prompt SAL model to segment objects outside of existing Lidar dataset class vocabularies. + +![](images/5aeb54162997a1b95043f6c9a6f3d87021c2e9daf025cad3f5bfd29f5191cfa3.jpg) +Fig. 2: SAL overview: Given a Lidar scan and a class vocabulary prompt, specified as a list of per-class free-form text descriptions (left), SAL segments and classifiers objects (things and stuff classes). As labeled data for training such a model does not exist, we supervise SAL by distilling off-the-shelf vision foundation models to Lidar (right). + +![](images/3cc983aafba28d114fb4f567e6d9819b781280699abf68937ee548c90c2382f5.jpg) + +Contributions. As our main contribution, we re-think the established approach to Lidar Panoptic Segmentation and (i) present SAL for segmentation and classification of any object in a Lidar scan. As we do not utilize labeled Lidar data, we propose (ii) a pseudo-label engine that distills vision foundation models to the Lidar domain. The resulting pseudo-labels allow us to train (iii) a zero-shot LPS model without any human supervision. We demonstrate (iv) encouraging results on standard LPS benchmarks and outline a clear path towards universal, promptable segmentation foundation models for Lidar data. + +# 2 Related Work + +Lidar has played a pivotal role since the dawn of embodied navigation [60,78,79]. Recent data-driven efforts in Lidar perception have been pushing boundaries in semantic segmentation [3,12,35,47,69,75,82,83,85,98], object detection [32,40, 42,57-59,88,90,95], and tracking [25,49,76]. + +Lidar panoptic segmentation. Recently, Lidar Panoptic Segmentation [4,6, 7,18] has emerged as a holistic approach to Lidar-based dynamic scene understanding. Prior works [2,4,19,26,28,31,34,43-45,68,89,96,97] learn to group and classify points specified in the training data according to the target class vocabularies, while methods for open-set instance [24,49,76] and panoptic segmentation [81] rely on bottom-up grouping based on Euclidean distance between points. While these developments have been impressive, end-to-end methods remain limited to specific target classes that appear in the training data, while bottom-up, clustering-based approaches [24,27,49,76,81] are sensitive to a particular choice of clustering parameters and are unable to improve their performance in a data-driven fashion. In contrast, the SAL zero-shot model not only learns to segment objects but is also capable of zero-shot classification. + +Self-supervised learning for Lidar. As labeled data in the Lidar domain is scarce, several works investigate how contrastive learning, proven effective in the image domain [10, 23], can reduce the need for labeled data. SegContrast [52] learns to align representations of point clouds and their augmented views, utilizing density-based clustering (DBSCAN [17]) to pool and contrast + +features. Rather than using DBSCAN, [39, 70] leverage image-based methods [1] and foundation models [30] to perform contrastive alignment between image and corresponding Lidar features. The aforementioned methods only provide pretraining recipes for potential segmentation downstream tasks. Our SAL method goes a step further by training a full zero-shot Lidar Panoptic Segmentation model that could benefit from such pre-training recipes. + +Self-supervised (3D) object detection. Several methods (self-)supervise object detectors with (RGB-D) videos [22, 53-55, 61, 62], or Lidar sequences [50, 71, 92] to discover objects in sensory streams. However, such motion-based supervision provides no semantic information. Therefore, [51] distills per-point (PCA quantized) CLIP [65] features to Lidar for zero-shot classification of detected objects. Significantly different from [51], we tackle the more general and challenging task of LPS, which segments and classifies both moving and stationary things and stuff classes. To this end, we distill CLIP features not per point but per object instance. This provides a more holistic object-centric semantic distillation and allows us to supervise our model with non-quantized CLIP features. + +Zero-shot recognition. Zero-shot recognition (ZSR) [84] methods (recently also referred to as "open-vocabulary recognition") tackle the recognition of distinct, yet related, semantic concepts that are either not labeled in the training data (transductive) or not observed at all (inductive). Such methods assume a dataset with labels for some classes, whereas unseen instances are supervised via attributes or class names. The latter is often used in conjunction with word embeddings [46]. By learning to align image-language features, ZSR methods can infer class labels for unseen objects [65]. ZSR has also been explored in the context of object detection, semantic segmentation, and panoptic segmentation [16,86] using word embedding models [5,8,48,66], or CLIP vision-language embeddings [20,21,33,36,67,87,91,93,94]. In the context of semantic segmentation in RGB-D and Lidar data, [41,56] augment image and language embeddings with features extracted from point clouds. Both rely on lifting dense image features from images to 3D during inference, which limits their applicability to setups with dense camera coverage. Similarly, [74] segments instances in accumulated RGB-D point clouds (ScanNet [14]) and relies on manual supervision in conjunction with image-based dense CLIP features for zero-shot classification. Moreover, datasets such as ScanNet [14] were recorded with RGB-D sensors in static environments and fused into 3D reconstructions based on scans, taken from multiple viewpoints. In contrast, SAL directly applies to full Lidar point clouds while requiring no image features at the inference time. + +# 3 SAL: Segment Anything in Lidar + +This section outlines the key challenges and components towards a Lidar model that segments and classifies any object. We first define the underlying problem formation in Sec. 3.1 and then present our (Segmenting Anything in Lidar) SAL model in Sec. 3.2. + +![](images/2c58dbd436e1694c79a318974d52cd36f7a1e35f1f775233fbc35fde74144a5f.jpg) +(a) Pseudo-label engine + +![](images/f37f165dfcb9f08cd9cb6920825715429cf2640b099b3c8d3b7ff2958784bc1a.jpg) +(b) Zero-shot model +Fig. 3: Our pseudo-label engine (Fig. 3a) utilizes SAM [30] to estimate segmentation masks in images, MaskCLIP [16] to estimate corresponding per-mask CLIP features, and a calibrated sensory setup to transfer them to the Lidar domain. We distill these pseudo-labels to our zero-shot model (Fig. 3b), which segments and classifies Lidar point clouds. The SAL zero-shot model employs a sparse-convolutional backbone [12], followed by a Transformer decoder that predicts objectness scores, segmentation masks, and CLIP tokens for each query. To (optionally) perform zero-shot classification, we forward the dataset class vocabulary through the CLIP text encoder and match the encoded vocabulary with predicted CLIP tokens. Our model requires no retraining for different vocabularies and no image features at inference time. + +# 3.1 Zero-Shot Lidar Panoptic Segmentation + +The task. We discuss and evaluate SAL in the context of Lidar Panoptic Segmentation (LPS), which tackles both instance segmentation and object recognition. LPS methods take as input an unordered point cloud $P \in \mathbb{R}^{N \times 4}$ encoding spatial coordinates and their sensor intensity. The currently established problem setting [7,18] assumes supervision and task outputs as per-point semantic class and instance identity labels. The class labels are confined to a pre-defined and fixed class vocabulary $\mathcal{V}$ . In contrast, we tackle LPS in a generalized, zero-shot setting, where the class vocabulary is only provided at inference, not training time. Such a vocabulary $\mathcal{V}$ specifies target classes as a list of $C$ class prompts. Each prompt $c_{i} \in \mathcal{V}$ is specified via free-form text, e.g., with a class name and optional class description. To perform zero-shot LPS a model must be designed to segment and classify any object in a Lidar scan. + +The challenge. Where do we get the data to train such a model? Image-based models [16,21,30,91,93] rely on large and diverse datasets for pre-training [15], dense prediction [13,37], and foundation models that align images with textual descriptions [65] - a commodity not available in the Lidar domain. + +# 3.2 SAL Overview + +Our 2D-to-Lidar distillation method consists of two core components: (i) The pseudo-label engine transfers 2D vision foundation models into Lidar pseudolabels using multi-modal inputs from a calibrated sensory setup, shown in Fig. 3a. (ii) Our zero-shot model is trained on the generated pseudo-labels and is able to perform class-agnostic segmentation and zero-shot classification via text prompts. Given a semantic dataset vocabulary, this allows us to tackle zero-shot LPS as illustrated in Fig. 3b. + +# 3.3 SAL Pseudo-Label Engine + +The SAL pseudo-label engine relies on a calibrated multi-modal sensory setup with $k \geq 1$ RGB cameras. Furthermore, a sufficient overlap between the Lidar sensing area and each camera view is paramount. + +Mask generation: To pseudo-label object instances, we utilize the Segment Anything (SAM) model [30], which generates an overlapping set of segmentation masks. We flatten SAM's output mask hierarchy by non-maxima suppression (NMS) with a minimal overlap threshold to ensure mutually exclusive masks, suppressing object parts and subparts in favor of objects. This way, we obtain a set of non-overlapping binary masks $m_{i}^{k} \in \{0,1\}^{W \times H}$ for each camera view $k$ with image plane of size $W \times H$ . + +CLIP image token generation: As shown by CLIP [65], aligning image and text features allows vision-language foundation models to perform zero-shot image classification based on arbitrary text prompts. To transfer this capability to the Lidar domain, we generate a localized CLIP image feature token $f_{i}^{k} \in \mathbb{R}^{C_{t}}$ for each binary SAM mask $m_{i}^{k}$ . To obtain image tokens for a masked region of the input image, we utilize [16] and their relative mask attention in the CLIP image encoder feature space. Note that we never classify segments in the image domain but merely distill CLIP image feature tokens to Lidar to facilitate zero-shot classification during inference where we do not use any image features. + +Image-to-Lidar unprojection: From the Lidar perspective, we unproject each image mask $m_i^k \in \{0,1\}^{W \times H}$ to a binary Lidar segmentation mask $\tilde{m}_i \in \{0,1\}^N$ by transforming the respective camera coordinate frame to the Lidar space. For datasets with multiple cameras, such as nuScenes [18], we process each image independently, followed by a cross-camera fusion of masks with a sufficient IoU overlap and averaging of their CLIP image tokens. The unprojection yields pairs $\{\tilde{m}_i^k, f_i^k\}$ of Lidar masks and their corresponding CLIP features. + +Refinement via density-based clustering: Image-to-Lidar transformation (Fig. 4a→Fig. 4b) is inherently noisy due to imperfect calibration and issues with synchronization and Lidar rolling shutter. We improve our pseudo-label quality by creating an ensemble of DBSCAN [17] clusters $\tilde{m}_l \in \{0,1\}^N$ , obtained by varying the density threshold to compensate for varying density in Lidar point clouds (details in the appendix). We replace each $\tilde{m}_i$ with its best-matching $\tilde{m}_l$ in case their IoU exceeds a minimal overlap threshold and retain the original mask otherwise to obtain a refined set of pseudo-labels (Fig. 4c) that retain their original cardinality and associated CLIP features. + +# 3.4 SAL Zero-Shot Model + +The universal SAL zero-shot model deconstructs LPS into class-agnostic segmentation and zero-shot classification via text prompts. + +Universal architecture. Instead of relying on highly specialized and engineered LPS models, we base the SAL model on a universal Transformer decoder + +![](images/b54c146f43ff079276c0b5fa2bd9ea445b35766b2aaee6e933c85664775d83ce.jpg) +(a) SAM masks + +![](images/5e738e9d95690a48a4678403483432bed70b632f7484385335f3b48b3778b380.jpg) +(b) SAM unprojected +Fig. 4: Refinement via clustering. After transferring image masks (Fig. 4a) to Lidar (Fig. 4b), we obtain pseudo-labels that suffer from sensory misalignment-related issues. Our geometric refinement (Fig. 4c) improves localization. + +![](images/e948e3f8612409e66dc5ff7741d878924f11a71b688833788f4fba3f7ee0688f.jpg) +(c) SAM + DBSCAN + +architecture (Fig. 3b), similar to [43]. Its 2D counterparts [9, 11] provide top-performing results across semantic, instance, and panoptic segmentation for images. With increasing amounts of (pseudo-)labeled data, Transformer decoders have the potential to achieve a similar impact in the Lidar domain. Following [43], we deploy a Minkowski U-Net [12] backbone for feature extraction followed by a Transformer decoder architecture with object query to point/voxel feature crossattention. Training our model to segment and classify any object and perform in a zero-shot evaluation setting requires a unique design of the final task heads. + +Class-agnostic segmentation. To localize segments in a point cloud, we rely on two model heads that predict segmentation masks and objectness scores for each query. The architecture of the former follows [43] and predicts binary output masks $\hat{m}_i\in \{0,1\} ^N$ by computing the dot product between queries and point features. The objectness head reduces the multi-class problem to a binary object or no-object decision. + +Zero-shot classification. To equip our model with zero-shot classification capabilities without relying on any image input, we learn to predict CLIP [65] Lidar tokens, i.e., features in the CLIP space obtained from Lidar inputs, for each query. The token head consists of a three-layer MLP that directly regresses tokens $\hat{f}_i^j$ . At inference, the predicted tokens are matched to text prompts via the CLIP text encoder. The matching yields a probability distribution over the prompts. To perform zero-shot LPS on a pre-defined dataset, we use its class vocabulary as input text prompts (Fig. 3b). + +We train our model jointly for both class-agnostic segmentation and zero-shot classification with the following loss: + +$$ +\mathcal {L} _ {\mathrm {S A L}} = \mathcal {L} _ {\text {o b j}} + \mathcal {L} _ {\text {s e g}} + \mathcal {L} _ {\text {t o k e n}}, \tag {1} +$$ + +with binary cross-entropy loss $\mathcal{L}_{obj}$ and a cosine distance loss $\mathcal{L}_{token}$ . The segmentation loss $\mathcal{L}_{seg}$ follows [43] and consists of a binary mask cross-entropy and dice loss. The $\mathcal{L}_{\mathrm{SAL}}$ loss is supervised by pseudo-label pairs $\{\tilde{m}_i^k, f_i^k\}$ obtained from our label engine and thereby distills both 2D foundation models (SAM and CLIP) into our LPS model. + +How to train with partial labels? Partially (pseudo-)labeled point clouds present a challenge for SAL model training. For example, in SemanticKITTI, only $14\%$ of points are pseudo-labeled due to low camera coverage (see appendix). If we naively train SAL on partially (pseudo-)labeled point clouds, the objectness + +loss $\mathcal{L}_{obj}$ would penalize any segmentation in these regions as a false positive, thereby teaching the model to ignore these points entirely. Empirically, we determined that the most effective training strategy is to (i) remove unlabeled points from point clouds, (ii) utilize standard data augmentations (rotations, flipping, scaling, and translations), in conjunction with proposed FrankenFruustum augmentation (detailed in the appendix) that mimics fully-labeled point clouds during training by randomly removing unlabeled points and replicating labeled frustum regions around the vertical axis. This augmentation does not increase the overall label coverage, but our ablations in Sec. 4.2 show its effectiveness in reducing the domain gap between training and inference input clouds. + +# 4 Experiments + +In the following section, we outline our experimental setup (Sec. 4.1) used to ablate the SAL model and its training strategies (Sec. 4.2). Since SAL presents the first Zero-Shot Lidar Panoptic Segmentation (ZS-LPS) model, we design and compare with multiple hand-crafted baselines that implement concepts from related works. We conclude our analysis by comparing our zero-shot performance (Sec. 4.4) to fully-supervised LPS methods. + +# 4.1 Experimental Setup + +Datasets. We evaluate SAL on two public Lidar Panoptic Segmentation (LPS) datasets, SemanticKITTI [6, 7] and nuScenes [18]. We utilize provided, human-annotated ground-truth (GT) panoptic segmentation labels solely for evaluation purposes, linear probing ablations, and baselines. We train SAL model only using our self-supervised pseudo-labels (Sec. 3.3). To demonstrate the general applicability of SAL to other domains and datasets, we show qualitative results of SAL trained on Waymo Open [73] dataset, which contains no GT LPS labels. + +Metrics. For the evaluation, we utilize the standard Panoptic Segmentation [29] metrics. The Panoptic Quality $PQ = RQ \times SQ$ combines Segmentation Quality (SQ) and Recognition Quality (RQ), i.e., F-1 score. True positives (TPs) are mask predictions with sufficient intersection-over-union (IoU) overlap with GT masks of the same class. + +Semantic Oracle (SO). The aforementioned metrics assume GT labels and predictions with instance IDs and semantic classes for each point. Therefore, they are not suitable for evaluating class-agnostic segmentation. To assess the class-agnostic segmentation independently of semantics, we apply SO during the evaluation by assigning ground truth semantic classes $c_{i}$ to predicted masks $m_{i}$ via majority voting. Evaluation with SO is only needed for evaluation of class-agnostic segmentation - our zero-shot models provide both instance and semantic class predictions. + +Stuff Merging (SM). SO allows us to assess class-agnostic instance segmentation for things classes. However, this approach is unsuitable for assessing + +the performance of stuff classes, as existing Lidar datasets merge instances of stuff classes into a single instance. The things- stuff separation can be unintuitive and inconsistent across datasets (e.g., SemanticKITTI merges instances of traffic signs into a single stuff class). In contrast, we fully embrace the philosophy that all classes can be segmented into individual instances (e.g., segmentation models should localize individual trees/bushes in the vegetation class or distinguish individual buildings). To evaluate our models on the target datasets in the SO regime on Lidar datasets that do not provide instance-level annotations for all classes, we additionally report results using a merge strategy (SM), which merges all instances of a particular stuff class into a single mask to ensure outputs of our models are consistent with the format of target datasets. + +# 4.2 SAL Ablations + +In this section, we discuss how to train SAL zero-shot segmentation model using self-generated pseudo-labels (Tab. 1) and discuss design decisions behind our instance (Tab. 2) and semantic distillation (Tab. 3). + +Learning with partial labels. For the single-camera setup [6,7], our pseudolabels only cover $14\%$ of the input point clouds (see appendix). Training on partial label coverage presents a challenge for our model. In Tab. 1, we only evaluate class-agnostic segmentation by applying SO and SM (see Sec. 4.1). We compare SAL models trained using pseudo and GT labels. The Frustum Filter removes all unlabeled points not visible in the camera frustum during training and/or evaluation. As shown in Tab. 1, removing unlabeled points from pseudo-labeled point clouds during training has a significant benefit on the performance ( $3^{rd}$ row, 59.3 PQ) as compared to the variant, where we merely ignore unlabeled region during the training ( $2^{nd}$ row, 22.2 PQ). This is likely due to data imbalance in a single-camera setup, leading to a class imbalance between labeled/unlabeled points, incentivizing SAL to suppress predictions. + +While utilizing standard augmentations (rotations, flipping, scaling, and translations) is crucial, we observe that performing our FrankenFrustum augmentation, which concatenates the visible portion of the point cloud around the z-axis (see Sec. 3.4), significantly improves results ( $4^{th}$ row, 62.5 PQ). Mixing point clouds from different scans ( $5^{th}$ row, 62.8 PQ) further boosts results. Overall, we obtain 69 PQ from GT supervision (1st row) and 62.8 PQ when using our pseudo-labels (5th row). Remarkably, our labels yield $91\%$ of the GT performance while covering only $14\%$ of the input point cloud (see appendix). To further contextualize these results, we note that this evaluation is performed only on classes for which GT instance labels are available. As shown in Fig. 1, our model learns to segment a much larger variety of classes. Evaluating only the subset of points visible in the camera ( $14\%$ of all points, rows 6&7), the gap between the GT-supervised model (71.8 PQ) and pseudo-supervised model (70.7 PQ) shrinks further. We conclude that by training SAL model using pseudo-labels, we distill the notion of objectness from SAM [30] into our model and obtain segmentation capabilities, similar to the model trained on GT data. + +Table 1: Class-agnostic segmentation. By cropping unlabeled points and performing data augmentations (in combination with our FrankenFrostum), SAL successfully learns to segment full point clouds even when only $14\%$ of points are (pseudo)-labeled. + +
LabelsFrust.-Filter TrainEvalFranken Frust.PQSQ\( PQ^{Th} \)\( PQ^{St} \)
GT69.083.581.659.7
Pseu.22.267.545.15.5
Pseu.✓59.378.265.854.5
Pseu.✓✓62.579.167.658.8
Pseu.✓✓(mix)62.879.069.058.3
GT✓71.884.884.662.5
Pseu.✓✓✓(mix)70.781.975.467.3
+ +Table 2: Scaling. In contrast to predefined GT data, increasing the number of queries SAL improves performance on stuff classes. By increasing the amount of labeled data, we further improve segmentation performance. + +
TrainNum.PQSQ\( PQ^{Th} \)\( PQ^{St} \)
LabelsSetQuer.
GTtrain10067.882.479.759.2
GTtrain20067.983.679.859.3
GTtrain30069.083.581.659.7
Pseu.train10053.975.362.247.8
Pseu.train20060.877.066.756.5
Pseu.train30062.879.069.058.3
Pseu.bigtrain30065.379.571.960.5
+ +Scaling queries and data. Analogous to the grid inference of SAM, which operates with hundreds of spatial queries (prompts), we analyze the effect of different numbers of decoder queries. We subsample instances during training to train models with fewer queries than the maximum number of segments per scan. As visualized in Tab. 2, increasing the number of decoder queries improves the recognition of things for both training with pseudo and GT labels. However, we observe a significant performance boost with pseudo labels for stuff classes $(58.3\mathrm{PQ}^{St}, +10.8)$ , while the improvement for GT labels is marginal $(59.7\mathrm{PQ}^{St}, +0.5)$ . Our model benefits from a larger number of queries for stuff classes as it learns a fine-grained segmentation model that is not limited to a prefixed set of things classes. In the last row, we concatenate pseudo-labeled train and test sets (neither used for the validation) into a bigtrain set and observe an improvement of $+2.5$ PQ, suggesting that scaling the amount of training data has the potential to improve the performance of our models further. + +Table 3: Semantic distillation. We linearly probe the SAL model, trained with and without semantic distillation loss $\mathcal{L}_{token}$ . We train a linear classifier with GT labels while keeping backbone and decoder features frozen. As can be seen, $\mathcal{L}_{token}$ successfully distills a notion of semantics into our model. + +
Linear prob. LtokenPQRQSQmIoU
×20.026.355.223.4
××33.141.968.340.0
×24.832.366.829.7
+ +Semantic distillation. At inference, SAL predicts a binary segmentation mask and objectness score for each query. To perform zero-shot classification, our model additionally predicts CLIP image tokens. This allows us to not rely on image features during inference but prompt our predicted tokens with arbitrary texts, i.e., object classes. Does the token head loss $\mathcal{L}_{token}$ successfully distill a notion of object semantics from CLIP into our model? To verify this, we perform linear probing experiments: we replace the token distillation head of a trained model with a linear classifier (backbone and instance de + +coder remain frozen) and train it using GT labels. As can be seen in Tab. 3, distilling CLIP features into our model significantly improves linear probing + +Table 4: Zero-shot panoptic segmentation. We utilize prior efforts in the image domain [30] and Lidar [56] domain to craft multiple baselines that only unproject segmentation masks and lift image features to Lidar. By contrast, SAL distills outputs of such baselines (pseudo-labels) into a stronger Lidar segmentation model. With ImageFeat, we denote methods that require image features at inference time, and Frust. Eval. denotes the evaluation of a subset of points visible in the camera. + +
MethodFrost. ImageDefault classesSuper classes
EvalFeat.PQSQPQThPQStmIoUPQSQPQThPQStmIoU
Class-agnostic Segmentation (Semantic Oracle)
SAM✓✓46.072.149.743.4------
SAM+Erosion✓✓42.269.445.639.6------
SAM+DBS (filter)✓✓46.770.376.824.8------
SAM+DBS (replace)✓✓48.773.753.145.4------
SAL✓X70.781.975.467.3------
Zero-Shot Lidar Panoptic Segmentation
SAM+DBS+CLIP✓✓27.571.531.724.530.651.177.571.241.054.3
SAL✓X33.171.422.840.533.563.984.288.351.766.4
SAM+DBS+CLIPXX8.256.418.60.67.511.547.60.017.311.2
SALXX24.866.817.430.228.748.578.880.432.652.8
+ +performance (33.1 PQ, 40 mIoU) compared to the baseline, not trained with the semantic distillation loss $\mathcal{L}_{token}$ (20.0 PQ, 23.4 mIoU). This confirms that by distilling CLIP features, we inject a notion of semantics into our model, even though our model is not explicitly supervised with any labels containing semantic information - the notion of semantics is learned implicitly via CLIP feature distillation. Without this loss, our zero-shot model (24.8 PQ, 29.7 mIoU) outperforms the linearly probed model tuned using GT semantic labels. + +# 4.3 Class-Agnostic and Zero-Shot Lidar Panoptic Segmentation + +As the first study tackling Zero-Shot Lidar Panoptic Segmentation (ZS-LPS), we devise several strong baselines inspired by prior works in RGB-D semantic segmentation. During inference, these methods utilize image-based models and lift image features to 3D [56]. To ensure these are evaluated fairly, we report results in Tab. 4 on the subset of points visible in at least one camera (Frost. Eval.). Our SAL model does not have this limitation. Therefore, we additionally report results evaluated on full point clouds in the bottom of Tab. 4. + +Class-agnostic segmentation. Our first baseline (SAM [30]) generates masks in images and unprojects them to Lidar (see Sec. 3.3), and leads to 46 PQ. We observe bleeding edges (see Fig. 4b) after unprojecting from the cameras to Lidar caused by imperfect calibrations. To mitigate these artifacts, we experiment with slightly eroding SAM predictions in the image domain. However, this leads to a performance drop (42.2 PQ), likely because smaller segments after erosion fail to pass $>0.5$ overlap with labeled instances. + +Density-based filtering: Alternatively, to mitigate unprojection errors, we experiment with density-based clustering methods. We generate a large pool of DBSCAN (DBS) clusters and test two strategies to improve our SAM-based segmentation labels: (i) filtering segments without a sufficient overlap with any DBSCAN cluster or (ii) replacing segments with large overlaps. + +SAM+DBS (filter): Intuitively, the filter approach removed Lidar segments only segmentable in the image domain, e.g., a SAM mask that segments a shadow on a flat wall. While this strategy improves precision for things classes (76.8 PQTh, +27.1), it significantly degrades performance on stuff classes (24.8 PQSt -18.6) and overall degrades PQ to 46.7. + +SAM+DBS (replace): The replace strategy (that replaces SAM masks with DBSCAN clusters with sufficient mutual overlap), on the other hand, improves both things and stuff classes to an overall PQ of 48.7 (+3.9) and SQ 73.7 (+1.1), respectively. This improvement can be visually verified in Fig. 4c. + +SAL: We train our SAL model using top-performing $SAM+DBS$ (replace) and observe a significant improvement (70.7 PQ) in terms of both things and stuff classes. Even though SAL obtains significantly higher PQ than hand-crafted baselines, it also improves in terms of SQ (81.9 vs. 73.7 closest competitor), suggesting that the distilled model is robust to projection artifacts. + +Do we need a vision segmentation foundation model? To generate a set with a sufficiently high recall, we need to cluster points with varying DBSCAN radius parameters (details in the appendix). This leads to a large set of predicted masks per scan (5,413 avg.). While most of the correct segments are in this set, separating the signal from the noise is difficult. Prior works, such as [50], are limited to moving thing classes as they rely on motion cues to filter static point cloud regions and DBSCAN to group remaining points. Our core insight is that we can utilize vision foundation models that already learned a notion of objectness: SAM generates only 171 masks per image, reduced to 45 via flattening, and to 39 masks after the unprojection, achieving a high recall with only a few instances, or in other words, a high signal-to-noise ratio. + +Zero-shot segmentation. So far, we discussed class-agnostic segmentation. To obtain semantic class predictions per mask in a zero-shot fashion, we devise a simple baseline inspired by [56]. Once masks are unprojected from the image to Lidar space, we utilize associated CLIP features (directly extracted from images using [16]) to perform zero-shot classification via dot product between encoded class prompts and (lifted) CLIP features [65]. This baseline (SAM+DBS+CLIP) can be understood as an approach that directly lifts image features from state-of-the-art image-based models [16] to Lidar. We detail prompts in the appendix. + +There is a significant performance drop between $SAM+DBS$ (replace) baseline, evaluated with Semantic Oracle ( $5^{th}$ row, 48.7 PQ), and zero-shot baseline ( $SAM+DBS$ (replace)+CLIP), $7^{th}$ row, 27.5 PQ. This is not surprising, as zero-shot segmentation is a challenging problem in both image [16] and Lidar domains. To this end, even when transferring features of state-of-the-art image-based models to Lidar we observe a significant performance drop. Remarkably, + +Table 5: Lidar Panoptic Segmentation (LPS) on SemanticKITTI and nuScenes validation sets. We prompt our zero-shot SAL model with the respective class vocabularies and compare its performance to fully-supervised baselines. On SemanticKITTI and nuScenes, we reach $42\%$ and $54\%$ of the fully-supervised model, respectively. This gap reduces significantly when we evaluate super classes. + +
MethodSupervisionDefault classesSuper classes
PQRQSQ\( PQ^{Th} \)\( PQ^{St} \)mIoUPQRQSQmIoU
SemanticKITT1DS-Net [26]Full57.768.077.661.854.863.5---
PolarSeg [96]Full59.170.278.365.754.364.5---
EfficientLPS [72]Full59.269.875.058.060.964.9---
GP-S3Net [68]Full63.375.981.470.258.373.0---
MaskPLS [43]Full59.869.076.3---78.487.188.2
SALFull59.569.275.762.357.463.881.790.089.2
SALZero-shot24.832.366.817.430.228.748.559.478.8
mScenesPHNet [34]Full74.784.288.274.075.979.7---
DS-Net [26]Full51.259.086.138.472.373.5---
GP-S3Net [68]Full61.072.084.156.066.075.8---
EfficientLPS [72]Full62.073.983.456.870.665.6---
PolarSeg [96]Full63.475.383.959.270.466.9---
MaskPLS [43]Full57.766.071.864.452.262.571.581.086.2
SALFull70.580.885.979.461.772.874.282.787.1
SALZero-shot38.447.877.247.529.233.952.663.577.3
+ +SAL distills such noisy image-based features to a stronger model ( $8^{th}$ row, 33.1 PQ, +5.6 PQ). Finally, to emphasize the inherent limitation of baselines, relying on image features, we evaluate on full $360^{\circ}$ point clouds. SAL performs similarly in both settings. The baseline, on the other hand, relies on image features and is, therefore, unable to predict labels out of the camera frustum. This problem can be mitigated by utilizing a setup with denser camera coverage. + +Evaluation on super-classes. We observe that the CLIP classifier often confuses related classes, e.g., car and other vehicle. We additionally evaluate our semantics by prompting with super-classes (as defined in [6]). This alone increases PQ to 63.9 (48.5 when evaluating on full point clouds) and suggests significant potential for improving prompting and text-to-image alignment. + +# 4.4 Lidar Panoptic Segmentation Evaluation + +Finally, we compare SAL with state-of-the-art LPS methods on standard benchmarks [6,18]. We focus this discussion on comparing SAL, trained with labeled data, and our SAL zero-shot model that specifies classes in benchmarks [6,18] as text prompts. Importantly, zero-shot results are obtained without training on any labeled data. In contrast, all baselines are trained using human-labeled data, and are constrained to their respective, pre-defined class vocabularies. + +SemanticKITTI. When evaluating with the default class vocabulary (Tab. 5), our zero-shot model reaches 24.8 PQ or $42\%$ of the fully-supervised model (59.5 PQ). As discussed in Sec. 4.2, our method is mainly limited by the quality of the generated CLIP features. In practice, our zero-shot semantic classifier mainly confuses objects within super classes, e.g., car vs. other-vehicle. Therefore, + +without any retraining, we prompt our zero-shot model with super classes and obtain 48.5 PQ, $59\%$ of the supervised model (81.7 PQ), + +nuScenes. We report our result in Tab. 5 and reach similar conclusions. SAL reaches 38.4 PQ on default classes, $54\%$ of the supervised model (70.5 PQ). As with SemanticKITTI, results consistently improve when evaluated with superclasses. In this setting, zero-shot model yields 52.6 PQ. Note that SAL model, supervised with labeled data, significantly outperforms MaskPLS. + +Towards closing the gap. Compared to SemanticKITTI (24.8 PQ, $42\%$ ), we observe a significantly smaller gap between the model trained on pseudo- and ground truth labels for nuScenes (38.4 PQ, $54\%$ ). Despite its sparse Lidar signal and more diverse input scenarios, SAL performs better on this challenging Lidar dataset. We attribute this to a larger pseudo-label coverage of nuScenes ( $48\%$ , compared to $14\%$ coverage on SemanticKITTI), and, therefore, effectively more labeled data, which is also consistent with our observations on SemanticKITTI w.r.t. increasing the amount of labeled data (Tab. 2). This trend suggests a clear path forward: applying SAL to pseudo-label more data with zero annotation cost. + +Beyond labeled datasets. Datasets, such as SemanticKITTI and nuScenes, provide dense, point-level panoptic segmentation labels, which we use for analysis and the evaluation of our method on a small set of classes that were labeled in the respective dataset (14/11 things/stuff in SemanticKITTI, respectively, and 23/6 in nuScenes). To demonstrate the versatility of SAL, we additionally pseudolabel and train a model on Waymo Open [73] dataset and show qualitative results in Fig. 1 and more in the appendix. This effort takes less than a week of pseudolabeling and model training efforts (we provide details in the appendix). + +The good, the (breaking) bad, and the ugly. While we are thrilled about class-agnostic segmentation performance, our experimental analysis reveals a substantial gap between full supervision and our zero-shot model. Finally, at the moment, SAL is limited to training on specific datasets. As in fully-supervised Lidar perception, cross-sensor generalization remains a challenge for SAL. This could be addressed by utilizing temporal context, data scaling efforts and investigating sensor-agnostic backbone networks. + +# 5 Conclusions + +We proposed SAL, a method for Zero-Shot Lidar Panoptic Segmentation. The key SAL components are a pseudo-label engine that utilizes vision foundation models and a zero-shot model trained via self-supervision. While SAL is the first step in the direction of learning general Lidar segmentation models by distilling image-based foundation to Lidar, we believe we just scratched the surface and opened the door for training Lidar segmentation models without manual supervision. + +Acknowledgments. This project was funded, in parts, by ERC Starting Grant DynAI (ERC-101043189). We are grateful to Žan Gojčić, Guillem Braso, Cristiano Saltori, Sérgio Agostinho, and Jonas Schult for their feedback on the paper and their insightful comments. Special thanks to Maxim Maximov for his help on figures. + +# References + +1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell. 34(11), 2274-2282 (2012) +2. Agarwalla, A., Huang, X., Ziglar, J., Ferroni, F., Laura, L.T., Hays, J., Osep, A., Ramanan, D.: Lidar panoptic segmentation and tracking without bells and whistles. In: Int. Conf. Intel. Rob. Sys. (2023) +3. Aksoy, E.E., Baci, S., Cavidar, S.: Salsanet: Fast road and vehicle segmentation in lidar point clouds for autonomous driving. In: Intel. Veh. Symp. (2020) +4. Aygiun, M., Osep, A., Weber, M., Maximov, M., Stachniss, C., Behley, J., Leal-Taixe, L.: 4d panoptic lidar segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +5. Bansal, A., Sikka, K., Sharma, G., Chellappa, R., Divakaran, A.: Zero-shot object detection. In: Eur. Conf. Comput. Vis. (2018) +6. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In: Int. Conf. Comput. Vis. (2019) +7. Behley, J., Milioto, A., Stachniss, C.: A Benchmark for LiDAR-based Panoptic Segmentation based on KITTI. In: Int. Conf. Rob. Automat. (2021) +8. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. Adv. Neural Inform. Process. Syst. (2019) +9. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. (2020) +0. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Adv. Neural Inform. Process. Syst. (2020) +1. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +2. Choy, C., Gwak, J., Savarese, S.: 4D spatio-temporal convnets: Minkowski convolutional neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) +3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016) +4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017) +5. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. (2009) +6. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip. In: Int. Conf. Mach. Learn. (2023) +7. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Rob. Sci. Sys. (1996) +8. Fong, W.K., Mohan, R., Hurtado, J.V., Zhou, L., Caesar, H., Beijbom, O., Valada, A.: Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Rob. Automat. Letters (2021) +9. Gasperini, S., Mahani, M.A.N., Marcos-Ramiro, A., Navab, N., Tombari, F.: Panost-ter: End-to-end panoptic segmentation of lidar point clouds. IEEE Rob. Automat. Letters (2021) + +20. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Eur. Conf. Comput. Vis. (2022) +21. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021) +22. Harley, A.W., Zuo, Y., Wen, J., Mangal, A., Potdar, S., Chaudhry, R., Fragkiadaki, K.: Track, check, repeat: An em approach to unsupervised tracking. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +24. Held, D., Guillory, D., Rebsamen, B., Thrun, S., Savarese, S.: A probabilistic framework for real-time 3d segmentation using spatial, temporal, and semantic cues. In: Rob. Sci. Sys. (2016) +25. Held, D., Levinson, J., Thrun, S., Savarese, S.: Combining 3d shape, color, and motion for robust anytime tracking. In: Rob. Sci. Sys. (2014) +26. Hong, F., Zhou, H., Zhu, X., Li, H., Liu, Z.: Lidar-based panoptic segmentation via dynamic shifting network. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +27. Hu, P., Held, D., Ramanan, D.: Learning to optimally segment point clouds. IEEE Robotics and Automation Letters 5(2), 875-882 (2020) +28. Hurtado, J.V., Mohan, R., Valada, A.: Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189 (2020) +29. Kirillov, A., He, K., Girshick, R.B., Rother, C., Dollar, P.: Panoptic segmentation. IEEE Conf. Comput. Vis. Pattern Recog. (2018) +30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Int. Conf. Comput. Vis. (2023) +31. Kreuzberg, L., Zulfikar, I.E., Mahadevan, S., Engelmann, F., Leibe, B.: 4d-stop: Panoptic segmentation of 4d lidar using spatio-temporal object proposal generation and aggregation. In: ECCV AVVision Workshop (2022) +32. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) +33. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: Int. Conf. Learn. Represent. (2022) +34. Li, J., He, X., Wen, Y., Gao, Y., Cheng, Y., Zhang, D.: Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +35. Li, S., Chen, X., Liu, Y., Dai, D., Stachniss, C., Gall, J.: Multi-scale interaction for real-time lidar data segmentation on an embedded platform. IEEE Rob. Automat. Letters 7(2), 738-745 (2021) +36. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +37. Lin, T., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. (2014) +38. Lin, Z., Pathak, D., Wang, Y.X., Ramanan, D., Kong, S.: Continual learning with evolving class ontologies. Adv. Neural Inform. Process. Syst. (2022) +39. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023) + +40. Liu, Z., Zhang, Z., Cao, Y., Hu, H., Tong, X.: Group-free 3d object detection via transformers. In: Int. Conf. Comput. Vis. (2021) +41. Lu, Y., Jiang, Q., Chen, R., Hou, Y., Zhu, X., Ma, Y.: See more and know more: Zero-shot point cloud segmentation via multi-modal visual data. In: Int. Conf. Comput. Vis. (2023) +42. Ma, Y., Peri, N., Wei, S., Hua, W., Ramanan, D., Li, Y., Kong, S.: Long-tailed 3d detection via 2d late fusion. arXiv preprint arXiv:2312.10986 (2023) +43. Marcuzzi, R., Nunes, L., Wiesmann, L., Behley, J., Stachniss, C.: Mask-based panoptic lidar segmentation for autonomous driving. IEEE Rob. Automat. Letters 8(2), 1141-1148 (2023) +44. Marcuzzi, R., Nunes, L., Wiesmann, L., Marks, E., Behley, J., Stachniss, C.: Mask4d: End-to-end mask-based 4d panoptic segmentation for lidar sequences. IEEE Rob. Automat. Letters (2023) +45. Marcuzzi, R., Nunes, L., Wiesmann, L., Vizzo, I., Behley, J., Stachniss, C.: Contrastive instance association for 4d panoptic segmentation using sequences of 3d lidar scans. IEEE Rob. Automat. Letters (2022) +46. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013) +47. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: RangeNet++: Fast and Accurate LiDAR Semantic Segmentation. In: Int. Conf. Intel. Rob. Sys. (2019) +48. Miller, D., Nicholson, L., Dayoub, F., Sünderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: Int. Conf. Rob. Automat. (2018) +49. Moosmann, F., Stiller, C.: Joint self-localization and tracking of generic objects in 3d range data. In: Int. Conf. Rob. Automat. (2013) +50. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: Eur. Conf. Comput. Vis. (2022) +51. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: Int. Conf. Comput. Vis. (2023) +52. Nunes, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Segcontrast: 3d point cloud feature representation learning through self-supervised segment discrimination. IEEE Rob. Automat. Letters 7(2), 2116-2123 (2022) +53. Osep, A., Voigtlaender, P., Luiten, J., Breuers, S., Leibe, B.: Towards large-scale video video object mining. In: ECCV Workshop on Interactive and Adaptive Learning in an Open World (2018) +54. Osep, A., Mehner, W., Voigtlaender, P., Leibe, B.: Track, then decide: Category-agnostic vision-based multi-object tracking. In: Int. Conf. Rob. Automat. (2018) +55. Osep, A., Voigtlaender, P., Luiten, J., Breuers, S., Leibe, B.: Large-scale object mining for object discovery from unlabeled video. In: Int. Conf. Rob. Automat. (2019) +56. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +57. Peri, N., Dave, A., Ramanan, D., Kong, S.: Towards long-tailed 3d detection. In: Conf. Rob. Learn. (2023) +58. Peri, N., Li, M., Wilson, B., Wang, Y.X., Hays, J., Ramanan, D.: An empirical analysis of range for 3d object detection. In: ICCV Workshops (2023) +59. Peri, N., Luiten, J., Li, M., Osep, A., Leal-Taixe, L., Ramanan, D.: Forecasting from lidar via future object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) + +60. Petrovskaya, A., Thrun, S.: Model based vehicle detection and tracking for autonomous urban driving. Aut. Rob. 26, 123-139 (2009) +61. Pot, E., Toshev, A., Kosecka, J.: Self-supervisory signals for object discovery and detection. arXiv preprint arXiv:1806.03370 (2018) +62. Prest, A., Leistner, C., Civera, J., Schmid, C., Ferrari, V.: Learning object class detectors from weakly annotated video. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012) +63. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017) +64. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Adv. Neural Inform. Process. Syst. (2017) +65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021) +66. Rahman, S., Khan, S.H., Porikli, F.: Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. Asian Conf. Comput. Vis. (2018) +67. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +68. Razani, R., Cheng, R., Li, E., Taghavi, E., Ren, Y., Bingbing, L.: Gp-s3net: Graph-based panoptic sparse semantic segmentation network. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +69. Razani, R., Cheng, R., Taghavi, E., Bingbing, L.: Lite-hdseg: Lidar semantic segmentation using lite harmonic dense convolutions. In: Int. Conf. Rob. Automat. (2021) +70. Sautier, C., Puy, G., Gidaris, S., Boulch, A., Bursuc, A., Marlet, R.: Image-to-lidar self-supervised distillation for autonomous driving data. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +71. Seidenschwarz, J., Osep, A., Ferroni, F., Lucey, S., Leal-Taixe, L.: Semoli: What moves together belongs together. IEEE Conf. Comput. Vis. Pattern Recog. (2024) +72. Sirohi, K., Mohan, R., Bülicher, D., Burgard, W., Valada, A.: Efficientlps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics (2021) +73. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +74. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023) +75. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: Eur. Conf. Comput. Vis. (2020) +76. Teichman, A., Levinson, J., Thrun, S.: Towards 3D object recognition via classification of arbitrary object tracks. In: Int. Conf. Rob. Automat. (2011) +77. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Int. Conf. Comput. Vis. (2019) +78. Thorpe, C., Herbert, M., Kanade, T., Shafer, S.: Toward autonomous driving: the cmu navlab. i. perception. IEEE expert 6(4), 31-42 (1991) + +79. Thrun, S., Montemerlo, M., Dahlkamp, H., Stavens, D., Aron, A., Diebel, J., Fong, P., Gale, J., Halpenny, M., Hoffmann, G.: Stanley: The robot that won the darpa grand challenge. Journal of field Robotics (2006) +80. Wang, Y., Chen, X., You, Y., Li, L., Hariharan, B., Campbell, M., Weinberger, K., Chao, W.: Train in germany, test in the usa: Making 3d object detectors generalize. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +81. Wong, K., Wang, S., Ren, M., Liang, M., Urtasun, R.: Identifying unknown instances for autonomous driving. In: Conference on Robot Learning. pp. 384-393. PMLR (2020) +82. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: Int. Conf. Rob. Automat. (2018) +83. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: Int. Conf. Rob. Automat. (2019) +84. Xian, Y., Lampert, C.H., Schiele, B., Akata, Z.: Zero-shot learning - a comprehensive evaluation of the good, the bad and the ugly. IEEE Trans. Pattern Anal. Mach. Intell. (2018) +85. Xiong, X., Munoz, D., Bagnell, J.A., Hebert, M.: 3-D Scene Analysis via Sequenced Predictions over Points and Regions. In: Int. Conf. Rob. Automat. pp. 2609-2616 (2011) +86. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +87. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +88. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018) +89. Yilmaz, K., Schult, J., Nekrasov, A., Leibe, B.: Mask4d: Mask transformer for 4d panoptic segmentation. arXiv preprint arXiv:2309.16133 (2023) +90. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +91. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +92. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +93. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +94. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: Eur. Conf. Comput. Vis. (2022) +95. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) +96. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +97. Zhu, M., Han, S., Cai, H., Borse, S., Ghaffari, M., Porikli, F.: 4d panoptic segmentation as invariant and equivariant field prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) + +98. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) \ No newline at end of file diff --git a/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/images.zip b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3bcb993c02a42bd8a0e135a973e3f63e77d5e05d --- /dev/null +++ b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43e24fe2636213d1bfaf008008e9f5d4316636acefcd91fa02ba45db63f12688 +size 409095 diff --git a/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/layout.json b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f597a74b9ae0b66be96a3bf32b5f244bd8e2b831 --- /dev/null +++ b/2024/Better Call SAL_ Towards Learning to Segment Anything in Lidar/layout.json @@ -0,0 +1,10407 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "spans": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "text", + "content": "Better Call SAL: Towards Learning to Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 168, + 211, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 168, + 211, + 182 + ], + "spans": [ + { + "bbox": [ + 149, + 168, + 211, + 182 + ], + "type": "text", + "content": "Aljosa Osep1*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 222, + 169, + 300, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 169, + 300, + 181 + ], + "spans": [ + { + "bbox": [ + 222, + 169, + 300, + 181 + ], + "type": "text", + "content": "Tim Meinhardt" + }, + { + "bbox": [ + 222, + 169, + 300, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 312, + 169, + 394, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 169, + 394, + 181 + ], + "spans": [ + { + "bbox": [ + 312, + 169, + 394, + 181 + ], + "type": "text", + "content": "Francesco Ferroni" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 406, + 169, + 465, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 169, + 465, + 181 + ], + "spans": [ + { + "bbox": [ + 406, + 169, + 465, + 181 + ], + "type": "text", + "content": "Neehar Peri²" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 182, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 182, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 222, + 182, + 296, + 193 + ], + "type": "text", + "content": "Deva Ramanan" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 182, + 387, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 182, + 387, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 387, + 193 + ], + "type": "text", + "content": "Laura Leal-Taixe1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 266, + 198, + 348, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 198, + 348, + 209 + ], + "spans": [ + { + "bbox": [ + 266, + 198, + 348, + 209 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 262, + 217, + 351, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 217, + 351, + 228 + ], + "spans": [ + { + "bbox": [ + 262, + 217, + 351, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 262, + 217, + 351, + 228 + ], + "type": "text", + "content": "NVIDIA " + }, + { + "bbox": [ + 262, + 217, + 351, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 262, + 217, + 351, + 228 + ], + "type": "text", + "content": "CMU" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 137, + 264, + 218, + 343 + ], + "blocks": [ + { + "bbox": [ + 137, + 264, + 218, + 343 + ], + "lines": [ + { + "bbox": [ + 137, + 264, + 218, + 343 + ], + "spans": [ + { + "bbox": [ + 137, + 264, + 218, + 343 + ], + "type": "image", + "image_path": "8ff034901fcb4faedefe67f0f01f1518971ef5a39c4d50b3d34aa76493c5910c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 345, + 217, + 361 + ], + "lines": [ + { + "bbox": [ + 138, + 345, + 217, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 345, + 217, + 361 + ], + "type": "text", + "content": "Class-agnostic instance segmentation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 222, + 264, + 304, + 344 + ], + "blocks": [ + { + "bbox": [ + 222, + 264, + 304, + 344 + ], + "lines": [ + { + "bbox": [ + 222, + 264, + 304, + 344 + ], + "spans": [ + { + "bbox": [ + 222, + 264, + 304, + 344 + ], + "type": "image", + "image_path": "6bc6d52bfa739a8ef74b704c25dc5ba34fdd0ae6ed9335ab53532e9a9985f403.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 345, + 298, + 361 + ], + "lines": [ + { + "bbox": [ + 229, + 345, + 298, + 361 + ], + "spans": [ + { + "bbox": [ + 229, + 345, + 298, + 361 + ], + "type": "text", + "content": "Text prompts: {car, building,...}" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 264, + 390, + 344 + ], + "blocks": [ + { + "bbox": [ + 308, + 264, + 390, + 344 + ], + "lines": [ + { + "bbox": [ + 308, + 264, + 390, + 344 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 390, + 344 + ], + "type": "image", + "image_path": "20cf5d720d00f4ba55a989460329df6254c01ef76753b7495edf7d090aeb0969.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 345, + 373, + 361 + ], + "lines": [ + { + "bbox": [ + 324, + 345, + 373, + 361 + ], + "spans": [ + { + "bbox": [ + 324, + 345, + 373, + 361 + ], + "type": "text", + "content": "Text prompts: {trash bin}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 370, + 482, + 415 + ], + "lines": [ + { + "bbox": [ + 131, + 370, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 482, + 415 + ], + "type": "text", + "content": "Fig. 1: The SAL (Segment Anything in Lidar) model performs class-agnostic instance segmentation (i) and zero-shot classification via text prompting. This allows us to not only predict panoptic segmentation (ii) for fixed class vocabularies but segment any object (iii and iv) in a given Lidar scan." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 395, + 264, + 477, + 344 + ], + "blocks": [ + { + "bbox": [ + 395, + 264, + 477, + 344 + ], + "lines": [ + { + "bbox": [ + 395, + 264, + 477, + 344 + ], + "spans": [ + { + "bbox": [ + 395, + 264, + 477, + 344 + ], + "type": "image", + "image_path": "1512bc5e8da8666c27fc3de232e224f6db5e4c3d742bdc5c88a4d2855cbbd9db.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 345, + 460, + 361 + ], + "lines": [ + { + "bbox": [ + 410, + 345, + 460, + 361 + ], + "spans": [ + { + "bbox": [ + 410, + 345, + 460, + 361 + ], + "type": "text", + "content": "Text prompts: {streetcar}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 159, + 468, + 455, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 468, + 455, + 665 + ], + "spans": [ + { + "bbox": [ + 159, + 468, + 455, + 665 + ], + "type": "text", + "content": "Abstract. We propose the SAL (Segment Anything in Lidar) method consisting of a text-promptable zero-shot model for segmenting and classifying any object in Lidar, and a pseudo-labeling engine that facilitates model training without manual supervision. While the established paradigm for Lidar Panoptic Segmentation (LPS) relies on manual supervision for a handful of object classes defined a priori, we utilize 2D vision foundation models to generate 3D supervision \"for free\". Our pseudolabels consist of instance masks and corresponding CLIP tokens, which we lift to Lidar using calibrated multi-modal data. By training our model on these labels, we distill the 2D foundation models into our Lidar SAL model. Even without manual labels, our model achieves " + }, + { + "bbox": [ + 159, + 468, + 455, + 665 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 159, + 468, + 455, + 665 + ], + "type": "text", + "content": " in terms of class-agnostic segmentation and " + }, + { + "bbox": [ + 159, + 468, + 455, + 665 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 159, + 468, + 455, + 665 + ], + "type": "text", + "content": " in terms of zero-shot LPS of the fully supervised state-of-the-art. Furthermore, we outperform several baselines that do not distill but only lift image features to 3D. More importantly, we demonstrate that SAL supports arbitrary class prompts, can be easily extended to new datasets, and shows significant potential to improve with increasing amounts of self-labeled data. We release all models and the code." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 175 + ], + "type": "text", + "content": "We tackle segmentation and recognition of objects in Lidar point clouds, a task commonly tackled via Lidar Panoptic Segmentation (LPS)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 180, + 481, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 180, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 481, + 277 + ], + "type": "text", + "content": "Status quo. LPS has been gaining significant attention in the community due to its role in scene understanding, which is vital for safe autonomous navigation. Thanks to the availability of labeled datasets [6, 18] and advances in learning representations from unordered point sets [12, 63, 64, 77], the Lidar community made significant progress in learning to segment and classify instances of predefined and manually-labeled classes. While this progress has been impressive, existing models [2, 26, 34, 43, 68, 72, 96] cannot adapt to continually evolving class ontologies [38] that may even vary in different geographic regions [80]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 282, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 342 + ], + "type": "text", + "content": "Stirring the pot. We challenge this well-established approach and investigate how we can train general LPS models that can be prompted to segment point clouds according to any object class vocabulary. Towards such a promptable Lidar segmentation approach, we first need to be able to segment any object, a capability that remains elusive in the Lidar domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 348, + 481, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 539 + ], + "type": "text", + "content": "Towards segmenting anything in Lidar. To address these challenges, we propose SAL, which consists of a text-promptable zero-shot model (Fig. 2, left) for panoptic segmentation of arbitrary objects (Fig. 1), and pseudo-labeling engine (Fig. 2, right) that facilitates model training directly from raw sensory data. The SAL pseudo-label engine automatically labels Lidar sequences using image segmentation [30] and vision-language models [65]. We utilize SAM [30] to generate class-agnostic masks in images, and CLIP [65] to generate per-mask tokens that connect visual features to language, and finally, transfer both to Lidar using a calibrated sensory setup. Even though generated pseudo-labels only partially cover Lidar scans and are inherently noisy due to errors in the image-level generation process and imperfect sensory calibration, we demonstrate their effectiveness as self-supervised training data for the SAL zero-shot model. In contrast to prior work in LPS, SAL is the first model that can be prompted with any new vocabulary without re-training. Different from recent advances [56] that lift image features to 3D without distillation into a model, SAL does not require image features during inference and segments full " + }, + { + "bbox": [ + 132, + 348, + 481, + 539 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 132, + 348, + 481, + 539 + ], + "type": "text", + "content": " point clouds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": "Talk to your SAL. Our model predicts Lidar segmentation masks and their corresponding CLIP tokens, which allow us to perform zero-shot classification of segmented objects using arbitrary text prompts (Fig. 1). We evaluate SAL by prompting our zero-shot model on standard benchmarks for LPS. For class-agnostic segmentation, we reach " + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": " of the performance of manually supervised baselines. By classifying segmented objects in a zero-shot manner, we report the first (and very encouraging) results for zero-shot LPS and reach " + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": " of the supervised baselines trained on SemanticKITTI and nuScenes, respectively. Beyond that, as shown in Fig. 1, we can prompt SAL model to segment objects outside of existing Lidar dataset class vocabularies." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 266, + 217 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 266, + 217 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 266, + 217 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 266, + 217 + ], + "type": "image", + "image_path": "5aeb54162997a1b95043f6c9a6f3d87021c2e9daf025cad3f5bfd29f5191cfa3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 220, + 480, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 480, + 266 + ], + "type": "text", + "content": "Fig. 2: SAL overview: Given a Lidar scan and a class vocabulary prompt, specified as a list of per-class free-form text descriptions (left), SAL segments and classifiers objects (things and stuff classes). As labeled data for training such a model does not exist, we supervise SAL by distilling off-the-shelf vision foundation models to Lidar (right)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 271, + 116, + 477, + 217 + ], + "blocks": [ + { + "bbox": [ + 271, + 116, + 477, + 217 + ], + "lines": [ + { + "bbox": [ + 271, + 116, + 477, + 217 + ], + "spans": [ + { + "bbox": [ + 271, + 116, + 477, + 217 + ], + "type": "image", + "image_path": "3cc983aafba28d114fb4f567e6d9819b781280699abf68937ee548c90c2382f5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "content": "Contributions. As our main contribution, we re-think the established approach to Lidar Panoptic Segmentation and (i) present SAL for segmentation and classification of any object in a Lidar scan. As we do not utilize labeled Lidar data, we propose (ii) a pseudo-label engine that distills vision foundation models to the Lidar domain. The resulting pseudo-labels allow us to train (iii) a zero-shot LPS model without any human supervision. We demonstrate (iv) encouraging results on standard LPS benchmarks and outline a clear path towards universal, promptable segmentation foundation models for Lidar data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 382, + 236, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 236, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 236, + 395 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 406, + 479, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 479, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 479, + 455 + ], + "type": "text", + "content": "Lidar has played a pivotal role since the dawn of embodied navigation [60,78,79]. Recent data-driven efforts in Lidar perception have been pushing boundaries in semantic segmentation [3,12,35,47,69,75,82,83,85,98], object detection [32,40, 42,57-59,88,90,95], and tracking [25,49,76]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 458, + 480, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 480, + 601 + ], + "type": "text", + "content": "Lidar panoptic segmentation. Recently, Lidar Panoptic Segmentation [4,6, 7,18] has emerged as a holistic approach to Lidar-based dynamic scene understanding. Prior works [2,4,19,26,28,31,34,43-45,68,89,96,97] learn to group and classify points specified in the training data according to the target class vocabularies, while methods for open-set instance [24,49,76] and panoptic segmentation [81] rely on bottom-up grouping based on Euclidean distance between points. While these developments have been impressive, end-to-end methods remain limited to specific target classes that appear in the training data, while bottom-up, clustering-based approaches [24,27,49,76,81] are sensitive to a particular choice of clustering parameters and are unable to improve their performance in a data-driven fashion. In contrast, the SAL zero-shot model not only learns to segment objects but is also capable of zero-shot classification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "Self-supervised learning for Lidar. As labeled data in the Lidar domain is scarce, several works investigate how contrastive learning, proven effective in the image domain [10, 23], can reduce the need for labeled data. SegContrast [52] learns to align representations of point clouds and their augmented views, utilizing density-based clustering (DBSCAN [17]) to pool and contrast" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": "features. Rather than using DBSCAN, [39, 70] leverage image-based methods [1] and foundation models [30] to perform contrastive alignment between image and corresponding Lidar features. The aforementioned methods only provide pretraining recipes for potential segmentation downstream tasks. Our SAL method goes a step further by training a full zero-shot Lidar Panoptic Segmentation model that could benefit from such pre-training recipes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 192, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 312 + ], + "type": "text", + "content": "Self-supervised (3D) object detection. Several methods (self-)supervise object detectors with (RGB-D) videos [22, 53-55, 61, 62], or Lidar sequences [50, 71, 92] to discover objects in sensory streams. However, such motion-based supervision provides no semantic information. Therefore, [51] distills per-point (PCA quantized) CLIP [65] features to Lidar for zero-shot classification of detected objects. Significantly different from [51], we tackle the more general and challenging task of LPS, which segments and classifies both moving and stationary things and stuff classes. To this end, we distill CLIP features not per point but per object instance. This provides a more holistic object-centric semantic distillation and allows us to supervise our model with non-quantized CLIP features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 316, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 569 + ], + "type": "text", + "content": "Zero-shot recognition. Zero-shot recognition (ZSR) [84] methods (recently also referred to as \"open-vocabulary recognition\") tackle the recognition of distinct, yet related, semantic concepts that are either not labeled in the training data (transductive) or not observed at all (inductive). Such methods assume a dataset with labels for some classes, whereas unseen instances are supervised via attributes or class names. The latter is often used in conjunction with word embeddings [46]. By learning to align image-language features, ZSR methods can infer class labels for unseen objects [65]. ZSR has also been explored in the context of object detection, semantic segmentation, and panoptic segmentation [16,86] using word embedding models [5,8,48,66], or CLIP vision-language embeddings [20,21,33,36,67,87,91,93,94]. In the context of semantic segmentation in RGB-D and Lidar data, [41,56] augment image and language embeddings with features extracted from point clouds. Both rely on lifting dense image features from images to 3D during inference, which limits their applicability to setups with dense camera coverage. Similarly, [74] segments instances in accumulated RGB-D point clouds (ScanNet [14]) and relies on manual supervision in conjunction with image-based dense CLIP features for zero-shot classification. Moreover, datasets such as ScanNet [14] were recorded with RGB-D sensors in static environments and fused into 3D reconstructions based on scans, taken from multiple viewpoints. In contrast, SAL directly applies to full Lidar point clouds while requiring no image features at the inference time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 589, + 350, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 589, + 350, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 589, + 350, + 604 + ], + "type": "text", + "content": "3 SAL: Segment Anything in Lidar" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "This section outlines the key challenges and components towards a Lidar model that segments and classifies any object. We first define the underlying problem formation in Sec. 3.1 and then present our (Segmenting Anything in Lidar) SAL model in Sec. 3.2." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 265, + 193 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 265, + 193 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 265, + 193 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 265, + 193 + ], + "type": "image", + "image_path": "2c58dbd436e1694c79a318974d52cd36f7a1e35f1f775233fbc35fde74144a5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 196, + 242, + 205 + ], + "lines": [ + { + "bbox": [ + 157, + 196, + 242, + 205 + ], + "spans": [ + { + "bbox": [ + 157, + 196, + 242, + 205 + ], + "type": "text", + "content": "(a) Pseudo-label engine" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 270, + 114, + 478, + 193 + ], + "blocks": [ + { + "bbox": [ + 270, + 114, + 478, + 193 + ], + "lines": [ + { + "bbox": [ + 270, + 114, + 478, + 193 + ], + "spans": [ + { + "bbox": [ + 270, + 114, + 478, + 193 + ], + "type": "image", + "image_path": "f37f165dfcb9f08cd9cb6920825715429cf2640b099b3c8d3b7ff2958784bc1a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 196, + 411, + 205 + ], + "lines": [ + { + "bbox": [ + 339, + 196, + 411, + 205 + ], + "spans": [ + { + "bbox": [ + 339, + 196, + 411, + 205 + ], + "type": "text", + "content": "(b) Zero-shot model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 215, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 326 + ], + "type": "text", + "content": "Fig. 3: Our pseudo-label engine (Fig. 3a) utilizes SAM [30] to estimate segmentation masks in images, MaskCLIP [16] to estimate corresponding per-mask CLIP features, and a calibrated sensory setup to transfer them to the Lidar domain. We distill these pseudo-labels to our zero-shot model (Fig. 3b), which segments and classifies Lidar point clouds. The SAL zero-shot model employs a sparse-convolutional backbone [12], followed by a Transformer decoder that predicts objectness scores, segmentation masks, and CLIP tokens for each query. To (optionally) perform zero-shot classification, we forward the dataset class vocabulary through the CLIP text encoder and match the encoded vocabulary with predicted CLIP tokens. Our model requires no retraining for different vocabularies and no image features at inference time." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 330, + 361, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 330, + 361, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 330, + 361, + 342 + ], + "type": "text", + "content": "3.1 Zero-Shot Lidar Panoptic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "text", + "content": "The task. We discuss and evaluate SAL in the context of Lidar Panoptic Segmentation (LPS), which tackles both instance segmentation and object recognition. LPS methods take as input an unordered point cloud " + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times 4}" + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "text", + "content": " encoding spatial coordinates and their sensor intensity. The currently established problem setting [7,18] assumes supervision and task outputs as per-point semantic class and instance identity labels. The class labels are confined to a pre-defined and fixed class vocabulary " + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "text", + "content": ". In contrast, we tackle LPS in a generalized, zero-shot setting, where the class vocabulary is only provided at inference, not training time. Such a vocabulary " + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "text", + "content": " specifies target classes as a list of " + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "text", + "content": " class prompts. Each prompt " + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "inline_equation", + "content": "c_{i} \\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "text", + "content": " is specified via free-form text, e.g., with a class name and optional class description. To perform zero-shot LPS a model must be designed to segment and classify any object in a Lidar scan." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 548 + ], + "type": "text", + "content": "The challenge. Where do we get the data to train such a model? Image-based models [16,21,30,91,93] rely on large and diverse datasets for pre-training [15], dense prediction [13,37], and foundation models that align images with textual descriptions [65] - a commodity not available in the Lidar domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 563, + 228, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 228, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 228, + 574 + ], + "type": "text", + "content": "3.2 SAL Overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Our 2D-to-Lidar distillation method consists of two core components: (i) The pseudo-label engine transfers 2D vision foundation models into Lidar pseudolabels using multi-modal inputs from a calibrated sensory setup, shown in Fig. 3a. (ii) Our zero-shot model is trained on the generated pseudo-labels and is able to perform class-agnostic segmentation and zero-shot classification via text prompts. Given a semantic dataset vocabulary, this allows us to tackle zero-shot LPS as illustrated in Fig. 3b." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 287, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 287, + 128 + ], + "type": "text", + "content": "3.3 SAL Pseudo-Label Engine" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 480, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 480, + 173 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 480, + 173 + ], + "type": "text", + "content": "The SAL pseudo-label engine relies on a calibrated multi-modal sensory setup with " + }, + { + "bbox": [ + 130, + 137, + 480, + 173 + ], + "type": "inline_equation", + "content": "k \\geq 1" + }, + { + "bbox": [ + 130, + 137, + 480, + 173 + ], + "type": "text", + "content": " RGB cameras. Furthermore, a sufficient overlap between the Lidar sensing area and each camera view is paramount." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 481, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 481, + 261 + ], + "type": "text", + "content": "Mask generation: To pseudo-label object instances, we utilize the Segment Anything (SAM) model [30], which generates an overlapping set of segmentation masks. We flatten SAM's output mask hierarchy by non-maxima suppression (NMS) with a minimal overlap threshold to ensure mutually exclusive masks, suppressing object parts and subparts in favor of objects. This way, we obtain a set of non-overlapping binary masks " + }, + { + "bbox": [ + 130, + 177, + 481, + 261 + ], + "type": "inline_equation", + "content": "m_{i}^{k} \\in \\{0,1\\}^{W \\times H}" + }, + { + "bbox": [ + 130, + 177, + 481, + 261 + ], + "type": "text", + "content": " for each camera view " + }, + { + "bbox": [ + 130, + 177, + 481, + 261 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 177, + 481, + 261 + ], + "type": "text", + "content": " with image plane of size " + }, + { + "bbox": [ + 130, + 177, + 481, + 261 + ], + "type": "inline_equation", + "content": "W \\times H" + }, + { + "bbox": [ + 130, + 177, + 481, + 261 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "type": "text", + "content": "CLIP image token generation: As shown by CLIP [65], aligning image and text features allows vision-language foundation models to perform zero-shot image classification based on arbitrary text prompts. To transfer this capability to the Lidar domain, we generate a localized CLIP image feature token " + }, + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "type": "inline_equation", + "content": "f_{i}^{k} \\in \\mathbb{R}^{C_{t}}" + }, + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "type": "text", + "content": " for each binary SAM mask " + }, + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "type": "inline_equation", + "content": "m_{i}^{k}" + }, + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "type": "text", + "content": ". To obtain image tokens for a masked region of the input image, we utilize [16] and their relative mask attention in the CLIP image encoder feature space. Note that we never classify segments in the image domain but merely distill CLIP image feature tokens to Lidar to facilitate zero-shot classification during inference where we do not use any image features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "text", + "content": "Image-to-Lidar unprojection: From the Lidar perspective, we unproject each image mask " + }, + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "inline_equation", + "content": "m_i^k \\in \\{0,1\\}^{W \\times H}" + }, + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "text", + "content": " to a binary Lidar segmentation mask " + }, + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\tilde{m}_i \\in \\{0,1\\}^N" + }, + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "text", + "content": " by transforming the respective camera coordinate frame to the Lidar space. For datasets with multiple cameras, such as nuScenes [18], we process each image independently, followed by a cross-camera fusion of masks with a sufficient IoU overlap and averaging of their CLIP image tokens. The unprojection yields pairs " + }, + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\{\\tilde{m}_i^k, f_i^k\\}" + }, + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "text", + "content": " of Lidar masks and their corresponding CLIP features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 465, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 574 + ], + "type": "text", + "content": "Refinement via density-based clustering: Image-to-Lidar transformation (Fig. 4a→Fig. 4b) is inherently noisy due to imperfect calibration and issues with synchronization and Lidar rolling shutter. We improve our pseudo-label quality by creating an ensemble of DBSCAN [17] clusters " + }, + { + "bbox": [ + 130, + 465, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\tilde{m}_l \\in \\{0,1\\}^N" + }, + { + "bbox": [ + 130, + 465, + 482, + 574 + ], + "type": "text", + "content": ", obtained by varying the density threshold to compensate for varying density in Lidar point clouds (details in the appendix). We replace each " + }, + { + "bbox": [ + 130, + 465, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\tilde{m}_i" + }, + { + "bbox": [ + 130, + 465, + 482, + 574 + ], + "type": "text", + "content": " with its best-matching " + }, + { + "bbox": [ + 130, + 465, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\tilde{m}_l" + }, + { + "bbox": [ + 130, + 465, + 482, + 574 + ], + "type": "text", + "content": " in case their IoU exceeds a minimal overlap threshold and retain the original mask otherwise to obtain a refined set of pseudo-labels (Fig. 4c) that retain their original cardinality and associated CLIP features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 592, + 266, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 266, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 266, + 604 + ], + "type": "text", + "content": "3.4 SAL Zero-Shot Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 613, + 480, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 480, + 637 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 480, + 637 + ], + "type": "text", + "content": "The universal SAL zero-shot model deconstructs LPS into class-agnostic segmentation and zero-shot classification via text prompts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Universal architecture. Instead of relying on highly specialized and engineered LPS models, we base the SAL model on a universal Transformer decoder" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 246, + 164 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 164 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 164 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 164 + ], + "type": "image", + "image_path": "b54c146f43ff079276c0b5fa2bd9ea445b35766b2aaee6e933c85664775d83ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 165, + 218, + 175 + ], + "lines": [ + { + "bbox": [ + 161, + 165, + 218, + 175 + ], + "spans": [ + { + "bbox": [ + 161, + 165, + 218, + 175 + ], + "type": "text", + "content": "(a) SAM masks" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 364, + 164 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 364, + 164 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 364, + 164 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 364, + 164 + ], + "type": "image", + "image_path": "5e738e9d95690a48a4678403483432bed70b632f7484385335f3b48b3778b380.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 165, + 345, + 175 + ], + "lines": [ + { + "bbox": [ + 268, + 165, + 345, + 175 + ], + "spans": [ + { + "bbox": [ + 268, + 165, + 345, + 175 + ], + "type": "text", + "content": "(b) SAM unprojected" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 183, + 480, + 217 + ], + "lines": [ + { + "bbox": [ + 131, + 183, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 183, + 480, + 217 + ], + "type": "text", + "content": "Fig. 4: Refinement via clustering. After transferring image masks (Fig. 4a) to Lidar (Fig. 4b), we obtain pseudo-labels that suffer from sensory misalignment-related issues. Our geometric refinement (Fig. 4c) improves localization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 368, + 114, + 482, + 164 + ], + "blocks": [ + { + "bbox": [ + 368, + 114, + 482, + 164 + ], + "lines": [ + { + "bbox": [ + 368, + 114, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 368, + 114, + 482, + 164 + ], + "type": "image", + "image_path": "e948e3f8612409e66dc5ff7741d878924f11a71b688833788f4fba3f7ee0688f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 165, + 463, + 175 + ], + "lines": [ + { + "bbox": [ + 386, + 165, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 386, + 165, + 463, + 175 + ], + "type": "text", + "content": "(c) SAM + DBSCAN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 223, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 480, + 319 + ], + "type": "text", + "content": "architecture (Fig. 3b), similar to [43]. Its 2D counterparts [9, 11] provide top-performing results across semantic, instance, and panoptic segmentation for images. With increasing amounts of (pseudo-)labeled data, Transformer decoders have the potential to achieve a similar impact in the Lidar domain. Following [43], we deploy a Minkowski U-Net [12] backbone for feature extraction followed by a Transformer decoder architecture with object query to point/voxel feature crossattention. Training our model to segment and classify any object and perform in a zero-shot evaluation setting requires a unique design of the final task heads." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 323, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 480, + 396 + ], + "type": "text", + "content": "Class-agnostic segmentation. To localize segments in a point cloud, we rely on two model heads that predict segmentation masks and objectness scores for each query. The architecture of the former follows [43] and predicts binary output masks " + }, + { + "bbox": [ + 130, + 323, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\hat{m}_i\\in \\{0,1\\} ^N" + }, + { + "bbox": [ + 130, + 323, + 480, + 396 + ], + "type": "text", + "content": " by computing the dot product between queries and point features. The objectness head reduces the multi-class problem to a binary object or no-object decision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 400, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 480, + 495 + ], + "type": "text", + "content": "Zero-shot classification. To equip our model with zero-shot classification capabilities without relying on any image input, we learn to predict CLIP [65] Lidar tokens, i.e., features in the CLIP space obtained from Lidar inputs, for each query. The token head consists of a three-layer MLP that directly regresses tokens " + }, + { + "bbox": [ + 130, + 400, + 480, + 495 + ], + "type": "inline_equation", + "content": "\\hat{f}_i^j" + }, + { + "bbox": [ + 130, + 400, + 480, + 495 + ], + "type": "text", + "content": ". At inference, the predicted tokens are matched to text prompts via the CLIP text encoder. The matching yields a probability distribution over the prompts. To perform zero-shot LPS on a pre-defined dataset, we use its class vocabulary as input text prompts (Fig. 3b)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 496, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 480, + 520 + ], + "type": "text", + "content": "We train our model jointly for both class-agnostic segmentation and zero-shot classification with the following loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 242, + 531, + 480, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 531, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 242, + 531, + 480, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S A L}} = \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {\\text {t o k e n}}, \\tag {1}", + "image_path": "fd074231ea5a8ec6663d1c5d21f75eef37260a195649869b258364ca582cb857.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "text", + "content": "with binary cross-entropy loss " + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "text", + "content": " and a cosine distance loss " + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{token}" + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "text", + "content": ". The segmentation loss " + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "text", + "content": " follows [43] and consists of a binary mask cross-entropy and dice loss. The " + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SAL}}" + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "text", + "content": " loss is supervised by pseudo-label pairs " + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "inline_equation", + "content": "\\{\\tilde{m}_i^k, f_i^k\\}" + }, + { + "bbox": [ + 130, + 553, + 480, + 613 + ], + "type": "text", + "content": " obtained from our label engine and thereby distills both 2D foundation models (SAM and CLIP) into our LPS model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": "How to train with partial labels? Partially (pseudo-)labeled point clouds present a challenge for SAL model training. For example, in SemanticKITTI, only " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " of points are pseudo-labeled due to low camera coverage (see appendix). If we naively train SAL on partially (pseudo-)labeled point clouds, the objectness" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " would penalize any segmentation in these regions as a false positive, thereby teaching the model to ignore these points entirely. Empirically, we determined that the most effective training strategy is to (i) remove unlabeled points from point clouds, (ii) utilize standard data augmentations (rotations, flipping, scaling, and translations), in conjunction with proposed FrankenFruustum augmentation (detailed in the appendix) that mimics fully-labeled point clouds during training by randomly removing unlabeled points and replicating labeled frustum regions around the vertical axis. This augmentation does not increase the overall label coverage, but our ablations in Sec. 4.2 show its effectiveness in reducing the domain gap between training and inference input clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 263, + 230, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 230, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 230, + 277 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 360 + ], + "type": "text", + "content": "In the following section, we outline our experimental setup (Sec. 4.1) used to ablate the SAL model and its training strategies (Sec. 4.2). Since SAL presents the first Zero-Shot Lidar Panoptic Segmentation (ZS-LPS) model, we design and compare with multiple hand-crafted baselines that implement concepts from related works. We conclude our analysis by comparing our zero-shot performance (Sec. 4.4) to fully-supervised LPS methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 373, + 262, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 262, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 262, + 385 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "content": "Datasets. We evaluate SAL on two public Lidar Panoptic Segmentation (LPS) datasets, SemanticKITTI [6, 7] and nuScenes [18]. We utilize provided, human-annotated ground-truth (GT) panoptic segmentation labels solely for evaluation purposes, linear probing ablations, and baselines. We train SAL model only using our self-supervised pseudo-labels (Sec. 3.3). To demonstrate the general applicability of SAL to other domains and datasets, we show qualitative results of SAL trained on Waymo Open [73] dataset, which contains no GT LPS labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "text", + "content": "Metrics. For the evaluation, we utilize the standard Panoptic Segmentation [29] metrics. The Panoptic Quality " + }, + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "inline_equation", + "content": "PQ = RQ \\times SQ" + }, + { + "bbox": [ + 130, + 477, + 482, + 537 + ], + "type": "text", + "content": " combines Segmentation Quality (SQ) and Recognition Quality (RQ), i.e., F-1 score. True positives (TPs) are mask predictions with sufficient intersection-over-union (IoU) overlap with GT masks of the same class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "type": "text", + "content": "Semantic Oracle (SO). The aforementioned metrics assume GT labels and predictions with instance IDs and semantic classes for each point. Therefore, they are not suitable for evaluating class-agnostic segmentation. To assess the class-agnostic segmentation independently of semantics, we apply SO during the evaluation by assigning ground truth semantic classes " + }, + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "type": "text", + "content": " to predicted masks " + }, + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "type": "text", + "content": " via majority voting. Evaluation with SO is only needed for evaluation of class-agnostic segmentation - our zero-shot models provide both instance and semantic class predictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Stuff Merging (SM). SO allows us to assess class-agnostic instance segmentation for things classes. However, this approach is unsuitable for assessing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "the performance of stuff classes, as existing Lidar datasets merge instances of stuff classes into a single instance. The things- stuff separation can be unintuitive and inconsistent across datasets (e.g., SemanticKITTI merges instances of traffic signs into a single stuff class). In contrast, we fully embrace the philosophy that all classes can be segmented into individual instances (e.g., segmentation models should localize individual trees/bushes in the vegetation class or distinguish individual buildings). To evaluate our models on the target datasets in the SO regime on Lidar datasets that do not provide instance-level annotations for all classes, we additionally report results using a merge strategy (SM), which merges all instances of a particular stuff class into a single mask to ensure outputs of our models are consistent with the format of target datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 267, + 229, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 229, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 229, + 278 + ], + "type": "text", + "content": "4.2 SAL Ablations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 290, + 481, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 481, + 326 + ], + "type": "text", + "content": "In this section, we discuss how to train SAL zero-shot segmentation model using self-generated pseudo-labels (Tab. 1) and discuss design decisions behind our instance (Tab. 2) and semantic distillation (Tab. 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 474 + ], + "type": "text", + "content": "Learning with partial labels. For the single-camera setup [6,7], our pseudolabels only cover " + }, + { + "bbox": [ + 130, + 330, + 482, + 474 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 330, + 482, + 474 + ], + "type": "text", + "content": " of the input point clouds (see appendix). Training on partial label coverage presents a challenge for our model. In Tab. 1, we only evaluate class-agnostic segmentation by applying SO and SM (see Sec. 4.1). We compare SAL models trained using pseudo and GT labels. The Frustum Filter removes all unlabeled points not visible in the camera frustum during training and/or evaluation. As shown in Tab. 1, removing unlabeled points from pseudo-labeled point clouds during training has a significant benefit on the performance (" + }, + { + "bbox": [ + 130, + 330, + 482, + 474 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 130, + 330, + 482, + 474 + ], + "type": "text", + "content": " row, 59.3 PQ) as compared to the variant, where we merely ignore unlabeled region during the training (" + }, + { + "bbox": [ + 130, + 330, + 482, + 474 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 130, + 330, + 482, + 474 + ], + "type": "text", + "content": " row, 22.2 PQ). This is likely due to data imbalance in a single-camera setup, leading to a class imbalance between labeled/unlabeled points, incentivizing SAL to suppress predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "While utilizing standard augmentations (rotations, flipping, scaling, and translations) is crucial, we observe that performing our FrankenFrustum augmentation, which concatenates the visible portion of the point cloud around the z-axis (see Sec. 3.4), significantly improves results (" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " row, 62.5 PQ). Mixing point clouds from different scans (" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " row, 62.8 PQ) further boosts results. Overall, we obtain 69 PQ from GT supervision (1st row) and 62.8 PQ when using our pseudo-labels (5th row). Remarkably, our labels yield " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " of the GT performance while covering only " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " of the input point cloud (see appendix). To further contextualize these results, we note that this evaluation is performed only on classes for which GT instance labels are available. As shown in Fig. 1, our model learns to segment a much larger variety of classes. Evaluating only the subset of points visible in the camera (" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " of all points, rows 6&7), the gap between the GT-supervised model (71.8 PQ) and pseudo-supervised model (70.7 PQ) shrinks further. We conclude that by training SAL model using pseudo-labels, we distill the notion of objectness from SAM [30] into our model and obtain segmentation capabilities, similar to the model trained on GT data." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 190, + 317, + 290 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 318, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 318, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 318, + 180 + ], + "type": "text", + "content": "Table 1: Class-agnostic segmentation. By cropping unlabeled points and performing data augmentations (in combination with our FrankenFrostum), SAL successfully learns to segment full point clouds even when only " + }, + { + "bbox": [ + 130, + 114, + 318, + 180 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 114, + 318, + 180 + ], + "type": "text", + "content": " of points are (pseudo)-labeled." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 190, + 317, + 290 + ], + "lines": [ + { + "bbox": [ + 132, + 190, + 317, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 317, + 290 + ], + "type": "table", + "html": "
LabelsFrust.-Filter TrainEvalFranken Frust.PQSQ\\( PQ^{Th} \\)\\( PQ^{St} \\)
GT69.083.581.659.7
Pseu.22.267.545.15.5
Pseu.✓59.378.265.854.5
Pseu.✓✓62.579.167.658.8
Pseu.✓✓(mix)62.879.069.058.3
GT✓71.884.884.662.5
Pseu.✓✓✓(mix)70.781.975.467.3
", + "image_path": "b24378f03d874df4f99196aeb219411d2749539c1c572431a9907bbf2480fba2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 320, + 190, + 477, + 284 + ], + "blocks": [ + { + "bbox": [ + 320, + 114, + 479, + 181 + ], + "lines": [ + { + "bbox": [ + 320, + 114, + 479, + 181 + ], + "spans": [ + { + "bbox": [ + 320, + 114, + 479, + 181 + ], + "type": "text", + "content": "Table 2: Scaling. In contrast to predefined GT data, increasing the number of queries SAL improves performance on stuff classes. By increasing the amount of labeled data, we further improve segmentation performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 190, + 477, + 284 + ], + "lines": [ + { + "bbox": [ + 320, + 190, + 477, + 284 + ], + "spans": [ + { + "bbox": [ + 320, + 190, + 477, + 284 + ], + "type": "table", + "html": "
TrainNum.PQSQ\\( PQ^{Th} \\)\\( PQ^{St} \\)
LabelsSetQuer.
GTtrain10067.882.479.759.2
GTtrain20067.983.679.859.3
GTtrain30069.083.581.659.7
Pseu.train10053.975.362.247.8
Pseu.train20060.877.066.756.5
Pseu.train30062.879.069.058.3
Pseu.bigtrain30065.379.571.960.5
", + "image_path": "2a939d7f05942b6f906506af9c6f7e2892a0b21a39d7446d8a5ec0379e98953b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 300, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 468 + ], + "type": "text", + "content": "Scaling queries and data. Analogous to the grid inference of SAM, which operates with hundreds of spatial queries (prompts), we analyze the effect of different numbers of decoder queries. We subsample instances during training to train models with fewer queries than the maximum number of segments per scan. As visualized in Tab. 2, increasing the number of decoder queries improves the recognition of things for both training with pseudo and GT labels. However, we observe a significant performance boost with pseudo labels for stuff classes " + }, + { + "bbox": [ + 130, + 300, + 482, + 468 + ], + "type": "inline_equation", + "content": "(58.3\\mathrm{PQ}^{St}, +10.8)" + }, + { + "bbox": [ + 130, + 300, + 482, + 468 + ], + "type": "text", + "content": ", while the improvement for GT labels is marginal " + }, + { + "bbox": [ + 130, + 300, + 482, + 468 + ], + "type": "inline_equation", + "content": "(59.7\\mathrm{PQ}^{St}, +0.5)" + }, + { + "bbox": [ + 130, + 300, + 482, + 468 + ], + "type": "text", + "content": ". Our model benefits from a larger number of queries for stuff classes as it learns a fine-grained segmentation model that is not limited to a prefixed set of things classes. In the last row, we concatenate pseudo-labeled train and test sets (neither used for the validation) into a bigtrain set and observe an improvement of " + }, + { + "bbox": [ + 130, + 300, + 482, + 468 + ], + "type": "inline_equation", + "content": "+2.5" + }, + { + "bbox": [ + 130, + 300, + 482, + 468 + ], + "type": "text", + "content": " PQ, suggesting that scaling the amount of training data has the potential to improve the performance of our models further." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 586, + 278, + 635 + ], + "blocks": [ + { + "bbox": [ + 130, + 487, + 280, + 585 + ], + "lines": [ + { + "bbox": [ + 130, + 487, + 280, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 280, + 585 + ], + "type": "text", + "content": "Table 3: Semantic distillation. We linearly probe the SAL model, trained with and without semantic distillation loss " + }, + { + "bbox": [ + 130, + 487, + 280, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{token}" + }, + { + "bbox": [ + 130, + 487, + 280, + 585 + ], + "type": "text", + "content": ". We train a linear classifier with GT labels while keeping backbone and decoder features frozen. As can be seen, " + }, + { + "bbox": [ + 130, + 487, + 280, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{token}" + }, + { + "bbox": [ + 130, + 487, + 280, + 585 + ], + "type": "text", + "content": " successfully distills a notion of semantics into our model." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 586, + 278, + 635 + ], + "lines": [ + { + "bbox": [ + 132, + 586, + 278, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 278, + 635 + ], + "type": "table", + "html": "
Linear prob. LtokenPQRQSQmIoU
×20.026.355.223.4
××33.141.968.340.0
×24.832.366.829.7
", + "image_path": "c61cf55e92c0284e2b30610e115595330dda41318a1dadd0e2f5ddb7a3c1b21b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 286, + 474, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 474, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 286, + 474, + 482, + 641 + ], + "type": "text", + "content": "Semantic distillation. At inference, SAL predicts a binary segmentation mask and objectness score for each query. To perform zero-shot classification, our model additionally predicts CLIP image tokens. This allows us to not rely on image features during inference but prompt our predicted tokens with arbitrary texts, i.e., object classes. Does the token head loss " + }, + { + "bbox": [ + 286, + 474, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{token}" + }, + { + "bbox": [ + 286, + 474, + 482, + 641 + ], + "type": "text", + "content": " successfully distill a notion of object semantics from CLIP into our model? To verify this, we perform linear probing experiments: we replace the token distillation head of a trained model with a linear classifier (backbone and instance de" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "coder remain frozen) and train it using GT labels. As can be seen in Tab. 3, distilling CLIP features into our model significantly improves linear probing" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 190, + 480, + 348 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 4: Zero-shot panoptic segmentation. We utilize prior efforts in the image domain [30] and Lidar [56] domain to craft multiple baselines that only unproject segmentation masks and lift image features to Lidar. By contrast, SAL distills outputs of such baselines (pseudo-labels) into a stronger Lidar segmentation model. With ImageFeat, we denote methods that require image features at inference time, and Frust. Eval. denotes the evaluation of a subset of points visible in the camera." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 190, + 480, + 348 + ], + "lines": [ + { + "bbox": [ + 132, + 190, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 480, + 348 + ], + "type": "table", + "html": "
MethodFrost. ImageDefault classesSuper classes
EvalFeat.PQSQPQThPQStmIoUPQSQPQThPQStmIoU
Class-agnostic Segmentation (Semantic Oracle)
SAM✓✓46.072.149.743.4------
SAM+Erosion✓✓42.269.445.639.6------
SAM+DBS (filter)✓✓46.770.376.824.8------
SAM+DBS (replace)✓✓48.773.753.145.4------
SAL✓X70.781.975.467.3------
Zero-Shot Lidar Panoptic Segmentation
SAM+DBS+CLIP✓✓27.571.531.724.530.651.177.571.241.054.3
SAL✓X33.171.422.840.533.563.984.288.351.766.4
SAM+DBS+CLIPXX8.256.418.60.67.511.547.60.017.311.2
SALXX24.866.817.430.228.748.578.880.432.652.8
", + "image_path": "144cadf95b7313c8a13e48091d9cddd53459183980961368daf392f02b875354.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 441 + ], + "type": "text", + "content": "performance (33.1 PQ, 40 mIoU) compared to the baseline, not trained with the semantic distillation loss " + }, + { + "bbox": [ + 130, + 356, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{token}" + }, + { + "bbox": [ + 130, + 356, + 482, + 441 + ], + "type": "text", + "content": " (20.0 PQ, 23.4 mIoU). This confirms that by distilling CLIP features, we inject a notion of semantics into our model, even though our model is not explicitly supervised with any labels containing semantic information - the notion of semantics is learned implicitly via CLIP feature distillation. Without this loss, our zero-shot model (24.8 PQ, 29.7 mIoU) outperforms the linearly probed model tuned using GT semantic labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 464, + 460, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 460, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 460, + 478 + ], + "type": "text", + "content": "4.3 Class-Agnostic and Zero-Shot Lidar Panoptic Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 492, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 577 + ], + "type": "text", + "content": "As the first study tackling Zero-Shot Lidar Panoptic Segmentation (ZS-LPS), we devise several strong baselines inspired by prior works in RGB-D semantic segmentation. During inference, these methods utilize image-based models and lift image features to 3D [56]. To ensure these are evaluated fairly, we report results in Tab. 4 on the subset of points visible in at least one camera (Frost. Eval.). Our SAL model does not have this limitation. Therefore, we additionally report results evaluated on full point clouds in the bottom of Tab. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Class-agnostic segmentation. Our first baseline (SAM [30]) generates masks in images and unprojects them to Lidar (see Sec. 3.3), and leads to 46 PQ. We observe bleeding edges (see Fig. 4b) after unprojecting from the cameras to Lidar caused by imperfect calibrations. To mitigate these artifacts, we experiment with slightly eroding SAM predictions in the image domain. However, this leads to a performance drop (42.2 PQ), likely because smaller segments after erosion fail to pass " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " overlap with labeled instances." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "Density-based filtering: Alternatively, to mitigate unprojection errors, we experiment with density-based clustering methods. We generate a large pool of DBSCAN (DBS) clusters and test two strategies to improve our SAM-based segmentation labels: (i) filtering segments without a sufficient overlap with any DBSCAN cluster or (ii) replacing segments with large overlaps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 180, + 482, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 482, + 241 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 482, + 241 + ], + "type": "text", + "content": "SAM+DBS (filter): Intuitively, the filter approach removed Lidar segments only segmentable in the image domain, e.g., a SAM mask that segments a shadow on a flat wall. While this strategy improves precision for things classes (76.8 PQTh, +27.1), it significantly degrades performance on stuff classes (24.8 PQSt -18.6) and overall degrades PQ to 46.7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 244, + 482, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 294 + ], + "type": "text", + "content": "SAM+DBS (replace): The replace strategy (that replaces SAM masks with DBSCAN clusters with sufficient mutual overlap), on the other hand, improves both things and stuff classes to an overall PQ of 48.7 (+3.9) and SQ 73.7 (+1.1), respectively. This improvement can be visually verified in Fig. 4c." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": "SAL: We train our SAL model using top-performing " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "SAM+DBS" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": " (replace) and observe a significant improvement (70.7 PQ) in terms of both things and stuff classes. Even though SAL obtains significantly higher PQ than hand-crafted baselines, it also improves in terms of SQ (81.9 vs. 73.7 closest competitor), suggesting that the distilled model is robust to projection artifacts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 494 + ], + "type": "text", + "content": "Do we need a vision segmentation foundation model? To generate a set with a sufficiently high recall, we need to cluster points with varying DBSCAN radius parameters (details in the appendix). This leads to a large set of predicted masks per scan (5,413 avg.). While most of the correct segments are in this set, separating the signal from the noise is difficult. Prior works, such as [50], are limited to moving thing classes as they rely on motion cues to filter static point cloud regions and DBSCAN to group remaining points. Our core insight is that we can utilize vision foundation models that already learned a notion of objectness: SAM generates only 171 masks per image, reduced to 45 via flattening, and to 39 masks after the unprojection, achieving a high recall with only a few instances, or in other words, a high signal-to-noise ratio." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "type": "text", + "content": "Zero-shot segmentation. So far, we discussed class-agnostic segmentation. To obtain semantic class predictions per mask in a zero-shot fashion, we devise a simple baseline inspired by [56]. Once masks are unprojected from the image to Lidar space, we utilize associated CLIP features (directly extracted from images using [16]) to perform zero-shot classification via dot product between encoded class prompts and (lifted) CLIP features [65]. This baseline (SAM+DBS+CLIP) can be understood as an approach that directly lifts image features from state-of-the-art image-based models [16] to Lidar. We detail prompts in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "There is a significant performance drop between " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "SAM+DBS" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " (replace) baseline, evaluated with Semantic Oracle (" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " row, 48.7 PQ), and zero-shot baseline (" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "SAM+DBS" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " (replace)+CLIP), " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "7^{th}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " row, 27.5 PQ. This is not surprising, as zero-shot segmentation is a challenging problem in both image [16] and Lidar domains. To this end, even when transferring features of state-of-the-art image-based models to Lidar we observe a significant performance drop. Remarkably," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 179, + 460, + 348 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 5: Lidar Panoptic Segmentation (LPS) on SemanticKITTI and nuScenes validation sets. We prompt our zero-shot SAL model with the respective class vocabularies and compare its performance to fully-supervised baselines. On SemanticKITTI and nuScenes, we reach " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " of the fully-supervised model, respectively. This gap reduces significantly when we evaluate super classes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 179, + 460, + 348 + ], + "lines": [ + { + "bbox": [ + 151, + 179, + 460, + 348 + ], + "spans": [ + { + "bbox": [ + 151, + 179, + 460, + 348 + ], + "type": "table", + "html": "
MethodSupervisionDefault classesSuper classes
PQRQSQ\\( PQ^{Th} \\)\\( PQ^{St} \\)mIoUPQRQSQmIoU
SemanticKITT1DS-Net [26]Full57.768.077.661.854.863.5---
PolarSeg [96]Full59.170.278.365.754.364.5---
EfficientLPS [72]Full59.269.875.058.060.964.9---
GP-S3Net [68]Full63.375.981.470.258.373.0---
MaskPLS [43]Full59.869.076.3---78.487.188.2
SALFull59.569.275.762.357.463.881.790.089.2
SALZero-shot24.832.366.817.430.228.748.559.478.8
mScenesPHNet [34]Full74.784.288.274.075.979.7---
DS-Net [26]Full51.259.086.138.472.373.5---
GP-S3Net [68]Full61.072.084.156.066.075.8---
EfficientLPS [72]Full62.073.983.456.870.665.6---
PolarSeg [96]Full63.475.383.959.270.466.9---
MaskPLS [43]Full57.766.071.864.452.262.571.581.086.2
SALFull70.580.885.979.461.772.874.282.787.1
SALZero-shot38.447.877.247.529.233.952.663.577.3
", + "image_path": "bc047352df77dc94781bfa399903d099fdfac172b540eceabd4ca2efe8b619f4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": "SAL distills such noisy image-based features to a stronger model (" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "8^{th}" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": " row, 33.1 PQ, +5.6 PQ). Finally, to emphasize the inherent limitation of baselines, relying on image features, we evaluate on full " + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": " point clouds. SAL performs similarly in both settings. The baseline, on the other hand, relies on image features and is, therefore, unable to predict labels out of the camera frustum. This problem can be mitigated by utilizing a setup with denser camera coverage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 491 + ], + "type": "text", + "content": "Evaluation on super-classes. We observe that the CLIP classifier often confuses related classes, e.g., car and other vehicle. We additionally evaluate our semantics by prompting with super-classes (as defined in [6]). This alone increases PQ to 63.9 (48.5 when evaluating on full point clouds) and suggests significant potential for improving prompting and text-to-image alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 509, + 365, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 365, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 365, + 521 + ], + "type": "text", + "content": "4.4 Lidar Panoptic Segmentation Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 529, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 601 + ], + "type": "text", + "content": "Finally, we compare SAL with state-of-the-art LPS methods on standard benchmarks [6,18]. We focus this discussion on comparing SAL, trained with labeled data, and our SAL zero-shot model that specifies classes in benchmarks [6,18] as text prompts. Importantly, zero-shot results are obtained without training on any labeled data. In contrast, all baselines are trained using human-labeled data, and are constrained to their respective, pre-defined class vocabularies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "SemanticKITTI. When evaluating with the default class vocabulary (Tab. 5), our zero-shot model reaches 24.8 PQ or " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " of the fully-supervised model (59.5 PQ). As discussed in Sec. 4.2, our method is mainly limited by the quality of the generated CLIP features. In practice, our zero-shot semantic classifier mainly confuses objects within super classes, e.g., car vs. other-vehicle. Therefore," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "text", + "content": "without any retraining, we prompt our zero-shot model with super classes and obtain 48.5 PQ, " + }, + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "inline_equation", + "content": "59\\%" + }, + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "text", + "content": " of the supervised model (81.7 PQ)," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 144, + 481, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 144, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 481, + 204 + ], + "type": "text", + "content": "nuScenes. We report our result in Tab. 5 and reach similar conclusions. SAL reaches 38.4 PQ on default classes, " + }, + { + "bbox": [ + 132, + 144, + 481, + 204 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 132, + 144, + 481, + 204 + ], + "type": "text", + "content": " of the supervised model (70.5 PQ). As with SemanticKITTI, results consistently improve when evaluated with superclasses. In this setting, zero-shot model yields 52.6 PQ. Note that SAL model, supervised with labeled data, significantly outperforms MaskPLS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "text", + "content": "Towards closing the gap. Compared to SemanticKITTI (24.8 PQ, " + }, + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "text", + "content": "), we observe a significantly smaller gap between the model trained on pseudo- and ground truth labels for nuScenes (38.4 PQ, " + }, + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "text", + "content": "). Despite its sparse Lidar signal and more diverse input scenarios, SAL performs better on this challenging Lidar dataset. We attribute this to a larger pseudo-label coverage of nuScenes (" + }, + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "inline_equation", + "content": "48\\%" + }, + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "text", + "content": ", compared to " + }, + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 132, + 207, + 481, + 316 + ], + "type": "text", + "content": " coverage on SemanticKITTI), and, therefore, effectively more labeled data, which is also consistent with our observations on SemanticKITTI w.r.t. increasing the amount of labeled data (Tab. 2). This trend suggests a clear path forward: applying SAL to pseudo-label more data with zero annotation cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 319, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 415 + ], + "type": "text", + "content": "Beyond labeled datasets. Datasets, such as SemanticKITTI and nuScenes, provide dense, point-level panoptic segmentation labels, which we use for analysis and the evaluation of our method on a small set of classes that were labeled in the respective dataset (14/11 things/stuff in SemanticKITTI, respectively, and 23/6 in nuScenes). To demonstrate the versatility of SAL, we additionally pseudolabel and train a model on Waymo Open [73] dataset and show qualitative results in Fig. 1 and more in the appendix. This effort takes less than a week of pseudolabeling and model training efforts (we provide details in the appendix)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 418, + 481, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 502 + ], + "type": "text", + "content": "The good, the (breaking) bad, and the ugly. While we are thrilled about class-agnostic segmentation performance, our experimental analysis reveals a substantial gap between full supervision and our zero-shot model. Finally, at the moment, SAL is limited to training on specific datasets. As in fully-supervised Lidar perception, cross-sensor generalization remains a challenge for SAL. This could be addressed by utilizing temporal context, data scaling efforts and investigating sensor-agnostic backbone networks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 521, + 224, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 224, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 224, + 533 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 545, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 617 + ], + "type": "text", + "content": "We proposed SAL, a method for Zero-Shot Lidar Panoptic Segmentation. The key SAL components are a pseudo-label engine that utilizes vision foundation models and a zero-shot model trained via self-supervision. While SAL is the first step in the direction of learning general Lidar segmentation models by distilling image-based foundation to Lidar, we believe we just scratched the surface and opened the door for training Lidar segmentation models without manual supervision." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 620, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 481, + 665 + ], + "type": "text", + "content": "Acknowledgments. This project was funded, in parts, by ERC Starting Grant DynAI (ERC-101043189). We are grateful to Žan Gojčić, Guillem Braso, Cristiano Saltori, Sérgio Agostinho, and Jonas Schult for their feedback on the paper and their insightful comments. Special thanks to Maxim Maximov for his help on figures." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell. 34(11), 2274-2282 (2012)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Agarwalla, A., Huang, X., Ziglar, J., Ferroni, F., Laura, L.T., Hays, J., Osep, A., Ramanan, D.: Lidar panoptic segmentation and tracking without bells and whistles. In: Int. Conf. Intel. Rob. Sys. (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "text", + "content": "3. Aksoy, E.E., Baci, S., Cavidar, S.: Salsanet: Fast road and vehicle segmentation in lidar point clouds for autonomous driving. In: Intel. Veh. Symp. (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "text", + "content": "4. Aygiun, M., Osep, A., Weber, M., Maximov, M., Stachniss, C., Behley, J., Leal-Taixe, L.: 4d panoptic lidar segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 480, + 281 + ], + "type": "text", + "content": "5. Bansal, A., Sikka, K., Sharma, G., Chellappa, R., Divakaran, A.: Zero-shot object detection. In: Eur. Conf. Comput. Vis. (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 313 + ], + "type": "text", + "content": "6. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In: Int. Conf. Comput. Vis. (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "text", + "content": "7. Behley, J., Milioto, A., Stachniss, C.: A Benchmark for LiDAR-based Panoptic Segmentation based on KITTI. In: Int. Conf. Rob. Automat. (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "type": "text", + "content": "8. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. Adv. Neural Inform. Process. Syst. (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "type": "text", + "content": "9. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 380, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 412 + ], + "type": "text", + "content": "0. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Adv. Neural Inform. Process. Syst. (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "text", + "content": "1. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 468 + ], + "type": "text", + "content": "2. Choy, C., Gwak, J., Savarese, S.: 4D spatio-temporal convnets: Minkowski convolutional neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "text", + "content": "3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "text", + "content": "4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "5. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. (2009)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "text", + "content": "6. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip. In: Int. Conf. Mach. Learn. (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 578, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 598 + ], + "type": "text", + "content": "7. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Rob. Sci. Sys. (1996)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 600, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 631 + ], + "type": "text", + "content": "8. Fong, W.K., Mohan, R., Hurtado, J.V., Zhou, L., Caesar, H., Beijbom, O., Valada, A.: Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Rob. Automat. Letters (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 632, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 664 + ], + "type": "text", + "content": "9. Gasperini, S., Mahani, M.A.N., Marcos-Ramiro, A., Navab, N., Tombari, F.: Panost-ter: End-to-end panoptic segmentation of lidar point clouds. IEEE Rob. Automat. Letters (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "20. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Eur. Conf. Comput. Vis. (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "21. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "type": "text", + "content": "22. Harley, A.W., Zuo, Y., Wen, J., Mangal, A., Potdar, S., Chaudhry, R., Fragkiadaki, K.: Track, check, repeat: An em approach to unsupervised tracking. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "24. Held, D., Guillory, D., Rebsamen, B., Thrun, S., Savarese, S.: A probabilistic framework for real-time 3d segmentation using spatial, temporal, and semantic cues. In: Rob. Sci. Sys. (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "25. Held, D., Levinson, J., Thrun, S., Savarese, S.: Combining 3d shape, color, and motion for robust anytime tracking. In: Rob. Sci. Sys. (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "26. Hong, F., Zhou, H., Zhu, X., Li, H., Liu, Z.: Lidar-based panoptic segmentation via dynamic shifting network. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "27. Hu, P., Held, D., Ramanan, D.: Learning to optimally segment point clouds. IEEE Robotics and Automation Letters 5(2), 875-882 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "28. Hurtado, J.V., Mohan, R., Valada, A.: Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "29. Kirillov, A., He, K., Girshick, R.B., Rother, C., Dollar, P.: Panoptic segmentation. IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "31. Kreuzberg, L., Zulfikar, I.E., Mahadevan, S., Engelmann, F., Leibe, B.: 4d-stop: Panoptic segmentation of 4d lidar using spatio-temporal object proposal generation and aggregation. In: ECCV AVVision Workshop (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "32. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "text", + "content": "33. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "34. Li, J., He, X., Wen, Y., Gao, Y., Cheng, Y., Zhang, D.: Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "35. Li, S., Chen, X., Liu, Y., Dai, D., Stachniss, C., Gall, J.: Multi-scale interaction for real-time lidar data segmentation on an embedded platform. IEEE Rob. Automat. Letters 7(2), 738-745 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "36. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "type": "text", + "content": "37. Lin, T., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. (2014)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "38. Lin, Z., Pathak, D., Wang, Y.X., Ramanan, D., Kong, S.: Continual learning with evolving class ontologies. Adv. Neural Inform. Process. Syst. (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "39. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "40. Liu, Z., Zhang, Z., Cao, Y., Hu, H., Tong, X.: Group-free 3d object detection via transformers. In: Int. Conf. Comput. Vis. (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "41. Lu, Y., Jiang, Q., Chen, R., Hou, Y., Zhu, X., Ma, Y.: See more and know more: Zero-shot point cloud segmentation via multi-modal visual data. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "text", + "content": "42. Ma, Y., Peri, N., Wei, S., Hua, W., Ramanan, D., Li, Y., Kong, S.: Long-tailed 3d detection via 2d late fusion. arXiv preprint arXiv:2312.10986 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "43. Marcuzzi, R., Nunes, L., Wiesmann, L., Behley, J., Stachniss, C.: Mask-based panoptic lidar segmentation for autonomous driving. IEEE Rob. Automat. Letters 8(2), 1141-1148 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "text", + "content": "44. Marcuzzi, R., Nunes, L., Wiesmann, L., Marks, E., Behley, J., Stachniss, C.: Mask4d: End-to-end mask-based 4d panoptic segmentation for lidar sequences. IEEE Rob. Automat. Letters (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "45. Marcuzzi, R., Nunes, L., Wiesmann, L., Vizzo, I., Behley, J., Stachniss, C.: Contrastive instance association for 4d panoptic segmentation using sequences of 3d lidar scans. IEEE Rob. Automat. Letters (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "text", + "content": "46. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "text", + "content": "47. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: RangeNet++: Fast and Accurate LiDAR Semantic Segmentation. In: Int. Conf. Intel. Rob. Sys. (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "text", + "content": "48. Miller, D., Nicholson, L., Dayoub, F., Sünderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: Int. Conf. Rob. Automat. (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "49. Moosmann, F., Stiller, C.: Joint self-localization and tracking of generic objects in 3d range data. In: Int. Conf. Rob. Automat. (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "50. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: Eur. Conf. Comput. Vis. (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "text", + "content": "51. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "text", + "content": "52. Nunes, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Segcontrast: 3d point cloud feature representation learning through self-supervised segment discrimination. IEEE Rob. Automat. Letters 7(2), 2116-2123 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "53. Osep, A., Voigtlaender, P., Luiten, J., Breuers, S., Leibe, B.: Towards large-scale video video object mining. In: ECCV Workshop on Interactive and Adaptive Learning in an Open World (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "54. Osep, A., Mehner, W., Voigtlaender, P., Leibe, B.: Track, then decide: Category-agnostic vision-based multi-object tracking. In: Int. Conf. Rob. Automat. (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "55. Osep, A., Voigtlaender, P., Luiten, J., Breuers, S., Leibe, B.: Large-scale object mining for object discovery from unlabeled video. In: Int. Conf. Rob. Automat. (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "text", + "content": "56. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "57. Peri, N., Dave, A., Ramanan, D., Kong, S.: Towards long-tailed 3d detection. In: Conf. Rob. Learn. (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "text", + "content": "58. Peri, N., Li, M., Wilson, B., Wang, Y.X., Hays, J., Ramanan, D.: An empirical analysis of range for 3d object detection. In: ICCV Workshops (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "59. Peri, N., Luiten, J., Li, M., Osep, A., Leal-Taixe, L., Ramanan, D.: Forecasting from lidar via future object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "60. Petrovskaya, A., Thrun, S.: Model based vehicle detection and tracking for autonomous urban driving. Aut. Rob. 26, 123-139 (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "61. Pot, E., Toshev, A., Kosecka, J.: Self-supervisory signals for object discovery and detection. arXiv preprint arXiv:1806.03370 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 162, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 480, + 195 + ], + "type": "text", + "content": "62. Prest, A., Leistner, C., Civera, J., Schmid, C., Ferrari, V.: Learning object class detectors from weakly annotated video. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 196, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 480, + 228 + ], + "type": "text", + "content": "63. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "text", + "content": "64. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Adv. Neural Inform. Process. Syst. (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 252, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 480, + 285 + ], + "type": "text", + "content": "65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "text", + "content": "66. Rahman, S., Khan, S.H., Porikli, F.: Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. Asian Conf. Comput. Vis. (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "text", + "content": "67. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "text", + "content": "68. Razani, R., Cheng, R., Li, E., Taghavi, E., Ren, Y., Bingbing, L.: Gp-s3net: Graph-based panoptic sparse semantic segmentation network. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "69. Razani, R., Cheng, R., Taghavi, E., Bingbing, L.: Lite-hdseg: Lidar semantic segmentation using lite harmonic dense convolutions. In: Int. Conf. Rob. Automat. (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "text", + "content": "70. Sautier, C., Puy, G., Gidaris, S., Boulch, A., Bursuc, A., Marlet, R.: Image-to-lidar self-supervised distillation for autonomous driving data. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "text", + "content": "71. Seidenschwarz, J., Osep, A., Ferroni, F., Lucey, S., Leal-Taixe, L.: Semoli: What moves together belongs together. IEEE Conf. Comput. Vis. Pattern Recog. (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "text", + "content": "72. Sirohi, K., Mohan, R., Bülicher, D., Burgard, W., Valada, A.: Efficientlps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "type": "text", + "content": "73. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "text", + "content": "74. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "text", + "content": "75. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: Eur. Conf. Comput. Vis. (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "text", + "content": "76. Teichman, A., Levinson, J., Thrun, S.: Towards 3D object recognition via classification of arbitrary object tracks. In: Int. Conf. Rob. Automat. (2011)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "text", + "content": "77. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Int. Conf. Comput. Vis. (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "78. Thorpe, C., Herbert, M., Kanade, T., Shafer, S.: Toward autonomous driving: the cmu navlab. i. perception. IEEE expert 6(4), 31-42 (1991)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "79. Thrun, S., Montemerlo, M., Dahlkamp, H., Stavens, D., Aron, A., Diebel, J., Fong, P., Gale, J., Halpenny, M., Hoffmann, G.: Stanley: The robot that won the darpa grand challenge. Journal of field Robotics (2006)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "80. Wang, Y., Chen, X., You, Y., Li, L., Hariharan, B., Campbell, M., Weinberger, K., Chao, W.: Train in germany, test in the usa: Making 3d object detectors generalize. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "81. Wong, K., Wang, S., Ren, M., Liang, M., Urtasun, R.: Identifying unknown instances for autonomous driving. In: Conference on Robot Learning. pp. 384-393. PMLR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "text", + "content": "82. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: Int. Conf. Rob. Automat. (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "type": "text", + "content": "83. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: Int. Conf. Rob. Automat. (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 282, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 314 + ], + "type": "text", + "content": "84. Xian, Y., Lampert, C.H., Schiele, B., Akata, Z.: Zero-shot learning - a comprehensive evaluation of the good, the bad and the ugly. IEEE Trans. Pattern Anal. Mach. Intell. (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "text", + "content": "85. Xiong, X., Munoz, D., Bagnell, J.A., Hebert, M.: 3-D Scene Analysis via Sequenced Predictions over Points and Regions. In: Int. Conf. Rob. Automat. pp. 2609-2616 (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "text", + "content": "86. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "87. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "text", + "content": "88. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 482, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 457 + ], + "type": "text", + "content": "89. Yilmaz, K., Schult, J., Nekrasov, A., Leibe, B.: Mask4d: Mask transformer for 4d panoptic segmentation. arXiv preprint arXiv:2309.16133 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 457, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 479 + ], + "type": "text", + "content": "90. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "type": "text", + "content": "91. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "text", + "content": "92. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 534, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 567 + ], + "type": "text", + "content": "93. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "type": "text", + "content": "94. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: Eur. Conf. Comput. Vis. (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 589, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 482, + 611 + ], + "type": "text", + "content": "95. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 611, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 633 + ], + "type": "text", + "content": "96. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "text", + "content": "97. Zhu, M., Han, S., Cai, H., Borse, S., Ghaffari, M., Porikli, F.: 4d panoptic segmentation as invariant and equivariant field prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Segment Anything in Lidar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "text", + "content": "98. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Osep et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_content_list.json b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..571be45ed91888ece52fac6655fe4f24d9d4f593 --- /dev/null +++ b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_content_list.json @@ -0,0 +1,1738 @@ +[ + { + "type": "text", + "text": "Reg-TTA3D: Better Regression Makes Better Test-time Adaptive 3D Object Detection", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiakang Yuan $^{1}$ , Bo Zhang $^{2}$ , Kaixiong Gong $^{3}$ , Xiangyu Yue $^{3}$ , Botian Shi $^{2}$ , Yu Qiao $^{2}$ , and Tao Chen $^{1,\\text{忍}}$", + "bbox": [ + 225, + 212, + 774, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Information Science and Technology, Fudan University, Shanghai, China jkyuan22@m.fudan.edu.cn, eetchen@fudan.edu.cn", + "bbox": [ + 220, + 253, + 781, + 282 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{2}$ Shanghai Artificial Intelligence Laboratory, Shanghai, China", + "3 The Chinese University of Hong Kong, Central Ave, Hong Kong" + ], + "bbox": [ + 279, + 282, + 720, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Domain Adaptation (DA) has been widely explored and made significant progress on cross-domain 3D tasks recently. Despite being effective, existing works fail to deal with rapidly changing domains due to the unpredictable test time scenarios and meanwhile fast response time requirement. Thus, we explore a new task named test-time domain adaptive 3D object detection and propose Reg-TTA3D, a pseudo-label-based test-time adaptative 3D object detection method. By investigating the factor that limits the detection accuracy, we find that regression is essential in this task. To make better regression, we first design a noise-consistency pseudo-label generation process to filter pseudo-labels with instability under noise interference and obtain reliable pseudo-labels. Then, confidence-guided regression refinement is introduced, which uses the box regression results of high-confidence boxes to supervise boxes with relatively low confidence, further making the predicted box size gradually approach the distribution of the target domain. Finally, to better update the regression layer and alleviate the class-imbalance issue, a class-balance EMA updating strategy is proposed. Experimental results on multiple cross-domain scenarios including cross-beam, cross-location, and cross-weather demonstrate that Reg-TTA3D can achieve comparable or even better performance compared to unsupervised domain adaptation works by only updating less than $0.1\\%$ parameters within less than $1\\%$ time.", + "bbox": [ + 259, + 353, + 743, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Test-time adaptation $\\cdot$ 3D object detection $\\cdot$ Regression", + "bbox": [ + 261, + 672, + 723, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 714, + 375, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR-based 3D object detection [12, 18, 20, 21, 32, 36, 43] aims to predict the location and category of each object based on LiDAR point clouds and has attracted increasing attention recently due to its practical applications to autonomous driving. Although significant progress has been made, the stability of detectors under environmental changes, namely cross-domain 3D object detection, remains a challenge. However, the stability is important in practical usage", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27ce13672240b86d1095e3e95a07ea32faf712d1906c86435f7246bd609cfb4a.jpg", + "image_caption": [ + "Fig.1: Comparison between (a) previous domain adaptive 3D object detection and (b) Reg-TTA3D. Previous DA works perform training and testing in a split way. In contrast, Reg-TTA3D tackles domain discrepancies in an online manner and can deal with domain changes rapidly." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 477, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0efd9b555507b097d540c0eb58dbfabe98fb0ae5369fe0ac70e76194b148cc3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 143, + 785, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "since the surrounding environment dynamically changes. For example, weather and geographical environment changes may be encountered in road scenes.", + "bbox": [ + 212, + 337, + 782, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by unsupervised domain adaption (UDA) techniques in 2D cross-domain tasks [3,5,8,13,39-41,44,45], some works [6,15,30,31,33,38] begin to deal with 3D domain discrepancies. ST3D [33] is a pioneer in using self-training to gradually transfer the knowledge learned from the source domain to the target domain. ReDB [6] attempts to deal with multi-class problems by designing cross-domain examination and overlapped boxes counting. Despite achieving large performance gains, the existing UDA model still requires a long training time to be adapted to the target domain, lacking the ability to rapidly deal with environmental changes in an online manner as shown in Fig. 1. Actually, in real applications, models are expected to be continuously updated at a low cost and quickly adapt to changing domains.", + "bbox": [ + 212, + 369, + 787, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To make the model quickly adapt to the changing environment in an online manner, test-time adaptation (TTA) [1,14,16,27,28] is introduced to adapt the model during test-time with only source model and unlabeled target data. Tab. 1 shows differences between TTA and other adaptation methods. In fact, the TTA technique has been explored in 2D classification tasks such as TENT [27] and SHOT [14] recently, but it is still under-explored in 3D vision fields, especially in detection tasks. To verify the effectiveness of previous TTA methods in the domain adaptive 3D object detection tasks, we conduct experiments by directly applying such methods (e.g., TENT [27], SHOT [14], MATE [16]) to our baseline detector (i.e., SECOND-IoU [32]). As shown in Tab. 3, directly applying previous methods results in a large performance drop, even significantly behind the results of source only which directly evaluates with the pre-trained source model. For instance, TENT [27] only achieves $31.34\\%$ / $19.47\\%$ in $\\mathrm{AP}_{\\mathrm{BEV}} / \\mathrm{AP}_{\\mathrm{3D}}$ , with $15.30\\%$ / $9.36\\%$ declination comparing to source only results.", + "bbox": [ + 212, + 536, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accordingly, previous TTA methods are hard to apply to domain adaptive 3D detection tasks due to two major reasons. Firstly, from the perspective of updating parameter selection, previous methods often only update batch normalization (BN) layers to alleviate the domain shifts in the statistical information of network. However, this is not well suited for 3D detection task that also needs rich localization information. Besides, the optimization objectives of previous", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Yuan et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/516bf718f7e4fad5fa1e7c175f0778f0e14a38a174559d3bfa02a3bf2f50de67.jpg", + "table_caption": [ + "Table 1: Comparison of different settings to alleviate domain discrepancies." + ], + "table_footnote": [], + "table_body": "
SettingSource data Target data Online
Fine-tuningXxt,ytX
Unsupervised domain adaptationxs,s,ysxtX
Source-free domain adaptationXxtX
Test-time adaptationXxt✓
", + "bbox": [ + 303, + 164, + 694, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods are designed for classification tasks, which would make the model optimize for extracting features with more category semantic information rather than regression information.", + "bbox": [ + 212, + 253, + 782, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to tackle the challenges mentioned above, in this work, we propose Reg-TTA3D to perform test-time adaptation on 3D object detection tasks. To deal with the problem of updated parameter selection, we conduct extensive experiments and observe that high classification accuracy is already achieved by the source model, and the main factor limiting the performance improvement is the inaccurate regression results. Further, by only adapting regression layers, the model can achieve high performance at a low training cost as shown in Tab. 2 which motivates us to focus on regression adaptation in this work. Therefore, to make better regression, we propose noise-consistency pseudo-label generation (NPG) and confidence-guided regression refinement (CRR). The former module aims to filter boxes with instability under noise interference, i.e., noisy point clouds, while the latter one makes the box regression results align better with the distribution of the target domain box. Finally, we introduce a class-balanced EMA updating strategy (CBU) that can effectively update the regression layers, to improve the performance on multi-class simultaneously.", + "bbox": [ + 212, + 300, + 784, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contribution can be summarized as follows:", + "bbox": [ + 238, + 526, + 586, + 540 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We present a new task, namely test-time domain adaptive 3D object detection, and propose Reg-TTA3D to solve this task from the novel view of focusing on adapting only regression parameters.", + "2. We propose a noise-consistency pseudo-label generation process, a confidence-guided regression refinement loss, and a class-balanced EMA updating strategy, to obtain better detection results and boost the model's performance in an efficient and effective way.", + "3. Experiments show that Reg-TTA3D reaches a comparable or even better performance by updating less than $0.1\\%$ parameters within less than $1\\%$ time compared to UDA works, and significantly outperforms previous TTA methods on multiple cross-domain scenarios." + ], + "bbox": [ + 220, + 541, + 785, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 728, + 397, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 LiDAR-based 3D Object Detection", + "text_level": 1, + "bbox": [ + 215, + 756, + 552, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the development of autonomous driving, LiDAR-based 3D object detection has attracted increasing attention. Current prevailing LiDAR-based 3D detectors can be roughly divided into point-based, voxel-based, and point-voxel-based methods according to the point cloud processing procedure. Point-based", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reg-TTA3D", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods [20, 35, 43] first sample a subset of point clouds, then extract features and generate proposals from sampled points. Among them, PointRCNN [20] uses a two-stage framework that is composed of a bottom-up proposal generation process and a proposal refinement process. In contrast, voxel-based methods [7, 12, 32, 36, 46] first divide disordered points into regular grids which are then fed into the convolutional network. SECOND [32] is a pioneer in using the 3D sparse backbone to extract features. CenterPoint [36] introduces a center-based detector and achieves promising performance. Point-voxel-based [18, 19] methods integrate the features extracted from points and voxels to obtain better feature representations. However, existing 3D detectors fail to deal with domain gaps which is essential in practical usage.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Test-time Adaptation", + "text_level": 1, + "bbox": [ + 215, + 340, + 439, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Test-time adaptation [1, 4, 10, 22, 24, 27, 28] aims to rapidly adapt the source model to target domain without accessing source domain data during test time, which is more practical in the real world since it can handle dynamic domain shifts. TENT [27] explores test-time adaptation for the first time and updates the batch normalization layers using designed entropy minimization loss. SHOT [14] deals with such a challenging task by maximizing the mutual information. More recently, MATE [16] extends test-time adaptation to 3D point clouds and finds that the model can be adapted effectively by masking and reconstruction.", + "bbox": [ + 212, + 371, + 787, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although TTA has been extensively explored in 2D scenarios, it is still under-explored in 3D tasks, especially in large-scale point cloud scenarios (e.g., autonomous driving scenarios). Besides, most previous TTA works mainly focus on classification tasks, such as classification [1, 16, 27] and segmentation [22], and they fail to tackle detection tasks since detection considers classification and box regression at the same time. In contrast, Reg-TTA3D considers TTA for 3D object detection tasks from the perspective of regression, which is more suitable for 3D detection tasks.", + "bbox": [ + 212, + 493, + 787, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Domain Adaptation for 3D Object Detection", + "text_level": 1, + "bbox": [ + 215, + 642, + 632, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D detectors often suffer large performance drops when training and testing under different domains, which makes them unstable when the surrounding environment dynamically changes. Recently, researchers have tried to tackle such a problem using domain adaptation techniques [6,17,29,33,34,38]. SN [29] proposes a statistical method to align the size of bounding boxes between source and target domains. ST3D [33] and $\\mathrm{ST3D + + }$ [34] design a self-training pipeline and mitigate domain gaps using a memory bank and curriculum data augmentation. To deal with the class imbalance issue, ReDB [6] proposes a pseudo-label generation pipeline to obtain reliable, diverse, and class-balanced pseudo-labels. However, the existing works cannot deal with the emergency changes in the environment due to the time-consuming and large amount of calculation.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Yuan et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a3d38071571ab017ada483d50284fe3b00b27f6fff31e034fb526f4ce9211441.jpg", + "table_caption": [ + "Table 2: Performance of updating different parameters." + ], + "table_footnote": [], + "table_body": "
Tuning layersWaymo → KITTInuScenes → KITTIParameters
Batch normalization52.84 / 31.7731.86 / 18.259.08k
Classification layers52.82 / 36.3731.51 / 19.761.17k
Regression layers59.38 / 45.5433.63 / 21.802.73k
Detection head59.62 / 46.0832.51 / 20.394.68k
All parameters60.78 / 46.7133.23 / 19.6012.12M
", + "bbox": [ + 289, + 171, + 707, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 266, + 330, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall framework of Reg-TTA3D is shown in Fig. 2. To better illustrate our proposed method, we start by introducing the problem formulation and our baseline in Sec. 3.1. Then, we give an analysis of the updated parameters selection in Sec. 3.2. Finally, we detail each module of Reg-TTA3D and give the overall objective and adaptation strategy in Sec. 3.3 and Sec. 3.4.", + "bbox": [ + 212, + 297, + 787, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 393, + 361, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Formulation. Following typical test-time adaptation setting [22, 27], given the model $M_S$ pre-trained on the source domain $D_S = \\{(x_i^S, y_i^S)\\}_{i=1}^{n_S}$ and an unlabeled target set $D_T = \\{x_j^T\\}_{j=1}^{n_T}$ , where the source and target domain follow different distributions, our goal is to adapt the source model to the target domain without accessing source domain data in an online manner. During adaptation, in each timestep, a batch of unlabeled data $[x_1^t, x_2^t, \\dots, x_N^t]$ is sampled from $D_T$ to update the model, where $N$ is batch size.", + "bbox": [ + 212, + 417, + 787, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline Introduction. Following [6, 34], we use SECOND-IoU [32] as our baseline detector which is composed of a feature extractor (i.e., 3D and 2D backbone) to obtain bird's eye view (BEV) features, a region proposal network (RPN) to generate proposals and a detection head to provide the classification and regression results. The detector is trained using the following loss function:", + "bbox": [ + 212, + 523, + 787, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e t} = \\mathcal {L} _ {r p n} + \\mathcal {L} _ {r c n n}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 611, + 784, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{rpn}$ represents the loss of RPN and $\\mathcal{L}_{rcnn}$ denotes box refinement loss.", + "bbox": [ + 212, + 628, + 779, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Updated Parameters Selection", + "text_level": 1, + "bbox": [ + 215, + 664, + 513, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The aim of Reg-TTA3D is to transfer the source model to the target domain rapidly so that the model can tackle the continuous changes in the surrounding environment. To achieve this goal, the amount of updated parameters we select needs to be relatively small, and the accuracy needs to be largely improved by only updating selected parameters.", + "bbox": [ + 212, + 688, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous TTA works [22, 27] mainly focus on adapting batch normalization (BN) layers to mitigate the domain discrepancies, since BN layers contain the statistics of domains. Such an adaptation strategy is effective for domain adaptive classification tasks, as more accurate semantic representations can be obtained, which is especially important for classification. However, only adapting", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reg-TTA3D", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/28479010401af4e409f3208040817cf94873654a0ae8bdf1a5e9b88ef27a9663.jpg", + "image_caption": [ + "Fig. 2: The overall framework of the proposed Reg-TTA3D, which employs SECOND-IoU [32] as the baseline detector and consists of noisy-consistency pseudo-label generation process (NPG), confidence-guided regression refinement (CRR) and class-balanced EMA updating strategy. The data are first fed into the teacher model to generate pseudo-labels as the supervision to the student model using NPG. Then, the student model is trained using data from another view and supervised by detection loss and refinement loss obtained by CRR. Finally, the class-balanced EMA updating strategy is used to update the teacher model. Note that the only difference between the teacher model and the student model lies in the regression layers." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BN layers cannot rapidly transfer the source detector to the target domain since detection tasks contain both classification and regression. Some works [16, 26] try to solve this problem by updating the entire model but suffer from huge computational costs. To explore an effective and efficient method, by conducting extensive experiments, we find that the model can adapt to the target domain efficiently by only updating regression layers. As shown in Tab. 2, we fine-tune the parameters using $1\\%$ KITTI training data for 1 epoch and observe that only fine-tuning the regression layers can reach relatively high performance and even surpass the performance of fine-tuning the entire model. This is because in the autonomous driving scenario, there exists a large difference between categories such as car and pedestrian, the source model can already achieve high classification accuracy and the main factor for limiting the detection accuracy is the inaccurate regression results. Based on the observation, in this work, we focus on studying how to only adapt the parameters of regression layers and make better regression.", + "bbox": [ + 212, + 446, + 787, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Reg-TTA3D", + "text_level": 1, + "bbox": [ + 215, + 694, + 364, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As mentioned above, the quality of box regression is essential to TTA for 3D object detection. To make better regression, we propose noise-consistency pseudolabel generation to obtain better regression supervision, and confidence-guided regression refinement to further refine the regression results. Besides, to effectively update the regression layers, we design a class-balanced EMA updating strategy to avoid overfitting the model to a single class.", + "bbox": [ + 212, + 719, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Noise-consistency Pseudo-label Generation. As mentioned in [34, 38], the pseudo-labels directly generated by the source model contain a lot of noise due", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Yuan et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to domain discrepancies. Previous methods try to alleviate pseudo-label noise by weighting pseudo-labels obtained in different epochs [33, 34], or adding the groundtruth instances from the source domain to target domain frames [6]. However, these methods cannot apply to test-time adaptation due to the inaccessibility of source data and online adaptation.", + "bbox": [ + 212, + 146, + 784, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inspired by test-time augmentation methods which comprehensively consider the model predictions under different data augmentation methods to provide final results. Specifically, the model's prediction of real objects is relatively stable, while being uncertain for the noise regions. To this end, we propose noise-consistency pseudo-label generation to obtain more accurate pseudo-labels, by measuring the stability of model predictions when adding random noise to regions of interest (RoIs). In detail, given proposals generated by the teacher model $B^{t} = \\{(c_{x}^{t}, c_{y}^{t}, c_{z}^{t}, l^{t}, w^{t}, h^{t}, \\theta^{t})_{i}\\}_{i=1}^{n}$ , its corresponding features $\\{f_{i}^{t}\\}_{i=1}^{n}$ and confidence scores $\\{s_{i}^{t}\\}_{i=1}^{n}$ where $(c_{x}^{t}, c_{y}^{t}, c_{z}^{t})$ and $(l^{t}, w^{t}, h^{t})$ denote the center and the size of the box respectively, $\\theta^{t}$ is the rotation angle and $n$ is the number of boxes. We can obtain boxes with noise $B^{t'} = \\{(c_{x}^{t'}, c_{y}^{t'}, c_{z}^{t'}, l^{t'}, w^{t'}, h^{t'}, \\theta^{t'})\\}_{i=1}^{n}$ , corresponding features $\\{f_{i}^{t'}\\}_{i=1}^{n}$ and scores $\\{s_{i}^{t'}\\}_{i=1}^{n}$ , by adding random noise to boxes' size and angle. To measure the stability of RoIs, we calculate the inconsistency rate as follows:", + "bbox": [ + 212, + 222, + 787, + 434 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i n c o n s i s t e n c y, i} = 1 - \\frac {\\left| s _ {i} ^ {t} - s _ {i} ^ {t ^ {\\prime}} \\right|}{\\operatorname {I o U} \\left(b _ {i} ^ {t} , b _ {i} ^ {t ^ {\\prime}}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 439, + 785, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{IoU}(b_i^t, b_i^{t'})$ denotes the intersection of union (IoU) between $b_i^t \\in B^t$ and $b_i^{t'} \\in B^{t'}$ . A high inconsistency rate represents that the model prediction is unstable to the current object which is more likely to be a noise region. Therefore, we filter the boxes with high inconsistency rate by a threshold $\\tau$ to get boxes and scores $\\hat{B}, \\{\\hat{s}_i\\}_{i=1}^n$ and $\\hat{B}', \\{\\hat{s}_i'\\}_{i=1}^n$ with stability under noise interference. Then, we ensemble $\\hat{B}$ and $\\hat{B}'$ by confidence scores as follows:", + "bbox": [ + 214, + 477, + 787, + 568 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {B} = \\left\\{\\tilde {b _ {i}} \\right\\}, \\quad \\tilde {b _ {i}} = \\left\\{ \\begin{array}{l l} \\hat {b} _ {i}, & \\hat {s} _ {i} > \\hat {s} _ {i} ^ {\\prime}, \\hat {b} _ {i} \\in \\hat {B} \\\\ \\hat {b} _ {i} ^ {\\prime}, & \\hat {s} _ {i} < \\hat {s} _ {i} ^ {\\prime}, \\hat {b} _ {i} ^ {\\prime} \\in \\hat {B} ^ {\\prime} \\end{array} , \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 574, + 785, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After obtaining $\\tilde{B}$ , the boxes are filtered by non-maximum suppression (NMS) with a relatively high confidence threshold, to remove redundant boxes and boxes with low confidence. Finally, the pseudo-labels $B^{npg}$ and $\\{l_i^{npg}\\}_{i=1}^n$ with high stability and confidence can be acquired where $\\{l_i^{npg}\\}_{i=1}^n$ is classification results. Confidence-guided Regression Refinement. Although lots of boxes containing noisy point cloud can be removed by measuring the stability under noise injection, there still exist the problems of inaccurate box regression results. This is because the distributions of bounding box size are inconsistent between source and target domains as mentioned in [29,37,42], and it is difficult to mitigate such domain discrepancy with only NPG. By analyzing the distribution of predicted bounding boxes, we find that boxes with higher confidence scores align better with the distribution of the target domain boxes as shown in Fig. 3, since precise boxes correspond to high-quality RoI features that are less disturbed by background representations. Based on this observation, we design a confidence-guided regression refinement to obtain more precise regression results of box size.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reg-TTA3D", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64c255b7c3a24accc9e7467597ecaa4b71e4e0d5cc58abe748fbcd8370c03c77.jpg", + "image_caption": [ + "Fig. 3: Distribution of pseudo box size under different thresholds. PS denotes pseudo boxes." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 733, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In particular, given predicted boxes generated by the student model $B^{s} = \\{(c_{x}^{s}, c_{y}^{s}, c_{z}^{s}, l^{s}, w^{s}, h^{s}, \\theta^{s})_{j}\\}_{j=1}^{k}$ and scores $\\{s_{j}^{s}\\}_{j=1}^{k}$ of a batch of frames, where $k$ is the total number of predicted boxes. We sort boxes belonging to the same class according to scores in descending order, and sample a proportion of boxes with high scores to calculate the mean value of sample boxes size $(\\bar{l}^{s}, \\bar{w}^{s}, \\bar{h}^{s})$ . Please refer to our supplementary material for the study of the proportion of sampled boxes. Once obtained $(\\bar{l}^{s}, \\bar{w}^{s}, \\bar{h}^{s})$ , we can use it as a supervision signal to make the box regression value closer to the distribution of the target domain as follows:", + "bbox": [ + 214, + 320, + 787, + 441 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e f i n e}} = \\mathbb {E} _ {i = 1} ^ {n} \\Sigma_ {d \\in [ d _ {l}, d _ {w}, d _ {h} ]} d,\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 454, + 612, + 470 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e d = \\left\\{ \\begin{array}{l l} 0, & | \\left(z ^ {s} - \\bar {z} ^ {s}\\right) | < p \\\\ \\left(z ^ {s} - \\bar {z} ^ {s}\\right) ^ {2}, & | \\left(z ^ {s} - \\bar {z} ^ {s}\\right) | \\geq p \\end{array} , \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 473, + 784, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $z \\in [l^s, w^s, h^s]$ , $\\overline{z} \\in [\\overline{l}^s, \\overline{w}^s, \\overline{h}^s]$ , $n$ is batch size and $p$ is set 0.1 to allow the size of objects to fluctuate within a certain range. The model can predict more accurate box size when using $\\mathcal{L}_{\\text{refine}}$ for optimization.", + "bbox": [ + 214, + 521, + 787, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class-balanced EMA Updating. To make the model more stable, we update the teacher regression layers by exponential moving average (EMA). However, the model tends to overfit to the category with a large number of instances (i.e., car). As a result, the detection performance on categories with few instances will decrease. To make the model better balance the performance improvement for each class, we propose a class-balanced EMA updating strategy to make the model update more when the class is balanced, while updating less when a single category dominates. Specifically, given the pseudo-labels obtained from NPG $B^{npg} = \\{b_i^{npg} \\in \\mathbb{R}^7\\}$ and its corresponding class labels $\\{c_i^{npg}\\}$ . We calculate the proportion $\\{p_c\\}_{c=1}^C$ of each category in pseudo-labels according to $\\{c_i^{npg}\\}$ , where $C$ is the number of total categories and $\\Sigma_{c=1}^C p_c = 1$ . Then the variance of $\\{p_c\\}_{c=1}^C$ is calculated and normalized to [0.99, 0.999] to obtain the momentum $\\alpha$ . A high value of $\\alpha$ means that a severe class-imbalance issue exists in the current batch and the model should update less. Finally the parameter of the teacher regression layer $\\omega_{reg}^t$ can be updated by the following formula:", + "bbox": [ + 214, + 566, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {r e g} ^ {t} \\leftarrow \\alpha \\omega_ {r e g} ^ {t} + (1 - \\alpha) \\omega_ {r e g} ^ {s}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 806, + 784, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\omega_{reg}^{s}$ is the parameters of student regression layers.", + "bbox": [ + 214, + 828, + 625, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Yuan et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Algorithm 1 Single iteration of Reg-TTA3D strategy", + "text_level": 1, + "bbox": [ + 217, + 146, + 609, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Input: A batch of frames $\\{x_{i}\\}_{i = 1}^{n}$ and $\\{\\dot{x}_i\\}_{i = 1}^n$ , where $\\dot{x}_i$ is obtained by performing data augmentation on $x_{i}$", + "bbox": [ + 222, + 162, + 619, + 189 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Output: Predictions of current batch $B^{t}$ and $\\{l_i^t\\}_{i = 1}^n$", + "bbox": [ + 222, + 190, + 580, + 205 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Get the predictions $B^{t} = \\{b_{i}^{t}\\in \\mathbb{R}^{7}\\}_{i = 1}^{n}$ and $\\{l_i^t\\}_{i = 1}^n$ of $\\{x_{i}\\}_{i = 1}^{n}$ using teacher model, $B^{t}$ and $\\{l_i^t\\}_{i = 1}^n$ are the detection results of current batch.", + "2: Use NPG to obtain boxes $B^{npg}$ and $\\{l_i^{npg}\\}_{i=1}^n$ with stability under noise interference as pseudo-labels.", + "3: Get the predictions of student model $B^{s}$ and $\\{s_{i}^{s}\\}_{i = 1}^{n}$ of $\\{\\dot{x}_i\\}_{i = 1}^n$ , supervise the student model by $B^{npg}$ and $\\{l_i^{npg}\\}_{i = 1}^n$ to get $\\mathcal{L}_{det}$ .", + "4: Calculate $\\mathcal{L}_{\\text{refine}}$ using CRR.", + "5: Optimize the student model using Eq. (6).", + "6: Update the teacher model using CBU and Eq. (5).", + "7: Return: Current predictions $B^{t}$ and $\\{l_i^t\\}_{i = 1}^n$ ." + ], + "bbox": [ + 222, + 205, + 785, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Overall Objective and Test-time Adaptation Strategy", + "text_level": 1, + "bbox": [ + 215, + 378, + 700, + 395 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Overall Objective. The overall objective to optimize the student regression layers can be formulated as follows:", + "bbox": [ + 212, + 404, + 785, + 434 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {d e t}} + \\mathcal {L} _ {\\text {r e f i n e}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 446, + 785, + 464 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{det}$ and $\\mathcal{L}_{refine}$ are defined in Eq. (1) and Eq. (4).", + "bbox": [ + 212, + 472, + 630, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Test-time Adaptation Strategy. The test-time adaptation strategy of Reg-TTA3D can be summarized in Algorithm 1.", + "bbox": [ + 212, + 488, + 784, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 542, + 375, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 573, + 426, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets and Cross-domain Settings. We conduct experiments on four commonly-used datasets: Waymo [23], KITTI [9], nuScenes [2] and nuScenes-C [11]. To verify the effectiveness of Reg-TTA3D, we consider three cross-domain settings including cross-location (i.e., Waymo $\\rightarrow$ KITTI), cross-beam (i.e., Waymo $\\rightarrow$ nuScenes, nuScenes $\\rightarrow$ KITTI) and cross-weather (i.e., nuScenes $\\rightarrow$ nuScenes-C). Note that in the cross-weather setting, we use the medium-difficulty fog and snow scenarios in the nuScenes-C dataset. Following ReDB [6], in the first two cross-domain settings, we use the KITTI evaluation metric to evaluate for three common classes in autonomous driving scenarios (i.e., car, pedestrian, and cyclist), we report the average precision in both 3D (i.e., $\\mathrm{AP}_{\\mathrm{3D}}$ ) and BEV (i.e., $\\mathrm{AP}_{\\mathrm{BEV}}$ ) over 40 recall positions, with IoU threshold 0.7 for car and 0.5 for pedestrian and cyclist. In the cross-weather setting, we use the official nuScenes evaluation metric and report mean average precision (mAP).", + "bbox": [ + 212, + 598, + 794, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. Following [6], we evaluate the proposed Reg-TTA3D on SECOND-IoU [32], and the source model is trained for all categories using random object scaling (ROS) augmentation method. We set hyperparameter $\\tau$", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reg-TTA3D", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to 1.5 and $p$ to 0.1 for all cross-domain settings and use Adam optimizer with learning rate of $1 \\times 10^{-3}$ . We utilize commonly-used data augmentation methods (e.g., random world flip, random world rotation, random world scaling) to get $\\{\\dot{x}_i\\}_{i=1}^n$ . Our code is built on 3DTrans [25].", + "bbox": [ + 212, + 146, + 785, + 207 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison Baselines. To verify the effectiveness of Reg-TTA3D, we mainly compare our methods to unsupervised domain adaptive 3D object detection and previous test-time adaptation methods that we apply to 3D object detection. UDA methods: (1) ST3D [33] is a self-training method that continuously updates the pseudo-labels, (2) $\\mathrm{ST3D + + }$ [34] provides more insight analysis and extends ST3D to get better results, (3) ReDB [6] is a pioneer to perform domain adaptive 3D detection on multiple categories. TTA methods: (1) TENT [27] proposes an entropy minimization method which focuses on adapting BN layers, (2) SHOT [14] optimizes feature extraction module using both information maximization and pseudo-labeling. (3) MATE [16] uses masked autoencoder to apply test-time adaptation to 3D classification.", + "bbox": [ + 212, + 207, + 785, + 373 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 393, + 372, + 407 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on cross-location and cross-beam scenarios. We first evaluate Reg-TTA3D on cross-location and cross-beam scenarios and compare the results with previous UDA methods. As shown in Tab. 3, Reg-TTA3D can greatly improve the detection accuracy on Waymo $\\rightarrow$ KITTI and nuScenes $\\rightarrow$ KITTI settings, largely narrow the performance gap between source only and oracle. For example, in the Waymo $\\rightarrow$ KITTI setting, our Reg-TTA3D can achieve $61.78\\%$ / $49.96\\%$ in $\\mathrm{AP}_{\\mathrm{BEV}} / \\mathrm{AP}_{\\mathrm{3D}}$ , even surpass UDA methods such as ST3D [33] and ST3D++ [34] and it is comparable to the state-of-the-art UDA method (i.e., $61.14\\%$ / $50.10\\%$ in $\\mathrm{AP}_{\\mathrm{BEV}} / \\mathrm{AP}_{\\mathrm{3D}}$ reported in ReDB [6]). In nuScenes $\\rightarrow$ KITTI and Waymo $\\rightarrow$ nuScenes cross-domain scenarios, the performance of Reg-TTA3D is also better than some UDA methods (e.g., ST3D [33]) which needs source domain data and training for multiple epochs. In contrast, Reg-TTA3D only needs target data and one epoch of adaptation which greatly reduces the time consumption.", + "bbox": [ + 212, + 417, + 787, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Further, we compare Reg-TTA3D with the previous TTA methods in Tab. 3. We find that previous TTA methods suffer performance drops when applied to 3D detection tasks (e.g., TENT [27] reduce the detection accuracy to $31.34\\%$ / $19.47\\%$ in the Waymo $\\rightarrow$ KITTI setting). This is mainly because the detection task is composed of classification and regression and existing methods only consider the classification tasks. As a result, the model will concentrate on extracting effective representations for the classification tasks and ignore the localization information. However, our Reg-TTA3D fully considers the challenges in detection tasks and can improve the performance on multiple categories.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on the cross-weather scenario. To more comprehensively evaluate the effectiveness of our proposed method, we conduct experiments in a more realistic scenario (i.e., cross-weather). Since nuScenes-C dataset only has the validation set, we only compare Reg-TTA3D with existing TTA methods. It can be seen in Tab. 4 that Reg-TTA3D can improve the performance on both snow", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Yuan et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/312fd5513c337188287bdf4b09a26cff50959b10f35f873a37a7f37fc4ac398e.jpg", + "table_caption": [ + "Table 3: Results on cross-location and cross-beam scenarios. Following [6, 33], we use SECOND-IoU [32] as out baseline detector. Source Only represents that the pretrained detector is directly evaluated on the target domain, and Oracle denotes the detection results obtained using the fully annotated target training data. Red: Best UDA method, Blue: Best TTA method." + ], + "table_footnote": [], + "table_body": "
TASKMETHODSETTINGCARPEDESTRIANCYCLISTMEAN AP
Waymo → KITTISOURCE ONLY-51.48 / 19.7840.80 / 31.2647.63 / 35.4546.64 / 28.83
SN76.61 / 54.1452.48 / 48.2034.56 / 32.7454.55 / 45.03
ST3DUDA77.62 / 49.2444.45 / 42.0447.74 / 42.9556.60 / 44.70
ST3D++77.68 / 50.0349.09 / 46.1951.50 / 47.7059.42 / 47.97
ReDB80.37 / 54.1251.01 / 48.2052.05 / 47.9761.14 / 50.10
TENTTTA71.98 / 39.014.54 / 3.9517.51 / 15.4531.34 / 19.47
SHOT79.52 / 41.3133.52 / 29.0240.79 / 35.8351.27 / 35.38
MATE75.17 / 30.6439.94 / 36.9243.72 / 40.2252.94 / 35.92
REG-TTA3D81.60 / 56.0350.89 / 45.4352.85 / 48.4361.78 / 49.96
ORACLE-83.29 / 73.4546.64 / 41.3362.92 / 60.3264.28 / 58.37
Waymo → nuScenesSOURCE ONLY-30.64 / 17.181.81 / 0.580.97 / 0.8811.14 / 6.22
SN29.56 / 17.781.33 / 1.072.92 / 2.6111.27 / 7.15
ST3DUDA28.42 / 17.831.64 / 1.394.01 / 3.5411.36 / 7.59
ST3D++28.87 / 19.151.82 / 1.584.09 / 3.7411.60 / 8.16
ReDB30.12 / 18.562.47 / 2.146.56 / 5.1913.05 / 8.63
TENTTTA16.40 / 3.430.82 / 0.250.28 / 0.215.83 / 1.29
SHOT12.26 / 1.421.56 / 1.360.22 / 0.204.68 / 0.99
MATE8.64 / 3.991.72 / 0.970.24 / 0.083.53 / 1.68
REG-TTA3D30.54 / 18.152.04 / 1.755.01 / 4.7012.53 / 8.20
ORACLE-51.88 / 34.8725.24 / 18.9215.06 / 11.7330.73 / 21.84
nuScenes → KITTISOURCE ONLY-39.15 / 7.6521.54 / 16.876.31 / 2.4422.33 / 8.98
SN56.08 / 28.6723.05 / 16.845.11 / 2.3228.08 / 15.94
ST3DUDA71.50 / 48.0922.64 / 17.617.86 / 5.2034.00 / 23.64
ST3D++69.90 / 44.6224.11 / 18.2010.14 / 6.3933.75 / 23.07
ReDB74.23 / 51.3125.95 / 18.3813.82 / 8.6438.00 / 26.11
TENTTTA43.35 / 19.839.40 / 5.772.71 / 2.7018.49 / 9.43
SHOT60.24 / 23.0117.51 / 9.442.47 / 1.6626.74 / 11.37
MATE61.89 / 27.1315.98 / 13.112.94 / 2.4526.93 / 14.23
REG-TTA3D68.73 / 44.5622.78 / 17.7110.21 / 7.1233.90 / 23.13
ORACLE-83.29 / 73.4546.64 / 41.3362.92 / 60.3264.28 / 58.37
", + "bbox": [ + 220, + 229, + 782, + 608 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "and fog scenarios (e.g., $45.68\\% \\rightarrow 47.09\\%$ mAP on the Normal $\\rightarrow$ Snow setting). The results verify that our proposed method can handle dynamic environment changes in road scenes and can be applied in more practical scenarios.", + "bbox": [ + 212, + 619, + 787, + 667 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Insight Analyses", + "text_level": 1, + "bbox": [ + 215, + 691, + 398, + 707 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sensitivity to Detector Architecture. Following [6], to verify the effectiveness of Reg-TTA3D on different types of detectors, we further conduct experiments on nuScenes $\\rightarrow$ KITTI cross-domain scenario using PointRCNN [20] as the baseline detector. As show in Tab. 5, our Reg-TTA3D outperforms all TTA methods and some UDA methods. Note that compared to the source-free domain adaptation method (i.e., SF-UDA $^{3D}$ [17]) which is inaccessible to the source data and need to adapt for multiple epochs, our Reg-TTA3D adapted on multi-class can achieve a better result on the car category (e.g., $50.87\\%$ compared to $49.80\\%$", + "bbox": [ + 212, + 718, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reg-TTA3D", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7083bd166e04f7b904cfe18fb4f0e134ee8e57fbe56d3bdf8f80cae621748033.jpg", + "table_caption": [ + "Table 4: Results on cross-weather scenario (i.e., nuScenes Normal $\\rightarrow$ Fog and Normal $\\rightarrow$ Snow). Fog and snow data are from nuScenes-C [11] dataset. We use CenterPoint [36] as our baseline detector since it is the most widely used detector on nuScenes." + ], + "table_footnote": [], + "table_body": "
METHODTaskmAPCARTRUCKCV.BUSTRAILERBARRIERMOTOR.BICYCLEPED.TC.
SOURCE ONLY48.4079.1252.588.1067.5331.0064.407.3026.6581.2865.97
TENT41.8373.6215.237.5858.0421.8560.451.5623.8477.5952.24
SHOTFog42.3172.6340.707.5958.1822.2461.710.8526.4577.8454.87
MATE23.9764.6124.183.3427.2614.1120.450.007.6749.6728.33
REG-TTA3D49.6780.3054.4310.8668.6534.5363.527.7927.0782.2067.28
SOURCE ONLY45.8680.7448.774.6864.1528.8462.629.0224.7278.3756.62
TENTSnow40.5175.0939.376.5157.5719.9659.821.8923.3376.7844.73
SHOT42.1574.6639.747.0158.1221.4662.941.1927.3078.3950.66
MATE23.4363.9822.573.2025.3013.0622.790.007.640.5024.81
REG-TTA3D47.0982.0149.717.1968.1630.5462.966.3326.2980.3757.30
", + "bbox": [ + 220, + 191, + 781, + 325 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/14237526b6157033b1f02026eca35937362ee18b9e045515e4f9ce1a69abd8c7.jpg", + "table_caption": [ + "Table 5: Performance comparisons (AP3D) with PointRCNN as baseline detector on nuScenes $\\rightarrow$ KITTI task. $\\ddagger$ indicates the results reported in the original paper. Red: Best UDA method, Blue: Best TTA method." + ], + "table_footnote": [], + "table_body": "
MethodSettingCARPEDESTRIANCYCLISTAVERAGE
EASYMOD.HARDEASYMOD.HARDEASYMOD.HARDEASYMOD.HARD
SOURCE ONLY-42.7732.1128.7543.2637.2933.163.284.094.1829.7724.6022.03
SN66.5650.3245.9242.9637.1532.459.077.577.4239.5331.6828.60
ST3DUDA48.8541.9038.9245.6738.7133.0926.5019.3518.3840.3433.3230.13
ST3D++60.4549.3645.8850.7742.4336.6427.2017.9416.5246.1436.5833.01
REDB71.4557.953.9152.3244.3337.9545.1332.9331.0556.3045.0540.97
SF-UDA3D‡SFDA68.8049.8045.00---------
TENTTTA42.5532.4127.6130.7126.0923.784.243.663.3225.8320.7018.23
SHOT30.0923.0920.4119.2416.2914.312.431.701.5117.2513.6912.07
REG-TTA3D65.2050.8746.7245.8042.5534.9040.2527.8725.2950.4140.4335.63
", + "bbox": [ + 220, + 388, + 781, + 534 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "reported by SF-UDA $^{3D}$ in moderate level) in one epoch of adaptation even if SF-UDA $^{3D}$ is trained on the unfair single class. The results further show that our proposed methods can apply to different detector architectures.", + "bbox": [ + 212, + 549, + 785, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Studies. To further verify the effectiveness of Reg-TTA3D, we conduct extensive experiments to investigate the impact of each component we proposed including NPG, CRR and CBU. As shown in Tab. 6, the improvement (i.e., $0.92\\% \\mathrm{AP}_{3\\mathrm{D}}$ ) of using naive pseudo-labels is relatively small compared to the performance of source only since pseudo-labels contain lots of noise samples and the regression results are not accurate due to domain discrepancies. Reg-TTA can effectively handle such a problem by proposing NPG and CRR modules, with $2.90\\%$ and $3.60\\%$ improvement respectively compared to source only. Besides, when combining these two modules, the performance can be further improved because NPG focuses on filter noise samples and CRR concentrates on making more accurate regression results. Further, by combining with CBU which makes the model update in a class-balance manner, the performance can be further boosted and achieve $49.96\\% \\mathrm{AP}_{3\\mathrm{D}}$ .", + "bbox": [ + 212, + 595, + 785, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sensitivity to Hyperparameters. We ablate the impact of hyperparameters $\\tau$ (i.e., the threshold of inconsistency rate) and $p$ in this subsection. As shown in Tab. 7, we first set $\\tau$ to 1.0, 1.5, 2.0 respectively and observe that the fluctuation", + "bbox": [ + 214, + 794, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Yuan et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a0a0ab4ff1ddb621ab8788b162acd6f4dff89c20e643d7db92d925722b5c8dcb.jpg", + "table_caption": [ + "Table 6: Ablation studies on each component. PS, NPG, CRR and CBU represent using naive pseudo-labels, noise-consistency pseudo-label generation, confidence-guided regression refinement and class-balanced EMA updating, respectively. The experiments are conducted on Waymo $\\rightarrow$ KITTI scenario using SECOND-IoU as the baseline detector." + ], + "table_footnote": [], + "table_body": "
PS NPGCRRCBUmean APBEVmean AP3D
---57.8942.01
✓--58.7342.93
✓✓-58.8944.91
✓-✓59.8445.61
✓--59.7344.05
✓✓✓60.5747.97
✓✓-60.1246.62
✓-✓60.5948.39
✓✓✓61.7849.96
", + "bbox": [ + 246, + 250, + 539, + 388 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/92a2f7351331055a74914c034f5f01733bd036a2ce8e0992cd5d302ec4cc9b7f.jpg", + "table_caption": [ + "Table 7: Ablation studies of hyperparameters $\\tau$ and $p$ . The experiments are conducted on Waymo $\\rightarrow$ KITTI scenario using SECOND-IoU. Note that when ablate on $\\tau$ , we set $p = 0.1$ and when ablate on $p$ , we set $\\tau = 0.15$ ." + ], + "table_footnote": [], + "table_body": "
τAPBEV / AP3D
1.060.92 / 49.12
1.561.78 / 49.96
2.060.44 / 48.88
pAPBEV / AP3D
0.0560.29 / 48.26
0.1061.78 / 49.96
0.1561.56 / 49.63
", + "bbox": [ + 617, + 258, + 751, + 388 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8d3ba54ca5c5066e56034224777598f6050e0ca924f67c7e7ed8926fd83e99d6.jpg", + "image_caption": [ + "Fig. 4: The relationship between detection accuracy and the iteration number." + ], + "image_footnote": [], + "bbox": [ + 217, + 406, + 496, + 525 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/94ce73de8fc48177ceaf95ac06c30abe40bf625c5135ce68dd85348608d48611.jpg", + "table_caption": [ + "Table 8: Ablation studies on costs. The experiments are conducted on $\\mathrm{W} \\rightarrow \\mathrm{K}$ scenario using SECOND-IoU. We use Tesla A100 to conduct the experiments." + ], + "table_footnote": [], + "table_body": "
MethodTime Cost N_GPUsParam.Results
ST3D4h412.12M 56.60 / 44.70
ST3D++3.5h412.12M 59.42 / 47.97
ReDB3.2h412.12M 61.14 / 50.10
TENT4min19.08K 31.34 / 19.47
MATE6min10.71M 52.94 / 35.92
Reg-TTA3D4.5min12.73K 61.78 / 49.96
", + "bbox": [ + 503, + 479, + 777, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of the performance is $1.34\\% / 1.08\\%$ . Then, by setting $p$ to different values (i.e., 0.05, 0.1, 0.15), we find that there is only little fluctuation and the best performance can be achieved when set to 0.1. The results further show that our method is robust to these hyperparameters.", + "bbox": [ + 212, + 578, + 785, + 640 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis on Pseudo-labeling Accuracy. In this section, we measure the accuracy of each iteration during adaptation. Fig. 4 reflects the detection accuracy on Waymo $\\rightarrow$ KITTI as iteration increases. It can be seen that the accuracy continuously improves since the model is gradually adapted to the target domain which also verifies the stability of Reg-TTA3D during adaptation.", + "bbox": [ + 212, + 641, + 787, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis on Costs. Further, we report the time cost and total updated parameters in Tab. 8. Notably, Reg-TTA3D achieves comparable results to UDA methods adapting only less than $0.1\\%$ of parameters within less than $1\\%$ of time. When compared to TTA methods, our proposed method can achieve much higher performance with less updated parameters and comparable time cost which further verifies that our Reg-TTA3D is an effective and efficient method.", + "bbox": [ + 212, + 717, + 787, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation Studies of the momentum. We ablate the momentum in EMA updating and compare it to our class-balance EMA updating strategy. Specifi-", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reg-TTA3D", + "bbox": [ + 645, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0b0854b088a69e6979bc96ee42082480df6f82a4057f122ec79e13a8aee0450c.jpg", + "image_caption": [ + "Fig. 5: The impact of momentum and com- Fig.6: ablation studies on the proportions with CBU. tion of selected samples." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 480, + 290 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5d8d709697e33b8ee75c10460202c65a1f9776a3559567967521f95982b8025b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 142, + 764, + 290 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "cally, we conduct the experiment for momentum between 0.99 and 0.999 at 0.001 intervals. As shown in Fig. 5, the result demonstrates that CBU can better update the regression layer due to it can balance update the model and improve the performance on multiple classes at the same time.", + "bbox": [ + 212, + 340, + 784, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation Studies on the proportion of Selected samples in CRR. In our manuscript, we use $20\\%$ boxes with the highest scores to supervise box sizes with relatively low scores. Here, we analyze the impact on the proportion of selected samples. As shown in Fig. 6, when the proportion is relatively high, the average box size will be distributed far from the target domain, and when set to relatively small values, the small number of samples results in the averages being affected by individual instances. Based on the experiments and analysis, we set the proportion to $20\\%$ .", + "bbox": [ + 212, + 402, + 785, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 553, + 357, + 569 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, for the first time, we explore test-time adaptation in 3D object detection tasks and propose Reg-TTA3D, which handles rapid changes in the surrounding environment from the perspective of updating parameters selection and adaptation algorithm. By adapting using the proposed noise-consistency pseudo-label generation process, confidence-guided regression refinement loss and class-balanced EMA updating strategy and only updating regression layers, Reg-TTA3D can achieve comparable performance to some UDA methods at a low cost.", + "bbox": [ + 212, + 590, + 787, + 712 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitation", + "text_level": 1, + "bbox": [ + 215, + 741, + 354, + 757 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Although the proposed Reg-TTA3D can handle dynamic changes in the surrounding environments such as continuously changing weather. It cannot deal with emergencies such as some corner cases which is also important in real-world autonomous driving.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Yuan et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 393, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by National Key Research and Development Program of China (No. 2022ZD0160101), National Natural Science Foundation of China (No. 62071127, and 62101137), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No.2021SHZDZX0103). This work is also supported by the National Key R&D Program of China (Grant No. 2022ZD0160104), and Shanghai Rising Star Program (Grant No. 23QD1401000). The computations in this research were performed using the CFFF platform of Fudan University.", + "bbox": [ + 212, + 179, + 808, + 301 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 327, + 321, + 343 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8344-8353 (2022)", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)", + "3. Cai, Q., Pan, Y., Ngo, C.W., Tian, X., Duan, L., Yao, T.: Exploring object relation in mean teacher for cross-domain detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11457-11466 (2019)", + "4. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 295-305 (2022)", + "5. Chen, Y., Li, W., Sakaridis, C., Dai, D., Van Gool, L.: Domain adaptive faster r-cnn for object detection in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3339-3348 (2018)", + "6. Chen, Z., Luo, Y., Huang, Z., Wang, Z., Baktashmotlagh, M.: Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudolabeling. arXiv preprint arXiv:2307.07944 (2023)", + "7. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1201-1209 (2021)", + "8. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: International conference on machine learning. pp. 1180-1189. PMLR (2015)", + "9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (2012)", + "10. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems 35, 6204-6218 (2022)", + "11. Kong, L., Liu, Y., Li, X., Chen, R., Zhang, W., Ren, J., Pan, L., Chen, K., Liu, Z.: Robo3d: Towards robust and reliable 3d perception against corruptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 19994-20006 (2023)" + ], + "bbox": [ + 217, + 361, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reg-TTA3D", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)", + "13. Liang, J., Gong, K., Li, S., Liu, C.H., Li, H., Liu, D., Wang, G., et al.: Pareto domain adaptation. Advances in Neural Information Processing Systems 34, 12917-12929 (2021)", + "14. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)", + "15. Luo, Z., Cai, Z., Zhou, C., Zhang, G., Zhao, H., Yi, S., Lu, S., Li, H., Zhang, S., Liu, Z.: Unsupervised domain adaptive 3d detection with multi-level consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8866-8875 (2021)", + "16. Mirza, M.J., Shin, I., Lin, W., Schriebl, A., Sun, K., Choe, J., Kozinski, M., Possegger, H., Kweon, I.S., Yoon, K.J., et al.: Mate: Masked autoencoders are online 3d test-time learners. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16709-16718 (2023)", + "17. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771-780. IEEE (2020)", + "18. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529–10538 (2020)", + "19. Shi, S., Jiang, L., Deng, J., Wang, Z., Guo, C., Shi, J., Wang, X., Li, H.: Pvrccnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. International Journal of Computer Vision 131(2), 531-551 (2023)", + "20. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)", + "21. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence 43(8), 2647-2664 (2020)", + "22. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022)", + "23. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2446-2454 (2020)", + "24. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)", + "25. Team, D.D.: 3dtrans: An open-source codebase for exploring transferable autonomous driving perception task. https://github.com/PJLab-ADG/3DTrans (2023)", + "26. VS, V., Oza, P., Patel, V.M.: Towards online domain adaptive object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 478-488 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Yuan et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726 (2020)", + "28. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7201-7211 (2022)", + "29. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in germany, test in the usa: Making 3d object detectors generalize. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11713-11723 (2020)", + "30. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. arXiv preprint arXiv:2203.14956 (2022)", + "31. Xu, Q., Zhou, Y., Wang, W., Qi, C.R., Anguelov, D.: Spg: Unsupervised domain adaptation for 3d object detection via semantic point generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15446-15456 (2021)", + "32. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)", + "33. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10368-10378 (2021)", + "34. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d++: Denoised self-training for unsupervised domain adaptation on 3d object detection. IEEE transactions on pattern analysis and machine intelligence 45(5), 6354-6371 (2022)", + "35. Yang, Z., Sun, Y., Liu, S., Shen, X., Jia, J.: Std: Sparse-to-dense 3d object detector for point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1951-1960 (2019)", + "36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2021)", + "37. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Ad-pt: Autonomous driving pre-training with large-scale point cloud dataset. arXiv preprint arXiv:2306.00612 (2023)", + "38. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Bi3d: Bidomain active learning for cross-domain 3d object detection. arXiv preprint arXiv:2303.05886 (2023)", + "39. Yue, X., Zhang, Y., Zhao, S., Sangiovanni-Vincentelli, A., Keutzer, K., Gong, B.: Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2100-2110 (2019)", + "40. Yue, X., Zheng, Z., Zhang, S., Gao, Y., Darrell, T., Keutzer, K., Vincentelli, A.S.: Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13834-13844 (2021)", + "41. Zhang, B., Chen, T., Wang, B., Li, R.: Joint distribution alignment via adversarial learning for domain adaptive object detection. IEEE Transactions on Multimedia (2021)", + "42. Zhang, B., Yuan, J., Shi, B., Chen, T., Li, Y., Qiao, Y.: Uni3d: A unified baseline for multi-dataset 3d object detection. arXiv preprint arXiv:2303.06880 (2023)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reg-TTA3D", + "bbox": [ + 645, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022)", + "44. Zhao, S., Li, B., Yue, X., Gu, Y., Xu, P., Hu, R., Chai, H., Keutzer, K.: Multi-source domain adaptation for semantic segmentation. Advances in neural information processing systems 32 (2019)", + "45. Zheng, Y., Huang, D., Liu, S., Wang, Y.: Cross-domain object detection through coarse-to-fine feature adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13766-13775 (2020)", + "46. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 327 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Yuan et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_model.json b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6474ce267b8c82260901cc07601dfa42008a1f5e --- /dev/null +++ b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_model.json @@ -0,0 +1,2392 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.242, + 0.141, + 0.761, + 0.186 + ], + "angle": 0, + "content": "Reg-TTA3D: Better Regression Makes Better Test-time Adaptive 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.213, + 0.776, + 0.244 + ], + "angle": 0, + "content": "Jiakang Yuan\\(^{1}\\), Bo Zhang\\(^{2}\\), Kaixiong Gong\\(^{3}\\), Xiangyu Yue\\(^{3}\\), Botian Shi\\(^{2}\\), Yu Qiao\\(^{2}\\), and Tao Chen\\(^{1,\\text{忍}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.255, + 0.782, + 0.283 + ], + "angle": 0, + "content": "1 School of Information Science and Technology, Fudan University, Shanghai, China jkyuan22@m.fudan.edu.cn, eetchen@fudan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.283, + 0.71, + 0.298 + ], + "angle": 0, + "content": "\\(^{2}\\) Shanghai Artificial Intelligence Laboratory, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.298, + 0.721, + 0.312 + ], + "angle": 0, + "content": "3 The Chinese University of Hong Kong, Central Ave, Hong Kong" + }, + { + "type": "list", + "bbox": [ + 0.281, + 0.283, + 0.721, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.354, + 0.744, + 0.658 + ], + "angle": 0, + "content": "Abstract. Domain Adaptation (DA) has been widely explored and made significant progress on cross-domain 3D tasks recently. Despite being effective, existing works fail to deal with rapidly changing domains due to the unpredictable test time scenarios and meanwhile fast response time requirement. Thus, we explore a new task named test-time domain adaptive 3D object detection and propose Reg-TTA3D, a pseudo-label-based test-time adaptative 3D object detection method. By investigating the factor that limits the detection accuracy, we find that regression is essential in this task. To make better regression, we first design a noise-consistency pseudo-label generation process to filter pseudo-labels with instability under noise interference and obtain reliable pseudo-labels. Then, confidence-guided regression refinement is introduced, which uses the box regression results of high-confidence boxes to supervise boxes with relatively low confidence, further making the predicted box size gradually approach the distribution of the target domain. Finally, to better update the regression layer and alleviate the class-imbalance issue, a class-balance EMA updating strategy is proposed. Experimental results on multiple cross-domain scenarios including cross-beam, cross-location, and cross-weather demonstrate that Reg-TTA3D can achieve comparable or even better performance compared to unsupervised domain adaptation works by only updating less than \\(0.1\\%\\) parameters within less than \\(1\\%\\) time." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.673, + 0.724, + 0.687 + ], + "angle": 0, + "content": "Keywords: Test-time adaptation \\(\\cdot\\) 3D object detection \\(\\cdot\\) Regression" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.715, + 0.377, + 0.731 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "LiDAR-based 3D object detection [12, 18, 20, 21, 32, 36, 43] aims to predict the location and category of each object based on LiDAR point clouds and has attracted increasing attention recently due to its practical applications to autonomous driving. Although significant progress has been made, the stability of detectors under environmental changes, namely cross-domain 3D object detection, remains a challenge. However, the stability is important in practical usage" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Yuan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.478, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.145, + 0.787, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig.1: Comparison between (a) previous domain adaptive 3D object detection and (b) Reg-TTA3D. Previous DA works perform training and testing in a split way. In contrast, Reg-TTA3D tackles domain discrepancies in an online manner and can deal with domain changes rapidly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.784, + 0.368 + ], + "angle": 0, + "content": "since the surrounding environment dynamically changes. For example, weather and geographical environment changes may be encountered in road scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.789, + 0.536 + ], + "angle": 0, + "content": "Inspired by unsupervised domain adaption (UDA) techniques in 2D cross-domain tasks [3,5,8,13,39-41,44,45], some works [6,15,30,31,33,38] begin to deal with 3D domain discrepancies. ST3D [33] is a pioneer in using self-training to gradually transfer the knowledge learned from the source domain to the target domain. ReDB [6] attempts to deal with multi-class problems by designing cross-domain examination and overlapped boxes counting. Despite achieving large performance gains, the existing UDA model still requires a long training time to be adapted to the target domain, lacking the ability to rapidly deal with environmental changes in an online manner as shown in Fig. 1. Actually, in real applications, models are expected to be continuously updated at a low cost and quickly adapt to changing domains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.789, + 0.749 + ], + "angle": 0, + "content": "To make the model quickly adapt to the changing environment in an online manner, test-time adaptation (TTA) [1,14,16,27,28] is introduced to adapt the model during test-time with only source model and unlabeled target data. Tab. 1 shows differences between TTA and other adaptation methods. In fact, the TTA technique has been explored in 2D classification tasks such as TENT [27] and SHOT [14] recently, but it is still under-explored in 3D vision fields, especially in detection tasks. To verify the effectiveness of previous TTA methods in the domain adaptive 3D object detection tasks, we conduct experiments by directly applying such methods (e.g., TENT [27], SHOT [14], MATE [16]) to our baseline detector (i.e., SECOND-IoU [32]). As shown in Tab. 3, directly applying previous methods results in a large performance drop, even significantly behind the results of source only which directly evaluates with the pre-trained source model. For instance, TENT [27] only achieves \\(31.34\\%\\) / \\(19.47\\%\\) in \\(\\mathrm{AP}_{\\mathrm{BEV}} / \\mathrm{AP}_{\\mathrm{3D}}\\), with \\(15.30\\%\\) / \\(9.36\\%\\) declination comparing to source only results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Accordingly, previous TTA methods are hard to apply to domain adaptive 3D detection tasks due to two major reasons. Firstly, from the perspective of updating parameter selection, previous methods often only update batch normalization (BN) layers to alleviate the domain shifts in the statistical information of network. However, this is not well suited for 3D detection task that also needs rich localization information. Besides, the optimization objectives of previous" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reg-TTA3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table_caption", + "bbox": [ + 0.247, + 0.145, + 0.756, + 0.16 + ], + "angle": 0, + "content": "Table 1: Comparison of different settings to alleviate domain discrepancies." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.165, + 0.695, + 0.248 + ], + "angle": 0, + "content": "
SettingSource data Target data Online
Fine-tuningXxt,ytX
Unsupervised domain adaptationxs,s,ysxtX
Source-free domain adaptationXxtX
Test-time adaptationXxt✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.784, + 0.3 + ], + "angle": 0, + "content": "methods are designed for classification tasks, which would make the model optimize for extracting features with more category semantic information rather than regression information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.785, + 0.526 + ], + "angle": 0, + "content": "In order to tackle the challenges mentioned above, in this work, we propose Reg-TTA3D to perform test-time adaptation on 3D object detection tasks. To deal with the problem of updated parameter selection, we conduct extensive experiments and observe that high classification accuracy is already achieved by the source model, and the main factor limiting the performance improvement is the inaccurate regression results. Further, by only adapting regression layers, the model can achieve high performance at a low training cost as shown in Tab. 2 which motivates us to focus on regression adaptation in this work. Therefore, to make better regression, we propose noise-consistency pseudo-label generation (NPG) and confidence-guided regression refinement (CRR). The former module aims to filter boxes with instability under noise interference, i.e., noisy point clouds, while the latter one makes the box regression results align better with the distribution of the target domain box. Finally, we introduce a class-balanced EMA updating strategy (CBU) that can effectively update the regression layers, to improve the performance on multi-class simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.527, + 0.588, + 0.541 + ], + "angle": 0, + "content": "Our contribution can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.542, + 0.787, + 0.587 + ], + "angle": 0, + "content": "1. We present a new task, namely test-time domain adaptive 3D object detection, and propose Reg-TTA3D to solve this task from the novel view of focusing on adapting only regression parameters." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.588, + 0.787, + 0.648 + ], + "angle": 0, + "content": "2. We propose a noise-consistency pseudo-label generation process, a confidence-guided regression refinement loss, and a class-balanced EMA updating strategy, to obtain better detection results and boost the model's performance in an efficient and effective way." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.648, + 0.787, + 0.709 + ], + "angle": 0, + "content": "3. Experiments show that Reg-TTA3D reaches a comparable or even better performance by updating less than \\(0.1\\%\\) parameters within less than \\(1\\%\\) time compared to UDA works, and significantly outperforms previous TTA methods on multiple cross-domain scenarios." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.542, + 0.787, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.729, + 0.398, + 0.745 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.553, + 0.773 + ], + "angle": 0, + "content": "2.1 LiDAR-based 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "With the development of autonomous driving, LiDAR-based 3D object detection has attracted increasing attention. Current prevailing LiDAR-based 3D detectors can be roughly divided into point-based, voxel-based, and point-voxel-based methods according to the point cloud processing procedure. Point-based" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Yuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "methods [20, 35, 43] first sample a subset of point clouds, then extract features and generate proposals from sampled points. Among them, PointRCNN [20] uses a two-stage framework that is composed of a bottom-up proposal generation process and a proposal refinement process. In contrast, voxel-based methods [7, 12, 32, 36, 46] first divide disordered points into regular grids which are then fed into the convolutional network. SECOND [32] is a pioneer in using the 3D sparse backbone to extract features. CenterPoint [36] introduces a center-based detector and achieves promising performance. Point-voxel-based [18, 19] methods integrate the features extracted from points and voxels to obtain better feature representations. However, existing 3D detectors fail to deal with domain gaps which is essential in practical usage." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.341, + 0.441, + 0.357 + ], + "angle": 0, + "content": "2.2 Test-time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Test-time adaptation [1, 4, 10, 22, 24, 27, 28] aims to rapidly adapt the source model to target domain without accessing source domain data during test time, which is more practical in the real world since it can handle dynamic domain shifts. TENT [27] explores test-time adaptation for the first time and updates the batch normalization layers using designed entropy minimization loss. SHOT [14] deals with such a challenging task by maximizing the mutual information. More recently, MATE [16] extends test-time adaptation to 3D point clouds and finds that the model can be adapted effectively by masking and reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Although TTA has been extensively explored in 2D scenarios, it is still under-explored in 3D tasks, especially in large-scale point cloud scenarios (e.g., autonomous driving scenarios). Besides, most previous TTA works mainly focus on classification tasks, such as classification [1, 16, 27] and segmentation [22], and they fail to tackle detection tasks since detection considers classification and box regression at the same time. In contrast, Reg-TTA3D considers TTA for 3D object detection tasks from the perspective of regression, which is more suitable for 3D detection tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.633, + 0.659 + ], + "angle": 0, + "content": "2.3 Domain Adaptation for 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "3D detectors often suffer large performance drops when training and testing under different domains, which makes them unstable when the surrounding environment dynamically changes. Recently, researchers have tried to tackle such a problem using domain adaptation techniques [6,17,29,33,34,38]. SN [29] proposes a statistical method to align the size of bounding boxes between source and target domains. ST3D [33] and \\(\\mathrm{ST3D + + }\\) [34] design a self-training pipeline and mitigate domain gaps using a memory bank and curriculum data augmentation. To deal with the class imbalance issue, ReDB [6] proposes a pseudo-label generation pipeline to obtain reliable, diverse, and class-balanced pseudo-labels. However, the existing works cannot deal with the emergency changes in the environment due to the time-consuming and large amount of calculation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reg-TTA3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.312, + 0.145, + 0.689, + 0.16 + ], + "angle": 0, + "content": "Table 2: Performance of updating different parameters." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.172, + 0.709, + 0.26 + ], + "angle": 0, + "content": "
Tuning layersWaymo → KITTInuScenes → KITTIParameters
Batch normalization52.84 / 31.7731.86 / 18.259.08k
Classification layers52.82 / 36.3731.51 / 19.761.17k
Regression layers59.38 / 45.5433.63 / 21.802.73k
Detection head59.62 / 46.0832.51 / 20.394.68k
All parameters60.78 / 46.7133.23 / 19.6012.12M
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.267, + 0.331, + 0.284 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.375 + ], + "angle": 0, + "content": "The overall framework of Reg-TTA3D is shown in Fig. 2. To better illustrate our proposed method, we start by introducing the problem formulation and our baseline in Sec. 3.1. Then, we give an analysis of the updated parameters selection in Sec. 3.2. Finally, we detail each module of Reg-TTA3D and give the overall objective and adaptation strategy in Sec. 3.3 and Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.394, + 0.362, + 0.41 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.524 + ], + "angle": 0, + "content": "Problem Formulation. Following typical test-time adaptation setting [22, 27], given the model \\( M_S \\) pre-trained on the source domain \\( D_S = \\{(x_i^S, y_i^S)\\}_{i=1}^{n_S} \\) and an unlabeled target set \\( D_T = \\{x_j^T\\}_{j=1}^{n_T} \\), where the source and target domain follow different distributions, our goal is to adapt the source model to the target domain without accessing source domain data in an online manner. During adaptation, in each timestep, a batch of unlabeled data \\( [x_1^t, x_2^t, \\dots, x_N^t] \\) is sampled from \\( D_T \\) to update the model, where \\( N \\) is batch size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Baseline Introduction. Following [6, 34], we use SECOND-IoU [32] as our baseline detector which is composed of a feature extractor (i.e., 3D and 2D backbone) to obtain bird's eye view (BEV) features, a region proposal network (RPN) to generate proposals and a detection head to provide the classification and regression results. The detector is trained using the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.612, + 0.785, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e t} = \\mathcal {L} _ {r p n} + \\mathcal {L} _ {r c n n}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.78, + 0.646 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{rpn}\\) represents the loss of RPN and \\(\\mathcal{L}_{rcnn}\\) denotes box refinement loss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.514, + 0.681 + ], + "angle": 0, + "content": "3.2 Updated Parameters Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.764 + ], + "angle": 0, + "content": "The aim of Reg-TTA3D is to transfer the source model to the target domain rapidly so that the model can tackle the continuous changes in the surrounding environment. To achieve this goal, the amount of updated parameters we select needs to be relatively small, and the accuracy needs to be largely improved by only updating selected parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Previous TTA works [22, 27] mainly focus on adapting batch normalization (BN) layers to mitigate the domain discrepancies, since BN layers contain the statistics of domains. Such an adaptation strategy is effective for domain adaptive classification tasks, as more accurate semantic representations can be obtained, which is especially important for classification. However, only adapting" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Yuan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.437 + ], + "angle": 0, + "content": "Fig. 2: The overall framework of the proposed Reg-TTA3D, which employs SECOND-IoU [32] as the baseline detector and consists of noisy-consistency pseudo-label generation process (NPG), confidence-guided regression refinement (CRR) and class-balanced EMA updating strategy. The data are first fed into the teacher model to generate pseudo-labels as the supervision to the student model using NPG. Then, the student model is trained using data from another view and supervised by detection loss and refinement loss obtained by CRR. Finally, the class-balanced EMA updating strategy is used to update the teacher model. Note that the only difference between the teacher model and the student model lies in the regression layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.789, + 0.675 + ], + "angle": 0, + "content": "BN layers cannot rapidly transfer the source detector to the target domain since detection tasks contain both classification and regression. Some works [16, 26] try to solve this problem by updating the entire model but suffer from huge computational costs. To explore an effective and efficient method, by conducting extensive experiments, we find that the model can adapt to the target domain efficiently by only updating regression layers. As shown in Tab. 2, we fine-tune the parameters using \\(1\\%\\) KITTI training data for 1 epoch and observe that only fine-tuning the regression layers can reach relatively high performance and even surpass the performance of fine-tuning the entire model. This is because in the autonomous driving scenario, there exists a large difference between categories such as car and pedestrian, the source model can already achieve high classification accuracy and the main factor for limiting the detection accuracy is the inaccurate regression results. Based on the observation, in this work, we focus on studying how to only adapt the parameters of regression layers and make better regression." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.365, + 0.71 + ], + "angle": 0, + "content": "3.3 Reg-TTA3D" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.809 + ], + "angle": 0, + "content": "As mentioned above, the quality of box regression is essential to TTA for 3D object detection. To make better regression, we propose noise-consistency pseudolabel generation to obtain better regression supervision, and confidence-guided regression refinement to further refine the regression results. Besides, to effectively update the regression layers, we design a class-balanced EMA updating strategy to avoid overfitting the model to a single class." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Noise-consistency Pseudo-label Generation. As mentioned in [34, 38], the pseudo-labels directly generated by the source model contain a lot of noise due" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reg-TTA3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.223 + ], + "angle": 0, + "content": "to domain discrepancies. Previous methods try to alleviate pseudo-label noise by weighting pseudo-labels obtained in different epochs [33, 34], or adding the groundtruth instances from the source domain to target domain frames [6]. However, these methods cannot apply to test-time adaptation due to the inaccessibility of source data and online adaptation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Inspired by test-time augmentation methods which comprehensively consider the model predictions under different data augmentation methods to provide final results. Specifically, the model's prediction of real objects is relatively stable, while being uncertain for the noise regions. To this end, we propose noise-consistency pseudo-label generation to obtain more accurate pseudo-labels, by measuring the stability of model predictions when adding random noise to regions of interest (RoIs). In detail, given proposals generated by the teacher model \\( B^{t} = \\{(c_{x}^{t}, c_{y}^{t}, c_{z}^{t}, l^{t}, w^{t}, h^{t}, \\theta^{t})_{i}\\}_{i=1}^{n} \\), its corresponding features \\( \\{f_{i}^{t}\\}_{i=1}^{n} \\) and confidence scores \\( \\{s_{i}^{t}\\}_{i=1}^{n} \\) where \\( (c_{x}^{t}, c_{y}^{t}, c_{z}^{t}) \\) and \\( (l^{t}, w^{t}, h^{t}) \\) denote the center and the size of the box respectively, \\( \\theta^{t} \\) is the rotation angle and \\( n \\) is the number of boxes. We can obtain boxes with noise \\( B^{t'} = \\{(c_{x}^{t'}, c_{y}^{t'}, c_{z}^{t'}, l^{t'}, w^{t'}, h^{t'}, \\theta^{t'})\\}_{i=1}^{n} \\), corresponding features \\( \\{f_{i}^{t'}\\}_{i=1}^{n} \\) and scores \\( \\{s_{i}^{t'}\\}_{i=1}^{n} \\), by adding random noise to boxes' size and angle. To measure the stability of RoIs, we calculate the inconsistency rate as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.44, + 0.786, + 0.476 + ], + "angle": 0, + "content": "\\[\nR _ {i n c o n s i s t e n c y, i} = 1 - \\frac {\\left| s _ {i} ^ {t} - s _ {i} ^ {t ^ {\\prime}} \\right|}{\\operatorname {I o U} \\left(b _ {i} ^ {t} , b _ {i} ^ {t ^ {\\prime}}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.478, + 0.788, + 0.569 + ], + "angle": 0, + "content": "where \\(\\mathrm{IoU}(b_i^t, b_i^{t'})\\) denotes the intersection of union (IoU) between \\(b_i^t \\in B^t\\) and \\(b_i^{t'} \\in B^{t'}\\). A high inconsistency rate represents that the model prediction is unstable to the current object which is more likely to be a noise region. Therefore, we filter the boxes with high inconsistency rate by a threshold \\(\\tau\\) to get boxes and scores \\(\\hat{B}, \\{\\hat{s}_i\\}_{i=1}^n\\) and \\(\\hat{B}', \\{\\hat{s}_i'\\}_{i=1}^n\\) with stability under noise interference. Then, we ensemble \\(\\hat{B}\\) and \\(\\hat{B}'\\) by confidence scores as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.575, + 0.786, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\tilde {B} = \\left\\{\\tilde {b _ {i}} \\right\\}, \\quad \\tilde {b _ {i}} = \\left\\{ \\begin{array}{l l} \\hat {b} _ {i}, & \\hat {s} _ {i} > \\hat {s} _ {i} ^ {\\prime}, \\hat {b} _ {i} \\in \\hat {B} \\\\ \\hat {b} _ {i} ^ {\\prime}, & \\hat {s} _ {i} < \\hat {s} _ {i} ^ {\\prime}, \\hat {b} _ {i} ^ {\\prime} \\in \\hat {B} ^ {\\prime} \\end{array} , \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "After obtaining \\(\\tilde{B}\\), the boxes are filtered by non-maximum suppression (NMS) with a relatively high confidence threshold, to remove redundant boxes and boxes with low confidence. Finally, the pseudo-labels \\(B^{npg}\\) and \\(\\{l_i^{npg}\\}_{i=1}^n\\) with high stability and confidence can be acquired where \\(\\{l_i^{npg}\\}_{i=1}^n\\) is classification results. Confidence-guided Regression Refinement. Although lots of boxes containing noisy point cloud can be removed by measuring the stability under noise injection, there still exist the problems of inaccurate box regression results. This is because the distributions of bounding box size are inconsistent between source and target domains as mentioned in [29,37,42], and it is difficult to mitigate such domain discrepancy with only NPG. By analyzing the distribution of predicted bounding boxes, we find that boxes with higher confidence scores align better with the distribution of the target domain boxes as shown in Fig. 3, since precise boxes correspond to high-quality RoI features that are less disturbed by background representations. Based on this observation, we design a confidence-guided regression refinement to obtain more precise regression results of box size." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "Yuan et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.734, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.28, + 0.789, + 0.308 + ], + "angle": 0, + "content": "Fig. 3: Distribution of pseudo box size under different thresholds. PS denotes pseudo boxes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.321, + 0.789, + 0.442 + ], + "angle": 0, + "content": "In particular, given predicted boxes generated by the student model \\( B^{s} = \\{(c_{x}^{s}, c_{y}^{s}, c_{z}^{s}, l^{s}, w^{s}, h^{s}, \\theta^{s})_{j}\\}_{j=1}^{k} \\) and scores \\( \\{s_{j}^{s}\\}_{j=1}^{k} \\) of a batch of frames, where \\( k \\) is the total number of predicted boxes. We sort boxes belonging to the same class according to scores in descending order, and sample a proportion of boxes with high scores to calculate the mean value of sample boxes size \\( (\\bar{l}^{s}, \\bar{w}^{s}, \\bar{h}^{s}) \\). Please refer to our supplementary material for the study of the proportion of sampled boxes. Once obtained \\( (\\bar{l}^{s}, \\bar{w}^{s}, \\bar{h}^{s}) \\), we can use it as a supervision signal to make the box regression value closer to the distribution of the target domain as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.455, + 0.613, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e f i n e}} = \\mathbb {E} _ {i = 1} ^ {n} \\Sigma_ {d \\in [ d _ {l}, d _ {w}, d _ {h} ]} d,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "\\[\nw h e r e d = \\left\\{ \\begin{array}{l l} 0, & | \\left(z ^ {s} - \\bar {z} ^ {s}\\right) | < p \\\\ \\left(z ^ {s} - \\bar {z} ^ {s}\\right) ^ {2}, & | \\left(z ^ {s} - \\bar {z} ^ {s}\\right) | \\geq p \\end{array} , \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.522, + 0.789, + 0.568 + ], + "angle": 0, + "content": "where \\(z \\in [l^s, w^s, h^s]\\), \\(\\overline{z} \\in [\\overline{l}^s, \\overline{w}^s, \\overline{h}^s]\\), \\(n\\) is batch size and \\(p\\) is set 0.1 to allow the size of objects to fluctuate within a certain range. The model can predict more accurate box size when using \\(\\mathcal{L}_{\\text{refine}}\\) for optimization." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.568, + 0.789, + 0.795 + ], + "angle": 0, + "content": "Class-balanced EMA Updating. To make the model more stable, we update the teacher regression layers by exponential moving average (EMA). However, the model tends to overfit to the category with a large number of instances (i.e., car). As a result, the detection performance on categories with few instances will decrease. To make the model better balance the performance improvement for each class, we propose a class-balanced EMA updating strategy to make the model update more when the class is balanced, while updating less when a single category dominates. Specifically, given the pseudo-labels obtained from NPG \\( B^{npg} = \\{b_i^{npg} \\in \\mathbb{R}^7\\} \\) and its corresponding class labels \\( \\{c_i^{npg}\\} \\). We calculate the proportion \\( \\{p_c\\}_{c=1}^C \\) of each category in pseudo-labels according to \\( \\{c_i^{npg}\\} \\), where \\( C \\) is the number of total categories and \\( \\Sigma_{c=1}^C p_c = 1 \\). Then the variance of \\( \\{p_c\\}_{c=1}^C \\) is calculated and normalized to [0.99, 0.999] to obtain the momentum \\( \\alpha \\). A high value of \\( \\alpha \\) means that a severe class-imbalance issue exists in the current batch and the model should update less. Finally the parameter of the teacher regression layer \\( \\omega_{reg}^t \\) can be updated by the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.807, + 0.785, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\omega_ {r e g} ^ {t} \\leftarrow \\alpha \\omega_ {r e g} ^ {t} + (1 - \\alpha) \\omega_ {r e g} ^ {s}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.829, + 0.627, + 0.845 + ], + "angle": 0, + "content": "where \\(\\omega_{reg}^{s}\\) is the parameters of student regression layers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reg-TTA3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.147, + 0.61, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Single iteration of Reg-TTA3D strategy" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.164, + 0.62, + 0.19 + ], + "angle": 0, + "content": "Input: A batch of frames \\(\\{x_{i}\\}_{i = 1}^{n}\\) and \\(\\{\\dot{x}_i\\}_{i = 1}^n\\) , where \\(\\dot{x}_i\\) is obtained by performing data augmentation on \\(x_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.191, + 0.581, + 0.206 + ], + "angle": 0, + "content": "Output: Predictions of current batch \\( B^{t} \\) and \\( \\{l_i^t\\}_{i = 1}^n \\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.207, + 0.786, + 0.235 + ], + "angle": 0, + "content": "1: Get the predictions \\( B^{t} = \\{b_{i}^{t}\\in \\mathbb{R}^{7}\\}_{i = 1}^{n} \\) and \\( \\{l_i^t\\}_{i = 1}^n \\) of \\( \\{x_{i}\\}_{i = 1}^{n} \\) using teacher model, \\( B^{t} \\) and \\( \\{l_i^t\\}_{i = 1}^n \\) are the detection results of current batch." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.235, + 0.786, + 0.262 + ], + "angle": 0, + "content": "2: Use NPG to obtain boxes \\( B^{npg} \\) and \\( \\{l_i^{npg}\\}_{i=1}^n \\) with stability under noise interference as pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.263, + 0.786, + 0.292 + ], + "angle": 0, + "content": "3: Get the predictions of student model \\( B^{s} \\) and \\( \\{s_{i}^{s}\\}_{i = 1}^{n} \\) of \\( \\{\\dot{x}_i\\}_{i = 1}^n \\), supervise the student model by \\( B^{npg} \\) and \\( \\{l_i^{npg}\\}_{i = 1}^n \\) to get \\( \\mathcal{L}_{det} \\)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.292, + 0.442, + 0.305 + ], + "angle": 0, + "content": "4: Calculate \\(\\mathcal{L}_{\\text{refine}}\\) using CRR." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.305, + 0.526, + 0.319 + ], + "angle": 0, + "content": "5: Optimize the student model using Eq. (6)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.319, + 0.58, + 0.333 + ], + "angle": 0, + "content": "6: Update the teacher model using CBU and Eq. (5)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.333, + 0.545, + 0.347 + ], + "angle": 0, + "content": "7: Return: Current predictions \\( B^{t} \\) and \\( \\{l_i^t\\}_{i = 1}^n \\)." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.207, + 0.786, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.701, + 0.396 + ], + "angle": 0, + "content": "3.4 Overall Objective and Test-time Adaptation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.786, + 0.435 + ], + "angle": 0, + "content": "Overall Objective. The overall objective to optimize the student regression layers can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.448, + 0.786, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {d e t}} + \\mathcal {L} _ {\\text {r e f i n e}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.631, + 0.489 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{det}\\) and \\(\\mathcal{L}_{refine}\\) are defined in Eq. (1) and Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.519 + ], + "angle": 0, + "content": "Test-time Adaptation Strategy. The test-time adaptation strategy of Reg-TTA3D can be summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.376, + 0.56 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.428, + 0.59 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.795, + 0.795 + ], + "angle": 0, + "content": "Datasets and Cross-domain Settings. We conduct experiments on four commonly-used datasets: Waymo [23], KITTI [9], nuScenes [2] and nuScenes-C [11]. To verify the effectiveness of Reg-TTA3D, we consider three cross-domain settings including cross-location (i.e., Waymo \\(\\rightarrow\\) KITTI), cross-beam (i.e., Waymo \\(\\rightarrow\\) nuScenes, nuScenes \\(\\rightarrow\\) KITTI) and cross-weather (i.e., nuScenes \\(\\rightarrow\\) nuScenes-C). Note that in the cross-weather setting, we use the medium-difficulty fog and snow scenarios in the nuScenes-C dataset. Following ReDB [6], in the first two cross-domain settings, we use the KITTI evaluation metric to evaluate for three common classes in autonomous driving scenarios (i.e., car, pedestrian, and cyclist), we report the average precision in both 3D (i.e., \\(\\mathrm{AP}_{\\mathrm{3D}}\\)) and BEV (i.e., \\(\\mathrm{AP}_{\\mathrm{BEV}}\\)) over 40 recall positions, with IoU threshold 0.7 for car and 0.5 for pedestrian and cyclist. In the cross-weather setting, we use the official nuScenes evaluation metric and report mean average precision (mAP)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Implementation Details. Following [6], we evaluate the proposed Reg-TTA3D on SECOND-IoU [32], and the source model is trained for all categories using random object scaling (ROS) augmentation method. We set hyperparameter \\(\\tau\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Yuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.208 + ], + "angle": 0, + "content": "to 1.5 and \\( p \\) to 0.1 for all cross-domain settings and use Adam optimizer with learning rate of \\( 1 \\times 10^{-3} \\). We utilize commonly-used data augmentation methods (e.g., random world flip, random world rotation, random world scaling) to get \\( \\{\\dot{x}_i\\}_{i=1}^n \\). Our code is built on 3DTrans [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.374 + ], + "angle": 0, + "content": "Comparison Baselines. To verify the effectiveness of Reg-TTA3D, we mainly compare our methods to unsupervised domain adaptive 3D object detection and previous test-time adaptation methods that we apply to 3D object detection. UDA methods: (1) ST3D [33] is a self-training method that continuously updates the pseudo-labels, (2) \\(\\mathrm{ST3D + + }\\) [34] provides more insight analysis and extends ST3D to get better results, (3) ReDB [6] is a pioneer to perform domain adaptive 3D detection on multiple categories. TTA methods: (1) TENT [27] proposes an entropy minimization method which focuses on adapting BN layers, (2) SHOT [14] optimizes feature extraction module using both information maximization and pseudo-labeling. (3) MATE [16] uses masked autoencoder to apply test-time adaptation to 3D classification." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.394, + 0.373, + 0.408 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.789, + 0.629 + ], + "angle": 0, + "content": "Results on cross-location and cross-beam scenarios. We first evaluate Reg-TTA3D on cross-location and cross-beam scenarios and compare the results with previous UDA methods. As shown in Tab. 3, Reg-TTA3D can greatly improve the detection accuracy on Waymo \\(\\rightarrow\\) KITTI and nuScenes \\(\\rightarrow\\) KITTI settings, largely narrow the performance gap between source only and oracle. For example, in the Waymo \\(\\rightarrow\\) KITTI setting, our Reg-TTA3D can achieve \\(61.78\\%\\) / \\(49.96\\%\\) in \\(\\mathrm{AP}_{\\mathrm{BEV}} / \\mathrm{AP}_{\\mathrm{3D}}\\), even surpass UDA methods such as ST3D [33] and ST3D++ [34] and it is comparable to the state-of-the-art UDA method (i.e., \\(61.14\\%\\) / \\(50.10\\%\\) in \\(\\mathrm{AP}_{\\mathrm{BEV}} / \\mathrm{AP}_{\\mathrm{3D}}\\) reported in ReDB [6]). In nuScenes \\(\\rightarrow\\) KITTI and Waymo \\(\\rightarrow\\) nuScenes cross-domain scenarios, the performance of Reg-TTA3D is also better than some UDA methods (e.g., ST3D [33]) which needs source domain data and training for multiple epochs. In contrast, Reg-TTA3D only needs target data and one epoch of adaptation which greatly reduces the time consumption." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Further, we compare Reg-TTA3D with the previous TTA methods in Tab. 3. We find that previous TTA methods suffer performance drops when applied to 3D detection tasks (e.g., TENT [27] reduce the detection accuracy to \\(31.34\\%\\) / \\(19.47\\%\\) in the Waymo \\(\\rightarrow\\) KITTI setting). This is mainly because the detection task is composed of classification and regression and existing methods only consider the classification tasks. As a result, the model will concentrate on extracting effective representations for the classification tasks and ignore the localization information. However, our Reg-TTA3D fully considers the challenges in detection tasks and can improve the performance on multiple categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Results on the cross-weather scenario. To more comprehensively evaluate the effectiveness of our proposed method, we conduct experiments in a more realistic scenario (i.e., cross-weather). Since nuScenes-C dataset only has the validation set, we only compare Reg-TTA3D with existing TTA methods. It can be seen in Tab. 4 that Reg-TTA3D can improve the performance on both snow" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reg-TTA3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.219 + ], + "angle": 0, + "content": "Table 3: Results on cross-location and cross-beam scenarios. Following [6, 33], we use SECOND-IoU [32] as out baseline detector. Source Only represents that the pretrained detector is directly evaluated on the target domain, and Oracle denotes the detection results obtained using the fully annotated target training data. Red: Best UDA method, Blue: Best TTA method." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.23, + 0.784, + 0.609 + ], + "angle": 0, + "content": "
TASKMETHODSETTINGCARPEDESTRIANCYCLISTMEAN AP
Waymo → KITTISOURCE ONLY-51.48 / 19.7840.80 / 31.2647.63 / 35.4546.64 / 28.83
SN76.61 / 54.1452.48 / 48.2034.56 / 32.7454.55 / 45.03
ST3DUDA77.62 / 49.2444.45 / 42.0447.74 / 42.9556.60 / 44.70
ST3D++77.68 / 50.0349.09 / 46.1951.50 / 47.7059.42 / 47.97
ReDB80.37 / 54.1251.01 / 48.2052.05 / 47.9761.14 / 50.10
TENTTTA71.98 / 39.014.54 / 3.9517.51 / 15.4531.34 / 19.47
SHOT79.52 / 41.3133.52 / 29.0240.79 / 35.8351.27 / 35.38
MATE75.17 / 30.6439.94 / 36.9243.72 / 40.2252.94 / 35.92
REG-TTA3D81.60 / 56.0350.89 / 45.4352.85 / 48.4361.78 / 49.96
ORACLE-83.29 / 73.4546.64 / 41.3362.92 / 60.3264.28 / 58.37
Waymo → nuScenesSOURCE ONLY-30.64 / 17.181.81 / 0.580.97 / 0.8811.14 / 6.22
SN29.56 / 17.781.33 / 1.072.92 / 2.6111.27 / 7.15
ST3DUDA28.42 / 17.831.64 / 1.394.01 / 3.5411.36 / 7.59
ST3D++28.87 / 19.151.82 / 1.584.09 / 3.7411.60 / 8.16
ReDB30.12 / 18.562.47 / 2.146.56 / 5.1913.05 / 8.63
TENTTTA16.40 / 3.430.82 / 0.250.28 / 0.215.83 / 1.29
SHOT12.26 / 1.421.56 / 1.360.22 / 0.204.68 / 0.99
MATE8.64 / 3.991.72 / 0.970.24 / 0.083.53 / 1.68
REG-TTA3D30.54 / 18.152.04 / 1.755.01 / 4.7012.53 / 8.20
ORACLE-51.88 / 34.8725.24 / 18.9215.06 / 11.7330.73 / 21.84
nuScenes → KITTISOURCE ONLY-39.15 / 7.6521.54 / 16.876.31 / 2.4422.33 / 8.98
SN56.08 / 28.6723.05 / 16.845.11 / 2.3228.08 / 15.94
ST3DUDA71.50 / 48.0922.64 / 17.617.86 / 5.2034.00 / 23.64
ST3D++69.90 / 44.6224.11 / 18.2010.14 / 6.3933.75 / 23.07
ReDB74.23 / 51.3125.95 / 18.3813.82 / 8.6438.00 / 26.11
TENTTTA43.35 / 19.839.40 / 5.772.71 / 2.7018.49 / 9.43
SHOT60.24 / 23.0117.51 / 9.442.47 / 1.6626.74 / 11.37
MATE61.89 / 27.1315.98 / 13.112.94 / 2.4526.93 / 14.23
REG-TTA3D68.73 / 44.5622.78 / 17.7110.21 / 7.1233.90 / 23.13
ORACLE-83.29 / 73.4546.64 / 41.3362.92 / 60.3264.28 / 58.37
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.669 + ], + "angle": 0, + "content": "and fog scenarios (e.g., \\(45.68\\% \\rightarrow 47.09\\%\\) mAP on the Normal \\(\\rightarrow\\) Snow setting). The results verify that our proposed method can handle dynamic environment changes in road scenes and can be applied in more practical scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.4, + 0.708 + ], + "angle": 0, + "content": "4.3 Insight Analyses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Sensitivity to Detector Architecture. Following [6], to verify the effectiveness of Reg-TTA3D on different types of detectors, we further conduct experiments on nuScenes \\(\\rightarrow\\) KITTI cross-domain scenario using PointRCNN [20] as the baseline detector. As show in Tab. 5, our Reg-TTA3D outperforms all TTA methods and some UDA methods. Note that compared to the source-free domain adaptation method (i.e., SF-UDA\\(^{3D}\\) [17]) which is inaccessible to the source data and need to adapt for multiple epochs, our Reg-TTA3D adapted on multi-class can achieve a better result on the car category (e.g., \\(50.87\\%\\) compared to \\(49.80\\%\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Yuan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Results on cross-weather scenario (i.e., nuScenes Normal \\(\\rightarrow\\) Fog and Normal \\(\\rightarrow\\) Snow). Fog and snow data are from nuScenes-C [11] dataset. We use CenterPoint [36] as our baseline detector since it is the most widely used detector on nuScenes." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.193, + 0.782, + 0.326 + ], + "angle": 0, + "content": "
METHODTaskmAPCARTRUCKCV.BUSTRAILERBARRIERMOTOR.BICYCLEPED.TC.
SOURCE ONLY48.4079.1252.588.1067.5331.0064.407.3026.6581.2865.97
TENT41.8373.6215.237.5858.0421.8560.451.5623.8477.5952.24
SHOTFog42.3172.6340.707.5958.1822.2461.710.8526.4577.8454.87
MATE23.9764.6124.183.3427.2614.1120.450.007.6749.6728.33
REG-TTA3D49.6780.3054.4310.8668.6534.5363.527.7927.0782.2067.28
SOURCE ONLY45.8680.7448.774.6864.1528.8462.629.0224.7278.3756.62
TENTSnow40.5175.0939.376.5157.5719.9659.821.8923.3376.7844.73
SHOT42.1574.6639.747.0158.1221.4662.941.1927.3078.3950.66
MATE23.4363.9822.573.2025.3013.0622.790.007.640.5024.81
REG-TTA3D47.0982.0149.717.1968.1630.5462.966.3326.2980.3757.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.334, + 0.787, + 0.38 + ], + "angle": 0, + "content": "Table 5: Performance comparisons (AP3D) with PointRCNN as baseline detector on nuScenes \\(\\rightarrow\\) KITTI task. \\(\\ddagger\\) indicates the results reported in the original paper. Red: Best UDA method, Blue: Best TTA method." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.389, + 0.782, + 0.535 + ], + "angle": 0, + "content": "
MethodSettingCARPEDESTRIANCYCLISTAVERAGE
EASYMOD.HARDEASYMOD.HARDEASYMOD.HARDEASYMOD.HARD
SOURCE ONLY-42.7732.1128.7543.2637.2933.163.284.094.1829.7724.6022.03
SN66.5650.3245.9242.9637.1532.459.077.577.4239.5331.6828.60
ST3DUDA48.8541.9038.9245.6738.7133.0926.5019.3518.3840.3433.3230.13
ST3D++60.4549.3645.8850.7742.4336.6427.2017.9416.5246.1436.5833.01
REDB71.4557.953.9152.3244.3337.9545.1332.9331.0556.3045.0540.97
SF-UDA3D‡SFDA68.8049.8045.00---------
TENTTTA42.5532.4127.6130.7126.0923.784.243.663.3225.8320.7018.23
SHOT30.0923.0920.4119.2416.2914.312.431.701.5117.2513.6912.07
REG-TTA3D65.2050.8746.7245.8042.5534.9040.2527.8725.2950.4140.4335.63
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.596 + ], + "angle": 0, + "content": "reported by SF-UDA\\(^{3D}\\) in moderate level) in one epoch of adaptation even if SF-UDA\\(^{3D}\\) is trained on the unfair single class. The results further show that our proposed methods can apply to different detector architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.793 + ], + "angle": 0, + "content": "Ablation Studies. To further verify the effectiveness of Reg-TTA3D, we conduct extensive experiments to investigate the impact of each component we proposed including NPG, CRR and CBU. As shown in Tab. 6, the improvement (i.e., \\(0.92\\% \\mathrm{AP}_{3\\mathrm{D}}\\)) of using naive pseudo-labels is relatively small compared to the performance of source only since pseudo-labels contain lots of noise samples and the regression results are not accurate due to domain discrepancies. Reg-TTA can effectively handle such a problem by proposing NPG and CRR modules, with \\(2.90\\%\\) and \\(3.60\\%\\) improvement respectively compared to source only. Besides, when combining these two modules, the performance can be further improved because NPG focuses on filter noise samples and CRR concentrates on making more accurate regression results. Further, by combining with CBU which makes the model update in a class-balance manner, the performance can be further boosted and achieve \\(49.96\\% \\mathrm{AP}_{3\\mathrm{D}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Sensitivity to Hyperparameters. We ablate the impact of hyperparameters \\(\\tau\\) (i.e., the threshold of inconsistency rate) and \\(p\\) in this subsection. As shown in Tab. 7, we first set \\(\\tau\\) to 1.0, 1.5, 2.0 respectively and observe that the fluctuation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Reg-TTA3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.147, + 0.574, + 0.244 + ], + "angle": 0, + "content": "Table 6: Ablation studies on each component. PS, NPG, CRR and CBU represent using naive pseudo-labels, noise-consistency pseudo-label generation, confidence-guided regression refinement and class-balanced EMA updating, respectively. The experiments are conducted on Waymo \\(\\rightarrow\\) KITTI scenario using SECOND-IoU as the baseline detector." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.25, + 0.54, + 0.389 + ], + "angle": 0, + "content": "
PS NPGCRRCBUmean APBEVmean AP3D
---57.8942.01
✓--58.7342.93
✓✓-58.8944.91
✓-✓59.8445.61
✓--59.7344.05
✓✓✓60.5747.97
✓✓-60.1246.62
✓-✓60.5948.39
✓✓✓61.7849.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.582, + 0.145, + 0.788, + 0.255 + ], + "angle": 0, + "content": "Table 7: Ablation studies of hyperparameters \\(\\tau\\) and \\(p\\). The experiments are conducted on Waymo \\(\\rightarrow\\) KITTI scenario using SECOND-IoU. Note that when ablate on \\(\\tau\\), we set \\(p = 0.1\\) and when ablate on \\(p\\), we set \\(\\tau = 0.15\\)." + }, + { + "type": "table", + "bbox": [ + 0.618, + 0.26, + 0.753, + 0.389 + ], + "angle": 0, + "content": "
τAPBEV / AP3D
1.060.92 / 49.12
1.561.78 / 49.96
2.060.44 / 48.88
pAPBEV / AP3D
0.0560.29 / 48.26
0.1061.78 / 49.96
0.1561.56 / 49.63
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.407, + 0.497, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.536, + 0.498, + 0.565 + ], + "angle": 0, + "content": "Fig. 4: The relationship between detection accuracy and the iteration number." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.424, + 0.782, + 0.478 + ], + "angle": 0, + "content": "Table 8: Ablation studies on costs. The experiments are conducted on \\( \\mathrm{W} \\rightarrow \\mathrm{K} \\) scenario using SECOND-IoU. We use Tesla A100 to conduct the experiments." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.48, + 0.779, + 0.563 + ], + "angle": 0, + "content": "
MethodTime Cost N_GPUsParam.Results
ST3D4h412.12M 56.60 / 44.70
ST3D++3.5h412.12M 59.42 / 47.97
ReDB3.2h412.12M 61.14 / 50.10
TENT4min19.08K 31.34 / 19.47
MATE6min10.71M 52.94 / 35.92
Reg-TTA3D4.5min12.73K 61.78 / 49.96
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.787, + 0.641 + ], + "angle": 0, + "content": "of the performance is \\( 1.34\\% / 1.08\\% \\). Then, by setting \\( p \\) to different values (i.e., 0.05, 0.1, 0.15), we find that there is only little fluctuation and the best performance can be achieved when set to 0.1. The results further show that our method is robust to these hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Analysis on Pseudo-labeling Accuracy. In this section, we measure the accuracy of each iteration during adaptation. Fig. 4 reflects the detection accuracy on Waymo \\(\\rightarrow\\) KITTI as iteration increases. It can be seen that the accuracy continuously improves since the model is gradually adapted to the target domain which also verifies the stability of Reg-TTA3D during adaptation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Analysis on Costs. Further, we report the time cost and total updated parameters in Tab. 8. Notably, Reg-TTA3D achieves comparable results to UDA methods adapting only less than \\(0.1\\%\\) of parameters within less than \\(1\\%\\) of time. When compared to TTA methods, our proposed method can achieve much higher performance with less updated parameters and comparable time cost which further verifies that our Reg-TTA3D is an effective and efficient method." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Ablation Studies of the momentum. We ablate the momentum in EMA updating and compare it to our class-balance EMA updating strategy. Specifi-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Yuan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.481, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.143, + 0.766, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.299, + 0.782, + 0.328 + ], + "angle": 0, + "content": "Fig. 5: The impact of momentum and com- Fig.6: ablation studies on the proportions with CBU. tion of selected samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.785, + 0.403 + ], + "angle": 0, + "content": "cally, we conduct the experiment for momentum between 0.99 and 0.999 at 0.001 intervals. As shown in Fig. 5, the result demonstrates that CBU can better update the regression layer due to it can balance update the model and improve the performance on multiple classes at the same time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.525 + ], + "angle": 0, + "content": "Ablation Studies on the proportion of Selected samples in CRR. In our manuscript, we use \\(20\\%\\) boxes with the highest scores to supervise box sizes with relatively low scores. Here, we analyze the impact on the proportion of selected samples. As shown in Fig. 6, when the proportion is relatively high, the average box size will be distributed far from the target domain, and when set to relatively small values, the small number of samples results in the averages being affected by individual instances. Based on the experiments and analysis, we set the proportion to \\(20\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.359, + 0.57 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.713 + ], + "angle": 0, + "content": "In this work, for the first time, we explore test-time adaptation in 3D object detection tasks and propose Reg-TTA3D, which handles rapid changes in the surrounding environment from the perspective of updating parameters selection and adaptation algorithm. By adapting using the proposed noise-consistency pseudo-label generation process, confidence-guided regression refinement loss and class-balanced EMA updating strategy and only updating regression layers, Reg-TTA3D can achieve comparable performance to some UDA methods at a low cost." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.742, + 0.356, + 0.758 + ], + "angle": 0, + "content": "6 Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Although the proposed Reg-TTA3D can handle dynamic changes in the surrounding environments such as continuously changing weather. It cannot deal with emergencies such as some corner cases which is also important in real-world autonomous driving." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reg-TTA3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.395, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.81, + 0.302 + ], + "angle": 0, + "content": "This work is supported by National Key Research and Development Program of China (No. 2022ZD0160101), National Natural Science Foundation of China (No. 62071127, and 62101137), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No.2021SHZDZX0103). This work is also supported by the National Key R&D Program of China (Grant No. 2022ZD0160104), and Shanghai Rising Star Program (Grant No. 23QD1401000). The computations in this research were performed using the CFFF platform of Fudan University." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.328, + 0.323, + 0.344 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.787, + 0.405 + ], + "angle": 0, + "content": "1. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8344-8353 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.46 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.787, + 0.503 + ], + "angle": 0, + "content": "3. Cai, Q., Pan, Y., Ngo, C.W., Tian, X., Duan, L., Yao, T.: Exploring object relation in mean teacher for cross-domain detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11457-11466 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.787, + 0.544 + ], + "angle": 0, + "content": "4. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 295-305 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.787, + 0.587 + ], + "angle": 0, + "content": "5. Chen, Y., Li, W., Sakaridis, C., Dai, D., Van Gool, L.: Domain adaptive faster r-cnn for object detection in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3339-3348 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.787, + 0.629 + ], + "angle": 0, + "content": "6. Chen, Z., Luo, Y., Huang, Z., Wang, Z., Baktashmotlagh, M.: Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudolabeling. arXiv preprint arXiv:2307.07944 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.787, + 0.672 + ], + "angle": 0, + "content": "7. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1201-1209 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.701 + ], + "angle": 0, + "content": "8. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: International conference on machine learning. pp. 1180-1189. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.787, + 0.742 + ], + "angle": 0, + "content": "9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.787, + 0.784 + ], + "angle": 0, + "content": "10. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems 35, 6204-6218 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "11. Kong, L., Liu, Y., Li, X., Chen, R., Zhang, W., Ren, J., Pan, L., Chen, K., Liu, Z.: Robo3d: Towards robust and reliable 3d perception against corruptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 19994-20006 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.362, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Yuan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "12. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "13. Liang, J., Gong, K., Li, S., Liu, C.H., Li, H., Liu, D., Wang, G., et al.: Pareto domain adaptation. Advances in Neural Information Processing Systems 34, 12917-12929 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "14. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.327 + ], + "angle": 0, + "content": "15. Luo, Z., Cai, Z., Zhou, C., Zhang, G., Zhao, H., Yi, S., Lu, S., Li, H., Zhang, S., Liu, Z.: Unsupervised domain adaptive 3d detection with multi-level consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8866-8875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.383 + ], + "angle": 0, + "content": "16. Mirza, M.J., Shin, I., Lin, W., Schriebl, A., Sun, K., Choe, J., Kozinski, M., Possegger, H., Kweon, I.S., Yoon, K.J., et al.: Mate: Masked autoencoders are online 3d test-time learners. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16709-16718 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "17. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771-780. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.466 + ], + "angle": 0, + "content": "18. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529–10538 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "19. Shi, S., Jiang, L., Deng, J., Wang, Z., Guo, C., Shi, J., Wang, X., Li, H.: Pvrccnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. International Journal of Computer Vision 131(2), 531-551 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "20. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "21. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence 43(8), 2647-2664 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.66 + ], + "angle": 0, + "content": "22. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.715 + ], + "angle": 0, + "content": "23. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "24. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "25. Team, D.D.: 3dtrans: An open-source codebase for exploring transferable autonomous driving perception task. https://github.com/PJLab-ADG/3DTrans (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "26. VS, V., Oza, P., Patel, V.M.: Towards online domain adaptive object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 478-488 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Reg-TTA3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "27. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "28. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7201-7211 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.274 + ], + "angle": 0, + "content": "29. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in germany, test in the usa: Making 3d object detectors generalize. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11713-11723 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "30. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. arXiv preprint arXiv:2203.14956 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.374 + ], + "angle": 0, + "content": "31. Xu, Q., Zhou, Y., Wang, W., Qi, C.R., Anguelov, D.: Spg: Unsupervised domain adaptation for 3d object detection via semantic point generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15446-15456 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.402 + ], + "angle": 0, + "content": "32. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.445 + ], + "angle": 0, + "content": "33. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10368-10378 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "34. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d++: Denoised self-training for unsupervised domain adaptation on 3d object detection. IEEE transactions on pattern analysis and machine intelligence 45(5), 6354-6371 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "35. Yang, Z., Sun, Y., Liu, S., Shen, X., Jia, J.: Std: Sparse-to-dense 3d object detector for point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1951-1960 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.572 + ], + "angle": 0, + "content": "36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.614 + ], + "angle": 0, + "content": "37. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Ad-pt: Autonomous driving pre-training with large-scale point cloud dataset. arXiv preprint arXiv:2306.00612 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.657 + ], + "angle": 0, + "content": "38. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Bi3d: Bidomain active learning for cross-domain 3d object detection. arXiv preprint arXiv:2303.05886 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.713 + ], + "angle": 0, + "content": "39. Yue, X., Zhang, Y., Zhao, S., Sangiovanni-Vincentelli, A., Keutzer, K., Gong, B.: Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2100-2110 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.769 + ], + "angle": 0, + "content": "40. Yue, X., Zheng, Z., Zhang, S., Gao, Y., Darrell, T., Keutzer, K., Vincentelli, A.S.: Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13834-13844 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.812 + ], + "angle": 0, + "content": "41. Zhang, B., Chen, T., Wang, B., Li, R.: Joint distribution alignment via adversarial learning for domain adaptive object detection. IEEE Transactions on Multimedia (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Zhang, B., Yuan, J., Shi, B., Chen, T., Li, Y., Qiao, Y.: Uni3d: A unified baseline for multi-dataset 3d object detection. arXiv preprint arXiv:2303.06880 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Yuan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "43. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "44. Zhao, S., Li, B., Yue, X., Gu, Y., Xu, P., Hu, R., Chai, H., Keutzer, K.: Multi-source domain adaptation for semantic segmentation. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "45. Zheng, Y., Huang, D., Liu, S., Wang, Y.: Cross-domain object detection through coarse-to-fine feature adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13766-13775 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "46. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_origin.pdf b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c2d15471654e3184e1e8db91457620989e59b114 --- /dev/null +++ b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/42cb7ad0-bed3-4fe6-bf59-566eca510270_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec37bc9570f24cc383e69effab162d6d287883fc33ee699f07578e034fb5f16a +size 801265 diff --git a/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/full.md b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9f68ab9eab744b41c91a5543b4305931930ec13d --- /dev/null +++ b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/full.md @@ -0,0 +1,303 @@ +# Reg-TTA3D: Better Regression Makes Better Test-time Adaptive 3D Object Detection + +Jiakang Yuan $^{1}$ , Bo Zhang $^{2}$ , Kaixiong Gong $^{3}$ , Xiangyu Yue $^{3}$ , Botian Shi $^{2}$ , Yu Qiao $^{2}$ , and Tao Chen $^{1,\text{忍}}$ + +1 School of Information Science and Technology, Fudan University, Shanghai, China jkyuan22@m.fudan.edu.cn, eetchen@fudan.edu.cn + +$^{2}$ Shanghai Artificial Intelligence Laboratory, Shanghai, China +3 The Chinese University of Hong Kong, Central Ave, Hong Kong + +Abstract. Domain Adaptation (DA) has been widely explored and made significant progress on cross-domain 3D tasks recently. Despite being effective, existing works fail to deal with rapidly changing domains due to the unpredictable test time scenarios and meanwhile fast response time requirement. Thus, we explore a new task named test-time domain adaptive 3D object detection and propose Reg-TTA3D, a pseudo-label-based test-time adaptative 3D object detection method. By investigating the factor that limits the detection accuracy, we find that regression is essential in this task. To make better regression, we first design a noise-consistency pseudo-label generation process to filter pseudo-labels with instability under noise interference and obtain reliable pseudo-labels. Then, confidence-guided regression refinement is introduced, which uses the box regression results of high-confidence boxes to supervise boxes with relatively low confidence, further making the predicted box size gradually approach the distribution of the target domain. Finally, to better update the regression layer and alleviate the class-imbalance issue, a class-balance EMA updating strategy is proposed. Experimental results on multiple cross-domain scenarios including cross-beam, cross-location, and cross-weather demonstrate that Reg-TTA3D can achieve comparable or even better performance compared to unsupervised domain adaptation works by only updating less than $0.1\%$ parameters within less than $1\%$ time. + +Keywords: Test-time adaptation $\cdot$ 3D object detection $\cdot$ Regression + +# 1 Introduction + +LiDAR-based 3D object detection [12, 18, 20, 21, 32, 36, 43] aims to predict the location and category of each object based on LiDAR point clouds and has attracted increasing attention recently due to its practical applications to autonomous driving. Although significant progress has been made, the stability of detectors under environmental changes, namely cross-domain 3D object detection, remains a challenge. However, the stability is important in practical usage + +![](images/27ce13672240b86d1095e3e95a07ea32faf712d1906c86435f7246bd609cfb4a.jpg) +Fig.1: Comparison between (a) previous domain adaptive 3D object detection and (b) Reg-TTA3D. Previous DA works perform training and testing in a split way. In contrast, Reg-TTA3D tackles domain discrepancies in an online manner and can deal with domain changes rapidly. + +![](images/0efd9b555507b097d540c0eb58dbfabe98fb0ae5369fe0ac70e76194b148cc3a.jpg) + +since the surrounding environment dynamically changes. For example, weather and geographical environment changes may be encountered in road scenes. + +Inspired by unsupervised domain adaption (UDA) techniques in 2D cross-domain tasks [3,5,8,13,39-41,44,45], some works [6,15,30,31,33,38] begin to deal with 3D domain discrepancies. ST3D [33] is a pioneer in using self-training to gradually transfer the knowledge learned from the source domain to the target domain. ReDB [6] attempts to deal with multi-class problems by designing cross-domain examination and overlapped boxes counting. Despite achieving large performance gains, the existing UDA model still requires a long training time to be adapted to the target domain, lacking the ability to rapidly deal with environmental changes in an online manner as shown in Fig. 1. Actually, in real applications, models are expected to be continuously updated at a low cost and quickly adapt to changing domains. + +To make the model quickly adapt to the changing environment in an online manner, test-time adaptation (TTA) [1,14,16,27,28] is introduced to adapt the model during test-time with only source model and unlabeled target data. Tab. 1 shows differences between TTA and other adaptation methods. In fact, the TTA technique has been explored in 2D classification tasks such as TENT [27] and SHOT [14] recently, but it is still under-explored in 3D vision fields, especially in detection tasks. To verify the effectiveness of previous TTA methods in the domain adaptive 3D object detection tasks, we conduct experiments by directly applying such methods (e.g., TENT [27], SHOT [14], MATE [16]) to our baseline detector (i.e., SECOND-IoU [32]). As shown in Tab. 3, directly applying previous methods results in a large performance drop, even significantly behind the results of source only which directly evaluates with the pre-trained source model. For instance, TENT [27] only achieves $31.34\%$ / $19.47\%$ in $\mathrm{AP}_{\mathrm{BEV}} / \mathrm{AP}_{\mathrm{3D}}$ , with $15.30\%$ / $9.36\%$ declination comparing to source only results. + +Accordingly, previous TTA methods are hard to apply to domain adaptive 3D detection tasks due to two major reasons. Firstly, from the perspective of updating parameter selection, previous methods often only update batch normalization (BN) layers to alleviate the domain shifts in the statistical information of network. However, this is not well suited for 3D detection task that also needs rich localization information. Besides, the optimization objectives of previous + +Table 1: Comparison of different settings to alleviate domain discrepancies. + +
SettingSource data Target data Online
Fine-tuningXxt,ytX
Unsupervised domain adaptationxs,s,ysxtX
Source-free domain adaptationXxtX
Test-time adaptationXxt✓
+ +methods are designed for classification tasks, which would make the model optimize for extracting features with more category semantic information rather than regression information. + +In order to tackle the challenges mentioned above, in this work, we propose Reg-TTA3D to perform test-time adaptation on 3D object detection tasks. To deal with the problem of updated parameter selection, we conduct extensive experiments and observe that high classification accuracy is already achieved by the source model, and the main factor limiting the performance improvement is the inaccurate regression results. Further, by only adapting regression layers, the model can achieve high performance at a low training cost as shown in Tab. 2 which motivates us to focus on regression adaptation in this work. Therefore, to make better regression, we propose noise-consistency pseudo-label generation (NPG) and confidence-guided regression refinement (CRR). The former module aims to filter boxes with instability under noise interference, i.e., noisy point clouds, while the latter one makes the box regression results align better with the distribution of the target domain box. Finally, we introduce a class-balanced EMA updating strategy (CBU) that can effectively update the regression layers, to improve the performance on multi-class simultaneously. + +Our contribution can be summarized as follows: + +1. We present a new task, namely test-time domain adaptive 3D object detection, and propose Reg-TTA3D to solve this task from the novel view of focusing on adapting only regression parameters. +2. We propose a noise-consistency pseudo-label generation process, a confidence-guided regression refinement loss, and a class-balanced EMA updating strategy, to obtain better detection results and boost the model's performance in an efficient and effective way. +3. Experiments show that Reg-TTA3D reaches a comparable or even better performance by updating less than $0.1\%$ parameters within less than $1\%$ time compared to UDA works, and significantly outperforms previous TTA methods on multiple cross-domain scenarios. + +# 2 Related Works + +# 2.1 LiDAR-based 3D Object Detection + +With the development of autonomous driving, LiDAR-based 3D object detection has attracted increasing attention. Current prevailing LiDAR-based 3D detectors can be roughly divided into point-based, voxel-based, and point-voxel-based methods according to the point cloud processing procedure. Point-based + +methods [20, 35, 43] first sample a subset of point clouds, then extract features and generate proposals from sampled points. Among them, PointRCNN [20] uses a two-stage framework that is composed of a bottom-up proposal generation process and a proposal refinement process. In contrast, voxel-based methods [7, 12, 32, 36, 46] first divide disordered points into regular grids which are then fed into the convolutional network. SECOND [32] is a pioneer in using the 3D sparse backbone to extract features. CenterPoint [36] introduces a center-based detector and achieves promising performance. Point-voxel-based [18, 19] methods integrate the features extracted from points and voxels to obtain better feature representations. However, existing 3D detectors fail to deal with domain gaps which is essential in practical usage. + +# 2.2 Test-time Adaptation + +Test-time adaptation [1, 4, 10, 22, 24, 27, 28] aims to rapidly adapt the source model to target domain without accessing source domain data during test time, which is more practical in the real world since it can handle dynamic domain shifts. TENT [27] explores test-time adaptation for the first time and updates the batch normalization layers using designed entropy minimization loss. SHOT [14] deals with such a challenging task by maximizing the mutual information. More recently, MATE [16] extends test-time adaptation to 3D point clouds and finds that the model can be adapted effectively by masking and reconstruction. + +Although TTA has been extensively explored in 2D scenarios, it is still under-explored in 3D tasks, especially in large-scale point cloud scenarios (e.g., autonomous driving scenarios). Besides, most previous TTA works mainly focus on classification tasks, such as classification [1, 16, 27] and segmentation [22], and they fail to tackle detection tasks since detection considers classification and box regression at the same time. In contrast, Reg-TTA3D considers TTA for 3D object detection tasks from the perspective of regression, which is more suitable for 3D detection tasks. + +# 2.3 Domain Adaptation for 3D Object Detection + +3D detectors often suffer large performance drops when training and testing under different domains, which makes them unstable when the surrounding environment dynamically changes. Recently, researchers have tried to tackle such a problem using domain adaptation techniques [6,17,29,33,34,38]. SN [29] proposes a statistical method to align the size of bounding boxes between source and target domains. ST3D [33] and $\mathrm{ST3D + + }$ [34] design a self-training pipeline and mitigate domain gaps using a memory bank and curriculum data augmentation. To deal with the class imbalance issue, ReDB [6] proposes a pseudo-label generation pipeline to obtain reliable, diverse, and class-balanced pseudo-labels. However, the existing works cannot deal with the emergency changes in the environment due to the time-consuming and large amount of calculation. + +Table 2: Performance of updating different parameters. + +
Tuning layersWaymo → KITTInuScenes → KITTIParameters
Batch normalization52.84 / 31.7731.86 / 18.259.08k
Classification layers52.82 / 36.3731.51 / 19.761.17k
Regression layers59.38 / 45.5433.63 / 21.802.73k
Detection head59.62 / 46.0832.51 / 20.394.68k
All parameters60.78 / 46.7133.23 / 19.6012.12M
+ +# 3 Method + +The overall framework of Reg-TTA3D is shown in Fig. 2. To better illustrate our proposed method, we start by introducing the problem formulation and our baseline in Sec. 3.1. Then, we give an analysis of the updated parameters selection in Sec. 3.2. Finally, we detail each module of Reg-TTA3D and give the overall objective and adaptation strategy in Sec. 3.3 and Sec. 3.4. + +# 3.1 Preliminary + +Problem Formulation. Following typical test-time adaptation setting [22, 27], given the model $M_S$ pre-trained on the source domain $D_S = \{(x_i^S, y_i^S)\}_{i=1}^{n_S}$ and an unlabeled target set $D_T = \{x_j^T\}_{j=1}^{n_T}$ , where the source and target domain follow different distributions, our goal is to adapt the source model to the target domain without accessing source domain data in an online manner. During adaptation, in each timestep, a batch of unlabeled data $[x_1^t, x_2^t, \dots, x_N^t]$ is sampled from $D_T$ to update the model, where $N$ is batch size. + +Baseline Introduction. Following [6, 34], we use SECOND-IoU [32] as our baseline detector which is composed of a feature extractor (i.e., 3D and 2D backbone) to obtain bird's eye view (BEV) features, a region proposal network (RPN) to generate proposals and a detection head to provide the classification and regression results. The detector is trained using the following loss function: + +$$ +\mathcal {L} _ {d e t} = \mathcal {L} _ {r p n} + \mathcal {L} _ {r c n n}, \tag {1} +$$ + +where $\mathcal{L}_{rpn}$ represents the loss of RPN and $\mathcal{L}_{rcnn}$ denotes box refinement loss. + +# 3.2 Updated Parameters Selection + +The aim of Reg-TTA3D is to transfer the source model to the target domain rapidly so that the model can tackle the continuous changes in the surrounding environment. To achieve this goal, the amount of updated parameters we select needs to be relatively small, and the accuracy needs to be largely improved by only updating selected parameters. + +Previous TTA works [22, 27] mainly focus on adapting batch normalization (BN) layers to mitigate the domain discrepancies, since BN layers contain the statistics of domains. Such an adaptation strategy is effective for domain adaptive classification tasks, as more accurate semantic representations can be obtained, which is especially important for classification. However, only adapting + +![](images/28479010401af4e409f3208040817cf94873654a0ae8bdf1a5e9b88ef27a9663.jpg) +Fig. 2: The overall framework of the proposed Reg-TTA3D, which employs SECOND-IoU [32] as the baseline detector and consists of noisy-consistency pseudo-label generation process (NPG), confidence-guided regression refinement (CRR) and class-balanced EMA updating strategy. The data are first fed into the teacher model to generate pseudo-labels as the supervision to the student model using NPG. Then, the student model is trained using data from another view and supervised by detection loss and refinement loss obtained by CRR. Finally, the class-balanced EMA updating strategy is used to update the teacher model. Note that the only difference between the teacher model and the student model lies in the regression layers. + +BN layers cannot rapidly transfer the source detector to the target domain since detection tasks contain both classification and regression. Some works [16, 26] try to solve this problem by updating the entire model but suffer from huge computational costs. To explore an effective and efficient method, by conducting extensive experiments, we find that the model can adapt to the target domain efficiently by only updating regression layers. As shown in Tab. 2, we fine-tune the parameters using $1\%$ KITTI training data for 1 epoch and observe that only fine-tuning the regression layers can reach relatively high performance and even surpass the performance of fine-tuning the entire model. This is because in the autonomous driving scenario, there exists a large difference between categories such as car and pedestrian, the source model can already achieve high classification accuracy and the main factor for limiting the detection accuracy is the inaccurate regression results. Based on the observation, in this work, we focus on studying how to only adapt the parameters of regression layers and make better regression. + +# 3.3 Reg-TTA3D + +As mentioned above, the quality of box regression is essential to TTA for 3D object detection. To make better regression, we propose noise-consistency pseudolabel generation to obtain better regression supervision, and confidence-guided regression refinement to further refine the regression results. Besides, to effectively update the regression layers, we design a class-balanced EMA updating strategy to avoid overfitting the model to a single class. + +Noise-consistency Pseudo-label Generation. As mentioned in [34, 38], the pseudo-labels directly generated by the source model contain a lot of noise due + +to domain discrepancies. Previous methods try to alleviate pseudo-label noise by weighting pseudo-labels obtained in different epochs [33, 34], or adding the groundtruth instances from the source domain to target domain frames [6]. However, these methods cannot apply to test-time adaptation due to the inaccessibility of source data and online adaptation. + +Inspired by test-time augmentation methods which comprehensively consider the model predictions under different data augmentation methods to provide final results. Specifically, the model's prediction of real objects is relatively stable, while being uncertain for the noise regions. To this end, we propose noise-consistency pseudo-label generation to obtain more accurate pseudo-labels, by measuring the stability of model predictions when adding random noise to regions of interest (RoIs). In detail, given proposals generated by the teacher model $B^{t} = \{(c_{x}^{t}, c_{y}^{t}, c_{z}^{t}, l^{t}, w^{t}, h^{t}, \theta^{t})_{i}\}_{i=1}^{n}$ , its corresponding features $\{f_{i}^{t}\}_{i=1}^{n}$ and confidence scores $\{s_{i}^{t}\}_{i=1}^{n}$ where $(c_{x}^{t}, c_{y}^{t}, c_{z}^{t})$ and $(l^{t}, w^{t}, h^{t})$ denote the center and the size of the box respectively, $\theta^{t}$ is the rotation angle and $n$ is the number of boxes. We can obtain boxes with noise $B^{t'} = \{(c_{x}^{t'}, c_{y}^{t'}, c_{z}^{t'}, l^{t'}, w^{t'}, h^{t'}, \theta^{t'})\}_{i=1}^{n}$ , corresponding features $\{f_{i}^{t'}\}_{i=1}^{n}$ and scores $\{s_{i}^{t'}\}_{i=1}^{n}$ , by adding random noise to boxes' size and angle. To measure the stability of RoIs, we calculate the inconsistency rate as follows: + +$$ +R _ {i n c o n s i s t e n c y, i} = 1 - \frac {\left| s _ {i} ^ {t} - s _ {i} ^ {t ^ {\prime}} \right|}{\operatorname {I o U} \left(b _ {i} ^ {t} , b _ {i} ^ {t ^ {\prime}}\right)}, \tag {2} +$$ + +where $\mathrm{IoU}(b_i^t, b_i^{t'})$ denotes the intersection of union (IoU) between $b_i^t \in B^t$ and $b_i^{t'} \in B^{t'}$ . A high inconsistency rate represents that the model prediction is unstable to the current object which is more likely to be a noise region. Therefore, we filter the boxes with high inconsistency rate by a threshold $\tau$ to get boxes and scores $\hat{B}, \{\hat{s}_i\}_{i=1}^n$ and $\hat{B}', \{\hat{s}_i'\}_{i=1}^n$ with stability under noise interference. Then, we ensemble $\hat{B}$ and $\hat{B}'$ by confidence scores as follows: + +$$ +\tilde {B} = \left\{\tilde {b _ {i}} \right\}, \quad \tilde {b _ {i}} = \left\{ \begin{array}{l l} \hat {b} _ {i}, & \hat {s} _ {i} > \hat {s} _ {i} ^ {\prime}, \hat {b} _ {i} \in \hat {B} \\ \hat {b} _ {i} ^ {\prime}, & \hat {s} _ {i} < \hat {s} _ {i} ^ {\prime}, \hat {b} _ {i} ^ {\prime} \in \hat {B} ^ {\prime} \end{array} , \right. \tag {3} +$$ + +After obtaining $\tilde{B}$ , the boxes are filtered by non-maximum suppression (NMS) with a relatively high confidence threshold, to remove redundant boxes and boxes with low confidence. Finally, the pseudo-labels $B^{npg}$ and $\{l_i^{npg}\}_{i=1}^n$ with high stability and confidence can be acquired where $\{l_i^{npg}\}_{i=1}^n$ is classification results. Confidence-guided Regression Refinement. Although lots of boxes containing noisy point cloud can be removed by measuring the stability under noise injection, there still exist the problems of inaccurate box regression results. This is because the distributions of bounding box size are inconsistent between source and target domains as mentioned in [29,37,42], and it is difficult to mitigate such domain discrepancy with only NPG. By analyzing the distribution of predicted bounding boxes, we find that boxes with higher confidence scores align better with the distribution of the target domain boxes as shown in Fig. 3, since precise boxes correspond to high-quality RoI features that are less disturbed by background representations. Based on this observation, we design a confidence-guided regression refinement to obtain more precise regression results of box size. + +![](images/64c255b7c3a24accc9e7467597ecaa4b71e4e0d5cc58abe748fbcd8370c03c77.jpg) +Fig. 3: Distribution of pseudo box size under different thresholds. PS denotes pseudo boxes. + +In particular, given predicted boxes generated by the student model $B^{s} = \{(c_{x}^{s}, c_{y}^{s}, c_{z}^{s}, l^{s}, w^{s}, h^{s}, \theta^{s})_{j}\}_{j=1}^{k}$ and scores $\{s_{j}^{s}\}_{j=1}^{k}$ of a batch of frames, where $k$ is the total number of predicted boxes. We sort boxes belonging to the same class according to scores in descending order, and sample a proportion of boxes with high scores to calculate the mean value of sample boxes size $(\bar{l}^{s}, \bar{w}^{s}, \bar{h}^{s})$ . Please refer to our supplementary material for the study of the proportion of sampled boxes. Once obtained $(\bar{l}^{s}, \bar{w}^{s}, \bar{h}^{s})$ , we can use it as a supervision signal to make the box regression value closer to the distribution of the target domain as follows: + +$$ +\mathcal {L} _ {\text {r e f i n e}} = \mathbb {E} _ {i = 1} ^ {n} \Sigma_ {d \in [ d _ {l}, d _ {w}, d _ {h} ]} d, +$$ + +$$ +w h e r e d = \left\{ \begin{array}{l l} 0, & | \left(z ^ {s} - \bar {z} ^ {s}\right) | < p \\ \left(z ^ {s} - \bar {z} ^ {s}\right) ^ {2}, & | \left(z ^ {s} - \bar {z} ^ {s}\right) | \geq p \end{array} , \right. \tag {4} +$$ + +where $z \in [l^s, w^s, h^s]$ , $\overline{z} \in [\overline{l}^s, \overline{w}^s, \overline{h}^s]$ , $n$ is batch size and $p$ is set 0.1 to allow the size of objects to fluctuate within a certain range. The model can predict more accurate box size when using $\mathcal{L}_{\text{refine}}$ for optimization. + +Class-balanced EMA Updating. To make the model more stable, we update the teacher regression layers by exponential moving average (EMA). However, the model tends to overfit to the category with a large number of instances (i.e., car). As a result, the detection performance on categories with few instances will decrease. To make the model better balance the performance improvement for each class, we propose a class-balanced EMA updating strategy to make the model update more when the class is balanced, while updating less when a single category dominates. Specifically, given the pseudo-labels obtained from NPG $B^{npg} = \{b_i^{npg} \in \mathbb{R}^7\}$ and its corresponding class labels $\{c_i^{npg}\}$ . We calculate the proportion $\{p_c\}_{c=1}^C$ of each category in pseudo-labels according to $\{c_i^{npg}\}$ , where $C$ is the number of total categories and $\Sigma_{c=1}^C p_c = 1$ . Then the variance of $\{p_c\}_{c=1}^C$ is calculated and normalized to [0.99, 0.999] to obtain the momentum $\alpha$ . A high value of $\alpha$ means that a severe class-imbalance issue exists in the current batch and the model should update less. Finally the parameter of the teacher regression layer $\omega_{reg}^t$ can be updated by the following formula: + +$$ +\omega_ {r e g} ^ {t} \leftarrow \alpha \omega_ {r e g} ^ {t} + (1 - \alpha) \omega_ {r e g} ^ {s}, \tag {5} +$$ + +where $\omega_{reg}^{s}$ is the parameters of student regression layers. + +# Algorithm 1 Single iteration of Reg-TTA3D strategy + +Input: A batch of frames $\{x_{i}\}_{i = 1}^{n}$ and $\{\dot{x}_i\}_{i = 1}^n$ , where $\dot{x}_i$ is obtained by performing data augmentation on $x_{i}$ + +Output: Predictions of current batch $B^{t}$ and $\{l_i^t\}_{i = 1}^n$ + +1: Get the predictions $B^{t} = \{b_{i}^{t}\in \mathbb{R}^{7}\}_{i = 1}^{n}$ and $\{l_i^t\}_{i = 1}^n$ of $\{x_{i}\}_{i = 1}^{n}$ using teacher model, $B^{t}$ and $\{l_i^t\}_{i = 1}^n$ are the detection results of current batch. +2: Use NPG to obtain boxes $B^{npg}$ and $\{l_i^{npg}\}_{i=1}^n$ with stability under noise interference as pseudo-labels. +3: Get the predictions of student model $B^{s}$ and $\{s_{i}^{s}\}_{i = 1}^{n}$ of $\{\dot{x}_i\}_{i = 1}^n$ , supervise the student model by $B^{npg}$ and $\{l_i^{npg}\}_{i = 1}^n$ to get $\mathcal{L}_{det}$ . +4: Calculate $\mathcal{L}_{\text{refine}}$ using CRR. +5: Optimize the student model using Eq. (6). +6: Update the teacher model using CBU and Eq. (5). +7: Return: Current predictions $B^{t}$ and $\{l_i^t\}_{i = 1}^n$ . + +# 3.4 Overall Objective and Test-time Adaptation Strategy + +Overall Objective. The overall objective to optimize the student regression layers can be formulated as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {d e t}} + \mathcal {L} _ {\text {r e f i n e}}, \tag {6} +$$ + +where $\mathcal{L}_{det}$ and $\mathcal{L}_{refine}$ are defined in Eq. (1) and Eq. (4). + +Test-time Adaptation Strategy. The test-time adaptation strategy of Reg-TTA3D can be summarized in Algorithm 1. + +# 4 Experiments + +# 4.1 Experimental Setup + +Datasets and Cross-domain Settings. We conduct experiments on four commonly-used datasets: Waymo [23], KITTI [9], nuScenes [2] and nuScenes-C [11]. To verify the effectiveness of Reg-TTA3D, we consider three cross-domain settings including cross-location (i.e., Waymo $\rightarrow$ KITTI), cross-beam (i.e., Waymo $\rightarrow$ nuScenes, nuScenes $\rightarrow$ KITTI) and cross-weather (i.e., nuScenes $\rightarrow$ nuScenes-C). Note that in the cross-weather setting, we use the medium-difficulty fog and snow scenarios in the nuScenes-C dataset. Following ReDB [6], in the first two cross-domain settings, we use the KITTI evaluation metric to evaluate for three common classes in autonomous driving scenarios (i.e., car, pedestrian, and cyclist), we report the average precision in both 3D (i.e., $\mathrm{AP}_{\mathrm{3D}}$ ) and BEV (i.e., $\mathrm{AP}_{\mathrm{BEV}}$ ) over 40 recall positions, with IoU threshold 0.7 for car and 0.5 for pedestrian and cyclist. In the cross-weather setting, we use the official nuScenes evaluation metric and report mean average precision (mAP). + +Implementation Details. Following [6], we evaluate the proposed Reg-TTA3D on SECOND-IoU [32], and the source model is trained for all categories using random object scaling (ROS) augmentation method. We set hyperparameter $\tau$ + +to 1.5 and $p$ to 0.1 for all cross-domain settings and use Adam optimizer with learning rate of $1 \times 10^{-3}$ . We utilize commonly-used data augmentation methods (e.g., random world flip, random world rotation, random world scaling) to get $\{\dot{x}_i\}_{i=1}^n$ . Our code is built on 3DTrans [25]. + +Comparison Baselines. To verify the effectiveness of Reg-TTA3D, we mainly compare our methods to unsupervised domain adaptive 3D object detection and previous test-time adaptation methods that we apply to 3D object detection. UDA methods: (1) ST3D [33] is a self-training method that continuously updates the pseudo-labels, (2) $\mathrm{ST3D + + }$ [34] provides more insight analysis and extends ST3D to get better results, (3) ReDB [6] is a pioneer to perform domain adaptive 3D detection on multiple categories. TTA methods: (1) TENT [27] proposes an entropy minimization method which focuses on adapting BN layers, (2) SHOT [14] optimizes feature extraction module using both information maximization and pseudo-labeling. (3) MATE [16] uses masked autoencoder to apply test-time adaptation to 3D classification. + +# 4.2 Main Results + +Results on cross-location and cross-beam scenarios. We first evaluate Reg-TTA3D on cross-location and cross-beam scenarios and compare the results with previous UDA methods. As shown in Tab. 3, Reg-TTA3D can greatly improve the detection accuracy on Waymo $\rightarrow$ KITTI and nuScenes $\rightarrow$ KITTI settings, largely narrow the performance gap between source only and oracle. For example, in the Waymo $\rightarrow$ KITTI setting, our Reg-TTA3D can achieve $61.78\%$ / $49.96\%$ in $\mathrm{AP}_{\mathrm{BEV}} / \mathrm{AP}_{\mathrm{3D}}$ , even surpass UDA methods such as ST3D [33] and ST3D++ [34] and it is comparable to the state-of-the-art UDA method (i.e., $61.14\%$ / $50.10\%$ in $\mathrm{AP}_{\mathrm{BEV}} / \mathrm{AP}_{\mathrm{3D}}$ reported in ReDB [6]). In nuScenes $\rightarrow$ KITTI and Waymo $\rightarrow$ nuScenes cross-domain scenarios, the performance of Reg-TTA3D is also better than some UDA methods (e.g., ST3D [33]) which needs source domain data and training for multiple epochs. In contrast, Reg-TTA3D only needs target data and one epoch of adaptation which greatly reduces the time consumption. + +Further, we compare Reg-TTA3D with the previous TTA methods in Tab. 3. We find that previous TTA methods suffer performance drops when applied to 3D detection tasks (e.g., TENT [27] reduce the detection accuracy to $31.34\%$ / $19.47\%$ in the Waymo $\rightarrow$ KITTI setting). This is mainly because the detection task is composed of classification and regression and existing methods only consider the classification tasks. As a result, the model will concentrate on extracting effective representations for the classification tasks and ignore the localization information. However, our Reg-TTA3D fully considers the challenges in detection tasks and can improve the performance on multiple categories. + +Results on the cross-weather scenario. To more comprehensively evaluate the effectiveness of our proposed method, we conduct experiments in a more realistic scenario (i.e., cross-weather). Since nuScenes-C dataset only has the validation set, we only compare Reg-TTA3D with existing TTA methods. It can be seen in Tab. 4 that Reg-TTA3D can improve the performance on both snow + +Table 3: Results on cross-location and cross-beam scenarios. Following [6, 33], we use SECOND-IoU [32] as out baseline detector. Source Only represents that the pretrained detector is directly evaluated on the target domain, and Oracle denotes the detection results obtained using the fully annotated target training data. Red: Best UDA method, Blue: Best TTA method. + +
TASKMETHODSETTINGCARPEDESTRIANCYCLISTMEAN AP
Waymo → KITTISOURCE ONLY-51.48 / 19.7840.80 / 31.2647.63 / 35.4546.64 / 28.83
SN76.61 / 54.1452.48 / 48.2034.56 / 32.7454.55 / 45.03
ST3DUDA77.62 / 49.2444.45 / 42.0447.74 / 42.9556.60 / 44.70
ST3D++77.68 / 50.0349.09 / 46.1951.50 / 47.7059.42 / 47.97
ReDB80.37 / 54.1251.01 / 48.2052.05 / 47.9761.14 / 50.10
TENTTTA71.98 / 39.014.54 / 3.9517.51 / 15.4531.34 / 19.47
SHOT79.52 / 41.3133.52 / 29.0240.79 / 35.8351.27 / 35.38
MATE75.17 / 30.6439.94 / 36.9243.72 / 40.2252.94 / 35.92
REG-TTA3D81.60 / 56.0350.89 / 45.4352.85 / 48.4361.78 / 49.96
ORACLE-83.29 / 73.4546.64 / 41.3362.92 / 60.3264.28 / 58.37
Waymo → nuScenesSOURCE ONLY-30.64 / 17.181.81 / 0.580.97 / 0.8811.14 / 6.22
SN29.56 / 17.781.33 / 1.072.92 / 2.6111.27 / 7.15
ST3DUDA28.42 / 17.831.64 / 1.394.01 / 3.5411.36 / 7.59
ST3D++28.87 / 19.151.82 / 1.584.09 / 3.7411.60 / 8.16
ReDB30.12 / 18.562.47 / 2.146.56 / 5.1913.05 / 8.63
TENTTTA16.40 / 3.430.82 / 0.250.28 / 0.215.83 / 1.29
SHOT12.26 / 1.421.56 / 1.360.22 / 0.204.68 / 0.99
MATE8.64 / 3.991.72 / 0.970.24 / 0.083.53 / 1.68
REG-TTA3D30.54 / 18.152.04 / 1.755.01 / 4.7012.53 / 8.20
ORACLE-51.88 / 34.8725.24 / 18.9215.06 / 11.7330.73 / 21.84
nuScenes → KITTISOURCE ONLY-39.15 / 7.6521.54 / 16.876.31 / 2.4422.33 / 8.98
SN56.08 / 28.6723.05 / 16.845.11 / 2.3228.08 / 15.94
ST3DUDA71.50 / 48.0922.64 / 17.617.86 / 5.2034.00 / 23.64
ST3D++69.90 / 44.6224.11 / 18.2010.14 / 6.3933.75 / 23.07
ReDB74.23 / 51.3125.95 / 18.3813.82 / 8.6438.00 / 26.11
TENTTTA43.35 / 19.839.40 / 5.772.71 / 2.7018.49 / 9.43
SHOT60.24 / 23.0117.51 / 9.442.47 / 1.6626.74 / 11.37
MATE61.89 / 27.1315.98 / 13.112.94 / 2.4526.93 / 14.23
REG-TTA3D68.73 / 44.5622.78 / 17.7110.21 / 7.1233.90 / 23.13
ORACLE-83.29 / 73.4546.64 / 41.3362.92 / 60.3264.28 / 58.37
+ +and fog scenarios (e.g., $45.68\% \rightarrow 47.09\%$ mAP on the Normal $\rightarrow$ Snow setting). The results verify that our proposed method can handle dynamic environment changes in road scenes and can be applied in more practical scenarios. + +# 4.3 Insight Analyses + +Sensitivity to Detector Architecture. Following [6], to verify the effectiveness of Reg-TTA3D on different types of detectors, we further conduct experiments on nuScenes $\rightarrow$ KITTI cross-domain scenario using PointRCNN [20] as the baseline detector. As show in Tab. 5, our Reg-TTA3D outperforms all TTA methods and some UDA methods. Note that compared to the source-free domain adaptation method (i.e., SF-UDA $^{3D}$ [17]) which is inaccessible to the source data and need to adapt for multiple epochs, our Reg-TTA3D adapted on multi-class can achieve a better result on the car category (e.g., $50.87\%$ compared to $49.80\%$ + +Table 4: Results on cross-weather scenario (i.e., nuScenes Normal $\rightarrow$ Fog and Normal $\rightarrow$ Snow). Fog and snow data are from nuScenes-C [11] dataset. We use CenterPoint [36] as our baseline detector since it is the most widely used detector on nuScenes. + +
METHODTaskmAPCARTRUCKCV.BUSTRAILERBARRIERMOTOR.BICYCLEPED.TC.
SOURCE ONLY48.4079.1252.588.1067.5331.0064.407.3026.6581.2865.97
TENT41.8373.6215.237.5858.0421.8560.451.5623.8477.5952.24
SHOTFog42.3172.6340.707.5958.1822.2461.710.8526.4577.8454.87
MATE23.9764.6124.183.3427.2614.1120.450.007.6749.6728.33
REG-TTA3D49.6780.3054.4310.8668.6534.5363.527.7927.0782.2067.28
SOURCE ONLY45.8680.7448.774.6864.1528.8462.629.0224.7278.3756.62
TENTSnow40.5175.0939.376.5157.5719.9659.821.8923.3376.7844.73
SHOT42.1574.6639.747.0158.1221.4662.941.1927.3078.3950.66
MATE23.4363.9822.573.2025.3013.0622.790.007.640.5024.81
REG-TTA3D47.0982.0149.717.1968.1630.5462.966.3326.2980.3757.30
+ +Table 5: Performance comparisons (AP3D) with PointRCNN as baseline detector on nuScenes $\rightarrow$ KITTI task. $\ddagger$ indicates the results reported in the original paper. Red: Best UDA method, Blue: Best TTA method. + +
MethodSettingCARPEDESTRIANCYCLISTAVERAGE
EASYMOD.HARDEASYMOD.HARDEASYMOD.HARDEASYMOD.HARD
SOURCE ONLY-42.7732.1128.7543.2637.2933.163.284.094.1829.7724.6022.03
SN66.5650.3245.9242.9637.1532.459.077.577.4239.5331.6828.60
ST3DUDA48.8541.9038.9245.6738.7133.0926.5019.3518.3840.3433.3230.13
ST3D++60.4549.3645.8850.7742.4336.6427.2017.9416.5246.1436.5833.01
REDB71.4557.953.9152.3244.3337.9545.1332.9331.0556.3045.0540.97
SF-UDA3D‡SFDA68.8049.8045.00---------
TENTTTA42.5532.4127.6130.7126.0923.784.243.663.3225.8320.7018.23
SHOT30.0923.0920.4119.2416.2914.312.431.701.5117.2513.6912.07
REG-TTA3D65.2050.8746.7245.8042.5534.9040.2527.8725.2950.4140.4335.63
+ +reported by SF-UDA $^{3D}$ in moderate level) in one epoch of adaptation even if SF-UDA $^{3D}$ is trained on the unfair single class. The results further show that our proposed methods can apply to different detector architectures. + +Ablation Studies. To further verify the effectiveness of Reg-TTA3D, we conduct extensive experiments to investigate the impact of each component we proposed including NPG, CRR and CBU. As shown in Tab. 6, the improvement (i.e., $0.92\% \mathrm{AP}_{3\mathrm{D}}$ ) of using naive pseudo-labels is relatively small compared to the performance of source only since pseudo-labels contain lots of noise samples and the regression results are not accurate due to domain discrepancies. Reg-TTA can effectively handle such a problem by proposing NPG and CRR modules, with $2.90\%$ and $3.60\%$ improvement respectively compared to source only. Besides, when combining these two modules, the performance can be further improved because NPG focuses on filter noise samples and CRR concentrates on making more accurate regression results. Further, by combining with CBU which makes the model update in a class-balance manner, the performance can be further boosted and achieve $49.96\% \mathrm{AP}_{3\mathrm{D}}$ . + +Sensitivity to Hyperparameters. We ablate the impact of hyperparameters $\tau$ (i.e., the threshold of inconsistency rate) and $p$ in this subsection. As shown in Tab. 7, we first set $\tau$ to 1.0, 1.5, 2.0 respectively and observe that the fluctuation + +Table 6: Ablation studies on each component. PS, NPG, CRR and CBU represent using naive pseudo-labels, noise-consistency pseudo-label generation, confidence-guided regression refinement and class-balanced EMA updating, respectively. The experiments are conducted on Waymo $\rightarrow$ KITTI scenario using SECOND-IoU as the baseline detector. + +
PS NPGCRRCBUmean APBEVmean AP3D
---57.8942.01
✓--58.7342.93
✓✓-58.8944.91
✓-✓59.8445.61
✓--59.7344.05
✓✓✓60.5747.97
✓✓-60.1246.62
✓-✓60.5948.39
✓✓✓61.7849.96
+ +Table 7: Ablation studies of hyperparameters $\tau$ and $p$ . The experiments are conducted on Waymo $\rightarrow$ KITTI scenario using SECOND-IoU. Note that when ablate on $\tau$ , we set $p = 0.1$ and when ablate on $p$ , we set $\tau = 0.15$ . + +
τAPBEV / AP3D
1.060.92 / 49.12
1.561.78 / 49.96
2.060.44 / 48.88
pAPBEV / AP3D
0.0560.29 / 48.26
0.1061.78 / 49.96
0.1561.56 / 49.63
+ +![](images/8d3ba54ca5c5066e56034224777598f6050e0ca924f67c7e7ed8926fd83e99d6.jpg) +Fig. 4: The relationship between detection accuracy and the iteration number. + +Table 8: Ablation studies on costs. The experiments are conducted on $\mathrm{W} \rightarrow \mathrm{K}$ scenario using SECOND-IoU. We use Tesla A100 to conduct the experiments. + +
MethodTime Cost N_GPUsParam.Results
ST3D4h412.12M 56.60 / 44.70
ST3D++3.5h412.12M 59.42 / 47.97
ReDB3.2h412.12M 61.14 / 50.10
TENT4min19.08K 31.34 / 19.47
MATE6min10.71M 52.94 / 35.92
Reg-TTA3D4.5min12.73K 61.78 / 49.96
+ +of the performance is $1.34\% / 1.08\%$ . Then, by setting $p$ to different values (i.e., 0.05, 0.1, 0.15), we find that there is only little fluctuation and the best performance can be achieved when set to 0.1. The results further show that our method is robust to these hyperparameters. + +Analysis on Pseudo-labeling Accuracy. In this section, we measure the accuracy of each iteration during adaptation. Fig. 4 reflects the detection accuracy on Waymo $\rightarrow$ KITTI as iteration increases. It can be seen that the accuracy continuously improves since the model is gradually adapted to the target domain which also verifies the stability of Reg-TTA3D during adaptation. + +Analysis on Costs. Further, we report the time cost and total updated parameters in Tab. 8. Notably, Reg-TTA3D achieves comparable results to UDA methods adapting only less than $0.1\%$ of parameters within less than $1\%$ of time. When compared to TTA methods, our proposed method can achieve much higher performance with less updated parameters and comparable time cost which further verifies that our Reg-TTA3D is an effective and efficient method. + +Ablation Studies of the momentum. We ablate the momentum in EMA updating and compare it to our class-balance EMA updating strategy. Specifi- + +![](images/0b0854b088a69e6979bc96ee42082480df6f82a4057f122ec79e13a8aee0450c.jpg) +Fig. 5: The impact of momentum and com- Fig.6: ablation studies on the proportions with CBU. tion of selected samples. + +![](images/5d8d709697e33b8ee75c10460202c65a1f9776a3559567967521f95982b8025b.jpg) + +cally, we conduct the experiment for momentum between 0.99 and 0.999 at 0.001 intervals. As shown in Fig. 5, the result demonstrates that CBU can better update the regression layer due to it can balance update the model and improve the performance on multiple classes at the same time. + +Ablation Studies on the proportion of Selected samples in CRR. In our manuscript, we use $20\%$ boxes with the highest scores to supervise box sizes with relatively low scores. Here, we analyze the impact on the proportion of selected samples. As shown in Fig. 6, when the proportion is relatively high, the average box size will be distributed far from the target domain, and when set to relatively small values, the small number of samples results in the averages being affected by individual instances. Based on the experiments and analysis, we set the proportion to $20\%$ . + +# 5 Conclusion + +In this work, for the first time, we explore test-time adaptation in 3D object detection tasks and propose Reg-TTA3D, which handles rapid changes in the surrounding environment from the perspective of updating parameters selection and adaptation algorithm. By adapting using the proposed noise-consistency pseudo-label generation process, confidence-guided regression refinement loss and class-balanced EMA updating strategy and only updating regression layers, Reg-TTA3D can achieve comparable performance to some UDA methods at a low cost. + +# 6 Limitation + +Although the proposed Reg-TTA3D can handle dynamic changes in the surrounding environments such as continuously changing weather. It cannot deal with emergencies such as some corner cases which is also important in real-world autonomous driving. + +# Acknowledgement + +This work is supported by National Key Research and Development Program of China (No. 2022ZD0160101), National Natural Science Foundation of China (No. 62071127, and 62101137), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No.2021SHZDZX0103). This work is also supported by the National Key R&D Program of China (Grant No. 2022ZD0160104), and Shanghai Rising Star Program (Grant No. 23QD1401000). The computations in this research were performed using the CFFF platform of Fudan University. + +# References + +1. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8344-8353 (2022) +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020) +3. Cai, Q., Pan, Y., Ngo, C.W., Tian, X., Duan, L., Yao, T.: Exploring object relation in mean teacher for cross-domain detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11457-11466 (2019) +4. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 295-305 (2022) +5. Chen, Y., Li, W., Sakaridis, C., Dai, D., Van Gool, L.: Domain adaptive faster r-cnn for object detection in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3339-3348 (2018) +6. Chen, Z., Luo, Y., Huang, Z., Wang, Z., Baktashmotlagh, M.: Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudolabeling. arXiv preprint arXiv:2307.07944 (2023) +7. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1201-1209 (2021) +8. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: International conference on machine learning. pp. 1180-1189. PMLR (2015) +9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (2012) +10. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems 35, 6204-6218 (2022) +11. Kong, L., Liu, Y., Li, X., Chen, R., Zhang, W., Ren, J., Pan, L., Chen, K., Liu, Z.: Robo3d: Towards robust and reliable 3d perception against corruptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 19994-20006 (2023) + +12. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019) +13. Liang, J., Gong, K., Li, S., Liu, C.H., Li, H., Liu, D., Wang, G., et al.: Pareto domain adaptation. Advances in Neural Information Processing Systems 34, 12917-12929 (2021) +14. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020) +15. Luo, Z., Cai, Z., Zhou, C., Zhang, G., Zhao, H., Yi, S., Lu, S., Li, H., Zhang, S., Liu, Z.: Unsupervised domain adaptive 3d detection with multi-level consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8866-8875 (2021) +16. Mirza, M.J., Shin, I., Lin, W., Schriebl, A., Sun, K., Choe, J., Kozinski, M., Possegger, H., Kweon, I.S., Yoon, K.J., et al.: Mate: Masked autoencoders are online 3d test-time learners. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16709-16718 (2023) +17. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771-780. IEEE (2020) +18. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529–10538 (2020) +19. Shi, S., Jiang, L., Deng, J., Wang, Z., Guo, C., Shi, J., Wang, X., Li, H.: Pvrccnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. International Journal of Computer Vision 131(2), 531-551 (2023) +20. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019) +21. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence 43(8), 2647-2664 (2020) +22. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022) +23. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2446-2454 (2020) +24. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020) +25. Team, D.D.: 3dtrans: An open-source codebase for exploring transferable autonomous driving perception task. https://github.com/PJLab-ADG/3DTrans (2023) +26. VS, V., Oza, P., Patel, V.M.: Towards online domain adaptive object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 478-488 (2023) + +27. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726 (2020) +28. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7201-7211 (2022) +29. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in germany, test in the usa: Making 3d object detectors generalize. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11713-11723 (2020) +30. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. arXiv preprint arXiv:2203.14956 (2022) +31. Xu, Q., Zhou, Y., Wang, W., Qi, C.R., Anguelov, D.: Spg: Unsupervised domain adaptation for 3d object detection via semantic point generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15446-15456 (2021) +32. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018) +33. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10368-10378 (2021) +34. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d++: Denoised self-training for unsupervised domain adaptation on 3d object detection. IEEE transactions on pattern analysis and machine intelligence 45(5), 6354-6371 (2022) +35. Yang, Z., Sun, Y., Liu, S., Shen, X., Jia, J.: Std: Sparse-to-dense 3d object detector for point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1951-1960 (2019) +36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2021) +37. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Ad-pt: Autonomous driving pre-training with large-scale point cloud dataset. arXiv preprint arXiv:2306.00612 (2023) +38. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Bi3d: Bidomain active learning for cross-domain 3d object detection. arXiv preprint arXiv:2303.05886 (2023) +39. Yue, X., Zhang, Y., Zhao, S., Sangiovanni-Vincentelli, A., Keutzer, K., Gong, B.: Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2100-2110 (2019) +40. Yue, X., Zheng, Z., Zhang, S., Gao, Y., Darrell, T., Keutzer, K., Vincentelli, A.S.: Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13834-13844 (2021) +41. Zhang, B., Chen, T., Wang, B., Li, R.: Joint distribution alignment via adversarial learning for domain adaptive object detection. IEEE Transactions on Multimedia (2021) +42. Zhang, B., Yuan, J., Shi, B., Chen, T., Li, Y., Qiao, Y.: Uni3d: A unified baseline for multi-dataset 3d object detection. arXiv preprint arXiv:2303.06880 (2023) + +43. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022) +44. Zhao, S., Li, B., Yue, X., Gu, Y., Xu, P., Hu, R., Chai, H., Keutzer, K.: Multi-source domain adaptation for semantic segmentation. Advances in neural information processing systems 32 (2019) +45. Zheng, Y., Huang, D., Liu, S., Wang, Y.: Cross-domain object detection through coarse-to-fine feature adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13766-13775 (2020) +46. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018) \ No newline at end of file diff --git a/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/images.zip b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..088803c68d6102c2418ab617a79a60c82ec2d93f --- /dev/null +++ b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c1e5301f6192a174aad907bcec2ca30679998fbc880c7c5f01604b2f973edbc +size 596047 diff --git a/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/layout.json b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7150b765fa19413cd5f23bfa0ecac7a78e3b46fa --- /dev/null +++ b/2024/Better Regression Makes Better Test-time Adaptive 3D Object Detection/layout.json @@ -0,0 +1,10387 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 147 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 147 + ], + "type": "text", + "content": "Reg-TTA3D: Better Regression Makes Better Test-time Adaptive 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "text", + "content": "Jiakang Yuan" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "text", + "content": ", Bo Zhang" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "text", + "content": ", Kaixiong Gong" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "text", + "content": ", Xiangyu Yue" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "text", + "content": ", Botian Shi" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "text", + "content": ", Yu Qiao" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "text", + "content": ", and Tao Chen" + }, + { + "bbox": [ + 138, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{1,\\text{忍}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 201, + 478, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 201, + 478, + 224 + ], + "spans": [ + { + "bbox": [ + 135, + 201, + 478, + 224 + ], + "type": "text", + "content": "1 School of Information Science and Technology, Fudan University, Shanghai, China jkyuan22@m.fudan.edu.cn, eetchen@fudan.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 171, + 224, + 441, + 247 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 178, + 224, + 434, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 224, + 434, + 236 + ], + "spans": [ + { + "bbox": [ + 178, + 224, + 434, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 178, + 224, + 434, + 236 + ], + "type": "text", + "content": " Shanghai Artificial Intelligence Laboratory, Shanghai, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 236, + 441, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 236, + 441, + 247 + ], + "spans": [ + { + "bbox": [ + 171, + 236, + 441, + 247 + ], + "type": "text", + "content": "3 The Chinese University of Hong Kong, Central Ave, Hong Kong" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 280, + 455, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 280, + 455, + 521 + ], + "spans": [ + { + "bbox": [ + 159, + 280, + 455, + 521 + ], + "type": "text", + "content": "Abstract. Domain Adaptation (DA) has been widely explored and made significant progress on cross-domain 3D tasks recently. Despite being effective, existing works fail to deal with rapidly changing domains due to the unpredictable test time scenarios and meanwhile fast response time requirement. Thus, we explore a new task named test-time domain adaptive 3D object detection and propose Reg-TTA3D, a pseudo-label-based test-time adaptative 3D object detection method. By investigating the factor that limits the detection accuracy, we find that regression is essential in this task. To make better regression, we first design a noise-consistency pseudo-label generation process to filter pseudo-labels with instability under noise interference and obtain reliable pseudo-labels. Then, confidence-guided regression refinement is introduced, which uses the box regression results of high-confidence boxes to supervise boxes with relatively low confidence, further making the predicted box size gradually approach the distribution of the target domain. Finally, to better update the regression layer and alleviate the class-imbalance issue, a class-balance EMA updating strategy is proposed. Experimental results on multiple cross-domain scenarios including cross-beam, cross-location, and cross-weather demonstrate that Reg-TTA3D can achieve comparable or even better performance compared to unsupervised domain adaptation works by only updating less than " + }, + { + "bbox": [ + 159, + 280, + 455, + 521 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 159, + 280, + 455, + 521 + ], + "type": "text", + "content": " parameters within less than " + }, + { + "bbox": [ + 159, + 280, + 455, + 521 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 159, + 280, + 455, + 521 + ], + "type": "text", + "content": " time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 533, + 443, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 533, + 443, + 544 + ], + "spans": [ + { + "bbox": [ + 160, + 533, + 443, + 544 + ], + "type": "text", + "content": "Keywords: Test-time adaptation " + }, + { + "bbox": [ + 160, + 533, + 443, + 544 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 533, + 443, + 544 + ], + "type": "text", + "content": " 3D object detection " + }, + { + "bbox": [ + 160, + 533, + 443, + 544 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 533, + 443, + 544 + ], + "type": "text", + "content": " Regression" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 566, + 230, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 230, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 230, + 578 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "LiDAR-based 3D object detection [12, 18, 20, 21, 32, 36, 43] aims to predict the location and category of each object based on LiDAR point clouds and has attracted increasing attention recently due to its practical applications to autonomous driving. Although significant progress has been made, the stability of detectors under environmental changes, namely cross-domain 3D object detection, remains a challenge. However, the stability is important in practical usage" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 292, + 194 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 292, + 194 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 292, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 292, + 194 + ], + "type": "image", + "image_path": "27ce13672240b86d1095e3e95a07ea32faf712d1906c86435f7246bd609cfb4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "text", + "content": "Fig.1: Comparison between (a) previous domain adaptive 3D object detection and (b) Reg-TTA3D. Previous DA works perform training and testing in a split way. In contrast, Reg-TTA3D tackles domain discrepancies in an online manner and can deal with domain changes rapidly." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 293, + 114, + 481, + 194 + ], + "blocks": [ + { + "bbox": [ + 293, + 114, + 481, + 194 + ], + "lines": [ + { + "bbox": [ + 293, + 114, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 293, + 114, + 481, + 194 + ], + "type": "image", + "image_path": "0efd9b555507b097d540c0eb58dbfabe98fb0ae5369fe0ac70e76194b148cc3a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 479, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 479, + 291 + ], + "type": "text", + "content": "since the surrounding environment dynamically changes. For example, weather and geographical environment changes may be encountered in road scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 424 + ], + "type": "text", + "content": "Inspired by unsupervised domain adaption (UDA) techniques in 2D cross-domain tasks [3,5,8,13,39-41,44,45], some works [6,15,30,31,33,38] begin to deal with 3D domain discrepancies. ST3D [33] is a pioneer in using self-training to gradually transfer the knowledge learned from the source domain to the target domain. ReDB [6] attempts to deal with multi-class problems by designing cross-domain examination and overlapped boxes counting. Despite achieving large performance gains, the existing UDA model still requires a long training time to be adapted to the target domain, lacking the ability to rapidly deal with environmental changes in an online manner as shown in Fig. 1. Actually, in real applications, models are expected to be continuously updated at a low cost and quickly adapt to changing domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "text", + "content": "To make the model quickly adapt to the changing environment in an online manner, test-time adaptation (TTA) [1,14,16,27,28] is introduced to adapt the model during test-time with only source model and unlabeled target data. Tab. 1 shows differences between TTA and other adaptation methods. In fact, the TTA technique has been explored in 2D classification tasks such as TENT [27] and SHOT [14] recently, but it is still under-explored in 3D vision fields, especially in detection tasks. To verify the effectiveness of previous TTA methods in the domain adaptive 3D object detection tasks, we conduct experiments by directly applying such methods (e.g., TENT [27], SHOT [14], MATE [16]) to our baseline detector (i.e., SECOND-IoU [32]). As shown in Tab. 3, directly applying previous methods results in a large performance drop, even significantly behind the results of source only which directly evaluates with the pre-trained source model. For instance, TENT [27] only achieves " + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "inline_equation", + "content": "31.34\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "inline_equation", + "content": "19.47\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{BEV}} / \\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "inline_equation", + "content": "15.30\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "inline_equation", + "content": "9.36\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 593 + ], + "type": "text", + "content": " declination comparing to source only results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Accordingly, previous TTA methods are hard to apply to domain adaptive 3D detection tasks due to two major reasons. Firstly, from the perspective of updating parameter selection, previous methods often only update batch normalization (BN) layers to alleviate the domain shifts in the statistical information of network. However, this is not well suited for 3D detection task that also needs rich localization information. Besides, the optimization objectives of previous" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 130, + 425, + 196 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 126 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 126 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 126 + ], + "type": "text", + "content": "Table 1: Comparison of different settings to alleviate domain discrepancies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 130, + 425, + 196 + ], + "lines": [ + { + "bbox": [ + 186, + 130, + 425, + 196 + ], + "spans": [ + { + "bbox": [ + 186, + 130, + 425, + 196 + ], + "type": "table", + "html": "
SettingSource data Target data Online
Fine-tuningXxt,ytX
Unsupervised domain adaptationxs,s,ysxtX
Source-free domain adaptationXxtX
Test-time adaptationXxt✓
", + "image_path": "516bf718f7e4fad5fa1e7c175f0778f0e14a38a174559d3bfa02a3bf2f50de67.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 479, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 479, + 237 + ], + "type": "text", + "content": "methods are designed for classification tasks, which would make the model optimize for extracting features with more category semantic information rather than regression information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 238, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 480, + 416 + ], + "type": "text", + "content": "In order to tackle the challenges mentioned above, in this work, we propose Reg-TTA3D to perform test-time adaptation on 3D object detection tasks. To deal with the problem of updated parameter selection, we conduct extensive experiments and observe that high classification accuracy is already achieved by the source model, and the main factor limiting the performance improvement is the inaccurate regression results. Further, by only adapting regression layers, the model can achieve high performance at a low training cost as shown in Tab. 2 which motivates us to focus on regression adaptation in this work. Therefore, to make better regression, we propose noise-consistency pseudo-label generation (NPG) and confidence-guided regression refinement (CRR). The former module aims to filter boxes with instability under noise interference, i.e., noisy point clouds, while the latter one makes the box regression results align better with the distribution of the target domain box. Finally, we introduce a class-balanced EMA updating strategy (CBU) that can effectively update the regression layers, to improve the performance on multi-class simultaneously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 417, + 359, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 417, + 359, + 428 + ], + "spans": [ + { + "bbox": [ + 146, + 417, + 359, + 428 + ], + "type": "text", + "content": "Our contribution can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 429, + 481, + 561 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 429, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 429, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 136, + 429, + 481, + 464 + ], + "type": "text", + "content": "1. We present a new task, namely test-time domain adaptive 3D object detection, and propose Reg-TTA3D to solve this task from the novel view of focusing on adapting only regression parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 465, + 481, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 465, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 136, + 465, + 481, + 513 + ], + "type": "text", + "content": "2. We propose a noise-consistency pseudo-label generation process, a confidence-guided regression refinement loss, and a class-balanced EMA updating strategy, to obtain better detection results and boost the model's performance in an efficient and effective way." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 513, + 481, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 513, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 135, + 513, + 481, + 561 + ], + "type": "text", + "content": "3. Experiments show that Reg-TTA3D reaches a comparable or even better performance by updating less than " + }, + { + "bbox": [ + 135, + 513, + 481, + 561 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 135, + 513, + 481, + 561 + ], + "type": "text", + "content": " parameters within less than " + }, + { + "bbox": [ + 135, + 513, + 481, + 561 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 135, + 513, + 481, + 561 + ], + "type": "text", + "content": " time compared to UDA works, and significantly outperforms previous TTA methods on multiple cross-domain scenarios." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 577, + 243, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 243, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 243, + 590 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 599, + 338, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 338, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 338, + 612 + ], + "type": "text", + "content": "2.1 LiDAR-based 3D Object Detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "With the development of autonomous driving, LiDAR-based 3D object detection has attracted increasing attention. Current prevailing LiDAR-based 3D detectors can be roughly divided into point-based, voxel-based, and point-voxel-based methods according to the point cloud processing procedure. Point-based" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reg-TTA3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "methods [20, 35, 43] first sample a subset of point clouds, then extract features and generate proposals from sampled points. Among them, PointRCNN [20] uses a two-stage framework that is composed of a bottom-up proposal generation process and a proposal refinement process. In contrast, voxel-based methods [7, 12, 32, 36, 46] first divide disordered points into regular grids which are then fed into the convolutional network. SECOND [32] is a pioneer in using the 3D sparse backbone to extract features. CenterPoint [36] introduces a center-based detector and achieves promising performance. Point-voxel-based [18, 19] methods integrate the features extracted from points and voxels to obtain better feature representations. However, existing 3D detectors fail to deal with domain gaps which is essential in practical usage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 270, + 269, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 269, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 269, + 282 + ], + "type": "text", + "content": "2.2 Test-time Adaptation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 294, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 390 + ], + "type": "text", + "content": "Test-time adaptation [1, 4, 10, 22, 24, 27, 28] aims to rapidly adapt the source model to target domain without accessing source domain data during test time, which is more practical in the real world since it can handle dynamic domain shifts. TENT [27] explores test-time adaptation for the first time and updates the batch normalization layers using designed entropy minimization loss. SHOT [14] deals with such a challenging task by maximizing the mutual information. More recently, MATE [16] extends test-time adaptation to 3D point clouds and finds that the model can be adapted effectively by masking and reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 391, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 486 + ], + "type": "text", + "content": "Although TTA has been extensively explored in 2D scenarios, it is still under-explored in 3D tasks, especially in large-scale point cloud scenarios (e.g., autonomous driving scenarios). Besides, most previous TTA works mainly focus on classification tasks, such as classification [1, 16, 27] and segmentation [22], and they fail to tackle detection tasks since detection considers classification and box regression at the same time. In contrast, Reg-TTA3D considers TTA for 3D object detection tasks from the perspective of regression, which is more suitable for 3D detection tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 509, + 387, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 387, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 387, + 521 + ], + "type": "text", + "content": "2.3 Domain Adaptation for 3D Object Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "3D detectors often suffer large performance drops when training and testing under different domains, which makes them unstable when the surrounding environment dynamically changes. Recently, researchers have tried to tackle such a problem using domain adaptation techniques [6,17,29,33,34,38]. SN [29] proposes a statistical method to align the size of bounding boxes between source and target domains. ST3D [33] and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{ST3D + + }" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " [34] design a self-training pipeline and mitigate domain gaps using a memory bank and curriculum data augmentation. To deal with the class imbalance issue, ReDB [6] proposes a pseudo-label generation pipeline to obtain reliable, diverse, and class-balanced pseudo-labels. However, the existing works cannot deal with the emergency changes in the environment due to the time-consuming and large amount of calculation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 136, + 433, + 205 + ], + "blocks": [ + { + "bbox": [ + 190, + 114, + 421, + 126 + ], + "lines": [ + { + "bbox": [ + 190, + 114, + 421, + 126 + ], + "spans": [ + { + "bbox": [ + 190, + 114, + 421, + 126 + ], + "type": "text", + "content": "Table 2: Performance of updating different parameters." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 136, + 433, + 205 + ], + "lines": [ + { + "bbox": [ + 177, + 136, + 433, + 205 + ], + "spans": [ + { + "bbox": [ + 177, + 136, + 433, + 205 + ], + "type": "table", + "html": "
Tuning layersWaymo → KITTInuScenes → KITTIParameters
Batch normalization52.84 / 31.7731.86 / 18.259.08k
Classification layers52.82 / 36.3731.51 / 19.761.17k
Regression layers59.38 / 45.5433.63 / 21.802.73k
Detection head59.62 / 46.0832.51 / 20.394.68k
All parameters60.78 / 46.7133.23 / 19.6012.12M
", + "image_path": "a3d38071571ab017ada483d50284fe3b00b27f6fff31e034fb526f4ce9211441.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 211, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 211, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 211, + 202, + 224 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "content": "The overall framework of Reg-TTA3D is shown in Fig. 2. To better illustrate our proposed method, we start by introducing the problem formulation and our baseline in Sec. 3.1. Then, we give an analysis of the updated parameters selection in Sec. 3.2. Finally, we detail each module of Reg-TTA3D and give the overall objective and adaptation strategy in Sec. 3.3 and Sec. 3.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 312, + 221, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 221, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 221, + 324 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "text", + "content": "Problem Formulation. Following typical test-time adaptation setting [22, 27], given the model " + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "text", + "content": " pre-trained on the source domain " + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "inline_equation", + "content": "D_S = \\{(x_i^S, y_i^S)\\}_{i=1}^{n_S}" + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "text", + "content": " and an unlabeled target set " + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "inline_equation", + "content": "D_T = \\{x_j^T\\}_{j=1}^{n_T}" + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "text", + "content": ", where the source and target domain follow different distributions, our goal is to adapt the source model to the target domain without accessing source domain data in an online manner. During adaptation, in each timestep, a batch of unlabeled data " + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "inline_equation", + "content": "[x_1^t, x_2^t, \\dots, x_N^t]" + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "text", + "content": " is sampled from " + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "text", + "content": " to update the model, where " + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 331, + 482, + 415 + ], + "type": "text", + "content": " is batch size." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 475 + ], + "type": "text", + "content": "Baseline Introduction. Following [6, 34], we use SECOND-IoU [32] as our baseline detector which is composed of a feature extractor (i.e., 3D and 2D backbone) to obtain bird's eye view (BEV) features, a region proposal network (RPN) to generate proposals and a detection head to provide the classification and regression results. The detector is trained using the following loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 261, + 484, + 480, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 484, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 261, + 484, + 480, + 496 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e t} = \\mathcal {L} _ {r p n} + \\mathcal {L} _ {r c n n}, \\tag {1}", + "image_path": "842f1a3e47e6b5a23c7ba745d0cdd68bcfb21988729af7768bb813a36b245528.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 477, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 477, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 477, + 511 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 498, + 477, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rpn}" + }, + { + "bbox": [ + 130, + 498, + 477, + 511 + ], + "type": "text", + "content": " represents the loss of RPN and " + }, + { + "bbox": [ + 130, + 498, + 477, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rcnn}" + }, + { + "bbox": [ + 130, + 498, + 477, + 511 + ], + "type": "text", + "content": " denotes box refinement loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 526, + 314, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 314, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 314, + 539 + ], + "type": "text", + "content": "3.2 Updated Parameters Selection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "The aim of Reg-TTA3D is to transfer the source model to the target domain rapidly so that the model can tackle the continuous changes in the surrounding environment. To achieve this goal, the amount of updated parameters we select needs to be relatively small, and the accuracy needs to be largely improved by only updating selected parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Previous TTA works [22, 27] mainly focus on adapting batch normalization (BN) layers to mitigate the domain discrepancies, since BN layers contain the statistics of domains. Such an adaptation strategy is effective for domain adaptive classification tasks, as more accurate semantic representations can be obtained, which is especially important for classification. However, only adapting" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reg-TTA3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 232 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 232 + ], + "type": "image", + "image_path": "28479010401af4e409f3208040817cf94873654a0ae8bdf1a5e9b88ef27a9663.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 246, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 346 + ], + "type": "text", + "content": "Fig. 2: The overall framework of the proposed Reg-TTA3D, which employs SECOND-IoU [32] as the baseline detector and consists of noisy-consistency pseudo-label generation process (NPG), confidence-guided regression refinement (CRR) and class-balanced EMA updating strategy. The data are first fed into the teacher model to generate pseudo-labels as the supervision to the student model using NPG. Then, the student model is trained using data from another view and supervised by detection loss and refinement loss obtained by CRR. Finally, the class-balanced EMA updating strategy is used to update the teacher model. Note that the only difference between the teacher model and the student model lies in the regression layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "type": "text", + "content": "BN layers cannot rapidly transfer the source detector to the target domain since detection tasks contain both classification and regression. Some works [16, 26] try to solve this problem by updating the entire model but suffer from huge computational costs. To explore an effective and efficient method, by conducting extensive experiments, we find that the model can adapt to the target domain efficiently by only updating regression layers. As shown in Tab. 2, we fine-tune the parameters using " + }, + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "type": "text", + "content": " KITTI training data for 1 epoch and observe that only fine-tuning the regression layers can reach relatively high performance and even surpass the performance of fine-tuning the entire model. This is because in the autonomous driving scenario, there exists a large difference between categories such as car and pedestrian, the source model can already achieve high classification accuracy and the main factor for limiting the detection accuracy is the inaccurate regression results. Based on the observation, in this work, we focus on studying how to only adapt the parameters of regression layers and make better regression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 550, + 223, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 223, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 223, + 562 + ], + "type": "text", + "content": "3.3 Reg-TTA3D" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "content": "As mentioned above, the quality of box regression is essential to TTA for 3D object detection. To make better regression, we propose noise-consistency pseudolabel generation to obtain better regression supervision, and confidence-guided regression refinement to further refine the regression results. Besides, to effectively update the regression layers, we design a class-balanced EMA updating strategy to avoid overfitting the model to a single class." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "Noise-consistency Pseudo-label Generation. As mentioned in [34, 38], the pseudo-labels directly generated by the source model contain a lot of noise due" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "content": "to domain discrepancies. Previous methods try to alleviate pseudo-label noise by weighting pseudo-labels obtained in different epochs [33, 34], or adding the groundtruth instances from the source domain to target domain frames [6]. However, these methods cannot apply to test-time adaptation due to the inaccessibility of source data and online adaptation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": "Inspired by test-time augmentation methods which comprehensively consider the model predictions under different data augmentation methods to provide final results. Specifically, the model's prediction of real objects is relatively stable, while being uncertain for the noise regions. To this end, we propose noise-consistency pseudo-label generation to obtain more accurate pseudo-labels, by measuring the stability of model predictions when adding random noise to regions of interest (RoIs). In detail, given proposals generated by the teacher model " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "B^{t} = \\{(c_{x}^{t}, c_{y}^{t}, c_{z}^{t}, l^{t}, w^{t}, h^{t}, \\theta^{t})_{i}\\}_{i=1}^{n}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": ", its corresponding features " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\{f_{i}^{t}\\}_{i=1}^{n}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " and confidence scores " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\{s_{i}^{t}\\}_{i=1}^{n}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "(c_{x}^{t}, c_{y}^{t}, c_{z}^{t})" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "(l^{t}, w^{t}, h^{t})" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " denote the center and the size of the box respectively, " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\theta^{t}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " is the rotation angle and " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " is the number of boxes. We can obtain boxes with noise " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "B^{t'} = \\{(c_{x}^{t'}, c_{y}^{t'}, c_{z}^{t'}, l^{t'}, w^{t'}, h^{t'}, \\theta^{t'})\\}_{i=1}^{n}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": ", corresponding features " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\{f_{i}^{t'}\\}_{i=1}^{n}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " and scores " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\{s_{i}^{t'}\\}_{i=1}^{n}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": ", by adding random noise to boxes' size and angle. To measure the stability of RoIs, we calculate the inconsistency rate as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 348, + 481, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 348, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 235, + 348, + 481, + 376 + ], + "type": "interline_equation", + "content": "R _ {i n c o n s i s t e n c y, i} = 1 - \\frac {\\left| s _ {i} ^ {t} - s _ {i} ^ {t ^ {\\prime}} \\right|}{\\operatorname {I o U} \\left(b _ {i} ^ {t} , b _ {i} ^ {t ^ {\\prime}}\\right)}, \\tag {2}", + "image_path": "016b641d1effe181d185e46dcf468ec7ae26014bd6c1c97efed087a2efde0476.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}(b_i^t, b_i^{t'})" + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "content": " denotes the intersection of union (IoU) between " + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "inline_equation", + "content": "b_i^t \\in B^t" + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "inline_equation", + "content": "b_i^{t'} \\in B^{t'}" + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "content": ". A high inconsistency rate represents that the model prediction is unstable to the current object which is more likely to be a noise region. Therefore, we filter the boxes with high inconsistency rate by a threshold " + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "content": " to get boxes and scores " + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\hat{B}, \\{\\hat{s}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\hat{B}', \\{\\hat{s}_i'\\}_{i=1}^n" + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "content": " with stability under noise interference. Then, we ensemble " + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\hat{B}'" + }, + { + "bbox": [ + 131, + 378, + 482, + 450 + ], + "type": "text", + "content": " by confidence scores as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 223, + 455, + 481, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 455, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 223, + 455, + 481, + 486 + ], + "type": "interline_equation", + "content": "\\tilde {B} = \\left\\{\\tilde {b _ {i}} \\right\\}, \\quad \\tilde {b _ {i}} = \\left\\{ \\begin{array}{l l} \\hat {b} _ {i}, & \\hat {s} _ {i} > \\hat {s} _ {i} ^ {\\prime}, \\hat {b} _ {i} \\in \\hat {B} \\\\ \\hat {b} _ {i} ^ {\\prime}, & \\hat {s} _ {i} < \\hat {s} _ {i} ^ {\\prime}, \\hat {b} _ {i} ^ {\\prime} \\in \\hat {B} ^ {\\prime} \\end{array} , \\right. \\tag {3}", + "image_path": "03571c51df09ffb964abbff36462a3fd85c2098f8fba131a6ada72fdff0caf24.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "After obtaining " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{B}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", the boxes are filtered by non-maximum suppression (NMS) with a relatively high confidence threshold, to remove redundant boxes and boxes with low confidence. Finally, the pseudo-labels " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "B^{npg}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{l_i^{npg}\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " with high stability and confidence can be acquired where " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{l_i^{npg}\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " is classification results. Confidence-guided Regression Refinement. Although lots of boxes containing noisy point cloud can be removed by measuring the stability under noise injection, there still exist the problems of inaccurate box regression results. This is because the distributions of bounding box size are inconsistent between source and target domains as mentioned in [29,37,42], and it is difficult to mitigate such domain discrepancy with only NPG. By analyzing the distribution of predicted bounding boxes, we find that boxes with higher confidence scores align better with the distribution of the target domain boxes as shown in Fig. 3, since precise boxes correspond to high-quality RoI features that are less disturbed by background representations. Based on this observation, we design a confidence-guided regression refinement to obtain more precise regression results of box size." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reg-TTA3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 449, + 214 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 449, + 214 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 449, + 214 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 449, + 214 + ], + "type": "image", + "image_path": "64c255b7c3a24accc9e7467597ecaa4b71e4e0d5cc58abe748fbcd8370c03c77.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 221, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 131, + 221, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 221, + 482, + 243 + ], + "type": "text", + "content": "Fig. 3: Distribution of pseudo box size under different thresholds. PS denotes pseudo boxes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "text", + "content": "In particular, given predicted boxes generated by the student model " + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "inline_equation", + "content": "B^{s} = \\{(c_{x}^{s}, c_{y}^{s}, c_{z}^{s}, l^{s}, w^{s}, h^{s}, \\theta^{s})_{j}\\}_{j=1}^{k}" + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "text", + "content": " and scores " + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\{s_{j}^{s}\\}_{j=1}^{k}" + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "text", + "content": " of a batch of frames, where " + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "text", + "content": " is the total number of predicted boxes. We sort boxes belonging to the same class according to scores in descending order, and sample a proportion of boxes with high scores to calculate the mean value of sample boxes size " + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "inline_equation", + "content": "(\\bar{l}^{s}, \\bar{w}^{s}, \\bar{h}^{s})" + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "text", + "content": ". Please refer to our supplementary material for the study of the proportion of sampled boxes. Once obtained " + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "inline_equation", + "content": "(\\bar{l}^{s}, \\bar{w}^{s}, \\bar{h}^{s})" + }, + { + "bbox": [ + 131, + 254, + 482, + 350 + ], + "type": "text", + "content": ", we can use it as a supervision signal to make the box regression value closer to the distribution of the target domain as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 248, + 360, + 375, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 360, + 375, + 373 + ], + "spans": [ + { + "bbox": [ + 248, + 360, + 375, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e f i n e}} = \\mathbb {E} _ {i = 1} ^ {n} \\Sigma_ {d \\in [ d _ {l}, d _ {w}, d _ {h} ]} d,", + "image_path": "ca1e80adc8c95b6618cad0b49fb454fc71cf5da73d11ef7e48e67bf8ba2c2151.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 375, + 480, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 214, + 375, + 480, + 407 + ], + "type": "interline_equation", + "content": "w h e r e d = \\left\\{ \\begin{array}{l l} 0, & | \\left(z ^ {s} - \\bar {z} ^ {s}\\right) | < p \\\\ \\left(z ^ {s} - \\bar {z} ^ {s}\\right) ^ {2}, & | \\left(z ^ {s} - \\bar {z} ^ {s}\\right) | \\geq p \\end{array} , \\right. \\tag {4}", + "image_path": "77ec2d2bc69ccc16fad8a9e2fa735ab30d9d475d20f486dfc763bdc71161f48d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "inline_equation", + "content": "z \\in [l^s, w^s, h^s]" + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\overline{z} \\in [\\overline{l}^s, \\overline{w}^s, \\overline{h}^s]" + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "text", + "content": " is batch size and " + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "text", + "content": " is set 0.1 to allow the size of objects to fluctuate within a certain range. The model can predict more accurate box size when using " + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{refine}}" + }, + { + "bbox": [ + 131, + 413, + 482, + 449 + ], + "type": "text", + "content": " for optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": "Class-balanced EMA Updating. To make the model more stable, we update the teacher regression layers by exponential moving average (EMA). However, the model tends to overfit to the category with a large number of instances (i.e., car). As a result, the detection performance on categories with few instances will decrease. To make the model better balance the performance improvement for each class, we propose a class-balanced EMA updating strategy to make the model update more when the class is balanced, while updating less when a single category dominates. Specifically, given the pseudo-labels obtained from NPG " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "B^{npg} = \\{b_i^{npg} \\in \\mathbb{R}^7\\}" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": " and its corresponding class labels " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\{c_i^{npg}\\}" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": ". We calculate the proportion " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\{p_c\\}_{c=1}^C" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": " of each category in pseudo-labels according to " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\{c_i^{npg}\\}" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": " is the number of total categories and " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\Sigma_{c=1}^C p_c = 1" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": ". Then the variance of " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\{p_c\\}_{c=1}^C" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": " is calculated and normalized to [0.99, 0.999] to obtain the momentum " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": ". A high value of " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": " means that a severe class-imbalance issue exists in the current batch and the model should update less. Finally the parameter of the teacher regression layer " + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\omega_{reg}^t" + }, + { + "bbox": [ + 131, + 449, + 482, + 629 + ], + "type": "text", + "content": " can be updated by the following formula:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 639, + 480, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 639, + 480, + 653 + ], + "spans": [ + { + "bbox": [ + 242, + 639, + 480, + 653 + ], + "type": "interline_equation", + "content": "\\omega_ {r e g} ^ {t} \\leftarrow \\alpha \\omega_ {r e g} ^ {t} + (1 - \\alpha) \\omega_ {r e g} ^ {s}, \\tag {5}", + "image_path": "dda28d9ad93ed8b9e17ab395b65b86196c21b205407597b9e5f2bd634e6142e6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 656, + 383, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 656, + 383, + 669 + ], + "spans": [ + { + "bbox": [ + 131, + 656, + 383, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 656, + 383, + 669 + ], + "type": "inline_equation", + "content": "\\omega_{reg}^{s}" + }, + { + "bbox": [ + 131, + 656, + 383, + 669 + ], + "type": "text", + "content": " is the parameters of student regression layers." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 373, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 373, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 373, + 128 + ], + "type": "text", + "content": "Algorithm 1 Single iteration of Reg-TTA3D strategy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "spans": [ + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "type": "text", + "content": "Input: A batch of frames " + }, + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "type": "inline_equation", + "content": "\\{\\dot{x}_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "type": "inline_equation", + "content": "\\dot{x}_i" + }, + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "type": "text", + "content": " is obtained by performing data augmentation on " + }, + { + "bbox": [ + 136, + 129, + 379, + 150 + ], + "type": "inline_equation", + "content": "x_{i}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 151, + 355, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 151, + 355, + 163 + ], + "spans": [ + { + "bbox": [ + 136, + 151, + 355, + 163 + ], + "type": "text", + "content": "Output: Predictions of current batch " + }, + { + "bbox": [ + 136, + 151, + 355, + 163 + ], + "type": "inline_equation", + "content": "B^{t}" + }, + { + "bbox": [ + 136, + 151, + 355, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 151, + 355, + 163 + ], + "type": "inline_equation", + "content": "\\{l_i^t\\}_{i = 1}^n" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 163, + 481, + 274 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "text", + "content": "1: Get the predictions " + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "inline_equation", + "content": "B^{t} = \\{b_{i}^{t}\\in \\mathbb{R}^{7}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "inline_equation", + "content": "\\{l_i^t\\}_{i = 1}^n" + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "text", + "content": " using teacher model, " + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "inline_equation", + "content": "B^{t}" + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "inline_equation", + "content": "\\{l_i^t\\}_{i = 1}^n" + }, + { + "bbox": [ + 136, + 163, + 481, + 186 + ], + "type": "text", + "content": " are the detection results of current batch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 186, + 481, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 186, + 481, + 207 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 481, + 207 + ], + "type": "text", + "content": "2: Use NPG to obtain boxes " + }, + { + "bbox": [ + 136, + 186, + 481, + 207 + ], + "type": "inline_equation", + "content": "B^{npg}" + }, + { + "bbox": [ + 136, + 186, + 481, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 186, + 481, + 207 + ], + "type": "inline_equation", + "content": "\\{l_i^{npg}\\}_{i=1}^n" + }, + { + "bbox": [ + 136, + 186, + 481, + 207 + ], + "type": "text", + "content": " with stability under noise interference as pseudo-labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "text", + "content": "3: Get the predictions of student model " + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "inline_equation", + "content": "B^{s}" + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "inline_equation", + "content": "\\{s_{i}^{s}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "inline_equation", + "content": "\\{\\dot{x}_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "text", + "content": ", supervise the student model by " + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "inline_equation", + "content": "B^{npg}" + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "inline_equation", + "content": "\\{l_i^{npg}\\}_{i = 1}^n" + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{det}" + }, + { + "bbox": [ + 136, + 208, + 481, + 231 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 231, + 270, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 231, + 270, + 241 + ], + "spans": [ + { + "bbox": [ + 136, + 231, + 270, + 241 + ], + "type": "text", + "content": "4: Calculate " + }, + { + "bbox": [ + 136, + 231, + 270, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{refine}}" + }, + { + "bbox": [ + 136, + 231, + 270, + 241 + ], + "type": "text", + "content": " using CRR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 241, + 321, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 241, + 321, + 252 + ], + "spans": [ + { + "bbox": [ + 136, + 241, + 321, + 252 + ], + "type": "text", + "content": "5: Optimize the student model using Eq. (6)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 252, + 354, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 252, + 354, + 263 + ], + "spans": [ + { + "bbox": [ + 136, + 252, + 354, + 263 + ], + "type": "text", + "content": "6: Update the teacher model using CBU and Eq. (5)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 263, + 333, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 263, + 333, + 274 + ], + "spans": [ + { + "bbox": [ + 136, + 263, + 333, + 274 + ], + "type": "text", + "content": "7: Return: Current predictions " + }, + { + "bbox": [ + 136, + 263, + 333, + 274 + ], + "type": "inline_equation", + "content": "B^{t}" + }, + { + "bbox": [ + 136, + 263, + 333, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 263, + 333, + 274 + ], + "type": "inline_equation", + "content": "\\{l_i^t\\}_{i = 1}^n" + }, + { + "bbox": [ + 136, + 263, + 333, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 300, + 429, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 429, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 429, + 313 + ], + "type": "text", + "content": "3.4 Overall Objective and Test-time Adaptation Strategy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 320, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 481, + 344 + ], + "type": "text", + "content": "Overall Objective. The overall objective to optimize the student regression layers can be formulated as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 254, + 354, + 481, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 354, + 481, + 368 + ], + "spans": [ + { + "bbox": [ + 254, + 354, + 481, + 368 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {d e t}} + \\mathcal {L} _ {\\text {r e f i n e}}, \\tag {6}", + "image_path": "2da23a6f9adfc33582f8ce7ce241af705b20a1f94016d233a8b1bd640f0e74ba.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 374, + 386, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 386, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 386, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 374, + 386, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{det}" + }, + { + "bbox": [ + 130, + 374, + 386, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 374, + 386, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{refine}" + }, + { + "bbox": [ + 130, + 374, + 386, + 387 + ], + "type": "text", + "content": " are defined in Eq. (1) and Eq. (4)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 387, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 411 + ], + "type": "text", + "content": "Test-time Adaptation Strategy. The test-time adaptation strategy of Reg-TTA3D can be summarized in Algorithm 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 430, + 230, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 230, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 230, + 443 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 454, + 261, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 261, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 261, + 467 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "text", + "content": "Datasets and Cross-domain Settings. We conduct experiments on four commonly-used datasets: Waymo [23], KITTI [9], nuScenes [2] and nuScenes-C [11]. To verify the effectiveness of Reg-TTA3D, we consider three cross-domain settings including cross-location (i.e., Waymo " + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "text", + "content": " KITTI), cross-beam (i.e., Waymo " + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "text", + "content": " nuScenes, nuScenes " + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "text", + "content": " KITTI) and cross-weather (i.e., nuScenes " + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "text", + "content": " nuScenes-C). Note that in the cross-weather setting, we use the medium-difficulty fog and snow scenarios in the nuScenes-C dataset. Following ReDB [6], in the first two cross-domain settings, we use the KITTI evaluation metric to evaluate for three common classes in autonomous driving scenarios (i.e., car, pedestrian, and cyclist), we report the average precision in both 3D (i.e., " + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "text", + "content": ") and BEV (i.e., " + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{BEV}}" + }, + { + "bbox": [ + 130, + 474, + 486, + 629 + ], + "type": "text", + "content": ") over 40 recall positions, with IoU threshold 0.7 for car and 0.5 for pedestrian and cyclist. In the cross-weather setting, we use the official nuScenes evaluation metric and report mean average precision (mAP)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "Implementation Details. Following [6], we evaluate the proposed Reg-TTA3D on SECOND-IoU [32], and the source model is trained for all categories using random object scaling (ROS) augmentation method. We set hyperparameter " + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reg-TTA3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 164 + ], + "type": "text", + "content": "to 1.5 and " + }, + { + "bbox": [ + 130, + 116, + 481, + 164 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 481, + 164 + ], + "type": "text", + "content": " to 0.1 for all cross-domain settings and use Adam optimizer with learning rate of " + }, + { + "bbox": [ + 130, + 116, + 481, + 164 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 116, + 481, + 164 + ], + "type": "text", + "content": ". We utilize commonly-used data augmentation methods (e.g., random world flip, random world rotation, random world scaling) to get " + }, + { + "bbox": [ + 130, + 116, + 481, + 164 + ], + "type": "inline_equation", + "content": "\\{\\dot{x}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 116, + 481, + 164 + ], + "type": "text", + "content": ". Our code is built on 3DTrans [25]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 296 + ], + "type": "text", + "content": "Comparison Baselines. To verify the effectiveness of Reg-TTA3D, we mainly compare our methods to unsupervised domain adaptive 3D object detection and previous test-time adaptation methods that we apply to 3D object detection. UDA methods: (1) ST3D [33] is a self-training method that continuously updates the pseudo-labels, (2) " + }, + { + "bbox": [ + 130, + 164, + 481, + 296 + ], + "type": "inline_equation", + "content": "\\mathrm{ST3D + + }" + }, + { + "bbox": [ + 130, + 164, + 481, + 296 + ], + "type": "text", + "content": " [34] provides more insight analysis and extends ST3D to get better results, (3) ReDB [6] is a pioneer to perform domain adaptive 3D detection on multiple categories. TTA methods: (1) TENT [27] proposes an entropy minimization method which focuses on adapting BN layers, (2) SHOT [14] optimizes feature extraction module using both information maximization and pseudo-labeling. (3) MATE [16] uses masked autoencoder to apply test-time adaptation to 3D classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 312, + 228, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 228, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 228, + 323 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": "Results on cross-location and cross-beam scenarios. We first evaluate Reg-TTA3D on cross-location and cross-beam scenarios and compare the results with previous UDA methods. As shown in Tab. 3, Reg-TTA3D can greatly improve the detection accuracy on Waymo " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " KITTI and nuScenes " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " KITTI settings, largely narrow the performance gap between source only and oracle. For example, in the Waymo " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " KITTI setting, our Reg-TTA3D can achieve " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "61.78\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "49.96\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{BEV}} / \\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": ", even surpass UDA methods such as ST3D [33] and ST3D++ [34] and it is comparable to the state-of-the-art UDA method (i.e., " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "61.14\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "50.10\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{BEV}} / \\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " reported in ReDB [6]). In nuScenes " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " KITTI and Waymo " + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": " nuScenes cross-domain scenarios, the performance of Reg-TTA3D is also better than some UDA methods (e.g., ST3D [33]) which needs source domain data and training for multiple epochs. In contrast, Reg-TTA3D only needs target data and one epoch of adaptation which greatly reduces the time consumption." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "Further, we compare Reg-TTA3D with the previous TTA methods in Tab. 3. We find that previous TTA methods suffer performance drops when applied to 3D detection tasks (e.g., TENT [27] reduce the detection accuracy to " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "31.34\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "19.47\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " in the Waymo " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " KITTI setting). This is mainly because the detection task is composed of classification and regression and existing methods only consider the classification tasks. As a result, the model will concentrate on extracting effective representations for the classification tasks and ignore the localization information. However, our Reg-TTA3D fully considers the challenges in detection tasks and can improve the performance on multiple categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Results on the cross-weather scenario. To more comprehensively evaluate the effectiveness of our proposed method, we conduct experiments in a more realistic scenario (i.e., cross-weather). Since nuScenes-C dataset only has the validation set, we only compare Reg-TTA3D with existing TTA methods. It can be seen in Tab. 4 that Reg-TTA3D can improve the performance on both snow" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 182, + 479, + 482 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "text", + "content": "Table 3: Results on cross-location and cross-beam scenarios. Following [6, 33], we use SECOND-IoU [32] as out baseline detector. Source Only represents that the pretrained detector is directly evaluated on the target domain, and Oracle denotes the detection results obtained using the fully annotated target training data. Red: Best UDA method, Blue: Best TTA method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 182, + 479, + 482 + ], + "lines": [ + { + "bbox": [ + 135, + 182, + 479, + 482 + ], + "spans": [ + { + "bbox": [ + 135, + 182, + 479, + 482 + ], + "type": "table", + "html": "
TASKMETHODSETTINGCARPEDESTRIANCYCLISTMEAN AP
Waymo → KITTISOURCE ONLY-51.48 / 19.7840.80 / 31.2647.63 / 35.4546.64 / 28.83
SN76.61 / 54.1452.48 / 48.2034.56 / 32.7454.55 / 45.03
ST3DUDA77.62 / 49.2444.45 / 42.0447.74 / 42.9556.60 / 44.70
ST3D++77.68 / 50.0349.09 / 46.1951.50 / 47.7059.42 / 47.97
ReDB80.37 / 54.1251.01 / 48.2052.05 / 47.9761.14 / 50.10
TENTTTA71.98 / 39.014.54 / 3.9517.51 / 15.4531.34 / 19.47
SHOT79.52 / 41.3133.52 / 29.0240.79 / 35.8351.27 / 35.38
MATE75.17 / 30.6439.94 / 36.9243.72 / 40.2252.94 / 35.92
REG-TTA3D81.60 / 56.0350.89 / 45.4352.85 / 48.4361.78 / 49.96
ORACLE-83.29 / 73.4546.64 / 41.3362.92 / 60.3264.28 / 58.37
Waymo → nuScenesSOURCE ONLY-30.64 / 17.181.81 / 0.580.97 / 0.8811.14 / 6.22
SN29.56 / 17.781.33 / 1.072.92 / 2.6111.27 / 7.15
ST3DUDA28.42 / 17.831.64 / 1.394.01 / 3.5411.36 / 7.59
ST3D++28.87 / 19.151.82 / 1.584.09 / 3.7411.60 / 8.16
ReDB30.12 / 18.562.47 / 2.146.56 / 5.1913.05 / 8.63
TENTTTA16.40 / 3.430.82 / 0.250.28 / 0.215.83 / 1.29
SHOT12.26 / 1.421.56 / 1.360.22 / 0.204.68 / 0.99
MATE8.64 / 3.991.72 / 0.970.24 / 0.083.53 / 1.68
REG-TTA3D30.54 / 18.152.04 / 1.755.01 / 4.7012.53 / 8.20
ORACLE-51.88 / 34.8725.24 / 18.9215.06 / 11.7330.73 / 21.84
nuScenes → KITTISOURCE ONLY-39.15 / 7.6521.54 / 16.876.31 / 2.4422.33 / 8.98
SN56.08 / 28.6723.05 / 16.845.11 / 2.3228.08 / 15.94
ST3DUDA71.50 / 48.0922.64 / 17.617.86 / 5.2034.00 / 23.64
ST3D++69.90 / 44.6224.11 / 18.2010.14 / 6.3933.75 / 23.07
ReDB74.23 / 51.3125.95 / 18.3813.82 / 8.6438.00 / 26.11
TENTTTA43.35 / 19.839.40 / 5.772.71 / 2.7018.49 / 9.43
SHOT60.24 / 23.0117.51 / 9.442.47 / 1.6626.74 / 11.37
MATE61.89 / 27.1315.98 / 13.112.94 / 2.4526.93 / 14.23
REG-TTA3D68.73 / 44.5622.78 / 17.7110.21 / 7.1233.90 / 23.13
ORACLE-83.29 / 73.4546.64 / 41.3362.92 / 60.3264.28 / 58.37
", + "image_path": "312fd5513c337188287bdf4b09a26cff50959b10f35f873a37a7f37fc4ac398e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "text", + "content": "and fog scenarios (e.g., " + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "inline_equation", + "content": "45.68\\% \\rightarrow 47.09\\%" + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "text", + "content": " mAP on the Normal " + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "text", + "content": " Snow setting). The results verify that our proposed method can handle dynamic environment changes in road scenes and can be applied in more practical scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 548, + 244, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 244, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 244, + 560 + ], + "type": "text", + "content": "4.3 Insight Analyses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": "Sensitivity to Detector Architecture. Following [6], to verify the effectiveness of Reg-TTA3D on different types of detectors, we further conduct experiments on nuScenes " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " KITTI cross-domain scenario using PointRCNN [20] as the baseline detector. As show in Tab. 5, our Reg-TTA3D outperforms all TTA methods and some UDA methods. Note that compared to the source-free domain adaptation method (i.e., SF-UDA" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "^{3D}" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " [17]) which is inaccessible to the source data and need to adapt for multiple epochs, our Reg-TTA3D adapted on multi-class can achieve a better result on the car category (e.g., " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "50.87\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "49.80\\%" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reg-TTA3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 152, + 478, + 258 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Results on cross-weather scenario (i.e., nuScenes Normal " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " Fog and Normal " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " Snow). Fog and snow data are from nuScenes-C [11] dataset. We use CenterPoint [36] as our baseline detector since it is the most widely used detector on nuScenes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 152, + 478, + 258 + ], + "lines": [ + { + "bbox": [ + 135, + 152, + 478, + 258 + ], + "spans": [ + { + "bbox": [ + 135, + 152, + 478, + 258 + ], + "type": "table", + "html": "
METHODTaskmAPCARTRUCKCV.BUSTRAILERBARRIERMOTOR.BICYCLEPED.TC.
SOURCE ONLY48.4079.1252.588.1067.5331.0064.407.3026.6581.2865.97
TENT41.8373.6215.237.5858.0421.8560.451.5623.8477.5952.24
SHOTFog42.3172.6340.707.5958.1822.2461.710.8526.4577.8454.87
MATE23.9764.6124.183.3427.2614.1120.450.007.6749.6728.33
REG-TTA3D49.6780.3054.4310.8668.6534.5363.527.7927.0782.2067.28
SOURCE ONLY45.8680.7448.774.6864.1528.8462.629.0224.7278.3756.62
TENTSnow40.5175.0939.376.5157.5719.9659.821.8923.3376.7844.73
SHOT42.1574.6639.747.0158.1221.4662.941.1927.3078.3950.66
MATE23.4363.9822.573.2025.3013.0622.790.007.640.5024.81
REG-TTA3D47.0982.0149.717.1968.1630.5462.966.3326.2980.3757.30
", + "image_path": "7083bd166e04f7b904cfe18fb4f0e134ee8e57fbe56d3bdf8f80cae621748033.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 308, + 478, + 423 + ], + "blocks": [ + { + "bbox": [ + 132, + 264, + 481, + 300 + ], + "lines": [ + { + "bbox": [ + 132, + 264, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 481, + 300 + ], + "type": "text", + "content": "Table 5: Performance comparisons (AP3D) with PointRCNN as baseline detector on nuScenes " + }, + { + "bbox": [ + 132, + 264, + 481, + 300 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 132, + 264, + 481, + 300 + ], + "type": "text", + "content": " KITTI task. " + }, + { + "bbox": [ + 132, + 264, + 481, + 300 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 132, + 264, + 481, + 300 + ], + "type": "text", + "content": " indicates the results reported in the original paper. Red: Best UDA method, Blue: Best TTA method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 308, + 478, + 423 + ], + "lines": [ + { + "bbox": [ + 135, + 308, + 478, + 423 + ], + "spans": [ + { + "bbox": [ + 135, + 308, + 478, + 423 + ], + "type": "table", + "html": "
MethodSettingCARPEDESTRIANCYCLISTAVERAGE
EASYMOD.HARDEASYMOD.HARDEASYMOD.HARDEASYMOD.HARD
SOURCE ONLY-42.7732.1128.7543.2637.2933.163.284.094.1829.7724.6022.03
SN66.5650.3245.9242.9637.1532.459.077.577.4239.5331.6828.60
ST3DUDA48.8541.9038.9245.6738.7133.0926.5019.3518.3840.3433.3230.13
ST3D++60.4549.3645.8850.7742.4336.6427.2017.9416.5246.1436.5833.01
REDB71.4557.953.9152.3244.3337.9545.1332.9331.0556.3045.0540.97
SF-UDA3D‡SFDA68.8049.8045.00---------
TENTTTA42.5532.4127.6130.7126.0923.784.243.663.3225.8320.7018.23
SHOT30.0923.0920.4119.2416.2914.312.431.701.5117.2513.6912.07
REG-TTA3D65.2050.8746.7245.8042.5534.9040.2527.8725.2950.4140.4335.63
", + "image_path": "14237526b6157033b1f02026eca35937362ee18b9e045515e4f9ce1a69abd8c7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 435, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 472 + ], + "type": "text", + "content": "reported by SF-UDA" + }, + { + "bbox": [ + 130, + 435, + 481, + 472 + ], + "type": "inline_equation", + "content": "^{3D}" + }, + { + "bbox": [ + 130, + 435, + 481, + 472 + ], + "type": "text", + "content": " in moderate level) in one epoch of adaptation even if SF-UDA" + }, + { + "bbox": [ + 130, + 435, + 481, + 472 + ], + "type": "inline_equation", + "content": "^{3D}" + }, + { + "bbox": [ + 130, + 435, + 481, + 472 + ], + "type": "text", + "content": " is trained on the unfair single class. The results further show that our proposed methods can apply to different detector architectures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "text", + "content": "Ablation Studies. To further verify the effectiveness of Reg-TTA3D, we conduct extensive experiments to investigate the impact of each component we proposed including NPG, CRR and CBU. As shown in Tab. 6, the improvement (i.e., " + }, + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "inline_equation", + "content": "0.92\\% \\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "text", + "content": ") of using naive pseudo-labels is relatively small compared to the performance of source only since pseudo-labels contain lots of noise samples and the regression results are not accurate due to domain discrepancies. Reg-TTA can effectively handle such a problem by proposing NPG and CRR modules, with " + }, + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "inline_equation", + "content": "2.90\\%" + }, + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "inline_equation", + "content": "3.60\\%" + }, + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "text", + "content": " improvement respectively compared to source only. Besides, when combining these two modules, the performance can be further improved because NPG focuses on filter noise samples and CRR concentrates on making more accurate regression results. Further, by combining with CBU which makes the model update in a class-balance manner, the performance can be further boosted and achieve " + }, + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "inline_equation", + "content": "49.96\\% \\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 472, + 481, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "text", + "content": "Sensitivity to Hyperparameters. We ablate the impact of hyperparameters " + }, + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "text", + "content": " (i.e., the threshold of inconsistency rate) and " + }, + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "text", + "content": " in this subsection. As shown in Tab. 7, we first set " + }, + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "text", + "content": " to 1.0, 1.5, 2.0 respectively and observe that the fluctuation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 198, + 330, + 308 + ], + "blocks": [ + { + "bbox": [ + 131, + 116, + 351, + 193 + ], + "lines": [ + { + "bbox": [ + 131, + 116, + 351, + 193 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 351, + 193 + ], + "type": "text", + "content": "Table 6: Ablation studies on each component. PS, NPG, CRR and CBU represent using naive pseudo-labels, noise-consistency pseudo-label generation, confidence-guided regression refinement and class-balanced EMA updating, respectively. The experiments are conducted on Waymo " + }, + { + "bbox": [ + 131, + 116, + 351, + 193 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 131, + 116, + 351, + 193 + ], + "type": "text", + "content": " KITTI scenario using SECOND-IoU as the baseline detector." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 198, + 330, + 308 + ], + "lines": [ + { + "bbox": [ + 151, + 198, + 330, + 308 + ], + "spans": [ + { + "bbox": [ + 151, + 198, + 330, + 308 + ], + "type": "table", + "html": "
PS NPGCRRCBUmean APBEVmean AP3D
---57.8942.01
✓--58.7342.93
✓✓-58.8944.91
✓-✓59.8445.61
✓--59.7344.05
✓✓✓60.5747.97
✓✓-60.1246.62
✓-✓60.5948.39
✓✓✓61.7849.96
", + "image_path": "a0a0ab4ff1ddb621ab8788b162acd6f4dff89c20e643d7db92d925722b5c8dcb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 378, + 205, + 460, + 308 + ], + "blocks": [ + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "lines": [ + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "text", + "content": "Table 7: Ablation studies of hyperparameters " + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "text", + "content": ". The experiments are conducted on Waymo " + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "text", + "content": " KITTI scenario using SECOND-IoU. Note that when ablate on " + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "p = 0.1" + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "text", + "content": " and when ablate on " + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\tau = 0.15" + }, + { + "bbox": [ + 356, + 114, + 482, + 201 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 378, + 205, + 460, + 308 + ], + "lines": [ + { + "bbox": [ + 378, + 205, + 460, + 308 + ], + "spans": [ + { + "bbox": [ + 378, + 205, + 460, + 308 + ], + "type": "table", + "html": "
τAPBEV / AP3D
1.060.92 / 49.12
1.561.78 / 49.96
2.060.44 / 48.88
pAPBEV / AP3D
0.0560.29 / 48.26
0.1061.78 / 49.96
0.1561.56 / 49.63
", + "image_path": "92a2f7351331055a74914c034f5f01733bd036a2ce8e0992cd5d302ec4cc9b7f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 322, + 304, + 416 + ], + "blocks": [ + { + "bbox": [ + 133, + 322, + 304, + 416 + ], + "lines": [ + { + "bbox": [ + 133, + 322, + 304, + 416 + ], + "spans": [ + { + "bbox": [ + 133, + 322, + 304, + 416 + ], + "type": "image", + "image_path": "8d3ba54ca5c5066e56034224777598f6050e0ca924f67c7e7ed8926fd83e99d6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 424, + 304, + 447 + ], + "lines": [ + { + "bbox": [ + 132, + 424, + 304, + 447 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 304, + 447 + ], + "type": "text", + "content": "Fig. 4: The relationship between detection accuracy and the iteration number." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 380, + 476, + 445 + ], + "blocks": [ + { + "bbox": [ + 306, + 335, + 478, + 378 + ], + "lines": [ + { + "bbox": [ + 306, + 335, + 478, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 478, + 378 + ], + "type": "text", + "content": "Table 8: Ablation studies on costs. The experiments are conducted on " + }, + { + "bbox": [ + 306, + 335, + 478, + 378 + ], + "type": "inline_equation", + "content": "\\mathrm{W} \\rightarrow \\mathrm{K}" + }, + { + "bbox": [ + 306, + 335, + 478, + 378 + ], + "type": "text", + "content": " scenario using SECOND-IoU. We use Tesla A100 to conduct the experiments." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 380, + 476, + 445 + ], + "lines": [ + { + "bbox": [ + 308, + 380, + 476, + 445 + ], + "spans": [ + { + "bbox": [ + 308, + 380, + 476, + 445 + ], + "type": "table", + "html": "
MethodTime Cost N_GPUsParam.Results
ST3D4h412.12M 56.60 / 44.70
ST3D++3.5h412.12M 59.42 / 47.97
ReDB3.2h412.12M 61.14 / 50.10
TENT4min19.08K 31.34 / 19.47
MATE6min10.71M 52.94 / 35.92
Reg-TTA3D4.5min12.73K 61.78 / 49.96
", + "image_path": "94ce73de8fc48177ceaf95ac06c30abe40bf625c5135ce68dd85348608d48611.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 458, + 481, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 481, + 507 + ], + "type": "text", + "content": "of the performance is " + }, + { + "bbox": [ + 130, + 458, + 481, + 507 + ], + "type": "inline_equation", + "content": "1.34\\% / 1.08\\%" + }, + { + "bbox": [ + 130, + 458, + 481, + 507 + ], + "type": "text", + "content": ". Then, by setting " + }, + { + "bbox": [ + 130, + 458, + 481, + 507 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 458, + 481, + 507 + ], + "type": "text", + "content": " to different values (i.e., 0.05, 0.1, 0.15), we find that there is only little fluctuation and the best performance can be achieved when set to 0.1. The results further show that our method is robust to these hyperparameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "type": "text", + "content": "Analysis on Pseudo-labeling Accuracy. In this section, we measure the accuracy of each iteration during adaptation. Fig. 4 reflects the detection accuracy on Waymo " + }, + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "type": "text", + "content": " KITTI as iteration increases. It can be seen that the accuracy continuously improves since the model is gradually adapted to the target domain which also verifies the stability of Reg-TTA3D during adaptation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": "Analysis on Costs. Further, we report the time cost and total updated parameters in Tab. 8. Notably, Reg-TTA3D achieves comparable results to UDA methods adapting only less than " + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": " of parameters within less than " + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": " of time. When compared to TTA methods, our proposed method can achieve much higher performance with less updated parameters and comparable time cost which further verifies that our Reg-TTA3D is an effective and efficient method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Ablation Studies of the momentum. We ablate the momentum in EMA updating and compare it to our class-balance EMA updating strategy. Specifi-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 101 + ], + "type": "text", + "content": "Reg-TTA3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 294, + 230 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 294, + 230 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 294, + 230 + ], + "type": "image", + "image_path": "0b0854b088a69e6979bc96ee42082480df6f82a4057f122ec79e13a8aee0450c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 236, + 478, + 259 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 478, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 478, + 259 + ], + "type": "text", + "content": "Fig. 5: The impact of momentum and com- Fig.6: ablation studies on the proportions with CBU. tion of selected samples." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 113, + 468, + 230 + ], + "blocks": [ + { + "bbox": [ + 319, + 113, + 468, + 230 + ], + "lines": [ + { + "bbox": [ + 319, + 113, + 468, + 230 + ], + "spans": [ + { + "bbox": [ + 319, + 113, + 468, + 230 + ], + "type": "image", + "image_path": "5d8d709697e33b8ee75c10460202c65a1f9776a3559567967521f95982b8025b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 319 + ], + "type": "text", + "content": "cally, we conduct the experiment for momentum between 0.99 and 0.999 at 0.001 intervals. As shown in Fig. 5, the result demonstrates that CBU can better update the regression layer due to it can balance update the model and improve the performance on multiple classes at the same time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "text", + "content": "Ablation Studies on the proportion of Selected samples in CRR. In our manuscript, we use " + }, + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "text", + "content": " boxes with the highest scores to supervise box sizes with relatively low scores. Here, we analyze the impact on the proportion of selected samples. As shown in Fig. 6, when the proportion is relatively high, the average box size will be distributed far from the target domain, and when set to relatively small values, the small number of samples results in the averages being affected by individual instances. Based on the experiments and analysis, we set the proportion to " + }, + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 438, + 219, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 219, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 219, + 451 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 468, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 564 + ], + "type": "text", + "content": "In this work, for the first time, we explore test-time adaptation in 3D object detection tasks and propose Reg-TTA3D, which handles rapid changes in the surrounding environment from the perspective of updating parameters selection and adaptation algorithm. By adapting using the proposed noise-consistency pseudo-label generation process, confidence-guided regression refinement loss and class-balanced EMA updating strategy and only updating regression layers, Reg-TTA3D can achieve comparable performance to some UDA methods at a low cost." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 587, + 217, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 217, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 217, + 600 + ], + "type": "text", + "content": "6 Limitation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Although the proposed Reg-TTA3D can handle dynamic changes in the surrounding environments such as continuously changing weather. It cannot deal with emergencies such as some corner cases which is also important in real-world autonomous driving." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 495, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 495, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 495, + 239 + ], + "type": "text", + "content": "This work is supported by National Key Research and Development Program of China (No. 2022ZD0160101), National Natural Science Foundation of China (No. 62071127, and 62101137), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No.2021SHZDZX0103). This work is also supported by the National Key R&D Program of China (Grant No. 2022ZD0160104), and Shanghai Rising Star Program (Grant No. 23QD1401000). The computations in this research were performed using the CFFF platform of Fudan University." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 259, + 197, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 197, + 272 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 197, + 272 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 286, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 286, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 481, + 320 + ], + "type": "text", + "content": "1. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8344-8353 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 321, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 364 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "type": "text", + "content": "3. Cai, Q., Pan, Y., Ngo, C.W., Tian, X., Duan, L., Yao, T.: Exploring object relation in mean teacher for cross-domain detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11457-11466 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 399, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 430 + ], + "type": "text", + "content": "4. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 295-305 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 432, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 481, + 464 + ], + "type": "text", + "content": "5. Chen, Y., Li, W., Sakaridis, C., Dai, D., Van Gool, L.: Domain adaptive faster r-cnn for object detection in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3339-3348 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "type": "text", + "content": "6. Chen, Z., Luo, Y., Huang, Z., Wang, Z., Baktashmotlagh, M.: Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudolabeling. arXiv preprint arXiv:2307.07944 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 498, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 532 + ], + "type": "text", + "content": "7. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1201-1209 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "text", + "content": "8. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: International conference on machine learning. pp. 1180-1189. PMLR (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 588, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 481, + 620 + ], + "type": "text", + "content": "10. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems 35, 6204-6218 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "text", + "content": "11. Kong, L., Liu, Y., Li, X., Chen, R., Zhang, W., Ren, J., Pan, L., Chen, K., Liu, Z.: Robo3d: Towards robust and reliable 3d perception against corruptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 19994-20006 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reg-TTA3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "12. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "13. Liang, J., Gong, K., Li, S., Liu, C.H., Li, H., Liu, D., Wang, G., et al.: Pareto domain adaptation. Advances in Neural Information Processing Systems 34, 12917-12929 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "14. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "text", + "content": "15. Luo, Z., Cai, Z., Zhou, C., Zhang, G., Zhao, H., Yi, S., Lu, S., Li, H., Zhang, S., Liu, Z.: Unsupervised domain adaptive 3d detection with multi-level consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8866-8875 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "text", + "content": "16. Mirza, M.J., Shin, I., Lin, W., Schriebl, A., Sun, K., Choe, J., Kozinski, M., Possegger, H., Kweon, I.S., Yoon, K.J., et al.: Mate: Masked autoencoders are online 3d test-time learners. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16709-16718 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "17. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771-780. IEEE (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "18. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529–10538 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "text", + "content": "19. Shi, S., Jiang, L., Deng, J., Wang, Z., Guo, C., Shi, J., Wang, X., Li, H.: Pvrccnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. International Journal of Computer Vision 131(2), 531-551 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "20. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "21. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence 43(8), 2647-2664 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "text", + "content": "22. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "text", + "content": "23. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "24. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "25. Team, D.D.: 3dtrans: An open-source codebase for exploring transferable autonomous driving perception task. https://github.com/PJLab-ADG/3DTrans (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "26. VS, V., Oza, P., Patel, V.M.: Towards online domain adaptive object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 478-488 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "27. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "28. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7201-7211 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 217 + ], + "type": "text", + "content": "29. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in germany, test in the usa: Making 3d object detectors generalize. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11713-11723 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "text", + "content": "30. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. arXiv preprint arXiv:2203.14956 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 296 + ], + "type": "text", + "content": "31. Xu, Q., Zhou, Y., Wang, W., Qi, C.R., Anguelov, D.: Spg: Unsupervised domain adaptation for 3d object detection via semantic point generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15446-15456 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 318 + ], + "type": "text", + "content": "32. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "text", + "content": "33. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10368-10378 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "text", + "content": "34. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d++: Denoised self-training for unsupervised domain adaptation on 3d object detection. IEEE transactions on pattern analysis and machine intelligence 45(5), 6354-6371 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "35. Yang, Z., Sun, Y., Liu, S., Shen, X., Jia, J.: Std: Sparse-to-dense 3d object detector for point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1951-1960 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "text", + "content": "36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "text", + "content": "37. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Ad-pt: Autonomous driving pre-training with large-scale point cloud dataset. arXiv preprint arXiv:2306.00612 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "type": "text", + "content": "38. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Bi3d: Bidomain active learning for cross-domain 3d object detection. arXiv preprint arXiv:2303.05886 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 521, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 564 + ], + "type": "text", + "content": "39. Yue, X., Zhang, Y., Zhao, S., Sangiovanni-Vincentelli, A., Keutzer, K., Gong, B.: Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2100-2110 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 609 + ], + "type": "text", + "content": "40. Yue, X., Zheng, Z., Zhang, S., Gao, Y., Darrell, T., Keutzer, K., Vincentelli, A.S.: Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13834-13844 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 609, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 643 + ], + "type": "text", + "content": "41. Zhang, B., Chen, T., Wang, B., Li, R.: Joint distribution alignment via adversarial learning for domain adaptive object detection. IEEE Transactions on Multimedia (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "42. Zhang, B., Yuan, J., Shi, B., Chen, T., Li, Y., Qiao, Y.: Uni3d: A unified baseline for multi-dataset 3d object detection. arXiv preprint arXiv:2303.06880 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 101 + ], + "type": "text", + "content": "Reg-TTA3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 259 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "43. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "44. Zhao, S., Li, B., Yue, X., Gu, Y., Xu, P., Hu, R., Chai, H., Keutzer, K.: Multi-source domain adaptation for semantic segmentation. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "45. Zheng, Y., Huang, D., Liu, S., Wang, Y.: Cross-domain object detection through coarse-to-fine feature adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13766-13775 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "46. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Beyond MOT_ Semantic Multi-Object Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_content_list.json b/2024/Beyond MOT_ Semantic Multi-Object Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f887dce01ab11823d3fbaed5385540950af1b026 --- /dev/null +++ b/2024/Beyond MOT_ Semantic Multi-Object Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_content_list.json @@ -0,0 +1,1697 @@ +[ + { + "type": "text", + "text": "Beyond MOT: Semantic Multi-Object Tracking", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 165 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunhao Li $^{1,2}$ , Qin Li $^{1}$ , Hao Wang $^{2}$ , Xue Ma $^{1}$ , Jiali Yao $^{3}$ , Shaohua Dong $^{4}$ , Heng Fan $^{4\\dagger}$ , and Libo Zhang $^{1,2,3\\dagger\\sharp}$", + "bbox": [ + 235, + 188, + 767, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Software Chinese Academy of Sciences", + "bbox": [ + 323, + 231, + 676, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Chinese Academy of Sciences", + "bbox": [ + 349, + 246, + 651, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences", + "bbox": [ + 215, + 258, + 785, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Department of Computer Science & Engineering, University of North Texas", + "bbox": [ + 241, + 273, + 759, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current multi-object tracking (MOT) aims to predict trajectories of targets (i.e., \"where\") in videos. Yet, knowing merely \"where\" is insufficient in many crucial applications. In comparison, semantic understanding such as fine-grained behaviors, interactions, and overall summarized captions (i.e., \"what\") from videos, associated with \"where\", is highly-desired for comprehensive video analysis. Thus motivated, we introduce Semantic Multi-Object Tracking (SMOT), that aims to estimate object trajectories and meanwhile understand semantic details of associated trajectories including instance captions, instance interactions, and overall video captions, integrating \"where\" and \"what\" for tracking. In order to foster the exploration of SMOT, we propose BenSMOT, a largescale Benchmark for Semantic MOT. Specifically, BenSMOT comprises 3,292 videos with 151K frames, covering various scenarios for semantic tracking of humans. BenSMOT provides annotations for the trajectories of targets, along with associated instance captions in natural language, instance interactions, and overall caption for each video sequence. To our best knowledge, BenSMOT is the first publicly available benchmark for SMOT. Besides, to encourage future research, we present a novel tracker named SMOTer, which is specially designed and end-to-end trained for SMOT, showing promising performance. By releasing BenSMOT, we expect to go beyond conventional MOT by predicting \"where\" and \"what\" for SMOT, opening up a new direction in tracking for video understanding. We will release BenSMOT and SMOTer at https://github.com/Nathan-Li123/SMOTer.", + "bbox": [ + 259, + 323, + 743, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Semantic Multi-Object Tracking (SMOT) $\\cdot$ Benchmark", + "bbox": [ + 261, + 669, + 720, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 709, + 375, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-object tracking (MOT) is a fundamental problem in computer vision with many applications such as autonomous driving, robotics, and visual surveillance. Current MOT tasks (e.g., [9,18,33,38,48]) usually focus on predicting trajectories of the targets from videos, i.e., answering the question of \"where are the targets\". Despite considerable advancements in deep learning era (e.g., [17,32,44,55,60]),", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Equal advising and co-last author $\\sharp$ Corresponding author (libo@iscas.ac.cn)", + "bbox": [ + 230, + 824, + 750, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01923049d8a6a11cba42968437f6b15463ca28a601ae1427c37695d86bcaa1d0.jpg", + "image_caption": [ + "Fig. 1: Existing multi-object tracking (MOT) focusing on predicting trajectories only (see (a)) and our semantic multi-object tracking (SMOT) aiming at estimating trajectories and understanding their semantics (see (b)). Best viewed in color for all figures." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 772, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "knowing merely \"where\" in existing MOT is insufficient for video understanding in many important applications. For example, besides trajectories, knowing more trajectory-associated semantic details of the objects, such as the behaviors and their interactions with the surroundings (i.e., \"what\"), is greatly beneficial for visual surveillance and robotics, which grows the interest in expanding modern MOT (i.e., addressing \"where\") with trajectory-associated semantic analysis (i.e., understanding \"what\") for more comprehensive video understanding.", + "bbox": [ + 212, + 500, + 787, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus motivated, in this work we introduce a new type of MOT task, dubbed Semantic Multi-Object Tracking (SMOT) $^5$ , which aims to expand scope of MOT beyond merely predicting trajectories to capture rich semantic details of objects in videos. Specifically, besides trajectory estimation, SMOT comprises three additional trajectory-associated semantic understanding tasks, including instance captioning, instance interaction recognition, and video captioning, as illustrated in Fig. 1. In particular, instance captioning aims at describing objects and their behaviors in human language, answering \"what are the objects doing\"; instance interaction recognition is to capture relations between objects, answering \"what are the relations between objects\"; video captioning provides the overall scenario understanding based on trajectories, answering \"what is happening\". Note that, all these three additional tasks are associated with object trajectories in videos. Ultimately, the goal of SMOT is to integrate \"where\" (i.e., object trajectories)", + "bbox": [ + 212, + 604, + 787, + 801 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Here by \"semantic\", we emphasize high-level trajectory-based activity understanding in videos in the context of tracking, instead of category as in semantic segmentation.", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and \"what\" (i.e., trajectory-associated semantics) for video understanding, going beyond MOT predicting only trajectories (see again Fig. 1). Compared to conventional MOT, our SMOT is a multimodal task across vision and language, which is more challenging yet practical. It is worthy to notice that, SMOT is a natural extension of MOT. Despite sharing some similarities with video captioning [23], SMOT differs in that it aims at language understanding of dense target trajectories from videos in the context of instance-level tracking.", + "bbox": [ + 218, + 146, + 785, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To foster study of SMOT, we propose BenSMOT, a large-scale Benchmark for Semantic MOT. Specifically, BenSMOT contains 3,292 video sequences with 151K frames, captured from more than 40 diverse daily-life scenarios for human-centric semantic tracking. In BenSMOT, there exist more than 7.8K instances, labeled with over 335K bounding boxes for trajectory prediction. For each instance, we provide fine-grained human language to describe the behaviors for instance captioning, resulting in 7.8K sentences. In addition, to capture interactions between objects, BenSMOT provides more than 14K interaction annotations from a rich collection of 335 kinds of interactions. Moreover, for each sequence, a summary text is provided for overall understanding of instances and backgrounds, leading to 3,292 video caption texts in total in BenSMOT. To ensure the high quality of BenSMOT, annotations are manually labeled with careful inspection and refinement. To our best knowledge, BenSMOT is the first publicly available benchmark dedicated to SMOT. By releasing BenSMOT, we expect it to serve as a platform for advancing the research on SMOT.", + "bbox": [ + 218, + 257, + 785, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, in order to facilitate the development of SMOT algorithms on BenSMOT, we present a simple yet effective tracker dubbed SMOTer. Specifically, SMOTer is built upon a multi-object tracking model [53]. After generating object trajectories, we extract features for each trajectory and design additional prediction heads for semantic understanding of instance captioning, interaction recognition, and video captioning. Note, SMOTer is not a simple offline assembly of models for different tasks. Instead, it is specially designed for SMOT and trained in an end-to-end manner for predicting trajectories and understanding their semantics in videos. Despite simplicity, SMOTer shows promising performance for SMOT, and outperforms the offline-combination strategies, evidencing its effectiveness. We expect it to provide a reference for future research.", + "bbox": [ + 218, + 488, + 785, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We notice there exists a concurrent work combining multi-object tracking and trajectory captioning for dense video object captioning [58]. Compared with [58], our work differs in three aspects. First, task-wise, besides trajectory estimation and captioning, the SMOT in this work contains instance interaction and overall video captioning, providing more semantic details. Second, dataset-wise, we introduce a large-scale benchmark, BenSMOT, that is dedicated for SMOT and supports complete end-to-end model training, while the work of [58] borrows datasets from different tasks, which results in disjoint training of the model on different tasks and may thus degrade performance. Third, model-wise, owing to BenSMOT, we propose an end-to-end algorithm that demonstrates promising results on SMOT, while the approach in [58] is disjointly learned and does not support predicting instance interactions and overall video captioning.", + "bbox": [ + 218, + 657, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Beyond MOT: Semantic Multi-Object Tracking", + "bbox": [ + 415, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we make the following contributions: $\\spadesuit$ We introduce semantic multi-object tracking (SMOT), a new tracking paradigm that expands existing MOT task by integrating \"where\" and \"what\"; $\\spadesuit$ We present BenSMOT, a large-scale dataset with 3,292 videos and rich annotations for SMOT; $\\spadesuit$ We propose SMOTer, a simple but effective tracker to facilitate future research of SMOT; $\\spadesuit$ We show that SMOTer achieves promising performance for SMOT and conduct in-depth analysis on it to provide guidance for future algorithm design.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 277, + 397, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MOT Benchmarks. Benchmarks have played a crucial role in facilitating the development of MOT. PETS2009 [16] is one of the earliest benchmarks for multipedestrian tracking. Later, the popular MOT Challenge [10, 24, 33] has been introduced with more crowded videos and greatly advanced MOT. KITTI [18] and BDD100K [48] are specifically designed for object tracking in autonomous driving. ImageNet-Vid [11] provides trajectory annotations for 30 categories across over 1,000 videos, and TAO [9] further expands object classes to 833 for generic multi-object tracking. To foster MOT in specific dancing and sport scenarios, DanceTrack [38] and SportsMOT [8] have been presented for dancer and player tracking. AnimalTrack [52] focuses on tracking various animals in wild scenes. Moreover, UAVDT [12] and VisDrone [61] provide platforms for tracking targets with drones. Different from the above MOT benchmarks for object trajectory prediction only (\"where\"), BenSMOT is specially developed for the new task of SMOT to serve as a platform for simultaneous estimation of target trajectories and trajectory-associated semantic understanding. To this end, it offers rich annotations of not only object trajectories as in existing standard MOT datasets (\"where\") but also semantic details of objects (\"what\") such as instance trajectory captions, interactions and overall video captions (see Fig. 1 again).", + "bbox": [ + 212, + 310, + 787, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MOT Algorithms. MOT algorithms have seen great progress in recent years. One of the popular paradigms for multi-object tracking is the so-called Tracking-by-Detection. This paradigm involves initial object detection followed by association based on these detections, forming the basis for many representative methods [5,13,21,31,43,53]. In this case, MOT methods typically improve their performance by enhancing both detection and matching effectiveness. In addition, another prevalent paradigm is the Joint-Tracking-and-Detection [20,44,45,47,59], which integrates tracking and detection steps into a single process, enabling end-to-end training. Recently, Transformer [41] has been introduced into MOT and exhibited remarkable improvements over previous tackers [7,17,32,39,50,55,60]. Our SMOTer is related to but different than existing MOT methods. Specifically, going beyond merely predicting object trajectories, SMOTer also aims at semantic understanding of trajectories, integrating “where” and “what”.", + "bbox": [ + 212, + 583, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Captioning. Video captioning is a multimodal vision-language task that aims to automatically describe the video content using natural language. Owing to its important applications in video event commentary and human-computer interaction, video captioning has drawn increasing attention in the past decade", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with many models proposed (e.g., [23, 27, 36, 37, 46, 57]). Similar to the video captioning task, SMOT utilizes natural language to describe instances and the overall video content. However, the difference is that our SMOT aims at language comprehension for dense target trajectories in the context of multi-object tracking, which is more challenging due to the requirement of accurate trajectories yet provides more fine-grained information for video understanding.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual Relationship Detection. The goal of visual relationship detection task is to identify and comprehend the relationships (usually represented by a triplet of $<$ subject, predicate, object $\\rangle$ ) between objects from an image (e.g., [25, 29, 51]). Besides static images, recent works extend visual relationship detection in the video domain (e.g., [6, 28, 56]). Our work is relevant to video visual relationship detection but different by extending this task target-specific trajectories, and thus more challenging due to complicated multi-object tracking scenarios.", + "bbox": [ + 212, + 237, + 784, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The Proposed BenSMOT", + "text_level": 1, + "bbox": [ + 214, + 367, + 496, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Design Principle", + "text_level": 1, + "bbox": [ + 214, + 400, + 398, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BenSMOT aims at providing a new platform for exploring human-centric SMOT. In construction of BenSMOT, we follow the following rules:", + "bbox": [ + 212, + 426, + 782, + 455 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Dedicated benchmark. The key motivation of our BenSMOT is to provide a dedicated dataset for semantic multi-object tracking. In current deep learning era, a large number of videos are desired in benchmark construction for training robust tracking models. Considering this, we hope to build a dedicated dataset with more than 3,000 videos for human-centric semantic multi-object tracking.", + "(2) Diverse scenarios. For a dataset, diversity is crucial, in both training and evaluation, for developing general systems. In order to provide a diverse platform for SMOT, in BenSMOT we will include videos from more than 40 various scenes that range from daily life scenarios to dancing scenarios to sport scenarios.", + "(3) High-quality annotations. High-quality annotation is crucial for a benchmark in both training and assessing models. In BenSMOT, we ensure its high quality by providing manual annotations for each sequence including object trajectories, instance captions, instance interactions and overall video captions. In this process, multi-round inspections and refinements will be carried out." + ], + "bbox": [ + 212, + 455, + 784, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Data Acquisition", + "text_level": 1, + "bbox": [ + 214, + 691, + 401, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BenSMOT focuses on predicting the target trajectories and meanwhile understanding their semantics by instance captioning, interaction recognition, and overall video captioning. To this end, the video sequences in BenSMOT are desired to involve diverse multi-person activities and interactions, which is different from videos in existing benchmarks to some extent. Specifically, we first identify 47 common scenarios, involving different activities ranging from daily-life talk and play to different sports, that are suitable for tracking by drawing inspiration from [22]. Due to limited space, please refer to supplementary material for", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Beyond MOT: Semantic Multi-Object Tracking", + "bbox": [ + 413, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/551841633fcc6e6c14a841ea8331e9ad5ed2921602e5d438750db0a3705710fd.jpg", + "table_caption": [ + "Table 1: Summary of BenSMOT and its comparison with popular multi-object tracking benchmarks. \"n/a\" denotes that data is not available." + ], + "table_footnote": [], + "table_body": "
KITTI [18]MOT17 [33]MOT20 [10]BDD100k [48]TAO [9]GMOT-40 [1]DanceTrack [38]SportsMOT [8]BenSMOT (ours)
Videos501481,6002,907401002403,292
Min. length (s)n/a17.017.040.0n/a3.0n/an/a1.5
Avg. length (s)10.033.066.840.036.88.952.9n/a22.9
Max. length (s)n/a85.0133.040.0n/a24.2n/an/a116.0
Total length (s)498463535640,000106,9783565,292601575,499
Total tracks2,6001.3K3.83K131K17,2872,0269903,4017,792
Total boxes80K300K2,102K3,300K333K256Kn/a1,629K335K
Total frames15K11K13K318K2,674K9K106K150K151K
Instance CaptionsXXXXXXXX7,792
Instance InteractionsXXXXXXXX14K
Video SummariesXXXXXXXX3,292
", + "bbox": [ + 246, + 184, + 751, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "detailed scenarios in BenSMOT. Afterwards, we search for raw sequences for each scenario under the Creative Commons License from YouTube, the largest and most popular video platform with massive real-world videos. Please notice, for each raw video, instead of using the whole sequence, we usually choose a suitable clip for our semantic multi-object tracking task.", + "bbox": [ + 212, + 344, + 782, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Eventually, we compile a dataset for SMOT by including 3,292 videos with 151K frames. BenSMOT has an average video length of 23 seconds, with our current focus in this work not being on long-term SMOT. The total number of instance tracks is 7.8K, annotated with 335K bounding boxes for tracking. Besides bounding boxes, we provide 7.8K instance captions, 14K interactions, and 3,292 video summaries for trajectory-associated semantic understanding (how to annotate will be described in the following). To our best knowledge, BenSMOT by far is the first publicly available benchmark dedicated for SMOT. Tab. 1 summarizes our BenSMOT and compares it with popular MOT benchmarks.", + "bbox": [ + 212, + 421, + 784, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Data Annotation", + "text_level": 1, + "bbox": [ + 215, + 584, + 401, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to meet the requirements of SMOT, BenSMOT provides four types of annotations, including bounding box, instance caption, instance interaction, and the overall video caption. Specifically, for each trajectory in videos, similar to standard MOT benchmarks (e.g., [9,33]), we provide axis-aligned bounding boxes to indicate its spatial positions in the videos. For instance captioning, we label each trajectory with a precise sentence in natural language to describe the detailed behavior of the associated target. For interaction recognition, we first collect a set of 335 interactions (i.e., verbs), among them 327 from WordNet [15]. Note that, in our interaction set, the same verb may have different meanings and thus is assigned to different interaction types, e.g., hold(v.01) for “being the physical support of someone” and hold(v.02) for “someone in one’s hands or arm”, for precise relationship description. Due to limited space, we refer reader to supplementary material for the full set of interaction types. After this, we provide an directed interaction label, represented with a triplet of $<$ subject predicate, object $\\rightharpoondown$ (see Fig. 1 (b)), between two trajectories if associated objects", + "bbox": [ + 212, + 613, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/41fc0948e3f3b4719a1566e7dc91e1fb6ce3c9eec3c194f410c27907877d7474.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 143, + 406, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9959dbb8db30f6cbdaa45fcfd9201b5c977df662d35dc8d58a9036754777b6b1.jpg", + "image_caption": [ + "(d) Inst caption length distribution" + ], + "image_footnote": [], + "bbox": [ + 230, + 222, + 401, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f523d359a7ee4a387f43faac26e3cea9942a124cca07edc04411446f59c5627.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 143, + 593, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a1ca43adafa5b6fa864e9b4406e4218bf8a2152ef3966bf1f8ec84ec0fd1da3.jpg", + "image_caption": [ + "(e) Video caption length distribution" + ], + "image_footnote": [], + "bbox": [ + 419, + 222, + 581, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc21fb07e299f637ad7cda3e98e90058e4ed5c83144beb57f120cea72d724c61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 143, + 779, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ad3a9d96dbaa1096c26e248d75b301f9542ee1fb0e52bfdd158877638e333ec.jpg", + "image_caption": [ + "(f) WordCloud of caption words", + "Fig. 2: Representative statistics on BenSMOT, including distributions of target trajectory length (in seconds) in (a), sequence length (in seconds) in (b), number of interactions per sequence in (c), instance caption length in (d), video caption length in (e), and wordcloud of all caption words with prepositions excluded in (f)." + ], + "image_footnote": [], + "bbox": [ + 602, + 222, + 772, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are interacting. Finally, for the overall video captioning, we utilize a concise text summary to describe object trajectories in the video from a global perspective.", + "bbox": [ + 212, + 415, + 782, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure the high-quality annotations, we adopt a multi-round mechanism. Specifically, each video will be first manually annotated by a few volunteers who are familiar with the task and an expert working on related problems. After this initial round, all annotations of the video will be sent to a validation team former by more than two experts for inspection. If initial annotations are not unanimously agreed by the experts, they will be returned to the original labeling team for refinement using the feedback from the validation team. We repeat this procedure for each video until the annotations of all video sequences in BenSMOT are completed. An annotation example in our BenSMOT is shown in Fig. 1 (b), and more can be found in supplementary material.", + "bbox": [ + 212, + 445, + 784, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Statistics of annotation. In order to better understand BenSMOT, we show some representative statics of annotations in Fig. 2. Specifically, we demonstrate the distributions of object trajectory length, video length, number of instance interactions, instance caption length, video caption length, and wordcloud. From Fig. 2 (c), it is worth noting that different targets in videos have frequent interactions. In addition, for the instance captions in Fig. 2 (d), we provide relatively longer textual descriptions, which allows precise understanding.", + "bbox": [ + 212, + 595, + 784, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Dataset Split and Evaluation Metric", + "text_level": 1, + "bbox": [ + 214, + 724, + 560, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training/Test Split. BenSMOT contains 3,292 videos, captured from 47 different scenarios. Within each scenario, we use $70\\%$ of the sequences for training, and the rest $30\\%$ for test. During dataset split, we try to keep the distributions of training and test sets similar. Eventfully, the training set of BenSMOT comprises 2,284 sequences with 104K frames, and the test set consists of 1,008 videos with 47K frames. Please see supplementary material for more split details.", + "bbox": [ + 212, + 748, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Beyond MOT: Semantic Multi-Object Tracking", + "bbox": [ + 413, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e171706b4b13802126f70e12061a427159fe8b04983952bd097dce288d2a556.jpg", + "image_caption": [ + "Fig. 3: Illustration of SMOTer, which contains three components of trajectory estimation for tracking, feature fusion, and trajectory-associated semantic understanding." + ], + "image_footnote": [], + "bbox": [ + 246, + 145, + 754, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metric. We apply multiple metrics on BenSMOT for evaluation. Specifically, to assess the performance in object trajectory estimation, we employ higher order tracking accuracy (HOTA), association accuracy (AssA), detection accuracy (DetA), and localization accuracy (LocA) by following [30], CLEAR metrics [3] including multiple object tracking accuracy (MOTA), false positives (FP), false negatives (FN), and ID switches (IDs), and ID metrics [35] containing identification precision (IDP), identification recall (IDR) and related F1 score (IDF1). For instance and video captioning, we follow existing methods (e.g., [27, 37]) and employ BLEU [34], ROUGE [26], METEOR [2], and CIDEr [42] for thorough evaluation. Lastly, to assess performance of interaction recognition, we leverage classic metrics such as Precision (Prcn), Recall (Rcll), and F1, which are widely used in the context of classification tasks.", + "bbox": [ + 212, + 426, + 787, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Methodology: A New Baseline for SMOT", + "text_level": 1, + "bbox": [ + 214, + 628, + 653, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overview. To encourage development of SMOT algorithms, we propose a simple yet effective tracker on BenSMOT, dubbed SMOTer. As illustrated in Fig. 3, SMOTer logically consists of three key components. The first component, as detailed in Sec 4.1, is dedicated to object trajectory estimation for tracking, which is the basis for SMOT. Then, the second component, as explained in Sec 4.2, focuses on aggregating features from each frame into the overall video feature and meanwhile merging features of individual objects into target trajectory features using two distinct fusion modules, for subsequent semantic understanding. Finally, the third component, as described in Sec 4.3, aims to predict trajectory-associated semantic details in the video, including instance captions, instance interactions, and overall video caption. Please notice that, our SMOTer is specially designed for SMOT and trained in an end-to-end manner, as in Sec 4.4.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Target Trajectory Estimation for Tracking", + "text_level": 1, + "bbox": [ + 215, + 146, + 609, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Give a video with $N$ frames, SMOTer first uses a CNN backbone [49] to extract their features $\\{f_i\\}_{i=1}^N$ , where $f_i$ denotes the feature of the $i^{\\text{th}}$ frame. Then, an initial set of proposals is generated using proposal generator of a popular detection architecture [14] with threshold filtering by $\\tau_p$ . Afterwards, all reserved proposals are send to the association module (BYTE [53] is used in SMOTer for association) to obtain object trajectories. Finally, leveraging target coordinates within the trajectories and image features, target features are acquired for each target in trajectories through RoI Pooling [19]. Specifically, given feature $f_i$ of the $i^{\\text{th}}$ frame and target coordinates $B_j^i$ corresponding to the $j^{\\text{th}}$ object trajectory in this frame, we extract the corresponding target feature $t_j^i = \\text{RoIPooling}(f_i, B_j^i)$ and apply it in subsequent feature fusion for semantic understanding.", + "bbox": [ + 212, + 176, + 787, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In summary, at this stage, SMOTer predicts object trajectories for tracking in a given video and extracts image and target features for subsequent stages.", + "bbox": [ + 212, + 345, + 787, + 377 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Feature Fusion", + "text_level": 1, + "bbox": [ + 215, + 402, + 385, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Given image and target features from previous stage in Sec. 4.1, we design two feature fusion modules, including attention-based Video Fusion Module (VFM) and concatenation-based Trajectory Fusion Module (TFM). VFM aims to merge image-level features into overall video features while TFM focuses on integrating target features from each frame into target trajectory features.", + "bbox": [ + 212, + 433, + 787, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Video Fusion Module. VFM takes features $\\{f_i\\}_{i=1}^N$ of $N$ frames as input and sequentially feeds them into a cross-attention module for video feature fusion. In specific, the fused video feature $F_i$ up to the $i^{\\text{th}}$ ( $i > 1$ ) frame is computed as", + "bbox": [ + 212, + 508, + 787, + 556 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} = \\mathbb {C A} \\left(F _ {i - 1}, f _ {i}\\right) \\quad i > 1 \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 574, + 785, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{CA}(\\mathbf{z},\\mathbf{u})$ denotes cross-attention with $\\mathbf{z}$ generating query and $\\mathbf{u}$ key/value as in [41]. For the first frame ( $i = 1$ ), $F_{1} = f_{1}$ . Finally, with the help of VFM, we can generate the overall fused video feature $\\tilde{F} = F_{N}$ .", + "bbox": [ + 212, + 599, + 787, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Trajectory Fusion Module. Unlike VFM, TFM is to integrate multiple target features belonging to the same object trajectory, which are distributed across different frames, into a single trajectory feature. Concretely, given $t_i^j$ for object feature of trajectory $j$ in frame $i$ , we perform average pooling to yield a relatively simple 2D feature $\\hat{t}_i^j = \\mathsf{AvgPooling}(t_i^j)$ , where $\\hat{t}_i^j \\in \\mathbb{R}^{4 \\times d}$ with dimension $d = 256$ . Then, for the $j^{\\mathrm{th}}$ object trajectory, we first concatenate its discrete features into an integrated one $\\hat{T}_j = \\mathsf{Concat}(\\hat{t}_1^j, \\hat{t}_2^j, \\dots, \\hat{t}_{N_j}^j)$ , where $N_j$ denotes trajectory length. After that, self-attention is used to enhance trajectory feature as follows", + "bbox": [ + 212, + 645, + 787, + 767 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {T} _ {j} = \\mathbf {S A} (\\hat {T} _ {j}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 457, + 782, + 785, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{SA}(\\mathbf{z})$ denotes self-attention with $\\mathbf{z}$ generating query/key/value as in [41], and $\\bar{T}_j$ is the enhanced feature for trajectory $j$ .", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Beyond MOT: Semantic Multi-Object Tracking", + "bbox": [ + 413, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Semantic Understanding", + "text_level": 1, + "bbox": [ + 215, + 146, + 467, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the third phase, SMOTer applies the fused video feature $\\tilde{F}$ and trajectory features $\\tilde{T} = \\{\\tilde{T}_1,\\tilde{T}_2,\\dots,\\tilde{T}_M\\}$ , where $M$ denotes the number of trajectories in the video, to tackle trajectory-associated semantic understanding tasks in SMOT, i.e., video captioning, instance captioning, and interaction recognition.", + "bbox": [ + 212, + 170, + 782, + 231 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Video captioning. Unlike current video captioning, in SMOTer we incorporate trajectory features into describing video content and allow more comprehensive human-centric caption generation. Specifically, SMOTer sequentially injects trajectory features $\\tilde{T}$ into video feature $\\tilde{F}$ with cross-attention and then predicts captioning result $R_{\\mathrm{Vid}}$ via a linear project layer and a text decoder, as follows", + "bbox": [ + 212, + 232, + 782, + 308 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\mathrm {V i d}} = \\operatorname {D e c} _ {\\mathrm {V i d}} (\\operatorname {L i n e a r} (\\mathbf {C A} (\\tilde {F}, \\tilde {T}))) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 318, + 784, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathsf{Dec}_{\\mathrm{Vid}}(\\cdot)$ is a video text decoder for video captioning (please kindly refer to supplementary material for details), and $\\mathbf{Linear}(\\cdot)$ is a linear projection. Instance captioning. Different from video captioning, for instance captioning, we directly generate the results based on the target trajectory features. Specifically, given the feature $\\tilde{T}_j$ for the $j^{\\mathrm{th}}$ ( $1 \\leq j \\leq M$ ) trajectory, we apply a linear projection layer followed by the designed text decoder to predict the captioning result $R_{\\mathrm{Ins}}^j$ of each trajectory, depicted as follows,", + "bbox": [ + 212, + 345, + 784, + 454 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\text {I n s}} ^ {j} = \\operatorname {D e c} _ {\\text {I n s}} \\left(\\operatorname {L i n e a r} \\left(\\tilde {T} _ {j}\\right)\\right) \\quad (1 \\leq j \\leq M) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 464, + 784, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathsf{Dec}_{\\mathrm{Ins}}(\\cdot)$ is instance text decoder [46] for instance captioning. Please refer to supplementary material for more details because of space limitation.", + "bbox": [ + 212, + 492, + 782, + 522 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Interaction recognition. Besides captioning, SMOTer predicts the interaction between trajectories, which is crucial for video understanding. Particularly, in SMOTer we concentrate solely on interactions between two trajectories, as our current focus doesn't extend to understanding interactions involving three or more targets, leaving it for future research. More specifically, for two trajectories $j$ and $k$ , we assume the $j^{th}$ trajectory in a video sequence is the active trajectory and the $k^{th}$ trajectory the passive trajectory. Then, we first fuse the features of active and passive trajectories using cross-attention and then predict the interaction result $R_{\\mathrm{Int}}^{j,k}$ using multi-layer perception [40], as follows,", + "bbox": [ + 212, + 523, + 784, + 661 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\text {I n t}} ^ {j, k} = \\operatorname {M L P} (\\mathrm {C A} (\\tilde {T} _ {j}, \\tilde {T} _ {k})), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 671, + 784, + 690 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathsf{MLP}(\\cdot)$ denotes the multi-layer perception. Please note, interchanging the active and passive trajectories may yield disparate outcomes, i.e., $R_{\\mathrm{Int}}^{j,k}$ and $R_{\\mathrm{Int}}^{k,j}$ may exhibit dissimilar characteristics, as indicated by the formula above.", + "bbox": [ + 212, + 700, + 782, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 End-to-end Training", + "text_level": 1, + "bbox": [ + 215, + 768, + 429, + 786 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "End-to-end training of SMOTer is not easy due to conflict between two types of training requirements, i.e., the trajectory estimation is usually trained frame-by-frame, while the captioning tasks, and interaction recognition are often trained", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "at the entire video level, resulting in two different losses in training. In addition, due to the frame-by-frame training nature of trajectory estimation for tracking, the complete sequence is hard to be input at once during training, preventing the model from obtaining full trajectories thus hindering subsequent caption generation and interaction prediction. To address these issues, we draw inspiration from [60] and adopt a strategy of completing full trajectory association after detection on the entire video. Specifically, we first compute detection loss after each frame input and then perform trajectory association after the entire sequence is detected. Subsequently, with complete trajectories, we compute losses for trajectory-associated instance captioning, interaction recognition, and video captioning tasks, realizing end-to-end training of SMOTer. For the detailed training loss of SMOTer, please kindly refer to the supplementary material.", + "bbox": [ + 212, + 146, + 787, + 329 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 354, + 375, + 372 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experimental setup. We conduct experiments with 4 Nvidia Tesla V100 GPUs (32GB). The batch size is 1 per GPU. Each batch has a video clip with multiple frames. We use the AdamW optimizer with an initial learning rate of $5.0 \\times 10^{-4}$ . During training, we filter out proposals with scores lower than the threshold $\\tau_{p} = 0.3$ . For lost tracklets, we retain them for 30 frames in case they reappear. Due to memory constraints and limited accuracy of trajectory estimation in early training stages, we sample 6-8 frames per sequence for training detection model and only compute tracking related losses in the first 20K training iterations.", + "bbox": [ + 212, + 390, + 787, + 511 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As SMOT is new and no suitable methods exist for comparison with SMOTer, we logically divide SMOT into two components: trajectory estimation for tracking and semantic understanding (including instance/video captioning, and interaction recognition). Subsequently, we devise a series of two-stage models based on existing MOT frameworks. Leveraging several state-of-the-art and classic MOT models such as SORT [4], DeepSORT [43], OC-SORT [5], ByteTrack [53], TransTrack [39], MOTR [50], and MOTRv2 [55] as foundations, we integrate feature fusion and semantic generation functionalities akin to SMOTer and conduct comparative experiments on BenSMOT. To ensure fair experiments, we employ the same training framework for comparative experiments of two-stage approaches. Additionally, for models based on two-stage MOT methods like SORT, we use the same detector [14] as SMOTer during training and evaluation.", + "bbox": [ + 212, + 512, + 787, + 694 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Comparison on Trajectory Estimation for Tracking.", + "text_level": 1, + "bbox": [ + 214, + 720, + 687, + 736 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We first compare SMOTer with other MOT methods for trajectory estimation on BenSMOT. Specifically, SMOTer is fully end-to-end trained, while other two-stage MOT models are trained with their own strategies, concentrating solely on tracking. As depicted in Tab. 2, despite being at a certain degree of natural disadvantage, our SMOTer achieves comparable or even superior performance in tracking compared to state-of-the-art MOT models. SMOTer consistently", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Beyond MOT: Semantic Multi-Object Tracking", + "bbox": [ + 413, + 114, + 732, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bda312800e7a2c4da6b74cf7f07715037debe2d8cf8644801228d3e9ee6651fd.jpg", + "table_caption": [ + "Table 2: Comparison between SMOTer and two-stage MOT methods regarding tracking performance on BenSMOT. The best two results are shown in red and blue fonts." + ], + "table_footnote": [], + "table_body": "
MethodHOTA↑AssA↑DetA↑LocA↑MOTA↑FN↓FP↓IDs↓IDR↑IDP↑IDF1↑
SORT [4]48.4938.9560.9187.5053.582400151051387560.8548.4353.93
DeepSORT [43]50.1240.2361.4587.6754.292289055401127862.1051.1156.76
OC-SORT [5]51.0041.4263.3187.6155.192106153881504963.9253.1058.01
ByteTrack [53]68.8471.1567.1085.1573.87154197070171282.2574.8378.37
TransTrack [39]71.3173.3469.6791.3174.08201244420253085.6372.7578.67
MOTR [50]66.1073.1255.1486.3045.19312971117861772.3970.1268.97
MOTRv2 [55]65.2876.8251.3086.0945.52407652092343078.4765.5170.76
SMOTer (ours)71.9873.7170.7987.1177.71125346388170283.8277.9780.65
", + "bbox": [ + 246, + 184, + 751, + 275 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/74b1e24c58fb9994ad2bfb92766a3acf525839b588320eb0dbfbee39c36cfefb.jpg", + "table_caption": [ + "Table 3: Comparison of SMOTer against two-stage methods based on MOT models regarding semantic understanding. The best two results are highlighted in red and blue." + ], + "table_footnote": [], + "table_body": "
Video CaptionInstance CaptionInteraction
MethodBLEU↑ROUGE↑METEOR↑CIDEr↑BLEU↑ROUGE↑METEOR↑CIDEr↑Prec↑Rcll↑F1↑
SORT [4]0.2450.2240.2020.2980.2330.2450.2080.0560.3630.2590.302
DeepSORT [43]0.1980.2130.1870.3090.2380.2120.1990.0650.3650.2770.310
OC-SORT [5]0.2310.2520.2150.2420.2700.2050.1800.0330.3840.2910.331
ByteTrack [53]0.2240.2250.2120.2660.3040.2420.2240.0640.4430.2580.326
TransTrack [39]0.2470.2480.2090.2690.2830.2190.2010.0740.4060.3110.376
MOTR [50]0.1870.2540.2030.2440.2300.2090.1820.0610.4250.3140.354
MOTRv2 [55]0.2170.2580.2190.2480.2380.2410.2040.0590.3130.3950.349
SMOTer (ours)0.2450.2610.2230.3430.3060.2230.2090.0870.4340.3200.368
", + "bbox": [ + 246, + 327, + 751, + 426 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/083fe2dfb38a383cbc5f490de3535173529d8c656e2f6bd0cd72ec3703bfbb6e.jpg", + "table_caption": [ + "Table 4: Ablation experiments for evaluating various feature fusion strategies using results of three additional tasks. The best result is highlighted in red." + ], + "table_footnote": [], + "table_body": "
VFM (video fusion)TFM (trajectory fusion)
BLEU↑ROUGE↑METEOR↑CIDEr↑BLEU↑ROUGE↑METEOR↑CIDEr↑Pcn↑Rcll↑F1↑
Attention-based0.2450.2610.2230.3430.3030.2180.2080.0720.4060.3470.364
MLP-based0.2510.2230.2000.2200.3370.2190.2050.0720.3800.2520.303
Concatenation-based0.2300.2130.2150.3650.3060.2230.2090.0870.4340.3200.368
Addition-based0.1400.1570.1040.1270.1820.1370.1140.0410.2280.1630.205
", + "bbox": [ + 246, + 479, + 751, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "achieves top-tier performance across nearly all metrics, demonstrating impressive results in key metrics. For instance, it achieves a notable $71.98\\%$ in HOTA, $77.71\\%$ in MOTA, and $80.65\\%$ in IDF1. This suggests that subsequent training tasks do not degrade tracking performance, validating effectiveness of our end-to-end training strategy. In fact, by comparing with the standard ByteTrack, we observe that although SMOTer adopts a highly similar structure to ByteTrack in tracking, it achieves superior performance, showing $+3.15\\%$ in HOTA and $+3.74\\%$ in MOTA. Moreover, even in metrics where it does not excel, such as FP and LocA, SMOTer exhibits significant improvements compared to ByteTrack, implying that semantic understanding tasks can reciprocally aid in tracking.", + "bbox": [ + 212, + 568, + 787, + 720 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Comparison on Semantic Understanding", + "text_level": 1, + "bbox": [ + 215, + 739, + 596, + 757 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Besides tracking, we compare SMOTer with a set of two-stage models on their semantic understanding capabilities. Similarly, SMOTer undergoes complete end-to-end training. However, other two-stage models first train trackers on their own to acquire target trajectories and then use these trajectories to separately train the semantic understanding tasks (with the same feature fusion as SMOTer).", + "bbox": [ + 212, + 763, + 784, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2b4035949ae0bcf10d5fa8bf6ce10f71f8ba56bfc486d6cb00d78d7addd1c7b3.jpg", + "image_caption": [ + "(a) Comparison results on object tracking" + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 357, + 210 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eec43bb00846fecfcfb1c050803386a7698d65e8f4848611f014fe75ff238dcb.jpg", + "image_caption": [ + "Fig. 4: Ablation experiments for different association strategies, including comparison results on object tracking in image (a), on video captioning in image (b), on instance captioning in image (c), and on interaction recognition in image (d)." + ], + "image_footnote": [], + "bbox": [ + 362, + 145, + 500, + 210 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7338f1c6ec8beb8ee6e3521a509bc3f6f29779d9e894a4e925cdc13666f3023a.jpg", + "image_caption": [ + "(b) Comparison results on video captioning" + ], + "image_footnote": [], + "bbox": [ + 501, + 146, + 638, + 210 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c9a35591f8b1de1674dfbf715070f47fabaf6320a0bbcbf274d23326188c0c7a.jpg", + "image_caption": [ + "(c) Comparison results on instance captioning", + "(d) Comparison results on interaction recognition" + ], + "image_footnote": [], + "bbox": [ + 642, + 147, + 781, + 210 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "From Tab. 3, we can see SMOTer achieves highly favorable results on almost all metrics. Particularly, compared to the baseline method ByteTrack, SMOTer surpasses it on most metrics, including $+2.1\\%$ BLEU and $+7.7\\%$ CIDEr on video captioning, $+2.3\\%$ CIDEr on instance captioning, and $+4.2\\%$ F1 score on interaction recognition. More results can be seen in supplementary material.", + "bbox": [ + 212, + 311, + 784, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 409, + 387, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of different feature fusion. Feature fusion is a crucial component of SMOTer for generating overall video and enhanced trajectory features. In this work, we study four types of feature fusion strategies for VFM and TFM, including attention-based fusion (using cross-attention), MLP-based fusion (using MLP module), concatenation-based fusion, and addition-based fusion. Please refer to supplementary material for detailed architectures. To assess the impact for VFM, we compare the performance on video caption generation because the fused video feature from VFM is used for this sub-task, while for TFM, we measure the results on instance captioning and interaction recognition since these two sub-tasks rely on the fused trajectory features. Tab. 4 shows the results of different fusion mechanisms for VFM and TFM. We can see that attention-based fusion works generally better for VFM by achieving best ROUGE score of 0.261 and METEOR score of 0.223 and second best BLEU score of 0.245 and CIDEr score of 0.343 for video captioning. For TFM, concatenation-based fusion shows more superior results by obtaining best scores on ROUGE, METEOR, CIDEr for instance captioning and on Precision and F1 for interaction recognition.", + "bbox": [ + 212, + 431, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis on association mechanism. Association is necessary in SMOTer to generate target trajectories. In our method, we leverage the popular BYTE [54] for proposal association. To analyze the impact of different association strategies, we compare other representative manners, including SORT [4], DeepSORT [43], and OC-SORT [5], with the adopted BYTE. Fig. 4 demonstrates the comparison results of association mechanisms in SMOTer (using BYTE) and other approaches on four tasks, including object tracking (Fig. 4 (a)), video captioning (Fig. 4 (b)), instance captioning (Fig. 4 (c)), and interaction recognition (Fig. 4 (d)). From the results in Fig. 4 (a)-(d), we can observe that BYTE used in our SMOTer can achieve the best or competitive performance on different metrics for various tasks, validating its effectiveness for semantic multi-object tracking.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Beyond MOT: Semantic Multi-Object Tracking", + "bbox": [ + 413, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/322d364199767ad9cf9d8806f1e696b5b63885f30b903cba1207be6cdef00f0a.jpg", + "table_caption": [ + "Table 5: Study of different thresholds ${\\tau }_{p}$ . The best result is highlighted in red." + ], + "table_footnote": [], + "table_body": "
TrackingVideo CaptionInstance CaptionInteraction
τpHOTA↑MOTA↑IDF1↑METEOR↑CIDEr↑METEOR↑CIDEr↑Prcn↑Rcll↑F1↑
0.170.5172.3078.800.2190.2560.1920.0690.3780.3780.319
0.271.1073.8078.900.2150.2890.1990.0690.3840.3650.331
0.371.9877.7180.650.2240.3430.2090.0870.4340.3200.368
0.471.1976.3676.010.2140.3300.2150.0910.4200.3020.341
0.568.3074.3478.900.2240.3430.2160.0870.3960.2910.310
", + "bbox": [ + 267, + 171, + 733, + 247 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Study of threshold $\\tau_{p}$ . The parameter $\\tau_{p}$ is used to remove low-confidence proposals to ensure the quality of subsequent trajectories, and thus crucial for all tasks in SMOTer. Because of this, we conduct an ablation study for investigating different $\\tau_{p}$ , and Tab. 5 displays the results. From Tab. 5, we can observe that, when $\\tau_{p} = 0.3$ , SMOTer achieves the overall best performance for different tasks.", + "bbox": [ + 212, + 273, + 787, + 351 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Discussion", + "text_level": 1, + "bbox": [ + 214, + 369, + 349, + 386 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Semantic understanding for improving tracking. SMOT aims to go beyond just tracking by incorporating additional semantic understanding tasks including trajectory-associated instance/video captioning and interaction recognition. We observe that these extra goals greatly improve the tracking performance. For example, as shown in Tab. 2, we can see that joint training of different tasks leads to better performance of SMOTer (71.98%/77.71% in HOTA/MOTA) in tracking compared to its baseline ByteTrack (68.84%/73.87 in HOTA/MOTA). The possible reason is that SMOTer needs to comprehensively understand the video content for semantic multi-object tracking, and the learned semantic knowledge helps distinguish different object trajectories.", + "bbox": [ + 212, + 393, + 787, + 545 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Challenge in SMOT. One reason why SMOT is challenging lies in the complicated behaviors of targets, often leading to lengthy instance captions. For instance, a basketball player can be dribbling, shooting, and defending in a game. Models often miss some states in understanding. In BenSMOT, the average video length is about 23 seconds, with objects typically present for around 21 seconds, resulting in instance captions averaging over 35 words. This complexity poses a challenge for SMOT and indicates a potential area for future improvement.", + "bbox": [ + 212, + 546, + 789, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 672, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce SMOT to expand the scope of MOT by integrating \"where\" and \"what\". To facilitate the research of SMOT, we propose the large-scale BenSMOT by including 3,292 videos from more than 40 diverse scenarios. To our best knowledge, BenSMOT is the first benchmark dedicated to SMOT. Furthermore, to encourage algorithm development on BenSMOT, we introduce SMOTer, an end-to-end tracker designed for SMOT. Our results exhibit SMOTer surpasses offline-combination strategies, showing efficacy. By presenting BenSMOT and SMOTer, we hope to inspire more future research on SMOT.", + "bbox": [ + 212, + 703, + 787, + 823 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. Heng Fan was not supported by any fund for this work.", + "bbox": [ + 214, + 824, + 782, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bai, H., Cheng, W., Chu, P., Liu, J., Zhang, K., Ling, H.: Gmot-40: A benchmark for generic multiple object tracking. In: CVPR (2021)", + "2. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: ACL Workshop (2005)", + "3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. JIVP (2008)", + "4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: ICIP (2016)", + "5. Cao, J., Pang, J., Weng, X., Khirodkar, R., Kitani, K.: Observation-centric sort: Rethinking sort for robust multi-object tracking. In: CVPR (2023)", + "6. Chen, S., Shi, Z., Mettes, P., Snoek, C.G.: Social fabric: Tubelet compositions for video relation detection. In: ICCV (2021)", + "7. Chu, P., Wang, J., You, Q., Ling, H., Liu, Z.: Transmot: Spatial-temporal graph transformer for multiple object tracking. In: WACV (2023)", + "8. Cui, Y., Zeng, C., Zhao, X., Yang, Y., Wu, G., Wang, L.: Sportsmot: A large multi-object tracking dataset in multiple sports scenes. ICCV (2023)", + "9. Dave, A., Khurana, T., Tokmakov, P., Schmid, C., Ramanan, D.: Tao: A large-scale benchmark for tracking any object. In: ECCV (2020)", + "10. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J., Cremers, D., Reid, I., Roth, S., Schindler, K., Leal-Taixe, L.: Mot20: A benchmark for multi object tracking in crowded scenes. arXiv (2020)", + "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)", + "2. Du, D., Qi, Y., Yu, H., Yang, Y., Duan, K., Li, G., Zhang, W., Huang, Q., Tian, Q.: The unmanned aerial vehicle benchmark: Object detection and tracking. In: ECCV (2018)", + "3. Du, Y., Zhao, Z., Song, Y., Zhao, Y., Su, F., Gong, T., Meng, H.: Strongsort: Make deepsort great again. TMM (2023)", + "4. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: ICCV (2019)", + "5. Fellbaum, C.: WordNet: An electronic lexical database (1998)", + "6. Ferryman, J., Shahrokni, A.: Pets2009: Dataset and challenge. In: PET Workshop (2009)", + "7. Gao, R., Wang, L.: Memoritr: Long-term memory-augmented transformer for multi-object tracking. In: ICCV (2023)", + "8. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset (2013)", + "9. Girshick, R.: Fast r-cnn. In: ICCV (2015)", + "20. Han, X., Pasquier, T., Bates, A., Mickens, J., Seltzer, M.: Unicorn: Runtime provenance-based detector for advanced persistent threats. ECCV (2020)", + "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "22. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015)", + "23. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV (2017)", + "24. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv (2015)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Beyond MOT: Semantic Multi-Object Tracking", + "bbox": [ + 413, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "25. Liang, X., Lee, L., Xing, E.P.: Deep variation-structured reinforcement learning for visual relationship and attribute detection. In: CVPR (2017)", + "26. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL (2004)", + "27. Lin, K., Li, L., Lin, C.C., Ahmed, F., Gan, Z., Liu, Z., Lu, Y., Wang, L.: Swinbert: End-to-end transformers with sparse attention for video captioning. In: CVPR (2022)", + "28. Liu, C., Jin, Y., Xu, K., Gong, G., Mu, Y.: Beyond short-term snippet: Video relation detection with spatio-temporal global context. In: CVPR (2020)", + "29. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors. In: ECCV (2016)", + "30. Luiten, J., Osep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: Hota: A higher order metric for evaluating multi-object tracking. IJCV (2021)", + "31. Maggiolino, G., Ahmad, A., Cao, J., Kitani, K.: Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. ICIP (2023)", + "32. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: CVPR (2022)", + "33. Milan, A., Leal-Taixe, L., Reid, I., Roth, S., Schindler, K.: Mot16: A benchmark for multi-object tracking. arXiv (2016)", + "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002)", + "35. Ristani, E., Solera, F., Zou, R., Cucchiara, R., Tomasi, C.: Performance measures and a data set for multi-target, multi-camera tracking. In: ECCV (2016)", + "36. Seo, P.H., Nagrani, A., Arnab, A., Schmid, C.: End-to-end generative pretraining for multimodal video captioning. In: CVPR (2022)", + "37. Shen, Y., Gu, X., Xu, K., Fan, H., Wen, L., Zhang, L.: Accurate and fast compressed video captioning. In: ICCV (2023)", + "38. Sun, P., Cao, J., Jiang, Y., Yuan, Z., Bai, S., Kitani, K., Luo, P.: Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In: CVPR (2022)", + "39. Sun, P., Cao, J., Jiang, Y., Zhang, R., Xie, E., Yuan, Z., Wang, C., Luo, P.: Transtrack: Multiple object tracking with transformer. arXiv (2020)", + "40. Taud, H., Mas, J.: Multilayer perceptron (mlp). Geomatic approaches for modeling land change scenarios (2018)", + "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NIPS (2017)", + "42. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015)", + "43. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: ICIP (2017)", + "44. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: ECCV (2022)", + "45. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: CVPR (2023)", + "46. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023)", + "47. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: ECCV (2022)", + "48. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: CVPR (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: CVPR (2018)", + "50. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: ECCV (2022)", + "51. Zhang, H., Kyaw, Z., Chang, S.F., Chua, T.S.: Visual translation embedding network for visual relation detection. In: CVPR (2017)", + "52. Zhang, L., Gao, J., Xiao, Z., Fan, H.: Animaltrack: A benchmark for multi-animal tracking in the wild. IJCV (2023)", + "53. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: ECCV (2022)", + "54. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV (2021)", + "55. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In: CVPR (2023)", + "56. Zheng, S., Chen, S., Jin, Q.: Vrderformer: End-to-end video visual relation detection with transformers. In: CVPR (2022)", + "57. Zhou, L., Zhou, Y., Corso, J.J., Socher, R., Xiong, C.: End-to-end dense video captioning with masked transformer. In: CVPR (2018)", + "58. Zhou, X., Arnab, A., Sun, C., Schmid, C.: Dense video object captioning from disjoint supervision. arXiv (2023)", + "59. Zhou, X., Koltun, V., Krahenbihl, P.: Tracking objects as points. In: ECCV (2020)", + "60. Zhou, X., Yin, T., Koltun, V., Krahenbuhl, P.: Global tracking transformers. In: CVPR (2022)", + "61. Zhu, P., Wen, L., Du, D., Bian, X., Fan, H., Hu, Q., Ling, H.: Detection and tracking meet drones challenge. TPAMI (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 507 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Beyond MOT: Semantic Multi-Object Tracking", + "bbox": [ + 413, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Beyond MOT_ Semantic Multi-Object Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_model.json b/2024/Beyond MOT_ Semantic Multi-Object Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..19ce312e6265062d892028384e498920ab04ed18 --- /dev/null +++ b/2024/Beyond MOT_ Semantic Multi-Object Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_model.json @@ -0,0 +1,2445 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.231, + 0.141, + 0.773, + 0.166 + ], + "angle": 0, + "content": "Beyond MOT: Semantic Multi-Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.189, + 0.768, + 0.221 + ], + "angle": 0, + "content": "Yunhao Li\\(^{1,2}\\), Qin Li\\(^{1}\\), Hao Wang\\(^{2}\\), Xue Ma\\(^{1}\\), Jiali Yao\\(^{3}\\), Shaohua Dong\\(^{4}\\), Heng Fan\\(^{4\\dagger}\\), and Libo Zhang\\(^{1,2,3\\dagger\\sharp}\\)" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.232, + 0.677, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Software Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.247, + 0.652, + 0.26 + ], + "angle": 0, + "content": "\\(^{2}\\) University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.26, + 0.786, + 0.274 + ], + "angle": 0, + "content": "\\(^{3}\\) Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.275, + 0.761, + 0.289 + ], + "angle": 0, + "content": "\\(^{4}\\) Department of Computer Science & Engineering, University of North Texas" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.324, + 0.744, + 0.657 + ], + "angle": 0, + "content": "Abstract. Current multi-object tracking (MOT) aims to predict trajectories of targets (i.e., \"where\") in videos. Yet, knowing merely \"where\" is insufficient in many crucial applications. In comparison, semantic understanding such as fine-grained behaviors, interactions, and overall summarized captions (i.e., \"what\") from videos, associated with \"where\", is highly-desired for comprehensive video analysis. Thus motivated, we introduce Semantic Multi-Object Tracking (SMOT), that aims to estimate object trajectories and meanwhile understand semantic details of associated trajectories including instance captions, instance interactions, and overall video captions, integrating \"where\" and \"what\" for tracking. In order to foster the exploration of SMOT, we propose BenSMOT, a largescale Benchmark for Semantic MOT. Specifically, BenSMOT comprises 3,292 videos with 151K frames, covering various scenarios for semantic tracking of humans. BenSMOT provides annotations for the trajectories of targets, along with associated instance captions in natural language, instance interactions, and overall caption for each video sequence. To our best knowledge, BenSMOT is the first publicly available benchmark for SMOT. Besides, to encourage future research, we present a novel tracker named SMOTer, which is specially designed and end-to-end trained for SMOT, showing promising performance. By releasing BenSMOT, we expect to go beyond conventional MOT by predicting \"where\" and \"what\" for SMOT, opening up a new direction in tracking for video understanding. We will release BenSMOT and SMOTer at https://github.com/Nathan-Li123/SMOTer." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.67, + 0.721, + 0.685 + ], + "angle": 0, + "content": "Keywords: Semantic Multi-Object Tracking (SMOT) \\(\\cdot\\) Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.377, + 0.725 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Multi-object tracking (MOT) is a fundamental problem in computer vision with many applications such as autonomous driving, robotics, and visual surveillance. Current MOT tasks (e.g., [9,18,33,38,48]) usually focus on predicting trajectories of the targets from videos, i.e., answering the question of \"where are the targets\". Despite considerable advancements in deep learning era (e.g., [17,32,44,55,60])," + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.751, + 0.841 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Equal advising and co-last author \\(\\sharp\\)Corresponding author (libo@iscas.ac.cn)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.145, + 0.773, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.471 + ], + "angle": 0, + "content": "Fig. 1: Existing multi-object tracking (MOT) focusing on predicting trajectories only (see (a)) and our semantic multi-object tracking (SMOT) aiming at estimating trajectories and understanding their semantics (see (b)). Best viewed in color for all figures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.788, + 0.606 + ], + "angle": 0, + "content": "knowing merely \"where\" in existing MOT is insufficient for video understanding in many important applications. For example, besides trajectories, knowing more trajectory-associated semantic details of the objects, such as the behaviors and their interactions with the surroundings (i.e., \"what\"), is greatly beneficial for visual surveillance and robotics, which grows the interest in expanding modern MOT (i.e., addressing \"where\") with trajectory-associated semantic analysis (i.e., understanding \"what\") for more comprehensive video understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Thus motivated, in this work we introduce a new type of MOT task, dubbed Semantic Multi-Object Tracking (SMOT)\\(^5\\), which aims to expand scope of MOT beyond merely predicting trajectories to capture rich semantic details of objects in videos. Specifically, besides trajectory estimation, SMOT comprises three additional trajectory-associated semantic understanding tasks, including instance captioning, instance interaction recognition, and video captioning, as illustrated in Fig. 1. In particular, instance captioning aims at describing objects and their behaviors in human language, answering \"what are the objects doing\"; instance interaction recognition is to capture relations between objects, answering \"what are the relations between objects\"; video captioning provides the overall scenario understanding based on trajectories, answering \"what is happening\". Note that, all these three additional tasks are associated with object trajectories in videos. Ultimately, the goal of SMOT is to integrate \"where\" (i.e., object trajectories)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Here by \"semantic\", we emphasize high-level trajectory-based activity understanding in videos in the context of tracking, instead of category as in semantic segmentation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.416, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond MOT: Semantic Multi-Object Tracking" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.786, + 0.253 + ], + "angle": 0, + "content": "and \"what\" (i.e., trajectory-associated semantics) for video understanding, going beyond MOT predicting only trajectories (see again Fig. 1). Compared to conventional MOT, our SMOT is a multimodal task across vision and language, which is more challenging yet practical. It is worthy to notice that, SMOT is a natural extension of MOT. Despite sharing some similarities with video captioning [23], SMOT differs in that it aims at language understanding of dense target trajectories from videos in the context of instance-level tracking." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.258, + 0.787, + 0.484 + ], + "angle": 0, + "content": "To foster study of SMOT, we propose BenSMOT, a large-scale Benchmark for Semantic MOT. Specifically, BenSMOT contains 3,292 video sequences with 151K frames, captured from more than 40 diverse daily-life scenarios for human-centric semantic tracking. In BenSMOT, there exist more than 7.8K instances, labeled with over 335K bounding boxes for trajectory prediction. For each instance, we provide fine-grained human language to describe the behaviors for instance captioning, resulting in 7.8K sentences. In addition, to capture interactions between objects, BenSMOT provides more than 14K interaction annotations from a rich collection of 335 kinds of interactions. Moreover, for each sequence, a summary text is provided for overall understanding of instances and backgrounds, leading to 3,292 video caption texts in total in BenSMOT. To ensure the high quality of BenSMOT, annotations are manually labeled with careful inspection and refinement. To our best knowledge, BenSMOT is the first publicly available benchmark dedicated to SMOT. By releasing BenSMOT, we expect it to serve as a platform for advancing the research on SMOT." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.489, + 0.787, + 0.655 + ], + "angle": 0, + "content": "Furthermore, in order to facilitate the development of SMOT algorithms on BenSMOT, we present a simple yet effective tracker dubbed SMOTer. Specifically, SMOTer is built upon a multi-object tracking model [53]. After generating object trajectories, we extract features for each trajectory and design additional prediction heads for semantic understanding of instance captioning, interaction recognition, and video captioning. Note, SMOTer is not a simple offline assembly of models for different tasks. Instead, it is specially designed for SMOT and trained in an end-to-end manner for predicting trajectories and understanding their semantics in videos. Despite simplicity, SMOTer shows promising performance for SMOT, and outperforms the offline-combination strategies, evidencing its effectiveness. We expect it to provide a reference for future research." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.659, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We notice there exists a concurrent work combining multi-object tracking and trajectory captioning for dense video object captioning [58]. Compared with [58], our work differs in three aspects. First, task-wise, besides trajectory estimation and captioning, the SMOT in this work contains instance interaction and overall video captioning, providing more semantic details. Second, dataset-wise, we introduce a large-scale benchmark, BenSMOT, that is dedicated for SMOT and supports complete end-to-end model training, while the work of [58] borrows datasets from different tasks, which results in disjoint training of the model on different tasks and may thus degrade performance. Third, model-wise, owing to BenSMOT, we propose an end-to-end algorithm that demonstrates promising results on SMOT, while the approach in [58] is disjointly learned and does not support predicting instance interactions and overall video captioning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.129 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "In summary, we make the following contributions: \\(\\spadesuit\\) We introduce semantic multi-object tracking (SMOT), a new tracking paradigm that expands existing MOT task by integrating \"where\" and \"what\"; \\(\\spadesuit\\) We present BenSMOT, a large-scale dataset with 3,292 videos and rich annotations for SMOT; \\(\\spadesuit\\) We propose SMOTer, a simple but effective tracker to facilitate future research of SMOT; \\(\\spadesuit\\) We show that SMOTer achieves promising performance for SMOT and conduct in-depth analysis on it to provide guidance for future algorithm design." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.278, + 0.398, + 0.294 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.583 + ], + "angle": 0, + "content": "MOT Benchmarks. Benchmarks have played a crucial role in facilitating the development of MOT. PETS2009 [16] is one of the earliest benchmarks for multipedestrian tracking. Later, the popular MOT Challenge [10, 24, 33] has been introduced with more crowded videos and greatly advanced MOT. KITTI [18] and BDD100K [48] are specifically designed for object tracking in autonomous driving. ImageNet-Vid [11] provides trajectory annotations for 30 categories across over 1,000 videos, and TAO [9] further expands object classes to 833 for generic multi-object tracking. To foster MOT in specific dancing and sport scenarios, DanceTrack [38] and SportsMOT [8] have been presented for dancer and player tracking. AnimalTrack [52] focuses on tracking various animals in wild scenes. Moreover, UAVDT [12] and VisDrone [61] provide platforms for tracking targets with drones. Different from the above MOT benchmarks for object trajectory prediction only (\"where\"), BenSMOT is specially developed for the new task of SMOT to serve as a platform for simultaneous estimation of target trajectories and trajectory-associated semantic understanding. To this end, it offers rich annotations of not only object trajectories as in existing standard MOT datasets (\"where\") but also semantic details of objects (\"what\") such as instance trajectory captions, interactions and overall video captions (see Fig. 1 again)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.78 + ], + "angle": 0, + "content": "MOT Algorithms. MOT algorithms have seen great progress in recent years. One of the popular paradigms for multi-object tracking is the so-called Tracking-by-Detection. This paradigm involves initial object detection followed by association based on these detections, forming the basis for many representative methods [5,13,21,31,43,53]. In this case, MOT methods typically improve their performance by enhancing both detection and matching effectiveness. In addition, another prevalent paradigm is the Joint-Tracking-and-Detection [20,44,45,47,59], which integrates tracking and detection steps into a single process, enabling end-to-end training. Recently, Transformer [41] has been introduced into MOT and exhibited remarkable improvements over previous tackers [7,17,32,39,50,55,60]. Our SMOTer is related to but different than existing MOT methods. Specifically, going beyond merely predicting object trajectories, SMOTer also aims at semantic understanding of trajectories, integrating “where” and “what”." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Video Captioning. Video captioning is a multimodal vision-language task that aims to automatically describe the video content using natural language. Owing to its important applications in video event commentary and human-computer interaction, video captioning has drawn increasing attention in the past decade" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.415, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Beyond MOT: Semantic Multi-Object Tracking" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "with many models proposed (e.g., [23, 27, 36, 37, 46, 57]). Similar to the video captioning task, SMOT utilizes natural language to describe instances and the overall video content. However, the difference is that our SMOT aims at language comprehension for dense target trajectories in the context of multi-object tracking, which is more challenging due to the requirement of accurate trajectories yet provides more fine-grained information for video understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Visual Relationship Detection. The goal of visual relationship detection task is to identify and comprehend the relationships (usually represented by a triplet of \\(<\\) subject, predicate, object \\(\\rangle\\)) between objects from an image (e.g., [25, 29, 51]). Besides static images, recent works extend visual relationship detection in the video domain (e.g., [6, 28, 56]). Our work is relevant to video visual relationship detection but different by extending this task target-specific trajectories, and thus more challenging due to complicated multi-object tracking scenarios." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.368, + 0.498, + 0.385 + ], + "angle": 0, + "content": "3 The Proposed BenSMOT" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.401, + 0.4, + 0.416 + ], + "angle": 0, + "content": "3.1 Design Principle" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.784, + 0.457 + ], + "angle": 0, + "content": "BenSMOT aims at providing a new platform for exploring human-centric SMOT. In construction of BenSMOT, we follow the following rules:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.785, + 0.532 + ], + "angle": 0, + "content": "(1) Dedicated benchmark. The key motivation of our BenSMOT is to provide a dedicated dataset for semantic multi-object tracking. In current deep learning era, a large number of videos are desired in benchmark construction for training robust tracking models. Considering this, we hope to build a dedicated dataset with more than 3,000 videos for human-centric semantic multi-object tracking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.785, + 0.593 + ], + "angle": 0, + "content": "(2) Diverse scenarios. For a dataset, diversity is crucial, in both training and evaluation, for developing general systems. In order to provide a diverse platform for SMOT, in BenSMOT we will include videos from more than 40 various scenes that range from daily life scenarios to dancing scenarios to sport scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.785, + 0.669 + ], + "angle": 0, + "content": "(3) High-quality annotations. High-quality annotation is crucial for a benchmark in both training and assessing models. In BenSMOT, we ensure its high quality by providing manual annotations for each sequence including object trajectories, instance captions, instance interactions and overall video captions. In this process, multi-round inspections and refinements will be carried out." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.457, + 0.785, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.403, + 0.708 + ], + "angle": 0, + "content": "3.2 Data Acquisition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "BenSMOT focuses on predicting the target trajectories and meanwhile understanding their semantics by instance captioning, interaction recognition, and overall video captioning. To this end, the video sequences in BenSMOT are desired to involve diverse multi-person activities and interactions, which is different from videos in existing benchmarks to some extent. Specifically, we first identify 47 common scenarios, involving different activities ranging from daily-life talk and play to different sports, that are suitable for tracking by drawing inspiration from [22]. Due to limited space, please refer to supplementary material for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 1: Summary of BenSMOT and its comparison with popular multi-object tracking benchmarks. \"n/a\" denotes that data is not available." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.185, + 0.753, + 0.317 + ], + "angle": 0, + "content": "
KITTI [18]MOT17 [33]MOT20 [10]BDD100k [48]TAO [9]GMOT-40 [1]DanceTrack [38]SportsMOT [8]BenSMOT (ours)
Videos501481,6002,907401002403,292
Min. length (s)n/a17.017.040.0n/a3.0n/an/a1.5
Avg. length (s)10.033.066.840.036.88.952.9n/a22.9
Max. length (s)n/a85.0133.040.0n/a24.2n/an/a116.0
Total length (s)498463535640,000106,9783565,292601575,499
Total tracks2,6001.3K3.83K131K17,2872,0269903,4017,792
Total boxes80K300K2,102K3,300K333K256Kn/a1,629K335K
Total frames15K11K13K318K2,674K9K106K150K151K
Instance CaptionsXXXXXXXX7,792
Instance InteractionsXXXXXXXX14K
Video SummariesXXXXXXXX3,292
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.784, + 0.42 + ], + "angle": 0, + "content": "detailed scenarios in BenSMOT. Afterwards, we search for raw sequences for each scenario under the Creative Commons License from YouTube, the largest and most popular video platform with massive real-world videos. Please notice, for each raw video, instead of using the whole sequence, we usually choose a suitable clip for our semantic multi-object tracking task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.785, + 0.558 + ], + "angle": 0, + "content": "Eventually, we compile a dataset for SMOT by including 3,292 videos with 151K frames. BenSMOT has an average video length of 23 seconds, with our current focus in this work not being on long-term SMOT. The total number of instance tracks is 7.8K, annotated with 335K bounding boxes for tracking. Besides bounding boxes, we provide 7.8K instance captions, 14K interactions, and 3,292 video summaries for trajectory-associated semantic understanding (how to annotate will be described in the following). To our best knowledge, BenSMOT by far is the first publicly available benchmark dedicated for SMOT. Tab. 1 summarizes our BenSMOT and compares it with popular MOT benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.585, + 0.403, + 0.599 + ], + "angle": 0, + "content": "3.3 Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In order to meet the requirements of SMOT, BenSMOT provides four types of annotations, including bounding box, instance caption, instance interaction, and the overall video caption. Specifically, for each trajectory in videos, similar to standard MOT benchmarks (e.g., [9,33]), we provide axis-aligned bounding boxes to indicate its spatial positions in the videos. For instance captioning, we label each trajectory with a precise sentence in natural language to describe the detailed behavior of the associated target. For interaction recognition, we first collect a set of 335 interactions (i.e., verbs), among them 327 from WordNet [15]. Note that, in our interaction set, the same verb may have different meanings and thus is assigned to different interaction types, e.g., hold(v.01) for “being the physical support of someone” and hold(v.02) for “someone in one’s hands or arm”, for precise relationship description. Due to limited space, we refer reader to supplementary material for the full set of interaction types. After this, we provide an directed interaction label, represented with a triplet of \\(<\\) subject predicate, object \\(\\rightharpoondown\\) (see Fig. 1 (b)), between two trajectories if associated objects" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.415, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Beyond MOT: Semantic Multi-Object Tracking" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.408, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.223, + 0.402, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.308, + 0.403, + 0.318 + ], + "angle": 0, + "content": "(d) Inst caption length distribution" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.144, + 0.594, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.223, + 0.583, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.309, + 0.593, + 0.318 + ], + "angle": 0, + "content": "(e) Video caption length distribution" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.144, + 0.78, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.223, + 0.773, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.309, + 0.767, + 0.318 + ], + "angle": 0, + "content": "(f) WordCloud of caption words" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.329, + 0.785, + 0.386 + ], + "angle": 0, + "content": "Fig. 2: Representative statistics on BenSMOT, including distributions of target trajectory length (in seconds) in (a), sequence length (in seconds) in (b), number of interactions per sequence in (c), instance caption length in (d), video caption length in (e), and wordcloud of all caption words with prepositions excluded in (f)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.784, + 0.445 + ], + "angle": 0, + "content": "are interacting. Finally, for the overall video captioning, we utilize a concise text summary to describe object trajectories in the video from a global perspective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.785, + 0.596 + ], + "angle": 0, + "content": "To ensure the high-quality annotations, we adopt a multi-round mechanism. Specifically, each video will be first manually annotated by a few volunteers who are familiar with the task and an expert working on related problems. After this initial round, all annotations of the video will be sent to a validation team former by more than two experts for inspection. If initial annotations are not unanimously agreed by the experts, they will be returned to the original labeling team for refinement using the feedback from the validation team. We repeat this procedure for each video until the annotations of all video sequences in BenSMOT are completed. An annotation example in our BenSMOT is shown in Fig. 1 (b), and more can be found in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.785, + 0.703 + ], + "angle": 0, + "content": "Statistics of annotation. In order to better understand BenSMOT, we show some representative statics of annotations in Fig. 2. Specifically, we demonstrate the distributions of object trajectory length, video length, number of instance interactions, instance caption length, video caption length, and wordcloud. From Fig. 2 (c), it is worth noting that different targets in videos have frequent interactions. In addition, for the instance captions in Fig. 2 (d), we provide relatively longer textual descriptions, which allows precise understanding." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.562, + 0.74 + ], + "angle": 0, + "content": "3.4 Dataset Split and Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Training/Test Split. BenSMOT contains 3,292 videos, captured from 47 different scenarios. Within each scenario, we use \\(70\\%\\) of the sequences for training, and the rest \\(30\\%\\) for test. During dataset split, we try to keep the distributions of training and test sets similar. Eventfully, the training set of BenSMOT comprises 2,284 sequences with 104K frames, and the test set consists of 1,008 videos with 47K frames. Please see supplementary material for more split details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.756, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.371, + 0.785, + 0.4 + ], + "angle": 0, + "content": "Fig. 3: Illustration of SMOTer, which contains three components of trajectory estimation for tracking, feature fusion, and trajectory-associated semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.609 + ], + "angle": 0, + "content": "Evaluation Metric. We apply multiple metrics on BenSMOT for evaluation. Specifically, to assess the performance in object trajectory estimation, we employ higher order tracking accuracy (HOTA), association accuracy (AssA), detection accuracy (DetA), and localization accuracy (LocA) by following [30], CLEAR metrics [3] including multiple object tracking accuracy (MOTA), false positives (FP), false negatives (FN), and ID switches (IDs), and ID metrics [35] containing identification precision (IDP), identification recall (IDR) and related F1 score (IDF1). For instance and video captioning, we follow existing methods (e.g., [27, 37]) and employ BLEU [34], ROUGE [26], METEOR [2], and CIDEr [42] for thorough evaluation. Lastly, to assess performance of interaction recognition, we leverage classic metrics such as Precision (Prcn), Recall (Rcll), and F1, which are widely used in the context of classification tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.629, + 0.655, + 0.648 + ], + "angle": 0, + "content": "4 Methodology: A New Baseline for SMOT" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Overview. To encourage development of SMOT algorithms, we propose a simple yet effective tracker on BenSMOT, dubbed SMOTer. As illustrated in Fig. 3, SMOTer logically consists of three key components. The first component, as detailed in Sec 4.1, is dedicated to object trajectory estimation for tracking, which is the basis for SMOT. Then, the second component, as explained in Sec 4.2, focuses on aggregating features from each frame into the overall video feature and meanwhile merging features of individual objects into target trajectory features using two distinct fusion modules, for subsequent semantic understanding. Finally, the third component, as described in Sec 4.3, aims to predict trajectory-associated semantic details in the video, including instance captions, instance interactions, and overall video caption. Please notice that, our SMOTer is specially designed for SMOT and trained in an end-to-end manner, as in Sec 4.4." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.415, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Beyond MOT: Semantic Multi-Object Tracking" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.611, + 0.163 + ], + "angle": 0, + "content": "4.1 Target Trajectory Estimation for Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Give a video with \\(N\\) frames, SMOTer first uses a CNN backbone [49] to extract their features \\(\\{f_i\\}_{i=1}^N\\), where \\(f_i\\) denotes the feature of the \\(i^{\\text{th}}\\) frame. Then, an initial set of proposals is generated using proposal generator of a popular detection architecture [14] with threshold filtering by \\(\\tau_p\\). Afterwards, all reserved proposals are send to the association module (BYTE [53] is used in SMOTer for association) to obtain object trajectories. Finally, leveraging target coordinates within the trajectories and image features, target features are acquired for each target in trajectories through RoI Pooling [19]. Specifically, given feature \\(f_i\\) of the \\(i^{\\text{th}}\\) frame and target coordinates \\(B_j^i\\) corresponding to the \\(j^{\\text{th}}\\) object trajectory in this frame, we extract the corresponding target feature \\(t_j^i = \\text{RoIPooling}(f_i, B_j^i)\\) and apply it in subsequent feature fusion for semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.378 + ], + "angle": 0, + "content": "In summary, at this stage, SMOTer predicts object trajectories for tracking in a given video and extracts image and target features for subsequent stages." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.404, + 0.386, + 0.418 + ], + "angle": 0, + "content": "4.2 Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.509 + ], + "angle": 0, + "content": "Given image and target features from previous stage in Sec. 4.1, we design two feature fusion modules, including attention-based Video Fusion Module (VFM) and concatenation-based Trajectory Fusion Module (TFM). VFM aims to merge image-level features into overall video features while TFM focuses on integrating target features from each frame into target trajectory features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.557 + ], + "angle": 0, + "content": "Video Fusion Module. VFM takes features \\(\\{f_i\\}_{i=1}^N\\) of \\(N\\) frames as input and sequentially feeds them into a cross-attention module for video feature fusion. In specific, the fused video feature \\(F_i\\) up to the \\(i^{\\text{th}}\\) (\\(i > 1\\)) frame is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.575, + 0.786, + 0.59 + ], + "angle": 0, + "content": "\\[\nF _ {i} = \\mathbb {C A} \\left(F _ {i - 1}, f _ {i}\\right) \\quad i > 1 \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.645 + ], + "angle": 0, + "content": "where \\(\\mathbf{CA}(\\mathbf{z},\\mathbf{u})\\) denotes cross-attention with \\(\\mathbf{z}\\) generating query and \\(\\mathbf{u}\\) key/value as in [41]. For the first frame (\\(i = 1\\)), \\(F_{1} = f_{1}\\). Finally, with the help of VFM, we can generate the overall fused video feature \\(\\tilde{F} = F_{N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.768 + ], + "angle": 0, + "content": "Trajectory Fusion Module. Unlike VFM, TFM is to integrate multiple target features belonging to the same object trajectory, which are distributed across different frames, into a single trajectory feature. Concretely, given \\( t_i^j \\) for object feature of trajectory \\( j \\) in frame \\( i \\), we perform average pooling to yield a relatively simple 2D feature \\( \\hat{t}_i^j = \\mathsf{AvgPooling}(t_i^j) \\), where \\( \\hat{t}_i^j \\in \\mathbb{R}^{4 \\times d} \\) with dimension \\( d = 256 \\). Then, for the \\( j^{\\mathrm{th}} \\) object trajectory, we first concatenate its discrete features into an integrated one \\( \\hat{T}_j = \\mathsf{Concat}(\\hat{t}_1^j, \\hat{t}_2^j, \\dots, \\hat{t}_{N_j}^j) \\), where \\( N_j \\) denotes trajectory length. After that, self-attention is used to enhance trajectory feature as follows" + }, + { + "type": "equation", + "bbox": [ + 0.459, + 0.784, + 0.786, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\tilde {T} _ {j} = \\mathbf {S A} (\\hat {T} _ {j}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathbf{SA}(\\mathbf{z})\\) denotes self-attention with \\(\\mathbf{z}\\) generating query/key/value as in [41], and \\(\\bar{T}_j\\) is the enhanced feature for trajectory \\(j\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.468, + 0.163 + ], + "angle": 0, + "content": "4.3 Semantic Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.784, + 0.232 + ], + "angle": 0, + "content": "In the third phase, SMOTer applies the fused video feature \\(\\tilde{F}\\) and trajectory features \\(\\tilde{T} = \\{\\tilde{T}_1,\\tilde{T}_2,\\dots,\\tilde{T}_M\\}\\), where \\(M\\) denotes the number of trajectories in the video, to tackle trajectory-associated semantic understanding tasks in SMOT, i.e., video captioning, instance captioning, and interaction recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.233, + 0.784, + 0.309 + ], + "angle": 0, + "content": "Video captioning. Unlike current video captioning, in SMOTer we incorporate trajectory features into describing video content and allow more comprehensive human-centric caption generation. Specifically, SMOTer sequentially injects trajectory features \\(\\tilde{T}\\) into video feature \\(\\tilde{F}\\) with cross-attention and then predicts captioning result \\(R_{\\mathrm{Vid}}\\) via a linear project layer and a text decoder, as follows" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.319, + 0.785, + 0.337 + ], + "angle": 0, + "content": "\\[\nR _ {\\mathrm {V i d}} = \\operatorname {D e c} _ {\\mathrm {V i d}} (\\operatorname {L i n e a r} (\\mathbf {C A} (\\tilde {F}, \\tilde {T}))) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.785, + 0.455 + ], + "angle": 0, + "content": "where \\(\\mathsf{Dec}_{\\mathrm{Vid}}(\\cdot)\\) is a video text decoder for video captioning (please kindly refer to supplementary material for details), and \\(\\mathbf{Linear}(\\cdot)\\) is a linear projection. Instance captioning. Different from video captioning, for instance captioning, we directly generate the results based on the target trajectory features. Specifically, given the feature \\(\\tilde{T}_j\\) for the \\(j^{\\mathrm{th}}\\) (\\(1 \\leq j \\leq M\\)) trajectory, we apply a linear projection layer followed by the designed text decoder to predict the captioning result \\(R_{\\mathrm{Ins}}^j\\) of each trajectory, depicted as follows," + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.465, + 0.785, + 0.484 + ], + "angle": 0, + "content": "\\[\nR _ {\\text {I n s}} ^ {j} = \\operatorname {D e c} _ {\\text {I n s}} \\left(\\operatorname {L i n e a r} \\left(\\tilde {T} _ {j}\\right)\\right) \\quad (1 \\leq j \\leq M) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.784, + 0.523 + ], + "angle": 0, + "content": "where \\(\\mathsf{Dec}_{\\mathrm{Ins}}(\\cdot)\\) is instance text decoder [46] for instance captioning. Please refer to supplementary material for more details because of space limitation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.785, + 0.662 + ], + "angle": 0, + "content": "Interaction recognition. Besides captioning, SMOTer predicts the interaction between trajectories, which is crucial for video understanding. Particularly, in SMOTer we concentrate solely on interactions between two trajectories, as our current focus doesn't extend to understanding interactions involving three or more targets, leaving it for future research. More specifically, for two trajectories \\( j \\) and \\( k \\), we assume the \\( j^{th} \\) trajectory in a video sequence is the active trajectory and the \\( k^{th} \\) trajectory the passive trajectory. Then, we first fuse the features of active and passive trajectories using cross-attention and then predict the interaction result \\( R_{\\mathrm{Int}}^{j,k} \\) using multi-layer perception [40], as follows," + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.672, + 0.785, + 0.691 + ], + "angle": 0, + "content": "\\[\nR _ {\\text {I n t}} ^ {j, k} = \\operatorname {M L P} (\\mathrm {C A} (\\tilde {T} _ {j}, \\tilde {T} _ {k})), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.784, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mathsf{MLP}(\\cdot)\\) denotes the multi-layer perception. Please note, interchanging the active and passive trajectories may yield disparate outcomes, i.e., \\(R_{\\mathrm{Int}}^{j,k}\\) and \\(R_{\\mathrm{Int}}^{k,j}\\) may exhibit dissimilar characteristics, as indicated by the formula above." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.431, + 0.787 + ], + "angle": 0, + "content": "4.4 End-to-end Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "End-to-end training of SMOTer is not easy due to conflict between two types of training requirements, i.e., the trajectory estimation is usually trained frame-by-frame, while the captioning tasks, and interaction recognition are often trained" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.415, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Beyond MOT: Semantic Multi-Object Tracking" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.33 + ], + "angle": 0, + "content": "at the entire video level, resulting in two different losses in training. In addition, due to the frame-by-frame training nature of trajectory estimation for tracking, the complete sequence is hard to be input at once during training, preventing the model from obtaining full trajectories thus hindering subsequent caption generation and interaction prediction. To address these issues, we draw inspiration from [60] and adopt a strategy of completing full trajectory association after detection on the entire video. Specifically, we first compute detection loss after each frame input and then perform trajectory association after the entire sequence is detected. Subsequently, with complete trajectories, we compute losses for trajectory-associated instance captioning, interaction recognition, and video captioning tasks, realizing end-to-end training of SMOTer. For the detailed training loss of SMOTer, please kindly refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.356, + 0.377, + 0.373 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.512 + ], + "angle": 0, + "content": "Experimental setup. We conduct experiments with 4 Nvidia Tesla V100 GPUs (32GB). The batch size is 1 per GPU. Each batch has a video clip with multiple frames. We use the AdamW optimizer with an initial learning rate of \\(5.0 \\times 10^{-4}\\). During training, we filter out proposals with scores lower than the threshold \\(\\tau_{p} = 0.3\\). For lost tracklets, we retain them for 30 frames in case they reappear. Due to memory constraints and limited accuracy of trajectory estimation in early training stages, we sample 6-8 frames per sequence for training detection model and only compute tracking related losses in the first 20K training iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.695 + ], + "angle": 0, + "content": "As SMOT is new and no suitable methods exist for comparison with SMOTer, we logically divide SMOT into two components: trajectory estimation for tracking and semantic understanding (including instance/video captioning, and interaction recognition). Subsequently, we devise a series of two-stage models based on existing MOT frameworks. Leveraging several state-of-the-art and classic MOT models such as SORT [4], DeepSORT [43], OC-SORT [5], ByteTrack [53], TransTrack [39], MOTR [50], and MOTRv2 [55] as foundations, we integrate feature fusion and semantic generation functionalities akin to SMOTer and conduct comparative experiments on BenSMOT. To ensure fair experiments, we employ the same training framework for comparative experiments of two-stage approaches. Additionally, for models based on two-stage MOT methods like SORT, we use the same detector [14] as SMOTer during training and evaluation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.721, + 0.688, + 0.737 + ], + "angle": 0, + "content": "5.1 Comparison on Trajectory Estimation for Tracking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We first compare SMOTer with other MOT methods for trajectory estimation on BenSMOT. Specifically, SMOTer is fully end-to-end trained, while other two-stage MOT models are trained with their own strategies, concentrating solely on tracking. As depicted in Tab. 2, despite being at a certain degree of natural disadvantage, our SMOTer achieves comparable or even superior performance in tracking compared to state-of-the-art MOT models. SMOTer consistently" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 2: Comparison between SMOTer and two-stage MOT methods regarding tracking performance on BenSMOT. The best two results are shown in red and blue fonts." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.185, + 0.753, + 0.276 + ], + "angle": 0, + "content": "
MethodHOTA↑AssA↑DetA↑LocA↑MOTA↑FN↓FP↓IDs↓IDR↑IDP↑IDF1↑
SORT [4]48.4938.9560.9187.5053.582400151051387560.8548.4353.93
DeepSORT [43]50.1240.2361.4587.6754.292289055401127862.1051.1156.76
OC-SORT [5]51.0041.4263.3187.6155.192106153881504963.9253.1058.01
ByteTrack [53]68.8471.1567.1085.1573.87154197070171282.2574.8378.37
TransTrack [39]71.3173.3469.6791.3174.08201244420253085.6372.7578.67
MOTR [50]66.1073.1255.1486.3045.19312971117861772.3970.1268.97
MOTRv2 [55]65.2876.8251.3086.0945.52407652092343078.4765.5170.76
SMOTer (ours)71.9873.7170.7987.1177.71125346388170283.8277.9780.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.287, + 0.785, + 0.316 + ], + "angle": 0, + "content": "Table 3: Comparison of SMOTer against two-stage methods based on MOT models regarding semantic understanding. The best two results are highlighted in red and blue." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.328, + 0.753, + 0.428 + ], + "angle": 0, + "content": "
Video CaptionInstance CaptionInteraction
MethodBLEU↑ROUGE↑METEOR↑CIDEr↑BLEU↑ROUGE↑METEOR↑CIDEr↑Prec↑Rcll↑F1↑
SORT [4]0.2450.2240.2020.2980.2330.2450.2080.0560.3630.2590.302
DeepSORT [43]0.1980.2130.1870.3090.2380.2120.1990.0650.3650.2770.310
OC-SORT [5]0.2310.2520.2150.2420.2700.2050.1800.0330.3840.2910.331
ByteTrack [53]0.2240.2250.2120.2660.3040.2420.2240.0640.4430.2580.326
TransTrack [39]0.2470.2480.2090.2690.2830.2190.2010.0740.4060.3110.376
MOTR [50]0.1870.2540.2030.2440.2300.2090.1820.0610.4250.3140.354
MOTRv2 [55]0.2170.2580.2190.2480.2380.2410.2040.0590.3130.3950.349
SMOTer (ours)0.2450.2610.2230.3430.3060.2230.2090.0870.4340.3200.368
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.44, + 0.784, + 0.468 + ], + "angle": 0, + "content": "Table 4: Ablation experiments for evaluating various feature fusion strategies using results of three additional tasks. The best result is highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.48, + 0.753, + 0.543 + ], + "angle": 0, + "content": "
VFM (video fusion)TFM (trajectory fusion)
BLEU↑ROUGE↑METEOR↑CIDEr↑BLEU↑ROUGE↑METEOR↑CIDEr↑Pcn↑Rcll↑F1↑
Attention-based0.2450.2610.2230.3430.3030.2180.2080.0720.4060.3470.364
MLP-based0.2510.2230.2000.2200.3370.2190.2050.0720.3800.2520.303
Concatenation-based0.2300.2130.2150.3650.3060.2230.2090.0870.4340.3200.368
Addition-based0.1400.1570.1040.1270.1820.1370.1140.0410.2280.1630.205
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.721 + ], + "angle": 0, + "content": "achieves top-tier performance across nearly all metrics, demonstrating impressive results in key metrics. For instance, it achieves a notable \\(71.98\\%\\) in HOTA, \\(77.71\\%\\) in MOTA, and \\(80.65\\%\\) in IDF1. This suggests that subsequent training tasks do not degrade tracking performance, validating effectiveness of our end-to-end training strategy. In fact, by comparing with the standard ByteTrack, we observe that although SMOTer adopts a highly similar structure to ByteTrack in tracking, it achieves superior performance, showing \\(+3.15\\%\\) in HOTA and \\(+3.74\\%\\) in MOTA. Moreover, even in metrics where it does not excel, such as FP and LocA, SMOTer exhibits significant improvements compared to ByteTrack, implying that semantic understanding tasks can reciprocally aid in tracking." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.597, + 0.758 + ], + "angle": 0, + "content": "5.2 Comparison on Semantic Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Besides tracking, we compare SMOTer with a set of two-stage models on their semantic understanding capabilities. Similarly, SMOTer undergoes complete end-to-end training. However, other two-stage models first train trackers on their own to acquire target trajectories and then use these trajectories to separately train the semantic understanding tasks (with the same feature fusion as SMOTer)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.415, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Beyond MOT: Semantic Multi-Object Tracking" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.359, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.213, + 0.36, + 0.231 + ], + "angle": 0, + "content": "(a) Comparison results on object tracking" + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.146, + 0.501, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.213, + 0.501, + 0.231 + ], + "angle": 0, + "content": "(b) Comparison results on video captioning" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.147, + 0.64, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.213, + 0.64, + 0.231 + ], + "angle": 0, + "content": "(c) Comparison results on instance captioning" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.148, + 0.782, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.213, + 0.782, + 0.231 + ], + "angle": 0, + "content": "(d) Comparison results on interaction recognition" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.243, + 0.787, + 0.285 + ], + "angle": 0, + "content": "Fig. 4: Ablation experiments for different association strategies, including comparison results on object tracking in image (a), on video captioning in image (b), on instance captioning in image (c), and on interaction recognition in image (d)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.785, + 0.389 + ], + "angle": 0, + "content": "From Tab. 3, we can see SMOTer achieves highly favorable results on almost all metrics. Particularly, compared to the baseline method ByteTrack, SMOTer surpasses it on most metrics, including \\(+2.1\\%\\) BLEU and \\(+7.7\\%\\) CIDEr on video captioning, \\(+2.3\\%\\) CIDEr on instance captioning, and \\(+4.2\\%\\) F1 score on interaction recognition. More results can be seen in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.388, + 0.424 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Impact of different feature fusion. Feature fusion is a crucial component of SMOTer for generating overall video and enhanced trajectory features. In this work, we study four types of feature fusion strategies for VFM and TFM, including attention-based fusion (using cross-attention), MLP-based fusion (using MLP module), concatenation-based fusion, and addition-based fusion. Please refer to supplementary material for detailed architectures. To assess the impact for VFM, we compare the performance on video caption generation because the fused video feature from VFM is used for this sub-task, while for TFM, we measure the results on instance captioning and interaction recognition since these two sub-tasks rely on the fused trajectory features. Tab. 4 shows the results of different fusion mechanisms for VFM and TFM. We can see that attention-based fusion works generally better for VFM by achieving best ROUGE score of 0.261 and METEOR score of 0.223 and second best BLEU score of 0.245 and CIDEr score of 0.343 for video captioning. For TFM, concatenation-based fusion shows more superior results by obtaining best scores on ROUGE, METEOR, CIDEr for instance captioning and on Precision and F1 for interaction recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Analysis on association mechanism. Association is necessary in SMOTer to generate target trajectories. In our method, we leverage the popular BYTE [54] for proposal association. To analyze the impact of different association strategies, we compare other representative manners, including SORT [4], DeepSORT [43], and OC-SORT [5], with the adopted BYTE. Fig. 4 demonstrates the comparison results of association mechanisms in SMOTer (using BYTE) and other approaches on four tasks, including object tracking (Fig. 4 (a)), video captioning (Fig. 4 (b)), instance captioning (Fig. 4 (c)), and interaction recognition (Fig. 4 (d)). From the results in Fig. 4 (a)-(d), we can observe that BYTE used in our SMOTer can achieve the best or competitive performance on different metrics for various tasks, validating its effectiveness for semantic multi-object tracking." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.237, + 0.145, + 0.764, + 0.16 + ], + "angle": 0, + "content": "Table 5: Study of different thresholds \\( {\\tau }_{p} \\) . The best result is highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.268, + 0.172, + 0.734, + 0.248 + ], + "angle": 0, + "content": "
TrackingVideo CaptionInstance CaptionInteraction
τpHOTA↑MOTA↑IDF1↑METEOR↑CIDEr↑METEOR↑CIDEr↑Prcn↑Rcll↑F1↑
0.170.5172.3078.800.2190.2560.1920.0690.3780.3780.319
0.271.1073.8078.900.2150.2890.1990.0690.3840.3650.331
0.371.9877.7180.650.2240.3430.2090.0870.4340.3200.368
0.471.1976.3676.010.2140.3300.2150.0910.4200.3020.341
0.568.3074.3478.900.2240.3430.2160.0870.3960.2910.310
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Study of threshold \\(\\tau_{p}\\). The parameter \\(\\tau_{p}\\) is used to remove low-confidence proposals to ensure the quality of subsequent trajectories, and thus crucial for all tasks in SMOTer. Because of this, we conduct an ablation study for investigating different \\(\\tau_{p}\\), and Tab. 5 displays the results. From Tab. 5, we can observe that, when \\(\\tau_{p} = 0.3\\), SMOTer achieves the overall best performance for different tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.371, + 0.351, + 0.387 + ], + "angle": 0, + "content": "5.4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.546 + ], + "angle": 0, + "content": "Semantic understanding for improving tracking. SMOT aims to go beyond just tracking by incorporating additional semantic understanding tasks including trajectory-associated instance/video captioning and interaction recognition. We observe that these extra goals greatly improve the tracking performance. For example, as shown in Tab. 2, we can see that joint training of different tasks leads to better performance of SMOTer (71.98%/77.71% in HOTA/MOTA) in tracking compared to its baseline ByteTrack (68.84%/73.87 in HOTA/MOTA). The possible reason is that SMOTer needs to comprehensively understand the video content for semantic multi-object tracking, and the learned semantic knowledge helps distinguish different object trajectories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.79, + 0.653 + ], + "angle": 0, + "content": "Challenge in SMOT. One reason why SMOT is challenging lies in the complicated behaviors of targets, often leading to lengthy instance captions. For instance, a basketball player can be dribbling, shooting, and defending in a game. Models often miss some states in understanding. In BenSMOT, the average video length is about 23 seconds, with objects typically present for around 21 seconds, resulting in instance captions averaging over 35 words. This complexity poses a challenge for SMOT and indicates a potential area for future improvement." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.674, + 0.36, + 0.691 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.824 + ], + "angle": 0, + "content": "In this paper, we introduce SMOT to expand the scope of MOT by integrating \"where\" and \"what\". To facilitate the research of SMOT, we propose the large-scale BenSMOT by including 3,292 videos from more than 40 diverse scenarios. To our best knowledge, BenSMOT is the first benchmark dedicated to SMOT. Furthermore, to encourage algorithm development on BenSMOT, we introduce SMOTer, an end-to-end tracker designed for SMOT. Our results exhibit SMOTer surpasses offline-combination strategies, showing efficacy. By presenting BenSMOT and SMOTer, we hope to inspire more future research on SMOT." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Acknowledgements. Heng Fan was not supported by any fund for this work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.415, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond MOT: Semantic Multi-Object Tracking" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Bai, H., Cheng, W., Chu, P., Liu, J., Zhang, K., Ling, H.: Gmot-40: A benchmark for generic multiple object tracking. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: ACL Workshop (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.259 + ], + "angle": 0, + "content": "3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. JIVP (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: ICIP (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.314 + ], + "angle": 0, + "content": "5. Cao, J., Pang, J., Weng, X., Khirodkar, R., Kitani, K.: Observation-centric sort: Rethinking sort for robust multi-object tracking. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.342 + ], + "angle": 0, + "content": "6. Chen, S., Shi, Z., Mettes, P., Snoek, C.G.: Social fabric: Tubelet compositions for video relation detection. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "7. Chu, P., Wang, J., You, Q., Ling, H., Liu, Z.: Transmot: Spatial-temporal graph transformer for multiple object tracking. In: WACV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "8. Cui, Y., Zeng, C., Zhao, X., Yang, Y., Wu, G., Wang, L.: Sportsmot: A large multi-object tracking dataset in multiple sports scenes. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.425 + ], + "angle": 0, + "content": "9. Dave, A., Khurana, T., Tokmakov, P., Schmid, C., Ramanan, D.: Tao: A large-scale benchmark for tracking any object. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.466 + ], + "angle": 0, + "content": "10. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J., Cremers, D., Reid, I., Roth, S., Schindler, K., Leal-Taixe, L.: Mot20: A benchmark for multi object tracking in crowded scenes. arXiv (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.535 + ], + "angle": 0, + "content": "2. Du, D., Qi, Y., Yu, H., Yang, Y., Duan, K., Li, G., Zhang, W., Huang, Q., Tian, Q.: The unmanned aerial vehicle benchmark: Object detection and tracking. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "3. Du, Y., Zhao, Z., Song, Y., Zhao, Y., Su, F., Gong, T., Meng, H.: Strongsort: Make deepsort great again. TMM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "4. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.653, + 0.605 + ], + "angle": 0, + "content": "5. Fellbaum, C.: WordNet: An electronic lexical database (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.785, + 0.633 + ], + "angle": 0, + "content": "6. Ferryman, J., Shahrokni, A.: Pets2009: Dataset and challenge. In: PET Workshop (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "7. Gao, R., Wang, L.: Memoritr: Long-term memory-augmented transformer for multi-object tracking. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.687 + ], + "angle": 0, + "content": "8. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.519, + 0.701 + ], + "angle": 0, + "content": "9. Girshick, R.: Fast r-cnn. In: ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.729 + ], + "angle": 0, + "content": "20. Han, X., Pasquier, T., Bates, A., Mickens, J., Seltzer, M.: Unicorn: Runtime provenance-based detector for advanced persistent threats. ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.756 + ], + "angle": 0, + "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "22. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "23. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "24. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv (2015)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "25. Liang, X., Lee, L., Xing, E.P.: Deep variation-structured reinforcement learning for visual relationship and attribute detection. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.189 + ], + "angle": 0, + "content": "26. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "27. Lin, K., Li, L., Lin, C.C., Ahmed, F., Gan, Z., Liu, Z., Lu, Y., Wang, L.: Swinbert: End-to-end transformers with sparse attention for video captioning. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "28. Liu, C., Jin, Y., Xu, K., Gong, G., Mu, Y.: Beyond short-term snippet: Video relation detection with spatio-temporal global context. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.285 + ], + "angle": 0, + "content": "29. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "30. Luiten, J., Osep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: Hota: A higher order metric for evaluating multi-object tracking. IJCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "31. Maggiolino, G., Ahmad, A., Cao, J., Kitani, K.: Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. ICIP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "32. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "33. Milan, A., Leal-Taixe, L., Reid, I., Roth, S., Schindler, K.: Mot16: A benchmark for multi-object tracking. arXiv (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "35. Ristani, E., Solera, F., Zou, R., Cucchiara, R., Tomasi, C.: Performance measures and a data set for multi-target, multi-camera tracking. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "36. Seo, P.H., Nagrani, A., Arnab, A., Schmid, C.: End-to-end generative pretraining for multimodal video captioning. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "37. Shen, Y., Gu, X., Xu, K., Fan, H., Wen, L., Zhang, L.: Accurate and fast compressed video captioning. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "38. Sun, P., Cao, J., Jiang, Y., Yuan, Z., Bai, S., Kitani, K., Luo, P.: Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "39. Sun, P., Cao, J., Jiang, Y., Zhang, R., Xie, E., Yuan, Z., Wang, C., Luo, P.: Transtrack: Multiple object tracking with transformer. arXiv (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "40. Taud, H., Mas, J.: Multilayer perceptron (mlp). Geomatic approaches for modeling land change scenarios (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "42. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.673 + ], + "angle": 0, + "content": "43. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: ICIP (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "44. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "45. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "46. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "47. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "48. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.415, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond MOT: Semantic Multi-Object Tracking" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "49. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.203 + ], + "angle": 0, + "content": "50. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.231 + ], + "angle": 0, + "content": "51. Zhang, H., Kyaw, Z., Chang, S.F., Chua, T.S.: Visual translation embedding network for visual relation detection. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.787, + 0.259 + ], + "angle": 0, + "content": "52. Zhang, L., Gao, J., Xiao, Z., Fan, H.: Animaltrack: A benchmark for multi-animal tracking in the wild. IJCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "53. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "54. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "55. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.787, + 0.383 + ], + "angle": 0, + "content": "56. Zheng, S., Chen, S., Jin, Q.: Vrderformer: End-to-end video visual relation detection with transformers. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.787, + 0.411 + ], + "angle": 0, + "content": "57. Zhou, L., Zhou, Y., Corso, J.J., Socher, R., Xiong, C.: End-to-end dense video captioning with masked transformer. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.439 + ], + "angle": 0, + "content": "58. Zhou, X., Arnab, A., Sun, C., Schmid, C.: Dense video object captioning from disjoint supervision. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.787, + 0.453 + ], + "angle": 0, + "content": "59. Zhou, X., Koltun, V., Krahenbihl, P.: Tracking objects as points. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.787, + 0.481 + ], + "angle": 0, + "content": "60. Zhou, X., Yin, T., Koltun, V., Krahenbuhl, P.: Global tracking transformers. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "61. Zhu, P., Wen, L., Du, D., Bian, X., Fan, H., Hu, Q., Ling, H.: Detection and tracking meet drones challenge. TPAMI (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.508 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Beyond MOT_ Semantic Multi-Object Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_origin.pdf b/2024/Beyond MOT_ Semantic Multi-Object Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..581cc43ee40d4f50ea2daa61003c3de03613b94d --- /dev/null +++ b/2024/Beyond MOT_ Semantic Multi-Object Tracking/8c5079ad-9f90-4a59-8b24-122466753ff4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67b61480f726b8841f724343eeaa025a69546f32badd0080bfeaaacf1a48c9ba +size 2339890 diff --git a/2024/Beyond MOT_ Semantic Multi-Object Tracking/full.md b/2024/Beyond MOT_ Semantic Multi-Object Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b4529f171d593d10568e016451ec47dac50d9a61 --- /dev/null +++ b/2024/Beyond MOT_ Semantic Multi-Object Tracking/full.md @@ -0,0 +1,300 @@ +# Beyond MOT: Semantic Multi-Object Tracking + +Yunhao Li $^{1,2}$ , Qin Li $^{1}$ , Hao Wang $^{2}$ , Xue Ma $^{1}$ , Jiali Yao $^{3}$ , Shaohua Dong $^{4}$ , Heng Fan $^{4\dagger}$ , and Libo Zhang $^{1,2,3\dagger\sharp}$ + +$^{1}$ Institute of Software Chinese Academy of Sciences + +$^{2}$ University of Chinese Academy of Sciences + +$^{3}$ Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences + +$^{4}$ Department of Computer Science & Engineering, University of North Texas + +Abstract. Current multi-object tracking (MOT) aims to predict trajectories of targets (i.e., "where") in videos. Yet, knowing merely "where" is insufficient in many crucial applications. In comparison, semantic understanding such as fine-grained behaviors, interactions, and overall summarized captions (i.e., "what") from videos, associated with "where", is highly-desired for comprehensive video analysis. Thus motivated, we introduce Semantic Multi-Object Tracking (SMOT), that aims to estimate object trajectories and meanwhile understand semantic details of associated trajectories including instance captions, instance interactions, and overall video captions, integrating "where" and "what" for tracking. In order to foster the exploration of SMOT, we propose BenSMOT, a largescale Benchmark for Semantic MOT. Specifically, BenSMOT comprises 3,292 videos with 151K frames, covering various scenarios for semantic tracking of humans. BenSMOT provides annotations for the trajectories of targets, along with associated instance captions in natural language, instance interactions, and overall caption for each video sequence. To our best knowledge, BenSMOT is the first publicly available benchmark for SMOT. Besides, to encourage future research, we present a novel tracker named SMOTer, which is specially designed and end-to-end trained for SMOT, showing promising performance. By releasing BenSMOT, we expect to go beyond conventional MOT by predicting "where" and "what" for SMOT, opening up a new direction in tracking for video understanding. We will release BenSMOT and SMOTer at https://github.com/Nathan-Li123/SMOTer. + +Keywords: Semantic Multi-Object Tracking (SMOT) $\cdot$ Benchmark + +# 1 Introduction + +Multi-object tracking (MOT) is a fundamental problem in computer vision with many applications such as autonomous driving, robotics, and visual surveillance. Current MOT tasks (e.g., [9,18,33,38,48]) usually focus on predicting trajectories of the targets from videos, i.e., answering the question of "where are the targets". Despite considerable advancements in deep learning era (e.g., [17,32,44,55,60]), + +![](images/01923049d8a6a11cba42968437f6b15463ca28a601ae1427c37695d86bcaa1d0.jpg) +Fig. 1: Existing multi-object tracking (MOT) focusing on predicting trajectories only (see (a)) and our semantic multi-object tracking (SMOT) aiming at estimating trajectories and understanding their semantics (see (b)). Best viewed in color for all figures. + +knowing merely "where" in existing MOT is insufficient for video understanding in many important applications. For example, besides trajectories, knowing more trajectory-associated semantic details of the objects, such as the behaviors and their interactions with the surroundings (i.e., "what"), is greatly beneficial for visual surveillance and robotics, which grows the interest in expanding modern MOT (i.e., addressing "where") with trajectory-associated semantic analysis (i.e., understanding "what") for more comprehensive video understanding. + +Thus motivated, in this work we introduce a new type of MOT task, dubbed Semantic Multi-Object Tracking (SMOT) $^5$ , which aims to expand scope of MOT beyond merely predicting trajectories to capture rich semantic details of objects in videos. Specifically, besides trajectory estimation, SMOT comprises three additional trajectory-associated semantic understanding tasks, including instance captioning, instance interaction recognition, and video captioning, as illustrated in Fig. 1. In particular, instance captioning aims at describing objects and their behaviors in human language, answering "what are the objects doing"; instance interaction recognition is to capture relations between objects, answering "what are the relations between objects"; video captioning provides the overall scenario understanding based on trajectories, answering "what is happening". Note that, all these three additional tasks are associated with object trajectories in videos. Ultimately, the goal of SMOT is to integrate "where" (i.e., object trajectories) + +and "what" (i.e., trajectory-associated semantics) for video understanding, going beyond MOT predicting only trajectories (see again Fig. 1). Compared to conventional MOT, our SMOT is a multimodal task across vision and language, which is more challenging yet practical. It is worthy to notice that, SMOT is a natural extension of MOT. Despite sharing some similarities with video captioning [23], SMOT differs in that it aims at language understanding of dense target trajectories from videos in the context of instance-level tracking. + +To foster study of SMOT, we propose BenSMOT, a large-scale Benchmark for Semantic MOT. Specifically, BenSMOT contains 3,292 video sequences with 151K frames, captured from more than 40 diverse daily-life scenarios for human-centric semantic tracking. In BenSMOT, there exist more than 7.8K instances, labeled with over 335K bounding boxes for trajectory prediction. For each instance, we provide fine-grained human language to describe the behaviors for instance captioning, resulting in 7.8K sentences. In addition, to capture interactions between objects, BenSMOT provides more than 14K interaction annotations from a rich collection of 335 kinds of interactions. Moreover, for each sequence, a summary text is provided for overall understanding of instances and backgrounds, leading to 3,292 video caption texts in total in BenSMOT. To ensure the high quality of BenSMOT, annotations are manually labeled with careful inspection and refinement. To our best knowledge, BenSMOT is the first publicly available benchmark dedicated to SMOT. By releasing BenSMOT, we expect it to serve as a platform for advancing the research on SMOT. + +Furthermore, in order to facilitate the development of SMOT algorithms on BenSMOT, we present a simple yet effective tracker dubbed SMOTer. Specifically, SMOTer is built upon a multi-object tracking model [53]. After generating object trajectories, we extract features for each trajectory and design additional prediction heads for semantic understanding of instance captioning, interaction recognition, and video captioning. Note, SMOTer is not a simple offline assembly of models for different tasks. Instead, it is specially designed for SMOT and trained in an end-to-end manner for predicting trajectories and understanding their semantics in videos. Despite simplicity, SMOTer shows promising performance for SMOT, and outperforms the offline-combination strategies, evidencing its effectiveness. We expect it to provide a reference for future research. + +We notice there exists a concurrent work combining multi-object tracking and trajectory captioning for dense video object captioning [58]. Compared with [58], our work differs in three aspects. First, task-wise, besides trajectory estimation and captioning, the SMOT in this work contains instance interaction and overall video captioning, providing more semantic details. Second, dataset-wise, we introduce a large-scale benchmark, BenSMOT, that is dedicated for SMOT and supports complete end-to-end model training, while the work of [58] borrows datasets from different tasks, which results in disjoint training of the model on different tasks and may thus degrade performance. Third, model-wise, owing to BenSMOT, we propose an end-to-end algorithm that demonstrates promising results on SMOT, while the approach in [58] is disjointly learned and does not support predicting instance interactions and overall video captioning. + +In summary, we make the following contributions: $\spadesuit$ We introduce semantic multi-object tracking (SMOT), a new tracking paradigm that expands existing MOT task by integrating "where" and "what"; $\spadesuit$ We present BenSMOT, a large-scale dataset with 3,292 videos and rich annotations for SMOT; $\spadesuit$ We propose SMOTer, a simple but effective tracker to facilitate future research of SMOT; $\spadesuit$ We show that SMOTer achieves promising performance for SMOT and conduct in-depth analysis on it to provide guidance for future algorithm design. + +# 2 Related Works + +MOT Benchmarks. Benchmarks have played a crucial role in facilitating the development of MOT. PETS2009 [16] is one of the earliest benchmarks for multipedestrian tracking. Later, the popular MOT Challenge [10, 24, 33] has been introduced with more crowded videos and greatly advanced MOT. KITTI [18] and BDD100K [48] are specifically designed for object tracking in autonomous driving. ImageNet-Vid [11] provides trajectory annotations for 30 categories across over 1,000 videos, and TAO [9] further expands object classes to 833 for generic multi-object tracking. To foster MOT in specific dancing and sport scenarios, DanceTrack [38] and SportsMOT [8] have been presented for dancer and player tracking. AnimalTrack [52] focuses on tracking various animals in wild scenes. Moreover, UAVDT [12] and VisDrone [61] provide platforms for tracking targets with drones. Different from the above MOT benchmarks for object trajectory prediction only ("where"), BenSMOT is specially developed for the new task of SMOT to serve as a platform for simultaneous estimation of target trajectories and trajectory-associated semantic understanding. To this end, it offers rich annotations of not only object trajectories as in existing standard MOT datasets ("where") but also semantic details of objects ("what") such as instance trajectory captions, interactions and overall video captions (see Fig. 1 again). + +MOT Algorithms. MOT algorithms have seen great progress in recent years. One of the popular paradigms for multi-object tracking is the so-called Tracking-by-Detection. This paradigm involves initial object detection followed by association based on these detections, forming the basis for many representative methods [5,13,21,31,43,53]. In this case, MOT methods typically improve their performance by enhancing both detection and matching effectiveness. In addition, another prevalent paradigm is the Joint-Tracking-and-Detection [20,44,45,47,59], which integrates tracking and detection steps into a single process, enabling end-to-end training. Recently, Transformer [41] has been introduced into MOT and exhibited remarkable improvements over previous tackers [7,17,32,39,50,55,60]. Our SMOTer is related to but different than existing MOT methods. Specifically, going beyond merely predicting object trajectories, SMOTer also aims at semantic understanding of trajectories, integrating “where” and “what”. + +Video Captioning. Video captioning is a multimodal vision-language task that aims to automatically describe the video content using natural language. Owing to its important applications in video event commentary and human-computer interaction, video captioning has drawn increasing attention in the past decade + +with many models proposed (e.g., [23, 27, 36, 37, 46, 57]). Similar to the video captioning task, SMOT utilizes natural language to describe instances and the overall video content. However, the difference is that our SMOT aims at language comprehension for dense target trajectories in the context of multi-object tracking, which is more challenging due to the requirement of accurate trajectories yet provides more fine-grained information for video understanding. + +Visual Relationship Detection. The goal of visual relationship detection task is to identify and comprehend the relationships (usually represented by a triplet of $<$ subject, predicate, object $\rangle$ ) between objects from an image (e.g., [25, 29, 51]). Besides static images, recent works extend visual relationship detection in the video domain (e.g., [6, 28, 56]). Our work is relevant to video visual relationship detection but different by extending this task target-specific trajectories, and thus more challenging due to complicated multi-object tracking scenarios. + +# 3 The Proposed BenSMOT + +# 3.1 Design Principle + +BenSMOT aims at providing a new platform for exploring human-centric SMOT. In construction of BenSMOT, we follow the following rules: + +(1) Dedicated benchmark. The key motivation of our BenSMOT is to provide a dedicated dataset for semantic multi-object tracking. In current deep learning era, a large number of videos are desired in benchmark construction for training robust tracking models. Considering this, we hope to build a dedicated dataset with more than 3,000 videos for human-centric semantic multi-object tracking. +(2) Diverse scenarios. For a dataset, diversity is crucial, in both training and evaluation, for developing general systems. In order to provide a diverse platform for SMOT, in BenSMOT we will include videos from more than 40 various scenes that range from daily life scenarios to dancing scenarios to sport scenarios. +(3) High-quality annotations. High-quality annotation is crucial for a benchmark in both training and assessing models. In BenSMOT, we ensure its high quality by providing manual annotations for each sequence including object trajectories, instance captions, instance interactions and overall video captions. In this process, multi-round inspections and refinements will be carried out. + +# 3.2 Data Acquisition + +BenSMOT focuses on predicting the target trajectories and meanwhile understanding their semantics by instance captioning, interaction recognition, and overall video captioning. To this end, the video sequences in BenSMOT are desired to involve diverse multi-person activities and interactions, which is different from videos in existing benchmarks to some extent. Specifically, we first identify 47 common scenarios, involving different activities ranging from daily-life talk and play to different sports, that are suitable for tracking by drawing inspiration from [22]. Due to limited space, please refer to supplementary material for + +Table 1: Summary of BenSMOT and its comparison with popular multi-object tracking benchmarks. "n/a" denotes that data is not available. + +
KITTI [18]MOT17 [33]MOT20 [10]BDD100k [48]TAO [9]GMOT-40 [1]DanceTrack [38]SportsMOT [8]BenSMOT (ours)
Videos501481,6002,907401002403,292
Min. length (s)n/a17.017.040.0n/a3.0n/an/a1.5
Avg. length (s)10.033.066.840.036.88.952.9n/a22.9
Max. length (s)n/a85.0133.040.0n/a24.2n/an/a116.0
Total length (s)498463535640,000106,9783565,292601575,499
Total tracks2,6001.3K3.83K131K17,2872,0269903,4017,792
Total boxes80K300K2,102K3,300K333K256Kn/a1,629K335K
Total frames15K11K13K318K2,674K9K106K150K151K
Instance CaptionsXXXXXXXX7,792
Instance InteractionsXXXXXXXX14K
Video SummariesXXXXXXXX3,292
+ +detailed scenarios in BenSMOT. Afterwards, we search for raw sequences for each scenario under the Creative Commons License from YouTube, the largest and most popular video platform with massive real-world videos. Please notice, for each raw video, instead of using the whole sequence, we usually choose a suitable clip for our semantic multi-object tracking task. + +Eventually, we compile a dataset for SMOT by including 3,292 videos with 151K frames. BenSMOT has an average video length of 23 seconds, with our current focus in this work not being on long-term SMOT. The total number of instance tracks is 7.8K, annotated with 335K bounding boxes for tracking. Besides bounding boxes, we provide 7.8K instance captions, 14K interactions, and 3,292 video summaries for trajectory-associated semantic understanding (how to annotate will be described in the following). To our best knowledge, BenSMOT by far is the first publicly available benchmark dedicated for SMOT. Tab. 1 summarizes our BenSMOT and compares it with popular MOT benchmarks. + +# 3.3 Data Annotation + +In order to meet the requirements of SMOT, BenSMOT provides four types of annotations, including bounding box, instance caption, instance interaction, and the overall video caption. Specifically, for each trajectory in videos, similar to standard MOT benchmarks (e.g., [9,33]), we provide axis-aligned bounding boxes to indicate its spatial positions in the videos. For instance captioning, we label each trajectory with a precise sentence in natural language to describe the detailed behavior of the associated target. For interaction recognition, we first collect a set of 335 interactions (i.e., verbs), among them 327 from WordNet [15]. Note that, in our interaction set, the same verb may have different meanings and thus is assigned to different interaction types, e.g., hold(v.01) for “being the physical support of someone” and hold(v.02) for “someone in one’s hands or arm”, for precise relationship description. Due to limited space, we refer reader to supplementary material for the full set of interaction types. After this, we provide an directed interaction label, represented with a triplet of $<$ subject predicate, object $\rightharpoondown$ (see Fig. 1 (b)), between two trajectories if associated objects + +![](images/41fc0948e3f3b4719a1566e7dc91e1fb6ce3c9eec3c194f410c27907877d7474.jpg) + +![](images/9959dbb8db30f6cbdaa45fcfd9201b5c977df662d35dc8d58a9036754777b6b1.jpg) +(d) Inst caption length distribution + +![](images/0f523d359a7ee4a387f43faac26e3cea9942a124cca07edc04411446f59c5627.jpg) + +![](images/2a1ca43adafa5b6fa864e9b4406e4218bf8a2152ef3966bf1f8ec84ec0fd1da3.jpg) +(e) Video caption length distribution + +![](images/cc21fb07e299f637ad7cda3e98e90058e4ed5c83144beb57f120cea72d724c61.jpg) + +![](images/4ad3a9d96dbaa1096c26e248d75b301f9542ee1fb0e52bfdd158877638e333ec.jpg) +(f) WordCloud of caption words +Fig. 2: Representative statistics on BenSMOT, including distributions of target trajectory length (in seconds) in (a), sequence length (in seconds) in (b), number of interactions per sequence in (c), instance caption length in (d), video caption length in (e), and wordcloud of all caption words with prepositions excluded in (f). + +are interacting. Finally, for the overall video captioning, we utilize a concise text summary to describe object trajectories in the video from a global perspective. + +To ensure the high-quality annotations, we adopt a multi-round mechanism. Specifically, each video will be first manually annotated by a few volunteers who are familiar with the task and an expert working on related problems. After this initial round, all annotations of the video will be sent to a validation team former by more than two experts for inspection. If initial annotations are not unanimously agreed by the experts, they will be returned to the original labeling team for refinement using the feedback from the validation team. We repeat this procedure for each video until the annotations of all video sequences in BenSMOT are completed. An annotation example in our BenSMOT is shown in Fig. 1 (b), and more can be found in supplementary material. + +Statistics of annotation. In order to better understand BenSMOT, we show some representative statics of annotations in Fig. 2. Specifically, we demonstrate the distributions of object trajectory length, video length, number of instance interactions, instance caption length, video caption length, and wordcloud. From Fig. 2 (c), it is worth noting that different targets in videos have frequent interactions. In addition, for the instance captions in Fig. 2 (d), we provide relatively longer textual descriptions, which allows precise understanding. + +# 3.4 Dataset Split and Evaluation Metric + +Training/Test Split. BenSMOT contains 3,292 videos, captured from 47 different scenarios. Within each scenario, we use $70\%$ of the sequences for training, and the rest $30\%$ for test. During dataset split, we try to keep the distributions of training and test sets similar. Eventfully, the training set of BenSMOT comprises 2,284 sequences with 104K frames, and the test set consists of 1,008 videos with 47K frames. Please see supplementary material for more split details. + +![](images/9e171706b4b13802126f70e12061a427159fe8b04983952bd097dce288d2a556.jpg) +Fig. 3: Illustration of SMOTer, which contains three components of trajectory estimation for tracking, feature fusion, and trajectory-associated semantic understanding. + +Evaluation Metric. We apply multiple metrics on BenSMOT for evaluation. Specifically, to assess the performance in object trajectory estimation, we employ higher order tracking accuracy (HOTA), association accuracy (AssA), detection accuracy (DetA), and localization accuracy (LocA) by following [30], CLEAR metrics [3] including multiple object tracking accuracy (MOTA), false positives (FP), false negatives (FN), and ID switches (IDs), and ID metrics [35] containing identification precision (IDP), identification recall (IDR) and related F1 score (IDF1). For instance and video captioning, we follow existing methods (e.g., [27, 37]) and employ BLEU [34], ROUGE [26], METEOR [2], and CIDEr [42] for thorough evaluation. Lastly, to assess performance of interaction recognition, we leverage classic metrics such as Precision (Prcn), Recall (Rcll), and F1, which are widely used in the context of classification tasks. + +# 4 Methodology: A New Baseline for SMOT + +Overview. To encourage development of SMOT algorithms, we propose a simple yet effective tracker on BenSMOT, dubbed SMOTer. As illustrated in Fig. 3, SMOTer logically consists of three key components. The first component, as detailed in Sec 4.1, is dedicated to object trajectory estimation for tracking, which is the basis for SMOT. Then, the second component, as explained in Sec 4.2, focuses on aggregating features from each frame into the overall video feature and meanwhile merging features of individual objects into target trajectory features using two distinct fusion modules, for subsequent semantic understanding. Finally, the third component, as described in Sec 4.3, aims to predict trajectory-associated semantic details in the video, including instance captions, instance interactions, and overall video caption. Please notice that, our SMOTer is specially designed for SMOT and trained in an end-to-end manner, as in Sec 4.4. + +# 4.1 Target Trajectory Estimation for Tracking + +Give a video with $N$ frames, SMOTer first uses a CNN backbone [49] to extract their features $\{f_i\}_{i=1}^N$ , where $f_i$ denotes the feature of the $i^{\text{th}}$ frame. Then, an initial set of proposals is generated using proposal generator of a popular detection architecture [14] with threshold filtering by $\tau_p$ . Afterwards, all reserved proposals are send to the association module (BYTE [53] is used in SMOTer for association) to obtain object trajectories. Finally, leveraging target coordinates within the trajectories and image features, target features are acquired for each target in trajectories through RoI Pooling [19]. Specifically, given feature $f_i$ of the $i^{\text{th}}$ frame and target coordinates $B_j^i$ corresponding to the $j^{\text{th}}$ object trajectory in this frame, we extract the corresponding target feature $t_j^i = \text{RoIPooling}(f_i, B_j^i)$ and apply it in subsequent feature fusion for semantic understanding. + +In summary, at this stage, SMOTer predicts object trajectories for tracking in a given video and extracts image and target features for subsequent stages. + +# 4.2 Feature Fusion + +Given image and target features from previous stage in Sec. 4.1, we design two feature fusion modules, including attention-based Video Fusion Module (VFM) and concatenation-based Trajectory Fusion Module (TFM). VFM aims to merge image-level features into overall video features while TFM focuses on integrating target features from each frame into target trajectory features. + +Video Fusion Module. VFM takes features $\{f_i\}_{i=1}^N$ of $N$ frames as input and sequentially feeds them into a cross-attention module for video feature fusion. In specific, the fused video feature $F_i$ up to the $i^{\text{th}}$ ( $i > 1$ ) frame is computed as + +$$ +F _ {i} = \mathbb {C A} \left(F _ {i - 1}, f _ {i}\right) \quad i > 1 \tag {1} +$$ + +where $\mathbf{CA}(\mathbf{z},\mathbf{u})$ denotes cross-attention with $\mathbf{z}$ generating query and $\mathbf{u}$ key/value as in [41]. For the first frame ( $i = 1$ ), $F_{1} = f_{1}$ . Finally, with the help of VFM, we can generate the overall fused video feature $\tilde{F} = F_{N}$ . + +Trajectory Fusion Module. Unlike VFM, TFM is to integrate multiple target features belonging to the same object trajectory, which are distributed across different frames, into a single trajectory feature. Concretely, given $t_i^j$ for object feature of trajectory $j$ in frame $i$ , we perform average pooling to yield a relatively simple 2D feature $\hat{t}_i^j = \mathsf{AvgPooling}(t_i^j)$ , where $\hat{t}_i^j \in \mathbb{R}^{4 \times d}$ with dimension $d = 256$ . Then, for the $j^{\mathrm{th}}$ object trajectory, we first concatenate its discrete features into an integrated one $\hat{T}_j = \mathsf{Concat}(\hat{t}_1^j, \hat{t}_2^j, \dots, \hat{t}_{N_j}^j)$ , where $N_j$ denotes trajectory length. After that, self-attention is used to enhance trajectory feature as follows + +$$ +\tilde {T} _ {j} = \mathbf {S A} (\hat {T} _ {j}) \tag {2} +$$ + +where $\mathbf{SA}(\mathbf{z})$ denotes self-attention with $\mathbf{z}$ generating query/key/value as in [41], and $\bar{T}_j$ is the enhanced feature for trajectory $j$ . + +# 4.3 Semantic Understanding + +In the third phase, SMOTer applies the fused video feature $\tilde{F}$ and trajectory features $\tilde{T} = \{\tilde{T}_1,\tilde{T}_2,\dots,\tilde{T}_M\}$ , where $M$ denotes the number of trajectories in the video, to tackle trajectory-associated semantic understanding tasks in SMOT, i.e., video captioning, instance captioning, and interaction recognition. + +Video captioning. Unlike current video captioning, in SMOTer we incorporate trajectory features into describing video content and allow more comprehensive human-centric caption generation. Specifically, SMOTer sequentially injects trajectory features $\tilde{T}$ into video feature $\tilde{F}$ with cross-attention and then predicts captioning result $R_{\mathrm{Vid}}$ via a linear project layer and a text decoder, as follows + +$$ +R _ {\mathrm {V i d}} = \operatorname {D e c} _ {\mathrm {V i d}} (\operatorname {L i n e a r} (\mathbf {C A} (\tilde {F}, \tilde {T}))) \tag {3} +$$ + +where $\mathsf{Dec}_{\mathrm{Vid}}(\cdot)$ is a video text decoder for video captioning (please kindly refer to supplementary material for details), and $\mathbf{Linear}(\cdot)$ is a linear projection. Instance captioning. Different from video captioning, for instance captioning, we directly generate the results based on the target trajectory features. Specifically, given the feature $\tilde{T}_j$ for the $j^{\mathrm{th}}$ ( $1 \leq j \leq M$ ) trajectory, we apply a linear projection layer followed by the designed text decoder to predict the captioning result $R_{\mathrm{Ins}}^j$ of each trajectory, depicted as follows, + +$$ +R _ {\text {I n s}} ^ {j} = \operatorname {D e c} _ {\text {I n s}} \left(\operatorname {L i n e a r} \left(\tilde {T} _ {j}\right)\right) \quad (1 \leq j \leq M) \tag {4} +$$ + +where $\mathsf{Dec}_{\mathrm{Ins}}(\cdot)$ is instance text decoder [46] for instance captioning. Please refer to supplementary material for more details because of space limitation. + +Interaction recognition. Besides captioning, SMOTer predicts the interaction between trajectories, which is crucial for video understanding. Particularly, in SMOTer we concentrate solely on interactions between two trajectories, as our current focus doesn't extend to understanding interactions involving three or more targets, leaving it for future research. More specifically, for two trajectories $j$ and $k$ , we assume the $j^{th}$ trajectory in a video sequence is the active trajectory and the $k^{th}$ trajectory the passive trajectory. Then, we first fuse the features of active and passive trajectories using cross-attention and then predict the interaction result $R_{\mathrm{Int}}^{j,k}$ using multi-layer perception [40], as follows, + +$$ +R _ {\text {I n t}} ^ {j, k} = \operatorname {M L P} (\mathrm {C A} (\tilde {T} _ {j}, \tilde {T} _ {k})), \tag {5} +$$ + +where $\mathsf{MLP}(\cdot)$ denotes the multi-layer perception. Please note, interchanging the active and passive trajectories may yield disparate outcomes, i.e., $R_{\mathrm{Int}}^{j,k}$ and $R_{\mathrm{Int}}^{k,j}$ may exhibit dissimilar characteristics, as indicated by the formula above. + +# 4.4 End-to-end Training + +End-to-end training of SMOTer is not easy due to conflict between two types of training requirements, i.e., the trajectory estimation is usually trained frame-by-frame, while the captioning tasks, and interaction recognition are often trained + +at the entire video level, resulting in two different losses in training. In addition, due to the frame-by-frame training nature of trajectory estimation for tracking, the complete sequence is hard to be input at once during training, preventing the model from obtaining full trajectories thus hindering subsequent caption generation and interaction prediction. To address these issues, we draw inspiration from [60] and adopt a strategy of completing full trajectory association after detection on the entire video. Specifically, we first compute detection loss after each frame input and then perform trajectory association after the entire sequence is detected. Subsequently, with complete trajectories, we compute losses for trajectory-associated instance captioning, interaction recognition, and video captioning tasks, realizing end-to-end training of SMOTer. For the detailed training loss of SMOTer, please kindly refer to the supplementary material. + +# 5 Experiments + +Experimental setup. We conduct experiments with 4 Nvidia Tesla V100 GPUs (32GB). The batch size is 1 per GPU. Each batch has a video clip with multiple frames. We use the AdamW optimizer with an initial learning rate of $5.0 \times 10^{-4}$ . During training, we filter out proposals with scores lower than the threshold $\tau_{p} = 0.3$ . For lost tracklets, we retain them for 30 frames in case they reappear. Due to memory constraints and limited accuracy of trajectory estimation in early training stages, we sample 6-8 frames per sequence for training detection model and only compute tracking related losses in the first 20K training iterations. + +As SMOT is new and no suitable methods exist for comparison with SMOTer, we logically divide SMOT into two components: trajectory estimation for tracking and semantic understanding (including instance/video captioning, and interaction recognition). Subsequently, we devise a series of two-stage models based on existing MOT frameworks. Leveraging several state-of-the-art and classic MOT models such as SORT [4], DeepSORT [43], OC-SORT [5], ByteTrack [53], TransTrack [39], MOTR [50], and MOTRv2 [55] as foundations, we integrate feature fusion and semantic generation functionalities akin to SMOTer and conduct comparative experiments on BenSMOT. To ensure fair experiments, we employ the same training framework for comparative experiments of two-stage approaches. Additionally, for models based on two-stage MOT methods like SORT, we use the same detector [14] as SMOTer during training and evaluation. + +# 5.1 Comparison on Trajectory Estimation for Tracking. + +We first compare SMOTer with other MOT methods for trajectory estimation on BenSMOT. Specifically, SMOTer is fully end-to-end trained, while other two-stage MOT models are trained with their own strategies, concentrating solely on tracking. As depicted in Tab. 2, despite being at a certain degree of natural disadvantage, our SMOTer achieves comparable or even superior performance in tracking compared to state-of-the-art MOT models. SMOTer consistently + +Table 2: Comparison between SMOTer and two-stage MOT methods regarding tracking performance on BenSMOT. The best two results are shown in red and blue fonts. + +
MethodHOTA↑AssA↑DetA↑LocA↑MOTA↑FN↓FP↓IDs↓IDR↑IDP↑IDF1↑
SORT [4]48.4938.9560.9187.5053.582400151051387560.8548.4353.93
DeepSORT [43]50.1240.2361.4587.6754.292289055401127862.1051.1156.76
OC-SORT [5]51.0041.4263.3187.6155.192106153881504963.9253.1058.01
ByteTrack [53]68.8471.1567.1085.1573.87154197070171282.2574.8378.37
TransTrack [39]71.3173.3469.6791.3174.08201244420253085.6372.7578.67
MOTR [50]66.1073.1255.1486.3045.19312971117861772.3970.1268.97
MOTRv2 [55]65.2876.8251.3086.0945.52407652092343078.4765.5170.76
SMOTer (ours)71.9873.7170.7987.1177.71125346388170283.8277.9780.65
+ +Table 3: Comparison of SMOTer against two-stage methods based on MOT models regarding semantic understanding. The best two results are highlighted in red and blue. + +
Video CaptionInstance CaptionInteraction
MethodBLEU↑ROUGE↑METEOR↑CIDEr↑BLEU↑ROUGE↑METEOR↑CIDEr↑Prec↑Rcll↑F1↑
SORT [4]0.2450.2240.2020.2980.2330.2450.2080.0560.3630.2590.302
DeepSORT [43]0.1980.2130.1870.3090.2380.2120.1990.0650.3650.2770.310
OC-SORT [5]0.2310.2520.2150.2420.2700.2050.1800.0330.3840.2910.331
ByteTrack [53]0.2240.2250.2120.2660.3040.2420.2240.0640.4430.2580.326
TransTrack [39]0.2470.2480.2090.2690.2830.2190.2010.0740.4060.3110.376
MOTR [50]0.1870.2540.2030.2440.2300.2090.1820.0610.4250.3140.354
MOTRv2 [55]0.2170.2580.2190.2480.2380.2410.2040.0590.3130.3950.349
SMOTer (ours)0.2450.2610.2230.3430.3060.2230.2090.0870.4340.3200.368
+ +Table 4: Ablation experiments for evaluating various feature fusion strategies using results of three additional tasks. The best result is highlighted in red. + +
VFM (video fusion)TFM (trajectory fusion)
BLEU↑ROUGE↑METEOR↑CIDEr↑BLEU↑ROUGE↑METEOR↑CIDEr↑Pcn↑Rcll↑F1↑
Attention-based0.2450.2610.2230.3430.3030.2180.2080.0720.4060.3470.364
MLP-based0.2510.2230.2000.2200.3370.2190.2050.0720.3800.2520.303
Concatenation-based0.2300.2130.2150.3650.3060.2230.2090.0870.4340.3200.368
Addition-based0.1400.1570.1040.1270.1820.1370.1140.0410.2280.1630.205
+ +achieves top-tier performance across nearly all metrics, demonstrating impressive results in key metrics. For instance, it achieves a notable $71.98\%$ in HOTA, $77.71\%$ in MOTA, and $80.65\%$ in IDF1. This suggests that subsequent training tasks do not degrade tracking performance, validating effectiveness of our end-to-end training strategy. In fact, by comparing with the standard ByteTrack, we observe that although SMOTer adopts a highly similar structure to ByteTrack in tracking, it achieves superior performance, showing $+3.15\%$ in HOTA and $+3.74\%$ in MOTA. Moreover, even in metrics where it does not excel, such as FP and LocA, SMOTer exhibits significant improvements compared to ByteTrack, implying that semantic understanding tasks can reciprocally aid in tracking. + +# 5.2 Comparison on Semantic Understanding + +Besides tracking, we compare SMOTer with a set of two-stage models on their semantic understanding capabilities. Similarly, SMOTer undergoes complete end-to-end training. However, other two-stage models first train trackers on their own to acquire target trajectories and then use these trajectories to separately train the semantic understanding tasks (with the same feature fusion as SMOTer). + +![](images/2b4035949ae0bcf10d5fa8bf6ce10f71f8ba56bfc486d6cb00d78d7addd1c7b3.jpg) +(a) Comparison results on object tracking + +![](images/eec43bb00846fecfcfb1c050803386a7698d65e8f4848611f014fe75ff238dcb.jpg) +Fig. 4: Ablation experiments for different association strategies, including comparison results on object tracking in image (a), on video captioning in image (b), on instance captioning in image (c), and on interaction recognition in image (d). + +![](images/7338f1c6ec8beb8ee6e3521a509bc3f6f29779d9e894a4e925cdc13666f3023a.jpg) +(b) Comparison results on video captioning + +![](images/c9a35591f8b1de1674dfbf715070f47fabaf6320a0bbcbf274d23326188c0c7a.jpg) +(c) Comparison results on instance captioning +(d) Comparison results on interaction recognition + +From Tab. 3, we can see SMOTer achieves highly favorable results on almost all metrics. Particularly, compared to the baseline method ByteTrack, SMOTer surpasses it on most metrics, including $+2.1\%$ BLEU and $+7.7\%$ CIDEr on video captioning, $+2.3\%$ CIDEr on instance captioning, and $+4.2\%$ F1 score on interaction recognition. More results can be seen in supplementary material. + +# 5.3 Ablation Study + +Impact of different feature fusion. Feature fusion is a crucial component of SMOTer for generating overall video and enhanced trajectory features. In this work, we study four types of feature fusion strategies for VFM and TFM, including attention-based fusion (using cross-attention), MLP-based fusion (using MLP module), concatenation-based fusion, and addition-based fusion. Please refer to supplementary material for detailed architectures. To assess the impact for VFM, we compare the performance on video caption generation because the fused video feature from VFM is used for this sub-task, while for TFM, we measure the results on instance captioning and interaction recognition since these two sub-tasks rely on the fused trajectory features. Tab. 4 shows the results of different fusion mechanisms for VFM and TFM. We can see that attention-based fusion works generally better for VFM by achieving best ROUGE score of 0.261 and METEOR score of 0.223 and second best BLEU score of 0.245 and CIDEr score of 0.343 for video captioning. For TFM, concatenation-based fusion shows more superior results by obtaining best scores on ROUGE, METEOR, CIDEr for instance captioning and on Precision and F1 for interaction recognition. + +Analysis on association mechanism. Association is necessary in SMOTer to generate target trajectories. In our method, we leverage the popular BYTE [54] for proposal association. To analyze the impact of different association strategies, we compare other representative manners, including SORT [4], DeepSORT [43], and OC-SORT [5], with the adopted BYTE. Fig. 4 demonstrates the comparison results of association mechanisms in SMOTer (using BYTE) and other approaches on four tasks, including object tracking (Fig. 4 (a)), video captioning (Fig. 4 (b)), instance captioning (Fig. 4 (c)), and interaction recognition (Fig. 4 (d)). From the results in Fig. 4 (a)-(d), we can observe that BYTE used in our SMOTer can achieve the best or competitive performance on different metrics for various tasks, validating its effectiveness for semantic multi-object tracking. + +Table 5: Study of different thresholds ${\tau }_{p}$ . The best result is highlighted in red. + +
TrackingVideo CaptionInstance CaptionInteraction
τpHOTA↑MOTA↑IDF1↑METEOR↑CIDEr↑METEOR↑CIDEr↑Prcn↑Rcll↑F1↑
0.170.5172.3078.800.2190.2560.1920.0690.3780.3780.319
0.271.1073.8078.900.2150.2890.1990.0690.3840.3650.331
0.371.9877.7180.650.2240.3430.2090.0870.4340.3200.368
0.471.1976.3676.010.2140.3300.2150.0910.4200.3020.341
0.568.3074.3478.900.2240.3430.2160.0870.3960.2910.310
+ +Study of threshold $\tau_{p}$ . The parameter $\tau_{p}$ is used to remove low-confidence proposals to ensure the quality of subsequent trajectories, and thus crucial for all tasks in SMOTer. Because of this, we conduct an ablation study for investigating different $\tau_{p}$ , and Tab. 5 displays the results. From Tab. 5, we can observe that, when $\tau_{p} = 0.3$ , SMOTer achieves the overall best performance for different tasks. + +# 5.4 Discussion + +Semantic understanding for improving tracking. SMOT aims to go beyond just tracking by incorporating additional semantic understanding tasks including trajectory-associated instance/video captioning and interaction recognition. We observe that these extra goals greatly improve the tracking performance. For example, as shown in Tab. 2, we can see that joint training of different tasks leads to better performance of SMOTer (71.98%/77.71% in HOTA/MOTA) in tracking compared to its baseline ByteTrack (68.84%/73.87 in HOTA/MOTA). The possible reason is that SMOTer needs to comprehensively understand the video content for semantic multi-object tracking, and the learned semantic knowledge helps distinguish different object trajectories. + +Challenge in SMOT. One reason why SMOT is challenging lies in the complicated behaviors of targets, often leading to lengthy instance captions. For instance, a basketball player can be dribbling, shooting, and defending in a game. Models often miss some states in understanding. In BenSMOT, the average video length is about 23 seconds, with objects typically present for around 21 seconds, resulting in instance captions averaging over 35 words. This complexity poses a challenge for SMOT and indicates a potential area for future improvement. + +# 6 Conclusion + +In this paper, we introduce SMOT to expand the scope of MOT by integrating "where" and "what". To facilitate the research of SMOT, we propose the large-scale BenSMOT by including 3,292 videos from more than 40 diverse scenarios. To our best knowledge, BenSMOT is the first benchmark dedicated to SMOT. Furthermore, to encourage algorithm development on BenSMOT, we introduce SMOTer, an end-to-end tracker designed for SMOT. Our results exhibit SMOTer surpasses offline-combination strategies, showing efficacy. By presenting BenSMOT and SMOTer, we hope to inspire more future research on SMOT. + +Acknowledgements. Heng Fan was not supported by any fund for this work. + +# References + +1. Bai, H., Cheng, W., Chu, P., Liu, J., Zhang, K., Ling, H.: Gmot-40: A benchmark for generic multiple object tracking. In: CVPR (2021) +2. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: ACL Workshop (2005) +3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. JIVP (2008) +4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: ICIP (2016) +5. Cao, J., Pang, J., Weng, X., Khirodkar, R., Kitani, K.: Observation-centric sort: Rethinking sort for robust multi-object tracking. In: CVPR (2023) +6. Chen, S., Shi, Z., Mettes, P., Snoek, C.G.: Social fabric: Tubelet compositions for video relation detection. In: ICCV (2021) +7. Chu, P., Wang, J., You, Q., Ling, H., Liu, Z.: Transmot: Spatial-temporal graph transformer for multiple object tracking. In: WACV (2023) +8. Cui, Y., Zeng, C., Zhao, X., Yang, Y., Wu, G., Wang, L.: Sportsmot: A large multi-object tracking dataset in multiple sports scenes. ICCV (2023) +9. Dave, A., Khurana, T., Tokmakov, P., Schmid, C., Ramanan, D.: Tao: A large-scale benchmark for tracking any object. In: ECCV (2020) +10. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J., Cremers, D., Reid, I., Roth, S., Schindler, K., Leal-Taixe, L.: Mot20: A benchmark for multi object tracking in crowded scenes. arXiv (2020) +1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) +2. Du, D., Qi, Y., Yu, H., Yang, Y., Duan, K., Li, G., Zhang, W., Huang, Q., Tian, Q.: The unmanned aerial vehicle benchmark: Object detection and tracking. In: ECCV (2018) +3. Du, Y., Zhao, Z., Song, Y., Zhao, Y., Su, F., Gong, T., Meng, H.: Strongsort: Make deepsort great again. TMM (2023) +4. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: ICCV (2019) +5. Fellbaum, C.: WordNet: An electronic lexical database (1998) +6. Ferryman, J., Shahrokni, A.: Pets2009: Dataset and challenge. In: PET Workshop (2009) +7. Gao, R., Wang, L.: Memoritr: Long-term memory-augmented transformer for multi-object tracking. In: ICCV (2023) +8. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset (2013) +9. Girshick, R.: Fast r-cnn. In: ICCV (2015) +20. Han, X., Pasquier, T., Bates, A., Mickens, J., Seltzer, M.: Unicorn: Runtime provenance-based detector for advanced persistent threats. ECCV (2020) +21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +22. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015) +23. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV (2017) +24. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv (2015) + +25. Liang, X., Lee, L., Xing, E.P.: Deep variation-structured reinforcement learning for visual relationship and attribute detection. In: CVPR (2017) +26. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL (2004) +27. Lin, K., Li, L., Lin, C.C., Ahmed, F., Gan, Z., Liu, Z., Lu, Y., Wang, L.: Swinbert: End-to-end transformers with sparse attention for video captioning. In: CVPR (2022) +28. Liu, C., Jin, Y., Xu, K., Gong, G., Mu, Y.: Beyond short-term snippet: Video relation detection with spatio-temporal global context. In: CVPR (2020) +29. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors. In: ECCV (2016) +30. Luiten, J., Osep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: Hota: A higher order metric for evaluating multi-object tracking. IJCV (2021) +31. Maggiolino, G., Ahmad, A., Cao, J., Kitani, K.: Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. ICIP (2023) +32. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: CVPR (2022) +33. Milan, A., Leal-Taixe, L., Reid, I., Roth, S., Schindler, K.: Mot16: A benchmark for multi-object tracking. arXiv (2016) +34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002) +35. Ristani, E., Solera, F., Zou, R., Cucchiara, R., Tomasi, C.: Performance measures and a data set for multi-target, multi-camera tracking. In: ECCV (2016) +36. Seo, P.H., Nagrani, A., Arnab, A., Schmid, C.: End-to-end generative pretraining for multimodal video captioning. In: CVPR (2022) +37. Shen, Y., Gu, X., Xu, K., Fan, H., Wen, L., Zhang, L.: Accurate and fast compressed video captioning. In: ICCV (2023) +38. Sun, P., Cao, J., Jiang, Y., Yuan, Z., Bai, S., Kitani, K., Luo, P.: Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In: CVPR (2022) +39. Sun, P., Cao, J., Jiang, Y., Zhang, R., Xie, E., Yuan, Z., Wang, C., Luo, P.: Transtrack: Multiple object tracking with transformer. arXiv (2020) +40. Taud, H., Mas, J.: Multilayer perceptron (mlp). Geomatic approaches for modeling land change scenarios (2018) +41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NIPS (2017) +42. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015) +43. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: ICIP (2017) +44. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: ECCV (2022) +45. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: CVPR (2023) +46. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023) +47. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: ECCV (2022) +48. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: CVPR (2020) + +49. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: CVPR (2018) +50. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: ECCV (2022) +51. Zhang, H., Kyaw, Z., Chang, S.F., Chua, T.S.: Visual translation embedding network for visual relation detection. In: CVPR (2017) +52. Zhang, L., Gao, J., Xiao, Z., Fan, H.: Animaltrack: A benchmark for multi-animal tracking in the wild. IJCV (2023) +53. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: ECCV (2022) +54. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV (2021) +55. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In: CVPR (2023) +56. Zheng, S., Chen, S., Jin, Q.: Vrderformer: End-to-end video visual relation detection with transformers. In: CVPR (2022) +57. Zhou, L., Zhou, Y., Corso, J.J., Socher, R., Xiong, C.: End-to-end dense video captioning with masked transformer. In: CVPR (2018) +58. Zhou, X., Arnab, A., Sun, C., Schmid, C.: Dense video object captioning from disjoint supervision. arXiv (2023) +59. Zhou, X., Koltun, V., Krahenbihl, P.: Tracking objects as points. In: ECCV (2020) +60. Zhou, X., Yin, T., Koltun, V., Krahenbuhl, P.: Global tracking transformers. In: CVPR (2022) +61. Zhu, P., Wen, L., Du, D., Bian, X., Fan, H., Hu, Q., Ling, H.: Detection and tracking meet drones challenge. TPAMI (2021) \ No newline at end of file diff --git a/2024/Beyond MOT_ Semantic Multi-Object Tracking/images.zip b/2024/Beyond MOT_ Semantic Multi-Object Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7f07f046707eb0cdab9a7322c47f824e42333b5b --- /dev/null +++ b/2024/Beyond MOT_ Semantic Multi-Object Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13a0d1548c8eef190484b8bf8db46ed1eacaf4fedbfe258a37f39060248e2c51 +size 526028 diff --git a/2024/Beyond MOT_ Semantic Multi-Object Tracking/layout.json b/2024/Beyond MOT_ Semantic Multi-Object Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cab82849dceff878d3068d858086094c90fb2dbd --- /dev/null +++ b/2024/Beyond MOT_ Semantic Multi-Object Tracking/layout.json @@ -0,0 +1,9438 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 131 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 131 + ], + "type": "text", + "content": "Beyond MOT: Semantic Multi-Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "spans": [ + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "text", + "content": "Yunhao Li" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "text", + "content": ", Qin Li" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "text", + "content": ", Hao Wang" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "text", + "content": ", Xue Ma" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "text", + "content": ", Jiali Yao" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "text", + "content": ", Shaohua Dong" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "text", + "content": ", Heng Fan" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "text", + "content": ", and Libo Zhang" + }, + { + "bbox": [ + 144, + 149, + 470, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger\\sharp}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 183, + 414, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 183, + 414, + 194 + ], + "spans": [ + { + "bbox": [ + 198, + 183, + 414, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 198, + 183, + 414, + 194 + ], + "type": "text", + "content": " Institute of Software Chinese Academy of Sciences" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 195, + 399, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 195, + 399, + 205 + ], + "spans": [ + { + "bbox": [ + 214, + 195, + 399, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 214, + 195, + 399, + 205 + ], + "type": "text", + "content": " University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 205, + 481, + 217 + ], + "type": "text", + "content": " Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 217, + 465, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 217, + 465, + 228 + ], + "spans": [ + { + "bbox": [ + 148, + 217, + 465, + 228 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 148, + 217, + 465, + 228 + ], + "type": "text", + "content": " Department of Computer Science & Engineering, University of North Texas" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 256, + 455, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 256, + 455, + 520 + ], + "spans": [ + { + "bbox": [ + 159, + 256, + 455, + 520 + ], + "type": "text", + "content": "Abstract. Current multi-object tracking (MOT) aims to predict trajectories of targets (i.e., \"where\") in videos. Yet, knowing merely \"where\" is insufficient in many crucial applications. In comparison, semantic understanding such as fine-grained behaviors, interactions, and overall summarized captions (i.e., \"what\") from videos, associated with \"where\", is highly-desired for comprehensive video analysis. Thus motivated, we introduce Semantic Multi-Object Tracking (SMOT), that aims to estimate object trajectories and meanwhile understand semantic details of associated trajectories including instance captions, instance interactions, and overall video captions, integrating \"where\" and \"what\" for tracking. In order to foster the exploration of SMOT, we propose BenSMOT, a largescale Benchmark for Semantic MOT. Specifically, BenSMOT comprises 3,292 videos with 151K frames, covering various scenarios for semantic tracking of humans. BenSMOT provides annotations for the trajectories of targets, along with associated instance captions in natural language, instance interactions, and overall caption for each video sequence. To our best knowledge, BenSMOT is the first publicly available benchmark for SMOT. Besides, to encourage future research, we present a novel tracker named SMOTer, which is specially designed and end-to-end trained for SMOT, showing promising performance. By releasing BenSMOT, we expect to go beyond conventional MOT by predicting \"where\" and \"what\" for SMOT, opening up a new direction in tracking for video understanding. We will release BenSMOT and SMOTer at https://github.com/Nathan-Li123/SMOTer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 530, + 441, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 530, + 441, + 542 + ], + "spans": [ + { + "bbox": [ + 160, + 530, + 441, + 542 + ], + "type": "text", + "content": "Keywords: Semantic Multi-Object Tracking (SMOT) " + }, + { + "bbox": [ + 160, + 530, + 441, + 542 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 530, + 441, + 542 + ], + "type": "text", + "content": " Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 562, + 230, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 230, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 230, + 574 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Multi-object tracking (MOT) is a fundamental problem in computer vision with many applications such as autonomous driving, robotics, and visual surveillance. Current MOT tasks (e.g., [9,18,33,38,48]) usually focus on predicting trajectories of the targets from videos, i.e., answering the question of \"where are the targets\". Despite considerable advancements in deep learning era (e.g., [17,32,44,55,60])," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 459, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 459, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 459, + 666 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 141, + 653, + 459, + 666 + ], + "type": "text", + "content": "Equal advising and co-last author " + }, + { + "bbox": [ + 141, + 653, + 459, + 666 + ], + "type": "inline_equation", + "content": "\\sharp" + }, + { + "bbox": [ + 141, + 653, + 459, + 666 + ], + "type": "text", + "content": "Corresponding author (libo@iscas.ac.cn)" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 473, + 330 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 473, + 330 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 473, + 330 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 473, + 330 + ], + "type": "image", + "image_path": "01923049d8a6a11cba42968437f6b15463ca28a601ae1427c37695d86bcaa1d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 338, + 482, + 373 + ], + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 373 + ], + "type": "text", + "content": "Fig. 1: Existing multi-object tracking (MOT) focusing on predicting trajectories only (see (a)) and our semantic multi-object tracking (SMOT) aiming at estimating trajectories and understanding their semantics (see (b)). Best viewed in color for all figures." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 396, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 479 + ], + "type": "text", + "content": "knowing merely \"where\" in existing MOT is insufficient for video understanding in many important applications. For example, besides trajectories, knowing more trajectory-associated semantic details of the objects, such as the behaviors and their interactions with the surroundings (i.e., \"what\"), is greatly beneficial for visual surveillance and robotics, which grows the interest in expanding modern MOT (i.e., addressing \"where\") with trajectory-associated semantic analysis (i.e., understanding \"what\") for more comprehensive video understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 479, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 635 + ], + "type": "text", + "content": "Thus motivated, in this work we introduce a new type of MOT task, dubbed Semantic Multi-Object Tracking (SMOT)" + }, + { + "bbox": [ + 130, + 479, + 482, + 635 + ], + "type": "inline_equation", + "content": "^5" + }, + { + "bbox": [ + 130, + 479, + 482, + 635 + ], + "type": "text", + "content": ", which aims to expand scope of MOT beyond merely predicting trajectories to capture rich semantic details of objects in videos. Specifically, besides trajectory estimation, SMOT comprises three additional trajectory-associated semantic understanding tasks, including instance captioning, instance interaction recognition, and video captioning, as illustrated in Fig. 1. In particular, instance captioning aims at describing objects and their behaviors in human language, answering \"what are the objects doing\"; instance interaction recognition is to capture relations between objects, answering \"what are the relations between objects\"; video captioning provides the overall scenario understanding based on trajectories, answering \"what is happening\". Note that, all these three additional tasks are associated with object trajectories in videos. Ultimately, the goal of SMOT is to integrate \"where\" (i.e., object trajectories)" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "Here by \"semantic\", we emphasize high-level trajectory-based activity understanding in videos in the context of tracking, instead of category as in semantic segmentation." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 481, + 200 + ], + "type": "text", + "content": "and \"what\" (i.e., trajectory-associated semantics) for video understanding, going beyond MOT predicting only trajectories (see again Fig. 1). Compared to conventional MOT, our SMOT is a multimodal task across vision and language, which is more challenging yet practical. It is worthy to notice that, SMOT is a natural extension of MOT. Despite sharing some similarities with video captioning [23], SMOT differs in that it aims at language understanding of dense target trajectories from videos in the context of instance-level tracking." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 204, + 481, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 204, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 134, + 204, + 481, + 383 + ], + "type": "text", + "content": "To foster study of SMOT, we propose BenSMOT, a large-scale Benchmark for Semantic MOT. Specifically, BenSMOT contains 3,292 video sequences with 151K frames, captured from more than 40 diverse daily-life scenarios for human-centric semantic tracking. In BenSMOT, there exist more than 7.8K instances, labeled with over 335K bounding boxes for trajectory prediction. For each instance, we provide fine-grained human language to describe the behaviors for instance captioning, resulting in 7.8K sentences. In addition, to capture interactions between objects, BenSMOT provides more than 14K interaction annotations from a rich collection of 335 kinds of interactions. Moreover, for each sequence, a summary text is provided for overall understanding of instances and backgrounds, leading to 3,292 video caption texts in total in BenSMOT. To ensure the high quality of BenSMOT, annotations are manually labeled with careful inspection and refinement. To our best knowledge, BenSMOT is the first publicly available benchmark dedicated to SMOT. By releasing BenSMOT, we expect it to serve as a platform for advancing the research on SMOT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 387, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 387, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 134, + 387, + 481, + 518 + ], + "type": "text", + "content": "Furthermore, in order to facilitate the development of SMOT algorithms on BenSMOT, we present a simple yet effective tracker dubbed SMOTer. Specifically, SMOTer is built upon a multi-object tracking model [53]. After generating object trajectories, we extract features for each trajectory and design additional prediction heads for semantic understanding of instance captioning, interaction recognition, and video captioning. Note, SMOTer is not a simple offline assembly of models for different tasks. Instead, it is specially designed for SMOT and trained in an end-to-end manner for predicting trajectories and understanding their semantics in videos. Despite simplicity, SMOTer shows promising performance for SMOT, and outperforms the offline-combination strategies, evidencing its effectiveness. We expect it to provide a reference for future research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 481, + 665 + ], + "type": "text", + "content": "We notice there exists a concurrent work combining multi-object tracking and trajectory captioning for dense video object captioning [58]. Compared with [58], our work differs in three aspects. First, task-wise, besides trajectory estimation and captioning, the SMOT in this work contains instance interaction and overall video captioning, providing more semantic details. Second, dataset-wise, we introduce a large-scale benchmark, BenSMOT, that is dedicated for SMOT and supports complete end-to-end model training, while the work of [58] borrows datasets from different tasks, which results in disjoint training of the model on different tasks and may thus degrade performance. Third, model-wise, owing to BenSMOT, we propose an end-to-end algorithm that demonstrates promising results on SMOT, while the approach in [58] is disjointly learned and does not support predicting instance interactions and overall video captioning." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond MOT: Semantic Multi-Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "In summary, we make the following contributions: " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " We introduce semantic multi-object tracking (SMOT), a new tracking paradigm that expands existing MOT task by integrating \"where\" and \"what\"; " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " We present BenSMOT, a large-scale dataset with 3,292 videos and rich annotations for SMOT; " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " We propose SMOTer, a simple but effective tracker to facilitate future research of SMOT; " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " We show that SMOTer achieves promising performance for SMOT and conduct in-depth analysis on it to provide guidance for future algorithm design." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 220, + 243, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 243, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 243, + 232 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 246, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 461 + ], + "type": "text", + "content": "MOT Benchmarks. Benchmarks have played a crucial role in facilitating the development of MOT. PETS2009 [16] is one of the earliest benchmarks for multipedestrian tracking. Later, the popular MOT Challenge [10, 24, 33] has been introduced with more crowded videos and greatly advanced MOT. KITTI [18] and BDD100K [48] are specifically designed for object tracking in autonomous driving. ImageNet-Vid [11] provides trajectory annotations for 30 categories across over 1,000 videos, and TAO [9] further expands object classes to 833 for generic multi-object tracking. To foster MOT in specific dancing and sport scenarios, DanceTrack [38] and SportsMOT [8] have been presented for dancer and player tracking. AnimalTrack [52] focuses on tracking various animals in wild scenes. Moreover, UAVDT [12] and VisDrone [61] provide platforms for tracking targets with drones. Different from the above MOT benchmarks for object trajectory prediction only (\"where\"), BenSMOT is specially developed for the new task of SMOT to serve as a platform for simultaneous estimation of target trajectories and trajectory-associated semantic understanding. To this end, it offers rich annotations of not only object trajectories as in existing standard MOT datasets (\"where\") but also semantic details of objects (\"what\") such as instance trajectory captions, interactions and overall video captions (see Fig. 1 again)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "text", + "content": "MOT Algorithms. MOT algorithms have seen great progress in recent years. One of the popular paradigms for multi-object tracking is the so-called Tracking-by-Detection. This paradigm involves initial object detection followed by association based on these detections, forming the basis for many representative methods [5,13,21,31,43,53]. In this case, MOT methods typically improve their performance by enhancing both detection and matching effectiveness. In addition, another prevalent paradigm is the Joint-Tracking-and-Detection [20,44,45,47,59], which integrates tracking and detection steps into a single process, enabling end-to-end training. Recently, Transformer [41] has been introduced into MOT and exhibited remarkable improvements over previous tackers [7,17,32,39,50,55,60]. Our SMOTer is related to but different than existing MOT methods. Specifically, going beyond merely predicting object trajectories, SMOTer also aims at semantic understanding of trajectories, integrating “where” and “what”." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Video Captioning. Video captioning is a multimodal vision-language task that aims to automatically describe the video content using natural language. Owing to its important applications in video event commentary and human-computer interaction, video captioning has drawn increasing attention in the past decade" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "with many models proposed (e.g., [23, 27, 36, 37, 46, 57]). Similar to the video captioning task, SMOT utilizes natural language to describe instances and the overall video content. However, the difference is that our SMOT aims at language comprehension for dense target trajectories in the context of multi-object tracking, which is more challenging due to the requirement of accurate trajectories yet provides more fine-grained information for video understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "type": "text", + "content": "Visual Relationship Detection. The goal of visual relationship detection task is to identify and comprehend the relationships (usually represented by a triplet of " + }, + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "type": "text", + "content": " subject, predicate, object " + }, + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\rangle" + }, + { + "bbox": [ + 130, + 188, + 480, + 271 + ], + "type": "text", + "content": ") between objects from an image (e.g., [25, 29, 51]). Besides static images, recent works extend visual relationship detection in the video domain (e.g., [6, 28, 56]). Our work is relevant to video visual relationship detection but different by extending this task target-specific trajectories, and thus more challenging due to complicated multi-object tracking scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 291, + 304, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 291, + 304, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 291, + 304, + 304 + ], + "type": "text", + "content": "3 The Proposed BenSMOT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 317, + 244, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 317, + 244, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 317, + 244, + 329 + ], + "type": "text", + "content": "3.1 Design Principle" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 338, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 479, + 361 + ], + "type": "text", + "content": "BenSMOT aims at providing a new platform for exploring human-centric SMOT. In construction of BenSMOT, we follow the following rules:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 361, + 480, + 529 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 361, + 480, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 421 + ], + "type": "text", + "content": "(1) Dedicated benchmark. The key motivation of our BenSMOT is to provide a dedicated dataset for semantic multi-object tracking. In current deep learning era, a large number of videos are desired in benchmark construction for training robust tracking models. Considering this, we hope to build a dedicated dataset with more than 3,000 videos for human-centric semantic multi-object tracking." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 422, + 480, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 469 + ], + "type": "text", + "content": "(2) Diverse scenarios. For a dataset, diversity is crucial, in both training and evaluation, for developing general systems. In order to provide a diverse platform for SMOT, in BenSMOT we will include videos from more than 40 various scenes that range from daily life scenarios to dancing scenarios to sport scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 470, + 480, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 480, + 529 + ], + "type": "text", + "content": "(3) High-quality annotations. High-quality annotation is crucial for a benchmark in both training and assessing models. In BenSMOT, we ensure its high quality by providing manual annotations for each sequence including object trajectories, instance captions, instance interactions and overall video captions. In this process, multi-round inspections and refinements will be carried out." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 548, + 246, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 246, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 246, + 560 + ], + "type": "text", + "content": "3.2 Data Acquisition" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "BenSMOT focuses on predicting the target trajectories and meanwhile understanding their semantics by instance captioning, interaction recognition, and overall video captioning. To this end, the video sequences in BenSMOT are desired to involve diverse multi-person activities and interactions, which is different from videos in existing benchmarks to some extent. Specifically, we first identify 47 common scenarios, involving different activities ranging from daily-life talk and play to different sports, that are suitable for tracking by drawing inspiration from [22]. Due to limited space, please refer to supplementary material for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 448, + 102 + ], + "type": "text", + "content": "Beyond MOT: Semantic Multi-Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 146, + 460, + 251 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Summary of BenSMOT and its comparison with popular multi-object tracking benchmarks. \"n/a\" denotes that data is not available." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 146, + 460, + 251 + ], + "lines": [ + { + "bbox": [ + 151, + 146, + 460, + 251 + ], + "spans": [ + { + "bbox": [ + 151, + 146, + 460, + 251 + ], + "type": "table", + "html": "
KITTI [18]MOT17 [33]MOT20 [10]BDD100k [48]TAO [9]GMOT-40 [1]DanceTrack [38]SportsMOT [8]BenSMOT (ours)
Videos501481,6002,907401002403,292
Min. length (s)n/a17.017.040.0n/a3.0n/an/a1.5
Avg. length (s)10.033.066.840.036.88.952.9n/a22.9
Max. length (s)n/a85.0133.040.0n/a24.2n/an/a116.0
Total length (s)498463535640,000106,9783565,292601575,499
Total tracks2,6001.3K3.83K131K17,2872,0269903,4017,792
Total boxes80K300K2,102K3,300K333K256Kn/a1,629K335K
Total frames15K11K13K318K2,674K9K106K150K151K
Instance CaptionsXXXXXXXX7,792
Instance InteractionsXXXXXXXX14K
Video SummariesXXXXXXXX3,292
", + "image_path": "551841633fcc6e6c14a841ea8331e9ad5ed2921602e5d438750db0a3705710fd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "text", + "content": "detailed scenarios in BenSMOT. Afterwards, we search for raw sequences for each scenario under the Creative Commons License from YouTube, the largest and most popular video platform with massive real-world videos. Please notice, for each raw video, instead of using the whole sequence, we usually choose a suitable clip for our semantic multi-object tracking task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 480, + 441 + ], + "type": "text", + "content": "Eventually, we compile a dataset for SMOT by including 3,292 videos with 151K frames. BenSMOT has an average video length of 23 seconds, with our current focus in this work not being on long-term SMOT. The total number of instance tracks is 7.8K, annotated with 335K bounding boxes for tracking. Besides bounding boxes, we provide 7.8K instance captions, 14K interactions, and 3,292 video summaries for trajectory-associated semantic understanding (how to annotate will be described in the following). To our best knowledge, BenSMOT by far is the first publicly available benchmark dedicated for SMOT. Tab. 1 summarizes our BenSMOT and compares it with popular MOT benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 463, + 246, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 246, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 246, + 474 + ], + "type": "text", + "content": "3.3 Data Annotation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": "In order to meet the requirements of SMOT, BenSMOT provides four types of annotations, including bounding box, instance caption, instance interaction, and the overall video caption. Specifically, for each trajectory in videos, similar to standard MOT benchmarks (e.g., [9,33]), we provide axis-aligned bounding boxes to indicate its spatial positions in the videos. For instance captioning, we label each trajectory with a precise sentence in natural language to describe the detailed behavior of the associated target. For interaction recognition, we first collect a set of 335 interactions (i.e., verbs), among them 327 from WordNet [15]. Note that, in our interaction set, the same verb may have different meanings and thus is assigned to different interaction types, e.g., hold(v.01) for “being the physical support of someone” and hold(v.02) for “someone in one’s hands or arm”, for precise relationship description. Due to limited space, we refer reader to supplementary material for the full set of interaction types. After this, we provide an directed interaction label, represented with a triplet of " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": " subject predicate, object " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\rightharpoondown" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": " (see Fig. 1 (b)), between two trajectories if associated objects" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 249, + 175 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 249, + 175 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 249, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 249, + 175 + ], + "type": "image", + "image_path": "41fc0948e3f3b4719a1566e7dc91e1fb6ce3c9eec3c194f410c27907877d7474.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 176, + 246, + 242 + ], + "blocks": [ + { + "bbox": [ + 141, + 176, + 246, + 242 + ], + "lines": [ + { + "bbox": [ + 141, + 176, + 246, + 242 + ], + "spans": [ + { + "bbox": [ + 141, + 176, + 246, + 242 + ], + "type": "image", + "image_path": "9959dbb8db30f6cbdaa45fcfd9201b5c977df662d35dc8d58a9036754777b6b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 243, + 246, + 251 + ], + "lines": [ + { + "bbox": [ + 140, + 243, + 246, + 251 + ], + "spans": [ + { + "bbox": [ + 140, + 243, + 246, + 251 + ], + "type": "text", + "content": "(d) Inst caption length distribution" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 252, + 114, + 363, + 175 + ], + "blocks": [ + { + "bbox": [ + 252, + 114, + 363, + 175 + ], + "lines": [ + { + "bbox": [ + 252, + 114, + 363, + 175 + ], + "spans": [ + { + "bbox": [ + 252, + 114, + 363, + 175 + ], + "type": "image", + "image_path": "0f523d359a7ee4a387f43faac26e3cea9942a124cca07edc04411446f59c5627.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 257, + 176, + 356, + 243 + ], + "blocks": [ + { + "bbox": [ + 257, + 176, + 356, + 243 + ], + "lines": [ + { + "bbox": [ + 257, + 176, + 356, + 243 + ], + "spans": [ + { + "bbox": [ + 257, + 176, + 356, + 243 + ], + "type": "image", + "image_path": "2a1ca43adafa5b6fa864e9b4406e4218bf8a2152ef3966bf1f8ec84ec0fd1da3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 244, + 362, + 251 + ], + "lines": [ + { + "bbox": [ + 251, + 244, + 362, + 251 + ], + "spans": [ + { + "bbox": [ + 251, + 244, + 362, + 251 + ], + "type": "text", + "content": "(e) Video caption length distribution" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 477, + 175 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 477, + 175 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 477, + 175 + ], + "type": "image", + "image_path": "cc21fb07e299f637ad7cda3e98e90058e4ed5c83144beb57f120cea72d724c61.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 369, + 176, + 473, + 243 + ], + "blocks": [ + { + "bbox": [ + 369, + 176, + 473, + 243 + ], + "lines": [ + { + "bbox": [ + 369, + 176, + 473, + 243 + ], + "spans": [ + { + "bbox": [ + 369, + 176, + 473, + 243 + ], + "type": "image", + "image_path": "4ad3a9d96dbaa1096c26e248d75b301f9542ee1fb0e52bfdd158877638e333ec.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 244, + 469, + 251 + ], + "lines": [ + { + "bbox": [ + 372, + 244, + 469, + 251 + ], + "spans": [ + { + "bbox": [ + 372, + 244, + 469, + 251 + ], + "type": "text", + "content": "(f) WordCloud of caption words" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 260, + 480, + 305 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 480, + 305 + ], + "type": "text", + "content": "Fig. 2: Representative statistics on BenSMOT, including distributions of target trajectory length (in seconds) in (a), sequence length (in seconds) in (b), number of interactions per sequence in (c), instance caption length in (d), video caption length in (e), and wordcloud of all caption words with prepositions excluded in (f)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 329, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 479, + 352 + ], + "type": "text", + "content": "are interacting. Finally, for the overall video captioning, we utilize a concise text summary to describe object trajectories in the video from a global perspective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": "To ensure the high-quality annotations, we adopt a multi-round mechanism. Specifically, each video will be first manually annotated by a few volunteers who are familiar with the task and an expert working on related problems. After this initial round, all annotations of the video will be sent to a validation team former by more than two experts for inspection. If initial annotations are not unanimously agreed by the experts, they will be returned to the original labeling team for refinement using the feedback from the validation team. We repeat this procedure for each video until the annotations of all video sequences in BenSMOT are completed. An annotation example in our BenSMOT is shown in Fig. 1 (b), and more can be found in supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 472, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 480, + 556 + ], + "type": "text", + "content": "Statistics of annotation. In order to better understand BenSMOT, we show some representative statics of annotations in Fig. 2. Specifically, we demonstrate the distributions of object trajectory length, video length, number of instance interactions, instance caption length, video caption length, and wordcloud. From Fig. 2 (c), it is worth noting that different targets in videos have frequent interactions. In addition, for the instance captions in Fig. 2 (d), we provide relatively longer textual descriptions, which allows precise understanding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 574, + 343, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 343, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 343, + 586 + ], + "type": "text", + "content": "3.4 Dataset Split and Evaluation Metric" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "text", + "content": "Training/Test Split. BenSMOT contains 3,292 videos, captured from 47 different scenarios. Within each scenario, we use " + }, + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "text", + "content": " of the sequences for training, and the rest " + }, + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "text", + "content": " for test. During dataset split, we try to keep the distributions of training and test sets similar. Eventfully, the training set of BenSMOT comprises 2,284 sequences with 104K frames, and the test set consists of 1,008 videos with 47K frames. Please see supplementary material for more split details." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 448, + 102 + ], + "type": "text", + "content": "Beyond MOT: Semantic Multi-Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 462, + 284 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 462, + 284 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 462, + 284 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 462, + 284 + ], + "type": "image", + "image_path": "9e171706b4b13802126f70e12061a427159fe8b04983952bd097dce288d2a556.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 293, + 480, + 316 + ], + "lines": [ + { + "bbox": [ + 131, + 293, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 480, + 316 + ], + "type": "text", + "content": "Fig. 3: Illustration of SMOTer, which contains three components of trajectory estimation for tracking, feature fusion, and trajectory-associated semantic understanding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 482 + ], + "type": "text", + "content": "Evaluation Metric. We apply multiple metrics on BenSMOT for evaluation. Specifically, to assess the performance in object trajectory estimation, we employ higher order tracking accuracy (HOTA), association accuracy (AssA), detection accuracy (DetA), and localization accuracy (LocA) by following [30], CLEAR metrics [3] including multiple object tracking accuracy (MOTA), false positives (FP), false negatives (FN), and ID switches (IDs), and ID metrics [35] containing identification precision (IDP), identification recall (IDR) and related F1 score (IDF1). For instance and video captioning, we follow existing methods (e.g., [27, 37]) and employ BLEU [34], ROUGE [26], METEOR [2], and CIDEr [42] for thorough evaluation. Lastly, to assess performance of interaction recognition, we leverage classic metrics such as Precision (Prcn), Recall (Rcll), and F1, which are widely used in the context of classification tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 498, + 400, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 400, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 400, + 513 + ], + "type": "text", + "content": "4 Methodology: A New Baseline for SMOT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Overview. To encourage development of SMOT algorithms, we propose a simple yet effective tracker on BenSMOT, dubbed SMOTer. As illustrated in Fig. 3, SMOTer logically consists of three key components. The first component, as detailed in Sec 4.1, is dedicated to object trajectory estimation for tracking, which is the basis for SMOT. Then, the second component, as explained in Sec 4.2, focuses on aggregating features from each frame into the overall video feature and meanwhile merging features of individual objects into target trajectory features using two distinct fusion modules, for subsequent semantic understanding. Finally, the third component, as described in Sec 4.3, aims to predict trajectory-associated semantic details in the video, including instance captions, instance interactions, and overall video caption. Please notice that, our SMOTer is specially designed for SMOT and trained in an end-to-end manner, as in Sec 4.4." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 373, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 373, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 373, + 129 + ], + "type": "text", + "content": "4.1 Target Trajectory Estimation for Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": "Give a video with " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " frames, SMOTer first uses a CNN backbone [49] to extract their features " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " denotes the feature of the " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " frame. Then, an initial set of proposals is generated using proposal generator of a popular detection architecture [14] with threshold filtering by " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\tau_p" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": ". Afterwards, all reserved proposals are send to the association module (BYTE [53] is used in SMOTer for association) to obtain object trajectories. Finally, leveraging target coordinates within the trajectories and image features, target features are acquired for each target in trajectories through RoI Pooling [19]. Specifically, given feature " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " frame and target coordinates " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "B_j^i" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "j^{\\text{th}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " object trajectory in this frame, we extract the corresponding target feature " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "t_j^i = \\text{RoIPooling}(f_i, B_j^i)" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " and apply it in subsequent feature fusion for semantic understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 482, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 299 + ], + "type": "text", + "content": "In summary, at this stage, SMOTer predicts object trajectories for tracking in a given video and extracts image and target features for subsequent stages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 319, + 236, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 236, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 236, + 331 + ], + "type": "text", + "content": "4.2 Feature Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "type": "text", + "content": "Given image and target features from previous stage in Sec. 4.1, we design two feature fusion modules, including attention-based Video Fusion Module (VFM) and concatenation-based Trajectory Fusion Module (TFM). VFM aims to merge image-level features into overall video features while TFM focuses on integrating target features from each frame into target trajectory features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "text", + "content": "Video Fusion Module. VFM takes features " + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "text", + "content": " frames as input and sequentially feeds them into a cross-attention module for video feature fusion. In specific, the fused video feature " + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "text", + "content": " up to the " + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "inline_equation", + "content": "i > 1" + }, + { + "bbox": [ + 130, + 403, + 482, + 441 + ], + "type": "text", + "content": ") frame is computed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 252, + 455, + 481, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 455, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 252, + 455, + 481, + 467 + ], + "type": "interline_equation", + "content": "F _ {i} = \\mathbb {C A} \\left(F _ {i - 1}, f _ {i}\\right) \\quad i > 1 \\tag {1}", + "image_path": "fc07fcd32b75045b01bf1c84eb67bf367b74dd5b132e6caec2773486d8c3b00a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{CA}(\\mathbf{z},\\mathbf{u})" + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "text", + "content": " denotes cross-attention with " + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "text", + "content": " generating query and " + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "text", + "content": " key/value as in [41]. For the first frame (" + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "inline_equation", + "content": "F_{1} = f_{1}" + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "text", + "content": ". Finally, with the help of VFM, we can generate the overall fused video feature " + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{F} = F_{N}" + }, + { + "bbox": [ + 130, + 475, + 482, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": "Trajectory Fusion Module. Unlike VFM, TFM is to integrate multiple target features belonging to the same object trajectory, which are distributed across different frames, into a single trajectory feature. Concretely, given " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "t_i^j" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": " for object feature of trajectory " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": ", we perform average pooling to yield a relatively simple 2D feature " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\hat{t}_i^j = \\mathsf{AvgPooling}(t_i^j)" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\hat{t}_i^j \\in \\mathbb{R}^{4 \\times d}" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": " with dimension " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": ". Then, for the " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": " object trajectory, we first concatenate its discrete features into an integrated one " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\hat{T}_j = \\mathsf{Concat}(\\hat{t}_1^j, \\hat{t}_2^j, \\dots, \\hat{t}_{N_j}^j)" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "N_j" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": " denotes trajectory length. After that, self-attention is used to enhance trajectory feature as follows" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 280, + 620, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 620, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 280, + 620, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\tilde {T} _ {j} = \\mathbf {S A} (\\hat {T} _ {j}) \\tag {2}", + "image_path": "78597e4a15e1b07129f33665a27b7bca82c9ccd1c119b012f9b4450b14f6ed85.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{SA}(\\mathbf{z})" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " denotes self-attention with " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " generating query/key/value as in [41], and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\bar{T}_j" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " is the enhanced feature for trajectory " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 448, + 103 + ], + "type": "text", + "content": "Beyond MOT: Semantic Multi-Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 286, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 286, + 129 + ], + "type": "text", + "content": "4.3 Semantic Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 479, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 479, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 479, + 183 + ], + "type": "text", + "content": "In the third phase, SMOTer applies the fused video feature " + }, + { + "bbox": [ + 130, + 135, + 479, + 183 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 130, + 135, + 479, + 183 + ], + "type": "text", + "content": " and trajectory features " + }, + { + "bbox": [ + 130, + 135, + 479, + 183 + ], + "type": "inline_equation", + "content": "\\tilde{T} = \\{\\tilde{T}_1,\\tilde{T}_2,\\dots,\\tilde{T}_M\\}" + }, + { + "bbox": [ + 130, + 135, + 479, + 183 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 135, + 479, + 183 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 135, + 479, + 183 + ], + "type": "text", + "content": " denotes the number of trajectories in the video, to tackle trajectory-associated semantic understanding tasks in SMOT, i.e., video captioning, instance captioning, and interaction recognition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 479, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 479, + 244 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 479, + 244 + ], + "type": "text", + "content": "Video captioning. Unlike current video captioning, in SMOTer we incorporate trajectory features into describing video content and allow more comprehensive human-centric caption generation. Specifically, SMOTer sequentially injects trajectory features " + }, + { + "bbox": [ + 130, + 184, + 479, + 244 + ], + "type": "inline_equation", + "content": "\\tilde{T}" + }, + { + "bbox": [ + 130, + 184, + 479, + 244 + ], + "type": "text", + "content": " into video feature " + }, + { + "bbox": [ + 130, + 184, + 479, + 244 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 130, + 184, + 479, + 244 + ], + "type": "text", + "content": " with cross-attention and then predicts captioning result " + }, + { + "bbox": [ + 130, + 184, + 479, + 244 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{Vid}}" + }, + { + "bbox": [ + 130, + 184, + 479, + 244 + ], + "type": "text", + "content": " via a linear project layer and a text decoder, as follows" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 233, + 252, + 480, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 252, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 233, + 252, + 480, + 266 + ], + "type": "interline_equation", + "content": "R _ {\\mathrm {V i d}} = \\operatorname {D e c} _ {\\mathrm {V i d}} (\\operatorname {L i n e a r} (\\mathbf {C A} (\\tilde {F}, \\tilde {T}))) \\tag {3}", + "image_path": "839cb1ac851afa880706de5f8a518859df1a4cbafefed7cd0f36dd99ec5d2e9d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\mathsf{Dec}_{\\mathrm{Vid}}(\\cdot)" + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "text", + "content": " is a video text decoder for video captioning (please kindly refer to supplementary material for details), and " + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{Linear}(\\cdot)" + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "text", + "content": " is a linear projection. Instance captioning. Different from video captioning, for instance captioning, we directly generate the results based on the target trajectory features. Specifically, given the feature " + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\tilde{T}_j" + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "inline_equation", + "content": "1 \\leq j \\leq M" + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "text", + "content": ") trajectory, we apply a linear projection layer followed by the designed text decoder to predict the captioning result " + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{Ins}}^j" + }, + { + "bbox": [ + 130, + 274, + 480, + 360 + ], + "type": "text", + "content": " of each trajectory, depicted as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 215, + 368, + 480, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 368, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 215, + 368, + 480, + 383 + ], + "type": "interline_equation", + "content": "R _ {\\text {I n s}} ^ {j} = \\operatorname {D e c} _ {\\text {I n s}} \\left(\\operatorname {L i n e a r} \\left(\\tilde {T} _ {j}\\right)\\right) \\quad (1 \\leq j \\leq M) \\tag {4}", + "image_path": "07e5e7b28c69befdaa904fb9dadaa0ef49a0a63b9bf2f7f3b65997ebca4b5a48.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 479, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 479, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 479, + 414 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 390, + 479, + 414 + ], + "type": "inline_equation", + "content": "\\mathsf{Dec}_{\\mathrm{Ins}}(\\cdot)" + }, + { + "bbox": [ + 130, + 390, + 479, + 414 + ], + "type": "text", + "content": " is instance text decoder [46] for instance captioning. Please refer to supplementary material for more details because of space limitation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "text", + "content": "Interaction recognition. Besides captioning, SMOTer predicts the interaction between trajectories, which is crucial for video understanding. Particularly, in SMOTer we concentrate solely on interactions between two trajectories, as our current focus doesn't extend to understanding interactions involving three or more targets, leaving it for future research. More specifically, for two trajectories " + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "text", + "content": ", we assume the " + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "text", + "content": " trajectory in a video sequence is the active trajectory and the " + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "text", + "content": " trajectory the passive trajectory. Then, we first fuse the features of active and passive trajectories using cross-attention and then predict the interaction result " + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{Int}}^{j,k}" + }, + { + "bbox": [ + 130, + 415, + 480, + 524 + ], + "type": "text", + "content": " using multi-layer perception [40], as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 256, + 532, + 480, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 532, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 256, + 532, + 480, + 547 + ], + "type": "interline_equation", + "content": "R _ {\\text {I n t}} ^ {j, k} = \\operatorname {M L P} (\\mathrm {C A} (\\tilde {T} _ {j}, \\tilde {T} _ {k})), \\tag {5}", + "image_path": "f6252c0d1d02fd9132102feed1f10daba279a757859daefbc9a671e021bb698c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 555, + 479, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 479, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 479, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 555, + 479, + 593 + ], + "type": "inline_equation", + "content": "\\mathsf{MLP}(\\cdot)" + }, + { + "bbox": [ + 130, + 555, + 479, + 593 + ], + "type": "text", + "content": " denotes the multi-layer perception. Please note, interchanging the active and passive trajectories may yield disparate outcomes, i.e., " + }, + { + "bbox": [ + 130, + 555, + 479, + 593 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{Int}}^{j,k}" + }, + { + "bbox": [ + 130, + 555, + 479, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 555, + 479, + 593 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{Int}}^{k,j}" + }, + { + "bbox": [ + 130, + 555, + 479, + 593 + ], + "type": "text", + "content": " may exhibit dissimilar characteristics, as indicated by the formula above." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 609, + 263, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 263, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 263, + 623 + ], + "type": "text", + "content": "4.4 End-to-end Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "End-to-end training of SMOTer is not easy due to conflict between two types of training requirements, i.e., the trajectory estimation is usually trained frame-by-frame, while the captioning tasks, and interaction recognition are often trained" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": "at the entire video level, resulting in two different losses in training. In addition, due to the frame-by-frame training nature of trajectory estimation for tracking, the complete sequence is hard to be input at once during training, preventing the model from obtaining full trajectories thus hindering subsequent caption generation and interaction prediction. To address these issues, we draw inspiration from [60] and adopt a strategy of completing full trajectory association after detection on the entire video. Specifically, we first compute detection loss after each frame input and then perform trajectory association after the entire sequence is detected. Subsequently, with complete trajectories, we compute losses for trajectory-associated instance captioning, interaction recognition, and video captioning tasks, realizing end-to-end training of SMOTer. For the detailed training loss of SMOTer, please kindly refer to the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 281, + 230, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 230, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 230, + 295 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "type": "text", + "content": "Experimental setup. We conduct experiments with 4 Nvidia Tesla V100 GPUs (32GB). The batch size is 1 per GPU. Each batch has a video clip with multiple frames. We use the AdamW optimizer with an initial learning rate of " + }, + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "type": "inline_equation", + "content": "5.0 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "type": "text", + "content": ". During training, we filter out proposals with scores lower than the threshold " + }, + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\tau_{p} = 0.3" + }, + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "type": "text", + "content": ". For lost tracklets, we retain them for 30 frames in case they reappear. Due to memory constraints and limited accuracy of trajectory estimation in early training stages, we sample 6-8 frames per sequence for training detection model and only compute tracking related losses in the first 20K training iterations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 406, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 550 + ], + "type": "text", + "content": "As SMOT is new and no suitable methods exist for comparison with SMOTer, we logically divide SMOT into two components: trajectory estimation for tracking and semantic understanding (including instance/video captioning, and interaction recognition). Subsequently, we devise a series of two-stage models based on existing MOT frameworks. Leveraging several state-of-the-art and classic MOT models such as SORT [4], DeepSORT [43], OC-SORT [5], ByteTrack [53], TransTrack [39], MOTR [50], and MOTRv2 [55] as foundations, we integrate feature fusion and semantic generation functionalities akin to SMOTer and conduct comparative experiments on BenSMOT. To ensure fair experiments, we employ the same training framework for comparative experiments of two-stage approaches. Additionally, for models based on two-stage MOT methods like SORT, we use the same detector [14] as SMOTer during training and evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 571, + 421, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 421, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 421, + 583 + ], + "type": "text", + "content": "5.1 Comparison on Trajectory Estimation for Tracking." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We first compare SMOTer with other MOT methods for trajectory estimation on BenSMOT. Specifically, SMOTer is fully end-to-end trained, while other two-stage MOT models are trained with their own strategies, concentrating solely on tracking. As depicted in Tab. 2, despite being at a certain degree of natural disadvantage, our SMOTer achieves comparable or even superior performance in tracking compared to state-of-the-art MOT models. SMOTer consistently" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 448, + 103 + ], + "type": "text", + "content": "Beyond MOT: Semantic Multi-Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 146, + 460, + 218 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Comparison between SMOTer and two-stage MOT methods regarding tracking performance on BenSMOT. The best two results are shown in red and blue fonts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 146, + 460, + 218 + ], + "lines": [ + { + "bbox": [ + 151, + 146, + 460, + 218 + ], + "spans": [ + { + "bbox": [ + 151, + 146, + 460, + 218 + ], + "type": "table", + "html": "
MethodHOTA↑AssA↑DetA↑LocA↑MOTA↑FN↓FP↓IDs↓IDR↑IDP↑IDF1↑
SORT [4]48.4938.9560.9187.5053.582400151051387560.8548.4353.93
DeepSORT [43]50.1240.2361.4587.6754.292289055401127862.1051.1156.76
OC-SORT [5]51.0041.4263.3187.6155.192106153881504963.9253.1058.01
ByteTrack [53]68.8471.1567.1085.1573.87154197070171282.2574.8378.37
TransTrack [39]71.3173.3469.6791.3174.08201244420253085.6372.7578.67
MOTR [50]66.1073.1255.1486.3045.19312971117861772.3970.1268.97
MOTRv2 [55]65.2876.8251.3086.0945.52407652092343078.4765.5170.76
SMOTer (ours)71.9873.7170.7987.1177.71125346388170283.8277.9780.65
", + "image_path": "bda312800e7a2c4da6b74cf7f07715037debe2d8cf8644801228d3e9ee6651fd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 151, + 259, + 460, + 338 + ], + "blocks": [ + { + "bbox": [ + 132, + 227, + 480, + 250 + ], + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 250 + ], + "type": "text", + "content": "Table 3: Comparison of SMOTer against two-stage methods based on MOT models regarding semantic understanding. The best two results are highlighted in red and blue." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 259, + 460, + 338 + ], + "lines": [ + { + "bbox": [ + 151, + 259, + 460, + 338 + ], + "spans": [ + { + "bbox": [ + 151, + 259, + 460, + 338 + ], + "type": "table", + "html": "
Video CaptionInstance CaptionInteraction
MethodBLEU↑ROUGE↑METEOR↑CIDEr↑BLEU↑ROUGE↑METEOR↑CIDEr↑Prec↑Rcll↑F1↑
SORT [4]0.2450.2240.2020.2980.2330.2450.2080.0560.3630.2590.302
DeepSORT [43]0.1980.2130.1870.3090.2380.2120.1990.0650.3650.2770.310
OC-SORT [5]0.2310.2520.2150.2420.2700.2050.1800.0330.3840.2910.331
ByteTrack [53]0.2240.2250.2120.2660.3040.2420.2240.0640.4430.2580.326
TransTrack [39]0.2470.2480.2090.2690.2830.2190.2010.0740.4060.3110.376
MOTR [50]0.1870.2540.2030.2440.2300.2090.1820.0610.4250.3140.354
MOTRv2 [55]0.2170.2580.2190.2480.2380.2410.2040.0590.3130.3950.349
SMOTer (ours)0.2450.2610.2230.3430.3060.2230.2090.0870.4340.3200.368
", + "image_path": "74b1e24c58fb9994ad2bfb92766a3acf525839b588320eb0dbfbee39c36cfefb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 151, + 380, + 460, + 430 + ], + "blocks": [ + { + "bbox": [ + 132, + 348, + 479, + 370 + ], + "lines": [ + { + "bbox": [ + 132, + 348, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 479, + 370 + ], + "type": "text", + "content": "Table 4: Ablation experiments for evaluating various feature fusion strategies using results of three additional tasks. The best result is highlighted in red." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 380, + 460, + 430 + ], + "lines": [ + { + "bbox": [ + 151, + 380, + 460, + 430 + ], + "spans": [ + { + "bbox": [ + 151, + 380, + 460, + 430 + ], + "type": "table", + "html": "
VFM (video fusion)TFM (trajectory fusion)
BLEU↑ROUGE↑METEOR↑CIDEr↑BLEU↑ROUGE↑METEOR↑CIDEr↑Pcn↑Rcll↑F1↑
Attention-based0.2450.2610.2230.3430.3030.2180.2080.0720.4060.3470.364
MLP-based0.2510.2230.2000.2200.3370.2190.2050.0720.3800.2520.303
Concatenation-based0.2300.2130.2150.3650.3060.2230.2090.0870.4340.3200.368
Addition-based0.1400.1570.1040.1270.1820.1370.1140.0410.2280.1630.205
", + "image_path": "083fe2dfb38a383cbc5f490de3535173529d8c656e2f6bd0cd72ec3703bfbb6e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "text", + "content": "achieves top-tier performance across nearly all metrics, demonstrating impressive results in key metrics. For instance, it achieves a notable " + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "inline_equation", + "content": "71.98\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "text", + "content": " in HOTA, " + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "inline_equation", + "content": "77.71\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "text", + "content": " in MOTA, and " + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "inline_equation", + "content": "80.65\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "text", + "content": " in IDF1. This suggests that subsequent training tasks do not degrade tracking performance, validating effectiveness of our end-to-end training strategy. In fact, by comparing with the standard ByteTrack, we observe that although SMOTer adopts a highly similar structure to ByteTrack in tracking, it achieves superior performance, showing " + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "inline_equation", + "content": "+3.15\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "text", + "content": " in HOTA and " + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "inline_equation", + "content": "+3.74\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 571 + ], + "type": "text", + "content": " in MOTA. Moreover, even in metrics where it does not excel, such as FP and LocA, SMOTer exhibits significant improvements compared to ByteTrack, implying that semantic understanding tasks can reciprocally aid in tracking." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 586, + 365, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 365, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 365, + 600 + ], + "type": "text", + "content": "5.2 Comparison on Semantic Understanding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "Besides tracking, we compare SMOTer with a set of two-stage models on their semantic understanding capabilities. Similarly, SMOTer undergoes complete end-to-end training. However, other two-stage models first train trackers on their own to acquire target trajectories and then use these trajectories to separately train the semantic understanding tasks (with the same feature fusion as SMOTer)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 219, + 167 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 219, + 167 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 219, + 167 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 219, + 167 + ], + "type": "image", + "image_path": "2b4035949ae0bcf10d5fa8bf6ce10f71f8ba56bfc486d6cb00d78d7addd1c7b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 168, + 220, + 182 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 220, + 182 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 220, + 182 + ], + "type": "text", + "content": "(a) Comparison results on object tracking" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 115, + 306, + 167 + ], + "blocks": [ + { + "bbox": [ + 222, + 115, + 306, + 167 + ], + "lines": [ + { + "bbox": [ + 222, + 115, + 306, + 167 + ], + "spans": [ + { + "bbox": [ + 222, + 115, + 306, + 167 + ], + "type": "image", + "image_path": "eec43bb00846fecfcfb1c050803386a7698d65e8f4848611f014fe75ff238dcb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 192, + 481, + 225 + ], + "lines": [ + { + "bbox": [ + 131, + 192, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 131, + 192, + 481, + 225 + ], + "type": "text", + "content": "Fig. 4: Ablation experiments for different association strategies, including comparison results on object tracking in image (a), on video captioning in image (b), on instance captioning in image (c), and on interaction recognition in image (d)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 116, + 391, + 167 + ], + "blocks": [ + { + "bbox": [ + 221, + 168, + 306, + 182 + ], + "lines": [ + { + "bbox": [ + 221, + 168, + 306, + 182 + ], + "spans": [ + { + "bbox": [ + 221, + 168, + 306, + 182 + ], + "type": "text", + "content": "(b) Comparison results on video captioning" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 116, + 391, + 167 + ], + "lines": [ + { + "bbox": [ + 307, + 116, + 391, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 391, + 167 + ], + "type": "image", + "image_path": "7338f1c6ec8beb8ee6e3521a509bc3f6f29779d9e894a4e925cdc13666f3023a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 393, + 117, + 478, + 167 + ], + "blocks": [ + { + "bbox": [ + 306, + 168, + 391, + 182 + ], + "lines": [ + { + "bbox": [ + 306, + 168, + 391, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 391, + 182 + ], + "type": "text", + "content": "(c) Comparison results on instance captioning" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 393, + 117, + 478, + 167 + ], + "lines": [ + { + "bbox": [ + 393, + 117, + 478, + 167 + ], + "spans": [ + { + "bbox": [ + 393, + 117, + 478, + 167 + ], + "type": "image", + "image_path": "c9a35591f8b1de1674dfbf715070f47fabaf6320a0bbcbf274d23326188c0c7a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 168, + 478, + 182 + ], + "lines": [ + { + "bbox": [ + 392, + 168, + 478, + 182 + ], + "spans": [ + { + "bbox": [ + 392, + 168, + 478, + 182 + ], + "type": "text", + "content": "(d) Comparison results on interaction recognition" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "text", + "content": "From Tab. 3, we can see SMOTer achieves highly favorable results on almost all metrics. Particularly, compared to the baseline method ByteTrack, SMOTer surpasses it on most metrics, including " + }, + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "inline_equation", + "content": "+2.1\\%" + }, + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "text", + "content": " BLEU and " + }, + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "inline_equation", + "content": "+7.7\\%" + }, + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "text", + "content": " CIDEr on video captioning, " + }, + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "inline_equation", + "content": "+2.3\\%" + }, + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "text", + "content": " CIDEr on instance captioning, and " + }, + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "inline_equation", + "content": "+4.2\\%" + }, + { + "bbox": [ + 130, + 247, + 480, + 308 + ], + "type": "text", + "content": " F1 score on interaction recognition. More results can be seen in supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 324, + 237, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 237, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 237, + 335 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 342, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 533 + ], + "type": "text", + "content": "Impact of different feature fusion. Feature fusion is a crucial component of SMOTer for generating overall video and enhanced trajectory features. In this work, we study four types of feature fusion strategies for VFM and TFM, including attention-based fusion (using cross-attention), MLP-based fusion (using MLP module), concatenation-based fusion, and addition-based fusion. Please refer to supplementary material for detailed architectures. To assess the impact for VFM, we compare the performance on video caption generation because the fused video feature from VFM is used for this sub-task, while for TFM, we measure the results on instance captioning and interaction recognition since these two sub-tasks rely on the fused trajectory features. Tab. 4 shows the results of different fusion mechanisms for VFM and TFM. We can see that attention-based fusion works generally better for VFM by achieving best ROUGE score of 0.261 and METEOR score of 0.223 and second best BLEU score of 0.245 and CIDEr score of 0.343 for video captioning. For TFM, concatenation-based fusion shows more superior results by obtaining best scores on ROUGE, METEOR, CIDEr for instance captioning and on Precision and F1 for interaction recognition." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Analysis on association mechanism. Association is necessary in SMOTer to generate target trajectories. In our method, we leverage the popular BYTE [54] for proposal association. To analyze the impact of different association strategies, we compare other representative manners, including SORT [4], DeepSORT [43], and OC-SORT [5], with the adopted BYTE. Fig. 4 demonstrates the comparison results of association mechanisms in SMOTer (using BYTE) and other approaches on four tasks, including object tracking (Fig. 4 (a)), video captioning (Fig. 4 (b)), instance captioning (Fig. 4 (c)), and interaction recognition (Fig. 4 (d)). From the results in Fig. 4 (a)-(d), we can observe that BYTE used in our SMOTer can achieve the best or competitive performance on different metrics for various tasks, validating its effectiveness for semantic multi-object tracking." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 448, + 102 + ], + "type": "text", + "content": "Beyond MOT: Semantic Multi-Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 136, + 449, + 196 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 467, + 126 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 467, + 126 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 467, + 126 + ], + "type": "text", + "content": "Table 5: Study of different thresholds " + }, + { + "bbox": [ + 145, + 114, + 467, + 126 + ], + "type": "inline_equation", + "content": "{\\tau }_{p}" + }, + { + "bbox": [ + 145, + 114, + 467, + 126 + ], + "type": "text", + "content": " . The best result is highlighted in red." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 136, + 449, + 196 + ], + "lines": [ + { + "bbox": [ + 164, + 136, + 449, + 196 + ], + "spans": [ + { + "bbox": [ + 164, + 136, + 449, + 196 + ], + "type": "table", + "html": "
TrackingVideo CaptionInstance CaptionInteraction
τpHOTA↑MOTA↑IDF1↑METEOR↑CIDEr↑METEOR↑CIDEr↑Prcn↑Rcll↑F1↑
0.170.5172.3078.800.2190.2560.1920.0690.3780.3780.319
0.271.1073.8078.900.2150.2890.1990.0690.3840.3650.331
0.371.9877.7180.650.2240.3430.2090.0870.4340.3200.368
0.471.1976.3676.010.2140.3300.2150.0910.4200.3020.341
0.568.3074.3478.900.2240.3430.2160.0870.3960.2910.310
", + "image_path": "322d364199767ad9cf9d8806f1e696b5b63885f30b903cba1207be6cdef00f0a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "text", + "content": "Study of threshold " + }, + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "inline_equation", + "content": "\\tau_{p}" + }, + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "inline_equation", + "content": "\\tau_{p}" + }, + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "text", + "content": " is used to remove low-confidence proposals to ensure the quality of subsequent trajectories, and thus crucial for all tasks in SMOTer. Because of this, we conduct an ablation study for investigating different " + }, + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "inline_equation", + "content": "\\tau_{p}" + }, + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "text", + "content": ", and Tab. 5 displays the results. From Tab. 5, we can observe that, when " + }, + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "inline_equation", + "content": "\\tau_{p} = 0.3" + }, + { + "bbox": [ + 130, + 217, + 482, + 278 + ], + "type": "text", + "content": ", SMOTer achieves the overall best performance for different tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 293, + 214, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 214, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 214, + 306 + ], + "type": "text", + "content": "5.4 Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 432 + ], + "type": "text", + "content": "Semantic understanding for improving tracking. SMOT aims to go beyond just tracking by incorporating additional semantic understanding tasks including trajectory-associated instance/video captioning and interaction recognition. We observe that these extra goals greatly improve the tracking performance. For example, as shown in Tab. 2, we can see that joint training of different tasks leads to better performance of SMOTer (71.98%/77.71% in HOTA/MOTA) in tracking compared to its baseline ByteTrack (68.84%/73.87 in HOTA/MOTA). The possible reason is that SMOTer needs to comprehensively understand the video content for semantic multi-object tracking, and the learned semantic knowledge helps distinguish different object trajectories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 433, + 483, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 483, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 483, + 517 + ], + "type": "text", + "content": "Challenge in SMOT. One reason why SMOT is challenging lies in the complicated behaviors of targets, often leading to lengthy instance captions. For instance, a basketball player can be dribbling, shooting, and defending in a game. Models often miss some states in understanding. In BenSMOT, the average video length is about 23 seconds, with objects typically present for around 21 seconds, resulting in instance captions averaging over 35 words. This complexity poses a challenge for SMOT and indicates a potential area for future improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 533, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 220, + 547 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 652 + ], + "type": "text", + "content": "In this paper, we introduce SMOT to expand the scope of MOT by integrating \"where\" and \"what\". To facilitate the research of SMOT, we propose the large-scale BenSMOT by including 3,292 videos from more than 40 diverse scenarios. To our best knowledge, BenSMOT is the first benchmark dedicated to SMOT. Furthermore, to encourage algorithm development on BenSMOT, we introduce SMOTer, an end-to-end tracker designed for SMOT. Our results exhibit SMOTer surpasses offline-combination strategies, showing efficacy. By presenting BenSMOT and SMOTer, we hope to inspire more future research on SMOT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "content": "Acknowledgements. Heng Fan was not supported by any fund for this work." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Bai, H., Cheng, W., Chu, P., Liu, J., Zhang, K., Ling, H.: Gmot-40: A benchmark for generic multiple object tracking. In: CVPR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: ACL Workshop (2005)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "text", + "content": "3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. JIVP (2008)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "text", + "content": "4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: ICIP (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "text", + "content": "5. Cao, J., Pang, J., Weng, X., Khirodkar, R., Kitani, K.: Observation-centric sort: Rethinking sort for robust multi-object tracking. In: CVPR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "type": "text", + "content": "6. Chen, S., Shi, Z., Mettes, P., Snoek, C.G.: Social fabric: Tubelet compositions for video relation detection. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "7. Chu, P., Wang, J., You, Q., Ling, H., Liu, Z.: Transmot: Spatial-temporal graph transformer for multiple object tracking. In: WACV (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "text", + "content": "8. Cui, Y., Zeng, C., Zhao, X., Yang, Y., Wu, G., Wang, L.: Sportsmot: A large multi-object tracking dataset in multiple sports scenes. ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "text", + "content": "9. Dave, A., Khurana, T., Tokmakov, P., Schmid, C., Ramanan, D.: Tao: A large-scale benchmark for tracking any object. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "text", + "content": "10. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J., Cremers, D., Reid, I., Roth, S., Schindler, K., Leal-Taixe, L.: Mot20: A benchmark for multi object tracking in crowded scenes. arXiv (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "text", + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "text", + "content": "2. Du, D., Qi, Y., Yu, H., Yang, Y., Duan, K., Li, G., Zhang, W., Huang, Q., Tian, Q.: The unmanned aerial vehicle benchmark: Object detection and tracking. In: ECCV (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "text", + "content": "3. Du, Y., Zhao, Z., Song, Y., Zhao, Y., Su, F., Gong, T., Meng, H.: Strongsort: Make deepsort great again. TMM (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "4. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: ICCV (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 469, + 399, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 399, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 399, + 479 + ], + "type": "text", + "content": "5. Fellbaum, C.: WordNet: An electronic lexical database (1998)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 480, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 480, + 501 + ], + "type": "text", + "content": "6. Ferryman, J., Shahrokni, A.: Pets2009: Dataset and challenge. In: PET Workshop (2009)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "text", + "content": "7. Gao, R., Wang, L.: Memoritr: Long-term memory-augmented transformer for multi-object tracking. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "8. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset (2013)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 545, + 317, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 317, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 317, + 555 + ], + "type": "text", + "content": "9. Girshick, R.: Fast r-cnn. In: ICCV (2015)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 556, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 577 + ], + "type": "text", + "content": "20. Han, X., Pasquier, T., Bates, A., Mickens, J., Seltzer, M.: Unicorn: Runtime provenance-based detector for advanced persistent threats. ECCV (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 578, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 598 + ], + "type": "text", + "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 620 + ], + "type": "text", + "content": "22. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "23. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV (2017)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "24. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv (2015)" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond MOT: Semantic Multi-Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "25. Liang, X., Lee, L., Xing, E.P.: Deep variation-structured reinforcement learning for visual relationship and attribute detection. In: CVPR (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 149 + ], + "type": "text", + "content": "26. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: ACL (2004)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "27. Lin, K., Li, L., Lin, C.C., Ahmed, F., Gan, Z., Liu, Z., Lu, Y., Wang, L.: Swinbert: End-to-end transformers with sparse attention for video captioning. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "28. Liu, C., Jin, Y., Xu, K., Gong, G., Mu, Y.: Beyond short-term snippet: Video relation detection with spatio-temporal global context. In: CVPR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "text", + "content": "29. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors. In: ECCV (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "30. Luiten, J., Osep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: Hota: A higher order metric for evaluating multi-object tracking. IJCV (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "31. Maggiolino, G., Ahmad, A., Cao, J., Kitani, K.: Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. ICIP (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "32. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "33. Milan, A., Leal-Taixe, L., Reid, I., Roth, S., Schindler, K.: Mot16: A benchmark for multi-object tracking. arXiv (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "35. Ristani, E., Solera, F., Zou, R., Cucchiara, R., Tomasi, C.: Performance measures and a data set for multi-target, multi-camera tracking. In: ECCV (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "36. Seo, P.H., Nagrani, A., Arnab, A., Schmid, C.: End-to-end generative pretraining for multimodal video captioning. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "37. Shen, Y., Gu, X., Xu, K., Fan, H., Wen, L., Zhang, L.: Accurate and fast compressed video captioning. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "38. Sun, P., Cao, J., Jiang, Y., Yuan, Z., Bai, S., Kitani, K., Luo, P.: Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "39. Sun, P., Cao, J., Jiang, Y., Zhang, R., Xie, E., Yuan, Z., Wang, C., Luo, P.: Transtrack: Multiple object tracking with transformer. arXiv (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "40. Taud, H., Mas, J.: Multilayer perceptron (mlp). Geomatic approaches for modeling land change scenarios (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NIPS (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "42. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "43. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: ICIP (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "44. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: ECCV (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "text", + "content": "45. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "46. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "47. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: ECCV (2022)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "48. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: CVPR (2020)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 402 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "49. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: CVPR (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "text", + "content": "50. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: ECCV (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "text", + "content": "51. Zhang, H., Kyaw, Z., Chang, S.F., Chua, T.S.: Visual translation embedding network for visual relation detection. In: CVPR (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 205 + ], + "type": "text", + "content": "52. Zhang, L., Gao, J., Xiao, Z., Fan, H.: Animaltrack: A benchmark for multi-animal tracking in the wild. IJCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 237 + ], + "type": "text", + "content": "53. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: ECCV (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "type": "text", + "content": "54. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 281 + ], + "type": "text", + "content": "55. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 481, + 303 + ], + "type": "text", + "content": "56. Zheng, S., Chen, S., Jin, Q.: Vrderformer: End-to-end video visual relation detection with transformers. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 303, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 481, + 325 + ], + "type": "text", + "content": "57. Zhou, L., Zhou, Y., Corso, J.J., Socher, R., Xiong, C.: End-to-end dense video captioning with masked transformer. In: CVPR (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 325, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 347 + ], + "type": "text", + "content": "58. Zhou, X., Arnab, A., Sun, C., Schmid, C.: Dense video object captioning from disjoint supervision. arXiv (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 347, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 481, + 358 + ], + "type": "text", + "content": "59. Zhou, X., Koltun, V., Krahenbihl, P.: Tracking objects as points. In: ECCV (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 481, + 380 + ], + "type": "text", + "content": "60. Zhou, X., Yin, T., Koltun, V., Krahenbuhl, P.: Global tracking transformers. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 402 + ], + "type": "text", + "content": "61. Zhu, P., Wen, L., Du, D., Bian, X., Fan, H., Hu, Q., Ling, H.: Detection and tracking meet drones challenge. TPAMI (2021)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond MOT: Semantic Multi-Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_content_list.json b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a8ca3b28caed840b71d498bff3625d243d7170ee --- /dev/null +++ b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_content_list.json @@ -0,0 +1,1772 @@ +[ + { + "type": "text", + "text": "Beyond Pixels: Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prantik Howlader1, Srijan Das2, Hieu Le3, and Dimitris Samaras1", + "bbox": [ + 261, + 233, + 738, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Stony Brook University, New York, USA \n2 University of North Carolina at Charlotte, NC, USA \n3 EPFL, Lausanne, Switzerland", + "bbox": [ + 316, + 261, + 684, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Incorporating pixel contextual information is critical for accurate segmentation. In this paper, we show that an effective way to incorporate contextual information is through a patch-based classifier. This patch classifier is trained to identify classes present within an image region, which facilitates the elimination of distractors and enhances the classification of small object segments. Specifically, we introduce Multi-scale Patch-based Multi-label Classifier (MPMC), a novel plug-in module designed for existing semi-supervised segmentation (SSS) frameworks. MPMC offers patch-level supervision, enabling the discrimination of pixel regions of different classes within a patch. Furthermore, MPMC learns an adaptive pseudo-label weight, using patch-level classification to alleviate the impact of the teacher's noisy pseudo-label supervision on the student. This lightweight module can be integrated into any SSS framework, significantly enhancing their performance. We demonstrate the efficacy of our proposed MPMC by integrating it into four SSS methodologies and improving them across two natural image and one medical segmentation dataset, notably improving the segmentation results of the baselines across all the three datasets. Code is available at: https://github.com/cvlab-stonybrook/Beyond-Pixels", + "bbox": [ + 261, + 340, + 743, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Segmentation $\\cdot$ Semi-supervised $\\cdot$ Patches", + "bbox": [ + 261, + 617, + 629, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 657, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Supervised segmentation requires a large amount of pixel level annotations, which is laborious and time-consuming [53-56]. This paper focuses on semi-supervised segmentation (SSS) [1, 20], which alleviates the reliance on labeled annotations by leveraging a large quantity of unlabeled images, accompanied by a limited number of manually labeled images.", + "bbox": [ + 212, + 689, + 787, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The fundamental task in segmentation is a per-pixel classification. Thus contextual information of a pixel is important to improve its performance [31,32,68]. Context becomes more critical in SSS settings, which uses limited labeled data making the models more prone to overfitting. We note that recent approaches in SSS methods which predominantly leverage consistency regularization with", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/295d3fe0ebeeedd571d70a575cb4f782811626d04b8e4161aefeb1890fa9e0be.jpg", + "image_caption": [ + "Fig.1: Segmentations and Multi-label classifications of image patches (Cityscapes) After 20 epochs, (b) and (c) are the segmentations and classes in the segmentations respectively, (d) is classes predicted by multi-label classifier in MPMC. (f) and (g) are segmentations by Baseline (AugSeg [65]) and Baseline + MPMC at the end of training. White boxes represent regions where MPMC improves the baseline, while orange boxes represent regions where segmentation network predicts classes not present in the patch." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 756, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pseudo-labeling techniques [7,14,20,22,36,38,50,57,69] do not exploit contextual information explicitly.", + "bbox": [ + 212, + 486, + 784, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We hypothesize that patches are the simplest way to determine the context around pixels. The key observation is that the contextual information of an image patch are the classes present in it. For example, in the image patch in the first row of Fig. 1 (a) the classes present are car, road, vegetation and terrain. Thus, training a network to not only classify each pixel in an image but also predict the classes in the patch surrounding it, is a straight forward way to introduce contextual information in SSS pipeline. Moreover, utilizing multi-scale patch-level supervision can effectively encapsulate context at varying scales and distinguish pixels across different classes. To this end, we introduce a novel plugin module, termed as Multi-scale Patch-based Multilabel Classifier (MPMC). This module achieves enhanced discrimination of pixel regions between classes through additional multi-scale patch-level supervision. Notably, we observe that when training our baseline SSS model together with MPMC in limited labeled data scenario, the patch classifier exhibits greater resilience to noise than the segmentation network as illustrated in Fig. 1 (b) and (c). A key observation is when MPMC and baseline SSS model are trained together, the patch classifier provides patch-level labels, identifying the classes present within a region. This facilitates the elimination of distractors (first row in Fig. 1 (b), (c) and (d), while the segmentation network classifies pixels as truck which is not present in the patch, the multi-label classifier does not predict it) and enhances the classification of small object segments by the baseline method.", + "bbox": [ + 212, + 522, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Howlader et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another important issue plaguing SSS methods is noisy pseudo-labels generated by the teacher for unlabeled images to supervise the student. This can lead to the student model overfitting to these noisy labels, consequently impairing its performance. Recent approaches have attempted to mitigate the impact of noise in pseudo-labels through various strategies, such as implementing a hard threshold [14, 23, 38], applying dynamic re-weighting to enhance the influence of reliable pixels [20], and employing negative data mining by archiving unreliable pixels in a memory bank [50]. In contrast to these approaches, we use the MPMC predictions to learn an adaptive weight that reduces the weight of noisy pseudo-labels in the teacher-student framework. Our approach hinges on the underlying observation that MPMC exhibits greater uncertainty (lower confidence) for noisy labels in a patch (see supplementary for details).", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, MPMC is a parameterized module that can be plugged in the initial layers of the visual model, where it leverages the low-level pixel information to classify the classes present in each patch. This patch-level classification strategy forces the visual model to learn and preserve discriminative representations of classes, which are often prone to dilution in the deeper layers of the model. We demonstrate the efficacy of our proposed parameterized module, MPMC by integrating it into four SSS methodologies—UniMatch [57], AugSeg [65], AEL [20], and U2PL [50] - notably improving performance across all data partitions in two natural (Cityscapes and Pascal VOC) and one medical (ACDC) dataset. To summarize, our contributions are:", + "bbox": [ + 212, + 329, + 787, + 478 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a novel plug-in module, called MPMC that incorporates an auxiliary multi-scale multi-label classification task, enhancing pixel region discrimination across different classes via additional patch-level supervision.", + "2. MPMC consists of two adaptive learning weights that reduces the influence of noisy pseudo-labels from the teacher model.", + "3. We demonstrate MPMC's robustness by integrating in four SSS techniques, achieving consistent improvements across all protocols on Pascal VOC and Cityscapes. Notably, our method also shows improvement in medical image dataset ACDC, which indicates that our method is applicable beyond natural images." + ], + "bbox": [ + 220, + 486, + 784, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 657, + 387, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL) has been extensively researched in recent years. Popular SSL approaches include consistency regularization [2,10,11,14,38,43,58], entropy minimization [5,16], and pseudo-labeling [24-28, 46]. In this work, our emphasis is placed on the domains of pseudo-labeling and consistency regularization.", + "bbox": [ + 212, + 688, + 787, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pseudo labeling: Pseudo-labeling methods [28,45] alongside self-training techniques [35, 59] are designed to leverage labeled image data to produce pseudolabels for unlabeled data. An inherent problem that plagues pseudo-labeling is noise, thus most approaches [42,44,46,69] devise strategies to refine the pseudolabels. PseudoSeg [69], for instance, aims at enhancing pseudo-label accuracy", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Beyond Pixels", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using attention mechanisms based on grad-cam [44]. In pseudo-label selection, FixMatch [46] adopts a confidence threshold, whereas UPS [42] relies on uncertainty measures. $\\mathrm{ST}++$ [58] focuses on selecting unlabeled images more likely to yield accurate pseudo-pixels. Emerging approaches like U2PL [50] explore the idea of learning reliable pseudo-labels from their less reliable counterparts. CFCG [29] introduced a novel approach that combines cross-fusion and contour guidance for pseudo-label improvement. We differ from these methods by explicitly introducing patch-level contextual information in the SSS methods.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consistency Regularization: This approach enforces the model to have consistent predictions on different views of the unlabeled images. UniMatch [57] uses dual perturbations and feature perturbations to force the network to learn more consistent information. ICT [49] uses Mixup [64] augmentation for consistency regularization. Recent methods [14, 58, 62] use Cutout [9], Cutmix [63], Class-mix [37] as strong data augmentation. Consistency Regularization is used with pseudo-labeling techniques in Mixmatch [4] and TC-SSL [67]. In our work, we use a set of strong and weak data augmentations [1, 65] for generating pseudo-labels.", + "bbox": [ + 212, + 268, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-label classification: This approach classifies an image into more than one label present in the image. Most works in this field explore finding the correlation between the labels and the images [21,39,61]. Recent works mainly focus on solving data imbalance issues [17,51]. We use multi-label classification to help in segmentation. Some works [15,40] use multi-label classification for segmentation in weakly supervised scenarios, for improved CAM-based pseudo-segmentation. However, we are using multi-label classification in a semi-supervised setting to introduce patch-level contextual information and also use the classifier confidence to reduce the impact of noisy pseudo-labels.", + "bbox": [ + 212, + 388, + 787, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Semi-Supervised Segmentation Pipeline", + "text_level": 1, + "bbox": [ + 215, + 542, + 635, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a labeled set $\\mathcal{D}^l = \\{(x^l,y^l)\\}$ and an unlabeled set $\\mathcal{D}^u = \\{x^u\\}$ , the goal of semi-supervised semantic segmentation is to develop a model that effectively utilizes both labeled and unlabeled data. This section commences with an overview of the foundational framework for semi-supervised segmentation in Section 3.1. followed by an overview of our proposed MPMC in Section 3.2. The section concludes with an elaboration of MPMC's functionality (Section 3.3).", + "bbox": [ + 212, + 571, + 787, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 681, + 361, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Teacher-Student frameworks leverage weak-to-strong consistency regularization to leverage unlabeled data. As depicted in Fig. 2(a), each unlabeled image $x^{u}$ is simultaneously perturbed by two operators, i.e., weak perturbation $A^w$ (cropping, random flip, etc), and strong perturbation $A^s$ (color jitter, gaussian blur, contrast, etc). The framework's objective function is a convex combination of supervised loss $\\mathcal{L}_l$ for the labeled data and unsupervised loss $\\mathcal{L}_u$ for the unlabeled data as:", + "bbox": [ + 212, + 702, + 787, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {l} + \\alpha \\mathcal {L} _ {u} \\quad \\text {w h e r e ,} \\quad \\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum \\mathbb {1} \\left(p ^ {w} \\geq t h\\right) \\mathcal {E} \\left(p ^ {w}, p ^ {s}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 813, + 785, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Howlader et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9964a88ca55309adb44c838c691a9ad905479a6b4211da23387a4297b5e521e.jpg", + "image_caption": [ + "Fig. 2: Overall Pipeline of our novel multi-label classification based semantic segmentation: (a) Left: End-to-end Teacher-Student Pipeline with our novel method Multi-scale Patch-based Multi-label Classifier (MPMC). (b) Right: For unlabeled images, The teacher MPMC extracts features from a layer in the segmentation model's encoder to classify the feature's receptive field patch. The confidence of MPMC for a class in a patch is used to calculate two adaptive weights $\\lambda_{s}$ and $\\lambda_{m}$ which is used to reduce the influence of noisy predictions in that patch from the teacher to train both the student segmentation network and MPMC." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 493, + 299 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/33e39ec217d41dfae56c59a380a7ca5e742d5a2dacce2980edd14482500e8734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 142, + 787, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\mathcal{L}_l$ is the standard cross-entropy loss, comparing the model's predictions with ground truth. While $\\mathcal{L}_u$ aims to regularize the predictions for an unlabeled sample so that they remain consistent under both weak and strong augmentations ( $p^w$ and $p^s$ ). Here, $th$ is a predefined confidence threshold to filter noisy pseudo labels and $\\mathcal{E}$ minimizes the entropy between two probability distributions.", + "bbox": [ + 212, + 452, + 787, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Overview", + "text_level": 1, + "bbox": [ + 215, + 554, + 339, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2(a) illustrates the comprehensive training process of our proposed MPMC, which is tailored to enhance the discrimination of patches among different classes and to mitigate the impact of noisy predictions transmitted from the teacher to the student. MPMC can be integrated into both teacher and student models within the teacher-student framework. It is composed of a classifier that performs multi-label classifications of the multi-scale patch representations of the input image. The application of MPMC spans two distinct scenarios: 1) For labeled images, MPMC assigns classes to each patch representation based on the classes present within a patch. The multi-scale nature of these patch representations and their classification enhances the segmentor's ability to discriminate classes by capturing enhanced contextual information related to variations in shape and size. 2) In the case of unlabeled images, besides patch classification, MPMC computes a pseudo-label weight map to alleviate the impact of the teacher's noisy pseudo-label supervision on the student. This involves assigning weights to the pseudo-labels within each patch in consistency learning framework, which are determined by the confidence of the MPMC's predictions for each class in the patch, to enhance the reliability and accuracy of the pseudo-labeling process.", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Beyond Pixels", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Multi-scale Patch-based Multi-label Classifier (MPMC)", + "text_level": 1, + "bbox": [ + 215, + 146, + 718, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed MPMC module is specifically designed to address the shortcomings of SSS pipeline. MPMC can be plugged in at any layer within the teacher-student network, both within the teacher model, denoted as $g$ , and within the student model, denoted as $\\hat{g}$ , they share the same parameters. Below, we detail the training of MPMC on labeled images and then on unlabeled images.", + "bbox": [ + 212, + 171, + 787, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MPMC on Labeled Images. As illustrated in Fig. 2, MPMC incorporates patch-supervision into the segmentation framework, by training a multi-label classifier that takes as input a single pixel in the feature map of the $\\tau^{th}$ layer of the encoder. This single pixel-representation encodes the information from a patch (receptive field) in the original image. Further, to encapsulate context at varying scales we perform average pooling on this features map with kernels of different sizes before performing multi-label classification. Let us define a patch of image and the corresponding segmentation in a receptive field $r$ as $\\{x_r^l,y_r^l\\} \\in \\mathcal{D}^l$ . Note that $y_{r}^{l}$ is a multi-label target where the classes present in the receptive field $r$ are positive classes, and the classes not present are negative classes.", + "bbox": [ + 212, + 268, + 787, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The input to the student MPMC $\\hat{g}$ are the features extracted from the $\\tau^{th}$ layer of the encoder in student segmentation network $\\hat{f}$ , denoted by $z_{r} = \\hat{f}_{\\tau}(\\mathcal{A}_w(x_r^l))$ . MPMC captures the contextual information within a patch at different scales, $S_{i}$ , $i \\in \\{0,1\\}$ , where $s_0$ and $s_1$ are average pool kernel sizes. For all our experiments $s_0$ and $s_1$ are set to $7 \\times 7$ and $19 \\times 19$ . These pooling operations are performed with appropriate padding to maintain uniformity in the shape of the output features. MPMC $\\hat{g}$ is composed of the concatenation of these multiscale pooled features, followed by a convolutional network $\\mathcal{M}$ and a final linear layer. Thus, the output of the MPMC $\\hat{g}$ for the input patch $x_{r}^{l}$ is given by:", + "bbox": [ + 212, + 420, + 787, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {l} = \\sigma \\left(W ^ {T} \\mathcal {M} \\left(\\operatorname {c o n c a t} \\left(z _ {r}, z _ {r} ^ {S _ {0}}, z _ {r} ^ {S _ {1}}\\right)\\right) + b\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 566, + 784, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W$ and $b$ are linear layer weights and bias, $\\sigma$ is the sigmoid activation and $q^l = [q_1^l,\\dots,q_C^l]^T\\in \\mathbb{R}^C$ are the predicted probabilities for $C$ classes.", + "bbox": [ + 212, + 594, + 785, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While binary cross-entropy loss is applicable to train MPMC, it indiscriminately weighs all positive and negative classes equally. Considering the receptive field of a feature from the $\\tau^{th}$ layer of the encoder in the segmentation network $\\hat{f}$ covers only a small patch relative to the input image's size, the prevalence of negative classes within a patch significantly exceeds that of positive classes. This imbalance influences the binary cross-entropy loss to favor negative classes [51]. Inspired by long-tailed multi-label classification [30, 51] which mitigate such imbalances by diminishing the weight of negative classes, we adapt an asymmetric loss [41], a focal loss variant that assigns distinct $\\gamma$ values to positive and negative classes. For an image patch $x_{r}^{l}$ and the predicted class probabilities $q^{l}$ using our framework, the asymmetric focal loss is expressed as:", + "bbox": [ + 212, + 626, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nl _ {r} ^ {M} = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\left\\{ \\begin{array}{l l} (1 - q _ {c} ^ {l}) ^ {\\gamma^ {+}} \\log (q _ {c} ^ {l}), & y _ {r, c} ^ {l} = 1. \\\\ (q _ {c} ^ {l}) ^ {\\gamma^ {-}} \\log (1 - q _ {c} ^ {l}), & y _ {r, c} ^ {l} = 0. \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 801, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Howlader et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, we set $\\gamma^{+} = 0$ and $\\gamma^{-} = 1$ . Subsequently, MPMC $\\hat{g}$ undergoes training with multi-label supervision $\\mathcal{L}_l^M$ , formulated as:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l} ^ {M} = \\frac {1}{| D ^ {l} |} \\sum_ {\\left(x ^ {l}, y ^ {l}\\right) \\in D ^ {l}} \\sum_ {r = 1} ^ {R} l _ {r} ^ {M} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 191, + 785, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{R}$ is the number of receptive fields at $\\tau^{th}$ layer of the segmentor. This auxiliary task of patch classification augments the network's proficiency in learning and preserving class features, a critical function particularly pertinent as the fidelity of these features tends to diminish in the network's deeper layers.", + "bbox": [ + 212, + 248, + 784, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MPMC on Unlabeled images. In semi-supervised teacher-student framework, for unlabeled images, the prediction of the teacher network $f$ for the weak augmented view $\\mathcal{A}^w(x^u)$ is used to supervise the student network $\\hat{f}$ for the strong augmented view $\\mathcal{A}^s(x^u)$ . In the early iterations of semi-supervised learning, the teacher's predictions are inherently noisy. Common practices [13, 20, 33, 46, 70], involve employing pixel-wise confidence score thresholds (th) to filter out unreliable pseudo labels. However, a high confidence threshold does not entirely eliminate noise due to the teacher's initial miscalibration [42], allowing some noisy predictions to have high confidence. To refine this process, our approach does not solely depend on confidence-thresholded teacher segmentation predictions for reliable pseudo-label extraction. Instead, we leverage the teacher MPMC $g$ 's confidence for a patch to adjust the weights of the deemed reliable pseudo-labels within that patch. The underlying principle of our methodology is that MPMC exhibits greater uncertainty (lower confidence) for noisy labels.", + "bbox": [ + 212, + 335, + 787, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ the teacher MPMC $g$ 's predictions to define two adaptive learning weights $\\lambda_{m} \\in \\mathbb{R}^{R \\times C}$ and $\\lambda_{s} \\in \\mathbb{R}^{W \\times H}$ , where $R$ denotes the number of patches, sized according to the receptive field within the image, $W$ and $H$ represents the width and height of the input image, respectively. The learning weight $\\lambda_{m}$ is utilized to assign class weights to each patch for training the student MPMC, whereas $\\lambda_{s}$ is applied to diminish the impact of noisy predictions during training the student segmentation network.", + "bbox": [ + 212, + 547, + 787, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For an unlabeled image $x^{u}$ , consider a patch at the receptive field $r$ , denoted as $x_{r}^{u}$ . The teacher and student segmentation networks yield predictions $p^{w} = f(\\mathcal{A}^{w}(x_{r}^{u}))$ and $p^{s} = \\hat{f} (\\mathcal{A}^{s}(x_{r}^{u}))$ , respectively. Simultaneously, the teacher and student MPMC predict $q_{r}^{u} = g(f_{\\tau}(\\mathcal{A}^{w}(x_{r}^{u})))$ and $\\hat{q}_r^u = \\hat{g} (\\hat{f}\\tau (\\mathcal{A}^s (x_r^u)))$ using features from the $\\tau^{th}$ layer of the encoder. Let us define the confidence of teacher MPMC $g$ across all classes in patch $r$ as $\\gamma = \\mathrm{softmax}(q_r^u)$ , with $\\gamma \\in \\mathbb{R}^C$ . The weight of the patch $r$ , $\\lambda_{m,r}$ is determined by $\\gamma$ . For each pixel in patch $r$ associated with class $c$ , $\\lambda_{s}$ is assigned the weight $\\gamma_{c}$ . This approach assigns weights to pseudo-labels and pixels based on the teacher MPMC's confidence, enhancing the accuracy of the learning process.", + "bbox": [ + 212, + 655, + 787, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thus, the control parameters $\\lambda_{s}$ and $\\lambda_{m}$ are integrated to the unsupervised losses to train the student segmentation network and MPMC. The unsupervised", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Beyond Pixels", + "bbox": [ + 633, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "segmentation loss $L_{u}$ mentioned in equation 1 to train the student segmentation network $\\hat{f}$ is re-formulated as:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum_ {x ^ {u} \\in \\mathcal {D} ^ {u}} \\mathbb {1} \\left(p ^ {w} \\geq t h\\right) \\lambda_ {s} \\mathcal {E} \\left(p ^ {w}, p ^ {s}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 186, + 785, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ is the standard cross entropy loss and $th$ is the predefined segmentation confidence threshold. Finally, the unsupervised multi-label loss $\\mathcal{L}_u^M$ to train the student MPMC $\\hat{g}$ is given by:", + "bbox": [ + 212, + 233, + 782, + 280 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} ^ {M} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum_ {x ^ {u} \\in D ^ {u}} \\sum_ {r = 1} ^ {R} \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\lambda_ {m, r} ^ {i} \\mathcal {E} ^ {+} \\left(q _ {r} ^ {u}, \\hat {q} _ {r} ^ {u}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 291, + 785, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{E}^+$ is the binary cross entropy loss. Therefore, the overall loss function is defined as:", + "bbox": [ + 212, + 343, + 782, + 373 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {l} + \\mathcal {L} _ {l} ^ {M} + \\alpha \\mathcal {L} _ {u} + \\beta \\mathcal {L} _ {u} ^ {M} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 388, + 785, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ controls the contribution of $\\mathcal{L}_u$ and $\\mathcal{L}_u^M$ . In all our experiments $\\alpha$ and $\\beta$ are set to 0.1 and 0.25 respectively.", + "bbox": [ + 212, + 411, + 782, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 465, + 375, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 214, + 497, + 310, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets: PASCAL VOC 2012 [12] is a standard semantic segmentation dataset with 20 semantic classes and a background class. The training, validation, and testing sets consist of 1464, 1449, and 1556 images (Classic), respectively. Following [7, 65], we also include the additional augmented dataset [18] (Blender), which includes a collection of 10582 training images. We adopt the same partition protocols in [7, 65] to evaluate our method in both Classic and Blender sets. Cityscapes [8] dataset is designed for urban scene understanding, consisting of 30 classes, of which only 19 classes are used for scene parsing evaluation. It has 2975, 500, and 1525 images for training, validation, and testing, respectively. ACDC [3] dataset contains 200 annotated short-axis cardiac cineMR images from 100 patients. The segmentation masks consists of classes: left ventricle (LV), myocardium (Myo), and right ventricle (RV). 140 images from 70 patients, 60 images from 30 patients are randomly selected for training and validation respectively.", + "bbox": [ + 212, + 522, + 787, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details: For fair comparison, we adopt DeepLabv3+ [6] as the decoder in all of our experimental setups, and compare with both ResNet-101 and ResNet-50 [19] as the backbone architecture. In all our experiments with Cityscapes, Pascal, and ACDC datasets, the training resolution is 801, 513, and 224 respectively. Note, for maintaining fair comparison, our method \"UniMatch [57] + MPMC\" uses the training resolution of 321 which is used by UniMatch [57] when training on the Classic set of Pascal VOC. Further when training a MPMC", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Howlader et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "integrated baseline, we use the same weak and strong augmentations as used by the corresponding baselines. The convolutional network in MPMC is comprised of three convolution blocks. Each convolution block consists of a stack of two $3 \\times 3$ convolutions followed by a $1 \\times 1$ convolution. For all our experiments MPMC is integrated to the first layer of the ResNet Encoder.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 214, + 263, + 620, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conducted our experiments on PASCAL VOC 2012 and Cityscapes which are two popular natural image benchmarks, and a medical dataset ACDC. MPMC is integrated in four semi supervised methods UniMatch [57], AEL [20], U2PL [50] and AugSeg [65]. Note, UniMatch [57] is a consistency regularization based method. We observe that MPMC consistently improves over their baseline performance across all partition protocols on all datasets.", + "bbox": [ + 212, + 306, + 787, + 397 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results on PASCAL VOC 2012 Dataset. Table 1 shows the comparison with other SOTA methods for both the Classic and Blender sets. MPMC integrated to the four baseline methods improves them consistently across all data partitions for both Classic and Blender sets. We find that our method brings the most improvement for the least labeled data partition $(\\frac{1}{16})$ , improving AugSeg, AEL, U2PL and UniMatch by $1.8\\%$ , $2.5\\%$ , $2.0\\%$ and $2.1\\%$ respectively in Classic with a ResNet-101 encoder.", + "bbox": [ + 212, + 398, + 787, + 502 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results on Cityscapes Dataset. Table 2 shows the comparison with other SOTA methods. MPMC integrated to the four baseline methods improves them consistently across all data partitions. We observe that similar to results in PASCAL VOC 2012 dataset, our method brings the most improvement for the least labeled data partition $(\\frac{1}{16})$ , improving AugSeg, AEL, U2PL and UniMatch by $1.7\\%$ , $2.1\\%$ , $2.3\\%$ and $1.9\\%$ respectively in Classic for ResNet-101 encoder.", + "bbox": [ + 212, + 503, + 787, + 593 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results on ACDC Dataset. Table 3 shows the comparison with other SOTA methods. We observe that when MPMC is integrated to the four baseline methods, it significantly improves the performance for the \"3 cases\" data partition which has the least labeled images ( $< 10\\%$ of labeled images). Note we use ResNet-101 for all experiments in this dataset.", + "bbox": [ + 212, + 594, + 787, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The consistent improvement in SSS performance in both natural and medical datasets and its ability to integrate in different SSS methods substantiates the robustness of MPMC. Moreover, the notable improvements achieved in scenarios with limited data shows MPMC's vital efficacy within SSS pipelines.", + "bbox": [ + 212, + 670, + 787, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative Results In Fig. 3 we visualise the results of integration of our method (MPMC) to AugSeg [65]. Introducing patch-level supervision with the integration of MPMC, leads the SSS model to better segment confusing classes. Among comparisons we show that MPMC leads to better segmentation of pixels regions of confusing classes like bus, train and truck. These classes are often misclassified not only because of sharing visual similarity but also having low training instances.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Beyond Pixels", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f047d3a1dc07cc9d0bb857ea05c9dbd360be1f489e7eb1e6f4446e343ee70969.jpg", + "table_caption": [ + "Table 1: Quantitative results of different semi-supervised segmentation methods on Pascal VOC classic and blender set. We report Mean IoU under various partition protocols and show the improvements $(\\Delta)$ over corresponding baseline." + ], + "table_footnote": [], + "table_body": "
MethodClassicBlender
1/161/81/41/2Full1/161/81/4
ResNet-50
Supervised44.052.361.766.772.9---
PC2Seg [66][CVPR'21]56.964.667.670.972.3---
PseudoSeg [69][ICLR'21]56.964.667.670.972.2---
ST++ [58][CVPR'22]-----72.674.475.4
AugSeg [65][CVPR'21]64.272.176.177.478.877.278.278.2
AugSeg [65] + MPMC (Δ)65.973.777.678.579.779.079.879.7
(+1.7)(+1.6)(+1.5)(+1.1)(+0.9)(+1.8)(+1.6)(+1.5)
AEL [20][NeurIPS'21]62.964.170.372.774.074.176.177.9
AEL [20] + MPMC (Δ)65.066.172.174.275.376.178.079.6
(+2.1)(+2.0)(+1.8)(+1.5)(+1.3)(+2.0)(+1.9)(+1.7)
U2PL [50][CVPR'22]63.365.571.673.875.174.777.477.5
U2PL [50] + MPMC (Δ)65.267.473.375.276.276.579.179.1
(+1.9)(+1.9)(+1.7)(+1.4)(+1.1)(+1.8)(+1.7)(+1.6)
UniMatch [57][CVPR'23]71.972.576.077.478.778.179.079.1
UniMatch [57] + MPMC (Δ)73.574.077.478.579.479.680.480.4
(+1.6)(+1.5)(+1.4)(+1.1)(+0.7)(+1.5)(+1.4)(+1.3)
ResNet-101
Supervised45.155.364.869.773.570.675.076.5
CPS [7][CVPR'21]64.167.471.775.9-72.275.877.6
PS-MT [33][CVPR'22]65.869.676.678.480.075.578.278.7
PCR [52][NeurIPS'22]70.174.777.278.580.778.680.780.8
DAW [47][NeurIPS'23]74.877.479.580.681.578.578.979.6
CFCG [29][ICCV'23]-----77.479.480.4
AugSeg [65][CVPR'21]71.175.578.880.381.479.381.580.5
AugSeg [65] + MPMC (Δ)72.977.879.981.382.180.882.581.4
(+1.8)(+2.3)(+1.1)(+1.0)(+0.7)(+1.5)(+1.0)(+0.9)
AEL [20][NeurIPS'21]66.168.371.974.478.977.277.678.1
AEL [20] + MPMC (Δ)68.670.373.576.180.179.279.379.2
(+2.5)(+2.0)(+1.6)(+1.7)(+1.2)(+2.0)(+1.7)(+1.1)
U2PL [50][CVPR'22]68.069.273.776.279.577.279.079.3
U2PL [50] + MPMC (Δ)70.070.875.377.880.579.280.780.3
(+2.0)(+1.6)(+1.6)(+1.6)(+1.0)(+2.0)(+1.7)(+1.0)
UniMatch [57][CVPR'23]75.277.278.879.981.280.981.980.4
UniMatch [57] + MPMC (Δ)77.378.679.880.881.782.583.181.5
(+2.1)(+1.4)(+1.0)(+0.9)(+0.5)(+1.6)(+1.2)(+1.1)
", + "bbox": [ + 222, + 282, + 779, + 752 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Howlader et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bed05dbb3298e38f171f9ff4cd49a43ee97729d555d8f7cf552826fe975d7199.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 157, + 359, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/60cf4c0ef9b134097920c04cfa6f34a4885eec20acdae5a67d979188cf0351b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 222, + 357, + 284 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cc9a1e4163be66a3ce0f31047a6e3f451313d4f6fc5a954160c408fc210f94cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 285, + 357, + 347 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/adbb23c58d04ed1cfbad440b34db33fc59801a9e84ccfb40a6d51ab6fbd66619.jpg", + "image_caption": [ + "(a) Input Image" + ], + "image_footnote": [], + "bbox": [ + 218, + 348, + 359, + 410 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2e3599f5d7734cff4daa0062343dadfdcc3d7c89ace091c8cd097818991febef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 157, + 500, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/052331a1214c8da78af967a541f7a03ecf928ae1722e41867d3b71024951022d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 222, + 500, + 284 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7906a90b19af6415d8d8372f9ffe0bbe48e598f9eeecc46b57c34b3cef36dd07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 285, + 500, + 347 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/35a09bf93191eae0bf507e428bab2f6620770ed3d7c3e28256b377da6b53d469.jpg", + "image_caption": [ + "(b) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 361, + 348, + 500, + 410 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/844d0b6f789c3f408683c58dc93226c9aeea6adac0bede4e7f11677457667c15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 157, + 640, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1d6be199f70055a0cc81f08134cf7299523a01fad44d2bd9058e05a82ef528e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 222, + 640, + 284 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7501babaf86065a20750c0c62afbb5223dd0b780095956b969ce1a55f698f626.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 285, + 640, + 347 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1d8afd4a8d9642127033b56d7fae319d346e519153a63c78e08fa67785895d1d.jpg", + "image_caption": [ + "(c) AugSeg", + "Fig. 3: Qualitative Results on Cityscapes dataset: (a) is the original image, (b) is the Ground Truth, (c) are segmentations generated by AugSeg [65] compared to (d) which are segmentations generated when our method (MPMC) is integrated to AugSeg. White boxes show the areas where our method improves the baseline." + ], + "image_footnote": [], + "bbox": [ + 501, + 348, + 640, + 410 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/719b8463a09340c28f2f15e8434a4c25e60bca204b0dc80adf13545bddb8f90b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 157, + 784, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e35b9b096407e70c93ec3557f151f3d1340a987b148d6e5004c0a5bfaa5510cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 222, + 782, + 284 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d6587c8537d1345a9f4b1b6b452e85c51b7f9f49b60bfaf146375f59bd05a29a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 285, + 782, + 347 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/04a39c479d40f504f57cc89b08e1f8694e445c62790a1f6a13c130c09c57b9f2.jpg", + "image_caption": [ + "(d) AugSeg + MPMC" + ], + "image_footnote": [], + "bbox": [ + 642, + 348, + 782, + 410 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3c76d1a645d9df94852122794575bb28d4e8e96d3d521e7299874588907655c7.jpg", + "table_caption": [ + "Table 2: Quantitative results of different semi-supervised segmentation methods on Cityscapes validation set. We report Mean IoU under various partition protocols and show the improvements $(\\Delta)$ over the corresponding baseline." + ], + "table_footnote": [], + "table_body": "
MethodResNet-50ResNet-101
1/161/81/41/21/161/81/41/2
Supervised63.3468.7374.1476.6266.372.875.078.0
CPS [7][CVPR'21]69.7974.3976.8578.6469.874.374.676.8
PS-MT [33][CVPR'22]-75.7676.9277.64-76.977.679.1
PCR [52][NeurIPS'22]----73.476.378.479.1
CFCG [29][ICCV'23]76.178.979.380.177.879.680.480.9
AugSeg [65][CVPR'21]73.776.478.779.375.277.879.680.4
AugSeg [65] + MPMC (Δ)75.3 (+1.6)77.7 (+1.3)79.8 (+1.1)80.3 (+1.0)76.9 (+1.7)79.2 (+1.4)80.9 (+1.3)81.3 (+0.9)
AEL [20][NeurIPS'21]68.272.774.977.574.575.677.579.0
AEL [20] + MPMC (Δ)70.5 (+2.3)74.5 (+1.8)76.6 (+1.7)79.0 (+1.5)76.6 (+2.1)77.5 (+1.9)78.9 (+1.4)80.3 (+1.3)
U2PL [50][CVPR'22]69.073.076.378.674.976.578.579.1
U2PL [50] + MPMC (Δ)71.2 (+2.2)74.6 (+1.6)77.8 (+1.5)79.8 (+1.2)77.2 (+2.3)78.2 (+1.7)80.0 (+1.5)80.0 (+0.9)
UniMatch [57][CVPR'23]75.076.877.578.676.677.979.279.5
UniMatch [57] + MPMC (Δ)76.8 (+1.8)78.0 (+1.2)78.6 (+1.1)79.5 (+0.9)78.5 (+1.9)79.2 (+1.3)80.9 (+1.7)81.3 (+1.8)
", + "bbox": [ + 215, + 580, + 787, + 821 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Beyond Pixels", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/41447403b29a4af0acabf42a67726f8f0f15200c43eabd017ddc1921c8095e22.jpg", + "table_caption": [ + "Table 3: Quantitative results of different semi-supervised segmentation methods on ACDC dataset with 3/7 labeled cases. We report Dice Similarity Coefficient (DSC) metric averaged over 3 classes. We use ResNet-101 for all experiments in this dataset." + ], + "table_footnote": [], + "table_body": "
Method3 cases7 cases
Supervised41.562.5
UA-MT [60][MICCAI'19]61.081.5
CPS [7][CVPR'21]60.383.3
CNN & Trans [34][PMLR'22]65.686.4
AugSeg [65][CVPR'21]85.187.3
AugSeg [65] + MPMC (Δ)88.6 (+3.5)88.4 (+1.1)
AEL [20][NeurIPS'21]82.787.6
AEL [20] + MPMC (Δ)85.4 (+2.7)88.5 (+0.9)
U2PL [50][CVPR'22]84.289.2
U2PL [50] + MPMC (Δ)87.4 (+3.2)90.5 (+1.3)
UniMatch [57][CVPR'23]88.989.9
UniMatch [57] + MPMC (Δ)91.8 (+2.9)90.4 (+0.5)
", + "bbox": [ + 357, + 199, + 640, + 388 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 419, + 401, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct experiments to study the impact of various components of our approach. All experiments in this section were conducted on $\\frac{1}{16}$ data partition of the Cityscapes dataset, and use AugSeg [65] as the baseline.", + "bbox": [ + 212, + 446, + 785, + 492 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Segmentation performance at different scales. Here we analyse how scaling of input features for patch-classification, improves the performance of the SSS method. We consider two scales - average pool kernels $7 \\times 7$ and $19 \\times 19$ . Fig. 4 illustrates for two random classes pole and rider, that average pooling with a kernel size of $7 \\times 7$ is better at enhancing the pixel accuracy for instances with smaller areas, whereas the $19 \\times 19$ kernel size is better at improving accuracy for instances with larger areas. This can be attributed to the larger receptive field of the $19 \\times 19$ kernel, enabling more effective contextual learning for larger regions. While the smaller receptive field of $7 \\times 7$ kernel is better able to discern smaller objects.", + "bbox": [ + 212, + 492, + 785, + 643 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analysing class features The crux of our method is to better discriminate pixel regions between different classes. Here we analyze if MPMC integrated to AugSeg [65] can better discriminate features than the baseline for different classes in Cityscapes Dataset. We use the pixel features from the penultimate layer of the segmentation network to analyse via t-SNE. Fig 5 illustrates for four random classes - car, bus, train and truck, integration of MPMC leads to better discrimination between the classes, i.e., compared to AugSeg [65], the features of train and truck are more clustered and separated from the rest.", + "bbox": [ + 212, + 643, + 785, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Influence of MPMC on patch-level segmentation. Our method MPMC improves segmentation by introducing patch classification as an auxiliary task. We analyse at patch level the correlation between multi-label accuracy and the Mean IoU performance of the segmentation network. We use (1 - hamming loss) [48] as a measure of multi-label classification and use the basic teacher-", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Howlader et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5d3bccfe2eb4b7b2bad596ca3c896c83c810ee05e7923e4da7d8335435c1f3f5.jpg", + "image_caption": [ + "Fig. 4: Pixel accuracy across different scales with respect to instance size. For two different scales, we compare the pixel accuracy with respect to size of instances for two random classes: pole and rider." + ], + "image_footnote": [], + "bbox": [ + 267, + 142, + 498, + 328 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/200a277845cb106a9a3c7de7a37c13b633c4a9c85a68fb5915484fc33509372f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 142, + 735, + 328 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dbc537f6b47ae7aab8d16e8d61b3f3a63b42315d12ae7d28b493dca4d52bb9da.jpg", + "image_caption": [ + "Fig. 5: Analysis of class features (a) t-SNE of features predicted by AugSeg [65], (b) t-SNE of features predicted by AugSeg [65] + MPMC." + ], + "image_footnote": [], + "bbox": [ + 277, + 393, + 723, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "student framework mentioned in Section 3.1 as our baseline. Table 4 illustrates that the avg MIoU improvement is highest (4.05%) in patches having more than 90% multi-label classification accuracy showing that MPMC improves segmentation at the patch level.", + "bbox": [ + 212, + 627, + 784, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of MPMC on semantic segmentation in limited labeled data scenarios.", + "bbox": [ + 212, + 686, + 784, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Table 5 shows the impact of MPMC on segmentation when trained with only labeled data. We use the basic teacher-student framework mentioned in Section 3.1 as our baseline, which achieves MIoU of 66.3, 72.8 and 75.0 in 186, 372 and 744 partitions of labeled images respectively. After introducing PMC, there is an improvement of $1.9\\%$ , $1.5\\%$ and $1.3\\%$ in 186, 372 and 744 partitions of labeled images respectively. We observe that MPMC improves segmentation performance in limited-labeled data scenarios, with the improvement highest for least labeled data, i.e., $1.9\\%$ for 186 labeled images.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Beyond Pixels", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/03c0b0147e3f9971f4090d00062a8d3197739580c357b4a9165018de84e3c09b.jpg", + "table_caption": [ + "Table 4: Analysis of the influence of MPMC on segmentation at patch-level." + ], + "table_footnote": [], + "table_body": "
1 - hamming loss90+80-9070-8060-7050-6040-5030-4020-3010-200-10
% patches12.625.220.58.45.15.24.110.22.46.3
Avg M IoU Δ+4.05+1.57+0.88+1.55+0.91+1.15+0.49+0.10-1.90-2.20
", + "bbox": [ + 215, + 171, + 785, + 238 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/561bb522d34048178f7d43415401d343e334b051c73ebedbb9e6949c055c2321.jpg", + "table_caption": [ + "Table 5: Impact of MPMC on segmentation (MIoU) when trained with different volumes of labeled data. MPMC shows the highest improvement $(1.9\\%)$ in the least labeled data partition." + ], + "table_footnote": [], + "table_body": "
# Labeled Data
186372744
w/o MPMC66.372.875.0
w/ MPMC68.274.376.3
Gain/Δ+1.9+1.5+1.3
", + "bbox": [ + 243, + 333, + 444, + 407 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/b943ae342a5d21d728a5890672024142f63dd68dedd8b6e17182163c7ac150e4.jpg", + "table_caption": [ + "Table 6: Ablation study of different component: MPMC on labeled images (MPMCL) and MPMC on unlabeled images (MPMCUL)" + ], + "table_footnote": [], + "table_body": "
MPMCLMPMCUL1/16 (186)1/8 (372)1/4 (744)
68.473.574.2
✓70.875.475.7
✓71.376.076.6
✓✓73.176.777.3
", + "bbox": [ + 511, + 320, + 769, + 409 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Analysis of different components We ablate each component of MPMC with basic teacher-student framework (Section 3.1) as the baseline. MPMC on labeled images (MPMCL), which introduces multi-scale patch-based classification on only the labeled data, achieves an improvement of $2.4\\%$ , $1.9\\%$ , and $1.5\\%$ under $\\frac{1}{16}$ , $\\frac{1}{8}$ and $\\frac{1}{4}$ partition protocols respectively. MPMC on unlabeled images (MPMCUL) leads to a further improvement of $2.9\\%$ , $2.5\\%$ , and $2.4\\%$ under $\\frac{1}{16}$ , $\\frac{1}{8}$ and $\\frac{1}{4}$ partition protocols respectively. Thus, MPMC is effective on both labeled and unlabeled instances in SSS pipelines.", + "bbox": [ + 215, + 436, + 787, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 580, + 359, + 597 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel Multi-scale Patch-based Multi-label Classifier (MPMC), which introduces patch-level context information in the form of an auxiliary task of multi-scale patch-based multi-label classification. When jointly trained with a SSS model, the patch classifier provides patch-level labels, identifying the classes present within a region. This facilitates the elimination of distractors and enhances the classification of small object segments by the baseline method. Further, the MPMC predictions are used to learn an adaptive weight that reduces the influence of noisy pseudo-labels in the teacher-student framework. We performed a thorough analysis of the importance of multi-scale patch-level contextual information in improving segmentation especially in low data scenarios. Further, we demonstrated the efficacy of our proposed parameterized module MPMC by integrating it into four SSS methodologies and showing consistent improvements in Cityscapes, Pascal VOC and ACDC benchmarks.", + "bbox": [ + 212, + 613, + 785, + 809 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work was partially supported by the National Science Foundation (IIS-2245652, IIS-2123920, IIS-2212046).", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Howlader et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alonso, I., Sabater, A., Ferstl, D., Montesano, L., Murillo, A.C.: Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8219-8228 (2021)", + "2. Bachman, P., Alsharif, O., Precup, D.: Learning with pseudo-ensembles. Advances in neural information processing systems 27 (2014)", + "3. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37(11), 2514-2525 (2018)", + "4. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems 32 (2019)", + "5. Chen, H., Jin, Y., Jin, G., Zhu, C., Chen, E.: Semisupervised semantic segmentation by improving prediction confidence. IEEE Transactions on Neural Networks and Learning Systems (2021)", + "6. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018)", + "7. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021)", + "8. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)", + "9. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)", + "0. Durasov, N., Dorndorf, N., Le, H., Fua, P.: Zigzag: Universal sampling-free uncertainty estimation through two-step inference. Transactions on Machine Learning Research (2024)", + "1. Durasov, N., Oner, D., Donier, J., Le, H., Fua, P.: Enabling uncertainty estimation in iterative neural networks. In: Forty-first International Conference on Machine Learning (2024)", + "2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88(2), 303-338 (2010)", + "3. Fan, J., Gao, B., Jin, H., Jiang, L.: Ucc: Uncertainty guided cross-head co-training for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9947-9956 (2022)", + "4. French, G., Laine, S., Aila, T., Mackiewicz, M., Finlayson, G.: Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916 (2019)", + "5. Ge, W., Yang, S., Yu, Y.: Multi-evidence filtering and fusion for multi-label classification, object detection and semantic segmentation based on weakly supervised learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1277-1286 (2018)" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Beyond Pixels", + "bbox": [ + 633, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. Advances in neural information processing systems 17 (2004)", + "17. Guo, H., Wang, S.: Long-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15089-15098 (2021)", + "18. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011)", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "20. Hu, H., Wei, F., Hu, H., Ye, Q., Cui, J., Wang, L.: Semi-supervised semantic segmentation via adaptive equalization learning. Advances in Neural Information Processing Systems 34, 22106-22118 (2021)", + "21. Huang, J., Li, G., Huang, Q., Wu, X.: Learning label specific features for multi-label classification. In: 2015 IEEE International Conference on Data Mining. pp. 181-190. IEEE (2015)", + "22. Huo, X., Xie, L., He, J., Yang, Z., Zhou, W., Li, H., Tian, Q.: Atso: Asynchronous teacher-student optimization for semi-supervised image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1235-1244 (2021)", + "23. Ke, Z., Qiu, D., Li, K., Yan, Q., Lau, R.W.: Guided collaborative training for pixel-wise semi-supervised learning. In: European conference on computer vision. pp. 429-445. Springer (2020)", + "24. Le, H., Goncalves, B., Samaras, D., Lynch, H.: Weakly labeling the antarctic: The penguin colony case. In: CVPRW (June 2019)", + "25. Le, H., Samaras, D., Lynch, H.J.: A convolutional neural network architecture designed for the automated survey of seabird colonies. Remote Sensing in Ecology and Conservation 8(2), 251-262 (2022)", + "26. Le, H., Vicente, T.F.Y., Nguyen, V., Hoai, M., Samaras, D.: A+D Net: Training a shadow detector with adversarial shadow attenuation. In: European Conference on Computer Vision (ECCV) (2018)", + "27. Le, H., Yu, C.P., Zelinsky, G., Samaras, D.: Co-localization with category-consistent features and geodesic distance propagation. In: ICCVW (2017)", + "28. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896 (2013)", + "29. Li, S., He, Y., Zhang, W., Zhang, W., Tan, X., Han, J., Ding, E., Wang, J.: Ccfg: Semi-supervised semantic segmentation via cross-fusion and contour guidance supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16348-16358 (2023)", + "30. Lin, D.: Probability guided loss for long-tailed multi-label image classification. In: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI'23/IAAI'23/EAAI'23, AAAI Press (2023). https://doi.org/10.1609/aaai.v37i2.25244, https://doi.org/10.1609/aaai.v37i2.25244", + "31. Lin, G., Shen, C., Van Den Hengel, A., Reid, I.: Efficient piecewise training of deep structured models for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3194-3203 (2016)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Howlader et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Liu, W., Rabinovich, A., Berg, A.C.: Parsenet: Looking wider to see better. arXiv preprint arXiv:1506.04579 (2015)", + "33. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4258-4267 (2022)", + "34. Luo, X., Hu, M., Song, T., Wang, G., Zhang, S.: Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In: International Conference on Medical Imaging with Deep Learning. pp. 820-833. PMLR (2022)", + "35. McClosky, D., Charniak, E., Johnson, M.: Effective self-training for parsing. In: Proceedings of the Human Language Technology Conference of the NAACL, Main Conference. pp. 152-159 (2006)", + "36. Oliver, A., Odena, A., Raffel, C.A., Cubuk, E.D., Goodfellow, I.: Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems 31 (2018)", + "37. Olsson, V., Tranheden, W., Pinto, J., Svensson, L.: Classmix: Segmentation-based data augmentation for semi-supervised learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1369-1378 (2021)", + "38. Ouali, Y., Hudelot, C., Tami, M.: Semi-supervised semantic segmentation with cross-consistency training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12674-12684 (2020)", + "39. Read, J., Pfahringer, B., Holmes, G., Frank, E.: Classifier chains for multi-label classification. Machine learning 85, 333-359 (2011)", + "40. Reif, S., Seibold, C., Freytag, A., Rodner, E., Stiefelhagen, R.: Every annotation counts: Multi-label deep supervision for medical image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9532-9542 (2021)", + "41. Ridnik, T., Ben-Baruch, E., Zamir, N., Noy, A., Friedman, I., Protter, M., Zelnik-Manor, L.: Asymmetric loss for multi-label classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 82-91 (2021)", + "42. Rizve, M.N., Duarte, K., Rawat, Y.S., Shah, M.: In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329 (2021)", + "43. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems 29 (2016)", + "44. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)", + "45. Shi, W., Gong, Y., Ding, C., Tao, Z.M., Zheng, N.: Transductive semi-supervised deep learning using min-max features. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 299-315 (2018)", + "46. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems 33, 596-608 (2020)", + "47. Sun, R., Mai, H., Zhang, T., Wu, F.: DAW: Exploring the better weighting function for semi-supervised semantic segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=KR1G7NJUCD" + ], + "bbox": [ + 212, + 146, + 787, + 838 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Beyond Pixels", + "bbox": [ + 633, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Tsoumakas, G., Katakis, I.: Multi-label classification. Data Warehousing and Mining: Concepts, Methodologies, Tools, and Applications: Concepts, Methodologies, Tools, and Applications 3, 64 (2008)", + "49. Verma, V., Kawaguchi, K., Lamb, A., Kannala, J., Bengio, Y., Lopez-Paz, D.: Interpolation consistency training for semi-supervised learning. arXiv preprint arXiv:1903.03825 (2019)", + "50. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4248-4257 (2022)", + "51. Wu, T., Huang, Q., Liu, Z., Wang, Y., Lin, D.: Distribution-balanced loss for multi-label classification in long-tailed datasets. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 162-178. Springer (2020)", + "52. Xu, H.M., Liu, L., Bian, Q., Yang, Z.: Semi-supervised semantic segmentation with prototype-based consistency regularization. arXiv preprint arXiv:2210.04388 (2022)", + "53. Xu, J., Le, H.: Generating representative samples for few-shot classification. In: CVPR (2022)", + "54. Xu, J., Le, H., Huang, M., Athar, S., Samaras, D.: Variational feature disentangling for fine-grained few-shot classification. In: ICCV (2021)", + "55. Xu, J., Le, H., Samaras, D.: Generating features with increased crop-related diversity for few-shot object detection. In: CVPR (2023)", + "56. Xu, J., Le, H.M., Nguyen, V., Ranjan, V., Samaras, D.: Zero-shot object counting. CVPR pp. 15548-15557 (2023)", + "57. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023)", + "58. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4268-4277 (2022)", + "59. Yarowsky, D.: Unsupervised word sense disambiguation rivaling supervised methods. In: 33rd annual meeting of the association for computational linguistics. pp. 189-196 (1995)", + "60. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 605–613. Springer (2019)", + "61. Yu, Y., Pedrycz, W., Miao, D.: Multi-label classification by exploiting label correlations. Expert Systems with Applications 41(6), 2989-3004 (2014)", + "62. Yuan, J., Liu, Y., Shen, C., Wang, Z., Li, H.: A simple baseline for semi-supervised semantic segmentation with strong data augmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8229-8238 (2021)", + "63. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019)", + "64. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)", + "65. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: Pro" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P. Howlader et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11350-11359 (2023)", + "66. Zhong, Y., Yuan, B., Wu, H., Yuan, Z., Peng, J., Wang, Y.X.: Pixel contrastive-consistent semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7273-7282 (2021)", + "67. Zhou, T., Wang, S., Bilmes, J.: Time-consistent self-supervision for semi-supervised learning. In: International Conference on Machine Learning. pp. 11523-11533. PMLR (2020)", + "68. Zhou, Y., Sun, X., Zha, Z.J., Zeng, W.: Context-reinforced semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4046-4055 (2019)", + "69. Zou, Y., Zhang, Z., Zhang, H., Li, C.L., Bian, X., Huang, J.B., Pfister, T.: Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713 (2020)", + "70. Zuo, S., Yu, Y., Liang, C., Jiang, H., Er, S., Zhang, C., Zhao, T., Zha, H.: Self-training with differentiable teacher. arXiv preprint arXiv:2109.07049 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 383 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Beyond Pixels", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_model.json b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ad775e4d98a9c6990a8f35abb9d3d3a817ef326f --- /dev/null +++ b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_model.json @@ -0,0 +1,2592 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.207 + ], + "angle": 0, + "content": "Beyond Pixels: Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.234, + 0.739, + 0.252 + ], + "angle": 0, + "content": "Prantik Howlader1, Srijan Das2, Hieu Le3, and Dimitris Samaras1" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.262, + 0.685, + 0.306 + ], + "angle": 0, + "content": "1 Stony Brook University, New York, USA \n2 University of North Carolina at Charlotte, NC, USA \n3 EPFL, Lausanne, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.341, + 0.744, + 0.606 + ], + "angle": 0, + "content": "Abstract. Incorporating pixel contextual information is critical for accurate segmentation. In this paper, we show that an effective way to incorporate contextual information is through a patch-based classifier. This patch classifier is trained to identify classes present within an image region, which facilitates the elimination of distractors and enhances the classification of small object segments. Specifically, we introduce Multi-scale Patch-based Multi-label Classifier (MPMC), a novel plug-in module designed for existing semi-supervised segmentation (SSS) frameworks. MPMC offers patch-level supervision, enabling the discrimination of pixel regions of different classes within a patch. Furthermore, MPMC learns an adaptive pseudo-label weight, using patch-level classification to alleviate the impact of the teacher's noisy pseudo-label supervision on the student. This lightweight module can be integrated into any SSS framework, significantly enhancing their performance. We demonstrate the efficacy of our proposed MPMC by integrating it into four SSS methodologies and improving them across two natural image and one medical segmentation dataset, notably improving the segmentation results of the baselines across all the three datasets. Code is available at: https://github.com/cvlab-stonybrook/Beyond-Pixels" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.618, + 0.63, + 0.634 + ], + "angle": 0, + "content": "Keywords: Segmentation \\(\\cdot\\) Semi-supervised \\(\\cdot\\) Patches" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.658, + 0.377, + 0.674 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Supervised segmentation requires a large amount of pixel level annotations, which is laborious and time-consuming [53-56]. This paper focuses on semi-supervised segmentation (SSS) [1, 20], which alleviates the reliance on labeled annotations by leveraging a large quantity of unlabeled images, accompanied by a limited number of manually labeled images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The fundamental task in segmentation is a per-pixel classification. Thus contextual information of a pixel is important to improve its performance [31,32,68]. Context becomes more critical in SSS settings, which uses limited labeled data making the models more prone to overfitting. We note that recent approaches in SSS methods which predominantly leverage consistency regularization with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "P. Howlader et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.758, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig.1: Segmentations and Multi-label classifications of image patches (Cityscapes) After 20 epochs, (b) and (c) are the segmentations and classes in the segmentations respectively, (d) is classes predicted by multi-label classifier in MPMC. (f) and (g) are segmentations by Baseline (AugSeg [65]) and Baseline + MPMC at the end of training. White boxes represent regions where MPMC improves the baseline, while orange boxes represent regions where segmentation network predicts classes not present in the patch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.785, + 0.518 + ], + "angle": 0, + "content": "pseudo-labeling techniques [7,14,20,22,36,38,50,57,69] do not exploit contextual information explicitly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We hypothesize that patches are the simplest way to determine the context around pixels. The key observation is that the contextual information of an image patch are the classes present in it. For example, in the image patch in the first row of Fig. 1 (a) the classes present are car, road, vegetation and terrain. Thus, training a network to not only classify each pixel in an image but also predict the classes in the patch surrounding it, is a straight forward way to introduce contextual information in SSS pipeline. Moreover, utilizing multi-scale patch-level supervision can effectively encapsulate context at varying scales and distinguish pixels across different classes. To this end, we introduce a novel plugin module, termed as Multi-scale Patch-based Multilabel Classifier (MPMC). This module achieves enhanced discrimination of pixel regions between classes through additional multi-scale patch-level supervision. Notably, we observe that when training our baseline SSS model together with MPMC in limited labeled data scenario, the patch classifier exhibits greater resilience to noise than the segmentation network as illustrated in Fig. 1 (b) and (c). A key observation is when MPMC and baseline SSS model are trained together, the patch classifier provides patch-level labels, identifying the classes present within a region. This facilitates the elimination of distractors (first row in Fig. 1 (b), (c) and (d), while the segmentation network classifies pixels as truck which is not present in the patch, the multi-label classifier does not predict it) and enhances the classification of small object segments by the baseline method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Another important issue plaguing SSS methods is noisy pseudo-labels generated by the teacher for unlabeled images to supervise the student. This can lead to the student model overfitting to these noisy labels, consequently impairing its performance. Recent approaches have attempted to mitigate the impact of noise in pseudo-labels through various strategies, such as implementing a hard threshold [14, 23, 38], applying dynamic re-weighting to enhance the influence of reliable pixels [20], and employing negative data mining by archiving unreliable pixels in a memory bank [50]. In contrast to these approaches, we use the MPMC predictions to learn an adaptive weight that reduces the weight of noisy pseudo-labels in the teacher-student framework. Our approach hinges on the underlying observation that MPMC exhibits greater uncertainty (lower confidence) for noisy labels in a patch (see supplementary for details)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.479 + ], + "angle": 0, + "content": "In practice, MPMC is a parameterized module that can be plugged in the initial layers of the visual model, where it leverages the low-level pixel information to classify the classes present in each patch. This patch-level classification strategy forces the visual model to learn and preserve discriminative representations of classes, which are often prone to dilution in the deeper layers of the model. We demonstrate the efficacy of our proposed parameterized module, MPMC by integrating it into four SSS methodologies—UniMatch [57], AugSeg [65], AEL [20], and U2PL [50] - notably improving performance across all data partitions in two natural (Cityscapes and Pascal VOC) and one medical (ACDC) dataset. To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.487, + 0.785, + 0.531 + ], + "angle": 0, + "content": "1. We introduce a novel plug-in module, called MPMC that incorporates an auxiliary multi-scale multi-label classification task, enhancing pixel region discrimination across different classes via additional patch-level supervision." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.532, + 0.785, + 0.56 + ], + "angle": 0, + "content": "2. MPMC consists of two adaptive learning weights that reduces the influence of noisy pseudo-labels from the teacher model." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.562, + 0.785, + 0.636 + ], + "angle": 0, + "content": "3. We demonstrate MPMC's robustness by integrating in four SSS techniques, achieving consistent improvements across all protocols on Pascal VOC and Cityscapes. Notably, our method also shows improvement in medical image dataset ACDC, which indicates that our method is applicable beyond natural images." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.487, + 0.785, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.388, + 0.675 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Semi-supervised learning (SSL) has been extensively researched in recent years. Popular SSL approaches include consistency regularization [2,10,11,14,38,43,58], entropy minimization [5,16], and pseudo-labeling [24-28, 46]. In this work, our emphasis is placed on the domains of pseudo-labeling and consistency regularization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Pseudo labeling: Pseudo-labeling methods [28,45] alongside self-training techniques [35, 59] are designed to leverage labeled image data to produce pseudolabels for unlabeled data. An inherent problem that plagues pseudo-labeling is noise, thus most approaches [42,44,46,69] devise strategies to refine the pseudolabels. PseudoSeg [69], for instance, aims at enhancing pseudo-label accuracy" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Howlader et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "using attention mechanisms based on grad-cam [44]. In pseudo-label selection, FixMatch [46] adopts a confidence threshold, whereas UPS [42] relies on uncertainty measures. \\(\\mathrm{ST}++\\) [58] focuses on selecting unlabeled images more likely to yield accurate pseudo-pixels. Emerging approaches like U2PL [50] explore the idea of learning reliable pseudo-labels from their less reliable counterparts. CFCG [29] introduced a novel approach that combines cross-fusion and contour guidance for pseudo-label improvement. We differ from these methods by explicitly introducing patch-level contextual information in the SSS methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Consistency Regularization: This approach enforces the model to have consistent predictions on different views of the unlabeled images. UniMatch [57] uses dual perturbations and feature perturbations to force the network to learn more consistent information. ICT [49] uses Mixup [64] augmentation for consistency regularization. Recent methods [14, 58, 62] use Cutout [9], Cutmix [63], Class-mix [37] as strong data augmentation. Consistency Regularization is used with pseudo-labeling techniques in Mixmatch [4] and TC-SSL [67]. In our work, we use a set of strong and weak data augmentations [1, 65] for generating pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Multi-label classification: This approach classifies an image into more than one label present in the image. Most works in this field explore finding the correlation between the labels and the images [21,39,61]. Recent works mainly focus on solving data imbalance issues [17,51]. We use multi-label classification to help in segmentation. Some works [15,40] use multi-label classification for segmentation in weakly supervised scenarios, for improved CAM-based pseudo-segmentation. However, we are using multi-label classification in a semi-supervised setting to introduce patch-level contextual information and also use the classifier confidence to reduce the impact of noisy pseudo-labels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.637, + 0.562 + ], + "angle": 0, + "content": "3 Semi-Supervised Segmentation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Given a labeled set \\(\\mathcal{D}^l = \\{(x^l,y^l)\\}\\) and an unlabeled set \\(\\mathcal{D}^u = \\{x^u\\}\\), the goal of semi-supervised semantic segmentation is to develop a model that effectively utilizes both labeled and unlabeled data. This section commences with an overview of the foundational framework for semi-supervised segmentation in Section 3.1. followed by an overview of our proposed MPMC in Section 3.2. The section concludes with an elaboration of MPMC's functionality (Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.682, + 0.362, + 0.697 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Teacher-Student frameworks leverage weak-to-strong consistency regularization to leverage unlabeled data. As depicted in Fig. 2(a), each unlabeled image \\( x^{u} \\) is simultaneously perturbed by two operators, i.e., weak perturbation \\( A^w \\) (cropping, random flip, etc), and strong perturbation \\( A^s \\) (color jitter, gaussian blur, contrast, etc). The framework's objective function is a convex combination of supervised loss \\( \\mathcal{L}_l \\) for the labeled data and unsupervised loss \\( \\mathcal{L}_u \\) for the unlabeled data as:" + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.814, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {l} + \\alpha \\mathcal {L} _ {u} \\quad \\text {w h e r e ,} \\quad \\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum \\mathbb {1} \\left(p ^ {w} \\geq t h\\right) \\mathcal {E} \\left(p ^ {w}, p ^ {s}\\right) \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.143, + 0.495, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.143, + 0.788, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.789, + 0.421 + ], + "angle": 0, + "content": "Fig. 2: Overall Pipeline of our novel multi-label classification based semantic segmentation: (a) Left: End-to-end Teacher-Student Pipeline with our novel method Multi-scale Patch-based Multi-label Classifier (MPMC). (b) Right: For unlabeled images, The teacher MPMC extracts features from a layer in the segmentation model's encoder to classify the feature's receptive field patch. The confidence of MPMC for a class in a patch is used to calculate two adaptive weights \\(\\lambda_{s}\\) and \\(\\lambda_{m}\\) which is used to reduce the influence of noisy predictions in that patch from the teacher to train both the student segmentation network and MPMC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.529 + ], + "angle": 0, + "content": "where, \\(\\mathcal{L}_l\\) is the standard cross-entropy loss, comparing the model's predictions with ground truth. While \\(\\mathcal{L}_u\\) aims to regularize the predictions for an unlabeled sample so that they remain consistent under both weak and strong augmentations (\\(p^w\\) and \\(p^s\\)). Here, \\(th\\) is a predefined confidence threshold to filter noisy pseudo labels and \\(\\mathcal{E}\\) minimizes the entropy between two probability distributions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.555, + 0.341, + 0.569 + ], + "angle": 0, + "content": "3.2 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Fig. 2(a) illustrates the comprehensive training process of our proposed MPMC, which is tailored to enhance the discrimination of patches among different classes and to mitigate the impact of noisy predictions transmitted from the teacher to the student. MPMC can be integrated into both teacher and student models within the teacher-student framework. It is composed of a classifier that performs multi-label classifications of the multi-scale patch representations of the input image. The application of MPMC spans two distinct scenarios: 1) For labeled images, MPMC assigns classes to each patch representation based on the classes present within a patch. The multi-scale nature of these patch representations and their classification enhances the segmentor's ability to discriminate classes by capturing enhanced contextual information related to variations in shape and size. 2) In the case of unlabeled images, besides patch classification, MPMC computes a pseudo-label weight map to alleviate the impact of the teacher's noisy pseudo-label supervision on the student. This involves assigning weights to the pseudo-labels within each patch in consistency learning framework, which are determined by the confidence of the MPMC's predictions for each class in the patch, to enhance the reliability and accuracy of the pseudo-labeling process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Howlader et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.72, + 0.163 + ], + "angle": 0, + "content": "3.3 Multi-scale Patch-based Multi-label Classifier (MPMC)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.248 + ], + "angle": 0, + "content": "Our proposed MPMC module is specifically designed to address the shortcomings of SSS pipeline. MPMC can be plugged in at any layer within the teacher-student network, both within the teacher model, denoted as \\( g \\), and within the student model, denoted as \\( \\hat{g} \\), they share the same parameters. Below, we detail the training of MPMC on labeled images and then on unlabeled images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.789, + 0.42 + ], + "angle": 0, + "content": "MPMC on Labeled Images. As illustrated in Fig. 2, MPMC incorporates patch-supervision into the segmentation framework, by training a multi-label classifier that takes as input a single pixel in the feature map of the \\(\\tau^{th}\\) layer of the encoder. This single pixel-representation encodes the information from a patch (receptive field) in the original image. Further, to encapsulate context at varying scales we perform average pooling on this features map with kernels of different sizes before performing multi-label classification. Let us define a patch of image and the corresponding segmentation in a receptive field \\(r\\) as \\(\\{x_r^l,y_r^l\\} \\in \\mathcal{D}^l\\). Note that \\(y_{r}^{l}\\) is a multi-label target where the classes present in the receptive field \\(r\\) are positive classes, and the classes not present are negative classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.789, + 0.558 + ], + "angle": 0, + "content": "The input to the student MPMC \\(\\hat{g}\\) are the features extracted from the \\(\\tau^{th}\\) layer of the encoder in student segmentation network \\(\\hat{f}\\), denoted by \\(z_{r} = \\hat{f}_{\\tau}(\\mathcal{A}_w(x_r^l))\\). MPMC captures the contextual information within a patch at different scales, \\(S_{i}\\), \\(i \\in \\{0,1\\}\\), where \\(s_0\\) and \\(s_1\\) are average pool kernel sizes. For all our experiments \\(s_0\\) and \\(s_1\\) are set to \\(7 \\times 7\\) and \\(19 \\times 19\\). These pooling operations are performed with appropriate padding to maintain uniformity in the shape of the output features. MPMC \\(\\hat{g}\\) is composed of the concatenation of these multiscale pooled features, followed by a convolutional network \\(\\mathcal{M}\\) and a final linear layer. Thus, the output of the MPMC \\(\\hat{g}\\) for the input patch \\(x_{r}^{l}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.567, + 0.785, + 0.586 + ], + "angle": 0, + "content": "\\[\nq ^ {l} = \\sigma \\left(W ^ {T} \\mathcal {M} \\left(\\operatorname {c o n c a t} \\left(z _ {r}, z _ {r} ^ {S _ {0}}, z _ {r} ^ {S _ {1}}\\right)\\right) + b\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.787, + 0.626 + ], + "angle": 0, + "content": "where \\(W\\) and \\(b\\) are linear layer weights and bias, \\(\\sigma\\) is the sigmoid activation and \\(q^l = [q_1^l,\\dots,q_C^l]^T\\in \\mathbb{R}^C\\) are the predicted probabilities for \\(C\\) classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.789, + 0.793 + ], + "angle": 0, + "content": "While binary cross-entropy loss is applicable to train MPMC, it indiscriminately weighs all positive and negative classes equally. Considering the receptive field of a feature from the \\(\\tau^{th}\\) layer of the encoder in the segmentation network \\(\\hat{f}\\) covers only a small patch relative to the input image's size, the prevalence of negative classes within a patch significantly exceeds that of positive classes. This imbalance influences the binary cross-entropy loss to favor negative classes [51]. Inspired by long-tailed multi-label classification [30, 51] which mitigate such imbalances by diminishing the weight of negative classes, we adapt an asymmetric loss [41], a focal loss variant that assigns distinct \\(\\gamma\\) values to positive and negative classes. For an image patch \\(x_{r}^{l}\\) and the predicted class probabilities \\(q^{l}\\) using our framework, the asymmetric focal loss is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.803, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\nl _ {r} ^ {M} = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\left\\{ \\begin{array}{l l} (1 - q _ {c} ^ {l}) ^ {\\gamma^ {+}} \\log (q _ {c} ^ {l}), & y _ {r, c} ^ {l} = 1. \\\\ (q _ {c} ^ {l}) ^ {\\gamma^ {-}} \\log (1 - q _ {c} ^ {l}), & y _ {r, c} ^ {l} = 0. \\end{array} \\right. \\tag {3}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Beyond Pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "In our experiments, we set \\(\\gamma^{+} = 0\\) and \\(\\gamma^{-} = 1\\). Subsequently, MPMC \\(\\hat{g}\\) undergoes training with multi-label supervision \\(\\mathcal{L}_l^M\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.192, + 0.786, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l} ^ {M} = \\frac {1}{| D ^ {l} |} \\sum_ {\\left(x ^ {l}, y ^ {l}\\right) \\in D ^ {l}} \\sum_ {r = 1} ^ {R} l _ {r} ^ {M} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.785, + 0.311 + ], + "angle": 0, + "content": "where \\( \\mathrm{R} \\) is the number of receptive fields at \\( \\tau^{th} \\) layer of the segmentor. This auxiliary task of patch classification augments the network's proficiency in learning and preserving class features, a critical function particularly pertinent as the fidelity of these features tends to diminish in the network's deeper layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.548 + ], + "angle": 0, + "content": "MPMC on Unlabeled images. In semi-supervised teacher-student framework, for unlabeled images, the prediction of the teacher network \\( f \\) for the weak augmented view \\( \\mathcal{A}^w(x^u) \\) is used to supervise the student network \\( \\hat{f} \\) for the strong augmented view \\( \\mathcal{A}^s(x^u) \\). In the early iterations of semi-supervised learning, the teacher's predictions are inherently noisy. Common practices [13, 20, 33, 46, 70], involve employing pixel-wise confidence score thresholds (th) to filter out unreliable pseudo labels. However, a high confidence threshold does not entirely eliminate noise due to the teacher's initial miscalibration [42], allowing some noisy predictions to have high confidence. To refine this process, our approach does not solely depend on confidence-thresholded teacher segmentation predictions for reliable pseudo-label extraction. Instead, we leverage the teacher MPMC \\( g \\)'s confidence for a patch to adjust the weights of the deemed reliable pseudo-labels within that patch. The underlying principle of our methodology is that MPMC exhibits greater uncertainty (lower confidence) for noisy labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.653 + ], + "angle": 0, + "content": "We employ the teacher MPMC \\( g \\)'s predictions to define two adaptive learning weights \\( \\lambda_{m} \\in \\mathbb{R}^{R \\times C} \\) and \\( \\lambda_{s} \\in \\mathbb{R}^{W \\times H} \\), where \\( R \\) denotes the number of patches, sized according to the receptive field within the image, \\( W \\) and \\( H \\) represents the width and height of the input image, respectively. The learning weight \\( \\lambda_{m} \\) is utilized to assign class weights to each patch for training the student MPMC, whereas \\( \\lambda_{s} \\) is applied to diminish the impact of noisy predictions during training the student segmentation network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.809 + ], + "angle": 0, + "content": "For an unlabeled image \\( x^{u} \\), consider a patch at the receptive field \\( r \\), denoted as \\( x_{r}^{u} \\). The teacher and student segmentation networks yield predictions \\( p^{w} = f(\\mathcal{A}^{w}(x_{r}^{u})) \\) and \\( p^{s} = \\hat{f} (\\mathcal{A}^{s}(x_{r}^{u})) \\), respectively. Simultaneously, the teacher and student MPMC predict \\( q_{r}^{u} = g(f_{\\tau}(\\mathcal{A}^{w}(x_{r}^{u}))) \\) and \\( \\hat{q}_r^u = \\hat{g} (\\hat{f}\\tau (\\mathcal{A}^s (x_r^u))) \\) using features from the \\( \\tau^{th} \\) layer of the encoder. Let us define the confidence of teacher MPMC \\( g \\) across all classes in patch \\( r \\) as \\( \\gamma = \\mathrm{softmax}(q_r^u) \\), with \\( \\gamma \\in \\mathbb{R}^C \\). The weight of the patch \\( r \\), \\( \\lambda_{m,r} \\) is determined by \\( \\gamma \\). For each pixel in patch \\( r \\) associated with class \\( c \\), \\( \\lambda_{s} \\) is assigned the weight \\( \\gamma_{c} \\). This approach assigns weights to pseudo-labels and pixels based on the teacher MPMC's confidence, enhancing the accuracy of the learning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Thus, the control parameters \\(\\lambda_{s}\\) and \\(\\lambda_{m}\\) are integrated to the unsupervised losses to train the student segmentation network and MPMC. The unsupervised" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Howlader et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "segmentation loss \\( L_{u} \\) mentioned in equation 1 to train the student segmentation network \\( \\hat{f} \\) is re-formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.188, + 0.786, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum_ {x ^ {u} \\in \\mathcal {D} ^ {u}} \\mathbb {1} \\left(p ^ {w} \\geq t h\\right) \\lambda_ {s} \\mathcal {E} \\left(p ^ {w}, p ^ {s}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.234, + 0.784, + 0.281 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) is the standard cross entropy loss and \\(th\\) is the predefined segmentation confidence threshold. Finally, the unsupervised multi-label loss \\(\\mathcal{L}_u^M\\) to train the student MPMC \\(\\hat{g}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.292, + 0.786, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} ^ {M} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum_ {x ^ {u} \\in D ^ {u}} \\sum_ {r = 1} ^ {R} \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\lambda_ {m, r} ^ {i} \\mathcal {E} ^ {+} \\left(q _ {r} ^ {u}, \\hat {q} _ {r} ^ {u}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.784, + 0.374 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}^+\\) is the binary cross entropy loss. Therefore, the overall loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.389, + 0.786, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {l} + \\mathcal {L} _ {l} ^ {M} + \\alpha \\mathcal {L} _ {u} + \\beta \\mathcal {L} _ {u} ^ {M} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.784, + 0.444 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) controls the contribution of \\(\\mathcal{L}_u\\) and \\(\\mathcal{L}_u^M\\). In all our experiments \\(\\alpha\\) and \\(\\beta\\) are set to 0.1 and 0.25 respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.467, + 0.376, + 0.484 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.312, + 0.514 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Datasets: PASCAL VOC 2012 [12] is a standard semantic segmentation dataset with 20 semantic classes and a background class. The training, validation, and testing sets consist of 1464, 1449, and 1556 images (Classic), respectively. Following [7, 65], we also include the additional augmented dataset [18] (Blender), which includes a collection of 10582 training images. We adopt the same partition protocols in [7, 65] to evaluate our method in both Classic and Blender sets. Cityscapes [8] dataset is designed for urban scene understanding, consisting of 30 classes, of which only 19 classes are used for scene parsing evaluation. It has 2975, 500, and 1525 images for training, validation, and testing, respectively. ACDC [3] dataset contains 200 annotated short-axis cardiac cineMR images from 100 patients. The segmentation masks consists of classes: left ventricle (LV), myocardium (Myo), and right ventricle (RV). 140 images from 70 patients, 60 images from 30 patients are randomly selected for training and validation respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation Details: For fair comparison, we adopt DeepLabv3+ [6] as the decoder in all of our experimental setups, and compare with both ResNet-101 and ResNet-50 [19] as the backbone architecture. In all our experiments with Cityscapes, Pascal, and ACDC datasets, the training resolution is 801, 513, and 224 respectively. Note, for maintaining fair comparison, our method \"UniMatch [57] + MPMC\" uses the training resolution of 321 which is used by UniMatch [57] when training on the Classic set of Pascal VOC. Further when training a MPMC" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "integrated baseline, we use the same weak and strong augmentations as used by the corresponding baselines. The convolutional network in MPMC is comprised of three convolution blocks. Each convolution block consists of a stack of two \\(3 \\times 3\\) convolutions followed by a \\(1 \\times 1\\) convolution. For all our experiments MPMC is integrated to the first layer of the ResNet Encoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.264, + 0.622, + 0.279 + ], + "angle": 0, + "content": "4.2 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.398 + ], + "angle": 0, + "content": "We conducted our experiments on PASCAL VOC 2012 and Cityscapes which are two popular natural image benchmarks, and a medical dataset ACDC. MPMC is integrated in four semi supervised methods UniMatch [57], AEL [20], U2PL [50] and AugSeg [65]. Note, UniMatch [57] is a consistency regularization based method. We observe that MPMC consistently improves over their baseline performance across all partition protocols on all datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.788, + 0.503 + ], + "angle": 0, + "content": "Results on PASCAL VOC 2012 Dataset. Table 1 shows the comparison with other SOTA methods for both the Classic and Blender sets. MPMC integrated to the four baseline methods improves them consistently across all data partitions for both Classic and Blender sets. We find that our method brings the most improvement for the least labeled data partition \\((\\frac{1}{16})\\), improving AugSeg, AEL, U2PL and UniMatch by \\(1.8\\%\\), \\(2.5\\%\\), \\(2.0\\%\\) and \\(2.1\\%\\) respectively in Classic with a ResNet-101 encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Results on Cityscapes Dataset. Table 2 shows the comparison with other SOTA methods. MPMC integrated to the four baseline methods improves them consistently across all data partitions. We observe that similar to results in PASCAL VOC 2012 dataset, our method brings the most improvement for the least labeled data partition \\((\\frac{1}{16})\\), improving AugSeg, AEL, U2PL and UniMatch by \\(1.7\\%\\), \\(2.1\\%\\), \\(2.3\\%\\) and \\(1.9\\%\\) respectively in Classic for ResNet-101 encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Results on ACDC Dataset. Table 3 shows the comparison with other SOTA methods. We observe that when MPMC is integrated to the four baseline methods, it significantly improves the performance for the \"3 cases\" data partition which has the least labeled images (\\(< 10\\%\\) of labeled images). Note we use ResNet-101 for all experiments in this dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.732 + ], + "angle": 0, + "content": "The consistent improvement in SSS performance in both natural and medical datasets and its ability to integrate in different SSS methods substantiates the robustness of MPMC. Moreover, the notable improvements achieved in scenarios with limited data shows MPMC's vital efficacy within SSS pipelines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Qualitative Results In Fig. 3 we visualise the results of integration of our method (MPMC) to AugSeg [65]. Introducing patch-level supervision with the integration of MPMC, leads the SSS model to better segment confusing classes. Among comparisons we show that MPMC leads to better segmentation of pixels regions of confusing classes like bus, train and truck. These classes are often misclassified not only because of sharing visual similarity but also having low training instances." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Howlader et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.228, + 0.787, + 0.271 + ], + "angle": 0, + "content": "Table 1: Quantitative results of different semi-supervised segmentation methods on Pascal VOC classic and blender set. We report Mean IoU under various partition protocols and show the improvements \\((\\Delta)\\) over corresponding baseline." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.283, + 0.781, + 0.753 + ], + "angle": 0, + "content": "
MethodClassicBlender
1/161/81/41/2Full1/161/81/4
ResNet-50
Supervised44.052.361.766.772.9---
PC2Seg [66][CVPR'21]56.964.667.670.972.3---
PseudoSeg [69][ICLR'21]56.964.667.670.972.2---
ST++ [58][CVPR'22]-----72.674.475.4
AugSeg [65][CVPR'21]64.272.176.177.478.877.278.278.2
AugSeg [65] + MPMC (Δ)65.973.777.678.579.779.079.879.7
(+1.7)(+1.6)(+1.5)(+1.1)(+0.9)(+1.8)(+1.6)(+1.5)
AEL [20][NeurIPS'21]62.964.170.372.774.074.176.177.9
AEL [20] + MPMC (Δ)65.066.172.174.275.376.178.079.6
(+2.1)(+2.0)(+1.8)(+1.5)(+1.3)(+2.0)(+1.9)(+1.7)
U2PL [50][CVPR'22]63.365.571.673.875.174.777.477.5
U2PL [50] + MPMC (Δ)65.267.473.375.276.276.579.179.1
(+1.9)(+1.9)(+1.7)(+1.4)(+1.1)(+1.8)(+1.7)(+1.6)
UniMatch [57][CVPR'23]71.972.576.077.478.778.179.079.1
UniMatch [57] + MPMC (Δ)73.574.077.478.579.479.680.480.4
(+1.6)(+1.5)(+1.4)(+1.1)(+0.7)(+1.5)(+1.4)(+1.3)
ResNet-101
Supervised45.155.364.869.773.570.675.076.5
CPS [7][CVPR'21]64.167.471.775.9-72.275.877.6
PS-MT [33][CVPR'22]65.869.676.678.480.075.578.278.7
PCR [52][NeurIPS'22]70.174.777.278.580.778.680.780.8
DAW [47][NeurIPS'23]74.877.479.580.681.578.578.979.6
CFCG [29][ICCV'23]-----77.479.480.4
AugSeg [65][CVPR'21]71.175.578.880.381.479.381.580.5
AugSeg [65] + MPMC (Δ)72.977.879.981.382.180.882.581.4
(+1.8)(+2.3)(+1.1)(+1.0)(+0.7)(+1.5)(+1.0)(+0.9)
AEL [20][NeurIPS'21]66.168.371.974.478.977.277.678.1
AEL [20] + MPMC (Δ)68.670.373.576.180.179.279.379.2
(+2.5)(+2.0)(+1.6)(+1.7)(+1.2)(+2.0)(+1.7)(+1.1)
U2PL [50][CVPR'22]68.069.273.776.279.577.279.079.3
U2PL [50] + MPMC (Δ)70.070.875.377.880.579.280.780.3
(+2.0)(+1.6)(+1.6)(+1.6)(+1.0)(+2.0)(+1.7)(+1.0)
UniMatch [57][CVPR'23]75.277.278.879.981.280.981.980.4
UniMatch [57] + MPMC (Δ)77.378.679.880.881.782.583.181.5
(+2.1)(+1.4)(+1.0)(+0.9)(+0.5)(+1.6)(+1.2)(+1.1)
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.159, + 0.36, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.223, + 0.359, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.286, + 0.359, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.349, + 0.36, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.413, + 0.318, + 0.422 + ], + "angle": 0, + "content": "(a) Input Image" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.159, + 0.501, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.223, + 0.501, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.286, + 0.501, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.349, + 0.501, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.413, + 0.461, + 0.421 + ], + "angle": 0, + "content": "(b) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.159, + 0.642, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.223, + 0.642, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.286, + 0.642, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.349, + 0.642, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.413, + 0.585, + 0.422 + ], + "angle": 0, + "content": "(c) AugSeg" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.159, + 0.785, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.223, + 0.784, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.286, + 0.784, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.349, + 0.784, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.413, + 0.749, + 0.422 + ], + "angle": 0, + "content": "(d) AugSeg + MPMC" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.434, + 0.788, + 0.491 + ], + "angle": 0, + "content": "Fig. 3: Qualitative Results on Cityscapes dataset: (a) is the original image, (b) is the Ground Truth, (c) are segmentations generated by AugSeg [65] compared to (d) which are segmentations generated when our method (MPMC) is integrated to AugSeg. White boxes show the areas where our method improves the baseline." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.527, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Table 2: Quantitative results of different semi-supervised segmentation methods on Cityscapes validation set. We report Mean IoU under various partition protocols and show the improvements \\((\\Delta)\\) over the corresponding baseline." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.582, + 0.789, + 0.823 + ], + "angle": 0, + "content": "
MethodResNet-50ResNet-101
1/161/81/41/21/161/81/41/2
Supervised63.3468.7374.1476.6266.372.875.078.0
CPS [7][CVPR'21]69.7974.3976.8578.6469.874.374.676.8
PS-MT [33][CVPR'22]-75.7676.9277.64-76.977.679.1
PCR [52][NeurIPS'22]----73.476.378.479.1
CFCG [29][ICCV'23]76.178.979.380.177.879.680.480.9
AugSeg [65][CVPR'21]73.776.478.779.375.277.879.680.4
AugSeg [65] + MPMC (Δ)75.3 (+1.6)77.7 (+1.3)79.8 (+1.1)80.3 (+1.0)76.9 (+1.7)79.2 (+1.4)80.9 (+1.3)81.3 (+0.9)
AEL [20][NeurIPS'21]68.272.774.977.574.575.677.579.0
AEL [20] + MPMC (Δ)70.5 (+2.3)74.5 (+1.8)76.6 (+1.7)79.0 (+1.5)76.6 (+2.1)77.5 (+1.9)78.9 (+1.4)80.3 (+1.3)
U2PL [50][CVPR'22]69.073.076.378.674.976.578.579.1
U2PL [50] + MPMC (Δ)71.2 (+2.2)74.6 (+1.6)77.8 (+1.5)79.8 (+1.2)77.2 (+2.3)78.2 (+1.7)80.0 (+1.5)80.0 (+0.9)
UniMatch [57][CVPR'23]75.076.877.578.676.677.979.279.5
UniMatch [57] + MPMC (Δ)76.8 (+1.8)78.0 (+1.2)78.6 (+1.1)79.5 (+0.9)78.5 (+1.9)79.2 (+1.3)80.9 (+1.7)81.3 (+1.8)
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Howlader et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Quantitative results of different semi-supervised segmentation methods on ACDC dataset with 3/7 labeled cases. We report Dice Similarity Coefficient (DSC) metric averaged over 3 classes. We use ResNet-101 for all experiments in this dataset." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.2, + 0.642, + 0.389 + ], + "angle": 0, + "content": "
Method3 cases7 cases
Supervised41.562.5
UA-MT [60][MICCAI'19]61.081.5
CPS [7][CVPR'21]60.383.3
CNN & Trans [34][PMLR'22]65.686.4
AugSeg [65][CVPR'21]85.187.3
AugSeg [65] + MPMC (Δ)88.6 (+3.5)88.4 (+1.1)
AEL [20][NeurIPS'21]82.787.6
AEL [20] + MPMC (Δ)85.4 (+2.7)88.5 (+0.9)
U2PL [50][CVPR'22]84.289.2
U2PL [50] + MPMC (Δ)87.4 (+3.2)90.5 (+1.3)
UniMatch [57][CVPR'23]88.989.9
UniMatch [57] + MPMC (Δ)91.8 (+2.9)90.4 (+0.5)
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.42, + 0.403, + 0.433 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.787, + 0.493 + ], + "angle": 0, + "content": "We conduct experiments to study the impact of various components of our approach. All experiments in this section were conducted on \\(\\frac{1}{16}\\) data partition of the Cityscapes dataset, and use AugSeg [65] as the baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.644 + ], + "angle": 0, + "content": "Segmentation performance at different scales. Here we analyse how scaling of input features for patch-classification, improves the performance of the SSS method. We consider two scales - average pool kernels \\(7 \\times 7\\) and \\(19 \\times 19\\). Fig. 4 illustrates for two random classes pole and rider, that average pooling with a kernel size of \\(7 \\times 7\\) is better at enhancing the pixel accuracy for instances with smaller areas, whereas the \\(19 \\times 19\\) kernel size is better at improving accuracy for instances with larger areas. This can be attributed to the larger receptive field of the \\(19 \\times 19\\) kernel, enabling more effective contextual learning for larger regions. While the smaller receptive field of \\(7 \\times 7\\) kernel is better able to discern smaller objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.764 + ], + "angle": 0, + "content": "Analysing class features The crux of our method is to better discriminate pixel regions between different classes. Here we analyze if MPMC integrated to AugSeg [65] can better discriminate features than the baseline for different classes in Cityscapes Dataset. We use the pixel features from the penultimate layer of the segmentation network to analyse via t-SNE. Fig 5 illustrates for four random classes - car, bus, train and truck, integration of MPMC leads to better discrimination between the classes, i.e., compared to AugSeg [65], the features of train and truck are more clustered and separated from the rest." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Influence of MPMC on patch-level segmentation. Our method MPMC improves segmentation by introducing patch classification as an auxiliary task. We analyse at patch level the correlation between multi-label accuracy and the Mean IoU performance of the segmentation network. We use (1 - hamming loss) [48] as a measure of multi-label classification and use the basic teacher-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.143, + 0.499, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.143, + 0.736, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Fig. 4: Pixel accuracy across different scales with respect to instance size. For two different scales, we compare the pixel accuracy with respect to size of instances for two random classes: pole and rider." + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.395, + 0.725, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.568, + 0.785, + 0.597 + ], + "angle": 0, + "content": "Fig. 5: Analysis of class features (a) t-SNE of features predicted by AugSeg [65], (b) t-SNE of features predicted by AugSeg [65] + MPMC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.785, + 0.687 + ], + "angle": 0, + "content": "student framework mentioned in Section 3.1 as our baseline. Table 4 illustrates that the avg MIoU improvement is highest (4.05%) in patches having more than 90% multi-label classification accuracy showing that MPMC improves segmentation at the patch level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.785, + 0.717 + ], + "angle": 0, + "content": "Impact of MPMC on semantic segmentation in limited labeled data scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In Table 5 shows the impact of MPMC on segmentation when trained with only labeled data. We use the basic teacher-student framework mentioned in Section 3.1 as our baseline, which achieves MIoU of 66.3, 72.8 and 75.0 in 186, 372 and 744 partitions of labeled images respectively. After introducing PMC, there is an improvement of \\(1.9\\%\\), \\(1.5\\%\\) and \\(1.3\\%\\) in 186, 372 and 744 partitions of labeled images respectively. We observe that MPMC improves segmentation performance in limited-labeled data scenarios, with the improvement highest for least labeled data, i.e., \\(1.9\\%\\) for 186 labeled images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Howlader et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.244, + 0.145, + 0.757, + 0.16 + ], + "angle": 0, + "content": "Table 4: Analysis of the influence of MPMC on segmentation at patch-level." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.172, + 0.787, + 0.239 + ], + "angle": 0, + "content": "
1 - hamming loss90+80-9070-8060-7050-6040-5030-4020-3010-200-10
% patches12.625.220.58.45.15.24.110.22.46.3
Avg M IoU Δ+4.05+1.57+0.88+1.55+0.91+1.15+0.49+0.10-1.90-2.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.252, + 0.477, + 0.323 + ], + "angle": 0, + "content": "Table 5: Impact of MPMC on segmentation (MIoU) when trained with different volumes of labeled data. MPMC shows the highest improvement \\((1.9\\%)\\) in the least labeled data partition." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.334, + 0.445, + 0.409 + ], + "angle": 0, + "content": "
# Labeled Data
186372744
w/o MPMC66.372.875.0
w/ MPMC68.274.376.3
Gain/Δ+1.9+1.5+1.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.252, + 0.788, + 0.308 + ], + "angle": 0, + "content": "Table 6: Ablation study of different component: MPMC on labeled images (MPMCL) and MPMC on unlabeled images (MPMCUL)" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.321, + 0.77, + 0.41 + ], + "angle": 0, + "content": "
MPMCLMPMCUL1/16 (186)1/8 (372)1/4 (744)
68.473.574.2
✓70.875.475.7
✓71.376.076.6
✓✓73.176.777.3
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.437, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Analysis of different components We ablate each component of MPMC with basic teacher-student framework (Section 3.1) as the baseline. MPMC on labeled images (MPMCL), which introduces multi-scale patch-based classification on only the labeled data, achieves an improvement of \\(2.4\\%\\), \\(1.9\\%\\), and \\(1.5\\%\\) under \\(\\frac{1}{16}\\), \\(\\frac{1}{8}\\) and \\(\\frac{1}{4}\\) partition protocols respectively. MPMC on unlabeled images (MPMCUL) leads to a further improvement of \\(2.9\\%\\), \\(2.5\\%\\), and \\(2.4\\%\\) under \\(\\frac{1}{16}\\), \\(\\frac{1}{8}\\) and \\(\\frac{1}{4}\\) partition protocols respectively. Thus, MPMC is effective on both labeled and unlabeled instances in SSS pipelines." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.582, + 0.36, + 0.598 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.81 + ], + "angle": 0, + "content": "In this paper, we have proposed a novel Multi-scale Patch-based Multi-label Classifier (MPMC), which introduces patch-level context information in the form of an auxiliary task of multi-scale patch-based multi-label classification. When jointly trained with a SSS model, the patch classifier provides patch-level labels, identifying the classes present within a region. This facilitates the elimination of distractors and enhances the classification of small object segments by the baseline method. Further, the MPMC predictions are used to learn an adaptive weight that reduces the influence of noisy pseudo-labels in the teacher-student framework. We performed a thorough analysis of the importance of multi-scale patch-level contextual information in improving segmentation especially in low data scenarios. Further, we demonstrated the efficacy of our proposed parameterized module MPMC by integrating it into four SSS methodologies and showing consistent improvements in Cityscapes, Pascal VOC and ACDC benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Acknowledgement. This work was partially supported by the National Science Foundation (IIS-2245652, IIS-2123920, IIS-2212046)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Beyond Pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.236 + ], + "angle": 0, + "content": "1. Alonso, I., Sabater, A., Ferstl, D., Montesano, L., Murillo, A.C.: Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8219-8228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.786, + 0.264 + ], + "angle": 0, + "content": "2. Bachman, P., Alsharif, O., Precup, D.: Learning with pseudo-ensembles. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.786, + 0.32 + ], + "angle": 0, + "content": "3. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37(11), 2514-2525 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.786, + 0.362 + ], + "angle": 0, + "content": "4. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.786, + 0.404 + ], + "angle": 0, + "content": "5. Chen, H., Jin, Y., Jin, G., Zhu, C., Chen, E.: Semisupervised semantic segmentation by improving prediction confidence. IEEE Transactions on Neural Networks and Learning Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.447 + ], + "angle": 0, + "content": "6. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.786, + 0.489 + ], + "angle": 0, + "content": "7. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.786, + 0.545 + ], + "angle": 0, + "content": "8. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.786, + 0.573 + ], + "angle": 0, + "content": "9. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.786, + 0.615 + ], + "angle": 0, + "content": "0. Durasov, N., Dorndorf, N., Le, H., Fua, P.: Zigzag: Universal sampling-free uncertainty estimation through two-step inference. Transactions on Machine Learning Research (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.786, + 0.657 + ], + "angle": 0, + "content": "1. Durasov, N., Oner, D., Donier, J., Le, H., Fua, P.: Enabling uncertainty estimation in iterative neural networks. In: Forty-first International Conference on Machine Learning (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.786, + 0.699 + ], + "angle": 0, + "content": "2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88(2), 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.786, + 0.742 + ], + "angle": 0, + "content": "3. Fan, J., Gao, B., Jin, H., Jiang, L.: Ucc: Uncertainty guided cross-head co-training for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9947-9956 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.786, + 0.784 + ], + "angle": 0, + "content": "4. French, G., Laine, S., Aila, T., Mackiewicz, M., Finlayson, G.: Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Ge, W., Yang, S., Yu, Y.: Multi-evidence filtering and fusion for multi-label classification, object detection and semantic segmentation based on weakly supervised learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1277-1286 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "P. Howlader et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. Advances in neural information processing systems 17 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Guo, H., Wang, S.: Long-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15089-15098 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "18. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "20. Hu, H., Wei, F., Hu, H., Ye, Q., Cui, J., Wang, L.: Semi-supervised semantic segmentation via adaptive equalization learning. Advances in Neural Information Processing Systems 34, 22106-22118 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "21. Huang, J., Li, G., Huang, Q., Wu, X.: Learning label specific features for multi-label classification. In: 2015 IEEE International Conference on Data Mining. pp. 181-190. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.438 + ], + "angle": 0, + "content": "22. Huo, X., Xie, L., He, J., Yang, Z., Zhou, W., Li, H., Tian, Q.: Atso: Asynchronous teacher-student optimization for semi-supervised image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1235-1244 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "23. Ke, Z., Qiu, D., Li, K., Yan, Q., Lau, R.W.: Guided collaborative training for pixel-wise semi-supervised learning. In: European conference on computer vision. pp. 429-445. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "24. Le, H., Goncalves, B., Samaras, D., Lynch, H.: Weakly labeling the antarctic: The penguin colony case. In: CVPRW (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "25. Le, H., Samaras, D., Lynch, H.J.: A convolutional neural network architecture designed for the automated survey of seabird colonies. Remote Sensing in Ecology and Conservation 8(2), 251-262 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "26. Le, H., Vicente, T.F.Y., Nguyen, V., Hoai, M., Samaras, D.: A+D Net: Training a shadow detector with adversarial shadow attenuation. In: European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "27. Le, H., Yu, C.P., Zelinsky, G., Samaras, D.: Co-localization with category-consistent features and geodesic distance propagation. In: ICCVW (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "28. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "29. Li, S., He, Y., Zhang, W., Zhang, W., Tan, X., Han, J., Ding, E., Wang, J.: Ccfg: Semi-supervised semantic segmentation via cross-fusion and contour guidance supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16348-16358 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.798 + ], + "angle": 0, + "content": "30. Lin, D.: Probability guided loss for long-tailed multi-label image classification. In: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI'23/IAAI'23/EAAI'23, AAAI Press (2023). https://doi.org/10.1609/aaai.v37i2.25244, https://doi.org/10.1609/aaai.v37i2.25244" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Lin, G., Shen, C., Van Den Hengel, A., Reid, I.: Efficient piecewise training of deep structured models for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3194-3203 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Beyond Pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "32. Liu, W., Rabinovich, A., Berg, A.C.: Parsenet: Looking wider to see better. arXiv preprint arXiv:1506.04579 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.229 + ], + "angle": 0, + "content": "33. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4258-4267 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.271 + ], + "angle": 0, + "content": "34. Luo, X., Hu, M., Song, T., Wang, G., Zhang, S.: Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In: International Conference on Medical Imaging with Deep Learning. pp. 820-833. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.311 + ], + "angle": 0, + "content": "35. McClosky, D., Charniak, E., Johnson, M.: Effective self-training for parsing. In: Proceedings of the Human Language Technology Conference of the NAACL, Main Conference. pp. 152-159 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.352 + ], + "angle": 0, + "content": "36. Oliver, A., Odena, A., Raffel, C.A., Cubuk, E.D., Goodfellow, I.: Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.393 + ], + "angle": 0, + "content": "37. Olsson, V., Tranheden, W., Pinto, J., Svensson, L.: Classmix: Segmentation-based data augmentation for semi-supervised learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1369-1378 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.433 + ], + "angle": 0, + "content": "38. Ouali, Y., Hudelot, C., Tami, M.: Semi-supervised semantic segmentation with cross-consistency training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12674-12684 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.459 + ], + "angle": 0, + "content": "39. Read, J., Pfahringer, B., Holmes, G., Frank, E.: Classifier chains for multi-label classification. Machine learning 85, 333-359 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.514 + ], + "angle": 0, + "content": "40. Reif, S., Seibold, C., Freytag, A., Rodner, E., Stiefelhagen, R.: Every annotation counts: Multi-label deep supervision for medical image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9532-9542 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.555 + ], + "angle": 0, + "content": "41. Ridnik, T., Ben-Baruch, E., Zamir, N., Noy, A., Friedman, I., Protter, M., Zelnik-Manor, L.: Asymmetric loss for multi-label classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 82-91 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.596 + ], + "angle": 0, + "content": "42. Rizve, M.N., Duarte, K., Rawat, Y.S., Shah, M.: In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.636 + ], + "angle": 0, + "content": "43. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.69 + ], + "angle": 0, + "content": "44. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.731 + ], + "angle": 0, + "content": "45. Shi, W., Gong, Y., Ding, C., Tao, Z.M., Zheng, N.: Transductive semi-supervised deep learning using min-max features. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 299-315 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.785 + ], + "angle": 0, + "content": "46. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems 33, 596-608 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.839 + ], + "angle": 0, + "content": "47. Sun, R., Mai, H., Zhang, T., Wu, F.: DAW: Exploring the better weighting function for semi-supervised semantic segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=KR1G7NJUCD" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Howlader et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Tsoumakas, G., Katakis, I.: Multi-label classification. Data Warehousing and Mining: Concepts, Methodologies, Tools, and Applications: Concepts, Methodologies, Tools, and Applications 3, 64 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "49. Verma, V., Kawaguchi, K., Lamb, A., Kannala, J., Bengio, Y., Lopez-Paz, D.: Interpolation consistency training for semi-supervised learning. arXiv preprint arXiv:1903.03825 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "50. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4248-4257 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.785, + 0.342 + ], + "angle": 0, + "content": "51. Wu, T., Huang, Q., Liu, Z., Wang, Y., Lin, D.: Distribution-balanced loss for multi-label classification in long-tailed datasets. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 162-178. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "52. Xu, H.M., Liu, L., Bian, Q., Yang, Z.: Semi-supervised semantic segmentation with prototype-based consistency regularization. arXiv preprint arXiv:2210.04388 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.41 + ], + "angle": 0, + "content": "53. Xu, J., Le, H.: Generating representative samples for few-shot classification. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "54. Xu, J., Le, H., Huang, M., Athar, S., Samaras, D.: Variational feature disentangling for fine-grained few-shot classification. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "55. Xu, J., Le, H., Samaras, D.: Generating features with increased crop-related diversity for few-shot object detection. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "56. Xu, J., Le, H.M., Nguyen, V., Ranjan, V., Samaras, D.: Zero-shot object counting. CVPR pp. 15548-15557 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "57. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "58. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4268-4277 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "59. Yarowsky, D.: Unsupervised word sense disambiguation rivaling supervised methods. In: 33rd annual meeting of the association for computational linguistics. pp. 189-196 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.674 + ], + "angle": 0, + "content": "60. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 605–613. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "61. Yu, Y., Pedrycz, W., Miao, D.: Multi-label classification by exploiting label correlations. Expert Systems with Applications 41(6), 2989-3004 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "62. Yuan, J., Liu, Y., Shen, C., Wang, Z., Li, H.: A simple baseline for semi-supervised semantic segmentation with strong data augmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8229-8238 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "63. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "64. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "65. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: Pro" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11350-11359 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.232 + ], + "angle": 0, + "content": "66. Zhong, Y., Yuan, B., Wu, H., Yuan, Z., Peng, J., Wang, Y.X.: Pixel contrastive-consistent semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7273-7282 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.272 + ], + "angle": 0, + "content": "67. Zhou, T., Wang, S., Bilmes, J.: Time-consistent self-supervision for semi-supervised learning. In: International Conference on Machine Learning. pp. 11523-11533. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.313 + ], + "angle": 0, + "content": "68. Zhou, Y., Sun, X., Zha, Z.J., Zeng, W.: Context-reinforced semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4046-4055 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "69. Zou, Y., Zhang, Z., Zhang, H., Li, C.L., Bian, X., Huang, J.B., Pfister, T.: Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.384 + ], + "angle": 0, + "content": "70. Zuo, S., Yu, Y., Liang, C., Jiang, H., Er, S., Zhang, C., Zhao, T., Zha, H.: Self-training with differentiable teacher. arXiv preprint arXiv:2109.07049 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_origin.pdf b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dfd70acb6d1fed562986e54082b45bde392d5731 --- /dev/null +++ b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/4e2bc210-face-48cf-b49f-a28b6aafb611_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a54cec48c6935a2c66a0087971d133fc52c1cc8bda5ba79dfb609d4ff4c0fbb8 +size 2762059 diff --git a/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/full.md b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c0efcfea7ef97ce8ad81d4d50db0d1b55c6862f4 --- /dev/null +++ b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/full.md @@ -0,0 +1,318 @@ +# Beyond Pixels: Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier + +Prantik Howlader1, Srijan Das2, Hieu Le3, and Dimitris Samaras1 + +1 Stony Brook University, New York, USA +2 University of North Carolina at Charlotte, NC, USA +3 EPFL, Lausanne, Switzerland + +Abstract. Incorporating pixel contextual information is critical for accurate segmentation. In this paper, we show that an effective way to incorporate contextual information is through a patch-based classifier. This patch classifier is trained to identify classes present within an image region, which facilitates the elimination of distractors and enhances the classification of small object segments. Specifically, we introduce Multi-scale Patch-based Multi-label Classifier (MPMC), a novel plug-in module designed for existing semi-supervised segmentation (SSS) frameworks. MPMC offers patch-level supervision, enabling the discrimination of pixel regions of different classes within a patch. Furthermore, MPMC learns an adaptive pseudo-label weight, using patch-level classification to alleviate the impact of the teacher's noisy pseudo-label supervision on the student. This lightweight module can be integrated into any SSS framework, significantly enhancing their performance. We demonstrate the efficacy of our proposed MPMC by integrating it into four SSS methodologies and improving them across two natural image and one medical segmentation dataset, notably improving the segmentation results of the baselines across all the three datasets. Code is available at: https://github.com/cvlab-stonybrook/Beyond-Pixels + +Keywords: Segmentation $\cdot$ Semi-supervised $\cdot$ Patches + +# 1 Introduction + +Supervised segmentation requires a large amount of pixel level annotations, which is laborious and time-consuming [53-56]. This paper focuses on semi-supervised segmentation (SSS) [1, 20], which alleviates the reliance on labeled annotations by leveraging a large quantity of unlabeled images, accompanied by a limited number of manually labeled images. + +The fundamental task in segmentation is a per-pixel classification. Thus contextual information of a pixel is important to improve its performance [31,32,68]. Context becomes more critical in SSS settings, which uses limited labeled data making the models more prone to overfitting. We note that recent approaches in SSS methods which predominantly leverage consistency regularization with + +![](images/295d3fe0ebeeedd571d70a575cb4f782811626d04b8e4161aefeb1890fa9e0be.jpg) +Fig.1: Segmentations and Multi-label classifications of image patches (Cityscapes) After 20 epochs, (b) and (c) are the segmentations and classes in the segmentations respectively, (d) is classes predicted by multi-label classifier in MPMC. (f) and (g) are segmentations by Baseline (AugSeg [65]) and Baseline + MPMC at the end of training. White boxes represent regions where MPMC improves the baseline, while orange boxes represent regions where segmentation network predicts classes not present in the patch. + +pseudo-labeling techniques [7,14,20,22,36,38,50,57,69] do not exploit contextual information explicitly. + +We hypothesize that patches are the simplest way to determine the context around pixels. The key observation is that the contextual information of an image patch are the classes present in it. For example, in the image patch in the first row of Fig. 1 (a) the classes present are car, road, vegetation and terrain. Thus, training a network to not only classify each pixel in an image but also predict the classes in the patch surrounding it, is a straight forward way to introduce contextual information in SSS pipeline. Moreover, utilizing multi-scale patch-level supervision can effectively encapsulate context at varying scales and distinguish pixels across different classes. To this end, we introduce a novel plugin module, termed as Multi-scale Patch-based Multilabel Classifier (MPMC). This module achieves enhanced discrimination of pixel regions between classes through additional multi-scale patch-level supervision. Notably, we observe that when training our baseline SSS model together with MPMC in limited labeled data scenario, the patch classifier exhibits greater resilience to noise than the segmentation network as illustrated in Fig. 1 (b) and (c). A key observation is when MPMC and baseline SSS model are trained together, the patch classifier provides patch-level labels, identifying the classes present within a region. This facilitates the elimination of distractors (first row in Fig. 1 (b), (c) and (d), while the segmentation network classifies pixels as truck which is not present in the patch, the multi-label classifier does not predict it) and enhances the classification of small object segments by the baseline method. + +Another important issue plaguing SSS methods is noisy pseudo-labels generated by the teacher for unlabeled images to supervise the student. This can lead to the student model overfitting to these noisy labels, consequently impairing its performance. Recent approaches have attempted to mitigate the impact of noise in pseudo-labels through various strategies, such as implementing a hard threshold [14, 23, 38], applying dynamic re-weighting to enhance the influence of reliable pixels [20], and employing negative data mining by archiving unreliable pixels in a memory bank [50]. In contrast to these approaches, we use the MPMC predictions to learn an adaptive weight that reduces the weight of noisy pseudo-labels in the teacher-student framework. Our approach hinges on the underlying observation that MPMC exhibits greater uncertainty (lower confidence) for noisy labels in a patch (see supplementary for details). + +In practice, MPMC is a parameterized module that can be plugged in the initial layers of the visual model, where it leverages the low-level pixel information to classify the classes present in each patch. This patch-level classification strategy forces the visual model to learn and preserve discriminative representations of classes, which are often prone to dilution in the deeper layers of the model. We demonstrate the efficacy of our proposed parameterized module, MPMC by integrating it into four SSS methodologies—UniMatch [57], AugSeg [65], AEL [20], and U2PL [50] - notably improving performance across all data partitions in two natural (Cityscapes and Pascal VOC) and one medical (ACDC) dataset. To summarize, our contributions are: + +1. We introduce a novel plug-in module, called MPMC that incorporates an auxiliary multi-scale multi-label classification task, enhancing pixel region discrimination across different classes via additional patch-level supervision. +2. MPMC consists of two adaptive learning weights that reduces the influence of noisy pseudo-labels from the teacher model. +3. We demonstrate MPMC's robustness by integrating in four SSS techniques, achieving consistent improvements across all protocols on Pascal VOC and Cityscapes. Notably, our method also shows improvement in medical image dataset ACDC, which indicates that our method is applicable beyond natural images. + +# 2 Related Work + +Semi-supervised learning (SSL) has been extensively researched in recent years. Popular SSL approaches include consistency regularization [2,10,11,14,38,43,58], entropy minimization [5,16], and pseudo-labeling [24-28, 46]. In this work, our emphasis is placed on the domains of pseudo-labeling and consistency regularization. + +Pseudo labeling: Pseudo-labeling methods [28,45] alongside self-training techniques [35, 59] are designed to leverage labeled image data to produce pseudolabels for unlabeled data. An inherent problem that plagues pseudo-labeling is noise, thus most approaches [42,44,46,69] devise strategies to refine the pseudolabels. PseudoSeg [69], for instance, aims at enhancing pseudo-label accuracy + +using attention mechanisms based on grad-cam [44]. In pseudo-label selection, FixMatch [46] adopts a confidence threshold, whereas UPS [42] relies on uncertainty measures. $\mathrm{ST}++$ [58] focuses on selecting unlabeled images more likely to yield accurate pseudo-pixels. Emerging approaches like U2PL [50] explore the idea of learning reliable pseudo-labels from their less reliable counterparts. CFCG [29] introduced a novel approach that combines cross-fusion and contour guidance for pseudo-label improvement. We differ from these methods by explicitly introducing patch-level contextual information in the SSS methods. + +Consistency Regularization: This approach enforces the model to have consistent predictions on different views of the unlabeled images. UniMatch [57] uses dual perturbations and feature perturbations to force the network to learn more consistent information. ICT [49] uses Mixup [64] augmentation for consistency regularization. Recent methods [14, 58, 62] use Cutout [9], Cutmix [63], Class-mix [37] as strong data augmentation. Consistency Regularization is used with pseudo-labeling techniques in Mixmatch [4] and TC-SSL [67]. In our work, we use a set of strong and weak data augmentations [1, 65] for generating pseudo-labels. + +Multi-label classification: This approach classifies an image into more than one label present in the image. Most works in this field explore finding the correlation between the labels and the images [21,39,61]. Recent works mainly focus on solving data imbalance issues [17,51]. We use multi-label classification to help in segmentation. Some works [15,40] use multi-label classification for segmentation in weakly supervised scenarios, for improved CAM-based pseudo-segmentation. However, we are using multi-label classification in a semi-supervised setting to introduce patch-level contextual information and also use the classifier confidence to reduce the impact of noisy pseudo-labels. + +# 3 Semi-Supervised Segmentation Pipeline + +Given a labeled set $\mathcal{D}^l = \{(x^l,y^l)\}$ and an unlabeled set $\mathcal{D}^u = \{x^u\}$ , the goal of semi-supervised semantic segmentation is to develop a model that effectively utilizes both labeled and unlabeled data. This section commences with an overview of the foundational framework for semi-supervised segmentation in Section 3.1. followed by an overview of our proposed MPMC in Section 3.2. The section concludes with an elaboration of MPMC's functionality (Section 3.3). + +# 3.1 Preliminary + +Teacher-Student frameworks leverage weak-to-strong consistency regularization to leverage unlabeled data. As depicted in Fig. 2(a), each unlabeled image $x^{u}$ is simultaneously perturbed by two operators, i.e., weak perturbation $A^w$ (cropping, random flip, etc), and strong perturbation $A^s$ (color jitter, gaussian blur, contrast, etc). The framework's objective function is a convex combination of supervised loss $\mathcal{L}_l$ for the labeled data and unsupervised loss $\mathcal{L}_u$ for the unlabeled data as: + +$$ +\mathcal {L} = \mathcal {L} _ {l} + \alpha \mathcal {L} _ {u} \quad \text {w h e r e ,} \quad \mathcal {L} _ {u} = \frac {1}{| \mathcal {D} ^ {u} |} \sum \mathbb {1} \left(p ^ {w} \geq t h\right) \mathcal {E} \left(p ^ {w}, p ^ {s}\right) \tag {1} +$$ + +![](images/c9964a88ca55309adb44c838c691a9ad905479a6b4211da23387a4297b5e521e.jpg) +Fig. 2: Overall Pipeline of our novel multi-label classification based semantic segmentation: (a) Left: End-to-end Teacher-Student Pipeline with our novel method Multi-scale Patch-based Multi-label Classifier (MPMC). (b) Right: For unlabeled images, The teacher MPMC extracts features from a layer in the segmentation model's encoder to classify the feature's receptive field patch. The confidence of MPMC for a class in a patch is used to calculate two adaptive weights $\lambda_{s}$ and $\lambda_{m}$ which is used to reduce the influence of noisy predictions in that patch from the teacher to train both the student segmentation network and MPMC. + +![](images/33e39ec217d41dfae56c59a380a7ca5e742d5a2dacce2980edd14482500e8734.jpg) + +where, $\mathcal{L}_l$ is the standard cross-entropy loss, comparing the model's predictions with ground truth. While $\mathcal{L}_u$ aims to regularize the predictions for an unlabeled sample so that they remain consistent under both weak and strong augmentations ( $p^w$ and $p^s$ ). Here, $th$ is a predefined confidence threshold to filter noisy pseudo labels and $\mathcal{E}$ minimizes the entropy between two probability distributions. + +# 3.2 Overview + +Fig. 2(a) illustrates the comprehensive training process of our proposed MPMC, which is tailored to enhance the discrimination of patches among different classes and to mitigate the impact of noisy predictions transmitted from the teacher to the student. MPMC can be integrated into both teacher and student models within the teacher-student framework. It is composed of a classifier that performs multi-label classifications of the multi-scale patch representations of the input image. The application of MPMC spans two distinct scenarios: 1) For labeled images, MPMC assigns classes to each patch representation based on the classes present within a patch. The multi-scale nature of these patch representations and their classification enhances the segmentor's ability to discriminate classes by capturing enhanced contextual information related to variations in shape and size. 2) In the case of unlabeled images, besides patch classification, MPMC computes a pseudo-label weight map to alleviate the impact of the teacher's noisy pseudo-label supervision on the student. This involves assigning weights to the pseudo-labels within each patch in consistency learning framework, which are determined by the confidence of the MPMC's predictions for each class in the patch, to enhance the reliability and accuracy of the pseudo-labeling process. + +# 3.3 Multi-scale Patch-based Multi-label Classifier (MPMC) + +Our proposed MPMC module is specifically designed to address the shortcomings of SSS pipeline. MPMC can be plugged in at any layer within the teacher-student network, both within the teacher model, denoted as $g$ , and within the student model, denoted as $\hat{g}$ , they share the same parameters. Below, we detail the training of MPMC on labeled images and then on unlabeled images. + +MPMC on Labeled Images. As illustrated in Fig. 2, MPMC incorporates patch-supervision into the segmentation framework, by training a multi-label classifier that takes as input a single pixel in the feature map of the $\tau^{th}$ layer of the encoder. This single pixel-representation encodes the information from a patch (receptive field) in the original image. Further, to encapsulate context at varying scales we perform average pooling on this features map with kernels of different sizes before performing multi-label classification. Let us define a patch of image and the corresponding segmentation in a receptive field $r$ as $\{x_r^l,y_r^l\} \in \mathcal{D}^l$ . Note that $y_{r}^{l}$ is a multi-label target where the classes present in the receptive field $r$ are positive classes, and the classes not present are negative classes. + +The input to the student MPMC $\hat{g}$ are the features extracted from the $\tau^{th}$ layer of the encoder in student segmentation network $\hat{f}$ , denoted by $z_{r} = \hat{f}_{\tau}(\mathcal{A}_w(x_r^l))$ . MPMC captures the contextual information within a patch at different scales, $S_{i}$ , $i \in \{0,1\}$ , where $s_0$ and $s_1$ are average pool kernel sizes. For all our experiments $s_0$ and $s_1$ are set to $7 \times 7$ and $19 \times 19$ . These pooling operations are performed with appropriate padding to maintain uniformity in the shape of the output features. MPMC $\hat{g}$ is composed of the concatenation of these multiscale pooled features, followed by a convolutional network $\mathcal{M}$ and a final linear layer. Thus, the output of the MPMC $\hat{g}$ for the input patch $x_{r}^{l}$ is given by: + +$$ +q ^ {l} = \sigma \left(W ^ {T} \mathcal {M} \left(\operatorname {c o n c a t} \left(z _ {r}, z _ {r} ^ {S _ {0}}, z _ {r} ^ {S _ {1}}\right)\right) + b\right) \tag {2} +$$ + +where $W$ and $b$ are linear layer weights and bias, $\sigma$ is the sigmoid activation and $q^l = [q_1^l,\dots,q_C^l]^T\in \mathbb{R}^C$ are the predicted probabilities for $C$ classes. + +While binary cross-entropy loss is applicable to train MPMC, it indiscriminately weighs all positive and negative classes equally. Considering the receptive field of a feature from the $\tau^{th}$ layer of the encoder in the segmentation network $\hat{f}$ covers only a small patch relative to the input image's size, the prevalence of negative classes within a patch significantly exceeds that of positive classes. This imbalance influences the binary cross-entropy loss to favor negative classes [51]. Inspired by long-tailed multi-label classification [30, 51] which mitigate such imbalances by diminishing the weight of negative classes, we adapt an asymmetric loss [41], a focal loss variant that assigns distinct $\gamma$ values to positive and negative classes. For an image patch $x_{r}^{l}$ and the predicted class probabilities $q^{l}$ using our framework, the asymmetric focal loss is expressed as: + +$$ +l _ {r} ^ {M} = \frac {1}{C} \sum_ {c = 1} ^ {C} \left\{ \begin{array}{l l} (1 - q _ {c} ^ {l}) ^ {\gamma^ {+}} \log (q _ {c} ^ {l}), & y _ {r, c} ^ {l} = 1. \\ (q _ {c} ^ {l}) ^ {\gamma^ {-}} \log (1 - q _ {c} ^ {l}), & y _ {r, c} ^ {l} = 0. \end{array} \right. \tag {3} +$$ + +In our experiments, we set $\gamma^{+} = 0$ and $\gamma^{-} = 1$ . Subsequently, MPMC $\hat{g}$ undergoes training with multi-label supervision $\mathcal{L}_l^M$ , formulated as: + +$$ +\mathcal {L} _ {l} ^ {M} = \frac {1}{| D ^ {l} |} \sum_ {\left(x ^ {l}, y ^ {l}\right) \in D ^ {l}} \sum_ {r = 1} ^ {R} l _ {r} ^ {M} \tag {4} +$$ + +where $\mathrm{R}$ is the number of receptive fields at $\tau^{th}$ layer of the segmentor. This auxiliary task of patch classification augments the network's proficiency in learning and preserving class features, a critical function particularly pertinent as the fidelity of these features tends to diminish in the network's deeper layers. + +MPMC on Unlabeled images. In semi-supervised teacher-student framework, for unlabeled images, the prediction of the teacher network $f$ for the weak augmented view $\mathcal{A}^w(x^u)$ is used to supervise the student network $\hat{f}$ for the strong augmented view $\mathcal{A}^s(x^u)$ . In the early iterations of semi-supervised learning, the teacher's predictions are inherently noisy. Common practices [13, 20, 33, 46, 70], involve employing pixel-wise confidence score thresholds (th) to filter out unreliable pseudo labels. However, a high confidence threshold does not entirely eliminate noise due to the teacher's initial miscalibration [42], allowing some noisy predictions to have high confidence. To refine this process, our approach does not solely depend on confidence-thresholded teacher segmentation predictions for reliable pseudo-label extraction. Instead, we leverage the teacher MPMC $g$ 's confidence for a patch to adjust the weights of the deemed reliable pseudo-labels within that patch. The underlying principle of our methodology is that MPMC exhibits greater uncertainty (lower confidence) for noisy labels. + +We employ the teacher MPMC $g$ 's predictions to define two adaptive learning weights $\lambda_{m} \in \mathbb{R}^{R \times C}$ and $\lambda_{s} \in \mathbb{R}^{W \times H}$ , where $R$ denotes the number of patches, sized according to the receptive field within the image, $W$ and $H$ represents the width and height of the input image, respectively. The learning weight $\lambda_{m}$ is utilized to assign class weights to each patch for training the student MPMC, whereas $\lambda_{s}$ is applied to diminish the impact of noisy predictions during training the student segmentation network. + +For an unlabeled image $x^{u}$ , consider a patch at the receptive field $r$ , denoted as $x_{r}^{u}$ . The teacher and student segmentation networks yield predictions $p^{w} = f(\mathcal{A}^{w}(x_{r}^{u}))$ and $p^{s} = \hat{f} (\mathcal{A}^{s}(x_{r}^{u}))$ , respectively. Simultaneously, the teacher and student MPMC predict $q_{r}^{u} = g(f_{\tau}(\mathcal{A}^{w}(x_{r}^{u})))$ and $\hat{q}_r^u = \hat{g} (\hat{f}\tau (\mathcal{A}^s (x_r^u)))$ using features from the $\tau^{th}$ layer of the encoder. Let us define the confidence of teacher MPMC $g$ across all classes in patch $r$ as $\gamma = \mathrm{softmax}(q_r^u)$ , with $\gamma \in \mathbb{R}^C$ . The weight of the patch $r$ , $\lambda_{m,r}$ is determined by $\gamma$ . For each pixel in patch $r$ associated with class $c$ , $\lambda_{s}$ is assigned the weight $\gamma_{c}$ . This approach assigns weights to pseudo-labels and pixels based on the teacher MPMC's confidence, enhancing the accuracy of the learning process. + +Thus, the control parameters $\lambda_{s}$ and $\lambda_{m}$ are integrated to the unsupervised losses to train the student segmentation network and MPMC. The unsupervised + +segmentation loss $L_{u}$ mentioned in equation 1 to train the student segmentation network $\hat{f}$ is re-formulated as: + +$$ +\mathcal {L} _ {u} = \frac {1}{| \mathcal {D} ^ {u} |} \sum_ {x ^ {u} \in \mathcal {D} ^ {u}} \mathbb {1} \left(p ^ {w} \geq t h\right) \lambda_ {s} \mathcal {E} \left(p ^ {w}, p ^ {s}\right) \tag {5} +$$ + +where $\mathcal{E}$ is the standard cross entropy loss and $th$ is the predefined segmentation confidence threshold. Finally, the unsupervised multi-label loss $\mathcal{L}_u^M$ to train the student MPMC $\hat{g}$ is given by: + +$$ +\mathcal {L} _ {u} ^ {M} = \frac {1}{| \mathcal {D} ^ {u} |} \sum_ {x ^ {u} \in D ^ {u}} \sum_ {r = 1} ^ {R} \frac {1}{C} \sum_ {i = 1} ^ {C} \lambda_ {m, r} ^ {i} \mathcal {E} ^ {+} \left(q _ {r} ^ {u}, \hat {q} _ {r} ^ {u}\right) \tag {6} +$$ + +where $\mathcal{E}^+$ is the binary cross entropy loss. Therefore, the overall loss function is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {l} + \mathcal {L} _ {l} ^ {M} + \alpha \mathcal {L} _ {u} + \beta \mathcal {L} _ {u} ^ {M} \tag {7} +$$ + +where $\alpha$ and $\beta$ controls the contribution of $\mathcal{L}_u$ and $\mathcal{L}_u^M$ . In all our experiments $\alpha$ and $\beta$ are set to 0.1 and 0.25 respectively. + +# 4 Experiments + +# 4.1 Setup + +Datasets: PASCAL VOC 2012 [12] is a standard semantic segmentation dataset with 20 semantic classes and a background class. The training, validation, and testing sets consist of 1464, 1449, and 1556 images (Classic), respectively. Following [7, 65], we also include the additional augmented dataset [18] (Blender), which includes a collection of 10582 training images. We adopt the same partition protocols in [7, 65] to evaluate our method in both Classic and Blender sets. Cityscapes [8] dataset is designed for urban scene understanding, consisting of 30 classes, of which only 19 classes are used for scene parsing evaluation. It has 2975, 500, and 1525 images for training, validation, and testing, respectively. ACDC [3] dataset contains 200 annotated short-axis cardiac cineMR images from 100 patients. The segmentation masks consists of classes: left ventricle (LV), myocardium (Myo), and right ventricle (RV). 140 images from 70 patients, 60 images from 30 patients are randomly selected for training and validation respectively. + +Implementation Details: For fair comparison, we adopt DeepLabv3+ [6] as the decoder in all of our experimental setups, and compare with both ResNet-101 and ResNet-50 [19] as the backbone architecture. In all our experiments with Cityscapes, Pascal, and ACDC datasets, the training resolution is 801, 513, and 224 respectively. Note, for maintaining fair comparison, our method "UniMatch [57] + MPMC" uses the training resolution of 321 which is used by UniMatch [57] when training on the Classic set of Pascal VOC. Further when training a MPMC + +integrated baseline, we use the same weak and strong augmentations as used by the corresponding baselines. The convolutional network in MPMC is comprised of three convolution blocks. Each convolution block consists of a stack of two $3 \times 3$ convolutions followed by a $1 \times 1$ convolution. For all our experiments MPMC is integrated to the first layer of the ResNet Encoder. + +# 4.2 Comparison with State-of-the-Art Methods + +We conducted our experiments on PASCAL VOC 2012 and Cityscapes which are two popular natural image benchmarks, and a medical dataset ACDC. MPMC is integrated in four semi supervised methods UniMatch [57], AEL [20], U2PL [50] and AugSeg [65]. Note, UniMatch [57] is a consistency regularization based method. We observe that MPMC consistently improves over their baseline performance across all partition protocols on all datasets. + +Results on PASCAL VOC 2012 Dataset. Table 1 shows the comparison with other SOTA methods for both the Classic and Blender sets. MPMC integrated to the four baseline methods improves them consistently across all data partitions for both Classic and Blender sets. We find that our method brings the most improvement for the least labeled data partition $(\frac{1}{16})$ , improving AugSeg, AEL, U2PL and UniMatch by $1.8\%$ , $2.5\%$ , $2.0\%$ and $2.1\%$ respectively in Classic with a ResNet-101 encoder. + +Results on Cityscapes Dataset. Table 2 shows the comparison with other SOTA methods. MPMC integrated to the four baseline methods improves them consistently across all data partitions. We observe that similar to results in PASCAL VOC 2012 dataset, our method brings the most improvement for the least labeled data partition $(\frac{1}{16})$ , improving AugSeg, AEL, U2PL and UniMatch by $1.7\%$ , $2.1\%$ , $2.3\%$ and $1.9\%$ respectively in Classic for ResNet-101 encoder. + +Results on ACDC Dataset. Table 3 shows the comparison with other SOTA methods. We observe that when MPMC is integrated to the four baseline methods, it significantly improves the performance for the "3 cases" data partition which has the least labeled images ( $< 10\%$ of labeled images). Note we use ResNet-101 for all experiments in this dataset. + +The consistent improvement in SSS performance in both natural and medical datasets and its ability to integrate in different SSS methods substantiates the robustness of MPMC. Moreover, the notable improvements achieved in scenarios with limited data shows MPMC's vital efficacy within SSS pipelines. + +Qualitative Results In Fig. 3 we visualise the results of integration of our method (MPMC) to AugSeg [65]. Introducing patch-level supervision with the integration of MPMC, leads the SSS model to better segment confusing classes. Among comparisons we show that MPMC leads to better segmentation of pixels regions of confusing classes like bus, train and truck. These classes are often misclassified not only because of sharing visual similarity but also having low training instances. + +Table 1: Quantitative results of different semi-supervised segmentation methods on Pascal VOC classic and blender set. We report Mean IoU under various partition protocols and show the improvements $(\Delta)$ over corresponding baseline. + +
MethodClassicBlender
1/161/81/41/2Full1/161/81/4
ResNet-50
Supervised44.052.361.766.772.9---
PC2Seg [66][CVPR'21]56.964.667.670.972.3---
PseudoSeg [69][ICLR'21]56.964.667.670.972.2---
ST++ [58][CVPR'22]-----72.674.475.4
AugSeg [65][CVPR'21]64.272.176.177.478.877.278.278.2
AugSeg [65] + MPMC (Δ)65.973.777.678.579.779.079.879.7
(+1.7)(+1.6)(+1.5)(+1.1)(+0.9)(+1.8)(+1.6)(+1.5)
AEL [20][NeurIPS'21]62.964.170.372.774.074.176.177.9
AEL [20] + MPMC (Δ)65.066.172.174.275.376.178.079.6
(+2.1)(+2.0)(+1.8)(+1.5)(+1.3)(+2.0)(+1.9)(+1.7)
U2PL [50][CVPR'22]63.365.571.673.875.174.777.477.5
U2PL [50] + MPMC (Δ)65.267.473.375.276.276.579.179.1
(+1.9)(+1.9)(+1.7)(+1.4)(+1.1)(+1.8)(+1.7)(+1.6)
UniMatch [57][CVPR'23]71.972.576.077.478.778.179.079.1
UniMatch [57] + MPMC (Δ)73.574.077.478.579.479.680.480.4
(+1.6)(+1.5)(+1.4)(+1.1)(+0.7)(+1.5)(+1.4)(+1.3)
ResNet-101
Supervised45.155.364.869.773.570.675.076.5
CPS [7][CVPR'21]64.167.471.775.9-72.275.877.6
PS-MT [33][CVPR'22]65.869.676.678.480.075.578.278.7
PCR [52][NeurIPS'22]70.174.777.278.580.778.680.780.8
DAW [47][NeurIPS'23]74.877.479.580.681.578.578.979.6
CFCG [29][ICCV'23]-----77.479.480.4
AugSeg [65][CVPR'21]71.175.578.880.381.479.381.580.5
AugSeg [65] + MPMC (Δ)72.977.879.981.382.180.882.581.4
(+1.8)(+2.3)(+1.1)(+1.0)(+0.7)(+1.5)(+1.0)(+0.9)
AEL [20][NeurIPS'21]66.168.371.974.478.977.277.678.1
AEL [20] + MPMC (Δ)68.670.373.576.180.179.279.379.2
(+2.5)(+2.0)(+1.6)(+1.7)(+1.2)(+2.0)(+1.7)(+1.1)
U2PL [50][CVPR'22]68.069.273.776.279.577.279.079.3
U2PL [50] + MPMC (Δ)70.070.875.377.880.579.280.780.3
(+2.0)(+1.6)(+1.6)(+1.6)(+1.0)(+2.0)(+1.7)(+1.0)
UniMatch [57][CVPR'23]75.277.278.879.981.280.981.980.4
UniMatch [57] + MPMC (Δ)77.378.679.880.881.782.583.181.5
(+2.1)(+1.4)(+1.0)(+0.9)(+0.5)(+1.6)(+1.2)(+1.1)
+ +![](images/bed05dbb3298e38f171f9ff4cd49a43ee97729d555d8f7cf552826fe975d7199.jpg) + +![](images/60cf4c0ef9b134097920c04cfa6f34a4885eec20acdae5a67d979188cf0351b1.jpg) + +![](images/cc9a1e4163be66a3ce0f31047a6e3f451313d4f6fc5a954160c408fc210f94cc.jpg) + +![](images/adbb23c58d04ed1cfbad440b34db33fc59801a9e84ccfb40a6d51ab6fbd66619.jpg) +(a) Input Image + +![](images/2e3599f5d7734cff4daa0062343dadfdcc3d7c89ace091c8cd097818991febef.jpg) + +![](images/052331a1214c8da78af967a541f7a03ecf928ae1722e41867d3b71024951022d.jpg) + +![](images/7906a90b19af6415d8d8372f9ffe0bbe48e598f9eeecc46b57c34b3cef36dd07.jpg) + +![](images/35a09bf93191eae0bf507e428bab2f6620770ed3d7c3e28256b377da6b53d469.jpg) +(b) Ground Truth + +![](images/844d0b6f789c3f408683c58dc93226c9aeea6adac0bede4e7f11677457667c15.jpg) + +![](images/1d6be199f70055a0cc81f08134cf7299523a01fad44d2bd9058e05a82ef528e5.jpg) + +![](images/7501babaf86065a20750c0c62afbb5223dd0b780095956b969ce1a55f698f626.jpg) + +![](images/1d8afd4a8d9642127033b56d7fae319d346e519153a63c78e08fa67785895d1d.jpg) +(c) AugSeg +Fig. 3: Qualitative Results on Cityscapes dataset: (a) is the original image, (b) is the Ground Truth, (c) are segmentations generated by AugSeg [65] compared to (d) which are segmentations generated when our method (MPMC) is integrated to AugSeg. White boxes show the areas where our method improves the baseline. + +![](images/719b8463a09340c28f2f15e8434a4c25e60bca204b0dc80adf13545bddb8f90b.jpg) + +![](images/e35b9b096407e70c93ec3557f151f3d1340a987b148d6e5004c0a5bfaa5510cd.jpg) + +![](images/d6587c8537d1345a9f4b1b6b452e85c51b7f9f49b60bfaf146375f59bd05a29a.jpg) + +![](images/04a39c479d40f504f57cc89b08e1f8694e445c62790a1f6a13c130c09c57b9f2.jpg) +(d) AugSeg + MPMC + +Table 2: Quantitative results of different semi-supervised segmentation methods on Cityscapes validation set. We report Mean IoU under various partition protocols and show the improvements $(\Delta)$ over the corresponding baseline. + +
MethodResNet-50ResNet-101
1/161/81/41/21/161/81/41/2
Supervised63.3468.7374.1476.6266.372.875.078.0
CPS [7][CVPR'21]69.7974.3976.8578.6469.874.374.676.8
PS-MT [33][CVPR'22]-75.7676.9277.64-76.977.679.1
PCR [52][NeurIPS'22]----73.476.378.479.1
CFCG [29][ICCV'23]76.178.979.380.177.879.680.480.9
AugSeg [65][CVPR'21]73.776.478.779.375.277.879.680.4
AugSeg [65] + MPMC (Δ)75.3 (+1.6)77.7 (+1.3)79.8 (+1.1)80.3 (+1.0)76.9 (+1.7)79.2 (+1.4)80.9 (+1.3)81.3 (+0.9)
AEL [20][NeurIPS'21]68.272.774.977.574.575.677.579.0
AEL [20] + MPMC (Δ)70.5 (+2.3)74.5 (+1.8)76.6 (+1.7)79.0 (+1.5)76.6 (+2.1)77.5 (+1.9)78.9 (+1.4)80.3 (+1.3)
U2PL [50][CVPR'22]69.073.076.378.674.976.578.579.1
U2PL [50] + MPMC (Δ)71.2 (+2.2)74.6 (+1.6)77.8 (+1.5)79.8 (+1.2)77.2 (+2.3)78.2 (+1.7)80.0 (+1.5)80.0 (+0.9)
UniMatch [57][CVPR'23]75.076.877.578.676.677.979.279.5
UniMatch [57] + MPMC (Δ)76.8 (+1.8)78.0 (+1.2)78.6 (+1.1)79.5 (+0.9)78.5 (+1.9)79.2 (+1.3)80.9 (+1.7)81.3 (+1.8)
+ +Table 3: Quantitative results of different semi-supervised segmentation methods on ACDC dataset with 3/7 labeled cases. We report Dice Similarity Coefficient (DSC) metric averaged over 3 classes. We use ResNet-101 for all experiments in this dataset. + +
Method3 cases7 cases
Supervised41.562.5
UA-MT [60][MICCAI'19]61.081.5
CPS [7][CVPR'21]60.383.3
CNN & Trans [34][PMLR'22]65.686.4
AugSeg [65][CVPR'21]85.187.3
AugSeg [65] + MPMC (Δ)88.6 (+3.5)88.4 (+1.1)
AEL [20][NeurIPS'21]82.787.6
AEL [20] + MPMC (Δ)85.4 (+2.7)88.5 (+0.9)
U2PL [50][CVPR'22]84.289.2
U2PL [50] + MPMC (Δ)87.4 (+3.2)90.5 (+1.3)
UniMatch [57][CVPR'23]88.989.9
UniMatch [57] + MPMC (Δ)91.8 (+2.9)90.4 (+0.5)
+ +# 4.3 Ablation Studies + +We conduct experiments to study the impact of various components of our approach. All experiments in this section were conducted on $\frac{1}{16}$ data partition of the Cityscapes dataset, and use AugSeg [65] as the baseline. + +Segmentation performance at different scales. Here we analyse how scaling of input features for patch-classification, improves the performance of the SSS method. We consider two scales - average pool kernels $7 \times 7$ and $19 \times 19$ . Fig. 4 illustrates for two random classes pole and rider, that average pooling with a kernel size of $7 \times 7$ is better at enhancing the pixel accuracy for instances with smaller areas, whereas the $19 \times 19$ kernel size is better at improving accuracy for instances with larger areas. This can be attributed to the larger receptive field of the $19 \times 19$ kernel, enabling more effective contextual learning for larger regions. While the smaller receptive field of $7 \times 7$ kernel is better able to discern smaller objects. + +Analysing class features The crux of our method is to better discriminate pixel regions between different classes. Here we analyze if MPMC integrated to AugSeg [65] can better discriminate features than the baseline for different classes in Cityscapes Dataset. We use the pixel features from the penultimate layer of the segmentation network to analyse via t-SNE. Fig 5 illustrates for four random classes - car, bus, train and truck, integration of MPMC leads to better discrimination between the classes, i.e., compared to AugSeg [65], the features of train and truck are more clustered and separated from the rest. + +Influence of MPMC on patch-level segmentation. Our method MPMC improves segmentation by introducing patch classification as an auxiliary task. We analyse at patch level the correlation between multi-label accuracy and the Mean IoU performance of the segmentation network. We use (1 - hamming loss) [48] as a measure of multi-label classification and use the basic teacher- + +![](images/5d3bccfe2eb4b7b2bad596ca3c896c83c810ee05e7923e4da7d8335435c1f3f5.jpg) +Fig. 4: Pixel accuracy across different scales with respect to instance size. For two different scales, we compare the pixel accuracy with respect to size of instances for two random classes: pole and rider. + +![](images/200a277845cb106a9a3c7de7a37c13b633c4a9c85a68fb5915484fc33509372f.jpg) + +![](images/dbc537f6b47ae7aab8d16e8d61b3f3a63b42315d12ae7d28b493dca4d52bb9da.jpg) +Fig. 5: Analysis of class features (a) t-SNE of features predicted by AugSeg [65], (b) t-SNE of features predicted by AugSeg [65] + MPMC. + +student framework mentioned in Section 3.1 as our baseline. Table 4 illustrates that the avg MIoU improvement is highest (4.05%) in patches having more than 90% multi-label classification accuracy showing that MPMC improves segmentation at the patch level. + +Impact of MPMC on semantic segmentation in limited labeled data scenarios. + +In Table 5 shows the impact of MPMC on segmentation when trained with only labeled data. We use the basic teacher-student framework mentioned in Section 3.1 as our baseline, which achieves MIoU of 66.3, 72.8 and 75.0 in 186, 372 and 744 partitions of labeled images respectively. After introducing PMC, there is an improvement of $1.9\%$ , $1.5\%$ and $1.3\%$ in 186, 372 and 744 partitions of labeled images respectively. We observe that MPMC improves segmentation performance in limited-labeled data scenarios, with the improvement highest for least labeled data, i.e., $1.9\%$ for 186 labeled images. + +Table 4: Analysis of the influence of MPMC on segmentation at patch-level. + +
1 - hamming loss90+80-9070-8060-7050-6040-5030-4020-3010-200-10
% patches12.625.220.58.45.15.24.110.22.46.3
Avg M IoU Δ+4.05+1.57+0.88+1.55+0.91+1.15+0.49+0.10-1.90-2.20
+ +Table 5: Impact of MPMC on segmentation (MIoU) when trained with different volumes of labeled data. MPMC shows the highest improvement $(1.9\%)$ in the least labeled data partition. + +
# Labeled Data
186372744
w/o MPMC66.372.875.0
w/ MPMC68.274.376.3
Gain/Δ+1.9+1.5+1.3
+ +Table 6: Ablation study of different component: MPMC on labeled images (MPMCL) and MPMC on unlabeled images (MPMCUL) + +
MPMCLMPMCUL1/16 (186)1/8 (372)1/4 (744)
68.473.574.2
✓70.875.475.7
✓71.376.076.6
✓✓73.176.777.3
+ +Analysis of different components We ablate each component of MPMC with basic teacher-student framework (Section 3.1) as the baseline. MPMC on labeled images (MPMCL), which introduces multi-scale patch-based classification on only the labeled data, achieves an improvement of $2.4\%$ , $1.9\%$ , and $1.5\%$ under $\frac{1}{16}$ , $\frac{1}{8}$ and $\frac{1}{4}$ partition protocols respectively. MPMC on unlabeled images (MPMCUL) leads to a further improvement of $2.9\%$ , $2.5\%$ , and $2.4\%$ under $\frac{1}{16}$ , $\frac{1}{8}$ and $\frac{1}{4}$ partition protocols respectively. Thus, MPMC is effective on both labeled and unlabeled instances in SSS pipelines. + +# 5 Conclusion + +In this paper, we have proposed a novel Multi-scale Patch-based Multi-label Classifier (MPMC), which introduces patch-level context information in the form of an auxiliary task of multi-scale patch-based multi-label classification. When jointly trained with a SSS model, the patch classifier provides patch-level labels, identifying the classes present within a region. This facilitates the elimination of distractors and enhances the classification of small object segments by the baseline method. Further, the MPMC predictions are used to learn an adaptive weight that reduces the influence of noisy pseudo-labels in the teacher-student framework. We performed a thorough analysis of the importance of multi-scale patch-level contextual information in improving segmentation especially in low data scenarios. Further, we demonstrated the efficacy of our proposed parameterized module MPMC by integrating it into four SSS methodologies and showing consistent improvements in Cityscapes, Pascal VOC and ACDC benchmarks. + +Acknowledgement. This work was partially supported by the National Science Foundation (IIS-2245652, IIS-2123920, IIS-2212046). + +# References + +1. Alonso, I., Sabater, A., Ferstl, D., Montesano, L., Murillo, A.C.: Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8219-8228 (2021) +2. Bachman, P., Alsharif, O., Precup, D.: Learning with pseudo-ensembles. Advances in neural information processing systems 27 (2014) +3. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37(11), 2514-2525 (2018) +4. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems 32 (2019) +5. Chen, H., Jin, Y., Jin, G., Zhu, C., Chen, E.: Semisupervised semantic segmentation by improving prediction confidence. IEEE Transactions on Neural Networks and Learning Systems (2021) +6. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018) +7. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021) +8. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016) +9. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017) +0. Durasov, N., Dorndorf, N., Le, H., Fua, P.: Zigzag: Universal sampling-free uncertainty estimation through two-step inference. Transactions on Machine Learning Research (2024) +1. Durasov, N., Oner, D., Donier, J., Le, H., Fua, P.: Enabling uncertainty estimation in iterative neural networks. In: Forty-first International Conference on Machine Learning (2024) +2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88(2), 303-338 (2010) +3. Fan, J., Gao, B., Jin, H., Jiang, L.: Ucc: Uncertainty guided cross-head co-training for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9947-9956 (2022) +4. French, G., Laine, S., Aila, T., Mackiewicz, M., Finlayson, G.: Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916 (2019) +5. Ge, W., Yang, S., Yu, Y.: Multi-evidence filtering and fusion for multi-label classification, object detection and semantic segmentation based on weakly supervised learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1277-1286 (2018) + +16. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. Advances in neural information processing systems 17 (2004) +17. Guo, H., Wang, S.: Long-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15089-15098 (2021) +18. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011) +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +20. Hu, H., Wei, F., Hu, H., Ye, Q., Cui, J., Wang, L.: Semi-supervised semantic segmentation via adaptive equalization learning. Advances in Neural Information Processing Systems 34, 22106-22118 (2021) +21. Huang, J., Li, G., Huang, Q., Wu, X.: Learning label specific features for multi-label classification. In: 2015 IEEE International Conference on Data Mining. pp. 181-190. IEEE (2015) +22. Huo, X., Xie, L., He, J., Yang, Z., Zhou, W., Li, H., Tian, Q.: Atso: Asynchronous teacher-student optimization for semi-supervised image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1235-1244 (2021) +23. Ke, Z., Qiu, D., Li, K., Yan, Q., Lau, R.W.: Guided collaborative training for pixel-wise semi-supervised learning. In: European conference on computer vision. pp. 429-445. Springer (2020) +24. Le, H., Goncalves, B., Samaras, D., Lynch, H.: Weakly labeling the antarctic: The penguin colony case. In: CVPRW (June 2019) +25. Le, H., Samaras, D., Lynch, H.J.: A convolutional neural network architecture designed for the automated survey of seabird colonies. Remote Sensing in Ecology and Conservation 8(2), 251-262 (2022) +26. Le, H., Vicente, T.F.Y., Nguyen, V., Hoai, M., Samaras, D.: A+D Net: Training a shadow detector with adversarial shadow attenuation. In: European Conference on Computer Vision (ECCV) (2018) +27. Le, H., Yu, C.P., Zelinsky, G., Samaras, D.: Co-localization with category-consistent features and geodesic distance propagation. In: ICCVW (2017) +28. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896 (2013) +29. Li, S., He, Y., Zhang, W., Zhang, W., Tan, X., Han, J., Ding, E., Wang, J.: Ccfg: Semi-supervised semantic segmentation via cross-fusion and contour guidance supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16348-16358 (2023) +30. Lin, D.: Probability guided loss for long-tailed multi-label image classification. In: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI'23/IAAI'23/EAAI'23, AAAI Press (2023). https://doi.org/10.1609/aaai.v37i2.25244, https://doi.org/10.1609/aaai.v37i2.25244 +31. Lin, G., Shen, C., Van Den Hengel, A., Reid, I.: Efficient piecewise training of deep structured models for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3194-3203 (2016) + +32. Liu, W., Rabinovich, A., Berg, A.C.: Parsenet: Looking wider to see better. arXiv preprint arXiv:1506.04579 (2015) +33. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4258-4267 (2022) +34. Luo, X., Hu, M., Song, T., Wang, G., Zhang, S.: Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In: International Conference on Medical Imaging with Deep Learning. pp. 820-833. PMLR (2022) +35. McClosky, D., Charniak, E., Johnson, M.: Effective self-training for parsing. In: Proceedings of the Human Language Technology Conference of the NAACL, Main Conference. pp. 152-159 (2006) +36. Oliver, A., Odena, A., Raffel, C.A., Cubuk, E.D., Goodfellow, I.: Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems 31 (2018) +37. Olsson, V., Tranheden, W., Pinto, J., Svensson, L.: Classmix: Segmentation-based data augmentation for semi-supervised learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1369-1378 (2021) +38. Ouali, Y., Hudelot, C., Tami, M.: Semi-supervised semantic segmentation with cross-consistency training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12674-12684 (2020) +39. Read, J., Pfahringer, B., Holmes, G., Frank, E.: Classifier chains for multi-label classification. Machine learning 85, 333-359 (2011) +40. Reif, S., Seibold, C., Freytag, A., Rodner, E., Stiefelhagen, R.: Every annotation counts: Multi-label deep supervision for medical image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9532-9542 (2021) +41. Ridnik, T., Ben-Baruch, E., Zamir, N., Noy, A., Friedman, I., Protter, M., Zelnik-Manor, L.: Asymmetric loss for multi-label classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 82-91 (2021) +42. Rizve, M.N., Duarte, K., Rawat, Y.S., Shah, M.: In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329 (2021) +43. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems 29 (2016) +44. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) +45. Shi, W., Gong, Y., Ding, C., Tao, Z.M., Zheng, N.: Transductive semi-supervised deep learning using min-max features. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 299-315 (2018) +46. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems 33, 596-608 (2020) +47. Sun, R., Mai, H., Zhang, T., Wu, F.: DAW: Exploring the better weighting function for semi-supervised semantic segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=KR1G7NJUCD + +48. Tsoumakas, G., Katakis, I.: Multi-label classification. Data Warehousing and Mining: Concepts, Methodologies, Tools, and Applications: Concepts, Methodologies, Tools, and Applications 3, 64 (2008) +49. Verma, V., Kawaguchi, K., Lamb, A., Kannala, J., Bengio, Y., Lopez-Paz, D.: Interpolation consistency training for semi-supervised learning. arXiv preprint arXiv:1903.03825 (2019) +50. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4248-4257 (2022) +51. Wu, T., Huang, Q., Liu, Z., Wang, Y., Lin, D.: Distribution-balanced loss for multi-label classification in long-tailed datasets. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 162-178. Springer (2020) +52. Xu, H.M., Liu, L., Bian, Q., Yang, Z.: Semi-supervised semantic segmentation with prototype-based consistency regularization. arXiv preprint arXiv:2210.04388 (2022) +53. Xu, J., Le, H.: Generating representative samples for few-shot classification. In: CVPR (2022) +54. Xu, J., Le, H., Huang, M., Athar, S., Samaras, D.: Variational feature disentangling for fine-grained few-shot classification. In: ICCV (2021) +55. Xu, J., Le, H., Samaras, D.: Generating features with increased crop-related diversity for few-shot object detection. In: CVPR (2023) +56. Xu, J., Le, H.M., Nguyen, V., Ranjan, V., Samaras, D.: Zero-shot object counting. CVPR pp. 15548-15557 (2023) +57. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023) +58. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4268-4277 (2022) +59. Yarowsky, D.: Unsupervised word sense disambiguation rivaling supervised methods. In: 33rd annual meeting of the association for computational linguistics. pp. 189-196 (1995) +60. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 605–613. Springer (2019) +61. Yu, Y., Pedrycz, W., Miao, D.: Multi-label classification by exploiting label correlations. Expert Systems with Applications 41(6), 2989-3004 (2014) +62. Yuan, J., Liu, Y., Shen, C., Wang, Z., Li, H.: A simple baseline for semi-supervised semantic segmentation with strong data augmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8229-8238 (2021) +63. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) +64. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) +65. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: Pro + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11350-11359 (2023) +66. Zhong, Y., Yuan, B., Wu, H., Yuan, Z., Peng, J., Wang, Y.X.: Pixel contrastive-consistent semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7273-7282 (2021) +67. Zhou, T., Wang, S., Bilmes, J.: Time-consistent self-supervision for semi-supervised learning. In: International Conference on Machine Learning. pp. 11523-11533. PMLR (2020) +68. Zhou, Y., Sun, X., Zha, Z.J., Zeng, W.: Context-reinforced semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4046-4055 (2019) +69. Zou, Y., Zhang, Z., Zhang, H., Li, C.L., Bian, X., Huang, J.B., Pfister, T.: Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713 (2020) +70. Zuo, S., Yu, Y., Liang, C., Jiang, H., Er, S., Zhang, C., Zhao, T., Zha, H.: Self-training with differentiable teacher. arXiv preprint arXiv:2109.07049 (2021) \ No newline at end of file diff --git a/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/images.zip b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6a74d7ca2e9f2c746c835223cb3240948b4fd6c3 --- /dev/null +++ b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:932d10a8db3a636c9fda605c838d9fc2117d8a00e79e93040d9f8694f9c2d687 +size 675416 diff --git a/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/layout.json b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..08ab0f05e57bd5f7f1b74ff3a2db7172ece0910b --- /dev/null +++ b/2024/Beyond Pixels_ Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier/layout.json @@ -0,0 +1,10965 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 163 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 163 + ], + "type": "text", + "content": "Beyond Pixels: Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 160, + 185, + 452, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 185, + 452, + 199 + ], + "spans": [ + { + "bbox": [ + 160, + 185, + 452, + 199 + ], + "type": "text", + "content": "Prantik Howlader1, Srijan Das2, Hieu Le3, and Dimitris Samaras1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 207, + 419, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 207, + 419, + 242 + ], + "spans": [ + { + "bbox": [ + 194, + 207, + 419, + 242 + ], + "type": "text", + "content": "1 Stony Brook University, New York, USA \n2 University of North Carolina at Charlotte, NC, USA \n3 EPFL, Lausanne, Switzerland" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 270, + 455, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 270, + 455, + 479 + ], + "spans": [ + { + "bbox": [ + 160, + 270, + 455, + 479 + ], + "type": "text", + "content": "Abstract. Incorporating pixel contextual information is critical for accurate segmentation. In this paper, we show that an effective way to incorporate contextual information is through a patch-based classifier. This patch classifier is trained to identify classes present within an image region, which facilitates the elimination of distractors and enhances the classification of small object segments. Specifically, we introduce Multi-scale Patch-based Multi-label Classifier (MPMC), a novel plug-in module designed for existing semi-supervised segmentation (SSS) frameworks. MPMC offers patch-level supervision, enabling the discrimination of pixel regions of different classes within a patch. Furthermore, MPMC learns an adaptive pseudo-label weight, using patch-level classification to alleviate the impact of the teacher's noisy pseudo-label supervision on the student. This lightweight module can be integrated into any SSS framework, significantly enhancing their performance. We demonstrate the efficacy of our proposed MPMC by integrating it into four SSS methodologies and improving them across two natural image and one medical segmentation dataset, notably improving the segmentation results of the baselines across all the three datasets. Code is available at: https://github.com/cvlab-stonybrook/Beyond-Pixels" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 489, + 385, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 489, + 385, + 502 + ], + "spans": [ + { + "bbox": [ + 160, + 489, + 385, + 502 + ], + "type": "text", + "content": "Keywords: Segmentation " + }, + { + "bbox": [ + 160, + 489, + 385, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 489, + 385, + 502 + ], + "type": "text", + "content": " Semi-supervised " + }, + { + "bbox": [ + 160, + 489, + 385, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 489, + 385, + 502 + ], + "type": "text", + "content": " Patches" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 605 + ], + "type": "text", + "content": "Supervised segmentation requires a large amount of pixel level annotations, which is laborious and time-consuming [53-56]. This paper focuses on semi-supervised segmentation (SSS) [1, 20], which alleviates the reliance on labeled annotations by leveraging a large quantity of unlabeled images, accompanied by a limited number of manually labeled images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "The fundamental task in segmentation is a per-pixel classification. Thus contextual information of a pixel is important to improve its performance [31,32,68]. Context becomes more critical in SSS settings, which uses limited labeled data making the models more prone to overfitting. We note that recent approaches in SSS methods which predominantly leverage consistency regularization with" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 463, + 265 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 463, + 265 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 463, + 265 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 463, + 265 + ], + "type": "image", + "image_path": "295d3fe0ebeeedd571d70a575cb4f782811626d04b8e4161aefeb1890fa9e0be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 352 + ], + "type": "text", + "content": "Fig.1: Segmentations and Multi-label classifications of image patches (Cityscapes) After 20 epochs, (b) and (c) are the segmentations and classes in the segmentations respectively, (d) is classes predicted by multi-label classifier in MPMC. (f) and (g) are segmentations by Baseline (AugSeg [65]) and Baseline + MPMC at the end of training. White boxes represent regions where MPMC improves the baseline, while orange boxes represent regions where segmentation network predicts classes not present in the patch." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 385, + 480, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 480, + 410 + ], + "type": "text", + "content": "pseudo-labeling techniques [7,14,20,22,36,38,50,57,69] do not exploit contextual information explicitly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "type": "text", + "content": "We hypothesize that patches are the simplest way to determine the context around pixels. The key observation is that the contextual information of an image patch are the classes present in it. For example, in the image patch in the first row of Fig. 1 (a) the classes present are car, road, vegetation and terrain. Thus, training a network to not only classify each pixel in an image but also predict the classes in the patch surrounding it, is a straight forward way to introduce contextual information in SSS pipeline. Moreover, utilizing multi-scale patch-level supervision can effectively encapsulate context at varying scales and distinguish pixels across different classes. To this end, we introduce a novel plugin module, termed as Multi-scale Patch-based Multilabel Classifier (MPMC). This module achieves enhanced discrimination of pixel regions between classes through additional multi-scale patch-level supervision. Notably, we observe that when training our baseline SSS model together with MPMC in limited labeled data scenario, the patch classifier exhibits greater resilience to noise than the segmentation network as illustrated in Fig. 1 (b) and (c). A key observation is when MPMC and baseline SSS model are trained together, the patch classifier provides patch-level labels, identifying the classes present within a region. This facilitates the elimination of distractors (first row in Fig. 1 (b), (c) and (d), while the segmentation network classifies pixels as truck which is not present in the patch, the multi-label classifier does not predict it) and enhances the classification of small object segments by the baseline method." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "P. Howlader et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "Another important issue plaguing SSS methods is noisy pseudo-labels generated by the teacher for unlabeled images to supervise the student. This can lead to the student model overfitting to these noisy labels, consequently impairing its performance. Recent approaches have attempted to mitigate the impact of noise in pseudo-labels through various strategies, such as implementing a hard threshold [14, 23, 38], applying dynamic re-weighting to enhance the influence of reliable pixels [20], and employing negative data mining by archiving unreliable pixels in a memory bank [50]. In contrast to these approaches, we use the MPMC predictions to learn an adaptive weight that reduces the weight of noisy pseudo-labels in the teacher-student framework. Our approach hinges on the underlying observation that MPMC exhibits greater uncertainty (lower confidence) for noisy labels in a patch (see supplementary for details)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 379 + ], + "type": "text", + "content": "In practice, MPMC is a parameterized module that can be plugged in the initial layers of the visual model, where it leverages the low-level pixel information to classify the classes present in each patch. This patch-level classification strategy forces the visual model to learn and preserve discriminative representations of classes, which are often prone to dilution in the deeper layers of the model. We demonstrate the efficacy of our proposed parameterized module, MPMC by integrating it into four SSS methodologies—UniMatch [57], AugSeg [65], AEL [20], and U2PL [50] - notably improving performance across all data partitions in two natural (Cityscapes and Pascal VOC) and one medical (ACDC) dataset. To summarize, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 385, + 480, + 503 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 385, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 385, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 136, + 385, + 480, + 420 + ], + "type": "text", + "content": "1. We introduce a novel plug-in module, called MPMC that incorporates an auxiliary multi-scale multi-label classification task, enhancing pixel region discrimination across different classes via additional patch-level supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 421, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 421, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 136, + 421, + 480, + 443 + ], + "type": "text", + "content": "2. MPMC consists of two adaptive learning weights that reduces the influence of noisy pseudo-labels from the teacher model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 445, + 480, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 445, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 135, + 445, + 480, + 503 + ], + "type": "text", + "content": "3. We demonstrate MPMC's robustness by integrating in four SSS techniques, achieving consistent improvements across all protocols on Pascal VOC and Cityscapes. Notably, our method also shows improvement in medical image dataset ACDC, which indicates that our method is applicable beyond natural images." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 521, + 237, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 237, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 237, + 534 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 604 + ], + "type": "text", + "content": "Semi-supervised learning (SSL) has been extensively researched in recent years. Popular SSL approaches include consistency regularization [2,10,11,14,38,43,58], entropy minimization [5,16], and pseudo-labeling [24-28, 46]. In this work, our emphasis is placed on the domains of pseudo-labeling and consistency regularization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Pseudo labeling: Pseudo-labeling methods [28,45] alongside self-training techniques [35, 59] are designed to leverage labeled image data to produce pseudolabels for unlabeled data. An inherent problem that plagues pseudo-labeling is noise, thus most approaches [42,44,46,69] devise strategies to refine the pseudolabels. PseudoSeg [69], for instance, aims at enhancing pseudo-label accuracy" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Pixels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "using attention mechanisms based on grad-cam [44]. In pseudo-label selection, FixMatch [46] adopts a confidence threshold, whereas UPS [42] relies on uncertainty measures. " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathrm{ST}++" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " [58] focuses on selecting unlabeled images more likely to yield accurate pseudo-pixels. Emerging approaches like U2PL [50] explore the idea of learning reliable pseudo-labels from their less reliable counterparts. CFCG [29] introduced a novel approach that combines cross-fusion and contour guidance for pseudo-label improvement. We differ from these methods by explicitly introducing patch-level contextual information in the SSS methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": "Consistency Regularization: This approach enforces the model to have consistent predictions on different views of the unlabeled images. UniMatch [57] uses dual perturbations and feature perturbations to force the network to learn more consistent information. ICT [49] uses Mixup [64] augmentation for consistency regularization. Recent methods [14, 58, 62] use Cutout [9], Cutmix [63], Class-mix [37] as strong data augmentation. Consistency Regularization is used with pseudo-labeling techniques in Mixmatch [4] and TC-SSL [67]. In our work, we use a set of strong and weak data augmentations [1, 65] for generating pseudo-labels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 416 + ], + "type": "text", + "content": "Multi-label classification: This approach classifies an image into more than one label present in the image. Most works in this field explore finding the correlation between the labels and the images [21,39,61]. Recent works mainly focus on solving data imbalance issues [17,51]. We use multi-label classification to help in segmentation. Some works [15,40] use multi-label classification for segmentation in weakly supervised scenarios, for improved CAM-based pseudo-segmentation. However, we are using multi-label classification in a semi-supervised setting to introduce patch-level contextual information and also use the classifier confidence to reduce the impact of noisy pseudo-labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 430, + 389, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 389, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 389, + 445 + ], + "type": "text", + "content": "3 Semi-Supervised Segmentation Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "type": "text", + "content": "Given a labeled set " + }, + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l = \\{(x^l,y^l)\\}" + }, + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "type": "text", + "content": " and an unlabeled set " + }, + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u = \\{x^u\\}" + }, + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "type": "text", + "content": ", the goal of semi-supervised semantic segmentation is to develop a model that effectively utilizes both labeled and unlabeled data. This section commences with an overview of the foundational framework for semi-supervised segmentation in Section 3.1. followed by an overview of our proposed MPMC in Section 3.2. The section concludes with an elaboration of MPMC's functionality (Section 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 540, + 221, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 221, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 221, + 552 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": "Teacher-Student frameworks leverage weak-to-strong consistency regularization to leverage unlabeled data. As depicted in Fig. 2(a), each unlabeled image " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": " is simultaneously perturbed by two operators, i.e., weak perturbation " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "A^w" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": " (cropping, random flip, etc), and strong perturbation " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "A^s" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": " (color jitter, gaussian blur, contrast, etc). The framework's objective function is a convex combination of supervised loss " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": " for the labeled data and unsupervised loss " + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 130, + 556, + 482, + 640 + ], + "type": "text", + "content": " for the unlabeled data as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 176, + 644, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 644, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 176, + 644, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {l} + \\alpha \\mathcal {L} _ {u} \\quad \\text {w h e r e ,} \\quad \\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum \\mathbb {1} \\left(p ^ {w} \\geq t h\\right) \\mathcal {E} \\left(p ^ {w}, p ^ {s}\\right) \\tag {1}", + "image_path": "fddbfcdec757ec73d9f42e9cb5921a420663673da61aa6b43f683bf44667aa9d.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Howlader et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 302, + 237 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 302, + 237 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 302, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 302, + 237 + ], + "type": "image", + "image_path": "c9964a88ca55309adb44c838c691a9ad905479a6b4211da23387a4297b5e521e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "type": "text", + "content": "Fig. 2: Overall Pipeline of our novel multi-label classification based semantic segmentation: (a) Left: End-to-end Teacher-Student Pipeline with our novel method Multi-scale Patch-based Multi-label Classifier (MPMC). (b) Right: For unlabeled images, The teacher MPMC extracts features from a layer in the segmentation model's encoder to classify the feature's receptive field patch. The confidence of MPMC for a class in a patch is used to calculate two adaptive weights " + }, + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "type": "text", + "content": " which is used to reduce the influence of noisy predictions in that patch from the teacher to train both the student segmentation network and MPMC." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 113, + 482, + 237 + ], + "blocks": [ + { + "bbox": [ + 304, + 113, + 482, + 237 + ], + "lines": [ + { + "bbox": [ + 304, + 113, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 482, + 237 + ], + "type": "image", + "image_path": "33e39ec217d41dfae56c59a380a7ca5e742d5a2dacce2980edd14482500e8734.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l" + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "content": " is the standard cross-entropy loss, comparing the model's predictions with ground truth. While " + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "content": " aims to regularize the predictions for an unlabeled sample so that they remain consistent under both weak and strong augmentations (" + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "inline_equation", + "content": "p^w" + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "inline_equation", + "content": "p^s" + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "content": "). Here, " + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "inline_equation", + "content": "th" + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "content": " is a predefined confidence threshold to filter noisy pseudo labels and " + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 358, + 482, + 418 + ], + "type": "text", + "content": " minimizes the entropy between two probability distributions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 439, + 208, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 208, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 208, + 450 + ], + "type": "text", + "content": "3.2 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "Fig. 2(a) illustrates the comprehensive training process of our proposed MPMC, which is tailored to enhance the discrimination of patches among different classes and to mitigate the impact of noisy predictions transmitted from the teacher to the student. MPMC can be integrated into both teacher and student models within the teacher-student framework. It is composed of a classifier that performs multi-label classifications of the multi-scale patch representations of the input image. The application of MPMC spans two distinct scenarios: 1) For labeled images, MPMC assigns classes to each patch representation based on the classes present within a patch. The multi-scale nature of these patch representations and their classification enhances the segmentor's ability to discriminate classes by capturing enhanced contextual information related to variations in shape and size. 2) In the case of unlabeled images, besides patch classification, MPMC computes a pseudo-label weight map to alleviate the impact of the teacher's noisy pseudo-label supervision on the student. This involves assigning weights to the pseudo-labels within each patch in consistency learning framework, which are determined by the confidence of the MPMC's predictions for each class in the patch, to enhance the reliability and accuracy of the pseudo-labeling process." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Pixels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 440, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 440, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 440, + 129 + ], + "type": "text", + "content": "3.3 Multi-scale Patch-based Multi-label Classifier (MPMC)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "text", + "content": "Our proposed MPMC module is specifically designed to address the shortcomings of SSS pipeline. MPMC can be plugged in at any layer within the teacher-student network, both within the teacher model, denoted as " + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "text", + "content": ", and within the student model, denoted as " + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "text", + "content": ", they share the same parameters. Below, we detail the training of MPMC on labeled images and then on unlabeled images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "text", + "content": "MPMC on Labeled Images. As illustrated in Fig. 2, MPMC incorporates patch-supervision into the segmentation framework, by training a multi-label classifier that takes as input a single pixel in the feature map of the " + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\tau^{th}" + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "text", + "content": " layer of the encoder. This single pixel-representation encodes the information from a patch (receptive field) in the original image. Further, to encapsulate context at varying scales we perform average pooling on this features map with kernels of different sizes before performing multi-label classification. Let us define a patch of image and the corresponding segmentation in a receptive field " + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\{x_r^l,y_r^l\\} \\in \\mathcal{D}^l" + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "inline_equation", + "content": "y_{r}^{l}" + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "text", + "content": " is a multi-label target where the classes present in the receptive field " + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 213, + 482, + 332 + ], + "type": "text", + "content": " are positive classes, and the classes not present are negative classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": "The input to the student MPMC " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " are the features extracted from the " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\tau^{th}" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " layer of the encoder in student segmentation network " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "z_{r} = \\hat{f}_{\\tau}(\\mathcal{A}_w(x_r^l))" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": ". MPMC captures the contextual information within a patch at different scales, " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "i \\in \\{0,1\\}" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " are average pool kernel sizes. For all our experiments " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "19 \\times 19" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": ". These pooling operations are performed with appropriate padding to maintain uniformity in the shape of the output features. MPMC " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " is composed of the concatenation of these multiscale pooled features, followed by a convolutional network " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " and a final linear layer. Thus, the output of the MPMC " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " for the input patch " + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "inline_equation", + "content": "x_{r}^{l}" + }, + { + "bbox": [ + 130, + 333, + 482, + 441 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 220, + 449, + 480, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 449, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 220, + 449, + 480, + 464 + ], + "type": "interline_equation", + "content": "q ^ {l} = \\sigma \\left(W ^ {T} \\mathcal {M} \\left(\\operatorname {c o n c a t} \\left(z _ {r}, z _ {r} ^ {S _ {0}}, z _ {r} ^ {S _ {1}}\\right)\\right) + b\\right) \\tag {2}", + "image_path": "1a9b2afb2a9b47d159a168e637a181d198317ce67745ae527e7c0f41dc7985b5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "content": " are linear layer weights and bias, " + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "content": " is the sigmoid activation and " + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "inline_equation", + "content": "q^l = [q_1^l,\\dots,q_C^l]^T\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "content": " are the predicted probabilities for " + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "content": " classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "text", + "content": "While binary cross-entropy loss is applicable to train MPMC, it indiscriminately weighs all positive and negative classes equally. Considering the receptive field of a feature from the " + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\tau^{th}" + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "text", + "content": " layer of the encoder in the segmentation network " + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "text", + "content": " covers only a small patch relative to the input image's size, the prevalence of negative classes within a patch significantly exceeds that of positive classes. This imbalance influences the binary cross-entropy loss to favor negative classes [51]. Inspired by long-tailed multi-label classification [30, 51] which mitigate such imbalances by diminishing the weight of negative classes, we adapt an asymmetric loss [41], a focal loss variant that assigns distinct " + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "text", + "content": " values to positive and negative classes. For an image patch " + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "inline_equation", + "content": "x_{r}^{l}" + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "text", + "content": " and the predicted class probabilities " + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "inline_equation", + "content": "q^{l}" + }, + { + "bbox": [ + 130, + 496, + 482, + 628 + ], + "type": "text", + "content": " using our framework, the asymmetric focal loss is expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 213, + 635, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 635, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 213, + 635, + 481, + 669 + ], + "type": "interline_equation", + "content": "l _ {r} ^ {M} = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\left\\{ \\begin{array}{l l} (1 - q _ {c} ^ {l}) ^ {\\gamma^ {+}} \\log (q _ {c} ^ {l}), & y _ {r, c} ^ {l} = 1. \\\\ (q _ {c} ^ {l}) ^ {\\gamma^ {-}} \\log (1 - q _ {c} ^ {l}), & y _ {r, c} ^ {l} = 0. \\end{array} \\right. \\tag {3}", + "image_path": "3ab4faf38c1bbfab2154386f631e4d2123ca0d711f71f71e1a390bce552bb30b.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Howlader et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "In our experiments, we set " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\gamma^{+} = 0" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\gamma^{-} = 1" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ". Subsequently, MPMC " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " undergoes training with multi-label supervision " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l^M" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 247, + 152, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 152, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 247, + 152, + 481, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l} ^ {M} = \\frac {1}{| D ^ {l} |} \\sum_ {\\left(x ^ {l}, y ^ {l}\\right) \\in D ^ {l}} \\sum_ {r = 1} ^ {R} l _ {r} ^ {M} \\tag {4}", + "image_path": "ff30540a7ce5b570513bf3f844c16de8795d085e696fe3433c7b5f6f6e0676e2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 197, + 480, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 480, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 197, + 480, + 246 + ], + "type": "inline_equation", + "content": "\\mathrm{R}" + }, + { + "bbox": [ + 130, + 197, + 480, + 246 + ], + "type": "text", + "content": " is the number of receptive fields at " + }, + { + "bbox": [ + 130, + 197, + 480, + 246 + ], + "type": "inline_equation", + "content": "\\tau^{th}" + }, + { + "bbox": [ + 130, + 197, + 480, + 246 + ], + "type": "text", + "content": " layer of the segmentor. This auxiliary task of patch classification augments the network's proficiency in learning and preserving class features, a critical function particularly pertinent as the fidelity of these features tends to diminish in the network's deeper layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "text", + "content": "MPMC on Unlabeled images. In semi-supervised teacher-student framework, for unlabeled images, the prediction of the teacher network " + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "text", + "content": " for the weak augmented view " + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^w(x^u)" + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "text", + "content": " is used to supervise the student network " + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "text", + "content": " for the strong augmented view " + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^s(x^u)" + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "text", + "content": ". In the early iterations of semi-supervised learning, the teacher's predictions are inherently noisy. Common practices [13, 20, 33, 46, 70], involve employing pixel-wise confidence score thresholds (th) to filter out unreliable pseudo labels. However, a high confidence threshold does not entirely eliminate noise due to the teacher's initial miscalibration [42], allowing some noisy predictions to have high confidence. To refine this process, our approach does not solely depend on confidence-thresholded teacher segmentation predictions for reliable pseudo-label extraction. Instead, we leverage the teacher MPMC " + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 266, + 482, + 434 + ], + "type": "text", + "content": "'s confidence for a patch to adjust the weights of the deemed reliable pseudo-labels within that patch. The underlying principle of our methodology is that MPMC exhibits greater uncertainty (lower confidence) for noisy labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": "We employ the teacher MPMC " + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": "'s predictions to define two adaptive learning weights " + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\lambda_{m} \\in \\mathbb{R}^{R \\times C}" + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\lambda_{s} \\in \\mathbb{R}^{W \\times H}" + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": " denotes the number of patches, sized according to the receptive field within the image, " + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": " represents the width and height of the input image, respectively. The learning weight " + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": " is utilized to assign class weights to each patch for training the student MPMC, whereas " + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": " is applied to diminish the impact of noisy predictions during training the student segmentation network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": "For an unlabeled image " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ", consider a patch at the receptive field " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "x_{r}^{u}" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ". The teacher and student segmentation networks yield predictions " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "p^{w} = f(\\mathcal{A}^{w}(x_{r}^{u}))" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "p^{s} = \\hat{f} (\\mathcal{A}^{s}(x_{r}^{u}))" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ", respectively. Simultaneously, the teacher and student MPMC predict " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "q_{r}^{u} = g(f_{\\tau}(\\mathcal{A}^{w}(x_{r}^{u})))" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\hat{q}_r^u = \\hat{g} (\\hat{f}\\tau (\\mathcal{A}^s (x_r^u)))" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": " using features from the " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\tau^{th}" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": " layer of the encoder. Let us define the confidence of teacher MPMC " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": " across all classes in patch " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\gamma = \\mathrm{softmax}(q_r^u)" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ". The weight of the patch " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\lambda_{m,r}" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": " is determined by " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ". For each pixel in patch " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": " associated with class " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": " is assigned the weight " + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\gamma_{c}" + }, + { + "bbox": [ + 130, + 519, + 482, + 640 + ], + "type": "text", + "content": ". This approach assigns weights to pseudo-labels and pixels based on the teacher MPMC's confidence, enhancing the accuracy of the learning process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Thus, the control parameters " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " are integrated to the unsupervised losses to train the student segmentation network and MPMC. The unsupervised" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "text", + "content": "Beyond Pixels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "segmentation loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "L_{u}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " mentioned in equation 1 to train the student segmentation network " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is re-formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 148, + 481, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 148, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 219, + 148, + 481, + 177 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum_ {x ^ {u} \\in \\mathcal {D} ^ {u}} \\mathbb {1} \\left(p ^ {w} \\geq t h\\right) \\lambda_ {s} \\mathcal {E} \\left(p ^ {w}, p ^ {s}\\right) \\tag {5}", + "image_path": "85a6aea3adcc13b74458b6b59786ece3dea10f83855b22e1b4960f2aa7334ffd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "text", + "content": " is the standard cross entropy loss and " + }, + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "inline_equation", + "content": "th" + }, + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "text", + "content": " is the predefined segmentation confidence threshold. Finally, the unsupervised multi-label loss " + }, + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u^M" + }, + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "text", + "content": " to train the student MPMC " + }, + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 130, + 185, + 479, + 222 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 213, + 231, + 481, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 231, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 213, + 231, + 481, + 262 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} ^ {M} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum_ {x ^ {u} \\in D ^ {u}} \\sum_ {r = 1} ^ {R} \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\lambda_ {m, r} ^ {i} \\mathcal {E} ^ {+} \\left(q _ {r} ^ {u}, \\hat {q} _ {r} ^ {u}\\right) \\tag {6}", + "image_path": "9867e98ed10021b17748eaf2e8acdcd0890907dfb6e3fe32c81890e2a8440c9f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^+" + }, + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "text", + "content": " is the binary cross entropy loss. Therefore, the overall loss function is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 308, + 481, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 308, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 246, + 308, + 481, + 323 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {l} + \\mathcal {L} _ {l} ^ {M} + \\alpha \\mathcal {L} _ {u} + \\beta \\mathcal {L} _ {u} ^ {M} \\tag {7}", + "image_path": "4d67cef80f4bfaaa6d3ac980c304746a1e4a81da9588c37a009e457443aeb5dd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "text", + "content": " controls the contribution of " + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u^M" + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "text", + "content": ". In all our experiments " + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 326, + 479, + 351 + ], + "type": "text", + "content": " are set to 0.1 and 0.25 respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 369, + 230, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 369, + 230, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 230, + 383 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 394, + 190, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 190, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 190, + 407 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "type": "text", + "content": "Datasets: PASCAL VOC 2012 [12] is a standard semantic segmentation dataset with 20 semantic classes and a background class. The training, validation, and testing sets consist of 1464, 1449, and 1556 images (Classic), respectively. Following [7, 65], we also include the additional augmented dataset [18] (Blender), which includes a collection of 10582 training images. We adopt the same partition protocols in [7, 65] to evaluate our method in both Classic and Blender sets. Cityscapes [8] dataset is designed for urban scene understanding, consisting of 30 classes, of which only 19 classes are used for scene parsing evaluation. It has 2975, 500, and 1525 images for training, validation, and testing, respectively. ACDC [3] dataset contains 200 annotated short-axis cardiac cineMR images from 100 patients. The segmentation masks consists of classes: left ventricle (LV), myocardium (Myo), and right ventricle (RV). 140 images from 70 patients, 60 images from 30 patients are randomly selected for training and validation respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details: For fair comparison, we adopt DeepLabv3+ [6] as the decoder in all of our experimental setups, and compare with both ResNet-101 and ResNet-50 [19] as the backbone architecture. In all our experiments with Cityscapes, Pascal, and ACDC datasets, the training resolution is 801, 513, and 224 respectively. Note, for maintaining fair comparison, our method \"UniMatch [57] + MPMC\" uses the training resolution of 321 which is used by UniMatch [57] when training on the Classic set of Pascal VOC. Further when training a MPMC" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Howlader et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "integrated baseline, we use the same weak and strong augmentations as used by the corresponding baselines. The convolutional network in MPMC is comprised of three convolution blocks. Each convolution block consists of a stack of two " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " convolutions followed by a " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " convolution. For all our experiments MPMC is integrated to the first layer of the ResNet Encoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 209, + 380, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 209, + 380, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 380, + 220 + ], + "type": "text", + "content": "4.2 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 482, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 315 + ], + "type": "text", + "content": "We conducted our experiments on PASCAL VOC 2012 and Cityscapes which are two popular natural image benchmarks, and a medical dataset ACDC. MPMC is integrated in four semi supervised methods UniMatch [57], AEL [20], U2PL [50] and AugSeg [65]. Note, UniMatch [57] is a consistency regularization based method. We observe that MPMC consistently improves over their baseline performance across all partition protocols on all datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "text", + "content": "Results on PASCAL VOC 2012 Dataset. Table 1 shows the comparison with other SOTA methods for both the Classic and Blender sets. MPMC integrated to the four baseline methods improves them consistently across all data partitions for both Classic and Blender sets. We find that our method brings the most improvement for the least labeled data partition " + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "inline_equation", + "content": "(\\frac{1}{16})" + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "text", + "content": ", improving AugSeg, AEL, U2PL and UniMatch by " + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 316, + 482, + 398 + ], + "type": "text", + "content": " respectively in Classic with a ResNet-101 encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "text", + "content": "Results on Cityscapes Dataset. Table 2 shows the comparison with other SOTA methods. MPMC integrated to the four baseline methods improves them consistently across all data partitions. We observe that similar to results in PASCAL VOC 2012 dataset, our method brings the most improvement for the least labeled data partition " + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "inline_equation", + "content": "(\\frac{1}{16})" + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "text", + "content": ", improving AugSeg, AEL, U2PL and UniMatch by " + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 130, + 399, + 482, + 470 + ], + "type": "text", + "content": " respectively in Classic for ResNet-101 encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 471, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 529 + ], + "type": "text", + "content": "Results on ACDC Dataset. Table 3 shows the comparison with other SOTA methods. We observe that when MPMC is integrated to the four baseline methods, it significantly improves the performance for the \"3 cases\" data partition which has the least labeled images (" + }, + { + "bbox": [ + 130, + 471, + 482, + 529 + ], + "type": "inline_equation", + "content": "< 10\\%" + }, + { + "bbox": [ + 130, + 471, + 482, + 529 + ], + "type": "text", + "content": " of labeled images). Note we use ResNet-101 for all experiments in this dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 531, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 579 + ], + "type": "text", + "content": "The consistent improvement in SSS performance in both natural and medical datasets and its ability to integrate in different SSS methods substantiates the robustness of MPMC. Moreover, the notable improvements achieved in scenarios with limited data shows MPMC's vital efficacy within SSS pipelines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Qualitative Results In Fig. 3 we visualise the results of integration of our method (MPMC) to AugSeg [65]. Introducing patch-level supervision with the integration of MPMC, leads the SSS model to better segment confusing classes. Among comparisons we show that MPMC leads to better segmentation of pixels regions of confusing classes like bus, train and truck. These classes are often misclassified not only because of sharing visual similarity but also having low training instances." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Pixels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 224, + 477, + 596 + ], + "blocks": [ + { + "bbox": [ + 132, + 180, + 481, + 214 + ], + "lines": [ + { + "bbox": [ + 132, + 180, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 481, + 214 + ], + "type": "text", + "content": "Table 1: Quantitative results of different semi-supervised segmentation methods on Pascal VOC classic and blender set. We report Mean IoU under various partition protocols and show the improvements " + }, + { + "bbox": [ + 132, + 180, + 481, + 214 + ], + "type": "inline_equation", + "content": "(\\Delta)" + }, + { + "bbox": [ + 132, + 180, + 481, + 214 + ], + "type": "text", + "content": " over corresponding baseline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 224, + 477, + 596 + ], + "lines": [ + { + "bbox": [ + 136, + 224, + 477, + 596 + ], + "spans": [ + { + "bbox": [ + 136, + 224, + 477, + 596 + ], + "type": "table", + "html": "
MethodClassicBlender
1/161/81/41/2Full1/161/81/4
ResNet-50
Supervised44.052.361.766.772.9---
PC2Seg [66][CVPR'21]56.964.667.670.972.3---
PseudoSeg [69][ICLR'21]56.964.667.670.972.2---
ST++ [58][CVPR'22]-----72.674.475.4
AugSeg [65][CVPR'21]64.272.176.177.478.877.278.278.2
AugSeg [65] + MPMC (Δ)65.973.777.678.579.779.079.879.7
(+1.7)(+1.6)(+1.5)(+1.1)(+0.9)(+1.8)(+1.6)(+1.5)
AEL [20][NeurIPS'21]62.964.170.372.774.074.176.177.9
AEL [20] + MPMC (Δ)65.066.172.174.275.376.178.079.6
(+2.1)(+2.0)(+1.8)(+1.5)(+1.3)(+2.0)(+1.9)(+1.7)
U2PL [50][CVPR'22]63.365.571.673.875.174.777.477.5
U2PL [50] + MPMC (Δ)65.267.473.375.276.276.579.179.1
(+1.9)(+1.9)(+1.7)(+1.4)(+1.1)(+1.8)(+1.7)(+1.6)
UniMatch [57][CVPR'23]71.972.576.077.478.778.179.079.1
UniMatch [57] + MPMC (Δ)73.574.077.478.579.479.680.480.4
(+1.6)(+1.5)(+1.4)(+1.1)(+0.7)(+1.5)(+1.4)(+1.3)
ResNet-101
Supervised45.155.364.869.773.570.675.076.5
CPS [7][CVPR'21]64.167.471.775.9-72.275.877.6
PS-MT [33][CVPR'22]65.869.676.678.480.075.578.278.7
PCR [52][NeurIPS'22]70.174.777.278.580.778.680.780.8
DAW [47][NeurIPS'23]74.877.479.580.681.578.578.979.6
CFCG [29][ICCV'23]-----77.479.480.4
AugSeg [65][CVPR'21]71.175.578.880.381.479.381.580.5
AugSeg [65] + MPMC (Δ)72.977.879.981.382.180.882.581.4
(+1.8)(+2.3)(+1.1)(+1.0)(+0.7)(+1.5)(+1.0)(+0.9)
AEL [20][NeurIPS'21]66.168.371.974.478.977.277.678.1
AEL [20] + MPMC (Δ)68.670.373.576.180.179.279.379.2
(+2.5)(+2.0)(+1.6)(+1.7)(+1.2)(+2.0)(+1.7)(+1.1)
U2PL [50][CVPR'22]68.069.273.776.279.577.279.079.3
U2PL [50] + MPMC (Δ)70.070.875.377.880.579.280.780.3
(+2.0)(+1.6)(+1.6)(+1.6)(+1.0)(+2.0)(+1.7)(+1.0)
UniMatch [57][CVPR'23]75.277.278.879.981.280.981.980.4
UniMatch [57] + MPMC (Δ)77.378.679.880.881.782.583.181.5
(+2.1)(+1.4)(+1.0)(+0.9)(+0.5)(+1.6)(+1.2)(+1.1)
", + "image_path": "f047d3a1dc07cc9d0bb857ea05c9dbd360be1f489e7eb1e6f4446e343ee70969.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Howlader et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 125, + 220, + 175 + ], + "blocks": [ + { + "bbox": [ + 134, + 125, + 220, + 175 + ], + "lines": [ + { + "bbox": [ + 134, + 125, + 220, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 125, + 220, + 175 + ], + "type": "image", + "image_path": "bed05dbb3298e38f171f9ff4cd49a43ee97729d555d8f7cf552826fe975d7199.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 176, + 219, + 225 + ], + "blocks": [ + { + "bbox": [ + 134, + 176, + 219, + 225 + ], + "lines": [ + { + "bbox": [ + 134, + 176, + 219, + 225 + ], + "spans": [ + { + "bbox": [ + 134, + 176, + 219, + 225 + ], + "type": "image", + "image_path": "60cf4c0ef9b134097920c04cfa6f34a4885eec20acdae5a67d979188cf0351b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 226, + 219, + 275 + ], + "blocks": [ + { + "bbox": [ + 134, + 226, + 219, + 275 + ], + "lines": [ + { + "bbox": [ + 134, + 226, + 219, + 275 + ], + "spans": [ + { + "bbox": [ + 134, + 226, + 219, + 275 + ], + "type": "image", + "image_path": "cc9a1e4163be66a3ce0f31047a6e3f451313d4f6fc5a954160c408fc210f94cc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 276, + 220, + 325 + ], + "blocks": [ + { + "bbox": [ + 134, + 276, + 220, + 325 + ], + "lines": [ + { + "bbox": [ + 134, + 276, + 220, + 325 + ], + "spans": [ + { + "bbox": [ + 134, + 276, + 220, + 325 + ], + "type": "image", + "image_path": "adbb23c58d04ed1cfbad440b34db33fc59801a9e84ccfb40a6d51ab6fbd66619.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 327, + 194, + 334 + ], + "lines": [ + { + "bbox": [ + 160, + 327, + 194, + 334 + ], + "spans": [ + { + "bbox": [ + 160, + 327, + 194, + 334 + ], + "type": "text", + "content": "(a) Input Image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 221, + 125, + 306, + 175 + ], + "blocks": [ + { + "bbox": [ + 221, + 125, + 306, + 175 + ], + "lines": [ + { + "bbox": [ + 221, + 125, + 306, + 175 + ], + "spans": [ + { + "bbox": [ + 221, + 125, + 306, + 175 + ], + "type": "image", + "image_path": "2e3599f5d7734cff4daa0062343dadfdcc3d7c89ace091c8cd097818991febef.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 221, + 176, + 306, + 225 + ], + "blocks": [ + { + "bbox": [ + 221, + 176, + 306, + 225 + ], + "lines": [ + { + "bbox": [ + 221, + 176, + 306, + 225 + ], + "spans": [ + { + "bbox": [ + 221, + 176, + 306, + 225 + ], + "type": "image", + "image_path": "052331a1214c8da78af967a541f7a03ecf928ae1722e41867d3b71024951022d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 221, + 226, + 306, + 275 + ], + "blocks": [ + { + "bbox": [ + 221, + 226, + 306, + 275 + ], + "lines": [ + { + "bbox": [ + 221, + 226, + 306, + 275 + ], + "spans": [ + { + "bbox": [ + 221, + 226, + 306, + 275 + ], + "type": "image", + "image_path": "7906a90b19af6415d8d8372f9ffe0bbe48e598f9eeecc46b57c34b3cef36dd07.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 221, + 276, + 306, + 325 + ], + "blocks": [ + { + "bbox": [ + 221, + 276, + 306, + 325 + ], + "lines": [ + { + "bbox": [ + 221, + 276, + 306, + 325 + ], + "spans": [ + { + "bbox": [ + 221, + 276, + 306, + 325 + ], + "type": "image", + "image_path": "35a09bf93191eae0bf507e428bab2f6620770ed3d7c3e28256b377da6b53d469.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 327, + 282, + 333 + ], + "lines": [ + { + "bbox": [ + 246, + 327, + 282, + 333 + ], + "spans": [ + { + "bbox": [ + 246, + 327, + 282, + 333 + ], + "type": "text", + "content": "(b) Ground Truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 125, + 392, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 125, + 392, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 125, + 392, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 125, + 392, + 175 + ], + "type": "image", + "image_path": "844d0b6f789c3f408683c58dc93226c9aeea6adac0bede4e7f11677457667c15.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 176, + 392, + 225 + ], + "blocks": [ + { + "bbox": [ + 307, + 176, + 392, + 225 + ], + "lines": [ + { + "bbox": [ + 307, + 176, + 392, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 392, + 225 + ], + "type": "image", + "image_path": "1d6be199f70055a0cc81f08134cf7299523a01fad44d2bd9058e05a82ef528e5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 226, + 392, + 275 + ], + "blocks": [ + { + "bbox": [ + 307, + 226, + 392, + 275 + ], + "lines": [ + { + "bbox": [ + 307, + 226, + 392, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 226, + 392, + 275 + ], + "type": "image", + "image_path": "7501babaf86065a20750c0c62afbb5223dd0b780095956b969ce1a55f698f626.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 276, + 392, + 325 + ], + "blocks": [ + { + "bbox": [ + 307, + 276, + 392, + 325 + ], + "lines": [ + { + "bbox": [ + 307, + 276, + 392, + 325 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 392, + 325 + ], + "type": "image", + "image_path": "1d8afd4a8d9642127033b56d7fae319d346e519153a63c78e08fa67785895d1d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 327, + 358, + 334 + ], + "lines": [ + { + "bbox": [ + 334, + 327, + 358, + 334 + ], + "spans": [ + { + "bbox": [ + 334, + 327, + 358, + 334 + ], + "type": "text", + "content": "(c) AugSeg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 343, + 482, + 388 + ], + "lines": [ + { + "bbox": [ + 131, + 343, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 343, + 482, + 388 + ], + "type": "text", + "content": "Fig. 3: Qualitative Results on Cityscapes dataset: (a) is the original image, (b) is the Ground Truth, (c) are segmentations generated by AugSeg [65] compared to (d) which are segmentations generated when our method (MPMC) is integrated to AugSeg. White boxes show the areas where our method improves the baseline." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 393, + 125, + 480, + 175 + ], + "blocks": [ + { + "bbox": [ + 393, + 125, + 480, + 175 + ], + "lines": [ + { + "bbox": [ + 393, + 125, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 393, + 125, + 480, + 175 + ], + "type": "image", + "image_path": "719b8463a09340c28f2f15e8434a4c25e60bca204b0dc80adf13545bddb8f90b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 393, + 176, + 479, + 225 + ], + "blocks": [ + { + "bbox": [ + 393, + 176, + 479, + 225 + ], + "lines": [ + { + "bbox": [ + 393, + 176, + 479, + 225 + ], + "spans": [ + { + "bbox": [ + 393, + 176, + 479, + 225 + ], + "type": "image", + "image_path": "e35b9b096407e70c93ec3557f151f3d1340a987b148d6e5004c0a5bfaa5510cd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 393, + 226, + 479, + 275 + ], + "blocks": [ + { + "bbox": [ + 393, + 226, + 479, + 275 + ], + "lines": [ + { + "bbox": [ + 393, + 226, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 393, + 226, + 479, + 275 + ], + "type": "image", + "image_path": "d6587c8537d1345a9f4b1b6b452e85c51b7f9f49b60bfaf146375f59bd05a29a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 393, + 276, + 479, + 325 + ], + "blocks": [ + { + "bbox": [ + 393, + 276, + 479, + 325 + ], + "lines": [ + { + "bbox": [ + 393, + 276, + 479, + 325 + ], + "spans": [ + { + "bbox": [ + 393, + 276, + 479, + 325 + ], + "type": "image", + "image_path": "04a39c479d40f504f57cc89b08e1f8694e445c62790a1f6a13c130c09c57b9f2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 327, + 458, + 334 + ], + "lines": [ + { + "bbox": [ + 415, + 327, + 458, + 334 + ], + "spans": [ + { + "bbox": [ + 415, + 327, + 458, + 334 + ], + "type": "text", + "content": "(d) AugSeg + MPMC" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 132, + 460, + 482, + 651 + ], + "blocks": [ + { + "bbox": [ + 131, + 417, + 482, + 451 + ], + "lines": [ + { + "bbox": [ + 131, + 417, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 482, + 451 + ], + "type": "text", + "content": "Table 2: Quantitative results of different semi-supervised segmentation methods on Cityscapes validation set. We report Mean IoU under various partition protocols and show the improvements " + }, + { + "bbox": [ + 131, + 417, + 482, + 451 + ], + "type": "inline_equation", + "content": "(\\Delta)" + }, + { + "bbox": [ + 131, + 417, + 482, + 451 + ], + "type": "text", + "content": " over the corresponding baseline." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 460, + 482, + 651 + ], + "lines": [ + { + "bbox": [ + 132, + 460, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 482, + 651 + ], + "type": "table", + "html": "
MethodResNet-50ResNet-101
1/161/81/41/21/161/81/41/2
Supervised63.3468.7374.1476.6266.372.875.078.0
CPS [7][CVPR'21]69.7974.3976.8578.6469.874.374.676.8
PS-MT [33][CVPR'22]-75.7676.9277.64-76.977.679.1
PCR [52][NeurIPS'22]----73.476.378.479.1
CFCG [29][ICCV'23]76.178.979.380.177.879.680.480.9
AugSeg [65][CVPR'21]73.776.478.779.375.277.879.680.4
AugSeg [65] + MPMC (Δ)75.3 (+1.6)77.7 (+1.3)79.8 (+1.1)80.3 (+1.0)76.9 (+1.7)79.2 (+1.4)80.9 (+1.3)81.3 (+0.9)
AEL [20][NeurIPS'21]68.272.774.977.574.575.677.579.0
AEL [20] + MPMC (Δ)70.5 (+2.3)74.5 (+1.8)76.6 (+1.7)79.0 (+1.5)76.6 (+2.1)77.5 (+1.9)78.9 (+1.4)80.3 (+1.3)
U2PL [50][CVPR'22]69.073.076.378.674.976.578.579.1
U2PL [50] + MPMC (Δ)71.2 (+2.2)74.6 (+1.6)77.8 (+1.5)79.8 (+1.2)77.2 (+2.3)78.2 (+1.7)80.0 (+1.5)80.0 (+0.9)
UniMatch [57][CVPR'23]75.076.877.578.676.677.979.279.5
UniMatch [57] + MPMC (Δ)76.8 (+1.8)78.0 (+1.2)78.6 (+1.1)79.5 (+0.9)78.5 (+1.9)79.2 (+1.3)80.9 (+1.7)81.3 (+1.8)
", + "image_path": "3c76d1a645d9df94852122794575bb28d4e8e96d3d521e7299874588907655c7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Pixels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 219, + 158, + 392, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Quantitative results of different semi-supervised segmentation methods on ACDC dataset with 3/7 labeled cases. We report Dice Similarity Coefficient (DSC) metric averaged over 3 classes. We use ResNet-101 for all experiments in this dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 158, + 392, + 308 + ], + "lines": [ + { + "bbox": [ + 219, + 158, + 392, + 308 + ], + "spans": [ + { + "bbox": [ + 219, + 158, + 392, + 308 + ], + "type": "table", + "html": "
Method3 cases7 cases
Supervised41.562.5
UA-MT [60][MICCAI'19]61.081.5
CPS [7][CVPR'21]60.383.3
CNN & Trans [34][PMLR'22]65.686.4
AugSeg [65][CVPR'21]85.187.3
AugSeg [65] + MPMC (Δ)88.6 (+3.5)88.4 (+1.1)
AEL [20][NeurIPS'21]82.787.6
AEL [20] + MPMC (Δ)85.4 (+2.7)88.5 (+0.9)
U2PL [50][CVPR'22]84.289.2
U2PL [50] + MPMC (Δ)87.4 (+3.2)90.5 (+1.3)
UniMatch [57][CVPR'23]88.989.9
UniMatch [57] + MPMC (Δ)91.8 (+2.9)90.4 (+0.5)
", + "image_path": "41447403b29a4af0acabf42a67726f8f0f15200c43eabd017ddc1921c8095e22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 332, + 246, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 246, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 246, + 342 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 390 + ], + "type": "text", + "content": "We conduct experiments to study the impact of various components of our approach. All experiments in this section were conducted on " + }, + { + "bbox": [ + 130, + 354, + 481, + 390 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 130, + 354, + 481, + 390 + ], + "type": "text", + "content": " data partition of the Cityscapes dataset, and use AugSeg [65] as the baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": "Segmentation performance at different scales. Here we analyse how scaling of input features for patch-classification, improves the performance of the SSS method. We consider two scales - average pool kernels " + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "inline_equation", + "content": "19 \\times 19" + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": ". Fig. 4 illustrates for two random classes pole and rider, that average pooling with a kernel size of " + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": " is better at enhancing the pixel accuracy for instances with smaller areas, whereas the " + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "inline_equation", + "content": "19 \\times 19" + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": " kernel size is better at improving accuracy for instances with larger areas. This can be attributed to the larger receptive field of the " + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "inline_equation", + "content": "19 \\times 19" + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": " kernel, enabling more effective contextual learning for larger regions. While the smaller receptive field of " + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": " kernel is better able to discern smaller objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 605 + ], + "type": "text", + "content": "Analysing class features The crux of our method is to better discriminate pixel regions between different classes. Here we analyze if MPMC integrated to AugSeg [65] can better discriminate features than the baseline for different classes in Cityscapes Dataset. We use the pixel features from the penultimate layer of the segmentation network to analyse via t-SNE. Fig 5 illustrates for four random classes - car, bus, train and truck, integration of MPMC leads to better discrimination between the classes, i.e., compared to AugSeg [65], the features of train and truck are more clustered and separated from the rest." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Influence of MPMC on patch-level segmentation. Our method MPMC improves segmentation by introducing patch classification as an auxiliary task. We analyse at patch level the correlation between multi-label accuracy and the Mean IoU performance of the segmentation network. We use (1 - hamming loss) [48] as a measure of multi-label classification and use the basic teacher-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Howlader et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 113, + 305, + 260 + ], + "blocks": [ + { + "bbox": [ + 164, + 113, + 305, + 260 + ], + "lines": [ + { + "bbox": [ + 164, + 113, + 305, + 260 + ], + "spans": [ + { + "bbox": [ + 164, + 113, + 305, + 260 + ], + "type": "image", + "image_path": "5d3bccfe2eb4b7b2bad596ca3c896c83c810ee05e7923e4da7d8335435c1f3f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 268, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 301 + ], + "type": "text", + "content": "Fig. 4: Pixel accuracy across different scales with respect to instance size. For two different scales, we compare the pixel accuracy with respect to size of instances for two random classes: pole and rider." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 113, + 450, + 260 + ], + "blocks": [ + { + "bbox": [ + 307, + 113, + 450, + 260 + ], + "lines": [ + { + "bbox": [ + 307, + 113, + 450, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 450, + 260 + ], + "type": "image", + "image_path": "200a277845cb106a9a3c7de7a37c13b633c4a9c85a68fb5915484fc33509372f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 170, + 312, + 443, + 441 + ], + "blocks": [ + { + "bbox": [ + 170, + 312, + 443, + 441 + ], + "lines": [ + { + "bbox": [ + 170, + 312, + 443, + 441 + ], + "spans": [ + { + "bbox": [ + 170, + 312, + 443, + 441 + ], + "type": "image", + "image_path": "dbc537f6b47ae7aab8d16e8d61b3f3a63b42315d12ae7d28b493dca4d52bb9da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 449, + 480, + 472 + ], + "lines": [ + { + "bbox": [ + 131, + 449, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 480, + 472 + ], + "type": "text", + "content": "Fig. 5: Analysis of class features (a) t-SNE of features predicted by AugSeg [65], (b) t-SNE of features predicted by AugSeg [65] + MPMC." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": "student framework mentioned in Section 3.1 as our baseline. Table 4 illustrates that the avg MIoU improvement is highest (4.05%) in patches having more than 90% multi-label classification accuracy showing that MPMC improves segmentation at the patch level." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 480, + 567 + ], + "type": "text", + "content": "Impact of MPMC on semantic segmentation in limited labeled data scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "In Table 5 shows the impact of MPMC on segmentation when trained with only labeled data. We use the basic teacher-student framework mentioned in Section 3.1 as our baseline, which achieves MIoU of 66.3, 72.8 and 75.0 in 186, 372 and 744 partitions of labeled images respectively. After introducing PMC, there is an improvement of " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " in 186, 372 and 744 partitions of labeled images respectively. We observe that MPMC improves segmentation performance in limited-labeled data scenarios, with the improvement highest for least labeled data, i.e., " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " for 186 labeled images." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Pixels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 136, + 481, + 189 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 463, + 126 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 463, + 126 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 463, + 126 + ], + "type": "text", + "content": "Table 4: Analysis of the influence of MPMC on segmentation at patch-level." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 136, + 481, + 189 + ], + "lines": [ + { + "bbox": [ + 132, + 136, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 481, + 189 + ], + "type": "table", + "html": "
1 - hamming loss90+80-9070-8060-7050-6040-5030-4020-3010-200-10
% patches12.625.220.58.45.15.24.110.22.46.3
Avg M IoU Δ+4.05+1.57+0.88+1.55+0.91+1.15+0.49+0.10-1.90-2.20
", + "image_path": "03c0b0147e3f9971f4090d00062a8d3197739580c357b4a9165018de84e3c09b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 149, + 264, + 272, + 323 + ], + "blocks": [ + { + "bbox": [ + 132, + 199, + 291, + 255 + ], + "lines": [ + { + "bbox": [ + 132, + 199, + 291, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 199, + 291, + 255 + ], + "type": "text", + "content": "Table 5: Impact of MPMC on segmentation (MIoU) when trained with different volumes of labeled data. MPMC shows the highest improvement " + }, + { + "bbox": [ + 132, + 199, + 291, + 255 + ], + "type": "inline_equation", + "content": "(1.9\\%)" + }, + { + "bbox": [ + 132, + 199, + 291, + 255 + ], + "type": "text", + "content": " in the least labeled data partition." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 264, + 272, + 323 + ], + "lines": [ + { + "bbox": [ + 149, + 264, + 272, + 323 + ], + "spans": [ + { + "bbox": [ + 149, + 264, + 272, + 323 + ], + "type": "table", + "html": "
# Labeled Data
186372744
w/o MPMC66.372.875.0
w/ MPMC68.274.376.3
Gain/Δ+1.9+1.5+1.3
", + "image_path": "561bb522d34048178f7d43415401d343e334b051c73ebedbb9e6949c055c2321.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 254, + 471, + 324 + ], + "blocks": [ + { + "bbox": [ + 304, + 199, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 304, + 199, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 482, + 243 + ], + "type": "text", + "content": "Table 6: Ablation study of different component: MPMC on labeled images (MPMCL) and MPMC on unlabeled images (MPMCUL)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 254, + 471, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 254, + 471, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 471, + 324 + ], + "type": "table", + "html": "
MPMCLMPMCUL1/16 (186)1/8 (372)1/4 (744)
68.473.574.2
✓70.875.475.7
✓71.376.076.6
✓✓73.176.777.3
", + "image_path": "b943ae342a5d21d728a5890672024142f63dd68dedd8b6e17182163c7ac150e4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": "Analysis of different components We ablate each component of MPMC with basic teacher-student framework (Section 3.1) as the baseline. MPMC on labeled images (MPMCL), which introduces multi-scale patch-based classification on only the labeled data, achieves an improvement of " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": " partition protocols respectively. MPMC on unlabeled images (MPMCUL) leads to a further improvement of " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 132, + 346, + 482, + 441 + ], + "type": "text", + "content": " partition protocols respectively. Thus, MPMC is effective on both labeled and unlabeled instances in SSS pipelines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 460, + 220, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 220, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 220, + 473 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 641 + ], + "type": "text", + "content": "In this paper, we have proposed a novel Multi-scale Patch-based Multi-label Classifier (MPMC), which introduces patch-level context information in the form of an auxiliary task of multi-scale patch-based multi-label classification. When jointly trained with a SSS model, the patch classifier provides patch-level labels, identifying the classes present within a region. This facilitates the elimination of distractors and enhances the classification of small object segments by the baseline method. Further, the MPMC predictions are used to learn an adaptive weight that reduces the influence of noisy pseudo-labels in the teacher-student framework. We performed a thorough analysis of the importance of multi-scale patch-level contextual information in improving segmentation especially in low data scenarios. Further, we demonstrated the efficacy of our proposed parameterized module MPMC by integrating it into four SSS methodologies and showing consistent improvements in Cityscapes, Pascal VOC and ACDC benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "Acknowledgement. This work was partially supported by the National Science Foundation (IIS-2245652, IIS-2123920, IIS-2212046)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Howlader et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "text", + "content": "1. Alonso, I., Sabater, A., Ferstl, D., Montesano, L., Murillo, A.C.: Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8219-8228 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 187, + 481, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 481, + 209 + ], + "type": "text", + "content": "2. Bachman, P., Alsharif, O., Precup, D.: Learning with pseudo-ensembles. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 253 + ], + "type": "text", + "content": "3. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37(11), 2514-2525 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "text", + "content": "4. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "type": "text", + "content": "5. Chen, H., Jin, Y., Jin, G., Zhu, C., Chen, E.: Semisupervised semantic segmentation by improving prediction confidence. IEEE Transactions on Neural Networks and Learning Systems (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 354 + ], + "type": "text", + "content": "6. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 481, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 387 + ], + "type": "text", + "content": "7. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 388, + 481, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 481, + 431 + ], + "type": "text", + "content": "8. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 432, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 481, + 453 + ], + "type": "text", + "content": "9. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "text", + "content": "0. Durasov, N., Dorndorf, N., Le, H., Fua, P.: Zigzag: Universal sampling-free uncertainty estimation through two-step inference. Transactions on Machine Learning Research (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "text", + "content": "1. Durasov, N., Oner, D., Donier, J., Le, H., Fua, P.: Enabling uncertainty estimation in iterative neural networks. In: Forty-first International Conference on Machine Learning (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "text", + "content": "2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88(2), 303-338 (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "3. Fan, J., Gao, B., Jin, H., Jiang, L.: Ucc: Uncertainty guided cross-head co-training for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9947-9956 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 588, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 620 + ], + "type": "text", + "content": "4. French, G., Laine, S., Aila, T., Mackiewicz, M., Finlayson, G.: Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "5. Ge, W., Yang, S., Yu, Y.: Multi-evidence filtering and fusion for multi-label classification, object detection and semantic segmentation based on weakly supervised learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1277-1286 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "text", + "content": "Beyond Pixels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. Advances in neural information processing systems 17 (2004)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "17. Guo, H., Wang, S.: Long-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15089-15098 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "text", + "content": "18. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "text", + "content": "20. Hu, H., Wei, F., Hu, H., Ye, Q., Cui, J., Wang, L.: Semi-supervised semantic segmentation via adaptive equalization learning. Advances in Neural Information Processing Systems 34, 22106-22118 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "21. Huang, J., Li, G., Huang, Q., Wu, X.: Learning label specific features for multi-label classification. In: 2015 IEEE International Conference on Data Mining. pp. 181-190. IEEE (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "text", + "content": "22. Huo, X., Xie, L., He, J., Yang, Z., Zhou, W., Li, H., Tian, Q.: Atso: Asynchronous teacher-student optimization for semi-supervised image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1235-1244 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "23. Ke, Z., Qiu, D., Li, K., Yan, Q., Lau, R.W.: Guided collaborative training for pixel-wise semi-supervised learning. In: European conference on computer vision. pp. 429-445. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "24. Le, H., Goncalves, B., Samaras, D., Lynch, H.: Weakly labeling the antarctic: The penguin colony case. In: CVPRW (June 2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "25. Le, H., Samaras, D., Lynch, H.J.: A convolutional neural network architecture designed for the automated survey of seabird colonies. Remote Sensing in Ecology and Conservation 8(2), 251-262 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "26. Le, H., Vicente, T.F.Y., Nguyen, V., Hoai, M., Samaras, D.: A+D Net: Training a shadow detector with adversarial shadow attenuation. In: European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "27. Le, H., Yu, C.P., Zelinsky, G., Samaras, D.: Co-localization with category-consistent features and geodesic distance propagation. In: ICCVW (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "28. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896 (2013)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "text", + "content": "29. Li, S., He, Y., Zhang, W., Zhang, W., Tan, X., Han, J., Ding, E., Wang, J.: Ccfg: Semi-supervised semantic segmentation via cross-fusion and contour guidance supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16348-16358 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 632 + ], + "type": "text", + "content": "30. Lin, D.: Probability guided loss for long-tailed multi-label image classification. In: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI'23/IAAI'23/EAAI'23, AAAI Press (2023). https://doi.org/10.1609/aaai.v37i2.25244, https://doi.org/10.1609/aaai.v37i2.25244" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Lin, G., Shen, C., Van Den Hengel, A., Reid, I.: Efficient piecewise training of deep structured models for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3194-3203 (2016)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "P. Howlader et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 664 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "32. Liu, W., Rabinovich, A., Berg, A.C.: Parsenet: Looking wider to see better. arXiv preprint arXiv:1506.04579 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 181 + ], + "type": "text", + "content": "33. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4258-4267 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "type": "text", + "content": "34. Luo, X., Hu, M., Song, T., Wang, G., Zhang, S.: Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In: International Conference on Medical Imaging with Deep Learning. pp. 820-833. PMLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "text", + "content": "35. McClosky, D., Charniak, E., Johnson, M.: Effective self-training for parsing. In: Proceedings of the Human Language Technology Conference of the NAACL, Main Conference. pp. 152-159 (2006)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 278 + ], + "type": "text", + "content": "36. Oliver, A., Odena, A., Raffel, C.A., Cubuk, E.D., Goodfellow, I.: Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "type": "text", + "content": "37. Olsson, V., Tranheden, W., Pinto, J., Svensson, L.: Classmix: Segmentation-based data augmentation for semi-supervised learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1369-1378 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 311, + 482, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 342 + ], + "type": "text", + "content": "38. Ouali, Y., Hudelot, C., Tami, M.: Semi-supervised semantic segmentation with cross-consistency training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12674-12684 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 342, + 482, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 363 + ], + "type": "text", + "content": "39. Read, J., Pfahringer, B., Holmes, G., Frank, E.: Classifier chains for multi-label classification. Machine learning 85, 333-359 (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "text", + "content": "40. Reif, S., Seibold, C., Freytag, A., Rodner, E., Stiefelhagen, R.: Every annotation counts: Multi-label deep supervision for medical image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9532-9542 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 482, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 439 + ], + "type": "text", + "content": "41. Ridnik, T., Ben-Baruch, E., Zamir, N., Noy, A., Friedman, I., Protter, M., Zelnik-Manor, L.: Asymmetric loss for multi-label classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 82-91 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 439, + 482, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 472 + ], + "type": "text", + "content": "42. Rizve, M.N., Duarte, K., Rawat, Y.S., Shah, M.: In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 472, + 482, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 503 + ], + "type": "text", + "content": "43. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 503, + 482, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 546 + ], + "type": "text", + "content": "44. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 546, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 578 + ], + "type": "text", + "content": "45. Shi, W., Gong, Y., Ding, C., Tao, Z.M., Zheng, N.: Transductive semi-supervised deep learning using min-max features. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 299-315 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 578, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 621 + ], + "type": "text", + "content": "46. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems 33, 596-608 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 621, + 482, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 664 + ], + "type": "text", + "content": "47. Sun, R., Mai, H., Zhang, T., Wu, F.: DAW: Exploring the better weighting function for semi-supervised semantic segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=KR1G7NJUCD" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 101 + ], + "type": "text", + "content": "Beyond Pixels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "48. Tsoumakas, G., Katakis, I.: Multi-label classification. Data Warehousing and Mining: Concepts, Methodologies, Tools, and Applications: Concepts, Methodologies, Tools, and Applications 3, 64 (2008)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "49. Verma, V., Kawaguchi, K., Lamb, A., Kannala, J., Bengio, Y., Lopez-Paz, D.: Interpolation consistency training for semi-supervised learning. arXiv preprint arXiv:1903.03825 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "text", + "content": "50. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4248-4257 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 270 + ], + "type": "text", + "content": "51. Wu, T., Huang, Q., Liu, Z., Wang, Y., Lin, D.: Distribution-balanced loss for multi-label classification in long-tailed datasets. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 162-178. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "52. Xu, H.M., Liu, L., Bian, Q., Yang, Z.: Semi-supervised semantic segmentation with prototype-based consistency regularization. arXiv preprint arXiv:2210.04388 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 324 + ], + "type": "text", + "content": "53. Xu, J., Le, H.: Generating representative samples for few-shot classification. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "54. Xu, J., Le, H., Huang, M., Athar, S., Samaras, D.: Variational feature disentangling for fine-grained few-shot classification. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "55. Xu, J., Le, H., Samaras, D.: Generating features with increased crop-related diversity for few-shot object detection. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "56. Xu, J., Le, H.M., Nguyen, V., Ranjan, V., Samaras, D.: Zero-shot object counting. CVPR pp. 15548-15557 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "57. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "58. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4268-4277 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "59. Yarowsky, D.: Unsupervised word sense disambiguation rivaling supervised methods. In: 33rd annual meeting of the association for computational linguistics. pp. 189-196 (1995)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 533 + ], + "type": "text", + "content": "60. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 605–613. Springer (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "61. Yu, Y., Pedrycz, W., Miao, D.: Multi-label classification by exploiting label correlations. Expert Systems with Applications 41(6), 2989-3004 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "62. Yuan, J., Liu, Y., Shen, C., Wang, Z., Li, H.: A simple baseline for semi-supervised semantic segmentation with strong data augmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8229-8238 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "63. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "64. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "65. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: Pro" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Howlader et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 304 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11350-11359 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 183 + ], + "type": "text", + "content": "66. Zhong, Y., Yuan, B., Wu, H., Yuan, Z., Peng, J., Wang, Y.X.: Pixel contrastive-consistent semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7273-7282 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "67. Zhou, T., Wang, S., Bilmes, J.: Time-consistent self-supervision for semi-supervised learning. In: International Conference on Machine Learning. pp. 11523-11533. PMLR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "68. Zhou, Y., Sun, X., Zha, Z.J., Zeng, W.: Context-reinforced semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4046-4055 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "69. Zou, Y., Zhang, Z., Zhang, H., Li, C.L., Bian, X., Huang, J.B., Pfister, T.: Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "70. Zuo, S., Yu, Y., Liang, C., Jiang, H., Er, S., Zhang, C., Zhao, T., Zha, H.: Self-training with differentiable teacher. arXiv preprint arXiv:2109.07049 (2021)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Pixels" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_content_list.json b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..14d2b9b2ad73a593d9a9630524c601473d1f11ec --- /dev/null +++ b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_content_list.json @@ -0,0 +1,1824 @@ +[ + { + "type": "text", + "text": "Beyond Prompt Learning: Continual Adapter for Efficient Rehearsal-Free Continual Learning", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyuan Gao $^{1\\#}$ , Songlin Dong $^{2\\#}$ , Yuhang He $^{2\\star}$ , Qiang Wang $^{1}$ , and Yihong Gong $^{1,2}$", + "bbox": [ + 256, + 210, + 746, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ School of Software Engineering, Xi'an Jiaotong University", + "$^{2}$ College of Artificial Intelligence, Xi'an Jiaotong University {gxy010317, dsl97273141, qwang}@stu.xjtu.edu.cn, heyuhang@xjtu.edu.cn ygong@mail.xjtu.edu.cn" + ], + "bbox": [ + 240, + 253, + 759, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The problem of Rehearsal-Free Continual Learning (RFCL) aims to continually learn new knowledge while preventing forgetting of the old knowledge, without storing any old samples and prototypes. The latest methods leverage large-scale pre-trained models as the backbone and use key-query matching to generate trainable prompts to learn new knowledge. However, the domain gap between the pre-training dataset and the downstream datasets can easily lead to inaccuracies in key-query matching prompt selection when directly generating queries using the pre-trained model, which hampers learning new knowledge. Thus, in this paper, we propose a beyond prompt learning approach to the RFCL task, called Continual Adapter (C-ADA). It mainly comprises a parameter-extensible continual adapter layer (CAL) and a scaling and shifting (S&S) module in parallel with the pre-trained model. C-ADA flexibly extends specific weights in CAL to learn new knowledge for each task and freezes old weights to preserve prior knowledge, thereby avoiding matching errors and operational inefficiencies introduced by key-query matching. To reduce the gap, C-ADA employs an S&S module to transfer the feature space from pre-trained datasets to downstream datasets. Moreover, we propose an orthogonal loss to mitigate the interaction between old and new knowledge. Our approach achieves significantly improved performance and training speed, outperforming the current state-of-the-art (SOTA) method. Additionally, we conduct experiments on domain-incremental learning, surpassing the SOTA, and demonstrating the generality of our approach in different settings.", + "bbox": [ + 259, + 339, + 743, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Continual Learning, Rehearsal-Free", + "bbox": [ + 261, + 685, + 589, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 722, + 375, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, with the emerging applications of deploying DNN models to dynamic and continuous environments such as autonomous driving [36, 44]. Continual Learning (CL) [18, 35, 38, 51, 53] has drawn a rapidly increasing research", + "bbox": [ + 212, + 750, + 787, + 797 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Yuhang He is the corresponding author; # Xinyuan Gao and Songlin Dong are co-first authors", + "bbox": [ + 217, + 811, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8bf0cbaa9cdc6fc3f0fbbc7cea90d34e7dc484b8c28456748cd6a8f5f5ece7e.jpg", + "image_caption": [ + "(a): Prior Prompt-Based Approach" + ], + "image_footnote": [ + "Decision Boundary Feature Tuned Frozen S85 Scaling & Shifting First Pass-forward Second Pass-forward" + ], + "bbox": [ + 217, + 143, + 475, + 266 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9ac87f5cd12fe58cdb8832d1c8dc2c77510fc3b23e3f314ee8d843ffd36c651.jpg", + "image_caption": [ + "(b):Continual Adapter Approach" + ], + "image_footnote": [], + "bbox": [ + 488, + 142, + 782, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36a73c6bb81b6f73b1ab85a59a75640dc6f32ba131889bd45bd7e98d1749203b.jpg", + "image_caption": [ + "(c): The Results $(\\%)$ of each new task on 5-task ImageNet-R", + "Fig. 1: Prior prompt-based approaches pass through the pre-trained model to generate the query and employ key-query matching to select prompts, which are inserted into the pre-trained model again (each layer has a unique prompt). Continual Adapter approach (C-ADA) strategy eliminates the need for key-query matching by introducing the CAL and S&S, which brings significant improvement in learning the new knowledge. Moreover, C-ADA only needs to pass the pre-trained model once, highlighting the training speed." + ], + "image_footnote": [], + "bbox": [ + 511, + 210, + 767, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "interest both in the literature and industry. Classical CL methods focus on addressing the problem of catastrophic forgetting [13], wherein the models suffer dramatic performance decreases on previous tasks once training on new data. The majority of research studies [8, 15, 18, 35, 41, 53, 57, 58] attempt to reduce forgetting by retaining a subset of previous image samples or prototypes and replaying them in new tasks, referred to as the rehearsal-based methods. However, on the concerns of data privacy and deployment efficiency, the applications of these methods in real-world scenarios are hampered. Therefore, we focus on a more challenging rehearsal-free CL (RFCL) [23, 25, 38, 50, 51], which does not store any previous samples or prototypes.", + "bbox": [ + 212, + 414, + 787, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Latest approaches [38,40,50,51] leverage prompt learning techniques to solve the RFCL problem. Due to freezing the large-scale pre-trained model, learning new knowledge (i.e. plasticity) becomes more challenging rather than reducing the forgetting (i.e. stability) [38]. To acquire new knowledge, prompt-based methods utilize key-query matching to obtain trainable prompts. As depicted in Figure 1(a), these methods first generate a feature representation (i.e. query) for each instance using the pre-trained model. Subsequently, they select or generate prompts for each instance based on the cosine similarity between the query and keys, followed by a second pass through the pre-trained model. Despite their promising progress, they are faced with two challenges. 1) Due to the substantial gap between the pre-trained dataset and the downstream datasets, the query embeddings directly generated by the pre-trained model are prone to confusion, leading to inaccurate key-query matching prompt selection. The inaccurate selection seriously hampers the learning ability of the model. 2) These methods require two forward passes, one for calculating the query and another for logit, leading to relatively low efficiency. Therefore, there are inherent shortcomings in prompt-based methods for RFCL tasks, motivating us to explore a novel approach distinct from prompt-learning techniques.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taking into account the inherent drawbacks mentioned above, in this paper, we shift our focus to adapter tuning [19] and propose a simple yet effective adapter variant, named ContinuaIAdapter (C-ADA) for rehearsal-free continual learning. Compared to prompt-learning techniques, C-ADA eliminates the necessity for key-query matching, fundamentally avoiding issues arising from erroneous key-query matches that could lead to conflicts in training parameters. As depicted in Figure 1 (c), this results in a substantial enhancement in the accuracy of C-ADA for new tasks, and C-ADA achieves this with just a single forward pass, greatly boosting training speed. Moreover, in contrast to lacking incremental capability adapter fine-tuning methods [5,19], we replace the adapter layer with a plug-and-play, parameter-expandable adapter layer, named Continuous Adapter Layer (CAL). The CAL expands the specific learnable weights in a new dimension (middle dimension) to acquire new knowledge and freeze the previously learned weights to retain old knowledge, which empowers the model with the capacity to tackle complex continual scenarios. This fundamentally encourages knowledge reuse, as the previous projection weights will contribute to future projection weights, facilitating the learning of new knowledge. We also introduce an orthogonal loss function to alleviate the conflict between new weights and previous weights, which significantly reduces the forgetting of old tasks. To further solve the problem of the gap between pre-trained datasets and downstream datasets, C-ADA includes scaling and shifting (S&S) module with negligible parameters to transfer the feature space from pre-trained datasets to downstream datasets before the incremental phase. Compared to the current SOTA methods, our model demonstrates significant improvements in both performance and training speed under equivalent parameter counts for RFCL. Furthermore, to validate its generality, we extend the method to the Domain-incremental Learning (DIL) setting and once again outperform the SOTA methods. The key contributions of this paper are summarized as follows:", + "bbox": [ + 212, + 145, + 787, + 569 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a simple yet effective approach called C-ADA, which eliminates the necessity for key-query matching, for RFCL, characterized by better plasticity and more efficient training compared to existing continual prompting approaches.", + "- We design a novel CAL module, which comprises a parameter-extensible down-projection and up-projection. It facilitates the preservation of previously acquired knowledge and the learning of new knowledge.", + "- We develop an S&S module to further reduce the divergence between the pre-training dataset and the downstream datasets, thereby enhancing the learning of new tasks. Additionally, we introduce an orthogonal loss to mitigate the interaction between old and new knowledge, which significantly contributes to the preservation of old knowledge.", + "- We conduct extensive experiments on well-established rehearsal-free benchmarks and achieve a new SOTA performance. Meanwhile, we evaluate our approach on the DIL setting and again outperform the state of the art, highlighting the generality." + ], + "bbox": [ + 225, + 589, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Continual Adapter", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1Rehearsal-Based Continual Learning", + "text_level": 1, + "bbox": [ + 215, + 176, + 563, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[8, 11, 15, 39, 41, 45, 47–49, 53, 57, 58] retain prior information (old exemplars or prototypes) for subsequent tasks, aimed at mitigating catastrophic forgetting by replay them in the new tasks. Exemplars-based methods [8, 11, 15, 41, 45, 53] store a portion of old exemplars in a memory buffer. These exemplars are trained alongside the current samples to optimize the model. For instance, iCaRL [35] and its variants [18, 41, 53] prevent forgetting by employing the herding technique for exemplar selection and designing distinct distillation losses. However, exemplar-based methods not only require extra memory but also experience a significant performance decline as the buffer size decreases. Exemplars-free methods [57, 58] typically store a prototype per class to refine the classifier. For example, [57] memorizes and augments prototypes to maintain the decision boundary. Some latest methods SLCA [55], RanPAC [29], HiDe-Prompt [46] combine the pretrained model and prototypes to achieve better performance. Since they are not rehearsal-free continual learning, we have excluded comparisons with them.", + "bbox": [ + 212, + 200, + 787, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Nevertheless, preserving previous information may introduce privacy and data leakage risks, often impractical in real-world applications. This motivates us to concentrate on the rehearsal-free setting, which offers enhanced privacy protection and real-world application.", + "bbox": [ + 212, + 412, + 785, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2Rehearsal-FreeContinual Learning", + "text_level": 1, + "bbox": [ + 215, + 494, + 549, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Early rehearsal-free methods [1, 25, 54] utilize regularization to alleviate model forgetting. While these approaches demonstrate strong performance in task-incremental continual learning, their effectiveness diminishes in the more challenging class-incremental learning settings. Generative-based methods [6, 14, 37] employ deep model inversion to generate images from previous tasks. While these methods perform well in some scenarios, inversion is hampered by slow training speed and high computational costs. Recent studies [16, 38, 40, 50, 51, 56] have incorporated large pre-trained models and prompt learning into the realm of continual learning. Large-scale pre-trained models have demonstrated remarkable generalization performance and robust anti-forgetting capabilities. Therefore, prompt-based methods serve as our primary comparison benchmarks and they will be subsequently elaborated upon in detail.", + "bbox": [ + 212, + 520, + 787, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Prompting for Continual Learning", + "text_level": 1, + "bbox": [ + 215, + 724, + 545, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt learning has garnered widespread success in both natural language processing (NLP) [7, 20, 33] and computer vision (CV) [21, 32, 34, 42]. It leverages prompts to fine-tune large-scale pre-trained models for downstream tasks. Firstly, L2P [51] pioneers the integration of prompt learning into the field of continual learning field. It proposes a key-query matching strategy for prompt selection from the prompt pool, followed by the insertion of these prompts into", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the pre-trained model. The subsequent works are based on this key-query matching strategy. DualPrompt [50] employs a set of task-specific prompts and selects prompts based on the closest key-query matching during inference. Furthermore, CODA [38] utilizes key-query matching to assign weights to prompts and can be optimized in an end-to-end fashion. APG [40] proposes a learnable adaptive prompt generator eases the reliance on intensive pretraining. S-Prompts [49] focuses on domain-incremental learning by acquiring distinct prompts for each domain and utilizing KNN to select prompts during inference. In this paper, we thoroughly compared these methods to CIL and DIL protocols and extensively discussed their shortcomings.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 319, + 339, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Setting", + "text_level": 1, + "bbox": [ + 215, + 349, + 401, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The continual learning aims to enable the model to learn non-stationary data from sequential tasks while preserving the knowledge obtained from previous tasks. We define a set of tasks $\\mathcal{D}_1, \\mathcal{D}_2, \\dots, \\mathcal{D}_T$ , where $\\mathcal{D}_t = \\{(x_i^t, y_i^t)\\}_{i=1}^{N_i}$ is the $t$ -th task with $N_i$ samples, and $x_i^t$ and $y_i^t$ is the $i$ -th images and its label. The label sets of $\\mathcal{D}_t$ are defined as $\\mathcal{C}_t$ . The objective of rehearsal-free continual learning is to employ a unified model for classifying test images from all previously learned tasks while not storing any old samples or prototypes.", + "bbox": [ + 212, + 372, + 787, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Depending on the specific problems, continual learning can be divided into multiple settings, which differ in challenge and practicality. In this paper, we mainly focus on addressing the challenging class-incremental learning setting. In this scenario, the class of each task is non-overlapping, denoted as $\\mathcal{C}_1 \\cap \\mathcal{C}_2 \\cdots \\cap \\mathcal{C}_T = \\emptyset$ and the task identity can not be obtained in the inference phase. Furthermore, to assess the robustness and generalization of our approach, we extend our experiments to domain-incremental continual learning setting. In this scenario, each task has the same classes but exhibits distinct data distributions, expressed as $\\mathcal{C}_1 = \\mathcal{C}_2 = \\dots = \\mathcal{C}_T$ .", + "bbox": [ + 212, + 479, + 787, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Framework and Formulation", + "text_level": 1, + "bbox": [ + 215, + 635, + 496, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 illustrates the framework of the Continual Adapter (C-ADA) approach. The C-ADA incorporates the S&S module and the plug-and-play Continual Adapter Layer (CAL) into the pre-trained model, endowing it with robust capabilities for addressing continual learning challenges. Concretely, the S&S module is exclusively trained before the incremental training to mitigate the domain gap and is subsequently frozen during further training. The CAL comprises a parameter-extensible down-projection, a ReLU activation, and a parameter-extensible up-projection. During incremental learning, the down-projection and up-projection expand the trainable weights to learn the new knowledge and freeze the previous weights to maintain the old knowledge. Moreover, we propose an orthogonal loss to alleviate the conflict between previous weights and trainable weights in each CAL.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Continual Adapter", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/791ed5520d531daa20d9b6d20a5396cb1cdf36a3aee58e9740e233aae1e50028.jpg", + "image_caption": [ + "Fig. 2: The framework of our C-ADA approach. For simplicity, we omit the skip connection and Layernorm in the figure. or loss represents the orthogonal loss. We attach our S&S and CAL, which are in parallel with the projection layer and MLP, to the shallow N layers of ViT and freeze the pre-trained backbone. For each new task, we expand two trainable weights in the CAL to learn the new knowledge and freeze the previous weights. Different from prior works, our approach uses a novel adapter variant to eliminate the necessity for key-query matching. Only trainable weights and classifier parameters are optimized which is parameter efficient and no old information (old images or prototypes) are stored which is privacy preserving." + ], + "image_footnote": [], + "bbox": [ + 243, + 142, + 759, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The basic backbone consists of a pre-trained vision transformer feature extractor $f = f(\\cdot ;\\theta)$ and a classifier $g = g(\\cdot ;\\phi)$ . Given an input image $x \\in \\mathbb{R}^{H\\times W\\times C}$ from $\\mathcal{D}_t$ . The image is reshaped and projected by the patch embedding layer, resulting in the embedding feature $x_{e} \\in \\mathbb{R}^{L\\times D}$ , where $L$ is the token length, and $D$ is the embedding dimension. These embedding features are not only input to the pre-training module within the ViT block but also to the parallel CAL and S&S modules. The output of the backbone and parallel is added and fed into the next layer. Finally, the output features $x_{p}$ from the last ViT block are fed into the $g$ to perform classification tasks.", + "bbox": [ + 212, + 501, + 787, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Scale and Shift Module", + "text_level": 1, + "bbox": [ + 215, + 667, + 455, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the significant domain gap between the pre-training dataset and the downstream dataset, our C-ADA approach introduces a flexible and independent module, named scaling and shifting (S&S), to revise the feature space before incremental training for more efficient fine-tuning (higher plasticity).", + "bbox": [ + 212, + 700, + 785, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Concretely, the S&S module includes the trainable parameter $\\alpha \\in \\mathbb{R}^D$ and $\\beta \\in \\mathbb{R}^D$ as the scale and shift factors, respectively, to learn the domain information of downstream tasks. The S&S operation is calculated as follows:", + "bbox": [ + 212, + 763, + 785, + 809 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ny = \\alpha \\odot x + \\beta \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 825, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\odot$ is the dot product. We employ the S&S operation before the incremental training by using the data from the first phase (i.e. $\\mathcal{D}_1$ ). Then we freeze the trainable parameters during the incremental phases.", + "bbox": [ + 214, + 146, + 787, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Continual Adapter Layer", + "text_level": 1, + "bbox": [ + 214, + 213, + 467, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Structure of Continual Adapter Layer To address the challenges of efficiency and lightweight, the CAL is designed as a bottleneck module. It consists of a parameter-extensible down-projection, a ReLU layer for non-linear property, and a parameter-extensible up-projection. The down-projection embeds the old and new knowledge into a latent embedding along with the ReLU function, and the up-projection extracts a dimension-fixed knowledge representation from the latent embedding. For each new task, we first initialize and append two learnable linear weights to the down-projection and up-projection, respectively, and freeze the previous weights. Such design ensures that the CAL is extendable for learning new knowledge and stable in maintaining old knowledge, maintaining the input and output feature dimensions unchanged. Assuming the current task $t$ , we assign the weights in the CAL into two types: the frozen weights $\\mathbf{W}^{1},\\dots,\\mathbf{W}^{(t - 1)}$ , and the trainable weight $\\mathbf{W}^t$ ( $\\mathbf{W}^i$ represents both the $\\mathbf{W}_{dp}^i$ and $\\mathbf{W}_{up}^i$ ). The frozen weights effectively preserve the previous knowledge to mitigate catastrophic forgetting, while simultaneously optimizing the trainable parameters to learn new knowledge.", + "bbox": [ + 212, + 238, + 787, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without loss of generality, we define the input and output of the current layer as $x_{i}$ and $x_{i}^{\\prime}$ , respectively. In task t, we can adapt the input by the CAL as follows:", + "bbox": [ + 212, + 482, + 787, + 523 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {d p} = \\left[ \\begin{array}{c} \\mathbf {W} _ {d p} ^ {1}, \\dots , \\mathbf {W} _ {d p} ^ {t - 1}, \\mathbf {W} _ {d p} ^ {t} \\end{array} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 523, + 607, + 541 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {W} _ {u p} = \\left[ \\mathbf {W} _ {u p} ^ {1}, \\dots , \\mathbf {W} _ {u p} ^ {t - 1}, \\mathbf {W} _ {u p} ^ {t} \\right], \\\\ \\mathbf {W} _ {d p} ^ {i} \\in \\mathbb {R} ^ {D \\times d} \\quad \\mathbf {W} _ {u p} ^ {i} \\in \\mathbb {R} ^ {d \\times D} \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 545, + 784, + 583 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} ^ {\\prime} = \\operatorname {R e L U} (x _ {i} \\cdot \\mathbf {W} _ {d p}) \\cdot \\mathbf {W} _ {u p}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 588, + 589, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the next task, we freeze the $\\mathbf{W}_{dp}^{t}$ and $\\mathbf{W}_{up}^{t}$ to maintain the old knowledge and convert two new weights $\\mathbf{W}_{dp}^{t + 1}$ and $\\mathbf{W}_{up}^{t + 1}$ to the trainable weights.", + "bbox": [ + 214, + 612, + 785, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameters of Continual Adapter Layer. In this paper, we follow the standard RFCL setting where the total task number is known before training. To compare fairly, we adjust the value of $d$ to maintain a consistent number of parameters across different settings. For real-world applications, we can keep the size of $d$ constant and continually expand the specific weights for new tasks. Remarkably, even when $d$ is set to a minimal value, such as $d = 1$ , it still exhibits commendable performance. Therefore C-ADA is lightweight, and memory-efficient in real-world applications.", + "bbox": [ + 212, + 666, + 787, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Position of Continual Adapter Layer. As depicted in [9,43], the Multi-Head Self-Attention (MHSA) block employs the attention mechanism to capture global", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Continual Adapter", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "information, while the MLP block prevents the ViT output from degeneration. Both components play pivotal roles in the ViT structure. Consequently, we investigate performing a comprehensive tuning of both the MHSA block and MLP block, instead of exclusively emphasizing one over the other. To fine-tune the MHSA block, we choose to insert the CAL in parallel with the Projection Layer instead of the Attention module. This decision is motivated by the observation that, despite various advanced transformer-based models using different attention mechanisms within the MHSA block [12, 26, 28], they consistently include a projection layer. On the contrary, we insert the CAL in parallel with the entire MLP block to fine-tune the MLP block.", + "bbox": [ + 212, + 146, + 787, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Without loss of generality, we define the input as $x_{l}$ and initially feed it into the Attention module to obtain $x_{a}$ , which serves as the input to the Projection Layer and $\\mathrm{CAL}_1$ :", + "bbox": [ + 212, + 297, + 787, + 342 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} x _ {a} = \\operatorname {A t t e n} \\left(x _ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q \\cdot K ^ {T}}{\\sqrt {d}}\\right) V, \\\\ x _ {l} = \\operatorname {P r e s i} \\left(x _ {l}\\right) = x _ {l} = \\lambda_ {1} \\operatorname {C A L} \\left(S _ {l} - S \\left(x _ {l}\\right)\\right) \\end{array} \\tag {3} \\\\ x _ {l 1} = \\operatorname {P r o j} (x _ {a}), \\quad x _ {l 2} = \\lambda \\cdot \\operatorname {C A L} _ {1} (S \\& S (x _ {a})), \\\\ x _ {l} ^ {\\prime} = x _ {l 1} + x _ {l 2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 349, + 784, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then the output $x_{l}^{\\prime}$ is further sent to the MLP block and $\\mathrm{CAL}_2$ . This process is formally formulated as follows,", + "bbox": [ + 212, + 430, + 785, + 460 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {l 1} = \\operatorname {M L P} \\left(x _ {l} ^ {\\prime}\\right), \\quad x _ {l 2} = \\lambda \\cdot \\operatorname {C A L} _ {2} (S \\& S \\left(x _ {l} ^ {\\prime}\\right)), \\tag {4} \\\\ x _ {l + 1} = x _ {l 1} + x _ {l 2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 469, + 784, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{CAL}_1$ and $\\mathrm{CAL}_2$ are in parallel with the projection layer and MLP block, respectively. Proj represents the Projection Layer and $\\lambda$ represents a scale factor (typically set to 0.1). Then, the output $x_{l + 1}$ is sent to the next layer and repeats the above process.", + "bbox": [ + 212, + 508, + 787, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Loss Function", + "text_level": 1, + "bbox": [ + 214, + 589, + 377, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Orthogonal Loss. When training the new task, the learned new knowledge will interfere with old knowledge, potentially resulting in catastrophic forgetting. Thus we propose to solve this problem by initialization and loss function. We apply the Gram-Schmidt process to initialize the trainable parameters at the start of each new task. Orthogonal initialization contributes to the diminishment of the scope when the model parameter is updated, thereby fostering enhanced stability and overall performance. After the first task, we employ the orthogonal loss to keep trainable weights orthogonal to previous weights.", + "bbox": [ + 212, + 611, + 787, + 733 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o r} = \\left\\| \\mathbf {W} _ {d p} ^ {t} \\right. ^ {T} \\cdot \\left\\{\\mathbf {W} _ {d p} ^ {1}, \\dots , \\mathbf {W} _ {d p} ^ {t - 1} \\right\\} \\| _ {2} + \\left\\| \\mathbf {W} _ {u p} ^ {t} \\right. \\cdot \\left\\{\\mathbf {W} _ {u p} ^ {1}, \\dots , \\mathbf {W} _ {u p} ^ {t - 1} \\right\\} ^ {T} \\| _ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 738, + 784, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_2$ represent the $L_{2}$ -norm and $T$ represents the matrix transpose. The loss function forces the model to optimize current weight $\\mathbf{W}^t$ to be orthogonal to previous weight $\\mathbf{W}^1, \\dots, \\mathbf{W}^{t-1}$ . Intuitively, orthogonal vectors exhibit reduced mutual influence. As a result, this design aims to minimize interference between new and old knowledge to reduce catastrophic forgetting.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Classification Loss. Similar to the previous methods, we use the CE loss to optimize the model to learn the new knowledge", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = \\mathcal {L} (g (f (\\boldsymbol {x}; \\mathbf {W} _ {d p}, \\mathbf {W} _ {u p}, \\theta); \\phi), y) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 188, + 785, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $x$ is the input image and $y$ is the corresponding label. $\\phi$ is the parameters of the classifier and $\\theta$ is the parameters of the pre-train model.", + "bbox": [ + 214, + 215, + 782, + 246 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Total Optimization Objective. For task $t$ , our full optimization consists of a classification loss $\\mathcal{L}_{ce}$ and an orthogonal loss $\\mathcal{L}_{or}$", + "bbox": [ + 214, + 267, + 785, + 297 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {W} _ {\\mathrm {d p}} ^ {t}, \\mathbf {W} _ {\\mathrm {u p}} ^ {t}, \\phi} \\mathcal {L} _ {c e} + \\delta \\mathcal {L} _ {\\mathrm {o r}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 309, + 785, + 335 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{dp}^{t},\\mathbf{W}_{up}^{t}$ and $\\phi$ are the learnable parameters of task t. We set the $\\delta$ to 1 for all the experiments. During each training phase, our models can be optimized in an end-to-end fashion.", + "bbox": [ + 214, + 345, + 782, + 391 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 414, + 375, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we mainly aim to solve the rehearsal-free class-incremental learning problem. To evaluate our methods, we follow the RFCL settings proposed in previous works [38, 50, 51] and conduct comprehensive experiments.", + "bbox": [ + 214, + 446, + 782, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Besides, to prove the robustness and generality of our method on other continual learning settings, we conduct experiments compared with the SOTA methods [49] on the domain-incremental learning settings. Extensive ablation studies prove the effectiveness of our methods.", + "bbox": [ + 214, + 491, + 782, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Evaluation Benchmarks", + "text_level": 1, + "bbox": [ + 215, + 573, + 455, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We use the Split CIFAR-100 [24] and Split ImageNet-R [17] for class-incremental setting, CORe50 [27] and DomainNet [30] for domain-incremental setting.", + "bbox": [ + 214, + 597, + 782, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CIFAR-100 is a widely-used benchmark in continual learning, which contains 60000 images of $32 \\times 32$ size from 100 classes. We split CIFAR-100 into the 10-task benchmark following previous works.", + "bbox": [ + 214, + 643, + 782, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ImageNet-R has 200 classes and includes newly collected data of different styles. It provides a fair and challenging problem setting due to the significant gap between training data and pre-trained data. We split the dataset into the 5-task benchmark, the original 10-task benchmark and the 20-task benchmark.", + "bbox": [ + 214, + 688, + 782, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CORe50 has 50 classes from 11 distinct domains. Following the domain-incremental setting, we use 8 domains for (120000 images) for continual training and the rest of the domains for testing.", + "bbox": [ + 214, + 750, + 782, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "DomainNet is a large dataset with 345 classes and 600000 images. These images are split into 6 domains (Clipart, Infograph, Paint, Quickdraw, Real, and Sketch) for continual learning like S-Prompts [49].", + "bbox": [ + 214, + 794, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Continual Adapter", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3a198af5df8c2815e472c8543763db7e12191fde06bc950a90e38da51c27fce0.jpg", + "table_caption": [ + "Table 1: Results (\\%) on ImageNet-R and CIFAR-100. The results are all obtained by CODA [38] directly. P represents the total number of tasks. UB represents the full fine-tuning result from [10]. We report $A_N(\\%)$ , $Param(\\%)$ and $FLOPs(\\mathrm{G})$ to bring intuitive comparison. C-ADA uses fewer parameters and doubles the training speed, greatly surpassing other methods." + ], + "table_footnote": [], + "table_body": "
MethodsImageNet-RCIFAR-100FLOPs
P=5P=10P=20P=10
AN(↑)Param(↓)AN(↑)Param(↓)AN(↑)Param(↓)AN(↑)Param(↓)
UB77.13100/10077.13100/10077.13100/10089.30100/100-
ER (5000)71.72100/10064.43100/10052.43100/10076.20100/100-
FT18.74100/10010.12100/1004.75100/1009.92100/100-
FT++60.42100/10048.93100/10035.98100/10049.91100/100-
LwF.MC74.56100/10066.73100/10054.05100/10064.83100/100-
L2P70.830.7/100.769.290.7/100.765.890.7/100.782.500.7/100.735.2
Deep L2P73.939.6/109.671.669.6/109.668.429.6/109.684.309.6/109.635.2
DualPrompt73.050.5/100.571.320.8/100.867.871.3/101.383.050.8/100.835.2
CODA-P-S75.190.7/100.773.930.7/100.770.530.7/100.784.590.6/100.635.2
C-ADA77.930.7/100.776.660.7/100.773.470.7/100.787.180.6/100.617.6
", + "bbox": [ + 232, + 226, + 767, + 377 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparing Methods. We compare our method with several baselines and state-of-the-art (SOTA) continual learning methods. For a fair comparison, all the methods are based on a ViT-B/16 backbone pre-trained on ImageNet-1K. Under class-incremental settings, we regard L2P [51], DualPrompt [50], and CODA [38] as the main competitors, which are rehearsal-free methods. Under domain-incremental settings, despite S-Prompt [49] storing the prototypes of previous domains instead of the rehearsal-free method, it is still the main comparison method as the SOTA of domain-incremental learning.", + "bbox": [ + 212, + 401, + 787, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. Similar to CODA [38], we use the Adam [22] optimizer with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ , and a batch size of 128 images in four GPUs. For every image, we resize it to $224\\times 224$ and normalize them to [0,1]. The learning rate is set to $5e^{-5}$ for all the settings with cosine-decaying. For fair comparisons, we set the total middle dimension to 60 (sum of two CALs) and insert the CAL into layers 0-4, which achieves a similar parameter amount to previous works [38, 50]. We train the CIFAR-100 for 20 epochs, ImageNet-R for 50 epochs, CORe50 and DomainNet for 10 epoch. During the training phase, we substitute predictions from previous-task logits with negative infinity during the training of a new task. This design leads to a softmax prediction of \"0\" for these previous task classes and ensures that gradients do not affect the linear heads of past task classes like previous works. [38, 50, 51]", + "bbox": [ + 212, + 542, + 787, + 724 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 214, + 743, + 372, + 757 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Class-incremental learning. We report the results on two class-incremental datasets in Table 1 following the previous work [38].", + "bbox": [ + 212, + 765, + 785, + 797 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "3 For the sake of fairness, we primarily selected CODA-P-S with the same parameter count as C-ADA for comparison.", + "bbox": [ + 217, + 810, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8eae0fae0d65d7cf67346d7874826bc9c0bf3d6c0f7793c65a6073d07e5fcddf.jpg", + "table_caption": [ + "Table 2: Results (\\%) on CORe50 and DomainNet. $A_N$ gives the accuracy averaged over tasks and buffer size represents the number of old exemplars stored in the memory buffer. The results are all obtained by S-Prompt [49] directly. We reproduce the CODA on these settings for further comparison. We ensure that all methods have the same pre-trained model (ViT-B/16)." + ], + "table_footnote": [], + "table_body": "
MethodBuffer size\\(A_N(↑)\\)
Upper-Bound-84.01
ER80.10
GDumb74.92
BiC79.28
DER++50/class79.70
Co2L79.75
DyTox79.21
L2P81.07
EWC74.82
LwF75.45
L2P0/class78.33
CODA-P-S85.41
S-Prompts83.13
C-ADA89.45
", + "bbox": [ + 245, + 226, + 459, + 431 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/74060bde1773f12663c5c8d099cafcce386907d682122a303d4d3b47b536dcf1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBuffer sizeAN(↑)
Upper-Bound-63.22
DyTox50/class62.94
EWC47.62
LwF49.19
SimCLR44.20
BYOL49.70
Barlow Twins0/class48.90
Supervised Contrastive50.90
L2P40.15
CODA-P-S47.56
S-Prompts50.62
C-ADA53.00
", + "bbox": [ + 514, + 238, + 787, + 417 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ImageNet-R: We report the results with 5-tasks, 10-tasks, and 20-tasks to evaluate the robustness of the model under different settings in Table 1. It is worth noting that our method significantly surpasses the previous methods in average accuracy under different settings with as much as $+2.74\\%$ , $+2.73\\%$ and $+2.94\\%$ over CODA, $+4.88\\%$ , $+5.41\\%$ and $+5.60\\%$ over DualPrompt (second best method). This phenomenon shows that our method is robust to different class increment settings. Moreover, since our method only requires forward propagation one time. We achieved an obvious increase in training speed while improving accuracy. Combining accuracy and training speed brings better practicality in the real world.", + "bbox": [ + 212, + 455, + 787, + 607 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "CIFAR-100: We note that our approach also brings similar results to the ImageNet-R benchmarks, with an improvement of $+2.59\\%$ compared with CODA, $+4.13\\%$ compared with DualPrompt.", + "bbox": [ + 212, + 607, + 787, + 654 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Domain-incremental learning. We report the results on two domain-incremental datasets in Table 2. Besides, we reproduced CODA on these datasets to bring intuitive comparison.", + "bbox": [ + 212, + 672, + 808, + 718 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "CORe50 and DomainNet are widely used domain dataset for continual learning. Following the setting of S-Prompt [49], we divide the baseline methods into two groups, exemplars-based methods [2-4, 11, 31, 52] and exemplars-free methods [49]/rehearsal-free methods [38, 51]. Exemplars-based methods store 50 old exemplars per class in the memory buffer. These methods often have better accuracy but are accompanied by the risk of privacy leakage and memory overhead. Our method surpasses the CODA by $+4.04\\%$ and $+5.44\\%$ on CORe50 and DomainNet benchmark. Besides, it is worth noting that our method even out", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Continual Adapter", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b9d7987e367c2b72787914d04d9403be925148a5712b7e56136f7e56603427c0.jpg", + "image_caption": [ + "Fig. 3: Average accuracy $A_{N}$ (\\%) vs tuning parameters (\\%). We report the results on the 20-tasks ImageNet-R." + ], + "image_footnote": [], + "bbox": [ + 228, + 142, + 478, + 256 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/aba5c5f980a5b2d6d199b9b77661f0d1878c833e2d372d348ab9aa2e0f244cef.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR-100ImageNet-R
C-ADA87.3276.66
Ablate S&S86.1175.93
Ablate Lor82.8972.17
Ablate CAL76.9364.45
", + "bbox": [ + 535, + 157, + 767, + 224 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Ablation Results (%) on 10-task ImageNet-R and CIFAR-100. $A_N$ gives the accuracy averaged over tasks. We ablate the key components in turn and report the results.", + "bbox": [ + 508, + 234, + 787, + 306 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "performs S-Prompt, which is specifically designed for domain-incremental learning, with improvements of $+6.32\\%$ and $+2.38\\%$ . This interesting phenomenon demonstrates the robustness and effectiveness of our method in various continual learning settings.", + "bbox": [ + 215, + 349, + 787, + 411 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Key Components Ablation", + "text_level": 1, + "bbox": [ + 215, + 433, + 482, + 448 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted ablation experiments to analyze the importance of each component. We incrementally introduced key components on both the 10-tasks ImageNet-R and CIFAR-100 datasets, and the results are presented in Table 3. Firstly, we ablate the S&S module, the performance has dropped $(0.73\\% \\downarrow$ on ImageNet-R). This shows that revising the feature space before training is beneficial to improving subsequent learning ability. Secondly, removing the orthogonal loss leads to severe performance degradation on average accuracy $(3.76\\% \\downarrow$ on ImageNet-R). This outcome aligns with our expectations, as ablating the orthogonal loss results in our method lacking a mechanism to prevent forgetting, leading to severe catastrophic forgetting. Thirdly, when we ablated the CAL component, we observed a noticeable drop in performance $(7.72\\% \\downarrow$ on ImageNet-R). The removal of the CAL implies that our model relies solely on adjusting the classifier to tune the model, which obviously leads to unsatisfactory results, particularly when dealing with the downstream datasets having a significant gap from the pre-trained dataset.", + "bbox": [ + 212, + 458, + 787, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Scaling Trainable Parameters", + "text_level": 1, + "bbox": [ + 215, + 708, + 504, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct experiments to discuss the relationship between the trainable parameters amount and the performance.", + "bbox": [ + 212, + 734, + 785, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The trainable parameters amount can be adjusted by changing the middle dimension of the CAL. The total number of middle dimensions of the two CALs in each block is chosen from $\\{20, 40, 60, 80, 100, 120, 140, 160\\}$ , corresponding to $\\{0.35\\%, 0.54\\%, 0.71\\%, 0.90\\%, 1.12\\%, 1.27\\%, 1.45\\%, 1.61\\}$ trainable parameters. We report the results on 20-tasks ImageNet-R in Figure 3. The figure", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "shows that more trainable parameters contribute to achieving a higher accuracy. Moreover, we can notice that even with the fewer trainable parameters, our model consistently demonstrates superior performance when compared to CODA (71.92% $A_N$ , 0.35% $\\text{Param}$ of C-ADA vs 70.53% $A_N$ , 0.7% $\\text{Param}$ of CODA). This fully illustrates that our approach improves the plasticity compared with previous works.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Attaching Layer and Position", + "text_level": 1, + "bbox": [ + 214, + 258, + 504, + 273 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Intuitively, each layer of the ViT exhibits distinct feature extraction capabilities [50]. The impact of tuning different layers may be quite different in CL. Furthermore, each module of the layer plays different roles, which causes attaching CAL to various positions within the layer may yield divergent outcomes. Therefore, it is crucial to explore layers and positions to attach the CAL under the CL settings. Note that the total number of trainable parameters remains consistent across all experiments described below (middle $\\dim = d*t$ ).", + "bbox": [ + 212, + 282, + 787, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Layer of CAL. For simplicity, we assume that the attaching layers are contiguous. We report the result in Table 4.", + "bbox": [ + 212, + 410, + 785, + 439 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Firstly, we attach the CAL to all the layers to get a preliminary result (Layer 0 to Layer 11). Next, we use binary search to analyze the importance of shallow (Layer 0 to Layer 5) and deep (Layer 6 to Layer 11) layers. We observe that attaching to shallow layers gets superior results compared to attaching to deeper layers. This phenomenon is consistent with the observation from previous work [50]. Then we continue to reduce the number of attached layers in shallow layers. Experimental results show that attaching from layer $= 0$ to layer $= 4$ performs the best. We use this conclusion to analyze the attached modules.", + "bbox": [ + 212, + 440, + 787, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Position of CAL. By analyzing common parts of the ViT architecture, we select three alternative attaching positions: QKV attention layer $(W_{q}, W_{k}, W_{v})$ in the MHSA block, projection layer in the MHSA block, and the MLP block. We report the result in Table 5.", + "bbox": [ + 212, + 582, + 785, + 642 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Firstly, we analyze the effects of only attaching these three positions, respectively. We observe that attaching the CAL to the projection layer has a better effect than in the QKV attention layer. This phenomenon makes us attach the CAL to the projection layer to tune the MHSA block. Then we opt to combine the tuning of different positions. We observe that attaching the CAL to different positions resulted in certain improvements (about $0.4\\%$ ). Interestingly, it seems to have little to do with how to assign the middle dimensions. This shows to some extent that our method is robust.", + "bbox": [ + 212, + 643, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Extra Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 785, + 472, + 799 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Although we emphasize that $A_N$ , $\\text{Param}$ , and $FLOPs$ are the more important metrics, there are still some common metrics in the CL field.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Continual Adapter", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1e28a3ef1f19537e72a26c748fdc20a9eaa1dc35a70eb99530131a29fd96071b.jpg", + "table_caption": [ + "Table 4: Results $(\\%)$ on ImageNetR 10 tasks. We analyze the difference of attaching to different layers. We keep the total tunable parameters constant by modifying the middle dimension," + ], + "table_footnote": [], + "table_body": "
Attach Layermiddle dimAN(↑)
Layer 0 ~ Layer 113075.23
Layer 6 ~ Layer 115072.51
Layer 0 ~ Layer 55076.21
Layer 1 ~ Layer 56076.29
Layer 0 ~ Layer 46076.66
Layer 2 ~ Layer 58075.72
Layer 0 ~ Layer 38075.75
", + "bbox": [ + 230, + 232, + 473, + 339 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/d7bf108739df39f5fda6ff6463e99af49293618680529f79cf58eea3f35294db.jpg", + "table_caption": [ + "Table 5: Results $(\\%)$ on ImageNetR 10 tasks. We fix the attaching layers (layer 0 to layer 4) to analyze two aspects: the attaching positions and the middle dimension of each CAL." + ], + "table_footnote": [], + "table_body": "
Attach Positionmiddle dimAN(↑)
QKV Attention6076.04
MLP6076.23
Proj6076.25
Proj + MLP20 + 4076.52
Proj + MLP30 + 3076.57
Proj + MLP50 + 1076.51
Proj + MLP10 + 5076.61
Proj + MLP40 + 2076.66
", + "bbox": [ + 532, + 226, + 756, + 345 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "$F_{N}$ : Average forgetting metrics, represent the drop in task performance averaged over $N$ tasks.", + "bbox": [ + 212, + 371, + 782, + 398 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "bwt: Backward transfer, assesses the quality of representation.", + "bbox": [ + 238, + 401, + 686, + 415 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "fwt: Forward transfer, represents the influence of existing knowledge on the performance of subsequent concepts.", + "bbox": [ + 212, + 416, + 784, + 446 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Table 6, we can see that across various metrics, C-ADA demonstrates superior performance. Notably, in terms of $fwt$ , which represents the capacity to learn new knowledge, it significantly outperforms the other methods. This emphasizes the effectiveness of C-ADA in learning new knowledge.", + "bbox": [ + 212, + 446, + 785, + 506 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/14a7aa13c7e7bbc7e10b500b1ed180df0decc24a28ede18a81714389beb46236.jpg", + "table_caption": [ + "Table 6: Results (\\%) on ImageNet-R. The $F_N$ results are all obtained by CODA [38] directly. We reproduce the bwt and fwt metrics for further comparison." + ], + "table_footnote": [], + "table_body": "
MethodsP=5P=10P=20P=10P=10
FN(↓)FN(↓)FN(↓)bwt(↑)fwt(↑)
L2P3.36 ± 0.182.03 ± 0.191.24 ± 0.14-3.69 ± 0.07-0.32 ± 0.04
DualPrompt2.64 ± 0.171.71 ± 0.241.07 ± 0.14-3.48 ± 0.06-0.09 ± 0.04
CODA2.65 ± 0.151.60 ± 0.201.00 ± 0.15-3.24 ± 0.060.01 ± 0.03
C-ADA2.55 ± 0.141.58 ± 0.181.02 ± 0.12-3.12 ± 0.050.34 ± 0.03
", + "bbox": [ + 256, + 571, + 743, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 704, + 359, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present a novel adapter variant, named Continual Adapter (C-ADA), for RFCL. Our approach assembles a parameter-extensible adapter layer (CAL) and an S&S module in parallel with the pre-trained model. Importantly, C-ADA eliminates the necessity for key-query matching to achieve better plasticity and efficient training. We set a new SOTA on well-established benchmarks of RFCL and outperform the mainstream methods of DIL. This underscores the generality and robustness of our approach across various continual learning settings.", + "bbox": [ + 212, + 733, + 823, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was funded by the National Natural Science Foundation of China under Grant No.U21B2048 and No.62302382 and Shenzhen Key Technical Projects under Grant CJGJZD2022051714160501 and the CAAI-MindSpore Open Fund, developed on OpenI Community.", + "bbox": [ + 212, + 175, + 787, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV). pp. 139-154 (2018) 4", + "2. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems 33, 15920-15930 (2020) 11", + "3. Cha, H., Lee, J., Shin, J.: Co2l: Contrastive continual learning. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 9516-9525 (2021) 11", + "4. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019) 11", + "5. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 3", + "6. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021) 4", + "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4", + "8. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021) 2, 4", + "9. Dong, Y., Cordonnier, J.B., Loukas, A.: Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In: International Conference on Machine Learning. pp. 2793-2803. PMLR (2021) 7", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) 10", + "11. Douillard, A., Ramé, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9285-9295 (2022) 4, 11", + "12. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 8", + "13. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999) 2" + ], + "bbox": [ + 218, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Continual Adapter", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) 4", + "15. Gao, X., He, Y., Dong, S., Cheng, J., Wei, X., Gong, Y.: Dkt: Diverse knowledge transfer transformer for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24236-24245 (2023) 2, 4", + "16. Gao, Z., Cen, J., Chang, X.: Consistent prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28463-28473 (2024) 4", + "17. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021) 9", + "18. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 831-839 (2019) 1, 2, 4", + "19. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) 3", + "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 4", + "21. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) 4", + "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10", + "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) 2", + "24. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., Citeseer (2009) 9", + "25. Li, Z., Hoiem, D.: Learning without forgetting. T-PAMI 40(12), 2935-2947 (2018) 2, 4", + "26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) 8", + "27. Lomonaco, V., Maltoni, D.: Core50: a new dataset and benchmark for continuous object recognition. In: Conference on robot learning. pp. 17-26. PMLR (2017) 9", + "28. Mao, X., Qi, G., Chen, Y., Li, X., Duan, R., Ye, S., He, Y., Xue, H.: Towards robust vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12042-12051 (2022) 8", + "29. McDonnell, M.D., Gong, D., Parvaneh, A., Abbasnejad, E., van den Hengel, A.: Ranpac: Random projections and pre-trained models for continual learning. Advances in Neural Information Processing Systems 36 (2024) 4", + "30. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 9" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Prabhu, A., Torr, P.H., Dokania, P.K.: Gdumb: A simple approach that questions our progress in continual learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 524-540. Springer (2020) 11", + "32. Qin, Y., Chen, Y., Peng, D., Peng, X., Zhou, J.T., Hu, P.: Noisy-correspondence learning for text-to-image person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27197-27206 (2024) 4", + "33. Qin, Y., Peng, D., Peng, X., Wang, X., Hu, P.: Deep evidential learning with noisy correspondence for cross-modal retrieval. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4948-4956 (2022) 4", + "34. Qin, Y., Sun, Y., Peng, D., Zhou, J.T., Peng, X., Hu, P.: Cross-modal active complementary learning with self-refining correspondence. Advances in Neural Information Processing Systems 36 (2024) 4", + "35. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 1, 2, 4", + "36. Shieh, J.L., Haq, Q.M.u., Haq, M.A., Karam, S., Chondro, P., Gao, D.Q., Ruan, S.J.: Continual learning strategy in one-stage object detection framework based on experience replay for autonomous driving vehicle. Sensors 20(23), 6777 (2020) 1", + "37. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021) 4", + "38. Smith, J.S., Karlinsky, L., et al.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: CVPR. pp. 11909-11919 (2023) 1, 2, 4, 5, 9, 10, 11, 14", + "39. Song, X., He, Y., Dong, S., Gong, Y.: Non-exemplar domain incremental object detection via learning domain bias. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 15056-15065 (2024) 4", + "40. Tang, Y.M., et al.: When prompt-based incremental learning does not meet strong pretraining. In: CVPR (2023) 2, 4, 5", + "41. Tao, X., Chang, X., Hong, X., Wei, X., Gong, Y.: Topology-preserving class-incremental learning. In: European Conference on Computer Vision. pp. 254-270. Springer (2020) 2, 4", + "42. Tu, C.H., Mai, Z., Chao, W.L.: Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7725-7735 (2023) 4", + "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 7", + "44. Verwimp, E., Yang, K., Parisot, S., Hong, L., McDonagh, S., Pérez-Pellitero, E., De Lange, M., Tuytelaars, T.: Clad: A realistic continual learning benchmark for autonomous driving. Neural Networks 161, 659-669 (2023) 1", + "45. Wang, F.Y., Zhou, D.W., Ye, H.J., Zhan, D.C.: Foster: Feature boosting and compression for class-incremental learning. arXiv preprint arXiv:2204.04662 (2022) 4", + "46. Wang, L., et al.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. NeurIPS (2024) 4", + "47. Wang, S., Shi, W., Dong, S., Gao, X., Song, X., Gong, Y.: Semantic knowledge guided class-incremental learning. IEEE Transactions on Circuits and Systems for Video Technology 33(10), 5921-5931 (2023) 4" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Continual Adapter", + "bbox": [ + 604, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Wang, S., Shi, W., He, Y., Yu, Y., Gong, Y.: Non-exemplar class-incremental learning via adaptive old class reconstruction. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4524-4534 (2023) 4", + "49. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022) 4, 5, 9, 10, 11", + "50. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799 (2022) 2, 4, 5, 9, 10, 13", + "51. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022) 1, 2, 4, 9, 10, 11", + "52. Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., Fu, Y.: Large scale incremental learning. arXiv preprint arXiv:1905.13260 (2019) 11", + "53. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3014-3023 (2021) 1, 2, 4", + "54. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International conference on machine learning. pp. 3987-3995. PMLR (2017) 4", + "55. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023) 4", + "56. Zhou, D.W., et al.: Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. arXiv (2023) 4", + "57. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5871-5880 (2021) 2, 4", + "58. Zhu, K., Zhai, W., Cao, Y., Luo, J., Zha, Z.J.: Self-sustaining representation expansion for non-exemplar class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9296-9305 (2022) 2, 4" + ], + "bbox": [ + 215, + 146, + 784, + 603 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_model.json b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_model.json new file mode 100644 index 0000000000000000000000000000000000000000..712a2466a8fc00395eff6ec030ff1f4cca690283 --- /dev/null +++ b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_model.json @@ -0,0 +1,2535 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.187 + ], + "angle": 0, + "content": "Beyond Prompt Learning: Continual Adapter for Efficient Rehearsal-Free Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.212, + 0.747, + 0.245 + ], + "angle": 0, + "content": "Xinyuan Gao\\(^{1\\#}\\), Songlin Dong\\(^{2\\#}\\), Yuhang He\\(^{2\\star}\\), Qiang Wang\\(^{1}\\), and Yihong Gong\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.255, + 0.703, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Software Engineering, Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.27, + 0.761, + 0.312 + ], + "angle": 0, + "content": "\\(^{2}\\) College of Artificial Intelligence, Xi'an Jiaotong University {gxy010317, dsl97273141, qwang}@stu.xjtu.edu.cn, heyuhang@xjtu.edu.cn ygong@mail.xjtu.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.241, + 0.255, + 0.761, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.34, + 0.744, + 0.675 + ], + "angle": 0, + "content": "Abstract. The problem of Rehearsal-Free Continual Learning (RFCL) aims to continually learn new knowledge while preventing forgetting of the old knowledge, without storing any old samples and prototypes. The latest methods leverage large-scale pre-trained models as the backbone and use key-query matching to generate trainable prompts to learn new knowledge. However, the domain gap between the pre-training dataset and the downstream datasets can easily lead to inaccuracies in key-query matching prompt selection when directly generating queries using the pre-trained model, which hampers learning new knowledge. Thus, in this paper, we propose a beyond prompt learning approach to the RFCL task, called Continual Adapter (C-ADA). It mainly comprises a parameter-extensible continual adapter layer (CAL) and a scaling and shifting (S&S) module in parallel with the pre-trained model. C-ADA flexibly extends specific weights in CAL to learn new knowledge for each task and freezes old weights to preserve prior knowledge, thereby avoiding matching errors and operational inefficiencies introduced by key-query matching. To reduce the gap, C-ADA employs an S&S module to transfer the feature space from pre-trained datasets to downstream datasets. Moreover, we propose an orthogonal loss to mitigate the interaction between old and new knowledge. Our approach achieves significantly improved performance and training speed, outperforming the current state-of-the-art (SOTA) method. Additionally, we conduct experiments on domain-incremental learning, surpassing the SOTA, and demonstrating the generality of our approach in different settings." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.686, + 0.59, + 0.7 + ], + "angle": 0, + "content": "Keywords: Continual Learning, Rehearsal-Free" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.377, + 0.739 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.751, + 0.788, + 0.799 + ], + "angle": 0, + "content": "In recent years, with the emerging applications of deploying DNN models to dynamic and continuous environments such as autonomous driving [36, 44]. Continual Learning (CL) [18, 35, 38, 51, 53] has drawn a rapidly increasing research" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.812, + 0.789, + 0.84 + ], + "angle": 0, + "content": "* Yuhang He is the corresponding author; # Xinyuan Gao and Songlin Dong are co-first authors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Gao et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.476, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.269, + 0.432, + 0.28 + ], + "angle": 0, + "content": "(a): Prior Prompt-Based Approach" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.143, + 0.783, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.199, + 0.719, + 0.21 + ], + "angle": 0, + "content": "(b):Continual Adapter Approach" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.211, + 0.769, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.269, + 0.786, + 0.28 + ], + "angle": 0, + "content": "(c): The Results \\((\\%)\\) of each new task on 5-task ImageNet-R" + }, + { + "type": "image_footnote", + "bbox": [ + 0.238, + 0.282, + 0.756, + 0.294 + ], + "angle": 0, + "content": "Decision Boundary Feature Tuned Frozen S85 Scaling & Shifting First Pass-forward Second Pass-forward" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.3, + 0.789, + 0.398 + ], + "angle": 0, + "content": "Fig. 1: Prior prompt-based approaches pass through the pre-trained model to generate the query and employ key-query matching to select prompts, which are inserted into the pre-trained model again (each layer has a unique prompt). Continual Adapter approach (C-ADA) strategy eliminates the need for key-query matching by introducing the CAL and S&S, which brings significant improvement in learning the new knowledge. Moreover, C-ADA only needs to pass the pre-trained model once, highlighting the training speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.567 + ], + "angle": 0, + "content": "interest both in the literature and industry. Classical CL methods focus on addressing the problem of catastrophic forgetting [13], wherein the models suffer dramatic performance decreases on previous tasks once training on new data. The majority of research studies [8, 15, 18, 35, 41, 53, 57, 58] attempt to reduce forgetting by retaining a subset of previous image samples or prototypes and replaying them in new tasks, referred to as the rehearsal-based methods. However, on the concerns of data privacy and deployment efficiency, the applications of these methods in real-world scenarios are hampered. Therefore, we focus on a more challenging rehearsal-free CL (RFCL) [23, 25, 38, 50, 51], which does not store any previous samples or prototypes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Latest approaches [38,40,50,51] leverage prompt learning techniques to solve the RFCL problem. Due to freezing the large-scale pre-trained model, learning new knowledge (i.e. plasticity) becomes more challenging rather than reducing the forgetting (i.e. stability) [38]. To acquire new knowledge, prompt-based methods utilize key-query matching to obtain trainable prompts. As depicted in Figure 1(a), these methods first generate a feature representation (i.e. query) for each instance using the pre-trained model. Subsequently, they select or generate prompts for each instance based on the cosine similarity between the query and keys, followed by a second pass through the pre-trained model. Despite their promising progress, they are faced with two challenges. 1) Due to the substantial gap between the pre-trained dataset and the downstream datasets, the query embeddings directly generated by the pre-trained model are prone to confusion, leading to inaccurate key-query matching prompt selection. The inaccurate selection seriously hampers the learning ability of the model. 2) These methods require two forward passes, one for calculating the query and another for logit, leading to relatively low efficiency. Therefore, there are inherent shortcomings in prompt-based methods for RFCL tasks, motivating us to explore a novel approach distinct from prompt-learning techniques." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Continual Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Taking into account the inherent drawbacks mentioned above, in this paper, we shift our focus to adapter tuning [19] and propose a simple yet effective adapter variant, named ContinuaIAdapter (C-ADA) for rehearsal-free continual learning. Compared to prompt-learning techniques, C-ADA eliminates the necessity for key-query matching, fundamentally avoiding issues arising from erroneous key-query matches that could lead to conflicts in training parameters. As depicted in Figure 1 (c), this results in a substantial enhancement in the accuracy of C-ADA for new tasks, and C-ADA achieves this with just a single forward pass, greatly boosting training speed. Moreover, in contrast to lacking incremental capability adapter fine-tuning methods [5,19], we replace the adapter layer with a plug-and-play, parameter-expandable adapter layer, named Continuous Adapter Layer (CAL). The CAL expands the specific learnable weights in a new dimension (middle dimension) to acquire new knowledge and freeze the previously learned weights to retain old knowledge, which empowers the model with the capacity to tackle complex continual scenarios. This fundamentally encourages knowledge reuse, as the previous projection weights will contribute to future projection weights, facilitating the learning of new knowledge. We also introduce an orthogonal loss function to alleviate the conflict between new weights and previous weights, which significantly reduces the forgetting of old tasks. To further solve the problem of the gap between pre-trained datasets and downstream datasets, C-ADA includes scaling and shifting (S&S) module with negligible parameters to transfer the feature space from pre-trained datasets to downstream datasets before the incremental phase. Compared to the current SOTA methods, our model demonstrates significant improvements in both performance and training speed under equivalent parameter counts for RFCL. Furthermore, to validate its generality, we extend the method to the Domain-incremental Learning (DIL) setting and once again outperform the SOTA methods. The key contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.59, + 0.787, + 0.651 + ], + "angle": 0, + "content": "- We propose a simple yet effective approach called C-ADA, which eliminates the necessity for key-query matching, for RFCL, characterized by better plasticity and more efficient training compared to existing continual prompting approaches." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.654, + 0.787, + 0.699 + ], + "angle": 0, + "content": "- We design a novel CAL module, which comprises a parameter-extensible down-projection and up-projection. It facilitates the preservation of previously acquired knowledge and the learning of new knowledge." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.702, + 0.787, + 0.778 + ], + "angle": 0, + "content": "- We develop an S&S module to further reduce the divergence between the pre-training dataset and the downstream datasets, thereby enhancing the learning of new tasks. Additionally, we introduce an orthogonal loss to mitigate the interaction between old and new knowledge, which significantly contributes to the preservation of old knowledge." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We conduct extensive experiments on well-established rehearsal-free benchmarks and achieve a new SOTA performance. Meanwhile, we evaluate our approach on the DIL setting and again outperform the state of the art, highlighting the generality." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.59, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Gao et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.177, + 0.564, + 0.193 + ], + "angle": 0, + "content": "2.1Rehearsal-Based Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.201, + 0.788, + 0.413 + ], + "angle": 0, + "content": "[8, 11, 15, 39, 41, 45, 47–49, 53, 57, 58] retain prior information (old exemplars or prototypes) for subsequent tasks, aimed at mitigating catastrophic forgetting by replay them in the new tasks. Exemplars-based methods [8, 11, 15, 41, 45, 53] store a portion of old exemplars in a memory buffer. These exemplars are trained alongside the current samples to optimize the model. For instance, iCaRL [35] and its variants [18, 41, 53] prevent forgetting by employing the herding technique for exemplar selection and designing distinct distillation losses. However, exemplar-based methods not only require extra memory but also experience a significant performance decline as the buffer size decreases. Exemplars-free methods [57, 58] typically store a prototype per class to refine the classifier. For example, [57] memorizes and augments prototypes to maintain the decision boundary. Some latest methods SLCA [55], RanPAC [29], HiDe-Prompt [46] combine the pretrained model and prototypes to achieve better performance. Since they are not rehearsal-free continual learning, we have excluded comparisons with them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.787, + 0.475 + ], + "angle": 0, + "content": "Nevertheless, preserving previous information may introduce privacy and data leakage risks, often impractical in real-world applications. This motivates us to concentrate on the rehearsal-free setting, which offers enhanced privacy protection and real-world application." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.496, + 0.55, + 0.512 + ], + "angle": 0, + "content": "2.2Rehearsal-FreeContinual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Early rehearsal-free methods [1, 25, 54] utilize regularization to alleviate model forgetting. While these approaches demonstrate strong performance in task-incremental continual learning, their effectiveness diminishes in the more challenging class-incremental learning settings. Generative-based methods [6, 14, 37] employ deep model inversion to generate images from previous tasks. While these methods perform well in some scenarios, inversion is hampered by slow training speed and high computational costs. Recent studies [16, 38, 40, 50, 51, 56] have incorporated large pre-trained models and prompt learning into the realm of continual learning. Large-scale pre-trained models have demonstrated remarkable generalization performance and robust anti-forgetting capabilities. Therefore, prompt-based methods serve as our primary comparison benchmarks and they will be subsequently elaborated upon in detail." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.547, + 0.74 + ], + "angle": 0, + "content": "2.3 Prompting for Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Prompt learning has garnered widespread success in both natural language processing (NLP) [7, 20, 33] and computer vision (CV) [21, 32, 34, 42]. It leverages prompts to fine-tune large-scale pre-trained models for downstream tasks. Firstly, L2P [51] pioneers the integration of prompt learning into the field of continual learning field. It proposes a key-query matching strategy for prompt selection from the prompt pool, followed by the insertion of these prompts into" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Continual Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "the pre-trained model. The subsequent works are based on this key-query matching strategy. DualPrompt [50] employs a set of task-specific prompts and selects prompts based on the closest key-query matching during inference. Furthermore, CODA [38] utilizes key-query matching to assign weights to prompts and can be optimized in an end-to-end fashion. APG [40] proposes a learnable adaptive prompt generator eases the reliance on intensive pretraining. S-Prompts [49] focuses on domain-incremental learning by acquiring distinct prompts for each domain and utilizing KNN to select prompts during inference. In this paper, we thoroughly compared these methods to CIL and DIL protocols and extensively discussed their shortcomings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.32, + 0.341, + 0.336 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.35, + 0.403, + 0.366 + ], + "angle": 0, + "content": "3.1 Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.479 + ], + "angle": 0, + "content": "The continual learning aims to enable the model to learn non-stationary data from sequential tasks while preserving the knowledge obtained from previous tasks. We define a set of tasks \\(\\mathcal{D}_1, \\mathcal{D}_2, \\dots, \\mathcal{D}_T\\), where \\(\\mathcal{D}_t = \\{(x_i^t, y_i^t)\\}_{i=1}^{N_i}\\) is the \\(t\\)-th task with \\(N_i\\) samples, and \\(x_i^t\\) and \\(y_i^t\\) is the \\(i\\)-th images and its label. The label sets of \\(\\mathcal{D}_t\\) are defined as \\(\\mathcal{C}_t\\). The objective of rehearsal-free continual learning is to employ a unified model for classifying test images from all previously learned tasks while not storing any old samples or prototypes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.789, + 0.617 + ], + "angle": 0, + "content": "Depending on the specific problems, continual learning can be divided into multiple settings, which differ in challenge and practicality. In this paper, we mainly focus on addressing the challenging class-incremental learning setting. In this scenario, the class of each task is non-overlapping, denoted as \\(\\mathcal{C}_1 \\cap \\mathcal{C}_2 \\cdots \\cap \\mathcal{C}_T = \\emptyset\\) and the task identity can not be obtained in the inference phase. Furthermore, to assess the robustness and generalization of our approach, we extend our experiments to domain-incremental continual learning setting. In this scenario, each task has the same classes but exhibits distinct data distributions, expressed as \\(\\mathcal{C}_1 = \\mathcal{C}_2 = \\dots = \\mathcal{C}_T\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.498, + 0.65 + ], + "angle": 0, + "content": "3.2 Framework and Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Figure 2 illustrates the framework of the Continual Adapter (C-ADA) approach. The C-ADA incorporates the S&S module and the plug-and-play Continual Adapter Layer (CAL) into the pre-trained model, endowing it with robust capabilities for addressing continual learning challenges. Concretely, the S&S module is exclusively trained before the incremental training to mitigate the domain gap and is subsequently frozen during further training. The CAL comprises a parameter-extensible down-projection, a ReLU activation, and a parameter-extensible up-projection. During incremental learning, the down-projection and up-projection expand the trainable weights to learn the new knowledge and freeze the previous weights to maintain the old knowledge. Moreover, we propose an orthogonal loss to alleviate the conflict between previous weights and trainable weights in each CAL." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Gao et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.143, + 0.761, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 2: The framework of our C-ADA approach. For simplicity, we omit the skip connection and Layernorm in the figure. or loss represents the orthogonal loss. We attach our S&S and CAL, which are in parallel with the projection layer and MLP, to the shallow N layers of ViT and freeze the pre-trained backbone. For each new task, we expand two trainable weights in the CAL to learn the new knowledge and freeze the previous weights. Different from prior works, our approach uses a novel adapter variant to eliminate the necessity for key-query matching. Only trainable weights and classifier parameters are optimized which is parameter efficient and no old information (old images or prototypes) are stored which is privacy preserving." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.639 + ], + "angle": 0, + "content": "The basic backbone consists of a pre-trained vision transformer feature extractor \\( f = f(\\cdot ;\\theta) \\) and a classifier \\( g = g(\\cdot ;\\phi) \\). Given an input image \\( x \\in \\mathbb{R}^{H\\times W\\times C} \\) from \\( \\mathcal{D}_t \\). The image is reshaped and projected by the patch embedding layer, resulting in the embedding feature \\( x_{e} \\in \\mathbb{R}^{L\\times D} \\), where \\( L \\) is the token length, and \\( D \\) is the embedding dimension. These embedding features are not only input to the pre-training module within the ViT block but also to the parallel CAL and S&S modules. The output of the backbone and parallel is added and fed into the next layer. Finally, the output features \\( x_{p} \\) from the last ViT block are fed into the \\( g \\) to perform classification tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.668, + 0.456, + 0.683 + ], + "angle": 0, + "content": "3.3 Scale and Shift Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.763 + ], + "angle": 0, + "content": "Due to the significant domain gap between the pre-training dataset and the downstream dataset, our C-ADA approach introduces a flexible and independent module, named scaling and shifting (S&S), to revise the feature space before incremental training for more efficient fine-tuning (higher plasticity)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.764, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Concretely, the S&S module includes the trainable parameter \\(\\alpha \\in \\mathbb{R}^D\\) and \\(\\beta \\in \\mathbb{R}^D\\) as the scale and shift factors, respectively, to learn the domain information of downstream tasks. The S&S operation is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.448, + 0.826, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\ny = \\alpha \\odot x + \\beta \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Continual Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "where \\( \\odot \\) is the dot product. We employ the S&S operation before the incremental training by using the data from the first phase (i.e. \\( \\mathcal{D}_1 \\)). Then we freeze the trainable parameters during the incremental phases." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.214, + 0.468, + 0.23 + ], + "angle": 0, + "content": "3.4 Continual Adapter Layer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.789, + 0.481 + ], + "angle": 0, + "content": "Structure of Continual Adapter Layer To address the challenges of efficiency and lightweight, the CAL is designed as a bottleneck module. It consists of a parameter-extensible down-projection, a ReLU layer for non-linear property, and a parameter-extensible up-projection. The down-projection embeds the old and new knowledge into a latent embedding along with the ReLU function, and the up-projection extracts a dimension-fixed knowledge representation from the latent embedding. For each new task, we first initialize and append two learnable linear weights to the down-projection and up-projection, respectively, and freeze the previous weights. Such design ensures that the CAL is extendable for learning new knowledge and stable in maintaining old knowledge, maintaining the input and output feature dimensions unchanged. Assuming the current task \\( t \\), we assign the weights in the CAL into two types: the frozen weights \\( \\mathbf{W}^{1},\\dots,\\mathbf{W}^{(t - 1)} \\), and the trainable weight \\( \\mathbf{W}^t \\) (\\( \\mathbf{W}^i \\) represents both the \\( \\mathbf{W}_{dp}^i \\) and \\( \\mathbf{W}_{up}^i \\)). The frozen weights effectively preserve the previous knowledge to mitigate catastrophic forgetting, while simultaneously optimizing the trainable parameters to learn new knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.789, + 0.525 + ], + "angle": 0, + "content": "Without loss of generality, we define the input and output of the current layer as \\( x_{i} \\) and \\( x_{i}^{\\prime} \\), respectively. In task t, we can adapt the input by the CAL as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.525, + 0.608, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {d p} = \\left[ \\begin{array}{c} \\mathbf {W} _ {d p} ^ {1}, \\dots , \\mathbf {W} _ {d p} ^ {t - 1}, \\mathbf {W} _ {d p} ^ {t} \\end{array} \\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.546, + 0.785, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {W} _ {u p} = \\left[ \\mathbf {W} _ {u p} ^ {1}, \\dots , \\mathbf {W} _ {u p} ^ {t - 1}, \\mathbf {W} _ {u p} ^ {t} \\right], \\\\ \\mathbf {W} _ {d p} ^ {i} \\in \\mathbb {R} ^ {D \\times d} \\quad \\mathbf {W} _ {u p} ^ {i} \\in \\mathbb {R} ^ {d \\times D} \\end{array} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.589, + 0.591, + 0.606 + ], + "angle": 0, + "content": "\\[\nx _ {i} ^ {\\prime} = \\operatorname {R e L U} (x _ {i} \\cdot \\mathbf {W} _ {d p}) \\cdot \\mathbf {W} _ {u p}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.613, + 0.787, + 0.649 + ], + "angle": 0, + "content": "In the next task, we freeze the \\(\\mathbf{W}_{dp}^{t}\\) and \\(\\mathbf{W}_{up}^{t}\\) to maintain the old knowledge and convert two new weights \\(\\mathbf{W}_{dp}^{t + 1}\\) and \\(\\mathbf{W}_{up}^{t + 1}\\) to the trainable weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Parameters of Continual Adapter Layer. In this paper, we follow the standard RFCL setting where the total task number is known before training. To compare fairly, we adjust the value of \\( d \\) to maintain a consistent number of parameters across different settings. For real-world applications, we can keep the size of \\( d \\) constant and continually expand the specific weights for new tasks. Remarkably, even when \\( d \\) is set to a minimal value, such as \\( d = 1 \\), it still exhibits commendable performance. Therefore C-ADA is lightweight, and memory-efficient in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Position of Continual Adapter Layer. As depicted in [9,43], the Multi-Head Self-Attention (MHSA) block employs the attention mechanism to capture global" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.297 + ], + "angle": 0, + "content": "information, while the MLP block prevents the ViT output from degeneration. Both components play pivotal roles in the ViT structure. Consequently, we investigate performing a comprehensive tuning of both the MHSA block and MLP block, instead of exclusively emphasizing one over the other. To fine-tune the MHSA block, we choose to insert the CAL in parallel with the Projection Layer instead of the Attention module. This decision is motivated by the observation that, despite various advanced transformer-based models using different attention mechanisms within the MHSA block [12, 26, 28], they consistently include a projection layer. On the contrary, we insert the CAL in parallel with the entire MLP block to fine-tune the MLP block." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.343 + ], + "angle": 0, + "content": "Without loss of generality, we define the input as \\( x_{l} \\) and initially feed it into the Attention module to obtain \\( x_{a} \\), which serves as the input to the Projection Layer and \\( \\mathrm{CAL}_1 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.351, + 0.785, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} x _ {a} = \\operatorname {A t t e n} \\left(x _ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q \\cdot K ^ {T}}{\\sqrt {d}}\\right) V, \\\\ x _ {l} = \\operatorname {P r e s i} \\left(x _ {l}\\right) = x _ {l} = \\lambda_ {1} \\operatorname {C A L} \\left(S _ {l} - S \\left(x _ {l}\\right)\\right) \\end{array} \\tag {3} \\\\ x _ {l 1} = \\operatorname {P r o j} (x _ {a}), \\quad x _ {l 2} = \\lambda \\cdot \\operatorname {C A L} _ {1} (S \\& S (x _ {a})), \\\\ x _ {l} ^ {\\prime} = x _ {l 1} + x _ {l 2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.786, + 0.462 + ], + "angle": 0, + "content": "Then the output \\( x_{l}^{\\prime} \\) is further sent to the MLP block and \\( \\mathrm{CAL}_2 \\). This process is formally formulated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.47, + 0.785, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {l 1} = \\operatorname {M L P} \\left(x _ {l} ^ {\\prime}\\right), \\quad x _ {l 2} = \\lambda \\cdot \\operatorname {C A L} _ {2} (S \\& S \\left(x _ {l} ^ {\\prime}\\right)), \\tag {4} \\\\ x _ {l + 1} = x _ {l 1} + x _ {l 2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.572 + ], + "angle": 0, + "content": "where \\(\\mathrm{CAL}_1\\) and \\(\\mathrm{CAL}_2\\) are in parallel with the projection layer and MLP block, respectively. Proj represents the Projection Layer and \\(\\lambda\\) represents a scale factor (typically set to 0.1). Then, the output \\(x_{l + 1}\\) is sent to the next layer and repeats the above process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.379, + 0.605 + ], + "angle": 0, + "content": "3.5 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Orthogonal Loss. When training the new task, the learned new knowledge will interfere with old knowledge, potentially resulting in catastrophic forgetting. Thus we propose to solve this problem by initialization and loss function. We apply the Gram-Schmidt process to initialize the trainable parameters at the start of each new task. Orthogonal initialization contributes to the diminishment of the scope when the model parameter is updated, thereby fostering enhanced stability and overall performance. After the first task, we employ the orthogonal loss to keep trainable weights orthogonal to previous weights." + }, + { + "type": "equation", + "bbox": [ + 0.262, + 0.739, + 0.785, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o r} = \\left\\| \\mathbf {W} _ {d p} ^ {t} \\right. ^ {T} \\cdot \\left\\{\\mathbf {W} _ {d p} ^ {1}, \\dots , \\mathbf {W} _ {d p} ^ {t - 1} \\right\\} \\| _ {2} + \\left\\| \\mathbf {W} _ {u p} ^ {t} \\right. \\cdot \\left\\{\\mathbf {W} _ {u p} ^ {1}, \\dots , \\mathbf {W} _ {u p} ^ {t - 1} \\right\\} ^ {T} \\| _ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_2 \\) represent the \\( L_{2} \\)-norm and \\( T \\) represents the matrix transpose. The loss function forces the model to optimize current weight \\( \\mathbf{W}^t \\) to be orthogonal to previous weight \\( \\mathbf{W}^1, \\dots, \\mathbf{W}^{t-1} \\). Intuitively, orthogonal vectors exhibit reduced mutual influence. As a result, this design aims to minimize interference between new and old knowledge to reduce catastrophic forgetting." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Continual Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Classification Loss. Similar to the previous methods, we use the CE loss to optimize the model to learn the new knowledge" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.189, + 0.786, + 0.206 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = \\mathcal {L} (g (f (\\boldsymbol {x}; \\mathbf {W} _ {d p}, \\mathbf {W} _ {u p}, \\theta); \\phi), y) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.216, + 0.784, + 0.247 + ], + "angle": 0, + "content": "where \\( x \\) is the input image and \\( y \\) is the corresponding label. \\( \\phi \\) is the parameters of the classifier and \\( \\theta \\) is the parameters of the pre-train model." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.268, + 0.787, + 0.299 + ], + "angle": 0, + "content": "Total Optimization Objective. For task \\( t \\), our full optimization consists of a classification loss \\( \\mathcal{L}_{ce} \\) and an orthogonal loss \\( \\mathcal{L}_{or} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.31, + 0.786, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {W} _ {\\mathrm {d p}} ^ {t}, \\mathbf {W} _ {\\mathrm {u p}} ^ {t}, \\phi} \\mathcal {L} _ {c e} + \\delta \\mathcal {L} _ {\\mathrm {o r}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.347, + 0.784, + 0.392 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{dp}^{t},\\mathbf{W}_{up}^{t}\\) and \\(\\phi\\) are the learnable parameters of task t. We set the \\(\\delta\\) to 1 for all the experiments. During each training phase, our models can be optimized in an end-to-end fashion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.415, + 0.376, + 0.433 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.447, + 0.784, + 0.492 + ], + "angle": 0, + "content": "In this paper, we mainly aim to solve the rehearsal-free class-incremental learning problem. To evaluate our methods, we follow the RFCL settings proposed in previous works [38, 50, 51] and conduct comprehensive experiments." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.492, + 0.784, + 0.553 + ], + "angle": 0, + "content": "Besides, to prove the robustness and generality of our method on other continual learning settings, we conduct experiments compared with the SOTA methods [49] on the domain-incremental learning settings. Extensive ablation studies prove the effectiveness of our methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.457, + 0.589 + ], + "angle": 0, + "content": "4.1 Evaluation Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.598, + 0.784, + 0.643 + ], + "angle": 0, + "content": "Datasets. We use the Split CIFAR-100 [24] and Split ImageNet-R [17] for class-incremental setting, CORe50 [27] and DomainNet [30] for domain-incremental setting." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.644, + 0.784, + 0.689 + ], + "angle": 0, + "content": "CIFAR-100 is a widely-used benchmark in continual learning, which contains 60000 images of \\(32 \\times 32\\) size from 100 classes. We split CIFAR-100 into the 10-task benchmark following previous works." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.689, + 0.784, + 0.749 + ], + "angle": 0, + "content": "ImageNet-R has 200 classes and includes newly collected data of different styles. It provides a fair and challenging problem setting due to the significant gap between training data and pre-trained data. We split the dataset into the 5-task benchmark, the original 10-task benchmark and the 20-task benchmark." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.784, + 0.795 + ], + "angle": 0, + "content": "CORe50 has 50 classes from 11 distinct domains. Following the domain-incremental setting, we use 8 domains for (120000 images) for continual training and the rest of the domains for testing." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "DomainNet is a large dataset with 345 classes and 600000 images. These images are split into 6 domains (Clipart, Infograph, Paint, Quickdraw, Real, and Sketch) for continual learning like S-Prompts [49]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Gao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Results (\\%) on ImageNet-R and CIFAR-100. The results are all obtained by CODA [38] directly. P represents the total number of tasks. UB represents the full fine-tuning result from [10]. We report \\( A_N(\\%) \\), \\( Param(\\%) \\) and \\( FLOPs(\\mathrm{G}) \\) to bring intuitive comparison. C-ADA uses fewer parameters and doubles the training speed, greatly surpassing other methods." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.227, + 0.768, + 0.378 + ], + "angle": 0, + "content": "
MethodsImageNet-RCIFAR-100FLOPs
P=5P=10P=20P=10
AN(↑)Param(↓)AN(↑)Param(↓)AN(↑)Param(↓)AN(↑)Param(↓)
UB77.13100/10077.13100/10077.13100/10089.30100/100-
ER (5000)71.72100/10064.43100/10052.43100/10076.20100/100-
FT18.74100/10010.12100/1004.75100/1009.92100/100-
FT++60.42100/10048.93100/10035.98100/10049.91100/100-
LwF.MC74.56100/10066.73100/10054.05100/10064.83100/100-
L2P70.830.7/100.769.290.7/100.765.890.7/100.782.500.7/100.735.2
Deep L2P73.939.6/109.671.669.6/109.668.429.6/109.684.309.6/109.635.2
DualPrompt73.050.5/100.571.320.8/100.867.871.3/101.383.050.8/100.835.2
CODA-P-S75.190.7/100.773.930.7/100.770.530.7/100.784.590.6/100.635.2
C-ADA77.930.7/100.776.660.7/100.773.470.7/100.787.180.6/100.617.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Comparing Methods. We compare our method with several baselines and state-of-the-art (SOTA) continual learning methods. For a fair comparison, all the methods are based on a ViT-B/16 backbone pre-trained on ImageNet-1K. Under class-incremental settings, we regard L2P [51], DualPrompt [50], and CODA [38] as the main competitors, which are rehearsal-free methods. Under domain-incremental settings, despite S-Prompt [49] storing the prototypes of previous domains instead of the rehearsal-free method, it is still the main comparison method as the SOTA of domain-incremental learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.726 + ], + "angle": 0, + "content": "Implementation Details. Similar to CODA [38], we use the Adam [22] optimizer with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\), and a batch size of 128 images in four GPUs. For every image, we resize it to \\(224\\times 224\\) and normalize them to [0,1]. The learning rate is set to \\(5e^{-5}\\) for all the settings with cosine-decaying. For fair comparisons, we set the total middle dimension to 60 (sum of two CALs) and insert the CAL into layers 0-4, which achieves a similar parameter amount to previous works [38, 50]. We train the CIFAR-100 for 20 epochs, ImageNet-R for 50 epochs, CORe50 and DomainNet for 10 epoch. During the training phase, we substitute predictions from previous-task logits with negative infinity during the training of a new task. This design leads to a softmax prediction of \"0\" for these previous task classes and ensures that gradients do not affect the linear heads of past task classes like previous works. [38, 50, 51]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.744, + 0.373, + 0.758 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.787, + 0.798 + ], + "angle": 0, + "content": "Class-incremental learning. We report the results on two class-incremental datasets in Table 1 following the previous work [38]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.84 + ], + "angle": 0, + "content": "3 For the sake of fairness, we primarily selected CODA-P-S with the same parameter count as C-ADA for comparison." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Continual Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Results (\\%) on CORe50 and DomainNet. \\( A_N \\) gives the accuracy averaged over tasks and buffer size represents the number of old exemplars stored in the memory buffer. The results are all obtained by S-Prompt [49] directly. We reproduce the CODA on these settings for further comparison. We ensure that all methods have the same pre-trained model (ViT-B/16)." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.227, + 0.46, + 0.432 + ], + "angle": 0, + "content": "
MethodBuffer size\\(A_N(↑)\\)
Upper-Bound-84.01
ER80.10
GDumb74.92
BiC79.28
DER++50/class79.70
Co2L79.75
DyTox79.21
L2P81.07
EWC74.82
LwF75.45
L2P0/class78.33
CODA-P-S85.41
S-Prompts83.13
C-ADA89.45
" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.239, + 0.789, + 0.418 + ], + "angle": 0, + "content": "
MethodBuffer sizeAN(↑)
Upper-Bound-63.22
DyTox50/class62.94
EWC47.62
LwF49.19
SimCLR44.20
BYOL49.70
Barlow Twins0/class48.90
Supervised Contrastive50.90
L2P40.15
CODA-P-S47.56
S-Prompts50.62
C-ADA53.00
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.608 + ], + "angle": 0, + "content": "ImageNet-R: We report the results with 5-tasks, 10-tasks, and 20-tasks to evaluate the robustness of the model under different settings in Table 1. It is worth noting that our method significantly surpasses the previous methods in average accuracy under different settings with as much as \\(+2.74\\%\\), \\(+2.73\\%\\) and \\(+2.94\\%\\) over CODA, \\(+4.88\\%\\), \\(+5.41\\%\\) and \\(+5.60\\%\\) over DualPrompt (second best method). This phenomenon shows that our method is robust to different class increment settings. Moreover, since our method only requires forward propagation one time. We achieved an obvious increase in training speed while improving accuracy. Combining accuracy and training speed brings better practicality in the real world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.655 + ], + "angle": 0, + "content": "CIFAR-100: We note that our approach also brings similar results to the ImageNet-R benchmarks, with an improvement of \\(+2.59\\%\\) compared with CODA, \\(+4.13\\%\\) compared with DualPrompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.809, + 0.719 + ], + "angle": 0, + "content": "Domain-incremental learning. We report the results on two domain-incremental datasets in Table 2. Besides, we reproduced CODA on these datasets to bring intuitive comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "CORe50 and DomainNet are widely used domain dataset for continual learning. Following the setting of S-Prompt [49], we divide the baseline methods into two groups, exemplars-based methods [2-4, 11, 31, 52] and exemplars-free methods [49]/rehearsal-free methods [38, 51]. Exemplars-based methods store 50 old exemplars per class in the memory buffer. These methods often have better accuracy but are accompanied by the risk of privacy leakage and memory overhead. Our method surpasses the CODA by \\(+4.04\\%\\) and \\(+5.44\\%\\) on CORe50 and DomainNet benchmark. Besides, it is worth noting that our method even out" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Gao et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.143, + 0.48, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.278, + 0.493, + 0.321 + ], + "angle": 0, + "content": "Fig. 3: Average accuracy \\( A_{N} \\) (\\%) vs tuning parameters (\\%). We report the results on the 20-tasks ImageNet-R." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.158, + 0.768, + 0.226 + ], + "angle": 0, + "content": "
MethodCIFAR-100ImageNet-R
C-ADA87.3276.66
Ablate S&S86.1175.93
Ablate Lor82.8972.17
Ablate CAL76.9364.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.236, + 0.788, + 0.307 + ], + "angle": 0, + "content": "Table 3: Ablation Results (%) on 10-task ImageNet-R and CIFAR-100. \\(A_N\\) gives the accuracy averaged over tasks. We ablate the key components in turn and report the results." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.35, + 0.788, + 0.412 + ], + "angle": 0, + "content": "performs S-Prompt, which is specifically designed for domain-incremental learning, with improvements of \\(+6.32\\%\\) and \\(+2.38\\%\\). This interesting phenomenon demonstrates the robustness and effectiveness of our method in various continual learning settings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.434, + 0.483, + 0.449 + ], + "angle": 0, + "content": "4.3 Key Components Ablation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.789, + 0.687 + ], + "angle": 0, + "content": "We conducted ablation experiments to analyze the importance of each component. We incrementally introduced key components on both the 10-tasks ImageNet-R and CIFAR-100 datasets, and the results are presented in Table 3. Firstly, we ablate the S&S module, the performance has dropped \\((0.73\\% \\downarrow\\) on ImageNet-R). This shows that revising the feature space before training is beneficial to improving subsequent learning ability. Secondly, removing the orthogonal loss leads to severe performance degradation on average accuracy \\((3.76\\% \\downarrow\\) on ImageNet-R). This outcome aligns with our expectations, as ablating the orthogonal loss results in our method lacking a mechanism to prevent forgetting, leading to severe catastrophic forgetting. Thirdly, when we ablated the CAL component, we observed a noticeable drop in performance \\((7.72\\% \\downarrow\\) on ImageNet-R). The removal of the CAL implies that our model relies solely on adjusting the classifier to tune the model, which obviously leads to unsatisfactory results, particularly when dealing with the downstream datasets having a significant gap from the pre-trained dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.506, + 0.725 + ], + "angle": 0, + "content": "4.4 Scaling Trainable Parameters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.765 + ], + "angle": 0, + "content": "We conduct experiments to discuss the relationship between the trainable parameters amount and the performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The trainable parameters amount can be adjusted by changing the middle dimension of the CAL. The total number of middle dimensions of the two CALs in each block is chosen from \\(\\{20, 40, 60, 80, 100, 120, 140, 160\\}\\), corresponding to \\(\\{0.35\\%, 0.54\\%, 0.71\\%, 0.90\\%, 1.12\\%, 1.27\\%, 1.45\\%, 1.61\\}\\) trainable parameters. We report the results on 20-tasks ImageNet-R in Figure 3. The figure" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Continual Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "shows that more trainable parameters contribute to achieving a higher accuracy. Moreover, we can notice that even with the fewer trainable parameters, our model consistently demonstrates superior performance when compared to CODA (71.92% \\( A_N \\), 0.35% \\( \\text{Param} \\) of C-ADA vs 70.53% \\( A_N \\), 0.7% \\( \\text{Param} \\) of CODA). This fully illustrates that our approach improves the plasticity compared with previous works." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.259, + 0.506, + 0.274 + ], + "angle": 0, + "content": "4.5 Attaching Layer and Position" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Intuitively, each layer of the ViT exhibits distinct feature extraction capabilities [50]. The impact of tuning different layers may be quite different in CL. Furthermore, each module of the layer plays different roles, which causes attaching CAL to various positions within the layer may yield divergent outcomes. Therefore, it is crucial to explore layers and positions to attach the CAL under the CL settings. Note that the total number of trainable parameters remains consistent across all experiments described below (middle \\(\\dim = d*t\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.44 + ], + "angle": 0, + "content": "Layer of CAL. For simplicity, we assume that the attaching layers are contiguous. We report the result in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Firstly, we attach the CAL to all the layers to get a preliminary result (Layer 0 to Layer 11). Next, we use binary search to analyze the importance of shallow (Layer 0 to Layer 5) and deep (Layer 6 to Layer 11) layers. We observe that attaching to shallow layers gets superior results compared to attaching to deeper layers. This phenomenon is consistent with the observation from previous work [50]. Then we continue to reduce the number of attached layers in shallow layers. Experimental results show that attaching from layer \\(= 0\\) to layer \\(= 4\\) performs the best. We use this conclusion to analyze the attached modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.643 + ], + "angle": 0, + "content": "Position of CAL. By analyzing common parts of the ViT architecture, we select three alternative attaching positions: QKV attention layer \\((W_{q}, W_{k}, W_{v})\\) in the MHSA block, projection layer in the MHSA block, and the MLP block. We report the result in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Firstly, we analyze the effects of only attaching these three positions, respectively. We observe that attaching the CAL to the projection layer has a better effect than in the QKV attention layer. This phenomenon makes us attach the CAL to the projection layer to tune the MHSA block. Then we opt to combine the tuning of different positions. We observe that attaching the CAL to different positions resulted in certain improvements (about \\(0.4\\%\\)). Interestingly, it seems to have little to do with how to assign the middle dimensions. This shows to some extent that our method is robust." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.473, + 0.8 + ], + "angle": 0, + "content": "4.6 Extra Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Although we emphasize that \\( A_N \\), \\( \\text{Param} \\), and \\( FLOPs \\) are the more important metrics, there are still some common metrics in the CL field." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Gao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.151, + 0.493, + 0.221 + ], + "angle": 0, + "content": "Table 4: Results \\((\\%)\\) on ImageNetR 10 tasks. We analyze the difference of attaching to different layers. We keep the total tunable parameters constant by modifying the middle dimension," + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.233, + 0.475, + 0.34 + ], + "angle": 0, + "content": "
Attach Layermiddle dimAN(↑)
Layer 0 ~ Layer 113075.23
Layer 6 ~ Layer 115072.51
Layer 0 ~ Layer 55076.21
Layer 1 ~ Layer 56076.29
Layer 0 ~ Layer 46076.66
Layer 2 ~ Layer 58075.72
Layer 0 ~ Layer 38075.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 5: Results \\((\\%)\\) on ImageNetR 10 tasks. We fix the attaching layers (layer 0 to layer 4) to analyze two aspects: the attaching positions and the middle dimension of each CAL." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.227, + 0.758, + 0.346 + ], + "angle": 0, + "content": "
Attach Positionmiddle dimAN(↑)
QKV Attention6076.04
MLP6076.23
Proj6076.25
Proj + MLP20 + 4076.52
Proj + MLP30 + 3076.57
Proj + MLP50 + 1076.51
Proj + MLP10 + 5076.61
Proj + MLP40 + 2076.66
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.784, + 0.4 + ], + "angle": 0, + "content": "\\(F_{N}\\): Average forgetting metrics, represent the drop in task performance averaged over \\(N\\) tasks." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.402, + 0.687, + 0.416 + ], + "angle": 0, + "content": "bwt: Backward transfer, assesses the quality of representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.447 + ], + "angle": 0, + "content": "fwt: Forward transfer, represents the influence of existing knowledge on the performance of subsequent concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.787, + 0.507 + ], + "angle": 0, + "content": "In Table 6, we can see that across various metrics, C-ADA demonstrates superior performance. Notably, in terms of \\( fwt \\), which represents the capacity to learn new knowledge, it significantly outperforms the other methods. This emphasizes the effectiveness of C-ADA in learning new knowledge." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.532, + 0.788, + 0.561 + ], + "angle": 0, + "content": "Table 6: Results (\\%) on ImageNet-R. The \\( F_N \\) results are all obtained by CODA [38] directly. We reproduce the bwt and fwt metrics for further comparison." + }, + { + "type": "table", + "bbox": [ + 0.257, + 0.572, + 0.744, + 0.658 + ], + "angle": 0, + "content": "
MethodsP=5P=10P=20P=10P=10
FN(↓)FN(↓)FN(↓)bwt(↑)fwt(↑)
L2P3.36 ± 0.182.03 ± 0.191.24 ± 0.14-3.69 ± 0.07-0.32 ± 0.04
DualPrompt2.64 ± 0.171.71 ± 0.241.07 ± 0.14-3.48 ± 0.06-0.09 ± 0.04
CODA2.65 ± 0.151.60 ± 0.201.00 ± 0.15-3.24 ± 0.060.01 ± 0.03
C-ADA2.55 ± 0.141.58 ± 0.181.02 ± 0.12-3.12 ± 0.050.34 ± 0.03
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.705, + 0.36, + 0.721 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.825, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present a novel adapter variant, named Continual Adapter (C-ADA), for RFCL. Our approach assembles a parameter-extensible adapter layer (CAL) and an S&S module in parallel with the pre-trained model. Importantly, C-ADA eliminates the necessity for key-query matching to achieve better plasticity and efficient training. We set a new SOTA on well-established benchmarks of RFCL and outperform the mainstream methods of DIL. This underscores the generality and robustness of our approach across various continual learning settings." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Continual Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.237 + ], + "angle": 0, + "content": "This work was funded by the National Natural Science Foundation of China under Grant No.U21B2048 and No.62302382 and Shenzhen Key Technical Projects under Grant CJGJZD2022051714160501 and the CAAI-MindSpore Open Fund, developed on OpenI Community." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV). pp. 139-154 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.372 + ], + "angle": 0, + "content": "2. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems 33, 15920-15930 (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.412 + ], + "angle": 0, + "content": "3. Cha, H., Lee, J., Shin, J.: Co2l: Contrastive continual learning. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 9516-9525 (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.455 + ], + "angle": 0, + "content": "4. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.785, + 0.496 + ], + "angle": 0, + "content": "5. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.537 + ], + "angle": 0, + "content": "6. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.785, + 0.578 + ], + "angle": 0, + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.62 + ], + "angle": 0, + "content": "8. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.661 + ], + "angle": 0, + "content": "9. Dong, Y., Cordonnier, J.B., Loukas, A.: Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In: International Conference on Machine Learning. pp. 2793-2803. PMLR (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.662, + 0.785, + 0.717 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.785, + 0.77 + ], + "angle": 0, + "content": "11. Douillard, A., Ramé, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9285-9295 (2022) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "12. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999) 2" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.288, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Gao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.243 + ], + "angle": 0, + "content": "15. Gao, X., He, Y., Dong, S., Cheng, J., Wei, X., Gong, Y.: Dkt: Diverse knowledge transfer transformer for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24236-24245 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.285 + ], + "angle": 0, + "content": "16. Gao, Z., Cen, J., Chang, X.: Consistent prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28463-28473 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.339 + ], + "angle": 0, + "content": "17. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.38 + ], + "angle": 0, + "content": "18. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 831-839 (2019) 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.38, + 0.785, + 0.433 + ], + "angle": 0, + "content": "19. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.475 + ], + "angle": 0, + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.475, + 0.785, + 0.516 + ], + "angle": 0, + "content": "21. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.542 + ], + "angle": 0, + "content": "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.597 + ], + "angle": 0, + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.624 + ], + "angle": 0, + "content": "24. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., Citeseer (2009) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.65 + ], + "angle": 0, + "content": "25. Li, Z., Hoiem, D.: Learning without forgetting. T-PAMI 40(12), 2935-2947 (2018) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.692 + ], + "angle": 0, + "content": "26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.785, + 0.718 + ], + "angle": 0, + "content": "27. Lomonaco, V., Maltoni, D.: Core50: a new dataset and benchmark for continuous object recognition. In: Conference on robot learning. pp. 17-26. PMLR (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "28. Mao, X., Qi, G., Chen, Y., Li, X., Duan, R., Ye, S., He, Y., Xue, H.: Towards robust vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12042-12051 (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "29. McDonnell, M.D., Gong, D., Parvaneh, A., Abbasnejad, E., van den Hengel, A.: Ranpac: Random projections and pre-trained models for continual learning. Advances in Neural Information Processing Systems 36 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 9" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Continual Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "31. Prabhu, A., Torr, P.H., Dokania, P.K.: Gdumb: A simple approach that questions our progress in continual learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 524-540. Springer (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.258 + ], + "angle": 0, + "content": "32. Qin, Y., Chen, Y., Peng, D., Peng, X., Zhou, J.T., Hu, P.: Noisy-correspondence learning for text-to-image person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27197-27206 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.298 + ], + "angle": 0, + "content": "33. Qin, Y., Peng, D., Peng, X., Wang, X., Hu, P.: Deep evidential learning with noisy correspondence for cross-modal retrieval. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4948-4956 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.339 + ], + "angle": 0, + "content": "34. Qin, Y., Sun, Y., Peng, D., Zhou, J.T., Peng, X., Hu, P.: Cross-modal active complementary learning with self-refining correspondence. Advances in Neural Information Processing Systems 36 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.367 + ], + "angle": 0, + "content": "35. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.407 + ], + "angle": 0, + "content": "36. Shieh, J.L., Haq, Q.M.u., Haq, M.A., Karam, S., Chondro, P., Gao, D.Q., Ruan, S.J.: Continual learning strategy in one-stage object detection framework based on experience replay for autonomous driving vehicle. Sensors 20(23), 6777 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.46 + ], + "angle": 0, + "content": "37. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.502 + ], + "angle": 0, + "content": "38. Smith, J.S., Karlinsky, L., et al.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: CVPR. pp. 11909-11919 (2023) 1, 2, 4, 5, 9, 10, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.543 + ], + "angle": 0, + "content": "39. Song, X., He, Y., Dong, S., Gong, Y.: Non-exemplar domain incremental object detection via learning domain bias. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 15056-15065 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.569 + ], + "angle": 0, + "content": "40. Tang, Y.M., et al.: When prompt-based incremental learning does not meet strong pretraining. In: CVPR (2023) 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.61 + ], + "angle": 0, + "content": "41. Tao, X., Chang, X., Hong, X., Wei, X., Gong, Y.: Topology-preserving class-incremental learning. In: European Conference on Computer Vision. pp. 254-270. Springer (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.664 + ], + "angle": 0, + "content": "42. Tu, C.H., Mai, Z., Chao, W.L.: Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7725-7735 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.705 + ], + "angle": 0, + "content": "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.746 + ], + "angle": 0, + "content": "44. Verwimp, E., Yang, K., Parisot, S., Hong, L., McDonagh, S., Pérez-Pellitero, E., De Lange, M., Tuytelaars, T.: Clad: A realistic continual learning benchmark for autonomous driving. Neural Networks 161, 659-669 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.746, + 0.788, + 0.773 + ], + "angle": 0, + "content": "45. Wang, F.Y., Zhou, D.W., Ye, H.J., Zhan, D.C.: Foster: Feature boosting and compression for class-incremental learning. arXiv preprint arXiv:2204.04662 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.799 + ], + "angle": 0, + "content": "46. Wang, L., et al.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. NeurIPS (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "47. Wang, S., Shi, W., Dong, S., Gao, X., Song, X., Gong, Y.: Semantic knowledge guided class-incremental learning. IEEE Transactions on Circuits and Systems for Video Technology 33(10), 5921-5931 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Gao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Wang, S., Shi, W., He, Y., Yu, Y., Gong, Y.: Non-exemplar class-incremental learning via adaptive old class reconstruction. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4524-4534 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "49. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022) 4, 5, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "50. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799 (2022) 2, 4, 5, 9, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "51. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022) 1, 2, 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "52. Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., Fu, Y.: Large scale incremental learning. arXiv preprint arXiv:1905.13260 (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "53. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3014-3023 (2021) 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "54. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International conference on machine learning. pp. 3987-3995. PMLR (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "55. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "56. Zhou, D.W., et al.: Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. arXiv (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.548 + ], + "angle": 0, + "content": "57. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5871-5880 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.785, + 0.604 + ], + "angle": 0, + "content": "58. Zhu, K., Zhai, W., Cao, Y., Luo, J., Zha, Z.J.: Self-sustaining representation expansion for non-exemplar class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9296-9305 (2022) 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.604 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_origin.pdf b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..338a7b51a3cbd7dd76086a3bb65e9c65702bd0cf --- /dev/null +++ b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/3d9bc515-5343-4ca9-abfd-bfccdf4ba905_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b6541d7eb6422b2d36ba68a3dc73d8c18ba1228ee8a22fff4554ee2d7a5af0a +size 1451932 diff --git a/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/full.md b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..141ec9d175a8192b61b4d4df9b3eecac71325a9e --- /dev/null +++ b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/full.md @@ -0,0 +1,322 @@ +# Beyond Prompt Learning: Continual Adapter for Efficient Rehearsal-Free Continual Learning + +Xinyuan Gao $^{1\#}$ , Songlin Dong $^{2\#}$ , Yuhang He $^{2\star}$ , Qiang Wang $^{1}$ , and Yihong Gong $^{1,2}$ + +$^{1}$ School of Software Engineering, Xi'an Jiaotong University +$^{2}$ College of Artificial Intelligence, Xi'an Jiaotong University {gxy010317, dsl97273141, qwang}@stu.xjtu.edu.cn, heyuhang@xjtu.edu.cn ygong@mail.xjtu.edu.cn + +Abstract. The problem of Rehearsal-Free Continual Learning (RFCL) aims to continually learn new knowledge while preventing forgetting of the old knowledge, without storing any old samples and prototypes. The latest methods leverage large-scale pre-trained models as the backbone and use key-query matching to generate trainable prompts to learn new knowledge. However, the domain gap between the pre-training dataset and the downstream datasets can easily lead to inaccuracies in key-query matching prompt selection when directly generating queries using the pre-trained model, which hampers learning new knowledge. Thus, in this paper, we propose a beyond prompt learning approach to the RFCL task, called Continual Adapter (C-ADA). It mainly comprises a parameter-extensible continual adapter layer (CAL) and a scaling and shifting (S&S) module in parallel with the pre-trained model. C-ADA flexibly extends specific weights in CAL to learn new knowledge for each task and freezes old weights to preserve prior knowledge, thereby avoiding matching errors and operational inefficiencies introduced by key-query matching. To reduce the gap, C-ADA employs an S&S module to transfer the feature space from pre-trained datasets to downstream datasets. Moreover, we propose an orthogonal loss to mitigate the interaction between old and new knowledge. Our approach achieves significantly improved performance and training speed, outperforming the current state-of-the-art (SOTA) method. Additionally, we conduct experiments on domain-incremental learning, surpassing the SOTA, and demonstrating the generality of our approach in different settings. + +Keywords: Continual Learning, Rehearsal-Free + +# 1 Introduction + +In recent years, with the emerging applications of deploying DNN models to dynamic and continuous environments such as autonomous driving [36, 44]. Continual Learning (CL) [18, 35, 38, 51, 53] has drawn a rapidly increasing research + +(a): Prior Prompt-Based Approach +![](images/e8bf0cbaa9cdc6fc3f0fbbc7cea90d34e7dc484b8c28456748cd6a8f5f5ece7e.jpg) +Decision Boundary Feature Tuned Frozen S85 Scaling & Shifting First Pass-forward Second Pass-forward + +![](images/c9ac87f5cd12fe58cdb8832d1c8dc2c77510fc3b23e3f314ee8d843ffd36c651.jpg) +(b):Continual Adapter Approach + +![](images/36a73c6bb81b6f73b1ab85a59a75640dc6f32ba131889bd45bd7e98d1749203b.jpg) +(c): The Results $(\%)$ of each new task on 5-task ImageNet-R +Fig. 1: Prior prompt-based approaches pass through the pre-trained model to generate the query and employ key-query matching to select prompts, which are inserted into the pre-trained model again (each layer has a unique prompt). Continual Adapter approach (C-ADA) strategy eliminates the need for key-query matching by introducing the CAL and S&S, which brings significant improvement in learning the new knowledge. Moreover, C-ADA only needs to pass the pre-trained model once, highlighting the training speed. + +interest both in the literature and industry. Classical CL methods focus on addressing the problem of catastrophic forgetting [13], wherein the models suffer dramatic performance decreases on previous tasks once training on new data. The majority of research studies [8, 15, 18, 35, 41, 53, 57, 58] attempt to reduce forgetting by retaining a subset of previous image samples or prototypes and replaying them in new tasks, referred to as the rehearsal-based methods. However, on the concerns of data privacy and deployment efficiency, the applications of these methods in real-world scenarios are hampered. Therefore, we focus on a more challenging rehearsal-free CL (RFCL) [23, 25, 38, 50, 51], which does not store any previous samples or prototypes. + +Latest approaches [38,40,50,51] leverage prompt learning techniques to solve the RFCL problem. Due to freezing the large-scale pre-trained model, learning new knowledge (i.e. plasticity) becomes more challenging rather than reducing the forgetting (i.e. stability) [38]. To acquire new knowledge, prompt-based methods utilize key-query matching to obtain trainable prompts. As depicted in Figure 1(a), these methods first generate a feature representation (i.e. query) for each instance using the pre-trained model. Subsequently, they select or generate prompts for each instance based on the cosine similarity between the query and keys, followed by a second pass through the pre-trained model. Despite their promising progress, they are faced with two challenges. 1) Due to the substantial gap between the pre-trained dataset and the downstream datasets, the query embeddings directly generated by the pre-trained model are prone to confusion, leading to inaccurate key-query matching prompt selection. The inaccurate selection seriously hampers the learning ability of the model. 2) These methods require two forward passes, one for calculating the query and another for logit, leading to relatively low efficiency. Therefore, there are inherent shortcomings in prompt-based methods for RFCL tasks, motivating us to explore a novel approach distinct from prompt-learning techniques. + +Taking into account the inherent drawbacks mentioned above, in this paper, we shift our focus to adapter tuning [19] and propose a simple yet effective adapter variant, named ContinuaIAdapter (C-ADA) for rehearsal-free continual learning. Compared to prompt-learning techniques, C-ADA eliminates the necessity for key-query matching, fundamentally avoiding issues arising from erroneous key-query matches that could lead to conflicts in training parameters. As depicted in Figure 1 (c), this results in a substantial enhancement in the accuracy of C-ADA for new tasks, and C-ADA achieves this with just a single forward pass, greatly boosting training speed. Moreover, in contrast to lacking incremental capability adapter fine-tuning methods [5,19], we replace the adapter layer with a plug-and-play, parameter-expandable adapter layer, named Continuous Adapter Layer (CAL). The CAL expands the specific learnable weights in a new dimension (middle dimension) to acquire new knowledge and freeze the previously learned weights to retain old knowledge, which empowers the model with the capacity to tackle complex continual scenarios. This fundamentally encourages knowledge reuse, as the previous projection weights will contribute to future projection weights, facilitating the learning of new knowledge. We also introduce an orthogonal loss function to alleviate the conflict between new weights and previous weights, which significantly reduces the forgetting of old tasks. To further solve the problem of the gap between pre-trained datasets and downstream datasets, C-ADA includes scaling and shifting (S&S) module with negligible parameters to transfer the feature space from pre-trained datasets to downstream datasets before the incremental phase. Compared to the current SOTA methods, our model demonstrates significant improvements in both performance and training speed under equivalent parameter counts for RFCL. Furthermore, to validate its generality, we extend the method to the Domain-incremental Learning (DIL) setting and once again outperform the SOTA methods. The key contributions of this paper are summarized as follows: + +- We propose a simple yet effective approach called C-ADA, which eliminates the necessity for key-query matching, for RFCL, characterized by better plasticity and more efficient training compared to existing continual prompting approaches. +- We design a novel CAL module, which comprises a parameter-extensible down-projection and up-projection. It facilitates the preservation of previously acquired knowledge and the learning of new knowledge. +- We develop an S&S module to further reduce the divergence between the pre-training dataset and the downstream datasets, thereby enhancing the learning of new tasks. Additionally, we introduce an orthogonal loss to mitigate the interaction between old and new knowledge, which significantly contributes to the preservation of old knowledge. +- We conduct extensive experiments on well-established rehearsal-free benchmarks and achieve a new SOTA performance. Meanwhile, we evaluate our approach on the DIL setting and again outperform the state of the art, highlighting the generality. + +# 2 Related work + +# 2.1Rehearsal-Based Continual Learning + +[8, 11, 15, 39, 41, 45, 47–49, 53, 57, 58] retain prior information (old exemplars or prototypes) for subsequent tasks, aimed at mitigating catastrophic forgetting by replay them in the new tasks. Exemplars-based methods [8, 11, 15, 41, 45, 53] store a portion of old exemplars in a memory buffer. These exemplars are trained alongside the current samples to optimize the model. For instance, iCaRL [35] and its variants [18, 41, 53] prevent forgetting by employing the herding technique for exemplar selection and designing distinct distillation losses. However, exemplar-based methods not only require extra memory but also experience a significant performance decline as the buffer size decreases. Exemplars-free methods [57, 58] typically store a prototype per class to refine the classifier. For example, [57] memorizes and augments prototypes to maintain the decision boundary. Some latest methods SLCA [55], RanPAC [29], HiDe-Prompt [46] combine the pretrained model and prototypes to achieve better performance. Since they are not rehearsal-free continual learning, we have excluded comparisons with them. + +Nevertheless, preserving previous information may introduce privacy and data leakage risks, often impractical in real-world applications. This motivates us to concentrate on the rehearsal-free setting, which offers enhanced privacy protection and real-world application. + +# 2.2Rehearsal-FreeContinual Learning + +Early rehearsal-free methods [1, 25, 54] utilize regularization to alleviate model forgetting. While these approaches demonstrate strong performance in task-incremental continual learning, their effectiveness diminishes in the more challenging class-incremental learning settings. Generative-based methods [6, 14, 37] employ deep model inversion to generate images from previous tasks. While these methods perform well in some scenarios, inversion is hampered by slow training speed and high computational costs. Recent studies [16, 38, 40, 50, 51, 56] have incorporated large pre-trained models and prompt learning into the realm of continual learning. Large-scale pre-trained models have demonstrated remarkable generalization performance and robust anti-forgetting capabilities. Therefore, prompt-based methods serve as our primary comparison benchmarks and they will be subsequently elaborated upon in detail. + +# 2.3 Prompting for Continual Learning + +Prompt learning has garnered widespread success in both natural language processing (NLP) [7, 20, 33] and computer vision (CV) [21, 32, 34, 42]. It leverages prompts to fine-tune large-scale pre-trained models for downstream tasks. Firstly, L2P [51] pioneers the integration of prompt learning into the field of continual learning field. It proposes a key-query matching strategy for prompt selection from the prompt pool, followed by the insertion of these prompts into + +the pre-trained model. The subsequent works are based on this key-query matching strategy. DualPrompt [50] employs a set of task-specific prompts and selects prompts based on the closest key-query matching during inference. Furthermore, CODA [38] utilizes key-query matching to assign weights to prompts and can be optimized in an end-to-end fashion. APG [40] proposes a learnable adaptive prompt generator eases the reliance on intensive pretraining. S-Prompts [49] focuses on domain-incremental learning by acquiring distinct prompts for each domain and utilizing KNN to select prompts during inference. In this paper, we thoroughly compared these methods to CIL and DIL protocols and extensively discussed their shortcomings. + +# 3 Methods + +# 3.1 Problem Setting + +The continual learning aims to enable the model to learn non-stationary data from sequential tasks while preserving the knowledge obtained from previous tasks. We define a set of tasks $\mathcal{D}_1, \mathcal{D}_2, \dots, \mathcal{D}_T$ , where $\mathcal{D}_t = \{(x_i^t, y_i^t)\}_{i=1}^{N_i}$ is the $t$ -th task with $N_i$ samples, and $x_i^t$ and $y_i^t$ is the $i$ -th images and its label. The label sets of $\mathcal{D}_t$ are defined as $\mathcal{C}_t$ . The objective of rehearsal-free continual learning is to employ a unified model for classifying test images from all previously learned tasks while not storing any old samples or prototypes. + +Depending on the specific problems, continual learning can be divided into multiple settings, which differ in challenge and practicality. In this paper, we mainly focus on addressing the challenging class-incremental learning setting. In this scenario, the class of each task is non-overlapping, denoted as $\mathcal{C}_1 \cap \mathcal{C}_2 \cdots \cap \mathcal{C}_T = \emptyset$ and the task identity can not be obtained in the inference phase. Furthermore, to assess the robustness and generalization of our approach, we extend our experiments to domain-incremental continual learning setting. In this scenario, each task has the same classes but exhibits distinct data distributions, expressed as $\mathcal{C}_1 = \mathcal{C}_2 = \dots = \mathcal{C}_T$ . + +# 3.2 Framework and Formulation + +Figure 2 illustrates the framework of the Continual Adapter (C-ADA) approach. The C-ADA incorporates the S&S module and the plug-and-play Continual Adapter Layer (CAL) into the pre-trained model, endowing it with robust capabilities for addressing continual learning challenges. Concretely, the S&S module is exclusively trained before the incremental training to mitigate the domain gap and is subsequently frozen during further training. The CAL comprises a parameter-extensible down-projection, a ReLU activation, and a parameter-extensible up-projection. During incremental learning, the down-projection and up-projection expand the trainable weights to learn the new knowledge and freeze the previous weights to maintain the old knowledge. Moreover, we propose an orthogonal loss to alleviate the conflict between previous weights and trainable weights in each CAL. + +![](images/791ed5520d531daa20d9b6d20a5396cb1cdf36a3aee58e9740e233aae1e50028.jpg) +Fig. 2: The framework of our C-ADA approach. For simplicity, we omit the skip connection and Layernorm in the figure. or loss represents the orthogonal loss. We attach our S&S and CAL, which are in parallel with the projection layer and MLP, to the shallow N layers of ViT and freeze the pre-trained backbone. For each new task, we expand two trainable weights in the CAL to learn the new knowledge and freeze the previous weights. Different from prior works, our approach uses a novel adapter variant to eliminate the necessity for key-query matching. Only trainable weights and classifier parameters are optimized which is parameter efficient and no old information (old images or prototypes) are stored which is privacy preserving. + +The basic backbone consists of a pre-trained vision transformer feature extractor $f = f(\cdot ;\theta)$ and a classifier $g = g(\cdot ;\phi)$ . Given an input image $x \in \mathbb{R}^{H\times W\times C}$ from $\mathcal{D}_t$ . The image is reshaped and projected by the patch embedding layer, resulting in the embedding feature $x_{e} \in \mathbb{R}^{L\times D}$ , where $L$ is the token length, and $D$ is the embedding dimension. These embedding features are not only input to the pre-training module within the ViT block but also to the parallel CAL and S&S modules. The output of the backbone and parallel is added and fed into the next layer. Finally, the output features $x_{p}$ from the last ViT block are fed into the $g$ to perform classification tasks. + +# 3.3 Scale and Shift Module + +Due to the significant domain gap between the pre-training dataset and the downstream dataset, our C-ADA approach introduces a flexible and independent module, named scaling and shifting (S&S), to revise the feature space before incremental training for more efficient fine-tuning (higher plasticity). + +Concretely, the S&S module includes the trainable parameter $\alpha \in \mathbb{R}^D$ and $\beta \in \mathbb{R}^D$ as the scale and shift factors, respectively, to learn the domain information of downstream tasks. The S&S operation is calculated as follows: + +$$ +y = \alpha \odot x + \beta \tag {1} +$$ + +where $\odot$ is the dot product. We employ the S&S operation before the incremental training by using the data from the first phase (i.e. $\mathcal{D}_1$ ). Then we freeze the trainable parameters during the incremental phases. + +# 3.4 Continual Adapter Layer + +Structure of Continual Adapter Layer To address the challenges of efficiency and lightweight, the CAL is designed as a bottleneck module. It consists of a parameter-extensible down-projection, a ReLU layer for non-linear property, and a parameter-extensible up-projection. The down-projection embeds the old and new knowledge into a latent embedding along with the ReLU function, and the up-projection extracts a dimension-fixed knowledge representation from the latent embedding. For each new task, we first initialize and append two learnable linear weights to the down-projection and up-projection, respectively, and freeze the previous weights. Such design ensures that the CAL is extendable for learning new knowledge and stable in maintaining old knowledge, maintaining the input and output feature dimensions unchanged. Assuming the current task $t$ , we assign the weights in the CAL into two types: the frozen weights $\mathbf{W}^{1},\dots,\mathbf{W}^{(t - 1)}$ , and the trainable weight $\mathbf{W}^t$ ( $\mathbf{W}^i$ represents both the $\mathbf{W}_{dp}^i$ and $\mathbf{W}_{up}^i$ ). The frozen weights effectively preserve the previous knowledge to mitigate catastrophic forgetting, while simultaneously optimizing the trainable parameters to learn new knowledge. + +Without loss of generality, we define the input and output of the current layer as $x_{i}$ and $x_{i}^{\prime}$ , respectively. In task t, we can adapt the input by the CAL as follows: + +$$ +\mathbf {W} _ {d p} = \left[ \begin{array}{c} \mathbf {W} _ {d p} ^ {1}, \dots , \mathbf {W} _ {d p} ^ {t - 1}, \mathbf {W} _ {d p} ^ {t} \end{array} \right], +$$ + +$$ +\begin{array}{l} \mathbf {W} _ {u p} = \left[ \mathbf {W} _ {u p} ^ {1}, \dots , \mathbf {W} _ {u p} ^ {t - 1}, \mathbf {W} _ {u p} ^ {t} \right], \\ \mathbf {W} _ {d p} ^ {i} \in \mathbb {R} ^ {D \times d} \quad \mathbf {W} _ {u p} ^ {i} \in \mathbb {R} ^ {d \times D} \end{array} \tag {2} +$$ + +$$ +x _ {i} ^ {\prime} = \operatorname {R e L U} (x _ {i} \cdot \mathbf {W} _ {d p}) \cdot \mathbf {W} _ {u p} +$$ + +In the next task, we freeze the $\mathbf{W}_{dp}^{t}$ and $\mathbf{W}_{up}^{t}$ to maintain the old knowledge and convert two new weights $\mathbf{W}_{dp}^{t + 1}$ and $\mathbf{W}_{up}^{t + 1}$ to the trainable weights. + +Parameters of Continual Adapter Layer. In this paper, we follow the standard RFCL setting where the total task number is known before training. To compare fairly, we adjust the value of $d$ to maintain a consistent number of parameters across different settings. For real-world applications, we can keep the size of $d$ constant and continually expand the specific weights for new tasks. Remarkably, even when $d$ is set to a minimal value, such as $d = 1$ , it still exhibits commendable performance. Therefore C-ADA is lightweight, and memory-efficient in real-world applications. + +Position of Continual Adapter Layer. As depicted in [9,43], the Multi-Head Self-Attention (MHSA) block employs the attention mechanism to capture global + +information, while the MLP block prevents the ViT output from degeneration. Both components play pivotal roles in the ViT structure. Consequently, we investigate performing a comprehensive tuning of both the MHSA block and MLP block, instead of exclusively emphasizing one over the other. To fine-tune the MHSA block, we choose to insert the CAL in parallel with the Projection Layer instead of the Attention module. This decision is motivated by the observation that, despite various advanced transformer-based models using different attention mechanisms within the MHSA block [12, 26, 28], they consistently include a projection layer. On the contrary, we insert the CAL in parallel with the entire MLP block to fine-tune the MLP block. + +Without loss of generality, we define the input as $x_{l}$ and initially feed it into the Attention module to obtain $x_{a}$ , which serves as the input to the Projection Layer and $\mathrm{CAL}_1$ : + +$$ +\begin{array}{l} \begin{array}{l} x _ {a} = \operatorname {A t t e n} \left(x _ {l}\right) = \operatorname {S o f t m a x} \left(\frac {Q \cdot K ^ {T}}{\sqrt {d}}\right) V, \\ x _ {l} = \operatorname {P r e s i} \left(x _ {l}\right) = x _ {l} = \lambda_ {1} \operatorname {C A L} \left(S _ {l} - S \left(x _ {l}\right)\right) \end{array} \tag {3} \\ x _ {l 1} = \operatorname {P r o j} (x _ {a}), \quad x _ {l 2} = \lambda \cdot \operatorname {C A L} _ {1} (S \& S (x _ {a})), \\ x _ {l} ^ {\prime} = x _ {l 1} + x _ {l 2}, \\ \end{array} +$$ + +Then the output $x_{l}^{\prime}$ is further sent to the MLP block and $\mathrm{CAL}_2$ . This process is formally formulated as follows, + +$$ +\begin{array}{l} x _ {l 1} = \operatorname {M L P} \left(x _ {l} ^ {\prime}\right), \quad x _ {l 2} = \lambda \cdot \operatorname {C A L} _ {2} (S \& S \left(x _ {l} ^ {\prime}\right)), \tag {4} \\ x _ {l + 1} = x _ {l 1} + x _ {l 2}, \\ \end{array} +$$ + +where $\mathrm{CAL}_1$ and $\mathrm{CAL}_2$ are in parallel with the projection layer and MLP block, respectively. Proj represents the Projection Layer and $\lambda$ represents a scale factor (typically set to 0.1). Then, the output $x_{l + 1}$ is sent to the next layer and repeats the above process. + +# 3.5 Loss Function + +Orthogonal Loss. When training the new task, the learned new knowledge will interfere with old knowledge, potentially resulting in catastrophic forgetting. Thus we propose to solve this problem by initialization and loss function. We apply the Gram-Schmidt process to initialize the trainable parameters at the start of each new task. Orthogonal initialization contributes to the diminishment of the scope when the model parameter is updated, thereby fostering enhanced stability and overall performance. After the first task, we employ the orthogonal loss to keep trainable weights orthogonal to previous weights. + +$$ +\mathcal {L} _ {o r} = \left\| \mathbf {W} _ {d p} ^ {t} \right. ^ {T} \cdot \left\{\mathbf {W} _ {d p} ^ {1}, \dots , \mathbf {W} _ {d p} ^ {t - 1} \right\} \| _ {2} + \left\| \mathbf {W} _ {u p} ^ {t} \right. \cdot \left\{\mathbf {W} _ {u p} ^ {1}, \dots , \mathbf {W} _ {u p} ^ {t - 1} \right\} ^ {T} \| _ {2} \tag {5} +$$ + +where $\| \cdot \|_2$ represent the $L_{2}$ -norm and $T$ represents the matrix transpose. The loss function forces the model to optimize current weight $\mathbf{W}^t$ to be orthogonal to previous weight $\mathbf{W}^1, \dots, \mathbf{W}^{t-1}$ . Intuitively, orthogonal vectors exhibit reduced mutual influence. As a result, this design aims to minimize interference between new and old knowledge to reduce catastrophic forgetting. + +Classification Loss. Similar to the previous methods, we use the CE loss to optimize the model to learn the new knowledge + +$$ +\mathcal {L} _ {c e} = \mathcal {L} (g (f (\boldsymbol {x}; \mathbf {W} _ {d p}, \mathbf {W} _ {u p}, \theta); \phi), y) \tag {6} +$$ + +where $x$ is the input image and $y$ is the corresponding label. $\phi$ is the parameters of the classifier and $\theta$ is the parameters of the pre-train model. + +Total Optimization Objective. For task $t$ , our full optimization consists of a classification loss $\mathcal{L}_{ce}$ and an orthogonal loss $\mathcal{L}_{or}$ + +$$ +\min _ {\mathbf {W} _ {\mathrm {d p}} ^ {t}, \mathbf {W} _ {\mathrm {u p}} ^ {t}, \phi} \mathcal {L} _ {c e} + \delta \mathcal {L} _ {\mathrm {o r}} \tag {7} +$$ + +where $\mathbf{W}_{dp}^{t},\mathbf{W}_{up}^{t}$ and $\phi$ are the learnable parameters of task t. We set the $\delta$ to 1 for all the experiments. During each training phase, our models can be optimized in an end-to-end fashion. + +# 4 Experiments + +In this paper, we mainly aim to solve the rehearsal-free class-incremental learning problem. To evaluate our methods, we follow the RFCL settings proposed in previous works [38, 50, 51] and conduct comprehensive experiments. + +Besides, to prove the robustness and generality of our method on other continual learning settings, we conduct experiments compared with the SOTA methods [49] on the domain-incremental learning settings. Extensive ablation studies prove the effectiveness of our methods. + +# 4.1 Evaluation Benchmarks + +Datasets. We use the Split CIFAR-100 [24] and Split ImageNet-R [17] for class-incremental setting, CORe50 [27] and DomainNet [30] for domain-incremental setting. + +CIFAR-100 is a widely-used benchmark in continual learning, which contains 60000 images of $32 \times 32$ size from 100 classes. We split CIFAR-100 into the 10-task benchmark following previous works. + +ImageNet-R has 200 classes and includes newly collected data of different styles. It provides a fair and challenging problem setting due to the significant gap between training data and pre-trained data. We split the dataset into the 5-task benchmark, the original 10-task benchmark and the 20-task benchmark. + +CORe50 has 50 classes from 11 distinct domains. Following the domain-incremental setting, we use 8 domains for (120000 images) for continual training and the rest of the domains for testing. + +DomainNet is a large dataset with 345 classes and 600000 images. These images are split into 6 domains (Clipart, Infograph, Paint, Quickdraw, Real, and Sketch) for continual learning like S-Prompts [49]. + +Table 1: Results (\%) on ImageNet-R and CIFAR-100. The results are all obtained by CODA [38] directly. P represents the total number of tasks. UB represents the full fine-tuning result from [10]. We report $A_N(\%)$ , $Param(\%)$ and $FLOPs(\mathrm{G})$ to bring intuitive comparison. C-ADA uses fewer parameters and doubles the training speed, greatly surpassing other methods. + +
MethodsImageNet-RCIFAR-100FLOPs
P=5P=10P=20P=10
AN(↑)Param(↓)AN(↑)Param(↓)AN(↑)Param(↓)AN(↑)Param(↓)
UB77.13100/10077.13100/10077.13100/10089.30100/100-
ER (5000)71.72100/10064.43100/10052.43100/10076.20100/100-
FT18.74100/10010.12100/1004.75100/1009.92100/100-
FT++60.42100/10048.93100/10035.98100/10049.91100/100-
LwF.MC74.56100/10066.73100/10054.05100/10064.83100/100-
L2P70.830.7/100.769.290.7/100.765.890.7/100.782.500.7/100.735.2
Deep L2P73.939.6/109.671.669.6/109.668.429.6/109.684.309.6/109.635.2
DualPrompt73.050.5/100.571.320.8/100.867.871.3/101.383.050.8/100.835.2
CODA-P-S75.190.7/100.773.930.7/100.770.530.7/100.784.590.6/100.635.2
C-ADA77.930.7/100.776.660.7/100.773.470.7/100.787.180.6/100.617.6
+ +Comparing Methods. We compare our method with several baselines and state-of-the-art (SOTA) continual learning methods. For a fair comparison, all the methods are based on a ViT-B/16 backbone pre-trained on ImageNet-1K. Under class-incremental settings, we regard L2P [51], DualPrompt [50], and CODA [38] as the main competitors, which are rehearsal-free methods. Under domain-incremental settings, despite S-Prompt [49] storing the prototypes of previous domains instead of the rehearsal-free method, it is still the main comparison method as the SOTA of domain-incremental learning. + +Implementation Details. Similar to CODA [38], we use the Adam [22] optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ , and a batch size of 128 images in four GPUs. For every image, we resize it to $224\times 224$ and normalize them to [0,1]. The learning rate is set to $5e^{-5}$ for all the settings with cosine-decaying. For fair comparisons, we set the total middle dimension to 60 (sum of two CALs) and insert the CAL into layers 0-4, which achieves a similar parameter amount to previous works [38, 50]. We train the CIFAR-100 for 20 epochs, ImageNet-R for 50 epochs, CORe50 and DomainNet for 10 epoch. During the training phase, we substitute predictions from previous-task logits with negative infinity during the training of a new task. This design leads to a softmax prediction of "0" for these previous task classes and ensures that gradients do not affect the linear heads of past task classes like previous works. [38, 50, 51] + +# 4.2 Main Results + +Class-incremental learning. We report the results on two class-incremental datasets in Table 1 following the previous work [38]. + +Table 2: Results (\%) on CORe50 and DomainNet. $A_N$ gives the accuracy averaged over tasks and buffer size represents the number of old exemplars stored in the memory buffer. The results are all obtained by S-Prompt [49] directly. We reproduce the CODA on these settings for further comparison. We ensure that all methods have the same pre-trained model (ViT-B/16). + +
MethodBuffer size\(A_N(↑)\)
Upper-Bound-84.01
ER80.10
GDumb74.92
BiC79.28
DER++50/class79.70
Co2L79.75
DyTox79.21
L2P81.07
EWC74.82
LwF75.45
L2P0/class78.33
CODA-P-S85.41
S-Prompts83.13
C-ADA89.45
+ +
MethodBuffer sizeAN(↑)
Upper-Bound-63.22
DyTox50/class62.94
EWC47.62
LwF49.19
SimCLR44.20
BYOL49.70
Barlow Twins0/class48.90
Supervised Contrastive50.90
L2P40.15
CODA-P-S47.56
S-Prompts50.62
C-ADA53.00
+ +ImageNet-R: We report the results with 5-tasks, 10-tasks, and 20-tasks to evaluate the robustness of the model under different settings in Table 1. It is worth noting that our method significantly surpasses the previous methods in average accuracy under different settings with as much as $+2.74\%$ , $+2.73\%$ and $+2.94\%$ over CODA, $+4.88\%$ , $+5.41\%$ and $+5.60\%$ over DualPrompt (second best method). This phenomenon shows that our method is robust to different class increment settings. Moreover, since our method only requires forward propagation one time. We achieved an obvious increase in training speed while improving accuracy. Combining accuracy and training speed brings better practicality in the real world. + +CIFAR-100: We note that our approach also brings similar results to the ImageNet-R benchmarks, with an improvement of $+2.59\%$ compared with CODA, $+4.13\%$ compared with DualPrompt. + +Domain-incremental learning. We report the results on two domain-incremental datasets in Table 2. Besides, we reproduced CODA on these datasets to bring intuitive comparison. + +CORe50 and DomainNet are widely used domain dataset for continual learning. Following the setting of S-Prompt [49], we divide the baseline methods into two groups, exemplars-based methods [2-4, 11, 31, 52] and exemplars-free methods [49]/rehearsal-free methods [38, 51]. Exemplars-based methods store 50 old exemplars per class in the memory buffer. These methods often have better accuracy but are accompanied by the risk of privacy leakage and memory overhead. Our method surpasses the CODA by $+4.04\%$ and $+5.44\%$ on CORe50 and DomainNet benchmark. Besides, it is worth noting that our method even out + +![](images/b9d7987e367c2b72787914d04d9403be925148a5712b7e56136f7e56603427c0.jpg) +Fig. 3: Average accuracy $A_{N}$ (\%) vs tuning parameters (\%). We report the results on the 20-tasks ImageNet-R. + +
MethodCIFAR-100ImageNet-R
C-ADA87.3276.66
Ablate S&S86.1175.93
Ablate Lor82.8972.17
Ablate CAL76.9364.45
+ +Table 3: Ablation Results (%) on 10-task ImageNet-R and CIFAR-100. $A_N$ gives the accuracy averaged over tasks. We ablate the key components in turn and report the results. + +performs S-Prompt, which is specifically designed for domain-incremental learning, with improvements of $+6.32\%$ and $+2.38\%$ . This interesting phenomenon demonstrates the robustness and effectiveness of our method in various continual learning settings. + +# 4.3 Key Components Ablation + +We conducted ablation experiments to analyze the importance of each component. We incrementally introduced key components on both the 10-tasks ImageNet-R and CIFAR-100 datasets, and the results are presented in Table 3. Firstly, we ablate the S&S module, the performance has dropped $(0.73\% \downarrow$ on ImageNet-R). This shows that revising the feature space before training is beneficial to improving subsequent learning ability. Secondly, removing the orthogonal loss leads to severe performance degradation on average accuracy $(3.76\% \downarrow$ on ImageNet-R). This outcome aligns with our expectations, as ablating the orthogonal loss results in our method lacking a mechanism to prevent forgetting, leading to severe catastrophic forgetting. Thirdly, when we ablated the CAL component, we observed a noticeable drop in performance $(7.72\% \downarrow$ on ImageNet-R). The removal of the CAL implies that our model relies solely on adjusting the classifier to tune the model, which obviously leads to unsatisfactory results, particularly when dealing with the downstream datasets having a significant gap from the pre-trained dataset. + +# 4.4 Scaling Trainable Parameters + +We conduct experiments to discuss the relationship between the trainable parameters amount and the performance. + +The trainable parameters amount can be adjusted by changing the middle dimension of the CAL. The total number of middle dimensions of the two CALs in each block is chosen from $\{20, 40, 60, 80, 100, 120, 140, 160\}$ , corresponding to $\{0.35\%, 0.54\%, 0.71\%, 0.90\%, 1.12\%, 1.27\%, 1.45\%, 1.61\}$ trainable parameters. We report the results on 20-tasks ImageNet-R in Figure 3. The figure + +shows that more trainable parameters contribute to achieving a higher accuracy. Moreover, we can notice that even with the fewer trainable parameters, our model consistently demonstrates superior performance when compared to CODA (71.92% $A_N$ , 0.35% $\text{Param}$ of C-ADA vs 70.53% $A_N$ , 0.7% $\text{Param}$ of CODA). This fully illustrates that our approach improves the plasticity compared with previous works. + +# 4.5 Attaching Layer and Position + +Intuitively, each layer of the ViT exhibits distinct feature extraction capabilities [50]. The impact of tuning different layers may be quite different in CL. Furthermore, each module of the layer plays different roles, which causes attaching CAL to various positions within the layer may yield divergent outcomes. Therefore, it is crucial to explore layers and positions to attach the CAL under the CL settings. Note that the total number of trainable parameters remains consistent across all experiments described below (middle $\dim = d*t$ ). + +Layer of CAL. For simplicity, we assume that the attaching layers are contiguous. We report the result in Table 4. + +Firstly, we attach the CAL to all the layers to get a preliminary result (Layer 0 to Layer 11). Next, we use binary search to analyze the importance of shallow (Layer 0 to Layer 5) and deep (Layer 6 to Layer 11) layers. We observe that attaching to shallow layers gets superior results compared to attaching to deeper layers. This phenomenon is consistent with the observation from previous work [50]. Then we continue to reduce the number of attached layers in shallow layers. Experimental results show that attaching from layer $= 0$ to layer $= 4$ performs the best. We use this conclusion to analyze the attached modules. + +Position of CAL. By analyzing common parts of the ViT architecture, we select three alternative attaching positions: QKV attention layer $(W_{q}, W_{k}, W_{v})$ in the MHSA block, projection layer in the MHSA block, and the MLP block. We report the result in Table 5. + +Firstly, we analyze the effects of only attaching these three positions, respectively. We observe that attaching the CAL to the projection layer has a better effect than in the QKV attention layer. This phenomenon makes us attach the CAL to the projection layer to tune the MHSA block. Then we opt to combine the tuning of different positions. We observe that attaching the CAL to different positions resulted in certain improvements (about $0.4\%$ ). Interestingly, it seems to have little to do with how to assign the middle dimensions. This shows to some extent that our method is robust. + +# 4.6 Extra Evaluation Metrics + +Although we emphasize that $A_N$ , $\text{Param}$ , and $FLOPs$ are the more important metrics, there are still some common metrics in the CL field. + +Table 4: Results $(\%)$ on ImageNetR 10 tasks. We analyze the difference of attaching to different layers. We keep the total tunable parameters constant by modifying the middle dimension, + +
Attach Layermiddle dimAN(↑)
Layer 0 ~ Layer 113075.23
Layer 6 ~ Layer 115072.51
Layer 0 ~ Layer 55076.21
Layer 1 ~ Layer 56076.29
Layer 0 ~ Layer 46076.66
Layer 2 ~ Layer 58075.72
Layer 0 ~ Layer 38075.75
+ +Table 5: Results $(\%)$ on ImageNetR 10 tasks. We fix the attaching layers (layer 0 to layer 4) to analyze two aspects: the attaching positions and the middle dimension of each CAL. + +
Attach Positionmiddle dimAN(↑)
QKV Attention6076.04
MLP6076.23
Proj6076.25
Proj + MLP20 + 4076.52
Proj + MLP30 + 3076.57
Proj + MLP50 + 1076.51
Proj + MLP10 + 5076.61
Proj + MLP40 + 2076.66
+ +$F_{N}$ : Average forgetting metrics, represent the drop in task performance averaged over $N$ tasks. + +bwt: Backward transfer, assesses the quality of representation. + +fwt: Forward transfer, represents the influence of existing knowledge on the performance of subsequent concepts. + +In Table 6, we can see that across various metrics, C-ADA demonstrates superior performance. Notably, in terms of $fwt$ , which represents the capacity to learn new knowledge, it significantly outperforms the other methods. This emphasizes the effectiveness of C-ADA in learning new knowledge. + +Table 6: Results (\%) on ImageNet-R. The $F_N$ results are all obtained by CODA [38] directly. We reproduce the bwt and fwt metrics for further comparison. + +
MethodsP=5P=10P=20P=10P=10
FN(↓)FN(↓)FN(↓)bwt(↑)fwt(↑)
L2P3.36 ± 0.182.03 ± 0.191.24 ± 0.14-3.69 ± 0.07-0.32 ± 0.04
DualPrompt2.64 ± 0.171.71 ± 0.241.07 ± 0.14-3.48 ± 0.06-0.09 ± 0.04
CODA2.65 ± 0.151.60 ± 0.201.00 ± 0.15-3.24 ± 0.060.01 ± 0.03
C-ADA2.55 ± 0.141.58 ± 0.181.02 ± 0.12-3.12 ± 0.050.34 ± 0.03
+ +# 5 Conclusion + +In this paper, we present a novel adapter variant, named Continual Adapter (C-ADA), for RFCL. Our approach assembles a parameter-extensible adapter layer (CAL) and an S&S module in parallel with the pre-trained model. Importantly, C-ADA eliminates the necessity for key-query matching to achieve better plasticity and efficient training. We set a new SOTA on well-established benchmarks of RFCL and outperform the mainstream methods of DIL. This underscores the generality and robustness of our approach across various continual learning settings. + +# Acknowledgments + +This work was funded by the National Natural Science Foundation of China under Grant No.U21B2048 and No.62302382 and Shenzhen Key Technical Projects under Grant CJGJZD2022051714160501 and the CAAI-MindSpore Open Fund, developed on OpenI Community. + +# References + +1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV). pp. 139-154 (2018) 4 +2. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems 33, 15920-15930 (2020) 11 +3. Cha, H., Lee, J., Shin, J.: Co2l: Contrastive continual learning. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 9516-9525 (2021) 11 +4. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019) 11 +5. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 3 +6. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021) 4 +7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4 +8. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021) 2, 4 +9. Dong, Y., Cordonnier, J.B., Loukas, A.: Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In: International Conference on Machine Learning. pp. 2793-2803. PMLR (2021) 7 +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) 10 +11. Douillard, A., Ramé, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9285-9295 (2022) 4, 11 +12. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 8 +13. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999) 2 + +14. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) 4 +15. Gao, X., He, Y., Dong, S., Cheng, J., Wei, X., Gong, Y.: Dkt: Diverse knowledge transfer transformer for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24236-24245 (2023) 2, 4 +16. Gao, Z., Cen, J., Chang, X.: Consistent prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28463-28473 (2024) 4 +17. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021) 9 +18. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 831-839 (2019) 1, 2, 4 +19. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) 3 +20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 4 +21. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) 4 +22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10 +23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) 2 +24. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., Citeseer (2009) 9 +25. Li, Z., Hoiem, D.: Learning without forgetting. T-PAMI 40(12), 2935-2947 (2018) 2, 4 +26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) 8 +27. Lomonaco, V., Maltoni, D.: Core50: a new dataset and benchmark for continuous object recognition. In: Conference on robot learning. pp. 17-26. PMLR (2017) 9 +28. Mao, X., Qi, G., Chen, Y., Li, X., Duan, R., Ye, S., He, Y., Xue, H.: Towards robust vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12042-12051 (2022) 8 +29. McDonnell, M.D., Gong, D., Parvaneh, A., Abbasnejad, E., van den Hengel, A.: Ranpac: Random projections and pre-trained models for continual learning. Advances in Neural Information Processing Systems 36 (2024) 4 +30. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 9 + +31. Prabhu, A., Torr, P.H., Dokania, P.K.: Gdumb: A simple approach that questions our progress in continual learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 524-540. Springer (2020) 11 +32. Qin, Y., Chen, Y., Peng, D., Peng, X., Zhou, J.T., Hu, P.: Noisy-correspondence learning for text-to-image person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27197-27206 (2024) 4 +33. Qin, Y., Peng, D., Peng, X., Wang, X., Hu, P.: Deep evidential learning with noisy correspondence for cross-modal retrieval. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4948-4956 (2022) 4 +34. Qin, Y., Sun, Y., Peng, D., Zhou, J.T., Peng, X., Hu, P.: Cross-modal active complementary learning with self-refining correspondence. Advances in Neural Information Processing Systems 36 (2024) 4 +35. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 1, 2, 4 +36. Shieh, J.L., Haq, Q.M.u., Haq, M.A., Karam, S., Chondro, P., Gao, D.Q., Ruan, S.J.: Continual learning strategy in one-stage object detection framework based on experience replay for autonomous driving vehicle. Sensors 20(23), 6777 (2020) 1 +37. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021) 4 +38. Smith, J.S., Karlinsky, L., et al.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: CVPR. pp. 11909-11919 (2023) 1, 2, 4, 5, 9, 10, 11, 14 +39. Song, X., He, Y., Dong, S., Gong, Y.: Non-exemplar domain incremental object detection via learning domain bias. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 15056-15065 (2024) 4 +40. Tang, Y.M., et al.: When prompt-based incremental learning does not meet strong pretraining. In: CVPR (2023) 2, 4, 5 +41. Tao, X., Chang, X., Hong, X., Wei, X., Gong, Y.: Topology-preserving class-incremental learning. In: European Conference on Computer Vision. pp. 254-270. Springer (2020) 2, 4 +42. Tu, C.H., Mai, Z., Chao, W.L.: Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7725-7735 (2023) 4 +43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 7 +44. Verwimp, E., Yang, K., Parisot, S., Hong, L., McDonagh, S., Pérez-Pellitero, E., De Lange, M., Tuytelaars, T.: Clad: A realistic continual learning benchmark for autonomous driving. Neural Networks 161, 659-669 (2023) 1 +45. Wang, F.Y., Zhou, D.W., Ye, H.J., Zhan, D.C.: Foster: Feature boosting and compression for class-incremental learning. arXiv preprint arXiv:2204.04662 (2022) 4 +46. Wang, L., et al.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. NeurIPS (2024) 4 +47. Wang, S., Shi, W., Dong, S., Gao, X., Song, X., Gong, Y.: Semantic knowledge guided class-incremental learning. IEEE Transactions on Circuits and Systems for Video Technology 33(10), 5921-5931 (2023) 4 + +48. Wang, S., Shi, W., He, Y., Yu, Y., Gong, Y.: Non-exemplar class-incremental learning via adaptive old class reconstruction. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4524-4534 (2023) 4 +49. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022) 4, 5, 9, 10, 11 +50. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799 (2022) 2, 4, 5, 9, 10, 13 +51. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022) 1, 2, 4, 9, 10, 11 +52. Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., Fu, Y.: Large scale incremental learning. arXiv preprint arXiv:1905.13260 (2019) 11 +53. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3014-3023 (2021) 1, 2, 4 +54. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International conference on machine learning. pp. 3987-3995. PMLR (2017) 4 +55. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023) 4 +56. Zhou, D.W., et al.: Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. arXiv (2023) 4 +57. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5871-5880 (2021) 2, 4 +58. Zhu, K., Zhai, W., Cao, Y., Luo, J., Zha, Z.J.: Self-sustaining representation expansion for non-exemplar class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9296-9305 (2022) 2, 4 \ No newline at end of file diff --git a/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/images.zip b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9f61c3fb9e14352200854bdb27b4d4949e3d427f --- /dev/null +++ b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3407555c835e22ca5045106d338a74d76de00747dfdf4d24ff3f6b35bf76c760 +size 395945 diff --git a/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/layout.json b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eec3b31d456329a599bb4c798a671138583ecb9a --- /dev/null +++ b/2024/Beyond Prompt Learning_ Continual Adapter for Efficient Rehearsal-Free Continual Learning/layout.json @@ -0,0 +1,10051 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "text", + "content": "Beyond Prompt Learning: Continual Adapter for Efficient Rehearsal-Free Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "text", + "content": "Xinyuan Gao" + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "inline_equation", + "content": "^{1\\#}" + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "text", + "content": ", Songlin Dong" + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "inline_equation", + "content": "^{2\\#}" + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "text", + "content": ", Yuhang He" + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "text", + "content": ", Qiang Wang" + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "text", + "content": ", and Yihong Gong" + }, + { + "bbox": [ + 157, + 167, + 457, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 147, + 201, + 465, + 247 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 183, + 201, + 430, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 201, + 430, + 213 + ], + "spans": [ + { + "bbox": [ + 183, + 201, + 430, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 183, + 201, + 430, + 213 + ], + "type": "text", + "content": " School of Software Engineering, Xi'an Jiaotong University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 213, + 465, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 213, + 465, + 247 + ], + "spans": [ + { + "bbox": [ + 147, + 213, + 465, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 213, + 465, + 247 + ], + "type": "text", + "content": " College of Artificial Intelligence, Xi'an Jiaotong University {gxy010317, dsl97273141, qwang}@stu.xjtu.edu.cn, heyuhang@xjtu.edu.cn ygong@mail.xjtu.edu.cn" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 269, + 455, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 269, + 455, + 534 + ], + "spans": [ + { + "bbox": [ + 159, + 269, + 455, + 534 + ], + "type": "text", + "content": "Abstract. The problem of Rehearsal-Free Continual Learning (RFCL) aims to continually learn new knowledge while preventing forgetting of the old knowledge, without storing any old samples and prototypes. The latest methods leverage large-scale pre-trained models as the backbone and use key-query matching to generate trainable prompts to learn new knowledge. However, the domain gap between the pre-training dataset and the downstream datasets can easily lead to inaccuracies in key-query matching prompt selection when directly generating queries using the pre-trained model, which hampers learning new knowledge. Thus, in this paper, we propose a beyond prompt learning approach to the RFCL task, called Continual Adapter (C-ADA). It mainly comprises a parameter-extensible continual adapter layer (CAL) and a scaling and shifting (S&S) module in parallel with the pre-trained model. C-ADA flexibly extends specific weights in CAL to learn new knowledge for each task and freezes old weights to preserve prior knowledge, thereby avoiding matching errors and operational inefficiencies introduced by key-query matching. To reduce the gap, C-ADA employs an S&S module to transfer the feature space from pre-trained datasets to downstream datasets. Moreover, we propose an orthogonal loss to mitigate the interaction between old and new knowledge. Our approach achieves significantly improved performance and training speed, outperforming the current state-of-the-art (SOTA) method. Additionally, we conduct experiments on domain-incremental learning, surpassing the SOTA, and demonstrating the generality of our approach in different settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 543, + 361, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 543, + 361, + 554 + ], + "spans": [ + { + "bbox": [ + 160, + 543, + 361, + 554 + ], + "type": "text", + "content": "Keywords: Continual Learning, Rehearsal-Free" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 572, + 230, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 230, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 230, + 585 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 632 + ], + "type": "text", + "content": "In recent years, with the emerging applications of deploying DNN models to dynamic and continuous environments such as autonomous driving [36, 44]. Continual Learning (CL) [18, 35, 38, 51, 53] has drawn a rapidly increasing research" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 643, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 482, + 665 + ], + "type": "text", + "content": "* Yuhang He is the corresponding author; # Xinyuan Gao and Songlin Dong are co-first authors" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 291, + 211 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 291, + 211 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 291, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 291, + 211 + ], + "type": "image", + "image_path": "e8bf0cbaa9cdc6fc3f0fbbc7cea90d34e7dc484b8c28456748cd6a8f5f5ece7e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 213, + 264, + 221 + ], + "lines": [ + { + "bbox": [ + 157, + 213, + 264, + 221 + ], + "spans": [ + { + "bbox": [ + 157, + 213, + 264, + 221 + ], + "type": "text", + "content": "(a): Prior Prompt-Based Approach" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 223, + 462, + 232 + ], + "lines": [ + { + "bbox": [ + 145, + 223, + 462, + 232 + ], + "spans": [ + { + "bbox": [ + 145, + 223, + 462, + 232 + ], + "type": "text", + "content": "Decision Boundary Feature Tuned Frozen S85 Scaling & Shifting First Pass-forward Second Pass-forward" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 113, + 479, + 156 + ], + "blocks": [ + { + "bbox": [ + 299, + 113, + 479, + 156 + ], + "lines": [ + { + "bbox": [ + 299, + 113, + 479, + 156 + ], + "spans": [ + { + "bbox": [ + 299, + 113, + 479, + 156 + ], + "type": "image", + "image_path": "c9ac87f5cd12fe58cdb8832d1c8dc2c77510fc3b23e3f314ee8d843ffd36c651.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 157, + 440, + 166 + ], + "lines": [ + { + "bbox": [ + 338, + 157, + 440, + 166 + ], + "spans": [ + { + "bbox": [ + 338, + 157, + 440, + 166 + ], + "type": "text", + "content": "(b):Continual Adapter Approach" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 167, + 470, + 212 + ], + "blocks": [ + { + "bbox": [ + 313, + 167, + 470, + 212 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 470, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 470, + 212 + ], + "type": "image", + "image_path": "36a73c6bb81b6f73b1ab85a59a75640dc6f32ba131889bd45bd7e98d1749203b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 213, + 481, + 221 + ], + "lines": [ + { + "bbox": [ + 299, + 213, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 299, + 213, + 481, + 221 + ], + "type": "text", + "content": "(c): The Results " + }, + { + "bbox": [ + 299, + 213, + 481, + 221 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 299, + 213, + 481, + 221 + ], + "type": "text", + "content": " of each new task on 5-task ImageNet-R" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 237, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 315 + ], + "type": "text", + "content": "Fig. 1: Prior prompt-based approaches pass through the pre-trained model to generate the query and employ key-query matching to select prompts, which are inserted into the pre-trained model again (each layer has a unique prompt). Continual Adapter approach (C-ADA) strategy eliminates the need for key-query matching by introducing the CAL and S&S, which brings significant improvement in learning the new knowledge. Moreover, C-ADA only needs to pass the pre-trained model once, highlighting the training speed." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "type": "text", + "content": "interest both in the literature and industry. Classical CL methods focus on addressing the problem of catastrophic forgetting [13], wherein the models suffer dramatic performance decreases on previous tasks once training on new data. The majority of research studies [8, 15, 18, 35, 41, 53, 57, 58] attempt to reduce forgetting by retaining a subset of previous image samples or prototypes and replaying them in new tasks, referred to as the rehearsal-based methods. However, on the concerns of data privacy and deployment efficiency, the applications of these methods in real-world scenarios are hampered. Therefore, we focus on a more challenging rehearsal-free CL (RFCL) [23, 25, 38, 50, 51], which does not store any previous samples or prototypes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "Latest approaches [38,40,50,51] leverage prompt learning techniques to solve the RFCL problem. Due to freezing the large-scale pre-trained model, learning new knowledge (i.e. plasticity) becomes more challenging rather than reducing the forgetting (i.e. stability) [38]. To acquire new knowledge, prompt-based methods utilize key-query matching to obtain trainable prompts. As depicted in Figure 1(a), these methods first generate a feature representation (i.e. query) for each instance using the pre-trained model. Subsequently, they select or generate prompts for each instance based on the cosine similarity between the query and keys, followed by a second pass through the pre-trained model. Despite their promising progress, they are faced with two challenges. 1) Due to the substantial gap between the pre-trained dataset and the downstream datasets, the query embeddings directly generated by the pre-trained model are prone to confusion, leading to inaccurate key-query matching prompt selection. The inaccurate selection seriously hampers the learning ability of the model. 2) These methods require two forward passes, one for calculating the query and another for logit, leading to relatively low efficiency. Therefore, there are inherent shortcomings in prompt-based methods for RFCL tasks, motivating us to explore a novel approach distinct from prompt-learning techniques." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 451 + ], + "type": "text", + "content": "Taking into account the inherent drawbacks mentioned above, in this paper, we shift our focus to adapter tuning [19] and propose a simple yet effective adapter variant, named ContinuaIAdapter (C-ADA) for rehearsal-free continual learning. Compared to prompt-learning techniques, C-ADA eliminates the necessity for key-query matching, fundamentally avoiding issues arising from erroneous key-query matches that could lead to conflicts in training parameters. As depicted in Figure 1 (c), this results in a substantial enhancement in the accuracy of C-ADA for new tasks, and C-ADA achieves this with just a single forward pass, greatly boosting training speed. Moreover, in contrast to lacking incremental capability adapter fine-tuning methods [5,19], we replace the adapter layer with a plug-and-play, parameter-expandable adapter layer, named Continuous Adapter Layer (CAL). The CAL expands the specific learnable weights in a new dimension (middle dimension) to acquire new knowledge and freeze the previously learned weights to retain old knowledge, which empowers the model with the capacity to tackle complex continual scenarios. This fundamentally encourages knowledge reuse, as the previous projection weights will contribute to future projection weights, facilitating the learning of new knowledge. We also introduce an orthogonal loss function to alleviate the conflict between new weights and previous weights, which significantly reduces the forgetting of old tasks. To further solve the problem of the gap between pre-trained datasets and downstream datasets, C-ADA includes scaling and shifting (S&S) module with negligible parameters to transfer the feature space from pre-trained datasets to downstream datasets before the incremental phase. Compared to the current SOTA methods, our model demonstrates significant improvements in both performance and training speed under equivalent parameter counts for RFCL. Furthermore, to validate its generality, we extend the method to the Domain-incremental Learning (DIL) setting and once again outperform the SOTA methods. The key contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 467, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 467, + 481, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 515 + ], + "type": "text", + "content": "- We propose a simple yet effective approach called C-ADA, which eliminates the necessity for key-query matching, for RFCL, characterized by better plasticity and more efficient training compared to existing continual prompting approaches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 517, + 481, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 517, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 517, + 481, + 553 + ], + "type": "text", + "content": "- We design a novel CAL module, which comprises a parameter-extensible down-projection and up-projection. It facilitates the preservation of previously acquired knowledge and the learning of new knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 555, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 616 + ], + "type": "text", + "content": "- We develop an S&S module to further reduce the divergence between the pre-training dataset and the downstream datasets, thereby enhancing the learning of new tasks. Additionally, we introduce an orthogonal loss to mitigate the interaction between old and new knowledge, which significantly contributes to the preservation of old knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 481, + 665 + ], + "type": "text", + "content": "- We conduct extensive experiments on well-established rehearsal-free benchmarks and achieve a new SOTA performance. Meanwhile, we evaluate our approach on the DIL setting and again outperform the state of the art, highlighting the generality." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Continual Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 345, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 345, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 345, + 152 + ], + "type": "text", + "content": "2.1Rehearsal-Based Continual Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "content": "[8, 11, 15, 39, 41, 45, 47–49, 53, 57, 58] retain prior information (old exemplars or prototypes) for subsequent tasks, aimed at mitigating catastrophic forgetting by replay them in the new tasks. Exemplars-based methods [8, 11, 15, 41, 45, 53] store a portion of old exemplars in a memory buffer. These exemplars are trained alongside the current samples to optimize the model. For instance, iCaRL [35] and its variants [18, 41, 53] prevent forgetting by employing the herding technique for exemplar selection and designing distinct distillation losses. However, exemplar-based methods not only require extra memory but also experience a significant performance decline as the buffer size decreases. Exemplars-free methods [57, 58] typically store a prototype per class to refine the classifier. For example, [57] memorizes and augments prototypes to maintain the decision boundary. Some latest methods SLCA [55], RanPAC [29], HiDe-Prompt [46] combine the pretrained model and prototypes to achieve better performance. Since they are not rehearsal-free continual learning, we have excluded comparisons with them." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 481, + 376 + ], + "type": "text", + "content": "Nevertheless, preserving previous information may introduce privacy and data leakage risks, often impractical in real-world applications. This motivates us to concentrate on the rehearsal-free setting, which offers enhanced privacy protection and real-world application." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 392, + 336, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 336, + 405 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 336, + 405 + ], + "type": "text", + "content": "2.2Rehearsal-FreeContinual Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 412, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 556 + ], + "type": "text", + "content": "Early rehearsal-free methods [1, 25, 54] utilize regularization to alleviate model forgetting. While these approaches demonstrate strong performance in task-incremental continual learning, their effectiveness diminishes in the more challenging class-incremental learning settings. Generative-based methods [6, 14, 37] employ deep model inversion to generate images from previous tasks. While these methods perform well in some scenarios, inversion is hampered by slow training speed and high computational costs. Recent studies [16, 38, 40, 50, 51, 56] have incorporated large pre-trained models and prompt learning into the realm of continual learning. Large-scale pre-trained models have demonstrated remarkable generalization performance and robust anti-forgetting capabilities. Therefore, prompt-based methods serve as our primary comparison benchmarks and they will be subsequently elaborated upon in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 574, + 334, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 334, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 334, + 586 + ], + "type": "text", + "content": "2.3 Prompting for Continual Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Prompt learning has garnered widespread success in both natural language processing (NLP) [7, 20, 33] and computer vision (CV) [21, 32, 34, 42]. It leverages prompts to fine-tune large-scale pre-trained models for downstream tasks. Firstly, L2P [51] pioneers the integration of prompt learning into the field of continual learning field. It proposes a key-query matching strategy for prompt selection from the prompt pool, followed by the insertion of these prompts into" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "the pre-trained model. The subsequent works are based on this key-query matching strategy. DualPrompt [50] employs a set of task-specific prompts and selects prompts based on the closest key-query matching during inference. Furthermore, CODA [38] utilizes key-query matching to assign weights to prompts and can be optimized in an end-to-end fashion. APG [40] proposes a learnable adaptive prompt generator eases the reliance on intensive pretraining. S-Prompts [49] focuses on domain-incremental learning by acquiring distinct prompts for each domain and utilizing KNN to select prompts during inference. In this paper, we thoroughly compared these methods to CIL and DIL protocols and extensively discussed their shortcomings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 253, + 208, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 253, + 208, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 208, + 266 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 277, + 246, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 246, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 246, + 289 + ], + "type": "text", + "content": "3.1 Problem Setting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": "The continual learning aims to enable the model to learn non-stationary data from sequential tasks while preserving the knowledge obtained from previous tasks. We define a set of tasks " + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1, \\mathcal{D}_2, \\dots, \\mathcal{D}_T" + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{(x_i^t, y_i^t)\\}_{i=1}^{N_i}" + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": "-th task with " + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "inline_equation", + "content": "N_i" + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": " samples, and " + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "inline_equation", + "content": "x_i^t" + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "inline_equation", + "content": "y_i^t" + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": "-th images and its label. The label sets of " + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": " are defined as " + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 130, + 295, + 482, + 379 + ], + "type": "text", + "content": ". The objective of rehearsal-free continual learning is to employ a unified model for classifying test images from all previously learned tasks while not storing any old samples or prototypes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 488 + ], + "type": "text", + "content": "Depending on the specific problems, continual learning can be divided into multiple settings, which differ in challenge and practicality. In this paper, we mainly focus on addressing the challenging class-incremental learning setting. In this scenario, the class of each task is non-overlapping, denoted as " + }, + { + "bbox": [ + 130, + 380, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1 \\cap \\mathcal{C}_2 \\cdots \\cap \\mathcal{C}_T = \\emptyset" + }, + { + "bbox": [ + 130, + 380, + 482, + 488 + ], + "type": "text", + "content": " and the task identity can not be obtained in the inference phase. Furthermore, to assess the robustness and generalization of our approach, we extend our experiments to domain-incremental continual learning setting. In this scenario, each task has the same classes but exhibits distinct data distributions, expressed as " + }, + { + "bbox": [ + 130, + 380, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1 = \\mathcal{C}_2 = \\dots = \\mathcal{C}_T" + }, + { + "bbox": [ + 130, + 380, + 482, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 503, + 304, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 304, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 304, + 514 + ], + "type": "text", + "content": "3.2 Framework and Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Figure 2 illustrates the framework of the Continual Adapter (C-ADA) approach. The C-ADA incorporates the S&S module and the plug-and-play Continual Adapter Layer (CAL) into the pre-trained model, endowing it with robust capabilities for addressing continual learning challenges. Concretely, the S&S module is exclusively trained before the incremental training to mitigate the domain gap and is subsequently frozen during further training. The CAL comprises a parameter-extensible down-projection, a ReLU activation, and a parameter-extensible up-projection. During incremental learning, the down-projection and up-projection expand the trainable weights to learn the new knowledge and freeze the previous weights to maintain the old knowledge. Moreover, we propose an orthogonal loss to alleviate the conflict between previous weights and trainable weights in each CAL." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Continual Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 113, + 465, + 277 + ], + "blocks": [ + { + "bbox": [ + 149, + 113, + 465, + 277 + ], + "lines": [ + { + "bbox": [ + 149, + 113, + 465, + 277 + ], + "spans": [ + { + "bbox": [ + 149, + 113, + 465, + 277 + ], + "type": "image", + "image_path": "791ed5520d531daa20d9b6d20a5396cb1cdf36a3aee58e9740e233aae1e50028.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": "Fig. 2: The framework of our C-ADA approach. For simplicity, we omit the skip connection and Layernorm in the figure. or loss represents the orthogonal loss. We attach our S&S and CAL, which are in parallel with the projection layer and MLP, to the shallow N layers of ViT and freeze the pre-trained backbone. For each new task, we expand two trainable weights in the CAL to learn the new knowledge and freeze the previous weights. Different from prior works, our approach uses a novel adapter variant to eliminate the necessity for key-query matching. Only trainable weights and classifier parameters are optimized which is parameter efficient and no old information (old images or prototypes) are stored which is privacy preserving." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": "The basic backbone consists of a pre-trained vision transformer feature extractor " + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "inline_equation", + "content": "f = f(\\cdot ;\\theta)" + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": " and a classifier " + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "inline_equation", + "content": "g = g(\\cdot ;\\phi)" + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": ". Given an input image " + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": ". The image is reshaped and projected by the patch embedding layer, resulting in the embedding feature " + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "inline_equation", + "content": "x_{e} \\in \\mathbb{R}^{L\\times D}" + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": " is the token length, and " + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": " is the embedding dimension. These embedding features are not only input to the pre-training module within the ViT block but also to the parallel CAL and S&S modules. The output of the backbone and parallel is added and fed into the next layer. Finally, the output features " + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": " from the last ViT block are fed into the " + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": " to perform classification tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 529, + 279, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 279, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 279, + 540 + ], + "type": "text", + "content": "3.3 Scale and Shift Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "content": "Due to the significant domain gap between the pre-training dataset and the downstream dataset, our C-ADA approach introduces a flexible and independent module, named scaling and shifting (S&S), to revise the feature space before incremental training for more efficient fine-tuning (higher plasticity)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "content": "Concretely, the S&S module includes the trainable parameter " + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "content": " as the scale and shift factors, respectively, to learn the domain information of downstream tasks. The S&S operation is calculated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 274, + 654, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 654, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 274, + 654, + 481, + 666 + ], + "type": "interline_equation", + "content": "y = \\alpha \\odot x + \\beta \\tag {1}", + "image_path": "40d2c1418301aefe215607be2bc38c943ff10023712df96eb1d58c231ce999d1.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": " is the dot product. We employ the S&S operation before the incremental training by using the data from the first phase (i.e. " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": "). Then we freeze the trainable parameters during the incremental phases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 169, + 286, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 169, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 131, + 169, + 286, + 182 + ], + "type": "text", + "content": "3.4 Continual Adapter Layer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "text", + "content": "Structure of Continual Adapter Layer To address the challenges of efficiency and lightweight, the CAL is designed as a bottleneck module. It consists of a parameter-extensible down-projection, a ReLU layer for non-linear property, and a parameter-extensible up-projection. The down-projection embeds the old and new knowledge into a latent embedding along with the ReLU function, and the up-projection extracts a dimension-fixed knowledge representation from the latent embedding. For each new task, we first initialize and append two learnable linear weights to the down-projection and up-projection, respectively, and freeze the previous weights. Such design ensures that the CAL is extendable for learning new knowledge and stable in maintaining old knowledge, maintaining the input and output feature dimensions unchanged. Assuming the current task " + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "text", + "content": ", we assign the weights in the CAL into two types: the frozen weights " + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{1},\\dots,\\mathbf{W}^{(t - 1)}" + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "text", + "content": ", and the trainable weight " + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^t" + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^i" + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "text", + "content": " represents both the " + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{dp}^i" + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{up}^i" + }, + { + "bbox": [ + 130, + 189, + 482, + 380 + ], + "type": "text", + "content": "). The frozen weights effectively preserve the previous knowledge to mitigate catastrophic forgetting, while simultaneously optimizing the trainable parameters to learn new knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 382, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 415 + ], + "type": "text", + "content": "Without loss of generality, we define the input and output of the current layer as " + }, + { + "bbox": [ + 130, + 382, + 482, + 415 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 382, + 482, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 382, + 482, + 415 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 382, + 482, + 415 + ], + "type": "text", + "content": ", respectively. In task t, we can adapt the input by the CAL as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 415, + 372, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 415, + 372, + 429 + ], + "spans": [ + { + "bbox": [ + 241, + 415, + 372, + 429 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {d p} = \\left[ \\begin{array}{c} \\mathbf {W} _ {d p} ^ {1}, \\dots , \\mathbf {W} _ {d p} ^ {t - 1}, \\mathbf {W} _ {d p} ^ {t} \\end{array} \\right],", + "image_path": "ee55bbd0d9255299bec14878e7d45b8b8944c072614709ea8af2ef7ef64e3d48.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 241, + 432, + 480, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 432, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 241, + 432, + 480, + 462 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {W} _ {u p} = \\left[ \\mathbf {W} _ {u p} ^ {1}, \\dots , \\mathbf {W} _ {u p} ^ {t - 1}, \\mathbf {W} _ {u p} ^ {t} \\right], \\\\ \\mathbf {W} _ {d p} ^ {i} \\in \\mathbb {R} ^ {D \\times d} \\quad \\mathbf {W} _ {u p} ^ {i} \\in \\mathbb {R} ^ {d \\times D} \\end{array} \\tag {2}", + "image_path": "4b7bafaf79722d9619b228b3a7784299a830280b3e061f71fd05302ebf8313c5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 241, + 466, + 361, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 466, + 361, + 479 + ], + "spans": [ + { + "bbox": [ + 241, + 466, + 361, + 479 + ], + "type": "interline_equation", + "content": "x _ {i} ^ {\\prime} = \\operatorname {R e L U} (x _ {i} \\cdot \\mathbf {W} _ {d p}) \\cdot \\mathbf {W} _ {u p}", + "image_path": "01dec606fc30860cdfc2ca467ac58cdb57808ffe8a1f24ca67264ed20dbda218.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "text", + "content": "In the next task, we freeze the " + }, + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{dp}^{t}" + }, + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{up}^{t}" + }, + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "text", + "content": " to maintain the old knowledge and convert two new weights " + }, + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{dp}^{t + 1}" + }, + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{up}^{t + 1}" + }, + { + "bbox": [ + 131, + 485, + 481, + 514 + ], + "type": "text", + "content": " to the trainable weights." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": "Parameters of Continual Adapter Layer. In this paper, we follow the standard RFCL setting where the total task number is known before training. To compare fairly, we adjust the value of " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": " to maintain a consistent number of parameters across different settings. For real-world applications, we can keep the size of " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": " constant and continually expand the specific weights for new tasks. Remarkably, even when " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": " is set to a minimal value, such as " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "d = 1" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": ", it still exhibits commendable performance. Therefore C-ADA is lightweight, and memory-efficient in real-world applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Position of Continual Adapter Layer. As depicted in [9,43], the Multi-Head Self-Attention (MHSA) block employs the attention mechanism to capture global" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Continual Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 235 + ], + "type": "text", + "content": "information, while the MLP block prevents the ViT output from degeneration. Both components play pivotal roles in the ViT structure. Consequently, we investigate performing a comprehensive tuning of both the MHSA block and MLP block, instead of exclusively emphasizing one over the other. To fine-tune the MHSA block, we choose to insert the CAL in parallel with the Projection Layer instead of the Attention module. This decision is motivated by the observation that, despite various advanced transformer-based models using different attention mechanisms within the MHSA block [12, 26, 28], they consistently include a projection layer. On the contrary, we insert the CAL in parallel with the entire MLP block to fine-tune the MLP block." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 271 + ], + "type": "text", + "content": "Without loss of generality, we define the input as " + }, + { + "bbox": [ + 130, + 236, + 482, + 271 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 130, + 236, + 482, + 271 + ], + "type": "text", + "content": " and initially feed it into the Attention module to obtain " + }, + { + "bbox": [ + 130, + 236, + 482, + 271 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 130, + 236, + 482, + 271 + ], + "type": "text", + "content": ", which serves as the input to the Projection Layer and " + }, + { + "bbox": [ + 130, + 236, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathrm{CAL}_1" + }, + { + "bbox": [ + 130, + 236, + 482, + 271 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 277, + 480, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 277, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 211, + 277, + 480, + 333 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} x _ {a} = \\operatorname {A t t e n} \\left(x _ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q \\cdot K ^ {T}}{\\sqrt {d}}\\right) V, \\\\ x _ {l} = \\operatorname {P r e s i} \\left(x _ {l}\\right) = x _ {l} = \\lambda_ {1} \\operatorname {C A L} \\left(S _ {l} - S \\left(x _ {l}\\right)\\right) \\end{array} \\tag {3} \\\\ x _ {l 1} = \\operatorname {P r o j} (x _ {a}), \\quad x _ {l 2} = \\lambda \\cdot \\operatorname {C A L} _ {1} (S \\& S (x _ {a})), \\\\ x _ {l} ^ {\\prime} = x _ {l 1} + x _ {l 2}, \\\\ \\end{array}", + "image_path": "f06d1acbc91c460c6a0d354bb76097b6aea919044f88a861448b21b0e4b37e37.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 341, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 481, + 365 + ], + "type": "text", + "content": "Then the output " + }, + { + "bbox": [ + 130, + 341, + 481, + 365 + ], + "type": "inline_equation", + "content": "x_{l}^{\\prime}" + }, + { + "bbox": [ + 130, + 341, + 481, + 365 + ], + "type": "text", + "content": " is further sent to the MLP block and " + }, + { + "bbox": [ + 130, + 341, + 481, + 365 + ], + "type": "inline_equation", + "content": "\\mathrm{CAL}_2" + }, + { + "bbox": [ + 130, + 341, + 481, + 365 + ], + "type": "text", + "content": ". This process is formally formulated as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 372, + 480, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 372, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 211, + 372, + 480, + 398 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {l 1} = \\operatorname {M L P} \\left(x _ {l} ^ {\\prime}\\right), \\quad x _ {l 2} = \\lambda \\cdot \\operatorname {C A L} _ {2} (S \\& S \\left(x _ {l} ^ {\\prime}\\right)), \\tag {4} \\\\ x _ {l + 1} = x _ {l 1} + x _ {l 2}, \\\\ \\end{array}", + "image_path": "ff6f373b5038ad157025360230e0017a6e6c5afce2313663330748ae4f9370a1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathrm{CAL}_1" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathrm{CAL}_2" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": " are in parallel with the projection layer and MLP block, respectively. Proj represents the Projection Layer and " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": " represents a scale factor (typically set to 0.1). Then, the output " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "x_{l + 1}" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": " is sent to the next layer and repeats the above process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 467, + 231, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 231, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 231, + 479 + ], + "type": "text", + "content": "3.5 Loss Function" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 484, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 581 + ], + "type": "text", + "content": "Orthogonal Loss. When training the new task, the learned new knowledge will interfere with old knowledge, potentially resulting in catastrophic forgetting. Thus we propose to solve this problem by initialization and loss function. We apply the Gram-Schmidt process to initialize the trainable parameters at the start of each new task. Orthogonal initialization contributes to the diminishment of the scope when the model parameter is updated, thereby fostering enhanced stability and overall performance. After the first task, we employ the orthogonal loss to keep trainable weights orthogonal to previous weights." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 160, + 585, + 480, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 585, + 480, + 602 + ], + "spans": [ + { + "bbox": [ + 160, + 585, + 480, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o r} = \\left\\| \\mathbf {W} _ {d p} ^ {t} \\right. ^ {T} \\cdot \\left\\{\\mathbf {W} _ {d p} ^ {1}, \\dots , \\mathbf {W} _ {d p} ^ {t - 1} \\right\\} \\| _ {2} + \\left\\| \\mathbf {W} _ {u p} ^ {t} \\right. \\cdot \\left\\{\\mathbf {W} _ {u p} ^ {1}, \\dots , \\mathbf {W} _ {u p} ^ {t - 1} \\right\\} ^ {T} \\| _ {2} \\tag {5}", + "image_path": "1988ac703bec75be57a57e49b124c157c059f69f7cd9b0d44e4d5cbfab7a7125.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "-norm and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " represents the matrix transpose. The loss function forces the model to optimize current weight " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^t" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " to be orthogonal to previous weight " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^1, \\dots, \\mathbf{W}^{t-1}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". Intuitively, orthogonal vectors exhibit reduced mutual influence. As a result, this design aims to minimize interference between new and old knowledge to reduce catastrophic forgetting." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "Classification Loss. Similar to the previous methods, we use the CE loss to optimize the model to learn the new knowledge" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 149, + 481, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 149, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 225, + 149, + 481, + 163 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = \\mathcal {L} (g (f (\\boldsymbol {x}; \\mathbf {W} _ {d p}, \\mathbf {W} _ {u p}, \\theta); \\phi), y) \\tag {6}", + "image_path": "9a86abbd1a3c0a08f15dfa7ff92abc45f4183b62607f19720a4a7fa8fc753fc8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "text", + "content": " is the input image and " + }, + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "text", + "content": " is the corresponding label. " + }, + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "text", + "content": " is the parameters of the classifier and " + }, + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 171, + 479, + 195 + ], + "type": "text", + "content": " is the parameters of the pre-train model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 212, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 212, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 481, + 236 + ], + "type": "text", + "content": "Total Optimization Objective. For task " + }, + { + "bbox": [ + 131, + 212, + 481, + 236 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 212, + 481, + 236 + ], + "type": "text", + "content": ", our full optimization consists of a classification loss " + }, + { + "bbox": [ + 131, + 212, + 481, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 131, + 212, + 481, + 236 + ], + "type": "text", + "content": " and an orthogonal loss " + }, + { + "bbox": [ + 131, + 212, + 481, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{or}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 255, + 245, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 245, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 255, + 245, + 481, + 266 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {W} _ {\\mathrm {d p}} ^ {t}, \\mathbf {W} _ {\\mathrm {u p}} ^ {t}, \\phi} \\mathcal {L} _ {c e} + \\delta \\mathcal {L} _ {\\mathrm {o r}} \\tag {7}", + "image_path": "cbdbac302be201fa3d88afc35bbd3becbef6558b711fad2eb90fd45cad698998.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 274, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 274, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 479, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 274, + 479, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{dp}^{t},\\mathbf{W}_{up}^{t}" + }, + { + "bbox": [ + 131, + 274, + 479, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 274, + 479, + 310 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 131, + 274, + 479, + 310 + ], + "type": "text", + "content": " are the learnable parameters of task t. We set the " + }, + { + "bbox": [ + 131, + 274, + 479, + 310 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 131, + 274, + 479, + 310 + ], + "type": "text", + "content": " to 1 for all the experiments. During each training phase, our models can be optimized in an end-to-end fashion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 328, + 230, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 328, + 230, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 230, + 342 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 354, + 479, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 479, + 389 + ], + "type": "text", + "content": "In this paper, we mainly aim to solve the rehearsal-free class-incremental learning problem. To evaluate our methods, we follow the RFCL settings proposed in previous works [38, 50, 51] and conduct comprehensive experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "type": "text", + "content": "Besides, to prove the robustness and generality of our method on other continual learning settings, we conduct experiments compared with the SOTA methods [49] on the domain-incremental learning settings. Extensive ablation studies prove the effectiveness of our methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 454, + 279, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 279, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 279, + 466 + ], + "type": "text", + "content": "4.1 Evaluation Benchmarks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 473, + 479, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 479, + 509 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 479, + 509 + ], + "type": "text", + "content": "Datasets. We use the Split CIFAR-100 [24] and Split ImageNet-R [17] for class-incremental setting, CORe50 [27] and DomainNet [30] for domain-incremental setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 510, + 479, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 479, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 479, + 545 + ], + "type": "text", + "content": "CIFAR-100 is a widely-used benchmark in continual learning, which contains 60000 images of " + }, + { + "bbox": [ + 131, + 510, + 479, + 545 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 131, + 510, + 479, + 545 + ], + "type": "text", + "content": " size from 100 classes. We split CIFAR-100 into the 10-task benchmark following previous works." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 545, + 479, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 479, + 593 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 479, + 593 + ], + "type": "text", + "content": "ImageNet-R has 200 classes and includes newly collected data of different styles. It provides a fair and challenging problem setting due to the significant gap between training data and pre-trained data. We split the dataset into the 5-task benchmark, the original 10-task benchmark and the 20-task benchmark." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 594, + 479, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 479, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 479, + 629 + ], + "type": "text", + "content": "CORe50 has 50 classes from 11 distinct domains. Following the domain-incremental setting, we use 8 domains for (120000 images) for continual training and the rest of the domains for testing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "content": "DomainNet is a large dataset with 345 classes and 600000 images. These images are split into 6 domains (Clipart, Infograph, Paint, Quickdraw, Real, and Sketch) for continual learning like S-Prompts [49]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Continual Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 179, + 470, + 299 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Results (\\%) on ImageNet-R and CIFAR-100. The results are all obtained by CODA [38] directly. P represents the total number of tasks. UB represents the full fine-tuning result from [10]. We report " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "A_N(\\%)" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "Param(\\%)" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "FLOPs(\\mathrm{G})" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " to bring intuitive comparison. C-ADA uses fewer parameters and doubles the training speed, greatly surpassing other methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 179, + 470, + 299 + ], + "lines": [ + { + "bbox": [ + 142, + 179, + 470, + 299 + ], + "spans": [ + { + "bbox": [ + 142, + 179, + 470, + 299 + ], + "type": "table", + "html": "
MethodsImageNet-RCIFAR-100FLOPs
P=5P=10P=20P=10
AN(↑)Param(↓)AN(↑)Param(↓)AN(↑)Param(↓)AN(↑)Param(↓)
UB77.13100/10077.13100/10077.13100/10089.30100/100-
ER (5000)71.72100/10064.43100/10052.43100/10076.20100/100-
FT18.74100/10010.12100/1004.75100/1009.92100/100-
FT++60.42100/10048.93100/10035.98100/10049.91100/100-
LwF.MC74.56100/10066.73100/10054.05100/10064.83100/100-
L2P70.830.7/100.769.290.7/100.765.890.7/100.782.500.7/100.735.2
Deep L2P73.939.6/109.671.669.6/109.668.429.6/109.684.309.6/109.635.2
DualPrompt73.050.5/100.571.320.8/100.867.871.3/101.383.050.8/100.835.2
CODA-P-S75.190.7/100.773.930.7/100.770.530.7/100.784.590.6/100.635.2
C-ADA77.930.7/100.776.660.7/100.773.470.7/100.787.180.6/100.617.6
", + "image_path": "3a198af5df8c2815e472c8543763db7e12191fde06bc950a90e38da51c27fce0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "content": "Comparing Methods. We compare our method with several baselines and state-of-the-art (SOTA) continual learning methods. For a fair comparison, all the methods are based on a ViT-B/16 backbone pre-trained on ImageNet-1K. Under class-incremental settings, we regard L2P [51], DualPrompt [50], and CODA [38] as the main competitors, which are rehearsal-free methods. Under domain-incremental settings, despite S-Prompt [49] storing the prototypes of previous domains instead of the rehearsal-free method, it is still the main comparison method as the SOTA of domain-incremental learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "text", + "content": "Implementation Details. Similar to CODA [38], we use the Adam [22] optimizer with " + }, + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "text", + "content": ", and a batch size of 128 images in four GPUs. For every image, we resize it to " + }, + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "text", + "content": " and normalize them to [0,1]. The learning rate is set to " + }, + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "inline_equation", + "content": "5e^{-5}" + }, + { + "bbox": [ + 130, + 430, + 482, + 574 + ], + "type": "text", + "content": " for all the settings with cosine-decaying. For fair comparisons, we set the total middle dimension to 60 (sum of two CALs) and insert the CAL into layers 0-4, which achieves a similar parameter amount to previous works [38, 50]. We train the CIFAR-100 for 20 epochs, ImageNet-R for 50 epochs, CORe50 and DomainNet for 10 epoch. During the training phase, we substitute predictions from previous-task logits with negative infinity during the training of a new task. This design leads to a softmax prediction of \"0\" for these previous task classes and ensures that gradients do not affect the linear heads of past task classes like previous works. [38, 50, 51]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 589, + 228, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 589, + 228, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 589, + 228, + 600 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 481, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 632 + ], + "type": "text", + "content": "Class-incremental learning. We report the results on two class-incremental datasets in Table 1 following the previous work [38]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Gao et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "text", + "content": "3 For the sake of fairness, we primarily selected CODA-P-S with the same parameter count as C-ADA for comparison." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 179, + 281, + 342 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Results (\\%) on CORe50 and DomainNet. " + }, + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "type": "text", + "content": " gives the accuracy averaged over tasks and buffer size represents the number of old exemplars stored in the memory buffer. The results are all obtained by S-Prompt [49] directly. We reproduce the CODA on these settings for further comparison. We ensure that all methods have the same pre-trained model (ViT-B/16)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 179, + 281, + 342 + ], + "lines": [ + { + "bbox": [ + 150, + 179, + 281, + 342 + ], + "spans": [ + { + "bbox": [ + 150, + 179, + 281, + 342 + ], + "type": "table", + "html": "
MethodBuffer size\\(A_N(↑)\\)
Upper-Bound-84.01
ER80.10
GDumb74.92
BiC79.28
DER++50/class79.70
Co2L79.75
DyTox79.21
L2P81.07
EWC74.82
LwF75.45
L2P0/class78.33
CODA-P-S85.41
S-Prompts83.13
C-ADA89.45
", + "image_path": "8eae0fae0d65d7cf67346d7874826bc9c0bf3d6c0f7793c65a6073d07e5fcddf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 315, + 189, + 482, + 331 + ], + "blocks": [ + { + "bbox": [ + 315, + 189, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 315, + 189, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 315, + 189, + 482, + 331 + ], + "type": "table", + "html": "
MethodBuffer sizeAN(↑)
Upper-Bound-63.22
DyTox50/class62.94
EWC47.62
LwF49.19
SimCLR44.20
BYOL49.70
Barlow Twins0/class48.90
Supervised Contrastive50.90
L2P40.15
CODA-P-S47.56
S-Prompts50.62
C-ADA53.00
", + "image_path": "74060bde1773f12663c5c8d099cafcce386907d682122a303d4d3b47b536dcf1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "text", + "content": "ImageNet-R: We report the results with 5-tasks, 10-tasks, and 20-tasks to evaluate the robustness of the model under different settings in Table 1. It is worth noting that our method significantly surpasses the previous methods in average accuracy under different settings with as much as " + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "inline_equation", + "content": "+2.74\\%" + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "inline_equation", + "content": "+2.73\\%" + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "inline_equation", + "content": "+2.94\\%" + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "text", + "content": " over CODA, " + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "inline_equation", + "content": "+4.88\\%" + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "inline_equation", + "content": "+5.41\\%" + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "inline_equation", + "content": "+5.60\\%" + }, + { + "bbox": [ + 130, + 361, + 482, + 481 + ], + "type": "text", + "content": " over DualPrompt (second best method). This phenomenon shows that our method is robust to different class increment settings. Moreover, since our method only requires forward propagation one time. We achieved an obvious increase in training speed while improving accuracy. Combining accuracy and training speed brings better practicality in the real world." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "content": "CIFAR-100: We note that our approach also brings similar results to the ImageNet-R benchmarks, with an improvement of " + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "inline_equation", + "content": "+2.59\\%" + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "content": " compared with CODA, " + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "inline_equation", + "content": "+4.13\\%" + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "content": " compared with DualPrompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 495, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 495, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 495, + 569 + ], + "type": "text", + "content": "Domain-incremental learning. We report the results on two domain-incremental datasets in Table 2. Besides, we reproduced CODA on these datasets to bring intuitive comparison." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "CORe50 and DomainNet are widely used domain dataset for continual learning. Following the setting of S-Prompt [49], we divide the baseline methods into two groups, exemplars-based methods [2-4, 11, 31, 52] and exemplars-free methods [49]/rehearsal-free methods [38, 51]. Exemplars-based methods store 50 old exemplars per class in the memory buffer. These methods often have better accuracy but are accompanied by the risk of privacy leakage and memory overhead. Our method surpasses the CODA by " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "+4.04\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "+5.44\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " on CORe50 and DomainNet benchmark. Besides, it is worth noting that our method even out" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Continual Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 113, + 293, + 203 + ], + "blocks": [ + { + "bbox": [ + 140, + 113, + 293, + 203 + ], + "lines": [ + { + "bbox": [ + 140, + 113, + 293, + 203 + ], + "spans": [ + { + "bbox": [ + 140, + 113, + 293, + 203 + ], + "type": "image", + "image_path": "b9d7987e367c2b72787914d04d9403be925148a5712b7e56136f7e56603427c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 220, + 301, + 254 + ], + "lines": [ + { + "bbox": [ + 132, + 220, + 301, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 301, + 254 + ], + "type": "text", + "content": "Fig. 3: Average accuracy " + }, + { + "bbox": [ + 132, + 220, + 301, + 254 + ], + "type": "inline_equation", + "content": "A_{N}" + }, + { + "bbox": [ + 132, + 220, + 301, + 254 + ], + "type": "text", + "content": " (\\%) vs tuning parameters (\\%). We report the results on the 20-tasks ImageNet-R." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 328, + 125, + 470, + 178 + ], + "blocks": [ + { + "bbox": [ + 328, + 125, + 470, + 178 + ], + "lines": [ + { + "bbox": [ + 328, + 125, + 470, + 178 + ], + "spans": [ + { + "bbox": [ + 328, + 125, + 470, + 178 + ], + "type": "table", + "html": "
MethodCIFAR-100ImageNet-R
C-ADA87.3276.66
Ablate S&S86.1175.93
Ablate Lor82.8972.17
Ablate CAL76.9364.45
", + "image_path": "aba5c5f980a5b2d6d199b9b77661f0d1878c833e2d372d348ab9aa2e0f244cef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 311, + 186, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 311, + 186, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 311, + 186, + 482, + 243 + ], + "type": "text", + "content": "Table 3: Ablation Results (%) on 10-task ImageNet-R and CIFAR-100. " + }, + { + "bbox": [ + 311, + 186, + 482, + 243 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 311, + 186, + 482, + 243 + ], + "type": "text", + "content": " gives the accuracy averaged over tasks. We ablate the key components in turn and report the results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 277, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 482, + 326 + ], + "type": "text", + "content": "performs S-Prompt, which is specifically designed for domain-incremental learning, with improvements of " + }, + { + "bbox": [ + 132, + 277, + 482, + 326 + ], + "type": "inline_equation", + "content": "+6.32\\%" + }, + { + "bbox": [ + 132, + 277, + 482, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 277, + 482, + 326 + ], + "type": "inline_equation", + "content": "+2.38\\%" + }, + { + "bbox": [ + 132, + 277, + 482, + 326 + ], + "type": "text", + "content": ". This interesting phenomenon demonstrates the robustness and effectiveness of our method in various continual learning settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 343, + 295, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 295, + 355 + ], + "type": "text", + "content": "4.3 Key Components Ablation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 363, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 544 + ], + "type": "text", + "content": "We conducted ablation experiments to analyze the importance of each component. We incrementally introduced key components on both the 10-tasks ImageNet-R and CIFAR-100 datasets, and the results are presented in Table 3. Firstly, we ablate the S&S module, the performance has dropped " + }, + { + "bbox": [ + 130, + 363, + 482, + 544 + ], + "type": "inline_equation", + "content": "(0.73\\% \\downarrow" + }, + { + "bbox": [ + 130, + 363, + 482, + 544 + ], + "type": "text", + "content": " on ImageNet-R). This shows that revising the feature space before training is beneficial to improving subsequent learning ability. Secondly, removing the orthogonal loss leads to severe performance degradation on average accuracy " + }, + { + "bbox": [ + 130, + 363, + 482, + 544 + ], + "type": "inline_equation", + "content": "(3.76\\% \\downarrow" + }, + { + "bbox": [ + 130, + 363, + 482, + 544 + ], + "type": "text", + "content": " on ImageNet-R). This outcome aligns with our expectations, as ablating the orthogonal loss results in our method lacking a mechanism to prevent forgetting, leading to severe catastrophic forgetting. Thirdly, when we ablated the CAL component, we observed a noticeable drop in performance " + }, + { + "bbox": [ + 130, + 363, + 482, + 544 + ], + "type": "inline_equation", + "content": "(7.72\\% \\downarrow" + }, + { + "bbox": [ + 130, + 363, + 482, + 544 + ], + "type": "text", + "content": " on ImageNet-R). The removal of the CAL implies that our model relies solely on adjusting the classifier to tune the model, which obviously leads to unsatisfactory results, particularly when dealing with the downstream datasets having a significant gap from the pre-trained dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 561, + 309, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 309, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 309, + 574 + ], + "type": "text", + "content": "4.4 Scaling Trainable Parameters" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 605 + ], + "type": "text", + "content": "We conduct experiments to discuss the relationship between the trainable parameters amount and the performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The trainable parameters amount can be adjusted by changing the middle dimension of the CAL. The total number of middle dimensions of the two CALs in each block is chosen from " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{20, 40, 60, 80, 100, 120, 140, 160\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", corresponding to " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{0.35\\%, 0.54\\%, 0.71\\%, 0.90\\%, 1.12\\%, 1.27\\%, 1.45\\%, 1.61\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " trainable parameters. We report the results on 20-tasks ImageNet-R in Figure 3. The figure" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "shows that more trainable parameters contribute to achieving a higher accuracy. Moreover, we can notice that even with the fewer trainable parameters, our model consistently demonstrates superior performance when compared to CODA (71.92% " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", 0.35% " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\text{Param}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " of C-ADA vs 70.53% " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", 0.7% " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\text{Param}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " of CODA). This fully illustrates that our approach improves the plasticity compared with previous works." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 309, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 309, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 309, + 217 + ], + "type": "text", + "content": "4.5 Attaching Layer and Position" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": "Intuitively, each layer of the ViT exhibits distinct feature extraction capabilities [50]. The impact of tuning different layers may be quite different in CL. Furthermore, each module of the layer plays different roles, which causes attaching CAL to various positions within the layer may yield divergent outcomes. Therefore, it is crucial to explore layers and positions to attach the CAL under the CL settings. Note that the total number of trainable parameters remains consistent across all experiments described below (middle " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\dim = d*t" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 325, + 481, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 348 + ], + "type": "text", + "content": "Layer of CAL. For simplicity, we assume that the attaching layers are contiguous. We report the result in Table 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "type": "text", + "content": "Firstly, we attach the CAL to all the layers to get a preliminary result (Layer 0 to Layer 11). Next, we use binary search to analyze the importance of shallow (Layer 0 to Layer 5) and deep (Layer 6 to Layer 11) layers. We observe that attaching to shallow layers gets superior results compared to attaching to deeper layers. This phenomenon is consistent with the observation from previous work [50]. Then we continue to reduce the number of attached layers in shallow layers. Experimental results show that attaching from layer " + }, + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "type": "text", + "content": " to layer " + }, + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "type": "inline_equation", + "content": "= 4" + }, + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "type": "text", + "content": " performs the best. We use this conclusion to analyze the attached modules." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 509 + ], + "type": "text", + "content": "Position of CAL. By analyzing common parts of the ViT architecture, we select three alternative attaching positions: QKV attention layer " + }, + { + "bbox": [ + 130, + 461, + 481, + 509 + ], + "type": "inline_equation", + "content": "(W_{q}, W_{k}, W_{v})" + }, + { + "bbox": [ + 130, + 461, + 481, + 509 + ], + "type": "text", + "content": " in the MHSA block, projection layer in the MHSA block, and the MLP block. We report the result in Table 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": "Firstly, we analyze the effects of only attaching these three positions, respectively. We observe that attaching the CAL to the projection layer has a better effect than in the QKV attention layer. This phenomenon makes us attach the CAL to the projection layer to tune the MHSA block. Then we opt to combine the tuning of different positions. We observe that attaching the CAL to different positions resulted in certain improvements (about " + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": "). Interestingly, it seems to have little to do with how to assign the middle dimensions. This shows to some extent that our method is robust." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 622, + 289, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 289, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 289, + 633 + ], + "type": "text", + "content": "4.6 Extra Evaluation Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Although we emphasize that " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\text{Param}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "FLOPs" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " are the more important metrics, there are still some common metrics in the CL field." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Continual Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 184, + 290, + 269 + ], + "blocks": [ + { + "bbox": [ + 132, + 119, + 301, + 175 + ], + "lines": [ + { + "bbox": [ + 132, + 119, + 301, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 119, + 301, + 175 + ], + "type": "text", + "content": "Table 4: Results " + }, + { + "bbox": [ + 132, + 119, + 301, + 175 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 132, + 119, + 301, + 175 + ], + "type": "text", + "content": " on ImageNetR 10 tasks. We analyze the difference of attaching to different layers. We keep the total tunable parameters constant by modifying the middle dimension," + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 184, + 290, + 269 + ], + "lines": [ + { + "bbox": [ + 141, + 184, + 290, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 184, + 290, + 269 + ], + "type": "table", + "html": "
Attach Layermiddle dimAN(↑)
Layer 0 ~ Layer 113075.23
Layer 6 ~ Layer 115072.51
Layer 0 ~ Layer 55076.21
Layer 1 ~ Layer 56076.29
Layer 0 ~ Layer 46076.66
Layer 2 ~ Layer 58075.72
Layer 0 ~ Layer 38075.75
", + "image_path": "1e28a3ef1f19537e72a26c748fdc20a9eaa1dc35a70eb99530131a29fd96071b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 326, + 179, + 463, + 274 + ], + "blocks": [ + { + "bbox": [ + 311, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 311, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 311, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 5: Results " + }, + { + "bbox": [ + 311, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 311, + 114, + 482, + 169 + ], + "type": "text", + "content": " on ImageNetR 10 tasks. We fix the attaching layers (layer 0 to layer 4) to analyze two aspects: the attaching positions and the middle dimension of each CAL." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 179, + 463, + 274 + ], + "lines": [ + { + "bbox": [ + 326, + 179, + 463, + 274 + ], + "spans": [ + { + "bbox": [ + 326, + 179, + 463, + 274 + ], + "type": "table", + "html": "
Attach Positionmiddle dimAN(↑)
QKV Attention6076.04
MLP6076.23
Proj6076.25
Proj + MLP20 + 4076.52
Proj + MLP30 + 3076.57
Proj + MLP50 + 1076.51
Proj + MLP10 + 5076.61
Proj + MLP40 + 2076.66
", + "image_path": "d7bf108739df39f5fda6ff6463e99af49293618680529f79cf58eea3f35294db.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 294, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 479, + 316 + ], + "type": "inline_equation", + "content": "F_{N}" + }, + { + "bbox": [ + 130, + 294, + 479, + 316 + ], + "type": "text", + "content": ": Average forgetting metrics, represent the drop in task performance averaged over " + }, + { + "bbox": [ + 130, + 294, + 479, + 316 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 294, + 479, + 316 + ], + "type": "text", + "content": " tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 318, + 420, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 318, + 420, + 329 + ], + "spans": [ + { + "bbox": [ + 146, + 318, + 420, + 329 + ], + "type": "text", + "content": "bwt: Backward transfer, assesses the quality of representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 354 + ], + "type": "text", + "content": "fwt: Forward transfer, represents the influence of existing knowledge on the performance of subsequent concepts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 481, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 401 + ], + "type": "text", + "content": "In Table 6, we can see that across various metrics, C-ADA demonstrates superior performance. Notably, in terms of " + }, + { + "bbox": [ + 130, + 354, + 481, + 401 + ], + "type": "inline_equation", + "content": "fwt" + }, + { + "bbox": [ + 130, + 354, + 481, + 401 + ], + "type": "text", + "content": ", which represents the capacity to learn new knowledge, it significantly outperforms the other methods. This emphasizes the effectiveness of C-ADA in learning new knowledge." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 157, + 453, + 455, + 521 + ], + "blocks": [ + { + "bbox": [ + 132, + 421, + 482, + 444 + ], + "lines": [ + { + "bbox": [ + 132, + 421, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 482, + 444 + ], + "type": "text", + "content": "Table 6: Results (\\%) on ImageNet-R. The " + }, + { + "bbox": [ + 132, + 421, + 482, + 444 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 132, + 421, + 482, + 444 + ], + "type": "text", + "content": " results are all obtained by CODA [38] directly. We reproduce the bwt and fwt metrics for further comparison." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 453, + 455, + 521 + ], + "lines": [ + { + "bbox": [ + 157, + 453, + 455, + 521 + ], + "spans": [ + { + "bbox": [ + 157, + 453, + 455, + 521 + ], + "type": "table", + "html": "
MethodsP=5P=10P=20P=10P=10
FN(↓)FN(↓)FN(↓)bwt(↑)fwt(↑)
L2P3.36 ± 0.182.03 ± 0.191.24 ± 0.14-3.69 ± 0.07-0.32 ± 0.04
DualPrompt2.64 ± 0.171.71 ± 0.241.07 ± 0.14-3.48 ± 0.06-0.09 ± 0.04
CODA2.65 ± 0.151.60 ± 0.201.00 ± 0.15-3.24 ± 0.060.01 ± 0.03
C-ADA2.55 ± 0.141.58 ± 0.181.02 ± 0.12-3.12 ± 0.050.34 ± 0.03
", + "image_path": "14a7aa13c7e7bbc7e10b500b1ed180df0decc24a28ede18a81714389beb46236.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 558, + 220, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 220, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 220, + 571 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 581, + 504, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 504, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 504, + 666 + ], + "type": "text", + "content": "In this paper, we present a novel adapter variant, named Continual Adapter (C-ADA), for RFCL. Our approach assembles a parameter-extensible adapter layer (CAL) and an S&S module in parallel with the pre-trained model. Importantly, C-ADA eliminates the necessity for key-query matching to achieve better plasticity and efficient training. We set a new SOTA on well-established benchmarks of RFCL and outperform the mainstream methods of DIL. This underscores the generality and robustness of our approach across various continual learning settings." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "content": "This work was funded by the National Natural Science Foundation of China under Grant No.U21B2048 and No.62302382 and Shenzhen Key Technical Projects under Grant CJGJZD2022051714160501 and the CAAI-MindSpore Open Fund, developed on OpenI Community." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV). pp. 139-154 (2018) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "text", + "content": "2. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems 33, 15920-15930 (2020) 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 326 + ], + "type": "text", + "content": "3. Cha, H., Lee, J., Shin, J.: Co2l: Contrastive continual learning. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 9516-9525 (2021) 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 360 + ], + "type": "text", + "content": "4. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019) 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "text", + "content": "5. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "type": "text", + "content": "6. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "type": "text", + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "type": "text", + "content": "8. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021) 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "text", + "content": "9. Dong, Y., Cordonnier, J.B., Loukas, A.: Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In: International Conference on Machine Learning. pp. 2793-2803. PMLR (2021) 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 568, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 480, + 609 + ], + "type": "text", + "content": "11. Douillard, A., Ramé, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9285-9295 (2022) 4, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 480, + 643 + ], + "type": "text", + "content": "12. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "13. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999) 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Continual Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "text", + "content": "15. Gao, X., He, Y., Dong, S., Cheng, J., Wei, X., Gong, Y.: Dkt: Diverse knowledge transfer transformer for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24236-24245 (2023) 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 225 + ], + "type": "text", + "content": "16. Gao, Z., Cen, J., Chang, X.: Consistent prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28463-28473 (2024) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 268 + ], + "type": "text", + "content": "17. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021) 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "text", + "content": "18. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 831-839 (2019) 1, 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "text", + "content": "19. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 342, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 376 + ], + "type": "text", + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 376, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 408 + ], + "type": "text", + "content": "21. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 429 + ], + "type": "text", + "content": "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 472 + ], + "type": "text", + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 494 + ], + "type": "text", + "content": "24. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., Citeseer (2009) 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "text", + "content": "25. Li, Z., Hoiem, D.: Learning without forgetting. T-PAMI 40(12), 2935-2947 (2018) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 514, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 548 + ], + "type": "text", + "content": "26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 548, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 480, + 568 + ], + "type": "text", + "content": "27. Lomonaco, V., Maltoni, D.: Core50: a new dataset and benchmark for continuous object recognition. In: Conference on robot learning. pp. 17-26. PMLR (2017) 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "28. Mao, X., Qi, G., Chen, Y., Li, X., Duan, R., Ye, S., He, Y., Xue, H.: Towards robust vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12042-12051 (2022) 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "29. McDonnell, M.D., Gong, D., Parvaneh, A., Abbasnejad, E., van den Hengel, A.: Ranpac: Random projections and pre-trained models for continual learning. Advances in Neural Information Processing Systems 36 (2024) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "30. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 9" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "31. Prabhu, A., Torr, P.H., Dokania, P.K.: Gdumb: A simple approach that questions our progress in continual learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 524-540. Springer (2020) 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 482, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 204 + ], + "type": "text", + "content": "32. Qin, Y., Chen, Y., Peng, D., Peng, X., Zhou, J.T., Hu, P.: Noisy-correspondence learning for text-to-image person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27197-27206 (2024) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 482, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 236 + ], + "type": "text", + "content": "33. Qin, Y., Peng, D., Peng, X., Wang, X., Hu, P.: Deep evidential learning with noisy correspondence for cross-modal retrieval. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4948-4956 (2022) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 482, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 268 + ], + "type": "text", + "content": "34. Qin, Y., Sun, Y., Peng, D., Zhou, J.T., Peng, X., Hu, P.: Cross-modal active complementary learning with self-refining correspondence. Advances in Neural Information Processing Systems 36 (2024) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 268, + 482, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 290 + ], + "type": "text", + "content": "35. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 1, 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 290, + 482, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 322 + ], + "type": "text", + "content": "36. Shieh, J.L., Haq, Q.M.u., Haq, M.A., Karam, S., Chondro, P., Gao, D.Q., Ruan, S.J.: Continual learning strategy in one-stage object detection framework based on experience replay for autonomous driving vehicle. Sensors 20(23), 6777 (2020) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 322, + 482, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 364 + ], + "type": "text", + "content": "37. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 364, + 482, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 397 + ], + "type": "text", + "content": "38. Smith, J.S., Karlinsky, L., et al.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: CVPR. pp. 11909-11919 (2023) 1, 2, 4, 5, 9, 10, 11, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 397, + 482, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 430 + ], + "type": "text", + "content": "39. Song, X., He, Y., Dong, S., Gong, Y.: Non-exemplar domain incremental object detection via learning domain bias. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 15056-15065 (2024) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 430, + 482, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 450 + ], + "type": "text", + "content": "40. Tang, Y.M., et al.: When prompt-based incremental learning does not meet strong pretraining. In: CVPR (2023) 2, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "text", + "content": "41. Tao, X., Chang, X., Hong, X., Wei, X., Gong, Y.: Topology-preserving class-incremental learning. In: European Conference on Computer Vision. pp. 254-270. Springer (2020) 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 483, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 525 + ], + "type": "text", + "content": "42. Tu, C.H., Mai, Z., Chao, W.L.: Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7725-7735 (2023) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 525, + 482, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 558 + ], + "type": "text", + "content": "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 558, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 590 + ], + "type": "text", + "content": "44. Verwimp, E., Yang, K., Parisot, S., Hong, L., McDonagh, S., Pérez-Pellitero, E., De Lange, M., Tuytelaars, T.: Clad: A realistic continual learning benchmark for autonomous driving. Neural Networks 161, 659-669 (2023) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 590, + 482, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 612 + ], + "type": "text", + "content": "45. Wang, F.Y., Zhou, D.W., Ye, H.J., Zhan, D.C.: Foster: Feature boosting and compression for class-incremental learning. arXiv preprint arXiv:2204.04662 (2022) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 612, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 632 + ], + "type": "text", + "content": "46. Wang, L., et al.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. NeurIPS (2024) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "47. Wang, S., Shi, W., Dong, S., Gao, X., Song, X., Gong, Y.: Semantic knowledge guided class-incremental learning. IEEE Transactions on Circuits and Systems for Video Technology 33(10), 5921-5931 (2023) 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 102 + ], + "type": "text", + "content": "Continual Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 478 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "48. Wang, S., Shi, W., He, Y., Yu, Y., Gong, Y.: Non-exemplar class-incremental learning via adaptive old class reconstruction. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4524-4534 (2023) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "49. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022) 4, 5, 9, 10, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "50. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799 (2022) 2, 4, 5, 9, 10, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "text", + "content": "51. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022) 1, 2, 4, 9, 10, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "52. Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., Fu, Y.: Large scale incremental learning. arXiv preprint arXiv:1905.13260 (2019) 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "53. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3014-3023 (2021) 1, 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "54. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International conference on machine learning. pp. 3987-3995. PMLR (2017) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "55. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "56. Zhou, D.W., et al.: Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. arXiv (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 434 + ], + "type": "text", + "content": "57. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5871-5880 (2021) 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 434, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 480, + 478 + ], + "type": "text", + "content": "58. Zhu, K., Zhai, W., Cao, Y., Luo, J., Zha, Z.J.: Self-sustaining representation expansion for non-exemplar class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9296-9305 (2022) 2, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_content_list.json b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..92546119f070beb8ba816ceaadfceb5231743740 --- /dev/null +++ b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_content_list.json @@ -0,0 +1,1649 @@ +[ + { + "type": "text", + "text": "Beyond Viewpoint: Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation", + "text_level": 1, + "bbox": [ + 217, + 140, + 782, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linlong Fan $^{\\text{日}}$ , Ye Huang $^{\\text{日}^{\\star}}$ , Yanqi Ge $^{\\text{日}}$ , Wen Li $^{\\text{日}}$ , and Lixin Duan $^{\\text{日}}$", + "bbox": [ + 253, + 234, + 750, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China fanlinlong703@163.com", + "bbox": [ + 308, + 262, + 692, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing view-based methods excel at recognizing 3D objects from predefined viewpoints, but their exploration of recognition under arbitrary views is limited. This is a challenging and realistic setting because each object has different viewpoint positions and quantities, and their poses are not aligned. However, most view-based methods, which aggregate multiple view features to obtain a global feature representation, hard to address 3D object recognition under arbitrary views. Due to the unaligned inputs from arbitrary views, it is challenging to robustly aggregate features, leading to performance degradation. In this paper, we introduce a novel Part-aware Network (PANet), which is a part-based representation, to address these issues. This part-based representation aims to localize and understand different parts of 3D objects, such as airplane wings and tails. It has properties such as viewpoint invariance and rotation robustness, which give it an advantage in addressing the 3D object recognition problem under arbitrary views. Our results on benchmark datasets clearly demonstrate that our proposed method outperforms existing view-based aggregation baselines for the task of 3D object recognition under arbitrary views, even surpassing most fixed viewpoint methods.", + "bbox": [ + 261, + 335, + 738, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D object recognition $\\cdot$ weakly-supervised learning", + "bbox": [ + 261, + 613, + 687, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 651, + 375, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning 3D with multi-view supervision has received widespread attention. It covers various research domains such as NeRFs [15,40], BEV detection [30,62], 3D generation [33], 3D reasoning [9,21], etc., among which 3D object recognition is considered one of the most fundamental problems. Thanks to mature and transferable 2D vision models [11,47] and large-scale image datasets [10], rendering 3D objects into multiple 2D images and applying 2D networks to indirectly recognize 3D objects has become the state-of-the-art method [57]. Furthermore, compared with direct methods (e.g., point-based [1,61], voxel-based [38]), view-based methods are closer to human perception of 3D objects in the real world.", + "bbox": [ + 212, + 681, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3e50804339c196a8b81f296bd240f8cc9be1db52975e43fe2dde2591e373aff.jpg", + "image_caption": [ + "Fig. 1: In practical scenarios, 3D objects are observed from arbitrary views(right). The definition of arbitrary views is that each object is unaligned, and the viewpoint positions and quantities vary for each object. Previous works have focused on 3D object recognition in aligned views(left), where the pose of each object is aligned, and the positions and quantities of viewpoints are predefined. Rotated views(middle) is an extension that introduces random rotation for each 3D object while keeping the viewpoint positions and quantities unchanged. Our work focuses on arbitrary views." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 784, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous view-based methods [5,6,49,50,56] focus on addressing 3D object recognition tasks under fixed viewpoints, while recognition under arbitrary views still lacks extensive research. As shown in Fig. 1, arbitrary views represent a more challenging setting where each object may have different viewpoint positions and quantities, and the poses of each object are not aligned.", + "bbox": [ + 212, + 354, + 787, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, most view-based methods [5,6,13,18,26,41,49,50,55-57,60,63-65], which aggregate multiple view features to obtain a global feature representation, hard to effectively address 3D object recognition under arbitrary views. This is because the misaligned nature of arbitrary views complicates the representation and aggregation of view features, which leads to performance degradation. For example, when applied to arbitrary views, the performance of CVR [55] drops by $6.31\\%$ and $14.23\\%$ in ModelNet [59] and ScanObjectNN [53] datasets. Even some multi-view aggregation methods [26, 56] cannot be applied to arbitrary views because they require fixed viewpoint position information. Furthermore, view aggregation methods [13, 49] typically jointly optimize feature extraction and aggregation modules to maximize performance, but this may lead the extraction module to overly focus on certain views while ignoring information from others, resulting in information loss.", + "bbox": [ + 212, + 431, + 787, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a novel part-based representation, as shown in Fig. 2. This idea is inspired by cognitive science and biology [20], which aims to understand different part features or specific regions in a multi-view sequence to better represent 3D objects. For example, a person can be composed of multiple parts with unique attributes, such as the head, arms, body, legs, and so on. By locating and integrating the features of these parts, we can more accurately represent a person. Compared with the multi-view aggregation features, part features offer several advantages. Firstly, an individual view contains abundant part information. Even if only a few views are utilized, we can still extract stable part features, thereby adapting to changes in the number of viewpoints in arbitrary view settings. Secondly, the part features are more stable than the multi-view aggregation features [37], which can better adapt to changes in object rotation and viewpoint positions. Therefore, the part-based representation method is robust and can better handle the challenge of 3D object recogni", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan", + "bbox": [ + 269, + 114, + 720, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a987c554f2db6f2aec74bd725338d32c96de92fb3f82578c570f2d509551cf4f.jpg", + "image_caption": [ + "Fig. 2: Comparison between multi-view aggregation representation (a) and part-based representation (b). View aggregation methods integrate features from different views and use the aggregated features to represent 3D objects. However, since the views are not aligned, the aggregated features lack robustness, making it difficult to handle 3D recognition on arbitrary views. Additionally, view aggregation methods suffer from information loss, thus unable to fully utilize the features from each view. We newly propose a part-based representation that focuses on multi-view part awareness and combines multiple parts to robustly represent 3D objects." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 782, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion under arbitrary views. We also provide comprehensive experimental results in Sec. 4 to demonstrate the effectiveness of part-based representation. When switching to arbitrary views, we do not observe significant performance degradation in our method, and it even outperforms most fixed viewpoint methods.", + "bbox": [ + 212, + 467, + 784, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, this paper proposes a novel part-aware network (PANet). Firstly, it employs weakly supervised learning to locate and recognize parts in each view, obtaining local view part features. Then, local view parts are input into the adaptive part refinement (APR) module. This module adaptively perceives the correlation of local view parts and further integrates these view parts to obtain global parts, which aims to adapt to the viewpoint changes of parts across different views. Finally, we combine these global parts to represent a 3D object. It is worth noting that we also introduce a regularization loss in the part localization module to ensure that part features can better focus on the object.", + "bbox": [ + 212, + 529, + 787, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The primary contributions of this work can be summarized as follows:", + "bbox": [ + 238, + 667, + 743, + 683 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Unlike existing view aggregation methods, we propose a novel part-based representation. We consider a 3D object as a collection of part features rather than an abstract multi-view aggregated feature. This part-based representation is not limited by the position and quantity of viewpoints or affected by object poses, making it more flexibly applicable to various viewpoint scenarios, especially in arbitrary views.", + "- Further, we newly develop a Part-Aware Network (PANet), for extracting parts in multiple views. The network employs a weakly supervised method to localize parts on each view, enabling it to obtain local view parts. Then, we further refines these view parts through the proposed Adaptive Part Refinement" + ], + "bbox": [ + 212, + 686, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Beyond Viewpoint", + "bbox": [ + 606, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(APR) module to obtain global parts. These global parts can better adapt to differences in parts across viewpoints. Finally, integrating these global parts forms a comprehensive description of a 3D object.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We conduct experiments on various datasets, including ScanObjectNN, ModelNet, and RGBD. Our method outperforms existing methods in both fixed views and arbitrary views. In particular, the performance in arbitrary views exceeds most existing fixed viewpoint methods based on fixed views, achieving robust recognition of 3D objects.", + "bbox": [ + 212, + 191, + 787, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Relative Work", + "text_level": 1, + "bbox": [ + 215, + 292, + 395, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D object Recognition methods: 3D object recognition using deep neural networks can be broadly divided into three main categories: point-based methods [17,42,43], voxel-based methods [28,38,46], and view-based methods. View-based methods render 3D objects into multiple views and utilize multiple views to jointly represent 3D objects. Benefiting from mature and transferable 2D vision models and large-scale image datasets, they can better handle 3D object recognition tasks.", + "bbox": [ + 212, + 325, + 782, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MVCNN [48] is a pioneering multi-view network for 3D object classification. It utilizes a view pooling strategy to integrate features from different views. Yang et al. [63] proposes a relational network to effectively connect corresponding regions from different viewpoints, thus enhancing the information of each view. 3D2Seq2views [18] method treats multiple views as a sequence and employs a convolutional neural network with novel hierarchical attention aggregation to aggregate sequential views. DRCNN [50] utilizes the dynamic routing algorithm of capsule networks to construct a new Dynamic Routing Layer (DRL), which integrates features from each view. View-GCN [56] employs multiple views as nodes in a view graph, leveraging a graph convolutional network (GCN) architecture to hierarchically aggregate information and learn robust 3D shape descriptors. These methods assume that each object has a predefined set of viewpoints and are trained and tested on these viewpoints. This assumption may limit model performance when dealing with 3D object recognition in arbitrary views. As far as we know, there are few studies [31, 34, 55] that have made significant progress beyond fixed viewpoint settings. The MVTN [34] implements viewpoint regression for 3D object recognition based on the latest research progress of differentiable rendering. The motivation of CVR [55] is to restore the intrinsic alignment of arbitrary views, transform it into aligned representations and then perform aggregation. In our work, we newly propose part-based representation. Part features exhibit viewpoint invariance and rotation robustness, making them more suitable for handling changes in viewpoint positions and quantities in arbitrary views. Moreover, the key idea of part-based representation is not to rely on multi-view aggregation, thus avoiding information loss.", + "bbox": [ + 212, + 431, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discriminative Part Localization: The perception and localization of object part-level features are usually challenging. Many existing approaches [25,35, 66, 67] suggest using bounding boxes and additional annotations to accurately", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan", + "bbox": [ + 271, + 113, + 720, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "locate discriminative regions of objects. However, this method often requires expert knowledge and extensive annotation efforts, rendering it impractical for real-world applications. Fortunately, more and more research [14, 23, 45, 68, 69] has shown that important object regions can also be localized using weakly supervised methods. These methods rely solely on image-level supervision to identify part features. Early works usually locate class-specific discriminative regions via CAM [69]. Fu et al. proposes Recursive Attention CNN (RACNN) [14] to iteratively focus on predicting the locations of attention regions and extract discriminative features. Zheng et al. proposes Multi-Attention CNN (MACNN) [68], which proposes channel grouping loss to generate multiple parts by clustering. WSDAN [23] proposes weakly supervised attention learning to generate the distribution of parts. Meng et al. [39] introduces an end-to-end part discovery model (PDM) that aims to uncover robust and diverse object parts within a unified framework. Our PANet focuses on perceiving multiple discriminative regions in each view, extracting sufficient local view parts, and proposing the Adaptive Part Refinement (APR) module to further refine them to obtain global parts. These global parts can adapt to changes in viewpoint for parts across different views, making them more suitable for 3D objects.", + "bbox": [ + 212, + 146, + 787, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 436, + 426, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we newly propose a Part-Aware Network (PANet) that aims to solve the challenge of 3D object recognition (especially under arbitrary views). The overall architecture of the proposed PANet is shown in Fig. 3, consisting of three modules: weakly supervised part localization, adaptive part refinement, and joint part representation. We first introduce a weakly supervised method to perform part localization and feature extraction on each view (Sec. 3.1). Secondly, to adapt to the variations of part features across views, we propose the Adaptive Part Refinement (APR) module, which refines redundant part features from multiple views into more compact representations (Sec. 3.2). Finally, we integrate these refined features to represent a 3D object (Sec. 3.3).", + "bbox": [ + 212, + 463, + 787, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Weakly-Supervised Part Localization", + "text_level": 1, + "bbox": [ + 215, + 632, + 568, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is worth noting that we do not rely on supervised positional annotations [67], such as bounding boxes or keypoints, for part localization. Instead, we employ a more cost-efficient weakly-supervised learning approach [23, 68] that utilizes image-level annotations to identify discriminative regions, which requires no expertise and no extensive annotation effort.", + "bbox": [ + 212, + 652, + 785, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We generate a multi-view sequence $\\mathcal{X} = \\{\\mathbf{X}_1, \\mathbf{X}_2, \\dots, \\mathbf{X}_v\\}$ by rendering a 3D object from different viewpoints. Here, $v$ denotes the number of views, $\\mathbf{X}_i$ represents the $i$ -th view, where $i \\in [1, v]$ . Then, each view is encoded by an image classification network e.g., ResNet [19] to obtain view features $\\mathbf{I} = \\{\\mathbf{I}_1, \\mathbf{I}_2, \\dots, \\mathbf{I}_v\\}$ , where $\\mathbf{I}_i \\in \\mathbb{R}^{H \\times W \\times C-1}$ , $i \\in [1, v]$ is corresponding to the feature map of the $i$ -th", + "bbox": [ + 212, + 728, + 787, + 805 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Beyond Viewpoint", + "bbox": [ + 606, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 H, W, and C represent the height, width, and number of channels of the feature layer, respectively", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e33d06aaec7b2ac8d1c4d8f0d9edab0e9654c7410868806a741831bf2bc8715a.jpg", + "image_caption": [ + "Fig. 3: Overview of our proposed Part-Aware Network (PANet). The network takes multi-view images as input, encodes them by CNN, obtains view features $\\mathbf{I}$ , and uses the cross-view association (CVA) module to enhance the view features. Firstly, employ weakly-supervised methods to perceive part regions and generate view part sequences $\\mathbf{T}$ . Then, given some learnable part tokens $\\mathbf{P}$ , the sequences $\\mathbf{T}$ and $\\mathbf{P}$ are concatenated and input to the adaptive part refinement (APR) module. The module aims to refine the sequence $\\mathbf{T}$ into a more compact representation, resulting in the global part features. Finally, multiple global parts $\\underline{\\mathbf{P}}$ together serve as a representation of a 3D object." + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 782, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "view $\\mathbf{X}_i$ . To align the features from multiple views into the same feature space, we share the weights among each of the convolutional neural network.", + "bbox": [ + 212, + 550, + 784, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Enhance Multi-view Features: However, the associations between view features $\\mathbf{I}$ are not explicitly modeled. Therefore, we develop a cross-view association (CVA) module to obtain enhanced view features $\\mathbf{F}_i$ , enabling each view to incorporate information from other views.", + "bbox": [ + 212, + 580, + 784, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each view feature $\\mathbf{I}_i$ , the cross-view association (CVA) module calculates the matching score between it and the overall view feature $\\mathbf{I}$ . Here, the dot product function is selected as the matching function. Then, $\\mathbf{I}_i$ is enhanced based on the correlation scores with other views. The enhanced view feature $\\mathbf{F}_i \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W} \\times \\mathrm{C}}$ can be expressed in detail as follows:", + "bbox": [ + 212, + 642, + 787, + 717 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} = \\operatorname {s o f t m a x} \\left(\\omega (\\mathbf {I} _ {i}) ^ {T} \\cdot \\omega (\\mathbf {I})\\right) \\mathbf {I}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 723, + 785, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\omega (\\mathbf{I}_i)\\in \\mathbb{R}^{\\mathrm{C}}$ represents the use of average pooling operations.", + "bbox": [ + 212, + 748, + 696, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Part Sampler: After passing through the cross-view association module, the enhanced view features $\\mathbf{F}$ comprehensively incorporate information from other views. Next, we sample parts from the enhanced view features $\\mathbf{F}$ .", + "bbox": [ + 212, + 763, + 784, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by [23, 45], we implement the attention model $\\psi$ to learn the spatial distribution of each part within each view. For view $i$ , the distribution of", + "bbox": [ + 212, + 810, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan", + "bbox": [ + 269, + 113, + 720, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "view parts is represented by the attention map $\\mathbf{A}_i\\in \\mathbb{R}^{\\mathrm{H}\\times \\mathrm{W}\\times \\mathrm{M}}$ , which can be computed by:", + "bbox": [ + 212, + 145, + 782, + 175 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {i} = \\psi (\\mathbf {F} _ {i}) = \\left\\{\\mathbf {A} _ {i, 1}, \\mathbf {A} _ {i, 2}, \\dots , \\mathbf {A} _ {i, \\mathrm {M}} \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 176, + 785, + 193 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{A}_{i,j} \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W}}$ , $i \\in [1,v]$ , $j \\in [1,\\mathrm{M}]$ represents the $j$ -th part region or visual pattern of the object's $i$ -th views, such as a car wheel or an airplane wing. M denotes the number of attention maps, each view extracts the same number of regions of interest.", + "bbox": [ + 212, + 198, + 782, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we employ the attention map $\\mathbf{A}_{i,j}$ to softly weight the enhanced feature maps $\\mathbf{F}_i$ , followed by the global average pooling operation $\\omega$ to obtain the parts on view $i$ (See Eq. 3).", + "bbox": [ + 212, + 258, + 784, + 304 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {t} _ {i, j} = \\omega \\left(\\mathbf {F} _ {i} \\otimes \\mathbf {A} _ {i, j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 306, + 785, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{t}_{i,j} \\in \\mathbb{R}^{\\mathrm{C}}$ , $\\otimes$ denotes element-wise multiplication between two tensors.", + "bbox": [ + 212, + 325, + 764, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we concatenate all the view part features together, resulting in a 3D multi-view part sequence $\\mathbf{T} \\in \\mathbb{R}^{N \\times C}$ (see Eq. 4).", + "bbox": [ + 212, + 342, + 782, + 372 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} = \\left[ \\boldsymbol {t} _ {0, 0}, \\boldsymbol {t} _ {0, 1}, \\dots , \\boldsymbol {t} _ {0, \\mathrm {M}}, \\boldsymbol {t} _ {1, 0}, \\boldsymbol {t} _ {1, 1}, \\dots , \\boldsymbol {t} _ {1, \\mathrm {M}}, \\dots , \\boldsymbol {t} _ {v, 0}, \\boldsymbol {t} _ {v, 1}, \\dots , \\boldsymbol {t} _ {v, \\mathrm {M}} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 385, + 785, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{N}$ represents the total number of parts $(v\\times \\mathrm{M})$ in the multi-view.", + "bbox": [ + 214, + 411, + 735, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Adaptive Part Refinement", + "text_level": 1, + "bbox": [ + 214, + 449, + 478, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Sec. 3.1, we have obtained a multi-view part sequence $\\mathbf{T}$ . However, there are differences in parts across different views, where each element $t_{ij}$ in $\\mathbf{T}$ can only represent local view features. It is challenging to generalize the overall features of a 3D object relying solely on local view parts. Therefore, we newly propose an Adaptive Part Refinement (APR) module, which aims to explicitly model the correlations between parts across views, thus enhancing the robustness of part features to viewpoint variations. This module adaptively integrates part information from multiple views and refines it to obtain global part features that are more suitable for 3D objects.", + "bbox": [ + 212, + 473, + 784, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Formally, the APR (Adaptive Part Refinement) module contains a transformer block. Given some global part tokens [4,7,11] $\\mathbf{P} \\in \\mathbb{R}^{\\mathrm{L} \\times \\mathrm{C}}$ , where $\\mathrm{L}$ represents the number of parts. We concatenate $\\mathbf{P}$ and $\\mathbf{T}$ as input to the APR module. After passing through the APR module, local parts $\\pmb{t}_{i,j}$ can learn cross-view correlations and obtain global parts $\\underline{\\mathbf{P}} \\in \\mathbb{R}^{\\mathrm{L} \\times \\mathrm{C}}$ suitable for 3D objects. We also observe the diversity of global part features $\\underline{\\mathbf{P}}$ (See Fig. 4).", + "bbox": [ + 212, + 609, + 782, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Joint Multi-Part Representation", + "text_level": 1, + "bbox": [ + 214, + 722, + 529, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relying solely on individual parts may not be sufficient to represent a 3D object. To overcome this limitation, we jointly describe the 3D object using multiple parts. We make predictions for each part and then average them to obtain the prediction for the 3D object. Each 3D object can be represented as Eq. 5:", + "bbox": [ + 212, + 747, + 782, + 808 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\underline {{\\mathbf {P}}} = \\left[ \\underline {{\\boldsymbol {p}}} _ {1}, \\underline {{\\boldsymbol {p}}} _ {2}, \\dots , \\underline {{\\boldsymbol {p}}} _ {\\mathrm {L}} \\right], \\quad \\text {w h e r e} \\underline {{\\mathbf {P}}} \\in \\mathbb {R} ^ {\\mathrm {L} \\times \\mathrm {C}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 819, + 785, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Beyond Viewpoint", + "bbox": [ + 606, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The proposed PANet is optimized through supervised learning, including instance classification loss $\\mathcal{L}_{ce}$ and part-aware loss $\\mathcal{L}_{awe}$ . The loss function for multi-view $\\mathcal{X}$ is defined as follows:", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p ^ {i} = \\operatorname {s o f t m a x} \\left(\\underline {{\\boldsymbol {p}}} _ {i} \\boldsymbol {W} _ {p} + \\boldsymbol {b} _ {p}\\right), \\quad \\text {w h e r e} \\underline {{\\boldsymbol {p}}} _ {i} \\in \\mathbb {R} ^ {1 \\times \\mathrm {C}}, i \\in [ 1, \\mathrm {L} ] \\\\ q ^ {j} = \\operatorname {s o f t m a x} \\left(\\boldsymbol {t} _ {j} \\boldsymbol {W} _ {q} + \\boldsymbol {b} _ {q}\\right), \\quad \\text {w h e r e} \\boldsymbol {t} _ {j} \\in \\mathbb {R} ^ {1 \\times (\\mathrm {M} \\times \\mathrm {C})}, j \\in [ 1, v ], \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 200, + 784, + 255 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {a w e} = - \\sum_ {k = 1} ^ {K} y _ {k} \\log \\hat {p} _ {k} + \\gamma \\left(- \\frac {1}{v} \\sum_ {i = 1} ^ {v} \\sum_ {k = 1} ^ {K} y _ {k} \\log q _ {k} ^ {j}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 256, + 709, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $W_{p}, W_{q}, b_{p}$ , and $b_{q}$ are learnable parameters. $p^{i}$ represents the predicted probability of the $i$ -th global part, $\\hat{p}_{k}$ represents the average of the predicted probabilities for L global parts, and $q^{j}$ represents the predicted probability for the $j$ -th view part. $y_{k}$ is the ground truth label of class $k$ . $\\mathcal{L}_{ce}$ represents the cross-entropy loss of the final parts. $\\mathcal{L}_{awe}$ represents the classification loss of M part features in each view, which is used to constrain the generation of parts. Furthermore, we apply techniques such as flipping and random erasure to augment the image for more robust part results.", + "bbox": [ + 212, + 306, + 787, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 450, + 375, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets and Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 482, + 539, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an extensive evaluation of PANet on several benchmark datasets, namely ScanObjectNN [53], ModelNet [59], and RGBD [29]. In order to facilitate fair comparisons, we strictly adhere to the methodologies proposed in prior works [26, 48, 55] for preparing the training and testing data on the different datasets. Specifically, we adopt the training scheme introduced by MVCNN [48] and adopt the same viewpoint generation and rendering method as RotationNet [26] to handle the multi-view images from the ModelNet and ScanObjectNN datasets. Experiment on rotated datasets also follows CVR [55] setting. For arbitrary views, the number of views for each 3D object ranges randomly from 10 to 20. Throughout the training process, we utilize the AdamW [36] optimizer with a learning rate of 1e-5. We select ResNet50 as our backbone network and utilize label smoothing regularization [51] to enhance the generalization capability of our model. $\\gamma$ is the weighting factor for the part-aware loss as in Eq. 6, which we set $\\gamma = 1$ for the experiments. Please refer to the Appendix for implementation details of our method.", + "bbox": [ + 212, + 506, + 787, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Comparison with State-of-the-Art Method", + "text_level": 1, + "bbox": [ + 215, + 753, + 612, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ScanObjectNN Datasets. As a recently introduced real-world 3D object classification dataset, ScanObjectNN [53] provides scanned indoor scene data. The dataset comprises approximately 15,000 objects, spanning 15 categories, and includes 2,902 unique object instances. Compared with CAD handcrafted", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan", + "bbox": [ + 269, + 113, + 720, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c79fb4f3f8a203411f917406b6b550f0026cbcd093ab1199dc116a4759ec19fa.jpg", + "table_caption": [ + "Table 1: Classification on ScanObjectNN. 'Aligned' means that every 3D object is in the same pose. 'Rotated' means that each 3D object is misaligned. '-' indicates that the specific result is not reported." + ], + "table_footnote": [], + "table_body": "
MethodTypeAlignedRotated
Per Class Acc.Per Ins. Acc.Per Class Acc.Per Ins. Acc.
PointNet++ [43]Point82.10%84.30%--
DGCNN [54]84.00%86.20%--
MVCNN [48]85.71%87.82%78.21%80.62%
GVCNN [13]86.64%88.68%82.86%83.70%
RotationNet [26]View (fixed)84.88%86.90%74.68%76.16%
View-GCN [56]88.67%90.39%81.99%83.50%
CVR [55]88.39%90.74%84.70%85.59%
View-GCN++ [57]89.10%91.30%85.80%86.80%
Ours93.21%93.99%86.15%87.82%
CVR [55](arbitrary)--68.07%71.36%
Ours91.77%93.48%84.08%86.45%
", + "bbox": [ + 217, + 198, + 785, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "datasets, ScanObjectNN offers more realistic challenges as it reflects the noise, complexity, and incompleteness present in real-world environments, incorporating the appearance and geometric shapes of actual objects. Consequently, it is better for assessing the robustness of the model. We train PANet under different viewpoint settings and compare with point-based methods [43, 54] and view-based methods [13, 26, 48, 55-57], as shown in Tab. 1. View-based methods primarily rely on view aggregation, and it is difficult to robustly aggregate features from structurally misaligned views, especially for arbitrary views. When switching to arbitrary views, the performance of these methods drops sharply. Our PANet is observed to significantly outperform previous approaches in both class accuracy and instance accuracy, whether in aligned views, rotated views, or arbitrary views. In aligned views, compared with the current state-of-the-art methods e.g., View-GCN++ [57] and CVR [55], PANet improves instance accuracy and class accuracy by $2.69\\%$ and $4.11\\%$ . PANet also achieves significant improvements of $1.02\\%$ and $0.35\\%$ in rotated views. Surprisingly, even under arbitrary views, our performance surpasses that of most fixed viewpoint methods, demonstrating that PANet can robustly recognize 3D objects unconstrained by the scene. Please refer to the Appendix for more detailed comparisons.", + "bbox": [ + 212, + 372, + 787, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The ModelNet Datasets. ModelNet [59] is a dataset of hand-drawn point clouds generated from CAD models. It comprises 12,311 3D shapes, covering 40 categories, with 9,483 models allocated for training and 2,468 for testing. The experimental results on the ModelNet dataset are presented in Tab. 2. We first compare with point-based and voxel-based methods, including PointNet++ [43], Kd-Networks [27], RS-CNN [22], KP-Convs [52], MVTN [34], 3DShapeNets [58], VoxNet [38], and VRENsemble [3]. Our PANet outperforms them by more than $3.93\\%$ and $4.6\\%$ in instance accuracy and class accuracy and is also $3.73\\%$ and $4.37\\%$ higher in free viewpoints. We also compared with view-based approaches, including MVCNN-new [48], GVCNN [13], HEAR [5], RotationNet [26], View-GCN [56], MVT [6], HMVCM [32], MVD [24], DAN [41], DRCNN [50], EMV [12], View-GCN++ [57], CVR [55], and HGNN+ [16]. PANet consistently outper", + "bbox": [ + 212, + 657, + 789, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Beyond Viewpoint", + "bbox": [ + 607, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f4b4bfd3fb70a04470e07b7d7ef5b3e193ba4b9c713213cd8ff63d06e8880f04.jpg", + "table_caption": [ + "Table 2: Classification on Aligned ModelNet40. '- indicates that the specific result is not reported, 'fixed' denotes sampling from a fixed viewpoint. 'free' means that the arbitrary number of viewpoints." + ], + "table_footnote": [], + "table_body": "
Aligned ModelNet40
MethodTypeClass Acc.Ins. Acc.
3DShapeNets [58]77.3%-
VoxNet [38]Voxel83.0%-
VRENsemble [3]91.4%93.8%
PointNet++ [43]-91.9%
Kd-Networks [27]88.5%91.8%
RS-CNN [22]Point-93.6%
KP-Convs [52]-92.9%
MVTN [34]92.0%93.8%
MVCNN-new [48]92.40%95.00%
GVCNN [13]90.70%93.10%
HEAR [5]95.20%96.70%
RotationNet [26]-97.37%
View-GCN [56]96.50%97.60%
MVT [6]-97.50%
HMVCM [32]-94.57%
MVD [24]View92.61%94.41%
DAN [41]-93.50%
DRCNN [50]94.86%96.84%
View-GCN++ [57]96.50%97.60%
CVR [55]95.77%97.16%
HGNN+ [16]-96.92%
Ours96.60%97.73%
Ours(free)96.37%97.53%
", + "bbox": [ + 217, + 224, + 488, + 520 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a48c58fec3a11daf4c3e18d21abc723b2b792a8f7b0a6a73001e2abc1529683c.jpg", + "table_caption": [ + "Table 3: The accuracy of 3D object classification on Rotated ModelNet40." + ], + "table_footnote": [], + "table_body": "
Rotated ModelNet40
Method#ViewClass Acc.Ins. Acc.
MVCNN [48]87.95%88.17%
GVCNN [13]89.69%88.10%
EMV [12]88.80%90.00%
RotationNet [26]fixed84.74%85.29%
View-GCN [56]85.90%88.25%
CVR [55]91.12%92.22%
View-GCN++ [57]90.70%92.80%
Ours90.35%92.87%
CVR [55]arbitrary84.01%86.91%
Ours90.69%92.23%
", + "bbox": [ + 511, + 184, + 784, + 332 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5df074cc37ac3eff97bb27ba5cc7229a9650ff9eedbb317fdaf46919d2bdef93.jpg", + "table_caption": [ + "Table 4: The accuracy of object classification on RGBD (in%)." + ], + "table_footnote": [], + "table_body": "
Method#ViewIns. Acc.
MDSICNN [2]≥ 12089.6%
CFK [8]≥ 12086.8%
MMDCNN [44]≥ 12086.8%
RotationNet [26]1289.3%
View-GCN [56]1294.3%
View-GCN++ [57]1292.5%
Ours1296.5%
MVCNN [48]89.0%
GVCNN [13]89.8%
CVR [55]4-1291.8%
Ours95.9%
", + "bbox": [ + 514, + 375, + 782, + 518 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "forms these methods in fixed views. Its performance at free viewpoints surpasses that of most existing approaches. We next report the robustness of PANet in rotated views, as shown in Tab. 3. Under the rotated views, PANet also achieves the highest score in instance accuracy.", + "bbox": [ + 212, + 526, + 784, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The RGBD Datasets. The RGBD [29] dataset comprises 51 categories of domestic objects captured in the real world, encompassing RGBD views from multiple viewpoints. To ensure a fair comparison, we adopt the identical experimental setup as RotationNet [26] and View-GCN [56]. Specifically, we uniformly sample 12 RGB images from the same circle for each 3D object as input and report the average results using ten-fold cross-validation. We keep the same settings as CVR [55] in arbitrary views, sampling images from 4 to 12 viewpoints as input to the network. The experimental results are summarized in Tab. 4. The performance of PANet in arbitrary views is impressive. Compared with work in arbitrary views e.g., CVR [55], PANet significantly improves by $4.1\\%$ . Additionally, compared with the work in fixed views, such as view-GCN [56], PANet has also improved by $1.6\\%$ . This further validates the superiority of our method.", + "bbox": [ + 212, + 585, + 787, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Model Analysis", + "text_level": 1, + "bbox": [ + 215, + 787, + 390, + 803 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "All experiments are conducted on the ScanObjectNN dataset in rotated views. Please refer to the Appendix for detailed experimental analyses.", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan", + "bbox": [ + 271, + 113, + 718, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c8ab5804a3db7cc39f0d2e0fe951b9c25564269d8d80930d3fd41747820e7a72.jpg", + "image_caption": [ + "Fig.4: Visualization of correlations between part features $\\underline{\\mathbf{P}}$ . The left figure illustrates the correlation between $\\underline{\\mathbf{P}}$ after applying $\\mathcal{L}_{awe}$ ." + ], + "image_footnote": [], + "bbox": [ + 228, + 147, + 336, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d7abb86743ead2a22ae1876985ac2207d96a77cebac32c138c597ee86f87b1d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 148, + 477, + 229 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/140c643ec3725e71d9b2d3f298af1b29fa3dab64fef08772bc3c93897cd211af.jpg", + "table_caption": [ + "Table 5: PANet component ablations. The first line serves as the baseline. $\\mathcal{L}_{awe}$ refers to the weakly supervised part-aware loss function. CVA stands for Cross-View Association module." + ], + "table_footnote": [], + "table_body": "
CVALaweClass Acc.Ins. Acc.
✓83.76%84.56%
84.99%85.25%
✓85.45%86.96%
", + "bbox": [ + 511, + 234, + 785, + 292 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c6246b8ddbf77713fbd11525b1f0f5dcf1c5fcb3020bce85dc9d3c0eeef5aa0c.jpg", + "table_caption": [ + "Table 6: Explore the impact of different viewpoint sampling on performance." + ], + "table_footnote": [], + "table_body": "
DatasetMethodAlignedRotated
Class Acc.Ins. Acc.Class Acc.Ins. Acc.
ModelNet40random sampling96.37%97.53%90.69%92.23%
furthest point sampling95.44%97.41%89.73%92.26%
ScanObjectNNrandom sampling91.77%93.48%84.08%86.45%
furthest point sampling92.97%93.48%84.80%85.93%
", + "bbox": [ + 225, + 332, + 772, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "PANet component ablations We evaluate the role of the cross-view association module and $\\mathcal{L}_{awe}$ , as shown in Tab. 5. The CVA module plays a positive role in enhancing the performance of PANet, improving instance accuracy and class accuracy by $0.69\\%$ and $1.23\\%$ . The results show that the CVA module enables each view to more effectively integrate information from other views, further enhancing the representation capability of the view features. Our method also improves by $2.40\\%$ and $1.76\\%$ in two metrics by introducing the $\\mathcal{L}_{awe}$ constraint. In addition, we also conduct correlation analysis between part features $\\underline{\\mathbf{P}}$ , as shown in Fig. 4. We observe that without applying $\\mathcal{L}_{awe}$ constraints, the correlation between parts exhibits a flat similarity matrix, indicating that the parts are unvaried and cannot be distinguished in the feature space. In contrast, the part features become more diverse with $\\mathcal{L}_{awe}$ constraint. Therefore, we conclude that $\\mathcal{L}_{awe}$ can supervise the perception of parts to obtain diverse part features.", + "bbox": [ + 212, + 428, + 785, + 625 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Influence of Different Viewpoint Sampling We introduce two different sampling methods: random sampling and furthest point sampling (FPS). Random sampling refers to the process of randomly selecting 10 to 20 views from each object. This method is closer to real-world scenarios, and we showcased PANet's performance in Tab. 1, 2, 3, 4 based on this sampling method. FPS selects 10 to 20 views based on the furthest point sampling principle. It provides a uniform sampling method that covers every orientation of the 3D objects. As shown in Tab. 6, we observed that the performance of random sampling and FPS across different datasets is nearly similar. This result shows that PANet is not sensitive to the choice of sampling methods. The model demonstrates robustness in adapting to changes in viewpoint caused by different sampling methods while maintaining stable performance.", + "bbox": [ + 212, + 627, + 787, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Influence of the Number of the Views We further evaluate PANet by considering different backbone networks and numbers of views, as shown in Tab. 7.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Beyond Viewpoint", + "bbox": [ + 606, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2b450ae56c007bb867e489c90cead97370fe8a3b9570410f5cc8dba9a9663544.jpg", + "table_caption": [ + "Table 7: Explore the impact of number of views and backbones on performance." + ], + "table_footnote": [], + "table_body": "
BackboneViews
1 views5 views10 views15 views20 views
resnet1845.63%70.84%75.66%79.90%82.85%
resnet5067.07%79.59%82.50%85.93%86.62%
", + "bbox": [ + 264, + 170, + 738, + 239 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/423c88ff53efb21ba3e3c5440d7554b6d73c8b218d52e659c9f17a0cdad99ba5.jpg", + "table_caption": [ + "Table 8: Exploring the effect of the number of attention maps, M denotes the number of attention maps per view." + ], + "table_footnote": [], + "table_body": "
#Attention MapsClass Acc.Ins. Acc.
M = 1684.57%86.28%
M = 3285.36%87.14%
M = 6486.50%87.31%
M = 12886.38%87.14%
", + "bbox": [ + 220, + 297, + 483, + 364 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bf45da096ebaf91f5bce0daa749db953fe5104d94015819d7befe47ff2d95ba5.jpg", + "table_caption": [ + "Table 9: Exploring the effect of the number of part tokens, L indicates the number of parts." + ], + "table_footnote": [], + "table_body": "
#Number of PartsClass Acc.Ins. Acc.
L = 185.50%86.28%
L = 886.20%87.14%
L = 1686.50%87.31%
L = 3285.63%86.31%
", + "bbox": [ + 511, + 299, + 779, + 364 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In previous studies, such as View-GCN [56], because view aggregation leads to feature loss. Choosing a stronger backbone network may not necessarily lead to performance improvement in certain dataset. However, our observations reveal that the choice of backbone significantly affects PANet performance. This is because PANet directly integrates multiple parts to characterize 3D objects. Therefore, a stronger backbone network is more advantageous for perceiving part features. Furthermore, we notice a positive correlation between the number of views and the performance of PANet. It is worth noting that previous studies on multi-view aggregation methods, such as MHBN [65], MVLADN [64], and GVCNN [13], have indicated that increasing the number of views does not necessarily lead to better representation of 3D objects. PANet combines multiple parts for representation, and the APR module can also filter out some redundant parts. This means that PANet is not affected by redundant views. Our experimental results also verify this idea.", + "bbox": [ + 212, + 372, + 785, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Influence of the Number of the Attention Maps and Parts We show the impact of the number of attention maps M and the number of part tokens L, as shown in Tab. 8 and Tab. 9. The number of attention maps M determines the discriminative region in each view, and the part tokens L determine the number of global parts. More M and L usually contribute to better performance. MACNN [68] and WSDAN [23] report similar conclusions. The experiments in Tab. 8 show that as the value of M increases, the classification accuracy gradually improves. When $\\mathrm{M} = 64$ , the performance tends to stabilize. The results in Tab. 9 show that the performance improves as the number of parts increases and reaches a peak at $\\mathrm{L} = 16$ . Compared with a single part, multiple parts provide richer information and enhance the generalization of the model.", + "bbox": [ + 212, + 584, + 787, + 750 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Visualization", + "text_level": 1, + "bbox": [ + 215, + 770, + 370, + 784 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We randomly select some 3D models from a test set and compare the semantic differences between the features extracted by MVCNN and PANet, as shown in Fig. 5. MVCNN [48] is a classic feature aggregation method. We notice that the", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan", + "bbox": [ + 269, + 113, + 720, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "features extracted by MVCNN are not entirely focused on the object. This is because the uneven weight distribution leads to view aggregation methods ignoring certain view information. It is worth noting that PANet does not discard any views but instead considers the features of each view, leveraging the valuable information within each view. We can observe that different views emphasize various part locations, with each view capturing multiple discriminative regions (e.g., the airplane nose, body, and wings). This observation further validates the diversity of captured parts. Furthermore, we observe that the view aggregation method does not explicitly model the unique attributes of 3D objects, leading to feature ambiguity. In contrast, part-based representation reveals the inherent properties of objects, enhancing the interpretability of features and making recognition easier. Please refer to the Appendix for more visualization results.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 349, + 359, + 367 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paper, we propose a novel part-aware network (PANet), which is a part-based representation. Compared with view aggregation methods, this part-based representation has properties such as viewpoint invariance and rotation robustness, which can more effectively address the 3D object recognition problem under arbitrary views. Moreover, we proposed the part-aware loss to ensure that part features can better focus on the object, thereby fully utilizing the information from each view. We evaluate the performance of PANet on several 3D datasets, including ScanObjectNN, ModelNet, and RGBD. The experimental results show that PANet has made significant progress, especially in arbitrary views. We also conduct ablation and visualization experiments to study the impact of different components in our method. It is worth noting that in these experiments, we found that our method is not affected by redundant views (Sec. 4.3), and the part features can better reveal the inherent properties of objects (Sec. 4.4). In conclusion, the experimental results prove the effectiveness of our method.", + "bbox": [ + 212, + 383, + 787, + 594 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Acknowledgement This research is supported by the Shenzhen Science and Technology Plan Project No. KJZD20230923113800002 and Shenzhen LongHua Fundamental Research Program No. 10162A20230325B73A546.", + "bbox": [ + 214, + 608, + 785, + 650 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Beyond Viewpoint", + "bbox": [ + 607, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9f3fe613b8e83a64611c25212ff7ed83aafbdc9c41e62438e38e43e91867e945.jpg", + "image_caption": [ + "Fig. 5: Comparing the semantic differences between view-based aggregation methods and PANet in feature extraction." + ], + "image_footnote": [], + "bbox": [ + 222, + 198, + 493, + 734 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/bf8ea68dca1f99451d4bde25a8a38882d323cd4d796afa8e2fa1d31136628ff5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 198, + 779, + 734 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan", + "bbox": [ + 271, + 114, + 720, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Afham, M., Dissanayake, I., Dissanayake, D., Dharmasiri, A., Thilakarathna, K., Rodrigo, R.: Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9902-9912 (2022)", + "2. Asif, U., Bennamoun, M., Sohel, F.A.: A multi-modal, discriminative and spatially invariant cnn for rgb-d object labeling. IEEE transactions on pattern analysis and machine intelligence 40(9), 2051-2065 (2017)", + "3. Brock, A., Lim, T., Ritchie, J.M., Weston, N.: Generative and discriminative voxel modeling with convolutional neural networks. arXiv preprint arXiv:1608.04236 (2016)", + "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020)", + "5. Chen, J., Qin, J., Shen, Y., Liu, L., Zhu, F., Shao, L.: Learning attentive and hierarchical representations for 3d shape recognition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 105-122. Springer (2020)", + "6. Chen, S., Yu, T., Li, P.: Mvt: Multi-view vision transformer for 3d object recognition. arXiv preprint arXiv:2110.13083 (2021)", + "7. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)", + "8. Cheng, Y., Cai, R., Zhao, X., Huang, K.: Convolutional fisher kernels for rgb-d object recognition. In: 2015 International Conference on 3D Vision. pp. 135-143. IEEE (2015)", + "9. Delitzas, A., Parelli, M., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329 (2023)", + "0. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "2. Esteves, C., Xu, Y., Allen-Blanchette, C., Daniilidis, K.: Equivariant multi-view networks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1568-1577 (2019)", + "3. Feng, Y., Zhang, Z., Zhao, X., Ji, R., Gao, Y.: GvCNN: Group-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 264-272 (2018)", + "4. Fu, J., Zheng, H., Mei, T.: Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4438-4446 (2017)", + "5. Gao, K., Gao, Y., He, H., Lu, D., Xu, L., Li, J.: Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379 (2022)" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Beyond Viewpoint", + "bbox": [ + 606, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Gao, Y., Feng, Y., Ji, S., Ji, R.: Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3181-3199 (2022)", + "17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020)", + "18. Han, Z., Lu, H., Liu, Z., Vong, C.M., Liu, Y.S., Zwicker, M., Han, J., Chen, C.P.: 3d2seqviews: Aggregating sequential views for 3d global feature learning by cnn with hierarchical attention aggregation. IEEE Transactions on Image Processing 28(8), 3986-3999 (2019)", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "20. Herreras, E.B.: Cognitive neuroscience; the biology of the mind. Cuadernos de Neuropsicología/Panamerican Journal of Neuropsychology 4(1), 87-90 (2010)", + "21. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9202-9212 (2023)", + "22. Hu, L., Qin, M., Zhang, F., Du, Z., Liu, R.: Rscnn: a cnn-based method to enhance low-light remote-sensing images. Remote Sensing 13(1), 62 (2020)", + "23. Hu, T., Qi, H., Huang, Q., Lu, Y.: See better before looking closer: Weakly supervised data augmentation network for fine-grained visual classification. arXiv preprint arXiv:1901.09891 (2019)", + "24. Huang, J., Yan, W., Li, T., Liu, S., Li, G.: Learning the global descriptor for 3-d object recognition based on multiple views decomposition. IEEE Transactions on Multimedia 24, 188-201 (2022). https://doi.org/10.1109/TMM.2020.3047762", + "25. Huang, S., Xu, Z., Tao, D., Zhang, Y.: Part-stacked cnn for fine-grained visual categorization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1173-1182 (2016)", + "26. Kanezaki, A., Matsushita, Y., Nishida, Y.: Rotationnet for joint object categorization and unsupervised pose estimation from multi-view images. IEEE transactions on pattern analysis and machine intelligence 43(1), 269-283 (2019)", + "27. Klokov, R., Lempitsky, V.: Escape from cells: Deep kd-networks for the recognition of 3d point cloud models. In: Proceedings of the IEEE international conference on computer vision. pp. 863-872 (2017)", + "28. Kumawat, S., Raman, S.: Lp-3dcnn: Unveiling local phase in 3d convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4903-4912 (2019)", + "29. Lai, K., Bo, L., Ren, X., Fox, D.: A large-scale hierarchical multi-view rgb-d object dataset. In: 2011 IEEE international conference on robotics and automation. pp. 1817-1824. IEEE (2011)", + "30. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)", + "31. Lin, Y., Gou, Y., Liu, X., Bai, J., Lv, J., Peng, X.: Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4447-4461 (2022)", + "32. Liu, A.A., Zhou, H., Nie, W., Liu, Z., Liu, W., Xie, H., Mao, Z., Li, X., Song, D.: Hierarchical multi-view context modelling for 3d object classification and retrieval. Information Sciences 547, 984-995 (2021)" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan", + "bbox": [ + 271, + 114, + 720, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)", + "34. Liu, S., Nguyen, V., Rehg, I., Tu, Z.: Recognizing objects from any view with object and viewer-centered representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2020)", + "35. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing 30, 1744-1758 (2021)", + "36. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "37. Lowe, D.: Object recognition from local scale-invariant features. In: Proceedings of the Seventh IEEE International Conference on Computer Vision. vol. 2, pp. 1150-1157 vol.2 (1999). https://doi.org/10.1109/ICCV.1999.790410", + "38. Maturana, D., Scherer, S.: Voxnet: A 3d convolutional neural network for real-time object recognition. In: 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 922-928. IEEE (2015)", + "39. Meng, M., Zhang, T., Yang, W., Zhao, J., Zhang, Y., Wu, F.: Diverse complementary part mining for weakly supervised object localization. IEEE Transactions on Image Processing 31, 1774-1788 (2022). https://doi.org/10.1109/TIP.2022.3145238", + "40. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "41. Nie, W., Zhao, Y., Song, D., Gao, Y.: Dan: Deep-attention network for 3d shape recognition. IEEE Transactions on Image Processing 30, 4371-4383 (2021)", + "42. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)", + "43. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "44. Rahman, M.M., Tan, Y., Xue, J., Lu, K.: Rgb-d object recognition with multimodal deep convolutional neural networks. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 991-996. IEEE (2017)", + "45. Rao, Y., Chen, G., Lu, J., Zhou, J.: Counterfactual attention learning for fine-grained visual categorization and re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1025–1034 (2021)", + "46. Riegler, G., Osman Ulusoy, A., Geiger, A.: Octnet: Learning deep 3d representations at high resolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3577-3586 (2017)", + "47. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "48. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015)", + "49. Su, J.C., Gadelha, M., Wang, R., Maji, S.: A deeper look at 3d shape classifiers. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Beyond Viewpoint", + "bbox": [ + 606, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Sun, K., Zhang, J., Liu, J., Yu, R., Song, Z.: Drcnn: Dynamic routing convolutional neural network for multi-view 3d object recognition. IEEE Transactions on Image Processing 30, 868-877 (2020)", + "51. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)", + "52. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019)", + "53. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1588-1597 (2019)", + "54. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog) 38(5), 1-12 (2019)", + "55. Wei, X., Gong, Y., Wang, F., Sun, X., Sun, J.: Learning canonical view representation for 3d shape recognition with arbitrary views. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 407-416 (2021)", + "56. Wei, X., Yu, R., Sun, J.: View-gcn: View-based graph convolutional network for 3d shape analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1850-1859 (2020)", + "57. Wei, X., Yu, R., Sun, J.: Learning view-based graph convolutional network for multi-view 3d shape analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "58. Wu, S., Khosla, Y., Zhang, T.: D shapenets: A deep representation for volumetric shape modeling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA. pp. 7-12 (2015)", + "59. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)", + "60. Xu, Y., Zheng, C., Xu, R., Quan, Y., Ling, H.: Multi-view 3d shape recognition via correspondence-aware deep learning. IEEE Transactions on Image Processing 30, 5299-5312 (2021)", + "61. Xue, L., Gao, M., Xing, C., Martin-Martin, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning unified representation of language, image and point cloud for 3d understanding. arXiv preprint arXiv:2212.05171 (2022)", + "62. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)", + "63. Yang, Z., Wang, L.: Learning relationships for multi-view 3d object recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7505-7514 (2019)", + "64. Yu, T., Meng, J., Yang, M., Yuan, J.: 3d object representation learning: A set-to-set matching perspective. IEEE Transactions on Image Processing 30, 2168-2179 (2021). https://doi.org/10.1109/TIP.2021.3049968", + "65. Yu, T., Meng, J., Yuan, J.: Multi-view harmonized bilinear network for 3d object recognition. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 186-194 (2018). https://doi.org/10.1109/CVPR.2018.00027" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan", + "bbox": [ + 271, + 114, + 720, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Zhang, H., Xu, T., Elhoseiny, M., Huang, X., Zhang, S., Elgammal, A., Metaxas, D.: Spda-cnn: Unifying semantic part detection and abstraction for fine-grained recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1143-1152 (2016)", + "67. Zhang, N., Donahue, J., Girshick, R., Darrell, T.: Part-based r-cnns for fine-grained category detection. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 834-849. Springer (2014)", + "68. Zheng, H., Fu, J., Mei, T., Luo, J.: Learning multi-attention convolutional neural network for fine-grained image recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 5209-5217 (2017)", + "69. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016)" + ], + "bbox": [ + 212, + 146, + 787, + 340 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Beyond Viewpoint", + "bbox": [ + 606, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_model.json b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2bd8f22d9d6f4206e55c73354d267d5aebed6663 --- /dev/null +++ b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_model.json @@ -0,0 +1,2449 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.784, + 0.208 + ], + "angle": 0, + "content": "Beyond Viewpoint: Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.235, + 0.75, + 0.252 + ], + "angle": 0, + "content": "Linlong Fan\\(^{\\text{日}}\\), Ye Huang\\(^{\\text{日}^{\\star}}\\), Yanqi Ge\\(^{\\text{日}}\\), Wen Li\\(^{\\text{日}}\\), and Lixin Duan\\(^{\\text{日}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.263, + 0.693, + 0.306 + ], + "angle": 0, + "content": "Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China fanlinlong703@163.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.337, + 0.74, + 0.601 + ], + "angle": 0, + "content": "Abstract. Existing view-based methods excel at recognizing 3D objects from predefined viewpoints, but their exploration of recognition under arbitrary views is limited. This is a challenging and realistic setting because each object has different viewpoint positions and quantities, and their poses are not aligned. However, most view-based methods, which aggregate multiple view features to obtain a global feature representation, hard to address 3D object recognition under arbitrary views. Due to the unaligned inputs from arbitrary views, it is challenging to robustly aggregate features, leading to performance degradation. In this paper, we introduce a novel Part-aware Network (PANet), which is a part-based representation, to address these issues. This part-based representation aims to localize and understand different parts of 3D objects, such as airplane wings and tails. It has properties such as viewpoint invariance and rotation robustness, which give it an advantage in addressing the 3D object recognition problem under arbitrary views. Our results on benchmark datasets clearly demonstrate that our proposed method outperforms existing view-based aggregation baselines for the task of 3D object recognition under arbitrary views, even surpassing most fixed viewpoint methods." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.614, + 0.688, + 0.628 + ], + "angle": 0, + "content": "Keywords: 3D object recognition \\(\\cdot\\) weakly-supervised learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.652, + 0.377, + 0.669 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Learning 3D with multi-view supervision has received widespread attention. It covers various research domains such as NeRFs [15,40], BEV detection [30,62], 3D generation [33], 3D reasoning [9,21], etc., among which 3D object recognition is considered one of the most fundamental problems. Thanks to mature and transferable 2D vision models [11,47] and large-scale image datasets [10], rendering 3D objects into multiple 2D images and applying 2D networks to indirectly recognize 3D objects has become the state-of-the-art method [57]. Furthermore, compared with direct methods (e.g., point-based [1,61], voxel-based [38]), view-based methods are closer to human perception of 3D objects in the real world." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.721, + 0.129 + ], + "angle": 0, + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.147, + 0.785, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig. 1: In practical scenarios, 3D objects are observed from arbitrary views(right). The definition of arbitrary views is that each object is unaligned, and the viewpoint positions and quantities vary for each object. Previous works have focused on 3D object recognition in aligned views(left), where the pose of each object is aligned, and the positions and quantities of viewpoints are predefined. Rotated views(middle) is an extension that introduces random rotation for each 3D object while keeping the viewpoint positions and quantities unchanged. Our work focuses on arbitrary views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.431 + ], + "angle": 0, + "content": "Previous view-based methods [5,6,49,50,56] focus on addressing 3D object recognition tasks under fixed viewpoints, while recognition under arbitrary views still lacks extensive research. As shown in Fig. 1, arbitrary views represent a more challenging setting where each object may have different viewpoint positions and quantities, and the poses of each object are not aligned." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.628 + ], + "angle": 0, + "content": "However, most view-based methods [5,6,13,18,26,41,49,50,55-57,60,63-65], which aggregate multiple view features to obtain a global feature representation, hard to effectively address 3D object recognition under arbitrary views. This is because the misaligned nature of arbitrary views complicates the representation and aggregation of view features, which leads to performance degradation. For example, when applied to arbitrary views, the performance of CVR [55] drops by \\(6.31\\%\\) and \\(14.23\\%\\) in ModelNet [59] and ScanObjectNN [53] datasets. Even some multi-view aggregation methods [26, 56] cannot be applied to arbitrary views because they require fixed viewpoint position information. Furthermore, view aggregation methods [13, 49] typically jointly optimize feature extraction and aggregation modules to maximize performance, but this may lead the extraction module to overly focus on certain views while ignoring information from others, resulting in information loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To address these challenges, we propose a novel part-based representation, as shown in Fig. 2. This idea is inspired by cognitive science and biology [20], which aims to understand different part features or specific regions in a multi-view sequence to better represent 3D objects. For example, a person can be composed of multiple parts with unique attributes, such as the head, arms, body, legs, and so on. By locating and integrating the features of these parts, we can more accurately represent a person. Compared with the multi-view aggregation features, part features offer several advantages. Firstly, an individual view contains abundant part information. Even if only a few views are utilized, we can still extract stable part features, thereby adapting to changes in the number of viewpoints in arbitrary view settings. Secondly, the part features are more stable than the multi-view aggregation features [37], which can better adapt to changes in object rotation and viewpoint positions. Therefore, the part-based representation method is robust and can better handle the challenge of 3D object recogni" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.784, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Fig. 2: Comparison between multi-view aggregation representation (a) and part-based representation (b). View aggregation methods integrate features from different views and use the aggregated features to represent 3D objects. However, since the views are not aligned, the aggregated features lack robustness, making it difficult to handle 3D recognition on arbitrary views. Additionally, view aggregation methods suffer from information loss, thus unable to fully utilize the features from each view. We newly propose a part-based representation that focuses on multi-view part awareness and combines multiple parts to robustly represent 3D objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.785, + 0.529 + ], + "angle": 0, + "content": "tion under arbitrary views. We also provide comprehensive experimental results in Sec. 4 to demonstrate the effectiveness of part-based representation. When switching to arbitrary views, we do not observe significant performance degradation in our method, and it even outperforms most fixed viewpoint methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.667 + ], + "angle": 0, + "content": "Specifically, this paper proposes a novel part-aware network (PANet). Firstly, it employs weakly supervised learning to locate and recognize parts in each view, obtaining local view part features. Then, local view parts are input into the adaptive part refinement (APR) module. This module adaptively perceives the correlation of local view parts and further integrates these view parts to obtain global parts, which aims to adapt to the viewpoint changes of parts across different views. Finally, we combine these global parts to represent a 3D object. It is worth noting that we also introduce a regularization loss in the part localization module to ensure that part features can better focus on the object." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.669, + 0.744, + 0.684 + ], + "angle": 0, + "content": "The primary contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.778 + ], + "angle": 0, + "content": "- Unlike existing view aggregation methods, we propose a novel part-based representation. We consider a 3D object as a collection of part features rather than an abstract multi-view aggregated feature. This part-based representation is not limited by the position and quantity of viewpoints or affected by object poses, making it more flexibly applicable to various viewpoint scenarios, especially in arbitrary views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "- Further, we newly develop a Part-Aware Network (PANet), for extracting parts in multiple views. The network employs a weakly supervised method to localize parts on each view, enabling it to obtain local view parts. Then, we further refines these view parts through the proposed Adaptive Part Refinement" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.722, + 0.129 + ], + "angle": 0, + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "(APR) module to obtain global parts. These global parts can better adapt to differences in parts across viewpoints. Finally, integrating these global parts forms a comprehensive description of a 3D object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.269 + ], + "angle": 0, + "content": "- We conduct experiments on various datasets, including ScanObjectNN, ModelNet, and RGBD. Our method outperforms existing methods in both fixed views and arbitrary views. In particular, the performance in arbitrary views exceeds most existing fixed viewpoint methods based on fixed views, achieving robust recognition of 3D objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.396, + 0.309 + ], + "angle": 0, + "content": "2 Relative Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.784, + 0.431 + ], + "angle": 0, + "content": "3D object Recognition methods: 3D object recognition using deep neural networks can be broadly divided into three main categories: point-based methods [17,42,43], voxel-based methods [28,38,46], and view-based methods. View-based methods render 3D objects into multiple views and utilize multiple views to jointly represent 3D objects. Benefiting from mature and transferable 2D vision models and large-scale image datasets, they can better handle 3D object recognition tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.794 + ], + "angle": 0, + "content": "MVCNN [48] is a pioneering multi-view network for 3D object classification. It utilizes a view pooling strategy to integrate features from different views. Yang et al. [63] proposes a relational network to effectively connect corresponding regions from different viewpoints, thus enhancing the information of each view. 3D2Seq2views [18] method treats multiple views as a sequence and employs a convolutional neural network with novel hierarchical attention aggregation to aggregate sequential views. DRCNN [50] utilizes the dynamic routing algorithm of capsule networks to construct a new Dynamic Routing Layer (DRL), which integrates features from each view. View-GCN [56] employs multiple views as nodes in a view graph, leveraging a graph convolutional network (GCN) architecture to hierarchically aggregate information and learn robust 3D shape descriptors. These methods assume that each object has a predefined set of viewpoints and are trained and tested on these viewpoints. This assumption may limit model performance when dealing with 3D object recognition in arbitrary views. As far as we know, there are few studies [31, 34, 55] that have made significant progress beyond fixed viewpoint settings. The MVTN [34] implements viewpoint regression for 3D object recognition based on the latest research progress of differentiable rendering. The motivation of CVR [55] is to restore the intrinsic alignment of arbitrary views, transform it into aligned representations and then perform aggregation. In our work, we newly propose part-based representation. Part features exhibit viewpoint invariance and rotation robustness, making them more suitable for handling changes in viewpoint positions and quantities in arbitrary views. Moreover, the key idea of part-based representation is not to rely on multi-view aggregation, thus avoiding information loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Discriminative Part Localization: The perception and localization of object part-level features are usually challenging. Many existing approaches [25,35, 66, 67] suggest using bounding boxes and additional annotations to accurately" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.42 + ], + "angle": 0, + "content": "locate discriminative regions of objects. However, this method often requires expert knowledge and extensive annotation efforts, rendering it impractical for real-world applications. Fortunately, more and more research [14, 23, 45, 68, 69] has shown that important object regions can also be localized using weakly supervised methods. These methods rely solely on image-level supervision to identify part features. Early works usually locate class-specific discriminative regions via CAM [69]. Fu et al. proposes Recursive Attention CNN (RACNN) [14] to iteratively focus on predicting the locations of attention regions and extract discriminative features. Zheng et al. proposes Multi-Attention CNN (MACNN) [68], which proposes channel grouping loss to generate multiple parts by clustering. WSDAN [23] proposes weakly supervised attention learning to generate the distribution of parts. Meng et al. [39] introduces an end-to-end part discovery model (PDM) that aims to uncover robust and diverse object parts within a unified framework. Our PANet focuses on perceiving multiple discriminative regions in each view, extracting sufficient local view parts, and proposing the Adaptive Part Refinement (APR) module to further refine them to obtain global parts. These global parts can adapt to changes in viewpoint for parts across different views, making them more suitable for 3D objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.438, + 0.427, + 0.455 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.616 + ], + "angle": 0, + "content": "In this work, we newly propose a Part-Aware Network (PANet) that aims to solve the challenge of 3D object recognition (especially under arbitrary views). The overall architecture of the proposed PANet is shown in Fig. 3, consisting of three modules: weakly supervised part localization, adaptive part refinement, and joint part representation. We first introduce a weakly supervised method to perform part localization and feature extraction on each view (Sec. 3.1). Secondly, to adapt to the variations of part features across views, we propose the Adaptive Part Refinement (APR) module, which refines redundant part features from multiple views into more compact representations (Sec. 3.2). Finally, we integrate these refined features to represent a 3D object (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.569, + 0.649 + ], + "angle": 0, + "content": "3.1 Weakly-Supervised Part Localization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.787, + 0.728 + ], + "angle": 0, + "content": "It is worth noting that we do not rely on supervised positional annotations [67], such as bounding boxes or keypoints, for part localization. Instead, we employ a more cost-efficient weakly-supervised learning approach [23, 68] that utilizes image-level annotations to identify discriminative regions, which requires no expertise and no extensive annotation effort." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.788, + 0.806 + ], + "angle": 0, + "content": "We generate a multi-view sequence \\(\\mathcal{X} = \\{\\mathbf{X}_1, \\mathbf{X}_2, \\dots, \\mathbf{X}_v\\}\\) by rendering a 3D object from different viewpoints. Here, \\(v\\) denotes the number of views, \\(\\mathbf{X}_i\\) represents the \\(i\\)-th view, where \\(i \\in [1, v]\\). Then, each view is encoded by an image classification network e.g., ResNet [19] to obtain view features \\(\\mathbf{I} = \\{\\mathbf{I}_1, \\mathbf{I}_2, \\dots, \\mathbf{I}_v\\}\\), where \\(\\mathbf{I}_i \\in \\mathbb{R}^{H \\times W \\times C-1}\\), \\(i \\in [1, v]\\) is corresponding to the feature map of the \\(i\\)-th" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "1 H, W, and C represent the height, width, and number of channels of the feature layer, respectively" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.722, + 0.129 + ], + "angle": 0, + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.149, + 0.784, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Fig. 3: Overview of our proposed Part-Aware Network (PANet). The network takes multi-view images as input, encodes them by CNN, obtains view features \\(\\mathbf{I}\\), and uses the cross-view association (CVA) module to enhance the view features. Firstly, employ weakly-supervised methods to perceive part regions and generate view part sequences \\(\\mathbf{T}\\). Then, given some learnable part tokens \\(\\mathbf{P}\\), the sequences \\(\\mathbf{T}\\) and \\(\\mathbf{P}\\) are concatenated and input to the adaptive part refinement (APR) module. The module aims to refine the sequence \\(\\mathbf{T}\\) into a more compact representation, resulting in the global part features. Finally, multiple global parts \\(\\underline{\\mathbf{P}}\\) together serve as a representation of a 3D object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.785, + 0.581 + ], + "angle": 0, + "content": "view \\(\\mathbf{X}_i\\). To align the features from multiple views into the same feature space, we share the weights among each of the convolutional neural network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.785, + 0.642 + ], + "angle": 0, + "content": "Enhance Multi-view Features: However, the associations between view features \\(\\mathbf{I}\\) are not explicitly modeled. Therefore, we develop a cross-view association (CVA) module to obtain enhanced view features \\(\\mathbf{F}_i\\), enabling each view to incorporate information from other views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.718 + ], + "angle": 0, + "content": "For each view feature \\(\\mathbf{I}_i\\), the cross-view association (CVA) module calculates the matching score between it and the overall view feature \\(\\mathbf{I}\\). Here, the dot product function is selected as the matching function. Then, \\(\\mathbf{I}_i\\) is enhanced based on the correlation scores with other views. The enhanced view feature \\(\\mathbf{F}_i \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W} \\times \\mathrm{C}}\\) can be expressed in detail as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.724, + 0.787, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} = \\operatorname {s o f t m a x} \\left(\\omega (\\mathbf {I} _ {i}) ^ {T} \\cdot \\omega (\\mathbf {I})\\right) \\mathbf {I}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.697, + 0.765 + ], + "angle": 0, + "content": "where \\(\\omega (\\mathbf{I}_i)\\in \\mathbb{R}^{\\mathrm{C}}\\) represents the use of average pooling operations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Part Sampler: After passing through the cross-view association module, the enhanced view features \\(\\mathbf{F}\\) comprehensively incorporate information from other views. Next, we sample parts from the enhanced view features \\(\\mathbf{F}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Inspired by [23, 45], we implement the attention model \\(\\psi\\) to learn the spatial distribution of each part within each view. For view \\(i\\), the distribution of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.176 + ], + "angle": 0, + "content": "view parts is represented by the attention map \\(\\mathbf{A}_i\\in \\mathbb{R}^{\\mathrm{H}\\times \\mathrm{W}\\times \\mathrm{M}}\\), which can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.178, + 0.786, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {i} = \\psi (\\mathbf {F} _ {i}) = \\left\\{\\mathbf {A} _ {i, 1}, \\mathbf {A} _ {i, 2}, \\dots , \\mathbf {A} _ {i, \\mathrm {M}} \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.199, + 0.784, + 0.259 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}_{i,j} \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W}}\\), \\(i \\in [1,v]\\), \\(j \\in [1,\\mathrm{M}]\\) represents the \\(j\\)-th part region or visual pattern of the object's \\(i\\)-th views, such as a car wheel or an airplane wing. M denotes the number of attention maps, each view extracts the same number of regions of interest." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.26, + 0.785, + 0.305 + ], + "angle": 0, + "content": "Next, we employ the attention map \\(\\mathbf{A}_{i,j}\\) to softly weight the enhanced feature maps \\(\\mathbf{F}_i\\), followed by the global average pooling operation \\(\\omega\\) to obtain the parts on view \\(i\\) (See Eq. 3)." + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.307, + 0.786, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {t} _ {i, j} = \\omega \\left(\\mathbf {F} _ {i} \\otimes \\mathbf {A} _ {i, j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.766, + 0.342 + ], + "angle": 0, + "content": "where \\(\\pmb{t}_{i,j} \\in \\mathbb{R}^{\\mathrm{C}}\\), \\(\\otimes\\) denotes element-wise multiplication between two tensors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.784, + 0.373 + ], + "angle": 0, + "content": "Finally, we concatenate all the view part features together, resulting in a 3D multi-view part sequence \\(\\mathbf{T} \\in \\mathbb{R}^{N \\times C}\\) (see Eq. 4)." + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.386, + 0.786, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} = \\left[ \\boldsymbol {t} _ {0, 0}, \\boldsymbol {t} _ {0, 1}, \\dots , \\boldsymbol {t} _ {0, \\mathrm {M}}, \\boldsymbol {t} _ {1, 0}, \\boldsymbol {t} _ {1, 1}, \\dots , \\boldsymbol {t} _ {1, \\mathrm {M}}, \\dots , \\boldsymbol {t} _ {v, 0}, \\boldsymbol {t} _ {v, 1}, \\dots , \\boldsymbol {t} _ {v, \\mathrm {M}} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.412, + 0.736, + 0.428 + ], + "angle": 0, + "content": "where \\( \\mathrm{N} \\) represents the total number of parts \\( (v\\times \\mathrm{M}) \\) in the multi-view." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.45, + 0.48, + 0.466 + ], + "angle": 0, + "content": "3.2 Adaptive Part Refinement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.609 + ], + "angle": 0, + "content": "In Sec. 3.1, we have obtained a multi-view part sequence \\(\\mathbf{T}\\). However, there are differences in parts across different views, where each element \\(t_{ij}\\) in \\(\\mathbf{T}\\) can only represent local view features. It is challenging to generalize the overall features of a 3D object relying solely on local view parts. Therefore, we newly propose an Adaptive Part Refinement (APR) module, which aims to explicitly model the correlations between parts across views, thus enhancing the robustness of part features to viewpoint variations. This module adaptively integrates part information from multiple views and refines it to obtain global part features that are more suitable for 3D objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.784, + 0.701 + ], + "angle": 0, + "content": "Formally, the APR (Adaptive Part Refinement) module contains a transformer block. Given some global part tokens [4,7,11] \\(\\mathbf{P} \\in \\mathbb{R}^{\\mathrm{L} \\times \\mathrm{C}}\\), where \\(\\mathrm{L}\\) represents the number of parts. We concatenate \\(\\mathbf{P}\\) and \\(\\mathbf{T}\\) as input to the APR module. After passing through the APR module, local parts \\(\\pmb{t}_{i,j}\\) can learn cross-view correlations and obtain global parts \\(\\underline{\\mathbf{P}} \\in \\mathbb{R}^{\\mathrm{L} \\times \\mathrm{C}}\\) suitable for 3D objects. We also observe the diversity of global part features \\(\\underline{\\mathbf{P}}\\) (See Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.53, + 0.739 + ], + "angle": 0, + "content": "3.3 Joint Multi-Part Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.748, + 0.784, + 0.809 + ], + "angle": 0, + "content": "Relying solely on individual parts may not be sufficient to represent a 3D object. To overcome this limitation, we jointly describe the 3D object using multiple parts. We make predictions for each part and then average them to obtain the prediction for the 3D object. Each 3D object can be represented as Eq. 5:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.82, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\underline {{\\mathbf {P}}} = \\left[ \\underline {{\\boldsymbol {p}}} _ {1}, \\underline {{\\boldsymbol {p}}} _ {2}, \\dots , \\underline {{\\boldsymbol {p}}} _ {\\mathrm {L}} \\right], \\quad \\text {w h e r e} \\underline {{\\mathbf {P}}} \\in \\mathbb {R} ^ {\\mathrm {L} \\times \\mathrm {C}} \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.722, + 0.13 + ], + "angle": 0, + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "The proposed PANet is optimized through supervised learning, including instance classification loss \\(\\mathcal{L}_{ce}\\) and part-aware loss \\(\\mathcal{L}_{awe}\\). The loss function for multi-view \\(\\mathcal{X}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.275, + 0.201, + 0.785, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p ^ {i} = \\operatorname {s o f t m a x} \\left(\\underline {{\\boldsymbol {p}}} _ {i} \\boldsymbol {W} _ {p} + \\boldsymbol {b} _ {p}\\right), \\quad \\text {w h e r e} \\underline {{\\boldsymbol {p}}} _ {i} \\in \\mathbb {R} ^ {1 \\times \\mathrm {C}}, i \\in [ 1, \\mathrm {L} ] \\\\ q ^ {j} = \\operatorname {s o f t m a x} \\left(\\boldsymbol {t} _ {j} \\boldsymbol {W} _ {q} + \\boldsymbol {b} _ {q}\\right), \\quad \\text {w h e r e} \\boldsymbol {t} _ {j} \\in \\mathbb {R} ^ {1 \\times (\\mathrm {M} \\times \\mathrm {C})}, j \\in [ 1, v ], \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.275, + 0.257, + 0.71, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {a w e} = - \\sum_ {k = 1} ^ {K} y _ {k} \\log \\hat {p} _ {k} + \\gamma \\left(- \\frac {1}{v} \\sum_ {i = 1} ^ {v} \\sum_ {k = 1} ^ {K} y _ {k} \\log q _ {k} ^ {j}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.789, + 0.43 + ], + "angle": 0, + "content": "where \\( W_{p}, W_{q}, b_{p} \\), and \\( b_{q} \\) are learnable parameters. \\( p^{i} \\) represents the predicted probability of the \\( i \\)-th global part, \\( \\hat{p}_{k} \\) represents the average of the predicted probabilities for L global parts, and \\( q^{j} \\) represents the predicted probability for the \\( j \\)-th view part. \\( y_{k} \\) is the ground truth label of class \\( k \\). \\( \\mathcal{L}_{ce} \\) represents the cross-entropy loss of the final parts. \\( \\mathcal{L}_{awe} \\) represents the classification loss of M part features in each view, which is used to constrain the generation of parts. Furthermore, we apply techniques such as flipping and random erasure to augment the image for more robust part results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.452, + 0.377, + 0.469 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.54, + 0.499 + ], + "angle": 0, + "content": "4.1 Datasets and Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.735 + ], + "angle": 0, + "content": "We conduct an extensive evaluation of PANet on several benchmark datasets, namely ScanObjectNN [53], ModelNet [59], and RGBD [29]. In order to facilitate fair comparisons, we strictly adhere to the methodologies proposed in prior works [26, 48, 55] for preparing the training and testing data on the different datasets. Specifically, we adopt the training scheme introduced by MVCNN [48] and adopt the same viewpoint generation and rendering method as RotationNet [26] to handle the multi-view images from the ModelNet and ScanObjectNN datasets. Experiment on rotated datasets also follows CVR [55] setting. For arbitrary views, the number of views for each 3D object ranges randomly from 10 to 20. Throughout the training process, we utilize the AdamW [36] optimizer with a learning rate of 1e-5. We select ResNet50 as our backbone network and utilize label smoothing regularization [51] to enhance the generalization capability of our model. \\(\\gamma\\) is the weighting factor for the part-aware loss as in Eq. 6, which we set \\(\\gamma = 1\\) for the experiments. Please refer to the Appendix for implementation details of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.614, + 0.771 + ], + "angle": 0, + "content": "4.2 Comparison with State-of-the-Art Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The ScanObjectNN Datasets. As a recently introduced real-world 3D object classification dataset, ScanObjectNN [53] provides scanned indoor scene data. The dataset comprises approximately 15,000 objects, spanning 15 categories, and includes 2,902 unique object instances. Compared with CAD handcrafted" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.608, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Classification on ScanObjectNN. 'Aligned' means that every 3D object is in the same pose. 'Rotated' means that each 3D object is misaligned. '-' indicates that the specific result is not reported." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.787, + 0.366 + ], + "angle": 0, + "content": "
MethodTypeAlignedRotated
Per Class Acc.Per Ins. Acc.Per Class Acc.Per Ins. Acc.
PointNet++ [43]Point82.10%84.30%--
DGCNN [54]84.00%86.20%--
MVCNN [48]85.71%87.82%78.21%80.62%
GVCNN [13]86.64%88.68%82.86%83.70%
RotationNet [26]View (fixed)84.88%86.90%74.68%76.16%
View-GCN [56]88.67%90.39%81.99%83.50%
CVR [55]88.39%90.74%84.70%85.59%
View-GCN++ [57]89.10%91.30%85.80%86.80%
Ours93.21%93.99%86.15%87.82%
CVR [55](arbitrary)--68.07%71.36%
Ours91.77%93.48%84.08%86.45%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.789, + 0.646 + ], + "angle": 0, + "content": "datasets, ScanObjectNN offers more realistic challenges as it reflects the noise, complexity, and incompleteness present in real-world environments, incorporating the appearance and geometric shapes of actual objects. Consequently, it is better for assessing the robustness of the model. We train PANet under different viewpoint settings and compare with point-based methods [43, 54] and view-based methods [13, 26, 48, 55-57], as shown in Tab. 1. View-based methods primarily rely on view aggregation, and it is difficult to robustly aggregate features from structurally misaligned views, especially for arbitrary views. When switching to arbitrary views, the performance of these methods drops sharply. Our PANet is observed to significantly outperform previous approaches in both class accuracy and instance accuracy, whether in aligned views, rotated views, or arbitrary views. In aligned views, compared with the current state-of-the-art methods e.g., View-GCN++ [57] and CVR [55], PANet improves instance accuracy and class accuracy by \\(2.69\\%\\) and \\(4.11\\%\\). PANet also achieves significant improvements of \\(1.02\\%\\) and \\(0.35\\%\\) in rotated views. Surprisingly, even under arbitrary views, our performance surpasses that of most fixed viewpoint methods, demonstrating that PANet can robustly recognize 3D objects unconstrained by the scene. Please refer to the Appendix for more detailed comparisons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.79, + 0.844 + ], + "angle": 0, + "content": "The ModelNet Datasets. ModelNet [59] is a dataset of hand-drawn point clouds generated from CAD models. It comprises 12,311 3D shapes, covering 40 categories, with 9,483 models allocated for training and 2,468 for testing. The experimental results on the ModelNet dataset are presented in Tab. 2. We first compare with point-based and voxel-based methods, including PointNet++ [43], Kd-Networks [27], RS-CNN [22], KP-Convs [52], MVTN [34], 3DShapeNets [58], VoxNet [38], and VRENsemble [3]. Our PANet outperforms them by more than \\(3.93\\%\\) and \\(4.6\\%\\) in instance accuracy and class accuracy and is also \\(3.73\\%\\) and \\(4.37\\%\\) higher in free viewpoints. We also compared with view-based approaches, including MVCNN-new [48], GVCNN [13], HEAR [5], RotationNet [26], View-GCN [56], MVT [6], HMVCM [32], MVD [24], DAN [41], DRCNN [50], EMV [12], View-GCN++ [57], CVR [55], and HGNN+ [16]. PANet consistently outper" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.72, + 0.129 + ], + "angle": 0, + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.493, + 0.215 + ], + "angle": 0, + "content": "Table 2: Classification on Aligned ModelNet40. '- indicates that the specific result is not reported, 'fixed' denotes sampling from a fixed viewpoint. 'free' means that the arbitrary number of viewpoints." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.226, + 0.49, + 0.521 + ], + "angle": 0, + "content": "
Aligned ModelNet40
MethodTypeClass Acc.Ins. Acc.
3DShapeNets [58]77.3%-
VoxNet [38]Voxel83.0%-
VRENsemble [3]91.4%93.8%
PointNet++ [43]-91.9%
Kd-Networks [27]88.5%91.8%
RS-CNN [22]Point-93.6%
KP-Convs [52]-92.9%
MVTN [34]92.0%93.8%
MVCNN-new [48]92.40%95.00%
GVCNN [13]90.70%93.10%
HEAR [5]95.20%96.70%
RotationNet [26]-97.37%
View-GCN [56]96.50%97.60%
MVT [6]-97.50%
HMVCM [32]-94.57%
MVD [24]View92.61%94.41%
DAN [41]-93.50%
DRCNN [50]94.86%96.84%
View-GCN++ [57]96.50%97.60%
CVR [55]95.77%97.16%
HGNN+ [16]-96.92%
Ours96.60%97.73%
Ours(free)96.37%97.53%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.145, + 0.786, + 0.172 + ], + "angle": 0, + "content": "Table 3: The accuracy of 3D object classification on Rotated ModelNet40." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.185, + 0.785, + 0.333 + ], + "angle": 0, + "content": "
Rotated ModelNet40
Method#ViewClass Acc.Ins. Acc.
MVCNN [48]87.95%88.17%
GVCNN [13]89.69%88.10%
EMV [12]88.80%90.00%
RotationNet [26]fixed84.74%85.29%
View-GCN [56]85.90%88.25%
CVR [55]91.12%92.22%
View-GCN++ [57]90.70%92.80%
Ours90.35%92.87%
CVR [55]arbitrary84.01%86.91%
Ours90.69%92.23%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.335, + 0.785, + 0.362 + ], + "angle": 0, + "content": "Table 4: The accuracy of object classification on RGBD (in%)." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.375, + 0.784, + 0.52 + ], + "angle": 0, + "content": "
Method#ViewIns. Acc.
MDSICNN [2]≥ 12089.6%
CFK [8]≥ 12086.8%
MMDCNN [44]≥ 12086.8%
RotationNet [26]1289.3%
View-GCN [56]1294.3%
View-GCN++ [57]1292.5%
Ours1296.5%
MVCNN [48]89.0%
GVCNN [13]89.8%
CVR [55]4-1291.8%
Ours95.9%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.785, + 0.586 + ], + "angle": 0, + "content": "forms these methods in fixed views. Its performance at free viewpoints surpasses that of most existing approaches. We next report the robustness of PANet in rotated views, as shown in Tab. 3. Under the rotated views, PANet also achieves the highest score in instance accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.769 + ], + "angle": 0, + "content": "The RGBD Datasets. The RGBD [29] dataset comprises 51 categories of domestic objects captured in the real world, encompassing RGBD views from multiple viewpoints. To ensure a fair comparison, we adopt the identical experimental setup as RotationNet [26] and View-GCN [56]. Specifically, we uniformly sample 12 RGB images from the same circle for each 3D object as input and report the average results using ten-fold cross-validation. We keep the same settings as CVR [55] in arbitrary views, sampling images from 4 to 12 viewpoints as input to the network. The experimental results are summarized in Tab. 4. The performance of PANet in arbitrary views is impressive. Compared with work in arbitrary views e.g., CVR [55], PANet significantly improves by \\(4.1\\%\\). Additionally, compared with the work in fixed views, such as view-GCN [56], PANet has also improved by \\(1.6\\%\\). This further validates the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.391, + 0.804 + ], + "angle": 0, + "content": "4.3 Model Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "All experiments are conducted on the ScanObjectNN dataset in rotated views. Please refer to the Appendix for detailed experimental analyses." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.148, + 0.337, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.149, + 0.478, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.238, + 0.493, + 0.294 + ], + "angle": 0, + "content": "Fig.4: Visualization of correlations between part features \\(\\underline{\\mathbf{P}}\\). The left figure illustrates the correlation between \\(\\underline{\\mathbf{P}}\\) after applying \\(\\mathcal{L}_{awe}\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.147, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 5: PANet component ablations. The first line serves as the baseline. \\( \\mathcal{L}_{awe} \\) refers to the weakly supervised part-aware loss function. CVA stands for Cross-View Association module." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.235, + 0.787, + 0.293 + ], + "angle": 0, + "content": "
CVALaweClass Acc.Ins. Acc.
✓83.76%84.56%
84.99%85.25%
✓85.45%86.96%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.306, + 0.76, + 0.321 + ], + "angle": 0, + "content": "Table 6: Explore the impact of different viewpoint sampling on performance." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.333, + 0.774, + 0.423 + ], + "angle": 0, + "content": "
DatasetMethodAlignedRotated
Class Acc.Ins. Acc.Class Acc.Ins. Acc.
ModelNet40random sampling96.37%97.53%90.69%92.23%
furthest point sampling95.44%97.41%89.73%92.26%
ScanObjectNNrandom sampling91.77%93.48%84.08%86.45%
furthest point sampling92.97%93.48%84.80%85.93%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.787, + 0.626 + ], + "angle": 0, + "content": "PANet component ablations We evaluate the role of the cross-view association module and \\(\\mathcal{L}_{awe}\\), as shown in Tab. 5. The CVA module plays a positive role in enhancing the performance of PANet, improving instance accuracy and class accuracy by \\(0.69\\%\\) and \\(1.23\\%\\). The results show that the CVA module enables each view to more effectively integrate information from other views, further enhancing the representation capability of the view features. Our method also improves by \\(2.40\\%\\) and \\(1.76\\%\\) in two metrics by introducing the \\(\\mathcal{L}_{awe}\\) constraint. In addition, we also conduct correlation analysis between part features \\(\\underline{\\mathbf{P}}\\), as shown in Fig. 4. We observe that without applying \\(\\mathcal{L}_{awe}\\) constraints, the correlation between parts exhibits a flat similarity matrix, indicating that the parts are unvaried and cannot be distinguished in the feature space. In contrast, the part features become more diverse with \\(\\mathcal{L}_{awe}\\) constraint. Therefore, we conclude that \\(\\mathcal{L}_{awe}\\) can supervise the perception of parts to obtain diverse part features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Influence of Different Viewpoint Sampling We introduce two different sampling methods: random sampling and furthest point sampling (FPS). Random sampling refers to the process of randomly selecting 10 to 20 views from each object. This method is closer to real-world scenarios, and we showcased PANet's performance in Tab. 1, 2, 3, 4 based on this sampling method. FPS selects 10 to 20 views based on the furthest point sampling principle. It provides a uniform sampling method that covers every orientation of the 3D objects. As shown in Tab. 6, we observed that the performance of random sampling and FPS across different datasets is nearly similar. This result shows that PANet is not sensitive to the choice of sampling methods. The model demonstrates robustness in adapting to changes in viewpoint caused by different sampling methods while maintaining stable performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Influence of the Number of the Views We further evaluate PANet by considering different backbone networks and numbers of views, as shown in Tab. 7." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.721, + 0.129 + ], + "angle": 0, + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.145, + 0.772, + 0.16 + ], + "angle": 0, + "content": "Table 7: Explore the impact of number of views and backbones on performance." + }, + { + "type": "table", + "bbox": [ + 0.266, + 0.171, + 0.74, + 0.24 + ], + "angle": 0, + "content": "
BackboneViews
1 views5 views10 views15 views20 views
resnet1845.63%70.84%75.66%79.90%82.85%
resnet5067.07%79.59%82.50%85.93%86.62%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.245, + 0.493, + 0.287 + ], + "angle": 0, + "content": "Table 8: Exploring the effect of the number of attention maps, M denotes the number of attention maps per view." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.299, + 0.484, + 0.366 + ], + "angle": 0, + "content": "
#Attention MapsClass Acc.Ins. Acc.
M = 1684.57%86.28%
M = 3285.36%87.14%
M = 6486.50%87.31%
M = 12886.38%87.14%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.245, + 0.787, + 0.287 + ], + "angle": 0, + "content": "Table 9: Exploring the effect of the number of part tokens, L indicates the number of parts." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.3, + 0.78, + 0.366 + ], + "angle": 0, + "content": "
#Number of PartsClass Acc.Ins. Acc.
L = 185.50%86.28%
L = 886.20%87.14%
L = 1686.50%87.31%
L = 3285.63%86.31%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.787, + 0.584 + ], + "angle": 0, + "content": "In previous studies, such as View-GCN [56], because view aggregation leads to feature loss. Choosing a stronger backbone network may not necessarily lead to performance improvement in certain dataset. However, our observations reveal that the choice of backbone significantly affects PANet performance. This is because PANet directly integrates multiple parts to characterize 3D objects. Therefore, a stronger backbone network is more advantageous for perceiving part features. Furthermore, we notice a positive correlation between the number of views and the performance of PANet. It is worth noting that previous studies on multi-view aggregation methods, such as MHBN [65], MVLADN [64], and GVCNN [13], have indicated that increasing the number of views does not necessarily lead to better representation of 3D objects. PANet combines multiple parts for representation, and the APR module can also filter out some redundant parts. This means that PANet is not affected by redundant views. Our experimental results also verify this idea." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Influence of the Number of the Attention Maps and Parts We show the impact of the number of attention maps M and the number of part tokens L, as shown in Tab. 8 and Tab. 9. The number of attention maps M determines the discriminative region in each view, and the part tokens L determine the number of global parts. More M and L usually contribute to better performance. MACNN [68] and WSDAN [23] report similar conclusions. The experiments in Tab. 8 show that as the value of M increases, the classification accuracy gradually improves. When \\( \\mathrm{M} = 64 \\), the performance tends to stabilize. The results in Tab. 9 show that the performance improves as the number of parts increases and reaches a peak at \\( \\mathrm{L} = 16 \\). Compared with a single part, multiple parts provide richer information and enhance the generalization of the model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.371, + 0.785 + ], + "angle": 0, + "content": "4.4 Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We randomly select some 3D models from a test set and compare the semantic differences between the features extracted by MVCNN and PANet, as shown in Fig. 5. MVCNN [48] is a classic feature aggregation method. We notice that the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.608, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "features extracted by MVCNN are not entirely focused on the object. This is because the uneven weight distribution leads to view aggregation methods ignoring certain view information. It is worth noting that PANet does not discard any views but instead considers the features of each view, leveraging the valuable information within each view. We can observe that different views emphasize various part locations, with each view capturing multiple discriminative regions (e.g., the airplane nose, body, and wings). This observation further validates the diversity of captured parts. Furthermore, we observe that the view aggregation method does not explicitly model the unique attributes of 3D objects, leading to feature ambiguity. In contrast, part-based representation reveals the inherent properties of objects, enhancing the interpretability of features and making recognition easier. Please refer to the Appendix for more visualization results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.36, + 0.368 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.595 + ], + "angle": 0, + "content": "In this paper, we propose a novel part-aware network (PANet), which is a part-based representation. Compared with view aggregation methods, this part-based representation has properties such as viewpoint invariance and rotation robustness, which can more effectively address the 3D object recognition problem under arbitrary views. Moreover, we proposed the part-aware loss to ensure that part features can better focus on the object, thereby fully utilizing the information from each view. We evaluate the performance of PANet on several 3D datasets, including ScanObjectNN, ModelNet, and RGBD. The experimental results show that PANet has made significant progress, especially in arbitrary views. We also conduct ablation and visualization experiments to study the impact of different components in our method. It is worth noting that in these experiments, we found that our method is not affected by redundant views (Sec. 4.3), and the part features can better reveal the inherent properties of objects (Sec. 4.4). In conclusion, the experimental results prove the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.609, + 0.787, + 0.651 + ], + "angle": 0, + "content": "Acknowledgement This research is supported by the Shenzhen Science and Technology Plan Project No. KJZD20230923113800002 and Shenzhen LongHua Fundamental Research Program No. 10162A20230325B73A546." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.721, + 0.129 + ], + "angle": 0, + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.199, + 0.495, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.199, + 0.781, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.757, + 0.788, + 0.784 + ], + "angle": 0, + "content": "Fig. 5: Comparing the semantic differences between view-based aggregation methods and PANet in feature extraction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.236 + ], + "angle": 0, + "content": "1. Afham, M., Dissanayake, I., Dissanayake, D., Dharmasiri, A., Thilakarathna, K., Rodrigo, R.: Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9902-9912 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.786, + 0.278 + ], + "angle": 0, + "content": "2. Asif, U., Bennamoun, M., Sohel, F.A.: A multi-modal, discriminative and spatially invariant cnn for rgb-d object labeling. IEEE transactions on pattern analysis and machine intelligence 40(9), 2051-2065 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.786, + 0.32 + ], + "angle": 0, + "content": "3. Brock, A., Lim, T., Ritchie, J.M., Weston, N.: Generative and discriminative voxel modeling with convolutional neural networks. arXiv preprint arXiv:1608.04236 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.786, + 0.377 + ], + "angle": 0, + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.786, + 0.433 + ], + "angle": 0, + "content": "5. Chen, J., Qin, J., Shen, Y., Liu, L., Zhu, F., Shao, L.: Learning attentive and hierarchical representations for 3d shape recognition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 105-122. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.786, + 0.46 + ], + "angle": 0, + "content": "6. Chen, S., Yu, T., Li, P.: Mvt: Multi-view vision transformer for 3d object recognition. arXiv preprint arXiv:2110.13083 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.786, + 0.502 + ], + "angle": 0, + "content": "7. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.503, + 0.786, + 0.544 + ], + "angle": 0, + "content": "8. Cheng, Y., Cai, R., Zhao, X., Huang, K.: Convolutional fisher kernels for rgb-d object recognition. In: 2015 International Conference on 3D Vision. pp. 135-143. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.786, + 0.587 + ], + "angle": 0, + "content": "9. Delitzas, A., Parelli, M., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.786, + 0.629 + ], + "angle": 0, + "content": "0. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.786, + 0.685 + ], + "angle": 0, + "content": "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.786, + 0.728 + ], + "angle": 0, + "content": "2. Esteves, C., Xu, Y., Allen-Blanchette, C., Daniilidis, K.: Equivariant multi-view networks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1568-1577 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.786, + 0.77 + ], + "angle": 0, + "content": "3. Feng, Y., Zhang, Z., Zhao, X., Ji, R., Gao, Y.: GvCNN: Group-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 264-272 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.786, + 0.812 + ], + "angle": 0, + "content": "4. Fu, J., Zheng, H., Mei, T.: Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4438-4446 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Gao, K., Gao, Y., He, H., Lu, D., Xu, L., Li, J.: Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.721, + 0.129 + ], + "angle": 0, + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Gao, Y., Feng, Y., Ji, S., Ji, R.: Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3181-3199 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.285 + ], + "angle": 0, + "content": "18. Han, Z., Lu, H., Liu, Z., Vong, C.M., Liu, Y.S., Zwicker, M., Han, J., Chen, C.P.: 3d2seqviews: Aggregating sequential views for 3d global feature learning by cnn with hierarchical attention aggregation. IEEE Transactions on Image Processing 28(8), 3986-3999 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.787, + 0.326 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.326, + 0.787, + 0.353 + ], + "angle": 0, + "content": "20. Herreras, E.B.: Cognitive neuroscience; the biology of the mind. Cuadernos de Neuropsicología/Panamerican Journal of Neuropsychology 4(1), 87-90 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.353, + 0.787, + 0.393 + ], + "angle": 0, + "content": "21. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9202-9212 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.393, + 0.787, + 0.42 + ], + "angle": 0, + "content": "22. Hu, L., Qin, M., Zhang, F., Du, Z., Liu, R.: Rscnn: a cnn-based method to enhance low-light remote-sensing images. Remote Sensing 13(1), 62 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "23. Hu, T., Qi, H., Huang, Q., Lu, Y.: See better before looking closer: Weakly supervised data augmentation network for fine-grained visual classification. arXiv preprint arXiv:1901.09891 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.787, + 0.501 + ], + "angle": 0, + "content": "24. Huang, J., Yan, W., Li, T., Liu, S., Li, G.: Learning the global descriptor for 3-d object recognition based on multiple views decomposition. IEEE Transactions on Multimedia 24, 188-201 (2022). https://doi.org/10.1109/TMM.2020.3047762" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "25. Huang, S., Xu, Z., Tao, D., Zhang, Y.: Part-stacked cnn for fine-grained visual categorization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1173-1182 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "26. Kanezaki, A., Matsushita, Y., Nishida, Y.: Rotationnet for joint object categorization and unsupervised pose estimation from multi-view images. IEEE transactions on pattern analysis and machine intelligence 43(1), 269-283 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "27. Klokov, R., Lempitsky, V.: Escape from cells: Deep kd-networks for the recognition of 3d point cloud models. In: Proceedings of the IEEE international conference on computer vision. pp. 863-872 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.623, + 0.787, + 0.663 + ], + "angle": 0, + "content": "28. Kumawat, S., Raman, S.: Lp-3dcnn: Unveiling local phase in 3d convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4903-4912 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.663, + 0.787, + 0.704 + ], + "angle": 0, + "content": "29. Lai, K., Bo, L., Ren, X., Fox, D.: A large-scale hierarchical multi-view rgb-d object dataset. In: 2011 IEEE international conference on robotics and automation. pp. 1817-1824. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.787, + 0.759 + ], + "angle": 0, + "content": "30. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "31. Lin, Y., Gou, Y., Liu, X., Bai, J., Lv, J., Peng, X.: Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4447-4461 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "32. Liu, A.A., Zhou, H., Nie, W., Liu, Z., Liu, W., Xie, H., Mao, Z., Li, X., Song, D.: Hierarchical multi-view context modelling for 3d object classification and retrieval. Information Sciences 547, 984-995 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "34. Liu, S., Nguyen, V., Rehg, I., Tu, Z.: Recognizing objects from any view with object and viewer-centered representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "35. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing 30, 1744-1758 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "36. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "37. Lowe, D.: Object recognition from local scale-invariant features. In: Proceedings of the Seventh IEEE International Conference on Computer Vision. vol. 2, pp. 1150-1157 vol.2 (1999). https://doi.org/10.1109/ICCV.1999.790410" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.384 + ], + "angle": 0, + "content": "38. Maturana, D., Scherer, S.: Voxnet: A 3d convolutional neural network for real-time object recognition. In: 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 922-928. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.437 + ], + "angle": 0, + "content": "39. Meng, M., Zhang, T., Yang, W., Zhao, J., Zhang, Y., Wu, F.: Diverse complementary part mining for weakly supervised object localization. IEEE Transactions on Image Processing 31, 1774-1788 (2022). https://doi.org/10.1109/TIP.2022.3145238" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.481 + ], + "angle": 0, + "content": "40. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.509 + ], + "angle": 0, + "content": "41. Nie, W., Zhao, Y., Song, D., Gao, Y.: Dan: Deep-attention network for 3d shape recognition. IEEE Transactions on Image Processing 30, 4371-4383 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.55 + ], + "angle": 0, + "content": "42. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.592 + ], + "angle": 0, + "content": "43. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.633 + ], + "angle": 0, + "content": "44. Rahman, M.M., Tan, Y., Xue, J., Lu, K.: Rgb-d object recognition with multimodal deep convolutional neural networks. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 991-996. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.675 + ], + "angle": 0, + "content": "45. Rao, Y., Chen, G., Lu, J., Zhou, J.: Counterfactual attention learning for fine-grained visual categorization and re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1025–1034 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.716 + ], + "angle": 0, + "content": "46. Riegler, G., Osman Ulusoy, A., Geiger, A.: Octnet: Learning deep 3d representations at high resolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3577-3586 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.758 + ], + "angle": 0, + "content": "47. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.799 + ], + "angle": 0, + "content": "48. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.842 + ], + "angle": 0, + "content": "49. Su, J.C., Gadelha, M., Wang, R., Maji, S.: A deeper look at 3d shape classifiers. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.721, + 0.129 + ], + "angle": 0, + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "50. Sun, K., Zhang, J., Liu, J., Yu, R., Song, Z.: Drcnn: Dynamic routing convolutional neural network for multi-view 3d object recognition. IEEE Transactions on Image Processing 30, 868-877 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "51. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "52. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "53. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1588-1597 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "54. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog) 38(5), 1-12 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "55. Wei, X., Gong, Y., Wang, F., Sun, X., Sun, J.: Learning canonical view representation for 3d shape recognition with arbitrary views. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 407-416 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "56. Wei, X., Yu, R., Sun, J.: View-gcn: View-based graph convolutional network for 3d shape analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1850-1859 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "57. Wei, X., Yu, R., Sun, J.: Learning view-based graph convolutional network for multi-view 3d shape analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "58. Wu, S., Khosla, Y., Zhang, T.: D shapenets: A deep representation for volumetric shape modeling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA. pp. 7-12 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "59. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "60. Xu, Y., Zheng, C., Xu, R., Quan, Y., Ling, H.: Multi-view 3d shape recognition via correspondence-aware deep learning. IEEE Transactions on Image Processing 30, 5299-5312 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "61. Xue, L., Gao, M., Xing, C., Martin-Martin, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning unified representation of language, image and point cloud for 3d understanding. arXiv preprint arXiv:2212.05171 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "62. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "63. Yang, Z., Wang, L.: Learning relationships for multi-view 3d object recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7505-7514 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "64. Yu, T., Meng, J., Yang, M., Yuan, J.: 3d object representation learning: A set-to-set matching perspective. IEEE Transactions on Image Processing 30, 2168-2179 (2021). https://doi.org/10.1109/TIP.2021.3049968" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "65. Yu, T., Meng, J., Yuan, J.: Multi-view harmonized bilinear network for 3d object recognition. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 186-194 (2018). https://doi.org/10.1109/CVPR.2018.00027" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.607, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond Viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "66. Zhang, H., Xu, T., Elhoseiny, M., Huang, X., Zhang, S., Elgammal, A., Metaxas, D.: Spda-cnn: Unifying semantic part detection and abstraction for fine-grained recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1143-1152 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.259 + ], + "angle": 0, + "content": "67. Zhang, N., Donahue, J., Girshick, R., Darrell, T.: Part-based r-cnns for fine-grained category detection. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 834-849. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "68. Zheng, H., Fu, J., Mei, T., Luo, J.: Learning multi-attention convolutional neural network for fine-grained image recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 5209-5217 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "69. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_origin.pdf b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ce888f76dec4977079c5daa6af6ed85a393cbd6b --- /dev/null +++ b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/884854c6-757b-45ba-8fec-fe48f8559eb7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cd9edc7f0dd293ed929bdadbf89ed8e7e7e5dc236ca717ef5e7b20f0a1fed5c +size 7284354 diff --git a/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/full.md b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8b023a7fe92c4f2115b1a432a7d6663404213917 --- /dev/null +++ b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/full.md @@ -0,0 +1,293 @@ +# Beyond Viewpoint: Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation + +Linlong Fan $^{\text{日}}$ , Ye Huang $^{\text{日}^{\star}}$ , Yanqi Ge $^{\text{日}}$ , Wen Li $^{\text{日}}$ , and Lixin Duan $^{\text{日}}$ + +Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China fanlinlong703@163.com + +Abstract. Existing view-based methods excel at recognizing 3D objects from predefined viewpoints, but their exploration of recognition under arbitrary views is limited. This is a challenging and realistic setting because each object has different viewpoint positions and quantities, and their poses are not aligned. However, most view-based methods, which aggregate multiple view features to obtain a global feature representation, hard to address 3D object recognition under arbitrary views. Due to the unaligned inputs from arbitrary views, it is challenging to robustly aggregate features, leading to performance degradation. In this paper, we introduce a novel Part-aware Network (PANet), which is a part-based representation, to address these issues. This part-based representation aims to localize and understand different parts of 3D objects, such as airplane wings and tails. It has properties such as viewpoint invariance and rotation robustness, which give it an advantage in addressing the 3D object recognition problem under arbitrary views. Our results on benchmark datasets clearly demonstrate that our proposed method outperforms existing view-based aggregation baselines for the task of 3D object recognition under arbitrary views, even surpassing most fixed viewpoint methods. + +Keywords: 3D object recognition $\cdot$ weakly-supervised learning + +# 1 Introduction + +Learning 3D with multi-view supervision has received widespread attention. It covers various research domains such as NeRFs [15,40], BEV detection [30,62], 3D generation [33], 3D reasoning [9,21], etc., among which 3D object recognition is considered one of the most fundamental problems. Thanks to mature and transferable 2D vision models [11,47] and large-scale image datasets [10], rendering 3D objects into multiple 2D images and applying 2D networks to indirectly recognize 3D objects has become the state-of-the-art method [57]. Furthermore, compared with direct methods (e.g., point-based [1,61], voxel-based [38]), view-based methods are closer to human perception of 3D objects in the real world. + +![](images/b3e50804339c196a8b81f296bd240f8cc9be1db52975e43fe2dde2591e373aff.jpg) +Fig. 1: In practical scenarios, 3D objects are observed from arbitrary views(right). The definition of arbitrary views is that each object is unaligned, and the viewpoint positions and quantities vary for each object. Previous works have focused on 3D object recognition in aligned views(left), where the pose of each object is aligned, and the positions and quantities of viewpoints are predefined. Rotated views(middle) is an extension that introduces random rotation for each 3D object while keeping the viewpoint positions and quantities unchanged. Our work focuses on arbitrary views. + +Previous view-based methods [5,6,49,50,56] focus on addressing 3D object recognition tasks under fixed viewpoints, while recognition under arbitrary views still lacks extensive research. As shown in Fig. 1, arbitrary views represent a more challenging setting where each object may have different viewpoint positions and quantities, and the poses of each object are not aligned. + +However, most view-based methods [5,6,13,18,26,41,49,50,55-57,60,63-65], which aggregate multiple view features to obtain a global feature representation, hard to effectively address 3D object recognition under arbitrary views. This is because the misaligned nature of arbitrary views complicates the representation and aggregation of view features, which leads to performance degradation. For example, when applied to arbitrary views, the performance of CVR [55] drops by $6.31\%$ and $14.23\%$ in ModelNet [59] and ScanObjectNN [53] datasets. Even some multi-view aggregation methods [26, 56] cannot be applied to arbitrary views because they require fixed viewpoint position information. Furthermore, view aggregation methods [13, 49] typically jointly optimize feature extraction and aggregation modules to maximize performance, but this may lead the extraction module to overly focus on certain views while ignoring information from others, resulting in information loss. + +To address these challenges, we propose a novel part-based representation, as shown in Fig. 2. This idea is inspired by cognitive science and biology [20], which aims to understand different part features or specific regions in a multi-view sequence to better represent 3D objects. For example, a person can be composed of multiple parts with unique attributes, such as the head, arms, body, legs, and so on. By locating and integrating the features of these parts, we can more accurately represent a person. Compared with the multi-view aggregation features, part features offer several advantages. Firstly, an individual view contains abundant part information. Even if only a few views are utilized, we can still extract stable part features, thereby adapting to changes in the number of viewpoints in arbitrary view settings. Secondly, the part features are more stable than the multi-view aggregation features [37], which can better adapt to changes in object rotation and viewpoint positions. Therefore, the part-based representation method is robust and can better handle the challenge of 3D object recogni + +![](images/a987c554f2db6f2aec74bd725338d32c96de92fb3f82578c570f2d509551cf4f.jpg) +Fig. 2: Comparison between multi-view aggregation representation (a) and part-based representation (b). View aggregation methods integrate features from different views and use the aggregated features to represent 3D objects. However, since the views are not aligned, the aggregated features lack robustness, making it difficult to handle 3D recognition on arbitrary views. Additionally, view aggregation methods suffer from information loss, thus unable to fully utilize the features from each view. We newly propose a part-based representation that focuses on multi-view part awareness and combines multiple parts to robustly represent 3D objects. + +tion under arbitrary views. We also provide comprehensive experimental results in Sec. 4 to demonstrate the effectiveness of part-based representation. When switching to arbitrary views, we do not observe significant performance degradation in our method, and it even outperforms most fixed viewpoint methods. + +Specifically, this paper proposes a novel part-aware network (PANet). Firstly, it employs weakly supervised learning to locate and recognize parts in each view, obtaining local view part features. Then, local view parts are input into the adaptive part refinement (APR) module. This module adaptively perceives the correlation of local view parts and further integrates these view parts to obtain global parts, which aims to adapt to the viewpoint changes of parts across different views. Finally, we combine these global parts to represent a 3D object. It is worth noting that we also introduce a regularization loss in the part localization module to ensure that part features can better focus on the object. + +The primary contributions of this work can be summarized as follows: + +- Unlike existing view aggregation methods, we propose a novel part-based representation. We consider a 3D object as a collection of part features rather than an abstract multi-view aggregated feature. This part-based representation is not limited by the position and quantity of viewpoints or affected by object poses, making it more flexibly applicable to various viewpoint scenarios, especially in arbitrary views. +- Further, we newly develop a Part-Aware Network (PANet), for extracting parts in multiple views. The network employs a weakly supervised method to localize parts on each view, enabling it to obtain local view parts. Then, we further refines these view parts through the proposed Adaptive Part Refinement + +(APR) module to obtain global parts. These global parts can better adapt to differences in parts across viewpoints. Finally, integrating these global parts forms a comprehensive description of a 3D object. + +- We conduct experiments on various datasets, including ScanObjectNN, ModelNet, and RGBD. Our method outperforms existing methods in both fixed views and arbitrary views. In particular, the performance in arbitrary views exceeds most existing fixed viewpoint methods based on fixed views, achieving robust recognition of 3D objects. + +# 2 Relative Work + +3D object Recognition methods: 3D object recognition using deep neural networks can be broadly divided into three main categories: point-based methods [17,42,43], voxel-based methods [28,38,46], and view-based methods. View-based methods render 3D objects into multiple views and utilize multiple views to jointly represent 3D objects. Benefiting from mature and transferable 2D vision models and large-scale image datasets, they can better handle 3D object recognition tasks. + +MVCNN [48] is a pioneering multi-view network for 3D object classification. It utilizes a view pooling strategy to integrate features from different views. Yang et al. [63] proposes a relational network to effectively connect corresponding regions from different viewpoints, thus enhancing the information of each view. 3D2Seq2views [18] method treats multiple views as a sequence and employs a convolutional neural network with novel hierarchical attention aggregation to aggregate sequential views. DRCNN [50] utilizes the dynamic routing algorithm of capsule networks to construct a new Dynamic Routing Layer (DRL), which integrates features from each view. View-GCN [56] employs multiple views as nodes in a view graph, leveraging a graph convolutional network (GCN) architecture to hierarchically aggregate information and learn robust 3D shape descriptors. These methods assume that each object has a predefined set of viewpoints and are trained and tested on these viewpoints. This assumption may limit model performance when dealing with 3D object recognition in arbitrary views. As far as we know, there are few studies [31, 34, 55] that have made significant progress beyond fixed viewpoint settings. The MVTN [34] implements viewpoint regression for 3D object recognition based on the latest research progress of differentiable rendering. The motivation of CVR [55] is to restore the intrinsic alignment of arbitrary views, transform it into aligned representations and then perform aggregation. In our work, we newly propose part-based representation. Part features exhibit viewpoint invariance and rotation robustness, making them more suitable for handling changes in viewpoint positions and quantities in arbitrary views. Moreover, the key idea of part-based representation is not to rely on multi-view aggregation, thus avoiding information loss. + +Discriminative Part Localization: The perception and localization of object part-level features are usually challenging. Many existing approaches [25,35, 66, 67] suggest using bounding boxes and additional annotations to accurately + +locate discriminative regions of objects. However, this method often requires expert knowledge and extensive annotation efforts, rendering it impractical for real-world applications. Fortunately, more and more research [14, 23, 45, 68, 69] has shown that important object regions can also be localized using weakly supervised methods. These methods rely solely on image-level supervision to identify part features. Early works usually locate class-specific discriminative regions via CAM [69]. Fu et al. proposes Recursive Attention CNN (RACNN) [14] to iteratively focus on predicting the locations of attention regions and extract discriminative features. Zheng et al. proposes Multi-Attention CNN (MACNN) [68], which proposes channel grouping loss to generate multiple parts by clustering. WSDAN [23] proposes weakly supervised attention learning to generate the distribution of parts. Meng et al. [39] introduces an end-to-end part discovery model (PDM) that aims to uncover robust and diverse object parts within a unified framework. Our PANet focuses on perceiving multiple discriminative regions in each view, extracting sufficient local view parts, and proposing the Adaptive Part Refinement (APR) module to further refine them to obtain global parts. These global parts can adapt to changes in viewpoint for parts across different views, making them more suitable for 3D objects. + +# 3 Proposed Method + +In this work, we newly propose a Part-Aware Network (PANet) that aims to solve the challenge of 3D object recognition (especially under arbitrary views). The overall architecture of the proposed PANet is shown in Fig. 3, consisting of three modules: weakly supervised part localization, adaptive part refinement, and joint part representation. We first introduce a weakly supervised method to perform part localization and feature extraction on each view (Sec. 3.1). Secondly, to adapt to the variations of part features across views, we propose the Adaptive Part Refinement (APR) module, which refines redundant part features from multiple views into more compact representations (Sec. 3.2). Finally, we integrate these refined features to represent a 3D object (Sec. 3.3). + +# 3.1 Weakly-Supervised Part Localization + +It is worth noting that we do not rely on supervised positional annotations [67], such as bounding boxes or keypoints, for part localization. Instead, we employ a more cost-efficient weakly-supervised learning approach [23, 68] that utilizes image-level annotations to identify discriminative regions, which requires no expertise and no extensive annotation effort. + +We generate a multi-view sequence $\mathcal{X} = \{\mathbf{X}_1, \mathbf{X}_2, \dots, \mathbf{X}_v\}$ by rendering a 3D object from different viewpoints. Here, $v$ denotes the number of views, $\mathbf{X}_i$ represents the $i$ -th view, where $i \in [1, v]$ . Then, each view is encoded by an image classification network e.g., ResNet [19] to obtain view features $\mathbf{I} = \{\mathbf{I}_1, \mathbf{I}_2, \dots, \mathbf{I}_v\}$ , where $\mathbf{I}_i \in \mathbb{R}^{H \times W \times C-1}$ , $i \in [1, v]$ is corresponding to the feature map of the $i$ -th + +![](images/e33d06aaec7b2ac8d1c4d8f0d9edab0e9654c7410868806a741831bf2bc8715a.jpg) +Fig. 3: Overview of our proposed Part-Aware Network (PANet). The network takes multi-view images as input, encodes them by CNN, obtains view features $\mathbf{I}$ , and uses the cross-view association (CVA) module to enhance the view features. Firstly, employ weakly-supervised methods to perceive part regions and generate view part sequences $\mathbf{T}$ . Then, given some learnable part tokens $\mathbf{P}$ , the sequences $\mathbf{T}$ and $\mathbf{P}$ are concatenated and input to the adaptive part refinement (APR) module. The module aims to refine the sequence $\mathbf{T}$ into a more compact representation, resulting in the global part features. Finally, multiple global parts $\underline{\mathbf{P}}$ together serve as a representation of a 3D object. + +view $\mathbf{X}_i$ . To align the features from multiple views into the same feature space, we share the weights among each of the convolutional neural network. + +Enhance Multi-view Features: However, the associations between view features $\mathbf{I}$ are not explicitly modeled. Therefore, we develop a cross-view association (CVA) module to obtain enhanced view features $\mathbf{F}_i$ , enabling each view to incorporate information from other views. + +For each view feature $\mathbf{I}_i$ , the cross-view association (CVA) module calculates the matching score between it and the overall view feature $\mathbf{I}$ . Here, the dot product function is selected as the matching function. Then, $\mathbf{I}_i$ is enhanced based on the correlation scores with other views. The enhanced view feature $\mathbf{F}_i \in \mathbb{R}^{\mathrm{H} \times \mathrm{W} \times \mathrm{C}}$ can be expressed in detail as follows: + +$$ +\mathbf {F} _ {i} = \operatorname {s o f t m a x} \left(\omega (\mathbf {I} _ {i}) ^ {T} \cdot \omega (\mathbf {I})\right) \mathbf {I}, \tag {1} +$$ + +where $\omega (\mathbf{I}_i)\in \mathbb{R}^{\mathrm{C}}$ represents the use of average pooling operations. + +Part Sampler: After passing through the cross-view association module, the enhanced view features $\mathbf{F}$ comprehensively incorporate information from other views. Next, we sample parts from the enhanced view features $\mathbf{F}$ . + +Inspired by [23, 45], we implement the attention model $\psi$ to learn the spatial distribution of each part within each view. For view $i$ , the distribution of + +view parts is represented by the attention map $\mathbf{A}_i\in \mathbb{R}^{\mathrm{H}\times \mathrm{W}\times \mathrm{M}}$ , which can be computed by: + +$$ +\mathbf {A} _ {i} = \psi (\mathbf {F} _ {i}) = \left\{\mathbf {A} _ {i, 1}, \mathbf {A} _ {i, 2}, \dots , \mathbf {A} _ {i, \mathrm {M}} \right\}, \tag {2} +$$ + +where $\mathbf{A}_{i,j} \in \mathbb{R}^{\mathrm{H} \times \mathrm{W}}$ , $i \in [1,v]$ , $j \in [1,\mathrm{M}]$ represents the $j$ -th part region or visual pattern of the object's $i$ -th views, such as a car wheel or an airplane wing. M denotes the number of attention maps, each view extracts the same number of regions of interest. + +Next, we employ the attention map $\mathbf{A}_{i,j}$ to softly weight the enhanced feature maps $\mathbf{F}_i$ , followed by the global average pooling operation $\omega$ to obtain the parts on view $i$ (See Eq. 3). + +$$ +\boldsymbol {t} _ {i, j} = \omega \left(\mathbf {F} _ {i} \otimes \mathbf {A} _ {i, j}\right), \tag {3} +$$ + +where $\pmb{t}_{i,j} \in \mathbb{R}^{\mathrm{C}}$ , $\otimes$ denotes element-wise multiplication between two tensors. + +Finally, we concatenate all the view part features together, resulting in a 3D multi-view part sequence $\mathbf{T} \in \mathbb{R}^{N \times C}$ (see Eq. 4). + +$$ +\mathbf {T} = \left[ \boldsymbol {t} _ {0, 0}, \boldsymbol {t} _ {0, 1}, \dots , \boldsymbol {t} _ {0, \mathrm {M}}, \boldsymbol {t} _ {1, 0}, \boldsymbol {t} _ {1, 1}, \dots , \boldsymbol {t} _ {1, \mathrm {M}}, \dots , \boldsymbol {t} _ {v, 0}, \boldsymbol {t} _ {v, 1}, \dots , \boldsymbol {t} _ {v, \mathrm {M}} \right], \tag {4} +$$ + +where $\mathrm{N}$ represents the total number of parts $(v\times \mathrm{M})$ in the multi-view. + +# 3.2 Adaptive Part Refinement + +In Sec. 3.1, we have obtained a multi-view part sequence $\mathbf{T}$ . However, there are differences in parts across different views, where each element $t_{ij}$ in $\mathbf{T}$ can only represent local view features. It is challenging to generalize the overall features of a 3D object relying solely on local view parts. Therefore, we newly propose an Adaptive Part Refinement (APR) module, which aims to explicitly model the correlations between parts across views, thus enhancing the robustness of part features to viewpoint variations. This module adaptively integrates part information from multiple views and refines it to obtain global part features that are more suitable for 3D objects. + +Formally, the APR (Adaptive Part Refinement) module contains a transformer block. Given some global part tokens [4,7,11] $\mathbf{P} \in \mathbb{R}^{\mathrm{L} \times \mathrm{C}}$ , where $\mathrm{L}$ represents the number of parts. We concatenate $\mathbf{P}$ and $\mathbf{T}$ as input to the APR module. After passing through the APR module, local parts $\pmb{t}_{i,j}$ can learn cross-view correlations and obtain global parts $\underline{\mathbf{P}} \in \mathbb{R}^{\mathrm{L} \times \mathrm{C}}$ suitable for 3D objects. We also observe the diversity of global part features $\underline{\mathbf{P}}$ (See Fig. 4). + +# 3.3 Joint Multi-Part Representation + +Relying solely on individual parts may not be sufficient to represent a 3D object. To overcome this limitation, we jointly describe the 3D object using multiple parts. We make predictions for each part and then average them to obtain the prediction for the 3D object. Each 3D object can be represented as Eq. 5: + +$$ +\underline {{\mathbf {P}}} = \left[ \underline {{\boldsymbol {p}}} _ {1}, \underline {{\boldsymbol {p}}} _ {2}, \dots , \underline {{\boldsymbol {p}}} _ {\mathrm {L}} \right], \quad \text {w h e r e} \underline {{\mathbf {P}}} \in \mathbb {R} ^ {\mathrm {L} \times \mathrm {C}} \tag {5} +$$ + +The proposed PANet is optimized through supervised learning, including instance classification loss $\mathcal{L}_{ce}$ and part-aware loss $\mathcal{L}_{awe}$ . The loss function for multi-view $\mathcal{X}$ is defined as follows: + +$$ +\begin{array}{l} p ^ {i} = \operatorname {s o f t m a x} \left(\underline {{\boldsymbol {p}}} _ {i} \boldsymbol {W} _ {p} + \boldsymbol {b} _ {p}\right), \quad \text {w h e r e} \underline {{\boldsymbol {p}}} _ {i} \in \mathbb {R} ^ {1 \times \mathrm {C}}, i \in [ 1, \mathrm {L} ] \\ q ^ {j} = \operatorname {s o f t m a x} \left(\boldsymbol {t} _ {j} \boldsymbol {W} _ {q} + \boldsymbol {b} _ {q}\right), \quad \text {w h e r e} \boldsymbol {t} _ {j} \in \mathbb {R} ^ {1 \times (\mathrm {M} \times \mathrm {C})}, j \in [ 1, v ], \tag {6} \\ \end{array} +$$ + +$$ +\mathcal {L} = \mathcal {L} _ {c e} + \mathcal {L} _ {a w e} = - \sum_ {k = 1} ^ {K} y _ {k} \log \hat {p} _ {k} + \gamma \left(- \frac {1}{v} \sum_ {i = 1} ^ {v} \sum_ {k = 1} ^ {K} y _ {k} \log q _ {k} ^ {j}\right), +$$ + +where $W_{p}, W_{q}, b_{p}$ , and $b_{q}$ are learnable parameters. $p^{i}$ represents the predicted probability of the $i$ -th global part, $\hat{p}_{k}$ represents the average of the predicted probabilities for L global parts, and $q^{j}$ represents the predicted probability for the $j$ -th view part. $y_{k}$ is the ground truth label of class $k$ . $\mathcal{L}_{ce}$ represents the cross-entropy loss of the final parts. $\mathcal{L}_{awe}$ represents the classification loss of M part features in each view, which is used to constrain the generation of parts. Furthermore, we apply techniques such as flipping and random erasure to augment the image for more robust part results. + +# 4 Experiments + +# 4.1 Datasets and Experimental Setup + +We conduct an extensive evaluation of PANet on several benchmark datasets, namely ScanObjectNN [53], ModelNet [59], and RGBD [29]. In order to facilitate fair comparisons, we strictly adhere to the methodologies proposed in prior works [26, 48, 55] for preparing the training and testing data on the different datasets. Specifically, we adopt the training scheme introduced by MVCNN [48] and adopt the same viewpoint generation and rendering method as RotationNet [26] to handle the multi-view images from the ModelNet and ScanObjectNN datasets. Experiment on rotated datasets also follows CVR [55] setting. For arbitrary views, the number of views for each 3D object ranges randomly from 10 to 20. Throughout the training process, we utilize the AdamW [36] optimizer with a learning rate of 1e-5. We select ResNet50 as our backbone network and utilize label smoothing regularization [51] to enhance the generalization capability of our model. $\gamma$ is the weighting factor for the part-aware loss as in Eq. 6, which we set $\gamma = 1$ for the experiments. Please refer to the Appendix for implementation details of our method. + +# 4.2 Comparison with State-of-the-Art Method + +The ScanObjectNN Datasets. As a recently introduced real-world 3D object classification dataset, ScanObjectNN [53] provides scanned indoor scene data. The dataset comprises approximately 15,000 objects, spanning 15 categories, and includes 2,902 unique object instances. Compared with CAD handcrafted + +Table 1: Classification on ScanObjectNN. 'Aligned' means that every 3D object is in the same pose. 'Rotated' means that each 3D object is misaligned. '-' indicates that the specific result is not reported. + +
MethodTypeAlignedRotated
Per Class Acc.Per Ins. Acc.Per Class Acc.Per Ins. Acc.
PointNet++ [43]Point82.10%84.30%--
DGCNN [54]84.00%86.20%--
MVCNN [48]85.71%87.82%78.21%80.62%
GVCNN [13]86.64%88.68%82.86%83.70%
RotationNet [26]View (fixed)84.88%86.90%74.68%76.16%
View-GCN [56]88.67%90.39%81.99%83.50%
CVR [55]88.39%90.74%84.70%85.59%
View-GCN++ [57]89.10%91.30%85.80%86.80%
Ours93.21%93.99%86.15%87.82%
CVR [55](arbitrary)--68.07%71.36%
Ours91.77%93.48%84.08%86.45%
+ +datasets, ScanObjectNN offers more realistic challenges as it reflects the noise, complexity, and incompleteness present in real-world environments, incorporating the appearance and geometric shapes of actual objects. Consequently, it is better for assessing the robustness of the model. We train PANet under different viewpoint settings and compare with point-based methods [43, 54] and view-based methods [13, 26, 48, 55-57], as shown in Tab. 1. View-based methods primarily rely on view aggregation, and it is difficult to robustly aggregate features from structurally misaligned views, especially for arbitrary views. When switching to arbitrary views, the performance of these methods drops sharply. Our PANet is observed to significantly outperform previous approaches in both class accuracy and instance accuracy, whether in aligned views, rotated views, or arbitrary views. In aligned views, compared with the current state-of-the-art methods e.g., View-GCN++ [57] and CVR [55], PANet improves instance accuracy and class accuracy by $2.69\%$ and $4.11\%$ . PANet also achieves significant improvements of $1.02\%$ and $0.35\%$ in rotated views. Surprisingly, even under arbitrary views, our performance surpasses that of most fixed viewpoint methods, demonstrating that PANet can robustly recognize 3D objects unconstrained by the scene. Please refer to the Appendix for more detailed comparisons. + +The ModelNet Datasets. ModelNet [59] is a dataset of hand-drawn point clouds generated from CAD models. It comprises 12,311 3D shapes, covering 40 categories, with 9,483 models allocated for training and 2,468 for testing. The experimental results on the ModelNet dataset are presented in Tab. 2. We first compare with point-based and voxel-based methods, including PointNet++ [43], Kd-Networks [27], RS-CNN [22], KP-Convs [52], MVTN [34], 3DShapeNets [58], VoxNet [38], and VRENsemble [3]. Our PANet outperforms them by more than $3.93\%$ and $4.6\%$ in instance accuracy and class accuracy and is also $3.73\%$ and $4.37\%$ higher in free viewpoints. We also compared with view-based approaches, including MVCNN-new [48], GVCNN [13], HEAR [5], RotationNet [26], View-GCN [56], MVT [6], HMVCM [32], MVD [24], DAN [41], DRCNN [50], EMV [12], View-GCN++ [57], CVR [55], and HGNN+ [16]. PANet consistently outper + +Table 2: Classification on Aligned ModelNet40. '- indicates that the specific result is not reported, 'fixed' denotes sampling from a fixed viewpoint. 'free' means that the arbitrary number of viewpoints. + +
Aligned ModelNet40
MethodTypeClass Acc.Ins. Acc.
3DShapeNets [58]77.3%-
VoxNet [38]Voxel83.0%-
VRENsemble [3]91.4%93.8%
PointNet++ [43]-91.9%
Kd-Networks [27]88.5%91.8%
RS-CNN [22]Point-93.6%
KP-Convs [52]-92.9%
MVTN [34]92.0%93.8%
MVCNN-new [48]92.40%95.00%
GVCNN [13]90.70%93.10%
HEAR [5]95.20%96.70%
RotationNet [26]-97.37%
View-GCN [56]96.50%97.60%
MVT [6]-97.50%
HMVCM [32]-94.57%
MVD [24]View92.61%94.41%
DAN [41]-93.50%
DRCNN [50]94.86%96.84%
View-GCN++ [57]96.50%97.60%
CVR [55]95.77%97.16%
HGNN+ [16]-96.92%
Ours96.60%97.73%
Ours(free)96.37%97.53%
+ +Table 3: The accuracy of 3D object classification on Rotated ModelNet40. + +
Rotated ModelNet40
Method#ViewClass Acc.Ins. Acc.
MVCNN [48]87.95%88.17%
GVCNN [13]89.69%88.10%
EMV [12]88.80%90.00%
RotationNet [26]fixed84.74%85.29%
View-GCN [56]85.90%88.25%
CVR [55]91.12%92.22%
View-GCN++ [57]90.70%92.80%
Ours90.35%92.87%
CVR [55]arbitrary84.01%86.91%
Ours90.69%92.23%
+ +Table 4: The accuracy of object classification on RGBD (in%). + +
Method#ViewIns. Acc.
MDSICNN [2]≥ 12089.6%
CFK [8]≥ 12086.8%
MMDCNN [44]≥ 12086.8%
RotationNet [26]1289.3%
View-GCN [56]1294.3%
View-GCN++ [57]1292.5%
Ours1296.5%
MVCNN [48]89.0%
GVCNN [13]89.8%
CVR [55]4-1291.8%
Ours95.9%
+ +forms these methods in fixed views. Its performance at free viewpoints surpasses that of most existing approaches. We next report the robustness of PANet in rotated views, as shown in Tab. 3. Under the rotated views, PANet also achieves the highest score in instance accuracy. + +The RGBD Datasets. The RGBD [29] dataset comprises 51 categories of domestic objects captured in the real world, encompassing RGBD views from multiple viewpoints. To ensure a fair comparison, we adopt the identical experimental setup as RotationNet [26] and View-GCN [56]. Specifically, we uniformly sample 12 RGB images from the same circle for each 3D object as input and report the average results using ten-fold cross-validation. We keep the same settings as CVR [55] in arbitrary views, sampling images from 4 to 12 viewpoints as input to the network. The experimental results are summarized in Tab. 4. The performance of PANet in arbitrary views is impressive. Compared with work in arbitrary views e.g., CVR [55], PANet significantly improves by $4.1\%$ . Additionally, compared with the work in fixed views, such as view-GCN [56], PANet has also improved by $1.6\%$ . This further validates the superiority of our method. + +# 4.3 Model Analysis + +All experiments are conducted on the ScanObjectNN dataset in rotated views. Please refer to the Appendix for detailed experimental analyses. + +![](images/c8ab5804a3db7cc39f0d2e0fe951b9c25564269d8d80930d3fd41747820e7a72.jpg) +Fig.4: Visualization of correlations between part features $\underline{\mathbf{P}}$ . The left figure illustrates the correlation between $\underline{\mathbf{P}}$ after applying $\mathcal{L}_{awe}$ . + +![](images/d7abb86743ead2a22ae1876985ac2207d96a77cebac32c138c597ee86f87b1d3.jpg) + +Table 5: PANet component ablations. The first line serves as the baseline. $\mathcal{L}_{awe}$ refers to the weakly supervised part-aware loss function. CVA stands for Cross-View Association module. + +
CVALaweClass Acc.Ins. Acc.
✓83.76%84.56%
84.99%85.25%
✓85.45%86.96%
+ +Table 6: Explore the impact of different viewpoint sampling on performance. + +
DatasetMethodAlignedRotated
Class Acc.Ins. Acc.Class Acc.Ins. Acc.
ModelNet40random sampling96.37%97.53%90.69%92.23%
furthest point sampling95.44%97.41%89.73%92.26%
ScanObjectNNrandom sampling91.77%93.48%84.08%86.45%
furthest point sampling92.97%93.48%84.80%85.93%
+ +PANet component ablations We evaluate the role of the cross-view association module and $\mathcal{L}_{awe}$ , as shown in Tab. 5. The CVA module plays a positive role in enhancing the performance of PANet, improving instance accuracy and class accuracy by $0.69\%$ and $1.23\%$ . The results show that the CVA module enables each view to more effectively integrate information from other views, further enhancing the representation capability of the view features. Our method also improves by $2.40\%$ and $1.76\%$ in two metrics by introducing the $\mathcal{L}_{awe}$ constraint. In addition, we also conduct correlation analysis between part features $\underline{\mathbf{P}}$ , as shown in Fig. 4. We observe that without applying $\mathcal{L}_{awe}$ constraints, the correlation between parts exhibits a flat similarity matrix, indicating that the parts are unvaried and cannot be distinguished in the feature space. In contrast, the part features become more diverse with $\mathcal{L}_{awe}$ constraint. Therefore, we conclude that $\mathcal{L}_{awe}$ can supervise the perception of parts to obtain diverse part features. + +Influence of Different Viewpoint Sampling We introduce two different sampling methods: random sampling and furthest point sampling (FPS). Random sampling refers to the process of randomly selecting 10 to 20 views from each object. This method is closer to real-world scenarios, and we showcased PANet's performance in Tab. 1, 2, 3, 4 based on this sampling method. FPS selects 10 to 20 views based on the furthest point sampling principle. It provides a uniform sampling method that covers every orientation of the 3D objects. As shown in Tab. 6, we observed that the performance of random sampling and FPS across different datasets is nearly similar. This result shows that PANet is not sensitive to the choice of sampling methods. The model demonstrates robustness in adapting to changes in viewpoint caused by different sampling methods while maintaining stable performance. + +Influence of the Number of the Views We further evaluate PANet by considering different backbone networks and numbers of views, as shown in Tab. 7. + +Table 7: Explore the impact of number of views and backbones on performance. + +
BackboneViews
1 views5 views10 views15 views20 views
resnet1845.63%70.84%75.66%79.90%82.85%
resnet5067.07%79.59%82.50%85.93%86.62%
+ +Table 8: Exploring the effect of the number of attention maps, M denotes the number of attention maps per view. + +
#Attention MapsClass Acc.Ins. Acc.
M = 1684.57%86.28%
M = 3285.36%87.14%
M = 6486.50%87.31%
M = 12886.38%87.14%
+ +Table 9: Exploring the effect of the number of part tokens, L indicates the number of parts. + +
#Number of PartsClass Acc.Ins. Acc.
L = 185.50%86.28%
L = 886.20%87.14%
L = 1686.50%87.31%
L = 3285.63%86.31%
+ +In previous studies, such as View-GCN [56], because view aggregation leads to feature loss. Choosing a stronger backbone network may not necessarily lead to performance improvement in certain dataset. However, our observations reveal that the choice of backbone significantly affects PANet performance. This is because PANet directly integrates multiple parts to characterize 3D objects. Therefore, a stronger backbone network is more advantageous for perceiving part features. Furthermore, we notice a positive correlation between the number of views and the performance of PANet. It is worth noting that previous studies on multi-view aggregation methods, such as MHBN [65], MVLADN [64], and GVCNN [13], have indicated that increasing the number of views does not necessarily lead to better representation of 3D objects. PANet combines multiple parts for representation, and the APR module can also filter out some redundant parts. This means that PANet is not affected by redundant views. Our experimental results also verify this idea. + +Influence of the Number of the Attention Maps and Parts We show the impact of the number of attention maps M and the number of part tokens L, as shown in Tab. 8 and Tab. 9. The number of attention maps M determines the discriminative region in each view, and the part tokens L determine the number of global parts. More M and L usually contribute to better performance. MACNN [68] and WSDAN [23] report similar conclusions. The experiments in Tab. 8 show that as the value of M increases, the classification accuracy gradually improves. When $\mathrm{M} = 64$ , the performance tends to stabilize. The results in Tab. 9 show that the performance improves as the number of parts increases and reaches a peak at $\mathrm{L} = 16$ . Compared with a single part, multiple parts provide richer information and enhance the generalization of the model. + +# 4.4 Visualization + +We randomly select some 3D models from a test set and compare the semantic differences between the features extracted by MVCNN and PANet, as shown in Fig. 5. MVCNN [48] is a classic feature aggregation method. We notice that the + +features extracted by MVCNN are not entirely focused on the object. This is because the uneven weight distribution leads to view aggregation methods ignoring certain view information. It is worth noting that PANet does not discard any views but instead considers the features of each view, leveraging the valuable information within each view. We can observe that different views emphasize various part locations, with each view capturing multiple discriminative regions (e.g., the airplane nose, body, and wings). This observation further validates the diversity of captured parts. Furthermore, we observe that the view aggregation method does not explicitly model the unique attributes of 3D objects, leading to feature ambiguity. In contrast, part-based representation reveals the inherent properties of objects, enhancing the interpretability of features and making recognition easier. Please refer to the Appendix for more visualization results. + +# 5 Conclusion + +In this paper, we propose a novel part-aware network (PANet), which is a part-based representation. Compared with view aggregation methods, this part-based representation has properties such as viewpoint invariance and rotation robustness, which can more effectively address the 3D object recognition problem under arbitrary views. Moreover, we proposed the part-aware loss to ensure that part features can better focus on the object, thereby fully utilizing the information from each view. We evaluate the performance of PANet on several 3D datasets, including ScanObjectNN, ModelNet, and RGBD. The experimental results show that PANet has made significant progress, especially in arbitrary views. We also conduct ablation and visualization experiments to study the impact of different components in our method. It is worth noting that in these experiments, we found that our method is not affected by redundant views (Sec. 4.3), and the part features can better reveal the inherent properties of objects (Sec. 4.4). In conclusion, the experimental results prove the effectiveness of our method. + +Acknowledgement This research is supported by the Shenzhen Science and Technology Plan Project No. KJZD20230923113800002 and Shenzhen LongHua Fundamental Research Program No. 10162A20230325B73A546. + +![](images/9f3fe613b8e83a64611c25212ff7ed83aafbdc9c41e62438e38e43e91867e945.jpg) +Fig. 5: Comparing the semantic differences between view-based aggregation methods and PANet in feature extraction. + +![](images/bf8ea68dca1f99451d4bde25a8a38882d323cd4d796afa8e2fa1d31136628ff5.jpg) + +# References + +1. Afham, M., Dissanayake, I., Dissanayake, D., Dharmasiri, A., Thilakarathna, K., Rodrigo, R.: Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9902-9912 (2022) +2. Asif, U., Bennamoun, M., Sohel, F.A.: A multi-modal, discriminative and spatially invariant cnn for rgb-d object labeling. IEEE transactions on pattern analysis and machine intelligence 40(9), 2051-2065 (2017) +3. Brock, A., Lim, T., Ritchie, J.M., Weston, N.: Generative and discriminative voxel modeling with convolutional neural networks. arXiv preprint arXiv:1608.04236 (2016) +4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) +5. Chen, J., Qin, J., Shen, Y., Liu, L., Zhu, F., Shao, L.: Learning attentive and hierarchical representations for 3d shape recognition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 105-122. Springer (2020) +6. Chen, S., Yu, T., Li, P.: Mvt: Multi-view vision transformer for 3d object recognition. arXiv preprint arXiv:2110.13083 (2021) +7. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021) +8. Cheng, Y., Cai, R., Zhao, X., Huang, K.: Convolutional fisher kernels for rgb-d object recognition. In: 2015 International Conference on 3D Vision. pp. 135-143. IEEE (2015) +9. Delitzas, A., Parelli, M., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329 (2023) +0. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +2. Esteves, C., Xu, Y., Allen-Blanchette, C., Daniilidis, K.: Equivariant multi-view networks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1568-1577 (2019) +3. Feng, Y., Zhang, Z., Zhao, X., Ji, R., Gao, Y.: GvCNN: Group-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 264-272 (2018) +4. Fu, J., Zheng, H., Mei, T.: Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4438-4446 (2017) +5. Gao, K., Gao, Y., He, H., Lu, D., Xu, L., Li, J.: Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379 (2022) + +16. Gao, Y., Feng, Y., Ji, S., Ji, R.: Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3181-3199 (2022) +17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020) +18. Han, Z., Lu, H., Liu, Z., Vong, C.M., Liu, Y.S., Zwicker, M., Han, J., Chen, C.P.: 3d2seqviews: Aggregating sequential views for 3d global feature learning by cnn with hierarchical attention aggregation. IEEE Transactions on Image Processing 28(8), 3986-3999 (2019) +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +20. Herreras, E.B.: Cognitive neuroscience; the biology of the mind. Cuadernos de Neuropsicología/Panamerican Journal of Neuropsychology 4(1), 87-90 (2010) +21. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9202-9212 (2023) +22. Hu, L., Qin, M., Zhang, F., Du, Z., Liu, R.: Rscnn: a cnn-based method to enhance low-light remote-sensing images. Remote Sensing 13(1), 62 (2020) +23. Hu, T., Qi, H., Huang, Q., Lu, Y.: See better before looking closer: Weakly supervised data augmentation network for fine-grained visual classification. arXiv preprint arXiv:1901.09891 (2019) +24. Huang, J., Yan, W., Li, T., Liu, S., Li, G.: Learning the global descriptor for 3-d object recognition based on multiple views decomposition. IEEE Transactions on Multimedia 24, 188-201 (2022). https://doi.org/10.1109/TMM.2020.3047762 +25. Huang, S., Xu, Z., Tao, D., Zhang, Y.: Part-stacked cnn for fine-grained visual categorization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1173-1182 (2016) +26. Kanezaki, A., Matsushita, Y., Nishida, Y.: Rotationnet for joint object categorization and unsupervised pose estimation from multi-view images. IEEE transactions on pattern analysis and machine intelligence 43(1), 269-283 (2019) +27. Klokov, R., Lempitsky, V.: Escape from cells: Deep kd-networks for the recognition of 3d point cloud models. In: Proceedings of the IEEE international conference on computer vision. pp. 863-872 (2017) +28. Kumawat, S., Raman, S.: Lp-3dcnn: Unveiling local phase in 3d convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4903-4912 (2019) +29. Lai, K., Bo, L., Ren, X., Fox, D.: A large-scale hierarchical multi-view rgb-d object dataset. In: 2011 IEEE international conference on robotics and automation. pp. 1817-1824. IEEE (2011) +30. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022) +31. Lin, Y., Gou, Y., Liu, X., Bai, J., Lv, J., Peng, X.: Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4447-4461 (2022) +32. Liu, A.A., Zhou, H., Nie, W., Liu, Z., Liu, W., Xie, H., Mao, Z., Li, X., Song, D.: Hierarchical multi-view context modelling for 3d object classification and retrieval. Information Sciences 547, 984-995 (2021) + +33. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023) +34. Liu, S., Nguyen, V., Rehg, I., Tu, Z.: Recognizing objects from any view with object and viewer-centered representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2020) +35. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing 30, 1744-1758 (2021) +36. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +37. Lowe, D.: Object recognition from local scale-invariant features. In: Proceedings of the Seventh IEEE International Conference on Computer Vision. vol. 2, pp. 1150-1157 vol.2 (1999). https://doi.org/10.1109/ICCV.1999.790410 +38. Maturana, D., Scherer, S.: Voxnet: A 3d convolutional neural network for real-time object recognition. In: 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 922-928. IEEE (2015) +39. Meng, M., Zhang, T., Yang, W., Zhao, J., Zhang, Y., Wu, F.: Diverse complementary part mining for weakly supervised object localization. IEEE Transactions on Image Processing 31, 1774-1788 (2022). https://doi.org/10.1109/TIP.2022.3145238 +40. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +41. Nie, W., Zhao, Y., Song, D., Gao, Y.: Dan: Deep-attention network for 3d shape recognition. IEEE Transactions on Image Processing 30, 4371-4383 (2021) +42. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017) +43. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +44. Rahman, M.M., Tan, Y., Xue, J., Lu, K.: Rgb-d object recognition with multimodal deep convolutional neural networks. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 991-996. IEEE (2017) +45. Rao, Y., Chen, G., Lu, J., Zhou, J.: Counterfactual attention learning for fine-grained visual categorization and re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1025–1034 (2021) +46. Riegler, G., Osman Ulusoy, A., Geiger, A.: Octnet: Learning deep 3d representations at high resolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3577-3586 (2017) +47. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +48. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015) +49. Su, J.C., Gadelha, M., Wang, R., Maji, S.: A deeper look at 3d shape classifiers. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018) + +50. Sun, K., Zhang, J., Liu, J., Yu, R., Song, Z.: Drcnn: Dynamic routing convolutional neural network for multi-view 3d object recognition. IEEE Transactions on Image Processing 30, 868-877 (2020) +51. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016) +52. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019) +53. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1588-1597 (2019) +54. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog) 38(5), 1-12 (2019) +55. Wei, X., Gong, Y., Wang, F., Sun, X., Sun, J.: Learning canonical view representation for 3d shape recognition with arbitrary views. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 407-416 (2021) +56. Wei, X., Yu, R., Sun, J.: View-gcn: View-based graph convolutional network for 3d shape analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1850-1859 (2020) +57. Wei, X., Yu, R., Sun, J.: Learning view-based graph convolutional network for multi-view 3d shape analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +58. Wu, S., Khosla, Y., Zhang, T.: D shapenets: A deep representation for volumetric shape modeling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA. pp. 7-12 (2015) +59. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015) +60. Xu, Y., Zheng, C., Xu, R., Quan, Y., Ling, H.: Multi-view 3d shape recognition via correspondence-aware deep learning. IEEE Transactions on Image Processing 30, 5299-5312 (2021) +61. Xue, L., Gao, M., Xing, C., Martin-Martin, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning unified representation of language, image and point cloud for 3d understanding. arXiv preprint arXiv:2212.05171 (2022) +62. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023) +63. Yang, Z., Wang, L.: Learning relationships for multi-view 3d object recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7505-7514 (2019) +64. Yu, T., Meng, J., Yang, M., Yuan, J.: 3d object representation learning: A set-to-set matching perspective. IEEE Transactions on Image Processing 30, 2168-2179 (2021). https://doi.org/10.1109/TIP.2021.3049968 +65. Yu, T., Meng, J., Yuan, J.: Multi-view harmonized bilinear network for 3d object recognition. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 186-194 (2018). https://doi.org/10.1109/CVPR.2018.00027 + +66. Zhang, H., Xu, T., Elhoseiny, M., Huang, X., Zhang, S., Elgammal, A., Metaxas, D.: Spda-cnn: Unifying semantic part detection and abstraction for fine-grained recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1143-1152 (2016) +67. Zhang, N., Donahue, J., Girshick, R., Darrell, T.: Part-based r-cnns for fine-grained category detection. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 834-849. Springer (2014) +68. Zheng, H., Fu, J., Mei, T., Luo, J.: Learning multi-attention convolutional neural network for fine-grained image recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 5209-5217 (2017) +69. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016) \ No newline at end of file diff --git a/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/images.zip b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c7718de1b6e83fd9aa1992d54063a4d112a3244 --- /dev/null +++ b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c404ba82868a4a319b18757d98f00c609bdd3d8b1f60e7c1d14ccc847427659 +size 607800 diff --git a/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/layout.json b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce9a105da8bc41624ccf729a81a13440b98657a5 --- /dev/null +++ b/2024/Beyond Viewpoint_ Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation/layout.json @@ -0,0 +1,9586 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 479, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 479, + 164 + ], + "type": "text", + "content": "Beyond Viewpoint: Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "spans": [ + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "text", + "content": "Linlong Fan" + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{\\text{日}}" + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "text", + "content": ", Ye Huang" + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{\\text{日}^{\\star}}" + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "text", + "content": ", Yanqi Ge" + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{\\text{日}}" + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "text", + "content": ", Wen Li" + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{\\text{日}}" + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "text", + "content": ", and Lixin Duan" + }, + { + "bbox": [ + 155, + 186, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{\\text{日}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 189, + 208, + 424, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 208, + 424, + 242 + ], + "spans": [ + { + "bbox": [ + 189, + 208, + 424, + 242 + ], + "type": "text", + "content": "Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China fanlinlong703@163.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 266, + 452, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 266, + 452, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 266, + 452, + 475 + ], + "type": "text", + "content": "Abstract. Existing view-based methods excel at recognizing 3D objects from predefined viewpoints, but their exploration of recognition under arbitrary views is limited. This is a challenging and realistic setting because each object has different viewpoint positions and quantities, and their poses are not aligned. However, most view-based methods, which aggregate multiple view features to obtain a global feature representation, hard to address 3D object recognition under arbitrary views. Due to the unaligned inputs from arbitrary views, it is challenging to robustly aggregate features, leading to performance degradation. In this paper, we introduce a novel Part-aware Network (PANet), which is a part-based representation, to address these issues. This part-based representation aims to localize and understand different parts of 3D objects, such as airplane wings and tails. It has properties such as viewpoint invariance and rotation robustness, which give it an advantage in addressing the 3D object recognition problem under arbitrary views. Our results on benchmark datasets clearly demonstrate that our proposed method outperforms existing view-based aggregation baselines for the task of 3D object recognition under arbitrary views, even surpassing most fixed viewpoint methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 486, + 421, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 486, + 421, + 497 + ], + "spans": [ + { + "bbox": [ + 160, + 486, + 421, + 497 + ], + "type": "text", + "content": "Keywords: 3D object recognition " + }, + { + "bbox": [ + 160, + 486, + 421, + 497 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 486, + 421, + 497 + ], + "type": "text", + "content": " weakly-supervised learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 516, + 230, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 230, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 230, + 529 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 540, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 647 + ], + "type": "text", + "content": "Learning 3D with multi-view supervision has received widespread attention. It covers various research domains such as NeRFs [15,40], BEV detection [30,62], 3D generation [33], 3D reasoning [9,21], etc., among which 3D object recognition is considered one of the most fundamental problems. Thanks to mature and transferable 2D vision models [11,47] and large-scale image datasets [10], rendering 3D objects into multiple 2D images and applying 2D networks to indirectly recognize 3D objects has become the state-of-the-art method [57]. Furthermore, compared with direct methods (e.g., point-based [1,61], voxel-based [38]), view-based methods are closer to human perception of 3D objects in the real world." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 480, + 186 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 186 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 480, + 186 + ], + "type": "image", + "image_path": "b3e50804339c196a8b81f296bd240f8cc9be1db52975e43fe2dde2591e373aff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 193, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 270 + ], + "type": "text", + "content": "Fig. 1: In practical scenarios, 3D objects are observed from arbitrary views(right). The definition of arbitrary views is that each object is unaligned, and the viewpoint positions and quantities vary for each object. Previous works have focused on 3D object recognition in aligned views(left), where the pose of each object is aligned, and the positions and quantities of viewpoints are predefined. Rotated views(middle) is an extension that introduces random rotation for each 3D object while keeping the viewpoint positions and quantities unchanged. Our work focuses on arbitrary views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "text", + "content": "Previous view-based methods [5,6,49,50,56] focus on addressing 3D object recognition tasks under fixed viewpoints, while recognition under arbitrary views still lacks extensive research. As shown in Fig. 1, arbitrary views represent a more challenging setting where each object may have different viewpoint positions and quantities, and the poses of each object are not aligned." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "text", + "content": "However, most view-based methods [5,6,13,18,26,41,49,50,55-57,60,63-65], which aggregate multiple view features to obtain a global feature representation, hard to effectively address 3D object recognition under arbitrary views. This is because the misaligned nature of arbitrary views complicates the representation and aggregation of view features, which leads to performance degradation. For example, when applied to arbitrary views, the performance of CVR [55] drops by " + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "inline_equation", + "content": "6.31\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "inline_equation", + "content": "14.23\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "text", + "content": " in ModelNet [59] and ScanObjectNN [53] datasets. Even some multi-view aggregation methods [26, 56] cannot be applied to arbitrary views because they require fixed viewpoint position information. Furthermore, view aggregation methods [13, 49] typically jointly optimize feature extraction and aggregation modules to maximize performance, but this may lead the extraction module to overly focus on certain views while ignoring information from others, resulting in information loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "To address these challenges, we propose a novel part-based representation, as shown in Fig. 2. This idea is inspired by cognitive science and biology [20], which aims to understand different part features or specific regions in a multi-view sequence to better represent 3D objects. For example, a person can be composed of multiple parts with unique attributes, such as the head, arms, body, legs, and so on. By locating and integrating the features of these parts, we can more accurately represent a person. Compared with the multi-view aggregation features, part features offer several advantages. Firstly, an individual view contains abundant part information. Even if only a few views are utilized, we can still extract stable part features, thereby adapting to changes in the number of viewpoints in arbitrary view settings. Secondly, the part features are more stable than the multi-view aggregation features [37], which can better adapt to changes in object rotation and viewpoint positions. Therefore, the part-based representation method is robust and can better handle the challenge of 3D object recogni" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 441, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 441, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 441, + 102 + ], + "type": "text", + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 479, + 264 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 479, + 264 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 479, + 264 + ], + "type": "image", + "image_path": "a987c554f2db6f2aec74bd725338d32c96de92fb3f82578c570f2d509551cf4f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 360 + ], + "type": "text", + "content": "Fig. 2: Comparison between multi-view aggregation representation (a) and part-based representation (b). View aggregation methods integrate features from different views and use the aggregated features to represent 3D objects. However, since the views are not aligned, the aggregated features lack robustness, making it difficult to handle 3D recognition on arbitrary views. Additionally, view aggregation methods suffer from information loss, thus unable to fully utilize the features from each view. We newly propose a part-based representation that focuses on multi-view part awareness and combines multiple parts to robustly represent 3D objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 370, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 480, + 418 + ], + "type": "text", + "content": "tion under arbitrary views. We also provide comprehensive experimental results in Sec. 4 to demonstrate the effectiveness of part-based representation. When switching to arbitrary views, we do not observe significant performance degradation in our method, and it even outperforms most fixed viewpoint methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 419, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 528 + ], + "type": "text", + "content": "Specifically, this paper proposes a novel part-aware network (PANet). Firstly, it employs weakly supervised learning to locate and recognize parts in each view, obtaining local view part features. Then, local view parts are input into the adaptive part refinement (APR) module. This module adaptively perceives the correlation of local view parts and further integrates these view parts to obtain global parts, which aims to adapt to the viewpoint changes of parts across different views. Finally, we combine these global parts to represent a 3D object. It is worth noting that we also introduce a regularization loss in the part localization module to ensure that part features can better focus on the object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 529, + 455, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 529, + 455, + 541 + ], + "spans": [ + { + "bbox": [ + 146, + 529, + 455, + 541 + ], + "type": "text", + "content": "The primary contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": "- Unlike existing view aggregation methods, we propose a novel part-based representation. We consider a 3D object as a collection of part features rather than an abstract multi-view aggregated feature. This part-based representation is not limited by the position and quantity of viewpoints or affected by object poses, making it more flexibly applicable to various viewpoint scenarios, especially in arbitrary views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "- Further, we newly develop a Part-Aware Network (PANet), for extracting parts in multiple views. The network employs a weakly supervised method to localize parts on each view, enabling it to obtain local view parts. Then, we further refines these view parts through the proposed Adaptive Part Refinement" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Viewpoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "(APR) module to obtain global parts. These global parts can better adapt to differences in parts across viewpoints. Finally, integrating these global parts forms a comprehensive description of a 3D object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 213 + ], + "type": "text", + "content": "- We conduct experiments on various datasets, including ScanObjectNN, ModelNet, and RGBD. Our method outperforms existing methods in both fixed views and arbitrary views. In particular, the performance in arbitrary views exceeds most existing fixed viewpoint methods based on fixed views, achieving robust recognition of 3D objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 242, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 242, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 242, + 244 + ], + "type": "text", + "content": "2 Relative Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 479, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 341 + ], + "type": "text", + "content": "3D object Recognition methods: 3D object recognition using deep neural networks can be broadly divided into three main categories: point-based methods [17,42,43], voxel-based methods [28,38,46], and view-based methods. View-based methods render 3D objects into multiple views and utilize multiple views to jointly represent 3D objects. Benefiting from mature and transferable 2D vision models and large-scale image datasets, they can better handle 3D object recognition tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": "MVCNN [48] is a pioneering multi-view network for 3D object classification. It utilizes a view pooling strategy to integrate features from different views. Yang et al. [63] proposes a relational network to effectively connect corresponding regions from different viewpoints, thus enhancing the information of each view. 3D2Seq2views [18] method treats multiple views as a sequence and employs a convolutional neural network with novel hierarchical attention aggregation to aggregate sequential views. DRCNN [50] utilizes the dynamic routing algorithm of capsule networks to construct a new Dynamic Routing Layer (DRL), which integrates features from each view. View-GCN [56] employs multiple views as nodes in a view graph, leveraging a graph convolutional network (GCN) architecture to hierarchically aggregate information and learn robust 3D shape descriptors. These methods assume that each object has a predefined set of viewpoints and are trained and tested on these viewpoints. This assumption may limit model performance when dealing with 3D object recognition in arbitrary views. As far as we know, there are few studies [31, 34, 55] that have made significant progress beyond fixed viewpoint settings. The MVTN [34] implements viewpoint regression for 3D object recognition based on the latest research progress of differentiable rendering. The motivation of CVR [55] is to restore the intrinsic alignment of arbitrary views, transform it into aligned representations and then perform aggregation. In our work, we newly propose part-based representation. Part features exhibit viewpoint invariance and rotation robustness, making them more suitable for handling changes in viewpoint positions and quantities in arbitrary views. Moreover, the key idea of part-based representation is not to rely on multi-view aggregation, thus avoiding information loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Discriminative Part Localization: The perception and localization of object part-level features are usually challenging. Many existing approaches [25,35, 66, 67] suggest using bounding boxes and additional annotations to accurately" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 441, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 441, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 441, + 102 + ], + "type": "text", + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": "locate discriminative regions of objects. However, this method often requires expert knowledge and extensive annotation efforts, rendering it impractical for real-world applications. Fortunately, more and more research [14, 23, 45, 68, 69] has shown that important object regions can also be localized using weakly supervised methods. These methods rely solely on image-level supervision to identify part features. Early works usually locate class-specific discriminative regions via CAM [69]. Fu et al. proposes Recursive Attention CNN (RACNN) [14] to iteratively focus on predicting the locations of attention regions and extract discriminative features. Zheng et al. proposes Multi-Attention CNN (MACNN) [68], which proposes channel grouping loss to generate multiple parts by clustering. WSDAN [23] proposes weakly supervised attention learning to generate the distribution of parts. Meng et al. [39] introduces an end-to-end part discovery model (PDM) that aims to uncover robust and diverse object parts within a unified framework. Our PANet focuses on perceiving multiple discriminative regions in each view, extracting sufficient local view parts, and proposing the Adaptive Part Refinement (APR) module to further refine them to obtain global parts. These global parts can adapt to changes in viewpoint for parts across different views, making them more suitable for 3D objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 346, + 261, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 261, + 360 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 261, + 360 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": "In this work, we newly propose a Part-Aware Network (PANet) that aims to solve the challenge of 3D object recognition (especially under arbitrary views). The overall architecture of the proposed PANet is shown in Fig. 3, consisting of three modules: weakly supervised part localization, adaptive part refinement, and joint part representation. We first introduce a weakly supervised method to perform part localization and feature extraction on each view (Sec. 3.1). Secondly, to adapt to the variations of part features across views, we propose the Adaptive Part Refinement (APR) module, which refines redundant part features from multiple views into more compact representations (Sec. 3.2). Finally, we integrate these refined features to represent a 3D object (Sec. 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 501, + 348, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 348, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 348, + 514 + ], + "type": "text", + "content": "3.1 Weakly-Supervised Part Localization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 517, + 481, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 481, + 576 + ], + "type": "text", + "content": "It is worth noting that we do not rely on supervised positional annotations [67], such as bounding boxes or keypoints, for part localization. Instead, we employ a more cost-efficient weakly-supervised learning approach [23, 68] that utilizes image-level annotations to identify discriminative regions, which requires no expertise and no extensive annotation effort." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": "We generate a multi-view sequence " + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathbf{X}_1, \\mathbf{X}_2, \\dots, \\mathbf{X}_v\\}" + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": " by rendering a 3D object from different viewpoints. Here, " + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": " denotes the number of views, " + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": "-th view, where " + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "inline_equation", + "content": "i \\in [1, v]" + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": ". Then, each view is encoded by an image classification network e.g., ResNet [19] to obtain view features " + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{\\mathbf{I}_1, \\mathbf{I}_2, \\dots, \\mathbf{I}_v\\}" + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i \\in \\mathbb{R}^{H \\times W \\times C-1}" + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "inline_equation", + "content": "i \\in [1, v]" + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": " is corresponding to the feature map of the " + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 577, + 482, + 638 + ], + "type": "text", + "content": "-th" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Viewpoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "1 H, W, and C represent the height, width, and number of channels of the feature layer, respectively" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 479, + 331 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 479, + 331 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 479, + 331 + ], + "type": "image", + "image_path": "e33d06aaec7b2ac8d1c4d8f0d9edab0e9654c7410868806a741831bf2bc8715a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "text", + "content": "Fig. 3: Overview of our proposed Part-Aware Network (PANet). The network takes multi-view images as input, encodes them by CNN, obtains view features " + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "text", + "content": ", and uses the cross-view association (CVA) module to enhance the view features. Firstly, employ weakly-supervised methods to perceive part regions and generate view part sequences " + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "text", + "content": ". Then, given some learnable part tokens " + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "text", + "content": ", the sequences " + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "text", + "content": " are concatenated and input to the adaptive part refinement (APR) module. The module aims to refine the sequence " + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "text", + "content": " into a more compact representation, resulting in the global part features. Finally, multiple global parts " + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\underline{\\mathbf{P}}" + }, + { + "bbox": [ + 130, + 337, + 482, + 426 + ], + "type": "text", + "content": " together serve as a representation of a 3D object." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 436, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 480, + 460 + ], + "type": "text", + "content": "view " + }, + { + "bbox": [ + 130, + 436, + 480, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 130, + 436, + 480, + 460 + ], + "type": "text", + "content": ". To align the features from multiple views into the same feature space, we share the weights among each of the convolutional neural network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 460, + 480, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 480, + 508 + ], + "type": "text", + "content": "Enhance Multi-view Features: However, the associations between view features " + }, + { + "bbox": [ + 130, + 460, + 480, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 460, + 480, + 508 + ], + "type": "text", + "content": " are not explicitly modeled. Therefore, we develop a cross-view association (CVA) module to obtain enhanced view features " + }, + { + "bbox": [ + 130, + 460, + 480, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 130, + 460, + 480, + 508 + ], + "type": "text", + "content": ", enabling each view to incorporate information from other views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": "For each view feature " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": ", the cross-view association (CVA) module calculates the matching score between it and the overall view feature " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": ". Here, the dot product function is selected as the matching function. Then, " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": " is enhanced based on the correlation scores with other views. The enhanced view feature " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W} \\times \\mathrm{C}}" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": " can be expressed in detail as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 573, + 481, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 573, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 236, + 573, + 481, + 588 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} = \\operatorname {s o f t m a x} \\left(\\omega (\\mathbf {I} _ {i}) ^ {T} \\cdot \\omega (\\mathbf {I})\\right) \\mathbf {I}, \\tag {1}", + "image_path": "0aa0c86cf43f2c4bf0167c2fdb0c5f68a84d521f626082164523b3f07793f2b9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 593, + 426, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 426, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 426, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 593, + 426, + 605 + ], + "type": "inline_equation", + "content": "\\omega (\\mathbf{I}_i)\\in \\mathbb{R}^{\\mathrm{C}}" + }, + { + "bbox": [ + 130, + 593, + 426, + 605 + ], + "type": "text", + "content": " represents the use of average pooling operations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "type": "text", + "content": "Part Sampler: After passing through the cross-view association module, the enhanced view features " + }, + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "type": "text", + "content": " comprehensively incorporate information from other views. Next, we sample parts from the enhanced view features " + }, + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": "Inspired by [23, 45], we implement the attention model " + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": " to learn the spatial distribution of each part within each view. For view " + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": ", the distribution of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 441, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 441, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 441, + 102 + ], + "type": "text", + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "text", + "content": "view parts is represented by the attention map " + }, + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\in \\mathbb{R}^{\\mathrm{H}\\times \\mathrm{W}\\times \\mathrm{M}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "text", + "content": ", which can be computed by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 140, + 481, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 140, + 481, + 153 + ], + "spans": [ + { + "bbox": [ + 225, + 140, + 481, + 153 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {i} = \\psi (\\mathbf {F} _ {i}) = \\left\\{\\mathbf {A} _ {i, 1}, \\mathbf {A} _ {i, 2}, \\dots , \\mathbf {A} _ {i, \\mathrm {M}} \\right\\}, \\tag {2}", + "image_path": "e778c74f8f10c764988dc14a99b3082fc400cb17f70d3ad3413b309d5ba273be.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,j} \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W}}" + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "inline_equation", + "content": "i \\in [1,v]" + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "inline_equation", + "content": "j \\in [1,\\mathrm{M}]" + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "text", + "content": "-th part region or visual pattern of the object's " + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 157, + 479, + 205 + ], + "type": "text", + "content": "-th views, such as a car wheel or an airplane wing. M denotes the number of attention maps, each view extracts the same number of regions of interest." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "text", + "content": "Next, we employ the attention map " + }, + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,j}" + }, + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "text", + "content": " to softly weight the enhanced feature maps " + }, + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "text", + "content": ", followed by the global average pooling operation " + }, + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "text", + "content": " to obtain the parts on view " + }, + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 205, + 480, + 241 + ], + "type": "text", + "content": " (See Eq. 3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 243, + 481, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 243, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 261, + 243, + 481, + 255 + ], + "type": "interline_equation", + "content": "\\boldsymbol {t} _ {i, j} = \\omega \\left(\\mathbf {F} _ {i} \\otimes \\mathbf {A} _ {i, j}\\right), \\tag {3}", + "image_path": "c24d420e3a05b89efef1cb46404a02efa77a958d07deb0fc25c0177a09af90ca.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 258, + 468, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 468, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 468, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 258, + 468, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{i,j} \\in \\mathbb{R}^{\\mathrm{C}}" + }, + { + "bbox": [ + 130, + 258, + 468, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 258, + 468, + 270 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 258, + 468, + 270 + ], + "type": "text", + "content": " denotes element-wise multiplication between two tensors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 271, + 479, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 479, + 295 + ], + "type": "text", + "content": "Finally, we concatenate all the view part features together, resulting in a 3D multi-view part sequence " + }, + { + "bbox": [ + 130, + 271, + 479, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 271, + 479, + 295 + ], + "type": "text", + "content": " (see Eq. 4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 169, + 305, + 481, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 305, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 169, + 305, + 481, + 318 + ], + "type": "interline_equation", + "content": "\\mathbf {T} = \\left[ \\boldsymbol {t} _ {0, 0}, \\boldsymbol {t} _ {0, 1}, \\dots , \\boldsymbol {t} _ {0, \\mathrm {M}}, \\boldsymbol {t} _ {1, 0}, \\boldsymbol {t} _ {1, 1}, \\dots , \\boldsymbol {t} _ {1, \\mathrm {M}}, \\dots , \\boldsymbol {t} _ {v, 0}, \\boldsymbol {t} _ {v, 1}, \\dots , \\boldsymbol {t} _ {v, \\mathrm {M}} \\right], \\tag {4}", + "image_path": "7e5ddcf4418ca7003b200d17f60ef1c7514f05670797acac9706e26888441dca.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 326, + 450, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 450, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 450, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 326, + 450, + 338 + ], + "type": "inline_equation", + "content": "\\mathrm{N}" + }, + { + "bbox": [ + 131, + 326, + 450, + 338 + ], + "type": "text", + "content": " represents the total number of parts " + }, + { + "bbox": [ + 131, + 326, + 450, + 338 + ], + "type": "inline_equation", + "content": "(v\\times \\mathrm{M})" + }, + { + "bbox": [ + 131, + 326, + 450, + 338 + ], + "type": "text", + "content": " in the multi-view." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 356, + 293, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 356, + 293, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 356, + 293, + 369 + ], + "type": "text", + "content": "3.2 Adaptive Part Refinement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 375, + 480, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 482 + ], + "type": "text", + "content": "In Sec. 3.1, we have obtained a multi-view part sequence " + }, + { + "bbox": [ + 130, + 375, + 480, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 375, + 480, + 482 + ], + "type": "text", + "content": ". However, there are differences in parts across different views, where each element " + }, + { + "bbox": [ + 130, + 375, + 480, + 482 + ], + "type": "inline_equation", + "content": "t_{ij}" + }, + { + "bbox": [ + 130, + 375, + 480, + 482 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 375, + 480, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 375, + 480, + 482 + ], + "type": "text", + "content": " can only represent local view features. It is challenging to generalize the overall features of a 3D object relying solely on local view parts. Therefore, we newly propose an Adaptive Part Refinement (APR) module, which aims to explicitly model the correlations between parts across views, thus enhancing the robustness of part features to viewpoint variations. This module adaptively integrates part information from multiple views and refines it to obtain global part features that are more suitable for 3D objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "content": "Formally, the APR (Adaptive Part Refinement) module contains a transformer block. Given some global part tokens [4,7,11] " + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{\\mathrm{L} \\times \\mathrm{C}}" + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{L}" + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "content": " represents the number of parts. We concatenate " + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "content": " as input to the APR module. After passing through the APR module, local parts " + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{i,j}" + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "content": " can learn cross-view correlations and obtain global parts " + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\underline{\\mathbf{P}} \\in \\mathbb{R}^{\\mathrm{L} \\times \\mathrm{C}}" + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "content": " suitable for 3D objects. We also observe the diversity of global part features " + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\underline{\\mathbf{P}}" + }, + { + "bbox": [ + 130, + 483, + 479, + 555 + ], + "type": "text", + "content": " (See Fig. 4)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 572, + 324, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 324, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 324, + 585 + ], + "type": "text", + "content": "3.3 Joint Multi-Part Representation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 592, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 479, + 640 + ], + "type": "text", + "content": "Relying solely on individual parts may not be sufficient to represent a 3D object. To overcome this limitation, we jointly describe the 3D object using multiple parts. We make predictions for each part and then average them to obtain the prediction for the 3D object. Each 3D object can be represented as Eq. 5:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 216, + 649, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 649, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 216, + 649, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\underline {{\\mathbf {P}}} = \\left[ \\underline {{\\boldsymbol {p}}} _ {1}, \\underline {{\\boldsymbol {p}}} _ {2}, \\dots , \\underline {{\\boldsymbol {p}}} _ {\\mathrm {L}} \\right], \\quad \\text {w h e r e} \\underline {{\\mathbf {P}}} \\in \\mathbb {R} ^ {\\mathrm {L} \\times \\mathrm {C}} \\tag {5}", + "image_path": "5c2c188673d28ea59c857393684fdfbed28cd188fc100b36a4b3a8ffb3d4e3df.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Viewpoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "The proposed PANet is optimized through supervised learning, including instance classification loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and part-aware loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{awe}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ". The loss function for multi-view " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 159, + 480, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 159, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 168, + 159, + 480, + 202 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p ^ {i} = \\operatorname {s o f t m a x} \\left(\\underline {{\\boldsymbol {p}}} _ {i} \\boldsymbol {W} _ {p} + \\boldsymbol {b} _ {p}\\right), \\quad \\text {w h e r e} \\underline {{\\boldsymbol {p}}} _ {i} \\in \\mathbb {R} ^ {1 \\times \\mathrm {C}}, i \\in [ 1, \\mathrm {L} ] \\\\ q ^ {j} = \\operatorname {s o f t m a x} \\left(\\boldsymbol {t} _ {j} \\boldsymbol {W} _ {q} + \\boldsymbol {b} _ {q}\\right), \\quad \\text {w h e r e} \\boldsymbol {t} _ {j} \\in \\mathbb {R} ^ {1 \\times (\\mathrm {M} \\times \\mathrm {C})}, j \\in [ 1, v ], \\tag {6} \\\\ \\end{array}", + "image_path": "d71529f3aa88da5c7b7d57904d31db1a4e7a526eec67efe712bcab188eb61623.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 203, + 434, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 203, + 434, + 235 + ], + "spans": [ + { + "bbox": [ + 168, + 203, + 434, + 235 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {a w e} = - \\sum_ {k = 1} ^ {K} y _ {k} \\log \\hat {p} _ {k} + \\gamma \\left(- \\frac {1}{v} \\sum_ {i = 1} ^ {v} \\sum_ {k = 1} ^ {K} y _ {k} \\log q _ {k} ^ {j}\\right),", + "image_path": "ef0f31c06904e77aeacd404f091e5374c16dd3b1dc1ae627f7d2317f989763af.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "W_{p}, W_{q}, b_{p}" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "b_{q}" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": " are learnable parameters. " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "p^{i}" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": " represents the predicted probability of the " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": "-th global part, " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\hat{p}_{k}" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": " represents the average of the predicted probabilities for L global parts, and " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "q^{j}" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": " represents the predicted probability for the " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": "-th view part. " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": " is the ground truth label of class " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": " represents the cross-entropy loss of the final parts. " + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{awe}" + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": " represents the classification loss of M part features in each view, which is used to constrain the generation of parts. Furthermore, we apply techniques such as flipping and random erasure to augment the image for more robust part results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 357, + 230, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 230, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 230, + 371 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 382, + 330, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 330, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 330, + 395 + ], + "type": "text", + "content": "4.1 Datasets and Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": "We conduct an extensive evaluation of PANet on several benchmark datasets, namely ScanObjectNN [53], ModelNet [59], and RGBD [29]. In order to facilitate fair comparisons, we strictly adhere to the methodologies proposed in prior works [26, 48, 55] for preparing the training and testing data on the different datasets. Specifically, we adopt the training scheme introduced by MVCNN [48] and adopt the same viewpoint generation and rendering method as RotationNet [26] to handle the multi-view images from the ModelNet and ScanObjectNN datasets. Experiment on rotated datasets also follows CVR [55] setting. For arbitrary views, the number of views for each 3D object ranges randomly from 10 to 20. Throughout the training process, we utilize the AdamW [36] optimizer with a learning rate of 1e-5. We select ResNet50 as our backbone network and utilize label smoothing regularization [51] to enhance the generalization capability of our model. " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": " is the weighting factor for the part-aware loss as in Eq. 6, which we set " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": " for the experiments. Please refer to the Appendix for implementation details of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 597, + 375, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 375, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 375, + 610 + ], + "type": "text", + "content": "4.2 Comparison with State-of-the-Art Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The ScanObjectNN Datasets. As a recently introduced real-world 3D object classification dataset, ScanObjectNN [53] provides scanned indoor scene data. The dataset comprises approximately 15,000 objects, spanning 15 categories, and includes 2,902 unique object instances. Compared with CAD handcrafted" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 441, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 441, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 441, + 102 + ], + "type": "text", + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 481, + 289 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Classification on ScanObjectNN. 'Aligned' means that every 3D object is in the same pose. 'Rotated' means that each 3D object is misaligned. '-' indicates that the specific result is not reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 481, + 289 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 481, + 289 + ], + "type": "table", + "html": "
MethodTypeAlignedRotated
Per Class Acc.Per Ins. Acc.Per Class Acc.Per Ins. Acc.
PointNet++ [43]Point82.10%84.30%--
DGCNN [54]84.00%86.20%--
MVCNN [48]85.71%87.82%78.21%80.62%
GVCNN [13]86.64%88.68%82.86%83.70%
RotationNet [26]View (fixed)84.88%86.90%74.68%76.16%
View-GCN [56]88.67%90.39%81.99%83.50%
CVR [55]88.39%90.74%84.70%85.59%
View-GCN++ [57]89.10%91.30%85.80%86.80%
Ours93.21%93.99%86.15%87.82%
CVR [55](arbitrary)--68.07%71.36%
Ours91.77%93.48%84.08%86.45%
", + "image_path": "c79fb4f3f8a203411f917406b6b550f0026cbcd093ab1199dc116a4759ec19fa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "text", + "content": "datasets, ScanObjectNN offers more realistic challenges as it reflects the noise, complexity, and incompleteness present in real-world environments, incorporating the appearance and geometric shapes of actual objects. Consequently, it is better for assessing the robustness of the model. We train PANet under different viewpoint settings and compare with point-based methods [43, 54] and view-based methods [13, 26, 48, 55-57], as shown in Tab. 1. View-based methods primarily rely on view aggregation, and it is difficult to robustly aggregate features from structurally misaligned views, especially for arbitrary views. When switching to arbitrary views, the performance of these methods drops sharply. Our PANet is observed to significantly outperform previous approaches in both class accuracy and instance accuracy, whether in aligned views, rotated views, or arbitrary views. In aligned views, compared with the current state-of-the-art methods e.g., View-GCN++ [57] and CVR [55], PANet improves instance accuracy and class accuracy by " + }, + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "inline_equation", + "content": "2.69\\%" + }, + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "inline_equation", + "content": "4.11\\%" + }, + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "text", + "content": ". PANet also achieves significant improvements of " + }, + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "inline_equation", + "content": "1.02\\%" + }, + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "inline_equation", + "content": "0.35\\%" + }, + { + "bbox": [ + 130, + 295, + 482, + 511 + ], + "type": "text", + "content": " in rotated views. Surprisingly, even under arbitrary views, our performance surpasses that of most fixed viewpoint methods, demonstrating that PANet can robustly recognize 3D objects unconstrained by the scene. Please refer to the Appendix for more detailed comparisons." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "text", + "content": "The ModelNet Datasets. ModelNet [59] is a dataset of hand-drawn point clouds generated from CAD models. It comprises 12,311 3D shapes, covering 40 categories, with 9,483 models allocated for training and 2,468 for testing. The experimental results on the ModelNet dataset are presented in Tab. 2. We first compare with point-based and voxel-based methods, including PointNet++ [43], Kd-Networks [27], RS-CNN [22], KP-Convs [52], MVTN [34], 3DShapeNets [58], VoxNet [38], and VRENsemble [3]. Our PANet outperforms them by more than " + }, + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "inline_equation", + "content": "3.93\\%" + }, + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "text", + "content": " in instance accuracy and class accuracy and is also " + }, + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "inline_equation", + "content": "3.73\\%" + }, + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "inline_equation", + "content": "4.37\\%" + }, + { + "bbox": [ + 130, + 521, + 483, + 668 + ], + "type": "text", + "content": " higher in free viewpoints. We also compared with view-based approaches, including MVCNN-new [48], GVCNN [13], HEAR [5], RotationNet [26], View-GCN [56], MVT [6], HMVCM [32], MVD [24], DAN [41], DRCNN [50], EMV [12], View-GCN++ [57], CVR [55], and HGNN+ [16]. PANet consistently outper" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 372, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 372, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Viewpoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 178, + 299, + 412 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 301, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 301, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 301, + 170 + ], + "type": "text", + "content": "Table 2: Classification on Aligned ModelNet40. '- indicates that the specific result is not reported, 'fixed' denotes sampling from a fixed viewpoint. 'free' means that the arbitrary number of viewpoints." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 178, + 299, + 412 + ], + "lines": [ + { + "bbox": [ + 133, + 178, + 299, + 412 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 299, + 412 + ], + "type": "table", + "html": "
Aligned ModelNet40
MethodTypeClass Acc.Ins. Acc.
3DShapeNets [58]77.3%-
VoxNet [38]Voxel83.0%-
VRENsemble [3]91.4%93.8%
PointNet++ [43]-91.9%
Kd-Networks [27]88.5%91.8%
RS-CNN [22]Point-93.6%
KP-Convs [52]-92.9%
MVTN [34]92.0%93.8%
MVCNN-new [48]92.40%95.00%
GVCNN [13]90.70%93.10%
HEAR [5]95.20%96.70%
RotationNet [26]-97.37%
View-GCN [56]96.50%97.60%
MVT [6]-97.50%
HMVCM [32]-94.57%
MVD [24]View92.61%94.41%
DAN [41]-93.50%
DRCNN [50]94.86%96.84%
View-GCN++ [57]96.50%97.60%
CVR [55]95.77%97.16%
HGNN+ [16]-96.92%
Ours96.60%97.73%
Ours(free)96.37%97.53%
", + "image_path": "f4b4bfd3fb70a04470e07b7d7ef5b3e193ba4b9c713213cd8ff63d06e8880f04.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 313, + 146, + 480, + 263 + ], + "blocks": [ + { + "bbox": [ + 311, + 114, + 481, + 136 + ], + "lines": [ + { + "bbox": [ + 311, + 114, + 481, + 136 + ], + "spans": [ + { + "bbox": [ + 311, + 114, + 481, + 136 + ], + "type": "text", + "content": "Table 3: The accuracy of 3D object classification on Rotated ModelNet40." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 146, + 480, + 263 + ], + "lines": [ + { + "bbox": [ + 313, + 146, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 480, + 263 + ], + "type": "table", + "html": "
Rotated ModelNet40
Method#ViewClass Acc.Ins. Acc.
MVCNN [48]87.95%88.17%
GVCNN [13]89.69%88.10%
EMV [12]88.80%90.00%
RotationNet [26]fixed84.74%85.29%
View-GCN [56]85.90%88.25%
CVR [55]91.12%92.22%
View-GCN++ [57]90.70%92.80%
Ours90.35%92.87%
CVR [55]arbitrary84.01%86.91%
Ours90.69%92.23%
", + "image_path": "a48c58fec3a11daf4c3e18d21abc723b2b792a8f7b0a6a73001e2abc1529683c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 297, + 479, + 411 + ], + "blocks": [ + { + "bbox": [ + 312, + 265, + 480, + 286 + ], + "lines": [ + { + "bbox": [ + 312, + 265, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 312, + 265, + 480, + 286 + ], + "type": "text", + "content": "Table 4: The accuracy of object classification on RGBD (in%)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 297, + 479, + 411 + ], + "lines": [ + { + "bbox": [ + 315, + 297, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 315, + 297, + 479, + 411 + ], + "type": "table", + "html": "
Method#ViewIns. Acc.
MDSICNN [2]≥ 12089.6%
CFK [8]≥ 12086.8%
MMDCNN [44]≥ 12086.8%
RotationNet [26]1289.3%
View-GCN [56]1294.3%
View-GCN++ [57]1292.5%
Ours1296.5%
MVCNN [48]89.0%
GVCNN [13]89.8%
CVR [55]4-1291.8%
Ours95.9%
", + "image_path": "5df074cc37ac3eff97bb27ba5cc7229a9650ff9eedbb317fdaf46919d2bdef93.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 417, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 480, + 464 + ], + "type": "text", + "content": "forms these methods in fixed views. Its performance at free viewpoints surpasses that of most existing approaches. We next report the robustness of PANet in rotated views, as shown in Tab. 3. Under the rotated views, PANet also achieves the highest score in instance accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "type": "text", + "content": "The RGBD Datasets. The RGBD [29] dataset comprises 51 categories of domestic objects captured in the real world, encompassing RGBD views from multiple viewpoints. To ensure a fair comparison, we adopt the identical experimental setup as RotationNet [26] and View-GCN [56]. Specifically, we uniformly sample 12 RGB images from the same circle for each 3D object as input and report the average results using ten-fold cross-validation. We keep the same settings as CVR [55] in arbitrary views, sampling images from 4 to 12 viewpoints as input to the network. The experimental results are summarized in Tab. 4. The performance of PANet in arbitrary views is impressive. Compared with work in arbitrary views e.g., CVR [55], PANet significantly improves by " + }, + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "type": "text", + "content": ". Additionally, compared with the work in fixed views, such as view-GCN [56], PANet has also improved by " + }, + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 464, + 482, + 609 + ], + "type": "text", + "content": ". This further validates the superiority of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 239, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 239, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 239, + 636 + ], + "type": "text", + "content": "4.3 Model Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "All experiments are conducted on the ScanObjectNN dataset in rotated views. Please refer to the Appendix for detailed experimental analyses." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 440, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 440, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 440, + 102 + ], + "type": "text", + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 117, + 206, + 182 + ], + "blocks": [ + { + "bbox": [ + 140, + 117, + 206, + 182 + ], + "lines": [ + { + "bbox": [ + 140, + 117, + 206, + 182 + ], + "spans": [ + { + "bbox": [ + 140, + 117, + 206, + 182 + ], + "type": "image", + "image_path": "c8ab5804a3db7cc39f0d2e0fe951b9c25564269d8d80930d3fd41747820e7a72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 188, + 301, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 188, + 301, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 188, + 301, + 232 + ], + "type": "text", + "content": "Fig.4: Visualization of correlations between part features " + }, + { + "bbox": [ + 132, + 188, + 301, + 232 + ], + "type": "inline_equation", + "content": "\\underline{\\mathbf{P}}" + }, + { + "bbox": [ + 132, + 188, + 301, + 232 + ], + "type": "text", + "content": ". The left figure illustrates the correlation between " + }, + { + "bbox": [ + 132, + 188, + 301, + 232 + ], + "type": "inline_equation", + "content": "\\underline{\\mathbf{P}}" + }, + { + "bbox": [ + 132, + 188, + 301, + 232 + ], + "type": "text", + "content": " after applying " + }, + { + "bbox": [ + 132, + 188, + 301, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{awe}" + }, + { + "bbox": [ + 132, + 188, + 301, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 210, + 118, + 292, + 182 + ], + "blocks": [ + { + "bbox": [ + 210, + 118, + 292, + 182 + ], + "lines": [ + { + "bbox": [ + 210, + 118, + 292, + 182 + ], + "spans": [ + { + "bbox": [ + 210, + 118, + 292, + 182 + ], + "type": "image", + "image_path": "d7abb86743ead2a22ae1876985ac2207d96a77cebac32c138c597ee86f87b1d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 313, + 186, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 311, + 116, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 311, + 116, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 311, + 116, + 482, + 170 + ], + "type": "text", + "content": "Table 5: PANet component ablations. The first line serves as the baseline. " + }, + { + "bbox": [ + 311, + 116, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{awe}" + }, + { + "bbox": [ + 311, + 116, + 482, + 170 + ], + "type": "text", + "content": " refers to the weakly supervised part-aware loss function. CVA stands for Cross-View Association module." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 186, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 481, + 232 + ], + "type": "table", + "html": "
CVALaweClass Acc.Ins. Acc.
✓83.76%84.56%
84.99%85.25%
✓85.45%86.96%
", + "image_path": "140c643ec3725e71d9b2d3f298af1b29fa3dab64fef08772bc3c93897cd211af.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 138, + 263, + 473, + 335 + ], + "blocks": [ + { + "bbox": [ + 147, + 242, + 465, + 254 + ], + "lines": [ + { + "bbox": [ + 147, + 242, + 465, + 254 + ], + "spans": [ + { + "bbox": [ + 147, + 242, + 465, + 254 + ], + "type": "text", + "content": "Table 6: Explore the impact of different viewpoint sampling on performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 263, + 473, + 335 + ], + "lines": [ + { + "bbox": [ + 138, + 263, + 473, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 473, + 335 + ], + "type": "table", + "html": "
DatasetMethodAlignedRotated
Class Acc.Ins. Acc.Class Acc.Ins. Acc.
ModelNet40random sampling96.37%97.53%90.69%92.23%
furthest point sampling95.44%97.41%89.73%92.26%
ScanObjectNNrandom sampling91.77%93.48%84.08%86.45%
furthest point sampling92.97%93.48%84.80%85.93%
", + "image_path": "c6246b8ddbf77713fbd11525b1f0f5dcf1c5fcb3020bce85dc9d3c0eeef5aa0c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": "PANet component ablations We evaluate the role of the cross-view association module and " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{awe}" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": ", as shown in Tab. 5. The CVA module plays a positive role in enhancing the performance of PANet, improving instance accuracy and class accuracy by " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "0.69\\%" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "1.23\\%" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": ". The results show that the CVA module enables each view to more effectively integrate information from other views, further enhancing the representation capability of the view features. Our method also improves by " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "2.40\\%" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "1.76\\%" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": " in two metrics by introducing the " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{awe}" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": " constraint. In addition, we also conduct correlation analysis between part features " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\underline{\\mathbf{P}}" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": ", as shown in Fig. 4. We observe that without applying " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{awe}" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": " constraints, the correlation between parts exhibits a flat similarity matrix, indicating that the parts are unvaried and cannot be distinguished in the feature space. In contrast, the part features become more diverse with " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{awe}" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": " constraint. Therefore, we conclude that " + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{awe}" + }, + { + "bbox": [ + 130, + 339, + 481, + 495 + ], + "type": "text", + "content": " can supervise the perception of parts to obtain diverse part features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 497, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 640 + ], + "type": "text", + "content": "Influence of Different Viewpoint Sampling We introduce two different sampling methods: random sampling and furthest point sampling (FPS). Random sampling refers to the process of randomly selecting 10 to 20 views from each object. This method is closer to real-world scenarios, and we showcased PANet's performance in Tab. 1, 2, 3, 4 based on this sampling method. FPS selects 10 to 20 views based on the furthest point sampling principle. It provides a uniform sampling method that covers every orientation of the 3D objects. As shown in Tab. 6, we observed that the performance of random sampling and FPS across different datasets is nearly similar. This result shows that PANet is not sensitive to the choice of sampling methods. The model demonstrates robustness in adapting to changes in viewpoint caused by different sampling methods while maintaining stable performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Influence of the Number of the Views We further evaluate PANet by considering different backbone networks and numbers of views, as shown in Tab. 7." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Viewpoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 162, + 135, + 452, + 190 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "type": "text", + "content": "Table 7: Explore the impact of number of views and backbones on performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 135, + 452, + 190 + ], + "lines": [ + { + "bbox": [ + 162, + 135, + 452, + 190 + ], + "spans": [ + { + "bbox": [ + 162, + 135, + 452, + 190 + ], + "type": "table", + "html": "
BackboneViews
1 views5 views10 views15 views20 views
resnet1845.63%70.84%75.66%79.90%82.85%
resnet5067.07%79.59%82.50%85.93%86.62%
", + "image_path": "2b450ae56c007bb867e489c90cead97370fe8a3b9570410f5cc8dba9a9663544.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 236, + 296, + 289 + ], + "blocks": [ + { + "bbox": [ + 132, + 194, + 301, + 227 + ], + "lines": [ + { + "bbox": [ + 132, + 194, + 301, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 301, + 227 + ], + "type": "text", + "content": "Table 8: Exploring the effect of the number of attention maps, M denotes the number of attention maps per view." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 236, + 296, + 289 + ], + "lines": [ + { + "bbox": [ + 135, + 236, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 135, + 236, + 296, + 289 + ], + "type": "table", + "html": "
#Attention MapsClass Acc.Ins. Acc.
M = 1684.57%86.28%
M = 3285.36%87.14%
M = 6486.50%87.31%
M = 12886.38%87.14%
", + "image_path": "423c88ff53efb21ba3e3c5440d7554b6d73c8b218d52e659c9f17a0cdad99ba5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 237, + 477, + 289 + ], + "blocks": [ + { + "bbox": [ + 312, + 194, + 481, + 227 + ], + "lines": [ + { + "bbox": [ + 312, + 194, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 194, + 481, + 227 + ], + "type": "text", + "content": "Table 9: Exploring the effect of the number of part tokens, L indicates the number of parts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 237, + 477, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 237, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 477, + 289 + ], + "type": "table", + "html": "
#Number of PartsClass Acc.Ins. Acc.
L = 185.50%86.28%
L = 886.20%87.14%
L = 1686.50%87.31%
L = 3285.63%86.31%
", + "image_path": "bf45da096ebaf91f5bce0daa749db953fe5104d94015819d7befe47ff2d95ba5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 295, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 481, + 462 + ], + "type": "text", + "content": "In previous studies, such as View-GCN [56], because view aggregation leads to feature loss. Choosing a stronger backbone network may not necessarily lead to performance improvement in certain dataset. However, our observations reveal that the choice of backbone significantly affects PANet performance. This is because PANet directly integrates multiple parts to characterize 3D objects. Therefore, a stronger backbone network is more advantageous for perceiving part features. Furthermore, we notice a positive correlation between the number of views and the performance of PANet. It is worth noting that previous studies on multi-view aggregation methods, such as MHBN [65], MVLADN [64], and GVCNN [13], have indicated that increasing the number of views does not necessarily lead to better representation of 3D objects. PANet combines multiple parts for representation, and the APR module can also filter out some redundant parts. This means that PANet is not affected by redundant views. Our experimental results also verify this idea." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 463, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 594 + ], + "type": "text", + "content": "Influence of the Number of the Attention Maps and Parts We show the impact of the number of attention maps M and the number of part tokens L, as shown in Tab. 8 and Tab. 9. The number of attention maps M determines the discriminative region in each view, and the part tokens L determine the number of global parts. More M and L usually contribute to better performance. MACNN [68] and WSDAN [23] report similar conclusions. The experiments in Tab. 8 show that as the value of M increases, the classification accuracy gradually improves. When " + }, + { + "bbox": [ + 130, + 463, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{M} = 64" + }, + { + "bbox": [ + 130, + 463, + 482, + 594 + ], + "type": "text", + "content": ", the performance tends to stabilize. The results in Tab. 9 show that the performance improves as the number of parts increases and reaches a peak at " + }, + { + "bbox": [ + 130, + 463, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{L} = 16" + }, + { + "bbox": [ + 130, + 463, + 482, + 594 + ], + "type": "text", + "content": ". Compared with a single part, multiple parts provide richer information and enhance the generalization of the model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 610, + 227, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 227, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 227, + 621 + ], + "type": "text", + "content": "4.4 Visualization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We randomly select some 3D models from a test set and compare the semantic differences between the features extracted by MVCNN and PANet, as shown in Fig. 5. MVCNN [48] is a classic feature aggregation method. We notice that the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 441, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 441, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 441, + 102 + ], + "type": "text", + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "features extracted by MVCNN are not entirely focused on the object. This is because the uneven weight distribution leads to view aggregation methods ignoring certain view information. It is worth noting that PANet does not discard any views but instead considers the features of each view, leveraging the valuable information within each view. We can observe that different views emphasize various part locations, with each view capturing multiple discriminative regions (e.g., the airplane nose, body, and wings). This observation further validates the diversity of captured parts. Furthermore, we observe that the view aggregation method does not explicitly model the unique attributes of 3D objects, leading to feature ambiguity. In contrast, part-based representation reveals the inherent properties of objects, enhancing the interpretability of features and making recognition easier. Please refer to the Appendix for more visualization results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 277, + 220, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 220, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 220, + 291 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 471 + ], + "type": "text", + "content": "In this paper, we propose a novel part-aware network (PANet), which is a part-based representation. Compared with view aggregation methods, this part-based representation has properties such as viewpoint invariance and rotation robustness, which can more effectively address the 3D object recognition problem under arbitrary views. Moreover, we proposed the part-aware loss to ensure that part features can better focus on the object, thereby fully utilizing the information from each view. We evaluate the performance of PANet on several 3D datasets, including ScanObjectNN, ModelNet, and RGBD. The experimental results show that PANet has made significant progress, especially in arbitrary views. We also conduct ablation and visualization experiments to study the impact of different components in our method. It is worth noting that in these experiments, we found that our method is not affected by redundant views (Sec. 4.3), and the part features can better reveal the inherent properties of objects (Sec. 4.4). In conclusion, the experimental results prove the effectiveness of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 482, + 481, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 482, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 482, + 481, + 515 + ], + "type": "text", + "content": "Acknowledgement This research is supported by the Shenzhen Science and Technology Plan Project No. KJZD20230923113800002 and Shenzhen LongHua Fundamental Research Program No. 10162A20230325B73A546." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 372, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 372, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Viewpoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 157, + 302, + 582 + ], + "blocks": [ + { + "bbox": [ + 136, + 157, + 302, + 582 + ], + "lines": [ + { + "bbox": [ + 136, + 157, + 302, + 582 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 302, + 582 + ], + "type": "image", + "image_path": "9f3fe613b8e83a64611c25212ff7ed83aafbdc9c41e62438e38e43e91867e945.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 599, + 482, + 620 + ], + "lines": [ + { + "bbox": [ + 132, + 599, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 482, + 620 + ], + "type": "text", + "content": "Fig. 5: Comparing the semantic differences between view-based aggregation methods and PANet in feature extraction." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 157, + 477, + 582 + ], + "blocks": [ + { + "bbox": [ + 310, + 157, + 477, + 582 + ], + "lines": [ + { + "bbox": [ + 310, + 157, + 477, + 582 + ], + "spans": [ + { + "bbox": [ + 310, + 157, + 477, + 582 + ], + "type": "image", + "image_path": "bf8ea68dca1f99451d4bde25a8a38882d323cd4d796afa8e2fa1d31136628ff5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 441, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 441, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 441, + 102 + ], + "type": "text", + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "text", + "content": "1. Afham, M., Dissanayake, I., Dissanayake, D., Dharmasiri, A., Thilakarathna, K., Rodrigo, R.: Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9902-9912 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "type": "text", + "content": "2. Asif, U., Bennamoun, M., Sohel, F.A.: A multi-modal, discriminative and spatially invariant cnn for rgb-d object labeling. IEEE transactions on pattern analysis and machine intelligence 40(9), 2051-2065 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "text", + "content": "3. Brock, A., Lim, T., Ritchie, J.M., Weston, N.: Generative and discriminative voxel modeling with convolutional neural networks. arXiv preprint arXiv:1608.04236 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 481, + 298 + ], + "type": "text", + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "text", + "content": "5. Chen, J., Qin, J., Shen, Y., Liu, L., Zhu, F., Shao, L.: Learning attentive and hierarchical representations for 3d shape recognition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 105-122. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 343, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 364 + ], + "type": "text", + "content": "6. Chen, S., Yu, T., Li, P.: Mvt: Multi-view vision transformer for 3d object recognition. arXiv preprint arXiv:2110.13083 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "text", + "content": "7. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 398, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 398, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 398, + 481, + 430 + ], + "type": "text", + "content": "8. Cheng, Y., Cai, R., Zhao, X., Huang, K.: Convolutional fisher kernels for rgb-d object recognition. In: 2015 International Conference on 3D Vision. pp. 135-143. IEEE (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 432, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 481, + 464 + ], + "type": "text", + "content": "9. Delitzas, A., Parelli, M., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "type": "text", + "content": "0. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 498, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 542 + ], + "type": "text", + "content": "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "text", + "content": "2. Esteves, C., Xu, Y., Allen-Blanchette, C., Daniilidis, K.: Equivariant multi-view networks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1568-1577 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "text", + "content": "3. Feng, Y., Zhang, Z., Zhao, X., Ji, R., Gao, Y.: GvCNN: Group-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 264-272 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "text", + "content": "4. Fu, J., Zheng, H., Mei, T.: Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4438-4446 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "5. Gao, K., Gao, Y., He, H., Lu, D., Xu, L., Li, J.: Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Viewpoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Gao, Y., Feng, Y., Ji, S., Ji, R.: Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3181-3199 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 225 + ], + "type": "text", + "content": "18. Han, Z., Lu, H., Liu, Z., Vong, C.M., Liu, Y.S., Zwicker, M., Han, J., Chen, C.P.: 3d2seqviews: Aggregating sequential views for 3d global feature learning by cnn with hierarchical attention aggregation. IEEE Transactions on Image Processing 28(8), 3986-3999 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 225, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 481, + 258 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 258, + 481, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 258, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 481, + 279 + ], + "type": "text", + "content": "20. Herreras, E.B.: Cognitive neuroscience; the biology of the mind. Cuadernos de Neuropsicología/Panamerican Journal of Neuropsychology 4(1), 87-90 (2010)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 279, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 279, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 279, + 481, + 311 + ], + "type": "text", + "content": "21. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9202-9212 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 311, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 311, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 481, + 332 + ], + "type": "text", + "content": "22. Hu, L., Qin, M., Zhang, F., Du, Z., Liu, R.: Rscnn: a cnn-based method to enhance low-light remote-sensing images. Remote Sensing 13(1), 62 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 481, + 365 + ], + "type": "text", + "content": "23. Hu, T., Qi, H., Huang, Q., Lu, Y.: See better before looking closer: Weakly supervised data augmentation network for fine-grained visual classification. arXiv preprint arXiv:1901.09891 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 365, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 481, + 396 + ], + "type": "text", + "content": "24. Huang, J., Yan, W., Li, T., Liu, S., Li, G.: Learning the global descriptor for 3-d object recognition based on multiple views decomposition. IEEE Transactions on Multimedia 24, 188-201 (2022). https://doi.org/10.1109/TMM.2020.3047762" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 481, + 429 + ], + "type": "text", + "content": "25. Huang, S., Xu, Z., Tao, D., Zhang, Y.: Part-stacked cnn for fine-grained visual categorization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1173-1182 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 429, + 481, + 460 + ], + "type": "text", + "content": "26. Kanezaki, A., Matsushita, Y., Nishida, Y.: Rotationnet for joint object categorization and unsupervised pose estimation from multi-view images. IEEE transactions on pattern analysis and machine intelligence 43(1), 269-283 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 133, + 460, + 481, + 493 + ], + "type": "text", + "content": "27. Klokov, R., Lempitsky, V.: Escape from cells: Deep kd-networks for the recognition of 3d point cloud models. In: Proceedings of the IEEE international conference on computer vision. pp. 863-872 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 493, + 481, + 525 + ], + "type": "text", + "content": "28. Kumawat, S., Raman, S.: Lp-3dcnn: Unveiling local phase in 3d convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4903-4912 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 481, + 557 + ], + "type": "text", + "content": "29. Lai, K., Bo, L., Ren, X., Fox, D.: A large-scale hierarchical multi-view rgb-d object dataset. In: 2011 IEEE international conference on robotics and automation. pp. 1817-1824. IEEE (2011)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 481, + 601 + ], + "type": "text", + "content": "30. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "type": "text", + "content": "31. Lin, Y., Gou, Y., Liu, X., Bai, J., Lv, J., Peng, X.: Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4447-4461 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 666 + ], + "type": "text", + "content": "32. Liu, A.A., Zhou, H., Nie, W., Liu, Z., Liu, W., Xie, H., Mao, Z., Li, X., Song, D.: Hierarchical multi-view context modelling for 3d object classification and retrieval. Information Sciences 547, 984-995 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 441, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 441, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 441, + 102 + ], + "type": "text", + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "34. Liu, S., Nguyen, V., Rehg, I., Tu, Z.: Recognizing objects from any view with object and viewer-centered representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "35. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing 30, 1744-1758 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "text", + "content": "36. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "37. Lowe, D.: Object recognition from local scale-invariant features. In: Proceedings of the Seventh IEEE International Conference on Computer Vision. vol. 2, pp. 1150-1157 vol.2 (1999). https://doi.org/10.1109/ICCV.1999.790410" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "38. Maturana, D., Scherer, S.: Voxnet: A 3d convolutional neural network for real-time object recognition. In: 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 922-928. IEEE (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 346 + ], + "type": "text", + "content": "39. Meng, M., Zhang, T., Yang, W., Zhao, J., Zhang, Y., Wu, F.: Diverse complementary part mining for weakly supervised object localization. IEEE Transactions on Image Processing 31, 1774-1788 (2022). https://doi.org/10.1109/TIP.2022.3145238" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "type": "text", + "content": "40. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 482, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 403 + ], + "type": "text", + "content": "41. Nie, W., Zhao, Y., Song, D., Gao, Y.: Dan: Deep-attention network for 3d shape recognition. IEEE Transactions on Image Processing 30, 4371-4383 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 403, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 435 + ], + "type": "text", + "content": "42. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "43. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "44. Rahman, M.M., Tan, Y., Xue, J., Lu, K.: Rgb-d object recognition with multimodal deep convolutional neural networks. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 991-996. IEEE (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "text", + "content": "45. Rao, Y., Chen, G., Lu, J., Zhou, J.: Counterfactual attention learning for fine-grained visual categorization and re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1025–1034 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 534, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 567 + ], + "type": "text", + "content": "46. Riegler, G., Osman Ulusoy, A., Geiger, A.: Octnet: Learning deep 3d representations at high resolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3577-3586 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "type": "text", + "content": "47. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "text", + "content": "48. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "49. Su, J.C., Gadelha, M., Wang, R., Maji, S.: A deeper look at 3d shape classifiers. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Viewpoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "50. Sun, K., Zhang, J., Liu, J., Yu, R., Song, Z.: Drcnn: Dynamic routing convolutional neural network for multi-view 3d object recognition. IEEE Transactions on Image Processing 30, 868-877 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "51. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "52. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "text", + "content": "53. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1588-1597 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "54. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog) 38(5), 1-12 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "55. Wei, X., Gong, Y., Wang, F., Sun, X., Sun, J.: Learning canonical view representation for 3d shape recognition with arbitrary views. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 407-416 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "text", + "content": "56. Wei, X., Yu, R., Sun, J.: View-gcn: View-based graph convolutional network for 3d shape analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1850-1859 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "57. Wei, X., Yu, R., Sun, J.: Learning view-based graph convolutional network for multi-view 3d shape analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "58. Wu, S., Khosla, Y., Zhang, T.: D shapenets: A deep representation for volumetric shape modeling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA. pp. 7-12 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "59. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "60. Xu, Y., Zheng, C., Xu, R., Quan, Y., Ling, H.: Multi-view 3d shape recognition via correspondence-aware deep learning. IEEE Transactions on Image Processing 30, 5299-5312 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "61. Xue, L., Gao, M., Xing, C., Martin-Martin, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning unified representation of language, image and point cloud for 3d understanding. arXiv preprint arXiv:2212.05171 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "text", + "content": "62. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "63. Yang, Z., Wang, L.: Learning relationships for multi-view 3d object recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7505-7514 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "64. Yu, T., Meng, J., Yang, M., Yuan, J.: 3d object representation learning: A set-to-set matching perspective. IEEE Transactions on Image Processing 30, 2168-2179 (2021). https://doi.org/10.1109/TIP.2021.3049968" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "65. Yu, T., Meng, J., Yuan, J.: Multi-view harmonized bilinear network for 3d object recognition. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 186-194 (2018). https://doi.org/10.1109/CVPR.2018.00027" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 441, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 441, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 441, + 102 + ], + "type": "text", + "content": "Linlong Fan, Ye Huang, Yanqi Ge, Wen Li, and Lixin Duan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "66. Zhang, H., Xu, T., Elhoseiny, M., Huang, X., Zhang, S., Elgammal, A., Metaxas, D.: Spda-cnn: Unifying semantic part detection and abstraction for fine-grained recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1143-1152 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "text", + "content": "67. Zhang, N., Donahue, J., Girshick, R., Darrell, T.: Part-based r-cnns for fine-grained category detection. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 834-849. Springer (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 481, + 237 + ], + "type": "text", + "content": "68. Zheng, H., Fu, J., Mei, T., Luo, J.: Learning multi-attention convolutional neural network for fine-grained image recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 5209-5217 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 481, + 270 + ], + "type": "text", + "content": "69. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 371, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond Viewpoint" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/4cf86d6f-6634-4c59-b4d4-c22212431643_content_list.json b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/4cf86d6f-6634-4c59-b4d4-c22212431643_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cef29637e388cea3a20b59f661acfb39dda3a424 --- /dev/null +++ b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/4cf86d6f-6634-4c59-b4d4-c22212431643_content_list.json @@ -0,0 +1,1741 @@ +[ + { + "type": "text", + "text": "Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyeonwoo $\\mathrm{Kim}^{1*}$ , Sookwan Han $^{1*}$ , Patrick Kwon $^{2}$ , and Hanbyul Joo $^{1}$", + "bbox": [ + 250, + 233, + 750, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Seoul National University", + "bbox": [ + 406, + 261, + 594, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Naver Webtoon AI", + "bbox": [ + 429, + 277, + 571, + 289 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36ebae95e9d9b0eb350d9c84445a18c8b0280fec06dd412637ea18e4232cbb5b.jpg", + "image_caption": [ + "Fig. 1: Given a 3D object, we generate numerous 3D Human-Object Interaction (HOI) samples using text prompts, and learn a novel affordance representation called Comprehensive Affordance (ComA) which models both contact and non-contact HOI patterns." + ], + "image_footnote": [], + "bbox": [ + 217, + 299, + 787, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Understanding the inherent human knowledge in interacting with a given environment (e.g., affordance) is essential for improving AI to better assist humans. While existing approaches primarily focus on human-object contacts during interactions, such affordance representation cannot fully address other important aspects of human-object interactions (HOIs), i.e. patterns of relative positions and orientations. In this paper, we introduce a novel affordance representation, named Comprehensive Affordance (ComA). Given a 3D object mesh, ComA models the distribution of relative orientation and proximity of vertices in interacting human meshes, capturing plausible patterns of contact, relative orientations, and spatial relationships. To construct the distribution, we present a novel pipeline that synthesizes diverse and realistic 3D HOI samples given any 3D target object mesh. The pipeline leverages a pre-trained 2D inpainting diffusion model to generate HOI images from object renderings and lifts them into 3D. To avoid the generation of false affordances, we propose a new inpainting framework, Adaptive Mask Inpainting. Since ComA is built on synthetic samples, it can extend to any object in an unbounded manner. Through extensive experiments, we demonstrate that ComA outperforms competitors that rely on human annotations in modeling contact-based affordance. Importantly, we also showcase the potential of ComA to reconstruct human-object interactions in 3D through an optimization framework, highlighting its advantage in incorporating both contact and non-contact properties.", + "bbox": [ + 259, + 474, + 740, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Affordance $\\cdot$ Human-Object Interaction $\\cdot$ Diffusion Models", + "bbox": [ + 261, + 805, + 740, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Indicates equal contribution", + "bbox": [ + 215, + 824, + 413, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Humans possess an innate ability to perceive the functionalities provided by the environment or objects and efficiently utilize them—a concept known as affordance [17]. Affordance incorporates a variety of patterns typically observable in human-object interactions (HOIs), including not only plausible physical contact but also spatial and orientational non-contact relationships. For example, while using a laptop, the orientation and distance patterns between human face or other body parts and the screen can be considered part of affordance, alongside the hand regions touching the laptop keyboard. Teaching such affordance knowledge to machines is crucial for enabling them to better understand or mimic human behaviors. Consequently, numerous studies have been conducted to implement such affordance knowledge into AI and robotics [1,20,23,37,40,63,79,87,89].", + "bbox": [ + 212, + 185, + 785, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most previous approaches however, primarily focus on a limited spectrum of affordances. Some approaches represent the human and object regions in contact using contact scores on 2D images [1,40], 3D objects [87], and 3D humans [23,79] using direct supervision from human annotations. Others represent affordances by inferring [63] or generating [37,89] plausible human-object pairs in contact-oriented object categories. While a few recent studies tackle learning the spatial relations [20,88] between a human and an object, these methods mainly focus on object categories where physical contact is dominant than non-contact patterns.", + "bbox": [ + 212, + 353, + 785, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce Comprehensive Affordance (ComA), a novel representation of an affordance encompassing the various aspects of human-object interaction, including orientational tendencies and relative positions. The motivation for our design arises from the observation that there exist typical 3D relations between entire body parts and objects during HOIs. For instance, human faces, torsos, and arms often exhibit specific distances and orientations, with some variability, when interacting with certain object categories, as shown in Fig. 2. Our ComA is designed to model these patterns by constructing a distribution between each pair of object surface points and human surface points in 3D. This distribution represents the likelihood of human body parts displaying specific spatial and orientational relations with object points, as shown in Fig. 3.", + "bbox": [ + 212, + 476, + 785, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning ComA requires various 3D HOI samples showing the way humans use the target object in 3D, from which pairwise distributions of 3D HOI relations can be constructed. Manually annotating such cues is challenging, compared to the annotations focusing solely on contact regions [40, 63, 79, 85, 87, 88]. As a key solution, we present a pipeline to synthesize a large-scale synthetic 3D HOI samples leveraging a pre-trained 2D diffusion model. Interestingly, we observe that a pre-trained 2D diffusion model captures the affordance knowledge in the form of 2D images, as shown in Fig. 2, where the images are synthesized from text prompts. However, leveraging these 2D HOI cues for construction of ComA (which requires 3D HOI samples) presents the following challenges: (1) the diffusion model needs to be applied for inpainting plausible humans without altering the original shape and pose of the target object, and (2) 3D HOI samples should be lifted from the synthesized 2D images. To address these challenges, we present", + "bbox": [ + 212, + 643, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d78d3eae5f629e72c5458a6734b1ddd26aa640641cbe34a733b21046f564dfb8.jpg", + "image_caption": [ + "Text Prompt: 1 person using a laptop" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 498, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cc1e59e40fb8408604b261e917e8179ad2c0d9e567e71a8a22c8a53b30813162.jpg", + "image_caption": [ + "Text Prompt: 1 person using a telescope", + "Fig. 2: Typically, people (1) view the screen (2) from relatively distant distance while using a laptop (Left), whereas they (1) peer into it (2) from a close distance while using a telescope (Right). Pre-trained diffusion model has a knowledge of these (1) orientational and (2) spatial relation between human and object during interaction." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 787, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adaptive Mask Inpainting to insert humans without altering the appearance of target objects in 2D, along with a 3D lifting pipeline from 2D HOI cues.", + "bbox": [ + 212, + 340, + 784, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We demonstrate the efficacy of our approach by learning ComA on 100 3D object meshes in diverse categories collected from various sources [2,9,29,52,71, 83]. We compare our ComA with existing approaches [79,87] on the BEHAV dataset [2], demonstrating that ours learned from synthetic cues outperforms competitors trained on manual annotations for contact-based affordance. Additionally, we verify the advantage of ComA against contact-only representations, in the scenario of reconstructing HOIs via an optimization framework. We will release the results and our source code1.", + "bbox": [ + 212, + 369, + 784, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our main contributions can be summarized as follows: (1) we introduce a new representation ComA, which models the distribution of both explicit contact and non-contact patterns during HOIs; (2) we present a pipeline to build ComA from synthetic datasets by leveraging a pre-trained 2D diffusion model, scalable to any 3D object without requiring laborious annotations; and (3) we propose Adaptive Mask Inpainting, a method for preserving the original context of an image in using inpainting diffusion model.", + "bbox": [ + 212, + 491, + 785, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 619, + 387, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning Visual Affordance. Affordance was introduced by Gibson [17] as a set of functionalities that the environment furnishes to an agent (e.g., human, animal). The concept extends to the vision and robotics area, where the focus is on teaching embodied agents to interact with scenes [1, 15, 39, 56, 84], mimicking human-object [2, 7, 10, 13, 23, 28, 33, 42, 54, 63, 65, 69, 74, 91, 92], or hand-object [3-5, 76, 96] interactions. Earlier methods focused on learning action category labels [6, 25, 38] with bounding boxes but lacked a detailed description of affordance. Due to the limited affordance information available from text descriptions or labels, some studies depicted affordance as contact regions or heatmaps on objects [1, 40, 87] and humans [23, 79]. However, learning affordances beyond the contact was challenging since existing datasets created through a multi-camera system [2, 29],", + "bbox": [ + 212, + 648, + 787, + 816 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Discovering Comprehensive Affordance for 3D Objects", + "bbox": [ + 367, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1https://github.com/snuvclab/coma", + "bbox": [ + 215, + 824, + 472, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "motion capture [22], and manual process [12,31,79,87] focused only on annotating contact information rather than capturing the holistic aspect of HOI. Following studies generate plausible HOI samples in 2D [37,89] or 3D [85,93] which has the potential for learning such broader HOI knowledge, but do not address the problem directly. In contrast, our method generates 3D HOI samples and extracts ComA for modelling beyond the contact.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Synthesis for Learning. There have been approaches to supplement data in fields lacking efficient annotation process by leveraging generative models, irrespective of the research area. Many methods employ GANs [18, 34] or diffusion models [27, 67, 73] to augment data [24, 78] for various vision tasks, including perception and representation learning [82], classification [8, 47, 77], segmentation [49, 80, 95], dense visual alignment [62], and further extending to 3D tasks, such as neural rendering [21, 94], shape reconstruction [60], and so forth. The major challenge in leveraging synthesized data (or generative models for cues) is controllability, as the generator, even when conditioned, typically produces free samples. Ali et al. [30] presents a method to control the viewpoint of the synthesized image by modeling the viewpoint-free latent space. CHORUS [20] applies cascaded filtering to the generated dataset to learn 3D human-object spatial relations, similar to our approach.", + "bbox": [ + 212, + 241, + 787, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Model for Synthesizing HOI Images. While diffusion models [27, 72, 73] excel at generating realistic images [64, 67], few address the generation of 2D human-object pairs (i.e., HOI images). Inpainting models [46, 67], while a common choice for human insertion, often compromise scene details, resulting in images with false affordance. Image editing techniques [26, 50, 53] may offer an alternative but tend to prioritize style changes and struggle with creating new geometry. Recently, Ye et al. [89] employs an additional layout network to determine the inpainting region, and Kulal et al. [37] trains a diffusion model on video clips to generate HOI images. In contrast to these approaches, we create HOI images by inserting humans while maintaining the context of the original image, without additional network training. Concurrent work by Li et al. [41] introduces Dynamic Mask Inpainting, which uses attention masks to change inpainting masks over timesteps to insert human into a 3D scene, while we explicitly apply segmentation model to adapt the inpainting mask.", + "bbox": [ + 212, + 441, + 789, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 696, + 331, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input 3D object mesh, our goal is to learn our novel affordance representation, ComA, which models the distribution of relative proximity and orientation between human and the object surface points. Our pipeline begins with creating 2D HOI images (Sec. 3.1). These images are then lifted to 3D, resulting large-scale synthetic 3D HOI samples (Sec. 3.2). From these 3D samples, we finally learn ComA (Sec. 3.3). An overview of our method is shown in Fig. 3.", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f875f4cc7482eba382bc3cb312decb0bf0b4096204ce340eb2c2eb9d0ad60f0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning Comprehensive Affordance from Generated 3D HOI Samples", + "text_level": 1, + "bbox": [ + 341, + 243, + 661, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/19a5cb419826cb1fc8f8e4fa4874d771501f7ae1d96f6a9a6f8ffaca6ec9dd07.jpg", + "image_caption": [ + "Fig. 3: Method Overview. Our method can be divided into two parts: (1) Generating 3D HOI Samples and (2) Learning ComA from Generated 3D HOI Samples. In the first step, we utilize an inpainting diffusion model with our Adaptive Mask Inpainting to create 2D HOI images, and generate 3D HOI samples via uplifting pipeline. In the second step, the generated 3D HOI samples are aggregated to create distributions for relative proximity and orientation, which can be derived into various affordance forms." + ], + "image_footnote": [], + "bbox": [ + 218, + 253, + 784, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 2D HOI Image Generation", + "text_level": 1, + "bbox": [ + 215, + 507, + 482, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering Object from Multi-Viewpoints. Given a 3D object, we first render it from multiple viewpoints. We place the object and camera differently for two object types: static and dynamic. For objects that are presumed to be fixed on the ground during interaction (e.g., chair), we fix the object on the ground plane, whereas objects with dynamic pose during interaction (e.g., cup) are perturbed with sampled rotation and translation. Type of the object is inferred automatically by rendering the object at any direction first and then utilize vision-language model [59] for predicting. The rendering process can be fully automated, but we empirically find it beneficial to leverage a small amount of human labor, such as restricting the range of perturbation due to the inherent bias of pre-trained diffusion models. For rendering, we place weak perspective cameras $\\{\\varPi_c\\}_{c=1}^N$ around the object with equal azimuth interval and same elevation, where $N$ is set as 8 for static, 40 for dynamic objects.", + "bbox": [ + 212, + 537, + 787, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inpainting Mask Selection. After rendering the images, we select the masks to inpaint a human interacting with the object. While our inpainting pipeline is quite robust to initial masks, we found that initial mask selection is beneficial for avoiding failure cases such as generating hallucinated objects or ambiguous interactions. Specifically, we move sliding windows around the object and choose the windows whose Intersection over Union (IoU) with the object segmentation mask falls within a specific range. See Supp. Mat. for more details.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Discovering Comprehensive Affordance for 3D Objects", + "bbox": [ + 367, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7b5047e38edee739660cfe10ce09d4b33f4fce88ba1b11be9c72e0e0bcdfb089.jpg", + "image_caption": [ + "Fig. 4: Adaptive Mask Inpainting. Without adaptive mask inpainting, the original object is damaged when inserting humans, resulting false affordances." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 781, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt Generation. Inspired by CHORUS [20], we automatically generate the prompts that describe the possible interactions with humans and the given object, which are taken as guidance for inpainting. Specifically, we utilize the vision-language model [59] to infer the HOI prompt using the rendered object image and predefined template. We generate 3 prompts per given object.", + "bbox": [ + 212, + 306, + 782, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adaptive Mask Inpainting for Human Insertion. We synthesize 2D images of human-object interaction by inpainting the human into rendered object images using publicly available text-conditional inpainting diffusion models [67]. However, the challenge arises as the object geometry and details within the mask region are not preserved during inpainting, resulting false affordance. To mitigate this problem, we present Adaptive Mask Inpainting to progressively specify the inpainting region over diffusion timesteps. Let the sequence of denoised image latent be $\\{x_{t}\\}_{t = T}^{0}$ , where $x_{T}$ is the fully random noise and $x_0$ is the denoised latent. Inspired by the observation that the quality of the predicted denoised image $\\hat{x}_0$ improves over progress of timestep, and creates a low-level structure of the target prompt (in this case, human) even at the early steps as shown in Fig. 4, we aim to adapt the inpainting region over timesteps by spatially discovering the mask from the low-level structure. Specifically, we apply PointRend [36] to decoded latent $\\mathbf{D}(\\hat{x}_0)$ at specific timesteps to predict the human mask, and consequently use as the mask for the next denoising step.", + "bbox": [ + 212, + 383, + 787, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Lifting 2D Affordance to 3D", + "text_level": 1, + "bbox": [ + 215, + 638, + 496, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From the previous section, we obtain a set of 2D HOI images $\\{I_d\\}_{d=1}^D$ . Even though we have multiple views with geometrically consistent object renderings, the inserted humans have diversity without multi-view consistency, and thus it is non-trivial to directly lift 3D humans from these 2D images.", + "bbox": [ + 212, + 670, + 787, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this, we first utilize a single view 3D human prediction model to obtain a corresponding 3D human pose and shape of the image, and then estimate the depth of the human with weak auxiliary cue provided by the set of 2D HOI images, generating a 3D HOI sample corresponds to the image.", + "bbox": [ + 212, + 732, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single View 3D Human Prediction. To uplift the 2D HOI images into 3D, we first predict the pose and shape of 3D humans from the images. We apply off-the-shelf 3D human prediction model [55] (denoted as $\\mathbf{F}_{\\mathrm{human}}$ ) to predict", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SMPL-X humans [61] from generated images:", + "bbox": [ + 214, + 146, + 545, + 162 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {\\theta} _ {d}, \\boldsymbol {\\beta} _ {d}, \\mathbf {j} _ {d}, \\boldsymbol {R} _ {d} ^ {h}, \\mathbf {s} _ {d} ^ {h}, x _ {d}, y _ {d} \\right\\} = \\mathbf {F} _ {\\text {h u m a n}} (I _ {d}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 172, + 785, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": ", where $\\pmb{\\theta}_d \\in \\mathbb{R}^{54 \\times 3}, \\pmb{\\beta}_d \\in \\mathbb{R}^{10}, \\mathbf{j}_d \\in \\mathbb{R}^{67 \\times 3}$ are the predicted SMPL-X pose, shape, and joints (see Supp. Mat. for the increased number of joints), respectively. $\\pmb{R}_d^h \\in \\mathrm{SO}(3)$ is global rotation, and the $\\mathbf{s}_d^h, x_d, y_d$ are scale, and $x, y$ direction offsets representing the parameters of weak perspective camera $\\varPi_{\\gamma(d)}$ assigned to image $I_d$ . $\\gamma(\\cdot)$ is the function that maps the index of the 2D HOI image to the index of the weak perspective camera from which it is rendered. We additionally define $\\mathbf{j}_{d^\\dagger}$ as joints in world coordinate, obtained by transforming inferred joints $\\mathbf{j}_d$ with the camera rotation and translation of $\\varPi_{\\gamma(d)}$ . As 3D human lies within the weak perspective camera framework in world coordinate, there remains depth ambiguity to solve.", + "bbox": [ + 212, + 200, + 787, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Depth Optimization using Weak Auxiliary Cue. To address the depth ambiguity of 3D human obtained from reference image, we leverage 2D HOI images generated in advance. Since the human poses in images are geometrically inconsistent, we select the largest subset inliers $\\mathcal{I}$ among them which are \"semiconsistent\" with the reference image by applying RANSAC [14] in terms of joint re-projection error. The joint triangulation is performed on two images, including the reference image. Note that we generate sufficiently large amounts of 2D HOI images to ensure the existence of the inlier set. The inliers $\\mathcal{I}$ is used to estimate the human depth $z_{d}$ by optimizing the joint re-projection loss defined below:", + "bbox": [ + 212, + 352, + 787, + 489 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e - p r o j e c t i o n}} = \\frac {1}{| \\mathcal {I} |} \\sum_ {I _ {\\delta} \\in \\mathcal {I}} \\| \\Pi_ {\\gamma (\\delta)} \\left(\\mathbf {j} _ {d ^ {\\dagger}} + z _ {d} \\boldsymbol {f} _ {d}\\right) - \\Pi_ {\\gamma (\\delta)} \\left(\\mathbf {j} _ {\\delta^ {\\dagger}}\\right) \\| ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 500, + 785, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{f}_d \\in \\mathbb{R}^3$ is the normalized direction of the camera $\\varPi_{\\gamma(d)}$ . Inspired by Jiang et al. [32], we leverage the human segmentation masks to reason about occlusion and promote a initial depth for human. We initialize SMPL-X [61] models along the camera direction and select the model with the best IOU between the rendered mask (regarding occlusion) and the segmented region to be an initial of optimization. This is useful when the optimal object position is in between two human body parts along the camera direction (e.g., a motorcycle body between the legs while riding) as we use a collision loss term during optimization.", + "bbox": [ + 212, + 542, + 787, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The collision loss is used to estimate fine-grained depth, as the collision cue can provide the physical plausibility of 3D HOI samples. We adopt collision loss from COAP [51] where the human body is represented as SDF [57] field. We set object vertices as query points to reduce the collision between human and object, namely $\\mathcal{L}_{\\mathrm{collision}}$ . We optimize depth $z_{d}$ to minimize the total loss:", + "bbox": [ + 212, + 664, + 787, + 739 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e - p r o j e c t i o n}} + \\lambda_ {\\text {c o l l i s i o n}} \\mathcal {L} _ {\\text {c o l l i s i o n}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 753, + 785, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Filtering. We filter out potentially low-quality 3D human samples in the cases of: (1) human rendering (regarding occlusion) and predicted human segmentation do not overlap much; (2) number of inliers during RANSAC [14] is small; (3) human and object collides significantly in 3D.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Discovering Comprehensive Affordance for 3D Objects", + "bbox": [ + 367, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Learning Comprehensive Affordance", + "text_level": 1, + "bbox": [ + 215, + 146, + 563, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our main objective is to learn ComA, a new affordance representation which encompasses orientational and spatial affordance, along with contact-based affordance. In order to learn ComA from the previously generated 3D HOI samples, we first apply poisson disk sampling [48] to uniformly select 3D surface points for object surface and uniformly sample mesh vertices from SMPL-X [61] human, constructing ordered point sets for each object and human.", + "bbox": [ + 212, + 170, + 787, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the ordered point sets, the density function on the $i$ -th object point and the $j$ -th human point, which we denote as $\\mathcal{P}_{ij}(\\mathbf{p},\\mathbf{n})$ , represents the joint probability of the relative position $\\mathbf{p}\\in \\mathbb{R}^3$ and relative normal orientation $\\mathbf{n}\\in \\mathbb{S}^2$ of the $j$ -th human point with respect to $i$ -th object point within the 3D HOI samples. We obtain $\\mathbf{p}$ and $\\mathbf{n}$ by canonicalizing the orientation of the human surface normal $\\mathbf{n}_j^h$ and relative position $\\mathbf{p}^{o\\rightarrow h}$ , assuming the object surface normal $\\mathbf{n}_i^o$ is rotated to face $\\hat{\\mathbf{n}} = [0,0,1]^T$ (see Supp. Mat. for details on canonicalization). The distribution $\\mathcal{P}_{ij}$ sums up to 1 for all possible $\\mathbf{p}$ and $\\mathbf{n}$ :", + "bbox": [ + 212, + 261, + 789, + 383 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\int_ {\\mathbb {R} ^ {3}} \\int_ {\\mathbb {S} ^ {2}} \\mathcal {P} _ {i j} (\\mathbf {p}, \\mathbf {n}) d \\mathbf {n} d \\mathbf {p} = 1 \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 393, + 785, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In practice, we set the domain of $\\mathbf{p}$ as voxelized grid and domain of $\\mathbf{n}$ as equispaced grid on $\\mathbb{S}^2$ , obtained via Fibonacci Spirals [16], and compute the discrete probabilities by aggregating Gaussian Kernels calculated for $\\mathbf{n}$ and $\\mathbf{p}$ (we use geodesic metrics for $\\mathbf{n}$ ). Under the probability distribution $\\mathcal{P}_{ij}(\\mathbf{p},\\mathbf{n})$ , we define three types of functions $f(\\mathbf{p},\\mathbf{n})$ capturing different aspects of affordances: (1) Contact-based Affordance, (2) Oriental Affordance, (3) Spatial Affordance. Given each function $f(\\mathbf{p},\\mathbf{n})$ , the pointwise affordance of $i$ -th object point with respect to $j$ -th human point is determined by computing its expectation:", + "bbox": [ + 212, + 433, + 787, + 555 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathbf {p}, \\mathbf {n} \\sim \\mathcal {P} _ {i j}} [ f (\\mathbf {p}, \\mathbf {n}) ] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 565, + 785, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contact-based Affordance. From ComA, we can infer the contact score for human-object point pairs regarding proximity and normal alignment as:", + "bbox": [ + 212, + 590, + 785, + 622 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {c o n t a c t}} (\\mathbf {p}, \\mathbf {n}) = \\left(\\frac {1 - \\mathbf {n} \\cdot \\hat {\\mathbf {n}}}{2}\\right) e ^ {- \\| \\mathbf {p} \\|} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 630, + 785, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $e^{-\\| \\mathbf{p}\\|}$ term encourages high score when distance $\\| \\mathbf{p}\\|$ is close. The $\\frac{1 - \\mathbf{n}\\cdot\\hat{\\mathbf{n}}}{2}$ term, motivated by the fact that physical interactions are conducted via exerted normal force between contact points, encourages high score when the object normal $\\mathbf{n}_i^o$ and human normal $\\mathbf{n}_j^h$ align in an antiparallel direction. The normal alignment term improves the precision of the contact map, especially when small body parts (e.g., fingertips) interact with an object and allows robust affordance learning even when using noisy HOI datasets, compared to prior work that only uses proximity for contact terms [2, 22, 23, 90, 93] or normal for computing only penetration [19, 86], without considering the alignment of surface normals.", + "bbox": [ + 212, + 672, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Orientational Affordance. We aim to capture the pattern and tendency of the orientation of body parts with respect to the object based on the concept", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of Shannon entropy [70]. Intuitively, low entropy means there exists more typical patterns of human body orientations during the interactions. To quantify the orientational tendency for ComA, we use negated normalized Shannon entropy [70] to enforce high value when the human surface normal shows consistent orientational tendency (low variance) with respect to the object surface normal:", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {o r i e n t a t i o n}} (\\mathbf {n}) = 1 + \\frac {\\log \\mathcal {P} _ {i j} (\\mathbf {n})}{\\log n _ {b}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 233, + 785, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $n_b$ denotes the number of discretized bins of probability measure. Intuitively, the orientational affordance defined by $\\mathbb{E}_{\\mathbf{n} \\sim \\mathcal{P}_{ij}}[f_{\\mathrm{orientation}}(\\mathbf{n})]$ has a low value when marginalized distribution $\\mathcal{P}_{ij}(\\mathbf{n})$ is close to the uniform distribution, meaning that there is no dominant orientational pattern. Note that Eq. 7 is agnostic to the proximity term $\\mathbf{p}$ , which allows us to model the nonphysical orientational effects for even far-distance points.", + "bbox": [ + 212, + 268, + 784, + 358 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Spatial Affordance. Following CHORUS [20], we also consider the 3D spatial occupancy distribution of a human surface point with respect to the object points by simply counting the occurrence:", + "bbox": [ + 212, + 359, + 782, + 404 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {s p a t i a l}} (\\mathbf {p}) = \\mathbb {1} (\\mathbf {x} - \\mathbf {p}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 415, + 784, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot):\\mathbb{R}^3\\to \\{0,1\\}$ is a binary function that returns 1 if the argument is zero vector, else 0. Thus the spatial affordance $\\mathbb{E}_{\\mathbf{p}\\sim \\mathcal{P}_{ij}}[f_{\\mathrm{spatial}}(\\mathbf{p})]$ outputs scalar occupancy for the spatial vector $\\mathbf{x}\\in \\mathbb{R}^3$ . In practice, we implement the function as voxel array, counting 1 to the voxel that contains $\\mathbf{p}$ . Note that $f_{\\mathrm{spatial}}$ is agnostic to the normal direction $\\mathbf{n}$ via marginalization. Learning the occupancy informs us about macro positioning of the human relative to the given object.", + "bbox": [ + 212, + 440, + 784, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 553, + 375, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to demonstrate the efficacy of our representation ComA, and the method for learning it. In Sec. 4.3, we illustrate how ComA extends beyond contact-based affordance to learn distributions for orientational tendency and spatial relation, by visualizing it across various object categories. In Sec. 4.4, we perform quantitative comparisons with other baselines and verify the efficacy of our Adaptive Mask Inpainting. Finally, in Sec. 4.5, we introduce an optimization framework achievable through the knowledge provided by ComA, demonstrating the potential of our new affordance representation.", + "bbox": [ + 212, + 583, + 784, + 704 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 724, + 334, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3D Object Datasets. We obtain 3D object meshes from various 3D object datasets for qualitative evaluation. Specifically, We utilize 20 from BEHAV [2], 10 from InterCap [29], 2 from ShapeNet [9], and 8 from SAPIEN [83] to learn ComA individually. As textures were not available in InterCap [29], making the diffusion model difficult to inpaint, we used TEXTure [66] to generate textures on the meshes using text prompt automatically generated from ChatGPT [58].", + "bbox": [ + 212, + 748, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Discovering Comprehensive Affordance for 3D Objects", + "bbox": [ + 367, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For quantitative evaluation, we utilize BEHAVE [2] as a ground-truth dataset to verify the quality of generated 3D HOI samples, assuming that interactions captured in lab-controlled environments exhibit the upper-bound quality of HOI. BEHAVE [2] contains video sequences of HOI scenarios for 20 object categories. Each video frame is annotated with a single 3D human and object, where each human is given as SMPL-H [44,68] format and objects as rotation and translation applied to the canonical mesh. We preprocess the dataset by transferring the SMPL-H annotations into SMPL-X format following Choutas et al. [61]. We use the 3D human-object pair in each frame as a sample, compute vertex-wise contact scores in the same way as our method, and use them as ground truth.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Internet Search Datasets. We also collect 60 free 3D object meshes from Internet source, SketchFab [71] and learn ComA.", + "bbox": [ + 212, + 297, + 787, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Baselines and Metric", + "text_level": 1, + "bbox": [ + 215, + 348, + 434, + 363 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As previous studies focused on contact rather than learning the overall aspect of affordances, we use contact-based affordance derived from ComA to compare with them. However, most of them output 3D contact of a single HOI situation by inferring affordance from a given HOI sample (mostly an image), while we produce an overall affordance distribution for the given 3D object. To address this, we treat the contact inferred by state-of-the-art (SOTA) methods, DECO [79] (for the human) and IAGNet [87] (for the object) from the HOI image as a single sample, and aggregate them across multiple images to construct a distribution for comparison against BEHAVE [2].", + "bbox": [ + 212, + 372, + 787, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Specifically, we aggregate the results of DECO [79] and IAGNet [87] on (1) 2D HOI images we've generated beforehand and (2) BEHAVE [2] test images, as these represent the usage of the BEHAVE [2] object. The aggregated scores are then normalized to form a distribution (summing up to 1), considering only the relative probability of potential contact, as the original scale of the scores varies between methods. Subsequently, we compare similarity scores (SIM) [75] and mean absolute error (MAE) [81] within these distributions.", + "bbox": [ + 212, + 508, + 787, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 633, + 419, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Contact for Various Objects. Similar to other studies, ComA can be derived into contact-based affordance as shown in Fig. 5. As our representation provides contact scores for all pairs of human-object vertices, we assign the contact score of each vertex as the maximum score among the pairs that include it. We can observe the conventional object usage patterns as heatmap, which fits with general human usages. Specifically, in case Fig. 5(b), where significant contact is absent, we can infer that the interaction does not frequently appear by contact.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Orientational Tendency for Various Objects. ComA also contains information on orientational tendencies during the HOI. The last column in Fig. 5 visualizes the negated entropy of the vertex normal direction of human relative to the object as a heatmap. Fig. 5(a), Fig. 5(b), and Fig. 5(c) exhibit relatively high orientational tendency, while the rest show lower. This aligns with the fact that", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6dc32b982e0fe503cdae927533964f58d50400ace08b13ce01fbc16e72faf083.jpg", + "image_caption": [ + "Fig. 5: Qualitative Results. ComA can model distributions of contact, orientation, and spatial relation exhibited during the interaction between humans and novel objects." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 774, + 597 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the human's orientation is fixed by the attachment to the back as in Fig. 5(a), and by the restrictions present on the back and side of the object as in Fig. 5(c). In Fig. 5(b), there exists an orientational tendency without attachments or restrictions, which arises due to long-distance interaction: viewing.", + "bbox": [ + 212, + 637, + 787, + 699 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Spatial Relation for Various Objects. The relative position of the human during HOI can be also derived from ComA. Fig. 5 visualizes the distribution of the human's full body positions relative to the object. Specifically, as shown in Fig. 5(d), Fig. 5(e), and Fig. 5(f), we find that the spatial relations for the object with multiple utilities and geometries (e.g., in the case of shovel, handle for holding and blade for digging) align with general human usage, demonstrating the plausibility of the results. Especially, in the case of tennis racket shown in Fig. 5(e), we can observe a relatively weak tendency in spatial relation, as the racket can move almost freely relative to humans during the interaction.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Discovering Comprehensive Affordance for 3D Objects", + "bbox": [ + 367, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f0661fbc5771ff412d48f6b6b69b41f374393cb90807d56cbb6d0b2e242e587f.jpg", + "image_caption": [ + "(a) Finding Corresponding Contact" + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 504, + 244 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2011eecb84d85428aac33a08a2e8ee6fc77b994aa1d0b79b0cd28930a2b3143c.jpg", + "image_caption": [ + "(b) Evaluating Plausibility of Orientation", + "Fig. 6: Pointwise Affordance Knowledge. ComA can provide pointwise affordance knowledge such as corresponding contacts or plausible orientation of body parts." + ], + "image_footnote": [], + "bbox": [ + 509, + 140, + 754, + 244 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e6159081e78d4bef11ed2f01b8b90d8c6e99e339e114e82ac4939c44a7282d53.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSIMHuman(%)↑SIMObject(%)↑MAEHuman(×100)↓MAEObject(×100)↓
Ours52.8270.550.1260.0288
IAGNet [87] on Our 2D HOI Images-66.34-0.0329
IAGNet [87] on BEHAVECT Images-63.86-0.0353
DECO [79] on Our 2D HOI Images21.66-0.246-
DECO [79] on BEHAVECT Images17.39-0.259-
", + "bbox": [ + 230, + 300, + 767, + 378 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 1: Quantitative Results on BEHAV. We report SIM, MAE to evaluate the similarity of contact distribution (normalized contact scores) on BEHAV sequences.", + "bbox": [ + 214, + 383, + 784, + 411 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Pointwise Affordance Knowledge. Different from previous studies, ComA models the distribution of relative orientation and proximity for all human-object vertex pairs, allowing it to provide pointwise affordance knowledge such as corresponding contacts and plausible orientation direction of specific body parts. As shown in Fig. 6(a), we can infer how each human parts and object parts contact. For the case of motorcycle, we can see that saddle and handle each contacts with human's hips and hands; while for the case of chair, the seat and backrest each contacts with human's hips and back. Also, we can utilize ComA to infer the probable direction of the head while using the monitor by computing the cosine similarity between the head direction and the vertex normal with highest orientational tendency. As illustrated in Fig. 6(b), we can see that the similarity maximizes when gazing towards the screen; which is a plausible head direction when interacting with the monitor.", + "bbox": [ + 212, + 417, + 785, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation on Adaptive Mask Inpainting. As shown in Fig. 4, without Adaptive Mask Inpainting, the original geometry of the object changes (left) or gets replaced by hallucinated objects (right), resulting in false affordances unfaithful to the original object. Consequently, the difference between the object in 2D and the original one existing in 3D causes the uplifted 3D HOI samples to fail to represent the right usage. This shows the importance of Adaptive Mask Inpainting as it prevents 2D inpainting diffusion model from generating false affordances.", + "bbox": [ + 212, + 614, + 787, + 720 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Quantitative Results", + "text_level": 1, + "bbox": [ + 214, + 739, + 433, + 755 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with Baselines. We evaluate the similarity of contact distribution between BEHAV [2] and each of DECO [79], IAGNet [87], and ours. We sample 2048 points from the 3D object mesh surface following IAGNet [87] and only consider the contact underneath the head for DECO [79], as DECO [79] uses SMPL [44] format, which only has vertex correspondence with SMPL-X [61] in", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3a3fa07afa08dd59ab24dca80b68695d8429e885329774f4931c3663adfe7290.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Inpainting MethodsRMSEBackground↓ mIoU↑mIoUOCclusionAware ↑
Vanilla42.230.5350.631
Adaptive Mask15.330.7060.815
", + "bbox": [ + 316, + 143, + 679, + 188 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 2: Effects of Adaptive Mask Inpainting We evaluate the performance of background preservation in vanilla inpainting and our Adaptive Mask Inpainting.", + "bbox": [ + 215, + 193, + 785, + 220 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "non-head regions. As shown in Tab. 1, the contact distribution from IAGNet [87] and DECO [79] using our 2D HOI images outperforms the one modeled from BEHAVE test images. This means that our generated HOI images represent the real-world affordance as well as BEHAVE [2] test images, demonstrating the efficacy of leveraging pre-trained diffusion model for generating affordance. Additionally, the contact distribution from our method outperforms the one modeled using IAGNet [87] and DECO [79] with our generated HOI images. Since we extract 3D contacts from the same image sets, we argue that our pipeline is better than other models trained on annotations for uplifting 2D affordance to 3D.", + "bbox": [ + 212, + 229, + 784, + 366 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Efficacy of Adaptive Mask Inpainting. To verify the effectiveness of our Adaptive Mask Inpainting, we compare the amount of background preservation for our generated 2D HOI images with the vanilla inpainting diffusion model. We sample a maximum of 2 inpainted images with a single human detected, per each object categories from BEHAV [2]; resulting in 691 images for vanilla inpainting and 692 images for Adaptive Mask Inpainting. We compute 3 metrics for these images; (1) $\\mathrm{RMSE}_{\\mathrm{Background}}$ : pixel error between images excluding the predicted human region; (2) mIoU: average IoU between object predictions in each inpainted image and original object image; (3) mIoU $_{\\mathrm{Occlusion\\_Aware}}$ : same metric as (2) but excluding the predicted human region during computation. To ensure fairness, we use open vocabulary segmentation model [35,43] to predict detailed human/object segmentation. As shown in Tab. 2, we demonstrate that our adaptive mask algorithm better preserves the background, not only pixelwise $(\\mathrm{RMSE}_{\\mathrm{Background}})$ , but also semantic-region-wise (mIOU, mIOU $_{\\mathrm{Occlusion\\_Aware}}$ ) than the vanilla inpainting baseline. This means that the generated images from adaptive mask inpainting are more likely to preserve the original object.", + "bbox": [ + 212, + 367, + 785, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Application", + "text_level": 1, + "bbox": [ + 215, + 632, + 357, + 647 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As ComA is a newly proposed affordance representation, we demonstrate its potential by showcasing an application. One possible task is to reconstruct HOI by optimizing the human pose and position to interact with an object. As shown in Fig. 7(a), it is nearly impossible to achieve the task with traditional methods which relies on contact affordance only. In contrast, ComA provides both relative orientation and contact during the HOI, making it possible to generate plausible 3D HOI samples from T-posed humans. For the objects with multiple modes of affordances, we apply mode seeking algorithm [11] on 3D HOI samples with their vertex normals and learn ComA separately; for example, swing in Fig. 7(a).", + "bbox": [ + 212, + 657, + 785, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "However, ComA learned on a specific object cannot be directly applied to other objects with different geometries. To address this, we transfer ComA through the process of (1) object canonicalization and (2) finding vertex cor-", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Discovering Comprehensive Affordance for 3D Objects", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/52af60b288a31d15ff981a926274ef539f88e6bf5316891784d2e77fed48469d.jpg", + "image_caption": [ + "Fig. 7: Application. ComA can reconstruct HOI through optimization, which can be scaled to other objects in the same category by transferring such knowledge between them." + ], + "image_footnote": [], + "bbox": [ + 256, + 141, + 738, + 435 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "respondences. Fig. 7(b) provides an example of transferring ComA learned from a specific motorcycle to other motorcycles with different geometries. We use objects from ShapeNet [9] which are already canonicalized and apply CPNet [45] to find vertex correspondences automatically for transferring the vertices with contact. As object canonicalization and vertex correspondence can each transfer macro orientation information and local contact information, we can transfer ComA and scale the human optimization task to other objects in the same category without any additional learning.", + "bbox": [ + 212, + 473, + 787, + 594 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 614, + 359, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we point out the importance of learning non-contact patterns in HOIs and propose a novel affordance representation called ComA which models both contact and non-contact patterns. To learn ComA, we utilize a pre-trained 2D diffusion model which already possesses prior knowledge of affordance. To leverage the affordance knowledge, we designed a Rendering-Inpainting-Uplifting pipeline to generate 3D HOI samples and learn ComA from the samples. The results show that ComA effectively models the common usage of objects in terms of contact, orientational, and spatial relations. Within the pipeline, we propose Adaptive Mask Inpainting, which allows insertion of human without damaging the original image, where we evaluate its efficacy both qualitatively and quantitatively. We expect ComA to be used as a prior knowledge for various downstream tasks, where we demonstrate its potential through human optimization framework and the possibility of transfer to various objects.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by Naver Webtoon. The work of SNU members was also supported by NRF grant funded by the Korean government (MSIT) (No. 2022R1A2C2092724 and No. RS-2023-00218601), and IITP grant funded by the Korean government (MSIT) [RS-2021-II211343, AI Graduate School Program (SNU)]. H. Joo is the corresponding author.", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 277, + 323, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bahl, S., Mendonca, R., Chen, L., Jain, U., Pathak, D.: Affordances from human videos as a versatile representation for robotics. In: CVPR (2023)", + "2. Bhatnagar, B.L., Xie, X., Petrov, I., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Behave: Dataset and method for tracking human object interactions. In: CVPR (2022)", + "3. Brahmbhatt, S., Ham, C., Kemp, C.C., Hays, J.: Contactdb: Analyzing and predicting grasp contact via thermal imaging. In: CVPR (2019)", + "4. Brahmbhatt, S., Handa, A., Hays, J., Fox, D.: Contactgrasp: Functional multifinger grasp synthesis from contact. In: IROS (2019)", + "5. Brahmbhatt, S., Tang, C., Twigg, C.D., Kemp, C.C., Hays, J.: Contactpose: A dataset of grasps with object contact and hand pose. In: ECCV (2020)", + "6. Cai, M., Kitani, K.M., Sato, Y.: Understanding hand-object manipulation with grasp types and object attributes. In: RSS (2016)", + "7. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV (2020)", + "8. Chai, L., Zhu, J.Y., Shechtman, E., Isola, P., Zhang, R.: Ensembling with deep generative views. In: CVPR (2021)", + "9. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 (2015)", + "10. Chen, Y., Huang, S., Yuan, T., Qi, S., Zhu, Y., Zhu, S.C.: Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In: ICCV (2019)", + "11. Comaniciu, D., Meer, P.: Mean shift: a robust approach toward feature space analysis. In: IEEE TPAMI (2002)", + "12. Deng, S., Xu, X., Wu, C., Chen, K., Jia, K.: 3d affordancenet: A benchmark for visual object affordance understanding. In: CVPR (2021)", + "13. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: CVPR (2020)", + "14. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. In: CACM (1981)", + "15. Gao, W., Tedrake, R.: kpm-sc: Generalizable manipulation planning using keypoint affordance and shape completion. In: ICRA (2021)", + "16. Garg, M., Garg, P., Vohra, R.: Advanced fibonacci sequence with golden ratio. In: IJSER (2014)", + "17. Gibson, J.J.: The Ecological Approach to Visual Perception. Houghton Mifflin (1979)" + ], + "bbox": [ + 218, + 310, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Discovering Comprehensive Affordance for 3D Objects", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)", + "19. Grady, P., Tang, C., Twigg, C.D., Vo, M., Brahmbhatt, S., Kemp, C.C.: ContactOpt: Optimizing contact to improve grasps. In: CVPR (2021)", + "20. Han, S., Joo, H.: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In: ICCV (2023)", + "21. Hao, Z., Mallya, A., Belongie, S., Liu, M.Y.: Gancraft: Unsupervised 3d neural rendering of apache worlds. In: ICCV (2021)", + "22. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV (2019)", + "23. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: CVPR (2021)", + "24. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? In: arXiv:2210.07574 (2022)", + "25. Hermans, T., Rehg, J.M., Bobick, A.: Affordance prediction via learned object attributes. In: ICRA (2011)", + "26. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: arXiv:2208.01626 (2022)", + "27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "28. Huang, C.H.P., Yi, H., Hoschle, M., Safroshkin, M., Alexiadis, T., Polikovsky, S., Scharstein, D., Black, M.J.: Capturing and inferring dense full-body human-scene contact. In: CVPR (2022)", + "29. Huang, Y., Taheri, O., Black, M.J., Tzionas, D.: Intercap: Joint markerless 3d tracking of humans and objects in interaction from multi-view rgb-d images. In: IJCV (2024)", + "30. Jahanian, A., Puig, X., Tian, Y., Isola, P.: Generative models as a data source for multiview representation learning. In: arXiv:2106.05258 (2021)", + "31. Jian, J., Liu, X., Li, M., Hu, R., Liu, J.: Affordpose: A large-scale dataset of handobject interactions with affordance-driven hand pose. In: ICCV (2023)", + "32. Jiang, W., Kolotouros, N., Pavlakos, G., Zhou, X., Daniilidis, K.: Coherent reconstruction of multiple humans from a single image. In: CVPR (2020)", + "33. Jiang, Y., Jiang, S., Sun, G., Su, Z., Guo, K., Wu, M., Yu, J., Xu, L.: Neuralhofusion: Neural volumetric rendering under human-object interactions. In: CVPR (2022)", + "34. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)", + "35. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023)", + "36. Kirillov, A., Wu, Y., He, K., Girshick, R.: Pointrend: Image segmentation as rendering. In: CVPR (2020)", + "37. Kulal, S., Brooks, T., Aiken, A., Wu, J., Yang, J., Lu, J., Efros, A.A., Singh, K.K.: Putting people in their place: Affordance-aware human insertion into scenes. In: CVPR (2023)", + "38. Lee, Y.J., Grauman, K.: Predicting important objects for egocentric video summarization. In: IJCV (2015)", + "39. Levine, S., Shah, D.: Learning robotic navigation from experience: principles, methods and recent results. In: Philos. Trans. R. Soc. B (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Li, G., Jampani, V., Sun, D., Sevilla-Lara, L.: Locate: Localize and transfer object parts for weakly supervised affordance grounding. In: CVPR (2023)", + "41. Li, L., Dai, A.: GenZI: Zero-shot 3D human-scene interaction generation. In: CVPR (2024)", + "42. Li, Z., Sedlar, J., Carpentier, J., Laptev, I., Mansard, N., Sivic, J.: Estimating 3d motion and forces of person-object interactions from monocular video. In: CVPR (2019)", + "43. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: arXiv:2303.05499 (2023)", + "44. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. In: Proc. ACM SIGGRAPH Asia (2015)", + "45. Lou, Y., You, Y., Li, C., Cheng, Z., Li, L., Ma, L., Wang, W., Lu, C.: Human correspondence consensus for 3d object semantic understanding. In: ECCV (2020)", + "46. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)", + "47. Mao, C., Cha, A., Gupta, A., Wang, H., Yang, J., Vondrick, C.: Generative interventions for causal learning. In: CVPR (2021)", + "48. McCool, M., Fiume, E.: Hierarchical poisson disk sampling distributions. In: Graphics Interface (1992)", + "49. Melas-Kyriazi, L., Rupprecht, C., Laina, I., Vedaldi, A.: Finding an unsupervised image segmenter in each of your deep generative models. In: arXiv:2105.08127 (2021)", + "50. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)", + "51. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: COAP: Compositional articulated occupancy of people. In: CVPR (2022)", + "52. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019)", + "53. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)", + "54. Monszpart, A., Guerrero, P., Ceylan, D., Yumer, E., Mitra, N.J.: imapper: interaction-guided scene mapping from monocular videos. In: ACM TOG (2019)", + "55. Moon, G., Choi, H., Lee, K.M.: Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In: CVPRW (2022)", + "56. Nagarajan, T., Grauman, K.: Learning affordance landscapes for interaction exploration in 3d environments. In: NeurIPS (2020)", + "57. Oleynikova, H., Millane, A., Taylor, Z., Galceran, E., Nieto, J., Siegwart, R.: Signed distance fields: A natural representation for both mapping and planning. In: RSS Workshop (2016)", + "58. OpenAI: Chatgpt: Optimizing language models for dialogue. https://openai.com/blog/chatgpt/ (2023)", + "59. OpenAI: Gpt-4v(ision) system card. https://openai.com/research/gpt-4v-system-card (2023)", + "60. Pan, X., Dai, B., Liu, Z., Loy, C.C., Luo, P.: Do 2d gans know 3d shape? unsupervised 3d shape reconstruction from 2d image gans. In: arXiv:2011.00844 (2020)", + "61. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Discovering Comprehensive Affordance for 3D Objects", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Peebles, W., Zhu, J.Y., Zhang, R., Torralba, A., Efros, A.A., Shechtman, E.: Gansupervised dense visual alignment. In: CVPR (2022)", + "63. Petrov, I.A., Marin, R., Chibane, J., Pons-Moll, G.: Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In: CVPR (2023)", + "64. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: arXiv:2204.06125 (2022)", + "65. Rempe, D., Guibas, L.J., Hertzmann, A., Russell, B., Villegas, R., Yang, J.: Contact and human dynamics from monocular video. In: ECCV (2020)", + "66. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. In: Proc. ACM SIGGRAPH (2023)", + "67. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "68. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. In: Proc. ACM SIGGRAPH Asia (2017)", + "69. Savva, M., Chang, A.X., Hanrahan, P., Fisher, M., Nießner, M.: Pigraphs: learning interaction snapshots from observations. In: ACM TOG (2016)", + "70. Shannon, C.E.: A mathematical theory of communication. In: ACM SIGMOBILE (2001)", + "71. Sketchfab: Sketchfab. https://sketchfab.com/ (2023)", + "72. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)", + "73. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: arXiv:2010.02502 (2020)", + "74. Sun, G., Chen, X., Chen, Y., Pang, A., Lin, P., Jiang, Y., Xu, L., Yu, J., Wang, J.: Neural free-viewpoint performance rendering under complex human-object interactions. In: ACM MM (2021)", + "75. Swain, M.J., Ballard, D.H.: Color indexing. In: IJCV (1991)", + "76. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: ECCV (2020)", + "77. Tanaka, F.H.K.d.S., Aranha, C.: Data augmentation using gans. In: arXiv:1904.09135 (2019)", + "78. Trabucco, B., Doherty, K., Gurinas, M., Salakhutdinov, R.: Effective data augmentation with diffusion models. In: arXiv:2302.07944 (2023)", + "79. Tripathi, S., Chatterjee, A., Passy, J.C., Yi, H., Tzionas, D., Black, M.J.: DECO: Dense estimation of 3D human-scene contact in the wild. In: ICCV (2023)", + "80. Tritrong, N., Rewatbowornwong, P., Suwajanakorn, S.: Repurposing gans for one-shot semantic part segmentation. In: CVPR (2021)", + "81. Willmott, C.J., Matsuura, K.: Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. In: Climate Research (2005)", + "82. Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetdm: Synthesizing data with perception annotations using diffusion models. In: NeurIPS (2023)", + "83. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: CVPR (2020)", + "84. Xu, J., Zhu, S., Guo, H., Wu, S.: Automated labeling for robotic autonomous navigation through multi-sensory semi-supervised learning on big data. In: IEEE TBD (2021)", + "85. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In: ICCV (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "86. Yang, L., Zhan, X., Li, K., Xu, W., Li, J., Lu, C.: CPF: Learning a contact potential field to model the hand-object interaction. In: ICCV (2021)", + "87. Yang, Y., Zhai, W., Luo, H., Cao, Y., Luo, J., Zha, Z.J.: Grounding 3d object affordance from 2d interactions in images. In: ICCV (2023)", + "88. Yang, Y., Zhai, W., Luo, H., Cao, Y., Zha, Z.J.: Lemon: Learning 3d human-object interaction relation from 2d images. In: CVPR (2024)", + "89. Ye, Y., Li, X., Gupta, A., Mello, S.D., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: CVPR (2023)", + "90. Yi, H., Huang, C.H.P., Tzionas, D., Kocabas, M., Hassan, M., Tang, S., Thies, J., Black, M.J.: Human-aware object placement for visual environment reconstruction. In: CVPR (2022)", + "91. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 3DV (2020)", + "92. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: ECCV (2022)", + "93. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: CVPR (2020)", + "94. Zhang, Y., Chen, W., Ling, H., Gao, J., Zhang, Y., Torralba, A., Fidler, S.: Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In: ICLR (2021)", + "95. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021)", + "96. Zhou, K., Bhatnagar, B.L., Lenssen, J.E., Pons-Moll, G.: Toch: Spatio-temporal object correspondence to hand for motion refinement. In: ECCV (2022)" + ], + "bbox": [ + 212, + 146, + 784, + 493 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Discovering Comprehensive Affordance for 3D Objects", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/4cf86d6f-6634-4c59-b4d4-c22212431643_model.json b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/4cf86d6f-6634-4c59-b4d4-c22212431643_model.json new file mode 100644 index 0000000000000000000000000000000000000000..89c37cf6292f0234318156e671adced04f11ecd3 --- /dev/null +++ b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/4cf86d6f-6634-4c59-b4d4-c22212431643_model.json @@ -0,0 +1,2779 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.207 + ], + "angle": 0, + "content": "Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.234, + 0.751, + 0.252 + ], + "angle": 0, + "content": "Hyeonwoo \\(\\mathrm{Kim}^{1*}\\), Sookwan Han\\(^{1*}\\), Patrick Kwon\\(^{2}\\), and Hanbyul Joo\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.408, + 0.262, + 0.596, + 0.277 + ], + "angle": 0, + "content": "1 Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.278, + 0.573, + 0.29 + ], + "angle": 0, + "content": "2 Naver Webtoon AI" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.3, + 0.789, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.421, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Fig. 1: Given a 3D object, we generate numerous 3D Human-Object Interaction (HOI) samples using text prompts, and learn a novel affordance representation called Comprehensive Affordance (ComA) which models both contact and non-contact HOI patterns." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.475, + 0.741, + 0.794 + ], + "angle": 0, + "content": "Abstract. Understanding the inherent human knowledge in interacting with a given environment (e.g., affordance) is essential for improving AI to better assist humans. While existing approaches primarily focus on human-object contacts during interactions, such affordance representation cannot fully address other important aspects of human-object interactions (HOIs), i.e. patterns of relative positions and orientations. In this paper, we introduce a novel affordance representation, named Comprehensive Affordance (ComA). Given a 3D object mesh, ComA models the distribution of relative orientation and proximity of vertices in interacting human meshes, capturing plausible patterns of contact, relative orientations, and spatial relationships. To construct the distribution, we present a novel pipeline that synthesizes diverse and realistic 3D HOI samples given any 3D target object mesh. The pipeline leverages a pre-trained 2D inpainting diffusion model to generate HOI images from object renderings and lifts them into 3D. To avoid the generation of false affordances, we propose a new inpainting framework, Adaptive Mask Inpainting. Since ComA is built on synthetic samples, it can extend to any object in an unbounded manner. Through extensive experiments, we demonstrate that ComA outperforms competitors that rely on human annotations in modeling contact-based affordance. Importantly, we also showcase the potential of ComA to reconstruct human-object interactions in 3D through an optimization framework, highlighting its advantage in incorporating both contact and non-contact properties." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.806, + 0.741, + 0.821 + ], + "angle": 0, + "content": "Keywords: Affordance \\(\\cdot\\) Human-Object Interaction \\(\\cdot\\) Diffusion Models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.825, + 0.414, + 0.84 + ], + "angle": 0, + "content": "*Indicates equal contribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.186, + 0.787, + 0.352 + ], + "angle": 0, + "content": "Humans possess an innate ability to perceive the functionalities provided by the environment or objects and efficiently utilize them—a concept known as affordance [17]. Affordance incorporates a variety of patterns typically observable in human-object interactions (HOIs), including not only plausible physical contact but also spatial and orientational non-contact relationships. For example, while using a laptop, the orientation and distance patterns between human face or other body parts and the screen can be considered part of affordance, alongside the hand regions touching the laptop keyboard. Teaching such affordance knowledge to machines is crucial for enabling them to better understand or mimic human behaviors. Consequently, numerous studies have been conducted to implement such affordance knowledge into AI and robotics [1,20,23,37,40,63,79,87,89]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.787, + 0.475 + ], + "angle": 0, + "content": "Most previous approaches however, primarily focus on a limited spectrum of affordances. Some approaches represent the human and object regions in contact using contact scores on 2D images [1,40], 3D objects [87], and 3D humans [23,79] using direct supervision from human annotations. Others represent affordances by inferring [63] or generating [37,89] plausible human-object pairs in contact-oriented object categories. While a few recent studies tackle learning the spatial relations [20,88] between a human and an object, these methods mainly focus on object categories where physical contact is dominant than non-contact patterns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.787, + 0.643 + ], + "angle": 0, + "content": "In this paper, we introduce Comprehensive Affordance (ComA), a novel representation of an affordance encompassing the various aspects of human-object interaction, including orientational tendencies and relative positions. The motivation for our design arises from the observation that there exist typical 3D relations between entire body parts and objects during HOIs. For instance, human faces, torsos, and arms often exhibit specific distances and orientations, with some variability, when interacting with certain object categories, as shown in Fig. 2. Our ComA is designed to model these patterns by constructing a distribution between each pair of object surface points and human surface points in 3D. This distribution represents the likelihood of human body parts displaying specific spatial and orientational relations with object points, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Learning ComA requires various 3D HOI samples showing the way humans use the target object in 3D, from which pairwise distributions of 3D HOI relations can be constructed. Manually annotating such cues is challenging, compared to the annotations focusing solely on contact regions [40, 63, 79, 85, 87, 88]. As a key solution, we present a pipeline to synthesize a large-scale synthetic 3D HOI samples leveraging a pre-trained 2D diffusion model. Interestingly, we observe that a pre-trained 2D diffusion model captures the affordance knowledge in the form of 2D images, as shown in Fig. 2, where the images are synthesized from text prompts. However, leveraging these 2D HOI cues for construction of ComA (which requires 3D HOI samples) presents the following challenges: (1) the diffusion model needs to be applied for inpainting plausible humans without altering the original shape and pose of the target object, and (2) 3D HOI samples should be lifted from the synthesized 2D images. To address these challenges, we present" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Discovering Comprehensive Affordance for 3D Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.499, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.256, + 0.466, + 0.269 + ], + "angle": 0, + "content": "Text Prompt: 1 person using a laptop" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.144, + 0.788, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.256, + 0.761, + 0.269 + ], + "angle": 0, + "content": "Text Prompt: 1 person using a telescope" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.785, + 0.331 + ], + "angle": 0, + "content": "Fig. 2: Typically, people (1) view the screen (2) from relatively distant distance while using a laptop (Left), whereas they (1) peer into it (2) from a close distance while using a telescope (Right). Pre-trained diffusion model has a knowledge of these (1) orientational and (2) spatial relation between human and object during interaction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.371 + ], + "angle": 0, + "content": "Adaptive Mask Inpainting to insert humans without altering the appearance of target objects in 2D, along with a 3D lifting pipeline from 2D HOI cues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.785, + 0.49 + ], + "angle": 0, + "content": "We demonstrate the efficacy of our approach by learning ComA on 100 3D object meshes in diverse categories collected from various sources [2,9,29,52,71, 83]. We compare our ComA with existing approaches [79,87] on the BEHAV dataset [2], demonstrating that ours learned from synthetic cues outperforms competitors trained on manual annotations for contact-based affordance. Additionally, we verify the advantage of ComA against contact-only representations, in the scenario of reconstructing HOIs via an optimization framework. We will release the results and our source code1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.598 + ], + "angle": 0, + "content": "In summary, our main contributions can be summarized as follows: (1) we introduce a new representation ComA, which models the distribution of both explicit contact and non-contact patterns during HOIs; (2) we present a pipeline to build ComA from synthetic datasets by leveraging a pre-trained 2D diffusion model, scalable to any 3D object without requiring laborious annotations; and (3) we propose Adaptive Mask Inpainting, a method for preserving the original context of an image in using inpainting diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.388, + 0.636 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Learning Visual Affordance. Affordance was introduced by Gibson [17] as a set of functionalities that the environment furnishes to an agent (e.g., human, animal). The concept extends to the vision and robotics area, where the focus is on teaching embodied agents to interact with scenes [1, 15, 39, 56, 84], mimicking human-object [2, 7, 10, 13, 23, 28, 33, 42, 54, 63, 65, 69, 74, 91, 92], or hand-object [3-5, 76, 96] interactions. Earlier methods focused on learning action category labels [6, 25, 38] with bounding boxes but lacked a detailed description of affordance. Due to the limited affordance information available from text descriptions or labels, some studies depicted affordance as contact regions or heatmaps on objects [1, 40, 87] and humans [23, 79]. However, learning affordances beyond the contact was challenging since existing datasets created through a multi-camera system [2, 29]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.825, + 0.473, + 0.84 + ], + "angle": 0, + "content": "1https://github.com/snuvclab/coma" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.129 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "motion capture [22], and manual process [12,31,79,87] focused only on annotating contact information rather than capturing the holistic aspect of HOI. Following studies generate plausible HOI samples in 2D [37,89] or 3D [85,93] which has the potential for learning such broader HOI knowledge, but do not address the problem directly. In contrast, our method generates 3D HOI samples and extracts ComA for modelling beyond the contact." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.439 + ], + "angle": 0, + "content": "Data Synthesis for Learning. There have been approaches to supplement data in fields lacking efficient annotation process by leveraging generative models, irrespective of the research area. Many methods employ GANs [18, 34] or diffusion models [27, 67, 73] to augment data [24, 78] for various vision tasks, including perception and representation learning [82], classification [8, 47, 77], segmentation [49, 80, 95], dense visual alignment [62], and further extending to 3D tasks, such as neural rendering [21, 94], shape reconstruction [60], and so forth. The major challenge in leveraging synthesized data (or generative models for cues) is controllability, as the generator, even when conditioned, typically produces free samples. Ali et al. [30] presents a method to control the viewpoint of the synthesized image by modeling the viewpoint-free latent space. CHORUS [20] applies cascaded filtering to the generated dataset to learn 3D human-object spatial relations, similar to our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.79, + 0.655 + ], + "angle": 0, + "content": "Diffusion Model for Synthesizing HOI Images. While diffusion models [27, 72, 73] excel at generating realistic images [64, 67], few address the generation of 2D human-object pairs (i.e., HOI images). Inpainting models [46, 67], while a common choice for human insertion, often compromise scene details, resulting in images with false affordance. Image editing techniques [26, 50, 53] may offer an alternative but tend to prioritize style changes and struggle with creating new geometry. Recently, Ye et al. [89] employs an additional layout network to determine the inpainting region, and Kulal et al. [37] trains a diffusion model on video clips to generate HOI images. In contrast to these approaches, we create HOI images by inserting humans while maintaining the context of the original image, without additional network training. Concurrent work by Li et al. [41] introduces Dynamic Mask Inpainting, which uses attention masks to change inpainting masks over timesteps to insert human into a 3D scene, while we explicitly apply segmentation model to adapt the inpainting mask." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.697, + 0.333, + 0.714 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Given an input 3D object mesh, our goal is to learn our novel affordance representation, ComA, which models the distribution of relative proximity and orientation between human and the object surface points. Our pipeline begins with creating 2D HOI images (Sec. 3.1). These images are then lifted to 3D, resulting large-scale synthetic 3D HOI samples (Sec. 3.2). From these 3D samples, we finally learn ComA (Sec. 3.3). An overview of our method is shown in Fig. 3." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Discovering Comprehensive Affordance for 3D Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.785, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.342, + 0.244, + 0.662, + 0.255 + ], + "angle": 0, + "content": "Learning Comprehensive Affordance from Generated 3D HOI Samples" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.255, + 0.785, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Fig. 3: Method Overview. Our method can be divided into two parts: (1) Generating 3D HOI Samples and (2) Learning ComA from Generated 3D HOI Samples. In the first step, we utilize an inpainting diffusion model with our Adaptive Mask Inpainting to create 2D HOI images, and generate 3D HOI samples via uplifting pipeline. In the second step, the generated 3D HOI samples are aggregated to create distributions for relative proximity and orientation, which can be derived into various affordance forms." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.508, + 0.483, + 0.525 + ], + "angle": 0, + "content": "3.1 2D HOI Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Rendering Object from Multi-Viewpoints. Given a 3D object, we first render it from multiple viewpoints. We place the object and camera differently for two object types: static and dynamic. For objects that are presumed to be fixed on the ground during interaction (e.g., chair), we fix the object on the ground plane, whereas objects with dynamic pose during interaction (e.g., cup) are perturbed with sampled rotation and translation. Type of the object is inferred automatically by rendering the object at any direction first and then utilize vision-language model [59] for predicting. The rendering process can be fully automated, but we empirically find it beneficial to leverage a small amount of human labor, such as restricting the range of perturbation due to the inherent bias of pre-trained diffusion models. For rendering, we place weak perspective cameras \\(\\{\\varPi_c\\}_{c=1}^N\\) around the object with equal azimuth interval and same elevation, where \\(N\\) is set as 8 for static, 40 for dynamic objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Inpainting Mask Selection. After rendering the images, we select the masks to inpaint a human interacting with the object. While our inpainting pipeline is quite robust to initial masks, we found that initial mask selection is beneficial for avoiding failure cases such as generating hallucinated objects or ambiguous interactions. Specifically, we move sliding windows around the object and choose the windows whose Intersection over Union (IoU) with the object segmentation mask falls within a specific range. See Supp. Mat. for more details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.782, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.266, + 0.788, + 0.296 + ], + "angle": 0, + "content": "Fig. 4: Adaptive Mask Inpainting. Without adaptive mask inpainting, the original object is damaged when inserting humans, resulting false affordances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.784, + 0.384 + ], + "angle": 0, + "content": "Prompt Generation. Inspired by CHORUS [20], we automatically generate the prompts that describe the possible interactions with humans and the given object, which are taken as guidance for inpainting. Specifically, we utilize the vision-language model [59] to infer the HOI prompt using the rendered object image and predefined template. We generate 3 prompts per given object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.789, + 0.613 + ], + "angle": 0, + "content": "Adaptive Mask Inpainting for Human Insertion. We synthesize 2D images of human-object interaction by inpainting the human into rendered object images using publicly available text-conditional inpainting diffusion models [67]. However, the challenge arises as the object geometry and details within the mask region are not preserved during inpainting, resulting false affordance. To mitigate this problem, we present Adaptive Mask Inpainting to progressively specify the inpainting region over diffusion timesteps. Let the sequence of denoised image latent be \\(\\{x_{t}\\}_{t = T}^{0}\\), where \\(x_{T}\\) is the fully random noise and \\(x_0\\) is the denoised latent. Inspired by the observation that the quality of the predicted denoised image \\(\\hat{x}_0\\) improves over progress of timestep, and creates a low-level structure of the target prompt (in this case, human) even at the early steps as shown in Fig. 4, we aim to adapt the inpainting region over timesteps by spatially discovering the mask from the low-level structure. Specifically, we apply PointRend [36] to decoded latent \\(\\mathbf{D}(\\hat{x}_0)\\) at specific timesteps to predict the human mask, and consequently use as the mask for the next denoising step." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.64, + 0.498, + 0.656 + ], + "angle": 0, + "content": "3.2 Lifting 2D Affordance to 3D" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.732 + ], + "angle": 0, + "content": "From the previous section, we obtain a set of 2D HOI images \\(\\{I_d\\}_{d=1}^D\\). Even though we have multiple views with geometrically consistent object renderings, the inserted humans have diversity without multi-view consistency, and thus it is non-trivial to directly lift 3D humans from these 2D images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.794 + ], + "angle": 0, + "content": "To address this, we first utilize a single view 3D human prediction model to obtain a corresponding 3D human pose and shape of the image, and then estimate the depth of the human with weak auxiliary cue provided by the set of 2D HOI images, generating a 3D HOI sample corresponds to the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Single View 3D Human Prediction. To uplift the 2D HOI images into 3D, we first predict the pose and shape of 3D humans from the images. We apply off-the-shelf 3D human prediction model [55] (denoted as \\(\\mathbf{F}_{\\mathrm{human}}\\)) to predict" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Discovering Comprehensive Affordance for 3D Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.547, + 0.163 + ], + "angle": 0, + "content": "SMPL-X humans [61] from generated images:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.174, + 0.786, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {\\theta} _ {d}, \\boldsymbol {\\beta} _ {d}, \\mathbf {j} _ {d}, \\boldsymbol {R} _ {d} ^ {h}, \\mathbf {s} _ {d} ^ {h}, x _ {d}, y _ {d} \\right\\} = \\mathbf {F} _ {\\text {h u m a n}} (I _ {d}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.201, + 0.788, + 0.353 + ], + "angle": 0, + "content": ", where \\(\\pmb{\\theta}_d \\in \\mathbb{R}^{54 \\times 3}, \\pmb{\\beta}_d \\in \\mathbb{R}^{10}, \\mathbf{j}_d \\in \\mathbb{R}^{67 \\times 3}\\) are the predicted SMPL-X pose, shape, and joints (see Supp. Mat. for the increased number of joints), respectively. \\(\\pmb{R}_d^h \\in \\mathrm{SO}(3)\\) is global rotation, and the \\(\\mathbf{s}_d^h, x_d, y_d\\) are scale, and \\(x, y\\) direction offsets representing the parameters of weak perspective camera \\(\\varPi_{\\gamma(d)}\\) assigned to image \\(I_d\\). \\(\\gamma(\\cdot)\\) is the function that maps the index of the 2D HOI image to the index of the weak perspective camera from which it is rendered. We additionally define \\(\\mathbf{j}_{d^\\dagger}\\) as joints in world coordinate, obtained by transforming inferred joints \\(\\mathbf{j}_d\\) with the camera rotation and translation of \\(\\varPi_{\\gamma(d)}\\). As 3D human lies within the weak perspective camera framework in world coordinate, there remains depth ambiguity to solve." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.491 + ], + "angle": 0, + "content": "Depth Optimization using Weak Auxiliary Cue. To address the depth ambiguity of 3D human obtained from reference image, we leverage 2D HOI images generated in advance. Since the human poses in images are geometrically inconsistent, we select the largest subset inliers \\(\\mathcal{I}\\) among them which are \"semiconsistent\" with the reference image by applying RANSAC [14] in terms of joint re-projection error. The joint triangulation is performed on two images, including the reference image. Note that we generate sufficiently large amounts of 2D HOI images to ensure the existence of the inlier set. The inliers \\(\\mathcal{I}\\) is used to estimate the human depth \\(z_{d}\\) by optimizing the joint re-projection loss defined below:" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.501, + 0.786, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e - p r o j e c t i o n}} = \\frac {1}{| \\mathcal {I} |} \\sum_ {I _ {\\delta} \\in \\mathcal {I}} \\| \\Pi_ {\\gamma (\\delta)} \\left(\\mathbf {j} _ {d ^ {\\dagger}} + z _ {d} \\boldsymbol {f} _ {d}\\right) - \\Pi_ {\\gamma (\\delta)} \\left(\\mathbf {j} _ {\\delta^ {\\dagger}}\\right) \\| ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.664 + ], + "angle": 0, + "content": "where \\( \\pmb{f}_d \\in \\mathbb{R}^3 \\) is the normalized direction of the camera \\( \\varPi_{\\gamma(d)} \\). Inspired by Jiang et al. [32], we leverage the human segmentation masks to reason about occlusion and promote a initial depth for human. We initialize SMPL-X [61] models along the camera direction and select the model with the best IOU between the rendered mask (regarding occlusion) and the segmented region to be an initial of optimization. This is useful when the optimal object position is in between two human body parts along the camera direction (e.g., a motorcycle body between the legs while riding) as we use a collision loss term during optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.741 + ], + "angle": 0, + "content": "The collision loss is used to estimate fine-grained depth, as the collision cue can provide the physical plausibility of 3D HOI samples. We adopt collision loss from COAP [51] where the human body is represented as SDF [57] field. We set object vertices as query points to reduce the collision between human and object, namely \\(\\mathcal{L}_{\\mathrm{collision}}\\). We optimize depth \\(z_{d}\\) to minimize the total loss:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.754, + 0.786, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e - p r o j e c t i o n}} + \\lambda_ {\\text {c o l l i s i o n}} \\mathcal {L} _ {\\text {c o l l i s i o n}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Filtering. We filter out potentially low-quality 3D human samples in the cases of: (1) human rendering (regarding occlusion) and predicted human segmentation do not overlap much; (2) number of inliers during RANSAC [14] is small; (3) human and object collides significantly in 3D." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.564, + 0.163 + ], + "angle": 0, + "content": "3.3 Learning Comprehensive Affordance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.261 + ], + "angle": 0, + "content": "Our main objective is to learn ComA, a new affordance representation which encompasses orientational and spatial affordance, along with contact-based affordance. In order to learn ComA from the previously generated 3D HOI samples, we first apply poisson disk sampling [48] to uniformly select 3D surface points for object surface and uniformly sample mesh vertices from SMPL-X [61] human, constructing ordered point sets for each object and human." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.262, + 0.79, + 0.385 + ], + "angle": 0, + "content": "Given the ordered point sets, the density function on the \\(i\\)-th object point and the \\(j\\)-th human point, which we denote as \\(\\mathcal{P}_{ij}(\\mathbf{p},\\mathbf{n})\\), represents the joint probability of the relative position \\(\\mathbf{p}\\in \\mathbb{R}^3\\) and relative normal orientation \\(\\mathbf{n}\\in \\mathbb{S}^2\\) of the \\(j\\)-th human point with respect to \\(i\\)-th object point within the 3D HOI samples. We obtain \\(\\mathbf{p}\\) and \\(\\mathbf{n}\\) by canonicalizing the orientation of the human surface normal \\(\\mathbf{n}_j^h\\) and relative position \\(\\mathbf{p}^{o\\rightarrow h}\\), assuming the object surface normal \\(\\mathbf{n}_i^o\\) is rotated to face \\(\\hat{\\mathbf{n}} = [0,0,1]^T\\) (see Supp. Mat. for details on canonicalization). The distribution \\(\\mathcal{P}_{ij}\\) sums up to 1 for all possible \\(\\mathbf{p}\\) and \\(\\mathbf{n}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.395, + 0.787, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\int_ {\\mathbb {R} ^ {3}} \\int_ {\\mathbb {S} ^ {2}} \\mathcal {P} _ {i j} (\\mathbf {p}, \\mathbf {n}) d \\mathbf {n} d \\mathbf {p} = 1 \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.556 + ], + "angle": 0, + "content": "In practice, we set the domain of \\(\\mathbf{p}\\) as voxelized grid and domain of \\(\\mathbf{n}\\) as equispaced grid on \\(\\mathbb{S}^2\\), obtained via Fibonacci Spirals [16], and compute the discrete probabilities by aggregating Gaussian Kernels calculated for \\(\\mathbf{n}\\) and \\(\\mathbf{p}\\) (we use geodesic metrics for \\(\\mathbf{n}\\)). Under the probability distribution \\(\\mathcal{P}_{ij}(\\mathbf{p},\\mathbf{n})\\), we define three types of functions \\(f(\\mathbf{p},\\mathbf{n})\\) capturing different aspects of affordances: (1) Contact-based Affordance, (2) Oriental Affordance, (3) Spatial Affordance. Given each function \\(f(\\mathbf{p},\\mathbf{n})\\), the pointwise affordance of \\(i\\)-th object point with respect to \\(j\\)-th human point is determined by computing its expectation:" + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.566, + 0.786, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathbf {p}, \\mathbf {n} \\sim \\mathcal {P} _ {i j}} [ f (\\mathbf {p}, \\mathbf {n}) ] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.623 + ], + "angle": 0, + "content": "Contact-based Affordance. From ComA, we can infer the contact score for human-object point pairs regarding proximity and normal alignment as:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.631, + 0.787, + 0.665 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {c o n t a c t}} (\\mathbf {p}, \\mathbf {n}) = \\left(\\frac {1 - \\mathbf {n} \\cdot \\hat {\\mathbf {n}}}{2}\\right) e ^ {- \\| \\mathbf {p} \\|} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.81 + ], + "angle": 0, + "content": "where \\( e^{-\\| \\mathbf{p}\\|} \\) term encourages high score when distance \\( \\| \\mathbf{p}\\| \\) is close. The \\( \\frac{1 - \\mathbf{n}\\cdot\\hat{\\mathbf{n}}}{2} \\) term, motivated by the fact that physical interactions are conducted via exerted normal force between contact points, encourages high score when the object normal \\( \\mathbf{n}_i^o \\) and human normal \\( \\mathbf{n}_j^h \\) align in an antiparallel direction. The normal alignment term improves the precision of the contact map, especially when small body parts (e.g., fingertips) interact with an object and allows robust affordance learning even when using noisy HOI datasets, compared to prior work that only uses proximity for contact terms [2, 22, 23, 90, 93] or normal for computing only penetration [19, 86], without considering the alignment of surface normals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Orientational Affordance. We aim to capture the pattern and tendency of the orientation of body parts with respect to the object based on the concept" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Discovering Comprehensive Affordance for 3D Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "of Shannon entropy [70]. Intuitively, low entropy means there exists more typical patterns of human body orientations during the interactions. To quantify the orientational tendency for ComA, we use negated normalized Shannon entropy [70] to enforce high value when the human surface normal shows consistent orientational tendency (low variance) with respect to the object surface normal:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.234, + 0.786, + 0.265 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {o r i e n t a t i o n}} (\\mathbf {n}) = 1 + \\frac {\\log \\mathcal {P} _ {i j} (\\mathbf {n})}{\\log n _ {b}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.785, + 0.359 + ], + "angle": 0, + "content": "where \\( n_b \\) denotes the number of discretized bins of probability measure. Intuitively, the orientational affordance defined by \\( \\mathbb{E}_{\\mathbf{n} \\sim \\mathcal{P}_{ij}}[f_{\\mathrm{orientation}}(\\mathbf{n})] \\) has a low value when marginalized distribution \\( \\mathcal{P}_{ij}(\\mathbf{n}) \\) is close to the uniform distribution, meaning that there is no dominant orientational pattern. Note that Eq. 7 is agnostic to the proximity term \\( \\mathbf{p} \\), which allows us to model the nonphysical orientational effects for even far-distance points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.784, + 0.405 + ], + "angle": 0, + "content": "Spatial Affordance. Following CHORUS [20], we also consider the 3D spatial occupancy distribution of a human surface point with respect to the object points by simply counting the occurrence:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.416, + 0.785, + 0.432 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {s p a t i a l}} (\\mathbf {p}) = \\mathbb {1} (\\mathbf {x} - \\mathbf {p}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.785, + 0.532 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot):\\mathbb{R}^3\\to \\{0,1\\}\\) is a binary function that returns 1 if the argument is zero vector, else 0. Thus the spatial affordance \\(\\mathbb{E}_{\\mathbf{p}\\sim \\mathcal{P}_{ij}}[f_{\\mathrm{spatial}}(\\mathbf{p})]\\) outputs scalar occupancy for the spatial vector \\(\\mathbf{x}\\in \\mathbb{R}^3\\). In practice, we implement the function as voxel array, counting 1 to the voxel that contains \\(\\mathbf{p}\\). Note that \\(f_{\\mathrm{spatial}}\\) is agnostic to the normal direction \\(\\mathbf{n}\\) via marginalization. Learning the occupancy informs us about macro positioning of the human relative to the given object." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.554, + 0.376, + 0.571 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.705 + ], + "angle": 0, + "content": "In this section, we conduct experiments to demonstrate the efficacy of our representation ComA, and the method for learning it. In Sec. 4.3, we illustrate how ComA extends beyond contact-based affordance to learn distributions for orientational tendency and spatial relation, by visualizing it across various object categories. In Sec. 4.4, we perform quantitative comparisons with other baselines and verify the efficacy of our Adaptive Mask Inpainting. Finally, in Sec. 4.5, we introduce an optimization framework achievable through the knowledge provided by ComA, demonstrating the potential of our new affordance representation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.336, + 0.74 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.784, + 0.841 + ], + "angle": 0, + "content": "3D Object Datasets. We obtain 3D object meshes from various 3D object datasets for qualitative evaluation. Specifically, We utilize 20 from BEHAV [2], 10 from InterCap [29], 2 from ShapeNet [9], and 8 from SAPIEN [83] to learn ComA individually. As textures were not available in InterCap [29], making the diffusion model difficult to inpaint, we used TEXTure [66] to generate textures on the meshes using text prompt automatically generated from ChatGPT [58]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "For quantitative evaluation, we utilize BEHAVE [2] as a ground-truth dataset to verify the quality of generated 3D HOI samples, assuming that interactions captured in lab-controlled environments exhibit the upper-bound quality of HOI. BEHAVE [2] contains video sequences of HOI scenarios for 20 object categories. Each video frame is annotated with a single 3D human and object, where each human is given as SMPL-H [44,68] format and objects as rotation and translation applied to the canonical mesh. We preprocess the dataset by transferring the SMPL-H annotations into SMPL-X format following Choutas et al. [61]. We use the 3D human-object pair in each frame as a sample, compute vertex-wise contact scores in the same way as our method, and use them as ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.33 + ], + "angle": 0, + "content": "Internet Search Datasets. We also collect 60 free 3D object meshes from Internet source, SketchFab [71] and learn ComA." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.349, + 0.436, + 0.364 + ], + "angle": 0, + "content": "4.2 Baselines and Metric" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.509 + ], + "angle": 0, + "content": "As previous studies focused on contact rather than learning the overall aspect of affordances, we use contact-based affordance derived from ComA to compare with them. However, most of them output 3D contact of a single HOI situation by inferring affordance from a given HOI sample (mostly an image), while we produce an overall affordance distribution for the given 3D object. To address this, we treat the contact inferred by state-of-the-art (SOTA) methods, DECO [79] (for the human) and IAGNet [87] (for the object) from the HOI image as a single sample, and aggregate them across multiple images to construct a distribution for comparison against BEHAVE [2]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.616 + ], + "angle": 0, + "content": "Specifically, we aggregate the results of DECO [79] and IAGNet [87] on (1) 2D HOI images we've generated beforehand and (2) BEHAVE [2] test images, as these represent the usage of the BEHAVE [2] object. The aggregated scores are then normalized to form a distribution (summing up to 1), considering only the relative probability of potential contact, as the original scale of the scores varies between methods. Subsequently, we compare similarity scores (SIM) [75] and mean absolute error (MAE) [81] within these distributions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.421, + 0.65 + ], + "angle": 0, + "content": "4.3 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Contact for Various Objects. Similar to other studies, ComA can be derived into contact-based affordance as shown in Fig. 5. As our representation provides contact scores for all pairs of human-object vertices, we assign the contact score of each vertex as the maximum score among the pairs that include it. We can observe the conventional object usage patterns as heatmap, which fits with general human usages. Specifically, in case Fig. 5(b), where significant contact is absent, we can infer that the interaction does not frequently appear by contact." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Orientational Tendency for Various Objects. ComA also contains information on orientational tendencies during the HOI. The last column in Fig. 5 visualizes the negated entropy of the vertex normal direction of human relative to the object as a heatmap. Fig. 5(a), Fig. 5(b), and Fig. 5(c) exhibit relatively high orientational tendency, while the rest show lower. This aligns with the fact that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Discovering Comprehensive Affordance for 3D Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.144, + 0.775, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.601, + 0.785, + 0.63 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Results. ComA can model distributions of contact, orientation, and spatial relation exhibited during the interaction between humans and novel objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.7 + ], + "angle": 0, + "content": "the human's orientation is fixed by the attachment to the back as in Fig. 5(a), and by the restrictions present on the back and side of the object as in Fig. 5(c). In Fig. 5(b), there exists an orientational tendency without attachments or restrictions, which arises due to long-distance interaction: viewing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Spatial Relation for Various Objects. The relative position of the human during HOI can be also derived from ComA. Fig. 5 visualizes the distribution of the human's full body positions relative to the object. Specifically, as shown in Fig. 5(d), Fig. 5(e), and Fig. 5(f), we find that the spatial relations for the object with multiple utilities and geometries (e.g., in the case of shovel, handle for holding and blade for digging) align with general human usage, demonstrating the plausibility of the results. Especially, in the case of tennis racket shown in Fig. 5(e), we can observe a relatively weak tendency in spatial relation, as the racket can move almost freely relative to humans during the interaction." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.506, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.248, + 0.471, + 0.261 + ], + "angle": 0, + "content": "(a) Finding Corresponding Contact" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.141, + 0.756, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.248, + 0.744, + 0.261 + ], + "angle": 0, + "content": "(b) Evaluating Plausibility of Orientation" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.266, + 0.787, + 0.294 + ], + "angle": 0, + "content": "Fig. 6: Pointwise Affordance Knowledge. ComA can provide pointwise affordance knowledge such as corresponding contacts or plausible orientation of body parts." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.301, + 0.768, + 0.38 + ], + "angle": 0, + "content": "
MethodsSIMHuman(%)↑SIMObject(%)↑MAEHuman(×100)↓MAEObject(×100)↓
Ours52.8270.550.1260.0288
IAGNet [87] on Our 2D HOI Images-66.34-0.0329
IAGNet [87] on BEHAVECT Images-63.86-0.0353
DECO [79] on Our 2D HOI Images21.66-0.246-
DECO [79] on BEHAVECT Images17.39-0.259-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.385, + 0.785, + 0.412 + ], + "angle": 0, + "content": "Table 1: Quantitative Results on BEHAV. We report SIM, MAE to evaluate the similarity of contact distribution (normalized contact scores) on BEHAV sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.787, + 0.613 + ], + "angle": 0, + "content": "Pointwise Affordance Knowledge. Different from previous studies, ComA models the distribution of relative orientation and proximity for all human-object vertex pairs, allowing it to provide pointwise affordance knowledge such as corresponding contacts and plausible orientation direction of specific body parts. As shown in Fig. 6(a), we can infer how each human parts and object parts contact. For the case of motorcycle, we can see that saddle and handle each contacts with human's hips and hands; while for the case of chair, the seat and backrest each contacts with human's hips and back. Also, we can utilize ComA to infer the probable direction of the head while using the monitor by computing the cosine similarity between the head direction and the vertex normal with highest orientational tendency. As illustrated in Fig. 6(b), we can see that the similarity maximizes when gazing towards the screen; which is a plausible head direction when interacting with the monitor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Ablation on Adaptive Mask Inpainting. As shown in Fig. 4, without Adaptive Mask Inpainting, the original geometry of the object changes (left) or gets replaced by hallucinated objects (right), resulting in false affordances unfaithful to the original object. Consequently, the difference between the object in 2D and the original one existing in 3D causes the uplifted 3D HOI samples to fail to represent the right usage. This shows the importance of Adaptive Mask Inpainting as it prevents 2D inpainting diffusion model from generating false affordances." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.434, + 0.756 + ], + "angle": 0, + "content": "4.4 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Comparison with Baselines. We evaluate the similarity of contact distribution between BEHAV [2] and each of DECO [79], IAGNet [87], and ours. We sample 2048 points from the 3D object mesh surface following IAGNet [87] and only consider the contact underneath the head for DECO [79], as DECO [79] uses SMPL [44] format, which only has vertex correspondence with SMPL-X [61] in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Discovering Comprehensive Affordance for 3D Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.144, + 0.681, + 0.189 + ], + "angle": 0, + "content": "
Inpainting MethodsRMSEBackground↓ mIoU↑mIoUOCclusionAware ↑
Vanilla42.230.5350.631
Adaptive Mask15.330.7060.815
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.194, + 0.787, + 0.222 + ], + "angle": 0, + "content": "Table 2: Effects of Adaptive Mask Inpainting We evaluate the performance of background preservation in vanilla inpainting and our Adaptive Mask Inpainting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.785, + 0.367 + ], + "angle": 0, + "content": "non-head regions. As shown in Tab. 1, the contact distribution from IAGNet [87] and DECO [79] using our 2D HOI images outperforms the one modeled from BEHAVE test images. This means that our generated HOI images represent the real-world affordance as well as BEHAVE [2] test images, demonstrating the efficacy of leveraging pre-trained diffusion model for generating affordance. Additionally, the contact distribution from our method outperforms the one modeled using IAGNet [87] and DECO [79] with our generated HOI images. Since we extract 3D contacts from the same image sets, we argue that our pipeline is better than other models trained on annotations for uplifting 2D affordance to 3D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.787, + 0.61 + ], + "angle": 0, + "content": "Efficacy of Adaptive Mask Inpainting. To verify the effectiveness of our Adaptive Mask Inpainting, we compare the amount of background preservation for our generated 2D HOI images with the vanilla inpainting diffusion model. We sample a maximum of 2 inpainted images with a single human detected, per each object categories from BEHAV [2]; resulting in 691 images for vanilla inpainting and 692 images for Adaptive Mask Inpainting. We compute 3 metrics for these images; (1) \\(\\mathrm{RMSE}_{\\mathrm{Background}}\\): pixel error between images excluding the predicted human region; (2) mIoU: average IoU between object predictions in each inpainted image and original object image; (3) mIoU\\(_{\\mathrm{Occlusion\\_Aware}}\\): same metric as (2) but excluding the predicted human region during computation. To ensure fairness, we use open vocabulary segmentation model [35,43] to predict detailed human/object segmentation. As shown in Tab. 2, we demonstrate that our adaptive mask algorithm better preserves the background, not only pixelwise \\((\\mathrm{RMSE}_{\\mathrm{Background}})\\), but also semantic-region-wise (mIOU, mIOU\\(_{\\mathrm{Occlusion\\_Aware}}\\)) than the vanilla inpainting baseline. This means that the generated images from adaptive mask inpainting are more likely to preserve the original object." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.358, + 0.648 + ], + "angle": 0, + "content": "4.5 Application" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.795 + ], + "angle": 0, + "content": "As ComA is a newly proposed affordance representation, we demonstrate its potential by showcasing an application. One possible task is to reconstruct HOI by optimizing the human pose and position to interact with an object. As shown in Fig. 7(a), it is nearly impossible to achieve the task with traditional methods which relies on contact affordance only. In contrast, ComA provides both relative orientation and contact during the HOI, making it possible to generate plausible 3D HOI samples from T-posed humans. For the objects with multiple modes of affordances, we apply mode seeking algorithm [11] on 3D HOI samples with their vertex normals and learn ComA separately; for example, swing in Fig. 7(a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "However, ComA learned on a specific object cannot be directly applied to other objects with different geometries. To address this, we transfer ComA through the process of (1) object canonicalization and (2) finding vertex cor-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.142, + 0.74, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Fig. 7: Application. ComA can reconstruct HOI through optimization, which can be scaled to other objects in the same category by transferring such knowledge between them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.595 + ], + "angle": 0, + "content": "respondences. Fig. 7(b) provides an example of transferring ComA learned from a specific motorcycle to other motorcycles with different geometries. We use objects from ShapeNet [9] which are already canonicalized and apply CPNet [45] to find vertex correspondences automatically for transferring the vertices with contact. As object canonicalization and vertex correspondence can each transfer macro orientation information and local contact information, we can transfer ComA and scale the human optimization task to other objects in the same category without any additional learning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.36, + 0.631 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we point out the importance of learning non-contact patterns in HOIs and propose a novel affordance representation called ComA which models both contact and non-contact patterns. To learn ComA, we utilize a pre-trained 2D diffusion model which already possesses prior knowledge of affordance. To leverage the affordance knowledge, we designed a Rendering-Inpainting-Uplifting pipeline to generate 3D HOI samples and learn ComA from the samples. The results show that ComA effectively models the common usage of objects in terms of contact, orientational, and spatial relations. Within the pipeline, we propose Adaptive Mask Inpainting, which allows insertion of human without damaging the original image, where we evaluate its efficacy both qualitatively and quantitatively. We expect ComA to be used as a prior knowledge for various downstream tasks, where we demonstrate its potential through human optimization framework and the possibility of transfer to various objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Discovering Comprehensive Affordance for 3D Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.255 + ], + "angle": 0, + "content": "This work was supported by Naver Webtoon. The work of SNU members was also supported by NRF grant funded by the Korean government (MSIT) (No. 2022R1A2C2092724 and No. RS-2023-00218601), and IITP grant funded by the Korean government (MSIT) [RS-2021-II211343, AI Graduate School Program (SNU)]. H. Joo is the corresponding author." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.278, + 0.324, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Bahl, S., Mendonca, R., Chen, L., Jain, U., Pathak, D.: Affordances from human videos as a versatile representation for robotics. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.38 + ], + "angle": 0, + "content": "2. Bhatnagar, B.L., Xie, X., Petrov, I., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Behave: Dataset and method for tracking human object interactions. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.382, + 0.785, + 0.408 + ], + "angle": 0, + "content": "3. Brahmbhatt, S., Ham, C., Kemp, C.C., Hays, J.: Contactdb: Analyzing and predicting grasp contact via thermal imaging. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.785, + 0.435 + ], + "angle": 0, + "content": "4. Brahmbhatt, S., Handa, A., Hays, J., Fox, D.: Contactgrasp: Functional multifinger grasp synthesis from contact. In: IROS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.464 + ], + "angle": 0, + "content": "5. Brahmbhatt, S., Tang, C., Twigg, C.D., Kemp, C.C., Hays, J.: Contactpose: A dataset of grasps with object contact and hand pose. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.785, + 0.492 + ], + "angle": 0, + "content": "6. Cai, M., Kitani, K.M., Sato, Y.: Understanding hand-object manipulation with grasp types and object attributes. In: RSS (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.494, + 0.785, + 0.52 + ], + "angle": 0, + "content": "7. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.548 + ], + "angle": 0, + "content": "8. Chai, L., Zhu, J.Y., Shechtman, E., Isola, P., Zhang, R.: Ensembling with deep generative views. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.59 + ], + "angle": 0, + "content": "9. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "10. Chen, Y., Huang, S., Yuan, T., Qi, S., Zhu, Y., Zhu, S.C.: Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.633, + 0.785, + 0.659 + ], + "angle": 0, + "content": "11. Comaniciu, D., Meer, P.: Mean shift: a robust approach toward feature space analysis. In: IEEE TPAMI (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.66, + 0.785, + 0.687 + ], + "angle": 0, + "content": "12. Deng, S., Xu, X., Wu, C., Chen, K., Jia, K.: 3d affordancenet: A benchmark for visual object affordance understanding. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.688, + 0.785, + 0.715 + ], + "angle": 0, + "content": "13. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "14. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. In: CACM (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "15. Gao, W., Tedrake, R.: kpm-sc: Generalizable manipulation planning using keypoint affordance and shape completion. In: ICRA (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "16. Garg, M., Garg, P., Vohra, R.: Advanced fibonacci sequence with golden ratio. In: IJSER (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "17. Gibson, J.J.: The Ecological Approach to Visual Perception. Houghton Mifflin (1979)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.311, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "19. Grady, P., Tang, C., Twigg, C.D., Vo, M., Brahmbhatt, S., Kemp, C.C.: ContactOpt: Optimizing contact to improve grasps. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "20. Han, S., Joo, H.: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "21. Hao, Z., Mallya, A., Belongie, S., Liu, M.Y.: Gancraft: Unsupervised 3d neural rendering of apache worlds. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "22. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.31 + ], + "angle": 0, + "content": "23. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.351 + ], + "angle": 0, + "content": "24. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? In: arXiv:2210.07574 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.378 + ], + "angle": 0, + "content": "25. Hermans, T., Rehg, J.M., Bobick, A.: Affordance prediction via learned object attributes. In: ICRA (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "26. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: arXiv:2208.01626 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.487 + ], + "angle": 0, + "content": "28. Huang, C.H.P., Yi, H., Hoschle, M., Safroshkin, M., Alexiadis, T., Polikovsky, S., Scharstein, D., Black, M.J.: Capturing and inferring dense full-body human-scene contact. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "29. Huang, Y., Taheri, O., Black, M.J., Tzionas, D.: Intercap: Joint markerless 3d tracking of humans and objects in interaction from multi-view rgb-d images. In: IJCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.555 + ], + "angle": 0, + "content": "30. Jahanian, A., Puig, X., Tian, Y., Isola, P.: Generative models as a data source for multiview representation learning. In: arXiv:2106.05258 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.582 + ], + "angle": 0, + "content": "31. Jian, J., Liu, X., Li, M., Hu, R., Liu, J.: Affordpose: A large-scale dataset of handobject interactions with affordance-driven hand pose. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "32. Jiang, W., Kolotouros, N., Pavlakos, G., Zhou, X., Daniilidis, K.: Coherent reconstruction of multiple humans from a single image. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.65 + ], + "angle": 0, + "content": "33. Jiang, Y., Jiang, S., Sun, G., Su, Z., Guo, K., Wu, M., Yu, J., Xu, L.: Neuralhofusion: Neural volumetric rendering under human-object interactions. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.676 + ], + "angle": 0, + "content": "34. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.718 + ], + "angle": 0, + "content": "35. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.746 + ], + "angle": 0, + "content": "36. Kirillov, A., Wu, Y., He, K., Girshick, R.: Pointrend: Image segmentation as rendering. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.786 + ], + "angle": 0, + "content": "37. Kulal, S., Brooks, T., Aiken, A., Wu, J., Yang, J., Lu, J., Efros, A.A., Singh, K.K.: Putting people in their place: Affordance-aware human insertion into scenes. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.787, + 0.785, + 0.813 + ], + "angle": 0, + "content": "38. Lee, Y.J., Grauman, K.: Predicting important objects for egocentric video summarization. In: IJCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Levine, S., Shah, D.: Learning robotic navigation from experience: principles, methods and recent results. In: Philos. Trans. R. Soc. B (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Discovering Comprehensive Affordance for 3D Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "40. Li, G., Jampani, V., Sun, D., Sevilla-Lara, L.: Locate: Localize and transfer object parts for weakly supervised affordance grounding. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "41. Li, L., Dai, A.: GenZI: Zero-shot 3D human-scene interaction generation. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.245 + ], + "angle": 0, + "content": "42. Li, Z., Sedlar, J., Carpentier, J., Laptev, I., Mansard, N., Sivic, J.: Estimating 3d motion and forces of person-object interactions from monocular video. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.285 + ], + "angle": 0, + "content": "43. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.313 + ], + "angle": 0, + "content": "44. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. In: Proc. ACM SIGGRAPH Asia (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "45. Lou, Y., You, Y., Li, C., Cheng, Z., Li, L., Ma, L., Wang, W., Lu, C.: Human correspondence consensus for 3d object semantic understanding. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "46. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "47. Mao, C., Cha, A., Gupta, A., Wang, H., Yang, J., Vondrick, C.: Generative interventions for causal learning. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "48. McCool, M., Fiume, E.: Hierarchical poisson disk sampling distributions. In: Graphics Interface (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "49. Melas-Kyriazi, L., Rupprecht, C., Laina, I., Vedaldi, A.: Finding an unsupervised image segmenter in each of your deep generative models. In: arXiv:2105.08127 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "50. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "51. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: COAP: Compositional articulated occupancy of people. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "52. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "53. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "54. Monszpart, A., Guerrero, P., Ceylan, D., Yumer, E., Mitra, N.J.: imapper: interaction-guided scene mapping from monocular videos. In: ACM TOG (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "55. Moon, G., Choi, H., Lee, K.M.: Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In: CVPRW (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "56. Nagarajan, T., Grauman, K.: Learning affordance landscapes for interaction exploration in 3d environments. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "57. Oleynikova, H., Millane, A., Taylor, Z., Galceran, E., Nieto, J., Siegwart, R.: Signed distance fields: A natural representation for both mapping and planning. In: RSS Workshop (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "58. OpenAI: Chatgpt: Optimizing language models for dialogue. https://openai.com/blog/chatgpt/ (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "59. OpenAI: Gpt-4v(ision) system card. https://openai.com/research/gpt-4v-system-card (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "60. Pan, X., Dai, B., Liu, Z., Loy, C.C., Luo, P.: Do 2d gans know 3d shape? unsupervised 3d shape reconstruction from 2d image gans. In: arXiv:2011.00844 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "61. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "62. Peebles, W., Zhu, J.Y., Zhang, R., Torralba, A., Efros, A.A., Shechtman, E.: Gansupervised dense visual alignment. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.202 + ], + "angle": 0, + "content": "63. Petrov, I.A., Marin, R., Chibane, J., Pons-Moll, G.: Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.23 + ], + "angle": 0, + "content": "64. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.257 + ], + "angle": 0, + "content": "65. Rempe, D., Guibas, L.J., Hertzmann, A., Russell, B., Villegas, R., Yang, J.: Contact and human dynamics from monocular video. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "66. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. In: Proc. ACM SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "67. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "68. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. In: Proc. ACM SIGGRAPH Asia (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "69. Savva, M., Chang, A.X., Hanrahan, P., Fisher, M., Nießner, M.: Pigraphs: learning interaction snapshots from observations. In: ACM TOG (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.392 + ], + "angle": 0, + "content": "70. Shannon, C.E.: A mathematical theory of communication. In: ACM SIGMOBILE (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.61, + 0.405 + ], + "angle": 0, + "content": "71. Sketchfab: Sketchfab. https://sketchfab.com/ (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.785, + 0.432 + ], + "angle": 0, + "content": "72. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.459 + ], + "angle": 0, + "content": "73. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.5 + ], + "angle": 0, + "content": "74. Sun, G., Chen, X., Chen, Y., Pang, A., Lin, P., Jiang, Y., Xu, L., Yu, J., Wang, J.: Neural free-viewpoint performance rendering under complex human-object interactions. In: ACM MM (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.5, + 0.643, + 0.514 + ], + "angle": 0, + "content": "75. Swain, M.J., Ballard, D.H.: Color indexing. In: IJCV (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.541 + ], + "angle": 0, + "content": "76. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.568 + ], + "angle": 0, + "content": "77. Tanaka, F.H.K.d.S., Aranha, C.: Data augmentation using gans. In: arXiv:1904.09135 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.595 + ], + "angle": 0, + "content": "78. Trabucco, B., Doherty, K., Gurinas, M., Salakhutdinov, R.: Effective data augmentation with diffusion models. In: arXiv:2302.07944 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.622 + ], + "angle": 0, + "content": "79. Tripathi, S., Chatterjee, A., Passy, J.C., Yi, H., Tzionas, D., Black, M.J.: DECO: Dense estimation of 3D human-scene contact in the wild. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.785, + 0.649 + ], + "angle": 0, + "content": "80. Tritrong, N., Rewatbowornwong, P., Suwajanakorn, S.: Repurposing gans for one-shot semantic part segmentation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.649, + 0.785, + 0.69 + ], + "angle": 0, + "content": "81. Willmott, C.J., Matsuura, K.: Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. In: Climate Research (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.785, + 0.731 + ], + "angle": 0, + "content": "82. Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetdm: Synthesizing data with perception annotations using diffusion models. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.772 + ], + "angle": 0, + "content": "83. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "84. Xu, J., Zhu, S., Guo, H., Wu, S.: Automated labeling for robotic autonomous navigation through multi-sensory semi-supervised learning on big data. In: IEEE TBD (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "85. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Discovering Comprehensive Affordance for 3D Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "86. Yang, L., Zhan, X., Li, K., Xu, W., Li, J., Lu, C.: CPF: Learning a contact potential field to model the hand-object interaction. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "87. Yang, Y., Zhai, W., Luo, H., Cao, Y., Luo, J., Zha, Z.J.: Grounding 3d object affordance from 2d interactions in images. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "88. Yang, Y., Zhai, W., Luo, H., Cao, Y., Zha, Z.J.: Lemon: Learning 3d human-object interaction relation from 2d images. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "89. Ye, Y., Li, X., Gupta, A., Mello, S.D., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.299 + ], + "angle": 0, + "content": "90. Yi, H., Huang, C.H.P., Tzionas, D., Kocabas, M., Hassan, M., Tang, S., Thies, J., Black, M.J.: Human-aware object placement for visual environment reconstruction. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.327 + ], + "angle": 0, + "content": "91. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 3DV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "92. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "93. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "94. Zhang, Y., Chen, W., Ling, H., Gao, J., Zhang, Y., Torralba, A., Fidler, S.: Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "95. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "96. Zhou, K., Bhatnagar, B.L., Lenssen, J.E., Pons-Moll, G.: Toch: Spatio-temporal object correspondence to hand for motion refinement. In: ECCV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.494 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/4cf86d6f-6634-4c59-b4d4-c22212431643_origin.pdf b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/4cf86d6f-6634-4c59-b4d4-c22212431643_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..77bc661892e5750643031cd06ea1503133f9ddf3 --- /dev/null +++ b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/4cf86d6f-6634-4c59-b4d4-c22212431643_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8a986fd0238d20aa251465ab1d07887f8ce9249b6b23623137e5a1cb6a0e3d2 +size 9678987 diff --git a/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/full.md b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..47f3291ca348ca4ffbd1dcc509f30a973d75b749 --- /dev/null +++ b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/full.md @@ -0,0 +1,325 @@ +# Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models + +Hyeonwoo $\mathrm{Kim}^{1*}$ , Sookwan Han $^{1*}$ , Patrick Kwon $^{2}$ , and Hanbyul Joo $^{1}$ + +1 Seoul National University + +2 Naver Webtoon AI + +![](images/36ebae95e9d9b0eb350d9c84445a18c8b0280fec06dd412637ea18e4232cbb5b.jpg) +Fig. 1: Given a 3D object, we generate numerous 3D Human-Object Interaction (HOI) samples using text prompts, and learn a novel affordance representation called Comprehensive Affordance (ComA) which models both contact and non-contact HOI patterns. + +Abstract. Understanding the inherent human knowledge in interacting with a given environment (e.g., affordance) is essential for improving AI to better assist humans. While existing approaches primarily focus on human-object contacts during interactions, such affordance representation cannot fully address other important aspects of human-object interactions (HOIs), i.e. patterns of relative positions and orientations. In this paper, we introduce a novel affordance representation, named Comprehensive Affordance (ComA). Given a 3D object mesh, ComA models the distribution of relative orientation and proximity of vertices in interacting human meshes, capturing plausible patterns of contact, relative orientations, and spatial relationships. To construct the distribution, we present a novel pipeline that synthesizes diverse and realistic 3D HOI samples given any 3D target object mesh. The pipeline leverages a pre-trained 2D inpainting diffusion model to generate HOI images from object renderings and lifts them into 3D. To avoid the generation of false affordances, we propose a new inpainting framework, Adaptive Mask Inpainting. Since ComA is built on synthetic samples, it can extend to any object in an unbounded manner. Through extensive experiments, we demonstrate that ComA outperforms competitors that rely on human annotations in modeling contact-based affordance. Importantly, we also showcase the potential of ComA to reconstruct human-object interactions in 3D through an optimization framework, highlighting its advantage in incorporating both contact and non-contact properties. + +Keywords: Affordance $\cdot$ Human-Object Interaction $\cdot$ Diffusion Models + +# 1 Introduction + +Humans possess an innate ability to perceive the functionalities provided by the environment or objects and efficiently utilize them—a concept known as affordance [17]. Affordance incorporates a variety of patterns typically observable in human-object interactions (HOIs), including not only plausible physical contact but also spatial and orientational non-contact relationships. For example, while using a laptop, the orientation and distance patterns between human face or other body parts and the screen can be considered part of affordance, alongside the hand regions touching the laptop keyboard. Teaching such affordance knowledge to machines is crucial for enabling them to better understand or mimic human behaviors. Consequently, numerous studies have been conducted to implement such affordance knowledge into AI and robotics [1,20,23,37,40,63,79,87,89]. + +Most previous approaches however, primarily focus on a limited spectrum of affordances. Some approaches represent the human and object regions in contact using contact scores on 2D images [1,40], 3D objects [87], and 3D humans [23,79] using direct supervision from human annotations. Others represent affordances by inferring [63] or generating [37,89] plausible human-object pairs in contact-oriented object categories. While a few recent studies tackle learning the spatial relations [20,88] between a human and an object, these methods mainly focus on object categories where physical contact is dominant than non-contact patterns. + +In this paper, we introduce Comprehensive Affordance (ComA), a novel representation of an affordance encompassing the various aspects of human-object interaction, including orientational tendencies and relative positions. The motivation for our design arises from the observation that there exist typical 3D relations between entire body parts and objects during HOIs. For instance, human faces, torsos, and arms often exhibit specific distances and orientations, with some variability, when interacting with certain object categories, as shown in Fig. 2. Our ComA is designed to model these patterns by constructing a distribution between each pair of object surface points and human surface points in 3D. This distribution represents the likelihood of human body parts displaying specific spatial and orientational relations with object points, as shown in Fig. 3. + +Learning ComA requires various 3D HOI samples showing the way humans use the target object in 3D, from which pairwise distributions of 3D HOI relations can be constructed. Manually annotating such cues is challenging, compared to the annotations focusing solely on contact regions [40, 63, 79, 85, 87, 88]. As a key solution, we present a pipeline to synthesize a large-scale synthetic 3D HOI samples leveraging a pre-trained 2D diffusion model. Interestingly, we observe that a pre-trained 2D diffusion model captures the affordance knowledge in the form of 2D images, as shown in Fig. 2, where the images are synthesized from text prompts. However, leveraging these 2D HOI cues for construction of ComA (which requires 3D HOI samples) presents the following challenges: (1) the diffusion model needs to be applied for inpainting plausible humans without altering the original shape and pose of the target object, and (2) 3D HOI samples should be lifted from the synthesized 2D images. To address these challenges, we present + +![](images/d78d3eae5f629e72c5458a6734b1ddd26aa640641cbe34a733b21046f564dfb8.jpg) +Text Prompt: 1 person using a laptop + +![](images/cc1e59e40fb8408604b261e917e8179ad2c0d9e567e71a8a22c8a53b30813162.jpg) +Text Prompt: 1 person using a telescope +Fig. 2: Typically, people (1) view the screen (2) from relatively distant distance while using a laptop (Left), whereas they (1) peer into it (2) from a close distance while using a telescope (Right). Pre-trained diffusion model has a knowledge of these (1) orientational and (2) spatial relation between human and object during interaction. + +Adaptive Mask Inpainting to insert humans without altering the appearance of target objects in 2D, along with a 3D lifting pipeline from 2D HOI cues. + +We demonstrate the efficacy of our approach by learning ComA on 100 3D object meshes in diverse categories collected from various sources [2,9,29,52,71, 83]. We compare our ComA with existing approaches [79,87] on the BEHAV dataset [2], demonstrating that ours learned from synthetic cues outperforms competitors trained on manual annotations for contact-based affordance. Additionally, we verify the advantage of ComA against contact-only representations, in the scenario of reconstructing HOIs via an optimization framework. We will release the results and our source code1. + +In summary, our main contributions can be summarized as follows: (1) we introduce a new representation ComA, which models the distribution of both explicit contact and non-contact patterns during HOIs; (2) we present a pipeline to build ComA from synthetic datasets by leveraging a pre-trained 2D diffusion model, scalable to any 3D object without requiring laborious annotations; and (3) we propose Adaptive Mask Inpainting, a method for preserving the original context of an image in using inpainting diffusion model. + +# 2 Related Work + +Learning Visual Affordance. Affordance was introduced by Gibson [17] as a set of functionalities that the environment furnishes to an agent (e.g., human, animal). The concept extends to the vision and robotics area, where the focus is on teaching embodied agents to interact with scenes [1, 15, 39, 56, 84], mimicking human-object [2, 7, 10, 13, 23, 28, 33, 42, 54, 63, 65, 69, 74, 91, 92], or hand-object [3-5, 76, 96] interactions. Earlier methods focused on learning action category labels [6, 25, 38] with bounding boxes but lacked a detailed description of affordance. Due to the limited affordance information available from text descriptions or labels, some studies depicted affordance as contact regions or heatmaps on objects [1, 40, 87] and humans [23, 79]. However, learning affordances beyond the contact was challenging since existing datasets created through a multi-camera system [2, 29], + +motion capture [22], and manual process [12,31,79,87] focused only on annotating contact information rather than capturing the holistic aspect of HOI. Following studies generate plausible HOI samples in 2D [37,89] or 3D [85,93] which has the potential for learning such broader HOI knowledge, but do not address the problem directly. In contrast, our method generates 3D HOI samples and extracts ComA for modelling beyond the contact. + +Data Synthesis for Learning. There have been approaches to supplement data in fields lacking efficient annotation process by leveraging generative models, irrespective of the research area. Many methods employ GANs [18, 34] or diffusion models [27, 67, 73] to augment data [24, 78] for various vision tasks, including perception and representation learning [82], classification [8, 47, 77], segmentation [49, 80, 95], dense visual alignment [62], and further extending to 3D tasks, such as neural rendering [21, 94], shape reconstruction [60], and so forth. The major challenge in leveraging synthesized data (or generative models for cues) is controllability, as the generator, even when conditioned, typically produces free samples. Ali et al. [30] presents a method to control the viewpoint of the synthesized image by modeling the viewpoint-free latent space. CHORUS [20] applies cascaded filtering to the generated dataset to learn 3D human-object spatial relations, similar to our approach. + +Diffusion Model for Synthesizing HOI Images. While diffusion models [27, 72, 73] excel at generating realistic images [64, 67], few address the generation of 2D human-object pairs (i.e., HOI images). Inpainting models [46, 67], while a common choice for human insertion, often compromise scene details, resulting in images with false affordance. Image editing techniques [26, 50, 53] may offer an alternative but tend to prioritize style changes and struggle with creating new geometry. Recently, Ye et al. [89] employs an additional layout network to determine the inpainting region, and Kulal et al. [37] trains a diffusion model on video clips to generate HOI images. In contrast to these approaches, we create HOI images by inserting humans while maintaining the context of the original image, without additional network training. Concurrent work by Li et al. [41] introduces Dynamic Mask Inpainting, which uses attention masks to change inpainting masks over timesteps to insert human into a 3D scene, while we explicitly apply segmentation model to adapt the inpainting mask. + +# 3 Method + +Given an input 3D object mesh, our goal is to learn our novel affordance representation, ComA, which models the distribution of relative proximity and orientation between human and the object surface points. Our pipeline begins with creating 2D HOI images (Sec. 3.1). These images are then lifted to 3D, resulting large-scale synthetic 3D HOI samples (Sec. 3.2). From these 3D samples, we finally learn ComA (Sec. 3.3). An overview of our method is shown in Fig. 3. + +![](images/f875f4cc7482eba382bc3cb312decb0bf0b4096204ce340eb2c2eb9d0ad60f0b.jpg) + +# Learning Comprehensive Affordance from Generated 3D HOI Samples + +![](images/19a5cb419826cb1fc8f8e4fa4874d771501f7ae1d96f6a9a6f8ffaca6ec9dd07.jpg) +Fig. 3: Method Overview. Our method can be divided into two parts: (1) Generating 3D HOI Samples and (2) Learning ComA from Generated 3D HOI Samples. In the first step, we utilize an inpainting diffusion model with our Adaptive Mask Inpainting to create 2D HOI images, and generate 3D HOI samples via uplifting pipeline. In the second step, the generated 3D HOI samples are aggregated to create distributions for relative proximity and orientation, which can be derived into various affordance forms. + +# 3.1 2D HOI Image Generation + +Rendering Object from Multi-Viewpoints. Given a 3D object, we first render it from multiple viewpoints. We place the object and camera differently for two object types: static and dynamic. For objects that are presumed to be fixed on the ground during interaction (e.g., chair), we fix the object on the ground plane, whereas objects with dynamic pose during interaction (e.g., cup) are perturbed with sampled rotation and translation. Type of the object is inferred automatically by rendering the object at any direction first and then utilize vision-language model [59] for predicting. The rendering process can be fully automated, but we empirically find it beneficial to leverage a small amount of human labor, such as restricting the range of perturbation due to the inherent bias of pre-trained diffusion models. For rendering, we place weak perspective cameras $\{\varPi_c\}_{c=1}^N$ around the object with equal azimuth interval and same elevation, where $N$ is set as 8 for static, 40 for dynamic objects. + +Inpainting Mask Selection. After rendering the images, we select the masks to inpaint a human interacting with the object. While our inpainting pipeline is quite robust to initial masks, we found that initial mask selection is beneficial for avoiding failure cases such as generating hallucinated objects or ambiguous interactions. Specifically, we move sliding windows around the object and choose the windows whose Intersection over Union (IoU) with the object segmentation mask falls within a specific range. See Supp. Mat. for more details. + +![](images/7b5047e38edee739660cfe10ce09d4b33f4fce88ba1b11be9c72e0e0bcdfb089.jpg) +Fig. 4: Adaptive Mask Inpainting. Without adaptive mask inpainting, the original object is damaged when inserting humans, resulting false affordances. + +Prompt Generation. Inspired by CHORUS [20], we automatically generate the prompts that describe the possible interactions with humans and the given object, which are taken as guidance for inpainting. Specifically, we utilize the vision-language model [59] to infer the HOI prompt using the rendered object image and predefined template. We generate 3 prompts per given object. + +Adaptive Mask Inpainting for Human Insertion. We synthesize 2D images of human-object interaction by inpainting the human into rendered object images using publicly available text-conditional inpainting diffusion models [67]. However, the challenge arises as the object geometry and details within the mask region are not preserved during inpainting, resulting false affordance. To mitigate this problem, we present Adaptive Mask Inpainting to progressively specify the inpainting region over diffusion timesteps. Let the sequence of denoised image latent be $\{x_{t}\}_{t = T}^{0}$ , where $x_{T}$ is the fully random noise and $x_0$ is the denoised latent. Inspired by the observation that the quality of the predicted denoised image $\hat{x}_0$ improves over progress of timestep, and creates a low-level structure of the target prompt (in this case, human) even at the early steps as shown in Fig. 4, we aim to adapt the inpainting region over timesteps by spatially discovering the mask from the low-level structure. Specifically, we apply PointRend [36] to decoded latent $\mathbf{D}(\hat{x}_0)$ at specific timesteps to predict the human mask, and consequently use as the mask for the next denoising step. + +# 3.2 Lifting 2D Affordance to 3D + +From the previous section, we obtain a set of 2D HOI images $\{I_d\}_{d=1}^D$ . Even though we have multiple views with geometrically consistent object renderings, the inserted humans have diversity without multi-view consistency, and thus it is non-trivial to directly lift 3D humans from these 2D images. + +To address this, we first utilize a single view 3D human prediction model to obtain a corresponding 3D human pose and shape of the image, and then estimate the depth of the human with weak auxiliary cue provided by the set of 2D HOI images, generating a 3D HOI sample corresponds to the image. + +Single View 3D Human Prediction. To uplift the 2D HOI images into 3D, we first predict the pose and shape of 3D humans from the images. We apply off-the-shelf 3D human prediction model [55] (denoted as $\mathbf{F}_{\mathrm{human}}$ ) to predict + +SMPL-X humans [61] from generated images: + +$$ +\left\{\boldsymbol {\theta} _ {d}, \boldsymbol {\beta} _ {d}, \mathbf {j} _ {d}, \boldsymbol {R} _ {d} ^ {h}, \mathbf {s} _ {d} ^ {h}, x _ {d}, y _ {d} \right\} = \mathbf {F} _ {\text {h u m a n}} (I _ {d}) \tag {1} +$$ + +, where $\pmb{\theta}_d \in \mathbb{R}^{54 \times 3}, \pmb{\beta}_d \in \mathbb{R}^{10}, \mathbf{j}_d \in \mathbb{R}^{67 \times 3}$ are the predicted SMPL-X pose, shape, and joints (see Supp. Mat. for the increased number of joints), respectively. $\pmb{R}_d^h \in \mathrm{SO}(3)$ is global rotation, and the $\mathbf{s}_d^h, x_d, y_d$ are scale, and $x, y$ direction offsets representing the parameters of weak perspective camera $\varPi_{\gamma(d)}$ assigned to image $I_d$ . $\gamma(\cdot)$ is the function that maps the index of the 2D HOI image to the index of the weak perspective camera from which it is rendered. We additionally define $\mathbf{j}_{d^\dagger}$ as joints in world coordinate, obtained by transforming inferred joints $\mathbf{j}_d$ with the camera rotation and translation of $\varPi_{\gamma(d)}$ . As 3D human lies within the weak perspective camera framework in world coordinate, there remains depth ambiguity to solve. + +Depth Optimization using Weak Auxiliary Cue. To address the depth ambiguity of 3D human obtained from reference image, we leverage 2D HOI images generated in advance. Since the human poses in images are geometrically inconsistent, we select the largest subset inliers $\mathcal{I}$ among them which are "semiconsistent" with the reference image by applying RANSAC [14] in terms of joint re-projection error. The joint triangulation is performed on two images, including the reference image. Note that we generate sufficiently large amounts of 2D HOI images to ensure the existence of the inlier set. The inliers $\mathcal{I}$ is used to estimate the human depth $z_{d}$ by optimizing the joint re-projection loss defined below: + +$$ +\mathcal {L} _ {\text {r e - p r o j e c t i o n}} = \frac {1}{| \mathcal {I} |} \sum_ {I _ {\delta} \in \mathcal {I}} \| \Pi_ {\gamma (\delta)} \left(\mathbf {j} _ {d ^ {\dagger}} + z _ {d} \boldsymbol {f} _ {d}\right) - \Pi_ {\gamma (\delta)} \left(\mathbf {j} _ {\delta^ {\dagger}}\right) \| ^ {2} \tag {2} +$$ + +where $\pmb{f}_d \in \mathbb{R}^3$ is the normalized direction of the camera $\varPi_{\gamma(d)}$ . Inspired by Jiang et al. [32], we leverage the human segmentation masks to reason about occlusion and promote a initial depth for human. We initialize SMPL-X [61] models along the camera direction and select the model with the best IOU between the rendered mask (regarding occlusion) and the segmented region to be an initial of optimization. This is useful when the optimal object position is in between two human body parts along the camera direction (e.g., a motorcycle body between the legs while riding) as we use a collision loss term during optimization. + +The collision loss is used to estimate fine-grained depth, as the collision cue can provide the physical plausibility of 3D HOI samples. We adopt collision loss from COAP [51] where the human body is represented as SDF [57] field. We set object vertices as query points to reduce the collision between human and object, namely $\mathcal{L}_{\mathrm{collision}}$ . We optimize depth $z_{d}$ to minimize the total loss: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e - p r o j e c t i o n}} + \lambda_ {\text {c o l l i s i o n}} \mathcal {L} _ {\text {c o l l i s i o n}} \tag {3} +$$ + +Filtering. We filter out potentially low-quality 3D human samples in the cases of: (1) human rendering (regarding occlusion) and predicted human segmentation do not overlap much; (2) number of inliers during RANSAC [14] is small; (3) human and object collides significantly in 3D. + +# 3.3 Learning Comprehensive Affordance + +Our main objective is to learn ComA, a new affordance representation which encompasses orientational and spatial affordance, along with contact-based affordance. In order to learn ComA from the previously generated 3D HOI samples, we first apply poisson disk sampling [48] to uniformly select 3D surface points for object surface and uniformly sample mesh vertices from SMPL-X [61] human, constructing ordered point sets for each object and human. + +Given the ordered point sets, the density function on the $i$ -th object point and the $j$ -th human point, which we denote as $\mathcal{P}_{ij}(\mathbf{p},\mathbf{n})$ , represents the joint probability of the relative position $\mathbf{p}\in \mathbb{R}^3$ and relative normal orientation $\mathbf{n}\in \mathbb{S}^2$ of the $j$ -th human point with respect to $i$ -th object point within the 3D HOI samples. We obtain $\mathbf{p}$ and $\mathbf{n}$ by canonicalizing the orientation of the human surface normal $\mathbf{n}_j^h$ and relative position $\mathbf{p}^{o\rightarrow h}$ , assuming the object surface normal $\mathbf{n}_i^o$ is rotated to face $\hat{\mathbf{n}} = [0,0,1]^T$ (see Supp. Mat. for details on canonicalization). The distribution $\mathcal{P}_{ij}$ sums up to 1 for all possible $\mathbf{p}$ and $\mathbf{n}$ : + +$$ +\int_ {\mathbb {R} ^ {3}} \int_ {\mathbb {S} ^ {2}} \mathcal {P} _ {i j} (\mathbf {p}, \mathbf {n}) d \mathbf {n} d \mathbf {p} = 1 \tag {4} +$$ + +In practice, we set the domain of $\mathbf{p}$ as voxelized grid and domain of $\mathbf{n}$ as equispaced grid on $\mathbb{S}^2$ , obtained via Fibonacci Spirals [16], and compute the discrete probabilities by aggregating Gaussian Kernels calculated for $\mathbf{n}$ and $\mathbf{p}$ (we use geodesic metrics for $\mathbf{n}$ ). Under the probability distribution $\mathcal{P}_{ij}(\mathbf{p},\mathbf{n})$ , we define three types of functions $f(\mathbf{p},\mathbf{n})$ capturing different aspects of affordances: (1) Contact-based Affordance, (2) Oriental Affordance, (3) Spatial Affordance. Given each function $f(\mathbf{p},\mathbf{n})$ , the pointwise affordance of $i$ -th object point with respect to $j$ -th human point is determined by computing its expectation: + +$$ +\mathbb {E} _ {\mathbf {p}, \mathbf {n} \sim \mathcal {P} _ {i j}} [ f (\mathbf {p}, \mathbf {n}) ] \tag {5} +$$ + +Contact-based Affordance. From ComA, we can infer the contact score for human-object point pairs regarding proximity and normal alignment as: + +$$ +f _ {\mathrm {c o n t a c t}} (\mathbf {p}, \mathbf {n}) = \left(\frac {1 - \mathbf {n} \cdot \hat {\mathbf {n}}}{2}\right) e ^ {- \| \mathbf {p} \|} \tag {6} +$$ + +where $e^{-\| \mathbf{p}\|}$ term encourages high score when distance $\| \mathbf{p}\|$ is close. The $\frac{1 - \mathbf{n}\cdot\hat{\mathbf{n}}}{2}$ term, motivated by the fact that physical interactions are conducted via exerted normal force between contact points, encourages high score when the object normal $\mathbf{n}_i^o$ and human normal $\mathbf{n}_j^h$ align in an antiparallel direction. The normal alignment term improves the precision of the contact map, especially when small body parts (e.g., fingertips) interact with an object and allows robust affordance learning even when using noisy HOI datasets, compared to prior work that only uses proximity for contact terms [2, 22, 23, 90, 93] or normal for computing only penetration [19, 86], without considering the alignment of surface normals. + +Orientational Affordance. We aim to capture the pattern and tendency of the orientation of body parts with respect to the object based on the concept + +of Shannon entropy [70]. Intuitively, low entropy means there exists more typical patterns of human body orientations during the interactions. To quantify the orientational tendency for ComA, we use negated normalized Shannon entropy [70] to enforce high value when the human surface normal shows consistent orientational tendency (low variance) with respect to the object surface normal: + +$$ +f _ {\text {o r i e n t a t i o n}} (\mathbf {n}) = 1 + \frac {\log \mathcal {P} _ {i j} (\mathbf {n})}{\log n _ {b}} \tag {7} +$$ + +where $n_b$ denotes the number of discretized bins of probability measure. Intuitively, the orientational affordance defined by $\mathbb{E}_{\mathbf{n} \sim \mathcal{P}_{ij}}[f_{\mathrm{orientation}}(\mathbf{n})]$ has a low value when marginalized distribution $\mathcal{P}_{ij}(\mathbf{n})$ is close to the uniform distribution, meaning that there is no dominant orientational pattern. Note that Eq. 7 is agnostic to the proximity term $\mathbf{p}$ , which allows us to model the nonphysical orientational effects for even far-distance points. + +Spatial Affordance. Following CHORUS [20], we also consider the 3D spatial occupancy distribution of a human surface point with respect to the object points by simply counting the occurrence: + +$$ +f _ {\text {s p a t i a l}} (\mathbf {p}) = \mathbb {1} (\mathbf {x} - \mathbf {p}) \tag {8} +$$ + +where $\mathbb{1}(\cdot):\mathbb{R}^3\to \{0,1\}$ is a binary function that returns 1 if the argument is zero vector, else 0. Thus the spatial affordance $\mathbb{E}_{\mathbf{p}\sim \mathcal{P}_{ij}}[f_{\mathrm{spatial}}(\mathbf{p})]$ outputs scalar occupancy for the spatial vector $\mathbf{x}\in \mathbb{R}^3$ . In practice, we implement the function as voxel array, counting 1 to the voxel that contains $\mathbf{p}$ . Note that $f_{\mathrm{spatial}}$ is agnostic to the normal direction $\mathbf{n}$ via marginalization. Learning the occupancy informs us about macro positioning of the human relative to the given object. + +# 4 Experiments + +In this section, we conduct experiments to demonstrate the efficacy of our representation ComA, and the method for learning it. In Sec. 4.3, we illustrate how ComA extends beyond contact-based affordance to learn distributions for orientational tendency and spatial relation, by visualizing it across various object categories. In Sec. 4.4, we perform quantitative comparisons with other baselines and verify the efficacy of our Adaptive Mask Inpainting. Finally, in Sec. 4.5, we introduce an optimization framework achievable through the knowledge provided by ComA, demonstrating the potential of our new affordance representation. + +# 4.1 Datasets + +3D Object Datasets. We obtain 3D object meshes from various 3D object datasets for qualitative evaluation. Specifically, We utilize 20 from BEHAV [2], 10 from InterCap [29], 2 from ShapeNet [9], and 8 from SAPIEN [83] to learn ComA individually. As textures were not available in InterCap [29], making the diffusion model difficult to inpaint, we used TEXTure [66] to generate textures on the meshes using text prompt automatically generated from ChatGPT [58]. + +For quantitative evaluation, we utilize BEHAVE [2] as a ground-truth dataset to verify the quality of generated 3D HOI samples, assuming that interactions captured in lab-controlled environments exhibit the upper-bound quality of HOI. BEHAVE [2] contains video sequences of HOI scenarios for 20 object categories. Each video frame is annotated with a single 3D human and object, where each human is given as SMPL-H [44,68] format and objects as rotation and translation applied to the canonical mesh. We preprocess the dataset by transferring the SMPL-H annotations into SMPL-X format following Choutas et al. [61]. We use the 3D human-object pair in each frame as a sample, compute vertex-wise contact scores in the same way as our method, and use them as ground truth. + +Internet Search Datasets. We also collect 60 free 3D object meshes from Internet source, SketchFab [71] and learn ComA. + +# 4.2 Baselines and Metric + +As previous studies focused on contact rather than learning the overall aspect of affordances, we use contact-based affordance derived from ComA to compare with them. However, most of them output 3D contact of a single HOI situation by inferring affordance from a given HOI sample (mostly an image), while we produce an overall affordance distribution for the given 3D object. To address this, we treat the contact inferred by state-of-the-art (SOTA) methods, DECO [79] (for the human) and IAGNet [87] (for the object) from the HOI image as a single sample, and aggregate them across multiple images to construct a distribution for comparison against BEHAVE [2]. + +Specifically, we aggregate the results of DECO [79] and IAGNet [87] on (1) 2D HOI images we've generated beforehand and (2) BEHAVE [2] test images, as these represent the usage of the BEHAVE [2] object. The aggregated scores are then normalized to form a distribution (summing up to 1), considering only the relative probability of potential contact, as the original scale of the scores varies between methods. Subsequently, we compare similarity scores (SIM) [75] and mean absolute error (MAE) [81] within these distributions. + +# 4.3 Qualitative Results + +Contact for Various Objects. Similar to other studies, ComA can be derived into contact-based affordance as shown in Fig. 5. As our representation provides contact scores for all pairs of human-object vertices, we assign the contact score of each vertex as the maximum score among the pairs that include it. We can observe the conventional object usage patterns as heatmap, which fits with general human usages. Specifically, in case Fig. 5(b), where significant contact is absent, we can infer that the interaction does not frequently appear by contact. + +Orientational Tendency for Various Objects. ComA also contains information on orientational tendencies during the HOI. The last column in Fig. 5 visualizes the negated entropy of the vertex normal direction of human relative to the object as a heatmap. Fig. 5(a), Fig. 5(b), and Fig. 5(c) exhibit relatively high orientational tendency, while the rest show lower. This aligns with the fact that + +![](images/6dc32b982e0fe503cdae927533964f58d50400ace08b13ce01fbc16e72faf083.jpg) +Fig. 5: Qualitative Results. ComA can model distributions of contact, orientation, and spatial relation exhibited during the interaction between humans and novel objects. + +the human's orientation is fixed by the attachment to the back as in Fig. 5(a), and by the restrictions present on the back and side of the object as in Fig. 5(c). In Fig. 5(b), there exists an orientational tendency without attachments or restrictions, which arises due to long-distance interaction: viewing. + +Spatial Relation for Various Objects. The relative position of the human during HOI can be also derived from ComA. Fig. 5 visualizes the distribution of the human's full body positions relative to the object. Specifically, as shown in Fig. 5(d), Fig. 5(e), and Fig. 5(f), we find that the spatial relations for the object with multiple utilities and geometries (e.g., in the case of shovel, handle for holding and blade for digging) align with general human usage, demonstrating the plausibility of the results. Especially, in the case of tennis racket shown in Fig. 5(e), we can observe a relatively weak tendency in spatial relation, as the racket can move almost freely relative to humans during the interaction. + +![](images/f0661fbc5771ff412d48f6b6b69b41f374393cb90807d56cbb6d0b2e242e587f.jpg) +(a) Finding Corresponding Contact + +![](images/2011eecb84d85428aac33a08a2e8ee6fc77b994aa1d0b79b0cd28930a2b3143c.jpg) +(b) Evaluating Plausibility of Orientation +Fig. 6: Pointwise Affordance Knowledge. ComA can provide pointwise affordance knowledge such as corresponding contacts or plausible orientation of body parts. + +
MethodsSIMHuman(%)↑SIMObject(%)↑MAEHuman(×100)↓MAEObject(×100)↓
Ours52.8270.550.1260.0288
IAGNet [87] on Our 2D HOI Images-66.34-0.0329
IAGNet [87] on BEHAVECT Images-63.86-0.0353
DECO [79] on Our 2D HOI Images21.66-0.246-
DECO [79] on BEHAVECT Images17.39-0.259-
+ +Table 1: Quantitative Results on BEHAV. We report SIM, MAE to evaluate the similarity of contact distribution (normalized contact scores) on BEHAV sequences. + +Pointwise Affordance Knowledge. Different from previous studies, ComA models the distribution of relative orientation and proximity for all human-object vertex pairs, allowing it to provide pointwise affordance knowledge such as corresponding contacts and plausible orientation direction of specific body parts. As shown in Fig. 6(a), we can infer how each human parts and object parts contact. For the case of motorcycle, we can see that saddle and handle each contacts with human's hips and hands; while for the case of chair, the seat and backrest each contacts with human's hips and back. Also, we can utilize ComA to infer the probable direction of the head while using the monitor by computing the cosine similarity between the head direction and the vertex normal with highest orientational tendency. As illustrated in Fig. 6(b), we can see that the similarity maximizes when gazing towards the screen; which is a plausible head direction when interacting with the monitor. + +Ablation on Adaptive Mask Inpainting. As shown in Fig. 4, without Adaptive Mask Inpainting, the original geometry of the object changes (left) or gets replaced by hallucinated objects (right), resulting in false affordances unfaithful to the original object. Consequently, the difference between the object in 2D and the original one existing in 3D causes the uplifted 3D HOI samples to fail to represent the right usage. This shows the importance of Adaptive Mask Inpainting as it prevents 2D inpainting diffusion model from generating false affordances. + +# 4.4 Quantitative Results + +Comparison with Baselines. We evaluate the similarity of contact distribution between BEHAV [2] and each of DECO [79], IAGNet [87], and ours. We sample 2048 points from the 3D object mesh surface following IAGNet [87] and only consider the contact underneath the head for DECO [79], as DECO [79] uses SMPL [44] format, which only has vertex correspondence with SMPL-X [61] in + +
Inpainting MethodsRMSEBackground↓ mIoU↑mIoUOCclusionAware ↑
Vanilla42.230.5350.631
Adaptive Mask15.330.7060.815
+ +Table 2: Effects of Adaptive Mask Inpainting We evaluate the performance of background preservation in vanilla inpainting and our Adaptive Mask Inpainting. + +non-head regions. As shown in Tab. 1, the contact distribution from IAGNet [87] and DECO [79] using our 2D HOI images outperforms the one modeled from BEHAVE test images. This means that our generated HOI images represent the real-world affordance as well as BEHAVE [2] test images, demonstrating the efficacy of leveraging pre-trained diffusion model for generating affordance. Additionally, the contact distribution from our method outperforms the one modeled using IAGNet [87] and DECO [79] with our generated HOI images. Since we extract 3D contacts from the same image sets, we argue that our pipeline is better than other models trained on annotations for uplifting 2D affordance to 3D. + +Efficacy of Adaptive Mask Inpainting. To verify the effectiveness of our Adaptive Mask Inpainting, we compare the amount of background preservation for our generated 2D HOI images with the vanilla inpainting diffusion model. We sample a maximum of 2 inpainted images with a single human detected, per each object categories from BEHAV [2]; resulting in 691 images for vanilla inpainting and 692 images for Adaptive Mask Inpainting. We compute 3 metrics for these images; (1) $\mathrm{RMSE}_{\mathrm{Background}}$ : pixel error between images excluding the predicted human region; (2) mIoU: average IoU between object predictions in each inpainted image and original object image; (3) mIoU $_{\mathrm{Occlusion\_Aware}}$ : same metric as (2) but excluding the predicted human region during computation. To ensure fairness, we use open vocabulary segmentation model [35,43] to predict detailed human/object segmentation. As shown in Tab. 2, we demonstrate that our adaptive mask algorithm better preserves the background, not only pixelwise $(\mathrm{RMSE}_{\mathrm{Background}})$ , but also semantic-region-wise (mIOU, mIOU $_{\mathrm{Occlusion\_Aware}}$ ) than the vanilla inpainting baseline. This means that the generated images from adaptive mask inpainting are more likely to preserve the original object. + +# 4.5 Application + +As ComA is a newly proposed affordance representation, we demonstrate its potential by showcasing an application. One possible task is to reconstruct HOI by optimizing the human pose and position to interact with an object. As shown in Fig. 7(a), it is nearly impossible to achieve the task with traditional methods which relies on contact affordance only. In contrast, ComA provides both relative orientation and contact during the HOI, making it possible to generate plausible 3D HOI samples from T-posed humans. For the objects with multiple modes of affordances, we apply mode seeking algorithm [11] on 3D HOI samples with their vertex normals and learn ComA separately; for example, swing in Fig. 7(a). + +However, ComA learned on a specific object cannot be directly applied to other objects with different geometries. To address this, we transfer ComA through the process of (1) object canonicalization and (2) finding vertex cor- + +![](images/52af60b288a31d15ff981a926274ef539f88e6bf5316891784d2e77fed48469d.jpg) +Fig. 7: Application. ComA can reconstruct HOI through optimization, which can be scaled to other objects in the same category by transferring such knowledge between them. + +respondences. Fig. 7(b) provides an example of transferring ComA learned from a specific motorcycle to other motorcycles with different geometries. We use objects from ShapeNet [9] which are already canonicalized and apply CPNet [45] to find vertex correspondences automatically for transferring the vertices with contact. As object canonicalization and vertex correspondence can each transfer macro orientation information and local contact information, we can transfer ComA and scale the human optimization task to other objects in the same category without any additional learning. + +# 5 Conclusion + +In this paper, we point out the importance of learning non-contact patterns in HOIs and propose a novel affordance representation called ComA which models both contact and non-contact patterns. To learn ComA, we utilize a pre-trained 2D diffusion model which already possesses prior knowledge of affordance. To leverage the affordance knowledge, we designed a Rendering-Inpainting-Uplifting pipeline to generate 3D HOI samples and learn ComA from the samples. The results show that ComA effectively models the common usage of objects in terms of contact, orientational, and spatial relations. Within the pipeline, we propose Adaptive Mask Inpainting, which allows insertion of human without damaging the original image, where we evaluate its efficacy both qualitatively and quantitatively. We expect ComA to be used as a prior knowledge for various downstream tasks, where we demonstrate its potential through human optimization framework and the possibility of transfer to various objects. + +# Acknowledgements + +This work was supported by Naver Webtoon. The work of SNU members was also supported by NRF grant funded by the Korean government (MSIT) (No. 2022R1A2C2092724 and No. RS-2023-00218601), and IITP grant funded by the Korean government (MSIT) [RS-2021-II211343, AI Graduate School Program (SNU)]. H. Joo is the corresponding author. + +# References + +1. Bahl, S., Mendonca, R., Chen, L., Jain, U., Pathak, D.: Affordances from human videos as a versatile representation for robotics. In: CVPR (2023) +2. Bhatnagar, B.L., Xie, X., Petrov, I., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Behave: Dataset and method for tracking human object interactions. In: CVPR (2022) +3. Brahmbhatt, S., Ham, C., Kemp, C.C., Hays, J.: Contactdb: Analyzing and predicting grasp contact via thermal imaging. In: CVPR (2019) +4. Brahmbhatt, S., Handa, A., Hays, J., Fox, D.: Contactgrasp: Functional multifinger grasp synthesis from contact. In: IROS (2019) +5. Brahmbhatt, S., Tang, C., Twigg, C.D., Kemp, C.C., Hays, J.: Contactpose: A dataset of grasps with object contact and hand pose. In: ECCV (2020) +6. Cai, M., Kitani, K.M., Sato, Y.: Understanding hand-object manipulation with grasp types and object attributes. In: RSS (2016) +7. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV (2020) +8. Chai, L., Zhu, J.Y., Shechtman, E., Isola, P., Zhang, R.: Ensembling with deep generative views. In: CVPR (2021) +9. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 (2015) +10. Chen, Y., Huang, S., Yuan, T., Qi, S., Zhu, Y., Zhu, S.C.: Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In: ICCV (2019) +11. Comaniciu, D., Meer, P.: Mean shift: a robust approach toward feature space analysis. In: IEEE TPAMI (2002) +12. Deng, S., Xu, X., Wu, C., Chen, K., Jia, K.: 3d affordancenet: A benchmark for visual object affordance understanding. In: CVPR (2021) +13. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: CVPR (2020) +14. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. In: CACM (1981) +15. Gao, W., Tedrake, R.: kpm-sc: Generalizable manipulation planning using keypoint affordance and shape completion. In: ICRA (2021) +16. Garg, M., Garg, P., Vohra, R.: Advanced fibonacci sequence with golden ratio. In: IJSER (2014) +17. Gibson, J.J.: The Ecological Approach to Visual Perception. Houghton Mifflin (1979) + +18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014) +19. Grady, P., Tang, C., Twigg, C.D., Vo, M., Brahmbhatt, S., Kemp, C.C.: ContactOpt: Optimizing contact to improve grasps. In: CVPR (2021) +20. Han, S., Joo, H.: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In: ICCV (2023) +21. Hao, Z., Mallya, A., Belongie, S., Liu, M.Y.: Gancraft: Unsupervised 3d neural rendering of apache worlds. In: ICCV (2021) +22. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV (2019) +23. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: CVPR (2021) +24. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? In: arXiv:2210.07574 (2022) +25. Hermans, T., Rehg, J.M., Bobick, A.: Affordance prediction via learned object attributes. In: ICRA (2011) +26. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: arXiv:2208.01626 (2022) +27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +28. Huang, C.H.P., Yi, H., Hoschle, M., Safroshkin, M., Alexiadis, T., Polikovsky, S., Scharstein, D., Black, M.J.: Capturing and inferring dense full-body human-scene contact. In: CVPR (2022) +29. Huang, Y., Taheri, O., Black, M.J., Tzionas, D.: Intercap: Joint markerless 3d tracking of humans and objects in interaction from multi-view rgb-d images. In: IJCV (2024) +30. Jahanian, A., Puig, X., Tian, Y., Isola, P.: Generative models as a data source for multiview representation learning. In: arXiv:2106.05258 (2021) +31. Jian, J., Liu, X., Li, M., Hu, R., Liu, J.: Affordpose: A large-scale dataset of handobject interactions with affordance-driven hand pose. In: ICCV (2023) +32. Jiang, W., Kolotouros, N., Pavlakos, G., Zhou, X., Daniilidis, K.: Coherent reconstruction of multiple humans from a single image. In: CVPR (2020) +33. Jiang, Y., Jiang, S., Sun, G., Su, Z., Guo, K., Wu, M., Yu, J., Xu, L.: Neuralhofusion: Neural volumetric rendering under human-object interactions. In: CVPR (2022) +34. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) +35. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) +36. Kirillov, A., Wu, Y., He, K., Girshick, R.: Pointrend: Image segmentation as rendering. In: CVPR (2020) +37. Kulal, S., Brooks, T., Aiken, A., Wu, J., Yang, J., Lu, J., Efros, A.A., Singh, K.K.: Putting people in their place: Affordance-aware human insertion into scenes. In: CVPR (2023) +38. Lee, Y.J., Grauman, K.: Predicting important objects for egocentric video summarization. In: IJCV (2015) +39. Levine, S., Shah, D.: Learning robotic navigation from experience: principles, methods and recent results. In: Philos. Trans. R. Soc. B (2022) + +40. Li, G., Jampani, V., Sun, D., Sevilla-Lara, L.: Locate: Localize and transfer object parts for weakly supervised affordance grounding. In: CVPR (2023) +41. Li, L., Dai, A.: GenZI: Zero-shot 3D human-scene interaction generation. In: CVPR (2024) +42. Li, Z., Sedlar, J., Carpentier, J., Laptev, I., Mansard, N., Sivic, J.: Estimating 3d motion and forces of person-object interactions from monocular video. In: CVPR (2019) +43. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: arXiv:2303.05499 (2023) +44. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. In: Proc. ACM SIGGRAPH Asia (2015) +45. Lou, Y., You, Y., Li, C., Cheng, Z., Li, L., Ma, L., Wang, W., Lu, C.: Human correspondence consensus for 3d object semantic understanding. In: ECCV (2020) +46. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022) +47. Mao, C., Cha, A., Gupta, A., Wang, H., Yang, J., Vondrick, C.: Generative interventions for causal learning. In: CVPR (2021) +48. McCool, M., Fiume, E.: Hierarchical poisson disk sampling distributions. In: Graphics Interface (1992) +49. Melas-Kyriazi, L., Rupprecht, C., Laina, I., Vedaldi, A.: Finding an unsupervised image segmenter in each of your deep generative models. In: arXiv:2105.08127 (2021) +50. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022) +51. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: COAP: Compositional articulated occupancy of people. In: CVPR (2022) +52. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019) +53. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023) +54. Monszpart, A., Guerrero, P., Ceylan, D., Yumer, E., Mitra, N.J.: imapper: interaction-guided scene mapping from monocular videos. In: ACM TOG (2019) +55. Moon, G., Choi, H., Lee, K.M.: Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In: CVPRW (2022) +56. Nagarajan, T., Grauman, K.: Learning affordance landscapes for interaction exploration in 3d environments. In: NeurIPS (2020) +57. Oleynikova, H., Millane, A., Taylor, Z., Galceran, E., Nieto, J., Siegwart, R.: Signed distance fields: A natural representation for both mapping and planning. In: RSS Workshop (2016) +58. OpenAI: Chatgpt: Optimizing language models for dialogue. https://openai.com/blog/chatgpt/ (2023) +59. OpenAI: Gpt-4v(ision) system card. https://openai.com/research/gpt-4v-system-card (2023) +60. Pan, X., Dai, B., Liu, Z., Loy, C.C., Luo, P.: Do 2d gans know 3d shape? unsupervised 3d shape reconstruction from 2d image gans. In: arXiv:2011.00844 (2020) +61. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019) + +62. Peebles, W., Zhu, J.Y., Zhang, R., Torralba, A., Efros, A.A., Shechtman, E.: Gansupervised dense visual alignment. In: CVPR (2022) +63. Petrov, I.A., Marin, R., Chibane, J., Pons-Moll, G.: Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In: CVPR (2023) +64. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: arXiv:2204.06125 (2022) +65. Rempe, D., Guibas, L.J., Hertzmann, A., Russell, B., Villegas, R., Yang, J.: Contact and human dynamics from monocular video. In: ECCV (2020) +66. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. In: Proc. ACM SIGGRAPH (2023) +67. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +68. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. In: Proc. ACM SIGGRAPH Asia (2017) +69. Savva, M., Chang, A.X., Hanrahan, P., Fisher, M., Nießner, M.: Pigraphs: learning interaction snapshots from observations. In: ACM TOG (2016) +70. Shannon, C.E.: A mathematical theory of communication. In: ACM SIGMOBILE (2001) +71. Sketchfab: Sketchfab. https://sketchfab.com/ (2023) +72. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) +73. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: arXiv:2010.02502 (2020) +74. Sun, G., Chen, X., Chen, Y., Pang, A., Lin, P., Jiang, Y., Xu, L., Yu, J., Wang, J.: Neural free-viewpoint performance rendering under complex human-object interactions. In: ACM MM (2021) +75. Swain, M.J., Ballard, D.H.: Color indexing. In: IJCV (1991) +76. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: ECCV (2020) +77. Tanaka, F.H.K.d.S., Aranha, C.: Data augmentation using gans. In: arXiv:1904.09135 (2019) +78. Trabucco, B., Doherty, K., Gurinas, M., Salakhutdinov, R.: Effective data augmentation with diffusion models. In: arXiv:2302.07944 (2023) +79. Tripathi, S., Chatterjee, A., Passy, J.C., Yi, H., Tzionas, D., Black, M.J.: DECO: Dense estimation of 3D human-scene contact in the wild. In: ICCV (2023) +80. Tritrong, N., Rewatbowornwong, P., Suwajanakorn, S.: Repurposing gans for one-shot semantic part segmentation. In: CVPR (2021) +81. Willmott, C.J., Matsuura, K.: Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. In: Climate Research (2005) +82. Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetdm: Synthesizing data with perception annotations using diffusion models. In: NeurIPS (2023) +83. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: CVPR (2020) +84. Xu, J., Zhu, S., Guo, H., Wu, S.: Automated labeling for robotic autonomous navigation through multi-sensory semi-supervised learning on big data. In: IEEE TBD (2021) +85. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In: ICCV (2023) + +86. Yang, L., Zhan, X., Li, K., Xu, W., Li, J., Lu, C.: CPF: Learning a contact potential field to model the hand-object interaction. In: ICCV (2021) +87. Yang, Y., Zhai, W., Luo, H., Cao, Y., Luo, J., Zha, Z.J.: Grounding 3d object affordance from 2d interactions in images. In: ICCV (2023) +88. Yang, Y., Zhai, W., Luo, H., Cao, Y., Zha, Z.J.: Lemon: Learning 3d human-object interaction relation from 2d images. In: CVPR (2024) +89. Ye, Y., Li, X., Gupta, A., Mello, S.D., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: CVPR (2023) +90. Yi, H., Huang, C.H.P., Tzionas, D., Kocabas, M., Hassan, M., Tang, S., Thies, J., Black, M.J.: Human-aware object placement for visual environment reconstruction. In: CVPR (2022) +91. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 3DV (2020) +92. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: ECCV (2022) +93. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: CVPR (2020) +94. Zhang, Y., Chen, W., Ling, H., Gao, J., Zhang, Y., Torralba, A., Fidler, S.: Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In: ICLR (2021) +95. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021) +96. Zhou, K., Bhatnagar, B.L., Lenssen, J.E., Pons-Moll, G.: Toch: Spatio-temporal object correspondence to hand for motion refinement. In: ECCV (2022) \ No newline at end of file diff --git a/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/images.zip b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d7d606ce5db80d11e7412b85804dcfbea95935c0 --- /dev/null +++ b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fd1553d2ec99e4f98e818671e6d3535b3b2488e6c1c530f0e9334c1654ebf88 +size 491160 diff --git a/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/layout.json b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eea6a502acdb60669b2f09a0b1d6007af97b3e2d --- /dev/null +++ b/2024/Beyond the Contact_ Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models/layout.json @@ -0,0 +1,10041 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 163 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 163 + ], + "type": "text", + "content": "Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "spans": [ + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "type": "text", + "content": "Hyeonwoo " + }, + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "type": "inline_equation", + "content": "\\mathrm{Kim}^{1*}" + }, + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "type": "text", + "content": ", Sookwan Han" + }, + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "type": "text", + "content": ", Patrick Kwon" + }, + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "type": "text", + "content": ", and Hanbyul Joo" + }, + { + "bbox": [ + 153, + 185, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 249, + 207, + 364, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 207, + 364, + 219 + ], + "spans": [ + { + "bbox": [ + 249, + 207, + 364, + 219 + ], + "type": "text", + "content": "1 Seoul National University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 263, + 220, + 350, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 220, + 350, + 229 + ], + "spans": [ + { + "bbox": [ + 263, + 220, + 350, + 229 + ], + "type": "text", + "content": "2 Naver Webtoon AI" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 237, + 482, + 326 + ], + "blocks": [ + { + "bbox": [ + 133, + 237, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 133, + 237, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 482, + 326 + ], + "type": "image", + "image_path": "36ebae95e9d9b0eb350d9c84445a18c8b0280fec06dd412637ea18e4232cbb5b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 333, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 132, + 333, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 482, + 367 + ], + "type": "text", + "content": "Fig. 1: Given a 3D object, we generate numerous 3D Human-Object Interaction (HOI) samples using text prompts, and learn a novel affordance representation called Comprehensive Affordance (ComA) which models both contact and non-contact HOI patterns." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 376, + 453, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 376, + 453, + 628 + ], + "spans": [ + { + "bbox": [ + 159, + 376, + 453, + 628 + ], + "type": "text", + "content": "Abstract. Understanding the inherent human knowledge in interacting with a given environment (e.g., affordance) is essential for improving AI to better assist humans. While existing approaches primarily focus on human-object contacts during interactions, such affordance representation cannot fully address other important aspects of human-object interactions (HOIs), i.e. patterns of relative positions and orientations. In this paper, we introduce a novel affordance representation, named Comprehensive Affordance (ComA). Given a 3D object mesh, ComA models the distribution of relative orientation and proximity of vertices in interacting human meshes, capturing plausible patterns of contact, relative orientations, and spatial relationships. To construct the distribution, we present a novel pipeline that synthesizes diverse and realistic 3D HOI samples given any 3D target object mesh. The pipeline leverages a pre-trained 2D inpainting diffusion model to generate HOI images from object renderings and lifts them into 3D. To avoid the generation of false affordances, we propose a new inpainting framework, Adaptive Mask Inpainting. Since ComA is built on synthetic samples, it can extend to any object in an unbounded manner. Through extensive experiments, we demonstrate that ComA outperforms competitors that rely on human annotations in modeling contact-based affordance. Importantly, we also showcase the potential of ComA to reconstruct human-object interactions in 3D through an optimization framework, highlighting its advantage in incorporating both contact and non-contact properties." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 638, + 453, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 638, + 453, + 650 + ], + "spans": [ + { + "bbox": [ + 160, + 638, + 453, + 650 + ], + "type": "text", + "content": "Keywords: Affordance " + }, + { + "bbox": [ + 160, + 638, + 453, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 453, + 650 + ], + "type": "text", + "content": " Human-Object Interaction " + }, + { + "bbox": [ + 160, + 638, + 453, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 453, + 650 + ], + "type": "text", + "content": " Diffusion Models" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 653, + 253, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 253, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 253, + 665 + ], + "type": "text", + "content": "*Indicates equal contribution" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 147, + 481, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 147, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 147, + 481, + 278 + ], + "type": "text", + "content": "Humans possess an innate ability to perceive the functionalities provided by the environment or objects and efficiently utilize them—a concept known as affordance [17]. Affordance incorporates a variety of patterns typically observable in human-object interactions (HOIs), including not only plausible physical contact but also spatial and orientational non-contact relationships. For example, while using a laptop, the orientation and distance patterns between human face or other body parts and the screen can be considered part of affordance, alongside the hand regions touching the laptop keyboard. Teaching such affordance knowledge to machines is crucial for enabling them to better understand or mimic human behaviors. Consequently, numerous studies have been conducted to implement such affordance knowledge into AI and robotics [1,20,23,37,40,63,79,87,89]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 280, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 481, + 376 + ], + "type": "text", + "content": "Most previous approaches however, primarily focus on a limited spectrum of affordances. Some approaches represent the human and object regions in contact using contact scores on 2D images [1,40], 3D objects [87], and 3D humans [23,79] using direct supervision from human annotations. Others represent affordances by inferring [63] or generating [37,89] plausible human-object pairs in contact-oriented object categories. While a few recent studies tackle learning the spatial relations [20,88] between a human and an object, these methods mainly focus on object categories where physical contact is dominant than non-contact patterns." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 481, + 509 + ], + "type": "text", + "content": "In this paper, we introduce Comprehensive Affordance (ComA), a novel representation of an affordance encompassing the various aspects of human-object interaction, including orientational tendencies and relative positions. The motivation for our design arises from the observation that there exist typical 3D relations between entire body parts and objects during HOIs. For instance, human faces, torsos, and arms often exhibit specific distances and orientations, with some variability, when interacting with certain object categories, as shown in Fig. 2. Our ComA is designed to model these patterns by constructing a distribution between each pair of object surface points and human surface points in 3D. This distribution represents the likelihood of human body parts displaying specific spatial and orientational relations with object points, as shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "content": "Learning ComA requires various 3D HOI samples showing the way humans use the target object in 3D, from which pairwise distributions of 3D HOI relations can be constructed. Manually annotating such cues is challenging, compared to the annotations focusing solely on contact regions [40, 63, 79, 85, 87, 88]. As a key solution, we present a pipeline to synthesize a large-scale synthetic 3D HOI samples leveraging a pre-trained 2D diffusion model. Interestingly, we observe that a pre-trained 2D diffusion model captures the affordance knowledge in the form of 2D images, as shown in Fig. 2, where the images are synthesized from text prompts. However, leveraging these 2D HOI cues for construction of ComA (which requires 3D HOI samples) presents the following challenges: (1) the diffusion model needs to be applied for inpainting plausible humans without altering the original shape and pose of the target object, and (2) 3D HOI samples should be lifted from the synthesized 2D images. To address these challenges, we present" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 305, + 201 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 305, + 201 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 305, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 305, + 201 + ], + "type": "image", + "image_path": "d78d3eae5f629e72c5458a6734b1ddd26aa640641cbe34a733b21046f564dfb8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 202, + 285, + 213 + ], + "lines": [ + { + "bbox": [ + 153, + 202, + 285, + 213 + ], + "spans": [ + { + "bbox": [ + 153, + 202, + 285, + 213 + ], + "type": "text", + "content": "Text Prompt: 1 person using a laptop" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 482, + 201 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 482, + 201 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 482, + 201 + ], + "type": "image", + "image_path": "cc1e59e40fb8408604b261e917e8179ad2c0d9e567e71a8a22c8a53b30813162.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 202, + 465, + 213 + ], + "lines": [ + { + "bbox": [ + 323, + 202, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 323, + 202, + 465, + 213 + ], + "type": "text", + "content": "Text Prompt: 1 person using a telescope" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 217, + 480, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 480, + 262 + ], + "type": "text", + "content": "Fig. 2: Typically, people (1) view the screen (2) from relatively distant distance while using a laptop (Left), whereas they (1) peer into it (2) from a close distance while using a telescope (Right). Pre-trained diffusion model has a knowledge of these (1) orientational and (2) spatial relation between human and object during interaction." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 480, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 293 + ], + "type": "text", + "content": "Adaptive Mask Inpainting to insert humans without altering the appearance of target objects in 2D, along with a 3D lifting pipeline from 2D HOI cues." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 388 + ], + "type": "text", + "content": "We demonstrate the efficacy of our approach by learning ComA on 100 3D object meshes in diverse categories collected from various sources [2,9,29,52,71, 83]. We compare our ComA with existing approaches [79,87] on the BEHAV dataset [2], demonstrating that ours learned from synthetic cues outperforms competitors trained on manual annotations for contact-based affordance. Additionally, we verify the advantage of ComA against contact-only representations, in the scenario of reconstructing HOIs via an optimization framework. We will release the results and our source code1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 389, + 481, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 473 + ], + "type": "text", + "content": "In summary, our main contributions can be summarized as follows: (1) we introduce a new representation ComA, which models the distribution of both explicit contact and non-contact patterns during HOIs; (2) we present a pipeline to build ComA from synthetic datasets by leveraging a pre-trained 2D diffusion model, scalable to any 3D object without requiring laborious annotations; and (3) we propose Adaptive Mask Inpainting, a method for preserving the original context of an image in using inpainting diffusion model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 491, + 237, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 237, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 237, + 503 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 514, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 647 + ], + "type": "text", + "content": "Learning Visual Affordance. Affordance was introduced by Gibson [17] as a set of functionalities that the environment furnishes to an agent (e.g., human, animal). The concept extends to the vision and robotics area, where the focus is on teaching embodied agents to interact with scenes [1, 15, 39, 56, 84], mimicking human-object [2, 7, 10, 13, 23, 28, 33, 42, 54, 63, 65, 69, 74, 91, 92], or hand-object [3-5, 76, 96] interactions. Earlier methods focused on learning action category labels [6, 25, 38] with bounding boxes but lacked a detailed description of affordance. Due to the limited affordance information available from text descriptions or labels, some studies depicted affordance as contact regions or heatmaps on objects [1, 40, 87] and humans [23, 79]. However, learning affordances beyond the contact was challenging since existing datasets created through a multi-camera system [2, 29]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 448, + 102 + ], + "type": "text", + "content": "Discovering Comprehensive Affordance for 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 653, + 289, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 289, + 665 + ], + "type": "text", + "content": "1https://github.com/snuvclab/coma" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "motion capture [22], and manual process [12,31,79,87] focused only on annotating contact information rather than capturing the holistic aspect of HOI. Following studies generate plausible HOI samples in 2D [37,89] or 3D [85,93] which has the potential for learning such broader HOI knowledge, but do not address the problem directly. In contrast, our method generates 3D HOI samples and extracts ComA for modelling beyond the contact." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 191, + 482, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 347 + ], + "type": "text", + "content": "Data Synthesis for Learning. There have been approaches to supplement data in fields lacking efficient annotation process by leveraging generative models, irrespective of the research area. Many methods employ GANs [18, 34] or diffusion models [27, 67, 73] to augment data [24, 78] for various vision tasks, including perception and representation learning [82], classification [8, 47, 77], segmentation [49, 80, 95], dense visual alignment [62], and further extending to 3D tasks, such as neural rendering [21, 94], shape reconstruction [60], and so forth. The major challenge in leveraging synthesized data (or generative models for cues) is controllability, as the generator, even when conditioned, typically produces free samples. Ali et al. [30] presents a method to control the viewpoint of the synthesized image by modeling the viewpoint-free latent space. CHORUS [20] applies cascaded filtering to the generated dataset to learn 3D human-object spatial relations, similar to our approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 483, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 483, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 483, + 518 + ], + "type": "text", + "content": "Diffusion Model for Synthesizing HOI Images. While diffusion models [27, 72, 73] excel at generating realistic images [64, 67], few address the generation of 2D human-object pairs (i.e., HOI images). Inpainting models [46, 67], while a common choice for human insertion, often compromise scene details, resulting in images with false affordance. Image editing techniques [26, 50, 53] may offer an alternative but tend to prioritize style changes and struggle with creating new geometry. Recently, Ye et al. [89] employs an additional layout network to determine the inpainting region, and Kulal et al. [37] trains a diffusion model on video clips to generate HOI images. In contrast to these approaches, we create HOI images by inserting humans while maintaining the context of the original image, without additional network training. Concurrent work by Li et al. [41] introduces Dynamic Mask Inpainting, which uses attention masks to change inpainting masks over timesteps to insert human into a 3D scene, while we explicitly apply segmentation model to adapt the inpainting mask." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 552, + 203, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 203, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 203, + 565 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Given an input 3D object mesh, our goal is to learn our novel affordance representation, ComA, which models the distribution of relative proximity and orientation between human and the object surface points. Our pipeline begins with creating 2D HOI images (Sec. 3.1). These images are then lifted to 3D, resulting large-scale synthetic 3D HOI samples (Sec. 3.2). From these 3D samples, we finally learn ComA (Sec. 3.3). An overview of our method is shown in Fig. 3." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 191 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 191 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 191 + ], + "type": "image", + "image_path": "f875f4cc7482eba382bc3cb312decb0bf0b4096204ce340eb2c2eb9d0ad60f0b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 193, + 405, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 193, + 405, + 201 + ], + "spans": [ + { + "bbox": [ + 209, + 193, + 405, + 201 + ], + "type": "text", + "content": "Learning Comprehensive Affordance from Generated 3D HOI Samples" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 201, + 480, + 326 + ], + "blocks": [ + { + "bbox": [ + 134, + 201, + 480, + 326 + ], + "lines": [ + { + "bbox": [ + 134, + 201, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 134, + 201, + 480, + 326 + ], + "type": "image", + "image_path": "19a5cb419826cb1fc8f8e4fa4874d771501f7ae1d96f6a9a6f8ffaca6ec9dd07.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 331, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 397 + ], + "type": "text", + "content": "Fig. 3: Method Overview. Our method can be divided into two parts: (1) Generating 3D HOI Samples and (2) Learning ComA from Generated 3D HOI Samples. In the first step, we utilize an inpainting diffusion model with our Adaptive Mask Inpainting to create 2D HOI images, and generate 3D HOI samples via uplifting pipeline. In the second step, the generated 3D HOI samples are aggregated to create distributions for relative proximity and orientation, which can be derived into various affordance forms." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 402, + 295, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 295, + 415 + ], + "type": "text", + "content": "3.1 2D HOI Image Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "text", + "content": "Rendering Object from Multi-Viewpoints. Given a 3D object, we first render it from multiple viewpoints. We place the object and camera differently for two object types: static and dynamic. For objects that are presumed to be fixed on the ground during interaction (e.g., chair), we fix the object on the ground plane, whereas objects with dynamic pose during interaction (e.g., cup) are perturbed with sampled rotation and translation. Type of the object is inferred automatically by rendering the object at any direction first and then utilize vision-language model [59] for predicting. The rendering process can be fully automated, but we empirically find it beneficial to leverage a small amount of human labor, such as restricting the range of perturbation due to the inherent bias of pre-trained diffusion models. For rendering, we place weak perspective cameras " + }, + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\{\\varPi_c\\}_{c=1}^N" + }, + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "text", + "content": " around the object with equal azimuth interval and same elevation, where " + }, + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "text", + "content": " is set as 8 for static, 40 for dynamic objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Inpainting Mask Selection. After rendering the images, we select the masks to inpaint a human interacting with the object. While our inpainting pipeline is quite robust to initial masks, we found that initial mask selection is beneficial for avoiding failure cases such as generating hallucinated objects or ambiguous interactions. Specifically, we move sliding windows around the object and choose the windows whose Intersection over Union (IoU) with the object segmentation mask falls within a specific range. See Supp. Mat. for more details." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 448, + 102 + ], + "type": "text", + "content": "Discovering Comprehensive Affordance for 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 478, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 478, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 478, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 478, + 205 + ], + "type": "image", + "image_path": "7b5047e38edee739660cfe10ce09d4b33f4fce88ba1b11be9c72e0e0bcdfb089.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 210, + 482, + 234 + ], + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 234 + ], + "type": "text", + "content": "Fig. 4: Adaptive Mask Inpainting. Without adaptive mask inpainting, the original object is damaged when inserting humans, resulting false affordances." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 243, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 304 + ], + "type": "text", + "content": "Prompt Generation. Inspired by CHORUS [20], we automatically generate the prompts that describe the possible interactions with humans and the given object, which are taken as guidance for inpainting. Specifically, we utilize the vision-language model [59] to infer the HOI prompt using the rendered object image and predefined template. We generate 3 prompts per given object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "text", + "content": "Adaptive Mask Inpainting for Human Insertion. We synthesize 2D images of human-object interaction by inpainting the human into rendered object images using publicly available text-conditional inpainting diffusion models [67]. However, the challenge arises as the object geometry and details within the mask region are not preserved during inpainting, resulting false affordance. To mitigate this problem, we present Adaptive Mask Inpainting to progressively specify the inpainting region over diffusion timesteps. Let the sequence of denoised image latent be " + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\{x_{t}\\}_{t = T}^{0}" + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "text", + "content": " is the fully random noise and " + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "text", + "content": " is the denoised latent. Inspired by the observation that the quality of the predicted denoised image " + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "text", + "content": " improves over progress of timestep, and creates a low-level structure of the target prompt (in this case, human) even at the early steps as shown in Fig. 4, we aim to adapt the inpainting region over timesteps by spatially discovering the mask from the low-level structure. Specifically, we apply PointRend [36] to decoded latent " + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(\\hat{x}_0)" + }, + { + "bbox": [ + 130, + 304, + 482, + 485 + ], + "type": "text", + "content": " at specific timesteps to predict the human mask, and consequently use as the mask for the next denoising step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 506, + 304, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 304, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 304, + 519 + ], + "type": "text", + "content": "3.2 Lifting 2D Affordance to 3D" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 531, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 579 + ], + "type": "text", + "content": "From the previous section, we obtain a set of 2D HOI images " + }, + { + "bbox": [ + 130, + 531, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\{I_d\\}_{d=1}^D" + }, + { + "bbox": [ + 130, + 531, + 482, + 579 + ], + "type": "text", + "content": ". Even though we have multiple views with geometrically consistent object renderings, the inserted humans have diversity without multi-view consistency, and thus it is non-trivial to directly lift 3D humans from these 2D images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 580, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 628 + ], + "type": "text", + "content": "To address this, we first utilize a single view 3D human prediction model to obtain a corresponding 3D human pose and shape of the image, and then estimate the depth of the human with weak auxiliary cue provided by the set of 2D HOI images, generating a 3D HOI sample corresponds to the image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Single View 3D Human Prediction. To uplift the 2D HOI images into 3D, we first predict the pose and shape of 3D humans from the images. We apply off-the-shelf 3D human prediction model [55] (denoted as " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{human}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ") to predict" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 334, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 334, + 129 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 334, + 129 + ], + "type": "text", + "content": "SMPL-X humans [61] from generated images:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 137, + 481, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 137, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 222, + 137, + 481, + 152 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {\\theta} _ {d}, \\boldsymbol {\\beta} _ {d}, \\mathbf {j} _ {d}, \\boldsymbol {R} _ {d} ^ {h}, \\mathbf {s} _ {d} ^ {h}, x _ {d}, y _ {d} \\right\\} = \\mathbf {F} _ {\\text {h u m a n}} (I _ {d}) \\tag {1}", + "image_path": "f02fe63546d0c2d613612d9b79f4e9193c00d4441a846378089479ee37dcfa3b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_d \\in \\mathbb{R}^{54 \\times 3}, \\pmb{\\beta}_d \\in \\mathbb{R}^{10}, \\mathbf{j}_d \\in \\mathbb{R}^{67 \\times 3}" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": " are the predicted SMPL-X pose, shape, and joints (see Supp. Mat. for the increased number of joints), respectively. " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\pmb{R}_d^h \\in \\mathrm{SO}(3)" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": " is global rotation, and the " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_d^h, x_d, y_d" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": " are scale, and " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": " direction offsets representing the parameters of weak perspective camera " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\varPi_{\\gamma(d)}" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": " assigned to image " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\gamma(\\cdot)" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": " is the function that maps the index of the 2D HOI image to the index of the weak perspective camera from which it is rendered. We additionally define " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{j}_{d^\\dagger}" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": " as joints in world coordinate, obtained by transforming inferred joints " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{j}_d" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": " with the camera rotation and translation of " + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "inline_equation", + "content": "\\varPi_{\\gamma(d)}" + }, + { + "bbox": [ + 130, + 159, + 482, + 279 + ], + "type": "text", + "content": ". As 3D human lies within the weak perspective camera framework in world coordinate, there remains depth ambiguity to solve." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 279, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 388 + ], + "type": "text", + "content": "Depth Optimization using Weak Auxiliary Cue. To address the depth ambiguity of 3D human obtained from reference image, we leverage 2D HOI images generated in advance. Since the human poses in images are geometrically inconsistent, we select the largest subset inliers " + }, + { + "bbox": [ + 130, + 279, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 279, + 482, + 388 + ], + "type": "text", + "content": " among them which are \"semiconsistent\" with the reference image by applying RANSAC [14] in terms of joint re-projection error. The joint triangulation is performed on two images, including the reference image. Note that we generate sufficiently large amounts of 2D HOI images to ensure the existence of the inlier set. The inliers " + }, + { + "bbox": [ + 130, + 279, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 279, + 482, + 388 + ], + "type": "text", + "content": " is used to estimate the human depth " + }, + { + "bbox": [ + 130, + 279, + 482, + 388 + ], + "type": "inline_equation", + "content": "z_{d}" + }, + { + "bbox": [ + 130, + 279, + 482, + 388 + ], + "type": "text", + "content": " by optimizing the joint re-projection loss defined below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 184, + 396, + 481, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 396, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 184, + 396, + 481, + 424 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e - p r o j e c t i o n}} = \\frac {1}{| \\mathcal {I} |} \\sum_ {I _ {\\delta} \\in \\mathcal {I}} \\| \\Pi_ {\\gamma (\\delta)} \\left(\\mathbf {j} _ {d ^ {\\dagger}} + z _ {d} \\boldsymbol {f} _ {d}\\right) - \\Pi_ {\\gamma (\\delta)} \\left(\\mathbf {j} _ {\\delta^ {\\dagger}}\\right) \\| ^ {2} \\tag {2}", + "image_path": "23733baa57c7003168f0b842ab30b4131e91d9406cfd05b0c68694750dc8421d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\pmb{f}_d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "content": " is the normalized direction of the camera " + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\varPi_{\\gamma(d)}" + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "content": ". Inspired by Jiang et al. [32], we leverage the human segmentation masks to reason about occlusion and promote a initial depth for human. We initialize SMPL-X [61] models along the camera direction and select the model with the best IOU between the rendered mask (regarding occlusion) and the segmented region to be an initial of optimization. This is useful when the optimal object position is in between two human body parts along the camera direction (e.g., a motorcycle body between the legs while riding) as we use a collision loss term during optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 526, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 586 + ], + "type": "text", + "content": "The collision loss is used to estimate fine-grained depth, as the collision cue can provide the physical plausibility of 3D HOI samples. We adopt collision loss from COAP [51] where the human body is represented as SDF [57] field. We set object vertices as query points to reduce the collision between human and object, namely " + }, + { + "bbox": [ + 130, + 526, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{collision}}" + }, + { + "bbox": [ + 130, + 526, + 482, + 586 + ], + "type": "text", + "content": ". We optimize depth " + }, + { + "bbox": [ + 130, + 526, + 482, + 586 + ], + "type": "inline_equation", + "content": "z_{d}" + }, + { + "bbox": [ + 130, + 526, + 482, + 586 + ], + "type": "text", + "content": " to minimize the total loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 230, + 597, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 597, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 230, + 597, + 481, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e - p r o j e c t i o n}} + \\lambda_ {\\text {c o l l i s i o n}} \\mathcal {L} _ {\\text {c o l l i s i o n}} \\tag {3}", + "image_path": "512ab652c34c69622ae1151050a1305d9ce5d5c3394774e76dd0124090c790b1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Filtering. We filter out potentially low-quality 3D human samples in the cases of: (1) human rendering (regarding occlusion) and predicted human segmentation do not overlap much; (2) number of inliers during RANSAC [14] is small; (3) human and object collides significantly in 3D." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 448, + 103 + ], + "type": "text", + "content": "Discovering Comprehensive Affordance for 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 345, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 345, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 345, + 129 + ], + "type": "text", + "content": "3.3 Learning Comprehensive Affordance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 206 + ], + "type": "text", + "content": "Our main objective is to learn ComA, a new affordance representation which encompasses orientational and spatial affordance, along with contact-based affordance. In order to learn ComA from the previously generated 3D HOI samples, we first apply poisson disk sampling [48] to uniformly select 3D surface points for object surface and uniformly sample mesh vertices from SMPL-X [61] human, constructing ordered point sets for each object and human." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": "Given the ordered point sets, the density function on the " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": "-th object point and the " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": "-th human point, which we denote as " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{ij}(\\mathbf{p},\\mathbf{n})" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": ", represents the joint probability of the relative position " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{p}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": " and relative normal orientation " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{n}\\in \\mathbb{S}^2" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": "-th human point with respect to " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": "-th object point within the 3D HOI samples. We obtain " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": " by canonicalizing the orientation of the human surface normal " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_j^h" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": " and relative position " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{o\\rightarrow h}" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": ", assuming the object surface normal " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i^o" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": " is rotated to face " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}} = [0,0,1]^T" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": " (see Supp. Mat. for details on canonicalization). The distribution " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{ij}" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": " sums up to 1 for all possible " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 207, + 483, + 304 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 312, + 481, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 312, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 250, + 312, + 481, + 338 + ], + "type": "interline_equation", + "content": "\\int_ {\\mathbb {R} ^ {3}} \\int_ {\\mathbb {S} ^ {2}} \\mathcal {P} _ {i j} (\\mathbf {p}, \\mathbf {n}) d \\mathbf {n} d \\mathbf {p} = 1 \\tag {4}", + "image_path": "11c9cadb3e90c9ca371ad53233b11efc835c1515e2f7b5d1dbfe5cf04462ceb9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": "In practice, we set the domain of " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": " as voxelized grid and domain of " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": " as equispaced grid on " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^2" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": ", obtained via Fibonacci Spirals [16], and compute the discrete probabilities by aggregating Gaussian Kernels calculated for " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": " (we use geodesic metrics for " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": "). Under the probability distribution " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{ij}(\\mathbf{p},\\mathbf{n})" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": ", we define three types of functions " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "f(\\mathbf{p},\\mathbf{n})" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": " capturing different aspects of affordances: (1) Contact-based Affordance, (2) Oriental Affordance, (3) Spatial Affordance. Given each function " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "f(\\mathbf{p},\\mathbf{n})" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": ", the pointwise affordance of " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": "-th object point with respect to " + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 343, + 482, + 440 + ], + "type": "text", + "content": "-th human point is determined by computing its expectation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 269, + 448, + 481, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 448, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 269, + 448, + 481, + 461 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathbf {p}, \\mathbf {n} \\sim \\mathcal {P} _ {i j}} [ f (\\mathbf {p}, \\mathbf {n}) ] \\tag {5}", + "image_path": "1929f452a2f68a9920892d474cbda3d5708009cae3153b0b6fd384ac3f7afe67.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 468, + 481, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 493 + ], + "type": "text", + "content": "Contact-based Affordance. From ComA, we can infer the contact score for human-object point pairs regarding proximity and normal alignment as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 233, + 499, + 481, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 499, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 233, + 499, + 481, + 526 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {c o n t a c t}} (\\mathbf {p}, \\mathbf {n}) = \\left(\\frac {1 - \\mathbf {n} \\cdot \\hat {\\mathbf {n}}}{2}\\right) e ^ {- \\| \\mathbf {p} \\|} \\tag {6}", + "image_path": "345f2501aa3855454075092eec6d135f495a4115b3b329dfe52f2fb3fc196b75.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "e^{-\\| \\mathbf{p}\\|}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " term encourages high score when distance " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{p}\\|" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " is close. The " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\frac{1 - \\mathbf{n}\\cdot\\hat{\\mathbf{n}}}{2}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " term, motivated by the fact that physical interactions are conducted via exerted normal force between contact points, encourages high score when the object normal " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i^o" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " and human normal " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_j^h" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " align in an antiparallel direction. The normal alignment term improves the precision of the contact map, especially when small body parts (e.g., fingertips) interact with an object and allows robust affordance learning even when using noisy HOI datasets, compared to prior work that only uses proximity for contact terms [2, 22, 23, 90, 93] or normal for computing only penetration [19, 86], without considering the alignment of surface normals." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Orientational Affordance. We aim to capture the pattern and tendency of the orientation of body parts with respect to the object based on the concept" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "of Shannon entropy [70]. Intuitively, low entropy means there exists more typical patterns of human body orientations during the interactions. To quantify the orientational tendency for ComA, we use negated normalized Shannon entropy [70] to enforce high value when the human surface normal shows consistent orientational tendency (low variance) with respect to the object surface normal:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 240, + 185, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 185, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 240, + 185, + 481, + 209 + ], + "type": "interline_equation", + "content": "f _ {\\text {o r i e n t a t i o n}} (\\mathbf {n}) = 1 + \\frac {\\log \\mathcal {P} _ {i j} (\\mathbf {n})}{\\log n _ {b}} \\tag {7}", + "image_path": "7416a633e737ba515a4c46d4494667c08028392176443780e6c4dbe844a1be86.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "text", + "content": " denotes the number of discretized bins of probability measure. Intuitively, the orientational affordance defined by " + }, + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathbf{n} \\sim \\mathcal{P}_{ij}}[f_{\\mathrm{orientation}}(\\mathbf{n})]" + }, + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "text", + "content": " has a low value when marginalized distribution " + }, + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{ij}(\\mathbf{n})" + }, + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "text", + "content": " is close to the uniform distribution, meaning that there is no dominant orientational pattern. Note that Eq. 7 is agnostic to the proximity term " + }, + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 213, + 480, + 284 + ], + "type": "text", + "content": ", which allows us to model the nonphysical orientational effects for even far-distance points." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 285, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 479, + 320 + ], + "type": "text", + "content": "Spatial Affordance. Following CHORUS [20], we also consider the 3D spatial occupancy distribution of a human surface point with respect to the object points by simply counting the occurrence:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 260, + 329, + 480, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 329, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 260, + 329, + 480, + 342 + ], + "type": "interline_equation", + "content": "f _ {\\text {s p a t i a l}} (\\mathbf {p}) = \\mathbb {1} (\\mathbf {x} - \\mathbf {p}) \\tag {8}", + "image_path": "e0266e6c708ce219bc21dc59cf1dbfd84b7af073ec1f8cf706d3a649426cfe66.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot):\\mathbb{R}^3\\to \\{0,1\\}" + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "content": " is a binary function that returns 1 if the argument is zero vector, else 0. Thus the spatial affordance " + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathbf{p}\\sim \\mathcal{P}_{ij}}[f_{\\mathrm{spatial}}(\\mathbf{p})]" + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "content": " outputs scalar occupancy for the spatial vector " + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "content": ". In practice, we implement the function as voxel array, counting 1 to the voxel that contains " + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{spatial}}" + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "content": " is agnostic to the normal direction " + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "content": " via marginalization. Learning the occupancy informs us about macro positioning of the human relative to the given object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 438, + 230, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 438, + 230, + 452 + ], + "spans": [ + { + "bbox": [ + 131, + 438, + 230, + 452 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 480, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 558 + ], + "type": "text", + "content": "In this section, we conduct experiments to demonstrate the efficacy of our representation ComA, and the method for learning it. In Sec. 4.3, we illustrate how ComA extends beyond contact-based affordance to learn distributions for orientational tendency and spatial relation, by visualizing it across various object categories. In Sec. 4.4, we perform quantitative comparisons with other baselines and verify the efficacy of our Adaptive Mask Inpainting. Finally, in Sec. 4.5, we introduce an optimization framework achievable through the knowledge provided by ComA, demonstrating the potential of our new affordance representation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 574, + 205, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 205, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 205, + 586 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 593, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 479, + 666 + ], + "type": "text", + "content": "3D Object Datasets. We obtain 3D object meshes from various 3D object datasets for qualitative evaluation. Specifically, We utilize 20 from BEHAV [2], 10 from InterCap [29], 2 from ShapeNet [9], and 8 from SAPIEN [83] to learn ComA individually. As textures were not available in InterCap [29], making the diffusion model difficult to inpaint, we used TEXTure [66] to generate textures on the meshes using text prompt automatically generated from ChatGPT [58]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 448, + 103 + ], + "type": "text", + "content": "Discovering Comprehensive Affordance for 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "For quantitative evaluation, we utilize BEHAVE [2] as a ground-truth dataset to verify the quality of generated 3D HOI samples, assuming that interactions captured in lab-controlled environments exhibit the upper-bound quality of HOI. BEHAVE [2] contains video sequences of HOI scenarios for 20 object categories. Each video frame is annotated with a single 3D human and object, where each human is given as SMPL-H [44,68] format and objects as rotation and translation applied to the canonical mesh. We preprocess the dataset by transferring the SMPL-H annotations into SMPL-X format following Choutas et al. [61]. We use the 3D human-object pair in each frame as a sample, compute vertex-wise contact scores in the same way as our method, and use them as ground truth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 261 + ], + "type": "text", + "content": "Internet Search Datasets. We also collect 60 free 3D object meshes from Internet source, SketchFab [71] and learn ComA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 276, + 266, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 276, + 266, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 276, + 266, + 288 + ], + "type": "text", + "content": "4.2 Baselines and Metric" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 403 + ], + "type": "text", + "content": "As previous studies focused on contact rather than learning the overall aspect of affordances, we use contact-based affordance derived from ComA to compare with them. However, most of them output 3D contact of a single HOI situation by inferring affordance from a given HOI sample (mostly an image), while we produce an overall affordance distribution for the given 3D object. To address this, we treat the contact inferred by state-of-the-art (SOTA) methods, DECO [79] (for the human) and IAGNet [87] (for the object) from the HOI image as a single sample, and aggregate them across multiple images to construct a distribution for comparison against BEHAVE [2]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 487 + ], + "type": "text", + "content": "Specifically, we aggregate the results of DECO [79] and IAGNet [87] on (1) 2D HOI images we've generated beforehand and (2) BEHAVE [2] test images, as these represent the usage of the BEHAVE [2] object. The aggregated scores are then normalized to form a distribution (summing up to 1), considering only the relative probability of potential contact, as the original scale of the scores varies between methods. Subsequently, we compare similarity scores (SIM) [75] and mean absolute error (MAE) [81] within these distributions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 502, + 257, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 257, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 257, + 514 + ], + "type": "text", + "content": "4.3 Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Contact for Various Objects. Similar to other studies, ComA can be derived into contact-based affordance as shown in Fig. 5. As our representation provides contact scores for all pairs of human-object vertices, we assign the contact score of each vertex as the maximum score among the pairs that include it. We can observe the conventional object usage patterns as heatmap, which fits with general human usages. Specifically, in case Fig. 5(b), where significant contact is absent, we can infer that the interaction does not frequently appear by contact." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Orientational Tendency for Various Objects. ComA also contains information on orientational tendencies during the HOI. The last column in Fig. 5 visualizes the negated entropy of the vertex normal direction of human relative to the object as a heatmap. Fig. 5(a), Fig. 5(b), and Fig. 5(c) exhibit relatively high orientational tendency, while the rest show lower. This aligns with the fact that" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 474, + 473 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 474, + 473 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 474, + 473 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 474, + 473 + ], + "type": "image", + "image_path": "6dc32b982e0fe503cdae927533964f58d50400ace08b13ce01fbc16e72faf083.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "type": "text", + "content": "Fig. 5: Qualitative Results. ComA can model distributions of contact, orientation, and spatial relation exhibited during the interaction between humans and novel objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 505, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 554 + ], + "type": "text", + "content": "the human's orientation is fixed by the attachment to the back as in Fig. 5(a), and by the restrictions present on the back and side of the object as in Fig. 5(c). In Fig. 5(b), there exists an orientational tendency without attachments or restrictions, which arises due to long-distance interaction: viewing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Spatial Relation for Various Objects. The relative position of the human during HOI can be also derived from ComA. Fig. 5 visualizes the distribution of the human's full body positions relative to the object. Specifically, as shown in Fig. 5(d), Fig. 5(e), and Fig. 5(f), we find that the spatial relations for the object with multiple utilities and geometries (e.g., in the case of shovel, handle for holding and blade for digging) align with general human usage, demonstrating the plausibility of the results. Especially, in the case of tennis racket shown in Fig. 5(e), we can observe a relatively weak tendency in spatial relation, as the racket can move almost freely relative to humans during the interaction." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 448, + 102 + ], + "type": "text", + "content": "Discovering Comprehensive Affordance for 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 309, + 194 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 309, + 194 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 309, + 194 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 309, + 194 + ], + "type": "image", + "image_path": "f0661fbc5771ff412d48f6b6b69b41f374393cb90807d56cbb6d0b2e242e587f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 196, + 288, + 206 + ], + "lines": [ + { + "bbox": [ + 171, + 196, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 171, + 196, + 288, + 206 + ], + "type": "text", + "content": "(a) Finding Corresponding Contact" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 111, + 462, + 194 + ], + "blocks": [ + { + "bbox": [ + 312, + 111, + 462, + 194 + ], + "lines": [ + { + "bbox": [ + 312, + 111, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 111, + 462, + 194 + ], + "type": "image", + "image_path": "2011eecb84d85428aac33a08a2e8ee6fc77b994aa1d0b79b0cd28930a2b3143c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 196, + 455, + 206 + ], + "lines": [ + { + "bbox": [ + 320, + 196, + 455, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 196, + 455, + 206 + ], + "type": "text", + "content": "(b) Evaluating Plausibility of Orientation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 210, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 131, + 210, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 481, + 232 + ], + "type": "text", + "content": "Fig. 6: Pointwise Affordance Knowledge. ComA can provide pointwise affordance knowledge such as corresponding contacts or plausible orientation of body parts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 141, + 238, + 470, + 300 + ], + "blocks": [ + { + "bbox": [ + 141, + 238, + 470, + 300 + ], + "lines": [ + { + "bbox": [ + 141, + 238, + 470, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 238, + 470, + 300 + ], + "type": "table", + "html": "
MethodsSIMHuman(%)↑SIMObject(%)↑MAEHuman(×100)↓MAEObject(×100)↓
Ours52.8270.550.1260.0288
IAGNet [87] on Our 2D HOI Images-66.34-0.0329
IAGNet [87] on BEHAVECT Images-63.86-0.0353
DECO [79] on Our 2D HOI Images21.66-0.246-
DECO [79] on BEHAVECT Images17.39-0.259-
", + "image_path": "e6159081e78d4bef11ed2f01b8b90d8c6e99e339e114e82ac4939c44a7282d53.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 304, + 480, + 326 + ], + "lines": [ + { + "bbox": [ + 131, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 480, + 326 + ], + "type": "text", + "content": "Table 1: Quantitative Results on BEHAV. We report SIM, MAE to evaluate the similarity of contact distribution (normalized contact scores) on BEHAV sequences." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 331, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 485 + ], + "type": "text", + "content": "Pointwise Affordance Knowledge. Different from previous studies, ComA models the distribution of relative orientation and proximity for all human-object vertex pairs, allowing it to provide pointwise affordance knowledge such as corresponding contacts and plausible orientation direction of specific body parts. As shown in Fig. 6(a), we can infer how each human parts and object parts contact. For the case of motorcycle, we can see that saddle and handle each contacts with human's hips and hands; while for the case of chair, the seat and backrest each contacts with human's hips and back. Also, we can utilize ComA to infer the probable direction of the head while using the monitor by computing the cosine similarity between the head direction and the vertex normal with highest orientational tendency. As illustrated in Fig. 6(b), we can see that the similarity maximizes when gazing towards the screen; which is a plausible head direction when interacting with the monitor." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "type": "text", + "content": "Ablation on Adaptive Mask Inpainting. As shown in Fig. 4, without Adaptive Mask Inpainting, the original geometry of the object changes (left) or gets replaced by hallucinated objects (right), resulting in false affordances unfaithful to the original object. Consequently, the difference between the object in 2D and the original one existing in 3D causes the uplifted 3D HOI samples to fail to represent the right usage. This shows the importance of Adaptive Mask Inpainting as it prevents 2D inpainting diffusion model from generating false affordances." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 586, + 265, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 265, + 598 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 265, + 598 + ], + "type": "text", + "content": "4.4 Quantitative Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Comparison with Baselines. We evaluate the similarity of contact distribution between BEHAV [2] and each of DECO [79], IAGNet [87], and ours. We sample 2048 points from the 3D object mesh surface following IAGNet [87] and only consider the contact underneath the head for DECO [79], as DECO [79] uses SMPL [44] format, which only has vertex correspondence with SMPL-X [61] in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 194, + 114, + 416, + 149 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 416, + 149 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 416, + 149 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 416, + 149 + ], + "type": "table", + "html": "
Inpainting MethodsRMSEBackground↓ mIoU↑mIoUOCclusionAware ↑
Vanilla42.230.5350.631
Adaptive Mask15.330.7060.815
", + "image_path": "3a3fa07afa08dd59ab24dca80b68695d8429e885329774f4931c3663adfe7290.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 153, + 481, + 175 + ], + "lines": [ + { + "bbox": [ + 132, + 153, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 153, + 481, + 175 + ], + "type": "text", + "content": "Table 2: Effects of Adaptive Mask Inpainting We evaluate the performance of background preservation in vanilla inpainting and our Adaptive Mask Inpainting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 182, + 480, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 290 + ], + "type": "text", + "content": "non-head regions. As shown in Tab. 1, the contact distribution from IAGNet [87] and DECO [79] using our 2D HOI images outperforms the one modeled from BEHAVE test images. This means that our generated HOI images represent the real-world affordance as well as BEHAVE [2] test images, demonstrating the efficacy of leveraging pre-trained diffusion model for generating affordance. Additionally, the contact distribution from our method outperforms the one modeled using IAGNet [87] and DECO [79] with our generated HOI images. Since we extract 3D contacts from the same image sets, we argue that our pipeline is better than other models trained on annotations for uplifting 2D affordance to 3D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "text", + "content": "Efficacy of Adaptive Mask Inpainting. To verify the effectiveness of our Adaptive Mask Inpainting, we compare the amount of background preservation for our generated 2D HOI images with the vanilla inpainting diffusion model. We sample a maximum of 2 inpainted images with a single human detected, per each object categories from BEHAV [2]; resulting in 691 images for vanilla inpainting and 692 images for Adaptive Mask Inpainting. We compute 3 metrics for these images; (1) " + }, + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{RMSE}_{\\mathrm{Background}}" + }, + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "text", + "content": ": pixel error between images excluding the predicted human region; (2) mIoU: average IoU between object predictions in each inpainted image and original object image; (3) mIoU" + }, + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "inline_equation", + "content": "_{\\mathrm{Occlusion\\_Aware}}" + }, + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "text", + "content": ": same metric as (2) but excluding the predicted human region during computation. To ensure fairness, we use open vocabulary segmentation model [35,43] to predict detailed human/object segmentation. As shown in Tab. 2, we demonstrate that our adaptive mask algorithm better preserves the background, not only pixelwise " + }, + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "inline_equation", + "content": "(\\mathrm{RMSE}_{\\mathrm{Background}})" + }, + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "text", + "content": ", but also semantic-region-wise (mIOU, mIOU" + }, + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "inline_equation", + "content": "_{\\mathrm{Occlusion\\_Aware}}" + }, + { + "bbox": [ + 130, + 291, + 481, + 483 + ], + "type": "text", + "content": ") than the vanilla inpainting baseline. This means that the generated images from adaptive mask inpainting are more likely to preserve the original object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 501, + 219, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 219, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 219, + 513 + ], + "type": "text", + "content": "4.5 Application" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "type": "text", + "content": "As ComA is a newly proposed affordance representation, we demonstrate its potential by showcasing an application. One possible task is to reconstruct HOI by optimizing the human pose and position to interact with an object. As shown in Fig. 7(a), it is nearly impossible to achieve the task with traditional methods which relies on contact affordance only. In contrast, ComA provides both relative orientation and contact during the HOI, making it possible to generate plausible 3D HOI samples from T-posed humans. For the objects with multiple modes of affordances, we apply mode seeking algorithm [11] on 3D HOI samples with their vertex normals and learn ComA separately; for example, swing in Fig. 7(a)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "However, ComA learned on a specific object cannot be directly applied to other objects with different geometries. To address this, we transfer ComA through the process of (1) object canonicalization and (2) finding vertex cor-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "Discovering Comprehensive Affordance for 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 112, + 452, + 345 + ], + "blocks": [ + { + "bbox": [ + 157, + 112, + 452, + 345 + ], + "lines": [ + { + "bbox": [ + 157, + 112, + 452, + 345 + ], + "spans": [ + { + "bbox": [ + 157, + 112, + 452, + 345 + ], + "type": "image", + "image_path": "52af60b288a31d15ff981a926274ef539f88e6bf5316891784d2e77fed48469d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 348, + 482, + 371 + ], + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 371 + ], + "type": "text", + "content": "Fig. 7: Application. ComA can reconstruct HOI through optimization, which can be scaled to other objects in the same category by transferring such knowledge between them." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 375, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 471 + ], + "type": "text", + "content": "respondences. Fig. 7(b) provides an example of transferring ComA learned from a specific motorcycle to other motorcycles with different geometries. We use objects from ShapeNet [9] which are already canonicalized and apply CPNet [45] to find vertex correspondences automatically for transferring the vertices with contact. As object canonicalization and vertex correspondence can each transfer macro orientation information and local contact information, we can transfer ComA and scale the human optimization task to other objects in the same category without any additional learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we point out the importance of learning non-contact patterns in HOIs and propose a novel affordance representation called ComA which models both contact and non-contact patterns. To learn ComA, we utilize a pre-trained 2D diffusion model which already possesses prior knowledge of affordance. To leverage the affordance knowledge, we designed a Rendering-Inpainting-Uplifting pipeline to generate 3D HOI samples and learn ComA from the samples. The results show that ComA effectively models the common usage of objects in terms of contact, orientational, and spatial relations. Within the pipeline, we propose Adaptive Mask Inpainting, which allows insertion of human without damaging the original image, where we evaluate its efficacy both qualitatively and quantitatively. We expect ComA to be used as a prior knowledge for various downstream tasks, where we demonstrate its potential through human optimization framework and the possibility of transfer to various objects." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "This work was supported by Naver Webtoon. The work of SNU members was also supported by NRF grant funded by the Korean government (MSIT) (No. 2022R1A2C2092724 and No. RS-2023-00218601), and IITP grant funded by the Korean government (MSIT) [RS-2021-II211343, AI Graduate School Program (SNU)]. H. Joo is the corresponding author." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 220, + 198, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 198, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 198, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 246, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "type": "text", + "content": "1. Bahl, S., Mendonca, R., Chen, L., Jain, U., Pathak, D.: Affordances from human videos as a versatile representation for robotics. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "text", + "content": "2. Bhatnagar, B.L., Xie, X., Petrov, I., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Behave: Dataset and method for tracking human object interactions. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "type": "text", + "content": "3. Brahmbhatt, S., Ham, C., Kemp, C.C., Hays, J.: Contactdb: Analyzing and predicting grasp contact via thermal imaging. In: CVPR (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 324, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 480, + 344 + ], + "type": "text", + "content": "4. Brahmbhatt, S., Handa, A., Hays, J., Fox, D.: Contactgrasp: Functional multifinger grasp synthesis from contact. In: IROS (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 346, + 480, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 367 + ], + "type": "text", + "content": "5. Brahmbhatt, S., Tang, C., Twigg, C.D., Kemp, C.C., Hays, J.: Contactpose: A dataset of grasps with object contact and hand pose. In: ECCV (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 368, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 480, + 389 + ], + "type": "text", + "content": "6. Cai, M., Kitani, K.M., Sato, Y.: Understanding hand-object manipulation with grasp types and object attributes. In: RSS (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 391, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 391, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 391, + 480, + 411 + ], + "type": "text", + "content": "7. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "type": "text", + "content": "8. Chai, L., Zhu, J.Y., Shechtman, E., Isola, P., Zhang, R.: Ensembling with deep generative views. In: CVPR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 434, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 467 + ], + "type": "text", + "content": "9. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 480, + 500 + ], + "type": "text", + "content": "10. Chen, Y., Huang, S., Yuan, T., Qi, S., Zhu, Y., Zhu, S.C.: Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In: ICCV (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 501, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 501, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 134, + 501, + 480, + 521 + ], + "type": "text", + "content": "11. Comaniciu, D., Meer, P.: Mean shift: a robust approach toward feature space analysis. In: IEEE TPAMI (2002)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 134, + 522, + 480, + 544 + ], + "type": "text", + "content": "12. Deng, S., Xu, X., Wu, C., Chen, K., Jia, K.: 3d affordancenet: A benchmark for visual object affordance understanding. In: CVPR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 544, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 480, + 566 + ], + "type": "text", + "content": "13. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: CVPR (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "text", + "content": "14. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. In: CACM (1981)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "text", + "content": "15. Gao, W., Tedrake, R.: kpm-sc: Generalizable manipulation planning using keypoint affordance and shape completion. In: ICRA (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "text", + "content": "16. Garg, M., Garg, P., Vohra, R.: Advanced fibonacci sequence with golden ratio. In: IJSER (2014)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 666 + ], + "type": "text", + "content": "17. Gibson, J.J.: The Ecological Approach to Visual Perception. Houghton Mifflin (1979)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "Discovering Comprehensive Affordance for 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "19. Grady, P., Tang, C., Twigg, C.D., Vo, M., Brahmbhatt, S., Kemp, C.C.: ContactOpt: Optimizing contact to improve grasps. In: CVPR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "20. Han, S., Joo, H.: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In: ICCV (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "21. Hao, Z., Mallya, A., Belongie, S., Liu, M.Y.: Gancraft: Unsupervised 3d neural rendering of apache worlds. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "22. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "type": "text", + "content": "23. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: CVPR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 277 + ], + "type": "text", + "content": "24. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? In: arXiv:2210.07574 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "type": "text", + "content": "25. Hermans, T., Rehg, J.M., Bobick, A.: Affordance prediction via learned object attributes. In: ICRA (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "text", + "content": "26. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: arXiv:2208.01626 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "text", + "content": "27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "text", + "content": "28. Huang, C.H.P., Yi, H., Hoschle, M., Safroshkin, M., Alexiadis, T., Polikovsky, S., Scharstein, D., Black, M.J.: Capturing and inferring dense full-body human-scene contact. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "29. Huang, Y., Taheri, O., Black, M.J., Tzionas, D.: Intercap: Joint markerless 3d tracking of humans and objects in interaction from multi-view rgb-d images. In: IJCV (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "30. Jahanian, A., Puig, X., Tian, Y., Isola, P.: Generative models as a data source for multiview representation learning. In: arXiv:2106.05258 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "text", + "content": "31. Jian, J., Liu, X., Li, M., Hu, R., Liu, J.: Affordpose: A large-scale dataset of handobject interactions with affordance-driven hand pose. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "32. Jiang, W., Kolotouros, N., Pavlakos, G., Zhou, X., Daniilidis, K.: Coherent reconstruction of multiple humans from a single image. In: CVPR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "text", + "content": "33. Jiang, Y., Jiang, S., Sun, G., Su, Z., Guo, K., Wu, M., Yu, J., Xu, L.: Neuralhofusion: Neural volumetric rendering under human-object interactions. In: CVPR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 515, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 535 + ], + "type": "text", + "content": "34. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "35. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "36. Kirillov, A., Wu, Y., He, K., Girshick, R.: Pointrend: Image segmentation as rendering. In: CVPR (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "text", + "content": "37. Kulal, S., Brooks, T., Aiken, A., Wu, J., Yang, J., Lu, J., Efros, A.A., Singh, K.K.: Putting people in their place: Affordance-aware human insertion into scenes. In: CVPR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "text", + "content": "38. Lee, Y.J., Grauman, K.: Predicting important objects for egocentric video summarization. In: IJCV (2015)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "39. Levine, S., Shah, D.: Learning robotic navigation from experience: principles, methods and recent results. In: Philos. Trans. R. Soc. B (2022)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "40. Li, G., Jampani, V., Sun, D., Sevilla-Lara, L.: Locate: Localize and transfer object parts for weakly supervised affordance grounding. In: CVPR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "41. Li, L., Dai, A.: GenZI: Zero-shot 3D human-scene interaction generation. In: CVPR (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "text", + "content": "42. Li, Z., Sedlar, J., Carpentier, J., Laptev, I., Mansard, N., Sivic, J.: Estimating 3d motion and forces of person-object interactions from monocular video. In: CVPR (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "43. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "type": "text", + "content": "44. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. In: Proc. ACM SIGGRAPH Asia (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "45. Lou, Y., You, Y., Li, C., Cheng, Z., Li, L., Ma, L., Wang, W., Lu, C.: Human correspondence consensus for 3d object semantic understanding. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "46. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "47. Mao, C., Cha, A., Gupta, A., Wang, H., Yang, J., Vondrick, C.: Generative interventions for causal learning. In: CVPR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "48. McCool, M., Fiume, E.: Hierarchical poisson disk sampling distributions. In: Graphics Interface (1992)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "49. Melas-Kyriazi, L., Rupprecht, C., Laina, I., Vedaldi, A.: Finding an unsupervised image segmenter in each of your deep generative models. In: arXiv:2105.08127 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "50. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "51. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: COAP: Compositional articulated occupancy of people. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "52. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "53. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "54. Monszpart, A., Guerrero, P., Ceylan, D., Yumer, E., Mitra, N.J.: imapper: interaction-guided scene mapping from monocular videos. In: ACM TOG (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "55. Moon, G., Choi, H., Lee, K.M.: Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In: CVPRW (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "56. Nagarajan, T., Grauman, K.: Learning affordance landscapes for interaction exploration in 3d environments. In: NeurIPS (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "57. Oleynikova, H., Millane, A., Taylor, Z., Galceran, E., Nieto, J., Siegwart, R.: Signed distance fields: A natural representation for both mapping and planning. In: RSS Workshop (2016)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "58. OpenAI: Chatgpt: Optimizing language models for dialogue. https://openai.com/blog/chatgpt/ (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "59. OpenAI: Gpt-4v(ision) system card. https://openai.com/research/gpt-4v-system-card (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "60. Pan, X., Dai, B., Liu, Z., Loy, C.C., Luo, P.: Do 2d gans know 3d shape? unsupervised 3d shape reconstruction from 2d image gans. In: arXiv:2011.00844 (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "61. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "Discovering Comprehensive Affordance for 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "62. Peebles, W., Zhu, J.Y., Zhang, R., Torralba, A., Efros, A.A., Shechtman, E.: Gansupervised dense visual alignment. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "text", + "content": "63. Petrov, I.A., Marin, R., Chibane, J., Pons-Moll, G.: Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "text", + "content": "64. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "text", + "content": "65. Rempe, D., Guibas, L.J., Hertzmann, A., Russell, B., Villegas, R., Yang, J.: Contact and human dynamics from monocular video. In: ECCV (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "66. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. In: Proc. ACM SIGGRAPH (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "67. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "68. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. In: Proc. ACM SIGGRAPH Asia (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "text", + "content": "69. Savva, M., Chang, A.X., Hanrahan, P., Fisher, M., Nießner, M.: Pigraphs: learning interaction snapshots from observations. In: ACM TOG (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "type": "text", + "content": "70. Shannon, C.E.: A mathematical theory of communication. In: ACM SIGMOBILE (2001)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 310, + 373, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 373, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 373, + 320 + ], + "type": "text", + "content": "71. Sketchfab: Sketchfab. https://sketchfab.com/ (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 320, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 480, + 342 + ], + "type": "text", + "content": "72. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "text", + "content": "73. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 363, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 396 + ], + "type": "text", + "content": "74. Sun, G., Chen, X., Chen, Y., Pang, A., Lin, P., Jiang, Y., Xu, L., Yu, J., Wang, J.: Neural free-viewpoint performance rendering under complex human-object interactions. In: ACM MM (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 396, + 393, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 393, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 393, + 407 + ], + "type": "text", + "content": "75. Swain, M.J., Ballard, D.H.: Color indexing. In: IJCV (1991)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "text", + "content": "76. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: ECCV (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "type": "text", + "content": "77. Tanaka, F.H.K.d.S., Aranha, C.: Data augmentation using gans. In: arXiv:1904.09135 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "text", + "content": "78. Trabucco, B., Doherty, K., Gurinas, M., Salakhutdinov, R.: Effective data augmentation with diffusion models. In: arXiv:2302.07944 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 471, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 492 + ], + "type": "text", + "content": "79. Tripathi, S., Chatterjee, A., Passy, J.C., Yi, H., Tzionas, D., Black, M.J.: DECO: Dense estimation of 3D human-scene contact in the wild. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 492, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 480, + 514 + ], + "type": "text", + "content": "80. Tritrong, N., Rewatbowornwong, P., Suwajanakorn, S.: Repurposing gans for one-shot semantic part segmentation. In: CVPR (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "81. Willmott, C.J., Matsuura, K.: Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. In: Climate Research (2005)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 546, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 480, + 578 + ], + "type": "text", + "content": "82. Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetdm: Synthesizing data with perception annotations using diffusion models. In: NeurIPS (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 578, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 611 + ], + "type": "text", + "content": "83. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: CVPR (2020)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "84. Xu, J., Zhu, S., Guo, H., Wu, S.: Automated labeling for robotic autonomous navigation through multi-sensory semi-supervised learning on big data. In: IEEE TBD (2021)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "85. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In: ICCV (2023)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 391 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "86. Yang, L., Zhan, X., Li, K., Xu, W., Li, J., Lu, C.: CPF: Learning a contact potential field to model the hand-object interaction. In: ICCV (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "text", + "content": "87. Yang, Y., Zhai, W., Luo, H., Cao, Y., Luo, J., Zha, Z.J.: Grounding 3d object affordance from 2d interactions in images. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "88. Yang, Y., Zhai, W., Luo, H., Cao, Y., Zha, Z.J.: Lemon: Learning 3d human-object interaction relation from 2d images. In: CVPR (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "89. Ye, Y., Li, X., Gupta, A., Mello, S.D., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 205, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 236 + ], + "type": "text", + "content": "90. Yi, H., Huang, C.H.P., Tzionas, D., Kocabas, M., Hassan, M., Tang, S., Thies, J., Black, M.J.: Human-aware object placement for visual environment reconstruction. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 237, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 258 + ], + "type": "text", + "content": "91. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 3DV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "type": "text", + "content": "92. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 480, + 303 + ], + "type": "text", + "content": "93. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: CVPR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 304, + 480, + 335 + ], + "type": "text", + "content": "94. Zhang, Y., Chen, W., Ling, H., Gao, J., Zhang, Y., Torralba, A., Fidler, S.: Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In: ICLR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "type": "text", + "content": "95. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "96. Zhou, K., Bhatnagar, B.L., Lenssen, J.E., Pons-Moll, G.: Toch: Spatio-temporal object correspondence to hand for motion refinement. In: ECCV (2022)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "Discovering Comprehensive Affordance for 3D Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_content_list.json b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..00b1f1fc8b701f43d0b4612fa0d32f3b67670773 --- /dev/null +++ b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "Beyond the Data Imbalance: Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction", + "text_level": 1, + "bbox": [ + 238, + 140, + 767, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyeongseok Jeon $^{1}$ , Sanmin Kim $^{2}$ , Abi Rahman Syamil $^{2}$ , Junsoo Kim $^{3}$ , and Dongsuk Kum $^{2}$", + "bbox": [ + 228, + 233, + 774, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MORAI Inc., Seoul, South Korea \n $^{2}$ Korea Advanced Institute of Science and Technology, Daejeon, South Korea \n $^{3}$ Hyundai Motor Company, Seoul, South Korea \nhsjeon@morai.ai dskum@kaist.ac.kr", + "bbox": [ + 243, + 276, + 759, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Predicting the maneuvers of surrounding vehicles is imperative for the safe navigation of autonomous vehicles. However, naturalistic driving datasets tend to be highly imbalanced, with a bias towards the \"going straight\" maneuver. Consequently, learning and accurately predicting turning maneuvers pose significant challenges. In this study, we propose a novel two-stage maneuver learning method that can overcome such strong biases by leveraging two heterogeneous datasets in a complementary manner. In the first training phase, we utilize an intersection-centric dataset characterized by balanced distribution of maneuver classes to learn the representations of each maneuver. Subsequently, in the second training phase, we incorporate an ego-centric driving dataset to account for various geometrical road shapes, by transferring the knowledge of geometric diversity to the maneuver prediction model. To facilitate this, we constructed an in-house intersection-centric trajectory dataset with a well-balanced maneuver distribution. By harnessing the power of heterogeneous datasets, our framework significantly improves maneuver prediction performance, particularly for minority maneuver classes such as turning maneuvers. The dataset is available at https://github.com/KAIST-VDCLab/VDC-Trajectory-Dataset.", + "bbox": [ + 261, + 367, + 743, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Maneuver Prediction $\\cdot$ Long-tail Problem $\\cdot$ Autonomous Driving", + "bbox": [ + 259, + 643, + 743, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 718, + 377, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Predicting the motion of surrounding vehicles is a critical component for ensuring the safety of autonomous driving. In recent years, numerous studies have introduced motion prediction models [19, 21, 27, 33, 38] that primarily focus on generating detailed trajectories. These approaches are trained and evaluated using trajectory-level metrics such as ADE/FDE (Average/Final Distance Error), assessing how closely predicted trajectories align with actual driving trajectories.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fcb5d650a1df8109d0f06a43fe28dbedf22964c2f9ef0eea659f72a9407dfe49.jpg", + "image_caption": [ + "Fig. 1: Importance of maneuver prediction. In a intersection scenario, accurate prediction of vehicles' turning maneuver is crucial." + ], + "image_footnote": [], + "bbox": [ + 248, + 146, + 754, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, from the perspective of motion planning, it becomes clear that accurately predicting the high-level intentions of surrounding vehicles holds greater importance than providing fine-grained future predictions. For example, in intersection scenarios where vehicle interactions are maximized, precisely determining a vehicle's turning maneuver is more crucial than outputting detailed future trajectories, which may end up being a wrong maneuver, as shown in Fig. 1. Hence, prioritizing the accurate prediction of vehicle intentions over detailed trajectory predictions becomes paramount for safe motion planning in autonomous driving, particularly in intricate scenarios like intersections.", + "bbox": [ + 212, + 398, + 784, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various datasets have been collected and released for the motion prediction task, gathered by driving sensor-equipped vehicles in real-world environments to ensure scenario diversity. While these datasets provide ample trajectories for training motion prediction models, they exhibit highly biased distributions due to the inherent nature of driving trajectories, particularly concerning maneuvers. Straight roads dominate the driving environment, leading to most trajectory datasets being biased towards straight trajectories. Consequently, maneuver classification models trained with biased datasets tend to be biased toward scenarios involving straight driving while learning less about other maneuvers.", + "bbox": [ + 212, + 536, + 785, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several approaches have been proposed in order to address the long-tail problem. Sampling strategies [5, 10, 28, 31] have attempted to handle this issue by either oversampling rare cases or downsampling dominant cases. Alternatively, loss function reformulation strategies [12, 20, 23, 35] have been explored to adjust gradients and enforce balanced supervision during training. However, these approaches have shown limited success, as they merely manipulate the training dataset to achieve balance or regulate the learning process within the existing dataset. Consequently, the fundamental problem stemming from the long-tailed nature of the data has not been effectively addressed.", + "bbox": [ + 212, + 672, + 785, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we present a novel two-stage learning framework that harnesses the power of heterogeneous datasets. Our approach aims to", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Jeon et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b889c40c26a60307a8166ca05a4149ed604c0ac927c586e3a804e8b22867c523.jpg", + "image_caption": [ + "Fig. 2: Pros and Cons of the two different datasets: An intersection-centric dataset acquired from the bird's eye view and an ego-centric dataset that collected in a vehicle view." + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 500, + 272 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/98ff3f616863790c54b32c55ad4bc8682cc912b45b9608457e95bdfcec17b7d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 148, + 781, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "complement the naturally imbalanced driving dataset by incorporating a cost-efficient and easily acquirable intersection-centric dataset. In our proposed framework, we recognize the importance of capturing various turning maneuvers commonly observed in intersections. Towards that end, we introduce a drone-based intersection-centric dataset, specifically chosen to mitigate the limitations of the ego-centric driving dataset. We utilized the in-house intersection-centric dataset by focusing on a representative intersection area with the aim of learning maneuver representations. To effectively combine these heterogeneous datasets for training maneuver prediction models, we first extract valuable knowledge of each maneuver class from the in-house intersection-centric trajectory dataset. This knowledge, which is challenging to learn from ego-centric driving datasets alone, is then transferred to the downstream network. Here, the model learns the geometric diversity leveraging the ego-centric driving dataset. This approach ensures that our model will initially acquire rich intersection-oriented maneuver knowledge in a balanced manner, while the downstream network then adapts the maneuver representations to diverse road geometries. As a result, our model simultaneously learns maneuver representations from the intersection-centric dataset and features of diverse road geometries from the ego-centric driving dataset.", + "bbox": [ + 212, + 342, + 787, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, this study is the first attempt to train the maneuver (or trajectory) prediction model with the heterogeneous datasets in complementary manner. The main contributions of our work are outlined as follows:", + "bbox": [ + 212, + 614, + 787, + 674 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel two-stage maneuver prediction framework that effectively leverages heterogeneous datasets to address the long-tail problem in motion prediction.", + "- We introduce a novel trajectory dataset collected specifically at intersections, ensuring a balanced distribution of maneuvers." + ], + "bbox": [ + 223, + 679, + 784, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 777, + 387, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the data imbalance problem in deep learning approaches appears widely, general solutions for handling the long tail problem is introduced. However, those", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Beyond the Data Imbalance in Vehicle Maneuver Prediction", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "approaches are only focused on formulating the training dataset or loss functions based on the given imbalanced dataset regardless of the data diversity. Therefore, the diversity of the scenarios for training the network still remain insufficient.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Sampling-based Approaches", + "text_level": 1, + "bbox": [ + 215, + 215, + 493, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In sampling-based approaches, the training dataset is selectively or redundantly reformulated by choosing the data from the original dataset. During this reformulation process, over- and/or under-sampling methods are employed to prevent overfitting resulting from imbalanced datasets.", + "bbox": [ + 212, + 243, + 782, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In studies such as [10, 11, 14], balanced training datasets are constructed through an undersampling mechanism, which involves discarding excessive data instances from the major class of scenarios in the given dataset. In these approaches, data instances from the dominant class are underused to balance the distribution of the training dataset. In contrast, [8,28,31] repeatedly sample minorities in datasets to formulate evenly distributed training datasets using various oversampling algorithms. [31] constructs an equally-distributed mini-batch with a class-balanced composition named class-aware sampling. In [28], a soft-balance sampling method is employed for probabilistic oversampling, while [8] adopts a two-stage training strategy that utilizes both the original dataset and the balanced dataset with different learning rates. Furthermore, to account for randomness in the oversampling process, [5] proposes a synthetic data oversampling algorithm named SMOTE (Synthetic Minority Over-sampling TEchnique), with many variants suggested in subsequent studies [5,7,9,34]. However, these sampling-based data balancing approaches have shown unsatisfactory performance in predicting the future maneuver of vehicles. This is because sampling-based methods merely reconstruct the training dataset within the imbalanced raw dataset. However, in autonomous driving, maneuver prediction networks should observe as many diverse samples as possible to create a generalized prediction model.", + "bbox": [ + 212, + 304, + 787, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Loss Function Reformulation", + "text_level": 1, + "bbox": [ + 215, + 631, + 501, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another approach to address imbalanced datasets is through loss function reformulation, as demonstrated in studies such as [12, 20, 23, 35]. These approaches introduce auxiliary losses to address minority cases. In [12] and [35], a simple weighting method is employed, where larger weights are assigned to gradients for minor classes to equalize the backpropagating effect across all classes. In other studies like [20] and [23], various loss functions are introduced to reshape the minority distribution. In [20], the learning of easy and difficult examples is reprioritized by modeling behavior with the Generalized Pareto Distribution (GPD) using the Pareto Loss. Conversely, in [23], the representation of internal maneuver context is reshaped using contrastive loss, ensuring that difficult examples are well-distributed throughout the representation space. While these loss reformulation approaches attempt to address the data imbalance problem during the", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Jeon et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/335c603e8d752667c71c3e4d679f9db015be3a9611c243e79591634c962456b0.jpg", + "image_caption": [ + "Fig.3: Overall architecture of the proposed method. Maneuver representations are extracted from the intersection-centric dataset and adapted to the diverse geometries with the ego-centric driving dataset. Using the maneuver context distribution form intersection dataset, each maneuver score is calculated based on the Gaussian KDE." + ], + "image_footnote": [], + "bbox": [ + 222, + 151, + 782, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "training stage, they still share the same limitation as sampling approaches in that they have a marginal impact on dataset diversity.", + "bbox": [ + 212, + 420, + 784, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Long-tail Problem in Motion Prediction", + "text_level": 1, + "bbox": [ + 214, + 469, + 591, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building an unbiased maneuver prediction model is challenging due to the vehicle maneuvers' inherent long-tail characteristics, even if they are trained with the numerous driving datasets already released. In [25], the influence of dataset composition on performance is comprehensively analyzed, revealing that a balanced dataset containing evenly distributed maneuvers, such as lane changes and straight movements, could enhance overall performance. As a result, several maneuver and trajectory prediction studies [2, 36] have employed oversampling approaches to avoid biased training. These frameworks have demonstrated effectiveness in the field of vehicle maneuver prediction [15, 30]. However, sampling approaches alone are insufficient to overcome the fundamental limitations of long-tailed datasets, and the maneuver prediction task is no exception. Therefore, a supplementary dataset containing numerous samples of minor maneuvers, such as turning maneuvers, is necessary to augment the observable trajectory samples in the training dataset.", + "bbox": [ + 212, + 489, + 787, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Our Approach", + "text_level": 1, + "bbox": [ + 214, + 720, + 393, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To establish a comprehensive approach that encompasses both balanced maneuver distribution and geometric diversity, we introduce a two-stage maneuver prediction framework leveraging heterogeneous datasets. Firstly, we construct an intersection-centric trajectory dataset characterized by a well-balanced distribution of maneuvers. This dataset serves as the foundation for obtaining maneuver representations, achieved through training a maneuver backbone using", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Beyond the Data Imbalance in Vehicle Maneuver Prediction", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8f6c867d6ba9b8c0f27c2a888cad53cbccc2926d52897e606ad1bbb27ddd3d59.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation on the in-house drone dataset with various conventional datasets from the fixed view point." + ], + "table_footnote": [], + "table_body": "
DatasetLocationTrajectory CountsLocation CountsObjectsFPSMethod
NGSIM [16]boulevard33664car10HzRSU
Ko-PER [32]intersection3501pedestrian, bicycle, car25HzRSU
VRU [13]intersection32781pedestrian, bicycle25HzRSU
DUT [37]campus18622pedestrian, vehicles24Hzdrone
SDD [29]campus102408pedestrian, bicycle, car, skateboard, cart, bus25Hzdrone
InD [3]intersection135994pedestrian, bicycle, car, truck, bus25Hzdrone
Oursintersection, lane merge1547910pedestrian, bicycle, car30Hzdrone
", + "bbox": [ + 228, + 184, + 782, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae4b9f693c83eebb4e642c08ed51f9052d8436e1e63352a484174daf22298135.jpg", + "image_caption": [ + "Fig. 4: Sample scenes in our in-house intersection-centric dataset. Car, bicycle and pedestrian are represented as yellow, cyan and green." + ], + "image_footnote": [], + "bbox": [ + 233, + 393, + 496, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf614ba3f78b58b66e08fd0f7d8833c84f78b65f733abec8817e5b94e032ad65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 395, + 767, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a self-supervised learning approach. Additionally, we derive probability density functions representing each maneuver by applying Kernel Density Function to the training set of the intersection dataset. Then, we bolster the model's scalability by injecting geometric diversity via the utilization of an ego-centric driving dataset. This step enriches the model's comprehension of diverse road geometries, thereby enhancing its predictive accuracy. During the inference stage, maneuver scores are computed by estimating the likelihood of the given trajectory based on the obtained maneuver probability density functions. The overall diagram for our proposed maneuver prediction framework is depicted in Fig. 3.", + "bbox": [ + 212, + 568, + 787, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Learning Maneuver Representations", + "text_level": 1, + "bbox": [ + 214, + 726, + 563, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An Intersection-centric In-house dataset. To ensure a balanced distribution of vehicle maneuvers in our dataset, we construct an in-house intersection-centric dataset. This dataset leverages stationary drone videos, which are easily acquired and require minimal labeling costs compared to sensor-based driving datasets. Samples from the in-house dataset, showcasing abundant turning maneuvers with vehicles, bicycles, and pedestrians depicted in yellow, green, and", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Jeon et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a9f96296c7a8fdf6455975be73926ac6fdfd570a370029eae8dead639220e391.jpg", + "image_caption": [ + "Fig. 5: Network architecture of the maneuver backbone (left) and auxiliary module for training the backbone network (right)." + ], + "image_footnote": [], + "bbox": [ + 223, + 148, + 782, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cyan colors respectively, are illustrated in Fig. 4. The dataset comprises 3.5 hours of video footage captured across 10 locations in 4K resolution at 30 frames per second. Additionally, the dataset includes not only agent trajectories but also high-definition (HD) map data. Furthermore, a digital twin environment is constructed to simulate scenarios within the dataset. A comprehensive evaluation of the in-house drone dataset, compared with other bird's eye view (BEV) datasets, is summarized in Tab. 1. Utilizing this dataset, we train the vehicle's maneuver knowledge in a balanced manner.", + "bbox": [ + 212, + 338, + 787, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Maneuver Backbone. In the first training phase, we extract maneuver features by utilizing the intersection-centric dataset, which is characterized by uniformly distributed maneuvers. To capture maneuver representations from the observed historical trajectory of each vehicle, we employ a self-supervised learning approach, as illustrated in Fig. 5.", + "bbox": [ + 212, + 479, + 787, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, the observed historical trajectory is encoded into a state embedding using an encoder:", + "bbox": [ + 212, + 555, + 787, + 585 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} ^ {t} = f _ {e} \\left(s _ {i} ^ {t} = \\left(x _ {i} ^ {t}, y _ {i} ^ {t}, \\theta_ {i} ^ {t}\\right) \\mid_ {t = - t h \\dots , 0}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 599, + 785, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $f_{e}$ represents the encoder, $h_{i}^{t}$ denotes a state embedding, $s_{i}^{t} = (x_{i}^{t},y_{i}^{t},\\theta_{i}^{t})$ denotes the state of the $i^{th}$ vehicle at timestep $t$ , and $t_h$ represents a observation time horizon. We employ 1D CNN and MLP for the encoder.", + "bbox": [ + 212, + 621, + 785, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The state embeddings are then fed into an autoregressive model (e.g. GRU [6]) to extract temporal trajectory features. Subsequently, an interaction model (e.g. ScaleNet [17]) is employed to account for interaction among vehicles in multi-vehicle scenarios:", + "bbox": [ + 214, + 667, + 787, + 727 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nc = f _ {v} \\left(f _ {a} \\left(h _ {i} ^ {t} | _ {t = - t _ {h}, \\dots , 0}\\right) | _ {i = 1, \\dots , n}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 741, + 785, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $c = \\{c_1, \\dots, c_n\\}$ , $f_v$ and $f_a$ represent the extracted maneuver representations, an interaction module, and an autoregressive model, respectively. $n$ denotes the number of vehicles in the scene.", + "bbox": [ + 212, + 763, + 787, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To learn the internal hidden representation of each maneuver, we employ a self-supervised learning framework. Based on the Contrastive Predictive Coding", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Beyond the Data Imbalance in Vehicle Maneuver Prediction", + "bbox": [ + 330, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "framework [26], the maneuver backbone is trained to maximize mutual information (MI) between the interaction-aware maneuver context of the historical view, $c_{i}$ , and the state embedding in the future $h_i^t|_{t\\geq 0}$ . By maximizing the MI between historical trajectory and future position of the vehicle, the network can learn the slow-varying high level maneuver context of the vehicles.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the definition, MI between $c_{i}$ and $h_i^t|_{t\\geq 0}$ , $I(h_i^t,c_i)$ is formulated as follow:", + "bbox": [ + 212, + 222, + 782, + 251 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(h _ {i} ^ {t}, c _ {i}\\right) = \\sum_ {i = 1} p \\left(h _ {i} ^ {t}, c _ {i}\\right) \\log \\frac {p \\left(h _ {i} ^ {t} , c _ {i}\\right)}{p \\left(h _ {i} ^ {t}\\right)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 263, + 785, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As the MI is hardly tractable within the batch training process, the auxiliary function named as MI score function is defined as Eq. (4) with trainable auxiliary linear transformation coefficients $w^{t}$ :", + "bbox": [ + 212, + 306, + 782, + 351 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nf _ {k} \\left(h _ {i} ^ {t}, c _ {i}\\right) \\equiv e x p \\left(\\left(h _ {i} ^ {t}\\right) ^ {T} w ^ {t} c _ {i}\\right) \\propto \\frac {p \\left(h _ {i} ^ {t} , c _ {i}\\right)}{p \\left(h _ {i} ^ {t _ {d}}\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 364, + 785, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since MI score function is proportional to $I(h_i^t, c_i)$ as formulated in Eq. (4), MI score function is maximized for the same vehicle and minimized for different vehicles using the infoNCE loss [26] function, formulated as:", + "bbox": [ + 212, + 405, + 782, + 450 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left[ \\frac {1}{t _ {f}} \\sum_ {t = 1} ^ {t _ {f}} \\log \\frac {f _ {k} \\left(h _ {i} ^ {t} , c _ {i}\\right)}{\\Sigma_ {j = 1} ^ {n} f _ {k} \\left(h _ {j} ^ {t} , c _ {i}\\right)} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 464, + 785, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $n$ denotes the number of data instances in a single batch and $t_f$ is the future time horizon. With this loss function, the MI between the history and future states of the identical vehicle is enlarged, whereas the MI with other vehicles is suppressed.", + "bbox": [ + 212, + 510, + 782, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Generalizing for Geometric Diversity", + "text_level": 1, + "bbox": [ + 214, + 594, + 568, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "An Ego-centric Driving Dataset. The intersection-centric dataset used for maneuver representation learning has a limitation in that it lacks geometric diversity. To address this limitation and ensure the model's generalizability across various geometric shapes of driving environments, we incorporate an ego-centric driving dataset [4] alongside the intersection-centric dataset. The ego-centric driving dataset is constructed using sensor-equipped vehicles, providing a diverse range of driving environments with various road shapes. By integrating this dataset into our training pipeline, we can effectively train the map adaptation module and enhance the model's ability to generalize to different geometric scenarios.", + "bbox": [ + 212, + 619, + 784, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Map Adaptation Module. In the proposed maneuver prediction model, we introduce the map adaptation module as a downstream network aimed at generalizing the pre-trained maneuver representations to various road shapes. For map", + "bbox": [ + 212, + 794, + 782, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Jeon et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbf818a36a5dbd51b352dcfc6b08abdb1b92e36fe9917d6be129f5ca5cda31fb.jpg", + "image_caption": [ + "Fig. 6: Overall architecture of the map adaptation modules." + ], + "image_footnote": [], + "bbox": [ + 236, + 152, + 767, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "adaptation, we adopt an off-the-shelf map-based prediction model, LaneGCN [21]. LaneGCN utilizes lane graphs and Graph Convolutional Networks (GCN) to encode road geometries. It consists of multiple modules for encoding and fusing geometric and trajectory features, including MapNet and Lane to Lane for map encoding based on GCN, and Actor to Lane and Lane to Actor to fuse map and trajectory features using an attention mechanism. The overall network is trained using cross-entropy loss as defined in Eq. (6), where $y_{i}$ and $p_i$ represent the ground truth and prediction of the vehicle maneuver, respectively.", + "bbox": [ + 212, + 430, + 787, + 551 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = - \\sum_ {i = 1} ^ {k} y _ {i} \\log \\left(p _ {i}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 563, + 785, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "It is important to note that the maneuver head is employed only during the training process. This is due to the potential bias that may arise in the model outputting predominantly straight maneuvers during the map adaptation process. This bias stems from the maneuver imbalance present in the ego-centric driving dataset, which can lead to biased training.", + "bbox": [ + 212, + 609, + 787, + 684 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Instead, during the inference stage, we utilize probability density functions for each maneuver based on the intersection-centric dataset to estimate the likelihood of the given map-adapted maneuver context. This approach helps mitigate bias and ensures more accurate maneuver predictions during inference.", + "bbox": [ + 212, + 686, + 787, + 746 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Estimating Likelihood of Maneuvers", + "text_level": 1, + "bbox": [ + 214, + 768, + 560, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Given that the intersection-centric dataset exhibits a well-distributed range of maneuvers, we estimate the likelihood of maneuvers for a given historical trajectory based on the maneuver distribution within this dataset. To achieve this,", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Beyond the Data Imbalance in Vehicle Maneuver Prediction", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fc6d6a87118ee18a72acf114536225eb4a3135a2b217353066390952c807e0fa.jpg", + "image_caption": [ + "Fig. 7: Distribution of the training dataset for each maneuver class and probability density function." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 408, + 263 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/90cc490c362b7667404d674711de456daf9c8ea1ecc482aaa62871cfce709754.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 146, + 594, + 263 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/af82451d2a79d7309d4c0974e15461f497e10309e8c3cd7a35e3f4057a8c552d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 146, + 781, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "we first obtain probability density functions for each maneuver using Gaussian Kernel Density Estimation (Gaussian KDE) [24]. We reduce the dimensionality of maneuver features and then define the probability density function, $f(x, H_d)$ , for each maneuver class using the following equations.", + "bbox": [ + 214, + 329, + 782, + 390 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nK (u) = \\frac {1}{2 \\pi} \\exp \\left(- \\frac {1}{2} u ^ {T} u\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 400, + 785, + 433 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nK _ {H _ {d}} (u) = \\det (H _ {d}) ^ {- 1 / 2} K (H _ {d} ^ {- 1 / 2} u), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 443, + 785, + 465 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nf (x, H _ {d}) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} K _ {H _ {d}} \\left(x - x _ {i}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 479, + 785, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $H_{d}$ is bandwidth matrix parameter for each maneuver $d \\in \\{ST, RT, LT\\}$ and $x_{i=1,\\dots,N}$ is the training data points consist of each maneuver class. Here, $ST, RT$ and $LT$ denote going straight, right turn and left turn respectively. The Epanechnikov kernel method is employed whose bandwidth parameter is defined in Eq. (10):", + "bbox": [ + 214, + 526, + 782, + 602 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nH _ {d} = \\left(\\frac {4 \\sigma_ {d} ^ {5}}{3 N}\\right) ^ {\\frac {1}{5}} \\approx 1. 0 6 \\sigma_ {d} N ^ {- \\frac {1}{5}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 613, + 785, + 651 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\sigma_{d}$ denotes the standard deviation.", + "bbox": [ + 215, + 656, + 514, + 670 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Result and Case Study", + "text_level": 1, + "bbox": [ + 215, + 693, + 473, + 710 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Quantitative Result", + "text_level": 1, + "bbox": [ + 215, + 724, + 426, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The performance of the proposed maneuver classification scheme enhanced by employing heterogeneous datasets is quantitatively compared with the conventional various sampling strategies: a vanilla dataset, oversampled dataset and dataset augmented with SMOTE [5]. We categorize the future maneuvers in Left turn (LT), Straight (ST), and Right turn (RT) from the observed historical trajectories of all surrounding vehicles in the scene.", + "bbox": [ + 214, + 750, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Jeon et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e63e70f73769bc3ecc64cc9debfba6cf0a598c85374c8472f0426333163774a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 338, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/116d4e25ea15a630a29f6ca371bb42a82000450b82bfc760b96f7227d2fcb229.jpg", + "image_caption": [ + "Fig. 8: Baselines and the proposed method in preparing the training dataset and general process in maneuver prediction." + ], + "image_footnote": [], + "bbox": [ + 359, + 143, + 500, + 382 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1979c9a83a1a56a78e476fbb9dbfd023e0d93a6d227f36bfc206dc786ba057e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 143, + 633, + 381 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a9250907369fae345d3faab02555f81fb1b2c7be57fc74709a83396419575d17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 143, + 779, + 382 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/beddf18691df7c3d30b9427bcf05d1b21087d83a85b867c0c623ab7dfdae192f.jpg", + "table_caption": [ + "Table 2: Overall maneuver classification performance of the proposed method comparing with various conventional approaches" + ], + "table_footnote": [], + "table_body": "
Training DatasetLT validation setST validation setRT validation set
RecallPrec.F1AUCRecall.Prec.F1AUCRecallPrec.F1AUC
Vanilla \nArgo.0.0000.5000.0000.6931.0000.8240.9040.5000.0000.5000.0000.594
Argo. + \nOver.0.7430.7590.7510.8790.9390.9430.9410.8490.6960.6440.6690.875
Argo. + \nSMOTE0.7530.7030.7270.9390.8800.9550.9160.9160.8150.4590.5870.927
Ours0.9420.6810.7900.9520.9050.9870.9440.9440.9890.6610.7730.952
", + "bbox": [ + 228, + 488, + 782, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset compositions and training processes of the baselines are illustrated in Fig. 8 by comparing with the proposed framework. To evaluate the maneuver classification performance, we employed the AUC (Area Under Curve) of the ROC (Receiver Operating Characteristic) curve for each maneuver class and supplementary metrics, Precision (Prec.), Recall, and F1 Score (F1). The threshold for the precision and recall is 0.5.", + "bbox": [ + 212, + 700, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Tab. 2, the AUC, which is a representative performance metric, is improved compared to the conventional sampling methods in all maneuver classes. We found that when the model is trained with the vanilla Argoverse", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Beyond the Data Imbalance in Vehicle Maneuver Prediction", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/48dd99aa2223520b7e8daa01b0e6192876bae021436f6c9df42a5005fdf90d7d.jpg", + "image_caption": [ + "Fig. 9: Data distribution comparison according to the temporal distance to the end of the intersection." + ], + "image_footnote": [], + "bbox": [ + 233, + 146, + 491, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fccd5a0cbcc9232247e111b83557b3772410f5801f401c95b6e77c325e9f0d7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 146, + 769, + 299 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "dataset (only ego-centric), it predicts most samples as ST, thereby achieving a high recall but a low precision, which the model is extremely biased toward outputing ST. The zero recall in LT and RT (which means there are no predictions for LT and RT with a score larger than 0.5.) also proves the biased training.", + "bbox": [ + 212, + 367, + 782, + 428 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Although the model achieves better performance than the vanilla dataset case by employing the oversampling strategy or SMOTE, they are still inferior than the proposed heterogeneous dataset result. Therefore, We argue that sampling-based methods which try to uniformize only the number of samples while ignoring the importance of sample diversity, is unable to overcome the imbalanced dataset problem. Conclusively, introducing the higher variety in the training samples by leveraging the proposed intersection-centric dataset is required for classifying the future maneuver of vehicles without bias.", + "bbox": [ + 212, + 429, + 787, + 550 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Qualitative Result", + "text_level": 1, + "bbox": [ + 215, + 580, + 413, + 594 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Maneuver Representation Learning. We visualized the learned representations for each maneuver class using [22] to analyze the behavior of the maneuver classification model. By reducing the dimension of the interaction-aware maneuver context, $c$ , every samples in the training dataset are mapped to the two dimensional space as shown in Fig. 9.", + "bbox": [ + 212, + 611, + 782, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Fig. 9, the maneuver contexts for each maneuver class are scattered according to the time gap to the intersection. The points in blue, cyan, and green denote going straight (ST), left turn (LT) and right turn (RT) maneuver, respectively. The blue points show that the ST maneuver is less spread compared to other maneuvers even at the 4 seconds before the intersection, which implies that the ST maneuver is much easier to distinguish than other turning maneuvers even if remarkable maneuver feature is not observed. On the other hand, turning maneuvers with cyan and green points are widely distributed invading the ST area in the left plot in Fig. 9, which means that the turning maneuvers are likely to be misclassified as the ST.", + "bbox": [ + 212, + 689, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Jeon et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4fe9eee2a288241b3582f58eccc0a9bc0ddf915c77f98c6f1a8b028db89e1dbc.jpg", + "image_caption": [ + "Fig. 10: Data distribution comparison according to the training dataset." + ], + "image_footnote": [], + "bbox": [ + 233, + 146, + 491, + 299 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e6cd5575eae06949054bb168c049a3babca287a246e336dc6e879ae1da1348c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 146, + 769, + 299 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Furthermore, as shown in Fig. 10, the distribution of data points is prominently different according to the dataset. In Fig. 10, the data points in the red circles indicate the challenging scenarios: all of three maneuvers are blended in this area so the maneuver representation is hard to be categorized. However, in the imbalanced dataset, the distance from the ambiguous area (the red circle) to the center of the ST maneuver cluster is smaller than other maneuvers, whereas the distances toward each maneuver class are similar in the balanced dataset. We argue that this characteristic of maneuver representations induce the model to generate the biased predictions toward the ST when trained with the imbalanced dataset.", + "bbox": [ + 212, + 351, + 787, + 501 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Case Study on Maneuver Classification. In Fig. 11, use case of the proposed maneuver prediction model is illustrated. Using the historical trajectory, blue lines, each maneuver probability, bar graph, is calculated considering road shape and multiple surrounding vehicles, black dotted lines and grey dots, respectively. Since blue dotted line denotes the ground truth future trajectory, predicted maneuver probability can be qualitatively evaluated.", + "bbox": [ + 212, + 525, + 787, + 616 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in (a) and (b) in Fig. 11, LT maneuver is well predicted even though the turning maneuver is not shown in the historical trajectory; maneuver probability of the LT class takes the highest likelihood. Furthermore, despite the ground truth label of LT class, ST and/or RT probability is also generated considering the road shape, which enables the planning algorithm to consider low probability cases so that abnormal situations can be handled. Similarly, ST and RT maneuver are well-predicted as shown in 11: (c), (d) for ST class and (e), (f) for RT class.", + "bbox": [ + 212, + 616, + 787, + 738 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 762, + 359, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the deep learning based autonomous driving system, the imbalanced data was the critical bottleneck in generalizing the model because of the nature of naturalistic driving dataset: long-tail problem. In the proposed research, for overcoming", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Beyond the Data Imbalance in Vehicle Maneuver Prediction", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f91f8630a54df143e0b847570433331cb251a5008b63e764e7ff27055b59226e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 233, + 162, + 410, + 277 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a0f48d55c7d400b28ee5f405d06faeae9ad405ca647503eaa8d4c64afd5d422e.jpg", + "image_caption": [ + ": Historical trajectory (input) : Current position : Future trajectory (GT)", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 413, + 162, + 588, + 277 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/82f819ddd300d2df4c2f15ba92acde8133001b5b658ba9c0d6cbd41bc8d6a2fa.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 591, + 164, + 767, + 277 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/41b651fe1d23f6b197391a5ef9bc436d02bf121de608cb6527b204b3be95162c.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 233, + 295, + 410, + 409 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/afc04b818a66a6a01ee282596b54682dc237409a221a9e1cea5595bb67339c8d.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 413, + 296, + 588, + 409 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6e86913b894c11f150af32424a9fc5819fbbed9358408614eae4b392ada42cb8.jpg", + "image_caption": [ + "(f)", + "Fig. 11: Prediction results for the representative scenarios. Left turn $(a, b)$ , going straight $(c, d)$ and right turn $(e, f)$ maneuver can be predicted as a stochastic manner." + ], + "image_footnote": [], + "bbox": [ + 591, + 296, + 767, + 409 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the data imbalance and biased prediction model, two dataset from difference modality is employed: in-house drone dataset and Argoverse dataset. The in-house Drone dataset has quantitatively abundant scenarios including a bunch of turning maneuvers, where the Argoverse dataset contains various road shape. By merging the strength of the heterogeneous dataset into a maneuver prediction model by introducing a two-stage training scheme with self-supervised learning, geometrically generalizable prediction model can be trained in balance in terms of the vehicle maneuver. Additionally, the probabilistic approach can provide additional information on the unlikely to happen maneuver so that the planning algorithm can consider abnormal driving situation.", + "bbox": [ + 212, + 487, + 787, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "However, the proposed prediction method can not consider the dependency between the future of the surrounding vehicle and the action plan of the ego vehicle. In real world, vehicles are continuously interacting each other and update their own driving strategy in real time. Therefore, for highly intelligent prediction network, the effect of the ego vehicle's future action plan has to be considered in predicting the maneuver of the other vehicles [18]. In this stage, simulation based approaches (e.g. MORAI SIM: Drive [1]) is highly promising to handle both of the data imbalance problem and reactiveness of the actors.", + "bbox": [ + 212, + 641, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by Autonomous Driving Development Center, AVP Division, Hyundai Motor Company and the National Research Foundation of Korea(NRF) funded by the Korea government(MSIT) under Grants 2022R1A2C200494413.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Jeon et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Morai simulator, https://www.morai.ai/drive 14", + "2. Azadani, M.N., Boukerche, A.: A novel multimodal vehicle path prediction method based on temporal convolutional networks. IEEE Transactions on Intelligent Transportation Systems (2022) 5", + "3. Bock, J., Krajewski, R., Moers, T., Runde, S., Vater, L., Eckstein, L.: The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 1929-1934. IEEE (2020) 6", + "4. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8748-8757 (2019) 8", + "5. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research 16, 321-357 (2002) 2, 4, 10", + "6. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014) 7", + "7. Cieslak, D.A., Chawla, N.V., Striegel, A.: Combating imbalance in network intrusion datasets. In: GrC. pp. 732-737. Citeseer (2006) 4", + "8. Cui, Y., Song, Y., Sun, C., Howard, A., Belongie, S.: Large scale fine-grained categorization and domain-specific transfer learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4109-4118 (2018) 4", + "9. Deepa, T., Punithavalli, M.: An e-smote technique for feature selection in high-dimensional imbalanced dataset. In: 2011 3rd International Conference on Electronics Computer Technology. vol. 2, pp. 322-324. IEEE (2011) 4", + "10. Drummond, C., Holte, R.C., et al.: C4. 5, class imbalance, and cost sensitivity: why under-sampling beats over-sampling. In: Workshop on learning from imbalanced datasets II. vol. 11, pp. 1-8. Citeseer (2003) 2, 4", + "11. Feng, X., Cen, Z., Hu, J., Zhang, Y.: Vehicle trajectory prediction using intention-based conditional variational autoencoder. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 3514-3519. IEEE (2019) 4", + "12. Girase, H., Gang, H., Malla, S., Li, J., Kanehara, A., Mangalam, K., Choi, C.: Loki: Long term and key intentions for trajectory prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9803-9812 (2021) 2, 4", + "13. Goldhammer, M., Kohler, S., Zernetsch, S., Doll, K., Sick, B., Dietmayer, K.: Intentions of vulnerable road users—detection and forecasting by means of machine learning. IEEE transactions on intelligent transportation systems 21(7), 3035-3045 (2019) 6", + "14. He, H., Garcia, E.A.: Learning from imbalanced data. IEEE Transactions on knowledge and data engineering 21(9), 1263-1284 (2009) 4", + "15. Hu, Y., Li, Y., Huang, H., Lee, J., Yuan, C., Zou, G.: A high-resolution trajectory data driven method for real-time evaluation of traffic safety. Accident Analysis & Prevention 165, 106503 (2022) 5", + "16. J. Colyar, J.H.: Us highway 101 dataset. Federal Highway Administration (FHWA), Tech. Rep. FHWA-HRT-07-030 (2007) 6" + ], + "bbox": [ + 218, + 179, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Beyond the Data Imbalance in Vehicle Maneuver Prediction", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Jeon, H., Choi, J., Kum, D.: Scale-net: Scalable vehicle trajectory prediction network under random number of interacting vehicles via edge-enhanced graph convolutional neural network. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2095-2102. IEEE (2020) 7", + "18. Jeon, H., Kim, S., Lee, K., Kang, D., Choi, J., Kum, D.: Are reactions to ego vehicles predictable without data?: A semi-supervised approach. IEEE Transactions on Intelligent Transportation Systems 24(6), 6477-6490 (2023) 14", + "19. Kim, S., Jeon, H., Choi, J.W., Kum, D.: Diverse multiple trajectory prediction using a two-stage prediction network trained with lane loss. IEEE Robotics and Automation Letters 8(4), 2038-2045 (2022) 1", + "20. Kozerawski, J., Sharan, M., Yu, R.: Taming the long tail of deep probabilistic forecasting. arXiv preprint arXiv:2202.13418 (2022) 2, 4", + "21. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: European Conference on Computer Vision. pp. 541-556. Springer (2020) 1, 9", + "22. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 12", + "23. Makansi, O., Çiçek, Ö., Marrakchi, Y., Brox, T.: On exposing the challenging long tail in future prediction of traffic actors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13147-13157 (2021) 2, 4", + "24. Marron, J.S., Wand, M.P.: Exact mean integrated squared error. The Annals of Statistics 20(2), 712-736 (1992) 10", + "25. Mirus, F., Stewart, T.C., Conradt, J.: The importance of balanced data sets: Analyzing a vehicle trajectory prediction model based on neural networks and distributed representations. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020) 5", + "26. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 8", + "27. Park, D., Ryu, H., Yang, Y., Cho, J., Kim, J., Yoon, K.J.: Leveraging future relationship reasoning for vehicle trajectory prediction. arXiv preprint arXiv:2305.14715 (2023) 1", + "28. Peng, J., Bu, X., Sun, M., Zhang, Z., Tan, T., Yan, J.: Large-scale object detection in the wild from imbalanced multi-labels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9709-9718 (2020) 2, 4", + "29. Robicquet, A., Sadeghian, A., Alahi, A., Savarese, S.: Learning social etiquette: Human trajectory prediction in crowded scenes. In: European Conference on Computer Vision (ECCV). vol. 2 (2020) 6", + "30. Shangguan, Q., Fu, T., Wang, J., Fu, L., et al.: A proactive lane-changing risk prediction framework considering driving intention recognition and different lane-changing patterns. Accident Analysis & Prevention 164, 106500 (2022) 5", + "31. Shen, L., Lin, Z., Huang, Q.: Relay backpropagation for effective learning of deep convolutional neural networks. In: European conference on computer vision. pp. 467-482. Springer (2016) 2, 4", + "32. Strigel, E., Meissner, D., Seeliger, F., Wilking, B., Dietmayer, K.: The ko-per intersection laserscanner and video dataset. In: 17th International IEEE Conference on Intelligent Transportation Systems (ITSC). pp. 1900-1901. IEEE (2014) 6", + "33. Varadarajan, B., Hefny, A., Srivastava, A., Refaat, K.S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C.P., Anguelov, D., et al.: Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 7814-7821. IEEE (2022) 1" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Jeon et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Wang, J., Xu, M., Wang, H., Zhang, J.: Classification of imbalanced data by using the smote algorithm and locally linear embedding. In: 2006 8th international Conference on Signal Processing. vol. 3. IEEE (2006) 4", + "35. Wang, K., Xue, Q., Lu, J.J.: Risky driver recognition with class imbalance data and automated machine learning framework. International journal of environmental research and public health 18(14), 7534 (2021) 2, 4", + "36. Westny, T., Frisk, E., Olofsson, B.: Vehicle behavior prediction and generalization using imbalanced learning techniques. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2003-2010. IEEE (2021) 5", + "37. Yang, D., Li, L., Redmill, K., Özgün, Ü.: Top-view trajectories: A pedestrian dataset of vehicle-crowd interaction from controlled experiments and crowded campus. In: 2019 IEEE Intelligent Vehicles Symposium (IV). pp. 899-904. IEEE (2019) 6", + "38. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022) 1" + ], + "bbox": [ + 212, + 146, + 787, + 369 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Beyond the Data Imbalance in Vehicle Maneuver Prediction", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_model.json b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ef4bd7846d92fd19b30c01b3d48a872a62bd11d2 --- /dev/null +++ b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_model.json @@ -0,0 +1,2280 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.768, + 0.208 + ], + "angle": 0, + "content": "Beyond the Data Imbalance: Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.234, + 0.776, + 0.267 + ], + "angle": 0, + "content": "Hyeongseok Jeon\\(^{1}\\), Sanmin Kim\\(^{2}\\), Abi Rahman Syamil\\(^{2}\\), Junsoo Kim\\(^{3}\\), and Dongsuk Kum\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.277, + 0.761, + 0.337 + ], + "angle": 0, + "content": "\\(^{1}\\) MORAI Inc., Seoul, South Korea \n\\(^{2}\\) Korea Advanced Institute of Science and Technology, Daejeon, South Korea \n\\(^{3}\\) Hyundai Motor Company, Seoul, South Korea \nhsjeon@morai.ai dskum@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.368, + 0.744, + 0.633 + ], + "angle": 0, + "content": "Abstract. Predicting the maneuvers of surrounding vehicles is imperative for the safe navigation of autonomous vehicles. However, naturalistic driving datasets tend to be highly imbalanced, with a bias towards the \"going straight\" maneuver. Consequently, learning and accurately predicting turning maneuvers pose significant challenges. In this study, we propose a novel two-stage maneuver learning method that can overcome such strong biases by leveraging two heterogeneous datasets in a complementary manner. In the first training phase, we utilize an intersection-centric dataset characterized by balanced distribution of maneuver classes to learn the representations of each maneuver. Subsequently, in the second training phase, we incorporate an ego-centric driving dataset to account for various geometrical road shapes, by transferring the knowledge of geometric diversity to the maneuver prediction model. To facilitate this, we constructed an in-house intersection-centric trajectory dataset with a well-balanced maneuver distribution. By harnessing the power of heterogeneous datasets, our framework significantly improves maneuver prediction performance, particularly for minority maneuver classes such as turning maneuvers. The dataset is available at https://github.com/KAIST-VDCLab/VDC-Trajectory-Dataset." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.645, + 0.744, + 0.674 + ], + "angle": 0, + "content": "Keywords: Maneuver Prediction \\(\\cdot\\) Long-tail Problem \\(\\cdot\\) Autonomous Driving" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.719, + 0.378, + 0.735 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Predicting the motion of surrounding vehicles is a critical component for ensuring the safety of autonomous driving. In recent years, numerous studies have introduced motion prediction models [19, 21, 27, 33, 38] that primarily focus on generating detailed trajectories. These approaches are trained and evaluated using trajectory-level metrics such as ADE/FDE (Average/Final Distance Error), assessing how closely predicted trajectories align with actual driving trajectories." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.344, + 0.127 + ], + "angle": 0, + "content": "Jeon et al." + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.147, + 0.756, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.352, + 0.785, + 0.38 + ], + "angle": 0, + "content": "Fig. 1: Importance of maneuver prediction. In a intersection scenario, accurate prediction of vehicles' turning maneuver is crucial." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.534 + ], + "angle": 0, + "content": "However, from the perspective of motion planning, it becomes clear that accurately predicting the high-level intentions of surrounding vehicles holds greater importance than providing fine-grained future predictions. For example, in intersection scenarios where vehicle interactions are maximized, precisely determining a vehicle's turning maneuver is more crucial than outputting detailed future trajectories, which may end up being a wrong maneuver, as shown in Fig. 1. Hence, prioritizing the accurate prediction of vehicle intentions over detailed trajectory predictions becomes paramount for safe motion planning in autonomous driving, particularly in intricate scenarios like intersections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.787, + 0.672 + ], + "angle": 0, + "content": "Various datasets have been collected and released for the motion prediction task, gathered by driving sensor-equipped vehicles in real-world environments to ensure scenario diversity. While these datasets provide ample trajectories for training motion prediction models, they exhibit highly biased distributions due to the inherent nature of driving trajectories, particularly concerning maneuvers. Straight roads dominate the driving environment, leading to most trajectory datasets being biased towards straight trajectories. Consequently, maneuver classification models trained with biased datasets tend to be biased toward scenarios involving straight driving while learning less about other maneuvers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Several approaches have been proposed in order to address the long-tail problem. Sampling strategies [5, 10, 28, 31] have attempted to handle this issue by either oversampling rare cases or downsampling dominant cases. Alternatively, loss function reformulation strategies [12, 20, 23, 35] have been explored to adjust gradients and enforce balanced supervision during training. However, these approaches have shown limited success, as they merely manipulate the training dataset to achieve balance or regulate the learning process within the existing dataset. Consequently, the fundamental problem stemming from the long-tailed nature of the data has not been effectively addressed." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To address these challenges, we present a novel two-stage learning framework that harnesses the power of heterogeneous datasets. Our approach aims to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.149, + 0.501, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.149, + 0.782, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.284, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Fig. 2: Pros and Cons of the two different datasets: An intersection-centric dataset acquired from the bird's eye view and an ego-centric dataset that collected in a vehicle view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.614 + ], + "angle": 0, + "content": "complement the naturally imbalanced driving dataset by incorporating a cost-efficient and easily acquirable intersection-centric dataset. In our proposed framework, we recognize the importance of capturing various turning maneuvers commonly observed in intersections. Towards that end, we introduce a drone-based intersection-centric dataset, specifically chosen to mitigate the limitations of the ego-centric driving dataset. We utilized the in-house intersection-centric dataset by focusing on a representative intersection area with the aim of learning maneuver representations. To effectively combine these heterogeneous datasets for training maneuver prediction models, we first extract valuable knowledge of each maneuver class from the in-house intersection-centric trajectory dataset. This knowledge, which is challenging to learn from ego-centric driving datasets alone, is then transferred to the downstream network. Here, the model learns the geometric diversity leveraging the ego-centric driving dataset. This approach ensures that our model will initially acquire rich intersection-oriented maneuver knowledge in a balanced manner, while the downstream network then adapts the maneuver representations to diverse road geometries. As a result, our model simultaneously learns maneuver representations from the intersection-centric dataset and features of diverse road geometries from the ego-centric driving dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.675 + ], + "angle": 0, + "content": "To the best of our knowledge, this study is the first attempt to train the maneuver (or trajectory) prediction model with the heterogeneous datasets in complementary manner. The main contributions of our work are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.68, + 0.785, + 0.724 + ], + "angle": 0, + "content": "- We propose a novel two-stage maneuver prediction framework that effectively leverages heterogeneous datasets to address the long-tail problem in motion prediction." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.725, + 0.785, + 0.755 + ], + "angle": 0, + "content": "- We introduce a novel trajectory dataset collected specifically at intersections, ensuring a balanced distribution of maneuvers." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.68, + 0.785, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.778, + 0.388, + 0.794 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Since the data imbalance problem in deep learning approaches appears widely, general solutions for handling the long tail problem is introduced. However, those" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.344, + 0.128 + ], + "angle": 0, + "content": "Jeon et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "approaches are only focused on formulating the training dataset or loss functions based on the given imbalanced dataset regardless of the data diversity. Therefore, the diversity of the scenarios for training the network still remain insufficient." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.217, + 0.494, + 0.234 + ], + "angle": 0, + "content": "2.1 Sampling-based Approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.784, + 0.304 + ], + "angle": 0, + "content": "In sampling-based approaches, the training dataset is selectively or redundantly reformulated by choosing the data from the original dataset. During this reformulation process, over- and/or under-sampling methods are employed to prevent overfitting resulting from imbalanced datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.789, + 0.606 + ], + "angle": 0, + "content": "In studies such as [10, 11, 14], balanced training datasets are constructed through an undersampling mechanism, which involves discarding excessive data instances from the major class of scenarios in the given dataset. In these approaches, data instances from the dominant class are underused to balance the distribution of the training dataset. In contrast, [8,28,31] repeatedly sample minorities in datasets to formulate evenly distributed training datasets using various oversampling algorithms. [31] constructs an equally-distributed mini-batch with a class-balanced composition named class-aware sampling. In [28], a soft-balance sampling method is employed for probabilistic oversampling, while [8] adopts a two-stage training strategy that utilizes both the original dataset and the balanced dataset with different learning rates. Furthermore, to account for randomness in the oversampling process, [5] proposes a synthetic data oversampling algorithm named SMOTE (Synthetic Minority Over-sampling TEchnique), with many variants suggested in subsequent studies [5,7,9,34]. However, these sampling-based data balancing approaches have shown unsatisfactory performance in predicting the future maneuver of vehicles. This is because sampling-based methods merely reconstruct the training dataset within the imbalanced raw dataset. However, in autonomous driving, maneuver prediction networks should observe as many diverse samples as possible to create a generalized prediction model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.502, + 0.646 + ], + "angle": 0, + "content": "2.2 Loss Function Reformulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Another approach to address imbalanced datasets is through loss function reformulation, as demonstrated in studies such as [12, 20, 23, 35]. These approaches introduce auxiliary losses to address minority cases. In [12] and [35], a simple weighting method is employed, where larger weights are assigned to gradients for minor classes to equalize the backpropagating effect across all classes. In other studies like [20] and [23], various loss functions are introduced to reshape the minority distribution. In [20], the learning of easy and difficult examples is reprioritized by modeling behavior with the Generalized Pareto Distribution (GPD) using the Pareto Loss. Conversely, in [23], the representation of internal maneuver context is reshaped using contrastive loss, ensuring that difficult examples are well-distributed throughout the representation space. While these loss reformulation approaches attempt to address the data imbalance problem during the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.152, + 0.784, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Fig.3: Overall architecture of the proposed method. Maneuver representations are extracted from the intersection-centric dataset and adapted to the diverse geometries with the ego-centric driving dataset. Using the maneuver context distribution form intersection dataset, each maneuver score is calculated based on the Gaussian KDE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.452 + ], + "angle": 0, + "content": "training stage, they still share the same limitation as sampling approaches in that they have a marginal impact on dataset diversity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.47, + 0.593, + 0.485 + ], + "angle": 0, + "content": "2.3 Long-tail Problem in Motion Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.703 + ], + "angle": 0, + "content": "Building an unbiased maneuver prediction model is challenging due to the vehicle maneuvers' inherent long-tail characteristics, even if they are trained with the numerous driving datasets already released. In [25], the influence of dataset composition on performance is comprehensively analyzed, revealing that a balanced dataset containing evenly distributed maneuvers, such as lane changes and straight movements, could enhance overall performance. As a result, several maneuver and trajectory prediction studies [2, 36] have employed oversampling approaches to avoid biased training. These frameworks have demonstrated effectiveness in the field of vehicle maneuver prediction [15, 30]. However, sampling approaches alone are insufficient to overcome the fundamental limitations of long-tailed datasets, and the maneuver prediction task is no exception. Therefore, a supplementary dataset containing numerous samples of minor maneuvers, such as turning maneuvers, is necessary to augment the observable trajectory samples in the training dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.722, + 0.394, + 0.74 + ], + "angle": 0, + "content": "3 Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To establish a comprehensive approach that encompasses both balanced maneuver distribution and geometric diversity, we introduce a two-stage maneuver prediction framework leveraging heterogeneous datasets. Firstly, we construct an intersection-centric trajectory dataset characterized by a well-balanced distribution of maneuvers. This dataset serves as the foundation for obtaining maneuver representations, achieved through training a maneuver backbone using" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.344, + 0.128 + ], + "angle": 0, + "content": "Jeon et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation on the in-house drone dataset with various conventional datasets from the fixed view point." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.185, + 0.784, + 0.382 + ], + "angle": 0, + "content": "
DatasetLocationTrajectory CountsLocation CountsObjectsFPSMethod
NGSIM [16]boulevard33664car10HzRSU
Ko-PER [32]intersection3501pedestrian, bicycle, car25HzRSU
VRU [13]intersection32781pedestrian, bicycle25HzRSU
DUT [37]campus18622pedestrian, vehicles24Hzdrone
SDD [29]campus102408pedestrian, bicycle, car, skateboard, cart, bus25Hzdrone
InD [3]intersection135994pedestrian, bicycle, car, truck, bus25Hzdrone
Oursintersection, lane merge1547910pedestrian, bicycle, car30Hzdrone
" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.395, + 0.497, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.396, + 0.768, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.525, + 0.785, + 0.555 + ], + "angle": 0, + "content": "Fig. 4: Sample scenes in our in-house intersection-centric dataset. Car, bicycle and pedestrian are represented as yellow, cyan and green." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.707 + ], + "angle": 0, + "content": "a self-supervised learning approach. Additionally, we derive probability density functions representing each maneuver by applying Kernel Density Function to the training set of the intersection dataset. Then, we bolster the model's scalability by injecting geometric diversity via the utilization of an ego-centric driving dataset. This step enriches the model's comprehension of diverse road geometries, thereby enhancing its predictive accuracy. During the inference stage, maneuver scores are computed by estimating the likelihood of the given trajectory based on the obtained maneuver probability density functions. The overall diagram for our proposed maneuver prediction framework is depicted in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.727, + 0.564, + 0.743 + ], + "angle": 0, + "content": "3.1 Learning Maneuver Representations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "An Intersection-centric In-house dataset. To ensure a balanced distribution of vehicle maneuvers in our dataset, we construct an in-house intersection-centric dataset. This dataset leverages stationary drone videos, which are easily acquired and require minimal labeling costs compared to sensor-based driving datasets. Samples from the in-house dataset, showcasing abundant turning maneuvers with vehicles, bicycles, and pedestrians depicted in yellow, green, and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.149, + 0.784, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.294, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 5: Network architecture of the maneuver backbone (left) and auxiliary module for training the backbone network (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.46 + ], + "angle": 0, + "content": "cyan colors respectively, are illustrated in Fig. 4. The dataset comprises 3.5 hours of video footage captured across 10 locations in 4K resolution at 30 frames per second. Additionally, the dataset includes not only agent trajectories but also high-definition (HD) map data. Furthermore, a digital twin environment is constructed to simulate scenarios within the dataset. A comprehensive evaluation of the in-house drone dataset, compared with other bird's eye view (BEV) datasets, is summarized in Tab. 1. Utilizing this dataset, we train the vehicle's maneuver knowledge in a balanced manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Maneuver Backbone. In the first training phase, we extract maneuver features by utilizing the intersection-centric dataset, which is characterized by uniformly distributed maneuvers. To capture maneuver representations from the observed historical trajectory of each vehicle, we employ a self-supervised learning approach, as illustrated in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.587 + ], + "angle": 0, + "content": "First, the observed historical trajectory is encoded into a state embedding using an encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.6, + 0.787, + 0.618 + ], + "angle": 0, + "content": "\\[\nh _ {i} ^ {t} = f _ {e} \\left(s _ {i} ^ {t} = \\left(x _ {i} ^ {t}, y _ {i} ^ {t}, \\theta_ {i} ^ {t}\\right) \\mid_ {t = - t h \\dots , 0}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.787, + 0.667 + ], + "angle": 0, + "content": "where \\( f_{e} \\) represents the encoder, \\( h_{i}^{t} \\) denotes a state embedding, \\( s_{i}^{t} = (x_{i}^{t},y_{i}^{t},\\theta_{i}^{t}) \\) denotes the state of the \\( i^{th} \\) vehicle at timestep \\( t \\), and \\( t_h \\) represents a observation time horizon. We employ 1D CNN and MLP for the encoder." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.668, + 0.788, + 0.728 + ], + "angle": 0, + "content": "The state embeddings are then fed into an autoregressive model (e.g. GRU [6]) to extract temporal trajectory features. Subsequently, an interaction model (e.g. ScaleNet [17]) is employed to account for interaction among vehicles in multi-vehicle scenarios:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.742, + 0.787, + 0.761 + ], + "angle": 0, + "content": "\\[\nc = f _ {v} \\left(f _ {a} \\left(h _ {i} ^ {t} | _ {t = - t _ {h}, \\dots , 0}\\right) | _ {i = 1, \\dots , n}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.809 + ], + "angle": 0, + "content": "where \\( c = \\{c_1, \\dots, c_n\\} \\), \\( f_v \\) and \\( f_a \\) represent the extracted maneuver representations, an interaction module, and an autoregressive model, respectively. \\( n \\) denotes the number of vehicles in the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To learn the internal hidden representation of each maneuver, we employ a self-supervised learning framework. Based on the Contrastive Predictive Coding" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.128 + ], + "angle": 0, + "content": "Jeon et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "framework [26], the maneuver backbone is trained to maximize mutual information (MI) between the interaction-aware maneuver context of the historical view, \\( c_{i} \\), and the state embedding in the future \\( h_i^t|_{t\\geq 0} \\). By maximizing the MI between historical trajectory and future position of the vehicle, the network can learn the slow-varying high level maneuver context of the vehicles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.252 + ], + "angle": 0, + "content": "Based on the definition, MI between \\( c_{i} \\) and \\( h_i^t|_{t\\geq 0} \\), \\( I(h_i^t,c_i) \\) is formulated as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.265, + 0.786, + 0.302 + ], + "angle": 0, + "content": "\\[\nI \\left(h _ {i} ^ {t}, c _ {i}\\right) = \\sum_ {i = 1} p \\left(h _ {i} ^ {t}, c _ {i}\\right) \\log \\frac {p \\left(h _ {i} ^ {t} , c _ {i}\\right)}{p \\left(h _ {i} ^ {t}\\right)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.352 + ], + "angle": 0, + "content": "As the MI is hardly tractable within the batch training process, the auxiliary function named as MI score function is defined as Eq. (4) with trainable auxiliary linear transformation coefficients \\( w^{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.365, + 0.786, + 0.4 + ], + "angle": 0, + "content": "\\[\nf _ {k} \\left(h _ {i} ^ {t}, c _ {i}\\right) \\equiv e x p \\left(\\left(h _ {i} ^ {t}\\right) ^ {T} w ^ {t} c _ {i}\\right) \\propto \\frac {p \\left(h _ {i} ^ {t} , c _ {i}\\right)}{p \\left(h _ {i} ^ {t _ {d}}\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.784, + 0.452 + ], + "angle": 0, + "content": "Since MI score function is proportional to \\( I(h_i^t, c_i) \\) as formulated in Eq. (4), MI score function is maximized for the same vehicle and minimized for different vehicles using the infoNCE loss [26] function, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.465, + 0.786, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left[ \\frac {1}{t _ {f}} \\sum_ {t = 1} ^ {t _ {f}} \\log \\frac {f _ {k} \\left(h _ {i} ^ {t} , c _ {i}\\right)}{\\Sigma_ {j = 1} ^ {n} f _ {k} \\left(h _ {j} ^ {t} , c _ {i}\\right)} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.784, + 0.572 + ], + "angle": 0, + "content": "where \\( n \\) denotes the number of data instances in a single batch and \\( t_f \\) is the future time horizon. With this loss function, the MI between the history and future states of the identical vehicle is enlarged, whereas the MI with other vehicles is suppressed." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.595, + 0.569, + 0.61 + ], + "angle": 0, + "content": "3.2 Generalizing for Geometric Diversity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.785, + 0.77 + ], + "angle": 0, + "content": "An Ego-centric Driving Dataset. The intersection-centric dataset used for maneuver representation learning has a limitation in that it lacks geometric diversity. To address this limitation and ensure the model's generalizability across various geometric shapes of driving environments, we incorporate an ego-centric driving dataset [4] alongside the intersection-centric dataset. The ego-centric driving dataset is constructed using sensor-equipped vehicles, providing a diverse range of driving environments with various road shapes. By integrating this dataset into our training pipeline, we can effectively train the map adaptation module and enhance the model's ability to generalize to different geometric scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.843 + ], + "angle": 0, + "content": "Map Adaptation Module. In the proposed maneuver prediction model, we introduce the map adaptation module as a downstream network aimed at generalizing the pre-trained maneuver representations to various road shapes. For map" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.153, + 0.768, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.393, + 0.7, + 0.407 + ], + "angle": 0, + "content": "Fig. 6: Overall architecture of the map adaptation modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.553 + ], + "angle": 0, + "content": "adaptation, we adopt an off-the-shelf map-based prediction model, LaneGCN [21]. LaneGCN utilizes lane graphs and Graph Convolutional Networks (GCN) to encode road geometries. It consists of multiple modules for encoding and fusing geometric and trajectory features, including MapNet and Lane to Lane for map encoding based on GCN, and Actor to Lane and Lane to Actor to fuse map and trajectory features using an attention mechanism. The overall network is trained using cross-entropy loss as defined in Eq. (6), where \\( y_{i} \\) and \\( p_i \\) represent the ground truth and prediction of the vehicle maneuver, respectively." + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = - \\sum_ {i = 1} ^ {k} y _ {i} \\log \\left(p _ {i}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.685 + ], + "angle": 0, + "content": "It is important to note that the maneuver head is employed only during the training process. This is due to the potential bias that may arise in the model outputting predominantly straight maneuvers during the map adaptation process. This bias stems from the maneuver imbalance present in the ego-centric driving dataset, which can lead to biased training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Instead, during the inference stage, we utilize probability density functions for each maneuver based on the intersection-centric dataset to estimate the likelihood of the given map-adapted maneuver context. This approach helps mitigate bias and ensures more accurate maneuver predictions during inference." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.562, + 0.785 + ], + "angle": 0, + "content": "3.3 Estimating Likelihood of Maneuvers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Given that the intersection-centric dataset exhibits a well-distributed range of maneuvers, we estimate the likelihood of maneuvers for a given historical trajectory based on the maneuver distribution within this dataset. To achieve this," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.344, + 0.127 + ], + "angle": 0, + "content": "Jeon et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.409, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.147, + 0.596, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.147, + 0.782, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.279, + 0.785, + 0.307 + ], + "angle": 0, + "content": "Fig. 7: Distribution of the training dataset for each maneuver class and probability density function." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.33, + 0.784, + 0.391 + ], + "angle": 0, + "content": "we first obtain probability density functions for each maneuver using Gaussian Kernel Density Estimation (Gaussian KDE) [24]. We reduce the dimensionality of maneuver features and then define the probability density function, \\( f(x, H_d) \\), for each maneuver class using the following equations." + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.401, + 0.786, + 0.434 + ], + "angle": 0, + "content": "\\[\nK (u) = \\frac {1}{2 \\pi} \\exp \\left(- \\frac {1}{2} u ^ {T} u\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.444, + 0.786, + 0.466 + ], + "angle": 0, + "content": "\\[\nK _ {H _ {d}} (u) = \\det (H _ {d}) ^ {- 1 / 2} K (H _ {d} ^ {- 1 / 2} u), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.481, + 0.786, + 0.522 + ], + "angle": 0, + "content": "\\[\nf (x, H _ {d}) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} K _ {H _ {d}} \\left(x - x _ {i}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.527, + 0.784, + 0.603 + ], + "angle": 0, + "content": "where \\( H_{d} \\) is bandwidth matrix parameter for each maneuver \\( d \\in \\{ST, RT, LT\\} \\) and \\( x_{i=1,\\dots,N} \\) is the training data points consist of each maneuver class. Here, \\( ST, RT \\) and \\( LT \\) denote going straight, right turn and left turn respectively. The Epanechnikov kernel method is employed whose bandwidth parameter is defined in Eq. (10):" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.614, + 0.786, + 0.652 + ], + "angle": 0, + "content": "\\[\nH _ {d} = \\left(\\frac {4 \\sigma_ {d} ^ {5}}{3 N}\\right) ^ {\\frac {1}{5}} \\approx 1. 0 6 \\sigma_ {d} N ^ {- \\frac {1}{5}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.657, + 0.515, + 0.671 + ], + "angle": 0, + "content": "where \\(\\sigma_{d}\\) denotes the standard deviation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.475, + 0.712 + ], + "angle": 0, + "content": "4 Result and Case Study" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.427, + 0.74 + ], + "angle": 0, + "content": "4.1 Quantitative Result" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The performance of the proposed maneuver classification scheme enhanced by employing heterogeneous datasets is quantitatively compared with the conventional various sampling strategies: a vanilla dataset, oversampled dataset and dataset augmented with SMOTE [5]. We categorize the future maneuvers in Left turn (LT), Straight (ST), and Right turn (RT) from the observed historical trajectories of all surrounding vehicles in the scene." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.339, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.145, + 0.501, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.145, + 0.634, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.145, + 0.781, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.401, + 0.785, + 0.43 + ], + "angle": 0, + "content": "Fig. 8: Baselines and the proposed method in preparing the training dataset and general process in maneuver prediction." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.449, + 0.785, + 0.478 + ], + "angle": 0, + "content": "Table 2: Overall maneuver classification performance of the proposed method comparing with various conventional approaches" + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.489, + 0.784, + 0.673 + ], + "angle": 0, + "content": "
Training DatasetLT validation setST validation setRT validation set
RecallPrec.F1AUCRecall.Prec.F1AUCRecallPrec.F1AUC
Vanilla \nArgo.0.0000.5000.0000.6931.0000.8240.9040.5000.0000.5000.0000.594
Argo. + \nOver.0.7430.7590.7510.8790.9390.9430.9410.8490.6960.6440.6690.875
Argo. + \nSMOTE0.7530.7030.7270.9390.8800.9550.9160.9160.8150.4590.5870.927
Ours0.9420.6810.7900.9520.9050.9870.9440.9440.9890.6610.7730.952
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Dataset compositions and training processes of the baselines are illustrated in Fig. 8 by comparing with the proposed framework. To evaluate the maneuver classification performance, we employed the AUC (Area Under Curve) of the ROC (Receiver Operating Characteristic) curve for each maneuver class and supplementary metrics, Precision (Prec.), Recall, and F1 Score (F1). The threshold for the precision and recall is 0.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "As shown in Tab. 2, the AUC, which is a representative performance metric, is improved compared to the conventional sampling methods in all maneuver classes. We found that when the model is trained with the vanilla Argoverse" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.344, + 0.127 + ], + "angle": 0, + "content": "Jeon et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.147, + 0.493, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.147, + 0.77, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.313, + 0.789, + 0.34 + ], + "angle": 0, + "content": "Fig. 9: Data distribution comparison according to the temporal distance to the end of the intersection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.784, + 0.429 + ], + "angle": 0, + "content": "dataset (only ego-centric), it predicts most samples as ST, thereby achieving a high recall but a low precision, which the model is extremely biased toward outputing ST. The zero recall in LT and RT (which means there are no predictions for LT and RT with a score larger than 0.5.) also proves the biased training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.551 + ], + "angle": 0, + "content": "Although the model achieves better performance than the vanilla dataset case by employing the oversampling strategy or SMOTE, they are still inferior than the proposed heterogeneous dataset result. Therefore, We argue that sampling-based methods which try to uniformize only the number of samples while ignoring the importance of sample diversity, is unable to overcome the imbalanced dataset problem. Conclusively, introducing the higher variety in the training samples by leveraging the proposed intersection-centric dataset is required for classifying the future maneuver of vehicles without bias." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.581, + 0.414, + 0.595 + ], + "angle": 0, + "content": "4.2 Qualitative Result" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.784, + 0.688 + ], + "angle": 0, + "content": "Maneuver Representation Learning. We visualized the learned representations for each maneuver class using [22] to analyze the behavior of the maneuver classification model. By reducing the dimension of the interaction-aware maneuver context, \\( c \\), every samples in the training dataset are mapped to the two dimensional space as shown in Fig. 9." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In Fig. 9, the maneuver contexts for each maneuver class are scattered according to the time gap to the intersection. The points in blue, cyan, and green denote going straight (ST), left turn (LT) and right turn (RT) maneuver, respectively. The blue points show that the ST maneuver is less spread compared to other maneuvers even at the 4 seconds before the intersection, which implies that the ST maneuver is much easier to distinguish than other turning maneuvers even if remarkable maneuver feature is not observed. On the other hand, turning maneuvers with cyan and green points are widely distributed invading the ST area in the left plot in Fig. 9, which means that the turning maneuvers are likely to be misclassified as the ST." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.147, + 0.493, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.147, + 0.77, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.313, + 0.744, + 0.328 + ], + "angle": 0, + "content": "Fig. 10: Data distribution comparison according to the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Furthermore, as shown in Fig. 10, the distribution of data points is prominently different according to the dataset. In Fig. 10, the data points in the red circles indicate the challenging scenarios: all of three maneuvers are blended in this area so the maneuver representation is hard to be categorized. However, in the imbalanced dataset, the distance from the ambiguous area (the red circle) to the center of the ST maneuver cluster is smaller than other maneuvers, whereas the distances toward each maneuver class are similar in the balanced dataset. We argue that this characteristic of maneuver representations induce the model to generate the biased predictions toward the ST when trained with the imbalanced dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.617 + ], + "angle": 0, + "content": "Case Study on Maneuver Classification. In Fig. 11, use case of the proposed maneuver prediction model is illustrated. Using the historical trajectory, blue lines, each maneuver probability, bar graph, is calculated considering road shape and multiple surrounding vehicles, black dotted lines and grey dots, respectively. Since blue dotted line denotes the ground truth future trajectory, predicted maneuver probability can be qualitatively evaluated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.739 + ], + "angle": 0, + "content": "As shown in (a) and (b) in Fig. 11, LT maneuver is well predicted even though the turning maneuver is not shown in the historical trajectory; maneuver probability of the LT class takes the highest likelihood. Furthermore, despite the ground truth label of LT class, ST and/or RT probability is also generated considering the road shape, which enables the planning algorithm to consider low probability cases so that abnormal situations can be handled. Similarly, ST and RT maneuver are well-predicted as shown in 11: (c), (d) for ST class and (e), (f) for RT class." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.763, + 0.36, + 0.779 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In the deep learning based autonomous driving system, the imbalanced data was the critical bottleneck in generalizing the model because of the nature of naturalistic driving dataset: long-tail problem. In the proposed research, for overcoming" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.344, + 0.127 + ], + "angle": 0, + "content": "Jeon et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.147, + 0.761, + 0.159 + ], + "angle": 0, + "content": ": Historical trajectory (input) : Current position : Future trajectory (GT)" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.163, + 0.411, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.281, + 0.331, + 0.292 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.164, + 0.589, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.281, + 0.51, + 0.292 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.165, + 0.768, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.28, + 0.687, + 0.292 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.296, + 0.411, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.413, + 0.331, + 0.424 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.297, + 0.589, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.494, + 0.413, + 0.51, + 0.424 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.297, + 0.768, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.413, + 0.687, + 0.424 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Fig. 11: Prediction results for the representative scenarios. Left turn \\((a, b)\\), going straight \\((c, d)\\) and right turn \\((e, f)\\) maneuver can be predicted as a stochastic manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.639 + ], + "angle": 0, + "content": "the data imbalance and biased prediction model, two dataset from difference modality is employed: in-house drone dataset and Argoverse dataset. The in-house Drone dataset has quantitatively abundant scenarios including a bunch of turning maneuvers, where the Argoverse dataset contains various road shape. By merging the strength of the heterogeneous dataset into a maneuver prediction model by introducing a two-stage training scheme with self-supervised learning, geometrically generalizable prediction model can be trained in balance in terms of the vehicle maneuver. Additionally, the probabilistic approach can provide additional information on the unlikely to happen maneuver so that the planning algorithm can consider abnormal driving situation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.764 + ], + "angle": 0, + "content": "However, the proposed prediction method can not consider the dependency between the future of the surrounding vehicle and the action plan of the ego vehicle. In real world, vehicles are continuously interacting each other and update their own driving strategy in real time. Therefore, for highly intelligent prediction network, the effect of the ego vehicle's future action plan has to be considered in predicting the maneuver of the other vehicles [18]. In this stage, simulation based approaches (e.g. MORAI SIM: Drive [1]) is highly promising to handle both of the data imbalance problem and reactiveness of the actors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by Autonomous Driving Development Center, AVP Division, Hyundai Motor Company and the National Research Foundation of Korea(NRF) funded by the Korea government(MSIT) under Grants 2022R1A2C200494413." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.18, + 0.57, + 0.194 + ], + "angle": 0, + "content": "1. Morai simulator, https://www.morai.ai/drive 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.195, + 0.787, + 0.236 + ], + "angle": 0, + "content": "2. Azadani, M.N., Boukerche, A.: A novel multimodal vehicle path prediction method based on temporal convolutional networks. IEEE Transactions on Intelligent Transportation Systems (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.237, + 0.787, + 0.292 + ], + "angle": 0, + "content": "3. Bock, J., Krajewski, R., Moers, T., Runde, S., Vater, L., Eckstein, L.: The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 1929-1934. IEEE (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.787, + 0.348 + ], + "angle": 0, + "content": "4. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8748-8757 (2019) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.787, + 0.39 + ], + "angle": 0, + "content": "5. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research 16, 321-357 (2002) 2, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.787, + 0.431 + ], + "angle": 0, + "content": "6. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.787, + 0.461 + ], + "angle": 0, + "content": "7. Cieslak, D.A., Chawla, N.V., Striegel, A.: Combating imbalance in network intrusion datasets. In: GrC. pp. 732-737. Citeseer (2006) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.787, + 0.503 + ], + "angle": 0, + "content": "8. Cui, Y., Song, Y., Sun, C., Howard, A., Belongie, S.: Large scale fine-grained categorization and domain-specific transfer learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4109-4118 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.787, + 0.545 + ], + "angle": 0, + "content": "9. Deepa, T., Punithavalli, M.: An e-smote technique for feature selection in high-dimensional imbalanced dataset. In: 2011 3rd International Conference on Electronics Computer Technology. vol. 2, pp. 322-324. IEEE (2011) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.546, + 0.786, + 0.587 + ], + "angle": 0, + "content": "10. Drummond, C., Holte, R.C., et al.: C4. 5, class imbalance, and cost sensitivity: why under-sampling beats over-sampling. In: Workshop on learning from imbalanced datasets II. vol. 11, pp. 1-8. Citeseer (2003) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.588, + 0.786, + 0.629 + ], + "angle": 0, + "content": "11. Feng, X., Cen, Z., Hu, J., Zhang, Y.: Vehicle trajectory prediction using intention-based conditional variational autoencoder. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 3514-3519. IEEE (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.63, + 0.787, + 0.685 + ], + "angle": 0, + "content": "12. Girase, H., Gang, H., Malla, S., Li, J., Kanehara, A., Mangalam, K., Choi, C.: Loki: Long term and key intentions for trajectory prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9803-9812 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.686, + 0.787, + 0.741 + ], + "angle": 0, + "content": "13. Goldhammer, M., Kohler, S., Zernetsch, S., Doll, K., Sick, B., Dietmayer, K.: Intentions of vulnerable road users—detection and forecasting by means of machine learning. IEEE transactions on intelligent transportation systems 21(7), 3035-3045 (2019) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.742, + 0.787, + 0.77 + ], + "angle": 0, + "content": "14. He, H., Garcia, E.A.: Learning from imbalanced data. IEEE Transactions on knowledge and data engineering 21(9), 1263-1284 (2009) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "15. Hu, Y., Li, Y., Huang, H., Lee, J., Yuan, C., Zou, G.: A high-resolution trajectory data driven method for real-time evaluation of traffic safety. Accident Analysis & Prevention 165, 106503 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "16. J. Colyar, J.H.: Us highway 101 dataset. Federal Highway Administration (FHWA), Tech. Rep. FHWA-HRT-07-030 (2007) 6" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.18, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.343, + 0.127 + ], + "angle": 0, + "content": "Jeon et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "17. Jeon, H., Choi, J., Kum, D.: Scale-net: Scalable vehicle trajectory prediction network under random number of interacting vehicles via edge-enhanced graph convolutional neural network. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2095-2102. IEEE (2020) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "18. Jeon, H., Kim, S., Lee, K., Kang, D., Choi, J., Kum, D.: Are reactions to ego vehicles predictable without data?: A semi-supervised approach. IEEE Transactions on Intelligent Transportation Systems 24(6), 6477-6490 (2023) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.787, + 0.285 + ], + "angle": 0, + "content": "19. Kim, S., Jeon, H., Choi, J.W., Kum, D.: Diverse multiple trajectory prediction using a two-stage prediction network trained with lane loss. IEEE Robotics and Automation Letters 8(4), 2038-2045 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.311 + ], + "angle": 0, + "content": "20. Kozerawski, J., Sharan, M., Yu, R.: Taming the long tail of deep probabilistic forecasting. arXiv preprint arXiv:2202.13418 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "21. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: European Conference on Computer Vision. pp. 541-556. Springer (2020) 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.379 + ], + "angle": 0, + "content": "22. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "23. Makansi, O., Çiçek, Ö., Marrakchi, Y., Brox, T.: On exposing the challenging long tail in future prediction of traffic actors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13147-13157 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.446 + ], + "angle": 0, + "content": "24. Marron, J.S., Wand, M.P.: Exact mean integrated squared error. The Annals of Statistics 20(2), 712-736 (1992) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.5 + ], + "angle": 0, + "content": "25. Mirus, F., Stewart, T.C., Conradt, J.: The importance of balanced data sets: Analyzing a vehicle trajectory prediction model based on neural networks and distributed representations. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.528 + ], + "angle": 0, + "content": "26. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.568 + ], + "angle": 0, + "content": "27. Park, D., Ryu, H., Yang, Y., Cho, J., Kim, J., Yoon, K.J.: Leveraging future relationship reasoning for vehicle trajectory prediction. arXiv preprint arXiv:2305.14715 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.787, + 0.609 + ], + "angle": 0, + "content": "28. Peng, J., Bu, X., Sun, M., Zhang, Z., Tan, T., Yan, J.: Large-scale object detection in the wild from imbalanced multi-labels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9709-9718 (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "29. Robicquet, A., Sadeghian, A., Alahi, A., Savarese, S.: Learning social etiquette: Human trajectory prediction in crowded scenes. In: European Conference on Computer Vision (ECCV). vol. 2 (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.69 + ], + "angle": 0, + "content": "30. Shangguan, Q., Fu, T., Wang, J., Fu, L., et al.: A proactive lane-changing risk prediction framework considering driving intention recognition and different lane-changing patterns. Accident Analysis & Prevention 164, 106500 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.787, + 0.731 + ], + "angle": 0, + "content": "31. Shen, L., Lin, Z., Huang, Q.: Relay backpropagation for effective learning of deep convolutional neural networks. In: European conference on computer vision. pp. 467-482. Springer (2016) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "32. Strigel, E., Meissner, D., Seeliger, F., Wilking, B., Dietmayer, K.: The ko-per intersection laserscanner and video dataset. In: 17th International IEEE Conference on Intelligent Transportation Systems (ITSC). pp. 1900-1901. IEEE (2014) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.841 + ], + "angle": 0, + "content": "33. Varadarajan, B., Hefny, A., Srivastava, A., Refaat, K.S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C.P., Anguelov, D., et al.: Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 7814-7821. IEEE (2022) 1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Wang, J., Xu, M., Wang, H., Zhang, J.: Classification of imbalanced data by using the smote algorithm and locally linear embedding. In: 2006 8th international Conference on Signal Processing. vol. 3. IEEE (2006) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "35. Wang, K., Xue, Q., Lu, J.J.: Risky driver recognition with class imbalance data and automated machine learning framework. International journal of environmental research and public health 18(14), 7534 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "36. Westny, T., Frisk, E., Olofsson, B.: Vehicle behavior prediction and generalization using imbalanced learning techniques. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2003-2010. IEEE (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.327 + ], + "angle": 0, + "content": "37. Yang, D., Li, L., Redmill, K., Özgün, Ü.: Top-view trajectories: A pedestrian dataset of vehicle-crowd interaction from controlled experiments and crowded campus. In: 2019 IEEE Intelligent Vehicles Symposium (IV). pp. 899-904. IEEE (2019) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "38. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022) 1" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_origin.pdf b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a522fa19a21aa5f45bfb92b84f5e6ad0076f0c7 --- /dev/null +++ b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/3da36513-0449-41fb-a4dc-3dd8440b6aa5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e20c466932c5e43c7d01d20aa8dd25e6758459b0f9f5bf4213153dd4b0f5b78 +size 14797170 diff --git a/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/full.md b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cfc0e78a82ff5c92a14e03374ce608e834727cb4 --- /dev/null +++ b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/full.md @@ -0,0 +1,307 @@ +# Beyond the Data Imbalance: Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction + +Hyeongseok Jeon $^{1}$ , Sanmin Kim $^{2}$ , Abi Rahman Syamil $^{2}$ , Junsoo Kim $^{3}$ , and Dongsuk Kum $^{2}$ + +$^{1}$ MORAI Inc., Seoul, South Korea + $^{2}$ Korea Advanced Institute of Science and Technology, Daejeon, South Korea + $^{3}$ Hyundai Motor Company, Seoul, South Korea +hsjeon@morai.ai dskum@kaist.ac.kr + +Abstract. Predicting the maneuvers of surrounding vehicles is imperative for the safe navigation of autonomous vehicles. However, naturalistic driving datasets tend to be highly imbalanced, with a bias towards the "going straight" maneuver. Consequently, learning and accurately predicting turning maneuvers pose significant challenges. In this study, we propose a novel two-stage maneuver learning method that can overcome such strong biases by leveraging two heterogeneous datasets in a complementary manner. In the first training phase, we utilize an intersection-centric dataset characterized by balanced distribution of maneuver classes to learn the representations of each maneuver. Subsequently, in the second training phase, we incorporate an ego-centric driving dataset to account for various geometrical road shapes, by transferring the knowledge of geometric diversity to the maneuver prediction model. To facilitate this, we constructed an in-house intersection-centric trajectory dataset with a well-balanced maneuver distribution. By harnessing the power of heterogeneous datasets, our framework significantly improves maneuver prediction performance, particularly for minority maneuver classes such as turning maneuvers. The dataset is available at https://github.com/KAIST-VDCLab/VDC-Trajectory-Dataset. + +Keywords: Maneuver Prediction $\cdot$ Long-tail Problem $\cdot$ Autonomous Driving + +# 1 Introduction + +Predicting the motion of surrounding vehicles is a critical component for ensuring the safety of autonomous driving. In recent years, numerous studies have introduced motion prediction models [19, 21, 27, 33, 38] that primarily focus on generating detailed trajectories. These approaches are trained and evaluated using trajectory-level metrics such as ADE/FDE (Average/Final Distance Error), assessing how closely predicted trajectories align with actual driving trajectories. + +![](images/fcb5d650a1df8109d0f06a43fe28dbedf22964c2f9ef0eea659f72a9407dfe49.jpg) +Fig. 1: Importance of maneuver prediction. In a intersection scenario, accurate prediction of vehicles' turning maneuver is crucial. + +However, from the perspective of motion planning, it becomes clear that accurately predicting the high-level intentions of surrounding vehicles holds greater importance than providing fine-grained future predictions. For example, in intersection scenarios where vehicle interactions are maximized, precisely determining a vehicle's turning maneuver is more crucial than outputting detailed future trajectories, which may end up being a wrong maneuver, as shown in Fig. 1. Hence, prioritizing the accurate prediction of vehicle intentions over detailed trajectory predictions becomes paramount for safe motion planning in autonomous driving, particularly in intricate scenarios like intersections. + +Various datasets have been collected and released for the motion prediction task, gathered by driving sensor-equipped vehicles in real-world environments to ensure scenario diversity. While these datasets provide ample trajectories for training motion prediction models, they exhibit highly biased distributions due to the inherent nature of driving trajectories, particularly concerning maneuvers. Straight roads dominate the driving environment, leading to most trajectory datasets being biased towards straight trajectories. Consequently, maneuver classification models trained with biased datasets tend to be biased toward scenarios involving straight driving while learning less about other maneuvers. + +Several approaches have been proposed in order to address the long-tail problem. Sampling strategies [5, 10, 28, 31] have attempted to handle this issue by either oversampling rare cases or downsampling dominant cases. Alternatively, loss function reformulation strategies [12, 20, 23, 35] have been explored to adjust gradients and enforce balanced supervision during training. However, these approaches have shown limited success, as they merely manipulate the training dataset to achieve balance or regulate the learning process within the existing dataset. Consequently, the fundamental problem stemming from the long-tailed nature of the data has not been effectively addressed. + +To address these challenges, we present a novel two-stage learning framework that harnesses the power of heterogeneous datasets. Our approach aims to + +![](images/b889c40c26a60307a8166ca05a4149ed604c0ac927c586e3a804e8b22867c523.jpg) +Fig. 2: Pros and Cons of the two different datasets: An intersection-centric dataset acquired from the bird's eye view and an ego-centric dataset that collected in a vehicle view. + +![](images/98ff3f616863790c54b32c55ad4bc8682cc912b45b9608457e95bdfcec17b7d1.jpg) + +complement the naturally imbalanced driving dataset by incorporating a cost-efficient and easily acquirable intersection-centric dataset. In our proposed framework, we recognize the importance of capturing various turning maneuvers commonly observed in intersections. Towards that end, we introduce a drone-based intersection-centric dataset, specifically chosen to mitigate the limitations of the ego-centric driving dataset. We utilized the in-house intersection-centric dataset by focusing on a representative intersection area with the aim of learning maneuver representations. To effectively combine these heterogeneous datasets for training maneuver prediction models, we first extract valuable knowledge of each maneuver class from the in-house intersection-centric trajectory dataset. This knowledge, which is challenging to learn from ego-centric driving datasets alone, is then transferred to the downstream network. Here, the model learns the geometric diversity leveraging the ego-centric driving dataset. This approach ensures that our model will initially acquire rich intersection-oriented maneuver knowledge in a balanced manner, while the downstream network then adapts the maneuver representations to diverse road geometries. As a result, our model simultaneously learns maneuver representations from the intersection-centric dataset and features of diverse road geometries from the ego-centric driving dataset. + +To the best of our knowledge, this study is the first attempt to train the maneuver (or trajectory) prediction model with the heterogeneous datasets in complementary manner. The main contributions of our work are outlined as follows: + +- We propose a novel two-stage maneuver prediction framework that effectively leverages heterogeneous datasets to address the long-tail problem in motion prediction. +- We introduce a novel trajectory dataset collected specifically at intersections, ensuring a balanced distribution of maneuvers. + +# 2 Related Work + +Since the data imbalance problem in deep learning approaches appears widely, general solutions for handling the long tail problem is introduced. However, those + +approaches are only focused on formulating the training dataset or loss functions based on the given imbalanced dataset regardless of the data diversity. Therefore, the diversity of the scenarios for training the network still remain insufficient. + +# 2.1 Sampling-based Approaches + +In sampling-based approaches, the training dataset is selectively or redundantly reformulated by choosing the data from the original dataset. During this reformulation process, over- and/or under-sampling methods are employed to prevent overfitting resulting from imbalanced datasets. + +In studies such as [10, 11, 14], balanced training datasets are constructed through an undersampling mechanism, which involves discarding excessive data instances from the major class of scenarios in the given dataset. In these approaches, data instances from the dominant class are underused to balance the distribution of the training dataset. In contrast, [8,28,31] repeatedly sample minorities in datasets to formulate evenly distributed training datasets using various oversampling algorithms. [31] constructs an equally-distributed mini-batch with a class-balanced composition named class-aware sampling. In [28], a soft-balance sampling method is employed for probabilistic oversampling, while [8] adopts a two-stage training strategy that utilizes both the original dataset and the balanced dataset with different learning rates. Furthermore, to account for randomness in the oversampling process, [5] proposes a synthetic data oversampling algorithm named SMOTE (Synthetic Minority Over-sampling TEchnique), with many variants suggested in subsequent studies [5,7,9,34]. However, these sampling-based data balancing approaches have shown unsatisfactory performance in predicting the future maneuver of vehicles. This is because sampling-based methods merely reconstruct the training dataset within the imbalanced raw dataset. However, in autonomous driving, maneuver prediction networks should observe as many diverse samples as possible to create a generalized prediction model. + +# 2.2 Loss Function Reformulation + +Another approach to address imbalanced datasets is through loss function reformulation, as demonstrated in studies such as [12, 20, 23, 35]. These approaches introduce auxiliary losses to address minority cases. In [12] and [35], a simple weighting method is employed, where larger weights are assigned to gradients for minor classes to equalize the backpropagating effect across all classes. In other studies like [20] and [23], various loss functions are introduced to reshape the minority distribution. In [20], the learning of easy and difficult examples is reprioritized by modeling behavior with the Generalized Pareto Distribution (GPD) using the Pareto Loss. Conversely, in [23], the representation of internal maneuver context is reshaped using contrastive loss, ensuring that difficult examples are well-distributed throughout the representation space. While these loss reformulation approaches attempt to address the data imbalance problem during the + +![](images/335c603e8d752667c71c3e4d679f9db015be3a9611c243e79591634c962456b0.jpg) +Fig.3: Overall architecture of the proposed method. Maneuver representations are extracted from the intersection-centric dataset and adapted to the diverse geometries with the ego-centric driving dataset. Using the maneuver context distribution form intersection dataset, each maneuver score is calculated based on the Gaussian KDE. + +training stage, they still share the same limitation as sampling approaches in that they have a marginal impact on dataset diversity. + +# 2.3 Long-tail Problem in Motion Prediction + +Building an unbiased maneuver prediction model is challenging due to the vehicle maneuvers' inherent long-tail characteristics, even if they are trained with the numerous driving datasets already released. In [25], the influence of dataset composition on performance is comprehensively analyzed, revealing that a balanced dataset containing evenly distributed maneuvers, such as lane changes and straight movements, could enhance overall performance. As a result, several maneuver and trajectory prediction studies [2, 36] have employed oversampling approaches to avoid biased training. These frameworks have demonstrated effectiveness in the field of vehicle maneuver prediction [15, 30]. However, sampling approaches alone are insufficient to overcome the fundamental limitations of long-tailed datasets, and the maneuver prediction task is no exception. Therefore, a supplementary dataset containing numerous samples of minor maneuvers, such as turning maneuvers, is necessary to augment the observable trajectory samples in the training dataset. + +# 3 Our Approach + +To establish a comprehensive approach that encompasses both balanced maneuver distribution and geometric diversity, we introduce a two-stage maneuver prediction framework leveraging heterogeneous datasets. Firstly, we construct an intersection-centric trajectory dataset characterized by a well-balanced distribution of maneuvers. This dataset serves as the foundation for obtaining maneuver representations, achieved through training a maneuver backbone using + +Table 1: Quantitative evaluation on the in-house drone dataset with various conventional datasets from the fixed view point. + +
DatasetLocationTrajectory CountsLocation CountsObjectsFPSMethod
NGSIM [16]boulevard33664car10HzRSU
Ko-PER [32]intersection3501pedestrian, bicycle, car25HzRSU
VRU [13]intersection32781pedestrian, bicycle25HzRSU
DUT [37]campus18622pedestrian, vehicles24Hzdrone
SDD [29]campus102408pedestrian, bicycle, car, skateboard, cart, bus25Hzdrone
InD [3]intersection135994pedestrian, bicycle, car, truck, bus25Hzdrone
Oursintersection, lane merge1547910pedestrian, bicycle, car30Hzdrone
+ +![](images/ae4b9f693c83eebb4e642c08ed51f9052d8436e1e63352a484174daf22298135.jpg) +Fig. 4: Sample scenes in our in-house intersection-centric dataset. Car, bicycle and pedestrian are represented as yellow, cyan and green. + +![](images/bf614ba3f78b58b66e08fd0f7d8833c84f78b65f733abec8817e5b94e032ad65.jpg) + +a self-supervised learning approach. Additionally, we derive probability density functions representing each maneuver by applying Kernel Density Function to the training set of the intersection dataset. Then, we bolster the model's scalability by injecting geometric diversity via the utilization of an ego-centric driving dataset. This step enriches the model's comprehension of diverse road geometries, thereby enhancing its predictive accuracy. During the inference stage, maneuver scores are computed by estimating the likelihood of the given trajectory based on the obtained maneuver probability density functions. The overall diagram for our proposed maneuver prediction framework is depicted in Fig. 3. + +# 3.1 Learning Maneuver Representations + +An Intersection-centric In-house dataset. To ensure a balanced distribution of vehicle maneuvers in our dataset, we construct an in-house intersection-centric dataset. This dataset leverages stationary drone videos, which are easily acquired and require minimal labeling costs compared to sensor-based driving datasets. Samples from the in-house dataset, showcasing abundant turning maneuvers with vehicles, bicycles, and pedestrians depicted in yellow, green, and + +![](images/a9f96296c7a8fdf6455975be73926ac6fdfd570a370029eae8dead639220e391.jpg) +Fig. 5: Network architecture of the maneuver backbone (left) and auxiliary module for training the backbone network (right). + +cyan colors respectively, are illustrated in Fig. 4. The dataset comprises 3.5 hours of video footage captured across 10 locations in 4K resolution at 30 frames per second. Additionally, the dataset includes not only agent trajectories but also high-definition (HD) map data. Furthermore, a digital twin environment is constructed to simulate scenarios within the dataset. A comprehensive evaluation of the in-house drone dataset, compared with other bird's eye view (BEV) datasets, is summarized in Tab. 1. Utilizing this dataset, we train the vehicle's maneuver knowledge in a balanced manner. + +Maneuver Backbone. In the first training phase, we extract maneuver features by utilizing the intersection-centric dataset, which is characterized by uniformly distributed maneuvers. To capture maneuver representations from the observed historical trajectory of each vehicle, we employ a self-supervised learning approach, as illustrated in Fig. 5. + +First, the observed historical trajectory is encoded into a state embedding using an encoder: + +$$ +h _ {i} ^ {t} = f _ {e} \left(s _ {i} ^ {t} = \left(x _ {i} ^ {t}, y _ {i} ^ {t}, \theta_ {i} ^ {t}\right) \mid_ {t = - t h \dots , 0}\right), \tag {1} +$$ + +where $f_{e}$ represents the encoder, $h_{i}^{t}$ denotes a state embedding, $s_{i}^{t} = (x_{i}^{t},y_{i}^{t},\theta_{i}^{t})$ denotes the state of the $i^{th}$ vehicle at timestep $t$ , and $t_h$ represents a observation time horizon. We employ 1D CNN and MLP for the encoder. + +The state embeddings are then fed into an autoregressive model (e.g. GRU [6]) to extract temporal trajectory features. Subsequently, an interaction model (e.g. ScaleNet [17]) is employed to account for interaction among vehicles in multi-vehicle scenarios: + +$$ +c = f _ {v} \left(f _ {a} \left(h _ {i} ^ {t} | _ {t = - t _ {h}, \dots , 0}\right) | _ {i = 1, \dots , n}\right), \tag {2} +$$ + +where $c = \{c_1, \dots, c_n\}$ , $f_v$ and $f_a$ represent the extracted maneuver representations, an interaction module, and an autoregressive model, respectively. $n$ denotes the number of vehicles in the scene. + +To learn the internal hidden representation of each maneuver, we employ a self-supervised learning framework. Based on the Contrastive Predictive Coding + +framework [26], the maneuver backbone is trained to maximize mutual information (MI) between the interaction-aware maneuver context of the historical view, $c_{i}$ , and the state embedding in the future $h_i^t|_{t\geq 0}$ . By maximizing the MI between historical trajectory and future position of the vehicle, the network can learn the slow-varying high level maneuver context of the vehicles. + +Based on the definition, MI between $c_{i}$ and $h_i^t|_{t\geq 0}$ , $I(h_i^t,c_i)$ is formulated as follow: + +$$ +I \left(h _ {i} ^ {t}, c _ {i}\right) = \sum_ {i = 1} p \left(h _ {i} ^ {t}, c _ {i}\right) \log \frac {p \left(h _ {i} ^ {t} , c _ {i}\right)}{p \left(h _ {i} ^ {t}\right)}. \tag {3} +$$ + +As the MI is hardly tractable within the batch training process, the auxiliary function named as MI score function is defined as Eq. (4) with trainable auxiliary linear transformation coefficients $w^{t}$ : + +$$ +f _ {k} \left(h _ {i} ^ {t}, c _ {i}\right) \equiv e x p \left(\left(h _ {i} ^ {t}\right) ^ {T} w ^ {t} c _ {i}\right) \propto \frac {p \left(h _ {i} ^ {t} , c _ {i}\right)}{p \left(h _ {i} ^ {t _ {d}}\right)}. \tag {4} +$$ + +Since MI score function is proportional to $I(h_i^t, c_i)$ as formulated in Eq. (4), MI score function is maximized for the same vehicle and minimized for different vehicles using the infoNCE loss [26] function, formulated as: + +$$ +\mathcal {L} _ {m} = - \frac {1}{n} \sum_ {i = 1} ^ {n} \left[ \frac {1}{t _ {f}} \sum_ {t = 1} ^ {t _ {f}} \log \frac {f _ {k} \left(h _ {i} ^ {t} , c _ {i}\right)}{\Sigma_ {j = 1} ^ {n} f _ {k} \left(h _ {j} ^ {t} , c _ {i}\right)} \right], \tag {5} +$$ + +where $n$ denotes the number of data instances in a single batch and $t_f$ is the future time horizon. With this loss function, the MI between the history and future states of the identical vehicle is enlarged, whereas the MI with other vehicles is suppressed. + +# 3.2 Generalizing for Geometric Diversity + +An Ego-centric Driving Dataset. The intersection-centric dataset used for maneuver representation learning has a limitation in that it lacks geometric diversity. To address this limitation and ensure the model's generalizability across various geometric shapes of driving environments, we incorporate an ego-centric driving dataset [4] alongside the intersection-centric dataset. The ego-centric driving dataset is constructed using sensor-equipped vehicles, providing a diverse range of driving environments with various road shapes. By integrating this dataset into our training pipeline, we can effectively train the map adaptation module and enhance the model's ability to generalize to different geometric scenarios. + +Map Adaptation Module. In the proposed maneuver prediction model, we introduce the map adaptation module as a downstream network aimed at generalizing the pre-trained maneuver representations to various road shapes. For map + +![](images/cbf818a36a5dbd51b352dcfc6b08abdb1b92e36fe9917d6be129f5ca5cda31fb.jpg) +Fig. 6: Overall architecture of the map adaptation modules. + +adaptation, we adopt an off-the-shelf map-based prediction model, LaneGCN [21]. LaneGCN utilizes lane graphs and Graph Convolutional Networks (GCN) to encode road geometries. It consists of multiple modules for encoding and fusing geometric and trajectory features, including MapNet and Lane to Lane for map encoding based on GCN, and Actor to Lane and Lane to Actor to fuse map and trajectory features using an attention mechanism. The overall network is trained using cross-entropy loss as defined in Eq. (6), where $y_{i}$ and $p_i$ represent the ground truth and prediction of the vehicle maneuver, respectively. + +$$ +\mathcal {L} _ {g} = - \sum_ {i = 1} ^ {k} y _ {i} \log \left(p _ {i}\right). \tag {6} +$$ + +It is important to note that the maneuver head is employed only during the training process. This is due to the potential bias that may arise in the model outputting predominantly straight maneuvers during the map adaptation process. This bias stems from the maneuver imbalance present in the ego-centric driving dataset, which can lead to biased training. + +Instead, during the inference stage, we utilize probability density functions for each maneuver based on the intersection-centric dataset to estimate the likelihood of the given map-adapted maneuver context. This approach helps mitigate bias and ensures more accurate maneuver predictions during inference. + +# 3.3 Estimating Likelihood of Maneuvers + +Given that the intersection-centric dataset exhibits a well-distributed range of maneuvers, we estimate the likelihood of maneuvers for a given historical trajectory based on the maneuver distribution within this dataset. To achieve this, + +![](images/fc6d6a87118ee18a72acf114536225eb4a3135a2b217353066390952c807e0fa.jpg) +Fig. 7: Distribution of the training dataset for each maneuver class and probability density function. + +![](images/90cc490c362b7667404d674711de456daf9c8ea1ecc482aaa62871cfce709754.jpg) + +![](images/af82451d2a79d7309d4c0974e15461f497e10309e8c3cd7a35e3f4057a8c552d.jpg) + +we first obtain probability density functions for each maneuver using Gaussian Kernel Density Estimation (Gaussian KDE) [24]. We reduce the dimensionality of maneuver features and then define the probability density function, $f(x, H_d)$ , for each maneuver class using the following equations. + +$$ +K (u) = \frac {1}{2 \pi} \exp \left(- \frac {1}{2} u ^ {T} u\right), \tag {7} +$$ + +$$ +K _ {H _ {d}} (u) = \det (H _ {d}) ^ {- 1 / 2} K (H _ {d} ^ {- 1 / 2} u), \tag {8} +$$ + +$$ +f (x, H _ {d}) = \frac {1}{N} \sum_ {i = 1} ^ {N} K _ {H _ {d}} \left(x - x _ {i}\right), \tag {9} +$$ + +where $H_{d}$ is bandwidth matrix parameter for each maneuver $d \in \{ST, RT, LT\}$ and $x_{i=1,\dots,N}$ is the training data points consist of each maneuver class. Here, $ST, RT$ and $LT$ denote going straight, right turn and left turn respectively. The Epanechnikov kernel method is employed whose bandwidth parameter is defined in Eq. (10): + +$$ +H _ {d} = \left(\frac {4 \sigma_ {d} ^ {5}}{3 N}\right) ^ {\frac {1}{5}} \approx 1. 0 6 \sigma_ {d} N ^ {- \frac {1}{5}}, \tag {10} +$$ + +where $\sigma_{d}$ denotes the standard deviation. + +# 4 Result and Case Study + +# 4.1 Quantitative Result + +The performance of the proposed maneuver classification scheme enhanced by employing heterogeneous datasets is quantitatively compared with the conventional various sampling strategies: a vanilla dataset, oversampled dataset and dataset augmented with SMOTE [5]. We categorize the future maneuvers in Left turn (LT), Straight (ST), and Right turn (RT) from the observed historical trajectories of all surrounding vehicles in the scene. + +![](images/e63e70f73769bc3ecc64cc9debfba6cf0a598c85374c8472f0426333163774a0.jpg) + +![](images/116d4e25ea15a630a29f6ca371bb42a82000450b82bfc760b96f7227d2fcb229.jpg) +Fig. 8: Baselines and the proposed method in preparing the training dataset and general process in maneuver prediction. + +![](images/1979c9a83a1a56a78e476fbb9dbfd023e0d93a6d227f36bfc206dc786ba057e4.jpg) + +![](images/a9250907369fae345d3faab02555f81fb1b2c7be57fc74709a83396419575d17.jpg) + +Table 2: Overall maneuver classification performance of the proposed method comparing with various conventional approaches + +
Training DatasetLT validation setST validation setRT validation set
RecallPrec.F1AUCRecall.Prec.F1AUCRecallPrec.F1AUC
Vanilla +Argo.0.0000.5000.0000.6931.0000.8240.9040.5000.0000.5000.0000.594
Argo. + +Over.0.7430.7590.7510.8790.9390.9430.9410.8490.6960.6440.6690.875
Argo. + +SMOTE0.7530.7030.7270.9390.8800.9550.9160.9160.8150.4590.5870.927
Ours0.9420.6810.7900.9520.9050.9870.9440.9440.9890.6610.7730.952
+ +Dataset compositions and training processes of the baselines are illustrated in Fig. 8 by comparing with the proposed framework. To evaluate the maneuver classification performance, we employed the AUC (Area Under Curve) of the ROC (Receiver Operating Characteristic) curve for each maneuver class and supplementary metrics, Precision (Prec.), Recall, and F1 Score (F1). The threshold for the precision and recall is 0.5. + +As shown in Tab. 2, the AUC, which is a representative performance metric, is improved compared to the conventional sampling methods in all maneuver classes. We found that when the model is trained with the vanilla Argoverse + +![](images/48dd99aa2223520b7e8daa01b0e6192876bae021436f6c9df42a5005fdf90d7d.jpg) +Fig. 9: Data distribution comparison according to the temporal distance to the end of the intersection. + +![](images/fccd5a0cbcc9232247e111b83557b3772410f5801f401c95b6e77c325e9f0d7a.jpg) + +dataset (only ego-centric), it predicts most samples as ST, thereby achieving a high recall but a low precision, which the model is extremely biased toward outputing ST. The zero recall in LT and RT (which means there are no predictions for LT and RT with a score larger than 0.5.) also proves the biased training. + +Although the model achieves better performance than the vanilla dataset case by employing the oversampling strategy or SMOTE, they are still inferior than the proposed heterogeneous dataset result. Therefore, We argue that sampling-based methods which try to uniformize only the number of samples while ignoring the importance of sample diversity, is unable to overcome the imbalanced dataset problem. Conclusively, introducing the higher variety in the training samples by leveraging the proposed intersection-centric dataset is required for classifying the future maneuver of vehicles without bias. + +# 4.2 Qualitative Result + +Maneuver Representation Learning. We visualized the learned representations for each maneuver class using [22] to analyze the behavior of the maneuver classification model. By reducing the dimension of the interaction-aware maneuver context, $c$ , every samples in the training dataset are mapped to the two dimensional space as shown in Fig. 9. + +In Fig. 9, the maneuver contexts for each maneuver class are scattered according to the time gap to the intersection. The points in blue, cyan, and green denote going straight (ST), left turn (LT) and right turn (RT) maneuver, respectively. The blue points show that the ST maneuver is less spread compared to other maneuvers even at the 4 seconds before the intersection, which implies that the ST maneuver is much easier to distinguish than other turning maneuvers even if remarkable maneuver feature is not observed. On the other hand, turning maneuvers with cyan and green points are widely distributed invading the ST area in the left plot in Fig. 9, which means that the turning maneuvers are likely to be misclassified as the ST. + +![](images/4fe9eee2a288241b3582f58eccc0a9bc0ddf915c77f98c6f1a8b028db89e1dbc.jpg) +Fig. 10: Data distribution comparison according to the training dataset. + +![](images/e6cd5575eae06949054bb168c049a3babca287a246e336dc6e879ae1da1348c7.jpg) + +Furthermore, as shown in Fig. 10, the distribution of data points is prominently different according to the dataset. In Fig. 10, the data points in the red circles indicate the challenging scenarios: all of three maneuvers are blended in this area so the maneuver representation is hard to be categorized. However, in the imbalanced dataset, the distance from the ambiguous area (the red circle) to the center of the ST maneuver cluster is smaller than other maneuvers, whereas the distances toward each maneuver class are similar in the balanced dataset. We argue that this characteristic of maneuver representations induce the model to generate the biased predictions toward the ST when trained with the imbalanced dataset. + +Case Study on Maneuver Classification. In Fig. 11, use case of the proposed maneuver prediction model is illustrated. Using the historical trajectory, blue lines, each maneuver probability, bar graph, is calculated considering road shape and multiple surrounding vehicles, black dotted lines and grey dots, respectively. Since blue dotted line denotes the ground truth future trajectory, predicted maneuver probability can be qualitatively evaluated. + +As shown in (a) and (b) in Fig. 11, LT maneuver is well predicted even though the turning maneuver is not shown in the historical trajectory; maneuver probability of the LT class takes the highest likelihood. Furthermore, despite the ground truth label of LT class, ST and/or RT probability is also generated considering the road shape, which enables the planning algorithm to consider low probability cases so that abnormal situations can be handled. Similarly, ST and RT maneuver are well-predicted as shown in 11: (c), (d) for ST class and (e), (f) for RT class. + +# 5 Conclusion + +In the deep learning based autonomous driving system, the imbalanced data was the critical bottleneck in generalizing the model because of the nature of naturalistic driving dataset: long-tail problem. In the proposed research, for overcoming + +![](images/f91f8630a54df143e0b847570433331cb251a5008b63e764e7ff27055b59226e.jpg) +(a) + +![](images/a0f48d55c7d400b28ee5f405d06faeae9ad405ca647503eaa8d4c64afd5d422e.jpg) +: Historical trajectory (input) : Current position : Future trajectory (GT) +(b) + +![](images/82f819ddd300d2df4c2f15ba92acde8133001b5b658ba9c0d6cbd41bc8d6a2fa.jpg) +(c) + +![](images/41b651fe1d23f6b197391a5ef9bc436d02bf121de608cb6527b204b3be95162c.jpg) +(d) + +![](images/afc04b818a66a6a01ee282596b54682dc237409a221a9e1cea5595bb67339c8d.jpg) +(e) + +![](images/6e86913b894c11f150af32424a9fc5819fbbed9358408614eae4b392ada42cb8.jpg) +(f) +Fig. 11: Prediction results for the representative scenarios. Left turn $(a, b)$ , going straight $(c, d)$ and right turn $(e, f)$ maneuver can be predicted as a stochastic manner. + +the data imbalance and biased prediction model, two dataset from difference modality is employed: in-house drone dataset and Argoverse dataset. The in-house Drone dataset has quantitatively abundant scenarios including a bunch of turning maneuvers, where the Argoverse dataset contains various road shape. By merging the strength of the heterogeneous dataset into a maneuver prediction model by introducing a two-stage training scheme with self-supervised learning, geometrically generalizable prediction model can be trained in balance in terms of the vehicle maneuver. Additionally, the probabilistic approach can provide additional information on the unlikely to happen maneuver so that the planning algorithm can consider abnormal driving situation. + +However, the proposed prediction method can not consider the dependency between the future of the surrounding vehicle and the action plan of the ego vehicle. In real world, vehicles are continuously interacting each other and update their own driving strategy in real time. Therefore, for highly intelligent prediction network, the effect of the ego vehicle's future action plan has to be considered in predicting the maneuver of the other vehicles [18]. In this stage, simulation based approaches (e.g. MORAI SIM: Drive [1]) is highly promising to handle both of the data imbalance problem and reactiveness of the actors. + +Acknowledgements. This work was supported by Autonomous Driving Development Center, AVP Division, Hyundai Motor Company and the National Research Foundation of Korea(NRF) funded by the Korea government(MSIT) under Grants 2022R1A2C200494413. + +# References + +1. Morai simulator, https://www.morai.ai/drive 14 +2. Azadani, M.N., Boukerche, A.: A novel multimodal vehicle path prediction method based on temporal convolutional networks. IEEE Transactions on Intelligent Transportation Systems (2022) 5 +3. Bock, J., Krajewski, R., Moers, T., Runde, S., Vater, L., Eckstein, L.: The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 1929-1934. IEEE (2020) 6 +4. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8748-8757 (2019) 8 +5. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research 16, 321-357 (2002) 2, 4, 10 +6. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014) 7 +7. Cieslak, D.A., Chawla, N.V., Striegel, A.: Combating imbalance in network intrusion datasets. In: GrC. pp. 732-737. Citeseer (2006) 4 +8. Cui, Y., Song, Y., Sun, C., Howard, A., Belongie, S.: Large scale fine-grained categorization and domain-specific transfer learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4109-4118 (2018) 4 +9. Deepa, T., Punithavalli, M.: An e-smote technique for feature selection in high-dimensional imbalanced dataset. In: 2011 3rd International Conference on Electronics Computer Technology. vol. 2, pp. 322-324. IEEE (2011) 4 +10. Drummond, C., Holte, R.C., et al.: C4. 5, class imbalance, and cost sensitivity: why under-sampling beats over-sampling. In: Workshop on learning from imbalanced datasets II. vol. 11, pp. 1-8. Citeseer (2003) 2, 4 +11. Feng, X., Cen, Z., Hu, J., Zhang, Y.: Vehicle trajectory prediction using intention-based conditional variational autoencoder. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 3514-3519. IEEE (2019) 4 +12. Girase, H., Gang, H., Malla, S., Li, J., Kanehara, A., Mangalam, K., Choi, C.: Loki: Long term and key intentions for trajectory prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9803-9812 (2021) 2, 4 +13. Goldhammer, M., Kohler, S., Zernetsch, S., Doll, K., Sick, B., Dietmayer, K.: Intentions of vulnerable road users—detection and forecasting by means of machine learning. IEEE transactions on intelligent transportation systems 21(7), 3035-3045 (2019) 6 +14. He, H., Garcia, E.A.: Learning from imbalanced data. IEEE Transactions on knowledge and data engineering 21(9), 1263-1284 (2009) 4 +15. Hu, Y., Li, Y., Huang, H., Lee, J., Yuan, C., Zou, G.: A high-resolution trajectory data driven method for real-time evaluation of traffic safety. Accident Analysis & Prevention 165, 106503 (2022) 5 +16. J. Colyar, J.H.: Us highway 101 dataset. Federal Highway Administration (FHWA), Tech. Rep. FHWA-HRT-07-030 (2007) 6 + +17. Jeon, H., Choi, J., Kum, D.: Scale-net: Scalable vehicle trajectory prediction network under random number of interacting vehicles via edge-enhanced graph convolutional neural network. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2095-2102. IEEE (2020) 7 +18. Jeon, H., Kim, S., Lee, K., Kang, D., Choi, J., Kum, D.: Are reactions to ego vehicles predictable without data?: A semi-supervised approach. IEEE Transactions on Intelligent Transportation Systems 24(6), 6477-6490 (2023) 14 +19. Kim, S., Jeon, H., Choi, J.W., Kum, D.: Diverse multiple trajectory prediction using a two-stage prediction network trained with lane loss. IEEE Robotics and Automation Letters 8(4), 2038-2045 (2022) 1 +20. Kozerawski, J., Sharan, M., Yu, R.: Taming the long tail of deep probabilistic forecasting. arXiv preprint arXiv:2202.13418 (2022) 2, 4 +21. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: European Conference on Computer Vision. pp. 541-556. Springer (2020) 1, 9 +22. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 12 +23. Makansi, O., Çiçek, Ö., Marrakchi, Y., Brox, T.: On exposing the challenging long tail in future prediction of traffic actors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13147-13157 (2021) 2, 4 +24. Marron, J.S., Wand, M.P.: Exact mean integrated squared error. The Annals of Statistics 20(2), 712-736 (1992) 10 +25. Mirus, F., Stewart, T.C., Conradt, J.: The importance of balanced data sets: Analyzing a vehicle trajectory prediction model based on neural networks and distributed representations. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020) 5 +26. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 8 +27. Park, D., Ryu, H., Yang, Y., Cho, J., Kim, J., Yoon, K.J.: Leveraging future relationship reasoning for vehicle trajectory prediction. arXiv preprint arXiv:2305.14715 (2023) 1 +28. Peng, J., Bu, X., Sun, M., Zhang, Z., Tan, T., Yan, J.: Large-scale object detection in the wild from imbalanced multi-labels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9709-9718 (2020) 2, 4 +29. Robicquet, A., Sadeghian, A., Alahi, A., Savarese, S.: Learning social etiquette: Human trajectory prediction in crowded scenes. In: European Conference on Computer Vision (ECCV). vol. 2 (2020) 6 +30. Shangguan, Q., Fu, T., Wang, J., Fu, L., et al.: A proactive lane-changing risk prediction framework considering driving intention recognition and different lane-changing patterns. Accident Analysis & Prevention 164, 106500 (2022) 5 +31. Shen, L., Lin, Z., Huang, Q.: Relay backpropagation for effective learning of deep convolutional neural networks. In: European conference on computer vision. pp. 467-482. Springer (2016) 2, 4 +32. Strigel, E., Meissner, D., Seeliger, F., Wilking, B., Dietmayer, K.: The ko-per intersection laserscanner and video dataset. In: 17th International IEEE Conference on Intelligent Transportation Systems (ITSC). pp. 1900-1901. IEEE (2014) 6 +33. Varadarajan, B., Hefny, A., Srivastava, A., Refaat, K.S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C.P., Anguelov, D., et al.: Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 7814-7821. IEEE (2022) 1 + +34. Wang, J., Xu, M., Wang, H., Zhang, J.: Classification of imbalanced data by using the smote algorithm and locally linear embedding. In: 2006 8th international Conference on Signal Processing. vol. 3. IEEE (2006) 4 +35. Wang, K., Xue, Q., Lu, J.J.: Risky driver recognition with class imbalance data and automated machine learning framework. International journal of environmental research and public health 18(14), 7534 (2021) 2, 4 +36. Westny, T., Frisk, E., Olofsson, B.: Vehicle behavior prediction and generalization using imbalanced learning techniques. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2003-2010. IEEE (2021) 5 +37. Yang, D., Li, L., Redmill, K., Özgün, Ü.: Top-view trajectories: A pedestrian dataset of vehicle-crowd interaction from controlled experiments and crowded campus. In: 2019 IEEE Intelligent Vehicles Symposium (IV). pp. 899-904. IEEE (2019) 6 +38. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022) 1 \ No newline at end of file diff --git a/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/images.zip b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2cd049733fed0e72ab6b52ca2c43485d6d01bc66 --- /dev/null +++ b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2920f88bc6daa119074d5e7dd6b8134dbdd4b9451247573fbf55941daa97bfea +size 628661 diff --git a/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/layout.json b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..837476203c68b996f00e19b42cc276b717e42af4 --- /dev/null +++ b/2024/Beyond the Data Imbalance_ Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction/layout.json @@ -0,0 +1,8004 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 470, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 470, + 164 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 470, + 164 + ], + "type": "text", + "content": "Beyond the Data Imbalance: Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "spans": [ + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "text", + "content": "Hyeongseok Jeon" + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "text", + "content": ", Sanmin Kim" + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "text", + "content": ", Abi Rahman Syamil" + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "text", + "content": ", Junsoo Kim" + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "text", + "content": ", and Dongsuk Kum" + }, + { + "bbox": [ + 140, + 185, + 474, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 219, + 465, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 219, + 465, + 266 + ], + "spans": [ + { + "bbox": [ + 149, + 219, + 465, + 266 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 219, + 465, + 266 + ], + "type": "text", + "content": " MORAI Inc., Seoul, South Korea \n" + }, + { + "bbox": [ + 149, + 219, + 465, + 266 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 219, + 465, + 266 + ], + "type": "text", + "content": " Korea Advanced Institute of Science and Technology, Daejeon, South Korea \n" + }, + { + "bbox": [ + 149, + 219, + 465, + 266 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 219, + 465, + 266 + ], + "type": "text", + "content": " Hyundai Motor Company, Seoul, South Korea \nhsjeon@morai.ai dskum@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 291, + 455, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 291, + 455, + 501 + ], + "spans": [ + { + "bbox": [ + 160, + 291, + 455, + 501 + ], + "type": "text", + "content": "Abstract. Predicting the maneuvers of surrounding vehicles is imperative for the safe navigation of autonomous vehicles. However, naturalistic driving datasets tend to be highly imbalanced, with a bias towards the \"going straight\" maneuver. Consequently, learning and accurately predicting turning maneuvers pose significant challenges. In this study, we propose a novel two-stage maneuver learning method that can overcome such strong biases by leveraging two heterogeneous datasets in a complementary manner. In the first training phase, we utilize an intersection-centric dataset characterized by balanced distribution of maneuver classes to learn the representations of each maneuver. Subsequently, in the second training phase, we incorporate an ego-centric driving dataset to account for various geometrical road shapes, by transferring the knowledge of geometric diversity to the maneuver prediction model. To facilitate this, we constructed an in-house intersection-centric trajectory dataset with a well-balanced maneuver distribution. By harnessing the power of heterogeneous datasets, our framework significantly improves maneuver prediction performance, particularly for minority maneuver classes such as turning maneuvers. The dataset is available at https://github.com/KAIST-VDCLab/VDC-Trajectory-Dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 510, + 455, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 510, + 455, + 533 + ], + "spans": [ + { + "bbox": [ + 159, + 510, + 455, + 533 + ], + "type": "text", + "content": "Keywords: Maneuver Prediction " + }, + { + "bbox": [ + 159, + 510, + 455, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 510, + 455, + 533 + ], + "type": "text", + "content": " Long-tail Problem " + }, + { + "bbox": [ + 159, + 510, + 455, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 510, + 455, + 533 + ], + "type": "text", + "content": " Autonomous Driving" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 569, + 231, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 231, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 231, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Predicting the motion of surrounding vehicles is a critical component for ensuring the safety of autonomous driving. In recent years, numerous studies have introduced motion prediction models [19, 21, 27, 33, 38] that primarily focus on generating detailed trajectories. These approaches are trained and evaluated using trajectory-level metrics such as ADE/FDE (Average/Final Distance Error), assessing how closely predicted trajectories align with actual driving trajectories." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 116, + 462, + 268 + ], + "blocks": [ + { + "bbox": [ + 152, + 116, + 462, + 268 + ], + "lines": [ + { + "bbox": [ + 152, + 116, + 462, + 268 + ], + "spans": [ + { + "bbox": [ + 152, + 116, + 462, + 268 + ], + "type": "image", + "image_path": "fcb5d650a1df8109d0f06a43fe28dbedf22964c2f9ef0eea659f72a9407dfe49.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "text", + "content": "Fig. 1: Importance of maneuver prediction. In a intersection scenario, accurate prediction of vehicles' turning maneuver is crucial." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 422 + ], + "type": "text", + "content": "However, from the perspective of motion planning, it becomes clear that accurately predicting the high-level intentions of surrounding vehicles holds greater importance than providing fine-grained future predictions. For example, in intersection scenarios where vehicle interactions are maximized, precisely determining a vehicle's turning maneuver is more crucial than outputting detailed future trajectories, which may end up being a wrong maneuver, as shown in Fig. 1. Hence, prioritizing the accurate prediction of vehicle intentions over detailed trajectory predictions becomes paramount for safe motion planning in autonomous driving, particularly in intricate scenarios like intersections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 425, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 481, + 532 + ], + "type": "text", + "content": "Various datasets have been collected and released for the motion prediction task, gathered by driving sensor-equipped vehicles in real-world environments to ensure scenario diversity. While these datasets provide ample trajectories for training motion prediction models, they exhibit highly biased distributions due to the inherent nature of driving trajectories, particularly concerning maneuvers. Straight roads dominate the driving environment, leading to most trajectory datasets being biased towards straight trajectories. Consequently, maneuver classification models trained with biased datasets tend to be biased toward scenarios involving straight driving while learning less about other maneuvers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "content": "Several approaches have been proposed in order to address the long-tail problem. Sampling strategies [5, 10, 28, 31] have attempted to handle this issue by either oversampling rare cases or downsampling dominant cases. Alternatively, loss function reformulation strategies [12, 20, 23, 35] have been explored to adjust gradients and enforce balanced supervision during training. However, these approaches have shown limited success, as they merely manipulate the training dataset to achieve balance or regulate the learning process within the existing dataset. Consequently, the fundamental problem stemming from the long-tailed nature of the data has not been effectively addressed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "To address these challenges, we present a novel two-stage learning framework that harnesses the power of heterogeneous datasets. Our approach aims to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "Jeon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 306, + 216 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 306, + 216 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 306, + 216 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 306, + 216 + ], + "type": "image", + "image_path": "b889c40c26a60307a8166ca05a4149ed604c0ac927c586e3a804e8b22867c523.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 257 + ], + "type": "text", + "content": "Fig. 2: Pros and Cons of the two different datasets: An intersection-centric dataset acquired from the bird's eye view and an ego-centric dataset that collected in a vehicle view." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 118, + 478, + 216 + ], + "blocks": [ + { + "bbox": [ + 307, + 118, + 478, + 216 + ], + "lines": [ + { + "bbox": [ + 307, + 118, + 478, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 478, + 216 + ], + "type": "image", + "image_path": "98ff3f616863790c54b32c55ad4bc8682cc912b45b9608457e95bdfcec17b7d1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 486 + ], + "type": "text", + "content": "complement the naturally imbalanced driving dataset by incorporating a cost-efficient and easily acquirable intersection-centric dataset. In our proposed framework, we recognize the importance of capturing various turning maneuvers commonly observed in intersections. Towards that end, we introduce a drone-based intersection-centric dataset, specifically chosen to mitigate the limitations of the ego-centric driving dataset. We utilized the in-house intersection-centric dataset by focusing on a representative intersection area with the aim of learning maneuver representations. To effectively combine these heterogeneous datasets for training maneuver prediction models, we first extract valuable knowledge of each maneuver class from the in-house intersection-centric trajectory dataset. This knowledge, which is challenging to learn from ego-centric driving datasets alone, is then transferred to the downstream network. Here, the model learns the geometric diversity leveraging the ego-centric driving dataset. This approach ensures that our model will initially acquire rich intersection-oriented maneuver knowledge in a balanced manner, while the downstream network then adapts the maneuver representations to diverse road geometries. As a result, our model simultaneously learns maneuver representations from the intersection-centric dataset and features of diverse road geometries from the ego-centric driving dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "text", + "content": "To the best of our knowledge, this study is the first attempt to train the maneuver (or trajectory) prediction model with the heterogeneous datasets in complementary manner. The main contributions of our work are outlined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 538, + 480, + 597 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 137, + 538, + 480, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 538, + 480, + 573 + ], + "spans": [ + { + "bbox": [ + 137, + 538, + 480, + 573 + ], + "type": "text", + "content": "- We propose a novel two-stage maneuver prediction framework that effectively leverages heterogeneous datasets to address the long-tail problem in motion prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 574, + 480, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 574, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 137, + 574, + 480, + 597 + ], + "type": "text", + "content": "- We introduce a novel trajectory dataset collected specifically at intersections, ensuring a balanced distribution of maneuvers." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 616, + 237, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 237, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 237, + 628 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Since the data imbalance problem in deep learning approaches appears widely, general solutions for handling the long tail problem is introduced. However, those" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "approaches are only focused on formulating the training dataset or loss functions based on the given imbalanced dataset regardless of the data diversity. Therefore, the diversity of the scenarios for training the network still remain insufficient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 302, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 302, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 302, + 185 + ], + "type": "text", + "content": "2.1 Sampling-based Approaches" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 479, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 479, + 240 + ], + "type": "text", + "content": "In sampling-based approaches, the training dataset is selectively or redundantly reformulated by choosing the data from the original dataset. During this reformulation process, over- and/or under-sampling methods are employed to prevent overfitting resulting from imbalanced datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 241, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 479 + ], + "type": "text", + "content": "In studies such as [10, 11, 14], balanced training datasets are constructed through an undersampling mechanism, which involves discarding excessive data instances from the major class of scenarios in the given dataset. In these approaches, data instances from the dominant class are underused to balance the distribution of the training dataset. In contrast, [8,28,31] repeatedly sample minorities in datasets to formulate evenly distributed training datasets using various oversampling algorithms. [31] constructs an equally-distributed mini-batch with a class-balanced composition named class-aware sampling. In [28], a soft-balance sampling method is employed for probabilistic oversampling, while [8] adopts a two-stage training strategy that utilizes both the original dataset and the balanced dataset with different learning rates. Furthermore, to account for randomness in the oversampling process, [5] proposes a synthetic data oversampling algorithm named SMOTE (Synthetic Minority Over-sampling TEchnique), with many variants suggested in subsequent studies [5,7,9,34]. However, these sampling-based data balancing approaches have shown unsatisfactory performance in predicting the future maneuver of vehicles. This is because sampling-based methods merely reconstruct the training dataset within the imbalanced raw dataset. However, in autonomous driving, maneuver prediction networks should observe as many diverse samples as possible to create a generalized prediction model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 500, + 307, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 307, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 307, + 511 + ], + "type": "text", + "content": "2.2 Loss Function Reformulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Another approach to address imbalanced datasets is through loss function reformulation, as demonstrated in studies such as [12, 20, 23, 35]. These approaches introduce auxiliary losses to address minority cases. In [12] and [35], a simple weighting method is employed, where larger weights are assigned to gradients for minor classes to equalize the backpropagating effect across all classes. In other studies like [20] and [23], various loss functions are introduced to reshape the minority distribution. In [20], the learning of easy and difficult examples is reprioritized by modeling behavior with the Generalized Pareto Distribution (GPD) using the Pareto Loss. Conversely, in [23], the representation of internal maneuver context is reshaped using contrastive loss, ensuring that difficult examples are well-distributed throughout the representation space. While these loss reformulation approaches attempt to address the data imbalance problem during the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "Jeon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 120, + 479, + 258 + ], + "blocks": [ + { + "bbox": [ + 136, + 120, + 479, + 258 + ], + "lines": [ + { + "bbox": [ + 136, + 120, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 136, + 120, + 479, + 258 + ], + "type": "image", + "image_path": "335c603e8d752667c71c3e4d679f9db015be3a9611c243e79591634c962456b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 317 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 317 + ], + "type": "text", + "content": "Fig.3: Overall architecture of the proposed method. Maneuver representations are extracted from the intersection-centric dataset and adapted to the diverse geometries with the ego-centric driving dataset. Using the maneuver context distribution form intersection dataset, each maneuver score is calculated based on the Gaussian KDE." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "type": "text", + "content": "training stage, they still share the same limitation as sampling approaches in that they have a marginal impact on dataset diversity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 372, + 362, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 362, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 362, + 384 + ], + "type": "text", + "content": "2.3 Long-tail Problem in Motion Prediction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 556 + ], + "type": "text", + "content": "Building an unbiased maneuver prediction model is challenging due to the vehicle maneuvers' inherent long-tail characteristics, even if they are trained with the numerous driving datasets already released. In [25], the influence of dataset composition on performance is comprehensively analyzed, revealing that a balanced dataset containing evenly distributed maneuvers, such as lane changes and straight movements, could enhance overall performance. As a result, several maneuver and trajectory prediction studies [2, 36] have employed oversampling approaches to avoid biased training. These frameworks have demonstrated effectiveness in the field of vehicle maneuver prediction [15, 30]. However, sampling approaches alone are insufficient to overcome the fundamental limitations of long-tailed datasets, and the maneuver prediction task is no exception. Therefore, a supplementary dataset containing numerous samples of minor maneuvers, such as turning maneuvers, is necessary to augment the observable trajectory samples in the training dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 571, + 241, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 241, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 241, + 586 + ], + "type": "text", + "content": "3 Our Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To establish a comprehensive approach that encompasses both balanced maneuver distribution and geometric diversity, we introduce a two-stage maneuver prediction framework leveraging heterogeneous datasets. Firstly, we construct an intersection-centric trajectory dataset characterized by a well-balanced distribution of maneuvers. This dataset serves as the foundation for obtaining maneuver representations, achieved through training a maneuver backbone using" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 146, + 479, + 302 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation on the in-house drone dataset with various conventional datasets from the fixed view point." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 146, + 479, + 302 + ], + "lines": [ + { + "bbox": [ + 140, + 146, + 479, + 302 + ], + "spans": [ + { + "bbox": [ + 140, + 146, + 479, + 302 + ], + "type": "table", + "html": "
DatasetLocationTrajectory CountsLocation CountsObjectsFPSMethod
NGSIM [16]boulevard33664car10HzRSU
Ko-PER [32]intersection3501pedestrian, bicycle, car25HzRSU
VRU [13]intersection32781pedestrian, bicycle25HzRSU
DUT [37]campus18622pedestrian, vehicles24Hzdrone
SDD [29]campus102408pedestrian, bicycle, car, skateboard, cart, bus25Hzdrone
InD [3]intersection135994pedestrian, bicycle, car, truck, bus25Hzdrone
Oursintersection, lane merge1547910pedestrian, bicycle, car30Hzdrone
", + "image_path": "8f6c867d6ba9b8c0f27c2a888cad53cbccc2926d52897e606ad1bbb27ddd3d59.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 312, + 304, + 403 + ], + "blocks": [ + { + "bbox": [ + 143, + 312, + 304, + 403 + ], + "lines": [ + { + "bbox": [ + 143, + 312, + 304, + 403 + ], + "spans": [ + { + "bbox": [ + 143, + 312, + 304, + 403 + ], + "type": "image", + "image_path": "ae4b9f693c83eebb4e642c08ed51f9052d8436e1e63352a484174daf22298135.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 415, + 480, + 439 + ], + "lines": [ + { + "bbox": [ + 131, + 415, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 480, + 439 + ], + "type": "text", + "content": "Fig. 4: Sample scenes in our in-house intersection-centric dataset. Car, bicycle and pedestrian are represented as yellow, cyan and green." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 313, + 470, + 404 + ], + "blocks": [ + { + "bbox": [ + 308, + 313, + 470, + 404 + ], + "lines": [ + { + "bbox": [ + 308, + 313, + 470, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 313, + 470, + 404 + ], + "type": "image", + "image_path": "bf614ba3f78b58b66e08fd0f7d8833c84f78b65f733abec8817e5b94e032ad65.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 559 + ], + "type": "text", + "content": "a self-supervised learning approach. Additionally, we derive probability density functions representing each maneuver by applying Kernel Density Function to the training set of the intersection dataset. Then, we bolster the model's scalability by injecting geometric diversity via the utilization of an ego-centric driving dataset. This step enriches the model's comprehension of diverse road geometries, thereby enhancing its predictive accuracy. During the inference stage, maneuver scores are computed by estimating the likelihood of the given trajectory based on the obtained maneuver probability density functions. The overall diagram for our proposed maneuver prediction framework is depicted in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 575, + 345, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 345, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 345, + 588 + ], + "type": "text", + "content": "3.1 Learning Maneuver Representations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "An Intersection-centric In-house dataset. To ensure a balanced distribution of vehicle maneuvers in our dataset, we construct an in-house intersection-centric dataset. This dataset leverages stationary drone videos, which are easily acquired and require minimal labeling costs compared to sensor-based driving datasets. Samples from the in-house dataset, showcasing abundant turning maneuvers with vehicles, bicycles, and pedestrians depicted in yellow, green, and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "Jeon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 118, + 479, + 222 + ], + "blocks": [ + { + "bbox": [ + 137, + 118, + 479, + 222 + ], + "lines": [ + { + "bbox": [ + 137, + 118, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 479, + 222 + ], + "type": "image", + "image_path": "a9f96296c7a8fdf6455975be73926ac6fdfd570a370029eae8dead639220e391.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 232, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 482, + 257 + ], + "type": "text", + "content": "Fig. 5: Network architecture of the maneuver backbone (left) and auxiliary module for training the backbone network (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "type": "text", + "content": "cyan colors respectively, are illustrated in Fig. 4. The dataset comprises 3.5 hours of video footage captured across 10 locations in 4K resolution at 30 frames per second. Additionally, the dataset includes not only agent trajectories but also high-definition (HD) map data. Furthermore, a digital twin environment is constructed to simulate scenarios within the dataset. A comprehensive evaluation of the in-house drone dataset, compared with other bird's eye view (BEV) datasets, is summarized in Tab. 1. Utilizing this dataset, we train the vehicle's maneuver knowledge in a balanced manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 440 + ], + "type": "text", + "content": "Maneuver Backbone. In the first training phase, we extract maneuver features by utilizing the intersection-centric dataset, which is characterized by uniformly distributed maneuvers. To capture maneuver representations from the observed historical trajectory of each vehicle, we employ a self-supervised learning approach, as illustrated in Fig. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 440, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 464 + ], + "type": "text", + "content": "First, the observed historical trajectory is encoded into a state embedding using an encoder:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 475, + 481, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 475, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 230, + 475, + 481, + 489 + ], + "type": "interline_equation", + "content": "h _ {i} ^ {t} = f _ {e} \\left(s _ {i} ^ {t} = \\left(x _ {i} ^ {t}, y _ {i} ^ {t}, \\theta_ {i} ^ {t}\\right) \\mid_ {t = - t h \\dots , 0}\\right), \\tag {1}", + "image_path": "3d96bfd2aa327849a02f4805dcf6dedca9498f5ad3afa3b811da54df23be9d5d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "text", + "content": " represents the encoder, " + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "inline_equation", + "content": "h_{i}^{t}" + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "text", + "content": " denotes a state embedding, " + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "inline_equation", + "content": "s_{i}^{t} = (x_{i}^{t},y_{i}^{t},\\theta_{i}^{t})" + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "text", + "content": " denotes the state of the " + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "text", + "content": " vehicle at timestep " + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "inline_equation", + "content": "t_h" + }, + { + "bbox": [ + 130, + 492, + 481, + 528 + ], + "type": "text", + "content": " represents a observation time horizon. We employ 1D CNN and MLP for the encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 529, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 482, + 576 + ], + "type": "text", + "content": "The state embeddings are then fed into an autoregressive model (e.g. GRU [6]) to extract temporal trajectory features. Subsequently, an interaction model (e.g. ScaleNet [17]) is employed to account for interaction among vehicles in multi-vehicle scenarios:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 230, + 587, + 481, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 587, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 230, + 587, + 481, + 602 + ], + "type": "interline_equation", + "content": "c = f _ {v} \\left(f _ {a} \\left(h _ {i} ^ {t} | _ {t = - t _ {h}, \\dots , 0}\\right) | _ {i = 1, \\dots , n}\\right), \\tag {2}", + "image_path": "0623663d527ead5de9dfa65f2fea163fa92404c2a64fd627877d7ab42b793059.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "inline_equation", + "content": "c = \\{c_1, \\dots, c_n\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "inline_equation", + "content": "f_v" + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "inline_equation", + "content": "f_a" + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "text", + "content": " represent the extracted maneuver representations, an interaction module, and an autoregressive model, respectively. " + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "text", + "content": " denotes the number of vehicles in the scene." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "To learn the internal hidden representation of each maneuver, we employ a self-supervised learning framework. Based on the Contrastive Predictive Coding" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "text", + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "framework [26], the maneuver backbone is trained to maximize mutual information (MI) between the interaction-aware maneuver context of the historical view, " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", and the state embedding in the future " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "h_i^t|_{t\\geq 0}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ". By maximizing the MI between historical trajectory and future position of the vehicle, the network can learn the slow-varying high level maneuver context of the vehicles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "text", + "content": "Based on the definition, MI between " + }, + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "inline_equation", + "content": "h_i^t|_{t\\geq 0}" + }, + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "inline_equation", + "content": "I(h_i^t,c_i)" + }, + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "text", + "content": " is formulated as follow:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 230, + 209, + 481, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 209, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 230, + 209, + 481, + 239 + ], + "type": "interline_equation", + "content": "I \\left(h _ {i} ^ {t}, c _ {i}\\right) = \\sum_ {i = 1} p \\left(h _ {i} ^ {t}, c _ {i}\\right) \\log \\frac {p \\left(h _ {i} ^ {t} , c _ {i}\\right)}{p \\left(h _ {i} ^ {t}\\right)}. \\tag {3}", + "image_path": "4a9fecc78c963858e1301930147ddd751df8f714bee81b378d465eeb3f58b605.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 479, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 278 + ], + "type": "text", + "content": "As the MI is hardly tractable within the batch training process, the auxiliary function named as MI score function is defined as Eq. (4) with trainable auxiliary linear transformation coefficients " + }, + { + "bbox": [ + 130, + 243, + 479, + 278 + ], + "type": "inline_equation", + "content": "w^{t}" + }, + { + "bbox": [ + 130, + 243, + 479, + 278 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 289, + 481, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 289, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 219, + 289, + 481, + 316 + ], + "type": "interline_equation", + "content": "f _ {k} \\left(h _ {i} ^ {t}, c _ {i}\\right) \\equiv e x p \\left(\\left(h _ {i} ^ {t}\\right) ^ {T} w ^ {t} c _ {i}\\right) \\propto \\frac {p \\left(h _ {i} ^ {t} , c _ {i}\\right)}{p \\left(h _ {i} ^ {t _ {d}}\\right)}. \\tag {4}", + "image_path": "0939117172b365906bfbfbef2b33f506e4f328eb8dc4f983eb10876683b3fbd4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 321, + 479, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 479, + 357 + ], + "type": "text", + "content": "Since MI score function is proportional to " + }, + { + "bbox": [ + 130, + 321, + 479, + 357 + ], + "type": "inline_equation", + "content": "I(h_i^t, c_i)" + }, + { + "bbox": [ + 130, + 321, + 479, + 357 + ], + "type": "text", + "content": " as formulated in Eq. (4), MI score function is maximized for the same vehicle and minimized for different vehicles using the infoNCE loss [26] function, formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 214, + 368, + 481, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 368, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 214, + 368, + 481, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left[ \\frac {1}{t _ {f}} \\sum_ {t = 1} ^ {t _ {f}} \\log \\frac {f _ {k} \\left(h _ {i} ^ {t} , c _ {i}\\right)}{\\Sigma_ {j = 1} ^ {n} f _ {k} \\left(h _ {j} ^ {t} , c _ {i}\\right)} \\right], \\tag {5}", + "image_path": "1d8255a55fa8c683b2298e1c56aa641e41071799de54e17195d5bbc0d7e2f3df.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 404, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 479, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 404, + 479, + 453 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 404, + 479, + 453 + ], + "type": "text", + "content": " denotes the number of data instances in a single batch and " + }, + { + "bbox": [ + 130, + 404, + 479, + 453 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 130, + 404, + 479, + 453 + ], + "type": "text", + "content": " is the future time horizon. With this loss function, the MI between the history and future states of the identical vehicle is enlarged, whereas the MI with other vehicles is suppressed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 471, + 348, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 471, + 348, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 471, + 348, + 483 + ], + "type": "text", + "content": "3.2 Generalizing for Geometric Diversity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 491, + 480, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 480, + 609 + ], + "type": "text", + "content": "An Ego-centric Driving Dataset. The intersection-centric dataset used for maneuver representation learning has a limitation in that it lacks geometric diversity. To address this limitation and ensure the model's generalizability across various geometric shapes of driving environments, we incorporate an ego-centric driving dataset [4] alongside the intersection-centric dataset. The ego-centric driving dataset is constructed using sensor-equipped vehicles, providing a diverse range of driving environments with various road shapes. By integrating this dataset into our training pipeline, we can effectively train the map adaptation module and enhance the model's ability to generalize to different geometric scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 479, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 667 + ], + "type": "text", + "content": "Map Adaptation Module. In the proposed maneuver prediction model, we introduce the map adaptation module as a downstream network aimed at generalizing the pre-trained maneuver representations to various road shapes. For map" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "Jeon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 121, + 470, + 300 + ], + "blocks": [ + { + "bbox": [ + 145, + 121, + 470, + 300 + ], + "lines": [ + { + "bbox": [ + 145, + 121, + 470, + 300 + ], + "spans": [ + { + "bbox": [ + 145, + 121, + 470, + 300 + ], + "type": "image", + "image_path": "cbf818a36a5dbd51b352dcfc6b08abdb1b92e36fe9917d6be129f5ca5cda31fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 311, + 428, + 322 + ], + "lines": [ + { + "bbox": [ + 182, + 311, + 428, + 322 + ], + "spans": [ + { + "bbox": [ + 182, + 311, + 428, + 322 + ], + "type": "text", + "content": "Fig. 6: Overall architecture of the map adaptation modules." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": "adaptation, we adopt an off-the-shelf map-based prediction model, LaneGCN [21]. LaneGCN utilizes lane graphs and Graph Convolutional Networks (GCN) to encode road geometries. It consists of multiple modules for encoding and fusing geometric and trajectory features, including MapNet and Lane to Lane for map encoding based on GCN, and Actor to Lane and Lane to Actor to fuse map and trajectory features using an attention mechanism. The overall network is trained using cross-entropy loss as defined in Eq. (6), where " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": " represent the ground truth and prediction of the vehicle maneuver, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 259, + 446, + 481, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 259, + 446, + 481, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = - \\sum_ {i = 1} ^ {k} y _ {i} \\log \\left(p _ {i}\\right). \\tag {6}", + "image_path": "877011b022ce38406bdba56430710b5b929e89019226d1ae0dbea7ac9e8b8acc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 483, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 542 + ], + "type": "text", + "content": "It is important to note that the maneuver head is employed only during the training process. This is due to the potential bias that may arise in the model outputting predominantly straight maneuvers during the map adaptation process. This bias stems from the maneuver imbalance present in the ego-centric driving dataset, which can lead to biased training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 591 + ], + "type": "text", + "content": "Instead, during the inference stage, we utilize probability density functions for each maneuver based on the intersection-centric dataset to estimate the likelihood of the given map-adapted maneuver context. This approach helps mitigate bias and ensures more accurate maneuver predictions during inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 609, + 343, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 343, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 343, + 621 + ], + "type": "text", + "content": "3.3 Estimating Likelihood of Maneuvers" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Given that the intersection-centric dataset exhibits a well-distributed range of maneuvers, we estimate the likelihood of maneuvers for a given historical trajectory based on the maneuver distribution within this dataset. To achieve this," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 250, + 209 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 250, + 209 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 250, + 209 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 250, + 209 + ], + "type": "image", + "image_path": "fc6d6a87118ee18a72acf114536225eb4a3135a2b217353066390952c807e0fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "lines": [ + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "type": "text", + "content": "Fig. 7: Distribution of the training dataset for each maneuver class and probability density function." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 116, + 364, + 209 + ], + "blocks": [ + { + "bbox": [ + 256, + 116, + 364, + 209 + ], + "lines": [ + { + "bbox": [ + 256, + 116, + 364, + 209 + ], + "spans": [ + { + "bbox": [ + 256, + 116, + 364, + 209 + ], + "type": "image", + "image_path": "90cc490c362b7667404d674711de456daf9c8ea1ecc482aaa62871cfce709754.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 372, + 116, + 478, + 209 + ], + "blocks": [ + { + "bbox": [ + 372, + 116, + 478, + 209 + ], + "lines": [ + { + "bbox": [ + 372, + 116, + 478, + 209 + ], + "spans": [ + { + "bbox": [ + 372, + 116, + 478, + 209 + ], + "type": "image", + "image_path": "af82451d2a79d7309d4c0974e15461f497e10309e8c3cd7a35e3f4057a8c552d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 261, + 479, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 261, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 131, + 261, + 479, + 309 + ], + "type": "text", + "content": "we first obtain probability density functions for each maneuver using Gaussian Kernel Density Estimation (Gaussian KDE) [24]. We reduce the dimensionality of maneuver features and then define the probability density function, " + }, + { + "bbox": [ + 131, + 261, + 479, + 309 + ], + "type": "inline_equation", + "content": "f(x, H_d)" + }, + { + "bbox": [ + 131, + 261, + 479, + 309 + ], + "type": "text", + "content": ", for each maneuver class using the following equations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 317, + 481, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 317, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 246, + 317, + 481, + 343 + ], + "type": "interline_equation", + "content": "K (u) = \\frac {1}{2 \\pi} \\exp \\left(- \\frac {1}{2} u ^ {T} u\\right), \\tag {7}", + "image_path": "3c7f3f6812c26d7855e370628916de889f68eb268ae582f294a878a27d88485d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 229, + 351, + 481, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 351, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 229, + 351, + 481, + 369 + ], + "type": "interline_equation", + "content": "K _ {H _ {d}} (u) = \\det (H _ {d}) ^ {- 1 / 2} K (H _ {d} ^ {- 1 / 2} u), \\tag {8}", + "image_path": "d893b3d3798f718c73b1ccd57a56e21bc0cf61b344ca83b13c932b4210d8d984.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 236, + 380, + 481, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 236, + 380, + 481, + 413 + ], + "type": "interline_equation", + "content": "f (x, H _ {d}) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} K _ {H _ {d}} \\left(x - x _ {i}\\right), \\tag {9}", + "image_path": "f4ae37ef92df23437496ca8b237819d3b15728a1a5f44d1867b87593ff02e843.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "text", + "content": " is bandwidth matrix parameter for each maneuver " + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "inline_equation", + "content": "d \\in \\{ST, RT, LT\\}" + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "inline_equation", + "content": "x_{i=1,\\dots,N}" + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "text", + "content": " is the training data points consist of each maneuver class. Here, " + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "inline_equation", + "content": "ST, RT" + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "inline_equation", + "content": "LT" + }, + { + "bbox": [ + 131, + 417, + 479, + 477 + ], + "type": "text", + "content": " denote going straight, right turn and left turn respectively. The Epanechnikov kernel method is employed whose bandwidth parameter is defined in Eq. (10):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 239, + 486, + 481, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 486, + 481, + 516 + ], + "spans": [ + { + "bbox": [ + 239, + 486, + 481, + 516 + ], + "type": "interline_equation", + "content": "H _ {d} = \\left(\\frac {4 \\sigma_ {d} ^ {5}}{3 N}\\right) ^ {\\frac {1}{5}} \\approx 1. 0 6 \\sigma_ {d} N ^ {- \\frac {1}{5}}, \\tag {10}", + "image_path": "dc83fe08f810eba55ba83e8df302c716061788951c56be53fe929ef931b8dbdd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 520, + 315, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 315, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 315, + 531 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 520, + 315, + 531 + ], + "type": "inline_equation", + "content": "\\sigma_{d}" + }, + { + "bbox": [ + 132, + 520, + 315, + 531 + ], + "type": "text", + "content": " denotes the standard deviation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 549, + 290, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 290, + 563 + ], + "type": "text", + "content": "4 Result and Case Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 574, + 261, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 261, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 261, + 586 + ], + "type": "text", + "content": "4.1 Quantitative Result" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "text", + "content": "The performance of the proposed maneuver classification scheme enhanced by employing heterogeneous datasets is quantitatively compared with the conventional various sampling strategies: a vanilla dataset, oversampled dataset and dataset augmented with SMOTE [5]. We categorize the future maneuvers in Left turn (LT), Straight (ST), and Right turn (RT) from the observed historical trajectories of all surrounding vehicles in the scene." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "Jeon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 207, + 247 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 207, + 247 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 207, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 207, + 247 + ], + "type": "image", + "image_path": "e63e70f73769bc3ecc64cc9debfba6cf0a598c85374c8472f0426333163774a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 306, + 303 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 306, + 303 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 306, + 303 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 306, + 303 + ], + "type": "image", + "image_path": "116d4e25ea15a630a29f6ca371bb42a82000450b82bfc760b96f7227d2fcb229.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 317, + 480, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 317, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 480, + 340 + ], + "type": "text", + "content": "Fig. 8: Baselines and the proposed method in preparing the training dataset and general process in maneuver prediction." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 388, + 302 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 388, + 302 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 388, + 302 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 388, + 302 + ], + "type": "image", + "image_path": "1979c9a83a1a56a78e476fbb9dbfd023e0d93a6d227f36bfc206dc786ba057e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 402, + 114, + 477, + 303 + ], + "blocks": [ + { + "bbox": [ + 402, + 114, + 477, + 303 + ], + "lines": [ + { + "bbox": [ + 402, + 114, + 477, + 303 + ], + "spans": [ + { + "bbox": [ + 402, + 114, + 477, + 303 + ], + "type": "image", + "image_path": "a9250907369fae345d3faab02555f81fb1b2c7be57fc74709a83396419575d17.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 140, + 387, + 479, + 533 + ], + "blocks": [ + { + "bbox": [ + 130, + 355, + 480, + 378 + ], + "lines": [ + { + "bbox": [ + 130, + 355, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 480, + 378 + ], + "type": "text", + "content": "Table 2: Overall maneuver classification performance of the proposed method comparing with various conventional approaches" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 387, + 479, + 533 + ], + "lines": [ + { + "bbox": [ + 140, + 387, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 140, + 387, + 479, + 533 + ], + "type": "table", + "html": "
Training DatasetLT validation setST validation setRT validation set
RecallPrec.F1AUCRecall.Prec.F1AUCRecallPrec.F1AUC
Vanilla \nArgo.0.0000.5000.0000.6931.0000.8240.9040.5000.0000.5000.0000.594
Argo. + \nOver.0.7430.7590.7510.8790.9390.9430.9410.8490.6960.6440.6690.875
Argo. + \nSMOTE0.7530.7030.7270.9390.8800.9550.9160.9160.8150.4590.5870.927
Ours0.9420.6810.7900.9520.9050.9870.9440.9440.9890.6610.7730.952
", + "image_path": "beddf18691df7c3d30b9427bcf05d1b21087d83a85b867c0c623ab7dfdae192f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": "Dataset compositions and training processes of the baselines are illustrated in Fig. 8 by comparing with the proposed framework. To evaluate the maneuver classification performance, we employed the AUC (Area Under Curve) of the ROC (Receiver Operating Characteristic) curve for each maneuver class and supplementary metrics, Precision (Prec.), Recall, and F1 Score (F1). The threshold for the precision and recall is 0.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "As shown in Tab. 2, the AUC, which is a representative performance metric, is improved compared to the conventional sampling methods in all maneuver classes. We found that when the model is trained with the vanilla Argoverse" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 301, + 237 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 301, + 237 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 301, + 237 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 301, + 237 + ], + "type": "image", + "image_path": "48dd99aa2223520b7e8daa01b0e6192876bae021436f6c9df42a5005fdf90d7d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 482, + 269 + ], + "type": "text", + "content": "Fig. 9: Data distribution comparison according to the temporal distance to the end of the intersection." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 116, + 471, + 237 + ], + "blocks": [ + { + "bbox": [ + 314, + 116, + 471, + 237 + ], + "lines": [ + { + "bbox": [ + 314, + 116, + 471, + 237 + ], + "spans": [ + { + "bbox": [ + 314, + 116, + 471, + 237 + ], + "type": "image", + "image_path": "fccd5a0cbcc9232247e111b83557b3772410f5801f401c95b6e77c325e9f0d7a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 291, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 479, + 339 + ], + "type": "text", + "content": "dataset (only ego-centric), it predicts most samples as ST, thereby achieving a high recall but a low precision, which the model is extremely biased toward outputing ST. The zero recall in LT and RT (which means there are no predictions for LT and RT with a score larger than 0.5.) also proves the biased training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 340, + 482, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 436 + ], + "type": "text", + "content": "Although the model achieves better performance than the vanilla dataset case by employing the oversampling strategy or SMOTE, they are still inferior than the proposed heterogeneous dataset result. Therefore, We argue that sampling-based methods which try to uniformize only the number of samples while ignoring the importance of sample diversity, is unable to overcome the imbalanced dataset problem. Conclusively, introducing the higher variety in the training samples by leveraging the proposed intersection-centric dataset is required for classifying the future maneuver of vehicles without bias." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 460, + 253, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 253, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 253, + 471 + ], + "type": "text", + "content": "4.2 Qualitative Result" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "content": "Maneuver Representation Learning. We visualized the learned representations for each maneuver class using [22] to analyze the behavior of the maneuver classification model. By reducing the dimension of the interaction-aware maneuver context, " + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 484, + 479, + 544 + ], + "type": "text", + "content": ", every samples in the training dataset are mapped to the two dimensional space as shown in Fig. 9." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": "In Fig. 9, the maneuver contexts for each maneuver class are scattered according to the time gap to the intersection. The points in blue, cyan, and green denote going straight (ST), left turn (LT) and right turn (RT) maneuver, respectively. The blue points show that the ST maneuver is less spread compared to other maneuvers even at the 4 seconds before the intersection, which implies that the ST maneuver is much easier to distinguish than other turning maneuvers even if remarkable maneuver feature is not observed. On the other hand, turning maneuvers with cyan and green points are widely distributed invading the ST area in the left plot in Fig. 9, which means that the turning maneuvers are likely to be misclassified as the ST." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "Jeon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 301, + 237 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 301, + 237 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 301, + 237 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 301, + 237 + ], + "type": "image", + "image_path": "4fe9eee2a288241b3582f58eccc0a9bc0ddf915c77f98c6f1a8b028db89e1dbc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 247, + 455, + 259 + ], + "lines": [ + { + "bbox": [ + 157, + 247, + 455, + 259 + ], + "spans": [ + { + "bbox": [ + 157, + 247, + 455, + 259 + ], + "type": "text", + "content": "Fig. 10: Data distribution comparison according to the training dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 116, + 471, + 237 + ], + "blocks": [ + { + "bbox": [ + 313, + 116, + 471, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 116, + 471, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 471, + 237 + ], + "type": "image", + "image_path": "e6cd5575eae06949054bb168c049a3babca287a246e336dc6e879ae1da1348c7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 278, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 397 + ], + "type": "text", + "content": "Furthermore, as shown in Fig. 10, the distribution of data points is prominently different according to the dataset. In Fig. 10, the data points in the red circles indicate the challenging scenarios: all of three maneuvers are blended in this area so the maneuver representation is hard to be categorized. However, in the imbalanced dataset, the distance from the ambiguous area (the red circle) to the center of the ST maneuver cluster is smaller than other maneuvers, whereas the distances toward each maneuver class are similar in the balanced dataset. We argue that this characteristic of maneuver representations induce the model to generate the biased predictions toward the ST when trained with the imbalanced dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 416, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 488 + ], + "type": "text", + "content": "Case Study on Maneuver Classification. In Fig. 11, use case of the proposed maneuver prediction model is illustrated. Using the historical trajectory, blue lines, each maneuver probability, bar graph, is calculated considering road shape and multiple surrounding vehicles, black dotted lines and grey dots, respectively. Since blue dotted line denotes the ground truth future trajectory, predicted maneuver probability can be qualitatively evaluated." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 488, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 585 + ], + "type": "text", + "content": "As shown in (a) and (b) in Fig. 11, LT maneuver is well predicted even though the turning maneuver is not shown in the historical trajectory; maneuver probability of the LT class takes the highest likelihood. Furthermore, despite the ground truth label of LT class, ST and/or RT probability is also generated considering the road shape, which enables the planning algorithm to consider low probability cases so that abnormal situations can be handled. Similarly, ST and RT maneuver are well-predicted as shown in 11: (c), (d) for ST class and (e), (f) for RT class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 604, + 220, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 220, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 220, + 616 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "In the deep learning based autonomous driving system, the imbalanced data was the critical bottleneck in generalizing the model because of the nature of naturalistic driving dataset: long-tail problem. In the proposed research, for overcoming" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 129, + 251, + 220 + ], + "blocks": [ + { + "bbox": [ + 143, + 129, + 251, + 220 + ], + "lines": [ + { + "bbox": [ + 143, + 129, + 251, + 220 + ], + "spans": [ + { + "bbox": [ + 143, + 129, + 251, + 220 + ], + "type": "image", + "image_path": "f91f8630a54df143e0b847570433331cb251a5008b63e764e7ff27055b59226e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 222, + 202, + 231 + ], + "lines": [ + { + "bbox": [ + 193, + 222, + 202, + 231 + ], + "spans": [ + { + "bbox": [ + 193, + 222, + 202, + 231 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 253, + 129, + 360, + 220 + ], + "blocks": [ + { + "bbox": [ + 231, + 116, + 465, + 125 + ], + "lines": [ + { + "bbox": [ + 231, + 116, + 465, + 125 + ], + "spans": [ + { + "bbox": [ + 231, + 116, + 465, + 125 + ], + "type": "text", + "content": ": Historical trajectory (input) : Current position : Future trajectory (GT)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 129, + 360, + 220 + ], + "lines": [ + { + "bbox": [ + 253, + 129, + 360, + 220 + ], + "spans": [ + { + "bbox": [ + 253, + 129, + 360, + 220 + ], + "type": "image", + "image_path": "a0f48d55c7d400b28ee5f405d06faeae9ad405ca647503eaa8d4c64afd5d422e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 222, + 312, + 231 + ], + "lines": [ + { + "bbox": [ + 301, + 222, + 312, + 231 + ], + "spans": [ + { + "bbox": [ + 301, + 222, + 312, + 231 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 362, + 130, + 470, + 220 + ], + "blocks": [ + { + "bbox": [ + 362, + 130, + 470, + 220 + ], + "lines": [ + { + "bbox": [ + 362, + 130, + 470, + 220 + ], + "spans": [ + { + "bbox": [ + 362, + 130, + 470, + 220 + ], + "type": "image", + "image_path": "82f819ddd300d2df4c2f15ba92acde8133001b5b658ba9c0d6cbd41bc8d6a2fa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 221, + 420, + 231 + ], + "lines": [ + { + "bbox": [ + 410, + 221, + 420, + 231 + ], + "spans": [ + { + "bbox": [ + 410, + 221, + 420, + 231 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 143, + 234, + 251, + 324 + ], + "blocks": [ + { + "bbox": [ + 143, + 234, + 251, + 324 + ], + "lines": [ + { + "bbox": [ + 143, + 234, + 251, + 324 + ], + "spans": [ + { + "bbox": [ + 143, + 234, + 251, + 324 + ], + "type": "image", + "image_path": "41b651fe1d23f6b197391a5ef9bc436d02bf121de608cb6527b204b3be95162c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 327, + 202, + 335 + ], + "lines": [ + { + "bbox": [ + 193, + 327, + 202, + 335 + ], + "spans": [ + { + "bbox": [ + 193, + 327, + 202, + 335 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 253, + 235, + 360, + 324 + ], + "blocks": [ + { + "bbox": [ + 253, + 235, + 360, + 324 + ], + "lines": [ + { + "bbox": [ + 253, + 235, + 360, + 324 + ], + "spans": [ + { + "bbox": [ + 253, + 235, + 360, + 324 + ], + "type": "image", + "image_path": "afc04b818a66a6a01ee282596b54682dc237409a221a9e1cea5595bb67339c8d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 327, + 312, + 335 + ], + "lines": [ + { + "bbox": [ + 302, + 327, + 312, + 335 + ], + "spans": [ + { + "bbox": [ + 302, + 327, + 312, + 335 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 362, + 235, + 470, + 324 + ], + "blocks": [ + { + "bbox": [ + 362, + 235, + 470, + 324 + ], + "lines": [ + { + "bbox": [ + 362, + 235, + 470, + 324 + ], + "spans": [ + { + "bbox": [ + 362, + 235, + 470, + 324 + ], + "type": "image", + "image_path": "6e86913b894c11f150af32424a9fc5819fbbed9358408614eae4b392ada42cb8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 327, + 420, + 335 + ], + "lines": [ + { + "bbox": [ + 411, + 327, + 420, + 335 + ], + "spans": [ + { + "bbox": [ + 411, + 327, + 420, + 335 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "text", + "content": "Fig. 11: Prediction results for the representative scenarios. Left turn " + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "inline_equation", + "content": "(a, b)" + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "text", + "content": ", going straight " + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "inline_equation", + "content": "(c, d)" + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "text", + "content": " and right turn " + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "inline_equation", + "content": "(e, f)" + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "text", + "content": " maneuver can be predicted as a stochastic manner." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 386, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 506 + ], + "type": "text", + "content": "the data imbalance and biased prediction model, two dataset from difference modality is employed: in-house drone dataset and Argoverse dataset. The in-house Drone dataset has quantitatively abundant scenarios including a bunch of turning maneuvers, where the Argoverse dataset contains various road shape. By merging the strength of the heterogeneous dataset into a maneuver prediction model by introducing a two-stage training scheme with self-supervised learning, geometrically generalizable prediction model can be trained in balance in terms of the vehicle maneuver. Additionally, the probabilistic approach can provide additional information on the unlikely to happen maneuver so that the planning algorithm can consider abnormal driving situation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "type": "text", + "content": "However, the proposed prediction method can not consider the dependency between the future of the surrounding vehicle and the action plan of the ego vehicle. In real world, vehicles are continuously interacting each other and update their own driving strategy in real time. Therefore, for highly intelligent prediction network, the effect of the ego vehicle's future action plan has to be considered in predicting the maneuver of the other vehicles [18]. In this stage, simulation based approaches (e.g. MORAI SIM: Drive [1]) is highly promising to handle both of the data imbalance problem and reactiveness of the actors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by Autonomous Driving Development Center, AVP Division, Hyundai Motor Company and the National Research Foundation of Korea(NRF) funded by the Korea government(MSIT) under Grants 2022R1A2C200494413." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "Jeon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 142, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 137, + 142, + 348, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 142, + 348, + 153 + ], + "spans": [ + { + "bbox": [ + 137, + 142, + 348, + 153 + ], + "type": "text", + "content": "1. Morai simulator, https://www.morai.ai/drive 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 154, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 154, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 137, + 154, + 481, + 186 + ], + "type": "text", + "content": "2. Azadani, M.N., Boukerche, A.: A novel multimodal vehicle path prediction method based on temporal convolutional networks. IEEE Transactions on Intelligent Transportation Systems (2022) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 187, + 481, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 187, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 137, + 187, + 481, + 231 + ], + "type": "text", + "content": "3. Bock, J., Krajewski, R., Moers, T., Runde, S., Vater, L., Eckstein, L.: The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 1929-1934. IEEE (2020) 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 481, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 481, + 275 + ], + "type": "text", + "content": "4. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8748-8757 (2019) 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 276, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 481, + 308 + ], + "type": "text", + "content": "5. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research 16, 321-357 (2002) 2, 4, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 481, + 341 + ], + "type": "text", + "content": "6. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014) 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 365 + ], + "type": "text", + "content": "7. Cieslak, D.A., Chawla, N.V., Striegel, A.: Combating imbalance in network intrusion datasets. In: GrC. pp. 732-737. Citeseer (2006) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "type": "text", + "content": "8. Cui, Y., Song, Y., Sun, C., Howard, A., Belongie, S.: Large scale fine-grained categorization and domain-specific transfer learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4109-4118 (2018) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 481, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 431 + ], + "type": "text", + "content": "9. Deepa, T., Punithavalli, M.: An e-smote technique for feature selection in high-dimensional imbalanced dataset. In: 2011 3rd International Conference on Electronics Computer Technology. vol. 2, pp. 322-324. IEEE (2011) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 432, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 432, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 134, + 432, + 481, + 464 + ], + "type": "text", + "content": "10. Drummond, C., Holte, R.C., et al.: C4. 5, class imbalance, and cost sensitivity: why under-sampling beats over-sampling. In: Workshop on learning from imbalanced datasets II. vol. 11, pp. 1-8. Citeseer (2003) 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 465, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 465, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 134, + 465, + 481, + 498 + ], + "type": "text", + "content": "11. Feng, X., Cen, Z., Hu, J., Zhang, Y.: Vehicle trajectory prediction using intention-based conditional variational autoencoder. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 3514-3519. IEEE (2019) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 498, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 498, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 134, + 498, + 481, + 542 + ], + "type": "text", + "content": "12. Girase, H., Gang, H., Malla, S., Li, J., Kanehara, A., Mangalam, K., Choi, C.: Loki: Long term and key intentions for trajectory prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9803-9812 (2021) 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 543, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 543, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 543, + 481, + 586 + ], + "type": "text", + "content": "13. Goldhammer, M., Kohler, S., Zernetsch, S., Doll, K., Sick, B., Dietmayer, K.: Intentions of vulnerable road users—detection and forecasting by means of machine learning. IEEE transactions on intelligent transportation systems 21(7), 3035-3045 (2019) 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 587, + 481, + 609 + ], + "type": "text", + "content": "14. He, H., Garcia, E.A.: Learning from imbalanced data. IEEE Transactions on knowledge and data engineering 21(9), 1263-1284 (2009) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "type": "text", + "content": "15. Hu, Y., Li, Y., Huang, H., Lee, J., Yuan, C., Zou, G.: A high-resolution trajectory data driven method for real-time evaluation of traffic safety. Accident Analysis & Prevention 165, 106503 (2022) 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "16. J. Colyar, J.H.: Us highway 101 dataset. Federal Highway Administration (FHWA), Tech. Rep. FHWA-HRT-07-030 (2007) 6" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "17. Jeon, H., Choi, J., Kum, D.: Scale-net: Scalable vehicle trajectory prediction network under random number of interacting vehicles via edge-enhanced graph convolutional neural network. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2095-2102. IEEE (2020) 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "18. Jeon, H., Kim, S., Lee, K., Kang, D., Choi, J., Kum, D.: Are reactions to ego vehicles predictable without data?: A semi-supervised approach. IEEE Transactions on Intelligent Transportation Systems 24(6), 6477-6490 (2023) 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 225 + ], + "type": "text", + "content": "19. Kim, S., Jeon, H., Choi, J.W., Kum, D.: Diverse multiple trajectory prediction using a two-stage prediction network trained with lane loss. IEEE Robotics and Automation Letters 8(4), 2038-2045 (2022) 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "type": "text", + "content": "20. Kozerawski, J., Sharan, M., Yu, R.: Taming the long tail of deep probabilistic forecasting. arXiv preprint arXiv:2202.13418 (2022) 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "21. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: European Conference on Computer Vision. pp. 541-556. Springer (2020) 1, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "text", + "content": "22. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "23. Makansi, O., Çiçek, Ö., Marrakchi, Y., Brox, T.: On exposing the challenging long tail in future prediction of traffic actors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13147-13157 (2021) 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 353 + ], + "type": "text", + "content": "24. Marron, J.S., Wand, M.P.: Exact mean integrated squared error. The Annals of Statistics 20(2), 712-736 (1992) 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 396 + ], + "type": "text", + "content": "25. Mirus, F., Stewart, T.C., Conradt, J.: The importance of balanced data sets: Analyzing a vehicle trajectory prediction model based on neural networks and distributed representations. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020) 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "text", + "content": "26. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 449 + ], + "type": "text", + "content": "27. Park, D., Ryu, H., Yang, Y., Cho, J., Kim, J., Yoon, K.J.: Leveraging future relationship reasoning for vehicle trajectory prediction. arXiv preprint arXiv:2305.14715 (2023) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 449, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 481, + 482 + ], + "type": "text", + "content": "28. Peng, J., Bu, X., Sun, M., Zhang, Z., Tan, T., Yan, J.: Large-scale object detection in the wild from imbalanced multi-labels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9709-9718 (2020) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "text", + "content": "29. Robicquet, A., Sadeghian, A., Alahi, A., Savarese, S.: Learning social etiquette: Human trajectory prediction in crowded scenes. In: European Conference on Computer Vision (ECCV). vol. 2 (2020) 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 546 + ], + "type": "text", + "content": "30. Shangguan, Q., Fu, T., Wang, J., Fu, L., et al.: A proactive lane-changing risk prediction framework considering driving intention recognition and different lane-changing patterns. Accident Analysis & Prevention 164, 106500 (2022) 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "type": "text", + "content": "31. Shen, L., Lin, Z., Huang, Q.: Relay backpropagation for effective learning of deep convolutional neural networks. In: European conference on computer vision. pp. 467-482. Springer (2016) 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "32. Strigel, E., Meissner, D., Seeliger, F., Wilking, B., Dietmayer, K.: The ko-per intersection laserscanner and video dataset. In: 17th International IEEE Conference on Intelligent Transportation Systems (ITSC). pp. 1900-1901. IEEE (2014) 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 666 + ], + "type": "text", + "content": "33. Varadarajan, B., Hefny, A., Srivastava, A., Refaat, K.S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C.P., Anguelov, D., et al.: Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 7814-7821. IEEE (2022) 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "Jeon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 293 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Wang, J., Xu, M., Wang, H., Zhang, J.: Classification of imbalanced data by using the smote algorithm and locally linear embedding. In: 2006 8th international Conference on Signal Processing. vol. 3. IEEE (2006) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "35. Wang, K., Xue, Q., Lu, J.J.: Risky driver recognition with class imbalance data and automated machine learning framework. International journal of environmental research and public health 18(14), 7534 (2021) 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "text", + "content": "36. Westny, T., Frisk, E., Olofsson, B.: Vehicle behavior prediction and generalization using imbalanced learning techniques. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2003-2010. IEEE (2021) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 258 + ], + "type": "text", + "content": "37. Yang, D., Li, L., Redmill, K., Özgün, Ü.: Top-view trajectories: A pedestrian dataset of vehicle-crowd interaction from controlled experiments and crowded campus. In: 2019 IEEE Intelligent Vehicles Symposium (IV). pp. 899-904. IEEE (2019) 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "38. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022) 1" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Beyond the Data Imbalance in Vehicle Maneuver Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_content_list.json b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..19aeec293fb8feebcceb4c3319a88ea89694502d --- /dev/null +++ b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_content_list.json @@ -0,0 +1,1394 @@ +[ + { + "type": "text", + "text": "BeyondScene: Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gwanghyun Kim $^{1\\star}$ , Hayeon Kim $^{1\\star}$ , Hoigi Seo $^{1\\star}$ , Dong Un Kang $^{1\\star}$ , Se Young Chun $^{1,2\\dagger}$", + "bbox": [ + 238, + 210, + 764, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. of Electrical and Computer Engineering, $^{2}$ INMC & IPAI Seoul National University, Republic of Korea", + "bbox": [ + 285, + 253, + 712, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gwang.kim, khy5630, sehoiki3215, qkrtnsfk23, sychun}@snu.ac.kr", + "bbox": [ + 246, + 282, + 750, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Generating higher-resolution human-centric scenes with details and controls remains a challenge for existing text-to-image diffusion models. This challenge stems from limited training image size, text encoder capacity (limited tokens), and the inherent difficulty of generating complex scenes involving multiple humans. While current methods attempted to address training size limit only, they often yielded human-centric scenes with severe artifacts. We propose BeyondScene, a novel framework that overcomes prior limitations, generating exquisite higher-resolution (over 8K) human-centric scenes with exceptional text-image correspondence and naturalness using existing pretrained diffusion models. BeyondScene employs a staged and hierarchical approach to initially generate a detailed base image focusing on crucial elements in instance creation for multiple humans and detailed descriptions beyond token limit of diffusion model, and then to seamlessly convert the base image to a higher-resolution output, exceeding training image size and incorporating details aware of text and instances via our novel instance-aware hierarchical enlargement process that consists of our proposed high-frequency injected forward diffusion and adaptive joint diffusion. BeyondScene surpasses existing methods in terms of correspondence with detailed text descriptions and naturalness, paving the way for advanced applications in higher-resolution human-centric scene creation beyond the capacity of pretrained diffusion models without costly retraining. Project page: https://janeyon.github.io/beyond-scene.", + "bbox": [ + 261, + 332, + 743, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Human-centric scene generation $\\cdot$ Text-to-image diffusion model $\\cdot$ High-resolution", + "bbox": [ + 259, + 664, + 740, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 739, + 375, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human-centric scene generation [8, 12, 15, 16, 19-23, 27, 28, 32, 33, 38, 40, 42, 44-46], has emerged as a critical research area with significant academic and industrial applications. While recent advancements utilizing text-to-image (T2I) diffusion", + "bbox": [ + 212, + 770, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Authors contributed equally. † Corresponding author.", + "bbox": [ + 217, + 824, + 596, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/672bf6d39fbb5d8a3cf78b2843eb49f7347ba83a3116d7a182ae4c74a3e6b9b3.jpg", + "image_caption": [ + "Fig. 1: BeyondScene pushes the boundaries of high-resolution human-centric scene generation. Unlike existing methods that often suffer from unrealistic scenes, anatomical distortions, and limited text-to-image correspondence, BeyondScene excels in 1) highly detailed scenes, 2) natural and diverse humans, 3) fine-grained control. The color in each text represents the text for each instance that has the same color in the pose map." + ], + "image_footnote": [ + "\"In the sunny road, there are a clay animation character with lab coat, a clay character with red dotted dress and glasses, a clay character with short, chubby cheeks and glasses, a clay animation character with short, chubby checks wearing blue jeans, a clay character with spiky hair, a clay character with yellow T-shirts.\"" + ], + "bbox": [ + 230, + 143, + 772, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models [2, 26, 29] have yielded promising results in generating controllable human scenes [12, 15, 23, 45], scaling these methods to handle larger and more complex scenes such as involving multiple humans remains a significant challenge. These limitations primarily stem from following causes. 1) Training image size: Directly sampling from a limited training image size can introduce artifacts and constrain the final scene resolution. 2) Limited text encoder tokens: The restricted number of tokens in text encoders (typically 77 for Stable Diffusion) hinders the inclusion of detailed descriptions for multiple instances within the scene. 3) Inherent limitations of T2I diffusion models: These models struggle to generate complex scenes with several human figures and intricate details.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing attempts to address the training size issue, such as joint-process [6, 14, 47] and dilation-based methods [10], only partially address the problem. These methods often introduce new challenges specific to human-centric scenes as represented in Fig. 1, including: 1) Unrealistic scenes and objects: These methods can generate nonsensical scenarios with duplicated objects, humans defying gravity, and physically distorted environments. 2) Anatomical distortions: Generated scenes may exhibit unrealistic human anatomy, such as abnormal limbs or facial features. 3) Limited correspondence: Existing methods often fail to capture the complexity of scenes with multiple human instances and detailed descriptions, often generating single objects and lacking control over specific details like clothing or hairstyles.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose BeyondScene, a novel framework that overcomes these limitations, generating high-resolution (over 8K) human-centric scenes with exceptional text-image correspondence and naturalness. BeyondScene employs a staged and hierarchical approach, which looks similar to a classical multi-resolution manner, but it is closer to how artists establish a foundation before adding details.", + "bbox": [ + 212, + 332, + 789, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, BeyondScene generates a detailed base image focusing on essential elements, human poses and detailed descriptions beyond the token limits. This initial stage enables detailed instance creation for multiple humans, surpassing the limitations of text encoders. Secondly, the method leverages an instance-aware hierarchical enlargement process to convert this base image to a higher resolution output beyond the training image size. Unlike naive super-resolution methods that simply scale resolution without considering text and instances, our approach refines content and adds more details aware of text and instances. This is achieved through our proposed novel techniques, including 1) high frequency-injected forward diffusion that addresses the issue of blurred low-quality results in image-to-image translation with the upsampled image by adaptively injecting high-frequency details into the upsampled image, enhancing the final output while preserving content and 2) adaptive joint diffusion that facilitates efficient and robust joint diffusion while maintaining control over human characteristics like pose and mask. This approach utilizes view-wise conditioning of text and pose information, along with a variable stride for the joint process. Furthermore, similar to how artists add details progressively, our method allows for the addition of details at different stages, exploiting the changing receptive field of each view.", + "bbox": [ + 212, + 428, + 789, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We comprehensively evaluate BeyondScene using qualitative and quantitative metrics, along with user studies. The results demonstrate significant improvements over existing methods in terms of 1) correspondence with detailed text descriptions and 2) naturalness and reduction of artifacts. Furthermore, we showcase the result of $8192 \\times 8192$ image generation beyond 8K ultra high-resolution, demonstrating the capability for generating even higher-resolution images as displayed in Fig. 2. These advancements pave the way for exciting new applications in higher-resolution human-centric scene creation.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BeyondScene", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/58e3552bda3e6dd42e6614ddf0c60b36f5555c6a3b2bf4733604d06567f38fdb.jpg", + "image_caption": [ + "Fig. 2: Beyond 8K ultra-high resolution image. This $8192 \\times 8192$ image, generated by BeyondScene, surpasses the training resolution of SDXL by $64 \\times$ , while exceeding the technical classification of 8K ( $7680 \\times 4320$ )." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 784, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 217, + 662, + 382, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Controllable Human Generation", + "text_level": 1, + "bbox": [ + 217, + 694, + 527, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Early approaches to controllable human generation [8, 19-22, 27, 28, 40, 42, 44, 46] relied on pose guidance and source images, achieving success within specific scenarios but struggling with diverse scenes and arbitrary poses. Text-based conditioning methods emerged with the rise of large vision-language models [25], but initial attempts [16, 32, 38] faced limitations in vocabulary size and open vocabulary settings. Recent T2I diffusion [2, 24, 26, 30]-based methods like ControlNet [45], T2I-Adapter [23], GLIGEN [15], and HumanSD [12] introduced methods for incorporating diverse conditions, but scaling to larger scenes with", + "bbox": [ + 215, + 718, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "multiple individuals remains a challenge. This work proposes BeyondScene, a novel framework specifically designed to overcome these limitations and enable the generation of high-resolution human-centric scenes.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 Large Scene Generation Using Diffusion Models", + "text_level": 1, + "bbox": [ + 214, + 213, + 658, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Achieving high-resolution image generation [9, 26] presents significant hurdles. Training models from scratch or fine-tuning pretrained models [2, 7, 11, 24, 26, 30, 31, 34, 37, 43, 48] requires immense computational resources and struggles with the complexity of high-dimensional data. Recent exploration has ventured into training-free methods, with approaches like MultiDiffusion [6], SyncDiffusion [14] (joint diffusion), and ScaleCrafter [10] (dilation-based) emerging. However, these methods often introduce challenges specific to human-centric scenes, including unrealistic objects, anatomical distortions, and limited correspondence between text descriptions and generated images. BeyondScene, our proposed framework, addresses these limitations through detailed base image generation and instance-aware hierarchical enlargement.", + "bbox": [ + 212, + 238, + 789, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 BeyondScene", + "text_level": 1, + "bbox": [ + 214, + 428, + 380, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Current approaches to generating large scenes attempt to fill the entire canvas at once, often leading to challenges in capturing complex details and ensuring physical realism. Inspired by the artistic workflow of human painters, who establish a foundation with key elements and progressively refine them, we introduce BeyondScene, a novel framework that generates high-resolution human-centric images through a staged and hierarchical approach as represented in Fig. 3.", + "bbox": [ + 212, + 460, + 787, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BeyondScene operates in two key stages, 1) detailed base image generation (Sec. 3.1) that focuses on generating high-quality base images with precise control over human poses and detailed descriptions beyond typical token limitations and 2) instance-aware hierarchical enlargement (Sec. 3.2) where the base image is progressively enlarged while maintaining control and adding scene details.", + "bbox": [ + 212, + 551, + 787, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Detailed Base Image Generation", + "text_level": 1, + "bbox": [ + 214, + 648, + 532, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating human-centric scenes with multiple individuals using T2I diffusion models faces a key challenge: the limited number of tokens available in the text encoder restricts detailed descriptions to a few key elements. To overcome this, we generate detailed base image as illustrated in Fig. 3. Firstly, individual instances are generated in the training resolution of pose-guided T2I diffusion models [12, 15, 23, 45]. This allows for detailed descriptions beyond token limitations, tailored to each individual. Subsequently, the generated instances are segmented, cropped, and resized to the desired size. Next, a diffusion-based inpainting method is employed to create an instance-aware background that seamlessly integrates with the individual instances. This ensures a realistic and coherent background tailored to the specific scene. To address potential inconsistencies in brightness", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BeyondScene", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bf639b9fd424f6bca4a2b92dd0ecb315487dbb9e978d8336283b95e1d773f88d.jpg", + "image_caption": [ + "Fig. 3: BeyondScene generates high-resolution images in two stages. First, individual instances are created using pose-guided T2I diffusion models, segmented, cropped, and placed onto an inpainted background. The tone of image is then normalized. In the second stage (illustrated in Fig. 4), this base image is progressively enlarged while maintaining detail and quality, effectively refining the image and adding further details leveraging high-frequency injected forward diffusion and adaptive joint diffusion (AJ)." + ], + "image_footnote": [], + "bbox": [ + 232, + 145, + 769, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and color tones across individual elements, tone normalization through contrast limited adaptive histogram equalization (CLAHE) is performed. This involves first creating a grid histogram for pixel values within a specified grid, followed by adjusting the local contrast to align with the image's overall contrast. Notably, direct equalization in the presence of noise within the grid could induce the noise amplification. To prevent this, equalization is preceded by a redistribution process that includes limiting the maximum height of the grid histogram.", + "bbox": [ + 212, + 580, + 787, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Instance-Aware Hierarchical Enlargement", + "text_level": 1, + "bbox": [ + 214, + 709, + 606, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BeyondScene introduces a novel approach to address the challenge of generating high-resolution outputs beyond the training image size through a instance-aware hierarchical enlargement process, which surpasses mere scaling by actively refining content and incorporating additional details aware of text and instance.", + "bbox": [ + 212, + 734, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the initial step, a low-resolution image is upsampled, and a forward diffusion process is applied to create a noisy upsampled image. Subsequently, a joint diffusion process refines the image to a high resolution over the training resolution.", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 341, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af0efa79b24c3839a184bb36065655b15dd1e50b9c440d83e3bd58b3e7ad0fc9.jpg", + "image_caption": [ + "3.2 Instance-Aware Hierarchical Enlargement", + "a. High frequency-injected forward diffusion" + ], + "image_footnote": [], + "bbox": [ + 253, + 176, + 738, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c6c58c75699dce09680a47badcb2b2398098f112d481982d3571c04128c89af.jpg", + "image_caption": [ + "b. Adaptive joint diffusion (AJ) At the diffusion timestep $t$", + "Fig. 4: Our instance-aware hierarchical enlargement involves two crucial processes: 1) High frequency-injected forward diffusion, which enables to achieve high resolution through a joint diffusion employing adaptive pixel perturbation. 2) Adaptive joint diffusion, dynamically regulating stride and conditioning of pose and text based on the presence of instances." + ], + "image_footnote": [], + "bbox": [ + 233, + 318, + 763, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, a drawback of this naive strategy is the induction of blurriness due to the model assigning sufficiently high likelihood to low-quality outputs. Also, this strategy often encounters issues such as unnatural duplication of objects and humans, stemming from the uniform application of the same text prompt to all views without pose conditioning. To mitigate these, the paper proposes a novel high frequency-injected forward diffusion and adaptive joint diffusion as illustrated in Fig. 4.", + "bbox": [ + 212, + 542, + 787, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High frequency-injected forward diffusion High frequency-injected forward diffusion enhances the translation of noisy latents from the upsampled image to high resolution with intricate details. This is achieved through a joint diffusion process employing adaptive pixel perturbation. This technique injects high-frequency details into the upsampled image based on a Canny edge map. This process enhances the final output's quality while preserving content. While simply adding pixel noise sharpens the image, it can also introduce flickering artifacts at the borders. Our adaptive approach based on the Canny map significantly improves image quality and reduces these artifacts. Specifically, we first up-sample the low-resolution image and calculate Canny map. Then Gaussian blur is applied to the map and normalized it to get probability map $C$ . $\\epsilon_{i,j} \\sim \\mathcal{U}(0,1)$ is sampled", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BeyondScene", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on each pixel, $I_{i,j}$ which denotes $i$ -th row and $j$ -th column pixel of image $I$ . If $\\epsilon_{i,j} > C_{i,j}$ , the pixel value within $d_r$ pixel distance is replaced so that we could apply pixel perturbation adaptively to avoid flickering artifacts near borders.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive joint diffusion To mitigate duplication of objects and humans in the naive joint diffusion, we introduce adaptive conditioning and adaptive stride.", + "bbox": [ + 214, + 213, + 789, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive conditioning We introduce adaptive view-wise conditioning, which leverages the segmentation maps obtained from the base image generation stage. For each view, we check which human instances are present. If an instance is included, we add its pose and detailed description to the prompt used for that view. Additionally, if the detailed text description includes specific parts (head, face, upper body, etc.), we can apply these details to the corresponding regions using fine-grained segmentation maps. This facilitates efficient and robust joint diffusion while maintaining control over human characteristics like pose and appearance.", + "bbox": [ + 212, + 257, + 789, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive stride To effectively enhance the quality of the generated scene, we employ an adaptive stride for the joint process. We reduce the stride in regions containing humans, ensuring that these areas capture fine details. In contrast, the stride is increased in background regions, allowing for more efficient computation in these areas. This combination of adaptive view-wise conditioning and adaptive stride allows BeyondScene to effectively handle scenes with multiple humans while maintaining high-resolution detail and controllable image generation. The details and algorithms on our methods are presented in the supplementary material.", + "bbox": [ + 212, + 407, + 787, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 551, + 375, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 582, + 447, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We comprehensively evaluate our method using both qualitative and quantitative metrics. To ensure a fair comparison, all compared models are implemented using Stable Diffusion XL (SDXL) [24]-based architectures. We utilize SDXL-ControlNet-Openpose [3, 45], SDXL-inpainting [4], and Lang-Segment-Anything [1] for our pose-guided T2I diffusion, inpainting diffusion model, and segmentation models. The training resolution is set to $1024 \\times 1024$ , while inference resolutions are varied across $2048 \\times 2048$ ( $4 \\times 1:1$ ), $3584 \\times 2048$ ( $7 \\times 7:4$ ), $4096 \\times 2048$ ( $8 \\times 2:1$ ), $4096 \\times 4096$ ( $16 \\times 1:1$ ), and $8192 \\times 8192$ ( $64 \\times 1:1$ ).", + "bbox": [ + 212, + 606, + 787, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Testing data Existing T2I generation datasets often lack large-scale scenes featuring multiple people with detailed descriptions for each individual. To address this limitation, we leverage the CrowdCaption dataset [35] as our primary test set. This dataset features images containing large crowds of people, allowing us to evaluate the effectiveness of our method in generating complex scenes with numerous individuals. For each image in the CrowdCaption dataset, we obtain", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f11e7bd20de4f971039ab1b694fd0c4c7f4887a0531e9a26e839ff6b51ae8b8f.jpg", + "image_caption": [], + "image_footnote": [ + "\"In empty stage of opera house, there are a ballerina with pink suit, a ballerina wearing blue ballet suit, a ballerina with pink suit, a ballerina with blue suit, a ballerina with yellow sparkling ballet suit, a ballerina with blue suit, a ballerina with pink suit, a ballerina with blue suit, a ballerina with pink suit.\"" + ], + "bbox": [ + 233, + 145, + 769, + 340 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/03a6eefb3b08aa41ef3a7c7a0d705d101a33bbe38ce23245cb38901dd28dc91b.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison for generating high-resolution human scenes ( $3584 \\times 2048$ ). While existing approaches like T2I-Direct (SDXL [24]), T2I-Large (MultiDiffusion [6], SyncDiffusion [14], and ScaleCrafter [10]), and Visual+T2I (ControlNet [24, 45], T2IAdapter [23, 24], and R-MultiDiffusion [6]) models struggle with artifacts, our method achieves superior results by producing images with minimal artifacts, strong text-image correspondence, and a natural look. The color in each description represents the description for each instance that has the same color in the pose map." + ], + "image_footnote": [ + "\"In the alien planet with galaxy in the sky, there are a storm trooper, a storm trooper, a storm trooper, a Darth Vader, a storm trooper, a storm trooper, a storm trooper.\"" + ], + "bbox": [ + 233, + 372, + 767, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "descriptions and individual poses using GRIT [36] and ViTPose [39]. Additionally, we obtain a global caption generated using GPT4 [5]. We filter the CrowdCaption images based on two criteria: number of humans (max 8 to accommodate token limits of baselines) and aspect ratio (close to 1:1 and 2:1), which results in a collection of 100 images for each aspect ratio (1:1 and 2:1). Additionally, to showcase the versatility of our method beyond the CrowdCaption dataset, we incorporate custom examples for qualitative comparisons.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BeyondScene", + "bbox": [ + 642, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/468f81c0bbd2d152feaf101be98e08e6060e4b6f489a500eed9add25293d3c18.jpg", + "image_caption": [ + "Fig. 6: Examples of large scene synthesis $(4096\\times 4096)$ on the poses and text obtained from CrowdCaption [35] images. All baselines including direct high-resolution inference (SDXL [24]), T2I models for large scenes (MultiDiffusion [6], SyncDiffusion [14], ScaleCrafter [10]), Visual prior-guided T2I models (ControlNet [45], T2IAAdapter [23]) produce duplicated objects and artifacts in human anatomy, while our method succeeded in generation of high-resolution image with high text-image correspondence. Each color in the description corresponds to instances sharing the same color in the pose map." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics For an evaluation of text-image correspondence, we adopt a multimodal large language model [5, 17, 18]-based text-image correspondence metric, as proposed in VIEScore [13], alongside the global CLIP [25] score that shows its limitations in producing reliable scores for complex scenes with detailed descriptions due to limitations in input image size and the available number of input tokens. This MLLM-based metric, powered by GPT4 [5], assigns a score ranging from 0 to 10, where 0 indicates no correspondence between the generated image and the prompt and 10 indicates perfect alignment. This metric has been demonstrably well-correlated with human judgments and provides explanations for the assigned scores [13]. To assess the naturalness of the generated images, we adapt the existing MLLM-based naturalness metric to focus on human-centric", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ca1e15be944d6e55fd3e8b5cbe14ac255907627303cfe14e2af0704002cb750f.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of our method with various approaches, including direct high-resolution inference (T2I-Direct) [24,41], T2I models designed for large scenes (T2I-Large) [6,10,14], and Visual prior-guided T2I models (Visual+T2I) [6,23,45]." + ], + "table_footnote": [], + "table_body": "
Model typesModelsGlobal CLIPMLLM (GPT4)-based
Text-image correspondenceGlobal naturalnessHuman naturalness
2048×2048 (4×1:1)
T2I-DirectSDXL0.3242.1411.8381.417
RPG0.2611.9632.1472.785
T2I-LargeMultiDiffusion0.3473.8553.3812.466
SyncDiffusion0.3473.6553.4292.607
ScaleCrafter0.3456.0815.6675.062
Visual+T2IControlNet0.2981.7901.4661.117
T2IAdapter0.3283.0942.6601.728
R-MultiDiffusion0.3113.6361.7111.597
BeyondScene0.3057.0416.5356.114
4096×2048 (8×2:1)
T2I-DirectSDXL0.3011.0610.7710.607
RPGOOMOOMOOMOOM
T2I-LargeMultiDiffusion0.3382.1362.2281.329
SyncDiffusion0.3423.1282.6651.591
ScaleCrafter0.2963.5523.5693.359
Visual+T2IControlNet0.2781.5261.0160.657
T2IAdapter0.2821.1091.0100.568
R-MultiDiffusion0.3063.4661.4221.201
BeyondScene0.3047.1186.6126.375
4096×4096 (16×1:1)
T2I-DirectSDXL0.2770.3290.4110.118
RPGOOMOOMOOMOOM
T2I-LargeMultiDiffusion0.3191.3052.4991.591
SyncDiffusion0.3100.9122.3091.359
ScaleCrafter0.3122.9623.0482.442
Visual+T2IControlNet0.2440.8111.3090.729
T2IAdapter0.2690.4910.7190.352
R-MultiDiffusion0.2811.7920.8600.708
BeyondScene0.3016.8016.0745.627
", + "bbox": [ + 272, + 198, + 730, + 598 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "scenarios, leading to two separate scores: Global naturalness scores from 0 to 10, with 0 indicating an unnatural or unrealistic image and 10 indicating a completely natural image. This score considers factors such as overall inconsistencies, unrealistic physics, disconnectivity. Human naturalness scores from 0 to 10, with 0 indicating unnatural human anatomy, and 10 indicating the absence of any artifacts in the human anatomy. Additional details on implementation and evaluation are provided in the supplementary material.", + "bbox": [ + 212, + 626, + 787, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Result", + "text_level": 1, + "bbox": [ + 215, + 753, + 316, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our method is compared against three groups of existing SOTA approaches: 1) Direct high-resolution inference using T2I models (T2I-Direct) that directly perform high-resolution inference of SDXL [24] and RPG [41] over their training image size, 2) T2I diffusion models for large scenes (T2I-Large) that encom", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BeyondScene", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d626ce1f95af8ec578d524beef57652468aad03567cf809b85a8c71670f7c15e.jpg", + "table_caption": [ + "Table 2: Quantitative comparison by scene complexity. Our method shows robust performance (averaged across resolutions: $2048 \\times 2048$ , $4096 \\times 2048$ , $4096 \\times 4096$ ) even with increasing the number of humans, highlighting its ability to handle complex scenes." + ], + "table_footnote": [], + "table_body": "
Models2~4 humans5~8 humans
MLLM (GPT4)-basedMLLM (GPT4)-based
Text-image corr.Global nat.Human nat.Text-image corr.Global nat.Human nat.
SDXL1.1921.0390.8451.1610.9740.853
RPG3.7573.1383.0130.1691.1562.558
MultiDiffusion2.1422.3921.6922.7223.0131.911
SyncDiffusion2.2772.4281.6922.8533.1752.013
ScaleCrafter4.554.4374.0643.8463.7523.177
ControlNet1.1721.1940.8101.5791.3330.858
T2IAdapter1.5211.4490.9641.6081.4860.801
R-MultiDiffusion3.271.2861.3292.6591.3761.007
BeyondScene7.3086.4966.3766.6666.3195.701
", + "bbox": [ + 287, + 198, + 718, + 363 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/95f8a56712669a6f0bc14e87fa7ad77b359911d6837af37bfbabfc8d2c9d8688.jpg", + "table_caption": [ + "Table 3: User study result. Users consistently preferred our method over baselines, with MLLM scores closely aligning with human choices. However, the global CLIP score diverged, suggesting potential limitations." + ], + "table_footnote": [], + "table_body": "
ModelsPreference of Ours ↑
Text-image corr.Global Human nat.
vs ScaleCrafter0.7950.6500.806
vs MultiDiffusion0.7270.6580.822
vs ControlNet0.8700.6940.897
", + "bbox": [ + 362, + 430, + 635, + 522 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "pass joint diffusion-based methods (MultiDiffusion [6], SyncDiffusion [14]), and ScaleCrafter [10].) Visual prior-guided T2I generation models (Visual+T2I) that focuses on models including SDXL-ControlNet-Openpose [24, 45], SDXL-T2IAAdapter [23, 24], and R-MultiDiffusion [6] that leverage visual priors for T2I generation as our method. We provide more results and ablation studies in the supplementary material.", + "bbox": [ + 212, + 549, + 787, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative comparison As represented in Fig. 1, 5 and 6, baselines introduce noticeable instance duplication, artifacts in anatomy, facial features, making the scene unnatural rather than reflecting real-world physics. Also, they struggle to generate complex scenes corresponding to the detailed descriptions. Conversely, our method generates images with minimal artifacts, demonstrating superior text-image correspondence, global and human naturalness, particularly in human-centric scenes.", + "bbox": [ + 212, + 664, + 787, + 770 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative comparison Table 1 summarizes the performance across the resolution settings of $2048 \\times 2048$ ( $4 \\times 1:1$ ), $4096 \\times 2048$ ( $8 \\times 2:1$ ), and $4096 \\times 4096$ ( $16 \\times 1:1$ ). Our method consistently outperforms baselines on all metrics except", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ebab3973c207d74f9d1afe68c5ed99a9f6958dbeca1b2d54f1f89239430f1760.jpg", + "image_caption": [ + "(a) Original image" + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 330, + 208 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b3d88ff5f10bd423587f714d6bb84bee38d6a8cd37c41ad123785f57122bc3c7.jpg", + "image_caption": [ + "(b) 64x down-sampled" + ], + "image_footnote": [], + "bbox": [ + 331, + 143, + 416, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e5001ee46066886bbce1f62b5155721853b98e2c1354a5749ab31ba409beca21.jpg", + "image_caption": [ + "mage (L) / Its part (R) (c) Bilinear interpol." + ], + "image_footnote": [], + "bbox": [ + 418, + 143, + 501, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/42ebeae8fa2d34185d3c86930c12f5144be7a333cdd8a0bd5861f9dd74f675fe.jpg", + "image_caption": [ + "(d) Naive PP" + ], + "image_footnote": [], + "bbox": [ + 503, + 143, + 586, + 207 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(c) Bilinear interpol.", + "bbox": [ + 504, + 210, + 581, + 219 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/396b37184c12fbc7c81b30616c51b7554ffa503706e18b8743ec179175149ea8.jpg", + "image_caption": [ + "(e) Adaptive PP (Ours)" + ], + "image_footnote": [], + "bbox": [ + 586, + 143, + 668, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/600c8abbe8eece1ed5eaf4d03cb651d1cca7198b0623fe7622544ac789be9296.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 143, + 754, + 208 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0e014212246c6b20a3e3acb2149d7c5041f825f3709d630858cb6e32b04c0efd.jpg", + "image_caption": [ + "Fig. 7: Forward-reverse diffusion process with interpolation methods. (a) SDXL generates an image with the prompt \"figurine of Superman\". (b) The image is then $\\times 64$ down-sampled (Left) with a close-up view (Right). (c) Upsampled image with bilinear interpolation and forward-reverse diffusion exhibits blurry details. (d) Naive pixel perturbation (PP) yields sharper details but introduces flickering artifacts. (e) Our Canny map-based adaptive PP produces clear and highly detailed image.", + "Fig. 8: Effectiveness of adaptive conditioning. Omitting either adaptive pose or text conditioning leads to issues like inconsistent poses, bad anatomy (yellow boxes), unwanted object duplication (red boxes), and mixed descriptions. Each color in the description corresponds to instances sharing the same color in the pose map." + ], + "image_footnote": [], + "bbox": [ + 243, + 316, + 759, + 619 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the global CLIP score, which shows inconsistencies across methods. This suggests that our approach effectively enhances the original generation capabilities of the pretrained diffusion model for high-resolution image generation, achieving superior text-image correspondence, naturalness.", + "bbox": [ + 212, + 727, + 787, + 789 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3 further analyzes the performance of different methods based on the number of people depicted in the scene. Our approach demonstrates robustness, effectively managing complexity as people count rises.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BeyondScene", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "User study We conducted a user study to evaluate the faithfulness and naturalness of images generated by three methods (one from each group) and our method. Participants compared large-scene images based on the same criteria used in the MLLM metrics: text-image correspondence, global naturalness, and human naturalness. A total of 12,120 responses were collected from 101 participants. The result shows that our method was consistently preferred over baselines, with a larger margin for all attributes. Notably, MLLM scores demonstrate strong alignment with human scores, supporting their validity. However, the global CLIP score appears misaligned, suggesting its limitations in capturing human judgment. We provide the additional details in the supplementary material.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Beyond 8K ultra-high-resolution image As displayed in Fig. 2, BeyondScene excels in generating unparalleled resolution, producing images at an impressive $8192 \\times 8192$ dimensions. This achievement surpasses the training resolution of SDXL by a remarkable factor of $64 \\times$ . It surpasses the standard 8K resolution $(7680 \\times 4320)$ , demonstrating the model's capability to produce images with exceptional detail at even higher resolutions.", + "bbox": [ + 212, + 316, + 787, + 409 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 428, + 388, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "High frequency-injected forward diffusion As represented in 7, forward diffusion and reverse generation with naively up-sampled image lead to increased blurring in the image. By adopting the pixel perturbation, it makes the result sharpened. However, it introduced flickering artifacts at the image borders. Applying adaptive pixel perturbation based on the Canny map significantly improves image quality and reduces flickering artifacts.", + "bbox": [ + 212, + 450, + 787, + 542 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Adaptive conditioning Fig. 8 demonstrates the effectiveness of adaptive conditioning. Omitting adaptive pose conditioning leads to inconsistent poses, bad anatomy, and decreased controllability compared to the desired input pose (yellow boxes). Excluding adaptive text conditioning, where each view receives the same full text including global and all instance descriptions, results in unwanted object duplication (red boxes) that ignores existing object scales, creating an awkward scene. Additionally, the descriptions of each instance become mixed, resulting in low text-image correspondence and decreased controllability.", + "bbox": [ + 212, + 561, + 787, + 684 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 704, + 359, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce BeyondScene, a novel framework capable of generating high-resolution human-centric scenes with exceptional text-image correspondence resolving artifacts beyond the token limits and beyond the training image size. BeyondScene overcomes limitations in generating complex human-centric scenes with detailed control, surpassing existing methods in generating high-fidelity scenes with control over human characteristics and number, while maintaining naturalness and reducing artifacts.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) [NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)] and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF-2022R1A4A1030579, NRF-2022M3C1A309202211). Also, the authors acknowledged the financial support from the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University.", + "bbox": [ + 212, + 174, + 787, + 295 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 316, + 321, + 333 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Language segment anything. https://github.com/luca-medeiros/lang-segment-anything", + "2. Midjourney. https://www.midjourney.com", + "3. Sdsl-controlnet: Openpose (v2). https://huggingface.co/thibaud/controlnet-openpose-sdsl-1.0", + "4. Sdxl inpainting 0.1. https://huggingface.co/diffusers/stable-diffusion-x1-1.0-inpainting-0.1", + "5. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "6. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv:2302.08113 (2023)", + "7. Chen, T.: On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972 (2023)", + "8. Cheong, S.Y., Mustafa, A., Gilbert, A.: KPE: Keypoint pose encoding for transformer-based image generation. In: British Machine Vision Conference (BMVC) (2022)", + "9. Ding, Z., Zhang, M., Wu, J., Tu, Z.: Patched denoising diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2023)", + "0. He, Y., Yang, S., Chen, H., Cun, X., Xia, M., Zhang, Y., Wang, X., He, R., Chen, Q., Shan, Y.: Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In: The Twelfth International Conference on Learning Representations (2023)", + "1. Hoogeboom, E., Heek, J., Salimans, T.: simple diffusion: End-to-end diffusion for high resolution images. arXiv preprint arXiv:2301.11093 (2023)", + "2. Ju, X., Zeng, A., Zhao, C., Wang, J., Zhang, L., Xu, Q.: Humansd: A native skeleton-guided diffusion model for human image generation. arXiv preprint arXiv:2304.04269 (2023)", + "3. Ku, M., Jiang, D., Wei, C., Yue, X., Chen, W.: Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867 (2023)", + "4. Lee, Y., Kim, K., Kim, H., Sung, M.: Syncdiffusion: Coherent montage via synchronized joint diffusions. arXiv:2306.05178 (2023)", + "5. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: Openset grounded text-to-image generation. arXiv preprint arXiv:2301.07093 (2023)" + ], + "bbox": [ + 225, + 347, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BeyondScene", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Liu, D., Wu, L., Zheng, F., Liu, L., Wang, M.: Verbal-person nets: Pose-guided multi-granularity language-to-person generation. IEEE Transactions on Neural Networks and Learning Systems (2022)", + "17. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)", + "18. Liu, S., Cheng, H., Liu, H., Zhang, H., Li, F., Ren, T., Zou, X., Yang, J., Su, H., Zhu, J., et al.: Llava-plus: Learning to use tools for creating multimodal agents. arXiv preprint arXiv:2311.05437 (2023)", + "19. Lv, Z., Li, X., Li, X., Li, F., Lin, T., He, D., Zuo, W.: Learning semantic person image generation by region-adaptive normalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10806-10815 (2021)", + "20. Ma, L., Jia, X., Sun, Q., Schiele, B., Tuytelaars, T., Van Gool, L.: Pose guided person image generation. vol. 30 (2017)", + "21. Ma, T., Peng, B., Wang, W., Dong, J.: MUST-GAN: Multi-level statistics transfer for self-driven person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13622-13631 (2021)", + "22. Men, Y., Mao, Y., Jiang, Y., Ma, W.Y., Lian, Z.: Controllable person image synthesis with attribute-decomposed GAN. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5084-5093 (2020)", + "23. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)", + "24. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "25. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021)", + "26. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)", + "27. Ren, Y., Fan, X., Li, G., Liu, S., Li, T.H.: Neural texture extraction and distribution for controllable person image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13535-13544 (2022)", + "28. Ren, Y., Yu, X., Chen, J., Li, T.H., Li, G.: Deep image spatial transformation for person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7690-7699 (2020)", + "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)", + "30. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)", + "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)", + "32. Roy, P., Ghosh, S., Bhattacharya, S., Pal, U., Blumenstein, M.: TIPS: Text-induced pose synthesis. In: European Conference on Computer Vision (ECCV). pp. 161-178. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "34. Teng, J., Zheng, W., Ding, M., Hong, W., Wangni, J., Yang, Z., Tang, J.: Relay diffusion: Unifying diffusion process across resolutions for image synthesis. arXiv preprint arXiv:2309.03350 (2023)", + "35. Wang, L., Li, H., Hu, W., Zhang, X., Qiu, H., Meng, F., Wu, Q.: What happens in crowd scenes: A new dataset about crowd scenes for image captioning. IEEE Transactions on Multimedia (2022)", + "36. Wu, J., Wang, J., Yang, Z., Gan, Z., Liu, Z., Yuan, J., Wang, L.: Grit: A generative region-to-text transformer for object understanding. arXiv:2212.00280 (2022)", + "37. Xie, E., Yao, L., Shi, H., Liu, Z., Zhou, D., Liu, Z., Li, J., Li, Z.: Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. arXiv preprint arXiv:2304.06648 (2023)", + "38. Xu, X., Chen, Y.C., Tao, X., Jia, J.: Text-guided human image manipulation via image-text shared space. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) 44(10), 6486-6500 (2021)", + "39. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. NeurIPS 35, 38571-38584 (2022)", + "40. Yang, F., Lin, G.: CT-Net: Complementary transferring network for garment transfer with arbitrary geometric changes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9899-9908 (2021)", + "41. Yang, L., Yu, Z., Meng, C., Xu, M., Ermon, S., Cui, B.: Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. arXiv preprint arXiv:2401.11708 (2024)", + "42. Yang, L., Wang, P., Liu, C., Gao, Z., Ren, P., Zhang, X., Wang, S., Ma, S., Hua, X., Gao, W.: Towards fine-grained human pose transfer with detail replenishing network. IEEE Transactions on Image Processing 30, 2422-2435 (2021)", + "43. Zhang, B., Gu, S., Zhang, B., Bao, J., Chen, D., Wen, F., Wang, Y., Guo, B.: Styleswin: Transformer-based gan for high-resolution image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11304-11314 (2022)", + "44. Zhang, J., Li, K., Lai, Y.K., Yang, J.: PISE: Person image synthesis and editing with decoupled GAN. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7982-7990 (2021)", + "45. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)", + "46. Zhang, P., Yang, L., Lai, J.H., Xie, X.: Exploring dual-task correlation for pose guided person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7713-7722 (2022)", + "47. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. arXiv:2303.17076 (2023)", + "48. Zheng, Q., Guo, Y., Deng, J., Han, J., Li, Y., Xu, S., Xu, H.: Any-size-diffusion: Toward efficient text-driven synthesis for any-size hd images. arXiv preprint arXiv:2308.16582 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 784 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BeyondScene", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_model.json b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e0f1b7e4d8e96cc4a0b3d495d1c91db43fbc755 --- /dev/null +++ b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_model.json @@ -0,0 +1,2049 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.224, + 0.141, + 0.78, + 0.185 + ], + "angle": 0, + "content": "BeyondScene: Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.212, + 0.765, + 0.245 + ], + "angle": 0, + "content": "Gwanghyun Kim\\(^{1\\star}\\), Hayeon Kim\\(^{1\\star}\\), Hoigi Seo\\(^{1\\star}\\), Dong Un Kang\\(^{1\\star}\\), Se Young Chun\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.255, + 0.714, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\)Dept. of Electrical and Computer Engineering, \\(^{2}\\)INMC & IPAI Seoul National University, Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.284, + 0.751, + 0.298 + ], + "angle": 0, + "content": "{gwang.kim, khy5630, sehoiki3215, qkrtnsfk23, sychun}@snu.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.333, + 0.744, + 0.653 + ], + "angle": 0, + "content": "Abstract. Generating higher-resolution human-centric scenes with details and controls remains a challenge for existing text-to-image diffusion models. This challenge stems from limited training image size, text encoder capacity (limited tokens), and the inherent difficulty of generating complex scenes involving multiple humans. While current methods attempted to address training size limit only, they often yielded human-centric scenes with severe artifacts. We propose BeyondScene, a novel framework that overcomes prior limitations, generating exquisite higher-resolution (over 8K) human-centric scenes with exceptional text-image correspondence and naturalness using existing pretrained diffusion models. BeyondScene employs a staged and hierarchical approach to initially generate a detailed base image focusing on crucial elements in instance creation for multiple humans and detailed descriptions beyond token limit of diffusion model, and then to seamlessly convert the base image to a higher-resolution output, exceeding training image size and incorporating details aware of text and instances via our novel instance-aware hierarchical enlargement process that consists of our proposed high-frequency injected forward diffusion and adaptive joint diffusion. BeyondScene surpasses existing methods in terms of correspondence with detailed text descriptions and naturalness, paving the way for advanced applications in higher-resolution human-centric scene creation beyond the capacity of pretrained diffusion models without costly retraining. Project page: https://janeyon.github.io/beyond-scene." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.665, + 0.741, + 0.693 + ], + "angle": 0, + "content": "Keywords: Human-centric scene generation \\(\\cdot\\) Text-to-image diffusion model \\(\\cdot\\) High-resolution" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.377, + 0.756 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Human-centric scene generation [8, 12, 15, 16, 19-23, 27, 28, 32, 33, 38, 40, 42, 44-46], has emerged as a critical research area with significant academic and industrial applications. While recent advancements utilizing text-to-image (T2I) diffusion" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.597, + 0.841 + ], + "angle": 0, + "content": "* Authors contributed equally. † Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.145, + 0.773, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.245, + 0.522, + 0.758, + 0.551 + ], + "angle": 0, + "content": "\"In the sunny road, there are a clay animation character with lab coat, a clay character with red dotted dress and glasses, a clay character with short, chubby cheeks and glasses, a clay animation character with short, chubby checks wearing blue jeans, a clay character with spiky hair, a clay character with yellow T-shirts.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Fig. 1: BeyondScene pushes the boundaries of high-resolution human-centric scene generation. Unlike existing methods that often suffer from unrealistic scenes, anatomical distortions, and limited text-to-image correspondence, BeyondScene excels in 1) highly detailed scenes, 2) natural and diverse humans, 3) fine-grained control. The color in each text represents the text for each instance that has the same color in the pose map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.84 + ], + "angle": 0, + "content": "models [2, 26, 29] have yielded promising results in generating controllable human scenes [12, 15, 23, 45], scaling these methods to handle larger and more complex scenes such as involving multiple humans remains a significant challenge. These limitations primarily stem from following causes. 1) Training image size: Directly sampling from a limited training image size can introduce artifacts and constrain the final scene resolution. 2) Limited text encoder tokens: The restricted number of tokens in text encoders (typically 77 for Stable Diffusion) hinders the inclusion of detailed descriptions for multiple instances within the scene. 3) Inherent limitations of T2I diffusion models: These models struggle to generate complex scenes with several human figures and intricate details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BeyondScene" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Existing attempts to address the training size issue, such as joint-process [6, 14, 47] and dilation-based methods [10], only partially address the problem. These methods often introduce new challenges specific to human-centric scenes as represented in Fig. 1, including: 1) Unrealistic scenes and objects: These methods can generate nonsensical scenarios with duplicated objects, humans defying gravity, and physically distorted environments. 2) Anatomical distortions: Generated scenes may exhibit unrealistic human anatomy, such as abnormal limbs or facial features. 3) Limited correspondence: Existing methods often fail to capture the complexity of scenes with multiple human instances and detailed descriptions, often generating single objects and lacking control over specific details like clothing or hairstyles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.79, + 0.41 + ], + "angle": 0, + "content": "We propose BeyondScene, a novel framework that overcomes these limitations, generating high-resolution (over 8K) human-centric scenes with exceptional text-image correspondence and naturalness. BeyondScene employs a staged and hierarchical approach, which looks similar to a classical multi-resolution manner, but it is closer to how artists establish a foundation before adding details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.79, + 0.701 + ], + "angle": 0, + "content": "Firstly, BeyondScene generates a detailed base image focusing on essential elements, human poses and detailed descriptions beyond the token limits. This initial stage enables detailed instance creation for multiple humans, surpassing the limitations of text encoders. Secondly, the method leverages an instance-aware hierarchical enlargement process to convert this base image to a higher resolution output beyond the training image size. Unlike naive super-resolution methods that simply scale resolution without considering text and instances, our approach refines content and adds more details aware of text and instances. This is achieved through our proposed novel techniques, including 1) high frequency-injected forward diffusion that addresses the issue of blurred low-quality results in image-to-image translation with the upsampled image by adaptively injecting high-frequency details into the upsampled image, enhancing the final output while preserving content and 2) adaptive joint diffusion that facilitates efficient and robust joint diffusion while maintaining control over human characteristics like pose and mask. This approach utilizes view-wise conditioning of text and pose information, along with a variable stride for the joint process. Furthermore, similar to how artists add details progressively, our method allows for the addition of details at different stages, exploiting the changing receptive field of each view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We comprehensively evaluate BeyondScene using qualitative and quantitative metrics, along with user studies. The results demonstrate significant improvements over existing methods in terms of 1) correspondence with detailed text descriptions and 2) naturalness and reduction of artifacts. Furthermore, we showcase the result of \\(8192 \\times 8192\\) image generation beyond 8K ultra high-resolution, demonstrating the capability for generating even higher-resolution images as displayed in Fig. 2. These advancements pave the way for exciting new applications in higher-resolution human-centric scene creation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.785, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.594, + 0.788, + 0.636 + ], + "angle": 0, + "content": "Fig. 2: Beyond 8K ultra-high resolution image. This \\(8192 \\times 8192\\) image, generated by BeyondScene, surpasses the training resolution of SDXL by \\(64 \\times\\), while exceeding the technical classification of 8K (\\(7680 \\times 4320\\))." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.664, + 0.383, + 0.68 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.695, + 0.528, + 0.71 + ], + "angle": 0, + "content": "2.1 Controllable Human Generation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.719, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Early approaches to controllable human generation [8, 19-22, 27, 28, 40, 42, 44, 46] relied on pose guidance and source images, achieving success within specific scenarios but struggling with diverse scenes and arbitrary poses. Text-based conditioning methods emerged with the rise of large vision-language models [25], but initial attempts [16, 32, 38] faced limitations in vocabulary size and open vocabulary settings. Recent T2I diffusion [2, 24, 26, 30]-based methods like ControlNet [45], T2I-Adapter [23], GLIGEN [15], and HumanSD [12] introduced methods for incorporating diverse conditions, but scaling to larger scenes with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BeyondScene" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "multiple individuals remains a challenge. This work proposes BeyondScene, a novel framework specifically designed to overcome these limitations and enable the generation of high-resolution human-centric scenes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.214, + 0.659, + 0.23 + ], + "angle": 0, + "content": "2.2 Large Scene Generation Using Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.79, + 0.406 + ], + "angle": 0, + "content": "Achieving high-resolution image generation [9, 26] presents significant hurdles. Training models from scratch or fine-tuning pretrained models [2, 7, 11, 24, 26, 30, 31, 34, 37, 43, 48] requires immense computational resources and struggles with the complexity of high-dimensional data. Recent exploration has ventured into training-free methods, with approaches like MultiDiffusion [6], SyncDiffusion [14] (joint diffusion), and ScaleCrafter [10] (dilation-based) emerging. However, these methods often introduce challenges specific to human-centric scenes, including unrealistic objects, anatomical distortions, and limited correspondence between text descriptions and generated images. BeyondScene, our proposed framework, addresses these limitations through detailed base image generation and instance-aware hierarchical enlargement." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.429, + 0.382, + 0.447 + ], + "angle": 0, + "content": "3 BeyondScene" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Current approaches to generating large scenes attempt to fill the entire canvas at once, often leading to challenges in capturing complex details and ensuring physical realism. Inspired by the artistic workflow of human painters, who establish a foundation with key elements and progressively refine them, we introduce BeyondScene, a novel framework that generates high-resolution human-centric images through a staged and hierarchical approach as represented in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.627 + ], + "angle": 0, + "content": "BeyondScene operates in two key stages, 1) detailed base image generation (Sec. 3.1) that focuses on generating high-quality base images with precise control over human poses and detailed descriptions beyond typical token limitations and 2) instance-aware hierarchical enlargement (Sec. 3.2) where the base image is progressively enlarged while maintaining control and adding scene details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.649, + 0.533, + 0.665 + ], + "angle": 0, + "content": "3.1 Detailed Base Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Generating human-centric scenes with multiple individuals using T2I diffusion models faces a key challenge: the limited number of tokens available in the text encoder restricts detailed descriptions to a few key elements. To overcome this, we generate detailed base image as illustrated in Fig. 3. Firstly, individual instances are generated in the training resolution of pose-guided T2I diffusion models [12, 15, 23, 45]. This allows for detailed descriptions beyond token limitations, tailored to each individual. Subsequently, the generated instances are segmented, cropped, and resized to the desired size. Next, a diffusion-based inpainting method is employed to create an instance-aware background that seamlessly integrates with the individual instances. This ensures a realistic and coherent background tailored to the specific scene. To address potential inconsistencies in brightness" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.128 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.146, + 0.77, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Fig. 3: BeyondScene generates high-resolution images in two stages. First, individual instances are created using pose-guided T2I diffusion models, segmented, cropped, and placed onto an inpainted background. The tone of image is then normalized. In the second stage (illustrated in Fig. 4), this base image is progressively enlarged while maintaining detail and quality, effectively refining the image and adding further details leveraging high-frequency injected forward diffusion and adaptive joint diffusion (AJ)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.689 + ], + "angle": 0, + "content": "and color tones across individual elements, tone normalization through contrast limited adaptive histogram equalization (CLAHE) is performed. This involves first creating a grid histogram for pixel values within a specified grid, followed by adjusting the local contrast to align with the image's overall contrast. Notably, direct equalization in the presence of noise within the grid could induce the noise amplification. To prevent this, equalization is preceded by a redistribution process that includes limiting the maximum height of the grid histogram." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.607, + 0.726 + ], + "angle": 0, + "content": "3.2 Instance-Aware Hierarchical Enlargement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.795 + ], + "angle": 0, + "content": "BeyondScene introduces a novel approach to address the challenge of generating high-resolution outputs beyond the training image size through a instance-aware hierarchical enlargement process, which surpasses mere scaling by actively refining content and incorporating additional details aware of text and instance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In the initial step, a low-resolution image is upsampled, and a forward diffusion process is applied to create a noisy upsampled image. Subsequently, a joint diffusion process refines the image to a high resolution over the training resolution." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BeyondScene" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.149, + 0.538, + 0.162 + ], + "angle": 0, + "content": "3.2 Instance-Aware Hierarchical Enlargement" + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.163, + 0.483, + 0.175 + ], + "angle": 0, + "content": "a. High frequency-injected forward diffusion" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.177, + 0.739, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.305, + 0.506, + 0.317 + ], + "angle": 0, + "content": "b. Adaptive joint diffusion (AJ) At the diffusion timestep \\(t\\)" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.319, + 0.764, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.512 + ], + "angle": 0, + "content": "Fig. 4: Our instance-aware hierarchical enlargement involves two crucial processes: 1) High frequency-injected forward diffusion, which enables to achieve high resolution through a joint diffusion employing adaptive pixel perturbation. 2) Adaptive joint diffusion, dynamically regulating stride and conditioning of pose and text based on the presence of instances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.65 + ], + "angle": 0, + "content": "However, a drawback of this naive strategy is the induction of blurriness due to the model assigning sufficiently high likelihood to low-quality outputs. Also, this strategy often encounters issues such as unnatural duplication of objects and humans, stemming from the uniform application of the same text prompt to all views without pose conditioning. To mitigate these, the paper proposes a novel high frequency-injected forward diffusion and adaptive joint diffusion as illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "High frequency-injected forward diffusion High frequency-injected forward diffusion enhances the translation of noisy latents from the upsampled image to high resolution with intricate details. This is achieved through a joint diffusion process employing adaptive pixel perturbation. This technique injects high-frequency details into the upsampled image based on a Canny edge map. This process enhances the final output's quality while preserving content. While simply adding pixel noise sharpens the image, it can also introduce flickering artifacts at the borders. Our adaptive approach based on the Canny map significantly improves image quality and reduces these artifacts. Specifically, we first up-sample the low-resolution image and calculate Canny map. Then Gaussian blur is applied to the map and normalized it to get probability map \\(C\\). \\(\\epsilon_{i,j} \\sim \\mathcal{U}(0,1)\\) is sampled" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "on each pixel, \\( I_{i,j} \\) which denotes \\( i \\)-th row and \\( j \\)-th column pixel of image \\( I \\). If \\( \\epsilon_{i,j} > C_{i,j} \\), the pixel value within \\( d_r \\) pixel distance is replaced so that we could apply pixel perturbation adaptively to avoid flickering artifacts near borders." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.214, + 0.79, + 0.246 + ], + "angle": 0, + "content": "Adaptive joint diffusion To mitigate duplication of objects and humans in the naive joint diffusion, we introduce adaptive conditioning and adaptive stride." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.79, + 0.395 + ], + "angle": 0, + "content": "Adaptive conditioning We introduce adaptive view-wise conditioning, which leverages the segmentation maps obtained from the base image generation stage. For each view, we check which human instances are present. If an instance is included, we add its pose and detailed description to the prompt used for that view. Additionally, if the detailed text description includes specific parts (head, face, upper body, etc.), we can apply these details to the corresponding regions using fine-grained segmentation maps. This facilitates efficient and robust joint diffusion while maintaining control over human characteristics like pose and appearance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.789, + 0.531 + ], + "angle": 0, + "content": "Adaptive stride To effectively enhance the quality of the generated scene, we employ an adaptive stride for the joint process. We reduce the stride in regions containing humans, ensuring that these areas capture fine details. In contrast, the stride is increased in background regions, allowing for more efficient computation in these areas. This combination of adaptive view-wise conditioning and adaptive stride allows BeyondScene to effectively handle scenes with multiple humans while maintaining high-resolution detail and controllable image generation. The details and algorithms on our methods are presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.552, + 0.377, + 0.57 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.583, + 0.448, + 0.599 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.789, + 0.73 + ], + "angle": 0, + "content": "We comprehensively evaluate our method using both qualitative and quantitative metrics. To ensure a fair comparison, all compared models are implemented using Stable Diffusion XL (SDXL) [24]-based architectures. We utilize SDXL-ControlNet-Openpose [3, 45], SDXL-inpainting [4], and Lang-Segment-Anything [1] for our pose-guided T2I diffusion, inpainting diffusion model, and segmentation models. The training resolution is set to \\(1024 \\times 1024\\), while inference resolutions are varied across \\(2048 \\times 2048\\) (\\(4 \\times 1:1\\)), \\(3584 \\times 2048\\) (\\(7 \\times 7:4\\)), \\(4096 \\times 2048\\) (\\(8 \\times 2:1\\)), \\(4096 \\times 4096\\) (\\(16 \\times 1:1\\)), and \\(8192 \\times 8192\\) (\\(64 \\times 1:1\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Testing data Existing T2I generation datasets often lack large-scale scenes featuring multiple people with detailed descriptions for each individual. To address this limitation, we leverage the CrowdCaption dataset [35] as our primary test set. This dataset features images containing large crowds of people, allowing us to evaluate the effectiveness of our method in generating complex scenes with numerous individuals. For each image in the CrowdCaption dataset, we obtain" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "BeyondScene" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.146, + 0.77, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.245, + 0.342, + 0.764, + 0.369 + ], + "angle": 0, + "content": "\"In empty stage of opera house, there are a ballerina with pink suit, a ballerina wearing blue ballet suit, a ballerina with pink suit, a ballerina with blue suit, a ballerina with yellow sparkling ballet suit, a ballerina with blue suit, a ballerina with pink suit, a ballerina with blue suit, a ballerina with pink suit.\"" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.373, + 0.769, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.245, + 0.568, + 0.759, + 0.587 + ], + "angle": 0, + "content": "\"In the alien planet with galaxy in the sky, there are a storm trooper, a storm trooper, a storm trooper, a Darth Vader, a storm trooper, a storm trooper, a storm trooper.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison for generating high-resolution human scenes (\\(3584 \\times 2048\\)). While existing approaches like T2I-Direct (SDXL [24]), T2I-Large (MultiDiffusion [6], SyncDiffusion [14], and ScaleCrafter [10]), and Visual+T2I (ControlNet [24, 45], T2IAdapter [23, 24], and R-MultiDiffusion [6]) models struggle with artifacts, our method achieves superior results by producing images with minimal artifacts, strong text-image correspondence, and a natural look. The color in each description represents the description for each instance that has the same color in the pose map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "descriptions and individual poses using GRIT [36] and ViTPose [39]. Additionally, we obtain a global caption generated using GPT4 [5]. We filter the CrowdCaption images based on two criteria: number of humans (max 8 to accommodate token limits of baselines) and aspect ratio (close to 1:1 and 2:1), which results in a collection of 100 images for each aspect ratio (1:1 and 2:1). Additionally, to showcase the versatility of our method beyond the CrowdCaption dataset, we incorporate custom examples for qualitative comparisons." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.773, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Fig. 6: Examples of large scene synthesis \\((4096\\times 4096)\\) on the poses and text obtained from CrowdCaption [35] images. All baselines including direct high-resolution inference (SDXL [24]), T2I models for large scenes (MultiDiffusion [6], SyncDiffusion [14], ScaleCrafter [10]), Visual prior-guided T2I models (ControlNet [45], T2IAAdapter [23]) produce duplicated objects and artifacts in human anatomy, while our method succeeded in generation of high-resolution image with high text-image correspondence. Each color in the description corresponds to instances sharing the same color in the pose map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Evaluation metrics For an evaluation of text-image correspondence, we adopt a multimodal large language model [5, 17, 18]-based text-image correspondence metric, as proposed in VIEScore [13], alongside the global CLIP [25] score that shows its limitations in producing reliable scores for complex scenes with detailed descriptions due to limitations in input image size and the available number of input tokens. This MLLM-based metric, powered by GPT4 [5], assigns a score ranging from 0 to 10, where 0 indicates no correspondence between the generated image and the prompt and 10 indicates perfect alignment. This metric has been demonstrably well-correlated with human judgments and provides explanations for the assigned scores [13]. To assess the naturalness of the generated images, we adapt the existing MLLM-based naturalness metric to focus on human-centric" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BeyondScene" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of our method with various approaches, including direct high-resolution inference (T2I-Direct) [24,41], T2I models designed for large scenes (T2I-Large) [6,10,14], and Visual prior-guided T2I models (Visual+T2I) [6,23,45]." + }, + { + "type": "table", + "bbox": [ + 0.273, + 0.199, + 0.732, + 0.599 + ], + "angle": 0, + "content": "
Model typesModelsGlobal CLIPMLLM (GPT4)-based
Text-image correspondenceGlobal naturalnessHuman naturalness
2048×2048 (4×1:1)
T2I-DirectSDXL0.3242.1411.8381.417
RPG0.2611.9632.1472.785
T2I-LargeMultiDiffusion0.3473.8553.3812.466
SyncDiffusion0.3473.6553.4292.607
ScaleCrafter0.3456.0815.6675.062
Visual+T2IControlNet0.2981.7901.4661.117
T2IAdapter0.3283.0942.6601.728
R-MultiDiffusion0.3113.6361.7111.597
BeyondScene0.3057.0416.5356.114
4096×2048 (8×2:1)
T2I-DirectSDXL0.3011.0610.7710.607
RPGOOMOOMOOMOOM
T2I-LargeMultiDiffusion0.3382.1362.2281.329
SyncDiffusion0.3423.1282.6651.591
ScaleCrafter0.2963.5523.5693.359
Visual+T2IControlNet0.2781.5261.0160.657
T2IAdapter0.2821.1091.0100.568
R-MultiDiffusion0.3063.4661.4221.201
BeyondScene0.3047.1186.6126.375
4096×4096 (16×1:1)
T2I-DirectSDXL0.2770.3290.4110.118
RPGOOMOOMOOMOOM
T2I-LargeMultiDiffusion0.3191.3052.4991.591
SyncDiffusion0.3100.9122.3091.359
ScaleCrafter0.3122.9623.0482.442
Visual+T2IControlNet0.2440.8111.3090.729
T2IAdapter0.2690.4910.7190.352
R-MultiDiffusion0.2811.7920.8600.708
BeyondScene0.3016.8016.0745.627
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.734 + ], + "angle": 0, + "content": "scenarios, leading to two separate scores: Global naturalness scores from 0 to 10, with 0 indicating an unnatural or unrealistic image and 10 indicating a completely natural image. This score considers factors such as overall inconsistencies, unrealistic physics, disconnectivity. Human naturalness scores from 0 to 10, with 0 indicating unnatural human anatomy, and 10 indicating the absence of any artifacts in the human anatomy. Additional details on implementation and evaluation are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.317, + 0.769 + ], + "angle": 0, + "content": "4.2 Result" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Our method is compared against three groups of existing SOTA approaches: 1) Direct high-resolution inference using T2I models (T2I-Direct) that directly perform high-resolution inference of SDXL [24] and RPG [41] over their training image size, 2) T2I diffusion models for large scenes (T2I-Large) that encom" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison by scene complexity. Our method shows robust performance (averaged across resolutions: \\(2048 \\times 2048\\), \\(4096 \\times 2048\\), \\(4096 \\times 4096\\)) even with increasing the number of humans, highlighting its ability to handle complex scenes." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.199, + 0.719, + 0.364 + ], + "angle": 0, + "content": "
Models2~4 humans5~8 humans
MLLM (GPT4)-basedMLLM (GPT4)-based
Text-image corr.Global nat.Human nat.Text-image corr.Global nat.Human nat.
SDXL1.1921.0390.8451.1610.9740.853
RPG3.7573.1383.0130.1691.1562.558
MultiDiffusion2.1422.3921.6922.7223.0131.911
SyncDiffusion2.2772.4281.6922.8533.1752.013
ScaleCrafter4.554.4374.0643.8463.7523.177
ControlNet1.1721.1940.8101.5791.3330.858
T2IAdapter1.5211.4490.9641.6081.4860.801
R-MultiDiffusion3.271.2861.3292.6591.3761.007
BeyondScene7.3086.4966.3766.6666.3195.701
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.377, + 0.789, + 0.42 + ], + "angle": 0, + "content": "Table 3: User study result. Users consistently preferred our method over baselines, with MLLM scores closely aligning with human choices. However, the global CLIP score diverged, suggesting potential limitations." + }, + { + "type": "table", + "bbox": [ + 0.364, + 0.431, + 0.637, + 0.523 + ], + "angle": 0, + "content": "
ModelsPreference of Ours ↑
Text-image corr.Global Human nat.
vs ScaleCrafter0.7950.6500.806
vs MultiDiffusion0.7270.6580.822
vs ControlNet0.8700.6940.897
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.643 + ], + "angle": 0, + "content": "pass joint diffusion-based methods (MultiDiffusion [6], SyncDiffusion [14]), and ScaleCrafter [10].) Visual prior-guided T2I generation models (Visual+T2I) that focuses on models including SDXL-ControlNet-Openpose [24, 45], SDXL-T2IAAdapter [23, 24], and R-MultiDiffusion [6] that leverage visual priors for T2I generation as our method. We provide more results and ablation studies in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Qualitative comparison As represented in Fig. 1, 5 and 6, baselines introduce noticeable instance duplication, artifacts in anatomy, facial features, making the scene unnatural rather than reflecting real-world physics. Also, they struggle to generate complex scenes corresponding to the detailed descriptions. Conversely, our method generates images with minimal artifacts, demonstrating superior text-image correspondence, global and human naturalness, particularly in human-centric scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Quantitative comparison Table 1 summarizes the performance across the resolution settings of \\(2048 \\times 2048\\) (\\(4 \\times 1:1\\)), \\(4096 \\times 2048\\) (\\(8 \\times 2:1\\)), and \\(4096 \\times 4096\\) (\\(16 \\times 1:1\\)). Our method consistently outperforms baselines on all metrics except" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BeyondScene" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.144, + 0.331, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.211, + 0.326, + 0.22 + ], + "angle": 0, + "content": "(a) Original image" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.144, + 0.417, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.211, + 0.415, + 0.22 + ], + "angle": 0, + "content": "(b) 64x down-sampled" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.144, + 0.502, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.211, + 0.582, + 0.22 + ], + "angle": 0, + "content": "mage (L) / Its part (R) (c) Bilinear interpol." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.144, + 0.587, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.211, + 0.582, + 0.22 + ], + "angle": 0, + "content": "(c) Bilinear interpol." + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.144, + 0.669, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.211, + 0.648, + 0.22 + ], + "angle": 0, + "content": "(d) Naive PP" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.144, + 0.755, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.211, + 0.754, + 0.22 + ], + "angle": 0, + "content": "(e) Adaptive PP (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.316 + ], + "angle": 0, + "content": "Fig. 7: Forward-reverse diffusion process with interpolation methods. (a) SDXL generates an image with the prompt \"figurine of Superman\". (b) The image is then \\(\\times 64\\) down-sampled (Left) with a close-up view (Right). (c) Upsampled image with bilinear interpolation and forward-reverse diffusion exhibits blurry details. (d) Naive pixel perturbation (PP) yields sharper details but introduces flickering artifacts. (e) Our Canny map-based adaptive PP produces clear and highly detailed image." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.318, + 0.76, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Fig. 8: Effectiveness of adaptive conditioning. Omitting either adaptive pose or text conditioning leads to issues like inconsistent poses, bad anatomy (yellow boxes), unwanted object duplication (red boxes), and mixed descriptions. Each color in the description corresponds to instances sharing the same color in the pose map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.788, + 0.79 + ], + "angle": 0, + "content": "the global CLIP score, which shows inconsistencies across methods. This suggests that our approach effectively enhances the original generation capabilities of the pretrained diffusion model for high-resolution image generation, achieving superior text-image correspondence, naturalness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Table 3 further analyzes the performance of different methods based on the number of people depicted in the scene. Our approach demonstrates robustness, effectively managing complexity as people count rises." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "User study We conducted a user study to evaluate the faithfulness and naturalness of images generated by three methods (one from each group) and our method. Participants compared large-scene images based on the same criteria used in the MLLM metrics: text-image correspondence, global naturalness, and human naturalness. A total of 12,120 responses were collected from 101 participants. The result shows that our method was consistently preferred over baselines, with a larger margin for all attributes. Notably, MLLM scores demonstrate strong alignment with human scores, supporting their validity. However, the global CLIP score appears misaligned, suggesting its limitations in capturing human judgment. We provide the additional details in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Beyond 8K ultra-high-resolution image As displayed in Fig. 2, BeyondScene excels in generating unparalleled resolution, producing images at an impressive \\(8192 \\times 8192\\) dimensions. This achievement surpasses the training resolution of SDXL by a remarkable factor of \\(64 \\times\\). It surpasses the standard 8K resolution \\((7680 \\times 4320)\\), demonstrating the model's capability to produce images with exceptional detail at even higher resolutions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.429, + 0.39, + 0.445 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.543 + ], + "angle": 0, + "content": "High frequency-injected forward diffusion As represented in 7, forward diffusion and reverse generation with naively up-sampled image lead to increased blurring in the image. By adopting the pixel perturbation, it makes the result sharpened. However, it introduced flickering artifacts at the image borders. Applying adaptive pixel perturbation based on the Canny map significantly improves image quality and reduces flickering artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Adaptive conditioning Fig. 8 demonstrates the effectiveness of adaptive conditioning. Omitting adaptive pose conditioning leads to inconsistent poses, bad anatomy, and decreased controllability compared to the desired input pose (yellow boxes). Excluding adaptive text conditioning, where each view receives the same full text including global and all instance descriptions, results in unwanted object duplication (red boxes) that ignores existing object scales, creating an awkward scene. Additionally, the descriptions of each instance become mixed, resulting in low text-image correspondence and decreased controllability." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.705, + 0.36, + 0.722 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduce BeyondScene, a novel framework capable of generating high-resolution human-centric scenes with exceptional text-image correspondence resolving artifacts beyond the token limits and beyond the training image size. BeyondScene overcomes limitations in generating complex human-centric scenes with detailed control, surpassing existing methods in generating high-fidelity scenes with control over human characteristics and number, while maintaining naturalness and reducing artifacts." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BeyondScene" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.296 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) [NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)] and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF-2022R1A4A1030579, NRF-2022M3C1A309202211). Also, the authors acknowledged the financial support from the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.318, + 0.323, + 0.334 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.348, + 0.787, + 0.375 + ], + "angle": 0, + "content": "1. Language segment anything. https://github.com/luca-medeiros/lang-segment-anything" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.376, + 0.533, + 0.389 + ], + "angle": 0, + "content": "2. Midjourney. https://www.midjourney.com" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.787, + 0.416 + ], + "angle": 0, + "content": "3. Sdsl-controlnet: Openpose (v2). https://huggingface.co/thibaud/controlnet-openpose-sdsl-1.0" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.416, + 0.787, + 0.443 + ], + "angle": 0, + "content": "4. Sdxl inpainting 0.1. https://huggingface.co/diffusers/stable-diffusion-x1-1.0-inpainting-0.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.444, + 0.787, + 0.485 + ], + "angle": 0, + "content": "5. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.485, + 0.787, + 0.512 + ], + "angle": 0, + "content": "6. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv:2302.08113 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.512, + 0.787, + 0.539 + ], + "angle": 0, + "content": "7. Chen, T.: On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.787, + 0.58 + ], + "angle": 0, + "content": "8. Cheong, S.Y., Mustafa, A., Gilbert, A.: KPE: Keypoint pose encoding for transformer-based image generation. In: British Machine Vision Conference (BMVC) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.787, + 0.621 + ], + "angle": 0, + "content": "9. Ding, Z., Zhang, M., Wu, J., Tu, Z.: Patched denoising diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.787, + 0.676 + ], + "angle": 0, + "content": "0. He, Y., Yang, S., Chen, H., Cun, X., Xia, M., Zhang, Y., Wang, X., He, R., Chen, Q., Shan, Y.: Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.787, + 0.703 + ], + "angle": 0, + "content": "1. Hoogeboom, E., Heek, J., Salimans, T.: simple diffusion: End-to-end diffusion for high resolution images. arXiv preprint arXiv:2301.11093 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.787, + 0.744 + ], + "angle": 0, + "content": "2. Ju, X., Zeng, A., Zhao, C., Wang, J., Zhang, L., Xu, Q.: Humansd: A native skeleton-guided diffusion model for human image generation. arXiv preprint arXiv:2304.04269 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.786 + ], + "angle": 0, + "content": "3. Ku, M., Jiang, D., Wei, C., Yue, X., Chen, W.: Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "4. Lee, Y., Kim, K., Kim, H., Sung, M.: Syncdiffusion: Coherent montage via synchronized joint diffusions. arXiv:2306.05178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: Openset grounded text-to-image generation. arXiv preprint arXiv:2301.07093 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.348, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Liu, D., Wu, L., Zheng, F., Liu, L., Wang, M.: Verbal-person nets: Pose-guided multi-granularity language-to-person generation. IEEE Transactions on Neural Networks and Learning Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.786, + 0.219 + ], + "angle": 0, + "content": "17. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.22, + 0.787, + 0.261 + ], + "angle": 0, + "content": "18. Liu, S., Cheng, H., Liu, H., Zhang, H., Li, F., Ren, T., Zou, X., Yang, J., Su, H., Zhu, J., et al.: Llava-plus: Learning to use tools for creating multimodal agents. arXiv preprint arXiv:2311.05437 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.317 + ], + "angle": 0, + "content": "19. Lv, Z., Li, X., Li, X., Li, F., Lin, T., He, D., Zuo, W.: Learning semantic person image generation by region-adaptive normalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10806-10815 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.346 + ], + "angle": 0, + "content": "20. Ma, L., Jia, X., Sun, Q., Schiele, B., Tuytelaars, T., Van Gool, L.: Pose guided person image generation. vol. 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.787, + 0.389 + ], + "angle": 0, + "content": "21. Ma, T., Peng, B., Wang, W., Dong, J.: MUST-GAN: Multi-level statistics transfer for self-driven person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13622-13631 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.787, + 0.431 + ], + "angle": 0, + "content": "22. Men, Y., Mao, Y., Jiang, Y., Ma, W.Y., Lian, Z.: Controllable person image synthesis with attribute-decomposed GAN. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5084-5093 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.787, + 0.473 + ], + "angle": 0, + "content": "23. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.515 + ], + "angle": 0, + "content": "24. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.572 + ], + "angle": 0, + "content": "25. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.614 + ], + "angle": 0, + "content": "26. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.67 + ], + "angle": 0, + "content": "27. Ren, Y., Fan, X., Li, G., Liu, S., Li, T.H.: Neural texture extraction and distribution for controllable person image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13535-13544 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.712 + ], + "angle": 0, + "content": "28. Ren, Y., Yu, X., Chen, J., Li, T.H., Li, G.: Deep image spatial transformation for person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7690-7699 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.787, + 0.742 + ], + "angle": 0, + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.77 + ], + "angle": 0, + "content": "30. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.798 + ], + "angle": 0, + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "32. Roy, P., Ghosh, S., Bhattacharya, S., Pal, U., Blumenstein, M.: TIPS: Text-induced pose synthesis. In: European Conference on Computer Vision (ECCV). pp. 161-178. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BeyondScene" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "33. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "34. Teng, J., Zheng, W., Ding, M., Hong, W., Wangni, J., Yang, Z., Tang, J.: Relay diffusion: Unifying diffusion process across resolutions for image synthesis. arXiv preprint arXiv:2309.03350 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "35. Wang, L., Li, H., Hu, W., Zhang, X., Qiu, H., Meng, F., Wu, Q.: What happens in crowd scenes: A new dataset about crowd scenes for image captioning. IEEE Transactions on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "36. Wu, J., Wang, J., Yang, Z., Gan, Z., Liu, Z., Yuan, J., Wang, L.: Grit: A generative region-to-text transformer for object understanding. arXiv:2212.00280 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "37. Xie, E., Yao, L., Shi, H., Liu, Z., Zhou, D., Liu, Z., Li, J., Li, Z.: Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. arXiv preprint arXiv:2304.06648 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.397 + ], + "angle": 0, + "content": "38. Xu, X., Chen, Y.C., Tao, X., Jia, J.: Text-guided human image manipulation via image-text shared space. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) 44(10), 6486-6500 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.424 + ], + "angle": 0, + "content": "39. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. NeurIPS 35, 38571-38584 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.467 + ], + "angle": 0, + "content": "40. Yang, F., Lin, G.: CT-Net: Complementary transferring network for garment transfer with arbitrary geometric changes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9899-9908 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.786, + 0.508 + ], + "angle": 0, + "content": "41. Yang, L., Yu, Z., Meng, C., Xu, M., Ermon, S., Cui, B.: Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. arXiv preprint arXiv:2401.11708 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "42. Yang, L., Wang, P., Liu, C., Gao, Z., Ren, P., Zhang, X., Wang, S., Ma, S., Hua, X., Gao, W.: Towards fine-grained human pose transfer with detail replenishing network. IEEE Transactions on Image Processing 30, 2422-2435 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "43. Zhang, B., Gu, S., Zhang, B., Bao, J., Chen, D., Wen, F., Wang, Y., Guo, B.: Styleswin: Transformer-based gan for high-resolution image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11304-11314 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "44. Zhang, J., Li, K., Lai, Y.K., Yang, J.: PISE: Person image synthesis and editing with decoupled GAN. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7982-7990 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "45. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.716 + ], + "angle": 0, + "content": "46. Zhang, P., Yang, L., Lai, J.H., Xie, X.: Exploring dual-task correlation for pose guided person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7713-7722 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.786, + 0.743 + ], + "angle": 0, + "content": "47. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. arXiv:2303.17076 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "48. Zheng, Q., Guo, Y., Deng, J., Han, J., Li, Y., Xu, S., Xu, H.: Any-size-diffusion: Toward efficient text-driven synthesis for any-size hd images. arXiv preprint arXiv:2308.16582 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.785 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_origin.pdf b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b06fb674bd3b8ee49fa5d28e4d150b84ddd1fb6 --- /dev/null +++ b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/f6f57311-b85e-4b93-9433-9a2027f7d58d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c71b3d1f7848a5efd40cb7227a701a0e36e7b6a613df324c25ac301440f480dc +size 10394710 diff --git a/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/full.md b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..36f97274f29dedf6f6faa933941237e2bf624cba --- /dev/null +++ b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/full.md @@ -0,0 +1,229 @@ +# BeyondScene: Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion + +Gwanghyun Kim $^{1\star}$ , Hayeon Kim $^{1\star}$ , Hoigi Seo $^{1\star}$ , Dong Un Kang $^{1\star}$ , Se Young Chun $^{1,2\dagger}$ + +$^{1}$ Dept. of Electrical and Computer Engineering, $^{2}$ INMC & IPAI Seoul National University, Republic of Korea + +{gwang.kim, khy5630, sehoiki3215, qkrtnsfk23, sychun}@snu.ac.kr + +Abstract. Generating higher-resolution human-centric scenes with details and controls remains a challenge for existing text-to-image diffusion models. This challenge stems from limited training image size, text encoder capacity (limited tokens), and the inherent difficulty of generating complex scenes involving multiple humans. While current methods attempted to address training size limit only, they often yielded human-centric scenes with severe artifacts. We propose BeyondScene, a novel framework that overcomes prior limitations, generating exquisite higher-resolution (over 8K) human-centric scenes with exceptional text-image correspondence and naturalness using existing pretrained diffusion models. BeyondScene employs a staged and hierarchical approach to initially generate a detailed base image focusing on crucial elements in instance creation for multiple humans and detailed descriptions beyond token limit of diffusion model, and then to seamlessly convert the base image to a higher-resolution output, exceeding training image size and incorporating details aware of text and instances via our novel instance-aware hierarchical enlargement process that consists of our proposed high-frequency injected forward diffusion and adaptive joint diffusion. BeyondScene surpasses existing methods in terms of correspondence with detailed text descriptions and naturalness, paving the way for advanced applications in higher-resolution human-centric scene creation beyond the capacity of pretrained diffusion models without costly retraining. Project page: https://janeyon.github.io/beyond-scene. + +Keywords: Human-centric scene generation $\cdot$ Text-to-image diffusion model $\cdot$ High-resolution + +# 1 Introduction + +Human-centric scene generation [8, 12, 15, 16, 19-23, 27, 28, 32, 33, 38, 40, 42, 44-46], has emerged as a critical research area with significant academic and industrial applications. While recent advancements utilizing text-to-image (T2I) diffusion + +Fig. 1: BeyondScene pushes the boundaries of high-resolution human-centric scene generation. Unlike existing methods that often suffer from unrealistic scenes, anatomical distortions, and limited text-to-image correspondence, BeyondScene excels in 1) highly detailed scenes, 2) natural and diverse humans, 3) fine-grained control. The color in each text represents the text for each instance that has the same color in the pose map. +![](images/672bf6d39fbb5d8a3cf78b2843eb49f7347ba83a3116d7a182ae4c74a3e6b9b3.jpg) +"In the sunny road, there are a clay animation character with lab coat, a clay character with red dotted dress and glasses, a clay character with short, chubby cheeks and glasses, a clay animation character with short, chubby checks wearing blue jeans, a clay character with spiky hair, a clay character with yellow T-shirts." + +models [2, 26, 29] have yielded promising results in generating controllable human scenes [12, 15, 23, 45], scaling these methods to handle larger and more complex scenes such as involving multiple humans remains a significant challenge. These limitations primarily stem from following causes. 1) Training image size: Directly sampling from a limited training image size can introduce artifacts and constrain the final scene resolution. 2) Limited text encoder tokens: The restricted number of tokens in text encoders (typically 77 for Stable Diffusion) hinders the inclusion of detailed descriptions for multiple instances within the scene. 3) Inherent limitations of T2I diffusion models: These models struggle to generate complex scenes with several human figures and intricate details. + +Existing attempts to address the training size issue, such as joint-process [6, 14, 47] and dilation-based methods [10], only partially address the problem. These methods often introduce new challenges specific to human-centric scenes as represented in Fig. 1, including: 1) Unrealistic scenes and objects: These methods can generate nonsensical scenarios with duplicated objects, humans defying gravity, and physically distorted environments. 2) Anatomical distortions: Generated scenes may exhibit unrealistic human anatomy, such as abnormal limbs or facial features. 3) Limited correspondence: Existing methods often fail to capture the complexity of scenes with multiple human instances and detailed descriptions, often generating single objects and lacking control over specific details like clothing or hairstyles. + +We propose BeyondScene, a novel framework that overcomes these limitations, generating high-resolution (over 8K) human-centric scenes with exceptional text-image correspondence and naturalness. BeyondScene employs a staged and hierarchical approach, which looks similar to a classical multi-resolution manner, but it is closer to how artists establish a foundation before adding details. + +Firstly, BeyondScene generates a detailed base image focusing on essential elements, human poses and detailed descriptions beyond the token limits. This initial stage enables detailed instance creation for multiple humans, surpassing the limitations of text encoders. Secondly, the method leverages an instance-aware hierarchical enlargement process to convert this base image to a higher resolution output beyond the training image size. Unlike naive super-resolution methods that simply scale resolution without considering text and instances, our approach refines content and adds more details aware of text and instances. This is achieved through our proposed novel techniques, including 1) high frequency-injected forward diffusion that addresses the issue of blurred low-quality results in image-to-image translation with the upsampled image by adaptively injecting high-frequency details into the upsampled image, enhancing the final output while preserving content and 2) adaptive joint diffusion that facilitates efficient and robust joint diffusion while maintaining control over human characteristics like pose and mask. This approach utilizes view-wise conditioning of text and pose information, along with a variable stride for the joint process. Furthermore, similar to how artists add details progressively, our method allows for the addition of details at different stages, exploiting the changing receptive field of each view. + +We comprehensively evaluate BeyondScene using qualitative and quantitative metrics, along with user studies. The results demonstrate significant improvements over existing methods in terms of 1) correspondence with detailed text descriptions and 2) naturalness and reduction of artifacts. Furthermore, we showcase the result of $8192 \times 8192$ image generation beyond 8K ultra high-resolution, demonstrating the capability for generating even higher-resolution images as displayed in Fig. 2. These advancements pave the way for exciting new applications in higher-resolution human-centric scene creation. + +![](images/58e3552bda3e6dd42e6614ddf0c60b36f5555c6a3b2bf4733604d06567f38fdb.jpg) +Fig. 2: Beyond 8K ultra-high resolution image. This $8192 \times 8192$ image, generated by BeyondScene, surpasses the training resolution of SDXL by $64 \times$ , while exceeding the technical classification of 8K ( $7680 \times 4320$ ). + +# 2 Related work + +# 2.1 Controllable Human Generation + +Early approaches to controllable human generation [8, 19-22, 27, 28, 40, 42, 44, 46] relied on pose guidance and source images, achieving success within specific scenarios but struggling with diverse scenes and arbitrary poses. Text-based conditioning methods emerged with the rise of large vision-language models [25], but initial attempts [16, 32, 38] faced limitations in vocabulary size and open vocabulary settings. Recent T2I diffusion [2, 24, 26, 30]-based methods like ControlNet [45], T2I-Adapter [23], GLIGEN [15], and HumanSD [12] introduced methods for incorporating diverse conditions, but scaling to larger scenes with + +multiple individuals remains a challenge. This work proposes BeyondScene, a novel framework specifically designed to overcome these limitations and enable the generation of high-resolution human-centric scenes. + +# 2.2 Large Scene Generation Using Diffusion Models + +Achieving high-resolution image generation [9, 26] presents significant hurdles. Training models from scratch or fine-tuning pretrained models [2, 7, 11, 24, 26, 30, 31, 34, 37, 43, 48] requires immense computational resources and struggles with the complexity of high-dimensional data. Recent exploration has ventured into training-free methods, with approaches like MultiDiffusion [6], SyncDiffusion [14] (joint diffusion), and ScaleCrafter [10] (dilation-based) emerging. However, these methods often introduce challenges specific to human-centric scenes, including unrealistic objects, anatomical distortions, and limited correspondence between text descriptions and generated images. BeyondScene, our proposed framework, addresses these limitations through detailed base image generation and instance-aware hierarchical enlargement. + +# 3 BeyondScene + +Current approaches to generating large scenes attempt to fill the entire canvas at once, often leading to challenges in capturing complex details and ensuring physical realism. Inspired by the artistic workflow of human painters, who establish a foundation with key elements and progressively refine them, we introduce BeyondScene, a novel framework that generates high-resolution human-centric images through a staged and hierarchical approach as represented in Fig. 3. + +BeyondScene operates in two key stages, 1) detailed base image generation (Sec. 3.1) that focuses on generating high-quality base images with precise control over human poses and detailed descriptions beyond typical token limitations and 2) instance-aware hierarchical enlargement (Sec. 3.2) where the base image is progressively enlarged while maintaining control and adding scene details. + +# 3.1 Detailed Base Image Generation + +Generating human-centric scenes with multiple individuals using T2I diffusion models faces a key challenge: the limited number of tokens available in the text encoder restricts detailed descriptions to a few key elements. To overcome this, we generate detailed base image as illustrated in Fig. 3. Firstly, individual instances are generated in the training resolution of pose-guided T2I diffusion models [12, 15, 23, 45]. This allows for detailed descriptions beyond token limitations, tailored to each individual. Subsequently, the generated instances are segmented, cropped, and resized to the desired size. Next, a diffusion-based inpainting method is employed to create an instance-aware background that seamlessly integrates with the individual instances. This ensures a realistic and coherent background tailored to the specific scene. To address potential inconsistencies in brightness + +![](images/bf639b9fd424f6bca4a2b92dd0ecb315487dbb9e978d8336283b95e1d773f88d.jpg) +Fig. 3: BeyondScene generates high-resolution images in two stages. First, individual instances are created using pose-guided T2I diffusion models, segmented, cropped, and placed onto an inpainted background. The tone of image is then normalized. In the second stage (illustrated in Fig. 4), this base image is progressively enlarged while maintaining detail and quality, effectively refining the image and adding further details leveraging high-frequency injected forward diffusion and adaptive joint diffusion (AJ). + +and color tones across individual elements, tone normalization through contrast limited adaptive histogram equalization (CLAHE) is performed. This involves first creating a grid histogram for pixel values within a specified grid, followed by adjusting the local contrast to align with the image's overall contrast. Notably, direct equalization in the presence of noise within the grid could induce the noise amplification. To prevent this, equalization is preceded by a redistribution process that includes limiting the maximum height of the grid histogram. + +# 3.2 Instance-Aware Hierarchical Enlargement + +BeyondScene introduces a novel approach to address the challenge of generating high-resolution outputs beyond the training image size through a instance-aware hierarchical enlargement process, which surpasses mere scaling by actively refining content and incorporating additional details aware of text and instance. + +In the initial step, a low-resolution image is upsampled, and a forward diffusion process is applied to create a noisy upsampled image. Subsequently, a joint diffusion process refines the image to a high resolution over the training resolution. + +![](images/af0efa79b24c3839a184bb36065655b15dd1e50b9c440d83e3bd58b3e7ad0fc9.jpg) +3.2 Instance-Aware Hierarchical Enlargement +a. High frequency-injected forward diffusion + +![](images/9c6c58c75699dce09680a47badcb2b2398098f112d481982d3571c04128c89af.jpg) +b. Adaptive joint diffusion (AJ) At the diffusion timestep $t$ +Fig. 4: Our instance-aware hierarchical enlargement involves two crucial processes: 1) High frequency-injected forward diffusion, which enables to achieve high resolution through a joint diffusion employing adaptive pixel perturbation. 2) Adaptive joint diffusion, dynamically regulating stride and conditioning of pose and text based on the presence of instances. + +However, a drawback of this naive strategy is the induction of blurriness due to the model assigning sufficiently high likelihood to low-quality outputs. Also, this strategy often encounters issues such as unnatural duplication of objects and humans, stemming from the uniform application of the same text prompt to all views without pose conditioning. To mitigate these, the paper proposes a novel high frequency-injected forward diffusion and adaptive joint diffusion as illustrated in Fig. 4. + +High frequency-injected forward diffusion High frequency-injected forward diffusion enhances the translation of noisy latents from the upsampled image to high resolution with intricate details. This is achieved through a joint diffusion process employing adaptive pixel perturbation. This technique injects high-frequency details into the upsampled image based on a Canny edge map. This process enhances the final output's quality while preserving content. While simply adding pixel noise sharpens the image, it can also introduce flickering artifacts at the borders. Our adaptive approach based on the Canny map significantly improves image quality and reduces these artifacts. Specifically, we first up-sample the low-resolution image and calculate Canny map. Then Gaussian blur is applied to the map and normalized it to get probability map $C$ . $\epsilon_{i,j} \sim \mathcal{U}(0,1)$ is sampled + +on each pixel, $I_{i,j}$ which denotes $i$ -th row and $j$ -th column pixel of image $I$ . If $\epsilon_{i,j} > C_{i,j}$ , the pixel value within $d_r$ pixel distance is replaced so that we could apply pixel perturbation adaptively to avoid flickering artifacts near borders. + +Adaptive joint diffusion To mitigate duplication of objects and humans in the naive joint diffusion, we introduce adaptive conditioning and adaptive stride. + +Adaptive conditioning We introduce adaptive view-wise conditioning, which leverages the segmentation maps obtained from the base image generation stage. For each view, we check which human instances are present. If an instance is included, we add its pose and detailed description to the prompt used for that view. Additionally, if the detailed text description includes specific parts (head, face, upper body, etc.), we can apply these details to the corresponding regions using fine-grained segmentation maps. This facilitates efficient and robust joint diffusion while maintaining control over human characteristics like pose and appearance. + +Adaptive stride To effectively enhance the quality of the generated scene, we employ an adaptive stride for the joint process. We reduce the stride in regions containing humans, ensuring that these areas capture fine details. In contrast, the stride is increased in background regions, allowing for more efficient computation in these areas. This combination of adaptive view-wise conditioning and adaptive stride allows BeyondScene to effectively handle scenes with multiple humans while maintaining high-resolution detail and controllable image generation. The details and algorithms on our methods are presented in the supplementary material. + +# 4 Experiments + +# 4.1 Experimental Settings + +We comprehensively evaluate our method using both qualitative and quantitative metrics. To ensure a fair comparison, all compared models are implemented using Stable Diffusion XL (SDXL) [24]-based architectures. We utilize SDXL-ControlNet-Openpose [3, 45], SDXL-inpainting [4], and Lang-Segment-Anything [1] for our pose-guided T2I diffusion, inpainting diffusion model, and segmentation models. The training resolution is set to $1024 \times 1024$ , while inference resolutions are varied across $2048 \times 2048$ ( $4 \times 1:1$ ), $3584 \times 2048$ ( $7 \times 7:4$ ), $4096 \times 2048$ ( $8 \times 2:1$ ), $4096 \times 4096$ ( $16 \times 1:1$ ), and $8192 \times 8192$ ( $64 \times 1:1$ ). + +Testing data Existing T2I generation datasets often lack large-scale scenes featuring multiple people with detailed descriptions for each individual. To address this limitation, we leverage the CrowdCaption dataset [35] as our primary test set. This dataset features images containing large crowds of people, allowing us to evaluate the effectiveness of our method in generating complex scenes with numerous individuals. For each image in the CrowdCaption dataset, we obtain + +![](images/f11e7bd20de4f971039ab1b694fd0c4c7f4887a0531e9a26e839ff6b51ae8b8f.jpg) +"In empty stage of opera house, there are a ballerina with pink suit, a ballerina wearing blue ballet suit, a ballerina with pink suit, a ballerina with blue suit, a ballerina with yellow sparkling ballet suit, a ballerina with blue suit, a ballerina with pink suit, a ballerina with blue suit, a ballerina with pink suit." + +Fig. 5: Qualitative comparison for generating high-resolution human scenes ( $3584 \times 2048$ ). While existing approaches like T2I-Direct (SDXL [24]), T2I-Large (MultiDiffusion [6], SyncDiffusion [14], and ScaleCrafter [10]), and Visual+T2I (ControlNet [24, 45], T2IAdapter [23, 24], and R-MultiDiffusion [6]) models struggle with artifacts, our method achieves superior results by producing images with minimal artifacts, strong text-image correspondence, and a natural look. The color in each description represents the description for each instance that has the same color in the pose map. +![](images/03a6eefb3b08aa41ef3a7c7a0d705d101a33bbe38ce23245cb38901dd28dc91b.jpg) +"In the alien planet with galaxy in the sky, there are a storm trooper, a storm trooper, a storm trooper, a Darth Vader, a storm trooper, a storm trooper, a storm trooper." + +descriptions and individual poses using GRIT [36] and ViTPose [39]. Additionally, we obtain a global caption generated using GPT4 [5]. We filter the CrowdCaption images based on two criteria: number of humans (max 8 to accommodate token limits of baselines) and aspect ratio (close to 1:1 and 2:1), which results in a collection of 100 images for each aspect ratio (1:1 and 2:1). Additionally, to showcase the versatility of our method beyond the CrowdCaption dataset, we incorporate custom examples for qualitative comparisons. + +![](images/468f81c0bbd2d152feaf101be98e08e6060e4b6f489a500eed9add25293d3c18.jpg) +Fig. 6: Examples of large scene synthesis $(4096\times 4096)$ on the poses and text obtained from CrowdCaption [35] images. All baselines including direct high-resolution inference (SDXL [24]), T2I models for large scenes (MultiDiffusion [6], SyncDiffusion [14], ScaleCrafter [10]), Visual prior-guided T2I models (ControlNet [45], T2IAAdapter [23]) produce duplicated objects and artifacts in human anatomy, while our method succeeded in generation of high-resolution image with high text-image correspondence. Each color in the description corresponds to instances sharing the same color in the pose map. + +Evaluation metrics For an evaluation of text-image correspondence, we adopt a multimodal large language model [5, 17, 18]-based text-image correspondence metric, as proposed in VIEScore [13], alongside the global CLIP [25] score that shows its limitations in producing reliable scores for complex scenes with detailed descriptions due to limitations in input image size and the available number of input tokens. This MLLM-based metric, powered by GPT4 [5], assigns a score ranging from 0 to 10, where 0 indicates no correspondence between the generated image and the prompt and 10 indicates perfect alignment. This metric has been demonstrably well-correlated with human judgments and provides explanations for the assigned scores [13]. To assess the naturalness of the generated images, we adapt the existing MLLM-based naturalness metric to focus on human-centric + +Table 1: Quantitative comparison of our method with various approaches, including direct high-resolution inference (T2I-Direct) [24,41], T2I models designed for large scenes (T2I-Large) [6,10,14], and Visual prior-guided T2I models (Visual+T2I) [6,23,45]. + +
Model typesModelsGlobal CLIPMLLM (GPT4)-based
Text-image correspondenceGlobal naturalnessHuman naturalness
2048×2048 (4×1:1)
T2I-DirectSDXL0.3242.1411.8381.417
RPG0.2611.9632.1472.785
T2I-LargeMultiDiffusion0.3473.8553.3812.466
SyncDiffusion0.3473.6553.4292.607
ScaleCrafter0.3456.0815.6675.062
Visual+T2IControlNet0.2981.7901.4661.117
T2IAdapter0.3283.0942.6601.728
R-MultiDiffusion0.3113.6361.7111.597
BeyondScene0.3057.0416.5356.114
4096×2048 (8×2:1)
T2I-DirectSDXL0.3011.0610.7710.607
RPGOOMOOMOOMOOM
T2I-LargeMultiDiffusion0.3382.1362.2281.329
SyncDiffusion0.3423.1282.6651.591
ScaleCrafter0.2963.5523.5693.359
Visual+T2IControlNet0.2781.5261.0160.657
T2IAdapter0.2821.1091.0100.568
R-MultiDiffusion0.3063.4661.4221.201
BeyondScene0.3047.1186.6126.375
4096×4096 (16×1:1)
T2I-DirectSDXL0.2770.3290.4110.118
RPGOOMOOMOOMOOM
T2I-LargeMultiDiffusion0.3191.3052.4991.591
SyncDiffusion0.3100.9122.3091.359
ScaleCrafter0.3122.9623.0482.442
Visual+T2IControlNet0.2440.8111.3090.729
T2IAdapter0.2690.4910.7190.352
R-MultiDiffusion0.2811.7920.8600.708
BeyondScene0.3016.8016.0745.627
+ +scenarios, leading to two separate scores: Global naturalness scores from 0 to 10, with 0 indicating an unnatural or unrealistic image and 10 indicating a completely natural image. This score considers factors such as overall inconsistencies, unrealistic physics, disconnectivity. Human naturalness scores from 0 to 10, with 0 indicating unnatural human anatomy, and 10 indicating the absence of any artifacts in the human anatomy. Additional details on implementation and evaluation are provided in the supplementary material. + +# 4.2 Result + +Our method is compared against three groups of existing SOTA approaches: 1) Direct high-resolution inference using T2I models (T2I-Direct) that directly perform high-resolution inference of SDXL [24] and RPG [41] over their training image size, 2) T2I diffusion models for large scenes (T2I-Large) that encom + +Table 2: Quantitative comparison by scene complexity. Our method shows robust performance (averaged across resolutions: $2048 \times 2048$ , $4096 \times 2048$ , $4096 \times 4096$ ) even with increasing the number of humans, highlighting its ability to handle complex scenes. + +
Models2~4 humans5~8 humans
MLLM (GPT4)-basedMLLM (GPT4)-based
Text-image corr.Global nat.Human nat.Text-image corr.Global nat.Human nat.
SDXL1.1921.0390.8451.1610.9740.853
RPG3.7573.1383.0130.1691.1562.558
MultiDiffusion2.1422.3921.6922.7223.0131.911
SyncDiffusion2.2772.4281.6922.8533.1752.013
ScaleCrafter4.554.4374.0643.8463.7523.177
ControlNet1.1721.1940.8101.5791.3330.858
T2IAdapter1.5211.4490.9641.6081.4860.801
R-MultiDiffusion3.271.2861.3292.6591.3761.007
BeyondScene7.3086.4966.3766.6666.3195.701
+ +Table 3: User study result. Users consistently preferred our method over baselines, with MLLM scores closely aligning with human choices. However, the global CLIP score diverged, suggesting potential limitations. + +
ModelsPreference of Ours ↑
Text-image corr.Global Human nat.
vs ScaleCrafter0.7950.6500.806
vs MultiDiffusion0.7270.6580.822
vs ControlNet0.8700.6940.897
+ +pass joint diffusion-based methods (MultiDiffusion [6], SyncDiffusion [14]), and ScaleCrafter [10].) Visual prior-guided T2I generation models (Visual+T2I) that focuses on models including SDXL-ControlNet-Openpose [24, 45], SDXL-T2IAAdapter [23, 24], and R-MultiDiffusion [6] that leverage visual priors for T2I generation as our method. We provide more results and ablation studies in the supplementary material. + +Qualitative comparison As represented in Fig. 1, 5 and 6, baselines introduce noticeable instance duplication, artifacts in anatomy, facial features, making the scene unnatural rather than reflecting real-world physics. Also, they struggle to generate complex scenes corresponding to the detailed descriptions. Conversely, our method generates images with minimal artifacts, demonstrating superior text-image correspondence, global and human naturalness, particularly in human-centric scenes. + +Quantitative comparison Table 1 summarizes the performance across the resolution settings of $2048 \times 2048$ ( $4 \times 1:1$ ), $4096 \times 2048$ ( $8 \times 2:1$ ), and $4096 \times 4096$ ( $16 \times 1:1$ ). Our method consistently outperforms baselines on all metrics except + +![](images/ebab3973c207d74f9d1afe68c5ed99a9f6958dbeca1b2d54f1f89239430f1760.jpg) +(a) Original image + +![](images/b3d88ff5f10bd423587f714d6bb84bee38d6a8cd37c41ad123785f57122bc3c7.jpg) +(b) 64x down-sampled + +![](images/e5001ee46066886bbce1f62b5155721853b98e2c1354a5749ab31ba409beca21.jpg) +mage (L) / Its part (R) (c) Bilinear interpol. + +![](images/42ebeae8fa2d34185d3c86930c12f5144be7a333cdd8a0bd5861f9dd74f675fe.jpg) +(d) Naive PP + +(c) Bilinear interpol. + +![](images/396b37184c12fbc7c81b30616c51b7554ffa503706e18b8743ec179175149ea8.jpg) +(e) Adaptive PP (Ours) + +![](images/600c8abbe8eece1ed5eaf4d03cb651d1cca7198b0623fe7622544ac789be9296.jpg) + +![](images/0e014212246c6b20a3e3acb2149d7c5041f825f3709d630858cb6e32b04c0efd.jpg) +Fig. 7: Forward-reverse diffusion process with interpolation methods. (a) SDXL generates an image with the prompt "figurine of Superman". (b) The image is then $\times 64$ down-sampled (Left) with a close-up view (Right). (c) Upsampled image with bilinear interpolation and forward-reverse diffusion exhibits blurry details. (d) Naive pixel perturbation (PP) yields sharper details but introduces flickering artifacts. (e) Our Canny map-based adaptive PP produces clear and highly detailed image. +Fig. 8: Effectiveness of adaptive conditioning. Omitting either adaptive pose or text conditioning leads to issues like inconsistent poses, bad anatomy (yellow boxes), unwanted object duplication (red boxes), and mixed descriptions. Each color in the description corresponds to instances sharing the same color in the pose map. + +the global CLIP score, which shows inconsistencies across methods. This suggests that our approach effectively enhances the original generation capabilities of the pretrained diffusion model for high-resolution image generation, achieving superior text-image correspondence, naturalness. + +Table 3 further analyzes the performance of different methods based on the number of people depicted in the scene. Our approach demonstrates robustness, effectively managing complexity as people count rises. + +User study We conducted a user study to evaluate the faithfulness and naturalness of images generated by three methods (one from each group) and our method. Participants compared large-scene images based on the same criteria used in the MLLM metrics: text-image correspondence, global naturalness, and human naturalness. A total of 12,120 responses were collected from 101 participants. The result shows that our method was consistently preferred over baselines, with a larger margin for all attributes. Notably, MLLM scores demonstrate strong alignment with human scores, supporting their validity. However, the global CLIP score appears misaligned, suggesting its limitations in capturing human judgment. We provide the additional details in the supplementary material. + +Beyond 8K ultra-high-resolution image As displayed in Fig. 2, BeyondScene excels in generating unparalleled resolution, producing images at an impressive $8192 \times 8192$ dimensions. This achievement surpasses the training resolution of SDXL by a remarkable factor of $64 \times$ . It surpasses the standard 8K resolution $(7680 \times 4320)$ , demonstrating the model's capability to produce images with exceptional detail at even higher resolutions. + +# 4.3 Ablation Study + +High frequency-injected forward diffusion As represented in 7, forward diffusion and reverse generation with naively up-sampled image lead to increased blurring in the image. By adopting the pixel perturbation, it makes the result sharpened. However, it introduced flickering artifacts at the image borders. Applying adaptive pixel perturbation based on the Canny map significantly improves image quality and reduces flickering artifacts. + +Adaptive conditioning Fig. 8 demonstrates the effectiveness of adaptive conditioning. Omitting adaptive pose conditioning leads to inconsistent poses, bad anatomy, and decreased controllability compared to the desired input pose (yellow boxes). Excluding adaptive text conditioning, where each view receives the same full text including global and all instance descriptions, results in unwanted object duplication (red boxes) that ignores existing object scales, creating an awkward scene. Additionally, the descriptions of each instance become mixed, resulting in low text-image correspondence and decreased controllability. + +# 5 Conclusion + +We introduce BeyondScene, a novel framework capable of generating high-resolution human-centric scenes with exceptional text-image correspondence resolving artifacts beyond the token limits and beyond the training image size. BeyondScene overcomes limitations in generating complex human-centric scenes with detailed control, surpassing existing methods in generating high-fidelity scenes with control over human characteristics and number, while maintaining naturalness and reducing artifacts. + +# Acknowledgements + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) [NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)] and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF-2022R1A4A1030579, NRF-2022M3C1A309202211). Also, the authors acknowledged the financial support from the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University. + +# References + +1. Language segment anything. https://github.com/luca-medeiros/lang-segment-anything +2. Midjourney. https://www.midjourney.com +3. Sdsl-controlnet: Openpose (v2). https://huggingface.co/thibaud/controlnet-openpose-sdsl-1.0 +4. Sdxl inpainting 0.1. https://huggingface.co/diffusers/stable-diffusion-x1-1.0-inpainting-0.1 +5. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +6. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv:2302.08113 (2023) +7. Chen, T.: On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972 (2023) +8. Cheong, S.Y., Mustafa, A., Gilbert, A.: KPE: Keypoint pose encoding for transformer-based image generation. In: British Machine Vision Conference (BMVC) (2022) +9. Ding, Z., Zhang, M., Wu, J., Tu, Z.: Patched denoising diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2023) +0. He, Y., Yang, S., Chen, H., Cun, X., Xia, M., Zhang, Y., Wang, X., He, R., Chen, Q., Shan, Y.: Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In: The Twelfth International Conference on Learning Representations (2023) +1. Hoogeboom, E., Heek, J., Salimans, T.: simple diffusion: End-to-end diffusion for high resolution images. arXiv preprint arXiv:2301.11093 (2023) +2. Ju, X., Zeng, A., Zhao, C., Wang, J., Zhang, L., Xu, Q.: Humansd: A native skeleton-guided diffusion model for human image generation. arXiv preprint arXiv:2304.04269 (2023) +3. Ku, M., Jiang, D., Wei, C., Yue, X., Chen, W.: Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867 (2023) +4. Lee, Y., Kim, K., Kim, H., Sung, M.: Syncdiffusion: Coherent montage via synchronized joint diffusions. arXiv:2306.05178 (2023) +5. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: Openset grounded text-to-image generation. arXiv preprint arXiv:2301.07093 (2023) + +16. Liu, D., Wu, L., Zheng, F., Liu, L., Wang, M.: Verbal-person nets: Pose-guided multi-granularity language-to-person generation. IEEE Transactions on Neural Networks and Learning Systems (2022) +17. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024) +18. Liu, S., Cheng, H., Liu, H., Zhang, H., Li, F., Ren, T., Zou, X., Yang, J., Su, H., Zhu, J., et al.: Llava-plus: Learning to use tools for creating multimodal agents. arXiv preprint arXiv:2311.05437 (2023) +19. Lv, Z., Li, X., Li, X., Li, F., Lin, T., He, D., Zuo, W.: Learning semantic person image generation by region-adaptive normalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10806-10815 (2021) +20. Ma, L., Jia, X., Sun, Q., Schiele, B., Tuytelaars, T., Van Gool, L.: Pose guided person image generation. vol. 30 (2017) +21. Ma, T., Peng, B., Wang, W., Dong, J.: MUST-GAN: Multi-level statistics transfer for self-driven person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13622-13631 (2021) +22. Men, Y., Mao, Y., Jiang, Y., Ma, W.Y., Lian, Z.: Controllable person image synthesis with attribute-decomposed GAN. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5084-5093 (2020) +23. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) +24. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +25. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021) +26. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022) +27. Ren, Y., Fan, X., Li, G., Liu, S., Li, T.H.: Neural texture extraction and distribution for controllable person image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13535-13544 (2022) +28. Ren, Y., Yu, X., Chen, J., Li, T.H., Li, G.: Deep image spatial transformation for person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7690-7699 (2020) +29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) +30. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) +31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021) +32. Roy, P., Ghosh, S., Bhattacharya, S., Pal, U., Blumenstein, M.: TIPS: Text-induced pose synthesis. In: European Conference on Computer Vision (ECCV). pp. 161-178. Springer (2022) + +33. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +34. Teng, J., Zheng, W., Ding, M., Hong, W., Wangni, J., Yang, Z., Tang, J.: Relay diffusion: Unifying diffusion process across resolutions for image synthesis. arXiv preprint arXiv:2309.03350 (2023) +35. Wang, L., Li, H., Hu, W., Zhang, X., Qiu, H., Meng, F., Wu, Q.: What happens in crowd scenes: A new dataset about crowd scenes for image captioning. IEEE Transactions on Multimedia (2022) +36. Wu, J., Wang, J., Yang, Z., Gan, Z., Liu, Z., Yuan, J., Wang, L.: Grit: A generative region-to-text transformer for object understanding. arXiv:2212.00280 (2022) +37. Xie, E., Yao, L., Shi, H., Liu, Z., Zhou, D., Liu, Z., Li, J., Li, Z.: Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. arXiv preprint arXiv:2304.06648 (2023) +38. Xu, X., Chen, Y.C., Tao, X., Jia, J.: Text-guided human image manipulation via image-text shared space. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) 44(10), 6486-6500 (2021) +39. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. NeurIPS 35, 38571-38584 (2022) +40. Yang, F., Lin, G.: CT-Net: Complementary transferring network for garment transfer with arbitrary geometric changes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9899-9908 (2021) +41. Yang, L., Yu, Z., Meng, C., Xu, M., Ermon, S., Cui, B.: Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. arXiv preprint arXiv:2401.11708 (2024) +42. Yang, L., Wang, P., Liu, C., Gao, Z., Ren, P., Zhang, X., Wang, S., Ma, S., Hua, X., Gao, W.: Towards fine-grained human pose transfer with detail replenishing network. IEEE Transactions on Image Processing 30, 2422-2435 (2021) +43. Zhang, B., Gu, S., Zhang, B., Bao, J., Chen, D., Wen, F., Wang, Y., Guo, B.: Styleswin: Transformer-based gan for high-resolution image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11304-11314 (2022) +44. Zhang, J., Li, K., Lai, Y.K., Yang, J.: PISE: Person image synthesis and editing with decoupled GAN. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7982-7990 (2021) +45. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023) +46. Zhang, P., Yang, L., Lai, J.H., Xie, X.: Exploring dual-task correlation for pose guided person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7713-7722 (2022) +47. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. arXiv:2303.17076 (2023) +48. Zheng, Q., Guo, Y., Deng, J., Han, J., Li, Y., Xu, S., Xu, H.: Any-size-diffusion: Toward efficient text-driven synthesis for any-size hd images. arXiv preprint arXiv:2308.16582 (2023) \ No newline at end of file diff --git a/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/images.zip b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..137fb50caf7204d9ae46abe216ec59d4b430ec3a --- /dev/null +++ b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d622759c47c17af929c57a7203ee86e4f43f61d7698618d2a077ad61a9b397dd +size 1272119 diff --git a/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/layout.json b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9e7b571fb722cdd70a2f274ee771c547d15a75e6 --- /dev/null +++ b/2024/BeyondScene_ Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion/layout.json @@ -0,0 +1,7324 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 146 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 146 + ], + "type": "text", + "content": "BeyondScene: Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": "Gwanghyun Kim" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Hayeon Kim" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Hoigi Seo" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Dong Un Kang" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Se Young Chun" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "spans": [ + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "type": "text", + "content": "Dept. of Electrical and Computer Engineering, " + }, + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "type": "text", + "content": "INMC & IPAI Seoul National University, Republic of Korea" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 224, + 459, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 224, + 459, + 236 + ], + "spans": [ + { + "bbox": [ + 151, + 224, + 459, + 236 + ], + "type": "text", + "content": "{gwang.kim, khy5630, sehoiki3215, qkrtnsfk23, sychun}@snu.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 263, + 455, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 263, + 455, + 517 + ], + "spans": [ + { + "bbox": [ + 160, + 263, + 455, + 517 + ], + "type": "text", + "content": "Abstract. Generating higher-resolution human-centric scenes with details and controls remains a challenge for existing text-to-image diffusion models. This challenge stems from limited training image size, text encoder capacity (limited tokens), and the inherent difficulty of generating complex scenes involving multiple humans. While current methods attempted to address training size limit only, they often yielded human-centric scenes with severe artifacts. We propose BeyondScene, a novel framework that overcomes prior limitations, generating exquisite higher-resolution (over 8K) human-centric scenes with exceptional text-image correspondence and naturalness using existing pretrained diffusion models. BeyondScene employs a staged and hierarchical approach to initially generate a detailed base image focusing on crucial elements in instance creation for multiple humans and detailed descriptions beyond token limit of diffusion model, and then to seamlessly convert the base image to a higher-resolution output, exceeding training image size and incorporating details aware of text and instances via our novel instance-aware hierarchical enlargement process that consists of our proposed high-frequency injected forward diffusion and adaptive joint diffusion. BeyondScene surpasses existing methods in terms of correspondence with detailed text descriptions and naturalness, paving the way for advanced applications in higher-resolution human-centric scene creation beyond the capacity of pretrained diffusion models without costly retraining. Project page: https://janeyon.github.io/beyond-scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "spans": [ + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "text", + "content": "Keywords: Human-centric scene generation " + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "text", + "content": " Text-to-image diffusion model " + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "text", + "content": " High-resolution" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 586, + 230, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 230, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 230, + 598 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "type": "text", + "content": "Human-centric scene generation [8, 12, 15, 16, 19-23, 27, 28, 32, 33, 38, 40, 42, 44-46], has emerged as a critical research area with significant academic and industrial applications. While recent advancements utilizing text-to-image (T2I) diffusion" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 365, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 365, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 365, + 666 + ], + "type": "text", + "content": "* Authors contributed equally. † Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 473, + 413 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 473, + 413 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 473, + 413 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 473, + 413 + ], + "type": "image", + "image_path": "672bf6d39fbb5d8a3cf78b2843eb49f7347ba83a3116d7a182ae4c74a3e6b9b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 413, + 463, + 436 + ], + "lines": [ + { + "bbox": [ + 149, + 413, + 463, + 436 + ], + "spans": [ + { + "bbox": [ + 149, + 413, + 463, + 436 + ], + "type": "text", + "content": "\"In the sunny road, there are a clay animation character with lab coat, a clay character with red dotted dress and glasses, a clay character with short, chubby cheeks and glasses, a clay animation character with short, chubby checks wearing blue jeans, a clay character with spiky hair, a clay character with yellow T-shirts.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 445, + 482, + 501 + ], + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 501 + ], + "type": "text", + "content": "Fig. 1: BeyondScene pushes the boundaries of high-resolution human-centric scene generation. Unlike existing methods that often suffer from unrealistic scenes, anatomical distortions, and limited text-to-image correspondence, BeyondScene excels in 1) highly detailed scenes, 2) natural and diverse humans, 3) fine-grained control. The color in each text represents the text for each instance that has the same color in the pose map." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "models [2, 26, 29] have yielded promising results in generating controllable human scenes [12, 15, 23, 45], scaling these methods to handle larger and more complex scenes such as involving multiple humans remains a significant challenge. These limitations primarily stem from following causes. 1) Training image size: Directly sampling from a limited training image size can introduce artifacts and constrain the final scene resolution. 2) Limited text encoder tokens: The restricted number of tokens in text encoders (typically 77 for Stable Diffusion) hinders the inclusion of detailed descriptions for multiple instances within the scene. 3) Inherent limitations of T2I diffusion models: These models struggle to generate complex scenes with several human figures and intricate details." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "Existing attempts to address the training size issue, such as joint-process [6, 14, 47] and dilation-based methods [10], only partially address the problem. These methods often introduce new challenges specific to human-centric scenes as represented in Fig. 1, including: 1) Unrealistic scenes and objects: These methods can generate nonsensical scenarios with duplicated objects, humans defying gravity, and physically distorted environments. 2) Anatomical distortions: Generated scenes may exhibit unrealistic human anatomy, such as abnormal limbs or facial features. 3) Limited correspondence: Existing methods often fail to capture the complexity of scenes with multiple human instances and detailed descriptions, often generating single objects and lacking control over specific details like clothing or hairstyles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 263, + 483, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 483, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 483, + 324 + ], + "type": "text", + "content": "We propose BeyondScene, a novel framework that overcomes these limitations, generating high-resolution (over 8K) human-centric scenes with exceptional text-image correspondence and naturalness. BeyondScene employs a staged and hierarchical approach, which looks similar to a classical multi-resolution manner, but it is closer to how artists establish a foundation before adding details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 339, + 483, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 483, + 555 + ], + "type": "text", + "content": "Firstly, BeyondScene generates a detailed base image focusing on essential elements, human poses and detailed descriptions beyond the token limits. This initial stage enables detailed instance creation for multiple humans, surpassing the limitations of text encoders. Secondly, the method leverages an instance-aware hierarchical enlargement process to convert this base image to a higher resolution output beyond the training image size. Unlike naive super-resolution methods that simply scale resolution without considering text and instances, our approach refines content and adds more details aware of text and instances. This is achieved through our proposed novel techniques, including 1) high frequency-injected forward diffusion that addresses the issue of blurred low-quality results in image-to-image translation with the upsampled image by adaptively injecting high-frequency details into the upsampled image, enhancing the final output while preserving content and 2) adaptive joint diffusion that facilitates efficient and robust joint diffusion while maintaining control over human characteristics like pose and mask. This approach utilizes view-wise conditioning of text and pose information, along with a variable stride for the joint process. Furthermore, similar to how artists add details progressively, our method allows for the addition of details at different stages, exploiting the changing receptive field of each view." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "We comprehensively evaluate BeyondScene using qualitative and quantitative metrics, along with user studies. The results demonstrate significant improvements over existing methods in terms of 1) correspondence with detailed text descriptions and 2) naturalness and reduction of artifacts. Furthermore, we showcase the result of " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "8192 \\times 8192" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " image generation beyond 8K ultra high-resolution, demonstrating the capability for generating even higher-resolution images as displayed in Fig. 2. These advancements pave the way for exciting new applications in higher-resolution human-centric scene creation." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "BeyondScene" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 480, + 462 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 480, + 462 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 480, + 462 + ], + "type": "image", + "image_path": "58e3552bda3e6dd42e6614ddf0c60b36f5555c6a3b2bf4733604d06567f38fdb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 470, + 482, + 503 + ], + "lines": [ + { + "bbox": [ + 132, + 470, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 482, + 503 + ], + "type": "text", + "content": "Fig. 2: Beyond 8K ultra-high resolution image. This " + }, + { + "bbox": [ + 132, + 470, + 482, + 503 + ], + "type": "inline_equation", + "content": "8192 \\times 8192" + }, + { + "bbox": [ + 132, + 470, + 482, + 503 + ], + "type": "text", + "content": " image, generated by BeyondScene, surpasses the training resolution of SDXL by " + }, + { + "bbox": [ + 132, + 470, + 482, + 503 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 132, + 470, + 482, + 503 + ], + "type": "text", + "content": ", while exceeding the technical classification of 8K (" + }, + { + "bbox": [ + 132, + 470, + 482, + 503 + ], + "type": "inline_equation", + "content": "7680 \\times 4320" + }, + { + "bbox": [ + 132, + 470, + 482, + 503 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 525, + 234, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 234, + 538 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 234, + 538 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 550, + 323, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 550, + 323, + 562 + ], + "spans": [ + { + "bbox": [ + 133, + 550, + 323, + 562 + ], + "type": "text", + "content": "2.1 Controllable Human Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 569, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 482, + 665 + ], + "type": "text", + "content": "Early approaches to controllable human generation [8, 19-22, 27, 28, 40, 42, 44, 46] relied on pose guidance and source images, achieving success within specific scenarios but struggling with diverse scenes and arbitrary poses. Text-based conditioning methods emerged with the rise of large vision-language models [25], but initial attempts [16, 32, 38] faced limitations in vocabulary size and open vocabulary settings. Recent T2I diffusion [2, 24, 26, 30]-based methods like ControlNet [45], T2I-Adapter [23], GLIGEN [15], and HumanSD [12] introduced methods for incorporating diverse conditions, but scaling to larger scenes with" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "multiple individuals remains a challenge. This work proposes BeyondScene, a novel framework specifically designed to overcome these limitations and enable the generation of high-resolution human-centric scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 169, + 403, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 169, + 403, + 182 + ], + "spans": [ + { + "bbox": [ + 131, + 169, + 403, + 182 + ], + "type": "text", + "content": "2.2 Large Scene Generation Using Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 189, + 483, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 483, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 483, + 321 + ], + "type": "text", + "content": "Achieving high-resolution image generation [9, 26] presents significant hurdles. Training models from scratch or fine-tuning pretrained models [2, 7, 11, 24, 26, 30, 31, 34, 37, 43, 48] requires immense computational resources and struggles with the complexity of high-dimensional data. Recent exploration has ventured into training-free methods, with approaches like MultiDiffusion [6], SyncDiffusion [14] (joint diffusion), and ScaleCrafter [10] (dilation-based) emerging. However, these methods often introduce challenges specific to human-centric scenes, including unrealistic objects, anatomical distortions, and limited correspondence between text descriptions and generated images. BeyondScene, our proposed framework, addresses these limitations through detailed base image generation and instance-aware hierarchical enlargement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 339, + 233, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 233, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 233, + 354 + ], + "type": "text", + "content": "3 BeyondScene" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 435 + ], + "type": "text", + "content": "Current approaches to generating large scenes attempt to fill the entire canvas at once, often leading to challenges in capturing complex details and ensuring physical realism. Inspired by the artistic workflow of human painters, who establish a foundation with key elements and progressively refine them, we introduce BeyondScene, a novel framework that generates high-resolution human-centric images through a staged and hierarchical approach as represented in Fig. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": "BeyondScene operates in two key stages, 1) detailed base image generation (Sec. 3.1) that focuses on generating high-quality base images with precise control over human poses and detailed descriptions beyond typical token limitations and 2) instance-aware hierarchical enlargement (Sec. 3.2) where the base image is progressively enlarged while maintaining control and adding scene details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 514, + 326, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 326, + 526 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 326, + 526 + ], + "type": "text", + "content": "3.1 Detailed Base Image Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Generating human-centric scenes with multiple individuals using T2I diffusion models faces a key challenge: the limited number of tokens available in the text encoder restricts detailed descriptions to a few key elements. To overcome this, we generate detailed base image as illustrated in Fig. 3. Firstly, individual instances are generated in the training resolution of pose-guided T2I diffusion models [12, 15, 23, 45]. This allows for detailed descriptions beyond token limitations, tailored to each individual. Subsequently, the generated instances are segmented, cropped, and resized to the desired size. Next, a diffusion-based inpainting method is employed to create an instance-aware background that seamlessly integrates with the individual instances. This ensures a realistic and coherent background tailored to the specific scene. To address potential inconsistencies in brightness" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "BeyondScene" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 115, + 471, + 358 + ], + "blocks": [ + { + "bbox": [ + 142, + 115, + 471, + 358 + ], + "lines": [ + { + "bbox": [ + 142, + 115, + 471, + 358 + ], + "spans": [ + { + "bbox": [ + 142, + 115, + 471, + 358 + ], + "type": "image", + "image_path": "bf639b9fd424f6bca4a2b92dd0ecb315487dbb9e978d8336283b95e1d773f88d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 370, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 437 + ], + "type": "text", + "content": "Fig. 3: BeyondScene generates high-resolution images in two stages. First, individual instances are created using pose-guided T2I diffusion models, segmented, cropped, and placed onto an inpainted background. The tone of image is then normalized. In the second stage (illustrated in Fig. 4), this base image is progressively enlarged while maintaining detail and quality, effectively refining the image and adding further details leveraging high-frequency injected forward diffusion and adaptive joint diffusion (AJ)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "type": "text", + "content": "and color tones across individual elements, tone normalization through contrast limited adaptive histogram equalization (CLAHE) is performed. This involves first creating a grid histogram for pixel values within a specified grid, followed by adjusting the local contrast to align with the image's overall contrast. Notably, direct equalization in the presence of noise within the grid could induce the noise amplification. To prevent this, equalization is preceded by a redistribution process that includes limiting the maximum height of the grid histogram." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 562, + 371, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 371, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 371, + 574 + ], + "type": "text", + "content": "3.2 Instance-Aware Hierarchical Enlargement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": "BeyondScene introduces a novel approach to address the challenge of generating high-resolution outputs beyond the training image size through a instance-aware hierarchical enlargement process, which surpasses mere scaling by actively refining content and incorporating additional details aware of text and instance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "In the initial step, a low-resolution image is upsampled, and a forward diffusion process is applied to create a noisy upsampled image. Subsequently, a joint diffusion process refines the image to a high resolution over the training resolution." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 140, + 452, + 235 + ], + "blocks": [ + { + "bbox": [ + 147, + 118, + 329, + 128 + ], + "lines": [ + { + "bbox": [ + 147, + 118, + 329, + 128 + ], + "spans": [ + { + "bbox": [ + 147, + 118, + 329, + 128 + ], + "type": "text", + "content": "3.2 Instance-Aware Hierarchical Enlargement" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 147, + 129, + 295, + 138 + ], + "lines": [ + { + "bbox": [ + 147, + 129, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 129, + 295, + 138 + ], + "type": "text", + "content": "a. High frequency-injected forward diffusion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 140, + 452, + 235 + ], + "lines": [ + { + "bbox": [ + 155, + 140, + 452, + 235 + ], + "spans": [ + { + "bbox": [ + 155, + 140, + 452, + 235 + ], + "type": "image", + "image_path": "af0efa79b24c3839a184bb36065655b15dd1e50b9c440d83e3bd58b3e7ad0fc9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 143, + 252, + 467, + 340 + ], + "blocks": [ + { + "bbox": [ + 149, + 241, + 309, + 251 + ], + "lines": [ + { + "bbox": [ + 149, + 241, + 309, + 251 + ], + "spans": [ + { + "bbox": [ + 149, + 241, + 309, + 251 + ], + "type": "text", + "content": "b. Adaptive joint diffusion (AJ) At the diffusion timestep " + }, + { + "bbox": [ + 149, + 241, + 309, + 251 + ], + "type": "inline_equation", + "content": "t" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 252, + 467, + 340 + ], + "lines": [ + { + "bbox": [ + 143, + 252, + 467, + 340 + ], + "spans": [ + { + "bbox": [ + 143, + 252, + 467, + 340 + ], + "type": "image", + "image_path": "9c6c58c75699dce09680a47badcb2b2398098f112d481982d3571c04128c89af.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 350, + 482, + 405 + ], + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 405 + ], + "type": "text", + "content": "Fig. 4: Our instance-aware hierarchical enlargement involves two crucial processes: 1) High frequency-injected forward diffusion, which enables to achieve high resolution through a joint diffusion employing adaptive pixel perturbation. 2) Adaptive joint diffusion, dynamically regulating stride and conditioning of pose and text based on the presence of instances." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 514 + ], + "type": "text", + "content": "However, a drawback of this naive strategy is the induction of blurriness due to the model assigning sufficiently high likelihood to low-quality outputs. Also, this strategy often encounters issues such as unnatural duplication of objects and humans, stemming from the uniform application of the same text prompt to all views without pose conditioning. To mitigate these, the paper proposes a novel high frequency-injected forward diffusion and adaptive joint diffusion as illustrated in Fig. 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "High frequency-injected forward diffusion High frequency-injected forward diffusion enhances the translation of noisy latents from the upsampled image to high resolution with intricate details. This is achieved through a joint diffusion process employing adaptive pixel perturbation. This technique injects high-frequency details into the upsampled image based on a Canny edge map. This process enhances the final output's quality while preserving content. While simply adding pixel noise sharpens the image, it can also introduce flickering artifacts at the borders. Our adaptive approach based on the Canny map significantly improves image quality and reduces these artifacts. Specifically, we first up-sample the low-resolution image and calculate Canny map. Then Gaussian blur is applied to the map and normalized it to get probability map " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,j} \\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " is sampled" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "BeyondScene" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "on each pixel, " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "I_{i,j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " which denotes " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "-th column pixel of image " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,j} > C_{i,j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": ", the pixel value within " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "d_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " pixel distance is replaced so that we could apply pixel perturbation adaptively to avoid flickering artifacts near borders." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 169, + 483, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 169, + 483, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 169, + 483, + 194 + ], + "type": "text", + "content": "Adaptive joint diffusion To mitigate duplication of objects and humans in the naive joint diffusion, we introduce adaptive conditioning and adaptive stride." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 483, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 483, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 483, + 312 + ], + "type": "text", + "content": "Adaptive conditioning We introduce adaptive view-wise conditioning, which leverages the segmentation maps obtained from the base image generation stage. For each view, we check which human instances are present. If an instance is included, we add its pose and detailed description to the prompt used for that view. Additionally, if the detailed text description includes specific parts (head, face, upper body, etc.), we can apply these details to the corresponding regions using fine-grained segmentation maps. This facilitates efficient and robust joint diffusion while maintaining control over human characteristics like pose and appearance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 420 + ], + "type": "text", + "content": "Adaptive stride To effectively enhance the quality of the generated scene, we employ an adaptive stride for the joint process. We reduce the stride in regions containing humans, ensuring that these areas capture fine details. In contrast, the stride is increased in background regions, allowing for more efficient computation in these areas. This combination of adaptive view-wise conditioning and adaptive stride allows BeyondScene to effectively handle scenes with multiple humans while maintaining high-resolution detail and controllable image generation. The details and algorithms on our methods are presented in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 437, + 230, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 230, + 451 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 230, + 451 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 461, + 274, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 461, + 274, + 474 + ], + "spans": [ + { + "bbox": [ + 131, + 461, + 274, + 474 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": "We comprehensively evaluate our method using both qualitative and quantitative metrics. To ensure a fair comparison, all compared models are implemented using Stable Diffusion XL (SDXL) [24]-based architectures. We utilize SDXL-ControlNet-Openpose [3, 45], SDXL-inpainting [4], and Lang-Segment-Anything [1] for our pose-guided T2I diffusion, inpainting diffusion model, and segmentation models. The training resolution is set to " + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": ", while inference resolutions are varied across " + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "2048 \\times 2048" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "4 \\times 1:1" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "3584 \\times 2048" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "7 \\times 7:4" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "4096 \\times 2048" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "8 \\times 2:1" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "4096 \\times 4096" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "16 \\times 1:1" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "8192 \\times 8192" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "inline_equation", + "content": "64 \\times 1:1" + }, + { + "bbox": [ + 130, + 480, + 482, + 578 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Testing data Existing T2I generation datasets often lack large-scale scenes featuring multiple people with detailed descriptions for each individual. To address this limitation, we leverage the CrowdCaption dataset [35] as our primary test set. This dataset features images containing large crowds of people, allowing us to evaluate the effectiveness of our method in generating complex scenes with numerous individuals. For each image in the CrowdCaption dataset, we obtain" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 115, + 471, + 270 + ], + "blocks": [ + { + "bbox": [ + 143, + 115, + 471, + 270 + ], + "lines": [ + { + "bbox": [ + 143, + 115, + 471, + 270 + ], + "spans": [ + { + "bbox": [ + 143, + 115, + 471, + 270 + ], + "type": "image", + "image_path": "f11e7bd20de4f971039ab1b694fd0c4c7f4887a0531e9a26e839ff6b51ae8b8f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 270, + 467, + 292 + ], + "lines": [ + { + "bbox": [ + 149, + 270, + 467, + 292 + ], + "spans": [ + { + "bbox": [ + 149, + 270, + 467, + 292 + ], + "type": "text", + "content": "\"In empty stage of opera house, there are a ballerina with pink suit, a ballerina wearing blue ballet suit, a ballerina with pink suit, a ballerina with blue suit, a ballerina with yellow sparkling ballet suit, a ballerina with blue suit, a ballerina with pink suit, a ballerina with blue suit, a ballerina with pink suit.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 295, + 470, + 449 + ], + "blocks": [ + { + "bbox": [ + 143, + 295, + 470, + 449 + ], + "lines": [ + { + "bbox": [ + 143, + 295, + 470, + 449 + ], + "spans": [ + { + "bbox": [ + 143, + 295, + 470, + 449 + ], + "type": "image", + "image_path": "03a6eefb3b08aa41ef3a7c7a0d705d101a33bbe38ce23245cb38901dd28dc91b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 449, + 464, + 464 + ], + "lines": [ + { + "bbox": [ + 149, + 449, + 464, + 464 + ], + "spans": [ + { + "bbox": [ + 149, + 449, + 464, + 464 + ], + "type": "text", + "content": "\"In the alien planet with galaxy in the sky, there are a storm trooper, a storm trooper, a storm trooper, a Darth Vader, a storm trooper, a storm trooper, a storm trooper.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 475, + 482, + 552 + ], + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 552 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison for generating high-resolution human scenes (" + }, + { + "bbox": [ + 130, + 475, + 482, + 552 + ], + "type": "inline_equation", + "content": "3584 \\times 2048" + }, + { + "bbox": [ + 130, + 475, + 482, + 552 + ], + "type": "text", + "content": "). While existing approaches like T2I-Direct (SDXL [24]), T2I-Large (MultiDiffusion [6], SyncDiffusion [14], and ScaleCrafter [10]), and Visual+T2I (ControlNet [24, 45], T2IAdapter [23, 24], and R-MultiDiffusion [6]) models struggle with artifacts, our method achieves superior results by producing images with minimal artifacts, strong text-image correspondence, and a natural look. The color in each description represents the description for each instance that has the same color in the pose map." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "descriptions and individual poses using GRIT [36] and ViTPose [39]. Additionally, we obtain a global caption generated using GPT4 [5]. We filter the CrowdCaption images based on two criteria: number of humans (max 8 to accommodate token limits of baselines) and aspect ratio (close to 1:1 and 2:1), which results in a collection of 100 images for each aspect ratio (1:1 and 2:1). Additionally, to showcase the versatility of our method beyond the CrowdCaption dataset, we incorporate custom examples for qualitative comparisons." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "type": "text", + "content": "BeyondScene" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 422 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 422 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 422 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 422 + ], + "type": "image", + "image_path": "468f81c0bbd2d152feaf101be98e08e6060e4b6f489a500eed9add25293d3c18.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 430, + 482, + 509 + ], + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 509 + ], + "type": "text", + "content": "Fig. 6: Examples of large scene synthesis " + }, + { + "bbox": [ + 130, + 430, + 482, + 509 + ], + "type": "inline_equation", + "content": "(4096\\times 4096)" + }, + { + "bbox": [ + 130, + 430, + 482, + 509 + ], + "type": "text", + "content": " on the poses and text obtained from CrowdCaption [35] images. All baselines including direct high-resolution inference (SDXL [24]), T2I models for large scenes (MultiDiffusion [6], SyncDiffusion [14], ScaleCrafter [10]), Visual prior-guided T2I models (ControlNet [45], T2IAAdapter [23]) produce duplicated objects and artifacts in human anatomy, while our method succeeded in generation of high-resolution image with high text-image correspondence. Each color in the description corresponds to instances sharing the same color in the pose map." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Evaluation metrics For an evaluation of text-image correspondence, we adopt a multimodal large language model [5, 17, 18]-based text-image correspondence metric, as proposed in VIEScore [13], alongside the global CLIP [25] score that shows its limitations in producing reliable scores for complex scenes with detailed descriptions due to limitations in input image size and the available number of input tokens. This MLLM-based metric, powered by GPT4 [5], assigns a score ranging from 0 to 10, where 0 indicates no correspondence between the generated image and the prompt and 10 indicates perfect alignment. This metric has been demonstrably well-correlated with human judgments and provides explanations for the assigned scores [13]. To assess the naturalness of the generated images, we adapt the existing MLLM-based naturalness metric to focus on human-centric" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 157, + 447, + 474 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of our method with various approaches, including direct high-resolution inference (T2I-Direct) [24,41], T2I models designed for large scenes (T2I-Large) [6,10,14], and Visual prior-guided T2I models (Visual+T2I) [6,23,45]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 157, + 447, + 474 + ], + "lines": [ + { + "bbox": [ + 167, + 157, + 447, + 474 + ], + "spans": [ + { + "bbox": [ + 167, + 157, + 447, + 474 + ], + "type": "table", + "html": "
Model typesModelsGlobal CLIPMLLM (GPT4)-based
Text-image correspondenceGlobal naturalnessHuman naturalness
2048×2048 (4×1:1)
T2I-DirectSDXL0.3242.1411.8381.417
RPG0.2611.9632.1472.785
T2I-LargeMultiDiffusion0.3473.8553.3812.466
SyncDiffusion0.3473.6553.4292.607
ScaleCrafter0.3456.0815.6675.062
Visual+T2IControlNet0.2981.7901.4661.117
T2IAdapter0.3283.0942.6601.728
R-MultiDiffusion0.3113.6361.7111.597
BeyondScene0.3057.0416.5356.114
4096×2048 (8×2:1)
T2I-DirectSDXL0.3011.0610.7710.607
RPGOOMOOMOOMOOM
T2I-LargeMultiDiffusion0.3382.1362.2281.329
SyncDiffusion0.3423.1282.6651.591
ScaleCrafter0.2963.5523.5693.359
Visual+T2IControlNet0.2781.5261.0160.657
T2IAdapter0.2821.1091.0100.568
R-MultiDiffusion0.3063.4661.4221.201
BeyondScene0.3047.1186.6126.375
4096×4096 (16×1:1)
T2I-DirectSDXL0.2770.3290.4110.118
RPGOOMOOMOOMOOM
T2I-LargeMultiDiffusion0.3191.3052.4991.591
SyncDiffusion0.3100.9122.3091.359
ScaleCrafter0.3122.9623.0482.442
Visual+T2IControlNet0.2440.8111.3090.729
T2IAdapter0.2690.4910.7190.352
R-MultiDiffusion0.2811.7920.8600.708
BeyondScene0.3016.8016.0745.627
", + "image_path": "ca1e15be944d6e55fd3e8b5cbe14ac255907627303cfe14e2af0704002cb750f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 496, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 581 + ], + "type": "text", + "content": "scenarios, leading to two separate scores: Global naturalness scores from 0 to 10, with 0 indicating an unnatural or unrealistic image and 10 indicating a completely natural image. This score considers factors such as overall inconsistencies, unrealistic physics, disconnectivity. Human naturalness scores from 0 to 10, with 0 indicating unnatural human anatomy, and 10 indicating the absence of any artifacts in the human anatomy. Additional details on implementation and evaluation are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 597, + 194, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 194, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 194, + 609 + ], + "type": "text", + "content": "4.2 Result" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Our method is compared against three groups of existing SOTA approaches: 1) Direct high-resolution inference using T2I models (T2I-Direct) that directly perform high-resolution inference of SDXL [24] and RPG [41] over their training image size, 2) T2I diffusion models for large scenes (T2I-Large) that encom" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "BeyondScene" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 157, + 440, + 288 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 2: Quantitative comparison by scene complexity. Our method shows robust performance (averaged across resolutions: " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "2048 \\times 2048" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "4096 \\times 2048" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "4096 \\times 4096" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": ") even with increasing the number of humans, highlighting its ability to handle complex scenes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 157, + 440, + 288 + ], + "lines": [ + { + "bbox": [ + 176, + 157, + 440, + 288 + ], + "spans": [ + { + "bbox": [ + 176, + 157, + 440, + 288 + ], + "type": "table", + "html": "
Models2~4 humans5~8 humans
MLLM (GPT4)-basedMLLM (GPT4)-based
Text-image corr.Global nat.Human nat.Text-image corr.Global nat.Human nat.
SDXL1.1921.0390.8451.1610.9740.853
RPG3.7573.1383.0130.1691.1562.558
MultiDiffusion2.1422.3921.6922.7223.0131.911
SyncDiffusion2.2772.4281.6922.8533.1752.013
ScaleCrafter4.554.4374.0643.8463.7523.177
ControlNet1.1721.1940.8101.5791.3330.858
T2IAdapter1.5211.4490.9641.6081.4860.801
R-MultiDiffusion3.271.2861.3292.6591.3761.007
BeyondScene7.3086.4966.3766.6666.3195.701
", + "image_path": "d626ce1f95af8ec578d524beef57652468aad03567cf809b85a8c71670f7c15e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 222, + 341, + 389, + 414 + ], + "blocks": [ + { + "bbox": [ + 130, + 298, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 332 + ], + "type": "text", + "content": "Table 3: User study result. Users consistently preferred our method over baselines, with MLLM scores closely aligning with human choices. However, the global CLIP score diverged, suggesting potential limitations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 222, + 341, + 389, + 414 + ], + "lines": [ + { + "bbox": [ + 222, + 341, + 389, + 414 + ], + "spans": [ + { + "bbox": [ + 222, + 341, + 389, + 414 + ], + "type": "table", + "html": "
ModelsPreference of Ours ↑
Text-image corr.Global Human nat.
vs ScaleCrafter0.7950.6500.806
vs MultiDiffusion0.7270.6580.822
vs ControlNet0.8700.6940.897
", + "image_path": "95f8a56712669a6f0bc14e87fa7ad77b359911d6837af37bfbabfc8d2c9d8688.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 435, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 509 + ], + "type": "text", + "content": "pass joint diffusion-based methods (MultiDiffusion [6], SyncDiffusion [14]), and ScaleCrafter [10].) Visual prior-guided T2I generation models (Visual+T2I) that focuses on models including SDXL-ControlNet-Openpose [24, 45], SDXL-T2IAAdapter [23, 24], and R-MultiDiffusion [6] that leverage visual priors for T2I generation as our method. We provide more results and ablation studies in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 526, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 610 + ], + "type": "text", + "content": "Qualitative comparison As represented in Fig. 1, 5 and 6, baselines introduce noticeable instance duplication, artifacts in anatomy, facial features, making the scene unnatural rather than reflecting real-world physics. Also, they struggle to generate complex scenes corresponding to the detailed descriptions. Conversely, our method generates images with minimal artifacts, demonstrating superior text-image correspondence, global and human naturalness, particularly in human-centric scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Quantitative comparison Table 1 summarizes the performance across the resolution settings of " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "2048 \\times 2048" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "4 \\times 1:1" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "4096 \\times 2048" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "8 \\times 2:1" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "4096 \\times 4096" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "16 \\times 1:1" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "). Our method consistently outperforms baselines on all metrics except" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 202, + 165 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 202, + 165 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 202, + 165 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 202, + 165 + ], + "type": "image", + "image_path": "ebab3973c207d74f9d1afe68c5ed99a9f6958dbeca1b2d54f1f89239430f1760.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 167, + 199, + 174 + ], + "lines": [ + { + "bbox": [ + 156, + 167, + 199, + 174 + ], + "spans": [ + { + "bbox": [ + 156, + 167, + 199, + 174 + ], + "type": "text", + "content": "(a) Original image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 114, + 255, + 164 + ], + "blocks": [ + { + "bbox": [ + 203, + 114, + 255, + 164 + ], + "lines": [ + { + "bbox": [ + 203, + 114, + 255, + 164 + ], + "spans": [ + { + "bbox": [ + 203, + 114, + 255, + 164 + ], + "type": "image", + "image_path": "b3d88ff5f10bd423587f714d6bb84bee38d6a8cd37c41ad123785f57122bc3c7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 167, + 253, + 174 + ], + "lines": [ + { + "bbox": [ + 202, + 167, + 253, + 174 + ], + "spans": [ + { + "bbox": [ + 202, + 167, + 253, + 174 + ], + "type": "text", + "content": "(b) 64x down-sampled" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 114, + 307, + 164 + ], + "blocks": [ + { + "bbox": [ + 256, + 114, + 307, + 164 + ], + "lines": [ + { + "bbox": [ + 256, + 114, + 307, + 164 + ], + "spans": [ + { + "bbox": [ + 256, + 114, + 307, + 164 + ], + "type": "image", + "image_path": "e5001ee46066886bbce1f62b5155721853b98e2c1354a5749ab31ba409beca21.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 167, + 356, + 174 + ], + "lines": [ + { + "bbox": [ + 257, + 167, + 356, + 174 + ], + "spans": [ + { + "bbox": [ + 257, + 167, + 356, + 174 + ], + "type": "text", + "content": "mage (L) / Its part (R) (c) Bilinear interpol." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 114, + 359, + 164 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 359, + 164 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 359, + 164 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 359, + 164 + ], + "type": "image", + "image_path": "42ebeae8fa2d34185d3c86930c12f5144be7a333cdd8a0bd5861f9dd74f675fe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 167, + 396, + 174 + ], + "lines": [ + { + "bbox": [ + 366, + 167, + 396, + 174 + ], + "spans": [ + { + "bbox": [ + 366, + 167, + 396, + 174 + ], + "type": "text", + "content": "(d) Naive PP" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 309, + 167, + 356, + 174 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 167, + 356, + 174 + ], + "spans": [ + { + "bbox": [ + 309, + 167, + 356, + 174 + ], + "type": "text", + "content": "(c) Bilinear interpol." + } + ] + } + ], + "index": 9, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 359, + 114, + 409, + 164 + ], + "blocks": [ + { + "bbox": [ + 359, + 114, + 409, + 164 + ], + "lines": [ + { + "bbox": [ + 359, + 114, + 409, + 164 + ], + "spans": [ + { + "bbox": [ + 359, + 114, + 409, + 164 + ], + "type": "image", + "image_path": "396b37184c12fbc7c81b30616c51b7554ffa503706e18b8743ec179175149ea8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 167, + 461, + 174 + ], + "lines": [ + { + "bbox": [ + 410, + 167, + 461, + 174 + ], + "spans": [ + { + "bbox": [ + 410, + 167, + 461, + 174 + ], + "type": "text", + "content": "(e) Adaptive PP (Ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 462, + 165 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 462, + 165 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 462, + 165 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 462, + 165 + ], + "type": "image", + "image_path": "600c8abbe8eece1ed5eaf4d03cb651d1cca7198b0623fe7622544ac789be9296.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 149, + 251, + 465, + 491 + ], + "blocks": [ + { + "bbox": [ + 130, + 183, + 482, + 250 + ], + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 250 + ], + "type": "text", + "content": "Fig. 7: Forward-reverse diffusion process with interpolation methods. (a) SDXL generates an image with the prompt \"figurine of Superman\". (b) The image is then " + }, + { + "bbox": [ + 130, + 183, + 482, + 250 + ], + "type": "inline_equation", + "content": "\\times 64" + }, + { + "bbox": [ + 130, + 183, + 482, + 250 + ], + "type": "text", + "content": " down-sampled (Left) with a close-up view (Right). (c) Upsampled image with bilinear interpolation and forward-reverse diffusion exhibits blurry details. (d) Naive pixel perturbation (PP) yields sharper details but introduces flickering artifacts. (e) Our Canny map-based adaptive PP produces clear and highly detailed image." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 251, + 465, + 491 + ], + "lines": [ + { + "bbox": [ + 149, + 251, + 465, + 491 + ], + "spans": [ + { + "bbox": [ + 149, + 251, + 465, + 491 + ], + "type": "image", + "image_path": "0e014212246c6b20a3e3acb2149d7c5041f825f3709d630858cb6e32b04c0efd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 498, + 482, + 544 + ], + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 544 + ], + "type": "text", + "content": "Fig. 8: Effectiveness of adaptive conditioning. Omitting either adaptive pose or text conditioning leads to issues like inconsistent poses, bad anatomy (yellow boxes), unwanted object duplication (red boxes), and mixed descriptions. Each color in the description corresponds to instances sharing the same color in the pose map." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 576, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 625 + ], + "type": "text", + "content": "the global CLIP score, which shows inconsistencies across methods. This suggests that our approach effectively enhances the original generation capabilities of the pretrained diffusion model for high-resolution image generation, achieving superior text-image correspondence, naturalness." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Table 3 further analyzes the performance of different methods based on the number of people depicted in the scene. Our approach demonstrates robustness, effectively managing complexity as people count rises." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "BeyondScene" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "User study We conducted a user study to evaluate the faithfulness and naturalness of images generated by three methods (one from each group) and our method. Participants compared large-scene images based on the same criteria used in the MLLM metrics: text-image correspondence, global naturalness, and human naturalness. A total of 12,120 responses were collected from 101 participants. The result shows that our method was consistently preferred over baselines, with a larger margin for all attributes. Notably, MLLM scores demonstrate strong alignment with human scores, supporting their validity. However, the global CLIP score appears misaligned, suggesting its limitations in capturing human judgment. We provide the additional details in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 324 + ], + "type": "text", + "content": "Beyond 8K ultra-high-resolution image As displayed in Fig. 2, BeyondScene excels in generating unparalleled resolution, producing images at an impressive " + }, + { + "bbox": [ + 130, + 251, + 482, + 324 + ], + "type": "inline_equation", + "content": "8192 \\times 8192" + }, + { + "bbox": [ + 130, + 251, + 482, + 324 + ], + "type": "text", + "content": " dimensions. This achievement surpasses the training resolution of SDXL by a remarkable factor of " + }, + { + "bbox": [ + 130, + 251, + 482, + 324 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 130, + 251, + 482, + 324 + ], + "type": "text", + "content": ". It surpasses the standard 8K resolution " + }, + { + "bbox": [ + 130, + 251, + 482, + 324 + ], + "type": "inline_equation", + "content": "(7680 \\times 4320)" + }, + { + "bbox": [ + 130, + 251, + 482, + 324 + ], + "type": "text", + "content": ", demonstrating the model's capability to produce images with exceptional detail at even higher resolutions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 339, + 238, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 238, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 238, + 352 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "type": "text", + "content": "High frequency-injected forward diffusion As represented in 7, forward diffusion and reverse generation with naively up-sampled image lead to increased blurring in the image. By adopting the pixel perturbation, it makes the result sharpened. However, it introduced flickering artifacts at the image borders. Applying adaptive pixel perturbation based on the Canny map significantly improves image quality and reduces flickering artifacts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 445, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 542 + ], + "type": "text", + "content": "Adaptive conditioning Fig. 8 demonstrates the effectiveness of adaptive conditioning. Omitting adaptive pose conditioning leads to inconsistent poses, bad anatomy, and decreased controllability compared to the desired input pose (yellow boxes). Excluding adaptive text conditioning, where each view receives the same full text including global and all instance descriptions, results in unwanted object duplication (red boxes) that ignores existing object scales, creating an awkward scene. Additionally, the descriptions of each instance become mixed, resulting in low text-image correspondence and decreased controllability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 558, + 220, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 220, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 220, + 571 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We introduce BeyondScene, a novel framework capable of generating high-resolution human-centric scenes with exceptional text-image correspondence resolving artifacts beyond the token limits and beyond the training image size. BeyondScene overcomes limitations in generating complex human-centric scenes with detailed control, surpassing existing methods in generating high-fidelity scenes with control over human characteristics and number, while maintaining naturalness and reducing artifacts." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 234 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) [NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)] and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF-2022R1A4A1030579, NRF-2022M3C1A309202211). Also, the authors acknowledged the financial support from the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 251, + 197, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 251, + 197, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 251, + 197, + 264 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 275, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 275, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 275, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 275, + 481, + 297 + ], + "type": "text", + "content": "1. Language segment anything. https://github.com/luca-medeiros/lang-segment-anything" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 297, + 326, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 326, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 326, + 308 + ], + "type": "text", + "content": "2. Midjourney. https://www.midjourney.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 308, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 481, + 329 + ], + "type": "text", + "content": "3. Sdsl-controlnet: Openpose (v2). https://huggingface.co/thibaud/controlnet-openpose-sdsl-1.0" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 329, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 329, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 329, + 481, + 350 + ], + "type": "text", + "content": "4. Sdxl inpainting 0.1. https://huggingface.co/diffusers/stable-diffusion-x1-1.0-inpainting-0.1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 351, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 481, + 384 + ], + "type": "text", + "content": "5. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 384, + 481, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 481, + 405 + ], + "type": "text", + "content": "6. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv:2302.08113 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 405, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 405, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 405, + 481, + 426 + ], + "type": "text", + "content": "7. Chen, T.: On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 426, + 481, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 481, + 459 + ], + "type": "text", + "content": "8. Cheong, S.Y., Mustafa, A., Gilbert, A.: KPE: Keypoint pose encoding for transformer-based image generation. In: British Machine Vision Conference (BMVC) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 459, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 481, + 491 + ], + "type": "text", + "content": "9. Ding, Z., Zhang, M., Wu, J., Tu, Z.: Patched denoising diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 535 + ], + "type": "text", + "content": "0. He, Y., Yang, S., Chen, H., Cun, X., Xia, M., Zhang, Y., Wang, X., He, R., Chen, Q., Shan, Y.: Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 535, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 481, + 556 + ], + "type": "text", + "content": "1. Hoogeboom, E., Heek, J., Salimans, T.: simple diffusion: End-to-end diffusion for high resolution images. arXiv preprint arXiv:2301.11093 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 556, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 481, + 589 + ], + "type": "text", + "content": "2. Ju, X., Zeng, A., Zhao, C., Wang, J., Zhang, L., Xu, Q.: Humansd: A native skeleton-guided diffusion model for human image generation. arXiv preprint arXiv:2304.04269 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 622 + ], + "type": "text", + "content": "3. Ku, M., Jiang, D., Wei, C., Yue, X., Chen, W.: Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "text", + "content": "4. Lee, Y., Kim, K., Kim, H., Sung, M.: Syncdiffusion: Coherent montage via synchronized joint diffusions. arXiv:2306.05178 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "5. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: Openset grounded text-to-image generation. arXiv preprint arXiv:2301.07093 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "BeyondScene" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Liu, D., Wu, L., Zheng, F., Liu, L., Wang, M.: Verbal-person nets: Pose-guided multi-granularity language-to-person generation. IEEE Transactions on Neural Networks and Learning Systems (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 173 + ], + "type": "text", + "content": "17. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 174, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 481, + 206 + ], + "type": "text", + "content": "18. Liu, S., Cheng, H., Liu, H., Zhang, H., Li, F., Ren, T., Zou, X., Yang, J., Su, H., Zhu, J., et al.: Llava-plus: Learning to use tools for creating multimodal agents. arXiv preprint arXiv:2311.05437 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 251 + ], + "type": "text", + "content": "19. Lv, Z., Li, X., Li, X., Li, F., Lin, T., He, D., Zuo, W.: Learning semantic person image generation by region-adaptive normalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10806-10815 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "text", + "content": "20. Ma, L., Jia, X., Sun, Q., Schiele, B., Tuytelaars, T., Van Gool, L.: Pose guided person image generation. vol. 30 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "21. Ma, T., Peng, B., Wang, W., Dong, J.: MUST-GAN: Multi-level statistics transfer for self-driven person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13622-13631 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "22. Men, Y., Mao, Y., Jiang, Y., Ma, W.Y., Lian, Z.: Controllable person image synthesis with attribute-decomposed GAN. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5084-5093 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "23. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "24. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 453 + ], + "type": "text", + "content": "25. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "26. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 530 + ], + "type": "text", + "content": "27. Ren, Y., Fan, X., Li, G., Liu, S., Li, T.H.: Neural texture extraction and distribution for controllable person image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13535-13544 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "text", + "content": "28. Ren, Y., Yu, X., Chen, J., Li, T.H., Li, G.: Deep image spatial transformation for person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7690-7699 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 564, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 481, + 587 + ], + "type": "text", + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 609 + ], + "type": "text", + "content": "30. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "32. Roy, P., Ghosh, S., Bhattacharya, S., Pal, U., Blumenstein, M.: TIPS: Text-induced pose synthesis. In: European Conference on Computer Vision (ECCV). pp. 161-178. Springer (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 621 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "33. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "34. Teng, J., Zheng, W., Ding, M., Hong, W., Wangni, J., Yang, Z., Tang, J.: Relay diffusion: Unifying diffusion process across resolutions for image synthesis. arXiv preprint arXiv:2309.03350 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "35. Wang, L., Li, H., Hu, W., Zhang, X., Qiu, H., Meng, F., Wu, Q.: What happens in crowd scenes: A new dataset about crowd scenes for image captioning. IEEE Transactions on Multimedia (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "36. Wu, J., Wang, J., Yang, Z., Gan, Z., Liu, Z., Yuan, J., Wang, L.: Grit: A generative region-to-text transformer for object understanding. arXiv:2212.00280 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "37. Xie, E., Yao, L., Shi, H., Liu, Z., Zhou, D., Liu, Z., Li, J., Li, Z.: Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. arXiv preprint arXiv:2304.06648 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "38. Xu, X., Chen, Y.C., Tao, X., Jia, J.: Text-guided human image manipulation via image-text shared space. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) 44(10), 6486-6500 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "39. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. NeurIPS 35, 38571-38584 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "40. Yang, F., Lin, G.: CT-Net: Complementary transferring network for garment transfer with arbitrary geometric changes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9899-9908 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "text", + "content": "41. Yang, L., Yu, Z., Meng, C., Xu, M., Ermon, S., Cui, B.: Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. arXiv preprint arXiv:2401.11708 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "42. Yang, L., Wang, P., Liu, C., Gao, Z., Ren, P., Zhang, X., Wang, S., Ma, S., Hua, X., Gao, W.: Towards fine-grained human pose transfer with detail replenishing network. IEEE Transactions on Image Processing 30, 2422-2435 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "43. Zhang, B., Gu, S., Zhang, B., Bao, J., Chen, D., Wen, F., Wang, Y., Guo, B.: Styleswin: Transformer-based gan for high-resolution image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11304-11314 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "44. Zhang, J., Li, K., Lai, Y.K., Yang, J.: PISE: Person image synthesis and editing with decoupled GAN. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7982-7990 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "45. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "46. Zhang, P., Yang, L., Lai, J.H., Xie, X.: Exploring dual-task correlation for pose guided person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7713-7722 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "47. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. arXiv:2303.17076 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "48. Zheng, Q., Guo, Y., Deng, J., Han, J., Li, Y., Xu, S., Xu, H.: Any-size-diffusion: Toward efficient text-driven synthesis for any-size hd images. arXiv preprint arXiv:2308.16582 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "BeyondScene" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_content_list.json b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89a65227c2a36405d81f787cd49042d50d77a786 --- /dev/null +++ b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_content_list.json @@ -0,0 +1,1762 @@ +[ + { + "type": "text", + "text": "Bi-TTA: Bidirectional Test-Time Adapter for Remote Physiological Measurement", + "text_level": 1, + "bbox": [ + 241, + 141, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haodong Li $^{1}$ , Hao Lu $^{1}$ , and Ying-Cong Chen $^{1,2}$", + "bbox": [ + 328, + 210, + 673, + 228 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Hong Kong University of Science and Technology (Guangzhou)", + "$^{2}$ Hong Kong University of Science and Technology {hli736, hlu585}@connect.hkust-gz.edu.cn yingcongchen@ust.hk" + ], + "bbox": [ + 272, + 238, + 728, + 281 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6389c687b9a3b491fd6d49f3b19bf8e53b6f2e798a18fcd56311ab5aaccbac75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 333, + 395, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e04e8c0ee2cfdd7c4d6b29f3c2397700fd0a23b702a681053ad9fa7808826c06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 301, + 581, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6640b9c4e5ba28d39343e9493e0ed97ddbe9d668b0e52af050477b93b2117a84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 301, + 759, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f86273366369275de54b2e9a6c74260f00b6731a67c594228c23c85f25a87e1e.jpg", + "image_caption": [ + "(a) Pre-trained" + ], + "image_footnote": [], + "bbox": [ + 233, + 429, + 406, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3fcb87f0ce485180fbb04f219c41a0353f8e1701f0840dce5b6fc852becd0c17.jpg", + "image_caption": [ + "(b) Priors-based TTA" + ], + "image_footnote": [], + "bbox": [ + 415, + 430, + 588, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db83430207fa4fcb3793b29689e4a618b70ae5ba2ae8779f32013a62faf13e1a.jpg", + "image_caption": [ + "(c) Priors-based Bi-TTA", + "Fig.1: Visualization of the mean absolute error (MAE) result on down-sampled VIPL dataset using (a) the pre-trained rPPG model; (b) priors-based test-time adapted model; and (c) priors-based bidirectionally test-time adapted model. These visualizations demonstrate that our proposed priors and the Bi-TTA framework significantly enhance the generalization performance (reflected by the flatness of MAE field [26]) beyond the pre-trained model." + ], + "image_footnote": [], + "bbox": [ + 594, + 430, + 767, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Remote photoplethysmography (rPPG) is gaining prominence for its non-invasive approach to monitoring physiological signals using only cameras. Despite its promise, the adaptability of rPPG models to new, unseen domains is hindered due to the environmental sensitivity of physiological signals. To address this issue, we pioneer the Test-Time Adaptation (TTA) in rPPG, enabling the adaptation of pre-trained models to the target domain during inference, sidestepping the need for annotations or source data due to privacy considerations. Particularly, utilizing only the user's face video stream as the accessible target domain data, the rPPG model is adjusted by tuning on each single instance it encounters. However, 1) TTA algorithms are designed predominantly for classifi-", + "bbox": [ + 259, + 686, + 746, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cation tasks, ill-suited in regression tasks such as rPPG due to inadequate supervision. 2) Tuning pre-trained models in a single-instance manner introduces variability and instability, posing challenges to effectively filtering domain-relevant from domain-irrelevant features while simultaneously preserving the learned information. To overcome these challenges, we present Bi-TTA, a novel expert knowledge-based Bidirectional Test-Time Adapter framework. Specifically, leveraging two expert-knowledge priors for providing self-supervision, our Bi-TTA primarily comprises two modules: a prospective adaptation (PA) module using sharpness-aware minimization to eliminate domain-irrelevant noise, enhancing the stability and efficacy during the adaptation process, and a retrospective stabilization (RS) module to dynamically reinforce crucial learned model parameters, averting performance degradation caused by overfitting or catastrophic forgetting. To this end, we established a large-scale benchmark for rPPG tasks under TTA protocol, promoting advancements in both the rPPG and TTA fields. The experimental results demonstrate the significant superiority of our approach over the state-of-the-art (SoTA).", + "bbox": [ + 261, + 147, + 743, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Bidirectional Test-Time Adaptation $\\cdot$ Expert Knowledge-based Priors $\\cdot$ Remote Photoplethysmography", + "bbox": [ + 261, + 410, + 740, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 486, + 375, + 502 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The cardiac cycle refers to the process experienced by the cardiovascular system from the start of one heartbeat to the beginning of the next. Through the analysis of the cardiac cycle, important physiological indicators such as heart rate (HR), respiration rate (RF), and heart rate variability (HRV) can be derived, playing crucial roles in medical diagnostics, health surveillance, and forensic analysis [20, 36, 55].", + "bbox": [ + 212, + 517, + 787, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional physiological measurement devices like electrocardiograms (ECG), heart rate bands, and finger-clips, despite their accuracy, are often expensive and uncomfortable to wear. In contrast, rPPG offers a non-invasive and more convenient alternative by extracting blood volume pulses (BVP) from facial videos, analyzing the skin's light absorption variations to measure HR, HRV, and RF [44,69,70]. These monitors are especially important for tracking health status and sympathetic activity levels [20,36,55]. Nowadays, physiological measurements based on rPPG, using only ordinary cameras, are increasingly becoming the focus of research in computer vision community [5,9,33,35,43,47,50]. However, the periodic color changes in the skin captured by remote cameras are quite subtle, and variables like natural scene illumination and head movements can significantly impair the reliability of physiological indicator measurements.", + "bbox": [ + 212, + 607, + 787, + 789 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 This dataset is 1/20 equally spaced down-sampled from the full VIPL dataset [42], and contains 8349 samples in total.", + "bbox": [ + 217, + 796, + 785, + 825 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2 This visualization is generated following [26].", + "bbox": [ + 218, + 825, + 532, + 839 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/10619055dd8175ecb6a1b49574c013f82ad76c4aaa2dcdcf5dd754293df96f9c.jpg", + "image_caption": [ + "Fig. 2: Visualization of Domain Adaptation (DA) and Test-Time Adaptation (TTA) methodologies. DA utilizes batch learning with labeled target data, while TTA dynamically refines the model during inference without relying on target data labels or distribution. Both DA and TTA obviate the requirement for source data. Note that domain generalization (DG) is excluded as it does not focus on specific target domain adaptation." + ], + "image_footnote": [], + "bbox": [ + 289, + 143, + 712, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conventional rPPG methods [6,25,37,46,60,64] perform well only in constrained environments. Recent deep learning (DL) methods [2,43,49,53,68] have demonstrate promising results, however, the most of training data are collected in controlled laboratory environments with limited variations in illumination, camera parameters, etc. Therefore, these methods still struggle to generalize to unseen domains, necessitating model adaptation.", + "bbox": [ + 212, + 426, + 787, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For privacy considerations, the adaptation of rPPG models in unseen scenarios is required to be performed without accessing the source data, and target data annotations or distributions (i.e., large batches of target data samples). Addressing this, we firstly implement Test-Time Adaptation (TTA) in the field of rPPG. TTA aims to fine-tune a pre-trained source model during inference time, as illustrated in Fig. 2, without accessing the distribution and labeling of both source data and target data and naturally eliminate the need for intensive re-training [61]. This innovation allows for the automatic adaptation of pretrained rPPG models to unseen domains during inference, without the necessity for labelling or batches of target data.", + "bbox": [ + 212, + 520, + 787, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, applying TTA directly into rPPG faces two key limitations. 1) Predominantly designed for classification tasks, most TTA algorithms leverage entropy in normalization layers or pseudo labels. These strategies are not inherently suited for regression tasks like rPPG, leading to a lack of appropriate supervision. 2) In real-world scenarios, the pre-trained rPPG model is tuned during inference the user's face video stream, on an instance-by-instance context. This variability and instability introduced by biased instances pose challenges to effectively filtering out domain-irrelevant noise and simultaneously preserving crucial learned information as the model acquiring new knowledge. This nuanced balance is critical in preventing slow or ineffective adaptation under single-instance learning condition, ensuring an efficient and consistent adaptation process.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bi-TTA", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To navigate these challenges, we proposed a novel Bidirectional Test-Time Adapter (Bi-TTA) framework, grounded on two expert knowledge-based self-supervised priors. Firstly, two regularizations are proposed for offering effective supervision especially in rPPG tasks, addressing the aforementioned ill-suited problem of applying existing TTA algorithms into regression tasks. These two priors focus on temporal and spatial consistency respectively, which is derived from the characteristics of physiological signals. Then, two strategies are introduced for stable and effective fine-tuning process. The prospective adaptation (PA) strategy is designed for filtering out the irrelevant information of each encountered instance, aiming that only target-domain-essential feature is truly contributing to model adaptation. The retrospective stabilization (RS) strategy backtracks and consolidates essential model parameters dynamically, leveraging previous update gradients during the self-supervised tuning process, ensuring that the learned adaptation ability is preserved intact. We evaluate our method across various datasets, benchmarking against existing prevailing TTA algorithms, demonstrating the superiority of our Bi-TTA. Overall, our contributions can be summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the TTA protocol to rPPG tasks, paving the way for a new paradigm in self-supervised adapting of rPPG models during inference.", + "- We introduce two novel self-supervised priors to regularize temporal and spatial consistency, providing appropriate supervision.", + "- We propose a novel Bidirectional Test-Time Adapter (Bi-TTA) framework grounded on the introduced priors, which promotes effective and efficient adaptation, with enhanced stability.", + "- We established a large-scale benchmark for rPPG tasks under TTA protocol. Extensive experiments demonstrate significantly improved adaptation capacity of our approach." + ], + "bbox": [ + 225, + 412, + 784, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 585, + 397, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Remote Physiological Measurement", + "text_level": 1, + "bbox": [ + 215, + 618, + 555, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remote physiological measurement aims to quantify HR and HRV from BVP signals by interpreting skin chrominance fluctuations. Traditional techniques typically involve color space transformations or signal decomposition to obtain the BVP signal with high signal-to-noise ratio (SNR) [6, 7, 46, 65]. With powerful modeling capacity of deep learning (DL), there's been a significant shift towards employing DL models for remote physiological estimation, yielding impressive results. [4, 16, 41, 43, 52, 66, 69]. Many of these methods extract BVP signals from face videos typically by aggregating spatial and temporal information [44, 52, 69, 70]. Besides, many works utilize hand-designed 2D feature maps extracted from face videos to alleviate computational demands [16, 33, 43, 48, 51]. Recently, researchers have turned to exploring how to leverage task-specific knowledge and how to improve the generalization performance of the model into unseen scenarios.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-specific priors. rPPG is distinguished by its unique biological and physical mechanisms, which can be exploited to enhance the physiological signal extraction ability of the model [17,31,58,62]. These task-specific priors can be summed up as 1) spatial consistency: BVP signals extracted from different face areas should exhibit similar; 2) temporal consistency: BVP signals typically exhibit smooth and gradual changes over short time period; 3) heart rate range: The heart rate typically ranges between 40 and 250 beats per minute (bpm). Different from existing works, we provide a more granular solution leveraging the multi-scale latent features, rather than purely focusing on the prediction output of BVP signals.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The generalization of the model. With the domain shifts (e.g., lighting, head motion, environment, etc.), most methods tend to have significant performance degradation. To improve the generalization performance on unseen domains, NEST-rPPG established a large-scale DG evaluation protocol [34]. These DG methods, requiring the access of all source data, do not focus on the adaptation of a specific target domain, leading to sub-optimal performance [56]. Besides, some work utilizes the labels and distribution of the target domain data to improve the model's adaptation performance [10,24,30], which is impractical due to accessing the unseen domain in advance. Differently from the above rPPG methods, we address this domain gap by using the TTA protocol, facilitating the adaptation to specific target domains during inference, based solely on unlabeled instances from the target domain.", + "bbox": [ + 212, + 297, + 789, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 215, + 507, + 444, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test-time Adaptation (TTA), an emerging protocol, aims to adjust a pre-trained source model during inference by learning from the unlabeled target data, without accessing the distributions and labels of both source and target data, thereby naturally eliminate the need of intensive re-training [27, 61, 63]. Current TTA methods typically require aggregating a large number of samples from the target data to form a batch, and perform adaptation through 1) normalization calibration [1, 21, 22, 39, 74], 2) pseudo-labeling [8, 14], 3) input adaptation [15, 73], 4) dynamic inference that learns the network weights from a set of model parameters [72], 5) meta-learning [12], etc. These approaches can be considered as different forms of source-free domain adaptation (SFDA) [28] and are infeasible for regression tasks like rPPG where data arrives continuously and in a sequential manner. Some instance-level TTA approaches may either rely on specialized network architectures [1, 22, 61, 74] or incorporate additional auxiliary tasks applied during both pre-training and adaptation [11, 57]. By imposing specific non-general requirement to the pre-training stage, these methods narrow application scope of TTA. Moreover, some works even introduce a memory mechanism that turns instances into batches [21], lacking an in-time adaptation to each inference with increased computational cost. In contrast with above mentioned TTA methods, our Bi-TTA adapts the model during inference under each encountered instance from target domain without specific pre-training", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bi-TTA", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "or network architecture requirements, offering a flexible, effective, and efficient solution for rPPG model adaptation.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 200, + 330, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 233, + 339, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, we propose a novel expert knowledge-based Bi-directional Test-Time Adapter (Bi-TTA) to address the performance degradation in unseen target domains. In the following sections, we will firstly formulate the problem in Sec. 3.2. Secondly, Sec. 3.3 introduces the proposed self-supervised priors based on expert knowledge in the realm of rPPG. The bidirectional (prospective and retrospective) adaptation strategy of our Bi-TTA will be specified in Sec. 3.4.", + "bbox": [ + 212, + 258, + 784, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 372, + 372, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The source dataset $D_{\\mathrm{S}} \\triangleq \\cup_{i=1}^{n_{\\mathrm{S}}} \\left\\{(\\pmb{x}_i, y_i)\\right\\}$ is assumed to be drawn independently and identically distributed (i.i.d.) from distribution $X_{\\mathrm{S}}$ . Similarly, the target dataset $D_{\\mathrm{T}}$ is assumed to be sampled from a different distribution $X_{\\mathrm{T}}$ , reflecting the domain shift. With a model parameterized by $\\pmb{w} \\in \\mathcal{W}$ and considering the supervised loss function $L: \\mathcal{W} \\times \\mathcal{X} \\times \\mathcal{Y} \\rightarrow \\mathbb{R}_{+}$ for evaluation, we define the observation loss over dataset $D$ as an approximation of the population loss over distribution $X$ :", + "bbox": [ + 212, + 397, + 782, + 502 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} (\\boldsymbol {w}) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} L \\left(\\boldsymbol {w}, \\boldsymbol {x} _ {i}, y _ {i}\\right) \\approx \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim X} [ L (\\boldsymbol {w}, \\boldsymbol {x}, y) ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 516, + 784, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The objective of TTA, including our Bi-TTA, is to minimize the target observation loss $\\mathcal{L}_{D_{\\mathrm{T}}}(\\boldsymbol{w}_t)$ after fine-tuning the initial model weights from $\\boldsymbol{w}_0$ to $\\boldsymbol{w}_t$ after $t$ observed samples.", + "bbox": [ + 214, + 564, + 782, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Expert Knowledge-based Priors", + "text_level": 1, + "bbox": [ + 215, + 632, + 524, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the context of Bi-TTA, the unavailability of target data labels compels the use of self-supervised loss functions for providing the tuning gradient. Inspired by [17, 31,58,62], our Bi-TTA capitalizes on the characteristics of rPPG signals to craft these self-supervised loss functions, concentrating on the principles of spatial and temporal consistency. The introduced loss functions are termed as $L_{\\mathrm{s}}(\\pmb{w}_{t-1}, \\pmb{x}_t)$ and $L_{\\mathrm{t}}(\\pmb{w}_{t-1}, \\pmb{x}_t)$ , focusing on spatial and temporal consistency respectively. These functions play a foundational role in the self-supervised tuning process, providing appropriate supervision based on the expert knowledge, despite the lack of direct label guidance.", + "bbox": [ + 212, + 657, + 784, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial-temporal feature map (STMap). In line with [16, 33, 43, 48, 51], 2D STMap is firstly constructed to encapsulate the face video, as specified in Fig. 3a. This STMap integrates both temporal and spatial dimensions: the temporal", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40947f9f907e1baf30091467befe69c31d0af0ef43d8582f3daecea6d250c58d.jpg", + "image_caption": [ + "(a) STMap Construction" + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 380, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12b439476b69a08ce4ba6ac3915b0e01f74268ad969fe04902220cb5073dd7ee.jpg", + "image_caption": [ + "(b) Temporal Consistency Loss" + ], + "image_footnote": [], + "bbox": [ + 395, + 145, + 552, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04e3232af5c821d9716d0b3cb0ac88bcd077e1b6aae56f3663d33372d53bad20.jpg", + "image_caption": [ + "(c) Spatial Consistency Loss", + "Fig. 3: Illustration of STMap construction and the implementation of our proposed expert knowledge-based priors. (a) The process of generating STMap, encompassing face alignment $^5$ and cropping, local signal extraction, and the subsequent integration. (b) The calculation process of TCL, aimed at minimizing significant prediction discrepancies between original and temporally shifted HR predictions. (c) The calculation process the SCL, focused on penalizing pronounced disparities across different facial regions. Note that the boxes colored in pink represent the loss outcomes." + ], + "image_footnote": [], + "bbox": [ + 566, + 143, + 785, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dimension represents the time sequence and the spatial dimension corresponds to the patches of the cropped face video frames. The current sample $\\pmb{x}_t\\in \\mathbb{R}^{W\\times H}$ is referred to as a sliding window on the STMap, where $W$ represents the sequential length of the sliding window, and $H$ indicates the size of spatial dimension in the STMap.", + "bbox": [ + 212, + 544, + 787, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal consistency loss (TCL). BVP signals exhibit smooth and gradual fluctuations over short time spans. This fundamental attribute, known as the temporal consistency, is essential for accurate physiological signal analysis and the key to guarantee the predictability of physiological signal changes. To preserving this critical characteristic and derive a tuning gradient accordingly, the TCL is meticulously designed to enforce the model's prediction deviations of BVP signals when subjected to minor time perturbations under an acceptable tolerance. To compute the TCL, a random minor temporal perturbation $\\delta_{\\mathrm{T}}$ is introduced by shifting the original sample $\\pmb{x}_t$ along the STMap's temporal dimension. The shifted sample $\\pmb{x}_{t - \\delta_{\\mathrm{T}}}$ , in parallel with the origin sample, are then fed into the rPPG network $\\pmb{w}_{t - 1}$ for predicting two sets of BVP signals. Then, for penalizing the model from large discrepancies in the HR estimation results between the original and shifted samples, $L_{1}$ regularization is applied on the original and shifted HR prediction results, incorporating with a tolerance", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bi-TTA", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "constant $\\xi_{\\mathrm{T}}$ . The specific calculation process of TCL is formulated in Eq. 2, as illustrated in Fig. 3b.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {t, t - \\delta_ {\\mathrm {T}}} ^ {\\mathrm {H R}} = \\left\\| \\boldsymbol {w} _ {t - 1} ^ {\\mathrm {H R}} (\\boldsymbol {x} _ {t}) - \\boldsymbol {w} _ {t - 1} ^ {\\mathrm {H R}} (\\boldsymbol {x} _ {t - \\delta_ {\\mathrm {T}}}) \\right\\| _ {1} \\in \\mathbb {R} ^ {W \\times 1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 188, + 691, + 207 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {t}} \\left(\\boldsymbol {w} _ {t - 1}, \\boldsymbol {x} _ {t}\\right) = \\sum_ {i} ^ {W} \\max \\left(0, \\Delta_ {t, t - \\delta_ {\\mathrm {T}}} ^ {\\mathrm {H R} (i)} - \\xi_ {\\mathrm {T}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 312, + 210, + 784, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial consistency loss (SCL). Heart rate variations induce chromatic changes across the face, which are periodic and consistent across different facial regions. This phenomenon is referred to as spatial consistency, alongside temporal consistency, are crucial attributes in the rPPG field. Inspired by this phenomenon, SCL is crafted to promote the spatial consistency, obtaining correct adaptation gradient by encouraging the BVP signals from different facial areas to exhibit similarities. In particular, the SCL is calculated covering a series of latent feature maps, corresponding to various depths, (i.e., the number of residual blocks) in the adopted ResNet-18 [18] architecture. Given the current sample $\\pmb{x}_t\\in \\mathbb{R}^{W\\times H}$ , four latent feature maps $\\left\\{F_{i}\\in \\mathbb{R}^{W_{i}\\times H_{i}}\\mid i\\in [1,4]\\right\\}$ could be extracted from the rPPG network $\\pmb{w}_{t - 1}$ . As illustrated in Fig. 3c, the calculation process of SCL is specify as:", + "bbox": [ + 212, + 263, + 792, + 444 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {s}} = \\sum_ {i} ^ {n} \\Delta F _ {i} = \\sum_ {i} ^ {n} \\sum_ {j} ^ {W _ {i} - \\delta_ {\\mathrm {S}}} | | F _ {i, j} - F _ {i, j + \\delta_ {\\mathrm {S}}} | | _ {1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 446, + 784, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $F_{i,j} = F_i[:,j] \\in \\mathbb{R}^{W_i}$ , $n = 4$ represents the number of residual blocks, and $\\delta_S$ is the incremental spatial shift. Leveraging the fine-grained multi-level strategy, SCL effectively nurtures a comprehensive regularization of spatial features, ensuring that appropriate supervision is available for the adaptation during inference without annotations.", + "bbox": [ + 212, + 494, + 782, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After combining $L_{\\mathrm{t}}$ and $L_{\\mathrm{s}}$ , without direct target label reliance, the self-supervised loss is formulated by weighting and summing the spatial and temporal consistency priors as follows:", + "bbox": [ + 212, + 571, + 784, + 617 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {p}} = \\lambda_ {\\mathrm {s}} L _ {\\mathrm {s}} + \\lambda_ {\\mathrm {t}} L _ {\\mathrm {t}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 630, + 784, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Bidirectional Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 214, + 667, + 555, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As introduced above, based on the proposed priors, the Bi-TTA framework aims to ensure effective and robust model adaptation under the variability and instability of single-instance learning conditions, minimizing the impact of domain-irrelevant noise. Simultaneously adapting the model with new target samples, our Bi-TTA also focuses on preserving the acquired adaptation ability dynamically, preventing the performance degradation caused by overfitting or catastrophic forgetting, achieving a stable maintenance of the adaptation performance in the target domain.", + "bbox": [ + 212, + 694, + 784, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3 https://github.com/adrianb/face-alignment.", + "bbox": [ + 217, + 823, + 532, + 840 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5094995fdae497c6b6c8b35239887c626464363180a8057bba77fdded0a44d85.jpg", + "image_caption": [ + "Fig. 4: Illustration of the proposed Bidirectional Test-Time Adapter (Bi-TTA). Black arrows $\\rightarrow$ indicate the adaptation process purely with the proposed two priors, i.e., TCL and SCL. Orange arrows $\\rightarrow$ denote that the PA module adjusts model parameter using the gradient of representative neighborhood with a radius $\\rho$ . The green ones $\\rightarrow$ show that the RS is activated when there is an oscillation, which is a sign of performance degradation, for maintaining the essential learned adaptation ability with former tuning gradients. The gradient and learning rate are formulated as $g$ and $\\eta$ respectively." + ], + "image_footnote": [], + "bbox": [ + 359, + 143, + 643, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Prospective adaptation (PA). During the process of model adaptation, especially under single-instance learning conditions, the noisy irrelevant information in biased target samples can easily interfere the model's tuning process, leading to a declined performance. Therefore, inspired by [13,38], rather than seeking out model parameter $\\pmb{w}_t$ that simply have low adaptation loss $L_{\\mathrm{p}}(\\pmb{w}_t)$ , we seek out model parameter whose entire neighborhoods to have uniformly low adaptation loss.", + "bbox": [ + 212, + 555, + 787, + 659 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {p}} ^ {\\prime} (\\boldsymbol {w}) = \\max _ {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\rho} L _ {\\mathrm {p}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 661, + 785, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the $\\rho$ is the disturbance range. This strategy, known as sharpness-aware minimization, has been proved to be effective in various areas [3, 19, 45, 67]. For computing efficiency and differentiability, we approximate the $\\epsilon^{*} = \\arg \\max_{\\| \\epsilon \\|_{2}\\leqslant \\rho}L_{\\mathrm{p}}(\\boldsymbol {w} +$", + "bbox": [ + 212, + 698, + 802, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$\\epsilon)$ . Following 1-order Taylor series, we get:", + "bbox": [ + 212, + 753, + 524, + 768 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {\\epsilon} ^ {*} \\approx \\underset {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\varrho} {\\arg \\max } L _ {\\mathrm {p}} (\\boldsymbol {w}) + \\boldsymbol {\\epsilon} ^ {\\mathrm {T}} \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w}) \\\\ = \\underset {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\rho} {\\arg \\max \\boldsymbol {\\epsilon}} ^ {\\mathrm {T}} \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w}) = \\underset {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\rho} {\\arg \\max \\boldsymbol {\\epsilon}} ^ {\\mathrm {T}} \\boldsymbol {g} = \\hat {\\boldsymbol {\\epsilon}}, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 782, + 784, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bi-TTA", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": ", w.r.t. $\\rho$ around 0. Then employing the Cauchy-Schwarz inequality from Dual Norm theory: $\\epsilon^{\\mathrm{T}}\\pmb {g}\\leqslant \\| \\pmb {\\epsilon}\\|_{2}\\| \\pmb {g}\\|_{2}$ . $\\pmb{\\epsilon}$ should be col-linear with $\\pmb{g}$ to maximize $\\pmb{\\epsilon}^{\\mathrm{T}}\\pmb{g}$ . Considering $\\| \\pmb {\\epsilon}\\| _2\\leqslant \\rho$ , we can get:", + "bbox": [ + 215, + 146, + 787, + 193 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} = \\rho \\frac {\\operatorname {s i g n} (\\boldsymbol {g}) \\boldsymbol {g}}{\\| \\boldsymbol {g} \\| _ {2}} = \\rho \\frac {\\operatorname {s i g n} \\left(\\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w})\\right) \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w})}{\\| \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w}) \\| _ {2}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 205, + 785, + 239 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Consequently, we can derive the gradient approximation under Eq. 5:", + "bbox": [ + 215, + 250, + 715, + 265 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}} = \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} ^ {\\prime} (\\boldsymbol {w}) \\approx \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\hat {\\boldsymbol {\\epsilon}}) \\\\ = \\left. \\frac {d (\\boldsymbol {w} + \\hat {\\epsilon})}{d \\boldsymbol {w}} \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\right| _ {\\boldsymbol {w} + \\hat {\\epsilon}} \\approx \\left. \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\right| _ {\\boldsymbol {w} + \\hat {\\epsilon}} \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 277, + 784, + 333 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4, the gradient $\\pmb{g}_t^{\\mathrm{PA}}$ is adopted to replace the original gradient merely derived from two self-supervised priors. This strategy boosts the model's robustness to the undesired variability from instance-level conditions, safeguarding the tuning process against the influence from target-domain irrelevant information. Consequently, it secures both the efficacy and stability of the adaptation process.", + "bbox": [ + 212, + 345, + 787, + 436 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Retrospective stabilization (RS). The PA module, by considering the model's entire neighborhood as the optimization objective, effectively enhances the model's robustness against irrelevant disturbances in target samples. However, after the model has processed a certain number of samples, it may still capture the detrimental noise feature instead of latent, universally domain-essential patterns, leading to performance degradation, as evidenced in Fig. 6. Therefore, to ensure reliable physiological signal predictions throughout the entire adaptation process, we propose the retrospective stabilization (RS) module. RS initially introduces the concept of trend gradient, termed as $\\pmb{g}^{*}$ . Starting with $\\pmb{g}_0^* = \\mathbf{0}$ , the trend gradient after processing $t$ samples is updated via: $\\pmb{g}_t^* = \\frac{\\pmb{g}_{t - 1}^* + (1 / ||L_{\\mathrm{p}}||_1)\\pmb{g}_t^{\\mathrm{PA}}}{1 + 1 / ||L_{\\mathrm{p}}||_1}$ . For each $\\pmb{g}_t^{\\mathrm{PA}}$ , its orthogonal projection on $\\pmb{g}_t^*$ is computed, formulated as $\\pmb{g}_t^\\perp = \\frac{\\pmb{g}_t^{\\mathrm{PA}}\\cdot\\pmb{g}_t^*}{\\|\\pmb{g}_t^*\\|^2}\\pmb{g}_t^*$ . As illustrated in Fig. 4, RS is activated if this projection aligns in the opposite of the trend gradient, which signifies an gradient oscillation, indicative of a potential fall of the generalization performance. Specifically, upon detection of oscillation, the model's optimization gradient will be the product of $\\pmb{g}_t^\\perp$ and the backtracking degree $k$ . The orthogonal part of $\\pmb{g}_t$ on $\\pmb{g}_t^*$ is ignored.", + "bbox": [ + 212, + 438, + 789, + 696 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {t} ^ {\\mathrm {R S}} = \\left\\{ \\begin{array}{l} k \\cdot \\boldsymbol {g} _ {t} ^ {\\perp}, \\cos \\left(\\boldsymbol {g} _ {t} ^ {*}, \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}}\\right) < 0 \\\\ \\left(1 - \\lambda_ {t} ^ {\\mathrm {R S}}\\right) \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}} + \\lambda_ {t} ^ {\\mathrm {R S}} \\boldsymbol {g} _ {t} ^ {*}, \\cos \\left(\\boldsymbol {g} _ {t} ^ {*}, \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}}\\right) \\geqslant 0 \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 708, + 784, + 750 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda_t^{\\mathrm{RS}} = -1 + 2 \\cdot \\operatorname{Sigmoid}(-\\frac{4t}{\\Omega}) \\in \\mathbb{R}_+^{(0,1)}$ , is the anneal scaling factor, $\\Omega$ denotes the amount of target samples needed to ensure the fidelity of trend gradient, i.e., $\\lambda_t^{\\mathrm{RS}}$ can be approximated as 1.0 when $t \\geqslant \\Omega$ . The obtained $\\pmb{g}_t^{\\mathrm{RS}}$ represents the gradient that ultimately contributes to fine-tune the model for target domain adaptation.", + "bbox": [ + 212, + 762, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/339466737b800041674716e618c0e79fdd4605fc70b0cd9f638331671f04282e.jpg", + "table_caption": [ + "Table 1: HR estimation results. Typically, the model selected for adaptation on each dataset is initially pre-trained on the remaining four datasets utilizing the NEST [34] framework. Ours w/o P.R. denotes that only the expert knowledge-based priors is adopted, without the bidirectional adaptation strategy. The best results are highlighted in bold, and the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodVIPL [42]V4V [49]PURE [53]UBFC [2]BUAA [68]
M↓R↓r↑M↓R↓r↑M↓R↓r↑M↓R↓r↑M↓R↓r↑
Methods based on DG protocol
Coral [54]8.6811.910.5310.3214.420.327.5910.870.725.898.040.763.645.740.8
VReX [23]8.3711.620.549.8214.160.377.2410.140.785.597.680.813.275.010.86
NCDG [59]8.4711.810.5210.1414.460.347.3210.350.775.317.560.823.125.160.85
NEST [34]7.8611.150.589.2713.790.416.719.590.814.676.790.862.884.690.89
Methods based on TTA protocol
Tent [61]8.0911.320.559.9811.730.476.869.680.844.579.740.872.373.590.92
SAR [40]8.1311.450.539.6811.460.476.339.30.864.569.640.872.072.930.95
TTT++ [32]7.7311.080.569.5111.310.546.239.140.884.366.790.872.052.820.92
EATA [39]7.6911.030.579.3611.120.516.139.230.864.256.860.881.892.640.94
SHOT [29]7.7511.070.569.5511.320.515.818.860.874.056.450.871.872.470.96
AdaODM [72]7.8111.210.559.110.930.555.518.290.884.016.590.871.912.60.95
ConPhys4 [58]7.4310.70.619.310.960.526.098.830.883.926.610.8861.752.190.97
Ours w/o P.R.7.3110.640.628.9710.820.555.568.670.883.646.630.8851.682.040.99
Ours w/o RS7.1510.830.638.9310.790.555.248.320.8893.596.610.8851.551.940.99
Ours w/o PA7.2410.780.629.0410.970.545.398.450.883.546.550.8881.511.980.99
Ours7.0910.720.639.111.020.545.028.110.8903.536.240.8911.491.970.99
", + "bbox": [ + 215, + 224, + 805, + 473 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 501, + 375, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 534, + 334, + 547 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We elaborately select five rPPG datasets, encompassing diverse races, levels of head motion, camera conditions, and ambient lighting, to establish a comprehensive large-scale benchmark of rPPG task under TTA protocol.", + "bbox": [ + 212, + 559, + 785, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "VIPL-HR [42] have nine scenarios, three RGB cameras, different illumination conditions, and different levels of movement. It is worth mentioning that the BVP signal and video of the dataset do not match in the time dimension. We used the output of HR head as the evaluation result. Besides, we normalize STMap to 30 fps by cubic spline interpolation to solve the problem of the unstable frame rate of video [33,51,70].", + "bbox": [ + 212, + 604, + 785, + 696 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "V4V [49] is designed to collect data with the drastic changes of physiological indicators by simulating ten tasks such as a funny joke, 911 emergency call, and odor experience. It is worth mentioning that There are only heart rate labels in the dataset and no BVP signal labels. We used the output of HR head as the evaluation result.", + "bbox": [ + 212, + 696, + 785, + 770 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "BUAA [68] is proposed to evaluate the performance of the algorithm against various illumination. We only use data with illumination greater than or equal", + "bbox": [ + 215, + 771, + 785, + 801 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bi-TTA", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "4 The proposed self-supervised loss in ConPhys [58] is applied for a comparative analysis of our proposed priors.", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f73fb16e85592b530001867f3bc1db5fdf1c5346ea3582618f5438b7346c4740.jpg", + "image_caption": [ + "(a) $k$" + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 485, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ca0988a0006d3a1c41b677483a66cd3bc5bbd943feb8350b2bbdf1fd44fd65ce.jpg", + "image_caption": [ + "(d) $\\xi_{\\mathrm{T}}$" + ], + "image_footnote": [], + "bbox": [ + 232, + 220, + 486, + 286 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/09e95419af086bb65a14654621222dccf7bd02728e2028509c8760437b0761dd.jpg", + "image_caption": [ + "(b) $\\lambda_{t}$" + ], + "image_footnote": [], + "bbox": [ + 491, + 143, + 625, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/049ce48e4b3afd1cfaef5076b0f953022bbf08345fd39b0e6c2fe82cdd63ee80.jpg", + "image_caption": [ + "(e) $\\rho$", + "Fig. 5: Ablation experiments of hyper-parameters on down-sampled VIPL dataset5. Adheres to the default parameter configuration detailed in Sec.4.2, only one hyper-parameter is varied during each set of experiments. The pink arrows point to the lowest MAE result and the blue arrows mark the second if indiscernible." + ], + "image_footnote": [], + "bbox": [ + 491, + 222, + 625, + 287 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ded933a3ada4d1dc625115cf256cc2ce9dd46d2259274fc81fc9e9edeb1f5e51.jpg", + "image_caption": [ + "(c) $\\lambda_{s}$" + ], + "image_footnote": [], + "bbox": [ + 635, + 143, + 771, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/226cdb6d02d5db366b7fd637888e0dc80225c8aab3c46933125e5310a56c0fe3.jpg", + "image_caption": [ + "(f) $\\Omega$" + ], + "image_footnote": [], + "bbox": [ + 635, + 222, + 771, + 287 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to 10 lux because underexposed images require special algorithms that are not considered in this article. We used the output of BVP head as the evaluation for BUAA, PURE, and UBFC-rPPG datasets.", + "bbox": [ + 212, + 398, + 784, + 443 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "PURE [53] contains 60 RGB videos from 10 subjects with six different activities, specifically, sitting still, talking, and four rotating and moving head variations. The BVP signals are down-sampled from 60 to 30 fps with cubic spline interpolation to align with the videos.", + "bbox": [ + 212, + 445, + 782, + 505 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "UBFC-rPPG [2] containing 42 face videos with sunlight and indoor illumination, the ground-truth BVP signals, and HR values were collected by CMS50E.", + "bbox": [ + 214, + 507, + 782, + 537 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 568, + 455, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We implement our proposed priors and Bi-TTA using the PyTorch framework. The STMap sample is denoted as $\\pmb{x} \\in \\mathbb{R}^{256 \\times 25 \\times 3}$ , where 256 is the temporal length, i.e., the number of video frames, of the sliding window, and 25 is the spatial dimension of the STMap. It is firstly resized to $\\pmb{x}' \\in \\mathbb{R}^{256 \\times 64 \\times 3}$ for network input. The network utilizes ResNet-18 [18] as the base architecture, followed by an additional fully connected layer for HR estimation. During the TCL calculation, the temporal perturbation $\\delta_{\\mathrm{T}}$ is randomly sampled from a uniform distribution $\\mathbb{U}_{(0,59]}$ . Furthermore, we use stochastic gradient descent (SGD) with a momentum of 0.9 as the base optimizer. The bidirectional adaptation mechanism is implemented on top of the SGD. During the adaptation process, we set the learning rate to 0.0001 and the batch size is 1. The hyper-parameters are configured as follows: $\\lambda_{\\mathrm{s}} = 0.001$ , $\\xi_{\\mathrm{T}} = 8.0$ , $\\lambda_{\\mathrm{t}} = 0.01$ , $\\rho = 0.005$ , $k = -9.0$ , and $\\Omega = 4000$ . Please refer to Fig. 5 for detailed ablations.", + "bbox": [ + 212, + 599, + 787, + 796 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "5 Note that the performance of down-sampled VIPL dataset might be inferior than the full VIPL dataset, due to the larger interval between neighboring samples.", + "bbox": [ + 217, + 810, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Results", + "text_level": 1, + "bbox": [ + 215, + 146, + 323, + 159 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Following existing rPPG methods [4,33,34,44,51,70], mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (r) are used to evaluate the HR estimation, which inherently reflects the ability to extract physiological periodic signals. The results are organized and recorded in Tab. 1. It is evident that TTA methods generally perform better than DG approaches. Among these TTA methods, our Bi-TTA demonstrates significantly superior performance, showcasing our effectiveness especially when both prospective and retrospective adaptations are synergistically employed. This highlights Bi-TTA's robustness in adapting to various unseen domains, underscoring its potential for real-world applications.", + "bbox": [ + 212, + 169, + 787, + 321 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 342, + 388, + 357 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this sub-section, we will mainly discuss about the effectiveness of the proposed priors and the bidirectional adaptation strategy.", + "bbox": [ + 212, + 366, + 785, + 397 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3c4faa4af001fabbbdba4f202a0f41e9d91c3c5883dbc485122816b11610854e.jpg", + "image_caption": [ + "(a) Comparison between priors and GT.", + "Fig. 6: (a) Comparison between task-specific priors-based TTA and fully supervised adaptation. The evaluation metrics are MSE, RMSE, and r. Alongside the visualization of the loss distribution for both priors-based and supervised adaptation method. Loss values smaller than $10^{-6}$ are ignored. Note that for better comparison, the learning rate of supervised adaptation is set to 0.0005, while for the priors-based TTA it is 0.0001. (b) Ablation study of the adaptation performance among the proposed PA and RS modules in our Bi-TTA framework. This Ablation study demonstrates the effectiveness of the PA and RS, in achieving a more effective and stable adaptation process." + ], + "image_footnote": [], + "bbox": [ + 222, + 417, + 460, + 686 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eebbcab48fbd0376c3d4c5df22baafedeb0d42dadb2203a3b6bf8129ad13bac2.jpg", + "image_caption": [ + "(b) Ablation study of PA, RS, and, priors." + ], + "image_footnote": [], + "bbox": [ + 477, + 420, + 785, + 685 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bi-TTA", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Task-specific priors. As depicted in Fig. 6a, compared to ground-truth supervision, our task-specific priors provide effective adaptation supervision only on a significantly smaller number of samples, but induce considerably larger gradients. Thus, we can infer that the convergence rate under our self-supervised priors is markedly faster than ground-truth, despite a substantially lower learning rate. This phenomenon underscores the effectiveness of our priors.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As discussed in [71], modern deep learning models are prone to memorize training data due to their over-parameterized design, leading to overfitting risks. Merely minimizing standard loss functions on training data often falls short in ensuring satisfactory generalization ability [13]. Given that the rPPG model has undergone extensive pre-training with supervised learning, its parameters have likely been excessively optimized. In contrast, task knowledge-based priors are expected to offer fresh and correct guidance, steering the model towards improved generalization capability.", + "bbox": [ + 212, + 241, + 787, + 362 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Prospective and retrospective adaptation. As depicted in Fig. 6b, the initial convergence rate and adaptation performance of the four comparisons are basically consistent across the first 4000 samples. However, in the latter half, the priors-based method demonstrates noticeable performance degradation. In contrast, other three methods, through utilizing prospective or retrospective adaptations, substantially mitigate this problem, thereby enhancing the adaptation ability.", + "bbox": [ + 212, + 364, + 787, + 470 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Notably, between the $5000^{\\mathrm{th}}$ and $6000^{\\mathrm{th}}$ samples, while methods employing prospective adaptation exhibit rapid convergence and superior performance, they also display a significant degree of performance decay due to overfitting or catastrophic forgetting, which even intensifies as more samples are processed. On the other hand, the retrospective adaptation strategy ensures a more stable post-convergence maintenance in the target domain. The Bi-TTA, integrating both prospective and retrospective strategies, adeptly balances convergence speed, effectiveness, and stability.", + "bbox": [ + 212, + 474, + 787, + 597 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 637, + 359, + 655 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we address the challenge of significant performance degradation when adapting rPPG models to unfamiliar target domains. Considering privacy, we implement TTA for rPPG tasks, enabling the adaptation of pre-trained models to the target domain using only unannotated, instance-level target data. The models are fine-tuned during the inference of target samples. In particular, we introduce Bi-TTA, a method that leverages spatial and temporal consistency for effective self-supervision, coupled with pioneering prospective and retrospective adaptation strategies. This methodology streamlines the adaptation process and outperforms existing techniques, demonstrating substantial promise for real-world applications in non-invasive physiological monitoring.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bahmani, S., Hahn, O., Zamfir, E., Araslanov, N., Cremers, D., Roth, S.: Semantic self-adaptation: Enhancing generalization with a single sample. arXiv preprint arXiv:2208.05788 (2022)", + "2. Bobbia, S., Macwan, R., Benezeth, Y., Mansouri, A., Dubois, J.: Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters 124, 82-90 (2019)", + "3. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: International Conference on Machine Learning. pp. 1059-1071. PMLR (2021)", + "4. Chen, W., McDuff, D.: Deepphys: Video-based physiological measurement using convolutional attention networks. In: Proc. ECCV. pp. 349-365 (2018)", + "5. Das, A., Lu, H., Han, H., Dantcheva, A., Shan, S., Chen, X.: Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In: FG. pp. 01-08. IEEE (2021)", + "6. De Haan, G., Jeanne, V.: Robust pulse rate from chrominance-based rppg. IEEE Trans. Biomed. Eng. 60(10), 2878-2886 (2013)", + "7. De Haan, G., Van Leest, A.: Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiol. Meas. 35(9), 1913 (2014)", + "8. D'Innocente, A., Bucci, S., Caputo, B., Tommasi, T.: Learning to generalize one sample at a time with self-supervision. arXiv preprint arXiv:1910.03915 (2019)", + "9. Du, J., Liu, S.Q., Zhang, B., Yuen, P.C.: Dual-bridging with adversarial noise generation for domain adaptive rppg estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10355–10364 (June 2023)", + "0. Du, J., Liu, S.Q., Zhang, B., Yuen, P.C.: Dual-bridging with adversarial noise generation for domain adaptive rppg estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10355–10364 (2023)", + "1. D'Innocente, A., Borlino, F.C., Bucci, S., Caputo, B., Tommasi, T.: One-shot unsupervised cross-domain detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 732-748. Springer (2020)", + "12. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1126-1135. PMLR (06-11 Aug 2017), https://proceedings.mlr.press/v70/finn17a.html", + "13. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=6Tm1mposlrM", + "14. Gandelsman, Y., Sun, Y., Chen, X., Efros, A.A.: Test-time training with masked autoencoders. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=SHMi1b7sjXk", + "15. Gao, J., Zhang, J., Liu, X., Darrell, T., Shelhamer, E., Wang, D.: Back to the source: Diffusion-driven adaptation to test-time corruption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11786-11796 (2023)", + "16. Gee-Sern Hsu, ArulMurugan Ambikapathi, M.S.C.: Deep learning with time-frequency representation for pulse estimation. In: Proc. IJCB. pp. 642-650 (2017)" + ], + "bbox": [ + 218, + 175, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bi-TTA", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Gideon, J., Stent, S.: The way to my heart is through contrastive learning: Remote photoplethysmography from unlabeled video. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3995-4004 (2021)", + "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. IEEE CVPR. pp. 770-778 (2016)", + "19. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)", + "20. Kessler, V., Thiam, P., Amirian, M., Schwenker, F.: Pain recognition with camera photoplethysmography. In: IEEE IPTA. pp. 1-5 (2017)", + "21. Khurana, A., Paul, S., Rai, P., Biswas, S., Aggarwal, G.: Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355 (2021)", + "22. Klingner, M., Ayache, M., Fingscheidt, T.: Continual batchnorm adaptation (cbna) for semantic segmentation. IEEE Transactions on Intelligent Transportation Systems 23(11), 20899-20911 (2022)", + "23. Krueger, D., Caballero, E., Jacobsen, J.H., Zhang, A., Binas, J., Zhang, D., Le Priol, R., Courville, A.: Out-of-distribution generalization via risk extrapolation (rex). In: ICML. pp. 5815-5826. PMLR (2021)", + "24. Lee, E., Chen, E., Lee, C.Y.: Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16. pp. 392–409. Springer (2020)", + "25. Lewandowska, M., Ruminski, J., Kocejko, T., Nowak, J.: Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In: FedCSIS. pp. 405-410. IEEE (2011)", + "26. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)", + "27. Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361 (2023)", + "28. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International Conference on Machine Learning (ICML). pp. 6028-6039 (2020)", + "29. Liang, J., Hu, D., Wang, Y., He, R., Feng, J.: Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021), in Press", + "30. Liu, X., Jiang, Z., Fromm, J., Xu, X., Patel, S., McDuff, D.: Metaphys: few-shot adaptation for non-contact physiological measurement. In: Proceedings of the conference on health, inference, and learning. pp. 154-163 (2021)", + "31. Liu, X., Zhang, Y., Yu, Z., Lu, H., Yue, H., Yang, J.: rppg-mae: Self-supervised pretraining with masked autoencoders for remote physiological measurement. arXiv preprint arXiv:2306.02301 (2023)", + "32. Liu, Y., Kothari, P., van Delft, B.G., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)", + "33. Lu, H., Han, H., Zhou, S.K.: Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In: Proc. IEEE CVPR. pp. 12404-12413 (2021)", + "34. Lu, H., Yu, Z., Niu, X., Chen, Y.C.: Neuron structure modeling for generalizable remote physiological measurement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18589-18599 (June 2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. McDuff, D.: Camera measurement of physiological vital signs. CSUR (2021)", + "36. McDuff, D.: Applications of camera-based physiological measurement beyond healthcare. In: CVSM, pp. 165-177. Elsevier (2022)", + "37. McDuff, D., Gontarek, S., Picard, R.W.: Improvements in remote cardiopulmonary measurement using a five band digital camera. IEEE Trans. Biomed. Eng. 61(10), 2593-2601 (2014)", + "38. Mi, P., Shen, L., Ren, T., Zhou, Y., Sun, X., Ji, R., Tao, D.: Make sharpness-aware minimization stronger: A sparsified perturbation approach (2022)", + "39. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022)", + "40. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023)", + "41. Niu, X., Han, H., Shan, S., Chen, X.: Synrhythm: Learning a deep heart rate estimator from general to specific. In: Proc. IEEE ICPR. pp. 3580-3585 (2018)", + "42. Niu, X., Han, H., Shan, S., Chen, X.: VIPL-HR: A multi-modal database for pulse estimation from less-constrained face video. In: Proc. ACCV. pp. 562-576 (2018)", + "43. Niu, X., Shan, S., Han, H., Chen, X.: Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Trans. on Image Process. 29, 2409-2423 (2020)", + "44. Niu, X., Yu, Z., Han, H., Li, X., Shan, S., Zhao, G.: Video-based remote physiological measurement via cross-verified feature disentangling. In: Proc. ECCV (2020)", + "45. Pham, H., Dai, Z., Xie, Q., Le, Q.V.: Meta pseudo labels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11557-11568 (2021)", + "46. Poh, M.Z., McDuff, D.J., Picard, R.W.: Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express 18(10), 10762-10774 (2010)", + "47. Qiu, Y., Liu, Y., Arteaga-Falconi, J., Dong, H., El Saddik, A.: Evm-cnn: Real-time contactless heart rate estimation from facial video. IEEE Trans. Multimedia (2019)", + "48. Reiss, A., Indlekofer, I., Schmidt, P., Van Laerhoven, K.: Deep ppg: large-scale heart rate estimation with convolutional neural networks. Sensors 19(14), 3079 (2019)", + "49. Revanur, A., Li, Z., Ciftci, U.A., Yin, L., Jeni, L.A.: The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In: Proc. CVPR workshop. pp. 2760-2767 (2021)", + "50. Song, R., Chen, H., Cheng, J., Li, C., Liu, Y., Chen, X.: Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE J-BHI pp. 1-1 (2021)", + "51. Song, R., Zhang, S., Li, C., Zhang, Y., Cheng, J., Chen, X.: Heart rate estimation from facial videos using a spatiotemporal representation with convolutional neural networks. IEEE Trans. Instrum. Meas. (2020)", + "52. Spetlik, R., Franc, V., Matas, J.: Visual heart rate estimation with convolutional neural network. In: Proc. BMVC. pp. 3-6 (2018)", + "53. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: Proc. IEEE ISRHIC. pp. 1056-1062 (2014)", + "54. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: ECCV. pp. 443-450. Springer (2016)", + "55. Sun, B., Wei, Q., Li, L., Xu, Q., He, J., Yu, L.: LSTM for dynamic emotion and group emotion recognition in the wild. In: Proc. ACM ICMI. pp. 451-457 (2016)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bi-TTA", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Sun, W., Zhang, X., Lu, H., Chen, Y., Ge, Y., Huang, X., Yuan, J., Chen, Y.: Resolve domain conflicts for generalizable remote physiological measurement. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8214-8224 (2023)", + "57. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)", + "58. Sun, Z., Li, X.: Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In: European Conference on Computer Vision. pp. 492-510. Springer (2022)", + "59. Tian, C.X., Li, H., Xie, X., Liu, Y., Wang, S.: Neuron coverage-guided domain generalization. IEEE Trans. on Pattern Anal. Mach. Intell. (2022)", + "60. Verkruysse, W., Svaasand, L.O., Nelson, J.S.: Remote plethysmographic imaging using ambient light. Opt. Express 16(26), 21434-21445 (2008)", + "61. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c", + "62. Wang, H., Ahn, E., Kim, J.: Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2431-2439 (2022)", + "63. Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Yu, P.: Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering (2022)", + "64. Wang, W., den Brinker, A.C., Stuijk, S., de Haan, G.: Algorithmic principles of remote ppg. IEEE Trans. Biomed. Eng. 64(7), 1479-1491 (2017)", + "65. Wang, W., Stuijk, S., De Haan, G.: Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE Trans. Biomed. Eng. 62(2), 415-425 (2015)", + "66. Wang, Z.K., Kao, Y., Hsu, C.T.: Vision-based heart rate estimation via a two-stream cnn. In: Proc. IEEE ICIP. pp. 3327-3331 (2019)", + "67. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems 33, 2958-2969 (2020)", + "68. Xi, L., Chen, W., Zhao, C., Wu, X., Wang, J.: Image enhancement for remote photoplethysmography in a low-light environment. In: FG. pp. 1-7. IEEE (2020)", + "69. Yu, Z., Peng, W., Li, X., Hong, X., Zhao, G.: Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In: Proc. IEEE ICCV. pp. 151-160 (2019)", + "70. Yu, Z., Shen, Y., Shi, J., Zhao, H., Torr, P., Zhao, G.: Physformer: Facial videobased physiological measurement with temporal difference transformer. IEEE CVPR (2022)", + "71. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021)", + "72. Zhang, X., Chen, Y.C.: Adaptive domain generalization via online disagreement minimization. IEEE Transactions on Image Processing (2023)", + "73. Zhao, X., Liu, C., Sicilia, A., Hwang, S.J., Fu, Y.: Test-time fourier style calibration for domain generalization. arXiv preprint arXiv:2205.06427 (2022)", + "74. Zou, Y., Zhang, Z., Li, C.L., Zhang, H., Pfister, T., Huang, J.B.: Learning instance-specific adaptation for cross-domain segmentation. In: European Conference on Computer Vision. pp. 459-476. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 833 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_model.json b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3e40b63dcd2f9ac88c17fa82b1b28d71a10dddaa --- /dev/null +++ b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_model.json @@ -0,0 +1,2722 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.243, + 0.142, + 0.761, + 0.187 + ], + "angle": 0, + "content": "Bi-TTA: Bidirectional Test-Time Adapter for Remote Physiological Measurement" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.212, + 0.674, + 0.229 + ], + "angle": 0, + "content": "Haodong Li\\(^{1}\\), Hao Lu\\(^{1}\\), and Ying-Cong Chen\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.239, + 0.717, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\) Hong Kong University of Science and Technology (Guangzhou)" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.254, + 0.73, + 0.282 + ], + "angle": 0, + "content": "\\(^{2}\\) Hong Kong University of Science and Technology {hli736, hlu585}@connect.hkust-gz.edu.cn yingcongchen@ust.hk" + }, + { + "type": "list", + "bbox": [ + 0.274, + 0.239, + 0.73, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.334, + 0.397, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.302, + 0.583, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.303, + 0.761, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.43, + 0.408, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.536, + 0.366, + 0.55 + ], + "angle": 0, + "content": "(a) Pre-trained" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.431, + 0.589, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.424, + 0.536, + 0.569, + 0.55 + ], + "angle": 0, + "content": "(b) Priors-based TTA" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.431, + 0.769, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.536, + 0.759, + 0.55 + ], + "angle": 0, + "content": "(c) Priors-based Bi-TTA" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Fig.1: Visualization of the mean absolute error (MAE) result on down-sampled VIPL dataset using (a) the pre-trained rPPG model; (b) priors-based test-time adapted model; and (c) priors-based bidirectionally test-time adapted model. These visualizations demonstrate that our proposed priors and the Bi-TTA framework significantly enhance the generalization performance (reflected by the flatness of MAE field [26]) beyond the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.688, + 0.747, + 0.84 + ], + "angle": 0, + "content": "Abstract. Remote photoplethysmography (rPPG) is gaining prominence for its non-invasive approach to monitoring physiological signals using only cameras. Despite its promise, the adaptability of rPPG models to new, unseen domains is hindered due to the environmental sensitivity of physiological signals. To address this issue, we pioneer the Test-Time Adaptation (TTA) in rPPG, enabling the adaptation of pre-trained models to the target domain during inference, sidestepping the need for annotations or source data due to privacy considerations. Particularly, utilizing only the user's face video stream as the accessible target domain data, the rPPG model is adjusted by tuning on each single instance it encounters. However, 1) TTA algorithms are designed predominantly for classifi-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "H. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.148, + 0.744, + 0.398 + ], + "angle": 0, + "content": "cation tasks, ill-suited in regression tasks such as rPPG due to inadequate supervision. 2) Tuning pre-trained models in a single-instance manner introduces variability and instability, posing challenges to effectively filtering domain-relevant from domain-irrelevant features while simultaneously preserving the learned information. To overcome these challenges, we present Bi-TTA, a novel expert knowledge-based Bidirectional Test-Time Adapter framework. Specifically, leveraging two expert-knowledge priors for providing self-supervision, our Bi-TTA primarily comprises two modules: a prospective adaptation (PA) module using sharpness-aware minimization to eliminate domain-irrelevant noise, enhancing the stability and efficacy during the adaptation process, and a retrospective stabilization (RS) module to dynamically reinforce crucial learned model parameters, averting performance degradation caused by overfitting or catastrophic forgetting. To this end, we established a large-scale benchmark for rPPG tasks under TTA protocol, promoting advancements in both the rPPG and TTA fields. The experimental results demonstrate the significant superiority of our approach over the state-of-the-art (SoTA)." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.411, + 0.741, + 0.44 + ], + "angle": 0, + "content": "Keywords: Bidirectional Test-Time Adaptation \\(\\cdot\\) Expert Knowledge-based Priors \\(\\cdot\\) Remote Photoplethysmography" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.487, + 0.377, + 0.503 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.608 + ], + "angle": 0, + "content": "The cardiac cycle refers to the process experienced by the cardiovascular system from the start of one heartbeat to the beginning of the next. Through the analysis of the cardiac cycle, important physiological indicators such as heart rate (HR), respiration rate (RF), and heart rate variability (HRV) can be derived, playing crucial roles in medical diagnostics, health surveillance, and forensic analysis [20, 36, 55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.789, + 0.79 + ], + "angle": 0, + "content": "Traditional physiological measurement devices like electrocardiograms (ECG), heart rate bands, and finger-clips, despite their accuracy, are often expensive and uncomfortable to wear. In contrast, rPPG offers a non-invasive and more convenient alternative by extracting blood volume pulses (BVP) from facial videos, analyzing the skin's light absorption variations to measure HR, HRV, and RF [44,69,70]. These monitors are especially important for tracking health status and sympathetic activity levels [20,36,55]. Nowadays, physiological measurements based on rPPG, using only ordinary cameras, are increasingly becoming the focus of research in computer vision community [5,9,33,35,43,47,50]. However, the periodic color changes in the skin captured by remote cameras are quite subtle, and variables like natural scene illumination and head movements can significantly impair the reliability of physiological indicator measurements." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.787, + 0.826 + ], + "angle": 0, + "content": "1 This dataset is 1/20 equally spaced down-sampled from the full VIPL dataset [42], and contains 8349 samples in total." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.534, + 0.84 + ], + "angle": 0, + "content": "2 This visualization is generated following [26]." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Bi-TTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.145, + 0.713, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.309, + 0.789, + 0.394 + ], + "angle": 0, + "content": "Fig. 2: Visualization of Domain Adaptation (DA) and Test-Time Adaptation (TTA) methodologies. DA utilizes batch learning with labeled target data, while TTA dynamically refines the model during inference without relying on target data labels or distribution. Both DA and TTA obviate the requirement for source data. Note that domain generalization (DG) is excluded as it does not focus on specific target domain adaptation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.789, + 0.519 + ], + "angle": 0, + "content": "Conventional rPPG methods [6,25,37,46,60,64] perform well only in constrained environments. Recent deep learning (DL) methods [2,43,49,53,68] have demonstrate promising results, however, the most of training data are collected in controlled laboratory environments with limited variations in illumination, camera parameters, etc. Therefore, these methods still struggle to generalize to unseen domains, necessitating model adaptation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.789, + 0.672 + ], + "angle": 0, + "content": "For privacy considerations, the adaptation of rPPG models in unseen scenarios is required to be performed without accessing the source data, and target data annotations or distributions (i.e., large batches of target data samples). Addressing this, we firstly implement Test-Time Adaptation (TTA) in the field of rPPG. TTA aims to fine-tune a pre-trained source model during inference time, as illustrated in Fig. 2, without accessing the distribution and labeling of both source data and target data and naturally eliminate the need for intensive re-training [61]. This innovation allows for the automatic adaptation of pretrained rPPG models to unseen domains during inference, without the necessity for labelling or batches of target data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "However, applying TTA directly into rPPG faces two key limitations. 1) Predominantly designed for classification tasks, most TTA algorithms leverage entropy in normalization layers or pseudo labels. These strategies are not inherently suited for regression tasks like rPPG, leading to a lack of appropriate supervision. 2) In real-world scenarios, the pre-trained rPPG model is tuned during inference the user's face video stream, on an instance-by-instance context. This variability and instability introduced by biased instances pose challenges to effectively filtering out domain-irrelevant noise and simultaneously preserving crucial learned information as the model acquiring new knowledge. This nuanced balance is critical in preventing slow or ineffective adaptation under single-instance learning condition, ensuring an efficient and consistent adaptation process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "H. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.404 + ], + "angle": 0, + "content": "To navigate these challenges, we proposed a novel Bidirectional Test-Time Adapter (Bi-TTA) framework, grounded on two expert knowledge-based self-supervised priors. Firstly, two regularizations are proposed for offering effective supervision especially in rPPG tasks, addressing the aforementioned ill-suited problem of applying existing TTA algorithms into regression tasks. These two priors focus on temporal and spatial consistency respectively, which is derived from the characteristics of physiological signals. Then, two strategies are introduced for stable and effective fine-tuning process. The prospective adaptation (PA) strategy is designed for filtering out the irrelevant information of each encountered instance, aiming that only target-domain-essential feature is truly contributing to model adaptation. The retrospective stabilization (RS) strategy backtracks and consolidates essential model parameters dynamically, leveraging previous update gradients during the self-supervised tuning process, ensuring that the learned adaptation ability is preserved intact. We evaluate our method across various datasets, benchmarking against existing prevailing TTA algorithms, demonstrating the superiority of our Bi-TTA. Overall, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.414, + 0.785, + 0.442 + ], + "angle": 0, + "content": "- We introduce the TTA protocol to rPPG tasks, paving the way for a new paradigm in self-supervised adapting of rPPG models during inference." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.444, + 0.785, + 0.472 + ], + "angle": 0, + "content": "- We introduce two novel self-supervised priors to regularize temporal and spatial consistency, providing appropriate supervision." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.474, + 0.785, + 0.517 + ], + "angle": 0, + "content": "- We propose a novel Bidirectional Test-Time Adapter (Bi-TTA) framework grounded on the introduced priors, which promotes effective and efficient adaptation, with enhanced stability." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.519, + 0.785, + 0.563 + ], + "angle": 0, + "content": "- We established a large-scale benchmark for rPPG tasks under TTA protocol. Extensive experiments demonstrate significantly improved adaptation capacity of our approach." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.414, + 0.785, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.398, + 0.604 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.557, + 0.635 + ], + "angle": 0, + "content": "2.1 Remote Physiological Measurement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Remote physiological measurement aims to quantify HR and HRV from BVP signals by interpreting skin chrominance fluctuations. Traditional techniques typically involve color space transformations or signal decomposition to obtain the BVP signal with high signal-to-noise ratio (SNR) [6, 7, 46, 65]. With powerful modeling capacity of deep learning (DL), there's been a significant shift towards employing DL models for remote physiological estimation, yielding impressive results. [4, 16, 41, 43, 52, 66, 69]. Many of these methods extract BVP signals from face videos typically by aggregating spatial and temporal information [44, 52, 69, 70]. Besides, many works utilize hand-designed 2D feature maps extracted from face videos to alleviate computational demands [16, 33, 43, 48, 51]. Recently, researchers have turned to exploring how to leverage task-specific knowledge and how to improve the generalization performance of the model into unseen scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Bi-TTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Task-specific priors. rPPG is distinguished by its unique biological and physical mechanisms, which can be exploited to enhance the physiological signal extraction ability of the model [17,31,58,62]. These task-specific priors can be summed up as 1) spatial consistency: BVP signals extracted from different face areas should exhibit similar; 2) temporal consistency: BVP signals typically exhibit smooth and gradual changes over short time period; 3) heart rate range: The heart rate typically ranges between 40 and 250 beats per minute (bpm). Different from existing works, we provide a more granular solution leveraging the multi-scale latent features, rather than purely focusing on the prediction output of BVP signals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.79, + 0.481 + ], + "angle": 0, + "content": "The generalization of the model. With the domain shifts (e.g., lighting, head motion, environment, etc.), most methods tend to have significant performance degradation. To improve the generalization performance on unseen domains, NEST-rPPG established a large-scale DG evaluation protocol [34]. These DG methods, requiring the access of all source data, do not focus on the adaptation of a specific target domain, leading to sub-optimal performance [56]. Besides, some work utilizes the labels and distribution of the target domain data to improve the model's adaptation performance [10,24,30], which is impractical due to accessing the unseen domain in advance. Differently from the above rPPG methods, we address this domain gap by using the TTA protocol, facilitating the adaptation to specific target domains during inference, based solely on unlabeled instances from the target domain." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.508, + 0.446, + 0.525 + ], + "angle": 0, + "content": "2.2 Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Test-time Adaptation (TTA), an emerging protocol, aims to adjust a pre-trained source model during inference by learning from the unlabeled target data, without accessing the distributions and labels of both source and target data, thereby naturally eliminate the need of intensive re-training [27, 61, 63]. Current TTA methods typically require aggregating a large number of samples from the target data to form a batch, and perform adaptation through 1) normalization calibration [1, 21, 22, 39, 74], 2) pseudo-labeling [8, 14], 3) input adaptation [15, 73], 4) dynamic inference that learns the network weights from a set of model parameters [72], 5) meta-learning [12], etc. These approaches can be considered as different forms of source-free domain adaptation (SFDA) [28] and are infeasible for regression tasks like rPPG where data arrives continuously and in a sequential manner. Some instance-level TTA approaches may either rely on specialized network architectures [1, 22, 61, 74] or incorporate additional auxiliary tasks applied during both pre-training and adaptation [11, 57]. By imposing specific non-general requirement to the pre-training stage, these methods narrow application scope of TTA. Moreover, some works even introduce a memory mechanism that turns instances into batches [21], lacking an in-time adaptation to each inference with increased computational cost. In contrast with above mentioned TTA methods, our Bi-TTA adapts the model during inference under each encountered instance from target domain without specific pre-training" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "H. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "or network architecture requirements, offering a flexible, effective, and efficient solution for rPPG model adaptation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.201, + 0.331, + 0.217 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.234, + 0.341, + 0.248 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.785, + 0.351 + ], + "angle": 0, + "content": "In this paper, we propose a novel expert knowledge-based Bi-directional Test-Time Adapter (Bi-TTA) to address the performance degradation in unseen target domains. In the following sections, we will firstly formulate the problem in Sec. 3.2. Secondly, Sec. 3.3 introduces the proposed self-supervised priors based on expert knowledge in the realm of rPPG. The bidirectional (prospective and retrospective) adaptation strategy of our Bi-TTA will be specified in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.373, + 0.387 + ], + "angle": 0, + "content": "3.2 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.784, + 0.503 + ], + "angle": 0, + "content": "The source dataset \\(D_{\\mathrm{S}} \\triangleq \\cup_{i=1}^{n_{\\mathrm{S}}} \\left\\{(\\pmb{x}_i, y_i)\\right\\}\\) is assumed to be drawn independently and identically distributed (i.i.d.) from distribution \\(X_{\\mathrm{S}}\\). Similarly, the target dataset \\(D_{\\mathrm{T}}\\) is assumed to be sampled from a different distribution \\(X_{\\mathrm{T}}\\), reflecting the domain shift. With a model parameterized by \\(\\pmb{w} \\in \\mathcal{W}\\) and considering the supervised loss function \\(L: \\mathcal{W} \\times \\mathcal{X} \\times \\mathcal{Y} \\rightarrow \\mathbb{R}_{+}\\) for evaluation, we define the observation loss over dataset \\(D\\) as an approximation of the population loss over distribution \\(X\\):" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.517, + 0.785, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} (\\boldsymbol {w}) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} L \\left(\\boldsymbol {w}, \\boldsymbol {x} _ {i}, y _ {i}\\right) \\approx \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim X} [ L (\\boldsymbol {w}, \\boldsymbol {x}, y) ]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.565, + 0.784, + 0.61 + ], + "angle": 0, + "content": "The objective of TTA, including our Bi-TTA, is to minimize the target observation loss \\(\\mathcal{L}_{D_{\\mathrm{T}}}(\\boldsymbol{w}_t)\\) after fine-tuning the initial model weights from \\(\\boldsymbol{w}_0\\) to \\(\\boldsymbol{w}_t\\) after \\(t\\) observed samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.525, + 0.648 + ], + "angle": 0, + "content": "3.3 Expert Knowledge-based Priors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.795 + ], + "angle": 0, + "content": "In the context of Bi-TTA, the unavailability of target data labels compels the use of self-supervised loss functions for providing the tuning gradient. Inspired by [17, 31,58,62], our Bi-TTA capitalizes on the characteristics of rPPG signals to craft these self-supervised loss functions, concentrating on the principles of spatial and temporal consistency. The introduced loss functions are termed as \\(L_{\\mathrm{s}}(\\pmb{w}_{t-1}, \\pmb{x}_t)\\) and \\(L_{\\mathrm{t}}(\\pmb{w}_{t-1}, \\pmb{x}_t)\\), focusing on spatial and temporal consistency respectively. These functions play a foundational role in the self-supervised tuning process, providing appropriate supervision based on the expert knowledge, despite the lack of direct label guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Spatial-temporal feature map (STMap). In line with [16, 33, 43, 48, 51], 2D STMap is firstly constructed to encapsulate the face video, as specified in Fig. 3a. This STMap integrates both temporal and spatial dimensions: the temporal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Bi-TTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.145, + 0.382, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.367, + 0.375, + 0.378 + ], + "angle": 0, + "content": "(a) STMap Construction" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.146, + 0.553, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.367, + 0.56, + 0.379 + ], + "angle": 0, + "content": "(b) Temporal Consistency Loss" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.144, + 0.787, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.367, + 0.756, + 0.379 + ], + "angle": 0, + "content": "(c) Spatial Consistency Loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.498 + ], + "angle": 0, + "content": "Fig. 3: Illustration of STMap construction and the implementation of our proposed expert knowledge-based priors. (a) The process of generating STMap, encompassing face alignment\\(^5\\) and cropping, local signal extraction, and the subsequent integration. (b) The calculation process of TCL, aimed at minimizing significant prediction discrepancies between original and temporally shifted HR predictions. (c) The calculation process the SCL, focused on penalizing pronounced disparities across different facial regions. Note that the boxes colored in pink represent the loss outcomes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.788, + 0.621 + ], + "angle": 0, + "content": "dimension represents the time sequence and the spatial dimension corresponds to the patches of the cropped face video frames. The current sample \\(\\pmb{x}_t\\in \\mathbb{R}^{W\\times H}\\) is referred to as a sliding window on the STMap, where \\(W\\) represents the sequential length of the sliding window, and \\(H\\) indicates the size of spatial dimension in the STMap." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Temporal consistency loss (TCL). BVP signals exhibit smooth and gradual fluctuations over short time spans. This fundamental attribute, known as the temporal consistency, is essential for accurate physiological signal analysis and the key to guarantee the predictability of physiological signal changes. To preserving this critical characteristic and derive a tuning gradient accordingly, the TCL is meticulously designed to enforce the model's prediction deviations of BVP signals when subjected to minor time perturbations under an acceptable tolerance. To compute the TCL, a random minor temporal perturbation \\(\\delta_{\\mathrm{T}}\\) is introduced by shifting the original sample \\(\\pmb{x}_t\\) along the STMap's temporal dimension. The shifted sample \\(\\pmb{x}_{t - \\delta_{\\mathrm{T}}}\\), in parallel with the origin sample, are then fed into the rPPG network \\(\\pmb{w}_{t - 1}\\) for predicting two sets of BVP signals. Then, for penalizing the model from large discrepancies in the HR estimation results between the original and shifted samples, \\(L_{1}\\) regularization is applied on the original and shifted HR prediction results, incorporating with a tolerance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "H. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "constant \\(\\xi_{\\mathrm{T}}\\). The specific calculation process of TCL is formulated in Eq. 2, as illustrated in Fig. 3b." + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.189, + 0.692, + 0.208 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {t, t - \\delta_ {\\mathrm {T}}} ^ {\\mathrm {H R}} = \\left\\| \\boldsymbol {w} _ {t - 1} ^ {\\mathrm {H R}} (\\boldsymbol {x} _ {t}) - \\boldsymbol {w} _ {t - 1} ^ {\\mathrm {H R}} (\\boldsymbol {x} _ {t - \\delta_ {\\mathrm {T}}}) \\right\\| _ {1} \\in \\mathbb {R} ^ {W \\times 1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.313, + 0.211, + 0.785, + 0.252 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {t}} \\left(\\boldsymbol {w} _ {t - 1}, \\boldsymbol {x} _ {t}\\right) = \\sum_ {i} ^ {W} \\max \\left(0, \\Delta_ {t, t - \\delta_ {\\mathrm {T}}} ^ {\\mathrm {H R} (i)} - \\xi_ {\\mathrm {T}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.794, + 0.445 + ], + "angle": 0, + "content": "Spatial consistency loss (SCL). Heart rate variations induce chromatic changes across the face, which are periodic and consistent across different facial regions. This phenomenon is referred to as spatial consistency, alongside temporal consistency, are crucial attributes in the rPPG field. Inspired by this phenomenon, SCL is crafted to promote the spatial consistency, obtaining correct adaptation gradient by encouraging the BVP signals from different facial areas to exhibit similarities. In particular, the SCL is calculated covering a series of latent feature maps, corresponding to various depths, (i.e., the number of residual blocks) in the adopted ResNet-18 [18] architecture. Given the current sample \\(\\pmb{x}_t\\in \\mathbb{R}^{W\\times H}\\), four latent feature maps \\(\\left\\{F_{i}\\in \\mathbb{R}^{W_{i}\\times H_{i}}\\mid i\\in [1,4]\\right\\}\\) could be extracted from the rPPG network \\(\\pmb{w}_{t - 1}\\). As illustrated in Fig. 3c, the calculation process of SCL is specify as:" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.447, + 0.785, + 0.488 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {s}} = \\sum_ {i} ^ {n} \\Delta F _ {i} = \\sum_ {i} ^ {n} \\sum_ {j} ^ {W _ {i} - \\delta_ {\\mathrm {S}}} | | F _ {i, j} - F _ {i, j + \\delta_ {\\mathrm {S}}} | | _ {1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.784, + 0.57 + ], + "angle": 0, + "content": "where \\( F_{i,j} = F_i[:,j] \\in \\mathbb{R}^{W_i} \\), \\( n = 4 \\) represents the number of residual blocks, and \\( \\delta_S \\) is the incremental spatial shift. Leveraging the fine-grained multi-level strategy, SCL effectively nurtures a comprehensive regularization of spatial features, ensuring that appropriate supervision is available for the adaptation during inference without annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.785, + 0.618 + ], + "angle": 0, + "content": "After combining \\( L_{\\mathrm{t}} \\) and \\( L_{\\mathrm{s}} \\), without direct target label reliance, the self-supervised loss is formulated by weighting and summing the spatial and temporal consistency priors as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.438, + 0.631, + 0.785, + 0.647 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {p}} = \\lambda_ {\\mathrm {s}} L _ {\\mathrm {s}} + \\lambda_ {\\mathrm {t}} L _ {\\mathrm {t}}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.669, + 0.557, + 0.685 + ], + "angle": 0, + "content": "3.4 Bidirectional Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.785, + 0.817 + ], + "angle": 0, + "content": "As introduced above, based on the proposed priors, the Bi-TTA framework aims to ensure effective and robust model adaptation under the variability and instability of single-instance learning conditions, minimizing the impact of domain-irrelevant noise. Simultaneously adapting the model with new target samples, our Bi-TTA also focuses on preserving the acquired adaptation ability dynamically, preventing the performance degradation caused by overfitting or catastrophic forgetting, achieving a stable maintenance of the adaptation performance in the target domain." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.534, + 0.841 + ], + "angle": 0, + "content": "3 https://github.com/adrianb/face-alignment." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Bi-TTA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.145, + 0.645, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig. 4: Illustration of the proposed Bidirectional Test-Time Adapter (Bi-TTA). Black arrows \\(\\rightarrow\\) indicate the adaptation process purely with the proposed two priors, i.e., TCL and SCL. Orange arrows \\(\\rightarrow\\) denote that the PA module adjusts model parameter using the gradient of representative neighborhood with a radius \\(\\rho\\). The green ones \\(\\rightarrow\\) show that the RS is activated when there is an oscillation, which is a sign of performance degradation, for maintaining the essential learned adaptation ability with former tuning gradients. The gradient and learning rate are formulated as \\(g\\) and \\(\\eta\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.66 + ], + "angle": 0, + "content": "Prospective adaptation (PA). During the process of model adaptation, especially under single-instance learning conditions, the noisy irrelevant information in biased target samples can easily interfere the model's tuning process, leading to a declined performance. Therefore, inspired by [13,38], rather than seeking out model parameter \\(\\pmb{w}_t\\) that simply have low adaptation loss \\(L_{\\mathrm{p}}(\\pmb{w}_t)\\), we seek out model parameter whose entire neighborhoods to have uniformly low adaptation loss." + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.662, + 0.786, + 0.687 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {p}} ^ {\\prime} (\\boldsymbol {w}) = \\max _ {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\rho} L _ {\\mathrm {p}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.803, + 0.754 + ], + "angle": 0, + "content": "where the \\(\\rho\\) is the disturbance range. This strategy, known as sharpness-aware minimization, has been proved to be effective in various areas [3, 19, 45, 67]. For computing efficiency and differentiability, we approximate the \\(\\epsilon^{*} = \\arg \\max_{\\| \\epsilon \\|_{2}\\leqslant \\rho}L_{\\mathrm{p}}(\\boldsymbol {w} + \\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.525, + 0.77 + ], + "angle": 0, + "content": "\\(\\epsilon)\\). Following 1-order Taylor series, we get:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.784, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {\\epsilon} ^ {*} \\approx \\underset {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\varrho} {\\arg \\max } L _ {\\mathrm {p}} (\\boldsymbol {w}) + \\boldsymbol {\\epsilon} ^ {\\mathrm {T}} \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w}) \\\\ = \\underset {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\rho} {\\arg \\max \\boldsymbol {\\epsilon}} ^ {\\mathrm {T}} \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w}) = \\underset {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\rho} {\\arg \\max \\boldsymbol {\\epsilon}} ^ {\\mathrm {T}} \\boldsymbol {g} = \\hat {\\boldsymbol {\\epsilon}}, \\tag {6} \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "H. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": ", w.r.t. \\(\\rho\\) around 0. Then employing the Cauchy-Schwarz inequality from Dual Norm theory: \\(\\epsilon^{\\mathrm{T}}\\pmb {g}\\leqslant \\| \\pmb {\\epsilon}\\|_{2}\\| \\pmb {g}\\|_{2}\\). \\(\\pmb{\\epsilon}\\) should be col-linear with \\(\\pmb{g}\\) to maximize \\(\\pmb{\\epsilon}^{\\mathrm{T}}\\pmb{g}\\). Considering \\(\\| \\pmb {\\epsilon}\\| _2\\leqslant \\rho\\), we can get:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.206, + 0.786, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} = \\rho \\frac {\\operatorname {s i g n} (\\boldsymbol {g}) \\boldsymbol {g}}{\\| \\boldsymbol {g} \\| _ {2}} = \\rho \\frac {\\operatorname {s i g n} \\left(\\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w})\\right) \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w})}{\\| \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w}) \\| _ {2}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.251, + 0.716, + 0.266 + ], + "angle": 0, + "content": "Consequently, we can derive the gradient approximation under Eq. 5:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.279, + 0.785, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}} = \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} ^ {\\prime} (\\boldsymbol {w}) \\approx \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\hat {\\boldsymbol {\\epsilon}}) \\\\ = \\left. \\frac {d (\\boldsymbol {w} + \\hat {\\epsilon})}{d \\boldsymbol {w}} \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\right| _ {\\boldsymbol {w} + \\hat {\\epsilon}} \\approx \\left. \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\right| _ {\\boldsymbol {w} + \\hat {\\epsilon}} \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.789, + 0.437 + ], + "angle": 0, + "content": "As illustrated in Fig. 4, the gradient \\( \\pmb{g}_t^{\\mathrm{PA}} \\) is adopted to replace the original gradient merely derived from two self-supervised priors. This strategy boosts the model's robustness to the undesired variability from instance-level conditions, safeguarding the tuning process against the influence from target-domain irrelevant information. Consequently, it secures both the efficacy and stability of the adaptation process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.79, + 0.697 + ], + "angle": 0, + "content": "Retrospective stabilization (RS). The PA module, by considering the model's entire neighborhood as the optimization objective, effectively enhances the model's robustness against irrelevant disturbances in target samples. However, after the model has processed a certain number of samples, it may still capture the detrimental noise feature instead of latent, universally domain-essential patterns, leading to performance degradation, as evidenced in Fig. 6. Therefore, to ensure reliable physiological signal predictions throughout the entire adaptation process, we propose the retrospective stabilization (RS) module. RS initially introduces the concept of trend gradient, termed as \\( \\pmb{g}^{*} \\). Starting with \\( \\pmb{g}_0^* = \\mathbf{0} \\), the trend gradient after processing \\( t \\) samples is updated via: \\( \\pmb{g}_t^* = \\frac{\\pmb{g}_{t - 1}^* + (1 / ||L_{\\mathrm{p}}||_1)\\pmb{g}_t^{\\mathrm{PA}}}{1 + 1 / ||L_{\\mathrm{p}}||_1} \\). For each \\( \\pmb{g}_t^{\\mathrm{PA}} \\), its orthogonal projection on \\( \\pmb{g}_t^* \\) is computed, formulated as \\( \\pmb{g}_t^\\perp = \\frac{\\pmb{g}_t^{\\mathrm{PA}}\\cdot\\pmb{g}_t^*}{\\|\\pmb{g}_t^*\\|^2}\\pmb{g}_t^* \\). As illustrated in Fig. 4, RS is activated if this projection aligns in the opposite of the trend gradient, which signifies an gradient oscillation, indicative of a potential fall of the generalization performance. Specifically, upon detection of oscillation, the model's optimization gradient will be the product of \\( \\pmb{g}_t^\\perp \\) and the backtracking degree \\( k \\). The orthogonal part of \\( \\pmb{g}_t \\) on \\( \\pmb{g}_t^* \\) is ignored." + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.709, + 0.785, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {t} ^ {\\mathrm {R S}} = \\left\\{ \\begin{array}{l} k \\cdot \\boldsymbol {g} _ {t} ^ {\\perp}, \\cos \\left(\\boldsymbol {g} _ {t} ^ {*}, \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}}\\right) < 0 \\\\ \\left(1 - \\lambda_ {t} ^ {\\mathrm {R S}}\\right) \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}} + \\lambda_ {t} ^ {\\mathrm {R S}} \\boldsymbol {g} _ {t} ^ {*}, \\cos \\left(\\boldsymbol {g} _ {t} ^ {*}, \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}}\\right) \\geqslant 0 \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\lambda_t^{\\mathrm{RS}} = -1 + 2 \\cdot \\operatorname{Sigmoid}(-\\frac{4t}{\\Omega}) \\in \\mathbb{R}_+^{(0,1)}\\), is the anneal scaling factor, \\(\\Omega\\) denotes the amount of target samples needed to ensure the fidelity of trend gradient, i.e., \\(\\lambda_t^{\\mathrm{RS}}\\) can be approximated as 1.0 when \\(t \\geqslant \\Omega\\). The obtained \\(\\pmb{g}_t^{\\mathrm{RS}}\\) represents the gradient that ultimately contributes to fine-tune the model for target domain adaptation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Bi-TTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: HR estimation results. Typically, the model selected for adaptation on each dataset is initially pre-trained on the remaining four datasets utilizing the NEST [34] framework. Ours w/o P.R. denotes that only the expert knowledge-based priors is adopted, without the bidirectional adaptation strategy. The best results are highlighted in bold, and the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.226, + 0.806, + 0.474 + ], + "angle": 0, + "content": "
MethodVIPL [42]V4V [49]PURE [53]UBFC [2]BUAA [68]
M↓R↓r↑M↓R↓r↑M↓R↓r↑M↓R↓r↑M↓R↓r↑
Methods based on DG protocol
Coral [54]8.6811.910.5310.3214.420.327.5910.870.725.898.040.763.645.740.8
VReX [23]8.3711.620.549.8214.160.377.2410.140.785.597.680.813.275.010.86
NCDG [59]8.4711.810.5210.1414.460.347.3210.350.775.317.560.823.125.160.85
NEST [34]7.8611.150.589.2713.790.416.719.590.814.676.790.862.884.690.89
Methods based on TTA protocol
Tent [61]8.0911.320.559.9811.730.476.869.680.844.579.740.872.373.590.92
SAR [40]8.1311.450.539.6811.460.476.339.30.864.569.640.872.072.930.95
TTT++ [32]7.7311.080.569.5111.310.546.239.140.884.366.790.872.052.820.92
EATA [39]7.6911.030.579.3611.120.516.139.230.864.256.860.881.892.640.94
SHOT [29]7.7511.070.569.5511.320.515.818.860.874.056.450.871.872.470.96
AdaODM [72]7.8111.210.559.110.930.555.518.290.884.016.590.871.912.60.95
ConPhys4 [58]7.4310.70.619.310.960.526.098.830.883.926.610.8861.752.190.97
Ours w/o P.R.7.3110.640.628.9710.820.555.568.670.883.646.630.8851.682.040.99
Ours w/o RS7.1510.830.638.9310.790.555.248.320.8893.596.610.8851.551.940.99
Ours w/o PA7.2410.780.629.0410.970.545.398.450.883.546.550.8881.511.980.99
Ours7.0910.720.639.111.020.545.028.110.8903.536.240.8911.491.970.99
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.502, + 0.377, + 0.519 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.535, + 0.335, + 0.548 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.787, + 0.606 + ], + "angle": 0, + "content": "We elaborately select five rPPG datasets, encompassing diverse races, levels of head motion, camera conditions, and ambient lighting, to establish a comprehensive large-scale benchmark of rPPG task under TTA protocol." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.787, + 0.697 + ], + "angle": 0, + "content": "VIPL-HR [42] have nine scenarios, three RGB cameras, different illumination conditions, and different levels of movement. It is worth mentioning that the BVP signal and video of the dataset do not match in the time dimension. We used the output of HR head as the evaluation result. Besides, we normalize STMap to 30 fps by cubic spline interpolation to solve the problem of the unstable frame rate of video [33,51,70]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.787, + 0.771 + ], + "angle": 0, + "content": "V4V [49] is designed to collect data with the drastic changes of physiological indicators by simulating ten tasks such as a funny joke, 911 emergency call, and odor experience. It is worth mentioning that There are only heart rate labels in the dataset and no BVP signal labels. We used the output of HR head as the evaluation result." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.772, + 0.787, + 0.803 + ], + "angle": 0, + "content": "BUAA [68] is proposed to evaluate the performance of the algorithm against various illumination. We only use data with illumination greater than or equal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "4 The proposed self-supervised loss in ConPhys [58] is applied for a comparative analysis of our proposed priors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "H. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.486, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.212, + 0.368, + 0.221 + ], + "angle": 0, + "content": "(a) \\(k\\)" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.221, + 0.487, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.288, + 0.372, + 0.299 + ], + "angle": 0, + "content": "(d) \\(\\xi_{\\mathrm{T}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.144, + 0.627, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.212, + 0.578, + 0.222 + ], + "angle": 0, + "content": "(b) \\(\\lambda_{t}\\)" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.223, + 0.627, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.289, + 0.574, + 0.299 + ], + "angle": 0, + "content": "(e) \\(\\rho\\)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.144, + 0.772, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.212, + 0.731, + 0.222 + ], + "angle": 0, + "content": "(c) \\(\\lambda_{s}\\)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.223, + 0.772, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.289, + 0.728, + 0.298 + ], + "angle": 0, + "content": "(f) \\(\\Omega\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.311, + 0.787, + 0.366 + ], + "angle": 0, + "content": "Fig. 5: Ablation experiments of hyper-parameters on down-sampled VIPL dataset5. Adheres to the default parameter configuration detailed in Sec.4.2, only one hyper-parameter is varied during each set of experiments. The pink arrows point to the lowest MAE result and the blue arrows mark the second if indiscernible." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.785, + 0.444 + ], + "angle": 0, + "content": "to 10 lux because underexposed images require special algorithms that are not considered in this article. We used the output of BVP head as the evaluation for BUAA, PURE, and UBFC-rPPG datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.784, + 0.506 + ], + "angle": 0, + "content": "PURE [53] contains 60 RGB videos from 10 subjects with six different activities, specifically, sitting still, talking, and four rotating and moving head variations. The BVP signals are down-sampled from 60 to 30 fps with cubic spline interpolation to align with the videos." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.508, + 0.784, + 0.539 + ], + "angle": 0, + "content": "UBFC-rPPG [2] containing 42 face videos with sunlight and indoor illumination, the ground-truth BVP signals, and HR values were collected by CMS50E." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.456, + 0.584 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.797 + ], + "angle": 0, + "content": "We implement our proposed priors and Bi-TTA using the PyTorch framework. The STMap sample is denoted as \\(\\pmb{x} \\in \\mathbb{R}^{256 \\times 25 \\times 3}\\), where 256 is the temporal length, i.e., the number of video frames, of the sliding window, and 25 is the spatial dimension of the STMap. It is firstly resized to \\(\\pmb{x}' \\in \\mathbb{R}^{256 \\times 64 \\times 3}\\) for network input. The network utilizes ResNet-18 [18] as the base architecture, followed by an additional fully connected layer for HR estimation. During the TCL calculation, the temporal perturbation \\(\\delta_{\\mathrm{T}}\\) is randomly sampled from a uniform distribution \\(\\mathbb{U}_{(0,59]}\\). Furthermore, we use stochastic gradient descent (SGD) with a momentum of 0.9 as the base optimizer. The bidirectional adaptation mechanism is implemented on top of the SGD. During the adaptation process, we set the learning rate to 0.0001 and the batch size is 1. The hyper-parameters are configured as follows: \\(\\lambda_{\\mathrm{s}} = 0.001\\), \\(\\xi_{\\mathrm{T}} = 8.0\\), \\(\\lambda_{\\mathrm{t}} = 0.01\\), \\(\\rho = 0.005\\), \\(k = -9.0\\), and \\(\\Omega = 4000\\). Please refer to Fig. 5 for detailed ablations." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5 Note that the performance of down-sampled VIPL dataset might be inferior than the full VIPL dataset, due to the larger interval between neighboring samples." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Bi-TTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.324, + 0.16 + ], + "angle": 0, + "content": "4.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.323 + ], + "angle": 0, + "content": "Following existing rPPG methods [4,33,34,44,51,70], mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (r) are used to evaluate the HR estimation, which inherently reflects the ability to extract physiological periodic signals. The results are organized and recorded in Tab. 1. It is evident that TTA methods generally perform better than DG approaches. Among these TTA methods, our Bi-TTA demonstrates significantly superior performance, showcasing our effectiveness especially when both prospective and retrospective adaptations are synergistically employed. This highlights Bi-TTA's robustness in adapting to various unseen domains, underscoring its potential for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.343, + 0.39, + 0.358 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.786, + 0.398 + ], + "angle": 0, + "content": "In this sub-section, we will mainly discuss about the effectiveness of the proposed priors and the bidirectional adaptation strategy." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.419, + 0.462, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.688, + 0.455, + 0.7 + ], + "angle": 0, + "content": "(a) Comparison between priors and GT." + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.421, + 0.787, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.688, + 0.753, + 0.7 + ], + "angle": 0, + "content": "(b) Ablation study of PA, RS, and, priors." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 6: (a) Comparison between task-specific priors-based TTA and fully supervised adaptation. The evaluation metrics are MSE, RMSE, and r. Alongside the visualization of the loss distribution for both priors-based and supervised adaptation method. Loss values smaller than \\(10^{-6}\\) are ignored. Note that for better comparison, the learning rate of supervised adaptation is set to 0.0005, while for the priors-based TTA it is 0.0001. (b) Ablation study of the adaptation performance among the proposed PA and RS modules in our Bi-TTA framework. This Ablation study demonstrates the effectiveness of the PA and RS, in achieving a more effective and stable adaptation process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "H. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Task-specific priors. As depicted in Fig. 6a, compared to ground-truth supervision, our task-specific priors provide effective adaptation supervision only on a significantly smaller number of samples, but induce considerably larger gradients. Thus, we can infer that the convergence rate under our self-supervised priors is markedly faster than ground-truth, despite a substantially lower learning rate. This phenomenon underscores the effectiveness of our priors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.363 + ], + "angle": 0, + "content": "As discussed in [71], modern deep learning models are prone to memorize training data due to their over-parameterized design, leading to overfitting risks. Merely minimizing standard loss functions on training data often falls short in ensuring satisfactory generalization ability [13]. Given that the rPPG model has undergone extensive pre-training with supervised learning, its parameters have likely been excessively optimized. In contrast, task knowledge-based priors are expected to offer fresh and correct guidance, steering the model towards improved generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.472 + ], + "angle": 0, + "content": "Prospective and retrospective adaptation. As depicted in Fig. 6b, the initial convergence rate and adaptation performance of the four comparisons are basically consistent across the first 4000 samples. However, in the latter half, the priors-based method demonstrates noticeable performance degradation. In contrast, other three methods, through utilizing prospective or retrospective adaptations, substantially mitigate this problem, thereby enhancing the adaptation ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Notably, between the \\(5000^{\\mathrm{th}}\\) and \\(6000^{\\mathrm{th}}\\) samples, while methods employing prospective adaptation exhibit rapid convergence and superior performance, they also display a significant degree of performance decay due to overfitting or catastrophic forgetting, which even intensifies as more samples are processed. On the other hand, the retrospective adaptation strategy ensures a more stable post-convergence maintenance in the target domain. The Bi-TTA, integrating both prospective and retrospective strategies, adeptly balances convergence speed, effectiveness, and stability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.638, + 0.36, + 0.656 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we address the challenge of significant performance degradation when adapting rPPG models to unfamiliar target domains. Considering privacy, we implement TTA for rPPG tasks, enabling the adaptation of pre-trained models to the target domain using only unannotated, instance-level target data. The models are fine-tuned during the inference of target samples. In particular, we introduce Bi-TTA, a method that leverages spatial and temporal consistency for effective self-supervision, coupled with pioneering prospective and retrospective adaptation strategies. This methodology streamlines the adaptation process and outperforms existing techniques, demonstrating substantial promise for real-world applications in non-invasive physiological monitoring." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Bi-TTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Bahmani, S., Hahn, O., Zamfir, E., Araslanov, N., Cremers, D., Roth, S.: Semantic self-adaptation: Enhancing generalization with a single sample. arXiv preprint arXiv:2208.05788 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.259 + ], + "angle": 0, + "content": "2. Bobbia, S., Macwan, R., Benezeth, Y., Mansouri, A., Dubois, J.: Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters 124, 82-90 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: International Conference on Machine Learning. pp. 1059-1071. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.328 + ], + "angle": 0, + "content": "4. Chen, W., McDuff, D.: Deepphys: Video-based physiological measurement using convolutional attention networks. In: Proc. ECCV. pp. 349-365 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.355 + ], + "angle": 0, + "content": "5. Das, A., Lu, H., Han, H., Dantcheva, A., Shan, S., Chen, X.: Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In: FG. pp. 01-08. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "6. De Haan, G., Jeanne, V.: Robust pulse rate from chrominance-based rppg. IEEE Trans. Biomed. Eng. 60(10), 2878-2886 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.411 + ], + "angle": 0, + "content": "7. De Haan, G., Van Leest, A.: Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiol. Meas. 35(9), 1913 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.439 + ], + "angle": 0, + "content": "8. D'Innocente, A., Bucci, S., Caputo, B., Tommasi, T.: Learning to generalize one sample at a time with self-supervision. arXiv preprint arXiv:1910.03915 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.785, + 0.493 + ], + "angle": 0, + "content": "9. Du, J., Liu, S.Q., Zhang, B., Yuen, P.C.: Dual-bridging with adversarial noise generation for domain adaptive rppg estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10355–10364 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.535 + ], + "angle": 0, + "content": "0. Du, J., Liu, S.Q., Zhang, B., Yuen, P.C.: Dual-bridging with adversarial noise generation for domain adaptive rppg estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10355–10364 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.591 + ], + "angle": 0, + "content": "1. D'Innocente, A., Borlino, F.C., Bucci, S., Caputo, B., Tommasi, T.: One-shot unsupervised cross-domain detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 732-748. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.592, + 0.785, + 0.659 + ], + "angle": 0, + "content": "12. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1126-1135. PMLR (06-11 Aug 2017), https://proceedings.mlr.press/v70/finn17a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "13. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=6Tm1mposlrM" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.785, + 0.757 + ], + "angle": 0, + "content": "14. Gandelsman, Y., Sun, Y., Chen, X., Efros, A.A.: Test-time training with masked autoencoders. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=SHMi1b7sjXk" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.812 + ], + "angle": 0, + "content": "15. Gao, J., Zhang, J., Liu, X., Darrell, T., Shelhamer, E., Wang, D.: Back to the source: Diffusion-driven adaptation to test-time corruption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11786-11796 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "16. Gee-Sern Hsu, ArulMurugan Ambikapathi, M.S.C.: Deep learning with time-frequency representation for pulse estimation. In: Proc. IJCB. pp. 642-650 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.176, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "H. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Gideon, J., Stent, S.: The way to my heart is through contrastive learning: Remote photoplethysmography from unlabeled video. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3995-4004 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.216 + ], + "angle": 0, + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. IEEE CVPR. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.271 + ], + "angle": 0, + "content": "19. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.298 + ], + "angle": 0, + "content": "20. Kessler, V., Thiam, P., Amirian, M., Schwenker, F.: Pain recognition with camera photoplethysmography. In: IEEE IPTA. pp. 1-5 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.325 + ], + "angle": 0, + "content": "21. Khurana, A., Paul, S., Rai, P., Biswas, S., Aggarwal, G.: Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "22. Klingner, M., Ayache, M., Fingscheidt, T.: Continual batchnorm adaptation (cbna) for semantic segmentation. IEEE Transactions on Intelligent Transportation Systems 23(11), 20899-20911 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.407 + ], + "angle": 0, + "content": "23. Krueger, D., Caballero, E., Jacobsen, J.H., Zhang, A., Binas, J., Zhang, D., Le Priol, R., Courville, A.: Out-of-distribution generalization via risk extrapolation (rex). In: ICML. pp. 5815-5826. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.461 + ], + "angle": 0, + "content": "24. Lee, E., Chen, E., Lee, C.Y.: Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16. pp. 392–409. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.501 + ], + "angle": 0, + "content": "25. Lewandowska, M., Ruminski, J., Kocejko, T., Nowak, J.: Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In: FedCSIS. pp. 405-410. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "26. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.555 + ], + "angle": 0, + "content": "27. Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "28. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International Conference on Machine Learning (ICML). pp. 6028-6039 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "29. Liang, J., Hu, D., Wang, Y., He, R., Feng, J.: Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021), in Press" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "30. Liu, X., Jiang, Z., Fromm, J., Xu, X., Patel, S., McDuff, D.: Metaphys: few-shot adaptation for non-contact physiological measurement. In: Proceedings of the conference on health, inference, and learning. pp. 154-163 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.718 + ], + "angle": 0, + "content": "31. Liu, X., Zhang, Y., Yu, Z., Lu, H., Yue, H., Yang, J.: rppg-mae: Self-supervised pretraining with masked autoencoders for remote physiological measurement. arXiv preprint arXiv:2306.02301 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "32. Liu, Y., Kothari, P., van Delft, B.G., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "33. Lu, H., Han, H., Zhou, S.K.: Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In: Proc. IEEE CVPR. pp. 12404-12413 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Lu, H., Yu, Z., Niu, X., Chen, Y.C.: Neuron structure modeling for generalizable remote physiological measurement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18589-18599 (June 2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.677, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Bi-TTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.75, + 0.161 + ], + "angle": 0, + "content": "35. McDuff, D.: Camera measurement of physiological vital signs. CSUR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.162, + 0.785, + 0.188 + ], + "angle": 0, + "content": "36. McDuff, D.: Applications of camera-based physiological measurement beyond healthcare. In: CVSM, pp. 165-177. Elsevier (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "37. McDuff, D., Gontarek, S., Picard, R.W.: Improvements in remote cardiopulmonary measurement using a five band digital camera. IEEE Trans. Biomed. Eng. 61(10), 2593-2601 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "38. Mi, P., Shen, L., Ren, T., Zhou, Y., Sun, X., Ji, R., Tao, D.: Make sharpness-aware minimization stronger: A sparsified perturbation approach (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.297 + ], + "angle": 0, + "content": "39. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "40. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.365 + ], + "angle": 0, + "content": "41. Niu, X., Han, H., Shan, S., Chen, X.: Synrhythm: Learning a deep heart rate estimator from general to specific. In: Proc. IEEE ICPR. pp. 3580-3585 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.785, + 0.392 + ], + "angle": 0, + "content": "42. Niu, X., Han, H., Shan, S., Chen, X.: VIPL-HR: A multi-modal database for pulse estimation from less-constrained face video. In: Proc. ACCV. pp. 562-576 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.432 + ], + "angle": 0, + "content": "43. Niu, X., Shan, S., Han, H., Chen, X.: Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Trans. on Image Process. 29, 2409-2423 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.459 + ], + "angle": 0, + "content": "44. Niu, X., Yu, Z., Han, H., Li, X., Shan, S., Zhao, G.: Video-based remote physiological measurement via cross-verified feature disentangling. In: Proc. ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "45. Pham, H., Dai, Z., Xie, Q., Le, Q.V.: Meta pseudo labels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11557-11568 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.54 + ], + "angle": 0, + "content": "46. Poh, M.Z., McDuff, D.J., Picard, R.W.: Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express 18(10), 10762-10774 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.568 + ], + "angle": 0, + "content": "47. Qiu, Y., Liu, Y., Arteaga-Falconi, J., Dong, H., El Saddik, A.: Evm-cnn: Real-time contactless heart rate estimation from facial video. IEEE Trans. Multimedia (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "48. Reiss, A., Indlekofer, I., Schmidt, P., Van Laerhoven, K.: Deep ppg: large-scale heart rate estimation with convolutional neural networks. Sensors 19(14), 3079 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.65 + ], + "angle": 0, + "content": "49. Revanur, A., Li, Z., Ciftci, U.A., Yin, L., Jeni, L.A.: The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In: Proc. CVPR workshop. pp. 2760-2767 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.691 + ], + "angle": 0, + "content": "50. Song, R., Chen, H., Cheng, J., Li, C., Liu, Y., Chen, X.: Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE J-BHI pp. 1-1 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.785, + 0.732 + ], + "angle": 0, + "content": "51. Song, R., Zhang, S., Li, C., Zhang, Y., Cheng, J., Chen, X.: Heart rate estimation from facial videos using a spatiotemporal representation with convolutional neural networks. IEEE Trans. Instrum. Meas. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.785, + 0.759 + ], + "angle": 0, + "content": "52. Spetlik, R., Franc, V., Matas, J.: Visual heart rate estimation with convolutional neural network. In: Proc. BMVC. pp. 3-6 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.785, + 0.786 + ], + "angle": 0, + "content": "53. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: Proc. IEEE ISRHIC. pp. 1056-1062 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.787, + 0.785, + 0.813 + ], + "angle": 0, + "content": "54. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: ECCV. pp. 443-450. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "55. Sun, B., Wei, Q., Li, L., Xu, Q., He, J., Yu, L.: LSTM for dynamic emotion and group emotion recognition in the wild. In: Proc. ACM ICMI. pp. 451-457 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "H. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "56. Sun, W., Zhang, X., Lu, H., Chen, Y., Ge, Y., Huang, X., Yuan, J., Chen, Y.: Resolve domain conflicts for generalizable remote physiological measurement. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8214-8224 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.244 + ], + "angle": 0, + "content": "57. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.787, + 0.284 + ], + "angle": 0, + "content": "58. Sun, Z., Li, X.: Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In: European Conference on Computer Vision. pp. 492-510. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.311 + ], + "angle": 0, + "content": "59. Tian, C.X., Li, H., Xie, X., Liu, Y., Wang, S.: Neuron coverage-guided domain generalization. IEEE Trans. on Pattern Anal. Mach. Intell. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.337 + ], + "angle": 0, + "content": "60. Verkruysse, W., Svaasand, L.O., Nelson, J.S.: Remote plethysmographic imaging using ambient light. Opt. Express 16(26), 21434-21445 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.337, + 0.787, + 0.378 + ], + "angle": 0, + "content": "61. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.378, + 0.787, + 0.431 + ], + "angle": 0, + "content": "62. Wang, H., Ahn, E., Kim, J.: Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2431-2439 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "63. Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Yu, P.: Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.472, + 0.787, + 0.499 + ], + "angle": 0, + "content": "64. Wang, W., den Brinker, A.C., Stuijk, S., de Haan, G.: Algorithmic principles of remote ppg. IEEE Trans. Biomed. Eng. 64(7), 1479-1491 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.499, + 0.787, + 0.526 + ], + "angle": 0, + "content": "65. Wang, W., Stuijk, S., De Haan, G.: Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE Trans. Biomed. Eng. 62(2), 415-425 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.526, + 0.787, + 0.552 + ], + "angle": 0, + "content": "66. Wang, Z.K., Kao, Y., Hsu, C.T.: Vision-based heart rate estimation via a two-stream cnn. In: Proc. IEEE ICIP. pp. 3327-3331 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.552, + 0.787, + 0.592 + ], + "angle": 0, + "content": "67. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems 33, 2958-2969 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "68. Xi, L., Chen, W., Zhao, C., Wu, X., Wang, J.: Image enhancement for remote photoplethysmography in a low-light environment. In: FG. pp. 1-7. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.659 + ], + "angle": 0, + "content": "69. Yu, Z., Peng, W., Li, X., Hong, X., Zhao, G.: Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In: Proc. IEEE ICCV. pp. 151-160 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.659, + 0.787, + 0.699 + ], + "angle": 0, + "content": "70. Yu, Z., Shen, Y., Shi, J., Zhao, H., Torr, P., Zhao, G.: Physformer: Facial videobased physiological measurement with temporal difference transformer. IEEE CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.699, + 0.787, + 0.74 + ], + "angle": 0, + "content": "71. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.74, + 0.787, + 0.766 + ], + "angle": 0, + "content": "72. Zhang, X., Chen, Y.C.: Adaptive domain generalization via online disagreement minimization. IEEE Transactions on Image Processing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.766, + 0.787, + 0.793 + ], + "angle": 0, + "content": "73. Zhao, X., Liu, C., Sicilia, A., Hwang, S.J., Fu, Y.: Test-time fourier style calibration for domain generalization. arXiv preprint arXiv:2205.06427 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.793, + 0.787, + 0.834 + ], + "angle": 0, + "content": "74. Zou, Y., Zhang, Z., Li, C.L., Zhang, H., Pfister, T., Huang, J.B.: Learning instance-specific adaptation for cross-domain segmentation. In: European Conference on Computer Vision. pp. 459-476. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.834 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_origin.pdf b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..592a426e629a933ee850f44e7b696522b66b8042 --- /dev/null +++ b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/815c5bdf-6893-4c0c-b246-02135f42536b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8de2360ccb0c9985a861cfc2eb4a01d61095ba1b55b4ebb5a9637553661e1f6c +size 5667528 diff --git a/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/full.md b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d2ce009de3858fb317cb4d9b2c38e755b5b362d0 --- /dev/null +++ b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/full.md @@ -0,0 +1,321 @@ +# Bi-TTA: Bidirectional Test-Time Adapter for Remote Physiological Measurement + +Haodong Li $^{1}$ , Hao Lu $^{1}$ , and Ying-Cong Chen $^{1,2}$ + +$^{1}$ Hong Kong University of Science and Technology (Guangzhou) +$^{2}$ Hong Kong University of Science and Technology {hli736, hlu585}@connect.hkust-gz.edu.cn yingcongchen@ust.hk + +![](images/6389c687b9a3b491fd6d49f3b19bf8e53b6f2e798a18fcd56311ab5aaccbac75.jpg) + +![](images/e04e8c0ee2cfdd7c4d6b29f3c2397700fd0a23b702a681053ad9fa7808826c06.jpg) + +![](images/6640b9c4e5ba28d39343e9493e0ed97ddbe9d668b0e52af050477b93b2117a84.jpg) + +![](images/f86273366369275de54b2e9a6c74260f00b6731a67c594228c23c85f25a87e1e.jpg) +(a) Pre-trained + +![](images/3fcb87f0ce485180fbb04f219c41a0353f8e1701f0840dce5b6fc852becd0c17.jpg) +(b) Priors-based TTA + +![](images/db83430207fa4fcb3793b29689e4a618b70ae5ba2ae8779f32013a62faf13e1a.jpg) +(c) Priors-based Bi-TTA +Fig.1: Visualization of the mean absolute error (MAE) result on down-sampled VIPL dataset using (a) the pre-trained rPPG model; (b) priors-based test-time adapted model; and (c) priors-based bidirectionally test-time adapted model. These visualizations demonstrate that our proposed priors and the Bi-TTA framework significantly enhance the generalization performance (reflected by the flatness of MAE field [26]) beyond the pre-trained model. + +Abstract. Remote photoplethysmography (rPPG) is gaining prominence for its non-invasive approach to monitoring physiological signals using only cameras. Despite its promise, the adaptability of rPPG models to new, unseen domains is hindered due to the environmental sensitivity of physiological signals. To address this issue, we pioneer the Test-Time Adaptation (TTA) in rPPG, enabling the adaptation of pre-trained models to the target domain during inference, sidestepping the need for annotations or source data due to privacy considerations. Particularly, utilizing only the user's face video stream as the accessible target domain data, the rPPG model is adjusted by tuning on each single instance it encounters. However, 1) TTA algorithms are designed predominantly for classifi- + +cation tasks, ill-suited in regression tasks such as rPPG due to inadequate supervision. 2) Tuning pre-trained models in a single-instance manner introduces variability and instability, posing challenges to effectively filtering domain-relevant from domain-irrelevant features while simultaneously preserving the learned information. To overcome these challenges, we present Bi-TTA, a novel expert knowledge-based Bidirectional Test-Time Adapter framework. Specifically, leveraging two expert-knowledge priors for providing self-supervision, our Bi-TTA primarily comprises two modules: a prospective adaptation (PA) module using sharpness-aware minimization to eliminate domain-irrelevant noise, enhancing the stability and efficacy during the adaptation process, and a retrospective stabilization (RS) module to dynamically reinforce crucial learned model parameters, averting performance degradation caused by overfitting or catastrophic forgetting. To this end, we established a large-scale benchmark for rPPG tasks under TTA protocol, promoting advancements in both the rPPG and TTA fields. The experimental results demonstrate the significant superiority of our approach over the state-of-the-art (SoTA). + +Keywords: Bidirectional Test-Time Adaptation $\cdot$ Expert Knowledge-based Priors $\cdot$ Remote Photoplethysmography + +# 1 Introduction + +The cardiac cycle refers to the process experienced by the cardiovascular system from the start of one heartbeat to the beginning of the next. Through the analysis of the cardiac cycle, important physiological indicators such as heart rate (HR), respiration rate (RF), and heart rate variability (HRV) can be derived, playing crucial roles in medical diagnostics, health surveillance, and forensic analysis [20, 36, 55]. + +Traditional physiological measurement devices like electrocardiograms (ECG), heart rate bands, and finger-clips, despite their accuracy, are often expensive and uncomfortable to wear. In contrast, rPPG offers a non-invasive and more convenient alternative by extracting blood volume pulses (BVP) from facial videos, analyzing the skin's light absorption variations to measure HR, HRV, and RF [44,69,70]. These monitors are especially important for tracking health status and sympathetic activity levels [20,36,55]. Nowadays, physiological measurements based on rPPG, using only ordinary cameras, are increasingly becoming the focus of research in computer vision community [5,9,33,35,43,47,50]. However, the periodic color changes in the skin captured by remote cameras are quite subtle, and variables like natural scene illumination and head movements can significantly impair the reliability of physiological indicator measurements. + +![](images/10619055dd8175ecb6a1b49574c013f82ad76c4aaa2dcdcf5dd754293df96f9c.jpg) +Fig. 2: Visualization of Domain Adaptation (DA) and Test-Time Adaptation (TTA) methodologies. DA utilizes batch learning with labeled target data, while TTA dynamically refines the model during inference without relying on target data labels or distribution. Both DA and TTA obviate the requirement for source data. Note that domain generalization (DG) is excluded as it does not focus on specific target domain adaptation. + +Conventional rPPG methods [6,25,37,46,60,64] perform well only in constrained environments. Recent deep learning (DL) methods [2,43,49,53,68] have demonstrate promising results, however, the most of training data are collected in controlled laboratory environments with limited variations in illumination, camera parameters, etc. Therefore, these methods still struggle to generalize to unseen domains, necessitating model adaptation. + +For privacy considerations, the adaptation of rPPG models in unseen scenarios is required to be performed without accessing the source data, and target data annotations or distributions (i.e., large batches of target data samples). Addressing this, we firstly implement Test-Time Adaptation (TTA) in the field of rPPG. TTA aims to fine-tune a pre-trained source model during inference time, as illustrated in Fig. 2, without accessing the distribution and labeling of both source data and target data and naturally eliminate the need for intensive re-training [61]. This innovation allows for the automatic adaptation of pretrained rPPG models to unseen domains during inference, without the necessity for labelling or batches of target data. + +However, applying TTA directly into rPPG faces two key limitations. 1) Predominantly designed for classification tasks, most TTA algorithms leverage entropy in normalization layers or pseudo labels. These strategies are not inherently suited for regression tasks like rPPG, leading to a lack of appropriate supervision. 2) In real-world scenarios, the pre-trained rPPG model is tuned during inference the user's face video stream, on an instance-by-instance context. This variability and instability introduced by biased instances pose challenges to effectively filtering out domain-irrelevant noise and simultaneously preserving crucial learned information as the model acquiring new knowledge. This nuanced balance is critical in preventing slow or ineffective adaptation under single-instance learning condition, ensuring an efficient and consistent adaptation process. + +To navigate these challenges, we proposed a novel Bidirectional Test-Time Adapter (Bi-TTA) framework, grounded on two expert knowledge-based self-supervised priors. Firstly, two regularizations are proposed for offering effective supervision especially in rPPG tasks, addressing the aforementioned ill-suited problem of applying existing TTA algorithms into regression tasks. These two priors focus on temporal and spatial consistency respectively, which is derived from the characteristics of physiological signals. Then, two strategies are introduced for stable and effective fine-tuning process. The prospective adaptation (PA) strategy is designed for filtering out the irrelevant information of each encountered instance, aiming that only target-domain-essential feature is truly contributing to model adaptation. The retrospective stabilization (RS) strategy backtracks and consolidates essential model parameters dynamically, leveraging previous update gradients during the self-supervised tuning process, ensuring that the learned adaptation ability is preserved intact. We evaluate our method across various datasets, benchmarking against existing prevailing TTA algorithms, demonstrating the superiority of our Bi-TTA. Overall, our contributions can be summarized as follows: + +- We introduce the TTA protocol to rPPG tasks, paving the way for a new paradigm in self-supervised adapting of rPPG models during inference. +- We introduce two novel self-supervised priors to regularize temporal and spatial consistency, providing appropriate supervision. +- We propose a novel Bidirectional Test-Time Adapter (Bi-TTA) framework grounded on the introduced priors, which promotes effective and efficient adaptation, with enhanced stability. +- We established a large-scale benchmark for rPPG tasks under TTA protocol. Extensive experiments demonstrate significantly improved adaptation capacity of our approach. + +# 2 Related Works + +# 2.1 Remote Physiological Measurement + +Remote physiological measurement aims to quantify HR and HRV from BVP signals by interpreting skin chrominance fluctuations. Traditional techniques typically involve color space transformations or signal decomposition to obtain the BVP signal with high signal-to-noise ratio (SNR) [6, 7, 46, 65]. With powerful modeling capacity of deep learning (DL), there's been a significant shift towards employing DL models for remote physiological estimation, yielding impressive results. [4, 16, 41, 43, 52, 66, 69]. Many of these methods extract BVP signals from face videos typically by aggregating spatial and temporal information [44, 52, 69, 70]. Besides, many works utilize hand-designed 2D feature maps extracted from face videos to alleviate computational demands [16, 33, 43, 48, 51]. Recently, researchers have turned to exploring how to leverage task-specific knowledge and how to improve the generalization performance of the model into unseen scenarios. + +Task-specific priors. rPPG is distinguished by its unique biological and physical mechanisms, which can be exploited to enhance the physiological signal extraction ability of the model [17,31,58,62]. These task-specific priors can be summed up as 1) spatial consistency: BVP signals extracted from different face areas should exhibit similar; 2) temporal consistency: BVP signals typically exhibit smooth and gradual changes over short time period; 3) heart rate range: The heart rate typically ranges between 40 and 250 beats per minute (bpm). Different from existing works, we provide a more granular solution leveraging the multi-scale latent features, rather than purely focusing on the prediction output of BVP signals. + +The generalization of the model. With the domain shifts (e.g., lighting, head motion, environment, etc.), most methods tend to have significant performance degradation. To improve the generalization performance on unseen domains, NEST-rPPG established a large-scale DG evaluation protocol [34]. These DG methods, requiring the access of all source data, do not focus on the adaptation of a specific target domain, leading to sub-optimal performance [56]. Besides, some work utilizes the labels and distribution of the target domain data to improve the model's adaptation performance [10,24,30], which is impractical due to accessing the unseen domain in advance. Differently from the above rPPG methods, we address this domain gap by using the TTA protocol, facilitating the adaptation to specific target domains during inference, based solely on unlabeled instances from the target domain. + +# 2.2 Test-Time Adaptation + +Test-time Adaptation (TTA), an emerging protocol, aims to adjust a pre-trained source model during inference by learning from the unlabeled target data, without accessing the distributions and labels of both source and target data, thereby naturally eliminate the need of intensive re-training [27, 61, 63]. Current TTA methods typically require aggregating a large number of samples from the target data to form a batch, and perform adaptation through 1) normalization calibration [1, 21, 22, 39, 74], 2) pseudo-labeling [8, 14], 3) input adaptation [15, 73], 4) dynamic inference that learns the network weights from a set of model parameters [72], 5) meta-learning [12], etc. These approaches can be considered as different forms of source-free domain adaptation (SFDA) [28] and are infeasible for regression tasks like rPPG where data arrives continuously and in a sequential manner. Some instance-level TTA approaches may either rely on specialized network architectures [1, 22, 61, 74] or incorporate additional auxiliary tasks applied during both pre-training and adaptation [11, 57]. By imposing specific non-general requirement to the pre-training stage, these methods narrow application scope of TTA. Moreover, some works even introduce a memory mechanism that turns instances into batches [21], lacking an in-time adaptation to each inference with increased computational cost. In contrast with above mentioned TTA methods, our Bi-TTA adapts the model during inference under each encountered instance from target domain without specific pre-training + +or network architecture requirements, offering a flexible, effective, and efficient solution for rPPG model adaptation. + +# 3 Method + +# 3.1 Overview + +In this paper, we propose a novel expert knowledge-based Bi-directional Test-Time Adapter (Bi-TTA) to address the performance degradation in unseen target domains. In the following sections, we will firstly formulate the problem in Sec. 3.2. Secondly, Sec. 3.3 introduces the proposed self-supervised priors based on expert knowledge in the realm of rPPG. The bidirectional (prospective and retrospective) adaptation strategy of our Bi-TTA will be specified in Sec. 3.4. + +# 3.2 Preliminaries + +The source dataset $D_{\mathrm{S}} \triangleq \cup_{i=1}^{n_{\mathrm{S}}} \left\{(\pmb{x}_i, y_i)\right\}$ is assumed to be drawn independently and identically distributed (i.i.d.) from distribution $X_{\mathrm{S}}$ . Similarly, the target dataset $D_{\mathrm{T}}$ is assumed to be sampled from a different distribution $X_{\mathrm{T}}$ , reflecting the domain shift. With a model parameterized by $\pmb{w} \in \mathcal{W}$ and considering the supervised loss function $L: \mathcal{W} \times \mathcal{X} \times \mathcal{Y} \rightarrow \mathbb{R}_{+}$ for evaluation, we define the observation loss over dataset $D$ as an approximation of the population loss over distribution $X$ : + +$$ +\mathcal {L} _ {D} (\boldsymbol {w}) = \frac {1}{n} \sum_ {i = 1} ^ {n} L \left(\boldsymbol {w}, \boldsymbol {x} _ {i}, y _ {i}\right) \approx \mathbb {E} _ {(\boldsymbol {x}, y) \sim X} [ L (\boldsymbol {w}, \boldsymbol {x}, y) ]. \tag {1} +$$ + +The objective of TTA, including our Bi-TTA, is to minimize the target observation loss $\mathcal{L}_{D_{\mathrm{T}}}(\boldsymbol{w}_t)$ after fine-tuning the initial model weights from $\boldsymbol{w}_0$ to $\boldsymbol{w}_t$ after $t$ observed samples. + +# 3.3 Expert Knowledge-based Priors + +In the context of Bi-TTA, the unavailability of target data labels compels the use of self-supervised loss functions for providing the tuning gradient. Inspired by [17, 31,58,62], our Bi-TTA capitalizes on the characteristics of rPPG signals to craft these self-supervised loss functions, concentrating on the principles of spatial and temporal consistency. The introduced loss functions are termed as $L_{\mathrm{s}}(\pmb{w}_{t-1}, \pmb{x}_t)$ and $L_{\mathrm{t}}(\pmb{w}_{t-1}, \pmb{x}_t)$ , focusing on spatial and temporal consistency respectively. These functions play a foundational role in the self-supervised tuning process, providing appropriate supervision based on the expert knowledge, despite the lack of direct label guidance. + +Spatial-temporal feature map (STMap). In line with [16, 33, 43, 48, 51], 2D STMap is firstly constructed to encapsulate the face video, as specified in Fig. 3a. This STMap integrates both temporal and spatial dimensions: the temporal + +![](images/40947f9f907e1baf30091467befe69c31d0af0ef43d8582f3daecea6d250c58d.jpg) +(a) STMap Construction + +![](images/12b439476b69a08ce4ba6ac3915b0e01f74268ad969fe04902220cb5073dd7ee.jpg) +(b) Temporal Consistency Loss + +![](images/04e3232af5c821d9716d0b3cb0ac88bcd077e1b6aae56f3663d33372d53bad20.jpg) +(c) Spatial Consistency Loss +Fig. 3: Illustration of STMap construction and the implementation of our proposed expert knowledge-based priors. (a) The process of generating STMap, encompassing face alignment $^5$ and cropping, local signal extraction, and the subsequent integration. (b) The calculation process of TCL, aimed at minimizing significant prediction discrepancies between original and temporally shifted HR predictions. (c) The calculation process the SCL, focused on penalizing pronounced disparities across different facial regions. Note that the boxes colored in pink represent the loss outcomes. + +dimension represents the time sequence and the spatial dimension corresponds to the patches of the cropped face video frames. The current sample $\pmb{x}_t\in \mathbb{R}^{W\times H}$ is referred to as a sliding window on the STMap, where $W$ represents the sequential length of the sliding window, and $H$ indicates the size of spatial dimension in the STMap. + +Temporal consistency loss (TCL). BVP signals exhibit smooth and gradual fluctuations over short time spans. This fundamental attribute, known as the temporal consistency, is essential for accurate physiological signal analysis and the key to guarantee the predictability of physiological signal changes. To preserving this critical characteristic and derive a tuning gradient accordingly, the TCL is meticulously designed to enforce the model's prediction deviations of BVP signals when subjected to minor time perturbations under an acceptable tolerance. To compute the TCL, a random minor temporal perturbation $\delta_{\mathrm{T}}$ is introduced by shifting the original sample $\pmb{x}_t$ along the STMap's temporal dimension. The shifted sample $\pmb{x}_{t - \delta_{\mathrm{T}}}$ , in parallel with the origin sample, are then fed into the rPPG network $\pmb{w}_{t - 1}$ for predicting two sets of BVP signals. Then, for penalizing the model from large discrepancies in the HR estimation results between the original and shifted samples, $L_{1}$ regularization is applied on the original and shifted HR prediction results, incorporating with a tolerance + +constant $\xi_{\mathrm{T}}$ . The specific calculation process of TCL is formulated in Eq. 2, as illustrated in Fig. 3b. + +$$ +\Delta_ {t, t - \delta_ {\mathrm {T}}} ^ {\mathrm {H R}} = \left\| \boldsymbol {w} _ {t - 1} ^ {\mathrm {H R}} (\boldsymbol {x} _ {t}) - \boldsymbol {w} _ {t - 1} ^ {\mathrm {H R}} (\boldsymbol {x} _ {t - \delta_ {\mathrm {T}}}) \right\| _ {1} \in \mathbb {R} ^ {W \times 1} +$$ + +$$ +L _ {\mathrm {t}} \left(\boldsymbol {w} _ {t - 1}, \boldsymbol {x} _ {t}\right) = \sum_ {i} ^ {W} \max \left(0, \Delta_ {t, t - \delta_ {\mathrm {T}}} ^ {\mathrm {H R} (i)} - \xi_ {\mathrm {T}}\right). \tag {2} +$$ + +Spatial consistency loss (SCL). Heart rate variations induce chromatic changes across the face, which are periodic and consistent across different facial regions. This phenomenon is referred to as spatial consistency, alongside temporal consistency, are crucial attributes in the rPPG field. Inspired by this phenomenon, SCL is crafted to promote the spatial consistency, obtaining correct adaptation gradient by encouraging the BVP signals from different facial areas to exhibit similarities. In particular, the SCL is calculated covering a series of latent feature maps, corresponding to various depths, (i.e., the number of residual blocks) in the adopted ResNet-18 [18] architecture. Given the current sample $\pmb{x}_t\in \mathbb{R}^{W\times H}$ , four latent feature maps $\left\{F_{i}\in \mathbb{R}^{W_{i}\times H_{i}}\mid i\in [1,4]\right\}$ could be extracted from the rPPG network $\pmb{w}_{t - 1}$ . As illustrated in Fig. 3c, the calculation process of SCL is specify as: + +$$ +L _ {\mathrm {s}} = \sum_ {i} ^ {n} \Delta F _ {i} = \sum_ {i} ^ {n} \sum_ {j} ^ {W _ {i} - \delta_ {\mathrm {S}}} | | F _ {i, j} - F _ {i, j + \delta_ {\mathrm {S}}} | | _ {1}, \tag {3} +$$ + +where $F_{i,j} = F_i[:,j] \in \mathbb{R}^{W_i}$ , $n = 4$ represents the number of residual blocks, and $\delta_S$ is the incremental spatial shift. Leveraging the fine-grained multi-level strategy, SCL effectively nurtures a comprehensive regularization of spatial features, ensuring that appropriate supervision is available for the adaptation during inference without annotations. + +After combining $L_{\mathrm{t}}$ and $L_{\mathrm{s}}$ , without direct target label reliance, the self-supervised loss is formulated by weighting and summing the spatial and temporal consistency priors as follows: + +$$ +L _ {\mathrm {p}} = \lambda_ {\mathrm {s}} L _ {\mathrm {s}} + \lambda_ {\mathrm {t}} L _ {\mathrm {t}}. \tag {4} +$$ + +# 3.4 Bidirectional Test-Time Adaptation + +As introduced above, based on the proposed priors, the Bi-TTA framework aims to ensure effective and robust model adaptation under the variability and instability of single-instance learning conditions, minimizing the impact of domain-irrelevant noise. Simultaneously adapting the model with new target samples, our Bi-TTA also focuses on preserving the acquired adaptation ability dynamically, preventing the performance degradation caused by overfitting or catastrophic forgetting, achieving a stable maintenance of the adaptation performance in the target domain. + +![](images/5094995fdae497c6b6c8b35239887c626464363180a8057bba77fdded0a44d85.jpg) +Fig. 4: Illustration of the proposed Bidirectional Test-Time Adapter (Bi-TTA). Black arrows $\rightarrow$ indicate the adaptation process purely with the proposed two priors, i.e., TCL and SCL. Orange arrows $\rightarrow$ denote that the PA module adjusts model parameter using the gradient of representative neighborhood with a radius $\rho$ . The green ones $\rightarrow$ show that the RS is activated when there is an oscillation, which is a sign of performance degradation, for maintaining the essential learned adaptation ability with former tuning gradients. The gradient and learning rate are formulated as $g$ and $\eta$ respectively. + +Prospective adaptation (PA). During the process of model adaptation, especially under single-instance learning conditions, the noisy irrelevant information in biased target samples can easily interfere the model's tuning process, leading to a declined performance. Therefore, inspired by [13,38], rather than seeking out model parameter $\pmb{w}_t$ that simply have low adaptation loss $L_{\mathrm{p}}(\pmb{w}_t)$ , we seek out model parameter whose entire neighborhoods to have uniformly low adaptation loss. + +$$ +L _ {\mathrm {p}} ^ {\prime} (\boldsymbol {w}) = \max _ {\| \boldsymbol {\epsilon} \| _ {2} \leqslant \rho} L _ {\mathrm {p}} (\boldsymbol {w} + \boldsymbol {\epsilon}), \tag {5} +$$ + +where the $\rho$ is the disturbance range. This strategy, known as sharpness-aware minimization, has been proved to be effective in various areas [3, 19, 45, 67]. For computing efficiency and differentiability, we approximate the $\epsilon^{*} = \arg \max_{\| \epsilon \|_{2}\leqslant \rho}L_{\mathrm{p}}(\boldsymbol {w} +$ + +$\epsilon)$ . Following 1-order Taylor series, we get: + +$$ +\begin{array}{l} \boldsymbol {\epsilon} ^ {*} \approx \underset {\| \boldsymbol {\epsilon} \| _ {2} \leqslant \varrho} {\arg \max } L _ {\mathrm {p}} (\boldsymbol {w}) + \boldsymbol {\epsilon} ^ {\mathrm {T}} \nabla_ {\boldsymbol {w}} L _ {\mathrm {p}} (\boldsymbol {w}) \\ = \underset {\| \boldsymbol {\epsilon} \| _ {2} \leqslant \rho} {\arg \max \boldsymbol {\epsilon}} ^ {\mathrm {T}} \nabla_ {\boldsymbol {w}} L _ {\mathrm {p}} (\boldsymbol {w}) = \underset {\| \boldsymbol {\epsilon} \| _ {2} \leqslant \rho} {\arg \max \boldsymbol {\epsilon}} ^ {\mathrm {T}} \boldsymbol {g} = \hat {\boldsymbol {\epsilon}}, \tag {6} \\ \end{array} +$$ + +, w.r.t. $\rho$ around 0. Then employing the Cauchy-Schwarz inequality from Dual Norm theory: $\epsilon^{\mathrm{T}}\pmb {g}\leqslant \| \pmb {\epsilon}\|_{2}\| \pmb {g}\|_{2}$ . $\pmb{\epsilon}$ should be col-linear with $\pmb{g}$ to maximize $\pmb{\epsilon}^{\mathrm{T}}\pmb{g}$ . Considering $\| \pmb {\epsilon}\| _2\leqslant \rho$ , we can get: + +$$ +\hat {\epsilon} = \rho \frac {\operatorname {s i g n} (\boldsymbol {g}) \boldsymbol {g}}{\| \boldsymbol {g} \| _ {2}} = \rho \frac {\operatorname {s i g n} \left(\nabla_ {\boldsymbol {w}} L _ {\mathrm {p}} (\boldsymbol {w})\right) \nabla_ {\boldsymbol {w}} L _ {\mathrm {p}} (\boldsymbol {w})}{\| \nabla_ {\boldsymbol {w}} L _ {\mathrm {p}} (\boldsymbol {w}) \| _ {2}}. \tag {7} +$$ + +Consequently, we can derive the gradient approximation under Eq. 5: + +$$ +\begin{array}{l} \boldsymbol {g} _ {t} ^ {\mathrm {P A}} = \nabla_ {\boldsymbol {w}} L _ {\mathrm {p}} ^ {\prime} (\boldsymbol {w}) \approx \nabla_ {\boldsymbol {w}} L _ {\mathcal {S}} (\boldsymbol {w} + \hat {\boldsymbol {\epsilon}}) \\ = \left. \frac {d (\boldsymbol {w} + \hat {\epsilon})}{d \boldsymbol {w}} \nabla_ {\boldsymbol {w}} L _ {\mathcal {S}} (\boldsymbol {w}) \right| _ {\boldsymbol {w} + \hat {\epsilon}} \approx \left. \nabla_ {\boldsymbol {w}} L _ {\mathcal {S}} (\boldsymbol {w}) \right| _ {\boldsymbol {w} + \hat {\epsilon}} \tag {8} \\ \end{array} +$$ + +As illustrated in Fig. 4, the gradient $\pmb{g}_t^{\mathrm{PA}}$ is adopted to replace the original gradient merely derived from two self-supervised priors. This strategy boosts the model's robustness to the undesired variability from instance-level conditions, safeguarding the tuning process against the influence from target-domain irrelevant information. Consequently, it secures both the efficacy and stability of the adaptation process. + +Retrospective stabilization (RS). The PA module, by considering the model's entire neighborhood as the optimization objective, effectively enhances the model's robustness against irrelevant disturbances in target samples. However, after the model has processed a certain number of samples, it may still capture the detrimental noise feature instead of latent, universally domain-essential patterns, leading to performance degradation, as evidenced in Fig. 6. Therefore, to ensure reliable physiological signal predictions throughout the entire adaptation process, we propose the retrospective stabilization (RS) module. RS initially introduces the concept of trend gradient, termed as $\pmb{g}^{*}$ . Starting with $\pmb{g}_0^* = \mathbf{0}$ , the trend gradient after processing $t$ samples is updated via: $\pmb{g}_t^* = \frac{\pmb{g}_{t - 1}^* + (1 / ||L_{\mathrm{p}}||_1)\pmb{g}_t^{\mathrm{PA}}}{1 + 1 / ||L_{\mathrm{p}}||_1}$ . For each $\pmb{g}_t^{\mathrm{PA}}$ , its orthogonal projection on $\pmb{g}_t^*$ is computed, formulated as $\pmb{g}_t^\perp = \frac{\pmb{g}_t^{\mathrm{PA}}\cdot\pmb{g}_t^*}{\|\pmb{g}_t^*\|^2}\pmb{g}_t^*$ . As illustrated in Fig. 4, RS is activated if this projection aligns in the opposite of the trend gradient, which signifies an gradient oscillation, indicative of a potential fall of the generalization performance. Specifically, upon detection of oscillation, the model's optimization gradient will be the product of $\pmb{g}_t^\perp$ and the backtracking degree $k$ . The orthogonal part of $\pmb{g}_t$ on $\pmb{g}_t^*$ is ignored. + +$$ +\boldsymbol {g} _ {t} ^ {\mathrm {R S}} = \left\{ \begin{array}{l} k \cdot \boldsymbol {g} _ {t} ^ {\perp}, \cos \left(\boldsymbol {g} _ {t} ^ {*}, \boldsymbol {g} _ {t} ^ {\mathrm {P A}}\right) < 0 \\ \left(1 - \lambda_ {t} ^ {\mathrm {R S}}\right) \boldsymbol {g} _ {t} ^ {\mathrm {P A}} + \lambda_ {t} ^ {\mathrm {R S}} \boldsymbol {g} _ {t} ^ {*}, \cos \left(\boldsymbol {g} _ {t} ^ {*}, \boldsymbol {g} _ {t} ^ {\mathrm {P A}}\right) \geqslant 0 \end{array} \right. \tag {9} +$$ + +where $\lambda_t^{\mathrm{RS}} = -1 + 2 \cdot \operatorname{Sigmoid}(-\frac{4t}{\Omega}) \in \mathbb{R}_+^{(0,1)}$ , is the anneal scaling factor, $\Omega$ denotes the amount of target samples needed to ensure the fidelity of trend gradient, i.e., $\lambda_t^{\mathrm{RS}}$ can be approximated as 1.0 when $t \geqslant \Omega$ . The obtained $\pmb{g}_t^{\mathrm{RS}}$ represents the gradient that ultimately contributes to fine-tune the model for target domain adaptation. + +Table 1: HR estimation results. Typically, the model selected for adaptation on each dataset is initially pre-trained on the remaining four datasets utilizing the NEST [34] framework. Ours w/o P.R. denotes that only the expert knowledge-based priors is adopted, without the bidirectional adaptation strategy. The best results are highlighted in bold, and the second-best results are underlined. + +
MethodVIPL [42]V4V [49]PURE [53]UBFC [2]BUAA [68]
M↓R↓r↑M↓R↓r↑M↓R↓r↑M↓R↓r↑M↓R↓r↑
Methods based on DG protocol
Coral [54]8.6811.910.5310.3214.420.327.5910.870.725.898.040.763.645.740.8
VReX [23]8.3711.620.549.8214.160.377.2410.140.785.597.680.813.275.010.86
NCDG [59]8.4711.810.5210.1414.460.347.3210.350.775.317.560.823.125.160.85
NEST [34]7.8611.150.589.2713.790.416.719.590.814.676.790.862.884.690.89
Methods based on TTA protocol
Tent [61]8.0911.320.559.9811.730.476.869.680.844.579.740.872.373.590.92
SAR [40]8.1311.450.539.6811.460.476.339.30.864.569.640.872.072.930.95
TTT++ [32]7.7311.080.569.5111.310.546.239.140.884.366.790.872.052.820.92
EATA [39]7.6911.030.579.3611.120.516.139.230.864.256.860.881.892.640.94
SHOT [29]7.7511.070.569.5511.320.515.818.860.874.056.450.871.872.470.96
AdaODM [72]7.8111.210.559.110.930.555.518.290.884.016.590.871.912.60.95
ConPhys4 [58]7.4310.70.619.310.960.526.098.830.883.926.610.8861.752.190.97
Ours w/o P.R.7.3110.640.628.9710.820.555.568.670.883.646.630.8851.682.040.99
Ours w/o RS7.1510.830.638.9310.790.555.248.320.8893.596.610.8851.551.940.99
Ours w/o PA7.2410.780.629.0410.970.545.398.450.883.546.550.8881.511.980.99
Ours7.0910.720.639.111.020.545.028.110.8903.536.240.8911.491.970.99
+ +# 4 Experiments + +# 4.1 Datasets + +We elaborately select five rPPG datasets, encompassing diverse races, levels of head motion, camera conditions, and ambient lighting, to establish a comprehensive large-scale benchmark of rPPG task under TTA protocol. + +VIPL-HR [42] have nine scenarios, three RGB cameras, different illumination conditions, and different levels of movement. It is worth mentioning that the BVP signal and video of the dataset do not match in the time dimension. We used the output of HR head as the evaluation result. Besides, we normalize STMap to 30 fps by cubic spline interpolation to solve the problem of the unstable frame rate of video [33,51,70]. + +V4V [49] is designed to collect data with the drastic changes of physiological indicators by simulating ten tasks such as a funny joke, 911 emergency call, and odor experience. It is worth mentioning that There are only heart rate labels in the dataset and no BVP signal labels. We used the output of HR head as the evaluation result. + +BUAA [68] is proposed to evaluate the performance of the algorithm against various illumination. We only use data with illumination greater than or equal + +![](images/f73fb16e85592b530001867f3bc1db5fdf1c5346ea3582618f5438b7346c4740.jpg) +(a) $k$ + +![](images/ca0988a0006d3a1c41b677483a66cd3bc5bbd943feb8350b2bbdf1fd44fd65ce.jpg) +(d) $\xi_{\mathrm{T}}$ + +![](images/09e95419af086bb65a14654621222dccf7bd02728e2028509c8760437b0761dd.jpg) +(b) $\lambda_{t}$ + +![](images/049ce48e4b3afd1cfaef5076b0f953022bbf08345fd39b0e6c2fe82cdd63ee80.jpg) +(e) $\rho$ +Fig. 5: Ablation experiments of hyper-parameters on down-sampled VIPL dataset5. Adheres to the default parameter configuration detailed in Sec.4.2, only one hyper-parameter is varied during each set of experiments. The pink arrows point to the lowest MAE result and the blue arrows mark the second if indiscernible. + +![](images/ded933a3ada4d1dc625115cf256cc2ce9dd46d2259274fc81fc9e9edeb1f5e51.jpg) +(c) $\lambda_{s}$ + +![](images/226cdb6d02d5db366b7fd637888e0dc80225c8aab3c46933125e5310a56c0fe3.jpg) +(f) $\Omega$ + +to 10 lux because underexposed images require special algorithms that are not considered in this article. We used the output of BVP head as the evaluation for BUAA, PURE, and UBFC-rPPG datasets. + +PURE [53] contains 60 RGB videos from 10 subjects with six different activities, specifically, sitting still, talking, and four rotating and moving head variations. The BVP signals are down-sampled from 60 to 30 fps with cubic spline interpolation to align with the videos. + +UBFC-rPPG [2] containing 42 face videos with sunlight and indoor illumination, the ground-truth BVP signals, and HR values were collected by CMS50E. + +# 4.2 Implementation Details + +We implement our proposed priors and Bi-TTA using the PyTorch framework. The STMap sample is denoted as $\pmb{x} \in \mathbb{R}^{256 \times 25 \times 3}$ , where 256 is the temporal length, i.e., the number of video frames, of the sliding window, and 25 is the spatial dimension of the STMap. It is firstly resized to $\pmb{x}' \in \mathbb{R}^{256 \times 64 \times 3}$ for network input. The network utilizes ResNet-18 [18] as the base architecture, followed by an additional fully connected layer for HR estimation. During the TCL calculation, the temporal perturbation $\delta_{\mathrm{T}}$ is randomly sampled from a uniform distribution $\mathbb{U}_{(0,59]}$ . Furthermore, we use stochastic gradient descent (SGD) with a momentum of 0.9 as the base optimizer. The bidirectional adaptation mechanism is implemented on top of the SGD. During the adaptation process, we set the learning rate to 0.0001 and the batch size is 1. The hyper-parameters are configured as follows: $\lambda_{\mathrm{s}} = 0.001$ , $\xi_{\mathrm{T}} = 8.0$ , $\lambda_{\mathrm{t}} = 0.01$ , $\rho = 0.005$ , $k = -9.0$ , and $\Omega = 4000$ . Please refer to Fig. 5 for detailed ablations. + +# 4.3 Results + +Following existing rPPG methods [4,33,34,44,51,70], mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (r) are used to evaluate the HR estimation, which inherently reflects the ability to extract physiological periodic signals. The results are organized and recorded in Tab. 1. It is evident that TTA methods generally perform better than DG approaches. Among these TTA methods, our Bi-TTA demonstrates significantly superior performance, showcasing our effectiveness especially when both prospective and retrospective adaptations are synergistically employed. This highlights Bi-TTA's robustness in adapting to various unseen domains, underscoring its potential for real-world applications. + +# 4.4 Ablation Study + +In this sub-section, we will mainly discuss about the effectiveness of the proposed priors and the bidirectional adaptation strategy. + +![](images/3c4faa4af001fabbbdba4f202a0f41e9d91c3c5883dbc485122816b11610854e.jpg) +(a) Comparison between priors and GT. +Fig. 6: (a) Comparison between task-specific priors-based TTA and fully supervised adaptation. The evaluation metrics are MSE, RMSE, and r. Alongside the visualization of the loss distribution for both priors-based and supervised adaptation method. Loss values smaller than $10^{-6}$ are ignored. Note that for better comparison, the learning rate of supervised adaptation is set to 0.0005, while for the priors-based TTA it is 0.0001. (b) Ablation study of the adaptation performance among the proposed PA and RS modules in our Bi-TTA framework. This Ablation study demonstrates the effectiveness of the PA and RS, in achieving a more effective and stable adaptation process. + +![](images/eebbcab48fbd0376c3d4c5df22baafedeb0d42dadb2203a3b6bf8129ad13bac2.jpg) +(b) Ablation study of PA, RS, and, priors. + +Task-specific priors. As depicted in Fig. 6a, compared to ground-truth supervision, our task-specific priors provide effective adaptation supervision only on a significantly smaller number of samples, but induce considerably larger gradients. Thus, we can infer that the convergence rate under our self-supervised priors is markedly faster than ground-truth, despite a substantially lower learning rate. This phenomenon underscores the effectiveness of our priors. + +As discussed in [71], modern deep learning models are prone to memorize training data due to their over-parameterized design, leading to overfitting risks. Merely minimizing standard loss functions on training data often falls short in ensuring satisfactory generalization ability [13]. Given that the rPPG model has undergone extensive pre-training with supervised learning, its parameters have likely been excessively optimized. In contrast, task knowledge-based priors are expected to offer fresh and correct guidance, steering the model towards improved generalization capability. + +Prospective and retrospective adaptation. As depicted in Fig. 6b, the initial convergence rate and adaptation performance of the four comparisons are basically consistent across the first 4000 samples. However, in the latter half, the priors-based method demonstrates noticeable performance degradation. In contrast, other three methods, through utilizing prospective or retrospective adaptations, substantially mitigate this problem, thereby enhancing the adaptation ability. + +Notably, between the $5000^{\mathrm{th}}$ and $6000^{\mathrm{th}}$ samples, while methods employing prospective adaptation exhibit rapid convergence and superior performance, they also display a significant degree of performance decay due to overfitting or catastrophic forgetting, which even intensifies as more samples are processed. On the other hand, the retrospective adaptation strategy ensures a more stable post-convergence maintenance in the target domain. The Bi-TTA, integrating both prospective and retrospective strategies, adeptly balances convergence speed, effectiveness, and stability. + +# 5 Conclusion + +In this paper, we address the challenge of significant performance degradation when adapting rPPG models to unfamiliar target domains. Considering privacy, we implement TTA for rPPG tasks, enabling the adaptation of pre-trained models to the target domain using only unannotated, instance-level target data. The models are fine-tuned during the inference of target samples. In particular, we introduce Bi-TTA, a method that leverages spatial and temporal consistency for effective self-supervision, coupled with pioneering prospective and retrospective adaptation strategies. This methodology streamlines the adaptation process and outperforms existing techniques, demonstrating substantial promise for real-world applications in non-invasive physiological monitoring. + +# References + +1. Bahmani, S., Hahn, O., Zamfir, E., Araslanov, N., Cremers, D., Roth, S.: Semantic self-adaptation: Enhancing generalization with a single sample. arXiv preprint arXiv:2208.05788 (2022) +2. Bobbia, S., Macwan, R., Benezeth, Y., Mansouri, A., Dubois, J.: Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters 124, 82-90 (2019) +3. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: International Conference on Machine Learning. pp. 1059-1071. PMLR (2021) +4. Chen, W., McDuff, D.: Deepphys: Video-based physiological measurement using convolutional attention networks. In: Proc. ECCV. pp. 349-365 (2018) +5. Das, A., Lu, H., Han, H., Dantcheva, A., Shan, S., Chen, X.: Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In: FG. pp. 01-08. IEEE (2021) +6. De Haan, G., Jeanne, V.: Robust pulse rate from chrominance-based rppg. IEEE Trans. Biomed. Eng. 60(10), 2878-2886 (2013) +7. De Haan, G., Van Leest, A.: Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiol. Meas. 35(9), 1913 (2014) +8. D'Innocente, A., Bucci, S., Caputo, B., Tommasi, T.: Learning to generalize one sample at a time with self-supervision. arXiv preprint arXiv:1910.03915 (2019) +9. Du, J., Liu, S.Q., Zhang, B., Yuen, P.C.: Dual-bridging with adversarial noise generation for domain adaptive rppg estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10355–10364 (June 2023) +0. Du, J., Liu, S.Q., Zhang, B., Yuen, P.C.: Dual-bridging with adversarial noise generation for domain adaptive rppg estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10355–10364 (2023) +1. D'Innocente, A., Borlino, F.C., Bucci, S., Caputo, B., Tommasi, T.: One-shot unsupervised cross-domain detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 732-748. Springer (2020) +12. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1126-1135. PMLR (06-11 Aug 2017), https://proceedings.mlr.press/v70/finn17a.html +13. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=6Tm1mposlrM +14. Gandelsman, Y., Sun, Y., Chen, X., Efros, A.A.: Test-time training with masked autoencoders. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=SHMi1b7sjXk +15. Gao, J., Zhang, J., Liu, X., Darrell, T., Shelhamer, E., Wang, D.: Back to the source: Diffusion-driven adaptation to test-time corruption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11786-11796 (2023) +16. Gee-Sern Hsu, ArulMurugan Ambikapathi, M.S.C.: Deep learning with time-frequency representation for pulse estimation. In: Proc. IJCB. pp. 642-650 (2017) + +17. Gideon, J., Stent, S.: The way to my heart is through contrastive learning: Remote photoplethysmography from unlabeled video. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3995-4004 (2021) +18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. IEEE CVPR. pp. 770-778 (2016) +19. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021) +20. Kessler, V., Thiam, P., Amirian, M., Schwenker, F.: Pain recognition with camera photoplethysmography. In: IEEE IPTA. pp. 1-5 (2017) +21. Khurana, A., Paul, S., Rai, P., Biswas, S., Aggarwal, G.: Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355 (2021) +22. Klingner, M., Ayache, M., Fingscheidt, T.: Continual batchnorm adaptation (cbna) for semantic segmentation. IEEE Transactions on Intelligent Transportation Systems 23(11), 20899-20911 (2022) +23. Krueger, D., Caballero, E., Jacobsen, J.H., Zhang, A., Binas, J., Zhang, D., Le Priol, R., Courville, A.: Out-of-distribution generalization via risk extrapolation (rex). In: ICML. pp. 5815-5826. PMLR (2021) +24. Lee, E., Chen, E., Lee, C.Y.: Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16. pp. 392–409. Springer (2020) +25. Lewandowska, M., Ruminski, J., Kocejko, T., Nowak, J.: Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In: FedCSIS. pp. 405-410. IEEE (2011) +26. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018) +27. Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361 (2023) +28. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International Conference on Machine Learning (ICML). pp. 6028-6039 (2020) +29. Liang, J., Hu, D., Wang, Y., He, R., Feng, J.: Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021), in Press +30. Liu, X., Jiang, Z., Fromm, J., Xu, X., Patel, S., McDuff, D.: Metaphys: few-shot adaptation for non-contact physiological measurement. In: Proceedings of the conference on health, inference, and learning. pp. 154-163 (2021) +31. Liu, X., Zhang, Y., Yu, Z., Lu, H., Yue, H., Yang, J.: rppg-mae: Self-supervised pretraining with masked autoencoders for remote physiological measurement. arXiv preprint arXiv:2306.02301 (2023) +32. Liu, Y., Kothari, P., van Delft, B.G., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Thirty-Fifth Conference on Neural Information Processing Systems (2021) +33. Lu, H., Han, H., Zhou, S.K.: Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In: Proc. IEEE CVPR. pp. 12404-12413 (2021) +34. Lu, H., Yu, Z., Niu, X., Chen, Y.C.: Neuron structure modeling for generalizable remote physiological measurement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18589-18599 (June 2023) + +35. McDuff, D.: Camera measurement of physiological vital signs. CSUR (2021) +36. McDuff, D.: Applications of camera-based physiological measurement beyond healthcare. In: CVSM, pp. 165-177. Elsevier (2022) +37. McDuff, D., Gontarek, S., Picard, R.W.: Improvements in remote cardiopulmonary measurement using a five band digital camera. IEEE Trans. Biomed. Eng. 61(10), 2593-2601 (2014) +38. Mi, P., Shen, L., Ren, T., Zhou, Y., Sun, X., Ji, R., Tao, D.: Make sharpness-aware minimization stronger: A sparsified perturbation approach (2022) +39. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022) +40. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023) +41. Niu, X., Han, H., Shan, S., Chen, X.: Synrhythm: Learning a deep heart rate estimator from general to specific. In: Proc. IEEE ICPR. pp. 3580-3585 (2018) +42. Niu, X., Han, H., Shan, S., Chen, X.: VIPL-HR: A multi-modal database for pulse estimation from less-constrained face video. In: Proc. ACCV. pp. 562-576 (2018) +43. Niu, X., Shan, S., Han, H., Chen, X.: Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Trans. on Image Process. 29, 2409-2423 (2020) +44. Niu, X., Yu, Z., Han, H., Li, X., Shan, S., Zhao, G.: Video-based remote physiological measurement via cross-verified feature disentangling. In: Proc. ECCV (2020) +45. Pham, H., Dai, Z., Xie, Q., Le, Q.V.: Meta pseudo labels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11557-11568 (2021) +46. Poh, M.Z., McDuff, D.J., Picard, R.W.: Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express 18(10), 10762-10774 (2010) +47. Qiu, Y., Liu, Y., Arteaga-Falconi, J., Dong, H., El Saddik, A.: Evm-cnn: Real-time contactless heart rate estimation from facial video. IEEE Trans. Multimedia (2019) +48. Reiss, A., Indlekofer, I., Schmidt, P., Van Laerhoven, K.: Deep ppg: large-scale heart rate estimation with convolutional neural networks. Sensors 19(14), 3079 (2019) +49. Revanur, A., Li, Z., Ciftci, U.A., Yin, L., Jeni, L.A.: The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In: Proc. CVPR workshop. pp. 2760-2767 (2021) +50. Song, R., Chen, H., Cheng, J., Li, C., Liu, Y., Chen, X.: Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE J-BHI pp. 1-1 (2021) +51. Song, R., Zhang, S., Li, C., Zhang, Y., Cheng, J., Chen, X.: Heart rate estimation from facial videos using a spatiotemporal representation with convolutional neural networks. IEEE Trans. Instrum. Meas. (2020) +52. Spetlik, R., Franc, V., Matas, J.: Visual heart rate estimation with convolutional neural network. In: Proc. BMVC. pp. 3-6 (2018) +53. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: Proc. IEEE ISRHIC. pp. 1056-1062 (2014) +54. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: ECCV. pp. 443-450. Springer (2016) +55. Sun, B., Wei, Q., Li, L., Xu, Q., He, J., Yu, L.: LSTM for dynamic emotion and group emotion recognition in the wild. In: Proc. ACM ICMI. pp. 451-457 (2016) + +56. Sun, W., Zhang, X., Lu, H., Chen, Y., Ge, Y., Huang, X., Yuan, J., Chen, Y.: Resolve domain conflicts for generalizable remote physiological measurement. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8214-8224 (2023) +57. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020) +58. Sun, Z., Li, X.: Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In: European Conference on Computer Vision. pp. 492-510. Springer (2022) +59. Tian, C.X., Li, H., Xie, X., Liu, Y., Wang, S.: Neuron coverage-guided domain generalization. IEEE Trans. on Pattern Anal. Mach. Intell. (2022) +60. Verkruysse, W., Svaasand, L.O., Nelson, J.S.: Remote plethysmographic imaging using ambient light. Opt. Express 16(26), 21434-21445 (2008) +61. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c +62. Wang, H., Ahn, E., Kim, J.: Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2431-2439 (2022) +63. Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Yu, P.: Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering (2022) +64. Wang, W., den Brinker, A.C., Stuijk, S., de Haan, G.: Algorithmic principles of remote ppg. IEEE Trans. Biomed. Eng. 64(7), 1479-1491 (2017) +65. Wang, W., Stuijk, S., De Haan, G.: Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE Trans. Biomed. Eng. 62(2), 415-425 (2015) +66. Wang, Z.K., Kao, Y., Hsu, C.T.: Vision-based heart rate estimation via a two-stream cnn. In: Proc. IEEE ICIP. pp. 3327-3331 (2019) +67. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems 33, 2958-2969 (2020) +68. Xi, L., Chen, W., Zhao, C., Wu, X., Wang, J.: Image enhancement for remote photoplethysmography in a low-light environment. In: FG. pp. 1-7. IEEE (2020) +69. Yu, Z., Peng, W., Li, X., Hong, X., Zhao, G.: Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In: Proc. IEEE ICCV. pp. 151-160 (2019) +70. Yu, Z., Shen, Y., Shi, J., Zhao, H., Torr, P., Zhao, G.: Physformer: Facial videobased physiological measurement with temporal difference transformer. IEEE CVPR (2022) +71. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021) +72. Zhang, X., Chen, Y.C.: Adaptive domain generalization via online disagreement minimization. IEEE Transactions on Image Processing (2023) +73. Zhao, X., Liu, C., Sicilia, A., Hwang, S.J., Fu, Y.: Test-time fourier style calibration for domain generalization. arXiv preprint arXiv:2205.06427 (2022) +74. Zou, Y., Zhang, Z., Li, C.L., Zhang, H., Pfister, T., Huang, J.B.: Learning instance-specific adaptation for cross-domain segmentation. In: European Conference on Computer Vision. pp. 459-476. Springer (2022) \ No newline at end of file diff --git a/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/images.zip b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7aed4519602a05061b3d1d2088e59daba6272cfd --- /dev/null +++ b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46fd464256b3f08c37fa39a6b1b2cf2bdcedb26caf38d620905b4f92b18edd85 +size 567740 diff --git a/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/layout.json b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe1b6ae7ca8f0b1bca43cc4ffd8736a5cb5a4f32 --- /dev/null +++ b/2024/Bi-TTA_ Bidirectional Test-Time Adapter for Remote Physiological Measurement/layout.json @@ -0,0 +1,10089 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 112, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 112, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 148, + 112, + 465, + 148 + ], + "type": "text", + "content": "Bi-TTA: Bidirectional Test-Time Adapter for Remote Physiological Measurement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 167, + 412, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 167, + 412, + 181 + ], + "spans": [ + { + "bbox": [ + 201, + 167, + 412, + 181 + ], + "type": "text", + "content": "Haodong Li" + }, + { + "bbox": [ + 201, + 167, + 412, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 167, + 412, + 181 + ], + "type": "text", + "content": ", Hao Lu" + }, + { + "bbox": [ + 201, + 167, + 412, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 167, + 412, + 181 + ], + "type": "text", + "content": ", and Ying-Cong Chen" + }, + { + "bbox": [ + 201, + 167, + 412, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 167, + 189, + 446, + 223 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 171, + 189, + 438, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 189, + 438, + 201 + ], + "spans": [ + { + "bbox": [ + 171, + 189, + 438, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 189, + 438, + 201 + ], + "type": "text", + "content": " Hong Kong University of Science and Technology (Guangzhou)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 167, + 201, + 446, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 201, + 446, + 223 + ], + "spans": [ + { + "bbox": [ + 167, + 201, + 446, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 167, + 201, + 446, + 223 + ], + "type": "text", + "content": " Hong Kong University of Science and Technology {hli736, hlu585}@connect.hkust-gz.edu.cn yingcongchen@ust.hk" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 152, + 264, + 242, + 326 + ], + "blocks": [ + { + "bbox": [ + 152, + 264, + 242, + 326 + ], + "lines": [ + { + "bbox": [ + 152, + 264, + 242, + 326 + ], + "spans": [ + { + "bbox": [ + 152, + 264, + 242, + 326 + ], + "type": "image", + "image_path": "6389c687b9a3b491fd6d49f3b19bf8e53b6f2e798a18fcd56311ab5aaccbac75.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 261, + 239, + 356, + 330 + ], + "blocks": [ + { + "bbox": [ + 261, + 239, + 356, + 330 + ], + "lines": [ + { + "bbox": [ + 261, + 239, + 356, + 330 + ], + "spans": [ + { + "bbox": [ + 261, + 239, + 356, + 330 + ], + "type": "image", + "image_path": "e04e8c0ee2cfdd7c4d6b29f3c2397700fd0a23b702a681053ad9fa7808826c06.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 372, + 239, + 465, + 331 + ], + "blocks": [ + { + "bbox": [ + 372, + 239, + 465, + 331 + ], + "lines": [ + { + "bbox": [ + 372, + 239, + 465, + 331 + ], + "spans": [ + { + "bbox": [ + 372, + 239, + 465, + 331 + ], + "type": "image", + "image_path": "6640b9c4e5ba28d39343e9493e0ed97ddbe9d668b0e52af050477b93b2117a84.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 143, + 340, + 249, + 419 + ], + "blocks": [ + { + "bbox": [ + 143, + 340, + 249, + 419 + ], + "lines": [ + { + "bbox": [ + 143, + 340, + 249, + 419 + ], + "spans": [ + { + "bbox": [ + 143, + 340, + 249, + 419 + ], + "type": "image", + "image_path": "f86273366369275de54b2e9a6c74260f00b6731a67c594228c23c85f25a87e1e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 424, + 223, + 435 + ], + "lines": [ + { + "bbox": [ + 162, + 424, + 223, + 435 + ], + "spans": [ + { + "bbox": [ + 162, + 424, + 223, + 435 + ], + "type": "text", + "content": "(a) Pre-trained" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 254, + 341, + 360, + 419 + ], + "blocks": [ + { + "bbox": [ + 254, + 341, + 360, + 419 + ], + "lines": [ + { + "bbox": [ + 254, + 341, + 360, + 419 + ], + "spans": [ + { + "bbox": [ + 254, + 341, + 360, + 419 + ], + "type": "image", + "image_path": "3fcb87f0ce485180fbb04f219c41a0353f8e1701f0840dce5b6fc852becd0c17.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 259, + 424, + 348, + 435 + ], + "lines": [ + { + "bbox": [ + 259, + 424, + 348, + 435 + ], + "spans": [ + { + "bbox": [ + 259, + 424, + 348, + 435 + ], + "type": "text", + "content": "(b) Priors-based TTA" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 364, + 341, + 470, + 419 + ], + "blocks": [ + { + "bbox": [ + 364, + 341, + 470, + 419 + ], + "lines": [ + { + "bbox": [ + 364, + 341, + 470, + 419 + ], + "spans": [ + { + "bbox": [ + 364, + 341, + 470, + 419 + ], + "type": "image", + "image_path": "db83430207fa4fcb3793b29689e4a618b70ae5ba2ae8779f32013a62faf13e1a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 424, + 464, + 435 + ], + "lines": [ + { + "bbox": [ + 364, + 424, + 464, + 435 + ], + "spans": [ + { + "bbox": [ + 364, + 424, + 464, + 435 + ], + "type": "text", + "content": "(c) Priors-based Bi-TTA" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 453, + 482, + 518 + ], + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 518 + ], + "type": "text", + "content": "Fig.1: Visualization of the mean absolute error (MAE) result on down-sampled VIPL dataset using (a) the pre-trained rPPG model; (b) priors-based test-time adapted model; and (c) priors-based bidirectionally test-time adapted model. These visualizations demonstrate that our proposed priors and the Bi-TTA framework significantly enhance the generalization performance (reflected by the flatness of MAE field [26]) beyond the pre-trained model." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 159, + 544, + 457, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 544, + 457, + 665 + ], + "spans": [ + { + "bbox": [ + 159, + 544, + 457, + 665 + ], + "type": "text", + "content": "Abstract. Remote photoplethysmography (rPPG) is gaining prominence for its non-invasive approach to monitoring physiological signals using only cameras. Despite its promise, the adaptability of rPPG models to new, unseen domains is hindered due to the environmental sensitivity of physiological signals. To address this issue, we pioneer the Test-Time Adaptation (TTA) in rPPG, enabling the adaptation of pre-trained models to the target domain during inference, sidestepping the need for annotations or source data due to privacy considerations. Particularly, utilizing only the user's face video stream as the accessible target domain data, the rPPG model is adjusted by tuning on each single instance it encounters. However, 1) TTA algorithms are designed predominantly for classifi-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 160, + 117, + 455, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 455, + 315 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 455, + 315 + ], + "type": "text", + "content": "cation tasks, ill-suited in regression tasks such as rPPG due to inadequate supervision. 2) Tuning pre-trained models in a single-instance manner introduces variability and instability, posing challenges to effectively filtering domain-relevant from domain-irrelevant features while simultaneously preserving the learned information. To overcome these challenges, we present Bi-TTA, a novel expert knowledge-based Bidirectional Test-Time Adapter framework. Specifically, leveraging two expert-knowledge priors for providing self-supervision, our Bi-TTA primarily comprises two modules: a prospective adaptation (PA) module using sharpness-aware minimization to eliminate domain-irrelevant noise, enhancing the stability and efficacy during the adaptation process, and a retrospective stabilization (RS) module to dynamically reinforce crucial learned model parameters, averting performance degradation caused by overfitting or catastrophic forgetting. To this end, we established a large-scale benchmark for rPPG tasks under TTA protocol, promoting advancements in both the rPPG and TTA fields. The experimental results demonstrate the significant superiority of our approach over the state-of-the-art (SoTA)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 325, + 453, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 325, + 453, + 348 + ], + "spans": [ + { + "bbox": [ + 160, + 325, + 453, + 348 + ], + "type": "text", + "content": "Keywords: Bidirectional Test-Time Adaptation " + }, + { + "bbox": [ + 160, + 325, + 453, + 348 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 325, + 453, + 348 + ], + "type": "text", + "content": " Expert Knowledge-based Priors " + }, + { + "bbox": [ + 160, + 325, + 453, + 348 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 325, + 453, + 348 + ], + "type": "text", + "content": " Remote Photoplethysmography" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 385, + 230, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 230, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 230, + 398 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 410, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 481 + ], + "type": "text", + "content": "The cardiac cycle refers to the process experienced by the cardiovascular system from the start of one heartbeat to the beginning of the next. Through the analysis of the cardiac cycle, important physiological indicators such as heart rate (HR), respiration rate (RF), and heart rate variability (HRV) can be derived, playing crucial roles in medical diagnostics, health surveillance, and forensic analysis [20, 36, 55]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "text", + "content": "Traditional physiological measurement devices like electrocardiograms (ECG), heart rate bands, and finger-clips, despite their accuracy, are often expensive and uncomfortable to wear. In contrast, rPPG offers a non-invasive and more convenient alternative by extracting blood volume pulses (BVP) from facial videos, analyzing the skin's light absorption variations to measure HR, HRV, and RF [44,69,70]. These monitors are especially important for tracking health status and sympathetic activity levels [20,36,55]. Nowadays, physiological measurements based on rPPG, using only ordinary cameras, are increasingly becoming the focus of research in computer vision community [5,9,33,35,43,47,50]. However, the periodic color changes in the skin captured by remote cameras are quite subtle, and variables like natural scene illumination and head movements can significantly impair the reliability of physiological indicator measurements." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Li et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 481, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 481, + 654 + ], + "type": "text", + "content": "1 This dataset is 1/20 equally spaced down-sampled from the full VIPL dataset [42], and contains 8349 samples in total." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 654, + 326, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 326, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 326, + 665 + ], + "type": "text", + "content": "2 This visualization is generated following [26]." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 436, + 235 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 436, + 235 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 436, + 235 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 436, + 235 + ], + "type": "image", + "image_path": "10619055dd8175ecb6a1b49574c013f82ad76c4aaa2dcdcf5dd754293df96f9c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 244, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 131, + 244, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 244, + 482, + 312 + ], + "type": "text", + "content": "Fig. 2: Visualization of Domain Adaptation (DA) and Test-Time Adaptation (TTA) methodologies. DA utilizes batch learning with labeled target data, while TTA dynamically refines the model during inference without relying on target data labels or distribution. Both DA and TTA obviate the requirement for source data. Note that domain generalization (DG) is excluded as it does not focus on specific target domain adaptation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 411 + ], + "type": "text", + "content": "Conventional rPPG methods [6,25,37,46,60,64] perform well only in constrained environments. Recent deep learning (DL) methods [2,43,49,53,68] have demonstrate promising results, however, the most of training data are collected in controlled laboratory environments with limited variations in illumination, camera parameters, etc. Therefore, these methods still struggle to generalize to unseen domains, necessitating model adaptation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 412, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 532 + ], + "type": "text", + "content": "For privacy considerations, the adaptation of rPPG models in unseen scenarios is required to be performed without accessing the source data, and target data annotations or distributions (i.e., large batches of target data samples). Addressing this, we firstly implement Test-Time Adaptation (TTA) in the field of rPPG. TTA aims to fine-tune a pre-trained source model during inference time, as illustrated in Fig. 2, without accessing the distribution and labeling of both source data and target data and naturally eliminate the need for intensive re-training [61]. This innovation allows for the automatic adaptation of pretrained rPPG models to unseen domains during inference, without the necessity for labelling or batches of target data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "However, applying TTA directly into rPPG faces two key limitations. 1) Predominantly designed for classification tasks, most TTA algorithms leverage entropy in normalization layers or pseudo labels. These strategies are not inherently suited for regression tasks like rPPG, leading to a lack of appropriate supervision. 2) In real-world scenarios, the pre-trained rPPG model is tuned during inference the user's face video stream, on an instance-by-instance context. This variability and instability introduced by biased instances pose challenges to effectively filtering out domain-irrelevant noise and simultaneously preserving crucial learned information as the model acquiring new knowledge. This nuanced balance is critical in preventing slow or ineffective adaptation under single-instance learning condition, ensuring an efficient and consistent adaptation process." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "Bi-TTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": "To navigate these challenges, we proposed a novel Bidirectional Test-Time Adapter (Bi-TTA) framework, grounded on two expert knowledge-based self-supervised priors. Firstly, two regularizations are proposed for offering effective supervision especially in rPPG tasks, addressing the aforementioned ill-suited problem of applying existing TTA algorithms into regression tasks. These two priors focus on temporal and spatial consistency respectively, which is derived from the characteristics of physiological signals. Then, two strategies are introduced for stable and effective fine-tuning process. The prospective adaptation (PA) strategy is designed for filtering out the irrelevant information of each encountered instance, aiming that only target-domain-essential feature is truly contributing to model adaptation. The retrospective stabilization (RS) strategy backtracks and consolidates essential model parameters dynamically, leveraging previous update gradients during the self-supervised tuning process, ensuring that the learned adaptation ability is preserved intact. We evaluate our method across various datasets, benchmarking against existing prevailing TTA algorithms, demonstrating the superiority of our Bi-TTA. Overall, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 327, + 480, + 445 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 327, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 350 + ], + "type": "text", + "content": "- We introduce the TTA protocol to rPPG tasks, paving the way for a new paradigm in self-supervised adapting of rPPG models during inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 351, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 480, + 373 + ], + "type": "text", + "content": "- We introduce two novel self-supervised priors to regularize temporal and spatial consistency, providing appropriate supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 375, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 480, + 409 + ], + "type": "text", + "content": "- We propose a novel Bidirectional Test-Time Adapter (Bi-TTA) framework grounded on the introduced priors, which promotes effective and efficient adaptation, with enhanced stability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 411, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 445 + ], + "type": "text", + "content": "- We established a large-scale benchmark for rPPG tasks under TTA protocol. Extensive experiments demonstrate significantly improved adaptation capacity of our approach." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 464, + 243, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 243, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 243, + 478 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 490, + 340, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 340, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 340, + 502 + ], + "type": "text", + "content": "2.1 Remote Physiological Measurement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "Remote physiological measurement aims to quantify HR and HRV from BVP signals by interpreting skin chrominance fluctuations. Traditional techniques typically involve color space transformations or signal decomposition to obtain the BVP signal with high signal-to-noise ratio (SNR) [6, 7, 46, 65]. With powerful modeling capacity of deep learning (DL), there's been a significant shift towards employing DL models for remote physiological estimation, yielding impressive results. [4, 16, 41, 43, 52, 66, 69]. Many of these methods extract BVP signals from face videos typically by aggregating spatial and temporal information [44, 52, 69, 70]. Besides, many works utilize hand-designed 2D feature maps extracted from face videos to alleviate computational demands [16, 33, 43, 48, 51]. Recently, researchers have turned to exploring how to leverage task-specific knowledge and how to improve the generalization performance of the model into unseen scenarios." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "H. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "Task-specific priors. rPPG is distinguished by its unique biological and physical mechanisms, which can be exploited to enhance the physiological signal extraction ability of the model [17,31,58,62]. These task-specific priors can be summed up as 1) spatial consistency: BVP signals extracted from different face areas should exhibit similar; 2) temporal consistency: BVP signals typically exhibit smooth and gradual changes over short time period; 3) heart rate range: The heart rate typically ranges between 40 and 250 beats per minute (bpm). Different from existing works, we provide a more granular solution leveraging the multi-scale latent features, rather than purely focusing on the prediction output of BVP signals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 483, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 483, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 483, + 380 + ], + "type": "text", + "content": "The generalization of the model. With the domain shifts (e.g., lighting, head motion, environment, etc.), most methods tend to have significant performance degradation. To improve the generalization performance on unseen domains, NEST-rPPG established a large-scale DG evaluation protocol [34]. These DG methods, requiring the access of all source data, do not focus on the adaptation of a specific target domain, leading to sub-optimal performance [56]. Besides, some work utilizes the labels and distribution of the target domain data to improve the model's adaptation performance [10,24,30], which is impractical due to accessing the unseen domain in advance. Differently from the above rPPG methods, we address this domain gap by using the TTA protocol, facilitating the adaptation to specific target domains during inference, based solely on unlabeled instances from the target domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 402, + 272, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 272, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 272, + 415 + ], + "type": "text", + "content": "2.2 Test-Time Adaptation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "Test-time Adaptation (TTA), an emerging protocol, aims to adjust a pre-trained source model during inference by learning from the unlabeled target data, without accessing the distributions and labels of both source and target data, thereby naturally eliminate the need of intensive re-training [27, 61, 63]. Current TTA methods typically require aggregating a large number of samples from the target data to form a batch, and perform adaptation through 1) normalization calibration [1, 21, 22, 39, 74], 2) pseudo-labeling [8, 14], 3) input adaptation [15, 73], 4) dynamic inference that learns the network weights from a set of model parameters [72], 5) meta-learning [12], etc. These approaches can be considered as different forms of source-free domain adaptation (SFDA) [28] and are infeasible for regression tasks like rPPG where data arrives continuously and in a sequential manner. Some instance-level TTA approaches may either rely on specialized network architectures [1, 22, 61, 74] or incorporate additional auxiliary tasks applied during both pre-training and adaptation [11, 57]. By imposing specific non-general requirement to the pre-training stage, these methods narrow application scope of TTA. Moreover, some works even introduce a memory mechanism that turns instances into batches [21], lacking an in-time adaptation to each inference with increased computational cost. In contrast with above mentioned TTA methods, our Bi-TTA adapts the model during inference under each encountered instance from target domain without specific pre-training" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "Bi-TTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "or network architecture requirements, offering a flexible, effective, and efficient solution for rPPG model adaptation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 202, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 202, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 202, + 171 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 208, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 208, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 208, + 196 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 480, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 480, + 277 + ], + "type": "text", + "content": "In this paper, we propose a novel expert knowledge-based Bi-directional Test-Time Adapter (Bi-TTA) to address the performance degradation in unseen target domains. In the following sections, we will firstly formulate the problem in Sec. 3.2. Secondly, Sec. 3.3 introduces the proposed self-supervised priors based on expert knowledge in the realm of rPPG. The bidirectional (prospective and retrospective) adaptation strategy of our Bi-TTA will be specified in Sec. 3.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 295, + 228, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 228, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 228, + 306 + ], + "type": "text", + "content": "3.2 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "content": "The source dataset " + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{S}} \\triangleq \\cup_{i=1}^{n_{\\mathrm{S}}} \\left\\{(\\pmb{x}_i, y_i)\\right\\}" + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "content": " is assumed to be drawn independently and identically distributed (i.i.d.) from distribution " + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{S}}" + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "content": ". Similarly, the target dataset " + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "content": " is assumed to be sampled from a different distribution " + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "content": ", reflecting the domain shift. With a model parameterized by " + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "inline_equation", + "content": "\\pmb{w} \\in \\mathcal{W}" + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "content": " and considering the supervised loss function " + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "inline_equation", + "content": "L: \\mathcal{W} \\times \\mathcal{X} \\times \\mathcal{Y} \\rightarrow \\mathbb{R}_{+}" + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "content": " for evaluation, we define the observation loss over dataset " + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "content": " as an approximation of the population loss over distribution " + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 315, + 479, + 398 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 192, + 409, + 480, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 409, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 192, + 409, + 480, + 438 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} (\\boldsymbol {w}) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} L \\left(\\boldsymbol {w}, \\boldsymbol {x} _ {i}, y _ {i}\\right) \\approx \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim X} [ L (\\boldsymbol {w}, \\boldsymbol {x}, y) ]. \\tag {1}", + "image_path": "67946463059af5abffaded657aa821f77186452a4d9cbb2082713b17b6ae772f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "text", + "content": "The objective of TTA, including our Bi-TTA, is to minimize the target observation loss " + }, + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{D_{\\mathrm{T}}}(\\boldsymbol{w}_t)" + }, + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "text", + "content": " after fine-tuning the initial model weights from " + }, + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_0" + }, + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_t" + }, + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 447, + 479, + 483 + ], + "type": "text", + "content": " observed samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 501, + 321, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 321, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 321, + 513 + ], + "type": "text", + "content": "3.3 Expert Knowledge-based Priors" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "type": "text", + "content": "In the context of Bi-TTA, the unavailability of target data labels compels the use of self-supervised loss functions for providing the tuning gradient. Inspired by [17, 31,58,62], our Bi-TTA capitalizes on the characteristics of rPPG signals to craft these self-supervised loss functions, concentrating on the principles of spatial and temporal consistency. The introduced loss functions are termed as " + }, + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{s}}(\\pmb{w}_{t-1}, \\pmb{x}_t)" + }, + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{t}}(\\pmb{w}_{t-1}, \\pmb{x}_t)" + }, + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "type": "text", + "content": ", focusing on spatial and temporal consistency respectively. These functions play a foundational role in the self-supervised tuning process, providing appropriate supervision based on the expert knowledge, despite the lack of direct label guidance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "Spatial-temporal feature map (STMap). In line with [16, 33, 43, 48, 51], 2D STMap is firstly constructed to encapsulate the face video, as specified in Fig. 3a. This STMap integrates both temporal and spatial dimensions: the temporal" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "H. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 233, + 289 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 233, + 289 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 233, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 233, + 289 + ], + "type": "image", + "image_path": "40947f9f907e1baf30091467befe69c31d0af0ef43d8582f3daecea6d250c58d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 290, + 229, + 299 + ], + "lines": [ + { + "bbox": [ + 140, + 290, + 229, + 299 + ], + "spans": [ + { + "bbox": [ + 140, + 290, + 229, + 299 + ], + "type": "text", + "content": "(a) STMap Construction" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 115, + 338, + 289 + ], + "blocks": [ + { + "bbox": [ + 242, + 115, + 338, + 289 + ], + "lines": [ + { + "bbox": [ + 242, + 115, + 338, + 289 + ], + "spans": [ + { + "bbox": [ + 242, + 115, + 338, + 289 + ], + "type": "image", + "image_path": "12b439476b69a08ce4ba6ac3915b0e01f74268ad969fe04902220cb5073dd7ee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 290, + 342, + 300 + ], + "lines": [ + { + "bbox": [ + 241, + 290, + 342, + 300 + ], + "spans": [ + { + "bbox": [ + 241, + 290, + 342, + 300 + ], + "type": "text", + "content": "(b) Temporal Consistency Loss" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 347, + 114, + 481, + 289 + ], + "blocks": [ + { + "bbox": [ + 347, + 114, + 481, + 289 + ], + "lines": [ + { + "bbox": [ + 347, + 114, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 347, + 114, + 481, + 289 + ], + "type": "image", + "image_path": "04e3232af5c821d9716d0b3cb0ac88bcd077e1b6aae56f3663d33372d53bad20.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 290, + 462, + 300 + ], + "lines": [ + { + "bbox": [ + 361, + 290, + 462, + 300 + ], + "spans": [ + { + "bbox": [ + 361, + 290, + 462, + 300 + ], + "type": "text", + "content": "(c) Spatial Consistency Loss" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 316, + 482, + 394 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 394 + ], + "type": "text", + "content": "Fig. 3: Illustration of STMap construction and the implementation of our proposed expert knowledge-based priors. (a) The process of generating STMap, encompassing face alignment" + }, + { + "bbox": [ + 130, + 316, + 482, + 394 + ], + "type": "inline_equation", + "content": "^5" + }, + { + "bbox": [ + 130, + 316, + 482, + 394 + ], + "type": "text", + "content": " and cropping, local signal extraction, and the subsequent integration. (b) The calculation process of TCL, aimed at minimizing significant prediction discrepancies between original and temporally shifted HR predictions. (c) The calculation process the SCL, focused on penalizing pronounced disparities across different facial regions. Note that the boxes colored in pink represent the loss outcomes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "text", + "content": "dimension represents the time sequence and the spatial dimension corresponds to the patches of the cropped face video frames. The current sample " + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t\\in \\mathbb{R}^{W\\times H}" + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "text", + "content": " is referred to as a sliding window on the STMap, where " + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "text", + "content": " represents the sequential length of the sliding window, and " + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "text", + "content": " indicates the size of spatial dimension in the STMap." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Temporal consistency loss (TCL). BVP signals exhibit smooth and gradual fluctuations over short time spans. This fundamental attribute, known as the temporal consistency, is essential for accurate physiological signal analysis and the key to guarantee the predictability of physiological signal changes. To preserving this critical characteristic and derive a tuning gradient accordingly, the TCL is meticulously designed to enforce the model's prediction deviations of BVP signals when subjected to minor time perturbations under an acceptable tolerance. To compute the TCL, a random minor temporal perturbation " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " is introduced by shifting the original sample " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " along the STMap's temporal dimension. The shifted sample " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t - \\delta_{\\mathrm{T}}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", in parallel with the origin sample, are then fed into the rPPG network " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{t - 1}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " for predicting two sets of BVP signals. Then, for penalizing the model from large discrepancies in the HR estimation results between the original and shifted samples, " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " regularization is applied on the original and shifted HR prediction results, incorporating with a tolerance" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "Bi-TTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "constant " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ". The specific calculation process of TCL is formulated in Eq. 2, as illustrated in Fig. 3b." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 149, + 423, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 149, + 423, + 164 + ], + "spans": [ + { + "bbox": [ + 216, + 149, + 423, + 164 + ], + "type": "interline_equation", + "content": "\\Delta_ {t, t - \\delta_ {\\mathrm {T}}} ^ {\\mathrm {H R}} = \\left\\| \\boldsymbol {w} _ {t - 1} ^ {\\mathrm {H R}} (\\boldsymbol {x} _ {t}) - \\boldsymbol {w} _ {t - 1} ^ {\\mathrm {H R}} (\\boldsymbol {x} _ {t - \\delta_ {\\mathrm {T}}}) \\right\\| _ {1} \\in \\mathbb {R} ^ {W \\times 1}", + "image_path": "54f7d834b2f66a21f1dc39c76e04ba71c98c9db66d0d518f228ac5e98bcf746f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 191, + 167, + 480, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 167, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 191, + 167, + 480, + 199 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {t}} \\left(\\boldsymbol {w} _ {t - 1}, \\boldsymbol {x} _ {t}\\right) = \\sum_ {i} ^ {W} \\max \\left(0, \\Delta_ {t, t - \\delta_ {\\mathrm {T}}} ^ {\\mathrm {H R} (i)} - \\xi_ {\\mathrm {T}}\\right). \\tag {2}", + "image_path": "cb472948825c67cc18a9e469b3d81efd70343d7eefe93a3c780342b0ad7ed7de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 209, + 485, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 485, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 485, + 352 + ], + "type": "text", + "content": "Spatial consistency loss (SCL). Heart rate variations induce chromatic changes across the face, which are periodic and consistent across different facial regions. This phenomenon is referred to as spatial consistency, alongside temporal consistency, are crucial attributes in the rPPG field. Inspired by this phenomenon, SCL is crafted to promote the spatial consistency, obtaining correct adaptation gradient by encouraging the BVP signals from different facial areas to exhibit similarities. In particular, the SCL is calculated covering a series of latent feature maps, corresponding to various depths, (i.e., the number of residual blocks) in the adopted ResNet-18 [18] architecture. Given the current sample " + }, + { + "bbox": [ + 130, + 209, + 485, + 352 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t\\in \\mathbb{R}^{W\\times H}" + }, + { + "bbox": [ + 130, + 209, + 485, + 352 + ], + "type": "text", + "content": ", four latent feature maps " + }, + { + "bbox": [ + 130, + 209, + 485, + 352 + ], + "type": "inline_equation", + "content": "\\left\\{F_{i}\\in \\mathbb{R}^{W_{i}\\times H_{i}}\\mid i\\in [1,4]\\right\\}" + }, + { + "bbox": [ + 130, + 209, + 485, + 352 + ], + "type": "text", + "content": " could be extracted from the rPPG network " + }, + { + "bbox": [ + 130, + 209, + 485, + 352 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{t - 1}" + }, + { + "bbox": [ + 130, + 209, + 485, + 352 + ], + "type": "text", + "content": ". As illustrated in Fig. 3c, the calculation process of SCL is specify as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 354, + 480, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 211, + 354, + 480, + 386 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {s}} = \\sum_ {i} ^ {n} \\Delta F _ {i} = \\sum_ {i} ^ {n} \\sum_ {j} ^ {W _ {i} - \\delta_ {\\mathrm {S}}} | | F _ {i, j} - F _ {i, j + \\delta_ {\\mathrm {S}}} | | _ {1}, \\tag {3}", + "image_path": "d5956ca5b5090a23991927586a172d6e23cdbe8acdb6ef4065be631bd93c76d5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 392, + 479, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 479, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 479, + 451 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 392, + 479, + 451 + ], + "type": "inline_equation", + "content": "F_{i,j} = F_i[:,j] \\in \\mathbb{R}^{W_i}" + }, + { + "bbox": [ + 130, + 392, + 479, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 392, + 479, + 451 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 130, + 392, + 479, + 451 + ], + "type": "text", + "content": " represents the number of residual blocks, and " + }, + { + "bbox": [ + 130, + 392, + 479, + 451 + ], + "type": "inline_equation", + "content": "\\delta_S" + }, + { + "bbox": [ + 130, + 392, + 479, + 451 + ], + "type": "text", + "content": " is the incremental spatial shift. Leveraging the fine-grained multi-level strategy, SCL effectively nurtures a comprehensive regularization of spatial features, ensuring that appropriate supervision is available for the adaptation during inference without annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 480, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 480, + 489 + ], + "type": "text", + "content": "After combining " + }, + { + "bbox": [ + 130, + 453, + 480, + 489 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{t}}" + }, + { + "bbox": [ + 130, + 453, + 480, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 453, + 480, + 489 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 453, + 480, + 489 + ], + "type": "text", + "content": ", without direct target label reliance, the self-supervised loss is formulated by weighting and summing the spatial and temporal consistency priors as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 268, + 499, + 480, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 499, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 268, + 499, + 480, + 512 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {p}} = \\lambda_ {\\mathrm {s}} L _ {\\mathrm {s}} + \\lambda_ {\\mathrm {t}} L _ {\\mathrm {t}}. \\tag {4}", + "image_path": "f4902391a7262c2d9b714c5b7e55864a57cc5a13882ecf9af585370760244e0a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 529, + 340, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 340, + 542 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 340, + 542 + ], + "type": "text", + "content": "3.4 Bidirectional Test-Time Adaptation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 550, + 480, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 480, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 480, + 647 + ], + "type": "text", + "content": "As introduced above, based on the proposed priors, the Bi-TTA framework aims to ensure effective and robust model adaptation under the variability and instability of single-instance learning conditions, minimizing the impact of domain-irrelevant noise. Simultaneously adapting the model with new target samples, our Bi-TTA also focuses on preserving the acquired adaptation ability dynamically, preventing the performance degradation caused by overfitting or catastrophic forgetting, achieving a stable maintenance of the adaptation performance in the target domain." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "H. Li et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 326, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 326, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 326, + 666 + ], + "type": "text", + "content": "3 https://github.com/adrianb/face-alignment." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 220, + 114, + 394, + 316 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 394, + 316 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 394, + 316 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 394, + 316 + ], + "type": "image", + "image_path": "5094995fdae497c6b6c8b35239887c626464363180a8057bba77fdded0a44d85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "text", + "content": "Fig. 4: Illustration of the proposed Bidirectional Test-Time Adapter (Bi-TTA). Black arrows " + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "text", + "content": " indicate the adaptation process purely with the proposed two priors, i.e., TCL and SCL. Orange arrows " + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "text", + "content": " denote that the PA module adjusts model parameter using the gradient of representative neighborhood with a radius " + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "text", + "content": ". The green ones " + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "text", + "content": " show that the RS is activated when there is an oscillation, which is a sign of performance degradation, for maintaining the essential learned adaptation ability with former tuning gradients. The gradient and learning rate are formulated as " + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 325, + 482, + 413 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 440, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 522 + ], + "type": "text", + "content": "Prospective adaptation (PA). During the process of model adaptation, especially under single-instance learning conditions, the noisy irrelevant information in biased target samples can easily interfere the model's tuning process, leading to a declined performance. Therefore, inspired by [13,38], rather than seeking out model parameter " + }, + { + "bbox": [ + 130, + 440, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\pmb{w}_t" + }, + { + "bbox": [ + 130, + 440, + 482, + 522 + ], + "type": "text", + "content": " that simply have low adaptation loss " + }, + { + "bbox": [ + 130, + 440, + 482, + 522 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{p}}(\\pmb{w}_t)" + }, + { + "bbox": [ + 130, + 440, + 482, + 522 + ], + "type": "text", + "content": ", we seek out model parameter whose entire neighborhoods to have uniformly low adaptation loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 524, + 481, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 524, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 246, + 524, + 481, + 544 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {p}} ^ {\\prime} (\\boldsymbol {w}) = \\max _ {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\rho} L _ {\\mathrm {p}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}), \\tag {5}", + "image_path": "60f0d7531bee3b762777550cf0df468eaad7515eaab5a03208f489524d7b3dec.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 553, + 491, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 491, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 491, + 597 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 553, + 491, + 597 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 553, + 491, + 597 + ], + "type": "text", + "content": " is the disturbance range. This strategy, known as sharpness-aware minimization, has been proved to be effective in various areas [3, 19, 45, 67]. For computing efficiency and differentiability, we approximate the " + }, + { + "bbox": [ + 130, + 553, + 491, + 597 + ], + "type": "inline_equation", + "content": "\\epsilon^{*} = \\arg \\max_{\\| \\epsilon \\|_{2}\\leqslant \\rho}L_{\\mathrm{p}}(\\boldsymbol {w} +" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 597, + 321, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 321, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 321, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon)" + }, + { + "bbox": [ + 130, + 597, + 321, + 609 + ], + "type": "text", + "content": ". Following 1-order Taylor series, we get:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 205, + 620, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 620, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 205, + 620, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {\\epsilon} ^ {*} \\approx \\underset {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\varrho} {\\arg \\max } L _ {\\mathrm {p}} (\\boldsymbol {w}) + \\boldsymbol {\\epsilon} ^ {\\mathrm {T}} \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w}) \\\\ = \\underset {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\rho} {\\arg \\max \\boldsymbol {\\epsilon}} ^ {\\mathrm {T}} \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w}) = \\underset {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leqslant \\rho} {\\arg \\max \\boldsymbol {\\epsilon}} ^ {\\mathrm {T}} \\boldsymbol {g} = \\hat {\\boldsymbol {\\epsilon}}, \\tag {6} \\\\ \\end{array}", + "image_path": "0ab34cb09e3b71a132b405372fee8b9f9c82b7d6a5502bf8b3f4b2e6c918c7d6.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "Bi-TTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": ", w.r.t. " + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": " around 0. Then employing the Cauchy-Schwarz inequality from Dual Norm theory: " + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\epsilon^{\\mathrm{T}}\\pmb {g}\\leqslant \\| \\pmb {\\epsilon}\\|_{2}\\| \\pmb {g}\\|_{2}" + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": " should be col-linear with " + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": " to maximize " + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}^{\\mathrm{T}}\\pmb{g}" + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": ". Considering " + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\| \\pmb {\\epsilon}\\| _2\\leqslant \\rho" + }, + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": ", we can get:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 163, + 481, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 163, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 206, + 163, + 481, + 190 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} = \\rho \\frac {\\operatorname {s i g n} (\\boldsymbol {g}) \\boldsymbol {g}}{\\| \\boldsymbol {g} \\| _ {2}} = \\rho \\frac {\\operatorname {s i g n} \\left(\\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w})\\right) \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w})}{\\| \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} (\\boldsymbol {w}) \\| _ {2}}. \\tag {7}", + "image_path": "510bd73be38355102cc9c67ed571b9ca5870d369d089270fe1a0314f0c2a5216.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 198, + 438, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 198, + 438, + 210 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 438, + 210 + ], + "type": "text", + "content": "Consequently, we can derive the gradient approximation under Eq. 5:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 220, + 480, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 220, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 198, + 220, + 480, + 264 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}} = \\nabla_ {\\boldsymbol {w}} L _ {\\mathrm {p}} ^ {\\prime} (\\boldsymbol {w}) \\approx \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\hat {\\boldsymbol {\\epsilon}}) \\\\ = \\left. \\frac {d (\\boldsymbol {w} + \\hat {\\epsilon})}{d \\boldsymbol {w}} \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\right| _ {\\boldsymbol {w} + \\hat {\\epsilon}} \\approx \\left. \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\right| _ {\\boldsymbol {w} + \\hat {\\epsilon}} \\tag {8} \\\\ \\end{array}", + "image_path": "54162bc4fa674cd3ffddc3e5560f59768f6475fb49020b8b6f781a6bd156373b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "content": "As illustrated in Fig. 4, the gradient " + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t^{\\mathrm{PA}}" + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "content": " is adopted to replace the original gradient merely derived from two self-supervised priors. This strategy boosts the model's robustness to the undesired variability from instance-level conditions, safeguarding the tuning process against the influence from target-domain irrelevant information. Consequently, it secures both the efficacy and stability of the adaptation process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": "Retrospective stabilization (RS). The PA module, by considering the model's entire neighborhood as the optimization objective, effectively enhances the model's robustness against irrelevant disturbances in target samples. However, after the model has processed a certain number of samples, it may still capture the detrimental noise feature instead of latent, universally domain-essential patterns, leading to performance degradation, as evidenced in Fig. 6. Therefore, to ensure reliable physiological signal predictions throughout the entire adaptation process, we propose the retrospective stabilization (RS) module. RS initially introduces the concept of trend gradient, termed as " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{g}^{*}" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": ". Starting with " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{g}_0^* = \\mathbf{0}" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": ", the trend gradient after processing " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": " samples is updated via: " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t^* = \\frac{\\pmb{g}_{t - 1}^* + (1 / ||L_{\\mathrm{p}}||_1)\\pmb{g}_t^{\\mathrm{PA}}}{1 + 1 / ||L_{\\mathrm{p}}||_1}" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t^{\\mathrm{PA}}" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": ", its orthogonal projection on " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t^*" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": " is computed, formulated as " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t^\\perp = \\frac{\\pmb{g}_t^{\\mathrm{PA}}\\cdot\\pmb{g}_t^*}{\\|\\pmb{g}_t^*\\|^2}\\pmb{g}_t^*" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": ". As illustrated in Fig. 4, RS is activated if this projection aligns in the opposite of the trend gradient, which signifies an gradient oscillation, indicative of a potential fall of the generalization performance. Specifically, upon detection of oscillation, the model's optimization gradient will be the product of " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t^\\perp" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": " and the backtracking degree " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": ". The orthogonal part of " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t^*" + }, + { + "bbox": [ + 130, + 347, + 483, + 552 + ], + "type": "text", + "content": " is ignored." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 198, + 561, + 480, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 561, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 198, + 561, + 480, + 594 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {t} ^ {\\mathrm {R S}} = \\left\\{ \\begin{array}{l} k \\cdot \\boldsymbol {g} _ {t} ^ {\\perp}, \\cos \\left(\\boldsymbol {g} _ {t} ^ {*}, \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}}\\right) < 0 \\\\ \\left(1 - \\lambda_ {t} ^ {\\mathrm {R S}}\\right) \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}} + \\lambda_ {t} ^ {\\mathrm {R S}} \\boldsymbol {g} _ {t} ^ {*}, \\cos \\left(\\boldsymbol {g} _ {t} ^ {*}, \\boldsymbol {g} _ {t} ^ {\\mathrm {P A}}\\right) \\geqslant 0 \\end{array} \\right. \\tag {9}", + "image_path": "7afcfa4c97f5ba296c35c5854ac342a97e1e319932a9de3598a301f025c73536.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_t^{\\mathrm{RS}} = -1 + 2 \\cdot \\operatorname{Sigmoid}(-\\frac{4t}{\\Omega}) \\in \\mathbb{R}_+^{(0,1)}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": ", is the anneal scaling factor, " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " denotes the amount of target samples needed to ensure the fidelity of trend gradient, i.e., " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_t^{\\mathrm{RS}}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " can be approximated as 1.0 when " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "t \\geqslant \\Omega" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": ". The obtained " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t^{\\mathrm{RS}}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " represents the gradient that ultimately contributes to fine-tune the model for target domain adaptation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 178, + 493, + 375 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: HR estimation results. Typically, the model selected for adaptation on each dataset is initially pre-trained on the remaining four datasets utilizing the NEST [34] framework. Ours w/o P.R. denotes that only the expert knowledge-based priors is adopted, without the bidirectional adaptation strategy. The best results are highlighted in bold, and the second-best results are underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 178, + 493, + 375 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 493, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 493, + 375 + ], + "type": "table", + "html": "
MethodVIPL [42]V4V [49]PURE [53]UBFC [2]BUAA [68]
M↓R↓r↑M↓R↓r↑M↓R↓r↑M↓R↓r↑M↓R↓r↑
Methods based on DG protocol
Coral [54]8.6811.910.5310.3214.420.327.5910.870.725.898.040.763.645.740.8
VReX [23]8.3711.620.549.8214.160.377.2410.140.785.597.680.813.275.010.86
NCDG [59]8.4711.810.5210.1414.460.347.3210.350.775.317.560.823.125.160.85
NEST [34]7.8611.150.589.2713.790.416.719.590.814.676.790.862.884.690.89
Methods based on TTA protocol
Tent [61]8.0911.320.559.9811.730.476.869.680.844.579.740.872.373.590.92
SAR [40]8.1311.450.539.6811.460.476.339.30.864.569.640.872.072.930.95
TTT++ [32]7.7311.080.569.5111.310.546.239.140.884.366.790.872.052.820.92
EATA [39]7.6911.030.579.3611.120.516.139.230.864.256.860.881.892.640.94
SHOT [29]7.7511.070.569.5511.320.515.818.860.874.056.450.871.872.470.96
AdaODM [72]7.8111.210.559.110.930.555.518.290.884.016.590.871.912.60.95
ConPhys4 [58]7.4310.70.619.310.960.526.098.830.883.926.610.8861.752.190.97
Ours w/o P.R.7.3110.640.628.9710.820.555.568.670.883.646.630.8851.682.040.99
Ours w/o RS7.1510.830.638.9310.790.555.248.320.8893.596.610.8851.551.940.99
Ours w/o PA7.2410.780.629.0410.970.545.398.450.883.546.550.8881.511.980.99
Ours7.0910.720.639.111.020.545.028.110.8903.536.240.8911.491.970.99
", + "image_path": "339466737b800041674716e618c0e79fdd4605fc70b0cd9f638331671f04282e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 397, + 230, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 230, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 230, + 411 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 423, + 205, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 205, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 205, + 434 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 443, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 481, + 479 + ], + "type": "text", + "content": "We elaborately select five rPPG datasets, encompassing diverse races, levels of head motion, camera conditions, and ambient lighting, to establish a comprehensive large-scale benchmark of rPPG task under TTA protocol." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 479, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 552 + ], + "type": "text", + "content": "VIPL-HR [42] have nine scenarios, three RGB cameras, different illumination conditions, and different levels of movement. It is worth mentioning that the BVP signal and video of the dataset do not match in the time dimension. We used the output of HR head as the evaluation result. Besides, we normalize STMap to 30 fps by cubic spline interpolation to solve the problem of the unstable frame rate of video [33,51,70]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 552, + 481, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 481, + 610 + ], + "type": "text", + "content": "V4V [49] is designed to collect data with the drastic changes of physiological indicators by simulating ten tasks such as a funny joke, 911 emergency call, and odor experience. It is worth mentioning that There are only heart rate labels in the dataset and no BVP signal labels. We used the output of HR head as the evaluation result." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 611, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 635 + ], + "type": "text", + "content": "BUAA [68] is proposed to evaluate the performance of the algorithm against various illumination. We only use data with illumination greater than or equal" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "Bi-TTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "4 The proposed self-supervised loss in ConPhys [58] is applied for a comparative analysis of our proposed priors." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 297, + 167 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 297, + 167 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 297, + 167 + ], + "type": "image", + "image_path": "f73fb16e85592b530001867f3bc1db5fdf1c5346ea3582618f5438b7346c4740.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 167, + 225, + 175 + ], + "lines": [ + { + "bbox": [ + 209, + 167, + 225, + 175 + ], + "spans": [ + { + "bbox": [ + 209, + 167, + 225, + 175 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 209, + 167, + 225, + 175 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 175, + 298, + 227 + ], + "blocks": [ + { + "bbox": [ + 142, + 175, + 298, + 227 + ], + "lines": [ + { + "bbox": [ + 142, + 175, + 298, + 227 + ], + "spans": [ + { + "bbox": [ + 142, + 175, + 298, + 227 + ], + "type": "image", + "image_path": "ca0988a0006d3a1c41b677483a66cd3bc5bbd943feb8350b2bbdf1fd44fd65ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 228, + 227, + 236 + ], + "lines": [ + { + "bbox": [ + 209, + 228, + 227, + 236 + ], + "spans": [ + { + "bbox": [ + 209, + 228, + 227, + 236 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 209, + 228, + 227, + 236 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathrm{T}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 114, + 383, + 167 + ], + "blocks": [ + { + "bbox": [ + 301, + 114, + 383, + 167 + ], + "lines": [ + { + "bbox": [ + 301, + 114, + 383, + 167 + ], + "spans": [ + { + "bbox": [ + 301, + 114, + 383, + 167 + ], + "type": "image", + "image_path": "09e95419af086bb65a14654621222dccf7bd02728e2028509c8760437b0761dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 167, + 353, + 175 + ], + "lines": [ + { + "bbox": [ + 336, + 167, + 353, + 175 + ], + "spans": [ + { + "bbox": [ + 336, + 167, + 353, + 175 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 336, + 167, + 353, + 175 + ], + "type": "inline_equation", + "content": "\\lambda_{t}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 301, + 176, + 383, + 228 + ], + "blocks": [ + { + "bbox": [ + 301, + 176, + 383, + 228 + ], + "lines": [ + { + "bbox": [ + 301, + 176, + 383, + 228 + ], + "spans": [ + { + "bbox": [ + 301, + 176, + 383, + 228 + ], + "type": "image", + "image_path": "049ce48e4b3afd1cfaef5076b0f953022bbf08345fd39b0e6c2fe82cdd63ee80.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 228, + 351, + 236 + ], + "lines": [ + { + "bbox": [ + 336, + 228, + 351, + 236 + ], + "spans": [ + { + "bbox": [ + 336, + 228, + 351, + 236 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 336, + 228, + 351, + 236 + ], + "type": "inline_equation", + "content": "\\rho" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 246, + 481, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 246, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 481, + 289 + ], + "type": "text", + "content": "Fig. 5: Ablation experiments of hyper-parameters on down-sampled VIPL dataset5. Adheres to the default parameter configuration detailed in Sec.4.2, only one hyper-parameter is varied during each set of experiments. The pink arrows point to the lowest MAE result and the blue arrows mark the second if indiscernible." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 114, + 472, + 167 + ], + "blocks": [ + { + "bbox": [ + 389, + 114, + 472, + 167 + ], + "lines": [ + { + "bbox": [ + 389, + 114, + 472, + 167 + ], + "spans": [ + { + "bbox": [ + 389, + 114, + 472, + 167 + ], + "type": "image", + "image_path": "ded933a3ada4d1dc625115cf256cc2ce9dd46d2259274fc81fc9e9edeb1f5e51.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 167, + 447, + 175 + ], + "lines": [ + { + "bbox": [ + 429, + 167, + 447, + 175 + ], + "spans": [ + { + "bbox": [ + 429, + 167, + 447, + 175 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 429, + 167, + 447, + 175 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 389, + 176, + 472, + 228 + ], + "blocks": [ + { + "bbox": [ + 389, + 176, + 472, + 228 + ], + "lines": [ + { + "bbox": [ + 389, + 176, + 472, + 228 + ], + "spans": [ + { + "bbox": [ + 389, + 176, + 472, + 228 + ], + "type": "image", + "image_path": "226cdb6d02d5db366b7fd637888e0dc80225c8aab3c46933125e5310a56c0fe3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 228, + 445, + 236 + ], + "lines": [ + { + "bbox": [ + 429, + 228, + 445, + 236 + ], + "spans": [ + { + "bbox": [ + 429, + 228, + 445, + 236 + ], + "type": "text", + "content": "(f) " + }, + { + "bbox": [ + 429, + 228, + 445, + 236 + ], + "type": "inline_equation", + "content": "\\Omega" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "content": "to 10 lux because underexposed images require special algorithms that are not considered in this article. We used the output of BVP head as the evaluation for BUAA, PURE, and UBFC-rPPG datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 353, + 479, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 479, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 479, + 400 + ], + "type": "text", + "content": "PURE [53] contains 60 RGB videos from 10 subjects with six different activities, specifically, sitting still, talking, and four rotating and moving head variations. The BVP signals are down-sampled from 60 to 30 fps with cubic spline interpolation to align with the videos." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 402, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 479, + 426 + ], + "type": "text", + "content": "UBFC-rPPG [2] containing 42 face videos with sunlight and indoor illumination, the ground-truth BVP signals, and HR values were collected by CMS50E." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 450, + 279, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 279, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 279, + 462 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": "We implement our proposed priors and Bi-TTA using the PyTorch framework. The STMap sample is denoted as " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{256 \\times 25 \\times 3}" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": ", where 256 is the temporal length, i.e., the number of video frames, of the sliding window, and 25 is the spatial dimension of the STMap. It is firstly resized to " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\pmb{x}' \\in \\mathbb{R}^{256 \\times 64 \\times 3}" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": " for network input. The network utilizes ResNet-18 [18] as the base architecture, followed by an additional fully connected layer for HR estimation. During the TCL calculation, the temporal perturbation " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": " is randomly sampled from a uniform distribution " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbb{U}_{(0,59]}" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": ". Furthermore, we use stochastic gradient descent (SGD) with a momentum of 0.9 as the base optimizer. The bidirectional adaptation mechanism is implemented on top of the SGD. During the adaptation process, we set the learning rate to 0.0001 and the batch size is 1. The hyper-parameters are configured as follows: " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{s}} = 0.001" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathrm{T}} = 8.0" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{t}} = 0.01" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\rho = 0.005" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "k = -9.0" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\Omega = 4000" + }, + { + "bbox": [ + 130, + 475, + 482, + 631 + ], + "type": "text", + "content": ". Please refer to Fig. 5 for detailed ablations." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Li et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 480, + 665 + ], + "type": "text", + "content": "5 Note that the performance of down-sampled VIPL dataset might be inferior than the full VIPL dataset, due to the larger interval between neighboring samples." + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 198, + 126 + ], + "type": "text", + "content": "4.3 Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 255 + ], + "type": "text", + "content": "Following existing rPPG methods [4,33,34,44,51,70], mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (r) are used to evaluate the HR estimation, which inherently reflects the ability to extract physiological periodic signals. The results are organized and recorded in Tab. 1. It is evident that TTA methods generally perform better than DG approaches. Among these TTA methods, our Bi-TTA demonstrates significantly superior performance, showcasing our effectiveness especially when both prospective and retrospective adaptations are synergistically employed. This highlights Bi-TTA's robustness in adapting to various unseen domains, underscoring its potential for real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 271, + 238, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 238, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 238, + 283 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 481, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 481, + 315 + ], + "type": "text", + "content": "In this sub-section, we will mainly discuss about the effectiveness of the proposed priors and the bidirectional adaptation strategy." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 136, + 331, + 282, + 544 + ], + "blocks": [ + { + "bbox": [ + 136, + 331, + 282, + 544 + ], + "lines": [ + { + "bbox": [ + 136, + 331, + 282, + 544 + ], + "spans": [ + { + "bbox": [ + 136, + 331, + 282, + 544 + ], + "type": "image", + "image_path": "3c4faa4af001fabbbdba4f202a0f41e9d91c3c5883dbc485122816b11610854e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 544, + 278, + 554 + ], + "lines": [ + { + "bbox": [ + 135, + 544, + 278, + 554 + ], + "spans": [ + { + "bbox": [ + 135, + 544, + 278, + 554 + ], + "type": "text", + "content": "(a) Comparison between priors and GT." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "text", + "content": "Fig. 6: (a) Comparison between task-specific priors-based TTA and fully supervised adaptation. The evaluation metrics are MSE, RMSE, and r. Alongside the visualization of the loss distribution for both priors-based and supervised adaptation method. Loss values smaller than " + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "text", + "content": " are ignored. Note that for better comparison, the learning rate of supervised adaptation is set to 0.0005, while for the priors-based TTA it is 0.0001. (b) Ablation study of the adaptation performance among the proposed PA and RS modules in our Bi-TTA framework. This Ablation study demonstrates the effectiveness of the PA and RS, in achieving a more effective and stable adaptation process." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 292, + 333, + 481, + 543 + ], + "blocks": [ + { + "bbox": [ + 292, + 333, + 481, + 543 + ], + "lines": [ + { + "bbox": [ + 292, + 333, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 292, + 333, + 481, + 543 + ], + "type": "image", + "image_path": "eebbcab48fbd0376c3d4c5df22baafedeb0d42dadb2203a3b6bf8129ad13bac2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 544, + 460, + 554 + ], + "lines": [ + { + "bbox": [ + 310, + 544, + 460, + 554 + ], + "spans": [ + { + "bbox": [ + 310, + 544, + 460, + 554 + ], + "type": "text", + "content": "(b) Ablation study of PA, RS, and, priors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "Bi-TTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Task-specific priors. As depicted in Fig. 6a, compared to ground-truth supervision, our task-specific priors provide effective adaptation supervision only on a significantly smaller number of samples, but induce considerably larger gradients. Thus, we can infer that the convergence rate under our self-supervised priors is markedly faster than ground-truth, despite a substantially lower learning rate. This phenomenon underscores the effectiveness of our priors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 191, + 482, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 287 + ], + "type": "text", + "content": "As discussed in [71], modern deep learning models are prone to memorize training data due to their over-parameterized design, leading to overfitting risks. Merely minimizing standard loss functions on training data often falls short in ensuring satisfactory generalization ability [13]. Given that the rPPG model has undergone extensive pre-training with supervised learning, its parameters have likely been excessively optimized. In contrast, task knowledge-based priors are expected to offer fresh and correct guidance, steering the model towards improved generalization capability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "content": "Prospective and retrospective adaptation. As depicted in Fig. 6b, the initial convergence rate and adaptation performance of the four comparisons are basically consistent across the first 4000 samples. However, in the latter half, the priors-based method demonstrates noticeable performance degradation. In contrast, other three methods, through utilizing prospective or retrospective adaptations, substantially mitigate this problem, thereby enhancing the adaptation ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 473 + ], + "type": "text", + "content": "Notably, between the " + }, + { + "bbox": [ + 130, + 376, + 482, + 473 + ], + "type": "inline_equation", + "content": "5000^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 376, + 482, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 376, + 482, + 473 + ], + "type": "inline_equation", + "content": "6000^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 376, + 482, + 473 + ], + "type": "text", + "content": " samples, while methods employing prospective adaptation exhibit rapid convergence and superior performance, they also display a significant degree of performance decay due to overfitting or catastrophic forgetting, which even intensifies as more samples are processed. On the other hand, the retrospective adaptation strategy ensures a more stable post-convergence maintenance in the target domain. The Bi-TTA, integrating both prospective and retrospective strategies, adeptly balances convergence speed, effectiveness, and stability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 505, + 220, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 220, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 220, + 519 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we address the challenge of significant performance degradation when adapting rPPG models to unfamiliar target domains. Considering privacy, we implement TTA for rPPG tasks, enabling the adaptation of pre-trained models to the target domain using only unannotated, instance-level target data. The models are fine-tuned during the inference of target samples. In particular, we introduce Bi-TTA, a method that leverages spatial and temporal consistency for effective self-supervision, coupled with pioneering prospective and retrospective adaptation strategies. This methodology streamlines the adaptation process and outperforms existing techniques, demonstrating substantial promise for real-world applications in non-invasive physiological monitoring." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "H. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 139, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "text", + "content": "1. Bahmani, S., Hahn, O., Zamfir, E., Araslanov, N., Cremers, D., Roth, S.: Semantic self-adaptation: Enhancing generalization with a single sample. arXiv preprint arXiv:2208.05788 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Bobbia, S., Macwan, R., Benezeth, Y., Mansouri, A., Dubois, J.: Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters 124, 82-90 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "text", + "content": "3. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: International Conference on Machine Learning. pp. 1059-1071. PMLR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 259 + ], + "type": "text", + "content": "4. Chen, W., McDuff, D.: Deepphys: Video-based physiological measurement using convolutional attention networks. In: Proc. ECCV. pp. 349-365 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "text", + "content": "5. Das, A., Lu, H., Han, H., Dantcheva, A., Shan, S., Chen, X.: Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In: FG. pp. 01-08. IEEE (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 304 + ], + "type": "text", + "content": "6. De Haan, G., Jeanne, V.: Robust pulse rate from chrominance-based rppg. IEEE Trans. Biomed. Eng. 60(10), 2878-2886 (2013)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "text", + "content": "7. De Haan, G., Van Leest, A.: Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiol. Meas. 35(9), 1913 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "type": "text", + "content": "8. D'Innocente, A., Bucci, S., Caputo, B., Tommasi, T.: Learning to generalize one sample at a time with self-supervision. arXiv preprint arXiv:1910.03915 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 349, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 480, + 390 + ], + "type": "text", + "content": "9. Du, J., Liu, S.Q., Zhang, B., Yuen, P.C.: Dual-bridging with adversarial noise generation for domain adaptive rppg estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10355–10364 (June 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "text", + "content": "0. Du, J., Liu, S.Q., Zhang, B., Yuen, P.C.: Dual-bridging with adversarial noise generation for domain adaptive rppg estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10355–10364 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 425, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 468 + ], + "type": "text", + "content": "1. D'Innocente, A., Borlino, F.C., Bucci, S., Caputo, B., Tommasi, T.: One-shot unsupervised cross-domain detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 732-748. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 468, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 480, + 521 + ], + "type": "text", + "content": "12. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1126-1135. PMLR (06-11 Aug 2017), https://proceedings.mlr.press/v70/finn17a.html" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 523, + 480, + 555 + ], + "type": "text", + "content": "13. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=6Tm1mposlrM" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 556, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 599 + ], + "type": "text", + "content": "14. Gandelsman, Y., Sun, Y., Chen, X., Efros, A.A.: Test-time training with masked autoencoders. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=SHMi1b7sjXk" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 600, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 643 + ], + "type": "text", + "content": "15. Gao, J., Zhang, J., Liu, X., Darrell, T., Shelhamer, E., Wang, D.: Back to the source: Diffusion-driven adaptation to test-time corruption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11786-11796 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "16. Gee-Sern Hsu, ArulMurugan Ambikapathi, M.S.C.: Deep learning with time-frequency representation for pulse estimation. In: Proc. IJCB. pp. 642-650 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "Bi-TTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "17. Gideon, J., Stent, S.: The way to my heart is through contrastive learning: Remote photoplethysmography from unlabeled video. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3995-4004 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "text", + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. IEEE CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "text", + "content": "19. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "text", + "content": "20. Kessler, V., Thiam, P., Amirian, M., Schwenker, F.: Pain recognition with camera photoplethysmography. In: IEEE IPTA. pp. 1-5 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "text", + "content": "21. Khurana, A., Paul, S., Rai, P., Biswas, S., Aggarwal, G.: Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "22. Klingner, M., Ayache, M., Fingscheidt, T.: Continual batchnorm adaptation (cbna) for semantic segmentation. IEEE Transactions on Intelligent Transportation Systems 23(11), 20899-20911 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "text", + "content": "23. Krueger, D., Caballero, E., Jacobsen, J.H., Zhang, A., Binas, J., Zhang, D., Le Priol, R., Courville, A.: Out-of-distribution generalization via risk extrapolation (rex). In: ICML. pp. 5815-5826. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 365 + ], + "type": "text", + "content": "24. Lee, E., Chen, E., Lee, C.Y.: Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16. pp. 392–409. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "text", + "content": "25. Lewandowska, M., Ruminski, J., Kocejko, T., Nowak, J.: Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In: FedCSIS. pp. 405-410. IEEE (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "26. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "27. Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "text", + "content": "28. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International Conference on Machine Learning (ICML). pp. 6028-6039 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "29. Liang, J., Hu, D., Wang, Y., He, R., Feng, J.: Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021), in Press" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "30. Liu, X., Jiang, Z., Fromm, J., Xu, X., Patel, S., McDuff, D.: Metaphys: few-shot adaptation for non-contact physiological measurement. In: Proceedings of the conference on health, inference, and learning. pp. 154-163 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "31. Liu, X., Zhang, Y., Yu, Z., Lu, H., Yue, H., Yang, J.: rppg-mae: Self-supervised pretraining with masked autoencoders for remote physiological measurement. arXiv preprint arXiv:2306.02301 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "32. Liu, Y., Kothari, P., van Delft, B.G., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "33. Lu, H., Han, H., Zhou, S.K.: Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In: Proc. IEEE CVPR. pp. 12404-12413 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "34. Lu, H., Yu, Z., Niu, X., Chen, Y.C.: Neuron structure modeling for generalizable remote physiological measurement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18589-18599 (June 2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 459, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 459, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 459, + 127 + ], + "type": "text", + "content": "35. McDuff, D.: Camera measurement of physiological vital signs. CSUR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 128, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 128, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 480, + 148 + ], + "type": "text", + "content": "36. McDuff, D.: Applications of camera-based physiological measurement beyond healthcare. In: CVSM, pp. 165-177. Elsevier (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 181 + ], + "type": "text", + "content": "37. McDuff, D., Gontarek, S., Picard, R.W.: Improvements in remote cardiopulmonary measurement using a five band digital camera. IEEE Trans. Biomed. Eng. 61(10), 2593-2601 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "38. Mi, P., Shen, L., Ren, T., Zhou, Y., Sun, X., Ji, R., Tao, D.: Make sharpness-aware minimization stronger: A sparsified perturbation approach (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "type": "text", + "content": "39. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "40. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "41. Niu, X., Han, H., Shan, S., Chen, X.: Synrhythm: Learning a deep heart rate estimator from general to specific. In: Proc. IEEE ICPR. pp. 3580-3585 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "type": "text", + "content": "42. Niu, X., Han, H., Shan, S., Chen, X.: VIPL-HR: A multi-modal database for pulse estimation from less-constrained face video. In: Proc. ACCV. pp. 562-576 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "text", + "content": "43. Niu, X., Shan, S., Han, H., Chen, X.: Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Trans. on Image Process. 29, 2409-2423 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "text", + "content": "44. Niu, X., Yu, Z., Han, H., Li, X., Shan, S., Zhao, G.: Video-based remote physiological measurement via cross-verified feature disentangling. In: Proc. ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "45. Pham, H., Dai, Z., Xie, Q., Le, Q.V.: Meta pseudo labels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11557-11568 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 396, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 427 + ], + "type": "text", + "content": "46. Poh, M.Z., McDuff, D.J., Picard, R.W.: Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express 18(10), 10762-10774 (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "text", + "content": "47. Qiu, Y., Liu, Y., Arteaga-Falconi, J., Dong, H., El Saddik, A.: Evm-cnn: Real-time contactless heart rate estimation from facial video. IEEE Trans. Multimedia (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "text", + "content": "48. Reiss, A., Indlekofer, I., Schmidt, P., Van Laerhoven, K.: Deep ppg: large-scale heart rate estimation with convolutional neural networks. Sensors 19(14), 3079 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "text", + "content": "49. Revanur, A., Li, Z., Ciftci, U.A., Yin, L., Jeni, L.A.: The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In: Proc. CVPR workshop. pp. 2760-2767 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "text", + "content": "50. Song, R., Chen, H., Cheng, J., Li, C., Liu, Y., Chen, X.: Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE J-BHI pp. 1-1 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 548, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 480, + 579 + ], + "type": "text", + "content": "51. Song, R., Zhang, S., Li, C., Zhang, Y., Cheng, J., Chen, X.: Heart rate estimation from facial videos using a spatiotemporal representation with convolutional neural networks. IEEE Trans. Instrum. Meas. (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "text", + "content": "52. Spetlik, R., Franc, V., Matas, J.: Visual heart rate estimation with convolutional neural network. In: Proc. BMVC. pp. 3-6 (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "53. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: Proc. IEEE ISRHIC. pp. 1056-1062 (2014)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "text", + "content": "54. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: ECCV. pp. 443-450. Springer (2016)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "55. Sun, B., Wei, Q., Li, L., Xu, Q., He, J., Yu, L.: LSTM for dynamic emotion and group emotion recognition in the wild. In: Proc. ACM ICMI. pp. 451-457 (2016)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "Bi-TTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 660 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "text", + "content": "56. Sun, W., Zhang, X., Lu, H., Chen, Y., Ge, Y., Huang, X., Yuan, J., Chen, Y.: Resolve domain conflicts for generalizable remote physiological measurement. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8214-8224 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 193 + ], + "type": "text", + "content": "57. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "type": "text", + "content": "58. Sun, Z., Li, X.: Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In: European Conference on Computer Vision. pp. 492-510. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "text", + "content": "59. Tian, C.X., Li, H., Xie, X., Liu, Y., Wang, S.: Neuron coverage-guided domain generalization. IEEE Trans. on Pattern Anal. Mach. Intell. (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 266 + ], + "type": "text", + "content": "60. Verkruysse, W., Svaasand, L.O., Nelson, J.S.: Remote plethysmographic imaging using ambient light. Opt. Express 16(26), 21434-21445 (2008)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 266, + 481, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 481, + 299 + ], + "type": "text", + "content": "61. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 299, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 481, + 341 + ], + "type": "text", + "content": "62. Wang, H., Ahn, E., Kim, J.: Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2431-2439 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "text", + "content": "63. Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Yu, P.: Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 373, + 481, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 481, + 395 + ], + "type": "text", + "content": "64. Wang, W., den Brinker, A.C., Stuijk, S., de Haan, G.: Algorithmic principles of remote ppg. IEEE Trans. Biomed. Eng. 64(7), 1479-1491 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 395, + 481, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 395, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 481, + 416 + ], + "type": "text", + "content": "65. Wang, W., Stuijk, S., De Haan, G.: Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE Trans. Biomed. Eng. 62(2), 415-425 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 416, + 481, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 481, + 437 + ], + "type": "text", + "content": "66. Wang, Z.K., Kao, Y., Hsu, C.T.: Vision-based heart rate estimation via a two-stream cnn. In: Proc. IEEE ICIP. pp. 3327-3331 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 437, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 481, + 468 + ], + "type": "text", + "content": "67. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems 33, 2958-2969 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "68. Xi, L., Chen, W., Zhao, C., Wu, X., Wang, J.: Image enhancement for remote photoplethysmography in a low-light environment. In: FG. pp. 1-7. IEEE (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 521 + ], + "type": "text", + "content": "69. Yu, Z., Peng, W., Li, X., Hong, X., Zhao, G.: Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In: Proc. IEEE ICCV. pp. 151-160 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "70. Yu, Z., Shen, Y., Shi, J., Zhao, H., Torr, P., Zhao, G.: Physformer: Facial videobased physiological measurement with temporal difference transformer. IEEE CVPR (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 553, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 481, + 586 + ], + "type": "text", + "content": "71. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 586, + 481, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 481, + 606 + ], + "type": "text", + "content": "72. Zhang, X., Chen, Y.C.: Adaptive domain generalization via online disagreement minimization. IEEE Transactions on Image Processing (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 606, + 481, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 481, + 628 + ], + "type": "text", + "content": "73. Zhao, X., Liu, C., Sicilia, A., Hwang, S.J., Fu, Y.: Test-time fourier style calibration for domain generalization. arXiv preprint arXiv:2205.06427 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 628, + 481, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 628, + 481, + 660 + ], + "spans": [ + { + "bbox": [ + 132, + 628, + 481, + 660 + ], + "type": "text", + "content": "74. Zou, Y., Zhang, Z., Li, C.L., Zhang, H., Pfister, T., Huang, J.B.: Learning instance-specific adaptation for cross-domain segmentation. In: European Conference on Computer Vision. pp. 459-476. Springer (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bi-directional Contextual Attention for 3D Dense Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_content_list.json b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7219962315d347efcad73314cc7646106eba023c --- /dev/null +++ b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_content_list.json @@ -0,0 +1,1751 @@ +[ + { + "type": "text", + "text": "Bi-directional Contextual Attention for 3D Dense Captioning", + "text_level": 1, + "bbox": [ + 277, + 140, + 725, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minjung Kim $^{1,4,*}$ , Hyung Suk Lim $^{3}$ , Soonyoung Lee $^{2}$ , Bumsoo Kim $^{2,\\dagger}$ , and Gunhee Kim $^{1,\\dagger}$", + "bbox": [ + 295, + 210, + 707, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Seoul National University, Seoul, Korea minjung.kim@vision.snu.ac.kr, gunhee@snu.ac.kr", + "bbox": [ + 328, + 253, + 671, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ LG AI Research, Seoul, Korea {soonyoung.lee, bumsoo.kim}@lgresearch.ai", + "bbox": [ + 352, + 282, + 648, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Dioxide, Seoul, Korea hslim@dioxide.com", + "bbox": [ + 421, + 310, + 581, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 SNU-LG AI Research Center, Seoul, Korea", + "bbox": [ + 349, + 338, + 651, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D dense captioning is a task involving the localization of objects and the generation of descriptions for each object in a 3D scene. Recent approaches have attempted to incorporate contextual information by modeling relationships with object pairs or aggregating the nearest neighbor features of an object. However, the contextual information constructed in these scenarios is limited in two aspects: first, objects have multiple positional relationships that exist across the entire global scene, not only near the object itself. Second, it faces with contradicting objectives—where localization and attribute descriptions are generated better with tight localization, while descriptions involving global positional relations are generated better with contextualized features of the global scene. To overcome this challenge, we introduce BiCA, a transformer encoder-decoder pipeline that engages in 3D dense captioning for each object with Bi-directional Contextual Attention. Leveraging parallelly decoded instance queries for objects and context queries for non-object contexts, BiCA generates object-aware contexts, where the contexts relevant to each object is summarized, and context-aware objects, where the objects relevant to the summarized object-aware contexts are aggregated. This extension relieves previous methods from the contradicting objectives, enhancing both localization performance and enabling the aggregation of contextual features throughout the global scene; thus improving caption generation performance simultaneously. Extensive experiments on two of the most widely-used 3D dense captioning datasets demonstrate that our proposed method achieves a significant improvement over prior methods.", + "bbox": [ + 259, + 378, + 743, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 744, + 375, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D dense captioning is a task that requires 1) determining the location of all objects and 2) generating descriptive sentences for each object detected within a 3D", + "bbox": [ + 212, + 772, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during internship at LG AI Research", + "bbox": [ + 217, + 810, + 557, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 222, + 825, + 387, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dedbdc99e53d499fdc716426510288b55ffac3aac0427c6f53e73d1048ec3f5b.jpg", + "image_caption": [ + "Fig. 1: Conceptual illustration of the multi-stage pipeline of BiCA (best viewed in color)." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 782, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scene. Early approaches have utilized a two-stage process [4, 7, 11, 19, 33, 36, 37], where the objects are detected first, then captions are generated afterward by sequentially feeding the features of each detected object. Following work [9] has borrowed the end-to-end transformer encoder-decoder pipeline from object detection [5] to actively involve contextual information for 3D dense captioning. In these works, contextual information for predicting relationships is composed by combining object features (e.g., with dot product [4, 33] or transformer attention [9]) and aggregating local nearest neighbor visual features [9].", + "bbox": [ + 212, + 363, + 787, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their improved performance, the scope of contextual information in previous 3D dense captioning methods poses its own set of challenges. First, 3D dense captioning heavily involves various positional relationships throughout the global scene; thus, the context scope restricted to the combination of individual objects or their spatially nearest neighbors typically cannot provide sufficient information. Next, the endeavor to incorporate the contextual features of various positions into a single object encounters a conflicting objective since the object must encompass precise local features for localization while simultaneously capturing various positional relationships throughout the global context for caption generation. Clearly, these challenges limit the performance of 3D dense captioning, which raises an interesting research question—“Can we design an architecture for 3D dense captioning that can effectively aggregate relevant context features without harming localization performance?”", + "bbox": [ + 212, + 489, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a Bi-directional Contextual Attention (BiCA) network for 3D dense captioning. BiCA parallelly decodes a fixed set of object queries and context queries sampled from the geometric positions that do not overlap with each other. While the decoded object query performs standard object detection, the context queries are designed to distinctively capture non-object contexts and subsequently aggregate relevant contextual information for each object. Recognizing that non-object regions often lack sufficient visual features in 3D point clouds, BiCA incorporates a multi-stage pipeline where context features are enriched by relevant object features, thus implementing bidirectional attention: i) Contextual Attention of Objects for Context (O4C):", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Kim et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "context representation is represented by the attention summarization of global objects. ii) Contextual Attention of Contexts for Object (C4O): the object representation is adjusted by summarizing the non-object contexts.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, localization is inferred with object queries while the captions are generated with object-aware context and context-aware object features. By disentangling the attention for localization and the attention for incorporating contextual features, BiCA successfully resolves the aforementioned contradicting objectives for local and global features. Moreover, since the contextual attentions retrieve globally relevant contexts and objects throughout the scene, it could accurately produce descriptions that transcend the spatial boundary of localization and its nearest neighbor. The conceptual overview of our BiCA is illustrated in Figure 1.", + "bbox": [ + 212, + 193, + 784, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BiCA sets a new state-of-the-art standard for 3D dense captioning, while simultaneously improving 3D object detection performance. Extensive experiments on two widely used benchmarks in 3D dense captioning (i.e., ScanRefer [6] and Nr3D [1]) show that our proposed BiCA surpasses prior approaches by a large margin. The contribution of our paper can be summarized as:", + "bbox": [ + 212, + 315, + 785, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose a novel Bi-directional Contextual Attention network with multi-stage contextual attention for 3D dense captioning. This enables our model to capture relevant contexts throughout the global scene without being bound to single-object localization or their nearest neighbors.", + "bbox": [ + 225, + 405, + 782, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- By performing localization with decoded object queries and caption generation by including object-aware context features and context-aware object correspondence features, BiCA can simultaneously improve the performance of localization and caption generation for 3D dense captioning.", + "bbox": [ + 225, + 467, + 784, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Our BiCA achieves state-of-the-art performances across multiple evaluation metrics on two widely used benchmarks for 3D dense captioning: ScanRefer and Nr3D datasets.", + "bbox": [ + 225, + 529, + 784, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 214, + 604, + 382, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Image Captioning", + "text_level": 1, + "bbox": [ + 214, + 642, + 410, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image captioning is a fundamental task in visual language creation that automatically generates descriptive sentences for images. The main goal is to improve 2D scene understanding and generate captions that accurately reflect the content and context of the image [17]. However, with the advancement of deep learning in image captioning, there has been a marked shift towards the development and adoption of more sophisticated models [16]. Image captioning methods using deep learning adopt an encoder-decoder architecture, where the decoder generates a sentence from the visual features extracted by the encoder. Attention-based methods for grid regions [34] and detected objects [2] focus on specific image regions and use graph neural networks [35] or transformer layers to capture relationships between objects [14].", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bi-directional Contextual Attention for 3D Dense Captioning", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 3D Dense Captioning", + "text_level": 1, + "bbox": [ + 215, + 146, + 439, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D dense captioning, an emerging field focused on achieving detailed object-level understanding of 3D scenes through natural language descriptions, has garnered significant interest in recent years. This task involves converting 3D visual data [15] into a consistent set of bounding boxes and generating appropriate natural language descriptions for each identified object. This task presents a considerable challenge due to the complexity of 3D scene information.", + "bbox": [ + 212, + 174, + 785, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most 3D dense captioning models employ an encoder-decoder architecture comprising three main components: a scene encoder, a relational module, and a feature decoder. The scene encoder in 3D dense captioning utilizes 3D object detection methods such as 3DETR [24], PointNet++ [28], VoteNet [27], and PointGroup [18] to extract object-level visual features and contextual details from 3D point cloud datasets. The relational module is crucial in 3D dense captioning, modeling complex connections and cross-modal interactions among objects within indoor scenes. Relational modeling approaches vary based on task requirements and can be categorized into graph-based [7, 11, 19], transformer-based [4, 9, 33], and knowledge distillation-based [36] methods. The feature decoder generates bounding boxes and captions for candidate objects by considering attributes and relational features. GRU-based decoders with attention mechanisms are used in Scan2Cap [11], MORE [19], and D3Net [7], while transformer-based decoders are employed in SpaCap3D [33], $\\chi$ -Tran2Cap [36], 3DJCG [4], and Vote2Cap-DETR [9] to facilitate the caption generation process.", + "bbox": [ + 212, + 266, + 787, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, most existing methods use a single query set for object localization and caption generation, which often leads to conflicts in achieving the objectives of the two tasks. To address this, Vote2Cap-DETR++ [10] enhances task-specific feature capture by decoupling queries for localization and captioning, overcoming the limitations of a unified query system. Despite this improvement, challenges persist as its performance remains fundamentally bound by the precision of its object localization capabilities.", + "bbox": [ + 212, + 492, + 787, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 623, + 331, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to take a 3D scene as input, identify all the objects within it, and generate captions for each object. Previous research has adopted a pipeline to identify objects in a 3D scene and simultaneously generate captions from the same features. However, while each caption may describe features of the object itself, it can also be based on the relationships with surrounding objects or information about the overall scene.", + "bbox": [ + 212, + 657, + 785, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section introduces a novel transformer encoder-decoder approach with bi-directional contextual attention for the 3D dense captioning task. This approach distinguishes between features for instances and context information between instances, indirectly supervising the context features to enable end-to-end learning with the instance features. Instance queries aim to localize instances within a 3D scene and capture the characteristics of instances. In contrast, each", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Kim et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3cbb63a63752e186642596f7de173da8e76e2fd3de8941a26b3e0f5bd32b526.jpg", + "image_caption": [ + "Fig. 2: The overall pipeline of BiCA. We parallelly generate and decode two sets of queries (i.e., Instance Query and Context Query) that encodes the instance features and the non-object context features throughout the global scene, respectively. The object-aware contexts are calculated per each object by the weighted sum of the context queries, where the weights are calculated by the attention between the decoded instance query and context query. Then, with the object-aware context feature, the context-aware object feature is obtained by the weighted sum of the instances, which is weighted by the attention between the object-aware contexts." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 781, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "context query is designed to capture the non-object region within a scene, including positional relationships between instances in the 3D scene. The overall architecture of the newly proposed model is described in Figure 2.", + "bbox": [ + 212, + 470, + 787, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Encoder", + "text_level": 1, + "bbox": [ + 215, + 539, + 331, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous work [9], our model also adopts 3DETR [24] encoder as our scene encoder. This encoder is applied to transform the input point cloud into a set of encoded tokens that capture spatial, structural, and contextual information. Given the input point cloud $PC = [p_{\\mathrm{in}};f_{\\mathrm{in}}]\\in \\mathbb{R}^{N\\times (3 + F)}$ , it is initially tokenized by a set-abstraction layer of PointNet++ [28]. This tokenized output is subsequently fed into a masked transformer encoder incorporating the set-abstraction layer, followed by two additional encoder layers. The final encoded scene tokens are denoted as $p_{enc}\\in \\mathbb{R}^{1,024\\times 3}$ and $f_{enc}\\in \\mathbb{R}^{1,024\\times 256}$ .", + "bbox": [ + 212, + 564, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Query Generator", + "text_level": 1, + "bbox": [ + 215, + 709, + 405, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To separate captions bound to a single object from captions containing information relative to other objects or the global scene, we specify two separate Instance Query and Context Query from the encoded scene. While the instance query set is designed to capture individual features of objects bounded by their localization, the context query set is designed to encompass a broad area that consists of the positional information of a caption (e.g., next to, on the right of, is most far from, etc.).", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bi-directional Contextual Attention for 3D Dense Captioning", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance Query Generator. In the Instance Query Generator, instance queries are generated to detect all objects in the 3D scene and generate captions for each object's individual attributes. The instance query $(p^o,f^o)$ is defined as follows:", + "bbox": [ + 212, + 146, + 789, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n[ \\Delta p _ {\\text {v o t e}}; \\Delta f _ {\\text {v o t e}} ] = \\mathrm {F F N} _ {o} (f _ {\\text {e n c}}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 205, + 785, + 220 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left(p ^ {o}, f ^ {o}\\right) = \\mathrm {S A} _ {o} \\left(p _ {\\text {e n c}} + \\Delta p _ {\\text {v o t e}}, f _ {\\text {e n c}} + \\Delta f _ {\\text {v o t e}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 247, + 785, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $[\\varDelta p_{\\mathrm{vote}};\\varDelta f_{\\mathrm{vote}}]\\in\\mathbb{R}^{1,024\\times(3+256)}$ is an offset that learns to shift the encoded points to object's centers spatially by a feed-forward network $\\mathrm{FFN}_o$ , following [9]. $\\mathrm{SA}_o$ represents the set-abstraction layer with a radius of 0.3 and samples 16 points for $p^o$ . Since our Instance Query Generator extracts the features from the candidate coordinates after the voting, our instance queries are not focused on specific objects. However, they are distributed across the locations where objects are present throughout the 3D scene. As a result, 256 instance queries are extracted from our Instance Query Generator.", + "bbox": [ + 212, + 272, + 787, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Context Query Generator. 3D scenes are composed of 3D points arranged in XYZ coordinates, forming specific objects, and the spaces between objects signify space. We recognize that obtaining context features should not merely involve interpolating the spaces between Instance Queries but also encompassing how these points are arranged and the composition of empty spaces to represent a scene. Inspired by this observation, we define context information as the geometric details in a 3D scene that illustrate the relationships between objects and between objects and the scene itself. We decide to extract structural information centered on specific points in the 3D scene and call them Context Queries. This approach is distinct from the Instance Query Generator. While the Instance Query Generator uses a voting network to obtain object center coordinates that do not exist in the 3D scene and extracts features such as identification from them, the proposed Context Query Generator extracts structural information without disrupting the arrangement of the 3D points.", + "bbox": [ + 212, + 417, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the encoded scene tokens $(p_{\\mathrm{enc}}, f_{\\mathrm{enc}})$ , we sample 512 points $p_{\\mathrm{seed}}^c$ with farthest point sampling (FPS) [28] on $p_{\\mathrm{enc}}$ that each represent a unique geometry within the global scene. The context query $(p^c, f^c)$ is then defined as:", + "bbox": [ + 212, + 628, + 787, + 674 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n(p ^ {c}, f ^ {c}) = \\mathrm {S A} _ {c} (p _ {\\mathrm {e n c}}, f _ {\\mathrm {e n c}}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 688, + 785, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{SA}_c$ denotes the set-abstraction layer [28] with a radius of 1.2, sampling 64 points for $p^c$ . All hyper-parameters are determined experimentally.", + "bbox": [ + 212, + 714, + 787, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Decoder", + "text_level": 1, + "bbox": [ + 214, + 768, + 333, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instance Decoder and Context Decoder. Given the context queries $(p^c, f^c)$ and instance queries $(p^o, f^o)$ , we build a parallel decoding pipeline where the Context Decoder designates contextual information between objects and the", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Kim et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instance Decoder performs the localization and attribute description for the participating objects.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Each Instance Decoder and Context Decoder consists of a transformer decoder with 8 layers according to 3DETR [24], and Fourier positional encoding is applied to XYZ coordinates. Fourier positional encoding transforms the XYZ coordinates of each query into a Fourier position embedding [31] and then combines them with the corresponding query embeddings. This encoding transforms the input data into frequency functions, capturing invariant changes in complex shapes and structures. It aids the context queries in accurately identifying geometric shapes and the instance queries in detecting objects. The context queries $V^{c}$ are afterwards contextualized for each instance query $V^{o}$ , then retrieve relevant instance query features that are structurally involved.", + "bbox": [ + 212, + 176, + 787, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Bi-directional Contextual Attention", + "text_level": 1, + "bbox": [ + 214, + 349, + 555, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Captions must be generated for each object in relation to the scene context, and through positional encoding, it becomes possible to understand directional relationships, such as \"to the right\" with surrounding objects. Here, we gain the insight that information about the relative objects closely related to these relationships is also necessary, and we design the model to proceed simultaneously with a multi-stage contextualization. With the decoded queries, our method proceeds in two stages: i) Contextual Attention of Objects for Context (O4C): context representation is represented by the attention summarization of global objects. ii) Contextual Attention of Contexts for Object (C4O): the object representation is adjusted by summarizing the non-object contexts. Figure 1 and Figure 2 show detailed illustrations of obtaining the object-aware context feature and the context-aware object feature.", + "bbox": [ + 212, + 375, + 787, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contextual Attention of Objects for Context (O4C). To retrieve the structural relationship between objects throughout the global scene, we first construct a Object-aware Context $V_{a}^{c}$ per object by summating the attention of each context position feature with regard to each object. The attention result between the instance feature and the context feature is applied to the context feature, making a weighted summarization of the geometries. At this time, the ratio is adjusted using learnable gamma. We name this result $V_{a}^{c}$ .", + "bbox": [ + 212, + 575, + 787, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contextual Attention of Contexts for Object (C4O). Most contexts not only interact with the reference object but also have other objects that react against them. To achieve this, we go through a Contextual Attention of Contexts for Object process. This could specify the information that can simply be said to be 'next to' as 'next to the red chair'. To this end, the attention result of the context feature and instance feature with the attention weight applied is applied to the instance query feature. Likewise, it is adjusted using a learnable parameter lambda. We name this result $V_{a}^{o}$ . $V_{a}^{o}$ results from finding instance combinations that have a meaningful relationship with the contextual region focused on a", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bi-directional Contextual Attention for 3D Dense Captioning", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "specific instance. The concatenation of the instance token $V^o$ , contextualized geometry $V_a^c$ , and contextualized object $V_a^o$ , denoted as $V^a$ , is fed to the caption head to generate final captions.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 3D Dense Captioning", + "text_level": 1, + "bbox": [ + 215, + 214, + 439, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our final goal is to identify all objects within the input 3D scene and generate a caption for each object. We perform object detection and caption generation in parallel and group the results into one instance.", + "bbox": [ + 212, + 239, + 785, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Localization. To localize instances in the 3D scene, we use the decoded instance query $V^{o}$ . Through 5 MLP heads, we reformulate the box corner estimation as offset estimation from a query point to an object's center and box size regression. These localization heads are shared across the decoder layers.", + "bbox": [ + 212, + 308, + 785, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Caption Generation. For caption generation, we use a transformer decoder-based caption head based on GPT-2 [29], following the methods of Vote2CapDETR [9] and SpaCap3D [33]. Our caption head comprises two transformer decoder blocks, sinusoid positional embedding, and a linear classification layer. When generating captions for a proposal, we substitute the standard Start Of Sequence ('SOS') prefix with $V^a$ . During the inference process, our approach employs beam search to produce captions. The beam size is 5.", + "bbox": [ + 212, + 391, + 787, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 Training BiCA", + "text_level": 1, + "bbox": [ + 215, + 520, + 385, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instance Query Loss. To train instance query generator to find an object's center by shifting points $p_{enc}$ , we use the vote loss from VoteNet [27]. Given the generated instance query $(p^o, f^o)$ and the encoded scene tokens $(p_{enc}, f_{enc})$ , the vote loss $\\mathcal{L}^o$ is denoted as:", + "bbox": [ + 212, + 545, + 785, + 606 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {o} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {N _ {g t}} \\| p _ {i} ^ {o} - \\operatorname {c n t} _ {j} \\| _ {1} \\cdot \\mathbb {I} \\left(p _ {e n c} ^ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 616, + 785, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbb{I}(x)$ is an indicator function that equals 1 when $x \\in I_j$ and 0 otherwise, $N_{gt}$ is the number of instances in the 3D scene, $M$ is the number of $p^o$ , and $\\mathrm{cnt}_j$ is the center of $j$ -th instance $I_j$ .", + "bbox": [ + 214, + 670, + 785, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detection Loss. Following DETR [5], we use Hungarian matching [21] to assign each proposal with the ground truth. The detection loss $\\mathcal{L}_{\\mathrm{det}}$ is defined as:", + "bbox": [ + 212, + 738, + 785, + 770 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d e t}} = \\alpha_ {1} \\mathcal {L} _ {\\mathrm {g i o u}} + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {c l s}} + \\alpha_ {3} \\mathcal {L} _ {\\mathrm {c n t}} + \\alpha_ {4} \\mathcal {L} _ {\\mathrm {s i z e}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 782, + 785, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $\\alpha_{1} = 10, \\alpha_{2} = 1, \\alpha_{3} = 5, \\alpha_{4} = 1$ set heuristically. This loss is applied across all decoder layers for better convergence.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Kim et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Caption Loss. Following the standard image captioning protocol, we first train the caption heads using cross-entropy loss for Maximum Likelihood Estimation (MLE). During MLE training, the model predicts the $(t + 1)$ -th word $c_{n}^{t + 1}$ based on the first $t$ words $c_{n}^{[1:t]}$ and the visual features $\\mathcal{V}$ . The loss function for a final caption of length $T$ is defined as follows:", + "bbox": [ + 214, + 146, + 782, + 224 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c _ {n}} = \\sum_ {t = 1} ^ {T} \\mathcal {L} _ {c _ {n}} (t) = - \\sum_ {t = 1} ^ {T} \\log \\hat {P} \\left(c _ {n} ^ {t + 1} | \\mathcal {V}, c _ {n} ^ {[ 1: t ]}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 238, + 784, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After word-level supervision, the caption head is refined using Self-Critical Sequence Training (SCST) [30], where the model generates multiple captions $\\hat{c}_{1,\\dots,k}$ with a beam size of $k$ and an additional caption $\\hat{g}$ via greedy search. The loss function for SCST is formulated as follows:", + "bbox": [ + 214, + 290, + 782, + 349 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c _ {n}} = - \\sum_ {i = 1} ^ {k} \\left(R \\left(\\hat {c} _ {i}\\right) - R (\\hat {g})\\right) \\cdot \\frac {1}{\\left| \\hat {c} _ {i} \\right|} \\log \\hat {P} \\left(\\hat {c} _ {i} \\mid \\mathcal {V}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 362, + 784, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The reward function $R(\\cdot)$ is based on the CIDEr [32] metric for caption evaluation, and the logarithmic probability of the caption $\\hat{c}_i$ is normalized by its length $|\\hat{c}_i|$ , ensuring equal importance to captions of different lengths.", + "bbox": [ + 214, + 415, + 782, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Final Loss. Given the instance query Loss $\\mathcal{L}^o$ , the detection loss for the $i$ -th decoder layer as $\\mathcal{L}_{\\mathrm{det}}^i$ , and the average of the caption loss $\\mathcal{L}_{c_n}$ within a batch, denoted as $\\mathcal{L}_{\\mathrm{cap}}$ , the final loss $\\mathcal{L}$ is formulated as:", + "bbox": [ + 214, + 486, + 782, + 532 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\beta_ {1} \\mathcal {L} ^ {o} + \\beta_ {2} \\sum_ {i = 1} ^ {n _ {\\mathrm {d e c - l a y e r}}} \\mathcal {L} _ {\\mathrm {d e t}} ^ {i} + \\beta_ {3} \\mathcal {L} _ {\\mathrm {c a p}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 546, + 784, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\beta_{1} = 10$ , $\\beta_{2} = 1$ , and $\\beta_{3} = 5$ .", + "bbox": [ + 214, + 597, + 468, + 613 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 638, + 375, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Metrics", + "text_level": 1, + "bbox": [ + 214, + 675, + 437, + 689 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We focus on 3D dense captioning, leveraging two benchmark datasets: ScanRefer [6] and $\\mathrm{Nr3D}$ [1]. These datasets offer an extensive human-generated description of 3D scenes and objects. ScanRefer encompasses 36,665 descriptions covering 7,875 objects within 562 scenes, while $\\mathrm{Nr3D}$ contains 32,919 descriptions for 4,664 objects across 511 scenes. Both datasets draw their training data from the ScanNet [15] database, which includes 1,201 3D scenes. For evaluation, we use 9,508 descriptions for 2,068 objects across 141 scenes from ScanRefer and 8,584 descriptions for 1,214 objects in 130 scenes from $\\mathrm{Nr3D}$ , all sourced from the 312 3D scenes in the ScanNet validation set.", + "bbox": [ + 212, + 704, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bi-directional Contextual Attention for 3D Dense Captioning", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics. We evaluate model performance using four metrics: CIDEr [32], BLEU-4 [25],METEOR [3], and ROUGE-L [13], denoted as $\\mathbf{C}$ , B-4, $\\mathbf{M}$ , and $\\mathbf{R}$ , respectively. Following the previous studies [4,9,11,19,33], Non-Maximum Suppression (NMS) is initially applied to filter out duplicate object predictions among the proposals. Each proposal is represented as a pair consisting of a predicted bounding box $\\hat{b} i$ and its associated caption $\\hat{c} i$ . To accurately evaluate the model's capacity for object localization and caption generation, we employ the metric $m@k$ , setting the IoU thresholds at 0.25 and 0.5 for our experiments, following [11]:", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nm @ k = \\frac {1}{N} \\sum_ {i = 1} ^ {N} m \\left(\\hat {c} _ {i}, C _ {i}\\right) \\cdot \\mathbb {I} \\left\\{\\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {i}\\right) \\geq k \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 282, + 785, + 324 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $N$ denotes the total number of annotated objects in the evaluation set, and $m$ stands for the captioning metrics C, B-4, M, and R.", + "bbox": [ + 214, + 335, + 784, + 367 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 397, + 455, + 412 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Following the [9], our training comprises three stages. We pre-train our network without the caption head on the ScanNet dataset [15] for 1,080 epochs, using a batch size of 8. We minimize the loss function with an AdamW optimizer [22], starting with a learning rate of $5 \\times 10^{-4}$ that reduces to $10^{-6}$ using a cosine annealing schedule, along with a weight decay of 0.1 and gradient clipping at 0.1 as per [24]. We then proceed to jointly train the model using standard cross-entropy loss for 720 epochs on both ScanRefer [6] and Nr3D [1], maintaining the detector's learning rate at $10^{-6}$ and reducing the caption head from $10^{-4}$ to $10^{-6}$ to avoid overfitting. In the SCST [30], we adjust the caption head using a batch size of 2 while keeping the detector fixed over a span of 180 epochs and maintain a constant learning rate of $10^{-6}$ . Additionally, for experiments incorporating 2D data, we use the pre-trained ENet [8] to extract 128-dimensional multi-view features from ScanNet images as outlined in Scan2Cap [11]. The model has 16.9M parameters, and the average inference time on ScanRefer [6] is 1.8ms. All experiments are conducted on a single Titan RTX GPU using PyTorch [26].", + "bbox": [ + 212, + 429, + 787, + 657 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 214, + 686, + 555, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we evaluate our performance against state-of-the-art methods: Scan2Cap [11], D3Net [7], SpaCap3D [33], MORE [19], 3DJCG [4], Contextual [37], REMAN [23], 3D-VLP [20], $\\chi$ -Tran2Cap [36], Vote2Cap-DETR [9], Unit3D [12], and Vote2Cap-DETR++ [10]. We apply IoU thresholds of 0.25 and 0.5 for ScanRefer [6] as shown in Table 1 and an IoU threshold of 0.5 for Nr3D [1] indicated in Table 2. For the baselines, we present the evaluation results reported in the original papers, and \"-\" in Table 1 and Table 2 indicates that such results have not been reported in the original paper.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/09c33e2cd3a2d3f738e7b161af02ba8fe7bbad3b0a45560adbe7f2f135225525.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTrainingw/o additional 2D dataw/ additional 2D data
IoU=0.25IoU=0.50IoU=0.25IoU=0.50
C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑
Scan2Cap53.7334.2526.1454.9535.2022.3621.4443.5756.8234.1826.2955.2739.0823.3221.9744.78
D3Net------------46.0730.2924.3551.67
SpaCap3d58.0635.3026.1655.0342.7625.3822.8445.6663.3036.4626.7155.7144.0225.2622.3345.36
MORE58.8935.4126.3655.4138.9823.0121.6544.3362.9136.2526.7556.3340.9422.9321.6644.42
3DJCG60.8639.6727.4559.0247.6831.5324.2851.8064.7040.1727.6659.2349.4831.0324.2250.80
Contextual----42.7723.6022.0545.13----46.1125.4722.6445.96
REMANMLE--------62.0136.3727.7656.2545.0026.3122.6746.96
3D-VLP64.0939.8427.6558.7850.0231.8724.5351.1770.7341.0328.1459.7254.9432.3124.8351.51
Vote2Cap-DETR71.4539.3428.2559.3361.8134.4626.2254.4072.7939.1728.0659.2359.3232.4225.2852.38
Unit3D------------46.6927.2221.9145.98
Vote2Cap-DETR++76.3641.3728.7060.0067.5837.0526.8955.6477.0340.9928.5359.5964.3234.7326.0453.67
BiCA (Ours)78.4241.4628.8260.0268.4638.2327.5658.5678.3541.2028.8259.8066.4736.1326.7154.54
Scan2Cap------------48.3826.0922.1544.74
D3Net------------62.6435.6825.7253.90
x-Tran2Cap58.8134.1725.8154.1041.5223.8321.9044.9761.8335.6526.6154.7043.8725.0522.4645.28
ContextualSCST----50.2925.6422.5744.71----54.3027.2423.3045.81
Vote2Cap-DETR84.1542.5128.4759.2673.7738.2126.6454.7186.2842.6428.2759.0770.6335.6925.5152.28
Vote2Cap-DETR++88.2844.0728.7559.8978.1639.7226.9455.5288.5643.3028.6459.1974.4437.1826.2053.30
BiCA (Ours)89.7244.9728.9660.6980.1440.1627.7656.1089.3444.5628.7459.3376.3437.3426.6054.00
", + "bbox": [ + 217, + 143, + 782, + 351 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/cc2555b02eeb3b544d9665aaacd802969d5e2faee812f8f6a206cb44092f53d0.jpg", + "table_caption": [ + "Table 1: Experimental results on the ScanRefer [6]. C, B-4, M, and R represent the captioning metrics CIDEr [32], BLEU-4 [25],METEOR [3],and ROUGE-L [13], respectively. A higher score for each indicates better performance." + ], + "table_footnote": [], + "table_body": "
ModelTrainingC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
Scan2Cap27.4717.2421.8049.06
D3Net33.8520.7023.1353.38
SpaCap3d33.7119.9222.6150.50
3DJCG38.0622.8223.7752.99
ContextualMLE35.2620.4222.7750.78
REMAN34.8120.3723.0150.99
Vote2Cap-DETR43.8426.6825.4154.43
Vote2Cap-DETR++47.0827.7025.4455.22
BiCA (Ours)48.7728.3525.6055.81
D3Net38.4222.2224.7454.37
\\( \\chi \\)-Tran2Cap33.6219.2922.2750.00
Contextual37.3720.9622.8951.11
Vote2Cap-DETRSCST45.5326.8825.4354.76
Vote2Cap-DETR++47.6228.4125.6354.77
BiCA (Ours)49.8128.8325.8556.46
", + "bbox": [ + 357, + 419, + 640, + 573 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Experimental results on the Nr3D [1] with IoU threshold at 0.5.", + "bbox": [ + 254, + 573, + 746, + 587 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ScanRefer. Descriptions in the ScanRefer include the target object's attributes and spatial relationships with surrounding objects. As Table 1 shows, our method outperforms existing methods in all data settings and IoU thresholds, thanks to our multi-stage contextual attention method.", + "bbox": [ + 215, + 626, + 782, + 686 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nr3D. The Nr3D dataset evaluates the model's proficiency in interpreting human-spoken, free-form object descriptions. Our method shows a notable performance improvement over existing models in generating diverse object descriptions, as indicated in Table 2.", + "bbox": [ + 215, + 705, + 782, + 766 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Study and Discussion", + "text_level": 1, + "bbox": [ + 217, + 787, + 514, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The core components of our method are i) decomposition of the query set into Instance Query $V^{o}$ and Context Query $V^{c}$ , ii) Contextual Attention of Objects", + "bbox": [ + 215, + 809, + 782, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bi-directional Contextual Attention for 3D Dense Captioning", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4f669c5ef821d8687d62d4a3bb1341fc273185e5f2d0b926497b86b3cb810420.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelIoU=0.50
C↑B-4↑M↑R↑mAP↑AR↑
Vote2Cap-DETR [9]73.7738.2126.6454.7145.5667.77
BiCA using only Vo74.9040.6726.7655.3150.1269.49
Vote2Cap-DETR++ [10]78.1639.7226.9455.5255.4870.89
BiCA using Vo, KNN(Vc)79.0341.3627.3156.7655.9569.62
BiCA using Vo, Voc81.2240.9027.3957.9556.9170.38
BiCA using Vo, Voc, Vac(Ours)85.1442.2727.9859.3757.5872.68
", + "bbox": [ + 328, + 143, + 671, + 234 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Ablation study on the ScanRefer [6]. The core components of our BiCA are i) decomposition of the query set into Instance Query $V^{o}$ and Context Query $V^{c}$ , ii) the Object-aware Context feature $V_{a}^{c}$ , and iii) the Context-aware object feature $V_{a}^{o}$ .", + "bbox": [ + 212, + 234, + 785, + 275 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "for Context (O4C) ii) Contextual Attention of Contexts for Object (C4O). We demonstrate that all components of BiCA contribute positively to the final performance, as shown in Table 3.", + "bbox": [ + 212, + 316, + 784, + 363 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Instance Query Generator. We define BiCA using only the Instance Query Generator (i.e., BiCA using only $V^o$ in Table 3) as our baseline and compare it with Vote2Cap-DETR [9], an object-centric transformer encoder-decoder architecture. The major difference between our instance query generator and Vote2Cap-DETR is how we generate the query set for instances. Vote2Cap-DETR uses farthest point sampling (FPS) to generate queries before the query coordinates are adjusted through voting. Therefore, if the coordinates are mistakenly focused on a specific object after voting, features will be extracted from the same object. On the other hand, our instance query generator extracts the features from the candidate coordinates after the voting. This improves the number of matching candidates (e.g., for 2,068 objects in our evaluation set, our method has 1,540 matching proposals while Vote2Cap-DETR has 1,498). This enhancement boosts localization performance in terms of mean Average Precision (mAP) and Average Recall (AR), which directly contributes to improving dense captioning performance.", + "bbox": [ + 212, + 386, + 787, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Context Query Generator. To enable object features to focus on localization, we independently generate context and instance queries separately. In the Vote2Cap-DETR++ [10], the decoupled queries are projections of the object-centric queries and still entail the limitations of the object-centric design. In contrast, our context queries capture structural information from the entire 3D scene, effectively decoupling features from object localization, as shown in Table 3.", + "bbox": [ + 212, + 635, + 787, + 741 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "O4C module and C4O module. As shown in Table 3, utilizing the object-aware context feature $V_{a}^{c}$ and the context-aware object feature $V_{a}^{o}$ results in performance improvements across all aspects. Interestingly, we can see that performance improves even when surrounding context information is collected using KNN (See the results of BiCA using only $V^{o}$ , $\\mathrm{KNN}(V^{c})$ ). In the setting of BiCA", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d89b17529005707034de99f445ea0d980b47749ab42601b918362ec739f11505.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 243, + 147, + 401, + 229 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2f44dc37f07297e8898de53a341b44c00926ac19a439a3f909c85ca5330f960d.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 433, + 143, + 589, + 227 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bb521a03de15b02d1d94f803e14e6f092e977293f2fe23b8177ca85f6f612178.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 620, + 143, + 776, + 227 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7381bf2a40eb5ea169ca39c9397c957cd0f775aadf9789d756def8261295aca1.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 245, + 247, + 398, + 329 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/02c2283338ad9cb745153a0d78b89b97f81021f5a15a82f5c4b75cb82bc13db1.jpg", + "image_caption": [ + "(e)", + "(f)" + ], + "image_footnote": [], + "bbox": [ + 433, + 246, + 589, + 327 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2460846e40025705019eb5297088e322ca40cb894f6c71d7d470c38f067bb1c7.jpg", + "image_caption": [ + "Fig. 3: Visualization of (a) the input 3D scene, (b) input point cloud without color, (c) visualization of instance queries, (d) one sampled instance query, (e) geometry contextualization of (d), and (f) instance contextualization of (e) on the ScanRefer [6]." + ], + "image_footnote": [], + "bbox": [ + 620, + 244, + 776, + 327 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "using only $V^o$ , $\\mathrm{KNN}(V^c)$ , we set $K = 16$ . Additionally, aggregating contextualized object features improves the model performance significantly. It proves that when capturing context information for captioning, adding additional object information that matches the context is helpful.", + "bbox": [ + 214, + 430, + 784, + 491 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 512, + 429, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 3 shows the parts of the entire scene being focused on at each stage of the proposed model. In Figure 3c and Figure 3d, our instance queries are not concentrated on a specific object but are spread throughout the 3D scene and exist in the location where the object exists. Figure 3e shows the part of the context queries with the highest attention score. The green part of the Figure 3e indicates the surroundings of the object and the corners of the room further away, but it needs to be related to other instances or scenes to express the situation more clearly. The blue part of Figure 3f describes the part of related instance queries for the object-aware context feature selected in Figure 3e. Most instances are inside the green contextualized region with high scores of Figure 3e, and some instances on the other side are also included. This observation shows that our method can check all instances of the entire scene to generate captions for one object.", + "bbox": [ + 212, + 537, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the proposed method, we provide qualitative results with state-of-the-art models. SpaCap3D [33], 3DJCG [4], Vote2CapDETR [9], and Vote2Cap-DETR++ [10] have attempted to incorporate contextual information by modeling relationships with object pairs or aggregating nearest neighbor features. In the Figure 4, these methods generate captions limited to the object and its immediate relations in a fixed format (e.g., this is a white radiator. it is under right of the desk). Since their context is not collected based", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bi-directional Contextual Attention for 3D Dense Captioning", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0798132423552dd668a0422bb5fe9827e8153269831d12d1bb5c1c2cec98f507.jpg", + "image_caption": [ + "Fig. 4: Qualitative results on the ScanRefer [6]. The yellow-highlighted sections show information specific to the object itself, the green-highlighted sections describe the relationships between objects, and the blue-highlighted sections depict the spatial position of the object in the 3D scene. Captions underlined in red indicate incorrect descriptions. FTG. represent failures in caption generation due to low IoU." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 759, + 390 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "on the entire scene, it is challenging to localize relationships with surrounding objects or locations in the scene. BiCA accurately predicts the attributes (i.e., can obtain precise representations for regions that have localization boundaries) and captions that require a sufficient understanding of structural information throughout the scene (i.e., can retrieve geometrically substantial information within the scene that exceeds the boundary of localization constraints) in various formats (e.g., the chair is the northwest - most one in the room.).", + "bbox": [ + 212, + 518, + 787, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 645, + 359, + 660 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we propose BiCA, a novel end-to-end transformer encoder-decoder pipeline with multi-stage contextual attention for 3D dense captioning. Our BiCA parallelly decodes a fixed set of object queries that contain local features of individual objects and context queries which contain the non-object contexts in the 3D scene. This allows our model to capture relevant object-aware context and context-aware object features across the entire scene without being restricted to single object localization or their immediate surroundings. As the representation for localization and caption generation is disentangled, BiCA can improve both localization and contextual dense captioning performance. We validate the effectiveness of BiCA by showing that it outperforms the state-of-the-art across all metrics on two benchmarks for 3D dense captioning.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by LG AI Research and Institute of Information & Communications Technology Planning & Evaluation (IITP) grant (No. RS-2019-II191082, No. RS-2022-II220156) funded by the Korea government (MSIT).", + "bbox": [ + 215, + 172, + 784, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 422-440. Springer (2020)", + "2. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018)", + "3. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization. pp. 65-72 (2005)", + "4. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16464-16473 (2022)", + "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "6. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: European conference on computer vision. pp. 202-221. Springer (2020)", + "7. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In: European Conference on Computer Vision. pp. 487-505. Springer (2022)", + "8. Chen, J., Lei, B., Song, Q., Ying, H., Chen, D.Z., Wu, J.: A hierarchical graph network for 3d object detection on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 392-401 (2020)", + "9. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11124-11133 (2023)", + "10. Chen, S., Zhu, H., Li, M., Chen, X., Guo, P., Lei, Y., Gang, Y., Li, T., Chen, T.: Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)", + "11. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3193-3203 (2021)", + "12. Chen, Z., Hu, R., Chen, X., Nießner, M., Chang, A.X.: Unit3d: A unified transformer for 3d dense captioning and visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18109-18119 (2023)" + ], + "bbox": [ + 218, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bi-directional Contextual Attention for 3D Dense Captioning", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Chin-Yew, L.: Rouge: A package for automatic evaluation of summaries. In: Proceedings of the Workshop on Text Summarization Branches Out, 2004 (2004)", + "14. Cornia, M., Stefanini, M., Baraldi, L., Cucchiara, R.: Meshed-memory transformer for image captioning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10578-10587 (2020)", + "15. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition", + "16. Ghandi, T., Pourreza, H., Mahyar, H.: Deep learning approaches on image captioning: A review. arXiv preprint arXiv:2201.12944 (2022)", + "17. Hossain, M.Z., Sohel, F., Shiratuddin, M.F., Laga, H.: A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR) 51(6), 1-36 (2019)", + "18. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition. pp. 4867-4876 (2020)", + "19. Jiao, Y., Chen, S., Jie, Z., Chen, J., Ma, L., Jiang, Y.G.: More: Multi-order relation mining for dense captioning in 3d scenes. In: European Conference on Computer Vision. pp. 528-545. Springer (2022)", + "20. Jin, Z., Hayat, M., Yang, Y., Guo, Y., Lei, Y.: Context-aware alignment and mutual masking for 3d-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10984-10994 (2023)", + "21. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)", + "22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "23. Mao, A., Yang, Z., Chen, W., Yi, R., Liu, Y.j.: Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE Transactions on Visualization and Computer Graphics (2023)", + "24. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2906-2917 (2021)", + "25. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002)", + "26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., Devito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems 32, pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf", + "27. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep hough voting for 3d object detection in point clouds. In: proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9277-9286 (2019)", + "28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "29. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Rennie, S.J., Marcheret, E., Mroueh, Y., Ross, J., Goel, V.: Self-critical sequence training for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7008-7024 (2017)", + "31. Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems 33, 7537-7547 (2020)", + "32. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4566-4575 (2015)", + "33. Wang, H., Zhang, C., Yu, J., Cai, W.: Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688 (2022)", + "34. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: International conference on machine learning. pp. 2048-2057. PMLR (2015)", + "35. Yang, X., Tang, K., Zhang, H., Cai, J.: Auto-encoding scene graphs for image captioning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10685-10694 (2019)", + "36. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8563-8573 (2022)", + "37. Zhong, Y., Xu, L., Luo, J., Ma, L.: Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 494 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bi-directional Contextual Attention for 3D Dense Captioning", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Bi-directional Contextual Attention for 3D Dense Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_model.json b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2099aa9296b212accd44bac69dc6c14cef0f7c4b --- /dev/null +++ b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_model.json @@ -0,0 +1,2203 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.278, + 0.141, + 0.726, + 0.187 + ], + "angle": 0, + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.212, + 0.709, + 0.245 + ], + "angle": 0, + "content": "Minjung Kim\\(^{1,4,*}\\), Hyung Suk Lim\\(^{3}\\), Soonyoung Lee\\(^{2}\\), Bumsoo Kim\\(^{2,\\dagger}\\), and Gunhee Kim\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.255, + 0.672, + 0.283 + ], + "angle": 0, + "content": "1 Seoul National University, Seoul, Korea minjung.kim@vision.snu.ac.kr, gunhee@snu.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.283, + 0.65, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) LG AI Research, Seoul, Korea {soonyoung.lee, bumsoo.kim}@lgresearch.ai" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.311, + 0.582, + 0.338 + ], + "angle": 0, + "content": "3 Dioxide, Seoul, Korea hslim@dioxide.com" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.339, + 0.653, + 0.351 + ], + "angle": 0, + "content": "4 SNU-LG AI Research Center, Seoul, Korea" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.379, + 0.744, + 0.726 + ], + "angle": 0, + "content": "Abstract. 3D dense captioning is a task involving the localization of objects and the generation of descriptions for each object in a 3D scene. Recent approaches have attempted to incorporate contextual information by modeling relationships with object pairs or aggregating the nearest neighbor features of an object. However, the contextual information constructed in these scenarios is limited in two aspects: first, objects have multiple positional relationships that exist across the entire global scene, not only near the object itself. Second, it faces with contradicting objectives—where localization and attribute descriptions are generated better with tight localization, while descriptions involving global positional relations are generated better with contextualized features of the global scene. To overcome this challenge, we introduce BiCA, a transformer encoder-decoder pipeline that engages in 3D dense captioning for each object with Bi-directional Contextual Attention. Leveraging parallelly decoded instance queries for objects and context queries for non-object contexts, BiCA generates object-aware contexts, where the contexts relevant to each object is summarized, and context-aware objects, where the objects relevant to the summarized object-aware contexts are aggregated. This extension relieves previous methods from the contradicting objectives, enhancing both localization performance and enabling the aggregation of contextual features throughout the global scene; thus improving caption generation performance simultaneously. Extensive experiments on two of the most widely-used 3D dense captioning datasets demonstrate that our proposed method achieves a significant improvement over prior methods." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.746, + 0.377, + 0.763 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.806 + ], + "angle": 0, + "content": "3D dense captioning is a task that requires 1) determining the location of all objects and 2) generating descriptive sentences for each object detected within a 3D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.558, + 0.825 + ], + "angle": 0, + "content": "* Work done during internship at LG AI Research" + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.826, + 0.388, + 0.84 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.558, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "M. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.784, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.277, + 0.788, + 0.307 + ], + "angle": 0, + "content": "Fig. 1: Conceptual illustration of the multi-stage pipeline of BiCA (best viewed in color)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.788, + 0.488 + ], + "angle": 0, + "content": "scene. Early approaches have utilized a two-stage process [4, 7, 11, 19, 33, 36, 37], where the objects are detected first, then captions are generated afterward by sequentially feeding the features of each detected object. Following work [9] has borrowed the end-to-end transformer encoder-decoder pipeline from object detection [5] to actively involve contextual information for 3D dense captioning. In these works, contextual information for predicting relationships is composed by combining object features (e.g., with dot product [4, 33] or transformer attention [9]) and aggregating local nearest neighbor visual features [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.687 + ], + "angle": 0, + "content": "Despite their improved performance, the scope of contextual information in previous 3D dense captioning methods poses its own set of challenges. First, 3D dense captioning heavily involves various positional relationships throughout the global scene; thus, the context scope restricted to the combination of individual objects or their spatially nearest neighbors typically cannot provide sufficient information. Next, the endeavor to incorporate the contextual features of various positions into a single object encounters a conflicting objective since the object must encompass precise local features for localization while simultaneously capturing various positional relationships throughout the global context for caption generation. Clearly, these challenges limit the performance of 3D dense captioning, which raises an interesting research question—“Can we design an architecture for 3D dense captioning that can effectively aggregate relevant context features without harming localization performance?”" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To this end, we propose a Bi-directional Contextual Attention (BiCA) network for 3D dense captioning. BiCA parallelly decodes a fixed set of object queries and context queries sampled from the geometric positions that do not overlap with each other. While the decoded object query performs standard object detection, the context queries are designed to distinctively capture non-object contexts and subsequently aggregate relevant contextual information for each object. Recognizing that non-object regions often lack sufficient visual features in 3D point clouds, BiCA incorporates a multi-stage pipeline where context features are enriched by relevant object features, thus implementing bidirectional attention: i) Contextual Attention of Objects for Context (O4C):" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "context representation is represented by the attention summarization of global objects. ii) Contextual Attention of Contexts for Object (C4O): the object representation is adjusted by summarizing the non-object contexts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.194, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Then, localization is inferred with object queries while the captions are generated with object-aware context and context-aware object features. By disentangling the attention for localization and the attention for incorporating contextual features, BiCA successfully resolves the aforementioned contradicting objectives for local and global features. Moreover, since the contextual attentions retrieve globally relevant contexts and objects throughout the scene, it could accurately produce descriptions that transcend the spatial boundary of localization and its nearest neighbor. The conceptual overview of our BiCA is illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.787, + 0.392 + ], + "angle": 0, + "content": "BiCA sets a new state-of-the-art standard for 3D dense captioning, while simultaneously improving 3D object detection performance. Extensive experiments on two widely used benchmarks in 3D dense captioning (i.e., ScanRefer [6] and Nr3D [1]) show that our proposed BiCA surpasses prior approaches by a large margin. The contribution of our paper can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.406, + 0.784, + 0.467 + ], + "angle": 0, + "content": "- We propose a novel Bi-directional Contextual Attention network with multi-stage contextual attention for 3D dense captioning. This enables our model to capture relevant contexts throughout the global scene without being bound to single-object localization or their nearest neighbors." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.529 + ], + "angle": 0, + "content": "- By performing localization with decoded object queries and caption generation by including object-aware context features and context-aware object correspondence features, BiCA can simultaneously improve the performance of localization and caption generation for 3D dense captioning." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.53, + 0.785, + 0.575 + ], + "angle": 0, + "content": "- Our BiCA achieves state-of-the-art performances across multiple evaluation metrics on two widely used benchmarks for 3D dense captioning: ScanRefer and Nr3D datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.605, + 0.383, + 0.62 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.643, + 0.411, + 0.658 + ], + "angle": 0, + "content": "2.1 Image Captioning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Image captioning is a fundamental task in visual language creation that automatically generates descriptive sentences for images. The main goal is to improve 2D scene understanding and generate captions that accurately reflect the content and context of the image [17]. However, with the advancement of deep learning in image captioning, there has been a marked shift towards the development and adoption of more sophisticated models [16]. Image captioning methods using deep learning adopt an encoder-decoder architecture, where the decoder generates a sentence from the visual features extracted by the encoder. Attention-based methods for grid regions [34] and detected objects [2] focus on specific image regions and use graph neural networks [35] or transformer layers to capture relationships between objects [14]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "M. Kim et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.441, + 0.163 + ], + "angle": 0, + "content": "2.2 3D Dense Captioning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.265 + ], + "angle": 0, + "content": "3D dense captioning, an emerging field focused on achieving detailed object-level understanding of 3D scenes through natural language descriptions, has garnered significant interest in recent years. This task involves converting 3D visual data [15] into a consistent set of bounding boxes and generating appropriate natural language descriptions for each identified object. This task presents a considerable challenge due to the complexity of 3D scene information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.789, + 0.493 + ], + "angle": 0, + "content": "Most 3D dense captioning models employ an encoder-decoder architecture comprising three main components: a scene encoder, a relational module, and a feature decoder. The scene encoder in 3D dense captioning utilizes 3D object detection methods such as 3DETR [24], PointNet++ [28], VoteNet [27], and PointGroup [18] to extract object-level visual features and contextual details from 3D point cloud datasets. The relational module is crucial in 3D dense captioning, modeling complex connections and cross-modal interactions among objects within indoor scenes. Relational modeling approaches vary based on task requirements and can be categorized into graph-based [7, 11, 19], transformer-based [4, 9, 33], and knowledge distillation-based [36] methods. The feature decoder generates bounding boxes and captions for candidate objects by considering attributes and relational features. GRU-based decoders with attention mechanisms are used in Scan2Cap [11], MORE [19], and D3Net [7], while transformer-based decoders are employed in SpaCap3D [33], \\(\\chi\\)-Tran2Cap [36], 3DJCG [4], and Vote2Cap-DETR [9] to facilitate the caption generation process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.6 + ], + "angle": 0, + "content": "However, most existing methods use a single query set for object localization and caption generation, which often leads to conflicts in achieving the objectives of the two tasks. To address this, Vote2Cap-DETR++ [10] enhances task-specific feature capture by decoupling queries for localization and captioning, overcoming the limitations of a unified query system. Despite this improvement, challenges persist as its performance remains fundamentally bound by the precision of its object localization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.332, + 0.641 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.748 + ], + "angle": 0, + "content": "Our goal is to take a 3D scene as input, identify all the objects within it, and generate captions for each object. Previous research has adopted a pipeline to identify objects in a 3D scene and simultaneously generate captions from the same features. However, while each caption may describe features of the object itself, it can also be based on the relationships with surrounding objects or information about the overall scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This section introduces a novel transformer encoder-decoder approach with bi-directional contextual attention for the 3D dense captioning task. This approach distinguishes between features for instances and context information between instances, indirectly supervising the context features to enable end-to-end learning with the instance features. Instance queries aim to localize instances within a 3D scene and capture the characteristics of instances. In contrast, each" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.782, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.789, + 0.419 + ], + "angle": 0, + "content": "Fig. 2: The overall pipeline of BiCA. We parallelly generate and decode two sets of queries (i.e., Instance Query and Context Query) that encodes the instance features and the non-object context features throughout the global scene, respectively. The object-aware contexts are calculated per each object by the weighted sum of the context queries, where the weights are calculated by the attention between the decoded instance query and context query. Then, with the object-aware context feature, the context-aware object feature is obtained by the weighted sum of the instances, which is weighted by the attention between the object-aware contexts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.518 + ], + "angle": 0, + "content": "context query is designed to capture the non-object region within a scene, including positional relationships between instances in the 3D scene. The overall architecture of the newly proposed model is described in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.54, + 0.332, + 0.554 + ], + "angle": 0, + "content": "3.1 Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Following previous work [9], our model also adopts 3DETR [24] encoder as our scene encoder. This encoder is applied to transform the input point cloud into a set of encoded tokens that capture spatial, structural, and contextual information. Given the input point cloud \\(PC = [p_{\\mathrm{in}};f_{\\mathrm{in}}]\\in \\mathbb{R}^{N\\times (3 + F)}\\), it is initially tokenized by a set-abstraction layer of PointNet++ [28]. This tokenized output is subsequently fed into a masked transformer encoder incorporating the set-abstraction layer, followed by two additional encoder layers. The final encoded scene tokens are denoted as \\(p_{enc}\\in \\mathbb{R}^{1,024\\times 3}\\) and \\(f_{enc}\\in \\mathbb{R}^{1,024\\times 256}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.406, + 0.725 + ], + "angle": 0, + "content": "3.2 Query Generator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To separate captions bound to a single object from captions containing information relative to other objects or the global scene, we specify two separate Instance Query and Context Query from the encoded scene. While the instance query set is designed to capture individual features of objects bounded by their localization, the context query set is designed to encompass a broad area that consists of the positional information of a caption (e.g., next to, on the right of, is most far from, etc.)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "M. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.79, + 0.193 + ], + "angle": 0, + "content": "Instance Query Generator. In the Instance Query Generator, instance queries are generated to detect all objects in the 3D scene and generate captions for each object's individual attributes. The instance query \\((p^o,f^o)\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.206, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\n[ \\Delta p _ {\\text {v o t e}}; \\Delta f _ {\\text {v o t e}} ] = \\mathrm {F F N} _ {o} (f _ {\\text {e n c}}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.248, + 0.786, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\left(p ^ {o}, f ^ {o}\\right) = \\mathrm {S A} _ {o} \\left(p _ {\\text {e n c}} + \\Delta p _ {\\text {v o t e}}, f _ {\\text {e n c}} + \\Delta f _ {\\text {v o t e}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.789, + 0.395 + ], + "angle": 0, + "content": "where \\([\\varDelta p_{\\mathrm{vote}};\\varDelta f_{\\mathrm{vote}}]\\in\\mathbb{R}^{1,024\\times(3+256)}\\) is an offset that learns to shift the encoded points to object's centers spatially by a feed-forward network \\(\\mathrm{FFN}_o\\), following [9]. \\(\\mathrm{SA}_o\\) represents the set-abstraction layer with a radius of 0.3 and samples 16 points for \\(p^o\\). Since our Instance Query Generator extracts the features from the candidate coordinates after the voting, our instance queries are not focused on specific objects. However, they are distributed across the locations where objects are present throughout the 3D scene. As a result, 256 instance queries are extracted from our Instance Query Generator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.789, + 0.629 + ], + "angle": 0, + "content": "Context Query Generator. 3D scenes are composed of 3D points arranged in XYZ coordinates, forming specific objects, and the spaces between objects signify space. We recognize that obtaining context features should not merely involve interpolating the spaces between Instance Queries but also encompassing how these points are arranged and the composition of empty spaces to represent a scene. Inspired by this observation, we define context information as the geometric details in a 3D scene that illustrate the relationships between objects and between objects and the scene itself. We decide to extract structural information centered on specific points in the 3D scene and call them Context Queries. This approach is distinct from the Instance Query Generator. While the Instance Query Generator uses a voting network to obtain object center coordinates that do not exist in the 3D scene and extracts features such as identification from them, the proposed Context Query Generator extracts structural information without disrupting the arrangement of the 3D points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.675 + ], + "angle": 0, + "content": "Given the encoded scene tokens \\((p_{\\mathrm{enc}}, f_{\\mathrm{enc}})\\), we sample 512 points \\(p_{\\mathrm{seed}}^c\\) with farthest point sampling (FPS) [28] on \\(p_{\\mathrm{enc}}\\) that each represent a unique geometry within the global scene. The context query \\((p^c, f^c)\\) is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.689, + 0.786, + 0.705 + ], + "angle": 0, + "content": "\\[\n(p ^ {c}, f ^ {c}) = \\mathrm {S A} _ {c} (p _ {\\mathrm {e n c}}, f _ {\\mathrm {e n c}}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.747 + ], + "angle": 0, + "content": "where \\(\\mathrm{SA}_c\\) denotes the set-abstraction layer [28] with a radius of 1.2, sampling 64 points for \\(p^c\\). All hyper-parameters are determined experimentally." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.334, + 0.784 + ], + "angle": 0, + "content": "3.3 Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Instance Decoder and Context Decoder. Given the context queries \\((p^c, f^c)\\) and instance queries \\((p^o, f^o)\\), we build a parallel decoding pipeline where the Context Decoder designates contextual information between objects and the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Instance Decoder performs the localization and attribute description for the participating objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Each Instance Decoder and Context Decoder consists of a transformer decoder with 8 layers according to 3DETR [24], and Fourier positional encoding is applied to XYZ coordinates. Fourier positional encoding transforms the XYZ coordinates of each query into a Fourier position embedding [31] and then combines them with the corresponding query embeddings. This encoding transforms the input data into frequency functions, capturing invariant changes in complex shapes and structures. It aids the context queries in accurately identifying geometric shapes and the instance queries in detecting objects. The context queries \\( V^{c} \\) are afterwards contextualized for each instance query \\( V^{o} \\), then retrieve relevant instance query features that are structurally involved." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.35, + 0.557, + 0.365 + ], + "angle": 0, + "content": "3.4 Bi-directional Contextual Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Captions must be generated for each object in relation to the scene context, and through positional encoding, it becomes possible to understand directional relationships, such as \"to the right\" with surrounding objects. Here, we gain the insight that information about the relative objects closely related to these relationships is also necessary, and we design the model to proceed simultaneously with a multi-stage contextualization. With the decoded queries, our method proceeds in two stages: i) Contextual Attention of Objects for Context (O4C): context representation is represented by the attention summarization of global objects. ii) Contextual Attention of Contexts for Object (C4O): the object representation is adjusted by summarizing the non-object contexts. Figure 1 and Figure 2 show detailed illustrations of obtaining the object-aware context feature and the context-aware object feature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Contextual Attention of Objects for Context (O4C). To retrieve the structural relationship between objects throughout the global scene, we first construct a Object-aware Context \\( V_{a}^{c} \\) per object by summating the attention of each context position feature with regard to each object. The attention result between the instance feature and the context feature is applied to the context feature, making a weighted summarization of the geometries. At this time, the ratio is adjusted using learnable gamma. We name this result \\( V_{a}^{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Contextual Attention of Contexts for Object (C4O). Most contexts not only interact with the reference object but also have other objects that react against them. To achieve this, we go through a Contextual Attention of Contexts for Object process. This could specify the information that can simply be said to be 'next to' as 'next to the red chair'. To this end, the attention result of the context feature and instance feature with the attention weight applied is applied to the instance query feature. Likewise, it is adjusted using a learnable parameter lambda. We name this result \\( V_{a}^{o} \\). \\( V_{a}^{o} \\) results from finding instance combinations that have a meaningful relationship with the contextual region focused on a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "M. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "specific instance. The concatenation of the instance token \\( V^o \\), contextualized geometry \\( V_a^c \\), and contextualized object \\( V_a^o \\), denoted as \\( V^a \\), is fed to the caption head to generate final captions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.215, + 0.441, + 0.232 + ], + "angle": 0, + "content": "3.5 3D Dense Captioning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.786, + 0.287 + ], + "angle": 0, + "content": "Our final goal is to identify all objects within the input 3D scene and generate a caption for each object. We perform object detection and caption generation in parallel and group the results into one instance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.787, + 0.37 + ], + "angle": 0, + "content": "Localization. To localize instances in the 3D scene, we use the decoded instance query \\( V^{o} \\). Through 5 MLP heads, we reformulate the box corner estimation as offset estimation from a query point to an object's center and box size regression. These localization heads are shared across the decoder layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Caption Generation. For caption generation, we use a transformer decoder-based caption head based on GPT-2 [29], following the methods of Vote2CapDETR [9] and SpaCap3D [33]. Our caption head comprises two transformer decoder blocks, sinusoid positional embedding, and a linear classification layer. When generating captions for a proposal, we substitute the standard Start Of Sequence ('SOS') prefix with \\( V^a \\). During the inference process, our approach employs beam search to produce captions. The beam size is 5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.521, + 0.386, + 0.537 + ], + "angle": 0, + "content": "3.6 Training BiCA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.787, + 0.607 + ], + "angle": 0, + "content": "Instance Query Loss. To train instance query generator to find an object's center by shifting points \\( p_{enc} \\), we use the vote loss from VoteNet [27]. Given the generated instance query \\( (p^o, f^o) \\) and the encoded scene tokens \\( (p_{enc}, f_{enc}) \\), the vote loss \\( \\mathcal{L}^o \\) is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.617, + 0.786, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {o} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {N _ {g t}} \\| p _ {i} ^ {o} - \\operatorname {c n t} _ {j} \\| _ {1} \\cdot \\mathbb {I} \\left(p _ {e n c} ^ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.671, + 0.787, + 0.72 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}(x)\\) is an indicator function that equals 1 when \\(x \\in I_j\\) and 0 otherwise, \\(N_{gt}\\) is the number of instances in the 3D scene, \\(M\\) is the number of \\(p^o\\), and \\(\\mathrm{cnt}_j\\) is the center of \\(j\\)-th instance \\(I_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.739, + 0.787, + 0.771 + ], + "angle": 0, + "content": "Detection Loss. Following DETR [5], we use Hungarian matching [21] to assign each proposal with the ground truth. The detection loss \\(\\mathcal{L}_{\\mathrm{det}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.784, + 0.786, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d e t}} = \\alpha_ {1} \\mathcal {L} _ {\\mathrm {g i o u}} + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {c l s}} + \\alpha_ {3} \\mathcal {L} _ {\\mathrm {c n t}} + \\alpha_ {4} \\mathcal {L} _ {\\mathrm {s i z e}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "with \\(\\alpha_{1} = 10, \\alpha_{2} = 1, \\alpha_{3} = 5, \\alpha_{4} = 1\\) set heuristically. This loss is applied across all decoder layers for better convergence." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.225 + ], + "angle": 0, + "content": "Caption Loss. Following the standard image captioning protocol, we first train the caption heads using cross-entropy loss for Maximum Likelihood Estimation (MLE). During MLE training, the model predicts the \\((t + 1)\\)-th word \\(c_{n}^{t + 1}\\) based on the first \\(t\\) words \\(c_{n}^{[1:t]}\\) and the visual features \\(\\mathcal{V}\\). The loss function for a final caption of length \\(T\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.239, + 0.785, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c _ {n}} = \\sum_ {t = 1} ^ {T} \\mathcal {L} _ {c _ {n}} (t) = - \\sum_ {t = 1} ^ {T} \\log \\hat {P} \\left(c _ {n} ^ {t + 1} | \\mathcal {V}, c _ {n} ^ {[ 1: t ]}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.291, + 0.784, + 0.351 + ], + "angle": 0, + "content": "After word-level supervision, the caption head is refined using Self-Critical Sequence Training (SCST) [30], where the model generates multiple captions \\(\\hat{c}_{1,\\dots,k}\\) with a beam size of \\(k\\) and an additional caption \\(\\hat{g}\\) via greedy search. The loss function for SCST is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.363, + 0.785, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c _ {n}} = - \\sum_ {i = 1} ^ {k} \\left(R \\left(\\hat {c} _ {i}\\right) - R (\\hat {g})\\right) \\cdot \\frac {1}{\\left| \\hat {c} _ {i} \\right|} \\log \\hat {P} \\left(\\hat {c} _ {i} \\mid \\mathcal {V}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.416, + 0.784, + 0.462 + ], + "angle": 0, + "content": "The reward function \\( R(\\cdot) \\) is based on the CIDEr [32] metric for caption evaluation, and the logarithmic probability of the caption \\( \\hat{c}_i \\) is normalized by its length \\( |\\hat{c}_i| \\), ensuring equal importance to captions of different lengths." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.487, + 0.784, + 0.533 + ], + "angle": 0, + "content": "Final Loss. Given the instance query Loss \\(\\mathcal{L}^o\\), the detection loss for the \\(i\\)-th decoder layer as \\(\\mathcal{L}_{\\mathrm{det}}^i\\), and the average of the caption loss \\(\\mathcal{L}_{c_n}\\) within a batch, denoted as \\(\\mathcal{L}_{\\mathrm{cap}}\\), the final loss \\(\\mathcal{L}\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.547, + 0.785, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\beta_ {1} \\mathcal {L} ^ {o} + \\beta_ {2} \\sum_ {i = 1} ^ {n _ {\\mathrm {d e c - l a y e r}}} \\mathcal {L} _ {\\mathrm {d e t}} ^ {i} + \\beta_ {3} \\mathcal {L} _ {\\mathrm {c a p}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.598, + 0.47, + 0.614 + ], + "angle": 0, + "content": "where \\(\\beta_{1} = 10\\), \\(\\beta_{2} = 1\\), and \\(\\beta_{3} = 5\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.64, + 0.376, + 0.657 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.676, + 0.438, + 0.69 + ], + "angle": 0, + "content": "4.1 Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Datasets. We focus on 3D dense captioning, leveraging two benchmark datasets: ScanRefer [6] and \\(\\mathrm{Nr3D}\\) [1]. These datasets offer an extensive human-generated description of 3D scenes and objects. ScanRefer encompasses 36,665 descriptions covering 7,875 objects within 562 scenes, while \\(\\mathrm{Nr3D}\\) contains 32,919 descriptions for 4,664 objects across 511 scenes. Both datasets draw their training data from the ScanNet [15] database, which includes 1,201 3D scenes. For evaluation, we use 9,508 descriptions for 2,068 objects across 141 scenes from ScanRefer and 8,584 descriptions for 1,214 objects in 130 scenes from \\(\\mathrm{Nr3D}\\), all sourced from the 312 3D scenes in the ScanNet validation set." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "M. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.27 + ], + "angle": 0, + "content": "Metrics. We evaluate model performance using four metrics: CIDEr [32], BLEU-4 [25],METEOR [3], and ROUGE-L [13], denoted as \\(\\mathbf{C}\\), B-4, \\(\\mathbf{M}\\), and \\(\\mathbf{R}\\), respectively. Following the previous studies [4,9,11,19,33], Non-Maximum Suppression (NMS) is initially applied to filter out duplicate object predictions among the proposals. Each proposal is represented as a pair consisting of a predicted bounding box \\(\\hat{b} i\\) and its associated caption \\(\\hat{c} i\\). To accurately evaluate the model's capacity for object localization and caption generation, we employ the metric \\(m@k\\), setting the IoU thresholds at 0.25 and 0.5 for our experiments, following [11]:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.284, + 0.786, + 0.325 + ], + "angle": 0, + "content": "\\[\nm @ k = \\frac {1}{N} \\sum_ {i = 1} ^ {N} m \\left(\\hat {c} _ {i}, C _ {i}\\right) \\cdot \\mathbb {I} \\left\\{\\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {i}\\right) \\geq k \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.337, + 0.785, + 0.368 + ], + "angle": 0, + "content": "where \\(N\\) denotes the total number of annotated objects in the evaluation set, and \\(m\\) stands for the captioning metrics C, B-4, M, and R." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.398, + 0.457, + 0.413 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.658 + ], + "angle": 0, + "content": "Following the [9], our training comprises three stages. We pre-train our network without the caption head on the ScanNet dataset [15] for 1,080 epochs, using a batch size of 8. We minimize the loss function with an AdamW optimizer [22], starting with a learning rate of \\(5 \\times 10^{-4}\\) that reduces to \\(10^{-6}\\) using a cosine annealing schedule, along with a weight decay of 0.1 and gradient clipping at 0.1 as per [24]. We then proceed to jointly train the model using standard cross-entropy loss for 720 epochs on both ScanRefer [6] and Nr3D [1], maintaining the detector's learning rate at \\(10^{-6}\\) and reducing the caption head from \\(10^{-4}\\) to \\(10^{-6}\\) to avoid overfitting. In the SCST [30], we adjust the caption head using a batch size of 2 while keeping the detector fixed over a span of 180 epochs and maintain a constant learning rate of \\(10^{-6}\\). Additionally, for experiments incorporating 2D data, we use the pre-trained ENet [8] to extract 128-dimensional multi-view features from ScanNet images as outlined in Scan2Cap [11]. The model has 16.9M parameters, and the average inference time on ScanRefer [6] is 1.8ms. All experiments are conducted on a single Titan RTX GPU using PyTorch [26]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.687, + 0.557, + 0.703 + ], + "angle": 0, + "content": "4.3 Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this section, we evaluate our performance against state-of-the-art methods: Scan2Cap [11], D3Net [7], SpaCap3D [33], MORE [19], 3DJCG [4], Contextual [37], REMAN [23], 3D-VLP [20], \\(\\chi\\)-Tran2Cap [36], Vote2Cap-DETR [9], Unit3D [12], and Vote2Cap-DETR++ [10]. We apply IoU thresholds of 0.25 and 0.5 for ScanRefer [6] as shown in Table 1 and an IoU threshold of 0.5 for Nr3D [1] indicated in Table 2. For the baselines, we present the evaluation results reported in the original papers, and \"-\" in Table 1 and Table 2 indicates that such results have not been reported in the original paper." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.352 + ], + "angle": 0, + "content": "
ModelTrainingw/o additional 2D dataw/ additional 2D data
IoU=0.25IoU=0.50IoU=0.25IoU=0.50
C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑
Scan2Cap53.7334.2526.1454.9535.2022.3621.4443.5756.8234.1826.2955.2739.0823.3221.9744.78
D3Net------------46.0730.2924.3551.67
SpaCap3d58.0635.3026.1655.0342.7625.3822.8445.6663.3036.4626.7155.7144.0225.2622.3345.36
MORE58.8935.4126.3655.4138.9823.0121.6544.3362.9136.2526.7556.3340.9422.9321.6644.42
3DJCG60.8639.6727.4559.0247.6831.5324.2851.8064.7040.1727.6659.2349.4831.0324.2250.80
Contextual----42.7723.6022.0545.13----46.1125.4722.6445.96
REMANMLE--------62.0136.3727.7656.2545.0026.3122.6746.96
3D-VLP64.0939.8427.6558.7850.0231.8724.5351.1770.7341.0328.1459.7254.9432.3124.8351.51
Vote2Cap-DETR71.4539.3428.2559.3361.8134.4626.2254.4072.7939.1728.0659.2359.3232.4225.2852.38
Unit3D------------46.6927.2221.9145.98
Vote2Cap-DETR++76.3641.3728.7060.0067.5837.0526.8955.6477.0340.9928.5359.5964.3234.7326.0453.67
BiCA (Ours)78.4241.4628.8260.0268.4638.2327.5658.5678.3541.2028.8259.8066.4736.1326.7154.54
Scan2Cap------------48.3826.0922.1544.74
D3Net------------62.6435.6825.7253.90
x-Tran2Cap58.8134.1725.8154.1041.5223.8321.9044.9761.8335.6526.6154.7043.8725.0522.4645.28
ContextualSCST----50.2925.6422.5744.71----54.3027.2423.3045.81
Vote2Cap-DETR84.1542.5128.4759.2673.7738.2126.6454.7186.2842.6428.2759.0770.6335.6925.5152.28
Vote2Cap-DETR++88.2844.0728.7559.8978.1639.7226.9455.5288.5643.3028.6459.1974.4437.1826.2053.30
BiCA (Ours)89.7244.9728.9660.6980.1440.1627.7656.1089.3444.5628.7459.3376.3437.3426.6054.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.354, + 0.784, + 0.395 + ], + "angle": 0, + "content": "Table 1: Experimental results on the ScanRefer [6]. C, B-4, M, and R represent the captioning metrics CIDEr [32], BLEU-4 [25],METEOR [3],and ROUGE-L [13], respectively. A higher score for each indicates better performance." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.42, + 0.642, + 0.574 + ], + "angle": 0, + "content": "
ModelTrainingC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
Scan2Cap27.4717.2421.8049.06
D3Net33.8520.7023.1353.38
SpaCap3d33.7119.9222.6150.50
3DJCG38.0622.8223.7752.99
ContextualMLE35.2620.4222.7750.78
REMAN34.8120.3723.0150.99
Vote2Cap-DETR43.8426.6825.4154.43
Vote2Cap-DETR++47.0827.7025.4455.22
BiCA (Ours)48.7728.3525.6055.81
D3Net38.4222.2224.7454.37
\\( \\chi \\)-Tran2Cap33.6219.2922.2750.00
Contextual37.3720.9622.8951.11
Vote2Cap-DETRSCST45.5326.8825.4354.76
Vote2Cap-DETR++47.6228.4125.6354.77
BiCA (Ours)49.8128.8325.8556.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.574, + 0.747, + 0.588 + ], + "angle": 0, + "content": "Table 2: Experimental results on the Nr3D [1] with IoU threshold at 0.5." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.627, + 0.784, + 0.687 + ], + "angle": 0, + "content": "ScanRefer. Descriptions in the ScanRefer include the target object's attributes and spatial relationships with surrounding objects. As Table 1 shows, our method outperforms existing methods in all data settings and IoU thresholds, thanks to our multi-stage contextual attention method." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.707, + 0.784, + 0.767 + ], + "angle": 0, + "content": "Nr3D. The Nr3D dataset evaluates the model's proficiency in interpreting human-spoken, free-form object descriptions. Our method shows a notable performance improvement over existing models in generating diverse object descriptions, as indicated in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.788, + 0.516, + 0.803 + ], + "angle": 0, + "content": "4.4 Ablation Study and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The core components of our method are i) decomposition of the query set into Instance Query \\( V^{o} \\) and Context Query \\( V^{c} \\), ii) Contextual Attention of Objects" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "M. Kim et al." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.144, + 0.672, + 0.235 + ], + "angle": 0, + "content": "
ModelIoU=0.50
C↑B-4↑M↑R↑mAP↑AR↑
Vote2Cap-DETR [9]73.7738.2126.6454.7145.5667.77
BiCA using only Vo74.9040.6726.7655.3150.1269.49
Vote2Cap-DETR++ [10]78.1639.7226.9455.5255.4870.89
BiCA using Vo, KNN(Vc)79.0341.3627.3156.7655.9569.62
BiCA using Vo, Voc81.2240.9027.3957.9556.9170.38
BiCA using Vo, Voc, Vac(Ours)85.1442.2727.9859.3757.5872.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.235, + 0.787, + 0.276 + ], + "angle": 0, + "content": "Table 3: Ablation study on the ScanRefer [6]. The core components of our BiCA are i) decomposition of the query set into Instance Query \\( V^{o} \\) and Context Query \\( V^{c} \\), ii) the Object-aware Context feature \\( V_{a}^{c} \\), and iii) the Context-aware object feature \\( V_{a}^{o} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.785, + 0.364 + ], + "angle": 0, + "content": "for Context (O4C) ii) Contextual Attention of Contexts for Object (C4O). We demonstrate that all components of BiCA contribute positively to the final performance, as shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Instance Query Generator. We define BiCA using only the Instance Query Generator (i.e., BiCA using only \\( V^o \\) in Table 3) as our baseline and compare it with Vote2Cap-DETR [9], an object-centric transformer encoder-decoder architecture. The major difference between our instance query generator and Vote2Cap-DETR is how we generate the query set for instances. Vote2Cap-DETR uses farthest point sampling (FPS) to generate queries before the query coordinates are adjusted through voting. Therefore, if the coordinates are mistakenly focused on a specific object after voting, features will be extracted from the same object. On the other hand, our instance query generator extracts the features from the candidate coordinates after the voting. This improves the number of matching candidates (e.g., for 2,068 objects in our evaluation set, our method has 1,540 matching proposals while Vote2Cap-DETR has 1,498). This enhancement boosts localization performance in terms of mean Average Precision (mAP) and Average Recall (AR), which directly contributes to improving dense captioning performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.742 + ], + "angle": 0, + "content": "Context Query Generator. To enable object features to focus on localization, we independently generate context and instance queries separately. In the Vote2Cap-DETR++ [10], the decoupled queries are projections of the object-centric queries and still entail the limitations of the object-centric design. In contrast, our context queries capture structural information from the entire 3D scene, effectively decoupling features from object localization, as shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "O4C module and C4O module. As shown in Table 3, utilizing the object-aware context feature \\( V_{a}^{c} \\) and the context-aware object feature \\( V_{a}^{o} \\) results in performance improvements across all aspects. Interestingly, we can see that performance improves even when surrounding context information is collected using KNN (See the results of BiCA using only \\( V^{o} \\), \\( \\mathrm{KNN}(V^{c}) \\)). In the setting of BiCA" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.147, + 0.24, + 0.16 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.148, + 0.403, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.147, + 0.428, + 0.16 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.145, + 0.591, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.148, + 0.614, + 0.16 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.145, + 0.777, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.249, + 0.24, + 0.261 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.248, + 0.4, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.25, + 0.428, + 0.261 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.247, + 0.591, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.25, + 0.615, + 0.262 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.246, + 0.777, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.359, + 0.785, + 0.402 + ], + "angle": 0, + "content": "Fig. 3: Visualization of (a) the input 3D scene, (b) input point cloud without color, (c) visualization of instance queries, (d) one sampled instance query, (e) geometry contextualization of (d), and (f) instance contextualization of (e) on the ScanRefer [6]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.431, + 0.785, + 0.492 + ], + "angle": 0, + "content": "using only \\( V^o \\), \\( \\mathrm{KNN}(V^c) \\), we set \\( K = 16 \\). Additionally, aggregating contextualized object features improves the model performance significantly. It proves that when capturing context information for captioning, adding additional object information that matches the context is helpful." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.43, + 0.529 + ], + "angle": 0, + "content": "4.5 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Figure 3 shows the parts of the entire scene being focused on at each stage of the proposed model. In Figure 3c and Figure 3d, our instance queries are not concentrated on a specific object but are spread throughout the 3D scene and exist in the location where the object exists. Figure 3e shows the part of the context queries with the highest attention score. The green part of the Figure 3e indicates the surroundings of the object and the corners of the room further away, but it needs to be related to other instances or scenes to express the situation more clearly. The blue part of Figure 3f describes the part of related instance queries for the object-aware context feature selected in Figure 3e. Most instances are inside the green contextualized region with high scores of Figure 3e, and some instances on the other side are also included. This observation shows that our method can check all instances of the entire scene to generate captions for one object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the proposed method, we provide qualitative results with state-of-the-art models. SpaCap3D [33], 3DJCG [4], Vote2CapDETR [9], and Vote2Cap-DETR++ [10] have attempted to incorporate contextual information by modeling relationships with object pairs or aggregating nearest neighbor features. In the Figure 4, these methods generate captions limited to the object and its immediate relations in a fixed format (e.g., this is a white radiator. it is under right of the desk). Since their context is not collected based" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "M. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.761, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.471 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results on the ScanRefer [6]. The yellow-highlighted sections show information specific to the object itself, the green-highlighted sections describe the relationships between objects, and the blue-highlighted sections depict the spatial position of the object in the 3D scene. Captions underlined in red indicate incorrect descriptions. FTG. represent failures in caption generation due to low IoU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.626 + ], + "angle": 0, + "content": "on the entire scene, it is challenging to localize relationships with surrounding objects or locations in the scene. BiCA accurately predicts the attributes (i.e., can obtain precise representations for regions that have localization boundaries) and captions that require a sufficient understanding of structural information throughout the scene (i.e., can retrieve geometrically substantial information within the scene that exceeds the boundary of localization constraints) in various formats (e.g., the chair is the northwest - most one in the room.)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.36, + 0.661 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this work, we propose BiCA, a novel end-to-end transformer encoder-decoder pipeline with multi-stage contextual attention for 3D dense captioning. Our BiCA parallelly decodes a fixed set of object queries that contain local features of individual objects and context queries which contain the non-object contexts in the 3D scene. This allows our model to capture relevant object-aware context and context-aware object features across the entire scene without being restricted to single object localization or their immediate surroundings. As the representation for localization and caption generation is disentangled, BiCA can improve both localization and contextual dense captioning performance. We validate the effectiveness of BiCA by showing that it outperforms the state-of-the-art across all metrics on two benchmarks for 3D dense captioning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.785, + 0.22 + ], + "angle": 0, + "content": "This work was supported by LG AI Research and Institute of Information & Communications Technology Planning & Evaluation (IITP) grant (No. RS-2019-II191082, No. RS-2022-II220156) funded by the Korea government (MSIT)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.24, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.323 + ], + "angle": 0, + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 422-440. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.786, + 0.377 + ], + "angle": 0, + "content": "2. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.786, + 0.432 + ], + "angle": 0, + "content": "3. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization. pp. 65-72 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.786, + 0.486 + ], + "angle": 0, + "content": "4. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16464-16473 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.786, + 0.528 + ], + "angle": 0, + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.786, + 0.569 + ], + "angle": 0, + "content": "6. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: European conference on computer vision. pp. 202-221. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.786, + 0.609 + ], + "angle": 0, + "content": "7. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In: European Conference on Computer Vision. pp. 487-505. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.786, + 0.663 + ], + "angle": 0, + "content": "8. Chen, J., Lei, B., Song, Q., Ying, H., Chen, D.Z., Wu, J.: A hierarchical graph network for 3d object detection on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 392-401 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.786, + 0.704 + ], + "angle": 0, + "content": "9. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11124-11133 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.704, + 0.786, + 0.758 + ], + "angle": 0, + "content": "10. Chen, S., Zhu, H., Li, M., Chen, X., Guo, P., Lei, Y., Gang, Y., Li, T., Chen, T.: Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "11. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3193-3203 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "12. Chen, Z., Hu, R., Chen, X., Nießner, M., Chang, A.X.: Unit3d: A unified transformer for 3d dense captioning and visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18109-18119 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.267, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "M. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "13. Chin-Yew, L.: Rouge: A package for automatic evaluation of summaries. In: Proceedings of the Workshop on Text Summarization Branches Out, 2004 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "14. Cornia, M., Stefanini, M., Baraldi, L., Cucchiara, R.: Meshed-memory transformer for image captioning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10578-10587 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "15. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "16. Ghandi, T., Pourreza, H., Mahyar, H.: Deep learning approaches on image captioning: A review. arXiv preprint arXiv:2201.12944 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "17. Hossain, M.Z., Sohel, F., Shiratuddin, M.F., Laga, H.: A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR) 51(6), 1-36 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "18. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition. pp. 4867-4876 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "19. Jiao, Y., Chen, S., Jie, Z., Chen, J., Ma, L., Jiang, Y.G.: More: Multi-order relation mining for dense captioning in 3d scenes. In: European Conference on Computer Vision. pp. 528-545. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "20. Jin, Z., Hayat, M., Yang, Y., Guo, Y., Lei, Y.: Context-aware alignment and mutual masking for 3d-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10984-10994 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "21. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "23. Mao, A., Yang, Z., Chen, W., Yi, R., Liu, Y.j.: Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE Transactions on Visualization and Computer Graphics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.551, + 0.785, + 0.591 + ], + "angle": 0, + "content": "24. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2906-2917 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "25. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.634, + 0.785, + 0.729 + ], + "angle": 0, + "content": "26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., Devito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems 32, pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "27. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep hough voting for 3d object detection in point clouds. In: proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9277-9286 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "30. Rennie, S.J., Marcheret, E., Mroueh, Y., Ross, J., Goel, V.: Self-critical sequence training for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7008-7024 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "31. Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems 33, 7537-7547 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "32. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4566-4575 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.314 + ], + "angle": 0, + "content": "33. Wang, H., Zhang, C., Yu, J., Cai, W.: Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.369 + ], + "angle": 0, + "content": "34. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: International conference on machine learning. pp. 2048-2057. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.411 + ], + "angle": 0, + "content": "35. Yang, X., Tang, K., Zhang, H., Cai, J.: Auto-encoding scene graphs for image captioning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10685-10694 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.467 + ], + "angle": 0, + "content": "36. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8563-8573 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.495 + ], + "angle": 0, + "content": "37. Zhong, Y., Xu, L., Luo, J., Ma, L.: Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.495 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bi-directional Contextual Attention for 3D Dense Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_origin.pdf b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4628348bbff3f503ffb90325a570b73a7db786ef --- /dev/null +++ b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/11d538e5-8e1b-4634-b736-9569f6a7b89a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af1923680c377c14ec3831680eac04654a40f696002a2eecc69a7dd4ea476e72 +size 6346949 diff --git a/2024/Bi-directional Contextual Attention for 3D Dense Captioning/full.md b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d81c15ac015467e4a03abb3ede0070ea09f2a800 --- /dev/null +++ b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/full.md @@ -0,0 +1,291 @@ +# Bi-directional Contextual Attention for 3D Dense Captioning + +Minjung Kim $^{1,4,*}$ , Hyung Suk Lim $^{3}$ , Soonyoung Lee $^{2}$ , Bumsoo Kim $^{2,\dagger}$ , and Gunhee Kim $^{1,\dagger}$ + +1 Seoul National University, Seoul, Korea minjung.kim@vision.snu.ac.kr, gunhee@snu.ac.kr + +$^{2}$ LG AI Research, Seoul, Korea {soonyoung.lee, bumsoo.kim}@lgresearch.ai + +3 Dioxide, Seoul, Korea hslim@dioxide.com + +4 SNU-LG AI Research Center, Seoul, Korea + +Abstract. 3D dense captioning is a task involving the localization of objects and the generation of descriptions for each object in a 3D scene. Recent approaches have attempted to incorporate contextual information by modeling relationships with object pairs or aggregating the nearest neighbor features of an object. However, the contextual information constructed in these scenarios is limited in two aspects: first, objects have multiple positional relationships that exist across the entire global scene, not only near the object itself. Second, it faces with contradicting objectives—where localization and attribute descriptions are generated better with tight localization, while descriptions involving global positional relations are generated better with contextualized features of the global scene. To overcome this challenge, we introduce BiCA, a transformer encoder-decoder pipeline that engages in 3D dense captioning for each object with Bi-directional Contextual Attention. Leveraging parallelly decoded instance queries for objects and context queries for non-object contexts, BiCA generates object-aware contexts, where the contexts relevant to each object is summarized, and context-aware objects, where the objects relevant to the summarized object-aware contexts are aggregated. This extension relieves previous methods from the contradicting objectives, enhancing both localization performance and enabling the aggregation of contextual features throughout the global scene; thus improving caption generation performance simultaneously. Extensive experiments on two of the most widely-used 3D dense captioning datasets demonstrate that our proposed method achieves a significant improvement over prior methods. + +# 1 Introduction + +3D dense captioning is a task that requires 1) determining the location of all objects and 2) generating descriptive sentences for each object detected within a 3D + +![](images/dedbdc99e53d499fdc716426510288b55ffac3aac0427c6f53e73d1048ec3f5b.jpg) +Fig. 1: Conceptual illustration of the multi-stage pipeline of BiCA (best viewed in color). + +scene. Early approaches have utilized a two-stage process [4, 7, 11, 19, 33, 36, 37], where the objects are detected first, then captions are generated afterward by sequentially feeding the features of each detected object. Following work [9] has borrowed the end-to-end transformer encoder-decoder pipeline from object detection [5] to actively involve contextual information for 3D dense captioning. In these works, contextual information for predicting relationships is composed by combining object features (e.g., with dot product [4, 33] or transformer attention [9]) and aggregating local nearest neighbor visual features [9]. + +Despite their improved performance, the scope of contextual information in previous 3D dense captioning methods poses its own set of challenges. First, 3D dense captioning heavily involves various positional relationships throughout the global scene; thus, the context scope restricted to the combination of individual objects or their spatially nearest neighbors typically cannot provide sufficient information. Next, the endeavor to incorporate the contextual features of various positions into a single object encounters a conflicting objective since the object must encompass precise local features for localization while simultaneously capturing various positional relationships throughout the global context for caption generation. Clearly, these challenges limit the performance of 3D dense captioning, which raises an interesting research question—“Can we design an architecture for 3D dense captioning that can effectively aggregate relevant context features without harming localization performance?” + +To this end, we propose a Bi-directional Contextual Attention (BiCA) network for 3D dense captioning. BiCA parallelly decodes a fixed set of object queries and context queries sampled from the geometric positions that do not overlap with each other. While the decoded object query performs standard object detection, the context queries are designed to distinctively capture non-object contexts and subsequently aggregate relevant contextual information for each object. Recognizing that non-object regions often lack sufficient visual features in 3D point clouds, BiCA incorporates a multi-stage pipeline where context features are enriched by relevant object features, thus implementing bidirectional attention: i) Contextual Attention of Objects for Context (O4C): + +context representation is represented by the attention summarization of global objects. ii) Contextual Attention of Contexts for Object (C4O): the object representation is adjusted by summarizing the non-object contexts. + +Then, localization is inferred with object queries while the captions are generated with object-aware context and context-aware object features. By disentangling the attention for localization and the attention for incorporating contextual features, BiCA successfully resolves the aforementioned contradicting objectives for local and global features. Moreover, since the contextual attentions retrieve globally relevant contexts and objects throughout the scene, it could accurately produce descriptions that transcend the spatial boundary of localization and its nearest neighbor. The conceptual overview of our BiCA is illustrated in Figure 1. + +BiCA sets a new state-of-the-art standard for 3D dense captioning, while simultaneously improving 3D object detection performance. Extensive experiments on two widely used benchmarks in 3D dense captioning (i.e., ScanRefer [6] and Nr3D [1]) show that our proposed BiCA surpasses prior approaches by a large margin. The contribution of our paper can be summarized as: + +- We propose a novel Bi-directional Contextual Attention network with multi-stage contextual attention for 3D dense captioning. This enables our model to capture relevant contexts throughout the global scene without being bound to single-object localization or their nearest neighbors. + +- By performing localization with decoded object queries and caption generation by including object-aware context features and context-aware object correspondence features, BiCA can simultaneously improve the performance of localization and caption generation for 3D dense captioning. + +- Our BiCA achieves state-of-the-art performances across multiple evaluation metrics on two widely used benchmarks for 3D dense captioning: ScanRefer and Nr3D datasets. + +# 2 Related work + +# 2.1 Image Captioning + +Image captioning is a fundamental task in visual language creation that automatically generates descriptive sentences for images. The main goal is to improve 2D scene understanding and generate captions that accurately reflect the content and context of the image [17]. However, with the advancement of deep learning in image captioning, there has been a marked shift towards the development and adoption of more sophisticated models [16]. Image captioning methods using deep learning adopt an encoder-decoder architecture, where the decoder generates a sentence from the visual features extracted by the encoder. Attention-based methods for grid regions [34] and detected objects [2] focus on specific image regions and use graph neural networks [35] or transformer layers to capture relationships between objects [14]. + +# 2.2 3D Dense Captioning + +3D dense captioning, an emerging field focused on achieving detailed object-level understanding of 3D scenes through natural language descriptions, has garnered significant interest in recent years. This task involves converting 3D visual data [15] into a consistent set of bounding boxes and generating appropriate natural language descriptions for each identified object. This task presents a considerable challenge due to the complexity of 3D scene information. + +Most 3D dense captioning models employ an encoder-decoder architecture comprising three main components: a scene encoder, a relational module, and a feature decoder. The scene encoder in 3D dense captioning utilizes 3D object detection methods such as 3DETR [24], PointNet++ [28], VoteNet [27], and PointGroup [18] to extract object-level visual features and contextual details from 3D point cloud datasets. The relational module is crucial in 3D dense captioning, modeling complex connections and cross-modal interactions among objects within indoor scenes. Relational modeling approaches vary based on task requirements and can be categorized into graph-based [7, 11, 19], transformer-based [4, 9, 33], and knowledge distillation-based [36] methods. The feature decoder generates bounding boxes and captions for candidate objects by considering attributes and relational features. GRU-based decoders with attention mechanisms are used in Scan2Cap [11], MORE [19], and D3Net [7], while transformer-based decoders are employed in SpaCap3D [33], $\chi$ -Tran2Cap [36], 3DJCG [4], and Vote2Cap-DETR [9] to facilitate the caption generation process. + +However, most existing methods use a single query set for object localization and caption generation, which often leads to conflicts in achieving the objectives of the two tasks. To address this, Vote2Cap-DETR++ [10] enhances task-specific feature capture by decoupling queries for localization and captioning, overcoming the limitations of a unified query system. Despite this improvement, challenges persist as its performance remains fundamentally bound by the precision of its object localization capabilities. + +# 3 Method + +Our goal is to take a 3D scene as input, identify all the objects within it, and generate captions for each object. Previous research has adopted a pipeline to identify objects in a 3D scene and simultaneously generate captions from the same features. However, while each caption may describe features of the object itself, it can also be based on the relationships with surrounding objects or information about the overall scene. + +This section introduces a novel transformer encoder-decoder approach with bi-directional contextual attention for the 3D dense captioning task. This approach distinguishes between features for instances and context information between instances, indirectly supervising the context features to enable end-to-end learning with the instance features. Instance queries aim to localize instances within a 3D scene and capture the characteristics of instances. In contrast, each + +![](images/b3cbb63a63752e186642596f7de173da8e76e2fd3de8941a26b3e0f5bd32b526.jpg) +Fig. 2: The overall pipeline of BiCA. We parallelly generate and decode two sets of queries (i.e., Instance Query and Context Query) that encodes the instance features and the non-object context features throughout the global scene, respectively. The object-aware contexts are calculated per each object by the weighted sum of the context queries, where the weights are calculated by the attention between the decoded instance query and context query. Then, with the object-aware context feature, the context-aware object feature is obtained by the weighted sum of the instances, which is weighted by the attention between the object-aware contexts. + +context query is designed to capture the non-object region within a scene, including positional relationships between instances in the 3D scene. The overall architecture of the newly proposed model is described in Figure 2. + +# 3.1 Encoder + +Following previous work [9], our model also adopts 3DETR [24] encoder as our scene encoder. This encoder is applied to transform the input point cloud into a set of encoded tokens that capture spatial, structural, and contextual information. Given the input point cloud $PC = [p_{\mathrm{in}};f_{\mathrm{in}}]\in \mathbb{R}^{N\times (3 + F)}$ , it is initially tokenized by a set-abstraction layer of PointNet++ [28]. This tokenized output is subsequently fed into a masked transformer encoder incorporating the set-abstraction layer, followed by two additional encoder layers. The final encoded scene tokens are denoted as $p_{enc}\in \mathbb{R}^{1,024\times 3}$ and $f_{enc}\in \mathbb{R}^{1,024\times 256}$ . + +# 3.2 Query Generator + +To separate captions bound to a single object from captions containing information relative to other objects or the global scene, we specify two separate Instance Query and Context Query from the encoded scene. While the instance query set is designed to capture individual features of objects bounded by their localization, the context query set is designed to encompass a broad area that consists of the positional information of a caption (e.g., next to, on the right of, is most far from, etc.). + +Instance Query Generator. In the Instance Query Generator, instance queries are generated to detect all objects in the 3D scene and generate captions for each object's individual attributes. The instance query $(p^o,f^o)$ is defined as follows: + +$$ +[ \Delta p _ {\text {v o t e}}; \Delta f _ {\text {v o t e}} ] = \mathrm {F F N} _ {o} (f _ {\text {e n c}}), \tag {1} +$$ + +$$ +\left(p ^ {o}, f ^ {o}\right) = \mathrm {S A} _ {o} \left(p _ {\text {e n c}} + \Delta p _ {\text {v o t e}}, f _ {\text {e n c}} + \Delta f _ {\text {v o t e}}\right), \tag {2} +$$ + +where $[\varDelta p_{\mathrm{vote}};\varDelta f_{\mathrm{vote}}]\in\mathbb{R}^{1,024\times(3+256)}$ is an offset that learns to shift the encoded points to object's centers spatially by a feed-forward network $\mathrm{FFN}_o$ , following [9]. $\mathrm{SA}_o$ represents the set-abstraction layer with a radius of 0.3 and samples 16 points for $p^o$ . Since our Instance Query Generator extracts the features from the candidate coordinates after the voting, our instance queries are not focused on specific objects. However, they are distributed across the locations where objects are present throughout the 3D scene. As a result, 256 instance queries are extracted from our Instance Query Generator. + +Context Query Generator. 3D scenes are composed of 3D points arranged in XYZ coordinates, forming specific objects, and the spaces between objects signify space. We recognize that obtaining context features should not merely involve interpolating the spaces between Instance Queries but also encompassing how these points are arranged and the composition of empty spaces to represent a scene. Inspired by this observation, we define context information as the geometric details in a 3D scene that illustrate the relationships between objects and between objects and the scene itself. We decide to extract structural information centered on specific points in the 3D scene and call them Context Queries. This approach is distinct from the Instance Query Generator. While the Instance Query Generator uses a voting network to obtain object center coordinates that do not exist in the 3D scene and extracts features such as identification from them, the proposed Context Query Generator extracts structural information without disrupting the arrangement of the 3D points. + +Given the encoded scene tokens $(p_{\mathrm{enc}}, f_{\mathrm{enc}})$ , we sample 512 points $p_{\mathrm{seed}}^c$ with farthest point sampling (FPS) [28] on $p_{\mathrm{enc}}$ that each represent a unique geometry within the global scene. The context query $(p^c, f^c)$ is then defined as: + +$$ +(p ^ {c}, f ^ {c}) = \mathrm {S A} _ {c} (p _ {\mathrm {e n c}}, f _ {\mathrm {e n c}}), \tag {3} +$$ + +where $\mathrm{SA}_c$ denotes the set-abstraction layer [28] with a radius of 1.2, sampling 64 points for $p^c$ . All hyper-parameters are determined experimentally. + +# 3.3 Decoder + +Instance Decoder and Context Decoder. Given the context queries $(p^c, f^c)$ and instance queries $(p^o, f^o)$ , we build a parallel decoding pipeline where the Context Decoder designates contextual information between objects and the + +Instance Decoder performs the localization and attribute description for the participating objects. + +Each Instance Decoder and Context Decoder consists of a transformer decoder with 8 layers according to 3DETR [24], and Fourier positional encoding is applied to XYZ coordinates. Fourier positional encoding transforms the XYZ coordinates of each query into a Fourier position embedding [31] and then combines them with the corresponding query embeddings. This encoding transforms the input data into frequency functions, capturing invariant changes in complex shapes and structures. It aids the context queries in accurately identifying geometric shapes and the instance queries in detecting objects. The context queries $V^{c}$ are afterwards contextualized for each instance query $V^{o}$ , then retrieve relevant instance query features that are structurally involved. + +# 3.4 Bi-directional Contextual Attention + +Captions must be generated for each object in relation to the scene context, and through positional encoding, it becomes possible to understand directional relationships, such as "to the right" with surrounding objects. Here, we gain the insight that information about the relative objects closely related to these relationships is also necessary, and we design the model to proceed simultaneously with a multi-stage contextualization. With the decoded queries, our method proceeds in two stages: i) Contextual Attention of Objects for Context (O4C): context representation is represented by the attention summarization of global objects. ii) Contextual Attention of Contexts for Object (C4O): the object representation is adjusted by summarizing the non-object contexts. Figure 1 and Figure 2 show detailed illustrations of obtaining the object-aware context feature and the context-aware object feature. + +Contextual Attention of Objects for Context (O4C). To retrieve the structural relationship between objects throughout the global scene, we first construct a Object-aware Context $V_{a}^{c}$ per object by summating the attention of each context position feature with regard to each object. The attention result between the instance feature and the context feature is applied to the context feature, making a weighted summarization of the geometries. At this time, the ratio is adjusted using learnable gamma. We name this result $V_{a}^{c}$ . + +Contextual Attention of Contexts for Object (C4O). Most contexts not only interact with the reference object but also have other objects that react against them. To achieve this, we go through a Contextual Attention of Contexts for Object process. This could specify the information that can simply be said to be 'next to' as 'next to the red chair'. To this end, the attention result of the context feature and instance feature with the attention weight applied is applied to the instance query feature. Likewise, it is adjusted using a learnable parameter lambda. We name this result $V_{a}^{o}$ . $V_{a}^{o}$ results from finding instance combinations that have a meaningful relationship with the contextual region focused on a + +specific instance. The concatenation of the instance token $V^o$ , contextualized geometry $V_a^c$ , and contextualized object $V_a^o$ , denoted as $V^a$ , is fed to the caption head to generate final captions. + +# 3.5 3D Dense Captioning + +Our final goal is to identify all objects within the input 3D scene and generate a caption for each object. We perform object detection and caption generation in parallel and group the results into one instance. + +Localization. To localize instances in the 3D scene, we use the decoded instance query $V^{o}$ . Through 5 MLP heads, we reformulate the box corner estimation as offset estimation from a query point to an object's center and box size regression. These localization heads are shared across the decoder layers. + +Caption Generation. For caption generation, we use a transformer decoder-based caption head based on GPT-2 [29], following the methods of Vote2CapDETR [9] and SpaCap3D [33]. Our caption head comprises two transformer decoder blocks, sinusoid positional embedding, and a linear classification layer. When generating captions for a proposal, we substitute the standard Start Of Sequence ('SOS') prefix with $V^a$ . During the inference process, our approach employs beam search to produce captions. The beam size is 5. + +# 3.6 Training BiCA + +Instance Query Loss. To train instance query generator to find an object's center by shifting points $p_{enc}$ , we use the vote loss from VoteNet [27]. Given the generated instance query $(p^o, f^o)$ and the encoded scene tokens $(p_{enc}, f_{enc})$ , the vote loss $\mathcal{L}^o$ is denoted as: + +$$ +\mathcal {L} ^ {o} = \frac {1}{M} \sum_ {i = 1} ^ {M} \sum_ {j = 1} ^ {N _ {g t}} \| p _ {i} ^ {o} - \operatorname {c n t} _ {j} \| _ {1} \cdot \mathbb {I} \left(p _ {e n c} ^ {i}\right), \tag {4} +$$ + +where $\mathbb{I}(x)$ is an indicator function that equals 1 when $x \in I_j$ and 0 otherwise, $N_{gt}$ is the number of instances in the 3D scene, $M$ is the number of $p^o$ , and $\mathrm{cnt}_j$ is the center of $j$ -th instance $I_j$ . + +Detection Loss. Following DETR [5], we use Hungarian matching [21] to assign each proposal with the ground truth. The detection loss $\mathcal{L}_{\mathrm{det}}$ is defined as: + +$$ +\mathcal {L} _ {\mathrm {d e t}} = \alpha_ {1} \mathcal {L} _ {\mathrm {g i o u}} + \alpha_ {2} \mathcal {L} _ {\mathrm {c l s}} + \alpha_ {3} \mathcal {L} _ {\mathrm {c n t}} + \alpha_ {4} \mathcal {L} _ {\mathrm {s i z e}}, \tag {5} +$$ + +with $\alpha_{1} = 10, \alpha_{2} = 1, \alpha_{3} = 5, \alpha_{4} = 1$ set heuristically. This loss is applied across all decoder layers for better convergence. + +Caption Loss. Following the standard image captioning protocol, we first train the caption heads using cross-entropy loss for Maximum Likelihood Estimation (MLE). During MLE training, the model predicts the $(t + 1)$ -th word $c_{n}^{t + 1}$ based on the first $t$ words $c_{n}^{[1:t]}$ and the visual features $\mathcal{V}$ . The loss function for a final caption of length $T$ is defined as follows: + +$$ +\mathcal {L} _ {c _ {n}} = \sum_ {t = 1} ^ {T} \mathcal {L} _ {c _ {n}} (t) = - \sum_ {t = 1} ^ {T} \log \hat {P} \left(c _ {n} ^ {t + 1} | \mathcal {V}, c _ {n} ^ {[ 1: t ]}\right). \tag {6} +$$ + +After word-level supervision, the caption head is refined using Self-Critical Sequence Training (SCST) [30], where the model generates multiple captions $\hat{c}_{1,\dots,k}$ with a beam size of $k$ and an additional caption $\hat{g}$ via greedy search. The loss function for SCST is formulated as follows: + +$$ +\mathcal {L} _ {c _ {n}} = - \sum_ {i = 1} ^ {k} \left(R \left(\hat {c} _ {i}\right) - R (\hat {g})\right) \cdot \frac {1}{\left| \hat {c} _ {i} \right|} \log \hat {P} \left(\hat {c} _ {i} \mid \mathcal {V}\right). \tag {7} +$$ + +The reward function $R(\cdot)$ is based on the CIDEr [32] metric for caption evaluation, and the logarithmic probability of the caption $\hat{c}_i$ is normalized by its length $|\hat{c}_i|$ , ensuring equal importance to captions of different lengths. + +Final Loss. Given the instance query Loss $\mathcal{L}^o$ , the detection loss for the $i$ -th decoder layer as $\mathcal{L}_{\mathrm{det}}^i$ , and the average of the caption loss $\mathcal{L}_{c_n}$ within a batch, denoted as $\mathcal{L}_{\mathrm{cap}}$ , the final loss $\mathcal{L}$ is formulated as: + +$$ +\mathcal {L} = \beta_ {1} \mathcal {L} ^ {o} + \beta_ {2} \sum_ {i = 1} ^ {n _ {\mathrm {d e c - l a y e r}}} \mathcal {L} _ {\mathrm {d e t}} ^ {i} + \beta_ {3} \mathcal {L} _ {\mathrm {c a p}}, \tag {8} +$$ + +where $\beta_{1} = 10$ , $\beta_{2} = 1$ , and $\beta_{3} = 5$ . + +# 4 Experiments + +# 4.1 Datasets and Metrics + +Datasets. We focus on 3D dense captioning, leveraging two benchmark datasets: ScanRefer [6] and $\mathrm{Nr3D}$ [1]. These datasets offer an extensive human-generated description of 3D scenes and objects. ScanRefer encompasses 36,665 descriptions covering 7,875 objects within 562 scenes, while $\mathrm{Nr3D}$ contains 32,919 descriptions for 4,664 objects across 511 scenes. Both datasets draw their training data from the ScanNet [15] database, which includes 1,201 3D scenes. For evaluation, we use 9,508 descriptions for 2,068 objects across 141 scenes from ScanRefer and 8,584 descriptions for 1,214 objects in 130 scenes from $\mathrm{Nr3D}$ , all sourced from the 312 3D scenes in the ScanNet validation set. + +Metrics. We evaluate model performance using four metrics: CIDEr [32], BLEU-4 [25],METEOR [3], and ROUGE-L [13], denoted as $\mathbf{C}$ , B-4, $\mathbf{M}$ , and $\mathbf{R}$ , respectively. Following the previous studies [4,9,11,19,33], Non-Maximum Suppression (NMS) is initially applied to filter out duplicate object predictions among the proposals. Each proposal is represented as a pair consisting of a predicted bounding box $\hat{b} i$ and its associated caption $\hat{c} i$ . To accurately evaluate the model's capacity for object localization and caption generation, we employ the metric $m@k$ , setting the IoU thresholds at 0.25 and 0.5 for our experiments, following [11]: + +$$ +m @ k = \frac {1}{N} \sum_ {i = 1} ^ {N} m \left(\hat {c} _ {i}, C _ {i}\right) \cdot \mathbb {I} \left\{\operatorname {I o U} \left(\hat {b} _ {i}, b _ {i}\right) \geq k \right\}, \tag {9} +$$ + +where $N$ denotes the total number of annotated objects in the evaluation set, and $m$ stands for the captioning metrics C, B-4, M, and R. + +# 4.2 Implementation Details + +Following the [9], our training comprises three stages. We pre-train our network without the caption head on the ScanNet dataset [15] for 1,080 epochs, using a batch size of 8. We minimize the loss function with an AdamW optimizer [22], starting with a learning rate of $5 \times 10^{-4}$ that reduces to $10^{-6}$ using a cosine annealing schedule, along with a weight decay of 0.1 and gradient clipping at 0.1 as per [24]. We then proceed to jointly train the model using standard cross-entropy loss for 720 epochs on both ScanRefer [6] and Nr3D [1], maintaining the detector's learning rate at $10^{-6}$ and reducing the caption head from $10^{-4}$ to $10^{-6}$ to avoid overfitting. In the SCST [30], we adjust the caption head using a batch size of 2 while keeping the detector fixed over a span of 180 epochs and maintain a constant learning rate of $10^{-6}$ . Additionally, for experiments incorporating 2D data, we use the pre-trained ENet [8] to extract 128-dimensional multi-view features from ScanNet images as outlined in Scan2Cap [11]. The model has 16.9M parameters, and the average inference time on ScanRefer [6] is 1.8ms. All experiments are conducted on a single Titan RTX GPU using PyTorch [26]. + +# 4.3 Comparison with Existing Methods + +In this section, we evaluate our performance against state-of-the-art methods: Scan2Cap [11], D3Net [7], SpaCap3D [33], MORE [19], 3DJCG [4], Contextual [37], REMAN [23], 3D-VLP [20], $\chi$ -Tran2Cap [36], Vote2Cap-DETR [9], Unit3D [12], and Vote2Cap-DETR++ [10]. We apply IoU thresholds of 0.25 and 0.5 for ScanRefer [6] as shown in Table 1 and an IoU threshold of 0.5 for Nr3D [1] indicated in Table 2. For the baselines, we present the evaluation results reported in the original papers, and "-" in Table 1 and Table 2 indicates that such results have not been reported in the original paper. + +
ModelTrainingw/o additional 2D dataw/ additional 2D data
IoU=0.25IoU=0.50IoU=0.25IoU=0.50
C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑
Scan2Cap53.7334.2526.1454.9535.2022.3621.4443.5756.8234.1826.2955.2739.0823.3221.9744.78
D3Net------------46.0730.2924.3551.67
SpaCap3d58.0635.3026.1655.0342.7625.3822.8445.6663.3036.4626.7155.7144.0225.2622.3345.36
MORE58.8935.4126.3655.4138.9823.0121.6544.3362.9136.2526.7556.3340.9422.9321.6644.42
3DJCG60.8639.6727.4559.0247.6831.5324.2851.8064.7040.1727.6659.2349.4831.0324.2250.80
Contextual----42.7723.6022.0545.13----46.1125.4722.6445.96
REMANMLE--------62.0136.3727.7656.2545.0026.3122.6746.96
3D-VLP64.0939.8427.6558.7850.0231.8724.5351.1770.7341.0328.1459.7254.9432.3124.8351.51
Vote2Cap-DETR71.4539.3428.2559.3361.8134.4626.2254.4072.7939.1728.0659.2359.3232.4225.2852.38
Unit3D------------46.6927.2221.9145.98
Vote2Cap-DETR++76.3641.3728.7060.0067.5837.0526.8955.6477.0340.9928.5359.5964.3234.7326.0453.67
BiCA (Ours)78.4241.4628.8260.0268.4638.2327.5658.5678.3541.2028.8259.8066.4736.1326.7154.54
Scan2Cap------------48.3826.0922.1544.74
D3Net------------62.6435.6825.7253.90
x-Tran2Cap58.8134.1725.8154.1041.5223.8321.9044.9761.8335.6526.6154.7043.8725.0522.4645.28
ContextualSCST----50.2925.6422.5744.71----54.3027.2423.3045.81
Vote2Cap-DETR84.1542.5128.4759.2673.7738.2126.6454.7186.2842.6428.2759.0770.6335.6925.5152.28
Vote2Cap-DETR++88.2844.0728.7559.8978.1639.7226.9455.5288.5643.3028.6459.1974.4437.1826.2053.30
BiCA (Ours)89.7244.9728.9660.6980.1440.1627.7656.1089.3444.5628.7459.3376.3437.3426.6054.00
+ +Table 1: Experimental results on the ScanRefer [6]. C, B-4, M, and R represent the captioning metrics CIDEr [32], BLEU-4 [25],METEOR [3],and ROUGE-L [13], respectively. A higher score for each indicates better performance. + +
ModelTrainingC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
Scan2Cap27.4717.2421.8049.06
D3Net33.8520.7023.1353.38
SpaCap3d33.7119.9222.6150.50
3DJCG38.0622.8223.7752.99
ContextualMLE35.2620.4222.7750.78
REMAN34.8120.3723.0150.99
Vote2Cap-DETR43.8426.6825.4154.43
Vote2Cap-DETR++47.0827.7025.4455.22
BiCA (Ours)48.7728.3525.6055.81
D3Net38.4222.2224.7454.37
\( \chi \)-Tran2Cap33.6219.2922.2750.00
Contextual37.3720.9622.8951.11
Vote2Cap-DETRSCST45.5326.8825.4354.76
Vote2Cap-DETR++47.6228.4125.6354.77
BiCA (Ours)49.8128.8325.8556.46
+ +Table 2: Experimental results on the Nr3D [1] with IoU threshold at 0.5. + +ScanRefer. Descriptions in the ScanRefer include the target object's attributes and spatial relationships with surrounding objects. As Table 1 shows, our method outperforms existing methods in all data settings and IoU thresholds, thanks to our multi-stage contextual attention method. + +Nr3D. The Nr3D dataset evaluates the model's proficiency in interpreting human-spoken, free-form object descriptions. Our method shows a notable performance improvement over existing models in generating diverse object descriptions, as indicated in Table 2. + +# 4.4 Ablation Study and Discussion + +The core components of our method are i) decomposition of the query set into Instance Query $V^{o}$ and Context Query $V^{c}$ , ii) Contextual Attention of Objects + +
ModelIoU=0.50
C↑B-4↑M↑R↑mAP↑AR↑
Vote2Cap-DETR [9]73.7738.2126.6454.7145.5667.77
BiCA using only Vo74.9040.6726.7655.3150.1269.49
Vote2Cap-DETR++ [10]78.1639.7226.9455.5255.4870.89
BiCA using Vo, KNN(Vc)79.0341.3627.3156.7655.9569.62
BiCA using Vo, Voc81.2240.9027.3957.9556.9170.38
BiCA using Vo, Voc, Vac(Ours)85.1442.2727.9859.3757.5872.68
+ +Table 3: Ablation study on the ScanRefer [6]. The core components of our BiCA are i) decomposition of the query set into Instance Query $V^{o}$ and Context Query $V^{c}$ , ii) the Object-aware Context feature $V_{a}^{c}$ , and iii) the Context-aware object feature $V_{a}^{o}$ . + +for Context (O4C) ii) Contextual Attention of Contexts for Object (C4O). We demonstrate that all components of BiCA contribute positively to the final performance, as shown in Table 3. + +Instance Query Generator. We define BiCA using only the Instance Query Generator (i.e., BiCA using only $V^o$ in Table 3) as our baseline and compare it with Vote2Cap-DETR [9], an object-centric transformer encoder-decoder architecture. The major difference between our instance query generator and Vote2Cap-DETR is how we generate the query set for instances. Vote2Cap-DETR uses farthest point sampling (FPS) to generate queries before the query coordinates are adjusted through voting. Therefore, if the coordinates are mistakenly focused on a specific object after voting, features will be extracted from the same object. On the other hand, our instance query generator extracts the features from the candidate coordinates after the voting. This improves the number of matching candidates (e.g., for 2,068 objects in our evaluation set, our method has 1,540 matching proposals while Vote2Cap-DETR has 1,498). This enhancement boosts localization performance in terms of mean Average Precision (mAP) and Average Recall (AR), which directly contributes to improving dense captioning performance. + +Context Query Generator. To enable object features to focus on localization, we independently generate context and instance queries separately. In the Vote2Cap-DETR++ [10], the decoupled queries are projections of the object-centric queries and still entail the limitations of the object-centric design. In contrast, our context queries capture structural information from the entire 3D scene, effectively decoupling features from object localization, as shown in Table 3. + +O4C module and C4O module. As shown in Table 3, utilizing the object-aware context feature $V_{a}^{c}$ and the context-aware object feature $V_{a}^{o}$ results in performance improvements across all aspects. Interestingly, we can see that performance improves even when surrounding context information is collected using KNN (See the results of BiCA using only $V^{o}$ , $\mathrm{KNN}(V^{c})$ ). In the setting of BiCA + +![](images/d89b17529005707034de99f445ea0d980b47749ab42601b918362ec739f11505.jpg) +(a) + +![](images/2f44dc37f07297e8898de53a341b44c00926ac19a439a3f909c85ca5330f960d.jpg) +(b) + +![](images/bb521a03de15b02d1d94f803e14e6f092e977293f2fe23b8177ca85f6f612178.jpg) +(c) + +![](images/7381bf2a40eb5ea169ca39c9397c957cd0f775aadf9789d756def8261295aca1.jpg) +(d) + +![](images/02c2283338ad9cb745153a0d78b89b97f81021f5a15a82f5c4b75cb82bc13db1.jpg) +(e) +(f) + +![](images/2460846e40025705019eb5297088e322ca40cb894f6c71d7d470c38f067bb1c7.jpg) +Fig. 3: Visualization of (a) the input 3D scene, (b) input point cloud without color, (c) visualization of instance queries, (d) one sampled instance query, (e) geometry contextualization of (d), and (f) instance contextualization of (e) on the ScanRefer [6]. + +using only $V^o$ , $\mathrm{KNN}(V^c)$ , we set $K = 16$ . Additionally, aggregating contextualized object features improves the model performance significantly. It proves that when capturing context information for captioning, adding additional object information that matches the context is helpful. + +# 4.5 Qualitative Analysis + +Figure 3 shows the parts of the entire scene being focused on at each stage of the proposed model. In Figure 3c and Figure 3d, our instance queries are not concentrated on a specific object but are spread throughout the 3D scene and exist in the location where the object exists. Figure 3e shows the part of the context queries with the highest attention score. The green part of the Figure 3e indicates the surroundings of the object and the corners of the room further away, but it needs to be related to other instances or scenes to express the situation more clearly. The blue part of Figure 3f describes the part of related instance queries for the object-aware context feature selected in Figure 3e. Most instances are inside the green contextualized region with high scores of Figure 3e, and some instances on the other side are also included. This observation shows that our method can check all instances of the entire scene to generate captions for one object. + +To demonstrate the effectiveness of the proposed method, we provide qualitative results with state-of-the-art models. SpaCap3D [33], 3DJCG [4], Vote2CapDETR [9], and Vote2Cap-DETR++ [10] have attempted to incorporate contextual information by modeling relationships with object pairs or aggregating nearest neighbor features. In the Figure 4, these methods generate captions limited to the object and its immediate relations in a fixed format (e.g., this is a white radiator. it is under right of the desk). Since their context is not collected based + +![](images/0798132423552dd668a0422bb5fe9827e8153269831d12d1bb5c1c2cec98f507.jpg) +Fig. 4: Qualitative results on the ScanRefer [6]. The yellow-highlighted sections show information specific to the object itself, the green-highlighted sections describe the relationships between objects, and the blue-highlighted sections depict the spatial position of the object in the 3D scene. Captions underlined in red indicate incorrect descriptions. FTG. represent failures in caption generation due to low IoU. + +on the entire scene, it is challenging to localize relationships with surrounding objects or locations in the scene. BiCA accurately predicts the attributes (i.e., can obtain precise representations for regions that have localization boundaries) and captions that require a sufficient understanding of structural information throughout the scene (i.e., can retrieve geometrically substantial information within the scene that exceeds the boundary of localization constraints) in various formats (e.g., the chair is the northwest - most one in the room.). + +# 5 Conclusion + +In this work, we propose BiCA, a novel end-to-end transformer encoder-decoder pipeline with multi-stage contextual attention for 3D dense captioning. Our BiCA parallelly decodes a fixed set of object queries that contain local features of individual objects and context queries which contain the non-object contexts in the 3D scene. This allows our model to capture relevant object-aware context and context-aware object features across the entire scene without being restricted to single object localization or their immediate surroundings. As the representation for localization and caption generation is disentangled, BiCA can improve both localization and contextual dense captioning performance. We validate the effectiveness of BiCA by showing that it outperforms the state-of-the-art across all metrics on two benchmarks for 3D dense captioning. + +# Acknowledgement + +This work was supported by LG AI Research and Institute of Information & Communications Technology Planning & Evaluation (IITP) grant (No. RS-2019-II191082, No. RS-2022-II220156) funded by the Korea government (MSIT). + +# References + +1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 422-440. Springer (2020) +2. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018) +3. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization. pp. 65-72 (2005) +4. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16464-16473 (2022) +5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +6. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: European conference on computer vision. pp. 202-221. Springer (2020) +7. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In: European Conference on Computer Vision. pp. 487-505. Springer (2022) +8. Chen, J., Lei, B., Song, Q., Ying, H., Chen, D.Z., Wu, J.: A hierarchical graph network for 3d object detection on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 392-401 (2020) +9. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11124-11133 (2023) +10. Chen, S., Zhu, H., Li, M., Chen, X., Guo, P., Lei, Y., Gang, Y., Li, T., Chen, T.: Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) +11. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3193-3203 (2021) +12. Chen, Z., Hu, R., Chen, X., Nießner, M., Chang, A.X.: Unit3d: A unified transformer for 3d dense captioning and visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18109-18119 (2023) + +13. Chin-Yew, L.: Rouge: A package for automatic evaluation of summaries. In: Proceedings of the Workshop on Text Summarization Branches Out, 2004 (2004) +14. Cornia, M., Stefanini, M., Baraldi, L., Cucchiara, R.: Meshed-memory transformer for image captioning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10578-10587 (2020) +15. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition +16. Ghandi, T., Pourreza, H., Mahyar, H.: Deep learning approaches on image captioning: A review. arXiv preprint arXiv:2201.12944 (2022) +17. Hossain, M.Z., Sohel, F., Shiratuddin, M.F., Laga, H.: A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR) 51(6), 1-36 (2019) +18. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition. pp. 4867-4876 (2020) +19. Jiao, Y., Chen, S., Jie, Z., Chen, J., Ma, L., Jiang, Y.G.: More: Multi-order relation mining for dense captioning in 3d scenes. In: European Conference on Computer Vision. pp. 528-545. Springer (2022) +20. Jin, Z., Hayat, M., Yang, Y., Guo, Y., Lei, Y.: Context-aware alignment and mutual masking for 3d-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10984-10994 (2023) +21. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) +22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +23. Mao, A., Yang, Z., Chen, W., Yi, R., Liu, Y.j.: Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE Transactions on Visualization and Computer Graphics (2023) +24. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2906-2917 (2021) +25. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002) +26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., Devito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems 32, pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf +27. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep hough voting for 3d object detection in point clouds. In: proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9277-9286 (2019) +28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +29. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) + +30. Rennie, S.J., Marcheret, E., Mroueh, Y., Ross, J., Goel, V.: Self-critical sequence training for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7008-7024 (2017) +31. Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems 33, 7537-7547 (2020) +32. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4566-4575 (2015) +33. Wang, H., Zhang, C., Yu, J., Cai, W.: Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688 (2022) +34. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: International conference on machine learning. pp. 2048-2057. PMLR (2015) +35. Yang, X., Tang, K., Zhang, H., Cai, J.: Auto-encoding scene graphs for image captioning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10685-10694 (2019) +36. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8563-8573 (2022) +37. Zhong, Y., Xu, L., Luo, J., Ma, L.: Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925 (2022) \ No newline at end of file diff --git a/2024/Bi-directional Contextual Attention for 3D Dense Captioning/images.zip b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0f7364a9a79b514bd747d6d9547aafdd12f02aff --- /dev/null +++ b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4b2edcb1fddcb85f8c37b13f887a68a8293ebcec738551e03b5e37b397e81f3 +size 504644 diff --git a/2024/Bi-directional Contextual Attention for 3D Dense Captioning/layout.json b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a91a11cb6c09b11d188eb94428ef298169bbee1 --- /dev/null +++ b/2024/Bi-directional Contextual Attention for 3D Dense Captioning/layout.json @@ -0,0 +1,8780 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 111, + 444, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 111, + 444, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 111, + 444, + 148 + ], + "type": "text", + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "spans": [ + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "text", + "content": "Minjung Kim" + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "inline_equation", + "content": "^{1,4,*}" + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "text", + "content": ", Hyung Suk Lim" + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "text", + "content": ", Soonyoung Lee" + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "text", + "content": ", Bumsoo Kim" + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "text", + "content": ", and Gunhee Kim" + }, + { + "bbox": [ + 181, + 167, + 433, + 194 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 201, + 201, + 411, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 201, + 411, + 224 + ], + "spans": [ + { + "bbox": [ + 201, + 201, + 411, + 224 + ], + "type": "text", + "content": "1 Seoul National University, Seoul, Korea minjung.kim@vision.snu.ac.kr, gunhee@snu.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 224, + 397, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 224, + 397, + 246 + ], + "spans": [ + { + "bbox": [ + 216, + 224, + 397, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 216, + 224, + 397, + 246 + ], + "type": "text", + "content": " LG AI Research, Seoul, Korea {soonyoung.lee, bumsoo.kim}@lgresearch.ai" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 246, + 356, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 246, + 356, + 267 + ], + "spans": [ + { + "bbox": [ + 258, + 246, + 356, + 267 + ], + "type": "text", + "content": "3 Dioxide, Seoul, Korea hslim@dioxide.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 214, + 268, + 399, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 268, + 399, + 277 + ], + "spans": [ + { + "bbox": [ + 214, + 268, + 399, + 277 + ], + "type": "text", + "content": "4 SNU-LG AI Research Center, Seoul, Korea" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 300, + 455, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 300, + 455, + 574 + ], + "spans": [ + { + "bbox": [ + 159, + 300, + 455, + 574 + ], + "type": "text", + "content": "Abstract. 3D dense captioning is a task involving the localization of objects and the generation of descriptions for each object in a 3D scene. Recent approaches have attempted to incorporate contextual information by modeling relationships with object pairs or aggregating the nearest neighbor features of an object. However, the contextual information constructed in these scenarios is limited in two aspects: first, objects have multiple positional relationships that exist across the entire global scene, not only near the object itself. Second, it faces with contradicting objectives—where localization and attribute descriptions are generated better with tight localization, while descriptions involving global positional relations are generated better with contextualized features of the global scene. To overcome this challenge, we introduce BiCA, a transformer encoder-decoder pipeline that engages in 3D dense captioning for each object with Bi-directional Contextual Attention. Leveraging parallelly decoded instance queries for objects and context queries for non-object contexts, BiCA generates object-aware contexts, where the contexts relevant to each object is summarized, and context-aware objects, where the objects relevant to the summarized object-aware contexts are aggregated. This extension relieves previous methods from the contradicting objectives, enhancing both localization performance and enabling the aggregation of contextual features throughout the global scene; thus improving caption generation performance simultaneously. Extensive experiments on two of the most widely-used 3D dense captioning datasets demonstrate that our proposed method achieves a significant improvement over prior methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 590, + 230, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 230, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 230, + 604 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 612, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 638 + ], + "type": "text", + "content": "3D dense captioning is a task that requires 1) determining the location of all objects and 2) generating descriptive sentences for each object detected within a 3D" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 341, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 341, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 341, + 653 + ], + "type": "text", + "content": "* Work done during internship at LG AI Research" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 654, + 237, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 654, + 237, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 654, + 237, + 665 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 479, + 210 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 479, + 210 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 479, + 210 + ], + "type": "image", + "image_path": "dedbdc99e53d499fdc716426510288b55ffac3aac0427c6f53e73d1048ec3f5b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 219, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 243 + ], + "type": "text", + "content": "Fig. 1: Conceptual illustration of the multi-stage pipeline of BiCA (best viewed in color)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 288, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 386 + ], + "type": "text", + "content": "scene. Early approaches have utilized a two-stage process [4, 7, 11, 19, 33, 36, 37], where the objects are detected first, then captions are generated afterward by sequentially feeding the features of each detected object. Following work [9] has borrowed the end-to-end transformer encoder-decoder pipeline from object detection [5] to actively involve contextual information for 3D dense captioning. In these works, contextual information for predicting relationships is composed by combining object features (e.g., with dot product [4, 33] or transformer attention [9]) and aggregating local nearest neighbor visual features [9]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 544 + ], + "type": "text", + "content": "Despite their improved performance, the scope of contextual information in previous 3D dense captioning methods poses its own set of challenges. First, 3D dense captioning heavily involves various positional relationships throughout the global scene; thus, the context scope restricted to the combination of individual objects or their spatially nearest neighbors typically cannot provide sufficient information. Next, the endeavor to incorporate the contextual features of various positions into a single object encounters a conflicting objective since the object must encompass precise local features for localization while simultaneously capturing various positional relationships throughout the global context for caption generation. Clearly, these challenges limit the performance of 3D dense captioning, which raises an interesting research question—“Can we design an architecture for 3D dense captioning that can effectively aggregate relevant context features without harming localization performance?”" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "To this end, we propose a Bi-directional Contextual Attention (BiCA) network for 3D dense captioning. BiCA parallelly decodes a fixed set of object queries and context queries sampled from the geometric positions that do not overlap with each other. While the decoded object query performs standard object detection, the context queries are designed to distinctively capture non-object contexts and subsequently aggregate relevant contextual information for each object. Recognizing that non-object regions often lack sufficient visual features in 3D point clouds, BiCA incorporates a multi-stage pipeline where context features are enriched by relevant object features, thus implementing bidirectional attention: i) Contextual Attention of Objects for Context (O4C):" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "M. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "context representation is represented by the attention summarization of global objects. ii) Contextual Attention of Contexts for Object (C4O): the object representation is adjusted by summarizing the non-object contexts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 153, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 153, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 153, + 480, + 248 + ], + "type": "text", + "content": "Then, localization is inferred with object queries while the captions are generated with object-aware context and context-aware object features. By disentangling the attention for localization and the attention for incorporating contextual features, BiCA successfully resolves the aforementioned contradicting objectives for local and global features. Moreover, since the contextual attentions retrieve globally relevant contexts and objects throughout the scene, it could accurately produce descriptions that transcend the spatial boundary of localization and its nearest neighbor. The conceptual overview of our BiCA is illustrated in Figure 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 250, + 481, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 481, + 310 + ], + "type": "text", + "content": "BiCA sets a new state-of-the-art standard for 3D dense captioning, while simultaneously improving 3D object detection performance. Extensive experiments on two widely used benchmarks in 3D dense captioning (i.e., ScanRefer [6] and Nr3D [1]) show that our proposed BiCA surpasses prior approaches by a large margin. The contribution of our paper can be summarized as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 321, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 479, + 369 + ], + "type": "text", + "content": "- We propose a novel Bi-directional Contextual Attention network with multi-stage contextual attention for 3D dense captioning. This enables our model to capture relevant contexts throughout the global scene without being bound to single-object localization or their nearest neighbors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 370, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 418 + ], + "type": "text", + "content": "- By performing localization with decoded object queries and caption generation by including object-aware context features and context-aware object correspondence features, BiCA can simultaneously improve the performance of localization and caption generation for 3D dense captioning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 419, + 480, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 480, + 455 + ], + "type": "text", + "content": "- Our BiCA achieves state-of-the-art performances across multiple evaluation metrics on two widely used benchmarks for 3D dense captioning: ScanRefer and Nr3D datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 479, + 234, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 234, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 234, + 491 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 509, + 251, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 251, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 251, + 521 + ], + "type": "text", + "content": "2.1 Image Captioning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Image captioning is a fundamental task in visual language creation that automatically generates descriptive sentences for images. The main goal is to improve 2D scene understanding and generate captions that accurately reflect the content and context of the image [17]. However, with the advancement of deep learning in image captioning, there has been a marked shift towards the development and adoption of more sophisticated models [16]. Image captioning methods using deep learning adopt an encoder-decoder architecture, where the decoder generates a sentence from the visual features extracted by the encoder. Attention-based methods for grid regions [34] and detected objects [2] focus on specific image regions and use graph neural networks [35] or transformer layers to capture relationships between objects [14]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 269, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 269, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 269, + 129 + ], + "type": "text", + "content": "2.2 3D Dense Captioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 209 + ], + "type": "text", + "content": "3D dense captioning, an emerging field focused on achieving detailed object-level understanding of 3D scenes through natural language descriptions, has garnered significant interest in recent years. This task involves converting 3D visual data [15] into a consistent set of bounding boxes and generating appropriate natural language descriptions for each identified object. This task presents a considerable challenge due to the complexity of 3D scene information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 211, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 390 + ], + "type": "text", + "content": "Most 3D dense captioning models employ an encoder-decoder architecture comprising three main components: a scene encoder, a relational module, and a feature decoder. The scene encoder in 3D dense captioning utilizes 3D object detection methods such as 3DETR [24], PointNet++ [28], VoteNet [27], and PointGroup [18] to extract object-level visual features and contextual details from 3D point cloud datasets. The relational module is crucial in 3D dense captioning, modeling complex connections and cross-modal interactions among objects within indoor scenes. Relational modeling approaches vary based on task requirements and can be categorized into graph-based [7, 11, 19], transformer-based [4, 9, 33], and knowledge distillation-based [36] methods. The feature decoder generates bounding boxes and captions for candidate objects by considering attributes and relational features. GRU-based decoders with attention mechanisms are used in Scan2Cap [11], MORE [19], and D3Net [7], while transformer-based decoders are employed in SpaCap3D [33], " + }, + { + "bbox": [ + 130, + 211, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 130, + 211, + 482, + 390 + ], + "type": "text", + "content": "-Tran2Cap [36], 3DJCG [4], and Vote2Cap-DETR [9] to facilitate the caption generation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 475 + ], + "type": "text", + "content": "However, most existing methods use a single query set for object localization and caption generation, which often leads to conflicts in achieving the objectives of the two tasks. To address this, Vote2Cap-DETR++ [10] enhances task-specific feature capture by decoupling queries for localization and captioning, overcoming the limitations of a unified query system. Despite this improvement, challenges persist as its performance remains fundamentally bound by the precision of its object localization capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 494, + 203, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 203, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 203, + 507 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 592 + ], + "type": "text", + "content": "Our goal is to take a 3D scene as input, identify all the objects within it, and generate captions for each object. Previous research has adopted a pipeline to identify objects in a 3D scene and simultaneously generate captions from the same features. However, while each caption may describe features of the object itself, it can also be based on the relationships with surrounding objects or information about the overall scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "This section introduces a novel transformer encoder-decoder approach with bi-directional contextual attention for the 3D dense captioning task. This approach distinguishes between features for instances and context information between instances, indirectly supervising the context features to enable end-to-end learning with the instance features. Instance queries aim to localize instances within a 3D scene and capture the characteristics of instances. In contrast, each" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "M. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 478, + 236 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 478, + 236 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 478, + 236 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 478, + 236 + ], + "type": "image", + "image_path": "b3cbb63a63752e186642596f7de173da8e76e2fd3de8941a26b3e0f5bd32b526.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "type": "text", + "content": "Fig. 2: The overall pipeline of BiCA. We parallelly generate and decode two sets of queries (i.e., Instance Query and Context Query) that encodes the instance features and the non-object context features throughout the global scene, respectively. The object-aware contexts are calculated per each object by the weighted sum of the context queries, where the weights are calculated by the attention between the decoded instance query and context query. Then, with the object-aware context feature, the context-aware object feature is obtained by the weighted sum of the instances, which is weighted by the attention between the object-aware contexts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "text", + "content": "context query is designed to capture the non-object region within a scene, including positional relationships between instances in the 3D scene. The overall architecture of the newly proposed model is described in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 427, + 203, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 427, + 203, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 203, + 438 + ], + "type": "text", + "content": "3.1 Encoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "text", + "content": "Following previous work [9], our model also adopts 3DETR [24] encoder as our scene encoder. This encoder is applied to transform the input point cloud into a set of encoded tokens that capture spatial, structural, and contextual information. Given the input point cloud " + }, + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "inline_equation", + "content": "PC = [p_{\\mathrm{in}};f_{\\mathrm{in}}]\\in \\mathbb{R}^{N\\times (3 + F)}" + }, + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "text", + "content": ", it is initially tokenized by a set-abstraction layer of PointNet++ [28]. This tokenized output is subsequently fed into a masked transformer encoder incorporating the set-abstraction layer, followed by two additional encoder layers. The final encoded scene tokens are denoted as " + }, + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "inline_equation", + "content": "p_{enc}\\in \\mathbb{R}^{1,024\\times 3}" + }, + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "inline_equation", + "content": "f_{enc}\\in \\mathbb{R}^{1,024\\times 256}" + }, + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 562, + 248, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 248, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 248, + 574 + ], + "type": "text", + "content": "3.2 Query Generator" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "To separate captions bound to a single object from captions containing information relative to other objects or the global scene, we specify two separate Instance Query and Context Query from the encoded scene. While the instance query set is designed to capture individual features of objects bounded by their localization, the context query set is designed to encompass a broad area that consists of the positional information of a caption (e.g., next to, on the right of, is most far from, etc.)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 483, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 483, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 483, + 152 + ], + "type": "text", + "content": "Instance Query Generator. In the Instance Query Generator, instance queries are generated to detect all objects in the 3D scene and generate captions for each object's individual attributes. The instance query " + }, + { + "bbox": [ + 130, + 116, + 483, + 152 + ], + "type": "inline_equation", + "content": "(p^o,f^o)" + }, + { + "bbox": [ + 130, + 116, + 483, + 152 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 163, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 163, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 241, + 163, + 481, + 175 + ], + "type": "interline_equation", + "content": "[ \\Delta p _ {\\text {v o t e}}; \\Delta f _ {\\text {v o t e}} ] = \\mathrm {F F N} _ {o} (f _ {\\text {e n c}}), \\tag {1}", + "image_path": "e1342367c54bd2127c1880de8a655327d5a98c83ccbea9f3ccbf88a57f2ca287.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 210, + 196, + 481, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 196, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 210, + 196, + 481, + 211 + ], + "type": "interline_equation", + "content": "\\left(p ^ {o}, f ^ {o}\\right) = \\mathrm {S A} _ {o} \\left(p _ {\\text {e n c}} + \\Delta p _ {\\text {v o t e}}, f _ {\\text {e n c}} + \\Delta f _ {\\text {v o t e}}\\right), \\tag {2}", + "image_path": "b7a1fae56d50d56d794edf3210de393b6642f255739ce6f8bf21358002249a89.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "inline_equation", + "content": "[\\varDelta p_{\\mathrm{vote}};\\varDelta f_{\\mathrm{vote}}]\\in\\mathbb{R}^{1,024\\times(3+256)}" + }, + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "text", + "content": " is an offset that learns to shift the encoded points to object's centers spatially by a feed-forward network " + }, + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_o" + }, + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "text", + "content": ", following [9]. " + }, + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{SA}_o" + }, + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "text", + "content": " represents the set-abstraction layer with a radius of 0.3 and samples 16 points for " + }, + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "inline_equation", + "content": "p^o" + }, + { + "bbox": [ + 130, + 216, + 482, + 312 + ], + "type": "text", + "content": ". Since our Instance Query Generator extracts the features from the candidate coordinates after the voting, our instance queries are not focused on specific objects. However, they are distributed across the locations where objects are present throughout the 3D scene. As a result, 256 instance queries are extracted from our Instance Query Generator." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 498 + ], + "type": "text", + "content": "Context Query Generator. 3D scenes are composed of 3D points arranged in XYZ coordinates, forming specific objects, and the spaces between objects signify space. We recognize that obtaining context features should not merely involve interpolating the spaces between Instance Queries but also encompassing how these points are arranged and the composition of empty spaces to represent a scene. Inspired by this observation, we define context information as the geometric details in a 3D scene that illustrate the relationships between objects and between objects and the scene itself. We decide to extract structural information centered on specific points in the 3D scene and call them Context Queries. This approach is distinct from the Instance Query Generator. While the Instance Query Generator uses a voting network to obtain object center coordinates that do not exist in the 3D scene and extracts features such as identification from them, the proposed Context Query Generator extracts structural information without disrupting the arrangement of the 3D points." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "text", + "content": "Given the encoded scene tokens " + }, + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "inline_equation", + "content": "(p_{\\mathrm{enc}}, f_{\\mathrm{enc}})" + }, + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "text", + "content": ", we sample 512 points " + }, + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{seed}}^c" + }, + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "text", + "content": " with farthest point sampling (FPS) [28] on " + }, + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{enc}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "text", + "content": " that each represent a unique geometry within the global scene. The context query " + }, + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "inline_equation", + "content": "(p^c, f^c)" + }, + { + "bbox": [ + 130, + 498, + 482, + 534 + ], + "type": "text", + "content": " is then defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 252, + 545, + 481, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 545, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 252, + 545, + 481, + 558 + ], + "type": "interline_equation", + "content": "(p ^ {c}, f ^ {c}) = \\mathrm {S A} _ {c} (p _ {\\mathrm {e n c}}, f _ {\\mathrm {e n c}}), \\tag {3}", + "image_path": "e51397d0397f1e590b5664ee008955076619526b449745309ccbd43716f10ed5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 566, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 566, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathrm{SA}_c" + }, + { + "bbox": [ + 130, + 566, + 482, + 591 + ], + "type": "text", + "content": " denotes the set-abstraction layer [28] with a radius of 1.2, sampling 64 points for " + }, + { + "bbox": [ + 130, + 566, + 482, + 591 + ], + "type": "inline_equation", + "content": "p^c" + }, + { + "bbox": [ + 130, + 566, + 482, + 591 + ], + "type": "text", + "content": ". All hyper-parameters are determined experimentally." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 609, + 204, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 204, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 204, + 620 + ], + "type": "text", + "content": "3.3 Decoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Instance Decoder and Context Decoder. Given the context queries " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "(p^c, f^c)" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and instance queries " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "(p^o, f^o)" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", we build a parallel decoding pipeline where the Context Decoder designates contextual information between objects and the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "M. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Instance Decoder performs the localization and attribute description for the participating objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "content": "Each Instance Decoder and Context Decoder consists of a transformer decoder with 8 layers according to 3DETR [24], and Fourier positional encoding is applied to XYZ coordinates. Fourier positional encoding transforms the XYZ coordinates of each query into a Fourier position embedding [31] and then combines them with the corresponding query embeddings. This encoding transforms the input data into frequency functions, capturing invariant changes in complex shapes and structures. It aids the context queries in accurately identifying geometric shapes and the instance queries in detecting objects. The context queries " + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "inline_equation", + "content": "V^{c}" + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "content": " are afterwards contextualized for each instance query " + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "inline_equation", + "content": "V^{o}" + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "content": ", then retrieve relevant instance query features that are structurally involved." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 277, + 340, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 340, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 340, + 289 + ], + "type": "text", + "content": "3.4 Bi-directional Contextual Attention" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 439 + ], + "type": "text", + "content": "Captions must be generated for each object in relation to the scene context, and through positional encoding, it becomes possible to understand directional relationships, such as \"to the right\" with surrounding objects. Here, we gain the insight that information about the relative objects closely related to these relationships is also necessary, and we design the model to proceed simultaneously with a multi-stage contextualization. With the decoded queries, our method proceeds in two stages: i) Contextual Attention of Objects for Context (O4C): context representation is represented by the attention summarization of global objects. ii) Contextual Attention of Contexts for Object (C4O): the object representation is adjusted by summarizing the non-object contexts. Figure 1 and Figure 2 show detailed illustrations of obtaining the object-aware context feature and the context-aware object feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": "Contextual Attention of Objects for Context (O4C). To retrieve the structural relationship between objects throughout the global scene, we first construct a Object-aware Context " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "V_{a}^{c}" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": " per object by summating the attention of each context position feature with regard to each object. The attention result between the instance feature and the context feature is applied to the context feature, making a weighted summarization of the geometries. At this time, the ratio is adjusted using learnable gamma. We name this result " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "V_{a}^{c}" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Contextual Attention of Contexts for Object (C4O). Most contexts not only interact with the reference object but also have other objects that react against them. To achieve this, we go through a Contextual Attention of Contexts for Object process. This could specify the information that can simply be said to be 'next to' as 'next to the red chair'. To this end, the attention result of the context feature and instance feature with the attention weight applied is applied to the instance query feature. Likewise, it is adjusted using a learnable parameter lambda. We name this result " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "V_{a}^{o}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "V_{a}^{o}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " results from finding instance combinations that have a meaningful relationship with the contextual region focused on a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "specific instance. The concatenation of the instance token " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "V^o" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ", contextualized geometry " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "V_a^c" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ", and contextualized object " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "V_a^o" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "V^a" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ", is fed to the caption head to generate final captions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 170, + 269, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 269, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 269, + 183 + ], + "type": "text", + "content": "3.5 3D Dense Captioning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 481, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 481, + 227 + ], + "type": "text", + "content": "Our final goal is to identify all objects within the input 3D scene and generate a caption for each object. We perform object detection and caption generation in parallel and group the results into one instance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 244, + 481, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 481, + 293 + ], + "type": "text", + "content": "Localization. To localize instances in the 3D scene, we use the decoded instance query " + }, + { + "bbox": [ + 130, + 244, + 481, + 293 + ], + "type": "inline_equation", + "content": "V^{o}" + }, + { + "bbox": [ + 130, + 244, + 481, + 293 + ], + "type": "text", + "content": ". Through 5 MLP heads, we reformulate the box corner estimation as offset estimation from a query point to an object's center and box size regression. These localization heads are shared across the decoder layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 310, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 395 + ], + "type": "text", + "content": "Caption Generation. For caption generation, we use a transformer decoder-based caption head based on GPT-2 [29], following the methods of Vote2CapDETR [9] and SpaCap3D [33]. Our caption head comprises two transformer decoder blocks, sinusoid positional embedding, and a linear classification layer. When generating captions for a proposal, we substitute the standard Start Of Sequence ('SOS') prefix with " + }, + { + "bbox": [ + 130, + 310, + 482, + 395 + ], + "type": "inline_equation", + "content": "V^a" + }, + { + "bbox": [ + 130, + 310, + 482, + 395 + ], + "type": "text", + "content": ". During the inference process, our approach employs beam search to produce captions. The beam size is 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 412, + 236, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 412, + 236, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 236, + 425 + ], + "type": "text", + "content": "3.6 Training BiCA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "text", + "content": "Instance Query Loss. To train instance query generator to find an object's center by shifting points " + }, + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "inline_equation", + "content": "p_{enc}" + }, + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "text", + "content": ", we use the vote loss from VoteNet [27]. Given the generated instance query " + }, + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "inline_equation", + "content": "(p^o, f^o)" + }, + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "text", + "content": " and the encoded scene tokens " + }, + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "inline_equation", + "content": "(p_{enc}, f_{enc})" + }, + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "text", + "content": ", the vote loss " + }, + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^o" + }, + { + "bbox": [ + 130, + 432, + 481, + 480 + ], + "type": "text", + "content": " is denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 224, + 488, + 481, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 488, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 224, + 488, + 481, + 522 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {o} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {N _ {g t}} \\| p _ {i} ^ {o} - \\operatorname {c n t} _ {j} \\| _ {1} \\cdot \\mathbb {I} \\left(p _ {e n c} ^ {i}\\right), \\tag {4}", + "image_path": "725741cb60ea9491682ffdb8124bcd3c4b2784ea9f5a417fbd56d224da688109.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(x)" + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "content": " is an indicator function that equals 1 when " + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "inline_equation", + "content": "x \\in I_j" + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "content": " and 0 otherwise, " + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "inline_equation", + "content": "N_{gt}" + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "content": " is the number of instances in the 3D scene, " + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "content": " is the number of " + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "inline_equation", + "content": "p^o" + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{cnt}_j" + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "content": " is the center of " + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "content": "-th instance " + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 131, + 531, + 481, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 585, + 481, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 585, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 585, + 481, + 610 + ], + "type": "text", + "content": "Detection Loss. Following DETR [5], we use Hungarian matching [21] to assign each proposal with the ground truth. The detection loss " + }, + { + "bbox": [ + 130, + 585, + 481, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}" + }, + { + "bbox": [ + 130, + 585, + 481, + 610 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 212, + 620, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 620, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 212, + 620, + 481, + 633 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d e t}} = \\alpha_ {1} \\mathcal {L} _ {\\mathrm {g i o u}} + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {c l s}} + \\alpha_ {3} \\mathcal {L} _ {\\mathrm {c n t}} + \\alpha_ {4} \\mathcal {L} _ {\\mathrm {s i z e}}, \\tag {5}", + "image_path": "6aaaa4af2c18b2d770083ee7c1f465eb4891f4dd606f3695675f6e9fea27f5c4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 10, \\alpha_{2} = 1, \\alpha_{3} = 5, \\alpha_{4} = 1" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " set heuristically. This loss is applied across all decoder layers for better convergence." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "M. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "text", + "content": "Caption Loss. Following the standard image captioning protocol, we first train the caption heads using cross-entropy loss for Maximum Likelihood Estimation (MLE). During MLE training, the model predicts the " + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "(t + 1)" + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "text", + "content": "-th word " + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "c_{n}^{t + 1}" + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "text", + "content": " based on the first " + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "text", + "content": " words " + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "c_{n}^{[1:t]}" + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "text", + "content": " and the visual features " + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "text", + "content": ". The loss function for a final caption of length " + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 116, + 479, + 178 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 189, + 480, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 189, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 206, + 189, + 480, + 220 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c _ {n}} = \\sum_ {t = 1} ^ {T} \\mathcal {L} _ {c _ {n}} (t) = - \\sum_ {t = 1} ^ {T} \\log \\hat {P} \\left(c _ {n} ^ {t + 1} | \\mathcal {V}, c _ {n} ^ {[ 1: t ]}\\right). \\tag {6}", + "image_path": "0898e8f12ec65dabdf9ec0759cc0cd63017cd0443a08cc0a46da0058fb81d0e3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 230, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 479, + 277 + ], + "type": "text", + "content": "After word-level supervision, the caption head is refined using Self-Critical Sequence Training (SCST) [30], where the model generates multiple captions " + }, + { + "bbox": [ + 131, + 230, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\hat{c}_{1,\\dots,k}" + }, + { + "bbox": [ + 131, + 230, + 479, + 277 + ], + "type": "text", + "content": " with a beam size of " + }, + { + "bbox": [ + 131, + 230, + 479, + 277 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 230, + 479, + 277 + ], + "type": "text", + "content": " and an additional caption " + }, + { + "bbox": [ + 131, + 230, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 131, + 230, + 479, + 277 + ], + "type": "text", + "content": " via greedy search. The loss function for SCST is formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 287, + 480, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 287, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 211, + 287, + 480, + 319 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c _ {n}} = - \\sum_ {i = 1} ^ {k} \\left(R \\left(\\hat {c} _ {i}\\right) - R (\\hat {g})\\right) \\cdot \\frac {1}{\\left| \\hat {c} _ {i} \\right|} \\log \\hat {P} \\left(\\hat {c} _ {i} \\mid \\mathcal {V}\\right). \\tag {7}", + "image_path": "88eb8483aee774281b8f5fb6fba2912266aee6be696eb15cadaafd595b9a0198.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 329, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 329, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 329, + 479, + 365 + ], + "type": "text", + "content": "The reward function " + }, + { + "bbox": [ + 131, + 329, + 479, + 365 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 131, + 329, + 479, + 365 + ], + "type": "text", + "content": " is based on the CIDEr [32] metric for caption evaluation, and the logarithmic probability of the caption " + }, + { + "bbox": [ + 131, + 329, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 131, + 329, + 479, + 365 + ], + "type": "text", + "content": " is normalized by its length " + }, + { + "bbox": [ + 131, + 329, + 479, + 365 + ], + "type": "inline_equation", + "content": "|\\hat{c}_i|" + }, + { + "bbox": [ + 131, + 329, + 479, + 365 + ], + "type": "text", + "content": ", ensuring equal importance to captions of different lengths." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "text", + "content": "Final Loss. Given the instance query Loss " + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^o" + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "text", + "content": ", the detection loss for the " + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "text", + "content": "-th decoder layer as " + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}^i" + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "text", + "content": ", and the average of the caption loss " + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{c_n}" + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "text", + "content": " within a batch, denoted as " + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cap}}" + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "text", + "content": ", the final loss " + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 385, + 479, + 422 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 223, + 433, + 480, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 433, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 223, + 433, + 480, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\beta_ {1} \\mathcal {L} ^ {o} + \\beta_ {2} \\sum_ {i = 1} ^ {n _ {\\mathrm {d e c - l a y e r}}} \\mathcal {L} _ {\\mathrm {d e t}} ^ {i} + \\beta_ {3} \\mathcal {L} _ {\\mathrm {c a p}}, \\tag {8}", + "image_path": "7383bebccec68a604d47e5946bd2d7f49afa243609eadf577faaadb5cd02b6e4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 473, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 287, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 473, + 287, + 486 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 10" + }, + { + "bbox": [ + 131, + 473, + 287, + 486 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 473, + 287, + 486 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 1" + }, + { + "bbox": [ + 131, + 473, + 287, + 486 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 473, + 287, + 486 + ], + "type": "inline_equation", + "content": "\\beta_{3} = 5" + }, + { + "bbox": [ + 131, + 473, + 287, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 506, + 230, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 230, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 230, + 520 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 535, + 268, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 535, + 268, + 546 + ], + "spans": [ + { + "bbox": [ + 131, + 535, + 268, + 546 + ], + "type": "text", + "content": "4.1 Datasets and Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": "Datasets. We focus on 3D dense captioning, leveraging two benchmark datasets: ScanRefer [6] and " + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": " [1]. These datasets offer an extensive human-generated description of 3D scenes and objects. ScanRefer encompasses 36,665 descriptions covering 7,875 objects within 562 scenes, while " + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": " contains 32,919 descriptions for 4,664 objects across 511 scenes. Both datasets draw their training data from the ScanNet [15] database, which includes 1,201 3D scenes. For evaluation, we use 9,508 descriptions for 2,068 objects across 141 scenes from ScanRefer and 8,584 descriptions for 1,214 objects in 130 scenes from " + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": ", all sourced from the 312 3D scenes in the ScanNet validation set." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Metrics. We evaluate model performance using four metrics: CIDEr [32], BLEU-4 [25],METEOR [3], and ROUGE-L [13], denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", B-4, " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", respectively. Following the previous studies [4,9,11,19,33], Non-Maximum Suppression (NMS) is initially applied to filter out duplicate object predictions among the proposals. Each proposal is represented as a pair consisting of a predicted bounding box " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{b} i" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " and its associated caption " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{c} i" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". To accurately evaluate the model's capacity for object localization and caption generation, we employ the metric " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "m@k" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", setting the IoU thresholds at 0.25 and 0.5 for our experiments, following [11]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 224, + 481, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 224, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 205, + 224, + 481, + 257 + ], + "type": "interline_equation", + "content": "m @ k = \\frac {1}{N} \\sum_ {i = 1} ^ {N} m \\left(\\hat {c} _ {i}, C _ {i}\\right) \\cdot \\mathbb {I} \\left\\{\\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {i}\\right) \\geq k \\right\\}, \\tag {9}", + "image_path": "d392a913e45f9cc6c84a1d5539ca83a1eabdd79df41c7abc0c177480dc79339f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 266, + 480, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 266, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 480, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 266, + 480, + 291 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 266, + 480, + 291 + ], + "type": "text", + "content": " denotes the total number of annotated objects in the evaluation set, and " + }, + { + "bbox": [ + 131, + 266, + 480, + 291 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 266, + 480, + 291 + ], + "type": "text", + "content": " stands for the captioning metrics C, B-4, M, and R." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 315, + 279, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 315, + 279, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 315, + 279, + 327 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "text", + "content": "Following the [9], our training comprises three stages. We pre-train our network without the caption head on the ScanNet dataset [15] for 1,080 epochs, using a batch size of 8. We minimize the loss function with an AdamW optimizer [22], starting with a learning rate of " + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "text", + "content": " that reduces to " + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "text", + "content": " using a cosine annealing schedule, along with a weight decay of 0.1 and gradient clipping at 0.1 as per [24]. We then proceed to jointly train the model using standard cross-entropy loss for 720 epochs on both ScanRefer [6] and Nr3D [1], maintaining the detector's learning rate at " + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "text", + "content": " and reducing the caption head from " + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "text", + "content": " to avoid overfitting. In the SCST [30], we adjust the caption head using a batch size of 2 while keeping the detector fixed over a span of 180 epochs and maintain a constant learning rate of " + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 130, + 340, + 482, + 521 + ], + "type": "text", + "content": ". Additionally, for experiments incorporating 2D data, we use the pre-trained ENet [8] to extract 128-dimensional multi-view features from ScanNet images as outlined in Scan2Cap [11]. The model has 16.9M parameters, and the average inference time on ScanRefer [6] is 1.8ms. All experiments are conducted on a single Titan RTX GPU using PyTorch [26]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 544, + 340, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 340, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 340, + 556 + ], + "type": "text", + "content": "4.3 Comparison with Existing Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this section, we evaluate our performance against state-of-the-art methods: Scan2Cap [11], D3Net [7], SpaCap3D [33], MORE [19], 3DJCG [4], Contextual [37], REMAN [23], 3D-VLP [20], " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "-Tran2Cap [36], Vote2Cap-DETR [9], Unit3D [12], and Vote2Cap-DETR++ [10]. We apply IoU thresholds of 0.25 and 0.5 for ScanRefer [6] as shown in Table 1 and an IoU threshold of 0.5 for Nr3D [1] indicated in Table 2. For the baselines, we present the evaluation results reported in the original papers, and \"-\" in Table 1 and Table 2 indicates that such results have not been reported in the original paper." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "M. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 479, + 278 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 278 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 278 + ], + "type": "table", + "html": "
ModelTrainingw/o additional 2D dataw/ additional 2D data
IoU=0.25IoU=0.50IoU=0.25IoU=0.50
C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑
Scan2Cap53.7334.2526.1454.9535.2022.3621.4443.5756.8234.1826.2955.2739.0823.3221.9744.78
D3Net------------46.0730.2924.3551.67
SpaCap3d58.0635.3026.1655.0342.7625.3822.8445.6663.3036.4626.7155.7144.0225.2622.3345.36
MORE58.8935.4126.3655.4138.9823.0121.6544.3362.9136.2526.7556.3340.9422.9321.6644.42
3DJCG60.8639.6727.4559.0247.6831.5324.2851.8064.7040.1727.6659.2349.4831.0324.2250.80
Contextual----42.7723.6022.0545.13----46.1125.4722.6445.96
REMANMLE--------62.0136.3727.7656.2545.0026.3122.6746.96
3D-VLP64.0939.8427.6558.7850.0231.8724.5351.1770.7341.0328.1459.7254.9432.3124.8351.51
Vote2Cap-DETR71.4539.3428.2559.3361.8134.4626.2254.4072.7939.1728.0659.2359.3232.4225.2852.38
Unit3D------------46.6927.2221.9145.98
Vote2Cap-DETR++76.3641.3728.7060.0067.5837.0526.8955.6477.0340.9928.5359.5964.3234.7326.0453.67
BiCA (Ours)78.4241.4628.8260.0268.4638.2327.5658.5678.3541.2028.8259.8066.4736.1326.7154.54
Scan2Cap------------48.3826.0922.1544.74
D3Net------------62.6435.6825.7253.90
x-Tran2Cap58.8134.1725.8154.1041.5223.8321.9044.9761.8335.6526.6154.7043.8725.0522.4645.28
ContextualSCST----50.2925.6422.5744.71----54.3027.2423.3045.81
Vote2Cap-DETR84.1542.5128.4759.2673.7738.2126.6454.7186.2842.6428.2759.0770.6335.6925.5152.28
Vote2Cap-DETR++88.2844.0728.7559.8978.1639.7226.9455.5288.5643.3028.6459.1974.4437.1826.2053.30
BiCA (Ours)89.7244.9728.9660.6980.1440.1627.7656.1089.3444.5628.7459.3376.3437.3426.6054.00
", + "image_path": "09c33e2cd3a2d3f738e7b161af02ba8fe7bbad3b0a45560adbe7f2f135225525.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 219, + 332, + 392, + 454 + ], + "blocks": [ + { + "bbox": [ + 132, + 280, + 479, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 280, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 479, + 312 + ], + "type": "text", + "content": "Table 1: Experimental results on the ScanRefer [6]. C, B-4, M, and R represent the captioning metrics CIDEr [32], BLEU-4 [25],METEOR [3],and ROUGE-L [13], respectively. A higher score for each indicates better performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 332, + 392, + 454 + ], + "lines": [ + { + "bbox": [ + 219, + 332, + 392, + 454 + ], + "spans": [ + { + "bbox": [ + 219, + 332, + 392, + 454 + ], + "type": "table", + "html": "
ModelTrainingC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
Scan2Cap27.4717.2421.8049.06
D3Net33.8520.7023.1353.38
SpaCap3d33.7119.9222.6150.50
3DJCG38.0622.8223.7752.99
ContextualMLE35.2620.4222.7750.78
REMAN34.8120.3723.0150.99
Vote2Cap-DETR43.8426.6825.4154.43
Vote2Cap-DETR++47.0827.7025.4455.22
BiCA (Ours)48.7728.3525.6055.81
D3Net38.4222.2224.7454.37
\\( \\chi \\)-Tran2Cap33.6219.2922.2750.00
Contextual37.3720.9622.8951.11
Vote2Cap-DETRSCST45.5326.8825.4354.76
Vote2Cap-DETR++47.6228.4125.6354.77
BiCA (Ours)49.8128.8325.8556.46
", + "image_path": "cc2555b02eeb3b544d9665aaacd802969d5e2faee812f8f6a206cb44092f53d0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 454, + 457, + 465 + ], + "lines": [ + { + "bbox": [ + 156, + 454, + 457, + 465 + ], + "spans": [ + { + "bbox": [ + 156, + 454, + 457, + 465 + ], + "type": "text", + "content": "Table 2: Experimental results on the Nr3D [1] with IoU threshold at 0.5." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 496, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 496, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 496, + 479, + 544 + ], + "type": "text", + "content": "ScanRefer. Descriptions in the ScanRefer include the target object's attributes and spatial relationships with surrounding objects. As Table 1 shows, our method outperforms existing methods in all data settings and IoU thresholds, thanks to our multi-stage contextual attention method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 559, + 479, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 479, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 479, + 607 + ], + "type": "text", + "content": "Nr3D. The Nr3D dataset evaluates the model's proficiency in interpreting human-spoken, free-form object descriptions. Our method shows a notable performance improvement over existing models in generating diverse object descriptions, as indicated in Table 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 624, + 315, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 624, + 315, + 635 + ], + "spans": [ + { + "bbox": [ + 133, + 624, + 315, + 635 + ], + "type": "text", + "content": "4.4 Ablation Study and Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "content": "The core components of our method are i) decomposition of the query set into Instance Query " + }, + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "V^{o}" + }, + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "content": " and Context Query " + }, + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "V^{c}" + }, + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "content": ", ii) Contextual Attention of Objects" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 114, + 411, + 186 + ], + "blocks": [ + { + "bbox": [ + 201, + 114, + 411, + 186 + ], + "lines": [ + { + "bbox": [ + 201, + 114, + 411, + 186 + ], + "spans": [ + { + "bbox": [ + 201, + 114, + 411, + 186 + ], + "type": "table", + "html": "
ModelIoU=0.50
C↑B-4↑M↑R↑mAP↑AR↑
Vote2Cap-DETR [9]73.7738.2126.6454.7145.5667.77
BiCA using only Vo74.9040.6726.7655.3150.1269.49
Vote2Cap-DETR++ [10]78.1639.7226.9455.5255.4870.89
BiCA using Vo, KNN(Vc)79.0341.3627.3156.7655.9569.62
BiCA using Vo, Voc81.2240.9027.3957.9556.9170.38
BiCA using Vo, Voc, Vac(Ours)85.1442.2727.9859.3757.5872.68
", + "image_path": "4f669c5ef821d8687d62d4a3bb1341fc273185e5f2d0b926497b86b3cb810420.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "lines": [ + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "type": "text", + "content": "Table 3: Ablation study on the ScanRefer [6]. The core components of our BiCA are i) decomposition of the query set into Instance Query " + }, + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "type": "inline_equation", + "content": "V^{o}" + }, + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "type": "text", + "content": " and Context Query " + }, + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "type": "inline_equation", + "content": "V^{c}" + }, + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "type": "text", + "content": ", ii) the Object-aware Context feature " + }, + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "type": "inline_equation", + "content": "V_{a}^{c}" + }, + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "type": "text", + "content": ", and iii) the Context-aware object feature " + }, + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "type": "inline_equation", + "content": "V_{a}^{o}" + }, + { + "bbox": [ + 130, + 186, + 481, + 218 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 251, + 480, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 480, + 288 + ], + "type": "text", + "content": "for Context (O4C) ii) Contextual Attention of Contexts for Object (C4O). We demonstrate that all components of BiCA contribute positively to the final performance, as shown in Table 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 485 + ], + "type": "text", + "content": "Instance Query Generator. We define BiCA using only the Instance Query Generator (i.e., BiCA using only " + }, + { + "bbox": [ + 130, + 306, + 482, + 485 + ], + "type": "inline_equation", + "content": "V^o" + }, + { + "bbox": [ + 130, + 306, + 482, + 485 + ], + "type": "text", + "content": " in Table 3) as our baseline and compare it with Vote2Cap-DETR [9], an object-centric transformer encoder-decoder architecture. The major difference between our instance query generator and Vote2Cap-DETR is how we generate the query set for instances. Vote2Cap-DETR uses farthest point sampling (FPS) to generate queries before the query coordinates are adjusted through voting. Therefore, if the coordinates are mistakenly focused on a specific object after voting, features will be extracted from the same object. On the other hand, our instance query generator extracts the features from the candidate coordinates after the voting. This improves the number of matching candidates (e.g., for 2,068 objects in our evaluation set, our method has 1,540 matching proposals while Vote2Cap-DETR has 1,498). This enhancement boosts localization performance in terms of mean Average Precision (mAP) and Average Recall (AR), which directly contributes to improving dense captioning performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 503, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 587 + ], + "type": "text", + "content": "Context Query Generator. To enable object features to focus on localization, we independently generate context and instance queries separately. In the Vote2Cap-DETR++ [10], the decoupled queries are projections of the object-centric queries and still entail the limitations of the object-centric design. In contrast, our context queries capture structural information from the entire 3D scene, effectively decoupling features from object localization, as shown in Table 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "O4C module and C4O module. As shown in Table 3, utilizing the object-aware context feature " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "V_{a}^{c}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " and the context-aware object feature " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "V_{a}^{o}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " results in performance improvements across all aspects. Interestingly, we can see that performance improves even when surrounding context information is collected using KNN (See the results of BiCA using only " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "V^{o}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{KNN}(V^{c})" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "). In the setting of BiCA" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "M. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 117, + 246, + 182 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 146, + 126 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 146, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 146, + 126 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 117, + 246, + 182 + ], + "lines": [ + { + "bbox": [ + 149, + 117, + 246, + 182 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 246, + 182 + ], + "type": "image", + "image_path": "d89b17529005707034de99f445ea0d980b47749ab42601b918362ec739f11505.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 265, + 114, + 361, + 180 + ], + "blocks": [ + { + "bbox": [ + 248, + 116, + 261, + 126 + ], + "lines": [ + { + "bbox": [ + 248, + 116, + 261, + 126 + ], + "spans": [ + { + "bbox": [ + 248, + 116, + 261, + 126 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 265, + 114, + 361, + 180 + ], + "lines": [ + { + "bbox": [ + 265, + 114, + 361, + 180 + ], + "spans": [ + { + "bbox": [ + 265, + 114, + 361, + 180 + ], + "type": "image", + "image_path": "2f44dc37f07297e8898de53a341b44c00926ac19a439a3f909c85ca5330f960d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 380, + 114, + 475, + 180 + ], + "blocks": [ + { + "bbox": [ + 364, + 117, + 375, + 126 + ], + "lines": [ + { + "bbox": [ + 364, + 117, + 375, + 126 + ], + "spans": [ + { + "bbox": [ + 364, + 117, + 375, + 126 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 114, + 475, + 180 + ], + "lines": [ + { + "bbox": [ + 380, + 114, + 475, + 180 + ], + "spans": [ + { + "bbox": [ + 380, + 114, + 475, + 180 + ], + "type": "image", + "image_path": "bb521a03de15b02d1d94f803e14e6f092e977293f2fe23b8177ca85f6f612178.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 150, + 196, + 244, + 261 + ], + "blocks": [ + { + "bbox": [ + 133, + 197, + 146, + 206 + ], + "lines": [ + { + "bbox": [ + 133, + 197, + 146, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 146, + 206 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 150, + 196, + 244, + 261 + ], + "lines": [ + { + "bbox": [ + 150, + 196, + 244, + 261 + ], + "spans": [ + { + "bbox": [ + 150, + 196, + 244, + 261 + ], + "type": "image", + "image_path": "7381bf2a40eb5ea169ca39c9397c957cd0f775aadf9789d756def8261295aca1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 265, + 195, + 361, + 259 + ], + "blocks": [ + { + "bbox": [ + 249, + 198, + 261, + 206 + ], + "lines": [ + { + "bbox": [ + 249, + 198, + 261, + 206 + ], + "spans": [ + { + "bbox": [ + 249, + 198, + 261, + 206 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 265, + 195, + 361, + 259 + ], + "lines": [ + { + "bbox": [ + 265, + 195, + 361, + 259 + ], + "spans": [ + { + "bbox": [ + 265, + 195, + 361, + 259 + ], + "type": "image", + "image_path": "02c2283338ad9cb745153a0d78b89b97f81021f5a15a82f5c4b75cb82bc13db1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 198, + 376, + 207 + ], + "lines": [ + { + "bbox": [ + 364, + 198, + 376, + 207 + ], + "spans": [ + { + "bbox": [ + 364, + 198, + 376, + 207 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 380, + 194, + 475, + 259 + ], + "blocks": [ + { + "bbox": [ + 380, + 194, + 475, + 259 + ], + "lines": [ + { + "bbox": [ + 380, + 194, + 475, + 259 + ], + "spans": [ + { + "bbox": [ + 380, + 194, + 475, + 259 + ], + "type": "image", + "image_path": "2460846e40025705019eb5297088e322ca40cb894f6c71d7d470c38f067bb1c7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 284, + 480, + 318 + ], + "lines": [ + { + "bbox": [ + 131, + 284, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 131, + 284, + 480, + 318 + ], + "type": "text", + "content": "Fig. 3: Visualization of (a) the input 3D scene, (b) input point cloud without color, (c) visualization of instance queries, (d) one sampled instance query, (e) geometry contextualization of (d), and (f) instance contextualization of (e) on the ScanRefer [6]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 341, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 480, + 389 + ], + "type": "text", + "content": "using only " + }, + { + "bbox": [ + 131, + 341, + 480, + 389 + ], + "type": "inline_equation", + "content": "V^o" + }, + { + "bbox": [ + 131, + 341, + 480, + 389 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 341, + 480, + 389 + ], + "type": "inline_equation", + "content": "\\mathrm{KNN}(V^c)" + }, + { + "bbox": [ + 131, + 341, + 480, + 389 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 131, + 341, + 480, + 389 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 131, + 341, + 480, + 389 + ], + "type": "text", + "content": ". Additionally, aggregating contextualized object features improves the model performance significantly. It proves that when capturing context information for captioning, adding additional object information that matches the context is helpful." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 406, + 263, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 263, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 263, + 418 + ], + "type": "text", + "content": "4.5 Qualitative Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "text", + "content": "Figure 3 shows the parts of the entire scene being focused on at each stage of the proposed model. In Figure 3c and Figure 3d, our instance queries are not concentrated on a specific object but are spread throughout the 3D scene and exist in the location where the object exists. Figure 3e shows the part of the context queries with the highest attention score. The green part of the Figure 3e indicates the surroundings of the object and the corners of the room further away, but it needs to be related to other instances or scenes to express the situation more clearly. The blue part of Figure 3f describes the part of related instance queries for the object-aware context feature selected in Figure 3e. Most instances are inside the green contextualized region with high scores of Figure 3e, and some instances on the other side are also included. This observation shows that our method can check all instances of the entire scene to generate captions for one object." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the proposed method, we provide qualitative results with state-of-the-art models. SpaCap3D [33], 3DJCG [4], Vote2CapDETR [9], and Vote2Cap-DETR++ [10] have attempted to incorporate contextual information by modeling relationships with object pairs or aggregating nearest neighbor features. In the Figure 4, these methods generate captions limited to the object and its immediate relations in a fixed format (e.g., this is a white radiator. it is under right of the desk). Since their context is not collected based" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 465, + 309 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 465, + 309 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 465, + 309 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 465, + 309 + ], + "type": "image", + "image_path": "0798132423552dd668a0422bb5fe9827e8153269831d12d1bb5c1c2cec98f507.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 316, + 482, + 373 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 373 + ], + "type": "text", + "content": "Fig. 4: Qualitative results on the ScanRefer [6]. The yellow-highlighted sections show information specific to the object itself, the green-highlighted sections describe the relationships between objects, and the blue-highlighted sections depict the spatial position of the object in the 3D scene. Captions underlined in red indicate incorrect descriptions. FTG. represent failures in caption generation due to low IoU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "content": "on the entire scene, it is challenging to localize relationships with surrounding objects or locations in the scene. BiCA accurately predicts the attributes (i.e., can obtain precise representations for regions that have localization boundaries) and captions that require a sufficient understanding of structural information throughout the scene (i.e., can retrieve geometrically substantial information within the scene that exceeds the boundary of localization constraints) in various formats (e.g., the chair is the northwest - most one in the room.)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 511, + 220, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 220, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 220, + 523 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In this work, we propose BiCA, a novel end-to-end transformer encoder-decoder pipeline with multi-stage contextual attention for 3D dense captioning. Our BiCA parallelly decodes a fixed set of object queries that contain local features of individual objects and context queries which contain the non-object contexts in the 3D scene. This allows our model to capture relevant object-aware context and context-aware object features across the entire scene without being restricted to single object localization or their immediate surroundings. As the representation for localization and caption generation is disentangled, BiCA can improve both localization and contextual dense captioning performance. We validate the effectiveness of BiCA by showing that it outperforms the state-of-the-art across all metrics on two benchmarks for 3D dense captioning." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "M. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 480, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 480, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 480, + 174 + ], + "type": "text", + "content": "This work was supported by LG AI Research and Institute of Information & Communications Technology Planning & Evaluation (IITP) grant (No. RS-2019-II191082, No. RS-2022-II220156) funded by the Korea government (MSIT)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 255 + ], + "type": "text", + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.: Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 422-440. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 255, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 481, + 298 + ], + "type": "text", + "content": "2. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "text", + "content": "3. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization. pp. 65-72 (2005)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 342, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 384 + ], + "type": "text", + "content": "4. Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16464-16473 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 384, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 481, + 418 + ], + "type": "text", + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "text", + "content": "6. Chen, D.Z., Chang, A.X., Nießner, M.: Scanrefer: 3d object localization in rgb-d scans using natural language. In: European conference on computer vision. pp. 202-221. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "text", + "content": "7. Chen, D.Z., Wu, Q., Nießner, M., Chang, A.X.: D3net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In: European Conference on Computer Vision. pp. 487-505. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 482, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 481, + 525 + ], + "type": "text", + "content": "8. Chen, J., Lei, B., Song, Q., Ying, H., Chen, D.Z., Wu, J.: A hierarchical graph network for 3d object detection on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 392-401 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "text", + "content": "9. Chen, S., Zhu, H., Chen, X., Lei, Y., Yu, G., Chen, T.: End-to-end 3d dense captioning with vote2cap-detr. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11124-11133 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 557, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 557, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 557, + 481, + 600 + ], + "type": "text", + "content": "10. Chen, S., Zhu, H., Li, M., Chen, X., Guo, P., Lei, Y., Gang, Y., Li, T., Chen, T.: Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 481, + 633 + ], + "type": "text", + "content": "11. Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3193-3203 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "12. Chen, Z., Hu, R., Chen, X., Nießner, M., Chang, A.X.: Unit3d: A unified transformer for 3d dense captioning and visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18109-18119 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "13. Chin-Yew, L.: Rouge: A package for automatic evaluation of summaries. In: Proceedings of the Workshop on Text Summarization Branches Out, 2004 (2004)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "14. Cornia, M., Stefanini, M., Baraldi, L., Cucchiara, R.: Meshed-memory transformer for image captioning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10578-10587 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "15. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "type": "text", + "content": "16. Ghandi, T., Pourreza, H., Mahyar, H.: Deep learning approaches on image captioning: A review. arXiv preprint arXiv:2201.12944 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 480, + 258 + ], + "type": "text", + "content": "17. Hossain, M.Z., Sohel, F., Shiratuddin, M.F., Laga, H.: A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR) 51(6), 1-36 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "text", + "content": "18. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: Pointgroup: Dual-set point grouping for 3d instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition. pp. 4867-4876 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 480, + 325 + ], + "type": "text", + "content": "19. Jiao, Y., Chen, S., Jie, Z., Chen, J., Ma, L., Jiang, Y.G.: More: Multi-order relation mining for dense captioning in 3d scenes. In: European Conference on Computer Vision. pp. 528-545. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "text", + "content": "20. Jin, Z., Hayat, M., Yang, Y., Guo, Y., Lei, Y.: Context-aware alignment and mutual masking for 3d-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10984-10994 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 480, + 379 + ], + "type": "text", + "content": "21. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 480, + 402 + ], + "type": "text", + "content": "22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 133, + 403, + 480, + 435 + ], + "type": "text", + "content": "23. Mao, A., Yang, Z., Chen, W., Yi, R., Liu, Y.j.: Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE Transactions on Visualization and Computer Graphics (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 133, + 436, + 480, + 468 + ], + "type": "text", + "content": "24. Misra, I., Girdhar, R., Joulin, A.: An end-to-end transformer model for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2906-2917 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 480, + 501 + ], + "type": "text", + "content": "25. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 502, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 502, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 502, + 480, + 577 + ], + "type": "text", + "content": "26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., Devito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems 32, pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 480, + 610 + ], + "type": "text", + "content": "27. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep hough voting for 3d object detection in point clouds. In: proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9277-9286 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "text", + "content": "28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "29. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "M. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 392 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "30. Rennie, S.J., Marcheret, E., Mroueh, Y., Ross, J., Goel, V.: Self-critical sequence training for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7008-7024 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "31. Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems 33, 7537-7547 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "32. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4566-4575 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 248 + ], + "type": "text", + "content": "33. Wang, H., Zhang, C., Yu, J., Cai, W.: Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 292 + ], + "type": "text", + "content": "34. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: International conference on machine learning. pp. 2048-2057. PMLR (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "35. Yang, X., Tang, K., Zhang, H., Cai, J.: Auto-encoding scene graphs for image captioning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10685-10694 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 369 + ], + "type": "text", + "content": "36. Yuan, Z., Yan, X., Liao, Y., Guo, Y., Li, G., Cui, S., Li, Z.: X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8563-8573 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 370, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 392 + ], + "type": "text", + "content": "37. Zhong, Y., Xu, L., Luo, J., Ma, L.: Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925 (2022)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bi-directional Contextual Attention for 3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_content_list.json b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ae89d4dbc62cdb0057d8c949bc714b1cf536f434 --- /dev/null +++ b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_content_list.json @@ -0,0 +1,2146 @@ +[ + { + "type": "text", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "text_level": 1, + "bbox": [ + 266, + 140, + 738, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zichen Zhang $^{1}$ , Hongchen Luo $^{1}$ , Wei Zhai $^{1,*}$ , Yang Cao $^{1,2}$ , and Yu Kang $^{1,2}$", + "bbox": [ + 228, + 210, + 771, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Science and Technology of China", + "bbox": [ + 336, + 238, + 663, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center {zhangzichen@mail., lhc12@mail., wzhai056@, kangduyu@, forrest@}ustc.edu.cn", + "bbox": [ + 228, + 253, + 772, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Interaction intention anticipation aims to jointly predict future hand trajectories and interaction hotspots. Existing research often treated trajectory forecasting and interaction hotspots prediction as separate tasks or solely considered the impact of trajectories on interaction hotspots, which led to the accumulation of prediction errors over time. However, a deeper inherent connection exists between hand trajectories and interaction hotspots, which allows for continuous mutual correction between them. Building upon this relationship, a novel Bidirectional prOgressive Transformer (BOT), which introduces a Bidirectional Progressive mechanism into the anticipation of interaction intention is established. Initially, BOT maximizes the utilization of spatial information from the last observation frame through the Spatial-Temporal Reconstruction Module, mitigating conflicts arising from changes of view in first-person videos. Subsequently, based on two independent prediction branches, a Bidirectional Progressive Enhancement Module is introduced to mutually improve the prediction of hand trajectories and interaction hotspots over time to minimize error accumulation. Finally, acknowledging the intrinsic randomness in human natural behavior, we employ a Trajectory Stochastic Unit and a C-VAE to introduce appropriate uncertainty to trajectories and interaction hotspots, respectively. Our method achieves state-of-the-art results on three benchmark datasets Epic-Kitchens-100, EGO4D, and EGTEA Gaze+, demonstrating superior in complex scenarios.", + "bbox": [ + 259, + 327, + 743, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Interaction Intention $\\cdot$ Egocentric Videos $\\cdot$ Transformer", + "bbox": [ + 261, + 657, + 720, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 696, + 375, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The pursuit of future prediction on egocentric videos is centered on anticipating intentions and behaviors observed from a first-person perspective [59]. Interaction intention can be reflected in the trajectory of hand motion and contact points on the next-active object [39]. This task is novel but holds significant potential value, such as AR [34, 61], robot skill acquisition [3, 9, 24, 44, 48], hand reconstruction [30, 67] and pose estimation [27, 38].", + "bbox": [ + 212, + 726, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 228, + 825, + 393, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/729d4c622d0dc0f1bf249262094c561098ee63d23f74e22b99e3041986ba1e7c.jpg", + "image_caption": [ + "Error Accumulation Over Time" + ], + "image_footnote": [], + "bbox": [ + 316, + 141, + 431, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/73c8ca3256aaec031421fe424af6593f591d00b95aece1223c4db66e030ec5c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 141, + 547, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f2e97a34d637bcc7ca8e8e111115663f3a03eda4c86273ed331d0efeda79d2b2.jpg", + "image_caption": [ + "Fig. 1: Error accumulation when only considering the influence from hand trajectories to interaction hotspots. Red ones represent the ground truth, while Blue ones are predictions." + ], + "image_footnote": [], + "bbox": [ + 571, + 141, + 687, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing relevant research treated hand trajectory forecasting and interaction hotspots prediction as independent tasks. Bao et al. [4] proposed to predict the 3D hand trajectory from a first-person view. However, They overlooked the guiding role of interactable objects on trajectories. The prediction process was incomplete, leading to significant errors at the end of trajectories. On the other hand, there was extensive research on interaction hotspots [18,32,41,42,51,52,63]. Nevertheless, these studies often focused on the characteristics of the objects themselves, resulting in lower integration with subjective human behaviors in complex scenes and less accurate predictions. Other studies [37,39] jointly predicted future hand trajectories and interaction hotspots, improving their integration. However, these studies only considered the unidirectional influence of hand trajectories on contact points. Consequently, errors accumulated continuously with the prediction time steps (Fig. 1).", + "bbox": [ + 212, + 320, + 789, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In fact, this one-way causation was insufficient, as it disregarded the spatial correlation between the two factors. Conversely, there exists a mutual relationship between the predictions of hand trajectories and interaction hotspots. On one hand, the derived directions of hand trajectories suggest the approximate areas where contact points occur. On the other hand, there is an inherent connection between the spatial distribution of contact points and hand trajectory patterns, as in Fig. 2, which implies that examining the distribution of contact points can provide insights into the development of hand trajectories. Leveraging this relationship, to mitigate the process of error accumulation, we establish a Bidirectional Progressive (Bi-Progressive) mechanism, allowing continuous mutual refinement between trajectories and interaction hotspots throughout the entire prediction period, improving the accuracy of joint predictions.", + "bbox": [ + 212, + 520, + 789, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, an innovative Bidirectional prOgressive Transformer (BOT) is proposed, introducing the Bi-Progressive mechanism mentioned earlier between two independent predicting branches. Firstly, the raw video input undergoes a Spatial-Temporal Reconstruction Module, which integrates the spatial information from the last observation frame and the temporal information among all observation frames. This process mitigates spatial information conflicts arising from changes in the field of view in first-person videos. Secondly, we establish separate dual branches for hand trajectory forecasting and interaction hotspots prediction, endowing them with the capability to independently perform indi", + "bbox": [ + 212, + 704, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/deff622f82bdf84eaf74a92c45fed659133792a3d4ca0acd0f30840eb3a03024.jpg", + "image_caption": [ + "(a) Long-length trajectories" + ], + "image_footnote": [], + "bbox": [ + 331, + 143, + 421, + 193 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7775fc263d940aa62090d77e1489f88bdb64940a5f53fe34c931a284d37f139f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 143, + 513, + 193 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/904b37d7bce550a3152177872f041f9a750b1a700a85c0ae1c8b349713a4c5e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 143, + 604, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1dad43f57b2dba72cd157ecaa28db7456c94982bfb3fd3a240bc7dce26aa1a63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 143, + 669, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/510f94951bc8fb717922f0dafc397a3eab8e83bf2a3c966b991b1223adcfb7d7.jpg", + "image_caption": [ + "(b) Medium-length trajectories" + ], + "image_footnote": [], + "bbox": [ + 331, + 203, + 421, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75a02884eefae83d9f647dc935b346d9e58419ac33e0a684c204c4c0b01e9d0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 203, + 513, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d4757304e7c23018db96e89d3d46c7dbfc0a5d7cae50451f3e65fbbf1d8d8330.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 203, + 604, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f932b23785c83bbd6ab62b1bb0c2036f6aafad609740fdeeff8e9be084a4b20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 203, + 669, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8cd833900d4d17f76445756c6e29d50e1e5885cad39fd30444d9b655b7f92d2.jpg", + "image_caption": [ + "(c) Short-length trajectories" + ], + "image_footnote": [], + "bbox": [ + 331, + 263, + 421, + 311 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/44e598bc4896d72cac03bad8b06a4feb60ce84130da800404c160e7cd0dbdbfc.jpg", + "image_caption": [ + "Fig. 2: Inherent connection between different hand trajectory categories and their corresponding contact points distribution. This is conducted on Epic-Kitchens-100 [14]." + ], + "image_footnote": [], + "bbox": [ + 423, + 263, + 513, + 311 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/95ac88be95a6d3be921b581de02b3ca7a4b886741ddf8b8fa43338f4006f408c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 263, + 604, + 311 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9289fa77c6feacfb3779179cbb29ba6f5110b4730021aefb40da0096a775be4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 263, + 669, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vidual prediction tasks, laying the foundation for introducing the Bi-Progressive mechanism. Subsequently, we introduce cross-attention blocks as the primary implementation of the Bi-Progressive mechanism. At each time step, the predictions of hand trajectories and interaction hotspots are alternately updated, addressing inherent errors in individual predictions. Finally, following the logic of natural human behaviors [50, 57, 74], we incorporate a Trajectory Stochastic Unit and a C-VAE for both hand trajectories and interaction hotspots, respectively, introducing appropriate uncertainties to the final prediction results.", + "bbox": [ + 212, + 383, + 787, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate our approach on Epic-Kitchens-100 [14], EGO4D [28] and EGTEA Gaze+ [36] datasets. BOT demonstrates notable advantages in predicting both hand trajectories and interaction hotspots, achieving superior performance in anticipating interaction intentions. In summary, our contributions are as follows:", + "bbox": [ + 212, + 505, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel BOT is introduced to jointly predict future hand trajectories and interaction hotspots in egocentric videos, attaining the optimal predictive performance with more lenient input conditions.", + "- We introduce a novel Bidirectional Progressive mechanism that enhances the coherence between hand trajectories and interaction hotspots, thereby improving the completeness of the interaction intention.", + "- The experimental results demonstrate that our method achieves state-of-the-art results across three benchmark datasets Epic-Kitchens-100 [14], EGO4D [28], and EGTEA Gaze+ [36], demonstrating superior in complex scenarios." + ], + "bbox": [ + 225, + 575, + 785, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 748, + 387, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Action anticipation in egocentric videos. Anticipating actions in an egocentric vision is a challenging task that includes short-term and long-term action anticipation [56,59]. Short-term action anticipation aims to predict the next action in the near future [21-23,25,49], while long-term action anticipation focuses", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on the sequence of subsequent actions performed by the camera wearer [26, 40, 47, 53]. These endeavors contribute significantly to the comprehension of actions in first-person videos. However, concerning interaction intention, action anticipation still yields information of relatively high dimensionality. Anticipating a semantic label does not furnish direct guidance for the robot, as it lacks specific instructions on task execution. Comparatively, hand trajectories and interaction hotspots offer a more intuitive representation of interaction intention, constituting a lower-dimensional depiction.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human trajectory forecasting. The forecasting of human trajectories was extensively studied over the years [1,2,12,13,45,54]. However, most of them were based on third-person video datasets, falling short of capturing intentions related to human-object interactions. Since hands serve as the primary medium through which humans interact with the world, predicting hand trajectories in egocentric videos can aid embodied AI in comprehending human-object interaction intentions. The approaches in [4,37,39] addressed the issue of hand trajectory prediction in egocentric videos. However, they did not account for the influence of contact points with objects on trajectories. This limitation resulted in the continuous accumulation of prediction errors over time.", + "bbox": [ + 212, + 267, + 787, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interaction hotspots prediction. Interaction hotspots prediction aims to locate where the hand-object interaction happens. Some research aimed to predict the next interactive object [5, 6, 15, 17, 20], going further, some work anticipated specific interaction regions on objects [42, 43, 52, 72, 73, 75, 76], yielding more precise information for interaction. Differing from image input, our task aims to generate more precise predictions regarding the regions of contact points by utilizing egocentric video inputs. The incorporation of first-person videos is intended to focus on precisely localizing interaction hotspots for distinct human motions. In this task, we leverage the predicted hand trajectories to improve the accuracy of interaction hotspots predictions.", + "bbox": [ + 212, + 417, + 787, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 593, + 330, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Setup. Given an original input video $\\mathcal{I} = \\{I_1, I_2, \\dots, I_N\\}$ , where $I_{\\mathrm{N}}$ is the last observation frame, the task is predicting future hand trajectories $\\mathcal{H} = \\{h_{\\mathrm{N + 1}}, h_{\\mathrm{N + 2}}, \\dots, h_{\\mathrm{N + T - 1}}\\}$ in future keyframes $\\{I_{\\mathrm{N + 1}}, I_{\\mathrm{N + 2}}, \\dots, I_{\\mathrm{N + T - 1}}\\}$ and the contact point $\\mathcal{O}_{\\mathrm{C}}$ in the contact frame $I_{N + T}$ . We subsequently convert the contact point $\\mathcal{O}_{\\mathrm{C}}$ into a heatmap by centering a Gaussian distribution, generating interaction hotspots $\\mathcal{O}_{\\mathrm{I}}$ . To ensure the continuity of the trajectory as well as the consistency of the background, both $\\mathcal{H}$ and $\\mathcal{O}_{\\mathrm{I}}$ are projected onto the last observation frame, which enhances the overall coherence of the predictions [39].", + "bbox": [ + 212, + 625, + 787, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Network", + "text_level": 1, + "bbox": [ + 215, + 768, + 333, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. As shown in Fig. 3, given the original input video $\\mathcal{I}$ , we aim to predict the future hand trajectory $\\mathcal{H}$ and interaction hotspots $\\mathcal{O}_{\\mathrm{I}}$ . Initially, to alleviate information conflicts arising from changes in the field of view, the", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1d22189a85db342bb468c6b4afa5b756fc13ca2d042ab9455079a017bb1c51ab.jpg", + "image_caption": [ + "Fig. 3: Overview of the Bidirectional Progressive Transformer. It follows a dual-branch structure with a Bi-Progressive Enhancement Module between them to anticipate future hand trajectories and interaction hotspots." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "original input video undergoes a Spatial-Temporal Reconstruction Module that integrates spatial information from the last observation frame and temporal information from the entire video, producing spatial features for each future timestep. Subsequently, these features are separated, with the last one serving as the contact feature while the preceding ones are for trajectory features. These features are then fed into the interaction hotspots and trajectory prediction branches, respectively. The former employs an overlap patch embedding layer to enhance attention at the center of the field of view, while the latter introduces temporal order to trajectory features through Cross-Transformers [11]. Simultaneously, we employ a Bi-Progressive Enhancement Module consisting of cross-attention blocks between dual branches, introducing a Bi-Progressive mechanism to the joint prediction. This alternately and iteratively corrects the features of trajectories and interaction hotspots in chronological order, thereby mitigating the accumulation of prediction errors over time. Finally, considering the inherent uncertainty in natural human behaviors [50, 57, 74], we introduce a Stochastic Unit for trajectory forecasting while employing a C-VAE [62] for interaction hotspots, adding appropriate uncertainty during the completion of position decoding to achieve comprehensive predictions.", + "bbox": [ + 212, + 398, + 787, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial-Temporal Reconstruction. When dealing with egocentric videos, both 3D convolutions [19] and self-attention mechanisms [68] struggle to address spatial information conflicts arising from changes in views. Moreover, predictions are projected onto the last observation frame. Taking these factors into account, we establish the Spatial-Temporal Reconstruction Module. On one hand, we employ optical flow containing more comprehensive motion information as the temporal input, on the other hand, we isolate the last observation frame as the spatial input. Subsequently, the temporal input is processed by a pre-trained Temporal Segment Network [22,69], extracting low-dimensional temporal features $\\mathbf{X}$ , and the spatial input is passed through a Segformer-b2 [70], extracting spatial features $\\mathbf{S}$ from the last observation frame. $\\mathbf{X}$ and $\\mathbf{S}$ are then fed into a", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "bbox": [ + 227, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Transformer [11]. The $i$ -th block is expressed as:", + "bbox": [ + 215, + 146, + 609, + 162 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {0} = \\mathbf {s} _ {\\text {p a t c h}} + \\mathbf {s} _ {\\text {p o s}}, \\quad \\mathbf {x} = \\mathbf {x} + \\mathbf {x} _ {\\text {p o s}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 176, + 785, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {i} = \\mathbf {s} _ {i - 1} + \\mathbf {M C A} (\\mathbf {L N} (\\mathbf {s} _ {i - 1}), \\mathbf {L N} (\\mathbf {x})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 194, + 785, + 210 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {i} = \\mathbf {y} _ {i} + \\mathbf {F F N} (\\mathbf {L N} (\\mathbf {y} _ {i})). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 214, + 785, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\mathbf{s}_0$ and $\\mathbf{x}$ represent the original spatial and temporal inputs, respectively. $\\mathbf{s}_{patch}$ is encoded by overlap patch embedding layers from spatial features $\\mathbf{S}$ , enhancing the network's focus on the center of the field of view. $\\mathbf{x}$ is encoded by positional embedding layers from temporal features $\\mathbf{X}$ , improving the overall sequentiality. Furthermore, FFN refers to the Feed Forward Layer, comprising a two-layer MLP with an expanding ratio $r$ , and Layer normalization (LN) is applied before each block. MCA stands for Multi-head Cross-Attention, where query, key, and value are processed through distinct linear layers: MCA(X,Y) = Attention(W^QX, W^KY, W^VY). The output of this module corresponds to the spatial features for each prediction frame.", + "bbox": [ + 212, + 241, + 787, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interaction Hotspots Prediction Branch. Due to the spatial distribution of contact points closer to the central field of view, the contact feature $\\mathbf{F}_{\\mathrm{C}}$ undergoes processing through an overlap patch embedding layer:", + "bbox": [ + 212, + 393, + 785, + 438 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {o}} = \\text {o v e r l a p} (\\mathbf {F} _ {\\mathrm {C}}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 452, + 785, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\mathbf{F}_C^0$ represents the processed contact feature.", + "bbox": [ + 212, + 479, + 588, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hand Trajectory Prediction Branch. We employ original trajectory features as primary features and utilize features from the previously predicted frame as conditional features, subsequently merging them through the cross-attention block $f$ , ensuring the coherence of the trajectory over time. Assuming the trajectory features are $\\{\\mathbf{F}_{\\mathrm{N} + 1},\\mathbf{F}_{\\mathrm{N} + 2},\\dots ,\\mathbf{F}_{\\mathrm{N} + \\mathrm{T} - 1}\\}$ and the feature map of the last observation frame is $\\mathbf{F}_{\\mathrm{N}}$ , they are updated in chronological order:", + "bbox": [ + 212, + 496, + 787, + 585 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {t} ^ {\\mathrm {o}} = f \\left(\\mathbf {F} _ {\\mathrm {t}}, \\mathbf {F} _ {\\mathrm {t} - 1}\\right), \\quad t \\in [ \\mathrm {N} + 1, \\mathrm {N} + \\mathrm {T} - 1 ]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 599, + 785, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\mathbf{F}_t^{\\mathrm{o}}$ stands for the updated trajectory feature, $t$ is the prediction time step.", + "bbox": [ + 212, + 628, + 787, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "H-O Bi-Progressive Enhancement. Expanding on the groundwork laid by two distinct prediction branches, and to mitigate the accumulation of errors as predictions unfold over time, we introduce a Bi-Progressive Enhancement Module (Fig. 4), which alternately corrects the predictions of trajectories $\\mathcal{H}$ and interaction hotspots $\\mathcal{C}_{\\mathrm{I}}$ at each time step $t$ . Specifically, owing to observation frames and the continuity of hand trajectories, the trajectory feature $\\mathbf{F}_{\\mathrm{N} + 1}^{\\mathrm{o}}$ at the initial predicted time step $\\mathrm{N} + 1$ demonstrates a high level of accuracy. Exploiting this, we enhance the precision of the contact feature $\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{o}}$ through the cross-attention block $f$ , thereby minimizing the error with the ground truth. Moving to the subsequent time step $\\mathrm{N} + 2$ , we employ the first refined contact feature $\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{r}_1}$ from the previous time step $\\mathrm{N} + 1$ as a condition to correct the trajectory feature $\\mathbf{F}_{\\mathrm{N} + 2}^{\\mathrm{o}}$ . Following this, the corrected trajectory feature is utilized to further refine the contact feature. This iterative correction process persists until all trajectory", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "features have undergone a complete round of updates. It can be represented as a recursive procedure:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {N} + 1} ^ {\\mathrm {r}} = \\mathbf {F} _ {\\mathrm {N} + 1} ^ {\\mathrm {o}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 190, + 785, + 207 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {1}} = f \\left(\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {o}}, \\mathbf {F} _ {\\mathrm {N} + 1} ^ {\\mathrm {r}}\\right), \\quad \\mathbf {F} _ {\\mathrm {N} + 2} ^ {\\mathrm {r}} = f \\left(\\mathbf {F} _ {\\mathrm {N} + 2} ^ {\\mathrm {o}}, \\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {1}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 209, + 785, + 226 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\cdot \\cdot \\cdot\n$$\n", + "text_format": "latex", + "bbox": [ + 486, + 232, + 509, + 239 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {3}} = f \\left(\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {2}}, \\mathbf {F} _ {\\mathrm {N} + 3} ^ {\\mathrm {r}}\\right), \\quad \\mathbf {F} _ {\\mathrm {N} + 4} ^ {\\mathrm {r}} = f \\left(\\mathbf {F} _ {\\mathrm {N} + 4} ^ {\\mathrm {o}}, \\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {3}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 247, + 785, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Where $\\mathbf{F}_t^{\\mathrm{r}}$ represents the refined trajectory feature map at time step $t\\in [\\mathrm{N} + 1,\\mathrm{N} + \\mathrm{T} - 1]$ , $\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{r},i}$ denotes the contact feature after the $i$ -th refinement. Via this module, we attain refined features for both trajectories and interaction hotspots, significantly reducing error accumulation in predictions.", + "bbox": [ + 212, + 276, + 589, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interaction Hotspots Decoder. Given the presence of inherent uncertainty associated with interaction hotspots [39, 41, 42], we employ a C-VAE [62] as the decoder. Due to the excessive randomness associated with directly decoding the coordinates of contact points, we instruct the C-VAE to output a feature map and select the location of the maximum value from this map as the output, thereby confining the uncertainty within a reasonable range. The C-VAE follows an encoder-decoder structure. For the encoder block, we concatenate the input $\\varGamma$ with the condition $\\mathbf{C}$ through an encoding function $\\mathbf{P}_{\\mathrm{e}}$ , obtaining features in latent space $\\mathbf{Z}$ which parameterized by mean $\\mu$ and co-variance $\\sigma$ . In this context, the input $\\varGamma$ is the ground truth heatmap, while the condition $\\mathbf{C}$ is the refined contact feature $\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{r}_3}$ . In the decoder", + "bbox": [ + 212, + 367, + 589, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "block, the first step involves sampling $\\mathbf{Z}$ from the latent space and concatenating it with condition $\\mathbf{C}$ . Following that, the output heatmap $\\hat{\\Gamma}$ is reconstructed by a decoding function $\\mathbf{P}_d$ . In summary, the entire process can be formulated as follows:", + "bbox": [ + 212, + 609, + 785, + 667 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu , \\sigma = \\mathbf {P} _ {\\mathrm {e}} (\\Gamma , \\mathrm {C}), \\quad \\hat {\\Gamma} = \\mathbf {P} _ {\\mathrm {d}} (\\mathrm {Z}, \\mathrm {C}), \\quad \\mathrm {Z} \\sim \\mathcal {N} (\\mu , \\sigma). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 670, + 785, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Where $\\mathcal{N}$ represents a normal distribution, $\\hat{\\Gamma}$ is the reconstructed heatmap. Both encoding and decoding blocks consist of two MLP layers. We introduce KL-Divergence and the reconstruction error as the overall loss:", + "bbox": [ + 212, + 696, + 787, + 742 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} (\\Gamma , \\hat {\\Gamma}) = (\\Gamma - \\hat {\\Gamma}) ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 755, + 785, + 772 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k l} (\\mu , \\sigma) = - \\mathbf {K L} \\left[ \\mathcal {N} (\\mu , \\sigma) ^ {2} | | \\mathcal {N} (0, 1) \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 775, + 785, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda \\mathcal {L} _ {k l}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 796, + 785, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this paper, we set $\\lambda = 0.01$ .", + "bbox": [ + 214, + 824, + 437, + 840 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c48a0ee08548f20e60eeb879a31deebec6c2a74dcbd5c54e4734051926f5c09.jpg", + "image_caption": [ + "Fig. 4: Bi-Progressive Enhancement Module. The output feature map is enclosed in purple boxes." + ], + "image_footnote": [], + "bbox": [ + 602, + 306, + 785, + 503 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$h_t$ hand position $\\mathcal{O}_{\\mathrm{C}}$ contact point $\\mathcal{R}_{21}, \\mathcal{R}_{22}$ elliptical region $\\mathcal{R}_2$ uncertain region $\\theta$ ellipse tilt angle", + "bbox": [ + 276, + 143, + 723, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b824b0898f3f03875f2d568b7d2c2bd9ef704fca6b36a17b95f7ef5500a8f536.jpg", + "image_caption": [ + "Fig. 5: Uncertain regions of the hand trajectory. It is derived through the intersection of two ellipses, corresponding to the Eq. 14-19. (i.e., the purple region)." + ], + "image_footnote": [], + "bbox": [ + 274, + 160, + 733, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hand Trajectory Decoder. Similar to interaction hotspots, hand trajectories also exhibit uncertainty [29, 33, 46]. Given that trajectory feature maps with distinct distribution characteristics have been obtained through the Bi-Progressive Enhancement Module, in conjunction with sampling efficiency, we don't utilize C-VAE, rather, we introduce a Trajectory Stochastic Unit to incorporate uncertainty into trajectory forecasting by calculating uncertain regions for each prediction time step instead. Since optimizing directly using point coordinates for loss calculation can pose challenges, we initially introduce a Gaussian distribution for each ground truth hand position to generate the ground truth heatmap. Subsequently, we calculate the loss between the corrected feature map and the generated ground truth heatmap, thereby enabling the network to learn the distribution of potential hand positions. The associated loss function $\\mathcal{L}_h$ is based on MSE Loss:", + "bbox": [ + 212, + 316, + 789, + 513 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h} = \\frac {1}{T - 1} \\sum_ {t = N + 1} ^ {N + T - 1} \\left(\\mathbf {F} _ {t} ^ {\\mathrm {r}} - \\mathbf {G} _ {t}\\right) ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 534, + 787, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Where $\\mathbf{G}_t$ represents the ground truth heatmap at time step $t$ . We designate the coordinates of the maximum value within each feature map as the predicted trajectory points, which are then fed into the stochastic unit to generate uncertain regions. The size of the region is determined by the time step as well as the length of trajectory segments in preceding and subsequent stages. The shape is constrained by the original image dimensions and the orientation of the trajectory segments. Formally, the uncertain region (Fig. 5) can be shown as:", + "bbox": [ + 214, + 599, + 787, + 705 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {t} \\sim (t, h _ {t - 1}, h _ {t}, h _ {t + 1}, \\mathrm {W}, \\mathrm {L}). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 729, + 787, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Where $\\mathcal{R}_t$ is the uncertain region in future time step $t$ , W and L are the width and length of the input video, respectively. $\\mathcal{R}_t$ satisfies the following constraints:", + "bbox": [ + 214, + 768, + 787, + 800 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {t} = \\mathcal {R} _ {t 1} \\cap \\mathcal {R} _ {t 2}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 824, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Where $\\mathcal{R}_{t1}$ and $\\mathcal{R}_{t2}$ are elliptical regions constrained by $h_{t-1}$ and $h_{t+1}$ , which can be represented as:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\left(\\mathbf {x} _ {t} \\cos \\theta_ {t \\omega} - \\mathbf {y} _ {t} \\sin \\theta_ {t \\omega}\\right) ^ {2}}{\\mathbf {a} _ {t \\omega} ^ {2}} + \\frac {\\left(\\mathbf {x} _ {t} \\sin \\theta_ {t \\omega} + \\mathbf {y} _ {t} \\cos \\theta_ {t \\omega}\\right) ^ {2}}{\\mathbf {b} _ {t \\omega} ^ {2}} \\leq 1, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 191, + 785, + 224 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} _ {t \\omega} = \\alpha \\beta^ {t - 1} \\mathbf {d} ^ {\\frac {1}{2}} \\left(h _ {t + \\omega - 1}, h _ {t + \\omega - 2}\\right), \\quad \\mathbf {b} _ {t \\omega} = \\sqrt {\\frac {\\mathrm {L}}{\\mathrm {W}}} \\mathbf {a} _ {t \\omega}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 227, + 785, + 260 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t \\omega} = \\tan^ {- 1} (\\mathbf {k} \\left(h _ {t + \\omega - 1}, h _ {t + \\omega - 2}\\right)), \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 263, + 785, + 280 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\omega \\in [ 1, 2 ], t \\in [ 1, 2, 3, 4 ], \\mathcal {O} _ {\\mathrm {C}} = h _ {5}. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 282, + 785, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, $\\alpha$ denotes the overall scale of $\\mathcal{R}_t$ , $\\beta$ signifies the temporal growth rate, $\\theta$ stands for the ellipse tilt angle, $(\\mathbf{x}_t,\\mathbf{y}_t)$ represent the coordinates of $h_t$ . Additionally, $\\mathbf{k}(\\mathrm{X},\\mathrm{Y})$ and $\\mathbf{d}(\\mathrm{X},\\mathrm{Y})$ represent the slope and Euclidean distance between points $\\mathrm{X}$ and $\\mathrm{Y}$ , respectively. In BOT, we set $\\alpha = 1.5$ and $\\beta = 2$ .", + "bbox": [ + 214, + 315, + 787, + 377 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.2 Training and Inference", + "text_level": 1, + "bbox": [ + 215, + 407, + 450, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training. During training, we freeze parameters in TSN [69]. Training loss combines both trajectory loss and interaction hotspots loss:", + "bbox": [ + 214, + 441, + 782, + 472 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {h} + \\varphi \\mathcal {L} _ {o}. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 489, + 785, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we set $\\varphi = 1$ .", + "bbox": [ + 214, + 518, + 416, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inference. We use a Trajectory Stochastic Unit and a C-VAE to introduce uncertainty during inference. Following previous work [39, 46], we report the minimum among 20 samples for trajectory evaluation.", + "bbox": [ + 214, + 536, + 782, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 614, + 375, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 654, + 333, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conducted experiments on three first-person video datasets Epic-Kitchen s-100 [14], EGO4D [28], and EGTEA Gaze+ [36]. Epic-Kitchens-100 [14] and EGTEA Gaze+ [36] predominantly include kitchen scenes, while EGO4D [28] encompasses a broader range of activities such as handicrafts, painting, and outdoor tasks. For each dataset, the observation time is set to 2 seconds. We configure the prediction time horizon T to 5, corresponding to 4 trajectory points and one contact point for each hand. Specifically, we predict 1-second trajectories on Epic-Kitchens-100 [14] and EGO4D [28], while on EGTEA Gaze+ [36], 0.5-second trajectories are forecast. Additional dataset information is available in the supplementary materials.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 146, + 457, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In EGO4D [28], We use the method [39] to automatically generate hand trajectories, while contact points are manually annotated based on the last observation frame and the contact frame. We employ GMFLOW [71] to generate optical flow from EGO4D [28] and EGTEA Gaze+ [36]. For each dataset, we resize the original videos to $456 \\times 256$ size, extracting both trajectory and contact features of size $114 \\times 64$ . In the Spatial-Temporal Reconstruction Module, we set the dimension $D$ in Cross-Transformer to 1024, with depth $= 6$ and attention head number $= 8$ . For the trajectory forecasting branch and H-O Bi-Progressive Enhancement Module, we configure the dimension $D$ to 256, with depth $= 2$ and head number $= 8$ . Additionally, for the introduced C-VAE, we set the dimension of both input and condition to $114 \\times 64$ , while the hidden and latent sizes are set to 1024 and 2048, respectively. Our method is implemented by PyTorch and trained with the AdamW optimizer. We train the model for 100 epochs on 4 NVIDIA 3090 GPUs with an initial learning rate of $3e - 5$ and a batch size of 40.", + "bbox": [ + 212, + 170, + 787, + 382 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For comparison, We first select methods for individual tasks including Kalman Filter (KF) [8], Seq2Seq [65], Hotspots [52] and USST [4]. Additionally, the methods for image/video understanding are chosen as the compared methods, which lack a joint mechanism. The former includes VIT [16], HRNet [64], UNet [60], Segformer [70], while the latter contains I3D [10], Slowfast [19], Timesformer [7], Uniformer-V2 [35], StillFast [58]. Furthermore, we select methods with high task relevance and joint design, including FHOI [37] and OCT [39]. We conduct experiments using these approaches on Epic-kithens-100 [14], EGO4D [28], and EGTEA Gaze+ [36] datasets.", + "bbox": [ + 212, + 383, + 789, + 520 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 541, + 419, + 555 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hand Trajectory Evaluation. The predicted hand trajectories are referenced to the last observation frames. ADE [39] and FDE [39] are chosen as hand trajectory evaluation metrics, which respectively represent the overall accuracy and endpoint accuracy of trajectories. Detailed descriptions are available in the supplementary materials.", + "bbox": [ + 212, + 565, + 787, + 641 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Interaction Hotspots Evaluation. For all datasets, we downsample each predicted heatmap to the size of $32 \\times 32$ . We choose SIM [66], AUC-J [31], and NSS [55] as interaction hotspots evaluation metrics. For a detailed explanation of the metrics, please refer to the supplementary materials.", + "bbox": [ + 212, + 641, + 787, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Quantitative and Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 215, + 724, + 609, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Tab. 1, our method achieves state-of-the-art performance in both trajectory prediction and interaction hotspots anticipation, with comparable leading margins across all three datasets. Taking the EGO4D dataset [28] with the richest variety of scenes as an example. In trajectory prediction, to ensure fair comparison, we additionally present results for OCT and BOT with only one sampling iteration. Under this single-sampling condition, BOT outperforms", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c301abb7d3e794f12317efefa370d808e02cc57e071df447f3352c89c4ba3d81.jpg", + "table_caption": [ + "Table 1: Results of different models. Best results are depicted using red (sample 20 times) and blue (sample once), respectively. unilateral, non-joint and joint denote prediction approaches for individual tasks, for two tasks without a collaborative design, and with a joint module, respectively. “-” denotes the vacancy of the value. “*” represents the results obtained from 20 random samplings. ↑ and ↓ indicate a better performance when the value of the metric increases and decreases, respectively." + ], + "table_footnote": [], + "table_body": "
MethodEK100 [14]EGO4D [28]EGTEA [36]
ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑
unilateralCenter0.240.240.090.610.330.260.250.070.570.150.320.320.090.630.27
KF [8]0.330.32--------0.490.48---
Seq2Seq [65]0.180.14--------0.180.14---
Hotspots [52]--0.150.660.53-------0.150.710.69
USST [4]0.170.18---0.180.18---0.180.18---
non-jointI3D [10]0.220.230.130.560.460.220.220.120.570.420.230.220.130.600.49
SLOWFAST [19]0.220.240.140.620.470.230.220.130.580.440.210.230.140.580.46
Timesformer [7]0.230.230.140.600.420.240.240.120.590.400.210.220.140.580.45
Uniformer-V2 [35]0.200.200.130.590.430.210.220.120.620.400.200.210.130.580.46
StillFast [58]0.180.190.160.620.610.190.190.150.650.570.180.190.160.600.55
UNet [60]0.190.180.150.620.570.180.180.140.590.550.180.180.140.600.60
HRNet [64]0.180.190.160.620.630.190.200.140.620.580.180.180.160.630.65
VIT [16]0.180.190.140.650.630.210.210.130.590.450.190.190.150.630.61
CrossVIT [11]0.190.200.140.620.590.190.200.130.600.550.200.210.130.590.49
Segformer-b2 [70]0.170.170.160.640.660.160.160.150.640.620.170.170.160.630.65
jointFHOI [37]0.360.350.100.550.39-----0.340.340.120.630.42
OCT [39]0.150.17---0.150.16---0.160.16---
OCT* [39]0.120.110.190.720.720.120.120.220.730.880.140.140.230.751.01
Ours0.120.14---0.110.13---0.120.14---
Ours*0.100.080.340.812.070.090.080.320.771.700.090.080.320.802.03
", + "bbox": [ + 220, + 231, + 784, + 470 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Segformer-b2, the best method in image/video understanding, with respective leads of $31.3\\%$ and $18.8\\%$ in ADE and FDE. Compared to OCT, BOT maintains advantages of $26.7\\%$ and $18.8\\%$ in ADE and FDE, respectively. These advantages persist even with 20 samplings, while the lead in FDE expands to $33.3\\%$ , underscoring the rationale behind BOT's stochastic sampling mechanism. Regarding interaction hotspots, we achieve the best results across all metrics. Specifically, BOT surpasses the second-best approach (OCT) by $45.5\\%$ in SIM, $5.5\\%$ in AUC-J, and $93.2\\%$ in NSS, respectively. In summary, our method leverages the intrinsic correlations between hand trajectories and interaction hotspots more effectively, enhancing overall prediction accuracy of interaction intention, and exhibiting robustness across the three benchmark datasets.", + "bbox": [ + 212, + 505, + 787, + 671 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To further explore the subjective outcomes of models, we report the visualization results of three datasets, as shown in Fig. 6. Compared to other methods, our model provides more accurate prediction results in both hand trajectories and interaction hotspots. When there is significant movement of hands (column 1,3,8), the BOT's predictions exhibit the highest accuracy. It indicates that our Spatial-Temporal Reconstruction Module can efficiently utilize the human motion information to exploit the hand interaction intention and thus accurately predict the hand trajectory. In cases where there are no identifiable interactive objects in the field of view (column 6), our method anticipates more accurate interaction intentions. It illustrates that our Bi-Progressive mechanism can eliminate the uncertainty in the interaction process by mining the intrinsic connection between", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/579663f562ff6b4dae9a41557ab1dce9686104d2f72a1ad76cc94b526d628c84.jpg", + "image_caption": [ + "Fig. 6: Visualization of future hand trajectories and interaction hotspots on EK100, EGO4D and EGETA. We show the visualization results of our model, the joint prediction model (OCT) [39], the segmentation model (Segformer-b2) [70], and the video understanding model (Uniformer-V2) [35]. The green represents the trajectory of the left hand, while the red denotes the trajectory of the right hand." + ], + "image_footnote": [], + "bbox": [ + 205, + 143, + 782, + 349 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the hand trajectory and the contact points, which results in a more accurate inference in the interaction region and the hand trajectory. Moreover, our predictions closely align with the actual scenario when both hands simultaneously exhibit trajectories and interaction hotspots (columns 4,5), which suggests the strategy of differentiating between left and right hands during the prediction of interaction intentions, thereby mitigating cross-hand interference. More visualization results are available in the supplementary materials.", + "bbox": [ + 212, + 446, + 787, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 574, + 388, + 589 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Different joint approaches. To explore the impact of the Bi-Progressive mechanism on model performance, we experiment with five different joint configurations: Individual, Hand, Object, Bidirectional and Bi-Progressive. The results are shown in Tab. 2 and Fig. 7. It suggests that the network is challenging to accurately predict trajectories and interaction hotspots during individual predictions, and the separate incorporation of hands and objects can effectively improve the prediction accuracy of interaction hotspots. With the combination of hand and object information, the prediction performance of the trajectory and interaction region is greatly improved. However, there is still a gap compared to the Bi-Progressive mechanism. Further, Fig. 8 shows that the introduction of the progressive mechanism plays a pivotal role in mitigating the accumulation of prediction errors over time.", + "bbox": [ + 212, + 598, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Uncertainty in anticipation. Uncertainty is a crucial consideration in interaction intention anticipation. Following the logic of natural human behavior [50,57,74], both hand trajectories [29,33,46] and interaction hotspots [39,41,42] predictions inherently possess a degree of uncertainty. To investigate the impact", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cef00f82ab30ae2debec9d1d1af258fc7a9bc8b85920aae4df7e57c4a338679c.jpg", + "table_caption": [ + "Table 2: Results of different joint forecasting methods. Individual refers to separately predict two subtasks. Hand indicates that the hand trajectory is the condition of the interaction hotspots. Object stands for correcting the trajectory using predicted contact points. Bidirectional combines the methods Hand and Object." + ], + "table_footnote": [], + "table_body": "
MethodTrajectoryInteraction Hotspots
ADE ↓ FDE ↓SIM ↑ AUC-J ↑NSS ↑
Individual0.140.140.180.680.86
Hand0.130.120.250.721.19
Object0.120.120.230.701.04
Bidirectional0.100.090.280.731.48
Bi-Progressive0.090.080.320.771.70
", + "bbox": [ + 310, + 212, + 694, + 333 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1016746c7d59d5e7ce2f4e81cedb6def15212392e41bb8c8ed049b4ff5f675d3.jpg", + "image_caption": [ + "Fig. 7: Visualization results of different joint approaches. Based on the order from left to right and top to bottom, the corresponding methods are GT, Bi-Progressive, Bidirectional, Hand, Object, Individual." + ], + "image_footnote": [], + "bbox": [ + 218, + 352, + 488, + 452 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/682996daf8b4fa713de5d94d31ad6e7eceaa69b6bebd6287d29bb334f1713a76.jpg", + "image_caption": [ + "Fig. 8: The role of Progressive Mechanism. The red refers to the ground truth, while the blue represents the predictions. From top to bottom, Bi-Pro stands for Bi-Progressive while Bid refers to bidirectional." + ], + "image_footnote": [], + "bbox": [ + 513, + 352, + 787, + 452 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of uncertain region scale and the magnitude of uncertainty variations in hand trajectories, we modify $\\alpha$ and $\\beta$ in Eq. 17, with the resultant outcomes depicted in Fig. 9. $\\alpha$ denotes the overall scale of the uncertain region, when $\\alpha$ equals 0, the region's size is zero, and the network output yields a unique result. As $\\alpha$ increases, the uncertainty of trajectories grows, leading to improved accuracy due to random samplings. However, given the limitations in sampling times, as $\\alpha$ continues to increase, the accuracy naturally decreases. $\\beta$ determines the magnitude of trajectory uncertainty increasing from front to back, when $\\beta$ equals 1, each trajectory point has the same level of uncertainty. Considering the escalating complexity and uncertainty in trajectory prediction over time, an elevation in $\\beta$ initially corresponds to augmented accuracy, followed by a subsequent decline. This gradual increase in uncertainty also renders the FDE more sensitive to $\\alpha$ and $\\beta$ , as shown in Fig. 10a-10c. Regarding interaction hotspots, BOT exhibits relatively small uncertainty (Fig. 10d), as for a specific intention, the interaction area becomes increasingly influenced by object constraints. This is achieved by tightly controlling the uncertainty within a reasonably small range through the process of enabling C-VAE to generate feature maps and subsequently select", + "bbox": [ + 212, + 583, + 787, + 843 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9946323f2de0e43254fdf43aea0faba3f0fa204c520809e8b7a6145446835db4.jpg", + "image_caption": [ + "Fig. 9: Uncertainty in hand trajectory forecasting. As $\\alpha$ and $\\beta$ increase, they denote a heightened degree of overall trajectory uncertainty." + ], + "image_footnote": [], + "bbox": [ + 220, + 151, + 493, + 284 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/760a15b4c9d602b1549d2a669ce781c93211af3d6ea2a380cc0722dacc529d3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 151, + 781, + 286 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9706ceef5dfae98bd7e824999cb90f44c6997cd5ba8063a80c858205de88cd42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 338, + 310, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c7ce3df40f11f04fbece25e9d1f9b701778d47733f7e6ae3cf407cc3ad53a556.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 338, + 401, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/95fa90005cc49811a5cf2b25a6fb9c77aaa6e1aaa6bf2a5d19bc9e0772f5ede3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 338, + 493, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d3c792f8019df11cf04d6e88d920257bea4cca5f8a3ff7053b2d5c61f7270f19.jpg", + "image_caption": [ + "(a) Long-length trajectories", + "(c) Short-length trajectories", + "Fig. 10: Uncertainty in anticipation. (a), (b) and (c) respectively depict long-length trajectories, medium-length trajectories, and short-length trajectories, while (d) illustrates the uncertainty of interaction hotspots." + ], + "image_footnote": [], + "bbox": [ + 220, + 402, + 310, + 454 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0cfd00471519f5e232259737c6a0d580a05742e2aeca17f994ca306669647783.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 402, + 401, + 454 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c9a6ec9d8581cab8aac50879e50fcf55f95a0711048cc5419fb82077b82e2974.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 402, + 493, + 453 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/746bc5e6a08a005ffa5b9757a001cbe707d5f13573df33442dd5c655a059fe5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 338, + 599, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a9258eca08991ea1b76caf076842b07def6f5b0da924eaf3859f0d06a104f293.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 338, + 691, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a0f48ce9da74069ef86b2b4958fb6ef2119d64d984833938fb29d08a44654c81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 338, + 782, + 388 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1a2e7ba433910f827577e8f1c603fc9ee02180df9b39a6583be09a7034394358.jpg", + "image_caption": [ + "(b) Medium-length trajectories", + "(d) interaction hotspots" + ], + "image_footnote": [], + "bbox": [ + 509, + 402, + 601, + 454 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fbfe206760380f513fb903ce4148aeb988429e5e1d380b89bf3dcee908babc4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 402, + 691, + 454 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2e82da4de91af397b39ffff4f8c750e05a0d85ed2af730938172ef5d9ed0b8e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 402, + 782, + 454 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "points, thereby ensuring the network's high confidence in predicting interaction regions.", + "bbox": [ + 212, + 547, + 785, + 578 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 599, + 354, + 614 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion. We plan to anticipate interaction intentions, encompassing the joint prediction of future hand trajectories and interaction hotspots with objects. We propose a novel BOT framework to establish an inherent connection between hand trajectories and contact points by introducing a Bi-Progressive mechanism, which eliminates the effect of error accumulation and achieves more accurate intention anticipation. Experimental results on three challenging benchmarks achieve state-of-the-art performance, thereby proving the superiority of our method in intention anticipation.", + "bbox": [ + 212, + 628, + 787, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future Work. BOT currently predicts hand trajectories and interaction hotspots on 2D images. However, human-object interactions occur in 3D space. To enable the model to be applied to real-world scenarios, we will consider the geometric structure and spatial relationships between objects in the scene in the future, providing the model with the ability to anticipate interaction intentions for the real world.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62306295 and 62033012.", + "bbox": [ + 212, + 178, + 787, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 248, + 323, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adeli, V., Adeli, E., Reid, I., Niebles, J.C., Rezatofighi, H.: Socially and contextually aware human motion and pose forecasting. IEEE Robotics and Automation Letters 5(4), 6033-6040 (2020)", + "2. Adeli, V., Ehsanpour, M., Reid, I., Niebles, J.C., Savarese, S., Adeli, E., Rezatofighi, H.: Tripod: Human trajectory and pose dynamics forecasting in the wild. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13390-13400 (2021)", + "3. Bahl, S., Gupta, A., Pathak, D.: Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450 (2022)", + "4. Bao, W., Chen, L., Zeng, L., Li, Z., Xu, Y., Yuan, J., Kong, Y.: Uncertainty-aware state space transformer for egocentric 3d hand trajectory forecasting. arXiv preprint arXiv:2307.08243 (2023)", + "5. Bertasius, G., Park, H.S., Yu, S.X., Shi, J.: First person action-object detection with egonet. arXiv preprint arXiv:1603.04908 (2016)", + "6. Bertasius, G., Soo Park, H., Yu, S.X., Shi, J.: Unsupervised learning of important objects from first-person videos. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1956-1964 (2017)", + "7. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)", + "8. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE international conference on image processing (ICIP). pp. 3464-3468. IEEE (2016)", + "9. Bharadhwaj, H., Gupta, A., Tulsiani, S., Kumar, V.: Zero-shot robot manipulation from passive human videos. arXiv preprint arXiv:2302.02011 (2023)", + "0. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "1. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021)", + "2. Chen, G., Chen, Z., Fan, S., Zhang, K.: Unsupervised sampling promoting for stochastic human trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17874-17884 (2023)", + "3. Choi, C., Dariush, B.: Looking to relations for future trajectory forecast. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)", + "4. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Scaling egocentric vision: The epic-kitchens dataset. In: European Conference on Computer Vision (ECCV) (2018)" + ], + "bbox": [ + 225, + 281, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Dessalene, E., Devaraj, C., Maynard, M., Fermuller, C., Aloimonos, Y.: Forecasting action through contact representations from first person video. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)", + "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)", + "17. Fan, C., Lee, J., Ryoo, M.S.: Forecasting hands and objects in future frames. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)", + "18. Fang, K., Wu, T.L., Yang, D., Savarese, S., Lim, J.J.: Demo2vec: Reasoning object affordances from online videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2139-2147 (2018)", + "19. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)", + "20. Furnari, A., Battiato, S., Grauman, K., Farinella, G.M.: Next-active-object prediction from egocentric videos. Journal of Visual Communication and Image Representation 49, 401-411 (2017)", + "21. Furnari, A., Battiato, S., Maria Farinella, G.: Leveraging uncertainty to rethink loss functions and evaluation measures for egocentric action anticipation. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)", + "22. Furnari, A., Farinella, G.M.: What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6252-6261 (2019)", + "23. Furnari, A., Farinella, G.M.: Rolling-unrolling lstms for action anticipation from first-person video. IEEE transactions on pattern analysis and machine intelligence 43(11), 4021-4036 (2020)", + "24. Geng, Y., An, B., Geng, H., Chen, Y., Yang, Y., Dong, H.: End-to-end affordance learning for robotic manipulation. arXiv preprint arXiv:2209.12941 (2022)", + "25. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021)", + "26. Gong, D., Lee, J., Kim, M., Ha, S.J., Cho, M.: Future transformer for long-term action anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3052-3061 (2022)", + "27. Gong, J., Foo, L.G., Fan, Z., Ke, Q., Rahmani, H., Liu, J.: Diffpose: Toward more reliable 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13041-13051 (2023)", + "28. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022)", + "29. Gupta, A., Johnson, J., Fei-Fei, L., Savarese, S., Alahi, A.: Social gan: Socially acceptable trajectories with generative adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2255-2264 (2018)", + "30. Jiang, H., Liu, S., Wang, J., Wang, X.: Hand-object contact consistency reasoning for human grasps generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11107-11116 (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Judd, T., Ehinger, K., Durand, F., Torralba, A.: Learning to predict where humans look. In: 2009 IEEE 12th international conference on computer vision. pp. 2106-2113. IEEE (2009)", + "32. Kjellström, H., Romero, J., Kragic, D.: Visual object-action recognition: Inferring object affordances from human demonstration. Computer Vision and Image Understanding 115(1), 81-90 (2011)", + "33. Lee, N., Choi, W., Vernaza, P., Choy, C.B., Torr, P.H., Chandraker, M.: Desire: Distant future prediction in dynamic scenes with interacting agents. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 336-345 (2017)", + "34. Li, C., Zheng, P., Yin, Y., Pang, Y.M., Huo, S.: An ar-assisted deep reinforcement learning-based approach towards mutual-cognitive safe human-robot interaction. Robotics and Computer-Integrated Manufacturing 80, 102471 (2023)", + "35. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Wang, L., Qiao, Y.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022)", + "36. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018)", + "37. Liu, M., Tang, S., Li, Y., Rehg, J.M.: Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 704-721. Springer (2020)", + "38. Liu, S., Jiang, H., Xu, J., Liu, S., Wang, X.: Semi-supervised 3d hand-object poses estimation with interactions in time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14687-14697 (2021)", + "39. Liu, S., Tripathi, S., Majumdar, S., Wang, X.: Joint hand motion and interaction hotspots prediction from egocentric videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3282-3292 (2022)", + "40. Liu, T., Lam, K.M.: A hybrid egocentric activity anticipation framework via memory-augmented recurrent and one-shot representation forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13904-13913 (2022)", + "41. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Learning affordance grounding from exocentric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2252-2261 (June 2022)", + "42. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Learning visual affordance grounding from demonstration videos. IEEE Transactions on Neural Networks and Learning Systems (2023)", + "43. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Leverage interactive affinity for affordance learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6809-6819 (2023)", + "44. Mandlekar, A., Xu, D., Martín-Martín, R., Savarese, S., Fei-Fei, L.: Learning to generalize across long-horizon tasks from human demonstrations. arXiv preprint arXiv:2003.06085 (2020)", + "45. Mangalam, K., An, Y., Girase, H., Malik, J.: From goals, waypoints & paths to long term human trajectory forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15233-15242 (2021)", + "46. Mangalam, K., Girase, H., Agarwal, S., Lee, K.H., Adeli, E., Malik, J., Gaidon, A.: It is not the journey but the destination: Endpoint conditioned trajectory" + ], + "bbox": [ + 215, + 147, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "prediction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 759-776. Springer (2020)", + "47. Mascaró, E.V., Ahn, H., Lee, D.: Intention-conditioned long-term human egocentric action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6048-6057 (2023)", + "48. Mendonca, R., Bahl, S., Pathak, D.: Structured world models from human videos. arXiv preprint arXiv:2308.10901 (2023)", + "49. Miech, A., Laptev, I., Sivic, J., Wang, H., Torresani, L., Tran, D.: Leveraging the present to anticipate the future in videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "50. Mohamed, A., Qian, K., Elhoseiny, M., Claudel, C.: Social-stgcnn: A social spatiotemporal graph convolutional neural network for human trajectory prediction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14424-14432 (2020)", + "51. Myers, A., Teo, C.L., Fermüller, C., Aloimonos, Y.: Affordance detection of tool parts from geometric features. In: 2015 IEEE International Conference on Robotics and Automation (ICRA). pp. 1374-1381. IEEE (2015)", + "52. Nagarajan, T., Feichtenhofer, C., Grauman, K.: Grounded human-object interaction hotspots from video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8688-8697 (2019)", + "53. Nawhal, M., Jyothi, A.A., Mori, G.: Rethinking learning approaches for long-term action anticipation. In: European Conference on Computer Vision. pp. 558-576. Springer (2022)", + "54. Parsaeifard, B., Saadatnejad, S., Liu, Y., Mordan, T., Alahi, A.: Learning decoupled representations for human pose forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2294-2303 (2021)", + "55. Peters, R.J., Iyer, A., Itti, L., Koch, C.: Components of bottom-up gaze allocation in natural images. Vision research 45(18), 2397-2416 (2005)", + "56. Plizzari, C., Goletto, G., Furnari, A., Bansal, S., Ragusa, F., Farinella, G.M., Damen, D., Tommasi, T.: An outlook into the future of egocentric vision. arXiv preprint arXiv:2308.07123 (2023)", + "57. Qi, H., Wang, X., Pathak, D., Ma, Y., Malik, J.: Learning long-term visual dynamics with region proposal interaction networks. arXiv preprint arXiv:2008.02265 (2020)", + "58. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3636-3645 (2023)", + "59. Rodin, I., Furnari, A., Mavroeidis, D., Farinella, G.M.: Predicting the future from first person (egocentric) vision: A survey. Computer Vision and Image Understanding 211, 103252 (2021)", + "60. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "61. Shen, J., Dudley, J., Kristensson, P.O.: Fast and robust mid-air gesture typing for ar headsets using 3d trajectory decoding. IEEE Transactions on Visualization and Computer Graphics (2023)", + "62. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Song, D., Kyriazis, N., Oikonomidis, I., Papazov, C., Argyros, A., Burschka, D., Kragic, D.: Predicting human intention in visual observations of hand/object interactions. In: 2013 IEEE International Conference on Robotics and Automation. pp. 1608-1615. IEEE (2013)", + "64. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019)", + "65. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. Advances in neural information processing systems 27 (2014)", + "66. Swain, M.J., Ballard, D.H.: Color indexing. International journal of computer vision 7(1), 11-32 (1991)", + "67. Tu, Z., Huang, Z., Chen, Y., Kang, D., Bao, L., Yang, B., Yuan, J.: Consistent 3d hand reconstruction in video via self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "68. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "69. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016)", + "70. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021)", + "71. Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8121-8130 (2022)", + "72. Yang, Y., Zhai, W., Luo, H., Cao, Y., Luo, J., Zha, Z.J.: Grounding 3d object affordance from 2d interactions in images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10905-10915 (2023)", + "73. Yang, Y., Zhai, W., Luo, H., Cao, Y., Zha, Z.J.: Lemon: Learning 3d human-object interaction relation from 2d images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16284-16295 (2024)", + "74. Ye, Y., Singh, M., Gupta, A., Tulsiani, S.: Compositional video prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10353-10362 (2019)", + "75. Zhai, W., Cao, Y., Zhang, J., Xie, H., Tao, D., Zha, Z.J.: On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "76. Zhai, W., Luo, H., Zhang, J., Cao, Y., Tao, D.: One-shot object affordance detection in the wild. International Journal of Computer Vision 130(10), 2472-2500 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 702 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Bidirectional Progressive Transformer for Interaction Intention Anticipation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_model.json b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8b9258135eae19eb8fe699a9820c22c7ae9aae3e --- /dev/null +++ b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_model.json @@ -0,0 +1,3043 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.267, + 0.141, + 0.74, + 0.187 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.212, + 0.772, + 0.229 + ], + "angle": 0, + "content": "Zichen Zhang\\(^{1}\\), Hongchen Luo\\(^{1}\\), Wei Zhai\\(^{1,*}\\), Yang Cao\\(^{1,2}\\), and Yu Kang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.239, + 0.664, + 0.254 + ], + "angle": 0, + "content": "1 University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.254, + 0.774, + 0.296 + ], + "angle": 0, + "content": "\\(^{2}\\) Institute of Artificial Intelligence, Hefei Comprehensive National Science Center {zhangzichen@mail., lhc12@mail., wzhai056@, kangduyu@, forrest@}ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.328, + 0.744, + 0.646 + ], + "angle": 0, + "content": "Abstract. Interaction intention anticipation aims to jointly predict future hand trajectories and interaction hotspots. Existing research often treated trajectory forecasting and interaction hotspots prediction as separate tasks or solely considered the impact of trajectories on interaction hotspots, which led to the accumulation of prediction errors over time. However, a deeper inherent connection exists between hand trajectories and interaction hotspots, which allows for continuous mutual correction between them. Building upon this relationship, a novel Bidirectional prOgressive Transformer (BOT), which introduces a Bidirectional Progressive mechanism into the anticipation of interaction intention is established. Initially, BOT maximizes the utilization of spatial information from the last observation frame through the Spatial-Temporal Reconstruction Module, mitigating conflicts arising from changes of view in first-person videos. Subsequently, based on two independent prediction branches, a Bidirectional Progressive Enhancement Module is introduced to mutually improve the prediction of hand trajectories and interaction hotspots over time to minimize error accumulation. Finally, acknowledging the intrinsic randomness in human natural behavior, we employ a Trajectory Stochastic Unit and a C-VAE to introduce appropriate uncertainty to trajectories and interaction hotspots, respectively. Our method achieves state-of-the-art results on three benchmark datasets Epic-Kitchens-100, EGO4D, and EGTEA Gaze+, demonstrating superior in complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.659, + 0.721, + 0.674 + ], + "angle": 0, + "content": "Keywords: Interaction Intention \\(\\cdot\\) Egocentric Videos \\(\\cdot\\) Transformer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.697, + 0.377, + 0.714 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.819 + ], + "angle": 0, + "content": "The pursuit of future prediction on egocentric videos is centered on anticipating intentions and behaviors observed from a first-person perspective [59]. Interaction intention can be reflected in the trajectory of hand motion and contact points on the next-active object [39]. This task is novel but holds significant potential value, such as AR [34, 61], robot skill acquisition [3, 9, 24, 44, 48], hand reconstruction [30, 67] and pose estimation [27, 38]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.826, + 0.395, + 0.841 + ], + "angle": 0, + "content": "*Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.142, + 0.432, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.142, + 0.549, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.142, + 0.688, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.226, + 0.653, + 0.24 + ], + "angle": 0, + "content": "Error Accumulation Over Time" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.286 + ], + "angle": 0, + "content": "Fig. 1: Error accumulation when only considering the influence from hand trajectories to interaction hotspots. Red ones represent the ground truth, while Blue ones are predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.79, + 0.518 + ], + "angle": 0, + "content": "Most existing relevant research treated hand trajectory forecasting and interaction hotspots prediction as independent tasks. Bao et al. [4] proposed to predict the 3D hand trajectory from a first-person view. However, They overlooked the guiding role of interactable objects on trajectories. The prediction process was incomplete, leading to significant errors at the end of trajectories. On the other hand, there was extensive research on interaction hotspots [18,32,41,42,51,52,63]. Nevertheless, these studies often focused on the characteristics of the objects themselves, resulting in lower integration with subjective human behaviors in complex scenes and less accurate predictions. Other studies [37,39] jointly predicted future hand trajectories and interaction hotspots, improving their integration. However, these studies only considered the unidirectional influence of hand trajectories on contact points. Consequently, errors accumulated continuously with the prediction time steps (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.79, + 0.702 + ], + "angle": 0, + "content": "In fact, this one-way causation was insufficient, as it disregarded the spatial correlation between the two factors. Conversely, there exists a mutual relationship between the predictions of hand trajectories and interaction hotspots. On one hand, the derived directions of hand trajectories suggest the approximate areas where contact points occur. On the other hand, there is an inherent connection between the spatial distribution of contact points and hand trajectory patterns, as in Fig. 2, which implies that examining the distribution of contact points can provide insights into the development of hand trajectories. Leveraging this relationship, to mitigate the process of error accumulation, we establish a Bidirectional Progressive (Bi-Progressive) mechanism, allowing continuous mutual refinement between trajectories and interaction hotspots throughout the entire prediction period, improving the accuracy of joint predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.79, + 0.842 + ], + "angle": 0, + "content": "In this paper, an innovative Bidirectional prOgressive Transformer (BOT) is proposed, introducing the Bi-Progressive mechanism mentioned earlier between two independent predicting branches. Firstly, the raw video input undergoes a Spatial-Temporal Reconstruction Module, which integrates the spatial information from the last observation frame and the temporal information among all observation frames. This process mitigates spatial information conflicts arising from changes in the field of view in first-person videos. Secondly, we establish separate dual branches for hand trajectory forecasting and interaction hotspots prediction, endowing them with the capability to independently perform indi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.144, + 0.422, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.144, + 0.514, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.144, + 0.605, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.145, + 0.671, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.194, + 0.601, + 0.204 + ], + "angle": 0, + "content": "(a) Long-length trajectories" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.204, + 0.422, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.204, + 0.514, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.204, + 0.605, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.204, + 0.671, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.254, + 0.612, + 0.263 + ], + "angle": 0, + "content": "(b) Medium-length trajectories" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.264, + 0.422, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.264, + 0.514, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.264, + 0.605, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.264, + 0.671, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.314, + 0.604, + 0.325 + ], + "angle": 0, + "content": "(c) Short-length trajectories" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.328, + 0.785, + 0.357 + ], + "angle": 0, + "content": "Fig. 2: Inherent connection between different hand trajectory categories and their corresponding contact points distribution. This is conducted on Epic-Kitchens-100 [14]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.505 + ], + "angle": 0, + "content": "vidual prediction tasks, laying the foundation for introducing the Bi-Progressive mechanism. Subsequently, we introduce cross-attention blocks as the primary implementation of the Bi-Progressive mechanism. At each time step, the predictions of hand trajectories and interaction hotspots are alternately updated, addressing inherent errors in individual predictions. Finally, following the logic of natural human behaviors [50, 57, 74], we incorporate a Trajectory Stochastic Unit and a C-VAE for both hand trajectories and interaction hotspots, respectively, introducing appropriate uncertainties to the final prediction results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.567 + ], + "angle": 0, + "content": "We evaluate our approach on Epic-Kitchens-100 [14], EGO4D [28] and EGTEA Gaze+ [36] datasets. BOT demonstrates notable advantages in predicting both hand trajectories and interaction hotspots, achieving superior performance in anticipating interaction intentions. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.577, + 0.785, + 0.621 + ], + "angle": 0, + "content": "- A novel BOT is introduced to jointly predict future hand trajectories and interaction hotspots in egocentric videos, attaining the optimal predictive performance with more lenient input conditions." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.622, + 0.785, + 0.667 + ], + "angle": 0, + "content": "- We introduce a novel Bidirectional Progressive mechanism that enhances the coherence between hand trajectories and interaction hotspots, thereby improving the completeness of the interaction intention." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.667, + 0.787, + 0.726 + ], + "angle": 0, + "content": "- The experimental results demonstrate that our method achieves state-of-the-art results across three benchmark datasets Epic-Kitchens-100 [14], EGO4D [28], and EGTEA Gaze+ [36], demonstrating superior in complex scenarios." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.577, + 0.787, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.749, + 0.388, + 0.765 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Action anticipation in egocentric videos. Anticipating actions in an egocentric vision is a challenging task that includes short-term and long-term action anticipation [56,59]. Short-term action anticipation aims to predict the next action in the near future [21-23,25,49], while long-term action anticipation focuses" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "on the sequence of subsequent actions performed by the camera wearer [26, 40, 47, 53]. These endeavors contribute significantly to the comprehension of actions in first-person videos. However, concerning interaction intention, action anticipation still yields information of relatively high dimensionality. Anticipating a semantic label does not furnish direct guidance for the robot, as it lacks specific instructions on task execution. Comparatively, hand trajectories and interaction hotspots offer a more intuitive representation of interaction intention, constituting a lower-dimensional depiction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.789, + 0.418 + ], + "angle": 0, + "content": "Human trajectory forecasting. The forecasting of human trajectories was extensively studied over the years [1,2,12,13,45,54]. However, most of them were based on third-person video datasets, falling short of capturing intentions related to human-object interactions. Since hands serve as the primary medium through which humans interact with the world, predicting hand trajectories in egocentric videos can aid embodied AI in comprehending human-object interaction intentions. The approaches in [4,37,39] addressed the issue of hand trajectory prediction in egocentric videos. However, they did not account for the influence of contact points with objects on trajectories. This limitation resulted in the continuous accumulation of prediction errors over time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.789, + 0.571 + ], + "angle": 0, + "content": "Interaction hotspots prediction. Interaction hotspots prediction aims to locate where the hand-object interaction happens. Some research aimed to predict the next interactive object [5, 6, 15, 17, 20], going further, some work anticipated specific interaction regions on objects [42, 43, 52, 72, 73, 75, 76], yielding more precise information for interaction. Differing from image input, our task aims to generate more precise predictions regarding the regions of contact points by utilizing egocentric video inputs. The incorporation of first-person videos is intended to focus on precisely localizing interaction hotspots for distinct human motions. In this task, we leverage the predicted hand trajectories to improve the accuracy of interaction hotspots predictions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.331, + 0.61 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Problem Setup. Given an original input video \\(\\mathcal{I} = \\{I_1, I_2, \\dots, I_N\\}\\), where \\(I_{\\mathrm{N}}\\) is the last observation frame, the task is predicting future hand trajectories \\(\\mathcal{H} = \\{h_{\\mathrm{N + 1}}, h_{\\mathrm{N + 2}}, \\dots, h_{\\mathrm{N + T - 1}}\\}\\) in future keyframes \\(\\{I_{\\mathrm{N + 1}}, I_{\\mathrm{N + 2}}, \\dots, I_{\\mathrm{N + T - 1}}\\}\\) and the contact point \\(\\mathcal{O}_{\\mathrm{C}}\\) in the contact frame \\(I_{N + T}\\). We subsequently convert the contact point \\(\\mathcal{O}_{\\mathrm{C}}\\) into a heatmap by centering a Gaussian distribution, generating interaction hotspots \\(\\mathcal{O}_{\\mathrm{I}}\\). To ensure the continuity of the trajectory as well as the consistency of the background, both \\(\\mathcal{H}\\) and \\(\\mathcal{O}_{\\mathrm{I}}\\) are projected onto the last observation frame, which enhances the overall coherence of the predictions [39]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.334, + 0.784 + ], + "angle": 0, + "content": "3.1 Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Overview. As shown in Fig. 3, given the original input video \\(\\mathcal{I}\\), we aim to predict the future hand trajectory \\(\\mathcal{H}\\) and interaction hotspots \\(\\mathcal{O}_{\\mathrm{I}}\\). Initially, to alleviate information conflicts arising from changes in the field of view, the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.228, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.322, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Fig. 3: Overview of the Bidirectional Progressive Transformer. It follows a dual-branch structure with a Bi-Progressive Enhancement Module between them to anticipate future hand trajectories and interaction hotspots." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.672 + ], + "angle": 0, + "content": "original input video undergoes a Spatial-Temporal Reconstruction Module that integrates spatial information from the last observation frame and temporal information from the entire video, producing spatial features for each future timestep. Subsequently, these features are separated, with the last one serving as the contact feature while the preceding ones are for trajectory features. These features are then fed into the interaction hotspots and trajectory prediction branches, respectively. The former employs an overlap patch embedding layer to enhance attention at the center of the field of view, while the latter introduces temporal order to trajectory features through Cross-Transformers [11]. Simultaneously, we employ a Bi-Progressive Enhancement Module consisting of cross-attention blocks between dual branches, introducing a Bi-Progressive mechanism to the joint prediction. This alternately and iteratively corrects the features of trajectories and interaction hotspots in chronological order, thereby mitigating the accumulation of prediction errors over time. Finally, considering the inherent uncertainty in natural human behaviors [50, 57, 74], we introduce a Stochastic Unit for trajectory forecasting while employing a C-VAE [62] for interaction hotspots, adding appropriate uncertainty during the completion of position decoding to achieve comprehensive predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Spatial-Temporal Reconstruction. When dealing with egocentric videos, both 3D convolutions [19] and self-attention mechanisms [68] struggle to address spatial information conflicts arising from changes in views. Moreover, predictions are projected onto the last observation frame. Taking these factors into account, we establish the Spatial-Temporal Reconstruction Module. On one hand, we employ optical flow containing more comprehensive motion information as the temporal input, on the other hand, we isolate the last observation frame as the spatial input. Subsequently, the temporal input is processed by a pre-trained Temporal Segment Network [22,69], extracting low-dimensional temporal features \\(\\mathbf{X}\\), and the spatial input is passed through a Segformer-b2 [70], extracting spatial features \\(\\mathbf{S}\\) from the last observation frame. \\(\\mathbf{X}\\) and \\(\\mathbf{S}\\) are then fed into a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.61, + 0.163 + ], + "angle": 0, + "content": "Cross-Transformer [11]. The \\(i\\)-th block is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.177, + 0.786, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {0} = \\mathbf {s} _ {\\text {p a t c h}} + \\mathbf {s} _ {\\text {p o s}}, \\quad \\mathbf {x} = \\mathbf {x} + \\mathbf {x} _ {\\text {p o s}}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.195, + 0.786, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {i} = \\mathbf {s} _ {i - 1} + \\mathbf {M C A} (\\mathbf {L N} (\\mathbf {s} _ {i - 1}), \\mathbf {L N} (\\mathbf {x})), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.215, + 0.786, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {i} = \\mathbf {y} _ {i} + \\mathbf {F F N} (\\mathbf {L N} (\\mathbf {y} _ {i})). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Where \\(\\mathbf{s}_0\\) and \\(\\mathbf{x}\\) represent the original spatial and temporal inputs, respectively. \\(\\mathbf{s}_{patch}\\) is encoded by overlap patch embedding layers from spatial features \\(\\mathbf{S}\\), enhancing the network's focus on the center of the field of view. \\(\\mathbf{x}\\) is encoded by positional embedding layers from temporal features \\(\\mathbf{X}\\), improving the overall sequentiality. Furthermore, FFN refers to the Feed Forward Layer, comprising a two-layer MLP with an expanding ratio \\(r\\), and Layer normalization (LN) is applied before each block. MCA stands for Multi-head Cross-Attention, where query, key, and value are processed through distinct linear layers: MCA(X,Y) = Attention(W^QX, W^KY, W^VY). The output of this module corresponds to the spatial features for each prediction frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.787, + 0.439 + ], + "angle": 0, + "content": "Interaction Hotspots Prediction Branch. Due to the spatial distribution of contact points closer to the central field of view, the contact feature \\(\\mathbf{F}_{\\mathrm{C}}\\) undergoes processing through an overlap patch embedding layer:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.453, + 0.787, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {o}} = \\text {o v e r l a p} (\\mathbf {F} _ {\\mathrm {C}}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.589, + 0.495 + ], + "angle": 0, + "content": "Where \\(\\mathbf{F}_C^0\\) represents the processed contact feature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.789, + 0.587 + ], + "angle": 0, + "content": "Hand Trajectory Prediction Branch. We employ original trajectory features as primary features and utilize features from the previously predicted frame as conditional features, subsequently merging them through the cross-attention block \\( f \\), ensuring the coherence of the trajectory over time. Assuming the trajectory features are \\( \\{\\mathbf{F}_{\\mathrm{N} + 1},\\mathbf{F}_{\\mathrm{N} + 2},\\dots ,\\mathbf{F}_{\\mathrm{N} + \\mathrm{T} - 1}\\} \\) and the feature map of the last observation frame is \\( \\mathbf{F}_{\\mathrm{N}} \\), they are updated in chronological order:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.601, + 0.787, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {t} ^ {\\mathrm {o}} = f \\left(\\mathbf {F} _ {\\mathrm {t}}, \\mathbf {F} _ {\\mathrm {t} - 1}\\right), \\quad t \\in [ \\mathrm {N} + 1, \\mathrm {N} + \\mathrm {T} - 1 ]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Where \\(\\mathbf{F}_t^{\\mathrm{o}}\\) stands for the updated trajectory feature, \\(t\\) is the prediction time step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.841 + ], + "angle": 0, + "content": "H-O Bi-Progressive Enhancement. Expanding on the groundwork laid by two distinct prediction branches, and to mitigate the accumulation of errors as predictions unfold over time, we introduce a Bi-Progressive Enhancement Module (Fig. 4), which alternately corrects the predictions of trajectories \\(\\mathcal{H}\\) and interaction hotspots \\(\\mathcal{C}_{\\mathrm{I}}\\) at each time step \\(t\\). Specifically, owing to observation frames and the continuity of hand trajectories, the trajectory feature \\(\\mathbf{F}_{\\mathrm{N} + 1}^{\\mathrm{o}}\\) at the initial predicted time step \\(\\mathrm{N} + 1\\) demonstrates a high level of accuracy. Exploiting this, we enhance the precision of the contact feature \\(\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{o}}\\) through the cross-attention block \\(f\\), thereby minimizing the error with the ground truth. Moving to the subsequent time step \\(\\mathrm{N} + 2\\), we employ the first refined contact feature \\(\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{r}_1}\\) from the previous time step \\(\\mathrm{N} + 1\\) as a condition to correct the trajectory feature \\(\\mathbf{F}_{\\mathrm{N} + 2}^{\\mathrm{o}}\\). Following this, the corrected trajectory feature is utilized to further refine the contact feature. This iterative correction process persists until all trajectory" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "features have undergone a complete round of updates. It can be represented as a recursive procedure:" + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.191, + 0.786, + 0.208 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {N} + 1} ^ {\\mathrm {r}} = \\mathbf {F} _ {\\mathrm {N} + 1} ^ {\\mathrm {o}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.21, + 0.787, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {1}} = f \\left(\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {o}}, \\mathbf {F} _ {\\mathrm {N} + 1} ^ {\\mathrm {r}}\\right), \\quad \\mathbf {F} _ {\\mathrm {N} + 2} ^ {\\mathrm {r}} = f \\left(\\mathbf {F} _ {\\mathrm {N} + 2} ^ {\\mathrm {o}}, \\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {1}}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.487, + 0.233, + 0.51, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\cdot \\cdot \\cdot\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.248, + 0.786, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {3}} = f \\left(\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {2}}, \\mathbf {F} _ {\\mathrm {N} + 3} ^ {\\mathrm {r}}\\right), \\quad \\mathbf {F} _ {\\mathrm {N} + 4} ^ {\\mathrm {r}} = f \\left(\\mathbf {F} _ {\\mathrm {N} + 4} ^ {\\mathrm {o}}, \\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {3}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.591, + 0.367 + ], + "angle": 0, + "content": "Where \\(\\mathbf{F}_t^{\\mathrm{r}}\\) represents the refined trajectory feature map at time step \\(t\\in [\\mathrm{N} + 1,\\mathrm{N} + \\mathrm{T} - 1]\\), \\(\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{r},i}\\) denotes the contact feature after the \\(i\\)-th refinement. Via this module, we attain refined features for both trajectories and interaction hotspots, significantly reducing error accumulation in predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.591, + 0.61 + ], + "angle": 0, + "content": "Interaction Hotspots Decoder. Given the presence of inherent uncertainty associated with interaction hotspots [39, 41, 42], we employ a C-VAE [62] as the decoder. Due to the excessive randomness associated with directly decoding the coordinates of contact points, we instruct the C-VAE to output a feature map and select the location of the maximum value from this map as the output, thereby confining the uncertainty within a reasonable range. The C-VAE follows an encoder-decoder structure. For the encoder block, we concatenate the input \\(\\varGamma\\) with the condition \\(\\mathbf{C}\\) through an encoding function \\(\\mathbf{P}_{\\mathrm{e}}\\), obtaining features in latent space \\(\\mathbf{Z}\\) which parameterized by mean \\(\\mu\\) and co-variance \\(\\sigma\\). In this context, the input \\(\\varGamma\\) is the ground truth heatmap, while the condition \\(\\mathbf{C}\\) is the refined contact feature \\(\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{r}_3}\\). In the decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.787, + 0.668 + ], + "angle": 0, + "content": "block, the first step involves sampling \\(\\mathbf{Z}\\) from the latent space and concatenating it with condition \\(\\mathbf{C}\\). Following that, the output heatmap \\(\\hat{\\Gamma}\\) is reconstructed by a decoding function \\(\\mathbf{P}_d\\). In summary, the entire process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.671, + 0.787, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mu , \\sigma = \\mathbf {P} _ {\\mathrm {e}} (\\Gamma , \\mathrm {C}), \\quad \\hat {\\Gamma} = \\mathbf {P} _ {\\mathrm {d}} (\\mathrm {Z}, \\mathrm {C}), \\quad \\mathrm {Z} \\sim \\mathcal {N} (\\mu , \\sigma). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Where \\(\\mathcal{N}\\) represents a normal distribution, \\(\\hat{\\Gamma}\\) is the reconstructed heatmap. Both encoding and decoding blocks consist of two MLP layers. We introduce KL-Divergence and the reconstruction error as the overall loss:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.756, + 0.787, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} (\\Gamma , \\hat {\\Gamma}) = (\\Gamma - \\hat {\\Gamma}) ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.776, + 0.787, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k l} (\\mu , \\sigma) = - \\mathbf {K L} \\left[ \\mathcal {N} (\\mu , \\sigma) ^ {2} | | \\mathcal {N} (0, 1) \\right], \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.797, + 0.787, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda \\mathcal {L} _ {k l}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.438, + 0.841 + ], + "angle": 0, + "content": "In this paper, we set \\(\\lambda = 0.01\\)." + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.308, + 0.787, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.513, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Fig. 4: Bi-Progressive Enhancement Module. The output feature map is enclosed in purple boxes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.145, + 0.725, + 0.16 + ], + "angle": 0, + "content": "\\(h_t\\) hand position \\(\\mathcal{O}_{\\mathrm{C}}\\) contact point \\(\\mathcal{R}_{21}, \\mathcal{R}_{22}\\) elliptical region \\(\\mathcal{R}_2\\) uncertain region \\(\\theta\\) ellipse tilt angle" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.161, + 0.734, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.248, + 0.789, + 0.278 + ], + "angle": 0, + "content": "Fig. 5: Uncertain regions of the hand trajectory. It is derived through the intersection of two ellipses, corresponding to the Eq. 14-19. (i.e., the purple region)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.79, + 0.515 + ], + "angle": 0, + "content": "Hand Trajectory Decoder. Similar to interaction hotspots, hand trajectories also exhibit uncertainty [29, 33, 46]. Given that trajectory feature maps with distinct distribution characteristics have been obtained through the Bi-Progressive Enhancement Module, in conjunction with sampling efficiency, we don't utilize C-VAE, rather, we introduce a Trajectory Stochastic Unit to incorporate uncertainty into trajectory forecasting by calculating uncertain regions for each prediction time step instead. Since optimizing directly using point coordinates for loss calculation can pose challenges, we initially introduce a Gaussian distribution for each ground truth hand position to generate the ground truth heatmap. Subsequently, we calculate the loss between the corrected feature map and the generated ground truth heatmap, thereby enabling the network to learn the distribution of potential hand positions. The associated loss function \\(\\mathcal{L}_h\\) is based on MSE Loss:" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.535, + 0.788, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h} = \\frac {1}{T - 1} \\sum_ {t = N + 1} ^ {N + T - 1} \\left(\\mathbf {F} _ {t} ^ {\\mathrm {r}} - \\mathbf {G} _ {t}\\right) ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.6, + 0.789, + 0.707 + ], + "angle": 0, + "content": "Where \\(\\mathbf{G}_t\\) represents the ground truth heatmap at time step \\(t\\). We designate the coordinates of the maximum value within each feature map as the predicted trajectory points, which are then fed into the stochastic unit to generate uncertain regions. The size of the region is determined by the time step as well as the length of trajectory segments in preceding and subsequent stages. The shape is constrained by the original image dimensions and the orientation of the trajectory segments. Formally, the uncertain region (Fig. 5) can be shown as:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.731, + 0.788, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {t} \\sim (t, h _ {t - 1}, h _ {t}, h _ {t + 1}, \\mathrm {W}, \\mathrm {L}). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.77, + 0.788, + 0.801 + ], + "angle": 0, + "content": "Where \\(\\mathcal{R}_t\\) is the uncertain region in future time step \\(t\\), W and L are the width and length of the input video, respectively. \\(\\mathcal{R}_t\\) satisfies the following constraints:" + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.825, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {t} = \\mathcal {R} _ {t 1} \\cap \\mathcal {R} _ {t 2}. \\tag {15}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Where \\(\\mathcal{R}_{t1}\\) and \\(\\mathcal{R}_{t2}\\) are elliptical regions constrained by \\(h_{t-1}\\) and \\(h_{t+1}\\), which can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.305, + 0.192, + 0.786, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\frac {\\left(\\mathbf {x} _ {t} \\cos \\theta_ {t \\omega} - \\mathbf {y} _ {t} \\sin \\theta_ {t \\omega}\\right) ^ {2}}{\\mathbf {a} _ {t \\omega} ^ {2}} + \\frac {\\left(\\mathbf {x} _ {t} \\sin \\theta_ {t \\omega} + \\mathbf {y} _ {t} \\cos \\theta_ {t \\omega}\\right) ^ {2}}{\\mathbf {b} _ {t \\omega} ^ {2}} \\leq 1, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.228, + 0.786, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} _ {t \\omega} = \\alpha \\beta^ {t - 1} \\mathbf {d} ^ {\\frac {1}{2}} \\left(h _ {t + \\omega - 1}, h _ {t + \\omega - 2}\\right), \\quad \\mathbf {b} _ {t \\omega} = \\sqrt {\\frac {\\mathrm {L}}{\\mathrm {W}}} \\mathbf {a} _ {t \\omega}, \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.264, + 0.786, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t \\omega} = \\tan^ {- 1} (\\mathbf {k} \\left(h _ {t + \\omega - 1}, h _ {t + \\omega - 2}\\right)), \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.284, + 0.786, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\omega \\in [ 1, 2 ], t \\in [ 1, 2, 3, 4 ], \\mathcal {O} _ {\\mathrm {C}} = h _ {5}. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.316, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Specifically, \\(\\alpha\\) denotes the overall scale of \\(\\mathcal{R}_t\\), \\(\\beta\\) signifies the temporal growth rate, \\(\\theta\\) stands for the ellipse tilt angle, \\((\\mathbf{x}_t,\\mathbf{y}_t)\\) represent the coordinates of \\(h_t\\). Additionally, \\(\\mathbf{k}(\\mathrm{X},\\mathrm{Y})\\) and \\(\\mathbf{d}(\\mathrm{X},\\mathrm{Y})\\) represent the slope and Euclidean distance between points \\(\\mathrm{X}\\) and \\(\\mathrm{Y}\\), respectively. In BOT, we set \\(\\alpha = 1.5\\) and \\(\\beta = 2\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.408, + 0.451, + 0.423 + ], + "angle": 0, + "content": "3.2 Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.442, + 0.784, + 0.473 + ], + "angle": 0, + "content": "Training. During training, we freeze parameters in TSN [69]. Training loss combines both trajectory loss and interaction hotspots loss:" + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.49, + 0.786, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {h} + \\varphi \\mathcal {L} _ {o}. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.418, + 0.535 + ], + "angle": 0, + "content": "In this paper, we set \\(\\varphi = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.537, + 0.784, + 0.583 + ], + "angle": 0, + "content": "Inference. We use a Trajectory Stochastic Unit and a C-VAE to introduce uncertainty during inference. Following previous work [39, 46], we report the minimum among 20 samples for trajectory evaluation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.376, + 0.632 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.655, + 0.334, + 0.669 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conducted experiments on three first-person video datasets Epic-Kitchen s-100 [14], EGO4D [28], and EGTEA Gaze+ [36]. Epic-Kitchens-100 [14] and EGTEA Gaze+ [36] predominantly include kitchen scenes, while EGO4D [28] encompasses a broader range of activities such as handicrafts, painting, and outdoor tasks. For each dataset, the observation time is set to 2 seconds. We configure the prediction time horizon T to 5, corresponding to 4 trajectory points and one contact point for each hand. Specifically, we predict 1-second trajectories on Epic-Kitchens-100 [14] and EGO4D [28], while on EGTEA Gaze+ [36], 0.5-second trajectories are forecast. Additional dataset information is available in the supplementary materials." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.458, + 0.162 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.383 + ], + "angle": 0, + "content": "In EGO4D [28], We use the method [39] to automatically generate hand trajectories, while contact points are manually annotated based on the last observation frame and the contact frame. We employ GMFLOW [71] to generate optical flow from EGO4D [28] and EGTEA Gaze+ [36]. For each dataset, we resize the original videos to \\( 456 \\times 256 \\) size, extracting both trajectory and contact features of size \\( 114 \\times 64 \\). In the Spatial-Temporal Reconstruction Module, we set the dimension \\( D \\) in Cross-Transformer to 1024, with depth \\( = 6 \\) and attention head number \\( = 8 \\). For the trajectory forecasting branch and H-O Bi-Progressive Enhancement Module, we configure the dimension \\( D \\) to 256, with depth \\( = 2 \\) and head number \\( = 8 \\). Additionally, for the introduced C-VAE, we set the dimension of both input and condition to \\( 114 \\times 64 \\), while the hidden and latent sizes are set to 1024 and 2048, respectively. Our method is implemented by PyTorch and trained with the AdamW optimizer. We train the model for 100 epochs on 4 NVIDIA 3090 GPUs with an initial learning rate of \\( 3e - 5 \\) and a batch size of 40." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.79, + 0.521 + ], + "angle": 0, + "content": "For comparison, We first select methods for individual tasks including Kalman Filter (KF) [8], Seq2Seq [65], Hotspots [52] and USST [4]. Additionally, the methods for image/video understanding are chosen as the compared methods, which lack a joint mechanism. The former includes VIT [16], HRNet [64], UNet [60], Segformer [70], while the latter contains I3D [10], Slowfast [19], Timesformer [7], Uniformer-V2 [35], StillFast [58]. Furthermore, we select methods with high task relevance and joint design, including FHOI [37] and OCT [39]. We conduct experiments using these approaches on Epic-kithens-100 [14], EGO4D [28], and EGTEA Gaze+ [36] datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.542, + 0.421, + 0.556 + ], + "angle": 0, + "content": "4.3 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Hand Trajectory Evaluation. The predicted hand trajectories are referenced to the last observation frames. ADE [39] and FDE [39] are chosen as hand trajectory evaluation metrics, which respectively represent the overall accuracy and endpoint accuracy of trajectories. Detailed descriptions are available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Interaction Hotspots Evaluation. For all datasets, we downsample each predicted heatmap to the size of \\(32 \\times 32\\). We choose SIM [66], AUC-J [31], and NSS [55] as interaction hotspots evaluation metrics. For a detailed explanation of the metrics, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.611, + 0.74 + ], + "angle": 0, + "content": "4.4 Quantitative and Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As shown in Tab. 1, our method achieves state-of-the-art performance in both trajectory prediction and interaction hotspots anticipation, with comparable leading margins across all three datasets. Taking the EGO4D dataset [28] with the richest variety of scenes as an example. In trajectory prediction, to ensure fair comparison, we additionally present results for OCT and BOT with only one sampling iteration. Under this single-sampling condition, BOT outperforms" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1: Results of different models. Best results are depicted using red (sample 20 times) and blue (sample once), respectively. unilateral, non-joint and joint denote prediction approaches for individual tasks, for two tasks without a collaborative design, and with a joint module, respectively. “-” denotes the vacancy of the value. “*” represents the results obtained from 20 random samplings. ↑ and ↓ indicate a better performance when the value of the metric increases and decreases, respectively." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.232, + 0.785, + 0.471 + ], + "angle": 0, + "content": "
MethodEK100 [14]EGO4D [28]EGTEA [36]
ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑
unilateralCenter0.240.240.090.610.330.260.250.070.570.150.320.320.090.630.27
KF [8]0.330.32--------0.490.48---
Seq2Seq [65]0.180.14--------0.180.14---
Hotspots [52]--0.150.660.53-------0.150.710.69
USST [4]0.170.18---0.180.18---0.180.18---
non-jointI3D [10]0.220.230.130.560.460.220.220.120.570.420.230.220.130.600.49
SLOWFAST [19]0.220.240.140.620.470.230.220.130.580.440.210.230.140.580.46
Timesformer [7]0.230.230.140.600.420.240.240.120.590.400.210.220.140.580.45
Uniformer-V2 [35]0.200.200.130.590.430.210.220.120.620.400.200.210.130.580.46
StillFast [58]0.180.190.160.620.610.190.190.150.650.570.180.190.160.600.55
UNet [60]0.190.180.150.620.570.180.180.140.590.550.180.180.140.600.60
HRNet [64]0.180.190.160.620.630.190.200.140.620.580.180.180.160.630.65
VIT [16]0.180.190.140.650.630.210.210.130.590.450.190.190.150.630.61
CrossVIT [11]0.190.200.140.620.590.190.200.130.600.550.200.210.130.590.49
Segformer-b2 [70]0.170.170.160.640.660.160.160.150.640.620.170.170.160.630.65
jointFHOI [37]0.360.350.100.550.39-----0.340.340.120.630.42
OCT [39]0.150.17---0.150.16---0.160.16---
OCT* [39]0.120.110.190.720.720.120.120.220.730.880.140.140.230.751.01
Ours0.120.14---0.110.13---0.120.14---
Ours*0.100.080.340.812.070.090.080.320.771.700.090.080.320.802.03
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.789, + 0.672 + ], + "angle": 0, + "content": "Segformer-b2, the best method in image/video understanding, with respective leads of \\(31.3\\%\\) and \\(18.8\\%\\) in ADE and FDE. Compared to OCT, BOT maintains advantages of \\(26.7\\%\\) and \\(18.8\\%\\) in ADE and FDE, respectively. These advantages persist even with 20 samplings, while the lead in FDE expands to \\(33.3\\%\\), underscoring the rationale behind BOT's stochastic sampling mechanism. Regarding interaction hotspots, we achieve the best results across all metrics. Specifically, BOT surpasses the second-best approach (OCT) by \\(45.5\\%\\) in SIM, \\(5.5\\%\\) in AUC-J, and \\(93.2\\%\\) in NSS, respectively. In summary, our method leverages the intrinsic correlations between hand trajectories and interaction hotspots more effectively, enhancing overall prediction accuracy of interaction intention, and exhibiting robustness across the three benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To further explore the subjective outcomes of models, we report the visualization results of three datasets, as shown in Fig. 6. Compared to other methods, our model provides more accurate prediction results in both hand trajectories and interaction hotspots. When there is significant movement of hands (column 1,3,8), the BOT's predictions exhibit the highest accuracy. It indicates that our Spatial-Temporal Reconstruction Module can efficiently utilize the human motion information to exploit the hand interaction intention and thus accurately predict the hand trajectory. In cases where there are no identifiable interactive objects in the field of view (column 6), our method anticipates more accurate interaction intentions. It illustrates that our Bi-Progressive mechanism can eliminate the uncertainty in the interaction process by mining the intrinsic connection between" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.144, + 0.784, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Fig. 6: Visualization of future hand trajectories and interaction hotspots on EK100, EGO4D and EGETA. We show the visualization results of our model, the joint prediction model (OCT) [39], the segmentation model (Segformer-b2) [70], and the video understanding model (Uniformer-V2) [35]. The green represents the trajectory of the left hand, while the red denotes the trajectory of the right hand." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.553 + ], + "angle": 0, + "content": "the hand trajectory and the contact points, which results in a more accurate inference in the interaction region and the hand trajectory. Moreover, our predictions closely align with the actual scenario when both hands simultaneously exhibit trajectories and interaction hotspots (columns 4,5), which suggests the strategy of differentiating between left and right hands during the prediction of interaction intentions, thereby mitigating cross-hand interference. More visualization results are available in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.39, + 0.59 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.779 + ], + "angle": 0, + "content": "Different joint approaches. To explore the impact of the Bi-Progressive mechanism on model performance, we experiment with five different joint configurations: Individual, Hand, Object, Bidirectional and Bi-Progressive. The results are shown in Tab. 2 and Fig. 7. It suggests that the network is challenging to accurately predict trajectories and interaction hotspots during individual predictions, and the separate incorporation of hands and objects can effectively improve the prediction accuracy of interaction hotspots. With the combination of hand and object information, the prediction performance of the trajectory and interaction region is greatly improved. However, there is still a gap compared to the Bi-Progressive mechanism. Further, Fig. 8 shows that the introduction of the progressive mechanism plays a pivotal role in mitigating the accumulation of prediction errors over time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Uncertainty in anticipation. Uncertainty is a crucial consideration in interaction intention anticipation. Following the logic of natural human behavior [50,57,74], both hand trajectories [29,33,46] and interaction hotspots [39,41,42] predictions inherently possess a degree of uncertainty. To investigate the impact" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 2: Results of different joint forecasting methods. Individual refers to separately predict two subtasks. Hand indicates that the hand trajectory is the condition of the interaction hotspots. Object stands for correcting the trajectory using predicted contact points. Bidirectional combines the methods Hand and Object." + }, + { + "type": "table", + "bbox": [ + 0.312, + 0.213, + 0.695, + 0.334 + ], + "angle": 0, + "content": "
MethodTrajectoryInteraction Hotspots
ADE ↓ FDE ↓SIM ↑ AUC-J ↑NSS ↑
Individual0.140.140.180.680.86
Hand0.130.120.250.721.19
Object0.120.120.230.701.04
Bidirectional0.100.090.280.731.48
Bi-Progressive0.090.080.320.771.70
" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.353, + 0.49, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.464, + 0.496, + 0.547 + ], + "angle": 0, + "content": "Fig. 7: Visualization results of different joint approaches. Based on the order from left to right and top to bottom, the corresponding methods are GT, Bi-Progressive, Bidirectional, Hand, Object, Individual." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.353, + 0.789, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.464, + 0.79, + 0.546 + ], + "angle": 0, + "content": "Fig. 8: The role of Progressive Mechanism. The red refers to the ground truth, while the blue represents the predictions. From top to bottom, Bi-Pro stands for Bi-Progressive while Bid refers to bidirectional." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.844 + ], + "angle": 0, + "content": "of uncertain region scale and the magnitude of uncertainty variations in hand trajectories, we modify \\(\\alpha\\) and \\(\\beta\\) in Eq. 17, with the resultant outcomes depicted in Fig. 9. \\(\\alpha\\) denotes the overall scale of the uncertain region, when \\(\\alpha\\) equals 0, the region's size is zero, and the network output yields a unique result. As \\(\\alpha\\) increases, the uncertainty of trajectories grows, leading to improved accuracy due to random samplings. However, given the limitations in sampling times, as \\(\\alpha\\) continues to increase, the accuracy naturally decreases. \\(\\beta\\) determines the magnitude of trajectory uncertainty increasing from front to back, when \\(\\beta\\) equals 1, each trajectory point has the same level of uncertainty. Considering the escalating complexity and uncertainty in trajectory prediction over time, an elevation in \\(\\beta\\) initially corresponds to augmented accuracy, followed by a subsequent decline. This gradual increase in uncertainty also renders the FDE more sensitive to \\(\\alpha\\) and \\(\\beta\\), as shown in Fig. 10a-10c. Regarding interaction hotspots, BOT exhibits relatively small uncertainty (Fig. 10d), as for a specific intention, the interaction area becomes increasingly influenced by object constraints. This is achieved by tightly controlling the uncertainty within a reasonably small range through the process of enabling C-VAE to generate feature maps and subsequently select" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.152, + 0.495, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.152, + 0.782, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.298, + 0.789, + 0.327 + ], + "angle": 0, + "content": "Fig. 9: Uncertainty in hand trajectory forecasting. As \\(\\alpha\\) and \\(\\beta\\) increase, they denote a heightened degree of overall trajectory uncertainty." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.339, + 0.311, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.392, + 0.448, + 0.403 + ], + "angle": 0, + "content": "(a) Long-length trajectories" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.339, + 0.403, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.339, + 0.495, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.404, + 0.312, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.457, + 0.45, + 0.468 + ], + "angle": 0, + "content": "(c) Short-length trajectories" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.404, + 0.403, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.403, + 0.495, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.339, + 0.601, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.339, + 0.692, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.339, + 0.784, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.392, + 0.748, + 0.403 + ], + "angle": 0, + "content": "(b) Medium-length trajectories" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.404, + 0.602, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.404, + 0.692, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.404, + 0.784, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.457, + 0.724, + 0.468 + ], + "angle": 0, + "content": "(d) interaction hotspots" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.479, + 0.789, + 0.522 + ], + "angle": 0, + "content": "Fig. 10: Uncertainty in anticipation. (a), (b) and (c) respectively depict long-length trajectories, medium-length trajectories, and short-length trajectories, while (d) illustrates the uncertainty of interaction hotspots." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.786, + 0.579 + ], + "angle": 0, + "content": "points, thereby ensuring the network's high confidence in predicting interaction regions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.355, + 0.615 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Conclusion. We plan to anticipate interaction intentions, encompassing the joint prediction of future hand trajectories and interaction hotspots with objects. We propose a novel BOT framework to establish an inherent connection between hand trajectories and contact points by introducing a Bi-Progressive mechanism, which eliminates the effect of error accumulation and achieves more accurate intention anticipation. Experimental results on three challenging benchmarks achieve state-of-the-art performance, thereby proving the superiority of our method in intention anticipation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Future Work. BOT currently predicts hand trajectories and interaction hotspots on 2D images. However, human-object interactions occur in 3D space. To enable the model to be applied to real-world scenarios, we will consider the geometric structure and spatial relationships between objects in the scene in the future, providing the model with the ability to anticipate interaction intentions for the real world." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.21 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62306295 and 62033012." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.249, + 0.324, + 0.265 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.786, + 0.323 + ], + "angle": 0, + "content": "1. Adeli, V., Adeli, E., Reid, I., Niebles, J.C., Rezatofighi, H.: Socially and contextually aware human motion and pose forecasting. IEEE Robotics and Automation Letters 5(4), 6033-6040 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.787, + 0.379 + ], + "angle": 0, + "content": "2. Adeli, V., Ehsanpour, M., Reid, I., Niebles, J.C., Savarese, S., Adeli, E., Rezatofighi, H.: Tripod: Human trajectory and pose dynamics forecasting in the wild. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13390-13400 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.786, + 0.407 + ], + "angle": 0, + "content": "3. Bahl, S., Gupta, A., Pathak, D.: Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.786, + 0.449 + ], + "angle": 0, + "content": "4. Bao, W., Chen, L., Zeng, L., Li, Z., Xu, Y., Yuan, J., Kong, Y.: Uncertainty-aware state space transformer for egocentric 3d hand trajectory forecasting. arXiv preprint arXiv:2307.08243 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.786, + 0.478 + ], + "angle": 0, + "content": "5. Bertasius, G., Park, H.S., Yu, S.X., Shi, J.: First person action-object detection with egonet. arXiv preprint arXiv:1603.04908 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.786, + 0.52 + ], + "angle": 0, + "content": "6. Bertasius, G., Soo Park, H., Yu, S.X., Shi, J.: Unsupervised learning of important objects from first-person videos. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1956-1964 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.786, + 0.547 + ], + "angle": 0, + "content": "7. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.786, + 0.589 + ], + "angle": 0, + "content": "8. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE international conference on image processing (ICIP). pp. 3464-3468. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.786, + 0.617 + ], + "angle": 0, + "content": "9. Bharadhwaj, H., Gupta, A., Tulsiani, S., Kumar, V.: Zero-shot robot manipulation from passive human videos. arXiv preprint arXiv:2302.02011 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.786, + 0.659 + ], + "angle": 0, + "content": "0. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.786, + 0.701 + ], + "angle": 0, + "content": "1. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.742 + ], + "angle": 0, + "content": "2. Chen, G., Chen, Z., Fan, S., Zhang, K.: Unsupervised sampling promoting for stochastic human trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17874-17884 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.786, + 0.784 + ], + "angle": 0, + "content": "3. Choi, C., Dariush, B.: Looking to relations for future trajectory forecast. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.786, + 0.84 + ], + "angle": 0, + "content": "4. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Scaling egocentric vision: The epic-kitchens dataset. In: European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.282, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Dessalene, E., Devaraj, C., Maynard, M., Fermuller, C., Aloimonos, Y.: Forecasting action through contact representations from first person video. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.787, + 0.245 + ], + "angle": 0, + "content": "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.288 + ], + "angle": 0, + "content": "17. Fan, C., Lee, J., Ryoo, M.S.: Forecasting hands and objects in future frames. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "18. Fang, K., Wu, T.L., Yang, D., Savarese, S., Lim, J.J.: Demo2vec: Reasoning object affordances from online videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2139-2147 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.787, + 0.374 + ], + "angle": 0, + "content": "19. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "20. Furnari, A., Battiato, S., Grauman, K., Farinella, G.M.: Next-active-object prediction from egocentric videos. Journal of Visual Communication and Image Representation 49, 401-411 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.787, + 0.471 + ], + "angle": 0, + "content": "21. Furnari, A., Battiato, S., Maria Farinella, G.: Leveraging uncertainty to rethink loss functions and evaluation measures for egocentric action anticipation. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.514 + ], + "angle": 0, + "content": "22. Furnari, A., Farinella, G.M.: What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6252-6261 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.557 + ], + "angle": 0, + "content": "23. Furnari, A., Farinella, G.M.: Rolling-unrolling lstms for action anticipation from first-person video. IEEE transactions on pattern analysis and machine intelligence 43(11), 4021-4036 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.586 + ], + "angle": 0, + "content": "24. Geng, Y., An, B., Geng, H., Chen, Y., Yang, Y., Dong, H.: End-to-end affordance learning for robotic manipulation. arXiv preprint arXiv:2209.12941 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.787, + 0.614 + ], + "angle": 0, + "content": "25. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.787, + 0.656 + ], + "angle": 0, + "content": "26. Gong, D., Lee, J., Kim, M., Ha, S.J., Cho, M.: Future transformer for long-term action anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3052-3061 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.659, + 0.787, + 0.699 + ], + "angle": 0, + "content": "27. Gong, J., Foo, L.G., Fan, Z., Ke, Q., Rahmani, H., Liu, J.: Diffpose: Toward more reliable 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13041-13051 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.701, + 0.787, + 0.755 + ], + "angle": 0, + "content": "28. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "29. Gupta, A., Johnson, J., Fei-Fei, L., Savarese, S., Alahi, A.: Social gan: Socially acceptable trajectories with generative adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2255-2264 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Jiang, H., Liu, S., Wang, J., Wang, X.: Hand-object contact consistency reasoning for human grasps generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11107-11116 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Judd, T., Ehinger, K., Durand, F., Torralba, A.: Learning to predict where humans look. In: 2009 IEEE 12th international conference on computer vision. pp. 2106-2113. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "32. Kjellström, H., Romero, J., Kragic, D.: Visual object-action recognition: Inferring object affordances from human demonstration. Computer Vision and Image Understanding 115(1), 81-90 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "33. Lee, N., Choi, W., Vernaza, P., Choy, C.B., Torr, P.H., Chandraker, M.: Desire: Distant future prediction in dynamic scenes with interacting agents. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 336-345 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "34. Li, C., Zheng, P., Yin, Y., Pang, Y.M., Huo, S.: An ar-assisted deep reinforcement learning-based approach towards mutual-cognitive safe human-robot interaction. Robotics and Computer-Integrated Manufacturing 80, 102471 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "35. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Wang, L., Qiao, Y.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "36. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "37. Liu, M., Tang, S., Li, Y., Rehg, J.M.: Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 704-721. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.509 + ], + "angle": 0, + "content": "38. Liu, S., Jiang, H., Xu, J., Liu, S., Wang, X.: Semi-supervised 3d hand-object poses estimation with interactions in time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14687-14697 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "39. Liu, S., Tripathi, S., Majumdar, S., Wang, X.: Joint hand motion and interaction hotspots prediction from egocentric videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3282-3292 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "40. Liu, T., Lam, K.M.: A hybrid egocentric activity anticipation framework via memory-augmented recurrent and one-shot representation forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13904-13913 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.647 + ], + "angle": 0, + "content": "41. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Learning affordance grounding from exocentric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2252-2261 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "42. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Learning visual affordance grounding from demonstration videos. IEEE Transactions on Neural Networks and Learning Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "43. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Leverage interactive affinity for affordance learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6809-6819 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.772 + ], + "angle": 0, + "content": "44. Mandlekar, A., Xu, D., Martín-Martín, R., Savarese, S., Fei-Fei, L.: Learning to generalize across long-horizon tasks from human demonstrations. arXiv preprint arXiv:2003.06085 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "45. Mangalam, K., An, Y., Girase, H., Malik, J.: From goals, waypoints & paths to long term human trajectory forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15233-15242 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "46. Mangalam, K., Girase, H., Agarwal, S., Lee, K.H., Adeli, E., Malik, J., Gaidon, A.: It is not the journey but the destination: Endpoint conditioned trajectory" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.787, + 0.175 + ], + "angle": 0, + "content": "prediction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 759-776. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.786, + 0.219 + ], + "angle": 0, + "content": "47. Mascaró, E.V., Ahn, H., Lee, D.: Intention-conditioned long-term human egocentric action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6048-6057 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.787, + 0.247 + ], + "angle": 0, + "content": "48. Mendonca, R., Bahl, S., Pathak, D.: Structured world models from human videos. arXiv preprint arXiv:2308.10901 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.249, + 0.787, + 0.289 + ], + "angle": 0, + "content": "49. Miech, A., Laptev, I., Sivic, J., Wang, H., Torresani, L., Tran, D.: Leveraging the present to anticipate the future in videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.291, + 0.787, + 0.346 + ], + "angle": 0, + "content": "50. Mohamed, A., Qian, K., Elhoseiny, M., Claudel, C.: Social-stgcnn: A social spatiotemporal graph convolutional neural network for human trajectory prediction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14424-14432 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.786, + 0.388 + ], + "angle": 0, + "content": "51. Myers, A., Teo, C.L., Fermüller, C., Aloimonos, Y.: Affordance detection of tool parts from geometric features. In: 2015 IEEE International Conference on Robotics and Automation (ICRA). pp. 1374-1381. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.786, + 0.431 + ], + "angle": 0, + "content": "52. Nagarajan, T., Feichtenhofer, C., Grauman, K.: Grounded human-object interaction hotspots from video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8688-8697 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.787, + 0.472 + ], + "angle": 0, + "content": "53. Nawhal, M., Jyothi, A.A., Mori, G.: Rethinking learning approaches for long-term action anticipation. In: European Conference on Computer Vision. pp. 558-576. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.515 + ], + "angle": 0, + "content": "54. Parsaeifard, B., Saadatnejad, S., Liu, Y., Mordan, T., Alahi, A.: Learning decoupled representations for human pose forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2294-2303 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.544 + ], + "angle": 0, + "content": "55. Peters, R.J., Iyer, A., Itti, L., Koch, C.: Components of bottom-up gaze allocation in natural images. Vision research 45(18), 2397-2416 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.787, + 0.585 + ], + "angle": 0, + "content": "56. Plizzari, C., Goletto, G., Furnari, A., Bansal, S., Ragusa, F., Farinella, G.M., Damen, D., Tommasi, T.: An outlook into the future of egocentric vision. arXiv preprint arXiv:2308.07123 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.614 + ], + "angle": 0, + "content": "57. Qi, H., Wang, X., Pathak, D., Ma, Y., Malik, J.: Learning long-term visual dynamics with region proposal interaction networks. arXiv preprint arXiv:2008.02265 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.787, + 0.657 + ], + "angle": 0, + "content": "58. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3636-3645 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.659, + 0.787, + 0.699 + ], + "angle": 0, + "content": "59. Rodin, I., Furnari, A., Mavroeidis, D., Farinella, G.M.: Predicting the future from first person (egocentric) vision: A survey. Computer Vision and Image Understanding 211, 103252 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.701, + 0.787, + 0.755 + ], + "angle": 0, + "content": "60. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "61. Shen, J., Dudley, J., Kristensson, P.O.: Fast and robust mid-air gesture typing for ar headsets using 3d trajectory decoding. IEEE Transactions on Visualization and Computer Graphics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "62. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "63. Song, D., Kyriazis, N., Oikonomidis, I., Papazov, C., Argyros, A., Burschka, D., Kragic, D.: Predicting human intention in visual observations of hand/object interactions. In: 2013 IEEE International Conference on Robotics and Automation. pp. 1608-1615. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "64. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "65. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "66. Swain, M.J., Ballard, D.H.: Color indexing. International journal of computer vision 7(1), 11-32 (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.342 + ], + "angle": 0, + "content": "67. Tu, Z., Huang, Z., Chen, Y., Kang, D., Bao, L., Yang, B., Yuan, J.: Consistent 3d hand reconstruction in video via self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "68. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.786, + 0.425 + ], + "angle": 0, + "content": "69. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "70. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "71. Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8121-8130 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.55 + ], + "angle": 0, + "content": "72. Yang, Y., Zhai, W., Luo, H., Cao, Y., Luo, J., Zha, Z.J.: Grounding 3d object affordance from 2d interactions in images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10905-10915 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.592 + ], + "angle": 0, + "content": "73. Yang, Y., Zhai, W., Luo, H., Cao, Y., Zha, Z.J.: Lemon: Learning 3d human-object interaction relation from 2d images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16284-16295 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.632 + ], + "angle": 0, + "content": "74. Ye, Y., Singh, M., Gupta, A., Tulsiani, S.: Compositional video prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10353-10362 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.632, + 0.787, + 0.674 + ], + "angle": 0, + "content": "75. Zhai, W., Cao, Y., Zhang, J., Xie, H., Tao, D., Zha, Z.J.: On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.787, + 0.703 + ], + "angle": 0, + "content": "76. Zhai, W., Luo, H., Zhang, J., Cao, Y., Tao, D.: One-shot object affordance detection in the wild. International Journal of Computer Vision 130(10), 2472-2500 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.703 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_origin.pdf b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..12cd0daf844e12556a3189bd7c403f39d27b4964 --- /dev/null +++ b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/b323c4bb-333e-48a5-9c13-ed254133260c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aab143277aa95c56274d57cde0c26fd6083acacd06063903790401f506db775c +size 2567000 diff --git a/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/full.md b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..399e3f279251abeb420abea4939faef8e4262310 --- /dev/null +++ b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/full.md @@ -0,0 +1,409 @@ +# Bidirectional Progressive Transformer for Interaction Intention Anticipation + +Zichen Zhang $^{1}$ , Hongchen Luo $^{1}$ , Wei Zhai $^{1,*}$ , Yang Cao $^{1,2}$ , and Yu Kang $^{1,2}$ + +1 University of Science and Technology of China + +$^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center {zhangzichen@mail., lhc12@mail., wzhai056@, kangduyu@, forrest@}ustc.edu.cn + +Abstract. Interaction intention anticipation aims to jointly predict future hand trajectories and interaction hotspots. Existing research often treated trajectory forecasting and interaction hotspots prediction as separate tasks or solely considered the impact of trajectories on interaction hotspots, which led to the accumulation of prediction errors over time. However, a deeper inherent connection exists between hand trajectories and interaction hotspots, which allows for continuous mutual correction between them. Building upon this relationship, a novel Bidirectional prOgressive Transformer (BOT), which introduces a Bidirectional Progressive mechanism into the anticipation of interaction intention is established. Initially, BOT maximizes the utilization of spatial information from the last observation frame through the Spatial-Temporal Reconstruction Module, mitigating conflicts arising from changes of view in first-person videos. Subsequently, based on two independent prediction branches, a Bidirectional Progressive Enhancement Module is introduced to mutually improve the prediction of hand trajectories and interaction hotspots over time to minimize error accumulation. Finally, acknowledging the intrinsic randomness in human natural behavior, we employ a Trajectory Stochastic Unit and a C-VAE to introduce appropriate uncertainty to trajectories and interaction hotspots, respectively. Our method achieves state-of-the-art results on three benchmark datasets Epic-Kitchens-100, EGO4D, and EGTEA Gaze+, demonstrating superior in complex scenarios. + +Keywords: Interaction Intention $\cdot$ Egocentric Videos $\cdot$ Transformer + +# 1 Introduction + +The pursuit of future prediction on egocentric videos is centered on anticipating intentions and behaviors observed from a first-person perspective [59]. Interaction intention can be reflected in the trajectory of hand motion and contact points on the next-active object [39]. This task is novel but holds significant potential value, such as AR [34, 61], robot skill acquisition [3, 9, 24, 44, 48], hand reconstruction [30, 67] and pose estimation [27, 38]. + +![](images/729d4c622d0dc0f1bf249262094c561098ee63d23f74e22b99e3041986ba1e7c.jpg) +Error Accumulation Over Time + +![](images/73c8ca3256aaec031421fe424af6593f591d00b95aece1223c4db66e030ec5c7.jpg) + +![](images/f2e97a34d637bcc7ca8e8e111115663f3a03eda4c86273ed331d0efeda79d2b2.jpg) +Fig. 1: Error accumulation when only considering the influence from hand trajectories to interaction hotspots. Red ones represent the ground truth, while Blue ones are predictions. + +Most existing relevant research treated hand trajectory forecasting and interaction hotspots prediction as independent tasks. Bao et al. [4] proposed to predict the 3D hand trajectory from a first-person view. However, They overlooked the guiding role of interactable objects on trajectories. The prediction process was incomplete, leading to significant errors at the end of trajectories. On the other hand, there was extensive research on interaction hotspots [18,32,41,42,51,52,63]. Nevertheless, these studies often focused on the characteristics of the objects themselves, resulting in lower integration with subjective human behaviors in complex scenes and less accurate predictions. Other studies [37,39] jointly predicted future hand trajectories and interaction hotspots, improving their integration. However, these studies only considered the unidirectional influence of hand trajectories on contact points. Consequently, errors accumulated continuously with the prediction time steps (Fig. 1). + +In fact, this one-way causation was insufficient, as it disregarded the spatial correlation between the two factors. Conversely, there exists a mutual relationship between the predictions of hand trajectories and interaction hotspots. On one hand, the derived directions of hand trajectories suggest the approximate areas where contact points occur. On the other hand, there is an inherent connection between the spatial distribution of contact points and hand trajectory patterns, as in Fig. 2, which implies that examining the distribution of contact points can provide insights into the development of hand trajectories. Leveraging this relationship, to mitigate the process of error accumulation, we establish a Bidirectional Progressive (Bi-Progressive) mechanism, allowing continuous mutual refinement between trajectories and interaction hotspots throughout the entire prediction period, improving the accuracy of joint predictions. + +In this paper, an innovative Bidirectional prOgressive Transformer (BOT) is proposed, introducing the Bi-Progressive mechanism mentioned earlier between two independent predicting branches. Firstly, the raw video input undergoes a Spatial-Temporal Reconstruction Module, which integrates the spatial information from the last observation frame and the temporal information among all observation frames. This process mitigates spatial information conflicts arising from changes in the field of view in first-person videos. Secondly, we establish separate dual branches for hand trajectory forecasting and interaction hotspots prediction, endowing them with the capability to independently perform indi + +![](images/deff622f82bdf84eaf74a92c45fed659133792a3d4ca0acd0f30840eb3a03024.jpg) +(a) Long-length trajectories + +![](images/7775fc263d940aa62090d77e1489f88bdb64940a5f53fe34c931a284d37f139f.jpg) + +![](images/904b37d7bce550a3152177872f041f9a750b1a700a85c0ae1c8b349713a4c5e0.jpg) + +![](images/1dad43f57b2dba72cd157ecaa28db7456c94982bfb3fd3a240bc7dce26aa1a63.jpg) + +![](images/510f94951bc8fb717922f0dafc397a3eab8e83bf2a3c966b991b1223adcfb7d7.jpg) +(b) Medium-length trajectories + +![](images/75a02884eefae83d9f647dc935b346d9e58419ac33e0a684c204c4c0b01e9d0d.jpg) + +![](images/d4757304e7c23018db96e89d3d46c7dbfc0a5d7cae50451f3e65fbbf1d8d8330.jpg) + +![](images/5f932b23785c83bbd6ab62b1bb0c2036f6aafad609740fdeeff8e9be084a4b20.jpg) + +![](images/a8cd833900d4d17f76445756c6e29d50e1e5885cad39fd30444d9b655b7f92d2.jpg) +(c) Short-length trajectories + +![](images/44e598bc4896d72cac03bad8b06a4feb60ce84130da800404c160e7cd0dbdbfc.jpg) +Fig. 2: Inherent connection between different hand trajectory categories and their corresponding contact points distribution. This is conducted on Epic-Kitchens-100 [14]. + +![](images/95ac88be95a6d3be921b581de02b3ca7a4b886741ddf8b8fa43338f4006f408c.jpg) + +![](images/9289fa77c6feacfb3779179cbb29ba6f5110b4730021aefb40da0096a775be4e.jpg) + +vidual prediction tasks, laying the foundation for introducing the Bi-Progressive mechanism. Subsequently, we introduce cross-attention blocks as the primary implementation of the Bi-Progressive mechanism. At each time step, the predictions of hand trajectories and interaction hotspots are alternately updated, addressing inherent errors in individual predictions. Finally, following the logic of natural human behaviors [50, 57, 74], we incorporate a Trajectory Stochastic Unit and a C-VAE for both hand trajectories and interaction hotspots, respectively, introducing appropriate uncertainties to the final prediction results. + +We evaluate our approach on Epic-Kitchens-100 [14], EGO4D [28] and EGTEA Gaze+ [36] datasets. BOT demonstrates notable advantages in predicting both hand trajectories and interaction hotspots, achieving superior performance in anticipating interaction intentions. In summary, our contributions are as follows: + +- A novel BOT is introduced to jointly predict future hand trajectories and interaction hotspots in egocentric videos, attaining the optimal predictive performance with more lenient input conditions. +- We introduce a novel Bidirectional Progressive mechanism that enhances the coherence between hand trajectories and interaction hotspots, thereby improving the completeness of the interaction intention. +- The experimental results demonstrate that our method achieves state-of-the-art results across three benchmark datasets Epic-Kitchens-100 [14], EGO4D [28], and EGTEA Gaze+ [36], demonstrating superior in complex scenarios. + +# 2 Related Work + +Action anticipation in egocentric videos. Anticipating actions in an egocentric vision is a challenging task that includes short-term and long-term action anticipation [56,59]. Short-term action anticipation aims to predict the next action in the near future [21-23,25,49], while long-term action anticipation focuses + +on the sequence of subsequent actions performed by the camera wearer [26, 40, 47, 53]. These endeavors contribute significantly to the comprehension of actions in first-person videos. However, concerning interaction intention, action anticipation still yields information of relatively high dimensionality. Anticipating a semantic label does not furnish direct guidance for the robot, as it lacks specific instructions on task execution. Comparatively, hand trajectories and interaction hotspots offer a more intuitive representation of interaction intention, constituting a lower-dimensional depiction. + +Human trajectory forecasting. The forecasting of human trajectories was extensively studied over the years [1,2,12,13,45,54]. However, most of them were based on third-person video datasets, falling short of capturing intentions related to human-object interactions. Since hands serve as the primary medium through which humans interact with the world, predicting hand trajectories in egocentric videos can aid embodied AI in comprehending human-object interaction intentions. The approaches in [4,37,39] addressed the issue of hand trajectory prediction in egocentric videos. However, they did not account for the influence of contact points with objects on trajectories. This limitation resulted in the continuous accumulation of prediction errors over time. + +Interaction hotspots prediction. Interaction hotspots prediction aims to locate where the hand-object interaction happens. Some research aimed to predict the next interactive object [5, 6, 15, 17, 20], going further, some work anticipated specific interaction regions on objects [42, 43, 52, 72, 73, 75, 76], yielding more precise information for interaction. Differing from image input, our task aims to generate more precise predictions regarding the regions of contact points by utilizing egocentric video inputs. The incorporation of first-person videos is intended to focus on precisely localizing interaction hotspots for distinct human motions. In this task, we leverage the predicted hand trajectories to improve the accuracy of interaction hotspots predictions. + +# 3 Method + +Problem Setup. Given an original input video $\mathcal{I} = \{I_1, I_2, \dots, I_N\}$ , where $I_{\mathrm{N}}$ is the last observation frame, the task is predicting future hand trajectories $\mathcal{H} = \{h_{\mathrm{N + 1}}, h_{\mathrm{N + 2}}, \dots, h_{\mathrm{N + T - 1}}\}$ in future keyframes $\{I_{\mathrm{N + 1}}, I_{\mathrm{N + 2}}, \dots, I_{\mathrm{N + T - 1}}\}$ and the contact point $\mathcal{O}_{\mathrm{C}}$ in the contact frame $I_{N + T}$ . We subsequently convert the contact point $\mathcal{O}_{\mathrm{C}}$ into a heatmap by centering a Gaussian distribution, generating interaction hotspots $\mathcal{O}_{\mathrm{I}}$ . To ensure the continuity of the trajectory as well as the consistency of the background, both $\mathcal{H}$ and $\mathcal{O}_{\mathrm{I}}$ are projected onto the last observation frame, which enhances the overall coherence of the predictions [39]. + +# 3.1 Network + +Overview. As shown in Fig. 3, given the original input video $\mathcal{I}$ , we aim to predict the future hand trajectory $\mathcal{H}$ and interaction hotspots $\mathcal{O}_{\mathrm{I}}$ . Initially, to alleviate information conflicts arising from changes in the field of view, the + +![](images/1d22189a85db342bb468c6b4afa5b756fc13ca2d042ab9455079a017bb1c51ab.jpg) +Fig. 3: Overview of the Bidirectional Progressive Transformer. It follows a dual-branch structure with a Bi-Progressive Enhancement Module between them to anticipate future hand trajectories and interaction hotspots. + +original input video undergoes a Spatial-Temporal Reconstruction Module that integrates spatial information from the last observation frame and temporal information from the entire video, producing spatial features for each future timestep. Subsequently, these features are separated, with the last one serving as the contact feature while the preceding ones are for trajectory features. These features are then fed into the interaction hotspots and trajectory prediction branches, respectively. The former employs an overlap patch embedding layer to enhance attention at the center of the field of view, while the latter introduces temporal order to trajectory features through Cross-Transformers [11]. Simultaneously, we employ a Bi-Progressive Enhancement Module consisting of cross-attention blocks between dual branches, introducing a Bi-Progressive mechanism to the joint prediction. This alternately and iteratively corrects the features of trajectories and interaction hotspots in chronological order, thereby mitigating the accumulation of prediction errors over time. Finally, considering the inherent uncertainty in natural human behaviors [50, 57, 74], we introduce a Stochastic Unit for trajectory forecasting while employing a C-VAE [62] for interaction hotspots, adding appropriate uncertainty during the completion of position decoding to achieve comprehensive predictions. + +Spatial-Temporal Reconstruction. When dealing with egocentric videos, both 3D convolutions [19] and self-attention mechanisms [68] struggle to address spatial information conflicts arising from changes in views. Moreover, predictions are projected onto the last observation frame. Taking these factors into account, we establish the Spatial-Temporal Reconstruction Module. On one hand, we employ optical flow containing more comprehensive motion information as the temporal input, on the other hand, we isolate the last observation frame as the spatial input. Subsequently, the temporal input is processed by a pre-trained Temporal Segment Network [22,69], extracting low-dimensional temporal features $\mathbf{X}$ , and the spatial input is passed through a Segformer-b2 [70], extracting spatial features $\mathbf{S}$ from the last observation frame. $\mathbf{X}$ and $\mathbf{S}$ are then fed into a + +Cross-Transformer [11]. The $i$ -th block is expressed as: + +$$ +\mathbf {s} _ {0} = \mathbf {s} _ {\text {p a t c h}} + \mathbf {s} _ {\text {p o s}}, \quad \mathbf {x} = \mathbf {x} + \mathbf {x} _ {\text {p o s}}, \tag {1} +$$ + +$$ +\mathbf {y} _ {i} = \mathbf {s} _ {i - 1} + \mathbf {M C A} (\mathbf {L N} (\mathbf {s} _ {i - 1}), \mathbf {L N} (\mathbf {x})), \tag {2} +$$ + +$$ +\mathbf {s} _ {i} = \mathbf {y} _ {i} + \mathbf {F F N} (\mathbf {L N} (\mathbf {y} _ {i})). \tag {3} +$$ + +Where $\mathbf{s}_0$ and $\mathbf{x}$ represent the original spatial and temporal inputs, respectively. $\mathbf{s}_{patch}$ is encoded by overlap patch embedding layers from spatial features $\mathbf{S}$ , enhancing the network's focus on the center of the field of view. $\mathbf{x}$ is encoded by positional embedding layers from temporal features $\mathbf{X}$ , improving the overall sequentiality. Furthermore, FFN refers to the Feed Forward Layer, comprising a two-layer MLP with an expanding ratio $r$ , and Layer normalization (LN) is applied before each block. MCA stands for Multi-head Cross-Attention, where query, key, and value are processed through distinct linear layers: MCA(X,Y) = Attention(W^QX, W^KY, W^VY). The output of this module corresponds to the spatial features for each prediction frame. + +Interaction Hotspots Prediction Branch. Due to the spatial distribution of contact points closer to the central field of view, the contact feature $\mathbf{F}_{\mathrm{C}}$ undergoes processing through an overlap patch embedding layer: + +$$ +\mathbf {F} _ {\mathrm {C}} ^ {\mathrm {o}} = \text {o v e r l a p} (\mathbf {F} _ {\mathrm {C}}). \tag {4} +$$ + +Where $\mathbf{F}_C^0$ represents the processed contact feature. + +Hand Trajectory Prediction Branch. We employ original trajectory features as primary features and utilize features from the previously predicted frame as conditional features, subsequently merging them through the cross-attention block $f$ , ensuring the coherence of the trajectory over time. Assuming the trajectory features are $\{\mathbf{F}_{\mathrm{N} + 1},\mathbf{F}_{\mathrm{N} + 2},\dots ,\mathbf{F}_{\mathrm{N} + \mathrm{T} - 1}\}$ and the feature map of the last observation frame is $\mathbf{F}_{\mathrm{N}}$ , they are updated in chronological order: + +$$ +\mathbf {F} _ {t} ^ {\mathrm {o}} = f \left(\mathbf {F} _ {\mathrm {t}}, \mathbf {F} _ {\mathrm {t} - 1}\right), \quad t \in [ \mathrm {N} + 1, \mathrm {N} + \mathrm {T} - 1 ]. \tag {5} +$$ + +Where $\mathbf{F}_t^{\mathrm{o}}$ stands for the updated trajectory feature, $t$ is the prediction time step. + +H-O Bi-Progressive Enhancement. Expanding on the groundwork laid by two distinct prediction branches, and to mitigate the accumulation of errors as predictions unfold over time, we introduce a Bi-Progressive Enhancement Module (Fig. 4), which alternately corrects the predictions of trajectories $\mathcal{H}$ and interaction hotspots $\mathcal{C}_{\mathrm{I}}$ at each time step $t$ . Specifically, owing to observation frames and the continuity of hand trajectories, the trajectory feature $\mathbf{F}_{\mathrm{N} + 1}^{\mathrm{o}}$ at the initial predicted time step $\mathrm{N} + 1$ demonstrates a high level of accuracy. Exploiting this, we enhance the precision of the contact feature $\mathbf{F}_{\mathrm{C}}^{\mathrm{o}}$ through the cross-attention block $f$ , thereby minimizing the error with the ground truth. Moving to the subsequent time step $\mathrm{N} + 2$ , we employ the first refined contact feature $\mathbf{F}_{\mathrm{C}}^{\mathrm{r}_1}$ from the previous time step $\mathrm{N} + 1$ as a condition to correct the trajectory feature $\mathbf{F}_{\mathrm{N} + 2}^{\mathrm{o}}$ . Following this, the corrected trajectory feature is utilized to further refine the contact feature. This iterative correction process persists until all trajectory + +features have undergone a complete round of updates. It can be represented as a recursive procedure: + +$$ +\mathbf {F} _ {\mathrm {N} + 1} ^ {\mathrm {r}} = \mathbf {F} _ {\mathrm {N} + 1} ^ {\mathrm {o}}, \tag {6} +$$ + +$$ +\mathbf {F} _ {\mathrm {C}} ^ {\mathrm {r} _ {1}} = f \left(\mathbf {F} _ {\mathrm {C}} ^ {\mathrm {o}}, \mathbf {F} _ {\mathrm {N} + 1} ^ {\mathrm {r}}\right), \quad \mathbf {F} _ {\mathrm {N} + 2} ^ {\mathrm {r}} = f \left(\mathbf {F} _ {\mathrm {N} + 2} ^ {\mathrm {o}}, \mathbf {F} _ {\mathrm {C}} ^ {\mathrm {r} _ {1}}\right), \tag {7} +$$ + +$$ +\cdot \cdot \cdot +$$ + +$$ +\mathbf {F} _ {\mathrm {C}} ^ {\mathrm {r} _ {3}} = f \left(\mathbf {F} _ {\mathrm {C}} ^ {\mathrm {r} _ {2}}, \mathbf {F} _ {\mathrm {N} + 3} ^ {\mathrm {r}}\right), \quad \mathbf {F} _ {\mathrm {N} + 4} ^ {\mathrm {r}} = f \left(\mathbf {F} _ {\mathrm {N} + 4} ^ {\mathrm {o}}, \mathbf {F} _ {\mathrm {C}} ^ {\mathrm {r} _ {3}}\right). \tag {8} +$$ + +Where $\mathbf{F}_t^{\mathrm{r}}$ represents the refined trajectory feature map at time step $t\in [\mathrm{N} + 1,\mathrm{N} + \mathrm{T} - 1]$ , $\mathbf{F}_{\mathrm{C}}^{\mathrm{r},i}$ denotes the contact feature after the $i$ -th refinement. Via this module, we attain refined features for both trajectories and interaction hotspots, significantly reducing error accumulation in predictions. + +Interaction Hotspots Decoder. Given the presence of inherent uncertainty associated with interaction hotspots [39, 41, 42], we employ a C-VAE [62] as the decoder. Due to the excessive randomness associated with directly decoding the coordinates of contact points, we instruct the C-VAE to output a feature map and select the location of the maximum value from this map as the output, thereby confining the uncertainty within a reasonable range. The C-VAE follows an encoder-decoder structure. For the encoder block, we concatenate the input $\varGamma$ with the condition $\mathbf{C}$ through an encoding function $\mathbf{P}_{\mathrm{e}}$ , obtaining features in latent space $\mathbf{Z}$ which parameterized by mean $\mu$ and co-variance $\sigma$ . In this context, the input $\varGamma$ is the ground truth heatmap, while the condition $\mathbf{C}$ is the refined contact feature $\mathbf{F}_{\mathrm{C}}^{\mathrm{r}_3}$ . In the decoder + +block, the first step involves sampling $\mathbf{Z}$ from the latent space and concatenating it with condition $\mathbf{C}$ . Following that, the output heatmap $\hat{\Gamma}$ is reconstructed by a decoding function $\mathbf{P}_d$ . In summary, the entire process can be formulated as follows: + +$$ +\mu , \sigma = \mathbf {P} _ {\mathrm {e}} (\Gamma , \mathrm {C}), \quad \hat {\Gamma} = \mathbf {P} _ {\mathrm {d}} (\mathrm {Z}, \mathrm {C}), \quad \mathrm {Z} \sim \mathcal {N} (\mu , \sigma). \tag {9} +$$ + +Where $\mathcal{N}$ represents a normal distribution, $\hat{\Gamma}$ is the reconstructed heatmap. Both encoding and decoding blocks consist of two MLP layers. We introduce KL-Divergence and the reconstruction error as the overall loss: + +$$ +\mathcal {L} _ {\text {r e c o n}} (\Gamma , \hat {\Gamma}) = (\Gamma - \hat {\Gamma}) ^ {2}, \tag {10} +$$ + +$$ +\mathcal {L} _ {k l} (\mu , \sigma) = - \mathbf {K L} \left[ \mathcal {N} (\mu , \sigma) ^ {2} | | \mathcal {N} (0, 1) \right], \tag {11} +$$ + +$$ +\mathcal {L} _ {o} = \mathcal {L} _ {\text {r e c o n}} + \lambda \mathcal {L} _ {k l}. \tag {12} +$$ + +In this paper, we set $\lambda = 0.01$ . + +![](images/0c48a0ee08548f20e60eeb879a31deebec6c2a74dcbd5c54e4734051926f5c09.jpg) +Fig. 4: Bi-Progressive Enhancement Module. The output feature map is enclosed in purple boxes. + +$h_t$ hand position $\mathcal{O}_{\mathrm{C}}$ contact point $\mathcal{R}_{21}, \mathcal{R}_{22}$ elliptical region $\mathcal{R}_2$ uncertain region $\theta$ ellipse tilt angle + +![](images/b824b0898f3f03875f2d568b7d2c2bd9ef704fca6b36a17b95f7ef5500a8f536.jpg) +Fig. 5: Uncertain regions of the hand trajectory. It is derived through the intersection of two ellipses, corresponding to the Eq. 14-19. (i.e., the purple region). + +Hand Trajectory Decoder. Similar to interaction hotspots, hand trajectories also exhibit uncertainty [29, 33, 46]. Given that trajectory feature maps with distinct distribution characteristics have been obtained through the Bi-Progressive Enhancement Module, in conjunction with sampling efficiency, we don't utilize C-VAE, rather, we introduce a Trajectory Stochastic Unit to incorporate uncertainty into trajectory forecasting by calculating uncertain regions for each prediction time step instead. Since optimizing directly using point coordinates for loss calculation can pose challenges, we initially introduce a Gaussian distribution for each ground truth hand position to generate the ground truth heatmap. Subsequently, we calculate the loss between the corrected feature map and the generated ground truth heatmap, thereby enabling the network to learn the distribution of potential hand positions. The associated loss function $\mathcal{L}_h$ is based on MSE Loss: + +$$ +\mathcal {L} _ {h} = \frac {1}{T - 1} \sum_ {t = N + 1} ^ {N + T - 1} \left(\mathbf {F} _ {t} ^ {\mathrm {r}} - \mathbf {G} _ {t}\right) ^ {2}. \tag {13} +$$ + +Where $\mathbf{G}_t$ represents the ground truth heatmap at time step $t$ . We designate the coordinates of the maximum value within each feature map as the predicted trajectory points, which are then fed into the stochastic unit to generate uncertain regions. The size of the region is determined by the time step as well as the length of trajectory segments in preceding and subsequent stages. The shape is constrained by the original image dimensions and the orientation of the trajectory segments. Formally, the uncertain region (Fig. 5) can be shown as: + +$$ +\mathcal {R} _ {t} \sim (t, h _ {t - 1}, h _ {t}, h _ {t + 1}, \mathrm {W}, \mathrm {L}). \tag {14} +$$ + +Where $\mathcal{R}_t$ is the uncertain region in future time step $t$ , W and L are the width and length of the input video, respectively. $\mathcal{R}_t$ satisfies the following constraints: + +$$ +\mathcal {R} _ {t} = \mathcal {R} _ {t 1} \cap \mathcal {R} _ {t 2}. \tag {15} +$$ + +Where $\mathcal{R}_{t1}$ and $\mathcal{R}_{t2}$ are elliptical regions constrained by $h_{t-1}$ and $h_{t+1}$ , which can be represented as: + +$$ +\frac {\left(\mathbf {x} _ {t} \cos \theta_ {t \omega} - \mathbf {y} _ {t} \sin \theta_ {t \omega}\right) ^ {2}}{\mathbf {a} _ {t \omega} ^ {2}} + \frac {\left(\mathbf {x} _ {t} \sin \theta_ {t \omega} + \mathbf {y} _ {t} \cos \theta_ {t \omega}\right) ^ {2}}{\mathbf {b} _ {t \omega} ^ {2}} \leq 1, \tag {16} +$$ + +$$ +\mathbf {a} _ {t \omega} = \alpha \beta^ {t - 1} \mathbf {d} ^ {\frac {1}{2}} \left(h _ {t + \omega - 1}, h _ {t + \omega - 2}\right), \quad \mathbf {b} _ {t \omega} = \sqrt {\frac {\mathrm {L}}{\mathrm {W}}} \mathbf {a} _ {t \omega}, \tag {17} +$$ + +$$ +\theta_ {t \omega} = \tan^ {- 1} (\mathbf {k} \left(h _ {t + \omega - 1}, h _ {t + \omega - 2}\right)), \tag {18} +$$ + +$$ +\omega \in [ 1, 2 ], t \in [ 1, 2, 3, 4 ], \mathcal {O} _ {\mathrm {C}} = h _ {5}. \tag {19} +$$ + +Specifically, $\alpha$ denotes the overall scale of $\mathcal{R}_t$ , $\beta$ signifies the temporal growth rate, $\theta$ stands for the ellipse tilt angle, $(\mathbf{x}_t,\mathbf{y}_t)$ represent the coordinates of $h_t$ . Additionally, $\mathbf{k}(\mathrm{X},\mathrm{Y})$ and $\mathbf{d}(\mathrm{X},\mathrm{Y})$ represent the slope and Euclidean distance between points $\mathrm{X}$ and $\mathrm{Y}$ , respectively. In BOT, we set $\alpha = 1.5$ and $\beta = 2$ . + +# 3.2 Training and Inference + +Training. During training, we freeze parameters in TSN [69]. Training loss combines both trajectory loss and interaction hotspots loss: + +$$ +\mathcal {L} = \mathcal {L} _ {h} + \varphi \mathcal {L} _ {o}. \tag {20} +$$ + +In this paper, we set $\varphi = 1$ . + +Inference. We use a Trajectory Stochastic Unit and a C-VAE to introduce uncertainty during inference. Following previous work [39, 46], we report the minimum among 20 samples for trajectory evaluation. + +# 4 Experiments + +# 4.1 Datasets + +We conducted experiments on three first-person video datasets Epic-Kitchen s-100 [14], EGO4D [28], and EGTEA Gaze+ [36]. Epic-Kitchens-100 [14] and EGTEA Gaze+ [36] predominantly include kitchen scenes, while EGO4D [28] encompasses a broader range of activities such as handicrafts, painting, and outdoor tasks. For each dataset, the observation time is set to 2 seconds. We configure the prediction time horizon T to 5, corresponding to 4 trajectory points and one contact point for each hand. Specifically, we predict 1-second trajectories on Epic-Kitchens-100 [14] and EGO4D [28], while on EGTEA Gaze+ [36], 0.5-second trajectories are forecast. Additional dataset information is available in the supplementary materials. + +# 4.2 Implementation Details + +In EGO4D [28], We use the method [39] to automatically generate hand trajectories, while contact points are manually annotated based on the last observation frame and the contact frame. We employ GMFLOW [71] to generate optical flow from EGO4D [28] and EGTEA Gaze+ [36]. For each dataset, we resize the original videos to $456 \times 256$ size, extracting both trajectory and contact features of size $114 \times 64$ . In the Spatial-Temporal Reconstruction Module, we set the dimension $D$ in Cross-Transformer to 1024, with depth $= 6$ and attention head number $= 8$ . For the trajectory forecasting branch and H-O Bi-Progressive Enhancement Module, we configure the dimension $D$ to 256, with depth $= 2$ and head number $= 8$ . Additionally, for the introduced C-VAE, we set the dimension of both input and condition to $114 \times 64$ , while the hidden and latent sizes are set to 1024 and 2048, respectively. Our method is implemented by PyTorch and trained with the AdamW optimizer. We train the model for 100 epochs on 4 NVIDIA 3090 GPUs with an initial learning rate of $3e - 5$ and a batch size of 40. + +For comparison, We first select methods for individual tasks including Kalman Filter (KF) [8], Seq2Seq [65], Hotspots [52] and USST [4]. Additionally, the methods for image/video understanding are chosen as the compared methods, which lack a joint mechanism. The former includes VIT [16], HRNet [64], UNet [60], Segformer [70], while the latter contains I3D [10], Slowfast [19], Timesformer [7], Uniformer-V2 [35], StillFast [58]. Furthermore, we select methods with high task relevance and joint design, including FHOI [37] and OCT [39]. We conduct experiments using these approaches on Epic-kithens-100 [14], EGO4D [28], and EGTEA Gaze+ [36] datasets. + +# 4.3 Evaluation Metrics + +Hand Trajectory Evaluation. The predicted hand trajectories are referenced to the last observation frames. ADE [39] and FDE [39] are chosen as hand trajectory evaluation metrics, which respectively represent the overall accuracy and endpoint accuracy of trajectories. Detailed descriptions are available in the supplementary materials. + +Interaction Hotspots Evaluation. For all datasets, we downsample each predicted heatmap to the size of $32 \times 32$ . We choose SIM [66], AUC-J [31], and NSS [55] as interaction hotspots evaluation metrics. For a detailed explanation of the metrics, please refer to the supplementary materials. + +# 4.4 Quantitative and Qualitative Comparisons + +As shown in Tab. 1, our method achieves state-of-the-art performance in both trajectory prediction and interaction hotspots anticipation, with comparable leading margins across all three datasets. Taking the EGO4D dataset [28] with the richest variety of scenes as an example. In trajectory prediction, to ensure fair comparison, we additionally present results for OCT and BOT with only one sampling iteration. Under this single-sampling condition, BOT outperforms + +Table 1: Results of different models. Best results are depicted using red (sample 20 times) and blue (sample once), respectively. unilateral, non-joint and joint denote prediction approaches for individual tasks, for two tasks without a collaborative design, and with a joint module, respectively. “-” denotes the vacancy of the value. “*” represents the results obtained from 20 random samplings. ↑ and ↓ indicate a better performance when the value of the metric increases and decreases, respectively. + +
MethodEK100 [14]EGO4D [28]EGTEA [36]
ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑
unilateralCenter0.240.240.090.610.330.260.250.070.570.150.320.320.090.630.27
KF [8]0.330.32--------0.490.48---
Seq2Seq [65]0.180.14--------0.180.14---
Hotspots [52]--0.150.660.53-------0.150.710.69
USST [4]0.170.18---0.180.18---0.180.18---
non-jointI3D [10]0.220.230.130.560.460.220.220.120.570.420.230.220.130.600.49
SLOWFAST [19]0.220.240.140.620.470.230.220.130.580.440.210.230.140.580.46
Timesformer [7]0.230.230.140.600.420.240.240.120.590.400.210.220.140.580.45
Uniformer-V2 [35]0.200.200.130.590.430.210.220.120.620.400.200.210.130.580.46
StillFast [58]0.180.190.160.620.610.190.190.150.650.570.180.190.160.600.55
UNet [60]0.190.180.150.620.570.180.180.140.590.550.180.180.140.600.60
HRNet [64]0.180.190.160.620.630.190.200.140.620.580.180.180.160.630.65
VIT [16]0.180.190.140.650.630.210.210.130.590.450.190.190.150.630.61
CrossVIT [11]0.190.200.140.620.590.190.200.130.600.550.200.210.130.590.49
Segformer-b2 [70]0.170.170.160.640.660.160.160.150.640.620.170.170.160.630.65
jointFHOI [37]0.360.350.100.550.39-----0.340.340.120.630.42
OCT [39]0.150.17---0.150.16---0.160.16---
OCT* [39]0.120.110.190.720.720.120.120.220.730.880.140.140.230.751.01
Ours0.120.14---0.110.13---0.120.14---
Ours*0.100.080.340.812.070.090.080.320.771.700.090.080.320.802.03
+ +Segformer-b2, the best method in image/video understanding, with respective leads of $31.3\%$ and $18.8\%$ in ADE and FDE. Compared to OCT, BOT maintains advantages of $26.7\%$ and $18.8\%$ in ADE and FDE, respectively. These advantages persist even with 20 samplings, while the lead in FDE expands to $33.3\%$ , underscoring the rationale behind BOT's stochastic sampling mechanism. Regarding interaction hotspots, we achieve the best results across all metrics. Specifically, BOT surpasses the second-best approach (OCT) by $45.5\%$ in SIM, $5.5\%$ in AUC-J, and $93.2\%$ in NSS, respectively. In summary, our method leverages the intrinsic correlations between hand trajectories and interaction hotspots more effectively, enhancing overall prediction accuracy of interaction intention, and exhibiting robustness across the three benchmark datasets. + +To further explore the subjective outcomes of models, we report the visualization results of three datasets, as shown in Fig. 6. Compared to other methods, our model provides more accurate prediction results in both hand trajectories and interaction hotspots. When there is significant movement of hands (column 1,3,8), the BOT's predictions exhibit the highest accuracy. It indicates that our Spatial-Temporal Reconstruction Module can efficiently utilize the human motion information to exploit the hand interaction intention and thus accurately predict the hand trajectory. In cases where there are no identifiable interactive objects in the field of view (column 6), our method anticipates more accurate interaction intentions. It illustrates that our Bi-Progressive mechanism can eliminate the uncertainty in the interaction process by mining the intrinsic connection between + +![](images/579663f562ff6b4dae9a41557ab1dce9686104d2f72a1ad76cc94b526d628c84.jpg) +Fig. 6: Visualization of future hand trajectories and interaction hotspots on EK100, EGO4D and EGETA. We show the visualization results of our model, the joint prediction model (OCT) [39], the segmentation model (Segformer-b2) [70], and the video understanding model (Uniformer-V2) [35]. The green represents the trajectory of the left hand, while the red denotes the trajectory of the right hand. + +the hand trajectory and the contact points, which results in a more accurate inference in the interaction region and the hand trajectory. Moreover, our predictions closely align with the actual scenario when both hands simultaneously exhibit trajectories and interaction hotspots (columns 4,5), which suggests the strategy of differentiating between left and right hands during the prediction of interaction intentions, thereby mitigating cross-hand interference. More visualization results are available in the supplementary materials. + +# 4.5 Ablation Study + +Different joint approaches. To explore the impact of the Bi-Progressive mechanism on model performance, we experiment with five different joint configurations: Individual, Hand, Object, Bidirectional and Bi-Progressive. The results are shown in Tab. 2 and Fig. 7. It suggests that the network is challenging to accurately predict trajectories and interaction hotspots during individual predictions, and the separate incorporation of hands and objects can effectively improve the prediction accuracy of interaction hotspots. With the combination of hand and object information, the prediction performance of the trajectory and interaction region is greatly improved. However, there is still a gap compared to the Bi-Progressive mechanism. Further, Fig. 8 shows that the introduction of the progressive mechanism plays a pivotal role in mitigating the accumulation of prediction errors over time. + +Uncertainty in anticipation. Uncertainty is a crucial consideration in interaction intention anticipation. Following the logic of natural human behavior [50,57,74], both hand trajectories [29,33,46] and interaction hotspots [39,41,42] predictions inherently possess a degree of uncertainty. To investigate the impact + +Table 2: Results of different joint forecasting methods. Individual refers to separately predict two subtasks. Hand indicates that the hand trajectory is the condition of the interaction hotspots. Object stands for correcting the trajectory using predicted contact points. Bidirectional combines the methods Hand and Object. + +
MethodTrajectoryInteraction Hotspots
ADE ↓ FDE ↓SIM ↑ AUC-J ↑NSS ↑
Individual0.140.140.180.680.86
Hand0.130.120.250.721.19
Object0.120.120.230.701.04
Bidirectional0.100.090.280.731.48
Bi-Progressive0.090.080.320.771.70
+ +![](images/1016746c7d59d5e7ce2f4e81cedb6def15212392e41bb8c8ed049b4ff5f675d3.jpg) +Fig. 7: Visualization results of different joint approaches. Based on the order from left to right and top to bottom, the corresponding methods are GT, Bi-Progressive, Bidirectional, Hand, Object, Individual. + +![](images/682996daf8b4fa713de5d94d31ad6e7eceaa69b6bebd6287d29bb334f1713a76.jpg) +Fig. 8: The role of Progressive Mechanism. The red refers to the ground truth, while the blue represents the predictions. From top to bottom, Bi-Pro stands for Bi-Progressive while Bid refers to bidirectional. + +of uncertain region scale and the magnitude of uncertainty variations in hand trajectories, we modify $\alpha$ and $\beta$ in Eq. 17, with the resultant outcomes depicted in Fig. 9. $\alpha$ denotes the overall scale of the uncertain region, when $\alpha$ equals 0, the region's size is zero, and the network output yields a unique result. As $\alpha$ increases, the uncertainty of trajectories grows, leading to improved accuracy due to random samplings. However, given the limitations in sampling times, as $\alpha$ continues to increase, the accuracy naturally decreases. $\beta$ determines the magnitude of trajectory uncertainty increasing from front to back, when $\beta$ equals 1, each trajectory point has the same level of uncertainty. Considering the escalating complexity and uncertainty in trajectory prediction over time, an elevation in $\beta$ initially corresponds to augmented accuracy, followed by a subsequent decline. This gradual increase in uncertainty also renders the FDE more sensitive to $\alpha$ and $\beta$ , as shown in Fig. 10a-10c. Regarding interaction hotspots, BOT exhibits relatively small uncertainty (Fig. 10d), as for a specific intention, the interaction area becomes increasingly influenced by object constraints. This is achieved by tightly controlling the uncertainty within a reasonably small range through the process of enabling C-VAE to generate feature maps and subsequently select + +![](images/9946323f2de0e43254fdf43aea0faba3f0fa204c520809e8b7a6145446835db4.jpg) +Fig. 9: Uncertainty in hand trajectory forecasting. As $\alpha$ and $\beta$ increase, they denote a heightened degree of overall trajectory uncertainty. + +![](images/760a15b4c9d602b1549d2a669ce781c93211af3d6ea2a380cc0722dacc529d3d.jpg) + +![](images/9706ceef5dfae98bd7e824999cb90f44c6997cd5ba8063a80c858205de88cd42.jpg) + +![](images/c7ce3df40f11f04fbece25e9d1f9b701778d47733f7e6ae3cf407cc3ad53a556.jpg) + +![](images/95fa90005cc49811a5cf2b25a6fb9c77aaa6e1aaa6bf2a5d19bc9e0772f5ede3.jpg) + +![](images/d3c792f8019df11cf04d6e88d920257bea4cca5f8a3ff7053b2d5c61f7270f19.jpg) +(a) Long-length trajectories +(c) Short-length trajectories +Fig. 10: Uncertainty in anticipation. (a), (b) and (c) respectively depict long-length trajectories, medium-length trajectories, and short-length trajectories, while (d) illustrates the uncertainty of interaction hotspots. + +![](images/0cfd00471519f5e232259737c6a0d580a05742e2aeca17f994ca306669647783.jpg) + +![](images/c9a6ec9d8581cab8aac50879e50fcf55f95a0711048cc5419fb82077b82e2974.jpg) + +![](images/746bc5e6a08a005ffa5b9757a001cbe707d5f13573df33442dd5c655a059fe5e.jpg) + +![](images/a9258eca08991ea1b76caf076842b07def6f5b0da924eaf3859f0d06a104f293.jpg) + +![](images/a0f48ce9da74069ef86b2b4958fb6ef2119d64d984833938fb29d08a44654c81.jpg) + +![](images/1a2e7ba433910f827577e8f1c603fc9ee02180df9b39a6583be09a7034394358.jpg) +(b) Medium-length trajectories +(d) interaction hotspots + +![](images/fbfe206760380f513fb903ce4148aeb988429e5e1d380b89bf3dcee908babc4d.jpg) + +![](images/2e82da4de91af397b39ffff4f8c750e05a0d85ed2af730938172ef5d9ed0b8e5.jpg) + +points, thereby ensuring the network's high confidence in predicting interaction regions. + +# 5 Discussion + +Conclusion. We plan to anticipate interaction intentions, encompassing the joint prediction of future hand trajectories and interaction hotspots with objects. We propose a novel BOT framework to establish an inherent connection between hand trajectories and contact points by introducing a Bi-Progressive mechanism, which eliminates the effect of error accumulation and achieves more accurate intention anticipation. Experimental results on three challenging benchmarks achieve state-of-the-art performance, thereby proving the superiority of our method in intention anticipation. + +Future Work. BOT currently predicts hand trajectories and interaction hotspots on 2D images. However, human-object interactions occur in 3D space. To enable the model to be applied to real-world scenarios, we will consider the geometric structure and spatial relationships between objects in the scene in the future, providing the model with the ability to anticipate interaction intentions for the real world. + +# Acknowledgements + +This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62306295 and 62033012. + +# References + +1. Adeli, V., Adeli, E., Reid, I., Niebles, J.C., Rezatofighi, H.: Socially and contextually aware human motion and pose forecasting. IEEE Robotics and Automation Letters 5(4), 6033-6040 (2020) +2. Adeli, V., Ehsanpour, M., Reid, I., Niebles, J.C., Savarese, S., Adeli, E., Rezatofighi, H.: Tripod: Human trajectory and pose dynamics forecasting in the wild. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13390-13400 (2021) +3. Bahl, S., Gupta, A., Pathak, D.: Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450 (2022) +4. Bao, W., Chen, L., Zeng, L., Li, Z., Xu, Y., Yuan, J., Kong, Y.: Uncertainty-aware state space transformer for egocentric 3d hand trajectory forecasting. arXiv preprint arXiv:2307.08243 (2023) +5. Bertasius, G., Park, H.S., Yu, S.X., Shi, J.: First person action-object detection with egonet. arXiv preprint arXiv:1603.04908 (2016) +6. Bertasius, G., Soo Park, H., Yu, S.X., Shi, J.: Unsupervised learning of important objects from first-person videos. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1956-1964 (2017) +7. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021) +8. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE international conference on image processing (ICIP). pp. 3464-3468. IEEE (2016) +9. Bharadhwaj, H., Gupta, A., Tulsiani, S., Kumar, V.: Zero-shot robot manipulation from passive human videos. arXiv preprint arXiv:2302.02011 (2023) +0. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +1. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021) +2. Chen, G., Chen, Z., Fan, S., Zhang, K.: Unsupervised sampling promoting for stochastic human trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17874-17884 (2023) +3. Choi, C., Dariush, B.: Looking to relations for future trajectory forecast. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019) +4. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Scaling egocentric vision: The epic-kitchens dataset. In: European Conference on Computer Vision (ECCV) (2018) + +15. Dessalene, E., Devaraj, C., Maynard, M., Fermuller, C., Aloimonos, Y.: Forecasting action through contact representations from first person video. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) +16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) +17. Fan, C., Lee, J., Ryoo, M.S.: Forecasting hands and objects in future frames. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018) +18. Fang, K., Wu, T.L., Yang, D., Savarese, S., Lim, J.J.: Demo2vec: Reasoning object affordances from online videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2139-2147 (2018) +19. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) +20. Furnari, A., Battiato, S., Grauman, K., Farinella, G.M.: Next-active-object prediction from egocentric videos. Journal of Visual Communication and Image Representation 49, 401-411 (2017) +21. Furnari, A., Battiato, S., Maria Farinella, G.: Leveraging uncertainty to rethink loss functions and evaluation measures for egocentric action anticipation. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018) +22. Furnari, A., Farinella, G.M.: What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6252-6261 (2019) +23. Furnari, A., Farinella, G.M.: Rolling-unrolling lstms for action anticipation from first-person video. IEEE transactions on pattern analysis and machine intelligence 43(11), 4021-4036 (2020) +24. Geng, Y., An, B., Geng, H., Chen, Y., Yang, Y., Dong, H.: End-to-end affordance learning for robotic manipulation. arXiv preprint arXiv:2209.12941 (2022) +25. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021) +26. Gong, D., Lee, J., Kim, M., Ha, S.J., Cho, M.: Future transformer for long-term action anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3052-3061 (2022) +27. Gong, J., Foo, L.G., Fan, Z., Ke, Q., Rahmani, H., Liu, J.: Diffpose: Toward more reliable 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13041-13051 (2023) +28. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022) +29. Gupta, A., Johnson, J., Fei-Fei, L., Savarese, S., Alahi, A.: Social gan: Socially acceptable trajectories with generative adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2255-2264 (2018) +30. Jiang, H., Liu, S., Wang, J., Wang, X.: Hand-object contact consistency reasoning for human grasps generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11107-11116 (2021) + +31. Judd, T., Ehinger, K., Durand, F., Torralba, A.: Learning to predict where humans look. In: 2009 IEEE 12th international conference on computer vision. pp. 2106-2113. IEEE (2009) +32. Kjellström, H., Romero, J., Kragic, D.: Visual object-action recognition: Inferring object affordances from human demonstration. Computer Vision and Image Understanding 115(1), 81-90 (2011) +33. Lee, N., Choi, W., Vernaza, P., Choy, C.B., Torr, P.H., Chandraker, M.: Desire: Distant future prediction in dynamic scenes with interacting agents. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 336-345 (2017) +34. Li, C., Zheng, P., Yin, Y., Pang, Y.M., Huo, S.: An ar-assisted deep reinforcement learning-based approach towards mutual-cognitive safe human-robot interaction. Robotics and Computer-Integrated Manufacturing 80, 102471 (2023) +35. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Wang, L., Qiao, Y.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022) +36. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018) +37. Liu, M., Tang, S., Li, Y., Rehg, J.M.: Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 704-721. Springer (2020) +38. Liu, S., Jiang, H., Xu, J., Liu, S., Wang, X.: Semi-supervised 3d hand-object poses estimation with interactions in time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14687-14697 (2021) +39. Liu, S., Tripathi, S., Majumdar, S., Wang, X.: Joint hand motion and interaction hotspots prediction from egocentric videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3282-3292 (2022) +40. Liu, T., Lam, K.M.: A hybrid egocentric activity anticipation framework via memory-augmented recurrent and one-shot representation forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13904-13913 (2022) +41. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Learning affordance grounding from exocentric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2252-2261 (June 2022) +42. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Learning visual affordance grounding from demonstration videos. IEEE Transactions on Neural Networks and Learning Systems (2023) +43. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Leverage interactive affinity for affordance learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6809-6819 (2023) +44. Mandlekar, A., Xu, D., Martín-Martín, R., Savarese, S., Fei-Fei, L.: Learning to generalize across long-horizon tasks from human demonstrations. arXiv preprint arXiv:2003.06085 (2020) +45. Mangalam, K., An, Y., Girase, H., Malik, J.: From goals, waypoints & paths to long term human trajectory forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15233-15242 (2021) +46. Mangalam, K., Girase, H., Agarwal, S., Lee, K.H., Adeli, E., Malik, J., Gaidon, A.: It is not the journey but the destination: Endpoint conditioned trajectory + +prediction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 759-776. Springer (2020) +47. Mascaró, E.V., Ahn, H., Lee, D.: Intention-conditioned long-term human egocentric action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6048-6057 (2023) +48. Mendonca, R., Bahl, S., Pathak, D.: Structured world models from human videos. arXiv preprint arXiv:2308.10901 (2023) +49. Miech, A., Laptev, I., Sivic, J., Wang, H., Torresani, L., Tran, D.: Leveraging the present to anticipate the future in videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +50. Mohamed, A., Qian, K., Elhoseiny, M., Claudel, C.: Social-stgcnn: A social spatiotemporal graph convolutional neural network for human trajectory prediction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14424-14432 (2020) +51. Myers, A., Teo, C.L., Fermüller, C., Aloimonos, Y.: Affordance detection of tool parts from geometric features. In: 2015 IEEE International Conference on Robotics and Automation (ICRA). pp. 1374-1381. IEEE (2015) +52. Nagarajan, T., Feichtenhofer, C., Grauman, K.: Grounded human-object interaction hotspots from video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8688-8697 (2019) +53. Nawhal, M., Jyothi, A.A., Mori, G.: Rethinking learning approaches for long-term action anticipation. In: European Conference on Computer Vision. pp. 558-576. Springer (2022) +54. Parsaeifard, B., Saadatnejad, S., Liu, Y., Mordan, T., Alahi, A.: Learning decoupled representations for human pose forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2294-2303 (2021) +55. Peters, R.J., Iyer, A., Itti, L., Koch, C.: Components of bottom-up gaze allocation in natural images. Vision research 45(18), 2397-2416 (2005) +56. Plizzari, C., Goletto, G., Furnari, A., Bansal, S., Ragusa, F., Farinella, G.M., Damen, D., Tommasi, T.: An outlook into the future of egocentric vision. arXiv preprint arXiv:2308.07123 (2023) +57. Qi, H., Wang, X., Pathak, D., Ma, Y., Malik, J.: Learning long-term visual dynamics with region proposal interaction networks. arXiv preprint arXiv:2008.02265 (2020) +58. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3636-3645 (2023) +59. Rodin, I., Furnari, A., Mavroeidis, D., Farinella, G.M.: Predicting the future from first person (egocentric) vision: A survey. Computer Vision and Image Understanding 211, 103252 (2021) +60. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +61. Shen, J., Dudley, J., Kristensson, P.O.: Fast and robust mid-air gesture typing for ar headsets using 3d trajectory decoding. IEEE Transactions on Visualization and Computer Graphics (2023) +62. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015) + +63. Song, D., Kyriazis, N., Oikonomidis, I., Papazov, C., Argyros, A., Burschka, D., Kragic, D.: Predicting human intention in visual observations of hand/object interactions. In: 2013 IEEE International Conference on Robotics and Automation. pp. 1608-1615. IEEE (2013) +64. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019) +65. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. Advances in neural information processing systems 27 (2014) +66. Swain, M.J., Ballard, D.H.: Color indexing. International journal of computer vision 7(1), 11-32 (1991) +67. Tu, Z., Huang, Z., Chen, Y., Kang, D., Bao, L., Yang, B., Yuan, J.: Consistent 3d hand reconstruction in video via self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +68. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +69. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016) +70. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021) +71. Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8121-8130 (2022) +72. Yang, Y., Zhai, W., Luo, H., Cao, Y., Luo, J., Zha, Z.J.: Grounding 3d object affordance from 2d interactions in images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10905-10915 (2023) +73. Yang, Y., Zhai, W., Luo, H., Cao, Y., Zha, Z.J.: Lemon: Learning 3d human-object interaction relation from 2d images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16284-16295 (2024) +74. Ye, Y., Singh, M., Gupta, A., Tulsiani, S.: Compositional video prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10353-10362 (2019) +75. Zhai, W., Cao, Y., Zhang, J., Xie, H., Tao, D., Zha, Z.J.: On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +76. Zhai, W., Luo, H., Zhang, J., Cao, Y., Tao, D.: One-shot object affordance detection in the wild. International Journal of Computer Vision 130(10), 2472-2500 (2022) \ No newline at end of file diff --git a/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/images.zip b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d087c0a5549fc89842c6fffcb33c99a3c230bf56 --- /dev/null +++ b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c00c7da3d806e5373812c6dbd2a27a3da372e6463b3ed9beb4ff7e91f7822ab5 +size 631104 diff --git a/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/layout.json b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4c490aa26e54b248265f4b44a7158942d4b86871 --- /dev/null +++ b/2024/Bidirectional Progressive Transformer for Interaction Intention Anticipation/layout.json @@ -0,0 +1,12167 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 163, + 111, + 452, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 111, + 452, + 148 + ], + "spans": [ + { + "bbox": [ + 163, + 111, + 452, + 148 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": "Zichen Zhang" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", Hongchen Luo" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", Wei Zhai" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", Yang Cao" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", and Yu Kang" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 206, + 189, + 406, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 189, + 406, + 201 + ], + "spans": [ + { + "bbox": [ + 206, + 189, + 406, + 201 + ], + "type": "text", + "content": "1 University of Science and Technology of China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 201, + 473, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 201, + 473, + 234 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 473, + 234 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 201, + 473, + 234 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Hefei Comprehensive National Science Center {zhangzichen@mail., lhc12@mail., wzhai056@, kangduyu@, forrest@}ustc.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 259, + 455, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 259, + 455, + 511 + ], + "spans": [ + { + "bbox": [ + 159, + 259, + 455, + 511 + ], + "type": "text", + "content": "Abstract. Interaction intention anticipation aims to jointly predict future hand trajectories and interaction hotspots. Existing research often treated trajectory forecasting and interaction hotspots prediction as separate tasks or solely considered the impact of trajectories on interaction hotspots, which led to the accumulation of prediction errors over time. However, a deeper inherent connection exists between hand trajectories and interaction hotspots, which allows for continuous mutual correction between them. Building upon this relationship, a novel Bidirectional prOgressive Transformer (BOT), which introduces a Bidirectional Progressive mechanism into the anticipation of interaction intention is established. Initially, BOT maximizes the utilization of spatial information from the last observation frame through the Spatial-Temporal Reconstruction Module, mitigating conflicts arising from changes of view in first-person videos. Subsequently, based on two independent prediction branches, a Bidirectional Progressive Enhancement Module is introduced to mutually improve the prediction of hand trajectories and interaction hotspots over time to minimize error accumulation. Finally, acknowledging the intrinsic randomness in human natural behavior, we employ a Trajectory Stochastic Unit and a C-VAE to introduce appropriate uncertainty to trajectories and interaction hotspots, respectively. Our method achieves state-of-the-art results on three benchmark datasets Epic-Kitchens-100, EGO4D, and EGTEA Gaze+, demonstrating superior in complex scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 521, + 441, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 521, + 441, + 533 + ], + "spans": [ + { + "bbox": [ + 160, + 521, + 441, + 533 + ], + "type": "text", + "content": "Keywords: Interaction Intention " + }, + { + "bbox": [ + 160, + 521, + 441, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 521, + 441, + 533 + ], + "type": "text", + "content": " Egocentric Videos " + }, + { + "bbox": [ + 160, + 521, + 441, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 521, + 441, + 533 + ], + "type": "text", + "content": " Transformer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 552, + 230, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 230, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 230, + 565 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 575, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 648 + ], + "type": "text", + "content": "The pursuit of future prediction on egocentric videos is centered on anticipating intentions and behaviors observed from a first-person perspective [59]. Interaction intention can be reflected in the trajectory of hand motion and contact points on the next-active object [39]. This task is novel but holds significant potential value, such as AR [34, 61], robot skill acquisition [3, 9, 24, 44, 48], hand reconstruction [30, 67] and pose estimation [27, 38]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 654, + 241, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 654, + 241, + 666 + ], + "spans": [ + { + "bbox": [ + 140, + 654, + 241, + 666 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 194, + 112, + 264, + 175 + ], + "blocks": [ + { + "bbox": [ + 194, + 112, + 264, + 175 + ], + "lines": [ + { + "bbox": [ + 194, + 112, + 264, + 175 + ], + "spans": [ + { + "bbox": [ + 194, + 112, + 264, + 175 + ], + "type": "image", + "image_path": "729d4c622d0dc0f1bf249262094c561098ee63d23f74e22b99e3041986ba1e7c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 178, + 399, + 190 + ], + "lines": [ + { + "bbox": [ + 250, + 178, + 399, + 190 + ], + "spans": [ + { + "bbox": [ + 250, + 178, + 399, + 190 + ], + "type": "text", + "content": "Error Accumulation Over Time" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 265, + 112, + 335, + 175 + ], + "blocks": [ + { + "bbox": [ + 265, + 112, + 335, + 175 + ], + "lines": [ + { + "bbox": [ + 265, + 112, + 335, + 175 + ], + "spans": [ + { + "bbox": [ + 265, + 112, + 335, + 175 + ], + "type": "image", + "image_path": "73c8ca3256aaec031421fe424af6593f591d00b95aece1223c4db66e030ec5c7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 350, + 112, + 421, + 176 + ], + "blocks": [ + { + "bbox": [ + 350, + 112, + 421, + 176 + ], + "lines": [ + { + "bbox": [ + 350, + 112, + 421, + 176 + ], + "spans": [ + { + "bbox": [ + 350, + 112, + 421, + 176 + ], + "type": "image", + "image_path": "f2e97a34d637bcc7ca8e8e111115663f3a03eda4c86273ed331d0efeda79d2b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 226 + ], + "type": "text", + "content": "Fig. 1: Error accumulation when only considering the influence from hand trajectories to interaction hotspots. Red ones represent the ground truth, while Blue ones are predictions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 254, + 483, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 483, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 483, + 410 + ], + "type": "text", + "content": "Most existing relevant research treated hand trajectory forecasting and interaction hotspots prediction as independent tasks. Bao et al. [4] proposed to predict the 3D hand trajectory from a first-person view. However, They overlooked the guiding role of interactable objects on trajectories. The prediction process was incomplete, leading to significant errors at the end of trajectories. On the other hand, there was extensive research on interaction hotspots [18,32,41,42,51,52,63]. Nevertheless, these studies often focused on the characteristics of the objects themselves, resulting in lower integration with subjective human behaviors in complex scenes and less accurate predictions. Other studies [37,39] jointly predicted future hand trajectories and interaction hotspots, improving their integration. However, these studies only considered the unidirectional influence of hand trajectories on contact points. Consequently, errors accumulated continuously with the prediction time steps (Fig. 1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 412, + 483, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 483, + 555 + ], + "type": "text", + "content": "In fact, this one-way causation was insufficient, as it disregarded the spatial correlation between the two factors. Conversely, there exists a mutual relationship between the predictions of hand trajectories and interaction hotspots. On one hand, the derived directions of hand trajectories suggest the approximate areas where contact points occur. On the other hand, there is an inherent connection between the spatial distribution of contact points and hand trajectory patterns, as in Fig. 2, which implies that examining the distribution of contact points can provide insights into the development of hand trajectories. Leveraging this relationship, to mitigate the process of error accumulation, we establish a Bidirectional Progressive (Bi-Progressive) mechanism, allowing continuous mutual refinement between trajectories and interaction hotspots throughout the entire prediction period, improving the accuracy of joint predictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "content": "In this paper, an innovative Bidirectional prOgressive Transformer (BOT) is proposed, introducing the Bi-Progressive mechanism mentioned earlier between two independent predicting branches. Firstly, the raw video input undergoes a Spatial-Temporal Reconstruction Module, which integrates the spatial information from the last observation frame and the temporal information among all observation frames. This process mitigates spatial information conflicts arising from changes in the field of view in first-person videos. Secondly, we establish separate dual branches for hand trajectory forecasting and interaction hotspots prediction, endowing them with the capability to independently perform indi" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 203, + 114, + 258, + 153 + ], + "blocks": [ + { + "bbox": [ + 203, + 114, + 258, + 153 + ], + "lines": [ + { + "bbox": [ + 203, + 114, + 258, + 153 + ], + "spans": [ + { + "bbox": [ + 203, + 114, + 258, + 153 + ], + "type": "image", + "image_path": "deff622f82bdf84eaf74a92c45fed659133792a3d4ca0acd0f30840eb3a03024.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 153, + 367, + 161 + ], + "lines": [ + { + "bbox": [ + 258, + 153, + 367, + 161 + ], + "spans": [ + { + "bbox": [ + 258, + 153, + 367, + 161 + ], + "type": "text", + "content": "(a) Long-length trajectories" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 259, + 114, + 314, + 153 + ], + "blocks": [ + { + "bbox": [ + 259, + 114, + 314, + 153 + ], + "lines": [ + { + "bbox": [ + 259, + 114, + 314, + 153 + ], + "spans": [ + { + "bbox": [ + 259, + 114, + 314, + 153 + ], + "type": "image", + "image_path": "7775fc263d940aa62090d77e1489f88bdb64940a5f53fe34c931a284d37f139f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 114, + 370, + 152 + ], + "blocks": [ + { + "bbox": [ + 316, + 114, + 370, + 152 + ], + "lines": [ + { + "bbox": [ + 316, + 114, + 370, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 114, + 370, + 152 + ], + "type": "image", + "image_path": "904b37d7bce550a3152177872f041f9a750b1a700a85c0ae1c8b349713a4c5e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 372, + 114, + 410, + 152 + ], + "blocks": [ + { + "bbox": [ + 372, + 114, + 410, + 152 + ], + "lines": [ + { + "bbox": [ + 372, + 114, + 410, + 152 + ], + "spans": [ + { + "bbox": [ + 372, + 114, + 410, + 152 + ], + "type": "image", + "image_path": "1dad43f57b2dba72cd157ecaa28db7456c94982bfb3fd3a240bc7dce26aa1a63.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 203, + 161, + 258, + 201 + ], + "blocks": [ + { + "bbox": [ + 203, + 161, + 258, + 201 + ], + "lines": [ + { + "bbox": [ + 203, + 161, + 258, + 201 + ], + "spans": [ + { + "bbox": [ + 203, + 161, + 258, + 201 + ], + "type": "image", + "image_path": "510f94951bc8fb717922f0dafc397a3eab8e83bf2a3c966b991b1223adcfb7d7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 201, + 374, + 208 + ], + "lines": [ + { + "bbox": [ + 251, + 201, + 374, + 208 + ], + "spans": [ + { + "bbox": [ + 251, + 201, + 374, + 208 + ], + "type": "text", + "content": "(b) Medium-length trajectories" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 259, + 161, + 314, + 201 + ], + "blocks": [ + { + "bbox": [ + 259, + 161, + 314, + 201 + ], + "lines": [ + { + "bbox": [ + 259, + 161, + 314, + 201 + ], + "spans": [ + { + "bbox": [ + 259, + 161, + 314, + 201 + ], + "type": "image", + "image_path": "75a02884eefae83d9f647dc935b346d9e58419ac33e0a684c204c4c0b01e9d0d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 161, + 370, + 201 + ], + "blocks": [ + { + "bbox": [ + 316, + 161, + 370, + 201 + ], + "lines": [ + { + "bbox": [ + 316, + 161, + 370, + 201 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 370, + 201 + ], + "type": "image", + "image_path": "d4757304e7c23018db96e89d3d46c7dbfc0a5d7cae50451f3e65fbbf1d8d8330.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 372, + 161, + 410, + 201 + ], + "blocks": [ + { + "bbox": [ + 372, + 161, + 410, + 201 + ], + "lines": [ + { + "bbox": [ + 372, + 161, + 410, + 201 + ], + "spans": [ + { + "bbox": [ + 372, + 161, + 410, + 201 + ], + "type": "image", + "image_path": "5f932b23785c83bbd6ab62b1bb0c2036f6aafad609740fdeeff8e9be084a4b20.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 203, + 209, + 258, + 247 + ], + "blocks": [ + { + "bbox": [ + 203, + 209, + 258, + 247 + ], + "lines": [ + { + "bbox": [ + 203, + 209, + 258, + 247 + ], + "spans": [ + { + "bbox": [ + 203, + 209, + 258, + 247 + ], + "type": "image", + "image_path": "a8cd833900d4d17f76445756c6e29d50e1e5885cad39fd30444d9b655b7f92d2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 248, + 369, + 257 + ], + "lines": [ + { + "bbox": [ + 257, + 248, + 369, + 257 + ], + "spans": [ + { + "bbox": [ + 257, + 248, + 369, + 257 + ], + "type": "text", + "content": "(c) Short-length trajectories" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 259, + 209, + 314, + 247 + ], + "blocks": [ + { + "bbox": [ + 259, + 209, + 314, + 247 + ], + "lines": [ + { + "bbox": [ + 259, + 209, + 314, + 247 + ], + "spans": [ + { + "bbox": [ + 259, + 209, + 314, + 247 + ], + "type": "image", + "image_path": "44e598bc4896d72cac03bad8b06a4feb60ce84130da800404c160e7cd0dbdbfc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 259, + 480, + 282 + ], + "lines": [ + { + "bbox": [ + 131, + 259, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 131, + 259, + 480, + 282 + ], + "type": "text", + "content": "Fig. 2: Inherent connection between different hand trajectory categories and their corresponding contact points distribution. This is conducted on Epic-Kitchens-100 [14]." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 209, + 370, + 247 + ], + "blocks": [ + { + "bbox": [ + 316, + 209, + 370, + 247 + ], + "lines": [ + { + "bbox": [ + 316, + 209, + 370, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 370, + 247 + ], + "type": "image", + "image_path": "95ac88be95a6d3be921b581de02b3ca7a4b886741ddf8b8fa43338f4006f408c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 372, + 209, + 410, + 247 + ], + "blocks": [ + { + "bbox": [ + 372, + 209, + 410, + 247 + ], + "lines": [ + { + "bbox": [ + 372, + 209, + 410, + 247 + ], + "spans": [ + { + "bbox": [ + 372, + 209, + 410, + 247 + ], + "type": "image", + "image_path": "9289fa77c6feacfb3779179cbb29ba6f5110b4730021aefb40da0096a775be4e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 304, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 399 + ], + "type": "text", + "content": "vidual prediction tasks, laying the foundation for introducing the Bi-Progressive mechanism. Subsequently, we introduce cross-attention blocks as the primary implementation of the Bi-Progressive mechanism. At each time step, the predictions of hand trajectories and interaction hotspots are alternately updated, addressing inherent errors in individual predictions. Finally, following the logic of natural human behaviors [50, 57, 74], we incorporate a Trajectory Stochastic Unit and a C-VAE for both hand trajectories and interaction hotspots, respectively, introducing appropriate uncertainties to the final prediction results." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 400, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 449 + ], + "type": "text", + "content": "We evaluate our approach on Epic-Kitchens-100 [14], EGO4D [28] and EGTEA Gaze+ [36] datasets. BOT demonstrates notable advantages in predicting both hand trajectories and interaction hotspots, achieving superior performance in anticipating interaction intentions. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 456, + 481, + 574 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 456, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 491 + ], + "type": "text", + "content": "- A novel BOT is introduced to jointly predict future hand trajectories and interaction hotspots in egocentric videos, attaining the optimal predictive performance with more lenient input conditions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 139, + 492, + 480, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 492, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 139, + 492, + 480, + 528 + ], + "type": "text", + "content": "- We introduce a novel Bidirectional Progressive mechanism that enhances the coherence between hand trajectories and interaction hotspots, thereby improving the completeness of the interaction intention." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 139, + 528, + 481, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 528, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 139, + 528, + 481, + 574 + ], + "type": "text", + "content": "- The experimental results demonstrate that our method achieves state-of-the-art results across three benchmark datasets Epic-Kitchens-100 [14], EGO4D [28], and EGTEA Gaze+ [36], demonstrating superior in complex scenarios." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 593, + 237, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 237, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 237, + 605 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Action anticipation in egocentric videos. Anticipating actions in an egocentric vision is a challenging task that includes short-term and long-term action anticipation [56,59]. Short-term action anticipation aims to predict the next action in the near future [21-23,25,49], while long-term action anticipation focuses" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "on the sequence of subsequent actions performed by the camera wearer [26, 40, 47, 53]. These endeavors contribute significantly to the comprehension of actions in first-person videos. However, concerning interaction intention, action anticipation still yields information of relatively high dimensionality. Anticipating a semantic label does not furnish direct guidance for the robot, as it lacks specific instructions on task execution. Comparatively, hand trajectories and interaction hotspots offer a more intuitive representation of interaction intention, constituting a lower-dimensional depiction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 331 + ], + "type": "text", + "content": "Human trajectory forecasting. The forecasting of human trajectories was extensively studied over the years [1,2,12,13,45,54]. However, most of them were based on third-person video datasets, falling short of capturing intentions related to human-object interactions. Since hands serve as the primary medium through which humans interact with the world, predicting hand trajectories in egocentric videos can aid embodied AI in comprehending human-object interaction intentions. The approaches in [4,37,39] addressed the issue of hand trajectory prediction in egocentric videos. However, they did not account for the influence of contact points with objects on trajectories. This limitation resulted in the continuous accumulation of prediction errors over time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": "Interaction hotspots prediction. Interaction hotspots prediction aims to locate where the hand-object interaction happens. Some research aimed to predict the next interactive object [5, 6, 15, 17, 20], going further, some work anticipated specific interaction regions on objects [42, 43, 52, 72, 73, 75, 76], yielding more precise information for interaction. Differing from image input, our task aims to generate more precise predictions regarding the regions of contact points by utilizing egocentric video inputs. The incorporation of first-person videos is intended to focus on precisely localizing interaction hotspots for distinct human motions. In this task, we leverage the predicted hand trajectories to improve the accuracy of interaction hotspots predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 470, + 202, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 202, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 202, + 483 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": "Problem Setup. Given an original input video " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_1, I_2, \\dots, I_N\\}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{N}}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": " is the last observation frame, the task is predicting future hand trajectories " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = \\{h_{\\mathrm{N + 1}}, h_{\\mathrm{N + 2}}, \\dots, h_{\\mathrm{N + T - 1}}\\}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": " in future keyframes " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\{I_{\\mathrm{N + 1}}, I_{\\mathrm{N + 2}}, \\dots, I_{\\mathrm{N + T - 1}}\\}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": " and the contact point " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{C}}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": " in the contact frame " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "I_{N + T}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": ". We subsequently convert the contact point " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{C}}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": " into a heatmap by centering a Gaussian distribution, generating interaction hotspots " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{I}}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": ". To ensure the continuity of the trajectory as well as the consistency of the background, both " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{I}}" + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": " are projected onto the last observation frame, which enhances the overall coherence of the predictions [39]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 609, + 204, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 204, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 204, + 620 + ], + "type": "text", + "content": "3.1 Network" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Overview. As shown in Fig. 3, given the original input video " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", we aim to predict the future hand trajectory " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and interaction hotspots " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{I}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ". Initially, to alleviate information conflicts arising from changes in the field of view, the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 247 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 247 + ], + "type": "image", + "image_path": "1d22189a85db342bb468c6b4afa5b756fc13ca2d042ab9455079a017bb1c51ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 255, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 255, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 482, + 289 + ], + "type": "text", + "content": "Fig. 3: Overview of the Bidirectional Progressive Transformer. It follows a dual-branch structure with a Bi-Progressive Enhancement Module between them to anticipate future hand trajectories and interaction hotspots." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 532 + ], + "type": "text", + "content": "original input video undergoes a Spatial-Temporal Reconstruction Module that integrates spatial information from the last observation frame and temporal information from the entire video, producing spatial features for each future timestep. Subsequently, these features are separated, with the last one serving as the contact feature while the preceding ones are for trajectory features. These features are then fed into the interaction hotspots and trajectory prediction branches, respectively. The former employs an overlap patch embedding layer to enhance attention at the center of the field of view, while the latter introduces temporal order to trajectory features through Cross-Transformers [11]. Simultaneously, we employ a Bi-Progressive Enhancement Module consisting of cross-attention blocks between dual branches, introducing a Bi-Progressive mechanism to the joint prediction. This alternately and iteratively corrects the features of trajectories and interaction hotspots in chronological order, thereby mitigating the accumulation of prediction errors over time. Finally, considering the inherent uncertainty in natural human behaviors [50, 57, 74], we introduce a Stochastic Unit for trajectory forecasting while employing a C-VAE [62] for interaction hotspots, adding appropriate uncertainty during the completion of position decoding to achieve comprehensive predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "Spatial-Temporal Reconstruction. When dealing with egocentric videos, both 3D convolutions [19] and self-attention mechanisms [68] struggle to address spatial information conflicts arising from changes in views. Moreover, predictions are projected onto the last observation frame. Taking these factors into account, we establish the Spatial-Temporal Reconstruction Module. On one hand, we employ optical flow containing more comprehensive motion information as the temporal input, on the other hand, we isolate the last observation frame as the spatial input. Subsequently, the temporal input is processed by a pre-trained Temporal Segment Network [22,69], extracting low-dimensional temporal features " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": ", and the spatial input is passed through a Segformer-b2 [70], extracting spatial features " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": " from the last observation frame. " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": " are then fed into a" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 373, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 373, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 373, + 129 + ], + "type": "text", + "content": "Cross-Transformer [11]. The " + }, + { + "bbox": [ + 132, + 116, + 373, + 129 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 116, + 373, + 129 + ], + "type": "text", + "content": "-th block is expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 242, + 140, + 481, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 140, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 242, + 140, + 481, + 152 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {0} = \\mathbf {s} _ {\\text {p a t c h}} + \\mathbf {s} _ {\\text {p o s}}, \\quad \\mathbf {x} = \\mathbf {x} + \\mathbf {x} _ {\\text {p o s}}, \\tag {1}", + "image_path": "e157dff1ebe3dc2dcd8f57b26a166b6b3c06e3ac58193f8c9e2e953809780f1e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 224, + 154, + 481, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 154, + 481, + 167 + ], + "spans": [ + { + "bbox": [ + 224, + 154, + 481, + 167 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {i} = \\mathbf {s} _ {i - 1} + \\mathbf {M C A} (\\mathbf {L N} (\\mathbf {s} _ {i - 1}), \\mathbf {L N} (\\mathbf {x})), \\tag {2}", + "image_path": "397a36ac53f5cdeac27af3d03234ae0c5f7ced3d15ba599ec53439eeb78bf454.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 262, + 170, + 481, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 170, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 262, + 170, + 481, + 182 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {i} = \\mathbf {y} _ {i} + \\mathbf {F F N} (\\mathbf {L N} (\\mathbf {y} _ {i})). \\tag {3}", + "image_path": "aa12f85abe925bad1dc68b7374ac1a4423614861403f16e3527035f89dbad5ed.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "text", + "content": " represent the original spatial and temporal inputs, respectively. " + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{patch}" + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "text", + "content": " is encoded by overlap patch embedding layers from spatial features " + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "text", + "content": ", enhancing the network's focus on the center of the field of view. " + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "text", + "content": " is encoded by positional embedding layers from temporal features " + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "text", + "content": ", improving the overall sequentiality. Furthermore, FFN refers to the Feed Forward Layer, comprising a two-layer MLP with an expanding ratio " + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 191, + 482, + 310 + ], + "type": "text", + "content": ", and Layer normalization (LN) is applied before each block. MCA stands for Multi-head Cross-Attention, where query, key, and value are processed through distinct linear layers: MCA(X,Y) = Attention(W^QX, W^KY, W^VY). The output of this module corresponds to the spatial features for each prediction frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 312, + 481, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 481, + 347 + ], + "type": "text", + "content": "Interaction Hotspots Prediction Branch. Due to the spatial distribution of contact points closer to the central field of view, the contact feature " + }, + { + "bbox": [ + 130, + 312, + 481, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{C}}" + }, + { + "bbox": [ + 130, + 312, + 481, + 347 + ], + "type": "text", + "content": " undergoes processing through an overlap patch embedding layer:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 265, + 358, + 481, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 358, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 265, + 358, + 481, + 370 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {o}} = \\text {o v e r l a p} (\\mathbf {F} _ {\\mathrm {C}}). \\tag {4}", + "image_path": "014e1eaf70a7360d35af829f53c37b11e1cf872f6a41ef88d41dbb2c04399bc9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 360, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 360, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 360, + 392 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 380, + 360, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_C^0" + }, + { + "bbox": [ + 130, + 380, + 360, + 392 + ], + "type": "text", + "content": " represents the processed contact feature." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 393, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 464 + ], + "type": "text", + "content": "Hand Trajectory Prediction Branch. We employ original trajectory features as primary features and utilize features from the previously predicted frame as conditional features, subsequently merging them through the cross-attention block " + }, + { + "bbox": [ + 130, + 393, + 482, + 464 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 393, + 482, + 464 + ], + "type": "text", + "content": ", ensuring the coherence of the trajectory over time. Assuming the trajectory features are " + }, + { + "bbox": [ + 130, + 393, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{\\mathrm{N} + 1},\\mathbf{F}_{\\mathrm{N} + 2},\\dots ,\\mathbf{F}_{\\mathrm{N} + \\mathrm{T} - 1}\\}" + }, + { + "bbox": [ + 130, + 393, + 482, + 464 + ], + "type": "text", + "content": " and the feature map of the last observation frame is " + }, + { + "bbox": [ + 130, + 393, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{N}}" + }, + { + "bbox": [ + 130, + 393, + 482, + 464 + ], + "type": "text", + "content": ", they are updated in chronological order:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 214, + 475, + 481, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 475, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 214, + 475, + 481, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {t} ^ {\\mathrm {o}} = f \\left(\\mathbf {F} _ {\\mathrm {t}}, \\mathbf {F} _ {\\mathrm {t} - 1}\\right), \\quad t \\in [ \\mathrm {N} + 1, \\mathrm {N} + \\mathrm {T} - 1 ]. \\tag {5}", + "image_path": "c6023b91c75061209f58a782d74ae075c2b236dc6500cf785f57431fadaf3239.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 498, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 510 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 498, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^{\\mathrm{o}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 510 + ], + "type": "text", + "content": " stands for the updated trajectory feature, " + }, + { + "bbox": [ + 130, + 498, + 482, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 498, + 482, + 510 + ], + "type": "text", + "content": " is the prediction time step." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "H-O Bi-Progressive Enhancement. Expanding on the groundwork laid by two distinct prediction branches, and to mitigate the accumulation of errors as predictions unfold over time, we introduce a Bi-Progressive Enhancement Module (Fig. 4), which alternately corrects the predictions of trajectories " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and interaction hotspots " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{I}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". Specifically, owing to observation frames and the continuity of hand trajectories, the trajectory feature " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{N} + 1}^{\\mathrm{o}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " at the initial predicted time step " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{N} + 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " demonstrates a high level of accuracy. Exploiting this, we enhance the precision of the contact feature " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{o}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " through the cross-attention block " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", thereby minimizing the error with the ground truth. Moving to the subsequent time step " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{N} + 2" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", we employ the first refined contact feature " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{r}_1}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " from the previous time step " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{N} + 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " as a condition to correct the trajectory feature " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{N} + 2}^{\\mathrm{o}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". Following this, the corrected trajectory feature is utilized to further refine the contact feature. This iterative correction process persists until all trajectory" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "features have undergone a complete round of updates. It can be represented as a recursive procedure:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 270, + 151, + 481, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 151, + 481, + 164 + ], + "spans": [ + { + "bbox": [ + 270, + 151, + 481, + 164 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {N} + 1} ^ {\\mathrm {r}} = \\mathbf {F} _ {\\mathrm {N} + 1} ^ {\\mathrm {o}}, \\tag {6}", + "image_path": "29ecba94a5bbb26bfdc1a41685d60d1b3a940fcaa63b92302a007d18332796d4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 166, + 481, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 166, + 481, + 179 + ], + "spans": [ + { + "bbox": [ + 214, + 166, + 481, + 179 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {1}} = f \\left(\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {o}}, \\mathbf {F} _ {\\mathrm {N} + 1} ^ {\\mathrm {r}}\\right), \\quad \\mathbf {F} _ {\\mathrm {N} + 2} ^ {\\mathrm {r}} = f \\left(\\mathbf {F} _ {\\mathrm {N} + 2} ^ {\\mathrm {o}}, \\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {1}}\\right), \\tag {7}", + "image_path": "532ed1892d91e5659eab9e7a6921a3195d4fafec210d63f4cf06d07e2d85a4b1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 298, + 184, + 312, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 184, + 312, + 190 + ], + "spans": [ + { + "bbox": [ + 298, + 184, + 312, + 190 + ], + "type": "interline_equation", + "content": "\\cdot \\cdot \\cdot", + "image_path": "bed954ed968780a2d3d5a4452ac92e5ff192cd3a10fb2cd620125ab368be7b36.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 196, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 196, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 212, + 196, + 481, + 209 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {3}} = f \\left(\\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {2}}, \\mathbf {F} _ {\\mathrm {N} + 3} ^ {\\mathrm {r}}\\right), \\quad \\mathbf {F} _ {\\mathrm {N} + 4} ^ {\\mathrm {r}} = f \\left(\\mathbf {F} _ {\\mathrm {N} + 4} ^ {\\mathrm {o}}, \\mathbf {F} _ {\\mathrm {C}} ^ {\\mathrm {r} _ {3}}\\right). \\tag {8}", + "image_path": "9503a9a200d36b872405dc52e108d056bb876973b03b627bff3ab22db94fc670.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^{\\mathrm{r}}" + }, + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "text", + "content": " represents the refined trajectory feature map at time step " + }, + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "inline_equation", + "content": "t\\in [\\mathrm{N} + 1,\\mathrm{N} + \\mathrm{T} - 1]" + }, + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{r},i}" + }, + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "text", + "content": " denotes the contact feature after the " + }, + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 219, + 361, + 290 + ], + "type": "text", + "content": "-th refinement. Via this module, we attain refined features for both trajectories and interaction hotspots, significantly reducing error accumulation in predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": "Interaction Hotspots Decoder. Given the presence of inherent uncertainty associated with interaction hotspots [39, 41, 42], we employ a C-VAE [62] as the decoder. Due to the excessive randomness associated with directly decoding the coordinates of contact points, we instruct the C-VAE to output a feature map and select the location of the maximum value from this map as the output, thereby confining the uncertainty within a reasonable range. The C-VAE follows an encoder-decoder structure. For the encoder block, we concatenate the input " + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "inline_equation", + "content": "\\varGamma" + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": " with the condition " + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": " through an encoding function " + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{e}}" + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": ", obtaining features in latent space " + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": " which parameterized by mean " + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": " and co-variance " + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": ". In this context, the input " + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "inline_equation", + "content": "\\varGamma" + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": " is the ground truth heatmap, while the condition " + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": " is the refined contact feature " + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{C}}^{\\mathrm{r}_3}" + }, + { + "bbox": [ + 130, + 291, + 361, + 483 + ], + "type": "text", + "content": ". In the decoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "text", + "content": "block, the first step involves sampling " + }, + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "text", + "content": " from the latent space and concatenating it with condition " + }, + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "text", + "content": ". Following that, the output heatmap " + }, + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\Gamma}" + }, + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "text", + "content": " is reconstructed by a decoding function " + }, + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_d" + }, + { + "bbox": [ + 130, + 483, + 481, + 529 + ], + "type": "text", + "content": ". In summary, the entire process can be formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 198, + 531, + 481, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 531, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 198, + 531, + 481, + 546 + ], + "type": "interline_equation", + "content": "\\mu , \\sigma = \\mathbf {P} _ {\\mathrm {e}} (\\Gamma , \\mathrm {C}), \\quad \\hat {\\Gamma} = \\mathbf {P} _ {\\mathrm {d}} (\\mathrm {Z}, \\mathrm {C}), \\quad \\mathrm {Z} \\sim \\mathcal {N} (\\mu , \\sigma). \\tag {9}", + "image_path": "ea79621c00f81d4d1ee2c2b8036498a22ff1a50447d501f146607cfccd54c654.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": " represents a normal distribution, " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\hat{\\Gamma}" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": " is the reconstructed heatmap. Both encoding and decoding blocks consist of two MLP layers. We introduce KL-Divergence and the reconstruction error as the overall loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 252, + 598, + 481, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 598, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 252, + 598, + 481, + 612 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} (\\Gamma , \\hat {\\Gamma}) = (\\Gamma - \\hat {\\Gamma}) ^ {2}, \\tag {10}", + "image_path": "2a101e0a1a2c450503511d81321be22f5fb60b85043705cdba86b1e23025ef14.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 222, + 614, + 481, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 614, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 222, + 614, + 481, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k l} (\\mu , \\sigma) = - \\mathbf {K L} \\left[ \\mathcal {N} (\\mu , \\sigma) ^ {2} | | \\mathcal {N} (0, 1) \\right], \\tag {11}", + "image_path": "b329b26d4d108859c3b9c6b5931d13002e11f11c05d788421fef365d327f10d6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 263, + 631, + 481, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 631, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 263, + 631, + 481, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda \\mathcal {L} _ {k l}. \\tag {12}", + "image_path": "958130d1dc36bb61adec9687d0672d8158b96e29cc0f023d98129d86c891f78b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 653, + 268, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 268, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 268, + 666 + ], + "type": "text", + "content": "In this paper, we set " + }, + { + "bbox": [ + 131, + 653, + 268, + 666 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 131, + 653, + 268, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 369, + 243, + 481, + 399 + ], + "blocks": [ + { + "bbox": [ + 369, + 243, + 481, + 399 + ], + "lines": [ + { + "bbox": [ + 369, + 243, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 369, + 243, + 481, + 399 + ], + "type": "image", + "image_path": "0c48a0ee08548f20e60eeb879a31deebec6c2a74dcbd5c54e4734051926f5c09.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 406, + 482, + 451 + ], + "lines": [ + { + "bbox": [ + 366, + 406, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 366, + 406, + 482, + 451 + ], + "type": "text", + "content": "Fig. 4: Bi-Progressive Enhancement Module. The output feature map is enclosed in purple boxes." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "spans": [ + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "text", + "content": " hand position " + }, + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{C}}" + }, + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "text", + "content": " contact point " + }, + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{21}, \\mathcal{R}_{22}" + }, + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "text", + "content": " elliptical region " + }, + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_2" + }, + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "text", + "content": " uncertain region " + }, + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "text", + "content": " ellipse tilt angle" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 127, + 449, + 189 + ], + "blocks": [ + { + "bbox": [ + 168, + 127, + 449, + 189 + ], + "lines": [ + { + "bbox": [ + 168, + 127, + 449, + 189 + ], + "spans": [ + { + "bbox": [ + 168, + 127, + 449, + 189 + ], + "type": "image", + "image_path": "b824b0898f3f03875f2d568b7d2c2bd9ef704fca6b36a17b95f7ef5500a8f536.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 196, + 482, + 220 + ], + "lines": [ + { + "bbox": [ + 131, + 196, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 196, + 482, + 220 + ], + "type": "text", + "content": "Fig. 5: Uncertain regions of the hand trajectory. It is derived through the intersection of two ellipses, corresponding to the Eq. 14-19. (i.e., the purple region)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 483, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 483, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 483, + 407 + ], + "type": "text", + "content": "Hand Trajectory Decoder. Similar to interaction hotspots, hand trajectories also exhibit uncertainty [29, 33, 46]. Given that trajectory feature maps with distinct distribution characteristics have been obtained through the Bi-Progressive Enhancement Module, in conjunction with sampling efficiency, we don't utilize C-VAE, rather, we introduce a Trajectory Stochastic Unit to incorporate uncertainty into trajectory forecasting by calculating uncertain regions for each prediction time step instead. Since optimizing directly using point coordinates for loss calculation can pose challenges, we initially introduce a Gaussian distribution for each ground truth hand position to generate the ground truth heatmap. Subsequently, we calculate the loss between the corrected feature map and the generated ground truth heatmap, thereby enabling the network to learn the distribution of potential hand positions. The associated loss function " + }, + { + "bbox": [ + 130, + 251, + 483, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 130, + 251, + 483, + 407 + ], + "type": "text", + "content": " is based on MSE Loss:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 238, + 423, + 482, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 423, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 238, + 423, + 482, + 458 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h} = \\frac {1}{T - 1} \\sum_ {t = N + 1} ^ {N + T - 1} \\left(\\mathbf {F} _ {t} ^ {\\mathrm {r}} - \\mathbf {G} _ {t}\\right) ^ {2}. \\tag {13}", + "image_path": "2d57ae43da2171ecd3a705a9f2bd52c7c19b4386a839cec68e5c4d8ea7ac9e6b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 475, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 482, + 559 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 131, + 475, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 131, + 475, + 482, + 559 + ], + "type": "text", + "content": " represents the ground truth heatmap at time step " + }, + { + "bbox": [ + 131, + 475, + 482, + 559 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 475, + 482, + 559 + ], + "type": "text", + "content": ". We designate the coordinates of the maximum value within each feature map as the predicted trajectory points, which are then fed into the stochastic unit to generate uncertain regions. The size of the region is determined by the time step as well as the length of trajectory segments in preceding and subsequent stages. The shape is constrained by the original image dimensions and the orientation of the trajectory segments. Formally, the uncertain region (Fig. 5) can be shown as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 241, + 578, + 482, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 578, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 241, + 578, + 482, + 592 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {t} \\sim (t, h _ {t - 1}, h _ {t}, h _ {t + 1}, \\mathrm {W}, \\mathrm {L}). \\tag {14}", + "image_path": "61fa73de20923d4e5004648596664c6a4e45c1ede0e1076669155c3843bee2bf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 609, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 482, + 634 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 131, + 609, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 131, + 609, + 482, + 634 + ], + "type": "text", + "content": " is the uncertain region in future time step " + }, + { + "bbox": [ + 131, + 609, + 482, + 634 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 609, + 482, + 634 + ], + "type": "text", + "content": ", W and L are the width and length of the input video, respectively. " + }, + { + "bbox": [ + 131, + 609, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 131, + 609, + 482, + 634 + ], + "type": "text", + "content": " satisfies the following constraints:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 269, + 653, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 269, + 653, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {t} = \\mathcal {R} _ {t 1} \\cap \\mathcal {R} _ {t 2}. \\tag {15}", + "image_path": "bb026c3ab8735b8a35d9315b662f2e11cf010a225818ba8b1a6606029a68dd21.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{t1}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{t2}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " are elliptical regions constrained by " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "h_{t+1}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", which can be represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 152, + 481, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 152, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 186, + 152, + 481, + 178 + ], + "type": "interline_equation", + "content": "\\frac {\\left(\\mathbf {x} _ {t} \\cos \\theta_ {t \\omega} - \\mathbf {y} _ {t} \\sin \\theta_ {t \\omega}\\right) ^ {2}}{\\mathbf {a} _ {t \\omega} ^ {2}} + \\frac {\\left(\\mathbf {x} _ {t} \\sin \\theta_ {t \\omega} + \\mathbf {y} _ {t} \\cos \\theta_ {t \\omega}\\right) ^ {2}}{\\mathbf {b} _ {t \\omega} ^ {2}} \\leq 1, \\tag {16}", + "image_path": "202cf88f68cbc15ac66bb12a0f3a23ac751f0ec7128f5fb3237d2cfc545cf34c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 202, + 180, + 481, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 180, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 202, + 180, + 481, + 206 + ], + "type": "interline_equation", + "content": "\\mathbf {a} _ {t \\omega} = \\alpha \\beta^ {t - 1} \\mathbf {d} ^ {\\frac {1}{2}} \\left(h _ {t + \\omega - 1}, h _ {t + \\omega - 2}\\right), \\quad \\mathbf {b} _ {t \\omega} = \\sqrt {\\frac {\\mathrm {L}}{\\mathrm {W}}} \\mathbf {a} _ {t \\omega}, \\tag {17}", + "image_path": "916f6e3f2602d6405ad8319f6ae744f3fca90865422679b8dc1f631b91f91fa2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 237, + 209, + 481, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 209, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 237, + 209, + 481, + 222 + ], + "type": "interline_equation", + "content": "\\theta_ {t \\omega} = \\tan^ {- 1} (\\mathbf {k} \\left(h _ {t + \\omega - 1}, h _ {t + \\omega - 2}\\right)), \\tag {18}", + "image_path": "07062502f70780d22978ce670f76a9eabe220b010daf7a0da8bc8e5dda40c303.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 232, + 224, + 481, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 224, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 232, + 224, + 481, + 236 + ], + "type": "interline_equation", + "content": "\\omega \\in [ 1, 2 ], t \\in [ 1, 2, 3, 4 ], \\mathcal {O} _ {\\mathrm {C}} = h _ {5}. \\tag {19}", + "image_path": "898ed2b4888ea66d400a9c7980cab792bea2e31833917a680e01a230f0502c23.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": "Specifically, " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": " denotes the overall scale of " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": " signifies the temporal growth rate, " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": " stands for the ellipse tilt angle, " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_t,\\mathbf{y}_t)" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": " represent the coordinates of " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{k}(\\mathrm{X},\\mathrm{Y})" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{d}(\\mathrm{X},\\mathrm{Y})" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": " represent the slope and Euclidean distance between points " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathrm{X}" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathrm{Y}" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": ", respectively. In BOT, we set " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\alpha = 1.5" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\beta = 2" + }, + { + "bbox": [ + 131, + 250, + 482, + 299 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 323, + 276, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 276, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 276, + 335 + ], + "type": "text", + "content": "3.2 Training and Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 350, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 479, + 374 + ], + "type": "text", + "content": "Training. During training, we freeze parameters in TSN [69]. Training loss combines both trajectory loss and interaction hotspots loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 274, + 388, + 481, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 388, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 274, + 388, + 481, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {h} + \\varphi \\mathcal {L} _ {o}. \\tag {20}", + "image_path": "52cc1b946f1117d168611b0ebe8baa1f244ad36d4b22f89a504544faeda8fef6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 411, + 255, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 255, + 423 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 255, + 423 + ], + "type": "text", + "content": "In this paper, we set " + }, + { + "bbox": [ + 131, + 411, + 255, + 423 + ], + "type": "inline_equation", + "content": "\\varphi = 1" + }, + { + "bbox": [ + 131, + 411, + 255, + 423 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 425, + 479, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 425, + 479, + 461 + ], + "spans": [ + { + "bbox": [ + 131, + 425, + 479, + 461 + ], + "type": "text", + "content": "Inference. We use a Trajectory Stochastic Unit and a C-VAE to introduce uncertainty during inference. Following previous work [39, 46], we report the minimum among 20 samples for trajectory evaluation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 230, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 230, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 230, + 500 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 518, + 204, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 204, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 204, + 529 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We conducted experiments on three first-person video datasets Epic-Kitchen s-100 [14], EGO4D [28], and EGTEA Gaze+ [36]. Epic-Kitchens-100 [14] and EGTEA Gaze+ [36] predominantly include kitchen scenes, while EGO4D [28] encompasses a broader range of activities such as handicrafts, painting, and outdoor tasks. For each dataset, the observation time is set to 2 seconds. We configure the prediction time horizon T to 5, corresponding to 4 trajectory points and one contact point for each hand. Specifically, we predict 1-second trajectories on Epic-Kitchens-100 [14] and EGO4D [28], while on EGTEA Gaze+ [36], 0.5-second trajectories are forecast. Additional dataset information is available in the supplementary materials." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 280, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 280, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 280, + 128 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": "In EGO4D [28], We use the method [39] to automatically generate hand trajectories, while contact points are manually annotated based on the last observation frame and the contact frame. We employ GMFLOW [71] to generate optical flow from EGO4D [28] and EGTEA Gaze+ [36]. For each dataset, we resize the original videos to " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "456 \\times 256" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": " size, extracting both trajectory and contact features of size " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "114 \\times 64" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": ". In the Spatial-Temporal Reconstruction Module, we set the dimension " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": " in Cross-Transformer to 1024, with depth " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "= 6" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": " and attention head number " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "= 8" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": ". For the trajectory forecasting branch and H-O Bi-Progressive Enhancement Module, we configure the dimension " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": " to 256, with depth " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "= 2" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": " and head number " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "= 8" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": ". Additionally, for the introduced C-VAE, we set the dimension of both input and condition to " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "114 \\times 64" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": ", while the hidden and latent sizes are set to 1024 and 2048, respectively. Our method is implemented by PyTorch and trained with the AdamW optimizer. We train the model for 100 epochs on 4 NVIDIA 3090 GPUs with an initial learning rate of " + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "inline_equation", + "content": "3e - 5" + }, + { + "bbox": [ + 130, + 135, + 482, + 303 + ], + "type": "text", + "content": " and a batch size of 40." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 483, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 483, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 483, + 412 + ], + "type": "text", + "content": "For comparison, We first select methods for individual tasks including Kalman Filter (KF) [8], Seq2Seq [65], Hotspots [52] and USST [4]. Additionally, the methods for image/video understanding are chosen as the compared methods, which lack a joint mechanism. The former includes VIT [16], HRNet [64], UNet [60], Segformer [70], while the latter contains I3D [10], Slowfast [19], Timesformer [7], Uniformer-V2 [35], StillFast [58]. Furthermore, we select methods with high task relevance and joint design, including FHOI [37] and OCT [39]. We conduct experiments using these approaches on Epic-kithens-100 [14], EGO4D [28], and EGTEA Gaze+ [36] datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 429, + 257, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 257, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 257, + 440 + ], + "type": "text", + "content": "4.3 Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 508 + ], + "type": "text", + "content": "Hand Trajectory Evaluation. The predicted hand trajectories are referenced to the last observation frames. ADE [39] and FDE [39] are chosen as hand trajectory evaluation metrics, which respectively represent the overall accuracy and endpoint accuracy of trajectories. Detailed descriptions are available in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 508, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 556 + ], + "type": "text", + "content": "Interaction Hotspots Evaluation. For all datasets, we downsample each predicted heatmap to the size of " + }, + { + "bbox": [ + 130, + 508, + 482, + 556 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 508, + 482, + 556 + ], + "type": "text", + "content": ". We choose SIM [66], AUC-J [31], and NSS [55] as interaction hotspots evaluation metrics. For a detailed explanation of the metrics, please refer to the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 574, + 373, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 373, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 373, + 586 + ], + "type": "text", + "content": "4.4 Quantitative and Qualitative Comparisons" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "As shown in Tab. 1, our method achieves state-of-the-art performance in both trajectory prediction and interaction hotspots anticipation, with comparable leading margins across all three datasets. Taking the EGO4D dataset [28] with the richest variety of scenes as an example. In trajectory prediction, to ensure fair comparison, we additionally present results for OCT and BOT with only one sampling iteration. Under this single-sampling condition, BOT outperforms" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 183, + 480, + 373 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 1: Results of different models. Best results are depicted using red (sample 20 times) and blue (sample once), respectively. unilateral, non-joint and joint denote prediction approaches for individual tasks, for two tasks without a collaborative design, and with a joint module, respectively. “-” denotes the vacancy of the value. “*” represents the results obtained from 20 random samplings. ↑ and ↓ indicate a better performance when the value of the metric increases and decreases, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 183, + 480, + 373 + ], + "lines": [ + { + "bbox": [ + 135, + 183, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 135, + 183, + 480, + 373 + ], + "type": "table", + "html": "
MethodEK100 [14]EGO4D [28]EGTEA [36]
ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑ADE ↓ FDE ↓ SIM ↑ AUC-J ↑ NSS ↑
unilateralCenter0.240.240.090.610.330.260.250.070.570.150.320.320.090.630.27
KF [8]0.330.32--------0.490.48---
Seq2Seq [65]0.180.14--------0.180.14---
Hotspots [52]--0.150.660.53-------0.150.710.69
USST [4]0.170.18---0.180.18---0.180.18---
non-jointI3D [10]0.220.230.130.560.460.220.220.120.570.420.230.220.130.600.49
SLOWFAST [19]0.220.240.140.620.470.230.220.130.580.440.210.230.140.580.46
Timesformer [7]0.230.230.140.600.420.240.240.120.590.400.210.220.140.580.45
Uniformer-V2 [35]0.200.200.130.590.430.210.220.120.620.400.200.210.130.580.46
StillFast [58]0.180.190.160.620.610.190.190.150.650.570.180.190.160.600.55
UNet [60]0.190.180.150.620.570.180.180.140.590.550.180.180.140.600.60
HRNet [64]0.180.190.160.620.630.190.200.140.620.580.180.180.160.630.65
VIT [16]0.180.190.140.650.630.210.210.130.590.450.190.190.150.630.61
CrossVIT [11]0.190.200.140.620.590.190.200.130.600.550.200.210.130.590.49
Segformer-b2 [70]0.170.170.160.640.660.160.160.150.640.620.170.170.160.630.65
jointFHOI [37]0.360.350.100.550.39-----0.340.340.120.630.42
OCT [39]0.150.17---0.150.16---0.160.16---
OCT* [39]0.120.110.190.720.720.120.120.220.730.880.140.140.230.751.01
Ours0.120.14---0.110.13---0.120.14---
Ours*0.100.080.340.812.070.090.080.320.771.700.090.080.320.802.03
", + "image_path": "c301abb7d3e794f12317efefa370d808e02cc57e071df447f3352c89c4ba3d81.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": "Segformer-b2, the best method in image/video understanding, with respective leads of " + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "inline_equation", + "content": "31.3\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "inline_equation", + "content": "18.8\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": " in ADE and FDE. Compared to OCT, BOT maintains advantages of " + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "inline_equation", + "content": "26.7\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "inline_equation", + "content": "18.8\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": " in ADE and FDE, respectively. These advantages persist even with 20 samplings, while the lead in FDE expands to " + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "inline_equation", + "content": "33.3\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": ", underscoring the rationale behind BOT's stochastic sampling mechanism. Regarding interaction hotspots, we achieve the best results across all metrics. Specifically, BOT surpasses the second-best approach (OCT) by " + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "inline_equation", + "content": "45.5\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": " in SIM, " + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": " in AUC-J, and " + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "inline_equation", + "content": "93.2\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": " in NSS, respectively. In summary, our method leverages the intrinsic correlations between hand trajectories and interaction hotspots more effectively, enhancing overall prediction accuracy of interaction intention, and exhibiting robustness across the three benchmark datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "To further explore the subjective outcomes of models, we report the visualization results of three datasets, as shown in Fig. 6. Compared to other methods, our model provides more accurate prediction results in both hand trajectories and interaction hotspots. When there is significant movement of hands (column 1,3,8), the BOT's predictions exhibit the highest accuracy. It indicates that our Spatial-Temporal Reconstruction Module can efficiently utilize the human motion information to exploit the hand interaction intention and thus accurately predict the hand trajectory. In cases where there are no identifiable interactive objects in the field of view (column 6), our method anticipates more accurate interaction intentions. It illustrates that our Bi-Progressive mechanism can eliminate the uncertainty in the interaction process by mining the intrinsic connection between" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 114, + 479, + 277 + ], + "blocks": [ + { + "bbox": [ + 126, + 114, + 479, + 277 + ], + "lines": [ + { + "bbox": [ + 126, + 114, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 126, + 114, + 479, + 277 + ], + "type": "image", + "image_path": "579663f562ff6b4dae9a41557ab1dce9686104d2f72a1ad76cc94b526d628c84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 278, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 332 + ], + "type": "text", + "content": "Fig. 6: Visualization of future hand trajectories and interaction hotspots on EK100, EGO4D and EGETA. We show the visualization results of our model, the joint prediction model (OCT) [39], the segmentation model (Segformer-b2) [70], and the video understanding model (Uniformer-V2) [35]. The green represents the trajectory of the left hand, while the red denotes the trajectory of the right hand." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": "the hand trajectory and the contact points, which results in a more accurate inference in the interaction region and the hand trajectory. Moreover, our predictions closely align with the actual scenario when both hands simultaneously exhibit trajectories and interaction hotspots (columns 4,5), which suggests the strategy of differentiating between left and right hands during the prediction of interaction intentions, thereby mitigating cross-hand interference. More visualization results are available in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 455, + 238, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 238, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 238, + 467 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "content": "Different joint approaches. To explore the impact of the Bi-Progressive mechanism on model performance, we experiment with five different joint configurations: Individual, Hand, Object, Bidirectional and Bi-Progressive. The results are shown in Tab. 2 and Fig. 7. It suggests that the network is challenging to accurately predict trajectories and interaction hotspots during individual predictions, and the separate incorporation of hands and objects can effectively improve the prediction accuracy of interaction hotspots. With the combination of hand and object information, the prediction performance of the trajectory and interaction region is greatly improved. However, there is still a gap compared to the Bi-Progressive mechanism. Further, Fig. 8 shows that the introduction of the progressive mechanism plays a pivotal role in mitigating the accumulation of prediction errors over time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Uncertainty in anticipation. Uncertainty is a crucial consideration in interaction intention anticipation. Following the logic of natural human behavior [50,57,74], both hand trajectories [29,33,46] and interaction hotspots [39,41,42] predictions inherently possess a degree of uncertainty. To investigate the impact" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 190, + 168, + 425, + 264 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 2: Results of different joint forecasting methods. Individual refers to separately predict two subtasks. Hand indicates that the hand trajectory is the condition of the interaction hotspots. Object stands for correcting the trajectory using predicted contact points. Bidirectional combines the methods Hand and Object." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 190, + 168, + 425, + 264 + ], + "lines": [ + { + "bbox": [ + 190, + 168, + 425, + 264 + ], + "spans": [ + { + "bbox": [ + 190, + 168, + 425, + 264 + ], + "type": "table", + "html": "
MethodTrajectoryInteraction Hotspots
ADE ↓ FDE ↓SIM ↑ AUC-J ↑NSS ↑
Individual0.140.140.180.680.86
Hand0.130.120.250.721.19
Object0.120.120.230.701.04
Bidirectional0.100.090.280.731.48
Bi-Progressive0.090.080.320.771.70
", + "image_path": "cef00f82ab30ae2debec9d1d1af258fc7a9bc8b85920aae4df7e57c4a338679c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 279, + 299, + 358 + ], + "blocks": [ + { + "bbox": [ + 134, + 279, + 299, + 358 + ], + "lines": [ + { + "bbox": [ + 134, + 279, + 299, + 358 + ], + "spans": [ + { + "bbox": [ + 134, + 279, + 299, + 358 + ], + "type": "image", + "image_path": "1016746c7d59d5e7ce2f4e81cedb6def15212392e41bb8c8ed049b4ff5f675d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 367, + 303, + 433 + ], + "lines": [ + { + "bbox": [ + 130, + 367, + 303, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 303, + 433 + ], + "type": "text", + "content": "Fig. 7: Visualization results of different joint approaches. Based on the order from left to right and top to bottom, the corresponding methods are GT, Bi-Progressive, Bidirectional, Hand, Object, Individual." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 314, + 279, + 482, + 358 + ], + "blocks": [ + { + "bbox": [ + 314, + 279, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 314, + 279, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 279, + 482, + 358 + ], + "type": "image", + "image_path": "682996daf8b4fa713de5d94d31ad6e7eceaa69b6bebd6287d29bb334f1713a76.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 367, + 483, + 432 + ], + "lines": [ + { + "bbox": [ + 311, + 367, + 483, + 432 + ], + "spans": [ + { + "bbox": [ + 311, + 367, + 483, + 432 + ], + "type": "text", + "content": "Fig. 8: The role of Progressive Mechanism. The red refers to the ground truth, while the blue represents the predictions. From top to bottom, Bi-Pro stands for Bi-Progressive while Bid refers to bidirectional." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": "of uncertain region scale and the magnitude of uncertainty variations in hand trajectories, we modify " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " in Eq. 17, with the resultant outcomes depicted in Fig. 9. " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " denotes the overall scale of the uncertain region, when " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " equals 0, the region's size is zero, and the network output yields a unique result. As " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " increases, the uncertainty of trajectories grows, leading to improved accuracy due to random samplings. However, given the limitations in sampling times, as " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " continues to increase, the accuracy naturally decreases. " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " determines the magnitude of trajectory uncertainty increasing from front to back, when " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " equals 1, each trajectory point has the same level of uncertainty. Considering the escalating complexity and uncertainty in trajectory prediction over time, an elevation in " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " initially corresponds to augmented accuracy, followed by a subsequent decline. This gradual increase in uncertainty also renders the FDE more sensitive to " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": ", as shown in Fig. 10a-10c. Regarding interaction hotspots, BOT exhibits relatively small uncertainty (Fig. 10d), as for a specific intention, the interaction area becomes increasingly influenced by object constraints. This is achieved by tightly controlling the uncertainty within a reasonably small range through the process of enabling C-VAE to generate feature maps and subsequently select" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 120, + 302, + 225 + ], + "blocks": [ + { + "bbox": [ + 135, + 120, + 302, + 225 + ], + "lines": [ + { + "bbox": [ + 135, + 120, + 302, + 225 + ], + "spans": [ + { + "bbox": [ + 135, + 120, + 302, + 225 + ], + "type": "image", + "image_path": "9946323f2de0e43254fdf43aea0faba3f0fa204c520809e8b7a6145446835db4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 482, + 258 + ], + "type": "text", + "content": "Fig. 9: Uncertainty in hand trajectory forecasting. As " + }, + { + "bbox": [ + 131, + 236, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 236, + 482, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 236, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 236, + 482, + 258 + ], + "type": "text", + "content": " increase, they denote a heightened degree of overall trajectory uncertainty." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 311, + 120, + 478, + 227 + ], + "blocks": [ + { + "bbox": [ + 311, + 120, + 478, + 227 + ], + "lines": [ + { + "bbox": [ + 311, + 120, + 478, + 227 + ], + "spans": [ + { + "bbox": [ + 311, + 120, + 478, + 227 + ], + "type": "image", + "image_path": "760a15b4c9d602b1549d2a669ce781c93211af3d6ea2a380cc0722dacc529d3d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 268, + 190, + 308 + ], + "blocks": [ + { + "bbox": [ + 135, + 268, + 190, + 308 + ], + "lines": [ + { + "bbox": [ + 135, + 268, + 190, + 308 + ], + "spans": [ + { + "bbox": [ + 135, + 268, + 190, + 308 + ], + "type": "image", + "image_path": "9706ceef5dfae98bd7e824999cb90f44c6997cd5ba8063a80c858205de88cd42.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 191, + 268, + 246, + 308 + ], + "blocks": [ + { + "bbox": [ + 191, + 268, + 246, + 308 + ], + "lines": [ + { + "bbox": [ + 191, + 268, + 246, + 308 + ], + "spans": [ + { + "bbox": [ + 191, + 268, + 246, + 308 + ], + "type": "image", + "image_path": "c7ce3df40f11f04fbece25e9d1f9b701778d47733f7e6ae3cf407cc3ad53a556.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 247, + 268, + 302, + 308 + ], + "blocks": [ + { + "bbox": [ + 247, + 268, + 302, + 308 + ], + "lines": [ + { + "bbox": [ + 247, + 268, + 302, + 308 + ], + "spans": [ + { + "bbox": [ + 247, + 268, + 302, + 308 + ], + "type": "image", + "image_path": "95fa90005cc49811a5cf2b25a6fb9c77aaa6e1aaa6bf2a5d19bc9e0772f5ede3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 319, + 190, + 360 + ], + "blocks": [ + { + "bbox": [ + 163, + 310, + 274, + 319 + ], + "lines": [ + { + "bbox": [ + 163, + 310, + 274, + 319 + ], + "spans": [ + { + "bbox": [ + 163, + 310, + 274, + 319 + ], + "type": "text", + "content": "(a) Long-length trajectories" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 319, + 190, + 360 + ], + "lines": [ + { + "bbox": [ + 135, + 319, + 190, + 360 + ], + "spans": [ + { + "bbox": [ + 135, + 319, + 190, + 360 + ], + "type": "image", + "image_path": "d3c792f8019df11cf04d6e88d920257bea4cca5f8a3ff7053b2d5c61f7270f19.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 361, + 275, + 370 + ], + "lines": [ + { + "bbox": [ + 162, + 361, + 275, + 370 + ], + "spans": [ + { + "bbox": [ + 162, + 361, + 275, + 370 + ], + "type": "text", + "content": "(c) Short-length trajectories" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 379, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 131, + 379, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 482, + 413 + ], + "type": "text", + "content": "Fig. 10: Uncertainty in anticipation. (a), (b) and (c) respectively depict long-length trajectories, medium-length trajectories, and short-length trajectories, while (d) illustrates the uncertainty of interaction hotspots." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 191, + 319, + 246, + 360 + ], + "blocks": [ + { + "bbox": [ + 191, + 319, + 246, + 360 + ], + "lines": [ + { + "bbox": [ + 191, + 319, + 246, + 360 + ], + "spans": [ + { + "bbox": [ + 191, + 319, + 246, + 360 + ], + "type": "image", + "image_path": "0cfd00471519f5e232259737c6a0d580a05742e2aeca17f994ca306669647783.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 247, + 319, + 302, + 359 + ], + "blocks": [ + { + "bbox": [ + 247, + 319, + 302, + 359 + ], + "lines": [ + { + "bbox": [ + 247, + 319, + 302, + 359 + ], + "spans": [ + { + "bbox": [ + 247, + 319, + 302, + 359 + ], + "type": "image", + "image_path": "c9a6ec9d8581cab8aac50879e50fcf55f95a0711048cc5419fb82077b82e2974.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 312, + 268, + 367, + 308 + ], + "blocks": [ + { + "bbox": [ + 312, + 268, + 367, + 308 + ], + "lines": [ + { + "bbox": [ + 312, + 268, + 367, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 268, + 367, + 308 + ], + "type": "image", + "image_path": "746bc5e6a08a005ffa5b9757a001cbe707d5f13573df33442dd5c655a059fe5e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 368, + 268, + 423, + 308 + ], + "blocks": [ + { + "bbox": [ + 368, + 268, + 423, + 308 + ], + "lines": [ + { + "bbox": [ + 368, + 268, + 423, + 308 + ], + "spans": [ + { + "bbox": [ + 368, + 268, + 423, + 308 + ], + "type": "image", + "image_path": "a9258eca08991ea1b76caf076842b07def6f5b0da924eaf3859f0d06a104f293.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 424, + 268, + 479, + 308 + ], + "blocks": [ + { + "bbox": [ + 424, + 268, + 479, + 308 + ], + "lines": [ + { + "bbox": [ + 424, + 268, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 424, + 268, + 479, + 308 + ], + "type": "image", + "image_path": "a0f48ce9da74069ef86b2b4958fb6ef2119d64d984833938fb29d08a44654c81.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 312, + 319, + 368, + 360 + ], + "blocks": [ + { + "bbox": [ + 332, + 310, + 457, + 319 + ], + "lines": [ + { + "bbox": [ + 332, + 310, + 457, + 319 + ], + "spans": [ + { + "bbox": [ + 332, + 310, + 457, + 319 + ], + "type": "text", + "content": "(b) Medium-length trajectories" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 319, + 368, + 360 + ], + "lines": [ + { + "bbox": [ + 312, + 319, + 368, + 360 + ], + "spans": [ + { + "bbox": [ + 312, + 319, + 368, + 360 + ], + "type": "image", + "image_path": "1a2e7ba433910f827577e8f1c603fc9ee02180df9b39a6583be09a7034394358.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 361, + 443, + 370 + ], + "lines": [ + { + "bbox": [ + 347, + 361, + 443, + 370 + ], + "spans": [ + { + "bbox": [ + 347, + 361, + 443, + 370 + ], + "type": "text", + "content": "(d) interaction hotspots" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 368, + 319, + 423, + 360 + ], + "blocks": [ + { + "bbox": [ + 368, + 319, + 423, + 360 + ], + "lines": [ + { + "bbox": [ + 368, + 319, + 423, + 360 + ], + "spans": [ + { + "bbox": [ + 368, + 319, + 423, + 360 + ], + "type": "image", + "image_path": "fbfe206760380f513fb903ce4148aeb988429e5e1d380b89bf3dcee908babc4d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 424, + 319, + 479, + 360 + ], + "blocks": [ + { + "bbox": [ + 424, + 319, + 479, + 360 + ], + "lines": [ + { + "bbox": [ + 424, + 319, + 479, + 360 + ], + "spans": [ + { + "bbox": [ + 424, + 319, + 479, + 360 + ], + "type": "image", + "image_path": "2e82da4de91af397b39ffff4f8c750e05a0d85ed2af730938172ef5d9ed0b8e5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 434, + 481, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 481, + 458 + ], + "type": "text", + "content": "points, thereby ensuring the network's high confidence in predicting interaction regions." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 475, + 217, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 217, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 217, + 487 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": "Conclusion. We plan to anticipate interaction intentions, encompassing the joint prediction of future hand trajectories and interaction hotspots with objects. We propose a novel BOT framework to establish an inherent connection between hand trajectories and contact points by introducing a Bi-Progressive mechanism, which eliminates the effect of error accumulation and achieves more accurate intention anticipation. Experimental results on three challenging benchmarks achieve state-of-the-art performance, thereby proving the superiority of our method in intention anticipation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Future Work. BOT currently predicts hand trajectories and interaction hotspots on 2D images. However, human-object interactions occur in 3D space. To enable the model to be applied to real-world scenarios, we will consider the geometric structure and spatial relationships between objects in the scene in the future, providing the model with the ability to anticipate interaction intentions for the real world." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 166 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62306295 and 62033012." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 197, + 198, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 198, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 198, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 223, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 223, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 481, + 255 + ], + "type": "text", + "content": "1. Adeli, V., Adeli, E., Reid, I., Niebles, J.C., Rezatofighi, H.: Socially and contextually aware human motion and pose forecasting. IEEE Robotics and Automation Letters 5(4), 6033-6040 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 300 + ], + "type": "text", + "content": "2. Adeli, V., Ehsanpour, M., Reid, I., Niebles, J.C., Savarese, S., Adeli, E., Rezatofighi, H.: Tripod: Human trajectory and pose dynamics forecasting in the wild. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13390-13400 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "type": "text", + "content": "3. Bahl, S., Gupta, A., Pathak, D.: Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "text", + "content": "4. Bao, W., Chen, L., Zeng, L., Li, Z., Xu, Y., Yuan, J., Kong, Y.: Uncertainty-aware state space transformer for egocentric 3d hand trajectory forecasting. arXiv preprint arXiv:2307.08243 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 356, + 481, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 481, + 378 + ], + "type": "text", + "content": "5. Bertasius, G., Park, H.S., Yu, S.X., Shi, J.: First person action-object detection with egonet. arXiv preprint arXiv:1603.04908 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 379, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 481, + 411 + ], + "type": "text", + "content": "6. Bertasius, G., Soo Park, H., Yu, S.X., Shi, J.: Unsupervised learning of important objects from first-person videos. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1956-1964 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 412, + 481, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 481, + 433 + ], + "type": "text", + "content": "7. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "text", + "content": "8. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE international conference on image processing (ICIP). pp. 3464-3468. IEEE (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 467, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 488 + ], + "type": "text", + "content": "9. Bharadhwaj, H., Gupta, A., Tulsiani, S., Kumar, V.: Zero-shot robot manipulation from passive human videos. arXiv preprint arXiv:2302.02011 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 489, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 481, + 521 + ], + "type": "text", + "content": "0. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "type": "text", + "content": "1. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "2. Chen, G., Chen, Z., Fan, S., Zhang, K.: Unsupervised sampling promoting for stochastic human trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17874-17884 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 588, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 620 + ], + "type": "text", + "content": "3. Choi, C., Dariush, B.: Looking to relations for future trajectory forecast. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "4. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Scaling egocentric vision: The epic-kitchens dataset. In: European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "15. Dessalene, E., Devaraj, C., Maynard, M., Fermuller, C., Aloimonos, Y.: Forecasting action through contact representations from first person video. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 194 + ], + "type": "text", + "content": "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "17. Fan, C., Lee, J., Ryoo, M.S.: Forecasting hands and objects in future frames. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "18. Fang, K., Wu, T.L., Yang, D., Savarese, S., Lim, J.J.: Demo2vec: Reasoning object affordances from online videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2139-2147 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "text", + "content": "19. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "20. Furnari, A., Battiato, S., Grauman, K., Farinella, G.M.: Next-active-object prediction from egocentric videos. Journal of Visual Communication and Image Representation 49, 401-411 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 331, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 373 + ], + "type": "text", + "content": "21. Furnari, A., Battiato, S., Maria Farinella, G.: Leveraging uncertainty to rethink loss functions and evaluation measures for egocentric action anticipation. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "22. Furnari, A., Farinella, G.M.: What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6252-6261 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "23. Furnari, A., Farinella, G.M.: Rolling-unrolling lstms for action anticipation from first-person video. IEEE transactions on pattern analysis and machine intelligence 43(11), 4021-4036 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "text", + "content": "24. Geng, Y., An, B., Geng, H., Chen, Y., Yang, Y., Dong, H.: End-to-end affordance learning for robotic manipulation. arXiv preprint arXiv:2209.12941 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 465, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 481, + 486 + ], + "type": "text", + "content": "25. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 481, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 519 + ], + "type": "text", + "content": "26. Gong, D., Lee, J., Kim, M., Ha, S.J., Cho, M.: Future transformer for long-term action anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3052-3061 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "27. Gong, J., Foo, L.G., Fan, Z., Ke, Q., Rahmani, H., Liu, J.: Diffpose: Toward more reliable 3d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13041-13051 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 597 + ], + "type": "text", + "content": "28. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "29. Gupta, A., Johnson, J., Fei-Fei, L., Savarese, S., Alahi, A.: Social gan: Socially acceptable trajectories with generative adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2255-2264 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "30. Jiang, H., Liu, S., Wang, J., Wang, X.: Hand-object contact consistency reasoning for human grasps generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11107-11116 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 117, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 482, + 149 + ], + "type": "text", + "content": "31. Judd, T., Ehinger, K., Durand, F., Torralba, A.: Learning to predict where humans look. In: 2009 IEEE 12th international conference on computer vision. pp. 2106-2113. IEEE (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "32. Kjellström, H., Romero, J., Kragic, D.: Visual object-action recognition: Inferring object affordances from human demonstration. Computer Vision and Image Understanding 115(1), 81-90 (2011)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "33. Lee, N., Choi, W., Vernaza, P., Choy, C.B., Torr, P.H., Chandraker, M.: Desire: Distant future prediction in dynamic scenes with interacting agents. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 336-345 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "34. Li, C., Zheng, P., Yin, Y., Pang, Y.M., Huo, S.: An ar-assisted deep reinforcement learning-based approach towards mutual-cognitive safe human-robot interaction. Robotics and Computer-Integrated Manufacturing 80, 102471 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "35. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Wang, L., Qiao, Y.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "36. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "37. Liu, M., Tang, S., Li, Y., Rehg, J.M.: Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 704-721. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "text", + "content": "38. Liu, S., Jiang, H., Xu, J., Liu, S., Wang, X.: Semi-supervised 3d hand-object poses estimation with interactions in time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14687-14697 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "39. Liu, S., Tripathi, S., Majumdar, S., Wang, X.: Joint hand motion and interaction hotspots prediction from egocentric videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3282-3292 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "40. Liu, T., Lam, K.M.: A hybrid egocentric activity anticipation framework via memory-augmented recurrent and one-shot representation forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13904-13913 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "41. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Learning affordance grounding from exocentric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2252-2261 (June 2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "42. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Learning visual affordance grounding from demonstration videos. IEEE Transactions on Neural Networks and Learning Systems (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "43. Luo, H., Zhai, W., Zhang, J., Cao, Y., Tao, D.: Leverage interactive affinity for affordance learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6809-6819 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "44. Mandlekar, A., Xu, D., Martín-Martín, R., Savarese, S., Fei-Fei, L.: Learning to generalize across long-horizon tasks from human demonstrations. arXiv preprint arXiv:2003.06085 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "45. Mangalam, K., An, Y., Girase, H., Malik, J.: From goals, waypoints & paths to long term human trajectory forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15233-15242 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "46. Mangalam, K., Girase, H., Agarwal, S., Lee, K.H., Adeli, E., Malik, J., Gaidon, A.: It is not the journey but the destination: Endpoint conditioned trajectory" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "text", + "content": "prediction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 759-776. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 173 + ], + "type": "text", + "content": "47. Mascaró, E.V., Ahn, H., Lee, D.: Intention-conditioned long-term human egocentric action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6048-6057 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 481, + 195 + ], + "type": "text", + "content": "48. Mendonca, R., Bahl, S., Pathak, D.: Structured world models from human videos. arXiv preprint arXiv:2308.10901 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 197, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 197, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 134, + 197, + 481, + 228 + ], + "type": "text", + "content": "49. Miech, A., Laptev, I., Sivic, J., Wang, H., Torresani, L., Tran, D.: Leveraging the present to anticipate the future in videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 230, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 230, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 133, + 230, + 481, + 274 + ], + "type": "text", + "content": "50. Mohamed, A., Qian, K., Elhoseiny, M., Claudel, C.: Social-stgcnn: A social spatiotemporal graph convolutional neural network for human trajectory prediction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14424-14432 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "type": "text", + "content": "51. Myers, A., Teo, C.L., Fermüller, C., Aloimonos, Y.: Affordance detection of tool parts from geometric features. In: 2015 IEEE International Conference on Robotics and Automation (ICRA). pp. 1374-1381. IEEE (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "52. Nagarajan, T., Feichtenhofer, C., Grauman, K.: Grounded human-object interaction hotspots from video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8688-8697 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 373 + ], + "type": "text", + "content": "53. Nawhal, M., Jyothi, A.A., Mori, G.: Rethinking learning approaches for long-term action anticipation. In: European Conference on Computer Vision. pp. 558-576. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "54. Parsaeifard, B., Saadatnejad, S., Liu, Y., Mordan, T., Alahi, A.: Learning decoupled representations for human pose forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2294-2303 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 409, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 430 + ], + "type": "text", + "content": "55. Peters, R.J., Iyer, A., Itti, L., Koch, C.: Components of bottom-up gaze allocation in natural images. Vision research 45(18), 2397-2416 (2005)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "text", + "content": "56. Plizzari, C., Goletto, G., Furnari, A., Bansal, S., Ragusa, F., Farinella, G.M., Damen, D., Tommasi, T.: An outlook into the future of egocentric vision. arXiv preprint arXiv:2308.07123 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "type": "text", + "content": "57. Qi, H., Wang, X., Pathak, D., Ma, Y., Malik, J.: Learning long-term visual dynamics with region proposal interaction networks. arXiv preprint arXiv:2008.02265 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "text", + "content": "58. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3636-3645 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "59. Rodin, I., Furnari, A., Mavroeidis, D., Farinella, G.M.: Predicting the future from first person (egocentric) vision: A survey. Computer Vision and Image Understanding 211, 103252 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 597 + ], + "type": "text", + "content": "60. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "61. Shen, J., Dudley, J., Kristensson, P.O.: Fast and robust mid-air gesture typing for ar headsets using 3d trajectory decoding. IEEE Transactions on Visualization and Computer Graphics (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "62. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 556 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "type": "text", + "content": "63. Song, D., Kyriazis, N., Oikonomidis, I., Papazov, C., Argyros, A., Burschka, D., Kragic, D.: Predicting human intention in visual observations of hand/object interactions. In: 2013 IEEE International Conference on Robotics and Automation. pp. 1608-1615. IEEE (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "text", + "content": "64. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "65. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "66. Swain, M.J., Ballard, D.H.: Color indexing. International journal of computer vision 7(1), 11-32 (1991)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "67. Tu, Z., Huang, Z., Chen, Y., Kang, D., Bao, L., Yang, B., Yuan, J.: Consistent 3d hand reconstruction in video via self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "68. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "text", + "content": "69. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "70. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "71. Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8121-8130 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "72. Yang, Y., Zhai, W., Luo, H., Cao, Y., Luo, J., Zha, Z.J.: Grounding 3d object affordance from 2d interactions in images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10905-10915 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "73. Yang, Y., Zhai, W., Luo, H., Cao, Y., Zha, Z.J.: Lemon: Learning 3d human-object interaction relation from 2d images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16284-16295 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "74. Ye, Y., Singh, M., Gupta, A., Tulsiani, S.: Compositional video prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10353-10362 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 500, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 481, + 533 + ], + "type": "text", + "content": "75. Zhai, W., Cao, Y., Zhang, J., Xie, H., Tao, D., Zha, Z.J.: On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 556 + ], + "type": "text", + "content": "76. Zhai, W., Luo, H., Zhang, J., Cao, Y., Tao, D.: One-shot object affordance detection in the wild. International Journal of Computer Vision 130(10), 2472-2500 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bidirectional Progressive Transformer for Interaction Intention Anticipation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/e50db183-7c10-4237-a52b-c78459d9ff33_content_list.json b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/e50db183-7c10-4237-a52b-c78459d9ff33_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f1e66bfb266a476f6411f5a5d513ff2dfef4958b --- /dev/null +++ b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/e50db183-7c10-4237-a52b-c78459d9ff33_content_list.json @@ -0,0 +1,1673 @@ +[ + { + "type": "text", + "text": "Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhening Liu, Xinjie Zhang, Jiawei Shao, Zehong Lin, and Jun Zhang", + "bbox": [ + 336, + 212, + 661, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong SAR {zhening.liu,xzhangga,jiawei.shao}@connect.ust.hk, {eezhlin,eejzhang}@ust.hk", + "bbox": [ + 217, + 255, + 784, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. With the rapid advancement of stereo vision technologies, stereo image compression has emerged as a crucial field that continues to draw significant attention. Previous approaches have primarily employed a unidirectional paradigm, where the compression of one view is dependent on the other, resulting in imbalanced compression. To address this issue, we introduce a symmetric bidirectional stereo image compression architecture, named BiSIC. Specifically, we propose a 3D convolution based codec backbone to capture local features and incorporate bidirectional attention blocks to exploit global features. Moreover, we design a novel cross-dimensional entropy model that integrates various conditioning factors, including the spatial context, channel context, and stereo dependency, to effectively estimate the distribution of latent representations for entropy coding. Extensive experiments demonstrate that our proposed BiSIC outperforms conventional image/video compression standards, as well as state-of-the-art learning-based methods, in terms of both PSNR and MS-SSIM.", + "bbox": [ + 261, + 338, + 743, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Stereo Image Compression $\\cdot$ Bidirectional Architecture $\\cdot$ Cross-Dimensional Entropy Model", + "bbox": [ + 259, + 574, + 740, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 628, + 375, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stereo vision, which mimics human binocular vision, has emerged as an important area in computer vision. It enables a wide range of applications, such as 3D movies, remote sensing [29], autonomous driving [38], and augmented/virtual reality (AR/VR) [1]. With the rapid development of stereo cameras, stereo images have become essential for providing immersive and convincing visual presentations that enhance user experience and benefit downstream applications. Meanwhile, the proliferation of stereo cameras leads to a significant surge in the volume of stereo images, resulting in substantial storage costs and posing challenges to transmission efficiency. Consequently, stereo image compression plays a pivotal role in ensuring the practicality and effectiveness of stereo vision.", + "bbox": [ + 212, + 662, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, numerous methods have been developed for single image compression, including traditional CODECs [6, 34] and learning-based methods [2-4, 10, 24]. These methods, however, focus on extracting individual features for compression and fail to capture the cross-view correlation. Consequently, they are less effective for stereo image compression that requires the extraction of both intra-view information and inter-view information. To exploit the inherent shared information in stereo images, traditional methods for stereo image coding follow a predictive compression procedure of video standards [9, 31]. In this process, one view is compressed first and then used to predict the other view. Subsequently, the disparity between the predicted and actual views is compressed. Nonetheless, these methods typically rely on hand-crafted modules and lack end-to-end optimization, resulting in suboptimal performance.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the success of learning-based single image compression, researchers have proposed novel stereo image compression methods using deep neural networks (DNNs) to enhance compression performance [12, 13, 20, 21, 36, 37, 40, 41]. Initially, extensive research focuses on unidirectional learned stereo image compression methods [12, 13, 21, 36, 37, 40], where one view is compressed and used as a reference to compress the other view. Despite achieving high overall rate-distortion (RD) performance, these unidirectional learned compression methods yield a noticeable imbalance in compression quality between stereo views, similar to traditional predictive compression methods. This imbalance is unfavorable for human vision and undermines downstream machine vision tasks [23]. Moreover, some of these methods [12, 13, 37, 40] heavily rely on sequential coding order and accurate estimation of the explicit warping relationship between views. To avoid these issues, a recent study [20] proposes a bidirectional compression method that leverages 2D convolution to extract individual features and facilitate information sharing between views. Nonetheless, the compression performance is limited due to the separate processing of convolutions and the plain spatial context based entropy model.", + "bbox": [ + 212, + 342, + 789, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above discussion, we propose a novel bidirectional method for stereo image compression, named BiSIC, to address the aforementioned issues. BiSIC achieves improvements in four key aspects: encoder/decoder structure, mutual stereo learning, entropy model, and coding speed. Firstly, regarding the codec structure, although 2D convolutions have proven effective in various computer vision tasks, they fail to capture the aligned morphs and features between stereo views. To overcome this limitation, we propose using 3D convolution as the backbone for the encoder/decoder, which takes both views as input to capture the inter-view dependencies that are inaccessible in previous works based on 2D convolution. Secondly, for mutual stereo learning, we propose a mutual attention block to facilitate the transfer of features between views. This module enhances the integration and cooperation of the stereo views, leading to improved compression performance. Thirdly, for the entropy model, prior methods primarily use hyperprior and spatial context as conditions, overlooking the rich inter-view dependencies across other dimensions. To exploit these dependencies,", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/82a897adf9318952b36b7d0d6d6f253c71eab74f67470215fcd560c7c0138048.jpg", + "image_caption": [ + "Fig. 1: The proposed bidirectional stereo image compression architecture. AE and AD are arithmetic coder for entropy coding. The backbones of encoder $\\mathbf{E}$ , decoder $\\mathbf{D}$ , hyper encoder $h_a$ , and hyper decoder $h_s$ are constructed with 3D convolution to model local features. 3DCovTr denotes transposed 3D convolution. To enhance global feature extraction, bidirectional mutual attention blocks are inserted between 3D convolutional layers. Moreover, a novel cross-dimensional entropy model is utilized to capture complex inter-view dependencies." + ], + "image_footnote": [], + "bbox": [ + 274, + 142, + 728, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we construct a cross-dimensional entropy model that aggregates hyperprior, spatial context, channel context, and stereo dependency.", + "bbox": [ + 212, + 433, + 782, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we introduce a masked 3D convolution to capture the stereo spatial context and utilize an attention block to process the stereo channel context. Finally, to improve coding speed and reduce auto-regressive steps, we design a fast variant based on a stereo-checkerboard structure. This design converts auto-regressive inference into a two-fold operation, improving the coding efficiency. The contributions of this paper are summarized as follows:", + "bbox": [ + 212, + 463, + 784, + 553 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel 3D convolution based codec backbone that processes stereo images as a whole, enabling the learning of local stereo features and effective capture of shared patterns across views.", + "- We design a bidirectional mutual attention block that facilitates the exchange of global features between stereo views. This block produces compact attention maps and is robust to input images of different sizes.", + "- We develop a powerful symmetric cross-dimensional entropy model, which incorporates hyperpriors, spatial context, channel context, and stereo dependency as conditioning factors, to enhance the probability estimation.", + "- To accelerate coding, we propose a stereo-checkerboard design that transforms the spatial auto-regressive inference into a two-fold process. This design strikes a balance between time consumption and compression quality." + ], + "bbox": [ + 225, + 558, + 784, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 761, + 395, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Single Image Compression", + "text_level": 1, + "bbox": [ + 215, + 787, + 478, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image compression is an important topic that continuously attracts significant research interest. With the increasing demand for vision products and applica", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BiSIC: Bidirectional Stereo Image Compression", + "bbox": [ + 415, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions, steady progress has been made in traditional image compression standards, evolving from JPEG [28, 34] and BPG [6] to video compression standards such as HEVC [30] and VVC [9]. These traditional methods establish a standard compression pipeline comprising transform, quantization, and entropy coding. Over the past five years, deep learning has revolutionized various vision applications. As demonstrated in the seminal work [3], DNNs have the potential to replace components in traditional compression methods and enable end-to-end compression. This work introduces uniform noise to mimic the distortion caused by quantization, thereby paving the way for end-to-end optimization in image compression.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In learned image compression models [3,4,10,14-17,22,24-26,42], neural networks are employed to parameterize the probability distribution of transformed latents. Accurate estimation of this distribution is crucial for achieving compact compression via entropy coding. To enhance probability estimation, various conditions have been integrated into entropy models [2], including hyperpriors [4], spatial context [24], and channel-wise context [25], to capture a broader range of dependencies. To be specific, hyperprior is a lightweight subsidiarity information transmitted alongside the encoded latents. Both spatial context and channel-wise context leverage auto-regressive inference, which is conditioned on the preceding segments of data by iteratively interleaving the inference. Checkerboard methods [15], on the other hand, avoid the spatial auto-regressive process and accelerate the coding speed by constraining the spatial context to the near neighbors of each pixel instead of spatial causal components. These dependencies operate on diverse scales and can be progressively combined to further improve effectiveness, as demonstrated in [14, 17].", + "bbox": [ + 212, + 297, + 787, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Stereo Image Compression", + "text_level": 1, + "bbox": [ + 214, + 545, + 482, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stereo image compression is an extension of single image compression that considers capturing both intra-view features and inter-view correlations. Traditional approaches employ key-frame based predictive coding, akin to video coding [18, 33]. However, these methods rely on hand-crafted feature extractors and classical optimization techniques. As these modules are optimized individually, they often fail to achieve satisfactory RD performance. Moreover, the predictive compression pipeline also leads to imbalanced compression.", + "bbox": [ + 212, + 568, + 787, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In recent years, learned stereo image compression methods have gained increasing attention [12, 13, 20, 21, 36, 37, 40, 41]. These methods can be broadly categorized into unidirectional and bidirectional approaches. Unidirectional methods first compress one view image and then use it as a reference to compress the other view, significantly reducing the required bits. The pioneering work [21] introduces a skip module to warp the features from one view to encode the other. Subsequent works have utilized various techniques to represent the warping relationship between stereo views, such as homography matrices [12, 13], disparity maps [40], and horizontal shifts [37].", + "bbox": [ + 212, + 674, + 787, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These unidirectional methods heavily rely on the precise estimation of the spatial transformations. A very recent work, ECSIC [36], avoids explicit para", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "metric transformation in favor of constructing a unidirectional entropy model that conditions the right image on the known left image. However, the decompressed results of these unidirectional methods often have imbalanced quality due to the one-way dependency between stereo views. On the other hand, bidirectional methods, while remaining less explored, produce more balanced quality by extracting shared information from both views and compressing stereo images simultaneously. One such approach [20] introduces a contextual transform module for feature transfer between views. However, it employs separate down-sampling networks and only takes spatial context into account, which limits the compression performance. In contrast, to fully exploit the shared information and establish stronger dependencies, we employ a joint downsampling codec and a cross-dimensional entropy model.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Proposed Methods", + "text_level": 1, + "bbox": [ + 215, + 349, + 434, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed architecture for stereo image compression, as depicted in Fig. 1, consists of two main components: a joint codec and an entropy model. The codec employs 3D convolutions to transform the original stereo images into compact latent representations, effectively exploiting the correlation between the two views. Subsequently, the entropy model estimates the probability distribution of these latent representations, enabling the generation of a compact bitstream through entropy coding. In the following subsections, we elaborate on the codec and the entropy model.", + "bbox": [ + 212, + 381, + 787, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Joint Codec", + "text_level": 1, + "bbox": [ + 215, + 523, + 361, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Convolution Backbone. The codec acts as a non-linear transformation that downsamples the raw images into compact latent representations and generates hyperpriors for the entropy model. Different from previous approaches to stereo image compression that employ 2D convolutional layers, we adopt 3D convolutions as the backbone of our codec. Unlike 2D convolution, which operates separately on each view, 3D convolution concurrently processes both stereo images. Consequently, it inherently possesses the capability to extract inter-view correlations.", + "bbox": [ + 212, + 547, + 787, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our encoder, denoted by $\\mathbf{E}$ , comprises four 3D convolutional layers, with bidirectional mutual attention blocks inserted after the second and the fourth layers to facilitate global feature communication between the two views. The encoder transforms a pair of stereo images, i.e., the left and right images denoted by $\\pmb{x}_l$ , $\\pmb{x}_r \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}$ , into compact latent representations $\\pmb{y}_l$ , $\\pmb{y}_r \\in \\mathbb{R}^{B \\times N \\times \\frac{H}{16} \\times \\frac{W}{16}}$ and quantizes them into $\\hat{\\pmb{y}} := \\{\\hat{\\pmb{y}}_l, \\hat{\\pmb{y}}_r\\}$ for efficient transmission. Conversely, the decoder $\\mathbf{D}$ transforms the quantized latents $\\hat{\\pmb{y}}$ back into images $\\hat{\\pmb{x}}_l$ and $\\hat{\\pmb{x}}_r$ . The encoding and decoding processes are expressed as:", + "bbox": [ + 212, + 669, + 787, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {l}, \\boldsymbol {y} _ {r} = \\mathbf {E} \\left(\\boldsymbol {x} _ {l}, \\boldsymbol {x} _ {r}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 801, + 784, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {l}, \\hat {\\boldsymbol {x}} _ {r} = \\mathbf {D} (\\hat {\\boldsymbol {y}} _ {l}, \\hat {\\boldsymbol {y}} _ {r}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 820, + 784, + 835 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BiSIC: Bidirectional Stereo Image Compression", + "bbox": [ + 415, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To capture the dependencies among elements in the latents $\\hat{\\pmb{y}}$ for efficient entropy coding [4,7], hyperpriors are generated through another 3D convolution based hyper encoder $h_a$ . The generated hyperpriors $\\pmb{z}_l, \\pmb{z}_r \\in \\mathbb{R}^{B \\times M \\times \\frac{H}{64} \\times \\frac{W}{64}}$ are quantized into $\\hat{\\pmb{z}}_l, \\hat{\\pmb{z}}_r$ and then transmitted alongside the quantized latents $\\hat{\\pmb{y}}$ . At the decoder side, a 3D convolution based hyper decoder $h_s$ upscales the quantized hyperpriors $\\hat{\\pmb{z}}_l, \\hat{\\pmb{z}}_r$ into $\\tilde{\\pmb{z}}_l, \\tilde{\\pmb{z}}_r$ , which are subsequently utilized in the entropy model, as elaborated in Sec. 3.2. The processes of the hyper encoder and hyper decoder are expressed as:", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {l}, \\boldsymbol {z} _ {r} = h _ {a} (\\boldsymbol {y} _ {l}, \\boldsymbol {y} _ {r}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 277, + 784, + 292 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {z}} _ {l}, \\tilde {\\boldsymbol {z}} _ {r} = h _ {s} (\\hat {\\boldsymbol {z}} _ {l}, \\hat {\\boldsymbol {z}} _ {r}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 296, + 784, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bidirectional Mutual Attention Block. Convolutional layers have an excellent ability for local modeling, but their ability for long-range feature modeling is relatively limited. On the other hand, attention modules [32] excel at extracting long-range features. Therefore, we propose combining the attention mechanism with convolutional layers to compensate for the shortcomings of the latter and leverage the advantages of both convolution and attention. To this end, we design a novel bidirectional mutual attention block, as depicted in Fig. 2, to be inserted between convolutional layers as the pink blocks shown in Fig. 1.", + "bbox": [ + 212, + 316, + 782, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, the bidirectional mutual attention block contains two cross-attention stages, each followed by a self-attention layer. The two cross-attention stages differ in how they concentrate on the feature map and are referred to as cross-key attention and cross-query attention. Before each cross-attention stage, the input stereo data $\\pmb{y}_l, \\pmb{y}_r$ first pass through a residual block, which are then embedded into low-dimensional query, key, and value tensors $\\pmb{Q}, \\pmb{K}, \\pmb{V} \\in \\mathbb{R}^{B \\times C \\times (H \\times W)}$ . To avoid the prohibitive computation of classical attention, which yields an $(H \\times W) \\times (H \\times W)$ sized attention map, we adopt the efficient attention approach proposed in [27]. This approach computes the product of the key and the transposed value, thereby reducing the attention map size to $C_K \\times C_V$ .", + "bbox": [ + 212, + 439, + 784, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the cross-key attention stage, for each view, we generate an attention map using its key and the value from the other view, which is then queried by the current view. By doing so, we can extract the cross-key feature $\\Phi$ from the stereo views as:", + "bbox": [ + 212, + 604, + 784, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Phi} _ {r \\rightarrow l} = (\\sigma (\\boldsymbol {K} _ {r}) \\times \\boldsymbol {V} _ {l} ^ {T}) ^ {T} \\times \\sigma (\\boldsymbol {Q} _ {r}), \\quad \\boldsymbol {\\Phi} _ {l \\rightarrow r} = (\\sigma (\\boldsymbol {K} _ {l}) \\times \\boldsymbol {V} _ {r} ^ {T}) ^ {T} \\times \\sigma (\\boldsymbol {Q} _ {l}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 672, + 784, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma$ is the softmax function. This approach calculates the attention map using the key and value from different views, thereby facilitating inter-view alignment and the identification of common patterns. When a feature in one view is correlated with the corresponding feature in the other view, it is effectively captured through the attention map.", + "bbox": [ + 212, + 696, + 782, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the cross-query attention stage, the attention map is obtained using the key and value from a single view, and then queried by the other view, yielding the cross-query feature $\\Psi$ as:", + "bbox": [ + 212, + 772, + 787, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {r \\rightarrow l} = \\left(\\sigma \\left(\\boldsymbol {K} _ {l}\\right) \\times \\boldsymbol {V} _ {l} ^ {T}\\right) ^ {T} \\times \\sigma \\left(\\boldsymbol {Q} _ {r}\\right), \\Psi_ {l \\rightarrow r} = \\left(\\sigma \\left(\\boldsymbol {K} _ {r}\\right) \\times \\boldsymbol {V} _ {r} ^ {T}\\right) ^ {T} \\times \\sigma \\left(\\boldsymbol {Q} _ {l}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 823, + 784, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b30f87dee6598cab8a1b245a633985d65c548e00d77420fae0ed4730f3075ace.jpg", + "image_caption": [ + "Fig. 2: Overview of the proposed bidirectional mutual attention block. The blue and green lines represent features extracted from the left view and right view, respectively. The network structures of the residual block, the basic efficient attention unit, and the combine block are illustrated on the right, from top to bottom." + ], + "image_footnote": [], + "bbox": [ + 253, + 143, + 750, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The attention map in this stage is generated within one view, thus maintaining an inner-focused relationship. This process preserves the individual self-features of each view, while the other view serves as a reference for determining the allocation of attention weights.", + "bbox": [ + 212, + 388, + 784, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After obtaining the cross-key attention and cross-query attention features $\\Phi, \\Psi$ , we employ self-attention to augment these features as $\\tilde{\\Phi}, \\tilde{\\Psi}$ and use a shared-parameter combine block to get the desired mutual attention output, as shown in the bottom right of Fig. 2. Specifically, the combine block concatenates the augmented features $\\tilde{\\Phi}, \\tilde{\\Psi}$ and the residual component, and shrinks the channel number to be consistent with the block input.", + "bbox": [ + 212, + 449, + 787, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This bidirectional mutual attention block captures diverse features and provides a global receptive field for the codec. It is suitable for any stereo information transfer scenarios, provided that the inputs are two features of equal size, making it a generalized plug-and-play block. Moreover, the size of the attention maps is determined only by the embedding channel $C$ and does not vary with the size of input images.", + "bbox": [ + 212, + 541, + 787, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Cross-Dimensional Entropy Model", + "text_level": 1, + "bbox": [ + 214, + 659, + 545, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The proposed codec effectively transforms stereo images into small-sized latent representations. To achieve efficient compression, the next crucial step is to encode these latents into compact bitstreams using accurate probability distribution and entropy coding [4,24]. Thus, an entropy model is critical for parameterizing and estimating probabilities. In learned image compression, the probability distribution is commonly modeled as Gaussian distributions [24], with the mean $\\mu$ and variance $\\sigma^2$ being estimated. To achieve accurate estimation, it is critical to provide appropriate references. Apart from hyperpriors, previous stereo image compression methods primarily use spatial context [13,20,41] and unidirectional stereo dependency as references [21,36], while abundant features on the", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BiSIC: Bidirectional Stereo Image Compression", + "bbox": [ + 415, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe612b9e57ce8faa802be1658b78292465b36d3069593a0608d31bef63cde874.jpg", + "image_caption": [ + "Fig. 3: Illustration of the proposed symmetric cross-dimensional entropy model. It jointly aggregates the hyperprior (blue), stereo spatial context (red), and stereo channel context (yellow) as conditions for an effective probability distribution estimation. The hyper decoder $h_s$ for hyperprior is shown in Fig. 1. The masked 3D convolution for spatial context is provided in Fig. 4a. Moreover, the mutual attention block for channel context is detailed in Fig. 2." + ], + "image_footnote": [], + "bbox": [ + 250, + 142, + 754, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "channel dimension are overlooked. In this work, we propose a symmetric cross-dimensional entropy model that integrates hyperprior, spatial context, channel context, and stereo dependency in a fully symmetric bidirectional way.", + "bbox": [ + 212, + 419, + 782, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The architecture of the proposed cross-dimensional entropy model is depicted in Fig. 3. We use $\\hat{\\pmb{y}}_{l,i}^{k}$ to denote the $i$ -th $(i = 1,2,\\dots,n)$ entry of the $k$ -th $(k = 1,2,\\dots,K)$ channel slice from the received left view latent $\\hat{\\pmb{y}}_l$ , and likewise for the others. The cross-dimensional condition formulation is expressed as:", + "bbox": [ + 212, + 465, + 787, + 525 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\hat {\\boldsymbol {y}} _ {l}} \\left(\\hat {\\boldsymbol {y}} _ {l, i} ^ {k} \\mid \\hat {\\boldsymbol {z}} _ {l}, \\hat {\\boldsymbol {y}} _ {l} ^ {< k}, \\hat {\\boldsymbol {y}} _ {l, < i} ^ {k}, \\hat {\\boldsymbol {y}} _ {r} ^ {< k}, \\hat {\\boldsymbol {y}} _ {r, < i} ^ {k}\\right) \\sim \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {l}, \\boldsymbol {\\sigma} _ {l} ^ {2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 532, + 784, + 553 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\hat {\\boldsymbol {y}} _ {r}} \\left(\\hat {\\boldsymbol {y}} _ {r, i} ^ {k} \\mid \\hat {\\boldsymbol {z}} _ {r}, \\hat {\\boldsymbol {y}} _ {r} ^ {< k}, \\hat {\\boldsymbol {y}} _ {r, < i} ^ {k}, \\hat {\\boldsymbol {y}} _ {l} ^ {< k}, \\hat {\\boldsymbol {y}} _ {l, < i} ^ {k}\\right) \\sim \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {r}, \\boldsymbol {\\sigma} _ {r} ^ {2}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 555, + 784, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the hyperprior component, we generate the dependency $\\tilde{z}_l,\\tilde{z}_r$ from the hyper decoder using Eq. (4). To obtain the channel context, we evenly split the stereo latents on the channel dimension into $K$ slices. These latent slices share a great similarity, providing abundant conditions [25]. To exploit this similarity, the probability estimation is performed slice by slice, using the previous slices as references for the following one. In addition, the decoded slices of the other view also serve as references. Therefore, we propose to use the mutual attention block introduced in the previous subsection to aggregate the channel-wise features $\\Theta_l,\\Theta_r$ from both views, conditioning on $\\hat{\\pmb{y}}_l^{< k}$ and $\\hat{\\pmb{y}}_r^{< k}$ , as depicted by the yellow lines in Fig. 3.", + "bbox": [ + 212, + 583, + 787, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To capture the spatial context within each channel slice, we consistently employ 3D convolution to jointly learn spatial features from the causal areas $\\hat{\\pmb{y}}_{< i}^{k}$ . Specifically, during the coding process, the current compression/decompression only accesses the areas that have already been encoded/decoded. To maintain the causality of the compression, we mask out the uncoded part, which is referred to as masked 3D convolution, as shown in Fig. 4a. In this approach, for the $i$ -th entry $\\hat{\\pmb{y}}_i^k$ , the weights of uncoded part are set to zero, and we obtain spatial", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c798248cfad0cab460d4c8cbcdc3c349835e20719e14546d2fbe16c5977ce264.jpg", + "image_caption": [ + "(a) Masked 3D convolution." + ], + "image_footnote": [], + "bbox": [ + 217, + 161, + 310, + 217 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0f40eb1b7e984197b69af06fb341ef06dee00dd3644a7b311ffca5b4b53b5f31.jpg", + "image_caption": [ + "(b) Auto-regressive pattern in BiSIC entropy model." + ], + "image_footnote": [], + "bbox": [ + 349, + 143, + 537, + 217 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8ac0cdfffcea789de3a4319aec3819d5335fc413e5ea0e637255fc9ef3c32c68.jpg", + "image_caption": [ + "(c) Stereo-checkerboard pattern in BiSIC-Fast.", + "Fig. 4: (Left) Illustration of masked 3D convolution. The blue region is used as the condition for the probability estimation of the current target depicted in red. The weights of the convolution kernel in blue are valid, while the white region is masked to maintain the causal spatial context. (Middle) Demonstration of the auto-regressive process in the cross-dimensional entropy model. The yellow and red arrows represent the channel context and spatial context, respectively. The spatial context is processed entry by entry. (Right) Demonstration of the stereo-checkerboard in the fast variant, where the blue and red parts are each processed only once." + ], + "image_footnote": [], + "bbox": [ + 581, + 143, + 769, + 217 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "context features $\\mathbf{\\Upsilon}_{l,i},\\mathbf{\\Upsilon}_{r,i}$ by considering the causal context $\\hat{\\pmb{y}}_{l, < i}^{k}$ and $\\hat{\\pmb{y}}_{r, < i}^{k}$ from both views simultaneously, as depicted by the red lines in Fig. 3.", + "bbox": [ + 212, + 393, + 782, + 424 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Upon extracting various dependency features, we employ an aggregation network $\\mathbf{G}_{\\mathrm{ag}}$ to fuse the conditions and obtain the probability estimations as:", + "bbox": [ + 214, + 426, + 782, + 455 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {l}, \\boldsymbol {\\sigma} _ {l} ^ {2} = \\mathbf {G} _ {\\mathrm {a g}} \\left(\\tilde {\\boldsymbol {z}} _ {l}, \\boldsymbol {\\Theta} _ {l}, \\boldsymbol {\\Upsilon} _ {l}\\right), \\boldsymbol {\\mu} _ {r}, \\boldsymbol {\\sigma} _ {r} ^ {2} = \\mathbf {G} _ {\\mathrm {a g}} \\left(\\tilde {\\boldsymbol {z}} _ {r}, \\boldsymbol {\\Theta} _ {r}, \\boldsymbol {\\Upsilon} _ {r}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 467, + 784, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The entire procedure in the entropy model is designed to be fully symmetric and bidirectional, which aligns with our motivation for a bidirectional design.", + "bbox": [ + 214, + 497, + 782, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Loss Function", + "text_level": 1, + "bbox": [ + 214, + 551, + 375, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The proposed stereo image compression network is trained using the rate-distortion loss. The rate is defined by bit-per-pixel (BPP), which includes bits for transmitting latents $\\hat{\\pmb{y}}_l, \\hat{\\pmb{y}}_r$ and hyperprioris $\\hat{\\pmb{z}}_l, \\hat{\\pmb{z}}_r$ . The distortion is measured using the mean squared error (MSE) and the multi-scale structural similarity index measure (MS-SSIM) [35]. The overall loss function is given by:", + "bbox": [ + 212, + 579, + 797, + 655 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} := \\lambda D + R = \\lambda \\sum_ {l, r} d (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) + \\sum_ {l, r} (r (\\hat {\\boldsymbol {y}}) + r (\\hat {\\boldsymbol {z}})), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 667, + 784, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $d(\\cdot)$ denotes the distortion function with MSE or MS-SSIM, $r(\\cdot)$ calculates the bitrate using the estimated probability of $\\hat{\\pmb{y}}$ and $\\hat{\\pmb{z}}$ [4], and $\\lambda$ is a hyperparameter that balances the rate and distortion.", + "bbox": [ + 212, + 712, + 782, + 756 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Fast Variant", + "text_level": 1, + "bbox": [ + 214, + 782, + 362, + 796 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The proposed entropy model effectively captures multiple references and estimates the probability distribution. However, the auto-regressive iterations in", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BiSIC: Bidirectional Stereo Image Compression", + "bbox": [ + 415, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/986a5221979dcd8eaae64a0f629b81edad2c86a61503593047d1f0be819543ea.jpg", + "image_caption": [ + "Fig. 5: Illustration of the entropy model for the proposed fast variant based on stereo-checkerboard. The blue and red parts represent the stereo anchor and stereo non-anchor, respectively. By concurrently processing two views, the 3D convolution based anchor context network effectively captures the dependency from the stereo anchor." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 781, + 280 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the spatial context are time-consuming. To strike a better balance between compression runtime and performance, we need to reduce the auto-regressive executions while maintaining a strong condition for the entropy model. To this end, we propose a fast variant of our model, named BiSIC-Fast, by extending the checkerboard structure [15] into stereo-checkerboard and utilizing 3D convolutions to jointly model the spatial context of two views with only two iterations.", + "bbox": [ + 212, + 376, + 782, + 465 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This fast variant maintains the codec as BiSIC, while the entropy model network is modified as illustrated in Fig. 5. In this approach, we split the latents $\\pmb{y}$ into two parts, stereo anchor part and stereo non-anchor part, as represented in blue and red in Fig. 5, respectively. The entropy parameter estimation for the stereo anchor part relies on the hyperprior condition and channel context. The stereo anchor is decompressed first, and then the stereo non-anchor utilizes the decompressed stereo anchor as a dependency. Consequently, the entropy parameter estimation for the stereo non-anchor part is based on the hyperprior, channel context, and stereo anchor context.", + "bbox": [ + 212, + 467, + 784, + 601 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "By adopting this approach, we avoid the spatial auto-regressive process in favor of a two-fold operation, as demonstrated in Fig. 4b and Fig. 4c. Although the spatial auto-regressive process is omitted to save runtime, the neighbors of each pixel in the latents are still maintained, preserving a good perceptive field.", + "bbox": [ + 212, + 603, + 785, + 664 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 688, + 375, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Implementation", + "text_level": 1, + "bbox": [ + 214, + 722, + 509, + 738 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We evaluate the performance of our methods on two benchmark datasets: InStereo2K [5] and Cityscapes [11]. InStereo2K contains 2,060 indoor scene stereo image pairs of size $1,080 \\times 860$ . It is split into 2,010 image pairs for training and 50 for testing. Cityscapes contains 5,000 outdoor scene stereo image pairs of size $2,048 \\times 1,024$ , where 2,975 pairs are allocated for training and 1,525 for testing.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmarks. We compare our methods with traditional CODECs, such as BPG [6], H.265/HEVC [30], MV-HEVC [31], and H.266/VVC [9]. BPG encodes each view of the stereo pair independently, while H.265/HEVC and H.266/VVC compress the left and right view images jointly. In addition, we show the results of learning-based stereo image compression schemes, including HESIC+ [13], SASIC [37], BCSIC [20], LDMIC [41], and ECSIC [36].", + "bbox": [ + 218, + 145, + 785, + 236 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics. The quality of the reconstructed images is evaluated by the peak signal-to-noise ratio (PSNR) and MS-SSIM [35]. The rate is measured by BPP. Besides, we use the widely employed Bjøntegaard Delta bitrate (BDBR) and Bjøntegaard Delta PSNR (BD-PSNR) [8] to compare the rate-distortion (RD) performance among the methods. Specifically, BDBR measures the average bit saving at the same distortion level, while BD-PSNR reflects the relative PSNR gap at an equivalent bitrate. A lower value of BDBR and a higher value of BD-PSNR indicate better performance.", + "bbox": [ + 218, + 237, + 785, + 356 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. During the training process, we optimize the model using the MSE loss over 1.5 million steps. We employ the Adam optimizer [19] and set the batch size to 8. The learning rate is initialized at $10^{-4}$ and is reduced by half after every 300,000 steps. The hyperparameter $\\lambda$ is chosen from the set {256,512,1024,2048,3072,4096} for different bitrates. To achieve better MS-SSIM performance, we fine-tune the pre-trained models for an additional 500,000 steps. More implementation details are provided in the supplementary material.", + "bbox": [ + 218, + 357, + 785, + 463 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Results", + "text_level": 1, + "bbox": [ + 218, + 484, + 321, + 498 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "RD Performance. Fig. 6 shows the RD curves of different compression methods on the InStereo2K and Cityscapes datasets. In terms of the MS-SSIM metric, our proposed method BiSIC outperforms all the other compression methods on both datasets. Besides, BiSIC achieves higher PSNR in the high BPP range compared to the baselines, while maintaining competitive PSNR values at low bitrates. The BDBR results are presented in Tab. 1, where BPG is set as the baseline. Our BiSIC achieves more than $40\\%$ and $50\\%$ BDBR reduction on InStereo2K and Cityscapes, respectively. Notably, compared to the state-of-the-art video codec VVC, BiSIC achieves an additional $12.76\\%$ BDBR reduction for PSNR and an additional $30.08\\%$ BDBR reduction for MS-SSIM on InStereo2K. Compared with unidirectional methods, BiSIC achieves significant bit savings by observing a holistic view and mutually sharing features, which helps remove redundancies in each view. Moreover, compared with the state-of-the-art bidirectional coding scheme BCSIC [20], our method achieves an additional $6.5\\%$ to $15\\%$ BDBR reduction. This is because the employed 3D convolution captures more correlations than the separately deployed 2D convolution used in BCSIC [20], and our cross-dimensional entropy model effectively aggregates various abundant dependencies from different dimensions, providing accurate probability estimation. In addition, the proposed fast variant also demonstrates satisfactory performance, with a marginal degradation in compression performance compared to our BiSIC.", + "bbox": [ + 218, + 507, + 785, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 7 visualizes the reconstructed images of different methods on the Cityscapes dataset. For a fair comparison, all images are com", + "bbox": [ + 218, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BiSIC: Bidirectional Stereo Image Compression", + "bbox": [ + 415, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8ef643ad96c08df9f0355be8c18e1e930c4a0c876452815bcdfc02a1c883b28f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 152, + 444, + 287 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/390ec9a2c255c81e5da490d92ce89cbb58f592afdac005fe955c24d3ec41f720.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 152, + 756, + 287 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/75f78f98a40d55e15127b5cc088019f430362067f9f71edfc3a0b8b1363cf6ce.jpg", + "image_caption": [ + "Fig. 6: Rate-distortion curves in terms of PSNR and MS-SSIM on the InStereo2K and Cityscapes datasets." + ], + "image_footnote": [], + "bbox": [ + 223, + 296, + 444, + 431 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/34ce93292296d44c365028774ed7213c04926af87d59eb6adb2835a68e58ce4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 297, + 754, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "pressed to similar bitrates. We observe that the proposed method not only achieves higher PSNR for the reconstructed images at reduced bitrates but also maintains consistent PSNR across stereo views, thanks to its bidirectional architecture. In comparison, the SASIC method displays a PSNR discrepancy of 0.6487 dB between views, as it compresses the right image based on the left one. The VVC codec compresses images in a predictive manner, resulting in an even larger PSNR gap of 2.5769 dB between views. We leave additional qualitative analysis to the supplementary material.", + "bbox": [ + 212, + 506, + 787, + 626 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Runtime Comparison", + "text_level": 1, + "bbox": [ + 215, + 656, + 439, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this part, we evaluate the computational complexity of various compression algorithms. To ensure a fair comparison, each method is executed on an Intel Xeon Platinum 8370C CPU. Tab. 2 reports the average runtime on the InStereo2K test set, which includes the encoding and decoding latency. We observe that our BiSIC-Fast significantly reduces the runtime compared with the proposed BiSIC. This is because we replace the spatial auto-regressive interweaving in the entropy model with a two-fold stereo-checkerboard, which simplifies the dependency structure within the latent representations. Note that BiSIC-Fast still maintains competitive performance, as demonstrated in Tab. 1. Therefore, it achieves a better trade-off between the compression runtime and performance.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b4f0bee8356dc562677d9781fd5b9be6a799d79226d457e9d865b5d053cfaf7d.jpg", + "table_caption": [ + "Table 1: BDBR values of different compression methods. Bold indicates best results, and underlined values are the second-best ones." + ], + "table_footnote": [], + "table_body": "
MethodInStereo2KCityscapes
PSNRMS-SSIMPSNRMS-SSIM
HEVC [30]-18.80%-12.59%-27.46%-24.36%
MV-HEVC [31]-7.69%2.14%-18.02%-17.13%
VVC [9]-35.31%-31.05%-56.25%-44.04%
HESIC+ [13]-14.96%-43.22%-23.83%-50.79%
SASIC [37]-18.40%-23.87%-21.47%-29.78%
BCSIC [20]-41.22%-54.67%-42.62%-60.72%
LDMIC [41]-41.95%-58.98%-41.92%-61.90%
ECSIC [36]-43.71%-55.65%-52.06%-64.96%
BiSIC (ours)-48.07%-61.13%-57.49%-67.98%
BiSIC-Fast (ours)-45.35%-59.36%-51.96%-65.56%
", + "bbox": [ + 220, + 184, + 782, + 375 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e580b55551af116816dc0b6065fc193a99fd5a3bf27e0326ef507b11aaf4c7d7.jpg", + "image_caption": [ + "Fig. 7: Qualitative comparison of image reconstruction quality across various codecs." + ], + "image_footnote": [], + "bbox": [ + 218, + 385, + 782, + 502 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 556, + 388, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct ablation experiments on the InStereo2K dataset to analyze the contribution of the proposed modules. Our model comprises several key components, including the 3D convolution backbone, cross-dimensional entropy model, and mutual attention block. To evaluate their impact, we replace each module with a baseline counterpart or remove it from the model, and compare the performance in terms of PSNR. The experimental results are shown in Fig. 8 and Tab. 3.", + "bbox": [ + 212, + 582, + 785, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of 3D Convolution Backbone. To assess the capability of 3D convolutional layers in learning inter-view features, we replace 3D convolutional layers in codec with traditional 2D convolutional layers. The resultant model, referred to as Baseline (2D-Conv), experiences a PSNR degradation of 0.3218 dB at the same bitrate, as shown in Fig. 8 and Tab. 3. This indicates that the 3D convolutional layers are more effective at capturing the spatio-stereo information inherent in the data and reducing the redundancies between views.", + "bbox": [ + 212, + 672, + 787, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Cross-Dimensional Entropy Model. To validate the performance improvement offered by our proposed cross-dimensional entropy model, we compare it with the joint auto-regressive and hyperprior entropy model [24], which we refer to as Baseline (Minnen). As shown in Fig. 8 and Tab. 3, the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BiSIC: Bidirectional Stereo Image Compression", + "bbox": [ + 415, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4c3eb28c4105c778c0e615163de62e3a4bd195564e91e4188eeedff4fa3b5d70.jpg", + "table_caption": [ + "Table 2: Average runtime of different methods." + ], + "table_footnote": [], + "table_body": "
CodecBPG HEVC VVC SASIC BCSIC LDMIC BiSIC (ours)BiSIC-Fast (ours)
Runtime16.17s 28.16s 190.27s 20.24s 89.44s 81.35s 167.25s22.82s
", + "bbox": [ + 222, + 171, + 781, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5b612df122bf7c75ea5d93b798d3d1c872b161de567c77fb5223150abfcf1a09.jpg", + "image_caption": [ + "Fig. 8: Rate-distortion curves of different ablation baselines on InStereo2K." + ], + "image_footnote": [], + "bbox": [ + 261, + 234, + 470, + 364 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/11f4d648a16ef25d8c5f106d2cf21cb1cc8741748044c9093d68b5226fe8fab3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBD-PSNR
BiSIC (ours)0
Baseline (2D-Conv)-0.3218 dB
Baseline (Minnen)-0.3475 dB
Baseline (Ying)-0.6554 dB
Baseline (Lei)-0.4956 dB
Baseline (w/o Atten)-0.7882 dB
", + "bbox": [ + 544, + 253, + 785, + 368 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 3: BD-PSNR [8] results relative to the proposed BiSIC.", + "bbox": [ + 542, + 377, + 772, + 406 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "proposed cross-dimensional entropy model achieves better rate-distortion performance than Baseline (Minnen). This suggests that our model provides more accurate probability estimations, which in turn minimizes the coding overhead. Effectiveness of Mutual Attention Block. To evaluate the effectiveness of the proposed mutual attention block, we consider three baselines for comparisons, namely Baseline (Ying), Baseline (Lei), and Baseline (w/o Atten). Baseline (Ying) and Baseline (Lei) replace the mutual attention block with the stereo attention module [39] and the contextual transfer module [20], respectively. In particular, Baseline (w/o Atten) represents a variant without the attention module. As shown in Fig. 8 and Tab. 3, our proposed model outperforms all baselines by a large margin. Notably, removing the proposed mutual attention block results in a dramatic degradation, which demonstrates its significance.", + "bbox": [ + 212, + 436, + 787, + 618 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 641, + 359, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce a bidirectional stereo image compression network that employs 3D convolution to encode local stereo features and propose a mutual attention block to capture global features. Besides, we design a symmetric cross-dimensional entropy model that integrates hyperprior, spatial context, channel context, and stereo dependency. Our bidirectional design mitigates the issue of imbalanced compression quality intrinsic to unidirectional methods and eliminates the requirement for sequential coding. As demonstrated in the experimental results, our BiSIC achieves greater bit savings at the same level of compression quality compared to both traditional coding standards and existing learning-based methods. Moreover, we propose a fast variant that significantly reduces the compression runtime while maintaining competitive performance.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the General Research Fund (Project No. 16209622) from the Hong Kong Research Grants Council.", + "bbox": [ + 215, + 172, + 782, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 224, + 321, + 239 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arena, F., Collotta, M., Pau, G., Termine, F.: An overview of augmented reality. Computers 11(2), 28 (2022)", + "2. Balle, J., Chou, P.A., Minnen, D., Singh, S., Johnston, N., Agustsson, E., Hwang, S.J., Toderici, G.: Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing 15(2), 339-353 (2020)", + "3. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. arXiv preprint arXiv:1611.01704 (2016)", + "4. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 (2018)", + "5. Bao, W., Wang, W., Xu, Y., Guo, Y., Hong, S., Zhang, X.: Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences 63, 1-11 (2020)", + "6. Bellard, F.: BPG image format. Website (2014), https://bellard.org/bpg/", + "7. Bishop, C.M.: Latent variable models. In: Learning in graphical models, pp. 371-403. Springer (1998)", + "8. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33 (2001)", + "9. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology 31(10), 3736-3764 (2021)", + "10. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "11. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)", + "12. Deng, X., Deng, Y., Yang, R., Yang, W., Timofte, R., Xu, M.: MASIC: Deep mask stereo image compression. IEEE Transactions on Circuits and Systems for Video Technology (2023)", + "13. Deng, X., Yang, W., Yang, R., Xu, M., Liu, E., Feng, Q., Timofte, R.: Deep homography for efficient stereo image compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1492-1501 (June 2021)", + "14. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5718-5727 (2022)", + "15. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14771-14780 (2021)" + ], + "bbox": [ + 218, + 253, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BiSIC: Bidirectional Stereo Image Compression", + "bbox": [ + 415, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hu, Y., Yang, W., Liu, J.: Coarse-to-fine hyper-prior modeling for learned image compression. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11013-11020 (2020)", + "17. Jiang, W., Yang, J., Zhai, Y., Ning, P., Gao, F., Wang, R.: Mlic: Multi-reference entropy model for learned image compression. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7618-7627 (2023)", + "18. Kadaikar, A., Dauphin, G., Mokraoui, A.: Joint disparity and variable size-block optimization algorithm for stereoscopic image compression. Signal Processing: Image Communication 61, 1-8 (2018)", + "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "20. Lei, J., Liu, X., Peng, B., Jin, D., Li, W., Gu, J.: Deep stereo image compression via bi-directional coding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19669-19678 (June 2022)", + "21. Liu, J., Wang, S., Urtasun, R.: DSIC: Deep stereo image compression. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)", + "22. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14388-14397 (2023)", + "23. Liu, Y., Ren, J., Zhang, J., Liu, J., Lin, M.: Visually imbalanced stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2029-2038 (2020)", + "24. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems 31 (2018)", + "25. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3339-3343. IEEE (2020)", + "26. Pan, G., Lu, G., Hu, Z., Xu, D.: Content adaptive latents and decoder for neural image compression. In: European Conference on Computer Vision. pp. 556-573. Springer (2022)", + "27. Shen, Z., Zhang, M., Zhao, H., Yi, S., Li, H.: Efficient attention: Attention with linear complexities. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 3531-3539 (2021)", + "28. Skodras, A., Christopoulos, C., Ebrahimi, T.: The jpeg 2000 still image compression standard. IEEE Signal processing magazine 18(5), 36-58 (2001)", + "29. Song, X., Zhao, X., Fang, L., Hu, H., Yu, Y.: Edge stereo: An effective multi-task learning network for stereo matching and edge detection. International Journal of Computer Vision 128, 910-930 (2020)", + "30. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (HEVC) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012)", + "31. Tech, G., Chen, Y., Müller, K., Ohm, J.R., Vetro, A., Wang, Y.K.: Overview of the multiview and 3d extensions of high efficiency video coding. IEEE Transactions on Circuits and Systems for Video Technology 26(1), 35-49 (2015)", + "32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Vetro, A., Wiegand, T., Sullivan, G.J.: Overview of the stereo and multiview video coding extensions of the h. 264/mpeg-4 avc standard. Proceedings of the IEEE 99(4), 626-642 (2011)", + "34. Wallace, G.K.: The JPEG still picture compression standard. Communications of the ACM 34(4), 30-44 (1991)", + "35. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402. IEEE (2003)", + "36. Wödlinger, M., Kotera, J., Keglevic, M., Xu, J., Sablatnig, R.: ECSIC: Epipolar cross attention for stereo image compression. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3436-3445 (2024)", + "37. Wödlinger, M., Kotera, J., Xu, J., Sablatnig, R.: SASIC: Stereo image compression with latent shifts and stereo attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 661-670 (June 2022)", + "38. Yin, H., Wang, Y., Tang, L., Ding, X., Huang, S., Xiong, R.: 3d lidar map compression for efficient localization on resource constrained vehicles. IEEE transactions on intelligent transportation systems 22(2), 837-852 (2020)", + "39. Ying, X., Wang, Y., Wang, L., Sheng, W., An, W., Guo, Y.: A stereo attention module for stereo image super-resolution. IEEE Signal Processing Letters 27, 496-500 (2020)", + "40. Zhai, Y., Tang, L., Ma, Y., Peng, R., Wang, R.: Disparity-based stereo image compression with aligned cross-view priors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2351-2360 (2022)", + "41. Zhang, X., Shao, J., Zhang, J.: LDMIC: Learning-based distributed multi-view image coding. arXiv preprint arXiv:2301.09799 (2023)", + "42. Zhu, Y., Yang, Y., Cohen, T.: Transformer-based transform coding. In: International Conference on Learning Representations (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 534 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BiSIC: Bidirectional Stereo Image Compression", + "bbox": [ + 415, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/e50db183-7c10-4237-a52b-c78459d9ff33_model.json b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/e50db183-7c10-4237-a52b-c78459d9ff33_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ac5cc61b53aa333980c2309d658eeec6adc2675a --- /dev/null +++ b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/e50db183-7c10-4237-a52b-c78459d9ff33_model.json @@ -0,0 +1,2214 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.244, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.213, + 0.663, + 0.245 + ], + "angle": 0, + "content": "Zhening Liu, Xinjie Zhang, Jiawei Shao, Zehong Lin, and Jun Zhang" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.256, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong SAR {zhening.liu,xzhangga,jiawei.shao}@connect.ust.hk, {eezhlin,eejzhang}@ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.339, + 0.744, + 0.56 + ], + "angle": 0, + "content": "Abstract. With the rapid advancement of stereo vision technologies, stereo image compression has emerged as a crucial field that continues to draw significant attention. Previous approaches have primarily employed a unidirectional paradigm, where the compression of one view is dependent on the other, resulting in imbalanced compression. To address this issue, we introduce a symmetric bidirectional stereo image compression architecture, named BiSIC. Specifically, we propose a 3D convolution based codec backbone to capture local features and incorporate bidirectional attention blocks to exploit global features. Moreover, we design a novel cross-dimensional entropy model that integrates various conditioning factors, including the spatial context, channel context, and stereo dependency, to effectively estimate the distribution of latent representations for entropy coding. Extensive experiments demonstrate that our proposed BiSIC outperforms conventional image/video compression standards, as well as state-of-the-art learning-based methods, in terms of both PSNR and MS-SSIM." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.575, + 0.741, + 0.603 + ], + "angle": 0, + "content": "Keywords: Stereo Image Compression \\(\\cdot\\) Bidirectional Architecture \\(\\cdot\\) Cross-Dimensional Entropy Model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.63, + 0.377, + 0.646 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Stereo vision, which mimics human binocular vision, has emerged as an important area in computer vision. It enables a wide range of applications, such as 3D movies, remote sensing [29], autonomous driving [38], and augmented/virtual reality (AR/VR) [1]. With the rapid development of stereo cameras, stereo images have become essential for providing immersive and convincing visual presentations that enhance user experience and benefit downstream applications. Meanwhile, the proliferation of stereo cameras leads to a significant surge in the volume of stereo images, resulting in substantial storage costs and posing challenges to transmission efficiency. Consequently, stereo image compression plays a pivotal role in ensuring the practicality and effectiveness of stereo vision." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "In recent years, numerous methods have been developed for single image compression, including traditional CODECs [6, 34] and learning-based methods [2-4, 10, 24]. These methods, however, focus on extracting individual features for compression and fail to capture the cross-view correlation. Consequently, they are less effective for stereo image compression that requires the extraction of both intra-view information and inter-view information. To exploit the inherent shared information in stereo images, traditional methods for stereo image coding follow a predictive compression procedure of video standards [9, 31]. In this process, one view is compressed first and then used to predict the other view. Subsequently, the disparity between the predicted and actual views is compressed. Nonetheless, these methods typically rely on hand-crafted modules and lack end-to-end optimization, resulting in suboptimal performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.79, + 0.601 + ], + "angle": 0, + "content": "Inspired by the success of learning-based single image compression, researchers have proposed novel stereo image compression methods using deep neural networks (DNNs) to enhance compression performance [12, 13, 20, 21, 36, 37, 40, 41]. Initially, extensive research focuses on unidirectional learned stereo image compression methods [12, 13, 21, 36, 37, 40], where one view is compressed and used as a reference to compress the other view. Despite achieving high overall rate-distortion (RD) performance, these unidirectional learned compression methods yield a noticeable imbalance in compression quality between stereo views, similar to traditional predictive compression methods. This imbalance is unfavorable for human vision and undermines downstream machine vision tasks [23]. Moreover, some of these methods [12, 13, 37, 40] heavily rely on sequential coding order and accurate estimation of the explicit warping relationship between views. To avoid these issues, a recent study [20] proposes a bidirectional compression method that leverages 2D convolution to extract individual features and facilitate information sharing between views. Nonetheless, the compression performance is limited due to the separate processing of convolutions and the plain spatial context based entropy model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Based on the above discussion, we propose a novel bidirectional method for stereo image compression, named BiSIC, to address the aforementioned issues. BiSIC achieves improvements in four key aspects: encoder/decoder structure, mutual stereo learning, entropy model, and coding speed. Firstly, regarding the codec structure, although 2D convolutions have proven effective in various computer vision tasks, they fail to capture the aligned morphs and features between stereo views. To overcome this limitation, we propose using 3D convolution as the backbone for the encoder/decoder, which takes both views as input to capture the inter-view dependencies that are inaccessible in previous works based on 2D convolution. Secondly, for mutual stereo learning, we propose a mutual attention block to facilitate the transfer of features between views. This module enhances the integration and cooperation of the stereo views, leading to improved compression performance. Thirdly, for the entropy model, prior methods primarily use hyperprior and spatial context as conditions, overlooking the rich inter-view dependencies across other dimensions. To exploit these dependencies," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.416, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BiSIC: Bidirectional Stereo Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.143, + 0.73, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Fig. 1: The proposed bidirectional stereo image compression architecture. AE and AD are arithmetic coder for entropy coding. The backbones of encoder \\(\\mathbf{E}\\), decoder \\(\\mathbf{D}\\), hyper encoder \\(h_a\\), and hyper decoder \\(h_s\\) are constructed with 3D convolution to model local features. 3DCovTr denotes transposed 3D convolution. To enhance global feature extraction, bidirectional mutual attention blocks are inserted between 3D convolutional layers. Moreover, a novel cross-dimensional entropy model is utilized to capture complex inter-view dependencies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.784, + 0.463 + ], + "angle": 0, + "content": "we construct a cross-dimensional entropy model that aggregates hyperprior, spatial context, channel context, and stereo dependency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.785, + 0.554 + ], + "angle": 0, + "content": "Specifically, we introduce a masked 3D convolution to capture the stereo spatial context and utilize an attention block to process the stereo channel context. Finally, to improve coding speed and reduce auto-regressive steps, we design a fast variant based on a stereo-checkerboard structure. This design converts auto-regressive inference into a two-fold operation, improving the coding efficiency. The contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.559, + 0.785, + 0.603 + ], + "angle": 0, + "content": "- We propose a novel 3D convolution based codec backbone that processes stereo images as a whole, enabling the learning of local stereo features and effective capture of shared patterns across views." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.604, + 0.785, + 0.647 + ], + "angle": 0, + "content": "- We design a bidirectional mutual attention block that facilitates the exchange of global features between stereo views. This block produces compact attention maps and is robust to input images of different sizes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.691 + ], + "angle": 0, + "content": "- We develop a powerful symmetric cross-dimensional entropy model, which incorporates hyperpriors, spatial context, channel context, and stereo dependency as conditioning factors, to enhance the probability estimation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.692, + 0.785, + 0.736 + ], + "angle": 0, + "content": "- To accelerate coding, we propose a stereo-checkerboard design that transforms the spatial auto-regressive inference into a two-fold process. This design strikes a balance between time consumption and compression quality." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.559, + 0.785, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.762, + 0.396, + 0.777 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.48, + 0.804 + ], + "angle": 0, + "content": "2.1 Single Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Image compression is an important topic that continuously attracts significant research interest. With the increasing demand for vision products and applica" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "tions, steady progress has been made in traditional image compression standards, evolving from JPEG [28, 34] and BPG [6] to video compression standards such as HEVC [30] and VVC [9]. These traditional methods establish a standard compression pipeline comprising transform, quantization, and entropy coding. Over the past five years, deep learning has revolutionized various vision applications. As demonstrated in the seminal work [3], DNNs have the potential to replace components in traditional compression methods and enable end-to-end compression. This work introduces uniform noise to mimic the distortion caused by quantization, thereby paving the way for end-to-end optimization in image compression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.789, + 0.526 + ], + "angle": 0, + "content": "In learned image compression models [3,4,10,14-17,22,24-26,42], neural networks are employed to parameterize the probability distribution of transformed latents. Accurate estimation of this distribution is crucial for achieving compact compression via entropy coding. To enhance probability estimation, various conditions have been integrated into entropy models [2], including hyperpriors [4], spatial context [24], and channel-wise context [25], to capture a broader range of dependencies. To be specific, hyperprior is a lightweight subsidiarity information transmitted alongside the encoded latents. Both spatial context and channel-wise context leverage auto-regressive inference, which is conditioned on the preceding segments of data by iteratively interleaving the inference. Checkerboard methods [15], on the other hand, avoid the spatial auto-regressive process and accelerate the coding speed by constraining the spatial context to the near neighbors of each pixel instead of spatial causal components. These dependencies operate on diverse scales and can be progressively combined to further improve effectiveness, as demonstrated in [14, 17]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.546, + 0.483, + 0.56 + ], + "angle": 0, + "content": "2.2 Stereo Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Stereo image compression is an extension of single image compression that considers capturing both intra-view features and inter-view correlations. Traditional approaches employ key-frame based predictive coding, akin to video coding [18, 33]. However, these methods rely on hand-crafted feature extractors and classical optimization techniques. As these modules are optimized individually, they often fail to achieve satisfactory RD performance. Moreover, the predictive compression pipeline also leads to imbalanced compression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.81 + ], + "angle": 0, + "content": "In recent years, learned stereo image compression methods have gained increasing attention [12, 13, 20, 21, 36, 37, 40, 41]. These methods can be broadly categorized into unidirectional and bidirectional approaches. Unidirectional methods first compress one view image and then use it as a reference to compress the other view, significantly reducing the required bits. The pioneering work [21] introduces a skip module to warp the features from one view to encode the other. Subsequent works have utilized various techniques to represent the warping relationship between stereo views, such as homography matrices [12, 13], disparity maps [40], and horizontal shifts [37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "These unidirectional methods heavily rely on the precise estimation of the spatial transformations. A very recent work, ECSIC [36], avoids explicit para" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.416, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "BiSIC: Bidirectional Stereo Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "metric transformation in favor of constructing a unidirectional entropy model that conditions the right image on the known left image. However, the decompressed results of these unidirectional methods often have imbalanced quality due to the one-way dependency between stereo views. On the other hand, bidirectional methods, while remaining less explored, produce more balanced quality by extracting shared information from both views and compressing stereo images simultaneously. One such approach [20] introduces a contextual transform module for feature transfer between views. However, it employs separate down-sampling networks and only takes spatial context into account, which limits the compression performance. In contrast, to fully exploit the shared information and establish stronger dependencies, we employ a joint downsampling codec and a cross-dimensional entropy model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.436, + 0.368 + ], + "angle": 0, + "content": "3 Proposed Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.789, + 0.503 + ], + "angle": 0, + "content": "The proposed architecture for stereo image compression, as depicted in Fig. 1, consists of two main components: a joint codec and an entropy model. The codec employs 3D convolutions to transform the original stereo images into compact latent representations, effectively exploiting the correlation between the two views. Subsequently, the entropy model estimates the probability distribution of these latent representations, enabling the generation of a compact bitstream through entropy coding. In the following subsections, we elaborate on the codec and the entropy model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.524, + 0.362, + 0.538 + ], + "angle": 0, + "content": "3.1 Joint Codec" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.668 + ], + "angle": 0, + "content": "3D Convolution Backbone. The codec acts as a non-linear transformation that downsamples the raw images into compact latent representations and generates hyperpriors for the entropy model. Different from previous approaches to stereo image compression that employ 2D convolutional layers, we adopt 3D convolutions as the backbone of our codec. Unlike 2D convolution, which operates separately on each view, 3D convolution concurrently processes both stereo images. Consequently, it inherently possesses the capability to extract inter-view correlations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.789, + 0.791 + ], + "angle": 0, + "content": "Our encoder, denoted by \\(\\mathbf{E}\\), comprises four 3D convolutional layers, with bidirectional mutual attention blocks inserted after the second and the fourth layers to facilitate global feature communication between the two views. The encoder transforms a pair of stereo images, i.e., the left and right images denoted by \\(\\pmb{x}_l\\), \\(\\pmb{x}_r \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}\\), into compact latent representations \\(\\pmb{y}_l\\), \\(\\pmb{y}_r \\in \\mathbb{R}^{B \\times N \\times \\frac{H}{16} \\times \\frac{W}{16}}\\) and quantizes them into \\(\\hat{\\pmb{y}} := \\{\\hat{\\pmb{y}}_l, \\hat{\\pmb{y}}_r\\}\\) for efficient transmission. Conversely, the decoder \\(\\mathbf{D}\\) transforms the quantized latents \\(\\hat{\\pmb{y}}\\) back into images \\(\\hat{\\pmb{x}}_l\\) and \\(\\hat{\\pmb{x}}_r\\). The encoding and decoding processes are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.803, + 0.785, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {l}, \\boldsymbol {y} _ {r} = \\mathbf {E} \\left(\\boldsymbol {x} _ {l}, \\boldsymbol {x} _ {r}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.821, + 0.785, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {l}, \\hat {\\boldsymbol {x}} _ {r} = \\mathbf {D} (\\hat {\\boldsymbol {y}} _ {l}, \\hat {\\boldsymbol {y}} _ {r}). \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.27 + ], + "angle": 0, + "content": "To capture the dependencies among elements in the latents \\(\\hat{\\pmb{y}}\\) for efficient entropy coding [4,7], hyperpriors are generated through another 3D convolution based hyper encoder \\(h_a\\). The generated hyperpriors \\(\\pmb{z}_l, \\pmb{z}_r \\in \\mathbb{R}^{B \\times M \\times \\frac{H}{64} \\times \\frac{W}{64}}\\) are quantized into \\(\\hat{\\pmb{z}}_l, \\hat{\\pmb{z}}_r\\) and then transmitted alongside the quantized latents \\(\\hat{\\pmb{y}}\\). At the decoder side, a 3D convolution based hyper decoder \\(h_s\\) upscales the quantized hyperpriors \\(\\hat{\\pmb{z}}_l, \\hat{\\pmb{z}}_r\\) into \\(\\tilde{\\pmb{z}}_l, \\tilde{\\pmb{z}}_r\\), which are subsequently utilized in the entropy model, as elaborated in Sec. 3.2. The processes of the hyper encoder and hyper decoder are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.279, + 0.785, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {l}, \\boldsymbol {z} _ {r} = h _ {a} (\\boldsymbol {y} _ {l}, \\boldsymbol {y} _ {r}), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.297, + 0.785, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {z}} _ {l}, \\tilde {\\boldsymbol {z}} _ {r} = h _ {s} (\\hat {\\boldsymbol {z}} _ {l}, \\hat {\\boldsymbol {z}} _ {r}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.784, + 0.439 + ], + "angle": 0, + "content": "Bidirectional Mutual Attention Block. Convolutional layers have an excellent ability for local modeling, but their ability for long-range feature modeling is relatively limited. On the other hand, attention modules [32] excel at extracting long-range features. Therefore, we propose combining the attention mechanism with convolutional layers to compensate for the shortcomings of the latter and leverage the advantages of both convolution and attention. To this end, we design a novel bidirectional mutual attention block, as depicted in Fig. 2, to be inserted between convolutional layers as the pink blocks shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.605 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, the bidirectional mutual attention block contains two cross-attention stages, each followed by a self-attention layer. The two cross-attention stages differ in how they concentrate on the feature map and are referred to as cross-key attention and cross-query attention. Before each cross-attention stage, the input stereo data \\(\\pmb{y}_l, \\pmb{y}_r\\) first pass through a residual block, which are then embedded into low-dimensional query, key, and value tensors \\(\\pmb{Q}, \\pmb{K}, \\pmb{V} \\in \\mathbb{R}^{B \\times C \\times (H \\times W)}\\). To avoid the prohibitive computation of classical attention, which yields an \\((H \\times W) \\times (H \\times W)\\) sized attention map, we adopt the efficient attention approach proposed in [27]. This approach computes the product of the key and the transposed value, thereby reducing the attention map size to \\(C_K \\times C_V\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.785, + 0.664 + ], + "angle": 0, + "content": "In the cross-key attention stage, for each view, we generate an attention map using its key and the value from the other view, which is then queried by the current view. By doing so, we can extract the cross-key feature \\(\\Phi\\) from the stereo views as:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.673, + 0.785, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Phi} _ {r \\rightarrow l} = (\\sigma (\\boldsymbol {K} _ {r}) \\times \\boldsymbol {V} _ {l} ^ {T}) ^ {T} \\times \\sigma (\\boldsymbol {Q} _ {r}), \\quad \\boldsymbol {\\Phi} _ {l \\rightarrow r} = (\\sigma (\\boldsymbol {K} _ {l}) \\times \\boldsymbol {V} _ {r} ^ {T}) ^ {T} \\times \\sigma (\\boldsymbol {Q} _ {l}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.784, + 0.772 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the softmax function. This approach calculates the attention map using the key and value from different views, thereby facilitating inter-view alignment and the identification of common patterns. When a feature in one view is correlated with the corresponding feature in the other view, it is effectively captured through the attention map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.817 + ], + "angle": 0, + "content": "In the cross-query attention stage, the attention map is obtained using the key and value from a single view, and then queried by the other view, yielding the cross-query feature \\(\\Psi\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.824, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {r \\rightarrow l} = \\left(\\sigma \\left(\\boldsymbol {K} _ {l}\\right) \\times \\boldsymbol {V} _ {l} ^ {T}\\right) ^ {T} \\times \\sigma \\left(\\boldsymbol {Q} _ {r}\\right), \\Psi_ {l \\rightarrow r} = \\left(\\sigma \\left(\\boldsymbol {K} _ {r}\\right) \\times \\boldsymbol {V} _ {r} ^ {T}\\right) ^ {T} \\times \\sigma \\left(\\boldsymbol {Q} _ {l}\\right). \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.416, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "BiSIC: Bidirectional Stereo Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.144, + 0.75, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Fig. 2: Overview of the proposed bidirectional mutual attention block. The blue and green lines represent features extracted from the left view and right view, respectively. The network structures of the residual block, the basic efficient attention unit, and the combine block are illustrated on the right, from top to bottom." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.785, + 0.449 + ], + "angle": 0, + "content": "The attention map in this stage is generated within one view, thus maintaining an inner-focused relationship. This process preserves the individual self-features of each view, while the other view serves as a reference for determining the allocation of attention weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.542 + ], + "angle": 0, + "content": "After obtaining the cross-key attention and cross-query attention features \\(\\Phi, \\Psi\\), we employ self-attention to augment these features as \\(\\tilde{\\Phi}, \\tilde{\\Psi}\\) and use a shared-parameter combine block to get the desired mutual attention output, as shown in the bottom right of Fig. 2. Specifically, the combine block concatenates the augmented features \\(\\tilde{\\Phi}, \\tilde{\\Psi}\\) and the residual component, and shrinks the channel number to be consistent with the block input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.633 + ], + "angle": 0, + "content": "This bidirectional mutual attention block captures diverse features and provides a global receptive field for the codec. It is suitable for any stereo information transfer scenarios, provided that the inputs are two features of equal size, making it a generalized plug-and-play block. Moreover, the size of the attention maps is determined only by the embedding channel \\( C \\) and does not vary with the size of input images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.66, + 0.547, + 0.675 + ], + "angle": 0, + "content": "3.2 Cross-Dimensional Entropy Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The proposed codec effectively transforms stereo images into small-sized latent representations. To achieve efficient compression, the next crucial step is to encode these latents into compact bitstreams using accurate probability distribution and entropy coding [4,24]. Thus, an entropy model is critical for parameterizing and estimating probabilities. In learned image compression, the probability distribution is commonly modeled as Gaussian distributions [24], with the mean \\(\\mu\\) and variance \\(\\sigma^2\\) being estimated. To achieve accurate estimation, it is critical to provide appropriate references. Apart from hyperpriors, previous stereo image compression methods primarily use spatial context [13,20,41] and unidirectional stereo dependency as references [21,36], while abundant features on the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.143, + 0.756, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.789, + 0.392 + ], + "angle": 0, + "content": "Fig. 3: Illustration of the proposed symmetric cross-dimensional entropy model. It jointly aggregates the hyperprior (blue), stereo spatial context (red), and stereo channel context (yellow) as conditions for an effective probability distribution estimation. The hyper decoder \\( h_s \\) for hyperprior is shown in Fig. 1. The masked 3D convolution for spatial context is provided in Fig. 4a. Moreover, the mutual attention block for channel context is detailed in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.784, + 0.464 + ], + "angle": 0, + "content": "channel dimension are overlooked. In this work, we propose a symmetric cross-dimensional entropy model that integrates hyperprior, spatial context, channel context, and stereo dependency in a fully symmetric bidirectional way." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.526 + ], + "angle": 0, + "content": "The architecture of the proposed cross-dimensional entropy model is depicted in Fig. 3. We use \\(\\hat{\\pmb{y}}_{l,i}^{k}\\) to denote the \\(i\\)-th \\((i = 1,2,\\dots,n)\\) entry of the \\(k\\)-th \\((k = 1,2,\\dots,K)\\) channel slice from the received left view latent \\(\\hat{\\pmb{y}}_l\\), and likewise for the others. The cross-dimensional condition formulation is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.534, + 0.785, + 0.554 + ], + "angle": 0, + "content": "\\[\np _ {\\hat {\\boldsymbol {y}} _ {l}} \\left(\\hat {\\boldsymbol {y}} _ {l, i} ^ {k} \\mid \\hat {\\boldsymbol {z}} _ {l}, \\hat {\\boldsymbol {y}} _ {l} ^ {< k}, \\hat {\\boldsymbol {y}} _ {l, < i} ^ {k}, \\hat {\\boldsymbol {y}} _ {r} ^ {< k}, \\hat {\\boldsymbol {y}} _ {r, < i} ^ {k}\\right) \\sim \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {l}, \\boldsymbol {\\sigma} _ {l} ^ {2}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.556, + 0.785, + 0.575 + ], + "angle": 0, + "content": "\\[\np _ {\\hat {\\boldsymbol {y}} _ {r}} \\left(\\hat {\\boldsymbol {y}} _ {r, i} ^ {k} \\mid \\hat {\\boldsymbol {z}} _ {r}, \\hat {\\boldsymbol {y}} _ {r} ^ {< k}, \\hat {\\boldsymbol {y}} _ {r, < i} ^ {k}, \\hat {\\boldsymbol {y}} _ {l} ^ {< k}, \\hat {\\boldsymbol {y}} _ {l, < i} ^ {k}\\right) \\sim \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {r}, \\boldsymbol {\\sigma} _ {r} ^ {2}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.734 + ], + "angle": 0, + "content": "For the hyperprior component, we generate the dependency \\(\\tilde{z}_l,\\tilde{z}_r\\) from the hyper decoder using Eq. (4). To obtain the channel context, we evenly split the stereo latents on the channel dimension into \\(K\\) slices. These latent slices share a great similarity, providing abundant conditions [25]. To exploit this similarity, the probability estimation is performed slice by slice, using the previous slices as references for the following one. In addition, the decoded slices of the other view also serve as references. Therefore, we propose to use the mutual attention block introduced in the previous subsection to aggregate the channel-wise features \\(\\Theta_l,\\Theta_r\\) from both views, conditioning on \\(\\hat{\\pmb{y}}_l^{< k}\\) and \\(\\hat{\\pmb{y}}_r^{< k}\\), as depicted by the yellow lines in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To capture the spatial context within each channel slice, we consistently employ 3D convolution to jointly learn spatial features from the causal areas \\(\\hat{\\pmb{y}}_{< i}^{k}\\). Specifically, during the coding process, the current compression/decompression only accesses the areas that have already been encoded/decoded. To maintain the causality of the compression, we mask out the uncoded part, which is referred to as masked 3D convolution, as shown in Fig. 4a. In this approach, for the \\(i\\)-th entry \\(\\hat{\\pmb{y}}_i^k\\), the weights of uncoded part are set to zero, and we obtain spatial" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.416, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BiSIC: Bidirectional Stereo Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.162, + 0.311, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.22, + 0.321, + 0.24 + ], + "angle": 0, + "content": "(a) Masked 3D convolution." + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.145, + 0.538, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.22, + 0.551, + 0.241 + ], + "angle": 0, + "content": "(b) Auto-regressive pattern in BiSIC entropy model." + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.145, + 0.77, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.22, + 0.78, + 0.24 + ], + "angle": 0, + "content": "(c) Stereo-checkerboard pattern in BiSIC-Fast." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.364 + ], + "angle": 0, + "content": "Fig. 4: (Left) Illustration of masked 3D convolution. The blue region is used as the condition for the probability estimation of the current target depicted in red. The weights of the convolution kernel in blue are valid, while the white region is masked to maintain the causal spatial context. (Middle) Demonstration of the auto-regressive process in the cross-dimensional entropy model. The yellow and red arrows represent the channel context and spatial context, respectively. The spatial context is processed entry by entry. (Right) Demonstration of the stereo-checkerboard in the fast variant, where the blue and red parts are each processed only once." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.784, + 0.425 + ], + "angle": 0, + "content": "context features \\(\\mathbf{\\Upsilon}_{l,i},\\mathbf{\\Upsilon}_{r,i}\\) by considering the causal context \\(\\hat{\\pmb{y}}_{l, < i}^{k}\\) and \\(\\hat{\\pmb{y}}_{r, < i}^{k}\\) from both views simultaneously, as depicted by the red lines in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.427, + 0.784, + 0.456 + ], + "angle": 0, + "content": "Upon extracting various dependency features, we employ an aggregation network \\(\\mathbf{G}_{\\mathrm{ag}}\\) to fuse the conditions and obtain the probability estimations as:" + }, + { + "type": "equation", + "bbox": [ + 0.309, + 0.468, + 0.785, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {l}, \\boldsymbol {\\sigma} _ {l} ^ {2} = \\mathbf {G} _ {\\mathrm {a g}} \\left(\\tilde {\\boldsymbol {z}} _ {l}, \\boldsymbol {\\Theta} _ {l}, \\boldsymbol {\\Upsilon} _ {l}\\right), \\boldsymbol {\\mu} _ {r}, \\boldsymbol {\\sigma} _ {r} ^ {2} = \\mathbf {G} _ {\\mathrm {a g}} \\left(\\tilde {\\boldsymbol {z}} _ {r}, \\boldsymbol {\\Theta} _ {r}, \\boldsymbol {\\Upsilon} _ {r}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.498, + 0.784, + 0.528 + ], + "angle": 0, + "content": "The entire procedure in the entropy model is designed to be fully symmetric and bidirectional, which aligns with our motivation for a bidirectional design." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.553, + 0.377, + 0.566 + ], + "angle": 0, + "content": "3.3 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.798, + 0.656 + ], + "angle": 0, + "content": "The proposed stereo image compression network is trained using the rate-distortion loss. The rate is defined by bit-per-pixel (BPP), which includes bits for transmitting latents \\(\\hat{\\pmb{y}}_l, \\hat{\\pmb{y}}_r\\) and hyperprioris \\(\\hat{\\pmb{z}}_l, \\hat{\\pmb{z}}_r\\). The distortion is measured using the mean squared error (MSE) and the multi-scale structural similarity index measure (MS-SSIM) [35]. The overall loss function is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.668, + 0.785, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} := \\lambda D + R = \\lambda \\sum_ {l, r} d (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) + \\sum_ {l, r} (r (\\hat {\\boldsymbol {y}}) + r (\\hat {\\boldsymbol {z}})), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.784, + 0.757 + ], + "angle": 0, + "content": "where \\(d(\\cdot)\\) denotes the distortion function with MSE or MS-SSIM, \\(r(\\cdot)\\) calculates the bitrate using the estimated probability of \\(\\hat{\\pmb{y}}\\) and \\(\\hat{\\pmb{z}}\\) [4], and \\(\\lambda\\) is a hyperparameter that balances the rate and distortion." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.364, + 0.797 + ], + "angle": 0, + "content": "3.4 Fast Variant" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The proposed entropy model effectively captures multiple references and estimates the probability distribution. However, the auto-regressive iterations in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.782, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 5: Illustration of the entropy model for the proposed fast variant based on stereo-checkerboard. The blue and red parts represent the stereo anchor and stereo non-anchor, respectively. By concurrently processing two views, the 3D convolution based anchor context network effectively captures the dependency from the stereo anchor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.784, + 0.467 + ], + "angle": 0, + "content": "the spatial context are time-consuming. To strike a better balance between compression runtime and performance, we need to reduce the auto-regressive executions while maintaining a strong condition for the entropy model. To this end, we propose a fast variant of our model, named BiSIC-Fast, by extending the checkerboard structure [15] into stereo-checkerboard and utilizing 3D convolutions to jointly model the spatial context of two views with only two iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.785, + 0.602 + ], + "angle": 0, + "content": "This fast variant maintains the codec as BiSIC, while the entropy model network is modified as illustrated in Fig. 5. In this approach, we split the latents \\( \\pmb{y} \\) into two parts, stereo anchor part and stereo non-anchor part, as represented in blue and red in Fig. 5, respectively. The entropy parameter estimation for the stereo anchor part relies on the hyperprior condition and channel context. The stereo anchor is decompressed first, and then the stereo non-anchor utilizes the decompressed stereo anchor as a dependency. Consequently, the entropy parameter estimation for the stereo non-anchor part is based on the hyperprior, channel context, and stereo anchor context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.787, + 0.665 + ], + "angle": 0, + "content": "By adopting this approach, we avoid the spatial auto-regressive process in favor of a two-fold operation, as demonstrated in Fig. 4b and Fig. 4c. Although the spatial auto-regressive process is omitted to save runtime, the neighbors of each pixel in the latents are still maintained, preserving a good perceptive field." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.377, + 0.707 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.511, + 0.739 + ], + "angle": 0, + "content": "4.1 Experimental Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets. We evaluate the performance of our methods on two benchmark datasets: InStereo2K [5] and Cityscapes [11]. InStereo2K contains 2,060 indoor scene stereo image pairs of size \\(1,080 \\times 860\\). It is split into 2,010 image pairs for training and 50 for testing. Cityscapes contains 5,000 outdoor scene stereo image pairs of size \\(2,048 \\times 1,024\\), where 2,975 pairs are allocated for training and 1,525 for testing." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.416, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BiSIC: Bidirectional Stereo Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.146, + 0.786, + 0.237 + ], + "angle": 0, + "content": "Benchmarks. We compare our methods with traditional CODECs, such as BPG [6], H.265/HEVC [30], MV-HEVC [31], and H.266/VVC [9]. BPG encodes each view of the stereo pair independently, while H.265/HEVC and H.266/VVC compress the left and right view images jointly. In addition, we show the results of learning-based stereo image compression schemes, including HESIC+ [13], SASIC [37], BCSIC [20], LDMIC [41], and ECSIC [36]." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.238, + 0.787, + 0.357 + ], + "angle": 0, + "content": "Metrics. The quality of the reconstructed images is evaluated by the peak signal-to-noise ratio (PSNR) and MS-SSIM [35]. The rate is measured by BPP. Besides, we use the widely employed Bjøntegaard Delta bitrate (BDBR) and Bjøntegaard Delta PSNR (BD-PSNR) [8] to compare the rate-distortion (RD) performance among the methods. Specifically, BDBR measures the average bit saving at the same distortion level, while BD-PSNR reflects the relative PSNR gap at an equivalent bitrate. A lower value of BDBR and a higher value of BD-PSNR indicate better performance." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.358, + 0.786, + 0.464 + ], + "angle": 0, + "content": "Implementation Details. During the training process, we optimize the model using the MSE loss over 1.5 million steps. We employ the Adam optimizer [19] and set the batch size to 8. The learning rate is initialized at \\(10^{-4}\\) and is reduced by half after every 300,000 steps. The hyperparameter \\(\\lambda\\) is chosen from the set {256,512,1024,2048,3072,4096} for different bitrates. To achieve better MS-SSIM performance, we fine-tune the pre-trained models for an additional 500,000 steps. More implementation details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.485, + 0.323, + 0.499 + ], + "angle": 0, + "content": "4.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.508, + 0.786, + 0.81 + ], + "angle": 0, + "content": "RD Performance. Fig. 6 shows the RD curves of different compression methods on the InStereo2K and Cityscapes datasets. In terms of the MS-SSIM metric, our proposed method BiSIC outperforms all the other compression methods on both datasets. Besides, BiSIC achieves higher PSNR in the high BPP range compared to the baselines, while maintaining competitive PSNR values at low bitrates. The BDBR results are presented in Tab. 1, where BPG is set as the baseline. Our BiSIC achieves more than \\(40\\%\\) and \\(50\\%\\) BDBR reduction on InStereo2K and Cityscapes, respectively. Notably, compared to the state-of-the-art video codec VVC, BiSIC achieves an additional \\(12.76\\%\\) BDBR reduction for PSNR and an additional \\(30.08\\%\\) BDBR reduction for MS-SSIM on InStereo2K. Compared with unidirectional methods, BiSIC achieves significant bit savings by observing a holistic view and mutually sharing features, which helps remove redundancies in each view. Moreover, compared with the state-of-the-art bidirectional coding scheme BCSIC [20], our method achieves an additional \\(6.5\\%\\) to \\(15\\%\\) BDBR reduction. This is because the employed 3D convolution captures more correlations than the separately deployed 2D convolution used in BCSIC [20], and our cross-dimensional entropy model effectively aggregates various abundant dependencies from different dimensions, providing accurate probability estimation. In addition, the proposed fast variant also demonstrates satisfactory performance, with a marginal degradation in compression performance compared to our BiSIC." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 7 visualizes the reconstructed images of different methods on the Cityscapes dataset. For a fair comparison, all images are com" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.153, + 0.445, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.153, + 0.758, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.297, + 0.446, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.298, + 0.756, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.445, + 0.785, + 0.473 + ], + "angle": 0, + "content": "Fig. 6: Rate-distortion curves in terms of PSNR and MS-SSIM on the InStereo2K and Cityscapes datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.627 + ], + "angle": 0, + "content": "pressed to similar bitrates. We observe that the proposed method not only achieves higher PSNR for the reconstructed images at reduced bitrates but also maintains consistent PSNR across stereo views, thanks to its bidirectional architecture. In comparison, the SASIC method displays a PSNR discrepancy of 0.6487 dB between views, as it compresses the right image based on the left one. The VVC codec compresses images in a predictive manner, resulting in an even larger PSNR gap of 2.5769 dB between views. We leave additional qualitative analysis to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.44, + 0.673 + ], + "angle": 0, + "content": "4.3 Runtime Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this part, we evaluate the computational complexity of various compression algorithms. To ensure a fair comparison, each method is executed on an Intel Xeon Platinum 8370C CPU. Tab. 2 reports the average runtime on the InStereo2K test set, which includes the encoding and decoding latency. We observe that our BiSIC-Fast significantly reduces the runtime compared with the proposed BiSIC. This is because we replace the spatial auto-regressive interweaving in the entropy model with a two-fold stereo-checkerboard, which simplifies the dependency structure within the latent representations. Note that BiSIC-Fast still maintains competitive performance, as demonstrated in Tab. 1. Therefore, it achieves a better trade-off between the compression runtime and performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.416, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BiSIC: Bidirectional Stereo Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: BDBR values of different compression methods. Bold indicates best results, and underlined values are the second-best ones." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.185, + 0.784, + 0.376 + ], + "angle": 0, + "content": "
MethodInStereo2KCityscapes
PSNRMS-SSIMPSNRMS-SSIM
HEVC [30]-18.80%-12.59%-27.46%-24.36%
MV-HEVC [31]-7.69%2.14%-18.02%-17.13%
VVC [9]-35.31%-31.05%-56.25%-44.04%
HESIC+ [13]-14.96%-43.22%-23.83%-50.79%
SASIC [37]-18.40%-23.87%-21.47%-29.78%
BCSIC [20]-41.22%-54.67%-42.62%-60.72%
LDMIC [41]-41.95%-58.98%-41.92%-61.90%
ECSIC [36]-43.71%-55.65%-52.06%-64.96%
BiSIC (ours)-48.07%-61.13%-57.49%-67.98%
BiSIC-Fast (ours)-45.35%-59.36%-51.96%-65.56%
" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.386, + 0.784, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.512, + 0.784, + 0.527 + ], + "angle": 0, + "content": "Fig. 7: Qualitative comparison of image reconstruction quality across various codecs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.389, + 0.573 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.674 + ], + "angle": 0, + "content": "We conduct ablation experiments on the InStereo2K dataset to analyze the contribution of the proposed modules. Our model comprises several key components, including the 3D convolution backbone, cross-dimensional entropy model, and mutual attention block. To evaluate their impact, we replace each module with a baseline counterpart or remove it from the model, and compare the performance in terms of PSNR. The experimental results are shown in Fig. 8 and Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Effectiveness of 3D Convolution Backbone. To assess the capability of 3D convolutional layers in learning inter-view features, we replace 3D convolutional layers in codec with traditional 2D convolutional layers. The resultant model, referred to as Baseline (2D-Conv), experiences a PSNR degradation of 0.3218 dB at the same bitrate, as shown in Fig. 8 and Tab. 3. This indicates that the 3D convolutional layers are more effective at capturing the spatio-stereo information inherent in the data and reducing the redundancies between views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of Cross-Dimensional Entropy Model. To validate the performance improvement offered by our proposed cross-dimensional entropy model, we compare it with the joint auto-regressive and hyperprior entropy model [24], which we refer to as Baseline (Minnen). As shown in Fig. 8 and Tab. 3, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.341, + 0.145, + 0.661, + 0.159 + ], + "angle": 0, + "content": "Table 2: Average runtime of different methods." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.172, + 0.782, + 0.215 + ], + "angle": 0, + "content": "
CodecBPG HEVC VVC SASIC BCSIC LDMIC BiSIC (ours)BiSIC-Fast (ours)
Runtime16.17s 28.16s 190.27s 20.24s 89.44s 81.35s 167.25s22.82s
" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.236, + 0.472, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.378, + 0.531, + 0.406 + ], + "angle": 0, + "content": "Fig. 8: Rate-distortion curves of different ablation baselines on InStereo2K." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.254, + 0.787, + 0.369 + ], + "angle": 0, + "content": "
MethodBD-PSNR
BiSIC (ours)0
Baseline (2D-Conv)-0.3218 dB
Baseline (Minnen)-0.3475 dB
Baseline (Ying)-0.6554 dB
Baseline (Lei)-0.4956 dB
Baseline (w/o Atten)-0.7882 dB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.543, + 0.378, + 0.774, + 0.407 + ], + "angle": 0, + "content": "Table 3: BD-PSNR [8] results relative to the proposed BiSIC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.619 + ], + "angle": 0, + "content": "proposed cross-dimensional entropy model achieves better rate-distortion performance than Baseline (Minnen). This suggests that our model provides more accurate probability estimations, which in turn minimizes the coding overhead. Effectiveness of Mutual Attention Block. To evaluate the effectiveness of the proposed mutual attention block, we consider three baselines for comparisons, namely Baseline (Ying), Baseline (Lei), and Baseline (w/o Atten). Baseline (Ying) and Baseline (Lei) replace the mutual attention block with the stereo attention module [39] and the contextual transfer module [20], respectively. In particular, Baseline (w/o Atten) represents a variant without the attention module. As shown in Fig. 8 and Tab. 3, our proposed model outperforms all baselines by a large margin. Notably, removing the proposed mutual attention block results in a dramatic degradation, which demonstrates its significance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.36, + 0.658 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we introduce a bidirectional stereo image compression network that employs 3D convolution to encode local stereo features and propose a mutual attention block to capture global features. Besides, we design a symmetric cross-dimensional entropy model that integrates hyperprior, spatial context, channel context, and stereo dependency. Our bidirectional design mitigates the issue of imbalanced compression quality intrinsic to unidirectional methods and eliminates the requirement for sequential coding. As demonstrated in the experimental results, our BiSIC achieves greater bit savings at the same level of compression quality compared to both traditional coding standards and existing learning-based methods. Moreover, we propose a fast variant that significantly reduces the compression runtime while maintaining competitive performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.416, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BiSIC: Bidirectional Stereo Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.784, + 0.205 + ], + "angle": 0, + "content": "This work was supported by the General Research Fund (Project No. 16209622) from the Hong Kong Research Grants Council." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.226, + 0.323, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.254, + 0.784, + 0.281 + ], + "angle": 0, + "content": "1. Arena, F., Collotta, M., Pau, G., Termine, F.: An overview of augmented reality. Computers 11(2), 28 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.322 + ], + "angle": 0, + "content": "2. Balle, J., Chou, P.A., Minnen, D., Singh, S., Johnston, N., Agustsson, E., Hwang, S.J., Toderici, G.: Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing 15(2), 339-353 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.785, + 0.35 + ], + "angle": 0, + "content": "3. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. arXiv preprint arXiv:1611.01704 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.377 + ], + "angle": 0, + "content": "4. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.417 + ], + "angle": 0, + "content": "5. Bao, W., Wang, W., Xu, Y., Guo, Y., Hong, S., Zhang, X.: Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences 63, 1-11 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.754, + 0.431 + ], + "angle": 0, + "content": "6. Bellard, F.: BPG image format. Website (2014), https://bellard.org/bpg/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.785, + 0.457 + ], + "angle": 0, + "content": "7. Bishop, C.M.: Latent variable models. In: Learning in graphical models, pp. 371-403. Springer (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.458, + 0.785, + 0.484 + ], + "angle": 0, + "content": "8. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.785, + 0.539 + ], + "angle": 0, + "content": "9. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology 31(10), 3736-3764 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.54, + 0.785, + 0.594 + ], + "angle": 0, + "content": "10. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.595, + 0.785, + 0.649 + ], + "angle": 0, + "content": "11. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "12. Deng, X., Deng, Y., Yang, R., Yang, W., Timofte, R., Xu, M.: MASIC: Deep mask stereo image compression. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.691, + 0.785, + 0.745 + ], + "angle": 0, + "content": "13. Deng, X., Yang, W., Yang, R., Xu, M., Liu, E., Feng, Q., Timofte, R.: Deep homography for efficient stereo image compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1492-1501 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.746, + 0.785, + 0.799 + ], + "angle": 0, + "content": "14. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5718-5727 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14771-14780 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.254, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Hu, Y., Yang, W., Liu, J.: Coarse-to-fine hyper-prior modeling for learned image compression. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11013-11020 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.233 + ], + "angle": 0, + "content": "17. Jiang, W., Yang, J., Zhai, Y., Ning, P., Gao, F., Wang, R.: Mlic: Multi-reference entropy model for learned image compression. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7618-7627 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.274 + ], + "angle": 0, + "content": "18. Kadaikar, A., Dauphin, G., Mokraoui, A.: Joint disparity and variable size-block optimization algorithm for stereoscopic image compression. Signal Processing: Image Communication 61, 1-8 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.302 + ], + "angle": 0, + "content": "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "20. Lei, J., Liu, X., Peng, B., Jin, D., Li, W., Gu, J.: Deep stereo image compression via bi-directional coding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19669-19678 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.388 + ], + "angle": 0, + "content": "21. Liu, J., Wang, S., Urtasun, R.: DSIC: Deep stereo image compression. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.785, + 0.431 + ], + "angle": 0, + "content": "22. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14388-14397 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.472 + ], + "angle": 0, + "content": "23. Liu, Y., Ren, J., Zhang, J., Liu, J., Lin, M.: Visually imbalanced stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2029-2038 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "24. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.558 + ], + "angle": 0, + "content": "25. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3339-3343. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.6 + ], + "angle": 0, + "content": "26. Pan, G., Lu, G., Hu, Z., Xu, D.: Content adaptive latents and decoder for neural image compression. In: European Conference on Computer Vision. pp. 556-573. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "27. Shen, Z., Zhang, M., Zhao, H., Yi, S., Li, H.: Efficient attention: Attention with linear complexities. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 3531-3539 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.671 + ], + "angle": 0, + "content": "28. Skodras, A., Christopoulos, C., Ebrahimi, T.: The jpeg 2000 still image compression standard. IEEE Signal processing magazine 18(5), 36-58 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.713 + ], + "angle": 0, + "content": "29. Song, X., Zhao, X., Fang, L., Hu, H., Yu, Y.: Edge stereo: An effective multi-task learning network for stereo matching and edge detection. International Journal of Computer Vision 128, 910-930 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "30. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (HEVC) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "31. Tech, G., Chen, Y., Müller, K., Ohm, J.R., Vetro, A., Wang, Y.K.: Overview of the multiview and 3d extensions of high efficiency video coding. IEEE Transactions on Circuits and Systems for Video Technology 26(1), 35-49 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.416, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BiSIC: Bidirectional Stereo Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Vetro, A., Wiegand, T., Sullivan, G.J.: Overview of the stereo and multiview video coding extensions of the h. 264/mpeg-4 avc standard. Proceedings of the IEEE 99(4), 626-642 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.217 + ], + "angle": 0, + "content": "34. Wallace, G.K.: The JPEG still picture compression standard. Communications of the ACM 34(4), 30-44 (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "35. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402. IEEE (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "36. Wödlinger, M., Kotera, J., Keglevic, M., Xu, J., Sablatnig, R.: ECSIC: Epipolar cross attention for stereo image compression. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3436-3445 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.355 + ], + "angle": 0, + "content": "37. Wödlinger, M., Kotera, J., Xu, J., Sablatnig, R.: SASIC: Stereo image compression with latent shifts and stereo attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 661-670 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "38. Yin, H., Wang, Y., Tang, L., Ding, X., Huang, S., Xiong, R.: 3d lidar map compression for efficient localization on resource constrained vehicles. IEEE transactions on intelligent transportation systems 22(2), 837-852 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "39. Ying, X., Wang, Y., Wang, L., Sheng, W., An, W., Guo, Y.: A stereo attention module for stereo image super-resolution. IEEE Signal Processing Letters 27, 496-500 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.481 + ], + "angle": 0, + "content": "40. Zhai, Y., Tang, L., Ma, Y., Peng, R., Wang, R.: Disparity-based stereo image compression with aligned cross-view priors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2351-2360 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.482, + 0.787, + 0.508 + ], + "angle": 0, + "content": "41. Zhang, X., Shao, J., Zhang, J.: LDMIC: Learning-based distributed multi-view image coding. arXiv preprint arXiv:2301.09799 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.535 + ], + "angle": 0, + "content": "42. Zhu, Y., Yang, Y., Cohen, T.: Transformer-based transform coding. In: International Conference on Learning Representations (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.535 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/e50db183-7c10-4237-a52b-c78459d9ff33_origin.pdf b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/e50db183-7c10-4237-a52b-c78459d9ff33_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d0df8c179bbc37b685104d3ae95c6b8edeabc41b --- /dev/null +++ b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/e50db183-7c10-4237-a52b-c78459d9ff33_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bae4fd33f77f7cdcf905ba0cdee2cecf6b931ead2eb3758f1f490dcf45d0ead +size 1574894 diff --git a/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/full.md b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d7d16b65bc918edaa4c14f17f7694e959633f4b --- /dev/null +++ b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/full.md @@ -0,0 +1,288 @@ +# Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model + +Zhening Liu, Xinjie Zhang, Jiawei Shao, Zehong Lin, and Jun Zhang + +Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong SAR {zhening.liu,xzhangga,jiawei.shao}@connect.ust.hk, {eezhlin,eejzhang}@ust.hk + +Abstract. With the rapid advancement of stereo vision technologies, stereo image compression has emerged as a crucial field that continues to draw significant attention. Previous approaches have primarily employed a unidirectional paradigm, where the compression of one view is dependent on the other, resulting in imbalanced compression. To address this issue, we introduce a symmetric bidirectional stereo image compression architecture, named BiSIC. Specifically, we propose a 3D convolution based codec backbone to capture local features and incorporate bidirectional attention blocks to exploit global features. Moreover, we design a novel cross-dimensional entropy model that integrates various conditioning factors, including the spatial context, channel context, and stereo dependency, to effectively estimate the distribution of latent representations for entropy coding. Extensive experiments demonstrate that our proposed BiSIC outperforms conventional image/video compression standards, as well as state-of-the-art learning-based methods, in terms of both PSNR and MS-SSIM. + +Keywords: Stereo Image Compression $\cdot$ Bidirectional Architecture $\cdot$ Cross-Dimensional Entropy Model + +# 1 Introduction + +Stereo vision, which mimics human binocular vision, has emerged as an important area in computer vision. It enables a wide range of applications, such as 3D movies, remote sensing [29], autonomous driving [38], and augmented/virtual reality (AR/VR) [1]. With the rapid development of stereo cameras, stereo images have become essential for providing immersive and convincing visual presentations that enhance user experience and benefit downstream applications. Meanwhile, the proliferation of stereo cameras leads to a significant surge in the volume of stereo images, resulting in substantial storage costs and posing challenges to transmission efficiency. Consequently, stereo image compression plays a pivotal role in ensuring the practicality and effectiveness of stereo vision. + +In recent years, numerous methods have been developed for single image compression, including traditional CODECs [6, 34] and learning-based methods [2-4, 10, 24]. These methods, however, focus on extracting individual features for compression and fail to capture the cross-view correlation. Consequently, they are less effective for stereo image compression that requires the extraction of both intra-view information and inter-view information. To exploit the inherent shared information in stereo images, traditional methods for stereo image coding follow a predictive compression procedure of video standards [9, 31]. In this process, one view is compressed first and then used to predict the other view. Subsequently, the disparity between the predicted and actual views is compressed. Nonetheless, these methods typically rely on hand-crafted modules and lack end-to-end optimization, resulting in suboptimal performance. + +Inspired by the success of learning-based single image compression, researchers have proposed novel stereo image compression methods using deep neural networks (DNNs) to enhance compression performance [12, 13, 20, 21, 36, 37, 40, 41]. Initially, extensive research focuses on unidirectional learned stereo image compression methods [12, 13, 21, 36, 37, 40], where one view is compressed and used as a reference to compress the other view. Despite achieving high overall rate-distortion (RD) performance, these unidirectional learned compression methods yield a noticeable imbalance in compression quality between stereo views, similar to traditional predictive compression methods. This imbalance is unfavorable for human vision and undermines downstream machine vision tasks [23]. Moreover, some of these methods [12, 13, 37, 40] heavily rely on sequential coding order and accurate estimation of the explicit warping relationship between views. To avoid these issues, a recent study [20] proposes a bidirectional compression method that leverages 2D convolution to extract individual features and facilitate information sharing between views. Nonetheless, the compression performance is limited due to the separate processing of convolutions and the plain spatial context based entropy model. + +Based on the above discussion, we propose a novel bidirectional method for stereo image compression, named BiSIC, to address the aforementioned issues. BiSIC achieves improvements in four key aspects: encoder/decoder structure, mutual stereo learning, entropy model, and coding speed. Firstly, regarding the codec structure, although 2D convolutions have proven effective in various computer vision tasks, they fail to capture the aligned morphs and features between stereo views. To overcome this limitation, we propose using 3D convolution as the backbone for the encoder/decoder, which takes both views as input to capture the inter-view dependencies that are inaccessible in previous works based on 2D convolution. Secondly, for mutual stereo learning, we propose a mutual attention block to facilitate the transfer of features between views. This module enhances the integration and cooperation of the stereo views, leading to improved compression performance. Thirdly, for the entropy model, prior methods primarily use hyperprior and spatial context as conditions, overlooking the rich inter-view dependencies across other dimensions. To exploit these dependencies, + +![](images/82a897adf9318952b36b7d0d6d6f253c71eab74f67470215fcd560c7c0138048.jpg) +Fig. 1: The proposed bidirectional stereo image compression architecture. AE and AD are arithmetic coder for entropy coding. The backbones of encoder $\mathbf{E}$ , decoder $\mathbf{D}$ , hyper encoder $h_a$ , and hyper decoder $h_s$ are constructed with 3D convolution to model local features. 3DCovTr denotes transposed 3D convolution. To enhance global feature extraction, bidirectional mutual attention blocks are inserted between 3D convolutional layers. Moreover, a novel cross-dimensional entropy model is utilized to capture complex inter-view dependencies. + +we construct a cross-dimensional entropy model that aggregates hyperprior, spatial context, channel context, and stereo dependency. + +Specifically, we introduce a masked 3D convolution to capture the stereo spatial context and utilize an attention block to process the stereo channel context. Finally, to improve coding speed and reduce auto-regressive steps, we design a fast variant based on a stereo-checkerboard structure. This design converts auto-regressive inference into a two-fold operation, improving the coding efficiency. The contributions of this paper are summarized as follows: + +- We propose a novel 3D convolution based codec backbone that processes stereo images as a whole, enabling the learning of local stereo features and effective capture of shared patterns across views. +- We design a bidirectional mutual attention block that facilitates the exchange of global features between stereo views. This block produces compact attention maps and is robust to input images of different sizes. +- We develop a powerful symmetric cross-dimensional entropy model, which incorporates hyperpriors, spatial context, channel context, and stereo dependency as conditioning factors, to enhance the probability estimation. +- To accelerate coding, we propose a stereo-checkerboard design that transforms the spatial auto-regressive inference into a two-fold process. This design strikes a balance between time consumption and compression quality. + +# 2 Related Works + +# 2.1 Single Image Compression + +Image compression is an important topic that continuously attracts significant research interest. With the increasing demand for vision products and applica + +tions, steady progress has been made in traditional image compression standards, evolving from JPEG [28, 34] and BPG [6] to video compression standards such as HEVC [30] and VVC [9]. These traditional methods establish a standard compression pipeline comprising transform, quantization, and entropy coding. Over the past five years, deep learning has revolutionized various vision applications. As demonstrated in the seminal work [3], DNNs have the potential to replace components in traditional compression methods and enable end-to-end compression. This work introduces uniform noise to mimic the distortion caused by quantization, thereby paving the way for end-to-end optimization in image compression. + +In learned image compression models [3,4,10,14-17,22,24-26,42], neural networks are employed to parameterize the probability distribution of transformed latents. Accurate estimation of this distribution is crucial for achieving compact compression via entropy coding. To enhance probability estimation, various conditions have been integrated into entropy models [2], including hyperpriors [4], spatial context [24], and channel-wise context [25], to capture a broader range of dependencies. To be specific, hyperprior is a lightweight subsidiarity information transmitted alongside the encoded latents. Both spatial context and channel-wise context leverage auto-regressive inference, which is conditioned on the preceding segments of data by iteratively interleaving the inference. Checkerboard methods [15], on the other hand, avoid the spatial auto-regressive process and accelerate the coding speed by constraining the spatial context to the near neighbors of each pixel instead of spatial causal components. These dependencies operate on diverse scales and can be progressively combined to further improve effectiveness, as demonstrated in [14, 17]. + +# 2.2 Stereo Image Compression + +Stereo image compression is an extension of single image compression that considers capturing both intra-view features and inter-view correlations. Traditional approaches employ key-frame based predictive coding, akin to video coding [18, 33]. However, these methods rely on hand-crafted feature extractors and classical optimization techniques. As these modules are optimized individually, they often fail to achieve satisfactory RD performance. Moreover, the predictive compression pipeline also leads to imbalanced compression. + +In recent years, learned stereo image compression methods have gained increasing attention [12, 13, 20, 21, 36, 37, 40, 41]. These methods can be broadly categorized into unidirectional and bidirectional approaches. Unidirectional methods first compress one view image and then use it as a reference to compress the other view, significantly reducing the required bits. The pioneering work [21] introduces a skip module to warp the features from one view to encode the other. Subsequent works have utilized various techniques to represent the warping relationship between stereo views, such as homography matrices [12, 13], disparity maps [40], and horizontal shifts [37]. + +These unidirectional methods heavily rely on the precise estimation of the spatial transformations. A very recent work, ECSIC [36], avoids explicit para + +metric transformation in favor of constructing a unidirectional entropy model that conditions the right image on the known left image. However, the decompressed results of these unidirectional methods often have imbalanced quality due to the one-way dependency between stereo views. On the other hand, bidirectional methods, while remaining less explored, produce more balanced quality by extracting shared information from both views and compressing stereo images simultaneously. One such approach [20] introduces a contextual transform module for feature transfer between views. However, it employs separate down-sampling networks and only takes spatial context into account, which limits the compression performance. In contrast, to fully exploit the shared information and establish stronger dependencies, we employ a joint downsampling codec and a cross-dimensional entropy model. + +# 3 Proposed Methods + +The proposed architecture for stereo image compression, as depicted in Fig. 1, consists of two main components: a joint codec and an entropy model. The codec employs 3D convolutions to transform the original stereo images into compact latent representations, effectively exploiting the correlation between the two views. Subsequently, the entropy model estimates the probability distribution of these latent representations, enabling the generation of a compact bitstream through entropy coding. In the following subsections, we elaborate on the codec and the entropy model. + +# 3.1 Joint Codec + +3D Convolution Backbone. The codec acts as a non-linear transformation that downsamples the raw images into compact latent representations and generates hyperpriors for the entropy model. Different from previous approaches to stereo image compression that employ 2D convolutional layers, we adopt 3D convolutions as the backbone of our codec. Unlike 2D convolution, which operates separately on each view, 3D convolution concurrently processes both stereo images. Consequently, it inherently possesses the capability to extract inter-view correlations. + +Our encoder, denoted by $\mathbf{E}$ , comprises four 3D convolutional layers, with bidirectional mutual attention blocks inserted after the second and the fourth layers to facilitate global feature communication between the two views. The encoder transforms a pair of stereo images, i.e., the left and right images denoted by $\pmb{x}_l$ , $\pmb{x}_r \in \mathbb{R}^{B \times 3 \times H \times W}$ , into compact latent representations $\pmb{y}_l$ , $\pmb{y}_r \in \mathbb{R}^{B \times N \times \frac{H}{16} \times \frac{W}{16}}$ and quantizes them into $\hat{\pmb{y}} := \{\hat{\pmb{y}}_l, \hat{\pmb{y}}_r\}$ for efficient transmission. Conversely, the decoder $\mathbf{D}$ transforms the quantized latents $\hat{\pmb{y}}$ back into images $\hat{\pmb{x}}_l$ and $\hat{\pmb{x}}_r$ . The encoding and decoding processes are expressed as: + +$$ +\boldsymbol {y} _ {l}, \boldsymbol {y} _ {r} = \mathbf {E} \left(\boldsymbol {x} _ {l}, \boldsymbol {x} _ {r}\right), \tag {1} +$$ + +$$ +\hat {\boldsymbol {x}} _ {l}, \hat {\boldsymbol {x}} _ {r} = \mathbf {D} (\hat {\boldsymbol {y}} _ {l}, \hat {\boldsymbol {y}} _ {r}). \tag {2} +$$ + +To capture the dependencies among elements in the latents $\hat{\pmb{y}}$ for efficient entropy coding [4,7], hyperpriors are generated through another 3D convolution based hyper encoder $h_a$ . The generated hyperpriors $\pmb{z}_l, \pmb{z}_r \in \mathbb{R}^{B \times M \times \frac{H}{64} \times \frac{W}{64}}$ are quantized into $\hat{\pmb{z}}_l, \hat{\pmb{z}}_r$ and then transmitted alongside the quantized latents $\hat{\pmb{y}}$ . At the decoder side, a 3D convolution based hyper decoder $h_s$ upscales the quantized hyperpriors $\hat{\pmb{z}}_l, \hat{\pmb{z}}_r$ into $\tilde{\pmb{z}}_l, \tilde{\pmb{z}}_r$ , which are subsequently utilized in the entropy model, as elaborated in Sec. 3.2. The processes of the hyper encoder and hyper decoder are expressed as: + +$$ +\boldsymbol {z} _ {l}, \boldsymbol {z} _ {r} = h _ {a} (\boldsymbol {y} _ {l}, \boldsymbol {y} _ {r}), \tag {3} +$$ + +$$ +\tilde {\boldsymbol {z}} _ {l}, \tilde {\boldsymbol {z}} _ {r} = h _ {s} (\hat {\boldsymbol {z}} _ {l}, \hat {\boldsymbol {z}} _ {r}). \tag {4} +$$ + +Bidirectional Mutual Attention Block. Convolutional layers have an excellent ability for local modeling, but their ability for long-range feature modeling is relatively limited. On the other hand, attention modules [32] excel at extracting long-range features. Therefore, we propose combining the attention mechanism with convolutional layers to compensate for the shortcomings of the latter and leverage the advantages of both convolution and attention. To this end, we design a novel bidirectional mutual attention block, as depicted in Fig. 2, to be inserted between convolutional layers as the pink blocks shown in Fig. 1. + +As illustrated in Fig. 2, the bidirectional mutual attention block contains two cross-attention stages, each followed by a self-attention layer. The two cross-attention stages differ in how they concentrate on the feature map and are referred to as cross-key attention and cross-query attention. Before each cross-attention stage, the input stereo data $\pmb{y}_l, \pmb{y}_r$ first pass through a residual block, which are then embedded into low-dimensional query, key, and value tensors $\pmb{Q}, \pmb{K}, \pmb{V} \in \mathbb{R}^{B \times C \times (H \times W)}$ . To avoid the prohibitive computation of classical attention, which yields an $(H \times W) \times (H \times W)$ sized attention map, we adopt the efficient attention approach proposed in [27]. This approach computes the product of the key and the transposed value, thereby reducing the attention map size to $C_K \times C_V$ . + +In the cross-key attention stage, for each view, we generate an attention map using its key and the value from the other view, which is then queried by the current view. By doing so, we can extract the cross-key feature $\Phi$ from the stereo views as: + +$$ +\boldsymbol {\Phi} _ {r \rightarrow l} = (\sigma (\boldsymbol {K} _ {r}) \times \boldsymbol {V} _ {l} ^ {T}) ^ {T} \times \sigma (\boldsymbol {Q} _ {r}), \quad \boldsymbol {\Phi} _ {l \rightarrow r} = (\sigma (\boldsymbol {K} _ {l}) \times \boldsymbol {V} _ {r} ^ {T}) ^ {T} \times \sigma (\boldsymbol {Q} _ {l}), \tag {5} +$$ + +where $\sigma$ is the softmax function. This approach calculates the attention map using the key and value from different views, thereby facilitating inter-view alignment and the identification of common patterns. When a feature in one view is correlated with the corresponding feature in the other view, it is effectively captured through the attention map. + +In the cross-query attention stage, the attention map is obtained using the key and value from a single view, and then queried by the other view, yielding the cross-query feature $\Psi$ as: + +$$ +\Psi_ {r \rightarrow l} = \left(\sigma \left(\boldsymbol {K} _ {l}\right) \times \boldsymbol {V} _ {l} ^ {T}\right) ^ {T} \times \sigma \left(\boldsymbol {Q} _ {r}\right), \Psi_ {l \rightarrow r} = \left(\sigma \left(\boldsymbol {K} _ {r}\right) \times \boldsymbol {V} _ {r} ^ {T}\right) ^ {T} \times \sigma \left(\boldsymbol {Q} _ {l}\right). \tag {6} +$$ + +![](images/b30f87dee6598cab8a1b245a633985d65c548e00d77420fae0ed4730f3075ace.jpg) +Fig. 2: Overview of the proposed bidirectional mutual attention block. The blue and green lines represent features extracted from the left view and right view, respectively. The network structures of the residual block, the basic efficient attention unit, and the combine block are illustrated on the right, from top to bottom. + +The attention map in this stage is generated within one view, thus maintaining an inner-focused relationship. This process preserves the individual self-features of each view, while the other view serves as a reference for determining the allocation of attention weights. + +After obtaining the cross-key attention and cross-query attention features $\Phi, \Psi$ , we employ self-attention to augment these features as $\tilde{\Phi}, \tilde{\Psi}$ and use a shared-parameter combine block to get the desired mutual attention output, as shown in the bottom right of Fig. 2. Specifically, the combine block concatenates the augmented features $\tilde{\Phi}, \tilde{\Psi}$ and the residual component, and shrinks the channel number to be consistent with the block input. + +This bidirectional mutual attention block captures diverse features and provides a global receptive field for the codec. It is suitable for any stereo information transfer scenarios, provided that the inputs are two features of equal size, making it a generalized plug-and-play block. Moreover, the size of the attention maps is determined only by the embedding channel $C$ and does not vary with the size of input images. + +# 3.2 Cross-Dimensional Entropy Model + +The proposed codec effectively transforms stereo images into small-sized latent representations. To achieve efficient compression, the next crucial step is to encode these latents into compact bitstreams using accurate probability distribution and entropy coding [4,24]. Thus, an entropy model is critical for parameterizing and estimating probabilities. In learned image compression, the probability distribution is commonly modeled as Gaussian distributions [24], with the mean $\mu$ and variance $\sigma^2$ being estimated. To achieve accurate estimation, it is critical to provide appropriate references. Apart from hyperpriors, previous stereo image compression methods primarily use spatial context [13,20,41] and unidirectional stereo dependency as references [21,36], while abundant features on the + +![](images/fe612b9e57ce8faa802be1658b78292465b36d3069593a0608d31bef63cde874.jpg) +Fig. 3: Illustration of the proposed symmetric cross-dimensional entropy model. It jointly aggregates the hyperprior (blue), stereo spatial context (red), and stereo channel context (yellow) as conditions for an effective probability distribution estimation. The hyper decoder $h_s$ for hyperprior is shown in Fig. 1. The masked 3D convolution for spatial context is provided in Fig. 4a. Moreover, the mutual attention block for channel context is detailed in Fig. 2. + +channel dimension are overlooked. In this work, we propose a symmetric cross-dimensional entropy model that integrates hyperprior, spatial context, channel context, and stereo dependency in a fully symmetric bidirectional way. + +The architecture of the proposed cross-dimensional entropy model is depicted in Fig. 3. We use $\hat{\pmb{y}}_{l,i}^{k}$ to denote the $i$ -th $(i = 1,2,\dots,n)$ entry of the $k$ -th $(k = 1,2,\dots,K)$ channel slice from the received left view latent $\hat{\pmb{y}}_l$ , and likewise for the others. The cross-dimensional condition formulation is expressed as: + +$$ +p _ {\hat {\boldsymbol {y}} _ {l}} \left(\hat {\boldsymbol {y}} _ {l, i} ^ {k} \mid \hat {\boldsymbol {z}} _ {l}, \hat {\boldsymbol {y}} _ {l} ^ {< k}, \hat {\boldsymbol {y}} _ {l, < i} ^ {k}, \hat {\boldsymbol {y}} _ {r} ^ {< k}, \hat {\boldsymbol {y}} _ {r, < i} ^ {k}\right) \sim \mathcal {N} \left(\boldsymbol {\mu} _ {l}, \boldsymbol {\sigma} _ {l} ^ {2}\right), \tag {7} +$$ + +$$ +p _ {\hat {\boldsymbol {y}} _ {r}} \left(\hat {\boldsymbol {y}} _ {r, i} ^ {k} \mid \hat {\boldsymbol {z}} _ {r}, \hat {\boldsymbol {y}} _ {r} ^ {< k}, \hat {\boldsymbol {y}} _ {r, < i} ^ {k}, \hat {\boldsymbol {y}} _ {l} ^ {< k}, \hat {\boldsymbol {y}} _ {l, < i} ^ {k}\right) \sim \mathcal {N} \left(\boldsymbol {\mu} _ {r}, \boldsymbol {\sigma} _ {r} ^ {2}\right). \tag {8} +$$ + +For the hyperprior component, we generate the dependency $\tilde{z}_l,\tilde{z}_r$ from the hyper decoder using Eq. (4). To obtain the channel context, we evenly split the stereo latents on the channel dimension into $K$ slices. These latent slices share a great similarity, providing abundant conditions [25]. To exploit this similarity, the probability estimation is performed slice by slice, using the previous slices as references for the following one. In addition, the decoded slices of the other view also serve as references. Therefore, we propose to use the mutual attention block introduced in the previous subsection to aggregate the channel-wise features $\Theta_l,\Theta_r$ from both views, conditioning on $\hat{\pmb{y}}_l^{< k}$ and $\hat{\pmb{y}}_r^{< k}$ , as depicted by the yellow lines in Fig. 3. + +To capture the spatial context within each channel slice, we consistently employ 3D convolution to jointly learn spatial features from the causal areas $\hat{\pmb{y}}_{< i}^{k}$ . Specifically, during the coding process, the current compression/decompression only accesses the areas that have already been encoded/decoded. To maintain the causality of the compression, we mask out the uncoded part, which is referred to as masked 3D convolution, as shown in Fig. 4a. In this approach, for the $i$ -th entry $\hat{\pmb{y}}_i^k$ , the weights of uncoded part are set to zero, and we obtain spatial + +![](images/c798248cfad0cab460d4c8cbcdc3c349835e20719e14546d2fbe16c5977ce264.jpg) +(a) Masked 3D convolution. + +![](images/0f40eb1b7e984197b69af06fb341ef06dee00dd3644a7b311ffca5b4b53b5f31.jpg) +(b) Auto-regressive pattern in BiSIC entropy model. + +![](images/8ac0cdfffcea789de3a4319aec3819d5335fc413e5ea0e637255fc9ef3c32c68.jpg) +(c) Stereo-checkerboard pattern in BiSIC-Fast. +Fig. 4: (Left) Illustration of masked 3D convolution. The blue region is used as the condition for the probability estimation of the current target depicted in red. The weights of the convolution kernel in blue are valid, while the white region is masked to maintain the causal spatial context. (Middle) Demonstration of the auto-regressive process in the cross-dimensional entropy model. The yellow and red arrows represent the channel context and spatial context, respectively. The spatial context is processed entry by entry. (Right) Demonstration of the stereo-checkerboard in the fast variant, where the blue and red parts are each processed only once. + +context features $\mathbf{\Upsilon}_{l,i},\mathbf{\Upsilon}_{r,i}$ by considering the causal context $\hat{\pmb{y}}_{l, < i}^{k}$ and $\hat{\pmb{y}}_{r, < i}^{k}$ from both views simultaneously, as depicted by the red lines in Fig. 3. + +Upon extracting various dependency features, we employ an aggregation network $\mathbf{G}_{\mathrm{ag}}$ to fuse the conditions and obtain the probability estimations as: + +$$ +\boldsymbol {\mu} _ {l}, \boldsymbol {\sigma} _ {l} ^ {2} = \mathbf {G} _ {\mathrm {a g}} \left(\tilde {\boldsymbol {z}} _ {l}, \boldsymbol {\Theta} _ {l}, \boldsymbol {\Upsilon} _ {l}\right), \boldsymbol {\mu} _ {r}, \boldsymbol {\sigma} _ {r} ^ {2} = \mathbf {G} _ {\mathrm {a g}} \left(\tilde {\boldsymbol {z}} _ {r}, \boldsymbol {\Theta} _ {r}, \boldsymbol {\Upsilon} _ {r}\right). \tag {9} +$$ + +The entire procedure in the entropy model is designed to be fully symmetric and bidirectional, which aligns with our motivation for a bidirectional design. + +# 3.3 Loss Function + +The proposed stereo image compression network is trained using the rate-distortion loss. The rate is defined by bit-per-pixel (BPP), which includes bits for transmitting latents $\hat{\pmb{y}}_l, \hat{\pmb{y}}_r$ and hyperprioris $\hat{\pmb{z}}_l, \hat{\pmb{z}}_r$ . The distortion is measured using the mean squared error (MSE) and the multi-scale structural similarity index measure (MS-SSIM) [35]. The overall loss function is given by: + +$$ +\mathcal {L} := \lambda D + R = \lambda \sum_ {l, r} d (\boldsymbol {x}, \hat {\boldsymbol {x}}) + \sum_ {l, r} (r (\hat {\boldsymbol {y}}) + r (\hat {\boldsymbol {z}})), \tag {10} +$$ + +where $d(\cdot)$ denotes the distortion function with MSE or MS-SSIM, $r(\cdot)$ calculates the bitrate using the estimated probability of $\hat{\pmb{y}}$ and $\hat{\pmb{z}}$ [4], and $\lambda$ is a hyperparameter that balances the rate and distortion. + +# 3.4 Fast Variant + +The proposed entropy model effectively captures multiple references and estimates the probability distribution. However, the auto-regressive iterations in + +![](images/986a5221979dcd8eaae64a0f629b81edad2c86a61503593047d1f0be819543ea.jpg) +Fig. 5: Illustration of the entropy model for the proposed fast variant based on stereo-checkerboard. The blue and red parts represent the stereo anchor and stereo non-anchor, respectively. By concurrently processing two views, the 3D convolution based anchor context network effectively captures the dependency from the stereo anchor. + +the spatial context are time-consuming. To strike a better balance between compression runtime and performance, we need to reduce the auto-regressive executions while maintaining a strong condition for the entropy model. To this end, we propose a fast variant of our model, named BiSIC-Fast, by extending the checkerboard structure [15] into stereo-checkerboard and utilizing 3D convolutions to jointly model the spatial context of two views with only two iterations. + +This fast variant maintains the codec as BiSIC, while the entropy model network is modified as illustrated in Fig. 5. In this approach, we split the latents $\pmb{y}$ into two parts, stereo anchor part and stereo non-anchor part, as represented in blue and red in Fig. 5, respectively. The entropy parameter estimation for the stereo anchor part relies on the hyperprior condition and channel context. The stereo anchor is decompressed first, and then the stereo non-anchor utilizes the decompressed stereo anchor as a dependency. Consequently, the entropy parameter estimation for the stereo non-anchor part is based on the hyperprior, channel context, and stereo anchor context. + +By adopting this approach, we avoid the spatial auto-regressive process in favor of a two-fold operation, as demonstrated in Fig. 4b and Fig. 4c. Although the spatial auto-regressive process is omitted to save runtime, the neighbors of each pixel in the latents are still maintained, preserving a good perceptive field. + +# 4 Experiments + +# 4.1 Experimental Implementation + +Datasets. We evaluate the performance of our methods on two benchmark datasets: InStereo2K [5] and Cityscapes [11]. InStereo2K contains 2,060 indoor scene stereo image pairs of size $1,080 \times 860$ . It is split into 2,010 image pairs for training and 50 for testing. Cityscapes contains 5,000 outdoor scene stereo image pairs of size $2,048 \times 1,024$ , where 2,975 pairs are allocated for training and 1,525 for testing. + +Benchmarks. We compare our methods with traditional CODECs, such as BPG [6], H.265/HEVC [30], MV-HEVC [31], and H.266/VVC [9]. BPG encodes each view of the stereo pair independently, while H.265/HEVC and H.266/VVC compress the left and right view images jointly. In addition, we show the results of learning-based stereo image compression schemes, including HESIC+ [13], SASIC [37], BCSIC [20], LDMIC [41], and ECSIC [36]. + +Metrics. The quality of the reconstructed images is evaluated by the peak signal-to-noise ratio (PSNR) and MS-SSIM [35]. The rate is measured by BPP. Besides, we use the widely employed Bjøntegaard Delta bitrate (BDBR) and Bjøntegaard Delta PSNR (BD-PSNR) [8] to compare the rate-distortion (RD) performance among the methods. Specifically, BDBR measures the average bit saving at the same distortion level, while BD-PSNR reflects the relative PSNR gap at an equivalent bitrate. A lower value of BDBR and a higher value of BD-PSNR indicate better performance. + +Implementation Details. During the training process, we optimize the model using the MSE loss over 1.5 million steps. We employ the Adam optimizer [19] and set the batch size to 8. The learning rate is initialized at $10^{-4}$ and is reduced by half after every 300,000 steps. The hyperparameter $\lambda$ is chosen from the set {256,512,1024,2048,3072,4096} for different bitrates. To achieve better MS-SSIM performance, we fine-tune the pre-trained models for an additional 500,000 steps. More implementation details are provided in the supplementary material. + +# 4.2 Results + +RD Performance. Fig. 6 shows the RD curves of different compression methods on the InStereo2K and Cityscapes datasets. In terms of the MS-SSIM metric, our proposed method BiSIC outperforms all the other compression methods on both datasets. Besides, BiSIC achieves higher PSNR in the high BPP range compared to the baselines, while maintaining competitive PSNR values at low bitrates. The BDBR results are presented in Tab. 1, where BPG is set as the baseline. Our BiSIC achieves more than $40\%$ and $50\%$ BDBR reduction on InStereo2K and Cityscapes, respectively. Notably, compared to the state-of-the-art video codec VVC, BiSIC achieves an additional $12.76\%$ BDBR reduction for PSNR and an additional $30.08\%$ BDBR reduction for MS-SSIM on InStereo2K. Compared with unidirectional methods, BiSIC achieves significant bit savings by observing a holistic view and mutually sharing features, which helps remove redundancies in each view. Moreover, compared with the state-of-the-art bidirectional coding scheme BCSIC [20], our method achieves an additional $6.5\%$ to $15\%$ BDBR reduction. This is because the employed 3D convolution captures more correlations than the separately deployed 2D convolution used in BCSIC [20], and our cross-dimensional entropy model effectively aggregates various abundant dependencies from different dimensions, providing accurate probability estimation. In addition, the proposed fast variant also demonstrates satisfactory performance, with a marginal degradation in compression performance compared to our BiSIC. + +Qualitative Results. Fig. 7 visualizes the reconstructed images of different methods on the Cityscapes dataset. For a fair comparison, all images are com + +![](images/8ef643ad96c08df9f0355be8c18e1e930c4a0c876452815bcdfc02a1c883b28f.jpg) + +![](images/390ec9a2c255c81e5da490d92ce89cbb58f592afdac005fe955c24d3ec41f720.jpg) + +![](images/75f78f98a40d55e15127b5cc088019f430362067f9f71edfc3a0b8b1363cf6ce.jpg) +Fig. 6: Rate-distortion curves in terms of PSNR and MS-SSIM on the InStereo2K and Cityscapes datasets. + +![](images/34ce93292296d44c365028774ed7213c04926af87d59eb6adb2835a68e58ce4c.jpg) + +pressed to similar bitrates. We observe that the proposed method not only achieves higher PSNR for the reconstructed images at reduced bitrates but also maintains consistent PSNR across stereo views, thanks to its bidirectional architecture. In comparison, the SASIC method displays a PSNR discrepancy of 0.6487 dB between views, as it compresses the right image based on the left one. The VVC codec compresses images in a predictive manner, resulting in an even larger PSNR gap of 2.5769 dB between views. We leave additional qualitative analysis to the supplementary material. + +# 4.3 Runtime Comparison + +In this part, we evaluate the computational complexity of various compression algorithms. To ensure a fair comparison, each method is executed on an Intel Xeon Platinum 8370C CPU. Tab. 2 reports the average runtime on the InStereo2K test set, which includes the encoding and decoding latency. We observe that our BiSIC-Fast significantly reduces the runtime compared with the proposed BiSIC. This is because we replace the spatial auto-regressive interweaving in the entropy model with a two-fold stereo-checkerboard, which simplifies the dependency structure within the latent representations. Note that BiSIC-Fast still maintains competitive performance, as demonstrated in Tab. 1. Therefore, it achieves a better trade-off between the compression runtime and performance. + +Table 1: BDBR values of different compression methods. Bold indicates best results, and underlined values are the second-best ones. + +
MethodInStereo2KCityscapes
PSNRMS-SSIMPSNRMS-SSIM
HEVC [30]-18.80%-12.59%-27.46%-24.36%
MV-HEVC [31]-7.69%2.14%-18.02%-17.13%
VVC [9]-35.31%-31.05%-56.25%-44.04%
HESIC+ [13]-14.96%-43.22%-23.83%-50.79%
SASIC [37]-18.40%-23.87%-21.47%-29.78%
BCSIC [20]-41.22%-54.67%-42.62%-60.72%
LDMIC [41]-41.95%-58.98%-41.92%-61.90%
ECSIC [36]-43.71%-55.65%-52.06%-64.96%
BiSIC (ours)-48.07%-61.13%-57.49%-67.98%
BiSIC-Fast (ours)-45.35%-59.36%-51.96%-65.56%
+ +![](images/e580b55551af116816dc0b6065fc193a99fd5a3bf27e0326ef507b11aaf4c7d7.jpg) +Fig. 7: Qualitative comparison of image reconstruction quality across various codecs. + +# 4.4 Ablation Study + +We conduct ablation experiments on the InStereo2K dataset to analyze the contribution of the proposed modules. Our model comprises several key components, including the 3D convolution backbone, cross-dimensional entropy model, and mutual attention block. To evaluate their impact, we replace each module with a baseline counterpart or remove it from the model, and compare the performance in terms of PSNR. The experimental results are shown in Fig. 8 and Tab. 3. + +Effectiveness of 3D Convolution Backbone. To assess the capability of 3D convolutional layers in learning inter-view features, we replace 3D convolutional layers in codec with traditional 2D convolutional layers. The resultant model, referred to as Baseline (2D-Conv), experiences a PSNR degradation of 0.3218 dB at the same bitrate, as shown in Fig. 8 and Tab. 3. This indicates that the 3D convolutional layers are more effective at capturing the spatio-stereo information inherent in the data and reducing the redundancies between views. + +Effectiveness of Cross-Dimensional Entropy Model. To validate the performance improvement offered by our proposed cross-dimensional entropy model, we compare it with the joint auto-regressive and hyperprior entropy model [24], which we refer to as Baseline (Minnen). As shown in Fig. 8 and Tab. 3, the + +Table 2: Average runtime of different methods. + +
CodecBPG HEVC VVC SASIC BCSIC LDMIC BiSIC (ours)BiSIC-Fast (ours)
Runtime16.17s 28.16s 190.27s 20.24s 89.44s 81.35s 167.25s22.82s
+ +![](images/5b612df122bf7c75ea5d93b798d3d1c872b161de567c77fb5223150abfcf1a09.jpg) +Fig. 8: Rate-distortion curves of different ablation baselines on InStereo2K. + +
MethodBD-PSNR
BiSIC (ours)0
Baseline (2D-Conv)-0.3218 dB
Baseline (Minnen)-0.3475 dB
Baseline (Ying)-0.6554 dB
Baseline (Lei)-0.4956 dB
Baseline (w/o Atten)-0.7882 dB
+ +Table 3: BD-PSNR [8] results relative to the proposed BiSIC. + +proposed cross-dimensional entropy model achieves better rate-distortion performance than Baseline (Minnen). This suggests that our model provides more accurate probability estimations, which in turn minimizes the coding overhead. Effectiveness of Mutual Attention Block. To evaluate the effectiveness of the proposed mutual attention block, we consider three baselines for comparisons, namely Baseline (Ying), Baseline (Lei), and Baseline (w/o Atten). Baseline (Ying) and Baseline (Lei) replace the mutual attention block with the stereo attention module [39] and the contextual transfer module [20], respectively. In particular, Baseline (w/o Atten) represents a variant without the attention module. As shown in Fig. 8 and Tab. 3, our proposed model outperforms all baselines by a large margin. Notably, removing the proposed mutual attention block results in a dramatic degradation, which demonstrates its significance. + +# 5 Conclusion + +In this work, we introduce a bidirectional stereo image compression network that employs 3D convolution to encode local stereo features and propose a mutual attention block to capture global features. Besides, we design a symmetric cross-dimensional entropy model that integrates hyperprior, spatial context, channel context, and stereo dependency. Our bidirectional design mitigates the issue of imbalanced compression quality intrinsic to unidirectional methods and eliminates the requirement for sequential coding. As demonstrated in the experimental results, our BiSIC achieves greater bit savings at the same level of compression quality compared to both traditional coding standards and existing learning-based methods. Moreover, we propose a fast variant that significantly reduces the compression runtime while maintaining competitive performance. + +# Acknowledgements + +This work was supported by the General Research Fund (Project No. 16209622) from the Hong Kong Research Grants Council. + +# References + +1. Arena, F., Collotta, M., Pau, G., Termine, F.: An overview of augmented reality. Computers 11(2), 28 (2022) +2. Balle, J., Chou, P.A., Minnen, D., Singh, S., Johnston, N., Agustsson, E., Hwang, S.J., Toderici, G.: Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing 15(2), 339-353 (2020) +3. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. arXiv preprint arXiv:1611.01704 (2016) +4. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 (2018) +5. Bao, W., Wang, W., Xu, Y., Guo, Y., Hong, S., Zhang, X.: Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences 63, 1-11 (2020) +6. Bellard, F.: BPG image format. Website (2014), https://bellard.org/bpg/ +7. Bishop, C.M.: Latent variable models. In: Learning in graphical models, pp. 371-403. Springer (1998) +8. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33 (2001) +9. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology 31(10), 3736-3764 (2021) +10. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +11. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016) +12. Deng, X., Deng, Y., Yang, R., Yang, W., Timofte, R., Xu, M.: MASIC: Deep mask stereo image compression. IEEE Transactions on Circuits and Systems for Video Technology (2023) +13. Deng, X., Yang, W., Yang, R., Xu, M., Liu, E., Feng, Q., Timofte, R.: Deep homography for efficient stereo image compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1492-1501 (June 2021) +14. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5718-5727 (2022) +15. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14771-14780 (2021) + +16. Hu, Y., Yang, W., Liu, J.: Coarse-to-fine hyper-prior modeling for learned image compression. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11013-11020 (2020) +17. Jiang, W., Yang, J., Zhai, Y., Ning, P., Gao, F., Wang, R.: Mlic: Multi-reference entropy model for learned image compression. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7618-7627 (2023) +18. Kadaikar, A., Dauphin, G., Mokraoui, A.: Joint disparity and variable size-block optimization algorithm for stereoscopic image compression. Signal Processing: Image Communication 61, 1-8 (2018) +19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +20. Lei, J., Liu, X., Peng, B., Jin, D., Li, W., Gu, J.: Deep stereo image compression via bi-directional coding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19669-19678 (June 2022) +21. Liu, J., Wang, S., Urtasun, R.: DSIC: Deep stereo image compression. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019) +22. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14388-14397 (2023) +23. Liu, Y., Ren, J., Zhang, J., Liu, J., Lin, M.: Visually imbalanced stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2029-2038 (2020) +24. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems 31 (2018) +25. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3339-3343. IEEE (2020) +26. Pan, G., Lu, G., Hu, Z., Xu, D.: Content adaptive latents and decoder for neural image compression. In: European Conference on Computer Vision. pp. 556-573. Springer (2022) +27. Shen, Z., Zhang, M., Zhao, H., Yi, S., Li, H.: Efficient attention: Attention with linear complexities. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 3531-3539 (2021) +28. Skodras, A., Christopoulos, C., Ebrahimi, T.: The jpeg 2000 still image compression standard. IEEE Signal processing magazine 18(5), 36-58 (2001) +29. Song, X., Zhao, X., Fang, L., Hu, H., Yu, Y.: Edge stereo: An effective multi-task learning network for stereo matching and edge detection. International Journal of Computer Vision 128, 910-930 (2020) +30. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (HEVC) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012) +31. Tech, G., Chen, Y., Müller, K., Ohm, J.R., Vetro, A., Wang, Y.K.: Overview of the multiview and 3d extensions of high efficiency video coding. IEEE Transactions on Circuits and Systems for Video Technology 26(1), 35-49 (2015) +32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) + +33. Vetro, A., Wiegand, T., Sullivan, G.J.: Overview of the stereo and multiview video coding extensions of the h. 264/mpeg-4 avc standard. Proceedings of the IEEE 99(4), 626-642 (2011) +34. Wallace, G.K.: The JPEG still picture compression standard. Communications of the ACM 34(4), 30-44 (1991) +35. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402. IEEE (2003) +36. Wödlinger, M., Kotera, J., Keglevic, M., Xu, J., Sablatnig, R.: ECSIC: Epipolar cross attention for stereo image compression. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3436-3445 (2024) +37. Wödlinger, M., Kotera, J., Xu, J., Sablatnig, R.: SASIC: Stereo image compression with latent shifts and stereo attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 661-670 (June 2022) +38. Yin, H., Wang, Y., Tang, L., Ding, X., Huang, S., Xiong, R.: 3d lidar map compression for efficient localization on resource constrained vehicles. IEEE transactions on intelligent transportation systems 22(2), 837-852 (2020) +39. Ying, X., Wang, Y., Wang, L., Sheng, W., An, W., Guo, Y.: A stereo attention module for stereo image super-resolution. IEEE Signal Processing Letters 27, 496-500 (2020) +40. Zhai, Y., Tang, L., Ma, Y., Peng, R., Wang, R.: Disparity-based stereo image compression with aligned cross-view priors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2351-2360 (2022) +41. Zhang, X., Shao, J., Zhang, J.: LDMIC: Learning-based distributed multi-view image coding. arXiv preprint arXiv:2301.09799 (2023) +42. Zhu, Y., Yang, Y., Cohen, T.: Transformer-based transform coding. In: International Conference on Learning Representations (2021) \ No newline at end of file diff --git a/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/images.zip b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c4efbd1537669e89348480b9ebbb2b270e69f6d4 --- /dev/null +++ b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7f075e95550e52512b870d91cba43ae3c416bb1ab4f0ad3ec0f62a62c658eb1 +size 553523 diff --git a/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/layout.json b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3f7de208d6f7a3e59abdd74852b7980f96ce3210 --- /dev/null +++ b/2024/Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model/layout.json @@ -0,0 +1,8336 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "text", + "content": "Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 206, + 168, + 405, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 168, + 405, + 194 + ], + "spans": [ + { + "bbox": [ + 206, + 168, + 405, + 194 + ], + "type": "text", + "content": "Zhening Liu, Xinjie Zhang, Jiawei Shao, Zehong Lin, and Jun Zhang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 202, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 202, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 202, + 480, + 236 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong SAR {zhening.liu,xzhangga,jiawei.shao}@connect.ust.hk, {eezhlin,eejzhang}@ust.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 268, + 455, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 268, + 455, + 443 + ], + "spans": [ + { + "bbox": [ + 160, + 268, + 455, + 443 + ], + "type": "text", + "content": "Abstract. With the rapid advancement of stereo vision technologies, stereo image compression has emerged as a crucial field that continues to draw significant attention. Previous approaches have primarily employed a unidirectional paradigm, where the compression of one view is dependent on the other, resulting in imbalanced compression. To address this issue, we introduce a symmetric bidirectional stereo image compression architecture, named BiSIC. Specifically, we propose a 3D convolution based codec backbone to capture local features and incorporate bidirectional attention blocks to exploit global features. Moreover, we design a novel cross-dimensional entropy model that integrates various conditioning factors, including the spatial context, channel context, and stereo dependency, to effectively estimate the distribution of latent representations for entropy coding. Extensive experiments demonstrate that our proposed BiSIC outperforms conventional image/video compression standards, as well as state-of-the-art learning-based methods, in terms of both PSNR and MS-SSIM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 455, + 453, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 455, + 453, + 477 + ], + "spans": [ + { + "bbox": [ + 159, + 455, + 453, + 477 + ], + "type": "text", + "content": "Keywords: Stereo Image Compression " + }, + { + "bbox": [ + 159, + 455, + 453, + 477 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 455, + 453, + 477 + ], + "type": "text", + "content": " Bidirectional Architecture " + }, + { + "bbox": [ + 159, + 455, + 453, + 477 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 455, + 453, + 477 + ], + "type": "text", + "content": " Cross-Dimensional Entropy Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 498, + 230, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 230, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 230, + 511 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 525, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 646 + ], + "type": "text", + "content": "Stereo vision, which mimics human binocular vision, has emerged as an important area in computer vision. It enables a wide range of applications, such as 3D movies, remote sensing [29], autonomous driving [38], and augmented/virtual reality (AR/VR) [1]. With the rapid development of stereo cameras, stereo images have become essential for providing immersive and convincing visual presentations that enhance user experience and benefit downstream applications. Meanwhile, the proliferation of stereo cameras leads to a significant surge in the volume of stereo images, resulting in substantial storage costs and posing challenges to transmission efficiency. Consequently, stereo image compression plays a pivotal role in ensuring the practicality and effectiveness of stereo vision." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "In recent years, numerous methods have been developed for single image compression, including traditional CODECs [6, 34] and learning-based methods [2-4, 10, 24]. These methods, however, focus on extracting individual features for compression and fail to capture the cross-view correlation. Consequently, they are less effective for stereo image compression that requires the extraction of both intra-view information and inter-view information. To exploit the inherent shared information in stereo images, traditional methods for stereo image coding follow a predictive compression procedure of video standards [9, 31]. In this process, one view is compressed first and then used to predict the other view. Subsequently, the disparity between the predicted and actual views is compressed. Nonetheless, these methods typically rely on hand-crafted modules and lack end-to-end optimization, resulting in suboptimal performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 271, + 483, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 483, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 483, + 475 + ], + "type": "text", + "content": "Inspired by the success of learning-based single image compression, researchers have proposed novel stereo image compression methods using deep neural networks (DNNs) to enhance compression performance [12, 13, 20, 21, 36, 37, 40, 41]. Initially, extensive research focuses on unidirectional learned stereo image compression methods [12, 13, 21, 36, 37, 40], where one view is compressed and used as a reference to compress the other view. Despite achieving high overall rate-distortion (RD) performance, these unidirectional learned compression methods yield a noticeable imbalance in compression quality between stereo views, similar to traditional predictive compression methods. This imbalance is unfavorable for human vision and undermines downstream machine vision tasks [23]. Moreover, some of these methods [12, 13, 37, 40] heavily rely on sequential coding order and accurate estimation of the explicit warping relationship between views. To avoid these issues, a recent study [20] proposes a bidirectional compression method that leverages 2D convolution to extract individual features and facilitate information sharing between views. Nonetheless, the compression performance is limited due to the separate processing of convolutions and the plain spatial context based entropy model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Based on the above discussion, we propose a novel bidirectional method for stereo image compression, named BiSIC, to address the aforementioned issues. BiSIC achieves improvements in four key aspects: encoder/decoder structure, mutual stereo learning, entropy model, and coding speed. Firstly, regarding the codec structure, although 2D convolutions have proven effective in various computer vision tasks, they fail to capture the aligned morphs and features between stereo views. To overcome this limitation, we propose using 3D convolution as the backbone for the encoder/decoder, which takes both views as input to capture the inter-view dependencies that are inaccessible in previous works based on 2D convolution. Secondly, for mutual stereo learning, we propose a mutual attention block to facilitate the transfer of features between views. This module enhances the integration and cooperation of the stereo views, leading to improved compression performance. Thirdly, for the entropy model, prior methods primarily use hyperprior and spatial context as conditions, overlooking the rich inter-view dependencies across other dimensions. To exploit these dependencies," + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 113, + 446, + 237 + ], + "blocks": [ + { + "bbox": [ + 168, + 113, + 446, + 237 + ], + "lines": [ + { + "bbox": [ + 168, + 113, + 446, + 237 + ], + "spans": [ + { + "bbox": [ + 168, + 113, + 446, + 237 + ], + "type": "image", + "image_path": "82a897adf9318952b36b7d0d6d6f253c71eab74f67470215fcd560c7c0138048.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "type": "text", + "content": "Fig. 1: The proposed bidirectional stereo image compression architecture. AE and AD are arithmetic coder for entropy coding. The backbones of encoder " + }, + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "type": "text", + "content": ", decoder " + }, + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "type": "text", + "content": ", hyper encoder " + }, + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "type": "text", + "content": ", and hyper decoder " + }, + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 130, + 245, + 482, + 322 + ], + "type": "text", + "content": " are constructed with 3D convolution to model local features. 3DCovTr denotes transposed 3D convolution. To enhance global feature extraction, bidirectional mutual attention blocks are inserted between 3D convolutional layers. Moreover, a novel cross-dimensional entropy model is utilized to capture complex inter-view dependencies." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 479, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 479, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 479, + 366 + ], + "type": "text", + "content": "we construct a cross-dimensional entropy model that aggregates hyperprior, spatial context, channel context, and stereo dependency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 367, + 480, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 480, + 438 + ], + "type": "text", + "content": "Specifically, we introduce a masked 3D convolution to capture the stereo spatial context and utilize an attention block to process the stereo channel context. Finally, to improve coding speed and reduce auto-regressive steps, we design a fast variant based on a stereo-checkerboard structure. This design converts auto-regressive inference into a two-fold operation, improving the coding efficiency. The contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 442, + 480, + 582 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 442, + 480, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 480, + 477 + ], + "type": "text", + "content": "- We propose a novel 3D convolution based codec backbone that processes stereo images as a whole, enabling the learning of local stereo features and effective capture of shared patterns across views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 478, + 480, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 480, + 512 + ], + "type": "text", + "content": "- We design a bidirectional mutual attention block that facilitates the exchange of global features between stereo views. This block produces compact attention maps and is robust to input images of different sizes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 513, + 480, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 513, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 513, + 480, + 547 + ], + "type": "text", + "content": "- We develop a powerful symmetric cross-dimensional entropy model, which incorporates hyperpriors, spatial context, channel context, and stereo dependency as conditioning factors, to enhance the probability estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 548, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 548, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 138, + 548, + 480, + 582 + ], + "type": "text", + "content": "- To accelerate coding, we propose a stereo-checkerboard design that transforms the spatial auto-regressive inference into a two-fold process. This design strikes a balance between time consumption and compression quality." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 603, + 242, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 242, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 242, + 615 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 624, + 293, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 293, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 293, + 636 + ], + "type": "text", + "content": "2.1 Single Image Compression" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Image compression is an important topic that continuously attracts significant research interest. With the increasing demand for vision products and applica" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "text", + "content": "BiSIC: Bidirectional Stereo Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "tions, steady progress has been made in traditional image compression standards, evolving from JPEG [28, 34] and BPG [6] to video compression standards such as HEVC [30] and VVC [9]. These traditional methods establish a standard compression pipeline comprising transform, quantization, and entropy coding. Over the past five years, deep learning has revolutionized various vision applications. As demonstrated in the seminal work [3], DNNs have the potential to replace components in traditional compression methods and enable end-to-end compression. This work introduces uniform noise to mimic the distortion caused by quantization, thereby paving the way for end-to-end optimization in image compression." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "text", + "content": "In learned image compression models [3,4,10,14-17,22,24-26,42], neural networks are employed to parameterize the probability distribution of transformed latents. Accurate estimation of this distribution is crucial for achieving compact compression via entropy coding. To enhance probability estimation, various conditions have been integrated into entropy models [2], including hyperpriors [4], spatial context [24], and channel-wise context [25], to capture a broader range of dependencies. To be specific, hyperprior is a lightweight subsidiarity information transmitted alongside the encoded latents. Both spatial context and channel-wise context leverage auto-regressive inference, which is conditioned on the preceding segments of data by iteratively interleaving the inference. Checkerboard methods [15], on the other hand, avoid the spatial auto-regressive process and accelerate the coding speed by constraining the spatial context to the near neighbors of each pixel instead of spatial causal components. These dependencies operate on diverse scales and can be progressively combined to further improve effectiveness, as demonstrated in [14, 17]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 432, + 295, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 432, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 295, + 443 + ], + "type": "text", + "content": "2.2 Stereo Image Compression" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "content": "Stereo image compression is an extension of single image compression that considers capturing both intra-view features and inter-view correlations. Traditional approaches employ key-frame based predictive coding, akin to video coding [18, 33]. However, these methods rely on hand-crafted feature extractors and classical optimization techniques. As these modules are optimized individually, they often fail to achieve satisfactory RD performance. Moreover, the predictive compression pipeline also leads to imbalanced compression." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "type": "text", + "content": "In recent years, learned stereo image compression methods have gained increasing attention [12, 13, 20, 21, 36, 37, 40, 41]. These methods can be broadly categorized into unidirectional and bidirectional approaches. Unidirectional methods first compress one view image and then use it as a reference to compress the other view, significantly reducing the required bits. The pioneering work [21] introduces a skip module to warp the features from one view to encode the other. Subsequent works have utilized various techniques to represent the warping relationship between stereo views, such as homography matrices [12, 13], disparity maps [40], and horizontal shifts [37]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "These unidirectional methods heavily rely on the precise estimation of the spatial transformations. A very recent work, ECSIC [36], avoids explicit para" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "metric transformation in favor of constructing a unidirectional entropy model that conditions the right image on the known left image. However, the decompressed results of these unidirectional methods often have imbalanced quality due to the one-way dependency between stereo views. On the other hand, bidirectional methods, while remaining less explored, produce more balanced quality by extracting shared information from both views and compressing stereo images simultaneously. One such approach [20] introduces a contextual transform module for feature transfer between views. However, it employs separate down-sampling networks and only takes spatial context into account, which limits the compression performance. In contrast, to fully exploit the shared information and establish stronger dependencies, we employ a joint downsampling codec and a cross-dimensional entropy model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 277, + 266, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 266, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 266, + 291 + ], + "type": "text", + "content": "3 Proposed Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 302, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 398 + ], + "type": "text", + "content": "The proposed architecture for stereo image compression, as depicted in Fig. 1, consists of two main components: a joint codec and an entropy model. The codec employs 3D convolutions to transform the original stereo images into compact latent representations, effectively exploiting the correlation between the two views. Subsequently, the entropy model estimates the probability distribution of these latent representations, enabling the generation of a compact bitstream through entropy coding. In the following subsections, we elaborate on the codec and the entropy model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 415, + 221, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 221, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 221, + 426 + ], + "type": "text", + "content": "3.1 Joint Codec" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "text", + "content": "3D Convolution Backbone. The codec acts as a non-linear transformation that downsamples the raw images into compact latent representations and generates hyperpriors for the entropy model. Different from previous approaches to stereo image compression that employ 2D convolutional layers, we adopt 3D convolutions as the backbone of our codec. Unlike 2D convolution, which operates separately on each view, 3D convolution concurrently processes both stereo images. Consequently, it inherently possesses the capability to extract inter-view correlations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": "Our encoder, denoted by " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": ", comprises four 3D convolutional layers, with bidirectional mutual attention blocks inserted after the second and the fourth layers to facilitate global feature communication between the two views. The encoder transforms a pair of stereo images, i.e., the left and right images denoted by " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{x}_l" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{x}_r \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": ", into compact latent representations " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{y}_l" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{y}_r \\in \\mathbb{R}^{B \\times N \\times \\frac{H}{16} \\times \\frac{W}{16}}" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": " and quantizes them into " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}} := \\{\\hat{\\pmb{y}}_l, \\hat{\\pmb{y}}_r\\}" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": " for efficient transmission. Conversely, the decoder " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": " transforms the quantized latents " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": " back into images " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_l" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_r" + }, + { + "bbox": [ + 130, + 530, + 482, + 626 + ], + "type": "text", + "content": ". The encoding and decoding processes are expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 265, + 635, + 480, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 635, + 480, + 647 + ], + "spans": [ + { + "bbox": [ + 265, + 635, + 480, + 647 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {l}, \\boldsymbol {y} _ {r} = \\mathbf {E} \\left(\\boldsymbol {x} _ {l}, \\boldsymbol {x} _ {r}\\right), \\tag {1}", + "image_path": "d07fec9ec5aab4ddb75a17336fee5191df390759d56dd71d54142cef18d96b13.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 265, + 650, + 480, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 650, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 265, + 650, + 480, + 662 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {l}, \\hat {\\boldsymbol {x}} _ {r} = \\mathbf {D} (\\hat {\\boldsymbol {y}} _ {l}, \\hat {\\boldsymbol {y}} _ {r}). \\tag {2}", + "image_path": "e606b7708ddb6d8e522afa0d1a843293c870c84bde36d13888708516bf30b11f.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 254, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 254, + 91, + 448, + 102 + ], + "type": "text", + "content": "BiSIC: Bidirectional Stereo Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "To capture the dependencies among elements in the latents " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " for efficient entropy coding [4,7], hyperpriors are generated through another 3D convolution based hyper encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". The generated hyperpriors " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\pmb{z}_l, \\pmb{z}_r \\in \\mathbb{R}^{B \\times M \\times \\frac{H}{64} \\times \\frac{W}{64}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " are quantized into " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}_l, \\hat{\\pmb{z}}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " and then transmitted alongside the quantized latents " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". At the decoder side, a 3D convolution based hyper decoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " upscales the quantized hyperpriors " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}_l, \\hat{\\pmb{z}}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{z}}_l, \\tilde{\\pmb{z}}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", which are subsequently utilized in the entropy model, as elaborated in Sec. 3.2. The processes of the hyper encoder and hyper decoder are expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 264, + 220, + 480, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 220, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 264, + 220, + 480, + 232 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {l}, \\boldsymbol {z} _ {r} = h _ {a} (\\boldsymbol {y} _ {l}, \\boldsymbol {y} _ {r}), \\tag {3}", + "image_path": "a555ffa84ce60ffcea1bb63dee75d5007d25f9b09591e5800fb0e11865b4d57f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 265, + 235, + 480, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 235, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 265, + 235, + 480, + 247 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {z}} _ {l}, \\tilde {\\boldsymbol {z}} _ {r} = h _ {s} (\\hat {\\boldsymbol {z}} _ {l}, \\hat {\\boldsymbol {z}} _ {r}). \\tag {4}", + "image_path": "694c853e5920fc44c905f33f009d3cad862f8dd138313270ad5fee8cbe5537de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 251, + 479, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 347 + ], + "type": "text", + "content": "Bidirectional Mutual Attention Block. Convolutional layers have an excellent ability for local modeling, but their ability for long-range feature modeling is relatively limited. On the other hand, attention modules [32] excel at extracting long-range features. Therefore, we propose combining the attention mechanism with convolutional layers to compensate for the shortcomings of the latter and leverage the advantages of both convolution and attention. To this end, we design a novel bidirectional mutual attention block, as depicted in Fig. 2, to be inserted between convolutional layers as the pink blocks shown in Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "text", + "content": "As illustrated in Fig. 2, the bidirectional mutual attention block contains two cross-attention stages, each followed by a self-attention layer. The two cross-attention stages differ in how they concentrate on the feature map and are referred to as cross-key attention and cross-query attention. Before each cross-attention stage, the input stereo data " + }, + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "inline_equation", + "content": "\\pmb{y}_l, \\pmb{y}_r" + }, + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "text", + "content": " first pass through a residual block, which are then embedded into low-dimensional query, key, and value tensors " + }, + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "inline_equation", + "content": "\\pmb{Q}, \\pmb{K}, \\pmb{V} \\in \\mathbb{R}^{B \\times C \\times (H \\times W)}" + }, + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "text", + "content": ". To avoid the prohibitive computation of classical attention, which yields an " + }, + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "inline_equation", + "content": "(H \\times W) \\times (H \\times W)" + }, + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "text", + "content": " sized attention map, we adopt the efficient attention approach proposed in [27]. This approach computes the product of the key and the transposed value, thereby reducing the attention map size to " + }, + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "inline_equation", + "content": "C_K \\times C_V" + }, + { + "bbox": [ + 130, + 348, + 480, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 479, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 480, + 525 + ], + "type": "text", + "content": "In the cross-key attention stage, for each view, we generate an attention map using its key and the value from the other view, which is then queried by the current view. By doing so, we can extract the cross-key feature " + }, + { + "bbox": [ + 130, + 479, + 480, + 525 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 130, + 479, + 480, + 525 + ], + "type": "text", + "content": " from the stereo views as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 145, + 533, + 480, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 533, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 145, + 533, + 480, + 546 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Phi} _ {r \\rightarrow l} = (\\sigma (\\boldsymbol {K} _ {r}) \\times \\boldsymbol {V} _ {l} ^ {T}) ^ {T} \\times \\sigma (\\boldsymbol {Q} _ {r}), \\quad \\boldsymbol {\\Phi} _ {l \\rightarrow r} = (\\sigma (\\boldsymbol {K} _ {l}) \\times \\boldsymbol {V} _ {r} ^ {T}) ^ {T} \\times \\sigma (\\boldsymbol {Q} _ {l}), \\tag {5}", + "image_path": "a902563ad181373d4d74c85c6918ac04cd513ebea9e3feab39bfeeadacf13d10.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 552, + 479, + 611 + ], + "type": "text", + "content": " is the softmax function. This approach calculates the attention map using the key and value from different views, thereby facilitating inter-view alignment and the identification of common patterns. When a feature in one view is correlated with the corresponding feature in the other view, it is effectively captured through the attention map." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 612, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 647 + ], + "type": "text", + "content": "In the cross-query attention stage, the attention map is obtained using the key and value from a single view, and then queried by the other view, yielding the cross-query feature " + }, + { + "bbox": [ + 130, + 612, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 130, + 612, + 482, + 647 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 144, + 652, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 652, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 144, + 652, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\Psi_ {r \\rightarrow l} = \\left(\\sigma \\left(\\boldsymbol {K} _ {l}\\right) \\times \\boldsymbol {V} _ {l} ^ {T}\\right) ^ {T} \\times \\sigma \\left(\\boldsymbol {Q} _ {r}\\right), \\Psi_ {l \\rightarrow r} = \\left(\\sigma \\left(\\boldsymbol {K} _ {r}\\right) \\times \\boldsymbol {V} _ {r} ^ {T}\\right) ^ {T} \\times \\sigma \\left(\\boldsymbol {Q} _ {l}\\right). \\tag {6}", + "image_path": "671e2e18ed5810c7cfeb36f918df15ac5504c36fad73dfb97bcc7c9c5bbd86d1.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 114, + 459, + 229 + ], + "blocks": [ + { + "bbox": [ + 155, + 114, + 459, + 229 + ], + "lines": [ + { + "bbox": [ + 155, + 114, + 459, + 229 + ], + "spans": [ + { + "bbox": [ + 155, + 114, + 459, + 229 + ], + "type": "image", + "image_path": "b30f87dee6598cab8a1b245a633985d65c548e00d77420fae0ed4730f3075ace.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 482, + 283 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 283 + ], + "type": "text", + "content": "Fig. 2: Overview of the proposed bidirectional mutual attention block. The blue and green lines represent features extracted from the left view and right view, respectively. The network structures of the residual block, the basic efficient attention unit, and the combine block are illustrated on the right, from top to bottom." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 355 + ], + "type": "text", + "content": "The attention map in this stage is generated within one view, thus maintaining an inner-focused relationship. This process preserves the individual self-features of each view, while the other view serves as a reference for determining the allocation of attention weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "content": "After obtaining the cross-key attention and cross-query attention features " + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\Phi, \\Psi" + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "content": ", we employ self-attention to augment these features as " + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\tilde{\\Phi}, \\tilde{\\Psi}" + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "content": " and use a shared-parameter combine block to get the desired mutual attention output, as shown in the bottom right of Fig. 2. Specifically, the combine block concatenates the augmented features " + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\tilde{\\Phi}, \\tilde{\\Psi}" + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "content": " and the residual component, and shrinks the channel number to be consistent with the block input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "type": "text", + "content": "This bidirectional mutual attention block captures diverse features and provides a global receptive field for the codec. It is suitable for any stereo information transfer scenarios, provided that the inputs are two features of equal size, making it a generalized plug-and-play block. Moreover, the size of the attention maps is determined only by the embedding channel " + }, + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "type": "text", + "content": " and does not vary with the size of input images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 522, + 334, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 334, + 534 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 334, + 534 + ], + "type": "text", + "content": "3.2 Cross-Dimensional Entropy Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "The proposed codec effectively transforms stereo images into small-sized latent representations. To achieve efficient compression, the next crucial step is to encode these latents into compact bitstreams using accurate probability distribution and entropy coding [4,24]. Thus, an entropy model is critical for parameterizing and estimating probabilities. In learned image compression, the probability distribution is commonly modeled as Gaussian distributions [24], with the mean " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " being estimated. To achieve accurate estimation, it is critical to provide appropriate references. Apart from hyperpriors, previous stereo image compression methods primarily use spatial context [13,20,41] and unidirectional stereo dependency as references [21,36], while abundant features on the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 254, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 254, + 91, + 448, + 102 + ], + "type": "text", + "content": "BiSIC: Bidirectional Stereo Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 113, + 462, + 236 + ], + "blocks": [ + { + "bbox": [ + 153, + 113, + 462, + 236 + ], + "lines": [ + { + "bbox": [ + 153, + 113, + 462, + 236 + ], + "spans": [ + { + "bbox": [ + 153, + 113, + 462, + 236 + ], + "type": "image", + "image_path": "fe612b9e57ce8faa802be1658b78292465b36d3069593a0608d31bef63cde874.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": "Fig. 3: Illustration of the proposed symmetric cross-dimensional entropy model. It jointly aggregates the hyperprior (blue), stereo spatial context (red), and stereo channel context (yellow) as conditions for an effective probability distribution estimation. The hyper decoder " + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": " for hyperprior is shown in Fig. 1. The masked 3D convolution for spatial context is provided in Fig. 4a. Moreover, the mutual attention block for channel context is detailed in Fig. 2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 332, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 479, + 367 + ], + "type": "text", + "content": "channel dimension are overlooked. In this work, we propose a symmetric cross-dimensional entropy model that integrates hyperprior, spatial context, channel context, and stereo dependency in a fully symmetric bidirectional way." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "content": "The architecture of the proposed cross-dimensional entropy model is depicted in Fig. 3. We use " + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_{l,i}^{k}" + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "content": " to denote the " + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "inline_equation", + "content": "(i = 1,2,\\dots,n)" + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "content": " entry of the " + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "inline_equation", + "content": "(k = 1,2,\\dots,K)" + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "content": " channel slice from the received left view latent " + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_l" + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "content": ", and likewise for the others. The cross-dimensional condition formulation is expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 422, + 480, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 422, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 205, + 422, + 480, + 438 + ], + "type": "interline_equation", + "content": "p _ {\\hat {\\boldsymbol {y}} _ {l}} \\left(\\hat {\\boldsymbol {y}} _ {l, i} ^ {k} \\mid \\hat {\\boldsymbol {z}} _ {l}, \\hat {\\boldsymbol {y}} _ {l} ^ {< k}, \\hat {\\boldsymbol {y}} _ {l, < i} ^ {k}, \\hat {\\boldsymbol {y}} _ {r} ^ {< k}, \\hat {\\boldsymbol {y}} _ {r, < i} ^ {k}\\right) \\sim \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {l}, \\boldsymbol {\\sigma} _ {l} ^ {2}\\right), \\tag {7}", + "image_path": "af4acbce36cfa29364370ada6af6a1ddfd80ea884d8dfefb5fd0acc262da7bd3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 201, + 440, + 480, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 440, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 201, + 440, + 480, + 455 + ], + "type": "interline_equation", + "content": "p _ {\\hat {\\boldsymbol {y}} _ {r}} \\left(\\hat {\\boldsymbol {y}} _ {r, i} ^ {k} \\mid \\hat {\\boldsymbol {z}} _ {r}, \\hat {\\boldsymbol {y}} _ {r} ^ {< k}, \\hat {\\boldsymbol {y}} _ {r, < i} ^ {k}, \\hat {\\boldsymbol {y}} _ {l} ^ {< k}, \\hat {\\boldsymbol {y}} _ {l, < i} ^ {k}\\right) \\sim \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {r}, \\boldsymbol {\\sigma} _ {r} ^ {2}\\right). \\tag {8}", + "image_path": "d8740a48f9dac63112fb2526019accbb2a9ec90eba3f721f7e84c38e05076cc6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": "For the hyperprior component, we generate the dependency " + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\tilde{z}_l,\\tilde{z}_r" + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": " from the hyper decoder using Eq. (4). To obtain the channel context, we evenly split the stereo latents on the channel dimension into " + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": " slices. These latent slices share a great similarity, providing abundant conditions [25]. To exploit this similarity, the probability estimation is performed slice by slice, using the previous slices as references for the following one. In addition, the decoded slices of the other view also serve as references. Therefore, we propose to use the mutual attention block introduced in the previous subsection to aggregate the channel-wise features " + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\Theta_l,\\Theta_r" + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": " from both views, conditioning on " + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_l^{< k}" + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_r^{< k}" + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": ", as depicted by the yellow lines in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "To capture the spatial context within each channel slice, we consistently employ 3D convolution to jointly learn spatial features from the causal areas " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_{< i}^{k}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". Specifically, during the coding process, the current compression/decompression only accesses the areas that have already been encoded/decoded. To maintain the causality of the compression, we mask out the uncoded part, which is referred to as masked 3D convolution, as shown in Fig. 4a. In this approach, for the " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "-th entry " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_i^k" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", the weights of uncoded part are set to zero, and we obtain spatial" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 128, + 190, + 172 + ], + "blocks": [ + { + "bbox": [ + 133, + 128, + 190, + 172 + ], + "lines": [ + { + "bbox": [ + 133, + 128, + 190, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 128, + 190, + 172 + ], + "type": "image", + "image_path": "c798248cfad0cab460d4c8cbcdc3c349835e20719e14546d2fbe16c5977ce264.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 174, + 196, + 190 + ], + "lines": [ + { + "bbox": [ + 133, + 174, + 196, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 196, + 190 + ], + "type": "text", + "content": "(a) Masked 3D convolution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 214, + 114, + 329, + 172 + ], + "blocks": [ + { + "bbox": [ + 214, + 114, + 329, + 172 + ], + "lines": [ + { + "bbox": [ + 214, + 114, + 329, + 172 + ], + "spans": [ + { + "bbox": [ + 214, + 114, + 329, + 172 + ], + "type": "image", + "image_path": "0f40eb1b7e984197b69af06fb341ef06dee00dd3644a7b311ffca5b4b53b5f31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 174, + 337, + 190 + ], + "lines": [ + { + "bbox": [ + 214, + 174, + 337, + 190 + ], + "spans": [ + { + "bbox": [ + 214, + 174, + 337, + 190 + ], + "type": "text", + "content": "(b) Auto-regressive pattern in BiSIC entropy model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 356, + 114, + 471, + 172 + ], + "blocks": [ + { + "bbox": [ + 356, + 114, + 471, + 172 + ], + "lines": [ + { + "bbox": [ + 356, + 114, + 471, + 172 + ], + "spans": [ + { + "bbox": [ + 356, + 114, + 471, + 172 + ], + "type": "image", + "image_path": "8ac0cdfffcea789de3a4319aec3819d5335fc413e5ea0e637255fc9ef3c32c68.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 174, + 477, + 190 + ], + "lines": [ + { + "bbox": [ + 355, + 174, + 477, + 190 + ], + "spans": [ + { + "bbox": [ + 355, + 174, + 477, + 190 + ], + "type": "text", + "content": "(c) Stereo-checkerboard pattern in BiSIC-Fast." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 200, + 480, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 288 + ], + "type": "text", + "content": "Fig. 4: (Left) Illustration of masked 3D convolution. The blue region is used as the condition for the probability estimation of the current target depicted in red. The weights of the convolution kernel in blue are valid, while the white region is masked to maintain the causal spatial context. (Middle) Demonstration of the auto-regressive process in the cross-dimensional entropy model. The yellow and red arrows represent the channel context and spatial context, respectively. The spatial context is processed entry by entry. (Right) Demonstration of the stereo-checkerboard in the fast variant, where the blue and red parts are each processed only once." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 312, + 479, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 479, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 479, + 336 + ], + "type": "text", + "content": "context features " + }, + { + "bbox": [ + 130, + 312, + 479, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{\\Upsilon}_{l,i},\\mathbf{\\Upsilon}_{r,i}" + }, + { + "bbox": [ + 130, + 312, + 479, + 336 + ], + "type": "text", + "content": " by considering the causal context " + }, + { + "bbox": [ + 130, + 312, + 479, + 336 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_{l, < i}^{k}" + }, + { + "bbox": [ + 130, + 312, + 479, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 312, + 479, + 336 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_{r, < i}^{k}" + }, + { + "bbox": [ + 130, + 312, + 479, + 336 + ], + "type": "text", + "content": " from both views simultaneously, as depicted by the red lines in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 338, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 479, + 361 + ], + "type": "text", + "content": "Upon extracting various dependency features, we employ an aggregation network " + }, + { + "bbox": [ + 131, + 338, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{ag}}" + }, + { + "bbox": [ + 131, + 338, + 479, + 361 + ], + "type": "text", + "content": " to fuse the conditions and obtain the probability estimations as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 189, + 370, + 480, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 370, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 189, + 370, + 480, + 384 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {l}, \\boldsymbol {\\sigma} _ {l} ^ {2} = \\mathbf {G} _ {\\mathrm {a g}} \\left(\\tilde {\\boldsymbol {z}} _ {l}, \\boldsymbol {\\Theta} _ {l}, \\boldsymbol {\\Upsilon} _ {l}\\right), \\boldsymbol {\\mu} _ {r}, \\boldsymbol {\\sigma} _ {r} ^ {2} = \\mathbf {G} _ {\\mathrm {a g}} \\left(\\tilde {\\boldsymbol {z}} _ {r}, \\boldsymbol {\\Theta} _ {r}, \\boldsymbol {\\Upsilon} _ {r}\\right). \\tag {9}", + "image_path": "7206ca1739949a47c0aeaf349722e6cdd9997d4b8bcd62bf91fb08a1023c79e1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 394, + 479, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 479, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 479, + 418 + ], + "type": "text", + "content": "The entire procedure in the entropy model is designed to be fully symmetric and bidirectional, which aligns with our motivation for a bidirectional design." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 437, + 230, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 230, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 230, + 448 + ], + "type": "text", + "content": "3.3 Loss Function" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 459, + 488, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 488, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 488, + 519 + ], + "type": "text", + "content": "The proposed stereo image compression network is trained using the rate-distortion loss. The rate is defined by bit-per-pixel (BPP), which includes bits for transmitting latents " + }, + { + "bbox": [ + 130, + 459, + 488, + 519 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_l, \\hat{\\pmb{y}}_r" + }, + { + "bbox": [ + 130, + 459, + 488, + 519 + ], + "type": "text", + "content": " and hyperprioris " + }, + { + "bbox": [ + 130, + 459, + 488, + 519 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}_l, \\hat{\\pmb{z}}_r" + }, + { + "bbox": [ + 130, + 459, + 488, + 519 + ], + "type": "text", + "content": ". The distortion is measured using the mean squared error (MSE) and the multi-scale structural similarity index measure (MS-SSIM) [35]. The overall loss function is given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 198, + 529, + 480, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 529, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 198, + 529, + 480, + 554 + ], + "type": "interline_equation", + "content": "\\mathcal {L} := \\lambda D + R = \\lambda \\sum_ {l, r} d (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) + \\sum_ {l, r} (r (\\hat {\\boldsymbol {y}}) + r (\\hat {\\boldsymbol {z}})), \\tag {10}", + "image_path": "8445c7612451e07decb8ef0048dd10b85ec57d4feda361f78b1756dfeb548e77.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "text", + "content": " denotes the distortion function with MSE or MS-SSIM, " + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "inline_equation", + "content": "r(\\cdot)" + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "text", + "content": " calculates the bitrate using the estimated probability of " + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}" + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "text", + "content": " [4], and " + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 564, + 479, + 599 + ], + "type": "text", + "content": " is a hyperparameter that balances the rate and distortion." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 620, + 222, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 222, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 222, + 631 + ], + "type": "text", + "content": "3.4 Fast Variant" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "The proposed entropy model effectively captures multiple references and estimates the probability distribution. However, the auto-regressive iterations in" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "text", + "content": "BiSIC: Bidirectional Stereo Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 478, + 222 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 222 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 222 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 222 + ], + "type": "image", + "image_path": "986a5221979dcd8eaae64a0f629b81edad2c86a61503593047d1f0be819543ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 275 + ], + "type": "text", + "content": "Fig. 5: Illustration of the entropy model for the proposed fast variant based on stereo-checkerboard. The blue and red parts represent the stereo anchor and stereo non-anchor, respectively. By concurrently processing two views, the 3D convolution based anchor context network effectively captures the dependency from the stereo anchor." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 298, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 479, + 369 + ], + "type": "text", + "content": "the spatial context are time-consuming. To strike a better balance between compression runtime and performance, we need to reduce the auto-regressive executions while maintaining a strong condition for the entropy model. To this end, we propose a fast variant of our model, named BiSIC-Fast, by extending the checkerboard structure [15] into stereo-checkerboard and utilizing 3D convolutions to jointly model the spatial context of two views with only two iterations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 370, + 480, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 480, + 476 + ], + "type": "text", + "content": "This fast variant maintains the codec as BiSIC, while the entropy model network is modified as illustrated in Fig. 5. In this approach, we split the latents " + }, + { + "bbox": [ + 130, + 370, + 480, + 476 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 370, + 480, + 476 + ], + "type": "text", + "content": " into two parts, stereo anchor part and stereo non-anchor part, as represented in blue and red in Fig. 5, respectively. The entropy parameter estimation for the stereo anchor part relies on the hyperprior condition and channel context. The stereo anchor is decompressed first, and then the stereo non-anchor utilizes the decompressed stereo anchor as a dependency. Consequently, the entropy parameter estimation for the stereo non-anchor part is based on the hyperprior, channel context, and stereo anchor context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 478, + 481, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 481, + 526 + ], + "type": "text", + "content": "By adopting this approach, we avoid the spatial auto-regressive process in favor of a two-fold operation, as demonstrated in Fig. 4b and Fig. 4c. Although the spatial auto-regressive process is omitted to save runtime, the neighbors of each pixel in the latents are still maintained, preserving a good perceptive field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 545, + 230, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 230, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 230, + 559 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 572, + 312, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 312, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 312, + 585 + ], + "type": "text", + "content": "4.1 Experimental Implementation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Datasets. We evaluate the performance of our methods on two benchmark datasets: InStereo2K [5] and Cityscapes [11]. InStereo2K contains 2,060 indoor scene stereo image pairs of size " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "1,080 \\times 860" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". It is split into 2,010 image pairs for training and 50 for testing. Cityscapes contains 5,000 outdoor scene stereo image pairs of size " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "2,048 \\times 1,024" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", where 2,975 pairs are allocated for training and 1,525 for testing." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 115, + 481, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 115, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 481, + 187 + ], + "type": "text", + "content": "Benchmarks. We compare our methods with traditional CODECs, such as BPG [6], H.265/HEVC [30], MV-HEVC [31], and H.266/VVC [9]. BPG encodes each view of the stereo pair independently, while H.265/HEVC and H.266/VVC compress the left and right view images jointly. In addition, we show the results of learning-based stereo image compression schemes, including HESIC+ [13], SASIC [37], BCSIC [20], LDMIC [41], and ECSIC [36]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 188, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 188, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 134, + 188, + 481, + 282 + ], + "type": "text", + "content": "Metrics. The quality of the reconstructed images is evaluated by the peak signal-to-noise ratio (PSNR) and MS-SSIM [35]. The rate is measured by BPP. Besides, we use the widely employed Bjøntegaard Delta bitrate (BDBR) and Bjøntegaard Delta PSNR (BD-PSNR) [8] to compare the rate-distortion (RD) performance among the methods. Specifically, BDBR measures the average bit saving at the same distortion level, while BD-PSNR reflects the relative PSNR gap at an equivalent bitrate. A lower value of BDBR and a higher value of BD-PSNR indicate better performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 283, + 481, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 283, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 134, + 283, + 481, + 367 + ], + "type": "text", + "content": "Implementation Details. During the training process, we optimize the model using the MSE loss over 1.5 million steps. We employ the Adam optimizer [19] and set the batch size to 8. The learning rate is initialized at " + }, + { + "bbox": [ + 134, + 283, + 481, + 367 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 134, + 283, + 481, + 367 + ], + "type": "text", + "content": " and is reduced by half after every 300,000 steps. The hyperparameter " + }, + { + "bbox": [ + 134, + 283, + 481, + 367 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 134, + 283, + 481, + 367 + ], + "type": "text", + "content": " is chosen from the set {256,512,1024,2048,3072,4096} for different bitrates. To achieve better MS-SSIM performance, we fine-tune the pre-trained models for an additional 500,000 steps. More implementation details are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 384, + 197, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 384, + 197, + 395 + ], + "spans": [ + { + "bbox": [ + 134, + 384, + 197, + 395 + ], + "type": "text", + "content": "4.2 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "text", + "content": "RD Performance. Fig. 6 shows the RD curves of different compression methods on the InStereo2K and Cityscapes datasets. In terms of the MS-SSIM metric, our proposed method BiSIC outperforms all the other compression methods on both datasets. Besides, BiSIC achieves higher PSNR in the high BPP range compared to the baselines, while maintaining competitive PSNR values at low bitrates. The BDBR results are presented in Tab. 1, where BPG is set as the baseline. Our BiSIC achieves more than " + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "text", + "content": " BDBR reduction on InStereo2K and Cityscapes, respectively. Notably, compared to the state-of-the-art video codec VVC, BiSIC achieves an additional " + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "inline_equation", + "content": "12.76\\%" + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "text", + "content": " BDBR reduction for PSNR and an additional " + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "inline_equation", + "content": "30.08\\%" + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "text", + "content": " BDBR reduction for MS-SSIM on InStereo2K. Compared with unidirectional methods, BiSIC achieves significant bit savings by observing a holistic view and mutually sharing features, which helps remove redundancies in each view. Moreover, compared with the state-of-the-art bidirectional coding scheme BCSIC [20], our method achieves an additional " + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 134, + 402, + 481, + 641 + ], + "type": "text", + "content": " BDBR reduction. This is because the employed 3D convolution captures more correlations than the separately deployed 2D convolution used in BCSIC [20], and our cross-dimensional entropy model effectively aggregates various abundant dependencies from different dimensions, providing accurate probability estimation. In addition, the proposed fast variant also demonstrates satisfactory performance, with a marginal degradation in compression performance compared to our BiSIC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 641, + 481, + 665 + ], + "type": "text", + "content": "Qualitative Results. Fig. 7 visualizes the reconstructed images of different methods on the Cityscapes dataset. For a fair comparison, all images are com" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "text", + "content": "BiSIC: Bidirectional Stereo Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 121, + 272, + 228 + ], + "blocks": [ + { + "bbox": [ + 138, + 121, + 272, + 228 + ], + "lines": [ + { + "bbox": [ + 138, + 121, + 272, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 121, + 272, + 228 + ], + "type": "image", + "image_path": "8ef643ad96c08df9f0355be8c18e1e930c4a0c876452815bcdfc02a1c883b28f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 327, + 121, + 463, + 228 + ], + "blocks": [ + { + "bbox": [ + 327, + 121, + 463, + 228 + ], + "lines": [ + { + "bbox": [ + 327, + 121, + 463, + 228 + ], + "spans": [ + { + "bbox": [ + 327, + 121, + 463, + 228 + ], + "type": "image", + "image_path": "390ec9a2c255c81e5da490d92ce89cbb58f592afdac005fe955c24d3ec41f720.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 235, + 272, + 342 + ], + "blocks": [ + { + "bbox": [ + 137, + 235, + 272, + 342 + ], + "lines": [ + { + "bbox": [ + 137, + 235, + 272, + 342 + ], + "spans": [ + { + "bbox": [ + 137, + 235, + 272, + 342 + ], + "type": "image", + "image_path": "75f78f98a40d55e15127b5cc088019f430362067f9f71edfc3a0b8b1363cf6ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 352, + 480, + 374 + ], + "lines": [ + { + "bbox": [ + 131, + 352, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 352, + 480, + 374 + ], + "type": "text", + "content": "Fig. 6: Rate-distortion curves in terms of PSNR and MS-SSIM on the InStereo2K and Cityscapes datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 327, + 236, + 462, + 342 + ], + "blocks": [ + { + "bbox": [ + 327, + 236, + 462, + 342 + ], + "lines": [ + { + "bbox": [ + 327, + 236, + 462, + 342 + ], + "spans": [ + { + "bbox": [ + 327, + 236, + 462, + 342 + ], + "type": "image", + "image_path": "34ce93292296d44c365028774ed7213c04926af87d59eb6adb2835a68e58ce4c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 496 + ], + "type": "text", + "content": "pressed to similar bitrates. We observe that the proposed method not only achieves higher PSNR for the reconstructed images at reduced bitrates but also maintains consistent PSNR across stereo views, thanks to its bidirectional architecture. In comparison, the SASIC method displays a PSNR discrepancy of 0.6487 dB between views, as it compresses the right image based on the left one. The VVC codec compresses images in a predictive manner, resulting in an even larger PSNR gap of 2.5769 dB between views. We leave additional qualitative analysis to the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 520, + 269, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 269, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 269, + 533 + ], + "type": "text", + "content": "4.3 Runtime Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "In this part, we evaluate the computational complexity of various compression algorithms. To ensure a fair comparison, each method is executed on an Intel Xeon Platinum 8370C CPU. Tab. 2 reports the average runtime on the InStereo2K test set, which includes the encoding and decoding latency. We observe that our BiSIC-Fast significantly reduces the runtime compared with the proposed BiSIC. This is because we replace the spatial auto-regressive interweaving in the entropy model with a two-fold stereo-checkerboard, which simplifies the dependency structure within the latent representations. Note that BiSIC-Fast still maintains competitive performance, as demonstrated in Tab. 1. Therefore, it achieves a better trade-off between the compression runtime and performance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 479, + 297 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: BDBR values of different compression methods. Bold indicates best results, and underlined values are the second-best ones." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 479, + 297 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 479, + 297 + ], + "type": "table", + "html": "
MethodInStereo2KCityscapes
PSNRMS-SSIMPSNRMS-SSIM
HEVC [30]-18.80%-12.59%-27.46%-24.36%
MV-HEVC [31]-7.69%2.14%-18.02%-17.13%
VVC [9]-35.31%-31.05%-56.25%-44.04%
HESIC+ [13]-14.96%-43.22%-23.83%-50.79%
SASIC [37]-18.40%-23.87%-21.47%-29.78%
BCSIC [20]-41.22%-54.67%-42.62%-60.72%
LDMIC [41]-41.95%-58.98%-41.92%-61.90%
ECSIC [36]-43.71%-55.65%-52.06%-64.96%
BiSIC (ours)-48.07%-61.13%-57.49%-67.98%
BiSIC-Fast (ours)-45.35%-59.36%-51.96%-65.56%
", + "image_path": "b4f0bee8356dc562677d9781fd5b9be6a799d79226d457e9d865b5d053cfaf7d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 305, + 479, + 398 + ], + "blocks": [ + { + "bbox": [ + 134, + 305, + 479, + 398 + ], + "lines": [ + { + "bbox": [ + 134, + 305, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 134, + 305, + 479, + 398 + ], + "type": "image", + "image_path": "e580b55551af116816dc0b6065fc193a99fd5a3bf27e0326ef507b11aaf4c7d7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 405, + 479, + 417 + ], + "lines": [ + { + "bbox": [ + 132, + 405, + 479, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 405, + 479, + 417 + ], + "type": "text", + "content": "Fig. 7: Qualitative comparison of image reconstruction quality across various codecs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 441, + 238, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 238, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 238, + 453 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "text", + "content": "We conduct ablation experiments on the InStereo2K dataset to analyze the contribution of the proposed modules. Our model comprises several key components, including the 3D convolution backbone, cross-dimensional entropy model, and mutual attention block. To evaluate their impact, we replace each module with a baseline counterpart or remove it from the model, and compare the performance in terms of PSNR. The experimental results are shown in Fig. 8 and Tab. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": "Effectiveness of 3D Convolution Backbone. To assess the capability of 3D convolutional layers in learning inter-view features, we replace 3D convolutional layers in codec with traditional 2D convolutional layers. The resultant model, referred to as Baseline (2D-Conv), experiences a PSNR degradation of 0.3218 dB at the same bitrate, as shown in Fig. 8 and Tab. 3. This indicates that the 3D convolutional layers are more effective at capturing the spatio-stereo information inherent in the data and reducing the redundancies between views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of Cross-Dimensional Entropy Model. To validate the performance improvement offered by our proposed cross-dimensional entropy model, we compare it with the joint auto-regressive and hyperprior entropy model [24], which we refer to as Baseline (Minnen). As shown in Fig. 8 and Tab. 3, the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "text", + "content": "BiSIC: Bidirectional Stereo Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 136, + 478, + 170 + ], + "blocks": [ + { + "bbox": [ + 208, + 114, + 404, + 125 + ], + "lines": [ + { + "bbox": [ + 208, + 114, + 404, + 125 + ], + "spans": [ + { + "bbox": [ + 208, + 114, + 404, + 125 + ], + "type": "text", + "content": "Table 2: Average runtime of different methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 136, + 478, + 170 + ], + "lines": [ + { + "bbox": [ + 136, + 136, + 478, + 170 + ], + "spans": [ + { + "bbox": [ + 136, + 136, + 478, + 170 + ], + "type": "table", + "html": "
CodecBPG HEVC VVC SASIC BCSIC LDMIC BiSIC (ours)BiSIC-Fast (ours)
Runtime16.17s 28.16s 190.27s 20.24s 89.44s 81.35s 167.25s22.82s
", + "image_path": "4c3eb28c4105c778c0e615163de62e3a4bd195564e91e4188eeedff4fa3b5d70.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 160, + 186, + 288, + 289 + ], + "blocks": [ + { + "bbox": [ + 160, + 186, + 288, + 289 + ], + "lines": [ + { + "bbox": [ + 160, + 186, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 160, + 186, + 288, + 289 + ], + "type": "image", + "image_path": "5b612df122bf7c75ea5d93b798d3d1c872b161de567c77fb5223150abfcf1a09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 299, + 324, + 321 + ], + "lines": [ + { + "bbox": [ + 132, + 299, + 324, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 324, + 321 + ], + "type": "text", + "content": "Fig. 8: Rate-distortion curves of different ablation baselines on InStereo2K." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 333, + 201, + 481, + 292 + ], + "blocks": [ + { + "bbox": [ + 333, + 201, + 481, + 292 + ], + "lines": [ + { + "bbox": [ + 333, + 201, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 333, + 201, + 481, + 292 + ], + "type": "table", + "html": "
MethodBD-PSNR
BiSIC (ours)0
Baseline (2D-Conv)-0.3218 dB
Baseline (Minnen)-0.3475 dB
Baseline (Ying)-0.6554 dB
Baseline (Lei)-0.4956 dB
Baseline (w/o Atten)-0.7882 dB
", + "image_path": "11f4d648a16ef25d8c5f106d2cf21cb1cc8741748044c9093d68b5226fe8fab3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 332, + 299, + 473, + 322 + ], + "lines": [ + { + "bbox": [ + 332, + 299, + 473, + 322 + ], + "spans": [ + { + "bbox": [ + 332, + 299, + 473, + 322 + ], + "type": "text", + "content": "Table 3: BD-PSNR [8] results relative to the proposed BiSIC." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "text", + "content": "proposed cross-dimensional entropy model achieves better rate-distortion performance than Baseline (Minnen). This suggests that our model provides more accurate probability estimations, which in turn minimizes the coding overhead. Effectiveness of Mutual Attention Block. To evaluate the effectiveness of the proposed mutual attention block, we consider three baselines for comparisons, namely Baseline (Ying), Baseline (Lei), and Baseline (w/o Atten). Baseline (Ying) and Baseline (Lei) replace the mutual attention block with the stereo attention module [39] and the contextual transfer module [20], respectively. In particular, Baseline (w/o Atten) represents a variant without the attention module. As shown in Fig. 8 and Tab. 3, our proposed model outperforms all baselines by a large margin. Notably, removing the proposed mutual attention block results in a dramatic degradation, which demonstrates its significance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In this work, we introduce a bidirectional stereo image compression network that employs 3D convolution to encode local stereo features and propose a mutual attention block to capture global features. Besides, we design a symmetric cross-dimensional entropy model that integrates hyperprior, spatial context, channel context, and stereo dependency. Our bidirectional design mitigates the issue of imbalanced compression quality intrinsic to unidirectional methods and eliminates the requirement for sequential coding. As demonstrated in the experimental results, our BiSIC achieves greater bit savings at the same level of compression quality compared to both traditional coding standards and existing learning-based methods. Moreover, we propose a fast variant that significantly reduces the compression runtime while maintaining competitive performance." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 479, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 479, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 479, + 162 + ], + "type": "text", + "content": "This work was supported by the General Research Fund (Project No. 16209622) from the Hong Kong Research Grants Council." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 201, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 201, + 479, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 479, + 222 + ], + "type": "text", + "content": "1. Arena, F., Collotta, M., Pau, G., Termine, F.: An overview of augmented reality. Computers 11(2), 28 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "text", + "content": "2. Balle, J., Chou, P.A., Minnen, D., Singh, S., Johnston, N., Agustsson, E., Hwang, S.J., Toderici, G.: Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing 15(2), 339-353 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 255, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 277 + ], + "type": "text", + "content": "3. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. arXiv preprint arXiv:1611.01704 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "text", + "content": "4. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 330 + ], + "type": "text", + "content": "5. Bao, W., Wang, W., Xu, Y., Guo, Y., Hong, S., Zhang, X.: Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences 63, 1-11 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 331, + 461, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 461, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 461, + 341 + ], + "type": "text", + "content": "6. Bellard, F.: BPG image format. Website (2014), https://bellard.org/bpg/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 342, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 361 + ], + "type": "text", + "content": "7. Bishop, C.M.: Latent variable models. In: Learning in graphical models, pp. 371-403. Springer (1998)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 362, + 480, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 362, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 138, + 362, + 480, + 383 + ], + "type": "text", + "content": "8. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33 (2001)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 384, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 426 + ], + "type": "text", + "content": "9. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology 31(10), 3736-3764 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 427, + 480, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 427, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 134, + 427, + 480, + 470 + ], + "type": "text", + "content": "10. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 471, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 471, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 134, + 471, + 480, + 514 + ], + "type": "text", + "content": "11. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 134, + 514, + 480, + 546 + ], + "type": "text", + "content": "12. Deng, X., Deng, Y., Yang, R., Yang, W., Timofte, R., Xu, M.: MASIC: Deep mask stereo image compression. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 547, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 547, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 134, + 547, + 480, + 590 + ], + "type": "text", + "content": "13. Deng, X., Yang, W., Yang, R., Xu, M., Liu, E., Feng, Q., Timofte, R.: Deep homography for efficient stereo image compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1492-1501 (June 2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 632 + ], + "type": "text", + "content": "14. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5718-5727 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "15. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14771-14780 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "text", + "content": "BiSIC: Bidirectional Stereo Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Hu, Y., Yang, W., Liu, J.: Coarse-to-fine hyper-prior modeling for learned image compression. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11013-11020 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 184 + ], + "type": "text", + "content": "17. Jiang, W., Yang, J., Zhai, Y., Ning, P., Gao, F., Wang, R.: Mlic: Multi-reference entropy model for learned image compression. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7618-7627 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "text", + "content": "18. Kadaikar, A., Dauphin, G., Mokraoui, A.: Joint disparity and variable size-block optimization algorithm for stereoscopic image compression. Signal Processing: Image Communication 61, 1-8 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "text", + "content": "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "text", + "content": "20. Lei, J., Liu, X., Peng, B., Jin, D., Li, W., Gu, J.: Deep stereo image compression via bi-directional coding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19669-19678 (June 2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "text", + "content": "21. Liu, J., Wang, S., Urtasun, R.: DSIC: Deep stereo image compression. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "text", + "content": "22. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14388-14397 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "text", + "content": "23. Liu, Y., Ren, J., Zhang, J., Liu, J., Lin, M.: Visually imbalanced stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2029-2038 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "24. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "text", + "content": "25. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3339-3343. IEEE (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "text", + "content": "26. Pan, G., Lu, G., Hu, Z., Xu, D.: Content adaptive latents and decoder for neural image compression. In: European Conference on Computer Vision. pp. 556-573. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "text", + "content": "27. Shen, Z., Zhang, M., Zhao, H., Yi, S., Li, H.: Efficient attention: Attention with linear complexities. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 3531-3539 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "text", + "content": "28. Skodras, A., Christopoulos, C., Ebrahimi, T.: The jpeg 2000 still image compression standard. IEEE Signal processing magazine 18(5), 36-58 (2001)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "text", + "content": "29. Song, X., Zhao, X., Fang, L., Hu, H., Yu, Y.: Edge stereo: An effective multi-task learning network for stereo matching and edge detection. International Journal of Computer Vision 128, 910-930 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "text", + "content": "30. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (HEVC) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "31. Tech, G., Chen, Y., Müller, K., Ohm, J.R., Vetro, A., Wang, Y.K.: Overview of the multiview and 3d extensions of high efficiency video coding. IEEE Transactions on Circuits and Systems for Video Technology 26(1), 35-49 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 423 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Vetro, A., Wiegand, T., Sullivan, G.J.: Overview of the stereo and multiview video coding extensions of the h. 264/mpeg-4 avc standard. Proceedings of the IEEE 99(4), 626-642 (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 171 + ], + "type": "text", + "content": "34. Wallace, G.K.: The JPEG still picture compression standard. Communications of the ACM 34(4), 30-44 (1991)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "35. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402. IEEE (2003)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "36. Wödlinger, M., Kotera, J., Keglevic, M., Xu, J., Sablatnig, R.: ECSIC: Epipolar cross attention for stereo image compression. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3436-3445 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "text", + "content": "37. Wödlinger, M., Kotera, J., Xu, J., Sablatnig, R.: SASIC: Stereo image compression with latent shifts and stereo attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 661-670 (June 2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "text", + "content": "38. Yin, H., Wang, Y., Tang, L., Ding, X., Huang, S., Xiong, R.: 3d lidar map compression for efficient localization on resource constrained vehicles. IEEE transactions on intelligent transportation systems 22(2), 837-852 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "39. Ying, X., Wang, Y., Wang, L., Sheng, W., An, W., Guo, Y.: A stereo attention module for stereo image super-resolution. IEEE Signal Processing Letters 27, 496-500 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "40. Zhai, Y., Tang, L., Ma, Y., Peng, R., Wang, R.: Disparity-based stereo image compression with aligned cross-view priors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2351-2360 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 381, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 481, + 402 + ], + "type": "text", + "content": "41. Zhang, X., Shao, J., Zhang, J.: LDMIC: Learning-based distributed multi-view image coding. arXiv preprint arXiv:2301.09799 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 423 + ], + "type": "text", + "content": "42. Zhu, Y., Yang, Y., Cohen, T.: Transformer-based transform coding. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 254, + 91, + 447, + 102 + ], + "type": "text", + "content": "BiSIC: Bidirectional Stereo Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_content_list.json b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..475609301d0382c699403f184444253f36468e90 --- /dev/null +++ b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_content_list.json @@ -0,0 +1,1592 @@ +[ + { + "type": "text", + "text": "Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen-Chen Zong $^{1}$ , Ye-Wen Wang $^{1}$ , Kun-Peng Ning $^{2}$ , Hai-Bo Ye $^{1}$ , and Sheng-Jun Huang $^{1}$ †", + "bbox": [ + 225, + 210, + 777, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanjing University of Aeronautics and Astronautics \n $^{2}$ Peking University", + "bbox": [ + 320, + 253, + 679, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Active learning (AL) in open set scenarios presents a novel challenge of identifying the most valuable examples in an unlabeled data pool that comprises data from both known and unknown classes. Traditional methods prioritize selecting informative examples with low confidence, with the risk of mistakenly selecting unknown-class examples with similarly low confidence. Recent methods favor the most probable known-class examples, with the risk of picking simple already mastered examples. In this paper, we attempt to query examples that are both likely from known classes and highly informative, and propose a Bidirectional Uncertainty-based Active Learning (BUAL) framework. Specifically, we achieve this by first pushing the unknown class examples toward regions with high-confidence predictions, i.e., the proposed Random Label Negative Learning method. Then, we propose a Bidirectional Uncertainty sampling strategy by jointly estimating uncertainty posed by both positive and negative learning to perform consistent and stable sampling. BUAL successfully extends existing uncertainty-based AL methods to complex open-set scenarios. Extensive experiments on multiple datasets with varying openness demonstrate that BUAL achieves state-of-the-art performance. The code is available at this link.", + "bbox": [ + 261, + 319, + 743, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Active learning $\\cdot$ Open-set annotation $\\cdot$ Negative learning $\\cdot$ Uncertainty estimation", + "bbox": [ + 261, + 595, + 740, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 648, + 375, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Labeling data can be costly and time-consuming, often requiring high levels of expertise from annotators [26]. This expense poses a significant challenge when dealing with insufficient labeled data in deep learning tasks. Recently, active learning (AL) has emerged as a prominent approach to tackle this issue and has gained widespread attention [9,17,21]. It iteratively selects the most informative examples from the unlabeled data pool and queries their labels from an oracle, enabling the learning of an effective model with reduced labeling costs.", + "bbox": [ + 212, + 680, + 787, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing AL methods [5,8,19,22,27,29,30] typically operate under the closed-set assumption, assuming that the label categories in the unlabeled data pool", + "bbox": [ + 212, + 786, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author: huangsj@nuaa.edu.cn", + "bbox": [ + 230, + 825, + 547, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "match those of the target task. However, this assumption often does not hold in practical scenarios. For example, consider a task that involves classifying images into two target categories, \"Dog\" and \"Cat\". Collecting training examples through keyword-based image search inevitably introduces irrelevant images from other categories (i.e., unknown classes), alongside the two target categories (i.e., known classes).", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In such open-set scenarios, many previous AL methods, which prefer querying examples with less confident predictions, may lead to failure since examples from unknown classes often receive uncertain predictions. To mitigate the impact of unknown class examples, some AL methods designed specifically for open-set scenarios attempt to query examples that are more likely to belong to known classes based on sample similarity [3] or model-predicted max activation value (MAV) [20]. However, since examples similar to the labeled ones may be already mastered by the model, they do not significantly benefit the target model. These methods only perform well when the proportion of unknown class examples is high. When the proportion is low, they tend to perform poorly than traditional AL methods, even inferior to random sampling (please refer to Figure 6). However, determining the proportion of unknown class examples in practical scenarios is often challenging. This further limits the usability of these methods.", + "bbox": [ + 212, + 237, + 784, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we start with a specific question: can we effectively distinguish the \"informative\" examples of known classes from examples of unknown classes? Intuitively, if we can push the unknown class examples toward regions with high-confidence predictions, existing uncertainty-based AL methods can be applied directly in open-set scenarios. To achieve this, we propose to fine-tune the model by performing negative learning (NL) [11, 14, 16, 32] on unlabeled examples. NL is an indirect learning manner that explores the utility of complementary labels, i.e., the label categories that an instance does not belong to. For a $K$ -classification problem, the NL loss is defined as:", + "bbox": [ + 212, + 433, + 785, + 568 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {N L} (f, \\bar {y}) = - \\sum_ {k = 1} ^ {K} \\bar {y} _ {k} \\log \\left(1 - p _ {k}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 574, + 784, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $f$ is the model we want to optimize, $\\bar{y}$ and $\\bar{\\pmb{y}} = [\\bar{y}_1,\\dots ,\\bar{y}_k,\\dots ,\\bar{y}_K]$ represent a complementary label and its corresponding one-hot form, respectively, and $p_k$ denotes the probability of the $k$ -th category.", + "bbox": [ + 212, + 598, + 782, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, the fine-tuning process comprises two parts. On one hand, for already labeled examples, we train them directly using Equation 1. On the other hand, for unlabeled examples, we first randomly assign labels to them in each training round and then train the model using Equation 1. Notably, the unlabeled known class example has a relatively higher chance of receiving the correct label, whereas the unknown class example will never be assigned the correct label. Once unlabeled known class data receive the correct labels, they suffer a larger penalty and are reduced confidence predictions by the model since they deviate from the distribution information obtained from labeled data. In contrast, unlabeled unknown class data are not constrained to move towards the high-confidence region for counteracting the update gradient produced by the labeled data.", + "bbox": [ + 212, + 643, + 785, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate this, we conducted preliminary experiments on CIFAR-10 [12] with 4 known and 6 unknown classes. Figure 1 illustrates the confidence statis", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Zong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/51f0717bad27da9b553adaa3ae6338739ea5d9c077930a51d622e0c5caf020d9.jpg", + "image_caption": [ + "(a) w/o NL" + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 483, + 309 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/65d474dc8457d970367c7a84e1887c2500e10c7d564a4059ad7c0465b076f113.jpg", + "image_caption": [ + "(b) w/ NL", + "Fig. 1: The statistics of prediction confidence before and after fine-tuning the model. In the zoomed-in area of Figure 1b, we swapped the display order of the two to prevent occlusion, allowing for a more intuitive view of how the distribution has changed." + ], + "image_footnote": [], + "bbox": [ + 517, + 146, + 779, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tics for all unlabeled examples in a fixed query round before and after fine-tuning the model. As expected, known class examples are more prevalent in the low-confidence region, while unknown class ones are more common in high-confidence regions. This distribution characteristic presents a potential solution to the aforementioned question and offers a promising approach to AL for open-set scenarios.", + "bbox": [ + 212, + 409, + 784, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this, we propose a Bidirectional Uncertainty-based Active Learning framework (BUAL). On the one hand, we propose the Random Label Negative Learning (RLNL) method to fine-tune the model and leverage information from the unlabeled data pool. Specifically, we first train a model using all labeled data as a positive classifier. Then, negative learning is performed to fine-tune the model as a negative classifier by randomly assigning labels to data from the unlabeled pool in each training iteration. This effectively distinguishes known class examples with lower confidence predictions from unknown class ones. On the other hand, we propose a Bidirectional Uncertainty (BU) sampling strategy for active selection, which estimates prediction uncertainty from both positive and negative classifiers. By selecting examples with the highest uncertainty, we expect to identify the most informative instances from the known classes. Experiments are performed on multiple datasets with different unknown class ratios. The results demonstrate that BUAL can query more informative known class examples and that the model performance obtained by BUAL is substantially improved compared with existing state-of-the-art methods.", + "bbox": [ + 212, + 484, + 785, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 748, + 387, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Active learning (AL) is a prominent approach aimed at reducing label costs by selecting a batch of examples that are most valuable for model training. Existing AL methods can be broadly categorized into three groups based on sample selection strategies: uncertainty-based, representative-based, and hybrid strategies", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BUAL for Open-Set Annotation", + "bbox": [ + 516, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which combine both aspects. Uncertainty-based strategies focus on sampling informative instances to reduce model uncertainty. Typical methods include Least Confident Sampling [15], Margin-based Sampling [2], and Entropy-based Sampling [7], etc. Representative-based strategies start from the sample distribution and aim to select representative instances that match the overall distribution. A typical method is Coreset [25]. Hybrid strategies combine uncertainty and representativeness by incorporating sample distribution information and the model's specific needs. Notable methods in this category include QUIRE [8] and BADGE [1], etc.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-set annotation (OSA) involves active learning under open-set scenarios [3,20]. Existing methods primarily focus on selecting examples that are most likely to belong to known classes. For instance, CCAL [3] employs contrastive learning to extract semantic and distinctive features of examples, facilitating discrimination of known class examples. LfOSA [20] introduces an auxiliary network to model the per-example max activation value (MAV) distribution and dynamically selects examples with the highest probability from known classes. However, these methods exhibit sensitivity to the openness of the dataset and may not consistently perform well. Open-set recognition (OSR) [18, 23, 24] is a related problem setting to OSA, aiming to predict correct labels for known class examples while simultaneously detecting examples from unknown classes. Nevertheless, direct use of OSR methods often falls short of expected effectiveness [20], mainly due to limited training examples and the inability to identify highly informative instances.", + "bbox": [ + 212, + 282, + 787, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Complementary labels are labels other than the ground-truth label assigned to an example. Complementary label learning (CLL) [4,10,31] was initially introduced in [10], where the authors leveraged the lower acquisition cost and higher correct labeling rate of complementary labels to address multi-class classification problems. However, complementary labels contain less information compared to the correct labels, which can result in slow convergence of the model and challenges in achieving the desired performance when directly using complementary labels for training. Authors in [11] made the first attempt to combine CLL with the noise labeling problem [33] and proposed an indirect and robust training schema called negative learning (NL). In this paper, we exploit the properties of NL and extend it to open-set and label-free settings for the first time.", + "bbox": [ + 212, + 493, + 787, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 681, + 380, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 713, + 372, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notations. In open-set annotation (OSA) settings, there are two labeled data pools: $D_{l}^{kn o} = \\{(x_{i}^{l},y_{i}^{l})\\}_{i = 1}^{n_{l}^{kn o}}$ for known classes and $D_{l}^{unk} = \\{(x_{i}^{l})\\}_{i = 1}^{n_{l}^{unk}}$ for unknown classes, as well as an unlabeled data pool: $D_{u} = \\{(x_{i}^{u})\\}_{i = 1}^{n_{u}^{kn o}} \\cup \\{(x_{i}^{u})\\}_{i = 1}^{n_{u}^{unk}}$ containing examples from both known and unknown classes. Each known class example belongs to one of $K$ known classes in the label space $\\mathcal{V} = \\{1,2,\\dots ,K\\}$ . The unknown class examples are uniformly grouped into one category, denoted", + "bbox": [ + 212, + 737, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Zong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d52dbf4c357580650dae3fdff056a3ade9f2f20e52bdc0af2ea36d24804ef04.jpg", + "image_caption": [ + "Fig. 2: The framework of BUAL. A two-stage $K$ class classifier is maintained, where the first stage is trained in a normal manner saved as $f_{p}(\\cdot)$ and the second stage is trained using the proposed random label negative learning method denoted as $f_{n}(\\cdot)$ . An auxiliary $K + 1$ class classifier $f_{aux}(\\cdot)$ is trained in parallel. By collecting the predicted uncertainty from $f_{p}(\\cdot)$ and $f_{n}(\\cdot)$ on each candidate example along with the global and local balancing factors, the proposed bidirectional sampling strategy can accurately estimate the potential utility of each example and perform effective sample sampling under complex open-set scenarios." + ], + "image_footnote": [], + "bbox": [ + 225, + 152, + 781, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as $y = \\emptyset$ . In each training round, active learning (AL) queries a batch of $b$ examples according to a given query strategy $\\mathcal{A}$ , denoted as $X_{query} = X_{query}^{kn}\\cup X_{query}^{unk}$ . Once the labeled feedback is obtained, we can calculate a ratio $r = \\frac{|X_{query}^{kn}|}{|X_{query}|}$ to represent the precision of known classes.", + "bbox": [ + 212, + 497, + 787, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview. The proposed Bidirectional Uncertainty-based Active Learning (BUAL) framework is depicted in Figure 2. In each iteration, there are in general three steps:", + "bbox": [ + 212, + 569, + 787, + 617 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Model Training: the algorithm first trains a target classifier in a normal learning manner, dubbed positive classifier $f_{p}(\\cdot)$ . Then, the algorithm finetunes the classifier with a new classifier head by the proposed Random Label Negative Learning method, dubbed negative classifier $f_{n}(\\cdot)$ . Similar to [20], a $K + 1$ auxiliary classifier $f_{aux}(\\cdot)$ is trained in parallel.", + "- Example Selection: the Bidirectional Uncertainty Sampling Strategy estimates uncertainty bidirectionally using both the positive and negative classifier heads. This estimation is combined with dynamic balance factors generated by the auxiliary classifier and query feedback to select the most informative known class examples.", + "- Oracle Labeling: the annotators assign class labels to the selected examples. Based on the feedback results, update the corresponding data pools accordingly." + ], + "bbox": [ + 223, + 637, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BUAL for Open-Set Annotation", + "bbox": [ + 514, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Random Label Negative Learning", + "text_level": 1, + "bbox": [ + 215, + 146, + 540, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In OSA scenarios, conventional uncertainty-based AL methods tend to be ineffective for example selection, mainly due to the unconfident predictions generated for the unknown class examples. Existing OSA methods prioritize sample purity without fully exploring sample informativeness, resulting in queries with too many already mastered examples that are not useful for model training. Thus, a key question arises: how to distinguish between unknown and \"informative\" known class examples.", + "bbox": [ + 212, + 172, + 787, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To cope with this problem, we propose a general method by pushing unknown class examples toward the high-confidence regions, while pushing known class examples to the low-confidence regions. If this separation is achieved, we can leverage existing uncertainty-based AL methods directly to handle various complex open-set scenarios. Fortunately, we achieve this to some extent by employing negative learning (NL), where we assign random labels to unlabeled examples and fine-tune the model with a new classifier head accordingly, dubbed Random Label Negative Learning (RLNL).", + "bbox": [ + 212, + 279, + 787, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, in the model training stage, we first train a $K$ -class classifier $f_{p}(\\cdot)$ in a normal training manner (e.g., cross-entropy loss) based on $D_{l}^{kn o}$ . Note that $f_{p}(\\cdot)$ can be the target model we eventually need to output. After training on the labeled data, $f_{p}(\\cdot)$ can own a good discriminative ability for known class examples at the representation level. Then, we replace the last layer of $f_{p}(\\cdot)$ and fine-tune a new classifier head $f_{n}(\\cdot)$ with Equation 1. In this phase, unlabeled examples in $D_{u}$ are involved in training to help achieve the goal mentioned above: pushing unknown and known class examples toward the high and low confidence regions, respectively. All labeled examples in $D_{l}^{kn o}$ are still involved to prevent them from being incorrectly shifted in the distribution. Eventually, we assign random labels to the unlabeled examples in $D_{u}$ , while using the complementary labels instead for the labeled known class examples, i.e., for an example $\\mathbf{x}$ :", + "bbox": [ + 212, + 400, + 787, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP (\\bar {y} = s) = \\frac {1}{| S |}, s \\in \\left\\{ \\begin{array}{l l} S = \\mathcal {Y} \\backslash y ^ {l}, & \\text {i f} \\boldsymbol {x} \\in D _ {l} ^ {\\text {k n o}}, \\\\ S = \\mathcal {Y}, & \\text {i f} \\boldsymbol {x} \\in D _ {u}. \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 593, + 785, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\bar{y}$ is uniformly sampled at each training iteration. Here, considering that the number of examples in $D_{u}$ is usually much larger than that in $D_{l}^{kn o}$ , a small subset $D_{sub}$ is randomly selected from the $D_{u}$ to save training cost.", + "bbox": [ + 212, + 643, + 787, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Why RLNL works? The first question one may hold is why it shouldn't be optimal for the unlabeled examples to be the uniform distribution as $\\bar{y}$ is sampled at each iteration. To explain this, we take two different kinds of updating for a single example in the binary classification scenario of \"Dog vs. Cat\" as illustrations: negative learning using all labels at each iteration and negative learning using only one random label at each iteration. One possible case is shown in Figure 3. If each iteration uses all labels for negative learning, i.e., such an example is neither a \"Cat\" nor a \"Dog\", it is apparent that an update gradient of 0 is only achieved when the example is at the decision boundary. In contrast, if each iteration gives only one random label for negative learning,", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Zong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77a46191c36b5738078276f88203ac4bc75aa78105bde38a1319bd1fd167ef0c.jpg", + "image_caption": [ + "Fig. 3: Use all labels per iteration for negative learning (left) vs. use one random label per iteration for negative learning (right)." + ], + "image_footnote": [], + "bbox": [ + 264, + 146, + 519, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4abca5303e80168f9e3a88dac26c47d8499efa862d5c582f5612b56a0ee9d42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 148, + 728, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/388cb1c695eac5e43f954282bcbc200ba686f40c136791cfc6e29a952bf3917c.jpg", + "image_caption": [ + "Fig. 4: The possible RLNL update scenario for unlabeled unknown class data in batch deep learning manner. The green \"→\" is the batch update gradient produced by the example itself, and the purple \"→\" is the update gradient produced by labeled data. Initially, the decision boundary is close to the left-hand category." + ], + "image_footnote": [], + "bbox": [ + 300, + 364, + 705, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "i.e., such an example is not a \"Cat\" or a \"Dog\", then for an instance that is not a \"Cat\", pushing it as much as possible towards the \"Dog\" side will result in a gradient update of 0, and vice versa. Obviously, there is no fixed optimal scenario for this type of data.", + "bbox": [ + 212, + 625, + 784, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In RLNL, we utilize this property by exploiting the prior knowledge contained in the labeled data. For unlabeled known class data, they might have overlapping regions with labeled data in the feature space, benefiting from the feature representations learned in the previous stage and the simultaneous introduction of labeled data for negative learning. The mapping of such examples in the feature space will remain or be close to the labeled ones in subsequent model updates, as they will receive invisible constraints provided by the prior knowledge from the labeled data, owing to similar features.", + "bbox": [ + 212, + 686, + 787, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 4, we present the possible update scenario for unlabeled unknown class examples in the commonly adopted batch deep learning manner.", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BUAL for Open-Set Annotation", + "bbox": [ + 514, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d965843017f0d70958e9f3b25f3de37e51e290cea9d29995474d94d3865322d.jpg", + "image_caption": [ + "(a) Before RLNL" + ], + "image_footnote": [], + "bbox": [ + 228, + 152, + 464, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8c05da7c5bcf0670046fd643aa281d966411bfa6b566d780a7e2f828adefabb.jpg", + "image_caption": [ + "(b) After RLNL", + "Fig. 5: The t-SNE feature visualization of labeled data, unlabeled known class data, and unlabeled unknown class data on CIFAR-10 with an openness ratio of 0.5 before and after performing RLNL. For a more intuitive visualization, we only show a single known class. More visualization results are shown in the supplementary file." + ], + "image_footnote": [], + "bbox": [ + 488, + 148, + 782, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, the green arrow indicates the batch update gradient produced by itself, while the purple arrow indicates the batch update gradient produced by labeled data. Different from Figure 3, where data oscillates on both sides of the decision boundary, the unlabeled unknown class examples will oscillate at uncharted away from the decision boundary to counteract the update gradient due to the labeled ones. In contrast, unlabeled known class examples will move much less than unknown class ones in magnitude in the feature space within the constraints of the prior knowledge provided by the labeled data. This is further confirmed in Figure 5, which illustrates the t-SNE feature visualization of labeled data, unlabeled known class data, and unlabeled unknown class data before and after performing RLNL. This ultimately leads to the result in Figure 1b and proves that RLNL does work.", + "bbox": [ + 212, + 388, + 787, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Bidirectional Uncertainty Sampling Strategy", + "text_level": 1, + "bbox": [ + 214, + 589, + 630, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the fine-tuning process, we observed that the predictions of $f_{n}(\\cdot)$ for examples in $D_{u}$ tend to oscillate between epochs. To ensure stable sampling, on one hand, we test all unlabeled examples $t$ times at $m$ round intervals to obtain the predicted probabilities $\\pmb{p}_{\\pmb{t}}^{-} = (p_{1}^{t},\\dots,p_{K}^{t})$ , which will further be averaged as $\\mathcal{P}^{-} = \\frac{1}{t}\\sum_{i = 1}^{t}\\pmb{p}_{i}^{-} = (\\frac{1}{t}\\sum_{i = 1}^{t}p_{1}^{i},\\dots,\\frac{1}{t}\\sum_{i = 1}^{t}p_{K}^{i})$ . On the other hand, we reuse the predictions of $f_{p}(\\cdot)$ to produce predicted probabilities $\\pmb{p}^{+} = (p_{1},\\dots,p_{K})$ for all unlabeled examples, as $f_{p}(\\cdot)$ are accurate for measuring known class ones.", + "bbox": [ + 212, + 611, + 787, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared to the positive head, the negative head is slightly biased for the measurement of sample uncertainty due to the unstable training. Therefore, if an example is likelier to belong to known classes, we prefer to utilize the sample uncertainty obtained from $f_{p}(\\cdot)$ . On the contrary, once an example has a higher risk of belonging to the unknown classes, the uncertainty obtained from $f_{p}(\\cdot)$ is unreliable, and thus the uncertainty by $f_{n}(\\cdot)$ should be given a higher weight.", + "bbox": [ + 212, + 719, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To achieve this, we introduce two balancing factors based on global and local observations. In open-set scenarios, some unknown class examples might be", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Zong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mistakenly selected for annotation. Although these examples cannot be directly used for training the target model, they are valuable for measuring sample purity. Therefore, similar to [20], we train a $K + 1$ classifier $f_{aux}(\\cdot)$ in a normal training manner based on the examples from both $D_l^{kno}$ and $D_l^{unk}$ . Then, we can obtain the predicted probability $p_{K + 1}^{aux}(\\boldsymbol{x}) = f_{aux}(\\emptyset \\mid \\boldsymbol{x})$ for each example. This can act as a local balancing factor, since the larger the value of $p_{K + 1}^{aux}(\\boldsymbol{x})$ the more likely $\\boldsymbol{x}$ is to belong to the unknown class.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Additionally, once the selected examples are sent to the oracle for annotation, we can calculate the ratio $r$ , which provides a rough estimate of the current openness of $D_{u}$ and can serve as a global balancing factor. With the two balancing factors, we propose a bidirectional uncertainty sampling strategy defined as follows:", + "bbox": [ + 212, + 252, + 782, + 325 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {*} = \\arg \\max _ {\\boldsymbol {x}} p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\operatorname {u n c} _ {n} + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\operatorname {u n c} _ {p}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 327, + 785, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $unc_{p}$ and $unc_{n}$ denote the uncertainty of $\\mathbf{x}$ to the positive classifier head and the negative classifier head, respectively. Noteworthy, this sampling strategy remains applicable even in closed-set settings. If there are no unknown class examples, the ratio $r$ will always be equal to 1, and $p_{K + 1}^{aux}(\\mathbf{x})$ will be 0, effectively making the strategy equivalent to a normal uncertainty sampling strategy.", + "bbox": [ + 212, + 352, + 782, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With Equation 3, extending the existing closed-set uncertainty-based active learning methods to open-set scenarios is possible. In this paper, we focus on three classical uncertainty sampling strategies: least confident sampling, margin-based sampling, and entropy-based sampling. The corresponding modified versions for open-set scenarios are as follows:", + "bbox": [ + 212, + 428, + 782, + 502 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Bidirectional Least Confident Sampling", + "bbox": [ + 225, + 512, + 571, + 527 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\arg \\max } p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\left[ 1 - \\boldsymbol {\\mathcal {P}} _ {y ^ {-}} ^ {-} (\\boldsymbol {x}) \\right] + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\left[ 1 - \\boldsymbol {\\mathcal {P}} _ {y ^ {+}} ^ {+} (\\boldsymbol {x}) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 536, + 784, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $y^{-} = \\arg \\max_{y}\\mathcal{P}_{y}^{-}(\\pmb {x})$ $y^{+} = \\arg \\max_{y}\\pmb{p}_{y}^{+}(\\pmb {x})$", + "bbox": [ + 214, + 569, + 589, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Bidirectional Margin-Based Sampling", + "bbox": [ + 225, + 595, + 555, + 609 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\arg \\max } p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\left[ \\boldsymbol {\\mathcal {P}} _ {y _ {1} ^ {-}} ^ {-} (\\boldsymbol {x}) - \\boldsymbol {\\mathcal {P}} _ {y _ {2} ^ {-}} ^ {-} (\\boldsymbol {x}) \\right] \\tag {5} \\\\ + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\left[ \\boldsymbol {p} _ {y _ {1} ^ {+}} ^ {+} (\\boldsymbol {x}) - \\boldsymbol {p} _ {y _ {2} ^ {+}} ^ {+} (\\boldsymbol {x}) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 619, + 784, + 674 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $y_1^- = \\arg \\max_y \\mathcal{P}_y^-(\\boldsymbol{x})$ , $y_1^+ = \\arg \\max_y \\mathbf{p}_y^+(\\boldsymbol{x})$ , $y_2^- = \\arg \\max_{y \\setminus y_1^-} \\mathcal{P}_y^-(\\boldsymbol{x})$ , $y_2^+ = \\arg \\max_{y \\setminus y_1^+} \\mathbf{p}_y^+(\\boldsymbol{x})$ .", + "bbox": [ + 214, + 679, + 782, + 719 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Bidirectional Entropy-Based Sampling", + "bbox": [ + 225, + 724, + 563, + 739 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\arg \\max } p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\left[ - \\boldsymbol {\\mathcal {P}} _ {y ^ {-}} ^ {-} (\\boldsymbol {x}) \\log \\boldsymbol {\\mathcal {P}} _ {y ^ {-}} ^ {-} (\\boldsymbol {x}) \\right] \\tag {6} \\\\ + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\left[ - \\boldsymbol {p} _ {y ^ {+}} ^ {+} (\\boldsymbol {x}) \\log \\boldsymbol {p} _ {y ^ {+}} ^ {+} (\\boldsymbol {x}) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 750, + 784, + 805 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $y^{-}$ and $y^{+}$ are consistent with the previous definition.", + "bbox": [ + 214, + 809, + 651, + 824 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The main procedures of BUAL are summarized in Algorithm 1.", + "bbox": [ + 238, + 824, + 697, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BUAL for Open-Set Annotation", + "bbox": [ + 514, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 BUAL Training Procedure" + ], + "code_body": "Input: $D_l^{kno}$ $D_{l}^{unk}$ $D_u$ $r$ $\\mathcal{A},b,m,t,$ subset size s, training epoch $E$ , optimizer $\\mathcal{O}$ \nOutput: $D_l^{kno}$ $D_{l}^{unk}$ $D_u$ $r$ , model parameters $\\theta_p$ \nProcess: \n1: # Model training \n2: for $j = 1;j\\leq E$ do \n3: $\\mathcal{L}_p = \\sum_{(x,y)\\in D_l^{kno}}\\ell (f_p(\\boldsymbol {x}),y);$ \n4: $\\mathcal{L}_{aux} = \\sum_{(x,y)\\in D_l^{kno}\\cup D_l^{unk}}\\ell (f_{aux}(\\boldsymbol {x}),y);$ \n5: Update $\\theta_p,\\theta_{aux} = \\mathcal{O}(\\mathcal{L}_p,\\theta_p),\\mathcal{O}(\\mathcal{L}_{aux},\\theta_{aux})$ \n6: end for \n7: Use $f_{aux}(\\pmb {x})$ to remove confidently unknown examples in $D_{u}$ and randomly select s examples as $D_{sub}$ \n8: for $j = 1;j\\leq E$ do \n9: Generate label $\\bar{y}$ for each example. \n10: $\\mathcal{L}_n = \\sum_{(x,y)\\in D_l^{kno}\\cup D_{sub}}\\ell_{NL}(f_n(\\boldsymbol {x}),\\bar{y})$ \n11: Update $\\theta_{n} = \\mathcal{O}(\\mathcal{L}_{n},\\theta_{n})$ \n12: if $j$ mod $m = 0$ then \n13: Obtain $\\pmb {p}_j^-$ for each sample in $D_{u}$ by $f_{n}(\\pmb {x})$ \n14: end if \n15: end for \n16: # Example selection \n17: Calculate $\\pmb{\\mathcal{P}}^{-}$ for each sample across rounds; \n18: Obtain $\\pmb{p}^{+}$ for each sample in $D_u$ by $f_{p}(\\pmb {x})$ \n19: Obtain $p_{K + 1}^{aux}$ for each sample in $D_u$ by $f_{aux}(\\pmb {x})$ \n20: Obtain $X_{query}$ using BU sampling strategy A; \n21: # Oracle labeling \n22: Ask for annotation and update $D_l^{kno}$ $D_l^{unk}$ $D_u$ and $r$", + "bbox": [ + 215, + 162, + 785, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 566, + 374, + 584 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on three benchmark datasets: CIFAR-10, CIFAR-100 [12], and Tiny-Imagenet [28]. Tiny-Imagenet is a subset of theImagenet [13], consisting of 200 classes with 500 training images per class. The openness of each dataset is defined as the ratio of unknown classes to the total number of classes, and we set its value to 0.2, 0.4, 0.6, and 0.8 for all datasets.", + "bbox": [ + 212, + 597, + 784, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparing methods. We select nine AL strategies for comparison, which can be further categorized into six groups: (1) Random: Randomly select examples from the unlabeled data pool for labeling. (2) Traditional uncertainty-based strategies: Least confident sampling (LC), Margin-based sampling (Margin), and Entropy-based sampling (Entropy). (3) Diversity-based strategy: Coreset. (4) Hybrid-based strategy: BADGE. (5) OSA methods: CCAL and LfOSA. (6) OSR method: DIAS [18]. Correspondingly, our methods are Bidirectional Least confident sampling (B-LC), Bidirectional Margin-based sampling (B-Margin), and Bidirectional Entropy-based sampling (B-Entropy).", + "bbox": [ + 212, + 674, + 785, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training details. On CIFAR-10, CIFAR-100, and Tiny-Imagenet, we randomly sample $1\\%$ , $8\\%$ , and $8\\%$ known class examples as the initial labeled data,", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Zong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cff1b691b242530b6d54613ab4b2921b1c28f5f0241bc7869989d239ebef8566.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 146, + 498, + 277 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/85e63ded2dd2fc13ac3e4569a550eb88de88b58420c269d8dd580b906e238166.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 146, + 746, + 277 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/371e5f2bc69e03e258b3cbd9aba53ae85296a2738e202810c37223d0fc0faa79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 282, + 496, + 412 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/40fa1eff4b66cabb8bbbc4d41e5c0b22cb79c7c2c92f8918f05a79ca264591f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 282, + 745, + 412 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/76d182aed50eab9644a1a43d8ab731d35d90b694ea85d9a8ec1fb1b184fc490f.jpg", + "image_caption": [ + "Fig. 6: Accuracy comparison on CIFAR-10 (first row), CIFAR-100 (second row), and Tiny-Imagenet (third row). The ratio of unknown class examples to the total number of examples is fixed at 0.4 (first column) and 0.6 (second column) for each dataset." + ], + "image_footnote": [], + "bbox": [ + 256, + 417, + 496, + 546 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/08d914508efc827b7fe8f1f2b24e5805f5a37b4d25792b2f5b5fd5c1b15e60c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 417, + 746, + 545 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "respectively. All the models involved in the experiments are ResNet18 [6], trained for 100 epochs, using SGD as the optimizer, where the learning rate is 0.01, momentum is 0.9, weight decay is 1e-4, and batch size is 128. We perform the experiments for 3 runs and report the average results. For CIFAR-10 and CIFAR-100, 5000 examples are randomly selected from the unlabeled pool as $D_{sub}$ , and 1500 examples are queried in each query round. Due to the doubling of data volume, we randomly select 10000 examples as $D_{sub}$ and query 3000 examples in each query round for Tiny-Imagenet.", + "bbox": [ + 212, + 628, + 787, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Performance Comparison", + "text_level": 1, + "bbox": [ + 214, + 771, + 472, + 787 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Figure 6 presents the variation curves for the classification accuracy of the proposed methods and the comparison methods. Here, to better observe the variation curves, we only show the B-Margin for the proposed methods and the", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BUAL for Open-Set Annotation", + "bbox": [ + 516, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/303966f68a775627a0b786729ec511d29692e370bf2085f1d75ff851228b9bc1.jpg", + "table_caption": [ + "Table 1: The final round average accuracy of different methods on CIFAR-10, CIFAR-100, and Tiny-Imagenet. The best performance is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
DatasetsCIFAR-10CIFAR-100Tiny-Imagenet
Openness Ratio0.20.40.60.80.20.40.60.80.20.40.60.8
(1)Random83.382.587.296.957.658.358.761.245.747.250.955.0
(2)LC84.381.687.596.255.854.654.056.244.845.948.451.6
Margin86.084.189.097.059.359.658.858.946.447.150.854.0
Entropy85.483.488.096.857.156.855.756.444.644.546.950.7
(3)Coreset85.081.886.497.460.261.261.864.246.247.851.854.0
(4)BADGE86.884.289.296.460.260.860.462.046.347.851.853.3
(5)LfOSA73.778.787.098.652.356.662.468.242.546.652.459.9
CCAL80.881.588.098.155.960.064.767.744.446.350.357.0
(6)DIAS81.880.783.094.055.756.156.957.243.145.147.554.4
OursB-LC87.087.292.599.159.362.867.572.145.748.754.760.6
B-Margin86.587.092.698.960.963.168.371.546.549.555.761.2
B-Entropy86.987.492.699.158.961.766.971.445.447.555.261.0
", + "bbox": [ + 223, + 185, + 782, + 417 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c01d33cdef9aaf1c78b23ac952bb348aa46fbc7bbd8a934c89db6939ad33ff61.jpg", + "image_caption": [ + "Fig. 7: The average recognition rate on CIFAR-10 (first column), CIFAR-100 (second column), and Tiny-Imagenet (third column)." + ], + "image_footnote": [], + "bbox": [ + 225, + 436, + 406, + 551 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/25f703ff279500275e32b25baf652250e16f41c56953375ea7a9e148bde62e9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 436, + 591, + 551 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4742a081036cd031165ec35a7202c312f0dff2010cb292f003ac6494176ce33a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 436, + 776, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Margin for the compared traditional uncertainty methods. Table 1 reports all methods' final round average accuracy.", + "bbox": [ + 212, + 626, + 784, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We can observe that our proposed methods consistently achieve the highest classification accuracy, demonstrating the effectiveness and superiority of our BUAL framework over other methods. Some notable observations are as follows: 1) The OSA methods, CCAL and LfOSA, gradually lose effectiveness as the proportion of known classes in the dataset increases. In contrast, traditional uncertainty-based methods perform better as the openness ratio decreases. This is consistent with our previous analysis. 2) The performance of Coreset and BADGE also deteriorates when the openness ratio is high. These methods tend to select examples with diverse characteristics. However, unknown class examples often differ significantly in characteristics from the labeled ones, making them more likely to be sampled and thus undermining the effectiveness of this type of method. 3) The OSR method DIAS does not perform well in all situations. The", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Zong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c185748c886ed0a9e0dd312ac2ebf0b8035bdd4b7c4829722300107fe6ad15da.jpg", + "image_caption": [ + "(a) LfOSA" + ], + "image_footnote": [], + "bbox": [ + 256, + 150, + 488, + 303 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/80f7fc80398aed826af2e304857124edcaed22458b5151a7be8d4b06ff79aa35.jpg", + "image_caption": [ + "(b) BUAL (ours)", + "Fig. 8: The t-SNE feature visualization of data from one query and labeled pool on CIFAR-10 with an openness ratio of 0.5." + ], + "image_footnote": [], + "bbox": [ + 511, + 151, + 754, + 303 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "limited labeled data prevents the model from learning a robust representation for identifying open-set examples effectively. Moreover, DIAS cannot identify highly informative examples, resulting in its queries often being simple and unhelpful examples. 4) All our methods remain stable and do not suffer significantly from changes in the openness ratio. This can demonstrate two points. The dynamic balance factors in the proposed framework can adaptively assign appropriate weights to positive and negative uncertainties regardless of the openness of the dataset, and negative uncertainties are indeed effective for querying highly informative known class examples.", + "bbox": [ + 212, + 398, + 784, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 7 shows the average recognition rate of known class examples across queries and openness ratios. We can observe that traditional AL methods all perform poorly due to their inability to recognize known class examples from those with high uncertainty and/or representative. As an OSR method, DIAS shows only marginal improvement compared to the traditional AL methods, unable to fully exploit its exceptional performance with limited training data. The recognition ability of CCAL decreases slightly with the increase of dataset categories compared to our method. It is noted that LfOSA achieved the highest query precision on all datasets. However, combining the results in Table 1, we can find that the performance of models trained by LfOSA is not satisfactory compared to other methods. To explain why, we visualize the queried examples' feature representation and labeled examples' feature representation for our method and LfOSA in Figure 8. We can observe that the features of examples queried by LfOSA are highly overlapping with the labeled ones, which are often examples that the model has already mastered and cannot provide an effective bonus for model training. On the contrary, the sample features queried by our method have little overlap and are more distributed in low-density regions, consistent with the goal of AL. These results validate the effectiveness of our approach in querying more informative examples of known classes and maintaining a high recognition rate.", + "bbox": [ + 212, + 537, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BUAL for Open-Set Annotation", + "bbox": [ + 514, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/10f04b45c307a01747d4f1447b5f6d10fef408eff19bc9cfa2b797169f030fcf.jpg", + "table_caption": [ + "Table 2: Final accuracy of each component in Equation 3 on CIFAR-10, CIFAR-100, and Tiny-Imagenet with an openness ratio of 0.6." + ], + "table_footnote": [], + "table_body": "
MethodDatasetuncpuncnw/o ww/o fauxB-LC
CIFAR-1087.589.490.891.392.5
CIFAR-10054.063.562.465.067.5
Tiny-Imagenet48.452.352.053.054.7
", + "bbox": [ + 259, + 185, + 743, + 290 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.2 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 316, + 388, + 330 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The ablation study is conducted on CIFAR-10 with an openness ratio of 0.6 to validate the effectiveness of each component in our proposed query strategy (Equation 3). The final round accuracy is shown in Table 2. Here, $\\text{unc}_n$ and $\\text{unc}_p$ indicate that only $\\text{unc}_n$ and $\\text{unc}_p$ is adopted for active sampling, respectively. $w / o$ denotes the removal of all balancing factors, i.e., both $f_{aux}$ and $r$ . $w / o$ means only removing the local balancing factor $f_{aux}$ .", + "bbox": [ + 212, + 337, + 785, + 428 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We can observe that removing any of the components leads to performance degradation. Although $\\mathbf{unc}_n$ and $\\mathbf{unc}_p$ are significantly less effective than B-LC, $\\mathbf{unc}_n$ has a substantial improvement compared to $\\mathbf{unc}_p$ due to its ability to distinguish informative examples of known classes from examples of unknown classes. The direct combination of $\\mathbf{unc}_n$ and $\\mathbf{unc}_p$ , i.e., $\\mathbf{w} / \\mathbf{o} \\mathbf{w}$ , works better than $\\mathbf{unc}_n$ , as models trained by RLNL may produce oscillate output and thus the uncertainty obtained in $\\mathbf{unc}_n$ is not necessarily accurate. By adding the dynamic global balancing factor $r$ , $\\mathbf{w} / \\mathbf{o} f_{aux}$ achieves a better performance. However, it still falls short in comparison to B-LC, which validates the effectiveness of local balancing factor $f_{aux}$ . These results further corroborate the soundness of our strategy design.", + "bbox": [ + 212, + 429, + 787, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 614, + 359, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we successfully expand the existing uncertainty-based active learning methods to complex and ever-changing open-set scenarios by proposing a Bidirectional uncertainty-based Active Learning (BUAL) framework. On one hand, to achieve the goal of distinguishing known and unknown class examples with high uncertainty, we propose a simple but effective Random Label Negative Learning (RLNL) method for pushing unknown and known class examples toward the high and low confidence regions respectively. On the other hand, to better measure sample uncertainty, we propose a Bidirectional Uncertainty (BU) sampling strategy by dynamically fusing the sample uncertainty obtained from positive learning and negative learning. The dynamic balancing factors in it can ensure that the strategy is effective under various openness ratios. Extensive experimental results show that the model trained with BUAL can achieve state-of-the-art performance under various open-set scenarios.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Zong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the Natural Science Foundation of Jiangsu Province of China (BK20222012, BK20211517), the National Key R&D Program of China (2020AAA0107000), and NSFC (62222605).", + "bbox": [ + 215, + 172, + 785, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ash, J.T., Zhang, C., Krishnamurthy, A., Langford, J., Agarwal, A.: Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671 (2019)", + "2. Balcan, M.F., Broder, A., Zhang, T.: Margin based active learning. In: International Conference on Computational Learning Theory. pp. 35-50. Springer (2007)", + "3. Du, P., Zhao, S., Chen, H., Chai, S., Chen, H., Li, C.: Contrastive coding for active learning under class distribution mismatch. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8927-8936 (2021)", + "4. Feng, L., Kaneko, T., Han, B., Niu, G., An, B., Sugiyama, M.: Learning with multiple complementary labels. In: International Conference on Machine Learning. pp. 3072-3081. PMLR (2020)", + "5. Fu, Y., Zhu, X., Li, B.: A survey on instance selection for active learning. Knowledge and information systems 35(2), 249-283 (2013)", + "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "7. Holub, A., Perona, P., Burl, M.C.: Entropy-based active learning for object recognition. In: 2008 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops. pp. 1-8. IEEE (2008)", + "8. Huang, S.J., Jin, R., Zhou, Z.H.: Active learning by querying informative and representative examples. Advances in neural information processing systems 23 (2010)", + "9. Huang, S.J., Zong, C.C., Ning, K.P., Ye, H.B.: Asynchronous active learning with distributed label querying. In: IJCAI. pp. 2570-2576 (2021)", + "0. Ishida, T., Niu, G., Hu, W., Sugiyama, M.: Learning from complementary labels. Advances in neural information processing systems 30 (2017)", + "1. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 101-110 (2019)", + "2. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "3. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)", + "4. Lee, J.H., Astrid, M., Zaheer, M.Z., Lee, S.I.: Deep visual anomaly detection with negative learning. In: International Workshop on Frontiers of Computer Vision. pp. 218-232. Springer (2021)", + "5. Li, M., Sethi, I.K.: Confidence-based active learning. IEEE transactions on pattern analysis and machine intelligence 28(8), 1251-1261 (2006)", + "6. Luo, X., Chen, W., Tan, Y., Li, C., He, Y., Jia, X.: Exploiting negative learning for implicit pseudo label rectification in source-free domain adaptive semantic segmentation. arXiv preprint arXiv:2106.12123 (2021)" + ], + "bbox": [ + 225, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BUAL for Open-Set Annotation", + "bbox": [ + 516, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Mahmood, R., Fidler, S., Law, M.T.: Low budget active learning via Wasserstein distance: An integer programming approach. arXiv preprint arXiv:2106.02968 (2021)", + "18. Moon, W., Park, J., Seong, H.S., Cho, C.H., Heo, J.P.: Difficulty-aware simulator for open set recognition. arXiv preprint arXiv:2207.10024 (2022)", + "19. Ning, K.P., Tao, L., Chen, S., Huang, S.J.: Improving model robustness by adaptively correcting perturbation levels with active queries. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 9161-9169 (2021)", + "20. Ning, K.P., Zhao, X., Li, Y., Huang, S.J.: Active learning for open-set annotation. arXiv preprint arXiv:2201.06758 (2022)", + "21. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Gupta, B.B., Chen, X., Wang, X.: A survey of deep active learning. ACM computing surveys (CSUR) 54(9), 1-40 (2021)", + "22. Roy, N., McCallum, A.: Toward optimal active learning through sampling estimation of error reduction. int. conf. on machine learning (2001)", + "23. Salehi, M., Mirzaei, H., Hendrycks, D., Li, Y., Rohban, M.H., Sabokrou, M.: A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051 (2021)", + "24. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012)", + "25. Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489 (2017)", + "26. Settles, B.: Active learning literature survey (2009)", + "27. Sinha, S., Ebrahimi, S., Darrell, T.: Variational adversarial active learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5972-5981 (2019)", + "28. Yao, L., Miller, J.: Tiny imagenet classification with convolutional neural networks. CS 231N 2(5), 8 (2015)", + "29. Yoo, D., Kweon, I.S.: Learning loss for active learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 93-102 (2019)", + "30. You, X., Wang, R., Tao, D.: Diverse expected gradient active learning for relative attributes. IEEE transactions on image processing 23(7), 3203-3217 (2014)", + "31. Yu, X., Liu, T., Gong, M., Tao, D.: Learning with biased complementary labels. In: Proceedings of the European conference on computer vision (ECCV). pp. 68-83 (2018)", + "32. Zong, C.C., Cao, Z.T., Guo, H.T., Du, Y., Xie, M.K., Li, S.Y., Huang, S.J.: Noise-robust bidirectional learning with dynamic sample reweighting. arXiv preprint arXiv:2209.01334 (2022)", + "33. Zong, C.C., Wang, Y.W., Xie, M.K., Huang, S.J.: Dirichlet-based prediction calibration for learning with noisy labels. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 17254-17262 (2024)" + ], + "bbox": [ + 215, + 147, + 784, + 742 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Zong et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_model.json b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5f7a31833ef1ca5bf42ca3a887b670af3c1fbae2 --- /dev/null +++ b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_model.json @@ -0,0 +1,2069 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.224, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.212, + 0.778, + 0.245 + ], + "angle": 0, + "content": "Chen-Chen Zong\\(^{1}\\), Ye-Wen Wang\\(^{1}\\), Kun-Peng Ning\\(^{2}\\), Hai-Bo Ye\\(^{1}\\), and Sheng-Jun Huang\\(^{1}\\)†" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.255, + 0.681, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) Nanjing University of Aeronautics and Astronautics \n\\(^{2}\\) Peking University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.32, + 0.744, + 0.584 + ], + "angle": 0, + "content": "Abstract. Active learning (AL) in open set scenarios presents a novel challenge of identifying the most valuable examples in an unlabeled data pool that comprises data from both known and unknown classes. Traditional methods prioritize selecting informative examples with low confidence, with the risk of mistakenly selecting unknown-class examples with similarly low confidence. Recent methods favor the most probable known-class examples, with the risk of picking simple already mastered examples. In this paper, we attempt to query examples that are both likely from known classes and highly informative, and propose a Bidirectional Uncertainty-based Active Learning (BUAL) framework. Specifically, we achieve this by first pushing the unknown class examples toward regions with high-confidence predictions, i.e., the proposed Random Label Negative Learning method. Then, we propose a Bidirectional Uncertainty sampling strategy by jointly estimating uncertainty posed by both positive and negative learning to perform consistent and stable sampling. BUAL successfully extends existing uncertainty-based AL methods to complex open-set scenarios. Extensive experiments on multiple datasets with varying openness demonstrate that BUAL achieves state-of-the-art performance. The code is available at this link." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.596, + 0.741, + 0.625 + ], + "angle": 0, + "content": "Keywords: Active learning \\(\\cdot\\) Open-set annotation \\(\\cdot\\) Negative learning \\(\\cdot\\) Uncertainty estimation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.65, + 0.377, + 0.667 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Labeling data can be costly and time-consuming, often requiring high levels of expertise from annotators [26]. This expense poses a significant challenge when dealing with insufficient labeled data in deep learning tasks. Recently, active learning (AL) has emerged as a prominent approach to tackle this issue and has gained widespread attention [9,17,21]. It iteratively selects the most informative examples from the unlabeled data pool and queries their labels from an oracle, enabling the learning of an effective model with reduced labeling costs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Existing AL methods [5,8,19,22,27,29,30] typically operate under the closed-set assumption, assuming that the label categories in the unlabeled data pool" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.548, + 0.841 + ], + "angle": 0, + "content": "† Corresponding Author: huangsj@nuaa.edu.cn" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Zong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "match those of the target task. However, this assumption often does not hold in practical scenarios. For example, consider a task that involves classifying images into two target categories, \"Dog\" and \"Cat\". Collecting training examples through keyword-based image search inevitably introduces irrelevant images from other categories (i.e., unknown classes), alongside the two target categories (i.e., known classes)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.433 + ], + "angle": 0, + "content": "In such open-set scenarios, many previous AL methods, which prefer querying examples with less confident predictions, may lead to failure since examples from unknown classes often receive uncertain predictions. To mitigate the impact of unknown class examples, some AL methods designed specifically for open-set scenarios attempt to query examples that are more likely to belong to known classes based on sample similarity [3] or model-predicted max activation value (MAV) [20]. However, since examples similar to the labeled ones may be already mastered by the model, they do not significantly benefit the target model. These methods only perform well when the proportion of unknown class examples is high. When the proportion is low, they tend to perform poorly than traditional AL methods, even inferior to random sampling (please refer to Figure 6). However, determining the proportion of unknown class examples in practical scenarios is often challenging. This further limits the usability of these methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.787, + 0.569 + ], + "angle": 0, + "content": "In this paper, we start with a specific question: can we effectively distinguish the \"informative\" examples of known classes from examples of unknown classes? Intuitively, if we can push the unknown class examples toward regions with high-confidence predictions, existing uncertainty-based AL methods can be applied directly in open-set scenarios. To achieve this, we propose to fine-tune the model by performing negative learning (NL) [11, 14, 16, 32] on unlabeled examples. NL is an indirect learning manner that explores the utility of complementary labels, i.e., the label categories that an instance does not belong to. For a \\(K\\)-classification problem, the NL loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.575, + 0.785, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\ell_ {N L} (f, \\bar {y}) = - \\sum_ {k = 1} ^ {K} \\bar {y} _ {k} \\log \\left(1 - p _ {k}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.784, + 0.643 + ], + "angle": 0, + "content": "where \\( f \\) is the model we want to optimize, \\( \\bar{y} \\) and \\( \\bar{\\pmb{y}} = [\\bar{y}_1,\\dots ,\\bar{y}_k,\\dots ,\\bar{y}_K] \\) represent a complementary label and its corresponding one-hot form, respectively, and \\( p_k \\) denotes the probability of the \\( k \\)-th category." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.786, + 0.809 + ], + "angle": 0, + "content": "Specifically, the fine-tuning process comprises two parts. On one hand, for already labeled examples, we train them directly using Equation 1. On the other hand, for unlabeled examples, we first randomly assign labels to them in each training round and then train the model using Equation 1. Notably, the unlabeled known class example has a relatively higher chance of receiving the correct label, whereas the unknown class example will never be assigned the correct label. Once unlabeled known class data receive the correct labels, they suffer a larger penalty and are reduced confidence predictions by the model since they deviate from the distribution information obtained from labeled data. In contrast, unlabeled unknown class data are not constrained to move towards the high-confidence region for counteracting the update gradient produced by the labeled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "To validate this, we conducted preliminary experiments on CIFAR-10 [12] with 4 known and 6 unknown classes. Figure 1 illustrates the confidence statis" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BUAL for Open-Set Annotation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.485, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.316, + 0.389, + 0.328 + ], + "angle": 0, + "content": "(a) w/o NL" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.147, + 0.781, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.316, + 0.683, + 0.328 + ], + "angle": 0, + "content": "(b) w/ NL" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.339, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Fig. 1: The statistics of prediction confidence before and after fine-tuning the model. In the zoomed-in area of Figure 1b, we swapped the display order of the two to prevent occlusion, allowing for a more intuitive view of how the distribution has changed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.785, + 0.485 + ], + "angle": 0, + "content": "tics for all unlabeled examples in a fixed query round before and after fine-tuning the model. As expected, known class examples are more prevalent in the low-confidence region, while unknown class ones are more common in high-confidence regions. This distribution characteristic presents a potential solution to the aforementioned question and offers a promising approach to AL for open-set scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.787, + 0.728 + ], + "angle": 0, + "content": "Based on this, we propose a Bidirectional Uncertainty-based Active Learning framework (BUAL). On the one hand, we propose the Random Label Negative Learning (RLNL) method to fine-tune the model and leverage information from the unlabeled data pool. Specifically, we first train a model using all labeled data as a positive classifier. Then, negative learning is performed to fine-tune the model as a negative classifier by randomly assigning labels to data from the unlabeled pool in each training iteration. This effectively distinguishes known class examples with lower confidence predictions from unknown class ones. On the other hand, we propose a Bidirectional Uncertainty (BU) sampling strategy for active selection, which estimates prediction uncertainty from both positive and negative classifiers. By selecting examples with the highest uncertainty, we expect to identify the most informative instances from the known classes. Experiments are performed on multiple datasets with different unknown class ratios. The results demonstrate that BUAL can query more informative known class examples and that the model performance obtained by BUAL is substantially improved compared with existing state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.749, + 0.388, + 0.765 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Active learning (AL) is a prominent approach aimed at reducing label costs by selecting a batch of examples that are most valuable for model training. Existing AL methods can be broadly categorized into three groups based on sample selection strategies: uncertainty-based, representative-based, and hybrid strategies" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Zong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "which combine both aspects. Uncertainty-based strategies focus on sampling informative instances to reduce model uncertainty. Typical methods include Least Confident Sampling [15], Margin-based Sampling [2], and Entropy-based Sampling [7], etc. Representative-based strategies start from the sample distribution and aim to select representative instances that match the overall distribution. A typical method is Coreset [25]. Hybrid strategies combine uncertainty and representativeness by incorporating sample distribution information and the model's specific needs. Notable methods in this category include QUIRE [8] and BADGE [1], etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Open-set annotation (OSA) involves active learning under open-set scenarios [3,20]. Existing methods primarily focus on selecting examples that are most likely to belong to known classes. For instance, CCAL [3] employs contrastive learning to extract semantic and distinctive features of examples, facilitating discrimination of known class examples. LfOSA [20] introduces an auxiliary network to model the per-example max activation value (MAV) distribution and dynamically selects examples with the highest probability from known classes. However, these methods exhibit sensitivity to the openness of the dataset and may not consistently perform well. Open-set recognition (OSR) [18, 23, 24] is a related problem setting to OSA, aiming to predict correct labels for known class examples while simultaneously detecting examples from unknown classes. Nevertheless, direct use of OSR methods often falls short of expected effectiveness [20], mainly due to limited training examples and the inability to identify highly informative instances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.66 + ], + "angle": 0, + "content": "Complementary labels are labels other than the ground-truth label assigned to an example. Complementary label learning (CLL) [4,10,31] was initially introduced in [10], where the authors leveraged the lower acquisition cost and higher correct labeling rate of complementary labels to address multi-class classification problems. However, complementary labels contain less information compared to the correct labels, which can result in slow convergence of the model and challenges in achieving the desired performance when directly using complementary labels for training. Authors in [11] made the first attempt to combine CLL with the noise labeling problem [33] and proposed an indirect and robust training schema called negative learning (NL). In this paper, we exploit the properties of NL and extend it to open-set and label-free settings for the first time." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.683, + 0.381, + 0.701 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.714, + 0.373, + 0.728 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.738, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Notations. In open-set annotation (OSA) settings, there are two labeled data pools: \\( D_{l}^{kn o} = \\{(x_{i}^{l},y_{i}^{l})\\}_{i = 1}^{n_{l}^{kn o}} \\) for known classes and \\( D_{l}^{unk} = \\{(x_{i}^{l})\\}_{i = 1}^{n_{l}^{unk}} \\) for unknown classes, as well as an unlabeled data pool: \\( D_{u} = \\{(x_{i}^{u})\\}_{i = 1}^{n_{u}^{kn o}} \\cup \\{(x_{i}^{u})\\}_{i = 1}^{n_{u}^{unk}} \\) containing examples from both known and unknown classes. Each known class example belongs to one of \\( K \\) known classes in the label space \\( \\mathcal{V} = \\{1,2,\\dots ,K\\} \\). The unknown class examples are uniformly grouped into one category, denoted" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BUAL for Open-Set Annotation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.154, + 0.782, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Fig. 2: The framework of BUAL. A two-stage \\( K \\) class classifier is maintained, where the first stage is trained in a normal manner saved as \\( f_{p}(\\cdot) \\) and the second stage is trained using the proposed random label negative learning method denoted as \\( f_{n}(\\cdot) \\). An auxiliary \\( K + 1 \\) class classifier \\( f_{aux}(\\cdot) \\) is trained in parallel. By collecting the predicted uncertainty from \\( f_{p}(\\cdot) \\) and \\( f_{n}(\\cdot) \\) on each candidate example along with the global and local balancing factors, the proposed bidirectional sampling strategy can accurately estimate the potential utility of each example and perform effective sample sampling under complex open-set scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.568 + ], + "angle": 0, + "content": "as \\( y = \\emptyset \\). In each training round, active learning (AL) queries a batch of \\( b \\) examples according to a given query strategy \\( \\mathcal{A} \\), denoted as \\( X_{query} = X_{query}^{kn}\\cup X_{query}^{unk} \\). Once the labeled feedback is obtained, we can calculate a ratio \\( r = \\frac{|X_{query}^{kn}|}{|X_{query}|} \\) to represent the precision of known classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Overview. The proposed Bidirectional Uncertainty-based Active Learning (BUAL) framework is depicted in Figure 2. In each iteration, there are in general three steps:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.638, + 0.788, + 0.715 + ], + "angle": 0, + "content": "- Model Training: the algorithm first trains a target classifier in a normal learning manner, dubbed positive classifier \\( f_{p}(\\cdot) \\). Then, the algorithm finetunes the classifier with a new classifier head by the proposed Random Label Negative Learning method, dubbed negative classifier \\( f_{n}(\\cdot) \\). Similar to [20], a \\( K + 1 \\) auxiliary classifier \\( f_{aux}(\\cdot) \\) is trained in parallel." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.716, + 0.788, + 0.793 + ], + "angle": 0, + "content": "- Example Selection: the Bidirectional Uncertainty Sampling Strategy estimates uncertainty bidirectionally using both the positive and negative classifier heads. This estimation is combined with dynamic balance factors generated by the auxiliary classifier and query feedback to select the most informative known class examples." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "- Oracle Labeling: the annotators assign class labels to the selected examples. Based on the feedback results, update the corresponding data pools accordingly." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.638, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Zong et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.541, + 0.163 + ], + "angle": 0, + "content": "3.2 Random Label Negative Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.279 + ], + "angle": 0, + "content": "In OSA scenarios, conventional uncertainty-based AL methods tend to be ineffective for example selection, mainly due to the unconfident predictions generated for the unknown class examples. Existing OSA methods prioritize sample purity without fully exploring sample informativeness, resulting in queries with too many already mastered examples that are not useful for model training. Thus, a key question arises: how to distinguish between unknown and \"informative\" known class examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.4 + ], + "angle": 0, + "content": "To cope with this problem, we propose a general method by pushing unknown class examples toward the high-confidence regions, while pushing known class examples to the low-confidence regions. If this separation is achieved, we can leverage existing uncertainty-based AL methods directly to handle various complex open-set scenarios. Fortunately, we achieve this to some extent by employing negative learning (NL), where we assign random labels to unlabeled examples and fine-tune the model with a new classifier head accordingly, dubbed Random Label Negative Learning (RLNL)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.788, + 0.582 + ], + "angle": 0, + "content": "Specifically, in the model training stage, we first train a \\(K\\)-class classifier \\(f_{p}(\\cdot)\\) in a normal training manner (e.g., cross-entropy loss) based on \\(D_{l}^{kn o}\\). Note that \\(f_{p}(\\cdot)\\) can be the target model we eventually need to output. After training on the labeled data, \\(f_{p}(\\cdot)\\) can own a good discriminative ability for known class examples at the representation level. Then, we replace the last layer of \\(f_{p}(\\cdot)\\) and fine-tune a new classifier head \\(f_{n}(\\cdot)\\) with Equation 1. In this phase, unlabeled examples in \\(D_{u}\\) are involved in training to help achieve the goal mentioned above: pushing unknown and known class examples toward the high and low confidence regions, respectively. All labeled examples in \\(D_{l}^{kn o}\\) are still involved to prevent them from being incorrectly shifted in the distribution. Eventually, we assign random labels to the unlabeled examples in \\(D_{u}\\), while using the complementary labels instead for the labeled known class examples, i.e., for an example \\(\\mathbf{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.594, + 0.787, + 0.634 + ], + "angle": 0, + "content": "\\[\nP (\\bar {y} = s) = \\frac {1}{| S |}, s \\in \\left\\{ \\begin{array}{l l} S = \\mathcal {Y} \\backslash y ^ {l}, & \\text {i f} \\boldsymbol {x} \\in D _ {l} ^ {\\text {k n o}}, \\\\ S = \\mathcal {Y}, & \\text {i f} \\boldsymbol {x} \\in D _ {u}. \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.689 + ], + "angle": 0, + "content": "where \\(\\bar{y}\\) is uniformly sampled at each training iteration. Here, considering that the number of examples in \\(D_{u}\\) is usually much larger than that in \\(D_{l}^{kn o}\\), a small subset \\(D_{sub}\\) is randomly selected from the \\(D_{u}\\) to save training cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Why RLNL works? The first question one may hold is why it shouldn't be optimal for the unlabeled examples to be the uniform distribution as \\(\\bar{y}\\) is sampled at each iteration. To explain this, we take two different kinds of updating for a single example in the binary classification scenario of \"Dog vs. Cat\" as illustrations: negative learning using all labels at each iteration and negative learning using only one random label at each iteration. One possible case is shown in Figure 3. If each iteration uses all labels for negative learning, i.e., such an example is neither a \"Cat\" nor a \"Dog\", it is apparent that an update gradient of 0 is only achieved when the example is at the decision boundary. In contrast, if each iteration gives only one random label for negative learning," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BUAL for Open-Set Annotation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.147, + 0.52, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.149, + 0.73, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.318, + 0.785, + 0.348 + ], + "angle": 0, + "content": "Fig. 3: Use all labels per iteration for negative learning (left) vs. use one random label per iteration for negative learning (right)." + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.366, + 0.706, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Fig. 4: The possible RLNL update scenario for unlabeled unknown class data in batch deep learning manner. The green \"→\" is the batch update gradient produced by the example itself, and the purple \"→\" is the update gradient produced by labeled data. Initially, the decision boundary is close to the left-hand category." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.785, + 0.686 + ], + "angle": 0, + "content": "i.e., such an example is not a \"Cat\" or a \"Dog\", then for an instance that is not a \"Cat\", pushing it as much as possible towards the \"Dog\" side will result in a gradient update of 0, and vice versa. Obviously, there is no fixed optimal scenario for this type of data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In RLNL, we utilize this property by exploiting the prior knowledge contained in the labeled data. For unlabeled known class data, they might have overlapping regions with labeled data in the feature space, benefiting from the feature representations learned in the previous stage and the simultaneous introduction of labeled data for negative learning. The mapping of such examples in the feature space will remain or be close to the labeled ones in subsequent model updates, as they will receive invisible constraints provided by the prior knowledge from the labeled data, owing to similar features." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "As shown in Figure 4, we present the possible update scenario for unlabeled unknown class examples in the commonly adopted batch deep learning manner." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "C. Zong et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.154, + 0.465, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.282, + 0.398, + 0.295 + ], + "angle": 0, + "content": "(a) Before RLNL" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.15, + 0.784, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.283, + 0.681, + 0.295 + ], + "angle": 0, + "content": "(b) After RLNL" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 5: The t-SNE feature visualization of labeled data, unlabeled known class data, and unlabeled unknown class data on CIFAR-10 with an openness ratio of 0.5 before and after performing RLNL. For a more intuitive visualization, we only show a single known class. More visualization results are shown in the supplementary file." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.789, + 0.571 + ], + "angle": 0, + "content": "Here, the green arrow indicates the batch update gradient produced by itself, while the purple arrow indicates the batch update gradient produced by labeled data. Different from Figure 3, where data oscillates on both sides of the decision boundary, the unlabeled unknown class examples will oscillate at uncharted away from the decision boundary to counteract the update gradient due to the labeled ones. In contrast, unlabeled known class examples will move much less than unknown class ones in magnitude in the feature space within the constraints of the prior knowledge provided by the labeled data. This is further confirmed in Figure 5, which illustrates the t-SNE feature visualization of labeled data, unlabeled known class data, and unlabeled unknown class data before and after performing RLNL. This ultimately leads to the result in Figure 1b and proves that RLNL does work." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.631, + 0.607 + ], + "angle": 0, + "content": "3.3 Bidirectional Uncertainty Sampling Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.72 + ], + "angle": 0, + "content": "During the fine-tuning process, we observed that the predictions of \\( f_{n}(\\cdot) \\) for examples in \\( D_{u} \\) tend to oscillate between epochs. To ensure stable sampling, on one hand, we test all unlabeled examples \\( t \\) times at \\( m \\) round intervals to obtain the predicted probabilities \\( \\pmb{p}_{\\pmb{t}}^{-} = (p_{1}^{t},\\dots,p_{K}^{t}) \\), which will further be averaged as \\( \\mathcal{P}^{-} = \\frac{1}{t}\\sum_{i = 1}^{t}\\pmb{p}_{i}^{-} = (\\frac{1}{t}\\sum_{i = 1}^{t}p_{1}^{i},\\dots,\\frac{1}{t}\\sum_{i = 1}^{t}p_{K}^{i}) \\). On the other hand, we reuse the predictions of \\( f_{p}(\\cdot) \\) to produce predicted probabilities \\( \\pmb{p}^{+} = (p_{1},\\dots,p_{K}) \\) for all unlabeled examples, as \\( f_{p}(\\cdot) \\) are accurate for measuring known class ones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Compared to the positive head, the negative head is slightly biased for the measurement of sample uncertainty due to the unstable training. Therefore, if an example is likelier to belong to known classes, we prefer to utilize the sample uncertainty obtained from \\( f_{p}(\\cdot) \\). On the contrary, once an example has a higher risk of belonging to the unknown classes, the uncertainty obtained from \\( f_{p}(\\cdot) \\) is unreliable, and thus the uncertainty by \\( f_{n}(\\cdot) \\) should be given a higher weight." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To achieve this, we introduce two balancing factors based on global and local observations. In open-set scenarios, some unknown class examples might be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BUAL for Open-Set Annotation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "mistakenly selected for annotation. Although these examples cannot be directly used for training the target model, they are valuable for measuring sample purity. Therefore, similar to [20], we train a \\(K + 1\\) classifier \\(f_{aux}(\\cdot)\\) in a normal training manner based on the examples from both \\(D_l^{kno}\\) and \\(D_l^{unk}\\). Then, we can obtain the predicted probability \\(p_{K + 1}^{aux}(\\boldsymbol{x}) = f_{aux}(\\emptyset \\mid \\boldsymbol{x})\\) for each example. This can act as a local balancing factor, since the larger the value of \\(p_{K + 1}^{aux}(\\boldsymbol{x})\\) the more likely \\(\\boldsymbol{x}\\) is to belong to the unknown class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.784, + 0.327 + ], + "angle": 0, + "content": "Additionally, once the selected examples are sent to the oracle for annotation, we can calculate the ratio \\( r \\), which provides a rough estimate of the current openness of \\( D_{u} \\) and can serve as a global balancing factor. With the two balancing factors, we propose a bidirectional uncertainty sampling strategy defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.328, + 0.786, + 0.349 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {*} = \\arg \\max _ {\\boldsymbol {x}} p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\operatorname {u n c} _ {n} + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\operatorname {u n c} _ {p}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.428 + ], + "angle": 0, + "content": "where \\(unc_{p}\\) and \\(unc_{n}\\) denote the uncertainty of \\(\\mathbf{x}\\) to the positive classifier head and the negative classifier head, respectively. Noteworthy, this sampling strategy remains applicable even in closed-set settings. If there are no unknown class examples, the ratio \\(r\\) will always be equal to 1, and \\(p_{K + 1}^{aux}(\\mathbf{x})\\) will be 0, effectively making the strategy equivalent to a normal uncertainty sampling strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.784, + 0.503 + ], + "angle": 0, + "content": "With Equation 3, extending the existing closed-set uncertainty-based active learning methods to open-set scenarios is possible. In this paper, we focus on three classical uncertainty sampling strategies: least confident sampling, margin-based sampling, and entropy-based sampling. The corresponding modified versions for open-set scenarios are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.513, + 0.573, + 0.529 + ], + "angle": 0, + "content": "- Bidirectional Least Confident Sampling" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.537, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\arg \\max } p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\left[ 1 - \\boldsymbol {\\mathcal {P}} _ {y ^ {-}} ^ {-} (\\boldsymbol {x}) \\right] + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\left[ 1 - \\boldsymbol {\\mathcal {P}} _ {y ^ {+}} ^ {+} (\\boldsymbol {x}) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.57, + 0.591, + 0.589 + ], + "angle": 0, + "content": "where \\(y^{-} = \\arg \\max_{y}\\mathcal{P}_{y}^{-}(\\pmb {x})\\) \\(y^{+} = \\arg \\max_{y}\\pmb{p}_{y}^{+}(\\pmb {x})\\)" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.596, + 0.557, + 0.611 + ], + "angle": 0, + "content": "- Bidirectional Margin-Based Sampling" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.62, + 0.785, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\arg \\max } p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\left[ \\boldsymbol {\\mathcal {P}} _ {y _ {1} ^ {-}} ^ {-} (\\boldsymbol {x}) - \\boldsymbol {\\mathcal {P}} _ {y _ {2} ^ {-}} ^ {-} (\\boldsymbol {x}) \\right] \\tag {5} \\\\ + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\left[ \\boldsymbol {p} _ {y _ {1} ^ {+}} ^ {+} (\\boldsymbol {x}) - \\boldsymbol {p} _ {y _ {2} ^ {+}} ^ {+} (\\boldsymbol {x}) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.68, + 0.784, + 0.72 + ], + "angle": 0, + "content": "where \\( y_1^- = \\arg \\max_y \\mathcal{P}_y^-(\\boldsymbol{x}) \\), \\( y_1^+ = \\arg \\max_y \\mathbf{p}_y^+(\\boldsymbol{x}) \\), \\( y_2^- = \\arg \\max_{y \\setminus y_1^-} \\mathcal{P}_y^-(\\boldsymbol{x}) \\), \\( y_2^+ = \\arg \\max_{y \\setminus y_1^+} \\mathbf{p}_y^+(\\boldsymbol{x}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.726, + 0.565, + 0.741 + ], + "angle": 0, + "content": "- Bidirectional Entropy-Based Sampling" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.75, + 0.785, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\arg \\max } p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\left[ - \\boldsymbol {\\mathcal {P}} _ {y ^ {-}} ^ {-} (\\boldsymbol {x}) \\log \\boldsymbol {\\mathcal {P}} _ {y ^ {-}} ^ {-} (\\boldsymbol {x}) \\right] \\tag {6} \\\\ + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\left[ - \\boldsymbol {p} _ {y ^ {+}} ^ {+} (\\boldsymbol {x}) \\log \\boldsymbol {p} _ {y ^ {+}} ^ {+} (\\boldsymbol {x}) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.652, + 0.825 + ], + "angle": 0, + "content": "where \\(y^{-}\\) and \\(y^{+}\\) are consistent with the previous definition." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.825, + 0.698, + 0.84 + ], + "angle": 0, + "content": "The main procedures of BUAL are summarized in Algorithm 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Zong et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.517, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 BUAL Training Procedure" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.164, + 0.787, + 0.538 + ], + "angle": 0, + "content": "Input: \\(D_l^{kno}\\) \\(D_{l}^{unk}\\) \\(D_u\\) \\(r\\) \\(\\mathcal{A},b,m,t,\\) subset size s, training epoch \\(E\\) , optimizer \\(\\mathcal{O}\\) \nOutput: \\(D_l^{kno}\\) \\(D_{l}^{unk}\\) \\(D_u\\) \\(r\\) , model parameters \\(\\theta_p\\) \nProcess: \n1: # Model training \n2: for \\(j = 1;j\\leq E\\) do \n3: \\(\\mathcal{L}_p = \\sum_{(x,y)\\in D_l^{kno}}\\ell (f_p(\\boldsymbol {x}),y);\\) \n4: \\(\\mathcal{L}_{aux} = \\sum_{(x,y)\\in D_l^{kno}\\cup D_l^{unk}}\\ell (f_{aux}(\\boldsymbol {x}),y);\\) \n5: Update \\(\\theta_p,\\theta_{aux} = \\mathcal{O}(\\mathcal{L}_p,\\theta_p),\\mathcal{O}(\\mathcal{L}_{aux},\\theta_{aux})\\) \n6: end for \n7: Use \\(f_{aux}(\\pmb {x})\\) to remove confidently unknown examples in \\(D_{u}\\) and randomly select s examples as \\(D_{sub}\\) \n8: for \\(j = 1;j\\leq E\\) do \n9: Generate label \\(\\bar{y}\\) for each example. \n10: \\(\\mathcal{L}_n = \\sum_{(x,y)\\in D_l^{kno}\\cup D_{sub}}\\ell_{NL}(f_n(\\boldsymbol {x}),\\bar{y})\\) \n11: Update \\(\\theta_{n} = \\mathcal{O}(\\mathcal{L}_{n},\\theta_{n})\\) \n12: if \\(j\\) mod \\(m = 0\\) then \n13: Obtain \\(\\pmb {p}_j^-\\) for each sample in \\(D_{u}\\) by \\(f_{n}(\\pmb {x})\\) \n14: end if \n15: end for \n16: # Example selection \n17: Calculate \\(\\pmb{\\mathcal{P}}^{-}\\) for each sample across rounds; \n18: Obtain \\(\\pmb{p}^{+}\\) for each sample in \\(D_u\\) by \\(f_{p}(\\pmb {x})\\) \n19: Obtain \\(p_{K + 1}^{aux}\\) for each sample in \\(D_u\\) by \\(f_{aux}(\\pmb {x})\\) \n20: Obtain \\(X_{query}\\) using BU sampling strategy A; \n21: # Oracle labeling \n22: Ask for annotation and update \\(D_l^{kno}\\) \\(D_l^{unk}\\) \\(D_u\\) and \\(r\\)" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.568, + 0.375, + 0.585 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.785, + 0.674 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on three benchmark datasets: CIFAR-10, CIFAR-100 [12], and Tiny-Imagenet [28]. Tiny-Imagenet is a subset of theImagenet [13], consisting of 200 classes with 500 training images per class. The openness of each dataset is defined as the ratio of unknown classes to the total number of classes, and we set its value to 0.2, 0.4, 0.6, and 0.8 for all datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Comparing methods. We select nine AL strategies for comparison, which can be further categorized into six groups: (1) Random: Randomly select examples from the unlabeled data pool for labeling. (2) Traditional uncertainty-based strategies: Least confident sampling (LC), Margin-based sampling (Margin), and Entropy-based sampling (Entropy). (3) Diversity-based strategy: Coreset. (4) Hybrid-based strategy: BADGE. (5) OSA methods: CCAL and LfOSA. (6) OSR method: DIAS [18]. Correspondingly, our methods are Bidirectional Least confident sampling (B-LC), Bidirectional Margin-based sampling (B-Margin), and Bidirectional Entropy-based sampling (B-Entropy)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Training details. On CIFAR-10, CIFAR-100, and Tiny-Imagenet, we randomly sample \\(1\\%\\), \\(8\\%\\), and \\(8\\%\\) known class examples as the initial labeled data," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BUAL for Open-Set Annotation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.147, + 0.499, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.147, + 0.747, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.283, + 0.498, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.283, + 0.746, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.419, + 0.498, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.419, + 0.747, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.604 + ], + "angle": 0, + "content": "Fig. 6: Accuracy comparison on CIFAR-10 (first row), CIFAR-100 (second row), and Tiny-Imagenet (third row). The ratio of unknown class examples to the total number of examples is fixed at 0.4 (first column) and 0.6 (second column) for each dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.753 + ], + "angle": 0, + "content": "respectively. All the models involved in the experiments are ResNet18 [6], trained for 100 epochs, using SGD as the optimizer, where the learning rate is 0.01, momentum is 0.9, weight decay is 1e-4, and batch size is 128. We perform the experiments for 3 runs and report the average results. For CIFAR-10 and CIFAR-100, 5000 examples are randomly selected from the unlabeled pool as \\( D_{sub} \\), and 1500 examples are queried in each query round. Due to the doubling of data volume, we randomly select 10000 examples as \\( D_{sub} \\) and query 3000 examples in each query round for Tiny-Imagenet." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.772, + 0.473, + 0.788 + ], + "angle": 0, + "content": "4.1 Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Figure 6 presents the variation curves for the classification accuracy of the proposed methods and the comparison methods. Here, to better observe the variation curves, we only show the B-Margin for the proposed methods and the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Zong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: The final round average accuracy of different methods on CIFAR-10, CIFAR-100, and Tiny-Imagenet. The best performance is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.186, + 0.784, + 0.419 + ], + "angle": 0, + "content": "
DatasetsCIFAR-10CIFAR-100Tiny-Imagenet
Openness Ratio0.20.40.60.80.20.40.60.80.20.40.60.8
(1)Random83.382.587.296.957.658.358.761.245.747.250.955.0
(2)LC84.381.687.596.255.854.654.056.244.845.948.451.6
Margin86.084.189.097.059.359.658.858.946.447.150.854.0
Entropy85.483.488.096.857.156.855.756.444.644.546.950.7
(3)Coreset85.081.886.497.460.261.261.864.246.247.851.854.0
(4)BADGE86.884.289.296.460.260.860.462.046.347.851.853.3
(5)LfOSA73.778.787.098.652.356.662.468.242.546.652.459.9
CCAL80.881.588.098.155.960.064.767.744.446.350.357.0
(6)DIAS81.880.783.094.055.756.156.957.243.145.147.554.4
OursB-LC87.087.292.599.159.362.867.572.145.748.754.760.6
B-Margin86.587.092.698.960.963.168.371.546.549.555.761.2
B-Entropy86.987.492.699.158.961.766.971.445.447.555.261.0
" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.437, + 0.407, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.437, + 0.592, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.437, + 0.777, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.566, + 0.785, + 0.595 + ], + "angle": 0, + "content": "Fig. 7: The average recognition rate on CIFAR-10 (first column), CIFAR-100 (second column), and Tiny-Imagenet (third column)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.785, + 0.658 + ], + "angle": 0, + "content": "Margin for the compared traditional uncertainty methods. Table 1 reports all methods' final round average accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We can observe that our proposed methods consistently achieve the highest classification accuracy, demonstrating the effectiveness and superiority of our BUAL framework over other methods. Some notable observations are as follows: 1) The OSA methods, CCAL and LfOSA, gradually lose effectiveness as the proportion of known classes in the dataset increases. In contrast, traditional uncertainty-based methods perform better as the openness ratio decreases. This is consistent with our previous analysis. 2) The performance of Coreset and BADGE also deteriorates when the openness ratio is high. These methods tend to select examples with diverse characteristics. However, unknown class examples often differ significantly in characteristics from the labeled ones, making them more likely to be sampled and thus undermining the effectiveness of this type of method. 3) The OSR method DIAS does not perform well in all situations. The" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BUAL for Open-Set Annotation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.151, + 0.49, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.311, + 0.404, + 0.323 + ], + "angle": 0, + "content": "(a) LfOSA" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.152, + 0.756, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.311, + 0.682, + 0.323 + ], + "angle": 0, + "content": "(b) BUAL (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.335, + 0.785, + 0.362 + ], + "angle": 0, + "content": "Fig. 8: The t-SNE feature visualization of data from one query and labeled pool on CIFAR-10 with an openness ratio of 0.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.534 + ], + "angle": 0, + "content": "limited labeled data prevents the model from learning a robust representation for identifying open-set examples effectively. Moreover, DIAS cannot identify highly informative examples, resulting in its queries often being simple and unhelpful examples. 4) All our methods remain stable and do not suffer significantly from changes in the openness ratio. This can demonstrate two points. The dynamic balance factors in the proposed framework can adaptively assign appropriate weights to positive and negative uncertainties regardless of the openness of the dataset, and negative uncertainties are indeed effective for querying highly informative known class examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Figure 7 shows the average recognition rate of known class examples across queries and openness ratios. We can observe that traditional AL methods all perform poorly due to their inability to recognize known class examples from those with high uncertainty and/or representative. As an OSR method, DIAS shows only marginal improvement compared to the traditional AL methods, unable to fully exploit its exceptional performance with limited training data. The recognition ability of CCAL decreases slightly with the increase of dataset categories compared to our method. It is noted that LfOSA achieved the highest query precision on all datasets. However, combining the results in Table 1, we can find that the performance of models trained by LfOSA is not satisfactory compared to other methods. To explain why, we visualize the queried examples' feature representation and labeled examples' feature representation for our method and LfOSA in Figure 8. We can observe that the features of examples queried by LfOSA are highly overlapping with the labeled ones, which are often examples that the model has already mastered and cannot provide an effective bonus for model training. On the contrary, the sample features queried by our method have little overlap and are more distributed in low-density regions, consistent with the goal of AL. These results validate the effectiveness of our approach in querying more informative examples of known classes and maintaining a high recognition rate." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Zong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 2: Final accuracy of each component in Equation 3 on CIFAR-10, CIFAR-100, and Tiny-Imagenet with an openness ratio of 0.6." + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.186, + 0.744, + 0.291 + ], + "angle": 0, + "content": "
MethodDatasetuncpuncnw/o ww/o fauxB-LC
CIFAR-1087.589.490.891.392.5
CIFAR-10054.063.562.465.067.5
Tiny-Imagenet48.452.352.053.054.7
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.317, + 0.389, + 0.332 + ], + "angle": 0, + "content": "4.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.787, + 0.429 + ], + "angle": 0, + "content": "The ablation study is conducted on CIFAR-10 with an openness ratio of 0.6 to validate the effectiveness of each component in our proposed query strategy (Equation 3). The final round accuracy is shown in Table 2. Here, \\( \\text{unc}_n \\) and \\( \\text{unc}_p \\) indicate that only \\( \\text{unc}_n \\) and \\( \\text{unc}_p \\) is adopted for active sampling, respectively. \\( w / o \\) denotes the removal of all balancing factors, i.e., both \\( f_{aux} \\) and \\( r \\). \\( w / o \\) means only removing the local balancing factor \\( f_{aux} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.596 + ], + "angle": 0, + "content": "We can observe that removing any of the components leads to performance degradation. Although \\( \\mathbf{unc}_n \\) and \\( \\mathbf{unc}_p \\) are significantly less effective than B-LC, \\( \\mathbf{unc}_n \\) has a substantial improvement compared to \\( \\mathbf{unc}_p \\) due to its ability to distinguish informative examples of known classes from examples of unknown classes. The direct combination of \\( \\mathbf{unc}_n \\) and \\( \\mathbf{unc}_p \\), i.e., \\( \\mathbf{w} / \\mathbf{o} \\mathbf{w} \\), works better than \\( \\mathbf{unc}_n \\), as models trained by RLNL may produce oscillate output and thus the uncertainty obtained in \\( \\mathbf{unc}_n \\) is not necessarily accurate. By adding the dynamic global balancing factor \\( r \\), \\( \\mathbf{w} / \\mathbf{o} f_{aux} \\) achieves a better performance. However, it still falls short in comparison to B-LC, which validates the effectiveness of local balancing factor \\( f_{aux} \\). These results further corroborate the soundness of our strategy design." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.36, + 0.631 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we successfully expand the existing uncertainty-based active learning methods to complex and ever-changing open-set scenarios by proposing a Bidirectional uncertainty-based Active Learning (BUAL) framework. On one hand, to achieve the goal of distinguishing known and unknown class examples with high uncertainty, we propose a simple but effective Random Label Negative Learning (RLNL) method for pushing unknown and known class examples toward the high and low confidence regions respectively. On the other hand, to better measure sample uncertainty, we propose a Bidirectional Uncertainty (BU) sampling strategy by dynamically fusing the sample uncertainty obtained from positive learning and negative learning. The dynamic balancing factors in it can ensure that the strategy is effective under various openness ratios. Extensive experimental results show that the model trained with BUAL can achieve state-of-the-art performance under various open-set scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BUAL for Open-Set Annotation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.787, + 0.22 + ], + "angle": 0, + "content": "This work was supported by the Natural Science Foundation of Jiangsu Province of China (BK20222012, BK20211517), the National Key R&D Program of China (2020AAA0107000), and NSFC (62222605)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.241, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.31 + ], + "angle": 0, + "content": "1. Ash, J.T., Zhang, C., Krishnamurthy, A., Langford, J., Agarwal, A.: Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.338 + ], + "angle": 0, + "content": "2. Balcan, M.F., Broder, A., Zhang, T.: Margin based active learning. In: International Conference on Computational Learning Theory. pp. 35-50. Springer (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.786, + 0.379 + ], + "angle": 0, + "content": "3. Du, P., Zhao, S., Chen, H., Chai, S., Chen, H., Li, C.: Contrastive coding for active learning under class distribution mismatch. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8927-8936 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.786, + 0.419 + ], + "angle": 0, + "content": "4. Feng, L., Kaneko, T., Han, B., Niu, G., An, B., Sugiyama, M.: Learning with multiple complementary labels. In: International Conference on Machine Learning. pp. 3072-3081. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.786, + 0.446 + ], + "angle": 0, + "content": "5. Fu, Y., Zhu, X., Li, B.: A survey on instance selection for active learning. Knowledge and information systems 35(2), 249-283 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.786, + 0.488 + ], + "angle": 0, + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.786, + 0.528 + ], + "angle": 0, + "content": "7. Holub, A., Perona, P., Burl, M.C.: Entropy-based active learning for object recognition. In: 2008 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops. pp. 1-8. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.786, + 0.569 + ], + "angle": 0, + "content": "8. Huang, S.J., Jin, R., Zhou, Z.H.: Active learning by querying informative and representative examples. Advances in neural information processing systems 23 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.786, + 0.596 + ], + "angle": 0, + "content": "9. Huang, S.J., Zong, C.C., Ning, K.P., Ye, H.B.: Asynchronous active learning with distributed label querying. In: IJCAI. pp. 2570-2576 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.786, + 0.623 + ], + "angle": 0, + "content": "0. Ishida, T., Niu, G., Hu, W., Sugiyama, M.: Learning from complementary labels. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.663 + ], + "angle": 0, + "content": "1. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 101-110 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.786, + 0.69 + ], + "angle": 0, + "content": "2. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.786, + 0.731 + ], + "angle": 0, + "content": "3. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.786, + 0.772 + ], + "angle": 0, + "content": "4. Lee, J.H., Astrid, M., Zaheer, M.Z., Lee, S.I.: Deep visual anomaly detection with negative learning. In: International Workshop on Frontiers of Computer Vision. pp. 218-232. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.8 + ], + "angle": 0, + "content": "5. Li, M., Sethi, I.K.: Confidence-based active learning. IEEE transactions on pattern analysis and machine intelligence 28(8), 1251-1261 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "6. Luo, X., Chen, W., Tan, Y., Li, C., He, Y., Jia, X.: Exploiting negative learning for implicit pseudo label rectification in source-free domain adaptive semantic segmentation. arXiv preprint arXiv:2106.12123 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "C. Zong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Mahmood, R., Fidler, S., Law, M.T.: Low budget active learning via Wasserstein distance: An integer programming approach. arXiv preprint arXiv:2106.02968 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "18. Moon, W., Park, J., Seong, H.S., Cho, C.H., Heo, J.P.: Difficulty-aware simulator for open set recognition. arXiv preprint arXiv:2207.10024 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "19. Ning, K.P., Tao, L., Chen, S., Huang, S.J.: Improving model robustness by adaptively correcting perturbation levels with active queries. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 9161-9169 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "20. Ning, K.P., Zhao, X., Li, Y., Huang, S.J.: Active learning for open-set annotation. arXiv preprint arXiv:2201.06758 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "21. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Gupta, B.B., Chen, X., Wang, X.: A survey of deep active learning. ACM computing surveys (CSUR) 54(9), 1-40 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "22. Roy, N., McCallum, A.: Toward optimal active learning through sampling estimation of error reduction. int. conf. on machine learning (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.398 + ], + "angle": 0, + "content": "23. Salehi, M., Mirzaei, H., Hendrycks, D., Li, Y., Rohban, M.H., Sabokrou, M.: A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "24. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "25. Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.585, + 0.481 + ], + "angle": 0, + "content": "26. Settles, B.: Active learning literature survey (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "27. Sinha, S., Ebrahimi, S., Darrell, T.: Variational adversarial active learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5972-5981 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "28. Yao, L., Miller, J.: Tiny imagenet classification with convolutional neural networks. CS 231N 2(5), 8 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "29. Yoo, D., Kweon, I.S.: Learning loss for active learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 93-102 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.619 + ], + "angle": 0, + "content": "30. You, X., Wang, R., Tao, D.: Diverse expected gradient active learning for relative attributes. IEEE transactions on image processing 23(7), 3203-3217 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "31. Yu, X., Liu, T., Gong, M., Tao, D.: Learning with biased complementary labels. In: Proceedings of the European conference on computer vision (ECCV). pp. 68-83 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "32. Zong, C.C., Cao, Z.T., Guo, H.T., Du, Y., Xie, M.K., Li, S.Y., Huang, S.J.: Noise-robust bidirectional learning with dynamic sample reweighting. arXiv preprint arXiv:2209.01334 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "33. Zong, C.C., Wang, Y.W., Xie, M.K., Huang, S.J.: Dirichlet-based prediction calibration for learning with noisy labels. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 17254-17262 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_origin.pdf b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..325962ab7e4e93dcd711fc1ea747692a0ef85130 --- /dev/null +++ b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/1f181b75-8c01-4fb8-ae75-f4b58a4586d2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3a831fedf63df66bfdfe1b239dcee856b5a7695368efc51bf08e51565cebd2a +size 3068426 diff --git a/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/full.md b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d982e1915d69c9bc9b19fbc8b5bab62bde8d666 --- /dev/null +++ b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/full.md @@ -0,0 +1,290 @@ +# Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation + +Chen-Chen Zong $^{1}$ , Ye-Wen Wang $^{1}$ , Kun-Peng Ning $^{2}$ , Hai-Bo Ye $^{1}$ , and Sheng-Jun Huang $^{1}$ † + +$^{1}$ Nanjing University of Aeronautics and Astronautics + $^{2}$ Peking University + +Abstract. Active learning (AL) in open set scenarios presents a novel challenge of identifying the most valuable examples in an unlabeled data pool that comprises data from both known and unknown classes. Traditional methods prioritize selecting informative examples with low confidence, with the risk of mistakenly selecting unknown-class examples with similarly low confidence. Recent methods favor the most probable known-class examples, with the risk of picking simple already mastered examples. In this paper, we attempt to query examples that are both likely from known classes and highly informative, and propose a Bidirectional Uncertainty-based Active Learning (BUAL) framework. Specifically, we achieve this by first pushing the unknown class examples toward regions with high-confidence predictions, i.e., the proposed Random Label Negative Learning method. Then, we propose a Bidirectional Uncertainty sampling strategy by jointly estimating uncertainty posed by both positive and negative learning to perform consistent and stable sampling. BUAL successfully extends existing uncertainty-based AL methods to complex open-set scenarios. Extensive experiments on multiple datasets with varying openness demonstrate that BUAL achieves state-of-the-art performance. The code is available at this link. + +Keywords: Active learning $\cdot$ Open-set annotation $\cdot$ Negative learning $\cdot$ Uncertainty estimation + +# 1 Introduction + +Labeling data can be costly and time-consuming, often requiring high levels of expertise from annotators [26]. This expense poses a significant challenge when dealing with insufficient labeled data in deep learning tasks. Recently, active learning (AL) has emerged as a prominent approach to tackle this issue and has gained widespread attention [9,17,21]. It iteratively selects the most informative examples from the unlabeled data pool and queries their labels from an oracle, enabling the learning of an effective model with reduced labeling costs. + +Existing AL methods [5,8,19,22,27,29,30] typically operate under the closed-set assumption, assuming that the label categories in the unlabeled data pool + +match those of the target task. However, this assumption often does not hold in practical scenarios. For example, consider a task that involves classifying images into two target categories, "Dog" and "Cat". Collecting training examples through keyword-based image search inevitably introduces irrelevant images from other categories (i.e., unknown classes), alongside the two target categories (i.e., known classes). + +In such open-set scenarios, many previous AL methods, which prefer querying examples with less confident predictions, may lead to failure since examples from unknown classes often receive uncertain predictions. To mitigate the impact of unknown class examples, some AL methods designed specifically for open-set scenarios attempt to query examples that are more likely to belong to known classes based on sample similarity [3] or model-predicted max activation value (MAV) [20]. However, since examples similar to the labeled ones may be already mastered by the model, they do not significantly benefit the target model. These methods only perform well when the proportion of unknown class examples is high. When the proportion is low, they tend to perform poorly than traditional AL methods, even inferior to random sampling (please refer to Figure 6). However, determining the proportion of unknown class examples in practical scenarios is often challenging. This further limits the usability of these methods. + +In this paper, we start with a specific question: can we effectively distinguish the "informative" examples of known classes from examples of unknown classes? Intuitively, if we can push the unknown class examples toward regions with high-confidence predictions, existing uncertainty-based AL methods can be applied directly in open-set scenarios. To achieve this, we propose to fine-tune the model by performing negative learning (NL) [11, 14, 16, 32] on unlabeled examples. NL is an indirect learning manner that explores the utility of complementary labels, i.e., the label categories that an instance does not belong to. For a $K$ -classification problem, the NL loss is defined as: + +$$ +\ell_ {N L} (f, \bar {y}) = - \sum_ {k = 1} ^ {K} \bar {y} _ {k} \log \left(1 - p _ {k}\right). \tag {1} +$$ + +where $f$ is the model we want to optimize, $\bar{y}$ and $\bar{\pmb{y}} = [\bar{y}_1,\dots ,\bar{y}_k,\dots ,\bar{y}_K]$ represent a complementary label and its corresponding one-hot form, respectively, and $p_k$ denotes the probability of the $k$ -th category. + +Specifically, the fine-tuning process comprises two parts. On one hand, for already labeled examples, we train them directly using Equation 1. On the other hand, for unlabeled examples, we first randomly assign labels to them in each training round and then train the model using Equation 1. Notably, the unlabeled known class example has a relatively higher chance of receiving the correct label, whereas the unknown class example will never be assigned the correct label. Once unlabeled known class data receive the correct labels, they suffer a larger penalty and are reduced confidence predictions by the model since they deviate from the distribution information obtained from labeled data. In contrast, unlabeled unknown class data are not constrained to move towards the high-confidence region for counteracting the update gradient produced by the labeled data. + +To validate this, we conducted preliminary experiments on CIFAR-10 [12] with 4 known and 6 unknown classes. Figure 1 illustrates the confidence statis + +![](images/51f0717bad27da9b553adaa3ae6338739ea5d9c077930a51d622e0c5caf020d9.jpg) +(a) w/o NL + +![](images/65d474dc8457d970367c7a84e1887c2500e10c7d564a4059ad7c0465b076f113.jpg) +(b) w/ NL +Fig. 1: The statistics of prediction confidence before and after fine-tuning the model. In the zoomed-in area of Figure 1b, we swapped the display order of the two to prevent occlusion, allowing for a more intuitive view of how the distribution has changed. + +tics for all unlabeled examples in a fixed query round before and after fine-tuning the model. As expected, known class examples are more prevalent in the low-confidence region, while unknown class ones are more common in high-confidence regions. This distribution characteristic presents a potential solution to the aforementioned question and offers a promising approach to AL for open-set scenarios. + +Based on this, we propose a Bidirectional Uncertainty-based Active Learning framework (BUAL). On the one hand, we propose the Random Label Negative Learning (RLNL) method to fine-tune the model and leverage information from the unlabeled data pool. Specifically, we first train a model using all labeled data as a positive classifier. Then, negative learning is performed to fine-tune the model as a negative classifier by randomly assigning labels to data from the unlabeled pool in each training iteration. This effectively distinguishes known class examples with lower confidence predictions from unknown class ones. On the other hand, we propose a Bidirectional Uncertainty (BU) sampling strategy for active selection, which estimates prediction uncertainty from both positive and negative classifiers. By selecting examples with the highest uncertainty, we expect to identify the most informative instances from the known classes. Experiments are performed on multiple datasets with different unknown class ratios. The results demonstrate that BUAL can query more informative known class examples and that the model performance obtained by BUAL is substantially improved compared with existing state-of-the-art methods. + +# 2 Related Work + +Active learning (AL) is a prominent approach aimed at reducing label costs by selecting a batch of examples that are most valuable for model training. Existing AL methods can be broadly categorized into three groups based on sample selection strategies: uncertainty-based, representative-based, and hybrid strategies + +which combine both aspects. Uncertainty-based strategies focus on sampling informative instances to reduce model uncertainty. Typical methods include Least Confident Sampling [15], Margin-based Sampling [2], and Entropy-based Sampling [7], etc. Representative-based strategies start from the sample distribution and aim to select representative instances that match the overall distribution. A typical method is Coreset [25]. Hybrid strategies combine uncertainty and representativeness by incorporating sample distribution information and the model's specific needs. Notable methods in this category include QUIRE [8] and BADGE [1], etc. + +Open-set annotation (OSA) involves active learning under open-set scenarios [3,20]. Existing methods primarily focus on selecting examples that are most likely to belong to known classes. For instance, CCAL [3] employs contrastive learning to extract semantic and distinctive features of examples, facilitating discrimination of known class examples. LfOSA [20] introduces an auxiliary network to model the per-example max activation value (MAV) distribution and dynamically selects examples with the highest probability from known classes. However, these methods exhibit sensitivity to the openness of the dataset and may not consistently perform well. Open-set recognition (OSR) [18, 23, 24] is a related problem setting to OSA, aiming to predict correct labels for known class examples while simultaneously detecting examples from unknown classes. Nevertheless, direct use of OSR methods often falls short of expected effectiveness [20], mainly due to limited training examples and the inability to identify highly informative instances. + +Complementary labels are labels other than the ground-truth label assigned to an example. Complementary label learning (CLL) [4,10,31] was initially introduced in [10], where the authors leveraged the lower acquisition cost and higher correct labeling rate of complementary labels to address multi-class classification problems. However, complementary labels contain less information compared to the correct labels, which can result in slow convergence of the model and challenges in achieving the desired performance when directly using complementary labels for training. Authors in [11] made the first attempt to combine CLL with the noise labeling problem [33] and proposed an indirect and robust training schema called negative learning (NL). In this paper, we exploit the properties of NL and extend it to open-set and label-free settings for the first time. + +# 3 Methodology + +# 3.1 Preliminaries + +Notations. In open-set annotation (OSA) settings, there are two labeled data pools: $D_{l}^{kn o} = \{(x_{i}^{l},y_{i}^{l})\}_{i = 1}^{n_{l}^{kn o}}$ for known classes and $D_{l}^{unk} = \{(x_{i}^{l})\}_{i = 1}^{n_{l}^{unk}}$ for unknown classes, as well as an unlabeled data pool: $D_{u} = \{(x_{i}^{u})\}_{i = 1}^{n_{u}^{kn o}} \cup \{(x_{i}^{u})\}_{i = 1}^{n_{u}^{unk}}$ containing examples from both known and unknown classes. Each known class example belongs to one of $K$ known classes in the label space $\mathcal{V} = \{1,2,\dots ,K\}$ . The unknown class examples are uniformly grouped into one category, denoted + +![](images/0d52dbf4c357580650dae3fdff056a3ade9f2f20e52bdc0af2ea36d24804ef04.jpg) +Fig. 2: The framework of BUAL. A two-stage $K$ class classifier is maintained, where the first stage is trained in a normal manner saved as $f_{p}(\cdot)$ and the second stage is trained using the proposed random label negative learning method denoted as $f_{n}(\cdot)$ . An auxiliary $K + 1$ class classifier $f_{aux}(\cdot)$ is trained in parallel. By collecting the predicted uncertainty from $f_{p}(\cdot)$ and $f_{n}(\cdot)$ on each candidate example along with the global and local balancing factors, the proposed bidirectional sampling strategy can accurately estimate the potential utility of each example and perform effective sample sampling under complex open-set scenarios. + +as $y = \emptyset$ . In each training round, active learning (AL) queries a batch of $b$ examples according to a given query strategy $\mathcal{A}$ , denoted as $X_{query} = X_{query}^{kn}\cup X_{query}^{unk}$ . Once the labeled feedback is obtained, we can calculate a ratio $r = \frac{|X_{query}^{kn}|}{|X_{query}|}$ to represent the precision of known classes. + +Overview. The proposed Bidirectional Uncertainty-based Active Learning (BUAL) framework is depicted in Figure 2. In each iteration, there are in general three steps: + +- Model Training: the algorithm first trains a target classifier in a normal learning manner, dubbed positive classifier $f_{p}(\cdot)$ . Then, the algorithm finetunes the classifier with a new classifier head by the proposed Random Label Negative Learning method, dubbed negative classifier $f_{n}(\cdot)$ . Similar to [20], a $K + 1$ auxiliary classifier $f_{aux}(\cdot)$ is trained in parallel. +- Example Selection: the Bidirectional Uncertainty Sampling Strategy estimates uncertainty bidirectionally using both the positive and negative classifier heads. This estimation is combined with dynamic balance factors generated by the auxiliary classifier and query feedback to select the most informative known class examples. +- Oracle Labeling: the annotators assign class labels to the selected examples. Based on the feedback results, update the corresponding data pools accordingly. + +# 3.2 Random Label Negative Learning + +In OSA scenarios, conventional uncertainty-based AL methods tend to be ineffective for example selection, mainly due to the unconfident predictions generated for the unknown class examples. Existing OSA methods prioritize sample purity without fully exploring sample informativeness, resulting in queries with too many already mastered examples that are not useful for model training. Thus, a key question arises: how to distinguish between unknown and "informative" known class examples. + +To cope with this problem, we propose a general method by pushing unknown class examples toward the high-confidence regions, while pushing known class examples to the low-confidence regions. If this separation is achieved, we can leverage existing uncertainty-based AL methods directly to handle various complex open-set scenarios. Fortunately, we achieve this to some extent by employing negative learning (NL), where we assign random labels to unlabeled examples and fine-tune the model with a new classifier head accordingly, dubbed Random Label Negative Learning (RLNL). + +Specifically, in the model training stage, we first train a $K$ -class classifier $f_{p}(\cdot)$ in a normal training manner (e.g., cross-entropy loss) based on $D_{l}^{kn o}$ . Note that $f_{p}(\cdot)$ can be the target model we eventually need to output. After training on the labeled data, $f_{p}(\cdot)$ can own a good discriminative ability for known class examples at the representation level. Then, we replace the last layer of $f_{p}(\cdot)$ and fine-tune a new classifier head $f_{n}(\cdot)$ with Equation 1. In this phase, unlabeled examples in $D_{u}$ are involved in training to help achieve the goal mentioned above: pushing unknown and known class examples toward the high and low confidence regions, respectively. All labeled examples in $D_{l}^{kn o}$ are still involved to prevent them from being incorrectly shifted in the distribution. Eventually, we assign random labels to the unlabeled examples in $D_{u}$ , while using the complementary labels instead for the labeled known class examples, i.e., for an example $\mathbf{x}$ : + +$$ +P (\bar {y} = s) = \frac {1}{| S |}, s \in \left\{ \begin{array}{l l} S = \mathcal {Y} \backslash y ^ {l}, & \text {i f} \boldsymbol {x} \in D _ {l} ^ {\text {k n o}}, \\ S = \mathcal {Y}, & \text {i f} \boldsymbol {x} \in D _ {u}. \end{array} \right. \tag {2} +$$ + +where $\bar{y}$ is uniformly sampled at each training iteration. Here, considering that the number of examples in $D_{u}$ is usually much larger than that in $D_{l}^{kn o}$ , a small subset $D_{sub}$ is randomly selected from the $D_{u}$ to save training cost. + +Why RLNL works? The first question one may hold is why it shouldn't be optimal for the unlabeled examples to be the uniform distribution as $\bar{y}$ is sampled at each iteration. To explain this, we take two different kinds of updating for a single example in the binary classification scenario of "Dog vs. Cat" as illustrations: negative learning using all labels at each iteration and negative learning using only one random label at each iteration. One possible case is shown in Figure 3. If each iteration uses all labels for negative learning, i.e., such an example is neither a "Cat" nor a "Dog", it is apparent that an update gradient of 0 is only achieved when the example is at the decision boundary. In contrast, if each iteration gives only one random label for negative learning, + +![](images/77a46191c36b5738078276f88203ac4bc75aa78105bde38a1319bd1fd167ef0c.jpg) +Fig. 3: Use all labels per iteration for negative learning (left) vs. use one random label per iteration for negative learning (right). + +![](images/e4abca5303e80168f9e3a88dac26c47d8499efa862d5c582f5612b56a0ee9d42.jpg) + +![](images/388cb1c695eac5e43f954282bcbc200ba686f40c136791cfc6e29a952bf3917c.jpg) +Fig. 4: The possible RLNL update scenario for unlabeled unknown class data in batch deep learning manner. The green "→" is the batch update gradient produced by the example itself, and the purple "→" is the update gradient produced by labeled data. Initially, the decision boundary is close to the left-hand category. + +i.e., such an example is not a "Cat" or a "Dog", then for an instance that is not a "Cat", pushing it as much as possible towards the "Dog" side will result in a gradient update of 0, and vice versa. Obviously, there is no fixed optimal scenario for this type of data. + +In RLNL, we utilize this property by exploiting the prior knowledge contained in the labeled data. For unlabeled known class data, they might have overlapping regions with labeled data in the feature space, benefiting from the feature representations learned in the previous stage and the simultaneous introduction of labeled data for negative learning. The mapping of such examples in the feature space will remain or be close to the labeled ones in subsequent model updates, as they will receive invisible constraints provided by the prior knowledge from the labeled data, owing to similar features. + +As shown in Figure 4, we present the possible update scenario for unlabeled unknown class examples in the commonly adopted batch deep learning manner. + +![](images/7d965843017f0d70958e9f3b25f3de37e51e290cea9d29995474d94d3865322d.jpg) +(a) Before RLNL + +![](images/a8c05da7c5bcf0670046fd643aa281d966411bfa6b566d780a7e2f828adefabb.jpg) +(b) After RLNL +Fig. 5: The t-SNE feature visualization of labeled data, unlabeled known class data, and unlabeled unknown class data on CIFAR-10 with an openness ratio of 0.5 before and after performing RLNL. For a more intuitive visualization, we only show a single known class. More visualization results are shown in the supplementary file. + +Here, the green arrow indicates the batch update gradient produced by itself, while the purple arrow indicates the batch update gradient produced by labeled data. Different from Figure 3, where data oscillates on both sides of the decision boundary, the unlabeled unknown class examples will oscillate at uncharted away from the decision boundary to counteract the update gradient due to the labeled ones. In contrast, unlabeled known class examples will move much less than unknown class ones in magnitude in the feature space within the constraints of the prior knowledge provided by the labeled data. This is further confirmed in Figure 5, which illustrates the t-SNE feature visualization of labeled data, unlabeled known class data, and unlabeled unknown class data before and after performing RLNL. This ultimately leads to the result in Figure 1b and proves that RLNL does work. + +# 3.3 Bidirectional Uncertainty Sampling Strategy + +During the fine-tuning process, we observed that the predictions of $f_{n}(\cdot)$ for examples in $D_{u}$ tend to oscillate between epochs. To ensure stable sampling, on one hand, we test all unlabeled examples $t$ times at $m$ round intervals to obtain the predicted probabilities $\pmb{p}_{\pmb{t}}^{-} = (p_{1}^{t},\dots,p_{K}^{t})$ , which will further be averaged as $\mathcal{P}^{-} = \frac{1}{t}\sum_{i = 1}^{t}\pmb{p}_{i}^{-} = (\frac{1}{t}\sum_{i = 1}^{t}p_{1}^{i},\dots,\frac{1}{t}\sum_{i = 1}^{t}p_{K}^{i})$ . On the other hand, we reuse the predictions of $f_{p}(\cdot)$ to produce predicted probabilities $\pmb{p}^{+} = (p_{1},\dots,p_{K})$ for all unlabeled examples, as $f_{p}(\cdot)$ are accurate for measuring known class ones. + +Compared to the positive head, the negative head is slightly biased for the measurement of sample uncertainty due to the unstable training. Therefore, if an example is likelier to belong to known classes, we prefer to utilize the sample uncertainty obtained from $f_{p}(\cdot)$ . On the contrary, once an example has a higher risk of belonging to the unknown classes, the uncertainty obtained from $f_{p}(\cdot)$ is unreliable, and thus the uncertainty by $f_{n}(\cdot)$ should be given a higher weight. + +To achieve this, we introduce two balancing factors based on global and local observations. In open-set scenarios, some unknown class examples might be + +mistakenly selected for annotation. Although these examples cannot be directly used for training the target model, they are valuable for measuring sample purity. Therefore, similar to [20], we train a $K + 1$ classifier $f_{aux}(\cdot)$ in a normal training manner based on the examples from both $D_l^{kno}$ and $D_l^{unk}$ . Then, we can obtain the predicted probability $p_{K + 1}^{aux}(\boldsymbol{x}) = f_{aux}(\emptyset \mid \boldsymbol{x})$ for each example. This can act as a local balancing factor, since the larger the value of $p_{K + 1}^{aux}(\boldsymbol{x})$ the more likely $\boldsymbol{x}$ is to belong to the unknown class. + +Additionally, once the selected examples are sent to the oracle for annotation, we can calculate the ratio $r$ , which provides a rough estimate of the current openness of $D_{u}$ and can serve as a global balancing factor. With the two balancing factors, we propose a bidirectional uncertainty sampling strategy defined as follows: + +$$ +\boldsymbol {x} ^ {*} = \arg \max _ {\boldsymbol {x}} p _ {K + 1} ^ {a u x} (\boldsymbol {x}) \operatorname {u n c} _ {n} + r \left[ 1 - p _ {K + 1} ^ {a u x} (\boldsymbol {x}) \right] \operatorname {u n c} _ {p}, \tag {3} +$$ + +where $unc_{p}$ and $unc_{n}$ denote the uncertainty of $\mathbf{x}$ to the positive classifier head and the negative classifier head, respectively. Noteworthy, this sampling strategy remains applicable even in closed-set settings. If there are no unknown class examples, the ratio $r$ will always be equal to 1, and $p_{K + 1}^{aux}(\mathbf{x})$ will be 0, effectively making the strategy equivalent to a normal uncertainty sampling strategy. + +With Equation 3, extending the existing closed-set uncertainty-based active learning methods to open-set scenarios is possible. In this paper, we focus on three classical uncertainty sampling strategies: least confident sampling, margin-based sampling, and entropy-based sampling. The corresponding modified versions for open-set scenarios are as follows: + +- Bidirectional Least Confident Sampling + +$$ +\boldsymbol {x} ^ {*} = \underset {\boldsymbol {x}} {\arg \max } p _ {K + 1} ^ {a u x} (\boldsymbol {x}) \left[ 1 - \boldsymbol {\mathcal {P}} _ {y ^ {-}} ^ {-} (\boldsymbol {x}) \right] + r \left[ 1 - p _ {K + 1} ^ {a u x} (\boldsymbol {x}) \right] \left[ 1 - \boldsymbol {\mathcal {P}} _ {y ^ {+}} ^ {+} (\boldsymbol {x}) \right], \tag {4} +$$ + +where $y^{-} = \arg \max_{y}\mathcal{P}_{y}^{-}(\pmb {x})$ $y^{+} = \arg \max_{y}\pmb{p}_{y}^{+}(\pmb {x})$ + +- Bidirectional Margin-Based Sampling + +$$ +\begin{array}{l} \boldsymbol {x} ^ {*} = \underset {\boldsymbol {x}} {\arg \max } p _ {K + 1} ^ {a u x} (\boldsymbol {x}) \left[ \boldsymbol {\mathcal {P}} _ {y _ {1} ^ {-}} ^ {-} (\boldsymbol {x}) - \boldsymbol {\mathcal {P}} _ {y _ {2} ^ {-}} ^ {-} (\boldsymbol {x}) \right] \tag {5} \\ + r \left[ 1 - p _ {K + 1} ^ {a u x} (\boldsymbol {x}) \right] \left[ \boldsymbol {p} _ {y _ {1} ^ {+}} ^ {+} (\boldsymbol {x}) - \boldsymbol {p} _ {y _ {2} ^ {+}} ^ {+} (\boldsymbol {x}) \right], \\ \end{array} +$$ + +where $y_1^- = \arg \max_y \mathcal{P}_y^-(\boldsymbol{x})$ , $y_1^+ = \arg \max_y \mathbf{p}_y^+(\boldsymbol{x})$ , $y_2^- = \arg \max_{y \setminus y_1^-} \mathcal{P}_y^-(\boldsymbol{x})$ , $y_2^+ = \arg \max_{y \setminus y_1^+} \mathbf{p}_y^+(\boldsymbol{x})$ . + +- Bidirectional Entropy-Based Sampling + +$$ +\begin{array}{l} \boldsymbol {x} ^ {*} = \underset {\boldsymbol {x}} {\arg \max } p _ {K + 1} ^ {a u x} (\boldsymbol {x}) \left[ - \boldsymbol {\mathcal {P}} _ {y ^ {-}} ^ {-} (\boldsymbol {x}) \log \boldsymbol {\mathcal {P}} _ {y ^ {-}} ^ {-} (\boldsymbol {x}) \right] \tag {6} \\ + r \left[ 1 - p _ {K + 1} ^ {a u x} (\boldsymbol {x}) \right] \left[ - \boldsymbol {p} _ {y ^ {+}} ^ {+} (\boldsymbol {x}) \log \boldsymbol {p} _ {y ^ {+}} ^ {+} (\boldsymbol {x}) \right], \\ \end{array} +$$ + +where $y^{-}$ and $y^{+}$ are consistent with the previous definition. + +The main procedures of BUAL are summarized in Algorithm 1. + +Algorithm 1 BUAL Training Procedure +Input: $D_l^{kno}$ $D_{l}^{unk}$ $D_u$ $r$ $\mathcal{A},b,m,t,$ subset size s, training epoch $E$ , optimizer $\mathcal{O}$ +Output: $D_l^{kno}$ $D_{l}^{unk}$ $D_u$ $r$ , model parameters $\theta_p$ +Process: +1: # Model training +2: for $j = 1;j\leq E$ do +3: $\mathcal{L}_p = \sum_{(x,y)\in D_l^{kno}}\ell (f_p(\boldsymbol {x}),y);$ +4: $\mathcal{L}_{aux} = \sum_{(x,y)\in D_l^{kno}\cup D_l^{unk}}\ell (f_{aux}(\boldsymbol {x}),y);$ +5: Update $\theta_p,\theta_{aux} = \mathcal{O}(\mathcal{L}_p,\theta_p),\mathcal{O}(\mathcal{L}_{aux},\theta_{aux})$ +6: end for +7: Use $f_{aux}(\pmb {x})$ to remove confidently unknown examples in $D_{u}$ and randomly select s examples as $D_{sub}$ +8: for $j = 1;j\leq E$ do +9: Generate label $\bar{y}$ for each example. +10: $\mathcal{L}_n = \sum_{(x,y)\in D_l^{kno}\cup D_{sub}}\ell_{NL}(f_n(\boldsymbol {x}),\bar{y})$ +11: Update $\theta_{n} = \mathcal{O}(\mathcal{L}_{n},\theta_{n})$ +12: if $j$ mod $m = 0$ then +13: Obtain $\pmb {p}_j^-$ for each sample in $D_{u}$ by $f_{n}(\pmb {x})$ +14: end if +15: end for +16: # Example selection +17: Calculate $\pmb{\mathcal{P}}^{-}$ for each sample across rounds; +18: Obtain $\pmb{p}^{+}$ for each sample in $D_u$ by $f_{p}(\pmb {x})$ +19: Obtain $p_{K + 1}^{aux}$ for each sample in $D_u$ by $f_{aux}(\pmb {x})$ +20: Obtain $X_{query}$ using BU sampling strategy A; +21: # Oracle labeling +22: Ask for annotation and update $D_l^{kno}$ $D_l^{unk}$ $D_u$ and $r$ + +# 4 Experiments + +Datasets. We conduct experiments on three benchmark datasets: CIFAR-10, CIFAR-100 [12], and Tiny-Imagenet [28]. Tiny-Imagenet is a subset of theImagenet [13], consisting of 200 classes with 500 training images per class. The openness of each dataset is defined as the ratio of unknown classes to the total number of classes, and we set its value to 0.2, 0.4, 0.6, and 0.8 for all datasets. + +Comparing methods. We select nine AL strategies for comparison, which can be further categorized into six groups: (1) Random: Randomly select examples from the unlabeled data pool for labeling. (2) Traditional uncertainty-based strategies: Least confident sampling (LC), Margin-based sampling (Margin), and Entropy-based sampling (Entropy). (3) Diversity-based strategy: Coreset. (4) Hybrid-based strategy: BADGE. (5) OSA methods: CCAL and LfOSA. (6) OSR method: DIAS [18]. Correspondingly, our methods are Bidirectional Least confident sampling (B-LC), Bidirectional Margin-based sampling (B-Margin), and Bidirectional Entropy-based sampling (B-Entropy). + +Training details. On CIFAR-10, CIFAR-100, and Tiny-Imagenet, we randomly sample $1\%$ , $8\%$ , and $8\%$ known class examples as the initial labeled data, + +![](images/cff1b691b242530b6d54613ab4b2921b1c28f5f0241bc7869989d239ebef8566.jpg) + +![](images/85e63ded2dd2fc13ac3e4569a550eb88de88b58420c269d8dd580b906e238166.jpg) + +![](images/371e5f2bc69e03e258b3cbd9aba53ae85296a2738e202810c37223d0fc0faa79.jpg) + +![](images/40fa1eff4b66cabb8bbbc4d41e5c0b22cb79c7c2c92f8918f05a79ca264591f3.jpg) + +![](images/76d182aed50eab9644a1a43d8ab731d35d90b694ea85d9a8ec1fb1b184fc490f.jpg) +Fig. 6: Accuracy comparison on CIFAR-10 (first row), CIFAR-100 (second row), and Tiny-Imagenet (third row). The ratio of unknown class examples to the total number of examples is fixed at 0.4 (first column) and 0.6 (second column) for each dataset. + +![](images/08d914508efc827b7fe8f1f2b24e5805f5a37b4d25792b2f5b5fd5c1b15e60c5.jpg) + +respectively. All the models involved in the experiments are ResNet18 [6], trained for 100 epochs, using SGD as the optimizer, where the learning rate is 0.01, momentum is 0.9, weight decay is 1e-4, and batch size is 128. We perform the experiments for 3 runs and report the average results. For CIFAR-10 and CIFAR-100, 5000 examples are randomly selected from the unlabeled pool as $D_{sub}$ , and 1500 examples are queried in each query round. Due to the doubling of data volume, we randomly select 10000 examples as $D_{sub}$ and query 3000 examples in each query round for Tiny-Imagenet. + +# 4.1 Performance Comparison + +Figure 6 presents the variation curves for the classification accuracy of the proposed methods and the comparison methods. Here, to better observe the variation curves, we only show the B-Margin for the proposed methods and the + +Table 1: The final round average accuracy of different methods on CIFAR-10, CIFAR-100, and Tiny-Imagenet. The best performance is highlighted in bold. + +
DatasetsCIFAR-10CIFAR-100Tiny-Imagenet
Openness Ratio0.20.40.60.80.20.40.60.80.20.40.60.8
(1)Random83.382.587.296.957.658.358.761.245.747.250.955.0
(2)LC84.381.687.596.255.854.654.056.244.845.948.451.6
Margin86.084.189.097.059.359.658.858.946.447.150.854.0
Entropy85.483.488.096.857.156.855.756.444.644.546.950.7
(3)Coreset85.081.886.497.460.261.261.864.246.247.851.854.0
(4)BADGE86.884.289.296.460.260.860.462.046.347.851.853.3
(5)LfOSA73.778.787.098.652.356.662.468.242.546.652.459.9
CCAL80.881.588.098.155.960.064.767.744.446.350.357.0
(6)DIAS81.880.783.094.055.756.156.957.243.145.147.554.4
OursB-LC87.087.292.599.159.362.867.572.145.748.754.760.6
B-Margin86.587.092.698.960.963.168.371.546.549.555.761.2
B-Entropy86.987.492.699.158.961.766.971.445.447.555.261.0
+ +![](images/c01d33cdef9aaf1c78b23ac952bb348aa46fbc7bbd8a934c89db6939ad33ff61.jpg) +Fig. 7: The average recognition rate on CIFAR-10 (first column), CIFAR-100 (second column), and Tiny-Imagenet (third column). + +![](images/25f703ff279500275e32b25baf652250e16f41c56953375ea7a9e148bde62e9b.jpg) + +![](images/4742a081036cd031165ec35a7202c312f0dff2010cb292f003ac6494176ce33a.jpg) + +Margin for the compared traditional uncertainty methods. Table 1 reports all methods' final round average accuracy. + +We can observe that our proposed methods consistently achieve the highest classification accuracy, demonstrating the effectiveness and superiority of our BUAL framework over other methods. Some notable observations are as follows: 1) The OSA methods, CCAL and LfOSA, gradually lose effectiveness as the proportion of known classes in the dataset increases. In contrast, traditional uncertainty-based methods perform better as the openness ratio decreases. This is consistent with our previous analysis. 2) The performance of Coreset and BADGE also deteriorates when the openness ratio is high. These methods tend to select examples with diverse characteristics. However, unknown class examples often differ significantly in characteristics from the labeled ones, making them more likely to be sampled and thus undermining the effectiveness of this type of method. 3) The OSR method DIAS does not perform well in all situations. The + +![](images/c185748c886ed0a9e0dd312ac2ebf0b8035bdd4b7c4829722300107fe6ad15da.jpg) +(a) LfOSA + +![](images/80f7fc80398aed826af2e304857124edcaed22458b5151a7be8d4b06ff79aa35.jpg) +(b) BUAL (ours) +Fig. 8: The t-SNE feature visualization of data from one query and labeled pool on CIFAR-10 with an openness ratio of 0.5. + +limited labeled data prevents the model from learning a robust representation for identifying open-set examples effectively. Moreover, DIAS cannot identify highly informative examples, resulting in its queries often being simple and unhelpful examples. 4) All our methods remain stable and do not suffer significantly from changes in the openness ratio. This can demonstrate two points. The dynamic balance factors in the proposed framework can adaptively assign appropriate weights to positive and negative uncertainties regardless of the openness of the dataset, and negative uncertainties are indeed effective for querying highly informative known class examples. + +Figure 7 shows the average recognition rate of known class examples across queries and openness ratios. We can observe that traditional AL methods all perform poorly due to their inability to recognize known class examples from those with high uncertainty and/or representative. As an OSR method, DIAS shows only marginal improvement compared to the traditional AL methods, unable to fully exploit its exceptional performance with limited training data. The recognition ability of CCAL decreases slightly with the increase of dataset categories compared to our method. It is noted that LfOSA achieved the highest query precision on all datasets. However, combining the results in Table 1, we can find that the performance of models trained by LfOSA is not satisfactory compared to other methods. To explain why, we visualize the queried examples' feature representation and labeled examples' feature representation for our method and LfOSA in Figure 8. We can observe that the features of examples queried by LfOSA are highly overlapping with the labeled ones, which are often examples that the model has already mastered and cannot provide an effective bonus for model training. On the contrary, the sample features queried by our method have little overlap and are more distributed in low-density regions, consistent with the goal of AL. These results validate the effectiveness of our approach in querying more informative examples of known classes and maintaining a high recognition rate. + +Table 2: Final accuracy of each component in Equation 3 on CIFAR-10, CIFAR-100, and Tiny-Imagenet with an openness ratio of 0.6. + +
MethodDatasetuncpuncnw/o ww/o fauxB-LC
CIFAR-1087.589.490.891.392.5
CIFAR-10054.063.562.465.067.5
Tiny-Imagenet48.452.352.053.054.7
+ +# 4.2 Ablation Study + +The ablation study is conducted on CIFAR-10 with an openness ratio of 0.6 to validate the effectiveness of each component in our proposed query strategy (Equation 3). The final round accuracy is shown in Table 2. Here, $\text{unc}_n$ and $\text{unc}_p$ indicate that only $\text{unc}_n$ and $\text{unc}_p$ is adopted for active sampling, respectively. $w / o$ denotes the removal of all balancing factors, i.e., both $f_{aux}$ and $r$ . $w / o$ means only removing the local balancing factor $f_{aux}$ . + +We can observe that removing any of the components leads to performance degradation. Although $\mathbf{unc}_n$ and $\mathbf{unc}_p$ are significantly less effective than B-LC, $\mathbf{unc}_n$ has a substantial improvement compared to $\mathbf{unc}_p$ due to its ability to distinguish informative examples of known classes from examples of unknown classes. The direct combination of $\mathbf{unc}_n$ and $\mathbf{unc}_p$ , i.e., $\mathbf{w} / \mathbf{o} \mathbf{w}$ , works better than $\mathbf{unc}_n$ , as models trained by RLNL may produce oscillate output and thus the uncertainty obtained in $\mathbf{unc}_n$ is not necessarily accurate. By adding the dynamic global balancing factor $r$ , $\mathbf{w} / \mathbf{o} f_{aux}$ achieves a better performance. However, it still falls short in comparison to B-LC, which validates the effectiveness of local balancing factor $f_{aux}$ . These results further corroborate the soundness of our strategy design. + +# 5 Conclusion + +In this paper, we successfully expand the existing uncertainty-based active learning methods to complex and ever-changing open-set scenarios by proposing a Bidirectional uncertainty-based Active Learning (BUAL) framework. On one hand, to achieve the goal of distinguishing known and unknown class examples with high uncertainty, we propose a simple but effective Random Label Negative Learning (RLNL) method for pushing unknown and known class examples toward the high and low confidence regions respectively. On the other hand, to better measure sample uncertainty, we propose a Bidirectional Uncertainty (BU) sampling strategy by dynamically fusing the sample uncertainty obtained from positive learning and negative learning. The dynamic balancing factors in it can ensure that the strategy is effective under various openness ratios. Extensive experimental results show that the model trained with BUAL can achieve state-of-the-art performance under various open-set scenarios. + +# Acknowledgements + +This work was supported by the Natural Science Foundation of Jiangsu Province of China (BK20222012, BK20211517), the National Key R&D Program of China (2020AAA0107000), and NSFC (62222605). + +# References + +1. Ash, J.T., Zhang, C., Krishnamurthy, A., Langford, J., Agarwal, A.: Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671 (2019) +2. Balcan, M.F., Broder, A., Zhang, T.: Margin based active learning. In: International Conference on Computational Learning Theory. pp. 35-50. Springer (2007) +3. Du, P., Zhao, S., Chen, H., Chai, S., Chen, H., Li, C.: Contrastive coding for active learning under class distribution mismatch. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8927-8936 (2021) +4. Feng, L., Kaneko, T., Han, B., Niu, G., An, B., Sugiyama, M.: Learning with multiple complementary labels. In: International Conference on Machine Learning. pp. 3072-3081. PMLR (2020) +5. Fu, Y., Zhu, X., Li, B.: A survey on instance selection for active learning. Knowledge and information systems 35(2), 249-283 (2013) +6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +7. Holub, A., Perona, P., Burl, M.C.: Entropy-based active learning for object recognition. In: 2008 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops. pp. 1-8. IEEE (2008) +8. Huang, S.J., Jin, R., Zhou, Z.H.: Active learning by querying informative and representative examples. Advances in neural information processing systems 23 (2010) +9. Huang, S.J., Zong, C.C., Ning, K.P., Ye, H.B.: Asynchronous active learning with distributed label querying. In: IJCAI. pp. 2570-2576 (2021) +0. Ishida, T., Niu, G., Hu, W., Sugiyama, M.: Learning from complementary labels. Advances in neural information processing systems 30 (2017) +1. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 101-110 (2019) +2. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +3. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012) +4. Lee, J.H., Astrid, M., Zaheer, M.Z., Lee, S.I.: Deep visual anomaly detection with negative learning. In: International Workshop on Frontiers of Computer Vision. pp. 218-232. Springer (2021) +5. Li, M., Sethi, I.K.: Confidence-based active learning. IEEE transactions on pattern analysis and machine intelligence 28(8), 1251-1261 (2006) +6. Luo, X., Chen, W., Tan, Y., Li, C., He, Y., Jia, X.: Exploiting negative learning for implicit pseudo label rectification in source-free domain adaptive semantic segmentation. arXiv preprint arXiv:2106.12123 (2021) + +17. Mahmood, R., Fidler, S., Law, M.T.: Low budget active learning via Wasserstein distance: An integer programming approach. arXiv preprint arXiv:2106.02968 (2021) +18. Moon, W., Park, J., Seong, H.S., Cho, C.H., Heo, J.P.: Difficulty-aware simulator for open set recognition. arXiv preprint arXiv:2207.10024 (2022) +19. Ning, K.P., Tao, L., Chen, S., Huang, S.J.: Improving model robustness by adaptively correcting perturbation levels with active queries. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 9161-9169 (2021) +20. Ning, K.P., Zhao, X., Li, Y., Huang, S.J.: Active learning for open-set annotation. arXiv preprint arXiv:2201.06758 (2022) +21. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Gupta, B.B., Chen, X., Wang, X.: A survey of deep active learning. ACM computing surveys (CSUR) 54(9), 1-40 (2021) +22. Roy, N., McCallum, A.: Toward optimal active learning through sampling estimation of error reduction. int. conf. on machine learning (2001) +23. Salehi, M., Mirzaei, H., Hendrycks, D., Li, Y., Rohban, M.H., Sabokrou, M.: A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051 (2021) +24. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012) +25. Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489 (2017) +26. Settles, B.: Active learning literature survey (2009) +27. Sinha, S., Ebrahimi, S., Darrell, T.: Variational adversarial active learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5972-5981 (2019) +28. Yao, L., Miller, J.: Tiny imagenet classification with convolutional neural networks. CS 231N 2(5), 8 (2015) +29. Yoo, D., Kweon, I.S.: Learning loss for active learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 93-102 (2019) +30. You, X., Wang, R., Tao, D.: Diverse expected gradient active learning for relative attributes. IEEE transactions on image processing 23(7), 3203-3217 (2014) +31. Yu, X., Liu, T., Gong, M., Tao, D.: Learning with biased complementary labels. In: Proceedings of the European conference on computer vision (ECCV). pp. 68-83 (2018) +32. Zong, C.C., Cao, Z.T., Guo, H.T., Du, Y., Xie, M.K., Li, S.Y., Huang, S.J.: Noise-robust bidirectional learning with dynamic sample reweighting. arXiv preprint arXiv:2209.01334 (2022) +33. Zong, C.C., Wang, Y.W., Xie, M.K., Huang, S.J.: Dirichlet-based prediction calibration for learning with noisy labels. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 17254-17262 (2024) \ No newline at end of file diff --git a/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/images.zip b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a6fe536058d079b26194cf064007305de9597c42 --- /dev/null +++ b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce7d0454b89bfd60f8e3d61e16d1283fb96cfd10aba93cc99d9a3bcc52c52b0e +size 563403 diff --git a/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/layout.json b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..45b99bd1cea5094395be0e937404ed238b46a0dc --- /dev/null +++ b/2024/Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation/layout.json @@ -0,0 +1,9453 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "text", + "content": "Chen-Chen Zong" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "text", + "content": ", Ye-Wen Wang" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "text", + "content": ", Kun-Peng Ning" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "text", + "content": ", Hai-Bo Ye" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "text", + "content": ", and Sheng-Jun Huang" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 167, + 476, + 194 + ], + "type": "text", + "content": "†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 196, + 201, + 416, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 201, + 416, + 224 + ], + "spans": [ + { + "bbox": [ + 196, + 201, + 416, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 196, + 201, + 416, + 224 + ], + "type": "text", + "content": " Nanjing University of Aeronautics and Astronautics \n" + }, + { + "bbox": [ + 196, + 201, + 416, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 196, + 201, + 416, + 224 + ], + "type": "text", + "content": " Peking University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 253, + 455, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 253, + 455, + 462 + ], + "spans": [ + { + "bbox": [ + 160, + 253, + 455, + 462 + ], + "type": "text", + "content": "Abstract. Active learning (AL) in open set scenarios presents a novel challenge of identifying the most valuable examples in an unlabeled data pool that comprises data from both known and unknown classes. Traditional methods prioritize selecting informative examples with low confidence, with the risk of mistakenly selecting unknown-class examples with similarly low confidence. Recent methods favor the most probable known-class examples, with the risk of picking simple already mastered examples. In this paper, we attempt to query examples that are both likely from known classes and highly informative, and propose a Bidirectional Uncertainty-based Active Learning (BUAL) framework. Specifically, we achieve this by first pushing the unknown class examples toward regions with high-confidence predictions, i.e., the proposed Random Label Negative Learning method. Then, we propose a Bidirectional Uncertainty sampling strategy by jointly estimating uncertainty posed by both positive and negative learning to perform consistent and stable sampling. BUAL successfully extends existing uncertainty-based AL methods to complex open-set scenarios. Extensive experiments on multiple datasets with varying openness demonstrate that BUAL achieves state-of-the-art performance. The code is available at this link." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 472, + 453, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 472, + 453, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 472, + 453, + 495 + ], + "type": "text", + "content": "Keywords: Active learning " + }, + { + "bbox": [ + 160, + 472, + 453, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 472, + 453, + 495 + ], + "type": "text", + "content": " Open-set annotation " + }, + { + "bbox": [ + 160, + 472, + 453, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 472, + 453, + 495 + ], + "type": "text", + "content": " Negative learning " + }, + { + "bbox": [ + 160, + 472, + 453, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 472, + 453, + 495 + ], + "type": "text", + "content": " Uncertainty estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 539, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 622 + ], + "type": "text", + "content": "Labeling data can be costly and time-consuming, often requiring high levels of expertise from annotators [26]. This expense poses a significant challenge when dealing with insufficient labeled data in deep learning tasks. Recently, active learning (AL) has emerged as a prominent approach to tackle this issue and has gained widespread attention [9,17,21]. It iteratively selects the most informative examples from the unlabeled data pool and queries their labels from an oracle, enabling the learning of an effective model with reduced labeling costs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "content": "Existing AL methods [5,8,19,22,27,29,30] typically operate under the closed-set assumption, assuming that the label categories in the unlabeled data pool" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 335, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 335, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 335, + 666 + ], + "type": "text", + "content": "† Corresponding Author: huangsj@nuaa.edu.cn" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "match those of the target task. However, this assumption often does not hold in practical scenarios. For example, consider a task that involves classifying images into two target categories, \"Dog\" and \"Cat\". Collecting training examples through keyword-based image search inevitably introduces irrelevant images from other categories (i.e., unknown classes), alongside the two target categories (i.e., known classes)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 342 + ], + "type": "text", + "content": "In such open-set scenarios, many previous AL methods, which prefer querying examples with less confident predictions, may lead to failure since examples from unknown classes often receive uncertain predictions. To mitigate the impact of unknown class examples, some AL methods designed specifically for open-set scenarios attempt to query examples that are more likely to belong to known classes based on sample similarity [3] or model-predicted max activation value (MAV) [20]. However, since examples similar to the labeled ones may be already mastered by the model, they do not significantly benefit the target model. These methods only perform well when the proportion of unknown class examples is high. When the proportion is low, they tend to perform poorly than traditional AL methods, even inferior to random sampling (please refer to Figure 6). However, determining the proportion of unknown class examples in practical scenarios is often challenging. This further limits the usability of these methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 481, + 450 + ], + "type": "text", + "content": "In this paper, we start with a specific question: can we effectively distinguish the \"informative\" examples of known classes from examples of unknown classes? Intuitively, if we can push the unknown class examples toward regions with high-confidence predictions, existing uncertainty-based AL methods can be applied directly in open-set scenarios. To achieve this, we propose to fine-tune the model by performing negative learning (NL) [11, 14, 16, 32] on unlabeled examples. NL is an indirect learning manner that explores the utility of complementary labels, i.e., the label categories that an instance does not belong to. For a " + }, + { + "bbox": [ + 130, + 343, + 481, + 450 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 343, + 481, + 450 + ], + "type": "text", + "content": "-classification problem, the NL loss is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 228, + 455, + 480, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 455, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 228, + 455, + 480, + 469 + ], + "type": "interline_equation", + "content": "\\ell_ {N L} (f, \\bar {y}) = - \\sum_ {k = 1} ^ {K} \\bar {y} _ {k} \\log \\left(1 - p _ {k}\\right). \\tag {1}", + "image_path": "79056d17c64e09521c99c0ff420abcc056d0bf7c1ef6148ad88db748f73f7d56.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "content": " is the model we want to optimize, " + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "inline_equation", + "content": "\\bar{y}" + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{y}} = [\\bar{y}_1,\\dots ,\\bar{y}_k,\\dots ,\\bar{y}_K]" + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "content": " represent a complementary label and its corresponding one-hot form, respectively, and " + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "content": " denotes the probability of the " + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "content": "-th category." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 640 + ], + "type": "text", + "content": "Specifically, the fine-tuning process comprises two parts. On one hand, for already labeled examples, we train them directly using Equation 1. On the other hand, for unlabeled examples, we first randomly assign labels to them in each training round and then train the model using Equation 1. Notably, the unlabeled known class example has a relatively higher chance of receiving the correct label, whereas the unknown class example will never be assigned the correct label. Once unlabeled known class data receive the correct labels, they suffer a larger penalty and are reduced confidence predictions by the model since they deviate from the distribution information obtained from labeled data. In contrast, unlabeled unknown class data are not constrained to move towards the high-confidence region for counteracting the update gradient produced by the labeled data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "To validate this, we conducted preliminary experiments on CIFAR-10 [12] with 4 known and 6 unknown classes. Figure 1 illustrates the confidence statis" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Zong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 296, + 245 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 296, + 245 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 296, + 245 + ], + "type": "image", + "image_path": "51f0717bad27da9b553adaa3ae6338739ea5d9c077930a51d622e0c5caf020d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 250, + 238, + 259 + ], + "lines": [ + { + "bbox": [ + 195, + 250, + 238, + 259 + ], + "spans": [ + { + "bbox": [ + 195, + 250, + 238, + 259 + ], + "type": "text", + "content": "(a) w/o NL" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 317, + 116, + 477, + 245 + ], + "blocks": [ + { + "bbox": [ + 317, + 116, + 477, + 245 + ], + "lines": [ + { + "bbox": [ + 317, + 116, + 477, + 245 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 477, + 245 + ], + "type": "image", + "image_path": "65d474dc8457d970367c7a84e1887c2500e10c7d564a4059ad7c0465b076f113.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 250, + 417, + 259 + ], + "lines": [ + { + "bbox": [ + 378, + 250, + 417, + 259 + ], + "spans": [ + { + "bbox": [ + 378, + 250, + 417, + 259 + ], + "type": "text", + "content": "(b) w/ NL" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 268, + 480, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 268, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 480, + 301 + ], + "type": "text", + "content": "Fig. 1: The statistics of prediction confidence before and after fine-tuning the model. In the zoomed-in area of Figure 1b, we swapped the display order of the two to prevent occlusion, allowing for a more intuitive view of how the distribution has changed." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 324, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 480, + 384 + ], + "type": "text", + "content": "tics for all unlabeled examples in a fixed query round before and after fine-tuning the model. As expected, known class examples are more prevalent in the low-confidence region, while unknown class ones are more common in high-confidence regions. This distribution characteristic presents a potential solution to the aforementioned question and offers a promising approach to AL for open-set scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 384, + 481, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 481, + 576 + ], + "type": "text", + "content": "Based on this, we propose a Bidirectional Uncertainty-based Active Learning framework (BUAL). On the one hand, we propose the Random Label Negative Learning (RLNL) method to fine-tune the model and leverage information from the unlabeled data pool. Specifically, we first train a model using all labeled data as a positive classifier. Then, negative learning is performed to fine-tune the model as a negative classifier by randomly assigning labels to data from the unlabeled pool in each training iteration. This effectively distinguishes known class examples with lower confidence predictions from unknown class ones. On the other hand, we propose a Bidirectional Uncertainty (BU) sampling strategy for active selection, which estimates prediction uncertainty from both positive and negative classifiers. By selecting examples with the highest uncertainty, we expect to identify the most informative instances from the known classes. Experiments are performed on multiple datasets with different unknown class ratios. The results demonstrate that BUAL can query more informative known class examples and that the model performance obtained by BUAL is substantially improved compared with existing state-of-the-art methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 593, + 237, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 237, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 237, + 605 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "Active learning (AL) is a prominent approach aimed at reducing label costs by selecting a batch of examples that are most valuable for model training. Existing AL methods can be broadly categorized into three groups based on sample selection strategies: uncertainty-based, representative-based, and hybrid strategies" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "text", + "content": "BUAL for Open-Set Annotation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "which combine both aspects. Uncertainty-based strategies focus on sampling informative instances to reduce model uncertainty. Typical methods include Least Confident Sampling [15], Margin-based Sampling [2], and Entropy-based Sampling [7], etc. Representative-based strategies start from the sample distribution and aim to select representative instances that match the overall distribution. A typical method is Coreset [25]. Hybrid strategies combine uncertainty and representativeness by incorporating sample distribution information and the model's specific needs. Notable methods in this category include QUIRE [8] and BADGE [1], etc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 390 + ], + "type": "text", + "content": "Open-set annotation (OSA) involves active learning under open-set scenarios [3,20]. Existing methods primarily focus on selecting examples that are most likely to belong to known classes. For instance, CCAL [3] employs contrastive learning to extract semantic and distinctive features of examples, facilitating discrimination of known class examples. LfOSA [20] introduces an auxiliary network to model the per-example max activation value (MAV) distribution and dynamically selects examples with the highest probability from known classes. However, these methods exhibit sensitivity to the openness of the dataset and may not consistently perform well. Open-set recognition (OSR) [18, 23, 24] is a related problem setting to OSA, aiming to predict correct labels for known class examples while simultaneously detecting examples from unknown classes. Nevertheless, direct use of OSR methods often falls short of expected effectiveness [20], mainly due to limited training examples and the inability to identify highly informative instances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 391, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 522 + ], + "type": "text", + "content": "Complementary labels are labels other than the ground-truth label assigned to an example. Complementary label learning (CLL) [4,10,31] was initially introduced in [10], where the authors leveraged the lower acquisition cost and higher correct labeling rate of complementary labels to address multi-class classification problems. However, complementary labels contain less information compared to the correct labels, which can result in slow convergence of the model and challenges in achieving the desired performance when directly using complementary labels for training. Authors in [11] made the first attempt to combine CLL with the noise labeling problem [33] and proposed an indirect and robust training schema called negative learning (NL). In this paper, we exploit the properties of NL and extend it to open-set and label-free settings for the first time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 540, + 233, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 233, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 233, + 555 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 565, + 228, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 228, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 228, + 576 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "content": "Notations. In open-set annotation (OSA) settings, there are two labeled data pools: " + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{l}^{kn o} = \\{(x_{i}^{l},y_{i}^{l})\\}_{i = 1}^{n_{l}^{kn o}}" + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "content": " for known classes and " + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{l}^{unk} = \\{(x_{i}^{l})\\}_{i = 1}^{n_{l}^{unk}}" + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "content": " for unknown classes, as well as an unlabeled data pool: " + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{u} = \\{(x_{i}^{u})\\}_{i = 1}^{n_{u}^{kn o}} \\cup \\{(x_{i}^{u})\\}_{i = 1}^{n_{u}^{unk}}" + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "content": " containing examples from both known and unknown classes. Each known class example belongs to one of " + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "content": " known classes in the label space " + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{1,2,\\dots ,K\\}" + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "content": ". The unknown class examples are uniformly grouped into one category, denoted" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Zong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 121, + 478, + 264 + ], + "blocks": [ + { + "bbox": [ + 138, + 121, + 478, + 264 + ], + "lines": [ + { + "bbox": [ + 138, + 121, + 478, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 121, + 478, + 264 + ], + "type": "image", + "image_path": "0d52dbf4c357580650dae3fdff056a3ade9f2f20e52bdc0af2ea36d24804ef04.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "text", + "content": "Fig. 2: The framework of BUAL. A two-stage " + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "text", + "content": " class classifier is maintained, where the first stage is trained in a normal manner saved as " + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "text", + "content": " and the second stage is trained using the proposed random label negative learning method denoted as " + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "inline_equation", + "content": "f_{n}(\\cdot)" + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "text", + "content": ". An auxiliary " + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "text", + "content": " class classifier " + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "inline_equation", + "content": "f_{aux}(\\cdot)" + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "text", + "content": " is trained in parallel. By collecting the predicted uncertainty from " + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "inline_equation", + "content": "f_{n}(\\cdot)" + }, + { + "bbox": [ + 130, + 277, + 482, + 365 + ], + "type": "text", + "content": " on each candidate example along with the global and local balancing factors, the proposed bidirectional sampling strategy can accurately estimate the potential utility of each example and perform effective sample sampling under complex open-set scenarios." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "inline_equation", + "content": "y = \\emptyset" + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "text", + "content": ". In each training round, active learning (AL) queries a batch of " + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "text", + "content": " examples according to a given query strategy " + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "inline_equation", + "content": "X_{query} = X_{query}^{kn}\\cup X_{query}^{unk}" + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "text", + "content": ". Once the labeled feedback is obtained, we can calculate a ratio " + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "inline_equation", + "content": "r = \\frac{|X_{query}^{kn}|}{|X_{query}|}" + }, + { + "bbox": [ + 130, + 394, + 482, + 449 + ], + "type": "text", + "content": " to represent the precision of known classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 451, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 489 + ], + "type": "text", + "content": "Overview. The proposed Bidirectional Uncertainty-based Active Learning (BUAL) framework is depicted in Figure 2. In each iteration, there are in general three steps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 505, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "text", + "content": "- Model Training: the algorithm first trains a target classifier in a normal learning manner, dubbed positive classifier " + }, + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "text", + "content": ". Then, the algorithm finetunes the classifier with a new classifier head by the proposed Random Label Negative Learning method, dubbed negative classifier " + }, + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "f_{n}(\\cdot)" + }, + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "text", + "content": ". Similar to [20], a " + }, + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "text", + "content": " auxiliary classifier " + }, + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "f_{aux}(\\cdot)" + }, + { + "bbox": [ + 138, + 505, + 482, + 566 + ], + "type": "text", + "content": " is trained in parallel." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 567, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 567, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 137, + 567, + 482, + 628 + ], + "type": "text", + "content": "- Example Selection: the Bidirectional Uncertainty Sampling Strategy estimates uncertainty bidirectionally using both the positive and negative classifier heads. This estimation is combined with dynamic balance factors generated by the auxiliary classifier and query feedback to select the most informative known class examples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "content": "- Oracle Labeling: the annotators assign class labels to the selected examples. Based on the feedback results, update the corresponding data pools accordingly." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "type": "text", + "content": "BUAL for Open-Set Annotation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 331, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 331, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 331, + 129 + ], + "type": "text", + "content": "3.2 Random Label Negative Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "content": "In OSA scenarios, conventional uncertainty-based AL methods tend to be ineffective for example selection, mainly due to the unconfident predictions generated for the unknown class examples. Existing OSA methods prioritize sample purity without fully exploring sample informativeness, resulting in queries with too many already mastered examples that are not useful for model training. Thus, a key question arises: how to distinguish between unknown and \"informative\" known class examples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 316 + ], + "type": "text", + "content": "To cope with this problem, we propose a general method by pushing unknown class examples toward the high-confidence regions, while pushing known class examples to the low-confidence regions. If this separation is achieved, we can leverage existing uncertainty-based AL methods directly to handle various complex open-set scenarios. Fortunately, we achieve this to some extent by employing negative learning (NL), where we assign random labels to unlabeled examples and fine-tune the model with a new classifier head accordingly, dubbed Random Label Negative Learning (RLNL)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": "Specifically, in the model training stage, we first train a " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": "-class classifier " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": " in a normal training manner (e.g., cross-entropy loss) based on " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "D_{l}^{kn o}" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": " can be the target model we eventually need to output. After training on the labeled data, " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": " can own a good discriminative ability for known class examples at the representation level. Then, we replace the last layer of " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": " and fine-tune a new classifier head " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "f_{n}(\\cdot)" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": " with Equation 1. In this phase, unlabeled examples in " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "D_{u}" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": " are involved in training to help achieve the goal mentioned above: pushing unknown and known class examples toward the high and low confidence regions, respectively. All labeled examples in " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "D_{l}^{kn o}" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": " are still involved to prevent them from being incorrectly shifted in the distribution. Eventually, we assign random labels to the unlabeled examples in " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "D_{u}" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": ", while using the complementary labels instead for the labeled known class examples, i.e., for an example " + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 317, + 482, + 460 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 470, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 470, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 201, + 470, + 481, + 502 + ], + "type": "interline_equation", + "content": "P (\\bar {y} = s) = \\frac {1}{| S |}, s \\in \\left\\{ \\begin{array}{l l} S = \\mathcal {Y} \\backslash y ^ {l}, & \\text {i f} \\boldsymbol {x} \\in D _ {l} ^ {\\text {k n o}}, \\\\ S = \\mathcal {Y}, & \\text {i f} \\boldsymbol {x} \\in D _ {u}. \\end{array} \\right. \\tag {2}", + "image_path": "524540d9dcf30ca4a0e9a2677c33e25b8af95076a338940805828b50ef3172b9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\bar{y}" + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "text", + "content": " is uniformly sampled at each training iteration. Here, considering that the number of examples in " + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "inline_equation", + "content": "D_{u}" + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "text", + "content": " is usually much larger than that in " + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "inline_equation", + "content": "D_{l}^{kn o}" + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "text", + "content": ", a small subset " + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "inline_equation", + "content": "D_{sub}" + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "text", + "content": " is randomly selected from the " + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "inline_equation", + "content": "D_{u}" + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "text", + "content": " to save training cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Why RLNL works? The first question one may hold is why it shouldn't be optimal for the unlabeled examples to be the uniform distribution as " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{y}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is sampled at each iteration. To explain this, we take two different kinds of updating for a single example in the binary classification scenario of \"Dog vs. Cat\" as illustrations: negative learning using all labels at each iteration and negative learning using only one random label at each iteration. One possible case is shown in Figure 3. If each iteration uses all labels for negative learning, i.e., such an example is neither a \"Cat\" nor a \"Dog\", it is apparent that an update gradient of 0 is only achieved when the example is at the decision boundary. In contrast, if each iteration gives only one random label for negative learning," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Zong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 162, + 116, + 318, + 242 + ], + "blocks": [ + { + "bbox": [ + 162, + 116, + 318, + 242 + ], + "lines": [ + { + "bbox": [ + 162, + 116, + 318, + 242 + ], + "spans": [ + { + "bbox": [ + 162, + 116, + 318, + 242 + ], + "type": "image", + "image_path": "77a46191c36b5738078276f88203ac4bc75aa78105bde38a1319bd1fd167ef0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 251, + 480, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 251, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 480, + 275 + ], + "type": "text", + "content": "Fig. 3: Use all labels per iteration for negative learning (left) vs. use one random label per iteration for negative learning (right)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 327, + 118, + 446, + 242 + ], + "blocks": [ + { + "bbox": [ + 327, + 118, + 446, + 242 + ], + "lines": [ + { + "bbox": [ + 327, + 118, + 446, + 242 + ], + "spans": [ + { + "bbox": [ + 327, + 118, + 446, + 242 + ], + "type": "image", + "image_path": "e4abca5303e80168f9e3a88dac26c47d8499efa862d5c582f5612b56a0ee9d42.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 184, + 289, + 432, + 414 + ], + "blocks": [ + { + "bbox": [ + 184, + 289, + 432, + 414 + ], + "lines": [ + { + "bbox": [ + 184, + 289, + 432, + 414 + ], + "spans": [ + { + "bbox": [ + 184, + 289, + 432, + 414 + ], + "type": "image", + "image_path": "388cb1c695eac5e43f954282bcbc200ba686f40c136791cfc6e29a952bf3917c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 424, + 482, + 470 + ], + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 470 + ], + "type": "text", + "content": "Fig. 4: The possible RLNL update scenario for unlabeled unknown class data in batch deep learning manner. The green \"→\" is the batch update gradient produced by the example itself, and the purple \"→\" is the update gradient produced by labeled data. Initially, the decision boundary is close to the left-hand category." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 495, + 480, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 480, + 543 + ], + "type": "text", + "content": "i.e., such an example is not a \"Cat\" or a \"Dog\", then for an instance that is not a \"Cat\", pushing it as much as possible towards the \"Dog\" side will result in a gradient update of 0, and vice versa. Obviously, there is no fixed optimal scenario for this type of data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": "In RLNL, we utilize this property by exploiting the prior knowledge contained in the labeled data. For unlabeled known class data, they might have overlapping regions with labeled data in the feature space, benefiting from the feature representations learned in the previous stage and the simultaneous introduction of labeled data for negative learning. The mapping of such examples in the feature space will remain or be close to the labeled ones in subsequent model updates, as they will receive invisible constraints provided by the prior knowledge from the labeled data, owing to similar features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "As shown in Figure 4, we present the possible update scenario for unlabeled unknown class examples in the commonly adopted batch deep learning manner." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "type": "text", + "content": "BUAL for Open-Set Annotation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 121, + 284, + 220 + ], + "blocks": [ + { + "bbox": [ + 140, + 121, + 284, + 220 + ], + "lines": [ + { + "bbox": [ + 140, + 121, + 284, + 220 + ], + "spans": [ + { + "bbox": [ + 140, + 121, + 284, + 220 + ], + "type": "image", + "image_path": "7d965843017f0d70958e9f3b25f3de37e51e290cea9d29995474d94d3865322d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 223, + 243, + 233 + ], + "lines": [ + { + "bbox": [ + 181, + 223, + 243, + 233 + ], + "spans": [ + { + "bbox": [ + 181, + 223, + 243, + 233 + ], + "type": "text", + "content": "(a) Before RLNL" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 118, + 479, + 220 + ], + "blocks": [ + { + "bbox": [ + 299, + 118, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 299, + 118, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 299, + 118, + 479, + 220 + ], + "type": "image", + "image_path": "a8c05da7c5bcf0670046fd643aa281d966411bfa6b566d780a7e2f828adefabb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 224, + 416, + 233 + ], + "lines": [ + { + "bbox": [ + 357, + 224, + 416, + 233 + ], + "spans": [ + { + "bbox": [ + 357, + 224, + 416, + 233 + ], + "type": "text", + "content": "(b) After RLNL" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 243, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 288 + ], + "type": "text", + "content": "Fig. 5: The t-SNE feature visualization of labeled data, unlabeled known class data, and unlabeled unknown class data on CIFAR-10 with an openness ratio of 0.5 before and after performing RLNL. For a more intuitive visualization, we only show a single known class. More visualization results are shown in the supplementary file." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": "Here, the green arrow indicates the batch update gradient produced by itself, while the purple arrow indicates the batch update gradient produced by labeled data. Different from Figure 3, where data oscillates on both sides of the decision boundary, the unlabeled unknown class examples will oscillate at uncharted away from the decision boundary to counteract the update gradient due to the labeled ones. In contrast, unlabeled known class examples will move much less than unknown class ones in magnitude in the feature space within the constraints of the prior knowledge provided by the labeled data. This is further confirmed in Figure 5, which illustrates the t-SNE feature visualization of labeled data, unlabeled known class data, and unlabeled unknown class data before and after performing RLNL. This ultimately leads to the result in Figure 1b and proves that RLNL does work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 467, + 386, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 386, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 386, + 480 + ], + "type": "text", + "content": "3.3 Bidirectional Uncertainty Sampling Strategy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": "During the fine-tuning process, we observed that the predictions of " + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "inline_equation", + "content": "f_{n}(\\cdot)" + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": " for examples in " + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "inline_equation", + "content": "D_{u}" + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": " tend to oscillate between epochs. To ensure stable sampling, on one hand, we test all unlabeled examples " + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": " times at " + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": " round intervals to obtain the predicted probabilities " + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{\\pmb{t}}^{-} = (p_{1}^{t},\\dots,p_{K}^{t})" + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": ", which will further be averaged as " + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{-} = \\frac{1}{t}\\sum_{i = 1}^{t}\\pmb{p}_{i}^{-} = (\\frac{1}{t}\\sum_{i = 1}^{t}p_{1}^{i},\\dots,\\frac{1}{t}\\sum_{i = 1}^{t}p_{K}^{i})" + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": ". On the other hand, we reuse the predictions of " + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": " to produce predicted probabilities " + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{+} = (p_{1},\\dots,p_{K})" + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": " for all unlabeled examples, as " + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 484, + 482, + 570 + ], + "type": "text", + "content": " are accurate for measuring known class ones." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": "Compared to the positive head, the negative head is slightly biased for the measurement of sample uncertainty due to the unstable training. Therefore, if an example is likelier to belong to known classes, we prefer to utilize the sample uncertainty obtained from " + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": ". On the contrary, once an example has a higher risk of belonging to the unknown classes, the uncertainty obtained from " + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": " is unreliable, and thus the uncertainty by " + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "inline_equation", + "content": "f_{n}(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": " should be given a higher weight." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "To achieve this, we introduce two balancing factors based on global and local observations. In open-set scenarios, some unknown class examples might be" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Zong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "mistakenly selected for annotation. Although these examples cannot be directly used for training the target model, they are valuable for measuring sample purity. Therefore, similar to [20], we train a " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " classifier " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "f_{aux}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " in a normal training manner based on the examples from both " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "D_l^{kno}" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "D_l^{unk}" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": ". Then, we can obtain the predicted probability " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "p_{K + 1}^{aux}(\\boldsymbol{x}) = f_{aux}(\\emptyset \\mid \\boldsymbol{x})" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " for each example. This can act as a local balancing factor, since the larger the value of " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "p_{K + 1}^{aux}(\\boldsymbol{x})" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " the more likely " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " is to belong to the unknown class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 479, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 479, + 258 + ], + "type": "text", + "content": "Additionally, once the selected examples are sent to the oracle for annotation, we can calculate the ratio " + }, + { + "bbox": [ + 130, + 200, + 479, + 258 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 200, + 479, + 258 + ], + "type": "text", + "content": ", which provides a rough estimate of the current openness of " + }, + { + "bbox": [ + 130, + 200, + 479, + 258 + ], + "type": "inline_equation", + "content": "D_{u}" + }, + { + "bbox": [ + 130, + 200, + 479, + 258 + ], + "type": "text", + "content": " and can serve as a global balancing factor. With the two balancing factors, we propose a bidirectional uncertainty sampling strategy defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 259, + 481, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 259, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 194, + 259, + 481, + 276 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {*} = \\arg \\max _ {\\boldsymbol {x}} p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\operatorname {u n c} _ {n} + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\operatorname {u n c} _ {p}, \\tag {3}", + "image_path": "6cdfc6e7ce3687fa1b3acb8d37d849ac7ec35296733aee4ae977bd204e391c2c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "inline_equation", + "content": "unc_{p}" + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "inline_equation", + "content": "unc_{n}" + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "text", + "content": " denote the uncertainty of " + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "text", + "content": " to the positive classifier head and the negative classifier head, respectively. Noteworthy, this sampling strategy remains applicable even in closed-set settings. If there are no unknown class examples, the ratio " + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "text", + "content": " will always be equal to 1, and " + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "inline_equation", + "content": "p_{K + 1}^{aux}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 279, + 479, + 338 + ], + "type": "text", + "content": " will be 0, effectively making the strategy equivalent to a normal uncertainty sampling strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 339, + 479, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 479, + 398 + ], + "type": "text", + "content": "With Equation 3, extending the existing closed-set uncertainty-based active learning methods to open-set scenarios is possible. In this paper, we focus on three classical uncertainty sampling strategies: least confident sampling, margin-based sampling, and entropy-based sampling. The corresponding modified versions for open-set scenarios are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 406, + 350, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 406, + 350, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 406, + 350, + 418 + ], + "type": "text", + "content": "- Bidirectional Least Confident Sampling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 148, + 425, + 480, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 425, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 148, + 425, + 480, + 447 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\arg \\max } p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\left[ 1 - \\boldsymbol {\\mathcal {P}} _ {y ^ {-}} ^ {-} (\\boldsymbol {x}) \\right] + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\left[ 1 - \\boldsymbol {\\mathcal {P}} _ {y ^ {+}} ^ {+} (\\boldsymbol {x}) \\right], \\tag {4}", + "image_path": "11c1fafedfdae656cb0d7fdce60e903df0056b892865ab59768da7807a2a6d6e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 451, + 361, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 451, + 361, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 451, + 361, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 451, + 361, + 466 + ], + "type": "inline_equation", + "content": "y^{-} = \\arg \\max_{y}\\mathcal{P}_{y}^{-}(\\pmb {x})" + }, + { + "bbox": [ + 131, + 451, + 361, + 466 + ], + "type": "inline_equation", + "content": "y^{+} = \\arg \\max_{y}\\pmb{p}_{y}^{+}(\\pmb {x})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 472, + 340, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 340, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 340, + 483 + ], + "type": "text", + "content": "- Bidirectional Margin-Based Sampling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 211, + 491, + 480, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 491, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 211, + 491, + 480, + 534 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\arg \\max } p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\left[ \\boldsymbol {\\mathcal {P}} _ {y _ {1} ^ {-}} ^ {-} (\\boldsymbol {x}) - \\boldsymbol {\\mathcal {P}} _ {y _ {2} ^ {-}} ^ {-} (\\boldsymbol {x}) \\right] \\tag {5} \\\\ + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\left[ \\boldsymbol {p} _ {y _ {1} ^ {+}} ^ {+} (\\boldsymbol {x}) - \\boldsymbol {p} _ {y _ {2} ^ {+}} ^ {+} (\\boldsymbol {x}) \\right], \\\\ \\end{array}", + "image_path": "5a1f8da5fa67e8aa733168ce06e4d6d4e7010f23c18566b37a4f893e8025adb5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "inline_equation", + "content": "y_1^- = \\arg \\max_y \\mathcal{P}_y^-(\\boldsymbol{x})" + }, + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "inline_equation", + "content": "y_1^+ = \\arg \\max_y \\mathbf{p}_y^+(\\boldsymbol{x})" + }, + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "inline_equation", + "content": "y_2^- = \\arg \\max_{y \\setminus y_1^-} \\mathcal{P}_y^-(\\boldsymbol{x})" + }, + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "inline_equation", + "content": "y_2^+ = \\arg \\max_{y \\setminus y_1^+} \\mathbf{p}_y^+(\\boldsymbol{x})" + }, + { + "bbox": [ + 131, + 538, + 479, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 574, + 345, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 574, + 345, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 574, + 345, + 586 + ], + "type": "text", + "content": "- Bidirectional Entropy-Based Sampling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 205, + 594, + 480, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 594, + 480, + 638 + ], + "spans": [ + { + "bbox": [ + 205, + 594, + 480, + 638 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\arg \\max } p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\left[ - \\boldsymbol {\\mathcal {P}} _ {y ^ {-}} ^ {-} (\\boldsymbol {x}) \\log \\boldsymbol {\\mathcal {P}} _ {y ^ {-}} ^ {-} (\\boldsymbol {x}) \\right] \\tag {6} \\\\ + r \\left[ 1 - p _ {K + 1} ^ {a u x} (\\boldsymbol {x}) \\right] \\left[ - \\boldsymbol {p} _ {y ^ {+}} ^ {+} (\\boldsymbol {x}) \\log \\boldsymbol {p} _ {y ^ {+}} ^ {+} (\\boldsymbol {x}) \\right], \\\\ \\end{array}", + "image_path": "75d13f550783c40108781e6d3c4a88fa269e4bb1c0c82c0506fa34006e9d00c1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 641, + 399, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 399, + 653 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 399, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 399, + 653 + ], + "type": "inline_equation", + "content": "y^{-}" + }, + { + "bbox": [ + 131, + 641, + 399, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 399, + 653 + ], + "type": "inline_equation", + "content": "y^{+}" + }, + { + "bbox": [ + 131, + 641, + 399, + 653 + ], + "type": "text", + "content": " are consistent with the previous definition." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 146, + 653, + 427, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 653, + 427, + 665 + ], + "spans": [ + { + "bbox": [ + 146, + 653, + 427, + 665 + ], + "type": "text", + "content": "The main procedures of BUAL are summarized in Algorithm 1." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "type": "text", + "content": "BUAL for Open-Set Annotation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 129, + 481, + 426 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 316, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 316, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 316, + 128 + ], + "type": "text", + "content": "Algorithm 1 BUAL Training Procedure" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_l^{kno}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_{l}^{unk}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_u" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{A},b,m,t," + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " subset size s, training epoch " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " , optimizer " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \nOutput: " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_l^{kno}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_{l}^{unk}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_u" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " , model parameters " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\theta_p" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \nProcess: \n1: # Model training \n2: for " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "j = 1;j\\leq E" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " do \n3: " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p = \\sum_{(x,y)\\in D_l^{kno}}\\ell (f_p(\\boldsymbol {x}),y);" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux} = \\sum_{(x,y)\\in D_l^{kno}\\cup D_l^{unk}}\\ell (f_{aux}(\\boldsymbol {x}),y);" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \n5: Update " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\theta_p,\\theta_{aux} = \\mathcal{O}(\\mathcal{L}_p,\\theta_p),\\mathcal{O}(\\mathcal{L}_{aux},\\theta_{aux})" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \n6: end for \n7: Use " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "f_{aux}(\\pmb {x})" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " to remove confidently unknown examples in " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_{u}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " and randomly select s examples as " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_{sub}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \n8: for " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "j = 1;j\\leq E" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " do \n9: Generate label " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\bar{y}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " for each example. \n10: " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_n = \\sum_{(x,y)\\in D_l^{kno}\\cup D_{sub}}\\ell_{NL}(f_n(\\boldsymbol {x}),\\bar{y})" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \n11: Update " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\theta_{n} = \\mathcal{O}(\\mathcal{L}_{n},\\theta_{n})" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \n12: if " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " mod " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "m = 0" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " then \n13: Obtain " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\pmb {p}_j^-" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " for each sample in " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_{u}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "f_{n}(\\pmb {x})" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \n14: end if \n15: end for \n16: # Example selection \n17: Calculate " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{\\mathcal{P}}^{-}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " for each sample across rounds; \n18: Obtain " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{+}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " for each sample in " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_u" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "f_{p}(\\pmb {x})" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \n19: Obtain " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "p_{K + 1}^{aux}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " for each sample in " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_u" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "f_{aux}(\\pmb {x})" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " \n20: Obtain " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "X_{query}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " using BU sampling strategy A; \n21: # Oracle labeling \n22: Ask for annotation and update " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_l^{kno}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_l^{unk}" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "D_u" + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 129, + 481, + 426 + ], + "type": "inline_equation", + "content": "r" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 449, + 229, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 229, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 229, + 463 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 480, + 533 + ], + "type": "text", + "content": "Datasets. We conduct experiments on three benchmark datasets: CIFAR-10, CIFAR-100 [12], and Tiny-Imagenet [28]. Tiny-Imagenet is a subset of theImagenet [13], consisting of 200 classes with 500 training images per class. The openness of each dataset is defined as the ratio of unknown classes to the total number of classes, and we set its value to 0.2, 0.4, 0.6, and 0.8 for all datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 641 + ], + "type": "text", + "content": "Comparing methods. We select nine AL strategies for comparison, which can be further categorized into six groups: (1) Random: Randomly select examples from the unlabeled data pool for labeling. (2) Traditional uncertainty-based strategies: Least confident sampling (LC), Margin-based sampling (Margin), and Entropy-based sampling (Entropy). (3) Diversity-based strategy: Coreset. (4) Hybrid-based strategy: BADGE. (5) OSA methods: CCAL and LfOSA. (6) OSR method: DIAS [18]. Correspondingly, our methods are Bidirectional Least confident sampling (B-LC), Bidirectional Margin-based sampling (B-Margin), and Bidirectional Entropy-based sampling (B-Entropy)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Training details. On CIFAR-10, CIFAR-100, and Tiny-Imagenet, we randomly sample " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " known class examples as the initial labeled data," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Zong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 116, + 305, + 220 + ], + "blocks": [ + { + "bbox": [ + 157, + 116, + 305, + 220 + ], + "lines": [ + { + "bbox": [ + 157, + 116, + 305, + 220 + ], + "spans": [ + { + "bbox": [ + 157, + 116, + 305, + 220 + ], + "type": "image", + "image_path": "cff1b691b242530b6d54613ab4b2921b1c28f5f0241bc7869989d239ebef8566.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 116, + 457, + 220 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 457, + 220 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 457, + 220 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 457, + 220 + ], + "type": "image", + "image_path": "85e63ded2dd2fc13ac3e4569a550eb88de88b58420c269d8dd580b906e238166.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 157, + 224, + 304, + 327 + ], + "blocks": [ + { + "bbox": [ + 157, + 224, + 304, + 327 + ], + "lines": [ + { + "bbox": [ + 157, + 224, + 304, + 327 + ], + "spans": [ + { + "bbox": [ + 157, + 224, + 304, + 327 + ], + "type": "image", + "image_path": "371e5f2bc69e03e258b3cbd9aba53ae85296a2738e202810c37223d0fc0faa79.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 224, + 456, + 327 + ], + "blocks": [ + { + "bbox": [ + 309, + 224, + 456, + 327 + ], + "lines": [ + { + "bbox": [ + 309, + 224, + 456, + 327 + ], + "spans": [ + { + "bbox": [ + 309, + 224, + 456, + 327 + ], + "type": "image", + "image_path": "40fa1eff4b66cabb8bbbc4d41e5c0b22cb79c7c2c92f8918f05a79ca264591f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 157, + 331, + 304, + 433 + ], + "blocks": [ + { + "bbox": [ + 157, + 331, + 304, + 433 + ], + "lines": [ + { + "bbox": [ + 157, + 331, + 304, + 433 + ], + "spans": [ + { + "bbox": [ + 157, + 331, + 304, + 433 + ], + "type": "image", + "image_path": "76d182aed50eab9644a1a43d8ab731d35d90b694ea85d9a8ec1fb1b184fc490f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 443, + 482, + 478 + ], + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 478 + ], + "type": "text", + "content": "Fig. 6: Accuracy comparison on CIFAR-10 (first row), CIFAR-100 (second row), and Tiny-Imagenet (third row). The ratio of unknown class examples to the total number of examples is fixed at 0.4 (first column) and 0.6 (second column) for each dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 331, + 457, + 432 + ], + "blocks": [ + { + "bbox": [ + 310, + 331, + 457, + 432 + ], + "lines": [ + { + "bbox": [ + 310, + 331, + 457, + 432 + ], + "spans": [ + { + "bbox": [ + 310, + 331, + 457, + 432 + ], + "type": "image", + "image_path": "08d914508efc827b7fe8f1f2b24e5805f5a37b4d25792b2f5b5fd5c1b15e60c5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 596 + ], + "type": "text", + "content": "respectively. All the models involved in the experiments are ResNet18 [6], trained for 100 epochs, using SGD as the optimizer, where the learning rate is 0.01, momentum is 0.9, weight decay is 1e-4, and batch size is 128. We perform the experiments for 3 runs and report the average results. For CIFAR-10 and CIFAR-100, 5000 examples are randomly selected from the unlabeled pool as " + }, + { + "bbox": [ + 130, + 498, + 482, + 596 + ], + "type": "inline_equation", + "content": "D_{sub}" + }, + { + "bbox": [ + 130, + 498, + 482, + 596 + ], + "type": "text", + "content": ", and 1500 examples are queried in each query round. Due to the doubling of data volume, we randomly select 10000 examples as " + }, + { + "bbox": [ + 130, + 498, + 482, + 596 + ], + "type": "inline_equation", + "content": "D_{sub}" + }, + { + "bbox": [ + 130, + 498, + 482, + 596 + ], + "type": "text", + "content": " and query 3000 examples in each query round for Tiny-Imagenet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 611, + 289, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 289, + 624 + ], + "type": "text", + "content": "4.1 Performance Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Figure 6 presents the variation curves for the classification accuracy of the proposed methods and the comparison methods. Here, to better observe the variation curves, we only show the B-Margin for the proposed methods and the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "text", + "content": "BUAL for Open-Set Annotation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 147, + 479, + 331 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: The final round average accuracy of different methods on CIFAR-10, CIFAR-100, and Tiny-Imagenet. The best performance is highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 147, + 479, + 331 + ], + "lines": [ + { + "bbox": [ + 137, + 147, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 137, + 147, + 479, + 331 + ], + "type": "table", + "html": "
DatasetsCIFAR-10CIFAR-100Tiny-Imagenet
Openness Ratio0.20.40.60.80.20.40.60.80.20.40.60.8
(1)Random83.382.587.296.957.658.358.761.245.747.250.955.0
(2)LC84.381.687.596.255.854.654.056.244.845.948.451.6
Margin86.084.189.097.059.359.658.858.946.447.150.854.0
Entropy85.483.488.096.857.156.855.756.444.644.546.950.7
(3)Coreset85.081.886.497.460.261.261.864.246.247.851.854.0
(4)BADGE86.884.289.296.460.260.860.462.046.347.851.853.3
(5)LfOSA73.778.787.098.652.356.662.468.242.546.652.459.9
CCAL80.881.588.098.155.960.064.767.744.446.350.357.0
(6)DIAS81.880.783.094.055.756.156.957.243.145.147.554.4
OursB-LC87.087.292.599.159.362.867.572.145.748.754.760.6
B-Margin86.587.092.698.960.963.168.371.546.549.555.761.2
B-Entropy86.987.492.699.158.961.766.971.445.447.555.261.0
", + "image_path": "303966f68a775627a0b786729ec511d29692e370bf2085f1d75ff851228b9bc1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 346, + 249, + 437 + ], + "blocks": [ + { + "bbox": [ + 138, + 346, + 249, + 437 + ], + "lines": [ + { + "bbox": [ + 138, + 346, + 249, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 249, + 437 + ], + "type": "image", + "image_path": "c01d33cdef9aaf1c78b23ac952bb348aa46fbc7bbd8a934c89db6939ad33ff61.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 448, + 480, + 471 + ], + "lines": [ + { + "bbox": [ + 130, + 448, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 480, + 471 + ], + "type": "text", + "content": "Fig. 7: The average recognition rate on CIFAR-10 (first column), CIFAR-100 (second column), and Tiny-Imagenet (third column)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 346, + 362, + 437 + ], + "blocks": [ + { + "bbox": [ + 253, + 346, + 362, + 437 + ], + "lines": [ + { + "bbox": [ + 253, + 346, + 362, + 437 + ], + "spans": [ + { + "bbox": [ + 253, + 346, + 362, + 437 + ], + "type": "image", + "image_path": "25f703ff279500275e32b25baf652250e16f41c56953375ea7a9e148bde62e9b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 366, + 346, + 475, + 437 + ], + "blocks": [ + { + "bbox": [ + 366, + 346, + 475, + 437 + ], + "lines": [ + { + "bbox": [ + 366, + 346, + 475, + 437 + ], + "spans": [ + { + "bbox": [ + 366, + 346, + 475, + 437 + ], + "type": "image", + "image_path": "4742a081036cd031165ec35a7202c312f0dff2010cb292f003ac6494176ce33a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 496, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 480, + 521 + ], + "type": "text", + "content": "Margin for the compared traditional uncertainty methods. Table 1 reports all methods' final round average accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "We can observe that our proposed methods consistently achieve the highest classification accuracy, demonstrating the effectiveness and superiority of our BUAL framework over other methods. Some notable observations are as follows: 1) The OSA methods, CCAL and LfOSA, gradually lose effectiveness as the proportion of known classes in the dataset increases. In contrast, traditional uncertainty-based methods perform better as the openness ratio decreases. This is consistent with our previous analysis. 2) The performance of Coreset and BADGE also deteriorates when the openness ratio is high. These methods tend to select examples with diverse characteristics. However, unknown class examples often differ significantly in characteristics from the labeled ones, making them more likely to be sampled and thus undermining the effectiveness of this type of method. 3) The OSR method DIAS does not perform well in all situations. The" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Zong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 119, + 299, + 240 + ], + "blocks": [ + { + "bbox": [ + 157, + 119, + 299, + 240 + ], + "lines": [ + { + "bbox": [ + 157, + 119, + 299, + 240 + ], + "spans": [ + { + "bbox": [ + 157, + 119, + 299, + 240 + ], + "type": "image", + "image_path": "c185748c886ed0a9e0dd312ac2ebf0b8035bdd4b7c4829722300107fe6ad15da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 246, + 247, + 255 + ], + "lines": [ + { + "bbox": [ + 208, + 246, + 247, + 255 + ], + "spans": [ + { + "bbox": [ + 208, + 246, + 247, + 255 + ], + "type": "text", + "content": "(a) LfOSA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 120, + 462, + 240 + ], + "blocks": [ + { + "bbox": [ + 313, + 120, + 462, + 240 + ], + "lines": [ + { + "bbox": [ + 313, + 120, + 462, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 462, + 240 + ], + "type": "image", + "image_path": "80f7fc80398aed826af2e304857124edcaed22458b5151a7be8d4b06ff79aa35.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 246, + 417, + 255 + ], + "lines": [ + { + "bbox": [ + 355, + 246, + 417, + 255 + ], + "spans": [ + { + "bbox": [ + 355, + 246, + 417, + 255 + ], + "type": "text", + "content": "(b) BUAL (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 265, + 480, + 286 + ], + "lines": [ + { + "bbox": [ + 131, + 265, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 480, + 286 + ], + "type": "text", + "content": "Fig. 8: The t-SNE feature visualization of data from one query and labeled pool on CIFAR-10 with an openness ratio of 0.5." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 422 + ], + "type": "text", + "content": "limited labeled data prevents the model from learning a robust representation for identifying open-set examples effectively. Moreover, DIAS cannot identify highly informative examples, resulting in its queries often being simple and unhelpful examples. 4) All our methods remain stable and do not suffer significantly from changes in the openness ratio. This can demonstrate two points. The dynamic balance factors in the proposed framework can adaptively assign appropriate weights to positive and negative uncertainties regardless of the openness of the dataset, and negative uncertainties are indeed effective for querying highly informative known class examples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 665 + ], + "type": "text", + "content": "Figure 7 shows the average recognition rate of known class examples across queries and openness ratios. We can observe that traditional AL methods all perform poorly due to their inability to recognize known class examples from those with high uncertainty and/or representative. As an OSR method, DIAS shows only marginal improvement compared to the traditional AL methods, unable to fully exploit its exceptional performance with limited training data. The recognition ability of CCAL decreases slightly with the increase of dataset categories compared to our method. It is noted that LfOSA achieved the highest query precision on all datasets. However, combining the results in Table 1, we can find that the performance of models trained by LfOSA is not satisfactory compared to other methods. To explain why, we visualize the queried examples' feature representation and labeled examples' feature representation for our method and LfOSA in Figure 8. We can observe that the features of examples queried by LfOSA are highly overlapping with the labeled ones, which are often examples that the model has already mastered and cannot provide an effective bonus for model training. On the contrary, the sample features queried by our method have little overlap and are more distributed in low-density regions, consistent with the goal of AL. These results validate the effectiveness of our approach in querying more informative examples of known classes and maintaining a high recognition rate." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 102 + ], + "type": "text", + "content": "BUAL for Open-Set Annotation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 147, + 455, + 230 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Final accuracy of each component in Equation 3 on CIFAR-10, CIFAR-100, and Tiny-Imagenet with an openness ratio of 0.6." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 147, + 455, + 230 + ], + "lines": [ + { + "bbox": [ + 159, + 147, + 455, + 230 + ], + "spans": [ + { + "bbox": [ + 159, + 147, + 455, + 230 + ], + "type": "table", + "html": "
MethodDatasetuncpuncnw/o ww/o fauxB-LC
CIFAR-1087.589.490.891.392.5
CIFAR-10054.063.562.465.067.5
Tiny-Imagenet48.452.352.053.054.7
", + "image_path": "10f04b45c307a01747d4f1447b5f6d10fef408eff19bc9cfa2b797169f030fcf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 251, + 238, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 238, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 238, + 262 + ], + "type": "text", + "content": "4.2 Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": "The ablation study is conducted on CIFAR-10 with an openness ratio of 0.6 to validate the effectiveness of each component in our proposed query strategy (Equation 3). The final round accuracy is shown in Table 2. Here, " + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "inline_equation", + "content": "\\text{unc}_n" + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "inline_equation", + "content": "\\text{unc}_p" + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": " indicate that only " + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "inline_equation", + "content": "\\text{unc}_n" + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "inline_equation", + "content": "\\text{unc}_p" + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": " is adopted for active sampling, respectively. " + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": " denotes the removal of all balancing factors, i.e., both " + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "inline_equation", + "content": "f_{aux}" + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": " means only removing the local balancing factor " + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "inline_equation", + "content": "f_{aux}" + }, + { + "bbox": [ + 130, + 267, + 481, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": "We can observe that removing any of the components leads to performance degradation. Although " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{unc}_n" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{unc}_p" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": " are significantly less effective than B-LC, " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{unc}_n" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": " has a substantial improvement compared to " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{unc}_p" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": " due to its ability to distinguish informative examples of known classes from examples of unknown classes. The direct combination of " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{unc}_n" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{unc}_p" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{w} / \\mathbf{o} \\mathbf{w}" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": ", works better than " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{unc}_n" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": ", as models trained by RLNL may produce oscillate output and thus the uncertainty obtained in " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{unc}_n" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": " is not necessarily accurate. By adding the dynamic global balancing factor " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{w} / \\mathbf{o} f_{aux}" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": " achieves a better performance. However, it still falls short in comparison to B-LC, which validates the effectiveness of local balancing factor " + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "inline_equation", + "content": "f_{aux}" + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": ". These results further corroborate the soundness of our strategy design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we successfully expand the existing uncertainty-based active learning methods to complex and ever-changing open-set scenarios by proposing a Bidirectional uncertainty-based Active Learning (BUAL) framework. On one hand, to achieve the goal of distinguishing known and unknown class examples with high uncertainty, we propose a simple but effective Random Label Negative Learning (RLNL) method for pushing unknown and known class examples toward the high and low confidence regions respectively. On the other hand, to better measure sample uncertainty, we propose a Bidirectional Uncertainty (BU) sampling strategy by dynamically fusing the sample uncertainty obtained from positive learning and negative learning. The dynamic balancing factors in it can ensure that the strategy is effective under various openness ratios. Extensive experimental results show that the model trained with BUAL can achieve state-of-the-art performance under various open-set scenarios." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Zong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 481, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 481, + 174 + ], + "type": "text", + "content": "This work was supported by the Natural Science Foundation of Jiangsu Province of China (BK20222012, BK20211517), the National Key R&D Program of China (2020AAA0107000), and NSFC (62222605)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "text", + "content": "1. Ash, J.T., Zhang, C., Krishnamurthy, A., Langford, J., Agarwal, A.: Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 267 + ], + "type": "text", + "content": "2. Balcan, M.F., Broder, A., Zhang, T.: Margin based active learning. In: International Conference on Computational Learning Theory. pp. 35-50. Springer (2007)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "type": "text", + "content": "3. Du, P., Zhao, S., Chen, H., Chai, S., Chen, H., Li, C.: Contrastive coding for active learning under class distribution mismatch. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8927-8936 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 331 + ], + "type": "text", + "content": "4. Feng, L., Kaneko, T., Han, B., Niu, G., An, B., Sugiyama, M.: Learning with multiple complementary labels. In: International Conference on Machine Learning. pp. 3072-3081. PMLR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "text", + "content": "5. Fu, Y., Zhu, X., Li, B.: A survey on instance selection for active learning. Knowledge and information systems 35(2), 249-283 (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 353, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 386 + ], + "type": "text", + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "text", + "content": "7. Holub, A., Perona, P., Burl, M.C.: Entropy-based active learning for object recognition. In: 2008 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops. pp. 1-8. IEEE (2008)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "text", + "content": "8. Huang, S.J., Jin, R., Zhou, Z.H.: Active learning by querying informative and representative examples. Advances in neural information processing systems 23 (2010)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 450, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 481, + 472 + ], + "type": "text", + "content": "9. Huang, S.J., Zong, C.C., Ning, K.P., Ye, H.B.: Asynchronous active learning with distributed label querying. In: IJCAI. pp. 2570-2576 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "type": "text", + "content": "0. Ishida, T., Niu, G., Hu, W., Sugiyama, M.: Learning from complementary labels. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "text", + "content": "1. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 101-110 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "type": "text", + "content": "2. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 546, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 481, + 578 + ], + "type": "text", + "content": "3. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "text", + "content": "4. Lee, J.H., Astrid, M., Zaheer, M.Z., Lee, S.I.: Deep visual anomaly detection with negative learning. In: International Workshop on Frontiers of Computer Vision. pp. 218-232. Springer (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "type": "text", + "content": "5. Li, M., Sethi, I.K.: Confidence-based active learning. IEEE transactions on pattern analysis and machine intelligence 28(8), 1251-1261 (2006)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "6. Luo, X., Chen, W., Tan, Y., Li, C., He, Y., Jia, X.: Exploiting negative learning for implicit pseudo label rectification in source-free domain adaptive semantic segmentation. arXiv preprint arXiv:2106.12123 (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "text", + "content": "BUAL for Open-Set Annotation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 588 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "17. Mahmood, R., Fidler, S., Law, M.T.: Low budget active learning via Wasserstein distance: An integer programming approach. arXiv preprint arXiv:2106.02968 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "18. Moon, W., Park, J., Seong, H.S., Cho, C.H., Heo, J.P.: Difficulty-aware simulator for open set recognition. arXiv preprint arXiv:2207.10024 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "19. Ning, K.P., Tao, L., Chen, S., Huang, S.J.: Improving model robustness by adaptively correcting perturbation levels with active queries. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 9161-9169 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "20. Ning, K.P., Zhao, X., Li, Y., Huang, S.J.: Active learning for open-set annotation. arXiv preprint arXiv:2201.06758 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "21. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Gupta, B.B., Chen, X., Wang, X.: A survey of deep active learning. ACM computing surveys (CSUR) 54(9), 1-40 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "22. Roy, N., McCallum, A.: Toward optimal active learning through sampling estimation of error reduction. int. conf. on machine learning (2001)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "text", + "content": "23. Salehi, M., Mirzaei, H., Hendrycks, D., Li, Y., Rohban, M.H., Sabokrou, M.: A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "24. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "25. Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 358, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 358, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 358, + 380 + ], + "type": "text", + "content": "26. Settles, B.: Active learning literature survey (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "27. Sinha, S., Ebrahimi, S., Darrell, T.: Variational adversarial active learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5972-5981 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "28. Yao, L., Miller, J.: Tiny imagenet classification with convolutional neural networks. CS 231N 2(5), 8 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "29. Yoo, D., Kweon, I.S.: Learning loss for active learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 93-102 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "30. You, X., Wang, R., Tao, D.: Diverse expected gradient active learning for relative attributes. IEEE transactions on image processing 23(7), 3203-3217 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "31. Yu, X., Liu, T., Gong, M., Tao, D.: Learning with biased complementary labels. In: Proceedings of the European conference on computer vision (ECCV). pp. 68-83 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "32. Zong, C.C., Cao, Z.T., Guo, H.T., Du, Y., Xie, M.K., Li, S.Y., Huang, S.J.: Noise-robust bidirectional learning with dynamic sample reweighting. arXiv preprint arXiv:2209.01334 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "33. Zong, C.C., Wang, Y.W., Xie, M.K., Huang, S.J.: Dirichlet-based prediction calibration for learning with noisy labels. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 17254-17262 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "C. Zong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_content_list.json b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..00e0ca1d894eb5c08f6c282d66df2f178e3dc627 --- /dev/null +++ b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_content_list.json @@ -0,0 +1,1927 @@ +[ + { + "type": "text", + "text": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Geyou Zhang $^{1}$ , Ce Zhu $^{1,*}$ , and Kai Liu $^{2}$", + "bbox": [ + 339, + 210, + 663, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Electronic Science and Technology of China, geyouzhang@foxmail.com, eczhu@uestc.edu.cn", + "bbox": [ + 302, + 238, + 699, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Sichuan University", + "bbox": [ + 431, + 268, + 571, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kailiu@scu.edu.cn", + "bbox": [ + 433, + 282, + 566, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Phase shifting profilometry (PSP) is favored in high-precision 3D scanning due to its high accuracy, robustness, and pixel-wise property. However, a fundamental assumption of PSP that the object should remain static is violated in dynamic measurement, making PSP susceptible to object moving, resulting in ripple-like errors in the point clouds. We propose a pixel-wise and frame-wise loopable binomial self-compensation (BSC) algorithm to effectively and flexibly eliminate motion error in the four-step PSP. Our mathematical model demonstrates that by summing successive motion-affected phase frames weighted by binomial coefficients, motion error exponentially diminishes as the binomial order increases, accomplishing automatic error compensation through the motion-affected phase sequence, without the assistance of any intermediate variable. Extensive experiments show that our BSC outperforms the existing methods in reducing motion error, while achieving a depth map frame rate equal to the camera's acquisition rate (90 fps), enabling high-accuracy 3D reconstruction with a quasi-single-shot frame rate. The code is available at https://github.com/GeyouZhang/BSC.", + "bbox": [ + 259, + 328, + 743, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Phase Shifting Profilometry $\\cdot$ Dynamic 3D Scanning $\\cdot$ Motion Error Compensation", + "bbox": [ + 259, + 575, + 740, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 628, + 375, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, structured light (SL) 3D imaging has been penetrated into diverse fields, e.g., industrial inspection [20] and design [32], medical 3D reconstruction [3,8,13,17], virtual reality [28], and digitalization of cultural heritage [9]. Phase shifting profilometry (PSP) [23], Fourier transform profilometry, binary code [19], and Gray code [22] are famous SL technologies that are widely applied in industry and scientific researches. Among the SL technologies above, PSP is especially favored in high-precision 3D scanning due to its high accuracy, robustness, and pixel-wise property [33]. Many studies have excavated the huge potentiality of PSP in various challenging scenes, such as low signal-to-noise [5,7], multi-path [29,30], high reflectivity [18], and global illumination [1,2,6], et. al.", + "bbox": [ + 212, + 657, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, a fundamental assumption of PSP that the object should remain static is violated in dynamic measurement, making PSP susceptible to object", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "moving, resulting in ripple-like errors in the point clouds. We observed that the solutions of PSP in dynamic scenes are overlooked within the computer vision community. Concurrently, there are some existing methods in the optics metrology community to address this challenge, most of them can be categorized into three types: 1) motion-induced phase shifting estimation, 2) object tracking, and 3) automatic compensation.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion-induced phase shifting estimation: In traditional PSP, the phase shifting of each pattern is equally distributed or pre-known values within one period of fringe, which becomes disorderly due to object motion. Thus, researchers estimate the motion-induced phase shifting according to the implicit motion information in the motion-affected phase frame [26] or frames [10, 12, 21, 25], thereby restoring the unknown motion-affected phase shifting value of each pattern, and finally solve the corrected phase by least squared method [11]. However, these methods may not effectively compensate for errors when the motion information can't be precisely predicted.", + "bbox": [ + 212, + 239, + 789, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object tracking: By tracking the position of the object by placing markers [15] or using image feature matching algorithm [16], the motion error can be corrected by identifying the trajectory of the object [14-16]. However, a limitation of the object tracking based methods is that they are tailored for 2D motion perpendicular to the sight line. Moreover, the accuracy and reliability of feature tracking are crucial issues when faced with textureless objects.", + "bbox": [ + 212, + 378, + 787, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Automatic compensation: In automatic compensation methods, instead of computing motion-induced phase shifting value or tracking the object, an appropriate operator is selected to conduct on the whole captured image [24, 27] or phase frame [4], and further directly obtain the motion-error-free phase, in which the motion error will be automatically compensated.", + "bbox": [ + 212, + 472, + 787, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous work on motion compensation typically estimates or generates certain intermediate variables, then calculates the motion-error-free phase based on these variables, either directly or indirectly. For example, in phase shifting estimation methods, the intermediate variables are motion-induced phase shifting values [10, 12, 21, 25]; in object tracking methods, the intermediate variable is the rigid transform matrix [14-16]; in automatic compensation methods, the intermediate variables are the captured images after undergoing Hilbert transform [24] or differential operator [27]. Therefore, two questions naturally arise: Why not utilize the motion-affected phase itself to compensate for motion error without depending on any intermediate variables? Would this approach be more effective and flexible?", + "bbox": [ + 212, + 550, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, without depending on any extra intermediate variable, we develop a binomial self-compensation (BSC) method that only utilizes the motion-affected phase sequence itself, thereby effectively and flexibly eliminating the motion error in four-step PSP. To achieve this, first, we design a paraxial binocular structured light system for reconstructing 3D point clouds using a single set of high-frequency fringe patterns. This system intentionally sets a very short baseline between two cameras to achieve a narrow disparity range, facilitating the use of high frequency fringes, while ensuring the baseline between the main", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "camera and projector is long enough for adequate accuracy. Subsequently, we cyclically [31] project high-frequency patterns with $\\pi /2$ phase shifting, and directly compute a sequence of phase frames from each three or four successive captured images. Finally, by adding up a sequence of successive phase frames weighted by binomial coefficients, we achieve self-compensation of motion errors. Our mathematical model demonstrates that motion error exponentially diminishes as the binomial order $K$ increases.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our BSC not only inherits the pixel-wise advantages of PSP, but also is of high temporal resolution to achieve quasi-single-shot 3D imaging frame rate because it's frame-wise loopable. Furthermore, our method offers flexibility and ease of deployment, acting as an enhanced, motion-error-free version of the four-step PSP, thus providing a plug-and-play enhancement for dynamic 3D scanning scenarios.", + "bbox": [ + 212, + 252, + 784, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 368, + 387, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the introduction section, based on specific implementation details, we categorize the motion error compensation methods into three classes. From another perspective, we can also distinguish them based on their 3D reconstruction performance on temporal and spatial resolution. We examine the motion compensation methods from two perspectives: property of pixel-wise and frame-wise loopable. Pixel-wise methods compute phase and point clouds pixel by pixel, achieving spatial independence through this detailed granularity. Frame-wise loopable methods accomplish the same depth map frame rate as the camera acquisition frame rate, enabling a quasi-single-shot 3D reconstruction frame rate.", + "bbox": [ + 212, + 401, + 782, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pixel-wise: Several motion error compensation methods preserve the pixelwise precision of traditional PSP, which utilizing two sampling phase frames to iteratively compute motion phase shifting [12, 25], automatically compensate for errors [4], tracking the position of the object by placing markers [15], using image feature matching algorithm [16] and correct motion error by identifying the trajectory of the object [14-16]. These approaches build on the granularity of pixel-level adjustments. In contrast, the non-pixel-wise methods conduct particular operator on local [10, 26, 27] or global image [21, 24, 34], e.g., Fourier transform [21, 34], Hilbert transform [24], differential operator [27], mean filter [10], optimization algorithm [26], from which estimate the extra phase shifting induced by motion and compensate error [10, 21] or directly correct motion error [24, 26, 27, 34]. However, the pixel-wise merit of PSP is not preserved in these methods, theoretically diminishing their robustness in handling depth discontinuities. Concurrently, introducing extra non-pixel-wise operators also escalates the computational time required.", + "bbox": [ + 212, + 537, + 784, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frame-wise loopable: By projecting high-frequency fringes and reconstructing 3D point clouds using a cyclic strategy [31], some motion error compensation methods [4, 10, 24, 26, 27] accomplish 3D reconstruction with quasi-single-shot frame rate, thus called frame-wise loopable methods in this paper. The cyclic strategy is not suitable for all motion compensation methods, some", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning", + "bbox": [ + 259, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ddaeef3a1d8f7f61d6ee8667b6953f57330d4b708b4880b65cd22b8c67a6ecb2.jpg", + "image_caption": [ + "Fig. 1: Diagram of our paraxial binocular structured light system." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 781, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "methods need to project an indivisible group of patterns [21, 25, 34], such that the depth map frame rate is reduced severalfold.", + "bbox": [ + 212, + 377, + 784, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 System Setup and Frequency Selection", + "text_level": 1, + "bbox": [ + 215, + 431, + 627, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our paraxial binocular structured light system comprises a main camera, an auxiliary camera, and a projector, as depicted in Fig. 1. Together, the main and auxiliary cameras form a paraxial setup to provide reference depth for phase unwrapping. Simultaneously, the combination of the main camera and projector facilitates precise 3D reconstruction. Following system calibration, we perform stereo rectification on the captured images to align the epipolar lines in the auxiliary system parallelly.", + "bbox": [ + 212, + 462, + 784, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure the simplicity of equations and convey intuitive geometrical relations, the following analysis in this section is under the assumption that the optical centers of the projector, main camera, and auxiliary camera are coplanar. Note that we build the world coordinate system to coincide with the main camera coordinate system. In reality, the projector may not be strictly coplanar to the other two, but the conclusions of our analysis are still effective.", + "bbox": [ + 212, + 568, + 784, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The primary benefit of the binocular structured light (SL) system lies in its ability to perform stereo phase unwrapping (SPU) through geometric constraints, substantially decreasing the needed number of fringe patterns. To avoid phase unwrapping ambiguities, it should be ensured that only a few candidate points are within the depth range. As a result, traditional SPU approaches typically necessitate a short baseline between the main camera and the projector, which unfortunately compromises the accuracy of point clouds.", + "bbox": [ + 212, + 657, + 784, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, in this paper, we shift the constraint from the camera projector baseline to the camera-camera baseline, thus allowing flexibility in the camera-projector distance and enhancing point cloud precision. This camera-camera baseline constraint is called the uniqueness constraint, ensuring that the phase remains single-valued within the feasible stereo matching range on the", + "bbox": [ + 212, + 765, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "auxiliary camera, from which guarantees the uniqueness of the matching point candidate.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given that stereo rectification has been applied to the cameras, we can disregard the Y axis. Assuming the 3D surface near the measured point is continuous and orthogonal to the Z axis, determining the depth range allows us to calculate the maximum fringe frequency. Consider a virtual segment, represented by a black dashed line in Fig. 1, cut by the auxiliary camera's feasible region cone. The length of this virtual segment (black solid line in Fig. 1) can be calculated as:", + "bbox": [ + 212, + 176, + 784, + 279 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd = \\frac {\\left(Z _ {\\operatorname* {m a x}} - Z _ {\\operatorname* {m i n}}\\right) Z L}{Z _ {\\operatorname* {m a x}} Z _ {\\operatorname* {m i n}}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 279, + 784, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As long as the projection of the virtual segment onto the projector space is distributed within one fringe period, the uniqueness constraint is satisfied, i.e.,", + "bbox": [ + 214, + 311, + 782, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {f _ {x} ^ {p} d}{Z - Z ^ {p}} < \\frac {W ^ {p}}{f}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 348, + 784, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{x}^{p}$ is the focal length of the projector, $W^{p}$ is the resolution of the projector along $X$ axis, $f$ is the frequency of fringe pattern.", + "bbox": [ + 212, + 385, + 782, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, according to uniqueness constraint, we can obtain the limiting frequency of fringe as", + "bbox": [ + 212, + 416, + 782, + 445 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf < \\left(1 - \\frac {Z ^ {p}}{Z}\\right) \\frac {Z _ {\\operatorname* {m i n}} Z _ {\\operatorname* {m a x}} W ^ {p}}{\\left(Z _ {\\operatorname* {m a x}} - Z _ {\\operatorname* {m i n}}\\right) L f _ {x} ^ {p}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 453, + 784, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Actually, the uniqueness constraint is derived from an aggressive postulate that the 3D surface around the measured point is continuous and orthogonal with the Z axis, so the uniqueness constraint is a weak constraint. Thus, the camera-camera baseline should be as short as possible to ensure the uniqueness of phase value in the feasible range. In our prototype, we simply placed the main and auxiliary cameras tightly against each other. In the future, by using a beam splitter, the camera baseline can be further shortened, breaking through the physical size limitations of industrial cameras.", + "bbox": [ + 212, + 491, + 784, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, utilizing the proposed paraxial binocular SL system enables the achievement of unambiguous 3D reconstruction within the specified depth range, using only a single set of high-frequency fringe patterns. As a result, we can employ the cyclic projection strategy [31] for dynamic 3D scanning.", + "bbox": [ + 212, + 612, + 784, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Binomial Self-compensation for Motion Error", + "text_level": 1, + "bbox": [ + 214, + 691, + 689, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we cyclically project a group of high-frequency fringe patterns with a phase shift of $\\pi /2$ . The 3D point corresponding to a camera pixel $(x^{c},y^{c})$ is the intersection point of the sight line and the object surface. The intersection point moves along with the sight line on the circumstance that measures moving objects. Assuming the object is of low-frequency texture, the $i$ th captured frame can be expressed as", + "bbox": [ + 212, + 720, + 782, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} \\left(x ^ {c}, y ^ {c}\\right) = A ^ {c} + B ^ {c} \\cos \\left(\\phi_ {0} - i \\frac {\\pi}{2} + x _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 818, + 784, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning", + "bbox": [ + 258, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A^c$ is the background intensity, $B^c$ is the modulation, $x_{i}$ represents the unknown phase offset induced by the motion from the datum frame to $i$ -th frame, note that the datum frame can be arbitrarily chosen, and the true phase value at $i$ th frame is $\\phi_i = \\phi_0 - i\\frac{\\pi}{2}$ . We choose a $\\pi /2$ phase shifting as it only transforms $\\sin (\\cdot)$ and $\\cos (\\cdot)$ terms into each other without inducing extra phase offset.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We notice that $x_{i}$ and its high order difference exactly correspond to the concepts of displacement, velocity, and acceleration, et. al, in kinematics, and can be expressed as", + "bbox": [ + 212, + 226, + 787, + 271 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\Delta^ {(K)} x _ {i} = \\Delta^ {(K - 1)} x _ {i + 1} - \\Delta^ {(K - 1)} x _ {i} \\\\ \\Delta^ {(K)} x _ {i} = \\sum_ {k = 0} ^ {K} (- 1) ^ {k} \\binom {K} {k} x _ {i + K - k}, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 291, + 785, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we define $\\varDelta^{(0)}x_{i} = x_{i}$ . Eq. (5) is the key to motion error compensation, and we will use it later.", + "bbox": [ + 212, + 375, + 785, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Utilizing a cyclic projection strategy [31] for high speed 3D scanning, a three or four-step phase shifting algorithm is typically employed to calculate the motion-affected phase as:", + "bbox": [ + 212, + 410, + 785, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\phi} _ {i} ^ {3} = \\tan^ {- 1} \\left(\\frac {\\tilde {S} _ {i} ^ {3}}{\\tilde {C} _ {i} ^ {3}}\\right) \\text {a n d} \\tilde {\\phi} _ {i} ^ {4} = \\tan^ {- 1} \\left(\\frac {\\tilde {S} _ {i} ^ {4}}{\\tilde {C} _ {i} ^ {4}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 476, + 785, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 535, + 263, + 547 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} S _ {i} ^ {3} = 2 I _ {i + 1} - I _ {i} - I _ {i + 2} \\\\ C _ {i} ^ {3} = I _ {i} - I _ {i + 2} \\end{array} \\right. \\text {a n d} \\left\\{ \\begin{array}{l} S _ {i} ^ {4} = I _ {i + 1} - I _ {i + 3} \\\\ C _ {i} ^ {4} = I _ {i} - I _ {i + 2} \\end{array} \\right.. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 559, + 785, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Please note that the \"three-step phase shifting\" discussed in this paper differs slightly from the traditional approach, as the phase shifting is $\\pi /2$ instead of the traditional $2\\pi /3$ . We substitute Eq. (4) into Eq. (6) to evaluate the motion error, for a small phase-shift error $x_{i}$ , $\\sin (x_i)\\approx x_i$ and $\\cos (x_i)\\approx 1$ , thus we have", + "bbox": [ + 212, + 617, + 787, + 678 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\epsilon_ {3} (\\phi_ {i}) = \\tilde {\\phi} _ {i} ^ {3} - \\phi_ {i} \\\\ \\approx \\tan^ {- 1} \\left(\\frac {\\left[ 1 + 0 . 5 \\left(x _ {i} - x _ {i + 2}\\right) \\right] \\sin \\left(\\phi_ {i}\\right) + \\cos \\left(\\phi_ {i}\\right) x _ {i + 1}}{\\cos \\left(\\phi_ {i}\\right) - 0 . 5 \\left(x _ {i} + x _ {i + 2}\\right) \\sin \\left(\\phi_ {i}\\right)}\\right) - \\tan^ {- 1} \\left(\\frac {\\sin \\left(\\phi_ {i}\\right)}{\\cos \\left(\\phi_ {i}\\right)}\\right) \\\\ \\approx \\frac {x _ {i} + 2 x _ {i + 1} + x _ {i + 2}}{4} + \\frac {2 x _ {i + 1} - (x _ {i} + x _ {i + 2})}{4} \\cos (2 \\phi_ {i}) + \\frac {x _ {i} - x _ {i + 2}}{4} \\sin (2 \\phi_ {i}) \\\\ = \\delta_ {i} ^ {3} + \\frac {2 x _ {i + 1} - \\left(x _ {i} + x _ {i + 2}\\right)}{4} \\cos \\left(2 \\phi_ {i}\\right) + \\frac {x _ {i} - x _ {i + 2}}{4} \\sin \\left(2 \\phi_ {i}\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 710, + 797, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d076194a6fc92fab7ca3b8df5ce59eda3f9f9880d48d5159e77be8a462c2b2d.jpg", + "image_caption": [ + "Fig. 2: Binomial self-compensation for motion error." + ], + "image_footnote": [], + "bbox": [ + 233, + 141, + 769, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 215, + 454, + 248, + 467 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\epsilon_ {4} \\left(\\phi_ {i}\\right) = \\tilde {\\phi} _ {i} ^ {4} - \\phi_ {i} \\\\ \\begin{array}{l} \\approx \\tan^ {- 1} \\left(\\frac {2 \\sin (\\phi_ {i}) + (x _ {i + 1} + x _ {i + 3}) \\cos (\\phi_ {i})}{2 \\cos (\\phi_ {i}) - (x _ {i} + x _ {i + 2}) \\sin (\\phi_ {i})}\\right) - \\tan^ {- 1} \\left(\\frac {\\sin (\\phi_ {i})}{\\cos (\\phi_ {i})}\\right) \\\\ x _ {1} + x _ {1 + 2} + x _ {1 + 3} + x _ {1 + 4} = x _ {1} + x _ {1 + 5} = x _ {1 + 6} \\end{array} \\\\ \\approx \\frac {x _ {i} + x _ {i + 2} + x _ {i + 1} + x _ {i + 3}}{4} + \\frac {x _ {i + 1} - x _ {i} + x _ {i + 3} - x _ {i + 2}}{4} \\cos \\left(2 \\phi_ {i}\\right) \\\\ = \\delta_ {i} ^ {4} + \\frac {x _ {i + 1} - x _ {i} + x _ {i + 3} - x _ {i + 2}}{4} \\cos (2 \\phi_ {i}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 484, + 756, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "According to Eq. (8) and (9), we discover that the motion error comprises a DC component and a harmonic component with twice the frequency of the wrapped phase. The DC component, manifesting as the lag of the point clouds due to object motion, is independent of the phase value, and thus is irrelevant to the ripple-shaped error. Consequently, the crux of motion error compensation lies in eliminating the harmonic.", + "bbox": [ + 212, + 619, + 787, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analyzing two successive phase frames $\\tilde{\\phi}_i$ and $\\tilde{\\phi}_{i + 1}$ in a discrete time series, we discover that the coefficients of trigonometric terms exhibit opposite signs owing to a phase shift of $\\pi /2$ . Consequently, summing each pair of successive phase frames yields higher-order difference terms in the coefficients of harmonics. Repeating this process with the resultant phase frames, as depicted in Fig. 2, naturally evokes the concept of Yang Hui's triangle.", + "bbox": [ + 212, + 714, + 787, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let's consider calculating the weighted sum of $K + 1$ successive phase frames in a time series as shown in Fig. 2, using the $(K + 1)$ -th row of Yang Hui's", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning", + "bbox": [ + 258, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4c72e424128c7b68823640969f5dc6afc7a7219b594590c488a485ba99b19c6.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 495, + 304 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90eeccef51c401369eb3fa93641584a80a8f34fdc5ee5568a551428f6cce4bca.jpg", + "image_caption": [ + "(b)", + "Fig. 3: Motion error converges as $K$ increases: (a) three-step VS four-step phase shifting, and (b) visualization of motion error in four-step phase shifting, the mean absolute error of phase exponentially diminishes as $K$ increases." + ], + "image_footnote": [], + "bbox": [ + 496, + 143, + 787, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "triangle, also as known as $K$ -th order binomial coefficients", + "bbox": [ + 214, + 401, + 637, + 415 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\phi} _ {i} (K) = 2 ^ {- K} \\sum_ {k = 0} ^ {K} \\binom {K} {k} \\tilde {\\phi} _ {i + K - k} (0), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 425, + 784, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tilde{\\phi}_i(0)$ is computed by Eq. (6). We substitute Eq. (5) and (8) into Eq. (10) to have", + "bbox": [ + 212, + 479, + 782, + 508 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\phi} _ {i} ^ {3} (K) = \\bar {\\phi} _ {i} ^ {3} (K) - 2 ^ {- (K + 2)} \\Delta^ {(K + 2)} x _ {i + 2} \\cos (2 \\phi_ {i}) \\\\ + 2 ^ {- (K + 2)} \\left(\\Delta^ {(K + 1)} x _ {i + 2} + \\Delta^ {(K + 1)} x _ {i + 1}\\right) \\sin \\left(2 \\phi_ {i}\\right), \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 516, + 784, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 574, + 263, + 585 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\phi} _ {i} ^ {3} (K) = 2 ^ {- (K + 2)} \\sum_ {k = 0} ^ {K} \\left[ \\binom {K} {k} \\left(\\phi_ {i + K - k} + \\delta_ {i + K - k} ^ {3}\\right) \\right]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 315, + 584, + 784, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarly, for four-step phase shifting we have", + "bbox": [ + 214, + 631, + 545, + 646 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\phi} _ {i} ^ {4} (K) = \\bar {\\phi} _ {i} ^ {4} (K) + 2 ^ {- (K + 2)} \\left(\\Delta^ {(K + 1)} x _ {i} + \\Delta^ {(K + 1)} x _ {i + 2}\\right) \\cos \\left(2 \\phi_ {i}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 254, + 656, + 784, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 691, + 263, + 704 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\phi} _ {i} ^ {4} (K) = 2 ^ {- (K + 2)} \\sum_ {k = 0} ^ {K} \\left[ \\binom {K} {k} \\left(\\phi_ {i + K - k} + \\delta_ {i + K - k} ^ {4}\\right) \\right]. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 315, + 702, + 784, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on Eq. (11) and (13), it is clear that increasing $K$ produces two main effects: first, the factor $2^{-(K + 2)}$ exponentially decreases the harmonic amplitudes; second, higher-order difference terms (approach zero as $K$ increases) occur in the harmonic amplitudes. Both effects significantly diminish the motion error.", + "bbox": [ + 212, + 750, + 782, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We measure a flat plate moving towards the camera using our binomial self-compensation (BSC) for motion error, the results visualized in Fig. 3(a) show", + "bbox": [ + 212, + 810, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that the ripples induced by motion in the self-compensated phase $\\tilde{\\phi}_i^4 (K)$ reduces significantly with the increment of $K$ . $K = 0$ represents the raw phase that is affected by motion error. The residual error is caused by the nonlinearity of the binary defocus technique as shown in Fig. 3(b). At least four frames of images are needed in BSC, and using eight images achieves a trade-off between accuracy and efficiency in practice. Note that the binomial order $K$ can be flexibly chosen according to the accuracy and time efficiency requirements.", + "bbox": [ + 212, + 145, + 782, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In practice, we decompose Eq. (11) into a pyramid-like procedure of adding up the successive two frames layer by layer, as shown in Fig. 2. However, due to the ambiguity of the wrapped phase, phase order error occurs around the phase jumping area. To address this issue, we define the $\\oplus$ operator as", + "bbox": [ + 212, + 252, + 784, + 313 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {a} \\oplus \\phi_ {b} := \\left\\{ \\begin{array}{l} 0. 5 \\left(\\phi_ {a} + \\phi_ {b}\\right), \\left| \\phi_ {a} - \\phi_ {b} \\right| > \\pi \\\\ 0. 5 \\left(\\phi_ {a} + \\phi_ {b}\\right) + \\pi , \\left| \\phi_ {a} - \\phi_ {b} \\right| < \\pi \\end{array} . \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 321, + 784, + 353 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This operation automatically compensates for the phase order error as long as the phase difference is greater than $\\pi$ . The motion-error-free phase $\\tilde{\\phi}_i(K)$ can be obtained by employing our binomial self-compensation (BSC), which is summarized as Algorithm 1, where \"mod\" represents modular operation. Note that we have corrected the constant phase shifting before adding up two phase frames $\\tilde{\\phi}_i(0)$ and $\\tilde{\\phi}_{i + 1}(0)$ . The MATLAB code for implementing the algorithm are given in the github repository.", + "bbox": [ + 214, + 361, + 787, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Stereo Phase Unwrapping", + "text_level": 1, + "bbox": [ + 215, + 487, + 503, + 506 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the previous section, we obtained the motion-error-free high-frequency phase by conducting BSC. Subsequently, we employ a sum of absolute difference algorithm on the wrapped phase to retrieve the correspondence between the main and auxiliary cameras within the disparity range obtained by pre-set depth range. Further, the projector space coordinates $(x^p,y^p)$ can be computed by directly re-projecting the binocular 3D coordinate onto the projector space. Finally, we can have the phase order and compute the unwrapped phase.", + "bbox": [ + 212, + 517, + 782, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Experimental Evaluations", + "text_level": 1, + "bbox": [ + 215, + 643, + 501, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The proposed algorithm is implemented on an Intel(R) Core(TM) i7-13790F @ 2.10 GHz with 32 GB RAM and written in MATLAB. Our experimental system (Fig. 4) consists of two AVT 1800U-120c cameras that have $1280 \\times 960$ resolution, equipped with $8\\mathrm{mm}$ Computer lenses, and a TI DLP4500 projector with a resolution of $912 \\times 1140$ . We cropped the cameras' field of view to $640 \\times 480$ to have a 90 fps image acquisition speed. The two cameras are synchronized by the trigger signal from the projector. In order to comprehensively demonstrate the superiority of our BSC, we conducted four groups of experiments on dynamic objects for evaluating: 1) absolute accuracy of 3D reconstruction, 2) temporal resolution of 3D imaging, 3) robustness on depth discontinuous scenes, and 4) generality on different dynamic objects.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning", + "bbox": [ + 258, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: BSC for motion error" + ], + "code_body": "Input: Captured image sequence: $\\{I_0,\\dots ,I_{T - 1}\\}$ Binomial order: $K$ Phase-shifting step: $N = 3$ or 4. Output: Motion-error-free phase sequence: $\\{\\bar{\\phi}_0(K),\\dots,\\bar{\\phi}_{T - K - N + 1}(K)\\}$ for $t\\gets 0$ to $T - N$ do $\\tilde{\\phi}_t(0)\\gets \\left\\{ \\begin{array}{ll}\\mathrm{mod}\\left[\\tan^{-1}\\left(\\frac{2I_{t + 1} - I_t - I_{t + 2}}{I_t - I_{t + 2}}\\right) + t\\frac{\\pi}{2},2\\pi \\right],N = 3\\\\ \\mathrm{mod}\\left[\\tan^{-1}\\left(\\frac{I_{t + 1} - I_{t + 3}}{I_t - I_{t + 2}}\\right) + t\\frac{\\pi}{2},2\\pi \\right],N = 4 \\end{array} \\right.$ end for $k\\gets 0$ to $K - 1$ do for $i\\gets 0$ to $T - k - N$ do $\\tilde{\\phi}_i(k + 1)\\gets \\tilde{\\phi}_i(k)\\oplus \\tilde{\\phi}_{i + 1}(k)$ end", + "bbox": [ + 225, + 166, + 660, + 393 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d77049b1c21816005f3775b76d5e4c5836b01ec44720bb4cfe3a798ed4ff3bd9.jpg", + "image_caption": [ + "Fig. 4: Experimental setup." + ], + "image_footnote": [], + "bbox": [ + 302, + 412, + 700, + 503 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.1 Absolute Accuracy of 3D Reconstruction", + "text_level": 1, + "bbox": [ + 212, + 558, + 599, + 574 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We measured a periodically waving plate at a distance of $500~\\mathrm{mm}$ by using our BSC (8 frames) and five representative methods for comparison: 1) traditional four-step PSP (4 frames), 2) Hilbert transform compensation (HTC, 4 frames) [24], 3) micro Fourier transform profilometry ( $\\mu$ -FTP, 4 frames) [34], 4) phase frame difference method (PFD, 8 frames) [10], 5) phase frame sum method (PFS, 5 frmaes) [4]. The fringe wavelength was set to 24 pixels for the cyclic $\\pi/2$ phase shifting patterns and was set to [22,24,26] for $\\mu$ -FTP. The plate moves within a distance range of [400,500] mm and a speed range of $[-150,150]$ mm/s, and our measurement lasted for 800 frames (about 8.9 seconds). We select three $160\\times 120$ rectangular windows in each depth map and fit the quadric surface as the ground truth, then we compute the root mean squared error (RMSE) within the selected regions. The trend of RMSE changing with motion speed is depicted in Fig. 5. The results show that our method exhibits the most satisfactory accuracy among existing methods when dealing with motion error.", + "bbox": [ + 212, + 582, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Then we selected a moment when the speed of the plate was $-88~\\mathrm{mm / s}$ , and depict the reconstructed point clouds in Fig. 6, our BSC effectively reduces the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b71efe0e4f51af127cb00e16143ffa3b98a23ad5f5c79421228a8de4277b823a.jpg", + "image_caption": [ + "Fig. 5: Measurement error VS motion speed of a periodically waving plate by traditional four-step phase shifting, HTC [24], $\\mu$ -FTP [34], PFD [10], PFS [4], and our BSC, the value in parentheses represents the mean error. Note the y-axis is in logarithmic scale." + ], + "image_footnote": [], + "bbox": [ + 300, + 143, + 702, + 359 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/66bda019f3fc400695d484b3f50cd4c80b4791adfd49caa711b8e5550fac6fe9.jpg", + "image_caption": [ + "(a)", + "(d)" + ], + "image_footnote": [], + "bbox": [ + 230, + 465, + 400, + 573 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/347d1287a6cf663eed11807d92a1ed1956249eb8a705ed30227b152ed613c02d.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 403, + 465, + 571, + 571 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/39bc771459a7d180cbc8ddbb2c3d8982a192b88b25c84e8508e9e8e9c99c87ad.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 576, + 453, + 774, + 573 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f3457610b00de9c11b86ec3f0c6db120ec8d187411961090490735753595db4b.jpg", + "image_caption": [ + "Fig. 6: Measurement results of a moving flat plane by (a) traditional four-step phase shifting, (b) HTC [24], (c) $\\mu$ -FTP [34], (d) PFD [10], (e) PFS [4], and (f) our BSC." + ], + "image_footnote": [], + "bbox": [ + 230, + 599, + 398, + 708 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5703c01ad67352a26821cf11148de1204f7ca35a91610687ccaf0b0693a9d015.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 403, + 598, + 571, + 708 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e51ec7ccd87c780c87d91503f748a423613f8174f01c401409fcf98fff4bec73.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 576, + 598, + 772, + 708 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "motion error from $324.2\\mu \\mathrm{m}$ to $54.78\\mu \\mathrm{m}$ compared with traditional four-step phase shifting. By counting the histogram of errors, it can be observed that the error distribution of our BSC visually presents a normal distribution, which means Gaussian noise (induced by unstable ambient light, camera/projector", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning", + "bbox": [ + 259, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6107b12975720eca2ae77f4acf1463066f32ce840464695bbff3b353d6cc941e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 172, + 403, + 279 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/38a2c2486c55c25c705cceefbd1d091de49c0babdfa1d5ae950a122651feb0d4.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 408, + 157, + 586, + 247 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d620481d207dcc81790361f60b4c291e2d5b57f4576048d206146b906cee0f65.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 410, + 250, + 589, + 304 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/48a0943cc52c726e001be28e78d7ff106498059226715656547e9db5beb52298.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 602, + 157, + 774, + 247 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/64ee32d814590d9392324ae4f104a47e7e53f3bf874fa82e6ea6d2b56e72cdfe.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 594, + 250, + 774, + 303 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2a576bdd7eb9d230f71017763154b3b94a1485661f7f790948fc94183e59f01b.jpg", + "image_caption": [ + "Fig. 7: Depth discontinuous scene (top left) and motion area labeled in white (bottom left). Reconstruction results and error at edges: (a) traditional four-step phase shifting, (b) HTC [24], (c) $\\mu$ -FTP [34], and (d) our BSC." + ], + "image_footnote": [], + "bbox": [ + 222, + 325, + 403, + 434 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/182ac23580f8950548a31da9dac2662a1e105214e08749d4f499af73bc49b8bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 325, + 586, + 404 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f033ad971d272589c2ff7d5783fd7c8fedee85824efad92d3a459d05e02e947e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 405, + 589, + 458 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6c769052cd8291da027c0237b67585ce5e79e780b3a020f771daa8e4e5fdeb62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 325, + 774, + 404 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/694599452683c4cc3f9eb4e44b4b18338aff1fd0796461c36f92ca1b09dbf66c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 406, + 774, + 458 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "flicker, quantization error, and sensor noise of the camera and projector) is the main component of error. In other words, the motion error is effectively suppressed by BSC. Meanwhile, the error distribution of other methods deviates from the normal distribution, indicating the existence of residual errors.", + "bbox": [ + 212, + 547, + 787, + 607 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.2 Robustness on Depth Discontinuous Scenes", + "text_level": 1, + "bbox": [ + 214, + 638, + 620, + 655 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "PSP's pixel-wise characteristic is one of its most crucial merits. Our BSC maintains this property, thereby exhibiting robustness to depth discontinuous scenes. We set up a scene with three static gypsum geometries and one moving plate. Then we employed our BSC, $\\mu$ -FTP, and HTC. The results are shown in Fig. 7. It can be seen that the two non-pixel-wise methods, i.e., $\\mu$ -FTP and HTC, exhibit visible deformation to the naked eye in the point cloud around the depth jumping areas. We selected three rectangular windows and quantitatively evaluate the RMSE, the results show that the RMSE of non-pixel-wise methods is significantly greater than our BSC. We can conclude that our BSC with pixel-wise property is more robust compared with non-pixel-wise methods when dealing with depth discontinuous scenes.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/479f16360ff1712f2274a892ee058779fb4532ad46bc4f665c5d30c450355394.jpg", + "image_caption": [ + "Fig. 8: The measurement result of a wooden butterfly with wings flapping, during $\\mathrm{T} = 0\\mathrm{ms}\\sim 88\\mathrm{ms}$ , Row 1 and 2: our BSC; Row 3 and 4: $\\mu$ -FTP [34]." + ], + "image_footnote": [], + "bbox": [ + 222, + 141, + 782, + 330 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.3 Temporal Resolution of 3D Imaging", + "text_level": 1, + "bbox": [ + 214, + 397, + 558, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our BSC employs cyclic $\\pi /2$ phase-shifting fringe patterns, any $K + N$ successive captured images can be used for reconstructing one frame of 3D point clouds. Therefore, our resulting frame rate of 3D point clouds is the same as the camera frame rate, thereby achieving a high temporal resolution of the measurement. That is the frame-wise loopable property we mentioned in the introduction. We successively conducted BSC and $\\mu$ -FTP to measure a wooden butterfly with wings flapping cyclically, and then manually aligned the timeline, the results are shown in Fig. 8. We noticed that our method can achieve 3D reconstruction with a frame rate of 90 fps (the same as the camera frame rate), while $\\mu$ -FTP reconstructing at 30 fps. Thus, our method is capable of distinguishing tinier motions.", + "bbox": [ + 212, + 421, + 787, + 587 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.4 Generality on Different Dynamic Objects", + "text_level": 1, + "bbox": [ + 214, + 609, + 602, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We captured 400 image frames (about 4 seconds) employing our pipeline on four different dynamic objects, and the results are shown in Fig. 9. The complete motion process video can be found in the supplementary materials. We observed that BSC can effectively eliminate the severe ripple errors caused by object motion.", + "bbox": [ + 212, + 633, + 787, + 709 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7 Limitations and Future Outlook", + "text_level": 1, + "bbox": [ + 214, + 732, + 563, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We presented a pixel-wise and frame-wise loopable binomial self-compensation algorithm to effectively and flexibly eliminate motion error in four-step PSP, without any intermediate variables assisted.", + "bbox": [ + 212, + 763, + 785, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "High-frequency textures: our BSC assumes low-frequency textures like most motion compensation algorithms, it is robust against moderate contrast", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning", + "bbox": [ + 258, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/13fdbcdc871228cc97262fd25b98c3e34fffe36bfba2435a90a42f7b8bde0d93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 150, + 359, + 243 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6a94379d7f65a90a1da9a8799401de7a68166c2a21d6fed9b24929ea1a67345d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 151, + 490, + 243 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4cd63007085a631bceb3a584acfba6966e663ad29fa08ef6488dea823dbc8505.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 151, + 620, + 243 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c6d9be6c48f0351dd74b52b6a0df6a67bcb9d255a895cbadc66bbc983125e150.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 151, + 779, + 244 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3a7391dfee698618eafd0c9a667535e24bdd4e36a9e7fbfee1100f7d7c1bfa12.jpg", + "image_caption": [ + "Fig. 9: The measurement results of four different dynamic objects. Column 1: wooden butterfly model with wings flapping, Column 2: moving statue, Column 3: waving hand, Column 4: hand showing scissors, stone, and cloth gestures. Top to bottom: our BSC and traditional four-step PSP." + ], + "image_footnote": [], + "bbox": [ + 232, + 258, + 357, + 344 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c36e7274a64d8b134b877edad16417321dea969a791b931128c159ff5ff30839.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 257, + 490, + 344 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2e501875e5e20f69ab84ab317ea8317095a0e0b2a80eee738284e5c2cf658c04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 256, + 620, + 344 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/55f779acca8e3d1cfe8f205ab58418b4f1d16780e80c821c9e675238ba4dadcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 258, + 779, + 344 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/655a29d17c5a91ddac4cb54e83f623dd6fccc98be6d55605163145e47d874d40.jpg", + "image_caption": [ + "Fig. 10: 3D error VS inter-frame phase-shifting error $|\\Delta x_{i}|$ ." + ], + "image_footnote": [], + "bbox": [ + 357, + 445, + 647, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "textures. However, when dealing with objects that have high contrast textures, our method induces artifacts similar to those seen with typical phase-shifting methods.", + "bbox": [ + 212, + 637, + 784, + 681 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Phase unwrapping error: we adopt a common block matching algorithm for stereo correspondence, inducing obvious phase unwrapping error when measuring complex surfaces. In the future, we will conduct a hierarchical adjustment for phase outliers to optimize the accuracy of the point cloud.", + "bbox": [ + 212, + 693, + 784, + 753 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Applicable speed range: Our BSC is predicated on the assumption that the phase-shift error $x_{i}$ is small. Consequently, as shown in Fig. 10, the residual motion error arises as $|\\Delta x_{i}|$ increases. Properly reducing the fringe frequency could decrease $|\\Delta x_{i}|$ , alleviating the performance loss of BSC for high-speed objects.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achar, S., Narasimhan, S.G.: Multi focus structured light for recovering scene shape and global illumination. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 205-219. Springer (2014)", + "2. Chen, T., Seidel, H.P., Lensch, H.P.: Modulated phase-shifting for 3d scanning. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1-8. IEEE (2008)", + "3. Geng, J.: Structured-light 3d surface imaging: a tutorial. Advances in optics and photonics 3(2), 128-160 (2011)", + "4. Guo, W., Wu, Z., Zhang, Q., Wang, Y.: Real-time motion-induced error compensation for 4-step phase-shifting profilometry. Optics Express 29(15), 23822-23834 (2021)", + "5. Gupta, M., Nakhate, N.: A geometric perspective on structured light coding. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 87-102 (2018)", + "6. Gupta, M., Tian, Y., Narasimhan, S.G., Zhang, L.: A combined theory of defocused illumination and global light transport. International Journal of Computer Vision 98, 146-167 (2012)", + "7. Gupta, M., Yin, Q., Nayar, S.K.: Structured light in sunlight. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 545-552 (2013)", + "8. Hirose, K., Kamei, H., Sugiyama, T., Kurosaka, Y.: $200 \\times 200 \\mu \\mathrm{m}$ 2 structured light source. Optics Express 28(25), 37307-37321 (2020)", + "9. Honrado, C.P., Larrabee Jr, W.F.: Update in three-dimensional imaging in facial plastic surgery. Current opinion in otolaryngology & head and neck surgery 12(4), 327-331 (2004)", + "0. Liu, X., Tao, T., Wan, Y., Kofman, J.: Real-time motion-induced-error compensation in 3d surface-shape measurement. Optics Express 27(18), 25265-25279 (2019)", + "1. Liu, Y., Pears, N., Rosin, P.L., Huber, P.: 3D imaging, analysis and applications. Springer (2012)", + "2. Liu, Z., Zibley, P.C., Zhang, S.: Motion-induced error compensation for phase shifting profilometry. Optics Express 26(10), 12632-12637 (2018)", + "3. Logozzo, S., Zanetti, E.M., Franceschini, G., Kilpelä, A., Mäkynen, A.: Recent advances in dental optics-part i: 3d intraoral scanners for restorative dentistry. Optics and Lasers in Engineering 54, 203-221 (2014)", + "4. Lu, L., Ding, Y., Luan, Y., Yin, Y., Liu, Q., Xi, J.: Automated approach for the surface profile measurement of moving objects based on psp. Optics Express 25(25), 32120-32131 (2017)", + "5. Lu, L., Xi, J., Yu, Y., Guo, Q.: New approach to improve the accuracy of 3-d shape measurement of moving object using phase shifting profilometry. Optics Express 21(25), 30610-30622 (2013)", + "6. Lu, L., Yin, Y., Su, Z., Ren, X., Luan, Y., Xi, J.: General model for phase shifting profilometry with an object in motion. Applied Optics 57(36), 10364-10369 (2018)", + "7. Lv, S., Kemao, Q.: Modeling the measurement precision of fringe projection profilometry. Light: Science & Applications 12(1), 257 (2023)", + "8. Nayar, S.K., Gupta, M.: Diffuse structured light. In: IEEE International Conference on Computational Photography (ICCP). pp. 1-11. IEEE (2012)", + "9. Posdamer, J.L., Altschuler, M.D.: Surface measurement by space-encoded projected beam systems. Computer Graphics and Image Processing 18(1), 1-17 (1982)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning", + "bbox": [ + 259, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Qian, J., Feng, S., Xu, M., Tao, T., Shang, Y., Chen, Q., Zuo, C.: High-resolution real-time 360 3d surface defect inspection with fringe projection profilometry. Optics and Lasers in Engineering 137, 106382 (2021)", + "21. Qian, J., Tao, T., Feng, S., Chen, Q., Zuo, C.: Motion-artifact-free dynamic 3d shape measurement with hybrid fourier-transform phase-shifting profilometry. Optics Express 27(3), 2713-2731 (2019)", + "22. Sansoni, G., Corini, S., Lazzari, S., Rodella, R., Docchio, F.: Three-dimensional imaging based on gray-code light projection: characterization of the measuring algorithm and development of a measuring system for industrial applications. Applied Optics 36(19), 4463-4472 (1997)", + "23. Srinivasan, V., Liu, H.C., Halioua, M.: Automated phase-measuring profilometry of 3-d diffuse objects. Applied optics 23(18), 3105-3108 (1984)", + "24. Wang, Y., Liu, Z., Jiang, C., Zhang, S.: Motion induced phase error reduction using a hilbert transform. Optics Express 26(26), 34224-34235 (2018)", + "25. Wang, Y., Suresh, V., Li, B.: Motion-induced error reduction for binary defocusing profilometry via additional temporal sampling. Optics Express 27(17), 23948-23958 (2019)", + "26. Weise, T., Leibe, B., Van Gool, L.: Fast 3d scanning with automatic motion compensation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1-8. IEEE (2007)", + "27. Wu, G., Yang, T., Liu, F., Qian, K.: Suppressing motion-induced phase error by using equal-step phase-shifting algorithms in fringe projection profilometry. Optics Express 30(11), 17980-17998 (2022)", + "28. Ye, Y., Chang, H., Song, Z., Zhao, J.: Accurate infrared structured light sensing system for dynamic 3d acquisition. Applied Optics 59(17), E80-E88 (2020)", + "29. Zhang, Y., Lau, D., Wipf, D.: Sparse multi-path corrections in fringe projection profilometry. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13344-13353 (2021)", + "30. Zhang, Y., Lau, D.L., Yu, Y.: Causes and corrections for bimodal multi-path scanning with structured light. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4431-4439 (2019)", + "31. Zhong, K., Li, Z., Shi, Y., Wang, C., Lei, Y.: Fast phase measurement profilometry for arbitrary shape objects without phase unwrapping. Optics and Lasers in Engineering 51(11), 1213-1222 (2013)", + "32. Zhu, J., Yang, F., Hu, J., Zhou, P.: High dynamic reflection surface 3d reconstruction with sharing phase demodulation mechanism and multi-indicators guided phase domain fusion. Optics Express 31(15), 25318-25338 (2023)", + "33. Zuo, C., Feng, S., Huang, L., Tao, T., Yin, W., Chen, Q.: Phase shifting algorithms for fringe projection profilometry: A review. Optics and lasers in engineering 109, 23-59 (2018)", + "34. Zuo, C., Tao, T., Feng, S., Huang, L., Asundi, A., Chen, Q.: Micro fourier transform profilometry (μftp): 3d shape measurement at 10,000 frames per second. Optics and Lasers in Engineering 102, 70-91 (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 742 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_model.json b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8136ace84c4b8102cf660749aabe96b07ef4df95 --- /dev/null +++ b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_model.json @@ -0,0 +1,2410 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.212, + 0.664, + 0.228 + ], + "angle": 0, + "content": "Geyou Zhang\\(^{1}\\), Ce Zhu\\(^{1,*}\\), and Kai Liu\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.239, + 0.7, + 0.268 + ], + "angle": 0, + "content": "1 University of Electronic Science and Technology of China, geyouzhang@foxmail.com, eczhu@uestc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.269, + 0.573, + 0.282 + ], + "angle": 0, + "content": "2 Sichuan University" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.284, + 0.568, + 0.295 + ], + "angle": 0, + "content": "kailiu@scu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.329, + 0.744, + 0.565 + ], + "angle": 0, + "content": "Abstract. Phase shifting profilometry (PSP) is favored in high-precision 3D scanning due to its high accuracy, robustness, and pixel-wise property. However, a fundamental assumption of PSP that the object should remain static is violated in dynamic measurement, making PSP susceptible to object moving, resulting in ripple-like errors in the point clouds. We propose a pixel-wise and frame-wise loopable binomial self-compensation (BSC) algorithm to effectively and flexibly eliminate motion error in the four-step PSP. Our mathematical model demonstrates that by summing successive motion-affected phase frames weighted by binomial coefficients, motion error exponentially diminishes as the binomial order increases, accomplishing automatic error compensation through the motion-affected phase sequence, without the assistance of any intermediate variable. Extensive experiments show that our BSC outperforms the existing methods in reducing motion error, while achieving a depth map frame rate equal to the camera's acquisition rate (90 fps), enabling high-accuracy 3D reconstruction with a quasi-single-shot frame rate. The code is available at https://github.com/GeyouZhang/BSC." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.577, + 0.741, + 0.606 + ], + "angle": 0, + "content": "Keywords: Phase Shifting Profilometry \\(\\cdot\\) Dynamic 3D Scanning \\(\\cdot\\) Motion Error Compensation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.377, + 0.646 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.81 + ], + "angle": 0, + "content": "In recent years, structured light (SL) 3D imaging has been penetrated into diverse fields, e.g., industrial inspection [20] and design [32], medical 3D reconstruction [3,8,13,17], virtual reality [28], and digitalization of cultural heritage [9]. Phase shifting profilometry (PSP) [23], Fourier transform profilometry, binary code [19], and Gray code [22] are famous SL technologies that are widely applied in industry and scientific researches. Among the SL technologies above, PSP is especially favored in high-precision 3D scanning due to its high accuracy, robustness, and pixel-wise property [33]. Many studies have excavated the huge potentiality of PSP in various challenging scenes, such as low signal-to-noise [5,7], multi-path [29,30], high reflectivity [18], and global illumination [1,2,6], et. al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "However, a fundamental assumption of PSP that the object should remain static is violated in dynamic measurement, making PSP susceptible to object" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "moving, resulting in ripple-like errors in the point clouds. We observed that the solutions of PSP in dynamic scenes are overlooked within the computer vision community. Concurrently, there are some existing methods in the optics metrology community to address this challenge, most of them can be categorized into three types: 1) motion-induced phase shifting estimation, 2) object tracking, and 3) automatic compensation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.79, + 0.377 + ], + "angle": 0, + "content": "Motion-induced phase shifting estimation: In traditional PSP, the phase shifting of each pattern is equally distributed or pre-known values within one period of fringe, which becomes disorderly due to object motion. Thus, researchers estimate the motion-induced phase shifting according to the implicit motion information in the motion-affected phase frame [26] or frames [10, 12, 21, 25], thereby restoring the unknown motion-affected phase shifting value of each pattern, and finally solve the corrected phase by least squared method [11]. However, these methods may not effectively compensate for errors when the motion information can't be precisely predicted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.47 + ], + "angle": 0, + "content": "Object tracking: By tracking the position of the object by placing markers [15] or using image feature matching algorithm [16], the motion error can be corrected by identifying the trajectory of the object [14-16]. However, a limitation of the object tracking based methods is that they are tailored for 2D motion perpendicular to the sight line. Moreover, the accuracy and reliability of feature tracking are crucial issues when faced with textureless objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.548 + ], + "angle": 0, + "content": "Automatic compensation: In automatic compensation methods, instead of computing motion-induced phase shifting value or tracking the object, an appropriate operator is selected to conduct on the whole captured image [24, 27] or phase frame [4], and further directly obtain the motion-error-free phase, in which the motion error will be automatically compensated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Previous work on motion compensation typically estimates or generates certain intermediate variables, then calculates the motion-error-free phase based on these variables, either directly or indirectly. For example, in phase shifting estimation methods, the intermediate variables are motion-induced phase shifting values [10, 12, 21, 25]; in object tracking methods, the intermediate variable is the rigid transform matrix [14-16]; in automatic compensation methods, the intermediate variables are the captured images after undergoing Hilbert transform [24] or differential operator [27]. Therefore, two questions naturally arise: Why not utilize the motion-affected phase itself to compensate for motion error without depending on any intermediate variables? Would this approach be more effective and flexible?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, without depending on any extra intermediate variable, we develop a binomial self-compensation (BSC) method that only utilizes the motion-affected phase sequence itself, thereby effectively and flexibly eliminating the motion error in four-step PSP. To achieve this, first, we design a paraxial binocular structured light system for reconstructing 3D point clouds using a single set of high-frequency fringe patterns. This system intentionally sets a very short baseline between two cameras to achieve a narrow disparity range, facilitating the use of high frequency fringes, while ensuring the baseline between the main" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.26, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "camera and projector is long enough for adequate accuracy. Subsequently, we cyclically [31] project high-frequency patterns with \\(\\pi /2\\) phase shifting, and directly compute a sequence of phase frames from each three or four successive captured images. Finally, by adding up a sequence of successive phase frames weighted by binomial coefficients, we achieve self-compensation of motion errors. Our mathematical model demonstrates that motion error exponentially diminishes as the binomial order \\(K\\) increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Our BSC not only inherits the pixel-wise advantages of PSP, but also is of high temporal resolution to achieve quasi-single-shot 3D imaging frame rate because it's frame-wise loopable. Furthermore, our method offers flexibility and ease of deployment, acting as an enhanced, motion-error-free version of the four-step PSP, thus providing a plug-and-play enhancement for dynamic 3D scanning scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.369, + 0.388, + 0.385 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.784, + 0.537 + ], + "angle": 0, + "content": "In the introduction section, based on specific implementation details, we categorize the motion error compensation methods into three classes. From another perspective, we can also distinguish them based on their 3D reconstruction performance on temporal and spatial resolution. We examine the motion compensation methods from two perspectives: property of pixel-wise and frame-wise loopable. Pixel-wise methods compute phase and point clouds pixel by pixel, achieving spatial independence through this detailed granularity. Frame-wise loopable methods accomplish the same depth map frame rate as the camera acquisition frame rate, enabling a quasi-single-shot 3D reconstruction frame rate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Pixel-wise: Several motion error compensation methods preserve the pixelwise precision of traditional PSP, which utilizing two sampling phase frames to iteratively compute motion phase shifting [12, 25], automatically compensate for errors [4], tracking the position of the object by placing markers [15], using image feature matching algorithm [16] and correct motion error by identifying the trajectory of the object [14-16]. These approaches build on the granularity of pixel-level adjustments. In contrast, the non-pixel-wise methods conduct particular operator on local [10, 26, 27] or global image [21, 24, 34], e.g., Fourier transform [21, 34], Hilbert transform [24], differential operator [27], mean filter [10], optimization algorithm [26], from which estimate the extra phase shifting induced by motion and compensate error [10, 21] or directly correct motion error [24, 26, 27, 34]. However, the pixel-wise merit of PSP is not preserved in these methods, theoretically diminishing their robustness in handling depth discontinuities. Concurrently, introducing extra non-pixel-wise operators also escalates the computational time required." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Frame-wise loopable: By projecting high-frequency fringes and reconstructing 3D point clouds using a cyclic strategy [31], some motion error compensation methods [4, 10, 24, 26, 27] accomplish 3D reconstruction with quasi-single-shot frame rate, thus called frame-wise loopable methods in this paper. The cyclic strategy is not suitable for all motion compensation methods, some" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.782, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.334, + 0.723, + 0.348 + ], + "angle": 0, + "content": "Fig. 1: Diagram of our paraxial binocular structured light system." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.785, + 0.409 + ], + "angle": 0, + "content": "methods need to project an indivisible group of patterns [21, 25, 34], such that the depth map frame rate is reduced severalfold." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.432, + 0.628, + 0.449 + ], + "angle": 0, + "content": "3 System Setup and Frequency Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.569 + ], + "angle": 0, + "content": "Our paraxial binocular structured light system comprises a main camera, an auxiliary camera, and a projector, as depicted in Fig. 1. Together, the main and auxiliary cameras form a paraxial setup to provide reference depth for phase unwrapping. Simultaneously, the combination of the main camera and projector facilitates precise 3D reconstruction. Following system calibration, we perform stereo rectification on the captured images to align the epipolar lines in the auxiliary system parallelly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.658 + ], + "angle": 0, + "content": "To ensure the simplicity of equations and convey intuitive geometrical relations, the following analysis in this section is under the assumption that the optical centers of the projector, main camera, and auxiliary camera are coplanar. Note that we build the world coordinate system to coincide with the main camera coordinate system. In reality, the projector may not be strictly coplanar to the other two, but the conclusions of our analysis are still effective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.764 + ], + "angle": 0, + "content": "The primary benefit of the binocular structured light (SL) system lies in its ability to perform stereo phase unwrapping (SPU) through geometric constraints, substantially decreasing the needed number of fringe patterns. To avoid phase unwrapping ambiguities, it should be ensured that only a few candidate points are within the depth range. As a result, traditional SPU approaches typically necessitate a short baseline between the main camera and the projector, which unfortunately compromises the accuracy of point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To address this, in this paper, we shift the constraint from the camera projector baseline to the camera-camera baseline, thus allowing flexibility in the camera-projector distance and enhancing point cloud precision. This camera-camera baseline constraint is called the uniqueness constraint, ensuring that the phase remains single-valued within the feasible stereo matching range on the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "auxiliary camera, from which guarantees the uniqueness of the matching point candidate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.28 + ], + "angle": 0, + "content": "Given that stereo rectification has been applied to the cameras, we can disregard the Y axis. Assuming the 3D surface near the measured point is continuous and orthogonal to the Z axis, determining the depth range allows us to calculate the maximum fringe frequency. Consider a virtual segment, represented by a black dashed line in Fig. 1, cut by the auxiliary camera's feasible region cone. The length of this virtual segment (black solid line in Fig. 1) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.28, + 0.785, + 0.311 + ], + "angle": 0, + "content": "\\[\nd = \\frac {\\left(Z _ {\\operatorname* {m a x}} - Z _ {\\operatorname* {m i n}}\\right) Z L}{Z _ {\\operatorname* {m a x}} Z _ {\\operatorname* {m i n}}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.313, + 0.784, + 0.343 + ], + "angle": 0, + "content": "As long as the projection of the virtual segment onto the projector space is distributed within one fringe period, the uniqueness constraint is satisfied, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.349, + 0.785, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\frac {f _ {x} ^ {p} d}{Z - Z ^ {p}} < \\frac {W ^ {p}}{f}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.784, + 0.416 + ], + "angle": 0, + "content": "where \\( f_{x}^{p} \\) is the focal length of the projector, \\( W^{p} \\) is the resolution of the projector along \\( X \\) axis, \\( f \\) is the frequency of fringe pattern." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.784, + 0.446 + ], + "angle": 0, + "content": "Finally, according to uniqueness constraint, we can obtain the limiting frequency of fringe as" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.454, + 0.785, + 0.486 + ], + "angle": 0, + "content": "\\[\nf < \\left(1 - \\frac {Z ^ {p}}{Z}\\right) \\frac {Z _ {\\operatorname* {m i n}} Z _ {\\operatorname* {m a x}} W ^ {p}}{\\left(Z _ {\\operatorname* {m a x}} - Z _ {\\operatorname* {m i n}}\\right) L f _ {x} ^ {p}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.611 + ], + "angle": 0, + "content": "Actually, the uniqueness constraint is derived from an aggressive postulate that the 3D surface around the measured point is continuous and orthogonal with the Z axis, so the uniqueness constraint is a weak constraint. Thus, the camera-camera baseline should be as short as possible to ensure the uniqueness of phase value in the feasible range. In our prototype, we simply placed the main and auxiliary cameras tightly against each other. In the future, by using a beam splitter, the camera baseline can be further shortened, breaking through the physical size limitations of industrial cameras." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.785, + 0.673 + ], + "angle": 0, + "content": "In summary, utilizing the proposed paraxial binocular SL system enables the achievement of unambiguous 3D reconstruction within the specified depth range, using only a single set of high-frequency fringe patterns. As a result, we can employ the cyclic projection strategy [31] for dynamic 3D scanning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.691, + 0.71 + ], + "angle": 0, + "content": "4 Binomial Self-compensation for Motion Error" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.721, + 0.784, + 0.811 + ], + "angle": 0, + "content": "In this paper, we cyclically project a group of high-frequency fringe patterns with a phase shift of \\(\\pi /2\\). The 3D point corresponding to a camera pixel \\((x^{c},y^{c})\\) is the intersection point of the sight line and the object surface. The intersection point moves along with the sight line on the circumstance that measures moving objects. Assuming the object is of low-frequency texture, the \\(i\\) th captured frame can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.819, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\nI _ {i} \\left(x ^ {c}, y ^ {c}\\right) = A ^ {c} + B ^ {c} \\cos \\left(\\phi_ {0} - i \\frac {\\pi}{2} + x _ {i}\\right), \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "where \\(A^c\\) is the background intensity, \\(B^c\\) is the modulation, \\(x_{i}\\) represents the unknown phase offset induced by the motion from the datum frame to \\(i\\)-th frame, note that the datum frame can be arbitrarily chosen, and the true phase value at \\(i\\) th frame is \\(\\phi_i = \\phi_0 - i\\frac{\\pi}{2}\\). We choose a \\(\\pi /2\\) phase shifting as it only transforms \\(\\sin (\\cdot)\\) and \\(\\cos (\\cdot)\\) terms into each other without inducing extra phase offset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.788, + 0.272 + ], + "angle": 0, + "content": "We notice that \\( x_{i} \\) and its high order difference exactly correspond to the concepts of displacement, velocity, and acceleration, et. al, in kinematics, and can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.292, + 0.786, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\Delta^ {(K)} x _ {i} = \\Delta^ {(K - 1)} x _ {i + 1} - \\Delta^ {(K - 1)} x _ {i} \\\\ \\Delta^ {(K)} x _ {i} = \\sum_ {k = 0} ^ {K} (- 1) ^ {k} \\binom {K} {k} x _ {i + K - k}, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.786, + 0.406 + ], + "angle": 0, + "content": "where we define \\(\\varDelta^{(0)}x_{i} = x_{i}\\). Eq. (5) is the key to motion error compensation, and we will use it later." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.456 + ], + "angle": 0, + "content": "Utilizing a cyclic projection strategy [31] for high speed 3D scanning, a three or four-step phase shifting algorithm is typically employed to calculate the motion-affected phase as:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.477, + 0.786, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\phi} _ {i} ^ {3} = \\tan^ {- 1} \\left(\\frac {\\tilde {S} _ {i} ^ {3}}{\\tilde {C} _ {i} ^ {3}}\\right) \\text {a n d} \\tilde {\\phi} _ {i} ^ {4} = \\tan^ {- 1} \\left(\\frac {\\tilde {S} _ {i} ^ {4}}{\\tilde {C} _ {i} ^ {4}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.536, + 0.264, + 0.549 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.56, + 0.786, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} S _ {i} ^ {3} = 2 I _ {i + 1} - I _ {i} - I _ {i + 2} \\\\ C _ {i} ^ {3} = I _ {i} - I _ {i + 2} \\end{array} \\right. \\text {a n d} \\left\\{ \\begin{array}{l} S _ {i} ^ {4} = I _ {i + 1} - I _ {i + 3} \\\\ C _ {i} ^ {4} = I _ {i} - I _ {i + 2} \\end{array} \\right.. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Please note that the \"three-step phase shifting\" discussed in this paper differs slightly from the traditional approach, as the phase shifting is \\(\\pi /2\\) instead of the traditional \\(2\\pi /3\\). We substitute Eq. (4) into Eq. (6) to evaluate the motion error, for a small phase-shift error \\(x_{i}\\), \\(\\sin (x_i)\\approx x_i\\) and \\(\\cos (x_i)\\approx 1\\), thus we have" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.711, + 0.798, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\epsilon_ {3} (\\phi_ {i}) = \\tilde {\\phi} _ {i} ^ {3} - \\phi_ {i} \\\\ \\approx \\tan^ {- 1} \\left(\\frac {\\left[ 1 + 0 . 5 \\left(x _ {i} - x _ {i + 2}\\right) \\right] \\sin \\left(\\phi_ {i}\\right) + \\cos \\left(\\phi_ {i}\\right) x _ {i + 1}}{\\cos \\left(\\phi_ {i}\\right) - 0 . 5 \\left(x _ {i} + x _ {i + 2}\\right) \\sin \\left(\\phi_ {i}\\right)}\\right) - \\tan^ {- 1} \\left(\\frac {\\sin \\left(\\phi_ {i}\\right)}{\\cos \\left(\\phi_ {i}\\right)}\\right) \\\\ \\approx \\frac {x _ {i} + 2 x _ {i + 1} + x _ {i + 2}}{4} + \\frac {2 x _ {i + 1} - (x _ {i} + x _ {i + 2})}{4} \\cos (2 \\phi_ {i}) + \\frac {x _ {i} - x _ {i + 2}}{4} \\sin (2 \\phi_ {i}) \\\\ = \\delta_ {i} ^ {3} + \\frac {2 x _ {i + 1} - \\left(x _ {i} + x _ {i + 2}\\right)}{4} \\cos \\left(2 \\phi_ {i}\\right) + \\frac {x _ {i} - x _ {i + 2}}{4} \\sin \\left(2 \\phi_ {i}\\right), \\tag {8} \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.142, + 0.77, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.402, + 0.677, + 0.417 + ], + "angle": 0, + "content": "Fig. 2: Binomial self-compensation for motion error." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.455, + 0.249, + 0.468 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.485, + 0.758, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\epsilon_ {4} \\left(\\phi_ {i}\\right) = \\tilde {\\phi} _ {i} ^ {4} - \\phi_ {i} \\\\ \\begin{array}{l} \\approx \\tan^ {- 1} \\left(\\frac {2 \\sin (\\phi_ {i}) + (x _ {i + 1} + x _ {i + 3}) \\cos (\\phi_ {i})}{2 \\cos (\\phi_ {i}) - (x _ {i} + x _ {i + 2}) \\sin (\\phi_ {i})}\\right) - \\tan^ {- 1} \\left(\\frac {\\sin (\\phi_ {i})}{\\cos (\\phi_ {i})}\\right) \\\\ x _ {1} + x _ {1 + 2} + x _ {1 + 3} + x _ {1 + 4} = x _ {1} + x _ {1 + 5} = x _ {1 + 6} \\end{array} \\\\ \\approx \\frac {x _ {i} + x _ {i + 2} + x _ {i + 1} + x _ {i + 3}}{4} + \\frac {x _ {i + 1} - x _ {i} + x _ {i + 3} - x _ {i + 2}}{4} \\cos \\left(2 \\phi_ {i}\\right) \\\\ = \\delta_ {i} ^ {4} + \\frac {x _ {i + 1} - x _ {i} + x _ {i + 3} - x _ {i + 2}}{4} \\cos (2 \\phi_ {i}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.711 + ], + "angle": 0, + "content": "According to Eq. (8) and (9), we discover that the motion error comprises a DC component and a harmonic component with twice the frequency of the wrapped phase. The DC component, manifesting as the lag of the point clouds due to object motion, is independent of the phase value, and thus is irrelevant to the ripple-shaped error. Consequently, the crux of motion error compensation lies in eliminating the harmonic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Analyzing two successive phase frames \\(\\tilde{\\phi}_i\\) and \\(\\tilde{\\phi}_{i + 1}\\) in a discrete time series, we discover that the coefficients of trigonometric terms exhibit opposite signs owing to a phase shift of \\(\\pi /2\\). Consequently, summing each pair of successive phase frames yields higher-order difference terms in the coefficients of harmonics. Repeating this process with the resultant phase frames, as depicted in Fig. 2, naturally evokes the concept of Yang Hui's triangle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Let's consider calculating the weighted sum of \\( K + 1 \\) successive phase frames in a time series as shown in Fig. 2, using the \\( (K + 1) \\)-th row of Yang Hui's" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.496, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.307, + 0.365, + 0.319 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.144, + 0.788, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.307, + 0.649, + 0.319 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.33, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Fig. 3: Motion error converges as \\( K \\) increases: (a) three-step VS four-step phase shifting, and (b) visualization of motion error in four-step phase shifting, the mean absolute error of phase exponentially diminishes as \\( K \\) increases." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.402, + 0.638, + 0.416 + ], + "angle": 0, + "content": "triangle, also as known as \\( K \\)-th order binomial coefficients" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.426, + 0.785, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\phi} _ {i} (K) = 2 ^ {- K} \\sum_ {k = 0} ^ {K} \\binom {K} {k} \\tilde {\\phi} _ {i + K - k} (0), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.784, + 0.51 + ], + "angle": 0, + "content": "where \\(\\tilde{\\phi}_i(0)\\) is computed by Eq. (6). We substitute Eq. (5) and (8) into Eq. (10) to have" + }, + { + "type": "equation", + "bbox": [ + 0.298, + 0.517, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\phi} _ {i} ^ {3} (K) = \\bar {\\phi} _ {i} ^ {3} (K) - 2 ^ {- (K + 2)} \\Delta^ {(K + 2)} x _ {i + 2} \\cos (2 \\phi_ {i}) \\\\ + 2 ^ {- (K + 2)} \\left(\\Delta^ {(K + 1)} x _ {i + 2} + \\Delta^ {(K + 1)} x _ {i + 1}\\right) \\sin \\left(2 \\phi_ {i}\\right), \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.575, + 0.264, + 0.587 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.316, + 0.585, + 0.785, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\bar {\\phi} _ {i} ^ {3} (K) = 2 ^ {- (K + 2)} \\sum_ {k = 0} ^ {K} \\left[ \\binom {K} {k} \\left(\\phi_ {i + K - k} + \\delta_ {i + K - k} ^ {3}\\right) \\right]. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.632, + 0.547, + 0.647 + ], + "angle": 0, + "content": "Similarly, for four-step phase shifting we have" + }, + { + "type": "equation", + "bbox": [ + 0.255, + 0.657, + 0.785, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\phi} _ {i} ^ {4} (K) = \\bar {\\phi} _ {i} ^ {4} (K) + 2 ^ {- (K + 2)} \\left(\\Delta^ {(K + 1)} x _ {i} + \\Delta^ {(K + 1)} x _ {i + 2}\\right) \\cos \\left(2 \\phi_ {i}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.692, + 0.264, + 0.705 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.316, + 0.703, + 0.785, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\bar {\\phi} _ {i} ^ {4} (K) = 2 ^ {- (K + 2)} \\sum_ {k = 0} ^ {K} \\left[ \\binom {K} {k} \\left(\\phi_ {i + K - k} + \\delta_ {i + K - k} ^ {4}\\right) \\right]. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.81 + ], + "angle": 0, + "content": "Based on Eq. (11) and (13), it is clear that increasing \\( K \\) produces two main effects: first, the factor \\( 2^{-(K + 2)} \\) exponentially decreases the harmonic amplitudes; second, higher-order difference terms (approach zero as \\( K \\) increases) occur in the harmonic amplitudes. Both effects significantly diminish the motion error." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We measure a flat plate moving towards the camera using our binomial self-compensation (BSC) for motion error, the results visualized in Fig. 3(a) show" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.253 + ], + "angle": 0, + "content": "that the ripples induced by motion in the self-compensated phase \\(\\tilde{\\phi}_i^4 (K)\\) reduces significantly with the increment of \\(K\\). \\(K = 0\\) represents the raw phase that is affected by motion error. The residual error is caused by the nonlinearity of the binary defocus technique as shown in Fig. 3(b). At least four frames of images are needed in BSC, and using eight images achieves a trade-off between accuracy and efficiency in practice. Note that the binomial order \\(K\\) can be flexibly chosen according to the accuracy and time efficiency requirements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.314 + ], + "angle": 0, + "content": "In practice, we decompose Eq. (11) into a pyramid-like procedure of adding up the successive two frames layer by layer, as shown in Fig. 2. However, due to the ambiguity of the wrapped phase, phase order error occurs around the phase jumping area. To address this issue, we define the \\(\\oplus\\) operator as" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.322, + 0.785, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\phi_ {a} \\oplus \\phi_ {b} := \\left\\{ \\begin{array}{l} 0. 5 \\left(\\phi_ {a} + \\phi_ {b}\\right), \\left| \\phi_ {a} - \\phi_ {b} \\right| > \\pi \\\\ 0. 5 \\left(\\phi_ {a} + \\phi_ {b}\\right) + \\pi , \\left| \\phi_ {a} - \\phi_ {b} \\right| < \\pi \\end{array} . \\right. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.362, + 0.788, + 0.468 + ], + "angle": 0, + "content": "This operation automatically compensates for the phase order error as long as the phase difference is greater than \\(\\pi\\). The motion-error-free phase \\(\\tilde{\\phi}_i(K)\\) can be obtained by employing our binomial self-compensation (BSC), which is summarized as Algorithm 1, where \"mod\" represents modular operation. Note that we have corrected the constant phase shifting before adding up two phase frames \\(\\tilde{\\phi}_i(0)\\) and \\(\\tilde{\\phi}_{i + 1}(0)\\). The MATLAB code for implementing the algorithm are given in the github repository." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.504, + 0.507 + ], + "angle": 0, + "content": "5 Stereo Phase Unwrapping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.784, + 0.624 + ], + "angle": 0, + "content": "In the previous section, we obtained the motion-error-free high-frequency phase by conducting BSC. Subsequently, we employ a sum of absolute difference algorithm on the wrapped phase to retrieve the correspondence between the main and auxiliary cameras within the disparity range obtained by pre-set depth range. Further, the projector space coordinates \\((x^p,y^p)\\) can be computed by directly re-projecting the binocular 3D coordinate onto the projector space. Finally, we can have the phase order and compute the unwrapped phase." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.645, + 0.502, + 0.663 + ], + "angle": 0, + "content": "6 Experimental Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The proposed algorithm is implemented on an Intel(R) Core(TM) i7-13790F @ 2.10 GHz with 32 GB RAM and written in MATLAB. Our experimental system (Fig. 4) consists of two AVT 1800U-120c cameras that have \\(1280 \\times 960\\) resolution, equipped with \\(8\\mathrm{mm}\\) Computer lenses, and a TI DLP4500 projector with a resolution of \\(912 \\times 1140\\). We cropped the cameras' field of view to \\(640 \\times 480\\) to have a 90 fps image acquisition speed. The two cameras are synchronized by the trigger signal from the projector. In order to comprehensively demonstrate the superiority of our BSC, we conducted four groups of experiments on dynamic objects for evaluating: 1) absolute accuracy of 3D reconstruction, 2) temporal resolution of 3D imaging, 3) robustness on depth discontinuous scenes, and 4) generality on different dynamic objects." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zhang et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.229, + 0.15, + 0.497, + 0.165 + ], + "angle": 0, + "content": "Algorithm 1: BSC for motion error" + }, + { + "type": "algorithm", + "bbox": [ + 0.226, + 0.167, + 0.661, + 0.395 + ], + "angle": 0, + "content": "Input: Captured image sequence: \\(\\{I_0,\\dots ,I_{T - 1}\\}\\) Binomial order: \\(K\\) Phase-shifting step: \\(N = 3\\) or 4. Output: Motion-error-free phase sequence: \\(\\{\\bar{\\phi}_0(K),\\dots,\\bar{\\phi}_{T - K - N + 1}(K)\\}\\) for \\(t\\gets 0\\) to \\(T - N\\) do \\(\\tilde{\\phi}_t(0)\\gets \\left\\{ \\begin{array}{ll}\\mathrm{mod}\\left[\\tan^{-1}\\left(\\frac{2I_{t + 1} - I_t - I_{t + 2}}{I_t - I_{t + 2}}\\right) + t\\frac{\\pi}{2},2\\pi \\right],N = 3\\\\ \\mathrm{mod}\\left[\\tan^{-1}\\left(\\frac{I_{t + 1} - I_{t + 3}}{I_t - I_{t + 2}}\\right) + t\\frac{\\pi}{2},2\\pi \\right],N = 4 \\end{array} \\right.\\) end for \\(k\\gets 0\\) to \\(K - 1\\) do for \\(i\\gets 0\\) to \\(T - k - N\\) do \\(\\tilde{\\phi}_i(k + 1)\\gets \\tilde{\\phi}_i(k)\\oplus \\tilde{\\phi}_{i + 1}(k)\\) end" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.413, + 0.701, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.516, + 0.595, + 0.53 + ], + "angle": 0, + "content": "Fig. 4: Experimental setup." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.559, + 0.6, + 0.575 + ], + "angle": 0, + "content": "6.1 Absolute Accuracy of 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.809 + ], + "angle": 0, + "content": "We measured a periodically waving plate at a distance of \\(500~\\mathrm{mm}\\) by using our BSC (8 frames) and five representative methods for comparison: 1) traditional four-step PSP (4 frames), 2) Hilbert transform compensation (HTC, 4 frames) [24], 3) micro Fourier transform profilometry (\\(\\mu\\)-FTP, 4 frames) [34], 4) phase frame difference method (PFD, 8 frames) [10], 5) phase frame sum method (PFS, 5 frmaes) [4]. The fringe wavelength was set to 24 pixels for the cyclic \\(\\pi/2\\) phase shifting patterns and was set to [22,24,26] for \\(\\mu\\)-FTP. The plate moves within a distance range of [400,500] mm and a speed range of \\([-150,150]\\) mm/s, and our measurement lasted for 800 frames (about 8.9 seconds). We select three \\(160\\times 120\\) rectangular windows in each depth map and fit the quadric surface as the ground truth, then we compute the root mean squared error (RMSE) within the selected regions. The trend of RMSE changing with motion speed is depicted in Fig. 5. The results show that our method exhibits the most satisfactory accuracy among existing methods when dealing with motion error." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Then we selected a moment when the speed of the plate was \\(-88~\\mathrm{mm / s}\\), and depict the reconstructed point clouds in Fig. 6, our BSC effectively reduces the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.26, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.144, + 0.703, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.371, + 0.788, + 0.425 + ], + "angle": 0, + "content": "Fig. 5: Measurement error VS motion speed of a periodically waving plate by traditional four-step phase shifting, HTC [24], \\(\\mu\\)-FTP [34], PFD [10], PFS [4], and our BSC, the value in parentheses represents the mean error. Note the y-axis is in logarithmic scale." + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.443, + 0.329, + 0.456 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.467, + 0.401, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.443, + 0.502, + 0.456 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.467, + 0.572, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.443, + 0.674, + 0.456 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.454, + 0.775, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.583, + 0.329, + 0.597 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.6, + 0.4, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.584, + 0.501, + 0.597 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.599, + 0.572, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.584, + 0.673, + 0.597 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.599, + 0.774, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.72, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Fig. 6: Measurement results of a moving flat plane by (a) traditional four-step phase shifting, (b) HTC [24], (c) \\(\\mu\\)-FTP [34], (d) PFD [10], (e) PFS [4], and (f) our BSC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "motion error from \\(324.2\\mu \\mathrm{m}\\) to \\(54.78\\mu \\mathrm{m}\\) compared with traditional four-step phase shifting. By counting the histogram of errors, it can be observed that the error distribution of our BSC visually presents a normal distribution, which means Gaussian noise (induced by unstable ambient light, camera/projector" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.174, + 0.405, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.147, + 0.51, + 0.157 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.159, + 0.588, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.251, + 0.59, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.147, + 0.707, + 0.157 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.159, + 0.775, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.25, + 0.776, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.327, + 0.405, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.305, + 0.509, + 0.314 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.326, + 0.588, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.406, + 0.59, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.305, + 0.707, + 0.314 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.326, + 0.775, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.407, + 0.776, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Fig. 7: Depth discontinuous scene (top left) and motion area labeled in white (bottom left). Reconstruction results and error at edges: (a) traditional four-step phase shifting, (b) HTC [24], (c) \\(\\mu\\)-FTP [34], and (d) our BSC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.608 + ], + "angle": 0, + "content": "flicker, quantization error, and sensor noise of the camera and projector) is the main component of error. In other words, the motion error is effectively suppressed by BSC. Meanwhile, the error distribution of other methods deviates from the normal distribution, indicating the existence of residual errors." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.64, + 0.622, + 0.656 + ], + "angle": 0, + "content": "6.2 Robustness on Depth Discontinuous Scenes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "PSP's pixel-wise characteristic is one of its most crucial merits. Our BSC maintains this property, thereby exhibiting robustness to depth discontinuous scenes. We set up a scene with three static gypsum geometries and one moving plate. Then we employed our BSC, \\(\\mu\\)-FTP, and HTC. The results are shown in Fig. 7. It can be seen that the two non-pixel-wise methods, i.e., \\(\\mu\\)-FTP and HTC, exhibit visible deformation to the naked eye in the point cloud around the depth jumping areas. We selected three rectangular windows and quantitatively evaluate the RMSE, the results show that the RMSE of non-pixel-wise methods is significantly greater than our BSC. We can conclude that our BSC with pixel-wise property is more robust compared with non-pixel-wise methods when dealing with depth discontinuous scenes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.259, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.142, + 0.784, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.34, + 0.788, + 0.369 + ], + "angle": 0, + "content": "Fig. 8: The measurement result of a wooden butterfly with wings flapping, during \\( \\mathrm{T} = 0\\mathrm{ms}\\sim 88\\mathrm{ms} \\), Row 1 and 2: our BSC; Row 3 and 4: \\( \\mu \\)-FTP [34]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.398, + 0.56, + 0.414 + ], + "angle": 0, + "content": "6.3 Temporal Resolution of 3D Imaging" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.588 + ], + "angle": 0, + "content": "Our BSC employs cyclic \\(\\pi /2\\) phase-shifting fringe patterns, any \\(K + N\\) successive captured images can be used for reconstructing one frame of 3D point clouds. Therefore, our resulting frame rate of 3D point clouds is the same as the camera frame rate, thereby achieving a high temporal resolution of the measurement. That is the frame-wise loopable property we mentioned in the introduction. We successively conducted BSC and \\(\\mu\\)-FTP to measure a wooden butterfly with wings flapping cyclically, and then manually aligned the timeline, the results are shown in Fig. 8. We noticed that our method can achieve 3D reconstruction with a frame rate of 90 fps (the same as the camera frame rate), while \\(\\mu\\)-FTP reconstructing at 30 fps. Thus, our method is capable of distinguishing tinier motions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.61, + 0.604, + 0.626 + ], + "angle": 0, + "content": "6.4 Generality on Different Dynamic Objects" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.71 + ], + "angle": 0, + "content": "We captured 400 image frames (about 4 seconds) employing our pipeline on four different dynamic objects, and the results are shown in Fig. 9. The complete motion process video can be found in the supplementary materials. We observed that BSC can effectively eliminate the severe ripple errors caused by object motion." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.733, + 0.564, + 0.75 + ], + "angle": 0, + "content": "7 Limitations and Future Outlook" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.809 + ], + "angle": 0, + "content": "We presented a pixel-wise and frame-wise loopable binomial self-compensation algorithm to effectively and flexibly eliminate motion error in four-step PSP, without any intermediate variables assisted." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "High-frequency textures: our BSC assumes low-frequency textures like most motion compensation algorithms, it is robust against moderate contrast" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.151, + 0.36, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.152, + 0.491, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.152, + 0.622, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.152, + 0.78, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.259, + 0.359, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.258, + 0.491, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.257, + 0.622, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.259, + 0.78, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 9: The measurement results of four different dynamic objects. Column 1: wooden butterfly model with wings flapping, Column 2: moving statue, Column 3: waving hand, Column 4: hand showing scissors, stone, and cloth gestures. Top to bottom: our BSC and traditional four-step PSP." + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.446, + 0.648, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.573, + 0.701, + 0.587 + ], + "angle": 0, + "content": "Fig. 10: 3D error VS inter-frame phase-shifting error \\( |\\Delta x_{i}| \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.785, + 0.682 + ], + "angle": 0, + "content": "textures. However, when dealing with objects that have high contrast textures, our method induces artifacts similar to those seen with typical phase-shifting methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.785, + 0.755 + ], + "angle": 0, + "content": "Phase unwrapping error: we adopt a common block matching algorithm for stereo correspondence, inducing obvious phase unwrapping error when measuring complex surfaces. In the future, we will conduct a hierarchical adjustment for phase outliers to optimize the accuracy of the point cloud." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Applicable speed range: Our BSC is predicated on the assumption that the phase-shift error \\( x_{i} \\) is small. Consequently, as shown in Fig. 10, the residual motion error arises as \\( |\\Delta x_{i}| \\) increases. Properly reducing the fringe frequency could decrease \\( |\\Delta x_{i}| \\), alleviating the performance loss of BSC for high-speed objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.26, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Achar, S., Narasimhan, S.G.: Multi focus structured light for recovering scene shape and global illumination. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 205-219. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.263 + ], + "angle": 0, + "content": "2. Chen, T., Seidel, H.P., Lensch, H.P.: Modulated phase-shifting for 3d scanning. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1-8. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Geng, J.: Structured-light 3d surface imaging: a tutorial. Advances in optics and photonics 3(2), 128-160 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.333 + ], + "angle": 0, + "content": "4. Guo, W., Wu, Z., Zhang, Q., Wang, Y.: Real-time motion-induced error compensation for 4-step phase-shifting profilometry. Optics Express 29(15), 23822-23834 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.335, + 0.785, + 0.375 + ], + "angle": 0, + "content": "5. Gupta, M., Nakhate, N.: A geometric perspective on structured light coding. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 87-102 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.417 + ], + "angle": 0, + "content": "6. Gupta, M., Tian, Y., Narasimhan, S.G., Zhang, L.: A combined theory of defocused illumination and global light transport. International Journal of Computer Vision 98, 146-167 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.446 + ], + "angle": 0, + "content": "7. Gupta, M., Yin, Q., Nayar, S.K.: Structured light in sunlight. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 545-552 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.473 + ], + "angle": 0, + "content": "8. Hirose, K., Kamei, H., Sugiyama, T., Kurosaka, Y.: \\(200 \\times 200 \\mu \\mathrm{m}\\) 2 structured light source. Optics Express 28(25), 37307-37321 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.516 + ], + "angle": 0, + "content": "9. Honrado, C.P., Larrabee Jr, W.F.: Update in three-dimensional imaging in facial plastic surgery. Current opinion in otolaryngology & head and neck surgery 12(4), 327-331 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.544 + ], + "angle": 0, + "content": "0. Liu, X., Tao, T., Wan, Y., Kofman, J.: Real-time motion-induced-error compensation in 3d surface-shape measurement. Optics Express 27(18), 25265-25279 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.572 + ], + "angle": 0, + "content": "1. Liu, Y., Pears, N., Rosin, P.L., Huber, P.: 3D imaging, analysis and applications. Springer (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.6 + ], + "angle": 0, + "content": "2. Liu, Z., Zibley, P.C., Zhang, S.: Motion-induced error compensation for phase shifting profilometry. Optics Express 26(10), 12632-12637 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.642 + ], + "angle": 0, + "content": "3. Logozzo, S., Zanetti, E.M., Franceschini, G., Kilpelä, A., Mäkynen, A.: Recent advances in dental optics-part i: 3d intraoral scanners for restorative dentistry. Optics and Lasers in Engineering 54, 203-221 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.644, + 0.785, + 0.685 + ], + "angle": 0, + "content": "4. Lu, L., Ding, Y., Luan, Y., Yin, Y., Liu, Q., Xi, J.: Automated approach for the surface profile measurement of moving objects based on psp. Optics Express 25(25), 32120-32131 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.727 + ], + "angle": 0, + "content": "5. Lu, L., Xi, J., Yu, Y., Guo, Q.: New approach to improve the accuracy of 3-d shape measurement of moving object using phase shifting profilometry. Optics Express 21(25), 30610-30622 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.755 + ], + "angle": 0, + "content": "6. Lu, L., Yin, Y., Su, Z., Ren, X., Luan, Y., Xi, J.: General model for phase shifting profilometry with an object in motion. Applied Optics 57(36), 10364-10369 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.783 + ], + "angle": 0, + "content": "7. Lv, S., Kemao, Q.: Modeling the measurement precision of fringe projection profilometry. Light: Science & Applications 12(1), 257 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.811 + ], + "angle": 0, + "content": "8. Nayar, S.K., Gupta, M.: Diffuse structured light. In: IEEE International Conference on Computational Photography (ICCP). pp. 1-11. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Posdamer, J.L., Altschuler, M.D.: Surface measurement by space-encoded projected beam systems. Computer Graphics and Image Processing 18(1), 1-17 (1982)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "G. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "20. Qian, J., Feng, S., Xu, M., Tao, T., Shang, Y., Chen, Q., Zuo, C.: High-resolution real-time 360 3d surface defect inspection with fringe projection profilometry. Optics and Lasers in Engineering 137, 106382 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "21. Qian, J., Tao, T., Feng, S., Chen, Q., Zuo, C.: Motion-artifact-free dynamic 3d shape measurement with hybrid fourier-transform phase-shifting profilometry. Optics Express 27(3), 2713-2731 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "22. Sansoni, G., Corini, S., Lazzari, S., Rodella, R., Docchio, F.: Three-dimensional imaging based on gray-code light projection: characterization of the measuring algorithm and development of a measuring system for industrial applications. Applied Optics 36(19), 4463-4472 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "23. Srinivasan, V., Liu, H.C., Halioua, M.: Automated phase-measuring profilometry of 3-d diffuse objects. Applied optics 23(18), 3105-3108 (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "24. Wang, Y., Liu, Z., Jiang, C., Zhang, S.: Motion induced phase error reduction using a hilbert transform. Optics Express 26(26), 34224-34235 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "25. Wang, Y., Suresh, V., Li, B.: Motion-induced error reduction for binary defocusing profilometry via additional temporal sampling. Optics Express 27(17), 23948-23958 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "26. Weise, T., Leibe, B., Van Gool, L.: Fast 3d scanning with automatic motion compensation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1-8. IEEE (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "27. Wu, G., Yang, T., Liu, F., Qian, K.: Suppressing motion-induced phase error by using equal-step phase-shifting algorithms in fringe projection profilometry. Optics Express 30(11), 17980-17998 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "28. Ye, Y., Chang, H., Song, Z., Zhao, J.: Accurate infrared structured light sensing system for dynamic 3d acquisition. Applied Optics 59(17), E80-E88 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "29. Zhang, Y., Lau, D., Wipf, D.: Sparse multi-path corrections in fringe projection profilometry. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13344-13353 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "30. Zhang, Y., Lau, D.L., Yu, Y.: Causes and corrections for bimodal multi-path scanning with structured light. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4431-4439 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "31. Zhong, K., Li, Z., Shi, Y., Wang, C., Lei, Y.: Fast phase measurement profilometry for arbitrary shape objects without phase unwrapping. Optics and Lasers in Engineering 51(11), 1213-1222 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "32. Zhu, J., Yang, F., Hu, J., Zhou, P.: High dynamic reflection surface 3d reconstruction with sharing phase demodulation mechanism and multi-indicators guided phase domain fusion. Optics Express 31(15), 25318-25338 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "33. Zuo, C., Feng, S., Huang, L., Tao, T., Yin, W., Chen, Q.: Phase shifting algorithms for fringe projection profilometry: A review. Optics and lasers in engineering 109, 23-59 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "34. Zuo, C., Tao, T., Feng, S., Huang, L., Asundi, A., Chen, Q.: Micro fourier transform profilometry (μftp): 3d shape measurement at 10,000 frames per second. Optics and Lasers in Engineering 102, 70-91 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_origin.pdf b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..13d511cd0431aefb7717b5f73eff7f94b51eef31 --- /dev/null +++ b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/fec97731-b9d0-4bd8-a038-75b0c34b0968_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f51eb463ef515f772203f949416b7d8ce94e9c6ebd1888d7d2f93b22d03659e0 +size 4853896 diff --git a/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/full.md b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..26f2c029fdb0f0e366ae5cd2c3b56f8891a4dd4a --- /dev/null +++ b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/full.md @@ -0,0 +1,342 @@ +# Binomial Self-compensation for Motion Error in Dynamic 3D Scanning + +Geyou Zhang $^{1}$ , Ce Zhu $^{1,*}$ , and Kai Liu $^{2}$ + +1 University of Electronic Science and Technology of China, geyouzhang@foxmail.com, eczhu@uestc.edu.cn + +2 Sichuan University + +kailiu@scu.edu.cn + +Abstract. Phase shifting profilometry (PSP) is favored in high-precision 3D scanning due to its high accuracy, robustness, and pixel-wise property. However, a fundamental assumption of PSP that the object should remain static is violated in dynamic measurement, making PSP susceptible to object moving, resulting in ripple-like errors in the point clouds. We propose a pixel-wise and frame-wise loopable binomial self-compensation (BSC) algorithm to effectively and flexibly eliminate motion error in the four-step PSP. Our mathematical model demonstrates that by summing successive motion-affected phase frames weighted by binomial coefficients, motion error exponentially diminishes as the binomial order increases, accomplishing automatic error compensation through the motion-affected phase sequence, without the assistance of any intermediate variable. Extensive experiments show that our BSC outperforms the existing methods in reducing motion error, while achieving a depth map frame rate equal to the camera's acquisition rate (90 fps), enabling high-accuracy 3D reconstruction with a quasi-single-shot frame rate. The code is available at https://github.com/GeyouZhang/BSC. + +Keywords: Phase Shifting Profilometry $\cdot$ Dynamic 3D Scanning $\cdot$ Motion Error Compensation + +# 1 Introduction + +In recent years, structured light (SL) 3D imaging has been penetrated into diverse fields, e.g., industrial inspection [20] and design [32], medical 3D reconstruction [3,8,13,17], virtual reality [28], and digitalization of cultural heritage [9]. Phase shifting profilometry (PSP) [23], Fourier transform profilometry, binary code [19], and Gray code [22] are famous SL technologies that are widely applied in industry and scientific researches. Among the SL technologies above, PSP is especially favored in high-precision 3D scanning due to its high accuracy, robustness, and pixel-wise property [33]. Many studies have excavated the huge potentiality of PSP in various challenging scenes, such as low signal-to-noise [5,7], multi-path [29,30], high reflectivity [18], and global illumination [1,2,6], et. al. + +However, a fundamental assumption of PSP that the object should remain static is violated in dynamic measurement, making PSP susceptible to object + +moving, resulting in ripple-like errors in the point clouds. We observed that the solutions of PSP in dynamic scenes are overlooked within the computer vision community. Concurrently, there are some existing methods in the optics metrology community to address this challenge, most of them can be categorized into three types: 1) motion-induced phase shifting estimation, 2) object tracking, and 3) automatic compensation. + +Motion-induced phase shifting estimation: In traditional PSP, the phase shifting of each pattern is equally distributed or pre-known values within one period of fringe, which becomes disorderly due to object motion. Thus, researchers estimate the motion-induced phase shifting according to the implicit motion information in the motion-affected phase frame [26] or frames [10, 12, 21, 25], thereby restoring the unknown motion-affected phase shifting value of each pattern, and finally solve the corrected phase by least squared method [11]. However, these methods may not effectively compensate for errors when the motion information can't be precisely predicted. + +Object tracking: By tracking the position of the object by placing markers [15] or using image feature matching algorithm [16], the motion error can be corrected by identifying the trajectory of the object [14-16]. However, a limitation of the object tracking based methods is that they are tailored for 2D motion perpendicular to the sight line. Moreover, the accuracy and reliability of feature tracking are crucial issues when faced with textureless objects. + +Automatic compensation: In automatic compensation methods, instead of computing motion-induced phase shifting value or tracking the object, an appropriate operator is selected to conduct on the whole captured image [24, 27] or phase frame [4], and further directly obtain the motion-error-free phase, in which the motion error will be automatically compensated. + +Previous work on motion compensation typically estimates or generates certain intermediate variables, then calculates the motion-error-free phase based on these variables, either directly or indirectly. For example, in phase shifting estimation methods, the intermediate variables are motion-induced phase shifting values [10, 12, 21, 25]; in object tracking methods, the intermediate variable is the rigid transform matrix [14-16]; in automatic compensation methods, the intermediate variables are the captured images after undergoing Hilbert transform [24] or differential operator [27]. Therefore, two questions naturally arise: Why not utilize the motion-affected phase itself to compensate for motion error without depending on any intermediate variables? Would this approach be more effective and flexible? + +In this paper, without depending on any extra intermediate variable, we develop a binomial self-compensation (BSC) method that only utilizes the motion-affected phase sequence itself, thereby effectively and flexibly eliminating the motion error in four-step PSP. To achieve this, first, we design a paraxial binocular structured light system for reconstructing 3D point clouds using a single set of high-frequency fringe patterns. This system intentionally sets a very short baseline between two cameras to achieve a narrow disparity range, facilitating the use of high frequency fringes, while ensuring the baseline between the main + +camera and projector is long enough for adequate accuracy. Subsequently, we cyclically [31] project high-frequency patterns with $\pi /2$ phase shifting, and directly compute a sequence of phase frames from each three or four successive captured images. Finally, by adding up a sequence of successive phase frames weighted by binomial coefficients, we achieve self-compensation of motion errors. Our mathematical model demonstrates that motion error exponentially diminishes as the binomial order $K$ increases. + +Our BSC not only inherits the pixel-wise advantages of PSP, but also is of high temporal resolution to achieve quasi-single-shot 3D imaging frame rate because it's frame-wise loopable. Furthermore, our method offers flexibility and ease of deployment, acting as an enhanced, motion-error-free version of the four-step PSP, thus providing a plug-and-play enhancement for dynamic 3D scanning scenarios. + +# 2 Related Work + +In the introduction section, based on specific implementation details, we categorize the motion error compensation methods into three classes. From another perspective, we can also distinguish them based on their 3D reconstruction performance on temporal and spatial resolution. We examine the motion compensation methods from two perspectives: property of pixel-wise and frame-wise loopable. Pixel-wise methods compute phase and point clouds pixel by pixel, achieving spatial independence through this detailed granularity. Frame-wise loopable methods accomplish the same depth map frame rate as the camera acquisition frame rate, enabling a quasi-single-shot 3D reconstruction frame rate. + +Pixel-wise: Several motion error compensation methods preserve the pixelwise precision of traditional PSP, which utilizing two sampling phase frames to iteratively compute motion phase shifting [12, 25], automatically compensate for errors [4], tracking the position of the object by placing markers [15], using image feature matching algorithm [16] and correct motion error by identifying the trajectory of the object [14-16]. These approaches build on the granularity of pixel-level adjustments. In contrast, the non-pixel-wise methods conduct particular operator on local [10, 26, 27] or global image [21, 24, 34], e.g., Fourier transform [21, 34], Hilbert transform [24], differential operator [27], mean filter [10], optimization algorithm [26], from which estimate the extra phase shifting induced by motion and compensate error [10, 21] or directly correct motion error [24, 26, 27, 34]. However, the pixel-wise merit of PSP is not preserved in these methods, theoretically diminishing their robustness in handling depth discontinuities. Concurrently, introducing extra non-pixel-wise operators also escalates the computational time required. + +Frame-wise loopable: By projecting high-frequency fringes and reconstructing 3D point clouds using a cyclic strategy [31], some motion error compensation methods [4, 10, 24, 26, 27] accomplish 3D reconstruction with quasi-single-shot frame rate, thus called frame-wise loopable methods in this paper. The cyclic strategy is not suitable for all motion compensation methods, some + +![](images/ddaeef3a1d8f7f61d6ee8667b6953f57330d4b708b4880b65cd22b8c67a6ecb2.jpg) +Fig. 1: Diagram of our paraxial binocular structured light system. + +methods need to project an indivisible group of patterns [21, 25, 34], such that the depth map frame rate is reduced severalfold. + +# 3 System Setup and Frequency Selection + +Our paraxial binocular structured light system comprises a main camera, an auxiliary camera, and a projector, as depicted in Fig. 1. Together, the main and auxiliary cameras form a paraxial setup to provide reference depth for phase unwrapping. Simultaneously, the combination of the main camera and projector facilitates precise 3D reconstruction. Following system calibration, we perform stereo rectification on the captured images to align the epipolar lines in the auxiliary system parallelly. + +To ensure the simplicity of equations and convey intuitive geometrical relations, the following analysis in this section is under the assumption that the optical centers of the projector, main camera, and auxiliary camera are coplanar. Note that we build the world coordinate system to coincide with the main camera coordinate system. In reality, the projector may not be strictly coplanar to the other two, but the conclusions of our analysis are still effective. + +The primary benefit of the binocular structured light (SL) system lies in its ability to perform stereo phase unwrapping (SPU) through geometric constraints, substantially decreasing the needed number of fringe patterns. To avoid phase unwrapping ambiguities, it should be ensured that only a few candidate points are within the depth range. As a result, traditional SPU approaches typically necessitate a short baseline between the main camera and the projector, which unfortunately compromises the accuracy of point clouds. + +To address this, in this paper, we shift the constraint from the camera projector baseline to the camera-camera baseline, thus allowing flexibility in the camera-projector distance and enhancing point cloud precision. This camera-camera baseline constraint is called the uniqueness constraint, ensuring that the phase remains single-valued within the feasible stereo matching range on the + +auxiliary camera, from which guarantees the uniqueness of the matching point candidate. + +Given that stereo rectification has been applied to the cameras, we can disregard the Y axis. Assuming the 3D surface near the measured point is continuous and orthogonal to the Z axis, determining the depth range allows us to calculate the maximum fringe frequency. Consider a virtual segment, represented by a black dashed line in Fig. 1, cut by the auxiliary camera's feasible region cone. The length of this virtual segment (black solid line in Fig. 1) can be calculated as: + +$$ +d = \frac {\left(Z _ {\operatorname* {m a x}} - Z _ {\operatorname* {m i n}}\right) Z L}{Z _ {\operatorname* {m a x}} Z _ {\operatorname* {m i n}}}. \tag {1} +$$ + +As long as the projection of the virtual segment onto the projector space is distributed within one fringe period, the uniqueness constraint is satisfied, i.e., + +$$ +\frac {f _ {x} ^ {p} d}{Z - Z ^ {p}} < \frac {W ^ {p}}{f}, \tag {2} +$$ + +where $f_{x}^{p}$ is the focal length of the projector, $W^{p}$ is the resolution of the projector along $X$ axis, $f$ is the frequency of fringe pattern. + +Finally, according to uniqueness constraint, we can obtain the limiting frequency of fringe as + +$$ +f < \left(1 - \frac {Z ^ {p}}{Z}\right) \frac {Z _ {\operatorname* {m i n}} Z _ {\operatorname* {m a x}} W ^ {p}}{\left(Z _ {\operatorname* {m a x}} - Z _ {\operatorname* {m i n}}\right) L f _ {x} ^ {p}}. \tag {3} +$$ + +Actually, the uniqueness constraint is derived from an aggressive postulate that the 3D surface around the measured point is continuous and orthogonal with the Z axis, so the uniqueness constraint is a weak constraint. Thus, the camera-camera baseline should be as short as possible to ensure the uniqueness of phase value in the feasible range. In our prototype, we simply placed the main and auxiliary cameras tightly against each other. In the future, by using a beam splitter, the camera baseline can be further shortened, breaking through the physical size limitations of industrial cameras. + +In summary, utilizing the proposed paraxial binocular SL system enables the achievement of unambiguous 3D reconstruction within the specified depth range, using only a single set of high-frequency fringe patterns. As a result, we can employ the cyclic projection strategy [31] for dynamic 3D scanning. + +# 4 Binomial Self-compensation for Motion Error + +In this paper, we cyclically project a group of high-frequency fringe patterns with a phase shift of $\pi /2$ . The 3D point corresponding to a camera pixel $(x^{c},y^{c})$ is the intersection point of the sight line and the object surface. The intersection point moves along with the sight line on the circumstance that measures moving objects. Assuming the object is of low-frequency texture, the $i$ th captured frame can be expressed as + +$$ +I _ {i} \left(x ^ {c}, y ^ {c}\right) = A ^ {c} + B ^ {c} \cos \left(\phi_ {0} - i \frac {\pi}{2} + x _ {i}\right), \tag {4} +$$ + +where $A^c$ is the background intensity, $B^c$ is the modulation, $x_{i}$ represents the unknown phase offset induced by the motion from the datum frame to $i$ -th frame, note that the datum frame can be arbitrarily chosen, and the true phase value at $i$ th frame is $\phi_i = \phi_0 - i\frac{\pi}{2}$ . We choose a $\pi /2$ phase shifting as it only transforms $\sin (\cdot)$ and $\cos (\cdot)$ terms into each other without inducing extra phase offset. + +We notice that $x_{i}$ and its high order difference exactly correspond to the concepts of displacement, velocity, and acceleration, et. al, in kinematics, and can be expressed as + +$$ +\left\{ \begin{array}{l} \Delta^ {(K)} x _ {i} = \Delta^ {(K - 1)} x _ {i + 1} - \Delta^ {(K - 1)} x _ {i} \\ \Delta^ {(K)} x _ {i} = \sum_ {k = 0} ^ {K} (- 1) ^ {k} \binom {K} {k} x _ {i + K - k}, \end{array} \right. \tag {5} +$$ + +where we define $\varDelta^{(0)}x_{i} = x_{i}$ . Eq. (5) is the key to motion error compensation, and we will use it later. + +Utilizing a cyclic projection strategy [31] for high speed 3D scanning, a three or four-step phase shifting algorithm is typically employed to calculate the motion-affected phase as: + +$$ +\tilde {\phi} _ {i} ^ {3} = \tan^ {- 1} \left(\frac {\tilde {S} _ {i} ^ {3}}{\tilde {C} _ {i} ^ {3}}\right) \text {a n d} \tilde {\phi} _ {i} ^ {4} = \tan^ {- 1} \left(\frac {\tilde {S} _ {i} ^ {4}}{\tilde {C} _ {i} ^ {4}}\right), \tag {6} +$$ + +where + +$$ +\left\{ \begin{array}{l} S _ {i} ^ {3} = 2 I _ {i + 1} - I _ {i} - I _ {i + 2} \\ C _ {i} ^ {3} = I _ {i} - I _ {i + 2} \end{array} \right. \text {a n d} \left\{ \begin{array}{l} S _ {i} ^ {4} = I _ {i + 1} - I _ {i + 3} \\ C _ {i} ^ {4} = I _ {i} - I _ {i + 2} \end{array} \right.. \tag {7} +$$ + +Please note that the "three-step phase shifting" discussed in this paper differs slightly from the traditional approach, as the phase shifting is $\pi /2$ instead of the traditional $2\pi /3$ . We substitute Eq. (4) into Eq. (6) to evaluate the motion error, for a small phase-shift error $x_{i}$ , $\sin (x_i)\approx x_i$ and $\cos (x_i)\approx 1$ , thus we have + +$$ +\begin{array}{l} \epsilon_ {3} (\phi_ {i}) = \tilde {\phi} _ {i} ^ {3} - \phi_ {i} \\ \approx \tan^ {- 1} \left(\frac {\left[ 1 + 0 . 5 \left(x _ {i} - x _ {i + 2}\right) \right] \sin \left(\phi_ {i}\right) + \cos \left(\phi_ {i}\right) x _ {i + 1}}{\cos \left(\phi_ {i}\right) - 0 . 5 \left(x _ {i} + x _ {i + 2}\right) \sin \left(\phi_ {i}\right)}\right) - \tan^ {- 1} \left(\frac {\sin \left(\phi_ {i}\right)}{\cos \left(\phi_ {i}\right)}\right) \\ \approx \frac {x _ {i} + 2 x _ {i + 1} + x _ {i + 2}}{4} + \frac {2 x _ {i + 1} - (x _ {i} + x _ {i + 2})}{4} \cos (2 \phi_ {i}) + \frac {x _ {i} - x _ {i + 2}}{4} \sin (2 \phi_ {i}) \\ = \delta_ {i} ^ {3} + \frac {2 x _ {i + 1} - \left(x _ {i} + x _ {i + 2}\right)}{4} \cos \left(2 \phi_ {i}\right) + \frac {x _ {i} - x _ {i + 2}}{4} \sin \left(2 \phi_ {i}\right), \tag {8} \\ \end{array} +$$ + +![](images/5d076194a6fc92fab7ca3b8df5ce59eda3f9f9880d48d5159e77be8a462c2b2d.jpg) +Fig. 2: Binomial self-compensation for motion error. + +and + +$$ +\begin{array}{l} \epsilon_ {4} \left(\phi_ {i}\right) = \tilde {\phi} _ {i} ^ {4} - \phi_ {i} \\ \begin{array}{l} \approx \tan^ {- 1} \left(\frac {2 \sin (\phi_ {i}) + (x _ {i + 1} + x _ {i + 3}) \cos (\phi_ {i})}{2 \cos (\phi_ {i}) - (x _ {i} + x _ {i + 2}) \sin (\phi_ {i})}\right) - \tan^ {- 1} \left(\frac {\sin (\phi_ {i})}{\cos (\phi_ {i})}\right) \\ x _ {1} + x _ {1 + 2} + x _ {1 + 3} + x _ {1 + 4} = x _ {1} + x _ {1 + 5} = x _ {1 + 6} \end{array} \\ \approx \frac {x _ {i} + x _ {i + 2} + x _ {i + 1} + x _ {i + 3}}{4} + \frac {x _ {i + 1} - x _ {i} + x _ {i + 3} - x _ {i + 2}}{4} \cos \left(2 \phi_ {i}\right) \\ = \delta_ {i} ^ {4} + \frac {x _ {i + 1} - x _ {i} + x _ {i + 3} - x _ {i + 2}}{4} \cos (2 \phi_ {i}) \\ \end{array} +$$ + +According to Eq. (8) and (9), we discover that the motion error comprises a DC component and a harmonic component with twice the frequency of the wrapped phase. The DC component, manifesting as the lag of the point clouds due to object motion, is independent of the phase value, and thus is irrelevant to the ripple-shaped error. Consequently, the crux of motion error compensation lies in eliminating the harmonic. + +Analyzing two successive phase frames $\tilde{\phi}_i$ and $\tilde{\phi}_{i + 1}$ in a discrete time series, we discover that the coefficients of trigonometric terms exhibit opposite signs owing to a phase shift of $\pi /2$ . Consequently, summing each pair of successive phase frames yields higher-order difference terms in the coefficients of harmonics. Repeating this process with the resultant phase frames, as depicted in Fig. 2, naturally evokes the concept of Yang Hui's triangle. + +Let's consider calculating the weighted sum of $K + 1$ successive phase frames in a time series as shown in Fig. 2, using the $(K + 1)$ -th row of Yang Hui's + +![](images/d4c72e424128c7b68823640969f5dc6afc7a7219b594590c488a485ba99b19c6.jpg) +(a) + +![](images/90eeccef51c401369eb3fa93641584a80a8f34fdc5ee5568a551428f6cce4bca.jpg) +(b) +Fig. 3: Motion error converges as $K$ increases: (a) three-step VS four-step phase shifting, and (b) visualization of motion error in four-step phase shifting, the mean absolute error of phase exponentially diminishes as $K$ increases. + +triangle, also as known as $K$ -th order binomial coefficients + +$$ +\tilde {\phi} _ {i} (K) = 2 ^ {- K} \sum_ {k = 0} ^ {K} \binom {K} {k} \tilde {\phi} _ {i + K - k} (0), \tag {10} +$$ + +where $\tilde{\phi}_i(0)$ is computed by Eq. (6). We substitute Eq. (5) and (8) into Eq. (10) to have + +$$ +\begin{array}{l} \tilde {\phi} _ {i} ^ {3} (K) = \bar {\phi} _ {i} ^ {3} (K) - 2 ^ {- (K + 2)} \Delta^ {(K + 2)} x _ {i + 2} \cos (2 \phi_ {i}) \\ + 2 ^ {- (K + 2)} \left(\Delta^ {(K + 1)} x _ {i + 2} + \Delta^ {(K + 1)} x _ {i + 1}\right) \sin \left(2 \phi_ {i}\right), \tag {11} \\ \end{array} +$$ + +where + +$$ +\bar {\phi} _ {i} ^ {3} (K) = 2 ^ {- (K + 2)} \sum_ {k = 0} ^ {K} \left[ \binom {K} {k} \left(\phi_ {i + K - k} + \delta_ {i + K - k} ^ {3}\right) \right]. \tag {12} +$$ + +Similarly, for four-step phase shifting we have + +$$ +\tilde {\phi} _ {i} ^ {4} (K) = \bar {\phi} _ {i} ^ {4} (K) + 2 ^ {- (K + 2)} \left(\Delta^ {(K + 1)} x _ {i} + \Delta^ {(K + 1)} x _ {i + 2}\right) \cos \left(2 \phi_ {i}\right), \tag {13} +$$ + +where + +$$ +\bar {\phi} _ {i} ^ {4} (K) = 2 ^ {- (K + 2)} \sum_ {k = 0} ^ {K} \left[ \binom {K} {k} \left(\phi_ {i + K - k} + \delta_ {i + K - k} ^ {4}\right) \right]. \tag {14} +$$ + +Based on Eq. (11) and (13), it is clear that increasing $K$ produces two main effects: first, the factor $2^{-(K + 2)}$ exponentially decreases the harmonic amplitudes; second, higher-order difference terms (approach zero as $K$ increases) occur in the harmonic amplitudes. Both effects significantly diminish the motion error. + +We measure a flat plate moving towards the camera using our binomial self-compensation (BSC) for motion error, the results visualized in Fig. 3(a) show + +that the ripples induced by motion in the self-compensated phase $\tilde{\phi}_i^4 (K)$ reduces significantly with the increment of $K$ . $K = 0$ represents the raw phase that is affected by motion error. The residual error is caused by the nonlinearity of the binary defocus technique as shown in Fig. 3(b). At least four frames of images are needed in BSC, and using eight images achieves a trade-off between accuracy and efficiency in practice. Note that the binomial order $K$ can be flexibly chosen according to the accuracy and time efficiency requirements. + +In practice, we decompose Eq. (11) into a pyramid-like procedure of adding up the successive two frames layer by layer, as shown in Fig. 2. However, due to the ambiguity of the wrapped phase, phase order error occurs around the phase jumping area. To address this issue, we define the $\oplus$ operator as + +$$ +\phi_ {a} \oplus \phi_ {b} := \left\{ \begin{array}{l} 0. 5 \left(\phi_ {a} + \phi_ {b}\right), \left| \phi_ {a} - \phi_ {b} \right| > \pi \\ 0. 5 \left(\phi_ {a} + \phi_ {b}\right) + \pi , \left| \phi_ {a} - \phi_ {b} \right| < \pi \end{array} . \right. \tag {15} +$$ + +This operation automatically compensates for the phase order error as long as the phase difference is greater than $\pi$ . The motion-error-free phase $\tilde{\phi}_i(K)$ can be obtained by employing our binomial self-compensation (BSC), which is summarized as Algorithm 1, where "mod" represents modular operation. Note that we have corrected the constant phase shifting before adding up two phase frames $\tilde{\phi}_i(0)$ and $\tilde{\phi}_{i + 1}(0)$ . The MATLAB code for implementing the algorithm are given in the github repository. + +# 5 Stereo Phase Unwrapping + +In the previous section, we obtained the motion-error-free high-frequency phase by conducting BSC. Subsequently, we employ a sum of absolute difference algorithm on the wrapped phase to retrieve the correspondence between the main and auxiliary cameras within the disparity range obtained by pre-set depth range. Further, the projector space coordinates $(x^p,y^p)$ can be computed by directly re-projecting the binocular 3D coordinate onto the projector space. Finally, we can have the phase order and compute the unwrapped phase. + +# 6 Experimental Evaluations + +The proposed algorithm is implemented on an Intel(R) Core(TM) i7-13790F @ 2.10 GHz with 32 GB RAM and written in MATLAB. Our experimental system (Fig. 4) consists of two AVT 1800U-120c cameras that have $1280 \times 960$ resolution, equipped with $8\mathrm{mm}$ Computer lenses, and a TI DLP4500 projector with a resolution of $912 \times 1140$ . We cropped the cameras' field of view to $640 \times 480$ to have a 90 fps image acquisition speed. The two cameras are synchronized by the trigger signal from the projector. In order to comprehensively demonstrate the superiority of our BSC, we conducted four groups of experiments on dynamic objects for evaluating: 1) absolute accuracy of 3D reconstruction, 2) temporal resolution of 3D imaging, 3) robustness on depth discontinuous scenes, and 4) generality on different dynamic objects. + +Algorithm 1: BSC for motion error +Input: Captured image sequence: $\{I_0,\dots ,I_{T - 1}\}$ Binomial order: $K$ Phase-shifting step: $N = 3$ or 4. Output: Motion-error-free phase sequence: $\{\bar{\phi}_0(K),\dots,\bar{\phi}_{T - K - N + 1}(K)\}$ for $t\gets 0$ to $T - N$ do $\tilde{\phi}_t(0)\gets \left\{ \begin{array}{ll}\mathrm{mod}\left[\tan^{-1}\left(\frac{2I_{t + 1} - I_t - I_{t + 2}}{I_t - I_{t + 2}}\right) + t\frac{\pi}{2},2\pi \right],N = 3\\ \mathrm{mod}\left[\tan^{-1}\left(\frac{I_{t + 1} - I_{t + 3}}{I_t - I_{t + 2}}\right) + t\frac{\pi}{2},2\pi \right],N = 4 \end{array} \right.$ end for $k\gets 0$ to $K - 1$ do for $i\gets 0$ to $T - k - N$ do $\tilde{\phi}_i(k + 1)\gets \tilde{\phi}_i(k)\oplus \tilde{\phi}_{i + 1}(k)$ end + +![](images/d77049b1c21816005f3775b76d5e4c5836b01ec44720bb4cfe3a798ed4ff3bd9.jpg) +Fig. 4: Experimental setup. + +# 6.1 Absolute Accuracy of 3D Reconstruction + +We measured a periodically waving plate at a distance of $500~\mathrm{mm}$ by using our BSC (8 frames) and five representative methods for comparison: 1) traditional four-step PSP (4 frames), 2) Hilbert transform compensation (HTC, 4 frames) [24], 3) micro Fourier transform profilometry ( $\mu$ -FTP, 4 frames) [34], 4) phase frame difference method (PFD, 8 frames) [10], 5) phase frame sum method (PFS, 5 frmaes) [4]. The fringe wavelength was set to 24 pixels for the cyclic $\pi/2$ phase shifting patterns and was set to [22,24,26] for $\mu$ -FTP. The plate moves within a distance range of [400,500] mm and a speed range of $[-150,150]$ mm/s, and our measurement lasted for 800 frames (about 8.9 seconds). We select three $160\times 120$ rectangular windows in each depth map and fit the quadric surface as the ground truth, then we compute the root mean squared error (RMSE) within the selected regions. The trend of RMSE changing with motion speed is depicted in Fig. 5. The results show that our method exhibits the most satisfactory accuracy among existing methods when dealing with motion error. + +Then we selected a moment when the speed of the plate was $-88~\mathrm{mm / s}$ , and depict the reconstructed point clouds in Fig. 6, our BSC effectively reduces the + +![](images/b71efe0e4f51af127cb00e16143ffa3b98a23ad5f5c79421228a8de4277b823a.jpg) +Fig. 5: Measurement error VS motion speed of a periodically waving plate by traditional four-step phase shifting, HTC [24], $\mu$ -FTP [34], PFD [10], PFS [4], and our BSC, the value in parentheses represents the mean error. Note the y-axis is in logarithmic scale. + +![](images/66bda019f3fc400695d484b3f50cd4c80b4791adfd49caa711b8e5550fac6fe9.jpg) +(a) +(d) + +![](images/347d1287a6cf663eed11807d92a1ed1956249eb8a705ed30227b152ed613c02d.jpg) +(b) + +![](images/39bc771459a7d180cbc8ddbb2c3d8982a192b88b25c84e8508e9e8e9c99c87ad.jpg) +(c) + +![](images/f3457610b00de9c11b86ec3f0c6db120ec8d187411961090490735753595db4b.jpg) +Fig. 6: Measurement results of a moving flat plane by (a) traditional four-step phase shifting, (b) HTC [24], (c) $\mu$ -FTP [34], (d) PFD [10], (e) PFS [4], and (f) our BSC. + +![](images/5703c01ad67352a26821cf11148de1204f7ca35a91610687ccaf0b0693a9d015.jpg) +(e) + +![](images/e51ec7ccd87c780c87d91503f748a423613f8174f01c401409fcf98fff4bec73.jpg) +(f) + +motion error from $324.2\mu \mathrm{m}$ to $54.78\mu \mathrm{m}$ compared with traditional four-step phase shifting. By counting the histogram of errors, it can be observed that the error distribution of our BSC visually presents a normal distribution, which means Gaussian noise (induced by unstable ambient light, camera/projector + +![](images/6107b12975720eca2ae77f4acf1463066f32ce840464695bbff3b353d6cc941e.jpg) + +![](images/38a2c2486c55c25c705cceefbd1d091de49c0babdfa1d5ae950a122651feb0d4.jpg) +(a) + +![](images/d620481d207dcc81790361f60b4c291e2d5b57f4576048d206146b906cee0f65.jpg) +(c) + +![](images/48a0943cc52c726e001be28e78d7ff106498059226715656547e9db5beb52298.jpg) +(b) + +![](images/64ee32d814590d9392324ae4f104a47e7e53f3bf874fa82e6ea6d2b56e72cdfe.jpg) +(d) + +![](images/2a576bdd7eb9d230f71017763154b3b94a1485661f7f790948fc94183e59f01b.jpg) +Fig. 7: Depth discontinuous scene (top left) and motion area labeled in white (bottom left). Reconstruction results and error at edges: (a) traditional four-step phase shifting, (b) HTC [24], (c) $\mu$ -FTP [34], and (d) our BSC. + +![](images/182ac23580f8950548a31da9dac2662a1e105214e08749d4f499af73bc49b8bc.jpg) + +![](images/f033ad971d272589c2ff7d5783fd7c8fedee85824efad92d3a459d05e02e947e.jpg) + +![](images/6c769052cd8291da027c0237b67585ce5e79e780b3a020f771daa8e4e5fdeb62.jpg) + +![](images/694599452683c4cc3f9eb4e44b4b18338aff1fd0796461c36f92ca1b09dbf66c.jpg) + +flicker, quantization error, and sensor noise of the camera and projector) is the main component of error. In other words, the motion error is effectively suppressed by BSC. Meanwhile, the error distribution of other methods deviates from the normal distribution, indicating the existence of residual errors. + +# 6.2 Robustness on Depth Discontinuous Scenes + +PSP's pixel-wise characteristic is one of its most crucial merits. Our BSC maintains this property, thereby exhibiting robustness to depth discontinuous scenes. We set up a scene with three static gypsum geometries and one moving plate. Then we employed our BSC, $\mu$ -FTP, and HTC. The results are shown in Fig. 7. It can be seen that the two non-pixel-wise methods, i.e., $\mu$ -FTP and HTC, exhibit visible deformation to the naked eye in the point cloud around the depth jumping areas. We selected three rectangular windows and quantitatively evaluate the RMSE, the results show that the RMSE of non-pixel-wise methods is significantly greater than our BSC. We can conclude that our BSC with pixel-wise property is more robust compared with non-pixel-wise methods when dealing with depth discontinuous scenes. + +![](images/479f16360ff1712f2274a892ee058779fb4532ad46bc4f665c5d30c450355394.jpg) +Fig. 8: The measurement result of a wooden butterfly with wings flapping, during $\mathrm{T} = 0\mathrm{ms}\sim 88\mathrm{ms}$ , Row 1 and 2: our BSC; Row 3 and 4: $\mu$ -FTP [34]. + +# 6.3 Temporal Resolution of 3D Imaging + +Our BSC employs cyclic $\pi /2$ phase-shifting fringe patterns, any $K + N$ successive captured images can be used for reconstructing one frame of 3D point clouds. Therefore, our resulting frame rate of 3D point clouds is the same as the camera frame rate, thereby achieving a high temporal resolution of the measurement. That is the frame-wise loopable property we mentioned in the introduction. We successively conducted BSC and $\mu$ -FTP to measure a wooden butterfly with wings flapping cyclically, and then manually aligned the timeline, the results are shown in Fig. 8. We noticed that our method can achieve 3D reconstruction with a frame rate of 90 fps (the same as the camera frame rate), while $\mu$ -FTP reconstructing at 30 fps. Thus, our method is capable of distinguishing tinier motions. + +# 6.4 Generality on Different Dynamic Objects + +We captured 400 image frames (about 4 seconds) employing our pipeline on four different dynamic objects, and the results are shown in Fig. 9. The complete motion process video can be found in the supplementary materials. We observed that BSC can effectively eliminate the severe ripple errors caused by object motion. + +# 7 Limitations and Future Outlook + +We presented a pixel-wise and frame-wise loopable binomial self-compensation algorithm to effectively and flexibly eliminate motion error in four-step PSP, without any intermediate variables assisted. + +High-frequency textures: our BSC assumes low-frequency textures like most motion compensation algorithms, it is robust against moderate contrast + +![](images/13fdbcdc871228cc97262fd25b98c3e34fffe36bfba2435a90a42f7b8bde0d93.jpg) + +![](images/6a94379d7f65a90a1da9a8799401de7a68166c2a21d6fed9b24929ea1a67345d.jpg) + +![](images/4cd63007085a631bceb3a584acfba6966e663ad29fa08ef6488dea823dbc8505.jpg) + +![](images/c6d9be6c48f0351dd74b52b6a0df6a67bcb9d255a895cbadc66bbc983125e150.jpg) + +![](images/3a7391dfee698618eafd0c9a667535e24bdd4e36a9e7fbfee1100f7d7c1bfa12.jpg) +Fig. 9: The measurement results of four different dynamic objects. Column 1: wooden butterfly model with wings flapping, Column 2: moving statue, Column 3: waving hand, Column 4: hand showing scissors, stone, and cloth gestures. Top to bottom: our BSC and traditional four-step PSP. + +![](images/c36e7274a64d8b134b877edad16417321dea969a791b931128c159ff5ff30839.jpg) + +![](images/2e501875e5e20f69ab84ab317ea8317095a0e0b2a80eee738284e5c2cf658c04.jpg) + +![](images/55f779acca8e3d1cfe8f205ab58418b4f1d16780e80c821c9e675238ba4dadcc.jpg) + +![](images/655a29d17c5a91ddac4cb54e83f623dd6fccc98be6d55605163145e47d874d40.jpg) +Fig. 10: 3D error VS inter-frame phase-shifting error $|\Delta x_{i}|$ . + +textures. However, when dealing with objects that have high contrast textures, our method induces artifacts similar to those seen with typical phase-shifting methods. + +Phase unwrapping error: we adopt a common block matching algorithm for stereo correspondence, inducing obvious phase unwrapping error when measuring complex surfaces. In the future, we will conduct a hierarchical adjustment for phase outliers to optimize the accuracy of the point cloud. + +Applicable speed range: Our BSC is predicated on the assumption that the phase-shift error $x_{i}$ is small. Consequently, as shown in Fig. 10, the residual motion error arises as $|\Delta x_{i}|$ increases. Properly reducing the fringe frequency could decrease $|\Delta x_{i}|$ , alleviating the performance loss of BSC for high-speed objects. + +# References + +1. Achar, S., Narasimhan, S.G.: Multi focus structured light for recovering scene shape and global illumination. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 205-219. Springer (2014) +2. Chen, T., Seidel, H.P., Lensch, H.P.: Modulated phase-shifting for 3d scanning. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1-8. IEEE (2008) +3. Geng, J.: Structured-light 3d surface imaging: a tutorial. Advances in optics and photonics 3(2), 128-160 (2011) +4. Guo, W., Wu, Z., Zhang, Q., Wang, Y.: Real-time motion-induced error compensation for 4-step phase-shifting profilometry. Optics Express 29(15), 23822-23834 (2021) +5. Gupta, M., Nakhate, N.: A geometric perspective on structured light coding. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 87-102 (2018) +6. Gupta, M., Tian, Y., Narasimhan, S.G., Zhang, L.: A combined theory of defocused illumination and global light transport. International Journal of Computer Vision 98, 146-167 (2012) +7. Gupta, M., Yin, Q., Nayar, S.K.: Structured light in sunlight. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 545-552 (2013) +8. Hirose, K., Kamei, H., Sugiyama, T., Kurosaka, Y.: $200 \times 200 \mu \mathrm{m}$ 2 structured light source. Optics Express 28(25), 37307-37321 (2020) +9. Honrado, C.P., Larrabee Jr, W.F.: Update in three-dimensional imaging in facial plastic surgery. Current opinion in otolaryngology & head and neck surgery 12(4), 327-331 (2004) +0. Liu, X., Tao, T., Wan, Y., Kofman, J.: Real-time motion-induced-error compensation in 3d surface-shape measurement. Optics Express 27(18), 25265-25279 (2019) +1. Liu, Y., Pears, N., Rosin, P.L., Huber, P.: 3D imaging, analysis and applications. Springer (2012) +2. Liu, Z., Zibley, P.C., Zhang, S.: Motion-induced error compensation for phase shifting profilometry. Optics Express 26(10), 12632-12637 (2018) +3. Logozzo, S., Zanetti, E.M., Franceschini, G., Kilpelä, A., Mäkynen, A.: Recent advances in dental optics-part i: 3d intraoral scanners for restorative dentistry. Optics and Lasers in Engineering 54, 203-221 (2014) +4. Lu, L., Ding, Y., Luan, Y., Yin, Y., Liu, Q., Xi, J.: Automated approach for the surface profile measurement of moving objects based on psp. Optics Express 25(25), 32120-32131 (2017) +5. Lu, L., Xi, J., Yu, Y., Guo, Q.: New approach to improve the accuracy of 3-d shape measurement of moving object using phase shifting profilometry. Optics Express 21(25), 30610-30622 (2013) +6. Lu, L., Yin, Y., Su, Z., Ren, X., Luan, Y., Xi, J.: General model for phase shifting profilometry with an object in motion. Applied Optics 57(36), 10364-10369 (2018) +7. Lv, S., Kemao, Q.: Modeling the measurement precision of fringe projection profilometry. Light: Science & Applications 12(1), 257 (2023) +8. Nayar, S.K., Gupta, M.: Diffuse structured light. In: IEEE International Conference on Computational Photography (ICCP). pp. 1-11. IEEE (2012) +9. Posdamer, J.L., Altschuler, M.D.: Surface measurement by space-encoded projected beam systems. Computer Graphics and Image Processing 18(1), 1-17 (1982) + +20. Qian, J., Feng, S., Xu, M., Tao, T., Shang, Y., Chen, Q., Zuo, C.: High-resolution real-time 360 3d surface defect inspection with fringe projection profilometry. Optics and Lasers in Engineering 137, 106382 (2021) +21. Qian, J., Tao, T., Feng, S., Chen, Q., Zuo, C.: Motion-artifact-free dynamic 3d shape measurement with hybrid fourier-transform phase-shifting profilometry. Optics Express 27(3), 2713-2731 (2019) +22. Sansoni, G., Corini, S., Lazzari, S., Rodella, R., Docchio, F.: Three-dimensional imaging based on gray-code light projection: characterization of the measuring algorithm and development of a measuring system for industrial applications. Applied Optics 36(19), 4463-4472 (1997) +23. Srinivasan, V., Liu, H.C., Halioua, M.: Automated phase-measuring profilometry of 3-d diffuse objects. Applied optics 23(18), 3105-3108 (1984) +24. Wang, Y., Liu, Z., Jiang, C., Zhang, S.: Motion induced phase error reduction using a hilbert transform. Optics Express 26(26), 34224-34235 (2018) +25. Wang, Y., Suresh, V., Li, B.: Motion-induced error reduction for binary defocusing profilometry via additional temporal sampling. Optics Express 27(17), 23948-23958 (2019) +26. Weise, T., Leibe, B., Van Gool, L.: Fast 3d scanning with automatic motion compensation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1-8. IEEE (2007) +27. Wu, G., Yang, T., Liu, F., Qian, K.: Suppressing motion-induced phase error by using equal-step phase-shifting algorithms in fringe projection profilometry. Optics Express 30(11), 17980-17998 (2022) +28. Ye, Y., Chang, H., Song, Z., Zhao, J.: Accurate infrared structured light sensing system for dynamic 3d acquisition. Applied Optics 59(17), E80-E88 (2020) +29. Zhang, Y., Lau, D., Wipf, D.: Sparse multi-path corrections in fringe projection profilometry. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13344-13353 (2021) +30. Zhang, Y., Lau, D.L., Yu, Y.: Causes and corrections for bimodal multi-path scanning with structured light. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4431-4439 (2019) +31. Zhong, K., Li, Z., Shi, Y., Wang, C., Lei, Y.: Fast phase measurement profilometry for arbitrary shape objects without phase unwrapping. Optics and Lasers in Engineering 51(11), 1213-1222 (2013) +32. Zhu, J., Yang, F., Hu, J., Zhou, P.: High dynamic reflection surface 3d reconstruction with sharing phase demodulation mechanism and multi-indicators guided phase domain fusion. Optics Express 31(15), 25318-25338 (2023) +33. Zuo, C., Feng, S., Huang, L., Tao, T., Yin, W., Chen, Q.: Phase shifting algorithms for fringe projection profilometry: A review. Optics and lasers in engineering 109, 23-59 (2018) +34. Zuo, C., Tao, T., Feng, S., Huang, L., Asundi, A., Chen, Q.: Micro fourier transform profilometry (μftp): 3d shape measurement at 10,000 frames per second. Optics and Lasers in Engineering 102, 70-91 (2018) \ No newline at end of file diff --git a/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/images.zip b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9fdeceae5a945734b83bf1c39e9eb4b997186c39 --- /dev/null +++ b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82cac7fda4edb6187499c4099a30aeb95a16d3ae27ea067b39231066e9b50696 +size 727900 diff --git a/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/layout.json b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1297c7b129aef4db3293e11ee205f372b189089d --- /dev/null +++ b/2024/Binomial Self-compensation for Motion Error in Dynamic 3D Scanning/layout.json @@ -0,0 +1,9533 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 208, + 167, + 406, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 167, + 406, + 180 + ], + "spans": [ + { + "bbox": [ + 208, + 167, + 406, + 180 + ], + "type": "text", + "content": "Geyou Zhang" + }, + { + "bbox": [ + 208, + 167, + 406, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 208, + 167, + 406, + 180 + ], + "type": "text", + "content": ", Ce Zhu" + }, + { + "bbox": [ + 208, + 167, + 406, + 180 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 208, + 167, + 406, + 180 + ], + "type": "text", + "content": ", and Kai Liu" + }, + { + "bbox": [ + 208, + 167, + 406, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 185, + 189, + 428, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 189, + 428, + 212 + ], + "spans": [ + { + "bbox": [ + 185, + 189, + 428, + 212 + ], + "type": "text", + "content": "1 University of Electronic Science and Technology of China, geyouzhang@foxmail.com, eczhu@uestc.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 264, + 213, + 350, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 213, + 350, + 223 + ], + "spans": [ + { + "bbox": [ + 264, + 213, + 350, + 223 + ], + "type": "text", + "content": "2 Sichuan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 265, + 224, + 347, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 224, + 347, + 233 + ], + "spans": [ + { + "bbox": [ + 265, + 224, + 347, + 233 + ], + "type": "text", + "content": "kailiu@scu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 260, + 455, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 260, + 455, + 447 + ], + "spans": [ + { + "bbox": [ + 159, + 260, + 455, + 447 + ], + "type": "text", + "content": "Abstract. Phase shifting profilometry (PSP) is favored in high-precision 3D scanning due to its high accuracy, robustness, and pixel-wise property. However, a fundamental assumption of PSP that the object should remain static is violated in dynamic measurement, making PSP susceptible to object moving, resulting in ripple-like errors in the point clouds. We propose a pixel-wise and frame-wise loopable binomial self-compensation (BSC) algorithm to effectively and flexibly eliminate motion error in the four-step PSP. Our mathematical model demonstrates that by summing successive motion-affected phase frames weighted by binomial coefficients, motion error exponentially diminishes as the binomial order increases, accomplishing automatic error compensation through the motion-affected phase sequence, without the assistance of any intermediate variable. Extensive experiments show that our BSC outperforms the existing methods in reducing motion error, while achieving a depth map frame rate equal to the camera's acquisition rate (90 fps), enabling high-accuracy 3D reconstruction with a quasi-single-shot frame rate. The code is available at https://github.com/GeyouZhang/BSC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "spans": [ + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "text", + "content": "Keywords: Phase Shifting Profilometry " + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "text", + "content": " Dynamic 3D Scanning " + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "text", + "content": " Motion Error Compensation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 498, + 230, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 230, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 230, + 511 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "In recent years, structured light (SL) 3D imaging has been penetrated into diverse fields, e.g., industrial inspection [20] and design [32], medical 3D reconstruction [3,8,13,17], virtual reality [28], and digitalization of cultural heritage [9]. Phase shifting profilometry (PSP) [23], Fourier transform profilometry, binary code [19], and Gray code [22] are famous SL technologies that are widely applied in industry and scientific researches. Among the SL technologies above, PSP is especially favored in high-precision 3D scanning due to its high accuracy, robustness, and pixel-wise property [33]. Many studies have excavated the huge potentiality of PSP in various challenging scenes, such as low signal-to-noise [5,7], multi-path [29,30], high reflectivity [18], and global illumination [1,2,6], et. al." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "However, a fundamental assumption of PSP that the object should remain static is violated in dynamic measurement, making PSP susceptible to object" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "moving, resulting in ripple-like errors in the point clouds. We observed that the solutions of PSP in dynamic scenes are overlooked within the computer vision community. Concurrently, there are some existing methods in the optics metrology community to address this challenge, most of them can be categorized into three types: 1) motion-induced phase shifting estimation, 2) object tracking, and 3) automatic compensation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 483, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 483, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 483, + 298 + ], + "type": "text", + "content": "Motion-induced phase shifting estimation: In traditional PSP, the phase shifting of each pattern is equally distributed or pre-known values within one period of fringe, which becomes disorderly due to object motion. Thus, researchers estimate the motion-induced phase shifting according to the implicit motion information in the motion-affected phase frame [26] or frames [10, 12, 21, 25], thereby restoring the unknown motion-affected phase shifting value of each pattern, and finally solve the corrected phase by least squared method [11]. However, these methods may not effectively compensate for errors when the motion information can't be precisely predicted." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 372 + ], + "type": "text", + "content": "Object tracking: By tracking the position of the object by placing markers [15] or using image feature matching algorithm [16], the motion error can be corrected by identifying the trajectory of the object [14-16]. However, a limitation of the object tracking based methods is that they are tailored for 2D motion perpendicular to the sight line. Moreover, the accuracy and reliability of feature tracking are crucial issues when faced with textureless objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "text", + "content": "Automatic compensation: In automatic compensation methods, instead of computing motion-induced phase shifting value or tracking the object, an appropriate operator is selected to conduct on the whole captured image [24, 27] or phase frame [4], and further directly obtain the motion-error-free phase, in which the motion error will be automatically compensated." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 436, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 567 + ], + "type": "text", + "content": "Previous work on motion compensation typically estimates or generates certain intermediate variables, then calculates the motion-error-free phase based on these variables, either directly or indirectly. For example, in phase shifting estimation methods, the intermediate variables are motion-induced phase shifting values [10, 12, 21, 25]; in object tracking methods, the intermediate variable is the rigid transform matrix [14-16]; in automatic compensation methods, the intermediate variables are the captured images after undergoing Hilbert transform [24] or differential operator [27]. Therefore, two questions naturally arise: Why not utilize the motion-affected phase itself to compensate for motion error without depending on any intermediate variables? Would this approach be more effective and flexible?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "In this paper, without depending on any extra intermediate variable, we develop a binomial self-compensation (BSC) method that only utilizes the motion-affected phase sequence itself, thereby effectively and flexibly eliminating the motion error in four-step PSP. To achieve this, first, we design a paraxial binocular structured light system for reconstructing 3D point clouds using a single set of high-frequency fringe patterns. This system intentionally sets a very short baseline between two cameras to achieve a narrow disparity range, facilitating the use of high frequency fringes, while ensuring the baseline between the main" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "camera and projector is long enough for adequate accuracy. Subsequently, we cyclically [31] project high-frequency patterns with " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " phase shifting, and directly compute a sequence of phase frames from each three or four successive captured images. Finally, by adding up a sequence of successive phase frames weighted by binomial coefficients, we achieve self-compensation of motion errors. Our mathematical model demonstrates that motion error exponentially diminishes as the binomial order " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 271 + ], + "type": "text", + "content": "Our BSC not only inherits the pixel-wise advantages of PSP, but also is of high temporal resolution to achieve quasi-single-shot 3D imaging frame rate because it's frame-wise loopable. Furthermore, our method offers flexibility and ease of deployment, acting as an enhanced, motion-error-free version of the four-step PSP, thus providing a plug-and-play enhancement for dynamic 3D scanning scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 292, + 237, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 292, + 237, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 237, + 304 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 318, + 479, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 479, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 479, + 425 + ], + "type": "text", + "content": "In the introduction section, based on specific implementation details, we categorize the motion error compensation methods into three classes. From another perspective, we can also distinguish them based on their 3D reconstruction performance on temporal and spatial resolution. We examine the motion compensation methods from two perspectives: property of pixel-wise and frame-wise loopable. Pixel-wise methods compute phase and point clouds pixel by pixel, achieving spatial independence through this detailed granularity. Frame-wise loopable methods accomplish the same depth map frame rate as the camera acquisition frame rate, enabling a quasi-single-shot 3D reconstruction frame rate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 604 + ], + "type": "text", + "content": "Pixel-wise: Several motion error compensation methods preserve the pixelwise precision of traditional PSP, which utilizing two sampling phase frames to iteratively compute motion phase shifting [12, 25], automatically compensate for errors [4], tracking the position of the object by placing markers [15], using image feature matching algorithm [16] and correct motion error by identifying the trajectory of the object [14-16]. These approaches build on the granularity of pixel-level adjustments. In contrast, the non-pixel-wise methods conduct particular operator on local [10, 26, 27] or global image [21, 24, 34], e.g., Fourier transform [21, 34], Hilbert transform [24], differential operator [27], mean filter [10], optimization algorithm [26], from which estimate the extra phase shifting induced by motion and compensate error [10, 21] or directly correct motion error [24, 26, 27, 34]. However, the pixel-wise merit of PSP is not preserved in these methods, theoretically diminishing their robustness in handling depth discontinuities. Concurrently, introducing extra non-pixel-wise operators also escalates the computational time required." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "Frame-wise loopable: By projecting high-frequency fringes and reconstructing 3D point clouds using a cyclic strategy [31], some motion error compensation methods [4, 10, 24, 26, 27] accomplish 3D reconstruction with quasi-single-shot frame rate, thus called frame-wise loopable methods in this paper. The cyclic strategy is not suitable for all motion compensation methods, some" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 447, + 103 + ], + "type": "text", + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 478, + 257 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 478, + 257 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 478, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 478, + 257 + ], + "type": "image", + "image_path": "ddaeef3a1d8f7f61d6ee8667b6953f57330d4b708b4880b65cd22b8c67a6ecb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 264, + 442, + 275 + ], + "lines": [ + { + "bbox": [ + 170, + 264, + 442, + 275 + ], + "spans": [ + { + "bbox": [ + 170, + 264, + 442, + 275 + ], + "type": "text", + "content": "Fig. 1: Diagram of our paraxial binocular structured light system." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 299, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 480, + 323 + ], + "type": "text", + "content": "methods need to project an indivisible group of patterns [21, 25, 34], such that the depth map frame rate is reduced severalfold." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 342, + 384, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 384, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 384, + 355 + ], + "type": "text", + "content": "3 System Setup and Frequency Selection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 480, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 450 + ], + "type": "text", + "content": "Our paraxial binocular structured light system comprises a main camera, an auxiliary camera, and a projector, as depicted in Fig. 1. Together, the main and auxiliary cameras form a paraxial setup to provide reference depth for phase unwrapping. Simultaneously, the combination of the main camera and projector facilitates precise 3D reconstruction. Following system calibration, we perform stereo rectification on the captured images to align the epipolar lines in the auxiliary system parallelly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 521 + ], + "type": "text", + "content": "To ensure the simplicity of equations and convey intuitive geometrical relations, the following analysis in this section is under the assumption that the optical centers of the projector, main camera, and auxiliary camera are coplanar. Note that we build the world coordinate system to coincide with the main camera coordinate system. In reality, the projector may not be strictly coplanar to the other two, but the conclusions of our analysis are still effective." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 605 + ], + "type": "text", + "content": "The primary benefit of the binocular structured light (SL) system lies in its ability to perform stereo phase unwrapping (SPU) through geometric constraints, substantially decreasing the needed number of fringe patterns. To avoid phase unwrapping ambiguities, it should be ensured that only a few candidate points are within the depth range. As a result, traditional SPU approaches typically necessitate a short baseline between the main camera and the projector, which unfortunately compromises the accuracy of point clouds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "content": "To address this, in this paper, we shift the constraint from the camera projector baseline to the camera-camera baseline, thus allowing flexibility in the camera-projector distance and enhancing point cloud precision. This camera-camera baseline constraint is called the uniqueness constraint, ensuring that the phase remains single-valued within the feasible stereo matching range on the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "auxiliary camera, from which guarantees the uniqueness of the matching point candidate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 221 + ], + "type": "text", + "content": "Given that stereo rectification has been applied to the cameras, we can disregard the Y axis. Assuming the 3D surface near the measured point is continuous and orthogonal to the Z axis, determining the depth range allows us to calculate the maximum fringe frequency. Consider a virtual segment, represented by a black dashed line in Fig. 1, cut by the auxiliary camera's feasible region cone. The length of this virtual segment (black solid line in Fig. 1) can be calculated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 256, + 221, + 480, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 221, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 256, + 221, + 480, + 246 + ], + "type": "interline_equation", + "content": "d = \\frac {\\left(Z _ {\\operatorname* {m a x}} - Z _ {\\operatorname* {m i n}}\\right) Z L}{Z _ {\\operatorname* {m a x}} Z _ {\\operatorname* {m i n}}}. \\tag {1}", + "image_path": "a4e4b54285a04571d3105d088df2f9ed8d07ca42a10e00979b6fa607f26697c9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 247, + 479, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 247, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 479, + 271 + ], + "type": "text", + "content": "As long as the projection of the virtual segment onto the projector space is distributed within one fringe period, the uniqueness constraint is satisfied, i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 274, + 276, + 480, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 276, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 274, + 276, + 480, + 301 + ], + "type": "interline_equation", + "content": "\\frac {f _ {x} ^ {p} d}{Z - Z ^ {p}} < \\frac {W ^ {p}}{f}, \\tag {2}", + "image_path": "ba84e5ea192fe3786197bd7599b145302e65bb0be1bae991afc467a621c3b070.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "inline_equation", + "content": "f_{x}^{p}" + }, + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "text", + "content": " is the focal length of the projector, " + }, + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "inline_equation", + "content": "W^{p}" + }, + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "text", + "content": " is the resolution of the projector along " + }, + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "text", + "content": " axis, " + }, + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 305, + 479, + 329 + ], + "type": "text", + "content": " is the frequency of fringe pattern." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 479, + 353 + ], + "type": "text", + "content": "Finally, according to uniqueness constraint, we can obtain the limiting frequency of fringe as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 230, + 359, + 480, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 359, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 230, + 359, + 480, + 384 + ], + "type": "interline_equation", + "content": "f < \\left(1 - \\frac {Z ^ {p}}{Z}\\right) \\frac {Z _ {\\operatorname* {m i n}} Z _ {\\operatorname* {m a x}} W ^ {p}}{\\left(Z _ {\\operatorname* {m a x}} - Z _ {\\operatorname* {m i n}}\\right) L f _ {x} ^ {p}}. \\tag {3}", + "image_path": "e4c73d44292af2a939e0ab3dde0df96001800a93d3c06364fb3ec321d8837929.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 389, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 483 + ], + "type": "text", + "content": "Actually, the uniqueness constraint is derived from an aggressive postulate that the 3D surface around the measured point is continuous and orthogonal with the Z axis, so the uniqueness constraint is a weak constraint. Thus, the camera-camera baseline should be as short as possible to ensure the uniqueness of phase value in the feasible range. In our prototype, we simply placed the main and auxiliary cameras tightly against each other. In the future, by using a beam splitter, the camera baseline can be further shortened, breaking through the physical size limitations of industrial cameras." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 485, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 480, + 533 + ], + "type": "text", + "content": "In summary, utilizing the proposed paraxial binocular SL system enables the achievement of unambiguous 3D reconstruction within the specified depth range, using only a single set of high-frequency fringe patterns. As a result, we can employ the cyclic projection strategy [31] for dynamic 3D scanning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 548, + 422, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 422, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 422, + 562 + ], + "type": "text", + "content": "4 Binomial Self-compensation for Motion Error" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 571, + 479, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 479, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 479, + 642 + ], + "type": "text", + "content": "In this paper, we cyclically project a group of high-frequency fringe patterns with a phase shift of " + }, + { + "bbox": [ + 130, + 571, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 571, + 479, + 642 + ], + "type": "text", + "content": ". The 3D point corresponding to a camera pixel " + }, + { + "bbox": [ + 130, + 571, + 479, + 642 + ], + "type": "inline_equation", + "content": "(x^{c},y^{c})" + }, + { + "bbox": [ + 130, + 571, + 479, + 642 + ], + "type": "text", + "content": " is the intersection point of the sight line and the object surface. The intersection point moves along with the sight line on the circumstance that measures moving objects. Assuming the object is of low-frequency texture, the " + }, + { + "bbox": [ + 130, + 571, + 479, + 642 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 571, + 479, + 642 + ], + "type": "text", + "content": " th captured frame can be expressed as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 217, + 648, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 648, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 217, + 648, + 480, + 669 + ], + "type": "interline_equation", + "content": "I _ {i} \\left(x ^ {c}, y ^ {c}\\right) = A ^ {c} + B ^ {c} \\cos \\left(\\phi_ {0} - i \\frac {\\pi}{2} + x _ {i}\\right), \\tag {4}", + "image_path": "14451410adf2f040c8d7951c1a597cafaa68cd5e9cf3439517175f8e79cecae0.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "text", + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "A^c" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is the background intensity, " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "B^c" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is the modulation, " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " represents the unknown phase offset induced by the motion from the datum frame to " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "-th frame, note that the datum frame can be arbitrarily chosen, and the true phase value at " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " th frame is " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\phi_i = \\phi_0 - i\\frac{\\pi}{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". We choose a " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " phase shifting as it only transforms " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\sin (\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " terms into each other without inducing extra phase offset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 179, + 482, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 482, + 215 + ], + "type": "text", + "content": "We notice that " + }, + { + "bbox": [ + 130, + 179, + 482, + 215 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 179, + 482, + 215 + ], + "type": "text", + "content": " and its high order difference exactly correspond to the concepts of displacement, velocity, and acceleration, et. al, in kinematics, and can be expressed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 231, + 481, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 231, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 225, + 231, + 481, + 281 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\Delta^ {(K)} x _ {i} = \\Delta^ {(K - 1)} x _ {i + 1} - \\Delta^ {(K - 1)} x _ {i} \\\\ \\Delta^ {(K)} x _ {i} = \\sum_ {k = 0} ^ {K} (- 1) ^ {k} \\binom {K} {k} x _ {i + K - k}, \\end{array} \\right. \\tag {5}", + "image_path": "2dc02ca2b685356b9509ba81b96d055d737f5516510cc859baeb7134ab3cf0b9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 481, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 481, + 321 + ], + "type": "text", + "content": "where we define " + }, + { + "bbox": [ + 130, + 297, + 481, + 321 + ], + "type": "inline_equation", + "content": "\\varDelta^{(0)}x_{i} = x_{i}" + }, + { + "bbox": [ + 130, + 297, + 481, + 321 + ], + "type": "text", + "content": ". Eq. (5) is the key to motion error compensation, and we will use it later." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": "Utilizing a cyclic projection strategy [31] for high speed 3D scanning, a three or four-step phase shifting algorithm is typically employed to calculate the motion-affected phase as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 377, + 481, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 377, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 211, + 377, + 481, + 410 + ], + "type": "interline_equation", + "content": "\\tilde {\\phi} _ {i} ^ {3} = \\tan^ {- 1} \\left(\\frac {\\tilde {S} _ {i} ^ {3}}{\\tilde {C} _ {i} ^ {3}}\\right) \\text {a n d} \\tilde {\\phi} _ {i} ^ {4} = \\tan^ {- 1} \\left(\\frac {\\tilde {S} _ {i} ^ {4}}{\\tilde {C} _ {i} ^ {4}}\\right), \\tag {6}", + "image_path": "9bad58750e6bf74d1253bc59beeb24be2bf0ee8c0f51499d6b8d83dd22e2683b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 424, + 161, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 161, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 161, + 434 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 197, + 443, + 481, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 443, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 197, + 443, + 481, + 475 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} S _ {i} ^ {3} = 2 I _ {i + 1} - I _ {i} - I _ {i + 2} \\\\ C _ {i} ^ {3} = I _ {i} - I _ {i + 2} \\end{array} \\right. \\text {a n d} \\left\\{ \\begin{array}{l} S _ {i} ^ {4} = I _ {i + 1} - I _ {i + 3} \\\\ C _ {i} ^ {4} = I _ {i} - I _ {i + 2} \\end{array} \\right.. \\tag {7}", + "image_path": "ee0e697de3acb162e70959b7a0f066cee6630d9076ef4cd0a64e2f0aaa2ac3ed.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "text", + "content": "Please note that the \"three-step phase shifting\" discussed in this paper differs slightly from the traditional approach, as the phase shifting is " + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "text", + "content": " instead of the traditional " + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "inline_equation", + "content": "2\\pi /3" + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "text", + "content": ". We substitute Eq. (4) into Eq. (6) to evaluate the motion error, for a small phase-shift error " + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\sin (x_i)\\approx x_i" + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\cos (x_i)\\approx 1" + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "text", + "content": ", thus we have" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 563, + 488, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 563, + 488, + 667 + ], + "spans": [ + { + "bbox": [ + 133, + 563, + 488, + 667 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\epsilon_ {3} (\\phi_ {i}) = \\tilde {\\phi} _ {i} ^ {3} - \\phi_ {i} \\\\ \\approx \\tan^ {- 1} \\left(\\frac {\\left[ 1 + 0 . 5 \\left(x _ {i} - x _ {i + 2}\\right) \\right] \\sin \\left(\\phi_ {i}\\right) + \\cos \\left(\\phi_ {i}\\right) x _ {i + 1}}{\\cos \\left(\\phi_ {i}\\right) - 0 . 5 \\left(x _ {i} + x _ {i + 2}\\right) \\sin \\left(\\phi_ {i}\\right)}\\right) - \\tan^ {- 1} \\left(\\frac {\\sin \\left(\\phi_ {i}\\right)}{\\cos \\left(\\phi_ {i}\\right)}\\right) \\\\ \\approx \\frac {x _ {i} + 2 x _ {i + 1} + x _ {i + 2}}{4} + \\frac {2 x _ {i + 1} - (x _ {i} + x _ {i + 2})}{4} \\cos (2 \\phi_ {i}) + \\frac {x _ {i} - x _ {i + 2}}{4} \\sin (2 \\phi_ {i}) \\\\ = \\delta_ {i} ^ {3} + \\frac {2 x _ {i + 1} - \\left(x _ {i} + x _ {i + 2}\\right)}{4} \\cos \\left(2 \\phi_ {i}\\right) + \\frac {x _ {i} - x _ {i + 2}}{4} \\sin \\left(2 \\phi_ {i}\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "289d700b63d4c9d905af35d4c5d08bfc44e1c9663922df154a484eacc57a5e54.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 112, + 471, + 310 + ], + "blocks": [ + { + "bbox": [ + 143, + 112, + 471, + 310 + ], + "lines": [ + { + "bbox": [ + 143, + 112, + 471, + 310 + ], + "spans": [ + { + "bbox": [ + 143, + 112, + 471, + 310 + ], + "type": "image", + "image_path": "5d076194a6fc92fab7ca3b8df5ce59eda3f9f9880d48d5159e77be8a462c2b2d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 318, + 414, + 330 + ], + "lines": [ + { + "bbox": [ + 198, + 318, + 414, + 330 + ], + "spans": [ + { + "bbox": [ + 198, + 318, + 414, + 330 + ], + "type": "text", + "content": "Fig. 2: Binomial self-compensation for motion error." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 360, + 152, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 152, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 152, + 370 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 384, + 463, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 384, + 463, + 474 + ], + "spans": [ + { + "bbox": [ + 141, + 384, + 463, + 474 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\epsilon_ {4} \\left(\\phi_ {i}\\right) = \\tilde {\\phi} _ {i} ^ {4} - \\phi_ {i} \\\\ \\begin{array}{l} \\approx \\tan^ {- 1} \\left(\\frac {2 \\sin (\\phi_ {i}) + (x _ {i + 1} + x _ {i + 3}) \\cos (\\phi_ {i})}{2 \\cos (\\phi_ {i}) - (x _ {i} + x _ {i + 2}) \\sin (\\phi_ {i})}\\right) - \\tan^ {- 1} \\left(\\frac {\\sin (\\phi_ {i})}{\\cos (\\phi_ {i})}\\right) \\\\ x _ {1} + x _ {1 + 2} + x _ {1 + 3} + x _ {1 + 4} = x _ {1} + x _ {1 + 5} = x _ {1 + 6} \\end{array} \\\\ \\approx \\frac {x _ {i} + x _ {i + 2} + x _ {i + 1} + x _ {i + 3}}{4} + \\frac {x _ {i + 1} - x _ {i} + x _ {i + 3} - x _ {i + 2}}{4} \\cos \\left(2 \\phi_ {i}\\right) \\\\ = \\delta_ {i} ^ {4} + \\frac {x _ {i + 1} - x _ {i} + x _ {i + 3} - x _ {i + 2}}{4} \\cos (2 \\phi_ {i}) \\\\ \\end{array}", + "image_path": "2ab159c1f68d11e4f8fcaeb5669ba7cdc57d356f98d80c802ef452ef0aed5bc7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "type": "text", + "content": "According to Eq. (8) and (9), we discover that the motion error comprises a DC component and a harmonic component with twice the frequency of the wrapped phase. The DC component, manifesting as the lag of the point clouds due to object motion, is independent of the phase value, and thus is irrelevant to the ripple-shaped error. Consequently, the crux of motion error compensation lies in eliminating the harmonic." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 566, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 639 + ], + "type": "text", + "content": "Analyzing two successive phase frames " + }, + { + "bbox": [ + 130, + 566, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_i" + }, + { + "bbox": [ + 130, + 566, + 482, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 566, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_{i + 1}" + }, + { + "bbox": [ + 130, + 566, + 482, + 639 + ], + "type": "text", + "content": " in a discrete time series, we discover that the coefficients of trigonometric terms exhibit opposite signs owing to a phase shift of " + }, + { + "bbox": [ + 130, + 566, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 566, + 482, + 639 + ], + "type": "text", + "content": ". Consequently, summing each pair of successive phase frames yields higher-order difference terms in the coefficients of harmonics. Repeating this process with the resultant phase frames, as depicted in Fig. 2, naturally evokes the concept of Yang Hui's triangle." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Let's consider calculating the weighted sum of " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " successive phase frames in a time series as shown in Fig. 2, using the " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "-th row of Yang Hui's" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "text", + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 303, + 241 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 303, + 241 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 303, + 241 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 303, + 241 + ], + "type": "image", + "image_path": "d4c72e424128c7b68823640969f5dc6afc7a7219b594590c488a485ba99b19c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 243, + 223, + 252 + ], + "lines": [ + { + "bbox": [ + 211, + 243, + 223, + 252 + ], + "spans": [ + { + "bbox": [ + 211, + 243, + 223, + 252 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 114, + 482, + 242 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 482, + 242 + ], + "type": "image", + "image_path": "90eeccef51c401369eb3fa93641584a80a8f34fdc5ee5568a551428f6cce4bca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 243, + 397, + 252 + ], + "lines": [ + { + "bbox": [ + 384, + 243, + 397, + 252 + ], + "spans": [ + { + "bbox": [ + 384, + 243, + 397, + 252 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "text", + "content": "Fig. 3: Motion error converges as " + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "text", + "content": " increases: (a) three-step VS four-step phase shifting, and (b) visualization of motion error in four-step phase shifting, the mean absolute error of phase exponentially diminishes as " + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 318, + 390, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 318, + 390, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 318, + 390, + 329 + ], + "type": "text", + "content": "triangle, also as known as " + }, + { + "bbox": [ + 131, + 318, + 390, + 329 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 318, + 390, + 329 + ], + "type": "text", + "content": "-th order binomial coefficients" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 226, + 337, + 480, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 337, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 226, + 337, + 480, + 371 + ], + "type": "interline_equation", + "content": "\\tilde {\\phi} _ {i} (K) = 2 ^ {- K} \\sum_ {k = 0} ^ {K} \\binom {K} {k} \\tilde {\\phi} _ {i + K - k} (0), \\tag {10}", + "image_path": "79b50e143bdbb7c71b74f13cefb131be63f99735774ee3be92b91c37d0380e6e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 479, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 380, + 479, + 403 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_i(0)" + }, + { + "bbox": [ + 130, + 380, + 479, + 403 + ], + "type": "text", + "content": " is computed by Eq. (6). We substitute Eq. (5) and (8) into Eq. (10) to have" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 182, + 409, + 480, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 409, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 182, + 409, + 480, + 447 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\phi} _ {i} ^ {3} (K) = \\bar {\\phi} _ {i} ^ {3} (K) - 2 ^ {- (K + 2)} \\Delta^ {(K + 2)} x _ {i + 2} \\cos (2 \\phi_ {i}) \\\\ + 2 ^ {- (K + 2)} \\left(\\Delta^ {(K + 1)} x _ {i + 2} + \\Delta^ {(K + 1)} x _ {i + 1}\\right) \\sin \\left(2 \\phi_ {i}\\right), \\tag {11} \\\\ \\end{array}", + "image_path": "054f636e414e4e0c1afeb00e5d98d28a502aa200388458f3802327ca888c13f1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 455, + 161, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 161, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 161, + 464 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 193, + 463, + 480, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 463, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 193, + 463, + 480, + 496 + ], + "type": "interline_equation", + "content": "\\bar {\\phi} _ {i} ^ {3} (K) = 2 ^ {- (K + 2)} \\sum_ {k = 0} ^ {K} \\left[ \\binom {K} {k} \\left(\\phi_ {i + K - k} + \\delta_ {i + K - k} ^ {3}\\right) \\right]. \\tag {12}", + "image_path": "8a4195d62b60ebfb45e78248f9dfb3636de68f08f47fd3b102028748ea095175.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 500, + 334, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 334, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 334, + 512 + ], + "type": "text", + "content": "Similarly, for four-step phase shifting we have" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 156, + 520, + 480, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 520, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 156, + 520, + 480, + 540 + ], + "type": "interline_equation", + "content": "\\tilde {\\phi} _ {i} ^ {4} (K) = \\bar {\\phi} _ {i} ^ {4} (K) + 2 ^ {- (K + 2)} \\left(\\Delta^ {(K + 1)} x _ {i} + \\Delta^ {(K + 1)} x _ {i + 2}\\right) \\cos \\left(2 \\phi_ {i}\\right), \\tag {13}", + "image_path": "84a3c41b21cea732d0425c7bfde2195c1d69ae46fad71c3a59f80c22fc7b2eb0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 548, + 161, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 161, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 161, + 558 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 193, + 556, + 480, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 556, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 193, + 556, + 480, + 590 + ], + "type": "interline_equation", + "content": "\\bar {\\phi} _ {i} ^ {4} (K) = 2 ^ {- (K + 2)} \\sum_ {k = 0} ^ {K} \\left[ \\binom {K} {k} \\left(\\phi_ {i + K - k} + \\delta_ {i + K - k} ^ {4}\\right) \\right]. \\tag {14}", + "image_path": "1cc77e202a8d81488d8b9341ee563f8967ee2efe8500e6eefc10e257b5e92bbb.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "text", + "content": "Based on Eq. (11) and (13), it is clear that increasing " + }, + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "text", + "content": " produces two main effects: first, the factor " + }, + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "inline_equation", + "content": "2^{-(K + 2)}" + }, + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "text", + "content": " exponentially decreases the harmonic amplitudes; second, higher-order difference terms (approach zero as " + }, + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "text", + "content": " increases) occur in the harmonic amplitudes. Both effects significantly diminish the motion error." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 642, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 479, + 665 + ], + "type": "text", + "content": "We measure a flat plate moving towards the camera using our binomial self-compensation (BSC) for motion error, the results visualized in Fig. 3(a) show" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": "that the ripples induced by motion in the self-compensated phase " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_i^4 (K)" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " reduces significantly with the increment of " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "K = 0" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " represents the raw phase that is affected by motion error. The residual error is caused by the nonlinearity of the binary defocus technique as shown in Fig. 3(b). At least four frames of images are needed in BSC, and using eight images achieves a trade-off between accuracy and efficiency in practice. Note that the binomial order " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " can be flexibly chosen according to the accuracy and time efficiency requirements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "text", + "content": "In practice, we decompose Eq. (11) into a pyramid-like procedure of adding up the successive two frames layer by layer, as shown in Fig. 2. However, due to the ambiguity of the wrapped phase, phase order error occurs around the phase jumping area. To address this issue, we define the " + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "text", + "content": " operator as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 207, + 255, + 480, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 255, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 207, + 255, + 480, + 280 + ], + "type": "interline_equation", + "content": "\\phi_ {a} \\oplus \\phi_ {b} := \\left\\{ \\begin{array}{l} 0. 5 \\left(\\phi_ {a} + \\phi_ {b}\\right), \\left| \\phi_ {a} - \\phi_ {b} \\right| > \\pi \\\\ 0. 5 \\left(\\phi_ {a} + \\phi_ {b}\\right) + \\pi , \\left| \\phi_ {a} - \\phi_ {b} \\right| < \\pi \\end{array} . \\right. \\tag {15}", + "image_path": "48a55bf2289dd3dd70af20ebf3b3713cc715eca58d4d6f9119bb3c42266e187a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "text", + "content": "This operation automatically compensates for the phase order error as long as the phase difference is greater than " + }, + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "text", + "content": ". The motion-error-free phase " + }, + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_i(K)" + }, + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "text", + "content": " can be obtained by employing our binomial self-compensation (BSC), which is summarized as Algorithm 1, where \"mod\" represents modular operation. Note that we have corrected the constant phase shifting before adding up two phase frames " + }, + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_i(0)" + }, + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_{i + 1}(0)" + }, + { + "bbox": [ + 131, + 286, + 482, + 370 + ], + "type": "text", + "content": ". The MATLAB code for implementing the algorithm are given in the github repository." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 386, + 308, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 308, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 308, + 401 + ], + "type": "text", + "content": "5 Stereo Phase Unwrapping" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 410, + 479, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 479, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 479, + 494 + ], + "type": "text", + "content": "In the previous section, we obtained the motion-error-free high-frequency phase by conducting BSC. Subsequently, we employ a sum of absolute difference algorithm on the wrapped phase to retrieve the correspondence between the main and auxiliary cameras within the disparity range obtained by pre-set depth range. Further, the projector space coordinates " + }, + { + "bbox": [ + 130, + 410, + 479, + 494 + ], + "type": "inline_equation", + "content": "(x^p,y^p)" + }, + { + "bbox": [ + 130, + 410, + 479, + 494 + ], + "type": "text", + "content": " can be computed by directly re-projecting the binocular 3D coordinate onto the projector space. Finally, we can have the phase order and compute the unwrapped phase." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 510, + 307, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 307, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 307, + 525 + ], + "type": "text", + "content": "6 Experimental Evaluations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "The proposed algorithm is implemented on an Intel(R) Core(TM) i7-13790F @ 2.10 GHz with 32 GB RAM and written in MATLAB. Our experimental system (Fig. 4) consists of two AVT 1800U-120c cameras that have " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "1280 \\times 960" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " resolution, equipped with " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "8\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " Computer lenses, and a TI DLP4500 projector with a resolution of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "912 \\times 1140" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". We cropped the cameras' field of view to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " to have a 90 fps image acquisition speed. The two cameras are synchronized by the trigger signal from the projector. In order to comprehensively demonstrate the superiority of our BSC, we conducted four groups of experiments on dynamic objects for evaluating: 1) absolute accuracy of 3D reconstruction, 2) temporal resolution of 3D imaging, 3) robustness on depth discontinuous scenes, and 4) generality on different dynamic objects." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 158, + 91, + 448, + 103 + ], + "type": "text", + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 138, + 132, + 404, + 312 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 304, + 130 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 304, + 130 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 304, + 130 + ], + "type": "text", + "content": "Algorithm 1: BSC for motion error" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "lines": [ + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": "Input: Captured image sequence: " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "\\{I_0,\\dots ,I_{T - 1}\\}" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " Binomial order: " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " Phase-shifting step: " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " or 4. Output: Motion-error-free phase sequence: " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\phi}_0(K),\\dots,\\bar{\\phi}_{T - K - N + 1}(K)\\}" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "t\\gets 0" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "T - N" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_t(0)\\gets \\left\\{ \\begin{array}{ll}\\mathrm{mod}\\left[\\tan^{-1}\\left(\\frac{2I_{t + 1} - I_t - I_{t + 2}}{I_t - I_{t + 2}}\\right) + t\\frac{\\pi}{2},2\\pi \\right],N = 3\\\\ \\mathrm{mod}\\left[\\tan^{-1}\\left(\\frac{I_{t + 1} - I_{t + 3}}{I_t - I_{t + 2}}\\right) + t\\frac{\\pi}{2},2\\pi \\right],N = 4 \\end{array} \\right." + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " end for " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "k\\gets 0" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " do for " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "i\\gets 0" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "T - k - N" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_i(k + 1)\\gets \\tilde{\\phi}_i(k)\\oplus \\tilde{\\phi}_{i + 1}(k)" + }, + { + "bbox": [ + 138, + 132, + 404, + 312 + ], + "type": "text", + "content": " end" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "type": "image", + "bbox": [ + 185, + 327, + 429, + 399 + ], + "blocks": [ + { + "bbox": [ + 185, + 327, + 429, + 399 + ], + "lines": [ + { + "bbox": [ + 185, + 327, + 429, + 399 + ], + "spans": [ + { + "bbox": [ + 185, + 327, + 429, + 399 + ], + "type": "image", + "image_path": "d77049b1c21816005f3775b76d5e4c5836b01ec44720bb4cfe3a798ed4ff3bd9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 408, + 364, + 419 + ], + "lines": [ + { + "bbox": [ + 248, + 408, + 364, + 419 + ], + "spans": [ + { + "bbox": [ + 248, + 408, + 364, + 419 + ], + "type": "text", + "content": "Fig. 4: Experimental setup." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 442, + 367, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 367, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 367, + 455 + ], + "type": "text", + "content": "6.1 Absolute Accuracy of 3D Reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "text", + "content": "We measured a periodically waving plate at a distance of " + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "inline_equation", + "content": "500~\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "text", + "content": " by using our BSC (8 frames) and five representative methods for comparison: 1) traditional four-step PSP (4 frames), 2) Hilbert transform compensation (HTC, 4 frames) [24], 3) micro Fourier transform profilometry (" + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "text", + "content": "-FTP, 4 frames) [34], 4) phase frame difference method (PFD, 8 frames) [10], 5) phase frame sum method (PFS, 5 frmaes) [4]. The fringe wavelength was set to 24 pixels for the cyclic " + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\pi/2" + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "text", + "content": " phase shifting patterns and was set to [22,24,26] for " + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "text", + "content": "-FTP. The plate moves within a distance range of [400,500] mm and a speed range of " + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "inline_equation", + "content": "[-150,150]" + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "text", + "content": " mm/s, and our measurement lasted for 800 frames (about 8.9 seconds). We select three " + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "inline_equation", + "content": "160\\times 120" + }, + { + "bbox": [ + 130, + 461, + 482, + 640 + ], + "type": "text", + "content": " rectangular windows in each depth map and fit the quadric surface as the ground truth, then we compute the root mean squared error (RMSE) within the selected regions. The trend of RMSE changing with motion speed is depicted in Fig. 5. The results show that our method exhibits the most satisfactory accuracy among existing methods when dealing with motion error." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Then we selected a moment when the speed of the plate was " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "-88~\\mathrm{mm / s}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ", and depict the reconstructed point clouds in Fig. 6, our BSC effectively reduces the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 114, + 430, + 285 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 430, + 285 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 430, + 285 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 430, + 285 + ], + "type": "image", + "image_path": "b71efe0e4f51af127cb00e16143ffa3b98a23ad5f5c79421228a8de4277b823a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 293, + 482, + 336 + ], + "lines": [ + { + "bbox": [ + 131, + 293, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 482, + 336 + ], + "type": "text", + "content": "Fig. 5: Measurement error VS motion speed of a periodically waving plate by traditional four-step phase shifting, HTC [24], " + }, + { + "bbox": [ + 131, + 293, + 482, + 336 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 131, + 293, + 482, + 336 + ], + "type": "text", + "content": "-FTP [34], PFD [10], PFS [4], and our BSC, the value in parentheses represents the mean error. Note the y-axis is in logarithmic scale." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 369, + 245, + 454 + ], + "blocks": [ + { + "bbox": [ + 189, + 350, + 201, + 361 + ], + "lines": [ + { + "bbox": [ + 189, + 350, + 201, + 361 + ], + "spans": [ + { + "bbox": [ + 189, + 350, + 201, + 361 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 369, + 245, + 454 + ], + "lines": [ + { + "bbox": [ + 141, + 369, + 245, + 454 + ], + "spans": [ + { + "bbox": [ + 141, + 369, + 245, + 454 + ], + "type": "image", + "image_path": "66bda019f3fc400695d484b3f50cd4c80b4791adfd49caa711b8e5550fac6fe9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 461, + 201, + 472 + ], + "lines": [ + { + "bbox": [ + 189, + 461, + 201, + 472 + ], + "spans": [ + { + "bbox": [ + 189, + 461, + 201, + 472 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 247, + 369, + 350, + 453 + ], + "blocks": [ + { + "bbox": [ + 295, + 350, + 307, + 361 + ], + "lines": [ + { + "bbox": [ + 295, + 350, + 307, + 361 + ], + "spans": [ + { + "bbox": [ + 295, + 350, + 307, + 361 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 369, + 350, + 453 + ], + "lines": [ + { + "bbox": [ + 247, + 369, + 350, + 453 + ], + "spans": [ + { + "bbox": [ + 247, + 369, + 350, + 453 + ], + "type": "image", + "image_path": "347d1287a6cf663eed11807d92a1ed1956249eb8a705ed30227b152ed613c02d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 353, + 359, + 474, + 454 + ], + "blocks": [ + { + "bbox": [ + 402, + 350, + 412, + 361 + ], + "lines": [ + { + "bbox": [ + 402, + 350, + 412, + 361 + ], + "spans": [ + { + "bbox": [ + 402, + 350, + 412, + 361 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 359, + 474, + 454 + ], + "lines": [ + { + "bbox": [ + 353, + 359, + 474, + 454 + ], + "spans": [ + { + "bbox": [ + 353, + 359, + 474, + 454 + ], + "type": "image", + "image_path": "39bc771459a7d180cbc8ddbb2c3d8982a192b88b25c84e8508e9e8e9c99c87ad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 141, + 475, + 244, + 561 + ], + "blocks": [ + { + "bbox": [ + 141, + 475, + 244, + 561 + ], + "lines": [ + { + "bbox": [ + 141, + 475, + 244, + 561 + ], + "spans": [ + { + "bbox": [ + 141, + 475, + 244, + 561 + ], + "type": "image", + "image_path": "f3457610b00de9c11b86ec3f0c6db120ec8d187411961090490735753595db4b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 570, + 480, + 593 + ], + "lines": [ + { + "bbox": [ + 131, + 570, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 480, + 593 + ], + "type": "text", + "content": "Fig. 6: Measurement results of a moving flat plane by (a) traditional four-step phase shifting, (b) HTC [24], (c) " + }, + { + "bbox": [ + 131, + 570, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 131, + 570, + 480, + 593 + ], + "type": "text", + "content": "-FTP [34], (d) PFD [10], (e) PFS [4], and (f) our BSC." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 247, + 474, + 350, + 561 + ], + "blocks": [ + { + "bbox": [ + 295, + 462, + 306, + 472 + ], + "lines": [ + { + "bbox": [ + 295, + 462, + 306, + 472 + ], + "spans": [ + { + "bbox": [ + 295, + 462, + 306, + 472 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 474, + 350, + 561 + ], + "lines": [ + { + "bbox": [ + 247, + 474, + 350, + 561 + ], + "spans": [ + { + "bbox": [ + 247, + 474, + 350, + 561 + ], + "type": "image", + "image_path": "5703c01ad67352a26821cf11148de1204f7ca35a91610687ccaf0b0693a9d015.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 353, + 474, + 473, + 561 + ], + "blocks": [ + { + "bbox": [ + 402, + 462, + 411, + 472 + ], + "lines": [ + { + "bbox": [ + 402, + 462, + 411, + 472 + ], + "spans": [ + { + "bbox": [ + 402, + 462, + 411, + 472 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 474, + 473, + 561 + ], + "lines": [ + { + "bbox": [ + 353, + 474, + 473, + 561 + ], + "spans": [ + { + "bbox": [ + 353, + 474, + 473, + 561 + ], + "type": "image", + "image_path": "e51ec7ccd87c780c87d91503f748a423613f8174f01c401409fcf98fff4bec73.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "motion error from " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "324.2\\mu \\mathrm{m}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "54.78\\mu \\mathrm{m}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " compared with traditional four-step phase shifting. By counting the histogram of errors, it can be observed that the error distribution of our BSC visually presents a normal distribution, which means Gaussian noise (induced by unstable ambient light, camera/projector" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "text", + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 137, + 247, + 221 + ], + "blocks": [ + { + "bbox": [ + 136, + 137, + 247, + 221 + ], + "lines": [ + { + "bbox": [ + 136, + 137, + 247, + 221 + ], + "spans": [ + { + "bbox": [ + 136, + 137, + 247, + 221 + ], + "type": "image", + "image_path": "6107b12975720eca2ae77f4acf1463066f32ce840464695bbff3b353d6cc941e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 125, + 359, + 196 + ], + "blocks": [ + { + "bbox": [ + 301, + 116, + 312, + 124 + ], + "lines": [ + { + "bbox": [ + 301, + 116, + 312, + 124 + ], + "spans": [ + { + "bbox": [ + 301, + 116, + 312, + 124 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 125, + 359, + 196 + ], + "lines": [ + { + "bbox": [ + 250, + 125, + 359, + 196 + ], + "spans": [ + { + "bbox": [ + 250, + 125, + 359, + 196 + ], + "type": "image", + "image_path": "38a2c2486c55c25c705cceefbd1d091de49c0babdfa1d5ae950a122651feb0d4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 251, + 198, + 361, + 241 + ], + "blocks": [ + { + "bbox": [ + 251, + 198, + 361, + 241 + ], + "lines": [ + { + "bbox": [ + 251, + 198, + 361, + 241 + ], + "spans": [ + { + "bbox": [ + 251, + 198, + 361, + 241 + ], + "type": "image", + "image_path": "d620481d207dcc81790361f60b4c291e2d5b57f4576048d206146b906cee0f65.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 241, + 311, + 248 + ], + "lines": [ + { + "bbox": [ + 301, + 241, + 311, + 248 + ], + "spans": [ + { + "bbox": [ + 301, + 241, + 311, + 248 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 369, + 125, + 474, + 196 + ], + "blocks": [ + { + "bbox": [ + 422, + 116, + 432, + 124 + ], + "lines": [ + { + "bbox": [ + 422, + 116, + 432, + 124 + ], + "spans": [ + { + "bbox": [ + 422, + 116, + 432, + 124 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 125, + 474, + 196 + ], + "lines": [ + { + "bbox": [ + 369, + 125, + 474, + 196 + ], + "spans": [ + { + "bbox": [ + 369, + 125, + 474, + 196 + ], + "type": "image", + "image_path": "48a0943cc52c726e001be28e78d7ff106498059226715656547e9db5beb52298.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 364, + 198, + 474, + 240 + ], + "blocks": [ + { + "bbox": [ + 364, + 198, + 474, + 240 + ], + "lines": [ + { + "bbox": [ + 364, + 198, + 474, + 240 + ], + "spans": [ + { + "bbox": [ + 364, + 198, + 474, + 240 + ], + "type": "image", + "image_path": "64ee32d814590d9392324ae4f104a47e7e53f3bf874fa82e6ea6d2b56e72cdfe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 241, + 432, + 248 + ], + "lines": [ + { + "bbox": [ + 422, + 241, + 432, + 248 + ], + "spans": [ + { + "bbox": [ + 422, + 241, + 432, + 248 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 136, + 258, + 247, + 344 + ], + "blocks": [ + { + "bbox": [ + 136, + 258, + 247, + 344 + ], + "lines": [ + { + "bbox": [ + 136, + 258, + 247, + 344 + ], + "spans": [ + { + "bbox": [ + 136, + 258, + 247, + 344 + ], + "type": "image", + "image_path": "2a576bdd7eb9d230f71017763154b3b94a1485661f7f790948fc94183e59f01b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 373, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 407 + ], + "type": "text", + "content": "Fig. 7: Depth discontinuous scene (top left) and motion area labeled in white (bottom left). Reconstruction results and error at edges: (a) traditional four-step phase shifting, (b) HTC [24], (c) " + }, + { + "bbox": [ + 130, + 373, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 373, + 482, + 407 + ], + "type": "text", + "content": "-FTP [34], and (d) our BSC." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 250, + 258, + 359, + 320 + ], + "blocks": [ + { + "bbox": [ + 250, + 258, + 359, + 320 + ], + "lines": [ + { + "bbox": [ + 250, + 258, + 359, + 320 + ], + "spans": [ + { + "bbox": [ + 250, + 258, + 359, + 320 + ], + "type": "image", + "image_path": "182ac23580f8950548a31da9dac2662a1e105214e08749d4f499af73bc49b8bc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 251, + 321, + 361, + 363 + ], + "blocks": [ + { + "bbox": [ + 251, + 321, + 361, + 363 + ], + "lines": [ + { + "bbox": [ + 251, + 321, + 361, + 363 + ], + "spans": [ + { + "bbox": [ + 251, + 321, + 361, + 363 + ], + "type": "image", + "image_path": "f033ad971d272589c2ff7d5783fd7c8fedee85824efad92d3a459d05e02e947e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 369, + 258, + 474, + 320 + ], + "blocks": [ + { + "bbox": [ + 369, + 258, + 474, + 320 + ], + "lines": [ + { + "bbox": [ + 369, + 258, + 474, + 320 + ], + "spans": [ + { + "bbox": [ + 369, + 258, + 474, + 320 + ], + "type": "image", + "image_path": "6c769052cd8291da027c0237b67585ce5e79e780b3a020f771daa8e4e5fdeb62.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 364, + 322, + 474, + 363 + ], + "blocks": [ + { + "bbox": [ + 364, + 322, + 474, + 363 + ], + "lines": [ + { + "bbox": [ + 364, + 322, + 474, + 363 + ], + "spans": [ + { + "bbox": [ + 364, + 322, + 474, + 363 + ], + "type": "image", + "image_path": "694599452683c4cc3f9eb4e44b4b18338aff1fd0796461c36f92ca1b09dbf66c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 434, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 481 + ], + "type": "text", + "content": "flicker, quantization error, and sensor noise of the camera and projector) is the main component of error. In other words, the motion error is effectively suppressed by BSC. Meanwhile, the error distribution of other methods deviates from the normal distribution, indicating the existence of residual errors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 506, + 380, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 380, + 519 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 380, + 519 + ], + "type": "text", + "content": "6.2 Robustness on Depth Discontinuous Scenes" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "PSP's pixel-wise characteristic is one of its most crucial merits. Our BSC maintains this property, thereby exhibiting robustness to depth discontinuous scenes. We set up a scene with three static gypsum geometries and one moving plate. Then we employed our BSC, " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "-FTP, and HTC. The results are shown in Fig. 7. It can be seen that the two non-pixel-wise methods, i.e., " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "-FTP and HTC, exhibit visible deformation to the naked eye in the point cloud around the depth jumping areas. We selected three rectangular windows and quantitatively evaluate the RMSE, the results show that the RMSE of non-pixel-wise methods is significantly greater than our BSC. We can conclude that our BSC with pixel-wise property is more robust compared with non-pixel-wise methods when dealing with depth discontinuous scenes." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 112, + 479, + 262 + ], + "blocks": [ + { + "bbox": [ + 136, + 112, + 479, + 262 + ], + "lines": [ + { + "bbox": [ + 136, + 112, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 136, + 112, + 479, + 262 + ], + "type": "image", + "image_path": "479f16360ff1712f2274a892ee058779fb4532ad46bc4f665c5d30c450355394.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 269, + 482, + 292 + ], + "lines": [ + { + "bbox": [ + 131, + 269, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 131, + 269, + 482, + 292 + ], + "type": "text", + "content": "Fig. 8: The measurement result of a wooden butterfly with wings flapping, during " + }, + { + "bbox": [ + 131, + 269, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 0\\mathrm{ms}\\sim 88\\mathrm{ms}" + }, + { + "bbox": [ + 131, + 269, + 482, + 292 + ], + "type": "text", + "content": ", Row 1 and 2: our BSC; Row 3 and 4: " + }, + { + "bbox": [ + 131, + 269, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 131, + 269, + 482, + 292 + ], + "type": "text", + "content": "-FTP [34]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 315, + 342, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 315, + 342, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 315, + 342, + 327 + ], + "type": "text", + "content": "6.3 Temporal Resolution of 3D Imaging" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": "Our BSC employs cyclic " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " phase-shifting fringe patterns, any " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "K + N" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " successive captured images can be used for reconstructing one frame of 3D point clouds. Therefore, our resulting frame rate of 3D point clouds is the same as the camera frame rate, thereby achieving a high temporal resolution of the measurement. That is the frame-wise loopable property we mentioned in the introduction. We successively conducted BSC and " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": "-FTP to measure a wooden butterfly with wings flapping cyclically, and then manually aligned the timeline, the results are shown in Fig. 8. We noticed that our method can achieve 3D reconstruction with a frame rate of 90 fps (the same as the camera frame rate), while " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": "-FTP reconstructing at 30 fps. Thus, our method is capable of distinguishing tinier motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 483, + 369, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 369, + 495 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 369, + 495 + ], + "type": "text", + "content": "6.4 Generality on Different Dynamic Objects" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 502, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 562 + ], + "type": "text", + "content": "We captured 400 image frames (about 4 seconds) employing our pipeline on four different dynamic objects, and the results are shown in Fig. 9. The complete motion process video can be found in the supplementary materials. We observed that BSC can effectively eliminate the severe ripple errors caused by object motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 580, + 345, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 580, + 345, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 580, + 345, + 594 + ], + "type": "text", + "content": "7 Limitations and Future Outlook" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "type": "text", + "content": "We presented a pixel-wise and frame-wise loopable binomial self-compensation algorithm to effectively and flexibly eliminate motion error in four-step PSP, without any intermediate variables assisted." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "High-frequency textures: our BSC assumes low-frequency textures like most motion compensation algorithms, it is robust against moderate contrast" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 158, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 158, + 91, + 448, + 102 + ], + "type": "text", + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 119, + 220, + 193 + ], + "blocks": [ + { + "bbox": [ + 142, + 119, + 220, + 193 + ], + "lines": [ + { + "bbox": [ + 142, + 119, + 220, + 193 + ], + "spans": [ + { + "bbox": [ + 142, + 119, + 220, + 193 + ], + "type": "image", + "image_path": "13fdbcdc871228cc97262fd25b98c3e34fffe36bfba2435a90a42f7b8bde0d93.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 120, + 300, + 193 + ], + "blocks": [ + { + "bbox": [ + 223, + 120, + 300, + 193 + ], + "lines": [ + { + "bbox": [ + 223, + 120, + 300, + 193 + ], + "spans": [ + { + "bbox": [ + 223, + 120, + 300, + 193 + ], + "type": "image", + "image_path": "6a94379d7f65a90a1da9a8799401de7a68166c2a21d6fed9b24929ea1a67345d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 303, + 120, + 380, + 193 + ], + "blocks": [ + { + "bbox": [ + 303, + 120, + 380, + 193 + ], + "lines": [ + { + "bbox": [ + 303, + 120, + 380, + 193 + ], + "spans": [ + { + "bbox": [ + 303, + 120, + 380, + 193 + ], + "type": "image", + "image_path": "4cd63007085a631bceb3a584acfba6966e663ad29fa08ef6488dea823dbc8505.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 384, + 120, + 477, + 194 + ], + "blocks": [ + { + "bbox": [ + 384, + 120, + 477, + 194 + ], + "lines": [ + { + "bbox": [ + 384, + 120, + 477, + 194 + ], + "spans": [ + { + "bbox": [ + 384, + 120, + 477, + 194 + ], + "type": "image", + "image_path": "c6d9be6c48f0351dd74b52b6a0df6a67bcb9d255a895cbadc66bbc983125e150.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 142, + 205, + 219, + 273 + ], + "blocks": [ + { + "bbox": [ + 142, + 205, + 219, + 273 + ], + "lines": [ + { + "bbox": [ + 142, + 205, + 219, + 273 + ], + "spans": [ + { + "bbox": [ + 142, + 205, + 219, + 273 + ], + "type": "image", + "image_path": "3a7391dfee698618eafd0c9a667535e24bdd4e36a9e7fbfee1100f7d7c1bfa12.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "text", + "content": "Fig. 9: The measurement results of four different dynamic objects. Column 1: wooden butterfly model with wings flapping, Column 2: moving statue, Column 3: waving hand, Column 4: hand showing scissors, stone, and cloth gestures. Top to bottom: our BSC and traditional four-step PSP." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 204, + 300, + 273 + ], + "blocks": [ + { + "bbox": [ + 223, + 204, + 300, + 273 + ], + "lines": [ + { + "bbox": [ + 223, + 204, + 300, + 273 + ], + "spans": [ + { + "bbox": [ + 223, + 204, + 300, + 273 + ], + "type": "image", + "image_path": "c36e7274a64d8b134b877edad16417321dea969a791b931128c159ff5ff30839.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 303, + 203, + 380, + 273 + ], + "blocks": [ + { + "bbox": [ + 303, + 203, + 380, + 273 + ], + "lines": [ + { + "bbox": [ + 303, + 203, + 380, + 273 + ], + "spans": [ + { + "bbox": [ + 303, + 203, + 380, + 273 + ], + "type": "image", + "image_path": "2e501875e5e20f69ab84ab317ea8317095a0e0b2a80eee738284e5c2cf658c04.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 384, + 205, + 477, + 273 + ], + "blocks": [ + { + "bbox": [ + 384, + 205, + 477, + 273 + ], + "lines": [ + { + "bbox": [ + 384, + 205, + 477, + 273 + ], + "spans": [ + { + "bbox": [ + 384, + 205, + 477, + 273 + ], + "type": "image", + "image_path": "55f779acca8e3d1cfe8f205ab58418b4f1d16780e80c821c9e675238ba4dadcc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 219, + 353, + 396, + 445 + ], + "blocks": [ + { + "bbox": [ + 219, + 353, + 396, + 445 + ], + "lines": [ + { + "bbox": [ + 219, + 353, + 396, + 445 + ], + "spans": [ + { + "bbox": [ + 219, + 353, + 396, + 445 + ], + "type": "image", + "image_path": "655a29d17c5a91ddac4cb54e83f623dd6fccc98be6d55605163145e47d874d40.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 453, + 429, + 464 + ], + "lines": [ + { + "bbox": [ + 184, + 453, + 429, + 464 + ], + "spans": [ + { + "bbox": [ + 184, + 453, + 429, + 464 + ], + "type": "text", + "content": "Fig. 10: 3D error VS inter-frame phase-shifting error " + }, + { + "bbox": [ + 184, + 453, + 429, + 464 + ], + "type": "inline_equation", + "content": "|\\Delta x_{i}|" + }, + { + "bbox": [ + 184, + 453, + 429, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 505, + 480, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 480, + 540 + ], + "type": "text", + "content": "textures. However, when dealing with objects that have high contrast textures, our method induces artifacts similar to those seen with typical phase-shifting methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 549, + 480, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 480, + 597 + ], + "type": "text", + "content": "Phase unwrapping error: we adopt a common block matching algorithm for stereo correspondence, inducing obvious phase unwrapping error when measuring complex surfaces. In the future, we will conduct a hierarchical adjustment for phase outliers to optimize the accuracy of the point cloud." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Applicable speed range: Our BSC is predicated on the assumption that the phase-shift error " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " is small. Consequently, as shown in Fig. 10, the residual motion error arises as " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "|\\Delta x_{i}|" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " increases. Properly reducing the fringe frequency could decrease " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "|\\Delta x_{i}|" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ", alleviating the performance loss of BSC for high-speed objects." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "text", + "content": "1. Achar, S., Narasimhan, S.G.: Multi focus structured light for recovering scene shape and global illumination. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 205-219. Springer (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "text", + "content": "2. Chen, T., Seidel, H.P., Lensch, H.P.: Modulated phase-shifting for 3d scanning. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1-8. IEEE (2008)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "text", + "content": "3. Geng, J.: Structured-light 3d surface imaging: a tutorial. Advances in optics and photonics 3(2), 128-160 (2011)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 263 + ], + "type": "text", + "content": "4. Guo, W., Wu, Z., Zhang, Q., Wang, Y.: Real-time motion-induced error compensation for 4-step phase-shifting profilometry. Optics Express 29(15), 23822-23834 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 265, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 265, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 265, + 480, + 297 + ], + "type": "text", + "content": "5. Gupta, M., Nakhate, N.: A geometric perspective on structured light coding. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 87-102 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "text", + "content": "6. Gupta, M., Tian, Y., Narasimhan, S.G., Zhang, L.: A combined theory of defocused illumination and global light transport. International Journal of Computer Vision 98, 146-167 (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "type": "text", + "content": "7. Gupta, M., Yin, Q., Nayar, S.K.: Structured light in sunlight. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 545-552 (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "text", + "content": "8. Hirose, K., Kamei, H., Sugiyama, T., Kurosaka, Y.: " + }, + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "inline_equation", + "content": "200 \\times 200 \\mu \\mathrm{m}" + }, + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "text", + "content": " 2 structured light source. Optics Express 28(25), 37307-37321 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 376, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 408 + ], + "type": "text", + "content": "9. Honrado, C.P., Larrabee Jr, W.F.: Update in three-dimensional imaging in facial plastic surgery. Current opinion in otolaryngology & head and neck surgery 12(4), 327-331 (2004)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "text", + "content": "0. Liu, X., Tao, T., Wan, Y., Kofman, J.: Real-time motion-induced-error compensation in 3d surface-shape measurement. Optics Express 27(18), 25265-25279 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "type": "text", + "content": "1. Liu, Y., Pears, N., Rosin, P.L., Huber, P.: 3D imaging, analysis and applications. Springer (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "type": "text", + "content": "2. Liu, Z., Zibley, P.C., Zhang, S.: Motion-induced error compensation for phase shifting profilometry. Optics Express 26(10), 12632-12637 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 476, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 508 + ], + "type": "text", + "content": "3. Logozzo, S., Zanetti, E.M., Franceschini, G., Kilpelä, A., Mäkynen, A.: Recent advances in dental optics-part i: 3d intraoral scanners for restorative dentistry. Optics and Lasers in Engineering 54, 203-221 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "type": "text", + "content": "4. Lu, L., Ding, Y., Luan, Y., Yin, Y., Liu, Q., Xi, J.: Automated approach for the surface profile measurement of moving objects based on psp. Optics Express 25(25), 32120-32131 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 544, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 575 + ], + "type": "text", + "content": "5. Lu, L., Xi, J., Yu, Y., Guo, Q.: New approach to improve the accuracy of 3-d shape measurement of moving object using phase shifting profilometry. Optics Express 21(25), 30610-30622 (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "type": "text", + "content": "6. Lu, L., Yin, Y., Su, Z., Ren, X., Luan, Y., Xi, J.: General model for phase shifting profilometry with an object in motion. Applied Optics 57(36), 10364-10369 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "7. Lv, S., Kemao, Q.: Modeling the measurement precision of fringe projection profilometry. Light: Science & Applications 12(1), 257 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 621, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 642 + ], + "type": "text", + "content": "8. Nayar, S.K., Gupta, M.: Diffuse structured light. In: IEEE International Conference on Computational Photography (ICCP). pp. 1-11. IEEE (2012)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "9. Posdamer, J.L., Altschuler, M.D.: Surface measurement by space-encoded projected beam systems. Computer Graphics and Image Processing 18(1), 1-17 (1982)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "text", + "content": "Binomial Self-compensation for Motion Error in Dynamic 3D Scanning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 588 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "20. Qian, J., Feng, S., Xu, M., Tao, T., Shang, Y., Chen, Q., Zuo, C.: High-resolution real-time 360 3d surface defect inspection with fringe projection profilometry. Optics and Lasers in Engineering 137, 106382 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "21. Qian, J., Tao, T., Feng, S., Chen, Q., Zuo, C.: Motion-artifact-free dynamic 3d shape measurement with hybrid fourier-transform phase-shifting profilometry. Optics Express 27(3), 2713-2731 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "text", + "content": "22. Sansoni, G., Corini, S., Lazzari, S., Rodella, R., Docchio, F.: Three-dimensional imaging based on gray-code light projection: characterization of the measuring algorithm and development of a measuring system for industrial applications. Applied Optics 36(19), 4463-4472 (1997)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "23. Srinivasan, V., Liu, H.C., Halioua, M.: Automated phase-measuring profilometry of 3-d diffuse objects. Applied optics 23(18), 3105-3108 (1984)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "24. Wang, Y., Liu, Z., Jiang, C., Zhang, S.: Motion induced phase error reduction using a hilbert transform. Optics Express 26(26), 34224-34235 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "25. Wang, Y., Suresh, V., Li, B.: Motion-induced error reduction for binary defocusing profilometry via additional temporal sampling. Optics Express 27(17), 23948-23958 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "text", + "content": "26. Weise, T., Leibe, B., Van Gool, L.: Fast 3d scanning with automatic motion compensation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1-8. IEEE (2007)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "27. Wu, G., Yang, T., Liu, F., Qian, K.: Suppressing motion-induced phase error by using equal-step phase-shifting algorithms in fringe projection profilometry. Optics Express 30(11), 17980-17998 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "28. Ye, Y., Chang, H., Song, Z., Zhao, J.: Accurate infrared structured light sensing system for dynamic 3d acquisition. Applied Optics 59(17), E80-E88 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "29. Zhang, Y., Lau, D., Wipf, D.: Sparse multi-path corrections in fringe projection profilometry. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13344-13353 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "30. Zhang, Y., Lau, D.L., Yu, Y.: Causes and corrections for bimodal multi-path scanning with structured light. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4431-4439 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "31. Zhong, K., Li, Z., Shi, Y., Wang, C., Lei, Y.: Fast phase measurement profilometry for arbitrary shape objects without phase unwrapping. Optics and Lasers in Engineering 51(11), 1213-1222 (2013)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "32. Zhu, J., Yang, F., Hu, J., Zhou, P.: High dynamic reflection surface 3d reconstruction with sharing phase demodulation mechanism and multi-indicators guided phase domain fusion. Optics Express 31(15), 25318-25338 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "33. Zuo, C., Feng, S., Huang, L., Tao, T., Yin, W., Chen, Q.: Phase shifting algorithms for fringe projection profilometry: A review. Optics and lasers in engineering 109, 23-59 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "34. Zuo, C., Tao, T., Feng, S., Huang, L., Asundi, A., Chen, Q.: Micro fourier transform profilometry (μftp): 3d shape measurement at 10,000 frames per second. Optics and Lasers in Engineering 102, 70-91 (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "G. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_content_list.json b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..64f86ceaa06571c771debac92f33a16a9ee1bc04 --- /dev/null +++ b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_content_list.json @@ -0,0 +1,1787 @@ +[ + { + "type": "text", + "text": "BlazeBVD: Make Scale-Time Equalization Great Again for Blind Video Deflickering", + "text_level": 1, + "bbox": [ + 222, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinmin Qiu $^{1}$ , Congying Han $^{1}$ , Zicheng Zhang $^{1}$ , Bonan Li $^{1}$ , Tiande Guo $^{1}$ , Pingyu Wang $^{2}$ , and Xuecheng Nie $^{3}$", + "bbox": [ + 228, + 218, + 772, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Chinese Academy of Sciences, Beijing, China", + "bbox": [ + 297, + 261, + 702, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Sichuan University, Sichuan, China", + "bbox": [ + 374, + 277, + 625, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 MT Lab, Meitu Inc., Beijing, China", + "bbox": [ + 374, + 290, + 627, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qiuxinmin21,zhangzicheng19}@mails.ucas.ac.cn,", + "bbox": [ + 325, + 305, + 676, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hancy,libonacci,tdguo}@ucas.ac.cn", + "bbox": [ + 375, + 319, + 627, + 332 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wangpingyu@scu.edu.cn, nxc@meitu.com", + "bbox": [ + 364, + 334, + 638, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/272906ba01293ee2a957fdc94b81aefe058b99b6cb2639e6fc31e94781068bcc.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 217, + 383, + 357, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2fdfc44e33dbeee4131ea7b7906a2ded145613c3a808b7f646bbe7a0aae299cc.jpg", + "image_caption": [ + "Deflicker" + ], + "image_footnote": [], + "bbox": [ + 359, + 383, + 500, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dc5085a5bafb6366d59691c93e040444202f2bc7010fa436628d37d269904027.jpg", + "image_caption": [ + "Ours", + "Fig. 1: Comparisons of the proposed BlazeBVD. We present flickering input, GT, Deflicker [21] and our BlazeBVD processed video frames, and illumination histograms along with KL divergence about GroundTruth. Our method recovers the illumination histograms well while avoiding the appearance of color artifacts and color distortions (such as the man's arm in the second column). Better see in color with $2 \\times$ zoom." + ], + "image_footnote": [], + "bbox": [ + 501, + 383, + 642, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/070bce0cc1aea6af146eebeb3bc234a883b6a86d1e8c442574bf12724f40bfeb.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 642, + 383, + 784, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Developing blind video deflickering (BVD) algorithms to enhance video temporal consistency, is gaining importance amid the flourish of image processing and video generation. However, the intricate nature of video data complicates the training of deep learning methods, leading to high resource consumption and instability, notably under severe lighting flicker. This underscores the critical need for a compact representation beyond pixel values to advance BVD research and applications. Inspired by the classic scale-time equalization (STE), our work introduces the histogram-assisted solution, called BlazeBVD, for high-fidelity and rapid BVD. Compared with STE, which directly corrects pixel values by temporally smoothing color histograms, BlazeBVD", + "bbox": [ + 259, + 661, + 740, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\mathbf{\\bar{\\Omega}}$ Corresponding Author", + "bbox": [ + 230, + 823, + 397, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leverages smoothed illumination histograms within STE filtering to ease the challenge of learning temporal data using neural networks. In technique, BlazeBVD begins by condensing pixel values into illumination histograms that precisely capture flickering and local exposure variations. These histograms are then smoothed to produce singular frames set, filtered illumination maps, and exposure maps. Resorting to these deflickering priors, BlazeBVD utilizes a 2D network to restore faithful and consistent texture impacted by lighting changes or localized exposure issues. BlazeBVD also incorporates a lightweight 3D network to amend slight temporal inconsistencies, avoiding the resource consumption issue. Comprehensive experiments on synthetic, real-world and generated videos, showcase the superior qualitative and quantitative results of BlazeBVD, achieving inference speeds up to $10 \\times$ faster than state-of-the-arts.", + "bbox": [ + 259, + 147, + 743, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Video de-flicker $\\cdot$ Histogram $\\cdot$ Temporal consistency", + "bbox": [ + 261, + 340, + 696, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 383, + 375, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As social applications and the multimedia industry continue to expand, videos have become an essential medium for conveying information in daily life [30]. Achieving a high-quality video demands both image clarity and temporal consistency, nevertheless, video flickering frequently compromises the temporal consistency, often resulting from shooting environment and camera hardware limitations [9, 18]. This issue is further exacerbated when image processing techniques are applied to video frames. Moreover, flickering artifacts and color distortion are also prevalent in recent video generation technologies, including those based on Generative Adversarial Networks (GANs) [7,35,36] and Diffusion Models (DMs) [25,42,46,52]. Since even certain types of flicker can significantly detract from the viewing experience, it is crucial to develop video deflickering techniques to eliminate flicker and preserve the integrity of video content given unknown degradation in various video processing scenarios.", + "bbox": [ + 212, + 416, + 787, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recovering visual content from unknown degradation is a key yet challenging topic [33, 43, 49]. Several seminal works [8, 20, 21] have addressed the blind video deflickering (BVD) task, in which methods refer to that (1) is agnostic to flickering patterns or levels (e.g., old movies, high-speed cameras, processing artifacts and color distortions), and (2) operates on a single flicker video without requiring additional guidance (e.g., flicker types, reference videos). In other words, BVD methods are blind to flicker types and guidance, making them widely applicable. Given these, the BVD task is challenging due to the lack of additional prior information guidance. Previous methods have explored traditional filtering [8, 31], forced temporal consistency [20, 22], and atlas adjustment [21], etc.", + "bbox": [ + 212, + 613, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite significant advancements by deep-learning methods in the BVD task, several critical issues hinder their broader application. First, these approaches demand substantial resources during both the training and inference stages, notably manifesting in slow processing speeds for individual videos. Second, their effectiveness is compromised by two main factors: (i) partial color artifacts and", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Qiu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distortion are present in the resulting video, causing differences in color fidelity and detail compared to the original video; (ii) Overexposure or underexposure due to lighting leads to significant loss of texture details and introduces new local flicker. These issues are largely attributed to the spatio-temporal complexity within video data, where the sheer volume of pixels presents a formidable obstacle for neural networks, including the advanced large vision models, to learn and maintain global visual consistency. Consequently, given that flicker fundamentally involves local or instantaneous shifts in illumination, effectively addressing the BVD task necessitates a representation more compact and adept at capturing illumination fluctuations than pixel values, whereas this aspect has not yet been explored by previous deep BVD methods.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we draw inspiration from the classic flicker removal method scale-time equalization (STE) to facilitate deep BVD with histogram representation. An image histogram is defined as a distribution of pixel values, and it has been widely applied in image manipulation to adjust image brightness or contrast. Given an arbitrary video, STE temporally smooths histograms with Gaussian filtering and corrects pixel values in each frame with histogram equalization, improving the visual stability of the video. Although STE is only effective for some mild flickers, it validates that 1) histogram, which is much more compact than pixel values, can well delineate the illumination and flicker information, and 2) the video with a smoothed histogram series is visually appealing without obvious flicker, as shown in Fig. 1. Therefore, it is promising to advance the quality and speed of deep BVD with the hints from STE and histogram.", + "bbox": [ + 212, + 316, + 789, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce Blaze Blind Video Deflickering, dubbed as BlazeBVD, which is a histogram-assisted approach to achieve fast and faithful texture restoration given illumination fluctuation and over-/under-exposure. Compared to previous deep methods, BlazeBVD is the first to leverage histogram meticulously to ease the learning complexity of the BVD task. Specifically, BlazeBVD comprises three stages: At first, we introduce STE to rectify the histogram series of video frames under the illumination space, and extract deflickering priors including singular frames set, filtered illumination maps, and exposure maps. Second, due to the stable temporal performance of filtered illumination maps, they will be used as prompt conditions of a global flicker removal module (GFRM) containing 2D spatial networks to guide the color correction of video frames. On the other hand, a local flicker removal module (LFRM) is based on the optical flow warp to restore local over-/under-exposure regions labeled by exposure maps. Finally, we introduce a lightweight spatio-temporal network to process all frames, where an adaptive mask weighted warping loss is designed to improve video coherence. Our contributions can be summarized as follows:", + "bbox": [ + 212, + 503, + 789, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) We present BlazeBVD, a histogram-assisted blind video deflickering method that simplifies the complexity and resource consumption of learning video data. At its core, BlazeBVD utilizes deflickering priors from STE, including filtered illumination maps for guiding the elimination of global flicker, singular frames set for identify the indexes of flicker frames, and exposure maps to identify regions affected locally by over-/under-exposure.", + "bbox": [ + 212, + 750, + 797, + 840 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) Leveraging deflickering priors, BlazeBVD incorporates a Global Flicker Removal Module (GFRM) and a Local Flicker Removal Module (LFRM). These modules work together to efficiently correct global illumination and locally exposed textures in individual adjacent frames, significantly reducing processing time compared to handling the entire video. For enhanced coherence across frames, a lightweight spatio-temporal network is also integrated, boosting performance without significant time consumption.", + "(3) Through comprehensive experiments on synthetic, real-world and generated videos, we showcase the superior qualitative and quantitative results of BlazeBVD, achieving model inference speeds up to $10 \\times$ faster than state-of-the-art. To the best of our knowledge, BlazeBVD also represents the first method to effectively tackle both illumination fluctuations and exposure challenges." + ], + "bbox": [ + 212, + 146, + 784, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 352, + 387, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video deflickering addresses the degradation in temporal consistency [26, 27, 29, 38, 48] resulting from abnormal camera, lighting, exposure or image processing algorithms. The relevant methods can be divided into three main categories. The first type of approach focuses on addressing specific flickering types, e.g., high-speed cameras [18], old film [9]. Delon [8] performs an axiomatic analysis of flicker, and proposed a fast method for global removal based on scale space theory to filter in color space directly. The second type of approach, blind temporal consistency [39, 53], aims to generate temporally stable output videos for arbitrary videos regardless of flickering or other artifacts. Bonneel et al. [4] calculate the gradient of the input frame as a guide to reduce the degree of inconsistency between frames. Lai et al. [20] input two consecutive frames as guidance. Lei et al. [22] directly learn the mapping function between input frames and processing frames to reach the purpose of inter-frame information consistency. Despite the generalization performance, such methods tend to be less efficient and hard to handle the complex flicker in practice. The third type of approach is blind video deflickering. Lei et al. [21] propose to solve universal blind video flicker using the neural atlas and video consistency. Meanwhile, some commercial software can remove flicker by integrating various known-flicker methods, such as DEFlicker [34] and Flicker Free [2]. However, most of these methods require users to have flickering type knowledge background or slow repair speed. Our approach does not require prior cognition of the specific information flashing, and achieves fast deflickering speeds so that most users can efficiently process more videos.", + "bbox": [ + 212, + 386, + 787, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exposure correction. The purpose of exposure correction is to enhance an improperly exposed image to achieve a satisfactory visual effect. Information is lost due to over-/under-exposure input images. How to solve this problem intelligently and automatically, there have been many techniques in the image domain [12, 13, 15, 24, 41]. ERL [14] connects the under-/over-exposed optimization processes by correlating and constraining the relationships in the mini-batch correction process. DA [44] proposes to decouple contrast enhancement and detail recovery in each convolution process and design detail perception units to", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Qiu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be inserted into the existing CNN-based exposure correction network. However, there is no general approach regarding local exposure in video tasks.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optical flow finds the corresponding relationship between the previous frame and the current frame by using the changes of pixels in the time domain and the correlation between adjacent frames, to calculate the motion information of objects between consecutive frames [10, 11, 16, 37]. In recent years, it has been used in video-related visual tasks [50]. The main target of flow-based methods is to produce accurate estimates of the optical flow through different network architectures, such as utilizing bidirectional optical flow estimation and refining inter-frame coherence.", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 316, + 380, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce basic definitions and notations, and review relevant techniques involving image histograms for use in BlazeBVD.", + "bbox": [ + 212, + 345, + 782, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image histogram is a graphical representation illustrating the distribution of pixels across different intensity levels within an image [3]. For a single-channel image $I$ with height $H$ and width $W$ , its histogram $\\mathcal{H}$ can be calculated as", + "bbox": [ + 212, + 375, + 785, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} (\\lambda) = \\frac {n _ {\\lambda}}{H W}, \\quad i f \\lambda \\in I o t h e r w i s e 0, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 426, + 785, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_{\\lambda}$ represents the count of pixels having the intensity $\\lambda$", + "bbox": [ + 212, + 457, + 666, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Histogram matching or equalization [1, 5, 47, 51], is a technique to adjust the contrast of an image by transforming its histogram towards a predefined distribution. The cumulative histogram of an image histogram hist is defined by $\\text{Hist}(\\lambda) = \\Sigma_{r=0}^{\\lambda} \\mathcal{H}(r)$ . Give a target histogram $\\mathcal{H}'$ and its cumulative histogram $\\text{Hist}'(\\lambda)$ , the histogram matching process can be described as", + "bbox": [ + 212, + 472, + 787, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} (\\lambda ; \\mathcal {H}, \\mathcal {H} ^ {\\prime}) = H i s t ^ {t - 1} (H i s t (\\lambda)), \\forall \\lambda \\in I, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 556, + 785, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This procedure yields an image $\\tilde{X}$ , whose histogram closely aligns with $Hist'$ . Scale-Time Equalization (STE) is a filtering approach [8] to adjust the intensity distribution across both spatial and temporal dimensions, mitigating flickers encountered in video sequences. The classic scale-space theory posits that convolving an image with a Gaussian kernel results in smoothing and the loss of image structure. Intuitively, while STE is based on scale-space theory, it applies Gaussian smoothing to sequences of histograms instead of the images. Given a flickering video sequence $\\{X_{t}\\}_{t=1}^{T}$ , leveraging Eq. (1) can get the histogram sequence $\\{\\mathcal{H}_t\\}_{t=1}^T$ . For arbitrary time index $t$ and intensity value $\\lambda$ , STE can be formulated as the convolution:", + "bbox": [ + 212, + 580, + 782, + 731 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS T E (\\lambda , t) = G _ {s} (\\tau) * \\mathcal {Q} (\\lambda ; \\mathcal {H} _ {t}, \\mathcal {H} _ {\\tau}) \\approx \\sum_ {\\tau = t - l} ^ {t + l} G _ {s} (\\tau - t) \\mathcal {Q} (\\lambda ; \\mathcal {H} _ {t}, \\mathcal {H} _ {\\tau}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 737, + 785, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $l$ denotes the window radius, $G_{s}(t) = \\frac{1}{\\sqrt{4\\pi s}} e^{-\\frac{t^{2}}{4s}}$ . After STE, we will get a sequence of histograms $\\{\\tilde{\\mathcal{H}}_t\\}_{t=1}^T$ and images $\\{\\tilde{X}_t\\}_{t=1}^T$ . Since histogram transformations do not create or cancel any image content, STE can achieve deflickering", + "bbox": [ + 212, + 787, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18c392babcb1e41a7661c6024ccdc5259d9009aabd33e6af8576ab947bec4219.jpg", + "image_caption": [ + "Fig. 2: The framework of our approach BlazeBVD. We first extract flicker prior information about the input video and correct the brightness representation with STE-assisted histogram filtering in illuminance space (Stage1). Then, the prior is leveraged to remove temporal flicker and over-/under-exposure flicker from the global and local perspectives (GFRM and LFRM in Stage2). Finally, the temporal consistency of the processed video is improved (TCM in Stage3)." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 781, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "results without compromising the integrity of the original video. However, this also indicates that STE cannot compensate for the lost texture in flickering regions and produces vivid effects, which impedes its practical application.", + "bbox": [ + 212, + 473, + 787, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 542, + 380, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $\\{X_{t}\\}_{t = 1}^{T}$ represent a video sequence affected by flickering defects and $\\{G_t\\}_{t = 1}^T$ represent the groundtruth clean video. In this section, we propose BlazeBVD as a universal solution to eliminate these flicker and produce a temporally consistent video $\\{O_t\\}_{t = 1}^T$ . It is important to note that the types of flicker addressed in our work can vary widely, involving long-/short-term flicker in color and brightness.", + "bbox": [ + 212, + 575, + 785, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Distinguished from existing prior BVD approaches, BlazeBVD employs a histogram-assisted approach, leveraging deflickering priors from Scale-Time Equalization (STE) [8] to streamline the complexity and resource demands of BVD task. Beyond ensuring stable temporal performance, BlazeBVD stands out as the first method capable of faithfully restoring the color and texture affected by illumination fluctuations and exposure challenges, including both over-/under-exposure.", + "bbox": [ + 212, + 652, + 795, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Overview of BlazeBVD", + "text_level": 1, + "bbox": [ + 215, + 782, + 454, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, the BlazeBVD pipeline is structured into three main stages to address video deflickering challenges effectively: (1) Deflickering priors", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Qiu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/12b2971deb2050878524729be7c5a753dd390ee4dc7bf9ed489d3ef2c3e41f7c.jpg", + "image_caption": [ + "Fig. 3: Pipeline of Stage1. The illumination maps are corrected in the temporal dimension by STE, which alleviates the difficulty of subsequent temporal data learning using networks. The deflickering priors can be extracted: $\\tilde{V}_t$ , $M_t$ and $S_{flicker}$ ." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "preparation. This initial stage leverages STE to correct the histogram within an illumination space representation. Key outputs from this stage include the filtered illumination map $\\{\\tilde{V}_t\\}_{t=1}^T$ , identifying the index set of singular frames $S_{flicker}$ , and generating exposure maps $\\{M_t\\}_{t=1}^T$ . These elements serve as crucial prior information, guiding the deflickering process in subsequent stages. (2) Global and local flicker removal. The second stage involves two flicker removal strategies. A global flicker removal module utilizes the filtered illumination map $\\{\\tilde{V}_t\\}_{t=1}^T$ to address temporal flicker issues across the video. A local over-/under-exposed region flicker removal module instructed by the singular frames set $S_{flicker}$ and exposure maps $\\{M_t\\}_{t=1}^T$ , restores the texture in the over-/under-exposure region to compensate for the lost high-frequency details. (3) Adaptive temporal consistency. The final stage introduces a spatio-temporal network under the guidance of an adaptive warping loss to improve video coherence and temporal consistency.", + "bbox": [ + 212, + 371, + 787, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These stages delineated in the following Sec. 4.2, Sec. 4.3, and Sec. 4.4, together provide a comprehensive solution to enhance the visual quality and temporal consistency of video content.", + "bbox": [ + 212, + 583, + 787, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Stage1: Deflickering priors preparation", + "text_level": 1, + "bbox": [ + 215, + 648, + 581, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatio-temporal challenge. An effective BVD model should strategically avoid the intricacies of spatio-temporal complexity across pixels, nevertheless, current deep learning approaches to BVD have not adequately addressed this challenge. For instance, the state-of-the-art BVD method, Deflicker [21], employs a neural atlas [19] as a unified representation of the entire video. While this can offer consistent guidance for the deflickering process, the effectiveness of neural atlases significantly depends on the video duration and the network capacity. Processing an 8-second video to create an atlas can take around 10 minutes, and longer videos exacerbate the demands on time, memory, and computational resources. Consequently, there is a further need for research into more compact representations that can facilitate deflickering without these limitations.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Illumination histogram. We introduce histogram representation [3], which utilizes statistical distribution to illustrate the global properties of image pixel values. As illustrated in Fig. 3, we first extract the illumination maps $\\{V_t\\}_{t=1}^T$ according to the conversion between HSV and RGB color space:", + "bbox": [ + 215, + 146, + 782, + 205 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nV _ {t} = \\max _ {R, G, B \\in X _ {t}} \\{R, G, B \\}, t \\in \\{1, 2, \\dots , T \\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 214, + 785, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, by applying Eq. (1) to each illumination map, we utilize illumination histograms $\\{H_t\\}_{t=1}^T$ to represent the illumination characteristics of video frames. These illumination histograms reflect important flickering information: As shown in Fig. 3, frames without distinct illumination changes have similar histograms, while the histogram of a flicker frame has a distribution distinguished from others. Furthermore, spikes in the head or tail area of the distribution indicate potential over-/under-exposure problems in the image. To this end, reasonable utilization of histogram information is significant for flicker elimination and alleviation of spatio-temporal complexity.", + "bbox": [ + 212, + 243, + 782, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Deflickering priors. We introduce deflickering priors extracted from illumination histograms to facilitate the deflickering process. Initially, considering that the illumination histograms of natural videos exhibit continuous changes in the temporal dimension, we utilize Scale-Time Equalization (STE) as defined in Eq. (3) to smooth the illumination histograms, as illustrated in Fig. 3. Through histogram matching, we further obtain filtered illumination maps $\\{\\tilde{V}\\}_{t=1}^{T}$ that exhibit no obvious flicker compared to the original ones, thereby enabling the subsequent independent processing of each frame by the neural network. Secondly, we identify the singular frames set $S_{flicker}$ containing the indexes of flicker frames, based on the divergence of histograms before and after STE:", + "bbox": [ + 212, + 378, + 784, + 530 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS _ {f l i c k e r} = \\left\\{t \\in [ 1, T ] \\mid \\mathrm {K L} \\left(\\tilde {H} _ {t} \\parallel H _ {t}\\right) > \\bar {\\epsilon} _ {t} \\right\\}, \\quad \\bar {\\epsilon} _ {t} = \\frac {1}{2 n + 1} \\sum_ {\\tau = t - n} ^ {t + n} \\mathrm {K L} \\left(\\tilde {H} _ {\\tau} \\parallel H _ {\\tau}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 537, + 784, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $n$ is the moving average radius. Finally, as shown in Fig. 2, we extract the exposure maps $\\{M_t \\in \\mathbb{R}^{H \\times W}\\}_{t=1}^T$ of local over-/under-exposure according to the filtered illumination map:", + "bbox": [ + 212, + 583, + 782, + 628 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nM _ {t} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\tilde {V} (i, j) < \\epsilon_ {1} \\text {o r} \\tilde {V} (i, j) > \\epsilon_ {2}, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 635, + 784, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\epsilon_{1}$ and $\\epsilon_{2}$ are darkness and exposure thresholds, respectively, and we consider illumination values belonging to $[\\epsilon_1,\\epsilon_2]$ as not posing an exposure risk.", + "bbox": [ + 212, + 679, + 782, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, we leverage illumination histograms to prepare deflickering priors including filtered illumination maps $\\{\\tilde{V}\\}_{t=1}^{T}$ , exposure maps $\\{M_t\\}_{t=1}^T$ , and the singular frames set $S_{flicker}$ , for alleviating the spatio-temporal complexity and enhancing deflickering in the subsequent learning stages.", + "bbox": [ + 212, + 710, + 782, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Stage2: Global and local flicker removal", + "text_level": 1, + "bbox": [ + 215, + 787, + 588, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this stage, simple 2D spatial neural networks are integrated with deflickering priors to eliminate both global and local flickering flaws. Here, \"global\" refers", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Qiu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to obvious temporal inconsistency due to illumination fluctuation, while \"local\" refers to lost texture details due to exposure and darkness. This approach allows for the direct generation of high-quality videos from flicker frames, regardless of the spatio-temporal complexity.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Global deflickering. BlazeBVD is equipped with a global flicker removal module (GFRM), which incorporates a 2D-Unet to correct each frame $X_{t}$ with the guidance of the filtered illumination map $\\tilde{V}_{t}$ . In other words, the corrected image $\\hat{X}_{t} = U\\mathrm{net}_{\\theta}(X_{t},\\tilde{V}_{t})$ , where $\\theta$ denotes the parameters of the 2D-Unet. The network is trained on the BVD dataset $\\mathcal{D}$ by minimizing the L2 loss function between the predicted images and the groundtruth images:", + "bbox": [ + 212, + 207, + 784, + 297 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\mathbb {E} _ {\\left(X _ {t}, G _ {t}\\right) \\in \\mathcal {D}} \\| U n e t _ {\\theta} \\left(X _ {t}, \\tilde {V} _ {t}\\right) - G _ {t} \\|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 308, + 784, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Local deflickering. BlazeBVD is also equipped with a local flicker removal module (LFRM) to enhance local texture details. As depicted in Fig. 5, overexposure can transform intricate texture in local regions into uniform pure, while local under-exposure can result in black areas. This loss of detailed texture cannot be compensated for by previous BVD methods, such as Deflicker [21]. We leverage LFRM to transfer detailed information from neighboring frames to these ill regions. Given the corrected singular frame $\\tilde{X}_t$ in $S_{\\text{flicker}}$ with local exposure, along with the previous frame $\\tilde{X}_{t-1}$ and the next frame $\\tilde{X}_{t+1}$ , we first compute the optical flow to extract part of the detailed texture:", + "bbox": [ + 212, + 333, + 784, + 470 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\no _ {\\tau \\rightarrow t} = \\operatorname {F l o w} \\left(X _ {\\tau}, X _ {t}\\right), \\tau \\in \\{t + 1, t - 1 \\} \\text {a n d} t \\in S _ {\\text {f l i c k e r}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 481, + 784, + 497 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here Flow is the optical flow estimation function, in our work we set it as the popular RAFT model [11, 37]. We migrate the possible detail textures of the preceding and following frames into the current frame:", + "bbox": [ + 212, + 506, + 782, + 551 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {t} = \\mathcal {F} \\left(M _ {t} \\odot \\mathcal {W} \\left(\\tilde {X} _ {t - 1}, o _ {t - 1 \\rightarrow t}\\right) + (1 - M _ {t}) \\odot \\tilde {X} _ {t}, \\tilde {X} _ {t}, \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 561, + 784, + 588 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nM _ {t} \\odot \\mathcal {W} \\left(\\tilde {X} _ {t - 1}, o _ {t + 1 \\rightarrow t}\\right) + \\left(1 - M _ {t}\\right) \\odot \\tilde {X} _ {t}), t \\in S _ {\\text {f l i c k e r}},\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 583, + 722, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{W}$ is the warping operator and $\\mathcal{F}$ is the fusion network from [41].", + "bbox": [ + 212, + 609, + 733, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Stage3: Adaptive temporal consistency", + "text_level": 1, + "bbox": [ + 214, + 646, + 581, + 661 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The former stages have well addressed the main illumination fluctuation and exposure problems in BVD task. Nevertheless, a video may also have content inconsistencies to largely degrade the visual effect, and the former stage may also cause this. Therefore, We introduce the Temporal Consistency Model (TCM) to refine video $\\{\\hat{X}_t\\}_{t=1}^T$ as the final outputs $\\{O_t\\}_{t=1}^T$ . TCM is developed based on the architecture by RTN [40], but with a temporal consistency loss of adaptive masks to refine local artifacts:", + "bbox": [ + 212, + 669, + 784, + 773 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {p a i r}} ^ {\\prime} \\left(O _ {t}, O _ {s}\\right) = \\left\\| W _ {t} \\odot \\left(M _ {t} + 1\\right) \\odot M _ {t, s} \\odot \\left(O _ {t} - \\mathcal {W} \\left(O _ {s}\\right)\\right) \\right\\| _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 782, + 705, + 799 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w a r p} = \\frac {1}{T - 1} \\sum_ {t = 2} ^ {T} \\left\\{E _ {\\text {p a i r}} ^ {\\prime} \\left(O _ {t}, O _ {1}\\right) + E _ {\\text {p a i r}} ^ {\\prime} \\left(O _ {t}, O _ {t - 1}\\right) \\right\\}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 804, + 784, + 842 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $W_{t}$ is the loss weight for the over-/under-exposure area and $M_{t}$ is the mask obtained from the former stage. $\\mathcal{W}$ is the warping operator and the frame after warping is computed by optical flow estimation [11,37]. $M_{t,s}$ represents the mask after optical flow estimation. This loss avoids the video blurring problem introduced in Deflicker [21]. The training loss of TCM consists of reconstruction loss $\\mathcal{L}_{rec}$ , perceptual loss $\\mathcal{L}_{per}$ [17], spatio-temporal adversarial loss $\\mathcal{L}_{adv}$ [6] and designed adaptive weighted warping loss $\\mathcal{L}_{warp}$ . Combining all the above losses, the overall objective to optimize TCM is:", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T C M} = \\alpha_ {1} * \\mathcal {L} _ {r e c} + \\alpha_ {2} * \\mathcal {L} _ {p e r} + \\alpha_ {3} * \\mathcal {L} _ {a d v} + \\alpha_ {4} * \\mathcal {L} _ {w a r p}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 281, + 785, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In our work, we set $\\alpha_{1} = 1.0$ , $\\alpha_{2} = 1.0$ , $\\alpha_{3} = 0.01$ and $\\alpha_{4} = 0.1$ .", + "bbox": [ + 214, + 308, + 676, + 324 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 347, + 375, + 366 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 380, + 334, + 393 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We inherit the training dataset DAVIS-2017-Train, MS-COCO and the testing dataset Blind Deflickering Dataset from Deflicker [21], including Synthetic videos from DAVIS-2017-Val, Real-world videos and Generation videos.", + "bbox": [ + 212, + 405, + 785, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training dataset. MS-COCO [23] contains 118287 images in the train set and is used for the pre-training of Unet in GFRM. DAVIS-2017-Train [20, 28, 32] consists of 60 sequences with a total of 4219 labeled frames, with a video frame rate of 24fps and a resolution of $480\\mathrm{p}$ . We create the synthetic process that provides the underlying facts for quantitative analysis. The training dataset is degraded by performing flicker on DAVIS-2017-Train, as shown in Eq. (12):", + "bbox": [ + 212, + 450, + 787, + 542 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} = G _ {t} + F _ {t}, t \\in \\{1, 2, \\dots , T \\}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 556, + 785, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\{G_t\\}_{t=1}^T$ are clean frames and $\\{F_t\\}_{t=1}^T$ are random synthesized flickering artifacts. The window size $W$ in $F_t$ has been set to denote the number of frames that share the same flickering artifacts, random sampling between [2, 12]. The value of $W$ controls the degree of flicker in the short- and long-term of the video.", + "bbox": [ + 212, + 582, + 787, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Testing dataset. We conduct Synthetic videos, Real-world videos and Generation videos respectively, from the Blind Deflickering Dataset [21] and DAVIS2017-Test processed by Eq. (12). Among them, Synthetic videos and DAVIS2017-Test have ground truth and the others do not. We can only use no-referenced metric and qualitative observation. Generation videos refer to the processed videos generated by generative algorithms, which may lead to flicker artifacts due to the imperfections of the algorithms. Real-world videos including old movies, old anime, slow-motion, etc. are directly captured, unprocessed videos in the real world.", + "bbox": [ + 212, + 643, + 787, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison methods. We compare our BlazeBVD with other 4 methods, including STE [8], ConvLSTM [20], DVP [22] and Deflicker [21]. We set Deflicker as the baseline, since it is the first publicly presented method for BVD task and is currently the most advanced approach for general videos.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Qiu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/075647226b6f736da970113ef77697ab5554a0abab0a1d77e24f5002c1902aef.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons between previous methods and our BlazeBVD. Our method removes flicker and restores details in over-exposed regions (red boxes in row 6 and row 8) while avoiding color artifacts (row 2 and row 4). Besides, BlazeBVD also ensures the fidelity of the video content and avoids color distortion (cyan boxes in row 5 and row 7). Zoom in for the best view and we recommend watching videos in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 498 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 215, + 635, + 622, + 650 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics. We quantitatively compare the differences between our method and others with reference and no-reference metrics, including PSNR, SSIM [45], and $E_{\\text{warp}}$ [22]. Among them, $E_{\\text{warp}}$ is a no-reference metric used in videos. For each frame $O_t$ , we calculate the warping error with frame $O_{t-1}$ and the first frame $O_1$ for considering both short-term and long-term consistency.", + "bbox": [ + 212, + 657, + 787, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative results. We compare our BlazeBVD with the state-of-the-art method Deflicker [21], and other consistency methods. Table 1 and Table 2 provide the comparative experimental results of various methods on DAVIS-2017-Test and Synthetic videos, where PSNR and SSIM of the resulting video are higher than those of the other methods. This shows that BlazeBVD contributes to the fidelity maintenance of video content, as well as the correctness of the missing details complement. The warping error of our method is a little higher", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0de3d15abc8656f74688e7957dd48bba86f2247460d10d92e7558e854235aa75.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons of synthetic videos in DAVIS-2017-Test. Our method performs best in terms of PSNR and SSIM and $E_{\\text{warp}}$ of all the methods." + ], + "table_footnote": [], + "table_body": "
MethodRawVideoConvLSTMDVPSTEDeflickerOursGT
PSNR↑-24.11024.13626.13823.93228.609-
SSIM↑-0.92560.92630.93210.92430.9638-
E_warp ↓0.14290.13690.14150.11170.08400.08250.1097
", + "bbox": [ + 215, + 184, + 785, + 258 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0a76f09d83b159d396f1858f0e7b49524d8de90be6356f5c095e4044946ee8d6.jpg", + "table_caption": [ + "Table 2: Quantitative comparisons in Synthetic videos from Blind Deflickering Dataset. $W$ is the window size and represents the number of frames that share the same flicker artifact. $L$ denotes the local seed window size." + ], + "table_footnote": [], + "table_body": "
VideoTypeRawE_warp↓PSNR↑
DeflickerOursGTDeflickerOursDeflickerOurs
Synthetic
-W=10.19330.08480.08950.105625.931229.67380.94030.9539
-W=30.15220.08440.08900.105625.499628.82720.93800.9555
-W=100.13070.08550.09040.105624.384226.99660.93440.9543
-L=30.11130.08750.09000.105625.387427.08810.92710.9517
Average0.14690.08560.08970.105625.300628.14640.93490.9538
", + "bbox": [ + 215, + 327, + 787, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "than Deflicker, but it is closer to the ground truth. Through analysis, we find that warping error has inaccuracy in the estimation of optical flow, and this error is masked by image blurring in Deflicker. Therefore, we believe that $E_{\\text{warp}}$ , which has the same magnitude as baseline and GT, is the valid experimental comparison result. In Tab. 3, we compare BlazeBVD and Deflicker on Real-world videos and Generation videos. Since the videos in these two datasets with no reference are inherently discontinuous, where $E_{\\text{warp}}$ cannot directly represent video continuity, we present the comparative results of user studies. To evaluate the perceived preference between the baseline Deflicker and the proposed method, each user needs to select a video with better perceptual quality between these two choices. In total, there are 50 users and 50 pairs in 7 datasets of comparisons.", + "bbox": [ + 212, + 492, + 785, + 656 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative results. Both ConvLSTM and DVP are temporal consistency models that require a reference, importing the input video as the reference, and the processed video cannot remove significant flicker, as shown in Fig. 4. Additionally, the flicker caused by over-/under-exposure is prone to lose the local texture (column 1 in Fig. 4). The atlas-based representation in Deflicker cannot accurately locate these points with different local features, resulting in the loss of all details on the processed video. Instead, our method designs local positioning and adjacent frame details transfer in LFRM to preserve part of the texture on the road, as shown in the red box of Fig. 4. We also reveal the color artifact in column 4 of Fig. 4 that appears at baseline, due to the excessive weight of warping error in the training loss when refining temporal consistency. Therefore, our designed adaptive mask weighted training loss in TCM improves the temporal", + "bbox": [ + 212, + 657, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Qiu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f4cfcdbced6ce05050e318b0b00a913e4a121aaa5f82ae883342a6dc5674c47e.jpg", + "table_caption": [ + "Table 3: User studies of Real-world videos and Generation videos. Since the videos in these two datasets are inherently discontinuous, $E_{\\text{warp}}$ cannot properly reflect temporal consistency. Here we only present the comparative results of the user study." + ], + "table_footnote": [], + "table_body": "
TypeExpertOldAnimeOldMovieSlowMotionTimeLapseVideoDMVideoLDM
Raw(E_warp)0.07540.10450.10990.07240.07320.12480.1601
Deflicker38.67%19.25%22.83%37.67%31.43%30.50%26.34%
Ours61.33%80.75%77.17%62.33%68.57%69.50%73.66%
", + "bbox": [ + 218, + 198, + 782, + 272 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "consistency while avoiding image blurring. More visualization results are in the supplementary material.", + "bbox": [ + 212, + 301, + 782, + 330 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 354, + 401, + 369 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "STE filtering. We contrast the difference between introducing STE filtering in illumination space and directly in color space. As shown in the red box of Fig. 5, directly performing histogram correction in color channels separately introduces color artifacts and color distortion. Because the filtering of non-gray images in the three channels is relatively independent, the recombination after nonlinear transformation would produce such defects.", + "bbox": [ + 212, + 381, + 784, + 470 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "GFRM and LFRM. We compare the important roles of two modules, GFRM and LFRM. As shown in Tab. 5, removing GFRM and LFRM reduces PSNR and SSIM and increases $E_{\\text{warp}}$ of inference videos. It can be inferred that GFRM needs to initially remove global flicker between frames, otherwise subsequent modules cannot maintain the temporal consistency, while LFRM maintains the texture fidelity of the video content, which is reflected in the cyan box of Fig. 5. Designed weighted loss. As shown in Tab. 5, the lightweight spatio-temporal network TCM reduces $E_{\\text{warp}}$ , which refines the temporal consistency of processed videos. We also compare the effect of using unweighted training loss and the adaptive mask weighted loss on this temporal network. Figure 5 provides the resulting frame with unweighted training, which is generally blurred and suffers from color distortion, while the weighted training keeps the frame clear and faithful in most areas. Adaptive mask weighting improves the temporal consistency of the inference video, improving the inter-frame continuity more pertinently.", + "bbox": [ + 212, + 473, + 787, + 684 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Discussion", + "text_level": 1, + "bbox": [ + 215, + 707, + 349, + 722 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Advantages and limitations. BlazeBVD is over $10 \\times$ faster than the inference speed of baseline using the atlas representation, which is specifically recorded in Tab. 4. Our method also takes advantage of explicit prior extraction in the BVD task, which provides directions for improvement in the subsequent work. And the cause of local flicker has been noted and effectively dealt with from both global and local perspectives. However, our method still has drawbacks in LFRM. Due to the inaccurate optical flow motion estimation, the fusion network", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Inference time and model parameters comparison. Our approach is faster than the baseline in inference time with less memory, and comparable in MACs and Params.", + "bbox": [ + 215, + 147, + 452, + 229 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/fc77fa3b8b0b70246e6f42d11b1f79850083434f3d98c24b5dec71e82c02bcf1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDeflickerOurs
Time/80p614.19s58.37s
Memory5204M1389M
MACs260.7G251.4G
Params12.48M17.77M
", + "bbox": [ + 217, + 243, + 452, + 329 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/694001cc9249d203195a39754f5001aad4db4f5c445af1ebeb4289fab833eeaa.jpg", + "table_caption": [ + "Table 5: Ablation studies in DAVIS-2017- Test. w/o STE replaces STE with mean filtering in Stage1 and w/o weighted denotes the warping loss without mask weighting in TCM." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑E_warp ↓
w/o STE26.44120.95170.1011
w/o GFRM23.89390.92440.1225
w/o LFRM27.13660.94620.1122
w/o TCM29.91300.95830.1093
w/o weighted28.10490.95230.0849
BlazeBVD28.60920.96380.0825
", + "bbox": [ + 468, + 212, + 799, + 333 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5740c8ba744f729da186f344f8354be7a3ea584b7ad58c1f6c4b14a18326bfae.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 217, + 348, + 357, + 410 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e24c203e70391a228155b2de7d6901dbcb8cb64e660a926e5b61c32771434e2b.jpg", + "image_caption": [ + "w/o GFRM" + ], + "image_footnote": [], + "bbox": [ + 359, + 348, + 500, + 410 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3d256e30443406c2ddd2af9e7de0467971f09eb7078a83aab616c3c952eb2a68.jpg", + "image_caption": [ + "w/o LFRM" + ], + "image_footnote": [], + "bbox": [ + 503, + 348, + 642, + 410 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/bb73a782de945ac40d6a1122a84cc82c9c520ba5445cc47725f55d91e9c4746a.jpg", + "image_caption": [ + "w/o STE" + ], + "image_footnote": [], + "bbox": [ + 645, + 348, + 785, + 410 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1741435d2c81bde25c30d85a62d1fde0e763015d13ddc8405ac50eb91c5db22f.jpg", + "image_caption": [ + "w/o TCM" + ], + "image_footnote": [], + "bbox": [ + 217, + 424, + 357, + 486 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c50d7a915ab75c7a4f8e9c2c68607b78d3b4d3b3715afe44f5f44a2402fed645.jpg", + "image_caption": [ + "w/o weighted mask" + ], + "image_footnote": [], + "bbox": [ + 359, + 424, + 500, + 484 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/efbd9145fe52c70a20930cc62857d55982fdb628d4cd3e937e29c9ae4bf55259.jpg", + "image_caption": [ + "Ours", + "Fig. 5: Qualitative ablation studies of key designs in BlazeBVD. Specifically, local detail loss is represented in cyan box, color artifact is represented in the red box, temporal inconsistency is represented in the black box, and color distortion is represented in the blue box. Zoom in for the best view." + ], + "image_footnote": [], + "bbox": [ + 503, + 424, + 642, + 484 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e169edf276ea632d6be578380b60d3ae9cc3e618564091f7c4a91567fb09c41b.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 645, + 424, + 785, + 484 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "cannot convey the local texture of adjacent frames, and there are still slight edge artifacts and color distortion that cannot be eliminated. We hope to continue exploring the transfer and fusion of local region information between frames [40].", + "bbox": [ + 214, + 616, + 787, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 693, + 359, + 709 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose BlazeBVD, a universal approach for the blind video deflickering task. The core of our method is to prepare flicker priors within the STE filter in illumination space first, and then utilize these priors to correct the global flicker and local exposed texture. GFRM and LFRM leverage the priors to remove global and local flicker in several adjacent frames with less computation. Finally, TCM is used to improve video coherence and inter-frame consistency. Extensive experiments validate the effectiveness of the designed modules.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Qiu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This paper is supported by the National Natural Science Foundation of China (Nos. U23B2012, 11991022), and the Fundamental Research Funds for the Central Universities (E3E41904X2).", + "bbox": [ + 215, + 176, + 787, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 247, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Affifi, M., Brubaker, M.A., Brown, M.S.: Histogram: Controlling colors of gangenerated and real images via color histograms. In: CVPR. pp. 7941-7950 (2021)", + "2. Anarchy, D.: Flicker free, https://digitalanarchy.com/Flicker/main.html", + "3. Bassiou, N., Kotropoulos, C.: Color image histogram equalization by absolute discounting back-off. Computer Vision and Image Understanding 107, 108-122 (Jul 2007)", + "4. Bonneel, N., Tompkin, J., Sunkavalli, K., Sun, D., Paris, S., Pfister, H.: Blind video temporal consistency. ACM Transactions on Graphics 34, 1-9 (Nov 2015)", + "5. Bottenus, N., Byram, B.C., Hyun, D.: Histogram matching for visual ultrasound image comparison. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control 68, 1487-1495 (May 2021)", + "6. Chang, Y.L., Liu, Z.Y., Lee, K.Y., Hsu, W.: Free-form video inpainting with 3d gated convolution and temporal patchgan. In: ICCV. pp. 9066-9075 (2019)", + "7. Chu, M., Xie, Y., Mayer, J., Leal-Taixe, L., Thuerey, N.: Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics 39, 75-1 (2020)", + "8. Delon, J.: Movie and video scale-time equalization application to flicker reduction. IEEE Transactions on Image Processing 15, 241-248 (Jan 2006)", + "9. Delon, J., Desolneux, A.: Stabilization of flicker-like effects in image sequences through local contrast correction. SIAM Journal on Imaging Sciences 3, 703-734 (Jan 2010)", + "0. Dosovitskiy, A., Fischer, P., Ilg, E., Hausser, P., Hazirbas, C., Golkov, V., Van Der Smagt, P., Cremers, D., Brox, T.: Flownet: Learning optical flow with convolutional networks. In: ICCV. pp. 2758-2766 (2015)", + "1. Eslami, N., Arefi, F., Mansourian, A.M., Kasaei, S.: Rethinking raft for efficient optical flow. arXiv preprint arXiv:2401.00833 (2024)", + "2. Huang, J., Liu, Y., Fu, X., Zhou, M., Wang, Y., Zhao, F., Xiong, Z.: Exposure normalization and compensation for multiple-exposure correction. In: CVPR. pp. 6043-6052 (2022)", + "3. Huang, J., Liu, Y., Zhao, F., Yan, K., Zhang, J., Huang, Y., Zhou, M., Xiong, Z.: Deep fourier-based exposure correction network with spatial-frequency interaction. In: ECCV. pp. 163-180 (2022)", + "4. Huang, J., Zhao, F., Zhou, M., Xiao, J., Zheng, N., Zheng, K., Xiong, Z.: Learning sample relationship for exposure correction. In: CVPR. pp. 9904-9913 (2023)", + "5. Huang, J., Zhou, M., Liu, Y., Yao, M., Zhao, F., Xiong, Z.: Exposure-consistency representation learning for exposure correction. In: ACM MM. pp. 6309-6317 (2022)", + "6. Ilg, E., Mayer, N., Saikia, T., Keuper, M., Dosovitskiy, A., Brox, T.: Flownet 2.0: Evolution of optical flow estimation with deep networks. In: CVPR. pp. 2462-2470 (2017)" + ], + "bbox": [ + 225, + 281, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV. pp. 694-711 (2016)", + "18. Kanj, A., Talbot, H., Luparello, R.R.: Flicker removal and superpixel-based motion tracking for high speed videos. In: ICIP. pp. 245-249 (2017)", + "19. Kasten, Y., Ofri, D., Wang, O., Dekel, T.: Layered neural atlases for consistent video editing. ACM Transactions on Graphics 40, 1-12 (2021)", + "20. Lai, W.S., Huang, J.B., Wang, O., Shechtman, E., Yumer, E., Yang, M.H.: Learning blind video temporal consistency. In: ECCV. pp. 170-185 (2018)", + "21. Lei, C., Ren, X., Zhang, Z., Chen, Q.: Blind video deflickering by neural filtering with a flawed atlas. In: CVPR. pp. 10439-10448 (2023)", + "22. Lei, C., Xing, Y., Chen, Q.: Blind video temporal consistency via deep video prior. In: NeurIPS. pp. 1083-1093 (2020)", + "23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014)", + "24. Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: CVPR. pp. 5637-5646 (2022)", + "25. Mei, K., Patel, V.: Vidm: Video implicit diffusion models. In: AAAI. pp. 9117-9125 (2023)", + "26. Moniz, J.R.A., Kang, E., Póczos, B.: Luciddream: Controlled temporally-consistent deepdream on videos. arXiv preprint arXiv:1911.11960 (2019)", + "27. Park, K., Woo, S., Kim, D., Cho, D., Kweon, I.S.: Preserving semantic and temporal consistency for unpaired video-to-video translation. In: ACM MM. pp. 1248-1257 (2019)", + "28. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: CVPR (Jun 2016)", + "29. Pérez-Pellitero, E., Sajjadi, M.S., Hirsch, M., Scholkopf, B.: Perceptual video super resolution with enhanced temporal consistency. arXiv preprint arXiv:1807.07930 (2018)", + "30. Pfeuffer, A., Dietmayer, K.: Separable convolutional lstms for faster video segmentation. In: ITSC. pp. 1072-1078 (2019)", + "31. Pitie, F., Kokaram, A., Dahyot, R.: Removing flicker from old movies. Master's thesis, University of Nice-Éophia Antipolis, France, Œe September (2002)", + "32. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv:1704.00675 (2017)", + "33. Qiu, X., Han, C., Zhang, Z., Li, B., Guo, T., Nie, X.: Diffbfr: Bootstrapping diffusion model for blind face restoration. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7785-7795 (2023)", + "34. RE:VISION: De:flicker, https://revisionfx.com/products/deflicker/", + "35. Saito, M., Saito, S., Koyama, M., Kobayashi, S.: Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. International Journal of Computer Vision p. 2586-2606 (Nov 2020)", + "36. Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In: CVPR. pp. 3626-3636 (2022)", + "37. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402-419 (2020)", + "38. Thasarathan, H., Nazeri, K., Ebrahimi, M.: Automatic temporally coherent video colorization. In: CRV. pp. 189-194 (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Qiu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Thimonier, H., Despois, J., Kips, R., Perrot, M.: Learning long-term style preserving blind video temporal consistency. In: ICME. pp. 1-6 (2021)", + "40. Wan, Z., Zhang, B., Chen, D., Liao, J.: Bringing old films back to life. In: CVPR. pp. 17694-17703 (2022)", + "41. Wang, H., Xu, K., Lau, R.W.: Local color distributions prior for image enhancement. In: ECCV. pp. 343-359 (2022)", + "42. Wang, W., Liu, J., Lin, Z., Yan, J., Chen, S., Low, C., Hoang, T., Wu, J., Liew, J.H., Yan, H., et al.: Magicvideo-v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468 (2024)", + "43. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9168-9178 (2021)", + "44. Wang, Y., Peng, L., Li, L., Cao, Y., Zha, Z.J.: Decoupling-and-aggregating for image exposure correction. In: CVPR. pp. 18115-18124 (2023)", + "45. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13, 600-612 (2004)", + "46. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV. vol. 39, pp. 7623-7633 (2023)", + "47. Xu, R., Tu, Z., Du, Y., Dong, X., Li, J., Meng, Z., Ma, J., Bovik, A., Yu, H.: Pik-fix: Restoring and colorizing old photos. In: WACV. pp. 1724-1734 (2023)", + "48. Xu, X., Li, M., Sun, W., Yang, M.H.: Learning spatial and spatio-temporal pixel aggregations for image and video denoising. IEEE Transactions on Image Processing 29, 7153-7165 (2020)", + "49. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 672-681 (2021)", + "50. Zhan, X., Pan, X., Liu, Z., Lin, D., Loy, C.C.: Self-supervised learning via conditional motion propagation. In: CVPR. pp. 1881-1889 (2019)", + "51. Zhang, F., Shao, Y., Sun, Y., Zhu, K., Gao, C., Sang, N.: Unsupervised low-light image enhancement via histogram equalization prior. arXiv preprint arXiv:2112.01766 (2021)", + "52. Zhang, Z., Li, B., Nie, X., Han, C., Guo, T., Liu, L.: Towards consistent video editing with text-to-image diffusion models. Advances in Neural Information Processing Systems 36 (2024)", + "53. Zhou, Y., Xu, X., Shen, F., Gao, L., Lu, H., Shen, H.T.: Temporal denoising mask synthesis network for learning blind video temporal consistency. In: ACM MM. pp. 475-483 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 686 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_model.json b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f69dd6a361012d513f88ff93e0965c38b98a620 --- /dev/null +++ b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_model.json @@ -0,0 +1,2478 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "BlazeBVD: Make Scale-Time Equalization Great Again for Blind Video Deflickering" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.219, + 0.773, + 0.252 + ], + "angle": 0, + "content": "Xinmin Qiu\\(^{1}\\), Congying Han\\(^{1}\\), Zicheng Zhang\\(^{1}\\), Bonan Li\\(^{1}\\), Tiande Guo\\(^{1}\\), Pingyu Wang\\(^{2}\\), and Xuecheng Nie\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.262, + 0.704, + 0.277 + ], + "angle": 0, + "content": "1 University of Chinese Academy of Sciences, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.278, + 0.626, + 0.29 + ], + "angle": 0, + "content": "\\(^{2}\\) Sichuan University, Sichuan, China" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.291, + 0.629, + 0.305 + ], + "angle": 0, + "content": "3 MT Lab, Meitu Inc., Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.306, + 0.678, + 0.319 + ], + "angle": 0, + "content": "{qiuxinmin21,zhangzicheng19}@mails.ucas.ac.cn," + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.32, + 0.628, + 0.333 + ], + "angle": 0, + "content": "{hancy,libonacci,tdguo}@ucas.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.335, + 0.639, + 0.347 + ], + "angle": 0, + "content": "wangpingyu@scu.edu.cn, nxc@meitu.com" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.385, + 0.359, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.504, + 0.297, + 0.515 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.385, + 0.501, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.504, + 0.453, + 0.515 + ], + "angle": 0, + "content": "Deflicker" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.385, + 0.643, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.504, + 0.588, + 0.514 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.385, + 0.785, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.504, + 0.721, + 0.513 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.538, + 0.788, + 0.608 + ], + "angle": 0, + "content": "Fig. 1: Comparisons of the proposed BlazeBVD. We present flickering input, GT, Deflicker [21] and our BlazeBVD processed video frames, and illumination histograms along with KL divergence about GroundTruth. Our method recovers the illumination histograms well while avoiding the appearance of color artifacts and color distortions (such as the man's arm in the second column). Better see in color with \\(2 \\times\\) zoom." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.662, + 0.741, + 0.815 + ], + "angle": 0, + "content": "Abstract. Developing blind video deflickering (BVD) algorithms to enhance video temporal consistency, is gaining importance amid the flourish of image processing and video generation. However, the intricate nature of video data complicates the training of deep learning methods, leading to high resource consumption and instability, notably under severe lighting flicker. This underscores the critical need for a compact representation beyond pixel values to advance BVD research and applications. Inspired by the classic scale-time equalization (STE), our work introduces the histogram-assisted solution, called BlazeBVD, for high-fidelity and rapid BVD. Compared with STE, which directly corrects pixel values by temporally smoothing color histograms, BlazeBVD" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.824, + 0.398, + 0.84 + ], + "angle": 0, + "content": "\\(\\mathbf{\\bar{\\Omega}}\\) Corresponding Author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "X. Qiu et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.329 + ], + "angle": 0, + "content": "leverages smoothed illumination histograms within STE filtering to ease the challenge of learning temporal data using neural networks. In technique, BlazeBVD begins by condensing pixel values into illumination histograms that precisely capture flickering and local exposure variations. These histograms are then smoothed to produce singular frames set, filtered illumination maps, and exposure maps. Resorting to these deflickering priors, BlazeBVD utilizes a 2D network to restore faithful and consistent texture impacted by lighting changes or localized exposure issues. BlazeBVD also incorporates a lightweight 3D network to amend slight temporal inconsistencies, avoiding the resource consumption issue. Comprehensive experiments on synthetic, real-world and generated videos, showcase the superior qualitative and quantitative results of BlazeBVD, achieving inference speeds up to \\(10 \\times\\) faster than state-of-the-arts." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.342, + 0.697, + 0.356 + ], + "angle": 0, + "content": "Keywords: Video de-flicker \\(\\cdot\\) Histogram \\(\\cdot\\) Temporal consistency" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.384, + 0.377, + 0.399 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.613 + ], + "angle": 0, + "content": "As social applications and the multimedia industry continue to expand, videos have become an essential medium for conveying information in daily life [30]. Achieving a high-quality video demands both image clarity and temporal consistency, nevertheless, video flickering frequently compromises the temporal consistency, often resulting from shooting environment and camera hardware limitations [9, 18]. This issue is further exacerbated when image processing techniques are applied to video frames. Moreover, flickering artifacts and color distortion are also prevalent in recent video generation technologies, including those based on Generative Adversarial Networks (GANs) [7,35,36] and Diffusion Models (DMs) [25,42,46,52]. Since even certain types of flicker can significantly detract from the viewing experience, it is crucial to develop video deflickering techniques to eliminate flicker and preserve the integrity of video content given unknown degradation in various video processing scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Recovering visual content from unknown degradation is a key yet challenging topic [33, 43, 49]. Several seminal works [8, 20, 21] have addressed the blind video deflickering (BVD) task, in which methods refer to that (1) is agnostic to flickering patterns or levels (e.g., old movies, high-speed cameras, processing artifacts and color distortions), and (2) operates on a single flicker video without requiring additional guidance (e.g., flicker types, reference videos). In other words, BVD methods are blind to flicker types and guidance, making them widely applicable. Given these, the BVD task is challenging due to the lack of additional prior information guidance. Previous methods have explored traditional filtering [8, 31], forced temporal consistency [20, 22], and atlas adjustment [21], etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Despite significant advancements by deep-learning methods in the BVD task, several critical issues hinder their broader application. First, these approaches demand substantial resources during both the training and inference stages, notably manifesting in slow processing speeds for individual videos. Second, their effectiveness is compromised by two main factors: (i) partial color artifacts and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "distortion are present in the resulting video, causing differences in color fidelity and detail compared to the original video; (ii) Overexposure or underexposure due to lighting leads to significant loss of texture details and introduces new local flicker. These issues are largely attributed to the spatio-temporal complexity within video data, where the sheer volume of pixels presents a formidable obstacle for neural networks, including the advanced large vision models, to learn and maintain global visual consistency. Consequently, given that flicker fundamentally involves local or instantaneous shifts in illumination, effectively addressing the BVD task necessitates a representation more compact and adept at capturing illumination fluctuations than pixel values, whereas this aspect has not yet been explored by previous deep BVD methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.79, + 0.501 + ], + "angle": 0, + "content": "In this paper, we draw inspiration from the classic flicker removal method scale-time equalization (STE) to facilitate deep BVD with histogram representation. An image histogram is defined as a distribution of pixel values, and it has been widely applied in image manipulation to adjust image brightness or contrast. Given an arbitrary video, STE temporally smooths histograms with Gaussian filtering and corrects pixel values in each frame with histogram equalization, improving the visual stability of the video. Although STE is only effective for some mild flickers, it validates that 1) histogram, which is much more compact than pixel values, can well delineate the illumination and flicker information, and 2) the video with a smoothed histogram series is visually appealing without obvious flicker, as shown in Fig. 1. Therefore, it is promising to advance the quality and speed of deep BVD with the hints from STE and histogram." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.79, + 0.744 + ], + "angle": 0, + "content": "We introduce Blaze Blind Video Deflickering, dubbed as BlazeBVD, which is a histogram-assisted approach to achieve fast and faithful texture restoration given illumination fluctuation and over-/under-exposure. Compared to previous deep methods, BlazeBVD is the first to leverage histogram meticulously to ease the learning complexity of the BVD task. Specifically, BlazeBVD comprises three stages: At first, we introduce STE to rectify the histogram series of video frames under the illumination space, and extract deflickering priors including singular frames set, filtered illumination maps, and exposure maps. Second, due to the stable temporal performance of filtered illumination maps, they will be used as prompt conditions of a global flicker removal module (GFRM) containing 2D spatial networks to guide the color correction of video frames. On the other hand, a local flicker removal module (LFRM) is based on the optical flow warp to restore local over-/under-exposure regions labeled by exposure maps. Finally, we introduce a lightweight spatio-temporal network to process all frames, where an adaptive mask weighted warping loss is designed to improve video coherence. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.799, + 0.842 + ], + "angle": 0, + "content": "(1) We present BlazeBVD, a histogram-assisted blind video deflickering method that simplifies the complexity and resource consumption of learning video data. At its core, BlazeBVD utilizes deflickering priors from STE, including filtered illumination maps for guiding the elimination of global flicker, singular frames set for identify the indexes of flicker frames, and exposure maps to identify regions affected locally by over-/under-exposure." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "X. Qiu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "(2) Leveraging deflickering priors, BlazeBVD incorporates a Global Flicker Removal Module (GFRM) and a Local Flicker Removal Module (LFRM). These modules work together to efficiently correct global illumination and locally exposed textures in individual adjacent frames, significantly reducing processing time compared to handling the entire video. For enhanced coherence across frames, a lightweight spatio-temporal network is also integrated, boosting performance without significant time consumption." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.329 + ], + "angle": 0, + "content": "(3) Through comprehensive experiments on synthetic, real-world and generated videos, we showcase the superior qualitative and quantitative results of BlazeBVD, achieving model inference speeds up to \\(10 \\times\\) faster than state-of-the-art. To the best of our knowledge, BlazeBVD also represents the first method to effectively tackle both illumination fluctuations and exposure challenges." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.353, + 0.388, + 0.37 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Video deflickering addresses the degradation in temporal consistency [26, 27, 29, 38, 48] resulting from abnormal camera, lighting, exposure or image processing algorithms. The relevant methods can be divided into three main categories. The first type of approach focuses on addressing specific flickering types, e.g., high-speed cameras [18], old film [9]. Delon [8] performs an axiomatic analysis of flicker, and proposed a fast method for global removal based on scale space theory to filter in color space directly. The second type of approach, blind temporal consistency [39, 53], aims to generate temporally stable output videos for arbitrary videos regardless of flickering or other artifacts. Bonneel et al. [4] calculate the gradient of the input frame as a guide to reduce the degree of inconsistency between frames. Lai et al. [20] input two consecutive frames as guidance. Lei et al. [22] directly learn the mapping function between input frames and processing frames to reach the purpose of inter-frame information consistency. Despite the generalization performance, such methods tend to be less efficient and hard to handle the complex flicker in practice. The third type of approach is blind video deflickering. Lei et al. [21] propose to solve universal blind video flicker using the neural atlas and video consistency. Meanwhile, some commercial software can remove flicker by integrating various known-flicker methods, such as DEFlicker [34] and Flicker Free [2]. However, most of these methods require users to have flickering type knowledge background or slow repair speed. Our approach does not require prior cognition of the specific information flashing, and achieves fast deflickering speeds so that most users can efficiently process more videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Exposure correction. The purpose of exposure correction is to enhance an improperly exposed image to achieve a satisfactory visual effect. Information is lost due to over-/under-exposure input images. How to solve this problem intelligently and automatically, there have been many techniques in the image domain [12, 13, 15, 24, 41]. ERL [14] connects the under-/over-exposed optimization processes by correlating and constraining the relationships in the mini-batch correction process. DA [44] proposes to decouple contrast enhancement and detail recovery in each convolution process and design detail perception units to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "be inserted into the existing CNN-based exposure correction network. However, there is no general approach regarding local exposure in video tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Optical flow finds the corresponding relationship between the previous frame and the current frame by using the changes of pixels in the time domain and the correlation between adjacent frames, to calculate the motion information of objects between consecutive frames [10, 11, 16, 37]. In recent years, it has been used in video-related visual tasks [50]. The main target of flow-based methods is to produce accurate estimates of the optical flow through different network architectures, such as utilizing bidirectional optical flow estimation and refining inter-frame coherence." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.317, + 0.382, + 0.334 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.784, + 0.375 + ], + "angle": 0, + "content": "In this section, we introduce basic definitions and notations, and review relevant techniques involving image histograms for use in BlazeBVD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.787, + 0.422 + ], + "angle": 0, + "content": "Image histogram is a graphical representation illustrating the distribution of pixels across different intensity levels within an image [3]. For a single-channel image \\(I\\) with height \\(H\\) and width \\(W\\), its histogram \\(\\mathcal{H}\\) can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.428, + 0.786, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} (\\lambda) = \\frac {n _ {\\lambda}}{H W}, \\quad i f \\lambda \\in I o t h e r w i s e 0, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.668, + 0.473 + ], + "angle": 0, + "content": "where \\(n_{\\lambda}\\) represents the count of pixels having the intensity \\(\\lambda\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Histogram matching or equalization [1, 5, 47, 51], is a technique to adjust the contrast of an image by transforming its histogram towards a predefined distribution. The cumulative histogram of an image histogram hist is defined by \\( \\text{Hist}(\\lambda) = \\Sigma_{r=0}^{\\lambda} \\mathcal{H}(r) \\). Give a target histogram \\( \\mathcal{H}' \\) and its cumulative histogram \\( \\text{Hist}'(\\lambda) \\), the histogram matching process can be described as" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.557, + 0.786, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} (\\lambda ; \\mathcal {H}, \\mathcal {H} ^ {\\prime}) = H i s t ^ {t - 1} (H i s t (\\lambda)), \\forall \\lambda \\in I, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.784, + 0.732 + ], + "angle": 0, + "content": "This procedure yields an image \\(\\tilde{X}\\), whose histogram closely aligns with \\(Hist'\\). Scale-Time Equalization (STE) is a filtering approach [8] to adjust the intensity distribution across both spatial and temporal dimensions, mitigating flickers encountered in video sequences. The classic scale-space theory posits that convolving an image with a Gaussian kernel results in smoothing and the loss of image structure. Intuitively, while STE is based on scale-space theory, it applies Gaussian smoothing to sequences of histograms instead of the images. Given a flickering video sequence \\(\\{X_{t}\\}_{t=1}^{T}\\), leveraging Eq. (1) can get the histogram sequence \\(\\{\\mathcal{H}_t\\}_{t=1}^T\\). For arbitrary time index \\(t\\) and intensity value \\(\\lambda\\), STE can be formulated as the convolution:" + }, + { + "type": "equation", + "bbox": [ + 0.262, + 0.738, + 0.786, + 0.779 + ], + "angle": 0, + "content": "\\[\nS T E (\\lambda , t) = G _ {s} (\\tau) * \\mathcal {Q} (\\lambda ; \\mathcal {H} _ {t}, \\mathcal {H} _ {\\tau}) \\approx \\sum_ {\\tau = t - l} ^ {t + l} G _ {s} (\\tau - t) \\mathcal {Q} (\\lambda ; \\mathcal {H} _ {t}, \\mathcal {H} _ {\\tau}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\(l\\) denotes the window radius, \\(G_{s}(t) = \\frac{1}{\\sqrt{4\\pi s}} e^{-\\frac{t^{2}}{4s}}\\). After STE, we will get a sequence of histograms \\(\\{\\tilde{\\mathcal{H}}_t\\}_{t=1}^T\\) and images \\(\\{\\tilde{X}_t\\}_{t=1}^T\\). Since histogram transformations do not create or cancel any image content, STE can achieve deflickering" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "X. Qiu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.782, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig. 2: The framework of our approach BlazeBVD. We first extract flicker prior information about the input video and correct the brightness representation with STE-assisted histogram filtering in illuminance space (Stage1). Then, the prior is leveraged to remove temporal flicker and over-/under-exposure flicker from the global and local perspectives (GFRM and LFRM in Stage2). Finally, the temporal consistency of the processed video is improved (TCM in Stage3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.522 + ], + "angle": 0, + "content": "results without compromising the integrity of the original video. However, this also indicates that STE cannot compensate for the lost texture in flickering regions and produces vivid effects, which impedes its practical application." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.381, + 0.563 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.787, + 0.653 + ], + "angle": 0, + "content": "Let \\(\\{X_{t}\\}_{t = 1}^{T}\\) represent a video sequence affected by flickering defects and \\(\\{G_t\\}_{t = 1}^T\\) represent the groundtruth clean video. In this section, we propose BlazeBVD as a universal solution to eliminate these flicker and produce a temporally consistent video \\(\\{O_t\\}_{t = 1}^T\\). It is important to note that the types of flicker addressed in our work can vary widely, involving long-/short-term flicker in color and brightness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.796, + 0.76 + ], + "angle": 0, + "content": "Distinguished from existing prior BVD approaches, BlazeBVD employs a histogram-assisted approach, leveraging deflickering priors from Scale-Time Equalization (STE) [8] to streamline the complexity and resource demands of BVD task. Beyond ensuring stable temporal performance, BlazeBVD stands out as the first method capable of faithfully restoring the color and texture affected by illumination fluctuations and exposure challenges, including both over-/under-exposure." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.455, + 0.798 + ], + "angle": 0, + "content": "4.1 Overview of BlazeBVD" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, the BlazeBVD pipeline is structured into three main stages to address video deflickering challenges effectively: (1) Deflickering priors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.301, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Fig. 3: Pipeline of Stage1. The illumination maps are corrected in the temporal dimension by STE, which alleviates the difficulty of subsequent temporal data learning using networks. The deflickering priors can be extracted: \\(\\tilde{V}_t\\), \\(M_t\\) and \\(S_{flicker}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.789, + 0.583 + ], + "angle": 0, + "content": "preparation. This initial stage leverages STE to correct the histogram within an illumination space representation. Key outputs from this stage include the filtered illumination map \\(\\{\\tilde{V}_t\\}_{t=1}^T\\), identifying the index set of singular frames \\(S_{flicker}\\), and generating exposure maps \\(\\{M_t\\}_{t=1}^T\\). These elements serve as crucial prior information, guiding the deflickering process in subsequent stages. (2) Global and local flicker removal. The second stage involves two flicker removal strategies. A global flicker removal module utilizes the filtered illumination map \\(\\{\\tilde{V}_t\\}_{t=1}^T\\) to address temporal flicker issues across the video. A local over-/under-exposed region flicker removal module instructed by the singular frames set \\(S_{flicker}\\) and exposure maps \\(\\{M_t\\}_{t=1}^T\\), restores the texture in the over-/under-exposure region to compensate for the lost high-frequency details. (3) Adaptive temporal consistency. The final stage introduces a spatio-temporal network under the guidance of an adaptive warping loss to improve video coherence and temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.63 + ], + "angle": 0, + "content": "These stages delineated in the following Sec. 4.2, Sec. 4.3, and Sec. 4.4, together provide a comprehensive solution to enhance the visual quality and temporal consistency of video content." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.582, + 0.667 + ], + "angle": 0, + "content": "4.2 Stage1: Deflickering priors preparation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Spatio-temporal challenge. An effective BVD model should strategically avoid the intricacies of spatio-temporal complexity across pixels, nevertheless, current deep learning approaches to BVD have not adequately addressed this challenge. For instance, the state-of-the-art BVD method, Deflicker [21], employs a neural atlas [19] as a unified representation of the entire video. While this can offer consistent guidance for the deflickering process, the effectiveness of neural atlases significantly depends on the video duration and the network capacity. Processing an 8-second video to create an atlas can take around 10 minutes, and longer videos exacerbate the demands on time, memory, and computational resources. Consequently, there is a further need for research into more compact representations that can facilitate deflickering without these limitations." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "X. Qiu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Illumination histogram. We introduce histogram representation [3], which utilizes statistical distribution to illustrate the global properties of image pixel values. As illustrated in Fig. 3, we first extract the illumination maps \\(\\{V_t\\}_{t=1}^T\\) according to the conversion between HSV and RGB color space:" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.215, + 0.786, + 0.238 + ], + "angle": 0, + "content": "\\[\nV _ {t} = \\max _ {R, G, B \\in X _ {t}} \\{R, G, B \\}, t \\in \\{1, 2, \\dots , T \\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.784, + 0.379 + ], + "angle": 0, + "content": "Next, by applying Eq. (1) to each illumination map, we utilize illumination histograms \\(\\{H_t\\}_{t=1}^T\\) to represent the illumination characteristics of video frames. These illumination histograms reflect important flickering information: As shown in Fig. 3, frames without distinct illumination changes have similar histograms, while the histogram of a flicker frame has a distribution distinguished from others. Furthermore, spikes in the head or tail area of the distribution indicate potential over-/under-exposure problems in the image. To this end, reasonable utilization of histogram information is significant for flicker elimination and alleviation of spatio-temporal complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.531 + ], + "angle": 0, + "content": "Deflickering priors. We introduce deflickering priors extracted from illumination histograms to facilitate the deflickering process. Initially, considering that the illumination histograms of natural videos exhibit continuous changes in the temporal dimension, we utilize Scale-Time Equalization (STE) as defined in Eq. (3) to smooth the illumination histograms, as illustrated in Fig. 3. Through histogram matching, we further obtain filtered illumination maps \\(\\{\\tilde{V}\\}_{t=1}^{T}\\) that exhibit no obvious flicker compared to the original ones, thereby enabling the subsequent independent processing of each frame by the neural network. Secondly, we identify the singular frames set \\(S_{flicker}\\) containing the indexes of flicker frames, based on the divergence of histograms before and after STE:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.538, + 0.785, + 0.577 + ], + "angle": 0, + "content": "\\[\nS _ {f l i c k e r} = \\left\\{t \\in [ 1, T ] \\mid \\mathrm {K L} \\left(\\tilde {H} _ {t} \\parallel H _ {t}\\right) > \\bar {\\epsilon} _ {t} \\right\\}, \\quad \\bar {\\epsilon} _ {t} = \\frac {1}{2 n + 1} \\sum_ {\\tau = t - n} ^ {t + n} \\mathrm {K L} \\left(\\tilde {H} _ {\\tau} \\parallel H _ {\\tau}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.784, + 0.629 + ], + "angle": 0, + "content": "where \\( n \\) is the moving average radius. Finally, as shown in Fig. 2, we extract the exposure maps \\( \\{M_t \\in \\mathbb{R}^{H \\times W}\\}_{t=1}^T \\) of local over-/under-exposure according to the filtered illumination map:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.636, + 0.785, + 0.675 + ], + "angle": 0, + "content": "\\[\nM _ {t} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\tilde {V} (i, j) < \\epsilon_ {1} \\text {o r} \\tilde {V} (i, j) > \\epsilon_ {2}, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.783, + 0.71 + ], + "angle": 0, + "content": "Here, \\(\\epsilon_{1}\\) and \\(\\epsilon_{2}\\) are darkness and exposure thresholds, respectively, and we consider illumination values belonging to \\([\\epsilon_1,\\epsilon_2]\\) as not posing an exposure risk." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.784, + 0.77 + ], + "angle": 0, + "content": "In summary, we leverage illumination histograms to prepare deflickering priors including filtered illumination maps \\(\\{\\tilde{V}\\}_{t=1}^{T}\\), exposure maps \\(\\{M_t\\}_{t=1}^T\\), and the singular frames set \\(S_{flicker}\\), for alleviating the spatio-temporal complexity and enhancing deflickering in the subsequent learning stages." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.589, + 0.804 + ], + "angle": 0, + "content": "4.3 Stage2: Global and local flicker removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this stage, simple 2D spatial neural networks are integrated with deflickering priors to eliminate both global and local flickering flaws. Here, \"global\" refers" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "to obvious temporal inconsistency due to illumination fluctuation, while \"local\" refers to lost texture details due to exposure and darkness. This approach allows for the direct generation of high-quality videos from flicker frames, regardless of the spatio-temporal complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Global deflickering. BlazeBVD is equipped with a global flicker removal module (GFRM), which incorporates a 2D-Unet to correct each frame \\( X_{t} \\) with the guidance of the filtered illumination map \\( \\tilde{V}_{t} \\). In other words, the corrected image \\( \\hat{X}_{t} = U\\mathrm{net}_{\\theta}(X_{t},\\tilde{V}_{t}) \\), where \\( \\theta \\) denotes the parameters of the 2D-Unet. The network is trained on the BVD dataset \\( \\mathcal{D} \\) by minimizing the L2 loss function between the predicted images and the groundtruth images:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.309, + 0.785, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\mathbb {E} _ {\\left(X _ {t}, G _ {t}\\right) \\in \\mathcal {D}} \\| U n e t _ {\\theta} \\left(X _ {t}, \\tilde {V} _ {t}\\right) - G _ {t} \\|. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.785, + 0.471 + ], + "angle": 0, + "content": "Local deflickering. BlazeBVD is also equipped with a local flicker removal module (LFRM) to enhance local texture details. As depicted in Fig. 5, overexposure can transform intricate texture in local regions into uniform pure, while local under-exposure can result in black areas. This loss of detailed texture cannot be compensated for by previous BVD methods, such as Deflicker [21]. We leverage LFRM to transfer detailed information from neighboring frames to these ill regions. Given the corrected singular frame \\(\\tilde{X}_t\\) in \\(S_{\\text{flicker}}\\) with local exposure, along with the previous frame \\(\\tilde{X}_{t-1}\\) and the next frame \\(\\tilde{X}_{t+1}\\), we first compute the optical flow to extract part of the detailed texture:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.482, + 0.785, + 0.498 + ], + "angle": 0, + "content": "\\[\no _ {\\tau \\rightarrow t} = \\operatorname {F l o w} \\left(X _ {\\tau}, X _ {t}\\right), \\tau \\in \\{t + 1, t - 1 \\} \\text {a n d} t \\in S _ {\\text {f l i c k e r}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.784, + 0.552 + ], + "angle": 0, + "content": "Here Flow is the optical flow estimation function, in our work we set it as the popular RAFT model [11, 37]. We migrate the possible detail textures of the preceding and following frames into the current frame:" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.563, + 0.785, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {t} = \\mathcal {F} \\left(M _ {t} \\odot \\mathcal {W} \\left(\\tilde {X} _ {t - 1}, o _ {t - 1 \\rightarrow t}\\right) + (1 - M _ {t}) \\odot \\tilde {X} _ {t}, \\tilde {X} _ {t}, \\right. \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.584, + 0.723, + 0.601 + ], + "angle": 0, + "content": "\\[\nM _ {t} \\odot \\mathcal {W} \\left(\\tilde {X} _ {t - 1}, o _ {t + 1 \\rightarrow t}\\right) + \\left(1 - M _ {t}\\right) \\odot \\tilde {X} _ {t}), t \\in S _ {\\text {f l i c k e r}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.734, + 0.626 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}\\) is the warping operator and \\(\\mathcal{F}\\) is the fusion network from [41]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.647, + 0.582, + 0.662 + ], + "angle": 0, + "content": "4.4 Stage3: Adaptive temporal consistency" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.785, + 0.774 + ], + "angle": 0, + "content": "The former stages have well addressed the main illumination fluctuation and exposure problems in BVD task. Nevertheless, a video may also have content inconsistencies to largely degrade the visual effect, and the former stage may also cause this. Therefore, We introduce the Temporal Consistency Model (TCM) to refine video \\(\\{\\hat{X}_t\\}_{t=1}^T\\) as the final outputs \\(\\{O_t\\}_{t=1}^T\\). TCM is developed based on the architecture by RTN [40], but with a temporal consistency loss of adaptive masks to refine local artifacts:" + }, + { + "type": "equation", + "bbox": [ + 0.278, + 0.784, + 0.707, + 0.8 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {p a i r}} ^ {\\prime} \\left(O _ {t}, O _ {s}\\right) = \\left\\| W _ {t} \\odot \\left(M _ {t} + 1\\right) \\odot M _ {t, s} \\odot \\left(O _ {t} - \\mathcal {W} \\left(O _ {s}\\right)\\right) \\right\\| _ {1},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.805, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w a r p} = \\frac {1}{T - 1} \\sum_ {t = 2} ^ {T} \\left\\{E _ {\\text {p a i r}} ^ {\\prime} \\left(O _ {t}, O _ {1}\\right) + E _ {\\text {p a i r}} ^ {\\prime} \\left(O _ {t}, O _ {t - 1}\\right) \\right\\}, \\tag {10}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "X. Qiu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "where \\( W_{t} \\) is the loss weight for the over-/under-exposure area and \\( M_{t} \\) is the mask obtained from the former stage. \\( \\mathcal{W} \\) is the warping operator and the frame after warping is computed by optical flow estimation [11,37]. \\( M_{t,s} \\) represents the mask after optical flow estimation. This loss avoids the video blurring problem introduced in Deflicker [21]. The training loss of TCM consists of reconstruction loss \\( \\mathcal{L}_{rec} \\), perceptual loss \\( \\mathcal{L}_{per} \\) [17], spatio-temporal adversarial loss \\( \\mathcal{L}_{adv} \\) [6] and designed adaptive weighted warping loss \\( \\mathcal{L}_{warp} \\). Combining all the above losses, the overall objective to optimize TCM is:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.282, + 0.786, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T C M} = \\alpha_ {1} * \\mathcal {L} _ {r e c} + \\alpha_ {2} * \\mathcal {L} _ {p e r} + \\alpha_ {3} * \\mathcal {L} _ {a d v} + \\alpha_ {4} * \\mathcal {L} _ {w a r p}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.309, + 0.677, + 0.325 + ], + "angle": 0, + "content": "In our work, we set \\(\\alpha_{1} = 1.0\\), \\(\\alpha_{2} = 1.0\\), \\(\\alpha_{3} = 0.01\\) and \\(\\alpha_{4} = 0.1\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.348, + 0.377, + 0.367 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.381, + 0.336, + 0.395 + ], + "angle": 0, + "content": "5.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.787, + 0.451 + ], + "angle": 0, + "content": "We inherit the training dataset DAVIS-2017-Train, MS-COCO and the testing dataset Blind Deflickering Dataset from Deflicker [21], including Synthetic videos from DAVIS-2017-Val, Real-world videos and Generation videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Training dataset. MS-COCO [23] contains 118287 images in the train set and is used for the pre-training of Unet in GFRM. DAVIS-2017-Train [20, 28, 32] consists of 60 sequences with a total of 4219 labeled frames, with a video frame rate of 24fps and a resolution of \\(480\\mathrm{p}\\). We create the synthetic process that provides the underlying facts for quantitative analysis. The training dataset is degraded by performing flicker on DAVIS-2017-Train, as shown in Eq. (12):" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.557, + 0.786, + 0.573 + ], + "angle": 0, + "content": "\\[\nX _ {t} = G _ {t} + F _ {t}, t \\in \\{1, 2, \\dots , T \\}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.644 + ], + "angle": 0, + "content": "where \\(\\{G_t\\}_{t=1}^T\\) are clean frames and \\(\\{F_t\\}_{t=1}^T\\) are random synthesized flickering artifacts. The window size \\(W\\) in \\(F_t\\) has been set to denote the number of frames that share the same flickering artifacts, random sampling between [2, 12]. The value of \\(W\\) controls the degree of flicker in the short- and long-term of the video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Testing dataset. We conduct Synthetic videos, Real-world videos and Generation videos respectively, from the Blind Deflickering Dataset [21] and DAVIS2017-Test processed by Eq. (12). Among them, Synthetic videos and DAVIS2017-Test have ground truth and the others do not. We can only use no-referenced metric and qualitative observation. Generation videos refer to the processed videos generated by generative algorithms, which may lead to flicker artifacts due to the imperfections of the algorithms. Real-world videos including old movies, old anime, slow-motion, etc. are directly captured, unprocessed videos in the real world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison methods. We compare our BlazeBVD with other 4 methods, including STE [8], ConvLSTM [20], DVP [22] and Deflicker [21]. We set Deflicker as the baseline, since it is the first publicly presented method for BVD task and is currently the most advanced approach for general videos." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.524, + 0.788, + 0.608 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons between previous methods and our BlazeBVD. Our method removes flicker and restores details in over-exposed regions (red boxes in row 6 and row 8) while avoiding color artifacts (row 2 and row 4). Besides, BlazeBVD also ensures the fidelity of the video content and avoids color distortion (cyan boxes in row 5 and row 7). Zoom in for the best view and we recommend watching videos in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.623, + 0.651 + ], + "angle": 0, + "content": "5.2 Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Metrics. We quantitatively compare the differences between our method and others with reference and no-reference metrics, including PSNR, SSIM [45], and \\( E_{\\text{warp}} \\) [22]. Among them, \\( E_{\\text{warp}} \\) is a no-reference metric used in videos. For each frame \\( O_t \\), we calculate the warping error with frame \\( O_{t-1} \\) and the first frame \\( O_1 \\) for considering both short-term and long-term consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Quantitative results. We compare our BlazeBVD with the state-of-the-art method Deflicker [21], and other consistency methods. Table 1 and Table 2 provide the comparative experimental results of various methods on DAVIS-2017-Test and Synthetic videos, where PSNR and SSIM of the resulting video are higher than those of the other methods. This shows that BlazeBVD contributes to the fidelity maintenance of video content, as well as the correctness of the missing details complement. The warping error of our method is a little higher" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "X. Qiu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons of synthetic videos in DAVIS-2017-Test. Our method performs best in terms of PSNR and SSIM and \\( E_{\\text{warp}} \\) of all the methods." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.185, + 0.787, + 0.259 + ], + "angle": 0, + "content": "
MethodRawVideoConvLSTMDVPSTEDeflickerOursGT
PSNR↑-24.11024.13626.13823.93228.609-
SSIM↑-0.92560.92630.93210.92430.9638-
E_warp ↓0.14290.13690.14150.11170.08400.08250.1097
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.316 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons in Synthetic videos from Blind Deflickering Dataset. \\( W \\) is the window size and represents the number of frames that share the same flicker artifact. \\( L \\) denotes the local seed window size." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.328, + 0.789, + 0.464 + ], + "angle": 0, + "content": "
VideoTypeRawE_warp↓PSNR↑
DeflickerOursGTDeflickerOursDeflickerOurs
Synthetic
-W=10.19330.08480.08950.105625.931229.67380.94030.9539
-W=30.15220.08440.08900.105625.499628.82720.93800.9555
-W=100.13070.08550.09040.105624.384226.99660.93440.9543
-L=30.11130.08750.09000.105625.387427.08810.92710.9517
Average0.14690.08560.08970.105625.300628.14640.93490.9538
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.657 + ], + "angle": 0, + "content": "than Deflicker, but it is closer to the ground truth. Through analysis, we find that warping error has inaccuracy in the estimation of optical flow, and this error is masked by image blurring in Deflicker. Therefore, we believe that \\( E_{\\text{warp}} \\), which has the same magnitude as baseline and GT, is the valid experimental comparison result. In Tab. 3, we compare BlazeBVD and Deflicker on Real-world videos and Generation videos. Since the videos in these two datasets with no reference are inherently discontinuous, where \\( E_{\\text{warp}} \\) cannot directly represent video continuity, we present the comparative results of user studies. To evaluate the perceived preference between the baseline Deflicker and the proposed method, each user needs to select a video with better perceptual quality between these two choices. In total, there are 50 users and 50 pairs in 7 datasets of comparisons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Qualitative results. Both ConvLSTM and DVP are temporal consistency models that require a reference, importing the input video as the reference, and the processed video cannot remove significant flicker, as shown in Fig. 4. Additionally, the flicker caused by over-/under-exposure is prone to lose the local texture (column 1 in Fig. 4). The atlas-based representation in Deflicker cannot accurately locate these points with different local features, resulting in the loss of all details on the processed video. Instead, our method designs local positioning and adjacent frame details transfer in LFRM to preserve part of the texture on the road, as shown in the red box of Fig. 4. We also reveal the color artifact in column 4 of Fig. 4 that appears at baseline, due to the excessive weight of warping error in the training loss when refining temporal consistency. Therefore, our designed adaptive mask weighted training loss in TCM improves the temporal" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: User studies of Real-world videos and Generation videos. Since the videos in these two datasets are inherently discontinuous, \\( E_{\\text{warp}} \\) cannot properly reflect temporal consistency. Here we only present the comparative results of the user study." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.784, + 0.273 + ], + "angle": 0, + "content": "
TypeExpertOldAnimeOldMovieSlowMotionTimeLapseVideoDMVideoLDM
Raw(E_warp)0.07540.10450.10990.07240.07320.12480.1601
Deflicker38.67%19.25%22.83%37.67%31.43%30.50%26.34%
Ours61.33%80.75%77.17%62.33%68.57%69.50%73.66%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.784, + 0.332 + ], + "angle": 0, + "content": "consistency while avoiding image blurring. More visualization results are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.356, + 0.402, + 0.37 + ], + "angle": 0, + "content": "5.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.472 + ], + "angle": 0, + "content": "STE filtering. We contrast the difference between introducing STE filtering in illumination space and directly in color space. As shown in the red box of Fig. 5, directly performing histogram correction in color channels separately introduces color artifacts and color distortion. Because the filtering of non-gray images in the three channels is relatively independent, the recombination after nonlinear transformation would produce such defects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.685 + ], + "angle": 0, + "content": "GFRM and LFRM. We compare the important roles of two modules, GFRM and LFRM. As shown in Tab. 5, removing GFRM and LFRM reduces PSNR and SSIM and increases \\( E_{\\text{warp}} \\) of inference videos. It can be inferred that GFRM needs to initially remove global flicker between frames, otherwise subsequent modules cannot maintain the temporal consistency, while LFRM maintains the texture fidelity of the video content, which is reflected in the cyan box of Fig. 5. Designed weighted loss. As shown in Tab. 5, the lightweight spatio-temporal network TCM reduces \\( E_{\\text{warp}} \\), which refines the temporal consistency of processed videos. We also compare the effect of using unweighted training loss and the adaptive mask weighted loss on this temporal network. Figure 5 provides the resulting frame with unweighted training, which is generally blurred and suffers from color distortion, while the weighted training keeps the frame clear and faithful in most areas. Adaptive mask weighting improves the temporal consistency of the inference video, improving the inter-frame continuity more pertinently." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.35, + 0.723 + ], + "angle": 0, + "content": "5.4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Advantages and limitations. BlazeBVD is over \\(10 \\times\\) faster than the inference speed of baseline using the atlas representation, which is specifically recorded in Tab. 4. Our method also takes advantage of explicit prior extraction in the BVD task, which provides directions for improvement in the subsequent work. And the cause of local flicker has been noted and effectively dealt with from both global and local perspectives. However, our method still has drawbacks in LFRM. Due to the inaccurate optical flow motion estimation, the fusion network" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "X. Qiu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.148, + 0.454, + 0.23 + ], + "angle": 0, + "content": "Table 4: Inference time and model parameters comparison. Our approach is faster than the baseline in inference time with less memory, and comparable in MACs and Params." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.244, + 0.453, + 0.33 + ], + "angle": 0, + "content": "
MethodDeflickerOurs
Time/80p614.19s58.37s
Memory5204M1389M
MACs260.7G251.4G
Params12.48M17.77M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.47, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 5: Ablation studies in DAVIS-2017- Test. w/o STE replaces STE with mean filtering in Stage1 and w/o weighted denotes the warping loss without mask weighting in TCM." + }, + { + "type": "table", + "bbox": [ + 0.47, + 0.213, + 0.8, + 0.334 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑E_warp ↓
w/o STE26.44120.95170.1011
w/o GFRM23.89390.92440.1225
w/o LFRM27.13660.94620.1122
w/o TCM29.91300.95830.1093
w/o weighted28.10490.95230.0849
BlazeBVD28.60920.96380.0825
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.349, + 0.358, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.412, + 0.303, + 0.424 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.349, + 0.5, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.412, + 0.466, + 0.424 + ], + "angle": 0, + "content": "w/o GFRM" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.349, + 0.643, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.412, + 0.607, + 0.423 + ], + "angle": 0, + "content": "w/o LFRM" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.349, + 0.786, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.412, + 0.737, + 0.423 + ], + "angle": 0, + "content": "w/o STE" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.425, + 0.358, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.488, + 0.317, + 0.499 + ], + "angle": 0, + "content": "w/o TCM" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.425, + 0.5, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.488, + 0.488, + 0.501 + ], + "angle": 0, + "content": "w/o weighted mask" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.425, + 0.643, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.488, + 0.582, + 0.499 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.425, + 0.786, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.488, + 0.724, + 0.498 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.527, + 0.789, + 0.584 + ], + "angle": 0, + "content": "Fig. 5: Qualitative ablation studies of key designs in BlazeBVD. Specifically, local detail loss is represented in cyan box, color artifact is represented in the red box, temporal inconsistency is represented in the black box, and color distortion is represented in the blue box. Zoom in for the best view." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.617, + 0.788, + 0.664 + ], + "angle": 0, + "content": "cannot convey the local texture of adjacent frames, and there are still slight edge artifacts and color distortion that cannot be eliminated. We hope to continue exploring the transfer and fusion of local region information between frames [40]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.36, + 0.71 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose BlazeBVD, a universal approach for the blind video deflickering task. The core of our method is to prepare flicker priors within the STE filter in illumination space first, and then utilize these priors to correct the global flicker and local exposed texture. GFRM and LFRM leverage the priors to remove global and local flicker in several adjacent frames with less computation. Finally, TCM is used to improve video coherence and inter-frame consistency. Extensive experiments validate the effectiveness of the designed modules." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.178, + 0.788, + 0.224 + ], + "angle": 0, + "content": "This paper is supported by the National Natural Science Foundation of China (Nos. U23B2012, 11991022), and the Fundamental Research Funds for the Central Universities (E3E41904X2)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.248, + 0.323, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.31 + ], + "angle": 0, + "content": "1. Affifi, M., Brubaker, M.A., Brown, M.S.: Histogram: Controlling colors of gangenerated and real images via color histograms. In: CVPR. pp. 7941-7950 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.757, + 0.323 + ], + "angle": 0, + "content": "2. Anarchy, D.: Flicker free, https://digitalanarchy.com/Flicker/main.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.786, + 0.365 + ], + "angle": 0, + "content": "3. Bassiou, N., Kotropoulos, C.: Color image histogram equalization by absolute discounting back-off. Computer Vision and Image Understanding 107, 108-122 (Jul 2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.786, + 0.394 + ], + "angle": 0, + "content": "4. Bonneel, N., Tompkin, J., Sunkavalli, K., Sun, D., Paris, S., Pfister, H.: Blind video temporal consistency. ACM Transactions on Graphics 34, 1-9 (Nov 2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.786, + 0.435 + ], + "angle": 0, + "content": "5. Bottenus, N., Byram, B.C., Hyun, D.: Histogram matching for visual ultrasound image comparison. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control 68, 1487-1495 (May 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.786, + 0.464 + ], + "angle": 0, + "content": "6. Chang, Y.L., Liu, Z.Y., Lee, K.Y., Hsu, W.: Free-form video inpainting with 3d gated convolution and temporal patchgan. In: ICCV. pp. 9066-9075 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.786, + 0.506 + ], + "angle": 0, + "content": "7. Chu, M., Xie, Y., Mayer, J., Leal-Taixe, L., Thuerey, N.: Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics 39, 75-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.786, + 0.533 + ], + "angle": 0, + "content": "8. Delon, J.: Movie and video scale-time equalization application to flicker reduction. IEEE Transactions on Image Processing 15, 241-248 (Jan 2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.786, + 0.575 + ], + "angle": 0, + "content": "9. Delon, J., Desolneux, A.: Stabilization of flicker-like effects in image sequences through local contrast correction. SIAM Journal on Imaging Sciences 3, 703-734 (Jan 2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.786, + 0.617 + ], + "angle": 0, + "content": "0. Dosovitskiy, A., Fischer, P., Ilg, E., Hausser, P., Hazirbas, C., Golkov, V., Van Der Smagt, P., Cremers, D., Brox, T.: Flownet: Learning optical flow with convolutional networks. In: ICCV. pp. 2758-2766 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.786, + 0.645 + ], + "angle": 0, + "content": "1. Eslami, N., Arefi, F., Mansourian, A.M., Kasaei, S.: Rethinking raft for efficient optical flow. arXiv preprint arXiv:2401.00833 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.786, + 0.686 + ], + "angle": 0, + "content": "2. Huang, J., Liu, Y., Fu, X., Zhou, M., Wang, Y., Zhao, F., Xiong, Z.: Exposure normalization and compensation for multiple-exposure correction. In: CVPR. pp. 6043-6052 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.786, + 0.728 + ], + "angle": 0, + "content": "3. Huang, J., Liu, Y., Zhao, F., Yan, K., Zhang, J., Huang, Y., Zhou, M., Xiong, Z.: Deep fourier-based exposure correction network with spatial-frequency interaction. In: ECCV. pp. 163-180 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.786, + 0.756 + ], + "angle": 0, + "content": "4. Huang, J., Zhao, F., Zhou, M., Xiao, J., Zheng, N., Zheng, K., Xiong, Z.: Learning sample relationship for exposure correction. In: CVPR. pp. 9904-9913 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.798 + ], + "angle": 0, + "content": "5. Huang, J., Zhou, M., Liu, Y., Yao, M., Zhao, F., Xiong, Z.: Exposure-consistency representation learning for exposure correction. In: ACM MM. pp. 6309-6317 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "6. Ilg, E., Mayer, N., Saikia, T., Keuper, M., Dosovitskiy, A., Brox, T.: Flownet 2.0: Evolution of optical flow estimation with deep networks. In: CVPR. pp. 2462-2470 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.282, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "X. Qiu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV. pp. 694-711 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "18. Kanj, A., Talbot, H., Luparello, R.R.: Flicker removal and superpixel-based motion tracking for high speed videos. In: ICIP. pp. 245-249 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.23 + ], + "angle": 0, + "content": "19. Kasten, Y., Ofri, D., Wang, O., Dekel, T.: Layered neural atlases for consistent video editing. ACM Transactions on Graphics 40, 1-12 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "20. Lai, W.S., Huang, J.B., Wang, O., Shechtman, E., Yumer, E., Yang, M.H.: Learning blind video temporal consistency. In: ECCV. pp. 170-185 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.283 + ], + "angle": 0, + "content": "21. Lei, C., Ren, X., Zhang, Z., Chen, Q.: Blind video deflickering by neural filtering with a flawed atlas. In: CVPR. pp. 10439-10448 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.31 + ], + "angle": 0, + "content": "22. Lei, C., Xing, Y., Chen, Q.: Blind video temporal consistency via deep video prior. In: NeurIPS. pp. 1083-1093 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.351 + ], + "angle": 0, + "content": "23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.378 + ], + "angle": 0, + "content": "24. Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: CVPR. pp. 5637-5646 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.404 + ], + "angle": 0, + "content": "25. Mei, K., Patel, V.: Vidm: Video implicit diffusion models. In: AAAI. pp. 9117-9125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.432 + ], + "angle": 0, + "content": "26. Moniz, J.R.A., Kang, E., Póczos, B.: Luciddream: Controlled temporally-consistent deepdream on videos. arXiv preprint arXiv:1911.11960 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.472 + ], + "angle": 0, + "content": "27. Park, K., Woo, S., Kim, D., Cho, D., Kweon, I.S.: Preserving semantic and temporal consistency for unpaired video-to-video translation. In: ACM MM. pp. 1248-1257 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "28. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: CVPR (Jun 2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.555 + ], + "angle": 0, + "content": "29. Pérez-Pellitero, E., Sajjadi, M.S., Hirsch, M., Scholkopf, B.: Perceptual video super resolution with enhanced temporal consistency. arXiv preprint arXiv:1807.07930 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.582 + ], + "angle": 0, + "content": "30. Pfeuffer, A., Dietmayer, K.: Separable convolutional lstms for faster video segmentation. In: ITSC. pp. 1072-1078 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "31. Pitie, F., Kokaram, A., Dahyot, R.: Removing flicker from old movies. Master's thesis, University of Nice-Éophia Antipolis, France, Œe September (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.65 + ], + "angle": 0, + "content": "32. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv:1704.00675 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.691 + ], + "angle": 0, + "content": "33. Qiu, X., Han, C., Zhang, Z., Li, B., Guo, T., Nie, X.: Diffbfr: Bootstrapping diffusion model for blind face restoration. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7785-7795 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.73, + 0.704 + ], + "angle": 0, + "content": "34. RE:VISION: De:flicker, https://revisionfx.com/products/deflicker/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.745 + ], + "angle": 0, + "content": "35. Saito, M., Saito, S., Koyama, M., Kobayashi, S.: Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. International Journal of Computer Vision p. 2586-2606 (Nov 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.786 + ], + "angle": 0, + "content": "36. Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In: CVPR. pp. 3626-3636 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.787, + 0.785, + 0.813 + ], + "angle": 0, + "content": "37. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402-419 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Thasarathan, H., Nazeri, K., Ebrahimi, M.: Automatic temporally coherent video colorization. In: CRV. pp. 189-194 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "39. Thimonier, H., Despois, J., Kips, R., Perrot, M.: Learning long-term style preserving blind video temporal consistency. In: ICME. pp. 1-6 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "40. Wan, Z., Zhang, B., Chen, D., Liao, J.: Bringing old films back to life. In: CVPR. pp. 17694-17703 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.231 + ], + "angle": 0, + "content": "41. Wang, H., Xu, K., Lau, R.W.: Local color distributions prior for image enhancement. In: ECCV. pp. 343-359 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "42. Wang, W., Liu, J., Lin, Z., Yan, J., Chen, S., Low, C., Hoang, T., Wu, J., Liew, J.H., Yan, H., et al.: Magicvideo-v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.313 + ], + "angle": 0, + "content": "43. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9168-9178 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.342 + ], + "angle": 0, + "content": "44. Wang, Y., Peng, L., Li, L., Cao, Y., Zha, Z.J.: Decoupling-and-aggregating for image exposure correction. In: CVPR. pp. 18115-18124 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.383 + ], + "angle": 0, + "content": "45. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13, 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.424 + ], + "angle": 0, + "content": "46. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV. vol. 39, pp. 7623-7633 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.452 + ], + "angle": 0, + "content": "47. Xu, R., Tu, Z., Du, Y., Dong, X., Li, J., Meng, Z., Ma, J., Bovik, A., Yu, H.: Pik-fix: Restoring and colorizing old photos. In: WACV. pp. 1724-1734 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.493 + ], + "angle": 0, + "content": "48. Xu, X., Li, M., Sun, W., Yang, M.H.: Learning spatial and spatio-temporal pixel aggregations for image and video denoising. IEEE Transactions on Image Processing 29, 7153-7165 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.535 + ], + "angle": 0, + "content": "49. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 672-681 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.563 + ], + "angle": 0, + "content": "50. Zhan, X., Pan, X., Liu, Z., Lin, D., Loy, C.C.: Self-supervised learning via conditional motion propagation. In: CVPR. pp. 1881-1889 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.604 + ], + "angle": 0, + "content": "51. Zhang, F., Shao, Y., Sun, Y., Zhu, K., Gao, C., Sang, N.: Unsupervised low-light image enhancement via histogram equalization prior. arXiv preprint arXiv:2112.01766 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.646 + ], + "angle": 0, + "content": "52. Zhang, Z., Li, B., Nie, X., Han, C., Guo, T., Liu, L.: Towards consistent video editing with text-to-image diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.688 + ], + "angle": 0, + "content": "53. Zhou, Y., Xu, X., Shen, F., Gao, L., Lu, H., Shen, H.T.: Temporal denoising mask synthesis network for learning blind video temporal consistency. In: ACM MM. pp. 475-483 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_origin.pdf b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..87958292054e8335e9693785281443676337e9fc --- /dev/null +++ b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/cea7a8fd-3be9-45e1-99c6-df772c59c608_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95cb25ae1678bb3caf85a95558741199553375087486eb86f5704416dd6027ca +size 3938553 diff --git a/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/full.md b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ba0c9a533bbc6d841fabd9c91f4904138e559643 --- /dev/null +++ b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/full.md @@ -0,0 +1,345 @@ +# BlazeBVD: Make Scale-Time Equalization Great Again for Blind Video Deflickering + +Xinmin Qiu $^{1}$ , Congying Han $^{1}$ , Zicheng Zhang $^{1}$ , Bonan Li $^{1}$ , Tiande Guo $^{1}$ , Pingyu Wang $^{2}$ , and Xuecheng Nie $^{3}$ + +1 University of Chinese Academy of Sciences, Beijing, China + +$^{2}$ Sichuan University, Sichuan, China + +3 MT Lab, Meitu Inc., Beijing, China + +{qiuxinmin21,zhangzicheng19}@mails.ucas.ac.cn, + +{hancy,libonacci,tdguo}@ucas.ac.cn + +wangpingyu@scu.edu.cn, nxc@meitu.com + +![](images/272906ba01293ee2a957fdc94b81aefe058b99b6cb2639e6fc31e94781068bcc.jpg) +Input + +![](images/2fdfc44e33dbeee4131ea7b7906a2ded145613c3a808b7f646bbe7a0aae299cc.jpg) +Deflicker + +![](images/dc5085a5bafb6366d59691c93e040444202f2bc7010fa436628d37d269904027.jpg) +Ours +Fig. 1: Comparisons of the proposed BlazeBVD. We present flickering input, GT, Deflicker [21] and our BlazeBVD processed video frames, and illumination histograms along with KL divergence about GroundTruth. Our method recovers the illumination histograms well while avoiding the appearance of color artifacts and color distortions (such as the man's arm in the second column). Better see in color with $2 \times$ zoom. + +![](images/070bce0cc1aea6af146eebeb3bc234a883b6a86d1e8c442574bf12724f40bfeb.jpg) +GT + +Abstract. Developing blind video deflickering (BVD) algorithms to enhance video temporal consistency, is gaining importance amid the flourish of image processing and video generation. However, the intricate nature of video data complicates the training of deep learning methods, leading to high resource consumption and instability, notably under severe lighting flicker. This underscores the critical need for a compact representation beyond pixel values to advance BVD research and applications. Inspired by the classic scale-time equalization (STE), our work introduces the histogram-assisted solution, called BlazeBVD, for high-fidelity and rapid BVD. Compared with STE, which directly corrects pixel values by temporally smoothing color histograms, BlazeBVD + +leverages smoothed illumination histograms within STE filtering to ease the challenge of learning temporal data using neural networks. In technique, BlazeBVD begins by condensing pixel values into illumination histograms that precisely capture flickering and local exposure variations. These histograms are then smoothed to produce singular frames set, filtered illumination maps, and exposure maps. Resorting to these deflickering priors, BlazeBVD utilizes a 2D network to restore faithful and consistent texture impacted by lighting changes or localized exposure issues. BlazeBVD also incorporates a lightweight 3D network to amend slight temporal inconsistencies, avoiding the resource consumption issue. Comprehensive experiments on synthetic, real-world and generated videos, showcase the superior qualitative and quantitative results of BlazeBVD, achieving inference speeds up to $10 \times$ faster than state-of-the-arts. + +Keywords: Video de-flicker $\cdot$ Histogram $\cdot$ Temporal consistency + +# 1 Introduction + +As social applications and the multimedia industry continue to expand, videos have become an essential medium for conveying information in daily life [30]. Achieving a high-quality video demands both image clarity and temporal consistency, nevertheless, video flickering frequently compromises the temporal consistency, often resulting from shooting environment and camera hardware limitations [9, 18]. This issue is further exacerbated when image processing techniques are applied to video frames. Moreover, flickering artifacts and color distortion are also prevalent in recent video generation technologies, including those based on Generative Adversarial Networks (GANs) [7,35,36] and Diffusion Models (DMs) [25,42,46,52]. Since even certain types of flicker can significantly detract from the viewing experience, it is crucial to develop video deflickering techniques to eliminate flicker and preserve the integrity of video content given unknown degradation in various video processing scenarios. + +Recovering visual content from unknown degradation is a key yet challenging topic [33, 43, 49]. Several seminal works [8, 20, 21] have addressed the blind video deflickering (BVD) task, in which methods refer to that (1) is agnostic to flickering patterns or levels (e.g., old movies, high-speed cameras, processing artifacts and color distortions), and (2) operates on a single flicker video without requiring additional guidance (e.g., flicker types, reference videos). In other words, BVD methods are blind to flicker types and guidance, making them widely applicable. Given these, the BVD task is challenging due to the lack of additional prior information guidance. Previous methods have explored traditional filtering [8, 31], forced temporal consistency [20, 22], and atlas adjustment [21], etc. + +Despite significant advancements by deep-learning methods in the BVD task, several critical issues hinder their broader application. First, these approaches demand substantial resources during both the training and inference stages, notably manifesting in slow processing speeds for individual videos. Second, their effectiveness is compromised by two main factors: (i) partial color artifacts and + +distortion are present in the resulting video, causing differences in color fidelity and detail compared to the original video; (ii) Overexposure or underexposure due to lighting leads to significant loss of texture details and introduces new local flicker. These issues are largely attributed to the spatio-temporal complexity within video data, where the sheer volume of pixels presents a formidable obstacle for neural networks, including the advanced large vision models, to learn and maintain global visual consistency. Consequently, given that flicker fundamentally involves local or instantaneous shifts in illumination, effectively addressing the BVD task necessitates a representation more compact and adept at capturing illumination fluctuations than pixel values, whereas this aspect has not yet been explored by previous deep BVD methods. + +In this paper, we draw inspiration from the classic flicker removal method scale-time equalization (STE) to facilitate deep BVD with histogram representation. An image histogram is defined as a distribution of pixel values, and it has been widely applied in image manipulation to adjust image brightness or contrast. Given an arbitrary video, STE temporally smooths histograms with Gaussian filtering and corrects pixel values in each frame with histogram equalization, improving the visual stability of the video. Although STE is only effective for some mild flickers, it validates that 1) histogram, which is much more compact than pixel values, can well delineate the illumination and flicker information, and 2) the video with a smoothed histogram series is visually appealing without obvious flicker, as shown in Fig. 1. Therefore, it is promising to advance the quality and speed of deep BVD with the hints from STE and histogram. + +We introduce Blaze Blind Video Deflickering, dubbed as BlazeBVD, which is a histogram-assisted approach to achieve fast and faithful texture restoration given illumination fluctuation and over-/under-exposure. Compared to previous deep methods, BlazeBVD is the first to leverage histogram meticulously to ease the learning complexity of the BVD task. Specifically, BlazeBVD comprises three stages: At first, we introduce STE to rectify the histogram series of video frames under the illumination space, and extract deflickering priors including singular frames set, filtered illumination maps, and exposure maps. Second, due to the stable temporal performance of filtered illumination maps, they will be used as prompt conditions of a global flicker removal module (GFRM) containing 2D spatial networks to guide the color correction of video frames. On the other hand, a local flicker removal module (LFRM) is based on the optical flow warp to restore local over-/under-exposure regions labeled by exposure maps. Finally, we introduce a lightweight spatio-temporal network to process all frames, where an adaptive mask weighted warping loss is designed to improve video coherence. Our contributions can be summarized as follows: + +(1) We present BlazeBVD, a histogram-assisted blind video deflickering method that simplifies the complexity and resource consumption of learning video data. At its core, BlazeBVD utilizes deflickering priors from STE, including filtered illumination maps for guiding the elimination of global flicker, singular frames set for identify the indexes of flicker frames, and exposure maps to identify regions affected locally by over-/under-exposure. + +(2) Leveraging deflickering priors, BlazeBVD incorporates a Global Flicker Removal Module (GFRM) and a Local Flicker Removal Module (LFRM). These modules work together to efficiently correct global illumination and locally exposed textures in individual adjacent frames, significantly reducing processing time compared to handling the entire video. For enhanced coherence across frames, a lightweight spatio-temporal network is also integrated, boosting performance without significant time consumption. +(3) Through comprehensive experiments on synthetic, real-world and generated videos, we showcase the superior qualitative and quantitative results of BlazeBVD, achieving model inference speeds up to $10 \times$ faster than state-of-the-art. To the best of our knowledge, BlazeBVD also represents the first method to effectively tackle both illumination fluctuations and exposure challenges. + +# 2 Related Work + +Video deflickering addresses the degradation in temporal consistency [26, 27, 29, 38, 48] resulting from abnormal camera, lighting, exposure or image processing algorithms. The relevant methods can be divided into three main categories. The first type of approach focuses on addressing specific flickering types, e.g., high-speed cameras [18], old film [9]. Delon [8] performs an axiomatic analysis of flicker, and proposed a fast method for global removal based on scale space theory to filter in color space directly. The second type of approach, blind temporal consistency [39, 53], aims to generate temporally stable output videos for arbitrary videos regardless of flickering or other artifacts. Bonneel et al. [4] calculate the gradient of the input frame as a guide to reduce the degree of inconsistency between frames. Lai et al. [20] input two consecutive frames as guidance. Lei et al. [22] directly learn the mapping function between input frames and processing frames to reach the purpose of inter-frame information consistency. Despite the generalization performance, such methods tend to be less efficient and hard to handle the complex flicker in practice. The third type of approach is blind video deflickering. Lei et al. [21] propose to solve universal blind video flicker using the neural atlas and video consistency. Meanwhile, some commercial software can remove flicker by integrating various known-flicker methods, such as DEFlicker [34] and Flicker Free [2]. However, most of these methods require users to have flickering type knowledge background or slow repair speed. Our approach does not require prior cognition of the specific information flashing, and achieves fast deflickering speeds so that most users can efficiently process more videos. + +Exposure correction. The purpose of exposure correction is to enhance an improperly exposed image to achieve a satisfactory visual effect. Information is lost due to over-/under-exposure input images. How to solve this problem intelligently and automatically, there have been many techniques in the image domain [12, 13, 15, 24, 41]. ERL [14] connects the under-/over-exposed optimization processes by correlating and constraining the relationships in the mini-batch correction process. DA [44] proposes to decouple contrast enhancement and detail recovery in each convolution process and design detail perception units to + +be inserted into the existing CNN-based exposure correction network. However, there is no general approach regarding local exposure in video tasks. + +Optical flow finds the corresponding relationship between the previous frame and the current frame by using the changes of pixels in the time domain and the correlation between adjacent frames, to calculate the motion information of objects between consecutive frames [10, 11, 16, 37]. In recent years, it has been used in video-related visual tasks [50]. The main target of flow-based methods is to produce accurate estimates of the optical flow through different network architectures, such as utilizing bidirectional optical flow estimation and refining inter-frame coherence. + +# 3 Preliminaries + +In this section, we introduce basic definitions and notations, and review relevant techniques involving image histograms for use in BlazeBVD. + +Image histogram is a graphical representation illustrating the distribution of pixels across different intensity levels within an image [3]. For a single-channel image $I$ with height $H$ and width $W$ , its histogram $\mathcal{H}$ can be calculated as + +$$ +\mathcal {H} (\lambda) = \frac {n _ {\lambda}}{H W}, \quad i f \lambda \in I o t h e r w i s e 0, \tag {1} +$$ + +where $n_{\lambda}$ represents the count of pixels having the intensity $\lambda$ + +Histogram matching or equalization [1, 5, 47, 51], is a technique to adjust the contrast of an image by transforming its histogram towards a predefined distribution. The cumulative histogram of an image histogram hist is defined by $\text{Hist}(\lambda) = \Sigma_{r=0}^{\lambda} \mathcal{H}(r)$ . Give a target histogram $\mathcal{H}'$ and its cumulative histogram $\text{Hist}'(\lambda)$ , the histogram matching process can be described as + +$$ +\mathcal {Q} (\lambda ; \mathcal {H}, \mathcal {H} ^ {\prime}) = H i s t ^ {t - 1} (H i s t (\lambda)), \forall \lambda \in I, \tag {2} +$$ + +This procedure yields an image $\tilde{X}$ , whose histogram closely aligns with $Hist'$ . Scale-Time Equalization (STE) is a filtering approach [8] to adjust the intensity distribution across both spatial and temporal dimensions, mitigating flickers encountered in video sequences. The classic scale-space theory posits that convolving an image with a Gaussian kernel results in smoothing and the loss of image structure. Intuitively, while STE is based on scale-space theory, it applies Gaussian smoothing to sequences of histograms instead of the images. Given a flickering video sequence $\{X_{t}\}_{t=1}^{T}$ , leveraging Eq. (1) can get the histogram sequence $\{\mathcal{H}_t\}_{t=1}^T$ . For arbitrary time index $t$ and intensity value $\lambda$ , STE can be formulated as the convolution: + +$$ +S T E (\lambda , t) = G _ {s} (\tau) * \mathcal {Q} (\lambda ; \mathcal {H} _ {t}, \mathcal {H} _ {\tau}) \approx \sum_ {\tau = t - l} ^ {t + l} G _ {s} (\tau - t) \mathcal {Q} (\lambda ; \mathcal {H} _ {t}, \mathcal {H} _ {\tau}), \tag {3} +$$ + +where $l$ denotes the window radius, $G_{s}(t) = \frac{1}{\sqrt{4\pi s}} e^{-\frac{t^{2}}{4s}}$ . After STE, we will get a sequence of histograms $\{\tilde{\mathcal{H}}_t\}_{t=1}^T$ and images $\{\tilde{X}_t\}_{t=1}^T$ . Since histogram transformations do not create or cancel any image content, STE can achieve deflickering + +![](images/18c392babcb1e41a7661c6024ccdc5259d9009aabd33e6af8576ab947bec4219.jpg) +Fig. 2: The framework of our approach BlazeBVD. We first extract flicker prior information about the input video and correct the brightness representation with STE-assisted histogram filtering in illuminance space (Stage1). Then, the prior is leveraged to remove temporal flicker and over-/under-exposure flicker from the global and local perspectives (GFRM and LFRM in Stage2). Finally, the temporal consistency of the processed video is improved (TCM in Stage3). + +results without compromising the integrity of the original video. However, this also indicates that STE cannot compensate for the lost texture in flickering regions and produces vivid effects, which impedes its practical application. + +# 4 Methodology + +Let $\{X_{t}\}_{t = 1}^{T}$ represent a video sequence affected by flickering defects and $\{G_t\}_{t = 1}^T$ represent the groundtruth clean video. In this section, we propose BlazeBVD as a universal solution to eliminate these flicker and produce a temporally consistent video $\{O_t\}_{t = 1}^T$ . It is important to note that the types of flicker addressed in our work can vary widely, involving long-/short-term flicker in color and brightness. + +Distinguished from existing prior BVD approaches, BlazeBVD employs a histogram-assisted approach, leveraging deflickering priors from Scale-Time Equalization (STE) [8] to streamline the complexity and resource demands of BVD task. Beyond ensuring stable temporal performance, BlazeBVD stands out as the first method capable of faithfully restoring the color and texture affected by illumination fluctuations and exposure challenges, including both over-/under-exposure. + +# 4.1 Overview of BlazeBVD + +As illustrated in Fig. 2, the BlazeBVD pipeline is structured into three main stages to address video deflickering challenges effectively: (1) Deflickering priors + +![](images/12b2971deb2050878524729be7c5a753dd390ee4dc7bf9ed489d3ef2c3e41f7c.jpg) +Fig. 3: Pipeline of Stage1. The illumination maps are corrected in the temporal dimension by STE, which alleviates the difficulty of subsequent temporal data learning using networks. The deflickering priors can be extracted: $\tilde{V}_t$ , $M_t$ and $S_{flicker}$ . + +preparation. This initial stage leverages STE to correct the histogram within an illumination space representation. Key outputs from this stage include the filtered illumination map $\{\tilde{V}_t\}_{t=1}^T$ , identifying the index set of singular frames $S_{flicker}$ , and generating exposure maps $\{M_t\}_{t=1}^T$ . These elements serve as crucial prior information, guiding the deflickering process in subsequent stages. (2) Global and local flicker removal. The second stage involves two flicker removal strategies. A global flicker removal module utilizes the filtered illumination map $\{\tilde{V}_t\}_{t=1}^T$ to address temporal flicker issues across the video. A local over-/under-exposed region flicker removal module instructed by the singular frames set $S_{flicker}$ and exposure maps $\{M_t\}_{t=1}^T$ , restores the texture in the over-/under-exposure region to compensate for the lost high-frequency details. (3) Adaptive temporal consistency. The final stage introduces a spatio-temporal network under the guidance of an adaptive warping loss to improve video coherence and temporal consistency. + +These stages delineated in the following Sec. 4.2, Sec. 4.3, and Sec. 4.4, together provide a comprehensive solution to enhance the visual quality and temporal consistency of video content. + +# 4.2 Stage1: Deflickering priors preparation + +Spatio-temporal challenge. An effective BVD model should strategically avoid the intricacies of spatio-temporal complexity across pixels, nevertheless, current deep learning approaches to BVD have not adequately addressed this challenge. For instance, the state-of-the-art BVD method, Deflicker [21], employs a neural atlas [19] as a unified representation of the entire video. While this can offer consistent guidance for the deflickering process, the effectiveness of neural atlases significantly depends on the video duration and the network capacity. Processing an 8-second video to create an atlas can take around 10 minutes, and longer videos exacerbate the demands on time, memory, and computational resources. Consequently, there is a further need for research into more compact representations that can facilitate deflickering without these limitations. + +Illumination histogram. We introduce histogram representation [3], which utilizes statistical distribution to illustrate the global properties of image pixel values. As illustrated in Fig. 3, we first extract the illumination maps $\{V_t\}_{t=1}^T$ according to the conversion between HSV and RGB color space: + +$$ +V _ {t} = \max _ {R, G, B \in X _ {t}} \{R, G, B \}, t \in \{1, 2, \dots , T \}. \tag {4} +$$ + +Next, by applying Eq. (1) to each illumination map, we utilize illumination histograms $\{H_t\}_{t=1}^T$ to represent the illumination characteristics of video frames. These illumination histograms reflect important flickering information: As shown in Fig. 3, frames without distinct illumination changes have similar histograms, while the histogram of a flicker frame has a distribution distinguished from others. Furthermore, spikes in the head or tail area of the distribution indicate potential over-/under-exposure problems in the image. To this end, reasonable utilization of histogram information is significant for flicker elimination and alleviation of spatio-temporal complexity. + +Deflickering priors. We introduce deflickering priors extracted from illumination histograms to facilitate the deflickering process. Initially, considering that the illumination histograms of natural videos exhibit continuous changes in the temporal dimension, we utilize Scale-Time Equalization (STE) as defined in Eq. (3) to smooth the illumination histograms, as illustrated in Fig. 3. Through histogram matching, we further obtain filtered illumination maps $\{\tilde{V}\}_{t=1}^{T}$ that exhibit no obvious flicker compared to the original ones, thereby enabling the subsequent independent processing of each frame by the neural network. Secondly, we identify the singular frames set $S_{flicker}$ containing the indexes of flicker frames, based on the divergence of histograms before and after STE: + +$$ +S _ {f l i c k e r} = \left\{t \in [ 1, T ] \mid \mathrm {K L} \left(\tilde {H} _ {t} \parallel H _ {t}\right) > \bar {\epsilon} _ {t} \right\}, \quad \bar {\epsilon} _ {t} = \frac {1}{2 n + 1} \sum_ {\tau = t - n} ^ {t + n} \mathrm {K L} \left(\tilde {H} _ {\tau} \parallel H _ {\tau}\right), \tag {5} +$$ + +where $n$ is the moving average radius. Finally, as shown in Fig. 2, we extract the exposure maps $\{M_t \in \mathbb{R}^{H \times W}\}_{t=1}^T$ of local over-/under-exposure according to the filtered illumination map: + +$$ +M _ {t} (i, j) = \left\{ \begin{array}{l l} 1 & \tilde {V} (i, j) < \epsilon_ {1} \text {o r} \tilde {V} (i, j) > \epsilon_ {2}, \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {6} +$$ + +Here, $\epsilon_{1}$ and $\epsilon_{2}$ are darkness and exposure thresholds, respectively, and we consider illumination values belonging to $[\epsilon_1,\epsilon_2]$ as not posing an exposure risk. + +In summary, we leverage illumination histograms to prepare deflickering priors including filtered illumination maps $\{\tilde{V}\}_{t=1}^{T}$ , exposure maps $\{M_t\}_{t=1}^T$ , and the singular frames set $S_{flicker}$ , for alleviating the spatio-temporal complexity and enhancing deflickering in the subsequent learning stages. + +# 4.3 Stage2: Global and local flicker removal + +In this stage, simple 2D spatial neural networks are integrated with deflickering priors to eliminate both global and local flickering flaws. Here, "global" refers + +to obvious temporal inconsistency due to illumination fluctuation, while "local" refers to lost texture details due to exposure and darkness. This approach allows for the direct generation of high-quality videos from flicker frames, regardless of the spatio-temporal complexity. + +Global deflickering. BlazeBVD is equipped with a global flicker removal module (GFRM), which incorporates a 2D-Unet to correct each frame $X_{t}$ with the guidance of the filtered illumination map $\tilde{V}_{t}$ . In other words, the corrected image $\hat{X}_{t} = U\mathrm{net}_{\theta}(X_{t},\tilde{V}_{t})$ , where $\theta$ denotes the parameters of the 2D-Unet. The network is trained on the BVD dataset $\mathcal{D}$ by minimizing the L2 loss function between the predicted images and the groundtruth images: + +$$ +\mathcal {L} _ {t} = \mathbb {E} _ {\left(X _ {t}, G _ {t}\right) \in \mathcal {D}} \| U n e t _ {\theta} \left(X _ {t}, \tilde {V} _ {t}\right) - G _ {t} \|. \tag {7} +$$ + +Local deflickering. BlazeBVD is also equipped with a local flicker removal module (LFRM) to enhance local texture details. As depicted in Fig. 5, overexposure can transform intricate texture in local regions into uniform pure, while local under-exposure can result in black areas. This loss of detailed texture cannot be compensated for by previous BVD methods, such as Deflicker [21]. We leverage LFRM to transfer detailed information from neighboring frames to these ill regions. Given the corrected singular frame $\tilde{X}_t$ in $S_{\text{flicker}}$ with local exposure, along with the previous frame $\tilde{X}_{t-1}$ and the next frame $\tilde{X}_{t+1}$ , we first compute the optical flow to extract part of the detailed texture: + +$$ +o _ {\tau \rightarrow t} = \operatorname {F l o w} \left(X _ {\tau}, X _ {t}\right), \tau \in \{t + 1, t - 1 \} \text {a n d} t \in S _ {\text {f l i c k e r}}. \tag {8} +$$ + +Here Flow is the optical flow estimation function, in our work we set it as the popular RAFT model [11, 37]. We migrate the possible detail textures of the preceding and following frames into the current frame: + +$$ +\hat {X} _ {t} = \mathcal {F} \left(M _ {t} \odot \mathcal {W} \left(\tilde {X} _ {t - 1}, o _ {t - 1 \rightarrow t}\right) + (1 - M _ {t}) \odot \tilde {X} _ {t}, \tilde {X} _ {t}, \right. \tag {9} +$$ + +$$ +M _ {t} \odot \mathcal {W} \left(\tilde {X} _ {t - 1}, o _ {t + 1 \rightarrow t}\right) + \left(1 - M _ {t}\right) \odot \tilde {X} _ {t}), t \in S _ {\text {f l i c k e r}}, +$$ + +where $\mathcal{W}$ is the warping operator and $\mathcal{F}$ is the fusion network from [41]. + +# 4.4 Stage3: Adaptive temporal consistency + +The former stages have well addressed the main illumination fluctuation and exposure problems in BVD task. Nevertheless, a video may also have content inconsistencies to largely degrade the visual effect, and the former stage may also cause this. Therefore, We introduce the Temporal Consistency Model (TCM) to refine video $\{\hat{X}_t\}_{t=1}^T$ as the final outputs $\{O_t\}_{t=1}^T$ . TCM is developed based on the architecture by RTN [40], but with a temporal consistency loss of adaptive masks to refine local artifacts: + +$$ +E _ {\text {p a i r}} ^ {\prime} \left(O _ {t}, O _ {s}\right) = \left\| W _ {t} \odot \left(M _ {t} + 1\right) \odot M _ {t, s} \odot \left(O _ {t} - \mathcal {W} \left(O _ {s}\right)\right) \right\| _ {1}, +$$ + +$$ +\mathcal {L} _ {w a r p} = \frac {1}{T - 1} \sum_ {t = 2} ^ {T} \left\{E _ {\text {p a i r}} ^ {\prime} \left(O _ {t}, O _ {1}\right) + E _ {\text {p a i r}} ^ {\prime} \left(O _ {t}, O _ {t - 1}\right) \right\}, \tag {10} +$$ + +where $W_{t}$ is the loss weight for the over-/under-exposure area and $M_{t}$ is the mask obtained from the former stage. $\mathcal{W}$ is the warping operator and the frame after warping is computed by optical flow estimation [11,37]. $M_{t,s}$ represents the mask after optical flow estimation. This loss avoids the video blurring problem introduced in Deflicker [21]. The training loss of TCM consists of reconstruction loss $\mathcal{L}_{rec}$ , perceptual loss $\mathcal{L}_{per}$ [17], spatio-temporal adversarial loss $\mathcal{L}_{adv}$ [6] and designed adaptive weighted warping loss $\mathcal{L}_{warp}$ . Combining all the above losses, the overall objective to optimize TCM is: + +$$ +\mathcal {L} _ {T C M} = \alpha_ {1} * \mathcal {L} _ {r e c} + \alpha_ {2} * \mathcal {L} _ {p e r} + \alpha_ {3} * \mathcal {L} _ {a d v} + \alpha_ {4} * \mathcal {L} _ {w a r p}. \tag {11} +$$ + +In our work, we set $\alpha_{1} = 1.0$ , $\alpha_{2} = 1.0$ , $\alpha_{3} = 0.01$ and $\alpha_{4} = 0.1$ . + +# 5 Experiments + +# 5.1 Datasets + +We inherit the training dataset DAVIS-2017-Train, MS-COCO and the testing dataset Blind Deflickering Dataset from Deflicker [21], including Synthetic videos from DAVIS-2017-Val, Real-world videos and Generation videos. + +Training dataset. MS-COCO [23] contains 118287 images in the train set and is used for the pre-training of Unet in GFRM. DAVIS-2017-Train [20, 28, 32] consists of 60 sequences with a total of 4219 labeled frames, with a video frame rate of 24fps and a resolution of $480\mathrm{p}$ . We create the synthetic process that provides the underlying facts for quantitative analysis. The training dataset is degraded by performing flicker on DAVIS-2017-Train, as shown in Eq. (12): + +$$ +X _ {t} = G _ {t} + F _ {t}, t \in \{1, 2, \dots , T \}, \tag {12} +$$ + +where $\{G_t\}_{t=1}^T$ are clean frames and $\{F_t\}_{t=1}^T$ are random synthesized flickering artifacts. The window size $W$ in $F_t$ has been set to denote the number of frames that share the same flickering artifacts, random sampling between [2, 12]. The value of $W$ controls the degree of flicker in the short- and long-term of the video. + +Testing dataset. We conduct Synthetic videos, Real-world videos and Generation videos respectively, from the Blind Deflickering Dataset [21] and DAVIS2017-Test processed by Eq. (12). Among them, Synthetic videos and DAVIS2017-Test have ground truth and the others do not. We can only use no-referenced metric and qualitative observation. Generation videos refer to the processed videos generated by generative algorithms, which may lead to flicker artifacts due to the imperfections of the algorithms. Real-world videos including old movies, old anime, slow-motion, etc. are directly captured, unprocessed videos in the real world. + +Comparison methods. We compare our BlazeBVD with other 4 methods, including STE [8], ConvLSTM [20], DVP [22] and Deflicker [21]. We set Deflicker as the baseline, since it is the first publicly presented method for BVD task and is currently the most advanced approach for general videos. + +![](images/075647226b6f736da970113ef77697ab5554a0abab0a1d77e24f5002c1902aef.jpg) +Fig. 4: Qualitative comparisons between previous methods and our BlazeBVD. Our method removes flicker and restores details in over-exposed regions (red boxes in row 6 and row 8) while avoiding color artifacts (row 2 and row 4). Besides, BlazeBVD also ensures the fidelity of the video content and avoids color distortion (cyan boxes in row 5 and row 7). Zoom in for the best view and we recommend watching videos in the supplementary materials. + +# 5.2 Comparisons with State-of-the-art Methods + +Metrics. We quantitatively compare the differences between our method and others with reference and no-reference metrics, including PSNR, SSIM [45], and $E_{\text{warp}}$ [22]. Among them, $E_{\text{warp}}$ is a no-reference metric used in videos. For each frame $O_t$ , we calculate the warping error with frame $O_{t-1}$ and the first frame $O_1$ for considering both short-term and long-term consistency. + +Quantitative results. We compare our BlazeBVD with the state-of-the-art method Deflicker [21], and other consistency methods. Table 1 and Table 2 provide the comparative experimental results of various methods on DAVIS-2017-Test and Synthetic videos, where PSNR and SSIM of the resulting video are higher than those of the other methods. This shows that BlazeBVD contributes to the fidelity maintenance of video content, as well as the correctness of the missing details complement. The warping error of our method is a little higher + +Table 1: Quantitative comparisons of synthetic videos in DAVIS-2017-Test. Our method performs best in terms of PSNR and SSIM and $E_{\text{warp}}$ of all the methods. + +
MethodRawVideoConvLSTMDVPSTEDeflickerOursGT
PSNR↑-24.11024.13626.13823.93228.609-
SSIM↑-0.92560.92630.93210.92430.9638-
E_warp ↓0.14290.13690.14150.11170.08400.08250.1097
+ +Table 2: Quantitative comparisons in Synthetic videos from Blind Deflickering Dataset. $W$ is the window size and represents the number of frames that share the same flicker artifact. $L$ denotes the local seed window size. + +
VideoTypeRawE_warp↓PSNR↑
DeflickerOursGTDeflickerOursDeflickerOurs
Synthetic
-W=10.19330.08480.08950.105625.931229.67380.94030.9539
-W=30.15220.08440.08900.105625.499628.82720.93800.9555
-W=100.13070.08550.09040.105624.384226.99660.93440.9543
-L=30.11130.08750.09000.105625.387427.08810.92710.9517
Average0.14690.08560.08970.105625.300628.14640.93490.9538
+ +than Deflicker, but it is closer to the ground truth. Through analysis, we find that warping error has inaccuracy in the estimation of optical flow, and this error is masked by image blurring in Deflicker. Therefore, we believe that $E_{\text{warp}}$ , which has the same magnitude as baseline and GT, is the valid experimental comparison result. In Tab. 3, we compare BlazeBVD and Deflicker on Real-world videos and Generation videos. Since the videos in these two datasets with no reference are inherently discontinuous, where $E_{\text{warp}}$ cannot directly represent video continuity, we present the comparative results of user studies. To evaluate the perceived preference between the baseline Deflicker and the proposed method, each user needs to select a video with better perceptual quality between these two choices. In total, there are 50 users and 50 pairs in 7 datasets of comparisons. + +Qualitative results. Both ConvLSTM and DVP are temporal consistency models that require a reference, importing the input video as the reference, and the processed video cannot remove significant flicker, as shown in Fig. 4. Additionally, the flicker caused by over-/under-exposure is prone to lose the local texture (column 1 in Fig. 4). The atlas-based representation in Deflicker cannot accurately locate these points with different local features, resulting in the loss of all details on the processed video. Instead, our method designs local positioning and adjacent frame details transfer in LFRM to preserve part of the texture on the road, as shown in the red box of Fig. 4. We also reveal the color artifact in column 4 of Fig. 4 that appears at baseline, due to the excessive weight of warping error in the training loss when refining temporal consistency. Therefore, our designed adaptive mask weighted training loss in TCM improves the temporal + +Table 3: User studies of Real-world videos and Generation videos. Since the videos in these two datasets are inherently discontinuous, $E_{\text{warp}}$ cannot properly reflect temporal consistency. Here we only present the comparative results of the user study. + +
TypeExpertOldAnimeOldMovieSlowMotionTimeLapseVideoDMVideoLDM
Raw(E_warp)0.07540.10450.10990.07240.07320.12480.1601
Deflicker38.67%19.25%22.83%37.67%31.43%30.50%26.34%
Ours61.33%80.75%77.17%62.33%68.57%69.50%73.66%
+ +consistency while avoiding image blurring. More visualization results are in the supplementary material. + +# 5.3 Ablation Studies + +STE filtering. We contrast the difference between introducing STE filtering in illumination space and directly in color space. As shown in the red box of Fig. 5, directly performing histogram correction in color channels separately introduces color artifacts and color distortion. Because the filtering of non-gray images in the three channels is relatively independent, the recombination after nonlinear transformation would produce such defects. + +GFRM and LFRM. We compare the important roles of two modules, GFRM and LFRM. As shown in Tab. 5, removing GFRM and LFRM reduces PSNR and SSIM and increases $E_{\text{warp}}$ of inference videos. It can be inferred that GFRM needs to initially remove global flicker between frames, otherwise subsequent modules cannot maintain the temporal consistency, while LFRM maintains the texture fidelity of the video content, which is reflected in the cyan box of Fig. 5. Designed weighted loss. As shown in Tab. 5, the lightweight spatio-temporal network TCM reduces $E_{\text{warp}}$ , which refines the temporal consistency of processed videos. We also compare the effect of using unweighted training loss and the adaptive mask weighted loss on this temporal network. Figure 5 provides the resulting frame with unweighted training, which is generally blurred and suffers from color distortion, while the weighted training keeps the frame clear and faithful in most areas. Adaptive mask weighting improves the temporal consistency of the inference video, improving the inter-frame continuity more pertinently. + +# 5.4 Discussion + +Advantages and limitations. BlazeBVD is over $10 \times$ faster than the inference speed of baseline using the atlas representation, which is specifically recorded in Tab. 4. Our method also takes advantage of explicit prior extraction in the BVD task, which provides directions for improvement in the subsequent work. And the cause of local flicker has been noted and effectively dealt with from both global and local perspectives. However, our method still has drawbacks in LFRM. Due to the inaccurate optical flow motion estimation, the fusion network + +Table 4: Inference time and model parameters comparison. Our approach is faster than the baseline in inference time with less memory, and comparable in MACs and Params. + +
MethodDeflickerOurs
Time/80p614.19s58.37s
Memory5204M1389M
MACs260.7G251.4G
Params12.48M17.77M
+ +Table 5: Ablation studies in DAVIS-2017- Test. w/o STE replaces STE with mean filtering in Stage1 and w/o weighted denotes the warping loss without mask weighting in TCM. + +
MethodPSNR↑SSIM↑E_warp ↓
w/o STE26.44120.95170.1011
w/o GFRM23.89390.92440.1225
w/o LFRM27.13660.94620.1122
w/o TCM29.91300.95830.1093
w/o weighted28.10490.95230.0849
BlazeBVD28.60920.96380.0825
+ +![](images/5740c8ba744f729da186f344f8354be7a3ea584b7ad58c1f6c4b14a18326bfae.jpg) +Input + +![](images/e24c203e70391a228155b2de7d6901dbcb8cb64e660a926e5b61c32771434e2b.jpg) +w/o GFRM + +![](images/3d256e30443406c2ddd2af9e7de0467971f09eb7078a83aab616c3c952eb2a68.jpg) +w/o LFRM + +![](images/bb73a782de945ac40d6a1122a84cc82c9c520ba5445cc47725f55d91e9c4746a.jpg) +w/o STE + +![](images/1741435d2c81bde25c30d85a62d1fde0e763015d13ddc8405ac50eb91c5db22f.jpg) +w/o TCM + +![](images/c50d7a915ab75c7a4f8e9c2c68607b78d3b4d3b3715afe44f5f44a2402fed645.jpg) +w/o weighted mask + +![](images/efbd9145fe52c70a20930cc62857d55982fdb628d4cd3e937e29c9ae4bf55259.jpg) +Ours +Fig. 5: Qualitative ablation studies of key designs in BlazeBVD. Specifically, local detail loss is represented in cyan box, color artifact is represented in the red box, temporal inconsistency is represented in the black box, and color distortion is represented in the blue box. Zoom in for the best view. + +![](images/e169edf276ea632d6be578380b60d3ae9cc3e618564091f7c4a91567fb09c41b.jpg) +GT + +cannot convey the local texture of adjacent frames, and there are still slight edge artifacts and color distortion that cannot be eliminated. We hope to continue exploring the transfer and fusion of local region information between frames [40]. + +# 6 Conclusion + +In this paper, we propose BlazeBVD, a universal approach for the blind video deflickering task. The core of our method is to prepare flicker priors within the STE filter in illumination space first, and then utilize these priors to correct the global flicker and local exposed texture. GFRM and LFRM leverage the priors to remove global and local flicker in several adjacent frames with less computation. Finally, TCM is used to improve video coherence and inter-frame consistency. Extensive experiments validate the effectiveness of the designed modules. + +# Acknowledgements + +This paper is supported by the National Natural Science Foundation of China (Nos. U23B2012, 11991022), and the Fundamental Research Funds for the Central Universities (E3E41904X2). + +# References + +1. Affifi, M., Brubaker, M.A., Brown, M.S.: Histogram: Controlling colors of gangenerated and real images via color histograms. In: CVPR. pp. 7941-7950 (2021) +2. Anarchy, D.: Flicker free, https://digitalanarchy.com/Flicker/main.html +3. Bassiou, N., Kotropoulos, C.: Color image histogram equalization by absolute discounting back-off. Computer Vision and Image Understanding 107, 108-122 (Jul 2007) +4. Bonneel, N., Tompkin, J., Sunkavalli, K., Sun, D., Paris, S., Pfister, H.: Blind video temporal consistency. ACM Transactions on Graphics 34, 1-9 (Nov 2015) +5. Bottenus, N., Byram, B.C., Hyun, D.: Histogram matching for visual ultrasound image comparison. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control 68, 1487-1495 (May 2021) +6. Chang, Y.L., Liu, Z.Y., Lee, K.Y., Hsu, W.: Free-form video inpainting with 3d gated convolution and temporal patchgan. In: ICCV. pp. 9066-9075 (2019) +7. Chu, M., Xie, Y., Mayer, J., Leal-Taixe, L., Thuerey, N.: Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics 39, 75-1 (2020) +8. Delon, J.: Movie and video scale-time equalization application to flicker reduction. IEEE Transactions on Image Processing 15, 241-248 (Jan 2006) +9. Delon, J., Desolneux, A.: Stabilization of flicker-like effects in image sequences through local contrast correction. SIAM Journal on Imaging Sciences 3, 703-734 (Jan 2010) +0. Dosovitskiy, A., Fischer, P., Ilg, E., Hausser, P., Hazirbas, C., Golkov, V., Van Der Smagt, P., Cremers, D., Brox, T.: Flownet: Learning optical flow with convolutional networks. In: ICCV. pp. 2758-2766 (2015) +1. Eslami, N., Arefi, F., Mansourian, A.M., Kasaei, S.: Rethinking raft for efficient optical flow. arXiv preprint arXiv:2401.00833 (2024) +2. Huang, J., Liu, Y., Fu, X., Zhou, M., Wang, Y., Zhao, F., Xiong, Z.: Exposure normalization and compensation for multiple-exposure correction. In: CVPR. pp. 6043-6052 (2022) +3. Huang, J., Liu, Y., Zhao, F., Yan, K., Zhang, J., Huang, Y., Zhou, M., Xiong, Z.: Deep fourier-based exposure correction network with spatial-frequency interaction. In: ECCV. pp. 163-180 (2022) +4. Huang, J., Zhao, F., Zhou, M., Xiao, J., Zheng, N., Zheng, K., Xiong, Z.: Learning sample relationship for exposure correction. In: CVPR. pp. 9904-9913 (2023) +5. Huang, J., Zhou, M., Liu, Y., Yao, M., Zhao, F., Xiong, Z.: Exposure-consistency representation learning for exposure correction. In: ACM MM. pp. 6309-6317 (2022) +6. Ilg, E., Mayer, N., Saikia, T., Keuper, M., Dosovitskiy, A., Brox, T.: Flownet 2.0: Evolution of optical flow estimation with deep networks. In: CVPR. pp. 2462-2470 (2017) + +17. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV. pp. 694-711 (2016) +18. Kanj, A., Talbot, H., Luparello, R.R.: Flicker removal and superpixel-based motion tracking for high speed videos. In: ICIP. pp. 245-249 (2017) +19. Kasten, Y., Ofri, D., Wang, O., Dekel, T.: Layered neural atlases for consistent video editing. ACM Transactions on Graphics 40, 1-12 (2021) +20. Lai, W.S., Huang, J.B., Wang, O., Shechtman, E., Yumer, E., Yang, M.H.: Learning blind video temporal consistency. In: ECCV. pp. 170-185 (2018) +21. Lei, C., Ren, X., Zhang, Z., Chen, Q.: Blind video deflickering by neural filtering with a flawed atlas. In: CVPR. pp. 10439-10448 (2023) +22. Lei, C., Xing, Y., Chen, Q.: Blind video temporal consistency via deep video prior. In: NeurIPS. pp. 1083-1093 (2020) +23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014) +24. Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: CVPR. pp. 5637-5646 (2022) +25. Mei, K., Patel, V.: Vidm: Video implicit diffusion models. In: AAAI. pp. 9117-9125 (2023) +26. Moniz, J.R.A., Kang, E., Póczos, B.: Luciddream: Controlled temporally-consistent deepdream on videos. arXiv preprint arXiv:1911.11960 (2019) +27. Park, K., Woo, S., Kim, D., Cho, D., Kweon, I.S.: Preserving semantic and temporal consistency for unpaired video-to-video translation. In: ACM MM. pp. 1248-1257 (2019) +28. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: CVPR (Jun 2016) +29. Pérez-Pellitero, E., Sajjadi, M.S., Hirsch, M., Scholkopf, B.: Perceptual video super resolution with enhanced temporal consistency. arXiv preprint arXiv:1807.07930 (2018) +30. Pfeuffer, A., Dietmayer, K.: Separable convolutional lstms for faster video segmentation. In: ITSC. pp. 1072-1078 (2019) +31. Pitie, F., Kokaram, A., Dahyot, R.: Removing flicker from old movies. Master's thesis, University of Nice-Éophia Antipolis, France, Œe September (2002) +32. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv:1704.00675 (2017) +33. Qiu, X., Han, C., Zhang, Z., Li, B., Guo, T., Nie, X.: Diffbfr: Bootstrapping diffusion model for blind face restoration. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7785-7795 (2023) +34. RE:VISION: De:flicker, https://revisionfx.com/products/deflicker/ +35. Saito, M., Saito, S., Koyama, M., Kobayashi, S.: Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. International Journal of Computer Vision p. 2586-2606 (Nov 2020) +36. Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In: CVPR. pp. 3626-3636 (2022) +37. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402-419 (2020) +38. Thasarathan, H., Nazeri, K., Ebrahimi, M.: Automatic temporally coherent video colorization. In: CRV. pp. 189-194 (2019) + +39. Thimonier, H., Despois, J., Kips, R., Perrot, M.: Learning long-term style preserving blind video temporal consistency. In: ICME. pp. 1-6 (2021) +40. Wan, Z., Zhang, B., Chen, D., Liao, J.: Bringing old films back to life. In: CVPR. pp. 17694-17703 (2022) +41. Wang, H., Xu, K., Lau, R.W.: Local color distributions prior for image enhancement. In: ECCV. pp. 343-359 (2022) +42. Wang, W., Liu, J., Lin, Z., Yan, J., Chen, S., Low, C., Hoang, T., Wu, J., Liew, J.H., Yan, H., et al.: Magicvideo-v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468 (2024) +43. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9168-9178 (2021) +44. Wang, Y., Peng, L., Li, L., Cao, Y., Zha, Z.J.: Decoupling-and-aggregating for image exposure correction. In: CVPR. pp. 18115-18124 (2023) +45. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13, 600-612 (2004) +46. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV. vol. 39, pp. 7623-7633 (2023) +47. Xu, R., Tu, Z., Du, Y., Dong, X., Li, J., Meng, Z., Ma, J., Bovik, A., Yu, H.: Pik-fix: Restoring and colorizing old photos. In: WACV. pp. 1724-1734 (2023) +48. Xu, X., Li, M., Sun, W., Yang, M.H.: Learning spatial and spatio-temporal pixel aggregations for image and video denoising. IEEE Transactions on Image Processing 29, 7153-7165 (2020) +49. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 672-681 (2021) +50. Zhan, X., Pan, X., Liu, Z., Lin, D., Loy, C.C.: Self-supervised learning via conditional motion propagation. In: CVPR. pp. 1881-1889 (2019) +51. Zhang, F., Shao, Y., Sun, Y., Zhu, K., Gao, C., Sang, N.: Unsupervised low-light image enhancement via histogram equalization prior. arXiv preprint arXiv:2112.01766 (2021) +52. Zhang, Z., Li, B., Nie, X., Han, C., Guo, T., Liu, L.: Towards consistent video editing with text-to-image diffusion models. Advances in Neural Information Processing Systems 36 (2024) +53. Zhou, Y., Xu, X., Shen, F., Gao, L., Lu, H., Shen, H.T.: Temporal denoising mask synthesis network for learning blind video temporal consistency. In: ACM MM. pp. 475-483 (2020) \ No newline at end of file diff --git a/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/images.zip b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bf723f50e55ceb37e4e1df46196be10e5fecdc02 --- /dev/null +++ b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8485ef70e64930e652f4d7ca27215319711bc46f201f2dd2c675d6629d6e9ca8 +size 712341 diff --git a/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/layout.json b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ef475fde561d0346ab47bb26b9805a6da94913c7 --- /dev/null +++ b/2024/BlazeBVD_ Make Scale-Time Equalization Great Again for Blind Video Deflickering/layout.json @@ -0,0 +1,9702 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "text", + "content": "BlazeBVD: Make Scale-Time Equalization Great Again for Blind Video Deflickering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "spans": [ + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "text", + "content": "Xinmin Qiu" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "text", + "content": ", Congying Han" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "text", + "content": ", Zicheng Zhang" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "text", + "content": ", Bonan Li" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "text", + "content": ", Tiande Guo" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "text", + "content": ", Pingyu Wang" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "text", + "content": ", and Xuecheng Nie" + }, + { + "bbox": [ + 140, + 173, + 473, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 182, + 207, + 430, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 207, + 430, + 219 + ], + "spans": [ + { + "bbox": [ + 182, + 207, + 430, + 219 + ], + "type": "text", + "content": "1 University of Chinese Academy of Sciences, Beijing, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 229, + 220, + 383, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 220, + 383, + 229 + ], + "spans": [ + { + "bbox": [ + 229, + 220, + 383, + 229 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 229, + 220, + 383, + 229 + ], + "type": "text", + "content": " Sichuan University, Sichuan, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 229, + 230, + 384, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 230, + 384, + 241 + ], + "spans": [ + { + "bbox": [ + 229, + 230, + 384, + 241 + ], + "type": "text", + "content": "3 MT Lab, Meitu Inc., Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 199, + 242, + 414, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 242, + 414, + 252 + ], + "spans": [ + { + "bbox": [ + 199, + 242, + 414, + 252 + ], + "type": "text", + "content": "{qiuxinmin21,zhangzicheng19}@mails.ucas.ac.cn," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 230, + 253, + 384, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 253, + 384, + 263 + ], + "spans": [ + { + "bbox": [ + 230, + 253, + 384, + 263 + ], + "type": "text", + "content": "{hancy,libonacci,tdguo}@ucas.ac.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 223, + 265, + 391, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 265, + 391, + 274 + ], + "spans": [ + { + "bbox": [ + 223, + 265, + 391, + 274 + ], + "type": "text", + "content": "wangpingyu@scu.edu.cn, nxc@meitu.com" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 133, + 304, + 219, + 398 + ], + "blocks": [ + { + "bbox": [ + 133, + 304, + 219, + 398 + ], + "lines": [ + { + "bbox": [ + 133, + 304, + 219, + 398 + ], + "spans": [ + { + "bbox": [ + 133, + 304, + 219, + 398 + ], + "type": "image", + "image_path": "272906ba01293ee2a957fdc94b81aefe058b99b6cb2639e6fc31e94781068bcc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 399, + 181, + 407 + ], + "lines": [ + { + "bbox": [ + 164, + 399, + 181, + 407 + ], + "spans": [ + { + "bbox": [ + 164, + 399, + 181, + 407 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 220, + 304, + 306, + 398 + ], + "blocks": [ + { + "bbox": [ + 220, + 304, + 306, + 398 + ], + "lines": [ + { + "bbox": [ + 220, + 304, + 306, + 398 + ], + "spans": [ + { + "bbox": [ + 220, + 304, + 306, + 398 + ], + "type": "image", + "image_path": "2fdfc44e33dbeee4131ea7b7906a2ded145613c3a808b7f646bbe7a0aae299cc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 399, + 277, + 407 + ], + "lines": [ + { + "bbox": [ + 248, + 399, + 277, + 407 + ], + "spans": [ + { + "bbox": [ + 248, + 399, + 277, + 407 + ], + "type": "text", + "content": "Deflicker" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 304, + 393, + 398 + ], + "blocks": [ + { + "bbox": [ + 307, + 304, + 393, + 398 + ], + "lines": [ + { + "bbox": [ + 307, + 304, + 393, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 393, + 398 + ], + "type": "image", + "image_path": "dc5085a5bafb6366d59691c93e040444202f2bc7010fa436628d37d269904027.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 399, + 359, + 407 + ], + "lines": [ + { + "bbox": [ + 343, + 399, + 359, + 407 + ], + "spans": [ + { + "bbox": [ + 343, + 399, + 359, + 407 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 426, + 482, + 481 + ], + "lines": [ + { + "bbox": [ + 131, + 426, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 482, + 481 + ], + "type": "text", + "content": "Fig. 1: Comparisons of the proposed BlazeBVD. We present flickering input, GT, Deflicker [21] and our BlazeBVD processed video frames, and illumination histograms along with KL divergence about GroundTruth. Our method recovers the illumination histograms well while avoiding the appearance of color artifacts and color distortions (such as the man's arm in the second column). Better see in color with " + }, + { + "bbox": [ + 131, + 426, + 482, + 481 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 131, + 426, + 482, + 481 + ], + "type": "text", + "content": " zoom." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 393, + 304, + 480, + 398 + ], + "blocks": [ + { + "bbox": [ + 393, + 304, + 480, + 398 + ], + "lines": [ + { + "bbox": [ + 393, + 304, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 393, + 304, + 480, + 398 + ], + "type": "image", + "image_path": "070bce0cc1aea6af146eebeb3bc234a883b6a86d1e8c442574bf12724f40bfeb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 399, + 441, + 406 + ], + "lines": [ + { + "bbox": [ + 430, + 399, + 441, + 406 + ], + "spans": [ + { + "bbox": [ + 430, + 399, + 441, + 406 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 159, + 524, + 453, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 524, + 453, + 645 + ], + "spans": [ + { + "bbox": [ + 159, + 524, + 453, + 645 + ], + "type": "text", + "content": "Abstract. Developing blind video deflickering (BVD) algorithms to enhance video temporal consistency, is gaining importance amid the flourish of image processing and video generation. However, the intricate nature of video data complicates the training of deep learning methods, leading to high resource consumption and instability, notably under severe lighting flicker. This underscores the critical need for a compact representation beyond pixel values to advance BVD research and applications. Inspired by the classic scale-time equalization (STE), our work introduces the histogram-assisted solution, called BlazeBVD, for high-fidelity and rapid BVD. Compared with STE, which directly corrects pixel values by temporally smoothing color histograms, BlazeBVD" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 652, + 243, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 652, + 243, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 652, + 243, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{\\bar{\\Omega}}" + }, + { + "bbox": [ + 141, + 652, + 243, + 665 + ], + "type": "text", + "content": " Corresponding Author" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 260 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 260 + ], + "type": "text", + "content": "leverages smoothed illumination histograms within STE filtering to ease the challenge of learning temporal data using neural networks. In technique, BlazeBVD begins by condensing pixel values into illumination histograms that precisely capture flickering and local exposure variations. These histograms are then smoothed to produce singular frames set, filtered illumination maps, and exposure maps. Resorting to these deflickering priors, BlazeBVD utilizes a 2D network to restore faithful and consistent texture impacted by lighting changes or localized exposure issues. BlazeBVD also incorporates a lightweight 3D network to amend slight temporal inconsistencies, avoiding the resource consumption issue. Comprehensive experiments on synthetic, real-world and generated videos, showcase the superior qualitative and quantitative results of BlazeBVD, achieving inference speeds up to " + }, + { + "bbox": [ + 159, + 117, + 455, + 260 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 159, + 117, + 455, + 260 + ], + "type": "text", + "content": " faster than state-of-the-arts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 270, + 426, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 270, + 426, + 281 + ], + "spans": [ + { + "bbox": [ + 160, + 270, + 426, + 281 + ], + "type": "text", + "content": "Keywords: Video de-flicker " + }, + { + "bbox": [ + 160, + 270, + 426, + 281 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 270, + 426, + 281 + ], + "type": "text", + "content": " Histogram " + }, + { + "bbox": [ + 160, + 270, + 426, + 281 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 270, + 426, + 281 + ], + "type": "text", + "content": " Temporal consistency" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 304, + 230, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 230, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 230, + 316 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 485 + ], + "type": "text", + "content": "As social applications and the multimedia industry continue to expand, videos have become an essential medium for conveying information in daily life [30]. Achieving a high-quality video demands both image clarity and temporal consistency, nevertheless, video flickering frequently compromises the temporal consistency, often resulting from shooting environment and camera hardware limitations [9, 18]. This issue is further exacerbated when image processing techniques are applied to video frames. Moreover, flickering artifacts and color distortion are also prevalent in recent video generation technologies, including those based on Generative Adversarial Networks (GANs) [7,35,36] and Diffusion Models (DMs) [25,42,46,52]. Since even certain types of flicker can significantly detract from the viewing experience, it is crucial to develop video deflickering techniques to eliminate flicker and preserve the integrity of video content given unknown degradation in various video processing scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "text", + "content": "Recovering visual content from unknown degradation is a key yet challenging topic [33, 43, 49]. Several seminal works [8, 20, 21] have addressed the blind video deflickering (BVD) task, in which methods refer to that (1) is agnostic to flickering patterns or levels (e.g., old movies, high-speed cameras, processing artifacts and color distortions), and (2) operates on a single flicker video without requiring additional guidance (e.g., flicker types, reference videos). In other words, BVD methods are blind to flicker types and guidance, making them widely applicable. Given these, the BVD task is challenging due to the lack of additional prior information guidance. Previous methods have explored traditional filtering [8, 31], forced temporal consistency [20, 22], and atlas adjustment [21], etc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Despite significant advancements by deep-learning methods in the BVD task, several critical issues hinder their broader application. First, these approaches demand substantial resources during both the training and inference stages, notably manifesting in slow processing speeds for individual videos. Second, their effectiveness is compromised by two main factors: (i) partial color artifacts and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "X. Qiu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "distortion are present in the resulting video, causing differences in color fidelity and detail compared to the original video; (ii) Overexposure or underexposure due to lighting leads to significant loss of texture details and introduces new local flicker. These issues are largely attributed to the spatio-temporal complexity within video data, where the sheer volume of pixels presents a formidable obstacle for neural networks, including the advanced large vision models, to learn and maintain global visual consistency. Consequently, given that flicker fundamentally involves local or instantaneous shifts in illumination, effectively addressing the BVD task necessitates a representation more compact and adept at capturing illumination fluctuations than pixel values, whereas this aspect has not yet been explored by previous deep BVD methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 251, + 483, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 483, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 483, + 396 + ], + "type": "text", + "content": "In this paper, we draw inspiration from the classic flicker removal method scale-time equalization (STE) to facilitate deep BVD with histogram representation. An image histogram is defined as a distribution of pixel values, and it has been widely applied in image manipulation to adjust image brightness or contrast. Given an arbitrary video, STE temporally smooths histograms with Gaussian filtering and corrects pixel values in each frame with histogram equalization, improving the visual stability of the video. Although STE is only effective for some mild flickers, it validates that 1) histogram, which is much more compact than pixel values, can well delineate the illumination and flicker information, and 2) the video with a smoothed histogram series is visually appealing without obvious flicker, as shown in Fig. 1. Therefore, it is promising to advance the quality and speed of deep BVD with the hints from STE and histogram." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 399, + 483, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 483, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 483, + 589 + ], + "type": "text", + "content": "We introduce Blaze Blind Video Deflickering, dubbed as BlazeBVD, which is a histogram-assisted approach to achieve fast and faithful texture restoration given illumination fluctuation and over-/under-exposure. Compared to previous deep methods, BlazeBVD is the first to leverage histogram meticulously to ease the learning complexity of the BVD task. Specifically, BlazeBVD comprises three stages: At first, we introduce STE to rectify the histogram series of video frames under the illumination space, and extract deflickering priors including singular frames set, filtered illumination maps, and exposure maps. Second, due to the stable temporal performance of filtered illumination maps, they will be used as prompt conditions of a global flicker removal module (GFRM) containing 2D spatial networks to guide the color correction of video frames. On the other hand, a local flicker removal module (LFRM) is based on the optical flow warp to restore local over-/under-exposure regions labeled by exposure maps. Finally, we introduce a lightweight spatio-temporal network to process all frames, where an adaptive mask weighted warping loss is designed to improve video coherence. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 594, + 488, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 488, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 488, + 666 + ], + "type": "text", + "content": "(1) We present BlazeBVD, a histogram-assisted blind video deflickering method that simplifies the complexity and resource consumption of learning video data. At its core, BlazeBVD utilizes deflickering priors from STE, including filtered illumination maps for guiding the elimination of global flicker, singular frames set for identify the indexes of flicker frames, and exposure maps to identify regions affected locally by over-/under-exposure." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 0 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 260 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "(2) Leveraging deflickering priors, BlazeBVD incorporates a Global Flicker Removal Module (GFRM) and a Local Flicker Removal Module (LFRM). These modules work together to efficiently correct global illumination and locally exposed textures in individual adjacent frames, significantly reducing processing time compared to handling the entire video. For enhanced coherence across frames, a lightweight spatio-temporal network is also integrated, boosting performance without significant time consumption." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "type": "text", + "content": "(3) Through comprehensive experiments on synthetic, real-world and generated videos, we showcase the superior qualitative and quantitative results of BlazeBVD, achieving model inference speeds up to " + }, + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "type": "text", + "content": " faster than state-of-the-art. To the best of our knowledge, BlazeBVD also represents the first method to effectively tackle both illumination fluctuations and exposure challenges." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 279, + 237, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 237, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 237, + 293 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 306, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 569 + ], + "type": "text", + "content": "Video deflickering addresses the degradation in temporal consistency [26, 27, 29, 38, 48] resulting from abnormal camera, lighting, exposure or image processing algorithms. The relevant methods can be divided into three main categories. The first type of approach focuses on addressing specific flickering types, e.g., high-speed cameras [18], old film [9]. Delon [8] performs an axiomatic analysis of flicker, and proposed a fast method for global removal based on scale space theory to filter in color space directly. The second type of approach, blind temporal consistency [39, 53], aims to generate temporally stable output videos for arbitrary videos regardless of flickering or other artifacts. Bonneel et al. [4] calculate the gradient of the input frame as a guide to reduce the degree of inconsistency between frames. Lai et al. [20] input two consecutive frames as guidance. Lei et al. [22] directly learn the mapping function between input frames and processing frames to reach the purpose of inter-frame information consistency. Despite the generalization performance, such methods tend to be less efficient and hard to handle the complex flicker in practice. The third type of approach is blind video deflickering. Lei et al. [21] propose to solve universal blind video flicker using the neural atlas and video consistency. Meanwhile, some commercial software can remove flicker by integrating various known-flicker methods, such as DEFlicker [34] and Flicker Free [2]. However, most of these methods require users to have flickering type knowledge background or slow repair speed. Our approach does not require prior cognition of the specific information flashing, and achieves fast deflickering speeds so that most users can efficiently process more videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Exposure correction. The purpose of exposure correction is to enhance an improperly exposed image to achieve a satisfactory visual effect. Information is lost due to over-/under-exposure input images. How to solve this problem intelligently and automatically, there have been many techniques in the image domain [12, 13, 15, 24, 41]. ERL [14] connects the under-/over-exposed optimization processes by correlating and constraining the relationships in the mini-batch correction process. DA [44] proposes to decouple contrast enhancement and detail recovery in each convolution process and design detail perception units to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "X. Qiu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "be inserted into the existing CNN-based exposure correction network. However, there is no general approach regarding local exposure in video tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "Optical flow finds the corresponding relationship between the previous frame and the current frame by using the changes of pixels in the time domain and the correlation between adjacent frames, to calculate the motion information of objects between consecutive frames [10, 11, 16, 37]. In recent years, it has been used in video-related visual tasks [50]. The main target of flow-based methods is to produce accurate estimates of the optical flow through different network architectures, such as utilizing bidirectional optical flow estimation and refining inter-frame coherence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 251, + 233, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 233, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 233, + 264 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 479, + 297 + ], + "type": "text", + "content": "In this section, we introduce basic definitions and notations, and review relevant techniques involving image histograms for use in BlazeBVD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "content": "Image histogram is a graphical representation illustrating the distribution of pixels across different intensity levels within an image [3]. For a single-channel image " + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "content": " with height " + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "content": ", its histogram " + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 297, + 481, + 334 + ], + "type": "text", + "content": " can be calculated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 224, + 338, + 481, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 338, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 224, + 338, + 481, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {H} (\\lambda) = \\frac {n _ {\\lambda}}{H W}, \\quad i f \\lambda \\in I o t h e r w i s e 0, \\tag {1}", + "image_path": "8d9a01ab2d7b22bb6fbca21ccad844a6bf5743b299c7dbcb09084c1f75a3932a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 362, + 408, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 408, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 408, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 362, + 408, + 374 + ], + "type": "inline_equation", + "content": "n_{\\lambda}" + }, + { + "bbox": [ + 130, + 362, + 408, + 374 + ], + "type": "text", + "content": " represents the count of pixels having the intensity " + }, + { + "bbox": [ + 130, + 362, + 408, + 374 + ], + "type": "inline_equation", + "content": "\\lambda" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "text", + "content": "Histogram matching or equalization [1, 5, 47, 51], is a technique to adjust the contrast of an image by transforming its histogram towards a predefined distribution. The cumulative histogram of an image histogram hist is defined by " + }, + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\text{Hist}(\\lambda) = \\Sigma_{r=0}^{\\lambda} \\mathcal{H}(r)" + }, + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "text", + "content": ". Give a target histogram " + }, + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{H}'" + }, + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "text", + "content": " and its cumulative histogram " + }, + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\text{Hist}'(\\lambda)" + }, + { + "bbox": [ + 130, + 374, + 482, + 434 + ], + "type": "text", + "content": ", the histogram matching process can be described as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 218, + 441, + 481, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 441, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 218, + 441, + 481, + 454 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} (\\lambda ; \\mathcal {H}, \\mathcal {H} ^ {\\prime}) = H i s t ^ {t - 1} (H i s t (\\lambda)), \\forall \\lambda \\in I, \\tag {2}", + "image_path": "1152ed6175e0385eef16b491dceaf0a94d97d6829b1247870c26fff293c115d0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "text", + "content": "This procedure yields an image " + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "inline_equation", + "content": "\\tilde{X}" + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "text", + "content": ", whose histogram closely aligns with " + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "inline_equation", + "content": "Hist'" + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "text", + "content": ". Scale-Time Equalization (STE) is a filtering approach [8] to adjust the intensity distribution across both spatial and temporal dimensions, mitigating flickers encountered in video sequences. The classic scale-space theory posits that convolving an image with a Gaussian kernel results in smoothing and the loss of image structure. Intuitively, while STE is based on scale-space theory, it applies Gaussian smoothing to sequences of histograms instead of the images. Given a flickering video sequence " + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "inline_equation", + "content": "\\{X_{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "text", + "content": ", leveraging Eq. (1) can get the histogram sequence " + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{H}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "text", + "content": ". For arbitrary time index " + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "text", + "content": " and intensity value " + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 460, + 479, + 579 + ], + "type": "text", + "content": ", STE can be formulated as the convolution:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 160, + 584, + 481, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 584, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 160, + 584, + 481, + 616 + ], + "type": "interline_equation", + "content": "S T E (\\lambda , t) = G _ {s} (\\tau) * \\mathcal {Q} (\\lambda ; \\mathcal {H} _ {t}, \\mathcal {H} _ {\\tau}) \\approx \\sum_ {\\tau = t - l} ^ {t + l} G _ {s} (\\tau - t) \\mathcal {Q} (\\lambda ; \\mathcal {H} _ {t}, \\mathcal {H} _ {\\tau}), \\tag {3}", + "image_path": "02078a6e1d3968690420a821a98176c842ded8242725e0f403525f39a4f76a51.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "text", + "content": " denotes the window radius, " + }, + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{s}(t) = \\frac{1}{\\sqrt{4\\pi s}} e^{-\\frac{t^{2}}{4s}}" + }, + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "text", + "content": ". After STE, we will get a sequence of histograms " + }, + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathcal{H}}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "text", + "content": " and images " + }, + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\tilde{X}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 624, + 482, + 666 + ], + "type": "text", + "content": ". Since histogram transformations do not create or cancel any image content, STE can achieve deflickering" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 0 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 478, + 264 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 478, + 264 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 478, + 264 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 478, + 264 + ], + "type": "image", + "image_path": "18c392babcb1e41a7661c6024ccdc5259d9009aabd33e6af8576ab947bec4219.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 285, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 352 + ], + "type": "text", + "content": "Fig. 2: The framework of our approach BlazeBVD. We first extract flicker prior information about the input video and correct the brightness representation with STE-assisted histogram filtering in illuminance space (Stage1). Then, the prior is leveraged to remove temporal flicker and over-/under-exposure flicker from the global and local perspectives (GFRM and LFRM in Stage2). Finally, the temporal consistency of the processed video is improved (TCM in Stage3)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 375, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 413 + ], + "type": "text", + "content": "results without compromising the integrity of the original video. However, this also indicates that STE cannot compensate for the lost texture in flickering regions and produces vivid effects, which impedes its practical application." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 430, + 233, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 233, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 233, + 445 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 456, + 481, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 481, + 517 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 456, + 481, + 517 + ], + "type": "inline_equation", + "content": "\\{X_{t}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 130, + 456, + 481, + 517 + ], + "type": "text", + "content": " represent a video sequence affected by flickering defects and " + }, + { + "bbox": [ + 130, + 456, + 481, + 517 + ], + "type": "inline_equation", + "content": "\\{G_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 130, + 456, + 481, + 517 + ], + "type": "text", + "content": " represent the groundtruth clean video. In this section, we propose BlazeBVD as a universal solution to eliminate these flicker and produce a temporally consistent video " + }, + { + "bbox": [ + 130, + 456, + 481, + 517 + ], + "type": "inline_equation", + "content": "\\{O_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 130, + 456, + 481, + 517 + ], + "type": "text", + "content": ". It is important to note that the types of flicker addressed in our work can vary widely, involving long-/short-term flicker in color and brightness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 487, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 487, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 487, + 601 + ], + "type": "text", + "content": "Distinguished from existing prior BVD approaches, BlazeBVD employs a histogram-assisted approach, leveraging deflickering priors from Scale-Time Equalization (STE) [8] to streamline the complexity and resource demands of BVD task. Beyond ensuring stable temporal performance, BlazeBVD stands out as the first method capable of faithfully restoring the color and texture affected by illumination fluctuations and exposure challenges, including both over-/under-exposure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 620, + 278, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 278, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 278, + 632 + ], + "type": "text", + "content": "4.1 Overview of BlazeBVD" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "As illustrated in Fig. 2, the BlazeBVD pipeline is structured into three main stages to address video deflickering challenges effectively: (1) Deflickering priors" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "X. Qiu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 220 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 220 + ], + "type": "image", + "image_path": "12b2971deb2050878524729be7c5a753dd390ee4dc7bf9ed489d3ef2c3e41f7c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "text", + "content": "Fig. 3: Pipeline of Stage1. The illumination maps are corrected in the temporal dimension by STE, which alleviates the difficulty of subsequent temporal data learning using networks. The deflickering priors can be extracted: " + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\tilde{V}_t" + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "inline_equation", + "content": "S_{flicker}" + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "content": "preparation. This initial stage leverages STE to correct the histogram within an illumination space representation. Key outputs from this stage include the filtered illumination map " + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\{\\tilde{V}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "content": ", identifying the index set of singular frames " + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "inline_equation", + "content": "S_{flicker}" + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "content": ", and generating exposure maps " + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\{M_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "content": ". These elements serve as crucial prior information, guiding the deflickering process in subsequent stages. (2) Global and local flicker removal. The second stage involves two flicker removal strategies. A global flicker removal module utilizes the filtered illumination map " + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\{\\tilde{V}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "content": " to address temporal flicker issues across the video. A local over-/under-exposed region flicker removal module instructed by the singular frames set " + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "inline_equation", + "content": "S_{flicker}" + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "content": " and exposure maps " + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\{M_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "content": ", restores the texture in the over-/under-exposure region to compensate for the lost high-frequency details. (3) Adaptive temporal consistency. The final stage introduces a spatio-temporal network under the guidance of an adaptive warping loss to improve video coherence and temporal consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 462, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 498 + ], + "type": "text", + "content": "These stages delineated in the following Sec. 4.2, Sec. 4.3, and Sec. 4.4, together provide a comprehensive solution to enhance the visual quality and temporal consistency of video content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 514, + 356, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 356, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 356, + 528 + ], + "type": "text", + "content": "4.2 Stage1: Deflickering priors preparation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Spatio-temporal challenge. An effective BVD model should strategically avoid the intricacies of spatio-temporal complexity across pixels, nevertheless, current deep learning approaches to BVD have not adequately addressed this challenge. For instance, the state-of-the-art BVD method, Deflicker [21], employs a neural atlas [19] as a unified representation of the entire video. While this can offer consistent guidance for the deflickering process, the effectiveness of neural atlases significantly depends on the video duration and the network capacity. Processing an 8-second video to create an atlas can take around 10 minutes, and longer videos exacerbate the demands on time, memory, and computational resources. Consequently, there is a further need for research into more compact representations that can facilitate deflickering without these limitations." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 0 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 163 + ], + "type": "text", + "content": "Illumination histogram. We introduce histogram representation [3], which utilizes statistical distribution to illustrate the global properties of image pixel values. As illustrated in Fig. 3, we first extract the illumination maps " + }, + { + "bbox": [ + 132, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\{V_t\\}_{t=1}^T" + }, + { + "bbox": [ + 132, + 116, + 479, + 163 + ], + "type": "text", + "content": " according to the conversion between HSV and RGB color space:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 170, + 481, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 170, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 217, + 170, + 481, + 188 + ], + "type": "interline_equation", + "content": "V _ {t} = \\max _ {R, G, B \\in X _ {t}} \\{R, G, B \\}, t \\in \\{1, 2, \\dots , T \\}. \\tag {4}", + "image_path": "e845a09840811d55d2775ee75251ed3c41dcec6ff82e7ff2f193c69a893c2cbd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 479, + 300 + ], + "type": "text", + "content": "Next, by applying Eq. (1) to each illumination map, we utilize illumination histograms " + }, + { + "bbox": [ + 130, + 193, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\{H_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 193, + 479, + 300 + ], + "type": "text", + "content": " to represent the illumination characteristics of video frames. These illumination histograms reflect important flickering information: As shown in Fig. 3, frames without distinct illumination changes have similar histograms, while the histogram of a flicker frame has a distribution distinguished from others. Furthermore, spikes in the head or tail area of the distribution indicate potential over-/under-exposure problems in the image. To this end, reasonable utilization of histogram information is significant for flicker elimination and alleviation of spatio-temporal complexity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "type": "text", + "content": "Deflickering priors. We introduce deflickering priors extracted from illumination histograms to facilitate the deflickering process. Initially, considering that the illumination histograms of natural videos exhibit continuous changes in the temporal dimension, we utilize Scale-Time Equalization (STE) as defined in Eq. (3) to smooth the illumination histograms, as illustrated in Fig. 3. Through histogram matching, we further obtain filtered illumination maps " + }, + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "type": "inline_equation", + "content": "\\{\\tilde{V}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "type": "text", + "content": " that exhibit no obvious flicker compared to the original ones, thereby enabling the subsequent independent processing of each frame by the neural network. Secondly, we identify the singular frames set " + }, + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "type": "inline_equation", + "content": "S_{flicker}" + }, + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "type": "text", + "content": " containing the indexes of flicker frames, based on the divergence of histograms before and after STE:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 426, + 480, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 426, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 139, + 426, + 480, + 456 + ], + "type": "interline_equation", + "content": "S _ {f l i c k e r} = \\left\\{t \\in [ 1, T ] \\mid \\mathrm {K L} \\left(\\tilde {H} _ {t} \\parallel H _ {t}\\right) > \\bar {\\epsilon} _ {t} \\right\\}, \\quad \\bar {\\epsilon} _ {t} = \\frac {1}{2 n + 1} \\sum_ {\\tau = t - n} ^ {t + n} \\mathrm {K L} \\left(\\tilde {H} _ {\\tau} \\parallel H _ {\\tau}\\right), \\tag {5}", + "image_path": "bc495ca5d9321a23fb0bb89bd3a7acd2bb9b0a0c1fc222aff3ae7f03a905f3b1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "text", + "content": " is the moving average radius. Finally, as shown in Fig. 2, we extract the exposure maps " + }, + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "inline_equation", + "content": "\\{M_t \\in \\mathbb{R}^{H \\times W}\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "text", + "content": " of local over-/under-exposure according to the filtered illumination map:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 503, + 480, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 503, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 212, + 503, + 480, + 534 + ], + "type": "interline_equation", + "content": "M _ {t} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\tilde {V} (i, j) < \\epsilon_ {1} \\text {o r} \\tilde {V} (i, j) > \\epsilon_ {2}, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}", + "image_path": "71f5f34e2275b63cd95a66473923255dfc22f493dd1a4de3ed51d57fcac5c0af.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 538, + 479, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 479, + 562 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 538, + 479, + 562 + ], + "type": "inline_equation", + "content": "\\epsilon_{1}" + }, + { + "bbox": [ + 130, + 538, + 479, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 538, + 479, + 562 + ], + "type": "inline_equation", + "content": "\\epsilon_{2}" + }, + { + "bbox": [ + 130, + 538, + 479, + 562 + ], + "type": "text", + "content": " are darkness and exposure thresholds, respectively, and we consider illumination values belonging to " + }, + { + "bbox": [ + 130, + 538, + 479, + 562 + ], + "type": "inline_equation", + "content": "[\\epsilon_1,\\epsilon_2]" + }, + { + "bbox": [ + 130, + 538, + 479, + 562 + ], + "type": "text", + "content": " as not posing an exposure risk." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 563, + 479, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 479, + 609 + ], + "type": "text", + "content": "In summary, we leverage illumination histograms to prepare deflickering priors including filtered illumination maps " + }, + { + "bbox": [ + 130, + 563, + 479, + 609 + ], + "type": "inline_equation", + "content": "\\{\\tilde{V}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 130, + 563, + 479, + 609 + ], + "type": "text", + "content": ", exposure maps " + }, + { + "bbox": [ + 130, + 563, + 479, + 609 + ], + "type": "inline_equation", + "content": "\\{M_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 563, + 479, + 609 + ], + "type": "text", + "content": ", and the singular frames set " + }, + { + "bbox": [ + 130, + 563, + 479, + 609 + ], + "type": "inline_equation", + "content": "S_{flicker}" + }, + { + "bbox": [ + 130, + 563, + 479, + 609 + ], + "type": "text", + "content": ", for alleviating the spatio-temporal complexity and enhancing deflickering in the subsequent learning stages." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 624, + 360, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 360, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 360, + 636 + ], + "type": "text", + "content": "4.3 Stage2: Global and local flicker removal" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "In this stage, simple 2D spatial neural networks are integrated with deflickering priors to eliminate both global and local flickering flaws. Here, \"global\" refers" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "X. Qiu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "to obvious temporal inconsistency due to illumination fluctuation, while \"local\" refers to lost texture details due to exposure and darkness. This approach allows for the direct generation of high-quality videos from flicker frames, regardless of the spatio-temporal complexity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": "Global deflickering. BlazeBVD is equipped with a global flicker removal module (GFRM), which incorporates a 2D-Unet to correct each frame " + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": " with the guidance of the filtered illumination map " + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\tilde{V}_{t}" + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": ". In other words, the corrected image " + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t} = U\\mathrm{net}_{\\theta}(X_{t},\\tilde{V}_{t})" + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": " denotes the parameters of the 2D-Unet. The network is trained on the BVD dataset " + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": " by minimizing the L2 loss function between the predicted images and the groundtruth images:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 244, + 480, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 244, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 224, + 244, + 480, + 258 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\mathbb {E} _ {\\left(X _ {t}, G _ {t}\\right) \\in \\mathcal {D}} \\| U n e t _ {\\theta} \\left(X _ {t}, \\tilde {V} _ {t}\\right) - G _ {t} \\|. \\tag {7}", + "image_path": "33d34d3e90ab1dd245471d264563c3d15b7598a6a956beba0af43884765af461.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "text", + "content": "Local deflickering. BlazeBVD is also equipped with a local flicker removal module (LFRM) to enhance local texture details. As depicted in Fig. 5, overexposure can transform intricate texture in local regions into uniform pure, while local under-exposure can result in black areas. This loss of detailed texture cannot be compensated for by previous BVD methods, such as Deflicker [21]. We leverage LFRM to transfer detailed information from neighboring frames to these ill regions. Given the corrected singular frame " + }, + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "inline_equation", + "content": "\\tilde{X}_t" + }, + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "inline_equation", + "content": "S_{\\text{flicker}}" + }, + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "text", + "content": " with local exposure, along with the previous frame " + }, + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "inline_equation", + "content": "\\tilde{X}_{t-1}" + }, + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "text", + "content": " and the next frame " + }, + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "inline_equation", + "content": "\\tilde{X}_{t+1}" + }, + { + "bbox": [ + 130, + 264, + 480, + 373 + ], + "type": "text", + "content": ", we first compute the optical flow to extract part of the detailed texture:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 182, + 381, + 480, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 381, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 182, + 381, + 480, + 394 + ], + "type": "interline_equation", + "content": "o _ {\\tau \\rightarrow t} = \\operatorname {F l o w} \\left(X _ {\\tau}, X _ {t}\\right), \\tau \\in \\{t + 1, t - 1 \\} \\text {a n d} t \\in S _ {\\text {f l i c k e r}}. \\tag {8}", + "image_path": "b5df41c944e55fcd5d90076da3e6205c6a1e611afcd3e934947e175fbbf503e4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "text", + "content": "Here Flow is the optical flow estimation function, in our work we set it as the popular RAFT model [11, 37]. We migrate the possible detail textures of the preceding and following frames into the current frame:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 172, + 445, + 480, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 445, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 172, + 445, + 480, + 466 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {t} = \\mathcal {F} \\left(M _ {t} \\odot \\mathcal {W} \\left(\\tilde {X} _ {t - 1}, o _ {t - 1 \\rightarrow t}\\right) + (1 - M _ {t}) \\odot \\tilde {X} _ {t}, \\tilde {X} _ {t}, \\right. \\tag {9}", + "image_path": "b50c729059cd3cae6a96d437e134d5a71d854db76648b4394b6be1f86445e06f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 209, + 462, + 442, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 462, + 442, + 475 + ], + "spans": [ + { + "bbox": [ + 209, + 462, + 442, + 475 + ], + "type": "interline_equation", + "content": "M _ {t} \\odot \\mathcal {W} \\left(\\tilde {X} _ {t - 1}, o _ {t + 1 \\rightarrow t}\\right) + \\left(1 - M _ {t}\\right) \\odot \\tilde {X} _ {t}), t \\in S _ {\\text {f l i c k e r}},", + "image_path": "457491d27f624f6bfaebd74a37385d2fb610254904a61b02f12ab4b12ac1ee7f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 483, + 449, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 449, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 449, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 483, + 449, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 483, + 449, + 495 + ], + "type": "text", + "content": " is the warping operator and " + }, + { + "bbox": [ + 130, + 483, + 449, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 483, + 449, + 495 + ], + "type": "text", + "content": " is the fusion network from [41]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 512, + 356, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 512, + 356, + 524 + ], + "spans": [ + { + "bbox": [ + 131, + 512, + 356, + 524 + ], + "type": "text", + "content": "4.4 Stage3: Adaptive temporal consistency" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 530, + 480, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 480, + 613 + ], + "type": "text", + "content": "The former stages have well addressed the main illumination fluctuation and exposure problems in BVD task. Nevertheless, a video may also have content inconsistencies to largely degrade the visual effect, and the former stage may also cause this. Therefore, We introduce the Temporal Consistency Model (TCM) to refine video " + }, + { + "bbox": [ + 130, + 530, + 480, + 613 + ], + "type": "inline_equation", + "content": "\\{\\hat{X}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 530, + 480, + 613 + ], + "type": "text", + "content": " as the final outputs " + }, + { + "bbox": [ + 130, + 530, + 480, + 613 + ], + "type": "inline_equation", + "content": "\\{O_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 530, + 480, + 613 + ], + "type": "text", + "content": ". TCM is developed based on the architecture by RTN [40], but with a temporal consistency loss of adaptive masks to refine local artifacts:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 170, + 620, + 432, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 620, + 432, + 633 + ], + "spans": [ + { + "bbox": [ + 170, + 620, + 432, + 633 + ], + "type": "interline_equation", + "content": "E _ {\\text {p a i r}} ^ {\\prime} \\left(O _ {t}, O _ {s}\\right) = \\left\\| W _ {t} \\odot \\left(M _ {t} + 1\\right) \\odot M _ {t, s} \\odot \\left(O _ {t} - \\mathcal {W} \\left(O _ {s}\\right)\\right) \\right\\| _ {1},", + "image_path": "cda94dc333e3746eca18ae7cafc2c807985ef742727b99acc5cbc7d9be1ac0d4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 205, + 637, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 637, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 205, + 637, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w a r p} = \\frac {1}{T - 1} \\sum_ {t = 2} ^ {T} \\left\\{E _ {\\text {p a i r}} ^ {\\prime} \\left(O _ {t}, O _ {1}\\right) + E _ {\\text {p a i r}} ^ {\\prime} \\left(O _ {t}, O _ {t - 1}\\right) \\right\\}, \\tag {10}", + "image_path": "2539ac582746630751cbebd82d31c224db3e9790b34c880f8f8c95de9184c3a4.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 0 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "W_{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is the loss weight for the over-/under-exposure area and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is the mask obtained from the former stage. " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is the warping operator and the frame after warping is computed by optical flow estimation [11,37]. " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "M_{t,s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " represents the mask after optical flow estimation. This loss avoids the video blurring problem introduced in Deflicker [21]. The training loss of TCM consists of reconstruction loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", perceptual loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{per}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " [17], spatio-temporal adversarial loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " [6] and designed adaptive weighted warping loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{warp}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". Combining all the above losses, the overall objective to optimize TCM is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 185, + 223, + 481, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 223, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 185, + 223, + 481, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T C M} = \\alpha_ {1} * \\mathcal {L} _ {r e c} + \\alpha_ {2} * \\mathcal {L} _ {p e r} + \\alpha_ {3} * \\mathcal {L} _ {a d v} + \\alpha_ {4} * \\mathcal {L} _ {w a r p}. \\tag {11}", + "image_path": "a41067e85a78f0540c2fe657b43d00822ab2231c472c27fa139eeac77c44c754.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "text", + "content": "In our work, we set " + }, + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 1.0" + }, + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "inline_equation", + "content": "\\alpha_{2} = 1.0" + }, + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "inline_equation", + "content": "\\alpha_{3} = 0.01" + }, + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "inline_equation", + "content": "\\alpha_{4} = 0.1" + }, + { + "bbox": [ + 131, + 244, + 414, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 275, + 230, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 230, + 290 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 230, + 290 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 301, + 205, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 205, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 205, + 312 + ], + "type": "text", + "content": "5.1 Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 321, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 481, + 357 + ], + "type": "text", + "content": "We inherit the training dataset DAVIS-2017-Train, MS-COCO and the testing dataset Blind Deflickering Dataset from Deflicker [21], including Synthetic videos from DAVIS-2017-Val, Real-world videos and Generation videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "type": "text", + "content": "Training dataset. MS-COCO [23] contains 118287 images in the train set and is used for the pre-training of Unet in GFRM. DAVIS-2017-Train [20, 28, 32] consists of 60 sequences with a total of 4219 labeled frames, with a video frame rate of 24fps and a resolution of " + }, + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "type": "inline_equation", + "content": "480\\mathrm{p}" + }, + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "type": "text", + "content": ". We create the synthetic process that provides the underlying facts for quantitative analysis. The training dataset is degraded by performing flicker on DAVIS-2017-Train, as shown in Eq. (12):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 240, + 441, + 481, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 441, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 240, + 441, + 481, + 453 + ], + "type": "interline_equation", + "content": "X _ {t} = G _ {t} + F _ {t}, t \\in \\{1, 2, \\dots , T \\}, \\tag {12}", + "image_path": "a2f1b95b989d3315b7f96f9a0ef758cb85cbb0895232b57a6310210ba868c45e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\{G_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": " are clean frames and " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\{F_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": " are random synthesized flickering artifacts. The window size " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "F_t" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": " has been set to denote the number of frames that share the same flickering artifacts, random sampling between [2, 12]. The value of " + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": " controls the degree of flicker in the short- and long-term of the video." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 616 + ], + "type": "text", + "content": "Testing dataset. We conduct Synthetic videos, Real-world videos and Generation videos respectively, from the Blind Deflickering Dataset [21] and DAVIS2017-Test processed by Eq. (12). Among them, Synthetic videos and DAVIS2017-Test have ground truth and the others do not. We can only use no-referenced metric and qualitative observation. Generation videos refer to the processed videos generated by generative algorithms, which may lead to flicker artifacts due to the imperfections of the algorithms. Real-world videos including old movies, old anime, slow-motion, etc. are directly captured, unprocessed videos in the real world." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Comparison methods. We compare our BlazeBVD with other 4 methods, including STE [8], ConvLSTM [20], DVP [22] and Deflicker [21]. We set Deflicker as the baseline, since it is the first publicly presented method for BVD task and is currently the most advanced approach for general videos." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "X. Qiu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 395 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 395 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 395 + ], + "type": "image", + "image_path": "075647226b6f736da970113ef77697ab5554a0abab0a1d77e24f5002c1902aef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 415, + 482, + 481 + ], + "lines": [ + { + "bbox": [ + 132, + 415, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 482, + 481 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons between previous methods and our BlazeBVD. Our method removes flicker and restores details in over-exposed regions (red boxes in row 6 and row 8) while avoiding color artifacts (row 2 and row 4). Besides, BlazeBVD also ensures the fidelity of the video content and avoids color distortion (cyan boxes in row 5 and row 7). Zoom in for the best view and we recommend watching videos in the supplementary materials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 503, + 381, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 381, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 381, + 515 + ], + "type": "text", + "content": "5.2 Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "Metrics. We quantitatively compare the differences between our method and others with reference and no-reference metrics, including PSNR, SSIM [45], and " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " [22]. Among them, " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " is a no-reference metric used in videos. For each frame " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "O_t" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": ", we calculate the warping error with frame " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "O_{t-1}" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " and the first frame " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "O_1" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " for considering both short-term and long-term consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Quantitative results. We compare our BlazeBVD with the state-of-the-art method Deflicker [21], and other consistency methods. Table 1 and Table 2 provide the comparative experimental results of various methods on DAVIS-2017-Test and Synthetic videos, where PSNR and SSIM of the resulting video are higher than those of the other methods. This shows that BlazeBVD contributes to the fidelity maintenance of video content, as well as the correctness of the missing details complement. The warping error of our method is a little higher" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 0 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 481, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons of synthetic videos in DAVIS-2017-Test. Our method performs best in terms of PSNR and SSIM and " + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": " of all the methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 481, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 481, + 205 + ], + "type": "table", + "html": "
MethodRawVideoConvLSTMDVPSTEDeflickerOursGT
PSNR↑-24.11024.13626.13823.93228.609-
SSIM↑-0.92560.92630.93210.92430.9638-
E_warp ↓0.14290.13690.14150.11170.08400.08250.1097
", + "image_path": "0de3d15abc8656f74688e7957dd48bba86f2247460d10d92e7558e854235aa75.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 259, + 482, + 367 + ], + "blocks": [ + { + "bbox": [ + 132, + 216, + 481, + 250 + ], + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 250 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons in Synthetic videos from Blind Deflickering Dataset. " + }, + { + "bbox": [ + 132, + 216, + 481, + 250 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 132, + 216, + 481, + 250 + ], + "type": "text", + "content": " is the window size and represents the number of frames that share the same flicker artifact. " + }, + { + "bbox": [ + 132, + 216, + 481, + 250 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 216, + 481, + 250 + ], + "type": "text", + "content": " denotes the local seed window size." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 259, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 132, + 259, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 482, + 367 + ], + "type": "table", + "html": "
VideoTypeRawE_warp↓PSNR↑
DeflickerOursGTDeflickerOursDeflickerOurs
Synthetic
-W=10.19330.08480.08950.105625.931229.67380.94030.9539
-W=30.15220.08440.08900.105625.499628.82720.93800.9555
-W=100.13070.08550.09040.105624.384226.99660.93440.9543
-L=30.11130.08750.09000.105625.387427.08810.92710.9517
Average0.14690.08560.08970.105625.300628.14640.93490.9538
", + "image_path": "0a76f09d83b159d396f1858f0e7b49524d8de90be6356f5c095e4044946ee8d6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 481, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 520 + ], + "type": "text", + "content": "than Deflicker, but it is closer to the ground truth. Through analysis, we find that warping error has inaccuracy in the estimation of optical flow, and this error is masked by image blurring in Deflicker. Therefore, we believe that " + }, + { + "bbox": [ + 130, + 390, + 481, + 520 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 130, + 390, + 481, + 520 + ], + "type": "text", + "content": ", which has the same magnitude as baseline and GT, is the valid experimental comparison result. In Tab. 3, we compare BlazeBVD and Deflicker on Real-world videos and Generation videos. Since the videos in these two datasets with no reference are inherently discontinuous, where " + }, + { + "bbox": [ + 130, + 390, + 481, + 520 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 130, + 390, + 481, + 520 + ], + "type": "text", + "content": " cannot directly represent video continuity, we present the comparative results of user studies. To evaluate the perceived preference between the baseline Deflicker and the proposed method, each user needs to select a video with better perceptual quality between these two choices. In total, there are 50 users and 50 pairs in 7 datasets of comparisons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "content": "Qualitative results. Both ConvLSTM and DVP are temporal consistency models that require a reference, importing the input video as the reference, and the processed video cannot remove significant flicker, as shown in Fig. 4. Additionally, the flicker caused by over-/under-exposure is prone to lose the local texture (column 1 in Fig. 4). The atlas-based representation in Deflicker cannot accurately locate these points with different local features, resulting in the loss of all details on the processed video. Instead, our method designs local positioning and adjacent frame details transfer in LFRM to preserve part of the texture on the road, as shown in the red box of Fig. 4. We also reveal the color artifact in column 4 of Fig. 4 that appears at baseline, due to the excessive weight of warping error in the training loss when refining temporal consistency. Therefore, our designed adaptive mask weighted training loss in TCM improves the temporal" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "X. Qiu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 479, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: User studies of Real-world videos and Generation videos. Since the videos in these two datasets are inherently discontinuous, " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " cannot properly reflect temporal consistency. Here we only present the comparative results of the user study." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 479, + 216 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 479, + 216 + ], + "type": "table", + "html": "
TypeExpertOldAnimeOldMovieSlowMotionTimeLapseVideoDMVideoLDM
Raw(E_warp)0.07540.10450.10990.07240.07320.12480.1601
Deflicker38.67%19.25%22.83%37.67%31.43%30.50%26.34%
Ours61.33%80.75%77.17%62.33%68.57%69.50%73.66%
", + "image_path": "f4cfcdbced6ce05050e318b0b00a913e4a121aaa5f82ae883342a6dc5674c47e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 239, + 479, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 262 + ], + "type": "text", + "content": "consistency while avoiding image blurring. More visualization results are in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 281, + 246, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 246, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 246, + 293 + ], + "type": "text", + "content": "5.3 Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 373 + ], + "type": "text", + "content": "STE filtering. We contrast the difference between introducing STE filtering in illumination space and directly in color space. As shown in the red box of Fig. 5, directly performing histogram correction in color channels separately introduces color artifacts and color distortion. Because the filtering of non-gray images in the three channels is relatively independent, the recombination after nonlinear transformation would produce such defects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 375, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 542 + ], + "type": "text", + "content": "GFRM and LFRM. We compare the important roles of two modules, GFRM and LFRM. As shown in Tab. 5, removing GFRM and LFRM reduces PSNR and SSIM and increases " + }, + { + "bbox": [ + 130, + 375, + 482, + 542 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 130, + 375, + 482, + 542 + ], + "type": "text", + "content": " of inference videos. It can be inferred that GFRM needs to initially remove global flicker between frames, otherwise subsequent modules cannot maintain the temporal consistency, while LFRM maintains the texture fidelity of the video content, which is reflected in the cyan box of Fig. 5. Designed weighted loss. As shown in Tab. 5, the lightweight spatio-temporal network TCM reduces " + }, + { + "bbox": [ + 130, + 375, + 482, + 542 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 130, + 375, + 482, + 542 + ], + "type": "text", + "content": ", which refines the temporal consistency of processed videos. We also compare the effect of using unweighted training loss and the adaptive mask weighted loss on this temporal network. Figure 5 provides the resulting frame with unweighted training, which is generally blurred and suffers from color distortion, while the weighted training keeps the frame clear and faithful in most areas. Adaptive mask weighting improves the temporal consistency of the inference video, improving the inter-frame continuity more pertinently." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 560, + 214, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 214, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 214, + 572 + ], + "type": "text", + "content": "5.4 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "Advantages and limitations. BlazeBVD is over " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": " faster than the inference speed of baseline using the atlas representation, which is specifically recorded in Tab. 4. Our method also takes advantage of explicit prior extraction in the BVD task, which provides directions for improvement in the subsequent work. And the cause of local flicker has been noted and effectively dealt with from both global and local perspectives. However, our method still has drawbacks in LFRM. Due to the inaccurate optical flow motion estimation, the fusion network" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 0 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 277, + 182 + ], + "lines": [ + { + "bbox": [ + 132, + 117, + 277, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 277, + 182 + ], + "type": "text", + "content": "Table 4: Inference time and model parameters comparison. Our approach is faster than the baseline in inference time with less memory, and comparable in MACs and Params." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 133, + 193, + 277, + 261 + ], + "blocks": [ + { + "bbox": [ + 133, + 193, + 277, + 261 + ], + "lines": [ + { + "bbox": [ + 133, + 193, + 277, + 261 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 277, + 261 + ], + "type": "table", + "html": "
MethodDeflickerOurs
Time/80p614.19s58.37s
Memory5204M1389M
MACs260.7G251.4G
Params12.48M17.77M
", + "image_path": "fc77fa3b8b0b70246e6f42d11b1f79850083434f3d98c24b5dec71e82c02bcf1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 287, + 168, + 489, + 264 + ], + "blocks": [ + { + "bbox": [ + 287, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 287, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 287, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 5: Ablation studies in DAVIS-2017- Test. w/o STE replaces STE with mean filtering in Stage1 and w/o weighted denotes the warping loss without mask weighting in TCM." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 287, + 168, + 489, + 264 + ], + "lines": [ + { + "bbox": [ + 287, + 168, + 489, + 264 + ], + "spans": [ + { + "bbox": [ + 287, + 168, + 489, + 264 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑E_warp ↓
w/o STE26.44120.95170.1011
w/o GFRM23.89390.92440.1225
w/o LFRM27.13660.94620.1122
w/o TCM29.91300.95830.1093
w/o weighted28.10490.95230.0849
BlazeBVD28.60920.96380.0825
", + "image_path": "694001cc9249d203195a39754f5001aad4db4f5c445af1ebeb4289fab833eeaa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 276, + 219, + 325 + ], + "blocks": [ + { + "bbox": [ + 133, + 276, + 219, + 325 + ], + "lines": [ + { + "bbox": [ + 133, + 276, + 219, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 276, + 219, + 325 + ], + "type": "image", + "image_path": "5740c8ba744f729da186f344f8354be7a3ea584b7ad58c1f6c4b14a18326bfae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 326, + 185, + 335 + ], + "lines": [ + { + "bbox": [ + 164, + 326, + 185, + 335 + ], + "spans": [ + { + "bbox": [ + 164, + 326, + 185, + 335 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 276, + 306, + 325 + ], + "blocks": [ + { + "bbox": [ + 220, + 276, + 306, + 325 + ], + "lines": [ + { + "bbox": [ + 220, + 276, + 306, + 325 + ], + "spans": [ + { + "bbox": [ + 220, + 276, + 306, + 325 + ], + "type": "image", + "image_path": "e24c203e70391a228155b2de7d6901dbcb8cb64e660a926e5b61c32771434e2b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 326, + 285, + 335 + ], + "lines": [ + { + "bbox": [ + 241, + 326, + 285, + 335 + ], + "spans": [ + { + "bbox": [ + 241, + 326, + 285, + 335 + ], + "type": "text", + "content": "w/o GFRM" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 276, + 393, + 325 + ], + "blocks": [ + { + "bbox": [ + 308, + 276, + 393, + 325 + ], + "lines": [ + { + "bbox": [ + 308, + 276, + 393, + 325 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 393, + 325 + ], + "type": "image", + "image_path": "3d256e30443406c2ddd2af9e7de0467971f09eb7078a83aab616c3c952eb2a68.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 326, + 371, + 335 + ], + "lines": [ + { + "bbox": [ + 329, + 326, + 371, + 335 + ], + "spans": [ + { + "bbox": [ + 329, + 326, + 371, + 335 + ], + "type": "text", + "content": "w/o LFRM" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 395, + 276, + 481, + 325 + ], + "blocks": [ + { + "bbox": [ + 395, + 276, + 481, + 325 + ], + "lines": [ + { + "bbox": [ + 395, + 276, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 395, + 276, + 481, + 325 + ], + "type": "image", + "image_path": "bb73a782de945ac40d6a1122a84cc82c9c520ba5445cc47725f55d91e9c4746a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 418, + 326, + 451, + 335 + ], + "lines": [ + { + "bbox": [ + 418, + 326, + 451, + 335 + ], + "spans": [ + { + "bbox": [ + 418, + 326, + 451, + 335 + ], + "type": "text", + "content": "w/o STE" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 336, + 219, + 385 + ], + "blocks": [ + { + "bbox": [ + 133, + 336, + 219, + 385 + ], + "lines": [ + { + "bbox": [ + 133, + 336, + 219, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 336, + 219, + 385 + ], + "type": "image", + "image_path": "1741435d2c81bde25c30d85a62d1fde0e763015d13ddc8405ac50eb91c5db22f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 386, + 194, + 395 + ], + "lines": [ + { + "bbox": [ + 156, + 386, + 194, + 395 + ], + "spans": [ + { + "bbox": [ + 156, + 386, + 194, + 395 + ], + "type": "text", + "content": "w/o TCM" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 220, + 336, + 306, + 384 + ], + "blocks": [ + { + "bbox": [ + 220, + 336, + 306, + 384 + ], + "lines": [ + { + "bbox": [ + 220, + 336, + 306, + 384 + ], + "spans": [ + { + "bbox": [ + 220, + 336, + 306, + 384 + ], + "type": "image", + "image_path": "c50d7a915ab75c7a4f8e9c2c68607b78d3b4d3b3715afe44f5f44a2402fed645.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 386, + 298, + 396 + ], + "lines": [ + { + "bbox": [ + 225, + 386, + 298, + 396 + ], + "spans": [ + { + "bbox": [ + 225, + 386, + 298, + 396 + ], + "type": "text", + "content": "w/o weighted mask" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 308, + 336, + 393, + 384 + ], + "blocks": [ + { + "bbox": [ + 308, + 336, + 393, + 384 + ], + "lines": [ + { + "bbox": [ + 308, + 336, + 393, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 336, + 393, + 384 + ], + "type": "image", + "image_path": "efbd9145fe52c70a20930cc62857d55982fdb628d4cd3e937e29c9ae4bf55259.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 386, + 356, + 395 + ], + "lines": [ + { + "bbox": [ + 335, + 386, + 356, + 395 + ], + "spans": [ + { + "bbox": [ + 335, + 386, + 356, + 395 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 417, + 482, + 462 + ], + "lines": [ + { + "bbox": [ + 131, + 417, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 482, + 462 + ], + "type": "text", + "content": "Fig. 5: Qualitative ablation studies of key designs in BlazeBVD. Specifically, local detail loss is represented in cyan box, color artifact is represented in the red box, temporal inconsistency is represented in the black box, and color distortion is represented in the blue box. Zoom in for the best view." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 395, + 336, + 481, + 384 + ], + "blocks": [ + { + "bbox": [ + 395, + 336, + 481, + 384 + ], + "lines": [ + { + "bbox": [ + 395, + 336, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 395, + 336, + 481, + 384 + ], + "type": "image", + "image_path": "e169edf276ea632d6be578380b60d3ae9cc3e618564091f7c4a91567fb09c41b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 386, + 443, + 394 + ], + "lines": [ + { + "bbox": [ + 429, + 386, + 443, + 394 + ], + "spans": [ + { + "bbox": [ + 429, + 386, + 443, + 394 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 488, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 488, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 482, + 525 + ], + "type": "text", + "content": "cannot convey the local texture of adjacent frames, and there are still slight edge artifacts and color distortion that cannot be eliminated. We hope to continue exploring the transfer and fusion of local region information between frames [40]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 549, + 220, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 220, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 220, + 562 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we propose BlazeBVD, a universal approach for the blind video deflickering task. The core of our method is to prepare flicker priors within the STE filter in illumination space first, and then utilize these priors to correct the global flicker and local exposed texture. GFRM and LFRM leverage the priors to remove global and local flicker in several adjacent frames with less computation. Finally, TCM is used to improve video coherence and inter-frame consistency. Extensive experiments validate the effectiveness of the designed modules." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "X. Qiu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 140, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 482, + 177 + ], + "type": "text", + "content": "This paper is supported by the National Natural Science Foundation of China (Nos. U23B2012, 11991022), and the Fundamental Research Funds for the Central Universities (E3E41904X2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 223, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 223, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 245 + ], + "type": "text", + "content": "1. Affifi, M., Brubaker, M.A., Brown, M.S.: Histogram: Controlling colors of gangenerated and real images via color histograms. In: CVPR. pp. 7941-7950 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 246, + 463, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 463, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 463, + 255 + ], + "type": "text", + "content": "2. Anarchy, D.: Flicker free, https://digitalanarchy.com/Flicker/main.html" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 481, + 289 + ], + "type": "text", + "content": "3. Bassiou, N., Kotropoulos, C.: Color image histogram equalization by absolute discounting back-off. Computer Vision and Image Understanding 107, 108-122 (Jul 2007)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 290, + 481, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 481, + 312 + ], + "type": "text", + "content": "4. Bonneel, N., Tompkin, J., Sunkavalli, K., Sun, D., Paris, S., Pfister, H.: Blind video temporal consistency. ACM Transactions on Graphics 34, 1-9 (Nov 2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "type": "text", + "content": "5. Bottenus, N., Byram, B.C., Hyun, D.: Histogram matching for visual ultrasound image comparison. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control 68, 1487-1495 (May 2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 346, + 481, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 481, + 367 + ], + "type": "text", + "content": "6. Chang, Y.L., Liu, Z.Y., Lee, K.Y., Hsu, W.: Free-form video inpainting with 3d gated convolution and temporal patchgan. In: ICCV. pp. 9066-9075 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 368, + 481, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 481, + 400 + ], + "type": "text", + "content": "7. Chu, M., Xie, Y., Mayer, J., Leal-Taixe, L., Thuerey, N.: Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics 39, 75-1 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 401, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 481, + 422 + ], + "type": "text", + "content": "8. Delon, J.: Movie and video scale-time equalization application to flicker reduction. IEEE Transactions on Image Processing 15, 241-248 (Jan 2006)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 423, + 481, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 455 + ], + "type": "text", + "content": "9. Delon, J., Desolneux, A.: Stabilization of flicker-like effects in image sequences through local contrast correction. SIAM Journal on Imaging Sciences 3, 703-734 (Jan 2010)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 456, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 488 + ], + "type": "text", + "content": "0. Dosovitskiy, A., Fischer, P., Ilg, E., Hausser, P., Hazirbas, C., Golkov, V., Van Der Smagt, P., Cremers, D., Brox, T.: Flownet: Learning optical flow with convolutional networks. In: ICCV. pp. 2758-2766 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 489, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 481, + 510 + ], + "type": "text", + "content": "1. Eslami, N., Arefi, F., Mansourian, A.M., Kasaei, S.: Rethinking raft for efficient optical flow. arXiv preprint arXiv:2401.00833 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 512, + 481, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 543 + ], + "type": "text", + "content": "2. Huang, J., Liu, Y., Fu, X., Zhou, M., Wang, Y., Zhao, F., Xiong, Z.: Exposure normalization and compensation for multiple-exposure correction. In: CVPR. pp. 6043-6052 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "text", + "content": "3. Huang, J., Liu, Y., Zhao, F., Yan, K., Zhang, J., Huang, Y., Zhou, M., Xiong, Z.: Deep fourier-based exposure correction network with spatial-frequency interaction. In: ECCV. pp. 163-180 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "type": "text", + "content": "4. Huang, J., Zhao, F., Zhou, M., Xiao, J., Zheng, N., Zheng, K., Xiong, Z.: Learning sample relationship for exposure correction. In: CVPR. pp. 9904-9913 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "5. Huang, J., Zhou, M., Liu, Y., Yao, M., Zhao, F., Xiong, Z.: Exposure-consistency representation learning for exposure correction. In: ACM MM. pp. 6309-6317 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "6. Ilg, E., Mayer, N., Saikia, T., Keuper, M., Dosovitskiy, A., Brox, T.: Flownet 2.0: Evolution of optical flow estimation with deep networks. In: CVPR. pp. 2462-2470 (2017)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 0 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV. pp. 694-711 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "18. Kanj, A., Talbot, H., Luparello, R.R.: Flicker removal and superpixel-based motion tracking for high speed videos. In: ICIP. pp. 245-249 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 182 + ], + "type": "text", + "content": "19. Kasten, Y., Ofri, D., Wang, O., Dekel, T.: Layered neural atlases for consistent video editing. ACM Transactions on Graphics 40, 1-12 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "20. Lai, W.S., Huang, J.B., Wang, O., Shechtman, E., Yumer, E., Yang, M.H.: Learning blind video temporal consistency. In: ECCV. pp. 170-185 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "21. Lei, C., Ren, X., Zhang, Z., Chen, Q.: Blind video deflickering by neural filtering with a flawed atlas. In: CVPR. pp. 10439-10448 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "type": "text", + "content": "22. Lei, C., Xing, Y., Chen, Q.: Blind video temporal consistency via deep video prior. In: NeurIPS. pp. 1083-1093 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 277 + ], + "type": "text", + "content": "23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "type": "text", + "content": "24. Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: CVPR. pp. 5637-5646 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 319 + ], + "type": "text", + "content": "25. Mei, K., Patel, V.: Vidm: Video implicit diffusion models. In: AAAI. pp. 9117-9125 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "26. Moniz, J.R.A., Kang, E., Póczos, B.: Luciddream: Controlled temporally-consistent deepdream on videos. arXiv preprint arXiv:1911.11960 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "text", + "content": "27. Park, K., Woo, S., Kim, D., Cho, D., Kweon, I.S.: Preserving semantic and temporal consistency for unpaired video-to-video translation. In: ACM MM. pp. 1248-1257 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "28. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: CVPR (Jun 2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "text", + "content": "29. Pérez-Pellitero, E., Sajjadi, M.S., Hirsch, M., Scholkopf, B.: Perceptual video super resolution with enhanced temporal consistency. arXiv preprint arXiv:1807.07930 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "text", + "content": "30. Pfeuffer, A., Dietmayer, K.: Separable convolutional lstms for faster video segmentation. In: ITSC. pp. 1072-1078 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "31. Pitie, F., Kokaram, A., Dahyot, R.: Removing flicker from old movies. Master's thesis, University of Nice-Éophia Antipolis, France, Œe September (2002)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "text", + "content": "32. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv:1704.00675 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "text", + "content": "33. Qiu, X., Han, C., Zhang, Z., Li, B., Guo, T., Nie, X.: Diffbfr: Bootstrapping diffusion model for blind face restoration. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7785-7795 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 548, + 446, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 446, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 446, + 557 + ], + "type": "text", + "content": "34. RE:VISION: De:flicker, https://revisionfx.com/products/deflicker/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "text", + "content": "35. Saito, M., Saito, S., Koyama, M., Kobayashi, S.: Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. International Journal of Computer Vision p. 2586-2606 (Nov 2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "text", + "content": "36. Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In: CVPR. pp. 3626-3636 (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "text", + "content": "37. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402-419 (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "38. Thasarathan, H., Nazeri, K., Ebrahimi, M.: Automatic temporally coherent video colorization. In: CRV. pp. 189-194 (2019)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "X. Qiu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 544 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "39. Thimonier, H., Despois, J., Kips, R., Perrot, M.: Learning long-term style preserving blind video temporal consistency. In: ICME. pp. 1-6 (2021)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "40. Wan, Z., Zhang, B., Chen, D., Liao, J.: Bringing old films back to life. In: CVPR. pp. 17694-17703 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "text", + "content": "41. Wang, H., Xu, K., Lau, R.W.: Local color distributions prior for image enhancement. In: ECCV. pp. 343-359 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "42. Wang, W., Liu, J., Lin, Z., Yan, J., Chen, S., Low, C., Hoang, T., Wu, J., Liew, J.H., Yan, H., et al.: Magicvideo-v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "text", + "content": "43. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9168-9178 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "text", + "content": "44. Wang, Y., Peng, L., Li, L., Cao, Y., Zha, Z.J.: Decoupling-and-aggregating for image exposure correction. In: CVPR. pp. 18115-18124 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "type": "text", + "content": "45. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13, 600-612 (2004)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 303, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 335 + ], + "type": "text", + "content": "46. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV. vol. 39, pp. 7623-7633 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "type": "text", + "content": "47. Xu, R., Tu, Z., Du, Y., Dong, X., Li, J., Meng, Z., Ma, J., Bovik, A., Yu, H.: Pik-fix: Restoring and colorizing old photos. In: WACV. pp. 1724-1734 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "text", + "content": "48. Xu, X., Li, M., Sun, W., Yang, M.H.: Learning spatial and spatio-temporal pixel aggregations for image and video denoising. IEEE Transactions on Image Processing 29, 7153-7165 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "type": "text", + "content": "49. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 672-681 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 423, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 445 + ], + "type": "text", + "content": "50. Zhan, X., Pan, X., Liu, Z., Lin, D., Loy, C.C.: Self-supervised learning via conditional motion propagation. In: CVPR. pp. 1881-1889 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 445, + 482, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 478 + ], + "type": "text", + "content": "51. Zhang, F., Shao, Y., Sun, Y., Zhu, K., Gao, C., Sang, N.: Unsupervised low-light image enhancement via histogram equalization prior. arXiv preprint arXiv:2112.01766 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "text", + "content": "52. Zhang, Z., Li, B., Nie, X., Han, C., Guo, T., Liu, L.: Towards consistent video editing with text-to-image diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 511, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 544 + ], + "type": "text", + "content": "53. Zhou, Y., Xu, X., Shen, F., Gao, L., Lu, H., Shen, H.T.: Temporal denoising mask synthesis network for learning blind video temporal consistency. In: ACM MM. pp. 475-483 (2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 0 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/19f5daea-7c9c-46d5-b550-8a14478e364b_content_list.json b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/19f5daea-7c9c-46d5-b550-8a14478e364b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb9ae5ba3be27ee5edbcbec7155d795d03812fb2 --- /dev/null +++ b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/19f5daea-7c9c-46d5-b550-8a14478e364b_content_list.json @@ -0,0 +1,1543 @@ +[ + { + "type": "text", + "text": "BlenderAlchemy: Editing 3D Graphics with Vision-Language Models", + "text_level": 1, + "bbox": [ + 253, + 140, + 751, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ian Huang, Guandao Yang, and Leonidas Guibas", + "bbox": [ + 297, + 212, + 702, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stanford University", + "bbox": [ + 433, + 239, + 566, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Graphics design is important for various applications, including movie production and game design. To create a high-quality scene, designers usually need to spend hours in software like Blender, in which they might need to interleave and repeat operations, such as connecting material nodes, hundreds of times. Moreover, slightly different design goals may require completely different sequences, making automation difficult. In this paper, we propose a system that leverages Vision-Language Models (VLMs), like GPT-4V, to intelligently search the design action space to arrive at an answer that can satisfy a user's intent. Specifically, we design a vision-based edit generator and state evaluator to work together to find the correct sequence of actions to achieve the goal. Inspired by the role of visual imagination in the human design process, we supplement the visual reasoning capabilities of VLMs with \"imagined\" reference images from image-generation models, providing visual grounding of abstract language descriptions. In this paper, we provide empirical evidence suggesting our system can produce simple but tedious Blender editing sequences for tasks such as editing procedural materials and geometry from text and/or reference images, as well as adjusting lighting configurations for product renderings in complex scenes.", + "bbox": [ + 261, + 294, + 738, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 585, + 375, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To produce the compelling graphics content we see in movies or video games, 3D artists usually need to spend hours in software like Blender to find appropriate surface materials, object placements, and lighting arrangements. These operations require the artist to create a mental picture of the target, experiment with different parameters, and visually examine whether their edits get closer to the end goal. One can imagine automating these processes by converting language or visual descriptions of user intent into edits that achieve a design goal. Such a system can improve the productivity of millions of 3D designers and impact various industries that depend on 3D graphic design.", + "bbox": [ + 212, + 619, + 787, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graphic design is very challenging because even a small design goal requires performing a variety of different tasks. For instance, modeling of a game environment requires the 3D artist to cycle between performing modeling, material", + "bbox": [ + 212, + 755, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 For project website and code, please go to: https://ianhuang0630.github.io/BlenderAlchemyWeb/", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8d18be0552d5a1e9b290069f16b51083f6875aa4f199495c78b9900e7b893f3.jpg", + "image_caption": [ + "Fig. 1: Overview of BlenderAlchemy. Given an input Blender state and a user intention specified using either language or reference images, BlenderAlchemy edits the Blender state to satisfy that intention by iteratively refining a Blender python program that executes to produce the final Blender state. Our system additionally leverages text-to-image generation for visual imagination, a step that expands a text-only user intention to a concrete visual target to improve program refinement." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "design, texture painting, animation, lighting, and scene composition. Prior attempts usually focus on specific editing tasks, like material synthesis [8, 62]. While these approaches show promising performance in the tasks they are designed for, it is non-trivial to combine these task-specific methods to satisfy an intended design goal. An alternative is to leverage Large Language Models (LLMs) [22, 27, 45] to digest user intent and suggest design actions by proposing intelligent combination of existing task-specific tools [36] or predicting edits to programs step-by-step [49, 57]. While LLMs have excellent abilities to understand user intentions and suggest sequences of actions to satisfy them, applying LLMs to graphical design remains challenging largely because language cannot capture the visual consequences of actions performed in software like Blender.", + "bbox": [ + 212, + 435, + 787, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One promising alternative is to leverage vision language models (VLM), such as LLaVA [23], GPT-4V [28], Gemini [44], and DallE-3 [7]. These VLMs have shown to be highly capable of understanding detailed visual information [13,56, 58, 59] and generating compelling images [7]. We posit that these VLMs can be leveraged to complete different kinds of design tasks within the Blender design environment, like editing materials, geometry and lighting setups.", + "bbox": [ + 212, + 607, + 787, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we provide a proof-of-concept system using the vision foundation model GPT-4V to generate and edit programs that modify the state of a Blender workspace to satisfy a user intention. Specifically, we first initialize the state of a workspace within Blender. The Blender state is parameterized as a short Python program, and a base Blender file. The user will then input a text description and potentially a reference image to communicate the desirable design outcome. The system is tasked to edit the program so that, when executed on the base Blender file, the rendered image can satisfy the user's intention. Figure 1 provides an illustration of the problem setup.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/faf85e5061ad7daf08a18921d7ff18c60663be6d15cecfc1d1fdd6904cb83b90.jpg", + "image_caption": [ + "Fig. 2: Iterative visual program editing employs a edit generator $G$ and a state evaluator $V$ in each iteration to explore and prune different potential program edits, where $G$ generates plausible variants of an input program and $V$ picks between two programs based on the consequences they have to the Blender visual state and their alignment to the user intention. Each iteration of the refinement explores variations of the most promising program from the previous iteration. See Algorithm 1 for details." + ], + "image_footnote": [], + "bbox": [ + 220, + 142, + 782, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Naively applying VLMs to this editing setting gives rise to many failure cases, possibly due to the fact that out-of-the-box VLMs have a poor understanding of the visual consequences of Blender program edits. To counter this, we propose a visually-guided program search procedure that combines a vision-aware edit generator and a visual state evaluator to iteratively search for a suitable program edit (Figure 2). Inspired by the human design process, our system performs guided trial-and-error, capped by some computation budget. Within each iteration, the visual program generator will propose several possible edits on the current program. These edits will be applied and executed in Blender to produce a rendered image. These rendered images will be provided to visual evaluator, which will select the best renders via pairwise comparison by assessing which choice better satisfies the design goal specified by text and reference image. The program achieving the best render will replace the current program as a starting point for the visual program generator. This iterative search process, however, has very low success rate because of the sparsity of correct program edits in the vast program space. To improve the success rate, we further propose two techniques. First, when the proposed program of the new iteration does not contain a viable candidate, we revert to the best candidate of the prior program. This reversion mechanism makes sure the search procedure will not diverge when facing a batch of bad edit candidates. Second, to facilitate our visual evaluator and generator to better understand user intent, we leverage the \"visual imagination\" of text-to-image generative model to imagine a reference image. We show that our method is capable of accomplishing graphical design tasks within Blender, guided by user intention in the form of text and images. We demonstrate the effectiveness of our system on material, geometry and lighting design tasks, all parts of the 3D design process where artists spend a significant amount of time, ranging from 20 hrs to 4-6 workdays per model [2,3]. We show that our method can outperform prior works designed for similar problem settings, such as BlenderGPT [1]. In summary, the main contributions of this paper include:", + "bbox": [ + 212, + 358, + 787, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. We propose BlenderAlchemy, a system that's able to edit visual programs based on user input in the form of text or images.", + "bbox": [ + 222, + 809, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BlenderAlchemy", + "bbox": [ + 620, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. We identify key components that make the system work: a visual state evaluator, a visual edit generator, a search algorithm with an edit reversion mechanism, and a visual imagination module to facilitate the search.", + "3. We provide evidence showing that our system can outperform prior works in text-based procedural material editing, as well as its applicability to other design tasks including editing geometry and lighting configurations." + ], + "bbox": [ + 220, + 146, + 785, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 260, + 390, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-specific tools for Material Design. Large bodies of works have been dedicated to using learning-based approaches to generate materials. Prior works exploit 2D diffusion models to generate texture maps either in a zero-shot way [8, 21, 33, 50, 54] or through fine-tuning [37, 60]. Though these works open the possibility of generating and modifying textures of objects using natural language for 3D meshes, these works do not model the material properties in a way that allows such objects to be relit. Other methods directly predict the physical properties of the surface of a material through learning, using diffusion [46, 47] or learning good latent representations from data [10, 17, 20, 25, 39, 62]. However, for all of the work mentioned so far, their fundamentally image-based or latent-based representations make the output materials difficult to edit in existing 3D graphics pipelines. People have also explored combining learning-based approaches with symbolic representations of materials [34, 43]. These works often involve creating differentiable representations of the procedural material graph often used in 3D graphical design pipelines, and backpropagating gradients throughout the graph to produce an image that can match the target [16, 19, 38]. Other works like Infinigen [32] use rule-based procedural generation over a library of procedural materials. However, no prior works in this direction have demonstrated the ability to edit procedural material graphs using user intention specified by language [34], a task that we are particularly interested in. Though the aforementioned works excel at material design, they aren't generalizable to other 3D graphics design task settings. BlenderAlchemy, on the other hand, aims to produce a system that can perform various design tasks according to user intents. This usually requires combining different methods together in a non-trivial way.", + "bbox": [ + 212, + 277, + 787, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLM as general problem solvers. Large language models (LLMs) like GPT-4 [27,28], Llama [45], and Mistral [22] have in recent years demonstrated unprecedented results in a variety of problem settings, like robotics [4,12,18,24,41,52,61], program synthesis [26,35,40], and graphic design [21,50,55]. Other works have shown that by extending such models with an external process like Chain-of-Thought [49], Tree-of-Thought [57] or memory/skill database [11, 29, 30, 48], or by embedding such systems within environments where it can perceive and act [11, 29, 36, 48], a range of new problems that require iterative refinement can be solved. Their application to visual problem settings, however, has mostly been limited due to the nonexistent visual perception capabilities of the LLMs [11, 15, 48, 50, 53, 55]. While this could be sidestepped by fully condensing the visual state of the environment using text [29] or some symbolic representation [11,48], doing so for 3D graphic design works poorly. For instance, the task", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of editing a Blender material graph to create a desired material requires many trial-and-error cycles and an accurate understanding of the consequences certain design actions can have on the visual output. Recent works that apply LLMs to graphical design settings [11, 15] and ones that more specifically do so within Blender (like BlenderGPT [1], 3D-GPT [42] and L3GO [53]) do not use visual information to inform or refine their system's outputs, leading to unsatisfactory results. BlenderAlchemy borrows ideas from existing LLM literature and tries to address this issue by inputting visual perception into the system.", + "bbox": [ + 212, + 146, + 782, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision-Language Models. State-of-the-art vision language models, such as LLaVA [23], GPT-4V [28], and Gemini [44] have demonstrated impressive understanding of the visual world and its connections to language and semantics, enabling many computer vision tasks like scene understanding, visual question answering, and object detection to be one API call away [13,14,56,58,59]. Works such as [6,51] suggest that such models can also be used as a replacement for human evaluators for a lot of tasks, positioning them as tools for guiding planning and search by acting as flexible reward functions. BlenderAlchemy takes the first steps to apply VLMs to solve 3D graphic design tasks, a novel yet challenging application rather unexplored by existing works.", + "bbox": [ + 212, + 270, + 784, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 446, + 330, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The goal of our system is to perform edits within the Blender 3D graphic design environment through iteratively refining programs that define a sequence of edits in Blender. This requires us to (1) decompose the input initial Blender input into a combination of programs and a \"base\" Blender state (Section 3.1) and (2) develop a procedure to edit each program to produce the desired visual state within Blender to match a user intention (Sections 3.2).", + "bbox": [ + 212, + 479, + 782, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Representation of the Blender Visual State", + "text_level": 1, + "bbox": [ + 215, + 593, + 617, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The state of the initial Blender design environment can be decomposed into an \"base\" Blender state $S_{\\mathrm{base}}$ and a set of initial programs $\\{p_0^{(1)}, p_0^{(2)}, \\dots, p_0^{(k)}\\}$ that acts on state $S_{\\mathrm{base}}$ to produce the initial Blender environment through a dynamics function $F$ that transitions from one state to another based on a set of programmatic actions:", + "bbox": [ + 212, + 619, + 782, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathrm {i n i t}} = F \\left(\\left\\{p _ {0} ^ {(i)} \\right\\} _ {i = 1 \\ldots k}, S _ {\\mathrm {b a s e}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 708, + 609, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our problem setting, $F$ is the python code executor within the Blender environment that executes $\\{p_0^{(i)}\\}_{i = 1\\dots k}$ in sequence. We set each initial program $p_0^{(i)}$ to be a program that concerns a single part of the 3D graphical design workflow - for instance, $p_0^{(1)}$ is in charge of the material on one mesh within the scene, and $p_0^{(2)}$ is in charge of the lighting setup of the entire scene. The decomposition of $S_{\\mathrm{init}}$ into $S_{\\mathrm{base}}$ and $p_0^1\\dots p_0^k$ can be done using techniques like the", + "bbox": [ + 212, + 739, + 784, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BlenderAlchemy", + "bbox": [ + 620, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "“node transpler” from Infinigen [32], which converts entities within the Blender instance into lines of Python code that can recreate a node graph, like a material shader graph. We develop a suite of tools to do this in our own problem setting.", + "bbox": [ + 212, + 146, + 785, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Though it's possible for all edits to be encompassed in a single program instead of $k$ programs, this is limiting in practice - either because the VLM's output length isn't large enough for the code necessary or because the VLM has a low success rate, due to the program search space exploding in size. Although it is possible that future VLMs will substantially mitigate this problem, splitting the program into $k$ task-specific programs may still be desirable, given the possibility of querying $k$ task-specific fine-tuned/expert VLMs in parallel.", + "bbox": [ + 212, + 194, + 787, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Iterative Refinement of Individual Visual Programs", + "text_level": 1, + "bbox": [ + 214, + 333, + 686, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Suppose that to complete a task like material editing for a single object, it suffices to decompose the initial state into $S_{\\mathrm{base}}$ and a single script, $p_0$ - that is, $S_{\\mathrm{init}} = F(\\{p_0\\}, S_{\\mathrm{base}})$ . Then our goal is to discover some edited version of $p_0$ , called $p_1$ , such that $F(\\{p_1\\}, S_{\\mathrm{base}})$ produces a visual state better aligned with some user intention $I$ . Our system assumes that the user intention $I$ is provided in the form of language and/or image references, leveraging the visual understanding of the latest VLM models to understand user intention.", + "bbox": [ + 212, + 369, + 787, + 474 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Iterative Refinement of Visual Programs" + ], + "code_body": "1: procedure TOURNAMENT(State candidates $\\{S_1,S_2,\\dots S_k\\}$ , Visual state evaluator $V$ ,User intention $I$ \n2: if $k > 2$ then \n3: $w_{1}\\gets$ TOURNAMENT( $\\{S_1,S_2,\\ldots S_{k / 2}\\}$ $V,I)$ \n4: $w_{2}\\gets$ TOURNAMENT( $\\{S_{k / 2},S_{k + 1},\\ldots S_k\\}$ $V,I)$ \n5: else \n6: $w_{1}\\gets S_{1},w_{2}\\gets S_{2}$ \n7: end if \n8: return $V(w_{1},w_{2},I)$ \n9: end procedure \n10: procedure REFINE(Depth $d$ , Breadth $b$ , Intention $I$ , Edit Generator $G$ , State \nEvaluator $V$ , Base state $S_{\\mathrm{base}}$ , Initial program $p_0$ , Dynamics Function $F$ \n11: $S_0\\gets F(p_0,S_{\\mathrm{base}})$ $S_{\\mathrm{best}}\\gets S_0$ $p_{\\mathrm{best}}\\gets p_0$ \n12: for $i\\gets 1$ to $d$ do \n13: $\\mathcal{P}_i = \\mathcal{N}(p_{\\mathrm{best}})$ if $i$ is odd else $\\mathcal{P}$ $\\triangleright$ \"Tweak\" or \"Leap\" edits \n14: $p_i^1,p_i^2\\dots p_i^b\\gets G(p_{\\mathrm{best}},S_{\\mathrm{best}},I,b,\\mathcal{P}_i)$ $\\triangleright$ Generate $b$ options \n15: $S_i^1\\gets F(p_i^1,S_0),\\dots,S_i^b\\gets F(p_i^b,S_0)$ $\\triangleright$ Observe the visual states \n16: $S_i^*\\gets$ TOURNAMENT( $\\{S_i^1,S_i^2\\dots S_i^b\\} \\cup \\{S_{\\mathrm{best}}\\}$ $V,I)$ $\\triangleright$ Choose the best \n17: $S_{\\mathrm{best}}\\gets S_i^*,p_{\\mathrm{best}}\\gets p_i^*$ $\\triangleright$ Best visual state and programs so far \n18: end for \n19: return $S_{\\mathrm{best}},p_{\\mathrm{best}}$ \n20: end procedure", + "bbox": [ + 215, + 526, + 787, + 834 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To discover a good edit to $p_0$ , we introduce the procedure outlined in Algorithm 1, an iterative refinement loop that repeatedly uses a visual state evaluator $V$ to select among the hypotheses from an edit generator $G$ . A single \"agent\" for a certain task like procedural material design can be fully described by $(G, V)$ .", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inspired by works like [51], we propose a visual state evaluator $V(S_{1}, S_{2}, I)$ , which is tasked with returning whichever of the two visual states ( $S_{1}$ or $S_{2}$ ) better matches the user intention $I$ . This evaluator is applied recursively to choose the most suitable visual state candidate among $b$ visual state candidates by making $\\mathcal{O}(\\log(b))$ queries, as done in TOURNAMENT in Algorithm 1.", + "bbox": [ + 212, + 207, + 785, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Though it seems straightforward to ask the same VLM to edit the code in a single pass, this leads to many failure cases (see more in the Appendix). Due to the VLM's lack of baked-in understanding of the visual consequences of various programs within Blender, a multi-hypothesis and multi-step approach is more appropriate. Extending Tree-of-Thoughts [57] to the visual domain, $G(p, S, I, b, \\mathcal{P})$ is a module tasked with generating $b$ different variations of program $p$ , conditioned on the current visual state $S$ and user intention $I$ , constrained such that the output programs fall within some family of programs $\\mathcal{P}$ , which can be used to instill useful priors to the edit generator.", + "bbox": [ + 212, + 282, + 785, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Below we describe some additional system design decisions that ensured better alignment of the resultant edited program to the user intention, either by improving the stability of the procedure or by supplementing the visual understanding of VLMs. The effect of each is investigated in the Appendix through ablations.", + "bbox": [ + 212, + 417, + 785, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hypothesis Reversion. To improve the stability of the edit discovery process, we add the visual state of the program being edited at every timestep ( $S_{\\mathrm{best}}$ for $p_{\\mathrm{best}}$ ) as an additional candidate to the selection process, providing the option for the process to revert to an earlier version if the search at a single iteration was unsuccessful. Line 16 in Algorithm 1 shows this.", + "bbox": [ + 214, + 508, + 785, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tweak and Leap Edits. An important characteristic of visual programs is that continuous values hard-coded within the program can modify the output just as much as structural changes. This is in contrast to non-visual program synthesis tasks based on unit-tests of I/O specs, like [5, 9], where foundation models are mostly tasked to produce the right structure of the program with minimal hard-coded values. Given a single visual program $p \\in \\mathcal{P}$ , the space of visual outputs achievable through only tweaking the numerical values to function parameters and variable assignments can cover a wide range, depending on the fields available in the program. In Algorithm 1, we refer this space as the \"neighborhood\" of $p$ or \"tweak\" edits, $\\mathcal{N}(p)$ . Though this can result in a very small change to the program, this can lead to a large visual difference in the final output, and in a few edits can change the \"wrong\" program into one more aligned with the user intention. On the other hand, more drastic changes (or \"leap\" edits) may be needed to accomplish a task. Consider for example the task of changing a perfectly smooth material to have a noised level of roughness scattered sparsely across the material surface. This may require the programmatic addition of the", + "bbox": [ + 214, + 598, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BlenderAlchemy", + "bbox": [ + 620, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "relevant nodes (e.g. color ramps or noise texture nodes), and thus the edited program falls outside of $\\mathcal{N}(p)$ .", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Empirically, we find that the optimal edits are often a mix of tweak and leap edits. As such, our procedure cycles between restricting the edits of $G$ to two different sets: the neighborhood of $p_{\\mathrm{best}}$ , and the whole program space $\\mathcal{P}$ (Line 13 in Algorithm 1). In practice, such restrictions are softly enforced through in-context prompting of VLMs, and though their inputs encourage them to abide by these constraints, the model can still produce more drastic \"tweak\" edits or conservative \"leap\" edits as needed.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Imagination. In the case when the user intention is communicated purely textually, their intention may be difficult for the VLM to turn into successful edits. Prior works like [21, 50, 55] have made similar observations of abstract language for 3D scenes. Consider, for example, the prompt \"make me a material that resembles a celestial nebula\". To do this, the VLM must know what a celestial nebula looks like, and how it should change the parameters of the material Shader nodes of, say, a wooden material. We find that in such cases, it's hard for the VLM to directly go from abstract descriptions to low-level program edits that affect low-level properties of the Blender visual state.", + "bbox": [ + 212, + 301, + 787, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instead, we propose supplementing the text-to-program understanding of VLM's with the text-to-image understanding in state-of-the-art image generation models. Intermediary visual artifacts (images generated using the user intention) are created and used to guide the refinement process towards a more plausible program edit to match the desired outcome, as shown in Figure 1. The generated images act as image references in addition to the textual intention provided by the user. This constitutes a simple visual chain of thought [49] for visual program editing, which not only creates a reference image for $G$ and $V$ to guide their low-level visual comparisons (e.g. color schemes, material roughness... etc.), but also provides a user-interpretable intermediary step to confirm the desired goals behind an otherwise vague user intention.", + "bbox": [ + 212, + 438, + 787, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 630, + 375, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate BlenderAlchemy on editing procedural materials, geometry and lighting setups within Blender, three of the most tedious parts of 3D design.", + "bbox": [ + 212, + 664, + 785, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Procedural Material Editing", + "text_level": 1, + "bbox": [ + 214, + 720, + 496, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Procedural material editing has characteristics that make it difficult for the same reason as a lot of other visual program settings: (1) small edit distances of programs may result in very large visual differences, contributing to potential instabilities in the edit discovery process, (2) it naturally requires trial-and-error when human users are editing them as the desired magnitude of edits depends on the visual output and (3) language descriptions of edit intent typically do not", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "contain low-level information for the editor to know immediately which part of the program to change.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We demonstrate this on two different kinds of editing tasks: (1) turning the same initial material (a synthetic wooden material) into other materials described by a list of language descriptions and (2) editing many different initial materials to resemble the same target material described by an image input. For the starting materials, we use the synthetic materials from Infinigen [32].", + "bbox": [ + 212, + 176, + 784, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Text-based material-editing An attractive application of our system is in the modification of preexisting procedural materials using natural language descriptions that communicate user intent, a desirable but thus far undemonstrated capability of neurosymbolic methods [34]. We demonstrate our system's capabilities by asking it to edit a wooden material from Infinigen [32] into many other materials according to diverse language descriptions of target materials that are, importantly, not wood. In reality, this is a very challenging task, since this may require a wide range in the size of edits even if the language describing desired target material may be very similar.", + "bbox": [ + 212, + 273, + 785, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 3 shows examples of edits of the same starter wood material using different language descriptions, and Figure 4 demonstrates the intermediary steps of the problem solving process for a single instance of the problem. Our system is composed of an edit generator that generates 8 hypotheses per iteration, for 4 iterations ( $d = 4$ , $b = 8$ ), cycling between tweak and leap edits. It uses GPT-4V for edit generation and state evaluation, and DallE-3 for visual imagination.", + "bbox": [ + 212, + 410, + 785, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare against BlenderGPT, the most recent open-sourced Blender AI agent that use GPT-4 to execute actions within the Blender environment through the Python API. We provide the same target material text prompt to BlenderGPT, as well as the starter code for the initial wood material for reference. We compare the CLIP similarity of their output material to the input text description against our system. BlenderGPT reasons only about how to edit the program using the input text description, doing so in a single pass without state evaluation or multi-hypothesis edit generation. To match the number of edit generator queries we make, we run their method a maximum of 32 times, using the first successful example as its final output. Everything else remains the same, including the starter material program, text description, base Blender state, and lighting setup.", + "bbox": [ + 212, + 501, + 785, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We find that qualitatively, BlenderGPT produces much shallower and more simplistic edits of the input material, resulting in low-quality output materials and poor alignment with the user intention. Examples can be seen in Figure 5. For instance, observe that for the \"digital camouflage\" example, BlenderAlchemy is able to produce the \"sharper angles\" that the original description requests (See Figure 3) whereas BlenderGPT produces the right colors but fails to create the sharp, digital look. For \"metallic swirl\" example, our system's visual state selection process would weed out insufficiently swirly examples such as the one", + "bbox": [ + 212, + 681, + 785, + 803 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BlenderAlchemy", + "bbox": [ + 620, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "$^{2}$ At the time of publishing, works like 3DGPT [42], L3GO [53] have not yet open-sourced their code.", + "bbox": [ + 217, + 809, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1b50a424840ecc234239ee56d36f7f5b8b0fc407ae700b8ca4e69a376061f120.jpg", + "image_caption": [ + "Fig. 3: Text-based Material Editing Results. The step-by-step edits of a 4x8 version of BlenderAlchemy to the same wooden material, given the text description on the left as the input user intention." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 785, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "given by BlenderGPT, enabling our method to produce a material closer to the prompt.", + "bbox": [ + 212, + 590, + 785, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1 demonstrates the comparison in terms of the average ViT-B/32 and ViT-L/14 CLIP similarity scores [31] with respect to the language description. Our system's ability to iteratively refine the edits based on multiple guesses at each step gives it the ability to make more substantive edits over the course of the process. Moreover, visual grounding provided both by the visual state evaluator as well as the output of the visual imagination stage guides the program search procedure to better align with the description. For ablations, see the Appendix.", + "bbox": [ + 212, + 625, + 787, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We additionally conducted a user preference study to compare BlenderAlchemy's performance with two material generation baselines that use diffusion: TEXTure [33] and Paint3D [60]. We collect 592 Mechanical Turk comparisons between BlenderAlchemy and the baselines from 24 Turkers on materials created using 32 different text prompts. Users must choose the material that best matches the text-prompt. The results show that BlenderAlchemy is preferred to Paint3D $73\\%$ of the time. Users picked BlenderAlchemy over TEXTure $56\\%$", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d49d38b56b44a2ac65002828b7cf48f21a8d1455cde24b5c02e63535e0b072dd.jpg", + "image_caption": [ + "Fig. 4: The edit discovery process of turning a wooden material into \"marbled granite\". Each column shows the hypotheses generated by $G$ , with the most promising candidates chosen by $V$ indicated by the highlights. Note that iteration 3 proved to be unfruitful according to $V$ , and the method reverts to the best candidate from iteration 2, before moving onto iteration 4." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 311 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/952e25b1509dd0f2f4354925c57068ec8b4f48cf090cbe7f38d2c1b270da82a0.jpg", + "table_caption": [ + "Table 1: CLIP scores of BlenderAlchemy vs. BlenderGPT for the text-based material editing task. We find that a version of our system that has no visual components (-Vision) still outperforms BlenderGPT. By adding vision to the state evaluator alone (-Vision G) and not the edit generator, a further improvement is observed." + ], + "table_footnote": [], + "table_body": "
MetricBlenderGPT-Vision-Vision GOurs
ViT-B/32 (↑)25.225.727.828.2
ViT-L/14 (↑)21.121.823.424.0
", + "bbox": [ + 299, + 468, + 699, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "of the time. We found that in $71\\%$ of comparisons, human users also prefer BlenderAlchemy outputs selected by our visual evaluator. This suggests that our visual evaluator makes decisions that align with users' preferences.", + "bbox": [ + 212, + 546, + 787, + 592 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Image-based material-editing Given an image of a desired material, the task is to convert the code of the starter material into a material that contains many of the visual attributes of the input image, akin to doing a kind of style transfer for procedural materials.", + "bbox": [ + 212, + 607, + 787, + 666 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "At each step, the edit generator is first asked to textually enumerate a list of obvious visual differences between the current material and the target, then asked to locate lines within the code that may be responsible for these visual differences (e.g. \"the target material looks more rough\" $\\rightarrow$ \"line 23 sets a roughness value, which we should try to increase\") before finally suggested an edited version of the program. As such, low-level visual differences (e.g. color discrepancies) are semantically compressed first into language descriptions (e.g. \"the target is more red\"), before being fed into the editing process, resulting in the behavior that our system produces variations of the input material that resembles the target image along many attributes, even if its outputs don't perfectly match the target image (See Figure 6). Our system is the same as for text-based material editing, but without the need for visual imagination.", + "bbox": [ + 212, + 667, + 787, + 849 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BlenderAlchemy", + "bbox": [ + 620, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/06d62e1ee7f5e2a1764eae8e3ec639b9ba3f2c1bc624fba2e0c299ce30f1a657.jpg", + "image_caption": [ + "Fig.5: Comparisons between our method and BlenderGPT for the text-based material editing task setting. Note how our materials align better with the original language prompts (See Figure 3 for the original prompts). The input material being edited is the same wooden material." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 339 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5beff7723f7e944e01a00ede3bb7c4fd8e30f653ccd477a57f7d854851964605.jpg", + "image_caption": [ + "Fig. 6: Material editing based on image inputs. Our edit intention is described by the target image shown on the right. 5 different Infinigen [32] initial materials are shown here, and the final edit. Note how in each case, certain attributes of the target material (metallicness, color, texture) are transferred." + ], + "image_footnote": [], + "bbox": [ + 217, + 402, + 643, + 551 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bf51817fbc787d52178671d2a1a31828fec0ccf5a815d83018407b31c6233c82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 422, + 774, + 534 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Geometry Editing", + "text_level": 1, + "bbox": [ + 215, + 625, + 413, + 641 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "BlenderAlchemy can also be used to manipulate geometry. We show qualitative examples of our system controlling geometry by programmatically 1) interpolating between preset blend shapes, 2) editing of geometry node graphs, and 3) the precise placement of objects within scenes. The blend weights, geometry node graphs and object placements can all be represented programmatically.", + "bbox": [ + 212, + 657, + 787, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As shown in Figure 7, given an input shape and pre-defined blend shapes, BlenderAlchemy can modify the shape by finding the appropriate blend weights between different shapes to match language, which is non-trivial for human users given the number of blend shapes to interpolate for complex characters and shapes. For instance, the facial expression editing example requires coordinated interpolation between 19 different blend shapes, each which controls a specific kind of motion in one part of the face (e.g. an left eye-brow raise). To aid it in", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d09363b83edf21f76638344c36f2a32c34d70614a5593b25a055cfca53295b2c.jpg", + "image_caption": [ + "Fig. 7: BlenderAlchemy editing geometry using blend shapes. Edits are made to match a description or a script line. Input shapes from BlenderKit." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 354 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/83d7ae0544c2666d144054ee01c7eab1da0865539f8420f1a95765ffc29cb5c1.jpg", + "image_caption": [ + "Fig. 8: BlenderAlchemy editing (a) the placement of assets within a living room scene and (b) the procedural geometry nodes of roses, to different text prompts." + ], + "image_footnote": [], + "bbox": [ + 217, + 393, + 787, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "this task, our task description to BlenderAlchemy includes language labels for each blend shape.", + "bbox": [ + 212, + 590, + 784, + 619 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Figure 8(a), BlenderAlchemy is presented with a scene layout editing task, where assets are initially dropped out of camera view and must be iteratively moved to fit the scene of a living room and the requirement that the couch is \"facing the window\", through calling a set of object placement helper functions. Figure 8(b) shows examples of BlenderAlchemy editing the procedural geometry of a set of roses to match a nuanced instruction that influences the number and angle of pedals, as well as the density of leaves. The full geometry node graph that BlenderAlchemy edits contains 60 geometry nodes.", + "bbox": [ + 212, + 622, + 787, + 742 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Lighting Setup Editing", + "text_level": 1, + "bbox": [ + 215, + 763, + 450, + 780 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We show that BlenderAlchemy can be used to adjust the lighting of scenes according to language instructions as well. Figure 9 shows this being done for an input product visualization designed by a Blender artist.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BlenderAlchemy", + "bbox": [ + 620, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2e8abb6e51680947847c7737b985d5dcae14705620d7c4c37890adb712bdb510.jpg", + "image_caption": [ + "Fig. 9: Optimizing the lighting of the scene setup to the text-based user intention. We base the initial Blender state input based on a product visualization downloaded from BlenderKit." + ], + "image_footnote": [], + "bbox": [ + 220, + 164, + 357, + 281 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e7a06a2f18c2f37165809061063a128b9c87b057265ca3f698dab805a9e39b31.jpg", + "image_caption": [ + "\"Hand lotion under disco lights, like at a nightclub.\"" + ], + "image_footnote": [], + "bbox": [ + 359, + 164, + 496, + 281 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ef951f71b27d874796cd4f9aec7f8023f312fe0b1828564b9a75e6b6a4124f52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 166, + 640, + 280 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c6691351d34c882058d3000c6a7ba4712adf001b9a94ad303a0e61c70f6d1b93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 167, + 782, + 280 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9af7aff3497f4411e76040759dde004da5dcfd8da8c8fdebb550d9360f1180ee.jpg", + "image_caption": [ + "Fig. 10: Optimizing the lighting and material iteratively within a product visualization scene, to satisfy the text-based user intention. Note the dimming of the environment lighting for the nighttime lighting, and the glowing-hot material the editing procedure has produced. We base the initial Blender state input based on a product visualization downloaded from BlenderKit." + ], + "image_footnote": [], + "bbox": [ + 220, + 347, + 357, + 465 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/28d334738c60eaebb3632557884ce3090d89ac71273dd029a0f2d6f49ad127ad.jpg", + "image_caption": [ + "\"Hand cream sitting on a hot comet outside in the night\"" + ], + "image_footnote": [], + "bbox": [ + 357, + 348, + 495, + 465 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/46a9fa8c4012ef34af8475d1b1deef355c382ecfe3c9f10a5e19207109cb2c1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 348, + 612, + 465 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/da92e80f5afdb9d9f48e3e88e4c8d9d6175080e297b59f5657d404b261c1fd8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 349, + 781, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As mentioned in Section 3.1, we can consider iteratively optimizing two separate programs, one controlling the lighting of the whole scene and another controlling the material of an object within the scene. That is, $S_{\\mathrm{init}} = F(\\{p_0^{(L)}, p_0^{(M)}\\}, S_{\\mathrm{base}})$ for initial lighting program $p_0^{(L)}$ and material program $p_0^{(M)}$ . Figure 10 shows an example of this, where two separate pairs of edit generators and state evaluators, $(G_M, V_M)$ and $(G_L, V_L)$ , are used to achieve an edit to the scene that aligns with the user intention. The full algorithm is shown in the Appendix.", + "bbox": [ + 212, + 551, + 787, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion & Discussion", + "text_level": 1, + "bbox": [ + 215, + 685, + 490, + 700 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce BlenderAlchemy, a system that performs edits within the Blender 3D design environment by leveraging vision-language models to iteratively refining a program to be more aligned with the user intention, by using visual information to both explore and prune possibilities within the program space. We equip our system with visual imagination by providing it access to text-to-image models, a tool it uses to guide itself towards program edits that better align with user intentions. We've demonstrated BlenderAlchemy on editing materials, geometry and lighting, and hope that future works will extend this to other workflows as well.", + "bbox": [ + 212, + 710, + 787, + 845 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We acknowledge the support of ARL grant W911NF-21-2-0104 and a Vannevar Bush Faculty Fellowship. We'd additionally like to thank Maneesh Agrawala for general discussions, and Purvi Goel, Mika Uy, Vishnu Sarukkai, Fan-yun Sun and Sharon Lee for feedback on paper revisions.", + "bbox": [ + 212, + 178, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 323, + 279 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Blendergt, https://github.com/gd3kr/BlenderGPT", + "2. How long does it take to create a 3d model?, https://3d-ace.com/blog/how-long-does-it-take-to-create-a-3d-model/", + "3. How long does it take to make a 3d model?, https://pixune.com/blog/how-long-does-it-take-to-create-a-3d-model/", + "4. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al.: Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 (2022)", + "5. Austin, J., Odena, A., Nye, M., Bosma, M., Michalewski, H., Dohan, D., Jiang, E., Cai, C., Terry, M., Le, Q., et al.: Program synthesis with large language models. arXiv preprint arXiv:2108.07732 (2021)", + "6. Baumli, K., Baveja, S., Behbahani, F., Chan, H., Comanici, G., Flennerhag, S., Gazeau, M., Holsheimer, K., Horgan, D., Laskin, M., et al.: Vision-language models as a source of rewards. arXiv preprint arXiv:2312.09187 (2023)", + "7. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opena.com/papers/dall-e-3.pdf 2(3), 8 (2023)", + "8. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023)", + "9. Chen, M., Tworek, J., Jun, H., Yuan, Q., de Oliveira Pinto, H.P., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F.P., Cummings, D., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W.H., Nichol, A., Paino, A., Tezak, N., Tang, J., Babuschkin, I., Balaji, S., Jain, S., Saunders, W., Hesse, C., Carr, A.N., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., Zaremba, W.: Evaluating large language models trained on code (2021)", + "10. Chen, Y., Chen, R., Lei, J., Zhang, Y., Jia, K.: Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems 35, 30923-30936 (2022)", + "11. De La Torre, F., Fang, C.M., Huang, H., Banburski-Fahey, A., Fernandez, J.A., Lanier, J.: Llmr: Real-time prompting of interactive worlds using large language models. arXiv preprint arXiv:2309.12276 (2023)", + "12. Firoozzi, R., Tucker, J., Tian, S., Majumdar, A., Sun, J., Liu, W., Zhu, Y., Song, S., Kapoor, A., Hausman, K., et al.: Foundation models in robotics: Applications, challenges, and the future. arXiv preprint arXiv:2312.07843 (2023)" + ], + "bbox": [ + 218, + 296, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BlenderAlchemy", + "bbox": [ + 620, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., Wu, Y., Ji, R.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)", + "14. Fu, C., Zhang, R., Wang, Z., Huang, Y., Zhang, Z., Qiu, L., Ye, G., Shen, Y., Zhang, M., Chen, P., Zhao, S., Lin, S., Jiang, D., Yin, D., Gao, P., Li, K., Li, H., Sun, X.: A challenger to gpt-4v? early explorations of gemini in visual expertise. arXiv preprint arXiv:2312.12436 (2023)", + "15. Goel, P., Wang, K.C., Liu, C.K., Fatahalian, K.: Iterative motion editing with natural language. arXiv preprint arXiv:2312.11538 (2023)", + "16. Guerrero, P., Hasan, M., Sunkavalli, K., Meech, R., Boubekeur, T., Mitra, N.J.: Matformer: A generative model for procedural materials. arXiv preprint arXiv:2207.01044 (2022)", + "17. Henzler, P., Deschaintre, V., Mitra, N.J., Ritschel, T.: Generative modelling of brdf textures from flash images. arXiv preprint arXiv:2102.11861 (2021)", + "18. Hu, Y., Xie, Q., Jain, V., Francis, J., Patrikar, J., Keetha, N., Kim, S., Xie, Y., Zhang, T., Zhao, Z., et al.: Toward general-purpose robots via foundation models: A survey and meta-analysis. arXiv preprint arXiv:2312.08782 (2023)", + "19. Hu, Y., Guerrero, P., Hasan, M., Rushmeier, H., Deschaintre, V.: Node graph optimization using differentiable proxies. In: ACM SIGGRAPH 2022 conference proceedings. pp. 1-9 (2022)", + "20. Hu, Y., He, C., Deschaintre, V., Dorsey, J., Rushmeier, H.: An inverse procedural modeling pipeline for svbrdf maps. ACM Transactions on Graphics (TOG) 41(2), 1-17 (2022)", + "21. Huang, I., Krishna, V., Atekha, O., Guibas, L.: Aladdin: Zero-shot hallucination of stylized 3d assets from abstract scene descriptions. arXiv preprint arXiv:2306.06212 (2023)", + "22. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)", + "23. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems 36 (2024)", + "24. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., Zeng, A.: Code as policies: Language model programs for embodied control. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9493-9500. IEEE (2023)", + "25. Liu, J., Gan, Y., Dong, J., Qi, L., Sun, X., Jian, M., Wang, L., Yu, H.: Perceptiondriven procedural texture generation from examples. Neurocomputing 291, 21-34 (2018)", + "26. Olausson, T.X., Inala, J.P., Wang, C., Gao, J., Solar-Lezama, A.: Is self-repair a silver bullet for code generation? In: The Twelfth International Conference on Learning Representations (2023)", + "27. OpenAI: Gpt-4 system card. OpenAI (2023), https://cdn.openai.com/papers/gpt-4-system-card.pdf", + "28. OpenAI: Gpt-4v(ison) system card. OpenAI (2023), https://apisemantic scholar.org/CorpusID:263218031", + "29. Park, J.S., O'Brien, J., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. pp. 1-22 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Patil, S.G., Zhang, T., Wang, X., Gonzalez, J.E.: Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334 (2023)", + "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "32. Raistrick, A., Lipson, L., Ma, Z., Mei, L., Wang, M., Zuo, Y., Kayan, K., Wen, H., Han, B., Wang, Y., et al.: Infinite photorealistic worlds using procedural generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12630-12641 (2023)", + "33. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721 (2023)", + "34. Ritchie, D., Guerrero, P., Jones, R.K., Mitra, N.J., Schulz, A., Willis, K.D., Wu, J.: Neurosymbolic models for computer graphics. In: Computer Graphics Forum. vol. 42, pp. 545-568. Wiley Online Library (2023)", + "35. Romera-Paredes, B., Barekatain, M., Novikov, A., Balog, M., Kumar, M.P., Dupont, E., Ruiz, F.J., Ellenberg, J.S., Wang, P., Fawzi, O., et al.: Mathematical discoveries from program search with large language models. Nature 625(7995), 468-475 (2024)", + "36. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Hambro, E., Zettle-moyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems 36 (2024)", + "37. Sharma, P., Jampani, V., Li, Y., Jia, X., Lagun, D., Durand, F., Freeman, W.T., Matthews, M.: Alchemist: Parametric control of material properties with diffusion models. arXiv preprint arXiv:2312.02970 (2023)", + "38. Shi, L., Li, B., Hasan, M., Sunkavalli, K., Boubekeur, T., Mech, R., Matusik, W.: Match: Differentiable material graphs for procedural material capture. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)", + "39. Shimizu, E., Fisher, M., Paris, S., McCann, J., Fatahalian, K.: Design adjectives: A framework for interactive model-guided exploration of parameterized design spaces. In: Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. pp. 261-278 (2020)", + "40. Shinn, N., Cassano, F., Labash, B., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Language agents with verbal reinforcement learning.(2023). arXiv preprint cs.AI/2303.11366 (2023)", + "41. Singh, I., Blukis, V., Mousavian, A., Goyal, A., Xu, D., Tremblay, J., Fox, D., Thomason, J., Garg, A.: Progprompt: Generating situated robot task plans using large language models. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11523-11530. IEEE (2023)", + "42. Sun, C., Han, J., Deng, W., Wang, X., Qin, Z., Gould, S.: 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945 (2023)", + "43. Tchapmi, L.P., Ray, T., Tchapmi, M., Shen, B., Martin-Martin, R., Savarese, S.: Generating procedural 3d materials from images using neural networks. In: 2022 4th International Conference on Image, Video and Signal Processing. pp. 32-40 (2022)", + "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BlenderAlchemy", + "bbox": [ + 620, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "46. Vecchio, G., Martin, R., Roullier, A., Kaiser, A., Rouffet, R., Deschaintre, V., Boubekeur, T.: Controlmat: A controlled generative approach to material capture. arXiv preprint arXiv:2309.01700 (2023)", + "47. Vecchio, G., Sortino, R., Palazzo, S., Spampinato, C.: Matfuse: Controllable material generation with diffusion models. arXiv preprint arXiv:2308.11408 (2023)", + "48. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)", + "49. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)", + "50. Wen, Z., Liu, Z., Sridhar, S., Fu, R.: Anyhome: Open-vocabulary generation of structured and textured 3d homes. arXiv preprint arXiv:2312.06644 (2023)", + "51. Wu, T., Yang, G., Li, Z., Zhang, K., Liu, Z., Guibas, L., Lin, D., Wetzstein, G.: Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. arXiv preprint arXiv:2401.04092 (2024)", + "52. Xiao, X., Liu, J., Wang, Z., Zhou, Y., Qi, Y., Cheng, Q., He, B., Jiang, S.: Robot learning in the era of foundation models: A survey. arXiv preprint arXiv:2311.14379 (2023)", + "53. Yamada, Y., Chandu, K., Lin, Y., Hessel, J., Yildirim, I., Choi, Y.: L3go: Language agents with chain-of-3d-thoughts for generating unconventional objects. arXiv preprint arXiv:2402.09052 (2024)", + "54. Yang, H., Chen, Y., Pan, Y., Yao, T., Chen, Z., Mei, T.: 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 6860-6868 (2023)", + "55. Yang, Y., Sun, F.Y., Weihs, L., VanderBilt, E., Herrasti, A., Han, W., Wu, J., Haber, N., Krishna, R., Liu, L., et al.: Holodeck: Language guided generation of 3d embodied ai environments. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024). vol. 30, pp. 20-25. IEEE/CVF (2024)", + "56. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)", + "57. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems 36 (2024)", + "58. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023)", + "59. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2310.16045 (2023)", + "60. Zeng, X.: Paint3d: Paint anything 3d with lighting-less texture diffusion models. arXiv preprint arXiv:2312.13913 (2023)", + "61. Zhou, H., Yao, X., Meng, Y., Sun, S., BIng, Z., Huang, K., Knoll, A.: Language-conditioned learning for robotic manipulation: A survey. arXiv preprint arXiv:2312.10807 (2023)", + "62. Zsolnai-Fehér, K., Wonka, P., Wimmer, M.: Gaussian material synthesis. arXiv preprint arXiv:1804.08369 (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 830 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/19f5daea-7c9c-46d5-b550-8a14478e364b_model.json b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/19f5daea-7c9c-46d5-b550-8a14478e364b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..35811d597a68e28a401a1d02700ab8adcba0e06e --- /dev/null +++ b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/19f5daea-7c9c-46d5-b550-8a14478e364b_model.json @@ -0,0 +1,2282 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.254, + 0.141, + 0.753, + 0.187 + ], + "angle": 0, + "content": "BlenderAlchemy: Editing 3D Graphics with Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.213, + 0.704, + 0.228 + ], + "angle": 0, + "content": "Ian Huang, Guandao Yang, and Leonidas Guibas" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.24, + 0.568, + 0.255 + ], + "angle": 0, + "content": "Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.295, + 0.74, + 0.558 + ], + "angle": 0, + "content": "Abstract. Graphics design is important for various applications, including movie production and game design. To create a high-quality scene, designers usually need to spend hours in software like Blender, in which they might need to interleave and repeat operations, such as connecting material nodes, hundreds of times. Moreover, slightly different design goals may require completely different sequences, making automation difficult. In this paper, we propose a system that leverages Vision-Language Models (VLMs), like GPT-4V, to intelligently search the design action space to arrive at an answer that can satisfy a user's intent. Specifically, we design a vision-based edit generator and state evaluator to work together to find the correct sequence of actions to achieve the goal. Inspired by the role of visual imagination in the human design process, we supplement the visual reasoning capabilities of VLMs with \"imagined\" reference images from image-generation models, providing visual grounding of abstract language descriptions. In this paper, we provide empirical evidence suggesting our system can produce simple but tedious Blender editing sequences for tasks such as editing procedural materials and geometry from text and/or reference images, as well as adjusting lighting configurations for product renderings in complex scenes." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.587, + 0.377, + 0.602 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.756 + ], + "angle": 0, + "content": "To produce the compelling graphics content we see in movies or video games, 3D artists usually need to spend hours in software like Blender to find appropriate surface materials, object placements, and lighting arrangements. These operations require the artist to create a mental picture of the target, experiment with different parameters, and visually examine whether their edits get closer to the end goal. One can imagine automating these processes by converting language or visual descriptions of user intent into edits that achieve a design goal. Such a system can improve the productivity of millions of 3D designers and impact various industries that depend on 3D graphic design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.802 + ], + "angle": 0, + "content": "Graphic design is very challenging because even a small design goal requires performing a variety of different tasks. For instance, modeling of a game environment requires the 3D artist to cycle between performing modeling, material" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "1 For project website and code, please go to: https://ianhuang0630.github.io/BlenderAlchemyWeb/" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Fig. 1: Overview of BlenderAlchemy. Given an input Blender state and a user intention specified using either language or reference images, BlenderAlchemy edits the Blender state to satisfy that intention by iteratively refining a Blender python program that executes to produce the final Blender state. Our system additionally leverages text-to-image generation for visual imagination, a step that expands a text-only user intention to a concrete visual target to improve program refinement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.603 + ], + "angle": 0, + "content": "design, texture painting, animation, lighting, and scene composition. Prior attempts usually focus on specific editing tasks, like material synthesis [8, 62]. While these approaches show promising performance in the tasks they are designed for, it is non-trivial to combine these task-specific methods to satisfy an intended design goal. An alternative is to leverage Large Language Models (LLMs) [22, 27, 45] to digest user intent and suggest design actions by proposing intelligent combination of existing task-specific tools [36] or predicting edits to programs step-by-step [49, 57]. While LLMs have excellent abilities to understand user intentions and suggest sequences of actions to satisfy them, applying LLMs to graphical design remains challenging largely because language cannot capture the visual consequences of actions performed in software like Blender." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.7 + ], + "angle": 0, + "content": "One promising alternative is to leverage vision language models (VLM), such as LLaVA [23], GPT-4V [28], Gemini [44], and DallE-3 [7]. These VLMs have shown to be highly capable of understanding detailed visual information [13,56, 58, 59] and generating compelling images [7]. We posit that these VLMs can be leveraged to complete different kinds of design tasks within the Blender design environment, like editing materials, geometry and lighting setups." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we provide a proof-of-concept system using the vision foundation model GPT-4V to generate and edit programs that modify the state of a Blender workspace to satisfy a user intention. Specifically, we first initialize the state of a workspace within Blender. The Blender state is parameterized as a short Python program, and a base Blender file. The user will then input a text description and potentially a reference image to communicate the desirable design outcome. The system is tasked to edit the program so that, when executed on the base Blender file, the rendered image can satisfy the user's intention. Figure 1 provides an illustration of the problem setup." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlenderAlchemy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.143, + 0.784, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 2: Iterative visual program editing employs a edit generator \\( G \\) and a state evaluator \\( V \\) in each iteration to explore and prune different potential program edits, where \\( G \\) generates plausible variants of an input program and \\( V \\) picks between two programs based on the consequences they have to the Blender visual state and their alignment to the user intention. Each iteration of the refinement explores variations of the most promising program from the previous iteration. See Algorithm 1 for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.789, + 0.798 + ], + "angle": 0, + "content": "Naively applying VLMs to this editing setting gives rise to many failure cases, possibly due to the fact that out-of-the-box VLMs have a poor understanding of the visual consequences of Blender program edits. To counter this, we propose a visually-guided program search procedure that combines a vision-aware edit generator and a visual state evaluator to iteratively search for a suitable program edit (Figure 2). Inspired by the human design process, our system performs guided trial-and-error, capped by some computation budget. Within each iteration, the visual program generator will propose several possible edits on the current program. These edits will be applied and executed in Blender to produce a rendered image. These rendered images will be provided to visual evaluator, which will select the best renders via pairwise comparison by assessing which choice better satisfies the design goal specified by text and reference image. The program achieving the best render will replace the current program as a starting point for the visual program generator. This iterative search process, however, has very low success rate because of the sparsity of correct program edits in the vast program space. To improve the success rate, we further propose two techniques. First, when the proposed program of the new iteration does not contain a viable candidate, we revert to the best candidate of the prior program. This reversion mechanism makes sure the search procedure will not diverge when facing a batch of bad edit candidates. Second, to facilitate our visual evaluator and generator to better understand user intent, we leverage the \"visual imagination\" of text-to-image generative model to imagine a reference image. We show that our method is capable of accomplishing graphical design tasks within Blender, guided by user intention in the form of text and images. We demonstrate the effectiveness of our system on material, geometry and lighting design tasks, all parts of the 3D design process where artists spend a significant amount of time, ranging from 20 hrs to 4-6 workdays per model [2,3]. We show that our method can outperform prior works designed for similar problem settings, such as BlenderGPT [1]. In summary, the main contributions of this paper include:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "1. We propose BlenderAlchemy, a system that's able to edit visual programs based on user input in the form of text or images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.147, + 0.786, + 0.191 + ], + "angle": 0, + "content": "2. We identify key components that make the system work: a visual state evaluator, a visual edit generator, a search algorithm with an edit reversion mechanism, and a visual imagination module to facilitate the search." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.192, + 0.786, + 0.238 + ], + "angle": 0, + "content": "3. We provide evidence showing that our system can outperform prior works in text-based procedural material editing, as well as its applicability to other design tasks including editing geometry and lighting configurations." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.147, + 0.786, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.261, + 0.392, + 0.276 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.789, + 0.643 + ], + "angle": 0, + "content": "Task-specific tools for Material Design. Large bodies of works have been dedicated to using learning-based approaches to generate materials. Prior works exploit 2D diffusion models to generate texture maps either in a zero-shot way [8, 21, 33, 50, 54] or through fine-tuning [37, 60]. Though these works open the possibility of generating and modifying textures of objects using natural language for 3D meshes, these works do not model the material properties in a way that allows such objects to be relit. Other methods directly predict the physical properties of the surface of a material through learning, using diffusion [46, 47] or learning good latent representations from data [10, 17, 20, 25, 39, 62]. However, for all of the work mentioned so far, their fundamentally image-based or latent-based representations make the output materials difficult to edit in existing 3D graphics pipelines. People have also explored combining learning-based approaches with symbolic representations of materials [34, 43]. These works often involve creating differentiable representations of the procedural material graph often used in 3D graphical design pipelines, and backpropagating gradients throughout the graph to produce an image that can match the target [16, 19, 38]. Other works like Infinigen [32] use rule-based procedural generation over a library of procedural materials. However, no prior works in this direction have demonstrated the ability to edit procedural material graphs using user intention specified by language [34], a task that we are particularly interested in. Though the aforementioned works excel at material design, they aren't generalizable to other 3D graphics design task settings. BlenderAlchemy, on the other hand, aims to produce a system that can perform various design tasks according to user intents. This usually requires combining different methods together in a non-trivial way." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "LLM as general problem solvers. Large language models (LLMs) like GPT-4 [27,28], Llama [45], and Mistral [22] have in recent years demonstrated unprecedented results in a variety of problem settings, like robotics [4,12,18,24,41,52,61], program synthesis [26,35,40], and graphic design [21,50,55]. Other works have shown that by extending such models with an external process like Chain-of-Thought [49], Tree-of-Thought [57] or memory/skill database [11, 29, 30, 48], or by embedding such systems within environments where it can perceive and act [11, 29, 36, 48], a range of new problems that require iterative refinement can be solved. Their application to visual problem settings, however, has mostly been limited due to the nonexistent visual perception capabilities of the LLMs [11, 15, 48, 50, 53, 55]. While this could be sidestepped by fully condensing the visual state of the environment using text [29] or some symbolic representation [11,48], doing so for 3D graphic design works poorly. For instance, the task" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "BlenderAlchemy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.268 + ], + "angle": 0, + "content": "of editing a Blender material graph to create a desired material requires many trial-and-error cycles and an accurate understanding of the consequences certain design actions can have on the visual output. Recent works that apply LLMs to graphical design settings [11, 15] and ones that more specifically do so within Blender (like BlenderGPT [1], 3D-GPT [42] and L3GO [53]) do not use visual information to inform or refine their system's outputs, leading to unsatisfactory results. BlenderAlchemy borrows ideas from existing LLM literature and tries to address this issue by inputting visual perception into the system." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.785, + 0.422 + ], + "angle": 0, + "content": "Vision-Language Models. State-of-the-art vision language models, such as LLaVA [23], GPT-4V [28], and Gemini [44] have demonstrated impressive understanding of the visual world and its connections to language and semantics, enabling many computer vision tasks like scene understanding, visual question answering, and object detection to be one API call away [13,14,56,58,59]. Works such as [6,51] suggest that such models can also be used as a replacement for human evaluators for a lot of tasks, positioning them as tools for guiding planning and search by acting as flexible reward functions. BlenderAlchemy takes the first steps to apply VLMs to solve 3D graphic design tasks, a novel yet challenging application rather unexplored by existing works." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.447, + 0.331, + 0.463 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.784, + 0.57 + ], + "angle": 0, + "content": "The goal of our system is to perform edits within the Blender 3D graphic design environment through iteratively refining programs that define a sequence of edits in Blender. This requires us to (1) decompose the input initial Blender input into a combination of programs and a \"base\" Blender state (Section 3.1) and (2) develop a procedure to edit each program to produce the desired visual state within Blender to match a user intention (Sections 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.618, + 0.609 + ], + "angle": 0, + "content": "3.1 Representation of the Blender Visual State" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.784, + 0.698 + ], + "angle": 0, + "content": "The state of the initial Blender design environment can be decomposed into an \"base\" Blender state \\( S_{\\mathrm{base}} \\) and a set of initial programs \\( \\{p_0^{(1)}, p_0^{(2)}, \\dots, p_0^{(k)}\\} \\) that acts on state \\( S_{\\mathrm{base}} \\) to produce the initial Blender environment through a dynamics function \\( F \\) that transitions from one state to another based on a set of programmatic actions:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.709, + 0.611, + 0.736 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathrm {i n i t}} = F \\left(\\left\\{p _ {0} ^ {(i)} \\right\\} _ {i = 1 \\ldots k}, S _ {\\mathrm {b a s e}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.785, + 0.843 + ], + "angle": 0, + "content": "In our problem setting, \\( F \\) is the python code executor within the Blender environment that executes \\( \\{p_0^{(i)}\\}_{i = 1\\dots k} \\) in sequence. We set each initial program \\( p_0^{(i)} \\) to be a program that concerns a single part of the 3D graphical design workflow - for instance, \\( p_0^{(1)} \\) is in charge of the material on one mesh within the scene, and \\( p_0^{(2)} \\) is in charge of the lighting setup of the entire scene. The decomposition of \\( S_{\\mathrm{init}} \\) into \\( S_{\\mathrm{base}} \\) and \\( p_0^1\\dots p_0^k \\) can be done using techniques like the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.786, + 0.193 + ], + "angle": 0, + "content": "“node transpler” from Infinigen [32], which converts entities within the Blender instance into lines of Python code that can recreate a node graph, like a material shader graph. We develop a suite of tools to do this in our own problem setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.195, + 0.788, + 0.302 + ], + "angle": 0, + "content": "Though it's possible for all edits to be encompassed in a single program instead of \\( k \\) programs, this is limiting in practice - either because the VLM's output length isn't large enough for the code necessary or because the VLM has a low success rate, due to the program search space exploding in size. Although it is possible that future VLMs will substantially mitigate this problem, splitting the program into \\( k \\) task-specific programs may still be desirable, given the possibility of querying \\( k \\) task-specific fine-tuned/expert VLMs in parallel." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.334, + 0.687, + 0.349 + ], + "angle": 0, + "content": "3.2 Iterative Refinement of Individual Visual Programs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Suppose that to complete a task like material editing for a single object, it suffices to decompose the initial state into \\( S_{\\mathrm{base}} \\) and a single script, \\( p_0 \\) - that is, \\( S_{\\mathrm{init}} = F(\\{p_0\\}, S_{\\mathrm{base}}) \\). Then our goal is to discover some edited version of \\( p_0 \\), called \\( p_1 \\), such that \\( F(\\{p_1\\}, S_{\\mathrm{base}}) \\) produces a visual state better aligned with some user intention \\( I \\). Our system assumes that the user intention \\( I \\) is provided in the form of language and/or image references, leveraging the visual understanding of the latest VLM models to understand user intention." + }, + { + "type": "code_caption", + "bbox": [ + 0.216, + 0.507, + 0.616, + 0.524 + ], + "angle": 0, + "content": "Algorithm 1 Iterative Refinement of Visual Programs" + }, + { + "type": "algorithm", + "bbox": [ + 0.216, + 0.527, + 0.789, + 0.835 + ], + "angle": 0, + "content": "1: procedure TOURNAMENT(State candidates \\(\\{S_1,S_2,\\dots S_k\\}\\) , Visual state evaluator \\(V\\) ,User intention \\(I\\) \n2: if \\(k > 2\\) then \n3: \\(w_{1}\\gets\\) TOURNAMENT( \\(\\{S_1,S_2,\\ldots S_{k / 2}\\}\\) \\(V,I)\\) \n4: \\(w_{2}\\gets\\) TOURNAMENT( \\(\\{S_{k / 2},S_{k + 1},\\ldots S_k\\}\\) \\(V,I)\\) \n5: else \n6: \\(w_{1}\\gets S_{1},w_{2}\\gets S_{2}\\) \n7: end if \n8: return \\(V(w_{1},w_{2},I)\\) \n9: end procedure \n10: procedure REFINE(Depth \\(d\\) , Breadth \\(b\\) , Intention \\(I\\) , Edit Generator \\(G\\) , State \nEvaluator \\(V\\) , Base state \\(S_{\\mathrm{base}}\\) , Initial program \\(p_0\\) , Dynamics Function \\(F\\) \n11: \\(S_0\\gets F(p_0,S_{\\mathrm{base}})\\) \\(S_{\\mathrm{best}}\\gets S_0\\) \\(p_{\\mathrm{best}}\\gets p_0\\) \n12: for \\(i\\gets 1\\) to \\(d\\) do \n13: \\(\\mathcal{P}_i = \\mathcal{N}(p_{\\mathrm{best}})\\) if \\(i\\) is odd else \\(\\mathcal{P}\\) \\(\\triangleright\\) \"Tweak\" or \"Leap\" edits \n14: \\(p_i^1,p_i^2\\dots p_i^b\\gets G(p_{\\mathrm{best}},S_{\\mathrm{best}},I,b,\\mathcal{P}_i)\\) \\(\\triangleright\\) Generate \\(b\\) options \n15: \\(S_i^1\\gets F(p_i^1,S_0),\\dots,S_i^b\\gets F(p_i^b,S_0)\\) \\(\\triangleright\\) Observe the visual states \n16: \\(S_i^*\\gets\\) TOURNAMENT( \\(\\{S_i^1,S_i^2\\dots S_i^b\\} \\cup \\{S_{\\mathrm{best}}\\}\\) \\(V,I)\\) \\(\\triangleright\\) Choose the best \n17: \\(S_{\\mathrm{best}}\\gets S_i^*,p_{\\mathrm{best}}\\gets p_i^*\\) \\(\\triangleright\\) Best visual state and programs so far \n18: end for \n19: return \\(S_{\\mathrm{best}},p_{\\mathrm{best}}\\) \n20: end procedure" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.622, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlenderAlchemy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "To discover a good edit to \\( p_0 \\), we introduce the procedure outlined in Algorithm 1, an iterative refinement loop that repeatedly uses a visual state evaluator \\( V \\) to select among the hypotheses from an edit generator \\( G \\). A single \"agent\" for a certain task like procedural material design can be fully described by \\( (G, V) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.282 + ], + "angle": 0, + "content": "Inspired by works like [51], we propose a visual state evaluator \\( V(S_{1}, S_{2}, I) \\), which is tasked with returning whichever of the two visual states (\\( S_{1} \\) or \\( S_{2} \\)) better matches the user intention \\( I \\). This evaluator is applied recursively to choose the most suitable visual state candidate among \\( b \\) visual state candidates by making \\( \\mathcal{O}(\\log(b)) \\) queries, as done in TOURNAMENT in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.787, + 0.418 + ], + "angle": 0, + "content": "Though it seems straightforward to ask the same VLM to edit the code in a single pass, this leads to many failure cases (see more in the Appendix). Due to the VLM's lack of baked-in understanding of the visual consequences of various programs within Blender, a multi-hypothesis and multi-step approach is more appropriate. Extending Tree-of-Thoughts [57] to the visual domain, \\( G(p, S, I, b, \\mathcal{P}) \\) is a module tasked with generating \\( b \\) different variations of program \\( p \\), conditioned on the current visual state \\( S \\) and user intention \\( I \\), constrained such that the output programs fall within some family of programs \\( \\mathcal{P} \\), which can be used to instill useful priors to the edit generator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.787, + 0.495 + ], + "angle": 0, + "content": "Below we describe some additional system design decisions that ensured better alignment of the resultant edited program to the user intention, either by improving the stability of the procedure or by supplementing the visual understanding of VLMs. The effect of each is investigated in the Appendix through ablations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.509, + 0.787, + 0.584 + ], + "angle": 0, + "content": "Hypothesis Reversion. To improve the stability of the edit discovery process, we add the visual state of the program being edited at every timestep (\\( S_{\\mathrm{best}} \\) for \\( p_{\\mathrm{best}} \\)) as an additional candidate to the selection process, providing the option for the process to revert to an earlier version if the search at a single iteration was unsuccessful. Line 16 in Algorithm 1 shows this." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.599, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Tweak and Leap Edits. An important characteristic of visual programs is that continuous values hard-coded within the program can modify the output just as much as structural changes. This is in contrast to non-visual program synthesis tasks based on unit-tests of I/O specs, like [5, 9], where foundation models are mostly tasked to produce the right structure of the program with minimal hard-coded values. Given a single visual program \\( p \\in \\mathcal{P} \\), the space of visual outputs achievable through only tweaking the numerical values to function parameters and variable assignments can cover a wide range, depending on the fields available in the program. In Algorithm 1, we refer this space as the \"neighborhood\" of \\( p \\) or \"tweak\" edits, \\( \\mathcal{N}(p) \\). Though this can result in a very small change to the program, this can lead to a large visual difference in the final output, and in a few edits can change the \"wrong\" program into one more aligned with the user intention. On the other hand, more drastic changes (or \"leap\" edits) may be needed to accomplish a task. Consider for example the task of changing a perfectly smooth material to have a noised level of roughness scattered sparsely across the material surface. This may require the programmatic addition of the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "relevant nodes (e.g. color ramps or noise texture nodes), and thus the edited program falls outside of \\(\\mathcal{N}(p)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Empirically, we find that the optimal edits are often a mix of tweak and leap edits. As such, our procedure cycles between restricting the edits of \\( G \\) to two different sets: the neighborhood of \\( p_{\\mathrm{best}} \\), and the whole program space \\( \\mathcal{P} \\) (Line 13 in Algorithm 1). In practice, such restrictions are softly enforced through in-context prompting of VLMs, and though their inputs encourage them to abide by these constraints, the model can still produce more drastic \"tweak\" edits or conservative \"leap\" edits as needed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.788, + 0.437 + ], + "angle": 0, + "content": "Visual Imagination. In the case when the user intention is communicated purely textually, their intention may be difficult for the VLM to turn into successful edits. Prior works like [21, 50, 55] have made similar observations of abstract language for 3D scenes. Consider, for example, the prompt \"make me a material that resembles a celestial nebula\". To do this, the VLM must know what a celestial nebula looks like, and how it should change the parameters of the material Shader nodes of, say, a wooden material. We find that in such cases, it's hard for the VLM to directly go from abstract descriptions to low-level program edits that affect low-level properties of the Blender visual state." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.605 + ], + "angle": 0, + "content": "Instead, we propose supplementing the text-to-program understanding of VLM's with the text-to-image understanding in state-of-the-art image generation models. Intermediary visual artifacts (images generated using the user intention) are created and used to guide the refinement process towards a more plausible program edit to match the desired outcome, as shown in Figure 1. The generated images act as image references in addition to the textual intention provided by the user. This constitutes a simple visual chain of thought [49] for visual program editing, which not only creates a reference image for \\( G \\) and \\( V \\) to guide their low-level visual comparisons (e.g. color schemes, material roughness... etc.), but also provides a user-interpretable intermediary step to confirm the desired goals behind an otherwise vague user intention." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.377, + 0.648 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.787, + 0.697 + ], + "angle": 0, + "content": "We demonstrate BlenderAlchemy on editing procedural materials, geometry and lighting setups within Blender, three of the most tedious parts of 3D design." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.721, + 0.498, + 0.737 + ], + "angle": 0, + "content": "4.1 Procedural Material Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Procedural material editing has characteristics that make it difficult for the same reason as a lot of other visual program settings: (1) small edit distances of programs may result in very large visual differences, contributing to potential instabilities in the edit discovery process, (2) it naturally requires trial-and-error when human users are editing them as the desired magnitude of edits depends on the visual output and (3) language descriptions of edit intent typically do not" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlenderAlchemy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "contain low-level information for the editor to know immediately which part of the program to change." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.254 + ], + "angle": 0, + "content": "We demonstrate this on two different kinds of editing tasks: (1) turning the same initial material (a synthetic wooden material) into other materials described by a list of language descriptions and (2) editing many different initial materials to resemble the same target material described by an image input. For the starting materials, we use the synthetic materials from Infinigen [32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.787, + 0.41 + ], + "angle": 0, + "content": "Text-based material-editing An attractive application of our system is in the modification of preexisting procedural materials using natural language descriptions that communicate user intent, a desirable but thus far undemonstrated capability of neurosymbolic methods [34]. We demonstrate our system's capabilities by asking it to edit a wooden material from Infinigen [32] into many other materials according to diverse language descriptions of target materials that are, importantly, not wood. In reality, this is a very challenging task, since this may require a wide range in the size of edits even if the language describing desired target material may be very similar." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.501 + ], + "angle": 0, + "content": "Figure 3 shows examples of edits of the same starter wood material using different language descriptions, and Figure 4 demonstrates the intermediary steps of the problem solving process for a single instance of the problem. Our system is composed of an edit generator that generates 8 hypotheses per iteration, for 4 iterations (\\(d = 4\\), \\(b = 8\\)), cycling between tweak and leap edits. It uses GPT-4V for edit generation and state evaluation, and DallE-3 for visual imagination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.787, + 0.682 + ], + "angle": 0, + "content": "We compare against BlenderGPT, the most recent open-sourced Blender AI agent that use GPT-4 to execute actions within the Blender environment through the Python API. We provide the same target material text prompt to BlenderGPT, as well as the starter code for the initial wood material for reference. We compare the CLIP similarity of their output material to the input text description against our system. BlenderGPT reasons only about how to edit the program using the input text description, doing so in a single pass without state evaluation or multi-hypothesis edit generation. To match the number of edit generator queries we make, we run their method a maximum of 32 times, using the first successful example as its final output. Everything else remains the same, including the starter material program, text description, base Blender state, and lighting setup." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.787, + 0.804 + ], + "angle": 0, + "content": "We find that qualitatively, BlenderGPT produces much shallower and more simplistic edits of the input material, resulting in low-quality output materials and poor alignment with the user intention. Examples can be seen in Figure 5. For instance, observe that for the \"digital camouflage\" example, BlenderAlchemy is able to produce the \"sharper angles\" that the original description requests (See Figure 3) whereas BlenderGPT produces the right colors but fails to create the sharp, digital look. For \"metallic swirl\" example, our system's visual state selection process would weed out insufficiently swirly examples such as the one" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "\\(^{2}\\) At the time of publishing, works like 3DGPT [42], L3GO [53] have not yet open-sourced their code." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.787, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.535, + 0.788, + 0.579 + ], + "angle": 0, + "content": "Fig. 3: Text-based Material Editing Results. The step-by-step edits of a 4x8 version of BlenderAlchemy to the same wooden material, given the text description on the left as the input user intention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.621 + ], + "angle": 0, + "content": "given by BlenderGPT, enabling our method to produce a material closer to the prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Table 1 demonstrates the comparison in terms of the average ViT-B/32 and ViT-L/14 CLIP similarity scores [31] with respect to the language description. Our system's ability to iteratively refine the edits based on multiple guesses at each step gives it the ability to make more substantive edits over the course of the process. Moreover, visual grounding provided both by the visual state evaluator as well as the output of the visual imagination stage guides the program search procedure to better align with the description. For ablations, see the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.84 + ], + "angle": 0, + "content": "We additionally conducted a user preference study to compare BlenderAlchemy's performance with two material generation baselines that use diffusion: TEXTure [33] and Paint3D [60]. We collect 592 Mechanical Turk comparisons between BlenderAlchemy and the baselines from 24 Turkers on materials created using 32 different text prompts. Users must choose the material that best matches the text-prompt. The results show that BlenderAlchemy is preferred to Paint3D \\(73\\%\\) of the time. Users picked BlenderAlchemy over TEXTure \\(56\\%\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlenderAlchemy" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 4: The edit discovery process of turning a wooden material into \"marbled granite\". Each column shows the hypotheses generated by \\( G \\), with the most promising candidates chosen by \\( V \\) indicated by the highlights. Note that iteration 3 proved to be unfruitful according to \\( V \\), and the method reverts to the best candidate from iteration 2, before moving onto iteration 4." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.401, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Table 1: CLIP scores of BlenderAlchemy vs. BlenderGPT for the text-based material editing task. We find that a version of our system that has no visual components (-Vision) still outperforms BlenderGPT. By adding vision to the state evaluator alone (-Vision G) and not the edit generator, a further improvement is observed." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.469, + 0.7, + 0.528 + ], + "angle": 0, + "content": "
MetricBlenderGPT-Vision-Vision GOurs
ViT-B/32 (↑)25.225.727.828.2
ViT-L/14 (↑)21.121.823.424.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.593 + ], + "angle": 0, + "content": "of the time. We found that in \\(71\\%\\) of comparisons, human users also prefer BlenderAlchemy outputs selected by our visual evaluator. This suggests that our visual evaluator makes decisions that align with users' preferences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.667 + ], + "angle": 0, + "content": "Image-based material-editing Given an image of a desired material, the task is to convert the code of the starter material into a material that contains many of the visual attributes of the input image, akin to doing a kind of style transfer for procedural materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.789, + 0.851 + ], + "angle": 0, + "content": "At each step, the edit generator is first asked to textually enumerate a list of obvious visual differences between the current material and the target, then asked to locate lines within the code that may be responsible for these visual differences (e.g. \"the target material looks more rough\" \\(\\rightarrow\\) \"line 23 sets a roughness value, which we should try to increase\") before finally suggested an edited version of the program. As such, low-level visual differences (e.g. color discrepancies) are semantically compressed first into language descriptions (e.g. \"the target is more red\"), before being fed into the editing process, resulting in the behavior that our system produces variations of the input material that resembles the target image along many attributes, even if its outputs don't perfectly match the target image (See Figure 6). Our system is the same as for text-based material editing, but without the need for visual imagination." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig.5: Comparisons between our method and BlenderGPT for the text-based material editing task setting. Note how our materials align better with the original language prompts (See Figure 3 for the original prompts). The input material being edited is the same wooden material." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.404, + 0.645, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.424, + 0.776, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Fig. 6: Material editing based on image inputs. Our edit intention is described by the target image shown on the right. 5 different Infinigen [32] initial materials are shown here, and the final edit. Note how in each case, certain attributes of the target material (metallicness, color, texture) are transferred." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.626, + 0.414, + 0.642 + ], + "angle": 0, + "content": "4.2 Geometry Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.734 + ], + "angle": 0, + "content": "BlenderAlchemy can also be used to manipulate geometry. We show qualitative examples of our system controlling geometry by programmatically 1) interpolating between preset blend shapes, 2) editing of geometry node graphs, and 3) the precise placement of objects within scenes. The blend weights, geometry node graphs and object placements can all be represented programmatically." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As shown in Figure 7, given an input shape and pre-defined blend shapes, BlenderAlchemy can modify the shape by finding the appropriate blend weights between different shapes to match language, which is non-trivial for human users given the number of blend shapes to interpolate for complex characters and shapes. For instance, the facial expression editing example requires coordinated interpolation between 19 different blend shapes, each which controls a specific kind of motion in one part of the face (e.g. an left eye-brow raise). To aid it in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlenderAlchemy" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.788, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.361, + 0.785, + 0.389 + ], + "angle": 0, + "content": "Fig. 7: BlenderAlchemy editing geometry using blend shapes. Edits are made to match a description or a script line. Input shapes from BlenderKit." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.394, + 0.788, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.548, + 0.785, + 0.577 + ], + "angle": 0, + "content": "Fig. 8: BlenderAlchemy editing (a) the placement of assets within a living room scene and (b) the procedural geometry nodes of roses, to different text prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.785, + 0.621 + ], + "angle": 0, + "content": "this task, our task description to BlenderAlchemy includes language labels for each blend shape." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.743 + ], + "angle": 0, + "content": "In Figure 8(a), BlenderAlchemy is presented with a scene layout editing task, where assets are initially dropped out of camera view and must be iteratively moved to fit the scene of a living room and the requirement that the couch is \"facing the window\", through calling a set of object placement helper functions. Figure 8(b) shows examples of BlenderAlchemy editing the procedural geometry of a set of roses to match a nuanced instruction that influences the number and angle of pedals, as well as the density of leaves. The full geometry node graph that BlenderAlchemy edits contains 60 geometry nodes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.764, + 0.452, + 0.781 + ], + "angle": 0, + "content": "4.3 Lighting Setup Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We show that BlenderAlchemy can be used to adjust the lighting of scenes according to language instructions as well. Figure 9 shows this being done for an input product visualization designed by a Blender artist." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.147, + 0.66, + 0.16 + ], + "angle": 0, + "content": "\"Hand lotion under disco lights, like at a nightclub.\"" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.165, + 0.359, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.165, + 0.497, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.167, + 0.641, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.168, + 0.783, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.285, + 0.784, + 0.325 + ], + "angle": 0, + "content": "Fig. 9: Optimizing the lighting of the scene setup to the text-based user intention. We base the initial Blender state input based on a product visualization downloaded from BlenderKit." + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.33, + 0.674, + 0.343 + ], + "angle": 0, + "content": "\"Hand cream sitting on a hot comet outside in the night\"" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.348, + 0.358, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.349, + 0.496, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.349, + 0.614, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.35, + 0.782, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.47, + 0.785, + 0.539 + ], + "angle": 0, + "content": "Fig. 10: Optimizing the lighting and material iteratively within a product visualization scene, to satisfy the text-based user intention. Note the dimming of the environment lighting for the nighttime lighting, and the glowing-hot material the editing procedure has produced. We base the initial Blender state input based on a product visualization downloaded from BlenderKit." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.678 + ], + "angle": 0, + "content": "As mentioned in Section 3.1, we can consider iteratively optimizing two separate programs, one controlling the lighting of the whole scene and another controlling the material of an object within the scene. That is, \\( S_{\\mathrm{init}} = F(\\{p_0^{(L)}, p_0^{(M)}\\}, S_{\\mathrm{base}}) \\) for initial lighting program \\( p_0^{(L)} \\) and material program \\( p_0^{(M)} \\). Figure 10 shows an example of this, where two separate pairs of edit generators and state evaluators, \\( (G_M, V_M) \\) and \\( (G_L, V_L) \\), are used to achieve an edit to the scene that aligns with the user intention. The full algorithm is shown in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.686, + 0.491, + 0.701 + ], + "angle": 0, + "content": "5 Conclusion & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.846 + ], + "angle": 0, + "content": "In this paper, we introduce BlenderAlchemy, a system that performs edits within the Blender 3D design environment by leveraging vision-language models to iteratively refining a program to be more aligned with the user intention, by using visual information to both explore and prune possibilities within the program space. We equip our system with visual imagination by providing it access to text-to-image models, a tool it uses to guide itself towards program edits that better align with user intentions. We've demonstrated BlenderAlchemy on editing materials, geometry and lighting, and hope that future works will extend this to other workflows as well." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlenderAlchemy" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.24 + ], + "angle": 0, + "content": "We acknowledge the support of ARL grant W911NF-21-2-0104 and a Vannevar Bush Faculty Fellowship. We'd additionally like to thank Maneesh Agrawala for general discussions, and Purvi Goel, Mika Uy, Vishnu Sarukkai, Fan-yun Sun and Sharon Lee for feedback on paper revisions." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.324, + 0.28 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.6, + 0.311 + ], + "angle": 0, + "content": "1. Blendergt, https://github.com/gd3kr/BlenderGPT" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.312, + 0.784, + 0.338 + ], + "angle": 0, + "content": "2. How long does it take to create a 3d model?, https://3d-ace.com/blog/how-long-does-it-take-to-create-a-3d-model/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.367 + ], + "angle": 0, + "content": "3. How long does it take to make a 3d model?, https://pixune.com/blog/how-long-does-it-take-to-create-a-3d-model/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.409 + ], + "angle": 0, + "content": "4. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al.: Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.785, + 0.451 + ], + "angle": 0, + "content": "5. Austin, J., Odena, A., Nye, M., Bosma, M., Michalewski, H., Dohan, D., Jiang, E., Cai, C., Terry, M., Le, Q., et al.: Program synthesis with large language models. arXiv preprint arXiv:2108.07732 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.785, + 0.492 + ], + "angle": 0, + "content": "6. Baumli, K., Baveja, S., Behbahani, F., Chan, H., Comanici, G., Flennerhag, S., Gazeau, M., Holsheimer, K., Horgan, D., Laskin, M., et al.: Vision-language models as a source of rewards. arXiv preprint arXiv:2312.09187 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.535 + ], + "angle": 0, + "content": "7. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opena.com/papers/dall-e-3.pdf 2(3), 8 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.576 + ], + "angle": 0, + "content": "8. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.715 + ], + "angle": 0, + "content": "9. Chen, M., Tworek, J., Jun, H., Yuan, Q., de Oliveira Pinto, H.P., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F.P., Cummings, D., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W.H., Nichol, A., Paino, A., Tezak, N., Tang, J., Babuschkin, I., Balaji, S., Jain, S., Saunders, W., Hesse, C., Carr, A.N., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., Zaremba, W.: Evaluating large language models trained on code (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "10. Chen, Y., Chen, R., Lei, J., Zhang, Y., Jia, K.: Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems 35, 30923-30936 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "11. De La Torre, F., Fang, C.M., Huang, H., Banburski-Fahey, A., Fernandez, J.A., Lanier, J.: Llmr: Real-time prompting of interactive worlds using large language models. arXiv preprint arXiv:2309.12276 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Firoozzi, R., Tucker, J., Tian, S., Majumdar, A., Sun, J., Liu, W., Zhu, Y., Song, S., Kapoor, A., Hausman, K., et al.: Foundation models in robotics: Applications, challenges, and the future. arXiv preprint arXiv:2312.07843 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.297, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., Wu, Y., Ji, R.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.244 + ], + "angle": 0, + "content": "14. Fu, C., Zhang, R., Wang, Z., Huang, Y., Zhang, Z., Qiu, L., Ye, G., Shen, Y., Zhang, M., Chen, P., Zhao, S., Lin, S., Jiang, D., Yin, D., Gao, P., Li, K., Li, H., Sun, X.: A challenger to gpt-4v? early explorations of gemini in visual expertise. arXiv preprint arXiv:2312.12436 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.787, + 0.271 + ], + "angle": 0, + "content": "15. Goel, P., Wang, K.C., Liu, C.K., Fatahalian, K.: Iterative motion editing with natural language. arXiv preprint arXiv:2312.11538 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "16. Guerrero, P., Hasan, M., Sunkavalli, K., Meech, R., Boubekeur, T., Mitra, N.J.: Matformer: A generative model for procedural materials. arXiv preprint arXiv:2207.01044 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.787, + 0.338 + ], + "angle": 0, + "content": "17. Henzler, P., Deschaintre, V., Mitra, N.J., Ritschel, T.: Generative modelling of brdf textures from flash images. arXiv preprint arXiv:2102.11861 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "18. Hu, Y., Xie, Q., Jain, V., Francis, J., Patrikar, J., Keetha, N., Kim, S., Xie, Y., Zhang, T., Zhao, Z., et al.: Toward general-purpose robots via foundation models: A survey and meta-analysis. arXiv preprint arXiv:2312.08782 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "19. Hu, Y., Guerrero, P., Hasan, M., Rushmeier, H., Deschaintre, V.: Node graph optimization using differentiable proxies. In: ACM SIGGRAPH 2022 conference proceedings. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.787, + 0.46 + ], + "angle": 0, + "content": "20. Hu, Y., He, C., Deschaintre, V., Dorsey, J., Rushmeier, H.: An inverse procedural modeling pipeline for svbrdf maps. ACM Transactions on Graphics (TOG) 41(2), 1-17 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "21. Huang, I., Krishna, V., Atekha, O., Guibas, L.: Aladdin: Zero-shot hallucination of stylized 3d assets from abstract scene descriptions. arXiv preprint arXiv:2306.06212 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "22. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.595 + ], + "angle": 0, + "content": "23. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.651 + ], + "angle": 0, + "content": "24. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., Zeng, A.: Code as policies: Language model programs for embodied control. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9493-9500. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.692 + ], + "angle": 0, + "content": "25. Liu, J., Gan, Y., Dong, J., Qi, L., Sun, X., Jian, M., Wang, L., Yu, H.: Perceptiondriven procedural texture generation from examples. Neurocomputing 291, 21-34 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.787, + 0.732 + ], + "angle": 0, + "content": "26. Olausson, T.X., Inala, J.P., Wang, C., Gao, J., Solar-Lezama, A.: Is self-repair a silver bullet for code generation? In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "27. OpenAI: Gpt-4 system card. OpenAI (2023), https://cdn.openai.com/papers/gpt-4-system-card.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.786 + ], + "angle": 0, + "content": "28. OpenAI: Gpt-4v(ison) system card. OpenAI (2023), https://apisemantic scholar.org/CorpusID:263218031" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "29. Park, J.S., O'Brien, J., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. pp. 1-22 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.622, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlenderAlchemy" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.176 + ], + "angle": 0, + "content": "30. Patil, S.G., Zhang, T., Wang, X., Gonzalez, J.E.: Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.232 + ], + "angle": 0, + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.288 + ], + "angle": 0, + "content": "32. Raistrick, A., Lipson, L., Ma, Z., Mei, L., Wang, M., Zuo, Y., Kayan, K., Wen, H., Han, B., Wang, Y., et al.: Infinite photorealistic worlds using procedural generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12630-12641 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.318 + ], + "angle": 0, + "content": "33. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.319, + 0.788, + 0.36 + ], + "angle": 0, + "content": "34. Ritchie, D., Guerrero, P., Jones, R.K., Mitra, N.J., Schulz, A., Willis, K.D., Wu, J.: Neurosymbolic models for computer graphics. In: Computer Graphics Forum. vol. 42, pp. 545-568. Wiley Online Library (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.416 + ], + "angle": 0, + "content": "35. Romera-Paredes, B., Barekatain, M., Novikov, A., Balog, M., Kumar, M.P., Dupont, E., Ruiz, F.J., Ellenberg, J.S., Wang, P., Fawzi, O., et al.: Mathematical discoveries from program search with large language models. Nature 625(7995), 468-475 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.472 + ], + "angle": 0, + "content": "36. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Hambro, E., Zettle-moyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.515 + ], + "angle": 0, + "content": "37. Sharma, P., Jampani, V., Li, Y., Jia, X., Lagun, D., Durand, F., Freeman, W.T., Matthews, M.: Alchemist: Parametric control of material properties with diffusion models. arXiv preprint arXiv:2312.02970 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.558 + ], + "angle": 0, + "content": "38. Shi, L., Li, B., Hasan, M., Sunkavalli, K., Boubekeur, T., Mech, R., Matusik, W.: Match: Differentiable material graphs for procedural material capture. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.613 + ], + "angle": 0, + "content": "39. Shimizu, E., Fisher, M., Paris, S., McCann, J., Fatahalian, K.: Design adjectives: A framework for interactive model-guided exploration of parameterized design spaces. In: Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. pp. 261-278 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.655 + ], + "angle": 0, + "content": "40. Shinn, N., Cassano, F., Labash, B., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Language agents with verbal reinforcement learning.(2023). arXiv preprint cs.AI/2303.11366 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.713 + ], + "angle": 0, + "content": "41. Singh, I., Blukis, V., Mousavian, A., Goyal, A., Xu, D., Tremblay, J., Fox, D., Thomason, J., Garg, A.: Progprompt: Generating situated robot task plans using large language models. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11523-11530. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.714, + 0.788, + 0.742 + ], + "angle": 0, + "content": "42. Sun, C., Han, J., Deng, W., Wang, X., Qin, Z., Gould, S.: 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.798 + ], + "angle": 0, + "content": "43. Tchapmi, L.P., Ray, T., Tchapmi, M., Shen, B., Martin-Martin, R., Savarese, S.: Generating procedural 3d materials from images using neural networks. In: 2022 4th International Conference on Image, Video and Signal Processing. pp. 32-40 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "46. Vecchio, G., Martin, R., Roullier, A., Kaiser, A., Rouffet, R., Deschaintre, V., Boubekeur, T.: Controlmat: A controlled generative approach to material capture. arXiv preprint arXiv:2309.01700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.257 + ], + "angle": 0, + "content": "47. Vecchio, G., Sortino, R., Palazzo, S., Spampinato, C.: Matfuse: Controllable material generation with diffusion models. arXiv preprint arXiv:2308.11408 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.298 + ], + "angle": 0, + "content": "48. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.339 + ], + "angle": 0, + "content": "49. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.366 + ], + "angle": 0, + "content": "50. Wen, Z., Liu, Z., Sridhar, S., Fu, R.: Anyhome: Open-vocabulary generation of structured and textured 3d homes. arXiv preprint arXiv:2312.06644 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "51. Wu, T., Yang, G., Li, Z., Zhang, K., Liu, Z., Guibas, L., Lin, D., Wetzstein, G.: Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. arXiv preprint arXiv:2401.04092 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.448 + ], + "angle": 0, + "content": "52. Xiao, X., Liu, J., Wang, Z., Zhou, Y., Qi, Y., Cheng, Q., He, B., Jiang, S.: Robot learning in the era of foundation models: A survey. arXiv preprint arXiv:2311.14379 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.448, + 0.785, + 0.489 + ], + "angle": 0, + "content": "53. Yamada, Y., Chandu, K., Lin, Y., Hessel, J., Yildirim, I., Choi, Y.: L3go: Language agents with chain-of-3d-thoughts for generating unconventional objects. arXiv preprint arXiv:2402.09052 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.785, + 0.53 + ], + "angle": 0, + "content": "54. Yang, H., Chen, Y., Pan, Y., Yao, T., Chen, Z., Mei, T.: 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 6860-6868 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.585 + ], + "angle": 0, + "content": "55. Yang, Y., Sun, F.Y., Weihs, L., VanderBilt, E., Herrasti, A., Han, W., Wu, J., Haber, N., Krishna, R., Liu, L., et al.: Holodeck: Language guided generation of 3d embodied ai environments. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024). vol. 30, pp. 20-25. IEEE/CVF (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.585, + 0.785, + 0.625 + ], + "angle": 0, + "content": "56. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.626, + 0.785, + 0.667 + ], + "angle": 0, + "content": "57. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.667, + 0.785, + 0.694 + ], + "angle": 0, + "content": "58. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.694, + 0.785, + 0.735 + ], + "angle": 0, + "content": "59. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2310.16045 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.735, + 0.785, + 0.762 + ], + "angle": 0, + "content": "60. Zeng, X.: Paint3d: Paint anything 3d with lighting-less texture diffusion models. arXiv preprint arXiv:2312.13913 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.762, + 0.785, + 0.803 + ], + "angle": 0, + "content": "61. Zhou, H., Yao, X., Meng, Y., Sun, S., BIng, Z., Huang, K., Knoll, A.: Language-conditioned learning for robotic manipulation: A survey. arXiv preprint arXiv:2312.10807 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.803, + 0.785, + 0.831 + ], + "angle": 0, + "content": "62. Zsolnai-Fehér, K., Wonka, P., Wimmer, M.: Gaussian material synthesis. arXiv preprint arXiv:1804.08369 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.831 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/19f5daea-7c9c-46d5-b550-8a14478e364b_origin.pdf b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/19f5daea-7c9c-46d5-b550-8a14478e364b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bc0f6a52b9ed53fffe7fa45aed7b7df00018dbf3 --- /dev/null +++ b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/19f5daea-7c9c-46d5-b550-8a14478e364b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:077527ecdfb56f1c372ef7d24893406708d209436d3af0b0ef04bf4326120163 +size 25847633 diff --git a/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/full.md b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1f73afc2c85501df2d485a2a60135698c39bd5bb --- /dev/null +++ b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/full.md @@ -0,0 +1,275 @@ +# BlenderAlchemy: Editing 3D Graphics with Vision-Language Models + +Ian Huang, Guandao Yang, and Leonidas Guibas + +Stanford University + +Abstract. Graphics design is important for various applications, including movie production and game design. To create a high-quality scene, designers usually need to spend hours in software like Blender, in which they might need to interleave and repeat operations, such as connecting material nodes, hundreds of times. Moreover, slightly different design goals may require completely different sequences, making automation difficult. In this paper, we propose a system that leverages Vision-Language Models (VLMs), like GPT-4V, to intelligently search the design action space to arrive at an answer that can satisfy a user's intent. Specifically, we design a vision-based edit generator and state evaluator to work together to find the correct sequence of actions to achieve the goal. Inspired by the role of visual imagination in the human design process, we supplement the visual reasoning capabilities of VLMs with "imagined" reference images from image-generation models, providing visual grounding of abstract language descriptions. In this paper, we provide empirical evidence suggesting our system can produce simple but tedious Blender editing sequences for tasks such as editing procedural materials and geometry from text and/or reference images, as well as adjusting lighting configurations for product renderings in complex scenes. + +# 1 Introduction + +To produce the compelling graphics content we see in movies or video games, 3D artists usually need to spend hours in software like Blender to find appropriate surface materials, object placements, and lighting arrangements. These operations require the artist to create a mental picture of the target, experiment with different parameters, and visually examine whether their edits get closer to the end goal. One can imagine automating these processes by converting language or visual descriptions of user intent into edits that achieve a design goal. Such a system can improve the productivity of millions of 3D designers and impact various industries that depend on 3D graphic design. + +Graphic design is very challenging because even a small design goal requires performing a variety of different tasks. For instance, modeling of a game environment requires the 3D artist to cycle between performing modeling, material + +![](images/b8d18be0552d5a1e9b290069f16b51083f6875aa4f199495c78b9900e7b893f3.jpg) +Fig. 1: Overview of BlenderAlchemy. Given an input Blender state and a user intention specified using either language or reference images, BlenderAlchemy edits the Blender state to satisfy that intention by iteratively refining a Blender python program that executes to produce the final Blender state. Our system additionally leverages text-to-image generation for visual imagination, a step that expands a text-only user intention to a concrete visual target to improve program refinement. + +design, texture painting, animation, lighting, and scene composition. Prior attempts usually focus on specific editing tasks, like material synthesis [8, 62]. While these approaches show promising performance in the tasks they are designed for, it is non-trivial to combine these task-specific methods to satisfy an intended design goal. An alternative is to leverage Large Language Models (LLMs) [22, 27, 45] to digest user intent and suggest design actions by proposing intelligent combination of existing task-specific tools [36] or predicting edits to programs step-by-step [49, 57]. While LLMs have excellent abilities to understand user intentions and suggest sequences of actions to satisfy them, applying LLMs to graphical design remains challenging largely because language cannot capture the visual consequences of actions performed in software like Blender. + +One promising alternative is to leverage vision language models (VLM), such as LLaVA [23], GPT-4V [28], Gemini [44], and DallE-3 [7]. These VLMs have shown to be highly capable of understanding detailed visual information [13,56, 58, 59] and generating compelling images [7]. We posit that these VLMs can be leveraged to complete different kinds of design tasks within the Blender design environment, like editing materials, geometry and lighting setups. + +In this paper, we provide a proof-of-concept system using the vision foundation model GPT-4V to generate and edit programs that modify the state of a Blender workspace to satisfy a user intention. Specifically, we first initialize the state of a workspace within Blender. The Blender state is parameterized as a short Python program, and a base Blender file. The user will then input a text description and potentially a reference image to communicate the desirable design outcome. The system is tasked to edit the program so that, when executed on the base Blender file, the rendered image can satisfy the user's intention. Figure 1 provides an illustration of the problem setup. + +![](images/faf85e5061ad7daf08a18921d7ff18c60663be6d15cecfc1d1fdd6904cb83b90.jpg) +Fig. 2: Iterative visual program editing employs a edit generator $G$ and a state evaluator $V$ in each iteration to explore and prune different potential program edits, where $G$ generates plausible variants of an input program and $V$ picks between two programs based on the consequences they have to the Blender visual state and their alignment to the user intention. Each iteration of the refinement explores variations of the most promising program from the previous iteration. See Algorithm 1 for details. + +Naively applying VLMs to this editing setting gives rise to many failure cases, possibly due to the fact that out-of-the-box VLMs have a poor understanding of the visual consequences of Blender program edits. To counter this, we propose a visually-guided program search procedure that combines a vision-aware edit generator and a visual state evaluator to iteratively search for a suitable program edit (Figure 2). Inspired by the human design process, our system performs guided trial-and-error, capped by some computation budget. Within each iteration, the visual program generator will propose several possible edits on the current program. These edits will be applied and executed in Blender to produce a rendered image. These rendered images will be provided to visual evaluator, which will select the best renders via pairwise comparison by assessing which choice better satisfies the design goal specified by text and reference image. The program achieving the best render will replace the current program as a starting point for the visual program generator. This iterative search process, however, has very low success rate because of the sparsity of correct program edits in the vast program space. To improve the success rate, we further propose two techniques. First, when the proposed program of the new iteration does not contain a viable candidate, we revert to the best candidate of the prior program. This reversion mechanism makes sure the search procedure will not diverge when facing a batch of bad edit candidates. Second, to facilitate our visual evaluator and generator to better understand user intent, we leverage the "visual imagination" of text-to-image generative model to imagine a reference image. We show that our method is capable of accomplishing graphical design tasks within Blender, guided by user intention in the form of text and images. We demonstrate the effectiveness of our system on material, geometry and lighting design tasks, all parts of the 3D design process where artists spend a significant amount of time, ranging from 20 hrs to 4-6 workdays per model [2,3]. We show that our method can outperform prior works designed for similar problem settings, such as BlenderGPT [1]. In summary, the main contributions of this paper include: + +1. We propose BlenderAlchemy, a system that's able to edit visual programs based on user input in the form of text or images. + +2. We identify key components that make the system work: a visual state evaluator, a visual edit generator, a search algorithm with an edit reversion mechanism, and a visual imagination module to facilitate the search. +3. We provide evidence showing that our system can outperform prior works in text-based procedural material editing, as well as its applicability to other design tasks including editing geometry and lighting configurations. + +# 2 Related works + +Task-specific tools for Material Design. Large bodies of works have been dedicated to using learning-based approaches to generate materials. Prior works exploit 2D diffusion models to generate texture maps either in a zero-shot way [8, 21, 33, 50, 54] or through fine-tuning [37, 60]. Though these works open the possibility of generating and modifying textures of objects using natural language for 3D meshes, these works do not model the material properties in a way that allows such objects to be relit. Other methods directly predict the physical properties of the surface of a material through learning, using diffusion [46, 47] or learning good latent representations from data [10, 17, 20, 25, 39, 62]. However, for all of the work mentioned so far, their fundamentally image-based or latent-based representations make the output materials difficult to edit in existing 3D graphics pipelines. People have also explored combining learning-based approaches with symbolic representations of materials [34, 43]. These works often involve creating differentiable representations of the procedural material graph often used in 3D graphical design pipelines, and backpropagating gradients throughout the graph to produce an image that can match the target [16, 19, 38]. Other works like Infinigen [32] use rule-based procedural generation over a library of procedural materials. However, no prior works in this direction have demonstrated the ability to edit procedural material graphs using user intention specified by language [34], a task that we are particularly interested in. Though the aforementioned works excel at material design, they aren't generalizable to other 3D graphics design task settings. BlenderAlchemy, on the other hand, aims to produce a system that can perform various design tasks according to user intents. This usually requires combining different methods together in a non-trivial way. + +LLM as general problem solvers. Large language models (LLMs) like GPT-4 [27,28], Llama [45], and Mistral [22] have in recent years demonstrated unprecedented results in a variety of problem settings, like robotics [4,12,18,24,41,52,61], program synthesis [26,35,40], and graphic design [21,50,55]. Other works have shown that by extending such models with an external process like Chain-of-Thought [49], Tree-of-Thought [57] or memory/skill database [11, 29, 30, 48], or by embedding such systems within environments where it can perceive and act [11, 29, 36, 48], a range of new problems that require iterative refinement can be solved. Their application to visual problem settings, however, has mostly been limited due to the nonexistent visual perception capabilities of the LLMs [11, 15, 48, 50, 53, 55]. While this could be sidestepped by fully condensing the visual state of the environment using text [29] or some symbolic representation [11,48], doing so for 3D graphic design works poorly. For instance, the task + +of editing a Blender material graph to create a desired material requires many trial-and-error cycles and an accurate understanding of the consequences certain design actions can have on the visual output. Recent works that apply LLMs to graphical design settings [11, 15] and ones that more specifically do so within Blender (like BlenderGPT [1], 3D-GPT [42] and L3GO [53]) do not use visual information to inform or refine their system's outputs, leading to unsatisfactory results. BlenderAlchemy borrows ideas from existing LLM literature and tries to address this issue by inputting visual perception into the system. + +Vision-Language Models. State-of-the-art vision language models, such as LLaVA [23], GPT-4V [28], and Gemini [44] have demonstrated impressive understanding of the visual world and its connections to language and semantics, enabling many computer vision tasks like scene understanding, visual question answering, and object detection to be one API call away [13,14,56,58,59]. Works such as [6,51] suggest that such models can also be used as a replacement for human evaluators for a lot of tasks, positioning them as tools for guiding planning and search by acting as flexible reward functions. BlenderAlchemy takes the first steps to apply VLMs to solve 3D graphic design tasks, a novel yet challenging application rather unexplored by existing works. + +# 3 Method + +The goal of our system is to perform edits within the Blender 3D graphic design environment through iteratively refining programs that define a sequence of edits in Blender. This requires us to (1) decompose the input initial Blender input into a combination of programs and a "base" Blender state (Section 3.1) and (2) develop a procedure to edit each program to produce the desired visual state within Blender to match a user intention (Sections 3.2). + +# 3.1 Representation of the Blender Visual State + +The state of the initial Blender design environment can be decomposed into an "base" Blender state $S_{\mathrm{base}}$ and a set of initial programs $\{p_0^{(1)}, p_0^{(2)}, \dots, p_0^{(k)}\}$ that acts on state $S_{\mathrm{base}}$ to produce the initial Blender environment through a dynamics function $F$ that transitions from one state to another based on a set of programmatic actions: + +$$ +S _ {\mathrm {i n i t}} = F \left(\left\{p _ {0} ^ {(i)} \right\} _ {i = 1 \ldots k}, S _ {\mathrm {b a s e}}\right) +$$ + +In our problem setting, $F$ is the python code executor within the Blender environment that executes $\{p_0^{(i)}\}_{i = 1\dots k}$ in sequence. We set each initial program $p_0^{(i)}$ to be a program that concerns a single part of the 3D graphical design workflow - for instance, $p_0^{(1)}$ is in charge of the material on one mesh within the scene, and $p_0^{(2)}$ is in charge of the lighting setup of the entire scene. The decomposition of $S_{\mathrm{init}}$ into $S_{\mathrm{base}}$ and $p_0^1\dots p_0^k$ can be done using techniques like the + +“node transpler” from Infinigen [32], which converts entities within the Blender instance into lines of Python code that can recreate a node graph, like a material shader graph. We develop a suite of tools to do this in our own problem setting. + +Though it's possible for all edits to be encompassed in a single program instead of $k$ programs, this is limiting in practice - either because the VLM's output length isn't large enough for the code necessary or because the VLM has a low success rate, due to the program search space exploding in size. Although it is possible that future VLMs will substantially mitigate this problem, splitting the program into $k$ task-specific programs may still be desirable, given the possibility of querying $k$ task-specific fine-tuned/expert VLMs in parallel. + +# 3.2 Iterative Refinement of Individual Visual Programs + +Suppose that to complete a task like material editing for a single object, it suffices to decompose the initial state into $S_{\mathrm{base}}$ and a single script, $p_0$ - that is, $S_{\mathrm{init}} = F(\{p_0\}, S_{\mathrm{base}})$ . Then our goal is to discover some edited version of $p_0$ , called $p_1$ , such that $F(\{p_1\}, S_{\mathrm{base}})$ produces a visual state better aligned with some user intention $I$ . Our system assumes that the user intention $I$ is provided in the form of language and/or image references, leveraging the visual understanding of the latest VLM models to understand user intention. + +Algorithm 1 Iterative Refinement of Visual Programs +1: procedure TOURNAMENT(State candidates $\{S_1,S_2,\dots S_k\}$ , Visual state evaluator $V$ ,User intention $I$ +2: if $k > 2$ then +3: $w_{1}\gets$ TOURNAMENT( $\{S_1,S_2,\ldots S_{k / 2}\}$ $V,I)$ +4: $w_{2}\gets$ TOURNAMENT( $\{S_{k / 2},S_{k + 1},\ldots S_k\}$ $V,I)$ +5: else +6: $w_{1}\gets S_{1},w_{2}\gets S_{2}$ +7: end if +8: return $V(w_{1},w_{2},I)$ +9: end procedure +10: procedure REFINE(Depth $d$ , Breadth $b$ , Intention $I$ , Edit Generator $G$ , State +Evaluator $V$ , Base state $S_{\mathrm{base}}$ , Initial program $p_0$ , Dynamics Function $F$ +11: $S_0\gets F(p_0,S_{\mathrm{base}})$ $S_{\mathrm{best}}\gets S_0$ $p_{\mathrm{best}}\gets p_0$ +12: for $i\gets 1$ to $d$ do +13: $\mathcal{P}_i = \mathcal{N}(p_{\mathrm{best}})$ if $i$ is odd else $\mathcal{P}$ $\triangleright$ "Tweak" or "Leap" edits +14: $p_i^1,p_i^2\dots p_i^b\gets G(p_{\mathrm{best}},S_{\mathrm{best}},I,b,\mathcal{P}_i)$ $\triangleright$ Generate $b$ options +15: $S_i^1\gets F(p_i^1,S_0),\dots,S_i^b\gets F(p_i^b,S_0)$ $\triangleright$ Observe the visual states +16: $S_i^*\gets$ TOURNAMENT( $\{S_i^1,S_i^2\dots S_i^b\} \cup \{S_{\mathrm{best}}\}$ $V,I)$ $\triangleright$ Choose the best +17: $S_{\mathrm{best}}\gets S_i^*,p_{\mathrm{best}}\gets p_i^*$ $\triangleright$ Best visual state and programs so far +18: end for +19: return $S_{\mathrm{best}},p_{\mathrm{best}}$ +20: end procedure + +To discover a good edit to $p_0$ , we introduce the procedure outlined in Algorithm 1, an iterative refinement loop that repeatedly uses a visual state evaluator $V$ to select among the hypotheses from an edit generator $G$ . A single "agent" for a certain task like procedural material design can be fully described by $(G, V)$ . + +Inspired by works like [51], we propose a visual state evaluator $V(S_{1}, S_{2}, I)$ , which is tasked with returning whichever of the two visual states ( $S_{1}$ or $S_{2}$ ) better matches the user intention $I$ . This evaluator is applied recursively to choose the most suitable visual state candidate among $b$ visual state candidates by making $\mathcal{O}(\log(b))$ queries, as done in TOURNAMENT in Algorithm 1. + +Though it seems straightforward to ask the same VLM to edit the code in a single pass, this leads to many failure cases (see more in the Appendix). Due to the VLM's lack of baked-in understanding of the visual consequences of various programs within Blender, a multi-hypothesis and multi-step approach is more appropriate. Extending Tree-of-Thoughts [57] to the visual domain, $G(p, S, I, b, \mathcal{P})$ is a module tasked with generating $b$ different variations of program $p$ , conditioned on the current visual state $S$ and user intention $I$ , constrained such that the output programs fall within some family of programs $\mathcal{P}$ , which can be used to instill useful priors to the edit generator. + +Below we describe some additional system design decisions that ensured better alignment of the resultant edited program to the user intention, either by improving the stability of the procedure or by supplementing the visual understanding of VLMs. The effect of each is investigated in the Appendix through ablations. + +Hypothesis Reversion. To improve the stability of the edit discovery process, we add the visual state of the program being edited at every timestep ( $S_{\mathrm{best}}$ for $p_{\mathrm{best}}$ ) as an additional candidate to the selection process, providing the option for the process to revert to an earlier version if the search at a single iteration was unsuccessful. Line 16 in Algorithm 1 shows this. + +Tweak and Leap Edits. An important characteristic of visual programs is that continuous values hard-coded within the program can modify the output just as much as structural changes. This is in contrast to non-visual program synthesis tasks based on unit-tests of I/O specs, like [5, 9], where foundation models are mostly tasked to produce the right structure of the program with minimal hard-coded values. Given a single visual program $p \in \mathcal{P}$ , the space of visual outputs achievable through only tweaking the numerical values to function parameters and variable assignments can cover a wide range, depending on the fields available in the program. In Algorithm 1, we refer this space as the "neighborhood" of $p$ or "tweak" edits, $\mathcal{N}(p)$ . Though this can result in a very small change to the program, this can lead to a large visual difference in the final output, and in a few edits can change the "wrong" program into one more aligned with the user intention. On the other hand, more drastic changes (or "leap" edits) may be needed to accomplish a task. Consider for example the task of changing a perfectly smooth material to have a noised level of roughness scattered sparsely across the material surface. This may require the programmatic addition of the + +relevant nodes (e.g. color ramps or noise texture nodes), and thus the edited program falls outside of $\mathcal{N}(p)$ . + +Empirically, we find that the optimal edits are often a mix of tweak and leap edits. As such, our procedure cycles between restricting the edits of $G$ to two different sets: the neighborhood of $p_{\mathrm{best}}$ , and the whole program space $\mathcal{P}$ (Line 13 in Algorithm 1). In practice, such restrictions are softly enforced through in-context prompting of VLMs, and though their inputs encourage them to abide by these constraints, the model can still produce more drastic "tweak" edits or conservative "leap" edits as needed. + +Visual Imagination. In the case when the user intention is communicated purely textually, their intention may be difficult for the VLM to turn into successful edits. Prior works like [21, 50, 55] have made similar observations of abstract language for 3D scenes. Consider, for example, the prompt "make me a material that resembles a celestial nebula". To do this, the VLM must know what a celestial nebula looks like, and how it should change the parameters of the material Shader nodes of, say, a wooden material. We find that in such cases, it's hard for the VLM to directly go from abstract descriptions to low-level program edits that affect low-level properties of the Blender visual state. + +Instead, we propose supplementing the text-to-program understanding of VLM's with the text-to-image understanding in state-of-the-art image generation models. Intermediary visual artifacts (images generated using the user intention) are created and used to guide the refinement process towards a more plausible program edit to match the desired outcome, as shown in Figure 1. The generated images act as image references in addition to the textual intention provided by the user. This constitutes a simple visual chain of thought [49] for visual program editing, which not only creates a reference image for $G$ and $V$ to guide their low-level visual comparisons (e.g. color schemes, material roughness... etc.), but also provides a user-interpretable intermediary step to confirm the desired goals behind an otherwise vague user intention. + +# 4 Experiments + +We demonstrate BlenderAlchemy on editing procedural materials, geometry and lighting setups within Blender, three of the most tedious parts of 3D design. + +# 4.1 Procedural Material Editing + +Procedural material editing has characteristics that make it difficult for the same reason as a lot of other visual program settings: (1) small edit distances of programs may result in very large visual differences, contributing to potential instabilities in the edit discovery process, (2) it naturally requires trial-and-error when human users are editing them as the desired magnitude of edits depends on the visual output and (3) language descriptions of edit intent typically do not + +contain low-level information for the editor to know immediately which part of the program to change. + +We demonstrate this on two different kinds of editing tasks: (1) turning the same initial material (a synthetic wooden material) into other materials described by a list of language descriptions and (2) editing many different initial materials to resemble the same target material described by an image input. For the starting materials, we use the synthetic materials from Infinigen [32]. + +Text-based material-editing An attractive application of our system is in the modification of preexisting procedural materials using natural language descriptions that communicate user intent, a desirable but thus far undemonstrated capability of neurosymbolic methods [34]. We demonstrate our system's capabilities by asking it to edit a wooden material from Infinigen [32] into many other materials according to diverse language descriptions of target materials that are, importantly, not wood. In reality, this is a very challenging task, since this may require a wide range in the size of edits even if the language describing desired target material may be very similar. + +Figure 3 shows examples of edits of the same starter wood material using different language descriptions, and Figure 4 demonstrates the intermediary steps of the problem solving process for a single instance of the problem. Our system is composed of an edit generator that generates 8 hypotheses per iteration, for 4 iterations ( $d = 4$ , $b = 8$ ), cycling between tweak and leap edits. It uses GPT-4V for edit generation and state evaluation, and DallE-3 for visual imagination. + +We compare against BlenderGPT, the most recent open-sourced Blender AI agent that use GPT-4 to execute actions within the Blender environment through the Python API. We provide the same target material text prompt to BlenderGPT, as well as the starter code for the initial wood material for reference. We compare the CLIP similarity of their output material to the input text description against our system. BlenderGPT reasons only about how to edit the program using the input text description, doing so in a single pass without state evaluation or multi-hypothesis edit generation. To match the number of edit generator queries we make, we run their method a maximum of 32 times, using the first successful example as its final output. Everything else remains the same, including the starter material program, text description, base Blender state, and lighting setup. + +We find that qualitatively, BlenderGPT produces much shallower and more simplistic edits of the input material, resulting in low-quality output materials and poor alignment with the user intention. Examples can be seen in Figure 5. For instance, observe that for the "digital camouflage" example, BlenderAlchemy is able to produce the "sharper angles" that the original description requests (See Figure 3) whereas BlenderGPT produces the right colors but fails to create the sharp, digital look. For "metallic swirl" example, our system's visual state selection process would weed out insufficiently swirly examples such as the one + +![](images/1b50a424840ecc234239ee56d36f7f5b8b0fc407ae700b8ca4e69a376061f120.jpg) +Fig. 3: Text-based Material Editing Results. The step-by-step edits of a 4x8 version of BlenderAlchemy to the same wooden material, given the text description on the left as the input user intention. + +given by BlenderGPT, enabling our method to produce a material closer to the prompt. + +Table 1 demonstrates the comparison in terms of the average ViT-B/32 and ViT-L/14 CLIP similarity scores [31] with respect to the language description. Our system's ability to iteratively refine the edits based on multiple guesses at each step gives it the ability to make more substantive edits over the course of the process. Moreover, visual grounding provided both by the visual state evaluator as well as the output of the visual imagination stage guides the program search procedure to better align with the description. For ablations, see the Appendix. + +We additionally conducted a user preference study to compare BlenderAlchemy's performance with two material generation baselines that use diffusion: TEXTure [33] and Paint3D [60]. We collect 592 Mechanical Turk comparisons between BlenderAlchemy and the baselines from 24 Turkers on materials created using 32 different text prompts. Users must choose the material that best matches the text-prompt. The results show that BlenderAlchemy is preferred to Paint3D $73\%$ of the time. Users picked BlenderAlchemy over TEXTure $56\%$ + +![](images/d49d38b56b44a2ac65002828b7cf48f21a8d1455cde24b5c02e63535e0b072dd.jpg) +Fig. 4: The edit discovery process of turning a wooden material into "marbled granite". Each column shows the hypotheses generated by $G$ , with the most promising candidates chosen by $V$ indicated by the highlights. Note that iteration 3 proved to be unfruitful according to $V$ , and the method reverts to the best candidate from iteration 2, before moving onto iteration 4. + +Table 1: CLIP scores of BlenderAlchemy vs. BlenderGPT for the text-based material editing task. We find that a version of our system that has no visual components (-Vision) still outperforms BlenderGPT. By adding vision to the state evaluator alone (-Vision G) and not the edit generator, a further improvement is observed. + +
MetricBlenderGPT-Vision-Vision GOurs
ViT-B/32 (↑)25.225.727.828.2
ViT-L/14 (↑)21.121.823.424.0
+ +of the time. We found that in $71\%$ of comparisons, human users also prefer BlenderAlchemy outputs selected by our visual evaluator. This suggests that our visual evaluator makes decisions that align with users' preferences. + +Image-based material-editing Given an image of a desired material, the task is to convert the code of the starter material into a material that contains many of the visual attributes of the input image, akin to doing a kind of style transfer for procedural materials. + +At each step, the edit generator is first asked to textually enumerate a list of obvious visual differences between the current material and the target, then asked to locate lines within the code that may be responsible for these visual differences (e.g. "the target material looks more rough" $\rightarrow$ "line 23 sets a roughness value, which we should try to increase") before finally suggested an edited version of the program. As such, low-level visual differences (e.g. color discrepancies) are semantically compressed first into language descriptions (e.g. "the target is more red"), before being fed into the editing process, resulting in the behavior that our system produces variations of the input material that resembles the target image along many attributes, even if its outputs don't perfectly match the target image (See Figure 6). Our system is the same as for text-based material editing, but without the need for visual imagination. + +![](images/06d62e1ee7f5e2a1764eae8e3ec639b9ba3f2c1bc624fba2e0c299ce30f1a657.jpg) +Fig.5: Comparisons between our method and BlenderGPT for the text-based material editing task setting. Note how our materials align better with the original language prompts (See Figure 3 for the original prompts). The input material being edited is the same wooden material. + +![](images/5beff7723f7e944e01a00ede3bb7c4fd8e30f653ccd477a57f7d854851964605.jpg) +Fig. 6: Material editing based on image inputs. Our edit intention is described by the target image shown on the right. 5 different Infinigen [32] initial materials are shown here, and the final edit. Note how in each case, certain attributes of the target material (metallicness, color, texture) are transferred. + +![](images/bf51817fbc787d52178671d2a1a31828fec0ccf5a815d83018407b31c6233c82.jpg) + +# 4.2 Geometry Editing + +BlenderAlchemy can also be used to manipulate geometry. We show qualitative examples of our system controlling geometry by programmatically 1) interpolating between preset blend shapes, 2) editing of geometry node graphs, and 3) the precise placement of objects within scenes. The blend weights, geometry node graphs and object placements can all be represented programmatically. + +As shown in Figure 7, given an input shape and pre-defined blend shapes, BlenderAlchemy can modify the shape by finding the appropriate blend weights between different shapes to match language, which is non-trivial for human users given the number of blend shapes to interpolate for complex characters and shapes. For instance, the facial expression editing example requires coordinated interpolation between 19 different blend shapes, each which controls a specific kind of motion in one part of the face (e.g. an left eye-brow raise). To aid it in + +![](images/d09363b83edf21f76638344c36f2a32c34d70614a5593b25a055cfca53295b2c.jpg) +Fig. 7: BlenderAlchemy editing geometry using blend shapes. Edits are made to match a description or a script line. Input shapes from BlenderKit. + +![](images/83d7ae0544c2666d144054ee01c7eab1da0865539f8420f1a95765ffc29cb5c1.jpg) +Fig. 8: BlenderAlchemy editing (a) the placement of assets within a living room scene and (b) the procedural geometry nodes of roses, to different text prompts. + +this task, our task description to BlenderAlchemy includes language labels for each blend shape. + +In Figure 8(a), BlenderAlchemy is presented with a scene layout editing task, where assets are initially dropped out of camera view and must be iteratively moved to fit the scene of a living room and the requirement that the couch is "facing the window", through calling a set of object placement helper functions. Figure 8(b) shows examples of BlenderAlchemy editing the procedural geometry of a set of roses to match a nuanced instruction that influences the number and angle of pedals, as well as the density of leaves. The full geometry node graph that BlenderAlchemy edits contains 60 geometry nodes. + +# 4.3 Lighting Setup Editing + +We show that BlenderAlchemy can be used to adjust the lighting of scenes according to language instructions as well. Figure 9 shows this being done for an input product visualization designed by a Blender artist. + +![](images/2e8abb6e51680947847c7737b985d5dcae14705620d7c4c37890adb712bdb510.jpg) +Fig. 9: Optimizing the lighting of the scene setup to the text-based user intention. We base the initial Blender state input based on a product visualization downloaded from BlenderKit. + +![](images/e7a06a2f18c2f37165809061063a128b9c87b057265ca3f698dab805a9e39b31.jpg) +"Hand lotion under disco lights, like at a nightclub." + +![](images/ef951f71b27d874796cd4f9aec7f8023f312fe0b1828564b9a75e6b6a4124f52.jpg) + +![](images/c6691351d34c882058d3000c6a7ba4712adf001b9a94ad303a0e61c70f6d1b93.jpg) + +![](images/9af7aff3497f4411e76040759dde004da5dcfd8da8c8fdebb550d9360f1180ee.jpg) +Fig. 10: Optimizing the lighting and material iteratively within a product visualization scene, to satisfy the text-based user intention. Note the dimming of the environment lighting for the nighttime lighting, and the glowing-hot material the editing procedure has produced. We base the initial Blender state input based on a product visualization downloaded from BlenderKit. + +![](images/28d334738c60eaebb3632557884ce3090d89ac71273dd029a0f2d6f49ad127ad.jpg) +"Hand cream sitting on a hot comet outside in the night" + +![](images/46a9fa8c4012ef34af8475d1b1deef355c382ecfe3c9f10a5e19207109cb2c1c.jpg) + +![](images/da92e80f5afdb9d9f48e3e88e4c8d9d6175080e297b59f5657d404b261c1fd8f.jpg) + +As mentioned in Section 3.1, we can consider iteratively optimizing two separate programs, one controlling the lighting of the whole scene and another controlling the material of an object within the scene. That is, $S_{\mathrm{init}} = F(\{p_0^{(L)}, p_0^{(M)}\}, S_{\mathrm{base}})$ for initial lighting program $p_0^{(L)}$ and material program $p_0^{(M)}$ . Figure 10 shows an example of this, where two separate pairs of edit generators and state evaluators, $(G_M, V_M)$ and $(G_L, V_L)$ , are used to achieve an edit to the scene that aligns with the user intention. The full algorithm is shown in the Appendix. + +# 5 Conclusion & Discussion + +In this paper, we introduce BlenderAlchemy, a system that performs edits within the Blender 3D design environment by leveraging vision-language models to iteratively refining a program to be more aligned with the user intention, by using visual information to both explore and prune possibilities within the program space. We equip our system with visual imagination by providing it access to text-to-image models, a tool it uses to guide itself towards program edits that better align with user intentions. We've demonstrated BlenderAlchemy on editing materials, geometry and lighting, and hope that future works will extend this to other workflows as well. + +# Acknowledgements + +We acknowledge the support of ARL grant W911NF-21-2-0104 and a Vannevar Bush Faculty Fellowship. We'd additionally like to thank Maneesh Agrawala for general discussions, and Purvi Goel, Mika Uy, Vishnu Sarukkai, Fan-yun Sun and Sharon Lee for feedback on paper revisions. + +# References + +1. Blendergt, https://github.com/gd3kr/BlenderGPT +2. How long does it take to create a 3d model?, https://3d-ace.com/blog/how-long-does-it-take-to-create-a-3d-model/ +3. How long does it take to make a 3d model?, https://pixune.com/blog/how-long-does-it-take-to-create-a-3d-model/ +4. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al.: Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 (2022) +5. Austin, J., Odena, A., Nye, M., Bosma, M., Michalewski, H., Dohan, D., Jiang, E., Cai, C., Terry, M., Le, Q., et al.: Program synthesis with large language models. arXiv preprint arXiv:2108.07732 (2021) +6. Baumli, K., Baveja, S., Behbahani, F., Chan, H., Comanici, G., Flennerhag, S., Gazeau, M., Holsheimer, K., Horgan, D., Laskin, M., et al.: Vision-language models as a source of rewards. arXiv preprint arXiv:2312.09187 (2023) +7. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opena.com/papers/dall-e-3.pdf 2(3), 8 (2023) +8. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023) +9. Chen, M., Tworek, J., Jun, H., Yuan, Q., de Oliveira Pinto, H.P., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F.P., Cummings, D., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W.H., Nichol, A., Paino, A., Tezak, N., Tang, J., Babuschkin, I., Balaji, S., Jain, S., Saunders, W., Hesse, C., Carr, A.N., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., Zaremba, W.: Evaluating large language models trained on code (2021) +10. Chen, Y., Chen, R., Lei, J., Zhang, Y., Jia, K.: Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems 35, 30923-30936 (2022) +11. De La Torre, F., Fang, C.M., Huang, H., Banburski-Fahey, A., Fernandez, J.A., Lanier, J.: Llmr: Real-time prompting of interactive worlds using large language models. arXiv preprint arXiv:2309.12276 (2023) +12. Firoozzi, R., Tucker, J., Tian, S., Majumdar, A., Sun, J., Liu, W., Zhu, Y., Song, S., Kapoor, A., Hausman, K., et al.: Foundation models in robotics: Applications, challenges, and the future. arXiv preprint arXiv:2312.07843 (2023) + +13. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., Wu, Y., Ji, R.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) +14. Fu, C., Zhang, R., Wang, Z., Huang, Y., Zhang, Z., Qiu, L., Ye, G., Shen, Y., Zhang, M., Chen, P., Zhao, S., Lin, S., Jiang, D., Yin, D., Gao, P., Li, K., Li, H., Sun, X.: A challenger to gpt-4v? early explorations of gemini in visual expertise. arXiv preprint arXiv:2312.12436 (2023) +15. Goel, P., Wang, K.C., Liu, C.K., Fatahalian, K.: Iterative motion editing with natural language. arXiv preprint arXiv:2312.11538 (2023) +16. Guerrero, P., Hasan, M., Sunkavalli, K., Meech, R., Boubekeur, T., Mitra, N.J.: Matformer: A generative model for procedural materials. arXiv preprint arXiv:2207.01044 (2022) +17. Henzler, P., Deschaintre, V., Mitra, N.J., Ritschel, T.: Generative modelling of brdf textures from flash images. arXiv preprint arXiv:2102.11861 (2021) +18. Hu, Y., Xie, Q., Jain, V., Francis, J., Patrikar, J., Keetha, N., Kim, S., Xie, Y., Zhang, T., Zhao, Z., et al.: Toward general-purpose robots via foundation models: A survey and meta-analysis. arXiv preprint arXiv:2312.08782 (2023) +19. Hu, Y., Guerrero, P., Hasan, M., Rushmeier, H., Deschaintre, V.: Node graph optimization using differentiable proxies. In: ACM SIGGRAPH 2022 conference proceedings. pp. 1-9 (2022) +20. Hu, Y., He, C., Deschaintre, V., Dorsey, J., Rushmeier, H.: An inverse procedural modeling pipeline for svbrdf maps. ACM Transactions on Graphics (TOG) 41(2), 1-17 (2022) +21. Huang, I., Krishna, V., Atekha, O., Guibas, L.: Aladdin: Zero-shot hallucination of stylized 3d assets from abstract scene descriptions. arXiv preprint arXiv:2306.06212 (2023) +22. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023) +23. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems 36 (2024) +24. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., Zeng, A.: Code as policies: Language model programs for embodied control. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9493-9500. IEEE (2023) +25. Liu, J., Gan, Y., Dong, J., Qi, L., Sun, X., Jian, M., Wang, L., Yu, H.: Perceptiondriven procedural texture generation from examples. Neurocomputing 291, 21-34 (2018) +26. Olausson, T.X., Inala, J.P., Wang, C., Gao, J., Solar-Lezama, A.: Is self-repair a silver bullet for code generation? In: The Twelfth International Conference on Learning Representations (2023) +27. OpenAI: Gpt-4 system card. OpenAI (2023), https://cdn.openai.com/papers/gpt-4-system-card.pdf +28. OpenAI: Gpt-4v(ison) system card. OpenAI (2023), https://apisemantic scholar.org/CorpusID:263218031 +29. Park, J.S., O'Brien, J., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. pp. 1-22 (2023) + +30. Patil, S.G., Zhang, T., Wang, X., Gonzalez, J.E.: Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334 (2023) +31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +32. Raistrick, A., Lipson, L., Ma, Z., Mei, L., Wang, M., Zuo, Y., Kayan, K., Wen, H., Han, B., Wang, Y., et al.: Infinite photorealistic worlds using procedural generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12630-12641 (2023) +33. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721 (2023) +34. Ritchie, D., Guerrero, P., Jones, R.K., Mitra, N.J., Schulz, A., Willis, K.D., Wu, J.: Neurosymbolic models for computer graphics. In: Computer Graphics Forum. vol. 42, pp. 545-568. Wiley Online Library (2023) +35. Romera-Paredes, B., Barekatain, M., Novikov, A., Balog, M., Kumar, M.P., Dupont, E., Ruiz, F.J., Ellenberg, J.S., Wang, P., Fawzi, O., et al.: Mathematical discoveries from program search with large language models. Nature 625(7995), 468-475 (2024) +36. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Hambro, E., Zettle-moyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems 36 (2024) +37. Sharma, P., Jampani, V., Li, Y., Jia, X., Lagun, D., Durand, F., Freeman, W.T., Matthews, M.: Alchemist: Parametric control of material properties with diffusion models. arXiv preprint arXiv:2312.02970 (2023) +38. Shi, L., Li, B., Hasan, M., Sunkavalli, K., Boubekeur, T., Mech, R., Matusik, W.: Match: Differentiable material graphs for procedural material capture. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020) +39. Shimizu, E., Fisher, M., Paris, S., McCann, J., Fatahalian, K.: Design adjectives: A framework for interactive model-guided exploration of parameterized design spaces. In: Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. pp. 261-278 (2020) +40. Shinn, N., Cassano, F., Labash, B., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Language agents with verbal reinforcement learning.(2023). arXiv preprint cs.AI/2303.11366 (2023) +41. Singh, I., Blukis, V., Mousavian, A., Goyal, A., Xu, D., Tremblay, J., Fox, D., Thomason, J., Garg, A.: Progprompt: Generating situated robot task plans using large language models. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11523-11530. IEEE (2023) +42. Sun, C., Han, J., Deng, W., Wang, X., Qin, Z., Gould, S.: 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945 (2023) +43. Tchapmi, L.P., Ray, T., Tchapmi, M., Shen, B., Martin-Martin, R., Savarese, S.: Generating procedural 3d materials from images using neural networks. In: 2022 4th International Conference on Image, Video and Signal Processing. pp. 32-40 (2022) +44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) + +45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +46. Vecchio, G., Martin, R., Roullier, A., Kaiser, A., Rouffet, R., Deschaintre, V., Boubekeur, T.: Controlmat: A controlled generative approach to material capture. arXiv preprint arXiv:2309.01700 (2023) +47. Vecchio, G., Sortino, R., Palazzo, S., Spampinato, C.: Matfuse: Controllable material generation with diffusion models. arXiv preprint arXiv:2308.11408 (2023) +48. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023) +49. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) +50. Wen, Z., Liu, Z., Sridhar, S., Fu, R.: Anyhome: Open-vocabulary generation of structured and textured 3d homes. arXiv preprint arXiv:2312.06644 (2023) +51. Wu, T., Yang, G., Li, Z., Zhang, K., Liu, Z., Guibas, L., Lin, D., Wetzstein, G.: Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. arXiv preprint arXiv:2401.04092 (2024) +52. Xiao, X., Liu, J., Wang, Z., Zhou, Y., Qi, Y., Cheng, Q., He, B., Jiang, S.: Robot learning in the era of foundation models: A survey. arXiv preprint arXiv:2311.14379 (2023) +53. Yamada, Y., Chandu, K., Lin, Y., Hessel, J., Yildirim, I., Choi, Y.: L3go: Language agents with chain-of-3d-thoughts for generating unconventional objects. arXiv preprint arXiv:2402.09052 (2024) +54. Yang, H., Chen, Y., Pan, Y., Yao, T., Chen, Z., Mei, T.: 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 6860-6868 (2023) +55. Yang, Y., Sun, F.Y., Weihs, L., VanderBilt, E., Herrasti, A., Han, W., Wu, J., Haber, N., Krishna, R., Liu, L., et al.: Holodeck: Language guided generation of 3d embodied ai environments. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024). vol. 30, pp. 20-25. IEEE/CVF (2024) +56. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023) +57. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems 36 (2024) +58. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023) +59. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2310.16045 (2023) +60. Zeng, X.: Paint3d: Paint anything 3d with lighting-less texture diffusion models. arXiv preprint arXiv:2312.13913 (2023) +61. Zhou, H., Yao, X., Meng, Y., Sun, S., BIng, Z., Huang, K., Knoll, A.: Language-conditioned learning for robotic manipulation: A survey. arXiv preprint arXiv:2312.10807 (2023) +62. Zsolnai-Fehér, K., Wonka, P., Wimmer, M.: Gaussian material synthesis. arXiv preprint arXiv:1804.08369 (2018) \ No newline at end of file diff --git a/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/images.zip b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5be3a185d4983e3fc5ad779f407addf9e30ad176 --- /dev/null +++ b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b933070b798e48e01ceabcfb7c70a2d0fc88173eaf1bd2b994ddf8d3a7337baf +size 490314 diff --git a/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/layout.json b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..12f5d9a972556bf01a4f1fc4fe67adf1412bb945 --- /dev/null +++ b/2024/BlenderAlchemy_ Editing 3D Graphics with Vision-Language Models/layout.json @@ -0,0 +1,9152 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 155, + 111, + 460, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 111, + 460, + 148 + ], + "spans": [ + { + "bbox": [ + 155, + 111, + 460, + 148 + ], + "type": "text", + "content": "BlenderAlchemy: Editing 3D Graphics with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 182, + 168, + 430, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 168, + 430, + 180 + ], + "spans": [ + { + "bbox": [ + 182, + 168, + 430, + 180 + ], + "type": "text", + "content": "Ian Huang, Guandao Yang, and Leonidas Guibas" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 265, + 190, + 347, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 190, + 347, + 201 + ], + "spans": [ + { + "bbox": [ + 265, + 190, + 347, + 201 + ], + "type": "text", + "content": "Stanford University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 233, + 452, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 233, + 452, + 441 + ], + "spans": [ + { + "bbox": [ + 160, + 233, + 452, + 441 + ], + "type": "text", + "content": "Abstract. Graphics design is important for various applications, including movie production and game design. To create a high-quality scene, designers usually need to spend hours in software like Blender, in which they might need to interleave and repeat operations, such as connecting material nodes, hundreds of times. Moreover, slightly different design goals may require completely different sequences, making automation difficult. In this paper, we propose a system that leverages Vision-Language Models (VLMs), like GPT-4V, to intelligently search the design action space to arrive at an answer that can satisfy a user's intent. Specifically, we design a vision-based edit generator and state evaluator to work together to find the correct sequence of actions to achieve the goal. Inspired by the role of visual imagination in the human design process, we supplement the visual reasoning capabilities of VLMs with \"imagined\" reference images from image-generation models, providing visual grounding of abstract language descriptions. In this paper, we provide empirical evidence suggesting our system can produce simple but tedious Blender editing sequences for tasks such as editing procedural materials and geometry from text and/or reference images, as well as adjusting lighting configurations for product renderings in complex scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 464, + 230, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 230, + 476 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 230, + 476 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 491, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 598 + ], + "type": "text", + "content": "To produce the compelling graphics content we see in movies or video games, 3D artists usually need to spend hours in software like Blender to find appropriate surface materials, object placements, and lighting arrangements. These operations require the artist to create a mental picture of the target, experiment with different parameters, and visually examine whether their edits get closer to the end goal. One can imagine automating these processes by converting language or visual descriptions of user intent into edits that achieve a design goal. Such a system can improve the productivity of millions of 3D designers and impact various industries that depend on 3D graphic design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 598, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 635 + ], + "type": "text", + "content": "Graphic design is very challenging because even a small design goal requires performing a variety of different tasks. For instance, modeling of a game environment requires the 3D artist to cycle between performing modeling, material" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "1 For project website and code, please go to: https://ianhuang0630.github.io/BlenderAlchemyWeb/" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 242 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 242 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 242 + ], + "type": "image", + "image_path": "b8d18be0552d5a1e9b290069f16b51083f6875aa4f199495c78b9900e7b893f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 321 + ], + "type": "text", + "content": "Fig. 1: Overview of BlenderAlchemy. Given an input Blender state and a user intention specified using either language or reference images, BlenderAlchemy edits the Blender state to satisfy that intention by iteratively refining a Blender python program that executes to produce the final Blender state. Our system additionally leverages text-to-image generation for visual imagination, a step that expands a text-only user intention to a concrete visual target to improve program refinement." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 345, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 477 + ], + "type": "text", + "content": "design, texture painting, animation, lighting, and scene composition. Prior attempts usually focus on specific editing tasks, like material synthesis [8, 62]. While these approaches show promising performance in the tasks they are designed for, it is non-trivial to combine these task-specific methods to satisfy an intended design goal. An alternative is to leverage Large Language Models (LLMs) [22, 27, 45] to digest user intent and suggest design actions by proposing intelligent combination of existing task-specific tools [36] or predicting edits to programs step-by-step [49, 57]. While LLMs have excellent abilities to understand user intentions and suggest sequences of actions to satisfy them, applying LLMs to graphical design remains challenging largely because language cannot capture the visual consequences of actions performed in software like Blender." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 481, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 554 + ], + "type": "text", + "content": "One promising alternative is to leverage vision language models (VLM), such as LLaVA [23], GPT-4V [28], Gemini [44], and DallE-3 [7]. These VLMs have shown to be highly capable of understanding detailed visual information [13,56, 58, 59] and generating compelling images [7]. We posit that these VLMs can be leveraged to complete different kinds of design tasks within the Blender design environment, like editing materials, geometry and lighting setups." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we provide a proof-of-concept system using the vision foundation model GPT-4V to generate and edit programs that modify the state of a Blender workspace to satisfy a user intention. Specifically, we first initialize the state of a workspace within Blender. The Blender state is parameterized as a short Python program, and a base Blender file. The user will then input a text description and potentially a reference image to communicate the desirable design outcome. The system is tasked to edit the program so that, when executed on the base Blender file, the rendered image can satisfy the user's intention. Figure 1 provides an illustration of the problem setup." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 479, + 193 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 479, + 193 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 479, + 193 + ], + "type": "image", + "image_path": "faf85e5061ad7daf08a18921d7ff18c60663be6d15cecfc1d1fdd6904cb83b90.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "text", + "content": "Fig. 2: Iterative visual program editing employs a edit generator " + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "text", + "content": " and a state evaluator " + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "text", + "content": " in each iteration to explore and prune different potential program edits, where " + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "text", + "content": " generates plausible variants of an input program and " + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "text", + "content": " picks between two programs based on the consequences they have to the Blender visual state and their alignment to the user intention. Each iteration of the refinement explores variations of the most promising program from the previous iteration. See Algorithm 1 for details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 632 + ], + "type": "text", + "content": "Naively applying VLMs to this editing setting gives rise to many failure cases, possibly due to the fact that out-of-the-box VLMs have a poor understanding of the visual consequences of Blender program edits. To counter this, we propose a visually-guided program search procedure that combines a vision-aware edit generator and a visual state evaluator to iteratively search for a suitable program edit (Figure 2). Inspired by the human design process, our system performs guided trial-and-error, capped by some computation budget. Within each iteration, the visual program generator will propose several possible edits on the current program. These edits will be applied and executed in Blender to produce a rendered image. These rendered images will be provided to visual evaluator, which will select the best renders via pairwise comparison by assessing which choice better satisfies the design goal specified by text and reference image. The program achieving the best render will replace the current program as a starting point for the visual program generator. This iterative search process, however, has very low success rate because of the sparsity of correct program edits in the vast program space. To improve the success rate, we further propose two techniques. First, when the proposed program of the new iteration does not contain a viable candidate, we revert to the best candidate of the prior program. This reversion mechanism makes sure the search procedure will not diverge when facing a batch of bad edit candidates. Second, to facilitate our visual evaluator and generator to better understand user intent, we leverage the \"visual imagination\" of text-to-image generative model to imagine a reference image. We show that our method is capable of accomplishing graphical design tasks within Blender, guided by user intention in the form of text and images. We demonstrate the effectiveness of our system on material, geometry and lighting design tasks, all parts of the 3D design process where artists spend a significant amount of time, ranging from 20 hrs to 4-6 workdays per model [2,3]. We show that our method can outperform prior works designed for similar problem settings, such as BlenderGPT [1]. In summary, the main contributions of this paper include:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 136, + 641, + 482, + 666 + ], + "type": "text", + "content": "1. We propose BlenderAlchemy, a system that's able to edit visual programs based on user input in the form of text or images." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlenderAlchemy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 135, + 116, + 481, + 188 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 135, + 116, + 481, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 116, + 481, + 151 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 481, + 151 + ], + "type": "text", + "content": "2. We identify key components that make the system work: a visual state evaluator, a visual edit generator, a search algorithm with an edit reversion mechanism, and a visual imagination module to facilitate the search." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 152, + 481, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 152, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 135, + 152, + 481, + 188 + ], + "type": "text", + "content": "3. We provide evidence showing that our system can outperform prior works in text-based procedural material editing, as well as its applicability to other design tasks including editing geometry and lighting configurations." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 206, + 239, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 239, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 239, + 218 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 220, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 509 + ], + "type": "text", + "content": "Task-specific tools for Material Design. Large bodies of works have been dedicated to using learning-based approaches to generate materials. Prior works exploit 2D diffusion models to generate texture maps either in a zero-shot way [8, 21, 33, 50, 54] or through fine-tuning [37, 60]. Though these works open the possibility of generating and modifying textures of objects using natural language for 3D meshes, these works do not model the material properties in a way that allows such objects to be relit. Other methods directly predict the physical properties of the surface of a material through learning, using diffusion [46, 47] or learning good latent representations from data [10, 17, 20, 25, 39, 62]. However, for all of the work mentioned so far, their fundamentally image-based or latent-based representations make the output materials difficult to edit in existing 3D graphics pipelines. People have also explored combining learning-based approaches with symbolic representations of materials [34, 43]. These works often involve creating differentiable representations of the procedural material graph often used in 3D graphical design pipelines, and backpropagating gradients throughout the graph to produce an image that can match the target [16, 19, 38]. Other works like Infinigen [32] use rule-based procedural generation over a library of procedural materials. However, no prior works in this direction have demonstrated the ability to edit procedural material graphs using user intention specified by language [34], a task that we are particularly interested in. Though the aforementioned works excel at material design, they aren't generalizable to other 3D graphics design task settings. BlenderAlchemy, on the other hand, aims to produce a system that can perform various design tasks according to user intents. This usually requires combining different methods together in a non-trivial way." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "LLM as general problem solvers. Large language models (LLMs) like GPT-4 [27,28], Llama [45], and Mistral [22] have in recent years demonstrated unprecedented results in a variety of problem settings, like robotics [4,12,18,24,41,52,61], program synthesis [26,35,40], and graphic design [21,50,55]. Other works have shown that by extending such models with an external process like Chain-of-Thought [49], Tree-of-Thought [57] or memory/skill database [11, 29, 30, 48], or by embedding such systems within environments where it can perceive and act [11, 29, 36, 48], a range of new problems that require iterative refinement can be solved. Their application to visual problem settings, however, has mostly been limited due to the nonexistent visual perception capabilities of the LLMs [11, 15, 48, 50, 53, 55]. While this could be sidestepped by fully condensing the visual state of the environment using text [29] or some symbolic representation [11,48], doing so for 3D graphic design works poorly. For instance, the task" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "content": "of editing a Blender material graph to create a desired material requires many trial-and-error cycles and an accurate understanding of the consequences certain design actions can have on the visual output. Recent works that apply LLMs to graphical design settings [11, 15] and ones that more specifically do so within Blender (like BlenderGPT [1], 3D-GPT [42] and L3GO [53]) do not use visual information to inform or refine their system's outputs, leading to unsatisfactory results. BlenderAlchemy borrows ideas from existing LLM literature and tries to address this issue by inputting visual perception into the system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 214, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 480, + 334 + ], + "type": "text", + "content": "Vision-Language Models. State-of-the-art vision language models, such as LLaVA [23], GPT-4V [28], and Gemini [44] have demonstrated impressive understanding of the visual world and its connections to language and semantics, enabling many computer vision tasks like scene understanding, visual question answering, and object detection to be one API call away [13,14,56,58,59]. Works such as [6,51] suggest that such models can also be used as a replacement for human evaluators for a lot of tasks, positioning them as tools for guiding planning and search by acting as flexible reward functions. BlenderAlchemy takes the first steps to apply VLMs to solve 3D graphic design tasks, a novel yet challenging application rather unexplored by existing works." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 354, + 202, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 202, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 202, + 366 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 479, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 479, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 479, + 451 + ], + "type": "text", + "content": "The goal of our system is to perform edits within the Blender 3D graphic design environment through iteratively refining programs that define a sequence of edits in Blender. This requires us to (1) decompose the input initial Blender input into a combination of programs and a \"base\" Blender state (Section 3.1) and (2) develop a procedure to edit each program to produce the desired visual state within Blender to match a user intention (Sections 3.2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 470, + 378, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 378, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 378, + 482 + ], + "type": "text", + "content": "3.1 Representation of the Blender Visual State" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "content": "The state of the initial Blender design environment can be decomposed into an \"base\" Blender state " + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{base}}" + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "content": " and a set of initial programs " + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "inline_equation", + "content": "\\{p_0^{(1)}, p_0^{(2)}, \\dots, p_0^{(k)}\\}" + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "content": " that acts on state " + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{base}}" + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "content": " to produce the initial Blender environment through a dynamics function " + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "content": " that transitions from one state to another based on a set of programmatic actions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 240, + 561, + 373, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 561, + 373, + 582 + ], + "spans": [ + { + "bbox": [ + 240, + 561, + 373, + 582 + ], + "type": "interline_equation", + "content": "S _ {\\mathrm {i n i t}} = F \\left(\\left\\{p _ {0} ^ {(i)} \\right\\} _ {i = 1 \\ldots k}, S _ {\\mathrm {b a s e}}\\right)", + "image_path": "a92294e4ab359c3e4c1ad82894ace44100547240345a72bbc14d29cb3fae4a87.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "content": "In our problem setting, " + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "content": " is the python code executor within the Blender environment that executes " + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\{p_0^{(i)}\\}_{i = 1\\dots k}" + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "content": " in sequence. We set each initial program " + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "inline_equation", + "content": "p_0^{(i)}" + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "content": " to be a program that concerns a single part of the 3D graphical design workflow - for instance, " + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "inline_equation", + "content": "p_0^{(1)}" + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "content": " is in charge of the material on one mesh within the scene, and " + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "inline_equation", + "content": "p_0^{(2)}" + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "content": " is in charge of the lighting setup of the entire scene. The decomposition of " + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{init}}" + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{base}}" + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "inline_equation", + "content": "p_0^1\\dots p_0^k" + }, + { + "bbox": [ + 130, + 586, + 480, + 667 + ], + "type": "text", + "content": " can be done using techniques like the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlenderAlchemy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": "“node transpler” from Infinigen [32], which converts entities within the Blender instance into lines of Python code that can recreate a node graph, like a material shader graph. We develop a suite of tools to do this in our own problem setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 154, + 482, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 154, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 482, + 239 + ], + "type": "text", + "content": "Though it's possible for all edits to be encompassed in a single program instead of " + }, + { + "bbox": [ + 130, + 154, + 482, + 239 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 154, + 482, + 239 + ], + "type": "text", + "content": " programs, this is limiting in practice - either because the VLM's output length isn't large enough for the code necessary or because the VLM has a low success rate, due to the program search space exploding in size. Although it is possible that future VLMs will substantially mitigate this problem, splitting the program into " + }, + { + "bbox": [ + 130, + 154, + 482, + 239 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 154, + 482, + 239 + ], + "type": "text", + "content": " task-specific programs may still be desirable, given the possibility of querying " + }, + { + "bbox": [ + 130, + 154, + 482, + 239 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 154, + 482, + 239 + ], + "type": "text", + "content": " task-specific fine-tuned/expert VLMs in parallel." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 264, + 420, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 264, + 420, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 264, + 420, + 276 + ], + "type": "text", + "content": "3.2 Iterative Refinement of Individual Visual Programs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": "Suppose that to complete a task like material editing for a single object, it suffices to decompose the initial state into " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{base}}" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": " and a single script, " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": " - that is, " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{init}} = F(\\{p_0\\}, S_{\\mathrm{base}})" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": ". Then our goal is to discover some edited version of " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": ", called " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "F(\\{p_1\\}, S_{\\mathrm{base}})" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": " produces a visual state better aligned with some user intention " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": ". Our system assumes that the user intention " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": " is provided in the form of language and/or image references, leveraging the visual understanding of the latest VLM models to understand user intention." + } + ] + } + ], + "index": 5 + }, + { + "type": "code", + "bbox": [ + 132, + 417, + 482, + 661 + ], + "blocks": [ + { + "bbox": [ + 132, + 401, + 376, + 415 + ], + "lines": [ + { + "bbox": [ + 132, + 401, + 376, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 376, + 415 + ], + "type": "text", + "content": "Algorithm 1 Iterative Refinement of Visual Programs" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "lines": [ + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": "1: procedure TOURNAMENT(State candidates " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\{S_1,S_2,\\dots S_k\\}" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " , Visual state evaluator " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " ,User intention " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " \n2: if " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "k > 2" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " then \n3: " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "w_{1}\\gets" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " TOURNAMENT( " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\{S_1,S_2,\\ldots S_{k / 2}\\}" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "V,I)" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "w_{2}\\gets" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " TOURNAMENT( " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\{S_{k / 2},S_{k + 1},\\ldots S_k\\}" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "V,I)" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " \n5: else \n6: " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "w_{1}\\gets S_{1},w_{2}\\gets S_{2}" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " \n7: end if \n8: return " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "V(w_{1},w_{2},I)" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " \n9: end procedure \n10: procedure REFINE(Depth " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " , Breadth " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " , Intention " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " , Edit Generator " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " , State \nEvaluator " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " , Base state " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{base}}" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " , Initial program " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " , Dynamics Function " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "S_0\\gets F(p_0,S_{\\mathrm{base}})" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{best}}\\gets S_0" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{best}}\\gets p_0" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " \n12: for " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "i\\gets 1" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " do \n13: " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = \\mathcal{N}(p_{\\mathrm{best}})" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " is odd else " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " \"Tweak\" or \"Leap\" edits \n14: " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "p_i^1,p_i^2\\dots p_i^b\\gets G(p_{\\mathrm{best}},S_{\\mathrm{best}},I,b,\\mathcal{P}_i)" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " Generate " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " options \n15: " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "S_i^1\\gets F(p_i^1,S_0),\\dots,S_i^b\\gets F(p_i^b,S_0)" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " Observe the visual states \n16: " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "S_i^*\\gets" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " TOURNAMENT( " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\{S_i^1,S_i^2\\dots S_i^b\\} \\cup \\{S_{\\mathrm{best}}\\}" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "V,I)" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " Choose the best \n17: " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{best}}\\gets S_i^*,p_{\\mathrm{best}}\\gets p_i^*" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " Best visual state and programs so far \n18: end for \n19: return " + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{best}},p_{\\mathrm{best}}" + }, + { + "bbox": [ + 132, + 417, + 482, + 661 + ], + "type": "text", + "content": " \n20: end procedure" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "algorithm" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "To discover a good edit to " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": ", we introduce the procedure outlined in Algorithm 1, an iterative refinement loop that repeatedly uses a visual state evaluator " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": " to select among the hypotheses from an edit generator " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": ". A single \"agent\" for a certain task like procedural material design can be fully described by " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "(G, V)" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "content": "Inspired by works like [51], we propose a visual state evaluator " + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "inline_equation", + "content": "V(S_{1}, S_{2}, I)" + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "content": ", which is tasked with returning whichever of the two visual states (" + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "content": ") better matches the user intention " + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "content": ". This evaluator is applied recursively to choose the most suitable visual state candidate among " + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "content": " visual state candidates by making " + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\log(b))" + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "content": " queries, as done in TOURNAMENT in Algorithm 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": "Though it seems straightforward to ask the same VLM to edit the code in a single pass, this leads to many failure cases (see more in the Appendix). Due to the VLM's lack of baked-in understanding of the visual consequences of various programs within Blender, a multi-hypothesis and multi-step approach is more appropriate. Extending Tree-of-Thoughts [57] to the visual domain, " + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "inline_equation", + "content": "G(p, S, I, b, \\mathcal{P})" + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": " is a module tasked with generating " + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": " different variations of program " + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": ", conditioned on the current visual state " + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": " and user intention " + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": ", constrained such that the output programs fall within some family of programs " + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": ", which can be used to instill useful priors to the edit generator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 392 + ], + "type": "text", + "content": "Below we describe some additional system design decisions that ensured better alignment of the resultant edited program to the user intention, either by improving the stability of the procedure or by supplementing the visual understanding of VLMs. The effect of each is investigated in the Appendix through ablations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 403, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 481, + 462 + ], + "type": "text", + "content": "Hypothesis Reversion. To improve the stability of the edit discovery process, we add the visual state of the program being edited at every timestep (" + }, + { + "bbox": [ + 131, + 403, + 481, + 462 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{best}}" + }, + { + "bbox": [ + 131, + 403, + 481, + 462 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 403, + 481, + 462 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{best}}" + }, + { + "bbox": [ + 131, + 403, + 481, + 462 + ], + "type": "text", + "content": ") as an additional candidate to the selection process, providing the option for the process to revert to an earlier version if the search at a single iteration was unsuccessful. Line 16 in Algorithm 1 shows this." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 474, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 474, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 474, + 481, + 665 + ], + "type": "text", + "content": "Tweak and Leap Edits. An important characteristic of visual programs is that continuous values hard-coded within the program can modify the output just as much as structural changes. This is in contrast to non-visual program synthesis tasks based on unit-tests of I/O specs, like [5, 9], where foundation models are mostly tasked to produce the right structure of the program with minimal hard-coded values. Given a single visual program " + }, + { + "bbox": [ + 131, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "p \\in \\mathcal{P}" + }, + { + "bbox": [ + 131, + 474, + 481, + 665 + ], + "type": "text", + "content": ", the space of visual outputs achievable through only tweaking the numerical values to function parameters and variable assignments can cover a wide range, depending on the fields available in the program. In Algorithm 1, we refer this space as the \"neighborhood\" of " + }, + { + "bbox": [ + 131, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 474, + 481, + 665 + ], + "type": "text", + "content": " or \"tweak\" edits, " + }, + { + "bbox": [ + 131, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p)" + }, + { + "bbox": [ + 131, + 474, + 481, + 665 + ], + "type": "text", + "content": ". Though this can result in a very small change to the program, this can lead to a large visual difference in the final output, and in a few edits can change the \"wrong\" program into one more aligned with the user intention. On the other hand, more drastic changes (or \"leap\" edits) may be needed to accomplish a task. Consider for example the task of changing a perfectly smooth material to have a noised level of roughness scattered sparsely across the material surface. This may require the programmatic addition of the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlenderAlchemy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "relevant nodes (e.g. color ramps or noise texture nodes), and thus the edited program falls outside of " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "Empirically, we find that the optimal edits are often a mix of tweak and leap edits. As such, our procedure cycles between restricting the edits of " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " to two different sets: the neighborhood of " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{best}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ", and the whole program space " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " (Line 13 in Algorithm 1). In practice, such restrictions are softly enforced through in-context prompting of VLMs, and though their inputs encourage them to abide by these constraints, the model can still produce more drastic \"tweak\" edits or conservative \"leap\" edits as needed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 346 + ], + "type": "text", + "content": "Visual Imagination. In the case when the user intention is communicated purely textually, their intention may be difficult for the VLM to turn into successful edits. Prior works like [21, 50, 55] have made similar observations of abstract language for 3D scenes. Consider, for example, the prompt \"make me a material that resembles a celestial nebula\". To do this, the VLM must know what a celestial nebula looks like, and how it should change the parameters of the material Shader nodes of, say, a wooden material. We find that in such cases, it's hard for the VLM to directly go from abstract descriptions to low-level program edits that affect low-level properties of the Blender visual state." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 347, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 479 + ], + "type": "text", + "content": "Instead, we propose supplementing the text-to-program understanding of VLM's with the text-to-image understanding in state-of-the-art image generation models. Intermediary visual artifacts (images generated using the user intention) are created and used to guide the refinement process towards a more plausible program edit to match the desired outcome, as shown in Figure 1. The generated images act as image references in addition to the textual intention provided by the user. This constitutes a simple visual chain of thought [49] for visual program editing, which not only creates a reference image for " + }, + { + "bbox": [ + 130, + 347, + 482, + 479 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 347, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 347, + 482, + 479 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 347, + 482, + 479 + ], + "type": "text", + "content": " to guide their low-level visual comparisons (e.g. color schemes, material roughness... etc.), but also provides a user-interpretable intermediary step to confirm the desired goals behind an otherwise vague user intention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 499, + 230, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 230, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 230, + 513 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 526, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 481, + 552 + ], + "type": "text", + "content": "We demonstrate BlenderAlchemy on editing procedural materials, geometry and lighting setups within Blender, three of the most tedious parts of 3D design." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 571, + 304, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 304, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 304, + 583 + ], + "type": "text", + "content": "4.1 Procedural Material Editing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Procedural material editing has characteristics that make it difficult for the same reason as a lot of other visual program settings: (1) small edit distances of programs may result in very large visual differences, contributing to potential instabilities in the edit discovery process, (2) it naturally requires trial-and-error when human users are editing them as the desired magnitude of edits depends on the visual output and (3) language descriptions of edit intent typically do not" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "contain low-level information for the editor to know immediately which part of the program to change." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": "We demonstrate this on two different kinds of editing tasks: (1) turning the same initial material (a synthetic wooden material) into other materials described by a list of language descriptions and (2) editing many different initial materials to resemble the same target material described by an image input. For the starting materials, we use the synthetic materials from Infinigen [32]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 481, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 481, + 324 + ], + "type": "text", + "content": "Text-based material-editing An attractive application of our system is in the modification of preexisting procedural materials using natural language descriptions that communicate user intent, a desirable but thus far undemonstrated capability of neurosymbolic methods [34]. We demonstrate our system's capabilities by asking it to edit a wooden material from Infinigen [32] into many other materials according to diverse language descriptions of target materials that are, importantly, not wood. In reality, this is a very challenging task, since this may require a wide range in the size of edits even if the language describing desired target material may be very similar." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "text", + "content": "Figure 3 shows examples of edits of the same starter wood material using different language descriptions, and Figure 4 demonstrates the intermediary steps of the problem solving process for a single instance of the problem. Our system is composed of an edit generator that generates 8 hypotheses per iteration, for 4 iterations (" + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "inline_equation", + "content": "d = 4" + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "inline_equation", + "content": "b = 8" + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "text", + "content": "), cycling between tweak and leap edits. It uses GPT-4V for edit generation and state evaluation, and DallE-3 for visual imagination." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 397, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 481, + 540 + ], + "type": "text", + "content": "We compare against BlenderGPT, the most recent open-sourced Blender AI agent that use GPT-4 to execute actions within the Blender environment through the Python API. We provide the same target material text prompt to BlenderGPT, as well as the starter code for the initial wood material for reference. We compare the CLIP similarity of their output material to the input text description against our system. BlenderGPT reasons only about how to edit the program using the input text description, doing so in a single pass without state evaluation or multi-hypothesis edit generation. To match the number of edit generator queries we make, we run their method a maximum of 32 times, using the first successful example as its final output. Everything else remains the same, including the starter material program, text description, base Blender state, and lighting setup." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 540, + 481, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 481, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 481, + 636 + ], + "type": "text", + "content": "We find that qualitatively, BlenderGPT produces much shallower and more simplistic edits of the input material, resulting in low-quality output materials and poor alignment with the user intention. Examples can be seen in Figure 5. For instance, observe that for the \"digital camouflage\" example, BlenderAlchemy is able to produce the \"sharper angles\" that the original description requests (See Figure 3) whereas BlenderGPT produces the right colors but fails to create the sharp, digital look. For \"metallic swirl\" example, our system's visual state selection process would weed out insufficiently swirly examples such as the one" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlenderAlchemy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 641, + 480, + 665 + ], + "type": "text", + "content": " At the time of publishing, works like 3DGPT [42], L3GO [53] have not yet open-sourced their code." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 481, + 422 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 481, + 422 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 481, + 422 + ], + "type": "image", + "image_path": "1b50a424840ecc234239ee56d36f7f5b8b0fc407ae700b8ca4e69a376061f120.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 423, + 482, + 458 + ], + "lines": [ + { + "bbox": [ + 132, + 423, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 482, + 458 + ], + "type": "text", + "content": "Fig. 3: Text-based Material Editing Results. The step-by-step edits of a 4x8 version of BlenderAlchemy to the same wooden material, given the text description on the left as the input user intention." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "text", + "content": "given by BlenderGPT, enabling our method to produce a material closer to the prompt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 495, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 579 + ], + "type": "text", + "content": "Table 1 demonstrates the comparison in terms of the average ViT-B/32 and ViT-L/14 CLIP similarity scores [31] with respect to the language description. Our system's ability to iteratively refine the edits based on multiple guesses at each step gives it the ability to make more substantive edits over the course of the process. Moreover, visual grounding provided both by the visual state evaluator as well as the output of the visual imagination stage guides the program search procedure to better align with the description. For ablations, see the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "We additionally conducted a user preference study to compare BlenderAlchemy's performance with two material generation baselines that use diffusion: TEXTure [33] and Paint3D [60]. We collect 592 Mechanical Turk comparisons between BlenderAlchemy and the baselines from 24 Turkers on materials created using 32 different text prompts. Users must choose the material that best matches the text-prompt. The results show that BlenderAlchemy is preferred to Paint3D " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " of the time. Users picked BlenderAlchemy over TEXTure " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "56\\%" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 247 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 247 + ], + "type": "image", + "image_path": "d49d38b56b44a2ac65002828b7cf48f21a8d1455cde24b5c02e63535e0b072dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 249, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 304 + ], + "type": "text", + "content": "Fig. 4: The edit discovery process of turning a wooden material into \"marbled granite\". Each column shows the hypotheses generated by " + }, + { + "bbox": [ + 130, + 249, + 482, + 304 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 249, + 482, + 304 + ], + "type": "text", + "content": ", with the most promising candidates chosen by " + }, + { + "bbox": [ + 130, + 249, + 482, + 304 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 249, + 482, + 304 + ], + "type": "text", + "content": " indicated by the highlights. Note that iteration 3 proved to be unfruitful according to " + }, + { + "bbox": [ + 130, + 249, + 482, + 304 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 249, + 482, + 304 + ], + "type": "text", + "content": ", and the method reverts to the best candidate from iteration 2, before moving onto iteration 4." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 183, + 371, + 428, + 418 + ], + "blocks": [ + { + "bbox": [ + 130, + 317, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 361 + ], + "type": "text", + "content": "Table 1: CLIP scores of BlenderAlchemy vs. BlenderGPT for the text-based material editing task. We find that a version of our system that has no visual components (-Vision) still outperforms BlenderGPT. By adding vision to the state evaluator alone (-Vision G) and not the edit generator, a further improvement is observed." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 371, + 428, + 418 + ], + "lines": [ + { + "bbox": [ + 183, + 371, + 428, + 418 + ], + "spans": [ + { + "bbox": [ + 183, + 371, + 428, + 418 + ], + "type": "table", + "html": "
MetricBlenderGPT-Vision-Vision GOurs
ViT-B/32 (↑)25.225.727.828.2
ViT-L/14 (↑)21.121.823.424.0
", + "image_path": "952e25b1509dd0f2f4354925c57068ec8b4f48cf090cbe7f38d2c1b270da82a0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "text", + "content": "of the time. We found that in " + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "text", + "content": " of comparisons, human users also prefer BlenderAlchemy outputs selected by our visual evaluator. This suggests that our visual evaluator makes decisions that align with users' preferences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 481, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 528 + ], + "type": "text", + "content": "Image-based material-editing Given an image of a desired material, the task is to convert the code of the starter material into a material that contains many of the visual attributes of the input image, akin to doing a kind of style transfer for procedural materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 529, + 482, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 673 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 673 + ], + "type": "text", + "content": "At each step, the edit generator is first asked to textually enumerate a list of obvious visual differences between the current material and the target, then asked to locate lines within the code that may be responsible for these visual differences (e.g. \"the target material looks more rough\" " + }, + { + "bbox": [ + 130, + 529, + 482, + 673 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 529, + 482, + 673 + ], + "type": "text", + "content": " \"line 23 sets a roughness value, which we should try to increase\") before finally suggested an edited version of the program. As such, low-level visual differences (e.g. color discrepancies) are semantically compressed first into language descriptions (e.g. \"the target is more red\"), before being fed into the editing process, resulting in the behavior that our system produces variations of the input material that resembles the target image along many attributes, even if its outputs don't perfectly match the target image (See Figure 6). Our system is the same as for text-based material editing, but without the need for visual imagination." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlenderAlchemy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 269 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 269 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 269 + ], + "type": "image", + "image_path": "06d62e1ee7f5e2a1764eae8e3ec639b9ba3f2c1bc624fba2e0c299ce30f1a657.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "type": "text", + "content": "Fig.5: Comparisons between our method and BlenderGPT for the text-based material editing task setting. Note how our materials align better with the original language prompts (See Figure 3 for the original prompts). The input material being edited is the same wooden material." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 319, + 394, + 437 + ], + "blocks": [ + { + "bbox": [ + 133, + 319, + 394, + 437 + ], + "lines": [ + { + "bbox": [ + 133, + 319, + 394, + 437 + ], + "spans": [ + { + "bbox": [ + 133, + 319, + 394, + 437 + ], + "type": "image", + "image_path": "5beff7723f7e944e01a00ede3bb7c4fd8e30f653ccd477a57f7d854851964605.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 439, + 482, + 483 + ], + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 483 + ], + "type": "text", + "content": "Fig. 6: Material editing based on image inputs. Our edit intention is described by the target image shown on the right. 5 different Infinigen [32] initial materials are shown here, and the final edit. Note how in each case, certain attributes of the target material (metallicness, color, texture) are transferred." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 335, + 474, + 423 + ], + "blocks": [ + { + "bbox": [ + 399, + 335, + 474, + 423 + ], + "lines": [ + { + "bbox": [ + 399, + 335, + 474, + 423 + ], + "spans": [ + { + "bbox": [ + 399, + 335, + 474, + 423 + ], + "type": "image", + "image_path": "bf51817fbc787d52178671d2a1a31828fec0ccf5a815d83018407b31c6233c82.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 495, + 253, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 253, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 253, + 508 + ], + "type": "text", + "content": "4.2 Geometry Editing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": "BlenderAlchemy can also be used to manipulate geometry. We show qualitative examples of our system controlling geometry by programmatically 1) interpolating between preset blend shapes, 2) editing of geometry node graphs, and 3) the precise placement of objects within scenes. The blend weights, geometry node graphs and object placements can all be represented programmatically." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "As shown in Figure 7, given an input shape and pre-defined blend shapes, BlenderAlchemy can modify the shape by finding the appropriate blend weights between different shapes to match language, which is non-trivial for human users given the number of blend shapes to interpolate for complex characters and shapes. For instance, the facial expression editing example requires coordinated interpolation between 19 different blend shapes, each which controls a specific kind of motion in one part of the face (e.g. an left eye-brow raise). To aid it in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 281 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 281 + ], + "type": "image", + "image_path": "d09363b83edf21f76638344c36f2a32c34d70614a5593b25a055cfca53295b2c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 285, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 308 + ], + "type": "text", + "content": "Fig. 7: BlenderAlchemy editing geometry using blend shapes. Edits are made to match a description or a script line. Input shapes from BlenderKit." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 312, + 482, + 429 + ], + "blocks": [ + { + "bbox": [ + 133, + 312, + 482, + 429 + ], + "lines": [ + { + "bbox": [ + 133, + 312, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 133, + 312, + 482, + 429 + ], + "type": "image", + "image_path": "83d7ae0544c2666d144054ee01c7eab1da0865539f8420f1a95765ffc29cb5c1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 434, + 480, + 456 + ], + "lines": [ + { + "bbox": [ + 132, + 434, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 480, + 456 + ], + "type": "text", + "content": "Fig. 8: BlenderAlchemy editing (a) the placement of assets within a living room scene and (b) the procedural geometry nodes of roses, to different text prompts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 468, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 480, + 491 + ], + "type": "text", + "content": "this task, our task description to BlenderAlchemy includes language labels for each blend shape." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 493, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 588 + ], + "type": "text", + "content": "In Figure 8(a), BlenderAlchemy is presented with a scene layout editing task, where assets are initially dropped out of camera view and must be iteratively moved to fit the scene of a living room and the requirement that the couch is \"facing the window\", through calling a set of object placement helper functions. Figure 8(b) shows examples of BlenderAlchemy editing the procedural geometry of a set of roses to match a nuanced instruction that influences the number and angle of pedals, as well as the density of leaves. The full geometry node graph that BlenderAlchemy edits contains 60 geometry nodes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 605, + 276, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 276, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 276, + 618 + ], + "type": "text", + "content": "4.3 Lighting Setup Editing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We show that BlenderAlchemy can be used to adjust the lighting of scenes according to language instructions as well. Figure 9 shows this being done for an input product visualization designed by a Blender artist." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlenderAlchemy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 130, + 219, + 223 + ], + "blocks": [ + { + "bbox": [ + 135, + 130, + 219, + 223 + ], + "lines": [ + { + "bbox": [ + 135, + 130, + 219, + 223 + ], + "spans": [ + { + "bbox": [ + 135, + 130, + 219, + 223 + ], + "type": "image", + "image_path": "2e8abb6e51680947847c7737b985d5dcae14705620d7c4c37890adb712bdb510.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 225, + 479, + 257 + ], + "lines": [ + { + "bbox": [ + 132, + 225, + 479, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 479, + 257 + ], + "type": "text", + "content": "Fig. 9: Optimizing the lighting of the scene setup to the text-based user intention. We base the initial Blender state input based on a product visualization downloaded from BlenderKit." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 130, + 304, + 223 + ], + "blocks": [ + { + "bbox": [ + 206, + 116, + 403, + 126 + ], + "lines": [ + { + "bbox": [ + 206, + 116, + 403, + 126 + ], + "spans": [ + { + "bbox": [ + 206, + 116, + 403, + 126 + ], + "type": "text", + "content": "\"Hand lotion under disco lights, like at a nightclub.\"" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 130, + 304, + 223 + ], + "lines": [ + { + "bbox": [ + 220, + 130, + 304, + 223 + ], + "spans": [ + { + "bbox": [ + 220, + 130, + 304, + 223 + ], + "type": "image", + "image_path": "e7a06a2f18c2f37165809061063a128b9c87b057265ca3f698dab805a9e39b31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 132, + 392, + 222 + ], + "blocks": [ + { + "bbox": [ + 306, + 132, + 392, + 222 + ], + "lines": [ + { + "bbox": [ + 306, + 132, + 392, + 222 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 392, + 222 + ], + "type": "image", + "image_path": "ef951f71b27d874796cd4f9aec7f8023f312fe0b1828564b9a75e6b6a4124f52.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 392, + 133, + 479, + 222 + ], + "blocks": [ + { + "bbox": [ + 392, + 133, + 479, + 222 + ], + "lines": [ + { + "bbox": [ + 392, + 133, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 392, + 133, + 479, + 222 + ], + "type": "image", + "image_path": "c6691351d34c882058d3000c6a7ba4712adf001b9a94ad303a0e61c70f6d1b93.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 135, + 275, + 219, + 369 + ], + "blocks": [ + { + "bbox": [ + 135, + 275, + 219, + 369 + ], + "lines": [ + { + "bbox": [ + 135, + 275, + 219, + 369 + ], + "spans": [ + { + "bbox": [ + 135, + 275, + 219, + 369 + ], + "type": "image", + "image_path": "9af7aff3497f4411e76040759dde004da5dcfd8da8c8fdebb550d9360f1180ee.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 372, + 480, + 426 + ], + "lines": [ + { + "bbox": [ + 130, + 372, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 480, + 426 + ], + "type": "text", + "content": "Fig. 10: Optimizing the lighting and material iteratively within a product visualization scene, to satisfy the text-based user intention. Note the dimming of the environment lighting for the nighttime lighting, and the glowing-hot material the editing procedure has produced. We base the initial Blender state input based on a product visualization downloaded from BlenderKit." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 219, + 276, + 303, + 369 + ], + "blocks": [ + { + "bbox": [ + 198, + 261, + 412, + 271 + ], + "lines": [ + { + "bbox": [ + 198, + 261, + 412, + 271 + ], + "spans": [ + { + "bbox": [ + 198, + 261, + 412, + 271 + ], + "type": "text", + "content": "\"Hand cream sitting on a hot comet outside in the night\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 276, + 303, + 369 + ], + "lines": [ + { + "bbox": [ + 219, + 276, + 303, + 369 + ], + "spans": [ + { + "bbox": [ + 219, + 276, + 303, + 369 + ], + "type": "image", + "image_path": "28d334738c60eaebb3632557884ce3090d89ac71273dd029a0f2d6f49ad127ad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 304, + 276, + 375, + 369 + ], + "blocks": [ + { + "bbox": [ + 304, + 276, + 375, + 369 + ], + "lines": [ + { + "bbox": [ + 304, + 276, + 375, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 375, + 369 + ], + "type": "image", + "image_path": "46a9fa8c4012ef34af8475d1b1deef355c382ecfe3c9f10a5e19207109cb2c1c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 376, + 277, + 478, + 369 + ], + "blocks": [ + { + "bbox": [ + 376, + 277, + 478, + 369 + ], + "lines": [ + { + "bbox": [ + 376, + 277, + 478, + 369 + ], + "spans": [ + { + "bbox": [ + 376, + 277, + 478, + 369 + ], + "type": "image", + "image_path": "da92e80f5afdb9d9f48e3e88e4c8d9d6175080e297b59f5657d404b261c1fd8f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "text", + "content": "As mentioned in Section 3.1, we can consider iteratively optimizing two separate programs, one controlling the lighting of the whole scene and another controlling the material of an object within the scene. That is, " + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{init}} = F(\\{p_0^{(L)}, p_0^{(M)}\\}, S_{\\mathrm{base}})" + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "text", + "content": " for initial lighting program " + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "inline_equation", + "content": "p_0^{(L)}" + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "text", + "content": " and material program " + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "inline_equation", + "content": "p_0^{(M)}" + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "text", + "content": ". Figure 10 shows an example of this, where two separate pairs of edit generators and state evaluators, " + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "inline_equation", + "content": "(G_M, V_M)" + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "inline_equation", + "content": "(G_L, V_L)" + }, + { + "bbox": [ + 130, + 437, + 482, + 536 + ], + "type": "text", + "content": ", are used to achieve an edit to the scene that aligns with the user intention. The full algorithm is shown in the Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 300, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 300, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 300, + 555 + ], + "type": "text", + "content": "5 Conclusion & Discussion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 563, + 482, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 670 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 670 + ], + "type": "text", + "content": "In this paper, we introduce BlenderAlchemy, a system that performs edits within the Blender 3D design environment by leveraging vision-language models to iteratively refining a program to be more aligned with the user intention, by using visual information to both explore and prune possibilities within the program space. We equip our system with visual imagination by providing it access to text-to-image models, a tool it uses to guide itself towards program edits that better align with user intentions. We've demonstrated BlenderAlchemy on editing materials, geometry and lighting, and hope that future works will extend this to other workflows as well." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "content": "We acknowledge the support of ARL grant W911NF-21-2-0104 and a Vannevar Bush Faculty Fellowship. We'd additionally like to thank Maneesh Agrawala for general discussions, and Purvi Goel, Mika Uy, Vishnu Sarukkai, Fan-yun Sun and Sharon Lee for feedback on paper revisions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 198, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 198, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 198, + 221 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 235, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 235, + 367, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 367, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 367, + 246 + ], + "type": "text", + "content": "1. Blendergt, https://github.com/gd3kr/BlenderGPT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 247, + 479, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 247, + 479, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 479, + 267 + ], + "type": "text", + "content": "2. How long does it take to create a 3d model?, https://3d-ace.com/blog/how-long-does-it-take-to-create-a-3d-model/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 269, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 290 + ], + "type": "text", + "content": "3. How long does it take to make a 3d model?, https://pixune.com/blog/how-long-does-it-take-to-create-a-3d-model/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 291, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 323 + ], + "type": "text", + "content": "4. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al.: Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 324, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 480, + 357 + ], + "type": "text", + "content": "5. Austin, J., Odena, A., Nye, M., Bosma, M., Michalewski, H., Dohan, D., Jiang, E., Cai, C., Terry, M., Le, Q., et al.: Program synthesis with large language models. arXiv preprint arXiv:2108.07732 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 389 + ], + "type": "text", + "content": "6. Baumli, K., Baveja, S., Behbahani, F., Chan, H., Comanici, G., Flennerhag, S., Gazeau, M., Holsheimer, K., Horgan, D., Laskin, M., et al.: Vision-language models as a source of rewards. arXiv preprint arXiv:2312.09187 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 390, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 423 + ], + "type": "text", + "content": "7. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opena.com/papers/dall-e-3.pdf 2(3), 8 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "text", + "content": "8. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 456, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 566 + ], + "type": "text", + "content": "9. Chen, M., Tworek, J., Jun, H., Yuan, Q., de Oliveira Pinto, H.P., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F.P., Cummings, D., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W.H., Nichol, A., Paino, A., Tezak, N., Tang, J., Babuschkin, I., Balaji, S., Jain, S., Saunders, W., Hesse, C., Carr, A.N., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., Zaremba, W.: Evaluating large language models trained on code (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "text", + "content": "10. Chen, Y., Chen, R., Lei, J., Zhang, Y., Jia, K.: Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems 35, 30923-30936 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "text", + "content": "11. De La Torre, F., Fang, C.M., Huang, H., Banburski-Fahey, A., Fernandez, J.A., Lanier, J.: Llmr: Real-time prompting of interactive worlds using large language models. arXiv preprint arXiv:2309.12276 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "12. Firoozzi, R., Tucker, J., Tian, S., Majumdar, A., Sun, J., Liu, W., Zhu, Y., Song, S., Kapoor, A., Hausman, K., et al.: Foundation models in robotics: Applications, challenges, and the future. arXiv preprint arXiv:2312.07843 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlenderAlchemy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., Wu, Y., Ji, R.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "type": "text", + "content": "14. Fu, C., Zhang, R., Wang, Z., Huang, Y., Zhang, Z., Qiu, L., Ye, G., Shen, Y., Zhang, M., Chen, P., Zhao, S., Lin, S., Jiang, D., Yin, D., Gao, P., Li, K., Li, H., Sun, X.: A challenger to gpt-4v? early explorations of gemini in visual expertise. arXiv preprint arXiv:2312.12436 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "text", + "content": "15. Goel, P., Wang, K.C., Liu, C.K., Fatahalian, K.: Iterative motion editing with natural language. arXiv preprint arXiv:2312.11538 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "text", + "content": "16. Guerrero, P., Hasan, M., Sunkavalli, K., Meech, R., Boubekeur, T., Mitra, N.J.: Matformer: A generative model for procedural materials. arXiv preprint arXiv:2207.01044 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 481, + 267 + ], + "type": "text", + "content": "17. Henzler, P., Deschaintre, V., Mitra, N.J., Ritschel, T.: Generative modelling of brdf textures from flash images. arXiv preprint arXiv:2102.11861 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "type": "text", + "content": "18. Hu, Y., Xie, Q., Jain, V., Francis, J., Patrikar, J., Keetha, N., Kim, S., Xie, Y., Zhang, T., Zhao, Z., et al.: Toward general-purpose robots via foundation models: A survey and meta-analysis. arXiv preprint arXiv:2312.08782 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 481, + 332 + ], + "type": "text", + "content": "19. Hu, Y., Guerrero, P., Hasan, M., Rushmeier, H., Deschaintre, V.: Node graph optimization using differentiable proxies. In: ACM SIGGRAPH 2022 conference proceedings. pp. 1-9 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 481, + 364 + ], + "type": "text", + "content": "20. Hu, Y., He, C., Deschaintre, V., Dorsey, J., Rushmeier, H.: An inverse procedural modeling pipeline for svbrdf maps. ACM Transactions on Graphics (TOG) 41(2), 1-17 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "text", + "content": "21. Huang, I., Krishna, V., Atekha, O., Guibas, L.: Aladdin: Zero-shot hallucination of stylized 3d assets from abstract scene descriptions. arXiv preprint arXiv:2306.06212 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "text", + "content": "22. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "type": "text", + "content": "23. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 471, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 515 + ], + "type": "text", + "content": "24. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., Zeng, A.: Code as policies: Language model programs for embodied control. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9493-9500. IEEE (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 515, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 548 + ], + "type": "text", + "content": "25. Liu, J., Gan, Y., Dong, J., Qi, L., Sun, X., Jian, M., Wang, L., Yu, H.: Perceptiondriven procedural texture generation from examples. Neurocomputing 291, 21-34 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 548, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 481, + 579 + ], + "type": "text", + "content": "26. Olausson, T.X., Inala, J.P., Wang, C., Gao, J., Solar-Lezama, A.: Is self-repair a silver bullet for code generation? In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "text", + "content": "27. OpenAI: Gpt-4 system card. OpenAI (2023), https://cdn.openai.com/papers/gpt-4-system-card.pdf" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "text", + "content": "28. OpenAI: Gpt-4v(ison) system card. OpenAI (2023), https://apisemantic scholar.org/CorpusID:263218031" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "29. Park, J.S., O'Brien, J., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. pp. 1-22 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "text", + "content": "30. Patil, S.G., Zhang, T., Wang, X., Gonzalez, J.E.: Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 183 + ], + "type": "text", + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 228 + ], + "type": "text", + "content": "32. Raistrick, A., Lipson, L., Ma, Z., Mei, L., Wang, M., Zuo, Y., Kayan, K., Wen, H., Han, B., Wang, Y., et al.: Infinite photorealistic worlds using procedural generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12630-12641 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 482, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 251 + ], + "type": "text", + "content": "33. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "text", + "content": "34. Ritchie, D., Guerrero, P., Jones, R.K., Mitra, N.J., Schulz, A., Willis, K.D., Wu, J.: Neurosymbolic models for computer graphics. In: Computer Graphics Forum. vol. 42, pp. 545-568. Wiley Online Library (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 285, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 329 + ], + "type": "text", + "content": "35. Romera-Paredes, B., Barekatain, M., Novikov, A., Balog, M., Kumar, M.P., Dupont, E., Ruiz, F.J., Ellenberg, J.S., Wang, P., Fawzi, O., et al.: Mathematical discoveries from program search with large language models. Nature 625(7995), 468-475 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 482, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 373 + ], + "type": "text", + "content": "36. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Hambro, E., Zettle-moyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "type": "text", + "content": "37. Sharma, P., Jampani, V., Li, Y., Jia, X., Lagun, D., Durand, F., Freeman, W.T., Matthews, M.: Alchemist: Parametric control of material properties with diffusion models. arXiv preprint arXiv:2312.02970 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 408, + 482, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 441 + ], + "type": "text", + "content": "38. Shi, L., Li, B., Hasan, M., Sunkavalli, K., Boubekeur, T., Mech, R., Matusik, W.: Match: Differentiable material graphs for procedural material capture. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 442, + 482, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 485 + ], + "type": "text", + "content": "39. Shimizu, E., Fisher, M., Paris, S., McCann, J., Fatahalian, K.: Design adjectives: A framework for interactive model-guided exploration of parameterized design spaces. In: Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. pp. 261-278 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 487, + 482, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 518 + ], + "type": "text", + "content": "40. Shinn, N., Cassano, F., Labash, B., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Language agents with verbal reinforcement learning.(2023). arXiv preprint cs.AI/2303.11366 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 520, + 482, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 564 + ], + "type": "text", + "content": "41. Singh, I., Blukis, V., Mousavian, A., Goyal, A., Xu, D., Tremblay, J., Fox, D., Thomason, J., Garg, A.: Progprompt: Generating situated robot task plans using large language models. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11523-11530. IEEE (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 565, + 482, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 482, + 587 + ], + "type": "text", + "content": "42. Sun, C., Han, J., Deng, W., Wang, X., Qin, Z., Gould, S.: 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "text", + "content": "43. Tchapmi, L.P., Ray, T., Tchapmi, M., Shen, B., Martin-Martin, R., Savarese, S.: Generating procedural 3d materials from images using neural networks. In: 2022 4th International Conference on Image, Video and Signal Processing. pp. 32-40 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "text", + "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlenderAlchemy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 658 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "46. Vecchio, G., Martin, R., Roullier, A., Kaiser, A., Rouffet, R., Deschaintre, V., Boubekeur, T.: Controlmat: A controlled generative approach to material capture. arXiv preprint arXiv:2309.01700 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "text", + "content": "47. Vecchio, G., Sortino, R., Palazzo, S., Spampinato, C.: Matfuse: Controllable material generation with diffusion models. arXiv preprint arXiv:2308.11408 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "type": "text", + "content": "48. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "type": "text", + "content": "49. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "50. Wen, Z., Liu, Z., Sridhar, S., Fu, R.: Anyhome: Open-vocabulary generation of structured and textured 3d homes. arXiv preprint arXiv:2312.06644 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "text", + "content": "51. Wu, T., Yang, G., Li, Z., Zhang, K., Liu, Z., Guibas, L., Lin, D., Wetzstein, G.: Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. arXiv preprint arXiv:2401.04092 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "text", + "content": "52. Xiao, X., Liu, J., Wang, Z., Zhou, Y., Qi, Y., Cheng, Q., He, B., Jiang, S.: Robot learning in the era of foundation models: A survey. arXiv preprint arXiv:2311.14379 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 387 + ], + "type": "text", + "content": "53. Yamada, Y., Chandu, K., Lin, Y., Hessel, J., Yildirim, I., Choi, Y.: L3go: Language agents with chain-of-3d-thoughts for generating unconventional objects. arXiv preprint arXiv:2402.09052 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "type": "text", + "content": "54. Yang, H., Chen, Y., Pan, Y., Yao, T., Chen, Z., Mei, T.: 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 6860-6868 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 463 + ], + "type": "text", + "content": "55. Yang, Y., Sun, F.Y., Weihs, L., VanderBilt, E., Herrasti, A., Han, W., Wu, J., Haber, N., Krishna, R., Liu, L., et al.: Holodeck: Language guided generation of 3d embodied ai environments. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024). vol. 30, pp. 20-25. IEEE/CVF (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 463, + 480, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 480, + 495 + ], + "type": "text", + "content": "56. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 495, + 480, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 480, + 528 + ], + "type": "text", + "content": "57. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 528, + 480, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 480, + 549 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 480, + 549 + ], + "type": "text", + "content": "58. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 549, + 480, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 480, + 582 + ], + "type": "text", + "content": "59. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2310.16045 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 582, + 480, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 480, + 603 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 480, + 603 + ], + "type": "text", + "content": "60. Zeng, X.: Paint3d: Paint anything 3d with lighting-less texture diffusion models. arXiv preprint arXiv:2312.13913 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 603, + 480, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 480, + 635 + ], + "type": "text", + "content": "61. Zhou, H., Yao, X., Meng, Y., Sun, S., BIng, Z., Huang, K., Knoll, A.: Language-conditioned learning for robotic manipulation: A survey. arXiv preprint arXiv:2312.10807 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 635, + 480, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 635, + 480, + 658 + ], + "spans": [ + { + "bbox": [ + 132, + 635, + 480, + 658 + ], + "type": "text", + "content": "62. Zsolnai-Fehér, K., Wonka, P., Wimmer, M.: Gaussian material synthesis. arXiv preprint arXiv:1804.08369 (2018)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_content_list.json b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fff1334ea8e116d16c2675434a149851bde2bd77 --- /dev/null +++ b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_content_list.json @@ -0,0 +1,2205 @@ +[ + { + "type": "text", + "text": "Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiangtao Zhang $^{1}$ , Zongsheng Yue $^{2}$ , Hui Wang $^{1}$ , Qian Zhao $^{1\\boxtimes}$ , and Deyu Meng $^{1,3}$", + "bbox": [ + 323, + 210, + 676, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xi'an Jiaotong University, Shaanxi, China", + "bbox": [ + 351, + 253, + 648, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ S-Lab, Nanyang Technological University, Singapore", + "bbox": [ + 318, + 268, + 681, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Pazhou Laboratory (Huangpu), Guangdong, China \nzhangjt2021@stu.xjtu.edu.cn, zsyzam@gmail.com,", + "bbox": [ + 321, + 282, + 676, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "huiwang2019@stu.xjtu.edu.cn, {timmy.zhaoqian, dymeng}@xjtu.edu.cn", + "bbox": [ + 250, + 311, + 751, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Blind image deconvolution (BID) is a classic yet challenging problem in the field of image processing. Recent advances in deep image prior (DIP) have motivated a series of DIP-based approaches, demonstrating remarkable success in BID. However, due to the high nonconvexity of the inherent optimization process, these methods are notorious for their sensitivity to the initialized kernel. To alleviate this issue and further improve their performance, we propose a new framework for BID that better considers the prior modeling and the initialization for blur kernels, leveraging a deep generative model. The proposed approach pre-trains a generative adversarial network-based kernel generator that aptly characterizes the kernel priors and a kernel initializer that facilitates a well-informed initialization for the blur kernel through latent space encoding. With the pre-trained kernel generator and initializer, one can obtain a high-quality initialization of the blur kernel, and enable optimization within a compact latent kernel manifold. Such a framework results in an evident performance improvement over existing DIP-based BID methods. Extensive experiments on different datasets demonstrate the effectiveness of the proposed method.", + "bbox": [ + 259, + 351, + 743, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Generative kernel prior $\\cdot$ Kernel initializer $\\cdot$ DIP", + "bbox": [ + 261, + 613, + 673, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 648, + 375, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Blind image deconvolution (BID), also known as deblurring, is a classical problem in image processing [8-10, 14, 29, 34], which aims at recovering the latent clean image from the observed blurry counterpart. This blurry counterpart can be approximately modeled in mathematics as (assuming the blur process is uniform and spatially invariant)", + "bbox": [ + 212, + 676, + 787, + 753 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} = \\boldsymbol {k} \\otimes \\boldsymbol {x} + \\boldsymbol {n}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 753, + 785, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\pmb{y}$ is the blurry image, $\\pmb{x}$ the underlying clean image, $\\pmb{k}$ the blur kernel, $\\pmb{n}$ usually the additive white Gaussian noise (AWGN), and $\\otimes$ denotes the 2D convolution operator. Then the goal of BID is to estimate $\\pmb{x}$ under this degradation", + "bbox": [ + 212, + 773, + 787, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code is available at https://github.com/jtaoz/GKPILE-Deconvolution.", + "bbox": [ + 220, + 825, + 712, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model without knowing the ground truth $\\pmb{k}$ . As is well-known, this problem is highly ill-posed, since different pairs of $\\pmb{x}$ and $\\pmb{k}$ may result in the same $\\pmb{y}$ .", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The BID task was conventionally formulated as an optimization problem, corresponding to a Bayesian posterior estimation problem from the probabilistic perspective. Thus, the maximum a postriori (MAP) [28, 29, 41, 45, 46, 55, 62] and variational inference (VI) [2, 14, 70] approaches were developed for point-wise and distribution-wise estimation, respectively. The core challenge of these approaches is to properly design the priors for both the clean image [28, 41, 45, 46, 62] and blur kernel [42, 70]. However, these handcrafted priors are relatively subjective and may not precisely characterize the intrinsic distributions of natural images and blur kernels in the real world, which limits the performance of the deblurring algorithm. Besides, the resulting optimization is typically endowed with highly non-convex characteristics, leading to potential convergence issues such as reaching a trivial solution of delta kernel or unsatisfactory local minimum [34].", + "bbox": [ + 212, + 180, + 785, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Owing to their wide and successful applications, deep learning technologies have been introduced to solve the BID problem. Early explorations [6, 16, 56] embedded deep neural networks (DNNs) into the traditional methods to facilitate the estimation of blur kernels. Recent works have pursued a more direct approach, wherein the mapping from a blurry image to its clean counterpart is directly learned [30, 31, 43, 58, 67, 68]. The latter strategies have achieved promising results on several benchmarks, mainly attributed to the well-designed network architectures. However, these methods tend to overfit the training datasets, and thus may not generalize well to unseen scenarios, particularly for images with large and complex motion blur kernels.", + "bbox": [ + 212, + 364, + 787, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, deep image prior (DIP) [59] has exhibited great potential in addressing the task of BID [22, 52], particularly under the scenario of motion blur. The core idea is to parameterize both the blur kernel and clean image as DNNs with random inputs, and then optimize these DNNs through an energy minimization problem conducted by Eq. (1). Compared with the traditional BID methods, these approaches employed more effective DNN-oriented priors and thus achieved superior performance. However, these methods did not fully explore the statistical characteristics of blur kernels, resulting in inaccurate kernel estimations, particularly in cases involving large kernel sizes. Additionally, since the optimization is highly non-convex, the random initialization of the blur kernel often leads to algorithmic instability and compromises the overall performance.", + "bbox": [ + 212, + 518, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate the aforementioned issues of DIP-based methods, we proposed a new approach to better characterize and initialize the blur kernel with deep generative prior (DGP) [7, 48, 60, 64]. Specifically, focusing on the BID of motion blurring, we train a kernel generator using generative adversarial network (GAN) [18] as kernel prior. Owing to GAN's powerful fitting capability on complex distributions, the learned generator is expected to well characterize the blur kernels in a low-dimensional latent space. Then a mapping from the blurry image to the latent code of the kernel generator is trained, acting as a kernel initializer for the subsequent BID task. Intuitively, such a kernel initializer is easier to learn as the latent kernel space is more compact than that of the original kernel. To", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/acf3ab469f5203478b2c95f7ecb49f1b9836ee055b8daadce34c439e8dbd809f.jpg", + "image_caption": [ + "Fig. 1: Illustration of the latent encoding-based kernel initialization and fine-tuning mechanism in our BID framework. The blur kernel is first initialized from the blurry image by a pre-trained kernel initializer via the latent code $\\mathbf{z}_0$ , which is expected to be close to the code of the ground truth kernel, $\\mathbf{z}_{\\mathrm{gt}}$ . The corresponding kernels $\\mathbf{k}_0$ and $\\mathbf{k}_{\\mathrm{gt}}$ in the kernel manifold is also expected to be close. Then the optimization is performed within the latent space, such that $\\mathbf{z}_0$ is fine-tuned to $\\mathbf{z}_{\\mathrm{opt}}$ , and the final estimated kernel $\\mathbf{k}_{\\mathrm{opt}}$ is closer to the ground truth $\\mathbf{k}_{\\mathrm{gt}}$ . See Sec. 3 for more details." + ], + "image_footnote": [], + "bbox": [ + 236, + 143, + 767, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "solve the problem of BID, we first obtain a coarse initialization for the blur kernel by the initializer from the blurry image, and then jointly fine-tune the DIP for the image and the DGP for the blur kernel. The mechanism of the latent encoding-based kernel initialization and fine-tuning is illustrated in Fig. 1. Since the kernel generator sufficiently fits the kernel distribution and the kernel initializer provides a better initial estimation, the overall BID performance can be significantly improved with faster convergence, especially for large blur kernels. Our main contributions are summarized as follows:", + "bbox": [ + 212, + 405, + 787, + 526 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We use GAN to learn the DGP for blur kernels, which can better depict the kernel structures and thus offer a more effective and compact kernel prior.", + "- We propose to learn a kernel initializer that maps from the blurry image to the latent code of the corresponding kernel. Attributed to the compactness of the latent kernel space, the proposed initializer can provide a more accurate kernel initialization for the subsequent BID process.", + "- By harnessing the designed kernel prior and initializer, our BID method achieves state-of-the-art (SotA) performance on the challenging benchmark dataset provided by Lai et al. [33], especially with large blur kernels." + ], + "bbox": [ + 225, + 535, + 784, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 693, + 382, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Blind image deconvolution", + "text_level": 1, + "bbox": [ + 215, + 724, + 480, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BID methods can be roughly divided into two main categories, namely optimization-based methods and deep learning-based methods.", + "bbox": [ + 212, + 750, + 803, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimization-based BID methods. Due to its high ill-posedness, the solution spaces of both the clean image and blur kernel should be properly constrained to make the BID problem solvable. To this aim, traditional optimization-based methods mainly focus on designing appropriate priors for clean images", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding", + "bbox": [ + 235, + 113, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and blur kernels, and plenty of studies have been conducted. For example, total variation [8], hyper-Laplacian prior [28], $l_{1} / l_{2}$ -norm [29], and transform-specific $l_{1}$ -norm [4, 62] were considered to model the image prior in the gradient domain, and patch-based priors [41, 57], low-rank prior [53] and dark/bright channel prior [46, 63] were directly applied in the image domain. As for blur kernels, the non-negative and normalization constraints result in a basic regularization. In addition, focusing on the properties of kernel, sparsity prior [47] and spectral prior [38] have been proposed. To get an ideal solution, tricks such as delayed normalization [50] and multi-scale implementation [71] for blur kernels are also suggested. However, these methods highly rely on handcrafted priors, which may not accurately characterize the intrinsic properties of the clean images and blur kernels, and thus are less competitive in this deep learning era.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deep learning-based BID methods. Having witnessed great success in a wide range of applications, deep learning techniques have also been adopted for the BID task. Early methods attempted to replace certain components of the traditional optimization-based methods with DNNs, by virtue of their flexibility. For example, some researchers introduced DNNs as kernel predictors [6,16,56,65]. As the computing resources increase, it is then more popular to directly learn a mapping from the blurred image to its clean counterpart in a fully supervised way, with paired training data and well-designed DNNs [5,21,30,31,43,44,58,67, 68]. Though achieved SotA performance on several benchmark datasets, such as GoPro [43] and RealBlur [54], these methods may encounter the limitations in generalization to images with large complex blur kernels that are not simulated in the pre-defined training sets.", + "bbox": [ + 212, + 328, + 787, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, focusing on motion blurring, a new type of deep learning methods for the BID problem that leverages the deep priors (will be reviewed in the next subsection), has attracted increasing attention. Along this line, Ren et al. [52] proposed the SelfDeblur method, which adopted DIP to parameterize both the clean image and blur kernel for the first time. After that, multiple variants and extensions, mainly paying attention to the image prior, have been proposed for further improvement of performance [22,35]. These methods achieved promising results in certain cases, especially that they outperform fully supervised deep learning methods on the challenging benchmark by Lai et al. [33]. However, the properties of the blur kernel were not explicitly explored, and the performance tends to be unstable for large kernels due to the non-convex optimization. In contrast to these studies that used DIP with untrained DNNs, Asim et al. [1] proposed to pre-train generators for both the clean images and blur kernels as DGPs, and then fine-tune the generators when doing BID.", + "bbox": [ + 212, + 508, + 787, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Deep prior for image processing", + "text_level": 1, + "bbox": [ + 214, + 739, + 527, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to their flexibility, DNNs have been used to characterize the image priors in recent years. There are mainly two kinds of such deep priors, namely DIP and DGP, which are briefly reviewed in the following.", + "bbox": [ + 212, + 763, + 785, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deep image prior. DIP was originally proposed by Ulyanov et al. [59] who trained a DNN to approximate the single target (maybe up to a transformation)", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with random noise as its input. Since there are complex structures and operators involved, the DNN is expected to well depict the manifold of images, playing a similar role as the prior or regularizer in traditional image processing methods. After being raised, DIP has attracted remarkable attention and been applied to various image processing tasks (other than BID mentioned before), including natural image denoising [59], super-resolution [36, 66], inpainting [59], image decomposition [15], low-light enhancement [69], PET image reconstruction [17] and hyperspectral image denoising [39].", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deep generative prior. Attributed to their powerful generation capabilities, the trained deep generators, such as GANs [18], can be seen as approximators to the distributions of images, hoping to provide reasonable priors for images. Such priors can be referred to as DGPs [48]. After a pre-training stage for the generator, DGP can be used as an estimator to approximate the target image via fine-tuning like DIP. There are two common ways to fine-tune DGP. The first is to fix the generator and optimize the input. For example, Menon et al. [40] used this idea for photo upsampling. This strategy is also closely related to GAN-inversion [61]. The second is to jointly fine-tune the latent code and the parameters of the pre-trained generator like [48], which may better locate the optimal estimation within the image manifold.", + "bbox": [ + 212, + 267, + 784, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to directly modeling the images, DGP was also used to characterize the degradation operators, such as blur kernels. For example, Asim et al. [1] proposed to pre-train a variational auto-encoder (VAE) [26] as DGP for motion blur kernels for the BID task; and Liang et al. [36] used the normalizing flow (NF) [11, 12, 27] as the DGP for Gaussian blur kernels, applying to blind superresolution. These studies highlight the effectiveness of DGP for blur kernels, and partially inspire our work.", + "bbox": [ + 212, + 433, + 785, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Proposed method", + "text_level": 1, + "bbox": [ + 215, + 563, + 423, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 DIP-based BID revisit", + "text_level": 1, + "bbox": [ + 215, + 595, + 447, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Before presenting our BID method, we first revisit DIP-based BID approaches, or more specifically SelfDeblur [52], and discuss the issues we try to address.", + "bbox": [ + 212, + 621, + 782, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In SelfDeblur [52], the clean image $\\pmb{x}$ in Eq. (1) was approximated by a convolutional network equipped with encoder-decoder architecture, and the blur kernel $\\pmb{k}$ was estimated by a fully-connected network with one hidden layer. The resulting optimization can then be formulated as", + "bbox": [ + 212, + 652, + 784, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {k}, \\theta_ {x}} \\| G _ {k} \\left(\\boldsymbol {z} _ {k}; \\theta_ {k}\\right) \\otimes G _ {x} \\left(\\boldsymbol {z} _ {x}; \\theta_ {x}\\right) - \\boldsymbol {y} \\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 720, + 784, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G_{k}(\\cdot ;\\theta_{k})$ and $G_{x}(\\cdot ;\\theta_{x})$ are networks for the kernel and image, respectively, and $\\pmb{z}_k$ and $\\pmb{z}_{x}$ are input random noise. Note that we ignore the additional constraints imposed on the blur kernel and clean image, positing that these constraints can be readily satisfied through thoughtful network design. Thus, the problem of estimating the blur kernel $\\pmb{k}$ and the clean image $\\pmb{x}$ is reduced to the optimization of parameters $\\theta_k^*$ and $\\theta_x^*$ , respectively.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding", + "bbox": [ + 235, + 113, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f123bdded2e99e8ca159a3a591f2effcf8bb2866a3eccc2b2b413042916fb096.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 138, + 331, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f70aa539da4dd02b6f92c50e2c544a8f15c8dc8175d5b59751fd2223375ffc1.jpg", + "image_caption": [ + "Initial Kernel" + ], + "image_footnote": [], + "bbox": [ + 243, + 210, + 331, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35b152e9ad2545a01a42c3c200b39ea487a51f178b117ac28534965d3cfb28e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 140, + 423, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d59f219d3afaeca0b9e85689c1ebeb398f3d56f11623fafff8073f6d75d93fc.jpg", + "image_caption": [ + "Estimated Kernel" + ], + "image_footnote": [], + "bbox": [ + 334, + 210, + 423, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2cd98dbca9e15011cc3b5a7b360d1845ee75a0308247d2a444bdd052f8e232df.jpg", + "image_caption": [ + "Fig.2: Illustration of the initialization effect of the DIP-based BID. The two rows correspond to two independent runs of SelfDeblur [52]. From left to right: the randomly initialized kernel, the finally estimated kernel, and the deblurred image." + ], + "image_footnote": [], + "bbox": [ + 426, + 140, + 562, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec8aa51fcca5c34bec8a3b25c3e8e99414b6f2f7932c7a408c0bd2a66b22ec82.jpg", + "image_caption": [ + "Deblurred Image" + ], + "image_footnote": [], + "bbox": [ + 426, + 210, + 562, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Owing to the powerful capabilities of DNNs, SelfDeblur obtained promising results on the BID task with motion blurring. Notably, it established a new SotA performance on the challenging benchmark introduced by Lai et al. [33]. Huo et al. [22] further improved SelfDeblur by incorporating a more sophisticated DIP within the variational Bayesian framework. However, these DIP-based BID methods did not fully explore the statistical structures or priors for blur kernels, thereby limiting their performance. A more important concern arises from the non-convexity of the optimization, where the randomness of $\\mathbf{z}_k$ causes unstable solutions, particularly when the blur kernel is large. Fig. 2 shows a typical example with two different random runs. Despite the initializations following the same distribution, the estimated blur kernels can be diverse, consequently leading to significantly different deblurring results.", + "bbox": [ + 212, + 294, + 787, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The aforementioned issues redirect our focus toward the blur kernels. Therefore, this study explores more precise statistical structures for blur kernels with DGP, followed by the development of a kernel initializer in the latent space.", + "bbox": [ + 212, + 474, + 785, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Overview of the proposed method", + "text_level": 1, + "bbox": [ + 215, + 542, + 542, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed method consists of two stages. The first is the pre-training stage, aiming to learn a kernel generator and a kernel initializer. The second stage is to solve the deconvolution problem following the paradigm of the DIP-based BID method, while with the aid of the learned generator and initializer.", + "bbox": [ + 212, + 568, + 785, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the pre-training stage, we train a kernel generator $G_{k}(\\cdot ;\\theta_{k}^{*})$ within the GAN [18] framework, where $\\theta_{k}^{*}$ denotes the generator parameters after training. Once learned, it can generate a blur kernel from the randomly sampled noise $z_{k}$ . This means the generator indeed models the distribution of the blur kernels and thus characterizes their statistical structures. Such a learned kernel generator can be referred to as the DGP of blur kernels, with $z_{k}$ serving as the kernel latent code. Next, with the generator $G_{k}(\\cdot ;\\theta_{k}^{*})$ fixed, we train an encoder $E(\\cdot ;\\theta_{E}^{*})$ that maps a blurry image $y$ to the corresponding kernel latent code through the technique of GAN-inversion [61]. The learned encoder allows for a relatively accurate prediction for the blur kernel based on its latent code, and thus acts as an effective initializer for the kernel DGP.", + "bbox": [ + 212, + 628, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the second stage, we first predict the kernel latent code corresponding to the input blurry image $\\pmb{y}$ using the pre-trained kernel initializer, which provides an initial value for the kernel DGP in the latent space. Then we jointly optimize", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4bb981d60bc2247efa3ff775c17c3195f2a3d374c36329a0e03ca30617ac8a7f.jpg", + "image_caption": [ + "Fig. 3: Overview of the proposed BID process, with the pre-trained kernel generator and initializer, of our proposed method." + ], + "image_footnote": [], + "bbox": [ + 233, + 145, + 763, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the parameters of DIP for the clean image and the latent code (or more precisely its mapped features as discussed in Sec. 3.5) of the DGP for the blur kernel. The overview of the proposed BID process is shown in Fig. 3.", + "bbox": [ + 212, + 369, + 784, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Sec. 3.3 and Sec. 3.4, we provide details for learning the kernel generator and kernel initializer, respectively. Then in Sec. 3.5, we discuss how to apply the pre-trained generator and initializer to the BID process.", + "bbox": [ + 212, + 415, + 785, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Kernel generator learning", + "text_level": 1, + "bbox": [ + 214, + 481, + 475, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To learn the kernel generator, a large amount of motion blur kernels are synthesized to simulate their distribution based on the physical generation mechanism proposed in [30] or [33]. Then we train the kernel generator $G_{k}(\\cdot;\\theta_{k}^{*})$ with these synthesized blur kernels. Note that our kernel generator is made up of several convolutional layers inspired by DCGAN [51]. Though can be seen as a straightforward application of GAN, this step is crucial for our method. On the one hand, the learned generator depicts the statistical structures of blur kernels, serving as a kernel DGP in the BID process; on the other hand, the learned kernel generator also forms the milestone for learning the kernel initializer in a low-dimensional latent space, which will be discussed in Sec. 3.4. Fig. 4 shows several blur kernels generated by $G_{k}(\\cdot;\\theta_{k}^{*})$ and the physical model according to [33]. As can be observed, the blur kernels generated by our learned generator are morphologically very similar to that synthesized by the physical model, indicating its capability to characterize the underlying manifold of the blur kernels.", + "bbox": [ + 212, + 506, + 787, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Remark. It should be mentioned there are some studies also considering learning DGP for blur kernels while with various generative models. For example, Asim et al. [1] used the VAE for the BID task, and Liang et al. [36] adopted the NF for the blind image super-resolution. In this work, we prefer GAN due to the following reasons. First, compared with the vanilla VAE that tends to generate blurry samples [3], GAN can provide sharper ones. Second, in contrast to NF, GAN imposes no constraints on the invertibility between the latent code and the generated sample, thereby facilitating easier learning of the kernel manifold in", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding", + "bbox": [ + 235, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6387327c8fceff6e67c5b2b73b81f241725a2cbc44791b4203cd62aa932f8eac.jpg", + "image_caption": [ + "Fig.4: Comparison of blur kernels synthesized according to physical model and generated by our pre-trained generator. Top row: blur kernels synthesized according to [33]. Bottom row: blur kernels generated by the pre-trained kernel generator." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 537, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a lower-dimensional latent space. This aspect is particularly advantageous when considering motion blur kernels, whose sizes can be relatively large (up to $75 \\times 75$ in the Lai dataset [33]).", + "bbox": [ + 212, + 275, + 787, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Kernel initializer learning", + "text_level": 1, + "bbox": [ + 215, + 342, + 475, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As discussed before, kernel initialization is important for the overall performance in the BID task since the optimization is highly non-convex. Therefore, we propose to initialize the blur kernel in the latent space using an encoder $E(\\cdot; \\theta_E^*)$ that maps the blurry image $\\mathbf{y}$ to the latent code $\\mathbf{z}$ of the learned kernel DGP.", + "bbox": [ + 212, + 366, + 782, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A direct way to train the encoder is to solve the following optimization:", + "bbox": [ + 238, + 426, + 756, + 441 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {E} ^ {*} = \\arg \\min _ {\\theta_ {E}} \\sum_ {n} \\ell \\left(G _ {k} \\left(E \\left(\\boldsymbol {y} _ {n}; \\theta_ {E}\\right); \\theta_ {k} ^ {*}\\right), \\boldsymbol {k} _ {n}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 444, + 784, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{y}_n$ denotes the $n$ -th blurry image, $\\pmb{k}_n$ the corresponding blur kernel, $\\theta_k^*$ is the parameters of the pre-trained kernel generator, and $\\ell(\\cdot, \\cdot)$ is some loss function. This optimization can be seen as training a predictor that directly estimates the blur kernel from the blurry image, which is not easy due to the relatively complex structures of blur kernels. On the other hand, if we can find the latent code $\\pmb{z}_n$ of $\\pmb{k}_n$ , and learn to predict $\\pmb{z}_n$ , the task could be simpler, since the latent space is expected to be much more compact than the original kernel space. Therefore, we propose to use a collaborative learning strategy motivated by [19], to guide the encoder training within the latent space. Specifically, we first formulate the following optimization to train the encoder:", + "bbox": [ + 212, + 469, + 787, + 619 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {E}} \\sum_ {n} \\left\\{\\| G _ {k} \\left(E \\left(\\boldsymbol {y} _ {n}; \\theta_ {E}\\right); \\theta_ {k} ^ {*}\\right) - G _ {k} \\left(\\boldsymbol {z} _ {n}; \\theta_ {k} ^ {*}\\right) \\| _ {1} + \\lambda \\| E \\left(\\boldsymbol {y} _ {n}; \\theta_ {E}\\right) - \\boldsymbol {z} _ {n} \\| _ {2} ^ {2} \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 622, + 784, + 657 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\boldsymbol {z} _ {n} = \\arg \\min _ {\\boldsymbol {z}} \\| G _ {k} (\\boldsymbol {z}; \\theta_ {k} ^ {*}) - \\boldsymbol {k} _ {n} \\| _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 659, + 540, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda$ is a tuning parameter set to 0.1 in this work. For convenience, we denote the objective functions of the outer and inner optimizations as $\\mathcal{L}_E(\\theta_E)$ and $\\mathcal{L}_z^n (z)$ , respectively, in the following.", + "bbox": [ + 212, + 689, + 784, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The optimization of Eq. (4) involves a sub-task of solving $\\mathbf{z}_n$ that corresponds to a GAN-inversion problem [61]. Following [19], we alternatively optimize $\\mathbf{z}_n$ and $\\theta_E$ in each iteration. Specifically, $\\mathbf{z}_n$ is firstly initialized as $E(\\mathbf{y}_n; \\theta_E)$ , and then updated by several gradient descent steps with respect to the lower-level objective $\\mathcal{L}_z^n(\\mathbf{z})$ . Based on the updated $\\mathbf{z}_n$ , $\\theta_E$ is subsequently optimized according to the upper-level objective $\\mathcal{L}_E(\\theta_E)$ , also by gradient descent steps. The whole procedure is summarized in Algorithm 1.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Kernel initializer learning" + ], + "code_body": "Input: Pre-trained kernel generator $G_{k}(\\cdot ;\\theta_{k}^{*})$ , blurry image-kernel pairs $\\{\\pmb {y}_n,\\pmb {k}_n\\}$ , step size $\\epsilon$ 1: Initialize $\\theta_E^{(0)}$ \n2: for $t = 1,\\dots ,T$ do \n3: For each $n$ .. \n4: $\\pmb{z}^{(0)} = E(\\pmb {y}_n;\\theta_E^{(t - 1)})$ \n5: for $s = 1,\\dots ,S$ do $\\pmb{z}^{(s)} = \\pmb{z}^{(s - 1)} - \\epsilon \\nabla_{\\pmb{z}}\\mathcal{L}_{\\pmb{z}}^{n}(\\pmb {z})|_{\\pmb{z} = \\pmb{z}^{(s - 1)}}$ end for \n6: $\\pmb{z}_{n}^{(t)} = \\pmb{z}^{(S)}$ \n7: $\\theta^{(0)} = \\theta_{E}^{(t - 1)}$ \n8: for $l = 1,\\dots ,L$ do $\\theta^{(l)} = \\theta^{(l - 1)} - \\epsilon \\nabla_{\\theta}\\mathcal{L}_{E}(\\theta)|_{\\theta = \\theta^{(l - 1)}}$ end for \n9: $\\theta_E^{(t)} = \\theta^{(L)}$ \n10: end for \nOutput: Kernel initializer $E(\\cdot ;\\theta_E^*)$ , where $\\theta_E^* = \\theta_E^{(T)}$", + "bbox": [ + 215, + 164, + 756, + 311 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5b55c6a713502069a8cda2752d5574d5ada2bcb1a3531085fabaffd27208ee70.jpg", + "image_caption": [ + "Fig.5: Illustration of the estimating ability of the pre-trained kernel initializer. Left: the blurry image. Top-right: the estimated blur kernel by the kernel initializer. Bottomright: the ground-truth blur kernel." + ], + "image_footnote": [], + "bbox": [ + 236, + 321, + 555, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Once trained, encoder $E(\\cdot; \\theta_E^*)$ can be regarded as an effective kernel initializer, providing a promising prediction of the blur kernel from the blurry image. Fig. 5 shows an example of the predicted kernel by the learned kernel initializer, in comparison with the ground truth. It can be observed that, though not perfect, the prediction is indeed very close to the ground truth, and thus is expected to be a good starting point for further processing.", + "bbox": [ + 212, + 453, + 787, + 545 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 BID Process", + "text_level": 1, + "bbox": [ + 215, + 566, + 366, + 579 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With the pre-trained kernel generator $G_{k}(\\cdot ;\\theta_{k}^{*})$ and initializer $E(\\cdot ;\\theta_{E}^{*})$ , we can formulate the BID problem as follows:", + "bbox": [ + 212, + 590, + 785, + 621 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {z} _ {k} ^ {*}, \\theta_ {x} ^ {*}\\right) = \\underset {\\boldsymbol {z} _ {k}, \\theta_ {x}} {\\arg \\min } \\| G _ {k} \\left(\\boldsymbol {z} _ {k}; \\theta_ {k} ^ {*}\\right) \\otimes G _ {x} \\left(\\boldsymbol {z} _ {x}; \\theta_ {x}\\right) - \\boldsymbol {y} \\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 628, + 785, + 652 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\pmb{z}_x$ is a random vector, $\\pmb{z}_k$ is initialized by $\\pmb{z}_k^{(0)} = E(\\pmb{y};\\theta_E^*)$ . After optimization, the desired clean image and blur kernel can be accessed via $\\hat{\\pmb{x}} = G_{x}(\\pmb{z}_{x};\\theta_{x}^{*})$ and $\\hat{\\pmb{k}} = G_{k}(\\pmb{z}_{k}^{*};\\theta_{k}^{*})$ , respectively. In fact, such a naive implementation still cannot achieve satisfactory results as empirically illustrated in Sec. 4.3, in particular when the blur kernel is large. This can be attributed to the relatively low dimensionality of $\\pmb{z}_k$ , compared with the blur kernel, which enforces too strong constraints and makes the optimization more difficult.", + "bbox": [ + 212, + 656, + 787, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To address this issue, let's first recall one small yet effective trick in StyleGAN [23]. It introduced a network to map the latent code $\\mathbf{z}$ to a style vector $\\mathbf{w}$ , enabling more precise control of the generated style. Inspired by this trick, we propose to optimize the feature map of the first layer in the kernel generator $G_{k}(\\cdot ;\\theta_{k}^{*})$ , whose dimension is relatively higher, instead of the latent code $\\mathbf{z}$ . By", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding", + "bbox": [ + 235, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 BID process with pre-trained kernel generator and initializer" + ], + "code_body": "Input: Pre-trained kernel generator $G_{k}(\\cdot ;\\theta_{k}^{*}) = G_{k}^{(w)}(g_{1}(\\cdot))$ pre-trained kernel initializer $E(\\cdot ;\\theta_E^*)$ blurry image $\\pmb{y}$ , random vector $\\pmb {z}_x$ , step size $\\epsilon$ \n1: Initialize $\\pmb{w}_{k}^{(0)} = g_{1}(E(\\pmb {y};\\theta_{E}^{*}))$ $\\theta_x^{(0)}$ , denote $\\mathcal{L}(\\boldsymbol {w},\\theta) = \\left\\| G_k^{(w)}(\\boldsymbol {w})\\otimes G_x(\\boldsymbol {z}_x;\\theta) - \\boldsymbol {y}\\right\\| ^2$ \n2: for $t = 1,\\dots ,T$ do \n3: $\\begin{array}{r}\\pmb{w}_{k}^{(t)} = \\pmb{w}_{k}^{(t - 1)} - \\epsilon \\nabla_{\\pmb{w}}\\mathcal{L}(\\pmb {w},\\theta)\\big|_{\\pmb {w} = \\pmb{w}_{k}^{(t - 1)},\\theta = \\theta_{x}^{(t - 1)}}\\\\ \\theta_{x}^{(t)} = \\theta_{x}^{(t - 1)} - \\epsilon \\nabla_{\\theta}\\mathcal{L}(\\pmb {w},\\theta)\\big|_{\\pmb {w} = \\pmb{w}_{k}^{(t - 1)},\\theta = \\theta_{x}^{(t - 1)}}\\\\ \\end{array}$ \n4: end for \n6: $\\pmb{w}_{k}^{*} = \\pmb{w}_{k}^{(T)}$ $\\theta_x^* = \\theta_x^{(T)}$ \nOutput: Estimated clean image $\\hat{\\pmb{x}} = G_{x}(\\pmb {z}_{x};\\theta_{x}^{*})$ , estimated blur kernel $\\hat{\\pmb{k}} = G_k^{(w)}(\\pmb {w}_k^*)$", + "bbox": [ + 215, + 164, + 785, + 292 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "denoting the first layer of $G_{k}(\\cdot ;\\theta_{k}^{*})$ as $g_{1}(\\cdot)$ and the reduced generator without $g_{1}(\\cdot)$ as $G_{k}^{(w)}(\\cdot)$ , the BID problem of Eq. (5) is reformulated as", + "bbox": [ + 212, + 310, + 782, + 344 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {w} _ {k} ^ {*}, \\theta_ {x} ^ {*}\\right) = \\underset {\\boldsymbol {w} _ {k}, \\theta_ {x}} {\\arg \\min } \\left\\| G _ {k} ^ {(w)} \\left(\\boldsymbol {w} _ {k}\\right) \\otimes G _ {x} \\left(\\boldsymbol {z} _ {x}; \\theta_ {x}\\right) - \\boldsymbol {y} \\right\\| ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 345, + 785, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\pmb{w}_k$ is initialized by $\\pmb{w}_k^{(0)} = g_1(E(\\pmb{y};\\theta_E^*))$ . The whole BID process for solving Eq. (6) is summarized in Algorithm 2.", + "bbox": [ + 212, + 382, + 782, + 412 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "It should be noted that the parameters of the generator network can also be fine-tuned in previous studies [48]. However, due to the highly non-convex nature of the BID optimization, too many to-be-optimized parameters may hinder its performance. We thus only update $\\pmb{w}_k$ for the blur kernel in this work. Empirical results in Sec. 4.3 substantiate the superiority of this strategy over updating the whole kernel generator.", + "bbox": [ + 212, + 414, + 784, + 503 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 529, + 375, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we present experimental results to demonstrate the effectiveness of the proposed method and ablation studies to analyze some key components in our framework. More results can be found in the supplementary material, together with discussions on the limitations of this work.", + "bbox": [ + 212, + 561, + 784, + 622 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets and implementation details", + "text_level": 1, + "bbox": [ + 215, + 645, + 563, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We evaluate our proposed framework mainly on two datasets: one is synthesized by ourselves, and the other is the well-known BID benchmark constructed by Lai et al. [33]. For the synthetic dataset, we randomly select 80 source images from MSCOCO [37] and apply the motion blur synthesized by [30] on them. The second Lai dataset consists of 25 clean images and 4 large-size blur kernels, resulting in 100 testing blur images in total. Beyond the two synthetic datasets, we also test our method real blurry images collected by Lai et al. [33].", + "bbox": [ + 212, + 671, + 785, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. As discussed in Sec. 3, our whole framework consists of a pre-training stage and a BID process. In the pre-training stage, we first train the kernel generator $G_{k}(\\cdot ;\\theta_{k}^{*})$ with DCGAN [51], using kernels synthesized following the way in [30] or [33]. Then we freeze the weights of $G_{k}(\\cdot ;\\theta_{k}^{*})$ and", + "bbox": [ + 212, + 779, + 785, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/23737c746a1c76f492c58a1b09c55c6cfac69b22e06675f7ca204d01aa6cc1bd.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons of various methods on our synthetic dataset." + ], + "table_footnote": [], + "table_body": "
MethodPSNRSSIMMethodPSNRSSIM
Pan et al. [46]24.090.808Zamir et al. [68]22.270.652
Dong et al. [13]23.850.799Ren et al. [52]25.480.765
Tao et al. [58]23.210.699Huo et al. [22]28.130.898
Kupyyn et al. [31]23.460.716Li et al. [35]24.490.744
Kaufman and Fattal [24]25.420.814Ours29.670.928
", + "bbox": [ + 217, + 160, + 781, + 224 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/43b8f0b5ebbebc77bc31136610cf7ee93a26f98090309b37e600c1d3c287d131.jpg", + "image_caption": [ + "Blurred" + ], + "image_footnote": [], + "bbox": [ + 217, + 231, + 310, + 285 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/af285bc82f7db4dae76fa6253c77d3b9818da4a493f338fa2ff83d2041e1c647.jpg", + "image_caption": [ + "Pan [46]" + ], + "image_footnote": [], + "bbox": [ + 313, + 232, + 405, + 285 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b393c5771c8a45fac7c1f70625e7a4d79251c3546bd2ffeb8c17d655d8ea7f39.jpg", + "image_caption": [ + "Dong [13]" + ], + "image_footnote": [], + "bbox": [ + 406, + 232, + 500, + 285 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/aae5f2cb585eea31b31ac13fc0d902c40cba9c4bdaae571a06218475c1fced5f.jpg", + "image_caption": [ + "Tao [58]" + ], + "image_footnote": [], + "bbox": [ + 501, + 232, + 594, + 285 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fcaae1773cf4ef864ef08222304fd4a05d4265cc259ea562e24c829ee5311746.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 232, + 687, + 285 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dd9c4692da8563a9d018080eadcf94875eb3563fd91df4f161f9899c9ccbf358.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 232, + 784, + 285 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/52c94a73e39b52b2928cd415dcca4f807d8740c471a8ce7663434f0cacbad417.jpg", + "image_caption": [ + "Zamir [67]" + ], + "image_footnote": [], + "bbox": [ + 217, + 299, + 310, + 351 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/25f9fe1055baa8a08732db49038e093c883ea5d42991c2b10e8a34c8f78b1b03.jpg", + "image_caption": [ + "Ren [52]" + ], + "image_footnote": [], + "bbox": [ + 313, + 299, + 405, + 351 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/99232cba2cc3c7f21c5649c049949f01c1f8d164a30860d74615b7f10d3fc736.jpg", + "image_caption": [ + "Huo [22]", + "Fig. 6: Visual results on our synthetic dataset. The estimated blur kernel is placed on the top-left corner for each method if available." + ], + "image_footnote": [], + "bbox": [ + 406, + 299, + 500, + 351 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/452d85d7c795f606bfa56b598783ddbbd8cbe9b5259179d59e0018384d9c8875.jpg", + "image_caption": [ + "Li [35]" + ], + "image_footnote": [], + "bbox": [ + 501, + 299, + 594, + 351 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ff405bbabdbfe218d7d8e9754d8feed9a5573812e86bdf538af7be5a83aab8c7.jpg", + "image_caption": [ + "Kupyn [31]", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 596, + 299, + 687, + 351 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3e06843ae7954a2f6466dc400f8b293317c556db22518cf4af43c54581da9dbb.jpg", + "image_caption": [ + "Kaufman [24]", + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 692, + 299, + 784, + 351 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "train the kernel initializer $E(\\cdot; \\theta_E^*)$ that is a ResNet-18 [20]. To train the kernel initializer, we randomly crop $256 \\times 256$ patches from the sources in OpenImages [32] and convolve them with blur kernels to get blurry images. The Adam optimizer [25] with its default configuration in PyTorch [49] is adopted. The initial learning rate is 1e-4 and reduced to 1e-5 as the loss becomes stable and finally to 1e-6. In the BID process, we jointly estimate the blur kernel $\\pmb{k}$ and the clean image $\\pmb{x}$ by optimizing $\\pmb{w}_k$ and $\\theta_x$ according to Eq. (6). The initial learning rate of $\\pmb{w}_k$ and $\\theta_x$ are set as 5e-4 and 1e-2, respectively, and decayed following [52]. For the architecture of the DIP network $G_x(\\cdot; \\theta_x)$ , we follow the settings in SelfDeblur [52].", + "bbox": [ + 212, + 398, + 787, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Experimental results", + "text_level": 1, + "bbox": [ + 215, + 573, + 433, + 588 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on our synthetic dataset. We first verify the effectiveness of the proposed method on the synthetic dataset by ourselves. Nine comparison methods are considered, including two traditional model-based ones (Pan et al. [46], Dong et al. [13]), four supervised deep learning ones (Tao et al. [58], Kupyn et al. [31], Kaufman and Fattal [24], Zamir et al. [68]), and three DIP-based ones (Ren et al. [52], Huo et al. [22], Li et al. [35]). We follow the default settings in their papers for these methods or tune them by ourselves for the best performance.", + "bbox": [ + 212, + 598, + 784, + 703 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The quantitative results of all competing methods are summarized in Table 1. As can be seen from this table, the proposed method achieves the best performance in terms of both PSNR and SSIM, demonstrating its effectiveness. The superiority of our method can be visually observed from Fig. 6. Specifically, the traditional model-based methods are able to obtain accurate estimations for the blur kernel, but the BID results are not very satisfactory due to the insufficiency of the handcrafted image priors. Most of the supervised deep learning methods generate over-smoothed outcomes, mainly attributed to the lack of exploitation of the physical blur model. Even though achieving relatively better results, the", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding", + "bbox": [ + 235, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7b835e463aba1fc986108423bfa82d6559ee48e4b7323261479fb31c3f3849eb.jpg", + "table_caption": [ + "Table 2: Quantitative results of various methods on the dataset by Lai et al. [33]." + ], + "table_footnote": [], + "table_body": "
MethodManmadeNaturalPeopleSaturatedTextAverage
Cho and Lee [10]16.35/0.38920.14/0.52019.90/0.55614.05/0.49314.87/0.44317.06/0.480
Krishnan et al. [29]15.73/0.40819.44/0.52621.50/0.64714.09/0.50715.40/0.48517.23/0.514
Xu et al. [62]17.99/0.59921.58/0.67924.40/0.81314.53/0.53817.64/0.66819.23/0.659
Perrone and Favaro [50]17.41/0.55121.04/0.67622.77/0.73514.24/0.51116.94/0.59318.48/0.613
Pan et al. [46]18.25/0.56122.15/0.67222.72/0.71016.01/0.59116.69/0.53019.16/0.613
Dong et al. [13]17.30/0.43221.18/0.58621.89/0.65215.85/0.55416.18/0.48018.48/0.541
Tao et al. [58]16.77/0.34219.90/0.45921.41/0.61615.12/0.50615.41/0.43717.72/0.472
Kupyn et al. [31]17.05/0.36720.37/0.48921.70/0.64215.20/0.52215.86/0.47618.04/0.499
Kaufman and Fattal [24]19.36/0.65023.14/0.75626.78/0.87316.40/0.65417.92/0.68320.72/0.723
Zamir et al. [68]16.73/0.34119.80/0.44720.88/0.58715.14/0.50514.49/0.38317.41/0.453
Ren et al. [52]19.00/0.61722.94/0.71526.43/0.78118.38/0.68822.82/0.77621.91/0.715
Huo et al. [22]22.51/0.82026.20/0.90131.38/0.95318.14/0.69127.72/0.93325.19/0.860
Li et al. [35]20.81/0.72124.07/0.78128.45/0.86014.81/0.52723.10/0.83422.25/0.744
Ours24.67/0.94228.12/0.96932.86/0.97318.98/0.73728.91/0.95126.71/0.914
BlurredCho [10]Krishnan [29]Xu [62]Perrone [50]Pan [46]Dong [13]
Kaufman [24]Zamir [67]Ren [52]Huo [22]Li [35]OursGround truth
", + "bbox": [ + 223, + 159, + 779, + 388 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 7: Visual results on the dataset by Lai et al. [33]. The estimated blur kernel is placed on the top-left corner of each method if available.", + "bbox": [ + 215, + 388, + 785, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "existing DIP-based methods still have some limitations. For example, Ren et al.'s method performs unsatisfactorily, and the structural details in the images by Huo et al. and Li et al.'s methods are not sharp enough. In contrast to these methods, our method can not only accurately estimate the blur kernel, but also recover more image details with fewer distortions.", + "bbox": [ + 212, + 422, + 784, + 498 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on the dataset by Lai et al. [33]. We evaluate our proposed method on the challenging BID benchmark by Lai et al. [33]. This dataset consists of 100 blurry images synthesized by applying 4 blur kernels, with sizes ranging from $31 \\times 31$ to $75 \\times 75$ , to 25 clean images. These 25 clean images are categorized into 5 groups, namely Manmade, Natural, People, Saturated and Text. We compare our method with 13 existing methods, including six traditional model-based methods (Cho and Li [10], Krishnan et al. [29], Xu et al. [62], Perrone and Favaro [50], Pan et al [46], Dong et al. [13]), four supervised deep learning methods (Tao et al. [58], Kupyn et al. [31], Kaufman and Fattal [24], Zamir et al. [68]), and three DIP-based methods (Ren et al. [52], Huo et al. [22], Li et al. [35]).", + "bbox": [ + 212, + 503, + 787, + 656 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2 reports the quantitative results of all competing methods, in terms of PSNR and SSIM. We can see that our method significantly outperforms existing methods in all of the five categories. Notably, our method is the only one that achieves an SSIM value higher than 0.9 on average. The visual results on one typical example are shown in Fig. 7. It can be observed that most of the traditional model-based methods fail to estimate the blur kernel due to the relatively larger kernel size in this dataset. The supervised deep learning methods still produce distorted results as before, because they do not carefully consider the physical blur model. In comparison, DIP-based methods, including ours, can accurately estimate the blur kernel and further achieve promising deblurring results. Among them, our method recovers more details, such as hairs, with reference to the ground truth image, showing its superiority.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fc7cceebc0da02d3d809e6c86bd3ab2625120508bbecd04b0b50727fc4a9760f.jpg", + "image_caption": [ + "Blurred" + ], + "image_footnote": [], + "bbox": [ + 248, + 143, + 331, + 186 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fd2971b66365791da9ea173a4ff31d367aca1ec139aecb9325500964eb6dba0b.jpg", + "image_caption": [ + "Pan [46]" + ], + "image_footnote": [], + "bbox": [ + 334, + 143, + 416, + 186 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6947626b66b3e2b2be3680527f9e40502b414739e2177ae844165fe0878b2ab4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 143, + 500, + 186 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/00f15976fa312fb43b9571d1acb2ffb6294bdc0e22588db5430b16435b11df4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 584, + 186 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8d8154e4e672373f55984ed03142bbabe49ace304d3021a14244ce625ee3f785.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 143, + 668, + 186 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7ace74815d9bd1b1bbc9f6516cbeaf182f28aef7c861931e8ecad437f2b2a43d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 143, + 751, + 186 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/297dd9de7ad9871b3605a403349973206ebaa15c1f53ebc25fba314ebfbd9d3e.jpg", + "image_caption": [ + "Zamir [67]" + ], + "image_footnote": [], + "bbox": [ + 292, + 199, + 372, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/33ea5e7566dbd8bfd5c56a0897edd31cace5b6d504dace6aa07bc5a8355a98c4.jpg", + "image_caption": [ + "Dong [13]", + "Ren [52]" + ], + "image_footnote": [], + "bbox": [ + 375, + 199, + 457, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/79482bbf21205a7733c160c8bfe4524911babea121430d68c474b45fe4db9796.jpg", + "image_caption": [ + "Tao [58]", + "Huo [22]" + ], + "image_footnote": [], + "bbox": [ + 460, + 199, + 542, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7faa62d02101d97806e48e440dc64acc46992d65113093dde0e909e457b2a23b.jpg", + "image_caption": [ + "Kupyn [31]", + "Li [35]" + ], + "image_footnote": [], + "bbox": [ + 544, + 199, + 627, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0b7d3b82dab6bb0f200c43fba73f74a3daf88f6cf1a517aa23cef40d81663e3a.jpg", + "image_caption": [ + "Kaufman [24]", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 629, + 199, + 710, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3f2c27a01f07e76cb7bac61d96328694bc5d8056011dfe647b0d9d69f545bdc2.jpg", + "image_caption": [ + "Fig. 8: Example results on the real blurry image provided by Lai et al. [33]. The estimated blur kernel is placed on the top-left corner of each method if available.", + "Fig.9: Performance comparison (PSNR) of the DIP-based BID method (SelfDeblur [52]) and our method in different iteration steps on the dataset by Lai et al. [33]." + ], + "image_footnote": [], + "bbox": [ + 227, + 294, + 467, + 436 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Example results on real blurry images. We further apply comparison BID methods to real blurry images collected by Lai et al. [33], and show one example in Fig. 8. As can be seen, on this very challenging image, none of the methods generate perfect results, while our method provides a relatively promising one.", + "bbox": [ + 214, + 441, + 784, + 503 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 523, + 385, + 540 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in the previous experiments, the proposed method has demonstrated its effectiveness on various blurry images. In this section, we provide further analyses for better understanding its advantageous properties.", + "bbox": [ + 212, + 549, + 784, + 594 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Convergence speed. One major methodological difference between our method and existing DIP-based methods is that we can initialize the blur kernel with pretrained models instead of random sampling. This property can not only result in more stable and better performance but also improve the convergence speed. To verify this point, we compare the PSNR values of our methods with that of Ren et al.'s [52] across various iterations on the Lai dataset in Fig. 9. It can be seen that the performance of our method in the 400-th iteration is comparable with that of Ren et al.'s method in the 1000-th iteration. This faster convergence can be mostly attributed to the better initialization and characterization of the blur kernel, since both of them adopt the same DIP network for the image.", + "bbox": [ + 212, + 595, + 785, + 747 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Optimization strategy for blur kernel. As discussed in Sec. 3.5, during the BID process, we optimize the higher-dimensional feature map $\\boldsymbol{w}_k$ instead of the original latent code $\\boldsymbol{z}_k$ for the blur kernel. Now we empirically show the different behaviors of the two optimization strategies. As a reference, the results by optimizing the entire kernel generator are also reported, which has a larger search space for optimization. The detailed comparison results on the Lai dataset are", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding", + "bbox": [ + 235, + 113, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/65e32587506f2ff2b799e97b8b718e6167cc9f284f2ac5996dba74ed975040da.jpg", + "table_caption": [ + "Table 3: Performance comparison (PSNR/SSIM) of different kernel optimization strategy on the dataset by Lai et al. [33]." + ], + "table_footnote": [], + "table_body": "
Kernel size31515575Average
wk27.89/0.93226.31/0.90125.91/0.91626.73/0.90826.71/0.914
zk27.11/0.92723.46/0.83022.75/0.82522.43/0.76323.94/0.836
zk and θk24.38/0.83022.33/0.71624.45/0.81924.17/0.81423.83/0.795
", + "bbox": [ + 217, + 172, + 785, + 215 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/124d902cc726db409cff3d21cc19ac65b119c5195817170da41fd0494dd23938.jpg", + "table_caption": [ + "Table 4: Performance comparison (PSNR/SSIM) of different kernel initialization strategies on the dataset by Lai et al. [33]." + ], + "table_footnote": [], + "table_body": "
BaselineRandomG(z0; θk*)G(E(y; θE*); θk*)
Metrics19.51/0.56323.00/0.74326.71/0.914
", + "bbox": [ + 256, + 250, + 745, + 273 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "summarized in Table 3. We can observe that optimizing the whole generator is not competitive due to the too-large search space. When the blur kernel is relatively small, the other two strategies have similar performance, while optimizing $\\pmb{w}_k$ is slightly better. As the blur kernel becomes larger, the performance by optimizing $\\pmb{z}_k$ significantly drops, while optimizing $\\pmb{w}_k$ still produces promising results. These observations demonstrate the reasonability of the proposed optimization strategy and also emphasize the importance of a proper search space.", + "bbox": [ + 212, + 284, + 785, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of kernel initialization. In our BID framework, a pre-trained encoder is employed to initialize the blur kernel. To show the necessity of such an initializer, we compare the proposed initialization strategy to two baselines. The first is the random initialization that randomly samples a latent code $\\mathbf{z}$ for the blur kernel. The second is to initialize the kernel with all elements being the same, which can be seen as an \"average\" kernel, and then find the corresponding $\\mathbf{z}_0$ by GAN-inversion. It should be noted that both of these two initialization strategies are independent of the blurry image. Table 4 lists the comparison results on the Lai dataset. We can see that the random initialization strategy produces the worst result since the randomly initialized kernel might largely deviate from the target without any constraint. The fixed \"average\" kernel performs better, and indeed already outperforms many other methods in Table 2. That's because this strategy, in some sense, can be regarded as an \"average\" approximation to any blur kernel. Not surprisingly, the proposed initialization with a pre-trained encoder obtains the best performance as the initialized kernel has been very close to the ground-truth one, as shown in Fig. 5. These results substantiate the significant benefits brought by the pre-trained kernel initializer for the BID task.", + "bbox": [ + 212, + 391, + 785, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 671, + 359, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have proposed a new framework for the BID task, by virtue of deep generative model. Within this framework, we first pre-train a kernel generator as a DGP for blur kernels and a kernel initializer that can offer a well-initialized kernel. Then during the BID process, the blur kernel is initialized in the latent feature space and jointly optimized with the DIP network for the final result. Comprehensive experiments have been conducted and demonstrated the effectiveness of the proposed method. The ablation study verifies the necessity of each component in our framework. In the future, we will try to generalize our method to non-uniform blur modeling, overcoming its current limitations.", + "bbox": [ + 212, + 703, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the NSFC Projects (No. 12226004, 62076196, 62331028, 62272375). The authors thank Prof. Jiangxin Dong for providing the code for generating the blur kernels with the type of Lai dataset.", + "bbox": [ + 212, + 172, + 787, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Asim, M., Shamshad, F., Ahmed, A.: Blind image deconvolution using deep generative priors. IEEE Transactions on Computational Imaging (TCI) 6, 1493-1506 (2020)", + "2. Bishop, T.E., Babacan, S.D., Amizic, B., Katsaggelos, A.K., Chan, T., Molina, R.: Blind image deconvolution: problem formulation and existing approaches. In: Blind Image Deconvolution, pp. 21-62. CRC press (2017)", + "3. Bredell, G., Flouris, K., Chaitanya, K., Erdil, E., Konukoglu, E.: Explicitly minimizing the blur error of variational autoencoders. In: Proceedings of International Conference on Learning Representations (ICLR) (2023)", + "4. Cai, J.F., Ji, H., Liu, C., Shen, Z.: Blind motion deblurring from a single image using sparse approximation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 104-111 (2009)", + "5. Cai, J., Zuo, W., Zhang, L.: Dark and bright channel prior embedded network for dynamic scene deblurring. IEEE Transactions on Image Processing (TIP) 29, 6885-6897 (2020)", + "6. Chakrabarti, A.: A neural approach to blind motion deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 221-235 (2016)", + "7. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14245-14254 (2021)", + "8. Chan, T.F., Wong, C.K.: Total variation blind deconvolution. IEEE Transactions on Image Processing (TIP) 7(3), 370-375 (1998)", + "9. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1742-1750 (2019)", + "0. Cho, S., Lee, S.: Fast motion deblurring. ACM Transactions on Graphics (TOG) 28(5), 1-8 (2009)", + "1. Dinh, L., Krueger, D., Bengio, Y.: NICE: Non-linear Independent Components Estimation (Apr 2015)", + "2. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using Real NVP. In: Proceedings of International Conference on Learning Representations (ICLR) (2017)", + "3. Dong, J., Pan, J., Su, Z., Yang, M.H.: Blind image deblurring with outlier handling. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2478-2486 (2017)", + "4. Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T.: Removing camera shake from a single photograph. ACM Transactions on Graphics (TOG) 25(3), 787-794 (2006)", + "5. Gandelsman, Y., Shocher, A., Irani, M.: \"Double-DIP\": Unsupervised image decomposition via coupled deep-image-priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11018-11027 (2019)" + ], + "bbox": [ + 225, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding", + "bbox": [ + 235, + 113, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Gong, D., Yang, J., Liu, L., Zhang, Y., Reid, I., Shen, C., Van Den Hengel, A., Shi, Q.: From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2319-2328 (2017)", + "17. Gong, K., Catana, C., Qi, J., Li, Q.: Pet image reconstruction using deep image prior. IEEE Transactions on Medical Imaging (TMI) 38(7), 1655-1665 (2019)", + "18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in Neural Information Processing Systems (NeurIPS) (2014)", + "19. Guan, S., Tai, Y., Ni, B., Zhu, F., Huang, F., Yang, X.: Collaborative learning for faster stylegan embedding. arXiv preprint arXiv:2007.01758 (2020)", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)", + "21. Hradis, M., Kotera, J., Zemcik, P., Sroubek, F.: Convolutional neural networks for direct text deblurring. In: Proceedings of British Machine Vision Conference (BMVC) (2015)", + "22. Huo, D., Masoumzadeh, A., Kushol, R., Yang, Y.H.: Blind image deconvolution using variational deep image prior. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(10), 11472-11483 (2023)", + "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4401-4410 (2019)", + "24. Kaufman, A., Fattal, R.: Deblurring using analysis-synthesis networks pair. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5811-5820 (2020)", + "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Proceedings of the International Conference for Learning Representations (ICLR) (2015)", + "26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: Proceedings of International Conference on Learning Representations (ICLR) (2014)", + "27. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. In: Advances in Neural Information Processing Systems (NeurIPS) (2018)", + "28. Krishnan, D., Fergus, R.: Fast image deconvolution using hyper-laplacian priors. In: Advances in Neural Information Processing Systems (NeurIPS) (2009)", + "29. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 233-240 (2011)", + "30. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8183-8192 (2018)", + "31. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8878-8887 (2019)", + "32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV) 128(7), 1956-1981 (2020)", + "33. Lai, W.S., Huang, J.B., Hu, Z., Ahuja, N., Yang, M.H.: A comparative study for single image blind deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1701-1709 (2016)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1964-1971 (2009)", + "35. Li, J., Wang, W., Nan, Y., Ji, H.: Self-supervised blind motion deblurring with deep expectation maximization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13986-13996 (2023)", + "36. Liang, J., Zhang, K., Gu, S., Gool, L.V., Timofte, R.: Flow-based kernel prior with application to blind super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10596-10605 (2021)", + "37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 740-755 (2014)", + "38. Liu, G., Chang, S., Ma, Y.: Blind image deblurring using spectral properties of convolution operators. IEEE Transactions on Image Processing (TIP) 23(12), 5047-5056 (2014)", + "39. Luo, Y.S., Zhao, X.L., Jiang, T.X., Zheng, Y.B., Chang, Y.: Hyperspectral mixed noise removal via spatial-spectral constrained unsupervised deep image prior. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS) 14, 9435-9449 (2021)", + "40. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2437-2445 (2020)", + "41. Michaeli, T., Irani, M.: Blind deblurring using internal patch recurrence. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 783-798 (2014)", + "42. Molina, R., Katsaggelos, A.K., Abad, J., Mateos, J.: A bayesian approach to blind deconvolution based on dirichlet distributions. In: Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP). vol. 4, pp. 2809-2812 (1997)", + "43. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3883-3891 (2017)", + "44. Pan, J., Dong, J., Liu, Y., Zhang, J., Ren, J., Tang, J., Tai, Y.W., Yang, M.H.: Physics-based generative adversarial models for image restoration and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 43(7), 2449-2462 (2021)", + "45. Pan, J., Hu, Z., Su, Z., Yang, M.H.: Deblurring text images via 10-regularized intensity and gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2901-2908 (2014)", + "46. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1628-1636 (2016)", + "47. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Deblurring images via dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(10), 2315-2328 (2017)", + "48. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 7474-7489 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding", + "bbox": [ + 235, + 113, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)", + "50. Perrone, D., Favaro, P.: Total variation blind deconvolution: The devil is in the details. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2909-2916 (2014)", + "51. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. In: Proceedings of International Conference on Learning Representations (ICLR) (2016)", + "52. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3341-3350 (2020)", + "53. Ren, W., Cao, X., Pan, J., Guo, X., Zuo, W., Yang, M.H.: Image deblurring via enhanced low-rank prior. IEEE Transactions on Image Processing (TIP) 25(7), 3426-3437 (2016)", + "54. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 184-201 (2020)", + "55. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. ACM Transactions on Graphics (TOG) 27(3), 1-10 (2008)", + "56. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 769-777 (2015)", + "57. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: Proceedings of the IEEE International Conference on Computational Photography (ICCP). pp. 1-8 (2013)", + "58. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8174-8182 (2018)", + "59. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9446-9454 (2018)", + "60. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9168-9178 (2021)", + "61. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(3), 3121-3138 (2022)", + "62. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1107-1114 (2013)", + "63. Yan, Y., Ren, W., Guo, Y., Wang, R., Cao, X.: Image deblurring via extreme channels prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4003-4011 (2017)", + "64. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 672-681 (2021)", + "65. Yue, Z., Yong, H., Zhao, Q., Zhang, L., Meng, D., Wong, K.Y.K.: Deep variational network toward blind image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Yue, Z., Zhao, Q., Xie, J., Zhang, L., Meng, D., Wong, K.Y.K.: Blind image superresolution with elaborate degradation modeling on noise and kernel. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2118-2128 (2022)", + "67. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5728-5739 (2022)", + "68. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14821-14831 (2021)", + "69. Zhao, Z., Xiong, B., Wang, L., Ou, Q., Yu, L., Kuang, F.: Retinexdip: A unified deep framework for low-light image enhancement. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) 32(3), 1076-1088 (2022)", + "70. Zhou, X., Mateos, J., Zhou, F., Molina, R., Katsaggelos, A.K.: Variational dirichlet blur kernel estimation. IEEE Transactions on Image Processing (TIP) 24(12), 5127-5139 (2015)", + "71. Zuo, W., Ren, D., Zhang, D., Gu, S., Zhang, L.: Learning iteration-wise generalized shrinkage-thresholding operators for blind deconvolution. IEEE Transactions on Image Processing (TIP) 25(4), 1751-1764 (2016)" + ], + "bbox": [ + 215, + 146, + 787, + 439 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding", + "bbox": [ + 235, + 113, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_model.json b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..711f7922fca7c40833067a4bbd5149381baa40fe --- /dev/null +++ b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_model.json @@ -0,0 +1,3230 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.225, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.212, + 0.677, + 0.245 + ], + "angle": 0, + "content": "Jiangtao Zhang\\(^{1}\\), Zongsheng Yue\\(^{2}\\), Hui Wang\\(^{1}\\), Qian Zhao\\(^{1\\boxtimes}\\), and Deyu Meng\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.255, + 0.65, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Xi'an Jiaotong University, Shaanxi, China" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.27, + 0.682, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\) S-Lab, Nanyang Technological University, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.284, + 0.678, + 0.311 + ], + "angle": 0, + "content": "3 Pazhou Laboratory (Huangpu), Guangdong, China \nzhangjt2021@stu.xjtu.edu.cn, zsyzam@gmail.com," + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.312, + 0.752, + 0.325 + ], + "angle": 0, + "content": "huiwang2019@stu.xjtu.edu.cn, {timmy.zhaoqian, dymeng}@xjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.352, + 0.744, + 0.603 + ], + "angle": 0, + "content": "Abstract. Blind image deconvolution (BID) is a classic yet challenging problem in the field of image processing. Recent advances in deep image prior (DIP) have motivated a series of DIP-based approaches, demonstrating remarkable success in BID. However, due to the high nonconvexity of the inherent optimization process, these methods are notorious for their sensitivity to the initialized kernel. To alleviate this issue and further improve their performance, we propose a new framework for BID that better considers the prior modeling and the initialization for blur kernels, leveraging a deep generative model. The proposed approach pre-trains a generative adversarial network-based kernel generator that aptly characterizes the kernel priors and a kernel initializer that facilitates a well-informed initialization for the blur kernel through latent space encoding. With the pre-trained kernel generator and initializer, one can obtain a high-quality initialization of the blur kernel, and enable optimization within a compact latent kernel manifold. Such a framework results in an evident performance improvement over existing DIP-based BID methods. Extensive experiments on different datasets demonstrate the effectiveness of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.614, + 0.674, + 0.629 + ], + "angle": 0, + "content": "Keywords: Generative kernel prior \\(\\cdot\\) Kernel initializer \\(\\cdot\\) DIP" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.377, + 0.667 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.754 + ], + "angle": 0, + "content": "Blind image deconvolution (BID), also known as deblurring, is a classical problem in image processing [8-10, 14, 29, 34], which aims at recovering the latent clean image from the observed blurry counterpart. This blurry counterpart can be approximately modeled in mathematics as (assuming the blur process is uniform and spatially invariant)" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.755, + 0.786, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} = \\boldsymbol {k} \\otimes \\boldsymbol {x} + \\boldsymbol {n}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.774, + 0.788, + 0.821 + ], + "angle": 0, + "content": "where \\(\\pmb{y}\\) is the blurry image, \\(\\pmb{x}\\) the underlying clean image, \\(\\pmb{k}\\) the blur kernel, \\(\\pmb{n}\\) usually the additive white Gaussian noise (AWGN), and \\(\\otimes\\) denotes the 2D convolution operator. Then the goal of BID is to estimate \\(\\pmb{x}\\) under this degradation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.826, + 0.714, + 0.841 + ], + "angle": 0, + "content": "Code is available at https://github.com/jtaoz/GKPILE-Deconvolution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "model without knowing the ground truth \\( \\pmb{k} \\). As is well-known, this problem is highly ill-posed, since different pairs of \\( \\pmb{x} \\) and \\( \\pmb{k} \\) may result in the same \\( \\pmb{y} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.787, + 0.363 + ], + "angle": 0, + "content": "The BID task was conventionally formulated as an optimization problem, corresponding to a Bayesian posterior estimation problem from the probabilistic perspective. Thus, the maximum a postriori (MAP) [28, 29, 41, 45, 46, 55, 62] and variational inference (VI) [2, 14, 70] approaches were developed for point-wise and distribution-wise estimation, respectively. The core challenge of these approaches is to properly design the priors for both the clean image [28, 41, 45, 46, 62] and blur kernel [42, 70]. However, these handcrafted priors are relatively subjective and may not precisely characterize the intrinsic distributions of natural images and blur kernels in the real world, which limits the performance of the deblurring algorithm. Besides, the resulting optimization is typically endowed with highly non-convex characteristics, leading to potential convergence issues such as reaching a trivial solution of delta kernel or unsatisfactory local minimum [34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Owing to their wide and successful applications, deep learning technologies have been introduced to solve the BID problem. Early explorations [6, 16, 56] embedded deep neural networks (DNNs) into the traditional methods to facilitate the estimation of blur kernels. Recent works have pursued a more direct approach, wherein the mapping from a blurry image to its clean counterpart is directly learned [30, 31, 43, 58, 67, 68]. The latter strategies have achieved promising results on several benchmarks, mainly attributed to the well-designed network architectures. However, these methods tend to overfit the training datasets, and thus may not generalize well to unseen scenarios, particularly for images with large and complex motion blur kernels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Recently, deep image prior (DIP) [59] has exhibited great potential in addressing the task of BID [22, 52], particularly under the scenario of motion blur. The core idea is to parameterize both the blur kernel and clean image as DNNs with random inputs, and then optimize these DNNs through an energy minimization problem conducted by Eq. (1). Compared with the traditional BID methods, these approaches employed more effective DNN-oriented priors and thus achieved superior performance. However, these methods did not fully explore the statistical characteristics of blur kernels, resulting in inaccurate kernel estimations, particularly in cases involving large kernel sizes. Additionally, since the optimization is highly non-convex, the random initialization of the blur kernel often leads to algorithmic instability and compromises the overall performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To alleviate the aforementioned issues of DIP-based methods, we proposed a new approach to better characterize and initialize the blur kernel with deep generative prior (DGP) [7, 48, 60, 64]. Specifically, focusing on the BID of motion blurring, we train a kernel generator using generative adversarial network (GAN) [18] as kernel prior. Owing to GAN's powerful fitting capability on complex distributions, the learned generator is expected to well characterize the blur kernels in a low-dimensional latent space. Then a mapping from the blurry image to the latent code of the kernel generator is trained, acting as a kernel initializer for the subsequent BID task. Intuitively, such a kernel initializer is easier to learn as the latent kernel space is more compact than that of the original kernel. To" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.145, + 0.768, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the latent encoding-based kernel initialization and fine-tuning mechanism in our BID framework. The blur kernel is first initialized from the blurry image by a pre-trained kernel initializer via the latent code \\( \\mathbf{z}_0 \\), which is expected to be close to the code of the ground truth kernel, \\( \\mathbf{z}_{\\mathrm{gt}} \\). The corresponding kernels \\( \\mathbf{k}_0 \\) and \\( \\mathbf{k}_{\\mathrm{gt}} \\) in the kernel manifold is also expected to be close. Then the optimization is performed within the latent space, such that \\( \\mathbf{z}_0 \\) is fine-tuned to \\( \\mathbf{z}_{\\mathrm{opt}} \\), and the final estimated kernel \\( \\mathbf{k}_{\\mathrm{opt}} \\) is closer to the ground truth \\( \\mathbf{k}_{\\mathrm{gt}} \\). See Sec. 3 for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.527 + ], + "angle": 0, + "content": "solve the problem of BID, we first obtain a coarse initialization for the blur kernel by the initializer from the blurry image, and then jointly fine-tune the DIP for the image and the DGP for the blur kernel. The mechanism of the latent encoding-based kernel initialization and fine-tuning is illustrated in Fig. 1. Since the kernel generator sufficiently fits the kernel distribution and the kernel initializer provides a better initial estimation, the overall BID performance can be significantly improved with faster convergence, especially for large blur kernels. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.565 + ], + "angle": 0, + "content": "- We use GAN to learn the DGP for blur kernels, which can better depict the kernel structures and thus offer a more effective and compact kernel prior." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.567, + 0.785, + 0.626 + ], + "angle": 0, + "content": "- We propose to learn a kernel initializer that maps from the blurry image to the latent code of the corresponding kernel. Attributed to the compactness of the latent kernel space, the proposed initializer can provide a more accurate kernel initialization for the subsequent BID process." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.626, + 0.785, + 0.672 + ], + "angle": 0, + "content": "- By harnessing the designed kernel prior and initializer, our BID method achieves state-of-the-art (SotA) performance on the challenging benchmark dataset provided by Lai et al. [33], especially with large blur kernels." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.383, + 0.71 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.482, + 0.74 + ], + "angle": 0, + "content": "2.1 Blind image deconvolution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.804, + 0.779 + ], + "angle": 0, + "content": "BID methods can be roughly divided into two main categories, namely optimization-based methods and deep learning-based methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Optimization-based BID methods. Due to its high ill-posedness, the solution spaces of both the clean image and blur kernel should be properly constrained to make the BID problem solvable. To this aim, traditional optimization-based methods mainly focus on designing appropriate priors for clean images" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "and blur kernels, and plenty of studies have been conducted. For example, total variation [8], hyper-Laplacian prior [28], \\( l_{1} / l_{2} \\)-norm [29], and transform-specific \\( l_{1} \\)-norm [4, 62] were considered to model the image prior in the gradient domain, and patch-based priors [41, 57], low-rank prior [53] and dark/bright channel prior [46, 63] were directly applied in the image domain. As for blur kernels, the non-negative and normalization constraints result in a basic regularization. In addition, focusing on the properties of kernel, sparsity prior [47] and spectral prior [38] have been proposed. To get an ideal solution, tricks such as delayed normalization [50] and multi-scale implementation [71] for blur kernels are also suggested. However, these methods highly rely on handcrafted priors, which may not accurately characterize the intrinsic properties of the clean images and blur kernels, and thus are less competitive in this deep learning era." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.509 + ], + "angle": 0, + "content": "Deep learning-based BID methods. Having witnessed great success in a wide range of applications, deep learning techniques have also been adopted for the BID task. Early methods attempted to replace certain components of the traditional optimization-based methods with DNNs, by virtue of their flexibility. For example, some researchers introduced DNNs as kernel predictors [6,16,56,65]. As the computing resources increase, it is then more popular to directly learn a mapping from the blurred image to its clean counterpart in a fully supervised way, with paired training data and well-designed DNNs [5,21,30,31,43,44,58,67, 68]. Though achieved SotA performance on several benchmark datasets, such as GoPro [43] and RealBlur [54], these methods may encounter the limitations in generalization to images with large complex blur kernels that are not simulated in the pre-defined training sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Recently, focusing on motion blurring, a new type of deep learning methods for the BID problem that leverages the deep priors (will be reviewed in the next subsection), has attracted increasing attention. Along this line, Ren et al. [52] proposed the SelfDeblur method, which adopted DIP to parameterize both the clean image and blur kernel for the first time. After that, multiple variants and extensions, mainly paying attention to the image prior, have been proposed for further improvement of performance [22,35]. These methods achieved promising results in certain cases, especially that they outperform fully supervised deep learning methods on the challenging benchmark by Lai et al. [33]. However, the properties of the blur kernel were not explicitly explored, and the performance tends to be unstable for large kernels due to the non-convex optimization. In contrast to these studies that used DIP with untrained DNNs, Asim et al. [1] proposed to pre-train generators for both the clean images and blur kernels as DGPs, and then fine-tune the generators when doing BID." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.528, + 0.758 + ], + "angle": 0, + "content": "2.2 Deep prior for image processing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Due to their flexibility, DNNs have been used to characterize the image priors in recent years. There are mainly two kinds of such deep priors, namely DIP and DGP, which are briefly reviewed in the following." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Deep image prior. DIP was originally proposed by Ulyanov et al. [59] who trained a DNN to approximate the single target (maybe up to a transformation)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "with random noise as its input. Since there are complex structures and operators involved, the DNN is expected to well depict the manifold of images, playing a similar role as the prior or regularizer in traditional image processing methods. After being raised, DIP has attracted remarkable attention and been applied to various image processing tasks (other than BID mentioned before), including natural image denoising [59], super-resolution [36, 66], inpainting [59], image decomposition [15], low-light enhancement [69], PET image reconstruction [17] and hyperspectral image denoising [39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.432 + ], + "angle": 0, + "content": "Deep generative prior. Attributed to their powerful generation capabilities, the trained deep generators, such as GANs [18], can be seen as approximators to the distributions of images, hoping to provide reasonable priors for images. Such priors can be referred to as DGPs [48]. After a pre-training stage for the generator, DGP can be used as an estimator to approximate the target image via fine-tuning like DIP. There are two common ways to fine-tune DGP. The first is to fix the generator and optimize the input. For example, Menon et al. [40] used this idea for photo upsampling. This strategy is also closely related to GAN-inversion [61]. The second is to jointly fine-tune the latent code and the parameters of the pre-trained generator like [48], which may better locate the optimal estimation within the image manifold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.787, + 0.541 + ], + "angle": 0, + "content": "In addition to directly modeling the images, DGP was also used to characterize the degradation operators, such as blur kernels. For example, Asim et al. [1] proposed to pre-train a variational auto-encoder (VAE) [26] as DGP for motion blur kernels for the BID task; and Liang et al. [36] used the normalizing flow (NF) [11, 12, 27] as the DGP for Gaussian blur kernels, applying to blind superresolution. These studies highlight the effectiveness of DGP for blur kernels, and partially inspire our work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.564, + 0.424, + 0.58 + ], + "angle": 0, + "content": "3 Proposed method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.448, + 0.61 + ], + "angle": 0, + "content": "3.1 DIP-based BID revisit" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.784, + 0.652 + ], + "angle": 0, + "content": "Before presenting our BID method, we first revisit DIP-based BID approaches, or more specifically SelfDeblur [52], and discuss the issues we try to address." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.785, + 0.711 + ], + "angle": 0, + "content": "In SelfDeblur [52], the clean image \\(\\pmb{x}\\) in Eq. (1) was approximated by a convolutional network equipped with encoder-decoder architecture, and the blur kernel \\(\\pmb{k}\\) was estimated by a fully-connected network with one hidden layer. The resulting optimization can then be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.721, + 0.785, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {k}, \\theta_ {x}} \\| G _ {k} \\left(\\boldsymbol {z} _ {k}; \\theta_ {k}\\right) \\otimes G _ {x} \\left(\\boldsymbol {z} _ {x}; \\theta_ {x}\\right) - \\boldsymbol {y} \\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( G_{k}(\\cdot ;\\theta_{k}) \\) and \\( G_{x}(\\cdot ;\\theta_{x}) \\) are networks for the kernel and image, respectively, and \\( \\pmb{z}_k \\) and \\( \\pmb{z}_{x} \\) are input random noise. Note that we ignore the additional constraints imposed on the blur kernel and clean image, positing that these constraints can be readily satisfied through thoughtful network design. Thus, the problem of estimating the blur kernel \\( \\pmb{k} \\) and the clean image \\( \\pmb{x} \\) is reduced to the optimization of parameters \\( \\theta_k^* \\) and \\( \\theta_x^* \\), respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.14, + 0.332, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.211, + 0.332, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.278, + 0.318, + 0.287 + ], + "angle": 0, + "content": "Initial Kernel" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.141, + 0.424, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.211, + 0.424, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.278, + 0.418, + 0.287 + ], + "angle": 0, + "content": "Estimated Kernel" + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.141, + 0.563, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.211, + 0.563, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.279, + 0.533, + 0.288 + ], + "angle": 0, + "content": "Deblurred Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.141, + 0.765, + 0.28 + ], + "angle": 0, + "content": "Fig.2: Illustration of the initialization effect of the DIP-based BID. The two rows correspond to two independent runs of SelfDeblur [52]. From left to right: the randomly initialized kernel, the finally estimated kernel, and the deblurred image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Owing to the powerful capabilities of DNNs, SelfDeblur obtained promising results on the BID task with motion blurring. Notably, it established a new SotA performance on the challenging benchmark introduced by Lai et al. [33]. Huo et al. [22] further improved SelfDeblur by incorporating a more sophisticated DIP within the variational Bayesian framework. However, these DIP-based BID methods did not fully explore the statistical structures or priors for blur kernels, thereby limiting their performance. A more important concern arises from the non-convexity of the optimization, where the randomness of \\( \\mathbf{z}_k \\) causes unstable solutions, particularly when the blur kernel is large. Fig. 2 shows a typical example with two different random runs. Despite the initializations following the same distribution, the estimated blur kernels can be diverse, consequently leading to significantly different deblurring results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.787, + 0.522 + ], + "angle": 0, + "content": "The aforementioned issues redirect our focus toward the blur kernels. Therefore, this study explores more precise statistical structures for blur kernels with DGP, followed by the development of a kernel initializer in the latent space." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.544, + 0.559 + ], + "angle": 0, + "content": "3.2 Overview of the proposed method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.629 + ], + "angle": 0, + "content": "Our proposed method consists of two stages. The first is the pre-training stage, aiming to learn a kernel generator and a kernel initializer. The second stage is to solve the deconvolution problem following the paradigm of the DIP-based BID method, while with the aid of the learned generator and initializer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.794 + ], + "angle": 0, + "content": "In the pre-training stage, we train a kernel generator \\( G_{k}(\\cdot ;\\theta_{k}^{*}) \\) within the GAN [18] framework, where \\( \\theta_{k}^{*} \\) denotes the generator parameters after training. Once learned, it can generate a blur kernel from the randomly sampled noise \\( z_{k} \\). This means the generator indeed models the distribution of the blur kernels and thus characterizes their statistical structures. Such a learned kernel generator can be referred to as the DGP of blur kernels, with \\( z_{k} \\) serving as the kernel latent code. Next, with the generator \\( G_{k}(\\cdot ;\\theta_{k}^{*}) \\) fixed, we train an encoder \\( E(\\cdot ;\\theta_{E}^{*}) \\) that maps a blurry image \\( y \\) to the corresponding kernel latent code through the technique of GAN-inversion [61]. The learned encoder allows for a relatively accurate prediction for the blur kernel based on its latent code, and thus acts as an effective initializer for the kernel DGP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the second stage, we first predict the kernel latent code corresponding to the input blurry image \\(\\pmb{y}\\) using the pre-trained kernel initializer, which provides an initial value for the kernel DGP in the latent space. Then we jointly optimize" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.146, + 0.764, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.334, + 0.785, + 0.363 + ], + "angle": 0, + "content": "Fig. 3: Overview of the proposed BID process, with the pre-trained kernel generator and initializer, of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.414 + ], + "angle": 0, + "content": "the parameters of DIP for the clean image and the latent code (or more precisely its mapped features as discussed in Sec. 3.5) of the DGP for the blur kernel. The overview of the proposed BID process is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.787, + 0.462 + ], + "angle": 0, + "content": "In Sec. 3.3 and Sec. 3.4, we provide details for learning the kernel generator and kernel initializer, respectively. Then in Sec. 3.5, we discuss how to apply the pre-trained generator and initializer to the BID process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.482, + 0.476, + 0.499 + ], + "angle": 0, + "content": "3.3 Kernel generator learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.718 + ], + "angle": 0, + "content": "To learn the kernel generator, a large amount of motion blur kernels are synthesized to simulate their distribution based on the physical generation mechanism proposed in [30] or [33]. Then we train the kernel generator \\( G_{k}(\\cdot;\\theta_{k}^{*}) \\) with these synthesized blur kernels. Note that our kernel generator is made up of several convolutional layers inspired by DCGAN [51]. Though can be seen as a straightforward application of GAN, this step is crucial for our method. On the one hand, the learned generator depicts the statistical structures of blur kernels, serving as a kernel DGP in the BID process; on the other hand, the learned kernel generator also forms the milestone for learning the kernel initializer in a low-dimensional latent space, which will be discussed in Sec. 3.4. Fig. 4 shows several blur kernels generated by \\( G_{k}(\\cdot;\\theta_{k}^{*}) \\) and the physical model according to [33]. As can be observed, the blur kernels generated by our learned generator are morphologically very similar to that synthesized by the physical model, indicating its capability to characterize the underlying manifold of the blur kernels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Remark. It should be mentioned there are some studies also considering learning DGP for blur kernels while with various generative models. For example, Asim et al. [1] used the VAE for the BID task, and Liang et al. [36] adopted the NF for the blind image super-resolution. In this work, we prefer GAN due to the following reasons. First, compared with the vanilla VAE that tends to generate blurry samples [3], GAN can provide sharper ones. Second, in contrast to NF, GAN imposes no constraints on the invertibility between the latent code and the generated sample, thereby facilitating easier learning of the kernel manifold in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.145, + 0.538, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.145, + 0.754, + 0.27 + ], + "angle": 0, + "content": "Fig.4: Comparison of blur kernels synthesized according to physical model and generated by our pre-trained generator. Top row: blur kernels synthesized according to [33]. Bottom row: blur kernels generated by the pre-trained kernel generator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.789, + 0.322 + ], + "angle": 0, + "content": "a lower-dimensional latent space. This aspect is particularly advantageous when considering motion blur kernels, whose sizes can be relatively large (up to \\(75 \\times 75\\) in the Lai dataset [33])." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.343, + 0.476, + 0.358 + ], + "angle": 0, + "content": "3.4 Kernel initializer learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.784, + 0.426 + ], + "angle": 0, + "content": "As discussed before, kernel initialization is important for the overall performance in the BID task since the optimization is highly non-convex. Therefore, we propose to initialize the blur kernel in the latent space using an encoder \\( E(\\cdot; \\theta_E^*) \\) that maps the blurry image \\( \\mathbf{y} \\) to the latent code \\( \\mathbf{z} \\) of the learned kernel DGP." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.428, + 0.758, + 0.442 + ], + "angle": 0, + "content": "A direct way to train the encoder is to solve the following optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.445, + 0.785, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\theta_ {E} ^ {*} = \\arg \\min _ {\\theta_ {E}} \\sum_ {n} \\ell \\left(G _ {k} \\left(E \\left(\\boldsymbol {y} _ {n}; \\theta_ {E}\\right); \\theta_ {k} ^ {*}\\right), \\boldsymbol {k} _ {n}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.621 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}_n\\) denotes the \\(n\\)-th blurry image, \\(\\pmb{k}_n\\) the corresponding blur kernel, \\(\\theta_k^*\\) is the parameters of the pre-trained kernel generator, and \\(\\ell(\\cdot, \\cdot)\\) is some loss function. This optimization can be seen as training a predictor that directly estimates the blur kernel from the blurry image, which is not easy due to the relatively complex structures of blur kernels. On the other hand, if we can find the latent code \\(\\pmb{z}_n\\) of \\(\\pmb{k}_n\\), and learn to predict \\(\\pmb{z}_n\\), the task could be simpler, since the latent space is expected to be much more compact than the original kernel space. Therefore, we propose to use a collaborative learning strategy motivated by [19], to guide the encoder training within the latent space. Specifically, we first formulate the following optimization to train the encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.27, + 0.623, + 0.785, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {E}} \\sum_ {n} \\left\\{\\| G _ {k} \\left(E \\left(\\boldsymbol {y} _ {n}; \\theta_ {E}\\right); \\theta_ {k} ^ {*}\\right) - G _ {k} \\left(\\boldsymbol {z} _ {n}; \\theta_ {k} ^ {*}\\right) \\| _ {1} + \\lambda \\| E \\left(\\boldsymbol {y} _ {n}; \\theta_ {E}\\right) - \\boldsymbol {z} _ {n} \\| _ {2} ^ {2} \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.66, + 0.541, + 0.68 + ], + "angle": 0, + "content": "\\[\ns. t. \\boldsymbol {z} _ {n} = \\arg \\min _ {\\boldsymbol {z}} \\| G _ {k} (\\boldsymbol {z}; \\theta_ {k} ^ {*}) - \\boldsymbol {k} _ {n} \\| _ {1},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.785, + 0.734 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a tuning parameter set to 0.1 in this work. For convenience, we denote the objective functions of the outer and inner optimizations as \\(\\mathcal{L}_E(\\theta_E)\\) and \\(\\mathcal{L}_z^n (z)\\), respectively, in the following." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The optimization of Eq. (4) involves a sub-task of solving \\( \\mathbf{z}_n \\) that corresponds to a GAN-inversion problem [61]. Following [19], we alternatively optimize \\( \\mathbf{z}_n \\) and \\( \\theta_E \\) in each iteration. Specifically, \\( \\mathbf{z}_n \\) is firstly initialized as \\( E(\\mathbf{y}_n; \\theta_E) \\), and then updated by several gradient descent steps with respect to the lower-level objective \\( \\mathcal{L}_z^n(\\mathbf{z}) \\). Based on the updated \\( \\mathbf{z}_n \\), \\( \\theta_E \\) is subsequently optimized according to the upper-level objective \\( \\mathcal{L}_E(\\theta_E) \\), also by gradient descent steps. The whole procedure is summarized in Algorithm 1." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.488, + 0.161 + ], + "angle": 0, + "content": "Algorithm 1 Kernel initializer learning" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.757, + 0.313 + ], + "angle": 0, + "content": "Input: Pre-trained kernel generator \\(G_{k}(\\cdot ;\\theta_{k}^{*})\\) , blurry image-kernel pairs \\(\\{\\pmb {y}_n,\\pmb {k}_n\\}\\) , step size \\(\\epsilon\\) 1: Initialize \\(\\theta_E^{(0)}\\) \n2: for \\(t = 1,\\dots ,T\\) do \n3: For each \\(n\\) .. \n4: \\(\\pmb{z}^{(0)} = E(\\pmb {y}_n;\\theta_E^{(t - 1)})\\) \n5: for \\(s = 1,\\dots ,S\\) do \\(\\pmb{z}^{(s)} = \\pmb{z}^{(s - 1)} - \\epsilon \\nabla_{\\pmb{z}}\\mathcal{L}_{\\pmb{z}}^{n}(\\pmb {z})|_{\\pmb{z} = \\pmb{z}^{(s - 1)}}\\) end for \n6: \\(\\pmb{z}_{n}^{(t)} = \\pmb{z}^{(S)}\\) \n7: \\(\\theta^{(0)} = \\theta_{E}^{(t - 1)}\\) \n8: for \\(l = 1,\\dots ,L\\) do \\(\\theta^{(l)} = \\theta^{(l - 1)} - \\epsilon \\nabla_{\\theta}\\mathcal{L}_{E}(\\theta)|_{\\theta = \\theta^{(l - 1)}}\\) end for \n9: \\(\\theta_E^{(t)} = \\theta^{(L)}\\) \n10: end for \nOutput: Kernel initializer \\(E(\\cdot ;\\theta_E^*)\\) , where \\(\\theta_E^* = \\theta_E^{(T)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.323, + 0.557, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.322, + 0.754, + 0.446 + ], + "angle": 0, + "content": "Fig.5: Illustration of the estimating ability of the pre-trained kernel initializer. Left: the blurry image. Top-right: the estimated blur kernel by the kernel initializer. Bottomright: the ground-truth blur kernel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.546 + ], + "angle": 0, + "content": "Once trained, encoder \\( E(\\cdot; \\theta_E^*) \\) can be regarded as an effective kernel initializer, providing a promising prediction of the blur kernel from the blurry image. Fig. 5 shows an example of the predicted kernel by the learned kernel initializer, in comparison with the ground truth. It can be observed that, though not perfect, the prediction is indeed very close to the ground truth, and thus is expected to be a good starting point for further processing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.367, + 0.58 + ], + "angle": 0, + "content": "3.5 BID Process" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.787, + 0.622 + ], + "angle": 0, + "content": "With the pre-trained kernel generator \\( G_{k}(\\cdot ;\\theta_{k}^{*}) \\) and initializer \\( E(\\cdot ;\\theta_{E}^{*}) \\), we can formulate the BID problem as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.629, + 0.786, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {z} _ {k} ^ {*}, \\theta_ {x} ^ {*}\\right) = \\underset {\\boldsymbol {z} _ {k}, \\theta_ {x}} {\\arg \\min } \\| G _ {k} \\left(\\boldsymbol {z} _ {k}; \\theta_ {k} ^ {*}\\right) \\otimes G _ {x} \\left(\\boldsymbol {z} _ {x}; \\theta_ {x}\\right) - \\boldsymbol {y} \\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.764 + ], + "angle": 0, + "content": "where \\(\\pmb{z}_x\\) is a random vector, \\(\\pmb{z}_k\\) is initialized by \\(\\pmb{z}_k^{(0)} = E(\\pmb{y};\\theta_E^*)\\). After optimization, the desired clean image and blur kernel can be accessed via \\(\\hat{\\pmb{x}} = G_{x}(\\pmb{z}_{x};\\theta_{x}^{*})\\) and \\(\\hat{\\pmb{k}} = G_{k}(\\pmb{z}_{k}^{*};\\theta_{k}^{*})\\), respectively. In fact, such a naive implementation still cannot achieve satisfactory results as empirically illustrated in Sec. 4.3, in particular when the blur kernel is large. This can be attributed to the relatively low dimensionality of \\(\\pmb{z}_k\\), compared with the blur kernel, which enforces too strong constraints and makes the optimization more difficult." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To address this issue, let's first recall one small yet effective trick in StyleGAN [23]. It introduced a network to map the latent code \\( \\mathbf{z} \\) to a style vector \\( \\mathbf{w} \\), enabling more precise control of the generated style. Inspired by this trick, we propose to optimize the feature map of the first layer in the kernel generator \\( G_{k}(\\cdot ;\\theta_{k}^{*}) \\), whose dimension is relatively higher, instead of the latent code \\( \\mathbf{z} \\). By" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.72, + 0.161 + ], + "angle": 0, + "content": "Algorithm 2 BID process with pre-trained kernel generator and initializer" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.787, + 0.294 + ], + "angle": 0, + "content": "Input: Pre-trained kernel generator \\(G_{k}(\\cdot ;\\theta_{k}^{*}) = G_{k}^{(w)}(g_{1}(\\cdot))\\) pre-trained kernel initializer \\(E(\\cdot ;\\theta_E^*)\\) blurry image \\(\\pmb{y}\\) , random vector \\(\\pmb {z}_x\\) , step size \\(\\epsilon\\) \n1: Initialize \\(\\pmb{w}_{k}^{(0)} = g_{1}(E(\\pmb {y};\\theta_{E}^{*}))\\) \\(\\theta_x^{(0)}\\) , denote \\(\\mathcal{L}(\\boldsymbol {w},\\theta) = \\left\\| G_k^{(w)}(\\boldsymbol {w})\\otimes G_x(\\boldsymbol {z}_x;\\theta) - \\boldsymbol {y}\\right\\| ^2\\) \n2: for \\(t = 1,\\dots ,T\\) do \n3: \\(\\begin{array}{r}\\pmb{w}_{k}^{(t)} = \\pmb{w}_{k}^{(t - 1)} - \\epsilon \\nabla_{\\pmb{w}}\\mathcal{L}(\\pmb {w},\\theta)\\big|_{\\pmb {w} = \\pmb{w}_{k}^{(t - 1)},\\theta = \\theta_{x}^{(t - 1)}}\\\\ \\theta_{x}^{(t)} = \\theta_{x}^{(t - 1)} - \\epsilon \\nabla_{\\theta}\\mathcal{L}(\\pmb {w},\\theta)\\big|_{\\pmb {w} = \\pmb{w}_{k}^{(t - 1)},\\theta = \\theta_{x}^{(t - 1)}}\\\\ \\end{array}\\) \n4: end for \n6: \\(\\pmb{w}_{k}^{*} = \\pmb{w}_{k}^{(T)}\\) \\(\\theta_x^* = \\theta_x^{(T)}\\) \nOutput: Estimated clean image \\(\\hat{\\pmb{x}} = G_{x}(\\pmb {z}_{x};\\theta_{x}^{*})\\) , estimated blur kernel \\(\\hat{\\pmb{k}} = G_k^{(w)}(\\pmb {w}_k^*)\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.784, + 0.345 + ], + "angle": 0, + "content": "denoting the first layer of \\( G_{k}(\\cdot ;\\theta_{k}^{*}) \\) as \\( g_{1}(\\cdot) \\) and the reduced generator without \\( g_{1}(\\cdot) \\) as \\( G_{k}^{(w)}(\\cdot) \\), the BID problem of Eq. (5) is reformulated as" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.346, + 0.786, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {w} _ {k} ^ {*}, \\theta_ {x} ^ {*}\\right) = \\underset {\\boldsymbol {w} _ {k}, \\theta_ {x}} {\\arg \\min } \\left\\| G _ {k} ^ {(w)} \\left(\\boldsymbol {w} _ {k}\\right) \\otimes G _ {x} \\left(\\boldsymbol {z} _ {x}; \\theta_ {x}\\right) - \\boldsymbol {y} \\right\\| ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.784, + 0.414 + ], + "angle": 0, + "content": "where \\(\\pmb{w}_k\\) is initialized by \\(\\pmb{w}_k^{(0)} = g_1(E(\\pmb{y};\\theta_E^*))\\). The whole BID process for solving Eq. (6) is summarized in Algorithm 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.505 + ], + "angle": 0, + "content": "It should be noted that the parameters of the generator network can also be fine-tuned in previous studies [48]. However, due to the highly non-convex nature of the BID optimization, too many to-be-optimized parameters may hinder its performance. We thus only update \\(\\pmb{w}_k\\) for the blur kernel in this work. Empirical results in Sec. 4.3 substantiate the superiority of this strategy over updating the whole kernel generator." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.53, + 0.376, + 0.547 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.785, + 0.623 + ], + "angle": 0, + "content": "In this section, we present experimental results to demonstrate the effectiveness of the proposed method and ablation studies to analyze some key components in our framework. More results can be found in the supplementary material, together with discussions on the limitations of this work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.564, + 0.663 + ], + "angle": 0, + "content": "4.1 Datasets and implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Datasets. We evaluate our proposed framework mainly on two datasets: one is synthesized by ourselves, and the other is the well-known BID benchmark constructed by Lai et al. [33]. For the synthetic dataset, we randomly select 80 source images from MSCOCO [37] and apply the motion blur synthesized by [30] on them. The second Lai dataset consists of 25 clean images and 4 large-size blur kernels, resulting in 100 testing blur images in total. Beyond the two synthetic datasets, we also test our method real blurry images collected by Lai et al. [33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Implementation details. As discussed in Sec. 3, our whole framework consists of a pre-training stage and a BID process. In the pre-training stage, we first train the kernel generator \\( G_{k}(\\cdot ;\\theta_{k}^{*}) \\) with DCGAN [51], using kernels synthesized following the way in [30] or [33]. Then we freeze the weights of \\( G_{k}(\\cdot ;\\theta_{k}^{*}) \\) and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.231, + 0.145, + 0.769, + 0.159 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons of various methods on our synthetic dataset." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.161, + 0.782, + 0.226 + ], + "angle": 0, + "content": "
MethodPSNRSSIMMethodPSNRSSIM
Pan et al. [46]24.090.808Zamir et al. [68]22.270.652
Dong et al. [13]23.850.799Ren et al. [52]25.480.765
Tao et al. [58]23.210.699Huo et al. [22]28.130.898
Kupyyn et al. [31]23.460.716Li et al. [35]24.490.744
Kaufman and Fattal [24]25.420.814Ours29.670.928
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.232, + 0.311, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.288, + 0.287, + 0.298 + ], + "angle": 0, + "content": "Blurred" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.233, + 0.406, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.288, + 0.383, + 0.299 + ], + "angle": 0, + "content": "Pan [46]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.233, + 0.5, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.288, + 0.482, + 0.299 + ], + "angle": 0, + "content": "Dong [13]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.233, + 0.596, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.288, + 0.572, + 0.299 + ], + "angle": 0, + "content": "Tao [58]" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.233, + 0.689, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.288, + 0.676, + 0.299 + ], + "angle": 0, + "content": "Kupyn [31]" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.233, + 0.785, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.288, + 0.778, + 0.299 + ], + "angle": 0, + "content": "Kaufman [24]" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.3, + 0.311, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.355, + 0.295, + 0.366 + ], + "angle": 0, + "content": "Zamir [67]" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.3, + 0.406, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.355, + 0.383, + 0.366 + ], + "angle": 0, + "content": "Ren [52]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.3, + 0.5, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.355, + 0.478, + 0.366 + ], + "angle": 0, + "content": "Huo [22]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.3, + 0.596, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.355, + 0.568, + 0.366 + ], + "angle": 0, + "content": "Li [35]" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.3, + 0.689, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.355, + 0.659, + 0.365 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.3, + 0.785, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.355, + 0.778, + 0.365 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.368, + 0.785, + 0.394 + ], + "angle": 0, + "content": "Fig. 6: Visual results on our synthetic dataset. The estimated blur kernel is placed on the top-left corner for each method if available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.552 + ], + "angle": 0, + "content": "train the kernel initializer \\( E(\\cdot; \\theta_E^*) \\) that is a ResNet-18 [20]. To train the kernel initializer, we randomly crop \\( 256 \\times 256 \\) patches from the sources in OpenImages [32] and convolve them with blur kernels to get blurry images. The Adam optimizer [25] with its default configuration in PyTorch [49] is adopted. The initial learning rate is 1e-4 and reduced to 1e-5 as the loss becomes stable and finally to 1e-6. In the BID process, we jointly estimate the blur kernel \\( \\pmb{k} \\) and the clean image \\( \\pmb{x} \\) by optimizing \\( \\pmb{w}_k \\) and \\( \\theta_x \\) according to Eq. (6). The initial learning rate of \\( \\pmb{w}_k \\) and \\( \\theta_x \\) are set as 5e-4 and 1e-2, respectively, and decayed following [52]. For the architecture of the DIP network \\( G_x(\\cdot; \\theta_x) \\), we follow the settings in SelfDeblur [52]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.434, + 0.589 + ], + "angle": 0, + "content": "4.2 Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.704 + ], + "angle": 0, + "content": "Results on our synthetic dataset. We first verify the effectiveness of the proposed method on the synthetic dataset by ourselves. Nine comparison methods are considered, including two traditional model-based ones (Pan et al. [46], Dong et al. [13]), four supervised deep learning ones (Tao et al. [58], Kupyn et al. [31], Kaufman and Fattal [24], Zamir et al. [68]), and three DIP-based ones (Ren et al. [52], Huo et al. [22], Li et al. [35]). We follow the default settings in their papers for these methods or tune them by ourselves for the best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The quantitative results of all competing methods are summarized in Table 1. As can be seen from this table, the proposed method achieves the best performance in terms of both PSNR and SSIM, demonstrating its effectiveness. The superiority of our method can be visually observed from Fig. 6. Specifically, the traditional model-based methods are able to obtain accurate estimations for the blur kernel, but the BID results are not very satisfactory due to the insufficiency of the handcrafted image priors. Most of the supervised deep learning methods generate over-smoothed outcomes, mainly attributed to the lack of exploitation of the physical blur model. Even though achieving relatively better results, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.145, + 0.776, + 0.16 + ], + "angle": 0, + "content": "Table 2: Quantitative results of various methods on the dataset by Lai et al. [33]." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.16, + 0.78, + 0.389 + ], + "angle": 0, + "content": "
MethodManmadeNaturalPeopleSaturatedTextAverage
Cho and Lee [10]16.35/0.38920.14/0.52019.90/0.55614.05/0.49314.87/0.44317.06/0.480
Krishnan et al. [29]15.73/0.40819.44/0.52621.50/0.64714.09/0.50715.40/0.48517.23/0.514
Xu et al. [62]17.99/0.59921.58/0.67924.40/0.81314.53/0.53817.64/0.66819.23/0.659
Perrone and Favaro [50]17.41/0.55121.04/0.67622.77/0.73514.24/0.51116.94/0.59318.48/0.613
Pan et al. [46]18.25/0.56122.15/0.67222.72/0.71016.01/0.59116.69/0.53019.16/0.613
Dong et al. [13]17.30/0.43221.18/0.58621.89/0.65215.85/0.55416.18/0.48018.48/0.541
Tao et al. [58]16.77/0.34219.90/0.45921.41/0.61615.12/0.50615.41/0.43717.72/0.472
Kupyn et al. [31]17.05/0.36720.37/0.48921.70/0.64215.20/0.52215.86/0.47618.04/0.499
Kaufman and Fattal [24]19.36/0.65023.14/0.75626.78/0.87316.40/0.65417.92/0.68320.72/0.723
Zamir et al. [68]16.73/0.34119.80/0.44720.88/0.58715.14/0.50514.49/0.38317.41/0.453
Ren et al. [52]19.00/0.61722.94/0.71526.43/0.78118.38/0.68822.82/0.77621.91/0.715
Huo et al. [22]22.51/0.82026.20/0.90131.38/0.95318.14/0.69127.72/0.93325.19/0.860
Li et al. [35]20.81/0.72124.07/0.78128.45/0.86014.81/0.52723.10/0.83422.25/0.744
Ours24.67/0.94228.12/0.96932.86/0.97318.98/0.73728.91/0.95126.71/0.914
BlurredCho [10]Krishnan [29]Xu [62]Perrone [50]Pan [46]Dong [13]
Kaufman [24]Zamir [67]Ren [52]Huo [22]Li [35]OursGround truth
" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.39, + 0.787, + 0.418 + ], + "angle": 0, + "content": "Fig. 7: Visual results on the dataset by Lai et al. [33]. The estimated blur kernel is placed on the top-left corner of each method if available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.499 + ], + "angle": 0, + "content": "existing DIP-based methods still have some limitations. For example, Ren et al.'s method performs unsatisfactorily, and the structural details in the images by Huo et al. and Li et al.'s methods are not sharp enough. In contrast to these methods, our method can not only accurately estimate the blur kernel, but also recover more image details with fewer distortions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Results on the dataset by Lai et al. [33]. We evaluate our proposed method on the challenging BID benchmark by Lai et al. [33]. This dataset consists of 100 blurry images synthesized by applying 4 blur kernels, with sizes ranging from \\(31 \\times 31\\) to \\(75 \\times 75\\), to 25 clean images. These 25 clean images are categorized into 5 groups, namely Manmade, Natural, People, Saturated and Text. We compare our method with 13 existing methods, including six traditional model-based methods (Cho and Li [10], Krishnan et al. [29], Xu et al. [62], Perrone and Favaro [50], Pan et al [46], Dong et al. [13]), four supervised deep learning methods (Tao et al. [58], Kupyn et al. [31], Kaufman and Fattal [24], Zamir et al. [68]), and three DIP-based methods (Ren et al. [52], Huo et al. [22], Li et al. [35])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Table 2 reports the quantitative results of all competing methods, in terms of PSNR and SSIM. We can see that our method significantly outperforms existing methods in all of the five categories. Notably, our method is the only one that achieves an SSIM value higher than 0.9 on average. The visual results on one typical example are shown in Fig. 7. It can be observed that most of the traditional model-based methods fail to estimate the blur kernel due to the relatively larger kernel size in this dataset. The supervised deep learning methods still produce distorted results as before, because they do not carefully consider the physical blur model. In comparison, DIP-based methods, including ours, can accurately estimate the blur kernel and further achieve promising deblurring results. Among them, our method recovers more details, such as hairs, with reference to the ground truth image, showing its superiority." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.144, + 0.333, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.188, + 0.314, + 0.199 + ], + "angle": 0, + "content": "Blurred" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.145, + 0.417, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.189, + 0.398, + 0.199 + ], + "angle": 0, + "content": "Pan [46]" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.145, + 0.501, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.189, + 0.487, + 0.199 + ], + "angle": 0, + "content": "Dong [13]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.586, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.189, + 0.567, + 0.199 + ], + "angle": 0, + "content": "Tao [58]" + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.145, + 0.669, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.189, + 0.659, + 0.199 + ], + "angle": 0, + "content": "Kupyn [31]" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.145, + 0.753, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.189, + 0.751, + 0.199 + ], + "angle": 0, + "content": "Kaufman [24]" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.2, + 0.374, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.245, + 0.364, + 0.256 + ], + "angle": 0, + "content": "Zamir [67]" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.2, + 0.459, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.245, + 0.442, + 0.256 + ], + "angle": 0, + "content": "Ren [52]" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.2, + 0.543, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.245, + 0.526, + 0.256 + ], + "angle": 0, + "content": "Huo [22]" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.2, + 0.628, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.245, + 0.605, + 0.256 + ], + "angle": 0, + "content": "Li [35]" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.2, + 0.711, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.245, + 0.684, + 0.256 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.257, + 0.785, + 0.285 + ], + "angle": 0, + "content": "Fig. 8: Example results on the real blurry image provided by Lai et al. [33]. The estimated blur kernel is placed on the top-left corner of each method if available." + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.295, + 0.468, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.297, + 0.753, + 0.368 + ], + "angle": 0, + "content": "Fig.9: Performance comparison (PSNR) of the DIP-based BID method (SelfDeblur [52]) and our method in different iteration steps on the dataset by Lai et al. [33]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.443, + 0.785, + 0.504 + ], + "angle": 0, + "content": "Example results on real blurry images. We further apply comparison BID methods to real blurry images collected by Lai et al. [33], and show one example in Fig. 8. As can be seen, on this very challenging image, none of the methods generate perfect results, while our method provides a relatively promising one." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.525, + 0.386, + 0.541 + ], + "angle": 0, + "content": "4.3 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.785, + 0.595 + ], + "angle": 0, + "content": "As shown in the previous experiments, the proposed method has demonstrated its effectiveness on various blurry images. In this section, we provide further analyses for better understanding its advantageous properties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.748 + ], + "angle": 0, + "content": "Convergence speed. One major methodological difference between our method and existing DIP-based methods is that we can initialize the blur kernel with pretrained models instead of random sampling. This property can not only result in more stable and better performance but also improve the convergence speed. To verify this point, we compare the PSNR values of our methods with that of Ren et al.'s [52] across various iterations on the Lai dataset in Fig. 9. It can be seen that the performance of our method in the 400-th iteration is comparable with that of Ren et al.'s method in the 1000-th iteration. This faster convergence can be mostly attributed to the better initialization and characterization of the blur kernel, since both of them adopt the same DIP network for the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Optimization strategy for blur kernel. As discussed in Sec. 3.5, during the BID process, we optimize the higher-dimensional feature map \\(\\boldsymbol{w}_k\\) instead of the original latent code \\(\\boldsymbol{z}_k\\) for the blur kernel. Now we empirically show the different behaviors of the two optimization strategies. As a reference, the results by optimizing the entire kernel generator are also reported, which has a larger search space for optimization. The detailed comparison results on the Lai dataset are" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 3: Performance comparison (PSNR/SSIM) of different kernel optimization strategy on the dataset by Lai et al. [33]." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.173, + 0.787, + 0.217 + ], + "angle": 0, + "content": "
Kernel size31515575Average
wk27.89/0.93226.31/0.90125.91/0.91626.73/0.90826.71/0.914
zk27.11/0.92723.46/0.83022.75/0.82522.43/0.76323.94/0.836
zk and θk24.38/0.83022.33/0.71624.45/0.81924.17/0.81423.83/0.795
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.222, + 0.784, + 0.249 + ], + "angle": 0, + "content": "Table 4: Performance comparison (PSNR/SSIM) of different kernel initialization strategies on the dataset by Lai et al. [33]." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.25, + 0.746, + 0.275 + ], + "angle": 0, + "content": "
BaselineRandomG(z0; θk*)G(E(y; θE*); θk*)
Metrics19.51/0.56323.00/0.74326.71/0.914
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.786, + 0.39 + ], + "angle": 0, + "content": "summarized in Table 3. We can observe that optimizing the whole generator is not competitive due to the too-large search space. When the blur kernel is relatively small, the other two strategies have similar performance, while optimizing \\(\\pmb{w}_k\\) is slightly better. As the blur kernel becomes larger, the performance by optimizing \\(\\pmb{z}_k\\) significantly drops, while optimizing \\(\\pmb{w}_k\\) still produces promising results. These observations demonstrate the reasonability of the proposed optimization strategy and also emphasize the importance of a proper search space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.787, + 0.649 + ], + "angle": 0, + "content": "Effect of kernel initialization. In our BID framework, a pre-trained encoder is employed to initialize the blur kernel. To show the necessity of such an initializer, we compare the proposed initialization strategy to two baselines. The first is the random initialization that randomly samples a latent code \\( \\mathbf{z} \\) for the blur kernel. The second is to initialize the kernel with all elements being the same, which can be seen as an \"average\" kernel, and then find the corresponding \\( \\mathbf{z}_0 \\) by GAN-inversion. It should be noted that both of these two initialization strategies are independent of the blurry image. Table 4 lists the comparison results on the Lai dataset. We can see that the random initialization strategy produces the worst result since the randomly initialized kernel might largely deviate from the target without any constraint. The fixed \"average\" kernel performs better, and indeed already outperforms many other methods in Table 2. That's because this strategy, in some sense, can be regarded as an \"average\" approximation to any blur kernel. Not surprisingly, the proposed initialization with a pre-trained encoder obtains the best performance as the initialized kernel has been very close to the ground-truth one, as shown in Fig. 5. These results substantiate the significant benefits brought by the pre-trained kernel initializer for the BID task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.36, + 0.687 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this paper, we have proposed a new framework for the BID task, by virtue of deep generative model. Within this framework, we first pre-train a kernel generator as a DGP for blur kernels and a kernel initializer that can offer a well-initialized kernel. Then during the BID process, the blur kernel is initialized in the latent feature space and jointly optimized with the DIP network for the final result. Comprehensive experiments have been conducted and demonstrated the effectiveness of the proposed method. The ablation study verifies the necessity of each component in our framework. In the future, we will try to generalize our method to non-uniform blur modeling, overcoming its current limitations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.22 + ], + "angle": 0, + "content": "This work was supported in part by the NSFC Projects (No. 12226004, 62076196, 62331028, 62272375). The authors thank Prof. Jiangxin Dong for providing the code for generating the blur kernels with the type of Lai dataset." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.241, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.309 + ], + "angle": 0, + "content": "1. Asim, M., Shamshad, F., Ahmed, A.: Blind image deconvolution using deep generative priors. IEEE Transactions on Computational Imaging (TCI) 6, 1493-1506 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.351 + ], + "angle": 0, + "content": "2. Bishop, T.E., Babacan, S.D., Amizic, B., Katsaggelos, A.K., Chan, T., Molina, R.: Blind image deconvolution: problem formulation and existing approaches. In: Blind Image Deconvolution, pp. 21-62. CRC press (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.787, + 0.392 + ], + "angle": 0, + "content": "3. Bredell, G., Flouris, K., Chaitanya, K., Erdil, E., Konukoglu, E.: Explicitly minimizing the blur error of variational autoencoders. In: Proceedings of International Conference on Learning Representations (ICLR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.787, + 0.433 + ], + "angle": 0, + "content": "4. Cai, J.F., Ji, H., Liu, C., Shen, Z.: Blind motion deblurring from a single image using sparse approximation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 104-111 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "5. Cai, J., Zuo, W., Zhang, L.: Dark and bright channel prior embedded network for dynamic scene deblurring. IEEE Transactions on Image Processing (TIP) 29, 6885-6897 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.787, + 0.5 + ], + "angle": 0, + "content": "6. Chakrabarti, A.: A neural approach to blind motion deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 221-235 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "7. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14245-14254 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.543, + 0.787, + 0.568 + ], + "angle": 0, + "content": "8. Chan, T.F., Wong, C.K.: Total variation blind deconvolution. IEEE Transactions on Image Processing (TIP) 7(3), 370-375 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "9. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1742-1750 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.787, + 0.636 + ], + "angle": 0, + "content": "0. Cho, S., Lee, S.: Fast motion deblurring. ACM Transactions on Graphics (TOG) 28(5), 1-8 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.787, + 0.663 + ], + "angle": 0, + "content": "1. Dinh, L., Krueger, D., Bengio, Y.: NICE: Non-linear Independent Components Estimation (Apr 2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.787, + 0.703 + ], + "angle": 0, + "content": "2. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using Real NVP. In: Proceedings of International Conference on Learning Representations (ICLR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.787, + 0.745 + ], + "angle": 0, + "content": "3. Dong, J., Pan, J., Su, Z., Yang, M.H.: Blind image deblurring with outlier handling. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2478-2486 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.746, + 0.787, + 0.785 + ], + "angle": 0, + "content": "4. Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T.: Removing camera shake from a single photograph. ACM Transactions on Graphics (TOG) 25(3), 787-794 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Gandelsman, Y., Shocher, A., Irani, M.: \"Double-DIP\": Unsupervised image decomposition via coupled deep-image-priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11018-11027 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "16. Gong, D., Yang, J., Liu, L., Zhang, Y., Reid, I., Shen, C., Van Den Hengel, A., Shi, Q.: From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2319-2328 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "17. Gong, K., Catana, C., Qi, J., Li, Q.: Pet image reconstruction using deep image prior. IEEE Transactions on Medical Imaging (TMI) 38(7), 1655-1665 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.231, + 0.787, + 0.271 + ], + "angle": 0, + "content": "18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in Neural Information Processing Systems (NeurIPS) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.298 + ], + "angle": 0, + "content": "19. Guan, S., Tai, Y., Ni, B., Zhu, F., Huang, F., Yang, X.: Collaborative learning for faster stylegan embedding. arXiv preprint arXiv:2007.01758 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "21. Hradis, M., Kotera, J., Zemcik, P., Sroubek, F.: Convolutional neural networks for direct text deblurring. In: Proceedings of British Machine Vision Conference (BMVC) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "22. Huo, D., Masoumzadeh, A., Kushol, R., Yang, Y.H.: Blind image deconvolution using variational deep image prior. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(10), 11472-11483 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4401-4410 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.502 + ], + "angle": 0, + "content": "24. Kaufman, A., Fattal, R.: Deblurring using analysis-synthesis networks pair. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5811-5820 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.529 + ], + "angle": 0, + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Proceedings of the International Conference for Learning Representations (ICLR) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.556 + ], + "angle": 0, + "content": "26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: Proceedings of International Conference on Learning Representations (ICLR) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.787, + 0.582 + ], + "angle": 0, + "content": "27. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. In: Advances in Neural Information Processing Systems (NeurIPS) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.609 + ], + "angle": 0, + "content": "28. Krishnan, D., Fergus, R.: Fast image deconvolution using hyper-laplacian priors. In: Advances in Neural Information Processing Systems (NeurIPS) (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "29. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 233-240 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.704 + ], + "angle": 0, + "content": "30. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8183-8192 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.745 + ], + "angle": 0, + "content": "31. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8878-8887 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.787, + 0.8 + ], + "angle": 0, + "content": "32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV) 128(7), 1956-1981 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "33. Lai, W.S., Huang, J.B., Hu, Z., Ahuja, N., Yang, M.H.: A comparative study for single image blind deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1701-1709 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "34. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1964-1971 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "35. Li, J., Wang, W., Nan, Y., Ji, H.: Self-supervised blind motion deblurring with deep expectation maximization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13986-13996 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.275 + ], + "angle": 0, + "content": "36. Liang, J., Zhang, K., Gu, S., Gool, L.V., Timofte, R.: Flow-based kernel prior with application to blind super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10596-10605 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.318 + ], + "angle": 0, + "content": "37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 740-755 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.787, + 0.359 + ], + "angle": 0, + "content": "38. Liu, G., Chang, S., Ma, Y.: Blind image deblurring using spectral properties of convolution operators. IEEE Transactions on Image Processing (TIP) 23(12), 5047-5056 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.416 + ], + "angle": 0, + "content": "39. Luo, Y.S., Zhao, X.L., Jiang, T.X., Zheng, Y.B., Chang, Y.: Hyperspectral mixed noise removal via spatial-spectral constrained unsupervised deep image prior. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS) 14, 9435-9449 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.787, + 0.473 + ], + "angle": 0, + "content": "40. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2437-2445 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.515 + ], + "angle": 0, + "content": "41. Michaeli, T., Irani, M.: Blind deblurring using internal patch recurrence. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 783-798 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.572 + ], + "angle": 0, + "content": "42. Molina, R., Katsaggelos, A.K., Abad, J., Mateos, J.: A bayesian approach to blind deconvolution based on dirichlet distributions. In: Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP). vol. 4, pp. 2809-2812 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.787, + 0.614 + ], + "angle": 0, + "content": "43. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3883-3891 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.67 + ], + "angle": 0, + "content": "44. Pan, J., Dong, J., Liu, Y., Zhang, J., Ren, J., Tang, J., Tai, Y.W., Yang, M.H.: Physics-based generative adversarial models for image restoration and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 43(7), 2449-2462 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.713 + ], + "angle": 0, + "content": "45. Pan, J., Hu, Z., Su, Z., Yang, M.H.: Deblurring text images via 10-regularized intensity and gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2901-2908 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "46. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1628-1636 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.787, + 0.798 + ], + "angle": 0, + "content": "47. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Deblurring images via dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(10), 2315-2328 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "48. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 7474-7489 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "49. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.786, + 0.243 + ], + "angle": 0, + "content": "50. Perrone, D., Favaro, P.: Total variation blind deconvolution: The devil is in the details. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2909-2916 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.786, + 0.285 + ], + "angle": 0, + "content": "51. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. In: Proceedings of International Conference on Learning Representations (ICLR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.786, + 0.325 + ], + "angle": 0, + "content": "52. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3341-3350 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.786, + 0.365 + ], + "angle": 0, + "content": "53. Ren, W., Cao, X., Pan, J., Guo, X., Zuo, W., Yang, M.H.: Image deblurring via enhanced low-rank prior. IEEE Transactions on Image Processing (TIP) 25(7), 3426-3437 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.786, + 0.407 + ], + "angle": 0, + "content": "54. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 184-201 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.786, + 0.434 + ], + "angle": 0, + "content": "55. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. ACM Transactions on Graphics (TOG) 27(3), 1-10 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.786, + 0.473 + ], + "angle": 0, + "content": "56. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 769-777 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.786, + 0.514 + ], + "angle": 0, + "content": "57. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: Proceedings of the IEEE International Conference on Computational Photography (ICCP). pp. 1-8 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.786, + 0.555 + ], + "angle": 0, + "content": "58. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8174-8182 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.786, + 0.596 + ], + "angle": 0, + "content": "59. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9446-9454 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.786, + 0.637 + ], + "angle": 0, + "content": "60. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9168-9178 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.786, + 0.677 + ], + "angle": 0, + "content": "61. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(3), 3121-3138 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.786, + 0.717 + ], + "angle": 0, + "content": "62. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1107-1114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.786, + 0.759 + ], + "angle": 0, + "content": "63. Yan, Y., Ren, W., Guo, Y., Wang, R., Cao, X.: Image deblurring via extreme channels prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4003-4011 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "64. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 672-681 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "65. Yue, Z., Yong, H., Zhao, Q., Zhang, L., Meng, D., Wong, K.Y.K.: Deep variational network toward blind image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.236, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "66. Yue, Z., Zhao, Q., Xie, J., Zhang, L., Meng, D., Wong, K.Y.K.: Blind image superresolution with elaborate degradation modeling on noise and kernel. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2118-2128 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.788, + 0.259 + ], + "angle": 0, + "content": "67. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5728-5739 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.788, + 0.313 + ], + "angle": 0, + "content": "68. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14821-14831 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.788, + 0.357 + ], + "angle": 0, + "content": "69. Zhao, Z., Xiong, B., Wang, L., Ou, Q., Yu, L., Kuang, F.: Retinexdip: A unified deep framework for low-light image enhancement. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) 32(3), 1076-1088 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.788, + 0.396 + ], + "angle": 0, + "content": "70. Zhou, X., Mateos, J., Zhou, F., Molina, R., Katsaggelos, A.K.: Variational dirichlet blur kernel estimation. IEEE Transactions on Image Processing (TIP) 24(12), 5127-5139 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.788, + 0.44 + ], + "angle": 0, + "content": "71. Zuo, W., Ren, D., Zhang, D., Gu, S., Zhang, L.: Learning iteration-wise generalized shrinkage-thresholding operators for blind deconvolution. IEEE Transactions on Image Processing (TIP) 25(4), 1751-1764 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.44 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_origin.pdf b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7ab3af1e07e0916e1cb59597a5b6fa1ec163f2a0 --- /dev/null +++ b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fcb7021a96abbce13260b4acf78f102a8b2b9e9ee8867491b02b01aae532eeb +size 18185538 diff --git a/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/full.md b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f9c79aab1bf64b506f23da0abc18adb24019532f --- /dev/null +++ b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/full.md @@ -0,0 +1,420 @@ +# Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding + +Jiangtao Zhang $^{1}$ , Zongsheng Yue $^{2}$ , Hui Wang $^{1}$ , Qian Zhao $^{1\boxtimes}$ , and Deyu Meng $^{1,3}$ + +$^{1}$ Xi'an Jiaotong University, Shaanxi, China + +$^{2}$ S-Lab, Nanyang Technological University, Singapore + +3 Pazhou Laboratory (Huangpu), Guangdong, China +zhangjt2021@stu.xjtu.edu.cn, zsyzam@gmail.com, + +huiwang2019@stu.xjtu.edu.cn, {timmy.zhaoqian, dymeng}@xjtu.edu.cn + +Abstract. Blind image deconvolution (BID) is a classic yet challenging problem in the field of image processing. Recent advances in deep image prior (DIP) have motivated a series of DIP-based approaches, demonstrating remarkable success in BID. However, due to the high nonconvexity of the inherent optimization process, these methods are notorious for their sensitivity to the initialized kernel. To alleviate this issue and further improve their performance, we propose a new framework for BID that better considers the prior modeling and the initialization for blur kernels, leveraging a deep generative model. The proposed approach pre-trains a generative adversarial network-based kernel generator that aptly characterizes the kernel priors and a kernel initializer that facilitates a well-informed initialization for the blur kernel through latent space encoding. With the pre-trained kernel generator and initializer, one can obtain a high-quality initialization of the blur kernel, and enable optimization within a compact latent kernel manifold. Such a framework results in an evident performance improvement over existing DIP-based BID methods. Extensive experiments on different datasets demonstrate the effectiveness of the proposed method. + +Keywords: Generative kernel prior $\cdot$ Kernel initializer $\cdot$ DIP + +# 1 Introduction + +Blind image deconvolution (BID), also known as deblurring, is a classical problem in image processing [8-10, 14, 29, 34], which aims at recovering the latent clean image from the observed blurry counterpart. This blurry counterpart can be approximately modeled in mathematics as (assuming the blur process is uniform and spatially invariant) + +$$ +\boldsymbol {y} = \boldsymbol {k} \otimes \boldsymbol {x} + \boldsymbol {n}, \tag {1} +$$ + +where $\pmb{y}$ is the blurry image, $\pmb{x}$ the underlying clean image, $\pmb{k}$ the blur kernel, $\pmb{n}$ usually the additive white Gaussian noise (AWGN), and $\otimes$ denotes the 2D convolution operator. Then the goal of BID is to estimate $\pmb{x}$ under this degradation + +model without knowing the ground truth $\pmb{k}$ . As is well-known, this problem is highly ill-posed, since different pairs of $\pmb{x}$ and $\pmb{k}$ may result in the same $\pmb{y}$ . + +The BID task was conventionally formulated as an optimization problem, corresponding to a Bayesian posterior estimation problem from the probabilistic perspective. Thus, the maximum a postriori (MAP) [28, 29, 41, 45, 46, 55, 62] and variational inference (VI) [2, 14, 70] approaches were developed for point-wise and distribution-wise estimation, respectively. The core challenge of these approaches is to properly design the priors for both the clean image [28, 41, 45, 46, 62] and blur kernel [42, 70]. However, these handcrafted priors are relatively subjective and may not precisely characterize the intrinsic distributions of natural images and blur kernels in the real world, which limits the performance of the deblurring algorithm. Besides, the resulting optimization is typically endowed with highly non-convex characteristics, leading to potential convergence issues such as reaching a trivial solution of delta kernel or unsatisfactory local minimum [34]. + +Owing to their wide and successful applications, deep learning technologies have been introduced to solve the BID problem. Early explorations [6, 16, 56] embedded deep neural networks (DNNs) into the traditional methods to facilitate the estimation of blur kernels. Recent works have pursued a more direct approach, wherein the mapping from a blurry image to its clean counterpart is directly learned [30, 31, 43, 58, 67, 68]. The latter strategies have achieved promising results on several benchmarks, mainly attributed to the well-designed network architectures. However, these methods tend to overfit the training datasets, and thus may not generalize well to unseen scenarios, particularly for images with large and complex motion blur kernels. + +Recently, deep image prior (DIP) [59] has exhibited great potential in addressing the task of BID [22, 52], particularly under the scenario of motion blur. The core idea is to parameterize both the blur kernel and clean image as DNNs with random inputs, and then optimize these DNNs through an energy minimization problem conducted by Eq. (1). Compared with the traditional BID methods, these approaches employed more effective DNN-oriented priors and thus achieved superior performance. However, these methods did not fully explore the statistical characteristics of blur kernels, resulting in inaccurate kernel estimations, particularly in cases involving large kernel sizes. Additionally, since the optimization is highly non-convex, the random initialization of the blur kernel often leads to algorithmic instability and compromises the overall performance. + +To alleviate the aforementioned issues of DIP-based methods, we proposed a new approach to better characterize and initialize the blur kernel with deep generative prior (DGP) [7, 48, 60, 64]. Specifically, focusing on the BID of motion blurring, we train a kernel generator using generative adversarial network (GAN) [18] as kernel prior. Owing to GAN's powerful fitting capability on complex distributions, the learned generator is expected to well characterize the blur kernels in a low-dimensional latent space. Then a mapping from the blurry image to the latent code of the kernel generator is trained, acting as a kernel initializer for the subsequent BID task. Intuitively, such a kernel initializer is easier to learn as the latent kernel space is more compact than that of the original kernel. To + +![](images/acf3ab469f5203478b2c95f7ecb49f1b9836ee055b8daadce34c439e8dbd809f.jpg) +Fig. 1: Illustration of the latent encoding-based kernel initialization and fine-tuning mechanism in our BID framework. The blur kernel is first initialized from the blurry image by a pre-trained kernel initializer via the latent code $\mathbf{z}_0$ , which is expected to be close to the code of the ground truth kernel, $\mathbf{z}_{\mathrm{gt}}$ . The corresponding kernels $\mathbf{k}_0$ and $\mathbf{k}_{\mathrm{gt}}$ in the kernel manifold is also expected to be close. Then the optimization is performed within the latent space, such that $\mathbf{z}_0$ is fine-tuned to $\mathbf{z}_{\mathrm{opt}}$ , and the final estimated kernel $\mathbf{k}_{\mathrm{opt}}$ is closer to the ground truth $\mathbf{k}_{\mathrm{gt}}$ . See Sec. 3 for more details. + +solve the problem of BID, we first obtain a coarse initialization for the blur kernel by the initializer from the blurry image, and then jointly fine-tune the DIP for the image and the DGP for the blur kernel. The mechanism of the latent encoding-based kernel initialization and fine-tuning is illustrated in Fig. 1. Since the kernel generator sufficiently fits the kernel distribution and the kernel initializer provides a better initial estimation, the overall BID performance can be significantly improved with faster convergence, especially for large blur kernels. Our main contributions are summarized as follows: + +- We use GAN to learn the DGP for blur kernels, which can better depict the kernel structures and thus offer a more effective and compact kernel prior. +- We propose to learn a kernel initializer that maps from the blurry image to the latent code of the corresponding kernel. Attributed to the compactness of the latent kernel space, the proposed initializer can provide a more accurate kernel initialization for the subsequent BID process. +- By harnessing the designed kernel prior and initializer, our BID method achieves state-of-the-art (SotA) performance on the challenging benchmark dataset provided by Lai et al. [33], especially with large blur kernels. + +# 2 Related work + +# 2.1 Blind image deconvolution + +BID methods can be roughly divided into two main categories, namely optimization-based methods and deep learning-based methods. + +Optimization-based BID methods. Due to its high ill-posedness, the solution spaces of both the clean image and blur kernel should be properly constrained to make the BID problem solvable. To this aim, traditional optimization-based methods mainly focus on designing appropriate priors for clean images + +and blur kernels, and plenty of studies have been conducted. For example, total variation [8], hyper-Laplacian prior [28], $l_{1} / l_{2}$ -norm [29], and transform-specific $l_{1}$ -norm [4, 62] were considered to model the image prior in the gradient domain, and patch-based priors [41, 57], low-rank prior [53] and dark/bright channel prior [46, 63] were directly applied in the image domain. As for blur kernels, the non-negative and normalization constraints result in a basic regularization. In addition, focusing on the properties of kernel, sparsity prior [47] and spectral prior [38] have been proposed. To get an ideal solution, tricks such as delayed normalization [50] and multi-scale implementation [71] for blur kernels are also suggested. However, these methods highly rely on handcrafted priors, which may not accurately characterize the intrinsic properties of the clean images and blur kernels, and thus are less competitive in this deep learning era. + +Deep learning-based BID methods. Having witnessed great success in a wide range of applications, deep learning techniques have also been adopted for the BID task. Early methods attempted to replace certain components of the traditional optimization-based methods with DNNs, by virtue of their flexibility. For example, some researchers introduced DNNs as kernel predictors [6,16,56,65]. As the computing resources increase, it is then more popular to directly learn a mapping from the blurred image to its clean counterpart in a fully supervised way, with paired training data and well-designed DNNs [5,21,30,31,43,44,58,67, 68]. Though achieved SotA performance on several benchmark datasets, such as GoPro [43] and RealBlur [54], these methods may encounter the limitations in generalization to images with large complex blur kernels that are not simulated in the pre-defined training sets. + +Recently, focusing on motion blurring, a new type of deep learning methods for the BID problem that leverages the deep priors (will be reviewed in the next subsection), has attracted increasing attention. Along this line, Ren et al. [52] proposed the SelfDeblur method, which adopted DIP to parameterize both the clean image and blur kernel for the first time. After that, multiple variants and extensions, mainly paying attention to the image prior, have been proposed for further improvement of performance [22,35]. These methods achieved promising results in certain cases, especially that they outperform fully supervised deep learning methods on the challenging benchmark by Lai et al. [33]. However, the properties of the blur kernel were not explicitly explored, and the performance tends to be unstable for large kernels due to the non-convex optimization. In contrast to these studies that used DIP with untrained DNNs, Asim et al. [1] proposed to pre-train generators for both the clean images and blur kernels as DGPs, and then fine-tune the generators when doing BID. + +# 2.2 Deep prior for image processing + +Due to their flexibility, DNNs have been used to characterize the image priors in recent years. There are mainly two kinds of such deep priors, namely DIP and DGP, which are briefly reviewed in the following. + +Deep image prior. DIP was originally proposed by Ulyanov et al. [59] who trained a DNN to approximate the single target (maybe up to a transformation) + +with random noise as its input. Since there are complex structures and operators involved, the DNN is expected to well depict the manifold of images, playing a similar role as the prior or regularizer in traditional image processing methods. After being raised, DIP has attracted remarkable attention and been applied to various image processing tasks (other than BID mentioned before), including natural image denoising [59], super-resolution [36, 66], inpainting [59], image decomposition [15], low-light enhancement [69], PET image reconstruction [17] and hyperspectral image denoising [39]. + +Deep generative prior. Attributed to their powerful generation capabilities, the trained deep generators, such as GANs [18], can be seen as approximators to the distributions of images, hoping to provide reasonable priors for images. Such priors can be referred to as DGPs [48]. After a pre-training stage for the generator, DGP can be used as an estimator to approximate the target image via fine-tuning like DIP. There are two common ways to fine-tune DGP. The first is to fix the generator and optimize the input. For example, Menon et al. [40] used this idea for photo upsampling. This strategy is also closely related to GAN-inversion [61]. The second is to jointly fine-tune the latent code and the parameters of the pre-trained generator like [48], which may better locate the optimal estimation within the image manifold. + +In addition to directly modeling the images, DGP was also used to characterize the degradation operators, such as blur kernels. For example, Asim et al. [1] proposed to pre-train a variational auto-encoder (VAE) [26] as DGP for motion blur kernels for the BID task; and Liang et al. [36] used the normalizing flow (NF) [11, 12, 27] as the DGP for Gaussian blur kernels, applying to blind superresolution. These studies highlight the effectiveness of DGP for blur kernels, and partially inspire our work. + +# 3 Proposed method + +# 3.1 DIP-based BID revisit + +Before presenting our BID method, we first revisit DIP-based BID approaches, or more specifically SelfDeblur [52], and discuss the issues we try to address. + +In SelfDeblur [52], the clean image $\pmb{x}$ in Eq. (1) was approximated by a convolutional network equipped with encoder-decoder architecture, and the blur kernel $\pmb{k}$ was estimated by a fully-connected network with one hidden layer. The resulting optimization can then be formulated as + +$$ +\min _ {\theta_ {k}, \theta_ {x}} \| G _ {k} \left(\boldsymbol {z} _ {k}; \theta_ {k}\right) \otimes G _ {x} \left(\boldsymbol {z} _ {x}; \theta_ {x}\right) - \boldsymbol {y} \| ^ {2}, \tag {2} +$$ + +where $G_{k}(\cdot ;\theta_{k})$ and $G_{x}(\cdot ;\theta_{x})$ are networks for the kernel and image, respectively, and $\pmb{z}_k$ and $\pmb{z}_{x}$ are input random noise. Note that we ignore the additional constraints imposed on the blur kernel and clean image, positing that these constraints can be readily satisfied through thoughtful network design. Thus, the problem of estimating the blur kernel $\pmb{k}$ and the clean image $\pmb{x}$ is reduced to the optimization of parameters $\theta_k^*$ and $\theta_x^*$ , respectively. + +![](images/f123bdded2e99e8ca159a3a591f2effcf8bb2866a3eccc2b2b413042916fb096.jpg) + +![](images/5f70aa539da4dd02b6f92c50e2c544a8f15c8dc8175d5b59751fd2223375ffc1.jpg) +Initial Kernel + +![](images/35b152e9ad2545a01a42c3c200b39ea487a51f178b117ac28534965d3cfb28e1.jpg) + +![](images/9d59f219d3afaeca0b9e85689c1ebeb398f3d56f11623fafff8073f6d75d93fc.jpg) +Estimated Kernel + +![](images/2cd98dbca9e15011cc3b5a7b360d1845ee75a0308247d2a444bdd052f8e232df.jpg) +Fig.2: Illustration of the initialization effect of the DIP-based BID. The two rows correspond to two independent runs of SelfDeblur [52]. From left to right: the randomly initialized kernel, the finally estimated kernel, and the deblurred image. + +![](images/ec8aa51fcca5c34bec8a3b25c3e8e99414b6f2f7932c7a408c0bd2a66b22ec82.jpg) +Deblurred Image + +Owing to the powerful capabilities of DNNs, SelfDeblur obtained promising results on the BID task with motion blurring. Notably, it established a new SotA performance on the challenging benchmark introduced by Lai et al. [33]. Huo et al. [22] further improved SelfDeblur by incorporating a more sophisticated DIP within the variational Bayesian framework. However, these DIP-based BID methods did not fully explore the statistical structures or priors for blur kernels, thereby limiting their performance. A more important concern arises from the non-convexity of the optimization, where the randomness of $\mathbf{z}_k$ causes unstable solutions, particularly when the blur kernel is large. Fig. 2 shows a typical example with two different random runs. Despite the initializations following the same distribution, the estimated blur kernels can be diverse, consequently leading to significantly different deblurring results. + +The aforementioned issues redirect our focus toward the blur kernels. Therefore, this study explores more precise statistical structures for blur kernels with DGP, followed by the development of a kernel initializer in the latent space. + +# 3.2 Overview of the proposed method + +Our proposed method consists of two stages. The first is the pre-training stage, aiming to learn a kernel generator and a kernel initializer. The second stage is to solve the deconvolution problem following the paradigm of the DIP-based BID method, while with the aid of the learned generator and initializer. + +In the pre-training stage, we train a kernel generator $G_{k}(\cdot ;\theta_{k}^{*})$ within the GAN [18] framework, where $\theta_{k}^{*}$ denotes the generator parameters after training. Once learned, it can generate a blur kernel from the randomly sampled noise $z_{k}$ . This means the generator indeed models the distribution of the blur kernels and thus characterizes their statistical structures. Such a learned kernel generator can be referred to as the DGP of blur kernels, with $z_{k}$ serving as the kernel latent code. Next, with the generator $G_{k}(\cdot ;\theta_{k}^{*})$ fixed, we train an encoder $E(\cdot ;\theta_{E}^{*})$ that maps a blurry image $y$ to the corresponding kernel latent code through the technique of GAN-inversion [61]. The learned encoder allows for a relatively accurate prediction for the blur kernel based on its latent code, and thus acts as an effective initializer for the kernel DGP. + +In the second stage, we first predict the kernel latent code corresponding to the input blurry image $\pmb{y}$ using the pre-trained kernel initializer, which provides an initial value for the kernel DGP in the latent space. Then we jointly optimize + +![](images/4bb981d60bc2247efa3ff775c17c3195f2a3d374c36329a0e03ca30617ac8a7f.jpg) +Fig. 3: Overview of the proposed BID process, with the pre-trained kernel generator and initializer, of our proposed method. + +the parameters of DIP for the clean image and the latent code (or more precisely its mapped features as discussed in Sec. 3.5) of the DGP for the blur kernel. The overview of the proposed BID process is shown in Fig. 3. + +In Sec. 3.3 and Sec. 3.4, we provide details for learning the kernel generator and kernel initializer, respectively. Then in Sec. 3.5, we discuss how to apply the pre-trained generator and initializer to the BID process. + +# 3.3 Kernel generator learning + +To learn the kernel generator, a large amount of motion blur kernels are synthesized to simulate their distribution based on the physical generation mechanism proposed in [30] or [33]. Then we train the kernel generator $G_{k}(\cdot;\theta_{k}^{*})$ with these synthesized blur kernels. Note that our kernel generator is made up of several convolutional layers inspired by DCGAN [51]. Though can be seen as a straightforward application of GAN, this step is crucial for our method. On the one hand, the learned generator depicts the statistical structures of blur kernels, serving as a kernel DGP in the BID process; on the other hand, the learned kernel generator also forms the milestone for learning the kernel initializer in a low-dimensional latent space, which will be discussed in Sec. 3.4. Fig. 4 shows several blur kernels generated by $G_{k}(\cdot;\theta_{k}^{*})$ and the physical model according to [33]. As can be observed, the blur kernels generated by our learned generator are morphologically very similar to that synthesized by the physical model, indicating its capability to characterize the underlying manifold of the blur kernels. + +Remark. It should be mentioned there are some studies also considering learning DGP for blur kernels while with various generative models. For example, Asim et al. [1] used the VAE for the BID task, and Liang et al. [36] adopted the NF for the blind image super-resolution. In this work, we prefer GAN due to the following reasons. First, compared with the vanilla VAE that tends to generate blurry samples [3], GAN can provide sharper ones. Second, in contrast to NF, GAN imposes no constraints on the invertibility between the latent code and the generated sample, thereby facilitating easier learning of the kernel manifold in + +![](images/6387327c8fceff6e67c5b2b73b81f241725a2cbc44791b4203cd62aa932f8eac.jpg) +Fig.4: Comparison of blur kernels synthesized according to physical model and generated by our pre-trained generator. Top row: blur kernels synthesized according to [33]. Bottom row: blur kernels generated by the pre-trained kernel generator. + +a lower-dimensional latent space. This aspect is particularly advantageous when considering motion blur kernels, whose sizes can be relatively large (up to $75 \times 75$ in the Lai dataset [33]). + +# 3.4 Kernel initializer learning + +As discussed before, kernel initialization is important for the overall performance in the BID task since the optimization is highly non-convex. Therefore, we propose to initialize the blur kernel in the latent space using an encoder $E(\cdot; \theta_E^*)$ that maps the blurry image $\mathbf{y}$ to the latent code $\mathbf{z}$ of the learned kernel DGP. + +A direct way to train the encoder is to solve the following optimization: + +$$ +\theta_ {E} ^ {*} = \arg \min _ {\theta_ {E}} \sum_ {n} \ell \left(G _ {k} \left(E \left(\boldsymbol {y} _ {n}; \theta_ {E}\right); \theta_ {k} ^ {*}\right), \boldsymbol {k} _ {n}\right), \tag {3} +$$ + +where $\mathbf{y}_n$ denotes the $n$ -th blurry image, $\pmb{k}_n$ the corresponding blur kernel, $\theta_k^*$ is the parameters of the pre-trained kernel generator, and $\ell(\cdot, \cdot)$ is some loss function. This optimization can be seen as training a predictor that directly estimates the blur kernel from the blurry image, which is not easy due to the relatively complex structures of blur kernels. On the other hand, if we can find the latent code $\pmb{z}_n$ of $\pmb{k}_n$ , and learn to predict $\pmb{z}_n$ , the task could be simpler, since the latent space is expected to be much more compact than the original kernel space. Therefore, we propose to use a collaborative learning strategy motivated by [19], to guide the encoder training within the latent space. Specifically, we first formulate the following optimization to train the encoder: + +$$ +\min _ {\theta_ {E}} \sum_ {n} \left\{\| G _ {k} \left(E \left(\boldsymbol {y} _ {n}; \theta_ {E}\right); \theta_ {k} ^ {*}\right) - G _ {k} \left(\boldsymbol {z} _ {n}; \theta_ {k} ^ {*}\right) \| _ {1} + \lambda \| E \left(\boldsymbol {y} _ {n}; \theta_ {E}\right) - \boldsymbol {z} _ {n} \| _ {2} ^ {2} \right\}, \tag {4} +$$ + +$$ +s. t. \boldsymbol {z} _ {n} = \arg \min _ {\boldsymbol {z}} \| G _ {k} (\boldsymbol {z}; \theta_ {k} ^ {*}) - \boldsymbol {k} _ {n} \| _ {1}, +$$ + +where $\lambda$ is a tuning parameter set to 0.1 in this work. For convenience, we denote the objective functions of the outer and inner optimizations as $\mathcal{L}_E(\theta_E)$ and $\mathcal{L}_z^n (z)$ , respectively, in the following. + +The optimization of Eq. (4) involves a sub-task of solving $\mathbf{z}_n$ that corresponds to a GAN-inversion problem [61]. Following [19], we alternatively optimize $\mathbf{z}_n$ and $\theta_E$ in each iteration. Specifically, $\mathbf{z}_n$ is firstly initialized as $E(\mathbf{y}_n; \theta_E)$ , and then updated by several gradient descent steps with respect to the lower-level objective $\mathcal{L}_z^n(\mathbf{z})$ . Based on the updated $\mathbf{z}_n$ , $\theta_E$ is subsequently optimized according to the upper-level objective $\mathcal{L}_E(\theta_E)$ , also by gradient descent steps. The whole procedure is summarized in Algorithm 1. + +Algorithm 1 Kernel initializer learning +Input: Pre-trained kernel generator $G_{k}(\cdot ;\theta_{k}^{*})$ , blurry image-kernel pairs $\{\pmb {y}_n,\pmb {k}_n\}$ , step size $\epsilon$ 1: Initialize $\theta_E^{(0)}$ +2: for $t = 1,\dots ,T$ do +3: For each $n$ .. +4: $\pmb{z}^{(0)} = E(\pmb {y}_n;\theta_E^{(t - 1)})$ +5: for $s = 1,\dots ,S$ do $\pmb{z}^{(s)} = \pmb{z}^{(s - 1)} - \epsilon \nabla_{\pmb{z}}\mathcal{L}_{\pmb{z}}^{n}(\pmb {z})|_{\pmb{z} = \pmb{z}^{(s - 1)}}$ end for +6: $\pmb{z}_{n}^{(t)} = \pmb{z}^{(S)}$ +7: $\theta^{(0)} = \theta_{E}^{(t - 1)}$ +8: for $l = 1,\dots ,L$ do $\theta^{(l)} = \theta^{(l - 1)} - \epsilon \nabla_{\theta}\mathcal{L}_{E}(\theta)|_{\theta = \theta^{(l - 1)}}$ end for +9: $\theta_E^{(t)} = \theta^{(L)}$ +10: end for +Output: Kernel initializer $E(\cdot ;\theta_E^*)$ , where $\theta_E^* = \theta_E^{(T)}$ + +![](images/5b55c6a713502069a8cda2752d5574d5ada2bcb1a3531085fabaffd27208ee70.jpg) +Fig.5: Illustration of the estimating ability of the pre-trained kernel initializer. Left: the blurry image. Top-right: the estimated blur kernel by the kernel initializer. Bottomright: the ground-truth blur kernel. + +Once trained, encoder $E(\cdot; \theta_E^*)$ can be regarded as an effective kernel initializer, providing a promising prediction of the blur kernel from the blurry image. Fig. 5 shows an example of the predicted kernel by the learned kernel initializer, in comparison with the ground truth. It can be observed that, though not perfect, the prediction is indeed very close to the ground truth, and thus is expected to be a good starting point for further processing. + +# 3.5 BID Process + +With the pre-trained kernel generator $G_{k}(\cdot ;\theta_{k}^{*})$ and initializer $E(\cdot ;\theta_{E}^{*})$ , we can formulate the BID problem as follows: + +$$ +\left(\boldsymbol {z} _ {k} ^ {*}, \theta_ {x} ^ {*}\right) = \underset {\boldsymbol {z} _ {k}, \theta_ {x}} {\arg \min } \| G _ {k} \left(\boldsymbol {z} _ {k}; \theta_ {k} ^ {*}\right) \otimes G _ {x} \left(\boldsymbol {z} _ {x}; \theta_ {x}\right) - \boldsymbol {y} \| ^ {2}, \tag {5} +$$ + +where $\pmb{z}_x$ is a random vector, $\pmb{z}_k$ is initialized by $\pmb{z}_k^{(0)} = E(\pmb{y};\theta_E^*)$ . After optimization, the desired clean image and blur kernel can be accessed via $\hat{\pmb{x}} = G_{x}(\pmb{z}_{x};\theta_{x}^{*})$ and $\hat{\pmb{k}} = G_{k}(\pmb{z}_{k}^{*};\theta_{k}^{*})$ , respectively. In fact, such a naive implementation still cannot achieve satisfactory results as empirically illustrated in Sec. 4.3, in particular when the blur kernel is large. This can be attributed to the relatively low dimensionality of $\pmb{z}_k$ , compared with the blur kernel, which enforces too strong constraints and makes the optimization more difficult. + +To address this issue, let's first recall one small yet effective trick in StyleGAN [23]. It introduced a network to map the latent code $\mathbf{z}$ to a style vector $\mathbf{w}$ , enabling more precise control of the generated style. Inspired by this trick, we propose to optimize the feature map of the first layer in the kernel generator $G_{k}(\cdot ;\theta_{k}^{*})$ , whose dimension is relatively higher, instead of the latent code $\mathbf{z}$ . By + +Algorithm 2 BID process with pre-trained kernel generator and initializer +Input: Pre-trained kernel generator $G_{k}(\cdot ;\theta_{k}^{*}) = G_{k}^{(w)}(g_{1}(\cdot))$ pre-trained kernel initializer $E(\cdot ;\theta_E^*)$ blurry image $\pmb{y}$ , random vector $\pmb {z}_x$ , step size $\epsilon$ +1: Initialize $\pmb{w}_{k}^{(0)} = g_{1}(E(\pmb {y};\theta_{E}^{*}))$ $\theta_x^{(0)}$ , denote $\mathcal{L}(\boldsymbol {w},\theta) = \left\| G_k^{(w)}(\boldsymbol {w})\otimes G_x(\boldsymbol {z}_x;\theta) - \boldsymbol {y}\right\| ^2$ +2: for $t = 1,\dots ,T$ do +3: $\begin{array}{r}\pmb{w}_{k}^{(t)} = \pmb{w}_{k}^{(t - 1)} - \epsilon \nabla_{\pmb{w}}\mathcal{L}(\pmb {w},\theta)\big|_{\pmb {w} = \pmb{w}_{k}^{(t - 1)},\theta = \theta_{x}^{(t - 1)}}\\ \theta_{x}^{(t)} = \theta_{x}^{(t - 1)} - \epsilon \nabla_{\theta}\mathcal{L}(\pmb {w},\theta)\big|_{\pmb {w} = \pmb{w}_{k}^{(t - 1)},\theta = \theta_{x}^{(t - 1)}}\\ \end{array}$ +4: end for +6: $\pmb{w}_{k}^{*} = \pmb{w}_{k}^{(T)}$ $\theta_x^* = \theta_x^{(T)}$ +Output: Estimated clean image $\hat{\pmb{x}} = G_{x}(\pmb {z}_{x};\theta_{x}^{*})$ , estimated blur kernel $\hat{\pmb{k}} = G_k^{(w)}(\pmb {w}_k^*)$ + +denoting the first layer of $G_{k}(\cdot ;\theta_{k}^{*})$ as $g_{1}(\cdot)$ and the reduced generator without $g_{1}(\cdot)$ as $G_{k}^{(w)}(\cdot)$ , the BID problem of Eq. (5) is reformulated as + +$$ +\left(\boldsymbol {w} _ {k} ^ {*}, \theta_ {x} ^ {*}\right) = \underset {\boldsymbol {w} _ {k}, \theta_ {x}} {\arg \min } \left\| G _ {k} ^ {(w)} \left(\boldsymbol {w} _ {k}\right) \otimes G _ {x} \left(\boldsymbol {z} _ {x}; \theta_ {x}\right) - \boldsymbol {y} \right\| ^ {2}, \tag {6} +$$ + +where $\pmb{w}_k$ is initialized by $\pmb{w}_k^{(0)} = g_1(E(\pmb{y};\theta_E^*))$ . The whole BID process for solving Eq. (6) is summarized in Algorithm 2. + +It should be noted that the parameters of the generator network can also be fine-tuned in previous studies [48]. However, due to the highly non-convex nature of the BID optimization, too many to-be-optimized parameters may hinder its performance. We thus only update $\pmb{w}_k$ for the blur kernel in this work. Empirical results in Sec. 4.3 substantiate the superiority of this strategy over updating the whole kernel generator. + +# 4 Experiments + +In this section, we present experimental results to demonstrate the effectiveness of the proposed method and ablation studies to analyze some key components in our framework. More results can be found in the supplementary material, together with discussions on the limitations of this work. + +# 4.1 Datasets and implementation details + +Datasets. We evaluate our proposed framework mainly on two datasets: one is synthesized by ourselves, and the other is the well-known BID benchmark constructed by Lai et al. [33]. For the synthetic dataset, we randomly select 80 source images from MSCOCO [37] and apply the motion blur synthesized by [30] on them. The second Lai dataset consists of 25 clean images and 4 large-size blur kernels, resulting in 100 testing blur images in total. Beyond the two synthetic datasets, we also test our method real blurry images collected by Lai et al. [33]. + +Implementation details. As discussed in Sec. 3, our whole framework consists of a pre-training stage and a BID process. In the pre-training stage, we first train the kernel generator $G_{k}(\cdot ;\theta_{k}^{*})$ with DCGAN [51], using kernels synthesized following the way in [30] or [33]. Then we freeze the weights of $G_{k}(\cdot ;\theta_{k}^{*})$ and + +Table 1: Quantitative comparisons of various methods on our synthetic dataset. + +
MethodPSNRSSIMMethodPSNRSSIM
Pan et al. [46]24.090.808Zamir et al. [68]22.270.652
Dong et al. [13]23.850.799Ren et al. [52]25.480.765
Tao et al. [58]23.210.699Huo et al. [22]28.130.898
Kupyyn et al. [31]23.460.716Li et al. [35]24.490.744
Kaufman and Fattal [24]25.420.814Ours29.670.928
+ +![](images/43b8f0b5ebbebc77bc31136610cf7ee93a26f98090309b37e600c1d3c287d131.jpg) +Blurred + +![](images/af285bc82f7db4dae76fa6253c77d3b9818da4a493f338fa2ff83d2041e1c647.jpg) +Pan [46] + +![](images/b393c5771c8a45fac7c1f70625e7a4d79251c3546bd2ffeb8c17d655d8ea7f39.jpg) +Dong [13] + +![](images/aae5f2cb585eea31b31ac13fc0d902c40cba9c4bdaae571a06218475c1fced5f.jpg) +Tao [58] + +![](images/fcaae1773cf4ef864ef08222304fd4a05d4265cc259ea562e24c829ee5311746.jpg) + +![](images/dd9c4692da8563a9d018080eadcf94875eb3563fd91df4f161f9899c9ccbf358.jpg) + +![](images/52c94a73e39b52b2928cd415dcca4f807d8740c471a8ce7663434f0cacbad417.jpg) +Zamir [67] + +![](images/25f9fe1055baa8a08732db49038e093c883ea5d42991c2b10e8a34c8f78b1b03.jpg) +Ren [52] + +![](images/99232cba2cc3c7f21c5649c049949f01c1f8d164a30860d74615b7f10d3fc736.jpg) +Huo [22] +Fig. 6: Visual results on our synthetic dataset. The estimated blur kernel is placed on the top-left corner for each method if available. + +![](images/452d85d7c795f606bfa56b598783ddbbd8cbe9b5259179d59e0018384d9c8875.jpg) +Li [35] + +![](images/ff405bbabdbfe218d7d8e9754d8feed9a5573812e86bdf538af7be5a83aab8c7.jpg) +Kupyn [31] +Ours + +![](images/3e06843ae7954a2f6466dc400f8b293317c556db22518cf4af43c54581da9dbb.jpg) +Kaufman [24] +Ground truth + +train the kernel initializer $E(\cdot; \theta_E^*)$ that is a ResNet-18 [20]. To train the kernel initializer, we randomly crop $256 \times 256$ patches from the sources in OpenImages [32] and convolve them with blur kernels to get blurry images. The Adam optimizer [25] with its default configuration in PyTorch [49] is adopted. The initial learning rate is 1e-4 and reduced to 1e-5 as the loss becomes stable and finally to 1e-6. In the BID process, we jointly estimate the blur kernel $\pmb{k}$ and the clean image $\pmb{x}$ by optimizing $\pmb{w}_k$ and $\theta_x$ according to Eq. (6). The initial learning rate of $\pmb{w}_k$ and $\theta_x$ are set as 5e-4 and 1e-2, respectively, and decayed following [52]. For the architecture of the DIP network $G_x(\cdot; \theta_x)$ , we follow the settings in SelfDeblur [52]. + +# 4.2 Experimental results + +Results on our synthetic dataset. We first verify the effectiveness of the proposed method on the synthetic dataset by ourselves. Nine comparison methods are considered, including two traditional model-based ones (Pan et al. [46], Dong et al. [13]), four supervised deep learning ones (Tao et al. [58], Kupyn et al. [31], Kaufman and Fattal [24], Zamir et al. [68]), and three DIP-based ones (Ren et al. [52], Huo et al. [22], Li et al. [35]). We follow the default settings in their papers for these methods or tune them by ourselves for the best performance. + +The quantitative results of all competing methods are summarized in Table 1. As can be seen from this table, the proposed method achieves the best performance in terms of both PSNR and SSIM, demonstrating its effectiveness. The superiority of our method can be visually observed from Fig. 6. Specifically, the traditional model-based methods are able to obtain accurate estimations for the blur kernel, but the BID results are not very satisfactory due to the insufficiency of the handcrafted image priors. Most of the supervised deep learning methods generate over-smoothed outcomes, mainly attributed to the lack of exploitation of the physical blur model. Even though achieving relatively better results, the + +Table 2: Quantitative results of various methods on the dataset by Lai et al. [33]. + +
MethodManmadeNaturalPeopleSaturatedTextAverage
Cho and Lee [10]16.35/0.38920.14/0.52019.90/0.55614.05/0.49314.87/0.44317.06/0.480
Krishnan et al. [29]15.73/0.40819.44/0.52621.50/0.64714.09/0.50715.40/0.48517.23/0.514
Xu et al. [62]17.99/0.59921.58/0.67924.40/0.81314.53/0.53817.64/0.66819.23/0.659
Perrone and Favaro [50]17.41/0.55121.04/0.67622.77/0.73514.24/0.51116.94/0.59318.48/0.613
Pan et al. [46]18.25/0.56122.15/0.67222.72/0.71016.01/0.59116.69/0.53019.16/0.613
Dong et al. [13]17.30/0.43221.18/0.58621.89/0.65215.85/0.55416.18/0.48018.48/0.541
Tao et al. [58]16.77/0.34219.90/0.45921.41/0.61615.12/0.50615.41/0.43717.72/0.472
Kupyn et al. [31]17.05/0.36720.37/0.48921.70/0.64215.20/0.52215.86/0.47618.04/0.499
Kaufman and Fattal [24]19.36/0.65023.14/0.75626.78/0.87316.40/0.65417.92/0.68320.72/0.723
Zamir et al. [68]16.73/0.34119.80/0.44720.88/0.58715.14/0.50514.49/0.38317.41/0.453
Ren et al. [52]19.00/0.61722.94/0.71526.43/0.78118.38/0.68822.82/0.77621.91/0.715
Huo et al. [22]22.51/0.82026.20/0.90131.38/0.95318.14/0.69127.72/0.93325.19/0.860
Li et al. [35]20.81/0.72124.07/0.78128.45/0.86014.81/0.52723.10/0.83422.25/0.744
Ours24.67/0.94228.12/0.96932.86/0.97318.98/0.73728.91/0.95126.71/0.914
BlurredCho [10]Krishnan [29]Xu [62]Perrone [50]Pan [46]Dong [13]
Kaufman [24]Zamir [67]Ren [52]Huo [22]Li [35]OursGround truth
+ +Fig. 7: Visual results on the dataset by Lai et al. [33]. The estimated blur kernel is placed on the top-left corner of each method if available. + +existing DIP-based methods still have some limitations. For example, Ren et al.'s method performs unsatisfactorily, and the structural details in the images by Huo et al. and Li et al.'s methods are not sharp enough. In contrast to these methods, our method can not only accurately estimate the blur kernel, but also recover more image details with fewer distortions. + +Results on the dataset by Lai et al. [33]. We evaluate our proposed method on the challenging BID benchmark by Lai et al. [33]. This dataset consists of 100 blurry images synthesized by applying 4 blur kernels, with sizes ranging from $31 \times 31$ to $75 \times 75$ , to 25 clean images. These 25 clean images are categorized into 5 groups, namely Manmade, Natural, People, Saturated and Text. We compare our method with 13 existing methods, including six traditional model-based methods (Cho and Li [10], Krishnan et al. [29], Xu et al. [62], Perrone and Favaro [50], Pan et al [46], Dong et al. [13]), four supervised deep learning methods (Tao et al. [58], Kupyn et al. [31], Kaufman and Fattal [24], Zamir et al. [68]), and three DIP-based methods (Ren et al. [52], Huo et al. [22], Li et al. [35]). + +Table 2 reports the quantitative results of all competing methods, in terms of PSNR and SSIM. We can see that our method significantly outperforms existing methods in all of the five categories. Notably, our method is the only one that achieves an SSIM value higher than 0.9 on average. The visual results on one typical example are shown in Fig. 7. It can be observed that most of the traditional model-based methods fail to estimate the blur kernel due to the relatively larger kernel size in this dataset. The supervised deep learning methods still produce distorted results as before, because they do not carefully consider the physical blur model. In comparison, DIP-based methods, including ours, can accurately estimate the blur kernel and further achieve promising deblurring results. Among them, our method recovers more details, such as hairs, with reference to the ground truth image, showing its superiority. + +![](images/fc7cceebc0da02d3d809e6c86bd3ab2625120508bbecd04b0b50727fc4a9760f.jpg) +Blurred + +![](images/fd2971b66365791da9ea173a4ff31d367aca1ec139aecb9325500964eb6dba0b.jpg) +Pan [46] + +![](images/6947626b66b3e2b2be3680527f9e40502b414739e2177ae844165fe0878b2ab4.jpg) + +![](images/00f15976fa312fb43b9571d1acb2ffb6294bdc0e22588db5430b16435b11df4e.jpg) + +![](images/8d8154e4e672373f55984ed03142bbabe49ace304d3021a14244ce625ee3f785.jpg) + +![](images/7ace74815d9bd1b1bbc9f6516cbeaf182f28aef7c861931e8ecad437f2b2a43d.jpg) + +![](images/297dd9de7ad9871b3605a403349973206ebaa15c1f53ebc25fba314ebfbd9d3e.jpg) +Zamir [67] + +![](images/33ea5e7566dbd8bfd5c56a0897edd31cace5b6d504dace6aa07bc5a8355a98c4.jpg) +Dong [13] +Ren [52] + +![](images/79482bbf21205a7733c160c8bfe4524911babea121430d68c474b45fe4db9796.jpg) +Tao [58] +Huo [22] + +![](images/7faa62d02101d97806e48e440dc64acc46992d65113093dde0e909e457b2a23b.jpg) +Kupyn [31] +Li [35] + +![](images/0b7d3b82dab6bb0f200c43fba73f74a3daf88f6cf1a517aa23cef40d81663e3a.jpg) +Kaufman [24] +Ours + +![](images/3f2c27a01f07e76cb7bac61d96328694bc5d8056011dfe647b0d9d69f545bdc2.jpg) +Fig. 8: Example results on the real blurry image provided by Lai et al. [33]. The estimated blur kernel is placed on the top-left corner of each method if available. +Fig.9: Performance comparison (PSNR) of the DIP-based BID method (SelfDeblur [52]) and our method in different iteration steps on the dataset by Lai et al. [33]. + +Example results on real blurry images. We further apply comparison BID methods to real blurry images collected by Lai et al. [33], and show one example in Fig. 8. As can be seen, on this very challenging image, none of the methods generate perfect results, while our method provides a relatively promising one. + +# 4.3 Ablation study + +As shown in the previous experiments, the proposed method has demonstrated its effectiveness on various blurry images. In this section, we provide further analyses for better understanding its advantageous properties. + +Convergence speed. One major methodological difference between our method and existing DIP-based methods is that we can initialize the blur kernel with pretrained models instead of random sampling. This property can not only result in more stable and better performance but also improve the convergence speed. To verify this point, we compare the PSNR values of our methods with that of Ren et al.'s [52] across various iterations on the Lai dataset in Fig. 9. It can be seen that the performance of our method in the 400-th iteration is comparable with that of Ren et al.'s method in the 1000-th iteration. This faster convergence can be mostly attributed to the better initialization and characterization of the blur kernel, since both of them adopt the same DIP network for the image. + +Optimization strategy for blur kernel. As discussed in Sec. 3.5, during the BID process, we optimize the higher-dimensional feature map $\boldsymbol{w}_k$ instead of the original latent code $\boldsymbol{z}_k$ for the blur kernel. Now we empirically show the different behaviors of the two optimization strategies. As a reference, the results by optimizing the entire kernel generator are also reported, which has a larger search space for optimization. The detailed comparison results on the Lai dataset are + +Table 3: Performance comparison (PSNR/SSIM) of different kernel optimization strategy on the dataset by Lai et al. [33]. + +
Kernel size31515575Average
wk27.89/0.93226.31/0.90125.91/0.91626.73/0.90826.71/0.914
zk27.11/0.92723.46/0.83022.75/0.82522.43/0.76323.94/0.836
zk and θk24.38/0.83022.33/0.71624.45/0.81924.17/0.81423.83/0.795
+ +Table 4: Performance comparison (PSNR/SSIM) of different kernel initialization strategies on the dataset by Lai et al. [33]. + +
BaselineRandomG(z0; θk*)G(E(y; θE*); θk*)
Metrics19.51/0.56323.00/0.74326.71/0.914
+ +summarized in Table 3. We can observe that optimizing the whole generator is not competitive due to the too-large search space. When the blur kernel is relatively small, the other two strategies have similar performance, while optimizing $\pmb{w}_k$ is slightly better. As the blur kernel becomes larger, the performance by optimizing $\pmb{z}_k$ significantly drops, while optimizing $\pmb{w}_k$ still produces promising results. These observations demonstrate the reasonability of the proposed optimization strategy and also emphasize the importance of a proper search space. + +Effect of kernel initialization. In our BID framework, a pre-trained encoder is employed to initialize the blur kernel. To show the necessity of such an initializer, we compare the proposed initialization strategy to two baselines. The first is the random initialization that randomly samples a latent code $\mathbf{z}$ for the blur kernel. The second is to initialize the kernel with all elements being the same, which can be seen as an "average" kernel, and then find the corresponding $\mathbf{z}_0$ by GAN-inversion. It should be noted that both of these two initialization strategies are independent of the blurry image. Table 4 lists the comparison results on the Lai dataset. We can see that the random initialization strategy produces the worst result since the randomly initialized kernel might largely deviate from the target without any constraint. The fixed "average" kernel performs better, and indeed already outperforms many other methods in Table 2. That's because this strategy, in some sense, can be regarded as an "average" approximation to any blur kernel. Not surprisingly, the proposed initialization with a pre-trained encoder obtains the best performance as the initialized kernel has been very close to the ground-truth one, as shown in Fig. 5. These results substantiate the significant benefits brought by the pre-trained kernel initializer for the BID task. + +# 5 Conclusion + +In this paper, we have proposed a new framework for the BID task, by virtue of deep generative model. Within this framework, we first pre-train a kernel generator as a DGP for blur kernels and a kernel initializer that can offer a well-initialized kernel. Then during the BID process, the blur kernel is initialized in the latent feature space and jointly optimized with the DIP network for the final result. Comprehensive experiments have been conducted and demonstrated the effectiveness of the proposed method. The ablation study verifies the necessity of each component in our framework. In the future, we will try to generalize our method to non-uniform blur modeling, overcoming its current limitations. + +# Acknowledgments + +This work was supported in part by the NSFC Projects (No. 12226004, 62076196, 62331028, 62272375). The authors thank Prof. Jiangxin Dong for providing the code for generating the blur kernels with the type of Lai dataset. + +# References + +1. Asim, M., Shamshad, F., Ahmed, A.: Blind image deconvolution using deep generative priors. IEEE Transactions on Computational Imaging (TCI) 6, 1493-1506 (2020) +2. Bishop, T.E., Babacan, S.D., Amizic, B., Katsaggelos, A.K., Chan, T., Molina, R.: Blind image deconvolution: problem formulation and existing approaches. In: Blind Image Deconvolution, pp. 21-62. CRC press (2017) +3. Bredell, G., Flouris, K., Chaitanya, K., Erdil, E., Konukoglu, E.: Explicitly minimizing the blur error of variational autoencoders. In: Proceedings of International Conference on Learning Representations (ICLR) (2023) +4. Cai, J.F., Ji, H., Liu, C., Shen, Z.: Blind motion deblurring from a single image using sparse approximation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 104-111 (2009) +5. Cai, J., Zuo, W., Zhang, L.: Dark and bright channel prior embedded network for dynamic scene deblurring. IEEE Transactions on Image Processing (TIP) 29, 6885-6897 (2020) +6. Chakrabarti, A.: A neural approach to blind motion deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 221-235 (2016) +7. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14245-14254 (2021) +8. Chan, T.F., Wong, C.K.: Total variation blind deconvolution. IEEE Transactions on Image Processing (TIP) 7(3), 370-375 (1998) +9. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1742-1750 (2019) +0. Cho, S., Lee, S.: Fast motion deblurring. ACM Transactions on Graphics (TOG) 28(5), 1-8 (2009) +1. Dinh, L., Krueger, D., Bengio, Y.: NICE: Non-linear Independent Components Estimation (Apr 2015) +2. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using Real NVP. In: Proceedings of International Conference on Learning Representations (ICLR) (2017) +3. Dong, J., Pan, J., Su, Z., Yang, M.H.: Blind image deblurring with outlier handling. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2478-2486 (2017) +4. Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T.: Removing camera shake from a single photograph. ACM Transactions on Graphics (TOG) 25(3), 787-794 (2006) +5. Gandelsman, Y., Shocher, A., Irani, M.: "Double-DIP": Unsupervised image decomposition via coupled deep-image-priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11018-11027 (2019) + +16. Gong, D., Yang, J., Liu, L., Zhang, Y., Reid, I., Shen, C., Van Den Hengel, A., Shi, Q.: From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2319-2328 (2017) +17. Gong, K., Catana, C., Qi, J., Li, Q.: Pet image reconstruction using deep image prior. IEEE Transactions on Medical Imaging (TMI) 38(7), 1655-1665 (2019) +18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in Neural Information Processing Systems (NeurIPS) (2014) +19. Guan, S., Tai, Y., Ni, B., Zhu, F., Huang, F., Yang, X.: Collaborative learning for faster stylegan embedding. arXiv preprint arXiv:2007.01758 (2020) +20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) +21. Hradis, M., Kotera, J., Zemcik, P., Sroubek, F.: Convolutional neural networks for direct text deblurring. In: Proceedings of British Machine Vision Conference (BMVC) (2015) +22. Huo, D., Masoumzadeh, A., Kushol, R., Yang, Y.H.: Blind image deconvolution using variational deep image prior. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(10), 11472-11483 (2023) +23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4401-4410 (2019) +24. Kaufman, A., Fattal, R.: Deblurring using analysis-synthesis networks pair. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5811-5820 (2020) +25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Proceedings of the International Conference for Learning Representations (ICLR) (2015) +26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: Proceedings of International Conference on Learning Representations (ICLR) (2014) +27. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. In: Advances in Neural Information Processing Systems (NeurIPS) (2018) +28. Krishnan, D., Fergus, R.: Fast image deconvolution using hyper-laplacian priors. In: Advances in Neural Information Processing Systems (NeurIPS) (2009) +29. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 233-240 (2011) +30. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8183-8192 (2018) +31. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8878-8887 (2019) +32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV) 128(7), 1956-1981 (2020) +33. Lai, W.S., Huang, J.B., Hu, Z., Ahuja, N., Yang, M.H.: A comparative study for single image blind deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1701-1709 (2016) + +34. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1964-1971 (2009) +35. Li, J., Wang, W., Nan, Y., Ji, H.: Self-supervised blind motion deblurring with deep expectation maximization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13986-13996 (2023) +36. Liang, J., Zhang, K., Gu, S., Gool, L.V., Timofte, R.: Flow-based kernel prior with application to blind super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10596-10605 (2021) +37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 740-755 (2014) +38. Liu, G., Chang, S., Ma, Y.: Blind image deblurring using spectral properties of convolution operators. IEEE Transactions on Image Processing (TIP) 23(12), 5047-5056 (2014) +39. Luo, Y.S., Zhao, X.L., Jiang, T.X., Zheng, Y.B., Chang, Y.: Hyperspectral mixed noise removal via spatial-spectral constrained unsupervised deep image prior. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS) 14, 9435-9449 (2021) +40. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2437-2445 (2020) +41. Michaeli, T., Irani, M.: Blind deblurring using internal patch recurrence. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 783-798 (2014) +42. Molina, R., Katsaggelos, A.K., Abad, J., Mateos, J.: A bayesian approach to blind deconvolution based on dirichlet distributions. In: Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP). vol. 4, pp. 2809-2812 (1997) +43. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3883-3891 (2017) +44. Pan, J., Dong, J., Liu, Y., Zhang, J., Ren, J., Tang, J., Tai, Y.W., Yang, M.H.: Physics-based generative adversarial models for image restoration and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 43(7), 2449-2462 (2021) +45. Pan, J., Hu, Z., Su, Z., Yang, M.H.: Deblurring text images via 10-regularized intensity and gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2901-2908 (2014) +46. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1628-1636 (2016) +47. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Deblurring images via dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(10), 2315-2328 (2017) +48. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 7474-7489 (2021) + +49. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) +50. Perrone, D., Favaro, P.: Total variation blind deconvolution: The devil is in the details. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2909-2916 (2014) +51. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. In: Proceedings of International Conference on Learning Representations (ICLR) (2016) +52. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3341-3350 (2020) +53. Ren, W., Cao, X., Pan, J., Guo, X., Zuo, W., Yang, M.H.: Image deblurring via enhanced low-rank prior. IEEE Transactions on Image Processing (TIP) 25(7), 3426-3437 (2016) +54. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 184-201 (2020) +55. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. ACM Transactions on Graphics (TOG) 27(3), 1-10 (2008) +56. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 769-777 (2015) +57. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: Proceedings of the IEEE International Conference on Computational Photography (ICCP). pp. 1-8 (2013) +58. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8174-8182 (2018) +59. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9446-9454 (2018) +60. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9168-9178 (2021) +61. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(3), 3121-3138 (2022) +62. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1107-1114 (2013) +63. Yan, Y., Ren, W., Guo, Y., Wang, R., Cao, X.: Image deblurring via extreme channels prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4003-4011 (2017) +64. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 672-681 (2021) +65. Yue, Z., Yong, H., Zhao, Q., Zhang, L., Meng, D., Wong, K.Y.K.: Deep variational network toward blind image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2024) + +66. Yue, Z., Zhao, Q., Xie, J., Zhang, L., Meng, D., Wong, K.Y.K.: Blind image superresolution with elaborate degradation modeling on noise and kernel. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2118-2128 (2022) +67. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5728-5739 (2022) +68. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14821-14831 (2021) +69. Zhao, Z., Xiong, B., Wang, L., Ou, Q., Yu, L., Kuang, F.: Retinexdip: A unified deep framework for low-light image enhancement. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) 32(3), 1076-1088 (2022) +70. Zhou, X., Mateos, J., Zhou, F., Molina, R., Katsaggelos, A.K.: Variational dirichlet blur kernel estimation. IEEE Transactions on Image Processing (TIP) 24(12), 5127-5139 (2015) +71. Zuo, W., Ren, D., Zhang, D., Gu, S., Zhang, L.: Learning iteration-wise generalized shrinkage-thresholding operators for blind deconvolution. IEEE Transactions on Image Processing (TIP) 25(4), 1751-1764 (2016) \ No newline at end of file diff --git a/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/images.zip b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f3ea7b906a479689bcb966f1057ba721ce3a5526 --- /dev/null +++ b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:673b221bc89acea63812525d04f9b7f8a0a02761b7707fa44848a4babbde63f3 +size 501532 diff --git a/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/layout.json b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3c0d9308dba8ea877144e10e490bdf6888cb1bb6 --- /dev/null +++ b/2024/Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding/layout.json @@ -0,0 +1,12974 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "spans": [ + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "text", + "content": "Jiangtao Zhang" + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "text", + "content": ", Zongsheng Yue" + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "text", + "content": ", Hui Wang" + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "text", + "content": ", Qian Zhao" + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "text", + "content": ", and Deyu Meng" + }, + { + "bbox": [ + 198, + 167, + 414, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 215, + 201, + 397, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 201, + 397, + 213 + ], + "spans": [ + { + "bbox": [ + 215, + 201, + 397, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 215, + 201, + 397, + 213 + ], + "type": "text", + "content": " Xi'an Jiaotong University, Shaanxi, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 213, + 417, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 213, + 417, + 224 + ], + "spans": [ + { + "bbox": [ + 195, + 213, + 417, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 195, + 213, + 417, + 224 + ], + "type": "text", + "content": " S-Lab, Nanyang Technological University, Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 197, + 224, + 414, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 224, + 414, + 246 + ], + "spans": [ + { + "bbox": [ + 197, + 224, + 414, + 246 + ], + "type": "text", + "content": "3 Pazhou Laboratory (Huangpu), Guangdong, China \nzhangjt2021@stu.xjtu.edu.cn, zsyzam@gmail.com," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 247, + 460, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 247, + 460, + 257 + ], + "spans": [ + { + "bbox": [ + 153, + 247, + 460, + 257 + ], + "type": "text", + "content": "huiwang2019@stu.xjtu.edu.cn, {timmy.zhaoqian, dymeng}@xjtu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 278, + 455, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 278, + 455, + 477 + ], + "spans": [ + { + "bbox": [ + 159, + 278, + 455, + 477 + ], + "type": "text", + "content": "Abstract. Blind image deconvolution (BID) is a classic yet challenging problem in the field of image processing. Recent advances in deep image prior (DIP) have motivated a series of DIP-based approaches, demonstrating remarkable success in BID. However, due to the high nonconvexity of the inherent optimization process, these methods are notorious for their sensitivity to the initialized kernel. To alleviate this issue and further improve their performance, we propose a new framework for BID that better considers the prior modeling and the initialization for blur kernels, leveraging a deep generative model. The proposed approach pre-trains a generative adversarial network-based kernel generator that aptly characterizes the kernel priors and a kernel initializer that facilitates a well-informed initialization for the blur kernel through latent space encoding. With the pre-trained kernel generator and initializer, one can obtain a high-quality initialization of the blur kernel, and enable optimization within a compact latent kernel manifold. Such a framework results in an evident performance improvement over existing DIP-based BID methods. Extensive experiments on different datasets demonstrate the effectiveness of the proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 486, + 412, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 486, + 412, + 498 + ], + "spans": [ + { + "bbox": [ + 160, + 486, + 412, + 498 + ], + "type": "text", + "content": "Keywords: Generative kernel prior " + }, + { + "bbox": [ + 160, + 486, + 412, + 498 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 486, + 412, + 498 + ], + "type": "text", + "content": " Kernel initializer " + }, + { + "bbox": [ + 160, + 486, + 412, + 498 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 486, + 412, + 498 + ], + "type": "text", + "content": " DIP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 536, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 597 + ], + "type": "text", + "content": "Blind image deconvolution (BID), also known as deblurring, is a classical problem in image processing [8-10, 14, 29, 34], which aims at recovering the latent clean image from the observed blurry counterpart. This blurry counterpart can be approximately modeled in mathematics as (assuming the blur process is uniform and spatially invariant)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 272, + 597, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 597, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 272, + 597, + 481, + 609 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} = \\boldsymbol {k} \\otimes \\boldsymbol {x} + \\boldsymbol {n}, \\tag {1}", + "image_path": "58a6754532eac35e287e1bbb84d89156e5bed39ddb2e6e63ee351269833b0fa6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "content": " is the blurry image, " + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "content": " the underlying clean image, " + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{k}" + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "content": " the blur kernel, " + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "content": " usually the additive white Gaussian noise (AWGN), and " + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "content": " denotes the 2D convolution operator. Then the goal of BID is to estimate " + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 613, + 482, + 650 + ], + "type": "text", + "content": " under this degradation" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 654, + 436, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 436, + 666 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 436, + 666 + ], + "type": "text", + "content": "Code is available at https://github.com/jtaoz/GKPILE-Deconvolution." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "model without knowing the ground truth " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{k}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ". As is well-known, this problem is highly ill-posed, since different pairs of " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{k}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " may result in the same " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 481, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 481, + 287 + ], + "type": "text", + "content": "The BID task was conventionally formulated as an optimization problem, corresponding to a Bayesian posterior estimation problem from the probabilistic perspective. Thus, the maximum a postriori (MAP) [28, 29, 41, 45, 46, 55, 62] and variational inference (VI) [2, 14, 70] approaches were developed for point-wise and distribution-wise estimation, respectively. The core challenge of these approaches is to properly design the priors for both the clean image [28, 41, 45, 46, 62] and blur kernel [42, 70]. However, these handcrafted priors are relatively subjective and may not precisely characterize the intrinsic distributions of natural images and blur kernels in the real world, which limits the performance of the deblurring algorithm. Besides, the resulting optimization is typically endowed with highly non-convex characteristics, leading to potential convergence issues such as reaching a trivial solution of delta kernel or unsatisfactory local minimum [34]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 409 + ], + "type": "text", + "content": "Owing to their wide and successful applications, deep learning technologies have been introduced to solve the BID problem. Early explorations [6, 16, 56] embedded deep neural networks (DNNs) into the traditional methods to facilitate the estimation of blur kernels. Recent works have pursued a more direct approach, wherein the mapping from a blurry image to its clean counterpart is directly learned [30, 31, 43, 58, 67, 68]. The latter strategies have achieved promising results on several benchmarks, mainly attributed to the well-designed network architectures. However, these methods tend to overfit the training datasets, and thus may not generalize well to unseen scenarios, particularly for images with large and complex motion blur kernels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 544 + ], + "type": "text", + "content": "Recently, deep image prior (DIP) [59] has exhibited great potential in addressing the task of BID [22, 52], particularly under the scenario of motion blur. The core idea is to parameterize both the blur kernel and clean image as DNNs with random inputs, and then optimize these DNNs through an energy minimization problem conducted by Eq. (1). Compared with the traditional BID methods, these approaches employed more effective DNN-oriented priors and thus achieved superior performance. However, these methods did not fully explore the statistical characteristics of blur kernels, resulting in inaccurate kernel estimations, particularly in cases involving large kernel sizes. Additionally, since the optimization is highly non-convex, the random initialization of the blur kernel often leads to algorithmic instability and compromises the overall performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "To alleviate the aforementioned issues of DIP-based methods, we proposed a new approach to better characterize and initialize the blur kernel with deep generative prior (DGP) [7, 48, 60, 64]. Specifically, focusing on the BID of motion blurring, we train a kernel generator using generative adversarial network (GAN) [18] as kernel prior. Owing to GAN's powerful fitting capability on complex distributions, the learned generator is expected to well characterize the blur kernels in a low-dimensional latent space. Then a mapping from the blurry image to the latent code of the kernel generator is trained, acting as a kernel initializer for the subsequent BID task. Intuitively, such a kernel initializer is easier to learn as the latent kernel space is more compact than that of the original kernel. To" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 114, + 470, + 235 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 470, + 235 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 470, + 235 + ], + "type": "image", + "image_path": "acf3ab469f5203478b2c95f7ecb49f1b9836ee055b8daadce34c439e8dbd809f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": "Fig. 1: Illustration of the latent encoding-based kernel initialization and fine-tuning mechanism in our BID framework. The blur kernel is first initialized from the blurry image by a pre-trained kernel initializer via the latent code " + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": ", which is expected to be close to the code of the ground truth kernel, " + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": ". The corresponding kernels " + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_0" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": " in the kernel manifold is also expected to be close. Then the optimization is performed within the latent space, such that " + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": " is fine-tuned to " + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{opt}}" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": ", and the final estimated kernel " + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{\\mathrm{opt}}" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": " is closer to the ground truth " + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": ". See Sec. 3 for more details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "content": "solve the problem of BID, we first obtain a coarse initialization for the blur kernel by the initializer from the blurry image, and then jointly fine-tune the DIP for the image and the DGP for the blur kernel. The mechanism of the latent encoding-based kernel initialization and fine-tuning is illustrated in Fig. 1. Since the kernel generator sufficiently fits the kernel distribution and the kernel initializer provides a better initial estimation, the overall BID performance can be significantly improved with faster convergence, especially for large blur kernels. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 424, + 480, + 532 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 424, + 480, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 447 + ], + "type": "text", + "content": "- We use GAN to learn the DGP for blur kernels, which can better depict the kernel structures and thus offer a more effective and compact kernel prior." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 449, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 495 + ], + "type": "text", + "content": "- We propose to learn a kernel initializer that maps from the blurry image to the latent code of the corresponding kernel. Attributed to the compactness of the latent kernel space, the proposed initializer can provide a more accurate kernel initialization for the subsequent BID process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 495, + 480, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 495, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 495, + 480, + 532 + ], + "type": "text", + "content": "- By harnessing the designed kernel prior and initializer, our BID method achieves state-of-the-art (SotA) performance on the challenging benchmark dataset provided by Lai et al. [33], especially with large blur kernels." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 549, + 234, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 234, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 234, + 562 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 574, + 294, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 294, + 586 + ], + "type": "text", + "content": "2.1 Blind image deconvolution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 492, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 492, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 492, + 616 + ], + "type": "text", + "content": "BID methods can be roughly divided into two main categories, namely optimization-based methods and deep learning-based methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Optimization-based BID methods. Due to its high ill-posedness, the solution spaces of both the clean image and blur kernel should be properly constrained to make the BID problem solvable. To this aim, traditional optimization-based methods mainly focus on designing appropriate priors for clean images" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 144, + 90, + 448, + 103 + ], + "type": "text", + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "and blur kernels, and plenty of studies have been conducted. For example, total variation [8], hyper-Laplacian prior [28], " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "-norm [29], and transform-specific " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "-norm [4, 62] were considered to model the image prior in the gradient domain, and patch-based priors [41, 57], low-rank prior [53] and dark/bright channel prior [46, 63] were directly applied in the image domain. As for blur kernels, the non-negative and normalization constraints result in a basic regularization. In addition, focusing on the properties of kernel, sparsity prior [47] and spectral prior [38] have been proposed. To get an ideal solution, tricks such as delayed normalization [50] and multi-scale implementation [71] for blur kernels are also suggested. However, these methods highly rely on handcrafted priors, which may not accurately characterize the intrinsic properties of the clean images and blur kernels, and thus are less competitive in this deep learning era." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 403 + ], + "type": "text", + "content": "Deep learning-based BID methods. Having witnessed great success in a wide range of applications, deep learning techniques have also been adopted for the BID task. Early methods attempted to replace certain components of the traditional optimization-based methods with DNNs, by virtue of their flexibility. For example, some researchers introduced DNNs as kernel predictors [6,16,56,65]. As the computing resources increase, it is then more popular to directly learn a mapping from the blurred image to its clean counterpart in a fully supervised way, with paired training data and well-designed DNNs [5,21,30,31,43,44,58,67, 68]. Though achieved SotA performance on several benchmark datasets, such as GoPro [43] and RealBlur [54], these methods may encounter the limitations in generalization to images with large complex blur kernels that are not simulated in the pre-defined training sets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 571 + ], + "type": "text", + "content": "Recently, focusing on motion blurring, a new type of deep learning methods for the BID problem that leverages the deep priors (will be reviewed in the next subsection), has attracted increasing attention. Along this line, Ren et al. [52] proposed the SelfDeblur method, which adopted DIP to parameterize both the clean image and blur kernel for the first time. After that, multiple variants and extensions, mainly paying attention to the image prior, have been proposed for further improvement of performance [22,35]. These methods achieved promising results in certain cases, especially that they outperform fully supervised deep learning methods on the challenging benchmark by Lai et al. [33]. However, the properties of the blur kernel were not explicitly explored, and the performance tends to be unstable for large kernels due to the non-convex optimization. In contrast to these studies that used DIP with untrained DNNs, Asim et al. [1] proposed to pre-train generators for both the clean images and blur kernels as DGPs, and then fine-tune the generators when doing BID." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 586, + 323, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 323, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 323, + 600 + ], + "type": "text", + "content": "2.2 Deep prior for image processing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "type": "text", + "content": "Due to their flexibility, DNNs have been used to characterize the image priors in recent years. There are mainly two kinds of such deep priors, namely DIP and DGP, which are briefly reviewed in the following." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Deep image prior. DIP was originally proposed by Ulyanov et al. [59] who trained a DNN to approximate the single target (maybe up to a transformation)" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "with random noise as its input. Since there are complex structures and operators involved, the DNN is expected to well depict the manifold of images, playing a similar role as the prior or regularizer in traditional image processing methods. After being raised, DIP has attracted remarkable attention and been applied to various image processing tasks (other than BID mentioned before), including natural image denoising [59], super-resolution [36, 66], inpainting [59], image decomposition [15], low-light enhancement [69], PET image reconstruction [17] and hyperspectral image denoising [39]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 342 + ], + "type": "text", + "content": "Deep generative prior. Attributed to their powerful generation capabilities, the trained deep generators, such as GANs [18], can be seen as approximators to the distributions of images, hoping to provide reasonable priors for images. Such priors can be referred to as DGPs [48]. After a pre-training stage for the generator, DGP can be used as an estimator to approximate the target image via fine-tuning like DIP. There are two common ways to fine-tune DGP. The first is to fix the generator and optimize the input. For example, Menon et al. [40] used this idea for photo upsampling. This strategy is also closely related to GAN-inversion [61]. The second is to jointly fine-tune the latent code and the parameters of the pre-trained generator like [48], which may better locate the optimal estimation within the image manifold." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 481, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 481, + 428 + ], + "type": "text", + "content": "In addition to directly modeling the images, DGP was also used to characterize the degradation operators, such as blur kernels. For example, Asim et al. [1] proposed to pre-train a variational auto-encoder (VAE) [26] as DGP for motion blur kernels for the BID task; and Liang et al. [36] used the normalizing flow (NF) [11, 12, 27] as the DGP for Gaussian blur kernels, applying to blind superresolution. These studies highlight the effectiveness of DGP for blur kernels, and partially inspire our work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 446, + 259, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 259, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 259, + 459 + ], + "type": "text", + "content": "3 Proposed method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 472, + 274, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 274, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 274, + 483 + ], + "type": "text", + "content": "3.1 DIP-based BID revisit" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 479, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 479, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 479, + 516 + ], + "type": "text", + "content": "Before presenting our BID method, we first revisit DIP-based BID approaches, or more specifically SelfDeblur [52], and discuss the issues we try to address." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 517, + 480, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 480, + 563 + ], + "type": "text", + "content": "In SelfDeblur [52], the clean image " + }, + { + "bbox": [ + 130, + 517, + 480, + 563 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 517, + 480, + 563 + ], + "type": "text", + "content": " in Eq. (1) was approximated by a convolutional network equipped with encoder-decoder architecture, and the blur kernel " + }, + { + "bbox": [ + 130, + 517, + 480, + 563 + ], + "type": "inline_equation", + "content": "\\pmb{k}" + }, + { + "bbox": [ + 130, + 517, + 480, + 563 + ], + "type": "text", + "content": " was estimated by a fully-connected network with one hidden layer. The resulting optimization can then be formulated as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 227, + 571, + 480, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 571, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 227, + 571, + 480, + 590 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {k}, \\theta_ {x}} \\| G _ {k} \\left(\\boldsymbol {z} _ {k}; \\theta_ {k}\\right) \\otimes G _ {x} \\left(\\boldsymbol {z} _ {x}; \\theta_ {x}\\right) - \\boldsymbol {y} \\| ^ {2}, \\tag {2}", + "image_path": "4bdef12665fd99c643c8564b99a0895d71adea5282d6059d19436a7c9ad44212.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k})" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{x}(\\cdot ;\\theta_{x})" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " are networks for the kernel and image, respectively, and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{z}_k" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{x}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " are input random noise. Note that we ignore the additional constraints imposed on the blur kernel and clean image, positing that these constraints can be readily satisfied through thoughtful network design. Thus, the problem of estimating the blur kernel " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{k}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and the clean image " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is reduced to the optimization of parameters " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta_k^*" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta_x^*" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 144, + 90, + 448, + 103 + ], + "type": "text", + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 110, + 203, + 165 + ], + "blocks": [ + { + "bbox": [ + 149, + 110, + 203, + 165 + ], + "lines": [ + { + "bbox": [ + 149, + 110, + 203, + 165 + ], + "spans": [ + { + "bbox": [ + 149, + 110, + 203, + 165 + ], + "type": "image", + "image_path": "f123bdded2e99e8ca159a3a591f2effcf8bb2866a3eccc2b2b413042916fb096.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 167, + 203, + 220 + ], + "blocks": [ + { + "bbox": [ + 149, + 167, + 203, + 220 + ], + "lines": [ + { + "bbox": [ + 149, + 167, + 203, + 220 + ], + "spans": [ + { + "bbox": [ + 149, + 167, + 203, + 220 + ], + "type": "image", + "image_path": "5f70aa539da4dd02b6f92c50e2c544a8f15c8dc8175d5b59751fd2223375ffc1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 220, + 194, + 227 + ], + "lines": [ + { + "bbox": [ + 159, + 220, + 194, + 227 + ], + "spans": [ + { + "bbox": [ + 159, + 220, + 194, + 227 + ], + "type": "text", + "content": "Initial Kernel" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 111, + 259, + 165 + ], + "blocks": [ + { + "bbox": [ + 205, + 111, + 259, + 165 + ], + "lines": [ + { + "bbox": [ + 205, + 111, + 259, + 165 + ], + "spans": [ + { + "bbox": [ + 205, + 111, + 259, + 165 + ], + "type": "image", + "image_path": "35b152e9ad2545a01a42c3c200b39ea487a51f178b117ac28534965d3cfb28e1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 205, + 167, + 259, + 220 + ], + "blocks": [ + { + "bbox": [ + 205, + 167, + 259, + 220 + ], + "lines": [ + { + "bbox": [ + 205, + 167, + 259, + 220 + ], + "spans": [ + { + "bbox": [ + 205, + 167, + 259, + 220 + ], + "type": "image", + "image_path": "9d59f219d3afaeca0b9e85689c1ebeb398f3d56f11623fafff8073f6d75d93fc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 220, + 255, + 227 + ], + "lines": [ + { + "bbox": [ + 209, + 220, + 255, + 227 + ], + "spans": [ + { + "bbox": [ + 209, + 220, + 255, + 227 + ], + "type": "text", + "content": "Estimated Kernel" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 261, + 111, + 344, + 165 + ], + "blocks": [ + { + "bbox": [ + 261, + 111, + 344, + 165 + ], + "lines": [ + { + "bbox": [ + 261, + 111, + 344, + 165 + ], + "spans": [ + { + "bbox": [ + 261, + 111, + 344, + 165 + ], + "type": "image", + "image_path": "2cd98dbca9e15011cc3b5a7b360d1845ee75a0308247d2a444bdd052f8e232df.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 111, + 468, + 221 + ], + "lines": [ + { + "bbox": [ + 359, + 111, + 468, + 221 + ], + "spans": [ + { + "bbox": [ + 359, + 111, + 468, + 221 + ], + "type": "text", + "content": "Fig.2: Illustration of the initialization effect of the DIP-based BID. The two rows correspond to two independent runs of SelfDeblur [52]. From left to right: the randomly initialized kernel, the finally estimated kernel, and the deblurred image." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 261, + 167, + 344, + 220 + ], + "blocks": [ + { + "bbox": [ + 261, + 167, + 344, + 220 + ], + "lines": [ + { + "bbox": [ + 261, + 167, + 344, + 220 + ], + "spans": [ + { + "bbox": [ + 261, + 167, + 344, + 220 + ], + "type": "image", + "image_path": "ec8aa51fcca5c34bec8a3b25c3e8e99414b6f2f7932c7a408c0bd2a66b22ec82.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 220, + 326, + 228 + ], + "lines": [ + { + "bbox": [ + 279, + 220, + 326, + 228 + ], + "spans": [ + { + "bbox": [ + 279, + 220, + 326, + 228 + ], + "type": "text", + "content": "Deblurred Image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 233, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 376 + ], + "type": "text", + "content": "Owing to the powerful capabilities of DNNs, SelfDeblur obtained promising results on the BID task with motion blurring. Notably, it established a new SotA performance on the challenging benchmark introduced by Lai et al. [33]. Huo et al. [22] further improved SelfDeblur by incorporating a more sophisticated DIP within the variational Bayesian framework. However, these DIP-based BID methods did not fully explore the statistical structures or priors for blur kernels, thereby limiting their performance. A more important concern arises from the non-convexity of the optimization, where the randomness of " + }, + { + "bbox": [ + 130, + 233, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_k" + }, + { + "bbox": [ + 130, + 233, + 482, + 376 + ], + "type": "text", + "content": " causes unstable solutions, particularly when the blur kernel is large. Fig. 2 shows a typical example with two different random runs. Despite the initializations following the same distribution, the estimated blur kernels can be diverse, consequently leading to significantly different deblurring results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 376, + 481, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 481, + 413 + ], + "type": "text", + "content": "The aforementioned issues redirect our focus toward the blur kernels. Therefore, this study explores more precise statistical structures for blur kernels with DGP, followed by the development of a kernel initializer in the latent space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 430, + 332, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 332, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 332, + 442 + ], + "type": "text", + "content": "3.2 Overview of the proposed method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 450, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 498 + ], + "type": "text", + "content": "Our proposed method consists of two stages. The first is the pre-training stage, aiming to learn a kernel generator and a kernel initializer. The second stage is to solve the deconvolution problem following the paradigm of the DIP-based BID method, while with the aid of the learned generator and initializer." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "content": "In the pre-training stage, we train a kernel generator " + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "content": " within the GAN [18] framework, where " + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\theta_{k}^{*}" + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "content": " denotes the generator parameters after training. Once learned, it can generate a blur kernel from the randomly sampled noise " + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "content": ". This means the generator indeed models the distribution of the blur kernels and thus characterizes their statistical structures. Such a learned kernel generator can be referred to as the DGP of blur kernels, with " + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "content": " serving as the kernel latent code. Next, with the generator " + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "content": " fixed, we train an encoder " + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "inline_equation", + "content": "E(\\cdot ;\\theta_{E}^{*})" + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "content": " that maps a blurry image " + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "content": " to the corresponding kernel latent code through the technique of GAN-inversion [61]. The learned encoder allows for a relatively accurate prediction for the blur kernel based on its latent code, and thus acts as an effective initializer for the kernel DGP." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "In the second stage, we first predict the kernel latent code corresponding to the input blurry image " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": " using the pre-trained kernel initializer, which provides an initial value for the kernel DGP in the latent space. Then we jointly optimize" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 115, + 467, + 260 + ], + "blocks": [ + { + "bbox": [ + 143, + 115, + 467, + 260 + ], + "lines": [ + { + "bbox": [ + 143, + 115, + 467, + 260 + ], + "spans": [ + { + "bbox": [ + 143, + 115, + 467, + 260 + ], + "type": "image", + "image_path": "4bb981d60bc2247efa3ff775c17c3195f2a3d374c36329a0e03ca30617ac8a7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 264, + 480, + 287 + ], + "lines": [ + { + "bbox": [ + 131, + 264, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 131, + 264, + 480, + 287 + ], + "type": "text", + "content": "Fig. 3: Overview of the proposed BID process, with the pre-trained kernel generator and initializer, of our proposed method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 327 + ], + "type": "text", + "content": "the parameters of DIP for the clean image and the latent code (or more precisely its mapped features as discussed in Sec. 3.5) of the DGP for the blur kernel. The overview of the proposed BID process is shown in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 481, + 365 + ], + "type": "text", + "content": "In Sec. 3.3 and Sec. 3.4, we provide details for learning the kernel generator and kernel initializer, respectively. Then in Sec. 3.5, we discuss how to apply the pre-trained generator and initializer to the BID process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 381, + 291, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 291, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 291, + 395 + ], + "type": "text", + "content": "3.3 Kernel generator learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 401, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 568 + ], + "type": "text", + "content": "To learn the kernel generator, a large amount of motion blur kernels are synthesized to simulate their distribution based on the physical generation mechanism proposed in [30] or [33]. Then we train the kernel generator " + }, + { + "bbox": [ + 130, + 401, + 482, + 568 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot;\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 401, + 482, + 568 + ], + "type": "text", + "content": " with these synthesized blur kernels. Note that our kernel generator is made up of several convolutional layers inspired by DCGAN [51]. Though can be seen as a straightforward application of GAN, this step is crucial for our method. On the one hand, the learned generator depicts the statistical structures of blur kernels, serving as a kernel DGP in the BID process; on the other hand, the learned kernel generator also forms the milestone for learning the kernel initializer in a low-dimensional latent space, which will be discussed in Sec. 3.4. Fig. 4 shows several blur kernels generated by " + }, + { + "bbox": [ + 130, + 401, + 482, + 568 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot;\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 401, + 482, + 568 + ], + "type": "text", + "content": " and the physical model according to [33]. As can be observed, the blur kernels generated by our learned generator are morphologically very similar to that synthesized by the physical model, indicating its capability to characterize the underlying manifold of the blur kernels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Remark. It should be mentioned there are some studies also considering learning DGP for blur kernels while with various generative models. For example, Asim et al. [1] used the VAE for the BID task, and Liang et al. [36] adopted the NF for the blind image super-resolution. In this work, we prefer GAN due to the following reasons. First, compared with the vanilla VAE that tends to generate blurry samples [3], GAN can provide sharper ones. Second, in contrast to NF, GAN imposes no constraints on the invertibility between the latent code and the generated sample, thereby facilitating easier learning of the kernel manifold in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 144, + 90, + 448, + 103 + ], + "type": "text", + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 329, + 213 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 329, + 213 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 329, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 329, + 213 + ], + "type": "image", + "image_path": "6387327c8fceff6e67c5b2b73b81f241725a2cbc44791b4203cd62aa932f8eac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 114, + 461, + 213 + ], + "lines": [ + { + "bbox": [ + 335, + 114, + 461, + 213 + ], + "spans": [ + { + "bbox": [ + 335, + 114, + 461, + 213 + ], + "type": "text", + "content": "Fig.4: Comparison of blur kernels synthesized according to physical model and generated by our pre-trained generator. Top row: blur kernels synthesized according to [33]. Bottom row: blur kernels generated by the pre-trained kernel generator." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 218, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 255 + ], + "type": "text", + "content": "a lower-dimensional latent space. This aspect is particularly advantageous when considering motion blur kernels, whose sizes can be relatively large (up to " + }, + { + "bbox": [ + 130, + 218, + 482, + 255 + ], + "type": "inline_equation", + "content": "75 \\times 75" + }, + { + "bbox": [ + 130, + 218, + 482, + 255 + ], + "type": "text", + "content": " in the Lai dataset [33])." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 271, + 291, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 291, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 291, + 283 + ], + "type": "text", + "content": "3.4 Kernel initializer learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 479, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 479, + 337 + ], + "type": "text", + "content": "As discussed before, kernel initialization is important for the overall performance in the BID task since the optimization is highly non-convex. Therefore, we propose to initialize the blur kernel in the latent space using an encoder " + }, + { + "bbox": [ + 130, + 290, + 479, + 337 + ], + "type": "inline_equation", + "content": "E(\\cdot; \\theta_E^*)" + }, + { + "bbox": [ + 130, + 290, + 479, + 337 + ], + "type": "text", + "content": " that maps the blurry image " + }, + { + "bbox": [ + 130, + 290, + 479, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 290, + 479, + 337 + ], + "type": "text", + "content": " to the latent code " + }, + { + "bbox": [ + 130, + 290, + 479, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 290, + 479, + 337 + ], + "type": "text", + "content": " of the learned kernel DGP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 338, + 463, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 338, + 463, + 350 + ], + "spans": [ + { + "bbox": [ + 146, + 338, + 463, + 350 + ], + "type": "text", + "content": "A direct way to train the encoder is to solve the following optimization:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 210, + 352, + 480, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 352, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 210, + 352, + 480, + 372 + ], + "type": "interline_equation", + "content": "\\theta_ {E} ^ {*} = \\arg \\min _ {\\theta_ {E}} \\sum_ {n} \\ell \\left(G _ {k} \\left(E \\left(\\boldsymbol {y} _ {n}; \\theta_ {E}\\right); \\theta_ {k} ^ {*}\\right), \\boldsymbol {k} _ {n}\\right), \\tag {3}", + "image_path": "e24d8512960a6e90b33e2c46b3c946f74cbe385e957202337e83bea877a95c73.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "content": "-th blurry image, " + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{k}_n" + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "content": " the corresponding blur kernel, " + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\theta_k^*" + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "content": " is the parameters of the pre-trained kernel generator, and " + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "content": " is some loss function. This optimization can be seen as training a predictor that directly estimates the blur kernel from the blurry image, which is not easy due to the relatively complex structures of blur kernels. On the other hand, if we can find the latent code " + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{z}_n" + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{k}_n" + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "content": ", and learn to predict " + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{z}_n" + }, + { + "bbox": [ + 130, + 372, + 482, + 491 + ], + "type": "text", + "content": ", the task could be simpler, since the latent space is expected to be much more compact than the original kernel space. Therefore, we propose to use a collaborative learning strategy motivated by [19], to guide the encoder training within the latent space. Specifically, we first formulate the following optimization to train the encoder:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 165, + 493, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 493, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 165, + 493, + 480, + 521 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {E}} \\sum_ {n} \\left\\{\\| G _ {k} \\left(E \\left(\\boldsymbol {y} _ {n}; \\theta_ {E}\\right); \\theta_ {k} ^ {*}\\right) - G _ {k} \\left(\\boldsymbol {z} _ {n}; \\theta_ {k} ^ {*}\\right) \\| _ {1} + \\lambda \\| E \\left(\\boldsymbol {y} _ {n}; \\theta_ {E}\\right) - \\boldsymbol {z} _ {n} \\| _ {2} ^ {2} \\right\\}, \\tag {4}", + "image_path": "49d57aecd9d2ee25bf5da83c6f0634a5793b9bd145ad0b3e1a095e98200cacba.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 166, + 522, + 331, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 522, + 331, + 538 + ], + "spans": [ + { + "bbox": [ + 166, + 522, + 331, + 538 + ], + "type": "interline_equation", + "content": "s. t. \\boldsymbol {z} _ {n} = \\arg \\min _ {\\boldsymbol {z}} \\| G _ {k} (\\boldsymbol {z}; \\theta_ {k} ^ {*}) - \\boldsymbol {k} _ {n} \\| _ {1},", + "image_path": "3c4c1b3d951da0ee5ea7778e8b79f2591e6ce59269c0b6b0ccab403366a0ecb2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 546, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 480, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 546, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 546, + 480, + 581 + ], + "type": "text", + "content": " is a tuning parameter set to 0.1 in this work. For convenience, we denote the objective functions of the outer and inner optimizations as " + }, + { + "bbox": [ + 130, + 546, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_E(\\theta_E)" + }, + { + "bbox": [ + 130, + 546, + 480, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_z^n (z)" + }, + { + "bbox": [ + 130, + 546, + 480, + 581 + ], + "type": "text", + "content": ", respectively, in the following." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "The optimization of Eq. (4) involves a sub-task of solving " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " that corresponds to a GAN-inversion problem [61]. Following [19], we alternatively optimize " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\theta_E" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " in each iteration. Specifically, " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " is firstly initialized as " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "E(\\mathbf{y}_n; \\theta_E)" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", and then updated by several gradient descent steps with respect to the lower-level objective " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_z^n(\\mathbf{z})" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ". Based on the updated " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\theta_E" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " is subsequently optimized according to the upper-level objective " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_E(\\theta_E)" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", also by gradient descent steps. The whole procedure is summarized in Algorithm 1." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 463, + 247 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 298, + 127 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 298, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 298, + 127 + ], + "type": "text", + "content": "Algorithm 1 Kernel initializer learning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": "Input: Pre-trained kernel generator " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k}^{*})" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " , blurry image-kernel pairs " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\{\\pmb {y}_n,\\pmb {k}_n\\}" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " , step size " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " 1: Initialize " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\theta_E^{(0)}" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,T" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " do \n3: For each " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " .. \n4: " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\pmb{z}^{(0)} = E(\\pmb {y}_n;\\theta_E^{(t - 1)})" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " \n5: for " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "s = 1,\\dots ,S" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\pmb{z}^{(s)} = \\pmb{z}^{(s - 1)} - \\epsilon \\nabla_{\\pmb{z}}\\mathcal{L}_{\\pmb{z}}^{n}(\\pmb {z})|_{\\pmb{z} = \\pmb{z}^{(s - 1)}}" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " end for \n6: " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{n}^{(t)} = \\pmb{z}^{(S)}" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\theta^{(0)} = \\theta_{E}^{(t - 1)}" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " \n8: for " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "l = 1,\\dots ,L" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\theta^{(l)} = \\theta^{(l - 1)} - \\epsilon \\nabla_{\\theta}\\mathcal{L}_{E}(\\theta)|_{\\theta = \\theta^{(l - 1)}}" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " end for \n9: " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\theta_E^{(t)} = \\theta^{(L)}" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " \n10: end for \nOutput: Kernel initializer " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "E(\\cdot ;\\theta_E^*)" + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 132, + 130, + 463, + 247 + ], + "type": "inline_equation", + "content": "\\theta_E^* = \\theta_E^{(T)}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "type": "image", + "bbox": [ + 145, + 255, + 340, + 353 + ], + "blocks": [ + { + "bbox": [ + 145, + 255, + 340, + 353 + ], + "lines": [ + { + "bbox": [ + 145, + 255, + 340, + 353 + ], + "spans": [ + { + "bbox": [ + 145, + 255, + 340, + 353 + ], + "type": "image", + "image_path": "5b55c6a713502069a8cda2752d5574d5ada2bcb1a3531085fabaffd27208ee70.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 255, + 461, + 353 + ], + "lines": [ + { + "bbox": [ + 352, + 255, + 461, + 353 + ], + "spans": [ + { + "bbox": [ + 352, + 255, + 461, + 353 + ], + "type": "text", + "content": "Fig.5: Illustration of the estimating ability of the pre-trained kernel initializer. Left: the blurry image. Top-right: the estimated blur kernel by the kernel initializer. Bottomright: the ground-truth blur kernel." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": "Once trained, encoder " + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "inline_equation", + "content": "E(\\cdot; \\theta_E^*)" + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": " can be regarded as an effective kernel initializer, providing a promising prediction of the blur kernel from the blurry image. Fig. 5 shows an example of the predicted kernel by the learned kernel initializer, in comparison with the ground truth. It can be observed that, though not perfect, the prediction is indeed very close to the ground truth, and thus is expected to be a good starting point for further processing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 449, + 224, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 224, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 224, + 459 + ], + "type": "text", + "content": "3.5 BID Process" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "type": "text", + "content": "With the pre-trained kernel generator " + }, + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "type": "text", + "content": " and initializer " + }, + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "type": "inline_equation", + "content": "E(\\cdot ;\\theta_{E}^{*})" + }, + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "type": "text", + "content": ", we can formulate the BID problem as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 198, + 498, + 481, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 498, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 198, + 498, + 481, + 517 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {z} _ {k} ^ {*}, \\theta_ {x} ^ {*}\\right) = \\underset {\\boldsymbol {z} _ {k}, \\theta_ {x}} {\\arg \\min } \\| G _ {k} \\left(\\boldsymbol {z} _ {k}; \\theta_ {k} ^ {*}\\right) \\otimes G _ {x} \\left(\\boldsymbol {z} _ {x}; \\theta_ {x}\\right) - \\boldsymbol {y} \\| ^ {2}, \\tag {5}", + "image_path": "60718c5a26e6079a8f98c075b38a971a8165cb88c55899c0b9a38ce685b226cc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{z}_x" + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": " is a random vector, " + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{z}_k" + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": " is initialized by " + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{z}_k^{(0)} = E(\\pmb{y};\\theta_E^*)" + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": ". After optimization, the desired clean image and blur kernel can be accessed via " + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}} = G_{x}(\\pmb{z}_{x};\\theta_{x}^{*})" + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{k}} = G_{k}(\\pmb{z}_{k}^{*};\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": ", respectively. In fact, such a naive implementation still cannot achieve satisfactory results as empirically illustrated in Sec. 4.3, in particular when the blur kernel is large. This can be attributed to the relatively low dimensionality of " + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{z}_k" + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": ", compared with the blur kernel, which enforces too strong constraints and makes the optimization more difficult." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "To address this issue, let's first recall one small yet effective trick in StyleGAN [23]. It introduced a network to map the latent code " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " to a style vector " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", enabling more precise control of the generated style. Inspired by this trick, we propose to optimize the feature map of the first layer in the kernel generator " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", whose dimension is relatively higher, instead of the latent code " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". By" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "type": "text", + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 440, + 127 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 440, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 440, + 127 + ], + "type": "text", + "content": "Algorithm 2 BID process with pre-trained kernel generator and initializer" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": "Input: Pre-trained kernel generator " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k}^{*}) = G_{k}^{(w)}(g_{1}(\\cdot))" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " pre-trained kernel initializer " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "E(\\cdot ;\\theta_E^*)" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " blurry image " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " , random vector " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\pmb {z}_x" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " , step size " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " \n1: Initialize " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{k}^{(0)} = g_{1}(E(\\pmb {y};\\theta_{E}^{*}))" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\theta_x^{(0)}" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " , denote " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\boldsymbol {w},\\theta) = \\left\\| G_k^{(w)}(\\boldsymbol {w})\\otimes G_x(\\boldsymbol {z}_x;\\theta) - \\boldsymbol {y}\\right\\| ^2" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,T" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " do \n3: " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\pmb{w}_{k}^{(t)} = \\pmb{w}_{k}^{(t - 1)} - \\epsilon \\nabla_{\\pmb{w}}\\mathcal{L}(\\pmb {w},\\theta)\\big|_{\\pmb {w} = \\pmb{w}_{k}^{(t - 1)},\\theta = \\theta_{x}^{(t - 1)}}\\\\ \\theta_{x}^{(t)} = \\theta_{x}^{(t - 1)} - \\epsilon \\nabla_{\\theta}\\mathcal{L}(\\pmb {w},\\theta)\\big|_{\\pmb {w} = \\pmb{w}_{k}^{(t - 1)},\\theta = \\theta_{x}^{(t - 1)}}\\\\ \\end{array}" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " \n4: end for \n6: " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{k}^{*} = \\pmb{w}_{k}^{(T)}" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\theta_x^* = \\theta_x^{(T)}" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " \nOutput: Estimated clean image " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}} = G_{x}(\\pmb {z}_{x};\\theta_{x}^{*})" + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "text", + "content": " , estimated blur kernel " + }, + { + "bbox": [ + 132, + 130, + 481, + 232 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{k}} = G_k^{(w)}(\\pmb {w}_k^*)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "text", + "content": "denoting the first layer of " + }, + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "inline_equation", + "content": "g_{1}(\\cdot)" + }, + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "text", + "content": " and the reduced generator without " + }, + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "inline_equation", + "content": "g_{1}(\\cdot)" + }, + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "inline_equation", + "content": "G_{k}^{(w)}(\\cdot)" + }, + { + "bbox": [ + 130, + 246, + 479, + 273 + ], + "type": "text", + "content": ", the BID problem of Eq. (5) is reformulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 274, + 481, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 274, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 198, + 274, + 481, + 297 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {w} _ {k} ^ {*}, \\theta_ {x} ^ {*}\\right) = \\underset {\\boldsymbol {w} _ {k}, \\theta_ {x}} {\\arg \\min } \\left\\| G _ {k} ^ {(w)} \\left(\\boldsymbol {w} _ {k}\\right) \\otimes G _ {x} \\left(\\boldsymbol {z} _ {x}; \\theta_ {x}\\right) - \\boldsymbol {y} \\right\\| ^ {2}, \\tag {6}", + "image_path": "7416a01bd87bd9e395654bcc18b77c2865970e5a46d8f0d2640a8f5914a10fc4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 303, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 479, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 303, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k" + }, + { + "bbox": [ + 130, + 303, + 479, + 327 + ], + "type": "text", + "content": " is initialized by " + }, + { + "bbox": [ + 130, + 303, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k^{(0)} = g_1(E(\\pmb{y};\\theta_E^*))" + }, + { + "bbox": [ + 130, + 303, + 479, + 327 + ], + "type": "text", + "content": ". The whole BID process for solving Eq. (6) is summarized in Algorithm 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 328, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 399 + ], + "type": "text", + "content": "It should be noted that the parameters of the generator network can also be fine-tuned in previous studies [48]. However, due to the highly non-convex nature of the BID optimization, too many to-be-optimized parameters may hinder its performance. We thus only update " + }, + { + "bbox": [ + 130, + 328, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k" + }, + { + "bbox": [ + 130, + 328, + 480, + 399 + ], + "type": "text", + "content": " for the blur kernel in this work. Empirical results in Sec. 4.3 substantiate the superiority of this strategy over updating the whole kernel generator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 419, + 230, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 230, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 230, + 433 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 445, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 480, + 493 + ], + "type": "text", + "content": "In this section, we present experimental results to demonstrate the effectiveness of the proposed method and ablation studies to analyze some key components in our framework. More results can be found in the supplementary material, together with discussions on the limitations of this work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 511, + 345, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 345, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 345, + 525 + ], + "type": "text", + "content": "4.1 Datasets and implementation details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 532, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 481, + 616 + ], + "type": "text", + "content": "Datasets. We evaluate our proposed framework mainly on two datasets: one is synthesized by ourselves, and the other is the well-known BID benchmark constructed by Lai et al. [33]. For the synthetic dataset, we randomly select 80 source images from MSCOCO [37] and apply the motion blur synthesized by [30] on them. The second Lai dataset consists of 25 clean images and 4 large-size blur kernels, resulting in 100 testing blur images in total. Beyond the two synthetic datasets, we also test our method real blurry images collected by Lai et al. [33]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 667 + ], + "type": "text", + "content": "Implementation details. As discussed in Sec. 3, our whole framework consists of a pre-training stage and a BID process. In the pre-training stage, we first train the kernel generator " + }, + { + "bbox": [ + 130, + 617, + 481, + 667 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 617, + 481, + 667 + ], + "type": "text", + "content": " with DCGAN [51], using kernels synthesized following the way in [30] or [33]. Then we freeze the weights of " + }, + { + "bbox": [ + 130, + 617, + 481, + 667 + ], + "type": "inline_equation", + "content": "G_{k}(\\cdot ;\\theta_{k}^{*})" + }, + { + "bbox": [ + 130, + 617, + 481, + 667 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 127, + 478, + 178 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 470, + 125 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 470, + 125 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 470, + 125 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons of various methods on our synthetic dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 127, + 478, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 127, + 478, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 127, + 478, + 178 + ], + "type": "table", + "html": "
MethodPSNRSSIMMethodPSNRSSIM
Pan et al. [46]24.090.808Zamir et al. [68]22.270.652
Dong et al. [13]23.850.799Ren et al. [52]25.480.765
Tao et al. [58]23.210.699Huo et al. [22]28.130.898
Kupyyn et al. [31]23.460.716Li et al. [35]24.490.744
Kaufman and Fattal [24]25.420.814Ours29.670.928
", + "image_path": "23737c746a1c76f492c58a1b09c55c6cfac69b22e06675f7ca204d01aa6cc1bd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 183, + 190, + 226 + ], + "blocks": [ + { + "bbox": [ + 133, + 183, + 190, + 226 + ], + "lines": [ + { + "bbox": [ + 133, + 183, + 190, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 190, + 226 + ], + "type": "image", + "image_path": "43b8f0b5ebbebc77bc31136610cf7ee93a26f98090309b37e600c1d3c287d131.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 228, + 175, + 236 + ], + "lines": [ + { + "bbox": [ + 148, + 228, + 175, + 236 + ], + "spans": [ + { + "bbox": [ + 148, + 228, + 175, + 236 + ], + "type": "text", + "content": "Blurred" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 192, + 184, + 248, + 226 + ], + "blocks": [ + { + "bbox": [ + 192, + 184, + 248, + 226 + ], + "lines": [ + { + "bbox": [ + 192, + 184, + 248, + 226 + ], + "spans": [ + { + "bbox": [ + 192, + 184, + 248, + 226 + ], + "type": "image", + "image_path": "af285bc82f7db4dae76fa6253c77d3b9818da4a493f338fa2ff83d2041e1c647.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 228, + 234, + 236 + ], + "lines": [ + { + "bbox": [ + 205, + 228, + 234, + 236 + ], + "spans": [ + { + "bbox": [ + 205, + 228, + 234, + 236 + ], + "type": "text", + "content": "Pan [46]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 249, + 184, + 306, + 226 + ], + "blocks": [ + { + "bbox": [ + 249, + 184, + 306, + 226 + ], + "lines": [ + { + "bbox": [ + 249, + 184, + 306, + 226 + ], + "spans": [ + { + "bbox": [ + 249, + 184, + 306, + 226 + ], + "type": "image", + "image_path": "b393c5771c8a45fac7c1f70625e7a4d79251c3546bd2ffeb8c17d655d8ea7f39.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 228, + 294, + 236 + ], + "lines": [ + { + "bbox": [ + 261, + 228, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 261, + 228, + 294, + 236 + ], + "type": "text", + "content": "Dong [13]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 184, + 364, + 226 + ], + "blocks": [ + { + "bbox": [ + 307, + 184, + 364, + 226 + ], + "lines": [ + { + "bbox": [ + 307, + 184, + 364, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 364, + 226 + ], + "type": "image", + "image_path": "aae5f2cb585eea31b31ac13fc0d902c40cba9c4bdaae571a06218475c1fced5f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 228, + 350, + 236 + ], + "lines": [ + { + "bbox": [ + 321, + 228, + 350, + 236 + ], + "spans": [ + { + "bbox": [ + 321, + 228, + 350, + 236 + ], + "type": "text", + "content": "Tao [58]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 184, + 421, + 226 + ], + "blocks": [ + { + "bbox": [ + 365, + 184, + 421, + 226 + ], + "lines": [ + { + "bbox": [ + 365, + 184, + 421, + 226 + ], + "spans": [ + { + "bbox": [ + 365, + 184, + 421, + 226 + ], + "type": "image", + "image_path": "fcaae1773cf4ef864ef08222304fd4a05d4265cc259ea562e24c829ee5311746.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 424, + 184, + 480, + 226 + ], + "blocks": [ + { + "bbox": [ + 424, + 184, + 480, + 226 + ], + "lines": [ + { + "bbox": [ + 424, + 184, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 424, + 184, + 480, + 226 + ], + "type": "image", + "image_path": "dd9c4692da8563a9d018080eadcf94875eb3563fd91df4f161f9899c9ccbf358.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 133, + 237, + 190, + 278 + ], + "blocks": [ + { + "bbox": [ + 133, + 237, + 190, + 278 + ], + "lines": [ + { + "bbox": [ + 133, + 237, + 190, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 190, + 278 + ], + "type": "image", + "image_path": "52c94a73e39b52b2928cd415dcca4f807d8740c471a8ce7663434f0cacbad417.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 281, + 180, + 289 + ], + "lines": [ + { + "bbox": [ + 143, + 281, + 180, + 289 + ], + "spans": [ + { + "bbox": [ + 143, + 281, + 180, + 289 + ], + "type": "text", + "content": "Zamir [67]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 192, + 237, + 248, + 278 + ], + "blocks": [ + { + "bbox": [ + 192, + 237, + 248, + 278 + ], + "lines": [ + { + "bbox": [ + 192, + 237, + 248, + 278 + ], + "spans": [ + { + "bbox": [ + 192, + 237, + 248, + 278 + ], + "type": "image", + "image_path": "25f9fe1055baa8a08732db49038e093c883ea5d42991c2b10e8a34c8f78b1b03.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 281, + 234, + 289 + ], + "lines": [ + { + "bbox": [ + 205, + 281, + 234, + 289 + ], + "spans": [ + { + "bbox": [ + 205, + 281, + 234, + 289 + ], + "type": "text", + "content": "Ren [52]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 249, + 237, + 306, + 278 + ], + "blocks": [ + { + "bbox": [ + 249, + 237, + 306, + 278 + ], + "lines": [ + { + "bbox": [ + 249, + 237, + 306, + 278 + ], + "spans": [ + { + "bbox": [ + 249, + 237, + 306, + 278 + ], + "type": "image", + "image_path": "99232cba2cc3c7f21c5649c049949f01c1f8d164a30860d74615b7f10d3fc736.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 281, + 292, + 289 + ], + "lines": [ + { + "bbox": [ + 263, + 281, + 292, + 289 + ], + "spans": [ + { + "bbox": [ + 263, + 281, + 292, + 289 + ], + "type": "text", + "content": "Huo [22]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 291, + 480, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 291, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 480, + 312 + ], + "type": "text", + "content": "Fig. 6: Visual results on our synthetic dataset. The estimated blur kernel is placed on the top-left corner for each method if available." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 307, + 237, + 364, + 278 + ], + "blocks": [ + { + "bbox": [ + 307, + 237, + 364, + 278 + ], + "lines": [ + { + "bbox": [ + 307, + 237, + 364, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 237, + 364, + 278 + ], + "type": "image", + "image_path": "452d85d7c795f606bfa56b598783ddbbd8cbe9b5259179d59e0018384d9c8875.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 281, + 347, + 289 + ], + "lines": [ + { + "bbox": [ + 324, + 281, + 347, + 289 + ], + "spans": [ + { + "bbox": [ + 324, + 281, + 347, + 289 + ], + "type": "text", + "content": "Li [35]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 365, + 237, + 421, + 278 + ], + "blocks": [ + { + "bbox": [ + 374, + 228, + 413, + 236 + ], + "lines": [ + { + "bbox": [ + 374, + 228, + 413, + 236 + ], + "spans": [ + { + "bbox": [ + 374, + 228, + 413, + 236 + ], + "type": "text", + "content": "Kupyn [31]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 237, + 421, + 278 + ], + "lines": [ + { + "bbox": [ + 365, + 237, + 421, + 278 + ], + "spans": [ + { + "bbox": [ + 365, + 237, + 421, + 278 + ], + "type": "image", + "image_path": "ff405bbabdbfe218d7d8e9754d8feed9a5573812e86bdf538af7be5a83aab8c7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 281, + 403, + 289 + ], + "lines": [ + { + "bbox": [ + 385, + 281, + 403, + 289 + ], + "spans": [ + { + "bbox": [ + 385, + 281, + 403, + 289 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 424, + 237, + 480, + 278 + ], + "blocks": [ + { + "bbox": [ + 429, + 228, + 476, + 236 + ], + "lines": [ + { + "bbox": [ + 429, + 228, + 476, + 236 + ], + "spans": [ + { + "bbox": [ + 429, + 228, + 476, + 236 + ], + "type": "text", + "content": "Kaufman [24]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 237, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 424, + 237, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 424, + 237, + 480, + 278 + ], + "type": "image", + "image_path": "3e06843ae7954a2f6466dc400f8b293317c556db22518cf4af43c54581da9dbb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 281, + 476, + 289 + ], + "lines": [ + { + "bbox": [ + 428, + 281, + 476, + 289 + ], + "spans": [ + { + "bbox": [ + 428, + 281, + 476, + 289 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": "train the kernel initializer " + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "inline_equation", + "content": "E(\\cdot; \\theta_E^*)" + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": " that is a ResNet-18 [20]. To train the kernel initializer, we randomly crop " + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": " patches from the sources in OpenImages [32] and convolve them with blur kernels to get blurry images. The Adam optimizer [25] with its default configuration in PyTorch [49] is adopted. The initial learning rate is 1e-4 and reduced to 1e-5 as the loss becomes stable and finally to 1e-6. In the BID process, we jointly estimate the blur kernel " + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{k}" + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": " and the clean image " + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": " by optimizing " + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k" + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\theta_x" + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": " according to Eq. (6). The initial learning rate of " + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k" + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\theta_x" + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": " are set as 5e-4 and 1e-2, respectively, and decayed following [52]. For the architecture of the DIP network " + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "inline_equation", + "content": "G_x(\\cdot; \\theta_x)" + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": ", we follow the settings in SelfDeblur [52]." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 454, + 265, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 265, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 265, + 466 + ], + "type": "text", + "content": "4.2 Experimental results" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 474, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 557 + ], + "type": "text", + "content": "Results on our synthetic dataset. We first verify the effectiveness of the proposed method on the synthetic dataset by ourselves. Nine comparison methods are considered, including two traditional model-based ones (Pan et al. [46], Dong et al. [13]), four supervised deep learning ones (Tao et al. [58], Kupyn et al. [31], Kaufman and Fattal [24], Zamir et al. [68]), and three DIP-based ones (Ren et al. [52], Huo et al. [22], Li et al. [35]). We follow the default settings in their papers for these methods or tune them by ourselves for the best performance." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "The quantitative results of all competing methods are summarized in Table 1. As can be seen from this table, the proposed method achieves the best performance in terms of both PSNR and SSIM, demonstrating its effectiveness. The superiority of our method can be visually observed from Fig. 6. Specifically, the traditional model-based methods are able to obtain accurate estimations for the blur kernel, but the BID results are not very satisfactory due to the insufficiency of the handcrafted image priors. Most of the supervised deep learning methods generate over-smoothed outcomes, mainly attributed to the lack of exploitation of the physical blur model. Even though achieving relatively better results, the" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 91, + 447, + 102 + ], + "type": "text", + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 126, + 477, + 308 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 474, + 126 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 474, + 126 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 474, + 126 + ], + "type": "text", + "content": "Table 2: Quantitative results of various methods on the dataset by Lai et al. [33]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 126, + 477, + 308 + ], + "lines": [ + { + "bbox": [ + 137, + 126, + 477, + 308 + ], + "spans": [ + { + "bbox": [ + 137, + 126, + 477, + 308 + ], + "type": "table", + "html": "
MethodManmadeNaturalPeopleSaturatedTextAverage
Cho and Lee [10]16.35/0.38920.14/0.52019.90/0.55614.05/0.49314.87/0.44317.06/0.480
Krishnan et al. [29]15.73/0.40819.44/0.52621.50/0.64714.09/0.50715.40/0.48517.23/0.514
Xu et al. [62]17.99/0.59921.58/0.67924.40/0.81314.53/0.53817.64/0.66819.23/0.659
Perrone and Favaro [50]17.41/0.55121.04/0.67622.77/0.73514.24/0.51116.94/0.59318.48/0.613
Pan et al. [46]18.25/0.56122.15/0.67222.72/0.71016.01/0.59116.69/0.53019.16/0.613
Dong et al. [13]17.30/0.43221.18/0.58621.89/0.65215.85/0.55416.18/0.48018.48/0.541
Tao et al. [58]16.77/0.34219.90/0.45921.41/0.61615.12/0.50615.41/0.43717.72/0.472
Kupyn et al. [31]17.05/0.36720.37/0.48921.70/0.64215.20/0.52215.86/0.47618.04/0.499
Kaufman and Fattal [24]19.36/0.65023.14/0.75626.78/0.87316.40/0.65417.92/0.68320.72/0.723
Zamir et al. [68]16.73/0.34119.80/0.44720.88/0.58715.14/0.50514.49/0.38317.41/0.453
Ren et al. [52]19.00/0.61722.94/0.71526.43/0.78118.38/0.68822.82/0.77621.91/0.715
Huo et al. [22]22.51/0.82026.20/0.90131.38/0.95318.14/0.69127.72/0.93325.19/0.860
Li et al. [35]20.81/0.72124.07/0.78128.45/0.86014.81/0.52723.10/0.83422.25/0.744
Ours24.67/0.94228.12/0.96932.86/0.97318.98/0.73728.91/0.95126.71/0.914
BlurredCho [10]Krishnan [29]Xu [62]Perrone [50]Pan [46]Dong [13]
Kaufman [24]Zamir [67]Ren [52]Huo [22]Li [35]OursGround truth
", + "image_path": "7b835e463aba1fc986108423bfa82d6559ee48e4b7323261479fb31c3f3849eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 308, + 481, + 331 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 331 + ], + "type": "text", + "content": "Fig. 7: Visual results on the dataset by Lai et al. [33]. The estimated blur kernel is placed on the top-left corner of each method if available." + } + ] + } + ], + "index": 4, + "type": "text" + }, + { + "bbox": [ + 130, + 335, + 480, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 395 + ], + "type": "text", + "content": "existing DIP-based methods still have some limitations. For example, Ren et al.'s method performs unsatisfactorily, and the structural details in the images by Huo et al. and Li et al.'s methods are not sharp enough. In contrast to these methods, our method can not only accurately estimate the blur kernel, but also recover more image details with fewer distortions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 520 + ], + "type": "text", + "content": "Results on the dataset by Lai et al. [33]. We evaluate our proposed method on the challenging BID benchmark by Lai et al. [33]. This dataset consists of 100 blurry images synthesized by applying 4 blur kernels, with sizes ranging from " + }, + { + "bbox": [ + 130, + 399, + 482, + 520 + ], + "type": "inline_equation", + "content": "31 \\times 31" + }, + { + "bbox": [ + 130, + 399, + 482, + 520 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 399, + 482, + 520 + ], + "type": "inline_equation", + "content": "75 \\times 75" + }, + { + "bbox": [ + 130, + 399, + 482, + 520 + ], + "type": "text", + "content": ", to 25 clean images. These 25 clean images are categorized into 5 groups, namely Manmade, Natural, People, Saturated and Text. We compare our method with 13 existing methods, including six traditional model-based methods (Cho and Li [10], Krishnan et al. [29], Xu et al. [62], Perrone and Favaro [50], Pan et al [46], Dong et al. [13]), four supervised deep learning methods (Tao et al. [58], Kupyn et al. [31], Kaufman and Fattal [24], Zamir et al. [68]), and three DIP-based methods (Ren et al. [52], Huo et al. [22], Li et al. [35])." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Table 2 reports the quantitative results of all competing methods, in terms of PSNR and SSIM. We can see that our method significantly outperforms existing methods in all of the five categories. Notably, our method is the only one that achieves an SSIM value higher than 0.9 on average. The visual results on one typical example are shown in Fig. 7. It can be observed that most of the traditional model-based methods fail to estimate the blur kernel due to the relatively larger kernel size in this dataset. The supervised deep learning methods still produce distorted results as before, because they do not carefully consider the physical blur model. In comparison, DIP-based methods, including ours, can accurately estimate the blur kernel and further achieve promising deblurring results. Among them, our method recovers more details, such as hairs, with reference to the ground truth image, showing its superiority." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 114, + 203, + 148 + ], + "blocks": [ + { + "bbox": [ + 152, + 114, + 203, + 148 + ], + "lines": [ + { + "bbox": [ + 152, + 114, + 203, + 148 + ], + "spans": [ + { + "bbox": [ + 152, + 114, + 203, + 148 + ], + "type": "image", + "image_path": "fc7cceebc0da02d3d809e6c86bd3ab2625120508bbecd04b0b50727fc4a9760f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 148, + 192, + 157 + ], + "lines": [ + { + "bbox": [ + 164, + 148, + 192, + 157 + ], + "spans": [ + { + "bbox": [ + 164, + 148, + 192, + 157 + ], + "type": "text", + "content": "Blurred" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 114, + 255, + 148 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 255, + 148 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 255, + 148 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 255, + 148 + ], + "type": "image", + "image_path": "fd2971b66365791da9ea173a4ff31d367aca1ec139aecb9325500964eb6dba0b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 149, + 243, + 157 + ], + "lines": [ + { + "bbox": [ + 215, + 149, + 243, + 157 + ], + "spans": [ + { + "bbox": [ + 215, + 149, + 243, + 157 + ], + "type": "text", + "content": "Pan [46]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 257, + 114, + 306, + 148 + ], + "blocks": [ + { + "bbox": [ + 257, + 114, + 306, + 148 + ], + "lines": [ + { + "bbox": [ + 257, + 114, + 306, + 148 + ], + "spans": [ + { + "bbox": [ + 257, + 114, + 306, + 148 + ], + "type": "image", + "image_path": "6947626b66b3e2b2be3680527f9e40502b414739e2177ae844165fe0878b2ab4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 358, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 358, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 358, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 358, + 148 + ], + "type": "image", + "image_path": "00f15976fa312fb43b9571d1acb2ffb6294bdc0e22588db5430b16435b11df4e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 359, + 114, + 409, + 148 + ], + "blocks": [ + { + "bbox": [ + 359, + 114, + 409, + 148 + ], + "lines": [ + { + "bbox": [ + 359, + 114, + 409, + 148 + ], + "spans": [ + { + "bbox": [ + 359, + 114, + 409, + 148 + ], + "type": "image", + "image_path": "8d8154e4e672373f55984ed03142bbabe49ace304d3021a14244ce625ee3f785.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 460, + 148 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 460, + 148 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 460, + 148 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 460, + 148 + ], + "type": "image", + "image_path": "7ace74815d9bd1b1bbc9f6516cbeaf182f28aef7c861931e8ecad437f2b2a43d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 179, + 158, + 228, + 192 + ], + "blocks": [ + { + "bbox": [ + 179, + 158, + 228, + 192 + ], + "lines": [ + { + "bbox": [ + 179, + 158, + 228, + 192 + ], + "spans": [ + { + "bbox": [ + 179, + 158, + 228, + 192 + ], + "type": "image", + "image_path": "297dd9de7ad9871b3605a403349973206ebaa15c1f53ebc25fba314ebfbd9d3e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 194, + 222, + 202 + ], + "lines": [ + { + "bbox": [ + 185, + 194, + 222, + 202 + ], + "spans": [ + { + "bbox": [ + 185, + 194, + 222, + 202 + ], + "type": "text", + "content": "Zamir [67]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 230, + 158, + 280, + 192 + ], + "blocks": [ + { + "bbox": [ + 264, + 149, + 298, + 157 + ], + "lines": [ + { + "bbox": [ + 264, + 149, + 298, + 157 + ], + "spans": [ + { + "bbox": [ + 264, + 149, + 298, + 157 + ], + "type": "text", + "content": "Dong [13]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 158, + 280, + 192 + ], + "lines": [ + { + "bbox": [ + 230, + 158, + 280, + 192 + ], + "spans": [ + { + "bbox": [ + 230, + 158, + 280, + 192 + ], + "type": "image", + "image_path": "33ea5e7566dbd8bfd5c56a0897edd31cace5b6d504dace6aa07bc5a8355a98c4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 194, + 270, + 202 + ], + "lines": [ + { + "bbox": [ + 240, + 194, + 270, + 202 + ], + "spans": [ + { + "bbox": [ + 240, + 194, + 270, + 202 + ], + "type": "text", + "content": "Ren [52]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 282, + 158, + 332, + 192 + ], + "blocks": [ + { + "bbox": [ + 318, + 149, + 347, + 157 + ], + "lines": [ + { + "bbox": [ + 318, + 149, + 347, + 157 + ], + "spans": [ + { + "bbox": [ + 318, + 149, + 347, + 157 + ], + "type": "text", + "content": "Tao [58]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 282, + 158, + 332, + 192 + ], + "lines": [ + { + "bbox": [ + 282, + 158, + 332, + 192 + ], + "spans": [ + { + "bbox": [ + 282, + 158, + 332, + 192 + ], + "type": "image", + "image_path": "79482bbf21205a7733c160c8bfe4524911babea121430d68c474b45fe4db9796.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 194, + 321, + 202 + ], + "lines": [ + { + "bbox": [ + 291, + 194, + 321, + 202 + ], + "spans": [ + { + "bbox": [ + 291, + 194, + 321, + 202 + ], + "type": "text", + "content": "Huo [22]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 333, + 158, + 384, + 192 + ], + "blocks": [ + { + "bbox": [ + 364, + 149, + 403, + 157 + ], + "lines": [ + { + "bbox": [ + 364, + 149, + 403, + 157 + ], + "spans": [ + { + "bbox": [ + 364, + 149, + 403, + 157 + ], + "type": "text", + "content": "Kupyn [31]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 158, + 384, + 192 + ], + "lines": [ + { + "bbox": [ + 333, + 158, + 384, + 192 + ], + "spans": [ + { + "bbox": [ + 333, + 158, + 384, + 192 + ], + "type": "image", + "image_path": "7faa62d02101d97806e48e440dc64acc46992d65113093dde0e909e457b2a23b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 194, + 370, + 202 + ], + "lines": [ + { + "bbox": [ + 347, + 194, + 370, + 202 + ], + "spans": [ + { + "bbox": [ + 347, + 194, + 370, + 202 + ], + "type": "text", + "content": "Li [35]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 385, + 158, + 435, + 192 + ], + "blocks": [ + { + "bbox": [ + 412, + 149, + 459, + 157 + ], + "lines": [ + { + "bbox": [ + 412, + 149, + 459, + 157 + ], + "spans": [ + { + "bbox": [ + 412, + 149, + 459, + 157 + ], + "type": "text", + "content": "Kaufman [24]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 385, + 158, + 435, + 192 + ], + "lines": [ + { + "bbox": [ + 385, + 158, + 435, + 192 + ], + "spans": [ + { + "bbox": [ + 385, + 158, + 435, + 192 + ], + "type": "image", + "image_path": "0b7d3b82dab6bb0f200c43fba73f74a3daf88f6cf1a517aa23cef40d81663e3a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 194, + 418, + 202 + ], + "lines": [ + { + "bbox": [ + 401, + 194, + 418, + 202 + ], + "spans": [ + { + "bbox": [ + 401, + 194, + 418, + 202 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 139, + 233, + 286, + 346 + ], + "blocks": [ + { + "bbox": [ + 131, + 203, + 480, + 225 + ], + "lines": [ + { + "bbox": [ + 131, + 203, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 131, + 203, + 480, + 225 + ], + "type": "text", + "content": "Fig. 8: Example results on the real blurry image provided by Lai et al. [33]. The estimated blur kernel is placed on the top-left corner of each method if available." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 233, + 286, + 346 + ], + "lines": [ + { + "bbox": [ + 139, + 233, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 139, + 233, + 286, + 346 + ], + "type": "image", + "image_path": "3f2c27a01f07e76cb7bac61d96328694bc5d8056011dfe647b0d9d69f545bdc2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 235, + 460, + 291 + ], + "lines": [ + { + "bbox": [ + 307, + 235, + 460, + 291 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 460, + 291 + ], + "type": "text", + "content": "Fig.9: Performance comparison (PSNR) of the DIP-based BID method (SelfDeblur [52]) and our method in different iteration steps on the dataset by Lai et al. [33]." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 131, + 350, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 480, + 399 + ], + "type": "text", + "content": "Example results on real blurry images. We further apply comparison BID methods to real blurry images collected by Lai et al. [33], and show one example in Fig. 8. As can be seen, on this very challenging image, none of the methods generate perfect results, while our method provides a relatively promising one." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 132, + 415, + 236, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 236, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 236, + 428 + ], + "type": "text", + "content": "4.3 Ablation study" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 435, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 480, + 471 + ], + "type": "text", + "content": "As shown in the previous experiments, the proposed method has demonstrated its effectiveness on various blurry images. In this section, we provide further analyses for better understanding its advantageous properties." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 472, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 592 + ], + "type": "text", + "content": "Convergence speed. One major methodological difference between our method and existing DIP-based methods is that we can initialize the blur kernel with pretrained models instead of random sampling. This property can not only result in more stable and better performance but also improve the convergence speed. To verify this point, we compare the PSNR values of our methods with that of Ren et al.'s [52] across various iterations on the Lai dataset in Fig. 9. It can be seen that the performance of our method in the 400-th iteration is comparable with that of Ren et al.'s method in the 1000-th iteration. This faster convergence can be mostly attributed to the better initialization and characterization of the blur kernel, since both of them adopt the same DIP network for the image." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Optimization strategy for blur kernel. As discussed in Sec. 3.5, during the BID process, we optimize the higher-dimensional feature map " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_k" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " instead of the original latent code " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_k" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for the blur kernel. Now we empirically show the different behaviors of the two optimization strategies. As a reference, the results by optimizing the entire kernel generator are also reported, which has a larger search space for optimization. The detailed comparison results on the Lai dataset are" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 90, + 448, + 102 + ], + "type": "text", + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 137, + 481, + 171 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 3: Performance comparison (PSNR/SSIM) of different kernel optimization strategy on the dataset by Lai et al. [33]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 137, + 481, + 171 + ], + "lines": [ + { + "bbox": [ + 133, + 137, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 137, + 481, + 171 + ], + "type": "table", + "html": "
Kernel size31515575Average
wk27.89/0.93226.31/0.90125.91/0.91626.73/0.90826.71/0.914
zk27.11/0.92723.46/0.83022.75/0.82522.43/0.76323.94/0.836
zk and θk24.38/0.83022.33/0.71624.45/0.81924.17/0.81423.83/0.795
", + "image_path": "65e32587506f2ff2b799e97b8b718e6167cc9f284f2ac5996dba74ed975040da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 157, + 198, + 456, + 217 + ], + "blocks": [ + { + "bbox": [ + 132, + 175, + 479, + 197 + ], + "lines": [ + { + "bbox": [ + 132, + 175, + 479, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 479, + 197 + ], + "type": "text", + "content": "Table 4: Performance comparison (PSNR/SSIM) of different kernel initialization strategies on the dataset by Lai et al. [33]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 198, + 456, + 217 + ], + "lines": [ + { + "bbox": [ + 157, + 198, + 456, + 217 + ], + "spans": [ + { + "bbox": [ + 157, + 198, + 456, + 217 + ], + "type": "table", + "html": "
BaselineRandomG(z0; θk*)G(E(y; θE*); θk*)
Metrics19.51/0.56323.00/0.74326.71/0.914
", + "image_path": "124d902cc726db409cff3d21cc19ac65b119c5195817170da41fd0494dd23938.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "text", + "content": "summarized in Table 3. We can observe that optimizing the whole generator is not competitive due to the too-large search space. When the blur kernel is relatively small, the other two strategies have similar performance, while optimizing " + }, + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k" + }, + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "text", + "content": " is slightly better. As the blur kernel becomes larger, the performance by optimizing " + }, + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{z}_k" + }, + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "text", + "content": " significantly drops, while optimizing " + }, + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k" + }, + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "text", + "content": " still produces promising results. These observations demonstrate the reasonability of the proposed optimization strategy and also emphasize the importance of a proper search space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 310, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 481, + 514 + ], + "type": "text", + "content": "Effect of kernel initialization. In our BID framework, a pre-trained encoder is employed to initialize the blur kernel. To show the necessity of such an initializer, we compare the proposed initialization strategy to two baselines. The first is the random initialization that randomly samples a latent code " + }, + { + "bbox": [ + 130, + 310, + 481, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 310, + 481, + 514 + ], + "type": "text", + "content": " for the blur kernel. The second is to initialize the kernel with all elements being the same, which can be seen as an \"average\" kernel, and then find the corresponding " + }, + { + "bbox": [ + 130, + 310, + 481, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 310, + 481, + 514 + ], + "type": "text", + "content": " by GAN-inversion. It should be noted that both of these two initialization strategies are independent of the blurry image. Table 4 lists the comparison results on the Lai dataset. We can see that the random initialization strategy produces the worst result since the randomly initialized kernel might largely deviate from the target without any constraint. The fixed \"average\" kernel performs better, and indeed already outperforms many other methods in Table 2. That's because this strategy, in some sense, can be regarded as an \"average\" approximation to any blur kernel. Not surprisingly, the proposed initialization with a pre-trained encoder obtains the best performance as the initialized kernel has been very close to the ground-truth one, as shown in Fig. 5. These results substantiate the significant benefits brought by the pre-trained kernel initializer for the BID task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": "In this paper, we have proposed a new framework for the BID task, by virtue of deep generative model. Within this framework, we first pre-train a kernel generator as a DGP for blur kernels and a kernel initializer that can offer a well-initialized kernel. Then during the BID process, the blur kernel is initialized in the latent feature space and jointly optimized with the DIP network for the final result. Comprehensive experiments have been conducted and demonstrated the effectiveness of the proposed method. The ablation study verifies the necessity of each component in our framework. In the future, we will try to generalize our method to non-uniform blur modeling, overcoming its current limitations." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 174 + ], + "type": "text", + "content": "This work was supported in part by the NSFC Projects (No. 12226004, 62076196, 62331028, 62272375). The authors thank Prof. Jiangxin Dong for providing the code for generating the blur kernels with the type of Lai dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 244 + ], + "type": "text", + "content": "1. Asim, M., Shamshad, F., Ahmed, A.: Blind image deconvolution using deep generative priors. IEEE Transactions on Computational Imaging (TCI) 6, 1493-1506 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "text", + "content": "2. Bishop, T.E., Babacan, S.D., Amizic, B., Katsaggelos, A.K., Chan, T., Molina, R.: Blind image deconvolution: problem formulation and existing approaches. In: Blind Image Deconvolution, pp. 21-62. CRC press (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 481, + 310 + ], + "type": "text", + "content": "3. Bredell, G., Flouris, K., Chaitanya, K., Erdil, E., Konukoglu, E.: Explicitly minimizing the blur error of variational autoencoders. In: Proceedings of International Conference on Learning Representations (ICLR) (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 311, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 481, + 342 + ], + "type": "text", + "content": "4. Cai, J.F., Ji, H., Liu, C., Shen, Z.: Blind motion deblurring from a single image using sparse approximation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 104-111 (2009)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "type": "text", + "content": "5. Cai, J., Zuo, W., Zhang, L.: Dark and bright channel prior embedded network for dynamic scene deblurring. IEEE Transactions on Image Processing (TIP) 29, 6885-6897 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 396 + ], + "type": "text", + "content": "6. Chakrabarti, A.: A neural approach to blind motion deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 221-235 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "text", + "content": "7. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14245-14254 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 430, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 430, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 430, + 481, + 449 + ], + "type": "text", + "content": "8. Chan, T.F., Wong, C.K.: Total variation blind deconvolution. IEEE Transactions on Image Processing (TIP) 7(3), 370-375 (1998)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 481, + 482 + ], + "type": "text", + "content": "9. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1742-1750 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 483, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 481, + 503 + ], + "type": "text", + "content": "0. Cho, S., Lee, S.: Fast motion deblurring. ACM Transactions on Graphics (TOG) 28(5), 1-8 (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "type": "text", + "content": "1. Dinh, L., Krueger, D., Bengio, Y.: NICE: Non-linear Independent Components Estimation (Apr 2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 525, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 556 + ], + "type": "text", + "content": "2. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using Real NVP. In: Proceedings of International Conference on Learning Representations (ICLR) (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 558, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 481, + 590 + ], + "type": "text", + "content": "3. Dong, J., Pan, J., Su, Z., Yang, M.H.: Blind image deblurring with outlier handling. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2478-2486 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "type": "text", + "content": "4. Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T.: Removing camera shake from a single photograph. ACM Transactions on Graphics (TOG) 25(3), 787-794 (2006)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "5. Gandelsman, Y., Shocher, A., Irani, M.: \"Double-DIP\": Unsupervised image decomposition via coupled deep-image-priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11018-11027 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "type": "text", + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "16. Gong, D., Yang, J., Liu, L., Zhang, Y., Reid, I., Shen, C., Van Den Hengel, A., Shi, Q.: From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2319-2328 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "text", + "content": "17. Gong, K., Catana, C., Qi, J., Li, Q.: Pet image reconstruction using deep image prior. IEEE Transactions on Medical Imaging (TMI) 38(7), 1655-1665 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "text", + "content": "18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in Neural Information Processing Systems (NeurIPS) (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 236 + ], + "type": "text", + "content": "19. Guan, S., Tai, Y., Ni, B., Zhu, F., Huang, F., Yang, X.: Collaborative learning for faster stylegan embedding. arXiv preprint arXiv:2007.01758 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "text", + "content": "21. Hradis, M., Kotera, J., Zemcik, P., Sroubek, F.: Convolutional neural networks for direct text deblurring. In: Proceedings of British Machine Vision Conference (BMVC) (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "22. Huo, D., Masoumzadeh, A., Kushol, R., Yang, Y.H.: Blind image deconvolution using variational deep image prior. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(10), 11472-11483 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4401-4410 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "text", + "content": "24. Kaufman, A., Fattal, R.: Deblurring using analysis-synthesis networks pair. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5811-5820 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "text", + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Proceedings of the International Conference for Learning Representations (ICLR) (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 440 + ], + "type": "text", + "content": "26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: Proceedings of International Conference on Learning Representations (ICLR) (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 440, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 460 + ], + "type": "text", + "content": "27. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. In: Advances in Neural Information Processing Systems (NeurIPS) (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "text", + "content": "28. Krishnan, D., Fergus, R.: Fast image deconvolution using hyper-laplacian priors. In: Advances in Neural Information Processing Systems (NeurIPS) (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "text", + "content": "29. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 233-240 (2011)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 514, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 557 + ], + "type": "text", + "content": "30. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8183-8192 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "text", + "content": "31. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8878-8887 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 590, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 633 + ], + "type": "text", + "content": "32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV) 128(7), 1956-1981 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "33. Lai, W.S., Huang, J.B., Hu, Z., Ahuja, N., Yang, M.H.: A comparative study for single image blind deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1701-1709 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "34. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1964-1971 (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "35. Li, J., Wang, W., Nan, Y., Ji, H.: Self-supervised blind motion deblurring with deep expectation maximization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13986-13996 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "36. Liang, J., Zhang, K., Gu, S., Gool, L.V., Timofte, R.: Flow-based kernel prior with application to blind super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10596-10605 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 740-755 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 252, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 481, + 284 + ], + "type": "text", + "content": "38. Liu, G., Chang, S., Ma, Y.: Blind image deblurring using spectral properties of convolution operators. IEEE Transactions on Image Processing (TIP) 23(12), 5047-5056 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 329 + ], + "type": "text", + "content": "39. Luo, Y.S., Zhao, X.L., Jiang, T.X., Zheng, Y.B., Chang, Y.: Hyperspectral mixed noise removal via spatial-spectral constrained unsupervised deep image prior. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS) 14, 9435-9449 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 331, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 374 + ], + "type": "text", + "content": "40. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2437-2445 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "41. Michaeli, T., Irani, M.: Blind deblurring using internal patch recurrence. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 783-798 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 453 + ], + "type": "text", + "content": "42. Molina, R., Katsaggelos, A.K., Abad, J., Mateos, J.: A bayesian approach to blind deconvolution based on dirichlet distributions. In: Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP). vol. 4, pp. 2809-2812 (1997)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "43. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3883-3891 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 487, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 530 + ], + "type": "text", + "content": "44. Pan, J., Dong, J., Liu, Y., Zhang, J., Ren, J., Tang, J., Tai, Y.W., Yang, M.H.: Physics-based generative adversarial models for image restoration and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 43(7), 2449-2462 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "text", + "content": "45. Pan, J., Hu, Z., Su, Z., Yang, M.H.: Deblurring text images via 10-regularized intensity and gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2901-2908 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "46. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1628-1636 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "text", + "content": "47. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Deblurring images via dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(10), 2315-2328 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "48. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 7474-7489 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "type": "text", + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "49. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "type": "text", + "content": "50. Perrone, D., Favaro, P.: Total variation blind deconvolution: The devil is in the details. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2909-2916 (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 225 + ], + "type": "text", + "content": "51. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. In: Proceedings of International Conference on Learning Representations (ICLR) (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "text", + "content": "52. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3341-3350 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "text", + "content": "53. Ren, W., Cao, X., Pan, J., Guo, X., Zuo, W., Yang, M.H.: Image deblurring via enhanced low-rank prior. IEEE Transactions on Image Processing (TIP) 25(7), 3426-3437 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "text", + "content": "54. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 184-201 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 322, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 343 + ], + "type": "text", + "content": "55. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. ACM Transactions on Graphics (TOG) 27(3), 1-10 (2008)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "text", + "content": "56. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 769-777 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "57. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: Proceedings of the IEEE International Conference on Computational Photography (ICCP). pp. 1-8 (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "58. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8174-8182 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "59. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9446-9454 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "60. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9168-9178 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "61. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(3), 3121-3138 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "text", + "content": "62. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1107-1114 (2013)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 601 + ], + "type": "text", + "content": "63. Yan, Y., Ren, W., Guo, Y., Wang, R., Cao, X.: Image deblurring via extreme channels prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4003-4011 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "text", + "content": "64. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 672-681 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "65. Yue, Z., Yong, H., Zhao, Q., Zhang, L., Meng, D., Wong, K.Y.K.: Deep variational network toward blind image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 348 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "66. Yue, Z., Zhao, Q., Xie, J., Zhang, L., Meng, D., Wong, K.Y.K.: Blind image superresolution with elaborate degradation modeling on noise and kernel. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2118-2128 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 482, + 205 + ], + "type": "text", + "content": "67. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 482, + 247 + ], + "type": "text", + "content": "68. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14821-14831 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 248, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 482, + 282 + ], + "type": "text", + "content": "69. Zhao, Z., Xiong, B., Wang, L., Ou, Q., Yu, L., Kuang, F.: Retinexdip: A unified deep framework for low-light image enhancement. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) 32(3), 1076-1088 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 482, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 482, + 313 + ], + "type": "text", + "content": "70. Zhou, X., Mateos, J., Zhou, F., Molina, R., Katsaggelos, A.K.: Variational dirichlet blur kernel estimation. IEEE Transactions on Image Processing (TIP) 24(12), 5127-5139 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 314, + 482, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 482, + 348 + ], + "type": "text", + "content": "71. Zuo, W., Ren, D., Zhang, D., Gu, S., Zhang, L.: Learning iteration-wise generalized shrinkage-thresholding operators for blind deconvolution. IEEE Transactions on Image Processing (TIP) 25(4), 1751-1764 (2016)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 144, + 90, + 447, + 102 + ], + "type": "text", + "content": "BID by Generative-based Kernel Prior and Initializer via Latent Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Blind image deblurring with noise-robust kernel estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_content_list.json b/2024/Blind image deblurring with noise-robust kernel estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2382f7465a9e933c85c60e509d578db01d2a9c8a --- /dev/null +++ b/2024/Blind image deblurring with noise-robust kernel estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_content_list.json @@ -0,0 +1,2173 @@ +[ + { + "type": "text", + "text": "Blind image deblurring with noise-robust kernel estimation", + "text_level": 1, + "bbox": [ + 227, + 140, + 776, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chanseok Lee $^{1}$ , Jeongsol Kim $^{1}$ , Seungmin Lee $^{1}$ , Jaehwang Jung $^{2}$ , Yunje Cho $^{2}$ , Taejoong Kim $^{2}$ , Taeyong Jo $^{2}$ , Myungjun Lee $^{2}$ , and Mooseok Jang $^{1,3}$", + "bbox": [ + 218, + 212, + 782, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Bio and Brain Engineering, KAIST, Daejeon, South Korea", + "bbox": [ + 243, + 268, + 753, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Advanced Process Development Team 4, Semiconductor R&D Center, Samsung Electronics, South Korea", + "bbox": [ + 225, + 282, + 772, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ KAIST Institute for Health Science and Technology, KAIST, Daejeon, South Korea {cslee, mooseok}@kaist.ac.kr", + "bbox": [ + 215, + 311, + 784, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Blind deblurring is an ill-posed inverse problem involving the retrieval of a clear image and blur kernel from a single blurry image. The challenge arises considerably when strong noise, where its level remains unknown, is introduced. Existing blind deblurring methods are highly susceptible to noise due to overfitting and disturbances in the solution space. Here, we propose a blind deblurring method based on a noise-robust kernel estimation function and deep image prior (DIP). Specifically, the proposed kernel estimation function effectively estimates the blur kernel even for strongly noisy blurry images given a clear image and optimal condition. Therefore, DIP is adopted for the generation of a clear image to leverage its natural image prior. Additionally, the multiple kernel estimation scheme is designed to address a wide range of unknown noise levels. Extensive experimental studies, including simulated images and real-world examples, demonstrate the superior deblurring performance of the proposed method. The official code is uploaded in https://github.com/csleemooo/BD_noise_robust_kernel_estimation.", + "bbox": [ + 259, + 378, + 740, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Blind deblurring $\\cdot$ Deep image prior $\\cdot$ Gaussian noise", + "bbox": [ + 261, + 614, + 702, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 655, + 375, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Optical imaging systems such as brightfield microscopes, fluorescence microscopes, or scanning electron microscopes (SEM) have a crucial role in histopathology, inspection, and clinical or biological research by providing complicated structural information of target imaging samples with high-resolution images. However, perturbation or misalignment (e.g. defocus or aberrations) of the imaging system due to its instability or systemic error brings blurry effects on the captured image. Furthermore, additive Gaussian Noise severely degrades the captured image quality. Such degradation process can be described as follows:", + "bbox": [ + 212, + 688, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\ny _ {m} = x * h + n _ {g} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 825, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": ", where $y_{m} \\in \\mathbb{R}^{N}$ is measured blurry and noisy image, $x \\in \\mathbb{R}^{N}$ is a clear image, * denotes the convolution operator, $h \\in \\mathbb{R}^{M}$ is blur kernel or point spread function (PSF) of the imaging system, and $n_g$ is additive noise. Retrieving clean and focused image $x$ as well as blur kernel $h$ from a single degraded image $y_{m}$ is well known as blind deconvolution or blind image deblurring and has been actively investigated for decades due to its ill-posedness. Although many previous approaches such as edge-selection-based method [5,16,37,38] or MAP-based approach [4,15,18,23,29-31,41,44] show promising results, deblurring performances significantly drop as the noise level increases.", + "bbox": [ + 212, + 145, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In general, there is a trade-off when simultaneously trying to eliminate both the blur effect and noise. For example, when focusing on removing the blur effect, the estimated image may be severely corrupted by noise or other high-frequency artifacts. On the other hand, when focusing on noise removal, there is a tendency to overly emphasize smoothness, compromising the fine details of the image. So, balancing the strength of deblurring and noise suppression is important to successfully reconstruct $x$ and $h$ . Moreover, in cases where the noise level cannot be accurately determined, the difficulty of the deblurring task increases significantly.", + "bbox": [ + 212, + 282, + 787, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper aims to address the challenge of blind deblurring in the presence of strong and an unknown noise level. Instead of relying on sharp edge-selective maps or prior information to estimate the blur kernel, we introduce a novel kernel estimation function. This function analytically computes the blur kernel when provided with a clear image, blurry image, and noise suppression factor. The first characteristic of the proposed function is its ability to estimate the blur kernel effectively even in scenarios with strong noise. Specifically, the noise suppression factor included in the kernel estimation function is pivotal in mitigating noise, and its optimal condition is derived accordingly. To further ameliorate this characteristic, we designed a multiple kernel estimation scheme, which effectively handles a wide range of noise levels, to perform blind deblurring even in situations where the noise level is unknown. The second characteristic is that the estimation of only the clear image is sufficient for estimating a blur kernel. In other words, the amount of information required for solving a blind deblurring problem is much less than that needed by conventional methods, and we validated this through maximum a posteriori (MAP) analysis. This characteristic is fully leveraged by employing DIP, known as one of the most effective methods for capturing priors of natural images and solving inverse mapping problems [39]. The inherent robustness of DIP to noise, combined with the noise-robust multiple kernel estimation scheme, enables superior performance in blind deblurring tasks.", + "bbox": [ + 212, + 417, + 787, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper can be summarized as follows:", + "bbox": [ + 240, + 734, + 730, + 750 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a noise-robust kernel estimation function along with its optimal conditions, enabling the analytical acquisition of an ideal blur kernel when provided with a clear image and a blurry image.", + "2. We design a multiple kernel estimation scheme to address a wide range of unknown noise levels." + ], + "bbox": [ + 222, + 763, + 785, + 838 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C.Lee et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We adopt DIP to leverage the natural image prior captured by the handcrafted neural network for the generation of both a clear image and intermediate images which are used for estimating blur kernels.", + "bbox": [ + 222, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. The experimental results and detailed analysis with simulation and experimental data consistently demonstrate that the proposed method outperforms in blind deblurring tasks, particularly in scenarios where strong noise exists.", + "bbox": [ + 222, + 196, + 784, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 287, + 390, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAP-based method The Blind deblurring problem has been actively investigated for decades as a blurred image can be explained as many pairs of clear images and blur kernels. One of the prevalent approaches involves addressing the problem within the MAP framework consisting of priors of clear image and blur kernel. For example, Krishnan et al. [18] proposed the normalized sparsity regularization $l_{1} / l_{2}$ to maximize the prior of clear image. Pan et al. [30] adopted the dark channel on the premise that the dark channel of a clean image exhibits much sparsity compared to the blurred image. Other sparsity constraints such as TV-regularization [31], $l_{0}$ norm [29], $l_{p}$ norm [15,44] have been proposed to regularize the solution space of a clear image. Also, there have been some prior works to regularize the blur kernel estimation such as sparsity priors [23, 44]. Also, equality constraint $\\sum h = 1$ and non-negativity $h \\geq 0$ are imposed to blur kernel estimation for the stability of convergence.", + "bbox": [ + 212, + 340, + 787, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, DIP has been incorporated into the MAP framework for blind deblurring problems. Specifically, the inherent capacity of a well-designed neural network to capture image statistics is leveraged by substituting the regularization function with the neural network itself. For example, Ren et al. [33] adopts double-DIP [11] to infer the latent information of clear image and blur kernel. The elaborately designed generators have shown state-of-the-art blind deblurring performance without requiring explicit knowledge of clear image and blur kernel. Bredell et al. [3] incorporates the Wiener filter as a guide during the training of the DIP, effectively improving deblurring performance.", + "bbox": [ + 212, + 537, + 787, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Noise-blind deblurring The deblurring problem becomes more challenging when the noise level is not accurately known. One expected approach is to estimate the noise level first and then apply state-of-the-art deblurring methods. However, the estimation of the noise level in blurry images has not yet been reported, making it difficult to apply. To remedy this problem, methods such as adopting Bayesian estimator [1,14], variational expected maximization-based method [28], and other few training-based methods [9,19,35] have been proposed and shown state-of-the-art deblurring performances.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Blind image deblurring with noise-robust kernel estimation", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 330, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Wiener deconvolution method", + "text_level": 1, + "bbox": [ + 215, + 174, + 511, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core idea of the Wiener deconvolution method is to find an inverse filter $w_{h\\rightarrow x}$ that can estimate true signal $\\hat{x}$ . The optimal inverse filter can be derived from error minimization of the difference between the inverse filtered image $\\hat{x}$ and target image $x$ : $\\epsilon = \\mathbb{E}[|X - \\hat{X}|^2]$ , where $\\mathbb{E}$ denotes expectation, $X$ and $\\hat{X}$ is Fourier transform of $x$ and $\\hat{x}$ , respectively. The estimated true signal can be acquired from the following equation:", + "bbox": [ + 212, + 198, + 787, + 289 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} = \\frac {\\bar {H}}{| H | ^ {2} + \\eta} Y _ {m} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 297, + 785, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": ", where $\\bar{\\cdot}$ denotes complex conjugation and $\\eta$ is the ratio between the mean power spectral density (mPSD) of noise and signal. As Eq. (2) is simple but effective in reducing noise and deblur the degraded image, many of previous works [3, 10, 24, 42] employed Wiener filter to their deblurring framework. Inspired by this, we derive the following kernel estimation function.", + "bbox": [ + 212, + 337, + 787, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Noise-robust kernel estimation function", + "text_level": 1, + "bbox": [ + 215, + 449, + 588, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Claim 1. Given blurry image $y_{m}$ and clean image $x$ , the blur kernel $\\widehat{h}$ can be estimated by $\\widehat{H} = \\frac{\\overline{X}}{|X|^2 + k} Y_m$ , where $\\widehat{H}, Y_m,$ and $X$ are Fourier transform of $\\widehat{h}, y_m,$ and $x$ , respectively, and $k$ is the ratio between the mean power spectral density of noise and blur kernel.", + "bbox": [ + 215, + 470, + 787, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. We formulated the following mean squared error to derive a kernel estimation function using an inverse filter $w_{x \\to h}$ .", + "bbox": [ + 215, + 549, + 787, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon = \\mathbb {E} [ | H - \\widehat {H} | ^ {2} ] = \\mathbb {E} [ | H - \\widehat {W} [ H X + N _ {g} ] | ^ {2} ] = \\mathbb {E} [ | [ 1 - \\widehat {W} X ] H - \\widehat {W} N _ {g} | ^ {2} ] \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 588, + 785, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the subscript of an inverse filter $w$ is dropped for the sake of simplicity. As we assume that the noise $n$ has zero mean and independent to blur kernel $k$ , Eq. (3) can be simplified as", + "bbox": [ + 215, + 614, + 787, + 661 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon = (1 - \\widehat {W} X) \\overline {{(1 - \\widehat {W} X)}} S _ {H} + \\widehat {W} \\overline {{\\widehat {W}}} S _ {N _ {g}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 672, + 785, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": ", where $S_{H} = \\mathbb{E}[|H|^{2}]$ and $S_{N_g} = \\mathbb{E}[|N_g|^2]$ are the mPSD of blur kernel and noise, respectively, and $\\bar{\\cdot}$ denotes complex conjugation. Eq. (4) can be solved by computing the derivative of $\\epsilon$ with respect to $\\widehat{W}$ to become zero. Now, we can calculate the $\\widehat{H}$ from $X$ and $Y_{m}$ as follows:", + "bbox": [ + 212, + 698, + 787, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {H} = \\widehat {W} Y _ {m} = \\frac {\\overline {{X}}}{| X | ^ {2} + k} Y _ {m} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 773, + 785, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": ", where noise suppression factor $k$ is the ratio between the mPSD of noise and blur kernel. The final $\\hat{h}$ can be acquired by applying an inverse Fourier transform", + "bbox": [ + 215, + 809, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C.Lee et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6bb09889a2e64afe4090400c1031ab64d604c77d0f9250035ecd1c624194f7b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 529, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/660d90fedd022a900149f760ac2f2508de8440a8ab91e3913acecb60f6aab7e5.jpg", + "image_caption": [ + "(b)", + "Fig. 1: Investigation of the working behavior of the proposed kernel estimation function. (a) Visualization of the estimated blur kernel from the Eq. (6), with $k = 0.1\\sigma^2$ , $\\sigma^2$ , $10\\sigma^2$ , and $100\\sigma^2$ , (b) Estimation error of blur kernel, and visual comparison between the estimated kernels under the optimal condition (solid line) and with those having minimum error values (dashed line)." + ], + "image_footnote": [], + "bbox": [ + 534, + 155, + 784, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is followed by absolute and cropping operation, $C:\\mathbb{R}^N\\to \\mathbb{R}^M$ to Eq. (5). The resultant PSF can be described as follows:", + "bbox": [ + 214, + 391, + 782, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh = C \\left(\\mathcal {W} \\left(y _ {m}, x, k\\right)\\right) = F \\left(y _ {m}, x, k\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 431, + 784, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": ", where $\\mathcal{W}$ represents inverse filtering of $y_{m}$ with $w$ . Although Eq. (6) is a closed-form solution, the uncertainty of mPSD of noise and blur kernel prevents computation. In the case of the Eq. (2), noise-to-signal ratio $\\eta$ should be carefully selected for successful reconstruction [8, 12]. Similarly, optimal blur kernel estimation requires the appropriate noise suppression factor $k$ . The optimal value for $k$ is derived and analyzed in the following.", + "bbox": [ + 212, + 455, + 787, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Claim 2. Given Gaussian noise $\\sim \\mathcal{N}(0, \\sigma^2)$ , the optimal noise suppression factor is given as $k \\simeq A(\\sigma)^2$ , where $A$ represents the smoothness of blur kernel.", + "bbox": [ + 214, + 582, + 787, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As we have discussed, the noise suppression factor in Eq. (5) is the ratio between mPSD of noise and blur kernel which is analytically derived. According to the Wiener-Khintchine theorem, the power spectral density of a random process can be expressed as the Fourier transform of auto-correlation [20,21]. In the case of Gaussian noise $\\sim \\mathcal{N}(0,\\sigma^2)$ , the auto-correlation is defined as $R_{nn}(\\tau) = \\sigma^2\\delta(\\tau)$ so its Fourier transform, i.e. mPSD, is: $S_{N_g} = \\sigma^2$ . Although the mPSD of the blur kernel rarely is known, it represents the smoothness of the blur kernel. Here, we empirically found that $A\\simeq 1$ for sparse blur kernel such as motion blur while $A\\simeq 100$ for smooth blur kernel such as Gaussian blur.", + "bbox": [ + 212, + 628, + 787, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To see how the noise suppression factor affects the reconstruction quality of the blur kernel, we reformulate the Eq. (5) as follows:", + "bbox": [ + 214, + 763, + 787, + 795 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {H} = \\frac {\\overline {{X}}}{| X | ^ {2} + k} (H X + N _ {g}) = H + H \\left(\\frac {- k + \\overline {{X}} N _ {g} / H}{| X | ^ {2} + k}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 803, + 785, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Blind image deblurring with noise-robust kernel estimation", + "bbox": [ + 338, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6fc89e2f65f6225ca74f7a4aabdd4f1455eabbe479055e3383163ef9320fe2c5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 509, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0dc747ae8ab989fd33035b51fcc46e1c4303c24aaac8cfdc881e230805bc57cd.jpg", + "image_caption": [ + "Fig. 2: (a) Illustration of the mean squared error between the measured blurry image $y_{m}$ and the estimated blurry image $y_{est}$ . (b) Visual comparison of the $y_{m}$ and $y_{m} - y_{est}$ for different $k$ . The inset shows the estimated blur kernel." + ], + "image_footnote": [], + "bbox": [ + 519, + 146, + 651, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9fb8f4e282714bb52c0d9acf4d7587594f06b6cd631e107a68a71577f0e58a2.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 651, + 147, + 782, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 1 (a) shows the reconstructed blur kernel for different noise suppression factors. When $k = \\sigma^2$ , the proposed kernel estimation function accurately reconstructs the kernel even in strong noise. However, when $k$ is small compared to the noise level, i.e. $k \\ll \\frac{\\overline{X}N_g}{H}$ , Eq. (7) can be expressed as $\\widehat{H} = H + \\frac{N_g}{X}$ . As shown in Fig. 1 (a) where $k = 0.1\\sigma^2$ , the Fourier spectrum of noise normalized by the spectrum of a clear image, i.e. $\\frac{N_g}{X}$ , is added to the blur kernel which leads to severe degradation of the reconstructed blur kernel. On the other hand, when $k$ is too large, i.e. $k \\gg \\frac{\\overline{X}N_g}{H}$ , Eq. (7) can be expressed as $\\widehat{H} = H(1 - \\frac{k}{|X|^2 + k})$ . As shown in Fig. 1 (a) where $k = 10\\sigma^2$ and $k = 100\\sigma^2$ , since the high-frequency spectrum of $H$ becomes zero due to the $\\frac{k}{|X|^2 + k} \\approx 1$ , the blur kernel exhibits as a low-pass filtered. Also, Fig. 1 (b) demonstrates that the proposed optimal condition closely approximates the minimization of kernel estimation error. The further analysis for $F$ is described in Supplementary S1.", + "bbox": [ + 212, + 391, + 789, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Multiple kernel estimation for unknown noise level", + "text_level": 1, + "bbox": [ + 214, + 628, + 681, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A classical strategy to solve inverse problems including blind deblurring is minimizing penalized least square (PLS) methods. The data-fidelity term in PLS formulation of the blind deblurring problem is: $\\mathrm{argmin}_{x,h}\\| y_m - x*h\\| _2^2$ . In our cases, as $h$ can be estimated through $F$ , the data-fidelity term can be expressed as $\\mathrm{argmin}_x\\| y_m - x*F(y_m,x,k)\\| _2^2$ , which requires setting of $k$ for optimization. However, because we cannot accurately determine the noise level, setting the optimal $k$ is extremely challenging. Nevertheless, since $F$ is a closed-form inverse mapping function, which minimizes Eq. (3), deviating from the optimal condition of $k$ results in increased noise or blurriness rather than leading to incorrect solutions. In other words, even if the $k$ value is not strictly set, as long as it falls within a certain boundary, the data-fidelity term can be minimized. We verified this by examining the mean squared error (MSE) between $y_{m}$ and", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C.Lee et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d42bcfe8ba0906ea41c9a6da6b2048be1f3bfee9e19c82ca1ca6586cee64dbc.jpg", + "image_caption": [ + "(a) Blur" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 310, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c30511b65de6bed8ecba639631a5f784b4a2b2b2535c11dd22c2073bbdcc678.jpg", + "image_caption": [ + "(b) GT", + "Fig. 3: Visualization of the blind deblurring results using the multiple kernel estimation scheme. (a) Blur image degraded with Gaussian noise level of $\\sigma = 20$ , (b) ground truth, (c) reconstructed clear image, and estimated intermediate image and blur kernels with (d) $k = 5^2$ , (e) $k = 15^2$ , and (f) $k = 25^2$ ." + ], + "image_footnote": [], + "bbox": [ + 313, + 143, + 406, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be1c66ed0e06eb4379033d92495a8377b3849677013cb888052a5163d061d748.jpg", + "image_caption": [ + "(c) $x_0$" + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 500, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ba296c635461c04756535dbb7c37414be49ab62f0b1ac4bc96486a903010545.jpg", + "image_caption": [ + "(d) $x_{1}$ and $h_1$" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 596, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edd4a63306f7f6565a0c46316bd60858c6f1e508cc4d56eba9b2f987f866c530.jpg", + "image_caption": [ + "(e) $x_{2}$ and $h_2$" + ], + "image_footnote": [], + "bbox": [ + 598, + 143, + 689, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67c8c496a96f29d32c1231f254587aa912621414330e5f49c51644bf34910e42.jpg", + "image_caption": [ + "(f) $x_{3}$ and $h_3$" + ], + "image_footnote": [], + "bbox": [ + 694, + 143, + 785, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$y_{est} = x * F(y_m, x, k)$ by changing $k$ . As shown in Fig. 2 (a), the MSE decreases as $k$ approaches the optimal condition. Also, the lower error value is maintained over a large range of $k$ , and there exists a range around the optimal suppression factor where $|y_m - x * F(y_m, x, k)| \\leq \\varepsilon$ even if $h$ is not optimally estimated (Fig. 2 (a) red box). Fig. 2 (b) further demonstrates this by visually comparing $y_{est}$ , $F(y_m, x, k)$ , and differential map $y_m - y_{est}$ for different $k$ values when noise level is $\\sigma = 10$ . Based on this observation, we propose a multiple kernel estimation scheme that handles a wide range of unknown noise levels by setting multiple values for $k$ , rather than a specific value as follows:", + "bbox": [ + 214, + 303, + 784, + 438 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} = F \\left(y _ {m}, x _ {i}, k _ {i}\\right), \\quad \\forall i = 1, \\dots , N _ {k} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 450, + 784, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": ", where $h_i$ is the estimated blur kernel given intermediate image $x_i$ and $k_i$ is noise suppression factors corresponding to different noise levels. The fundamental concept of this framework lies in managing unknown noise levels by setting $k_i$ from low to high, thus covering a broad range of noise that exists in nature. Here, we also introduce intermediate images, which are separately estimated from a clear image, to compensate for the error term in Eq. (7) caused by a suboptimal setting of $k$ , thereby stabilizing the optimization process. Fig. 3 shows an example of deblurred image $x_0$ , estimated intermediate images $x_i$ and blur kernels $h_i$ . A detailed analysis for intermediate images and the performance of multiple kernel estimation scheme, compared to case where noise level is precisely known, are presented in Supplementary S2. We demonstrate that the proposed multiple kernel estimation scheme exhibits deblurring performance comparable to cases where the noise level is precisely known, and see for the detailed analysis.", + "bbox": [ + 214, + 479, + 787, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 MAP approach for blind deblurring", + "text_level": 1, + "bbox": [ + 214, + 698, + 555, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conventionally, the MAP approach for blind deblurring is described as follows:", + "bbox": [ + 214, + 724, + 779, + 739 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np (x, h | y) \\propto p (y | x, h) p (x) p (h) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 752, + 784, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": ", where $p(y|x,h)$ , $p(x)$ , and $p(h)$ correspond to data-fidelity term, clear image prior, and blur kernel prior, respectively. Many previous works carefully designed the priors to strictly confine the feasible solution space.", + "bbox": [ + 212, + 779, + 784, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we show that leveraging the capabilities of DIP allows solving", + "bbox": [ + 238, + 824, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Blind image deblurring with noise-robust kernel estimation", + "bbox": [ + 338, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/292ba58fcfe7b1722cb91025744194da9cbf7f0bc7015cfd4686c93fed563122.jpg", + "image_caption": [ + "Fig. 4: Workflow of the proposed method. A single deep neural network generates the deblurred image $x_0$ , and the intermediate images $x_i$ which are used to estimate blur kernel $h_i$ with the noise suppression factor $k_i$ ." + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 785, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the blind deblurring problem by estimating solely a clear image even in the presence of noise. Specifically, as blur kernel can be analytically obtained from Eq. (6) if clear image $x$ and blurry image $y_{m}$ are given, Eq. (9) can be reformulated as follows:", + "bbox": [ + 212, + 353, + 782, + 411 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np (x, h | y) = p (h | x, y) p (x | y) \\propto p (y | x) p (x) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 414, + 785, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Eq. (10), it is evident that the blind deblurring problem can be addressed by solely focusing on estimating a clear image, while also effectively handling noise owing to the robustness of the proposed kernel estimation function. Therefore, we aim to utilize DIP, one of the optimal methods for estimating natural images.", + "bbox": [ + 212, + 434, + 787, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Blind deblurring based on noise-robust kernel estimation", + "text_level": 1, + "bbox": [ + 212, + 542, + 730, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We formulate the noise-robust blind deblurring method as follows:", + "bbox": [ + 212, + 564, + 692, + 579 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {x _ {0}, h _ {1}, \\dots , h _ {N _ {k}}} \\frac {1}{N _ {k}} \\sum_ {i = 1} ^ {N _ {k}} | | y _ {m} - x _ {0} * h _ {i} | | _ {2} ^ {2} + \\lambda R (x _ {0})\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 585, + 638, + 625 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ns. t. x _ {0} = g _ {\\phi_ {0}} \\left(G _ {\\theta} (z)\\right), h _ {i} = F \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 626, + 784, + 645 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} ^ {j} \\geq 0, \\sum_ {j} h _ {i} ^ {j} = 1 \\quad \\forall j = 0, 1, \\dots , M\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 648, + 588, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": ", where $z$ is fixed tensor randomly sampled from uniform distribution, $(\\cdot)^j$ denotes $j$ -th elements, $g_{\\phi_i}$ is $i$ -th group convolution filter convolved with the last feature layer of $G_{\\theta}$ to generate $x_{i}$ , $R(\\cdot)$ is regularization function for $x_0$ , and $\\lambda$ is regularization weight. Although we adopt total variation for $R(\\cdot)$ , other regularization functions such as $l_{0}$ -sparsity or $l_{1} / l_{2}$ norm can be used.", + "bbox": [ + 212, + 688, + 785, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Eq. (11), the blind deblurring problem can be solved by focusing on estimating $x$ . Also, it should be noted that the estimation of $x_0$ and $x_i$ can be achieved with a single neural network as $x_i$ should have a similar appearance with $x_0$ .", + "bbox": [ + 212, + 763, + 785, + 824 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C.Lee et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Blind deblurring with the noise-robust kernel estimation function" + ], + "code_body": "1: Input: Blurred image $y_{m}$ , pre-defined $\\epsilon, \\lambda_{1}, \\lambda_{2}, k_{1}, k_{2}, \\dots, k_{N_{k}}$ \n2: Output: clear image $x_{0}$ and estimated blur kernels $h_{1}, h_{2}, \\dots, h_{N_{k}}$ \n3: Sample $z$ from uniform distribution \n4: for $t = 1$ to $N_{iter}$ do \n5: $[f_{0}, f_{1}, \\dots, f_{N_{k}}] = G_{\\theta}(z)$ \n6: $x_{0} = g_{\\phi_{0}}(f_{0})$ \n7: $\\mathcal{L}_{tot} = 0$ \n8: for $i = 1$ to $N_{k}$ do \n9: $h_{i} = F(y_{m}, g_{\\phi_{i}}(f_{i}), k_{i})$ \n10: $h_{i} = |h_{i}|$ Non-negativity \n11: $h_{i} = h_{i} / (\\sum h_{i} + \\epsilon)$ Equiltity \n12: $\\mathcal{L}_{tot} = \\mathcal{L}_{tot} + \\frac{1}{N_{k}} \\left( ||y_{m} - x_{0} * h_{i}||_{2}^{2} + \\lambda_{1} ||W(y_{m}, g_{\\phi_{i}}(f_{i}), k_{i})||_{2}^{2} \\right)$ \n13: end for \n14: $\\mathcal{L}_{tot} = \\mathcal{L}_{tot} + \\lambda_{2} TV(x_{0})$ \n15: Update $\\{\\theta, \\phi\\}$ using $\\nabla_{\\theta, \\phi} \\mathcal{L}_{tot}$ \n16: end for \n17: return $x_{0}, h_{1}, h_{2}, \\dots, h_{N_{k}}$", + "bbox": [ + 218, + 164, + 785, + 406 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.6 Optimization of the proposed method", + "text_level": 1, + "bbox": [ + 215, + 446, + 571, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The workflow of the proposed method is illustrated in Fig. 4 and the network parameters $\\theta, \\Phi = \\{\\phi_0, \\dots, \\phi_{N_k}\\}$ are optimized by the following cost function:", + "bbox": [ + 214, + 497, + 784, + 529 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta , \\phi_ {0}, \\dots \\phi_ {N _ {k}}} \\frac {1}{N _ {k}} \\sum_ {i = 1} ^ {N _ {k}} \\left(\\left| \\left| y _ {m} - g _ {\\phi_ {0}} \\left(G _ {\\theta} (z)\\right) * F \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) \\right| \\right| _ {2} ^ {2} \\right. \\\\ + \\lambda_ {1} \\| \\mathcal {W} \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) \\| _ {2} ^ {2}) + \\lambda_ {2} T V \\left(g _ {\\phi_ {0}} \\left(G _ {\\theta} (z)\\right)\\right) \\tag {12} \\\\ s. t. F \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) ^ {J} \\geq 0, \\\\ \\sum_ {j} F (y _ {m}, g _ {\\phi_ {i}} (G _ {\\theta} (z)), k _ {i}) ^ {j} = 1 \\quad \\forall j = 0, 1, \\dots , M \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 551, + 784, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": ", where $TV$ is total variation, and $\\lambda_{1}$ and $\\lambda_{2}$ are regularization weights. The nonnegativity and equality constraints to the estimated blur kernel are induced by taking an absolute operation followed by dividing each element by the sum of all elements. Also, we additionally introduce a sparsity regularization term for the inverse filtered image, $\\mathcal{W}(y_m, g_{\\phi_i}(G_\\theta(z)), k_i)$ , as all pixels except for the blur kernel positioned should have a value of 0. It is important to emphasize that our proposed multiple kernel estimation scheme, rather than regularization terms, is the key enabler for robust deblurring (see the ablation study in Supplementary S7). The optimization of Eq. (12) is formally summarized in Algorithm 1.", + "bbox": [ + 212, + 689, + 785, + 825 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Blind image deblurring with noise-robust kernel estimation", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/13de169d7d55466941e68403acc4b420e99f8c39dffd00ed33a7e3748cb09b30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 310, + 215 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/88385f9821f2b47f409e11b8229b4758bf63cc6d0edcc20011064e35db71c31d.jpg", + "image_caption": [ + "(a) Blur" + ], + "image_footnote": [], + "bbox": [ + 217, + 215, + 310, + 286 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/180ab968c1be7815ea047dcfce4853e8cd56b45773b9fd785aac7dfaadcfecbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 143, + 406, + 215 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/83def309de633c2f0e302d10fa3f3fe04f7998ddbb4e533c1f2ae4ed796363ef.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 313, + 215, + 406, + 286 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/99c7b6449e71b28cbfabda034f4fc405b77816b39faef4dd85044b436ebb790c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 143, + 500, + 215 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3f298cdff34af1aa35e81681c2da16938bea4db79c346e4d9257dbb4a7475fbe.jpg", + "image_caption": [ + "(c) $l_{1} / l_{2}$" + ], + "image_footnote": [], + "bbox": [ + 408, + 215, + 500, + 286 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/346d94c32be2260eebcbc040721c37d490b172cbcf48853b63348fe2d77a25e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 596, + 215 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6ff3dce2b1d5c01ac5d88320064970b025abc774fc5e91410c10f5cc34430b79.jpg", + "image_caption": [ + "(d) Pure-let", + "Fig. 5: Qualitative comparison of blind deblurring results on AFHQ datasets [6]. (row 1): AFHQ-dog motion deblurring with a kernel size of $43 \\times 43$ and noise level of $\\sigma = 10$ , (row 2): AFHQ-cat motion deblurring with a kernel size of $39 \\times 39$ and noise level of $\\sigma = 20$ . Here, the estimated kernel for $k_{2} = 15^{2}$ is presented in (f) while others are presented in Supplementary S5." + ], + "image_footnote": [], + "bbox": [ + 501, + 215, + 596, + 286 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1345ee935c3bc39125403dfbefc00df9d4497c141548aaf2c8181063c96befd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 143, + 689, + 215 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b98692ca6737de00134b898ec4498bbf3db969247ddb874f0721680952063187.jpg", + "image_caption": [ + "(e)WDIP" + ], + "image_footnote": [], + "bbox": [ + 598, + 215, + 689, + 286 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/26c42b85b1e385e20935519c305803366a4c244debfb8e9db382d277b802f6ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 143, + 785, + 215 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f421912ad5d077e38ac54e32774b4721262fa0edc3dd37a79bfb2f11b68f8a5a.jpg", + "image_caption": [ + "(f) Ours" + ], + "image_footnote": [], + "bbox": [ + 694, + 215, + 785, + 286 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 409, + 374, + 425 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experiments on simulated image", + "text_level": 1, + "bbox": [ + 215, + 439, + 532, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Motion deblur on AFHQ dataset [6]. To demonstrate the deblurring performance on motion blur, we conduct experiments on AFHQ-dog $256 \\times 256$ and AFHQ-cat $256 \\times 256$ datasets. Specifically, we select 500 test images as an evaluation dataset. To synthesize motion-blurred images, motion-blur kernels of varying sizes, ranging from $21 \\times 21$ to $51 \\times 51$ , are randomly generated [7], and each kernel is then convolved with a clear image. Finally, Gaussian noise $\\sigma = 10$ and $\\sigma = 20$ are added to the synthesized blurry images.", + "bbox": [ + 212, + 460, + 787, + 568 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The proposed method (refer Supplementary S3 and S6 for implementation detail) is compared with other state-of-the-art blind deblurring methods including Sanghvi et al. [35], Xu et al. [40], Pure-let [24], Selfdeblur [33], WDIP [3], $l_{1} / l_{2}$ [18], Pan-DCP [30], Pan- $l_{0}$ [29], and Liang et al. [4], and see Supplementary S4 for implementation details. Note that we exclusively choose blind deblurring methods that do not utilize any training dataset to ensure a fair comparison. When optimizing the proposed method, $k_{1} = 5^{2}$ , $k_{2} = 15^{2}$ , and $k_{3} = 25^{2}$ are set to accommodate a wide range of unknown noise levels. The representative blind deblurring results from $l_{1} / l_{2}$ , Pure-let, WDIP, and the proposed method are qualitatively compared in Fig. 5. Due to the severe degradation caused by Gaussian noise in blurred images, the deblurring results obtained from the compared methods are greatly compromised. Even with laborious adjustment of hyperparameters, these methods tend to overfit strongly to the noise. In contrast, the deblurring results from the proposed method shown in Fig. 5 (f) demonstrate noise-robust deblurring by effectively reconstructing clear images and blur kernels simultaneously even in the absence of precise knowledge about the noise level. Furthermore, quantitative comparison presented in Tab. 1 indicates that the proposed method achieves state-of-the-art performance in terms of Peak", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C.Lee et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/85516dfeb3245e4e450a4a6beda993752d70c825abf4462db3c9f0880372331d.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of blind deblurring results on AFHQ datasets [6] with different noise levels ( $\\sigma = 10$ and $\\sigma = 20$ ). **Bold:** best." + ], + "table_footnote": [], + "table_body": "
MethodAFHQ-dogAFHQ-cat
σ = 10σ = 20σ = 10σ = 20
PSNR↑ SSIM↑ PSNR↑ SSIM↑ PSNR↑ SSIM↑ PSNR↑ SSIM
Proposed24.230.63222.900.53923.330.59022.110.497
SelfDeblur [33]13.540.10712.020.06312.830.08011.440.046
WDIP [3]14.110.11512.430.06613.430.08611.900.049
l1/l2 [18]19.690.33516.030.14619.360.30815.920.140
Pan-DCP [30]10.240.0607.6980.02410.170.0537.7130.022
Pan-l0 [29]8.3110.0316.8360.0168.2620.0296.9220.016
Liang et al. [4]10.680.0617.7110.02310.480.0577.6940.022
Sanghvi et al. [35]16.840.52016.840.52216.260.46516.380.471
Xu et al. [40]16.510.39815.010.29515.940.35714.340.258
Pure-let [24]21.020.53419.940.47720.780.48219.790.429
", + "bbox": [ + 217, + 184, + 785, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Signal-to-noise ratio (PSNR) and Structural Similarity Index Measure (SSIM). See Supplementary S5 for additional qualitative comparisons of deblurring results and quantitative comparisons at low-noise levels ( $\\sigma = 2$ and $\\sigma = 5$ ).", + "bbox": [ + 212, + 422, + 782, + 468 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sun et al. dataset [38]. To further showcase the capability of the proposed method, we utilize the Sun et al. dataset, comprising 80 natural images and 8 blur kernels, resulting in the generation of 640 blurry images. The representative blind deblurring results are presented in Fig. 6. The deblurring results obtained from compared methods - $l_{1} / l_{2}$ , Pan-DCP, and WDIP - are catastrophically corrupted when noise increases. Conversely, the proposed method shows superior deblurring results regardless of the noise level. Also, as shown in Tab. 2, the proposed method establishes the state-of-the-art deblurring performance.", + "bbox": [ + 212, + 489, + 784, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Experiments on scanning electron microscope", + "text_level": 1, + "bbox": [ + 215, + 633, + 629, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Scanning electron microscopes (SEM) provide sub-nanometer resolution images by detecting the back-scattered or emitted electrons from the sample. However, the incident electrons, which determine the shape of the PSF of SEM can be easily distorted when the electron passes through several lenses and coils, see Fig. 7 (a). Additionally, random fluctuations of the incident electrons cause Gaussian noise, significantly compromising image quality, see Fig. 8 (a). Correcting such degradation often requires manual adjustment of system parameters, such as working distance or stigmator, by experts. To automate this correction process image quality assessment based method [17,25,26], bayesian optimization based approach [2,34], and feedback-loop based deep learning approach [22,27,36] have been proposed. Nevertheless, previous methods necessitate iterative measurement of SEM images to determine correction parameters, rendering the process", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Blind image deblurring with noise-robust kernel estimation", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b417bb399ab85efeaf6b76735f329ba2e7fdc38e425e15a5485ecf7e4ebca35a.jpg", + "image_caption": [ + "(a) Blur" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 310, + 296 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/15c876766e8aa1963570ead27b4dad9613941ef3c8db283ad3a3813aa714a668.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 313, + 145, + 406, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7231a7f23e9842ca1fc11955b4e1acdefeb5697a83861bda5cb9cea164059eea.jpg", + "image_caption": [ + "(c) $l_{1} / l_{2}$" + ], + "image_footnote": [], + "bbox": [ + 408, + 145, + 500, + 296 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1c1472252408fed9d82c9d538ec2db7940835833975737b9ac9204263ae430cb.jpg", + "image_caption": [ + "(d) Pan-DCP", + "Fig.6: Qualitative comparison of blind deblurring results on Sun et al. datasets [38] with Gaussian noise level of (row 1): $\\sigma = 10$ and (row 2): $\\sigma = 20$ . Here, the estimated kernel for $k_{2} = 15^{2}$ is presented in (f) while others are presented in Supplementary S5." + ], + "image_footnote": [], + "bbox": [ + 501, + 145, + 594, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4230ca7878eb575dfe41b75ce757992611f7a97bfbf5f90369b3a2625e8329f3.jpg", + "image_caption": [ + "(e)WDIP" + ], + "image_footnote": [], + "bbox": [ + 596, + 145, + 689, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/54bd4a5fc665c271a8d67344cda2c077413179f941ac81f3a2298f067b1bcafe.jpg", + "image_caption": [ + "(f) Ours" + ], + "image_footnote": [], + "bbox": [ + 694, + 145, + 785, + 297 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/28e55c8e9e4dbe248f80731cd3d185a02d6ec6032cd22f54941fa46ef955c015.jpg", + "table_caption": [ + "Table 2: Quantitative comparison of blind deblurring results on Sun et al. datasets [38] with different noise levels ( $\\sigma = 10$ and $\\sigma = 20$ ). Bold: best." + ], + "table_footnote": [], + "table_body": "
Methodσ = 10σ = 20
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
Proposed25.940.67724.830.608
SelfDeblur [33]17.120.20215.440.136
WDIP [3]17.600.21412.890.103
l1/l2 [18]18.870.17314.970.073
Pan-DCP [30]11.620.0598.3210.021
Pan-l0 [29]9.5900.0308.0650.017
Liang et al. [4]11.960.0608.5220.022
Xu et al. [40]14.100.20713.450.128
Pure-let [24]18.870.25417.310.121
", + "bbox": [ + 334, + 421, + 661, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "time-consuming and energy-inefficient.", + "bbox": [ + 215, + 627, + 491, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Here, we demonstrate the effectiveness of the proposed blind deblurring method in correcting degraded SEM images. In general, well-focused PSF of SEM can be approximated to Gaussian shape [32, 43], we set $A = 100$ . Additionally, since the exact noise level of SEM images is rarely known, we set the noise suppression factor in a broad range, such as $k_{1} = 10000$ , $k_{2} = 20000$ , and $k_{3} = 30000$ . Fig. 7 and Fig. 8 presents visual comparisons of blind deblurring results. Although methods such as Pan-DCP or WDIP have successfully performed deblurring on images with weak noise and simple patterns, their performance noticeably deteriorates in cases of strong noise or complex patterns. In contrast, as shown in Fig. 7 (f) and Fig. 8 (f), the proposed method consistently reconstructs both the clear image and the PSF, regardless of the complexity of the image or the level of noise. Especially, as shown in the second row of Fig. 7, when", + "bbox": [ + 212, + 657, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C.Lee et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a8a32ea0ed20d00c1edbb9bf61b87552c778ee5a8c3858d5d82562750d4b858e.jpg", + "image_caption": [ + "(a) Blur" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 312, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f3eabd133a138ac205cb3f505c51e05379a239d028653354bf142678e97b7a7b.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 313, + 143, + 406, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a284f85bb5c52af4d3d1f556318e7a030b72dd9f1e2057610ca66cec4bf91c2c.jpg", + "image_caption": [ + "(c) $l_{1} / l_{2}$" + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 500, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f55ea2035b67b5d3467f2e5996b71ef014dc4b497740d3d5738416be2451e694.jpg", + "image_caption": [ + "(d) Pan-DCP" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 596, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a491bd3e03431584042738a7009bf306a40eee73c1a7a7a8617628912efe0972.jpg", + "image_caption": [ + "(e) WDIP" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 689, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/84a5ded5fd38026e56440933687be20df2fe1c10714d45a1d58cddfbf1c2618f.jpg", + "image_caption": [ + "(f) Ours" + ], + "image_footnote": [], + "bbox": [ + 692, + 143, + 785, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/59ddd69a082b932c61ee4729dbbc66e35c4a6535f3d6d30e62c24fb24b1f2500.jpg", + "image_caption": [ + "Fig. 7: Qualitative comparison of blind deblurring results on SEM image of semiconductor. (row 1): Circular ingredient of a semiconductor with the image size of $512 \\times 512$ , (row 2): Wire structure of semiconductor with the image size of $512 \\times 512$ . The ground truth PSFs are obtained by the Eq. (6) with $k = 10000$ .", + "(a) Blur", + "Fig. 8: Qualitative comparison of blind deblurring results on SEM image severely degraded by noise, with an image size of $256 \\times 256$ . The ground truth PSFs are obtained by the Eq. (6) with $k = 10000$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 382, + 310, + 527 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f83027979244cd36d8553bf403186df18474375e7392ff23c4a1492cdec100cf.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 313, + 382, + 406, + 527 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d68bf3db04f165df39cc526bd0f950fa0158c27e317052ef8c23a38e6640e642.jpg", + "image_caption": [ + "(c) $l_{1} / l_{2}$" + ], + "image_footnote": [], + "bbox": [ + 408, + 382, + 500, + 527 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1f13d2132b5a737c0a62f82328bfc5fd735f8db3ec5cfaa2b418dfe2c5bb4361.jpg", + "image_caption": [ + "(d) Pan-DCP" + ], + "image_footnote": [], + "bbox": [ + 501, + 382, + 596, + 527 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/75e57a03269707fc59dbd13f49c663cca7d5a3a849cc5adb9d0dd44ef066d392.jpg", + "image_caption": [ + "(e)WDIP" + ], + "image_footnote": [], + "bbox": [ + 596, + 382, + 689, + 527 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c5fe6898d6aadafdf7d597c8f0414f4e352a52085c576afa880d77a3923d2964.jpg", + "image_caption": [ + "(f) Ours" + ], + "image_footnote": [], + "bbox": [ + 692, + 382, + 785, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "examining an image with blurred wire patterns, predicting the structure can be challenging even for human observers. However, the deblurred images from the proposed method show a remarkable similarity to the ground truth images and one can effectively investigate the wire structures.", + "bbox": [ + 212, + 625, + 787, + 684 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Experiments on two-photon excitation fluorescence microscope", + "text_level": 1, + "bbox": [ + 214, + 708, + 781, + 724 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We further applied the proposed method as one of the real-world applications to a two-photon excitation fluorescence microscope (2PEF) image. In general, the imaging model of 2PEF follows Eq. (1) and its effective PSF can be expressed as the square of illumination PSF [13]. Also, due to the inherent characteristic of weak signals of fluorescence, the measured signal is often contaminated by noise. Here, we captured the fluorescence signal of $2\\mathrm{mm}$ beads from the custom-built 2PEF setup. As shown in Fig. 9 (a), the measured image from the 2PEF", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Blind image deblurring with noise-robust kernel estimation", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e7b06def865e5d6cfa51ce35bd4b2207d2ca25469b25ffe19ef0ee3058d5e036.jpg", + "image_caption": [ + "(a) Blur" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 357, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7073626d5e38656f58f5b5d45f5bd018aa37e59fe579d9c202a499e13c337cba.jpg", + "image_caption": [ + "Fig. 9: Visualization of blind deblurring results of the proposed method on blurry two-photon excitation fluorescence microscope image with an image size of $384 \\times 384$ ." + ], + "image_footnote": [], + "bbox": [ + 357, + 143, + 426, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/dad3b828ab8cfeecf50851802d675dc429011dba2f3720f4bf57136aeb07fac9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 143, + 496, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/23ecd2b9533b95656f370f792432d76e2bb282fa42ea07831e7f641fa38e5016.jpg", + "image_caption": [ + "(b) Ours" + ], + "image_footnote": [], + "bbox": [ + 506, + 143, + 645, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/05ee95bf8e5ffa926acb758183b9589edf218f9be8622082477962eba373e037.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 143, + 714, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d6b89de1eaad5ef1709da2548882248e9cd65b9b4fbdb686f6550b350df6f9ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 143, + 784, + 253 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "exhibits a blurry appearance, attributed to the diffraction limit of the illumination PSF, and is further corrupted by noise. However, when the proposed deblurring method is applied to the measured image, illustrated in Fig. 9 (b), it effectively reconstructs the sharp edges of beads and facilitates easy discernment of clustered beads.", + "bbox": [ + 212, + 335, + 784, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 435, + 367, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a noise-robust blind deblurring method combining the multiple kernel estimation scheme and the DIP. The most critical aspect of the proposed method, the noise-robust kernel estimation function, successfully restores the PSF even in situations of high noise when provided only with clear and blurry images. This function is developed to the multiple kernel estimation scheme that can be applied even in scenarios where noise level cannot be accurately predicted. Additionally, by effectively leveraging DIP, which acts as a powerful prior for natural images, we can estimate a clear image from a single well-designed neural network. The discussion of limitations of the proposed work and ablation study are presented in Supplementary S7. Our proposed method has been demonstrated to operate effectively in both simulated images and real-world applications and outperforms compared to the previous state-of-the-art methods.", + "bbox": [ + 212, + 467, + 787, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 702, + 401, + 719 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by the Samsung Research Funding and Incubation Center of Samsung Electronics grant SRFC-IT2002-03, Samsung Electronics Co., Ltd. (IO220908-02403-01), and the National Research Foundation of Korea grant funded by the Korean government (Grant Nos. NRF-2021R1A5A1032937, NRF-2021R1C1C1011307, RS-2023-00251628, and RS-2024-00397673).", + "bbox": [ + 212, + 733, + 784, + 809 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C.Lee et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arjomand Bigdeli, S., Zwicker, M., Favaro, P., Jin, M.: Deep mean-shift priors for image restoration. Advances in Neural Information Processing Systems 30 (2017)", + "2. Binding, J., Mikula, S., Denk, W.: Low-dosage maximum-a-posteriori focusing and stigmation. Microscopy and Microanalysis 19(1), 38-55 (2013)", + "3. Bredell, G., Erdil, E., Weber, B., Konukoglu, E.: Wiener guided dip for unsupervised blind image deconvolution. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3047-3056 (2023)", + "4. Chen, L., Fang, F., Lei, S., Li, F., Zhang, G.: Enhanced sparse model for blind deblurring. In: European Conference on Computer Vision. pp. 631-646. Springer (2020)", + "5. Cho, S., Lee, S.: Fast motion deblurring. In: ACM SIGGRAPH Asia 2009 papers, pp. 1-8 (2009)", + "6. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2020)", + "7. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6059-6069 (2023)", + "8. Dhawan, A.P., Rangayyan, R.M., Gordon, R.: Image restoration by wiener deconvolution in limited-view computed tomography. Applied optics 24(23), 4013-4020 (1985)", + "9. Dong, J., Pan, J., Sun, D., Su, Z., Yang, M.H.: Learning data terms for non-blind deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 748-763 (2018)", + "0. Dong, J., Roth, S., Schiele, B.: Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems 33, 1048-1059 (2020)", + "1. Gandelsman, Y., Shocher, A., Irani, M.: \"double-dip\": unsupervised image decomposition via coupled deep-image-priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11026-11035 (2019)", + "2. Gonzalez, R.C.: Digital image processing. Pearson education india (2009)", + "3. Helmchen, F., Denk, W.: Deep tissue two-photon microscopy. Nature methods 2(12), 932-940 (2005)", + "4. Jin, M., Roth, S., Favaro, P.: Noise-blind image deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3510-3518 (2017)", + "5. Jin, M., Roth, S., Favaro, P.: Normalized blind deconvolution. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 668-684 (2018)", + "6. Joshi, N., Szeliski, R., Kriegman, D.J.: Psf estimation using sharp edge prediction. In: 2008 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1-8. IEEE (2008)", + "7. Kawasaki, T., Nakano, T., Hirose, K.: Developing an aberration-corrected schottky emission sem and method for measuring aberration. Microelectronic engineering 86(4-6), 1017-1020 (2009)", + "8. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011)", + "9. Kruse, J., Rother, C., Schmidt, U.: Learning to push the limits of efficient ft-based image deconvolution. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4586-4594 (2017)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Blind image deblurring with noise-robust kernel estimation", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Kubo, R., Toda, M., Hashitsume, N.: Statistical physics II: nonequilibrium statistical mechanics, vol. 31. Springer Science & Business Media (2012)", + "21. Lampard, D.: Generalization of the wiener-khintchine theorem to nonstationary processes. Journal of Applied Physics 25(6), 802-803 (1954)", + "22. Lee, W., Nam, H.S., Kim, Y.G., Kim, Y.J., Lee, J.H., Yoo, H.: Robust autofocusing for scanning electron microscopy based on a dual deep learning network. Scientific reports 11(1), 20933 (2021)", + "23. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)", + "24. Li, J., Luisier, F., Blu, T.: Pure-let image deconvolution. IEEE Transactions on Image Processing 27(1), 92-105 (2017)", + "25. Lifshin, E., Kandel, Y.P., Moore, R.L.: Improving scanning electron microscope resolution for near planar samples through the use of image restoration. Microscopy and Microanalysis 20(1), 78-89 (2014)", + "26. Lu, Y., Zhang, X., Li, H.: A simplified focusing and astigmatism correction method for a scanning electron microscope. AIP Advances 8(1) (2018)", + "27. Na, J., Kim, G., Kang, S.H., Kim, S.J., Lee, S.: Deep learning-based discriminative refocusing of scanning electron microscopy images for materials science. Acta Materialia 214, 116987 (2021)", + "28. Nan, Y., Quan, Y., Ji, H.: Variational-em-based deep learning for noise-blind image deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3626-3635 (2020)", + "29. Pan, J., Hu, Z., Su, Z., Yang, M.H.: Deblurring text images via l0-regularized intensity and gradient prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2901-2908 (2014)", + "30. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1628-1636 (2016)", + "31. Perrone, D., Favaro, P.: Total variation blind deconvolution: The devil is in the details. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2909-2916 (2014)", + "32. Reimer, L.: Scanning electron microscopy: physics of image formation and microanalysis. Measurement Science and Technology 11(12), 1826-1826 (2000)", + "33. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3341-3350 (2020)", + "34. Roels, J., Aelterman, J., De Vylder, J., Luong, H., Saeys, Y., Philips, W.: Bayesian deconvolution of scanning electron microscopy images using point-spread function estimation and non-local regularization. In: 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). pp. 443-447. IEEE (2016)", + "35. Sanghvi, Y., Mao, Z., Chan, S.H.: Structured kernel estimation for photon-limited deconvolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9863-9872 (2023)", + "36. Schubert, P.J., Saxena, R., Kornfeld, J.: Deepfocus: Fast focus and astigmatism correction for electron microscopy. arXiv preprint arXiv:2305.04977 (2023)", + "37. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. Acm transactions on graphics (tog) 27(3), 1-10 (2008)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C.Lee et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)", + "39. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9446-9454 (2018)", + "40. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 157-170. Springer (2010)", + "41. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1107-1114 (2013)", + "42. Xue, F., Luisier, F., Blu, T.: Multi-wiener sure-let deconvolution. IEEE Transactions on Image Processing 22(5), 1954–1968 (2013)", + "43. Zotta, M.D., Nevins, M.C., Hailstone, R.K., Lifshin, E.: The determination and application of the point spread function in the scanning electron microscope. Microscopy and microanalysis 24(4), 396-405 (2018)", + "44. Zuo, W., Ren, D., Zhang, D., Gu, S., Zhang, L.: Learning iteration-wise generalized shrinkage-thresholding operators for blind deconvolution. IEEE Transactions on Image Processing 25(4), 1751-1764 (2016)" + ], + "bbox": [ + 212, + 146, + 787, + 424 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Blind image deblurring with noise-robust kernel estimation", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Blind image deblurring with noise-robust kernel estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_model.json b/2024/Blind image deblurring with noise-robust kernel estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_model.json new file mode 100644 index 0000000000000000000000000000000000000000..74e38df81d194a2c4e4737da41be4de476263f44 --- /dev/null +++ b/2024/Blind image deblurring with noise-robust kernel estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_model.json @@ -0,0 +1,2962 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.228, + 0.141, + 0.777, + 0.185 + ], + "angle": 0, + "content": "Blind image deblurring with noise-robust kernel estimation" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.213, + 0.784, + 0.259 + ], + "angle": 0, + "content": "Chanseok Lee\\(^{1}\\), Jeongsol Kim\\(^{1}\\), Seungmin Lee\\(^{1}\\), Jaehwang Jung\\(^{2}\\), Yunje Cho\\(^{2}\\), Taejoong Kim\\(^{2}\\), Taeyong Jo\\(^{2}\\), Myungjun Lee\\(^{2}\\), and Mooseok Jang\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.27, + 0.754, + 0.284 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Bio and Brain Engineering, KAIST, Daejeon, South Korea" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.284, + 0.774, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) Advanced Process Development Team 4, Semiconductor R&D Center, Samsung Electronics, South Korea" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.312, + 0.785, + 0.34 + ], + "angle": 0, + "content": "\\(^{3}\\) KAIST Institute for Health Science and Technology, KAIST, Daejeon, South Korea {cslee, mooseok}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.38, + 0.741, + 0.601 + ], + "angle": 0, + "content": "Abstract. Blind deblurring is an ill-posed inverse problem involving the retrieval of a clear image and blur kernel from a single blurry image. The challenge arises considerably when strong noise, where its level remains unknown, is introduced. Existing blind deblurring methods are highly susceptible to noise due to overfitting and disturbances in the solution space. Here, we propose a blind deblurring method based on a noise-robust kernel estimation function and deep image prior (DIP). Specifically, the proposed kernel estimation function effectively estimates the blur kernel even for strongly noisy blurry images given a clear image and optimal condition. Therefore, DIP is adopted for the generation of a clear image to leverage its natural image prior. Additionally, the multiple kernel estimation scheme is designed to address a wide range of unknown noise levels. Extensive experimental studies, including simulated images and real-world examples, demonstrate the superior deblurring performance of the proposed method. The official code is uploaded in https://github.com/csleemooo/BD_noise_robust_kernel_estimation." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.615, + 0.704, + 0.63 + ], + "angle": 0, + "content": "Keywords: Blind deblurring \\(\\cdot\\) Deep image prior \\(\\cdot\\) Gaussian noise" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.377, + 0.672 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Optical imaging systems such as brightfield microscopes, fluorescence microscopes, or scanning electron microscopes (SEM) have a crucial role in histopathology, inspection, and clinical or biological research by providing complicated structural information of target imaging samples with high-resolution images. However, perturbation or misalignment (e.g. defocus or aberrations) of the imaging system due to its instability or systemic error brings blurry effects on the captured image. Furthermore, additive Gaussian Noise severely degrades the captured image quality. Such degradation process can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.446, + 0.826, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\ny _ {m} = x * h + n _ {g} \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C.Lee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.283 + ], + "angle": 0, + "content": ", where \\( y_{m} \\in \\mathbb{R}^{N} \\) is measured blurry and noisy image, \\( x \\in \\mathbb{R}^{N} \\) is a clear image, * denotes the convolution operator, \\( h \\in \\mathbb{R}^{M} \\) is blur kernel or point spread function (PSF) of the imaging system, and \\( n_g \\) is additive noise. Retrieving clean and focused image \\( x \\) as well as blur kernel \\( h \\) from a single degraded image \\( y_{m} \\) is well known as blind deconvolution or blind image deblurring and has been actively investigated for decades due to its ill-posedness. Although many previous approaches such as edge-selection-based method [5,16,37,38] or MAP-based approach [4,15,18,23,29-31,41,44] show promising results, deblurring performances significantly drop as the noise level increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.418 + ], + "angle": 0, + "content": "In general, there is a trade-off when simultaneously trying to eliminate both the blur effect and noise. For example, when focusing on removing the blur effect, the estimated image may be severely corrupted by noise or other high-frequency artifacts. On the other hand, when focusing on noise removal, there is a tendency to overly emphasize smoothness, compromising the fine details of the image. So, balancing the strength of deblurring and noise suppression is important to successfully reconstruct \\( x \\) and \\( h \\). Moreover, in cases where the noise level cannot be accurately determined, the difficulty of the deblurring task increases significantly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.735 + ], + "angle": 0, + "content": "This paper aims to address the challenge of blind deblurring in the presence of strong and an unknown noise level. Instead of relying on sharp edge-selective maps or prior information to estimate the blur kernel, we introduce a novel kernel estimation function. This function analytically computes the blur kernel when provided with a clear image, blurry image, and noise suppression factor. The first characteristic of the proposed function is its ability to estimate the blur kernel effectively even in scenarios with strong noise. Specifically, the noise suppression factor included in the kernel estimation function is pivotal in mitigating noise, and its optimal condition is derived accordingly. To further ameliorate this characteristic, we designed a multiple kernel estimation scheme, which effectively handles a wide range of noise levels, to perform blind deblurring even in situations where the noise level is unknown. The second characteristic is that the estimation of only the clear image is sufficient for estimating a blur kernel. In other words, the amount of information required for solving a blind deblurring problem is much less than that needed by conventional methods, and we validated this through maximum a posteriori (MAP) analysis. This characteristic is fully leveraged by employing DIP, known as one of the most effective methods for capturing priors of natural images and solving inverse mapping problems [39]. The inherent robustness of DIP to noise, combined with the noise-robust multiple kernel estimation scheme, enables superior performance in blind deblurring tasks." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.736, + 0.731, + 0.751 + ], + "angle": 0, + "content": "The main contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.764, + 0.787, + 0.809 + ], + "angle": 0, + "content": "1. We propose a noise-robust kernel estimation function along with its optimal conditions, enabling the analytical acquisition of an ideal blur kernel when provided with a clear image and a blurry image." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.81, + 0.787, + 0.839 + ], + "angle": 0, + "content": "2. We design a multiple kernel estimation scheme to address a wide range of unknown noise levels." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.764, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Blind image deblurring with noise-robust kernel estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "3. We adopt DIP to leverage the natural image prior captured by the handcrafted neural network for the generation of both a clear image and intermediate images which are used for estimating blur kernels." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.197, + 0.785, + 0.242 + ], + "angle": 0, + "content": "4. The experimental results and detailed analysis with simulation and experimental data consistently demonstrate that the proposed method outperforms in blind deblurring tasks, particularly in scenarios where strong noise exists." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.288, + 0.391, + 0.304 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.538 + ], + "angle": 0, + "content": "MAP-based method The Blind deblurring problem has been actively investigated for decades as a blurred image can be explained as many pairs of clear images and blur kernels. One of the prevalent approaches involves addressing the problem within the MAP framework consisting of priors of clear image and blur kernel. For example, Krishnan et al. [18] proposed the normalized sparsity regularization \\( l_{1} / l_{2} \\) to maximize the prior of clear image. Pan et al. [30] adopted the dark channel on the premise that the dark channel of a clean image exhibits much sparsity compared to the blurred image. Other sparsity constraints such as TV-regularization [31], \\( l_{0} \\) norm [29], \\( l_{p} \\) norm [15,44] have been proposed to regularize the solution space of a clear image. Also, there have been some prior works to regularize the blur kernel estimation such as sparsity priors [23, 44]. Also, equality constraint \\( \\sum h = 1 \\) and non-negativity \\( h \\geq 0 \\) are imposed to blur kernel estimation for the stability of convergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Recently, DIP has been incorporated into the MAP framework for blind deblurring problems. Specifically, the inherent capacity of a well-designed neural network to capture image statistics is leveraged by substituting the regularization function with the neural network itself. For example, Ren et al. [33] adopts double-DIP [11] to infer the latent information of clear image and blur kernel. The elaborately designed generators have shown state-of-the-art blind deblurring performance without requiring explicit knowledge of clear image and blur kernel. Bredell et al. [3] incorporates the Wiener filter as a guide during the training of the DIP, effectively improving deblurring performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Noise-blind deblurring The deblurring problem becomes more challenging when the noise level is not accurately known. One expected approach is to estimate the noise level first and then apply state-of-the-art deblurring methods. However, the estimation of the noise level in blurry images has not yet been reported, making it difficult to apply. To remedy this problem, methods such as adopting Bayesian estimator [1,14], variational expected maximization-based method [28], and other few training-based methods [9,19,35] have been proposed and shown state-of-the-art deblurring performances." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C.Lee et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.331, + 0.161 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.175, + 0.513, + 0.19 + ], + "angle": 0, + "content": "3.1 Wiener deconvolution method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.199, + 0.788, + 0.29 + ], + "angle": 0, + "content": "The core idea of the Wiener deconvolution method is to find an inverse filter \\( w_{h\\rightarrow x} \\) that can estimate true signal \\( \\hat{x} \\). The optimal inverse filter can be derived from error minimization of the difference between the inverse filtered image \\( \\hat{x} \\) and target image \\( x \\): \\( \\epsilon = \\mathbb{E}[|X - \\hat{X}|^2] \\), where \\( \\mathbb{E} \\) denotes expectation, \\( X \\) and \\( \\hat{X} \\) is Fourier transform of \\( x \\) and \\( \\hat{x} \\), respectively. The estimated true signal can be acquired from the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.299, + 0.786, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\hat {X} = \\frac {\\bar {H}}{| H | ^ {2} + \\eta} Y _ {m} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.415 + ], + "angle": 0, + "content": ", where \\(\\bar{\\cdot}\\) denotes complex conjugation and \\(\\eta\\) is the ratio between the mean power spectral density (mPSD) of noise and signal. As Eq. (2) is simple but effective in reducing noise and deblur the degraded image, many of previous works [3, 10, 24, 42] employed Wiener filter to their deblurring framework. Inspired by this, we derive the following kernel estimation function." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.45, + 0.589, + 0.463 + ], + "angle": 0, + "content": "3.2 Noise-robust kernel estimation function" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.472, + 0.789, + 0.536 + ], + "angle": 0, + "content": "Claim 1. Given blurry image \\( y_{m} \\) and clean image \\( x \\), the blur kernel \\( \\widehat{h} \\) can be estimated by \\( \\widehat{H} = \\frac{\\overline{X}}{|X|^2 + k} Y_m \\), where \\( \\widehat{H}, Y_m, \\) and \\( X \\) are Fourier transform of \\( \\widehat{h}, y_m, \\) and \\( x \\), respectively, and \\( k \\) is the ratio between the mean power spectral density of noise and blur kernel." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.55, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Proof. We formulated the following mean squared error to derive a kernel estimation function using an inverse filter \\( w_{x \\to h} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.263, + 0.589, + 0.786, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\epsilon = \\mathbb {E} [ | H - \\widehat {H} | ^ {2} ] = \\mathbb {E} [ | H - \\widehat {W} [ H X + N _ {g} ] | ^ {2} ] = \\mathbb {E} [ | [ 1 - \\widehat {W} X ] H - \\widehat {W} N _ {g} | ^ {2} ] \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.616, + 0.788, + 0.662 + ], + "angle": 0, + "content": "Here, the subscript of an inverse filter \\( w \\) is dropped for the sake of simplicity. As we assume that the noise \\( n \\) has zero mean and independent to blur kernel \\( k \\), Eq. (3) can be simplified as" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.673, + 0.786, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\epsilon = (1 - \\widehat {W} X) \\overline {{(1 - \\widehat {W} X)}} S _ {H} + \\widehat {W} \\overline {{\\widehat {W}}} S _ {N _ {g}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.763 + ], + "angle": 0, + "content": ", where \\( S_{H} = \\mathbb{E}[|H|^{2}] \\) and \\( S_{N_g} = \\mathbb{E}[|N_g|^2] \\) are the mPSD of blur kernel and noise, respectively, and \\( \\bar{\\cdot} \\) denotes complex conjugation. Eq. (4) can be solved by computing the derivative of \\( \\epsilon \\) with respect to \\( \\widehat{W} \\) to become zero. Now, we can calculate the \\( \\widehat{H} \\) from \\( X \\) and \\( Y_{m} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.774, + 0.786, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\widehat {H} = \\widehat {W} Y _ {m} = \\frac {\\overline {{X}}}{| X | ^ {2} + k} Y _ {m} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": ", where noise suppression factor \\( k \\) is the ratio between the mPSD of noise and blur kernel. The final \\( \\hat{h} \\) can be acquired by applying an inverse Fourier transform" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Blind image deblurring with noise-robust kernel estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.53, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.271, + 0.383, + 0.283 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.156, + 0.785, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.271, + 0.668, + 0.283 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.295, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Fig. 1: Investigation of the working behavior of the proposed kernel estimation function. (a) Visualization of the estimated blur kernel from the Eq. (6), with \\( k = 0.1\\sigma^2 \\), \\( \\sigma^2 \\), \\( 10\\sigma^2 \\), and \\( 100\\sigma^2 \\), (b) Estimation error of blur kernel, and visual comparison between the estimated kernels under the optimal condition (solid line) and with those having minimum error values (dashed line)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.392, + 0.784, + 0.422 + ], + "angle": 0, + "content": "which is followed by absolute and cropping operation, \\(C:\\mathbb{R}^N\\to \\mathbb{R}^M\\) to Eq. (5). The resultant PSF can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.432, + 0.785, + 0.449 + ], + "angle": 0, + "content": "\\[\nh = C \\left(\\mathcal {W} \\left(y _ {m}, x, k\\right)\\right) = F \\left(y _ {m}, x, k\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.548 + ], + "angle": 0, + "content": ", where \\(\\mathcal{W}\\) represents inverse filtering of \\(y_{m}\\) with \\(w\\). Although Eq. (6) is a closed-form solution, the uncertainty of mPSD of noise and blur kernel prevents computation. In the case of the Eq. (2), noise-to-signal ratio \\(\\eta\\) should be carefully selected for successful reconstruction [8, 12]. Similarly, optimal blur kernel estimation requires the appropriate noise suppression factor \\(k\\). The optimal value for \\(k\\) is derived and analyzed in the following." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.583, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Claim 2. Given Gaussian noise \\(\\sim \\mathcal{N}(0, \\sigma^2)\\), the optimal noise suppression factor is given as \\(k \\simeq A(\\sigma)^2\\), where \\(A\\) represents the smoothness of blur kernel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.763 + ], + "angle": 0, + "content": "As we have discussed, the noise suppression factor in Eq. (5) is the ratio between mPSD of noise and blur kernel which is analytically derived. According to the Wiener-Khintchine theorem, the power spectral density of a random process can be expressed as the Fourier transform of auto-correlation [20,21]. In the case of Gaussian noise \\(\\sim \\mathcal{N}(0,\\sigma^2)\\), the auto-correlation is defined as \\(R_{nn}(\\tau) = \\sigma^2\\delta(\\tau)\\) so its Fourier transform, i.e. mPSD, is: \\(S_{N_g} = \\sigma^2\\). Although the mPSD of the blur kernel rarely is known, it represents the smoothness of the blur kernel. Here, we empirically found that \\(A\\simeq 1\\) for sparse blur kernel such as motion blur while \\(A\\simeq 100\\) for smooth blur kernel such as Gaussian blur." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.796 + ], + "angle": 0, + "content": "To see how the noise suppression factor affects the reconstruction quality of the blur kernel, we reformulate the Eq. (5) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.804, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\widehat {H} = \\frac {\\overline {{X}}}{| X | ^ {2} + k} (H X + N _ {g}) = H + H \\left(\\frac {- k + \\overline {{X}} N _ {g} / H}{| X | ^ {2} + k}\\right) \\tag {7}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C.Lee et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.511, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.295, + 0.374, + 0.308 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.147, + 0.652, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.148, + 0.783, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.295, + 0.66, + 0.308 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.319, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 2: (a) Illustration of the mean squared error between the measured blurry image \\( y_{m} \\) and the estimated blurry image \\( y_{est} \\). (b) Visual comparison of the \\( y_{m} \\) and \\( y_{m} - y_{est} \\) for different \\( k \\). The inset shows the estimated blur kernel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.79, + 0.606 + ], + "angle": 0, + "content": "Fig. 1 (a) shows the reconstructed blur kernel for different noise suppression factors. When \\( k = \\sigma^2 \\), the proposed kernel estimation function accurately reconstructs the kernel even in strong noise. However, when \\( k \\) is small compared to the noise level, i.e. \\( k \\ll \\frac{\\overline{X}N_g}{H} \\), Eq. (7) can be expressed as \\( \\widehat{H} = H + \\frac{N_g}{X} \\). As shown in Fig. 1 (a) where \\( k = 0.1\\sigma^2 \\), the Fourier spectrum of noise normalized by the spectrum of a clear image, i.e. \\( \\frac{N_g}{X} \\), is added to the blur kernel which leads to severe degradation of the reconstructed blur kernel. On the other hand, when \\( k \\) is too large, i.e. \\( k \\gg \\frac{\\overline{X}N_g}{H} \\), Eq. (7) can be expressed as \\( \\widehat{H} = H(1 - \\frac{k}{|X|^2 + k}) \\). As shown in Fig. 1 (a) where \\( k = 10\\sigma^2 \\) and \\( k = 100\\sigma^2 \\), since the high-frequency spectrum of \\( H \\) becomes zero due to the \\( \\frac{k}{|X|^2 + k} \\approx 1 \\), the blur kernel exhibits as a low-pass filtered. Also, Fig. 1 (b) demonstrates that the proposed optimal condition closely approximates the minimization of kernel estimation error. The further analysis for \\( F \\) is described in Supplementary S1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.683, + 0.646 + ], + "angle": 0, + "content": "3.3 Multiple kernel estimation for unknown noise level" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.842 + ], + "angle": 0, + "content": "A classical strategy to solve inverse problems including blind deblurring is minimizing penalized least square (PLS) methods. The data-fidelity term in PLS formulation of the blind deblurring problem is: \\(\\mathrm{argmin}_{x,h}\\| y_m - x*h\\| _2^2\\). In our cases, as \\(h\\) can be estimated through \\(F\\), the data-fidelity term can be expressed as \\(\\mathrm{argmin}_x\\| y_m - x*F(y_m,x,k)\\| _2^2\\), which requires setting of \\(k\\) for optimization. However, because we cannot accurately determine the noise level, setting the optimal \\(k\\) is extremely challenging. Nevertheless, since \\(F\\) is a closed-form inverse mapping function, which minimizes Eq. (3), deviating from the optimal condition of \\(k\\) results in increased noise or blurriness rather than leading to incorrect solutions. In other words, even if the \\(k\\) value is not strictly set, as long as it falls within a certain boundary, the data-fidelity term can be minimized. We verified this by examining the mean squared error (MSE) between \\(y_{m}\\) and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Blind image deblurring with noise-robust kernel estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.311, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.194, + 0.29, + 0.206 + ], + "angle": 0, + "content": "(a) Blur" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.145, + 0.407, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.194, + 0.382, + 0.206 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.145, + 0.501, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.194, + 0.473, + 0.206 + ], + "angle": 0, + "content": "(c) \\(x_0\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.597, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.194, + 0.589, + 0.206 + ], + "angle": 0, + "content": "(d) \\(x_{1}\\) and \\(h_1\\)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.145, + 0.691, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.194, + 0.685, + 0.206 + ], + "angle": 0, + "content": "(e) \\(x_{2}\\) and \\(h_2\\)" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.145, + 0.787, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.194, + 0.779, + 0.206 + ], + "angle": 0, + "content": "(f) \\(x_{3}\\) and \\(h_3\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.217, + 0.785, + 0.273 + ], + "angle": 0, + "content": "Fig. 3: Visualization of the blind deblurring results using the multiple kernel estimation scheme. (a) Blur image degraded with Gaussian noise level of \\(\\sigma = 20\\), (b) ground truth, (c) reconstructed clear image, and estimated intermediate image and blur kernels with (d) \\(k = 5^2\\), (e) \\(k = 15^2\\), and (f) \\(k = 25^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.304, + 0.785, + 0.439 + ], + "angle": 0, + "content": "\\(y_{est} = x * F(y_m, x, k)\\) by changing \\(k\\). As shown in Fig. 2 (a), the MSE decreases as \\(k\\) approaches the optimal condition. Also, the lower error value is maintained over a large range of \\(k\\), and there exists a range around the optimal suppression factor where \\(|y_m - x * F(y_m, x, k)| \\leq \\varepsilon\\) even if \\(h\\) is not optimally estimated (Fig. 2 (a) red box). Fig. 2 (b) further demonstrates this by visually comparing \\(y_{est}\\), \\(F(y_m, x, k)\\), and differential map \\(y_m - y_{est}\\) for different \\(k\\) values when noise level is \\(\\sigma = 10\\). Based on this observation, we propose a multiple kernel estimation scheme that handles a wide range of unknown noise levels by setting multiple values for \\(k\\), rather than a specific value as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.452, + 0.785, + 0.468 + ], + "angle": 0, + "content": "\\[\nh _ {i} = F \\left(y _ {m}, x _ {i}, k _ {i}\\right), \\quad \\forall i = 1, \\dots , N _ {k} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.48, + 0.788, + 0.678 + ], + "angle": 0, + "content": ", where \\( h_i \\) is the estimated blur kernel given intermediate image \\( x_i \\) and \\( k_i \\) is noise suppression factors corresponding to different noise levels. The fundamental concept of this framework lies in managing unknown noise levels by setting \\( k_i \\) from low to high, thus covering a broad range of noise that exists in nature. Here, we also introduce intermediate images, which are separately estimated from a clear image, to compensate for the error term in Eq. (7) caused by a suboptimal setting of \\( k \\), thereby stabilizing the optimization process. Fig. 3 shows an example of deblurred image \\( x_0 \\), estimated intermediate images \\( x_i \\) and blur kernels \\( h_i \\). A detailed analysis for intermediate images and the performance of multiple kernel estimation scheme, compared to case where noise level is precisely known, are presented in Supplementary S2. We demonstrate that the proposed multiple kernel estimation scheme exhibits deblurring performance comparable to cases where the noise level is precisely known, and see for the detailed analysis." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.699, + 0.557, + 0.714 + ], + "angle": 0, + "content": "3.4 MAP approach for blind deblurring" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.725, + 0.78, + 0.74 + ], + "angle": 0, + "content": "Conventionally, the MAP approach for blind deblurring is described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.753, + 0.785, + 0.769 + ], + "angle": 0, + "content": "\\[\np (x, h | y) \\propto p (y | x, h) p (x) p (h) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.825 + ], + "angle": 0, + "content": ", where \\( p(y|x,h) \\), \\( p(x) \\), and \\( p(h) \\) correspond to data-fidelity term, clear image prior, and blur kernel prior, respectively. Many previous works carefully designed the priors to strictly confine the feasible solution space." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.825, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this work, we show that leveraging the capabilities of DIP allows solving" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C.Lee et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.149, + 0.787, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.789, + 0.33 + ], + "angle": 0, + "content": "Fig. 4: Workflow of the proposed method. A single deep neural network generates the deblurred image \\( x_0 \\), and the intermediate images \\( x_i \\) which are used to estimate blur kernel \\( h_i \\) with the noise suppression factor \\( k_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.784, + 0.412 + ], + "angle": 0, + "content": "the blind deblurring problem by estimating solely a clear image even in the presence of noise. Specifically, as blur kernel can be analytically obtained from Eq. (6) if clear image \\( x \\) and blurry image \\( y_{m} \\) are given, Eq. (9) can be reformulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.415, + 0.786, + 0.431 + ], + "angle": 0, + "content": "\\[\np (x, h | y) = p (h | x, y) p (x | y) \\propto p (y | x) p (x) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.511 + ], + "angle": 0, + "content": "As shown in Eq. (10), it is evident that the blind deblurring problem can be addressed by solely focusing on estimating a clear image, while also effectively handling noise owing to the robustness of the proposed kernel estimation function. Therefore, we aim to utilize DIP, one of the optimal methods for estimating natural images." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.543, + 0.731, + 0.559 + ], + "angle": 0, + "content": "3.5 Blind deblurring based on noise-robust kernel estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.694, + 0.58 + ], + "angle": 0, + "content": "We formulate the noise-robust blind deblurring method as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.586, + 0.64, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\min _ {x _ {0}, h _ {1}, \\dots , h _ {N _ {k}}} \\frac {1}{N _ {k}} \\sum_ {i = 1} ^ {N _ {k}} | | y _ {m} - x _ {0} * h _ {i} | | _ {2} ^ {2} + \\lambda R (x _ {0})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.627, + 0.785, + 0.646 + ], + "angle": 0, + "content": "\\[\ns. t. x _ {0} = g _ {\\phi_ {0}} \\left(G _ {\\theta} (z)\\right), h _ {i} = F \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.649, + 0.589, + 0.682 + ], + "angle": 0, + "content": "\\[\nh _ {i} ^ {j} \\geq 0, \\sum_ {j} h _ {i} ^ {j} = 1 \\quad \\forall j = 0, 1, \\dots , M\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.764 + ], + "angle": 0, + "content": ", where \\(z\\) is fixed tensor randomly sampled from uniform distribution, \\((\\cdot)^j\\) denotes \\(j\\)-th elements, \\(g_{\\phi_i}\\) is \\(i\\)-th group convolution filter convolved with the last feature layer of \\(G_{\\theta}\\) to generate \\(x_{i}\\), \\(R(\\cdot)\\) is regularization function for \\(x_0\\), and \\(\\lambda\\) is regularization weight. Although we adopt total variation for \\(R(\\cdot)\\), other regularization functions such as \\(l_{0}\\)-sparsity or \\(l_{1} / l_{2}\\) norm can be used." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.825 + ], + "angle": 0, + "content": "As shown in Eq. (11), the blind deblurring problem can be solved by focusing on estimating \\( x \\). Also, it should be noted that the estimation of \\( x_0 \\) and \\( x_i \\) can be achieved with a single neural network as \\( x_i \\) should have a similar appearance with \\( x_0 \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Blind image deblurring with noise-robust kernel estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.786, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Blind deblurring with the noise-robust kernel estimation function" + }, + { + "type": "algorithm", + "bbox": [ + 0.219, + 0.165, + 0.787, + 0.407 + ], + "angle": 0, + "content": "1: Input: Blurred image \\( y_{m} \\), pre-defined \\( \\epsilon, \\lambda_{1}, \\lambda_{2}, k_{1}, k_{2}, \\dots, k_{N_{k}} \\) \n2: Output: clear image \\( x_{0} \\) and estimated blur kernels \\( h_{1}, h_{2}, \\dots, h_{N_{k}} \\) \n3: Sample \\( z \\) from uniform distribution \n4: for \\( t = 1 \\) to \\( N_{iter} \\) do \n5: \\( [f_{0}, f_{1}, \\dots, f_{N_{k}}] = G_{\\theta}(z) \\) \n6: \\( x_{0} = g_{\\phi_{0}}(f_{0}) \\) \n7: \\( \\mathcal{L}_{tot} = 0 \\) \n8: for \\( i = 1 \\) to \\( N_{k} \\) do \n9: \\( h_{i} = F(y_{m}, g_{\\phi_{i}}(f_{i}), k_{i}) \\) \n10: \\( h_{i} = |h_{i}| \\) Non-negativity \n11: \\( h_{i} = h_{i} / (\\sum h_{i} + \\epsilon) \\) Equiltity \n12: \\( \\mathcal{L}_{tot} = \\mathcal{L}_{tot} + \\frac{1}{N_{k}} \\left( ||y_{m} - x_{0} * h_{i}||_{2}^{2} + \\lambda_{1} ||W(y_{m}, g_{\\phi_{i}}(f_{i}), k_{i})||_{2}^{2} \\right) \\) \n13: end for \n14: \\( \\mathcal{L}_{tot} = \\mathcal{L}_{tot} + \\lambda_{2} TV(x_{0}) \\) \n15: Update \\( \\{\\theta, \\phi\\} \\) using \\( \\nabla_{\\theta, \\phi} \\mathcal{L}_{tot} \\) \n16: end for \n17: return \\( x_{0}, h_{1}, h_{2}, \\dots, h_{N_{k}} \\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.447, + 0.573, + 0.462 + ], + "angle": 0, + "content": "3.6 Optimization of the proposed method" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.498, + 0.785, + 0.53 + ], + "angle": 0, + "content": "The workflow of the proposed method is illustrated in Fig. 4 and the network parameters \\(\\theta, \\Phi = \\{\\phi_0, \\dots, \\phi_{N_k}\\}\\) are optimized by the following cost function:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.553, + 0.785, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta , \\phi_ {0}, \\dots \\phi_ {N _ {k}}} \\frac {1}{N _ {k}} \\sum_ {i = 1} ^ {N _ {k}} \\left(\\left| \\left| y _ {m} - g _ {\\phi_ {0}} \\left(G _ {\\theta} (z)\\right) * F \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) \\right| \\right| _ {2} ^ {2} \\right. \\\\ + \\lambda_ {1} \\| \\mathcal {W} \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) \\| _ {2} ^ {2}) + \\lambda_ {2} T V \\left(g _ {\\phi_ {0}} \\left(G _ {\\theta} (z)\\right)\\right) \\tag {12} \\\\ s. t. F \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) ^ {J} \\geq 0, \\\\ \\sum_ {j} F (y _ {m}, g _ {\\phi_ {i}} (G _ {\\theta} (z)), k _ {i}) ^ {j} = 1 \\quad \\forall j = 0, 1, \\dots , M \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.827 + ], + "angle": 0, + "content": ", where \\( TV \\) is total variation, and \\( \\lambda_{1} \\) and \\( \\lambda_{2} \\) are regularization weights. The nonnegativity and equality constraints to the estimated blur kernel are induced by taking an absolute operation followed by dividing each element by the sum of all elements. Also, we additionally introduce a sparsity regularization term for the inverse filtered image, \\( \\mathcal{W}(y_m, g_{\\phi_i}(G_\\theta(z)), k_i) \\), as all pixels except for the blur kernel positioned should have a value of 0. It is important to emphasize that our proposed multiple kernel estimation scheme, rather than regularization terms, is the key enabler for robust deblurring (see the ablation study in Supplementary S7). The optimization of Eq. (12) is formally summarized in Algorithm 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "C.Lee et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.311, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.216, + 0.311, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.289, + 0.29, + 0.301 + ], + "angle": 0, + "content": "(a) Blur" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.145, + 0.407, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.216, + 0.407, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.289, + 0.382, + 0.301 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.145, + 0.501, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.216, + 0.501, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.289, + 0.48, + 0.301 + ], + "angle": 0, + "content": "(c) \\(l_{1} / l_{2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.597, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.216, + 0.597, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.289, + 0.585, + 0.301 + ], + "angle": 0, + "content": "(d) Pure-let" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.145, + 0.691, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.216, + 0.69, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.289, + 0.675, + 0.301 + ], + "angle": 0, + "content": "(e)WDIP" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.145, + 0.786, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.216, + 0.786, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.289, + 0.765, + 0.301 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison of blind deblurring results on AFHQ datasets [6]. (row 1): AFHQ-dog motion deblurring with a kernel size of \\(43 \\times 43\\) and noise level of \\(\\sigma = 10\\), (row 2): AFHQ-cat motion deblurring with a kernel size of \\(39 \\times 39\\) and noise level of \\(\\sigma = 20\\). Here, the estimated kernel for \\(k_{2} = 15^{2}\\) is presented in (f) while others are presented in Supplementary S5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.375, + 0.426 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.44, + 0.533, + 0.456 + ], + "angle": 0, + "content": "4.1 Experiments on simulated image" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Motion deblur on AFHQ dataset [6]. To demonstrate the deblurring performance on motion blur, we conduct experiments on AFHQ-dog \\(256 \\times 256\\) and AFHQ-cat \\(256 \\times 256\\) datasets. Specifically, we select 500 test images as an evaluation dataset. To synthesize motion-blurred images, motion-blur kernels of varying sizes, ranging from \\(21 \\times 21\\) to \\(51 \\times 51\\), are randomly generated [7], and each kernel is then convolved with a clear image. Finally, Gaussian noise \\(\\sigma = 10\\) and \\(\\sigma = 20\\) are added to the synthesized blurry images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The proposed method (refer Supplementary S3 and S6 for implementation detail) is compared with other state-of-the-art blind deblurring methods including Sanghvi et al. [35], Xu et al. [40], Pure-let [24], Selfdeblur [33], WDIP [3], \\( l_{1} / l_{2} \\) [18], Pan-DCP [30], Pan-\\( l_{0} \\) [29], and Liang et al. [4], and see Supplementary S4 for implementation details. Note that we exclusively choose blind deblurring methods that do not utilize any training dataset to ensure a fair comparison. When optimizing the proposed method, \\( k_{1} = 5^{2} \\), \\( k_{2} = 15^{2} \\), and \\( k_{3} = 25^{2} \\) are set to accommodate a wide range of unknown noise levels. The representative blind deblurring results from \\( l_{1} / l_{2} \\), Pure-let, WDIP, and the proposed method are qualitatively compared in Fig. 5. Due to the severe degradation caused by Gaussian noise in blurred images, the deblurring results obtained from the compared methods are greatly compromised. Even with laborious adjustment of hyperparameters, these methods tend to overfit strongly to the noise. In contrast, the deblurring results from the proposed method shown in Fig. 5 (f) demonstrate noise-robust deblurring by effectively reconstructing clear images and blur kernels simultaneously even in the absence of precise knowledge about the noise level. Furthermore, quantitative comparison presented in Tab. 1 indicates that the proposed method achieves state-of-the-art performance in terms of Peak" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Blind image deblurring with noise-robust kernel estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of blind deblurring results on AFHQ datasets [6] with different noise levels (\\(\\sigma = 10\\) and \\(\\sigma = 20\\)). **Bold:** best." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.185, + 0.787, + 0.395 + ], + "angle": 0, + "content": "
MethodAFHQ-dogAFHQ-cat
σ = 10σ = 20σ = 10σ = 20
PSNR↑ SSIM↑ PSNR↑ SSIM↑ PSNR↑ SSIM↑ PSNR↑ SSIM
Proposed24.230.63222.900.53923.330.59022.110.497
SelfDeblur [33]13.540.10712.020.06312.830.08011.440.046
WDIP [3]14.110.11512.430.06613.430.08611.900.049
l1/l2 [18]19.690.33516.030.14619.360.30815.920.140
Pan-DCP [30]10.240.0607.6980.02410.170.0537.7130.022
Pan-l0 [29]8.3110.0316.8360.0168.2620.0296.9220.016
Liang et al. [4]10.680.0617.7110.02310.480.0577.6940.022
Sanghvi et al. [35]16.840.52016.840.52216.260.46516.380.471
Xu et al. [40]16.510.39815.010.29515.940.35714.340.258
Pure-let [24]21.020.53419.940.47720.780.48219.790.429
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.784, + 0.469 + ], + "angle": 0, + "content": "Signal-to-noise ratio (PSNR) and Structural Similarity Index Measure (SSIM). See Supplementary S5 for additional qualitative comparisons of deblurring results and quantitative comparisons at low-noise levels (\\(\\sigma = 2\\) and \\(\\sigma = 5\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.785, + 0.613 + ], + "angle": 0, + "content": "Sun et al. dataset [38]. To further showcase the capability of the proposed method, we utilize the Sun et al. dataset, comprising 80 natural images and 8 blur kernels, resulting in the generation of 640 blurry images. The representative blind deblurring results are presented in Fig. 6. The deblurring results obtained from compared methods - \\( l_{1} / l_{2} \\), Pan-DCP, and WDIP - are catastrophically corrupted when noise increases. Conversely, the proposed method shows superior deblurring results regardless of the noise level. Also, as shown in Tab. 2, the proposed method establishes the state-of-the-art deblurring performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.63, + 0.65 + ], + "angle": 0, + "content": "4.2 Experiments on scanning electron microscope" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Scanning electron microscopes (SEM) provide sub-nanometer resolution images by detecting the back-scattered or emitted electrons from the sample. However, the incident electrons, which determine the shape of the PSF of SEM can be easily distorted when the electron passes through several lenses and coils, see Fig. 7 (a). Additionally, random fluctuations of the incident electrons cause Gaussian noise, significantly compromising image quality, see Fig. 8 (a). Correcting such degradation often requires manual adjustment of system parameters, such as working distance or stigmator, by experts. To automate this correction process image quality assessment based method [17,25,26], bayesian optimization based approach [2,34], and feedback-loop based deep learning approach [22,27,36] have been proposed. Nevertheless, previous methods necessitate iterative measurement of SEM images to determine correction parameters, rendering the process" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "C.Lee et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.298, + 0.29, + 0.31 + ], + "angle": 0, + "content": "(a) Blur" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.146, + 0.407, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.299, + 0.382, + 0.31 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.146, + 0.501, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.299, + 0.479, + 0.31 + ], + "angle": 0, + "content": "(c) \\(l_{1} / l_{2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.146, + 0.596, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.299, + 0.589, + 0.31 + ], + "angle": 0, + "content": "(d) Pan-DCP" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.146, + 0.691, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.299, + 0.675, + 0.31 + ], + "angle": 0, + "content": "(e)WDIP" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.146, + 0.786, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.299, + 0.765, + 0.31 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.322, + 0.784, + 0.364 + ], + "angle": 0, + "content": "Fig.6: Qualitative comparison of blind deblurring results on Sun et al. datasets [38] with Gaussian noise level of (row 1): \\(\\sigma = 10\\) and (row 2): \\(\\sigma = 20\\). Here, the estimated kernel for \\(k_{2} = 15^{2}\\) is presented in (f) while others are presented in Supplementary S5." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.38, + 0.784, + 0.409 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison of blind deblurring results on Sun et al. datasets [38] with different noise levels (\\(\\sigma = 10\\) and \\(\\sigma = 20\\)). Bold: best." + }, + { + "type": "table", + "bbox": [ + 0.336, + 0.422, + 0.662, + 0.596 + ], + "angle": 0, + "content": "
Methodσ = 10σ = 20
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
Proposed25.940.67724.830.608
SelfDeblur [33]17.120.20215.440.136
WDIP [3]17.600.21412.890.103
l1/l2 [18]18.870.17314.970.073
Pan-DCP [30]11.620.0598.3210.021
Pan-l0 [29]9.5900.0308.0650.017
Liang et al. [4]11.960.0608.5220.022
Xu et al. [40]14.100.20713.450.128
Pure-let [24]18.870.25417.310.121
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.628, + 0.493, + 0.643 + ], + "angle": 0, + "content": "time-consuming and energy-inefficient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Here, we demonstrate the effectiveness of the proposed blind deblurring method in correcting degraded SEM images. In general, well-focused PSF of SEM can be approximated to Gaussian shape [32, 43], we set \\( A = 100 \\). Additionally, since the exact noise level of SEM images is rarely known, we set the noise suppression factor in a broad range, such as \\( k_{1} = 10000 \\), \\( k_{2} = 20000 \\), and \\( k_{3} = 30000 \\). Fig. 7 and Fig. 8 presents visual comparisons of blind deblurring results. Although methods such as Pan-DCP or WDIP have successfully performed deblurring on images with weak noise and simple patterns, their performance noticeably deteriorates in cases of strong noise or complex patterns. In contrast, as shown in Fig. 7 (f) and Fig. 8 (f), the proposed method consistently reconstructs both the clear image and the PSF, regardless of the complexity of the image or the level of noise. Especially, as shown in the second row of Fig. 7, when" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Blind image deblurring with noise-robust kernel estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.313, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.29, + 0.29, + 0.302 + ], + "angle": 0, + "content": "(a) Blur" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.144, + 0.407, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.29, + 0.382, + 0.302 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.144, + 0.501, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.29, + 0.479, + 0.302 + ], + "angle": 0, + "content": "(c) \\(l_{1} / l_{2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.597, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.29, + 0.589, + 0.302 + ], + "angle": 0, + "content": "(d) Pan-DCP" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.144, + 0.691, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.29, + 0.675, + 0.302 + ], + "angle": 0, + "content": "(e) WDIP" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.144, + 0.787, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.29, + 0.765, + 0.302 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.313, + 0.789, + 0.37 + ], + "angle": 0, + "content": "Fig. 7: Qualitative comparison of blind deblurring results on SEM image of semiconductor. (row 1): Circular ingredient of a semiconductor with the image size of \\(512 \\times 512\\), (row 2): Wire structure of semiconductor with the image size of \\(512 \\times 512\\). The ground truth PSFs are obtained by the Eq. (6) with \\(k = 10000\\)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.383, + 0.312, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.529, + 0.29, + 0.54 + ], + "angle": 0, + "content": "(a) Blur" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.383, + 0.408, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.529, + 0.382, + 0.54 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.383, + 0.501, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.529, + 0.479, + 0.54 + ], + "angle": 0, + "content": "(c) \\(l_{1} / l_{2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.383, + 0.597, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.529, + 0.589, + 0.54 + ], + "angle": 0, + "content": "(d) Pan-DCP" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.383, + 0.691, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.529, + 0.675, + 0.54 + ], + "angle": 0, + "content": "(e)WDIP" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.383, + 0.787, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.529, + 0.765, + 0.54 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.552, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Fig. 8: Qualitative comparison of blind deblurring results on SEM image severely degraded by noise, with an image size of \\(256 \\times 256\\). The ground truth PSFs are obtained by the Eq. (6) with \\(k = 10000\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.685 + ], + "angle": 0, + "content": "examining an image with blurred wire patterns, predicting the structure can be challenging even for human observers. However, the deblurred images from the proposed method show a remarkable similarity to the ground truth images and one can effectively investigate the wire structures." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.709, + 0.782, + 0.725 + ], + "angle": 0, + "content": "4.3 Experiments on two-photon excitation fluorescence microscope" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We further applied the proposed method as one of the real-world applications to a two-photon excitation fluorescence microscope (2PEF) image. In general, the imaging model of 2PEF follows Eq. (1) and its effective PSF can be expressed as the square of illumination PSF [13]. Also, due to the inherent characteristic of weak signals of fluorescence, the measured signal is often contaminated by noise. Here, we captured the fluorescence signal of \\(2\\mathrm{mm}\\) beads from the custom-built 2PEF setup. As shown in Fig. 9 (a), the measured image from the 2PEF" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "C.Lee et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.359, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.145, + 0.427, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.145, + 0.498, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.256, + 0.384, + 0.267 + ], + "angle": 0, + "content": "(a) Blur" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.145, + 0.646, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.145, + 0.715, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.145, + 0.785, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.256, + 0.674, + 0.268 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.279, + 0.785, + 0.308 + ], + "angle": 0, + "content": "Fig. 9: Visualization of blind deblurring results of the proposed method on blurry two-photon excitation fluorescence microscope image with an image size of \\(384 \\times 384\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.785, + 0.412 + ], + "angle": 0, + "content": "exhibits a blurry appearance, attributed to the diffraction limit of the illumination PSF, and is further corrupted by noise. However, when the proposed deblurring method is applied to the measured image, illustrated in Fig. 9 (b), it effectively reconstructs the sharp edges of beads and facilitates easy discernment of clustered beads." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.436, + 0.368, + 0.451 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.664 + ], + "angle": 0, + "content": "In this paper, we propose a noise-robust blind deblurring method combining the multiple kernel estimation scheme and the DIP. The most critical aspect of the proposed method, the noise-robust kernel estimation function, successfully restores the PSF even in situations of high noise when provided only with clear and blurry images. This function is developed to the multiple kernel estimation scheme that can be applied even in scenarios where noise level cannot be accurately predicted. Additionally, by effectively leveraging DIP, which acts as a powerful prior for natural images, we can estimate a clear image from a single well-designed neural network. The discussion of limitations of the proposed work and ablation study are presented in Supplementary S7. Our proposed method has been demonstrated to operate effectively in both simulated images and real-world applications and outperforms compared to the previous state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.703, + 0.403, + 0.72 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.785, + 0.81 + ], + "angle": 0, + "content": "This work was supported by the Samsung Research Funding and Incubation Center of Samsung Electronics grant SRFC-IT2002-03, Samsung Electronics Co., Ltd. (IO220908-02403-01), and the National Research Foundation of Korea grant funded by the Korean government (Grant Nos. NRF-2021R1A5A1032937, NRF-2021R1C1C1011307, RS-2023-00251628, and RS-2024-00397673)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Blind image deblurring with noise-robust kernel estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.201 + ], + "angle": 0, + "content": "1. Arjomand Bigdeli, S., Zwicker, M., Favaro, P., Jin, M.: Deep mean-shift priors for image restoration. Advances in Neural Information Processing Systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.228 + ], + "angle": 0, + "content": "2. Binding, J., Mikula, S., Denk, W.: Low-dosage maximum-a-posteriori focusing and stigmation. Microscopy and Microanalysis 19(1), 38-55 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.228, + 0.785, + 0.269 + ], + "angle": 0, + "content": "3. Bredell, G., Erdil, E., Weber, B., Konukoglu, E.: Wiener guided dip for unsupervised blind image deconvolution. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3047-3056 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.269, + 0.785, + 0.31 + ], + "angle": 0, + "content": "4. Chen, L., Fang, F., Lei, S., Li, F., Zhang, G.: Enhanced sparse model for blind deblurring. In: European Conference on Computer Vision. pp. 631-646. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.338 + ], + "angle": 0, + "content": "5. Cho, S., Lee, S.: Fast motion deblurring. In: ACM SIGGRAPH Asia 2009 papers, pp. 1-8 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.378 + ], + "angle": 0, + "content": "6. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.419 + ], + "angle": 0, + "content": "7. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6059-6069 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "8. Dhawan, A.P., Rangayyan, R.M., Gordon, R.: Image restoration by wiener deconvolution in limited-view computed tomography. Applied optics 24(23), 4013-4020 (1985)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "9. Dong, J., Pan, J., Sun, D., Su, Z., Yang, M.H.: Learning data terms for non-blind deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 748-763 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "0. Dong, J., Roth, S., Schiele, B.: Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems 33, 1048-1059 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "1. Gandelsman, Y., Shocher, A., Irani, M.: \"double-dip\": unsupervised image decomposition via coupled deep-image-priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11026-11035 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.731, + 0.596 + ], + "angle": 0, + "content": "2. Gonzalez, R.C.: Digital image processing. Pearson education india (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.785, + 0.622 + ], + "angle": 0, + "content": "3. Helmchen, F., Denk, W.: Deep tissue two-photon microscopy. Nature methods 2(12), 932-940 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.663 + ], + "angle": 0, + "content": "4. Jin, M., Roth, S., Favaro, P.: Noise-blind image deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3510-3518 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.691 + ], + "angle": 0, + "content": "5. Jin, M., Roth, S., Favaro, P.: Normalized blind deconvolution. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 668-684 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "6. Joshi, N., Szeliski, R., Kriegman, D.J.: Psf estimation using sharp edge prediction. In: 2008 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1-8. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "7. Kawasaki, T., Nakano, T., Hirose, K.: Developing an aberration-corrected schottky emission sem and method for measuring aberration. Microelectronic engineering 86(4-6), 1017-1020 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "8. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Kruse, J., Rother, C., Schmidt, U.: Learning to push the limits of efficient ft-based image deconvolution. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4586-4594 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "C.Lee et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Kubo, R., Toda, M., Hashitsume, N.: Statistical physics II: nonequilibrium statistical mechanics, vol. 31. Springer Science & Business Media (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "21. Lampard, D.: Generalization of the wiener-khintchine theorem to nonstationary processes. Journal of Applied Physics 25(6), 802-803 (1954)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "22. Lee, W., Nam, H.S., Kim, Y.G., Kim, Y.J., Lee, J.H., Yoo, H.: Robust autofocusing for scanning electron microscopy based on a dual deep learning network. Scientific reports 11(1), 20933 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.249, + 0.785, + 0.289 + ], + "angle": 0, + "content": "23. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.291, + 0.785, + 0.318 + ], + "angle": 0, + "content": "24. Li, J., Luisier, F., Blu, T.: Pure-let image deconvolution. IEEE Transactions on Image Processing 27(1), 92-105 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.32, + 0.785, + 0.36 + ], + "angle": 0, + "content": "25. Lifshin, E., Kandel, Y.P., Moore, R.L.: Improving scanning electron microscope resolution for near planar samples through the use of image restoration. Microscopy and Microanalysis 20(1), 78-89 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.362, + 0.785, + 0.388 + ], + "angle": 0, + "content": "26. Lu, Y., Zhang, X., Li, H.: A simplified focusing and astigmatism correction method for a scanning electron microscope. AIP Advances 8(1) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.785, + 0.431 + ], + "angle": 0, + "content": "27. Na, J., Kim, G., Kang, S.H., Kim, S.J., Lee, S.: Deep learning-based discriminative refocusing of scanning electron microscopy images for materials science. Acta Materialia 214, 116987 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "28. Nan, Y., Quan, Y., Ji, H.: Variational-em-based deep learning for noise-blind image deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3626-3635 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.475, + 0.785, + 0.515 + ], + "angle": 0, + "content": "29. Pan, J., Hu, Z., Su, Z., Yang, M.H.: Deblurring text images via l0-regularized intensity and gradient prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2901-2908 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.785, + 0.558 + ], + "angle": 0, + "content": "30. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1628-1636 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.6 + ], + "angle": 0, + "content": "31. Perrone, D., Favaro, P.: Total variation blind deconvolution: The devil is in the details. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2909-2916 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.785, + 0.628 + ], + "angle": 0, + "content": "32. Reimer, L.: Scanning electron microscopy: physics of image formation and microanalysis. Measurement Science and Technology 11(12), 1826-1826 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.671 + ], + "angle": 0, + "content": "33. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3341-3350 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.741 + ], + "angle": 0, + "content": "34. Roels, J., Aelterman, J., De Vylder, J., Luong, H., Saeys, Y., Philips, W.: Bayesian deconvolution of scanning electron microscopy images using point-spread function estimation and non-local regularization. In: 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). pp. 443-447. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.783 + ], + "angle": 0, + "content": "35. Sanghvi, Y., Mao, Z., Chan, S.H.: Structured kernel estimation for photon-limited deconvolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9863-9872 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "36. Schubert, P.J., Saxena, R., Kornfeld, J.: Deepfocus: Fast focus and astigmatism correction for electron microscopy. arXiv preprint arXiv:2305.04977 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "37. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. Acm transactions on graphics (tog) 27(3), 1-10 (2008)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Blind image deblurring with noise-robust kernel estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "39. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9446-9454 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "40. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 157-170. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "41. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1107-1114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "42. Xue, F., Luisier, F., Blu, T.: Multi-wiener sure-let deconvolution. IEEE Transactions on Image Processing 22(5), 1954–1968 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "43. Zotta, M.D., Nevins, M.C., Hailstone, R.K., Lifshin, E.: The determination and application of the point spread function in the scanning electron microscope. Microscopy and microanalysis 24(4), 396-405 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "44. Zuo, W., Ren, D., Zhang, D., Gu, S., Zhang, L.: Learning iteration-wise generalized shrinkage-thresholding operators for blind deconvolution. IEEE Transactions on Image Processing 25(4), 1751-1764 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.425 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Blind image deblurring with noise-robust kernel estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_origin.pdf b/2024/Blind image deblurring with noise-robust kernel estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8fe0d11c7d480f9def2ee7c1f8367ea45789d56b --- /dev/null +++ b/2024/Blind image deblurring with noise-robust kernel estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bf12e320e0a48a21f68e9186ab689a44974bb5a1c08dbf4f5067ae9bfb96b34 +size 7463030 diff --git a/2024/Blind image deblurring with noise-robust kernel estimation/full.md b/2024/Blind image deblurring with noise-robust kernel estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4f1d7b584af0c8387c47af71b888b5101497eec2 --- /dev/null +++ b/2024/Blind image deblurring with noise-robust kernel estimation/full.md @@ -0,0 +1,415 @@ +# Blind image deblurring with noise-robust kernel estimation + +Chanseok Lee $^{1}$ , Jeongsol Kim $^{1}$ , Seungmin Lee $^{1}$ , Jaehwang Jung $^{2}$ , Yunje Cho $^{2}$ , Taejoong Kim $^{2}$ , Taeyong Jo $^{2}$ , Myungjun Lee $^{2}$ , and Mooseok Jang $^{1,3}$ + +$^{1}$ Department of Bio and Brain Engineering, KAIST, Daejeon, South Korea + +$^{2}$ Advanced Process Development Team 4, Semiconductor R&D Center, Samsung Electronics, South Korea + +$^{3}$ KAIST Institute for Health Science and Technology, KAIST, Daejeon, South Korea {cslee, mooseok}@kaist.ac.kr + +Abstract. Blind deblurring is an ill-posed inverse problem involving the retrieval of a clear image and blur kernel from a single blurry image. The challenge arises considerably when strong noise, where its level remains unknown, is introduced. Existing blind deblurring methods are highly susceptible to noise due to overfitting and disturbances in the solution space. Here, we propose a blind deblurring method based on a noise-robust kernel estimation function and deep image prior (DIP). Specifically, the proposed kernel estimation function effectively estimates the blur kernel even for strongly noisy blurry images given a clear image and optimal condition. Therefore, DIP is adopted for the generation of a clear image to leverage its natural image prior. Additionally, the multiple kernel estimation scheme is designed to address a wide range of unknown noise levels. Extensive experimental studies, including simulated images and real-world examples, demonstrate the superior deblurring performance of the proposed method. The official code is uploaded in https://github.com/csleemooo/BD_noise_robust_kernel_estimation. + +Keywords: Blind deblurring $\cdot$ Deep image prior $\cdot$ Gaussian noise + +# 1 Introduction + +Optical imaging systems such as brightfield microscopes, fluorescence microscopes, or scanning electron microscopes (SEM) have a crucial role in histopathology, inspection, and clinical or biological research by providing complicated structural information of target imaging samples with high-resolution images. However, perturbation or misalignment (e.g. defocus or aberrations) of the imaging system due to its instability or systemic error brings blurry effects on the captured image. Furthermore, additive Gaussian Noise severely degrades the captured image quality. Such degradation process can be described as follows: + +$$ +y _ {m} = x * h + n _ {g} \tag {1} +$$ + +, where $y_{m} \in \mathbb{R}^{N}$ is measured blurry and noisy image, $x \in \mathbb{R}^{N}$ is a clear image, * denotes the convolution operator, $h \in \mathbb{R}^{M}$ is blur kernel or point spread function (PSF) of the imaging system, and $n_g$ is additive noise. Retrieving clean and focused image $x$ as well as blur kernel $h$ from a single degraded image $y_{m}$ is well known as blind deconvolution or blind image deblurring and has been actively investigated for decades due to its ill-posedness. Although many previous approaches such as edge-selection-based method [5,16,37,38] or MAP-based approach [4,15,18,23,29-31,41,44] show promising results, deblurring performances significantly drop as the noise level increases. + +In general, there is a trade-off when simultaneously trying to eliminate both the blur effect and noise. For example, when focusing on removing the blur effect, the estimated image may be severely corrupted by noise or other high-frequency artifacts. On the other hand, when focusing on noise removal, there is a tendency to overly emphasize smoothness, compromising the fine details of the image. So, balancing the strength of deblurring and noise suppression is important to successfully reconstruct $x$ and $h$ . Moreover, in cases where the noise level cannot be accurately determined, the difficulty of the deblurring task increases significantly. + +This paper aims to address the challenge of blind deblurring in the presence of strong and an unknown noise level. Instead of relying on sharp edge-selective maps or prior information to estimate the blur kernel, we introduce a novel kernel estimation function. This function analytically computes the blur kernel when provided with a clear image, blurry image, and noise suppression factor. The first characteristic of the proposed function is its ability to estimate the blur kernel effectively even in scenarios with strong noise. Specifically, the noise suppression factor included in the kernel estimation function is pivotal in mitigating noise, and its optimal condition is derived accordingly. To further ameliorate this characteristic, we designed a multiple kernel estimation scheme, which effectively handles a wide range of noise levels, to perform blind deblurring even in situations where the noise level is unknown. The second characteristic is that the estimation of only the clear image is sufficient for estimating a blur kernel. In other words, the amount of information required for solving a blind deblurring problem is much less than that needed by conventional methods, and we validated this through maximum a posteriori (MAP) analysis. This characteristic is fully leveraged by employing DIP, known as one of the most effective methods for capturing priors of natural images and solving inverse mapping problems [39]. The inherent robustness of DIP to noise, combined with the noise-robust multiple kernel estimation scheme, enables superior performance in blind deblurring tasks. + +The main contributions of this paper can be summarized as follows: + +1. We propose a noise-robust kernel estimation function along with its optimal conditions, enabling the analytical acquisition of an ideal blur kernel when provided with a clear image and a blurry image. +2. We design a multiple kernel estimation scheme to address a wide range of unknown noise levels. + +3. We adopt DIP to leverage the natural image prior captured by the handcrafted neural network for the generation of both a clear image and intermediate images which are used for estimating blur kernels. + +4. The experimental results and detailed analysis with simulation and experimental data consistently demonstrate that the proposed method outperforms in blind deblurring tasks, particularly in scenarios where strong noise exists. + +# 2 Related works + +MAP-based method The Blind deblurring problem has been actively investigated for decades as a blurred image can be explained as many pairs of clear images and blur kernels. One of the prevalent approaches involves addressing the problem within the MAP framework consisting of priors of clear image and blur kernel. For example, Krishnan et al. [18] proposed the normalized sparsity regularization $l_{1} / l_{2}$ to maximize the prior of clear image. Pan et al. [30] adopted the dark channel on the premise that the dark channel of a clean image exhibits much sparsity compared to the blurred image. Other sparsity constraints such as TV-regularization [31], $l_{0}$ norm [29], $l_{p}$ norm [15,44] have been proposed to regularize the solution space of a clear image. Also, there have been some prior works to regularize the blur kernel estimation such as sparsity priors [23, 44]. Also, equality constraint $\sum h = 1$ and non-negativity $h \geq 0$ are imposed to blur kernel estimation for the stability of convergence. + +Recently, DIP has been incorporated into the MAP framework for blind deblurring problems. Specifically, the inherent capacity of a well-designed neural network to capture image statistics is leveraged by substituting the regularization function with the neural network itself. For example, Ren et al. [33] adopts double-DIP [11] to infer the latent information of clear image and blur kernel. The elaborately designed generators have shown state-of-the-art blind deblurring performance without requiring explicit knowledge of clear image and blur kernel. Bredell et al. [3] incorporates the Wiener filter as a guide during the training of the DIP, effectively improving deblurring performance. + +Noise-blind deblurring The deblurring problem becomes more challenging when the noise level is not accurately known. One expected approach is to estimate the noise level first and then apply state-of-the-art deblurring methods. However, the estimation of the noise level in blurry images has not yet been reported, making it difficult to apply. To remedy this problem, methods such as adopting Bayesian estimator [1,14], variational expected maximization-based method [28], and other few training-based methods [9,19,35] have been proposed and shown state-of-the-art deblurring performances. + +# 3 Method + +# 3.1 Wiener deconvolution method + +The core idea of the Wiener deconvolution method is to find an inverse filter $w_{h\rightarrow x}$ that can estimate true signal $\hat{x}$ . The optimal inverse filter can be derived from error minimization of the difference between the inverse filtered image $\hat{x}$ and target image $x$ : $\epsilon = \mathbb{E}[|X - \hat{X}|^2]$ , where $\mathbb{E}$ denotes expectation, $X$ and $\hat{X}$ is Fourier transform of $x$ and $\hat{x}$ , respectively. The estimated true signal can be acquired from the following equation: + +$$ +\hat {X} = \frac {\bar {H}}{| H | ^ {2} + \eta} Y _ {m} \tag {2} +$$ + +, where $\bar{\cdot}$ denotes complex conjugation and $\eta$ is the ratio between the mean power spectral density (mPSD) of noise and signal. As Eq. (2) is simple but effective in reducing noise and deblur the degraded image, many of previous works [3, 10, 24, 42] employed Wiener filter to their deblurring framework. Inspired by this, we derive the following kernel estimation function. + +# 3.2 Noise-robust kernel estimation function + +Claim 1. Given blurry image $y_{m}$ and clean image $x$ , the blur kernel $\widehat{h}$ can be estimated by $\widehat{H} = \frac{\overline{X}}{|X|^2 + k} Y_m$ , where $\widehat{H}, Y_m,$ and $X$ are Fourier transform of $\widehat{h}, y_m,$ and $x$ , respectively, and $k$ is the ratio between the mean power spectral density of noise and blur kernel. + +Proof. We formulated the following mean squared error to derive a kernel estimation function using an inverse filter $w_{x \to h}$ . + +$$ +\epsilon = \mathbb {E} [ | H - \widehat {H} | ^ {2} ] = \mathbb {E} [ | H - \widehat {W} [ H X + N _ {g} ] | ^ {2} ] = \mathbb {E} [ | [ 1 - \widehat {W} X ] H - \widehat {W} N _ {g} | ^ {2} ] \qquad (3) +$$ + +Here, the subscript of an inverse filter $w$ is dropped for the sake of simplicity. As we assume that the noise $n$ has zero mean and independent to blur kernel $k$ , Eq. (3) can be simplified as + +$$ +\epsilon = (1 - \widehat {W} X) \overline {{(1 - \widehat {W} X)}} S _ {H} + \widehat {W} \overline {{\widehat {W}}} S _ {N _ {g}} \tag {4} +$$ + +, where $S_{H} = \mathbb{E}[|H|^{2}]$ and $S_{N_g} = \mathbb{E}[|N_g|^2]$ are the mPSD of blur kernel and noise, respectively, and $\bar{\cdot}$ denotes complex conjugation. Eq. (4) can be solved by computing the derivative of $\epsilon$ with respect to $\widehat{W}$ to become zero. Now, we can calculate the $\widehat{H}$ from $X$ and $Y_{m}$ as follows: + +$$ +\widehat {H} = \widehat {W} Y _ {m} = \frac {\overline {{X}}}{| X | ^ {2} + k} Y _ {m} \tag {5} +$$ + +, where noise suppression factor $k$ is the ratio between the mPSD of noise and blur kernel. The final $\hat{h}$ can be acquired by applying an inverse Fourier transform + +![](images/f6bb09889a2e64afe4090400c1031ab64d604c77d0f9250035ecd1c624194f7b.jpg) +(a) + +![](images/660d90fedd022a900149f760ac2f2508de8440a8ab91e3913acecb60f6aab7e5.jpg) +(b) +Fig. 1: Investigation of the working behavior of the proposed kernel estimation function. (a) Visualization of the estimated blur kernel from the Eq. (6), with $k = 0.1\sigma^2$ , $\sigma^2$ , $10\sigma^2$ , and $100\sigma^2$ , (b) Estimation error of blur kernel, and visual comparison between the estimated kernels under the optimal condition (solid line) and with those having minimum error values (dashed line). + +which is followed by absolute and cropping operation, $C:\mathbb{R}^N\to \mathbb{R}^M$ to Eq. (5). The resultant PSF can be described as follows: + +$$ +h = C \left(\mathcal {W} \left(y _ {m}, x, k\right)\right) = F \left(y _ {m}, x, k\right) \tag {6} +$$ + +, where $\mathcal{W}$ represents inverse filtering of $y_{m}$ with $w$ . Although Eq. (6) is a closed-form solution, the uncertainty of mPSD of noise and blur kernel prevents computation. In the case of the Eq. (2), noise-to-signal ratio $\eta$ should be carefully selected for successful reconstruction [8, 12]. Similarly, optimal blur kernel estimation requires the appropriate noise suppression factor $k$ . The optimal value for $k$ is derived and analyzed in the following. + +Claim 2. Given Gaussian noise $\sim \mathcal{N}(0, \sigma^2)$ , the optimal noise suppression factor is given as $k \simeq A(\sigma)^2$ , where $A$ represents the smoothness of blur kernel. + +As we have discussed, the noise suppression factor in Eq. (5) is the ratio between mPSD of noise and blur kernel which is analytically derived. According to the Wiener-Khintchine theorem, the power spectral density of a random process can be expressed as the Fourier transform of auto-correlation [20,21]. In the case of Gaussian noise $\sim \mathcal{N}(0,\sigma^2)$ , the auto-correlation is defined as $R_{nn}(\tau) = \sigma^2\delta(\tau)$ so its Fourier transform, i.e. mPSD, is: $S_{N_g} = \sigma^2$ . Although the mPSD of the blur kernel rarely is known, it represents the smoothness of the blur kernel. Here, we empirically found that $A\simeq 1$ for sparse blur kernel such as motion blur while $A\simeq 100$ for smooth blur kernel such as Gaussian blur. + +To see how the noise suppression factor affects the reconstruction quality of the blur kernel, we reformulate the Eq. (5) as follows: + +$$ +\widehat {H} = \frac {\overline {{X}}}{| X | ^ {2} + k} (H X + N _ {g}) = H + H \left(\frac {- k + \overline {{X}} N _ {g} / H}{| X | ^ {2} + k}\right) \tag {7} +$$ + +![](images/6fc89e2f65f6225ca74f7a4aabdd4f1455eabbe479055e3383163ef9320fe2c5.jpg) +(a) + +![](images/0dc747ae8ab989fd33035b51fcc46e1c4303c24aaac8cfdc881e230805bc57cd.jpg) +Fig. 2: (a) Illustration of the mean squared error between the measured blurry image $y_{m}$ and the estimated blurry image $y_{est}$ . (b) Visual comparison of the $y_{m}$ and $y_{m} - y_{est}$ for different $k$ . The inset shows the estimated blur kernel. + +![](images/c9fb8f4e282714bb52c0d9acf4d7587594f06b6cd631e107a68a71577f0e58a2.jpg) +(b) + +Fig. 1 (a) shows the reconstructed blur kernel for different noise suppression factors. When $k = \sigma^2$ , the proposed kernel estimation function accurately reconstructs the kernel even in strong noise. However, when $k$ is small compared to the noise level, i.e. $k \ll \frac{\overline{X}N_g}{H}$ , Eq. (7) can be expressed as $\widehat{H} = H + \frac{N_g}{X}$ . As shown in Fig. 1 (a) where $k = 0.1\sigma^2$ , the Fourier spectrum of noise normalized by the spectrum of a clear image, i.e. $\frac{N_g}{X}$ , is added to the blur kernel which leads to severe degradation of the reconstructed blur kernel. On the other hand, when $k$ is too large, i.e. $k \gg \frac{\overline{X}N_g}{H}$ , Eq. (7) can be expressed as $\widehat{H} = H(1 - \frac{k}{|X|^2 + k})$ . As shown in Fig. 1 (a) where $k = 10\sigma^2$ and $k = 100\sigma^2$ , since the high-frequency spectrum of $H$ becomes zero due to the $\frac{k}{|X|^2 + k} \approx 1$ , the blur kernel exhibits as a low-pass filtered. Also, Fig. 1 (b) demonstrates that the proposed optimal condition closely approximates the minimization of kernel estimation error. The further analysis for $F$ is described in Supplementary S1. + +# 3.3 Multiple kernel estimation for unknown noise level + +A classical strategy to solve inverse problems including blind deblurring is minimizing penalized least square (PLS) methods. The data-fidelity term in PLS formulation of the blind deblurring problem is: $\mathrm{argmin}_{x,h}\| y_m - x*h\| _2^2$ . In our cases, as $h$ can be estimated through $F$ , the data-fidelity term can be expressed as $\mathrm{argmin}_x\| y_m - x*F(y_m,x,k)\| _2^2$ , which requires setting of $k$ for optimization. However, because we cannot accurately determine the noise level, setting the optimal $k$ is extremely challenging. Nevertheless, since $F$ is a closed-form inverse mapping function, which minimizes Eq. (3), deviating from the optimal condition of $k$ results in increased noise or blurriness rather than leading to incorrect solutions. In other words, even if the $k$ value is not strictly set, as long as it falls within a certain boundary, the data-fidelity term can be minimized. We verified this by examining the mean squared error (MSE) between $y_{m}$ and + +![](images/5d42bcfe8ba0906ea41c9a6da6b2048be1f3bfee9e19c82ca1ca6586cee64dbc.jpg) +(a) Blur + +![](images/8c30511b65de6bed8ecba639631a5f784b4a2b2b2535c11dd22c2073bbdcc678.jpg) +(b) GT +Fig. 3: Visualization of the blind deblurring results using the multiple kernel estimation scheme. (a) Blur image degraded with Gaussian noise level of $\sigma = 20$ , (b) ground truth, (c) reconstructed clear image, and estimated intermediate image and blur kernels with (d) $k = 5^2$ , (e) $k = 15^2$ , and (f) $k = 25^2$ . + +![](images/be1c66ed0e06eb4379033d92495a8377b3849677013cb888052a5163d061d748.jpg) +(c) $x_0$ + +![](images/2ba296c635461c04756535dbb7c37414be49ab62f0b1ac4bc96486a903010545.jpg) +(d) $x_{1}$ and $h_1$ + +![](images/edd4a63306f7f6565a0c46316bd60858c6f1e508cc4d56eba9b2f987f866c530.jpg) +(e) $x_{2}$ and $h_2$ + +![](images/67c8c496a96f29d32c1231f254587aa912621414330e5f49c51644bf34910e42.jpg) +(f) $x_{3}$ and $h_3$ + +$y_{est} = x * F(y_m, x, k)$ by changing $k$ . As shown in Fig. 2 (a), the MSE decreases as $k$ approaches the optimal condition. Also, the lower error value is maintained over a large range of $k$ , and there exists a range around the optimal suppression factor where $|y_m - x * F(y_m, x, k)| \leq \varepsilon$ even if $h$ is not optimally estimated (Fig. 2 (a) red box). Fig. 2 (b) further demonstrates this by visually comparing $y_{est}$ , $F(y_m, x, k)$ , and differential map $y_m - y_{est}$ for different $k$ values when noise level is $\sigma = 10$ . Based on this observation, we propose a multiple kernel estimation scheme that handles a wide range of unknown noise levels by setting multiple values for $k$ , rather than a specific value as follows: + +$$ +h _ {i} = F \left(y _ {m}, x _ {i}, k _ {i}\right), \quad \forall i = 1, \dots , N _ {k} \tag {8} +$$ + +, where $h_i$ is the estimated blur kernel given intermediate image $x_i$ and $k_i$ is noise suppression factors corresponding to different noise levels. The fundamental concept of this framework lies in managing unknown noise levels by setting $k_i$ from low to high, thus covering a broad range of noise that exists in nature. Here, we also introduce intermediate images, which are separately estimated from a clear image, to compensate for the error term in Eq. (7) caused by a suboptimal setting of $k$ , thereby stabilizing the optimization process. Fig. 3 shows an example of deblurred image $x_0$ , estimated intermediate images $x_i$ and blur kernels $h_i$ . A detailed analysis for intermediate images and the performance of multiple kernel estimation scheme, compared to case where noise level is precisely known, are presented in Supplementary S2. We demonstrate that the proposed multiple kernel estimation scheme exhibits deblurring performance comparable to cases where the noise level is precisely known, and see for the detailed analysis. + +# 3.4 MAP approach for blind deblurring + +Conventionally, the MAP approach for blind deblurring is described as follows: + +$$ +p (x, h | y) \propto p (y | x, h) p (x) p (h) \tag {9} +$$ + +, where $p(y|x,h)$ , $p(x)$ , and $p(h)$ correspond to data-fidelity term, clear image prior, and blur kernel prior, respectively. Many previous works carefully designed the priors to strictly confine the feasible solution space. + +In this work, we show that leveraging the capabilities of DIP allows solving + +![](images/292ba58fcfe7b1722cb91025744194da9cbf7f0bc7015cfd4686c93fed563122.jpg) +Fig. 4: Workflow of the proposed method. A single deep neural network generates the deblurred image $x_0$ , and the intermediate images $x_i$ which are used to estimate blur kernel $h_i$ with the noise suppression factor $k_i$ . + +the blind deblurring problem by estimating solely a clear image even in the presence of noise. Specifically, as blur kernel can be analytically obtained from Eq. (6) if clear image $x$ and blurry image $y_{m}$ are given, Eq. (9) can be reformulated as follows: + +$$ +p (x, h | y) = p (h | x, y) p (x | y) \propto p (y | x) p (x) \tag {10} +$$ + +As shown in Eq. (10), it is evident that the blind deblurring problem can be addressed by solely focusing on estimating a clear image, while also effectively handling noise owing to the robustness of the proposed kernel estimation function. Therefore, we aim to utilize DIP, one of the optimal methods for estimating natural images. + +# 3.5 Blind deblurring based on noise-robust kernel estimation + +We formulate the noise-robust blind deblurring method as follows: + +$$ +\min _ {x _ {0}, h _ {1}, \dots , h _ {N _ {k}}} \frac {1}{N _ {k}} \sum_ {i = 1} ^ {N _ {k}} | | y _ {m} - x _ {0} * h _ {i} | | _ {2} ^ {2} + \lambda R (x _ {0}) +$$ + +$$ +s. t. x _ {0} = g _ {\phi_ {0}} \left(G _ {\theta} (z)\right), h _ {i} = F \left(y _ {m}, g _ {\phi_ {i}} \left(G _ {\theta} (z)\right), k _ {i}\right) \tag {11} +$$ + +$$ +h _ {i} ^ {j} \geq 0, \sum_ {j} h _ {i} ^ {j} = 1 \quad \forall j = 0, 1, \dots , M +$$ + +, where $z$ is fixed tensor randomly sampled from uniform distribution, $(\cdot)^j$ denotes $j$ -th elements, $g_{\phi_i}$ is $i$ -th group convolution filter convolved with the last feature layer of $G_{\theta}$ to generate $x_{i}$ , $R(\cdot)$ is regularization function for $x_0$ , and $\lambda$ is regularization weight. Although we adopt total variation for $R(\cdot)$ , other regularization functions such as $l_{0}$ -sparsity or $l_{1} / l_{2}$ norm can be used. + +As shown in Eq. (11), the blind deblurring problem can be solved by focusing on estimating $x$ . Also, it should be noted that the estimation of $x_0$ and $x_i$ can be achieved with a single neural network as $x_i$ should have a similar appearance with $x_0$ . + +Algorithm 1 Blind deblurring with the noise-robust kernel estimation function +1: Input: Blurred image $y_{m}$ , pre-defined $\epsilon, \lambda_{1}, \lambda_{2}, k_{1}, k_{2}, \dots, k_{N_{k}}$ +2: Output: clear image $x_{0}$ and estimated blur kernels $h_{1}, h_{2}, \dots, h_{N_{k}}$ +3: Sample $z$ from uniform distribution +4: for $t = 1$ to $N_{iter}$ do +5: $[f_{0}, f_{1}, \dots, f_{N_{k}}] = G_{\theta}(z)$ +6: $x_{0} = g_{\phi_{0}}(f_{0})$ +7: $\mathcal{L}_{tot} = 0$ +8: for $i = 1$ to $N_{k}$ do +9: $h_{i} = F(y_{m}, g_{\phi_{i}}(f_{i}), k_{i})$ +10: $h_{i} = |h_{i}|$ Non-negativity +11: $h_{i} = h_{i} / (\sum h_{i} + \epsilon)$ Equiltity +12: $\mathcal{L}_{tot} = \mathcal{L}_{tot} + \frac{1}{N_{k}} \left( ||y_{m} - x_{0} * h_{i}||_{2}^{2} + \lambda_{1} ||W(y_{m}, g_{\phi_{i}}(f_{i}), k_{i})||_{2}^{2} \right)$ +13: end for +14: $\mathcal{L}_{tot} = \mathcal{L}_{tot} + \lambda_{2} TV(x_{0})$ +15: Update $\{\theta, \phi\}$ using $\nabla_{\theta, \phi} \mathcal{L}_{tot}$ +16: end for +17: return $x_{0}, h_{1}, h_{2}, \dots, h_{N_{k}}$ + +# 3.6 Optimization of the proposed method + +The workflow of the proposed method is illustrated in Fig. 4 and the network parameters $\theta, \Phi = \{\phi_0, \dots, \phi_{N_k}\}$ are optimized by the following cost function: + +$$ +\begin{array}{l} \min _ {\theta , \phi_ {0}, \dots \phi_ {N _ {k}}} \frac {1}{N _ {k}} \sum_ {i = 1} ^ {N _ {k}} \left(\left| \left| y _ {m} - g _ {\phi_ {0}} \left(G _ {\theta} (z)\right) * F \left(y _ {m}, g _ {\phi_ {i}} \left(G _ {\theta} (z)\right), k _ {i}\right) \right| \right| _ {2} ^ {2} \right. \\ + \lambda_ {1} \| \mathcal {W} \left(y _ {m}, g _ {\phi_ {i}} \left(G _ {\theta} (z)\right), k _ {i}\right) \| _ {2} ^ {2}) + \lambda_ {2} T V \left(g _ {\phi_ {0}} \left(G _ {\theta} (z)\right)\right) \tag {12} \\ s. t. F \left(y _ {m}, g _ {\phi_ {i}} \left(G _ {\theta} (z)\right), k _ {i}\right) ^ {J} \geq 0, \\ \sum_ {j} F (y _ {m}, g _ {\phi_ {i}} (G _ {\theta} (z)), k _ {i}) ^ {j} = 1 \quad \forall j = 0, 1, \dots , M \\ \end{array} +$$ + +, where $TV$ is total variation, and $\lambda_{1}$ and $\lambda_{2}$ are regularization weights. The nonnegativity and equality constraints to the estimated blur kernel are induced by taking an absolute operation followed by dividing each element by the sum of all elements. Also, we additionally introduce a sparsity regularization term for the inverse filtered image, $\mathcal{W}(y_m, g_{\phi_i}(G_\theta(z)), k_i)$ , as all pixels except for the blur kernel positioned should have a value of 0. It is important to emphasize that our proposed multiple kernel estimation scheme, rather than regularization terms, is the key enabler for robust deblurring (see the ablation study in Supplementary S7). The optimization of Eq. (12) is formally summarized in Algorithm 1. + +![](images/13de169d7d55466941e68403acc4b420e99f8c39dffd00ed33a7e3748cb09b30.jpg) + +![](images/88385f9821f2b47f409e11b8229b4758bf63cc6d0edcc20011064e35db71c31d.jpg) +(a) Blur + +![](images/180ab968c1be7815ea047dcfce4853e8cd56b45773b9fd785aac7dfaadcfecbf.jpg) + +![](images/83def309de633c2f0e302d10fa3f3fe04f7998ddbb4e533c1f2ae4ed796363ef.jpg) +(b) GT + +![](images/99c7b6449e71b28cbfabda034f4fc405b77816b39faef4dd85044b436ebb790c.jpg) + +![](images/3f298cdff34af1aa35e81681c2da16938bea4db79c346e4d9257dbb4a7475fbe.jpg) +(c) $l_{1} / l_{2}$ + +![](images/346d94c32be2260eebcbc040721c37d490b172cbcf48853b63348fe2d77a25e4.jpg) + +![](images/6ff3dce2b1d5c01ac5d88320064970b025abc774fc5e91410c10f5cc34430b79.jpg) +(d) Pure-let +Fig. 5: Qualitative comparison of blind deblurring results on AFHQ datasets [6]. (row 1): AFHQ-dog motion deblurring with a kernel size of $43 \times 43$ and noise level of $\sigma = 10$ , (row 2): AFHQ-cat motion deblurring with a kernel size of $39 \times 39$ and noise level of $\sigma = 20$ . Here, the estimated kernel for $k_{2} = 15^{2}$ is presented in (f) while others are presented in Supplementary S5. + +![](images/1345ee935c3bc39125403dfbefc00df9d4497c141548aaf2c8181063c96befd9.jpg) + +![](images/b98692ca6737de00134b898ec4498bbf3db969247ddb874f0721680952063187.jpg) +(e)WDIP + +![](images/26c42b85b1e385e20935519c305803366a4c244debfb8e9db382d277b802f6ff.jpg) + +![](images/f421912ad5d077e38ac54e32774b4721262fa0edc3dd37a79bfb2f11b68f8a5a.jpg) +(f) Ours + +# 4 Experiments + +# 4.1 Experiments on simulated image + +Motion deblur on AFHQ dataset [6]. To demonstrate the deblurring performance on motion blur, we conduct experiments on AFHQ-dog $256 \times 256$ and AFHQ-cat $256 \times 256$ datasets. Specifically, we select 500 test images as an evaluation dataset. To synthesize motion-blurred images, motion-blur kernels of varying sizes, ranging from $21 \times 21$ to $51 \times 51$ , are randomly generated [7], and each kernel is then convolved with a clear image. Finally, Gaussian noise $\sigma = 10$ and $\sigma = 20$ are added to the synthesized blurry images. + +The proposed method (refer Supplementary S3 and S6 for implementation detail) is compared with other state-of-the-art blind deblurring methods including Sanghvi et al. [35], Xu et al. [40], Pure-let [24], Selfdeblur [33], WDIP [3], $l_{1} / l_{2}$ [18], Pan-DCP [30], Pan- $l_{0}$ [29], and Liang et al. [4], and see Supplementary S4 for implementation details. Note that we exclusively choose blind deblurring methods that do not utilize any training dataset to ensure a fair comparison. When optimizing the proposed method, $k_{1} = 5^{2}$ , $k_{2} = 15^{2}$ , and $k_{3} = 25^{2}$ are set to accommodate a wide range of unknown noise levels. The representative blind deblurring results from $l_{1} / l_{2}$ , Pure-let, WDIP, and the proposed method are qualitatively compared in Fig. 5. Due to the severe degradation caused by Gaussian noise in blurred images, the deblurring results obtained from the compared methods are greatly compromised. Even with laborious adjustment of hyperparameters, these methods tend to overfit strongly to the noise. In contrast, the deblurring results from the proposed method shown in Fig. 5 (f) demonstrate noise-robust deblurring by effectively reconstructing clear images and blur kernels simultaneously even in the absence of precise knowledge about the noise level. Furthermore, quantitative comparison presented in Tab. 1 indicates that the proposed method achieves state-of-the-art performance in terms of Peak + +Table 1: Quantitative comparison of blind deblurring results on AFHQ datasets [6] with different noise levels ( $\sigma = 10$ and $\sigma = 20$ ). **Bold:** best. + +
MethodAFHQ-dogAFHQ-cat
σ = 10σ = 20σ = 10σ = 20
PSNR↑ SSIM↑ PSNR↑ SSIM↑ PSNR↑ SSIM↑ PSNR↑ SSIM
Proposed24.230.63222.900.53923.330.59022.110.497
SelfDeblur [33]13.540.10712.020.06312.830.08011.440.046
WDIP [3]14.110.11512.430.06613.430.08611.900.049
l1/l2 [18]19.690.33516.030.14619.360.30815.920.140
Pan-DCP [30]10.240.0607.6980.02410.170.0537.7130.022
Pan-l0 [29]8.3110.0316.8360.0168.2620.0296.9220.016
Liang et al. [4]10.680.0617.7110.02310.480.0577.6940.022
Sanghvi et al. [35]16.840.52016.840.52216.260.46516.380.471
Xu et al. [40]16.510.39815.010.29515.940.35714.340.258
Pure-let [24]21.020.53419.940.47720.780.48219.790.429
+ +Signal-to-noise ratio (PSNR) and Structural Similarity Index Measure (SSIM). See Supplementary S5 for additional qualitative comparisons of deblurring results and quantitative comparisons at low-noise levels ( $\sigma = 2$ and $\sigma = 5$ ). + +Sun et al. dataset [38]. To further showcase the capability of the proposed method, we utilize the Sun et al. dataset, comprising 80 natural images and 8 blur kernels, resulting in the generation of 640 blurry images. The representative blind deblurring results are presented in Fig. 6. The deblurring results obtained from compared methods - $l_{1} / l_{2}$ , Pan-DCP, and WDIP - are catastrophically corrupted when noise increases. Conversely, the proposed method shows superior deblurring results regardless of the noise level. Also, as shown in Tab. 2, the proposed method establishes the state-of-the-art deblurring performance. + +# 4.2 Experiments on scanning electron microscope + +Scanning electron microscopes (SEM) provide sub-nanometer resolution images by detecting the back-scattered or emitted electrons from the sample. However, the incident electrons, which determine the shape of the PSF of SEM can be easily distorted when the electron passes through several lenses and coils, see Fig. 7 (a). Additionally, random fluctuations of the incident electrons cause Gaussian noise, significantly compromising image quality, see Fig. 8 (a). Correcting such degradation often requires manual adjustment of system parameters, such as working distance or stigmator, by experts. To automate this correction process image quality assessment based method [17,25,26], bayesian optimization based approach [2,34], and feedback-loop based deep learning approach [22,27,36] have been proposed. Nevertheless, previous methods necessitate iterative measurement of SEM images to determine correction parameters, rendering the process + +![](images/b417bb399ab85efeaf6b76735f329ba2e7fdc38e425e15a5485ecf7e4ebca35a.jpg) +(a) Blur + +![](images/15c876766e8aa1963570ead27b4dad9613941ef3c8db283ad3a3813aa714a668.jpg) +(b) GT + +![](images/7231a7f23e9842ca1fc11955b4e1acdefeb5697a83861bda5cb9cea164059eea.jpg) +(c) $l_{1} / l_{2}$ + +![](images/1c1472252408fed9d82c9d538ec2db7940835833975737b9ac9204263ae430cb.jpg) +(d) Pan-DCP +Fig.6: Qualitative comparison of blind deblurring results on Sun et al. datasets [38] with Gaussian noise level of (row 1): $\sigma = 10$ and (row 2): $\sigma = 20$ . Here, the estimated kernel for $k_{2} = 15^{2}$ is presented in (f) while others are presented in Supplementary S5. + +![](images/4230ca7878eb575dfe41b75ce757992611f7a97bfbf5f90369b3a2625e8329f3.jpg) +(e)WDIP + +![](images/54bd4a5fc665c271a8d67344cda2c077413179f941ac81f3a2298f067b1bcafe.jpg) +(f) Ours + +Table 2: Quantitative comparison of blind deblurring results on Sun et al. datasets [38] with different noise levels ( $\sigma = 10$ and $\sigma = 20$ ). Bold: best. + +
Methodσ = 10σ = 20
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
Proposed25.940.67724.830.608
SelfDeblur [33]17.120.20215.440.136
WDIP [3]17.600.21412.890.103
l1/l2 [18]18.870.17314.970.073
Pan-DCP [30]11.620.0598.3210.021
Pan-l0 [29]9.5900.0308.0650.017
Liang et al. [4]11.960.0608.5220.022
Xu et al. [40]14.100.20713.450.128
Pure-let [24]18.870.25417.310.121
+ +time-consuming and energy-inefficient. + +Here, we demonstrate the effectiveness of the proposed blind deblurring method in correcting degraded SEM images. In general, well-focused PSF of SEM can be approximated to Gaussian shape [32, 43], we set $A = 100$ . Additionally, since the exact noise level of SEM images is rarely known, we set the noise suppression factor in a broad range, such as $k_{1} = 10000$ , $k_{2} = 20000$ , and $k_{3} = 30000$ . Fig. 7 and Fig. 8 presents visual comparisons of blind deblurring results. Although methods such as Pan-DCP or WDIP have successfully performed deblurring on images with weak noise and simple patterns, their performance noticeably deteriorates in cases of strong noise or complex patterns. In contrast, as shown in Fig. 7 (f) and Fig. 8 (f), the proposed method consistently reconstructs both the clear image and the PSF, regardless of the complexity of the image or the level of noise. Especially, as shown in the second row of Fig. 7, when + +![](images/a8a32ea0ed20d00c1edbb9bf61b87552c778ee5a8c3858d5d82562750d4b858e.jpg) +(a) Blur + +![](images/f3eabd133a138ac205cb3f505c51e05379a239d028653354bf142678e97b7a7b.jpg) +(b) GT + +![](images/a284f85bb5c52af4d3d1f556318e7a030b72dd9f1e2057610ca66cec4bf91c2c.jpg) +(c) $l_{1} / l_{2}$ + +![](images/f55ea2035b67b5d3467f2e5996b71ef014dc4b497740d3d5738416be2451e694.jpg) +(d) Pan-DCP + +![](images/a491bd3e03431584042738a7009bf306a40eee73c1a7a7a8617628912efe0972.jpg) +(e) WDIP + +![](images/84a5ded5fd38026e56440933687be20df2fe1c10714d45a1d58cddfbf1c2618f.jpg) +(f) Ours + +![](images/59ddd69a082b932c61ee4729dbbc66e35c4a6535f3d6d30e62c24fb24b1f2500.jpg) +Fig. 7: Qualitative comparison of blind deblurring results on SEM image of semiconductor. (row 1): Circular ingredient of a semiconductor with the image size of $512 \times 512$ , (row 2): Wire structure of semiconductor with the image size of $512 \times 512$ . The ground truth PSFs are obtained by the Eq. (6) with $k = 10000$ . +(a) Blur +Fig. 8: Qualitative comparison of blind deblurring results on SEM image severely degraded by noise, with an image size of $256 \times 256$ . The ground truth PSFs are obtained by the Eq. (6) with $k = 10000$ . + +![](images/f83027979244cd36d8553bf403186df18474375e7392ff23c4a1492cdec100cf.jpg) +(b) GT + +![](images/d68bf3db04f165df39cc526bd0f950fa0158c27e317052ef8c23a38e6640e642.jpg) +(c) $l_{1} / l_{2}$ + +![](images/1f13d2132b5a737c0a62f82328bfc5fd735f8db3ec5cfaa2b418dfe2c5bb4361.jpg) +(d) Pan-DCP + +![](images/75e57a03269707fc59dbd13f49c663cca7d5a3a849cc5adb9d0dd44ef066d392.jpg) +(e)WDIP + +![](images/c5fe6898d6aadafdf7d597c8f0414f4e352a52085c576afa880d77a3923d2964.jpg) +(f) Ours + +examining an image with blurred wire patterns, predicting the structure can be challenging even for human observers. However, the deblurred images from the proposed method show a remarkable similarity to the ground truth images and one can effectively investigate the wire structures. + +# 4.3 Experiments on two-photon excitation fluorescence microscope + +We further applied the proposed method as one of the real-world applications to a two-photon excitation fluorescence microscope (2PEF) image. In general, the imaging model of 2PEF follows Eq. (1) and its effective PSF can be expressed as the square of illumination PSF [13]. Also, due to the inherent characteristic of weak signals of fluorescence, the measured signal is often contaminated by noise. Here, we captured the fluorescence signal of $2\mathrm{mm}$ beads from the custom-built 2PEF setup. As shown in Fig. 9 (a), the measured image from the 2PEF + +![](images/e7b06def865e5d6cfa51ce35bd4b2207d2ca25469b25ffe19ef0ee3058d5e036.jpg) +(a) Blur + +![](images/7073626d5e38656f58f5b5d45f5bd018aa37e59fe579d9c202a499e13c337cba.jpg) +Fig. 9: Visualization of blind deblurring results of the proposed method on blurry two-photon excitation fluorescence microscope image with an image size of $384 \times 384$ . + +![](images/dad3b828ab8cfeecf50851802d675dc429011dba2f3720f4bf57136aeb07fac9.jpg) + +![](images/23ecd2b9533b95656f370f792432d76e2bb282fa42ea07831e7f641fa38e5016.jpg) +(b) Ours + +![](images/05ee95bf8e5ffa926acb758183b9589edf218f9be8622082477962eba373e037.jpg) + +![](images/d6b89de1eaad5ef1709da2548882248e9cd65b9b4fbdb686f6550b350df6f9ea.jpg) + +exhibits a blurry appearance, attributed to the diffraction limit of the illumination PSF, and is further corrupted by noise. However, when the proposed deblurring method is applied to the measured image, illustrated in Fig. 9 (b), it effectively reconstructs the sharp edges of beads and facilitates easy discernment of clustered beads. + +# 5 Conclusions + +In this paper, we propose a noise-robust blind deblurring method combining the multiple kernel estimation scheme and the DIP. The most critical aspect of the proposed method, the noise-robust kernel estimation function, successfully restores the PSF even in situations of high noise when provided only with clear and blurry images. This function is developed to the multiple kernel estimation scheme that can be applied even in scenarios where noise level cannot be accurately predicted. Additionally, by effectively leveraging DIP, which acts as a powerful prior for natural images, we can estimate a clear image from a single well-designed neural network. The discussion of limitations of the proposed work and ablation study are presented in Supplementary S7. Our proposed method has been demonstrated to operate effectively in both simulated images and real-world applications and outperforms compared to the previous state-of-the-art methods. + +# Acknowledgements + +This work was supported by the Samsung Research Funding and Incubation Center of Samsung Electronics grant SRFC-IT2002-03, Samsung Electronics Co., Ltd. (IO220908-02403-01), and the National Research Foundation of Korea grant funded by the Korean government (Grant Nos. NRF-2021R1A5A1032937, NRF-2021R1C1C1011307, RS-2023-00251628, and RS-2024-00397673). + +# References + +1. Arjomand Bigdeli, S., Zwicker, M., Favaro, P., Jin, M.: Deep mean-shift priors for image restoration. Advances in Neural Information Processing Systems 30 (2017) +2. Binding, J., Mikula, S., Denk, W.: Low-dosage maximum-a-posteriori focusing and stigmation. Microscopy and Microanalysis 19(1), 38-55 (2013) +3. Bredell, G., Erdil, E., Weber, B., Konukoglu, E.: Wiener guided dip for unsupervised blind image deconvolution. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3047-3056 (2023) +4. Chen, L., Fang, F., Lei, S., Li, F., Zhang, G.: Enhanced sparse model for blind deblurring. In: European Conference on Computer Vision. pp. 631-646. Springer (2020) +5. Cho, S., Lee, S.: Fast motion deblurring. In: ACM SIGGRAPH Asia 2009 papers, pp. 1-8 (2009) +6. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2020) +7. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6059-6069 (2023) +8. Dhawan, A.P., Rangayyan, R.M., Gordon, R.: Image restoration by wiener deconvolution in limited-view computed tomography. Applied optics 24(23), 4013-4020 (1985) +9. Dong, J., Pan, J., Sun, D., Su, Z., Yang, M.H.: Learning data terms for non-blind deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 748-763 (2018) +0. Dong, J., Roth, S., Schiele, B.: Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems 33, 1048-1059 (2020) +1. Gandelsman, Y., Shocher, A., Irani, M.: "double-dip": unsupervised image decomposition via coupled deep-image-priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11026-11035 (2019) +2. Gonzalez, R.C.: Digital image processing. Pearson education india (2009) +3. Helmchen, F., Denk, W.: Deep tissue two-photon microscopy. Nature methods 2(12), 932-940 (2005) +4. Jin, M., Roth, S., Favaro, P.: Noise-blind image deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3510-3518 (2017) +5. Jin, M., Roth, S., Favaro, P.: Normalized blind deconvolution. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 668-684 (2018) +6. Joshi, N., Szeliski, R., Kriegman, D.J.: Psf estimation using sharp edge prediction. In: 2008 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1-8. IEEE (2008) +7. Kawasaki, T., Nakano, T., Hirose, K.: Developing an aberration-corrected schottky emission sem and method for measuring aberration. Microelectronic engineering 86(4-6), 1017-1020 (2009) +8. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011) +9. Kruse, J., Rother, C., Schmidt, U.: Learning to push the limits of efficient ft-based image deconvolution. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4586-4594 (2017) + +20. Kubo, R., Toda, M., Hashitsume, N.: Statistical physics II: nonequilibrium statistical mechanics, vol. 31. Springer Science & Business Media (2012) +21. Lampard, D.: Generalization of the wiener-khintchine theorem to nonstationary processes. Journal of Applied Physics 25(6), 802-803 (1954) +22. Lee, W., Nam, H.S., Kim, Y.G., Kim, Y.J., Lee, J.H., Yoo, H.: Robust autofocusing for scanning electron microscopy based on a dual deep learning network. Scientific reports 11(1), 20933 (2021) +23. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009) +24. Li, J., Luisier, F., Blu, T.: Pure-let image deconvolution. IEEE Transactions on Image Processing 27(1), 92-105 (2017) +25. Lifshin, E., Kandel, Y.P., Moore, R.L.: Improving scanning electron microscope resolution for near planar samples through the use of image restoration. Microscopy and Microanalysis 20(1), 78-89 (2014) +26. Lu, Y., Zhang, X., Li, H.: A simplified focusing and astigmatism correction method for a scanning electron microscope. AIP Advances 8(1) (2018) +27. Na, J., Kim, G., Kang, S.H., Kim, S.J., Lee, S.: Deep learning-based discriminative refocusing of scanning electron microscopy images for materials science. Acta Materialia 214, 116987 (2021) +28. Nan, Y., Quan, Y., Ji, H.: Variational-em-based deep learning for noise-blind image deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3626-3635 (2020) +29. Pan, J., Hu, Z., Su, Z., Yang, M.H.: Deblurring text images via l0-regularized intensity and gradient prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2901-2908 (2014) +30. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1628-1636 (2016) +31. Perrone, D., Favaro, P.: Total variation blind deconvolution: The devil is in the details. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2909-2916 (2014) +32. Reimer, L.: Scanning electron microscopy: physics of image formation and microanalysis. Measurement Science and Technology 11(12), 1826-1826 (2000) +33. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3341-3350 (2020) +34. Roels, J., Aelterman, J., De Vylder, J., Luong, H., Saeys, Y., Philips, W.: Bayesian deconvolution of scanning electron microscopy images using point-spread function estimation and non-local regularization. In: 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). pp. 443-447. IEEE (2016) +35. Sanghvi, Y., Mao, Z., Chan, S.H.: Structured kernel estimation for photon-limited deconvolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9863-9872 (2023) +36. Schubert, P.J., Saxena, R., Kornfeld, J.: Deepfocus: Fast focus and astigmatism correction for electron microscopy. arXiv preprint arXiv:2305.04977 (2023) +37. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. Acm transactions on graphics (tog) 27(3), 1-10 (2008) + +38. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013) +39. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9446-9454 (2018) +40. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 157-170. Springer (2010) +41. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1107-1114 (2013) +42. Xue, F., Luisier, F., Blu, T.: Multi-wiener sure-let deconvolution. IEEE Transactions on Image Processing 22(5), 1954–1968 (2013) +43. Zotta, M.D., Nevins, M.C., Hailstone, R.K., Lifshin, E.: The determination and application of the point spread function in the scanning electron microscope. Microscopy and microanalysis 24(4), 396-405 (2018) +44. Zuo, W., Ren, D., Zhang, D., Gu, S., Zhang, L.: Learning iteration-wise generalized shrinkage-thresholding operators for blind deconvolution. IEEE Transactions on Image Processing 25(4), 1751-1764 (2016) \ No newline at end of file diff --git a/2024/Blind image deblurring with noise-robust kernel estimation/images.zip b/2024/Blind image deblurring with noise-robust kernel estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f94605a29a741f15105a721be1304ee9e3290eea --- /dev/null +++ b/2024/Blind image deblurring with noise-robust kernel estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:872440efdecb1e9fff7f7509d983403b35273de5ae73fb58dbd53fbd31eed44f +size 751955 diff --git a/2024/Blind image deblurring with noise-robust kernel estimation/layout.json b/2024/Blind image deblurring with noise-robust kernel estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d50eb32aaf884cadabb745bd293aee70dca65203 --- /dev/null +++ b/2024/Blind image deblurring with noise-robust kernel estimation/layout.json @@ -0,0 +1,14103 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 111, + 475, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 111, + 475, + 146 + ], + "spans": [ + { + "bbox": [ + 139, + 111, + 475, + 146 + ], + "type": "text", + "content": "Blind image deblurring with noise-robust kernel estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "content": "Chanseok Lee" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "content": ", Jeongsol Kim" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "content": ", Seungmin Lee" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "content": ", Jaehwang Jung" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "content": ", Yunje Cho" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "content": ", Taejoong Kim" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "content": ", Taeyong Jo" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "content": ", Myungjun Lee" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "text", + "content": ", and Mooseok Jang" + }, + { + "bbox": [ + 134, + 168, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 213, + 461, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 213, + 461, + 224 + ], + "spans": [ + { + "bbox": [ + 149, + 213, + 461, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 213, + 461, + 224 + ], + "type": "text", + "content": " Department of Bio and Brain Engineering, KAIST, Daejeon, South Korea" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 224, + 473, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 473, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 473, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 224, + 473, + 246 + ], + "type": "text", + "content": " Advanced Process Development Team 4, Semiconductor R&D Center, Samsung Electronics, South Korea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 247, + 480, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 269 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 247, + 480, + 269 + ], + "type": "text", + "content": " KAIST Institute for Health Science and Technology, KAIST, Daejeon, South Korea {cslee, mooseok}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 300, + 453, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 300, + 453, + 475 + ], + "spans": [ + { + "bbox": [ + 159, + 300, + 453, + 475 + ], + "type": "text", + "content": "Abstract. Blind deblurring is an ill-posed inverse problem involving the retrieval of a clear image and blur kernel from a single blurry image. The challenge arises considerably when strong noise, where its level remains unknown, is introduced. Existing blind deblurring methods are highly susceptible to noise due to overfitting and disturbances in the solution space. Here, we propose a blind deblurring method based on a noise-robust kernel estimation function and deep image prior (DIP). Specifically, the proposed kernel estimation function effectively estimates the blur kernel even for strongly noisy blurry images given a clear image and optimal condition. Therefore, DIP is adopted for the generation of a clear image to leverage its natural image prior. Additionally, the multiple kernel estimation scheme is designed to address a wide range of unknown noise levels. Extensive experimental studies, including simulated images and real-world examples, demonstrate the superior deblurring performance of the proposed method. The official code is uploaded in https://github.com/csleemooo/BD_noise_robust_kernel_estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 487, + 430, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 487, + 430, + 498 + ], + "spans": [ + { + "bbox": [ + 160, + 487, + 430, + 498 + ], + "type": "text", + "content": "Keywords: Blind deblurring " + }, + { + "bbox": [ + 160, + 487, + 430, + 498 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 487, + 430, + 498 + ], + "type": "text", + "content": " Deep image prior " + }, + { + "bbox": [ + 160, + 487, + 430, + 498 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 487, + 430, + 498 + ], + "type": "text", + "content": " Gaussian noise" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": "Optical imaging systems such as brightfield microscopes, fluorescence microscopes, or scanning electron microscopes (SEM) have a crucial role in histopathology, inspection, and clinical or biological research by providing complicated structural information of target imaging samples with high-resolution images. However, perturbation or misalignment (e.g. defocus or aberrations) of the imaging system due to its instability or systemic error brings blurry effects on the captured image. Furthermore, additive Gaussian Noise severely degrades the captured image quality. Such degradation process can be described as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 272, + 654, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 654, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 272, + 654, + 481, + 666 + ], + "type": "interline_equation", + "content": "y _ {m} = x * h + n _ {g} \\tag {1}", + "image_path": "295501c6203f556f296bd91343ceb6375d60f6cbdf6046341b6bfa92ea4d75f6.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "y_{m} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " is measured blurry and noisy image, " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " is a clear image, * denotes the convolution operator, " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "h \\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " is blur kernel or point spread function (PSF) of the imaging system, and " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "n_g" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " is additive noise. Retrieving clean and focused image " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " as well as blur kernel " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " from a single degraded image " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " is well known as blind deconvolution or blind image deblurring and has been actively investigated for decades due to its ill-posedness. Although many previous approaches such as edge-selection-based method [5,16,37,38] or MAP-based approach [4,15,18,23,29-31,41,44] show promising results, deblurring performances significantly drop as the noise level increases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "text", + "content": "In general, there is a trade-off when simultaneously trying to eliminate both the blur effect and noise. For example, when focusing on removing the blur effect, the estimated image may be severely corrupted by noise or other high-frequency artifacts. On the other hand, when focusing on noise removal, there is a tendency to overly emphasize smoothness, compromising the fine details of the image. So, balancing the strength of deblurring and noise suppression is important to successfully reconstruct " + }, + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "text", + "content": ". Moreover, in cases where the noise level cannot be accurately determined, the difficulty of the deblurring task increases significantly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 582 + ], + "type": "text", + "content": "This paper aims to address the challenge of blind deblurring in the presence of strong and an unknown noise level. Instead of relying on sharp edge-selective maps or prior information to estimate the blur kernel, we introduce a novel kernel estimation function. This function analytically computes the blur kernel when provided with a clear image, blurry image, and noise suppression factor. The first characteristic of the proposed function is its ability to estimate the blur kernel effectively even in scenarios with strong noise. Specifically, the noise suppression factor included in the kernel estimation function is pivotal in mitigating noise, and its optimal condition is derived accordingly. To further ameliorate this characteristic, we designed a multiple kernel estimation scheme, which effectively handles a wide range of noise levels, to perform blind deblurring even in situations where the noise level is unknown. The second characteristic is that the estimation of only the clear image is sufficient for estimating a blur kernel. In other words, the amount of information required for solving a blind deblurring problem is much less than that needed by conventional methods, and we validated this through maximum a posteriori (MAP) analysis. This characteristic is fully leveraged by employing DIP, known as one of the most effective methods for capturing priors of natural images and solving inverse mapping problems [39]. The inherent robustness of DIP to noise, combined with the noise-robust multiple kernel estimation scheme, enables superior performance in blind deblurring tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 582, + 447, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 582, + 447, + 594 + ], + "spans": [ + { + "bbox": [ + 147, + 582, + 447, + 594 + ], + "type": "text", + "content": "The main contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 605, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 605, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 605, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 136, + 605, + 481, + 640 + ], + "type": "text", + "content": "1. We propose a noise-robust kernel estimation function along with its optimal conditions, enabling the analytical acquisition of an ideal blur kernel when provided with a clear image and a blurry image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 641, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 641, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 136, + 641, + 481, + 664 + ], + "type": "text", + "content": "2. We design a multiple kernel estimation scheme to address a wide range of unknown noise levels." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C.Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 152 + ], + "type": "text", + "content": "3. We adopt DIP to leverage the natural image prior captured by the handcrafted neural network for the generation of both a clear image and intermediate images which are used for estimating blur kernels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 156, + 480, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 156, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 136, + 156, + 480, + 191 + ], + "type": "text", + "content": "4. The experimental results and detailed analysis with simulation and experimental data consistently demonstrate that the proposed method outperforms in blind deblurring tasks, particularly in scenarios where strong noise exists." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 228, + 239, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 239, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 239, + 240 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": "MAP-based method The Blind deblurring problem has been actively investigated for decades as a blurred image can be explained as many pairs of clear images and blur kernels. One of the prevalent approaches involves addressing the problem within the MAP framework consisting of priors of clear image and blur kernel. For example, Krishnan et al. [18] proposed the normalized sparsity regularization " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " to maximize the prior of clear image. Pan et al. [30] adopted the dark channel on the premise that the dark channel of a clean image exhibits much sparsity compared to the blurred image. Other sparsity constraints such as TV-regularization [31], " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "l_{0}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " norm [29], " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " norm [15,44] have been proposed to regularize the solution space of a clear image. Also, there have been some prior works to regularize the blur kernel estimation such as sparsity priors [23, 44]. Also, equality constraint " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\sum h = 1" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " and non-negativity " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "h \\geq 0" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " are imposed to blur kernel estimation for the stability of convergence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "text", + "content": "Recently, DIP has been incorporated into the MAP framework for blind deblurring problems. Specifically, the inherent capacity of a well-designed neural network to capture image statistics is leveraged by substituting the regularization function with the neural network itself. For example, Ren et al. [33] adopts double-DIP [11] to infer the latent information of clear image and blur kernel. The elaborately designed generators have shown state-of-the-art blind deblurring performance without requiring explicit knowledge of clear image and blur kernel. Bredell et al. [3] incorporates the Wiener filter as a guide during the training of the DIP, effectively improving deblurring performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Noise-blind deblurring The deblurring problem becomes more challenging when the noise level is not accurately known. One expected approach is to estimate the noise level first and then apply state-of-the-art deblurring methods. However, the estimation of the noise level in blurry images has not yet been reported, making it difficult to apply. To remedy this problem, methods such as adopting Bayesian estimator [1,14], variational expected maximization-based method [28], and other few training-based methods [9,19,35] have been proposed and shown state-of-the-art deblurring performances." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "Blind image deblurring with noise-robust kernel estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 313, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 313, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 313, + 150 + ], + "type": "text", + "content": "3.1 Wiener deconvolution method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": "The core idea of the Wiener deconvolution method is to find an inverse filter " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "w_{h\\rightarrow x}" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": " that can estimate true signal " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": ". The optimal inverse filter can be derived from error minimization of the difference between the inverse filtered image " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": " and target image " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "\\epsilon = \\mathbb{E}[|X - \\hat{X}|^2]" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "\\mathbb{E}" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": " denotes expectation, " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": " is Fourier transform of " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 130, + 157, + 482, + 229 + ], + "type": "text", + "content": ", respectively. The estimated true signal can be acquired from the following equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 274, + 236, + 481, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 236, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 274, + 236, + 481, + 262 + ], + "type": "interline_equation", + "content": "\\hat {X} = \\frac {\\bar {H}}{| H | ^ {2} + \\eta} Y _ {m} \\tag {2}", + "image_path": "334f01591720dce11e249570fb83004d9b9b04b0ba39af736501aa311ccc9e3f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 267, + 482, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 328 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 267, + 482, + 328 + ], + "type": "inline_equation", + "content": "\\bar{\\cdot}" + }, + { + "bbox": [ + 130, + 267, + 482, + 328 + ], + "type": "text", + "content": " denotes complex conjugation and " + }, + { + "bbox": [ + 130, + 267, + 482, + 328 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 267, + 482, + 328 + ], + "type": "text", + "content": " is the ratio between the mean power spectral density (mPSD) of noise and signal. As Eq. (2) is simple but effective in reducing noise and deblur the degraded image, many of previous works [3, 10, 24, 42] employed Wiener filter to their deblurring framework. Inspired by this, we derive the following kernel estimation function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 356, + 360, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 356, + 360, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 356, + 360, + 366 + ], + "type": "text", + "content": "3.2 Noise-robust kernel estimation function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": "Claim 1. Given blurry image " + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": " and clean image " + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": ", the blur kernel " + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\widehat{h}" + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": " can be estimated by " + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\widehat{H} = \\frac{\\overline{X}}{|X|^2 + k} Y_m" + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\widehat{H}, Y_m," + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": " are Fourier transform of " + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\widehat{h}, y_m," + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 373, + 482, + 424 + ], + "type": "text", + "content": " is the ratio between the mean power spectral density of noise and blur kernel." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 435, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 482, + 460 + ], + "type": "text", + "content": "Proof. We formulated the following mean squared error to derive a kernel estimation function using an inverse filter " + }, + { + "bbox": [ + 132, + 435, + 482, + 460 + ], + "type": "inline_equation", + "content": "w_{x \\to h}" + }, + { + "bbox": [ + 132, + 435, + 482, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 160, + 466, + 481, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 466, + 481, + 481 + ], + "spans": [ + { + "bbox": [ + 160, + 466, + 481, + 481 + ], + "type": "interline_equation", + "content": "\\epsilon = \\mathbb {E} [ | H - \\widehat {H} | ^ {2} ] = \\mathbb {E} [ | H - \\widehat {W} [ H X + N _ {g} ] | ^ {2} ] = \\mathbb {E} [ | [ 1 - \\widehat {W} X ] H - \\widehat {W} N _ {g} | ^ {2} ] \\qquad (3)", + "image_path": "de12c3f2016cc67b67109b6f6502c55a07bcc97ed4390939f48b71dd249ecc26.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 487, + 482, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 482, + 524 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 482, + 524 + ], + "type": "text", + "content": "Here, the subscript of an inverse filter " + }, + { + "bbox": [ + 132, + 487, + 482, + 524 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 132, + 487, + 482, + 524 + ], + "type": "text", + "content": " is dropped for the sake of simplicity. As we assume that the noise " + }, + { + "bbox": [ + 132, + 487, + 482, + 524 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 487, + 482, + 524 + ], + "type": "text", + "content": " has zero mean and independent to blur kernel " + }, + { + "bbox": [ + 132, + 487, + 482, + 524 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 487, + 482, + 524 + ], + "type": "text", + "content": ", Eq. (3) can be simplified as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 230, + 533, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 533, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 230, + 533, + 481, + 550 + ], + "type": "interline_equation", + "content": "\\epsilon = (1 - \\widehat {W} X) \\overline {{(1 - \\widehat {W} X)}} S _ {H} + \\widehat {W} \\overline {{\\widehat {W}}} S _ {N _ {g}} \\tag {4}", + "image_path": "a87ecab41c558d77b4898b09c4418255e624c09b9ec108d8ee6836fb61b9e3a9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "inline_equation", + "content": "S_{H} = \\mathbb{E}[|H|^{2}]" + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "inline_equation", + "content": "S_{N_g} = \\mathbb{E}[|N_g|^2]" + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "content": " are the mPSD of blur kernel and noise, respectively, and " + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\bar{\\cdot}" + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "content": " denotes complex conjugation. Eq. (4) can be solved by computing the derivative of " + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\widehat{W}" + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "content": " to become zero. Now, we can calculate the " + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\widehat{H}" + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "inline_equation", + "content": "Y_{m}" + }, + { + "bbox": [ + 130, + 553, + 482, + 604 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 257, + 613, + 481, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 613, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 257, + 613, + 481, + 638 + ], + "type": "interline_equation", + "content": "\\widehat {H} = \\widehat {W} Y _ {m} = \\frac {\\overline {{X}}}{| X | ^ {2} + k} Y _ {m} \\tag {5}", + "image_path": "8ce3b0e7b743994221a776efc9db6e136420db378f6b61a76ce1d0d114168637.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "text", + "content": ", where noise suppression factor " + }, + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "text", + "content": " is the ratio between the mPSD of noise and blur kernel. The final " + }, + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{h}" + }, + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "text", + "content": " can be acquired by applying an inverse Fourier transform" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C.Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 324, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 324, + 213 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 324, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 324, + 213 + ], + "type": "image", + "image_path": "f6bb09889a2e64afe4090400c1031ab64d604c77d0f9250035ecd1c624194f7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 214, + 234, + 224 + ], + "lines": [ + { + "bbox": [ + 222, + 214, + 234, + 224 + ], + "spans": [ + { + "bbox": [ + 222, + 214, + 234, + 224 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 327, + 123, + 480, + 213 + ], + "blocks": [ + { + "bbox": [ + 327, + 123, + 480, + 213 + ], + "lines": [ + { + "bbox": [ + 327, + 123, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 327, + 123, + 480, + 213 + ], + "type": "image", + "image_path": "660d90fedd022a900149f760ac2f2508de8440a8ab91e3913acecb60f6aab7e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 214, + 408, + 224 + ], + "lines": [ + { + "bbox": [ + 396, + 214, + 408, + 224 + ], + "spans": [ + { + "bbox": [ + 396, + 214, + 408, + 224 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "text", + "content": "Fig. 1: Investigation of the working behavior of the proposed kernel estimation function. (a) Visualization of the estimated blur kernel from the Eq. (6), with " + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "inline_equation", + "content": "k = 0.1\\sigma^2" + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "inline_equation", + "content": "10\\sigma^2" + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "inline_equation", + "content": "100\\sigma^2" + }, + { + "bbox": [ + 131, + 233, + 482, + 289 + ], + "type": "text", + "content": ", (b) Estimation error of blur kernel, and visual comparison between the estimated kernels under the optimal condition (solid line) and with those having minimum error values (dashed line)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 310, + 479, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 310, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 310, + 479, + 334 + ], + "type": "text", + "content": "which is followed by absolute and cropping operation, " + }, + { + "bbox": [ + 131, + 310, + 479, + 334 + ], + "type": "inline_equation", + "content": "C:\\mathbb{R}^N\\to \\mathbb{R}^M" + }, + { + "bbox": [ + 131, + 310, + 479, + 334 + ], + "type": "text", + "content": " to Eq. (5). The resultant PSF can be described as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 342, + 480, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 342, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 233, + 342, + 480, + 355 + ], + "type": "interline_equation", + "content": "h = C \\left(\\mathcal {W} \\left(y _ {m}, x, k\\right)\\right) = F \\left(y _ {m}, x, k\\right) \\tag {6}", + "image_path": "28c46b1160ecb1da5682cf323cd11b39124c05be50dffbaffc6a50e9f045971d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "content": " represents inverse filtering of " + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "content": ". Although Eq. (6) is a closed-form solution, the uncertainty of mPSD of noise and blur kernel prevents computation. In the case of the Eq. (2), noise-to-signal ratio " + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "content": " should be carefully selected for successful reconstruction [8, 12]. Similarly, optimal blur kernel estimation requires the appropriate noise suppression factor " + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "content": ". The optimal value for " + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "content": " is derived and analyzed in the following." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 461, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 461, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 131, + 461, + 482, + 486 + ], + "type": "text", + "content": "Claim 2. Given Gaussian noise " + }, + { + "bbox": [ + 131, + 461, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\sim \\mathcal{N}(0, \\sigma^2)" + }, + { + "bbox": [ + 131, + 461, + 482, + 486 + ], + "type": "text", + "content": ", the optimal noise suppression factor is given as " + }, + { + "bbox": [ + 131, + 461, + 482, + 486 + ], + "type": "inline_equation", + "content": "k \\simeq A(\\sigma)^2" + }, + { + "bbox": [ + 131, + 461, + 482, + 486 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 461, + 482, + 486 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 461, + 482, + 486 + ], + "type": "text", + "content": " represents the smoothness of blur kernel." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "text", + "content": "As we have discussed, the noise suppression factor in Eq. (5) is the ratio between mPSD of noise and blur kernel which is analytically derived. According to the Wiener-Khintchine theorem, the power spectral density of a random process can be expressed as the Fourier transform of auto-correlation [20,21]. In the case of Gaussian noise " + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\sim \\mathcal{N}(0,\\sigma^2)" + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "text", + "content": ", the auto-correlation is defined as " + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "inline_equation", + "content": "R_{nn}(\\tau) = \\sigma^2\\delta(\\tau)" + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "text", + "content": " so its Fourier transform, i.e. mPSD, is: " + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "inline_equation", + "content": "S_{N_g} = \\sigma^2" + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "text", + "content": ". Although the mPSD of the blur kernel rarely is known, it represents the smoothness of the blur kernel. Here, we empirically found that " + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "inline_equation", + "content": "A\\simeq 1" + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "text", + "content": " for sparse blur kernel such as motion blur while " + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "inline_equation", + "content": "A\\simeq 100" + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "text", + "content": " for smooth blur kernel such as Gaussian blur." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "text", + "content": "To see how the noise suppression factor affects the reconstruction quality of the blur kernel, we reformulate the Eq. (5) as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 201, + 636, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 636, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 201, + 636, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\widehat {H} = \\frac {\\overline {{X}}}{| X | ^ {2} + k} (H X + N _ {g}) = H + H \\left(\\frac {- k + \\overline {{X}} N _ {g} / H}{| X | ^ {2} + k}\\right) \\tag {7}", + "image_path": "2f57388599ad1b995f87c591ba0fdb5b058695cb7e09245573946ba78e9e307c.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "text", + "content": "Blind image deblurring with noise-robust kernel estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 312, + 232 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 312, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 312, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 312, + 232 + ], + "type": "image", + "image_path": "6fc89e2f65f6225ca74f7a4aabdd4f1455eabbe479055e3383163ef9320fe2c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 233, + 228, + 243 + ], + "lines": [ + { + "bbox": [ + 216, + 233, + 228, + 243 + ], + "spans": [ + { + "bbox": [ + 216, + 233, + 228, + 243 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 318, + 116, + 399, + 228 + ], + "blocks": [ + { + "bbox": [ + 318, + 116, + 399, + 228 + ], + "lines": [ + { + "bbox": [ + 318, + 116, + 399, + 228 + ], + "spans": [ + { + "bbox": [ + 318, + 116, + 399, + 228 + ], + "type": "image", + "image_path": "0dc747ae8ab989fd33035b51fcc46e1c4303c24aaac8cfdc881e230805bc57cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "text", + "content": "Fig. 2: (a) Illustration of the mean squared error between the measured blurry image " + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "text", + "content": " and the estimated blurry image " + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "inline_equation", + "content": "y_{est}" + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "text", + "content": ". (b) Visual comparison of the " + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "inline_equation", + "content": "y_{m} - y_{est}" + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "text", + "content": " for different " + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 252, + 482, + 285 + ], + "type": "text", + "content": ". The inset shows the estimated blur kernel." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 117, + 479, + 227 + ], + "blocks": [ + { + "bbox": [ + 399, + 117, + 479, + 227 + ], + "lines": [ + { + "bbox": [ + 399, + 117, + 479, + 227 + ], + "spans": [ + { + "bbox": [ + 399, + 117, + 479, + 227 + ], + "type": "image", + "image_path": "c9fb8f4e282714bb52c0d9acf4d7587594f06b6cd631e107a68a71577f0e58a2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 233, + 403, + 243 + ], + "lines": [ + { + "bbox": [ + 391, + 233, + 403, + 243 + ], + "spans": [ + { + "bbox": [ + 391, + 233, + 403, + 243 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": "Fig. 1 (a) shows the reconstructed blur kernel for different noise suppression factors. When " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "k = \\sigma^2" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": ", the proposed kernel estimation function accurately reconstructs the kernel even in strong noise. However, when " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": " is small compared to the noise level, i.e. " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "k \\ll \\frac{\\overline{X}N_g}{H}" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": ", Eq. (7) can be expressed as " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\widehat{H} = H + \\frac{N_g}{X}" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": ". As shown in Fig. 1 (a) where " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "k = 0.1\\sigma^2" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": ", the Fourier spectrum of noise normalized by the spectrum of a clear image, i.e. " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\frac{N_g}{X}" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": ", is added to the blur kernel which leads to severe degradation of the reconstructed blur kernel. On the other hand, when " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": " is too large, i.e. " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "k \\gg \\frac{\\overline{X}N_g}{H}" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": ", Eq. (7) can be expressed as " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\widehat{H} = H(1 - \\frac{k}{|X|^2 + k})" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": ". As shown in Fig. 1 (a) where " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "k = 10\\sigma^2" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "k = 100\\sigma^2" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": ", since the high-frequency spectrum of " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": " becomes zero due to the " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\frac{k}{|X|^2 + k} \\approx 1" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": ", the blur kernel exhibits as a low-pass filtered. Also, Fig. 1 (b) demonstrates that the proposed optimal condition closely approximates the minimization of kernel estimation error. The further analysis for " + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 310, + 483, + 479 + ], + "type": "text", + "content": " is described in Supplementary S1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 498, + 417, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 417, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 417, + 511 + ], + "type": "text", + "content": "3.3 Multiple kernel estimation for unknown noise level" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "A classical strategy to solve inverse problems including blind deblurring is minimizing penalized least square (PLS) methods. The data-fidelity term in PLS formulation of the blind deblurring problem is: " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{argmin}_{x,h}\\| y_m - x*h\\| _2^2" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". In our cases, as " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " can be estimated through " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", the data-fidelity term can be expressed as " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{argmin}_x\\| y_m - x*F(y_m,x,k)\\| _2^2" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", which requires setting of " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " for optimization. However, because we cannot accurately determine the noise level, setting the optimal " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " is extremely challenging. Nevertheless, since " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " is a closed-form inverse mapping function, which minimizes Eq. (3), deviating from the optimal condition of " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " results in increased noise or blurriness rather than leading to incorrect solutions. In other words, even if the " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " value is not strictly set, as long as it falls within a certain boundary, the data-fidelity term can be minimized. We verified this by examining the mean squared error (MSE) between " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C.Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 190, + 152 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 152 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 152 + ], + "type": "image", + "image_path": "5d42bcfe8ba0906ea41c9a6da6b2048be1f3bfee9e19c82ca1ca6586cee64dbc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 153, + 177, + 163 + ], + "lines": [ + { + "bbox": [ + 146, + 153, + 177, + 163 + ], + "spans": [ + { + "bbox": [ + 146, + 153, + 177, + 163 + ], + "type": "text", + "content": "(a) Blur" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 249, + 152 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 249, + 152 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 249, + 152 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 249, + 152 + ], + "type": "image", + "image_path": "8c30511b65de6bed8ecba639631a5f784b4a2b2b2535c11dd22c2073bbdcc678.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 153, + 233, + 163 + ], + "lines": [ + { + "bbox": [ + 206, + 153, + 233, + 163 + ], + "spans": [ + { + "bbox": [ + 206, + 153, + 233, + 163 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "lines": [ + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "type": "text", + "content": "Fig. 3: Visualization of the blind deblurring results using the multiple kernel estimation scheme. (a) Blur image degraded with Gaussian noise level of " + }, + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "type": "inline_equation", + "content": "\\sigma = 20" + }, + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "type": "text", + "content": ", (b) ground truth, (c) reconstructed clear image, and estimated intermediate image and blur kernels with (d) " + }, + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "type": "inline_equation", + "content": "k = 5^2" + }, + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "type": "text", + "content": ", (e) " + }, + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "type": "inline_equation", + "content": "k = 15^2" + }, + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "type": "text", + "content": ", and (f) " + }, + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "type": "inline_equation", + "content": "k = 25^2" + }, + { + "bbox": [ + 131, + 171, + 480, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 306, + 152 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 306, + 152 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 306, + 152 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 306, + 152 + ], + "type": "image", + "image_path": "be1c66ed0e06eb4379033d92495a8377b3849677013cb888052a5163d061d748.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 267, + 153, + 289, + 163 + ], + "lines": [ + { + "bbox": [ + 267, + 153, + 289, + 163 + ], + "spans": [ + { + "bbox": [ + 267, + 153, + 289, + 163 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 267, + 153, + 289, + 163 + ], + "type": "inline_equation", + "content": "x_0" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 365, + 152 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 365, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 365, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 365, + 152 + ], + "type": "image", + "image_path": "2ba296c635461c04756535dbb7c37414be49ab62f0b1ac4bc96486a903010545.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 153, + 360, + 163 + ], + "lines": [ + { + "bbox": [ + 312, + 153, + 360, + 163 + ], + "spans": [ + { + "bbox": [ + 312, + 153, + 360, + 163 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 312, + 153, + 360, + 163 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 312, + 153, + 360, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 153, + 360, + 163 + ], + "type": "inline_equation", + "content": "h_1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 422, + 152 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 422, + 152 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 422, + 152 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 422, + 152 + ], + "type": "image", + "image_path": "edd4a63306f7f6565a0c46316bd60858c6f1e508cc4d56eba9b2f987f866c530.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 153, + 419, + 163 + ], + "lines": [ + { + "bbox": [ + 370, + 153, + 419, + 163 + ], + "spans": [ + { + "bbox": [ + 370, + 153, + 419, + 163 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 370, + 153, + 419, + 163 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 370, + 153, + 419, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 370, + 153, + 419, + 163 + ], + "type": "inline_equation", + "content": "h_2" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 114, + 481, + 152 + ], + "blocks": [ + { + "bbox": [ + 425, + 114, + 481, + 152 + ], + "lines": [ + { + "bbox": [ + 425, + 114, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 425, + 114, + 481, + 152 + ], + "type": "image", + "image_path": "67c8c496a96f29d32c1231f254587aa912621414330e5f49c51644bf34910e42.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 153, + 476, + 163 + ], + "lines": [ + { + "bbox": [ + 428, + 153, + 476, + 163 + ], + "spans": [ + { + "bbox": [ + 428, + 153, + 476, + 163 + ], + "type": "text", + "content": "(f) " + }, + { + "bbox": [ + 428, + 153, + 476, + 163 + ], + "type": "inline_equation", + "content": "x_{3}" + }, + { + "bbox": [ + 428, + 153, + 476, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 428, + 153, + 476, + 163 + ], + "type": "inline_equation", + "content": "h_3" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "y_{est} = x * F(y_m, x, k)" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": " by changing " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": ". As shown in Fig. 2 (a), the MSE decreases as " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": " approaches the optimal condition. Also, the lower error value is maintained over a large range of " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": ", and there exists a range around the optimal suppression factor where " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "|y_m - x * F(y_m, x, k)| \\leq \\varepsilon" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": " even if " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": " is not optimally estimated (Fig. 2 (a) red box). Fig. 2 (b) further demonstrates this by visually comparing " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "y_{est}" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "F(y_m, x, k)" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": ", and differential map " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "y_m - y_{est}" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": " for different " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": " values when noise level is " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "\\sigma = 10" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": ". Based on this observation, we propose a multiple kernel estimation scheme that handles a wide range of unknown noise levels by setting multiple values for " + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 240, + 480, + 347 + ], + "type": "text", + "content": ", rather than a specific value as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 229, + 357, + 480, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 357, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 229, + 357, + 480, + 370 + ], + "type": "interline_equation", + "content": "h _ {i} = F \\left(y _ {m}, x _ {i}, k _ {i}\\right), \\quad \\forall i = 1, \\dots , N _ {k} \\tag {8}", + "image_path": "861440a22ea9785f6e2b7194276090a9b7fedd9a39d9b01011ac4e48a7c4ef64.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "content": " is the estimated blur kernel given intermediate image " + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "content": " is noise suppression factors corresponding to different noise levels. The fundamental concept of this framework lies in managing unknown noise levels by setting " + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "content": " from low to high, thus covering a broad range of noise that exists in nature. Here, we also introduce intermediate images, which are separately estimated from a clear image, to compensate for the error term in Eq. (7) caused by a suboptimal setting of " + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "content": ", thereby stabilizing the optimization process. Fig. 3 shows an example of deblurred image " + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "content": ", estimated intermediate images " + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "content": " and blur kernels " + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 131, + 380, + 482, + 536 + ], + "type": "text", + "content": ". A detailed analysis for intermediate images and the performance of multiple kernel estimation scheme, compared to case where noise level is precisely known, are presented in Supplementary S2. We demonstrate that the proposed multiple kernel estimation scheme exhibits deblurring performance comparable to cases where the noise level is precisely known, and see for the detailed analysis." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 553, + 340, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 553, + 340, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 553, + 340, + 565 + ], + "type": "text", + "content": "3.4 MAP approach for blind deblurring" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 574, + 477, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 477, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 477, + 586 + ], + "type": "text", + "content": "Conventionally, the MAP approach for blind deblurring is described as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 242, + 596, + 480, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 596, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 242, + 596, + 480, + 609 + ], + "type": "interline_equation", + "content": "p (x, h | y) \\propto p (y | x, h) p (x) p (h) \\tag {9}", + "image_path": "dbc3a3236c30c2afc3bf61cc9340e91022bbc30bf9dfde330531912780d50384.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 617, + 480, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 617, + 480, + 653 + ], + "type": "inline_equation", + "content": "p(y|x,h)" + }, + { + "bbox": [ + 130, + 617, + 480, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 617, + 480, + 653 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 130, + 617, + 480, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 617, + 480, + 653 + ], + "type": "inline_equation", + "content": "p(h)" + }, + { + "bbox": [ + 130, + 617, + 480, + 653 + ], + "type": "text", + "content": " correspond to data-fidelity term, clear image prior, and blur kernel prior, respectively. Many previous works carefully designed the priors to strictly confine the feasible solution space." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 146, + 653, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 653, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 146, + 653, + 480, + 665 + ], + "type": "text", + "content": "In this work, we show that leveraging the capabilities of DIP allows solving" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "text", + "content": "Blind image deblurring with noise-robust kernel estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 481, + 219 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 481, + 219 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 481, + 219 + ], + "type": "image", + "image_path": "292ba58fcfe7b1722cb91025744194da9cbf7f0bc7015cfd4686c93fed563122.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "type": "text", + "content": "Fig. 4: Workflow of the proposed method. A single deep neural network generates the deblurred image " + }, + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "type": "text", + "content": ", and the intermediate images " + }, + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "type": "text", + "content": " which are used to estimate blur kernel " + }, + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "type": "text", + "content": " with the noise suppression factor " + }, + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 130, + 226, + 482, + 261 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 280, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 479, + 326 + ], + "type": "text", + "content": "the blind deblurring problem by estimating solely a clear image even in the presence of noise. Specifically, as blur kernel can be analytically obtained from Eq. (6) if clear image " + }, + { + "bbox": [ + 130, + 280, + 479, + 326 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 280, + 479, + 326 + ], + "type": "text", + "content": " and blurry image " + }, + { + "bbox": [ + 130, + 280, + 479, + 326 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 130, + 280, + 479, + 326 + ], + "type": "text", + "content": " are given, Eq. (9) can be reformulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 219, + 328, + 481, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 328, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 219, + 328, + 481, + 341 + ], + "type": "interline_equation", + "content": "p (x, h | y) = p (h | x, y) p (x | y) \\propto p (y | x) p (x) \\tag {10}", + "image_path": "3e9fba8382e49d45580929420aa865e14018a97a14e114d44fb4804d0136c70b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 344, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 404 + ], + "type": "text", + "content": "As shown in Eq. (10), it is evident that the blind deblurring problem can be addressed by solely focusing on estimating a clear image, while also effectively handling noise owing to the robustness of the proposed kernel estimation function. Therefore, we aim to utilize DIP, one of the optimal methods for estimating natural images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 447, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 447, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 447, + 442 + ], + "type": "text", + "content": "3.5 Blind deblurring based on noise-robust kernel estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 447, + 424, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 424, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 424, + 459 + ], + "type": "text", + "content": "We formulate the noise-robust blind deblurring method as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 204, + 464, + 391, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 464, + 391, + 495 + ], + "spans": [ + { + "bbox": [ + 204, + 464, + 391, + 495 + ], + "type": "interline_equation", + "content": "\\min _ {x _ {0}, h _ {1}, \\dots , h _ {N _ {k}}} \\frac {1}{N _ {k}} \\sum_ {i = 1} ^ {N _ {k}} | | y _ {m} - x _ {0} * h _ {i} | | _ {2} ^ {2} + \\lambda R (x _ {0})", + "image_path": "a504df519702a0b0ffb7f6c1f5787e86191791b6f96338d9b05539392355566c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 204, + 496, + 480, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 496, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 204, + 496, + 480, + 511 + ], + "type": "interline_equation", + "content": "s. t. x _ {0} = g _ {\\phi_ {0}} \\left(G _ {\\theta} (z)\\right), h _ {i} = F \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) \\tag {11}", + "image_path": "24c65801513d4136644c717bf91bb8a535bc65f9552b79db5cccaeab18bc4db6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 204, + 514, + 360, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 514, + 360, + 540 + ], + "spans": [ + { + "bbox": [ + 204, + 514, + 360, + 540 + ], + "type": "interline_equation", + "content": "h _ {i} ^ {j} \\geq 0, \\sum_ {j} h _ {i} ^ {j} = 1 \\quad \\forall j = 0, 1, \\dots , M", + "image_path": "437647f4a326d76af636859f51ee521765e462e676f9819a43595da91aae92e2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": " is fixed tensor randomly sampled from uniform distribution, " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "(\\cdot)^j" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": "-th elements, " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "g_{\\phi_i}" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": "-th group convolution filter convolved with the last feature layer of " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": " is regularization function for " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": " is regularization weight. Although we adopt total variation for " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": ", other regularization functions such as " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "l_{0}" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": "-sparsity or " + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": " norm can be used." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "text", + "content": "As shown in Eq. (11), the blind deblurring problem can be solved by focusing on estimating " + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "text", + "content": ". Also, it should be noted that the estimation of " + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "text", + "content": " can be achieved with a single neural network as " + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "text", + "content": " should have a similar appearance with " + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 605, + 481, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C.Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 134, + 130, + 481, + 322 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 128 + ], + "type": "text", + "content": "Algorithm 1 Blind deblurring with the noise-robust kernel estimation function" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "lines": [ + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": "1: Input: Blurred image " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": ", pre-defined " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\epsilon, \\lambda_{1}, \\lambda_{2}, k_{1}, k_{2}, \\dots, k_{N_{k}}" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " \n2: Output: clear image " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " and estimated blur kernels " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "h_{1}, h_{2}, \\dots, h_{N_{k}}" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " \n3: Sample " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " from uniform distribution \n4: for " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "N_{iter}" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " do \n5: " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "[f_{0}, f_{1}, \\dots, f_{N_{k}}] = G_{\\theta}(z)" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "x_{0} = g_{\\phi_{0}}(f_{0})" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tot} = 0" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " \n8: for " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " do \n9: " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "h_{i} = F(y_{m}, g_{\\phi_{i}}(f_{i}), k_{i})" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "h_{i} = |h_{i}|" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " Non-negativity \n11: " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "h_{i} = h_{i} / (\\sum h_{i} + \\epsilon)" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " Equiltity \n12: " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tot} = \\mathcal{L}_{tot} + \\frac{1}{N_{k}} \\left( ||y_{m} - x_{0} * h_{i}||_{2}^{2} + \\lambda_{1} ||W(y_{m}, g_{\\phi_{i}}(f_{i}), k_{i})||_{2}^{2} \\right)" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " \n13: end for \n14: " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tot} = \\mathcal{L}_{tot} + \\lambda_{2} TV(x_{0})" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " \n15: Update " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\{\\theta, \\phi\\}" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta, \\phi} \\mathcal{L}_{tot}" + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "text", + "content": " \n16: end for \n17: return " + }, + { + "bbox": [ + 134, + 130, + 481, + 322 + ], + "type": "inline_equation", + "content": "x_{0}, h_{1}, h_{2}, \\dots, h_{N_{k}}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 354, + 350, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 350, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 350, + 365 + ], + "type": "text", + "content": "3.6 Optimization of the proposed method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 394, + 480, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 480, + 419 + ], + "type": "text", + "content": "The workflow of the proposed method is illustrated in Fig. 4 and the network parameters " + }, + { + "bbox": [ + 131, + 394, + 480, + 419 + ], + "type": "inline_equation", + "content": "\\theta, \\Phi = \\{\\phi_0, \\dots, \\phi_{N_k}\\}" + }, + { + "bbox": [ + 131, + 394, + 480, + 419 + ], + "type": "text", + "content": " are optimized by the following cost function:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 144, + 437, + 480, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 437, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 144, + 437, + 480, + 529 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta , \\phi_ {0}, \\dots \\phi_ {N _ {k}}} \\frac {1}{N _ {k}} \\sum_ {i = 1} ^ {N _ {k}} \\left(\\left| \\left| y _ {m} - g _ {\\phi_ {0}} \\left(G _ {\\theta} (z)\\right) * F \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) \\right| \\right| _ {2} ^ {2} \\right. \\\\ + \\lambda_ {1} \\| \\mathcal {W} \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) \\| _ {2} ^ {2}) + \\lambda_ {2} T V \\left(g _ {\\phi_ {0}} \\left(G _ {\\theta} (z)\\right)\\right) \\tag {12} \\\\ s. t. F \\left(y _ {m}, g _ {\\phi_ {i}} \\left(G _ {\\theta} (z)\\right), k _ {i}\\right) ^ {J} \\geq 0, \\\\ \\sum_ {j} F (y _ {m}, g _ {\\phi_ {i}} (G _ {\\theta} (z)), k _ {i}) ^ {j} = 1 \\quad \\forall j = 0, 1, \\dots , M \\\\ \\end{array}", + "image_path": "dc412ad35fd2e3d41385ea25cebcb1ca776d218e2cfbd77a186a19fb1828e98c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "inline_equation", + "content": "TV" + }, + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "text", + "content": " is total variation, and " + }, + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "text", + "content": " are regularization weights. The nonnegativity and equality constraints to the estimated blur kernel are induced by taking an absolute operation followed by dividing each element by the sum of all elements. Also, we additionally introduce a sparsity regularization term for the inverse filtered image, " + }, + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(y_m, g_{\\phi_i}(G_\\theta(z)), k_i)" + }, + { + "bbox": [ + 130, + 546, + 481, + 654 + ], + "type": "text", + "content": ", as all pixels except for the blur kernel positioned should have a value of 0. It is important to emphasize that our proposed multiple kernel estimation scheme, rather than regularization terms, is the key enabler for robust deblurring (see the ablation study in Supplementary S7). The optimization of Eq. (12) is formally summarized in Algorithm 1." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "Blind image deblurring with noise-robust kernel estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 190, + 171 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 171 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 171 + ], + "type": "image", + "image_path": "13de169d7d55466941e68403acc4b420e99f8c39dffd00ed33a7e3748cb09b30.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 171, + 190, + 227 + ], + "blocks": [ + { + "bbox": [ + 133, + 171, + 190, + 227 + ], + "lines": [ + { + "bbox": [ + 133, + 171, + 190, + 227 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 190, + 227 + ], + "type": "image", + "image_path": "88385f9821f2b47f409e11b8229b4758bf63cc6d0edcc20011064e35db71c31d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 228, + 177, + 238 + ], + "lines": [ + { + "bbox": [ + 146, + 228, + 177, + 238 + ], + "spans": [ + { + "bbox": [ + 146, + 228, + 177, + 238 + ], + "type": "text", + "content": "(a) Blur" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 249, + 171 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 249, + 171 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 249, + 171 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 249, + 171 + ], + "type": "image", + "image_path": "180ab968c1be7815ea047dcfce4853e8cd56b45773b9fd785aac7dfaadcfecbf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 192, + 171, + 249, + 227 + ], + "blocks": [ + { + "bbox": [ + 192, + 171, + 249, + 227 + ], + "lines": [ + { + "bbox": [ + 192, + 171, + 249, + 227 + ], + "spans": [ + { + "bbox": [ + 192, + 171, + 249, + 227 + ], + "type": "image", + "image_path": "83def309de633c2f0e302d10fa3f3fe04f7998ddbb4e533c1f2ae4ed796363ef.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 228, + 233, + 238 + ], + "lines": [ + { + "bbox": [ + 206, + 228, + 233, + 238 + ], + "spans": [ + { + "bbox": [ + 206, + 228, + 233, + 238 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 306, + 171 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 306, + 171 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 306, + 171 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 306, + 171 + ], + "type": "image", + "image_path": "99c7b6449e71b28cbfabda034f4fc405b77816b39faef4dd85044b436ebb790c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 250, + 171, + 306, + 227 + ], + "blocks": [ + { + "bbox": [ + 250, + 171, + 306, + 227 + ], + "lines": [ + { + "bbox": [ + 250, + 171, + 306, + 227 + ], + "spans": [ + { + "bbox": [ + 250, + 171, + 306, + 227 + ], + "type": "image", + "image_path": "3f298cdff34af1aa35e81681c2da16938bea4db79c346e4d9257dbb4a7475fbe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 228, + 293, + 238 + ], + "lines": [ + { + "bbox": [ + 263, + 228, + 293, + 238 + ], + "spans": [ + { + "bbox": [ + 263, + 228, + 293, + 238 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 263, + 228, + 293, + 238 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 365, + 171 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 365, + 171 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 365, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 365, + 171 + ], + "type": "image", + "image_path": "346d94c32be2260eebcbc040721c37d490b172cbcf48853b63348fe2d77a25e4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 171, + 365, + 227 + ], + "blocks": [ + { + "bbox": [ + 307, + 171, + 365, + 227 + ], + "lines": [ + { + "bbox": [ + 307, + 171, + 365, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 365, + 227 + ], + "type": "image", + "image_path": "6ff3dce2b1d5c01ac5d88320064970b025abc774fc5e91410c10f5cc34430b79.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 228, + 358, + 238 + ], + "lines": [ + { + "bbox": [ + 314, + 228, + 358, + 238 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 358, + 238 + ], + "type": "text", + "content": "(d) Pure-let" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison of blind deblurring results on AFHQ datasets [6]. (row 1): AFHQ-dog motion deblurring with a kernel size of " + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "inline_equation", + "content": "43 \\times 43" + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "text", + "content": " and noise level of " + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\sigma = 10" + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "text", + "content": ", (row 2): AFHQ-cat motion deblurring with a kernel size of " + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "inline_equation", + "content": "39 \\times 39" + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "text", + "content": " and noise level of " + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\sigma = 20" + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "text", + "content": ". Here, the estimated kernel for " + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "inline_equation", + "content": "k_{2} = 15^{2}" + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "text", + "content": " is presented in (f) while others are presented in Supplementary S5." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 422, + 171 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 422, + 171 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 422, + 171 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 422, + 171 + ], + "type": "image", + "image_path": "1345ee935c3bc39125403dfbefc00df9d4497c141548aaf2c8181063c96befd9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 366, + 171, + 422, + 227 + ], + "blocks": [ + { + "bbox": [ + 366, + 171, + 422, + 227 + ], + "lines": [ + { + "bbox": [ + 366, + 171, + 422, + 227 + ], + "spans": [ + { + "bbox": [ + 366, + 171, + 422, + 227 + ], + "type": "image", + "image_path": "b98692ca6737de00134b898ec4498bbf3db969247ddb874f0721680952063187.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 228, + 413, + 238 + ], + "lines": [ + { + "bbox": [ + 376, + 228, + 413, + 238 + ], + "spans": [ + { + "bbox": [ + 376, + 228, + 413, + 238 + ], + "type": "text", + "content": "(e)WDIP" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 425, + 114, + 481, + 171 + ], + "blocks": [ + { + "bbox": [ + 425, + 114, + 481, + 171 + ], + "lines": [ + { + "bbox": [ + 425, + 114, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 425, + 114, + 481, + 171 + ], + "type": "image", + "image_path": "26c42b85b1e385e20935519c305803366a4c244debfb8e9db382d277b802f6ff.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 425, + 171, + 481, + 227 + ], + "blocks": [ + { + "bbox": [ + 425, + 171, + 481, + 227 + ], + "lines": [ + { + "bbox": [ + 425, + 171, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 425, + 171, + 481, + 227 + ], + "type": "image", + "image_path": "f421912ad5d077e38ac54e32774b4721262fa0edc3dd37a79bfb2f11b68f8a5a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 228, + 468, + 238 + ], + "lines": [ + { + "bbox": [ + 437, + 228, + 468, + 238 + ], + "spans": [ + { + "bbox": [ + 437, + 228, + 468, + 238 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 324, + 229, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 229, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 229, + 337 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 348, + 326, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 326, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 326, + 361 + ], + "type": "text", + "content": "4.1 Experiments on simulated image" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": "Motion deblur on AFHQ dataset [6]. To demonstrate the deblurring performance on motion blur, we conduct experiments on AFHQ-dog " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " and AFHQ-cat " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " datasets. Specifically, we select 500 test images as an evaluation dataset. To synthesize motion-blurred images, motion-blur kernels of varying sizes, ranging from " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "21 \\times 21" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "51 \\times 51" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": ", are randomly generated [7], and each kernel is then convolved with a clear image. Finally, Gaussian noise " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\sigma = 10" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\sigma = 20" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " are added to the synthesized blurry images." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "The proposed method (refer Supplementary S3 and S6 for implementation detail) is compared with other state-of-the-art blind deblurring methods including Sanghvi et al. [35], Xu et al. [40], Pure-let [24], Selfdeblur [33], WDIP [3], " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " [18], Pan-DCP [30], Pan-" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_{0}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " [29], and Liang et al. [4], and see Supplementary S4 for implementation details. Note that we exclusively choose blind deblurring methods that do not utilize any training dataset to ensure a fair comparison. When optimizing the proposed method, " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "k_{1} = 5^{2}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "k_{2} = 15^{2}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "k_{3} = 25^{2}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " are set to accommodate a wide range of unknown noise levels. The representative blind deblurring results from " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", Pure-let, WDIP, and the proposed method are qualitatively compared in Fig. 5. Due to the severe degradation caused by Gaussian noise in blurred images, the deblurring results obtained from the compared methods are greatly compromised. Even with laborious adjustment of hyperparameters, these methods tend to overfit strongly to the noise. In contrast, the deblurring results from the proposed method shown in Fig. 5 (f) demonstrate noise-robust deblurring by effectively reconstructing clear images and blur kernels simultaneously even in the absence of precise knowledge about the noise level. Furthermore, quantitative comparison presented in Tab. 1 indicates that the proposed method achieves state-of-the-art performance in terms of Peak" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "C.Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 146, + 481, + 312 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of blind deblurring results on AFHQ datasets [6] with different noise levels (" + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "inline_equation", + "content": "\\sigma = 10" + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "inline_equation", + "content": "\\sigma = 20" + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "). **Bold:** best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 146, + 481, + 312 + ], + "lines": [ + { + "bbox": [ + 133, + 146, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 481, + 312 + ], + "type": "table", + "html": "
MethodAFHQ-dogAFHQ-cat
σ = 10σ = 20σ = 10σ = 20
PSNR↑ SSIM↑ PSNR↑ SSIM↑ PSNR↑ SSIM↑ PSNR↑ SSIM
Proposed24.230.63222.900.53923.330.59022.110.497
SelfDeblur [33]13.540.10712.020.06312.830.08011.440.046
WDIP [3]14.110.11512.430.06613.430.08611.900.049
l1/l2 [18]19.690.33516.030.14619.360.30815.920.140
Pan-DCP [30]10.240.0607.6980.02410.170.0537.7130.022
Pan-l0 [29]8.3110.0316.8360.0168.2620.0296.9220.016
Liang et al. [4]10.680.0617.7110.02310.480.0577.6940.022
Sanghvi et al. [35]16.840.52016.840.52216.260.46516.380.471
Xu et al. [40]16.510.39815.010.29515.940.35714.340.258
Pure-let [24]21.020.53419.940.47720.780.48219.790.429
", + "image_path": "85516dfeb3245e4e450a4a6beda993752d70c825abf4462db3c9f0880372331d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "type": "text", + "content": "Signal-to-noise ratio (PSNR) and Structural Similarity Index Measure (SSIM). See Supplementary S5 for additional qualitative comparisons of deblurring results and quantitative comparisons at low-noise levels (" + }, + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "type": "inline_equation", + "content": "\\sigma = 2" + }, + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "type": "inline_equation", + "content": "\\sigma = 5" + }, + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 485 + ], + "type": "text", + "content": "Sun et al. dataset [38]. To further showcase the capability of the proposed method, we utilize the Sun et al. dataset, comprising 80 natural images and 8 blur kernels, resulting in the generation of 640 blurry images. The representative blind deblurring results are presented in Fig. 6. The deblurring results obtained from compared methods - " + }, + { + "bbox": [ + 130, + 388, + 480, + 485 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + }, + { + "bbox": [ + 130, + 388, + 480, + 485 + ], + "type": "text", + "content": ", Pan-DCP, and WDIP - are catastrophically corrupted when noise increases. Conversely, the proposed method shows superior deblurring results regardless of the noise level. Also, as shown in Tab. 2, the proposed method establishes the state-of-the-art deblurring performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 502, + 385, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 385, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 385, + 514 + ], + "type": "text", + "content": "4.2 Experiments on scanning electron microscope" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Scanning electron microscopes (SEM) provide sub-nanometer resolution images by detecting the back-scattered or emitted electrons from the sample. However, the incident electrons, which determine the shape of the PSF of SEM can be easily distorted when the electron passes through several lenses and coils, see Fig. 7 (a). Additionally, random fluctuations of the incident electrons cause Gaussian noise, significantly compromising image quality, see Fig. 8 (a). Correcting such degradation often requires manual adjustment of system parameters, such as working distance or stigmator, by experts. To automate this correction process image quality assessment based method [17,25,26], bayesian optimization based approach [2,34], and feedback-loop based deep learning approach [22,27,36] have been proposed. Nevertheless, previous methods necessitate iterative measurement of SEM images to determine correction parameters, rendering the process" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "Blind image deblurring with noise-robust kernel estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 190, + 235 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 235 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 235 + ], + "type": "image", + "image_path": "b417bb399ab85efeaf6b76735f329ba2e7fdc38e425e15a5485ecf7e4ebca35a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 236, + 177, + 245 + ], + "lines": [ + { + "bbox": [ + 146, + 236, + 177, + 245 + ], + "spans": [ + { + "bbox": [ + 146, + 236, + 177, + 245 + ], + "type": "text", + "content": "(a) Blur" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 115, + 249, + 236 + ], + "blocks": [ + { + "bbox": [ + 192, + 115, + 249, + 236 + ], + "lines": [ + { + "bbox": [ + 192, + 115, + 249, + 236 + ], + "spans": [ + { + "bbox": [ + 192, + 115, + 249, + 236 + ], + "type": "image", + "image_path": "15c876766e8aa1963570ead27b4dad9613941ef3c8db283ad3a3813aa714a668.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 236, + 233, + 245 + ], + "lines": [ + { + "bbox": [ + 206, + 236, + 233, + 245 + ], + "spans": [ + { + "bbox": [ + 206, + 236, + 233, + 245 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 250, + 115, + 306, + 235 + ], + "blocks": [ + { + "bbox": [ + 250, + 115, + 306, + 235 + ], + "lines": [ + { + "bbox": [ + 250, + 115, + 306, + 235 + ], + "spans": [ + { + "bbox": [ + 250, + 115, + 306, + 235 + ], + "type": "image", + "image_path": "7231a7f23e9842ca1fc11955b4e1acdefeb5697a83861bda5cb9cea164059eea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 236, + 293, + 245 + ], + "lines": [ + { + "bbox": [ + 263, + 236, + 293, + 245 + ], + "spans": [ + { + "bbox": [ + 263, + 236, + 293, + 245 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 263, + 236, + 293, + 245 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 115, + 364, + 236 + ], + "blocks": [ + { + "bbox": [ + 307, + 115, + 364, + 236 + ], + "lines": [ + { + "bbox": [ + 307, + 115, + 364, + 236 + ], + "spans": [ + { + "bbox": [ + 307, + 115, + 364, + 236 + ], + "type": "image", + "image_path": "1c1472252408fed9d82c9d538ec2db7940835833975737b9ac9204263ae430cb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 236, + 360, + 245 + ], + "lines": [ + { + "bbox": [ + 312, + 236, + 360, + 245 + ], + "spans": [ + { + "bbox": [ + 312, + 236, + 360, + 245 + ], + "type": "text", + "content": "(d) Pan-DCP" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 255, + 479, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 288 + ], + "type": "text", + "content": "Fig.6: Qualitative comparison of blind deblurring results on Sun et al. datasets [38] with Gaussian noise level of (row 1): " + }, + { + "bbox": [ + 130, + 255, + 479, + 288 + ], + "type": "inline_equation", + "content": "\\sigma = 10" + }, + { + "bbox": [ + 130, + 255, + 479, + 288 + ], + "type": "text", + "content": " and (row 2): " + }, + { + "bbox": [ + 130, + 255, + 479, + 288 + ], + "type": "inline_equation", + "content": "\\sigma = 20" + }, + { + "bbox": [ + 130, + 255, + 479, + 288 + ], + "type": "text", + "content": ". Here, the estimated kernel for " + }, + { + "bbox": [ + 130, + 255, + 479, + 288 + ], + "type": "inline_equation", + "content": "k_{2} = 15^{2}" + }, + { + "bbox": [ + 130, + 255, + 479, + 288 + ], + "type": "text", + "content": " is presented in (f) while others are presented in Supplementary S5." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 365, + 115, + 422, + 236 + ], + "blocks": [ + { + "bbox": [ + 365, + 115, + 422, + 236 + ], + "lines": [ + { + "bbox": [ + 365, + 115, + 422, + 236 + ], + "spans": [ + { + "bbox": [ + 365, + 115, + 422, + 236 + ], + "type": "image", + "image_path": "4230ca7878eb575dfe41b75ce757992611f7a97bfbf5f90369b3a2625e8329f3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 236, + 413, + 245 + ], + "lines": [ + { + "bbox": [ + 376, + 236, + 413, + 245 + ], + "spans": [ + { + "bbox": [ + 376, + 236, + 413, + 245 + ], + "type": "text", + "content": "(e)WDIP" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 115, + 481, + 236 + ], + "blocks": [ + { + "bbox": [ + 425, + 115, + 481, + 236 + ], + "lines": [ + { + "bbox": [ + 425, + 115, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 425, + 115, + 481, + 236 + ], + "type": "image", + "image_path": "54bd4a5fc665c271a8d67344cda2c077413179f941ac81f3a2298f067b1bcafe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 236, + 468, + 245 + ], + "lines": [ + { + "bbox": [ + 437, + 236, + 468, + 245 + ], + "spans": [ + { + "bbox": [ + 437, + 236, + 468, + 245 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 205, + 334, + 405, + 472 + ], + "blocks": [ + { + "bbox": [ + 130, + 300, + 479, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 479, + 323 + ], + "type": "text", + "content": "Table 2: Quantitative comparison of blind deblurring results on Sun et al. datasets [38] with different noise levels (" + }, + { + "bbox": [ + 130, + 300, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\sigma = 10" + }, + { + "bbox": [ + 130, + 300, + 479, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 300, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\sigma = 20" + }, + { + "bbox": [ + 130, + 300, + 479, + 323 + ], + "type": "text", + "content": "). Bold: best." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 205, + 334, + 405, + 472 + ], + "lines": [ + { + "bbox": [ + 205, + 334, + 405, + 472 + ], + "spans": [ + { + "bbox": [ + 205, + 334, + 405, + 472 + ], + "type": "table", + "html": "
Methodσ = 10σ = 20
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
Proposed25.940.67724.830.608
SelfDeblur [33]17.120.20215.440.136
WDIP [3]17.600.21412.890.103
l1/l2 [18]18.870.17314.970.073
Pan-DCP [30]11.620.0598.3210.021
Pan-l0 [29]9.5900.0308.0650.017
Liang et al. [4]11.960.0608.5220.022
Xu et al. [40]14.100.20713.450.128
Pure-let [24]18.870.25417.310.121
", + "image_path": "28e55c8e9e4dbe248f80731cd3d185a02d6ec6032cd22f54941fa46ef955c015.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 497, + 301, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 301, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 301, + 509 + ], + "type": "text", + "content": "time-consuming and energy-inefficient." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "content": "Here, we demonstrate the effectiveness of the proposed blind deblurring method in correcting degraded SEM images. In general, well-focused PSF of SEM can be approximated to Gaussian shape [32, 43], we set " + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "A = 100" + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "content": ". Additionally, since the exact noise level of SEM images is rarely known, we set the noise suppression factor in a broad range, such as " + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "k_{1} = 10000" + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "k_{2} = 20000" + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "k_{3} = 30000" + }, + { + "bbox": [ + 130, + 521, + 481, + 666 + ], + "type": "text", + "content": ". Fig. 7 and Fig. 8 presents visual comparisons of blind deblurring results. Although methods such as Pan-DCP or WDIP have successfully performed deblurring on images with weak noise and simple patterns, their performance noticeably deteriorates in cases of strong noise or complex patterns. In contrast, as shown in Fig. 7 (f) and Fig. 8 (f), the proposed method consistently reconstructs both the clear image and the PSF, regardless of the complexity of the image or the level of noise. Especially, as shown in the second row of Fig. 7, when" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "C.Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 191, + 228 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 191, + 228 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 191, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 191, + 228 + ], + "type": "image", + "image_path": "a8a32ea0ed20d00c1edbb9bf61b87552c778ee5a8c3858d5d82562750d4b858e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 229, + 177, + 239 + ], + "lines": [ + { + "bbox": [ + 146, + 229, + 177, + 239 + ], + "spans": [ + { + "bbox": [ + 146, + 229, + 177, + 239 + ], + "type": "text", + "content": "(a) Blur" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 249, + 228 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 249, + 228 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 249, + 228 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 249, + 228 + ], + "type": "image", + "image_path": "f3eabd133a138ac205cb3f505c51e05379a239d028653354bf142678e97b7a7b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 229, + 233, + 239 + ], + "lines": [ + { + "bbox": [ + 206, + 229, + 233, + 239 + ], + "spans": [ + { + "bbox": [ + 206, + 229, + 233, + 239 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 306, + 228 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 306, + 228 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 306, + 228 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 306, + 228 + ], + "type": "image", + "image_path": "a284f85bb5c52af4d3d1f556318e7a030b72dd9f1e2057610ca66cec4bf91c2c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 229, + 293, + 239 + ], + "lines": [ + { + "bbox": [ + 263, + 229, + 293, + 239 + ], + "spans": [ + { + "bbox": [ + 263, + 229, + 293, + 239 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 263, + 229, + 293, + 239 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 365, + 228 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 365, + 228 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 365, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 365, + 228 + ], + "type": "image", + "image_path": "f55ea2035b67b5d3467f2e5996b71ef014dc4b497740d3d5738416be2451e694.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 229, + 360, + 239 + ], + "lines": [ + { + "bbox": [ + 312, + 229, + 360, + 239 + ], + "spans": [ + { + "bbox": [ + 312, + 229, + 360, + 239 + ], + "type": "text", + "content": "(d) Pan-DCP" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 422, + 228 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 422, + 228 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 422, + 228 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 422, + 228 + ], + "type": "image", + "image_path": "a491bd3e03431584042738a7009bf306a40eee73c1a7a7a8617628912efe0972.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 229, + 413, + 239 + ], + "lines": [ + { + "bbox": [ + 376, + 229, + 413, + 239 + ], + "spans": [ + { + "bbox": [ + 376, + 229, + 413, + 239 + ], + "type": "text", + "content": "(e) WDIP" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 424, + 114, + 481, + 228 + ], + "blocks": [ + { + "bbox": [ + 424, + 114, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 424, + 114, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 424, + 114, + 481, + 228 + ], + "type": "image", + "image_path": "84a5ded5fd38026e56440933687be20df2fe1c10714d45a1d58cddfbf1c2618f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 229, + 468, + 239 + ], + "lines": [ + { + "bbox": [ + 437, + 229, + 468, + 239 + ], + "spans": [ + { + "bbox": [ + 437, + 229, + 468, + 239 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 303, + 190, + 418 + ], + "blocks": [ + { + "bbox": [ + 131, + 247, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 131, + 247, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 482, + 293 + ], + "type": "text", + "content": "Fig. 7: Qualitative comparison of blind deblurring results on SEM image of semiconductor. (row 1): Circular ingredient of a semiconductor with the image size of " + }, + { + "bbox": [ + 131, + 247, + 482, + 293 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 131, + 247, + 482, + 293 + ], + "type": "text", + "content": ", (row 2): Wire structure of semiconductor with the image size of " + }, + { + "bbox": [ + 131, + 247, + 482, + 293 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 131, + 247, + 482, + 293 + ], + "type": "text", + "content": ". The ground truth PSFs are obtained by the Eq. (6) with " + }, + { + "bbox": [ + 131, + 247, + 482, + 293 + ], + "type": "inline_equation", + "content": "k = 10000" + }, + { + "bbox": [ + 131, + 247, + 482, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 303, + 190, + 418 + ], + "lines": [ + { + "bbox": [ + 133, + 303, + 190, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 303, + 190, + 418 + ], + "type": "image", + "image_path": "59ddd69a082b932c61ee4729dbbc66e35c4a6535f3d6d30e62c24fb24b1f2500.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 418, + 177, + 427 + ], + "lines": [ + { + "bbox": [ + 146, + 418, + 177, + 427 + ], + "spans": [ + { + "bbox": [ + 146, + 418, + 177, + 427 + ], + "type": "text", + "content": "(a) Blur" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "lines": [ + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "type": "text", + "content": "Fig. 8: Qualitative comparison of blind deblurring results on SEM image severely degraded by noise, with an image size of " + }, + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "type": "text", + "content": ". The ground truth PSFs are obtained by the Eq. (6) with " + }, + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "type": "inline_equation", + "content": "k = 10000" + }, + { + "bbox": [ + 131, + 437, + 482, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 192, + 303, + 249, + 418 + ], + "blocks": [ + { + "bbox": [ + 192, + 303, + 249, + 418 + ], + "lines": [ + { + "bbox": [ + 192, + 303, + 249, + 418 + ], + "spans": [ + { + "bbox": [ + 192, + 303, + 249, + 418 + ], + "type": "image", + "image_path": "f83027979244cd36d8553bf403186df18474375e7392ff23c4a1492cdec100cf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 418, + 233, + 427 + ], + "lines": [ + { + "bbox": [ + 206, + 418, + 233, + 427 + ], + "spans": [ + { + "bbox": [ + 206, + 418, + 233, + 427 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 250, + 303, + 306, + 418 + ], + "blocks": [ + { + "bbox": [ + 250, + 303, + 306, + 418 + ], + "lines": [ + { + "bbox": [ + 250, + 303, + 306, + 418 + ], + "spans": [ + { + "bbox": [ + 250, + 303, + 306, + 418 + ], + "type": "image", + "image_path": "d68bf3db04f165df39cc526bd0f950fa0158c27e317052ef8c23a38e6640e642.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 418, + 293, + 427 + ], + "lines": [ + { + "bbox": [ + 263, + 418, + 293, + 427 + ], + "spans": [ + { + "bbox": [ + 263, + 418, + 293, + 427 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 263, + 418, + 293, + 427 + ], + "type": "inline_equation", + "content": "l_{1} / l_{2}" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 307, + 303, + 365, + 418 + ], + "blocks": [ + { + "bbox": [ + 307, + 303, + 365, + 418 + ], + "lines": [ + { + "bbox": [ + 307, + 303, + 365, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 365, + 418 + ], + "type": "image", + "image_path": "1f13d2132b5a737c0a62f82328bfc5fd735f8db3ec5cfaa2b418dfe2c5bb4361.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 418, + 360, + 427 + ], + "lines": [ + { + "bbox": [ + 312, + 418, + 360, + 427 + ], + "spans": [ + { + "bbox": [ + 312, + 418, + 360, + 427 + ], + "type": "text", + "content": "(d) Pan-DCP" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 365, + 303, + 422, + 418 + ], + "blocks": [ + { + "bbox": [ + 365, + 303, + 422, + 418 + ], + "lines": [ + { + "bbox": [ + 365, + 303, + 422, + 418 + ], + "spans": [ + { + "bbox": [ + 365, + 303, + 422, + 418 + ], + "type": "image", + "image_path": "75e57a03269707fc59dbd13f49c663cca7d5a3a849cc5adb9d0dd44ef066d392.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 418, + 413, + 427 + ], + "lines": [ + { + "bbox": [ + 376, + 418, + 413, + 427 + ], + "spans": [ + { + "bbox": [ + 376, + 418, + 413, + 427 + ], + "type": "text", + "content": "(e)WDIP" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 424, + 303, + 481, + 418 + ], + "blocks": [ + { + "bbox": [ + 424, + 303, + 481, + 418 + ], + "lines": [ + { + "bbox": [ + 424, + 303, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 424, + 303, + 481, + 418 + ], + "type": "image", + "image_path": "c5fe6898d6aadafdf7d597c8f0414f4e352a52085c576afa880d77a3923d2964.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 418, + 468, + 427 + ], + "lines": [ + { + "bbox": [ + 437, + 418, + 468, + 427 + ], + "spans": [ + { + "bbox": [ + 437, + 418, + 468, + 427 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 495, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 542 + ], + "type": "text", + "content": "examining an image with blurred wire patterns, predicting the structure can be challenging even for human observers. However, the deblurred images from the proposed method show a remarkable similarity to the ground truth images and one can effectively investigate the wire structures." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 131, + 561, + 478, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 561, + 478, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 561, + 478, + 574 + ], + "type": "text", + "content": "4.3 Experiments on two-photon excitation fluorescence microscope" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "We further applied the proposed method as one of the real-world applications to a two-photon excitation fluorescence microscope (2PEF) image. In general, the imaging model of 2PEF follows Eq. (1) and its effective PSF can be expressed as the square of illumination PSF [13]. Also, due to the inherent characteristic of weak signals of fluorescence, the measured signal is often contaminated by noise. Here, we captured the fluorescence signal of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "2\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " beads from the custom-built 2PEF setup. As shown in Fig. 9 (a), the measured image from the 2PEF" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "Blind image deblurring with noise-robust kernel estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 219, + 201 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 219, + 201 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 219, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 219, + 201 + ], + "type": "image", + "image_path": "e7b06def865e5d6cfa51ce35bd4b2207d2ca25469b25ffe19ef0ee3058d5e036.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 202, + 235, + 211 + ], + "lines": [ + { + "bbox": [ + 203, + 202, + 235, + 211 + ], + "spans": [ + { + "bbox": [ + 203, + 202, + 235, + 211 + ], + "type": "text", + "content": "(a) Blur" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 114, + 261, + 201 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 261, + 201 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 261, + 201 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 261, + 201 + ], + "type": "image", + "image_path": "7073626d5e38656f58f5b5d45f5bd018aa37e59fe579d9c202a499e13c337cba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "lines": [ + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "type": "text", + "content": "Fig. 9: Visualization of blind deblurring results of the proposed method on blurry two-photon excitation fluorescence microscope image with an image size of " + }, + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 261, + 114, + 304, + 201 + ], + "blocks": [ + { + "bbox": [ + 261, + 114, + 304, + 201 + ], + "lines": [ + { + "bbox": [ + 261, + 114, + 304, + 201 + ], + "spans": [ + { + "bbox": [ + 261, + 114, + 304, + 201 + ], + "type": "image", + "image_path": "dad3b828ab8cfeecf50851802d675dc429011dba2f3720f4bf57136aeb07fac9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 395, + 201 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 395, + 201 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 395, + 201 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 395, + 201 + ], + "type": "image", + "image_path": "23ecd2b9533b95656f370f792432d76e2bb282fa42ea07831e7f641fa38e5016.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 202, + 412, + 212 + ], + "lines": [ + { + "bbox": [ + 378, + 202, + 412, + 212 + ], + "spans": [ + { + "bbox": [ + 378, + 202, + 412, + 212 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 114, + 437, + 201 + ], + "blocks": [ + { + "bbox": [ + 395, + 114, + 437, + 201 + ], + "lines": [ + { + "bbox": [ + 395, + 114, + 437, + 201 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 437, + 201 + ], + "type": "image", + "image_path": "05ee95bf8e5ffa926acb758183b9589edf218f9be8622082477962eba373e037.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 437, + 114, + 480, + 201 + ], + "blocks": [ + { + "bbox": [ + 437, + 114, + 480, + 201 + ], + "lines": [ + { + "bbox": [ + 437, + 114, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 437, + 114, + 480, + 201 + ], + "type": "image", + "image_path": "d6b89de1eaad5ef1709da2548882248e9cd65b9b4fbdb686f6550b350df6f9ea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 266, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 326 + ], + "type": "text", + "content": "exhibits a blurry appearance, attributed to the diffraction limit of the illumination PSF, and is further corrupted by noise. However, when the proposed deblurring method is applied to the measured image, illustrated in Fig. 9 (b), it effectively reconstructs the sharp edges of beads and facilitates easy discernment of clustered beads." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 345, + 225, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 225, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 225, + 357 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 370, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 525 + ], + "type": "text", + "content": "In this paper, we propose a noise-robust blind deblurring method combining the multiple kernel estimation scheme and the DIP. The most critical aspect of the proposed method, the noise-robust kernel estimation function, successfully restores the PSF even in situations of high noise when provided only with clear and blurry images. This function is developed to the multiple kernel estimation scheme that can be applied even in scenarios where noise level cannot be accurately predicted. Additionally, by effectively leveraging DIP, which acts as a powerful prior for natural images, we can estimate a clear image from a single well-designed neural network. The discussion of limitations of the proposed work and ablation study are presented in Supplementary S7. Our proposed method has been demonstrated to operate effectively in both simulated images and real-world applications and outperforms compared to the previous state-of-the-art methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 556, + 246, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 246, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 246, + 570 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 581, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 480, + 641 + ], + "type": "text", + "content": "This work was supported by the Samsung Research Funding and Incubation Center of Samsung Electronics grant SRFC-IT2002-03, Samsung Electronics Co., Ltd. (IO220908-02403-01), and the National Research Foundation of Korea grant funded by the Korean government (Grant Nos. NRF-2021R1A5A1032937, NRF-2021R1C1C1011307, RS-2023-00251628, and RS-2024-00397673)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "C.Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "text", + "content": "1. Arjomand Bigdeli, S., Zwicker, M., Favaro, P., Jin, M.: Deep mean-shift priors for image restoration. Advances in Neural Information Processing Systems 30 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 180 + ], + "type": "text", + "content": "2. Binding, J., Mikula, S., Denk, W.: Low-dosage maximum-a-posteriori focusing and stigmation. Microscopy and Microanalysis 19(1), 38-55 (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 480, + 213 + ], + "type": "text", + "content": "3. Bredell, G., Erdil, E., Weber, B., Konukoglu, E.: Wiener guided dip for unsupervised blind image deconvolution. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3047-3056 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "text", + "content": "4. Chen, L., Fang, F., Lei, S., Li, F., Zhang, G.: Enhanced sparse model for blind deblurring. In: European Conference on Computer Vision. pp. 631-646. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 267 + ], + "type": "text", + "content": "5. Cho, S., Lee, S.: Fast motion deblurring. In: ACM SIGGRAPH Asia 2009 papers, pp. 1-8 (2009)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "type": "text", + "content": "6. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "text", + "content": "7. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6059-6069 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "text", + "content": "8. Dhawan, A.P., Rangayyan, R.M., Gordon, R.: Image restoration by wiener deconvolution in limited-view computed tomography. Applied optics 24(23), 4013-4020 (1985)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "text", + "content": "9. Dong, J., Pan, J., Sun, D., Su, Z., Yang, M.H.: Learning data terms for non-blind deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 748-763 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "text", + "content": "0. Dong, J., Roth, S., Schiele, B.: Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems 33, 1048-1059 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "text", + "content": "1. Gandelsman, Y., Shocher, A., Irani, M.: \"double-dip\": unsupervised image decomposition via coupled deep-image-priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11026-11035 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 460, + 447, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 447, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 447, + 472 + ], + "type": "text", + "content": "2. Gonzalez, R.C.: Digital image processing. Pearson education india (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 472, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 492 + ], + "type": "text", + "content": "3. Helmchen, F., Denk, W.: Deep tissue two-photon microscopy. Nature methods 2(12), 932-940 (2005)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 492, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 525 + ], + "type": "text", + "content": "4. Jin, M., Roth, S., Favaro, P.: Noise-blind image deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3510-3518 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "text", + "content": "5. Jin, M., Roth, S., Favaro, P.: Normalized blind deconvolution. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 668-684 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "text", + "content": "6. Joshi, N., Szeliski, R., Kriegman, D.J.: Psf estimation using sharp edge prediction. In: 2008 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1-8. IEEE (2008)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "text", + "content": "7. Kawasaki, T., Nakano, T., Hirose, K.: Developing an aberration-corrected schottky emission sem and method for measuring aberration. Microelectronic engineering 86(4-6), 1017-1020 (2009)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "text", + "content": "8. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "9. Kruse, J., Rother, C., Schmidt, U.: Learning to push the limits of efficient ft-based image deconvolution. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4586-4594 (2017)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "Blind image deblurring with noise-robust kernel estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "20. Kubo, R., Toda, M., Hashitsume, N.: Statistical physics II: nonequilibrium statistical mechanics, vol. 31. Springer Science & Business Media (2012)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "21. Lampard, D.: Generalization of the wiener-khintchine theorem to nonstationary processes. Journal of Applied Physics 25(6), 802-803 (1954)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "type": "text", + "content": "22. Lee, W., Nam, H.S., Kim, Y.G., Kim, Y.J., Lee, J.H., Yoo, H.: Robust autofocusing for scanning electron microscopy based on a dual deep learning network. Scientific reports 11(1), 20933 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 197, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 480, + 228 + ], + "type": "text", + "content": "23. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 230, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 480, + 251 + ], + "type": "text", + "content": "24. Li, J., Luisier, F., Blu, T.: Pure-let image deconvolution. IEEE Transactions on Image Processing 27(1), 92-105 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 253, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 253, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 480, + 285 + ], + "type": "text", + "content": "25. Lifshin, E., Kandel, Y.P., Moore, R.L.: Improving scanning electron microscope resolution for near planar samples through the use of image restoration. Microscopy and Microanalysis 20(1), 78-89 (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 286, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 480, + 307 + ], + "type": "text", + "content": "26. Lu, Y., Zhang, X., Li, H.: A simplified focusing and astigmatism correction method for a scanning electron microscope. AIP Advances 8(1) (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "text", + "content": "27. Na, J., Kim, G., Kang, S.H., Kim, S.J., Lee, S.: Deep learning-based discriminative refocusing of scanning electron microscopy images for materials science. Acta Materialia 214, 116987 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "28. Nan, Y., Quan, Y., Ji, H.: Variational-em-based deep learning for noise-blind image deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3626-3635 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "text", + "content": "29. Pan, J., Hu, Z., Su, Z., Yang, M.H.: Deblurring text images via l0-regularized intensity and gradient prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2901-2908 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 480, + 441 + ], + "type": "text", + "content": "30. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1628-1636 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "text", + "content": "31. Perrone, D., Favaro, P.: Total variation blind deconvolution: The devil is in the details. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2909-2916 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 476, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 480, + 497 + ], + "type": "text", + "content": "32. Reimer, L.: Scanning electron microscopy: physics of image formation and microanalysis. Measurement Science and Technology 11(12), 1826-1826 (2000)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "text", + "content": "33. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3341-3350 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 532, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 586 + ], + "type": "text", + "content": "34. Roels, J., Aelterman, J., De Vylder, J., Luong, H., Saeys, Y., Philips, W.: Bayesian deconvolution of scanning electron microscopy images using point-spread function estimation and non-local regularization. In: 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). pp. 443-447. IEEE (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 587, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 620 + ], + "type": "text", + "content": "35. Sanghvi, Y., Mao, Z., Chan, S.H.: Structured kernel estimation for photon-limited deconvolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9863-9872 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "36. Schubert, P.J., Saxena, R., Kornfeld, J.: Deepfocus: Fast focus and astigmatism correction for electron microscopy. arXiv preprint arXiv:2305.04977 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "37. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. Acm transactions on graphics (tog) 27(3), 1-10 (2008)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "C.Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 336 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "39. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9446-9454 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "40. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 157-170. Springer (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "41. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1107-1114 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "42. Xue, F., Luisier, F., Blu, T.: Multi-wiener sure-let deconvolution. IEEE Transactions on Image Processing 22(5), 1954–1968 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "43. Zotta, M.D., Nevins, M.C., Hailstone, R.K., Lifshin, E.: The determination and application of the point spread function in the scanning electron microscope. Microscopy and microanalysis 24(4), 396-405 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "44. Zuo, W., Ren, D., Zhang, D., Gu, S., Zhang, L.: Learning iteration-wise generalized shrinkage-thresholding operators for blind deconvolution. IEEE Transactions on Image Processing 25(4), 1751-1764 (2016)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "Blind image deblurring with noise-robust kernel estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_content_list.json b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1d3c11bc30ae666ab79455186eb89cf04a39bcb7 --- /dev/null +++ b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_content_list.json @@ -0,0 +1,1842 @@ +[ + { + "type": "text", + "text": "BlinkVision: A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yijin Li $^{1,2*}$ , Yichen Shen $^{1*}$ , Zhaoyang Huang $^{2*}$ , Shuo Chen $^{1}$ , Weikang Bian $^{3}$ , Xiaoyu Shi $^{3}$ , Fu-Yun Wang $^{3}$ , Keqiang Sun $^{3}$ , Hujun Bao $^{1}$ , Zhaopeng Cui $^{1}$ , Guofeng Zhang $^{1\\dagger}$ , and Hongsheng Li $^{3,4,5\\dagger}$", + "bbox": [ + 222, + 233, + 781, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Lab of CAD&CG, Zhejiang University", + "bbox": [ + 295, + 291, + 702, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Avolution AI $^{3}$ CUHK MMLab $^{4}$ Shanghai AI Laboratory", + "bbox": [ + 295, + 306, + 702, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 CPII under InnoHK", + "bbox": [ + 424, + 320, + 575, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advances in event-based vision suggest that they complement traditional cameras by providing continuous observation without frame rate limitations and high dynamic range which are well-suited for correspondence tasks such as optical flow and point tracking. However, so far there is still a lack of comprehensive benchmarks for correspondence tasks with both event data and images. To fill this gap, we propose BlinkVision, a large-scale and diverse benchmark with rich modality and dense annotation of correspondence. BlinkVision has several appealing properties: 1) Rich modalities: It encompasses both event data and RGB images. 2) Rich annotations: It provides dense per-pixel annotations covering optical flow, scene flow, and point tracking. 3) Large vocabulary: It incorporates 410 daily categories, sharing common classes with widely-used 2D and 3D datasets such as LVIS and ShapeNet. 4) Naturalistic: It delivers photorealism data and covers a variety of naturalistic factors such as camera shake and deformation. BlinkVision enables extensive benchmarks on three types of correspondence tasks (i.e., optical flow, point tracking and scene flow estimation) for both image-based methods and event-based methods, leading to new observations, practices, and insights for future research. The benchmark website is https://www.blinkvision.net/.", + "bbox": [ + 259, + 372, + 743, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 681, + 375, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern image-based computer vision technology still cannot match the accuracy and robustness of human vision in many areas. One possible reason is that traditional cameras suffer from motion blur and limited frame rates, and they often rely on well-lighted conditions. In contrast, event cameras [13, 36] detect changes in intensity at each pixel as a stream of asynchronous events, which", + "bbox": [ + 212, + 720, + 787, + 797 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Yijin Li, Yichen Shen, and Zhaoyang Huang contributed equally to this work.", + "bbox": [ + 218, + 806, + 707, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Hongsheng Li and Guofeng Zhang are the corresponding authors.", + "bbox": [ + 220, + 820, + 632, + 833 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0dc5854d567c50c6617ec6154781f8bdd52f6315f8ad9ca255ee96fc23579e7.jpg", + "image_caption": [ + "Fig. 1: BlinkVision is a large-scale and diverse benchmark with rich modality and dense annotation of correspondence. It covers 410 daily categories, sharing common classes with popular 2D and 3D datasets. The per-category object distributions, scene structure hierarchy, data samples, and supported applications of BlinkVision are shown in this figure." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 781, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eliminates frame rate limitations and enables operation within a high dynamic range. However, they can not capture fine-grained details as traditional cameras do. Looking at how human vision works, we find that two cells in the human retina, i.e., cones and rods, work similarly to these two types of cameras, respectively. According to the duplex theory of vision [45], which posits that rods and cones serve different functions, their combination ensures the robustness of human visual processing. Therefore, we believe that combining the advantages of traditional cameras and event cameras can significantly enhance computer vision systems, providing more comprehensive and adaptive vision capabilities.", + "bbox": [ + 212, + 446, + 787, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, there are currently only a few benchmarks [15, 32, 34] providing both event data and RGB frames, which hinder the development of algorithms that fully exploit the event data and fuse information from both modalities. This shortfall is particularly prominent in the domain of pixel correspondence estimation [8, 27, 28, 61], i.e., optical flow, scene flow, and point tracking estimation. Previous benchmarks built for pixel correspondence are either highly biased to specific scenes [15, 63] or rather simple [53]. A primary factor is that obtaining such precise pixel-wise annotations for these tasks is expensive.", + "bbox": [ + 212, + 589, + 787, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To boost the research in this area, we present BlinkVision, a synthetic benchmark for optical flow, scene flow and point tracking estimation using RGB frames and events. Our dataset has several appealing properties: 1) Rich modalities: BlinkVision encompasses three visual modalities: final RGB images, clean RGB images and event data. The final RGB images reflect real-world challenges like motion blur and limited dynamic range while the clean RGB images are devoid of such imperfections. The clean RGB images can be seen as the latent images [48]", + "bbox": [ + 212, + 717, + 787, + 824 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/d786e973c7eef1932a7028251f76ad7dc15b2e2194d84d7aa17975794bc7cf7f.jpg", + "table_caption": [ + "Table 1: A comparison between BlinkVision and other widely-used benchmarks on pixel correspondence estimation. \"Occ\", \"Chara\" and \"Cats\" is the abbreviation for occlusion, character and categories, respectively. $R^{lvis}$ denotes the ratio of the 1.2k LVIS [20] categories being covered. \"N/A\" denotes the dataset does not provide category labels." + ], + "table_footnote": [], + "table_body": "
DatasetsEventCleanFinalOptical FlowScene FlowPoint TrackingOccCharaAnimalCatsR1vis(%)
MVSEC [64]✓X✓✓XXXXXN/A0
DSEC [15]✓X✓✓XXXXXN/A0
BlinkFlow [34]✓✓X✓XX✓✓X554.1
EKubric [59]✓✓X✓✓✓✓XX170.9
Sintel [9]X✓X✓XX✓✓✓N/A0
KITTI [17]XX✓✓✓XXXXN/A0
FlyingThings [42]X✓X✓✓X✓XX554.1
Kubric [18]X✓X✓✓✓✓XX170.9
TAP-Vid [12]X✓XXX✓X✓XN/A0
PointOdyssey [62]X✓XXX✓✓✓✓N/A0
BlinkVision (Ours)✓✓✓✓✓✓✓✓✓41033.4
", + "bbox": [ + 217, + 224, + 781, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of event cameras. 2) Rich annotations: It provides annotations covering optical flow, scene flow, and point tracking. Unlike benchmarks [15,17,62] which only contain sparse annotations, our data provide dense per-pixel annotations of each image, covering objects including moving cars, deformable characters, and animals. 3) Large vocabulary: It incorporates 410 daily categories, sharing common classes with widely-used 2D and 3D datasets such as ImageNet [33], LVIS [20], and ShapeNet [10], as depicted in Fig. 1. To our knowledge, our data have the widest variety of objects in existing pixel correspondence benchmarks. This expansive vocabulary is pivotal, enabling rigorous exploration of algorithmic generalization across varied objects. 4) Naturalistic: BlinkVision employs high-quality assets and rendering tools and thus is able to deliver photorealism data. Besides, it covers a variety of naturalistic factors such as camera shake and deformation.", + "bbox": [ + 212, + 400, + 787, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To make full use of BlinkVision, we set up a public benchmark website that allows uploading results and provides a public leaderboard. Besides, we evaluated both existing image-based and event-based methods on three typical correspondence tasks (i.e. optical flow, point tracking, and scene flow estimation). The results reveal new observations and challenges and serve as the baseline for future approaches. Specifically, we first study the robustness of existing image-based methods under large frame intervals and extreme illumination and point out the new challenge for these methods. Second, the benchmark results on existing event-based methods show that current methods do not fully unleash the potential of event cameras. Third, we show that fine-tuning existing methods on the training set of BlinkVision significantly boosts the generalizability, demonstrating the vast diversity of BlinkVision. Finally, the broad diversity and accessible category labels in BlinkVision allow us, for the first time, to analyze the performance of correspondence tasks on different categories. We believe BlinkVision has the potential to serve as a cornerstone benchmark for advancing the development of more robust computer vision systems.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BlinkVision: A Benchmark for Correspondence Estimation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/faa359d42f2eb44ad9aa4bb22d01923d53e76a236e09fd281055ae554fc13d60.jpg", + "image_caption": [ + "Fig. 2: Scenes samples in the proposed BlinkVision benchmark." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 324, + 387, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A comprehensive overview and comparison between these benchmarks and BlinkVision is listed in Table 1.", + "bbox": [ + 212, + 363, + 797, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-based optical flow and scene flow benchmarks. Early evaluations depended on synthetic datasets, like the famous \"Yosemite\" sequence [7]. MPI Sintel [9] was one of the most representative synthetic benchmarks derived from a short open-source animated 3D movie and it became one of the most popular benchmark datasets. KITTI [17] was almost the most well-known among real data. It computed ground truth for static scenes through the data from a 3D laser scanner and the ego-motion data of the car. In a later version, KITTI extended the ground truth to rigidly moving cars by fitting CAD models of cars. FlyingThings3D [42] was a more recent synthetic dataset. It contained the ground truth of both optical flow and scene flow. FlyingThings3D and KITTI were the two most commonly used benchmarks in scene flow evaluation. However, on both datasets, ground truth about scene flow was not available in occluded regions. Furthermore, on KITTI, ground truth for foreground points was either missing or approximated by a fitted car CAD model.", + "bbox": [ + 212, + 401, + 787, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-based point tracking benchmarks. The first point-tracking dataset was FlyingThings++ [21], which was based on FlyingThings3D. It was initially designed for training the network. Due to the lack of evaluation benchmarks, it was also used for evaluation in the early stages. Later, Doersch et al. [12] proposed a real data benchmark named TAP-Vid, which was based on two real-world datasets: DAVIS [49] and Kinetics [30]. TAP-Vid relied on manual annotation and could not handle occlusions. PointOdyssey [62] was a newly proposed benchmark that was based on synthetic data and therefore could provide dense ground truth. However, it did not guarantee that every pixel had ground truth because it only tracked the mesh vertices of the object. Furthermore, the naturalism of PointOdyssey was limited to its interior parts. Its outdoor portion had no realistic layout, just a skybox with randomly dropped objects.", + "bbox": [ + 212, + 619, + 787, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event-based Benchmarks. Early event-based optical flow benchmarks [53] limited the camera to only rotational motion and inferred ground truth from", + "bbox": [ + 212, + 806, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the rotational motion of the camera. Two later benchmarks, DSEC [15] and MVSEC [64] computed the ground truth through LiDAR SLAM. Similar to KITTI, these two benchmarks limited the ground truth to the static elements of the scene. Besides, they had a rather limited motion pattern. More recently, Li et al. [34] proposed a large-scale diversified form synthetic benchmark named BlinkFlow, which presented more challenging scenes and complex motion patterns. As for the scene flow benchmark, Wan et al. [59] recently converted the existing FlyingThings [42] and Kubric [18] datasets to event datasets through video-to-event [14] technology. However, due to the limited frame rates in the original video data, this step inevitably produced artifacts in the event data. Wan et al. also extended the real-world event dataset DSEC with scene flow ground truth. As for the event-based point-tracking benchmark, existing methods usually employed Structure-from-Motion [54] (SfM) technology to associate keypoints across multiple frames to obtain correspondence ground truth. Event Camera Dataset [46] and EDS dataset [22] were two common benchmarks in this area. Limited by the sparsity and accuracy of SfM, these benchmarks did not support the evaluation of arbitrary point tracking.", + "bbox": [ + 212, + 145, + 787, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 BlinkVision Dataset", + "text_level": 1, + "bbox": [ + 215, + 430, + 444, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe how to build BlinkVision, including the scene setup, data rendering, and generation for the ground truth labels. The process is based on Blender [2] as it provides photorealistic rendering and a flexible data interface that allows us to obtain customized correspondence ground truth. At the last, we introduce the statistics and distribution of the BlinkVision data.", + "bbox": [ + 212, + 465, + 787, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Scene Setting", + "text_level": 1, + "bbox": [ + 215, + 568, + 375, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous synthetic benchmarks [9, 43] generally rely on open-source movies to avoid heavy scene construction. However, these movies are biased and do not cover diverse enough scenarios. In order to establish a comprehensive evaluation benchmark, we manually assemble a collection of scenarios that are as rich and diverse as possible. We first look for ready-made scenes that are photorealistic and visually diverse. Specifically, we purchased 40 indoor scenes and 13 outdoor scenes from Evermotion Archinteriors Collection [3] that cover common scenes such as living rooms, kitchens, offices, bedrooms, restaurants and gardens. The original scene is static. To enhance the realism, we procured 29 scanned and high-fidelity human bodies from ActorCore [1] and 88 artist-designed animals, together with 104 free characters from Mixamo [4]. These assets are rigged. We re-targeted them to a wide range of motion models (e.g., more than 100 human motions including dancing, walking, talking, etc.) and placed them in various locations. The human's motions mainly come from motion capture [1] and the animal's motions are designed by the artist to closely emulate their natural motions. Furthermore, we built 50 additional outdoor scenes that cover cases such", + "bbox": [ + 212, + 598, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BlinkVision: A Benchmark for Correspondence Estimation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as a drone flying through the forest and a camera following a human or car from the crowd, which are common but rarely included in existing correspondence benchmarks. We set the camera trajectory by referencing the shooting trajectories of real-world videos, including handheld shots, car shots, and drone shots that cover non-uniform motion such as sudden stops and sharp turns. Finally, we obtained 80 indoor sequences and 63 outdoor sequences, where 56 indoor sequences and 47 outdoor sequences are for testing and the others are used for fine-tuning. Some samples of our data are shown in Fig. 2. In the supplementary, we also provide video samples. All the assets used in the testing and fine-tuning are totally disjoint, even including trees.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Multi-modality Data Rendering", + "text_level": 1, + "bbox": [ + 215, + 324, + 524, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use Blender to render brightness $B$ in linear color space. The following simulations of real-world image capturing and event data are both based on the linear color space. We employ tone mapping and gamma correction on $B$ to obtain commonly used sRGB images. We call these sRGB images \"RGB (clean)\" because they do not suffer from effects like blurs or overexposed. On the contrary, we call the simulated real-world image capture \"RGB (final)\".", + "bbox": [ + 212, + 353, + 787, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Simulation of Real-world Image Capturing. Real-world image capture suffers from motion blur and limited dynamic range. While Blender supports the simulation of the former, the latter needs additional processing. The latter effect usually leads to overexposed or underexposed. To simulate it, we follow previous work [40] and employ the following steps: (1) Random exposure ratio. We uniformly sample the exposure ratio $w$ in the log2 space within $[-3, 3]$ and multiply it by $B$ to simulate underexposed or overexposed, which gives us the augmented HDR image $H = B \\times 2^w$ . (2) Dynamic range clipping. We clip $H$ according to the formulation $\\mathcal{C}(H) = \\min(H, 1)$ . This step leads to information loss for pixels in the overexposed regions. (3) Non-linear mapping. To align with how humans see a scene, a camera typically uses a non-linear camera response function (CRF) to modify the contrast of the captured image, which can be formulated by $I_n = \\mathcal{F}(I_c)$ . We randomly sample CRFs from an existing dataset [19]. (4) Quantization. The pixel values are quantized to 8 bits by $\\mathcal{Q}(I_n) = \\lfloor 255 \\times I_n + 0.5 \\rfloor / 255$ . This step causes information loss in underexposed and smooth gradient areas.", + "bbox": [ + 212, + 450, + 787, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Event Data. Event cameras work by responding to changes in the logarithmic brightness signal (i.e., $L = \\log B$ ) asynchronously and independently for each pixel [13]. An event is triggered when the change in brightness (either increase or decrease) since the last event at that pixel reaches a threshold of $\\pm T$ (with $T > 0$ ):", + "bbox": [ + 212, + 696, + 787, + 772 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {k} \\left(L \\left(u _ {k}, t _ {k}\\right) - L \\left(u _ {k}, t _ {k} - \\Delta t\\right)\\right) \\geq T, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 773, + 784, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Delta t$ is the time since the last event triggered at pixel location $u_{k}$ and at time $t_k$ . $p_k\\in \\{-1,1\\}$ is the polarity of the brightness change. To model the event generation process, we need access to a continuous representation of the", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33e38f8766414540ca2828fad75c6240fec90d90663cb8bf0c14a78b1f65fb68.jpg", + "image_caption": [ + "Fig. 3: Statistics of optical flow in BlinkVision." + ], + "image_footnote": [], + "bbox": [ + 220, + 147, + 352, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7871dd5c68c7fae8bde7cd421ec565d2288d0542eb6dff9bcbbcb8f8c0b67c08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 148, + 496, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f111e60c5c2783ef8312e798bc837738634169ed62cb823562f09deacb49dc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 148, + 638, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a9af71414961f4bea91647eed0cb78037f93e8562907b89e6b343b884fa1ad3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 148, + 781, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8aa04e1ff8f98f4f6a6a10e7717999c28a80b2053f4b1e241eba7e8ae8941fd.jpg", + "image_caption": [ + "Fig. 4: Statistics of point trajectories in BlinkVision. To save time, we sample the tracks using a grid with a size of 20. Trajectory segments are defined as contiguous sections of point trajectories, with interruptions caused by occlusion. The diameter is the maximum distance a point moves over time. We clip the trajectory diameter and divide it by the diagonal length of the image to obtain the ratio." + ], + "image_footnote": [], + "bbox": [ + 222, + 294, + 352, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b370f6293496785a4645d120e316adc729661801ed343c8c178344b0ee1bc280.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 295, + 493, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86231bbda378e7cd63d0d0fe98425d913927a299cce9eb4ad10fc2b61fb12cb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 295, + 635, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8862018015469665aa541f02eab472549cf493b168b480afefd79682dcac6dfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 295, + 777, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "visual signal for each pixel. In practice, it is approximated by rendering images at high frame rates for efficiency [34, 51]. Events are then synthesized based on frame-by-frame pixel differences. More specifically, we adaptively sample frames according to [34, 51] to ensure that the maximum pixel displacement between two sampling timestamps is bounded. In practice, we employ frame interpolation before simulating the event data to further reduce the pixel displacement. We use DVS-Voltmeter [37] to synthesize events because it can simulate complex noise effects (such as noise effects of temperature and parasitic photocurrent), thus generating realistic events.", + "bbox": [ + 212, + 508, + 787, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Ground Truth Generation", + "text_level": 1, + "bbox": [ + 214, + 674, + 477, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Blender provides optical flow data between two consecutive frames and a segmentation mask for each object. However, we cannot directly obtain these data from Blender: (1) pixel correspondence across multiple frames; (2) scene flow between frames; (3) semantic category for each object.", + "bbox": [ + 212, + 707, + 785, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pixel Correspondence. The data of (1) and (2) can be computed in a unified framework. Given camera poses $\\mathbf{T}$ (camera-to-world), depth maps $\\mathbf{Z}$ at frames $i$ and $j$ , and the camera intrinsic $K$ , for the pixel location $\\mathfrak{u}$ at frame $i$ , we first project it to world coordinate: $P = T_{i}Z_{i}(\\mathfrak{u})K^{-1}\\tilde{\\mathfrak{u}}$ , where $\\tilde{\\mathfrak{u}}$ is the homogeneous", + "bbox": [ + 212, + 773, + 787, + 835 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BlinkVision: A Benchmark for Correspondence Estimation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "coordinates of u. Then we re-project the 3D point $P$ to frame $j$ through $du_{j} = KT_{j}^{-1}P$ , where $\\mathfrak{u}_j$ is the corresponding pixel location and $d$ is the depth value at frame $j$ . In this way, we get the pixel correspondence between any two frames and the corresponding depth, i.e., (1) and (2).", + "bbox": [ + 215, + 146, + 784, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, this only works for static objects. To handle dynamic objects and even those objects with deformation motion such as humans and animals, we bake the face index and barycentric coordinate $(\\lambda_1,\\lambda_2,\\lambda_3)$ of each mesh face into textures. After rendering the texture, we locate the triangular face that the tracked pixel belongs to (indexed by the rendered face index) and obtain the vertices' 3D position $(V_{1},V_{2},V_{3})$ of that face via Blender's API. The 3D position of tracked pixels can be computed through $P = \\lambda_1V_1 + \\lambda_2V_2 + \\lambda_3V_3$ . This step replaces the back projection for the static objects. It is worth noting that we need to triangulate every mesh face in the scenes before baking.", + "bbox": [ + 212, + 213, + 787, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic Labeling. BlinkVision contains thousands of objects which makes manual semantic labeling expensive. To this end, we develop an automatic labeling framework that is based on open-vocabulary semantic understanding [50,60]. First, we pre-define a category list based on LVIS [20]. The category together with category descriptions is encoded into text embedding $F_{t} \\in \\mathbb{R}^{N \\times D}$ with a pre-trained text encoder, i.e., CLIP, where $N$ is the number of categories and $D$ is the feature dimension. For each object asset, we render it individually to a $224 \\times 224$ image. The process is rather fast while keeping photorealistic because it does not need to compute complex path tracing between objects. We use the pre-trained image encoder from CLIP [50] to extract the image embedding $F_{i} \\in \\mathbb{R}^{D}$ . The probability of the image belonging to one of the pre-defined categories is computed through $P(c|F_{i}) = \\text{softmax}\\left(F_{i} \\cdot (F_{t})^{\\top}\\right)_{c}$ .", + "bbox": [ + 212, + 357, + 787, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Dataset Overview", + "text_level": 1, + "bbox": [ + 215, + 568, + 410, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BlinkVision consists of 40 training sequences and 103 testing sequences. The training set provides 107,880 frame pairs for optical flow and scene flow, and 1025 sub-sequences for point tracking. The frame rate for RGB images and ground truth data is 20 FPS. The frame rate is not practically significant because we adjusted it to ensure sufficient motion between adjacent frames. This adjustment reduces the data size uploaded to the benchmark website. For each sequence of BlinkVision, we selected the sub-sequences by starting tracking from the first frame. Once the overlap with the reference frame (i.e., how many tracked points remain inside the image) is smaller than $40\\%$ , we start to track a new reference frame. If the overlap is smaller than $20\\%$ , we stop this track. Additionally, we discard short tracks that are less than 20 frames long because they are less challenging without long-term motion accumulation, which could result in the benchmark easily reaching saturation. As a result, we generate 1025 subsequences where each sub-sequence contains $640 \\times 480$ dense point trajectories. For the test set, we selected 12,804 frame pairs for evaluating optical flow and", + "bbox": [ + 212, + 599, + 787, + 828 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00ed5a18729acc643bb21eeccd2a63109cc39473c56cf53006ba05d5f5788de0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 145, + 346, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ec217484f460347a83417bf4b61be13b7f9345ec6a8eb73e818306994ff7fb8c.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 246, + 205, + 346, + 263 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6a62f30eb2cc1b147b0ef6c389f51d55993217a41c1166d54c646d760128a717.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 146, + 449, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b99d97f9f3ad5542bf0153e156c7ebcf68a522898923a7a0f5a999a61a301813.jpg", + "image_caption": [ + "Target Image" + ], + "image_footnote": [], + "bbox": [ + 349, + 205, + 449, + 263 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0f3e3dbfc1006eec3cbb8c19cfb8a1962001a2ff168c1fb62b95db57677f3e7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 145, + 550, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/916d4e5e3c8285620e4c6dd491c1115d00cebb57f1d54b5df118f1f39b2ee9a9.jpg", + "image_caption": [ + "Ground Truth", + "Fig. 5: Qualitative results of FlowFormer++ [55] before and after fine-tuning on the training set of BlinkVision." + ], + "image_footnote": [], + "bbox": [ + 450, + 205, + 550, + 263 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/37a3e35769199669e38113ddc173d1ee66152c37b17ee54ea9de0b4fb561ffa1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 145, + 651, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1bb862dc35943f791ba10b72447595bff9fa6fd797b6563268ce4ed619da5876.jpg", + "image_caption": [ + "Before Finetuning" + ], + "image_footnote": [], + "bbox": [ + 552, + 205, + 651, + 263 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/81d74fd5de321b0235f96293c83e90b23dbd16a1facf962037c9e2a64bf5d673.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 145, + 754, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b84c7682a80a7267e4b60b98fe7da5ce680eec9e3b658dce7fac5dc116011736.jpg", + "image_caption": [ + "After Finetuning" + ], + "image_footnote": [], + "bbox": [ + 653, + 205, + 754, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "scene flow, and 865 sub-sequences for evaluating point tracking. BlinkVision includes raw data on depth, normal, camera pose, etc. At this stage, we focus on pixel-level motion tasks. The additional raw data will be released later, which will benefit other tasks such as depth estimation [35] and SLAM [23, 24, 39].", + "bbox": [ + 212, + 344, + 782, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1 summarizes key statistics of BlinkVision compared to related works. BlinkVision contains more than 40K 3D models in 410 categories. It shares some common categories with 2D and 3D datasets [10,20], and bears a huge diversity in semantics, geometry, and appearance, enabling a wide range of research topics.", + "bbox": [ + 212, + 406, + 784, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We provide statistics of optical flow in Fig. 3. For benchmarking, we discard those pixels whose flow magnitudes are larger than half of the image diagonal. The latter situation usually occurs when the object is too close to the camera, and its magnitude can even approach infinity. As a result, the magnitude shown in Fig. 3 has an upper bound of 400. The broad distribution in Fig. 3 reveals the diverse motions in BlinkVision. Besides, the diverse distribution of \"X Variance\" and \"Y Variance\" also indicates the presence of complex object shapes in BlinkVision. We also provide statistics of point trajectories in Fig. 4. BlinkVision allows the evaluation of point tracking with different trajectories' lengths, from small to big, corresponding to different downstream applications like video editing [25, 41] or augmented reality [31, 39]. The broad distribution of \"occlusion ratio\" and \"number of segments\" indicates that BlinkVision brings diverse and challenging data. The occlusion ratio has an upper bound that is less than 100 percent due to our strategy of sub-sequences selection. To quantify the amount of motion a point undergoes, we calculate its trajectory diameter, which is the maximum distance between any two positions of the point over its entire trajectory. We observe that point motion in BlinkVision is diverse and nearly uniformly distributed across the entire image plane.", + "bbox": [ + 212, + 468, + 785, + 741 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Benchmark", + "text_level": 1, + "bbox": [ + 214, + 763, + 362, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We release all the data except for the ground truth for the test split. For a fair comparison, we create a public benchmark website with leaderboards. We also release the evaluation code for our online benchmarking.", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BlinkVision: A Benchmark for Correspondence Estimation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Optical flow result of RGB-frame-based methods. \"St.\" denotes Stride. \"†\" denotes fine-tuning. \"FF\" denotes FlowFormer [26] and \"FF++\" denotes FlowFormer++ [55].", + "bbox": [ + 214, + 143, + 483, + 215 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9c7f15d76058a88f6b90a9abea79470823041bc2f0c291bc823a80da4c875628.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEPE - clean
St. 1St. 2St. 4St. 8Avg.
RAFT [56]1.834.6210.1820.469.27
GMA [29]1.824.379.5218.688.60
FF [26]1.603.777.5716.037.24
FF++ [55]1.543.577.4316.267.20
RAFT† [56]1.332.966.3912.725.85
FF++† [55]1.092.354.8810.584.73
MethodEPE - final
St. 1St. 2St. 4St. 8Avg.
RAFT [56]2.545.6911.2722.6710.54
GMA [29]2.565.7211.2121.3510.21
FF [26]2.264.809.2418.238.63
FF++ [55]2.284.839.3919.088.90
RAFT† [56]1.943.887.3515.177.08
FF++† [55]1.663.346.4212.736.04
", + "bbox": [ + 215, + 226, + 485, + 426 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5d60b8387a531622172cb55b533af88f9d5e875459e3368bbb96c34d73bf10ac.jpg", + "table_caption": [ + "Table 3: Optical flow result of event-based methods. \"St.\" denotes Stride. \"†\" denotes fine-tuning." + ], + "table_footnote": [], + "table_body": "
Event
Methods|St. 1St. 2St. 4St. 8Avg.
E-RAFT [16]2.817.0417.8228.6014.07
STE-FlowNet [11]2.595.9413.1341.8915.89
E-FlowFormer [34]2.415.6612.5330.4512.76
E-RAFT† [16]1.683.637.4814.206.75
E-FlowFormer† [34]1.513.005.9613.606.02
Event + RGB (clean)
Methods|St. 1St. 2St. 4St. 8Avg.
DCEIFlow [58]3.3114.0734.7861.5628.43
DCEIFlow† [58]2.196.3012.5426.1111.79
Event + RGB (final)
Methods|St. 1St. 2St. 4St. 8Avg.
DCEIFlow [58]4.0814.2434.5461.2728.53
DCEIFlow† [58]2.526.8613.6428.4512.87
", + "bbox": [ + 491, + 199, + 790, + 426 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Optical Flow", + "text_level": 1, + "bbox": [ + 215, + 455, + 369, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we comprehensively evaluate the performance of existing optical flow methods on BlinkVision. Specifically, we first analyze the robustness of image-based methods under large frame intervals and extreme illumination. Large motion, severe occlusion, and information loss in these extreme cases pose great challenges to existing image-based methods. In contrast, event cameras naturally depict continuous pixel motion and possess a large dynamic range, thus offering great potential for solving these challenges. We then benchmark existing event-based methods, including event-only methods and event-RGB fusion methods. We find that existing event-based methods cannot fully unleash the potential of event cameras. We analyze the possible reasons and point to possible opportunities for future research.", + "bbox": [ + 212, + 487, + 787, + 654 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental Setup. To study the impact of frame intervals on existing methods, we select frames separated by 1, 2, 4, and 8 from the reference frame as target frames. To reduce the file size of results that are uploaded to the benchmark website, we uniformly sample a reference frame from ten consecutive frames. As a result, there are 12,804 frame pairs for testing in total. We follow the previous work [26, 47, 56] and report the EPE (end-point-error) as the evaluation metric.", + "bbox": [ + 212, + 660, + 787, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. We benchmark several image-based methods in Table 2. We make the following observations: 1) The error generally increases linearly with stride size. This is expected because doubling the stride yields motion that is twice as fast, resulting in an error that is also expected to be twice as high. 2) Extreme lighting has a heavy effect on existing image-based methods. In the supplementary", + "bbox": [ + 212, + 758, + 787, + 835 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/82e4b0ca3bb8f3ce553798088c4448dc04945152fb9e7aa59891021684b17892.jpg", + "table_caption": [ + "Table 4: Point tracking result of RGB-frame-based methods under different frame interval. \"St.\" denotes Stride." + ], + "table_footnote": [], + "table_body": "
RGB(clean)
MethodMetricSt. 1St. 2St. 4St. 8Avg.
PIPs [21]δavg37.7739.3436.3425.5034.74
Survival↑52.2855.3954.5547.5852.45
MTE↓50.7747.8149.5362.4252.63
PIPs++ [62]δavg40.0238.2530.2416.8431.34
Survival↑53.4453.1347.5739.7148.46
MTE↓45.5151.0259.8773.1757.39
Context-TAP [8]δavg38.7539.4736.5125.9935.18
Survival↑49.9853.0952.6146.1850.46
MTE↓51.4649.6151.8364.3954.32
RGB(final)
MethodMetricSt. 1St. 2St. 4St. 8Avg.
PIPs [21]δavg33.0234.4431.5621.7730.20
Survival↑48.3551.5650.7244.5648.80
MTE↓53.9451.5453.6566.1456.32
PIPs++ [62]δavg35.8334.2427.1115.3228.12
Survival↑50.4750.3645.3238.4646.15
MTE↓48.3753.7462.3474.5959.76
Context-TAP [8]δavg33.4634.1731.3922.0030.25
Survival↑45.5848.9148.5343.0646.52
MTE↓55.0953.7656.3368.1158.32
", + "bbox": [ + 215, + 196, + 549, + 477 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/dc9e9da325a8268b7cc931b3084281e3e38cecc6fc6d529787a30eb7ec2a2334.jpg", + "table_caption": [ + "Table 5: Point tracking result of event-based methods. \\* denotes that we track ORB [52] feature points rather than grid sampled points. DET\" denotes DeepETracker [44]" + ], + "table_footnote": [], + "table_body": "
Event
Methodδavg↑Survival↑MTE↓
AMH [5]18.9537.9875.44
HASTE [6]18.6037.4377.36
AMH* [5]26.7641.9362.48
HASTE* [6]27.6542.8762.52
Event + RGB (clean)
Methodδavg↑Survival↑MTE↓
DET [44]21.6235.4480.80
DET* [44]33.4648.8260.69
Event + RGB (final)
Methodδavg↑Survival↑MTE↓
DET [44]19.0533.5285.00
DET* [44]31.7848.6758.79
", + "bbox": [ + 557, + 241, + 785, + 477 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "materials, we analyze how different exposure factors lead to performance loss. 3) Fine-tuning on the training split of BlinkVision significantly improves the performance. In Fig. 5 we show qualitative results of FlowFormer++ before and after fine-tuning on BlinkVision. After fine-tuning, the flow predictions are more precise and clear, especially near the object boundary. As shown in Table 2, the benchmark is still challenging for methods after fine-tuning.", + "bbox": [ + 212, + 506, + 784, + 598 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We report the evaluation of event-based methods in Table 3. It is surprising that the performance of existing event-based methods is worse than image-based methods and they also suffer from severe degradation under large frame intervals. The reasons are mainly twofold. First, existing training data for events lag behind that of RGB. This can be verified by the performance of E-RAFT [16] and E-FlowFormer [34] after fine-tuning. These two methods are far behind RAFT [56] before fine-tuning, but exceed the fine-tuned RAFT [56] and even FlowFormer++ [55] after fine-tuning (in \"Stride-1\", \"Stride-2\" and \"Stride-4\" of the \"final\" case), indicating that the poor performance of the former mainly comes from limited training on existing event-based training data. Second, existing event-based methods usually convert events to regular 3D voxel grids before processing, which quantizes the event data and leads to information loss. This problem is more serious when dealing with event data within large frame intervals. As a result, it calls for a more powerful representation of event data and new algorithms that can deal with long-range optical flow estimation.", + "bbox": [ + 212, + 604, + 787, + 832 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BlinkVision: A Benchmark for Correspondence Estimation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a7982135da724b9ece0dbb755629e1e824cd25daa5db236eab6ecb16f88d09de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 321, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/07efd0dd90f5376b5e39d86965790341a9bd44ede33f289c9840c11f62ebf74e.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 217, + 205, + 321, + 265 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4092383402648846b974d562746ae5ac1ca1896fbcfb47113019481ce6ba48d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 143, + 424, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/06e0cb0547b1227a6b7b3847a431399063bebf1bd5322545bdf793a191d7594e.jpg", + "image_caption": [ + "Prediction Results" + ], + "image_footnote": [], + "bbox": [ + 321, + 205, + 424, + 265 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/61c8963d077cbbdfc03d9a3272b870c4677c82ea64270d7d0f35becf1dd76eea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 143, + 529, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5e671c5a44f5ff742642096ef68661b643ec01918282fc3184b7a90ece145536.jpg", + "image_caption": [ + "Error Visualization" + ], + "image_footnote": [], + "bbox": [ + 426, + 205, + 529, + 265 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9f9cc860f4f23b7950a3edd2454b554088f6e7e66410a9518421f850953ecf2d.jpg", + "image_caption": [ + "Fig. 6: Qualitative results of $\\mathbf{PIP}\\mathbf{s} + +$ [62]. Darker lines in the right figure (e.g., red) indicate larger errors.", + "Fig. 7: Performance distribution of FlowFormer++ [55]." + ], + "image_footnote": [], + "bbox": [ + 535, + 157, + 779, + 279 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6: Scene flow result of RGB-frame-based methods. \"St.\" denotes Stride.", + "bbox": [ + 214, + 344, + 509, + 385 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0afb1514deda57e71216f01ddc18b7cd29d159dd1115cf614662c93d26270c8a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EPE-2d
MethodRGBSt. 1St. 2St. 4St. 8
RAFT-3D [57]clean2.154.348.5916.08
final3.487.2910.7418.57
CamLiFlow [38]clean2.796.0512.3423.89
final4.288.2816.3030.61
EPE-3d
MethodRGBSt. 1St. 2St. 4St. 8
RAFT-3D [57]clean1.913.837.2512.18
final2.023.967.3914.21
CamLiFlow [38]clean1.452.765.369.12
final1.532.885.609.81
", + "bbox": [ + 217, + 398, + 509, + 554 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a421525148022ce46a250ed5a134ce0d2d5ee2bee3a9b155207a9bebe9fa202e.jpg", + "table_caption": [ + "Table 7: Scene flow result of event-based methods. \"St.\" denotes Stride." + ], + "table_footnote": [], + "table_body": "
EPE-2d
MethodRGBSt. 1St. 2St. 4St. 8
RPEFlow [59]clean1.533.507.3617.44
final1.924.108.1518.38
EPE-3d
MethodRGBSt. 1St. 2St. 4St. 8
RPEFlow [59]clean5.089.0715.7725.24
final5.209.3815.6725.03
", + "bbox": [ + 516, + 395, + 787, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Point Tracking", + "text_level": 1, + "bbox": [ + 215, + 583, + 385, + 599 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of existing methods for long-term point tracking on BlinkVision. Similar to what we analyzed on optical flow, we analyze the robustness under large frame intervals and extreme illumination, and the gaps between existing image-based methods and event-based methods.", + "bbox": [ + 212, + 614, + 787, + 676 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experimental Setup. Although BlinkVision provides per-pixel dense annotations for point tracking, we find that existing methods are not efficient enough to process so much data. As a result, we grid sample the tracked pixels with a grid size of 20 pixels. We follow PointOdyssey [62] and use $\\delta_{\\mathrm{avg}}$ , median trajectory error (MTE) and survival rate as the evaluation metrics.", + "bbox": [ + 212, + 681, + 787, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. We benchmark several image-based methods in Table 4. Similarly, we observe that extreme lighting has a major impact on existing image-based point tracking methods. In Table 4 we can see that larger frame intervals (such as 4 and 8) severely degrade performance. However, for PIPs [21] and Context TAP [8], increasing the interval from 1 to 2 makes the performance slightly better. We", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ba6d1424830e6c044829ff1099bf4ed4314cc6538e3df16388db906c4f3cdf25.jpg", + "table_caption": [ + "Table 8: Cross-dataset evaluation of optical flow methods fine-tuned on BlinkVision. \"BV\" denotes BlinkVision." + ], + "table_footnote": [], + "table_body": "
RGB
MethodDataSintel (clean)Sintel (final)KITTI
RAFT [56]- +BV2.08 1.693.41 3.045.10 4.66
Events
MethodDataE- BlenderFlying- ObjectsE- Tartan
E-RAFT [16]- +BV2.66 1.922.9 2.673.27 2.55
E-FlowFormer [34]- +BV2.38 1.752.89 2.582.91 2.36
Events + RGB
MethodDataE- BlenderFlying- ObjectsE- Tartan
DCEIFlow [58]- +BV8.96 4.769.58 2.377.44 2.88
", + "bbox": [ + 215, + 196, + 521, + 420 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5929037ea47c8af4838bdc7db04b7d3c44cf9ee5ac0daa9e2dfce66fc9cb2674.jpg", + "table_caption": [ + "Table 9: Performance evaluation for typical categories at each difficult level. \"FF\" denotes FlowFormer [26] and \"FF++\" denotes FlowFormer++ [55]. We use the EPE as the evaluation metric." + ], + "table_footnote": [], + "table_body": "
DifficultCategoryRAFT [56]FF [26]FF++ [55]
Hardperson9.828.728.61
tree9.799.7610.10
rhinoceros8.907.076.78
Mediumbench4.293.603.68
globe3.432.522.51
lantern2.441.921.95
Easynotebook1.080.570.84
coaster0.860.870.79
painting0.310.300.26
", + "bbox": [ + 529, + 241, + 784, + 419 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "guess it might be the limited temporal receptive field of these two methods (only 8 frames) that makes the accumulated error quickly increase when the stride is too small. Some qualitative results are shown in Fig. 6. BlinkVision contains many challenging cases that cannot be handled by the SOTA methods.", + "bbox": [ + 212, + 449, + 782, + 511 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We show the performance of event-based methods in Table 5. We find that event-based approaches perform particularly poorly. In addition to the reasons for insufficient training data and model design, we deduce that the receptive field of event-based point-tracking methods is relatively small and therefore cannot perform well on the task of tracking arbitrary points. To verify the claim, we replace the input of grid sampled positions with ORB [52] feature points, denoted by “*” in the table. The new results perform even better, validating our ideas. Although the new results cannot be strictly compared with image-based methods, it performs better under extreme frame interval, i.e., “Stride-8”, which shows the large potential of event-based methods.", + "bbox": [ + 212, + 516, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Scene Flow", + "text_level": 1, + "bbox": [ + 215, + 696, + 354, + 709 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Experimental Setup. We use the same pairs as the optical flow benchmark and we follow [38,57] to use 2D EPE and 3D EPE as the evaluation metrics.", + "bbox": [ + 212, + 727, + 782, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. We benchmark two state-of-the-art image-based methods, i.e., RAFT-3D [57] and CamliFlow [38] as shown in Table 6 and show the results of event-based methods in Table 7. The conclusions of the optical flow benchmark apply to the scene flow task and present similar challenges for existing methods in this area. We show qualitative results in the supplementary materials.", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BlinkVision: A Benchmark for Correspondence Estimation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Cross-dataset Evaluation", + "text_level": 1, + "bbox": [ + 215, + 146, + 467, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We also fine-tune existing optical flow methods on the BlinkVision training set and then evaluate them on existing representative benchmarks. The results are shown in Table 8. We observe that fine-tuning on BlinkVision brings significant improvement for both image-based methods (2.08 vs. 1.69) and event-based methods (2.38 vs. 1.75 and 9.58 vs. 2.37), which demonstrates the vast diversity of BlinkVision boosts the generalizability of these methods.", + "bbox": [ + 212, + 179, + 787, + 270 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Performance Distribution on Categories", + "text_level": 1, + "bbox": [ + 215, + 297, + 591, + 314 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Previous methods in optical flow mainly perform evaluations on biased and limited scenarios, which is not comprehensive and robust enough to demonstrate the ability of the methods for different categories of objects in different scenarios. Thanks to the vast diversity of data and semantic labels provided by BlinkVision, for the first time, we analyze the performance distribution of several image-based optical flow methods on different categories. The results are shown in Fig. 7. The average curve is imbalanced: hard categories usually include complex shapes (e.g., hammocks and shrubs) or with deformable motion (e.g., persons and animals). We thus split the categories into three levels of \"difficulty\" based on the average curve, and the performance evaluation for typical categories at each level is presented in Table 9. We believe such fine-grained analysis helps understand the generalization capacity of the methods.", + "bbox": [ + 212, + 330, + 787, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 542, + 359, + 559 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose BlinkVision, a large-scale diversified benchmark for three types of correspondence tasks, i.e., optical flow, point tracking, and scene flow estimation using RGB frames and events. Extensive benchmarks on BlinkVision point to new challenges for existing image-based approaches and show that existing event-based approaches are far from fully unlocking the potential of event cameras. BlinkVision reveals new observations, challenges, and opportunities for future research into more robust visual systems such as human vision.", + "bbox": [ + 212, + 582, + 787, + 688 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment. This project was funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR, by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. This work was also partially supported by NSF of China (No. 61932003).", + "bbox": [ + 212, + 695, + 787, + 801 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. ActorCore. https://actorcore.realusion.com/, accessed: 2023-11-17 5", + "2. Blender. https://www.blender.org/, accessed: 2023-11-17 5", + "3. Evermotion Archinteriors Collection. https://evermotion.org/, accessed: 2023-11-17 5", + "4. Mixamo. https://www MIXamo.com/, accessed: 2023-11-17 5", + "5. Alzugaray, I., Chli, M.: Asynchronous multi-hypothesis tracking of features with event cameras. In: 2019 International Conference on 3D Vision (3DV). pp. 269-278. IEEE (2019) 11", + "6. Alzugaray, I., Chli, M.: Haste: multi-hypothesis asynchronous speeded-up tracking of events. In: 31st British Machine Vision Virtual Conference (BMVC 2020). p. 744. ETH Zurich, Institute of Robotics and Intelligent Systems (2020) 11", + "7. Barron, J.L., Fleet, D.J., Beauchemin, S.S.: Performance of optical flow techniques. International journal of computer vision 12, 43-77 (1994) 4", + "8. Bian, W., Huang, Z., Shi, X., Dong, Y., Li, Y., Li, H.: Context-TAP: Tracking Any Point Demands Spatial Context Features. arXiv preprint arXiv:2306.02000 (2023) 2, 11, 12", + "9. Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12. pp. 611-625. Springer (2012) 3, 4, 5", + "0. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An Information-rich 3D Model Repository. arXiv preprint arXiv:1512.03012 (2015) 3, 9", + "1. Ding, Z., Zhao, R., Zhang, J., Gao, T., Xiong, R., Yu, Z., Huang, T.: Spatiotemporal Recurrent Networks for Event-based Optical Flow Estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 525-533 (2022) 10", + "2. Doersch, C., Gupta, A., Markeeva, L., Recasens, A., Smaira, L., Aytar, Y., Carreira, J., Zisserman, A., Yang, Y.: Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems 35, 13610–13626 (2022) 3, 4", + "3. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020) 1, 6", + "4. Gehrig, D., Gehrig, M., Hidalgo-Carrio, J., Scaramuzza, D.: Video to Events: Recycling Video Datasets for Event Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3586-3595 (2020) 5", + "5. Gehrig, M., Aarents, W., Gehrig, D., Scaramuzza, D.: DSEC: A Stereo Event Camera Dataset for Driving Scenarios. IEEE Robotics and Automation Letters 6(3), 4947-4954 (2021) 2, 3, 5", + "6. Gehrig, M., Millhäsler, M., Gehrig, D., Scaramuzza, D.: E-RAFT: Dense Optical Flow from Event Cameras. In: Proceedings of the International Conference on 3D Vision. pp. 197–206. IEEE (2021) 10, 11, 13", + "7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012) 3, 4", + "8. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., et al.: Kubric: A scalable dataset" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BlinkVision: A Benchmark for Correspondence Estimation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3749-3761 (2022) 3, 5", + "19. Grossberg, M.D., Nayar, S.K.: What is the space of camera response functions? In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. vol. 2, pp. II-602. IEEE (2003) 6", + "20. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) 3, 8, 9", + "21. Harley, A.W., Fang, Z., Fragkiadaki, K.: Particle video revisited: Tracking through occlusions using point trajectories. In: European Conference on Computer Vision. pp. 59-75. Springer (2022) 4, 11, 12", + "22. Hidalgo-Carrió, J., Gallego, G., Scaramuzza, D.: Event-aided direct sparse odometry. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5781-5790 (2022) 5", + "23. Hu, J., Chen, X., Feng, B., Li, G., Yang, L., Bao, H., Zhang, G., Cui, Z.: Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095 (2024) 9", + "24. Hu, J., Mao, M., Bao, H., Zhang, G., Cui, Z.: Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems 36 (2024) 9", + "25. Huang, Z., Pan, X., Pan, W., Bian, W., Xu, Y., Cheung, K.C., Zhang, G., Li, H.: Neuralmarker: A framework for learning general marker correspondence. ACM Transactions on Graphics (TOG) 41(6), 1-10 (2022) 9", + "26. Huang, Z., Shi, X., Zhang, C., Wang, Q., Cheung, K.C., Qin, H., Dai, J., Li, H.: Flowformer: A Transformer Architecture for Optical Flow. In: Proceedings of the European Conference on Computer Vision. pp. 668-685. Springer (2022) 10, 13", + "27. Huang, Z., Shi, X., Zhang, C., Wang, Q., Li, Y., Qin, H., Dai, J., Wang, X., Li, H.: FlowFormer: A Transformer Architecture and Its Masked Cost Volume Autoencoding for Optical Flow. arXiv preprint arXiv:2306.05442 (2023) 2", + "28. Huang, Z., Zhou, H., Li, Y., Yang, B., Xu, Y., Zhou, X., Bao, H., Zhang, G., Li, H.: Vs-net: Voting with segmentation for visual localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6101-6111 (2021) 2", + "29. Jiang, S., Campbell, D., Lu, Y., Li, H., Hartley, R.: Learning to estimate hidden motions with global motion aggregation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9772-9781 (2021) 10", + "30. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 4", + "31. Klein, G., Murray, D.: Parallel tracking and mapping for small ar workspaces. In: 2007 6th IEEE and ACM international symposium on mixed and augmented reality. pp. 225-234. IEEE (2007) 9", + "32. Klenk, S., Chui, J., Demmel, N., Cremers, D.: Tum-vie: The tum stereo visual-inertial event dataset. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8601-8608. IEEE (2021) 2", + "33. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012) 3", + "34. Li, Y., Huang, Z., Chen, S., Shi, X., Li, H., Bao, H., Cui, Z., Zhang, G.: Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv preprint arXiv:2303.07716 (2023) 2, 3, 5, 7, 10, 11, 13" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Li, Y., Liu, X., Dong, W., Zhou, H., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Deltar: Depth estimation from a light-weight tof sensor and rgb image. In: European conference on computer vision. pp. 619-636. Springer (2022) 9", + "36. Li, Y., Zhou, H., Yang, B., Zhang, Y., Cui, Z., Bao, H., Zhang, G.: Graph-based asynchronous event processing for rapid object recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 934-943 (2021) 1", + "37. Lin, S., Ma, Y., Guo, Z., Wen, B.: DVS-Voltmeter: Stochastic Process-Based Event Simulator for Dynamic Vision Sensors. In: Proceedings of the European Conference on Computer Vision. pp. 578-593. Springer (2022) 7", + "38. Liu, H., Lu, T., Xu, Y., Liu, J., Li, W., Chen, L.: Camliflow: Bidirectional cameralidar fusion for joint optical flow and scene flow estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5791-5801 (2022) 12, 13", + "39. Liu, X., Li, Y., Teng, Y., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. In: Proceedings of the IEEE/cvf international conference on computer vision. pp. 1-11 (2023) 9", + "40. Liu, Y.L., Lai, W.S., Chen, Y.S., Kao, Y.L., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Single-image hdr reconstruction by learning to reverse the camera pipeline. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1651-1660 (2020) 6", + "41. Luo, J., Huang, Z., Li, Y., Zhou, X., Zhang, G., Bao, H.: Niid-net: adapting surface normal knowledge for intrinsic image decomposition in indoor scenes. IEEE Transactions on Visualization and Computer Graphics 26(12), 3434-3445 (2020) 9", + "42. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4040-4048 (2016) 3, 4, 5", + "43. Mehl, L., Schmalfuss, J., Jahedi, A., Nalivayko, Y., Bruhn, A.: Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4981-4991 (2023) 5", + "44. Messikommer, N., Fang, C., Gehrig, M., Scaramuzza, D.: Data-driven feature tracking for event cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5642-5651 (2023) 11", + "45. Milner, D., Goodale, M.: The visual brain in action, vol. 27. OUP Oxford (2006) 2", + "46. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research 36(2), 142-149 (2017) 5", + "47. Ni, J., Li, Y., Huang, Z., Li, H., Bao, H., Cui, Z., Zhang, G.: PATS: Patch Area Transportation with Subdivision for Local Feature Matching. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 17776-17786 (2023) 10", + "48. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019) 2" + ], + "bbox": [ + 212, + 146, + 787, + 825 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BlinkVision: A Benchmark for Correspondence Estimation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016) 4", + "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 8", + "51. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: An Open Event Camera Simulator. In: Proceedings of the Conference on Robot Learning. pp. 969-982. PMLR (2018) 7", + "52. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: Orb: An efficient alternative to sift or surf. In: 2011 International conference on computer vision. pp. 2564-2571. IEEE (2011) 11, 13", + "53. Rueckauer, B., Delbruck, T.: Evaluation of Event-based Algorithms for Optical Flow with Ground-truth from Inertial Measurement Sensor. Frontiers in neuroscience 10, 176 (2016) 2, 4", + "54. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016) 5", + "55. Shi, X., Huang, Z., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Flowformer++: Masked Cost Volume Autoencoding for Pretraining Optical Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1599-1610 (2023) 9, 10, 11, 12, 13", + "56. Teed, Z., Deng, J.: RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In: Proceedings of the European Conference on Computer Vision. pp. 402-419. Springer (2020) 10, 11, 13", + "57. Teed, Z., Deng, J.: Raft-3d: Scene flow using rigid-motion embeddings. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8375–8384 (2021) 12, 13", + "58. Wan, Z., Dai, Y., Mao, Y.: Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing 31, 7237-7251 (2022) 10, 13", + "59. Wan, Z., Mao, Y., Zhang, J., Dai, Y.: Rpeflow: Multimodal fusion of rgbpointcloud-event for joint optical flow and scene flow estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10030-10040 (2023) 3, 5, 12", + "60. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023) 8", + "61. Yang, B., Huang, Z., Li, Y., Zhou, H., Li, H., Zhang, G., Bao, H.: Hybrid3d: learning 3d hybrid features with point clouds and multi-view images for point cloud registration. Science China Information Sciences 66(7), 172101 (2023) 2", + "62. Zheng, Y., Harley, A.W., Shen, B., Wetzstein, G., Guibas, L.J.: Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19855-19865 (2023) 3, 4, 11, 12", + "63. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The Multivehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters 3(3), 2032-2039 (2018) 2" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 17 + }, + { + "type": "ref_text", + "text": "64. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In: Kress-Gazit, H., Srinivasa, S.S., Howard, T., Atanasov, N. (eds.) Robotics: Science and Systems (2018) 3, 5", + "bbox": [ + 217, + 146, + 784, + 189 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BlinkVision: A Benchmark for Correspondence Estimation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_model.json b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8226f9edaba9e6f57e4926151c3248a9e2aa6009 --- /dev/null +++ b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_model.json @@ -0,0 +1,2625 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.224, + 0.141, + 0.78, + 0.207 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.234, + 0.782, + 0.283 + ], + "angle": 0, + "content": "Yijin Li\\(^{1,2*}\\), Yichen Shen\\(^{1*}\\), Zhaoyang Huang\\(^{2*}\\), Shuo Chen\\(^{1}\\), Weikang Bian\\(^{3}\\), Xiaoyu Shi\\(^{3}\\), Fu-Yun Wang\\(^{3}\\), Keqiang Sun\\(^{3}\\), Hujun Bao\\(^{1}\\), Zhaopeng Cui\\(^{1}\\), Guofeng Zhang\\(^{1\\dagger}\\), and Hongsheng Li\\(^{3,4,5\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.292, + 0.703, + 0.308 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Lab of CAD&CG, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.308, + 0.703, + 0.321 + ], + "angle": 0, + "content": "\\(^{2}\\) Avolution AI \\(^{3}\\) CUHK MMLab \\(^{4}\\) Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.426, + 0.321, + 0.576, + 0.334 + ], + "angle": 0, + "content": "5 CPII under InnoHK" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.373, + 0.744, + 0.651 + ], + "angle": 0, + "content": "Abstract. Recent advances in event-based vision suggest that they complement traditional cameras by providing continuous observation without frame rate limitations and high dynamic range which are well-suited for correspondence tasks such as optical flow and point tracking. However, so far there is still a lack of comprehensive benchmarks for correspondence tasks with both event data and images. To fill this gap, we propose BlinkVision, a large-scale and diverse benchmark with rich modality and dense annotation of correspondence. BlinkVision has several appealing properties: 1) Rich modalities: It encompasses both event data and RGB images. 2) Rich annotations: It provides dense per-pixel annotations covering optical flow, scene flow, and point tracking. 3) Large vocabulary: It incorporates 410 daily categories, sharing common classes with widely-used 2D and 3D datasets such as LVIS and ShapeNet. 4) Naturalistic: It delivers photorealism data and covers a variety of naturalistic factors such as camera shake and deformation. BlinkVision enables extensive benchmarks on three types of correspondence tasks (i.e., optical flow, point tracking and scene flow estimation) for both image-based methods and event-based methods, leading to new observations, practices, and insights for future research. The benchmark website is https://www.blinkvision.net/." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.683, + 0.377, + 0.699 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.788, + 0.799 + ], + "angle": 0, + "content": "Modern image-based computer vision technology still cannot match the accuracy and robustness of human vision in many areas. One possible reason is that traditional cameras suffer from motion blur and limited frame rates, and they often rely on well-lighted conditions. In contrast, event cameras [13, 36] detect changes in intensity at each pixel as a stream of asynchronous events, which" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.807, + 0.709, + 0.821 + ], + "angle": 0, + "content": "* Yijin Li, Yichen Shen, and Zhaoyang Huang contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.821, + 0.633, + 0.834 + ], + "angle": 0, + "content": "† Hongsheng Li and Guofeng Zhang are the corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.807, + 0.709, + 0.834 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.782, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 1: BlinkVision is a large-scale and diverse benchmark with rich modality and dense annotation of correspondence. It covers 410 daily categories, sharing common classes with popular 2D and 3D datasets. The per-category object distributions, scene structure hierarchy, data samples, and supported applications of BlinkVision are shown in this figure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.584 + ], + "angle": 0, + "content": "eliminates frame rate limitations and enables operation within a high dynamic range. However, they can not capture fine-grained details as traditional cameras do. Looking at how human vision works, we find that two cells in the human retina, i.e., cones and rods, work similarly to these two types of cameras, respectively. According to the duplex theory of vision [45], which posits that rods and cones serve different functions, their combination ensures the robustness of human visual processing. Therefore, we believe that combining the advantages of traditional cameras and event cameras can significantly enhance computer vision systems, providing more comprehensive and adaptive vision capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.711 + ], + "angle": 0, + "content": "However, there are currently only a few benchmarks [15, 32, 34] providing both event data and RGB frames, which hinder the development of algorithms that fully exploit the event data and fuse information from both modalities. This shortfall is particularly prominent in the domain of pixel correspondence estimation [8, 27, 28, 61], i.e., optical flow, scene flow, and point tracking estimation. Previous benchmarks built for pixel correspondence are either highly biased to specific scenes [15, 63] or rather simple [53]. A primary factor is that obtaining such precise pixel-wise annotations for these tasks is expensive." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.825 + ], + "angle": 0, + "content": "To boost the research in this area, we present BlinkVision, a synthetic benchmark for optical flow, scene flow and point tracking estimation using RGB frames and events. Our dataset has several appealing properties: 1) Rich modalities: BlinkVision encompasses three visual modalities: final RGB images, clean RGB images and event data. The final RGB images reflect real-world challenges like motion blur and limited dynamic range while the clean RGB images are devoid of such imperfections. The clean RGB images can be seen as the latent images [48]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: A comparison between BlinkVision and other widely-used benchmarks on pixel correspondence estimation. \"Occ\", \"Chara\" and \"Cats\" is the abbreviation for occlusion, character and categories, respectively. \\( R^{lvis} \\) denotes the ratio of the 1.2k LVIS [20] categories being covered. \"N/A\" denotes the dataset does not provide category labels." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.226, + 0.782, + 0.376 + ], + "angle": 0, + "content": "
DatasetsEventCleanFinalOptical FlowScene FlowPoint TrackingOccCharaAnimalCatsR1vis(%)
MVSEC [64]✓X✓✓XXXXXN/A0
DSEC [15]✓X✓✓XXXXXN/A0
BlinkFlow [34]✓✓X✓XX✓✓X554.1
EKubric [59]✓✓X✓✓✓✓XX170.9
Sintel [9]X✓X✓XX✓✓✓N/A0
KITTI [17]XX✓✓✓XXXXN/A0
FlyingThings [42]X✓X✓✓X✓XX554.1
Kubric [18]X✓X✓✓✓✓XX170.9
TAP-Vid [12]X✓XXX✓X✓XN/A0
PointOdyssey [62]X✓XXX✓✓✓✓N/A0
BlinkVision (Ours)✓✓✓✓✓✓✓✓✓41033.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.788, + 0.596 + ], + "angle": 0, + "content": "of event cameras. 2) Rich annotations: It provides annotations covering optical flow, scene flow, and point tracking. Unlike benchmarks [15,17,62] which only contain sparse annotations, our data provide dense per-pixel annotations of each image, covering objects including moving cars, deformable characters, and animals. 3) Large vocabulary: It incorporates 410 daily categories, sharing common classes with widely-used 2D and 3D datasets such as ImageNet [33], LVIS [20], and ShapeNet [10], as depicted in Fig. 1. To our knowledge, our data have the widest variety of objects in existing pixel correspondence benchmarks. This expansive vocabulary is pivotal, enabling rigorous exploration of algorithmic generalization across varied objects. 4) Naturalistic: BlinkVision employs high-quality assets and rendering tools and thus is able to deliver photorealism data. Besides, it covers a variety of naturalistic factors such as camera shake and deformation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To make full use of BlinkVision, we set up a public benchmark website that allows uploading results and provides a public leaderboard. Besides, we evaluated both existing image-based and event-based methods on three typical correspondence tasks (i.e. optical flow, point tracking, and scene flow estimation). The results reveal new observations and challenges and serve as the baseline for future approaches. Specifically, we first study the robustness of existing image-based methods under large frame intervals and extreme illumination and point out the new challenge for these methods. Second, the benchmark results on existing event-based methods show that current methods do not fully unleash the potential of event cameras. Third, we show that fine-tuning existing methods on the training set of BlinkVision significantly boosts the generalizability, demonstrating the vast diversity of BlinkVision. Finally, the broad diversity and accessible category labels in BlinkVision allow us, for the first time, to analyze the performance of correspondence tasks on different categories. We believe BlinkVision has the potential to serve as a cornerstone benchmark for advancing the development of more robust computer vision systems." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.784, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.294, + 0.742, + 0.309 + ], + "angle": 0, + "content": "Fig. 2: Scenes samples in the proposed BlinkVision benchmark." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.325, + 0.388, + 0.341 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.799, + 0.394 + ], + "angle": 0, + "content": "A comprehensive overview and comparison between these benchmarks and BlinkVision is listed in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.613 + ], + "angle": 0, + "content": "Image-based optical flow and scene flow benchmarks. Early evaluations depended on synthetic datasets, like the famous \"Yosemite\" sequence [7]. MPI Sintel [9] was one of the most representative synthetic benchmarks derived from a short open-source animated 3D movie and it became one of the most popular benchmark datasets. KITTI [17] was almost the most well-known among real data. It computed ground truth for static scenes through the data from a 3D laser scanner and the ego-motion data of the car. In a later version, KITTI extended the ground truth to rigidly moving cars by fitting CAD models of cars. FlyingThings3D [42] was a more recent synthetic dataset. It contained the ground truth of both optical flow and scene flow. FlyingThings3D and KITTI were the two most commonly used benchmarks in scene flow evaluation. However, on both datasets, ground truth about scene flow was not available in occluded regions. Furthermore, on KITTI, ground truth for foreground points was either missing or approximated by a fitted car CAD model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.802 + ], + "angle": 0, + "content": "Image-based point tracking benchmarks. The first point-tracking dataset was FlyingThings++ [21], which was based on FlyingThings3D. It was initially designed for training the network. Due to the lack of evaluation benchmarks, it was also used for evaluation in the early stages. Later, Doersch et al. [12] proposed a real data benchmark named TAP-Vid, which was based on two real-world datasets: DAVIS [49] and Kinetics [30]. TAP-Vid relied on manual annotation and could not handle occlusions. PointOdyssey [62] was a newly proposed benchmark that was based on synthetic data and therefore could provide dense ground truth. However, it did not guarantee that every pixel had ground truth because it only tracked the mesh vertices of the object. Furthermore, the naturalism of PointOdyssey was limited to its interior parts. Its outdoor portion had no realistic layout, just a skybox with randomly dropped objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.808, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Event-based Benchmarks. Early event-based optical flow benchmarks [53] limited the camera to only rotational motion and inferred ground truth from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.404 + ], + "angle": 0, + "content": "the rotational motion of the camera. Two later benchmarks, DSEC [15] and MVSEC [64] computed the ground truth through LiDAR SLAM. Similar to KITTI, these two benchmarks limited the ground truth to the static elements of the scene. Besides, they had a rather limited motion pattern. More recently, Li et al. [34] proposed a large-scale diversified form synthetic benchmark named BlinkFlow, which presented more challenging scenes and complex motion patterns. As for the scene flow benchmark, Wan et al. [59] recently converted the existing FlyingThings [42] and Kubric [18] datasets to event datasets through video-to-event [14] technology. However, due to the limited frame rates in the original video data, this step inevitably produced artifacts in the event data. Wan et al. also extended the real-world event dataset DSEC with scene flow ground truth. As for the event-based point-tracking benchmark, existing methods usually employed Structure-from-Motion [54] (SfM) technology to associate keypoints across multiple frames to obtain correspondence ground truth. Event Camera Dataset [46] and EDS dataset [22] were two common benchmarks in this area. Limited by the sparsity and accuracy of SfM, these benchmarks did not support the evaluation of arbitrary point tracking." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.431, + 0.446, + 0.447 + ], + "angle": 0, + "content": "3 BlinkVision Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.543 + ], + "angle": 0, + "content": "In this section, we describe how to build BlinkVision, including the scene setup, data rendering, and generation for the ground truth labels. The process is based on Blender [2] as it provides photorealistic rendering and a flexible data interface that allows us to obtain customized correspondence ground truth. At the last, we introduce the statistics and distribution of the BlinkVision data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.377, + 0.586 + ], + "angle": 0, + "content": "3.1 Scene Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Previous synthetic benchmarks [9, 43] generally rely on open-source movies to avoid heavy scene construction. However, these movies are biased and do not cover diverse enough scenarios. In order to establish a comprehensive evaluation benchmark, we manually assemble a collection of scenarios that are as rich and diverse as possible. We first look for ready-made scenes that are photorealistic and visually diverse. Specifically, we purchased 40 indoor scenes and 13 outdoor scenes from Evermotion Archinteriors Collection [3] that cover common scenes such as living rooms, kitchens, offices, bedrooms, restaurants and gardens. The original scene is static. To enhance the realism, we procured 29 scanned and high-fidelity human bodies from ActorCore [1] and 88 artist-designed animals, together with 104 free characters from Mixamo [4]. These assets are rigged. We re-targeted them to a wide range of motion models (e.g., more than 100 human motions including dancing, walking, talking, etc.) and placed them in various locations. The human's motions mainly come from motion capture [1] and the animal's motions are designed by the artist to closely emulate their natural motions. Furthermore, we built 50 additional outdoor scenes that cover cases such" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "as a drone flying through the forest and a camera following a human or car from the crowd, which are common but rarely included in existing correspondence benchmarks. We set the camera trajectory by referencing the shooting trajectories of real-world videos, including handheld shots, car shots, and drone shots that cover non-uniform motion such as sudden stops and sharp turns. Finally, we obtained 80 indoor sequences and 63 outdoor sequences, where 56 indoor sequences and 47 outdoor sequences are for testing and the others are used for fine-tuning. Some samples of our data are shown in Fig. 2. In the supplementary, we also provide video samples. All the assets used in the testing and fine-tuning are totally disjoint, even including trees." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.325, + 0.525, + 0.34 + ], + "angle": 0, + "content": "3.2 Multi-modality Data Rendering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.446 + ], + "angle": 0, + "content": "We use Blender to render brightness \\( B \\) in linear color space. The following simulations of real-world image capturing and event data are both based on the linear color space. We employ tone mapping and gamma correction on \\( B \\) to obtain commonly used sRGB images. We call these sRGB images \"RGB (clean)\" because they do not suffer from effects like blurs or overexposed. On the contrary, we call the simulated real-world image capture \"RGB (final)\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.789, + 0.693 + ], + "angle": 0, + "content": "Simulation of Real-world Image Capturing. Real-world image capture suffers from motion blur and limited dynamic range. While Blender supports the simulation of the former, the latter needs additional processing. The latter effect usually leads to overexposed or underexposed. To simulate it, we follow previous work [40] and employ the following steps: (1) Random exposure ratio. We uniformly sample the exposure ratio \\( w \\) in the log2 space within \\([-3, 3]\\) and multiply it by \\( B \\) to simulate underexposed or overexposed, which gives us the augmented HDR image \\( H = B \\times 2^w \\). (2) Dynamic range clipping. We clip \\( H \\) according to the formulation \\( \\mathcal{C}(H) = \\min(H, 1) \\). This step leads to information loss for pixels in the overexposed regions. (3) Non-linear mapping. To align with how humans see a scene, a camera typically uses a non-linear camera response function (CRF) to modify the contrast of the captured image, which can be formulated by \\( I_n = \\mathcal{F}(I_c) \\). We randomly sample CRFs from an existing dataset [19]. (4) Quantization. The pixel values are quantized to 8 bits by \\( \\mathcal{Q}(I_n) = \\lfloor 255 \\times I_n + 0.5 \\rfloor / 255 \\). This step causes information loss in underexposed and smooth gradient areas." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Event Data. Event cameras work by responding to changes in the logarithmic brightness signal (i.e., \\( L = \\log B \\)) asynchronously and independently for each pixel [13]. An event is triggered when the change in brightness (either increase or decrease) since the last event at that pixel reaches a threshold of \\( \\pm T \\) (with \\( T > 0 \\)):" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.774, + 0.785, + 0.789 + ], + "angle": 0, + "content": "\\[\np _ {k} \\left(L \\left(u _ {k}, t _ {k}\\right) - L \\left(u _ {k}, t _ {k} - \\Delta t\\right)\\right) \\geq T, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where \\(\\Delta t\\) is the time since the last event triggered at pixel location \\(u_{k}\\) and at time \\(t_k\\). \\(p_k\\in \\{-1,1\\}\\) is the polarity of the brightness change. To model the event generation process, we need access to a continuous representation of the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.148, + 0.353, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.149, + 0.497, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.149, + 0.639, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.149, + 0.782, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.264, + 0.678, + 0.278 + ], + "angle": 0, + "content": "Fig. 3: Statistics of optical flow in BlinkVision." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.295, + 0.353, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.296, + 0.495, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.296, + 0.637, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.296, + 0.779, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 4: Statistics of point trajectories in BlinkVision. To save time, we sample the tracks using a grid with a size of 20. Trajectory segments are defined as contiguous sections of point trajectories, with interruptions caused by occlusion. The diameter is the maximum distance a point moves over time. We clip the trajectory diameter and divide it by the diagonal length of the image to obtain the ratio." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.646 + ], + "angle": 0, + "content": "visual signal for each pixel. In practice, it is approximated by rendering images at high frame rates for efficiency [34, 51]. Events are then synthesized based on frame-by-frame pixel differences. More specifically, we adaptively sample frames according to [34, 51] to ensure that the maximum pixel displacement between two sampling timestamps is bounded. In practice, we employ frame interpolation before simulating the event data to further reduce the pixel displacement. We use DVS-Voltmeter [37] to synthesize events because it can simulate complex noise effects (such as noise effects of temperature and parasitic photocurrent), thus generating realistic events." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.675, + 0.478, + 0.689 + ], + "angle": 0, + "content": "3.3 Ground Truth Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.708, + 0.787, + 0.768 + ], + "angle": 0, + "content": "Blender provides optical flow data between two consecutive frames and a segmentation mask for each object. However, we cannot directly obtain these data from Blender: (1) pixel correspondence across multiple frames; (2) scene flow between frames; (3) semantic category for each object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Pixel Correspondence. The data of (1) and (2) can be computed in a unified framework. Given camera poses \\(\\mathbf{T}\\) (camera-to-world), depth maps \\(\\mathbf{Z}\\) at frames \\(i\\) and \\(j\\), and the camera intrinsic \\(K\\), for the pixel location \\(\\mathfrak{u}\\) at frame \\(i\\), we first project it to world coordinate: \\(P = T_{i}Z_{i}(\\mathfrak{u})K^{-1}\\tilde{\\mathfrak{u}}\\), where \\(\\tilde{\\mathfrak{u}}\\) is the homogeneous" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.208 + ], + "angle": 0, + "content": "coordinates of u. Then we re-project the 3D point \\( P \\) to frame \\( j \\) through \\( du_{j} = KT_{j}^{-1}P \\), where \\( \\mathfrak{u}_j \\) is the corresponding pixel location and \\( d \\) is the depth value at frame \\( j \\). In this way, we get the pixel correspondence between any two frames and the corresponding depth, i.e., (1) and (2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.214, + 0.788, + 0.351 + ], + "angle": 0, + "content": "However, this only works for static objects. To handle dynamic objects and even those objects with deformation motion such as humans and animals, we bake the face index and barycentric coordinate \\((\\lambda_1,\\lambda_2,\\lambda_3)\\) of each mesh face into textures. After rendering the texture, we locate the triangular face that the tracked pixel belongs to (indexed by the rendered face index) and obtain the vertices' 3D position \\((V_{1},V_{2},V_{3})\\) of that face via Blender's API. The 3D position of tracked pixels can be computed through \\(P = \\lambda_1V_1 + \\lambda_2V_2 + \\lambda_3V_3\\). This step replaces the back projection for the static objects. It is worth noting that we need to triangulate every mesh face in the scenes before baking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Semantic Labeling. BlinkVision contains thousands of objects which makes manual semantic labeling expensive. To this end, we develop an automatic labeling framework that is based on open-vocabulary semantic understanding [50,60]. First, we pre-define a category list based on LVIS [20]. The category together with category descriptions is encoded into text embedding \\( F_{t} \\in \\mathbb{R}^{N \\times D} \\) with a pre-trained text encoder, i.e., CLIP, where \\( N \\) is the number of categories and \\( D \\) is the feature dimension. For each object asset, we render it individually to a \\( 224 \\times 224 \\) image. The process is rather fast while keeping photorealistic because it does not need to compute complex path tracing between objects. We use the pre-trained image encoder from CLIP [50] to extract the image embedding \\( F_{i} \\in \\mathbb{R}^{D} \\). The probability of the image belonging to one of the pre-defined categories is computed through \\( P(c|F_{i}) = \\text{softmax}\\left(F_{i} \\cdot (F_{t})^{\\top}\\right)_{c} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.411, + 0.584 + ], + "angle": 0, + "content": "3.4 Dataset Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.829 + ], + "angle": 0, + "content": "BlinkVision consists of 40 training sequences and 103 testing sequences. The training set provides 107,880 frame pairs for optical flow and scene flow, and 1025 sub-sequences for point tracking. The frame rate for RGB images and ground truth data is 20 FPS. The frame rate is not practically significant because we adjusted it to ensure sufficient motion between adjacent frames. This adjustment reduces the data size uploaded to the benchmark website. For each sequence of BlinkVision, we selected the sub-sequences by starting tracking from the first frame. Once the overlap with the reference frame (i.e., how many tracked points remain inside the image) is smaller than \\(40\\%\\), we start to track a new reference frame. If the overlap is smaller than \\(20\\%\\), we stop this track. Additionally, we discard short tracks that are less than 20 frames long because they are less challenging without long-term motion accumulation, which could result in the benchmark easily reaching saturation. As a result, we generate 1025 subsequences where each sub-sequence contains \\(640 \\times 480\\) dense point trajectories. For the test set, we selected 12,804 frame pairs for evaluating optical flow and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.348, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.206, + 0.348, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.266, + 0.337, + 0.277 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.147, + 0.45, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.206, + 0.45, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.266, + 0.43, + 0.277 + ], + "angle": 0, + "content": "Target Image" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.146, + 0.551, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.206, + 0.551, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.266, + 0.532, + 0.276 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.146, + 0.653, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.206, + 0.653, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.266, + 0.644, + 0.277 + ], + "angle": 0, + "content": "Before Finetuning" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.146, + 0.755, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.206, + 0.755, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.266, + 0.742, + 0.277 + ], + "angle": 0, + "content": "After Finetuning" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.289, + 0.785, + 0.317 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of FlowFormer++ [55] before and after fine-tuning on the training set of BlinkVision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.784, + 0.406 + ], + "angle": 0, + "content": "scene flow, and 865 sub-sequences for evaluating point tracking. BlinkVision includes raw data on depth, normal, camera pose, etc. At this stage, we focus on pixel-level motion tasks. The additional raw data will be released later, which will benefit other tasks such as depth estimation [35] and SLAM [23, 24, 39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.785, + 0.468 + ], + "angle": 0, + "content": "Table 1 summarizes key statistics of BlinkVision compared to related works. BlinkVision contains more than 40K 3D models in 410 categories. It shares some common categories with 2D and 3D datasets [10,20], and bears a huge diversity in semantics, geometry, and appearance, enabling a wide range of research topics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.787, + 0.742 + ], + "angle": 0, + "content": "We provide statistics of optical flow in Fig. 3. For benchmarking, we discard those pixels whose flow magnitudes are larger than half of the image diagonal. The latter situation usually occurs when the object is too close to the camera, and its magnitude can even approach infinity. As a result, the magnitude shown in Fig. 3 has an upper bound of 400. The broad distribution in Fig. 3 reveals the diverse motions in BlinkVision. Besides, the diverse distribution of \"X Variance\" and \"Y Variance\" also indicates the presence of complex object shapes in BlinkVision. We also provide statistics of point trajectories in Fig. 4. BlinkVision allows the evaluation of point tracking with different trajectories' lengths, from small to big, corresponding to different downstream applications like video editing [25, 41] or augmented reality [31, 39]. The broad distribution of \"occlusion ratio\" and \"number of segments\" indicates that BlinkVision brings diverse and challenging data. The occlusion ratio has an upper bound that is less than 100 percent due to our strategy of sub-sequences selection. To quantify the amount of motion a point undergoes, we calculate its trajectory diameter, which is the maximum distance between any two positions of the point over its entire trajectory. We observe that point motion in BlinkVision is diverse and nearly uniformly distributed across the entire image plane." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.764, + 0.364, + 0.78 + ], + "angle": 0, + "content": "4 Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We release all the data except for the ground truth for the test split. For a fair comparison, we create a public benchmark website with leaderboards. We also release the evaluation code for our online benchmarking." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.484, + 0.216 + ], + "angle": 0, + "content": "Table 2: Optical flow result of RGB-frame-based methods. \"St.\" denotes Stride. \"†\" denotes fine-tuning. \"FF\" denotes FlowFormer [26] and \"FF++\" denotes FlowFormer++ [55]." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.227, + 0.486, + 0.427 + ], + "angle": 0, + "content": "
MethodEPE - clean
St. 1St. 2St. 4St. 8Avg.
RAFT [56]1.834.6210.1820.469.27
GMA [29]1.824.379.5218.688.60
FF [26]1.603.777.5716.037.24
FF++ [55]1.543.577.4316.267.20
RAFT† [56]1.332.966.3912.725.85
FF++† [55]1.092.354.8810.584.73
MethodEPE - final
St. 1St. 2St. 4St. 8Avg.
RAFT [56]2.545.6911.2722.6710.54
GMA [29]2.565.7211.2121.3510.21
FF [26]2.264.809.2418.238.63
FF++ [55]2.284.839.3919.088.90
RAFT† [56]1.943.887.3515.177.08
FF++† [55]1.663.346.4212.736.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.49, + 0.145, + 0.791, + 0.187 + ], + "angle": 0, + "content": "Table 3: Optical flow result of event-based methods. \"St.\" denotes Stride. \"†\" denotes fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.492, + 0.2, + 0.792, + 0.427 + ], + "angle": 0, + "content": "
Event
Methods|St. 1St. 2St. 4St. 8Avg.
E-RAFT [16]2.817.0417.8228.6014.07
STE-FlowNet [11]2.595.9413.1341.8915.89
E-FlowFormer [34]2.415.6612.5330.4512.76
E-RAFT† [16]1.683.637.4814.206.75
E-FlowFormer† [34]1.513.005.9613.606.02
Event + RGB (clean)
Methods|St. 1St. 2St. 4St. 8Avg.
DCEIFlow [58]3.3114.0734.7861.5628.43
DCEIFlow† [58]2.196.3012.5426.1111.79
Event + RGB (final)
Methods|St. 1St. 2St. 4St. 8Avg.
DCEIFlow [58]4.0814.2434.5461.2728.53
DCEIFlow† [58]2.526.8613.6428.4512.87
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.456, + 0.37, + 0.471 + ], + "angle": 0, + "content": "4.1 Optical Flow" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.655 + ], + "angle": 0, + "content": "In this section, we comprehensively evaluate the performance of existing optical flow methods on BlinkVision. Specifically, we first analyze the robustness of image-based methods under large frame intervals and extreme illumination. Large motion, severe occlusion, and information loss in these extreme cases pose great challenges to existing image-based methods. In contrast, event cameras naturally depict continuous pixel motion and possess a large dynamic range, thus offering great potential for solving these challenges. We then benchmark existing event-based methods, including event-only methods and event-RGB fusion methods. We find that existing event-based methods cannot fully unleash the potential of event cameras. We analyze the possible reasons and point to possible opportunities for future research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.754 + ], + "angle": 0, + "content": "Experimental Setup. To study the impact of frame intervals on existing methods, we select frames separated by 1, 2, 4, and 8 from the reference frame as target frames. To reduce the file size of results that are uploaded to the benchmark website, we uniformly sample a reference frame from ten consecutive frames. As a result, there are 12,804 frame pairs for testing in total. We follow the previous work [26, 47, 56] and report the EPE (end-point-error) as the evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Results. We benchmark several image-based methods in Table 2. We make the following observations: 1) The error generally increases linearly with stride size. This is expected because doubling the stride yields motion that is twice as fast, resulting in an error that is also expected to be twice as high. 2) Extreme lighting has a heavy effect on existing image-based methods. In the supplementary" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.549, + 0.186 + ], + "angle": 0, + "content": "Table 4: Point tracking result of RGB-frame-based methods under different frame interval. \"St.\" denotes Stride." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.198, + 0.55, + 0.478 + ], + "angle": 0, + "content": "
RGB(clean)
MethodMetricSt. 1St. 2St. 4St. 8Avg.
PIPs [21]δavg37.7739.3436.3425.5034.74
Survival↑52.2855.3954.5547.5852.45
MTE↓50.7747.8149.5362.4252.63
PIPs++ [62]δavg40.0238.2530.2416.8431.34
Survival↑53.4453.1347.5739.7148.46
MTE↓45.5151.0259.8773.1757.39
Context-TAP [8]δavg38.7539.4736.5125.9935.18
Survival↑49.9853.0952.6146.1850.46
MTE↓51.4649.6151.8364.3954.32
RGB(final)
MethodMetricSt. 1St. 2St. 4St. 8Avg.
PIPs [21]δavg33.0234.4431.5621.7730.20
Survival↑48.3551.5650.7244.5648.80
MTE↓53.9451.5453.6566.1456.32
PIPs++ [62]δavg35.8334.2427.1115.3228.12
Survival↑50.4750.3645.3238.4646.15
MTE↓48.3753.7462.3474.5959.76
Context-TAP [8]δavg33.4634.1731.3922.0030.25
Survival↑45.5848.9148.5343.0646.52
MTE↓55.0953.7656.3368.1158.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.145, + 0.793, + 0.23 + ], + "angle": 0, + "content": "Table 5: Point tracking result of event-based methods. \\* denotes that we track ORB [52] feature points rather than grid sampled points. DET\" denotes DeepETracker [44]" + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.242, + 0.787, + 0.478 + ], + "angle": 0, + "content": "
Event
Methodδavg↑Survival↑MTE↓
AMH [5]18.9537.9875.44
HASTE [6]18.6037.4377.36
AMH* [5]26.7641.9362.48
HASTE* [6]27.6542.8762.52
Event + RGB (clean)
Methodδavg↑Survival↑MTE↓
DET [44]21.6235.4480.80
DET* [44]33.4648.8260.69
Event + RGB (final)
Methodδavg↑Survival↑MTE↓
DET [44]19.0533.5285.00
DET* [44]31.7848.6758.79
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.599 + ], + "angle": 0, + "content": "materials, we analyze how different exposure factors lead to performance loss. 3) Fine-tuning on the training split of BlinkVision significantly improves the performance. In Fig. 5 we show qualitative results of FlowFormer++ before and after fine-tuning on BlinkVision. After fine-tuning, the flow predictions are more precise and clear, especially near the object boundary. As shown in Table 2, the benchmark is still challenging for methods after fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.833 + ], + "angle": 0, + "content": "We report the evaluation of event-based methods in Table 3. It is surprising that the performance of existing event-based methods is worse than image-based methods and they also suffer from severe degradation under large frame intervals. The reasons are mainly twofold. First, existing training data for events lag behind that of RGB. This can be verified by the performance of E-RAFT [16] and E-FlowFormer [34] after fine-tuning. These two methods are far behind RAFT [56] before fine-tuning, but exceed the fine-tuned RAFT [56] and even FlowFormer++ [55] after fine-tuning (in \"Stride-1\", \"Stride-2\" and \"Stride-4\" of the \"final\" case), indicating that the poor performance of the former mainly comes from limited training on existing event-based training data. Second, existing event-based methods usually convert events to regular 3D voxel grids before processing, which quantizes the event data and leads to information loss. This problem is more serious when dealing with event data within large frame intervals. As a result, it calls for a more powerful representation of event data and new algorithms that can deal with long-range optical flow estimation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.207, + 0.323, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.268, + 0.31, + 0.278 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.145, + 0.426, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.207, + 0.426, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.268, + 0.418, + 0.278 + ], + "angle": 0, + "content": "Prediction Results" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.145, + 0.531, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.207, + 0.531, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.436, + 0.268, + 0.524, + 0.278 + ], + "angle": 0, + "content": "Error Visualization" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.159, + 0.78, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.53, + 0.331 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results of \\(\\mathbf{PIP}\\mathbf{s} + +\\) [62]. Darker lines in the right figure (e.g., red) indicate larger errors." + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.29, + 0.789, + 0.32 + ], + "angle": 0, + "content": "Fig. 7: Performance distribution of FlowFormer++ [55]." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.345, + 0.51, + 0.386 + ], + "angle": 0, + "content": "Table 6: Scene flow result of RGB-frame-based methods. \"St.\" denotes Stride." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.399, + 0.51, + 0.555 + ], + "angle": 0, + "content": "
EPE-2d
MethodRGBSt. 1St. 2St. 4St. 8
RAFT-3D [57]clean2.154.348.5916.08
final3.487.2910.7418.57
CamLiFlow [38]clean2.796.0512.3423.89
final4.288.2816.3030.61
EPE-3d
MethodRGBSt. 1St. 2St. 4St. 8
RAFT-3D [57]clean1.913.837.2512.18
final2.023.967.3914.21
CamLiFlow [38]clean1.452.765.369.12
final1.532.885.609.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.356, + 0.787, + 0.384 + ], + "angle": 0, + "content": "Table 7: Scene flow result of event-based methods. \"St.\" denotes Stride." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.396, + 0.789, + 0.544 + ], + "angle": 0, + "content": "
EPE-2d
MethodRGBSt. 1St. 2St. 4St. 8
RPEFlow [59]clean1.533.507.3617.44
final1.924.108.1518.38
EPE-3d
MethodRGBSt. 1St. 2St. 4St. 8
RPEFlow [59]clean5.089.0715.7725.24
final5.209.3815.6725.03
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.584, + 0.386, + 0.6 + ], + "angle": 0, + "content": "4.2 Point Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.677 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of existing methods for long-term point tracking on BlinkVision. Similar to what we analyzed on optical flow, we analyze the robustness under large frame intervals and extreme illumination, and the gaps between existing image-based methods and event-based methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.788, + 0.759 + ], + "angle": 0, + "content": "Experimental Setup. Although BlinkVision provides per-pixel dense annotations for point tracking, we find that existing methods are not efficient enough to process so much data. As a result, we grid sample the tracked pixels with a grid size of 20 pixels. We follow PointOdyssey [62] and use \\(\\delta_{\\mathrm{avg}}\\), median trajectory error (MTE) and survival rate as the evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Results. We benchmark several image-based methods in Table 4. Similarly, we observe that extreme lighting has a major impact on existing image-based point tracking methods. In Table 4 we can see that larger frame intervals (such as 4 and 8) severely degrade performance. However, for PIPs [21] and Context TAP [8], increasing the interval from 1 to 2 makes the performance slightly better. We" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.522, + 0.186 + ], + "angle": 0, + "content": "Table 8: Cross-dataset evaluation of optical flow methods fine-tuned on BlinkVision. \"BV\" denotes BlinkVision." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.198, + 0.522, + 0.421 + ], + "angle": 0, + "content": "
RGB
MethodDataSintel (clean)Sintel (final)KITTI
RAFT [56]- +BV2.08 1.693.41 3.045.10 4.66
Events
MethodDataE- BlenderFlying- ObjectsE- Tartan
E-RAFT [16]- +BV2.66 1.922.9 2.673.27 2.55
E-FlowFormer [34]- +BV2.38 1.752.89 2.582.91 2.36
Events + RGB
MethodDataE- BlenderFlying- ObjectsE- Tartan
DCEIFlow [58]- +BV8.96 4.769.58 2.377.44 2.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.145, + 0.785, + 0.228 + ], + "angle": 0, + "content": "Table 9: Performance evaluation for typical categories at each difficult level. \"FF\" denotes FlowFormer [26] and \"FF++\" denotes FlowFormer++ [55]. We use the EPE as the evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.242, + 0.785, + 0.42 + ], + "angle": 0, + "content": "
DifficultCategoryRAFT [56]FF [26]FF++ [55]
Hardperson9.828.728.61
tree9.799.7610.10
rhinoceros8.907.076.78
Mediumbench4.293.603.68
globe3.432.522.51
lantern2.441.921.95
Easynotebook1.080.570.84
coaster0.860.870.79
painting0.310.300.26
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.784, + 0.512 + ], + "angle": 0, + "content": "guess it might be the limited temporal receptive field of these two methods (only 8 frames) that makes the accumulated error quickly increase when the stride is too small. Some qualitative results are shown in Fig. 6. BlinkVision contains many challenging cases that cannot be handled by the SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.669 + ], + "angle": 0, + "content": "We show the performance of event-based methods in Table 5. We find that event-based approaches perform particularly poorly. In addition to the reasons for insufficient training data and model design, we deduce that the receptive field of event-based point-tracking methods is relatively small and therefore cannot perform well on the task of tracking arbitrary points. To verify the claim, we replace the input of grid sampled positions with ORB [52] feature points, denoted by “*” in the table. The new results perform even better, validating our ideas. Although the new results cannot be strictly compared with image-based methods, it performs better under extreme frame interval, i.e., “Stride-8”, which shows the large potential of event-based methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.697, + 0.355, + 0.71 + ], + "angle": 0, + "content": "4.3 Scene Flow" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.784, + 0.759 + ], + "angle": 0, + "content": "Experimental Setup. We use the same pairs as the optical flow benchmark and we follow [38,57] to use 2D EPE and 3D EPE as the evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Results. We benchmark two state-of-the-art image-based methods, i.e., RAFT-3D [57] and CamliFlow [38] as shown in Table 6 and show the results of event-based methods in Table 7. The conclusions of the optical flow benchmark apply to the scene flow task and present similar challenges for existing methods in this area. We show qualitative results in the supplementary materials." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.468, + 0.162 + ], + "angle": 0, + "content": "4.4 Cross-dataset Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.271 + ], + "angle": 0, + "content": "We also fine-tune existing optical flow methods on the BlinkVision training set and then evaluate them on existing representative benchmarks. The results are shown in Table 8. We observe that fine-tuning on BlinkVision brings significant improvement for both image-based methods (2.08 vs. 1.69) and event-based methods (2.38 vs. 1.75 and 9.58 vs. 2.37), which demonstrates the vast diversity of BlinkVision boosts the generalizability of these methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.299, + 0.592, + 0.315 + ], + "angle": 0, + "content": "4.5 Performance Distribution on Categories" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.789, + 0.514 + ], + "angle": 0, + "content": "Previous methods in optical flow mainly perform evaluations on biased and limited scenarios, which is not comprehensive and robust enough to demonstrate the ability of the methods for different categories of objects in different scenarios. Thanks to the vast diversity of data and semantic labels provided by BlinkVision, for the first time, we analyze the performance distribution of several image-based optical flow methods on different categories. The results are shown in Fig. 7. The average curve is imbalanced: hard categories usually include complex shapes (e.g., hammocks and shrubs) or with deformable motion (e.g., persons and animals). We thus split the categories into three levels of \"difficulty\" based on the average curve, and the performance evaluation for typical categories at each level is presented in Table 9. We believe such fine-grained analysis helps understand the generalization capacity of the methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.36, + 0.56 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.689 + ], + "angle": 0, + "content": "We propose BlinkVision, a large-scale diversified benchmark for three types of correspondence tasks, i.e., optical flow, point tracking, and scene flow estimation using RGB frames and events. Extensive benchmarks on BlinkVision point to new challenges for existing image-based approaches and show that existing event-based approaches are far from fully unlocking the potential of event cameras. BlinkVision reveals new observations, challenges, and opportunities for future research into more robust visual systems such as human vision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Acknowledgment. This project was funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR, by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. This work was also partially supported by NSF of China (No. 61932003)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.735, + 0.19 + ], + "angle": 0, + "content": "1. ActorCore. https://actorcore.realusion.com/, accessed: 2023-11-17 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.65, + 0.204 + ], + "angle": 0, + "content": "2. Blender. https://www.blender.org/, accessed: 2023-11-17 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.786, + 0.231 + ], + "angle": 0, + "content": "3. Evermotion Archinteriors Collection. https://evermotion.org/, accessed: 2023-11-17 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.645, + 0.245 + ], + "angle": 0, + "content": "4. Mixamo. https://www MIXamo.com/, accessed: 2023-11-17 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "5. Alzugaray, I., Chli, M.: Asynchronous multi-hypothesis tracking of features with event cameras. In: 2019 International Conference on 3D Vision (3DV). pp. 269-278. IEEE (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.786, + 0.328 + ], + "angle": 0, + "content": "6. Alzugaray, I., Chli, M.: Haste: multi-hypothesis asynchronous speeded-up tracking of events. In: 31st British Machine Vision Virtual Conference (BMVC 2020). p. 744. ETH Zurich, Institute of Robotics and Intelligent Systems (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.786, + 0.356 + ], + "angle": 0, + "content": "7. Barron, J.L., Fleet, D.J., Beauchemin, S.S.: Performance of optical flow techniques. International journal of computer vision 12, 43-77 (1994) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.786, + 0.396 + ], + "angle": 0, + "content": "8. Bian, W., Huang, Z., Shi, X., Dong, Y., Li, Y., Li, H.: Context-TAP: Tracking Any Point Demands Spatial Context Features. arXiv preprint arXiv:2306.02000 (2023) 2, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.786, + 0.453 + ], + "angle": 0, + "content": "9. Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12. pp. 611-625. Springer (2012) 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.786, + 0.495 + ], + "angle": 0, + "content": "0. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An Information-rich 3D Model Repository. arXiv preprint arXiv:1512.03012 (2015) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.786, + 0.55 + ], + "angle": 0, + "content": "1. Ding, Z., Zhao, R., Zhang, J., Gao, T., Xiong, R., Yu, Z., Huang, T.: Spatiotemporal Recurrent Networks for Event-based Optical Flow Estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 525-533 (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.786, + 0.592 + ], + "angle": 0, + "content": "2. Doersch, C., Gupta, A., Markeeva, L., Recasens, A., Smaira, L., Aytar, Y., Carreira, J., Zisserman, A., Yang, Y.: Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems 35, 13610–13626 (2022) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.786, + 0.647 + ], + "angle": 0, + "content": "3. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020) 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "4. Gehrig, D., Gehrig, M., Hidalgo-Carrio, J., Scaramuzza, D.: Video to Events: Recycling Video Datasets for Event Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3586-3595 (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.786, + 0.73 + ], + "angle": 0, + "content": "5. Gehrig, M., Aarents, W., Gehrig, D., Scaramuzza, D.: DSEC: A Stereo Event Camera Dataset for Driving Scenarios. IEEE Robotics and Automation Letters 6(3), 4947-4954 (2021) 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.786, + 0.771 + ], + "angle": 0, + "content": "6. Gehrig, M., Millhäsler, M., Gehrig, D., Scaramuzza, D.: E-RAFT: Dense Optical Flow from Event Cameras. In: Proceedings of the International Conference on 3D Vision. pp. 197–206. IEEE (2021) 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.812 + ], + "angle": 0, + "content": "7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "8. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., et al.: Kubric: A scalable dataset" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3749-3761 (2022) 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.216 + ], + "angle": 0, + "content": "19. Grossberg, M.D., Nayar, S.K.: What is the space of camera response functions? In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. vol. 2, pp. II-602. IEEE (2003) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.216, + 0.786, + 0.257 + ], + "angle": 0, + "content": "20. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) 3, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.786, + 0.298 + ], + "angle": 0, + "content": "21. Harley, A.W., Fang, Z., Fragkiadaki, K.: Particle video revisited: Tracking through occlusions using point trajectories. In: European Conference on Computer Vision. pp. 59-75. Springer (2022) 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.786, + 0.338 + ], + "angle": 0, + "content": "22. Hidalgo-Carrió, J., Gallego, G., Scaramuzza, D.: Event-aided direct sparse odometry. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5781-5790 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.786, + 0.379 + ], + "angle": 0, + "content": "23. Hu, J., Chen, X., Feng, B., Li, G., Yang, L., Bao, H., Zhang, G., Cui, Z.: Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095 (2024) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.786, + 0.417 + ], + "angle": 0, + "content": "24. Hu, J., Mao, M., Bao, H., Zhang, G., Cui, Z.: Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems 36 (2024) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.786, + 0.461 + ], + "angle": 0, + "content": "25. Huang, Z., Pan, X., Pan, W., Bian, W., Xu, Y., Cheung, K.C., Zhang, G., Li, H.: Neuralmarker: A framework for learning general marker correspondence. ACM Transactions on Graphics (TOG) 41(6), 1-10 (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.786, + 0.502 + ], + "angle": 0, + "content": "26. Huang, Z., Shi, X., Zhang, C., Wang, Q., Cheung, K.C., Qin, H., Dai, J., Li, H.: Flowformer: A Transformer Architecture for Optical Flow. In: Proceedings of the European Conference on Computer Vision. pp. 668-685. Springer (2022) 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.786, + 0.542 + ], + "angle": 0, + "content": "27. Huang, Z., Shi, X., Zhang, C., Wang, Q., Li, Y., Qin, H., Dai, J., Wang, X., Li, H.: FlowFormer: A Transformer Architecture and Its Masked Cost Volume Autoencoding for Optical Flow. arXiv preprint arXiv:2306.05442 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.786, + 0.595 + ], + "angle": 0, + "content": "28. Huang, Z., Zhou, H., Li, Y., Yang, B., Xu, Y., Zhou, X., Bao, H., Zhang, G., Li, H.: Vs-net: Voting with segmentation for visual localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6101-6111 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.786, + 0.637 + ], + "angle": 0, + "content": "29. Jiang, S., Campbell, D., Lu, Y., Li, H., Hartley, R.: Learning to estimate hidden motions with global motion aggregation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9772-9781 (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.786, + 0.678 + ], + "angle": 0, + "content": "30. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.786, + 0.718 + ], + "angle": 0, + "content": "31. Klein, G., Murray, D.: Parallel tracking and mapping for small ar workspaces. In: 2007 6th IEEE and ACM international symposium on mixed and augmented reality. pp. 225-234. IEEE (2007) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "32. Klenk, S., Chui, J., Demmel, N., Cremers, D.: Tum-vie: The tum stereo visual-inertial event dataset. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8601-8608. IEEE (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "33. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "34. Li, Y., Huang, Z., Chen, S., Shi, X., Li, H., Bao, H., Cui, Z., Zhang, G.: Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv preprint arXiv:2303.07716 (2023) 2, 3, 5, 7, 10, 11, 13" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Li, Y., Liu, X., Dong, W., Zhou, H., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Deltar: Depth estimation from a light-weight tof sensor and rgb image. In: European conference on computer vision. pp. 619-636. Springer (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "36. Li, Y., Zhou, H., Yang, B., Zhang, Y., Cui, Z., Bao, H., Zhang, G.: Graph-based asynchronous event processing for rapid object recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 934-943 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "37. Lin, S., Ma, Y., Guo, Z., Wen, B.: DVS-Voltmeter: Stochastic Process-Based Event Simulator for Dynamic Vision Sensors. In: Proceedings of the European Conference on Computer Vision. pp. 578-593. Springer (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.327 + ], + "angle": 0, + "content": "38. Liu, H., Lu, T., Xu, Y., Liu, J., Li, W., Chen, L.: Camliflow: Bidirectional cameralidar fusion for joint optical flow and scene flow estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5791-5801 (2022) 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.383 + ], + "angle": 0, + "content": "39. Liu, X., Li, Y., Teng, Y., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. In: Proceedings of the IEEE/cvf international conference on computer vision. pp. 1-11 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.438 + ], + "angle": 0, + "content": "40. Liu, Y.L., Lai, W.S., Chen, Y.S., Kao, Y.L., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Single-image hdr reconstruction by learning to reverse the camera pipeline. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1651-1660 (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.492 + ], + "angle": 0, + "content": "41. Luo, J., Huang, Z., Li, Y., Zhou, X., Zhang, G., Bao, H.: Niid-net: adapting surface normal knowledge for intrinsic image decomposition in indoor scenes. IEEE Transactions on Visualization and Computer Graphics 26(12), 3434-3445 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.549 + ], + "angle": 0, + "content": "42. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4040-4048 (2016) 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.605 + ], + "angle": 0, + "content": "43. Mehl, L., Schmalfuss, J., Jahedi, A., Nalivayko, Y., Bruhn, A.: Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4981-4991 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.646 + ], + "angle": 0, + "content": "44. Messikommer, N., Fang, C., Gehrig, M., Scaramuzza, D.: Data-driven feature tracking for event cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5642-5651 (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.66 + ], + "angle": 0, + "content": "45. Milner, D., Goodale, M.: The visual brain in action, vol. 27. OUP Oxford (2006) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.715 + ], + "angle": 0, + "content": "46. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research 36(2), 142-149 (2017) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.771 + ], + "angle": 0, + "content": "47. Ni, J., Li, Y., Huang, Z., Li, H., Bao, H., Cui, Z., Zhang, G.: PATS: Patch Area Transportation with Subdivision for Local Feature Matching. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 17776-17786 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.827 + ], + "angle": 0, + "content": "48. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019) 2" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.827 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "49. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.257 + ], + "angle": 0, + "content": "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.786, + 0.297 + ], + "angle": 0, + "content": "51. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: An Open Event Camera Simulator. In: Proceedings of the Conference on Robot Learning. pp. 969-982. PMLR (2018) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.787, + 0.338 + ], + "angle": 0, + "content": "52. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: Orb: An efficient alternative to sift or surf. In: 2011 International conference on computer vision. pp. 2564-2571. IEEE (2011) 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "53. Rueckauer, B., Delbruck, T.: Evaluation of Event-based Algorithms for Optical Flow with Ground-truth from Inertial Measurement Sensor. Frontiers in neuroscience 10, 176 (2016) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "54. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.473 + ], + "angle": 0, + "content": "55. Shi, X., Huang, Z., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Flowformer++: Masked Cost Volume Autoencoding for Pretraining Optical Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1599-1610 (2023) 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "56. Teed, Z., Deng, J.: RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In: Proceedings of the European Conference on Computer Vision. pp. 402-419. Springer (2020) 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "57. Teed, Z., Deng, J.: Raft-3d: Scene flow using rigid-motion embeddings. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8375–8384 (2021) 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.594 + ], + "angle": 0, + "content": "58. Wan, Z., Dai, Y., Mao, Y.: Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing 31, 7237-7251 (2022) 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.65 + ], + "angle": 0, + "content": "59. Wan, Z., Mao, Y., Zhang, J., Dai, Y.: Rpeflow: Multimodal fusion of rgbpointcloud-event for joint optical flow and scene flow estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10030-10040 (2023) 3, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.703 + ], + "angle": 0, + "content": "60. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.787, + 0.745 + ], + "angle": 0, + "content": "61. Yang, B., Huang, Z., Li, Y., Zhou, H., Li, H., Zhang, G., Bao, H.: Hybrid3d: learning 3d hybrid features with point clouds and multi-view images for point cloud registration. Science China Information Sciences 66(7), 172101 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.787, + 0.798 + ], + "angle": 0, + "content": "62. Zheng, Y., Harley, A.W., Shen, B., Wetzstein, G., Guibas, L.J.: Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19855-19865 (2023) 3, 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "63. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The Multivehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters 3(3), 2032-2039 (2018) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.19 + ], + "angle": 0, + "content": "64. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In: Kress-Gazit, H., Srinivasa, S.S., Howard, T., Atanasov, N. (eds.) Robotics: Science and Systems (2018) 3, 5" + } + ] +] \ No newline at end of file diff --git a/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_origin.pdf b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c315338cb84bbc504a65a9cd63ecd3f938c9956a --- /dev/null +++ b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:823d7ba387ff1ccf89a7e7371528dfb31e3faf4c939e49b16fd51e16aca60328 +size 9787381 diff --git a/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/full.md b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a19aa2ebd9432799b642811efa2b62adf50da593 --- /dev/null +++ b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/full.md @@ -0,0 +1,310 @@ +# BlinkVision: A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events + +Yijin Li $^{1,2*}$ , Yichen Shen $^{1*}$ , Zhaoyang Huang $^{2*}$ , Shuo Chen $^{1}$ , Weikang Bian $^{3}$ , Xiaoyu Shi $^{3}$ , Fu-Yun Wang $^{3}$ , Keqiang Sun $^{3}$ , Hujun Bao $^{1}$ , Zhaopeng Cui $^{1}$ , Guofeng Zhang $^{1\dagger}$ , and Hongsheng Li $^{3,4,5\dagger}$ + +$^{1}$ State Key Lab of CAD&CG, Zhejiang University + +$^{2}$ Avolution AI $^{3}$ CUHK MMLab $^{4}$ Shanghai AI Laboratory + +5 CPII under InnoHK + +Abstract. Recent advances in event-based vision suggest that they complement traditional cameras by providing continuous observation without frame rate limitations and high dynamic range which are well-suited for correspondence tasks such as optical flow and point tracking. However, so far there is still a lack of comprehensive benchmarks for correspondence tasks with both event data and images. To fill this gap, we propose BlinkVision, a large-scale and diverse benchmark with rich modality and dense annotation of correspondence. BlinkVision has several appealing properties: 1) Rich modalities: It encompasses both event data and RGB images. 2) Rich annotations: It provides dense per-pixel annotations covering optical flow, scene flow, and point tracking. 3) Large vocabulary: It incorporates 410 daily categories, sharing common classes with widely-used 2D and 3D datasets such as LVIS and ShapeNet. 4) Naturalistic: It delivers photorealism data and covers a variety of naturalistic factors such as camera shake and deformation. BlinkVision enables extensive benchmarks on three types of correspondence tasks (i.e., optical flow, point tracking and scene flow estimation) for both image-based methods and event-based methods, leading to new observations, practices, and insights for future research. The benchmark website is https://www.blinkvision.net/. + +# 1 Introduction + +Modern image-based computer vision technology still cannot match the accuracy and robustness of human vision in many areas. One possible reason is that traditional cameras suffer from motion blur and limited frame rates, and they often rely on well-lighted conditions. In contrast, event cameras [13, 36] detect changes in intensity at each pixel as a stream of asynchronous events, which + +![](images/b0dc5854d567c50c6617ec6154781f8bdd52f6315f8ad9ca255ee96fc23579e7.jpg) +Fig. 1: BlinkVision is a large-scale and diverse benchmark with rich modality and dense annotation of correspondence. It covers 410 daily categories, sharing common classes with popular 2D and 3D datasets. The per-category object distributions, scene structure hierarchy, data samples, and supported applications of BlinkVision are shown in this figure. + +eliminates frame rate limitations and enables operation within a high dynamic range. However, they can not capture fine-grained details as traditional cameras do. Looking at how human vision works, we find that two cells in the human retina, i.e., cones and rods, work similarly to these two types of cameras, respectively. According to the duplex theory of vision [45], which posits that rods and cones serve different functions, their combination ensures the robustness of human visual processing. Therefore, we believe that combining the advantages of traditional cameras and event cameras can significantly enhance computer vision systems, providing more comprehensive and adaptive vision capabilities. + +However, there are currently only a few benchmarks [15, 32, 34] providing both event data and RGB frames, which hinder the development of algorithms that fully exploit the event data and fuse information from both modalities. This shortfall is particularly prominent in the domain of pixel correspondence estimation [8, 27, 28, 61], i.e., optical flow, scene flow, and point tracking estimation. Previous benchmarks built for pixel correspondence are either highly biased to specific scenes [15, 63] or rather simple [53]. A primary factor is that obtaining such precise pixel-wise annotations for these tasks is expensive. + +To boost the research in this area, we present BlinkVision, a synthetic benchmark for optical flow, scene flow and point tracking estimation using RGB frames and events. Our dataset has several appealing properties: 1) Rich modalities: BlinkVision encompasses three visual modalities: final RGB images, clean RGB images and event data. The final RGB images reflect real-world challenges like motion blur and limited dynamic range while the clean RGB images are devoid of such imperfections. The clean RGB images can be seen as the latent images [48] + +Table 1: A comparison between BlinkVision and other widely-used benchmarks on pixel correspondence estimation. "Occ", "Chara" and "Cats" is the abbreviation for occlusion, character and categories, respectively. $R^{lvis}$ denotes the ratio of the 1.2k LVIS [20] categories being covered. "N/A" denotes the dataset does not provide category labels. + +
DatasetsEventCleanFinalOptical FlowScene FlowPoint TrackingOccCharaAnimalCatsR1vis(%)
MVSEC [64]✓X✓✓XXXXXN/A0
DSEC [15]✓X✓✓XXXXXN/A0
BlinkFlow [34]✓✓X✓XX✓✓X554.1
EKubric [59]✓✓X✓✓✓✓XX170.9
Sintel [9]X✓X✓XX✓✓✓N/A0
KITTI [17]XX✓✓✓XXXXN/A0
FlyingThings [42]X✓X✓✓X✓XX554.1
Kubric [18]X✓X✓✓✓✓XX170.9
TAP-Vid [12]X✓XXX✓X✓XN/A0
PointOdyssey [62]X✓XXX✓✓✓✓N/A0
BlinkVision (Ours)✓✓✓✓✓✓✓✓✓41033.4
+ +of event cameras. 2) Rich annotations: It provides annotations covering optical flow, scene flow, and point tracking. Unlike benchmarks [15,17,62] which only contain sparse annotations, our data provide dense per-pixel annotations of each image, covering objects including moving cars, deformable characters, and animals. 3) Large vocabulary: It incorporates 410 daily categories, sharing common classes with widely-used 2D and 3D datasets such as ImageNet [33], LVIS [20], and ShapeNet [10], as depicted in Fig. 1. To our knowledge, our data have the widest variety of objects in existing pixel correspondence benchmarks. This expansive vocabulary is pivotal, enabling rigorous exploration of algorithmic generalization across varied objects. 4) Naturalistic: BlinkVision employs high-quality assets and rendering tools and thus is able to deliver photorealism data. Besides, it covers a variety of naturalistic factors such as camera shake and deformation. + +To make full use of BlinkVision, we set up a public benchmark website that allows uploading results and provides a public leaderboard. Besides, we evaluated both existing image-based and event-based methods on three typical correspondence tasks (i.e. optical flow, point tracking, and scene flow estimation). The results reveal new observations and challenges and serve as the baseline for future approaches. Specifically, we first study the robustness of existing image-based methods under large frame intervals and extreme illumination and point out the new challenge for these methods. Second, the benchmark results on existing event-based methods show that current methods do not fully unleash the potential of event cameras. Third, we show that fine-tuning existing methods on the training set of BlinkVision significantly boosts the generalizability, demonstrating the vast diversity of BlinkVision. Finally, the broad diversity and accessible category labels in BlinkVision allow us, for the first time, to analyze the performance of correspondence tasks on different categories. We believe BlinkVision has the potential to serve as a cornerstone benchmark for advancing the development of more robust computer vision systems. + +![](images/faa359d42f2eb44ad9aa4bb22d01923d53e76a236e09fd281055ae554fc13d60.jpg) +Fig. 2: Scenes samples in the proposed BlinkVision benchmark. + +# 2 Related Work + +A comprehensive overview and comparison between these benchmarks and BlinkVision is listed in Table 1. + +Image-based optical flow and scene flow benchmarks. Early evaluations depended on synthetic datasets, like the famous "Yosemite" sequence [7]. MPI Sintel [9] was one of the most representative synthetic benchmarks derived from a short open-source animated 3D movie and it became one of the most popular benchmark datasets. KITTI [17] was almost the most well-known among real data. It computed ground truth for static scenes through the data from a 3D laser scanner and the ego-motion data of the car. In a later version, KITTI extended the ground truth to rigidly moving cars by fitting CAD models of cars. FlyingThings3D [42] was a more recent synthetic dataset. It contained the ground truth of both optical flow and scene flow. FlyingThings3D and KITTI were the two most commonly used benchmarks in scene flow evaluation. However, on both datasets, ground truth about scene flow was not available in occluded regions. Furthermore, on KITTI, ground truth for foreground points was either missing or approximated by a fitted car CAD model. + +Image-based point tracking benchmarks. The first point-tracking dataset was FlyingThings++ [21], which was based on FlyingThings3D. It was initially designed for training the network. Due to the lack of evaluation benchmarks, it was also used for evaluation in the early stages. Later, Doersch et al. [12] proposed a real data benchmark named TAP-Vid, which was based on two real-world datasets: DAVIS [49] and Kinetics [30]. TAP-Vid relied on manual annotation and could not handle occlusions. PointOdyssey [62] was a newly proposed benchmark that was based on synthetic data and therefore could provide dense ground truth. However, it did not guarantee that every pixel had ground truth because it only tracked the mesh vertices of the object. Furthermore, the naturalism of PointOdyssey was limited to its interior parts. Its outdoor portion had no realistic layout, just a skybox with randomly dropped objects. + +Event-based Benchmarks. Early event-based optical flow benchmarks [53] limited the camera to only rotational motion and inferred ground truth from + +the rotational motion of the camera. Two later benchmarks, DSEC [15] and MVSEC [64] computed the ground truth through LiDAR SLAM. Similar to KITTI, these two benchmarks limited the ground truth to the static elements of the scene. Besides, they had a rather limited motion pattern. More recently, Li et al. [34] proposed a large-scale diversified form synthetic benchmark named BlinkFlow, which presented more challenging scenes and complex motion patterns. As for the scene flow benchmark, Wan et al. [59] recently converted the existing FlyingThings [42] and Kubric [18] datasets to event datasets through video-to-event [14] technology. However, due to the limited frame rates in the original video data, this step inevitably produced artifacts in the event data. Wan et al. also extended the real-world event dataset DSEC with scene flow ground truth. As for the event-based point-tracking benchmark, existing methods usually employed Structure-from-Motion [54] (SfM) technology to associate keypoints across multiple frames to obtain correspondence ground truth. Event Camera Dataset [46] and EDS dataset [22] were two common benchmarks in this area. Limited by the sparsity and accuracy of SfM, these benchmarks did not support the evaluation of arbitrary point tracking. + +# 3 BlinkVision Dataset + +In this section, we describe how to build BlinkVision, including the scene setup, data rendering, and generation for the ground truth labels. The process is based on Blender [2] as it provides photorealistic rendering and a flexible data interface that allows us to obtain customized correspondence ground truth. At the last, we introduce the statistics and distribution of the BlinkVision data. + +# 3.1 Scene Setting + +Previous synthetic benchmarks [9, 43] generally rely on open-source movies to avoid heavy scene construction. However, these movies are biased and do not cover diverse enough scenarios. In order to establish a comprehensive evaluation benchmark, we manually assemble a collection of scenarios that are as rich and diverse as possible. We first look for ready-made scenes that are photorealistic and visually diverse. Specifically, we purchased 40 indoor scenes and 13 outdoor scenes from Evermotion Archinteriors Collection [3] that cover common scenes such as living rooms, kitchens, offices, bedrooms, restaurants and gardens. The original scene is static. To enhance the realism, we procured 29 scanned and high-fidelity human bodies from ActorCore [1] and 88 artist-designed animals, together with 104 free characters from Mixamo [4]. These assets are rigged. We re-targeted them to a wide range of motion models (e.g., more than 100 human motions including dancing, walking, talking, etc.) and placed them in various locations. The human's motions mainly come from motion capture [1] and the animal's motions are designed by the artist to closely emulate their natural motions. Furthermore, we built 50 additional outdoor scenes that cover cases such + +as a drone flying through the forest and a camera following a human or car from the crowd, which are common but rarely included in existing correspondence benchmarks. We set the camera trajectory by referencing the shooting trajectories of real-world videos, including handheld shots, car shots, and drone shots that cover non-uniform motion such as sudden stops and sharp turns. Finally, we obtained 80 indoor sequences and 63 outdoor sequences, where 56 indoor sequences and 47 outdoor sequences are for testing and the others are used for fine-tuning. Some samples of our data are shown in Fig. 2. In the supplementary, we also provide video samples. All the assets used in the testing and fine-tuning are totally disjoint, even including trees. + +# 3.2 Multi-modality Data Rendering + +We use Blender to render brightness $B$ in linear color space. The following simulations of real-world image capturing and event data are both based on the linear color space. We employ tone mapping and gamma correction on $B$ to obtain commonly used sRGB images. We call these sRGB images "RGB (clean)" because they do not suffer from effects like blurs or overexposed. On the contrary, we call the simulated real-world image capture "RGB (final)". + +Simulation of Real-world Image Capturing. Real-world image capture suffers from motion blur and limited dynamic range. While Blender supports the simulation of the former, the latter needs additional processing. The latter effect usually leads to overexposed or underexposed. To simulate it, we follow previous work [40] and employ the following steps: (1) Random exposure ratio. We uniformly sample the exposure ratio $w$ in the log2 space within $[-3, 3]$ and multiply it by $B$ to simulate underexposed or overexposed, which gives us the augmented HDR image $H = B \times 2^w$ . (2) Dynamic range clipping. We clip $H$ according to the formulation $\mathcal{C}(H) = \min(H, 1)$ . This step leads to information loss for pixels in the overexposed regions. (3) Non-linear mapping. To align with how humans see a scene, a camera typically uses a non-linear camera response function (CRF) to modify the contrast of the captured image, which can be formulated by $I_n = \mathcal{F}(I_c)$ . We randomly sample CRFs from an existing dataset [19]. (4) Quantization. The pixel values are quantized to 8 bits by $\mathcal{Q}(I_n) = \lfloor 255 \times I_n + 0.5 \rfloor / 255$ . This step causes information loss in underexposed and smooth gradient areas. + +Event Data. Event cameras work by responding to changes in the logarithmic brightness signal (i.e., $L = \log B$ ) asynchronously and independently for each pixel [13]. An event is triggered when the change in brightness (either increase or decrease) since the last event at that pixel reaches a threshold of $\pm T$ (with $T > 0$ ): + +$$ +p _ {k} \left(L \left(u _ {k}, t _ {k}\right) - L \left(u _ {k}, t _ {k} - \Delta t\right)\right) \geq T, \tag {1} +$$ + +where $\Delta t$ is the time since the last event triggered at pixel location $u_{k}$ and at time $t_k$ . $p_k\in \{-1,1\}$ is the polarity of the brightness change. To model the event generation process, we need access to a continuous representation of the + +![](images/33e38f8766414540ca2828fad75c6240fec90d90663cb8bf0c14a78b1f65fb68.jpg) +Fig. 3: Statistics of optical flow in BlinkVision. + +![](images/7871dd5c68c7fae8bde7cd421ec565d2288d0542eb6dff9bcbbcb8f8c0b67c08.jpg) + +![](images/9f111e60c5c2783ef8312e798bc837738634169ed62cb823562f09deacb49dc7.jpg) + +![](images/4a9af71414961f4bea91647eed0cb78037f93e8562907b89e6b343b884fa1ad3.jpg) + +![](images/f8aa04e1ff8f98f4f6a6a10e7717999c28a80b2053f4b1e241eba7e8ae8941fd.jpg) +Fig. 4: Statistics of point trajectories in BlinkVision. To save time, we sample the tracks using a grid with a size of 20. Trajectory segments are defined as contiguous sections of point trajectories, with interruptions caused by occlusion. The diameter is the maximum distance a point moves over time. We clip the trajectory diameter and divide it by the diagonal length of the image to obtain the ratio. + +![](images/b370f6293496785a4645d120e316adc729661801ed343c8c178344b0ee1bc280.jpg) + +![](images/86231bbda378e7cd63d0d0fe98425d913927a299cce9eb4ad10fc2b61fb12cb3.jpg) + +![](images/8862018015469665aa541f02eab472549cf493b168b480afefd79682dcac6dfb.jpg) + +visual signal for each pixel. In practice, it is approximated by rendering images at high frame rates for efficiency [34, 51]. Events are then synthesized based on frame-by-frame pixel differences. More specifically, we adaptively sample frames according to [34, 51] to ensure that the maximum pixel displacement between two sampling timestamps is bounded. In practice, we employ frame interpolation before simulating the event data to further reduce the pixel displacement. We use DVS-Voltmeter [37] to synthesize events because it can simulate complex noise effects (such as noise effects of temperature and parasitic photocurrent), thus generating realistic events. + +# 3.3 Ground Truth Generation + +Blender provides optical flow data between two consecutive frames and a segmentation mask for each object. However, we cannot directly obtain these data from Blender: (1) pixel correspondence across multiple frames; (2) scene flow between frames; (3) semantic category for each object. + +Pixel Correspondence. The data of (1) and (2) can be computed in a unified framework. Given camera poses $\mathbf{T}$ (camera-to-world), depth maps $\mathbf{Z}$ at frames $i$ and $j$ , and the camera intrinsic $K$ , for the pixel location $\mathfrak{u}$ at frame $i$ , we first project it to world coordinate: $P = T_{i}Z_{i}(\mathfrak{u})K^{-1}\tilde{\mathfrak{u}}$ , where $\tilde{\mathfrak{u}}$ is the homogeneous + +coordinates of u. Then we re-project the 3D point $P$ to frame $j$ through $du_{j} = KT_{j}^{-1}P$ , where $\mathfrak{u}_j$ is the corresponding pixel location and $d$ is the depth value at frame $j$ . In this way, we get the pixel correspondence between any two frames and the corresponding depth, i.e., (1) and (2). + +However, this only works for static objects. To handle dynamic objects and even those objects with deformation motion such as humans and animals, we bake the face index and barycentric coordinate $(\lambda_1,\lambda_2,\lambda_3)$ of each mesh face into textures. After rendering the texture, we locate the triangular face that the tracked pixel belongs to (indexed by the rendered face index) and obtain the vertices' 3D position $(V_{1},V_{2},V_{3})$ of that face via Blender's API. The 3D position of tracked pixels can be computed through $P = \lambda_1V_1 + \lambda_2V_2 + \lambda_3V_3$ . This step replaces the back projection for the static objects. It is worth noting that we need to triangulate every mesh face in the scenes before baking. + +Semantic Labeling. BlinkVision contains thousands of objects which makes manual semantic labeling expensive. To this end, we develop an automatic labeling framework that is based on open-vocabulary semantic understanding [50,60]. First, we pre-define a category list based on LVIS [20]. The category together with category descriptions is encoded into text embedding $F_{t} \in \mathbb{R}^{N \times D}$ with a pre-trained text encoder, i.e., CLIP, where $N$ is the number of categories and $D$ is the feature dimension. For each object asset, we render it individually to a $224 \times 224$ image. The process is rather fast while keeping photorealistic because it does not need to compute complex path tracing between objects. We use the pre-trained image encoder from CLIP [50] to extract the image embedding $F_{i} \in \mathbb{R}^{D}$ . The probability of the image belonging to one of the pre-defined categories is computed through $P(c|F_{i}) = \text{softmax}\left(F_{i} \cdot (F_{t})^{\top}\right)_{c}$ . + +# 3.4 Dataset Overview + +BlinkVision consists of 40 training sequences and 103 testing sequences. The training set provides 107,880 frame pairs for optical flow and scene flow, and 1025 sub-sequences for point tracking. The frame rate for RGB images and ground truth data is 20 FPS. The frame rate is not practically significant because we adjusted it to ensure sufficient motion between adjacent frames. This adjustment reduces the data size uploaded to the benchmark website. For each sequence of BlinkVision, we selected the sub-sequences by starting tracking from the first frame. Once the overlap with the reference frame (i.e., how many tracked points remain inside the image) is smaller than $40\%$ , we start to track a new reference frame. If the overlap is smaller than $20\%$ , we stop this track. Additionally, we discard short tracks that are less than 20 frames long because they are less challenging without long-term motion accumulation, which could result in the benchmark easily reaching saturation. As a result, we generate 1025 subsequences where each sub-sequence contains $640 \times 480$ dense point trajectories. For the test set, we selected 12,804 frame pairs for evaluating optical flow and + +![](images/00ed5a18729acc643bb21eeccd2a63109cc39473c56cf53006ba05d5f5788de0.jpg) + +![](images/ec217484f460347a83417bf4b61be13b7f9345ec6a8eb73e818306994ff7fb8c.jpg) +Reference Image + +![](images/6a62f30eb2cc1b147b0ef6c389f51d55993217a41c1166d54c646d760128a717.jpg) + +![](images/b99d97f9f3ad5542bf0153e156c7ebcf68a522898923a7a0f5a999a61a301813.jpg) +Target Image + +![](images/0f3e3dbfc1006eec3cbb8c19cfb8a1962001a2ff168c1fb62b95db57677f3e7d.jpg) + +![](images/916d4e5e3c8285620e4c6dd491c1115d00cebb57f1d54b5df118f1f39b2ee9a9.jpg) +Ground Truth +Fig. 5: Qualitative results of FlowFormer++ [55] before and after fine-tuning on the training set of BlinkVision. + +![](images/37a3e35769199669e38113ddc173d1ee66152c37b17ee54ea9de0b4fb561ffa1.jpg) + +![](images/1bb862dc35943f791ba10b72447595bff9fa6fd797b6563268ce4ed619da5876.jpg) +Before Finetuning + +![](images/81d74fd5de321b0235f96293c83e90b23dbd16a1facf962037c9e2a64bf5d673.jpg) + +![](images/b84c7682a80a7267e4b60b98fe7da5ce680eec9e3b658dce7fac5dc116011736.jpg) +After Finetuning + +scene flow, and 865 sub-sequences for evaluating point tracking. BlinkVision includes raw data on depth, normal, camera pose, etc. At this stage, we focus on pixel-level motion tasks. The additional raw data will be released later, which will benefit other tasks such as depth estimation [35] and SLAM [23, 24, 39]. + +Table 1 summarizes key statistics of BlinkVision compared to related works. BlinkVision contains more than 40K 3D models in 410 categories. It shares some common categories with 2D and 3D datasets [10,20], and bears a huge diversity in semantics, geometry, and appearance, enabling a wide range of research topics. + +We provide statistics of optical flow in Fig. 3. For benchmarking, we discard those pixels whose flow magnitudes are larger than half of the image diagonal. The latter situation usually occurs when the object is too close to the camera, and its magnitude can even approach infinity. As a result, the magnitude shown in Fig. 3 has an upper bound of 400. The broad distribution in Fig. 3 reveals the diverse motions in BlinkVision. Besides, the diverse distribution of "X Variance" and "Y Variance" also indicates the presence of complex object shapes in BlinkVision. We also provide statistics of point trajectories in Fig. 4. BlinkVision allows the evaluation of point tracking with different trajectories' lengths, from small to big, corresponding to different downstream applications like video editing [25, 41] or augmented reality [31, 39]. The broad distribution of "occlusion ratio" and "number of segments" indicates that BlinkVision brings diverse and challenging data. The occlusion ratio has an upper bound that is less than 100 percent due to our strategy of sub-sequences selection. To quantify the amount of motion a point undergoes, we calculate its trajectory diameter, which is the maximum distance between any two positions of the point over its entire trajectory. We observe that point motion in BlinkVision is diverse and nearly uniformly distributed across the entire image plane. + +# 4 Benchmark + +We release all the data except for the ground truth for the test split. For a fair comparison, we create a public benchmark website with leaderboards. We also release the evaluation code for our online benchmarking. + +Table 2: Optical flow result of RGB-frame-based methods. "St." denotes Stride. "†" denotes fine-tuning. "FF" denotes FlowFormer [26] and "FF++" denotes FlowFormer++ [55]. + +
MethodEPE - clean
St. 1St. 2St. 4St. 8Avg.
RAFT [56]1.834.6210.1820.469.27
GMA [29]1.824.379.5218.688.60
FF [26]1.603.777.5716.037.24
FF++ [55]1.543.577.4316.267.20
RAFT† [56]1.332.966.3912.725.85
FF++† [55]1.092.354.8810.584.73
MethodEPE - final
St. 1St. 2St. 4St. 8Avg.
RAFT [56]2.545.6911.2722.6710.54
GMA [29]2.565.7211.2121.3510.21
FF [26]2.264.809.2418.238.63
FF++ [55]2.284.839.3919.088.90
RAFT† [56]1.943.887.3515.177.08
FF++† [55]1.663.346.4212.736.04
+ +Table 3: Optical flow result of event-based methods. "St." denotes Stride. "†" denotes fine-tuning. + +
Event
Methods|St. 1St. 2St. 4St. 8Avg.
E-RAFT [16]2.817.0417.8228.6014.07
STE-FlowNet [11]2.595.9413.1341.8915.89
E-FlowFormer [34]2.415.6612.5330.4512.76
E-RAFT† [16]1.683.637.4814.206.75
E-FlowFormer† [34]1.513.005.9613.606.02
Event + RGB (clean)
Methods|St. 1St. 2St. 4St. 8Avg.
DCEIFlow [58]3.3114.0734.7861.5628.43
DCEIFlow† [58]2.196.3012.5426.1111.79
Event + RGB (final)
Methods|St. 1St. 2St. 4St. 8Avg.
DCEIFlow [58]4.0814.2434.5461.2728.53
DCEIFlow† [58]2.526.8613.6428.4512.87
+ +# 4.1 Optical Flow + +In this section, we comprehensively evaluate the performance of existing optical flow methods on BlinkVision. Specifically, we first analyze the robustness of image-based methods under large frame intervals and extreme illumination. Large motion, severe occlusion, and information loss in these extreme cases pose great challenges to existing image-based methods. In contrast, event cameras naturally depict continuous pixel motion and possess a large dynamic range, thus offering great potential for solving these challenges. We then benchmark existing event-based methods, including event-only methods and event-RGB fusion methods. We find that existing event-based methods cannot fully unleash the potential of event cameras. We analyze the possible reasons and point to possible opportunities for future research. + +Experimental Setup. To study the impact of frame intervals on existing methods, we select frames separated by 1, 2, 4, and 8 from the reference frame as target frames. To reduce the file size of results that are uploaded to the benchmark website, we uniformly sample a reference frame from ten consecutive frames. As a result, there are 12,804 frame pairs for testing in total. We follow the previous work [26, 47, 56] and report the EPE (end-point-error) as the evaluation metric. + +Results. We benchmark several image-based methods in Table 2. We make the following observations: 1) The error generally increases linearly with stride size. This is expected because doubling the stride yields motion that is twice as fast, resulting in an error that is also expected to be twice as high. 2) Extreme lighting has a heavy effect on existing image-based methods. In the supplementary + +Table 4: Point tracking result of RGB-frame-based methods under different frame interval. "St." denotes Stride. + +
RGB(clean)
MethodMetricSt. 1St. 2St. 4St. 8Avg.
PIPs [21]δavg37.7739.3436.3425.5034.74
Survival↑52.2855.3954.5547.5852.45
MTE↓50.7747.8149.5362.4252.63
PIPs++ [62]δavg40.0238.2530.2416.8431.34
Survival↑53.4453.1347.5739.7148.46
MTE↓45.5151.0259.8773.1757.39
Context-TAP [8]δavg38.7539.4736.5125.9935.18
Survival↑49.9853.0952.6146.1850.46
MTE↓51.4649.6151.8364.3954.32
RGB(final)
MethodMetricSt. 1St. 2St. 4St. 8Avg.
PIPs [21]δavg33.0234.4431.5621.7730.20
Survival↑48.3551.5650.7244.5648.80
MTE↓53.9451.5453.6566.1456.32
PIPs++ [62]δavg35.8334.2427.1115.3228.12
Survival↑50.4750.3645.3238.4646.15
MTE↓48.3753.7462.3474.5959.76
Context-TAP [8]δavg33.4634.1731.3922.0030.25
Survival↑45.5848.9148.5343.0646.52
MTE↓55.0953.7656.3368.1158.32
+ +Table 5: Point tracking result of event-based methods. \* denotes that we track ORB [52] feature points rather than grid sampled points. DET" denotes DeepETracker [44] + +
Event
Methodδavg↑Survival↑MTE↓
AMH [5]18.9537.9875.44
HASTE [6]18.6037.4377.36
AMH* [5]26.7641.9362.48
HASTE* [6]27.6542.8762.52
Event + RGB (clean)
Methodδavg↑Survival↑MTE↓
DET [44]21.6235.4480.80
DET* [44]33.4648.8260.69
Event + RGB (final)
Methodδavg↑Survival↑MTE↓
DET [44]19.0533.5285.00
DET* [44]31.7848.6758.79
+ +materials, we analyze how different exposure factors lead to performance loss. 3) Fine-tuning on the training split of BlinkVision significantly improves the performance. In Fig. 5 we show qualitative results of FlowFormer++ before and after fine-tuning on BlinkVision. After fine-tuning, the flow predictions are more precise and clear, especially near the object boundary. As shown in Table 2, the benchmark is still challenging for methods after fine-tuning. + +We report the evaluation of event-based methods in Table 3. It is surprising that the performance of existing event-based methods is worse than image-based methods and they also suffer from severe degradation under large frame intervals. The reasons are mainly twofold. First, existing training data for events lag behind that of RGB. This can be verified by the performance of E-RAFT [16] and E-FlowFormer [34] after fine-tuning. These two methods are far behind RAFT [56] before fine-tuning, but exceed the fine-tuned RAFT [56] and even FlowFormer++ [55] after fine-tuning (in "Stride-1", "Stride-2" and "Stride-4" of the "final" case), indicating that the poor performance of the former mainly comes from limited training on existing event-based training data. Second, existing event-based methods usually convert events to regular 3D voxel grids before processing, which quantizes the event data and leads to information loss. This problem is more serious when dealing with event data within large frame intervals. As a result, it calls for a more powerful representation of event data and new algorithms that can deal with long-range optical flow estimation. + +![](images/a7982135da724b9ece0dbb755629e1e824cd25daa5db236eab6ecb16f88d09de.jpg) + +![](images/07efd0dd90f5376b5e39d86965790341a9bd44ede33f289c9840c11f62ebf74e.jpg) +Reference Image + +![](images/4092383402648846b974d562746ae5ac1ca1896fbcfb47113019481ce6ba48d9.jpg) + +![](images/06e0cb0547b1227a6b7b3847a431399063bebf1bd5322545bdf793a191d7594e.jpg) +Prediction Results + +![](images/61c8963d077cbbdfc03d9a3272b870c4677c82ea64270d7d0f35becf1dd76eea.jpg) + +![](images/5e671c5a44f5ff742642096ef68661b643ec01918282fc3184b7a90ece145536.jpg) +Error Visualization + +![](images/9f9cc860f4f23b7950a3edd2454b554088f6e7e66410a9518421f850953ecf2d.jpg) +Fig. 6: Qualitative results of $\mathbf{PIP}\mathbf{s} + +$ [62]. Darker lines in the right figure (e.g., red) indicate larger errors. +Fig. 7: Performance distribution of FlowFormer++ [55]. + +Table 6: Scene flow result of RGB-frame-based methods. "St." denotes Stride. + +
EPE-2d
MethodRGBSt. 1St. 2St. 4St. 8
RAFT-3D [57]clean2.154.348.5916.08
final3.487.2910.7418.57
CamLiFlow [38]clean2.796.0512.3423.89
final4.288.2816.3030.61
EPE-3d
MethodRGBSt. 1St. 2St. 4St. 8
RAFT-3D [57]clean1.913.837.2512.18
final2.023.967.3914.21
CamLiFlow [38]clean1.452.765.369.12
final1.532.885.609.81
+ +Table 7: Scene flow result of event-based methods. "St." denotes Stride. + +
EPE-2d
MethodRGBSt. 1St. 2St. 4St. 8
RPEFlow [59]clean1.533.507.3617.44
final1.924.108.1518.38
EPE-3d
MethodRGBSt. 1St. 2St. 4St. 8
RPEFlow [59]clean5.089.0715.7725.24
final5.209.3815.6725.03
+ +# 4.2 Point Tracking + +In this section, we evaluate the performance of existing methods for long-term point tracking on BlinkVision. Similar to what we analyzed on optical flow, we analyze the robustness under large frame intervals and extreme illumination, and the gaps between existing image-based methods and event-based methods. + +Experimental Setup. Although BlinkVision provides per-pixel dense annotations for point tracking, we find that existing methods are not efficient enough to process so much data. As a result, we grid sample the tracked pixels with a grid size of 20 pixels. We follow PointOdyssey [62] and use $\delta_{\mathrm{avg}}$ , median trajectory error (MTE) and survival rate as the evaluation metrics. + +Results. We benchmark several image-based methods in Table 4. Similarly, we observe that extreme lighting has a major impact on existing image-based point tracking methods. In Table 4 we can see that larger frame intervals (such as 4 and 8) severely degrade performance. However, for PIPs [21] and Context TAP [8], increasing the interval from 1 to 2 makes the performance slightly better. We + +Table 8: Cross-dataset evaluation of optical flow methods fine-tuned on BlinkVision. "BV" denotes BlinkVision. + +
RGB
MethodDataSintel (clean)Sintel (final)KITTI
RAFT [56]- +BV2.08 1.693.41 3.045.10 4.66
Events
MethodDataE- BlenderFlying- ObjectsE- Tartan
E-RAFT [16]- +BV2.66 1.922.9 2.673.27 2.55
E-FlowFormer [34]- +BV2.38 1.752.89 2.582.91 2.36
Events + RGB
MethodDataE- BlenderFlying- ObjectsE- Tartan
DCEIFlow [58]- +BV8.96 4.769.58 2.377.44 2.88
+ +Table 9: Performance evaluation for typical categories at each difficult level. "FF" denotes FlowFormer [26] and "FF++" denotes FlowFormer++ [55]. We use the EPE as the evaluation metric. + +
DifficultCategoryRAFT [56]FF [26]FF++ [55]
Hardperson9.828.728.61
tree9.799.7610.10
rhinoceros8.907.076.78
Mediumbench4.293.603.68
globe3.432.522.51
lantern2.441.921.95
Easynotebook1.080.570.84
coaster0.860.870.79
painting0.310.300.26
+ +guess it might be the limited temporal receptive field of these two methods (only 8 frames) that makes the accumulated error quickly increase when the stride is too small. Some qualitative results are shown in Fig. 6. BlinkVision contains many challenging cases that cannot be handled by the SOTA methods. + +We show the performance of event-based methods in Table 5. We find that event-based approaches perform particularly poorly. In addition to the reasons for insufficient training data and model design, we deduce that the receptive field of event-based point-tracking methods is relatively small and therefore cannot perform well on the task of tracking arbitrary points. To verify the claim, we replace the input of grid sampled positions with ORB [52] feature points, denoted by “*” in the table. The new results perform even better, validating our ideas. Although the new results cannot be strictly compared with image-based methods, it performs better under extreme frame interval, i.e., “Stride-8”, which shows the large potential of event-based methods. + +# 4.3 Scene Flow + +Experimental Setup. We use the same pairs as the optical flow benchmark and we follow [38,57] to use 2D EPE and 3D EPE as the evaluation metrics. + +Results. We benchmark two state-of-the-art image-based methods, i.e., RAFT-3D [57] and CamliFlow [38] as shown in Table 6 and show the results of event-based methods in Table 7. The conclusions of the optical flow benchmark apply to the scene flow task and present similar challenges for existing methods in this area. We show qualitative results in the supplementary materials. + +# 4.4 Cross-dataset Evaluation + +We also fine-tune existing optical flow methods on the BlinkVision training set and then evaluate them on existing representative benchmarks. The results are shown in Table 8. We observe that fine-tuning on BlinkVision brings significant improvement for both image-based methods (2.08 vs. 1.69) and event-based methods (2.38 vs. 1.75 and 9.58 vs. 2.37), which demonstrates the vast diversity of BlinkVision boosts the generalizability of these methods. + +# 4.5 Performance Distribution on Categories + +Previous methods in optical flow mainly perform evaluations on biased and limited scenarios, which is not comprehensive and robust enough to demonstrate the ability of the methods for different categories of objects in different scenarios. Thanks to the vast diversity of data and semantic labels provided by BlinkVision, for the first time, we analyze the performance distribution of several image-based optical flow methods on different categories. The results are shown in Fig. 7. The average curve is imbalanced: hard categories usually include complex shapes (e.g., hammocks and shrubs) or with deformable motion (e.g., persons and animals). We thus split the categories into three levels of "difficulty" based on the average curve, and the performance evaluation for typical categories at each level is presented in Table 9. We believe such fine-grained analysis helps understand the generalization capacity of the methods. + +# 5 Conclusion + +We propose BlinkVision, a large-scale diversified benchmark for three types of correspondence tasks, i.e., optical flow, point tracking, and scene flow estimation using RGB frames and events. Extensive benchmarks on BlinkVision point to new challenges for existing image-based approaches and show that existing event-based approaches are far from fully unlocking the potential of event cameras. BlinkVision reveals new observations, challenges, and opportunities for future research into more robust visual systems such as human vision. + +Acknowledgment. This project was funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR, by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. This work was also partially supported by NSF of China (No. 61932003). + +# References + +1. ActorCore. https://actorcore.realusion.com/, accessed: 2023-11-17 5 +2. Blender. https://www.blender.org/, accessed: 2023-11-17 5 +3. Evermotion Archinteriors Collection. https://evermotion.org/, accessed: 2023-11-17 5 +4. Mixamo. https://www MIXamo.com/, accessed: 2023-11-17 5 +5. Alzugaray, I., Chli, M.: Asynchronous multi-hypothesis tracking of features with event cameras. In: 2019 International Conference on 3D Vision (3DV). pp. 269-278. IEEE (2019) 11 +6. Alzugaray, I., Chli, M.: Haste: multi-hypothesis asynchronous speeded-up tracking of events. In: 31st British Machine Vision Virtual Conference (BMVC 2020). p. 744. ETH Zurich, Institute of Robotics and Intelligent Systems (2020) 11 +7. Barron, J.L., Fleet, D.J., Beauchemin, S.S.: Performance of optical flow techniques. International journal of computer vision 12, 43-77 (1994) 4 +8. Bian, W., Huang, Z., Shi, X., Dong, Y., Li, Y., Li, H.: Context-TAP: Tracking Any Point Demands Spatial Context Features. arXiv preprint arXiv:2306.02000 (2023) 2, 11, 12 +9. Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12. pp. 611-625. Springer (2012) 3, 4, 5 +0. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An Information-rich 3D Model Repository. arXiv preprint arXiv:1512.03012 (2015) 3, 9 +1. Ding, Z., Zhao, R., Zhang, J., Gao, T., Xiong, R., Yu, Z., Huang, T.: Spatiotemporal Recurrent Networks for Event-based Optical Flow Estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 525-533 (2022) 10 +2. Doersch, C., Gupta, A., Markeeva, L., Recasens, A., Smaira, L., Aytar, Y., Carreira, J., Zisserman, A., Yang, Y.: Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems 35, 13610–13626 (2022) 3, 4 +3. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020) 1, 6 +4. Gehrig, D., Gehrig, M., Hidalgo-Carrio, J., Scaramuzza, D.: Video to Events: Recycling Video Datasets for Event Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3586-3595 (2020) 5 +5. Gehrig, M., Aarents, W., Gehrig, D., Scaramuzza, D.: DSEC: A Stereo Event Camera Dataset for Driving Scenarios. IEEE Robotics and Automation Letters 6(3), 4947-4954 (2021) 2, 3, 5 +6. Gehrig, M., Millhäsler, M., Gehrig, D., Scaramuzza, D.: E-RAFT: Dense Optical Flow from Event Cameras. In: Proceedings of the International Conference on 3D Vision. pp. 197–206. IEEE (2021) 10, 11, 13 +7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012) 3, 4 +8. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., et al.: Kubric: A scalable dataset + +generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3749-3761 (2022) 3, 5 +19. Grossberg, M.D., Nayar, S.K.: What is the space of camera response functions? In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. vol. 2, pp. II-602. IEEE (2003) 6 +20. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) 3, 8, 9 +21. Harley, A.W., Fang, Z., Fragkiadaki, K.: Particle video revisited: Tracking through occlusions using point trajectories. In: European Conference on Computer Vision. pp. 59-75. Springer (2022) 4, 11, 12 +22. Hidalgo-Carrió, J., Gallego, G., Scaramuzza, D.: Event-aided direct sparse odometry. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5781-5790 (2022) 5 +23. Hu, J., Chen, X., Feng, B., Li, G., Yang, L., Bao, H., Zhang, G., Cui, Z.: Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095 (2024) 9 +24. Hu, J., Mao, M., Bao, H., Zhang, G., Cui, Z.: Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems 36 (2024) 9 +25. Huang, Z., Pan, X., Pan, W., Bian, W., Xu, Y., Cheung, K.C., Zhang, G., Li, H.: Neuralmarker: A framework for learning general marker correspondence. ACM Transactions on Graphics (TOG) 41(6), 1-10 (2022) 9 +26. Huang, Z., Shi, X., Zhang, C., Wang, Q., Cheung, K.C., Qin, H., Dai, J., Li, H.: Flowformer: A Transformer Architecture for Optical Flow. In: Proceedings of the European Conference on Computer Vision. pp. 668-685. Springer (2022) 10, 13 +27. Huang, Z., Shi, X., Zhang, C., Wang, Q., Li, Y., Qin, H., Dai, J., Wang, X., Li, H.: FlowFormer: A Transformer Architecture and Its Masked Cost Volume Autoencoding for Optical Flow. arXiv preprint arXiv:2306.05442 (2023) 2 +28. Huang, Z., Zhou, H., Li, Y., Yang, B., Xu, Y., Zhou, X., Bao, H., Zhang, G., Li, H.: Vs-net: Voting with segmentation for visual localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6101-6111 (2021) 2 +29. Jiang, S., Campbell, D., Lu, Y., Li, H., Hartley, R.: Learning to estimate hidden motions with global motion aggregation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9772-9781 (2021) 10 +30. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 4 +31. Klein, G., Murray, D.: Parallel tracking and mapping for small ar workspaces. In: 2007 6th IEEE and ACM international symposium on mixed and augmented reality. pp. 225-234. IEEE (2007) 9 +32. Klenk, S., Chui, J., Demmel, N., Cremers, D.: Tum-vie: The tum stereo visual-inertial event dataset. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8601-8608. IEEE (2021) 2 +33. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012) 3 +34. Li, Y., Huang, Z., Chen, S., Shi, X., Li, H., Bao, H., Cui, Z., Zhang, G.: Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv preprint arXiv:2303.07716 (2023) 2, 3, 5, 7, 10, 11, 13 + +35. Li, Y., Liu, X., Dong, W., Zhou, H., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Deltar: Depth estimation from a light-weight tof sensor and rgb image. In: European conference on computer vision. pp. 619-636. Springer (2022) 9 +36. Li, Y., Zhou, H., Yang, B., Zhang, Y., Cui, Z., Bao, H., Zhang, G.: Graph-based asynchronous event processing for rapid object recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 934-943 (2021) 1 +37. Lin, S., Ma, Y., Guo, Z., Wen, B.: DVS-Voltmeter: Stochastic Process-Based Event Simulator for Dynamic Vision Sensors. In: Proceedings of the European Conference on Computer Vision. pp. 578-593. Springer (2022) 7 +38. Liu, H., Lu, T., Xu, Y., Liu, J., Li, W., Chen, L.: Camliflow: Bidirectional cameralidar fusion for joint optical flow and scene flow estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5791-5801 (2022) 12, 13 +39. Liu, X., Li, Y., Teng, Y., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. In: Proceedings of the IEEE/cvf international conference on computer vision. pp. 1-11 (2023) 9 +40. Liu, Y.L., Lai, W.S., Chen, Y.S., Kao, Y.L., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Single-image hdr reconstruction by learning to reverse the camera pipeline. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1651-1660 (2020) 6 +41. Luo, J., Huang, Z., Li, Y., Zhou, X., Zhang, G., Bao, H.: Niid-net: adapting surface normal knowledge for intrinsic image decomposition in indoor scenes. IEEE Transactions on Visualization and Computer Graphics 26(12), 3434-3445 (2020) 9 +42. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4040-4048 (2016) 3, 4, 5 +43. Mehl, L., Schmalfuss, J., Jahedi, A., Nalivayko, Y., Bruhn, A.: Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4981-4991 (2023) 5 +44. Messikommer, N., Fang, C., Gehrig, M., Scaramuzza, D.: Data-driven feature tracking for event cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5642-5651 (2023) 11 +45. Milner, D., Goodale, M.: The visual brain in action, vol. 27. OUP Oxford (2006) 2 +46. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research 36(2), 142-149 (2017) 5 +47. Ni, J., Li, Y., Huang, Z., Li, H., Bao, H., Cui, Z., Zhang, G.: PATS: Patch Area Transportation with Subdivision for Local Feature Matching. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 17776-17786 (2023) 10 +48. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019) 2 + +49. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016) 4 +50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 8 +51. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: An Open Event Camera Simulator. In: Proceedings of the Conference on Robot Learning. pp. 969-982. PMLR (2018) 7 +52. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: Orb: An efficient alternative to sift or surf. In: 2011 International conference on computer vision. pp. 2564-2571. IEEE (2011) 11, 13 +53. Rueckauer, B., Delbruck, T.: Evaluation of Event-based Algorithms for Optical Flow with Ground-truth from Inertial Measurement Sensor. Frontiers in neuroscience 10, 176 (2016) 2, 4 +54. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016) 5 +55. Shi, X., Huang, Z., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Flowformer++: Masked Cost Volume Autoencoding for Pretraining Optical Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1599-1610 (2023) 9, 10, 11, 12, 13 +56. Teed, Z., Deng, J.: RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In: Proceedings of the European Conference on Computer Vision. pp. 402-419. Springer (2020) 10, 11, 13 +57. Teed, Z., Deng, J.: Raft-3d: Scene flow using rigid-motion embeddings. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8375–8384 (2021) 12, 13 +58. Wan, Z., Dai, Y., Mao, Y.: Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing 31, 7237-7251 (2022) 10, 13 +59. Wan, Z., Mao, Y., Zhang, J., Dai, Y.: Rpeflow: Multimodal fusion of rgbpointcloud-event for joint optical flow and scene flow estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10030-10040 (2023) 3, 5, 12 +60. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023) 8 +61. Yang, B., Huang, Z., Li, Y., Zhou, H., Li, H., Zhang, G., Bao, H.: Hybrid3d: learning 3d hybrid features with point clouds and multi-view images for point cloud registration. Science China Information Sciences 66(7), 172101 (2023) 2 +62. Zheng, Y., Harley, A.W., Shen, B., Wetzstein, G., Guibas, L.J.: Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19855-19865 (2023) 3, 4, 11, 12 +63. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The Multivehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters 3(3), 2032-2039 (2018) 2 + +64. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In: Kress-Gazit, H., Srinivasa, S.S., Howard, T., Atanasov, N. (eds.) Robotics: Science and Systems (2018) 3, 5 \ No newline at end of file diff --git a/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/images.zip b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db49cd2b23b5cde7cee3765d7706f14e6396a173 --- /dev/null +++ b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bef23e0d1961e49e06153326364e1dc9e7a9ced1a5f6ec9f28059530e908aafa +size 703860 diff --git a/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/layout.json b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a6b82ea823702d1e0d028235abdfe4b8e30fffd --- /dev/null +++ b/2024/BlinkVision_ A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events/layout.json @@ -0,0 +1,9533 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 163 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Optical Flow, Scene Flow and Point Tracking Estimation using RGB Frames and Events" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "spans": [ + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": "Yijin Li" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Yichen Shen" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Zhaoyang Huang" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Shuo Chen" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Weikang Bian" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Xiaoyu Shi" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Fu-Yun Wang" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Keqiang Sun" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Hujun Bao" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Zhaopeng Cui" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", Guofeng Zhang" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "text", + "content": ", and Hongsheng Li" + }, + { + "bbox": [ + 136, + 185, + 478, + 224 + ], + "type": "inline_equation", + "content": "^{3,4,5\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 181, + 231, + 430, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 231, + 430, + 243 + ], + "spans": [ + { + "bbox": [ + 181, + 231, + 430, + 243 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 231, + 430, + 243 + ], + "type": "text", + "content": " State Key Lab of CAD&CG, Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 243, + 430, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 243, + 430, + 254 + ], + "spans": [ + { + "bbox": [ + 181, + 243, + 430, + 254 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 181, + 243, + 430, + 254 + ], + "type": "text", + "content": " Avolution AI " + }, + { + "bbox": [ + 181, + 243, + 430, + 254 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 181, + 243, + 430, + 254 + ], + "type": "text", + "content": " CUHK MMLab " + }, + { + "bbox": [ + 181, + 243, + 430, + 254 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 181, + 243, + 430, + 254 + ], + "type": "text", + "content": " Shanghai AI Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 254, + 352, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 254, + 352, + 264 + ], + "spans": [ + { + "bbox": [ + 260, + 254, + 352, + 264 + ], + "type": "text", + "content": "5 CPII under InnoHK" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 295, + 455, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 295, + 455, + 515 + ], + "spans": [ + { + "bbox": [ + 159, + 295, + 455, + 515 + ], + "type": "text", + "content": "Abstract. Recent advances in event-based vision suggest that they complement traditional cameras by providing continuous observation without frame rate limitations and high dynamic range which are well-suited for correspondence tasks such as optical flow and point tracking. However, so far there is still a lack of comprehensive benchmarks for correspondence tasks with both event data and images. To fill this gap, we propose BlinkVision, a large-scale and diverse benchmark with rich modality and dense annotation of correspondence. BlinkVision has several appealing properties: 1) Rich modalities: It encompasses both event data and RGB images. 2) Rich annotations: It provides dense per-pixel annotations covering optical flow, scene flow, and point tracking. 3) Large vocabulary: It incorporates 410 daily categories, sharing common classes with widely-used 2D and 3D datasets such as LVIS and ShapeNet. 4) Naturalistic: It delivers photorealism data and covers a variety of naturalistic factors such as camera shake and deformation. BlinkVision enables extensive benchmarks on three types of correspondence tasks (i.e., optical flow, point tracking and scene flow estimation) for both image-based methods and event-based methods, leading to new observations, practices, and insights for future research. The benchmark website is https://www.blinkvision.net/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "content": "Modern image-based computer vision technology still cannot match the accuracy and robustness of human vision in many areas. One possible reason is that traditional cameras suffer from motion blur and limited frame rates, and they often rely on well-lighted conditions. In contrast, event cameras [13, 36] detect changes in intensity at each pixel as a stream of asynchronous events, which" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 639, + 433, + 650 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 639, + 433, + 650 + ], + "spans": [ + { + "bbox": [ + 134, + 639, + 433, + 650 + ], + "type": "text", + "content": "* Yijin Li, Yichen Shen, and Zhaoyang Huang contributed equally to this work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 650, + 387, + 660 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 650, + 387, + 660 + ], + "spans": [ + { + "bbox": [ + 135, + 650, + 387, + 660 + ], + "type": "text", + "content": "† Hongsheng Li and Guofeng Zhang are the corresponding authors." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 478, + 273 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 273 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 273 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 273 + ], + "type": "image", + "image_path": "b0dc5854d567c50c6617ec6154781f8bdd52f6315f8ad9ca255ee96fc23579e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "type": "text", + "content": "Fig. 1: BlinkVision is a large-scale and diverse benchmark with rich modality and dense annotation of correspondence. It covers 410 daily categories, sharing common classes with popular 2D and 3D datasets. The per-category object distributions, scene structure hierarchy, data samples, and supported applications of BlinkVision are shown in this figure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 354, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 462 + ], + "type": "text", + "content": "eliminates frame rate limitations and enables operation within a high dynamic range. However, they can not capture fine-grained details as traditional cameras do. Looking at how human vision works, we find that two cells in the human retina, i.e., cones and rods, work similarly to these two types of cameras, respectively. According to the duplex theory of vision [45], which posits that rods and cones serve different functions, their combination ensures the robustness of human visual processing. Therefore, we believe that combining the advantages of traditional cameras and event cameras can significantly enhance computer vision systems, providing more comprehensive and adaptive vision capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 467, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 563 + ], + "type": "text", + "content": "However, there are currently only a few benchmarks [15, 32, 34] providing both event data and RGB frames, which hinder the development of algorithms that fully exploit the event data and fuse information from both modalities. This shortfall is particularly prominent in the domain of pixel correspondence estimation [8, 27, 28, 61], i.e., optical flow, scene flow, and point tracking estimation. Previous benchmarks built for pixel correspondence are either highly biased to specific scenes [15, 63] or rather simple [53]. A primary factor is that obtaining such precise pixel-wise annotations for these tasks is expensive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 568, + 482, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 653 + ], + "type": "text", + "content": "To boost the research in this area, we present BlinkVision, a synthetic benchmark for optical flow, scene flow and point tracking estimation using RGB frames and events. Our dataset has several appealing properties: 1) Rich modalities: BlinkVision encompasses three visual modalities: final RGB images, clean RGB images and event data. The final RGB images reflect real-world challenges like motion blur and limited dynamic range while the clean RGB images are devoid of such imperfections. The clean RGB images can be seen as the latent images [48]" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 178, + 478, + 297 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: A comparison between BlinkVision and other widely-used benchmarks on pixel correspondence estimation. \"Occ\", \"Chara\" and \"Cats\" is the abbreviation for occlusion, character and categories, respectively. " + }, + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "R^{lvis}" + }, + { + "bbox": [ + 131, + 114, + 482, + 170 + ], + "type": "text", + "content": " denotes the ratio of the 1.2k LVIS [20] categories being covered. \"N/A\" denotes the dataset does not provide category labels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 178, + 478, + 297 + ], + "lines": [ + { + "bbox": [ + 133, + 178, + 478, + 297 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 478, + 297 + ], + "type": "table", + "html": "
DatasetsEventCleanFinalOptical FlowScene FlowPoint TrackingOccCharaAnimalCatsR1vis(%)
MVSEC [64]✓X✓✓XXXXXN/A0
DSEC [15]✓X✓✓XXXXXN/A0
BlinkFlow [34]✓✓X✓XX✓✓X554.1
EKubric [59]✓✓X✓✓✓✓XX170.9
Sintel [9]X✓X✓XX✓✓✓N/A0
KITTI [17]XX✓✓✓XXXXN/A0
FlyingThings [42]X✓X✓✓X✓XX554.1
Kubric [18]X✓X✓✓✓✓XX170.9
TAP-Vid [12]X✓XXX✓X✓XN/A0
PointOdyssey [62]X✓XXX✓✓✓✓N/A0
BlinkVision (Ours)✓✓✓✓✓✓✓✓✓41033.4
", + "image_path": "d786e973c7eef1932a7028251f76ad7dc15b2e2194d84d7aa17975794bc7cf7f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 317, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 472 + ], + "type": "text", + "content": "of event cameras. 2) Rich annotations: It provides annotations covering optical flow, scene flow, and point tracking. Unlike benchmarks [15,17,62] which only contain sparse annotations, our data provide dense per-pixel annotations of each image, covering objects including moving cars, deformable characters, and animals. 3) Large vocabulary: It incorporates 410 daily categories, sharing common classes with widely-used 2D and 3D datasets such as ImageNet [33], LVIS [20], and ShapeNet [10], as depicted in Fig. 1. To our knowledge, our data have the widest variety of objects in existing pixel correspondence benchmarks. This expansive vocabulary is pivotal, enabling rigorous exploration of algorithmic generalization across varied objects. 4) Naturalistic: BlinkVision employs high-quality assets and rendering tools and thus is able to deliver photorealism data. Besides, it covers a variety of naturalistic factors such as camera shake and deformation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "To make full use of BlinkVision, we set up a public benchmark website that allows uploading results and provides a public leaderboard. Besides, we evaluated both existing image-based and event-based methods on three typical correspondence tasks (i.e. optical flow, point tracking, and scene flow estimation). The results reveal new observations and challenges and serve as the baseline for future approaches. Specifically, we first study the robustness of existing image-based methods under large frame intervals and extreme illumination and point out the new challenge for these methods. Second, the benchmark results on existing event-based methods show that current methods do not fully unleash the potential of event cameras. Third, we show that fine-tuning existing methods on the training set of BlinkVision significantly boosts the generalizability, demonstrating the vast diversity of BlinkVision. Finally, the broad diversity and accessible category labels in BlinkVision allow us, for the first time, to analyze the performance of correspondence tasks on different categories. We believe BlinkVision has the potential to serve as a cornerstone benchmark for advancing the development of more robust computer vision systems." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 223 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 223 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 223 + ], + "type": "image", + "image_path": "faa359d42f2eb44ad9aa4bb22d01923d53e76a236e09fd281055ae554fc13d60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 232, + 454, + 244 + ], + "lines": [ + { + "bbox": [ + 158, + 232, + 454, + 244 + ], + "spans": [ + { + "bbox": [ + 158, + 232, + 454, + 244 + ], + "type": "text", + "content": "Fig. 2: Scenes samples in the proposed BlinkVision benchmark." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 257, + 237, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 237, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 237, + 270 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 288, + 488, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 488, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 488, + 312 + ], + "type": "text", + "content": "A comprehensive overview and comparison between these benchmarks and BlinkVision is listed in Table 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 485 + ], + "type": "text", + "content": "Image-based optical flow and scene flow benchmarks. Early evaluations depended on synthetic datasets, like the famous \"Yosemite\" sequence [7]. MPI Sintel [9] was one of the most representative synthetic benchmarks derived from a short open-source animated 3D movie and it became one of the most popular benchmark datasets. KITTI [17] was almost the most well-known among real data. It computed ground truth for static scenes through the data from a 3D laser scanner and the ego-motion data of the car. In a later version, KITTI extended the ground truth to rigidly moving cars by fitting CAD models of cars. FlyingThings3D [42] was a more recent synthetic dataset. It contained the ground truth of both optical flow and scene flow. FlyingThings3D and KITTI were the two most commonly used benchmarks in scene flow evaluation. However, on both datasets, ground truth about scene flow was not available in occluded regions. Furthermore, on KITTI, ground truth for foreground points was either missing or approximated by a fitted car CAD model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 635 + ], + "type": "text", + "content": "Image-based point tracking benchmarks. The first point-tracking dataset was FlyingThings++ [21], which was based on FlyingThings3D. It was initially designed for training the network. Due to the lack of evaluation benchmarks, it was also used for evaluation in the early stages. Later, Doersch et al. [12] proposed a real data benchmark named TAP-Vid, which was based on two real-world datasets: DAVIS [49] and Kinetics [30]. TAP-Vid relied on manual annotation and could not handle occlusions. PointOdyssey [62] was a newly proposed benchmark that was based on synthetic data and therefore could provide dense ground truth. However, it did not guarantee that every pixel had ground truth because it only tracked the mesh vertices of the object. Furthermore, the naturalism of PointOdyssey was limited to its interior parts. Its outdoor portion had no realistic layout, just a skybox with randomly dropped objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "text", + "content": "Event-based Benchmarks. Early event-based optical flow benchmarks [53] limited the camera to only rotational motion and inferred ground truth from" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": "the rotational motion of the camera. Two later benchmarks, DSEC [15] and MVSEC [64] computed the ground truth through LiDAR SLAM. Similar to KITTI, these two benchmarks limited the ground truth to the static elements of the scene. Besides, they had a rather limited motion pattern. More recently, Li et al. [34] proposed a large-scale diversified form synthetic benchmark named BlinkFlow, which presented more challenging scenes and complex motion patterns. As for the scene flow benchmark, Wan et al. [59] recently converted the existing FlyingThings [42] and Kubric [18] datasets to event datasets through video-to-event [14] technology. However, due to the limited frame rates in the original video data, this step inevitably produced artifacts in the event data. Wan et al. also extended the real-world event dataset DSEC with scene flow ground truth. As for the event-based point-tracking benchmark, existing methods usually employed Structure-from-Motion [54] (SfM) technology to associate keypoints across multiple frames to obtain correspondence ground truth. Event Camera Dataset [46] and EDS dataset [22] were two common benchmarks in this area. Limited by the sparsity and accuracy of SfM, these benchmarks did not support the evaluation of arbitrary point tracking." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 341, + 272, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 272, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 272, + 354 + ], + "type": "text", + "content": "3 BlinkVision Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": "In this section, we describe how to build BlinkVision, including the scene setup, data rendering, and generation for the ground truth labels. The process is based on Blender [2] as it provides photorealistic rendering and a flexible data interface that allows us to obtain customized correspondence ground truth. At the last, we introduce the statistics and distribution of the BlinkVision data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 450, + 230, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 230, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 230, + 464 + ], + "type": "text", + "content": "3.1 Scene Setting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "content": "Previous synthetic benchmarks [9, 43] generally rely on open-source movies to avoid heavy scene construction. However, these movies are biased and do not cover diverse enough scenarios. In order to establish a comprehensive evaluation benchmark, we manually assemble a collection of scenarios that are as rich and diverse as possible. We first look for ready-made scenes that are photorealistic and visually diverse. Specifically, we purchased 40 indoor scenes and 13 outdoor scenes from Evermotion Archinteriors Collection [3] that cover common scenes such as living rooms, kitchens, offices, bedrooms, restaurants and gardens. The original scene is static. To enhance the realism, we procured 29 scanned and high-fidelity human bodies from ActorCore [1] and 88 artist-designed animals, together with 104 free characters from Mixamo [4]. These assets are rigged. We re-targeted them to a wide range of motion models (e.g., more than 100 human motions including dancing, walking, talking, etc.) and placed them in various locations. The human's motions mainly come from motion capture [1] and the animal's motions are designed by the artist to closely emulate their natural motions. Furthermore, we built 50 additional outdoor scenes that cover cases such" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "as a drone flying through the forest and a camera following a human or car from the crowd, which are common but rarely included in existing correspondence benchmarks. We set the camera trajectory by referencing the shooting trajectories of real-world videos, including handheld shots, car shots, and drone shots that cover non-uniform motion such as sudden stops and sharp turns. Finally, we obtained 80 indoor sequences and 63 outdoor sequences, where 56 indoor sequences and 47 outdoor sequences are for testing and the others are used for fine-tuning. Some samples of our data are shown in Fig. 2. In the supplementary, we also provide video samples. All the assets used in the testing and fine-tuning are totally disjoint, even including trees." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 257, + 321, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 321, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 321, + 269 + ], + "type": "text", + "content": "3.2 Multi-modality Data Rendering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 280, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 353 + ], + "type": "text", + "content": "We use Blender to render brightness " + }, + { + "bbox": [ + 130, + 280, + 482, + 353 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 280, + 482, + 353 + ], + "type": "text", + "content": " in linear color space. The following simulations of real-world image capturing and event data are both based on the linear color space. We employ tone mapping and gamma correction on " + }, + { + "bbox": [ + 130, + 280, + 482, + 353 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 280, + 482, + 353 + ], + "type": "text", + "content": " to obtain commonly used sRGB images. We call these sRGB images \"RGB (clean)\" because they do not suffer from effects like blurs or overexposed. On the contrary, we call the simulated real-world image capture \"RGB (final)\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "content": "Simulation of Real-world Image Capturing. Real-world image capture suffers from motion blur and limited dynamic range. While Blender supports the simulation of the former, the latter needs additional processing. The latter effect usually leads to overexposed or underexposed. To simulate it, we follow previous work [40] and employ the following steps: (1) Random exposure ratio. We uniformly sample the exposure ratio " + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "content": " in the log2 space within " + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "inline_equation", + "content": "[-3, 3]" + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "content": " and multiply it by " + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "content": " to simulate underexposed or overexposed, which gives us the augmented HDR image " + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "inline_equation", + "content": "H = B \\times 2^w" + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "content": ". (2) Dynamic range clipping. We clip " + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "content": " according to the formulation " + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(H) = \\min(H, 1)" + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "content": ". This step leads to information loss for pixels in the overexposed regions. (3) Non-linear mapping. To align with how humans see a scene, a camera typically uses a non-linear camera response function (CRF) to modify the contrast of the captured image, which can be formulated by " + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "inline_equation", + "content": "I_n = \\mathcal{F}(I_c)" + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "content": ". We randomly sample CRFs from an existing dataset [19]. (4) Quantization. The pixel values are quantized to 8 bits by " + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(I_n) = \\lfloor 255 \\times I_n + 0.5 \\rfloor / 255" + }, + { + "bbox": [ + 130, + 357, + 482, + 548 + ], + "type": "text", + "content": ". This step causes information loss in underexposed and smooth gradient areas." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "text", + "content": "Event Data. Event cameras work by responding to changes in the logarithmic brightness signal (i.e., " + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "inline_equation", + "content": "L = \\log B" + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "text", + "content": ") asynchronously and independently for each pixel [13]. An event is triggered when the change in brightness (either increase or decrease) since the last event at that pixel reaches a threshold of " + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\pm T" + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "inline_equation", + "content": "T > 0" + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 613, + 480, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 613, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 230, + 613, + 480, + 624 + ], + "type": "interline_equation", + "content": "p _ {k} \\left(L \\left(u _ {k}, t _ {k}\\right) - L \\left(u _ {k}, t _ {k} - \\Delta t\\right)\\right) \\geq T, \\tag {1}", + "image_path": "c88090b945d8fd23412e349f43bdd3f54ad8b10325de406547936180d70eec36.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " is the time since the last event triggered at pixel location " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "u_{k}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " and at time " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "p_k\\in \\{-1,1\\}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " is the polarity of the brightness change. To model the event generation process, we need access to a continuous representation of the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 117, + 216, + 198 + ], + "blocks": [ + { + "bbox": [ + 135, + 117, + 216, + 198 + ], + "lines": [ + { + "bbox": [ + 135, + 117, + 216, + 198 + ], + "spans": [ + { + "bbox": [ + 135, + 117, + 216, + 198 + ], + "type": "image", + "image_path": "33e38f8766414540ca2828fad75c6240fec90d90663cb8bf0c14a78b1f65fb68.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 209, + 414, + 220 + ], + "lines": [ + { + "bbox": [ + 197, + 209, + 414, + 220 + ], + "spans": [ + { + "bbox": [ + 197, + 209, + 414, + 220 + ], + "type": "text", + "content": "Fig. 3: Statistics of optical flow in BlinkVision." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 118, + 304, + 198 + ], + "blocks": [ + { + "bbox": [ + 222, + 118, + 304, + 198 + ], + "lines": [ + { + "bbox": [ + 222, + 118, + 304, + 198 + ], + "spans": [ + { + "bbox": [ + 222, + 118, + 304, + 198 + ], + "type": "image", + "image_path": "7871dd5c68c7fae8bde7cd421ec565d2288d0542eb6dff9bcbbcb8f8c0b67c08.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 118, + 391, + 198 + ], + "blocks": [ + { + "bbox": [ + 310, + 118, + 391, + 198 + ], + "lines": [ + { + "bbox": [ + 310, + 118, + 391, + 198 + ], + "spans": [ + { + "bbox": [ + 310, + 118, + 391, + 198 + ], + "type": "image", + "image_path": "9f111e60c5c2783ef8312e798bc837738634169ed62cb823562f09deacb49dc7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 398, + 118, + 478, + 198 + ], + "blocks": [ + { + "bbox": [ + 398, + 118, + 478, + 198 + ], + "lines": [ + { + "bbox": [ + 398, + 118, + 478, + 198 + ], + "spans": [ + { + "bbox": [ + 398, + 118, + 478, + 198 + ], + "type": "image", + "image_path": "4a9af71414961f4bea91647eed0cb78037f93e8562907b89e6b343b884fa1ad3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 136, + 233, + 216, + 312 + ], + "blocks": [ + { + "bbox": [ + 136, + 233, + 216, + 312 + ], + "lines": [ + { + "bbox": [ + 136, + 233, + 216, + 312 + ], + "spans": [ + { + "bbox": [ + 136, + 233, + 216, + 312 + ], + "type": "image", + "image_path": "f8aa04e1ff8f98f4f6a6a10e7717999c28a80b2053f4b1e241eba7e8ae8941fd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 324, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 380 + ], + "type": "text", + "content": "Fig. 4: Statistics of point trajectories in BlinkVision. To save time, we sample the tracks using a grid with a size of 20. Trajectory segments are defined as contiguous sections of point trajectories, with interruptions caused by occlusion. The diameter is the maximum distance a point moves over time. We clip the trajectory diameter and divide it by the diagonal length of the image to obtain the ratio." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 223, + 234, + 302, + 312 + ], + "blocks": [ + { + "bbox": [ + 223, + 234, + 302, + 312 + ], + "lines": [ + { + "bbox": [ + 223, + 234, + 302, + 312 + ], + "spans": [ + { + "bbox": [ + 223, + 234, + 302, + 312 + ], + "type": "image", + "image_path": "b370f6293496785a4645d120e316adc729661801ed343c8c178344b0ee1bc280.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 234, + 389, + 312 + ], + "blocks": [ + { + "bbox": [ + 310, + 234, + 389, + 312 + ], + "lines": [ + { + "bbox": [ + 310, + 234, + 389, + 312 + ], + "spans": [ + { + "bbox": [ + 310, + 234, + 389, + 312 + ], + "type": "image", + "image_path": "86231bbda378e7cd63d0d0fe98425d913927a299cce9eb4ad10fc2b61fb12cb3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 397, + 234, + 476, + 312 + ], + "blocks": [ + { + "bbox": [ + 397, + 234, + 476, + 312 + ], + "lines": [ + { + "bbox": [ + 397, + 234, + 476, + 312 + ], + "spans": [ + { + "bbox": [ + 397, + 234, + 476, + 312 + ], + "type": "image", + "image_path": "8862018015469665aa541f02eab472549cf493b168b480afefd79682dcac6dfb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 403, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 511 + ], + "type": "text", + "content": "visual signal for each pixel. In practice, it is approximated by rendering images at high frame rates for efficiency [34, 51]. Events are then synthesized based on frame-by-frame pixel differences. More specifically, we adaptively sample frames according to [34, 51] to ensure that the maximum pixel displacement between two sampling timestamps is bounded. In practice, we employ frame interpolation before simulating the event data to further reduce the pixel displacement. We use DVS-Voltmeter [37] to synthesize events because it can simulate complex noise effects (such as noise effects of temperature and parasitic photocurrent), thus generating realistic events." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 534, + 292, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 534, + 292, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 534, + 292, + 545 + ], + "type": "text", + "content": "3.3 Ground Truth Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 560, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 481, + 608 + ], + "type": "text", + "content": "Blender provides optical flow data between two consecutive frames and a segmentation mask for each object. However, we cannot directly obtain these data from Blender: (1) pixel correspondence across multiple frames; (2) scene flow between frames; (3) semantic category for each object." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": "Pixel Correspondence. The data of (1) and (2) can be computed in a unified framework. Given camera poses " + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": " (camera-to-world), depth maps " + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": " at frames " + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": ", and the camera intrinsic " + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": ", for the pixel location " + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "inline_equation", + "content": "\\mathfrak{u}" + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": ", we first project it to world coordinate: " + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "inline_equation", + "content": "P = T_{i}Z_{i}(\\mathfrak{u})K^{-1}\\tilde{\\mathfrak{u}}" + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathfrak{u}}" + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": " is the homogeneous" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "text", + "content": "coordinates of u. Then we re-project the 3D point " + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "inline_equation", + "content": "du_{j} = KT_{j}^{-1}P" + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "inline_equation", + "content": "\\mathfrak{u}_j" + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "text", + "content": " is the corresponding pixel location and " + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "text", + "content": " is the depth value at frame " + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 132, + 116, + 480, + 164 + ], + "type": "text", + "content": ". In this way, we get the pixel correspondence between any two frames and the corresponding depth, i.e., (1) and (2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 169, + 482, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 169, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 482, + 277 + ], + "type": "text", + "content": "However, this only works for static objects. To handle dynamic objects and even those objects with deformation motion such as humans and animals, we bake the face index and barycentric coordinate " + }, + { + "bbox": [ + 130, + 169, + 482, + 277 + ], + "type": "inline_equation", + "content": "(\\lambda_1,\\lambda_2,\\lambda_3)" + }, + { + "bbox": [ + 130, + 169, + 482, + 277 + ], + "type": "text", + "content": " of each mesh face into textures. After rendering the texture, we locate the triangular face that the tracked pixel belongs to (indexed by the rendered face index) and obtain the vertices' 3D position " + }, + { + "bbox": [ + 130, + 169, + 482, + 277 + ], + "type": "inline_equation", + "content": "(V_{1},V_{2},V_{3})" + }, + { + "bbox": [ + 130, + 169, + 482, + 277 + ], + "type": "text", + "content": " of that face via Blender's API. The 3D position of tracked pixels can be computed through " + }, + { + "bbox": [ + 130, + 169, + 482, + 277 + ], + "type": "inline_equation", + "content": "P = \\lambda_1V_1 + \\lambda_2V_2 + \\lambda_3V_3" + }, + { + "bbox": [ + 130, + 169, + 482, + 277 + ], + "type": "text", + "content": ". This step replaces the back projection for the static objects. It is worth noting that we need to triangulate every mesh face in the scenes before baking." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "text", + "content": "Semantic Labeling. BlinkVision contains thousands of objects which makes manual semantic labeling expensive. To this end, we develop an automatic labeling framework that is based on open-vocabulary semantic understanding [50,60]. First, we pre-define a category list based on LVIS [20]. The category together with category descriptions is encoded into text embedding " + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "inline_equation", + "content": "F_{t} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "text", + "content": " with a pre-trained text encoder, i.e., CLIP, where " + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "text", + "content": " is the number of categories and " + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "text", + "content": " is the feature dimension. For each object asset, we render it individually to a " + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "text", + "content": " image. The process is rather fast while keeping photorealistic because it does not need to compute complex path tracing between objects. We use the pre-trained image encoder from CLIP [50] to extract the image embedding " + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "inline_equation", + "content": "F_{i} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "text", + "content": ". The probability of the image belonging to one of the pre-defined categories is computed through " + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "inline_equation", + "content": "P(c|F_{i}) = \\text{softmax}\\left(F_{i} \\cdot (F_{t})^{\\top}\\right)_{c}" + }, + { + "bbox": [ + 130, + 283, + 482, + 429 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 450, + 251, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 251, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 251, + 462 + ], + "type": "text", + "content": "3.4 Dataset Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "text", + "content": "BlinkVision consists of 40 training sequences and 103 testing sequences. The training set provides 107,880 frame pairs for optical flow and scene flow, and 1025 sub-sequences for point tracking. The frame rate for RGB images and ground truth data is 20 FPS. The frame rate is not practically significant because we adjusted it to ensure sufficient motion between adjacent frames. This adjustment reduces the data size uploaded to the benchmark website. For each sequence of BlinkVision, we selected the sub-sequences by starting tracking from the first frame. Once the overlap with the reference frame (i.e., how many tracked points remain inside the image) is smaller than " + }, + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "text", + "content": ", we start to track a new reference frame. If the overlap is smaller than " + }, + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "text", + "content": ", we stop this track. Additionally, we discard short tracks that are less than 20 frames long because they are less challenging without long-term motion accumulation, which could result in the benchmark easily reaching saturation. As a result, we generate 1025 subsequences where each sub-sequence contains " + }, + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "text", + "content": " dense point trajectories. For the test set, we selected 12,804 frame pairs for evaluating optical flow and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 212, + 162 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 212, + 162 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 212, + 162 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 212, + 162 + ], + "type": "image", + "image_path": "00ed5a18729acc643bb21eeccd2a63109cc39473c56cf53006ba05d5f5788de0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 163, + 212, + 209 + ], + "blocks": [ + { + "bbox": [ + 151, + 163, + 212, + 209 + ], + "lines": [ + { + "bbox": [ + 151, + 163, + 212, + 209 + ], + "spans": [ + { + "bbox": [ + 151, + 163, + 212, + 209 + ], + "type": "image", + "image_path": "ec217484f460347a83417bf4b61be13b7f9345ec6a8eb73e818306994ff7fb8c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 210, + 206, + 219 + ], + "lines": [ + { + "bbox": [ + 159, + 210, + 206, + 219 + ], + "spans": [ + { + "bbox": [ + 159, + 210, + 206, + 219 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 214, + 116, + 275, + 162 + ], + "blocks": [ + { + "bbox": [ + 214, + 116, + 275, + 162 + ], + "lines": [ + { + "bbox": [ + 214, + 116, + 275, + 162 + ], + "spans": [ + { + "bbox": [ + 214, + 116, + 275, + 162 + ], + "type": "image", + "image_path": "6a62f30eb2cc1b147b0ef6c389f51d55993217a41c1166d54c646d760128a717.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 214, + 163, + 275, + 209 + ], + "blocks": [ + { + "bbox": [ + 214, + 163, + 275, + 209 + ], + "lines": [ + { + "bbox": [ + 214, + 163, + 275, + 209 + ], + "spans": [ + { + "bbox": [ + 214, + 163, + 275, + 209 + ], + "type": "image", + "image_path": "b99d97f9f3ad5542bf0153e156c7ebcf68a522898923a7a0f5a999a61a301813.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 210, + 263, + 219 + ], + "lines": [ + { + "bbox": [ + 227, + 210, + 263, + 219 + ], + "spans": [ + { + "bbox": [ + 227, + 210, + 263, + 219 + ], + "type": "text", + "content": "Target Image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 276, + 115, + 337, + 162 + ], + "blocks": [ + { + "bbox": [ + 276, + 115, + 337, + 162 + ], + "lines": [ + { + "bbox": [ + 276, + 115, + 337, + 162 + ], + "spans": [ + { + "bbox": [ + 276, + 115, + 337, + 162 + ], + "type": "image", + "image_path": "0f3e3dbfc1006eec3cbb8c19cfb8a1962001a2ff168c1fb62b95db57677f3e7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 276, + 163, + 337, + 209 + ], + "blocks": [ + { + "bbox": [ + 276, + 163, + 337, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 163, + 337, + 209 + ], + "spans": [ + { + "bbox": [ + 276, + 163, + 337, + 209 + ], + "type": "image", + "image_path": "916d4e5e3c8285620e4c6dd491c1115d00cebb57f1d54b5df118f1f39b2ee9a9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 210, + 325, + 218 + ], + "lines": [ + { + "bbox": [ + 287, + 210, + 325, + 218 + ], + "spans": [ + { + "bbox": [ + 287, + 210, + 325, + 218 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 228, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 480, + 251 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of FlowFormer++ [55] before and after fine-tuning on the training set of BlinkVision." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 338, + 115, + 399, + 162 + ], + "blocks": [ + { + "bbox": [ + 338, + 115, + 399, + 162 + ], + "lines": [ + { + "bbox": [ + 338, + 115, + 399, + 162 + ], + "spans": [ + { + "bbox": [ + 338, + 115, + 399, + 162 + ], + "type": "image", + "image_path": "37a3e35769199669e38113ddc173d1ee66152c37b17ee54ea9de0b4fb561ffa1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 338, + 163, + 399, + 209 + ], + "blocks": [ + { + "bbox": [ + 338, + 163, + 399, + 209 + ], + "lines": [ + { + "bbox": [ + 338, + 163, + 399, + 209 + ], + "spans": [ + { + "bbox": [ + 338, + 163, + 399, + 209 + ], + "type": "image", + "image_path": "1bb862dc35943f791ba10b72447595bff9fa6fd797b6563268ce4ed619da5876.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 210, + 394, + 219 + ], + "lines": [ + { + "bbox": [ + 345, + 210, + 394, + 219 + ], + "spans": [ + { + "bbox": [ + 345, + 210, + 394, + 219 + ], + "type": "text", + "content": "Before Finetuning" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 400, + 115, + 462, + 162 + ], + "blocks": [ + { + "bbox": [ + 400, + 115, + 462, + 162 + ], + "lines": [ + { + "bbox": [ + 400, + 115, + 462, + 162 + ], + "spans": [ + { + "bbox": [ + 400, + 115, + 462, + 162 + ], + "type": "image", + "image_path": "81d74fd5de321b0235f96293c83e90b23dbd16a1facf962037c9e2a64bf5d673.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 400, + 163, + 462, + 209 + ], + "blocks": [ + { + "bbox": [ + 400, + 163, + 462, + 209 + ], + "lines": [ + { + "bbox": [ + 400, + 163, + 462, + 209 + ], + "spans": [ + { + "bbox": [ + 400, + 163, + 462, + 209 + ], + "type": "image", + "image_path": "b84c7682a80a7267e4b60b98fe7da5ce680eec9e3b658dce7fac5dc116011736.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 210, + 454, + 219 + ], + "lines": [ + { + "bbox": [ + 408, + 210, + 454, + 219 + ], + "spans": [ + { + "bbox": [ + 408, + 210, + 454, + 219 + ], + "type": "text", + "content": "After Finetuning" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 273, + 479, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 479, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 479, + 321 + ], + "type": "text", + "content": "scene flow, and 865 sub-sequences for evaluating point tracking. BlinkVision includes raw data on depth, normal, camera pose, etc. At this stage, we focus on pixel-level motion tasks. The additional raw data will be released later, which will benefit other tasks such as depth estimation [35] and SLAM [23, 24, 39]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 322, + 480, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 480, + 370 + ], + "type": "text", + "content": "Table 1 summarizes key statistics of BlinkVision compared to related works. BlinkVision contains more than 40K 3D models in 410 categories. It shares some common categories with 2D and 3D datasets [10,20], and bears a huge diversity in semantics, geometry, and appearance, enabling a wide range of research topics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 371, + 481, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 481, + 587 + ], + "type": "text", + "content": "We provide statistics of optical flow in Fig. 3. For benchmarking, we discard those pixels whose flow magnitudes are larger than half of the image diagonal. The latter situation usually occurs when the object is too close to the camera, and its magnitude can even approach infinity. As a result, the magnitude shown in Fig. 3 has an upper bound of 400. The broad distribution in Fig. 3 reveals the diverse motions in BlinkVision. Besides, the diverse distribution of \"X Variance\" and \"Y Variance\" also indicates the presence of complex object shapes in BlinkVision. We also provide statistics of point trajectories in Fig. 4. BlinkVision allows the evaluation of point tracking with different trajectories' lengths, from small to big, corresponding to different downstream applications like video editing [25, 41] or augmented reality [31, 39]. The broad distribution of \"occlusion ratio\" and \"number of segments\" indicates that BlinkVision brings diverse and challenging data. The occlusion ratio has an upper bound that is less than 100 percent due to our strategy of sub-sequences selection. To quantify the amount of motion a point undergoes, we calculate its trajectory diameter, which is the maximum distance between any two positions of the point over its entire trajectory. We observe that point motion in BlinkVision is diverse and nearly uniformly distributed across the entire image plane." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 605, + 222, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 222, + 617 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 222, + 617 + ], + "type": "text", + "content": "4 Benchmark" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "We release all the data except for the ground truth for the test split. For a fair comparison, we create a public benchmark website with leaderboards. We also release the evaluation code for our online benchmarking." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 114, + 296, + 171 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 296, + 171 + ], + "type": "text", + "content": "Table 2: Optical flow result of RGB-frame-based methods. \"St.\" denotes Stride. \"†\" denotes fine-tuning. \"FF\" denotes FlowFormer [26] and \"FF++\" denotes FlowFormer++ [55]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 132, + 179, + 297, + 338 + ], + "blocks": [ + { + "bbox": [ + 132, + 179, + 297, + 338 + ], + "lines": [ + { + "bbox": [ + 132, + 179, + 297, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 297, + 338 + ], + "type": "table", + "html": "
MethodEPE - clean
St. 1St. 2St. 4St. 8Avg.
RAFT [56]1.834.6210.1820.469.27
GMA [29]1.824.379.5218.688.60
FF [26]1.603.777.5716.037.24
FF++ [55]1.543.577.4316.267.20
RAFT† [56]1.332.966.3912.725.85
FF++† [55]1.092.354.8810.584.73
MethodEPE - final
St. 1St. 2St. 4St. 8Avg.
RAFT [56]2.545.6911.2722.6710.54
GMA [29]2.565.7211.2121.3510.21
FF [26]2.264.809.2418.238.63
FF++ [55]2.284.839.3919.088.90
RAFT† [56]1.943.887.3515.177.08
FF++† [55]1.663.346.4212.736.04
", + "image_path": "9c7f15d76058a88f6b90a9abea79470823041bc2f0c291bc823a80da4c875628.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 301, + 158, + 484, + 338 + ], + "blocks": [ + { + "bbox": [ + 299, + 114, + 484, + 148 + ], + "lines": [ + { + "bbox": [ + 299, + 114, + 484, + 148 + ], + "spans": [ + { + "bbox": [ + 299, + 114, + 484, + 148 + ], + "type": "text", + "content": "Table 3: Optical flow result of event-based methods. \"St.\" denotes Stride. \"†\" denotes fine-tuning." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 301, + 158, + 484, + 338 + ], + "lines": [ + { + "bbox": [ + 301, + 158, + 484, + 338 + ], + "spans": [ + { + "bbox": [ + 301, + 158, + 484, + 338 + ], + "type": "table", + "html": "
Event
Methods|St. 1St. 2St. 4St. 8Avg.
E-RAFT [16]2.817.0417.8228.6014.07
STE-FlowNet [11]2.595.9413.1341.8915.89
E-FlowFormer [34]2.415.6612.5330.4512.76
E-RAFT† [16]1.683.637.4814.206.75
E-FlowFormer† [34]1.513.005.9613.606.02
Event + RGB (clean)
Methods|St. 1St. 2St. 4St. 8Avg.
DCEIFlow [58]3.3114.0734.7861.5628.43
DCEIFlow† [58]2.196.3012.5426.1111.79
Event + RGB (final)
Methods|St. 1St. 2St. 4St. 8Avg.
DCEIFlow [58]4.0814.2434.5461.2728.53
DCEIFlow† [58]2.526.8613.6428.4512.87
", + "image_path": "5d60b8387a531622172cb55b533af88f9d5e875459e3368bbb96c34d73bf10ac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 361, + 226, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 226, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 226, + 373 + ], + "type": "text", + "content": "4.1 Optical Flow" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 386, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 518 + ], + "type": "text", + "content": "In this section, we comprehensively evaluate the performance of existing optical flow methods on BlinkVision. Specifically, we first analyze the robustness of image-based methods under large frame intervals and extreme illumination. Large motion, severe occlusion, and information loss in these extreme cases pose great challenges to existing image-based methods. In contrast, event cameras naturally depict continuous pixel motion and possess a large dynamic range, thus offering great potential for solving these challenges. We then benchmark existing event-based methods, including event-only methods and event-RGB fusion methods. We find that existing event-based methods cannot fully unleash the potential of event cameras. We analyze the possible reasons and point to possible opportunities for future research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 523, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 597 + ], + "type": "text", + "content": "Experimental Setup. To study the impact of frame intervals on existing methods, we select frames separated by 1, 2, 4, and 8 from the reference frame as target frames. To reduce the file size of results that are uploaded to the benchmark website, we uniformly sample a reference frame from ten consecutive frames. As a result, there are 12,804 frame pairs for testing in total. We follow the previous work [26, 47, 56] and report the EPE (end-point-error) as the evaluation metric." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 601, + 482, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 662 + ], + "type": "text", + "content": "Results. We benchmark several image-based methods in Table 2. We make the following observations: 1) The error generally increases linearly with stride size. This is expected because doubling the stride yields motion that is twice as fast, resulting in an error that is also expected to be twice as high. 2) Extreme lighting has a heavy effect on existing image-based methods. In the supplementary" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 156, + 336, + 378 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 335, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 335, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 335, + 147 + ], + "type": "text", + "content": "Table 4: Point tracking result of RGB-frame-based methods under different frame interval. \"St.\" denotes Stride." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 156, + 336, + 378 + ], + "lines": [ + { + "bbox": [ + 132, + 156, + 336, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 336, + 378 + ], + "type": "table", + "html": "
RGB(clean)
MethodMetricSt. 1St. 2St. 4St. 8Avg.
PIPs [21]δavg37.7739.3436.3425.5034.74
Survival↑52.2855.3954.5547.5852.45
MTE↓50.7747.8149.5362.4252.63
PIPs++ [62]δavg40.0238.2530.2416.8431.34
Survival↑53.4453.1347.5739.7148.46
MTE↓45.5151.0259.8773.1757.39
Context-TAP [8]δavg38.7539.4736.5125.9935.18
Survival↑49.9853.0952.6146.1850.46
MTE↓51.4649.6151.8364.3954.32
RGB(final)
MethodMetricSt. 1St. 2St. 4St. 8Avg.
PIPs [21]δavg33.0234.4431.5621.7730.20
Survival↑48.3551.5650.7244.5648.80
MTE↓53.9451.5453.6566.1456.32
PIPs++ [62]δavg35.8334.2427.1115.3228.12
Survival↑50.4750.3645.3238.4646.15
MTE↓48.3753.7462.3474.5959.76
Context-TAP [8]δavg33.4634.1731.3922.0030.25
Survival↑45.5848.9148.5343.0646.52
MTE↓55.0953.7656.3368.1158.32
", + "image_path": "82e4b0ca3bb8f3ce553798088c4448dc04945152fb9e7aa59891021684b17892.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 341, + 191, + 481, + 378 + ], + "blocks": [ + { + "bbox": [ + 339, + 114, + 485, + 182 + ], + "lines": [ + { + "bbox": [ + 339, + 114, + 485, + 182 + ], + "spans": [ + { + "bbox": [ + 339, + 114, + 485, + 182 + ], + "type": "text", + "content": "Table 5: Point tracking result of event-based methods. \\* denotes that we track ORB [52] feature points rather than grid sampled points. DET\" denotes DeepETracker [44]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 341, + 191, + 481, + 378 + ], + "lines": [ + { + "bbox": [ + 341, + 191, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 341, + 191, + 481, + 378 + ], + "type": "table", + "html": "
Event
Methodδavg↑Survival↑MTE↓
AMH [5]18.9537.9875.44
HASTE [6]18.6037.4377.36
AMH* [5]26.7641.9362.48
HASTE* [6]27.6542.8762.52
Event + RGB (clean)
Methodδavg↑Survival↑MTE↓
DET [44]21.6235.4480.80
DET* [44]33.4648.8260.69
Event + RGB (final)
Methodδavg↑Survival↑MTE↓
DET [44]19.0533.5285.00
DET* [44]31.7848.6758.79
", + "image_path": "dc9e9da325a8268b7cc931b3084281e3e38cecc6fc6d529787a30eb7ec2a2334.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 480, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 474 + ], + "type": "text", + "content": "materials, we analyze how different exposure factors lead to performance loss. 3) Fine-tuning on the training split of BlinkVision significantly improves the performance. In Fig. 5 we show qualitative results of FlowFormer++ before and after fine-tuning on BlinkVision. After fine-tuning, the flow predictions are more precise and clear, especially near the object boundary. As shown in Table 2, the benchmark is still challenging for methods after fine-tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 479, + 482, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 659 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 659 + ], + "type": "text", + "content": "We report the evaluation of event-based methods in Table 3. It is surprising that the performance of existing event-based methods is worse than image-based methods and they also suffer from severe degradation under large frame intervals. The reasons are mainly twofold. First, existing training data for events lag behind that of RGB. This can be verified by the performance of E-RAFT [16] and E-FlowFormer [34] after fine-tuning. These two methods are far behind RAFT [56] before fine-tuning, but exceed the fine-tuned RAFT [56] and even FlowFormer++ [55] after fine-tuning (in \"Stride-1\", \"Stride-2\" and \"Stride-4\" of the \"final\" case), indicating that the poor performance of the former mainly comes from limited training on existing event-based training data. Second, existing event-based methods usually convert events to regular 3D voxel grids before processing, which quantizes the event data and leads to information loss. This problem is more serious when dealing with event data within large frame intervals. As a result, it calls for a more powerful representation of event data and new algorithms that can deal with long-range optical flow estimation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 197, + 163 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 163 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 163 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 163 + ], + "type": "image", + "image_path": "a7982135da724b9ece0dbb755629e1e824cd25daa5db236eab6ecb16f88d09de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 163, + 197, + 210 + ], + "blocks": [ + { + "bbox": [ + 133, + 163, + 197, + 210 + ], + "lines": [ + { + "bbox": [ + 133, + 163, + 197, + 210 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 197, + 210 + ], + "type": "image", + "image_path": "07efd0dd90f5376b5e39d86965790341a9bd44ede33f289c9840c11f62ebf74e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 212, + 189, + 220 + ], + "lines": [ + { + "bbox": [ + 140, + 212, + 189, + 220 + ], + "spans": [ + { + "bbox": [ + 140, + 212, + 189, + 220 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 197, + 114, + 260, + 163 + ], + "blocks": [ + { + "bbox": [ + 197, + 114, + 260, + 163 + ], + "lines": [ + { + "bbox": [ + 197, + 114, + 260, + 163 + ], + "spans": [ + { + "bbox": [ + 197, + 114, + 260, + 163 + ], + "type": "image", + "image_path": "4092383402648846b974d562746ae5ac1ca1896fbcfb47113019481ce6ba48d9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 197, + 163, + 260, + 210 + ], + "blocks": [ + { + "bbox": [ + 197, + 163, + 260, + 210 + ], + "lines": [ + { + "bbox": [ + 197, + 163, + 260, + 210 + ], + "spans": [ + { + "bbox": [ + 197, + 163, + 260, + 210 + ], + "type": "image", + "image_path": "06e0cb0547b1227a6b7b3847a431399063bebf1bd5322545bdf793a191d7594e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 212, + 255, + 220 + ], + "lines": [ + { + "bbox": [ + 203, + 212, + 255, + 220 + ], + "spans": [ + { + "bbox": [ + 203, + 212, + 255, + 220 + ], + "type": "text", + "content": "Prediction Results" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 261, + 114, + 324, + 163 + ], + "blocks": [ + { + "bbox": [ + 261, + 114, + 324, + 163 + ], + "lines": [ + { + "bbox": [ + 261, + 114, + 324, + 163 + ], + "spans": [ + { + "bbox": [ + 261, + 114, + 324, + 163 + ], + "type": "image", + "image_path": "61c8963d077cbbdfc03d9a3272b870c4677c82ea64270d7d0f35becf1dd76eea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 261, + 163, + 324, + 210 + ], + "blocks": [ + { + "bbox": [ + 261, + 163, + 324, + 210 + ], + "lines": [ + { + "bbox": [ + 261, + 163, + 324, + 210 + ], + "spans": [ + { + "bbox": [ + 261, + 163, + 324, + 210 + ], + "type": "image", + "image_path": "5e671c5a44f5ff742642096ef68661b643ec01918282fc3184b7a90ece145536.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 212, + 320, + 220 + ], + "lines": [ + { + "bbox": [ + 266, + 212, + 320, + 220 + ], + "spans": [ + { + "bbox": [ + 266, + 212, + 320, + 220 + ], + "type": "text", + "content": "Error Visualization" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 125, + 477, + 221 + ], + "blocks": [ + { + "bbox": [ + 328, + 125, + 477, + 221 + ], + "lines": [ + { + "bbox": [ + 328, + 125, + 477, + 221 + ], + "spans": [ + { + "bbox": [ + 328, + 125, + 477, + 221 + ], + "type": "image", + "image_path": "9f9cc860f4f23b7950a3edd2454b554088f6e7e66410a9518421f850953ecf2d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 324, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 324, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 324, + 262 + ], + "type": "text", + "content": "Fig. 6: Qualitative results of " + }, + { + "bbox": [ + 130, + 228, + 324, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{PIP}\\mathbf{s} + +" + }, + { + "bbox": [ + 130, + 228, + 324, + 262 + ], + "type": "text", + "content": " [62]. Darker lines in the right figure (e.g., red) indicate larger errors." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 229, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 329, + 229, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 329, + 229, + 482, + 253 + ], + "type": "text", + "content": "Fig. 7: Performance distribution of FlowFormer++ [55]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 273, + 312, + 305 + ], + "lines": [ + { + "bbox": [ + 131, + 273, + 312, + 305 + ], + "spans": [ + { + "bbox": [ + 131, + 273, + 312, + 305 + ], + "type": "text", + "content": "Table 6: Scene flow result of RGB-frame-based methods. \"St.\" denotes Stride." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 133, + 316, + 312, + 439 + ], + "blocks": [ + { + "bbox": [ + 133, + 316, + 312, + 439 + ], + "lines": [ + { + "bbox": [ + 133, + 316, + 312, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 316, + 312, + 439 + ], + "type": "table", + "html": "
EPE-2d
MethodRGBSt. 1St. 2St. 4St. 8
RAFT-3D [57]clean2.154.348.5916.08
final3.487.2910.7418.57
CamLiFlow [38]clean2.796.0512.3423.89
final4.288.2816.3030.61
EPE-3d
MethodRGBSt. 1St. 2St. 4St. 8
RAFT-3D [57]clean1.913.837.2512.18
final2.023.967.3914.21
CamLiFlow [38]clean1.452.765.369.12
final1.532.885.609.81
", + "image_path": "0afb1514deda57e71216f01ddc18b7cd29d159dd1115cf614662c93d26270c8a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 316, + 313, + 482, + 430 + ], + "blocks": [ + { + "bbox": [ + 315, + 281, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 315, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 315, + 281, + 481, + 304 + ], + "type": "text", + "content": "Table 7: Scene flow result of event-based methods. \"St.\" denotes Stride." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 313, + 482, + 430 + ], + "lines": [ + { + "bbox": [ + 316, + 313, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 313, + 482, + 430 + ], + "type": "table", + "html": "
EPE-2d
MethodRGBSt. 1St. 2St. 4St. 8
RPEFlow [59]clean1.533.507.3617.44
final1.924.108.1518.38
EPE-3d
MethodRGBSt. 1St. 2St. 4St. 8
RPEFlow [59]clean5.089.0715.7725.24
final5.209.3815.6725.03
", + "image_path": "a421525148022ce46a250ed5a134ce0d2d5ee2bee3a9b155207a9bebe9fa202e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 462, + 236, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 236, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 236, + 475 + ], + "type": "text", + "content": "4.2 Point Tracking" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": "In this section, we evaluate the performance of existing methods for long-term point tracking on BlinkVision. Similar to what we analyzed on optical flow, we analyze the robustness under large frame intervals and extreme illumination, and the gaps between existing image-based methods and event-based methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 540, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 601 + ], + "type": "text", + "content": "Experimental Setup. Although BlinkVision provides per-pixel dense annotations for point tracking, we find that existing methods are not efficient enough to process so much data. As a result, we grid sample the tracked pixels with a grid size of 20 pixels. We follow PointOdyssey [62] and use " + }, + { + "bbox": [ + 130, + 540, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{avg}}" + }, + { + "bbox": [ + 130, + 540, + 482, + 601 + ], + "type": "text", + "content": ", median trajectory error (MTE) and survival rate as the evaluation metrics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Results. We benchmark several image-based methods in Table 4. Similarly, we observe that extreme lighting has a major impact on existing image-based point tracking methods. In Table 4 we can see that larger frame intervals (such as 4 and 8) severely degrade performance. However, for PIPs [21] and Context TAP [8], increasing the interval from 1 to 2 makes the performance slightly better. We" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 156, + 319, + 333 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 319, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 319, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 319, + 147 + ], + "type": "text", + "content": "Table 8: Cross-dataset evaluation of optical flow methods fine-tuned on BlinkVision. \"BV\" denotes BlinkVision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 156, + 319, + 333 + ], + "lines": [ + { + "bbox": [ + 132, + 156, + 319, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 319, + 333 + ], + "type": "table", + "html": "
RGB
MethodDataSintel (clean)Sintel (final)KITTI
RAFT [56]- +BV2.08 1.693.41 3.045.10 4.66
Events
MethodDataE- BlenderFlying- ObjectsE- Tartan
E-RAFT [16]- +BV2.66 1.922.9 2.673.27 2.55
E-FlowFormer [34]- +BV2.38 1.752.89 2.582.91 2.36
Events + RGB
MethodDataE- BlenderFlying- ObjectsE- Tartan
DCEIFlow [58]- +BV8.96 4.769.58 2.377.44 2.88
", + "image_path": "ba6d1424830e6c044829ff1099bf4ed4314cc6538e3df16388db906c4f3cdf25.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 324, + 191, + 480, + 332 + ], + "blocks": [ + { + "bbox": [ + 322, + 114, + 480, + 180 + ], + "lines": [ + { + "bbox": [ + 322, + 114, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 322, + 114, + 480, + 180 + ], + "type": "text", + "content": "Table 9: Performance evaluation for typical categories at each difficult level. \"FF\" denotes FlowFormer [26] and \"FF++\" denotes FlowFormer++ [55]. We use the EPE as the evaluation metric." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 191, + 480, + 332 + ], + "lines": [ + { + "bbox": [ + 324, + 191, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 324, + 191, + 480, + 332 + ], + "type": "table", + "html": "
DifficultCategoryRAFT [56]FF [26]FF++ [55]
Hardperson9.828.728.61
tree9.799.7610.10
rhinoceros8.907.076.78
Mediumbench4.293.603.68
globe3.432.522.51
lantern2.441.921.95
Easynotebook1.080.570.84
coaster0.860.870.79
painting0.310.300.26
", + "image_path": "5929037ea47c8af4838bdc7db04b7d3c44cf9ee5ac0daa9e2dfce66fc9cb2674.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 479, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 479, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 479, + 405 + ], + "type": "text", + "content": "guess it might be the limited temporal receptive field of these two methods (only 8 frames) that makes the accumulated error quickly increase when the stride is too small. Some qualitative results are shown in Fig. 6. BlinkVision contains many challenging cases that cannot be handled by the SOTA methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 409, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 529 + ], + "type": "text", + "content": "We show the performance of event-based methods in Table 5. We find that event-based approaches perform particularly poorly. In addition to the reasons for insufficient training data and model design, we deduce that the receptive field of event-based point-tracking methods is relatively small and therefore cannot perform well on the task of tracking arbitrary points. To verify the claim, we replace the input of grid sampled positions with ORB [52] feature points, denoted by “*” in the table. The new results perform even better, validating our ideas. Although the new results cannot be strictly compared with image-based methods, it performs better under extreme frame interval, i.e., “Stride-8”, which shows the large potential of event-based methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 552, + 217, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 217, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 217, + 562 + ], + "type": "text", + "content": "4.3 Scene Flow" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 576, + 479, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 479, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 479, + 601 + ], + "type": "text", + "content": "Experimental Setup. We use the same pairs as the optical flow benchmark and we follow [38,57] to use 2D EPE and 3D EPE as the evaluation metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Results. We benchmark two state-of-the-art image-based methods, i.e., RAFT-3D [57] and CamliFlow [38] as shown in Table 6 and show the results of event-based methods in Table 7. The conclusions of the optical flow benchmark apply to the scene flow task and present similar challenges for existing methods in this area. We show qualitative results in the supplementary materials." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 286, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 286, + 128 + ], + "type": "text", + "content": "4.4 Cross-dataset Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 214 + ], + "type": "text", + "content": "We also fine-tune existing optical flow methods on the BlinkVision training set and then evaluate them on existing representative benchmarks. The results are shown in Table 8. We observe that fine-tuning on BlinkVision brings significant improvement for both image-based methods (2.08 vs. 1.69) and event-based methods (2.38 vs. 1.75 and 9.58 vs. 2.37), which demonstrates the vast diversity of BlinkVision boosts the generalizability of these methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 236, + 362, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 362, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 362, + 249 + ], + "type": "text", + "content": "4.5 Performance Distribution on Categories" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 262, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 407 + ], + "type": "text", + "content": "Previous methods in optical flow mainly perform evaluations on biased and limited scenarios, which is not comprehensive and robust enough to demonstrate the ability of the methods for different categories of objects in different scenarios. Thanks to the vast diversity of data and semantic labels provided by BlinkVision, for the first time, we analyze the performance distribution of several image-based optical flow methods on different categories. The results are shown in Fig. 7. The average curve is imbalanced: hard categories usually include complex shapes (e.g., hammocks and shrubs) or with deformable motion (e.g., persons and animals). We thus split the categories into three levels of \"difficulty\" based on the average curve, and the performance evaluation for typical categories at each level is presented in Table 9. We believe such fine-grained analysis helps understand the generalization capacity of the methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 430, + 220, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 220, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 220, + 443 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 545 + ], + "type": "text", + "content": "We propose BlinkVision, a large-scale diversified benchmark for three types of correspondence tasks, i.e., optical flow, point tracking, and scene flow estimation using RGB frames and events. Extensive benchmarks on BlinkVision point to new challenges for existing image-based approaches and show that existing event-based approaches are far from fully unlocking the potential of event cameras. BlinkVision reveals new observations, challenges, and opportunities for future research into more robust visual systems such as human vision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 551, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 635 + ], + "type": "text", + "content": "Acknowledgment. This project was funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by Smart Traffic Fund PSRI/76/2311/PR, by RGC General Research Fund Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. This work was also partially supported by NSF of China (No. 61932003)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 140, + 449, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 449, + 150 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 449, + 150 + ], + "type": "text", + "content": "1. ActorCore. https://actorcore.realusion.com/, accessed: 2023-11-17 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 151, + 397, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 397, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 397, + 161 + ], + "type": "text", + "content": "2. Blender. https://www.blender.org/, accessed: 2023-11-17 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 162, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 481, + 182 + ], + "type": "text", + "content": "3. Evermotion Archinteriors Collection. https://evermotion.org/, accessed: 2023-11-17 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 183, + 394, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 394, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 394, + 194 + ], + "type": "text", + "content": "4. Mixamo. https://www MIXamo.com/, accessed: 2023-11-17 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 481, + 226 + ], + "type": "text", + "content": "5. Alzugaray, I., Chli, M.: Asynchronous multi-hypothesis tracking of features with event cameras. In: 2019 International Conference on 3D Vision (3DV). pp. 269-278. IEEE (2019) 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "text", + "content": "6. Alzugaray, I., Chli, M.: Haste: multi-hypothesis asynchronous speeded-up tracking of events. In: 31st British Machine Vision Virtual Conference (BMVC 2020). p. 744. ETH Zurich, Institute of Robotics and Intelligent Systems (2020) 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 261, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 481, + 281 + ], + "type": "text", + "content": "7. Barron, J.L., Fleet, D.J., Beauchemin, S.S.: Performance of optical flow techniques. International journal of computer vision 12, 43-77 (1994) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 282, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 481, + 313 + ], + "type": "text", + "content": "8. Bian, W., Huang, Z., Shi, X., Dong, Y., Li, Y., Li, H.: Context-TAP: Tracking Any Point Demands Spatial Context Features. arXiv preprint arXiv:2306.02000 (2023) 2, 11, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 481, + 358 + ], + "type": "text", + "content": "9. Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12. pp. 611-625. Springer (2012) 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 481, + 392 + ], + "type": "text", + "content": "0. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An Information-rich 3D Model Repository. arXiv preprint arXiv:1512.03012 (2015) 3, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 435 + ], + "type": "text", + "content": "1. Ding, Z., Zhao, R., Zhang, J., Gao, T., Xiong, R., Yu, Z., Huang, T.: Spatiotemporal Recurrent Networks for Event-based Optical Flow Estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 525-533 (2022) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "type": "text", + "content": "2. Doersch, C., Gupta, A., Markeeva, L., Recasens, A., Smaira, L., Aytar, Y., Carreira, J., Zisserman, A., Yang, Y.: Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems 35, 13610–13626 (2022) 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 512 + ], + "type": "text", + "content": "3. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020) 1, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "text", + "content": "4. Gehrig, D., Gehrig, M., Hidalgo-Carrio, J., Scaramuzza, D.: Video to Events: Recycling Video Datasets for Event Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3586-3595 (2020) 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "text", + "content": "5. Gehrig, M., Aarents, W., Gehrig, D., Scaramuzza, D.: DSEC: A Stereo Event Camera Dataset for Driving Scenarios. IEEE Robotics and Automation Letters 6(3), 4947-4954 (2021) 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "type": "text", + "content": "6. Gehrig, M., Millhäsler, M., Gehrig, D., Scaramuzza, D.: E-RAFT: Dense Optical Flow from Event Cameras. In: Proceedings of the International Conference on 3D Vision. pp. 197–206. IEEE (2021) 10, 11, 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "text", + "content": "7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012) 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "8. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., et al.: Kubric: A scalable dataset" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3749-3761 (2022) 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "text", + "content": "19. Grossberg, M.D., Nayar, S.K.: What is the space of camera response functions? In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. vol. 2, pp. II-602. IEEE (2003) 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "type": "text", + "content": "20. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) 3, 8, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "21. Harley, A.W., Fang, Z., Fragkiadaki, K.: Particle video revisited: Tracking through occlusions using point trajectories. In: European Conference on Computer Vision. pp. 59-75. Springer (2022) 4, 11, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "text", + "content": "22. Hidalgo-Carrió, J., Gallego, G., Scaramuzza, D.: Event-aided direct sparse odometry. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5781-5790 (2022) 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "text", + "content": "23. Hu, J., Chen, X., Feng, B., Li, G., Yang, L., Bao, H., Zhang, G., Cui, Z.: Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095 (2024) 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 330 + ], + "type": "text", + "content": "24. Hu, J., Mao, M., Bao, H., Zhang, G., Cui, Z.: Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems 36 (2024) 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 365 + ], + "type": "text", + "content": "25. Huang, Z., Pan, X., Pan, W., Bian, W., Xu, Y., Cheung, K.C., Zhang, G., Li, H.: Neuralmarker: A framework for learning general marker correspondence. ACM Transactions on Graphics (TOG) 41(6), 1-10 (2022) 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "text", + "content": "26. Huang, Z., Shi, X., Zhang, C., Wang, Q., Cheung, K.C., Qin, H., Dai, J., Li, H.: Flowformer: A Transformer Architecture for Optical Flow. In: Proceedings of the European Conference on Computer Vision. pp. 668-685. Springer (2022) 10, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "type": "text", + "content": "27. Huang, Z., Shi, X., Zhang, C., Wang, Q., Li, Y., Qin, H., Dai, J., Wang, X., Li, H.: FlowFormer: A Transformer Architecture and Its Masked Cost Volume Autoencoding for Optical Flow. arXiv preprint arXiv:2306.05442 (2023) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "type": "text", + "content": "28. Huang, Z., Zhou, H., Li, Y., Yang, B., Xu, Y., Zhou, X., Bao, H., Zhang, G., Li, H.: Vs-net: Voting with segmentation for visual localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6101-6111 (2021) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 471, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 504 + ], + "type": "text", + "content": "29. Jiang, S., Campbell, D., Lu, Y., Li, H., Hartley, R.: Learning to estimate hidden motions with global motion aggregation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9772-9781 (2021) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "30. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "31. Klein, G., Murray, D.: Parallel tracking and mapping for small ar workspaces. In: 2007 6th IEEE and ACM international symposium on mixed and augmented reality. pp. 225-234. IEEE (2007) 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "32. Klenk, S., Chui, J., Demmel, N., Cremers, D.: Tum-vie: The tum stereo visual-inertial event dataset. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8601-8608. IEEE (2021) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "33. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "34. Li, Y., Huang, Z., Chen, S., Shi, X., Li, H., Bao, H., Cui, Z., Zhang, G.: Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv preprint arXiv:2303.07716 (2023) 2, 3, 5, 7, 10, 11, 13" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 654 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Li, Y., Liu, X., Dong, W., Zhou, H., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Deltar: Depth estimation from a light-weight tof sensor and rgb image. In: European conference on computer vision. pp. 619-636. Springer (2022) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "36. Li, Y., Zhou, H., Yang, B., Zhang, Y., Cui, Z., Bao, H., Zhang, G.: Graph-based asynchronous event processing for rapid object recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 934-943 (2021) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "text", + "content": "37. Lin, S., Ma, Y., Guo, Z., Wen, B.: DVS-Voltmeter: Stochastic Process-Based Event Simulator for Dynamic Vision Sensors. In: Proceedings of the European Conference on Computer Vision. pp. 578-593. Springer (2022) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 258 + ], + "type": "text", + "content": "38. Liu, H., Lu, T., Xu, Y., Liu, J., Li, W., Chen, L.: Camliflow: Bidirectional cameralidar fusion for joint optical flow and scene flow estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5791-5801 (2022) 12, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "text", + "content": "39. Liu, X., Li, Y., Teng, Y., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. In: Proceedings of the IEEE/cvf international conference on computer vision. pp. 1-11 (2023) 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 346 + ], + "type": "text", + "content": "40. Liu, Y.L., Lai, W.S., Chen, Y.S., Kao, Y.L., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Single-image hdr reconstruction by learning to reverse the camera pipeline. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1651-1660 (2020) 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 347, + 482, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 389 + ], + "type": "text", + "content": "41. Luo, J., Huang, Z., Li, Y., Zhou, X., Zhang, G., Bao, H.: Niid-net: adapting surface normal knowledge for intrinsic image decomposition in indoor scenes. IEEE Transactions on Visualization and Computer Graphics 26(12), 3434-3445 (2020) 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 390, + 482, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 434 + ], + "type": "text", + "content": "42. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4040-4048 (2016) 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 434, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 479 + ], + "type": "text", + "content": "43. Mehl, L., Schmalfuss, J., Jahedi, A., Nalivayko, Y., Bruhn, A.: Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4981-4991 (2023) 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "text", + "content": "44. Messikommer, N., Fang, C., Gehrig, M., Scaramuzza, D.: Data-driven feature tracking for event cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5642-5651 (2023) 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 511, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 522 + ], + "type": "text", + "content": "45. Milner, D., Goodale, M.: The visual brain in action, vol. 27. OUP Oxford (2006) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 482, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 566 + ], + "type": "text", + "content": "46. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research 36(2), 142-149 (2017) 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 566, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 610 + ], + "type": "text", + "content": "47. Ni, J., Li, Y., Huang, Z., Li, H., Bao, H., Cui, Z., Zhang, G.: PATS: Patch Area Transportation with Subdivision for Local Feature Matching. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 17776-17786 (2023) 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 610, + 482, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 654 + ], + "type": "text", + "content": "48. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019) 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "49. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 203 + ], + "type": "text", + "content": "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "type": "text", + "content": "51. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: An Open Event Camera Simulator. In: Proceedings of the Conference on Robot Learning. pp. 969-982. PMLR (2018) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "type": "text", + "content": "52. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: Orb: An efficient alternative to sift or surf. In: 2011 International conference on computer vision. pp. 2564-2571. IEEE (2011) 11, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "text", + "content": "53. Rueckauer, B., Delbruck, T.: Evaluation of Event-based Algorithms for Optical Flow with Ground-truth from Inertial Measurement Sensor. Frontiers in neuroscience 10, 176 (2016) 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "54. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 374 + ], + "type": "text", + "content": "55. Shi, X., Huang, Z., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Flowformer++: Masked Cost Volume Autoencoding for Pretraining Optical Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1599-1610 (2023) 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "56. Teed, Z., Deng, J.: RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In: Proceedings of the European Conference on Computer Vision. pp. 402-419. Springer (2020) 10, 11, 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "57. Teed, Z., Deng, J.: Raft-3d: Scene flow using rigid-motion embeddings. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8375–8384 (2021) 12, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 439, + 481, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 470 + ], + "type": "text", + "content": "58. Wan, Z., Dai, Y., Mao, Y.: Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing 31, 7237-7251 (2022) 10, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 471, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 514 + ], + "type": "text", + "content": "59. Wan, Z., Mao, Y., Zhang, J., Dai, Y.: Rpeflow: Multimodal fusion of rgbpointcloud-event for joint optical flow and scene flow estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10030-10040 (2023) 3, 5, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 514, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 556 + ], + "type": "text", + "content": "60. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023) 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 556, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 590 + ], + "type": "text", + "content": "61. Yang, B., Huang, Z., Li, Y., Zhou, H., Li, H., Zhang, G., Bao, H.: Hybrid3d: learning 3d hybrid features with point clouds and multi-view images for point cloud registration. Science China Information Sciences 66(7), 172101 (2023) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 590, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 632 + ], + "type": "text", + "content": "62. Zheng, Y., Harley, A.W., Shen, B., Wetzstein, G., Guibas, L.J.: Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19855-19865 (2023) 3, 4, 11, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "63. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The Multivehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters 3(3), 2032-2039 (2018) 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 150 + ], + "type": "text", + "content": "64. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In: Kress-Gazit, H., Srinivasa, S.S., Howard, T., Atanasov, N. (eds.) Robotics: Science and Systems (2018) 3, 5" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "BlinkVision: A Benchmark for Correspondence Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file